diff --git a/.gitattributes b/.gitattributes index a1665104a489095f8893ff10cddebbafce3e2110..0aace96aa35cddf051055da9a5848d9190067608 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1703,3 +1703,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/PassionSR_[[:space:]]Post-Training[[:space:]]Quantization[[:space:]]with[[:space:]]Adaptive[[:space:]]Scale[[:space:]]in[[:space:]]One-Step[[:space:]]Diffusion[[:space:]]based[[:space:]]Image[[:space:]]Super-Resolution/12edb9f7-ce33-4011-ae44-3a97b6f825f6_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Patch[[:space:]]Matters_[[:space:]]Training-free[[:space:]]Fine-grained[[:space:]]Image[[:space:]]Caption[[:space:]]Enhancement[[:space:]]via[[:space:]]Local[[:space:]]Perception/ad7b7d61-e53d-436e-98be-4d8c90eb49b6_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/PatchDEMUX_[[:space:]]A[[:space:]]Certifiably[[:space:]]Robust[[:space:]]Framework[[:space:]]for[[:space:]]Multi-label[[:space:]]Classifiers[[:space:]]Against[[:space:]]Adversarial[[:space:]]Patches/b0e9e99e-06e6-468a-a8a3-673f58598de7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PatchDPO_[[:space:]]Patch-level[[:space:]]DPO[[:space:]]for[[:space:]]Finetuning-free[[:space:]]Personalized[[:space:]]Image[[:space:]]Generation/201fb178-5c6a-42cf-95fe-3e03de95821f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PatchGuard_[[:space:]]Adversarially[[:space:]]Robust[[:space:]]Anomaly[[:space:]]Detection[[:space:]]and[[:space:]]Localization[[:space:]]through[[:space:]]Vision[[:space:]]Transformers[[:space:]]and[[:space:]]Pseudo[[:space:]]Anomalies/6310723f-d72e-495d-87c7-6b59d06d7ca9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PatchVSR_[[:space:]]Breaking[[:space:]]Video[[:space:]]Diffusion[[:space:]]Resolution[[:space:]]Limits[[:space:]]with[[:space:]]Patch-wise[[:space:]]Video[[:space:]]Super-Resolution/7794bc7e-86d2-49a8-8adc-3430e39dd287_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Pathways[[:space:]]on[[:space:]]the[[:space:]]Image[[:space:]]Manifold_[[:space:]]Image[[:space:]]Editing[[:space:]]via[[:space:]]Video[[:space:]]Generation/f625c957-1581-449f-8cd5-463f4552272c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Patient-Level[[:space:]]Anatomy[[:space:]]Meets[[:space:]]Scanning-Level[[:space:]]Physics_[[:space:]]Personalized[[:space:]]Federated[[:space:]]Low-Dose[[:space:]]CT[[:space:]]Denoising[[:space:]]Empowered[[:space:]]by[[:space:]]Large[[:space:]]Language[[:space:]]Model/b8f485b4-2d7f-474b-acf2-8dc4a4e3047c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Pattern[[:space:]]Analogies_[[:space:]]Learning[[:space:]]to[[:space:]]Perform[[:space:]]Programmatic[[:space:]]Image[[:space:]]Edits[[:space:]]by[[:space:]]Analogy/86d3746e-1656-4dc5-ab96-b8f696f1d699_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Pay[[:space:]]Attention[[:space:]]to[[:space:]]the[[:space:]]Foreground[[:space:]]in[[:space:]]Object-Centric[[:space:]]Learning/b4bf29e9-ae1b-4c69-a186-15cc519fb707_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PerLA_[[:space:]]Perceptive[[:space:]]3D[[:space:]]Language[[:space:]]Assistant/54882f4a-04ec-4ab3-b66a-a66c8e83acf5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Percept,[[:space:]]Memory,[[:space:]]and[[:space:]]Imagine_[[:space:]]World[[:space:]]Feature[[:space:]]Simulating[[:space:]]for[[:space:]]Open-Domain[[:space:]]Unknown[[:space:]]Object[[:space:]]Detection/b34e32e0-849d-4c9e-8711-6ce6de1dc89c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Perception[[:space:]]Tokens[[:space:]]Enhance[[:space:]]Visual[[:space:]]Reasoning[[:space:]]in[[:space:]]Multimodal[[:space:]]Language[[:space:]]Models/a1c25909-0376-4a01-a3cb-0e20c3229024_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Perceptual[[:space:]]Inductive[[:space:]]Bias[[:space:]]Is[[:space:]]What[[:space:]]You[[:space:]]Need[[:space:]]Before[[:space:]]Contrastive[[:space:]]Learning/cdaa1dd9-bb97-42b0-9c82-a5fe76697ec3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Perceptual[[:space:]]Video[[:space:]]Compression[[:space:]]with[[:space:]]Neural[[:space:]]Wrapping/4f847378-cfb2-4455-91dd-a70a0227cf83_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Perceptually[[:space:]]Accurate[[:space:]]3D[[:space:]]Talking[[:space:]]Head[[:space:]]Generation_[[:space:]]New[[:space:]]Definitions,[[:space:]]Speech-Mesh[[:space:]]Representation,[[:space:]]and[[:space:]]Evaluation[[:space:]]Metrics/f0ca3209-a68f-41a0-a209-e41298a3de25_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Period-LLM_[[:space:]]Extending[[:space:]]the[[:space:]]Periodic[[:space:]]Capability[[:space:]]of[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Model/96cc4bcd-798e-44b6-b645-30ba105eaf18_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Person[[:space:]]De-reidentification_[[:space:]]A[[:space:]]Variation-guided[[:space:]]Identity[[:space:]]Shift[[:space:]]Modeling/1180818e-c94d-44fa-b5df-b1fc8919ac06_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PersonaBooth_[[:space:]]Personalized[[:space:]]Text-to-Motion[[:space:]]Generation/62946df9-a553-4225-bd84-84b1acdb1a5f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PersonaHOI_[[:space:]]Effortlessly[[:space:]]Improving[[:space:]]Face[[:space:]]Personalization[[:space:]]in[[:space:]]Human-Object[[:space:]]Interaction[[:space:]]Generation/00e10d7f-1169-4801-84c3-c1efaaab378c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Personalized[[:space:]]Preference[[:space:]]Fine-tuning[[:space:]]of[[:space:]]Diffusion[[:space:]]Models/005a06ec-5d3e-419e-a758-dd73f721c5f6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Perturb-and-Revise_[[:space:]]Flexible[[:space:]]3D[[:space:]]Editing[[:space:]]with[[:space:]]Generative[[:space:]]Trajectories/11693e39-cb59-4d3c-8f05-aea184580288_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PhD_[[:space:]]A[[:space:]]ChatGPT-Prompted[[:space:]]Visual[[:space:]]Hallucination[[:space:]]Evaluation[[:space:]]Dataset/0cae22af-0103-455d-a35f-1a510beabf2e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Phoenix_[[:space:]]A[[:space:]]Motion-based[[:space:]]Self-Reflection[[:space:]]Framework[[:space:]]for[[:space:]]Fine-grained[[:space:]]Robotic[[:space:]]Action[[:space:]]Correction/497d038e-fa23-45f9-aee3-d3bfa187eaf5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PhyS-EdiT_[[:space:]]Physics-aware[[:space:]]Semantic[[:space:]]Image[[:space:]]Editing[[:space:]]with[[:space:]]Text[[:space:]]Description/7ee131d2-1c01-43c3-8f06-0c7615cc10d7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PhyT2V_[[:space:]]LLM-Guided[[:space:]]Iterative[[:space:]]Self-Refinement[[:space:]]for[[:space:]]Physics-Grounded[[:space:]]Text-to-Video[[:space:]]Generation/d3a4c2ae-5aa4-4658-a693-f3535a28ce44_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PhysAnimator_[[:space:]]Physics-Guided[[:space:]]Generative[[:space:]]Cartoon[[:space:]]Animation/d8df7b02-5bad-4fed-8afa-940086807a3d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PhysGen3D_[[:space:]]Crafting[[:space:]]a[[:space:]]Miniature[[:space:]]Interactive[[:space:]]World[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image/1bf0def3-a278-47bf-9ffd-13a93c05e545_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PhysVLM_[[:space:]]Enabling[[:space:]]Visual[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]Understand[[:space:]]Robotic[[:space:]]Physical[[:space:]]Reachability/14176105-f8ea-4fe1-b4ab-02889a03b21d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Physical[[:space:]]Plausibility-aware[[:space:]]Trajectory[[:space:]]Prediction[[:space:]]via[[:space:]]Locomotion[[:space:]]Embodiment/88abcf25-51f5-40e1-9600-4a2e38b6cc9a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PhysicsGen_[[:space:]]Can[[:space:]]Generative[[:space:]]Models[[:space:]]Learn[[:space:]]from[[:space:]]Images[[:space:]]to[[:space:]]Predict[[:space:]]Complex[[:space:]]Physical[[:space:]]Relations_/d8bec7dd-2170-41a3-a07a-e8a3813c711e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PillarHist_[[:space:]]A[[:space:]]Quantization-aware[[:space:]]Pillar[[:space:]]Feature[[:space:]]Encoder[[:space:]]based[[:space:]]on[[:space:]]Height-aware[[:space:]]Histogram/18874676-27a7-4b87-b788-f15decae2835_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Pioneering[[:space:]]4-Bit[[:space:]]FP[[:space:]]Quantization[[:space:]]for[[:space:]]Diffusion[[:space:]]Models_[[:space:]]Mixup-Sign[[:space:]]Quantization[[:space:]]and[[:space:]]Timestep-Aware[[:space:]]Fine-Tuning/6042cf5b-516c-4fe8-bfad-ecfc854fc855_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Pippo_[[:space:]]High-Resolution[[:space:]]Multi-View[[:space:]]Humans[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Image/2679a0a3-d3fd-49c2-aeba-9db847b48300_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Pixel-aligned[[:space:]]RGB-NIR[[:space:]]Stereo[[:space:]]Imaging[[:space:]]and[[:space:]]Dataset[[:space:]]for[[:space:]]Robot[[:space:]]Vision/640cd516-44b5-47ce-aa6a-b65d88b8da0a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Pixel-level[[:space:]]and[[:space:]]Semantic-level[[:space:]]Adjustable[[:space:]]Super-resolution_[[:space:]]A[[:space:]]Dual-LoRA[[:space:]]Approach/6dc2236b-e40d-40ee-98e0-34b3414c88ae_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PlanarSplatting_[[:space:]]Accurate[[:space:]]Planar[[:space:]]Surface[[:space:]]Reconstruction[[:space:]]in[[:space:]]3[[:space:]]Minutes/3af07a07-d377-440f-9ac8-47f29938e643_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Playing[[:space:]]the[[:space:]]Fool_[[:space:]]Jailbreaking[[:space:]]LLMs[[:space:]]and[[:space:]]Multimodal[[:space:]]LLMs[[:space:]]with[[:space:]]Out-of-Distribution[[:space:]]Strategy/013bcc47-e134-4c31-adf0-ced109b7aaa0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Plug-and-Play[[:space:]]Interpretable[[:space:]]Responsible[[:space:]]Text-to-Image[[:space:]]Generation[[:space:]]via[[:space:]]Dual-Space[[:space:]]Multi-facet[[:space:]]Concept[[:space:]]Control/2c24c8f1-1822-42ba-8659-5feab5b2d726_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Plug-and-Play[[:space:]]PPO_[[:space:]]An[[:space:]]Adaptive[[:space:]]Point[[:space:]]Prompt[[:space:]]Optimizer[[:space:]]Making[[:space:]]SAM[[:space:]]Greater/fb8f3fbf-9e73-432d-9130-bbe1d55d00ee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Plug-and-Play[[:space:]]Versatile[[:space:]]Compressed[[:space:]]Video[[:space:]]Enhancement/e1321f09-cc21-461b-8861-4f42cf1e3c7e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Point[[:space:]]Cloud[[:space:]]Upsampling[[:space:]]Using[[:space:]]Conditional[[:space:]]Diffusion[[:space:]]Module[[:space:]]with[[:space:]]Adaptive[[:space:]]Noise[[:space:]]Suppression/96fbf65a-6eae-49b1-9bdc-b206a300f6d4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Point[[:space:]]Clouds[[:space:]]Meets[[:space:]]Physics_[[:space:]]Dynamic[[:space:]]Acoustic[[:space:]]Field[[:space:]]Fitting[[:space:]]Network[[:space:]]for[[:space:]]Point[[:space:]]Cloud[[:space:]]Understanding/7aa2860d-04df-426a-bde9-40b0ed00d4aa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Point-Cache_[[:space:]]Test-time[[:space:]]Dynamic[[:space:]]and[[:space:]]Hierarchical[[:space:]]Cache[[:space:]]for[[:space:]]Robust[[:space:]]and[[:space:]]Generalizable[[:space:]]Point[[:space:]]Cloud[[:space:]]Analysis/b2000dba-b471-4537-8cba-488e6bd7eaf3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Point-to-Region[[:space:]]Loss[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]Point-Based[[:space:]]Crowd[[:space:]]Counting/7bc00c8f-f0e1-486b-af9f-cab0b77e9a84_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Point2RBox-v2_[[:space:]]Rethinking[[:space:]]Point-supervised[[:space:]]Oriented[[:space:]]Object[[:space:]]Detection[[:space:]]with[[:space:]]Spatial[[:space:]]Layout[[:space:]]Among[[:space:]]Instances/d3a3a010-6834-441e-beb3-f3f784483984_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PointLoRA_[[:space:]]Low-Rank[[:space:]]Adaptation[[:space:]]with[[:space:]]Token[[:space:]]Selection[[:space:]]for[[:space:]]Point[[:space:]]Cloud[[:space:]]Learning/d2184bbd-78d6-4278-aed5-2e0864e5914c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PointSR_[[:space:]]Self-Regularized[[:space:]]Point[[:space:]]Supervision[[:space:]]for[[:space:]]Drone-View[[:space:]]Object[[:space:]]Detection/f94115e7-2e5c-44d9-81fe-c067d1bde80f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PolarFree_[[:space:]]Polarization-based[[:space:]]Reflection-Free[[:space:]]Imaging/e1567ff5-b726-4ec2-baeb-81cce9b4c3bf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PolarNeXt_[[:space:]]Rethink[[:space:]]Instance[[:space:]]Segmentation[[:space:]]with[[:space:]]Polar[[:space:]]Representation/a5f92f60-4941-4386-a4db-1f556681a113_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Polarized[[:space:]]Color[[:space:]]Screen[[:space:]]Matting/4bc8f905-0a33-46ab-838e-6d058731820a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Poly-Autoregressive[[:space:]]Prediction[[:space:]]for[[:space:]]Modeling[[:space:]]Interactions/43d051a6-6821-4e3b-bd9a-a4c244857daa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Population[[:space:]]Normalization[[:space:]]for[[:space:]]Federated[[:space:]]Learning/b6b8182e-2812-44b6-99cb-82e1b12982cb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Pos3R_[[:space:]]6D[[:space:]]Pose[[:space:]]Estimation[[:space:]]for[[:space:]]Unseen[[:space:]]Objects[[:space:]]Made[[:space:]]Easy/162f30eb-880f-42fb-95ca-56f10c93b7f6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Pose[[:space:]]Priors[[:space:]]from[[:space:]]Language[[:space:]]Models/9f701210-ac31-4714-b2b0-3e6ed2c2dfd9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Pose-Guided[[:space:]]Temporal[[:space:]]Enhancement[[:space:]]for[[:space:]]Robust[[:space:]]Low-Resolution[[:space:]]Hand[[:space:]]Reconstruction/61fa7d4b-03fc-4f0b-b84c-af353e8bd6fc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PoseBH_[[:space:]]Prototypical[[:space:]]Multi-Dataset[[:space:]]Training[[:space:]]Beyond[[:space:]]Human[[:space:]]Pose[[:space:]]Estimation/6abac681-ec51-468a-ba82-3e1de6acd774_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PoseTraj_[[:space:]]Pose-Aware[[:space:]]Trajectory[[:space:]]Control[[:space:]]in[[:space:]]Video[[:space:]]Diffusion/8cdccae4-d7bc-4e3e-9398-db2b0c310587_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Positive2Negative_[[:space:]]Breaking[[:space:]]the[[:space:]]Information-Lossy[[:space:]]Barrier[[:space:]]in[[:space:]]Self-Supervised[[:space:]]Single[[:space:]]Image[[:space:]]Denoising/867e9811-fe1d-4217-8c41-8e621b6da43b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Post-pre-training[[:space:]]for[[:space:]]Modality[[:space:]]Alignment[[:space:]]in[[:space:]]Vision-Language[[:space:]]Foundation[[:space:]]Models/1196e091-a513-40f9-85ef-51604d8681b7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PosterMaker_[[:space:]]Towards[[:space:]]High-Quality[[:space:]]Product[[:space:]]Poster[[:space:]]Generation[[:space:]]with[[:space:]]Accurate[[:space:]]Text[[:space:]]Rendering/b07ff10a-9a3d-44d4-80c1-c5cc5cfdc9e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PosterO_[[:space:]]Structuring[[:space:]]Layout[[:space:]]Trees[[:space:]]to[[:space:]]Enable[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]Generalized[[:space:]]Content-Aware[[:space:]]Layout[[:space:]]Generation/28953876-24ed-4c02-8fbd-02d889f7bc8b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Potential[[:space:]]Field[[:space:]]Based[[:space:]]Deep[[:space:]]Metric[[:space:]]Learning/c5f58da8-3e8b-4fa3-accf-6f5c33627040_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Pow3R_[[:space:]]Empowering[[:space:]]Unconstrained[[:space:]]3D[[:space:]]Reconstruction[[:space:]]with[[:space:]]Camera[[:space:]]and[[:space:]]Scene[[:space:]]Priors/fb9435c5-f90b-40a7-b204-e0d114acb0c9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/PrEditor3D_[[:space:]]Fast[[:space:]]and[[:space:]]Precise[[:space:]]3D[[:space:]]Shape[[:space:]]Editing/0f1a120c-8d95-4d35-a553-6cd4183aad0f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Practical[[:space:]]Solutions[[:space:]]to[[:space:]]the[[:space:]]Relative[[:space:]]Pose[[:space:]]of[[:space:]]Three[[:space:]]Calibrated[[:space:]]Cameras/3fec933e-2dde-4be1-8ad1-28710156467f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Precise[[:space:]]Event[[:space:]]Spotting[[:space:]]in[[:space:]]Sports[[:space:]]Videos_[[:space:]]Solving[[:space:]]Long-Range[[:space:]]Dependency[[:space:]]and[[:space:]]Class[[:space:]]Imbalance/f70c89aa-2e9b-4112-89fc-f4837e98e688_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/201fb178-5c6a-42cf-95fe-3e03de95821f_content_list.json b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/201fb178-5c6a-42cf-95fe-3e03de95821f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..26ad84f419066d2152af1a53e2339b34a70b1fbc --- /dev/null +++ b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/201fb178-5c6a-42cf-95fe-3e03de95821f_content_list.json @@ -0,0 +1,1654 @@ +[ + { + "type": "text", + "text": "PatchDPO: Patch-level DPO for Finetuning-free Personalized Image Generation", + "text_level": 1, + "bbox": [ + 94, + 130, + 901, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qihan Huang $^{1}$ , Long Chan $^{2}$ , Jinlong Liu $^{2}$ , Wanggui He $^{2}$ , Hao Jiang $^{2}$ , Mingli Song $^{1}$ , Jie Song $^{1,\\dagger}$ , Zhejiang University, $^{2}$ Alibaba Group", + "bbox": [ + 124, + 179, + 870, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qh.huang,sjie,brooksong}@zju.edu.cn,", + "bbox": [ + 334, + 219, + 658, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chenlong0104.chen,aoshu.jh}@alibaba-inc.com,", + "bbox": [ + 300, + 236, + 692, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LJLwykqh@126.com, wanggui.hwg@taobao.com", + "bbox": [ + 321, + 255, + 669, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Finetuning-free personalized image generation can synthesize customized images without test-time finetuning, attracting wide research interest owing to its high efficiency. Current finetuning-free methods simply adopt a single training stage with a simple image reconstruction task, and they typically generate low-quality images inconsistent with the reference images during test-time. To mitigate this problem, inspired by the recent DPO (i.e., direct preference optimization) technique, this work proposes an additional training stage to improve the pre-trained personalized generation models. However, traditional DPO only determines the overall superiority or inferiority of two samples, which is not suitable for personalized image generation because the generated images are commonly inconsistent with the reference images only in some local image patches. To tackle this problem, this work proposes PatchDPO that estimates the quality of image patches within each generated image and accordingly trains the model. To this end, PatchDPO first leverages the pre-trained vision model with a proposed self-supervised training method to estimate the patch quality. Next, PatchDPO adopts a weighted training approach to train the model with the estimated patch quality, which rewards the image patches with high quality while penalizing the image patches with low quality. Experiment results demonstrate that PatchDPO significantly improves the performance of multiple pre-trained personalized generation models, and achieves state-of-the-art performance on both single-object and multi-object personalized image generation. Our code is available at https://github.com/hqhQAQ/PatchDPO.", + "bbox": [ + 88, + 335, + 483, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 819, + 220, + 834 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Personalized image generation methods have garnered significant research interest, which generate images based on", + "bbox": [ + 89, + 845, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c0789e2cf1171686b82db9e823b5599638a91611dce20a1690acf23290feacf.jpg", + "image_caption": [ + "Reference Image", + "Figure 1. The generated images (Images 1 & 2 & 3) are commonly inconsistent with the reference image only in some local image patches (marked in the red boxes)." + ], + "image_footnote": [], + "bbox": [ + 521, + 301, + 612, + 372 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5de881bc67fbc6a9ac0297e2db3c6fd55e80a4a3e95cbb2bf6d9601ed8236edb.jpg", + "image_caption": [ + "Image 1" + ], + "image_footnote": [], + "bbox": [ + 617, + 301, + 707, + 371 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ced75c2d0b999b59bccf18bdc5fac34f81c87d0f0719a37c23a32c5faa90fe8c.jpg", + "image_caption": [ + "Image 2" + ], + "image_footnote": [], + "bbox": [ + 712, + 301, + 803, + 371 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/23d82d44c4a5f4f6dc88074b088e5d9874fa3d28f5676eba2205feb7f93a8b87.jpg", + "image_caption": [ + "Image 3" + ], + "image_footnote": [], + "bbox": [ + 810, + 301, + 900, + 371 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reference images that define specific details of the desired output. The methodology in this domain is progressively evolving from a finetuning-based approach (e.g., DreamBooth [35], Custom Diffusion [22]) towards a finetuning-free approach (e.g., IP-Adapter [45], Subject-Diffusion [26], JeDi [46]), as finetuning-free methods eliminate the need for finetuning during test-time, significantly reducing usage costs.", + "bbox": [ + 509, + 457, + 906, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current finetuning-free methods typically employ only a single training stage on a large-scale image dataset. During this stage, the model is trained with a simple image reconstruction task (i.e., taking each image as reference image to reconstruct itself). When generating images in different scenes from the reference images in test-time, existing models often generate images of lower quality (i.e., inconsistent with the reference images in local details).", + "bbox": [ + 509, + 580, + 908, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by the recent DPO technique (i.e., direct preference optimization [32]) that leverages human feedback to optimize the pre-trained model, in this work we strive to devise an additional training stage for improving the pretrained personalized generation models. Specifically, the DPO technique assigns human preference to each sample, and trains the model to generate outputs that align more closely with human preferences. However, traditional DPO only judges the overall superiority or inferiority of two samples, which is not suitable for personalized image generation because generated images are usually inconsistent with the reference images only in some local areas, leading to inaccuracies when comparing the overall quality of two im", + "bbox": [ + 509, + 703, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 114, + 887, + 245, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18369", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ages. For example, as shown in Figure 1, the generated images (1 & 2 & 3) are inconsistent with the reference image only in the head & back & leg, respectively. In this case, traditional DPO roughly categorizes these images into superior and inferior samples, which would lead to the model's predictions incorrectly approaching the low-quality parts in the superior sample while moving away from the high-quality parts in the inferior sample.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this problem, this work proposes PatchDPO, which estimates the quality (preference level) of each patch in the generated image and accordingly optimizes the model. PatchDPO can provide feedback for the model in a more refined way, enabling the model to retain high-quality patches within images while moving away from low-quality patches. To this end, PatchDPO can be divided into three main stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization.", + "bbox": [ + 89, + 213, + 480, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the first stage (data construction), PatchDPO requires constructing a training dataset that includes multiple pairs of reference image and generated image. Our preliminary experiments Table 4 show that the complex details of objects in natural images and the confusion between objects and backgrounds hinder model training. Therefore, this work constructs a high-quality dataset for the PatchDPO training. First, this work generates the reference images using the open-source Stable Diffusion model [33] with text prompts instructing the background to be cleaner. Next, the corresponding generated images are synthesized using the pre-trained personalized generation model, with the aforementioned reference images as input.", + "bbox": [ + 89, + 352, + 482, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the second stage (patch quality estimation), traditional DPO would require extensive labeling costs to estimate the preference level of samples. In the case of comparing patch details between reference and generated images, thanks to the excellent pre-trained vision models, PatchDPO ingeniously utilizes the pre-trained vision model to extract patch features from reference and generated images, and evaluates the quality of patches in the generated images through patch-to-patch comparisons with those in the reference images. Besides, due to vision models (e.g., classification models pre-trained on ImageNet [8]) typically being better at extracting overall image features instead of patch features, this work proposes a self-supervised training method to improve patch features extraction of the original vision models. We conduct a quantitative evaluation on the HPatches dataset [3] (a dataset with images of the same object from different perspectives and scenes), demonstrating that our method efficiently and accurately extracts patch features for patch-to-patch comparisons.", + "bbox": [ + 89, + 550, + 482, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the third stage (model optimization), PatchDPO utilizes the patch quality estimated from the previous stage to further train the generation model. Specifically, PatchDPO adopts a weighted training approach, which assigns higher", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "training weights to the image patches with higher quality, while imposing penalties on those of lower quality. Furthermore, this work also incorporates the original reference image as the ground-truth generated image into training. In this manner, for the patches with lower quality in the real generated images, we increase the training weight for their corresponding patches in the ground-truth generated image, thus encouraging the model to correct its predictions for those low-quality patches.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We perform comprehensive experiments to validate the performance of PatchDPO. Specifically, we apply PatchDPO to multiple pre-trained personalized generation models (e.g., IP-Adapter, ELITE) on both single-object and multi-object personalized image generation. Experiment results on the DreamBooth dataset [35] and the Concept101 dataset [22] demonstrate that PatchDPO significantly improves the performance of pre-trained models. In particular, PatchDPO achieves state-of-the-art performance on both these two tasks.", + "bbox": [ + 511, + 227, + 903, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, the main contributions of this work can be summarized as follows:", + "bbox": [ + 511, + 378, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We construct a high-quality dataset to facilitate the PatchDPO training on personalized image generation.", + "- We propose a patch quality estimation method, which adopts the pre-trained vision models with a proposed self-supervised training approach for assessing the quality of patches in the generated images.", + "- Based on the estimated patch quality, we propose a weighted training approach for the PatchDPO training on personalized image generation.", + "- Experiment results show that PatchDPO achieves state-of-the-art performance on both single-object and multi-object personalized image generation." + ], + "bbox": [ + 511, + 410, + 903, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 609, + 653, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Personalized image generation. Early personalized image generation methods (e.g., DreamBooth [35], Textual Inversion [10], Cones [24], Mix-of-Show [11]) require finetuning the original diffusion model with the reference images. Recently, finetuning-free methods (e.g., IP-Adapter [45], ELITE [40], Subject-Diffusion [26], BLIP-Diffusion [23], InstantBooth [36], FastComposer [41], Taming Encoder [18], SSR-Encoder [47], JeDi [46]) emerge and attract more research interest as they require no finetuning during test-time and significantly reduce the usage cost. However, finetuning-free methods employ only a single training stage with a simple image reconstruction task, leading to low-quality generated images inconsistent with the reference images. PatchDPO compensates for this deficiency using an additional training stage for model optimization from the feedback.", + "bbox": [ + 511, + 636, + 903, + 876 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Aligning diffusion models. The model alignment meth", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18370", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d5abd16fab8a7d461bb5efe09fd5dace00400556c50759018de25efe65ec6639.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 88, + 431, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f74927aa156ed0225f6ed2ef6ba2683495700aee3a9df95e030baed69b370ce5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 88, + 900, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b7579261ef90829e745c0244e5c0566ff98d3b29e7d4b03f18c94eb77b153ace.jpg", + "image_caption": [ + "(2.1) Self-Supervised Training", + "Figure 2. PatchDPO has three stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization. The stage (2) is split into (2.1) self-supervised training and (2.2) patch-to-patch comparison. Besides, in (3), $\\hat{\\epsilon}_{\\theta}(\\mathbf{x}_{\\mathrm{ref}}(t))$ abbreviates $\\epsilon_{\\theta}(\\mathbf{x}_{\\mathrm{ref}}(t),\\mathbf{c}_{\\mathrm{text}},\\mathbf{x}_{\\mathrm{ref}},t)$ and $\\hat{\\epsilon}_{\\theta}(\\mathbf{x}_{\\mathrm{gen}}(t))$ abbreviates $\\epsilon_{\\theta}(\\mathbf{x}_{\\mathrm{gen}}(t),\\mathbf{c}_{\\mathrm{text}},\\mathbf{x}_{\\mathrm{ref}},t)$ . $\\tilde{p} (\\mathbf{x}_{\\mathrm{ref}})$ and $\\tilde{p} (\\mathbf{x}_{\\mathrm{gen}})$ are upsampled from $p(\\mathbf{x}_{\\mathrm{ref}})$ and $p(\\mathbf{x}_{\\mathrm{gen}})$ , respectively." + ], + "image_footnote": [], + "bbox": [ + 96, + 227, + 442, + 383 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7bddcca2645637dee3e6955d85dc42f96ddd34d45d88731c8ab415023cc7b840.jpg", + "image_caption": [ + "(3) Model Optimization" + ], + "image_footnote": [], + "bbox": [ + 444, + 229, + 898, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ods (e.g., RLHF, DPO) first emerged in the field of large language model (LLM). Specifically, RLHF (Reinforcement Learning from Human Feedback) [2, 27] trains a reward function from comparative data of model outputs to reflect human preferences, and adopts reinforcement learning to align it with the original model. DPO (Direct Preference Optimization) [32] simplifies RLHF by aligning the original model directly on the feedback data, but matching RLHF in performance. Recently, some methods (e.g., DPOK [9], DDPO [4], Diffusion-DPO [38], DRaFT [7], AlignProp [31]) have integrated RLHF or DPO into diffusion models for improving image aesthetic. However, these methods simply estimate the overall quality of the entire image, which is not suitable for personalized image generation because the generated images are usually inconsistent with the reference images only in some local image patches. Therefore, in this work, we propose PatchDPO, an advanced model alignment method for personalized image generation by estimating patch quality instead of image quality for model training.", + "bbox": [ + 88, + 470, + 483, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 784, + 228, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion model. Existing personalized image generation models utilize diffusion model [12, 34] as the base model. Diffusion model comprises two processes: a diffusion process which gradually adds noise into the original image with a Markov chain in $T$ steps, and a denoising process which predicts the noise to reconstruct the image with a", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "deep neural network. Detailedly, personalized image generation methods synthesize images simultaneously conditioned on the text prompt and the reference images. Commonly, $\\epsilon_{\\theta}$ denotes the deep neural network for noise prediction, and the training loss of personalized diffusion model is calculated as below $(\\| \\cdot \\|_2$ denotes the L2 norm):", + "bbox": [ + 511, + 470, + 906, + 561 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m s e}} = \\mathbb {E} _ {\\boldsymbol {x} _ {0}, \\boldsymbol {\\epsilon} \\in \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {\\mathbf {I}}), \\boldsymbol {c} _ {\\text {t e x t}}, \\boldsymbol {c} _ {\\text {i m g}}} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} (t), \\boldsymbol {c} _ {\\text {t e x t}}, \\boldsymbol {c} _ {\\text {i m g}}, t) \\| _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 580, + 908, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{x}_0$ denotes the original real image, $t\\in [0,T]$ denotes the time step in the diffusion process, $\\pmb {x}(t) = \\alpha_{t}\\pmb{x}_{0} + \\sigma_{t}\\pmb{\\epsilon}$ , and $\\alpha_{t},\\sigma_{t}$ are predefined weights for step $t$ in the diffusion process. $c_{\\mathrm{text}}$ denotes the text condition, and $c_{\\mathrm{img}}$ denotes the reference image. After training, the model can generate images by gradually denoising Gaussian noise in multiple steps.", + "bbox": [ + 511, + 611, + 906, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reinforcement learning from human feedback (RLHF). RLHF [2, 27] trains the model by maximizing the reward of model output, while regularizing the KL-divergence between it and the original model. Specifically, RLHF trains a reward function $r(\\boldsymbol{c}, \\boldsymbol{x})$ that estimates the human preference on the generated sample $\\boldsymbol{x}$ conditioned on $\\boldsymbol{c}$ . Next, let $p_{\\theta}$ denote the model being optimized, $p_{\\mathrm{ref}}$ denote the original model, the training target of RLHF is calculated as below (note that $\\beta$ is the hyper-parameter):", + "bbox": [ + 511, + 723, + 908, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {p _ {\\theta}} \\mathbb {E} _ {\\boldsymbol {c}, \\boldsymbol {x}} [ r (\\boldsymbol {c}, \\boldsymbol {x}) ] - \\beta \\mathbb {D} _ {\\mathrm {K L}} [ p _ {\\theta} (\\boldsymbol {x} | \\boldsymbol {c}) | | p _ {\\mathrm {r e f}} (\\boldsymbol {x} | \\boldsymbol {c}) ]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 880, + 906, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18371", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Direct preference optimization (DPO). Direct preference optimization simplifies RLHF by training the model directly from human preferences. Detailedly, let $\\boldsymbol{x}^w$ and $\\boldsymbol{x}^l$ denote the \"winning\" and \"losing\" samples generated from the condition $c$ , then DPO optimizes the model by aligning its output closer to $\\boldsymbol{x}^w$ while distancing it from $\\boldsymbol{x}^l$ , and the DPO loss [32] $\\mathcal{L}_{\\mathrm{DPO}}$ is calculated as below (note that $\\sigma(\\cdot)$ denotes the sigmoid function):", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D P O}} = - \\mathbb {E} _ {\\boldsymbol {c}, \\boldsymbol {x} ^ {w}, \\boldsymbol {x} ^ {l}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} ^ {w} \\mid \\boldsymbol {c})}{p _ {\\text {r e f}} (\\boldsymbol {x} ^ {w} \\mid \\boldsymbol {c})} - \\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} ^ {l} \\mid \\boldsymbol {c})}{p _ {\\text {r e f}} (\\boldsymbol {x} ^ {l} \\mid \\boldsymbol {c})}\\right) \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 224, + 482, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. PatchDPO", + "text_level": 1, + "bbox": [ + 89, + 256, + 205, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PatchDPO consists of three stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization.", + "bbox": [ + 89, + 282, + 482, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Data Construction", + "text_level": 1, + "bbox": [ + 89, + 320, + 269, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PatchDPO requires constructing a training dataset comprising multiple pairs of reference image and generated image (the generated image is synthesized by the personalized generation model being optimized). Our preliminary experiments in Table 4 demonstrate that natural images typically contain images of low-quality, which are not suitable for the task of PatchDPO. Detailedly, these low-quality images comprise complex object details with the confused foreground and background, easily misleading the model training. Therefore, in this work, we choose to construct a high-quality training dataset generated from the open-source generation model using three steps.", + "bbox": [ + 89, + 342, + 482, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, this work utilizes ChatGPT to generate the text prompt for each image. The text prompt is in the format of \"An {object} in the {background}\", where {object} and {background} are imagined by ChatGPT. Second, this work feeds the generated text prompts into the open-source text-to-image generation model (e.g., Stable Diffusion) to generate the reference images. Besides, in addition to the original text prompt, the generation model is also instructed to generate simple backgrounds for the mitigation of confusion between object and background. Finally, this work employs the target personalized generation model to generate images, with the aforementioned text prompts and the corresponding reference images as input.", + "bbox": [ + 89, + 523, + 482, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Patch Quality Estimation", + "text_level": 1, + "bbox": [ + 89, + 727, + 321, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Traditional DPO simply estimates the overall quality of the entire generated image, which does not provide sufficiently fine and accurate feedback for personalized image generation, thus resulting in deficient performance. To address this problem, PatchDPO estimates the quality of each patch in the generated image to acquire precise feedback for model optimization. Besides, traditional DPO requires a large amount of annotation cost to estimate the quality of the samples. Instead, the patch quality in personalized image generation is evaluated by comparing the patch details between", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reference images and generated images, which can be conducted using the pre-trained vision models. To this end, this work proposes a patch-to-patch comparison method to estimate the patch quality, and proposes a self-supervised training method for further improvement.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1. Patch-to-Patch Comparison", + "text_level": 1, + "bbox": [ + 511, + 176, + 754, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by ProtoPNet [5, 13, 15, 16, 44] and SFD2 [43] that extract patch features from the deep feature maps, PatchDPO estimates the patch quality with a patch-to-patch comparison on the patch features extracted from the deep feature maps.", + "bbox": [ + 511, + 195, + 903, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, let $\\mathbf{x}_{\\mathrm{ref}}$ denote the reference image, $\\mathbf{x}_{\\mathrm{gen}}$ denote the corresponding generated image, $f$ denote the pre-trained vision model, then $f(\\mathbf{x}_{\\mathrm{ref}}) \\in \\mathbb{R}^{H \\times W \\times D}$ and $f(\\mathbf{x}_{\\mathrm{gen}}) \\in \\mathbb{R}^{H \\times W \\times D}$ are the feature maps extracted by $f$ (note that $H, W, D$ are the size of the feature map). To ensure the generalizability of PatchDPO, this work acquires the last feature maps extracted from the vision models pretrained on ImageNet as $f(\\mathbf{x}_{\\mathrm{ref}})$ and $f(\\mathbf{x}_{\\mathrm{gen}})$ . Because the model does not change the spatial position of feature maps during feature extraction, $f(\\mathbf{x})[h, w] \\in \\mathbb{R}^D$ represents the features of the patch $\\mathbf{x}[h, w]$ within the image $\\mathbf{x}$ . Note that $\\mathbf{x}[h, w]$ denotes the patch in the $h$ -th row and the $w$ -th column of $\\mathbf{x}$ , as shown in the right side of Figure 2 (2.2). Next, the quality of each patch $\\mathbf{x}_{\\mathrm{gen}}[h, w]$ is estimated according to the existence of a patch similar to it in the reference image $\\mathbf{x}_{\\mathrm{ref}}$ . Detailedly, the patch quality $p(\\mathbf{x}_{\\mathrm{gen}}[h, w])$ of $\\mathbf{x}_{\\mathrm{gen}}[h, w]$ is calculated as the maximum similarity between $f(\\mathbf{x}_{\\mathrm{gen}})[h, w]$ with all elements in $f(\\mathbf{x}_{\\mathrm{ref}})$ :", + "bbox": [ + 511, + 272, + 906, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\boldsymbol {x} _ {\\text {g e n}} [ h, w ]\\right) = \\max _ {i, j} \\frac {f \\left(\\boldsymbol {x} _ {\\text {g e n}}\\right) [ h , w ] \\cdot f \\left(\\boldsymbol {x} _ {\\text {r e f}}\\right) [ i , j ]}{\\| f \\left(\\boldsymbol {x} _ {\\text {g e n}}\\right) [ h , w ] \\| \\| f \\left(\\boldsymbol {x} _ {\\text {r e f}}\\right) [ i , j ] \\|}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 561, + 903, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ , $j$ iterate over all the indexes of elements in $f(\\pmb{x}_{\\mathrm{ref}})$ . Therefore, higher $p(\\pmb{x}_{\\mathrm{gen}}[h,w])$ indicates that $\\pmb{x}_{\\mathrm{gen}}[h,w]$ is more consistent with the corresponding patch in the reference image $\\pmb{x}_{\\mathrm{ref}}$ .", + "bbox": [ + 511, + 606, + 903, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Verification by $S_{\\mathrm{patch}}$ . To guarantee precise patch quality estimation, this work conducts a quantitative evaluation of the extracted patch features using the HPatches dataset [3]. Dually, the HPatches dataset consists of images from 108 groups, where the images of the same group contain the same object from different perspectives and scenes. Besides, for the same group of images, the HPatches dataset annotates the spatial correspondence between their image patches. Based on this dataset, this work adopts a patch matching score $S_{\\mathrm{patch}}$ to evaluate the extracted patch features. $S_{\\mathrm{patch}}$ is calculated in three steps: (1) Extract the patch features of all images in the dataset, using the pretrained vision model $f$ . (2) For each patch in the image, predict its most similar patch (calculated from the patch features) in other images from the same group. (3) Calculate", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18372", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/295f4fd766677828e76cde2bc73a49ad4e4d4357682ad74c33531328ccc3eb78.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model/Layer1471012
ViT-Base70.476.474.868.668.4
ViT-Base (After Training)72.782.783.777.475.8
", + "bbox": [ + 94, + 88, + 480, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. ${S}_{\\text{patch }}\\left( \\% \\right)$ estimated on the HPatches dataset.", + "bbox": [ + 122, + 156, + 450, + 171 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the matching accuracy of each patch by comparing the predicted patch with the ground-truth patch, and $S_{\\mathrm{patch}}$ is finally calculated by averaging them.", + "bbox": [ + 89, + 186, + 482, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Table 1 (a ViT-Base model with 12 layers is adopted here), $S_{\\mathrm{patch}}$ of the patch features extracted from the last feature maps achieves only $68.4\\%$ , which is not sufficient for patch quality estimation.", + "bbox": [ + 89, + 231, + 483, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2. Self-Supervised Training", + "text_level": 1, + "bbox": [ + 89, + 300, + 312, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To facilitate the patch quality estimation, this work strives to improve $S_{\\mathrm{patch}}$ from two aspects: (1) Extract patch features from the shallow layers instead of the latest layer. (2) Finetune the vision model $f$ with self-supervised training.", + "bbox": [ + 89, + 318, + 483, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the first aspect, the deep neurons in the deep neural networks have large effective receptive fields [1, 14, 19, 20, 25], meaning that each element in the deeper feature map perceives a larger region of the image rather than the image patch in the corresponding location. Therefore, this work explores extracting patch features from feature maps at different depths. As shown in Table 1, patch features extracted from shallow feature maps have higher $S_{\\mathrm{patch}}$ in general, and in particular, the patch features extracted from the 4-th layer have the highest $S_{\\mathrm{patch}}$ .", + "bbox": [ + 88, + 378, + 483, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the second aspect, the performance of the aforementioned patch features extraction is still limited, because the used vision models are typically trained for other vision tasks (e.g., image classification), which focus on extracting the overall image features instead of the patch features. Consequently, this work proposes a self-supervised training method to finetune the pre-trained vision model $f$ , towards improving the performance of patch features extraction without expensive labeling costs. This self-supervised method augments the input image through spatial transformation (i.e., image rotation, image flip) and then constrains the patch features at corresponding positions of the input image and the augmented image to be close. Specifically, let $\\operatorname{Aug}(\\cdot)$ denote the augmentation operation (e.g., $\\operatorname{Aug}(x)$ is the augmented image of the input image $x$ ), then the loss function $\\mathcal{L}_{\\mathrm{self}}$ of self-supervised training is an MSE loss with a regularization term calculated as below ( $f_{\\mathrm{ref}}$ denotes the original model that is frozen during training):", + "bbox": [ + 89, + 530, + 483, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathcal {L} _ {\\text {s e l f}} = \\mathcal {L} _ {\\text {a u g}} + \\mathcal {L} _ {\\text {r e g}}. \\\\ \\mathcal {L} _ {\\text {a u g}} = \\| \\operatorname {A u g} (f (\\boldsymbol {x})) - f (\\operatorname {A u g} (\\boldsymbol {x})) \\| _ {2} ^ {2}. \\\\ \\mathcal {L} _ {\\text {r e g}} = \\| f (\\boldsymbol {x}) - f _ {\\text {r e f}} (\\boldsymbol {x}) \\| _ {2} ^ {2}. \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 819, + 482, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, the regularization term could avoid excessive de", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "viation of the finetuned model from the original model, stabilizing model training. Besides, this work chooses the dataset constructed in subsection 4.1 for this finetuning, because the finetuned vision model $f$ will be finally employed for patch quality estimation in this dataset. After the self-supervised training, as shown in Table 1, $S_{\\mathrm{patch}}$ of patch features at different layers have shown a significant improvement. We select the patch features with the highest $S_{\\mathrm{patch}}$ (83.7%, from the 7th layer) for patch quality estimation, ensuring the performance of PatchDPO training.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Model Optimization", + "text_level": 1, + "bbox": [ + 511, + 250, + 705, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the vision model $f$ finetuned from the previous stage, PatchDPO estimates the patch quality $p(\\pmb{x}_{\\mathrm{gen}}) \\in \\mathbb{R}^{H \\times W}$ for all generated images in the training dataset. Note that $p(\\pmb{x}_{\\mathrm{gen}})[h,w] = p(\\pmb{x}_{\\mathrm{gen}}[h,w]) \\in \\mathbb{R}$ is the patch quality of image patch $\\pmb{x}_{\\mathrm{gen}}[h,w]$ . Next, different from traditional DPO simply aligning with the superior samples while distancing from the inferior samples, PatchDPO leverages a weighted training method that adopts a more precise approach for model optimization. Specifically, PatchDPO trains the original personalized generation model with an image reconstruction task (reconstructing the generated image according to the reference image), and then assigns higher training weights to the image patches with higher quality, while assigning lower training weights to the image patches with lower quality.", + "bbox": [ + 511, + 271, + 906, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, only reconstructing the generated image can lead the model to still generate low-quality patches in the generated images, instead of generating the corresponding correct patches in the reference images. To address this problem, PatchDPO simultaneously involves a task of reconstructing the reference image using the reference image, as the ground-truth to correct the low-quality patches in the generated image. To this end, PatchDPO estimates the patch quality $p(\\pmb{x}_{\\mathrm{ref}}) \\in \\mathbb{R}^{H \\times W}$ for the reference image by comparing the features of each patch with those in the generated image, in the same manner as Equation 3. Like $p(\\pmb{x}_{\\mathrm{gen}})$ , each $p(\\pmb{x}_{\\mathrm{ref}})[h,w]$ reflects the extent to which $\\pmb{x}_{\\mathrm{ref}}[h,w]$ exists in the generated image $\\pmb{x}_{\\mathrm{gen}}$ . Therefore, $\\pmb{x}_{\\mathrm{ref}}[h,w]$ with lower $p(\\pmb{x}_{\\mathrm{ref}})[h,w]$ indicates that the patch $\\pmb{x}_{\\mathrm{ref}}[h,w]$ has low generation quality in the generated image, and the training weight of this patch should be increased in the task of reconstructing the ground-truth image (i.e., the reference image) from the reference image. Finally, the loss $\\mathcal{L}_{\\mathrm{PatchDPO}}$ of PatchDPO is calculated as below:", + "bbox": [ + 511, + 500, + 908, + 785 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {P a t c h D P O}} = \\underbrace {\\| [ \\boldsymbol {\\epsilon} _ {\\text {g e n}} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {\\text {g e n}} (t) , \\boldsymbol {c} _ {\\text {t e x t}} , \\boldsymbol {x} _ {\\text {r e f}} , t) ]} _ {\\text {R e c o n s t r u c t} \\boldsymbol {x} _ {\\text {g e n w i t h}} \\boldsymbol {x} _ {\\text {r e f}}} \\odot \\tilde {p} (\\boldsymbol {x} _ {\\text {g e n}}) \\| _ {2} ^ {2} \\\\ + \\| \\underbrace {\\left[ \\boldsymbol {\\epsilon} _ {\\mathrm {r e f}} - \\boldsymbol {\\epsilon} _ {\\theta} \\big (\\boldsymbol {x} _ {\\mathrm {r e f}} (t) , \\boldsymbol {c} _ {\\mathrm {t e x t}} , \\boldsymbol {x} _ {\\mathrm {r e f}} , t \\big) \\right]} _ {\\text {R e c o n s t r u c t} \\boldsymbol {x} _ {\\mathrm {r e f}} \\text {w i t h} \\boldsymbol {x} _ {\\mathrm {r e f}}} \\odot (1 - \\tilde {p} (\\boldsymbol {x} _ {\\mathrm {r e f}})) \\| _ {2} ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 800, + 915, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $\\tilde{p}(\\boldsymbol{x}_{\\mathrm{gen}})$ and $\\tilde{p}(\\boldsymbol{x}_{\\mathrm{ref}})$ are upsampled from the", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18373", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a2c00432b322dae192e08a7ccaea500e64f368af9739a4bf81c8008279b57362.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDINOCLIP-ICLIP-TAvg.
Real Images [35]0.7740.885N/AN/A
Textual Inversion [10]0.5690.7800.2550.535
DreamBooth (Imagen) [35]0.6960.8120.3060.605
DreamBooth (SD) [35]0.6680.8030.3050.592
Custom Diffusion [22]0.6430.7900.3050.579
Re-Imagen [6]0.6000.7400.2700.537
λ-ECLIPSE [29]0.6130.7830.3070.568
ELITE [40]0.6520.7620.2550.556
IP-Adapter [45]0.6080.8090.2740.564
IP-Adapter-Plus [45]0.6920.8260.2810.600
SSR-Encoder [47]0.6120.8210.3080.580
BLIP-Diffusion [23]0.5940.7790.3000.558
MS-Diffusion [39]0.6710.7920.3210.595
Subject-Diffusion [26]0.7110.7870.2930.597
JeDi [46]0.6790.8140.2930.595
PatchDPO0.7270.8380.2920.619
", + "bbox": [ + 93, + 88, + 480, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "original $p(\\pmb{x}_{\\mathrm{gen}})$ and $p(\\pmb{x}_{\\mathrm{ref}})$ with a normalization operation to constrain the values within [0, 1], which have the same height and width as original images ( $\\pmb{x}_{\\mathrm{gen}} \\& \\pmb{x}_{\\mathrm{ref}}$ ) and noise ( $\\epsilon_{\\mathrm{gen}} \\& \\epsilon_{\\mathrm{ref}}$ ). Besides, $\\odot$ denotes element-wise multiplication that assigns the weights $(\\tilde{p}(\\pmb{x}_{\\mathrm{gen}}) \\& 1 - \\tilde{p}(\\pmb{x}_{\\mathrm{ref}}))$ to the corresponding patches in the reconstruction losses.", + "bbox": [ + 89, + 436, + 483, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 542, + 222, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Our main experiments are conducted on the pre-trained IP-Adapter-Plus [45] with SDXL model [30] as the text-to-image diffusion model and OpenCLIP ViT-H/14 as the image encoder. Note that IP-AdapterPlus is the advanced version of the original IP-Adapter with significantly superior performance, by using the Resampler [17] to extract reference image features. This work only estimates the patch quality of object in the image to eliminate the interference from the background. The parameters of the SDXL model and image encoder are frozen, and only the parameters for projecting image features are trainable. During training, we adopt AdamW optimizer with a learning rate of 3e-5, and train the model on 8 GPUs for 30,000 steps with a batch size of 4 per GPU. Besides, the self-supervised training of patch feature extraction is conducted for 10 epochs with a learning rate of 1e-1.", + "bbox": [ + 89, + 575, + 482, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training dataset. This work constructs the training dataset as illustrated in subsection 4.1. Detailedly, the datasets for single-object & multi-object personalized generation both consist of 50,000 images. More details of multi-object personalized generation are in S2.1 of the appendix.", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f550578c10de5a95c292775ab5f789b2af90f050b19569c9db6afc33fd6914c2.jpg", + "table_caption": [ + "Table 2. Performance comparison for single-object personalized generation on DreamBench. The upper methods are finetuning-based methods, the bottom methods are finetuning-free methods, and bold font denotes the best result. \"SD\" is Stable Diffusion." + ], + "table_footnote": [], + "table_body": "
MethodDINOCLIP-ICLIP-TAvg.
Kosmos-G [28] (single image)0.6940.8470.2870.609
Emu2-Gen [37] (single image)0.7660.8500.2870.634
OmniGen [42] (single image)0.8010.8470.3010.650
PatchDPO (single image)0.8310.8800.2880.666
", + "bbox": [ + 514, + 88, + 903, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Performance comparison for single-object personalized generation on DreamBench using the evaluation setting of Kosmos-G. In this setting, only one image is preserved for each object in DreamBench.", + "bbox": [ + 511, + 191, + 906, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test benchmark. For single-object personalized image generation, we adopt the famous DreamBench [35] as the benchmark. For multi-object personalized image generation, we follow the Concept101 [22] benchmark that has evaluated many methods. Besides, MultiDreamBench [26] is also adopted for comparison with Subject-Diffusion.", + "bbox": [ + 511, + 261, + 906, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We follow previous methods to adopt three metrics (CLIP-T, CLIP-I, and DINO) for evaluation. Specifically, CLIP-T evaluates the similarity between the generated images and given text prompts; CLIP-I and DINO evaluate the similarity between the generated images and the reference images. 5 images are generated for each prompt to ensure the evaluation stability. Besides, Avg. (the average of three metrics) is also calculated for a comprehensive comparison.", + "bbox": [ + 511, + 358, + 908, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline methods. We compare our method with both finetuning-based methods (e.g., Textual Inversion [10], DreamBooth [35], Custom Diffusion [22]) and finetuning-free methods (e.g., SSR-Encoder [47], Subject-Diffusion [26], JeDi [46]).", + "bbox": [ + 511, + 500, + 905, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Single-Object Personalized Generation", + "text_level": 1, + "bbox": [ + 511, + 585, + 846, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct both quantitative and qualitative comparisons between our method and baseline methods.", + "bbox": [ + 511, + 607, + 905, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparisons. Table 2 & Table 3 demonstrates the quantitative results of different methods on DreamBench. Note that Table 2 uses the original setting following most existing methods, where DINO, CLIP-I are calculated by comparing the generated image and all images of the same object. Table 3 uses the evaluation setting following Kosmos-G [28], where only one image is preserved for each object, and DINO, CLIP-I are calculated by comparing the generated image and only this image. The results of baseline methods are from their paper.", + "bbox": [ + 511, + 643, + 905, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 2 & Table 3, PatchDPO realizes significantly superior image similarity (DINO, CLIP-I) to the SOTA personalized generation methods, because PatchDPO provides very detailed patch-level feedback on the model's generated images, and facilitates the model to correct the low-quality patches that are inconsistent with those from the reference images. Furthermore, PatchDPO", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18374", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be3d417c3005442e537ba319b4c9febdf26be1cd70c6983755237ee38ca99790.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons of different methods on single-object & multi-object personalized image generation." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 906, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves text similarity (CLIP-T) comparable to SOTA methods, because each pair of reference images and generated images in the training dataset is from the same text prompt. Therefore, aligning the low-quality patches of generated images with reference images does not decrease the similarity between the generated images and the text prompt. Finally, our method also surpasses existing methods in average performance (Avg.) by a large margin.", + "bbox": [ + 88, + 465, + 482, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparisons. The upper part of Figure 3 demonstrates the qualitative results of different methods on DreamBench. Compared to existing methods, PatchDPO excels in preserving the local details of the reference image, thus achieving generation of higher quality.", + "bbox": [ + 89, + 595, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparisons. The bottom of Figure 3 presents the qualitative results of different methods on Concept101, showing that PatchDPO can better preserve the details of reference images in multi-object personalized generation.", + "bbox": [ + 89, + 679, + 482, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, the experiment results on Concept101 and MultiDreamBench (see S2.2 of the appendix) show that PatchDPO can also improve the performance of original IP-Adapter-Plus on multi-object personalized generation, and achieves superior performance to existing personalized generation methods on both Concept101 & MultiDreamBench.", + "bbox": [ + 89, + 741, + 482, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablation Experiments", + "text_level": 1, + "bbox": [ + 89, + 845, + 294, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct the main ablation experiments of PatchDPO on DreamBench, as demonstrated in Table 4.", + "bbox": [ + 89, + 869, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training datasets. This work compares our training dataset $\\mathcal{D}_{\\mathrm{ours}}$ of 50,000 images constructed in subsection 4.1 with a natural dataset $\\mathcal{D}_{\\mathrm{natural}}$ . $\\mathcal{D}_{\\mathrm{natural}}$ consists of also 50,000 images (with one main object in the image) randomly selected from the open-source SA-1B dataset [21].", + "bbox": [ + 511, + 465, + 906, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 4, Combination (1) seriously degrades the image-alignment (DINO, CLIP-I) of model and Combination (2) benefits the model performance, indicating that the low-quality natural images (with chaotic object details & confused foreground and background, as shown in S3 of the appendix) hinder the training of personalized generation. Note that $\\mathcal{L}_{\\mathrm{mse}}$ is the loss of original diffusion model.", + "bbox": [ + 511, + 545, + 906, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training strategies. This work compares three training strategies corresponding to three losses. $\\mathcal{L}_{\\mathrm{mse}}$ is the loss of original diffusion model. $\\mathcal{L}_{\\mathrm{DPO}}$ is the loss of traditional DPO. Detailedly, we leverage the Diffusion-DPO loss [38] that directly adapts the original DPO loss to diffusion model. Finally, $\\mathcal{L}_{\\mathrm{PatchDPO}}$ is the loss of PatchDPO.", + "bbox": [ + 511, + 660, + 905, + 751 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 4, Combination (3) (traditional DPO) fails to improve the performance of the original model, because traditional DPO would wrongly reward the low-quality patches in the superior sample, while wrongly punishing the high-quality patches in the inferior sample. Instead, Combination (4) (PatchDPO) correctly rewards the high-quality patches and punishes the low-quality patches, thus achieving a huge performance improvement.", + "bbox": [ + 511, + 753, + 905, + 875 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Patch features extraction. This work estimates the ex", + "bbox": [ + 511, + 885, + 903, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18375", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/35ad410886383e49259b4e3ea90283717741ee54d00efd5b10d3dcbfb193e2e9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CombinationDINOCLIP-ICLIP-T
Original model0.6920.8260.281
(1) Lmse + Dnatural0.6580.8180.287
(2) Lmse + Dours0.7080.8300.292
(3) LDPO + Dours0.6760.8190.284
(4) LPatchDPO + Dours + Original f0.7190.8350.291
(5) LPatchDPO + Dours + Trained f0.7270.8380.292
", + "bbox": [ + 93, + 88, + 482, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation experiments.", + "bbox": [ + 192, + 212, + 379, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ac78cc49cd6cf3cb3006219d453be4ba0dde71d2eb53dbb981118e82181da6b.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 99, + 228, + 176, + 285 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/23188fd71bdf8a3fae6f976bbb2a5a162ffeb10dc49423a1a2fc4933baee76da.jpg", + "image_caption": [ + "Figure 4. The matching heatmaps of target patch (cat's head in the original image) on other images of the same cat." + ], + "image_footnote": [], + "bbox": [ + 183, + 228, + 254, + 285 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10446d6563efc8d6a4e318aee5715c0b717740ac9ad096482f8d0f5def591548.jpg", + "image_caption": [ + "Other Images" + ], + "image_footnote": [], + "bbox": [ + 256, + 228, + 331, + 285 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a728a3007201608d012921c26d3ab2fefa3c9b5d51301d1a0b4141180e706cac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 228, + 406, + 285 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c99c897d09c7912ceb9009cc24bb2fde78680cd9ab352286427c522bd032d45e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 228, + 480, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tracted patch features with $S_{\\mathrm{patch}}$ in subsection 4.2, and here we compare the extracted patch features of low $S_{\\mathrm{patch}}$ (68.4%, from the last feature map of original vision model $f$ ) and high $S_{\\mathrm{patch}}$ (83.7%, from the shallow feature map of vision model $f$ after self-supervised training).", + "bbox": [ + 89, + 352, + 482, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 4, Combination (5) (PatchDPO with high $S_{\\mathrm{patch}}$ ) surpasses Combination (4) (PatchDPO with low $S_{\\mathrm{patch}}$ ), implying that patch features with higher $S_{\\mathrm{patch}}$ contribute to more precise patch quality estimation and provide more accurate feedback for the generation model.", + "bbox": [ + 89, + 428, + 482, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional ablation experiments. Besides, we provide more ablation experiments (e.g., PatchDPO on different personalized generation models) in S2.3 of the appendix.", + "bbox": [ + 89, + 510, + 483, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Visualization Analysis", + "text_level": 1, + "bbox": [ + 89, + 563, + 295, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Patch matching of extracted patch features. Here, this work visualizes the patch matching results of the extracted patch features. Specifically, for the target patch in the original image, we acquire a matching heatmap $z \\in \\mathbb{R}^{H \\times W}$ by calculating the similarity between its features and all patch features of another image. Note that $z[h, w] \\in \\mathbb{R}$ indicates the similarity between the target patch and the patch in the $h$ -th row and the $w$ -th column of another image. Next, we visualize $z$ by resizing it to the same size as the image, and overlapping them. As shown in Figure 4, the matching heatmap $z$ accurately highlights the correct patch corresponding to the target patch, implying that the extracted patch features accurately represent the corresponding patch.", + "bbox": [ + 89, + 592, + 482, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Patch quality estimation. Here, we visualize the estimated patch quality, $p(\\pmb{x}_{\\mathrm{ref}}) \\in \\mathbb{R}^{H \\times W}$ and $p(\\pmb{x}_{\\mathrm{gen}}) \\in \\mathbb{R}^{H \\times W}$ , in the same manner as visualizing the matching heatmap $z$ . As demonstrated in Figure 6, the dark patches in the image are inconsistent with the corresponding patches in another image, indicating that our method can accurately estimate the patch quality.", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f6f7388876e114f8a5c31e716efaee34b4a510f3ab64c955437061d0f5a25da5.jpg", + "image_caption": [ + "Figure 5. Qualitative ablation experiment." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ccfb4bd5ae8ecef60e5ed57835631165a01216eadcd4ac865a58a4023c6aa66a.jpg", + "image_caption": [ + "Figure 6. Three samples of patch quality estimation." + ], + "image_footnote": [], + "bbox": [ + 516, + 226, + 903, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Images before/after PatchDPO. Here, this work demonstrates the generated images from the model before/after the PatchDPO training. As shown in Figure 5, the images from the model after PatchDPO exhibit significantly enhanced quality in terms of image-alignment, highlighting the effectiveness of PatchDPO.", + "bbox": [ + 511, + 377, + 906, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional visualization results. Furthermore, we provide more visualization results in S3 of the appendix for a comprehensive understanding of our method.", + "bbox": [ + 511, + 474, + 905, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 536, + 633, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose PatchDPO (patch-level direct preference optimization), which leverages an additional training stage to improve the pre-trained personalized generation models. PatchDPO estimates the quality of image patches within each generated image and accordingly provides detailed feedback to the models. Specifically, we propose a patch quality estimation method based on the pre-trained vision model finetuned with a self-supervised training method. Next, we propose a weighted training approach to train the model with the estimated patch quality, which rewards high-quality image patches while penalizing those of low quality. Experiment results demonstrate that PatchDPO achieves state-of-the-art performance on both single-object and multi-object personalized image generation. We hope our method and dataset can contribute to the community of personalized image generation.", + "bbox": [ + 511, + 561, + 906, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by Zhejiang Province High-Level Talents Special Support Program \"Leading Talent of Technological Innovation of Ten-Thousands Talents Program\" (No. 2022R52046), and Alibaba-Zhejiang University Joint Research Institute of Frontier Technologies.", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18376", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] André Araujo, Wade Norris, and Jack Sim. Computing receptive fields of convolutional neural networks. Distill, 2019. https://distill.pub/2019/computing-receptive-fields. 5", + "[2] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3", + "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR 2017, pages 3852-3861. IEEE, 2017. 2, 4", + "[4] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In *ICLR* 2024. OpenReview.net, 2024. 3", + "[5] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan Su. This looks like that: Deep learning for interpretable image recognition. In NeurIPS 2019, pages 8928-8939, 2019. 4", + "[6] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W. Cohen. Re-imagen: Retrieval-augmented text-to-image generator. In ICLR 2023. OpenReview.net, 2023. 6", + "[7] Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. In ICLR 2024. OpenReview.net, 2024. 3", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR 2009, pages 248-255. IEEE, 2009. 2", + "[9] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS 2023, 2023. 3", + "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR 2023. OpenReview.net, 2023. 2, 6", + "[11] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, Yixiao Ge, Ying Shan, and Mike Zheng Shou. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. In NeurIPS 2023, 2023. 2", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS 2020, 2020. 3", + "[13] Qihan Huang, Mengqi Xue, Wenqi Huang, Haofei Zhang, Jie Song, Yongcheng Jing, and Mingli Song. Evaluation and improvement of interpretability for self-explainable part-prototype networks. In ICCV 2023, pages 2011–2020. IEEE, 2023. 4", + "[14] Qihan Huang, Siming Fu, Jinlong Liu, Hao Jiang, Yipeng Yu, and Jie Song. Resolving multi-condition confusion for finetuning-free personalized image generation. arXiv preprint arXiv:2409.17920, 2024. 5" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Qihan Huang, Jie Song, Jingwen Hu, Haofei Zhang, Yong Wang, and Mingli Song. On the concept trustworthiness in concept bottleneck models. In AAAI 2024, pages 21161-21168, 2024. 4", + "[16] Qihan Huang, Jie Song, Mengqi Xue, Haofei Zhang, Bingde Hu, Huiqiong Wang, Hao Jiang, Xingen Wang, and Mingli Song. LG-CAV: train any concept activation vector with language guidance. In NeurIPS 2024, 2024. 4", + "[17] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and João Carreira. Perceiver: General perception with iterative attention. In ICML 2021, pages 4651-4664. PMLR, 2021. 6", + "[18] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2", + "[19] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 5", + "[20] Yongcheng Jing, Chongbin Yuan, Li Ju, Yiding Yang, Xinchao Wang, and Dacheng Tao. Deep graph reprogramming. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24345-24354, 2023. 5", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In ICCV 2023, pages 3992-4003. IEEE, 2023. 7", + "[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR 2023, pages 1931-1941. IEEE, 2023. 1, 2, 6", + "[23] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS 2023, 2023. 2, 6", + "[24] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. In ICML 2023, pages 21548-21566. PMLR, 2023. 2", + "[25] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard S. Zemel. Understanding the effective receptive field in deep convolutional neural networks. In NeurIPS, pages 4898-4906, 2016. 5", + "[26] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In SIGGRAPH 2024, page 25. ACM, 2024. 1, 2, 6", + "[27] Jacob Menick, Maja Trebacz, Vladimir Mikulik, John Aslanides, Francis Song, Martin Chadwick, Mia Glaese, Susannah Young, Lucy Campbell-Gillingham, Geoffrey Irving, et al. Teaching language models to support answers with verified quotes. arXiv preprint arXiv:2203.11147, 2022. 3" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18377", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR 2024. OpenReview.net, 2024. 6", + "[29] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. lambda-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 6", + "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6", + "[31] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 3", + "[32] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS 2023, 2023. 1, 3, 4", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR 2022, pages 10674-10685. IEEE, 2022. 2", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR 2022, pages 10674-10685. IEEE, 2022. 3", + "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR 2023, pages 22500-22510. IEEE, 2023. 1, 2, 6", + "[36] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In CVPR 2024, pages 8543-8552. IEEE, 2024. 2", + "[37] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR 2024, pages 14398-14409. IEEE, 2024. 6", + "[38] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR 2024, pages 8228-8238. IEEE, 2024. 3, 7", + "[39] X Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance. arXiv preprint arXiv:2406.07209, 2024. 6", + "[40] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV 2023, pages 15897-15907. IEEE, 2023. 2, 6" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2", + "[42] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 6", + "[43] Fei Xue, Ignas Budvytis, and Roberto Cipolla. SFD2: semantic-guided feature detection and description. In CVPR 2023, pages 5206-5216. IEEE, 2023. 4", + "[44] Mengqi Xue, Qihan Huang, Haofei Zhang, Jingwen Hu, Jie Song, Mingli Song, and Canghong Jin. Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition. In IJCAI 2024, pages 1516-1524. ijcai.org, 2024. 4", + "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 6", + "[46] Yu Zeng, Vishal M. Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In CVPR 2024, pages 6786-6795. IEEE, 2024. 1, 2, 6", + "[47] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR 2024, pages 8069–8078. IEEE, 2024. 2, 6" + ], + "bbox": [ + 514, + 92, + 903, + 513 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18378", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/201fb178-5c6a-42cf-95fe-3e03de95821f_model.json b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/201fb178-5c6a-42cf-95fe-3e03de95821f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b747aa1b001297f45a5503abcc2fbadbd4fe9145 --- /dev/null +++ b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/201fb178-5c6a-42cf-95fe-3e03de95821f_model.json @@ -0,0 +1,2244 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.131, + 0.903, + 0.153 + ], + "angle": 0, + "content": "PatchDPO: Patch-level DPO for Finetuning-free Personalized Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.18, + 0.871, + 0.217 + ], + "angle": 0, + "content": "Qihan Huang\\(^{1}\\), Long Chan\\(^{2}\\), Jinlong Liu\\(^{2}\\), Wanggui He\\(^{2}\\), Hao Jiang\\(^{2}\\), Mingli Song\\(^{1}\\), Jie Song\\(^{1,\\dagger}\\), Zhejiang University, \\(^{2}\\) Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.22, + 0.66, + 0.235 + ], + "angle": 0, + "content": "{qh.huang,sjie,brooksong}@zju.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.237, + 0.694, + 0.252 + ], + "angle": 0, + "content": "{chenlong0104.chen,aoshu.jh}@alibaba-inc.com," + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.256, + 0.671, + 0.269 + ], + "angle": 0, + "content": "LJLwykqh@126.com, wanggui.hwg@taobao.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.485, + 0.79 + ], + "angle": 0, + "content": "Finetuning-free personalized image generation can synthesize customized images without test-time finetuning, attracting wide research interest owing to its high efficiency. Current finetuning-free methods simply adopt a single training stage with a simple image reconstruction task, and they typically generate low-quality images inconsistent with the reference images during test-time. To mitigate this problem, inspired by the recent DPO (i.e., direct preference optimization) technique, this work proposes an additional training stage to improve the pre-trained personalized generation models. However, traditional DPO only determines the overall superiority or inferiority of two samples, which is not suitable for personalized image generation because the generated images are commonly inconsistent with the reference images only in some local image patches. To tackle this problem, this work proposes PatchDPO that estimates the quality of image patches within each generated image and accordingly trains the model. To this end, PatchDPO first leverages the pre-trained vision model with a proposed self-supervised training method to estimate the patch quality. Next, PatchDPO adopts a weighted training approach to train the model with the estimated patch quality, which rewards the image patches with high quality while penalizing the image patches with low quality. Experiment results demonstrate that PatchDPO significantly improves the performance of multiple pre-trained personalized generation models, and achieves state-of-the-art performance on both single-object and multi-object personalized image generation. Our code is available at https://github.com/hqhQAQ/PatchDPO." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.82, + 0.222, + 0.835 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.846, + 0.483, + 0.876 + ], + "angle": 0, + "content": "Personalized image generation methods have garnered significant research interest, which generate images based on" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.303, + 0.613, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.373, + 0.618, + 0.386 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.303, + 0.708, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.373, + 0.686, + 0.385 + ], + "angle": 0, + "content": "Image 1" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.303, + 0.805, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.373, + 0.784, + 0.385 + ], + "angle": 0, + "content": "Image 2" + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.303, + 0.901, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.831, + 0.373, + 0.88, + 0.385 + ], + "angle": 0, + "content": "Image 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.396, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Figure 1. The generated images (Images 1 & 2 & 3) are commonly inconsistent with the reference image only in some local image patches (marked in the red boxes)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.458, + 0.907, + 0.579 + ], + "angle": 0, + "content": "reference images that define specific details of the desired output. The methodology in this domain is progressively evolving from a finetuning-based approach (e.g., DreamBooth [35], Custom Diffusion [22]) towards a finetuning-free approach (e.g., IP-Adapter [45], Subject-Diffusion [26], JeDi [46]), as finetuning-free methods eliminate the need for finetuning during test-time, significantly reducing usage costs." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.581, + 0.909, + 0.702 + ], + "angle": 0, + "content": "Current finetuning-free methods typically employ only a single training stage on a large-scale image dataset. During this stage, the model is trained with a simple image reconstruction task (i.e., taking each image as reference image to reconstruct itself). When generating images in different scenes from the reference images in test-time, existing models often generate images of lower quality (i.e., inconsistent with the reference images in local details)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.704, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Inspired by the recent DPO technique (i.e., direct preference optimization [32]) that leverages human feedback to optimize the pre-trained model, in this work we strive to devise an additional training stage for improving the pretrained personalized generation models. Specifically, the DPO technique assigns human preference to each sample, and trains the model to generate outputs that align more closely with human preferences. However, traditional DPO only judges the overall superiority or inferiority of two samples, which is not suitable for personalized image generation because generated images are usually inconsistent with the reference images only in some local areas, leading to inaccuracies when comparing the overall quality of two im" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.246, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18369" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "ages. For example, as shown in Figure 1, the generated images (1 & 2 & 3) are inconsistent with the reference image only in the head & back & leg, respectively. In this case, traditional DPO roughly categorizes these images into superior and inferior samples, which would lead to the model's predictions incorrectly approaching the low-quality parts in the superior sample while moving away from the high-quality parts in the inferior sample." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.482, + 0.35 + ], + "angle": 0, + "content": "To tackle this problem, this work proposes PatchDPO, which estimates the quality (preference level) of each patch in the generated image and accordingly optimizes the model. PatchDPO can provide feedback for the model in a more refined way, enabling the model to retain high-quality patches within images while moving away from low-quality patches. To this end, PatchDPO can be divided into three main stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.549 + ], + "angle": 0, + "content": "In the first stage (data construction), PatchDPO requires constructing a training dataset that includes multiple pairs of reference image and generated image. Our preliminary experiments Table 4 show that the complex details of objects in natural images and the confusion between objects and backgrounds hinder model training. Therefore, this work constructs a high-quality dataset for the PatchDPO training. First, this work generates the reference images using the open-source Stable Diffusion model [33] with text prompts instructing the background to be cleaner. Next, the corresponding generated images are synthesized using the pre-trained personalized generation model, with the aforementioned reference images as input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.838 + ], + "angle": 0, + "content": "In the second stage (patch quality estimation), traditional DPO would require extensive labeling costs to estimate the preference level of samples. In the case of comparing patch details between reference and generated images, thanks to the excellent pre-trained vision models, PatchDPO ingeniously utilizes the pre-trained vision model to extract patch features from reference and generated images, and evaluates the quality of patches in the generated images through patch-to-patch comparisons with those in the reference images. Besides, due to vision models (e.g., classification models pre-trained on ImageNet [8]) typically being better at extracting overall image features instead of patch features, this work proposes a self-supervised training method to improve patch features extraction of the original vision models. We conduct a quantitative evaluation on the HPatches dataset [3] (a dataset with images of the same object from different perspectives and scenes), demonstrating that our method efficiently and accurately extracts patch features for patch-to-patch comparisons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In the third stage (model optimization), PatchDPO utilizes the patch quality estimated from the previous stage to further train the generation model. Specifically, PatchDPO adopts a weighted training approach, which assigns higher" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "training weights to the image patches with higher quality, while imposing penalties on those of lower quality. Furthermore, this work also incorporates the original reference image as the ground-truth generated image into training. In this manner, for the patches with lower quality in the real generated images, we increase the training weight for their corresponding patches in the ground-truth generated image, thus encouraging the model to correct its predictions for those low-quality patches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.378 + ], + "angle": 0, + "content": "We perform comprehensive experiments to validate the performance of PatchDPO. Specifically, we apply PatchDPO to multiple pre-trained personalized generation models (e.g., IP-Adapter, ELITE) on both single-object and multi-object personalized image generation. Experiment results on the DreamBooth dataset [35] and the Concept101 dataset [22] demonstrate that PatchDPO significantly improves the performance of pre-trained models. In particular, PatchDPO achieves state-of-the-art performance on both these two tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.905, + 0.409 + ], + "angle": 0, + "content": "To sum up, the main contributions of this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.411, + 0.904, + 0.441 + ], + "angle": 0, + "content": "- We construct a high-quality dataset to facilitate the PatchDPO training on personalized image generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.443, + 0.905, + 0.503 + ], + "angle": 0, + "content": "- We propose a patch quality estimation method, which adopts the pre-trained vision models with a proposed self-supervised training approach for assessing the quality of patches in the generated images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.504, + 0.904, + 0.549 + ], + "angle": 0, + "content": "- Based on the estimated patch quality, we propose a weighted training approach for the PatchDPO training on personalized image generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.55, + 0.904, + 0.595 + ], + "angle": 0, + "content": "- Experiment results show that PatchDPO achieves state-of-the-art performance on both single-object and multi-object personalized image generation." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.411, + 0.905, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.611, + 0.655, + 0.627 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.905, + 0.877 + ], + "angle": 0, + "content": "Personalized image generation. Early personalized image generation methods (e.g., DreamBooth [35], Textual Inversion [10], Cones [24], Mix-of-Show [11]) require finetuning the original diffusion model with the reference images. Recently, finetuning-free methods (e.g., IP-Adapter [45], ELITE [40], Subject-Diffusion [26], BLIP-Diffusion [23], InstantBooth [36], FastComposer [41], Taming Encoder [18], SSR-Encoder [47], JeDi [46]) emerge and attract more research interest as they require no finetuning during test-time and significantly reduce the usage cost. However, finetuning-free methods employ only a single training stage with a simple image reconstruction task, leading to low-quality generated images inconsistent with the reference images. PatchDPO compensates for this deficiency using an additional training stage for model optimization from the feedback." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Aligning diffusion models. The model alignment meth" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18370" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.432, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.089, + 0.901, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.228, + 0.443, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.385, + 0.378, + 0.399 + ], + "angle": 0, + "content": "(2.1) Self-Supervised Training" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.23, + 0.9, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.385, + 0.756, + 0.399 + ], + "angle": 0, + "content": "(3) Model Optimization" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.416, + 0.907, + 0.46 + ], + "angle": 0, + "content": "Figure 2. PatchDPO has three stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization. The stage (2) is split into (2.1) self-supervised training and (2.2) patch-to-patch comparison. Besides, in (3), \\(\\hat{\\epsilon}_{\\theta}(\\mathbf{x}_{\\mathrm{ref}}(t))\\) abbreviates \\(\\epsilon_{\\theta}(\\mathbf{x}_{\\mathrm{ref}}(t),\\mathbf{c}_{\\mathrm{text}},\\mathbf{x}_{\\mathrm{ref}},t)\\) and \\(\\hat{\\epsilon}_{\\theta}(\\mathbf{x}_{\\mathrm{gen}}(t))\\) abbreviates \\(\\epsilon_{\\theta}(\\mathbf{x}_{\\mathrm{gen}}(t),\\mathbf{c}_{\\mathrm{text}},\\mathbf{x}_{\\mathrm{ref}},t)\\). \\(\\tilde{p} (\\mathbf{x}_{\\mathrm{ref}})\\) and \\(\\tilde{p} (\\mathbf{x}_{\\mathrm{gen}})\\) are upsampled from \\(p(\\mathbf{x}_{\\mathrm{ref}})\\) and \\(p(\\mathbf{x}_{\\mathrm{gen}})\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.485, + 0.775 + ], + "angle": 0, + "content": "ods (e.g., RLHF, DPO) first emerged in the field of large language model (LLM). Specifically, RLHF (Reinforcement Learning from Human Feedback) [2, 27] trains a reward function from comparative data of model outputs to reflect human preferences, and adopts reinforcement learning to align it with the original model. DPO (Direct Preference Optimization) [32] simplifies RLHF by aligning the original model directly on the feedback data, but matching RLHF in performance. Recently, some methods (e.g., DPOK [9], DDPO [4], Diffusion-DPO [38], DRaFT [7], AlignProp [31]) have integrated RLHF or DPO into diffusion models for improving image aesthetic. However, these methods simply estimate the overall quality of the entire image, which is not suitable for personalized image generation because the generated images are usually inconsistent with the reference images only in some local image patches. Therefore, in this work, we propose PatchDPO, an advanced model alignment method for personalized image generation by estimating patch quality instead of image quality for model training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.229, + 0.801 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Diffusion model. Existing personalized image generation models utilize diffusion model [12, 34] as the base model. Diffusion model comprises two processes: a diffusion process which gradually adds noise into the original image with a Markov chain in \\( T \\) steps, and a denoising process which predicts the noise to reconstruct the image with a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.563 + ], + "angle": 0, + "content": "deep neural network. Detailedly, personalized image generation methods synthesize images simultaneously conditioned on the text prompt and the reference images. Commonly, \\(\\epsilon_{\\theta}\\) denotes the deep neural network for noise prediction, and the training loss of personalized diffusion model is calculated as below \\((\\| \\cdot \\|_2\\) denotes the L2 norm):" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.582, + 0.909, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {m s e}} = \\mathbb {E} _ {\\boldsymbol {x} _ {0}, \\boldsymbol {\\epsilon} \\in \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {\\mathbf {I}}), \\boldsymbol {c} _ {\\text {t e x t}}, \\boldsymbol {c} _ {\\text {i m g}}} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} (t), \\boldsymbol {c} _ {\\text {t e x t}}, \\boldsymbol {c} _ {\\text {i m g}}, t) \\| _ {2} ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.718 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_0\\) denotes the original real image, \\(t\\in [0,T]\\) denotes the time step in the diffusion process, \\(\\pmb {x}(t) = \\alpha_{t}\\pmb{x}_{0} + \\sigma_{t}\\pmb{\\epsilon}\\), and \\(\\alpha_{t},\\sigma_{t}\\) are predefined weights for step \\(t\\) in the diffusion process. \\(c_{\\mathrm{text}}\\) denotes the text condition, and \\(c_{\\mathrm{img}}\\) denotes the reference image. After training, the model can generate images by gradually denoising Gaussian noise in multiple steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.724, + 0.909, + 0.861 + ], + "angle": 0, + "content": "Reinforcement learning from human feedback (RLHF). RLHF [2, 27] trains the model by maximizing the reward of model output, while regularizing the KL-divergence between it and the original model. Specifically, RLHF trains a reward function \\( r(\\boldsymbol{c}, \\boldsymbol{x}) \\) that estimates the human preference on the generated sample \\( \\boldsymbol{x} \\) conditioned on \\( \\boldsymbol{c} \\). Next, let \\( p_{\\theta} \\) denote the model being optimized, \\( p_{\\mathrm{ref}} \\) denote the original model, the training target of RLHF is calculated as below (note that \\( \\beta \\) is the hyper-parameter):" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.881, + 0.907, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\max _ {p _ {\\theta}} \\mathbb {E} _ {\\boldsymbol {c}, \\boldsymbol {x}} [ r (\\boldsymbol {c}, \\boldsymbol {x}) ] - \\beta \\mathbb {D} _ {\\mathrm {K L}} [ p _ {\\theta} (\\boldsymbol {x} | \\boldsymbol {c}) | | p _ {\\mathrm {r e f}} (\\boldsymbol {x} | \\boldsymbol {c}) ]. \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18371" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "Direct preference optimization (DPO). Direct preference optimization simplifies RLHF by training the model directly from human preferences. Detailedly, let \\( \\boldsymbol{x}^w \\) and \\( \\boldsymbol{x}^l \\) denote the \"winning\" and \"losing\" samples generated from the condition \\( c \\), then DPO optimizes the model by aligning its output closer to \\( \\boldsymbol{x}^w \\) while distancing it from \\( \\boldsymbol{x}^l \\), and the DPO loss [32] \\( \\mathcal{L}_{\\mathrm{DPO}} \\) is calculated as below (note that \\( \\sigma(\\cdot) \\) denotes the sigmoid function):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.226, + 0.483, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D P O}} = - \\mathbb {E} _ {\\boldsymbol {c}, \\boldsymbol {x} ^ {w}, \\boldsymbol {x} ^ {l}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} ^ {w} \\mid \\boldsymbol {c})}{p _ {\\text {r e f}} (\\boldsymbol {x} ^ {w} \\mid \\boldsymbol {c})} - \\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} ^ {l} \\mid \\boldsymbol {c})}{p _ {\\text {r e f}} (\\boldsymbol {x} ^ {l} \\mid \\boldsymbol {c})}\\right) \\right]. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.257, + 0.206, + 0.273 + ], + "angle": 0, + "content": "4. PatchDPO" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.483, + 0.314 + ], + "angle": 0, + "content": "PatchDPO consists of three stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.321, + 0.27, + 0.336 + ], + "angle": 0, + "content": "4.1. Data Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.483, + 0.524 + ], + "angle": 0, + "content": "PatchDPO requires constructing a training dataset comprising multiple pairs of reference image and generated image (the generated image is synthesized by the personalized generation model being optimized). Our preliminary experiments in Table 4 demonstrate that natural images typically contain images of low-quality, which are not suitable for the task of PatchDPO. Detailedly, these low-quality images comprise complex object details with the confused foreground and background, easily misleading the model training. Therefore, in this work, we choose to construct a high-quality training dataset generated from the open-source generation model using three steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.721 + ], + "angle": 0, + "content": "First, this work utilizes ChatGPT to generate the text prompt for each image. The text prompt is in the format of \"An {object} in the {background}\", where {object} and {background} are imagined by ChatGPT. Second, this work feeds the generated text prompts into the open-source text-to-image generation model (e.g., Stable Diffusion) to generate the reference images. Besides, in addition to the original text prompt, the generation model is also instructed to generate simple backgrounds for the mitigation of confusion between object and background. Finally, this work employs the target personalized generation model to generate images, with the aforementioned text prompts and the corresponding reference images as input." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.323, + 0.744 + ], + "angle": 0, + "content": "4.2. Patch Quality Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Traditional DPO simply estimates the overall quality of the entire generated image, which does not provide sufficiently fine and accurate feedback for personalized image generation, thus resulting in deficient performance. To address this problem, PatchDPO estimates the quality of each patch in the generated image to acquire precise feedback for model optimization. Besides, traditional DPO requires a large amount of annotation cost to estimate the quality of the samples. Instead, the patch quality in personalized image generation is evaluated by comparing the patch details between" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "reference images and generated images, which can be conducted using the pre-trained vision models. To this end, this work proposes a patch-to-patch comparison method to estimate the patch quality, and proposes a self-supervised training method for further improvement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.177, + 0.755, + 0.192 + ], + "angle": 0, + "content": "4.2.1. Patch-to-Patch Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.196, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Inspired by ProtoPNet [5, 13, 15, 16, 44] and SFD2 [43] that extract patch features from the deep feature maps, PatchDPO estimates the patch quality with a patch-to-patch comparison on the patch features extracted from the deep feature maps." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.907, + 0.546 + ], + "angle": 0, + "content": "Specifically, let \\( \\mathbf{x}_{\\mathrm{ref}} \\) denote the reference image, \\( \\mathbf{x}_{\\mathrm{gen}} \\) denote the corresponding generated image, \\( f \\) denote the pre-trained vision model, then \\( f(\\mathbf{x}_{\\mathrm{ref}}) \\in \\mathbb{R}^{H \\times W \\times D} \\) and \\( f(\\mathbf{x}_{\\mathrm{gen}}) \\in \\mathbb{R}^{H \\times W \\times D} \\) are the feature maps extracted by \\( f \\) (note that \\( H, W, D \\) are the size of the feature map). To ensure the generalizability of PatchDPO, this work acquires the last feature maps extracted from the vision models pretrained on ImageNet as \\( f(\\mathbf{x}_{\\mathrm{ref}}) \\) and \\( f(\\mathbf{x}_{\\mathrm{gen}}) \\). Because the model does not change the spatial position of feature maps during feature extraction, \\( f(\\mathbf{x})[h, w] \\in \\mathbb{R}^D \\) represents the features of the patch \\( \\mathbf{x}[h, w] \\) within the image \\( \\mathbf{x} \\). Note that \\( \\mathbf{x}[h, w] \\) denotes the patch in the \\( h \\)-th row and the \\( w \\)-th column of \\( \\mathbf{x} \\), as shown in the right side of Figure 2 (2.2). Next, the quality of each patch \\( \\mathbf{x}_{\\mathrm{gen}}[h, w] \\) is estimated according to the existence of a patch similar to it in the reference image \\( \\mathbf{x}_{\\mathrm{ref}} \\). Detailedly, the patch quality \\( p(\\mathbf{x}_{\\mathrm{gen}}[h, w]) \\) of \\( \\mathbf{x}_{\\mathrm{gen}}[h, w] \\) is calculated as the maximum similarity between \\( f(\\mathbf{x}_{\\mathrm{gen}})[h, w] \\) with all elements in \\( f(\\mathbf{x}_{\\mathrm{ref}}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.562, + 0.905, + 0.597 + ], + "angle": 0, + "content": "\\[\np \\left(\\boldsymbol {x} _ {\\text {g e n}} [ h, w ]\\right) = \\max _ {i, j} \\frac {f \\left(\\boldsymbol {x} _ {\\text {g e n}}\\right) [ h , w ] \\cdot f \\left(\\boldsymbol {x} _ {\\text {r e f}}\\right) [ i , j ]}{\\| f \\left(\\boldsymbol {x} _ {\\text {g e n}}\\right) [ h , w ] \\| \\| f \\left(\\boldsymbol {x} _ {\\text {r e f}}\\right) [ i , j ] \\|}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.607, + 0.905, + 0.669 + ], + "angle": 0, + "content": "where \\(i\\), \\(j\\) iterate over all the indexes of elements in \\(f(\\pmb{x}_{\\mathrm{ref}})\\). Therefore, higher \\(p(\\pmb{x}_{\\mathrm{gen}}[h,w])\\) indicates that \\(\\pmb{x}_{\\mathrm{gen}}[h,w]\\) is more consistent with the corresponding patch in the reference image \\(\\pmb{x}_{\\mathrm{ref}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Verification by \\( S_{\\mathrm{patch}} \\). To guarantee precise patch quality estimation, this work conducts a quantitative evaluation of the extracted patch features using the HPatches dataset [3]. Dually, the HPatches dataset consists of images from 108 groups, where the images of the same group contain the same object from different perspectives and scenes. Besides, for the same group of images, the HPatches dataset annotates the spatial correspondence between their image patches. Based on this dataset, this work adopts a patch matching score \\( S_{\\mathrm{patch}} \\) to evaluate the extracted patch features. \\( S_{\\mathrm{patch}} \\) is calculated in three steps: (1) Extract the patch features of all images in the dataset, using the pretrained vision model \\( f \\). (2) For each patch in the image, predict its most similar patch (calculated from the patch features) in other images from the same group. (3) Calculate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18372" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.148 + ], + "angle": 0, + "content": "
Model/Layer1471012
ViT-Base70.476.474.868.668.4
ViT-Base (After Training)72.782.783.777.475.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.123, + 0.157, + 0.452, + 0.172 + ], + "angle": 0, + "content": "Table 1. \\( {S}_{\\text{patch }}\\left( \\% \\right) \\) estimated on the HPatches dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.187, + 0.483, + 0.232 + ], + "angle": 0, + "content": "the matching accuracy of each patch by comparing the predicted patch with the ground-truth patch, and \\( S_{\\mathrm{patch}} \\) is finally calculated by averaging them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.232, + 0.484, + 0.293 + ], + "angle": 0, + "content": "As shown in Table 1 (a ViT-Base model with 12 layers is adopted here), \\( S_{\\mathrm{patch}} \\) of the patch features extracted from the last feature maps achieves only \\( 68.4\\% \\), which is not sufficient for patch quality estimation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.301, + 0.313, + 0.317 + ], + "angle": 0, + "content": "4.2.2. Self-Supervised Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.379 + ], + "angle": 0, + "content": "To facilitate the patch quality estimation, this work strives to improve \\( S_{\\mathrm{patch}} \\) from two aspects: (1) Extract patch features from the shallow layers instead of the latest layer. (2) Finetune the vision model \\( f \\) with self-supervised training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.38, + 0.484, + 0.53 + ], + "angle": 0, + "content": "In the first aspect, the deep neurons in the deep neural networks have large effective receptive fields [1, 14, 19, 20, 25], meaning that each element in the deeper feature map perceives a larger region of the image rather than the image patch in the corresponding location. Therefore, this work explores extracting patch features from feature maps at different depths. As shown in Table 1, patch features extracted from shallow feature maps have higher \\( S_{\\mathrm{patch}} \\) in general, and in particular, the patch features extracted from the 4-th layer have the highest \\( S_{\\mathrm{patch}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.803 + ], + "angle": 0, + "content": "In the second aspect, the performance of the aforementioned patch features extraction is still limited, because the used vision models are typically trained for other vision tasks (e.g., image classification), which focus on extracting the overall image features instead of the patch features. Consequently, this work proposes a self-supervised training method to finetune the pre-trained vision model \\( f \\), towards improving the performance of patch features extraction without expensive labeling costs. This self-supervised method augments the input image through spatial transformation (i.e., image rotation, image flip) and then constrains the patch features at corresponding positions of the input image and the augmented image to be close. Specifically, let \\( \\operatorname{Aug}(\\cdot) \\) denote the augmentation operation (e.g., \\( \\operatorname{Aug}(x) \\) is the augmented image of the input image \\( x \\)), then the loss function \\( \\mathcal{L}_{\\mathrm{self}} \\) of self-supervised training is an MSE loss with a regularization term calculated as below (\\( f_{\\mathrm{ref}} \\) denotes the original model that is frozen during training):" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.82, + 0.483, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathcal {L} _ {\\text {s e l f}} = \\mathcal {L} _ {\\text {a u g}} + \\mathcal {L} _ {\\text {r e g}}. \\\\ \\mathcal {L} _ {\\text {a u g}} = \\| \\operatorname {A u g} (f (\\boldsymbol {x})) - f (\\operatorname {A u g} (\\boldsymbol {x})) \\| _ {2} ^ {2}. \\\\ \\mathcal {L} _ {\\text {r e g}} = \\| f (\\boldsymbol {x}) - f _ {\\text {r e f}} (\\boldsymbol {x}) \\| _ {2} ^ {2}. \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Here, the regularization term could avoid excessive de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "viation of the finetuned model from the original model, stabilizing model training. Besides, this work chooses the dataset constructed in subsection 4.1 for this finetuning, because the finetuned vision model \\( f \\) will be finally employed for patch quality estimation in this dataset. After the self-supervised training, as shown in Table 1, \\( S_{\\mathrm{patch}} \\) of patch features at different layers have shown a significant improvement. We select the patch features with the highest \\( S_{\\mathrm{patch}} \\) (83.7%, from the 7th layer) for patch quality estimation, ensuring the performance of PatchDPO training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.25, + 0.707, + 0.267 + ], + "angle": 0, + "content": "4.3. Model Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.907, + 0.499 + ], + "angle": 0, + "content": "With the vision model \\( f \\) finetuned from the previous stage, PatchDPO estimates the patch quality \\( p(\\pmb{x}_{\\mathrm{gen}}) \\in \\mathbb{R}^{H \\times W} \\) for all generated images in the training dataset. Note that \\( p(\\pmb{x}_{\\mathrm{gen}})[h,w] = p(\\pmb{x}_{\\mathrm{gen}}[h,w]) \\in \\mathbb{R} \\) is the patch quality of image patch \\( \\pmb{x}_{\\mathrm{gen}}[h,w] \\). Next, different from traditional DPO simply aligning with the superior samples while distancing from the inferior samples, PatchDPO leverages a weighted training method that adopts a more precise approach for model optimization. Specifically, PatchDPO trains the original personalized generation model with an image reconstruction task (reconstructing the generated image according to the reference image), and then assigns higher training weights to the image patches with higher quality, while assigning lower training weights to the image patches with lower quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.909, + 0.786 + ], + "angle": 0, + "content": "However, only reconstructing the generated image can lead the model to still generate low-quality patches in the generated images, instead of generating the corresponding correct patches in the reference images. To address this problem, PatchDPO simultaneously involves a task of reconstructing the reference image using the reference image, as the ground-truth to correct the low-quality patches in the generated image. To this end, PatchDPO estimates the patch quality \\( p(\\pmb{x}_{\\mathrm{ref}}) \\in \\mathbb{R}^{H \\times W} \\) for the reference image by comparing the features of each patch with those in the generated image, in the same manner as Equation 3. Like \\( p(\\pmb{x}_{\\mathrm{gen}}) \\), each \\( p(\\pmb{x}_{\\mathrm{ref}})[h,w] \\) reflects the extent to which \\( \\pmb{x}_{\\mathrm{ref}}[h,w] \\) exists in the generated image \\( \\pmb{x}_{\\mathrm{gen}} \\). Therefore, \\( \\pmb{x}_{\\mathrm{ref}}[h,w] \\) with lower \\( p(\\pmb{x}_{\\mathrm{ref}})[h,w] \\) indicates that the patch \\( \\pmb{x}_{\\mathrm{ref}}[h,w] \\) has low generation quality in the generated image, and the training weight of this patch should be increased in the task of reconstructing the ground-truth image (i.e., the reference image) from the reference image. Finally, the loss \\( \\mathcal{L}_{\\mathrm{PatchDPO}} \\) of PatchDPO is calculated as below:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.801, + 0.916, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {P a t c h D P O}} = \\underbrace {\\| [ \\boldsymbol {\\epsilon} _ {\\text {g e n}} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {\\text {g e n}} (t) , \\boldsymbol {c} _ {\\text {t e x t}} , \\boldsymbol {x} _ {\\text {r e f}} , t) ]} _ {\\text {R e c o n s t r u c t} \\boldsymbol {x} _ {\\text {g e n w i t h}} \\boldsymbol {x} _ {\\text {r e f}}} \\odot \\tilde {p} (\\boldsymbol {x} _ {\\text {g e n}}) \\| _ {2} ^ {2} \\\\ + \\| \\underbrace {\\left[ \\boldsymbol {\\epsilon} _ {\\mathrm {r e f}} - \\boldsymbol {\\epsilon} _ {\\theta} \\big (\\boldsymbol {x} _ {\\mathrm {r e f}} (t) , \\boldsymbol {c} _ {\\mathrm {t e x t}} , \\boldsymbol {x} _ {\\mathrm {r e f}} , t \\big) \\right]} _ {\\text {R e c o n s t r u c t} \\boldsymbol {x} _ {\\mathrm {r e f}} \\text {w i t h} \\boldsymbol {x} _ {\\mathrm {r e f}}} \\odot (1 - \\tilde {p} (\\boldsymbol {x} _ {\\mathrm {r e f}})) \\| _ {2} ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Note that \\(\\tilde{p}(\\boldsymbol{x}_{\\mathrm{gen}})\\) and \\(\\tilde{p}(\\boldsymbol{x}_{\\mathrm{ref}})\\) are upsampled from the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18373" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.354 + ], + "angle": 0, + "content": "
MethodDINOCLIP-ICLIP-TAvg.
Real Images [35]0.7740.885N/AN/A
Textual Inversion [10]0.5690.7800.2550.535
DreamBooth (Imagen) [35]0.6960.8120.3060.605
DreamBooth (SD) [35]0.6680.8030.3050.592
Custom Diffusion [22]0.6430.7900.3050.579
Re-Imagen [6]0.6000.7400.2700.537
λ-ECLIPSE [29]0.6130.7830.3070.568
ELITE [40]0.6520.7620.2550.556
IP-Adapter [45]0.6080.8090.2740.564
IP-Adapter-Plus [45]0.6920.8260.2810.600
SSR-Encoder [47]0.6120.8210.3080.580
BLIP-Diffusion [23]0.5940.7790.3000.558
MS-Diffusion [39]0.6710.7920.3210.595
Subject-Diffusion [26]0.7110.7870.2930.597
JeDi [46]0.6790.8140.2930.595
PatchDPO0.7270.8380.2920.619
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.42 + ], + "angle": 0, + "content": "Table 2. Performance comparison for single-object personalized generation on DreamBench. The upper methods are finetuning-based methods, the bottom methods are finetuning-free methods, and bold font denotes the best result. \"SD\" is Stable Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.484, + 0.529 + ], + "angle": 0, + "content": "original \\( p(\\pmb{x}_{\\mathrm{gen}}) \\) and \\( p(\\pmb{x}_{\\mathrm{ref}}) \\) with a normalization operation to constrain the values within [0, 1], which have the same height and width as original images (\\( \\pmb{x}_{\\mathrm{gen}} \\& \\pmb{x}_{\\mathrm{ref}} \\)) and noise (\\( \\epsilon_{\\mathrm{gen}} \\& \\epsilon_{\\mathrm{ref}} \\)). Besides, \\( \\odot \\) denotes element-wise multiplication that assigns the weights \\( (\\tilde{p}(\\pmb{x}_{\\mathrm{gen}}) \\& 1 - \\tilde{p}(\\pmb{x}_{\\mathrm{ref}})) \\) to the corresponding patches in the reconstruction losses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.544, + 0.223, + 0.56 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.819 + ], + "angle": 0, + "content": "Implementation details. Our main experiments are conducted on the pre-trained IP-Adapter-Plus [45] with SDXL model [30] as the text-to-image diffusion model and OpenCLIP ViT-H/14 as the image encoder. Note that IP-AdapterPlus is the advanced version of the original IP-Adapter with significantly superior performance, by using the Resampler [17] to extract reference image features. This work only estimates the patch quality of object in the image to eliminate the interference from the background. The parameters of the SDXL model and image encoder are frozen, and only the parameters for projecting image features are trainable. During training, we adopt AdamW optimizer with a learning rate of 3e-5, and train the model on 8 GPUs for 30,000 steps with a batch size of 4 per GPU. Besides, the self-supervised training of patch feature extraction is conducted for 10 epochs with a learning rate of 1e-1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Training dataset. This work constructs the training dataset as illustrated in subsection 4.1. Detailedly, the datasets for single-object & multi-object personalized generation both consist of 50,000 images. More details of multi-object personalized generation are in S2.1 of the appendix." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.905, + 0.181 + ], + "angle": 0, + "content": "
MethodDINOCLIP-ICLIP-TAvg.
Kosmos-G [28] (single image)0.6940.8470.2870.609
Emu2-Gen [37] (single image)0.7660.8500.2870.634
OmniGen [42] (single image)0.8010.8470.3010.650
PatchDPO (single image)0.8310.8800.2880.666
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.192, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Table 3. Performance comparison for single-object personalized generation on DreamBench using the evaluation setting of Kosmos-G. In this setting, only one image is preserved for each object in DreamBench." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.907, + 0.354 + ], + "angle": 0, + "content": "Test benchmark. For single-object personalized image generation, we adopt the famous DreamBench [35] as the benchmark. For multi-object personalized image generation, we follow the Concept101 [22] benchmark that has evaluated many methods. Besides, MultiDreamBench [26] is also adopted for comparison with Subject-Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.909, + 0.496 + ], + "angle": 0, + "content": "Evaluation metrics. We follow previous methods to adopt three metrics (CLIP-T, CLIP-I, and DINO) for evaluation. Specifically, CLIP-T evaluates the similarity between the generated images and given text prompts; CLIP-I and DINO evaluate the similarity between the generated images and the reference images. 5 images are generated for each prompt to ensure the evaluation stability. Besides, Avg. (the average of three metrics) is also calculated for a comprehensive comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.577 + ], + "angle": 0, + "content": "Baseline methods. We compare our method with both finetuning-based methods (e.g., Textual Inversion [10], DreamBooth [35], Custom Diffusion [22]) and finetuning-free methods (e.g., SSR-Encoder [47], Subject-Diffusion [26], JeDi [46])." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.586, + 0.848, + 0.602 + ], + "angle": 0, + "content": "5.1. Single-Object Personalized Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.906, + 0.639 + ], + "angle": 0, + "content": "We conduct both quantitative and qualitative comparisons between our method and baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Quantitative comparisons. Table 2 & Table 3 demonstrates the quantitative results of different methods on DreamBench. Note that Table 2 uses the original setting following most existing methods, where DINO, CLIP-I are calculated by comparing the generated image and all images of the same object. Table 3 uses the evaluation setting following Kosmos-G [28], where only one image is preserved for each object, and DINO, CLIP-I are calculated by comparing the generated image and only this image. The results of baseline methods are from their paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As shown in Table 2 & Table 3, PatchDPO realizes significantly superior image similarity (DINO, CLIP-I) to the SOTA personalized generation methods, because PatchDPO provides very detailed patch-level feedback on the model's generated images, and facilitates the model to correct the low-quality patches that are inconsistent with those from the reference images. Furthermore, PatchDPO" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18374" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.907, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.439, + 0.848, + 0.453 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons of different methods on single-object & multi-object personalized image generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.467, + 0.483, + 0.588 + ], + "angle": 0, + "content": "achieves text similarity (CLIP-T) comparable to SOTA methods, because each pair of reference images and generated images in the training dataset is from the same text prompt. Therefore, aligning the low-quality patches of generated images with reference images does not decrease the similarity between the generated images and the text prompt. Finally, our method also surpasses existing methods in average performance (Avg.) by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Qualitative comparisons. The upper part of Figure 3 demonstrates the qualitative results of different methods on DreamBench. Compared to existing methods, PatchDPO excels in preserving the local details of the reference image, thus achieving generation of higher quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.74 + ], + "angle": 0, + "content": "Qualitative comparisons. The bottom of Figure 3 presents the qualitative results of different methods on Concept101, showing that PatchDPO can better preserve the details of reference images in multi-object personalized generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.483, + 0.833 + ], + "angle": 0, + "content": "Furthermore, the experiment results on Concept101 and MultiDreamBench (see S2.2 of the appendix) show that PatchDPO can also improve the performance of original IP-Adapter-Plus on multi-object personalized generation, and achieves superior performance to existing personalized generation methods on both Concept101 & MultiDreamBench." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.295, + 0.863 + ], + "angle": 0, + "content": "5.2. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "We conduct the main ablation experiments of PatchDPO on DreamBench, as demonstrated in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Training datasets. This work compares our training dataset \\(\\mathcal{D}_{\\mathrm{ours}}\\) of 50,000 images constructed in subsection 4.1 with a natural dataset \\(\\mathcal{D}_{\\mathrm{natural}}\\). \\(\\mathcal{D}_{\\mathrm{natural}}\\) consists of also 50,000 images (with one main object in the image) randomly selected from the open-source SA-1B dataset [21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.652 + ], + "angle": 0, + "content": "In Table 4, Combination (1) seriously degrades the image-alignment (DINO, CLIP-I) of model and Combination (2) benefits the model performance, indicating that the low-quality natural images (with chaotic object details & confused foreground and background, as shown in S3 of the appendix) hinder the training of personalized generation. Note that \\(\\mathcal{L}_{\\mathrm{mse}}\\) is the loss of original diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.906, + 0.752 + ], + "angle": 0, + "content": "Training strategies. This work compares three training strategies corresponding to three losses. \\(\\mathcal{L}_{\\mathrm{mse}}\\) is the loss of original diffusion model. \\(\\mathcal{L}_{\\mathrm{DPO}}\\) is the loss of traditional DPO. Detailedly, we leverage the Diffusion-DPO loss [38] that directly adapts the original DPO loss to diffusion model. Finally, \\(\\mathcal{L}_{\\mathrm{PatchDPO}}\\) is the loss of PatchDPO." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.906, + 0.876 + ], + "angle": 0, + "content": "In Table 4, Combination (3) (traditional DPO) fails to improve the performance of the original model, because traditional DPO would wrongly reward the low-quality patches in the superior sample, while wrongly punishing the high-quality patches in the inferior sample. Instead, Combination (4) (PatchDPO) correctly rewards the high-quality patches and punishes the low-quality patches, thus achieving a huge performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Patch features extraction. This work estimates the ex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18375" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.203 + ], + "angle": 0, + "content": "
CombinationDINOCLIP-ICLIP-T
Original model0.6920.8260.281
(1) Lmse + Dnatural0.6580.8180.287
(2) Lmse + Dours0.7080.8300.292
(3) LDPO + Dours0.6760.8190.284
(4) LPatchDPO + Dours + Original f0.7190.8350.291
(5) LPatchDPO + Dours + Trained f0.7270.8380.292
" + }, + { + "type": "table_caption", + "bbox": [ + 0.194, + 0.213, + 0.38, + 0.227 + ], + "angle": 0, + "content": "Table 4. Ablation experiments." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.229, + 0.178, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.287, + 0.184, + 0.3 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.229, + 0.256, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.229, + 0.332, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.229, + 0.407, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.287, + 0.372, + 0.3 + ], + "angle": 0, + "content": "Other Images" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.229, + 0.482, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.312, + 0.483, + 0.34 + ], + "angle": 0, + "content": "Figure 4. The matching heatmaps of target patch (cat's head in the original image) on other images of the same cat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.429 + ], + "angle": 0, + "content": "tracted patch features with \\( S_{\\mathrm{patch}} \\) in subsection 4.2, and here we compare the extracted patch features of low \\( S_{\\mathrm{patch}} \\) (68.4%, from the last feature map of original vision model \\( f \\)) and high \\( S_{\\mathrm{patch}} \\) (83.7%, from the shallow feature map of vision model \\( f \\) after self-supervised training)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.483, + 0.504 + ], + "angle": 0, + "content": "In Table 4, Combination (5) (PatchDPO with high \\( S_{\\mathrm{patch}} \\)) surpasses Combination (4) (PatchDPO with low \\( S_{\\mathrm{patch}} \\)), implying that patch features with higher \\( S_{\\mathrm{patch}} \\) contribute to more precise patch quality estimation and provide more accurate feedback for the generation model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.556 + ], + "angle": 0, + "content": "Additional ablation experiments. Besides, we provide more ablation experiments (e.g., PatchDPO on different personalized generation models) in S2.3 of the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.564, + 0.297, + 0.58 + ], + "angle": 0, + "content": "5.3. Visualization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.79 + ], + "angle": 0, + "content": "Patch matching of extracted patch features. Here, this work visualizes the patch matching results of the extracted patch features. Specifically, for the target patch in the original image, we acquire a matching heatmap \\( z \\in \\mathbb{R}^{H \\times W} \\) by calculating the similarity between its features and all patch features of another image. Note that \\( z[h, w] \\in \\mathbb{R} \\) indicates the similarity between the target patch and the patch in the \\( h \\)-th row and the \\( w \\)-th column of another image. Next, we visualize \\( z \\) by resizing it to the same size as the image, and overlapping them. As shown in Figure 4, the matching heatmap \\( z \\) accurately highlights the correct patch corresponding to the target patch, implying that the extracted patch features accurately represent the corresponding patch." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Patch quality estimation. Here, we visualize the estimated patch quality, \\( p(\\pmb{x}_{\\mathrm{ref}}) \\in \\mathbb{R}^{H \\times W} \\) and \\( p(\\pmb{x}_{\\mathrm{gen}}) \\in \\mathbb{R}^{H \\times W} \\), in the same manner as visualizing the matching heatmap \\( z \\). As demonstrated in Figure 6, the dark patches in the image are inconsistent with the corresponding patches in another image, indicating that our method can accurately estimate the patch quality." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.209, + 0.836, + 0.223 + ], + "angle": 0, + "content": "Figure 5. Qualitative ablation experiment." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.227, + 0.905, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.348, + 0.866, + 0.363 + ], + "angle": 0, + "content": "Figure 6. Three samples of patch quality estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.907, + 0.468 + ], + "angle": 0, + "content": "Images before/after PatchDPO. Here, this work demonstrates the generated images from the model before/after the PatchDPO training. As shown in Figure 5, the images from the model after PatchDPO exhibit significantly enhanced quality in terms of image-alignment, highlighting the effectiveness of PatchDPO." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.521 + ], + "angle": 0, + "content": "Additional visualization results. Furthermore, we provide more visualization results in S3 of the appendix for a comprehensive understanding of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.537, + 0.634, + 0.552 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.804 + ], + "angle": 0, + "content": "In this work, we propose PatchDPO (patch-level direct preference optimization), which leverages an additional training stage to improve the pre-trained personalized generation models. PatchDPO estimates the quality of image patches within each generated image and accordingly provides detailed feedback to the models. Specifically, we propose a patch quality estimation method based on the pre-trained vision model finetuned with a self-supervised training method. Next, we propose a weighted training approach to train the model with the estimated patch quality, which rewards high-quality image patches while penalizing those of low quality. Experiment results demonstrate that PatchDPO achieves state-of-the-art performance on both single-object and multi-object personalized image generation. We hope our method and dataset can contribute to the community of personalized image generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by Zhejiang Province High-Level Talents Special Support Program \"Leading Talent of Technological Innovation of Ten-Thousands Talents Program\" (No. 2022R52046), and Alibaba-Zhejiang University Joint Research Institute of Frontier Technologies." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18376" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] André Araujo, Wade Norris, and Jack Sim. Computing receptive fields of convolutional neural networks. Distill, 2019. https://distill.pub/2019/computing-receptive-fields. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.227 + ], + "angle": 0, + "content": "[2] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR 2017, pages 3852-3861. IEEE, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[4] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In *ICLR* 2024. OpenReview.net, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[5] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan Su. This looks like that: Deep learning for interpretable image recognition. In NeurIPS 2019, pages 8928-8939, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[6] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W. Cohen. Re-imagen: Retrieval-augmented text-to-image generator. In ICLR 2023. OpenReview.net, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.482, + 0.472 + ], + "angle": 0, + "content": "[7] Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. In ICLR 2024. OpenReview.net, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.475, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR 2009, pages 248-255. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[9] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS 2023, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.589, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR 2023. OpenReview.net, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[11] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, Yixiao Ge, Ying Shan, and Mike Zheng Shou. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. In NeurIPS 2023, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS 2020, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.841 + ], + "angle": 0, + "content": "[13] Qihan Huang, Mengqi Xue, Wenqi Huang, Haofei Zhang, Jie Song, Yongcheng Jing, and Mingli Song. Evaluation and improvement of interpretability for self-explainable part-prototype networks. In ICCV 2023, pages 2011–2020. IEEE, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Qihan Huang, Siming Fu, Jinlong Liu, Hao Jiang, Yipeng Yu, and Jie Song. Resolving multi-condition confusion for finetuning-free personalized image generation. arXiv preprint arXiv:2409.17920, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[15] Qihan Huang, Jie Song, Jingwen Hu, Haofei Zhang, Yong Wang, and Mingli Song. On the concept trustworthiness in concept bottleneck models. In AAAI 2024, pages 21161-21168, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[16] Qihan Huang, Jie Song, Mengqi Xue, Haofei Zhang, Bingde Hu, Huiqiong Wang, Hao Jiang, Xingen Wang, and Mingli Song. LG-CAV: train any concept activation vector with language guidance. In NeurIPS 2024, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[17] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and João Carreira. Perceiver: General perception with iterative attention. In ICML 2021, pages 4651-4664. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[18] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[19] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[20] Yongcheng Jing, Chongbin Yuan, Li Ju, Yiding Yang, Xinchao Wang, and Dacheng Tao. Deep graph reprogramming. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24345-24354, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In ICCV 2023, pages 3992-4003. IEEE, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR 2023, pages 1931-1941. IEEE, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[23] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS 2023, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[24] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. In ICML 2023, pages 21548-21566. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[25] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard S. Zemel. Understanding the effective receptive field in deep convolutional neural networks. In NeurIPS, pages 4898-4906, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[26] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In SIGGRAPH 2024, page 25. ACM, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Jacob Menick, Maja Trebacz, Vladimir Mikulik, John Aslanides, Francis Song, Martin Chadwick, Mia Glaese, Susannah Young, Lucy Campbell-Gillingham, Geoffrey Irving, et al. Teaching language models to support answers with verified quotes. arXiv preprint arXiv:2203.11147, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18377" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[28] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR 2024. OpenReview.net, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[29] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. lambda-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[31] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[32] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS 2023, 2023. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR 2022, pages 10674-10685. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR 2022, pages 10674-10685. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR 2023, pages 22500-22510. IEEE, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[36] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In CVPR 2024, pages 8543-8552. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[37] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR 2024, pages 14398-14409. IEEE, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[38] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR 2024, pages 8228-8238. IEEE, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[39] X Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance. arXiv preprint arXiv:2406.07209, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[40] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV 2023, pages 15897-15907. IEEE, 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[41] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[42] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[43] Fei Xue, Ignas Budvytis, and Roberto Cipolla. SFD2: semantic-guided feature detection and description. In CVPR 2023, pages 5206-5216. IEEE, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.249, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[44] Mengqi Xue, Qihan Huang, Haofei Zhang, Jingwen Hu, Jie Song, Mingli Song, and Canghong Jin. Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition. In IJCAI 2024, pages 1516-1524. ijcai.org, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.32, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.376, + 0.905, + 0.443 + ], + "angle": 0, + "content": "[46] Yu Zeng, Vishal M. Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In CVPR 2024, pages 6786-6795. IEEE, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.446, + 0.905, + 0.514 + ], + "angle": 0, + "content": "[47] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR 2024, pages 8069–8078. IEEE, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18378" + } + ] +] \ No newline at end of file diff --git a/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/201fb178-5c6a-42cf-95fe-3e03de95821f_origin.pdf b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/201fb178-5c6a-42cf-95fe-3e03de95821f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b9c214c61ebf521b87c96ad61bdc981513e99e52 --- /dev/null +++ b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/201fb178-5c6a-42cf-95fe-3e03de95821f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3e0a02de5aa052ba72f7c416352f5d16ca6e1376ab41219cb614effde2674a5 +size 9672663 diff --git a/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/full.md b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..85285f7c37739bf4d6a5ccf1c0d6a023f27f4313 --- /dev/null +++ b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/full.md @@ -0,0 +1,319 @@ +# PatchDPO: Patch-level DPO for Finetuning-free Personalized Image Generation + +Qihan Huang $^{1}$ , Long Chan $^{2}$ , Jinlong Liu $^{2}$ , Wanggui He $^{2}$ , Hao Jiang $^{2}$ , Mingli Song $^{1}$ , Jie Song $^{1,\dagger}$ , Zhejiang University, $^{2}$ Alibaba Group + +{qh.huang,sjie,brooksong}@zju.edu.cn, + +{chenlong0104.chen,aoshu.jh}@alibaba-inc.com, + +LJLwykqh@126.com, wanggui.hwg@taobao.com + +# Abstract + +Finetuning-free personalized image generation can synthesize customized images without test-time finetuning, attracting wide research interest owing to its high efficiency. Current finetuning-free methods simply adopt a single training stage with a simple image reconstruction task, and they typically generate low-quality images inconsistent with the reference images during test-time. To mitigate this problem, inspired by the recent DPO (i.e., direct preference optimization) technique, this work proposes an additional training stage to improve the pre-trained personalized generation models. However, traditional DPO only determines the overall superiority or inferiority of two samples, which is not suitable for personalized image generation because the generated images are commonly inconsistent with the reference images only in some local image patches. To tackle this problem, this work proposes PatchDPO that estimates the quality of image patches within each generated image and accordingly trains the model. To this end, PatchDPO first leverages the pre-trained vision model with a proposed self-supervised training method to estimate the patch quality. Next, PatchDPO adopts a weighted training approach to train the model with the estimated patch quality, which rewards the image patches with high quality while penalizing the image patches with low quality. Experiment results demonstrate that PatchDPO significantly improves the performance of multiple pre-trained personalized generation models, and achieves state-of-the-art performance on both single-object and multi-object personalized image generation. Our code is available at https://github.com/hqhQAQ/PatchDPO. + +# 1. Introduction + +Personalized image generation methods have garnered significant research interest, which generate images based on + +![](images/8c0789e2cf1171686b82db9e823b5599638a91611dce20a1690acf23290feacf.jpg) +Reference Image +Figure 1. The generated images (Images 1 & 2 & 3) are commonly inconsistent with the reference image only in some local image patches (marked in the red boxes). + +![](images/5de881bc67fbc6a9ac0297e2db3c6fd55e80a4a3e95cbb2bf6d9601ed8236edb.jpg) +Image 1 + +![](images/ced75c2d0b999b59bccf18bdc5fac34f81c87d0f0719a37c23a32c5faa90fe8c.jpg) +Image 2 + +![](images/23d82d44c4a5f4f6dc88074b088e5d9874fa3d28f5676eba2205feb7f93a8b87.jpg) +Image 3 + +reference images that define specific details of the desired output. The methodology in this domain is progressively evolving from a finetuning-based approach (e.g., DreamBooth [35], Custom Diffusion [22]) towards a finetuning-free approach (e.g., IP-Adapter [45], Subject-Diffusion [26], JeDi [46]), as finetuning-free methods eliminate the need for finetuning during test-time, significantly reducing usage costs. + +Current finetuning-free methods typically employ only a single training stage on a large-scale image dataset. During this stage, the model is trained with a simple image reconstruction task (i.e., taking each image as reference image to reconstruct itself). When generating images in different scenes from the reference images in test-time, existing models often generate images of lower quality (i.e., inconsistent with the reference images in local details). + +Inspired by the recent DPO technique (i.e., direct preference optimization [32]) that leverages human feedback to optimize the pre-trained model, in this work we strive to devise an additional training stage for improving the pretrained personalized generation models. Specifically, the DPO technique assigns human preference to each sample, and trains the model to generate outputs that align more closely with human preferences. However, traditional DPO only judges the overall superiority or inferiority of two samples, which is not suitable for personalized image generation because generated images are usually inconsistent with the reference images only in some local areas, leading to inaccuracies when comparing the overall quality of two im + +ages. For example, as shown in Figure 1, the generated images (1 & 2 & 3) are inconsistent with the reference image only in the head & back & leg, respectively. In this case, traditional DPO roughly categorizes these images into superior and inferior samples, which would lead to the model's predictions incorrectly approaching the low-quality parts in the superior sample while moving away from the high-quality parts in the inferior sample. + +To tackle this problem, this work proposes PatchDPO, which estimates the quality (preference level) of each patch in the generated image and accordingly optimizes the model. PatchDPO can provide feedback for the model in a more refined way, enabling the model to retain high-quality patches within images while moving away from low-quality patches. To this end, PatchDPO can be divided into three main stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization. + +In the first stage (data construction), PatchDPO requires constructing a training dataset that includes multiple pairs of reference image and generated image. Our preliminary experiments Table 4 show that the complex details of objects in natural images and the confusion between objects and backgrounds hinder model training. Therefore, this work constructs a high-quality dataset for the PatchDPO training. First, this work generates the reference images using the open-source Stable Diffusion model [33] with text prompts instructing the background to be cleaner. Next, the corresponding generated images are synthesized using the pre-trained personalized generation model, with the aforementioned reference images as input. + +In the second stage (patch quality estimation), traditional DPO would require extensive labeling costs to estimate the preference level of samples. In the case of comparing patch details between reference and generated images, thanks to the excellent pre-trained vision models, PatchDPO ingeniously utilizes the pre-trained vision model to extract patch features from reference and generated images, and evaluates the quality of patches in the generated images through patch-to-patch comparisons with those in the reference images. Besides, due to vision models (e.g., classification models pre-trained on ImageNet [8]) typically being better at extracting overall image features instead of patch features, this work proposes a self-supervised training method to improve patch features extraction of the original vision models. We conduct a quantitative evaluation on the HPatches dataset [3] (a dataset with images of the same object from different perspectives and scenes), demonstrating that our method efficiently and accurately extracts patch features for patch-to-patch comparisons. + +In the third stage (model optimization), PatchDPO utilizes the patch quality estimated from the previous stage to further train the generation model. Specifically, PatchDPO adopts a weighted training approach, which assigns higher + +training weights to the image patches with higher quality, while imposing penalties on those of lower quality. Furthermore, this work also incorporates the original reference image as the ground-truth generated image into training. In this manner, for the patches with lower quality in the real generated images, we increase the training weight for their corresponding patches in the ground-truth generated image, thus encouraging the model to correct its predictions for those low-quality patches. + +We perform comprehensive experiments to validate the performance of PatchDPO. Specifically, we apply PatchDPO to multiple pre-trained personalized generation models (e.g., IP-Adapter, ELITE) on both single-object and multi-object personalized image generation. Experiment results on the DreamBooth dataset [35] and the Concept101 dataset [22] demonstrate that PatchDPO significantly improves the performance of pre-trained models. In particular, PatchDPO achieves state-of-the-art performance on both these two tasks. + +To sum up, the main contributions of this work can be summarized as follows: + +- We construct a high-quality dataset to facilitate the PatchDPO training on personalized image generation. +- We propose a patch quality estimation method, which adopts the pre-trained vision models with a proposed self-supervised training approach for assessing the quality of patches in the generated images. +- Based on the estimated patch quality, we propose a weighted training approach for the PatchDPO training on personalized image generation. +- Experiment results show that PatchDPO achieves state-of-the-art performance on both single-object and multi-object personalized image generation. + +# 2. Related Work + +Personalized image generation. Early personalized image generation methods (e.g., DreamBooth [35], Textual Inversion [10], Cones [24], Mix-of-Show [11]) require finetuning the original diffusion model with the reference images. Recently, finetuning-free methods (e.g., IP-Adapter [45], ELITE [40], Subject-Diffusion [26], BLIP-Diffusion [23], InstantBooth [36], FastComposer [41], Taming Encoder [18], SSR-Encoder [47], JeDi [46]) emerge and attract more research interest as they require no finetuning during test-time and significantly reduce the usage cost. However, finetuning-free methods employ only a single training stage with a simple image reconstruction task, leading to low-quality generated images inconsistent with the reference images. PatchDPO compensates for this deficiency using an additional training stage for model optimization from the feedback. + +Aligning diffusion models. The model alignment meth + +![](images/d5abd16fab8a7d461bb5efe09fd5dace00400556c50759018de25efe65ec6639.jpg) + +![](images/f74927aa156ed0225f6ed2ef6ba2683495700aee3a9df95e030baed69b370ce5.jpg) + +![](images/b7579261ef90829e745c0244e5c0566ff98d3b29e7d4b03f18c94eb77b153ace.jpg) +(2.1) Self-Supervised Training +Figure 2. PatchDPO has three stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization. The stage (2) is split into (2.1) self-supervised training and (2.2) patch-to-patch comparison. Besides, in (3), $\hat{\epsilon}_{\theta}(\mathbf{x}_{\mathrm{ref}}(t))$ abbreviates $\epsilon_{\theta}(\mathbf{x}_{\mathrm{ref}}(t),\mathbf{c}_{\mathrm{text}},\mathbf{x}_{\mathrm{ref}},t)$ and $\hat{\epsilon}_{\theta}(\mathbf{x}_{\mathrm{gen}}(t))$ abbreviates $\epsilon_{\theta}(\mathbf{x}_{\mathrm{gen}}(t),\mathbf{c}_{\mathrm{text}},\mathbf{x}_{\mathrm{ref}},t)$ . $\tilde{p} (\mathbf{x}_{\mathrm{ref}})$ and $\tilde{p} (\mathbf{x}_{\mathrm{gen}})$ are upsampled from $p(\mathbf{x}_{\mathrm{ref}})$ and $p(\mathbf{x}_{\mathrm{gen}})$ , respectively. + +![](images/7bddcca2645637dee3e6955d85dc42f96ddd34d45d88731c8ab415023cc7b840.jpg) +(3) Model Optimization + +ods (e.g., RLHF, DPO) first emerged in the field of large language model (LLM). Specifically, RLHF (Reinforcement Learning from Human Feedback) [2, 27] trains a reward function from comparative data of model outputs to reflect human preferences, and adopts reinforcement learning to align it with the original model. DPO (Direct Preference Optimization) [32] simplifies RLHF by aligning the original model directly on the feedback data, but matching RLHF in performance. Recently, some methods (e.g., DPOK [9], DDPO [4], Diffusion-DPO [38], DRaFT [7], AlignProp [31]) have integrated RLHF or DPO into diffusion models for improving image aesthetic. However, these methods simply estimate the overall quality of the entire image, which is not suitable for personalized image generation because the generated images are usually inconsistent with the reference images only in some local image patches. Therefore, in this work, we propose PatchDPO, an advanced model alignment method for personalized image generation by estimating patch quality instead of image quality for model training. + +# 3. Preliminaries + +Diffusion model. Existing personalized image generation models utilize diffusion model [12, 34] as the base model. Diffusion model comprises two processes: a diffusion process which gradually adds noise into the original image with a Markov chain in $T$ steps, and a denoising process which predicts the noise to reconstruct the image with a + +deep neural network. Detailedly, personalized image generation methods synthesize images simultaneously conditioned on the text prompt and the reference images. Commonly, $\epsilon_{\theta}$ denotes the deep neural network for noise prediction, and the training loss of personalized diffusion model is calculated as below $(\| \cdot \|_2$ denotes the L2 norm): + +$$ +\mathcal {L} _ {\mathrm {m s e}} = \mathbb {E} _ {\boldsymbol {x} _ {0}, \boldsymbol {\epsilon} \in \mathcal {N} (\boldsymbol {0}, \boldsymbol {\mathbf {I}}), \boldsymbol {c} _ {\text {t e x t}}, \boldsymbol {c} _ {\text {i m g}}} \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {x} (t), \boldsymbol {c} _ {\text {t e x t}}, \boldsymbol {c} _ {\text {i m g}}, t) \| _ {2} ^ {2}, +$$ + +where $\pmb{x}_0$ denotes the original real image, $t\in [0,T]$ denotes the time step in the diffusion process, $\pmb {x}(t) = \alpha_{t}\pmb{x}_{0} + \sigma_{t}\pmb{\epsilon}$ , and $\alpha_{t},\sigma_{t}$ are predefined weights for step $t$ in the diffusion process. $c_{\mathrm{text}}$ denotes the text condition, and $c_{\mathrm{img}}$ denotes the reference image. After training, the model can generate images by gradually denoising Gaussian noise in multiple steps. + +Reinforcement learning from human feedback (RLHF). RLHF [2, 27] trains the model by maximizing the reward of model output, while regularizing the KL-divergence between it and the original model. Specifically, RLHF trains a reward function $r(\boldsymbol{c}, \boldsymbol{x})$ that estimates the human preference on the generated sample $\boldsymbol{x}$ conditioned on $\boldsymbol{c}$ . Next, let $p_{\theta}$ denote the model being optimized, $p_{\mathrm{ref}}$ denote the original model, the training target of RLHF is calculated as below (note that $\beta$ is the hyper-parameter): + +$$ +\max _ {p _ {\theta}} \mathbb {E} _ {\boldsymbol {c}, \boldsymbol {x}} [ r (\boldsymbol {c}, \boldsymbol {x}) ] - \beta \mathbb {D} _ {\mathrm {K L}} [ p _ {\theta} (\boldsymbol {x} | \boldsymbol {c}) | | p _ {\mathrm {r e f}} (\boldsymbol {x} | \boldsymbol {c}) ]. \tag {1} +$$ + +Direct preference optimization (DPO). Direct preference optimization simplifies RLHF by training the model directly from human preferences. Detailedly, let $\boldsymbol{x}^w$ and $\boldsymbol{x}^l$ denote the "winning" and "losing" samples generated from the condition $c$ , then DPO optimizes the model by aligning its output closer to $\boldsymbol{x}^w$ while distancing it from $\boldsymbol{x}^l$ , and the DPO loss [32] $\mathcal{L}_{\mathrm{DPO}}$ is calculated as below (note that $\sigma(\cdot)$ denotes the sigmoid function): + +$$ +\mathcal {L} _ {\mathrm {D P O}} = - \mathbb {E} _ {\boldsymbol {c}, \boldsymbol {x} ^ {w}, \boldsymbol {x} ^ {l}} \left[ \log \sigma \left(\beta \log \frac {p _ {\theta} (\boldsymbol {x} ^ {w} \mid \boldsymbol {c})}{p _ {\text {r e f}} (\boldsymbol {x} ^ {w} \mid \boldsymbol {c})} - \beta \log \frac {p _ {\theta} (\boldsymbol {x} ^ {l} \mid \boldsymbol {c})}{p _ {\text {r e f}} (\boldsymbol {x} ^ {l} \mid \boldsymbol {c})}\right) \right]. \tag {2} +$$ + +# 4. PatchDPO + +PatchDPO consists of three stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization. + +# 4.1. Data Construction + +PatchDPO requires constructing a training dataset comprising multiple pairs of reference image and generated image (the generated image is synthesized by the personalized generation model being optimized). Our preliminary experiments in Table 4 demonstrate that natural images typically contain images of low-quality, which are not suitable for the task of PatchDPO. Detailedly, these low-quality images comprise complex object details with the confused foreground and background, easily misleading the model training. Therefore, in this work, we choose to construct a high-quality training dataset generated from the open-source generation model using three steps. + +First, this work utilizes ChatGPT to generate the text prompt for each image. The text prompt is in the format of "An {object} in the {background}", where {object} and {background} are imagined by ChatGPT. Second, this work feeds the generated text prompts into the open-source text-to-image generation model (e.g., Stable Diffusion) to generate the reference images. Besides, in addition to the original text prompt, the generation model is also instructed to generate simple backgrounds for the mitigation of confusion between object and background. Finally, this work employs the target personalized generation model to generate images, with the aforementioned text prompts and the corresponding reference images as input. + +# 4.2. Patch Quality Estimation + +Traditional DPO simply estimates the overall quality of the entire generated image, which does not provide sufficiently fine and accurate feedback for personalized image generation, thus resulting in deficient performance. To address this problem, PatchDPO estimates the quality of each patch in the generated image to acquire precise feedback for model optimization. Besides, traditional DPO requires a large amount of annotation cost to estimate the quality of the samples. Instead, the patch quality in personalized image generation is evaluated by comparing the patch details between + +reference images and generated images, which can be conducted using the pre-trained vision models. To this end, this work proposes a patch-to-patch comparison method to estimate the patch quality, and proposes a self-supervised training method for further improvement. + +# 4.2.1. Patch-to-Patch Comparison + +Inspired by ProtoPNet [5, 13, 15, 16, 44] and SFD2 [43] that extract patch features from the deep feature maps, PatchDPO estimates the patch quality with a patch-to-patch comparison on the patch features extracted from the deep feature maps. + +Specifically, let $\mathbf{x}_{\mathrm{ref}}$ denote the reference image, $\mathbf{x}_{\mathrm{gen}}$ denote the corresponding generated image, $f$ denote the pre-trained vision model, then $f(\mathbf{x}_{\mathrm{ref}}) \in \mathbb{R}^{H \times W \times D}$ and $f(\mathbf{x}_{\mathrm{gen}}) \in \mathbb{R}^{H \times W \times D}$ are the feature maps extracted by $f$ (note that $H, W, D$ are the size of the feature map). To ensure the generalizability of PatchDPO, this work acquires the last feature maps extracted from the vision models pretrained on ImageNet as $f(\mathbf{x}_{\mathrm{ref}})$ and $f(\mathbf{x}_{\mathrm{gen}})$ . Because the model does not change the spatial position of feature maps during feature extraction, $f(\mathbf{x})[h, w] \in \mathbb{R}^D$ represents the features of the patch $\mathbf{x}[h, w]$ within the image $\mathbf{x}$ . Note that $\mathbf{x}[h, w]$ denotes the patch in the $h$ -th row and the $w$ -th column of $\mathbf{x}$ , as shown in the right side of Figure 2 (2.2). Next, the quality of each patch $\mathbf{x}_{\mathrm{gen}}[h, w]$ is estimated according to the existence of a patch similar to it in the reference image $\mathbf{x}_{\mathrm{ref}}$ . Detailedly, the patch quality $p(\mathbf{x}_{\mathrm{gen}}[h, w])$ of $\mathbf{x}_{\mathrm{gen}}[h, w]$ is calculated as the maximum similarity between $f(\mathbf{x}_{\mathrm{gen}})[h, w]$ with all elements in $f(\mathbf{x}_{\mathrm{ref}})$ : + +$$ +p \left(\boldsymbol {x} _ {\text {g e n}} [ h, w ]\right) = \max _ {i, j} \frac {f \left(\boldsymbol {x} _ {\text {g e n}}\right) [ h , w ] \cdot f \left(\boldsymbol {x} _ {\text {r e f}}\right) [ i , j ]}{\| f \left(\boldsymbol {x} _ {\text {g e n}}\right) [ h , w ] \| \| f \left(\boldsymbol {x} _ {\text {r e f}}\right) [ i , j ] \|}, \tag {3} +$$ + +where $i$ , $j$ iterate over all the indexes of elements in $f(\pmb{x}_{\mathrm{ref}})$ . Therefore, higher $p(\pmb{x}_{\mathrm{gen}}[h,w])$ indicates that $\pmb{x}_{\mathrm{gen}}[h,w]$ is more consistent with the corresponding patch in the reference image $\pmb{x}_{\mathrm{ref}}$ . + +Verification by $S_{\mathrm{patch}}$ . To guarantee precise patch quality estimation, this work conducts a quantitative evaluation of the extracted patch features using the HPatches dataset [3]. Dually, the HPatches dataset consists of images from 108 groups, where the images of the same group contain the same object from different perspectives and scenes. Besides, for the same group of images, the HPatches dataset annotates the spatial correspondence between their image patches. Based on this dataset, this work adopts a patch matching score $S_{\mathrm{patch}}$ to evaluate the extracted patch features. $S_{\mathrm{patch}}$ is calculated in three steps: (1) Extract the patch features of all images in the dataset, using the pretrained vision model $f$ . (2) For each patch in the image, predict its most similar patch (calculated from the patch features) in other images from the same group. (3) Calculate + +
Model/Layer1471012
ViT-Base70.476.474.868.668.4
ViT-Base (After Training)72.782.783.777.475.8
+ +Table 1. ${S}_{\text{patch }}\left( \% \right)$ estimated on the HPatches dataset. + +the matching accuracy of each patch by comparing the predicted patch with the ground-truth patch, and $S_{\mathrm{patch}}$ is finally calculated by averaging them. + +As shown in Table 1 (a ViT-Base model with 12 layers is adopted here), $S_{\mathrm{patch}}$ of the patch features extracted from the last feature maps achieves only $68.4\%$ , which is not sufficient for patch quality estimation. + +# 4.2.2. Self-Supervised Training + +To facilitate the patch quality estimation, this work strives to improve $S_{\mathrm{patch}}$ from two aspects: (1) Extract patch features from the shallow layers instead of the latest layer. (2) Finetune the vision model $f$ with self-supervised training. + +In the first aspect, the deep neurons in the deep neural networks have large effective receptive fields [1, 14, 19, 20, 25], meaning that each element in the deeper feature map perceives a larger region of the image rather than the image patch in the corresponding location. Therefore, this work explores extracting patch features from feature maps at different depths. As shown in Table 1, patch features extracted from shallow feature maps have higher $S_{\mathrm{patch}}$ in general, and in particular, the patch features extracted from the 4-th layer have the highest $S_{\mathrm{patch}}$ . + +In the second aspect, the performance of the aforementioned patch features extraction is still limited, because the used vision models are typically trained for other vision tasks (e.g., image classification), which focus on extracting the overall image features instead of the patch features. Consequently, this work proposes a self-supervised training method to finetune the pre-trained vision model $f$ , towards improving the performance of patch features extraction without expensive labeling costs. This self-supervised method augments the input image through spatial transformation (i.e., image rotation, image flip) and then constrains the patch features at corresponding positions of the input image and the augmented image to be close. Specifically, let $\operatorname{Aug}(\cdot)$ denote the augmentation operation (e.g., $\operatorname{Aug}(x)$ is the augmented image of the input image $x$ ), then the loss function $\mathcal{L}_{\mathrm{self}}$ of self-supervised training is an MSE loss with a regularization term calculated as below ( $f_{\mathrm{ref}}$ denotes the original model that is frozen during training): + +$$ +\left\{ \begin{array}{l} \mathcal {L} _ {\text {s e l f}} = \mathcal {L} _ {\text {a u g}} + \mathcal {L} _ {\text {r e g}}. \\ \mathcal {L} _ {\text {a u g}} = \| \operatorname {A u g} (f (\boldsymbol {x})) - f (\operatorname {A u g} (\boldsymbol {x})) \| _ {2} ^ {2}. \\ \mathcal {L} _ {\text {r e g}} = \| f (\boldsymbol {x}) - f _ {\text {r e f}} (\boldsymbol {x}) \| _ {2} ^ {2}. \end{array} \right. \tag {4} +$$ + +Here, the regularization term could avoid excessive de + +viation of the finetuned model from the original model, stabilizing model training. Besides, this work chooses the dataset constructed in subsection 4.1 for this finetuning, because the finetuned vision model $f$ will be finally employed for patch quality estimation in this dataset. After the self-supervised training, as shown in Table 1, $S_{\mathrm{patch}}$ of patch features at different layers have shown a significant improvement. We select the patch features with the highest $S_{\mathrm{patch}}$ (83.7%, from the 7th layer) for patch quality estimation, ensuring the performance of PatchDPO training. + +# 4.3. Model Optimization + +With the vision model $f$ finetuned from the previous stage, PatchDPO estimates the patch quality $p(\pmb{x}_{\mathrm{gen}}) \in \mathbb{R}^{H \times W}$ for all generated images in the training dataset. Note that $p(\pmb{x}_{\mathrm{gen}})[h,w] = p(\pmb{x}_{\mathrm{gen}}[h,w]) \in \mathbb{R}$ is the patch quality of image patch $\pmb{x}_{\mathrm{gen}}[h,w]$ . Next, different from traditional DPO simply aligning with the superior samples while distancing from the inferior samples, PatchDPO leverages a weighted training method that adopts a more precise approach for model optimization. Specifically, PatchDPO trains the original personalized generation model with an image reconstruction task (reconstructing the generated image according to the reference image), and then assigns higher training weights to the image patches with higher quality, while assigning lower training weights to the image patches with lower quality. + +However, only reconstructing the generated image can lead the model to still generate low-quality patches in the generated images, instead of generating the corresponding correct patches in the reference images. To address this problem, PatchDPO simultaneously involves a task of reconstructing the reference image using the reference image, as the ground-truth to correct the low-quality patches in the generated image. To this end, PatchDPO estimates the patch quality $p(\pmb{x}_{\mathrm{ref}}) \in \mathbb{R}^{H \times W}$ for the reference image by comparing the features of each patch with those in the generated image, in the same manner as Equation 3. Like $p(\pmb{x}_{\mathrm{gen}})$ , each $p(\pmb{x}_{\mathrm{ref}})[h,w]$ reflects the extent to which $\pmb{x}_{\mathrm{ref}}[h,w]$ exists in the generated image $\pmb{x}_{\mathrm{gen}}$ . Therefore, $\pmb{x}_{\mathrm{ref}}[h,w]$ with lower $p(\pmb{x}_{\mathrm{ref}})[h,w]$ indicates that the patch $\pmb{x}_{\mathrm{ref}}[h,w]$ has low generation quality in the generated image, and the training weight of this patch should be increased in the task of reconstructing the ground-truth image (i.e., the reference image) from the reference image. Finally, the loss $\mathcal{L}_{\mathrm{PatchDPO}}$ of PatchDPO is calculated as below: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {P a t c h D P O}} = \underbrace {\| [ \boldsymbol {\epsilon} _ {\text {g e n}} - \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {x} _ {\text {g e n}} (t) , \boldsymbol {c} _ {\text {t e x t}} , \boldsymbol {x} _ {\text {r e f}} , t) ]} _ {\text {R e c o n s t r u c t} \boldsymbol {x} _ {\text {g e n w i t h}} \boldsymbol {x} _ {\text {r e f}}} \odot \tilde {p} (\boldsymbol {x} _ {\text {g e n}}) \| _ {2} ^ {2} \\ + \| \underbrace {\left[ \boldsymbol {\epsilon} _ {\mathrm {r e f}} - \boldsymbol {\epsilon} _ {\theta} \big (\boldsymbol {x} _ {\mathrm {r e f}} (t) , \boldsymbol {c} _ {\mathrm {t e x t}} , \boldsymbol {x} _ {\mathrm {r e f}} , t \big) \right]} _ {\text {R e c o n s t r u c t} \boldsymbol {x} _ {\mathrm {r e f}} \text {w i t h} \boldsymbol {x} _ {\mathrm {r e f}}} \odot (1 - \tilde {p} (\boldsymbol {x} _ {\mathrm {r e f}})) \| _ {2} ^ {2}. \\ \end{array} +$$ + +Note that $\tilde{p}(\boldsymbol{x}_{\mathrm{gen}})$ and $\tilde{p}(\boldsymbol{x}_{\mathrm{ref}})$ are upsampled from the + +
MethodDINOCLIP-ICLIP-TAvg.
Real Images [35]0.7740.885N/AN/A
Textual Inversion [10]0.5690.7800.2550.535
DreamBooth (Imagen) [35]0.6960.8120.3060.605
DreamBooth (SD) [35]0.6680.8030.3050.592
Custom Diffusion [22]0.6430.7900.3050.579
Re-Imagen [6]0.6000.7400.2700.537
λ-ECLIPSE [29]0.6130.7830.3070.568
ELITE [40]0.6520.7620.2550.556
IP-Adapter [45]0.6080.8090.2740.564
IP-Adapter-Plus [45]0.6920.8260.2810.600
SSR-Encoder [47]0.6120.8210.3080.580
BLIP-Diffusion [23]0.5940.7790.3000.558
MS-Diffusion [39]0.6710.7920.3210.595
Subject-Diffusion [26]0.7110.7870.2930.597
JeDi [46]0.6790.8140.2930.595
PatchDPO0.7270.8380.2920.619
+ +original $p(\pmb{x}_{\mathrm{gen}})$ and $p(\pmb{x}_{\mathrm{ref}})$ with a normalization operation to constrain the values within [0, 1], which have the same height and width as original images ( $\pmb{x}_{\mathrm{gen}} \& \pmb{x}_{\mathrm{ref}}$ ) and noise ( $\epsilon_{\mathrm{gen}} \& \epsilon_{\mathrm{ref}}$ ). Besides, $\odot$ denotes element-wise multiplication that assigns the weights $(\tilde{p}(\pmb{x}_{\mathrm{gen}}) \& 1 - \tilde{p}(\pmb{x}_{\mathrm{ref}}))$ to the corresponding patches in the reconstruction losses. + +# 5. Experiments + +Implementation details. Our main experiments are conducted on the pre-trained IP-Adapter-Plus [45] with SDXL model [30] as the text-to-image diffusion model and OpenCLIP ViT-H/14 as the image encoder. Note that IP-AdapterPlus is the advanced version of the original IP-Adapter with significantly superior performance, by using the Resampler [17] to extract reference image features. This work only estimates the patch quality of object in the image to eliminate the interference from the background. The parameters of the SDXL model and image encoder are frozen, and only the parameters for projecting image features are trainable. During training, we adopt AdamW optimizer with a learning rate of 3e-5, and train the model on 8 GPUs for 30,000 steps with a batch size of 4 per GPU. Besides, the self-supervised training of patch feature extraction is conducted for 10 epochs with a learning rate of 1e-1. + +Training dataset. This work constructs the training dataset as illustrated in subsection 4.1. Detailedly, the datasets for single-object & multi-object personalized generation both consist of 50,000 images. More details of multi-object personalized generation are in S2.1 of the appendix. + +Table 2. Performance comparison for single-object personalized generation on DreamBench. The upper methods are finetuning-based methods, the bottom methods are finetuning-free methods, and bold font denotes the best result. "SD" is Stable Diffusion. + +
MethodDINOCLIP-ICLIP-TAvg.
Kosmos-G [28] (single image)0.6940.8470.2870.609
Emu2-Gen [37] (single image)0.7660.8500.2870.634
OmniGen [42] (single image)0.8010.8470.3010.650
PatchDPO (single image)0.8310.8800.2880.666
+ +Table 3. Performance comparison for single-object personalized generation on DreamBench using the evaluation setting of Kosmos-G. In this setting, only one image is preserved for each object in DreamBench. + +Test benchmark. For single-object personalized image generation, we adopt the famous DreamBench [35] as the benchmark. For multi-object personalized image generation, we follow the Concept101 [22] benchmark that has evaluated many methods. Besides, MultiDreamBench [26] is also adopted for comparison with Subject-Diffusion. + +Evaluation metrics. We follow previous methods to adopt three metrics (CLIP-T, CLIP-I, and DINO) for evaluation. Specifically, CLIP-T evaluates the similarity between the generated images and given text prompts; CLIP-I and DINO evaluate the similarity between the generated images and the reference images. 5 images are generated for each prompt to ensure the evaluation stability. Besides, Avg. (the average of three metrics) is also calculated for a comprehensive comparison. + +Baseline methods. We compare our method with both finetuning-based methods (e.g., Textual Inversion [10], DreamBooth [35], Custom Diffusion [22]) and finetuning-free methods (e.g., SSR-Encoder [47], Subject-Diffusion [26], JeDi [46]). + +# 5.1. Single-Object Personalized Generation + +We conduct both quantitative and qualitative comparisons between our method and baseline methods. + +Quantitative comparisons. Table 2 & Table 3 demonstrates the quantitative results of different methods on DreamBench. Note that Table 2 uses the original setting following most existing methods, where DINO, CLIP-I are calculated by comparing the generated image and all images of the same object. Table 3 uses the evaluation setting following Kosmos-G [28], where only one image is preserved for each object, and DINO, CLIP-I are calculated by comparing the generated image and only this image. The results of baseline methods are from their paper. + +As shown in Table 2 & Table 3, PatchDPO realizes significantly superior image similarity (DINO, CLIP-I) to the SOTA personalized generation methods, because PatchDPO provides very detailed patch-level feedback on the model's generated images, and facilitates the model to correct the low-quality patches that are inconsistent with those from the reference images. Furthermore, PatchDPO + +![](images/be3d417c3005442e537ba319b4c9febdf26be1cd70c6983755237ee38ca99790.jpg) +Figure 3. Qualitative comparisons of different methods on single-object & multi-object personalized image generation. + +achieves text similarity (CLIP-T) comparable to SOTA methods, because each pair of reference images and generated images in the training dataset is from the same text prompt. Therefore, aligning the low-quality patches of generated images with reference images does not decrease the similarity between the generated images and the text prompt. Finally, our method also surpasses existing methods in average performance (Avg.) by a large margin. + +Qualitative comparisons. The upper part of Figure 3 demonstrates the qualitative results of different methods on DreamBench. Compared to existing methods, PatchDPO excels in preserving the local details of the reference image, thus achieving generation of higher quality. + +Qualitative comparisons. The bottom of Figure 3 presents the qualitative results of different methods on Concept101, showing that PatchDPO can better preserve the details of reference images in multi-object personalized generation. + +Furthermore, the experiment results on Concept101 and MultiDreamBench (see S2.2 of the appendix) show that PatchDPO can also improve the performance of original IP-Adapter-Plus on multi-object personalized generation, and achieves superior performance to existing personalized generation methods on both Concept101 & MultiDreamBench. + +# 5.2. Ablation Experiments + +We conduct the main ablation experiments of PatchDPO on DreamBench, as demonstrated in Table 4. + +Training datasets. This work compares our training dataset $\mathcal{D}_{\mathrm{ours}}$ of 50,000 images constructed in subsection 4.1 with a natural dataset $\mathcal{D}_{\mathrm{natural}}$ . $\mathcal{D}_{\mathrm{natural}}$ consists of also 50,000 images (with one main object in the image) randomly selected from the open-source SA-1B dataset [21]. + +In Table 4, Combination (1) seriously degrades the image-alignment (DINO, CLIP-I) of model and Combination (2) benefits the model performance, indicating that the low-quality natural images (with chaotic object details & confused foreground and background, as shown in S3 of the appendix) hinder the training of personalized generation. Note that $\mathcal{L}_{\mathrm{mse}}$ is the loss of original diffusion model. + +Training strategies. This work compares three training strategies corresponding to three losses. $\mathcal{L}_{\mathrm{mse}}$ is the loss of original diffusion model. $\mathcal{L}_{\mathrm{DPO}}$ is the loss of traditional DPO. Detailedly, we leverage the Diffusion-DPO loss [38] that directly adapts the original DPO loss to diffusion model. Finally, $\mathcal{L}_{\mathrm{PatchDPO}}$ is the loss of PatchDPO. + +In Table 4, Combination (3) (traditional DPO) fails to improve the performance of the original model, because traditional DPO would wrongly reward the low-quality patches in the superior sample, while wrongly punishing the high-quality patches in the inferior sample. Instead, Combination (4) (PatchDPO) correctly rewards the high-quality patches and punishes the low-quality patches, thus achieving a huge performance improvement. + +Patch features extraction. This work estimates the ex + +
CombinationDINOCLIP-ICLIP-T
Original model0.6920.8260.281
(1) Lmse + Dnatural0.6580.8180.287
(2) Lmse + Dours0.7080.8300.292
(3) LDPO + Dours0.6760.8190.284
(4) LPatchDPO + Dours + Original f0.7190.8350.291
(5) LPatchDPO + Dours + Trained f0.7270.8380.292
+ +Table 4. Ablation experiments. + +![](images/6ac78cc49cd6cf3cb3006219d453be4ba0dde71d2eb53dbb981118e82181da6b.jpg) +Original Image + +![](images/23188fd71bdf8a3fae6f976bbb2a5a162ffeb10dc49423a1a2fc4933baee76da.jpg) +Figure 4. The matching heatmaps of target patch (cat's head in the original image) on other images of the same cat. + +![](images/10446d6563efc8d6a4e318aee5715c0b717740ac9ad096482f8d0f5def591548.jpg) +Other Images + +![](images/a728a3007201608d012921c26d3ab2fefa3c9b5d51301d1a0b4141180e706cac.jpg) + +![](images/c99c897d09c7912ceb9009cc24bb2fde78680cd9ab352286427c522bd032d45e.jpg) + +tracted patch features with $S_{\mathrm{patch}}$ in subsection 4.2, and here we compare the extracted patch features of low $S_{\mathrm{patch}}$ (68.4%, from the last feature map of original vision model $f$ ) and high $S_{\mathrm{patch}}$ (83.7%, from the shallow feature map of vision model $f$ after self-supervised training). + +In Table 4, Combination (5) (PatchDPO with high $S_{\mathrm{patch}}$ ) surpasses Combination (4) (PatchDPO with low $S_{\mathrm{patch}}$ ), implying that patch features with higher $S_{\mathrm{patch}}$ contribute to more precise patch quality estimation and provide more accurate feedback for the generation model. + +Additional ablation experiments. Besides, we provide more ablation experiments (e.g., PatchDPO on different personalized generation models) in S2.3 of the appendix. + +# 5.3. Visualization Analysis + +Patch matching of extracted patch features. Here, this work visualizes the patch matching results of the extracted patch features. Specifically, for the target patch in the original image, we acquire a matching heatmap $z \in \mathbb{R}^{H \times W}$ by calculating the similarity between its features and all patch features of another image. Note that $z[h, w] \in \mathbb{R}$ indicates the similarity between the target patch and the patch in the $h$ -th row and the $w$ -th column of another image. Next, we visualize $z$ by resizing it to the same size as the image, and overlapping them. As shown in Figure 4, the matching heatmap $z$ accurately highlights the correct patch corresponding to the target patch, implying that the extracted patch features accurately represent the corresponding patch. + +Patch quality estimation. Here, we visualize the estimated patch quality, $p(\pmb{x}_{\mathrm{ref}}) \in \mathbb{R}^{H \times W}$ and $p(\pmb{x}_{\mathrm{gen}}) \in \mathbb{R}^{H \times W}$ , in the same manner as visualizing the matching heatmap $z$ . As demonstrated in Figure 6, the dark patches in the image are inconsistent with the corresponding patches in another image, indicating that our method can accurately estimate the patch quality. + +![](images/f6f7388876e114f8a5c31e716efaee34b4a510f3ab64c955437061d0f5a25da5.jpg) +Figure 5. Qualitative ablation experiment. + +![](images/ccfb4bd5ae8ecef60e5ed57835631165a01216eadcd4ac865a58a4023c6aa66a.jpg) +Figure 6. Three samples of patch quality estimation. + +Images before/after PatchDPO. Here, this work demonstrates the generated images from the model before/after the PatchDPO training. As shown in Figure 5, the images from the model after PatchDPO exhibit significantly enhanced quality in terms of image-alignment, highlighting the effectiveness of PatchDPO. + +Additional visualization results. Furthermore, we provide more visualization results in S3 of the appendix for a comprehensive understanding of our method. + +# 6. Conclusion + +In this work, we propose PatchDPO (patch-level direct preference optimization), which leverages an additional training stage to improve the pre-trained personalized generation models. PatchDPO estimates the quality of image patches within each generated image and accordingly provides detailed feedback to the models. Specifically, we propose a patch quality estimation method based on the pre-trained vision model finetuned with a self-supervised training method. Next, we propose a weighted training approach to train the model with the estimated patch quality, which rewards high-quality image patches while penalizing those of low quality. Experiment results demonstrate that PatchDPO achieves state-of-the-art performance on both single-object and multi-object personalized image generation. We hope our method and dataset can contribute to the community of personalized image generation. + +Acknowledgements. This work is supported by Zhejiang Province High-Level Talents Special Support Program "Leading Talent of Technological Innovation of Ten-Thousands Talents Program" (No. 2022R52046), and Alibaba-Zhejiang University Joint Research Institute of Frontier Technologies. + +# References + +[1] André Araujo, Wade Norris, and Jack Sim. Computing receptive fields of convolutional neural networks. Distill, 2019. https://distill.pub/2019/computing-receptive-fields. 5 +[2] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3 +[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR 2017, pages 3852-3861. IEEE, 2017. 2, 4 +[4] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In *ICLR* 2024. OpenReview.net, 2024. 3 +[5] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan Su. This looks like that: Deep learning for interpretable image recognition. In NeurIPS 2019, pages 8928-8939, 2019. 4 +[6] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W. Cohen. Re-imagen: Retrieval-augmented text-to-image generator. In ICLR 2023. OpenReview.net, 2023. 6 +[7] Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. In ICLR 2024. OpenReview.net, 2024. 3 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR 2009, pages 248-255. IEEE, 2009. 2 +[9] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS 2023, 2023. 3 +[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR 2023. OpenReview.net, 2023. 2, 6 +[11] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, Yixiao Ge, Ying Shan, and Mike Zheng Shou. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. In NeurIPS 2023, 2023. 2 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS 2020, 2020. 3 +[13] Qihan Huang, Mengqi Xue, Wenqi Huang, Haofei Zhang, Jie Song, Yongcheng Jing, and Mingli Song. Evaluation and improvement of interpretability for self-explainable part-prototype networks. In ICCV 2023, pages 2011–2020. IEEE, 2023. 4 +[14] Qihan Huang, Siming Fu, Jinlong Liu, Hao Jiang, Yipeng Yu, and Jie Song. Resolving multi-condition confusion for finetuning-free personalized image generation. arXiv preprint arXiv:2409.17920, 2024. 5 + +[15] Qihan Huang, Jie Song, Jingwen Hu, Haofei Zhang, Yong Wang, and Mingli Song. On the concept trustworthiness in concept bottleneck models. In AAAI 2024, pages 21161-21168, 2024. 4 +[16] Qihan Huang, Jie Song, Mengqi Xue, Haofei Zhang, Bingde Hu, Huiqiong Wang, Hao Jiang, Xingen Wang, and Mingli Song. LG-CAV: train any concept activation vector with language guidance. In NeurIPS 2024, 2024. 4 +[17] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and João Carreira. Perceiver: General perception with iterative attention. In ICML 2021, pages 4651-4664. PMLR, 2021. 6 +[18] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2 +[19] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 5 +[20] Yongcheng Jing, Chongbin Yuan, Li Ju, Yiding Yang, Xinchao Wang, and Dacheng Tao. Deep graph reprogramming. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24345-24354, 2023. 5 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In ICCV 2023, pages 3992-4003. IEEE, 2023. 7 +[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR 2023, pages 1931-1941. IEEE, 2023. 1, 2, 6 +[23] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS 2023, 2023. 2, 6 +[24] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. In ICML 2023, pages 21548-21566. PMLR, 2023. 2 +[25] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard S. Zemel. Understanding the effective receptive field in deep convolutional neural networks. In NeurIPS, pages 4898-4906, 2016. 5 +[26] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In SIGGRAPH 2024, page 25. ACM, 2024. 1, 2, 6 +[27] Jacob Menick, Maja Trebacz, Vladimir Mikulik, John Aslanides, Francis Song, Martin Chadwick, Mia Glaese, Susannah Young, Lucy Campbell-Gillingham, Geoffrey Irving, et al. Teaching language models to support answers with verified quotes. arXiv preprint arXiv:2203.11147, 2022. 3 + +[28] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR 2024. OpenReview.net, 2024. 6 +[29] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. lambda-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 6 +[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6 +[31] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 3 +[32] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS 2023, 2023. 1, 3, 4 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR 2022, pages 10674-10685. IEEE, 2022. 2 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR 2022, pages 10674-10685. IEEE, 2022. 3 +[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR 2023, pages 22500-22510. IEEE, 2023. 1, 2, 6 +[36] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In CVPR 2024, pages 8543-8552. IEEE, 2024. 2 +[37] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR 2024, pages 14398-14409. IEEE, 2024. 6 +[38] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR 2024, pages 8228-8238. IEEE, 2024. 3, 7 +[39] X Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance. arXiv preprint arXiv:2406.07209, 2024. 6 +[40] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV 2023, pages 15897-15907. IEEE, 2023. 2, 6 + +[41] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2 +[42] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 6 +[43] Fei Xue, Ignas Budvytis, and Roberto Cipolla. SFD2: semantic-guided feature detection and description. In CVPR 2023, pages 5206-5216. IEEE, 2023. 4 +[44] Mengqi Xue, Qihan Huang, Haofei Zhang, Jingwen Hu, Jie Song, Mingli Song, and Canghong Jin. Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition. In IJCAI 2024, pages 1516-1524. ijcai.org, 2024. 4 +[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 6 +[46] Yu Zeng, Vishal M. Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In CVPR 2024, pages 6786-6795. IEEE, 2024. 1, 2, 6 +[47] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR 2024, pages 8069–8078. IEEE, 2024. 2, 6 \ No newline at end of file diff --git a/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/images.zip b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..585728d4ecd0ff5f65e69a076135ec31ffa78ad8 --- /dev/null +++ b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e730364ef381c54b02f3906c24d2b6d429b9d5d95445d378171bfee084343e6 +size 658154 diff --git a/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/layout.json b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a3e9ccc5d831dee6f84f4bb6eeb744977be415f8 --- /dev/null +++ b/2025/PatchDPO_ Patch-level DPO for Finetuning-free Personalized Image Generation/layout.json @@ -0,0 +1,9597 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 552, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 552, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 552, + 121 + ], + "type": "text", + "content": "PatchDPO: Patch-level DPO for Finetuning-free Personalized Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "spans": [ + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "text", + "content": "Qihan Huang" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "text", + "content": ", Long Chan" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "text", + "content": ", Jinlong Liu" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "text", + "content": ", Wanggui He" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "text", + "content": ", Hao Jiang" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "text", + "content": ", Mingli Song" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "text", + "content": ", Jie Song" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "text", + "content": ", Zhejiang University, " + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 76, + 142, + 533, + 171 + ], + "type": "text", + "content": " Alibaba Group" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 174, + 403, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 174, + 403, + 186 + ], + "spans": [ + { + "bbox": [ + 205, + 174, + 403, + 186 + ], + "type": "text", + "content": "{qh.huang,sjie,brooksong}@zju.edu.cn," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 184, + 187, + 424, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 187, + 424, + 199 + ], + "spans": [ + { + "bbox": [ + 184, + 187, + 424, + 199 + ], + "type": "text", + "content": "{chenlong0104.chen,aoshu.jh}@alibaba-inc.com," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 197, + 202, + 410, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 202, + 410, + 213 + ], + "spans": [ + { + "bbox": [ + 197, + 202, + 410, + 213 + ], + "type": "text", + "content": "LJLwykqh@126.com, wanggui.hwg@taobao.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 266, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 296, + 625 + ], + "type": "text", + "content": "Finetuning-free personalized image generation can synthesize customized images without test-time finetuning, attracting wide research interest owing to its high efficiency. Current finetuning-free methods simply adopt a single training stage with a simple image reconstruction task, and they typically generate low-quality images inconsistent with the reference images during test-time. To mitigate this problem, inspired by the recent DPO (i.e., direct preference optimization) technique, this work proposes an additional training stage to improve the pre-trained personalized generation models. However, traditional DPO only determines the overall superiority or inferiority of two samples, which is not suitable for personalized image generation because the generated images are commonly inconsistent with the reference images only in some local image patches. To tackle this problem, this work proposes PatchDPO that estimates the quality of image patches within each generated image and accordingly trains the model. To this end, PatchDPO first leverages the pre-trained vision model with a proposed self-supervised training method to estimate the patch quality. Next, PatchDPO adopts a weighted training approach to train the model with the estimated patch quality, which rewards the image patches with high quality while penalizing the image patches with low quality. Experiment results demonstrate that PatchDPO significantly improves the performance of multiple pre-trained personalized generation models, and achieves state-of-the-art performance on both single-object and multi-object personalized image generation. Our code is available at https://github.com/hqhQAQ/PatchDPO." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 649, + 135, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 135, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 135, + 661 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "type": "text", + "content": "Personalized image generation methods have garnered significant research interest, which generate images based on" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 239, + 375, + 295 + ], + "blocks": [ + { + "bbox": [ + 319, + 239, + 375, + 295 + ], + "lines": [ + { + "bbox": [ + 319, + 239, + 375, + 295 + ], + "spans": [ + { + "bbox": [ + 319, + 239, + 375, + 295 + ], + "type": "image", + "image_path": "8c0789e2cf1171686b82db9e823b5599638a91611dce20a1690acf23290feacf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 295, + 378, + 305 + ], + "lines": [ + { + "bbox": [ + 316, + 295, + 378, + 305 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 378, + 305 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 313, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 346 + ], + "type": "text", + "content": "Figure 1. The generated images (Images 1 & 2 & 3) are commonly inconsistent with the reference image only in some local image patches (marked in the red boxes)." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 378, + 239, + 433, + 294 + ], + "blocks": [ + { + "bbox": [ + 378, + 239, + 433, + 294 + ], + "lines": [ + { + "bbox": [ + 378, + 239, + 433, + 294 + ], + "spans": [ + { + "bbox": [ + 378, + 239, + 433, + 294 + ], + "type": "image", + "image_path": "5de881bc67fbc6a9ac0297e2db3c6fd55e80a4a3e95cbb2bf6d9601ed8236edb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 295, + 419, + 304 + ], + "lines": [ + { + "bbox": [ + 391, + 295, + 419, + 304 + ], + "spans": [ + { + "bbox": [ + 391, + 295, + 419, + 304 + ], + "type": "text", + "content": "Image 1" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 436, + 239, + 492, + 294 + ], + "blocks": [ + { + "bbox": [ + 436, + 239, + 492, + 294 + ], + "lines": [ + { + "bbox": [ + 436, + 239, + 492, + 294 + ], + "spans": [ + { + "bbox": [ + 436, + 239, + 492, + 294 + ], + "type": "image", + "image_path": "ced75c2d0b999b59bccf18bdc5fac34f81c87d0f0719a37c23a32c5faa90fe8c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 295, + 479, + 304 + ], + "lines": [ + { + "bbox": [ + 449, + 295, + 479, + 304 + ], + "spans": [ + { + "bbox": [ + 449, + 295, + 479, + 304 + ], + "type": "text", + "content": "Image 2" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 496, + 239, + 551, + 294 + ], + "blocks": [ + { + "bbox": [ + 496, + 239, + 551, + 294 + ], + "lines": [ + { + "bbox": [ + 496, + 239, + 551, + 294 + ], + "spans": [ + { + "bbox": [ + 496, + 239, + 551, + 294 + ], + "type": "image", + "image_path": "23d82d44c4a5f4f6dc88074b088e5d9874fa3d28f5676eba2205feb7f93a8b87.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 508, + 295, + 538, + 304 + ], + "lines": [ + { + "bbox": [ + 508, + 295, + 538, + 304 + ], + "spans": [ + { + "bbox": [ + 508, + 295, + 538, + 304 + ], + "type": "text", + "content": "Image 3" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 362, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 362, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 312, + 362, + 555, + 458 + ], + "type": "text", + "content": "reference images that define specific details of the desired output. The methodology in this domain is progressively evolving from a finetuning-based approach (e.g., DreamBooth [35], Custom Diffusion [22]) towards a finetuning-free approach (e.g., IP-Adapter [45], Subject-Diffusion [26], JeDi [46]), as finetuning-free methods eliminate the need for finetuning during test-time, significantly reducing usage costs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 460, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 556, + 555 + ], + "type": "text", + "content": "Current finetuning-free methods typically employ only a single training stage on a large-scale image dataset. During this stage, the model is trained with a simple image reconstruction task (i.e., taking each image as reference image to reconstruct itself). When generating images in different scenes from the reference images in test-time, existing models often generate images of lower quality (i.e., inconsistent with the reference images in local details)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 557, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 556, + 714 + ], + "type": "text", + "content": "Inspired by the recent DPO technique (i.e., direct preference optimization [32]) that leverages human feedback to optimize the pre-trained model, in this work we strive to devise an additional training stage for improving the pretrained personalized generation models. Specifically, the DPO technique assigns human preference to each sample, and trains the model to generate outputs that align more closely with human preferences. However, traditional DPO only judges the overall superiority or inferiority of two samples, which is not suitable for personalized image generation because generated images are usually inconsistent with the reference images only in some local areas, leading to inaccuracies when comparing the overall quality of two im" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 150, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 150, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 150, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18369" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "ages. For example, as shown in Figure 1, the generated images (1 & 2 & 3) are inconsistent with the reference image only in the head & back & leg, respectively. In this case, traditional DPO roughly categorizes these images into superior and inferior samples, which would lead to the model's predictions incorrectly approaching the low-quality parts in the superior sample while moving away from the high-quality parts in the inferior sample." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 294, + 277 + ], + "type": "text", + "content": "To tackle this problem, this work proposes PatchDPO, which estimates the quality (preference level) of each patch in the generated image and accordingly optimizes the model. PatchDPO can provide feedback for the model in a more refined way, enabling the model to retain high-quality patches within images while moving away from low-quality patches. To this end, PatchDPO can be divided into three main stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 434 + ], + "type": "text", + "content": "In the first stage (data construction), PatchDPO requires constructing a training dataset that includes multiple pairs of reference image and generated image. Our preliminary experiments Table 4 show that the complex details of objects in natural images and the confusion between objects and backgrounds hinder model training. Therefore, this work constructs a high-quality dataset for the PatchDPO training. First, this work generates the reference images using the open-source Stable Diffusion model [33] with text prompts instructing the background to be cleaner. Next, the corresponding generated images are synthesized using the pre-trained personalized generation model, with the aforementioned reference images as input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 436, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 663 + ], + "type": "text", + "content": "In the second stage (patch quality estimation), traditional DPO would require extensive labeling costs to estimate the preference level of samples. In the case of comparing patch details between reference and generated images, thanks to the excellent pre-trained vision models, PatchDPO ingeniously utilizes the pre-trained vision model to extract patch features from reference and generated images, and evaluates the quality of patches in the generated images through patch-to-patch comparisons with those in the reference images. Besides, due to vision models (e.g., classification models pre-trained on ImageNet [8]) typically being better at extracting overall image features instead of patch features, this work proposes a self-supervised training method to improve patch features extraction of the original vision models. We conduct a quantitative evaluation on the HPatches dataset [3] (a dataset with images of the same object from different perspectives and scenes), demonstrating that our method efficiently and accurately extracts patch features for patch-to-patch comparisons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "In the third stage (model optimization), PatchDPO utilizes the patch quality estimated from the previous stage to further train the generation model. Specifically, PatchDPO adopts a weighted training approach, which assigns higher" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "training weights to the image patches with higher quality, while imposing penalties on those of lower quality. Furthermore, this work also incorporates the original reference image as the ground-truth generated image into training. In this manner, for the patches with lower quality in the real generated images, we increase the training weight for their corresponding patches in the ground-truth generated image, thus encouraging the model to correct its predictions for those low-quality patches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 180, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 299 + ], + "type": "text", + "content": "We perform comprehensive experiments to validate the performance of PatchDPO. Specifically, we apply PatchDPO to multiple pre-trained personalized generation models (e.g., IP-Adapter, ELITE) on both single-object and multi-object personalized image generation. Experiment results on the DreamBooth dataset [35] and the Concept101 dataset [22] demonstrate that PatchDPO significantly improves the performance of pre-trained models. In particular, PatchDPO achieves state-of-the-art performance on both these two tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "type": "text", + "content": "To sum up, the main contributions of this work can be summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 325, + 553, + 471 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 325, + 553, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 553, + 349 + ], + "type": "text", + "content": "- We construct a high-quality dataset to facilitate the PatchDPO training on personalized image generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 350, + 553, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 553, + 398 + ], + "type": "text", + "content": "- We propose a patch quality estimation method, which adopts the pre-trained vision models with a proposed self-supervised training approach for assessing the quality of patches in the generated images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 553, + 434 + ], + "type": "text", + "content": "- Based on the estimated patch quality, we propose a weighted training approach for the PatchDPO training on personalized image generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "text", + "content": "- Experiment results show that PatchDPO achieves state-of-the-art performance on both single-object and multi-object personalized image generation." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 483, + 400, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 483, + 400, + 496 + ], + "spans": [ + { + "bbox": [ + 314, + 483, + 400, + 496 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 504, + 553, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 553, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 553, + 694 + ], + "type": "text", + "content": "Personalized image generation. Early personalized image generation methods (e.g., DreamBooth [35], Textual Inversion [10], Cones [24], Mix-of-Show [11]) require finetuning the original diffusion model with the reference images. Recently, finetuning-free methods (e.g., IP-Adapter [45], ELITE [40], Subject-Diffusion [26], BLIP-Diffusion [23], InstantBooth [36], FastComposer [41], Taming Encoder [18], SSR-Encoder [47], JeDi [46]) emerge and attract more research interest as they require no finetuning during test-time and significantly reduce the usage cost. However, finetuning-free methods employ only a single training stage with a simple image reconstruction task, leading to low-quality generated images inconsistent with the reference images. PatchDPO compensates for this deficiency using an additional training stage for model optimization from the feedback." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Aligning diffusion models. The model alignment meth" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18370" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 264, + 180 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 264, + 180 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 264, + 180 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 264, + 180 + ], + "type": "image", + "image_path": "d5abd16fab8a7d461bb5efe09fd5dace00400556c50759018de25efe65ec6639.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 272, + 70, + 551, + 180 + ], + "blocks": [ + { + "bbox": [ + 272, + 70, + 551, + 180 + ], + "lines": [ + { + "bbox": [ + 272, + 70, + 551, + 180 + ], + "spans": [ + { + "bbox": [ + 272, + 70, + 551, + 180 + ], + "type": "image", + "image_path": "f74927aa156ed0225f6ed2ef6ba2683495700aee3a9df95e030baed69b370ce5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 180, + 271, + 304 + ], + "blocks": [ + { + "bbox": [ + 59, + 180, + 271, + 304 + ], + "lines": [ + { + "bbox": [ + 59, + 180, + 271, + 304 + ], + "spans": [ + { + "bbox": [ + 59, + 180, + 271, + 304 + ], + "type": "image", + "image_path": "b7579261ef90829e745c0244e5c0566ff98d3b29e7d4b03f18c94eb77b153ace.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 304, + 231, + 316 + ], + "lines": [ + { + "bbox": [ + 104, + 304, + 231, + 316 + ], + "spans": [ + { + "bbox": [ + 104, + 304, + 231, + 316 + ], + "type": "text", + "content": "(2.1) Self-Supervised Training" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "lines": [ + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "text", + "content": "Figure 2. PatchDPO has three stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization. The stage (2) is split into (2.1) self-supervised training and (2.2) patch-to-patch comparison. Besides, in (3), " + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(\\mathbf{x}_{\\mathrm{ref}}(t))" + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "text", + "content": " abbreviates " + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_{\\mathrm{ref}}(t),\\mathbf{c}_{\\mathrm{text}},\\mathbf{x}_{\\mathrm{ref}},t)" + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(\\mathbf{x}_{\\mathrm{gen}}(t))" + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "text", + "content": " abbreviates " + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_{\\mathrm{gen}}(t),\\mathbf{c}_{\\mathrm{text}},\\mathbf{x}_{\\mathrm{ref}},t)" + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "inline_equation", + "content": "\\tilde{p} (\\mathbf{x}_{\\mathrm{ref}})" + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "inline_equation", + "content": "\\tilde{p} (\\mathbf{x}_{\\mathrm{gen}})" + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "text", + "content": " are upsampled from " + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_{\\mathrm{ref}})" + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_{\\mathrm{gen}})" + }, + { + "bbox": [ + 55, + 329, + 555, + 364 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 272, + 182, + 550, + 304 + ], + "blocks": [ + { + "bbox": [ + 272, + 182, + 550, + 304 + ], + "lines": [ + { + "bbox": [ + 272, + 182, + 550, + 304 + ], + "spans": [ + { + "bbox": [ + 272, + 182, + 550, + 304 + ], + "type": "image", + "image_path": "7bddcca2645637dee3e6955d85dc42f96ddd34d45d88731c8ab415023cc7b840.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 304, + 462, + 316 + ], + "lines": [ + { + "bbox": [ + 362, + 304, + 462, + 316 + ], + "spans": [ + { + "bbox": [ + 362, + 304, + 462, + 316 + ], + "type": "text", + "content": "(3) Model Optimization" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 373, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 296, + 613 + ], + "type": "text", + "content": "ods (e.g., RLHF, DPO) first emerged in the field of large language model (LLM). Specifically, RLHF (Reinforcement Learning from Human Feedback) [2, 27] trains a reward function from comparative data of model outputs to reflect human preferences, and adopts reinforcement learning to align it with the original model. DPO (Direct Preference Optimization) [32] simplifies RLHF by aligning the original model directly on the feedback data, but matching RLHF in performance. Recently, some methods (e.g., DPOK [9], DDPO [4], Diffusion-DPO [38], DRaFT [7], AlignProp [31]) have integrated RLHF or DPO into diffusion models for improving image aesthetic. However, these methods simply estimate the overall quality of the entire image, which is not suitable for personalized image generation because the generated images are usually inconsistent with the reference images only in some local image patches. Therefore, in this work, we propose PatchDPO, an advanced model alignment method for personalized image generation by estimating patch quality instead of image quality for model training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 621, + 140, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 140, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 140, + 634 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Diffusion model. Existing personalized image generation models utilize diffusion model [12, 34] as the base model. Diffusion model comprises two processes: a diffusion process which gradually adds noise into the original image with a Markov chain in " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " steps, and a denoising process which predicts the noise to reconstruct the image with a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 373, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 445 + ], + "type": "text", + "content": "deep neural network. Detailedly, personalized image generation methods synthesize images simultaneously conditioned on the text prompt and the reference images. Commonly, " + }, + { + "bbox": [ + 313, + 373, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 373, + 555, + 445 + ], + "type": "text", + "content": " denotes the deep neural network for noise prediction, and the training loss of personalized diffusion model is calculated as below " + }, + { + "bbox": [ + 313, + 373, + 555, + 445 + ], + "type": "inline_equation", + "content": "(\\| \\cdot \\|_2" + }, + { + "bbox": [ + 313, + 373, + 555, + 445 + ], + "type": "text", + "content": " denotes the L2 norm):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 460, + 556, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 460, + 556, + 476 + ], + "spans": [ + { + "bbox": [ + 314, + 460, + 556, + 476 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {m s e}} = \\mathbb {E} _ {\\boldsymbol {x} _ {0}, \\boldsymbol {\\epsilon} \\in \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {\\mathbf {I}}), \\boldsymbol {c} _ {\\text {t e x t}}, \\boldsymbol {c} _ {\\text {i m g}}} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} (t), \\boldsymbol {c} _ {\\text {t e x t}}, \\boldsymbol {c} _ {\\text {i m g}}, t) \\| _ {2} ^ {2},", + "image_path": "386c28a4c9adfa609d56840feb29c188d3c0daefe957bc2ad6b3f90625bcbff0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "text", + "content": " denotes the original real image, " + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "text", + "content": " denotes the time step in the diffusion process, " + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "inline_equation", + "content": "\\pmb {x}(t) = \\alpha_{t}\\pmb{x}_{0} + \\sigma_{t}\\pmb{\\epsilon}" + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "inline_equation", + "content": "\\alpha_{t},\\sigma_{t}" + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "text", + "content": " are predefined weights for step " + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "text", + "content": " in the diffusion process. " + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "text", + "content": " denotes the text condition, and " + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 484, + 555, + 568 + ], + "type": "text", + "content": " denotes the reference image. After training, the model can generate images by gradually denoising Gaussian noise in multiple steps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "text", + "content": "Reinforcement learning from human feedback (RLHF). RLHF [2, 27] trains the model by maximizing the reward of model output, while regularizing the KL-divergence between it and the original model. Specifically, RLHF trains a reward function " + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "inline_equation", + "content": "r(\\boldsymbol{c}, \\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "text", + "content": " that estimates the human preference on the generated sample " + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c}" + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "text", + "content": ". Next, let " + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "text", + "content": " denote the model being optimized, " + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "text", + "content": " denote the original model, the training target of RLHF is calculated as below (note that " + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 573, + 556, + 681 + ], + "type": "text", + "content": " is the hyper-parameter):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 697, + 555, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 697, + 555, + 716 + ], + "spans": [ + { + "bbox": [ + 326, + 697, + 555, + 716 + ], + "type": "interline_equation", + "content": "\\max _ {p _ {\\theta}} \\mathbb {E} _ {\\boldsymbol {c}, \\boldsymbol {x}} [ r (\\boldsymbol {c}, \\boldsymbol {x}) ] - \\beta \\mathbb {D} _ {\\mathrm {K L}} [ p _ {\\theta} (\\boldsymbol {x} | \\boldsymbol {c}) | | p _ {\\mathrm {r e f}} (\\boldsymbol {x} | \\boldsymbol {c}) ]. \\tag {1}", + "image_path": "aa8d87a333fd1c6c299386d7f79303d5965d1f08d1357c46043d37d8e07ce951.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18371" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "Direct preference optimization (DPO). Direct preference optimization simplifies RLHF by training the model directly from human preferences. Detailedly, let " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^w" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^l" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": " denote the \"winning\" and \"losing\" samples generated from the condition " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": ", then DPO optimizes the model by aligning its output closer to " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^w" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": " while distancing it from " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^l" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": ", and the DPO loss [32] " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DPO}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": " is calculated as below (note that " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": " denotes the sigmoid function):" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 178, + 295, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 178, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 61, + 178, + 295, + 199 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D P O}} = - \\mathbb {E} _ {\\boldsymbol {c}, \\boldsymbol {x} ^ {w}, \\boldsymbol {x} ^ {l}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} ^ {w} \\mid \\boldsymbol {c})}{p _ {\\text {r e f}} (\\boldsymbol {x} ^ {w} \\mid \\boldsymbol {c})} - \\beta \\log \\frac {p _ {\\theta} (\\boldsymbol {x} ^ {l} \\mid \\boldsymbol {c})}{p _ {\\text {r e f}} (\\boldsymbol {x} ^ {l} \\mid \\boldsymbol {c})}\\right) \\right]. \\tag {2}", + "image_path": "e230565cf23a0127861080814ed5bdbccc3ef4b3223192d2c119bb9ba0d7e9c6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 203, + 126, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 203, + 126, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 126, + 216 + ], + "type": "text", + "content": "4. PatchDPO" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 224, + 295, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 248 + ], + "type": "text", + "content": "PatchDPO consists of three stages: (1) Data construction; (2) Patch quality estimation; (3) Model optimization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 254, + 165, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 165, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 165, + 266 + ], + "type": "text", + "content": "4.1. Data Construction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 271, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 295, + 415 + ], + "type": "text", + "content": "PatchDPO requires constructing a training dataset comprising multiple pairs of reference image and generated image (the generated image is synthesized by the personalized generation model being optimized). Our preliminary experiments in Table 4 demonstrate that natural images typically contain images of low-quality, which are not suitable for the task of PatchDPO. Detailedly, these low-quality images comprise complex object details with the confused foreground and background, easily misleading the model training. Therefore, in this work, we choose to construct a high-quality training dataset generated from the open-source generation model using three steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 571 + ], + "type": "text", + "content": "First, this work utilizes ChatGPT to generate the text prompt for each image. The text prompt is in the format of \"An {object} in the {background}\", where {object} and {background} are imagined by ChatGPT. Second, this work feeds the generated text prompts into the open-source text-to-image generation model (e.g., Stable Diffusion) to generate the reference images. Besides, in addition to the original text prompt, the generation model is also instructed to generate simple backgrounds for the mitigation of confusion between object and background. Finally, this work employs the target personalized generation model to generate images, with the aforementioned text prompts and the corresponding reference images as input." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 197, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 197, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 197, + 589 + ], + "type": "text", + "content": "4.2. Patch Quality Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Traditional DPO simply estimates the overall quality of the entire generated image, which does not provide sufficiently fine and accurate feedback for personalized image generation, thus resulting in deficient performance. To address this problem, PatchDPO estimates the quality of each patch in the generated image to acquire precise feedback for model optimization. Besides, traditional DPO requires a large amount of annotation cost to estimate the quality of the samples. Instead, the patch quality in personalized image generation is evaluated by comparing the patch details between" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "reference images and generated images, which can be conducted using the pre-trained vision models. To this end, this work proposes a patch-to-patch comparison method to estimate the patch quality, and proposes a self-supervised training method for further improvement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 140, + 462, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 462, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 462, + 152 + ], + "type": "text", + "content": "4.2.1. Patch-to-Patch Comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 155, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 155, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 553, + 215 + ], + "type": "text", + "content": "Inspired by ProtoPNet [5, 13, 15, 16, 44] and SFD2 [43] that extract patch features from the deep feature maps, PatchDPO estimates the patch quality with a patch-to-patch comparison on the patch features extracted from the deep feature maps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": "Specifically, let " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " denote the reference image, " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{gen}}" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " denote the corresponding generated image, " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " denote the pre-trained vision model, then " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}_{\\mathrm{ref}}) \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}_{\\mathrm{gen}}) \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " are the feature maps extracted by " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " (note that " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "H, W, D" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " are the size of the feature map). To ensure the generalizability of PatchDPO, this work acquires the last feature maps extracted from the vision models pretrained on ImageNet as " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}_{\\mathrm{ref}})" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}_{\\mathrm{gen}})" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": ". Because the model does not change the spatial position of feature maps during feature extraction, " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})[h, w] \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " represents the features of the patch " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{x}[h, w]" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " within the image " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{x}[h, w]" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " denotes the patch in the " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": "-th row and the " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": "-th column of " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": ", as shown in the right side of Figure 2 (2.2). Next, the quality of each patch " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{gen}}[h, w]" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " is estimated according to the existence of a patch similar to it in the reference image " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": ". Detailedly, the patch quality " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_{\\mathrm{gen}}[h, w])" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{gen}}[h, w]" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " is calculated as the maximum similarity between " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}_{\\mathrm{gen}})[h, w]" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": " with all elements in " + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}_{\\mathrm{ref}})" + }, + { + "bbox": [ + 313, + 216, + 555, + 432 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 445, + 553, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 445, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 318, + 445, + 553, + 472 + ], + "type": "interline_equation", + "content": "p \\left(\\boldsymbol {x} _ {\\text {g e n}} [ h, w ]\\right) = \\max _ {i, j} \\frac {f \\left(\\boldsymbol {x} _ {\\text {g e n}}\\right) [ h , w ] \\cdot f \\left(\\boldsymbol {x} _ {\\text {r e f}}\\right) [ i , j ]}{\\| f \\left(\\boldsymbol {x} _ {\\text {g e n}}\\right) [ h , w ] \\| \\| f \\left(\\boldsymbol {x} _ {\\text {r e f}}\\right) [ i , j ] \\|}, \\tag {3}", + "image_path": "b2cd107376e30cba4e8a1815b9204a95f0e01fad1a7c8f716b2a7c1dce641130.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "text", + "content": " iterate over all the indexes of elements in " + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}_{\\mathrm{ref}})" + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "text", + "content": ". Therefore, higher " + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{\\mathrm{gen}}[h,w])" + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "text", + "content": " indicates that " + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{gen}}[h,w]" + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "text", + "content": " is more consistent with the corresponding patch in the reference image " + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 480, + 553, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "Verification by " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". To guarantee precise patch quality estimation, this work conducts a quantitative evaluation of the extracted patch features using the HPatches dataset [3]. Dually, the HPatches dataset consists of images from 108 groups, where the images of the same group contain the same object from different perspectives and scenes. Besides, for the same group of images, the HPatches dataset annotates the spatial correspondence between their image patches. Based on this dataset, this work adopts a patch matching score " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " to evaluate the extracted patch features. " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " is calculated in three steps: (1) Extract the patch features of all images in the dataset, using the pretrained vision model " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". (2) For each patch in the image, predict its most similar patch (calculated from the patch features) in other images from the same group. (3) Calculate" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18372" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 117 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 117 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 117 + ], + "type": "table", + "html": "
Model/Layer1471012
ViT-Base70.476.474.868.668.4
ViT-Base (After Training)72.782.783.777.475.8
", + "image_path": "295f4fd766677828e76cde2bc73a49ad4e4d4357682ad74c33531328ccc3eb78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 124, + 276, + 136 + ], + "lines": [ + { + "bbox": [ + 75, + 124, + 276, + 136 + ], + "spans": [ + { + "bbox": [ + 75, + 124, + 276, + 136 + ], + "type": "text", + "content": "Table 1. " + }, + { + "bbox": [ + 75, + 124, + 276, + 136 + ], + "type": "inline_equation", + "content": "{S}_{\\text{patch }}\\left( \\% \\right)" + }, + { + "bbox": [ + 75, + 124, + 276, + 136 + ], + "type": "text", + "content": " estimated on the HPatches dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 148, + 295, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 295, + 183 + ], + "type": "text", + "content": "the matching accuracy of each patch by comparing the predicted patch with the ground-truth patch, and " + }, + { + "bbox": [ + 55, + 148, + 295, + 183 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 55, + 148, + 295, + 183 + ], + "type": "text", + "content": " is finally calculated by averaging them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 183, + 296, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 183, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 296, + 232 + ], + "type": "text", + "content": "As shown in Table 1 (a ViT-Base model with 12 layers is adopted here), " + }, + { + "bbox": [ + 55, + 183, + 296, + 232 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 55, + 183, + 296, + 232 + ], + "type": "text", + "content": " of the patch features extracted from the last feature maps achieves only " + }, + { + "bbox": [ + 55, + 183, + 296, + 232 + ], + "type": "inline_equation", + "content": "68.4\\%" + }, + { + "bbox": [ + 55, + 183, + 296, + 232 + ], + "type": "text", + "content": ", which is not sufficient for patch quality estimation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 238, + 191, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 191, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 191, + 251 + ], + "type": "text", + "content": "4.2.2. Self-Supervised Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "type": "text", + "content": "To facilitate the patch quality estimation, this work strives to improve " + }, + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "type": "text", + "content": " from two aspects: (1) Extract patch features from the shallow layers instead of the latest layer. (2) Finetune the vision model " + }, + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 252, + 296, + 300 + ], + "type": "text", + "content": " with self-supervised training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 300, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 296, + 419 + ], + "type": "text", + "content": "In the first aspect, the deep neurons in the deep neural networks have large effective receptive fields [1, 14, 19, 20, 25], meaning that each element in the deeper feature map perceives a larger region of the image rather than the image patch in the corresponding location. Therefore, this work explores extracting patch features from feature maps at different depths. As shown in Table 1, patch features extracted from shallow feature maps have higher " + }, + { + "bbox": [ + 54, + 300, + 296, + 419 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 54, + 300, + 296, + 419 + ], + "type": "text", + "content": " in general, and in particular, the patch features extracted from the 4-th layer have the highest " + }, + { + "bbox": [ + 54, + 300, + 296, + 419 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 54, + 300, + 296, + 419 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "text", + "content": "In the second aspect, the performance of the aforementioned patch features extraction is still limited, because the used vision models are typically trained for other vision tasks (e.g., image classification), which focus on extracting the overall image features instead of the patch features. Consequently, this work proposes a self-supervised training method to finetune the pre-trained vision model " + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "text", + "content": ", towards improving the performance of patch features extraction without expensive labeling costs. This self-supervised method augments the input image through spatial transformation (i.e., image rotation, image flip) and then constrains the patch features at corresponding positions of the input image and the augmented image to be close. Specifically, let " + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\operatorname{Aug}(\\cdot)" + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "text", + "content": " denote the augmentation operation (e.g., " + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\operatorname{Aug}(x)" + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "text", + "content": " is the augmented image of the input image " + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "text", + "content": "), then the loss function " + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{self}}" + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "text", + "content": " of self-supervised training is an MSE loss with a regularization term calculated as below (" + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 420, + 296, + 635 + ], + "type": "text", + "content": " denotes the original model that is frozen during training):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 87, + 649, + 295, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 649, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 87, + 649, + 295, + 693 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathcal {L} _ {\\text {s e l f}} = \\mathcal {L} _ {\\text {a u g}} + \\mathcal {L} _ {\\text {r e g}}. \\\\ \\mathcal {L} _ {\\text {a u g}} = \\| \\operatorname {A u g} (f (\\boldsymbol {x})) - f (\\operatorname {A u g} (\\boldsymbol {x})) \\| _ {2} ^ {2}. \\\\ \\mathcal {L} _ {\\text {r e g}} = \\| f (\\boldsymbol {x}) - f _ {\\text {r e f}} (\\boldsymbol {x}) \\| _ {2} ^ {2}. \\end{array} \\right. \\tag {4}", + "image_path": "6d52c5cc64005d2ac05f73109376cfd120e70433442af794639d1c9b70433f23.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Here, the regularization term could avoid excessive de" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "viation of the finetuned model from the original model, stabilizing model training. Besides, this work chooses the dataset constructed in subsection 4.1 for this finetuning, because the finetuned vision model " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " will be finally employed for patch quality estimation in this dataset. After the self-supervised training, as shown in Table 1, " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " of patch features at different layers have shown a significant improvement. We select the patch features with the highest " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " (83.7%, from the 7th layer) for patch quality estimation, ensuring the performance of PatchDPO training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 198, + 432, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 432, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 432, + 211 + ], + "type": "text", + "content": "4.3. Model Optimization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 215, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 395 + ], + "type": "text", + "content": "With the vision model " + }, + { + "bbox": [ + 313, + 215, + 555, + 395 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 215, + 555, + 395 + ], + "type": "text", + "content": " finetuned from the previous stage, PatchDPO estimates the patch quality " + }, + { + "bbox": [ + 313, + 215, + 555, + 395 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{\\mathrm{gen}}) \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 215, + 555, + 395 + ], + "type": "text", + "content": " for all generated images in the training dataset. Note that " + }, + { + "bbox": [ + 313, + 215, + 555, + 395 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{\\mathrm{gen}})[h,w] = p(\\pmb{x}_{\\mathrm{gen}}[h,w]) \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 215, + 555, + 395 + ], + "type": "text", + "content": " is the patch quality of image patch " + }, + { + "bbox": [ + 313, + 215, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{gen}}[h,w]" + }, + { + "bbox": [ + 313, + 215, + 555, + 395 + ], + "type": "text", + "content": ". Next, different from traditional DPO simply aligning with the superior samples while distancing from the inferior samples, PatchDPO leverages a weighted training method that adopts a more precise approach for model optimization. Specifically, PatchDPO trains the original personalized generation model with an image reconstruction task (reconstructing the generated image according to the reference image), and then assigns higher training weights to the image patches with higher quality, while assigning lower training weights to the image patches with lower quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "text", + "content": "However, only reconstructing the generated image can lead the model to still generate low-quality patches in the generated images, instead of generating the corresponding correct patches in the reference images. To address this problem, PatchDPO simultaneously involves a task of reconstructing the reference image using the reference image, as the ground-truth to correct the low-quality patches in the generated image. To this end, PatchDPO estimates the patch quality " + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{\\mathrm{ref}}) \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "text", + "content": " for the reference image by comparing the features of each patch with those in the generated image, in the same manner as Equation 3. Like " + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{\\mathrm{gen}})" + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "text", + "content": ", each " + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{\\mathrm{ref}})[h,w]" + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "text", + "content": " reflects the extent to which " + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{ref}}[h,w]" + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "text", + "content": " exists in the generated image " + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{gen}}" + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "text", + "content": ". Therefore, " + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{ref}}[h,w]" + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "text", + "content": " with lower " + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{\\mathrm{ref}})[h,w]" + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "text", + "content": " indicates that the patch " + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{ref}}[h,w]" + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "text", + "content": " has low generation quality in the generated image, and the training weight of this patch should be increased in the task of reconstructing the ground-truth image (i.e., the reference image) from the reference image. Finally, the loss " + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{PatchDPO}}" + }, + { + "bbox": [ + 313, + 396, + 556, + 622 + ], + "type": "text", + "content": " of PatchDPO is calculated as below:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 634, + 560, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 634, + 560, + 696 + ], + "spans": [ + { + "bbox": [ + 315, + 634, + 560, + 696 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {P a t c h D P O}} = \\underbrace {\\| [ \\boldsymbol {\\epsilon} _ {\\text {g e n}} - \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {\\text {g e n}} (t) , \\boldsymbol {c} _ {\\text {t e x t}} , \\boldsymbol {x} _ {\\text {r e f}} , t) ]} _ {\\text {R e c o n s t r u c t} \\boldsymbol {x} _ {\\text {g e n w i t h}} \\boldsymbol {x} _ {\\text {r e f}}} \\odot \\tilde {p} (\\boldsymbol {x} _ {\\text {g e n}}) \\| _ {2} ^ {2} \\\\ + \\| \\underbrace {\\left[ \\boldsymbol {\\epsilon} _ {\\mathrm {r e f}} - \\boldsymbol {\\epsilon} _ {\\theta} \\big (\\boldsymbol {x} _ {\\mathrm {r e f}} (t) , \\boldsymbol {c} _ {\\mathrm {t e x t}} , \\boldsymbol {x} _ {\\mathrm {r e f}} , t \\big) \\right]} _ {\\text {R e c o n s t r u c t} \\boldsymbol {x} _ {\\mathrm {r e f}} \\text {w i t h} \\boldsymbol {x} _ {\\mathrm {r e f}}} \\odot (1 - \\tilde {p} (\\boldsymbol {x} _ {\\mathrm {r e f}})) \\| _ {2} ^ {2}. \\\\ \\end{array}", + "image_path": "317f2fdfc8b6cd932c25800971b2e8e2c87f280a52b60596d76392fb20d15c51.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{p}(\\boldsymbol{x}_{\\mathrm{gen}})" + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{p}(\\boldsymbol{x}_{\\mathrm{ref}})" + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": " are upsampled from the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18373" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 294, + 280 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 280 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 280 + ], + "type": "table", + "html": "
MethodDINOCLIP-ICLIP-TAvg.
Real Images [35]0.7740.885N/AN/A
Textual Inversion [10]0.5690.7800.2550.535
DreamBooth (Imagen) [35]0.6960.8120.3060.605
DreamBooth (SD) [35]0.6680.8030.3050.592
Custom Diffusion [22]0.6430.7900.3050.579
Re-Imagen [6]0.6000.7400.2700.537
λ-ECLIPSE [29]0.6130.7830.3070.568
ELITE [40]0.6520.7620.2550.556
IP-Adapter [45]0.6080.8090.2740.564
IP-Adapter-Plus [45]0.6920.8260.2810.600
SSR-Encoder [47]0.6120.8210.3080.580
BLIP-Diffusion [23]0.5940.7790.3000.558
MS-Diffusion [39]0.6710.7920.3210.595
Subject-Diffusion [26]0.7110.7870.2930.597
JeDi [46]0.6790.8140.2930.595
PatchDPO0.7270.8380.2920.619
", + "image_path": "a2c00432b322dae192e08a7ccaea500e64f368af9739a4bf81c8008279b57362.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "text", + "content": "original " + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{\\mathrm{gen}})" + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{\\mathrm{ref}})" + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "text", + "content": " with a normalization operation to constrain the values within [0, 1], which have the same height and width as original images (" + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{gen}} \\& \\pmb{x}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "text", + "content": ") and noise (" + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{gen}} \\& \\epsilon_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "text", + "content": "). Besides, " + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "text", + "content": " denotes element-wise multiplication that assigns the weights " + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "inline_equation", + "content": "(\\tilde{p}(\\pmb{x}_{\\mathrm{gen}}) \\& 1 - \\tilde{p}(\\pmb{x}_{\\mathrm{ref}}))" + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "text", + "content": " to the corresponding patches in the reconstruction losses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 430, + 136, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 136, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 136, + 443 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 295, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 648 + ], + "type": "text", + "content": "Implementation details. Our main experiments are conducted on the pre-trained IP-Adapter-Plus [45] with SDXL model [30] as the text-to-image diffusion model and OpenCLIP ViT-H/14 as the image encoder. Note that IP-AdapterPlus is the advanced version of the original IP-Adapter with significantly superior performance, by using the Resampler [17] to extract reference image features. This work only estimates the patch quality of object in the image to eliminate the interference from the background. The parameters of the SDXL model and image encoder are frozen, and only the parameters for projecting image features are trainable. During training, we adopt AdamW optimizer with a learning rate of 3e-5, and train the model on 8 GPUs for 30,000 steps with a batch size of 4 per GPU. Besides, the self-supervised training of patch feature extraction is conducted for 10 epochs with a learning rate of 1e-1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Training dataset. This work constructs the training dataset as illustrated in subsection 4.1. Detailedly, the datasets for single-object & multi-object personalized generation both consist of 50,000 images. More details of multi-object personalized generation are in S2.1 of the appendix." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 553, + 143 + ], + "blocks": [ + { + "bbox": [ + 55, + 288, + 295, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 332 + ], + "type": "text", + "content": "Table 2. Performance comparison for single-object personalized generation on DreamBench. The upper methods are finetuning-based methods, the bottom methods are finetuning-free methods, and bold font denotes the best result. \"SD\" is Stable Diffusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 553, + 143 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 143 + ], + "type": "table", + "html": "
MethodDINOCLIP-ICLIP-TAvg.
Kosmos-G [28] (single image)0.6940.8470.2870.609
Emu2-Gen [37] (single image)0.7660.8500.2870.634
OmniGen [42] (single image)0.8010.8470.3010.650
PatchDPO (single image)0.8310.8800.2880.666
", + "image_path": "f550578c10de5a95c292775ab5f789b2af90f050b19569c9db6afc33fd6914c2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 152, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 196 + ], + "type": "text", + "content": "Table 3. Performance comparison for single-object personalized generation on DreamBench using the evaluation setting of Kosmos-G. In this setting, only one image is preserved for each object in DreamBench." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 207, + 555, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 555, + 280 + ], + "type": "text", + "content": "Test benchmark. For single-object personalized image generation, we adopt the famous DreamBench [35] as the benchmark. For multi-object personalized image generation, we follow the Concept101 [22] benchmark that has evaluated many methods. Besides, MultiDreamBench [26] is also adopted for comparison with Subject-Diffusion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 284, + 556, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 556, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 556, + 392 + ], + "type": "text", + "content": "Evaluation metrics. We follow previous methods to adopt three metrics (CLIP-T, CLIP-I, and DINO) for evaluation. Specifically, CLIP-T evaluates the similarity between the generated images and given text prompts; CLIP-I and DINO evaluate the similarity between the generated images and the reference images. 5 images are generated for each prompt to ensure the evaluation stability. Besides, Avg. (the average of three metrics) is also calculated for a comprehensive comparison." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": "Baseline methods. We compare our method with both finetuning-based methods (e.g., Textual Inversion [10], DreamBooth [35], Custom Diffusion [22]) and finetuning-free methods (e.g., SSR-Encoder [47], Subject-Diffusion [26], JeDi [46])." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 464, + 518, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 518, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 518, + 476 + ], + "type": "text", + "content": "5.1. Single-Object Personalized Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 481, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 554, + 506 + ], + "type": "text", + "content": "We conduct both quantitative and qualitative comparisons between our method and baseline methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 629 + ], + "type": "text", + "content": "Quantitative comparisons. Table 2 & Table 3 demonstrates the quantitative results of different methods on DreamBench. Note that Table 2 uses the original setting following most existing methods, where DINO, CLIP-I are calculated by comparing the generated image and all images of the same object. Table 3 uses the evaluation setting following Kosmos-G [28], where only one image is preserved for each object, and DINO, CLIP-I are calculated by comparing the generated image and only this image. The results of baseline methods are from their paper." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "As shown in Table 2 & Table 3, PatchDPO realizes significantly superior image similarity (DINO, CLIP-I) to the SOTA personalized generation methods, because PatchDPO provides very detailed patch-level feedback on the model's generated images, and facilitates the model to correct the low-quality patches that are inconsistent with those from the reference images. Furthermore, PatchDPO" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18374" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 555, + 340 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 340 + ], + "type": "image", + "image_path": "be3d417c3005442e537ba319b4c9febdf26be1cd70c6983755237ee38ca99790.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 347, + 518, + 358 + ], + "lines": [ + { + "bbox": [ + 88, + 347, + 518, + 358 + ], + "spans": [ + { + "bbox": [ + 88, + 347, + 518, + 358 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons of different methods on single-object & multi-object personalized image generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 369, + 295, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 295, + 465 + ], + "type": "text", + "content": "achieves text similarity (CLIP-T) comparable to SOTA methods, because each pair of reference images and generated images in the training dataset is from the same text prompt. Therefore, aligning the low-quality patches of generated images with reference images does not decrease the similarity between the generated images and the text prompt. Finally, our method also surpasses existing methods in average performance (Avg.) by a large margin." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 533 + ], + "type": "text", + "content": "Qualitative comparisons. The upper part of Figure 3 demonstrates the qualitative results of different methods on DreamBench. Compared to existing methods, PatchDPO excels in preserving the local details of the reference image, thus achieving generation of higher quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 586 + ], + "type": "text", + "content": "Qualitative comparisons. The bottom of Figure 3 presents the qualitative results of different methods on Concept101, showing that PatchDPO can better preserve the details of reference images in multi-object personalized generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "text", + "content": "Furthermore, the experiment results on Concept101 and MultiDreamBench (see S2.2 of the appendix) show that PatchDPO can also improve the performance of original IP-Adapter-Plus on multi-object personalized generation, and achieves superior performance to existing personalized generation methods on both Concept101 & MultiDreamBench." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 670, + 180, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 180, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 180, + 683 + ], + "type": "text", + "content": "5.2. Ablation Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "We conduct the main ablation experiments of PatchDPO on DreamBench, as demonstrated in Table 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "content": "Training datasets. This work compares our training dataset " + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{ours}}" + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "content": " of 50,000 images constructed in subsection 4.1 with a natural dataset " + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{natural}}" + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{natural}}" + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "content": " consists of also 50,000 images (with one main object in the image) randomly selected from the open-source SA-1B dataset [21]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 432, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 516 + ], + "type": "text", + "content": "In Table 4, Combination (1) seriously degrades the image-alignment (DINO, CLIP-I) of model and Combination (2) benefits the model performance, indicating that the low-quality natural images (with chaotic object details & confused foreground and background, as shown in S3 of the appendix) hinder the training of personalized generation. Note that " + }, + { + "bbox": [ + 313, + 432, + 555, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mse}}" + }, + { + "bbox": [ + 313, + 432, + 555, + 516 + ], + "type": "text", + "content": " is the loss of original diffusion model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 523, + 554, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 595 + ], + "type": "text", + "content": "Training strategies. This work compares three training strategies corresponding to three losses. " + }, + { + "bbox": [ + 313, + 523, + 554, + 595 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mse}}" + }, + { + "bbox": [ + 313, + 523, + 554, + 595 + ], + "type": "text", + "content": " is the loss of original diffusion model. " + }, + { + "bbox": [ + 313, + 523, + 554, + 595 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DPO}}" + }, + { + "bbox": [ + 313, + 523, + 554, + 595 + ], + "type": "text", + "content": " is the loss of traditional DPO. Detailedly, we leverage the Diffusion-DPO loss [38] that directly adapts the original DPO loss to diffusion model. Finally, " + }, + { + "bbox": [ + 313, + 523, + 554, + 595 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{PatchDPO}}" + }, + { + "bbox": [ + 313, + 523, + 554, + 595 + ], + "type": "text", + "content": " is the loss of PatchDPO." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 597, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 554, + 693 + ], + "type": "text", + "content": "In Table 4, Combination (3) (traditional DPO) fails to improve the performance of the original model, because traditional DPO would wrongly reward the low-quality patches in the superior sample, while wrongly punishing the high-quality patches in the inferior sample. Instead, Combination (4) (PatchDPO) correctly rewards the high-quality patches and punishes the low-quality patches, thus achieving a huge performance improvement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 701, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 712 + ], + "type": "text", + "content": "Patch features extraction. This work estimates the ex" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18375" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 295, + 160 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 160 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 160 + ], + "type": "table", + "html": "
CombinationDINOCLIP-ICLIP-T
Original model0.6920.8260.281
(1) Lmse + Dnatural0.6580.8180.287
(2) Lmse + Dours0.7080.8300.292
(3) LDPO + Dours0.6760.8190.284
(4) LPatchDPO + Dours + Original f0.7190.8350.291
(5) LPatchDPO + Dours + Trained f0.7270.8380.292
", + "image_path": "35ad410886383e49259b4e3ea90283717741ee54d00efd5b10d3dcbfb193e2e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 168, + 232, + 179 + ], + "lines": [ + { + "bbox": [ + 118, + 168, + 232, + 179 + ], + "spans": [ + { + "bbox": [ + 118, + 168, + 232, + 179 + ], + "type": "text", + "content": "Table 4. Ablation experiments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 61, + 181, + 108, + 226 + ], + "blocks": [ + { + "bbox": [ + 61, + 181, + 108, + 226 + ], + "lines": [ + { + "bbox": [ + 61, + 181, + 108, + 226 + ], + "spans": [ + { + "bbox": [ + 61, + 181, + 108, + 226 + ], + "type": "image", + "image_path": "6ac78cc49cd6cf3cb3006219d453be4ba0dde71d2eb53dbb981118e82181da6b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 227, + 112, + 237 + ], + "lines": [ + { + "bbox": [ + 57, + 227, + 112, + 237 + ], + "spans": [ + { + "bbox": [ + 57, + 227, + 112, + 237 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 112, + 181, + 156, + 226 + ], + "blocks": [ + { + "bbox": [ + 112, + 181, + 156, + 226 + ], + "lines": [ + { + "bbox": [ + 112, + 181, + 156, + 226 + ], + "spans": [ + { + "bbox": [ + 112, + 181, + 156, + 226 + ], + "type": "image", + "image_path": "23188fd71bdf8a3fae6f976bbb2a5a162ffeb10dc49423a1a2fc4933baee76da.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 295, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 269 + ], + "type": "text", + "content": "Figure 4. The matching heatmaps of target patch (cat's head in the original image) on other images of the same cat." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 157, + 181, + 203, + 226 + ], + "blocks": [ + { + "bbox": [ + 157, + 181, + 203, + 226 + ], + "lines": [ + { + "bbox": [ + 157, + 181, + 203, + 226 + ], + "spans": [ + { + "bbox": [ + 157, + 181, + 203, + 226 + ], + "type": "image", + "image_path": "10446d6563efc8d6a4e318aee5715c0b717740ac9ad096482f8d0f5def591548.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 227, + 227, + 237 + ], + "lines": [ + { + "bbox": [ + 179, + 227, + 227, + 237 + ], + "spans": [ + { + "bbox": [ + 179, + 227, + 227, + 237 + ], + "type": "text", + "content": "Other Images" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 203, + 181, + 249, + 226 + ], + "blocks": [ + { + "bbox": [ + 203, + 181, + 249, + 226 + ], + "lines": [ + { + "bbox": [ + 203, + 181, + 249, + 226 + ], + "spans": [ + { + "bbox": [ + 203, + 181, + 249, + 226 + ], + "type": "image", + "image_path": "a728a3007201608d012921c26d3ab2fefa3c9b5d51301d1a0b4141180e706cac.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 249, + 181, + 294, + 227 + ], + "blocks": [ + { + "bbox": [ + 249, + 181, + 294, + 227 + ], + "lines": [ + { + "bbox": [ + 249, + 181, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 249, + 181, + 294, + 227 + ], + "type": "image", + "image_path": "c99c897d09c7912ceb9009cc24bb2fde78680cd9ab352286427c522bd032d45e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "text", + "content": "tracted patch features with " + }, + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "text", + "content": " in subsection 4.2, and here we compare the extracted patch features of low " + }, + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "text", + "content": " (68.4%, from the last feature map of original vision model " + }, + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "text", + "content": ") and high " + }, + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "text", + "content": " (83.7%, from the shallow feature map of vision model " + }, + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 279, + 295, + 339 + ], + "type": "text", + "content": " after self-supervised training)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 339, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 295, + 399 + ], + "type": "text", + "content": "In Table 4, Combination (5) (PatchDPO with high " + }, + { + "bbox": [ + 55, + 339, + 295, + 399 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 55, + 339, + 295, + 399 + ], + "type": "text", + "content": ") surpasses Combination (4) (PatchDPO with low " + }, + { + "bbox": [ + 55, + 339, + 295, + 399 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 55, + 339, + 295, + 399 + ], + "type": "text", + "content": "), implying that patch features with higher " + }, + { + "bbox": [ + 55, + 339, + 295, + 399 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{patch}}" + }, + { + "bbox": [ + 55, + 339, + 295, + 399 + ], + "type": "text", + "content": " contribute to more precise patch quality estimation and provide more accurate feedback for the generation model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 404, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 440 + ], + "type": "text", + "content": "Additional ablation experiments. Besides, we provide more ablation experiments (e.g., PatchDPO on different personalized generation models) in S2.3 of the appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 446, + 181, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 181, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 181, + 459 + ], + "type": "text", + "content": "5.3. Visualization Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "text", + "content": "Patch matching of extracted patch features. Here, this work visualizes the patch matching results of the extracted patch features. Specifically, for the target patch in the original image, we acquire a matching heatmap " + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "text", + "content": " by calculating the similarity between its features and all patch features of another image. Note that " + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "inline_equation", + "content": "z[h, w] \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "text", + "content": " indicates the similarity between the target patch and the patch in the " + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "text", + "content": "-th row and the " + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "text", + "content": "-th column of another image. Next, we visualize " + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "text", + "content": " by resizing it to the same size as the image, and overlapping them. As shown in Figure 4, the matching heatmap " + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 469, + 295, + 625 + ], + "type": "text", + "content": " accurately highlights the correct patch corresponding to the target patch, implying that the extracted patch features accurately represent the corresponding patch." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "Patch quality estimation. Here, we visualize the estimated patch quality, " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{\\mathrm{ref}}) \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{\\mathrm{gen}}) \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": ", in the same manner as visualizing the matching heatmap " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": ". As demonstrated in Figure 6, the dark patches in the image are inconsistent with the corresponding patches in another image, indicating that our method can accurately estimate the patch quality." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 158 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 158 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 158 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 158 + ], + "type": "image", + "image_path": "f6f7388876e114f8a5c31e716efaee34b4a510f3ab64c955437061d0f5a25da5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 165, + 511, + 176 + ], + "lines": [ + { + "bbox": [ + 357, + 165, + 511, + 176 + ], + "spans": [ + { + "bbox": [ + 357, + 165, + 511, + 176 + ], + "type": "text", + "content": "Figure 5. Qualitative ablation experiment." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 179, + 553, + 268 + ], + "blocks": [ + { + "bbox": [ + 316, + 179, + 553, + 268 + ], + "lines": [ + { + "bbox": [ + 316, + 179, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 316, + 179, + 553, + 268 + ], + "type": "image", + "image_path": "ccfb4bd5ae8ecef60e5ed57835631165a01216eadcd4ac865a58a4023c6aa66a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 275, + 529, + 287 + ], + "lines": [ + { + "bbox": [ + 338, + 275, + 529, + 287 + ], + "spans": [ + { + "bbox": [ + 338, + 275, + 529, + 287 + ], + "type": "text", + "content": "Figure 6. Three samples of patch quality estimation." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 299, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 370 + ], + "type": "text", + "content": "Images before/after PatchDPO. Here, this work demonstrates the generated images from the model before/after the PatchDPO training. As shown in Figure 5, the images from the model after PatchDPO exhibit significantly enhanced quality in terms of image-alignment, highlighting the effectiveness of PatchDPO." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 412 + ], + "type": "text", + "content": "Additional visualization results. Furthermore, we provide more visualization results in S3 of the appendix for a comprehensive understanding of our method." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 425, + 388, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 388, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 388, + 437 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 445, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 636 + ], + "type": "text", + "content": "In this work, we propose PatchDPO (patch-level direct preference optimization), which leverages an additional training stage to improve the pre-trained personalized generation models. PatchDPO estimates the quality of image patches within each generated image and accordingly provides detailed feedback to the models. Specifically, we propose a patch quality estimation method based on the pre-trained vision model finetuned with a self-supervised training method. Next, we propose a weighted training approach to train the model with the estimated patch quality, which rewards high-quality image patches while penalizing those of low quality. Experiment results demonstrate that PatchDPO achieves state-of-the-art performance on both single-object and multi-object personalized image generation. We hope our method and dataset can contribute to the community of personalized image generation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by Zhejiang Province High-Level Talents Special Support Program \"Leading Talent of Technological Innovation of Ten-Thousands Talents Program\" (No. 2022R52046), and Alibaba-Zhejiang University Joint Research Institute of Frontier Technologies." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18376" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] André Araujo, Wade Norris, and Jack Sim. Computing receptive fields of convolutional neural networks. Distill, 2019. https://distill.pub/2019/computing-receptive-fields. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "type": "text", + "content": "[2] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 225 + ], + "type": "text", + "content": "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR 2017, pages 3852-3861. IEEE, 2017. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 259 + ], + "type": "text", + "content": "[4] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In *ICLR* 2024. OpenReview.net, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 294, + 304 + ], + "type": "text", + "content": "[5] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan Su. This looks like that: Deep learning for interpretable image recognition. In NeurIPS 2019, pages 8928-8939, 2019. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 294, + 339 + ], + "type": "text", + "content": "[6] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W. Cohen. Re-imagen: Retrieval-augmented text-to-image generator. In ICLR 2023. OpenReview.net, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 294, + 373 + ], + "type": "text", + "content": "[7] Kevin Clark, Paul Vicol, Kevin Swersky, and David J. Fleet. Directly fine-tuning diffusion models on differentiable rewards. In ICLR 2024. OpenReview.net, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 376, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 408 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR 2009, pages 248-255. IEEE, 2009. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 410, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 294, + 464 + ], + "type": "text", + "content": "[9] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS 2023, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 466, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 466, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 57, + 466, + 294, + 520 + ], + "type": "text", + "content": "[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR 2023. OpenReview.net, 2023. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 522, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 586 + ], + "type": "text", + "content": "[11] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, Yixiao Ge, Ying Shan, and Mike Zheng Shou. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. In NeurIPS 2023, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 611 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS 2020, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 666 + ], + "type": "text", + "content": "[13] Qihan Huang, Mengqi Xue, Wenqi Huang, Haofei Zhang, Jie Song, Yongcheng Jing, and Mingli Song. Evaluation and improvement of interpretability for self-explainable part-prototype networks. In ICCV 2023, pages 2011–2020. IEEE, 2023. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 294, + 713 + ], + "type": "text", + "content": "[14] Qihan Huang, Siming Fu, Jinlong Liu, Hao Jiang, Yipeng Yu, and Jie Song. Resolving multi-condition confusion for finetuning-free personalized image generation. arXiv preprint arXiv:2409.17920, 2024. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[15] Qihan Huang, Jie Song, Jingwen Hu, Haofei Zhang, Yong Wang, and Mingli Song. On the concept trustworthiness in concept bottleneck models. In AAAI 2024, pages 21161-21168, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[16] Qihan Huang, Jie Song, Mengqi Xue, Haofei Zhang, Bingde Hu, Huiqiong Wang, Hao Jiang, Xingen Wang, and Mingli Song. LG-CAV: train any concept activation vector with language guidance. In NeurIPS 2024, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "text", + "content": "[17] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and João Carreira. Perceiver: General perception with iterative attention. In ICML 2021, pages 4651-4664. PMLR, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 259 + ], + "type": "text", + "content": "[18] Xuhui Jia, Yang Zhao, Kelvin CK Chan, Yandong Li, Han Zhang, Boqing Gong, Tingbo Hou, Huisheng Wang, and Yu-Chuan Su. Taming encoder for zero fine-tuning image customization with text-to-image diffusion models. arXiv preprint arXiv:2304.02642, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[19] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 369 + ], + "type": "text", + "content": "[20] Yongcheng Jing, Chongbin Yuan, Li Ju, Yiding Yang, Xinchao Wang, and Dacheng Tao. Deep graph reprogramming. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24345-24354, 2023. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In ICCV 2023, pages 3992-4003. IEEE, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "text", + "content": "[22] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In CVPR 2023, pages 1931-1941. IEEE, 2023. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[23] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In NeurIPS 2023, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "text", + "content": "[24] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept neurons in diffusion models for customized generation. In ICML 2023, pages 21548-21566. PMLR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "text", + "content": "[25] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard S. Zemel. Understanding the effective receptive field in deep convolutional neural networks. In NeurIPS, pages 4898-4906, 2016. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 658 + ], + "type": "text", + "content": "[26] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In SIGGRAPH 2024, page 25. ACM, 2024. 1, 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[27] Jacob Menick, Maja Trebacz, Vladimir Mikulik, John Aslanides, Francis Song, Martin Chadwick, Mia Glaese, Susannah Young, Lucy Campbell-Gillingham, Geoffrey Irving, et al. Teaching language models to support answers with verified quotes. arXiv preprint arXiv:2203.11147, 2022. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18377" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[28] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models. In ICLR 2024. OpenReview.net, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[29] Maitreya Patel, Sangmin Jung, Chitta Baral, and Yezhou Yang. lambda-eclipse: Multi-concept personalized text-to-image diffusion models by leveraging clip latent space. arXiv preprint arXiv:2402.05195, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[31] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[32] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D. Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS 2023, 2023. 1, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 352 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR 2022, pages 10674-10685. IEEE, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 397 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR 2022, pages 10674-10685. IEEE, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR 2023, pages 22500-22510. IEEE, 2023. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "text", + "content": "[36] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In CVPR 2024, pages 8543-8552. IEEE, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 554 + ], + "type": "text", + "content": "[37] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR 2024, pages 14398-14409. IEEE, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 611 + ], + "type": "text", + "content": "[38] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR 2024, pages 8228-8238. IEEE, 2024. 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "text", + "content": "[39] X Wang, Siming Fu, Qihan Huang, Wanggui He, and Hao Jiang. Ms-diffusion: Multi-subject zero-shot image personalization with layout guidance. arXiv preprint arXiv:2406.07209, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[40] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. ELITE: encoding visual concepts into textual embeddings for customized text-to-image generation. In ICCV 2023, pages 15897-15907. IEEE, 2023. 2, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 407 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "text", + "content": "[41] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. arXiv preprint arXiv:2305.10431, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "type": "text", + "content": "[42] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 553, + 195 + ], + "type": "text", + "content": "[43] Fei Xue, Ignas Budvytis, and Roberto Cipolla. SFD2: semantic-guided feature detection and description. In CVPR 2023, pages 5206-5216. IEEE, 2023. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 315, + 197, + 553, + 251 + ], + "type": "text", + "content": "[44] Mengqi Xue, Qihan Huang, Haofei Zhang, Jingwen Hu, Jie Song, Mingli Song, and Canghong Jin. Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition. In IJCAI 2024, pages 1516-1524. ijcai.org, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 253, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 253, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 315, + 253, + 553, + 295 + ], + "type": "text", + "content": "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 297, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 297, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 315, + 297, + 553, + 350 + ], + "type": "text", + "content": "[46] Yu Zeng, Vishal M. Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In CVPR 2024, pages 6786-6795. IEEE, 2024. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 353, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 353, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 315, + 353, + 553, + 407 + ], + "type": "text", + "content": "[47] Yuxuan Zhang, Yiren Song, Jiaming Liu, Rui Wang, Jinpeng Yu, Hao Tang, Huaxia Li, Xu Tang, Yao Hu, Han Pan, et al. Ssr-encoder: Encoding selective subject representation for subject-driven generation. In CVPR 2024, pages 8069–8078. IEEE, 2024. 2, 6" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18378" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/6310723f-d72e-495d-87c7-6b59d06d7ca9_content_list.json b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/6310723f-d72e-495d-87c7-6b59d06d7ca9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a0215a5b6be8abadad3e7544590273d950b70497 --- /dev/null +++ b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/6310723f-d72e-495d-87c7-6b59d06d7ca9_content_list.json @@ -0,0 +1,1532 @@ +[ + { + "type": "text", + "text": "PatchGuard: Adversarily Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mojtaba Nafez1 Amirhossein Koochakian1 * Arad Maleki1 * Jafar Habibi1 Mohammad Hossein Rohban1 1Sharif University of Technology, Iran", + "bbox": [ + 246, + 202, + 754, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mojtaba.nafez77, amir.kochakian12, arad.maleki02, rohban}@sharif.edu", + "bbox": [ + 225, + 260, + 767, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anomaly Detection (AD) and Anomaly Localization (AL) are crucial in fields that demand high reliability, such as medical imaging and industrial monitoring. However, current AD and AL approaches are often susceptible to adversarial attacks due to limitations in training data, which typically include only normal, unlabeled samples. This study introduces PatchGuard, an adversarily robust AD and AL method that incorporates pseudo anomalies with localization masks within a Vision Transformer (ViT)-based architecture to address these vulnerabilities. We begin by examining the essential properties of pseudo anomalies, and follow it by providing theoretical insights into the attention mechanisms required to enhance the adversarial robustness of AD and AL systems. We then present our approach, which leverages Foreground-Aware Pseudo-Anomalies to overcome the deficiencies of previous anomaly-aware methods. Our method incorporates these crafted pseudo-anomaly samples into a ViT-based framework, with adversarial training guided by a novel loss function designed to improve model robustness, as supported by our theoretical analysis. Experimental results on well-established industrial and medical datasets demonstrate that PatchGuard significantly outperforms previous methods in adversarial settings, achieving performance gains of $53.2\\%$ in AD and $68.5\\%$ in AL, while also maintaining competitive accuracy in non-adversarial settings. The code repository is available at here.", + "bbox": [ + 88, + 340, + 485, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, Anomaly Detection (AD) and Anomaly Localization (AL) have become critical techniques across various fields, including medical imaging, fraud detection, and industrial monitoring [22, 36, 42, 47, 68, 70, 73]. While state-of-the-art (SOTA) AD and AL methods exhibit near-perfect performance on standard benchmarks [6, 83], they", + "bbox": [ + 89, + 786, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f77e3f787947d6371e0f60b757d73fc6f18f6d3ff42807f4c70ce0a0b193ecb8.jpg", + "image_caption": [ + "Figure 1. Impact of Adversarial Attacks on Anomaly Localization Methods: Localization maps for multiple methods are shown before and after a PGD-1000 attack, illustrating the vulnerability of existing methods to adversarial attacks, even when performing perfectly in clean conditions. Our proposed method demonstrates enhanced robustness in these adversarial scenarios." + ], + "image_footnote": [], + "bbox": [ + 513, + 306, + 906, + 433 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "frequently face significant performance degradation when subjected to adversarial attacks, as illustrated in Figure 1 [12, 54, 64, 66, 74, 81]. These attacks involve introducing subtle, imperceptible perturbations into the input data, misleading models to misclassify normal1 samples as anomalies or to overlook actual anomalies [2, 24, 48, 65, 80]. To address this vulnerability, a variety of strategies have been proposed to strengthen the adversarial robustness of AD systems [3, 7, 10, 25, 45, 61]. However, despite these advancements, the resilience of these methods remains limited, with even mild adversarial attacks still causing noticeable performance drops, even on widely used anomaly benchmarks [6, 83]. Furthermore, adversarial robustness in AL remains an unexplored area, highlighting a critical gap in the current research.", + "bbox": [ + 511, + 556, + 908, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial training, a strategy designed to improve model robustness in classification tasks, incorporates adversarial samples into the training process [48, 76]. Despite its success in standard classification, this approach proves less effective in AD and AL scenarios [51]. The primary", + "bbox": [ + 511, + 785, + 908, + 861 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1In this study, the term 'normal' refers to inlier samples, not to be mistaken with Gaussian distribution.", + "bbox": [ + 511, + 875, + 906, + 898 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "*Equal Contribution", + "bbox": [ + 109, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20383", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "obstacle to achieving adversarial robustness in AD lies in the absence of anomalous samples during training [10, 11]. Without such samples, models cannot be trained on adversarial perturbations specific to anomalous data [5, 51, 56, 58], leaving them vulnerable to attacks on perturbed anomalies [3, 10, 11]. To address this issue in the AD setup, recent efforts have combined Outlier Exposure (OE) [32] with adversarial training [3, 10, 11]. However, this method struggles when applied to high-resolution, real-world datasets [51]. This challenge is even more pronounced in AL settings, as robust AL methods must perform pixel-level classification, which inherently requires pixel-level supervision, while at the same time operating within the same data constraints as the AD framework [35, 82]. Despite ongoing research, the adversarial robustness of AL methods remains an open problem, with no comprehensive solutions proposed to date.", + "bbox": [ + 89, + 90, + 485, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we identify two essential properties that OE techniques must satisfy for effective application in adversarially robust AD and AL. First, the outlier exposures must include localization maps to facilitate the training of robust AL models. This requirement aligns with prior research on adversarially robust segmentation models, which underscores the importance of generating localization maps to enhance model robustness [39, 46, 69]. Second, as noted by [51], OE samples should be \"near-distributed\" relative to normal data, meaning they should share similar semantic and stylistic attributes. This condition is consistent with the literature on adversarial robustness, which highlights the advantages of near-distribution decision boundaries samples for optimizing the model robustness [71].", + "bbox": [ + 89, + 335, + 485, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, the recent success of Vision Transformer (ViT)-based architectures in clean AD and AL tasks [15, 37, 41, 64], attributed to their global contextual modeling [19], motivated us to explore their potential in adversarial settings. For an input image $x$ and a fixed attention layer $\\ell$ , we define the Attention Degree of $\\ell$ under $x$ as the average number of input embeddings each output embedding attends to. As shown in Figure 2, we observe that images that induce higher attention degrees in the model are more resistant to adversarial attacks. Furthermore, adversarial training appears to implicitly steer the model toward achieving a high attention degree. We provide a theoretical analysis and justification for these findings in Section 4.", + "bbox": [ + 88, + 550, + 485, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these insights, we propose PatchGuard, an adversarially robust AD and AL method that addresses limitations in current OE-based methods and leverages attention-based insights. To address OE challenges, we generate pseudo-anomaly samples from normal data, ensuring they meet two essential criteria: availability of localization maps and near-distribution alignment with inliers. To create samples that exhibit anomalous characteristics while remaining close to the inlier distribution, we introduce Foreground-Aware Pseudo-Anomaly Generation. This approach iden", + "bbox": [ + 89, + 750, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tifies causal regions in normal samples using an enhanced Grad-CAM [60] feature attribution technique with tailored augmentations. These identified regions are then selectively distorted, drawing on insights from [55], to produce anomaly samples that retain stylistic and semantic coherence with the original data. By precisely controlling the distortion location, we can generate accurate localization maps for the pseudoanomalies, effectively meeting both criteria for robust AD and AL.", + "bbox": [ + 511, + 90, + 906, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the next step, we leverage normal samples and pseudoanomalies to perform adversarial training on a ViT-based architecture. Building on the provided insights, we introduce a novel loss function designed to increase the model's last-layer attention degree. As our theoretical justifications suggest, this property significantly enhances the model's adversarial robustness in both AD and AL settings.", + "bbox": [ + 511, + 227, + 908, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contribution. In this paper, we propose PatchGuard, an adversarially robust AD and AL method. To the best of our knowledge, we are the first to raise and effectively address the issue of adversarial robustness in AL. We provide insights into the shortcomings of previous OE-based methods and introduce the Foreground-Aware Pseudo-Anomaly Generation framework to address these limitations. Additionally, we provide insights and theoretical justifications on enhancing the adversarial robustness of ViT-based architectures, and propose a novel loss function aligned with these findings. We evaluate PatchGuard in both clean and adversarial settings, rigorously testing its robustness under several strong attacks, including PGD-1000 [48], $\\mathrm{A}^3$ [43], and SEA [17]. Results demonstrate that PatchGuard significantly outperforms existing methods in adversarial contexts, achieving an AUROC improvement of $53.2\\%$ in AD and $68.5\\%$ in AL, while maintaining competitive performance in standard conditions. Our evaluations span various well-established industrial and medical datasets, including MVTec [6], VisA [83], and BraTS2021 [4], highlighting PatchGuard's applicability.", + "bbox": [ + 511, + 333, + 908, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 664, + 656, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From an adversarial robustness perspective, anomaly detection and localization methods can be mainly classified into three categories: 1) methods with anomaly-free training, 2) methods that leverage anomalies at the embedding level, and 3) methods that leverage anomalies at the input level. We experimentally adapted methods from each category to adversarial training processes, and the results reveal they still exhibit significant vulnerabilities even after adversarial training. The first category consists of methods such as Recontrast[29], Transformaly [15], and PatchCore [57], that operate on the hypothesis that models trained exclusively on normal samples can reconstruct normal images well but fail to do so with abnormal regions. These methods lack robustness, since the model does not encounter any anomaly", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20384", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "data during adversarial training, meaning that when adversarial perturbations are added to anomaly data, the model still remains vulnerable. The second category, which includes methods such as SimpleNet [44], GeneralAD [64], UniAD [75], and DSR [79], utilize a feature extractor during training to obtain feature vectors of the normal set. By perturbing these features, they create anomaly features and primarily train a discriminator over these normal and anomaly features. However, these methods lack robustness due to their reliance on highly vulnerable feature extractors and the absence of anomaly data in the input space, which prevents proper training of a robust feature extractor. The third category consists of DRAEM [78], NSA [59], Cutpaste [40], and GLASS [12], generate pseudo-anomaly data during training. Nevertheless, some of these methods lack access to masks or do not use a foreground-aware generation process. Despite their astonishing success in clean settings, we empirically show in Table 3 that even when adversarily trained, they exhibit unsatisfactory results. Furthermore, while anomaly localization has not been explored in an adversarial setup, some other works have pursued anomaly detection in an adversarial setting, including PrincipaLS [45], ZARND [52], and Rodeo [51]. These approaches have achieved relatively better results by incorporating pseudo-anomaly generation processes and adversarial training. However, their performance falls short in real-world high-resolution datasets (as presented in Table 1). For more details about previous works, see Appendix A.", + "bbox": [ + 89, + 90, + 485, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 512, + 228, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anomaly Detection (AD) and Localization (AL): AD involves identifying instances that deviate from expected patterns, while AL focuses on localizing abnormal subregions within an image. In AD, a model assigns an anomaly score $S(x; f)$ to each sample $x$ using model $f$ ; if this score exceeds a predefined threshold, the sample is classified as anomalous. Conversely, lower scores indicate normal samples. In AL, a model generates an anomaly map $M(x; f)$ using model $f$ , where each pixel of the map corresponds to an anomaly score for the respective pixel in the input image.", + "bbox": [ + 89, + 537, + 483, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Robustness of Anomaly Detectors: An adversarial attack, primarily studied within classification tasks, involves the malicious alteration of a data sample $x$ with its label $y$ into a new sample $x^{*}$ that maximizes the loss function $\\ell(x^{*};y)$ [34, 72, 76]. This new sample $x^{*}$ is known as an adversarial example of $x$ , and the difference $(x^{*} - x)$ is referred to as adversarial perturbation. To prevent significant semantic changes, the $l_{p}$ norm of the adversarial noise is restricted by an upper limit $\\epsilon$ . Specifically, an adversarial example $x^{*}$ must satisfy the condition $x^{*} = \\arg \\max_{x^{\\prime}:\\|x - x^{\\prime}\\|_{p} \\leq \\epsilon} \\ell(x^{\\prime};y)$ . A widely used and effective attack method is the Projected Gradient Descent (PGD) [48] technique, which iteratively increases the loss function by moving in the direction of the gradient sign of", + "bbox": [ + 89, + 688, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\ell(x^{*};y)$ with a specified step size $\\alpha$ . When adapting adversarial attacks for AD, the objective shifts from maximizing the loss value to increasing $S(x,f)$ for normal samples and decreasing it for anomalies. The attack is formulated as $x_0^* = x$ , $x_{t+1}^* = x_t^* + y \\cdot \\alpha \\cdot \\mathrm{sign}(\\nabla_x S(x_t^*;f_\\theta))$ , and $x^* = x_k^*$ , where $y = +1$ for normal samples and $y = -1$ for anomaly samples. This approach is consistently applied to other attacks in the study [21, 76].", + "bbox": [ + 511, + 90, + 906, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Robustness of Anomaly Localization: Adversarial attacks aim to introduce perturbations that uniformly modify pixel scores in an anomaly map $M$ , flipping scores from 0 to 1 or vice versa. In the context of AL, the attack objective transitions from merely maximizing a loss function to specifically increasing the values of $M(x, f)$ for normal samples while decreasing them for anomalies. The attack is formalized as follows: starting with $x_0^* = x$ , the adversarial example at each step is updated as $x_{t+1}^* = x_t^* + Y \\cdot \\alpha \\cdot \\mathrm{sign}(\\nabla_x M(x_t^*; f_\\theta))$ , and the final adversarial example is $x^* = x_k^*$ . Here, $Y_{i,j} = +1$ for normal pixels and $Y_{i,j} = -1$ for anomaly pixels, where $i, j$ denote pixel coordinates. This strategy is extended to other advanced attack methods throughout the study [17]. Despite its significance, adversarial robustness in anomaly localization remains underexplored. Enhancing robustness is particularly critical for applications requiring spatial precision, such as medical diagnostics and industrial monitoring systems.", + "bbox": [ + 511, + 212, + 908, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Theoretical Insights", + "text_level": 1, + "bbox": [ + 513, + 500, + 705, + 518 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/79c34287f45507354ea2031fb5ff756f66282d5125e01d025f8cd07a61938e5a.jpg", + "image_caption": [ + "Figure 2. The figure demonstrates how increasing the average last-layer attention degree in the ViT-base architecture reduces vulnerability to adversarial attacks. Specifically, the BraTS dataset images were clustered into five groups based on their attention degree values (x-axis). The y-axis represents Vulnerability, measured as the absolute difference in the model's localization performance (AUROC%) for each cluster before and after a PGD attack. (a) illustrates the decreasing trend in a clean-trained model. (b) shows a similar decreasing trend in the adversarially trained model, where the attention degree is relatively higher than in the clean model." + ], + "image_footnote": [], + "bbox": [ + 516, + 542, + 705, + 657 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a0c1739ded82b78932cbf62868f2289f6788980eaa63019daa71bbb90600b27e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 541, + 903, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we want to show why a self-attention layer, in which each token attends to a higher number of other tokens is more adversarially robust compared to the case that the attention is distributed over a smaller number of tokens.", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20385", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let's consider the self-attention output embedding of the $i$ th token in the $l$ -th layer as $x_{i}^{(l)}$ , and such embedding after the perturbation is applied to the input as $\\hat{x}_i^{(l)}$ . We are seeking to set $x_{i}^{(2)} - \\hat{x}_{i}^{(2)} = r_{i}$ , where $r_i$ is the desired adversarial direction that most influences the layers next to the first layer of the transformer. According to the self-attention:", + "bbox": [ + 89, + 90, + 483, + 189 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i} ^ {(l + 1)} = V ^ {(l)}. \\text {s o f t m a x} \\left(q _ {i} ^ {(l) \\top} K / \\sqrt {d}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 200, + 415, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q_{i}^{(l)} = W_{q}x_{i}^{(l)},v_{j}^{(l)}$ the $j$ -th column of $V^{(l)}$ be $v_{j}^{(l)} = W_{v}x_{j}^{(l)},k_{j}^{(l)}$ the $j$ -th column of $\\mathbf{K}$ be $k_{j}^{(l)} = W_{k}x_{j}^{(l)}$ , and $d$ be the embedding dimension. For simplicity, we assume that the dimensions of embeddings of key, query, and values are all the same and are equal to the dimension of the original embeddings. Now, $W_{v}(X^{(1)}S - \\hat{X}^{(1)}\\hat{S}) = R$ , where the columns of $R$ and $X^{(l)}$ are represented by $r_i$ and $x_{i}^{(l)}$ , and the columns in $S$ and $\\hat{S}$ correspond to the softmax values of each query token calculated for $X$ and the perturbed inputs $\\hat{X}$ , respectively. Therefore,", + "bbox": [ + 89, + 223, + 483, + 387 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {v} (X ^ {(1)} S - \\hat {X} ^ {(1)} S) =\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 393, + 307, + 411 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {v} \\left(X ^ {(1)} S - \\hat {X} ^ {(1)} \\hat {S} + \\hat {X} ^ {(1)} \\hat {S} - \\hat {X} ^ {(1)} S\\right) =\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 414, + 431, + 433 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR + W _ {v} \\hat {X} ^ {(1)} (\\hat {S} - S) = R + W _ {v} \\hat {X} ^ {(1)} \\Delta S.\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 436, + 426, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta S\\coloneqq \\hat{S} -S$ . Hence, we have:", + "bbox": [ + 89, + 464, + 349, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {(1)} - \\hat {X} ^ {(1)} = \\underbrace {W _ {v} ^ {+} (R + W _ {v} \\hat {X} ^ {(1)} \\Delta S)} _ {U} S ^ {+}.\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 486, + 433, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $W_{v}^{+}$ and $S^{+}$ are the pseudo-inverses of the matrices $W_{v}$ and $S$ , respectively. Note that for the $\\delta := X^{(1)} - \\hat{X}^{(1)}$ perturbation in the input, in order to make the required perturbation in the output of the network, we get $\\| \\delta \\| \\leq \\lambda_{max} \\| U \\|$ , where $\\lambda_{max}$ is the maximum eigenvalue of $S^{+}$ . But we note that for a stochastic matrix $S$ , the maximum eigenvalue is 1. Therefore, for the inverse matrix, $S^{+}$ , the maximum eigenvalue would be larger than 1. As a result, we would expect the $\\| \\delta \\|$ to grow proportional to $\\lambda_{max} \\geq 1$ . We note that the smaller $\\| \\delta \\|$ would imply more vulnerability as it becomes easier to attack the network with a smaller perturbation length. On the other hand, for a purely localized $S$ , where each token only attends to itself $\\lambda_{max} = 1$ , showing that the increase in $\\| \\delta \\|$ would likely be small compared to the general case, leading to more vulnerability. Therefore, we conclude that a localized and overly sparse attention map would typically imply an easier attack on the network and hence higher vulnerability against adversarial attacks. Sparse attention values may also resemble what happens in convolutional networks that have small receptive fields.", + "bbox": [ + 89, + 530, + 483, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Methodology", + "text_level": 1, + "bbox": [ + 89, + 844, + 225, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. Current AD and AL methods perform poorly when subjected to adversarial attacks. Previous studies have", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "demonstrated the effectiveness of OE in enhancing robustness against such attacks [3, 10, 11, 51]. Additionally, ViTs have shown promising results in standard (clean) AD and AL tasks [15, 37, 41, 64], and our findings suggest their potential for adversarial settings with targeted modifications of attention degrees. Building on these insights and our theoretical foundations, we propose PatchGuard, a novel method designed to enhance adversarial robustness in AD and AL. The following subsections detail the key components of PatchGuard.", + "bbox": [ + 511, + 90, + 906, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Foreground-Aware Pseudo-Anomaly Generation", + "text_level": 1, + "bbox": [ + 511, + 251, + 906, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate anomalous data from normal samples, it is essential to apply targeted distortions that ensure the image becomes anomalous [8, 13, 14, 26]. Additionally, the generated pseudo-anomalies must satisfy two key requirements: (1) inclusion of localization maps to enhance the AL task, and (2) close alignment with the normal sample distribution [51, 53]. Meeting these requirements simultaneously necessitates precise but minimal alterations within the foreground. Thus, the proposed method must reliably identify a region of the image that we can confidently attribute to the foreground.", + "bbox": [ + 511, + 287, + 908, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this, we leverage insights from [55], which indicate that pretrained models benefit from style-agnostic processing to improve classification accuracy. To this end, we first introduce soft transformations, which are commonly used in self-supervised learning studies due to their ability to preserve semantic content [8, 13, 14, 26, 31]. Next, we apply Grad-CAM [60] to generate a saliency map using a standard pre-trained ResNet18 [30] model on a given input. Specifically, for a normal sample $x$ , we randomly select $k_{\\mathrm{soft}}$ soft transformations $t_1^s,\\ldots ,t_{k_{\\mathrm{soft}}}^s$ . We then compute saliency maps for $x,t_1^s (x),\\ldots ,t_{k_{\\mathrm{soft}}}^s (x)$ and take their element-wise product to produce a style-invariant saliency map, $G(x)$ . As shown in Figure 3 (part ①), $G(x)$ effectively assigns higher values to the foreground region.", + "bbox": [ + 511, + 439, + 908, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To distort the obtained region, we first randomly sample an anchor point from it. We then sample $k_{\\mathrm{hard}}$ hard transformations $t_1^h, \\ldots, t_{k_{\\mathrm{hard}}}^h$ , which are typically associated with altering semantic integrity, as highlighted in prior research [1, 9, 18, 23, 33, 38, 40, 62, 63, 66, 77]. We then design a rectangular mask centered on this chosen point, with its width $w$ and height $h$ satisfying $w \\sim U(0.05W, 0.3W)$ and $h \\sim U(0.05H, 0.3H)$ , where $W$ and $H$ are the image's width and height. Next, we rotate the mask by an angle of $\\theta \\sim U(-45^\\circ, 45^\\circ)$ . Finally, we sequentially apply the hard transformations only to the region within this mask, leaving the remaining unmasked areas of the image unchanged. Mathematically, the final anomaly is obtained by applying", + "bbox": [ + 511, + 650, + 910, + 848 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {\\prime} = t _ {1} ^ {h} \\left(\\dots \\left(t _ {k _ {\\text {h a r d}}} ^ {h} \\left(x _ {\\text {m a s k e d}}\\right) \\dots\\right) + \\left(x - x _ {\\text {m a s k e d}}\\right). \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 856, + 906, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An additional advantage of this pseudo-anomaly crafting ap-", + "bbox": [ + 511, + 885, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20386", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5c9deb95b88b8d22cd4d1f87d9d1a4fc393f0c62e6fcb33c765505e6e90afbe9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 88, + 459, + 285 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d5cb6698cfd038056731f52aa802055062cddea6a312e0abf28dd5f8bc324b6b.jpg", + "image_caption": [ + "Figure 3. Overview of the PatchGuard framework. (1) We use Grad-CAM to identify regions likely to belong to the foreground in an image. By applying $k_{\\mathrm{soft}} = 3$ augmentations to the image $x$ , we generate a combined saliency map $G(x)$ . (2) The input $x$ and its saliency map $G(x)$ are then passed to the anomaly generator to produce an anomaly sample $x'$ along with its ground-truth mask $x_m'$ . (3) For each normal sample $x_i$ , the input batch includes the image, its corresponding anomaly version $x_i'$ , and their adversarially attacked variants, along with their ground-truth masks. Each batch sample is processed through a Vision Transformer (ViT), where the Attention Discriminator assigns an anomaly score $p_i$ to each patch embedding. The final anomaly map is constructed from these scores and trained with a novel loss $\\mathcal{L}$ to replicate the ground truth." + ], + "image_footnote": [], + "bbox": [ + 101, + 305, + 439, + 450 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4d5ed744a1270fd6bb81971e28608eccaf9d6ed5241360a8cd2aa11bc6010e51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 89, + 897, + 452 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/827ad7a3dc58fc161fd52f1378f0b682782b1e7d5239446832ff9ea3c5e1bcd6.jpg", + "table_caption": [ + "Table 1. Detection performance of various AD methods, including PatchGuard, across benchmark datasets. Results are shown as “Clean / Adversarial”, representing performance on “Clean” and “Adversarial” data (PGD-1000, $\\epsilon = \\frac{8}{255}$ , measured by AUROC %), respectively." + ], + "table_footnote": [ + "*These works incorporated adversarial training into their proposed AD methods." + ], + "table_body": "
DatasetMethod
DRAEMPatchCoreSimpleNetReContrastGeneralADGLADGLASSPrincipaLS*ZARND*RODEO*Ours
MVTec AD98.0 / 5.199.1 / 3.999.6 / 2.399.5 / 0.099.2 / 13.799.3 / 11.899.9 / 6.763.8 / 10.671.6 / 11.061.5 / 4.988.1 / 71.1
VisA88.7 / 3.895.1 / 1.796.2 / 0.297.5 / 6.795.9 / 14.599.5 / 8.698.8 / 7.258.9 / 9.367.7 / 9.363.1 / 3.888.5 / 74.3
BTAD87.1 / 4.996.9 / 2.495.1 / 4.695.7 / 2.197.0 / 10.998.5 / 7.697.4 / 8.360.1 / 7.869.7 / 8.360.0 / 5.885.3 / 82.1
MPDD94.3 / 2.891.3 / 6.896.6 / 7.094.5 / 4.598.0 / 12.897.5 / 9.799.6 / 9.171.9 / 15.765.7 / 12.268.8 / 7.485.4 / 68.6
WFDD94.0 / 7.196.3 / 3.998.8 / 5.197.6 / 3.399.0 / 10.799.1 / 12.5100 / 6.066.3 / 14.176.1 / 16.371.2 / 7.884.2 / 65.0
DTD-Synthetic97.1 / 3.665.8 / 4.096.8 / 2.498.2 / 9.197.9 / 16.398.6 / 10.399.1 / 6.177.9 / 9.370.3 / 13.772.8 / 12.091.5 / 60.5
Head-CT72.1 / 8.192.4 / 5.184.9 / 4.886.1 / 6.389.1 / 9.193.1 / 14.294.7 / 8.773.2 / 16.064.3 / 9.785.7 / 61.097.3 / 90.4
BraTS202162.3 / 1.491.6 / 6.382.5 / 6.182.4 / 3.188.3 / 13.092.3 / 12.095.7 / 7.275.3 / 8.766.7 / 14.089.1 / 64.794.3 / 81.0
Average86.7 / 4.691.0 / 4.293.8 / 4.093.9 / 4.395.5 / 12.697.2 / 10.898.1 / 7.468.4 / 11.469.0 / 11.871.5 / 20.989.3 / 74.1
", + "bbox": [ + 91, + 632, + 906, + 854 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20387", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/532dd786bbcf664ffd4ca0b9427489b7c4063a650c736d7d4f9f891a4c6b1e4f.jpg", + "table_caption": [ + "Table 2. Localization performance of various AL methods, including PatchGuard, across benchmark datasets. Results are shown as \"Clean / Adversarial\", representing performance on \"Clean\" and \"Adversarial\" data (PGD-1000, $\\epsilon = \\frac{8}{255}$ , measured by AUROC %), respectively." + ], + "table_footnote": [], + "table_body": "
DatasetMethod
DRAEMPatchCoreSimpleNetReContrastGeneralADGLADGLASSMuScReConPatchRealNetOurs
MVTec AD97.3 / 3.998.1 / 4.298.1 / 6.898.4 / 4.897.7 / 12.398.7 / 15.399.3 / 8.197.3 / 4.699.2 / 2.899.0 / 8.792.7 / 73.8
VisA90.7 / 5.698.5 / 3.097.4 / 4.898.2 / 7.199.0 / 7.698.6 / 13.498.8 / 7.098.8 / 3.899.2 / 1.798.8 / 6.296.9 / 85.2
BTAD89.0 / 7.192.6 / 6.395.2 / 3.796.8 / 6.496.7 / 9.398.2 / 14.597.0 / 5.797.3 / 5.397.5 / 3.097.9 / 9.293.2 / 73.0
MPDD90.7 / 4.798.5 / 1.997.4 / 7.394.6 / 3.698.1 / 14.998.7 / 10.899.4 / 6.395.5 / 6.096.7 / 2.698.2 / 7.593.8 / 86.6
WFDD95.8 / 3.998.1 / 2.198.0 / 3.497.3 / 5.198.7 / 13.597.6 / 11.398.9 / 7.697.0 / 4.198.3 / 4.998.4 / 6.894.6 / 71.6
DTD-Synthetic96.1 / 4.895.6 / 6.797.0 / 2.996.8 / 7.697.9 / 10.897.3 / 13.898.1 / 7.097.8 / 5.796.3 / 3.998.0 / 8.395.9 / 74.8
Head-CT86.4 / 8.987.3 / 6.189.7 / 0.890.1 / 8.691.7 / 18.393.4 / 16.792.2 / 9.185.2 / 8.589.7 / 6.694.8 / 7.393.6 / 89.3
BraTS202182.2 / 6.196.9 / 4.594.7 / 6.795.3 / 11.196.8 / 14.795.1 / 13.594.0 / 6.367.5 / 6.397.5 / 5.996.9 / 10.297.7 / 94.5
Average91.0 / 5.695.7 / 4.395.9 / 4.595.9 / 6.797.0 / 12.697.2 / 13.697.2 / 7.192.0 / 5.596.8 / 3.997.7 / 8.094.8 / 81.1
", + "bbox": [ + 91, + 130, + 906, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "proach is that the location of the distorted region is precisely known, allowing us to generate an accurate localization map $x_{m}^{\\prime}$ for AL training, where each pixel's value equals 1 if it's anomalous, and 0 otherwise. By creating an anomalous counterpart $x^{\\prime}$ and its corresponding localization map $x_{m}^{\\prime}$ for each normal sample $x$ in the training set, we provide the model with both the anomaly and its location, as shown in Figure 3 (part ②). This pairing equips the model with the information needed to improve its localization accuracy, as we will discuss in Section 5.2. For details on the soft and hard transforms, and ablations on the hypers used in this approach, refer to Table 6 and Appendix B.", + "bbox": [ + 88, + 387, + 485, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Anomaly Detection and Localization with Vision Transformer", + "text_level": 1, + "bbox": [ + 89, + 575, + 483, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enhance adversarial robustness through the effective use of pseudo-anomaly samples and their corresponding localization maps, we employ a Vision Transformer (ViT) model. This model generates output embeddings that are processed by an attention-based discriminator network, which uses a multi-head attention (MLH) mechanism followed by a shared-weight multi-layer perceptron (MLP). The discriminator independently assigns an anomaly score to each patch in the input, providing a patch-level anomaly prediction as shown in Figure 3 (part ③).", + "bbox": [ + 88, + 613, + 485, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The model is trained using a combination of patch-level cross-entropy (CE) and an additional regularization term $R$ , designed to encourage a higher degree of attention in a specified ViT layer. As illustrated in Figure 2 and the theory in Section 4, this regularizer is expected to further enhance adversarial robustness. The cross-entropy loss component aligns the predicted anomaly scores with ground truth labels from the localization map, $x_{m}$ , where a patch label equals 1 if more than $5\\%$ of the pixels in the corresponding patch in", + "bbox": [ + 88, + 763, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$x_{m}$ are anomalous, and 0 otherwise:", + "bbox": [ + 511, + 387, + 756, + 402 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM = \\operatorname {M L P} (\\operatorname {M L H} (x)) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 415, + 906, + 431 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} (x, x _ {m}) = \\sum_ {i, j \\in x _ {\\text {p a t c h e s}}} \\mathrm {C E} \\left(M _ {(i, j)}, x _ {m, (i, j)}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 435, + 906, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $M$ denotes the predicted AL map, consisting of anomaly scores for each patch, $x$ is any input sample (normal or anomaly), and $i,j$ are the indices for individual patches. The cross-entropy loss term $\\mathcal{L}_{\\mathrm{CE}}$ ensures alignment between the patch $M_{(i,j)}$ and the ground truth anomaly map $x_{m,(i,j)}$ , promoting precise anomaly localization. The final AL map is obtained by upsampling $M$ to the original image size, and we use the average of the top- $k$ patch-level anomaly scores as the image-level anomaly score.", + "bbox": [ + 511, + 479, + 908, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further improve the robustness of the model, we introduce a regularization term $R$ , which encourages an increased attention degree in a specified layer $\\ell$ of the ViT, chosen here as the final layer. This regularization term is formulated as:", + "bbox": [ + 511, + 617, + 908, + 678 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nR (x) = \\frac {1}{\\sum_ {i} \\sum_ {j} \\sum_ {k} A _ {i j k} \\cdot \\mathbb {1} \\left(A _ {i j k} \\leq \\delta\\right)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 689, + 906, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $A_{ijk}$ represents the attention coefficient corresponding to the $i$ -th head, the $j$ -th output patch, and the $k$ -th input patch. Regularization will increase the values of attention coefficients that are below the threshold $\\delta$ . For a specific output patch, since a softmax function is applied to its corresponding attention coefficients, increasing the lower values naturally decreases the higher ones. With $\\delta = \\frac{1}{(\\text{number of input patches})}$ , the coefficient distribution gradually shifts toward a uniform distribution at this threshold, representing the maximum attention degree a patch can achieve. The total loss function $\\mathcal{L}$ used for training the model is the", + "bbox": [ + 511, + 734, + 910, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20388", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sum of the cross-entropy and the regularization term:", + "bbox": [ + 89, + 90, + 442, + 106 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(x, x _ {m}\\right) = \\mathcal {L} _ {\\mathrm {C E}} \\left(x, x _ {m}\\right) + \\alpha R (x) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 118, + 483, + 135 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\alpha$ is a scaling factor that controls the strength of the regularization. This combined objective encourages the model to not only accurately predict anomaly locations but also maintain an increased level of attention, enhancing both detection performance and adversarial robustness. It is then used in adversarial training, as explained in Section 5.3. Comprehensive ablation studies of hyperparameters including $\\delta$ and $\\alpha$ are provided in Appendix C.2 and Section 7.", + "bbox": [ + 89, + 147, + 483, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Adversarial Training and Testing", + "text_level": 1, + "bbox": [ + 89, + 277, + 382, + 295 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given an input sample $x$ , an adversarial example $x_{adv}$ is generated by introducing a perturbation $\\delta^{*}$ , optimized to maximize the final loss:", + "bbox": [ + 89, + 300, + 483, + 345 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\delta^ {*} = \\underset {\\| \\delta \\| _ {\\infty} \\leq \\epsilon} {\\operatorname {a r g m a x}} \\mathcal {L} (x + \\delta , x _ {m}), \\quad x _ {a d v} = x + \\delta^ {*} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 357, + 483, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These adversarial examples are then used in the training process alongside the original examples. The adversarial training objective is framed as a min-max problem, optimizing the model parameters $\\theta$ to minimize the expected loss over both clean and adversarial examples:", + "bbox": [ + 89, + 396, + 483, + 472 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {(x, y) \\in \\mathcal {B}} \\left[ \\max _ {\\| \\delta \\| _ {\\infty} \\leq \\epsilon} \\mathcal {L} (x + \\delta , x _ {m}; \\theta) \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 482, + 483, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 527, + 223, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of PatchGuard, we conduct extensive experiments assessing its performance alongside several established anomaly detection methods on multiple benchmarks for anomaly detection and localization. Our evaluation includes both adversarially trained models and those trained under standard conditions. To ensure a fair comparison, we assess all methods under both \"Clean\" and \"Adversarial\" test conditions, as shown in Table 1 and 2. In the Clean setting, test samples are unaltered, whereas the Adversarial setting includes samples perturbed by the PGD attack [48]. Additional results on PatchGuard's performance against alternative attacks are provided in Section 7.2. Datasets are detailed in Appendix D.", + "bbox": [ + 88, + 551, + 482, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Details. In the adversarial setting, we evaluate each method using the $\\ell_{\\infty}$ PGD-1000 attack with $\\epsilon = \\frac{8}{255}$ . Detection and localization results are shown in Tables 1 and 2, respectively. $\\ell_{2}$ norm evaluations are in Appendix J.", + "bbox": [ + 89, + 750, + 483, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adapting State-of-the-Art AD and AL Methods to Adversarial Settings. Prior to developing our novel Patch-Guard framework, we first investigated whether existing state-of-the-art methods in AD and AL could be enhanced through adversarial training—a technique where models are", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1cb3e5a91fa83810ea064588b0d795aa98b7b12e149c353c957df7bb1c550fa5.jpg", + "table_caption": [ + "Table 3. Performance of prior methods adapted for adversarial settings, shown as \"Clean / Adversarial\" (PGD-1000, $\\epsilon = \\frac{8}{255}$ )." + ], + "table_footnote": [], + "table_body": "
MethodTaskDataset
MVTec ADVisABTADBraTS2021
PatchCoreAD92.3 / 14.388.4 / 10.887.1 / 12.788.6 / 16.1
AL89.1 / 10.387.6 / 8.885.0 / 11.586.7 / 13.6
ReContrastAD93.8 / 5.193.7 / 16.789.9 / 13.478.2 / 13.1
AL94.5 / 17.292.5 / 18.288.6 / 17.990.8 / 14.2
SimpleNetAD91.1 / 18.189.2 / 14.988.6 / 12.775.3 / 14.2
AL94.1 / 19.589.4 / 15.888.2 / 11.790.1 / 17.0
GeneralADAD84.6 / 24.781.2 / 22.182.9 / 20.680.7 / 19.9
AL83.7 / 21.883.1 / 17.384.7 / 18.379.6 / 21.0
DRAEMAD83.0 / 19.281.7 / 16.879.1 / 17.057.4 / 12.2
AL86.1 / 17.578.6 / 15.080.9 / 18.673.6 / 14.9
GLASSAD84.2 / 26.783.5 / 19.984.7 / 25.483.7 / 20.1
AL85.9 / 21.287.1 / 21.783.0 / 19.085.8 / 23.6
OursAD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 / 85.293.2 / 73.097.7 / 94.5
", + "bbox": [ + 514, + 130, + 906, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trained on adversarial examples to improve their robustness. In Appendix E, we provide a detailed methodology for generating adversarial examples for each method and describe our optimized approach for improving their resilience. The empirical results, presented in Table 3, reveal that even after incorporating adversarial training, current SOTA AD and AL methods remain vulnerable to adversarial attacks and exhibit suboptimal performance. These limitations motivated the development of PatchGuard, our proposed solution.", + "bbox": [ + 511, + 446, + 906, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analyzing Results. As shown in Tables 1 and 2, prior SOTA methods such as Recontrast demonstrate substantial drops in performance under adversarial settings, despite excelling in clean conditions. Furthermore, as indicated in Table 3, even when adversarily trained, these methods struggle to achieve robust performance. By contrast, approaches specifically designed to address adversarial robustness, such as RODEO, still fall short when compared to PatchGuard. On average, PatchGuard improves robust detection across diverse datasets by up to $53.2\\%$ percent, with a notable robust improvement of $68.5\\%$ percent on the challenging MVtec dataset. Notably, we achieve a $71.8\\%$ improvement in AL on the challenging VisA dataset, with details in Appendix F. As prior research indicates, a slight decrease in clean performance is generally viewed as an acceptable tradeoff for the critical gains in robustness.", + "bbox": [ + 511, + 583, + 908, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We leverage a from-scratch ViT model in our pipeline to showcase our setup's independence from pre-train datasets. However, we conduct an ablation study on this manner in Appendix C.4. Further, we leverage PGD-10 with $\\epsilon = \\frac{8}{255}$ for adversarial training,", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20389", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/41981ffa0a29b65e7ee394009613ce0690d64355fe0d75eb6f4ab8facf10400e.jpg", + "table_caption": [ + "Table 4. Evaluation of PatchGuard's performance against advanced classification and segmentation attacks, showcasing its robustness in diverse attack scenarios." + ], + "table_footnote": [], + "table_body": "
DatasetTaskDetection AttacksLocalization Attacks
PGD-1000CAAAutoAttackA3SegPGDSEA
MVTec ADAD71.174.870.869.871.969.2
AL73.876.772.970.074.671.3
VisAAD74.378.574.673.175.073.6
AL85.287.485.584.785.083.7
BTADAD82.183.482.080.781.980.8
AL73.074.772.371.572.971.1
BraTS2021AD81.084.781.980.382.081.1
AL94.595.094.293.894.893.4
", + "bbox": [ + 91, + 143, + 483, + 303 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7f53e183a163b55bb96e2123b52fd3f020b0494b93d323b8607df5868a9285c8.jpg", + "table_caption": [ + "Table 5. Ablation study on the scaling factor $\\alpha$ , illustrating the effectiveness of our regularization." + ], + "table_footnote": [], + "table_body": "
Scaling Factor (α)TaskDataset
MVTec-ADVisABTADBraTS2021
0AD89.3 / 60.790.1 / 62.988.6 / 74.395.2 / 72.0
AL93.5 / 65.395.8 / 72.194.0 / 59.798.2 / 83.4
1 (Ours)AD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 / 85.293.2 / 73.097.7 / 94.5
2AD86.2 / 70.187.6 / 75.184.7 / 83.492.8 / 81.2
AL91.0 / 68.994.8 / 84.191.9 / 73.995.8 / 94.6
5AD80.1 / 64.178.6 / 65.473.8 / 69.682.5 / 73.2
AL84.7 / 65.681.1 / 78.984.0 / 65.386.3 / 79.9
", + "bbox": [ + 91, + 359, + 483, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "but the test results are generated with the PGD-1000 attack. For more details on the implementation parameters, refer to Appendix G.", + "bbox": [ + 89, + 556, + 485, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 616, + 256, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we provided further analysis of our choices of parameters and other components of our method.", + "bbox": [ + 89, + 641, + 483, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.1. Regularization", + "text_level": 1, + "bbox": [ + 89, + 681, + 243, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To experimentally demonstrate the effectiveness of our regularization term on the model's robustness and the impact of the scaling factor $\\alpha$ (set to 1 by default in our method), we conduct an ablation study on $\\alpha$ , as shown in Table 5. When $\\alpha = 0$ , the model's robustness against adversarial attacks decreases due to the lack of regularization. On the other hand, higher values of $\\alpha$ , such as $\\alpha = 5$ , cause the regularization term to dominate, leading to an unstable loss function and inappropriate results.", + "bbox": [ + 89, + 703, + 485, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.2. Advanced Attacks", + "text_level": 1, + "bbox": [ + 89, + 848, + 269, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our primary experiments in Section 6, we utilized variations of the PGD attack [48] for both model training and", + "bbox": [ + 89, + 869, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8c54935fba97c20127d235a19f8a7f7253beedc5a1d1ea5e08a01a2e0f401a20.jpg", + "table_caption": [ + "Table 6. An ablation study on our method's performance using various anomaly generation techniques that incorporate masks." + ], + "table_footnote": [], + "table_body": "
Pseudo-Anomaly GeneratorTaskDataset
MVTec ADVisABTADBraTS2021
CutPasteAD79.1 / 60.374.4 / 61.574.8 / 69.280.3 / 72.7
AL83.9 / 60.885.4 / 73.082.6 / 61.291.2 / 82.1
FPIAD78.9 / 62.175.1 / 64.372.9 / 71.084.1 / 74.1
AL84.3 / 63.882.9 / 71.886.3 / 60.593.3 / 84.8
NSAAD81.7 / 66.583.1 / 65.477.1 / 73.486.3 / 73.8
AL85.1 / 65.083.3 / 76.785.4 / 64.989.4 / 86.8
GLASSAD83.2 / 62.084.0 / 68.779.3 / 73.983.1 / 66.1
AL86.7 / 67.086.2 / 78.586.8 / 68.292.2 / 87.3
OursAD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 /85.293.2 /73.097.7 /94.5
", + "bbox": [ + 514, + 130, + 906, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "testing. To further illustrate the adaptability and robustness of our proposed method under diverse adversarial conditions, we also evaluate its performance against a variety of other attack types including Detection attacks and Localization attacks. The Detection attacks consist of black-box attacks [28], FGSM [24], CAA [49], AutoAttack [16], and $\\mathrm{A}^3$ (Adversarial Attack Automation) [43]. The localization attacks are adapted from the well-known segmentation attacks in literature, namely SegPGD [27] and SEA [17]. All results are available in Table 4, and the exact details on our adaptation approach for these attacks are available in Appendix H. It is important to note that the training process remains consistent, employing only the standard PGD-10 for simplicity and practicality.", + "bbox": [ + 511, + 358, + 908, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.3. Pseudo-Anomaly Generation Strategy", + "text_level": 1, + "bbox": [ + 511, + 580, + 841, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since our method targets both AD and AL using pseudo-anomalies, it is crucial to evaluate our pseudo-anomaly generation against compatible methods that produce localization maps. Thus, methods like MIXUP [33], FakeIt [50], and Dream-OOD [20] are excluded. We compare instead with CutPaste [40], NSA [59], GLASS [12], and FPI [67], keeping other pipeline components fixed. As shown in Table 6, our method outperforms these approaches, highlighting its effectiveness for robust anomaly localization.", + "bbox": [ + 511, + 603, + 908, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 753, + 633, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce PatchGuard, an adversarially robust method for AD and AL. It employs Foreground-Aware Pseudo-Anomaly Generation using segmentation masks. Additionally, we propose a regularization term, motivated by observations and theory, that strengthens robustness against adversarial attacks by increasing the attention degree. PatchGuard outperforms previous methods on AD and AL benchmarks under adversarial settings.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20390", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We are very grateful to Mohammad Javad Maheronnaghsh for his collaboration in the initial stages of this project and to the anonymous reviewers for their helpful discussions and valuable feedback.", + "bbox": [ + 89, + 114, + 485, + 171 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 202, + 187, + 217 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] M. Eren Akbiyik. Data augmentation in training cnns: Injecting noise to images. ArXiv, abs/2307.06855, 2019. 4", + "[2] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. *IEEE Access*, 6:14410–14430, 2018. 1", + "[3] Mohammad Azizmalayeri, Arshia Soltani Moakhar, Arman Zarei, Reihaneh Zohrabi, Mohammad Manzuri, and Mohammad Hossein Rohhan. Your out-of-distribution detection method is not robust! Advances in Neural Information Processing Systems, 35:4887-4901, 2022. 1, 2, 4", + "[4] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 2", + "[5] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1893-1902, 2015. 2", + "[6] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 2", + "[7] Louis Béthane, Paul Novello, Thibaut Boissin, Guillaume Coiffier, Mathieu Serrurier, Quentin Vincenot, and Andres Troya-Galvis. Robust one-class classification with signed distance function using 1-lipschitz neural networks. arXiv preprint arXiv:2303.01978, 2023. 1", + "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 4", + "[9] Chengwei Chen, Yuan Xie, Shaohui Lin, Ruizhi Qiao, Jian Zhou, Xin Tan, Yi Zhang, and Lizhuang Ma. Novelty detection via contrastive learning with negative data augmentation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI 2021, Virtual Event / Montreal, Canada, 19-27 August 2021, pages 606-614. ijcai.org, 2021. 4", + "[10] Jiefeng Chen, Yixuan Li, Xi Wu, Yingyu Liang, and Somesh Jha. Robust out-of-distribution detection for neural networks. arXiv preprint arXiv:2003.09711, 2020. 1, 2, 4", + "[11] Jiefeng Chen, Yixuan Li, Xi Wu, Yingyu Liang, and Somesh Jha. Atom: Robustifying out-of-distribution detection using" + ], + "bbox": [ + 93, + 226, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "outlier mining. In Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13-17, 2021, Proceedings, Part III 21, pages 430-445. Springer, 2021. 2, 4", + "bbox": [ + 544, + 92, + 906, + 148 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[12] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. arXiv preprint arXiv:2407.09359, 2024. 1, 3, 8", + "bbox": [ + 516, + 150, + 906, + 203 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4", + "bbox": [ + 517, + 205, + 906, + 260 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 4", + "[15] Matan Jacob Cohen and Shai Avidan. Transformally-two (feature spaces) are better than one. arXiv preprint arXiv:2112.04185, 2021. 2, 4", + "[16] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 8", + "[17] Francesco Croce, Naman D Singh, and Matthias Hein. Towards reliable evaluation and fast training of robust semantic segmentation models. In ECCV, 2024. 2, 3, 8", + "[18] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 4", + "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[20] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. arXiv preprint arXiv:2309.13415, 2023. 8", + "[21] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. 3", + "[22] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1", + "[23] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. arXiv preprint arXiv:2012.07177, 2020. 4", + "[24] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 8", + "[25] Adam Goodge, Bryan Hooi, See Kiong Ng, and Wee Siong Ng. Robustness of autoencoders for anomaly detection under adversarial impact. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 1244-1250, 2021. 1" + ], + "bbox": [ + 93, + 226, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20391", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4", + "[27] Jindong Gu, Hengshuang Zhao, Volker Tresp, and Philip HS Torr. Segpgd: An effective and efficient adversarial attack for evaluating and boosting segmentation robustness. In European Conference on Computer Vision, pages 308-325. Springer, 2022. 8", + "[28] Chuan Guo, Jacob Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Weinberger. Simple black-box adversarial attacks. In International Conference on Machine Learning, pages 2484-2493. PMLR, 2019. 8", + "[29] Jia Guo, Lize Jia, Weihang Zhang, Huiqi Li, et al. Recontrast: Domain-specific anomaly detection via contrastive reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4", + "[31] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4", + "[32] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606, 2018. 2", + "[33] Yann N. Dauphin Hongyi Zhang, Moustapha Cisse. mixup: Beyond empirical risk minimization. International Conference on Learning Representations, 2018. 4, 8", + "[34] Mohammad Hossein Rohban Hossein Mirzaei, ... Scanning trojaned models using out-of-distribution samples. In Advances in Neural Information Processing Systems, 2024. 3", + "[35] Jiande Huang, Xingxing Li, Xin Li, Jiaqi Wu, Keke Zhang, Yongqiang Yuan, and Wei Zhang. Real-time outlier detection of satellite orbit and clock products using reverse error estimation. GPS Solutions, 29(1):2, 2025. 2", + "[36] Nicholas Jeffrey, Qing Tan, and José R Villar. A review of anomaly detection strategies to detect threats to cyberphysical systems. *Electronics*, 12(15):3283, 2023. 1", + "[37] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2, 4", + "[38] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020. 4", + "[39] Jungbeom Lee, Eunj Kim, Jisoo Mok, and Sungroh Yoon. Anti-adversarily manipulated attributions for weakly supervised semantic segmentation and object localization. IEEE transactions on pattern analysis and machine intelligence, 46 (3):1618-1634, 2022. 2" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9664-9674, 2021. 3, 4, 8", + "[41] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. Musc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. arXiv preprint arXiv:2401.16753, 2024. 2, 4", + "[42] Jiaqi Liu, Guoyang Xie, Jinbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. Machine Intelligence Research, 21(1):104-135, 2024. 1", + "[43] Ye Liu, Yaya Cheng, Lianli Gao, Xianglong Liu, Qilong Zhang, and Jingkuan Song. Practical evaluation of adversarial robustness via adaptive auto attack, 2022. 2, 8", + "[44] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 3", + "[45] Shao-Yuan Lo, Poojan Oza, and Vishal M Patel. Adversarily robust one-class novelty detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3", + "[46] Pauline Luc, Camille Couprie, Soumith Chintala, and Jakob Verbeek. Semantic segmentation using adversarial networks. arXiv preprint arXiv:1611.08408, 2016. 2", + "[47] Yuan Luo, Ya Xiao, Long Cheng, Guojun Peng, and Danfeng Yao. Deep learning-based anomaly detection in cyberphysical systems: Progress and opportunities. ACM Computing Surveys (CSUR), 54(5):1-36, 2021. 1", + "[48] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2, 3, 7, 8", + "[49] Xiaofeng Mao, Yuefeng Chen, Shuhui Wang, Hang Su, Yuan He, and Hui Xue. Composite adversarial attacks. ArXiv, abs/2012.05434, 2020. 8", + "[50] Hossein Mirzaei, Mohammadreza Salehi, Sajjad Shahabi, Efstratios Gavves, Cees GM Snoek, Mohammad Sabokrou, and Mohammad Hossein Rohban. Fake it till you make it: Near-distribution novelty detection by score-based generative models. arXiv preprint arXiv:2205.14297, 2022. 8", + "[51] Hossein Mirzaei, Mohammad Jafari, Hamid Reza Dehbashi, Ali Ansari, Sepehr Ghobadi, Masoud Hadi, Arshia Soltani Moakhar, Mohammad Azizmalayeri, Mahdieh Soleymani Baghshah, and Mohammad Hossein Rohban. Rodeo: Robust outlier detection via exposing adaptive out-of-distribution samples. In Forty-first International Conference on Machine Learning, 2024. 1, 2, 3, 4", + "[52] Hossein Mirzaei, Mohammad Jafari, Hamid Reza Dehbashi, Zeinab Sadat Taghavi, Mohammad Sabokrou, and Mohammad Hossein Rohban. Killing it with zero-shot: Adversarially robust novelty detection. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7415-7419. IEEE, 2024. 3", + "[53] Hossein Mirzaei, Mojtaba Nafez, Mohammad Jafari, Mohammad Bagher Soltani, Mohammad Azizmalayeri, Jafar Habibi," + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20392", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mohammad Sabokrou, and Mohammad Hossein Rohban. Universal novelty detection through adaptive contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22914-22923, 2024. 4", + "[54] Arian Mousakhan, Thomas Brox, and Jawad Tayyyub. Anomaly detection with conditioned denoising diffusion models. arXiv preprint arXiv:2305.15956, 2023. 1", + "[55] Fahimeh Hosseini Noohdani, Parsa Hosseini, Aryan Yazdan Parast, Hamidreza Yaghoubi Araghi, and Mahdieh Soleymani Baghshah. Decompose-and-compose: A compositional approach to mitigating spurious correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27662-27671, 2024. 2, 4", + "[56] Pramuditha Perera, Poojan Oza, and Vishal M Patel. One-class classification: A survey. arXiv preprint arXiv:2101.03064, 2021. 2", + "[57] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection, 2021. 2", + "[58] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051, 2021. 2", + "[59] Hannah M Schlüter, Jeremy Tan, Benjamin Hou, and Bernhard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In European Conference on Computer Vision, pages 474-489. Springer, 2022. 3, 8", + "[60] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 4", + "[61] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16, pages 682-698. Springer, 2020. 1", + "[62] Abhishek Sinha, Kumar Ayush, Jiaming Song, Burak Uzkent, Hongxia Jin, and Stefano Ermon. Negative data augmentation. In International Conference on Learning Representations, 2021. 4", + "[63] Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister. Learning and evaluating representations for deep one-class classification. arXiv preprint arXiv:2011.02578, 2020. 4", + "[64] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Generalad: Anomaly detection across domains by attending to distorted features. arXiv preprint arXiv:2407.12427, 2024. 1, 2, 3, 4", + "[65] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1", + "[66] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distribu" + ], + "bbox": [ + 91, + 90, + 485, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tionally shifted instances. Advances in neural information processing systems, 33:11839-11852, 2020. 1, 4", + "[67] Jeremy Tan, Benjamin Hou, James Batten, Huaqi Qiu, and Bernhard Kainz. Detecting outliers with foreign patch interpolation. Machine Learning for Biomedical Imaging, 1(April 2022):1-27, 2022. 8", + "[68] Maximilian E Tschuchnig and Michael Gadermayr. Anomaly detection in medical imaging-a mini review. In Data Science-Analysics and Applications: Proceedings of the 4th International Data Science Conference-iDSC2021, pages 33-38. Springer, 2022. 1", + "[69] Qi Wang, Junyu Gao, and Xuelong Li. Weakly supervised adversarial domain adaptation for semantic segmentation in urban scenes. IEEE Transactions on Image Processing, 28 (9):4376-4386, 2019. 2", + "[70] Weiping Wang, Zhaorong Wang, Zhanfan Zhou, Haixia Deng, Weiliang Zhao, Chunyang Wang, and Yongzhen Guo. Anomaly detection of industrial control systems based on transfer learning. Tsinghua Science and Technology, 26(6): 821-832, 2021. 1", + "[71] Yue Xing, Qifan Song, and Guang Cheng. Why do artificially generated data help adversarial robustness. Advances in Neural Information Processing Systems, 35:954-966, 2022. 2", + "[72] Han Xu, Yao Ma, Haochen Liu, Debayan Deb, Hui Liu, Jiliang Tang, and Anil K. Jain. Adversarial attacks and defenses in images, graphs and text: A review, 2019. 3", + "[73] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, pages 1-28, 2024. 1", + "[74] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. arXiv preprint arXiv:2406.07487, 2024. 1", + "[75] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 3", + "[76] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE transactions on neural networks and learning systems, 30(9): 2805-2824, 2019. 1, 3", + "[77] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features, 2019. 4", + "[78] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 3", + "[79] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Dsr – a dual subspace re-projection network for surface anomaly detection, 2022. 3", + "[80] Chaoning Zhang, Philipp Benz, Chenguo Lin, Adil Karjauv, Jing Wu, and In So Kweon. A survey on universal adversarial attack. arXiv preprint arXiv:2103.01498, 2021. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20393", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Hui Zhang, Zheng Wang, Zuxuan Wu, and Yu-Gang Jiang. Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection. arXiv preprint arXiv:2303.08730, 2023. 1", + "[82] Lin Zhang and Yang Dai. A gan anomaly detection method based on multi-scale endogenous enhancement. In Blockchain and Web3 Technology Innovation and Application Exchange Conference, pages 269-281. Springer, 2024. 2", + "[83] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 1, 2" + ], + "bbox": [ + 91, + 90, + 485, + 273 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20394", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/6310723f-d72e-495d-87c7-6b59d06d7ca9_model.json b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/6310723f-d72e-495d-87c7-6b59d06d7ca9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a4c05425a76fefd51979d4b2ee80fe1c8386a6ff --- /dev/null +++ b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/6310723f-d72e-495d-87c7-6b59d06d7ca9_model.json @@ -0,0 +1,2402 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.908, + 0.175 + ], + "angle": 0, + "content": "PatchGuard: Adversarily Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.203, + 0.756, + 0.258 + ], + "angle": 0, + "content": "Mojtaba Nafez1 Amirhossein Koochakian1 * Arad Maleki1 * Jafar Habibi1 Mohammad Hossein Rohban1 1Sharif University of Technology, Iran" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.261, + 0.768, + 0.275 + ], + "angle": 0, + "content": "{mojtaba.nafez77, amir.kochakian12, arad.maleki02, rohban}@sharif.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.486, + 0.735 + ], + "angle": 0, + "content": "Anomaly Detection (AD) and Anomaly Localization (AL) are crucial in fields that demand high reliability, such as medical imaging and industrial monitoring. However, current AD and AL approaches are often susceptible to adversarial attacks due to limitations in training data, which typically include only normal, unlabeled samples. This study introduces PatchGuard, an adversarily robust AD and AL method that incorporates pseudo anomalies with localization masks within a Vision Transformer (ViT)-based architecture to address these vulnerabilities. We begin by examining the essential properties of pseudo anomalies, and follow it by providing theoretical insights into the attention mechanisms required to enhance the adversarial robustness of AD and AL systems. We then present our approach, which leverages Foreground-Aware Pseudo-Anomalies to overcome the deficiencies of previous anomaly-aware methods. Our method incorporates these crafted pseudo-anomaly samples into a ViT-based framework, with adversarial training guided by a novel loss function designed to improve model robustness, as supported by our theoretical analysis. Experimental results on well-established industrial and medical datasets demonstrate that PatchGuard significantly outperforms previous methods in adversarial settings, achieving performance gains of \\(53.2\\%\\) in AD and \\(68.5\\%\\) in AL, while also maintaining competitive accuracy in non-adversarial settings. The code repository is available at here." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.485, + 0.88 + ], + "angle": 0, + "content": "In recent years, Anomaly Detection (AD) and Anomaly Localization (AL) have become critical techniques across various fields, including medical imaging, fraud detection, and industrial monitoring [22, 36, 42, 47, 68, 70, 73]. While state-of-the-art (SOTA) AD and AL methods exhibit near-perfect performance on standard benchmarks [6, 83], they" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.307, + 0.907, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.445, + 0.909, + 0.53 + ], + "angle": 0, + "content": "Figure 1. Impact of Adversarial Attacks on Anomaly Localization Methods: Localization maps for multiple methods are shown before and after a PGD-1000 attack, illustrating the vulnerability of existing methods to adversarial attacks, even when performing perfectly in clean conditions. Our proposed method demonstrates enhanced robustness in these adversarial scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.909, + 0.784 + ], + "angle": 0, + "content": "frequently face significant performance degradation when subjected to adversarial attacks, as illustrated in Figure 1 [12, 54, 64, 66, 74, 81]. These attacks involve introducing subtle, imperceptible perturbations into the input data, misleading models to misclassify normal1 samples as anomalies or to overlook actual anomalies [2, 24, 48, 65, 80]. To address this vulnerability, a variety of strategies have been proposed to strengthen the adversarial robustness of AD systems [3, 7, 10, 25, 45, 61]. However, despite these advancements, the resilience of these methods remains limited, with even mild adversarial attacks still causing noticeable performance drops, even on widely used anomaly benchmarks [6, 83]. Furthermore, adversarial robustness in AL remains an unexplored area, highlighting a critical gap in the current research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.786, + 0.909, + 0.862 + ], + "angle": 0, + "content": "Adversarial training, a strategy designed to improve model robustness in classification tasks, incorporates adversarial samples into the training process [48, 76]. Despite its success in standard classification, this approach proves less effective in AD and AL scenarios [51]. The primary" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.908, + 0.9 + ], + "angle": 0, + "content": "1In this study, the term 'normal' refers to inlier samples, not to be mistaken with Gaussian distribution." + }, + { + "type": "footer", + "bbox": [ + 0.11, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20383" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.334 + ], + "angle": 0, + "content": "obstacle to achieving adversarial robustness in AD lies in the absence of anomalous samples during training [10, 11]. Without such samples, models cannot be trained on adversarial perturbations specific to anomalous data [5, 51, 56, 58], leaving them vulnerable to attacks on perturbed anomalies [3, 10, 11]. To address this issue in the AD setup, recent efforts have combined Outlier Exposure (OE) [32] with adversarial training [3, 10, 11]. However, this method struggles when applied to high-resolution, real-world datasets [51]. This challenge is even more pronounced in AL settings, as robust AL methods must perform pixel-level classification, which inherently requires pixel-level supervision, while at the same time operating within the same data constraints as the AD framework [35, 82]. Despite ongoing research, the adversarial robustness of AL methods remains an open problem, with no comprehensive solutions proposed to date." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.486, + 0.548 + ], + "angle": 0, + "content": "In this study, we identify two essential properties that OE techniques must satisfy for effective application in adversarially robust AD and AL. First, the outlier exposures must include localization maps to facilitate the training of robust AL models. This requirement aligns with prior research on adversarially robust segmentation models, which underscores the importance of generating localization maps to enhance model robustness [39, 46, 69]. Second, as noted by [51], OE samples should be \"near-distributed\" relative to normal data, meaning they should share similar semantic and stylistic attributes. This condition is consistent with the literature on adversarial robustness, which highlights the advantages of near-distribution decision boundaries samples for optimizing the model robustness [71]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.486, + 0.746 + ], + "angle": 0, + "content": "Moreover, the recent success of Vision Transformer (ViT)-based architectures in clean AD and AL tasks [15, 37, 41, 64], attributed to their global contextual modeling [19], motivated us to explore their potential in adversarial settings. For an input image \\( x \\) and a fixed attention layer \\( \\ell \\), we define the Attention Degree of \\( \\ell \\) under \\( x \\) as the average number of input embeddings each output embedding attends to. As shown in Figure 2, we observe that images that induce higher attention degrees in the model are more resistant to adversarial attacks. Furthermore, adversarial training appears to implicitly steer the model toward achieving a high attention degree. We provide a theoretical analysis and justification for these findings in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Building on these insights, we propose PatchGuard, an adversarially robust AD and AL method that addresses limitations in current OE-based methods and leverages attention-based insights. To address OE challenges, we generate pseudo-anomaly samples from normal data, ensuring they meet two essential criteria: availability of localization maps and near-distribution alignment with inliers. To create samples that exhibit anomalous characteristics while remaining close to the inlier distribution, we introduce Foreground-Aware Pseudo-Anomaly Generation. This approach iden" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.227 + ], + "angle": 0, + "content": "tifies causal regions in normal samples using an enhanced Grad-CAM [60] feature attribution technique with tailored augmentations. These identified regions are then selectively distorted, drawing on insights from [55], to produce anomaly samples that retain stylistic and semantic coherence with the original data. By precisely controlling the distortion location, we can generate accurate localization maps for the pseudoanomalies, effectively meeting both criteria for robust AD and AL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.909, + 0.333 + ], + "angle": 0, + "content": "In the next step, we leverage normal samples and pseudoanomalies to perform adversarial training on a ViT-based architecture. Building on the provided insights, we introduce a novel loss function designed to increase the model's last-layer attention degree. As our theoretical justifications suggest, this property significantly enhances the model's adversarial robustness in both AD and AL settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.909, + 0.651 + ], + "angle": 0, + "content": "Contribution. In this paper, we propose PatchGuard, an adversarially robust AD and AL method. To the best of our knowledge, we are the first to raise and effectively address the issue of adversarial robustness in AL. We provide insights into the shortcomings of previous OE-based methods and introduce the Foreground-Aware Pseudo-Anomaly Generation framework to address these limitations. Additionally, we provide insights and theoretical justifications on enhancing the adversarial robustness of ViT-based architectures, and propose a novel loss function aligned with these findings. We evaluate PatchGuard in both clean and adversarial settings, rigorously testing its robustness under several strong attacks, including PGD-1000 [48], \\(\\mathrm{A}^3\\) [43], and SEA [17]. Results demonstrate that PatchGuard significantly outperforms existing methods in adversarial contexts, achieving an AUROC improvement of \\(53.2\\%\\) in AD and \\(68.5\\%\\) in AL, while maintaining competitive performance in standard conditions. Our evaluations span various well-established industrial and medical datasets, including MVTec [6], VisA [83], and BraTS2021 [4], highlighting PatchGuard's applicability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.665, + 0.657, + 0.681 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "From an adversarial robustness perspective, anomaly detection and localization methods can be mainly classified into three categories: 1) methods with anomaly-free training, 2) methods that leverage anomalies at the embedding level, and 3) methods that leverage anomalies at the input level. We experimentally adapted methods from each category to adversarial training processes, and the results reveal they still exhibit significant vulnerabilities even after adversarial training. The first category consists of methods such as Recontrast[29], Transformaly [15], and PatchCore [57], that operate on the hypothesis that models trained exclusively on normal samples can reconstruct normal images well but fail to do so with abnormal regions. These methods lack robustness, since the model does not encounter any anomaly" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20384" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.501 + ], + "angle": 0, + "content": "data during adversarial training, meaning that when adversarial perturbations are added to anomaly data, the model still remains vulnerable. The second category, which includes methods such as SimpleNet [44], GeneralAD [64], UniAD [75], and DSR [79], utilize a feature extractor during training to obtain feature vectors of the normal set. By perturbing these features, they create anomaly features and primarily train a discriminator over these normal and anomaly features. However, these methods lack robustness due to their reliance on highly vulnerable feature extractors and the absence of anomaly data in the input space, which prevents proper training of a robust feature extractor. The third category consists of DRAEM [78], NSA [59], Cutpaste [40], and GLASS [12], generate pseudo-anomaly data during training. Nevertheless, some of these methods lack access to masks or do not use a foreground-aware generation process. Despite their astonishing success in clean settings, we empirically show in Table 3 that even when adversarily trained, they exhibit unsatisfactory results. Furthermore, while anomaly localization has not been explored in an adversarial setup, some other works have pursued anomaly detection in an adversarial setting, including PrincipaLS [45], ZARND [52], and Rodeo [51]. These approaches have achieved relatively better results by incorporating pseudo-anomaly generation processes and adversarial training. However, their performance falls short in real-world high-resolution datasets (as presented in Table 1). For more details about previous works, see Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.513, + 0.23, + 0.529 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.485, + 0.689 + ], + "angle": 0, + "content": "Anomaly Detection (AD) and Localization (AL): AD involves identifying instances that deviate from expected patterns, while AL focuses on localizing abnormal subregions within an image. In AD, a model assigns an anomaly score \\( S(x; f) \\) to each sample \\( x \\) using model \\( f \\); if this score exceeds a predefined threshold, the sample is classified as anomalous. Conversely, lower scores indicate normal samples. In AL, a model generates an anomaly map \\( M(x; f) \\) using model \\( f \\), where each pixel of the map corresponds to an anomaly score for the respective pixel in the input image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Adversarial Robustness of Anomaly Detectors: An adversarial attack, primarily studied within classification tasks, involves the malicious alteration of a data sample \\( x \\) with its label \\( y \\) into a new sample \\( x^{*} \\) that maximizes the loss function \\( \\ell(x^{*};y) \\) [34, 72, 76]. This new sample \\( x^{*} \\) is known as an adversarial example of \\( x \\), and the difference \\( (x^{*} - x) \\) is referred to as adversarial perturbation. To prevent significant semantic changes, the \\( l_{p} \\) norm of the adversarial noise is restricted by an upper limit \\( \\epsilon \\). Specifically, an adversarial example \\( x^{*} \\) must satisfy the condition \\( x^{*} = \\arg \\max_{x^{\\prime}:\\|x - x^{\\prime}\\|_{p} \\leq \\epsilon} \\ell(x^{\\prime};y) \\). A widely used and effective attack method is the Projected Gradient Descent (PGD) [48] technique, which iteratively increases the loss function by moving in the direction of the gradient sign of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.212 + ], + "angle": 0, + "content": "\\(\\ell(x^{*};y)\\) with a specified step size \\(\\alpha\\). When adapting adversarial attacks for AD, the objective shifts from maximizing the loss value to increasing \\(S(x,f)\\) for normal samples and decreasing it for anomalies. The attack is formulated as \\(x_0^* = x\\), \\(x_{t+1}^* = x_t^* + y \\cdot \\alpha \\cdot \\mathrm{sign}(\\nabla_x S(x_t^*;f_\\theta))\\), and \\(x^* = x_k^*\\), where \\(y = +1\\) for normal samples and \\(y = -1\\) for anomaly samples. This approach is consistently applied to other attacks in the study [21, 76]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.909, + 0.486 + ], + "angle": 0, + "content": "Adversarial Robustness of Anomaly Localization: Adversarial attacks aim to introduce perturbations that uniformly modify pixel scores in an anomaly map \\( M \\), flipping scores from 0 to 1 or vice versa. In the context of AL, the attack objective transitions from merely maximizing a loss function to specifically increasing the values of \\( M(x, f) \\) for normal samples while decreasing them for anomalies. The attack is formalized as follows: starting with \\( x_0^* = x \\), the adversarial example at each step is updated as \\( x_{t+1}^* = x_t^* + Y \\cdot \\alpha \\cdot \\mathrm{sign}(\\nabla_x M(x_t^*; f_\\theta)) \\), and the final adversarial example is \\( x^* = x_k^* \\). Here, \\( Y_{i,j} = +1 \\) for normal pixels and \\( Y_{i,j} = -1 \\) for anomaly pixels, where \\( i, j \\) denote pixel coordinates. This strategy is extended to other advanced attack methods throughout the study [17]. Despite its significance, adversarial robustness in anomaly localization remains underexplored. Enhancing robustness is particularly critical for applications requiring spatial precision, such as medical diagnostics and industrial monitoring systems." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.501, + 0.707, + 0.52 + ], + "angle": 0, + "content": "4. Theoretical Insights" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.543, + 0.707, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.542, + 0.904, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.671, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Figure 2. The figure demonstrates how increasing the average last-layer attention degree in the ViT-base architecture reduces vulnerability to adversarial attacks. Specifically, the BraTS dataset images were clustered into five groups based on their attention degree values (x-axis). The y-axis represents Vulnerability, measured as the absolute difference in the model's localization performance (AUROC%) for each cluster before and after a PGD attack. (a) illustrates the decreasing trend in a clean-trained model. (b) shows a similar decreasing trend in the adversarially trained model, where the attention degree is relatively higher than in the clean model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Here, we want to show why a self-attention layer, in which each token attends to a higher number of other tokens is more adversarially robust compared to the case that the attention is distributed over a smaller number of tokens." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20385" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.19 + ], + "angle": 0, + "content": "Let's consider the self-attention output embedding of the \\(i\\)th token in the \\(l\\)-th layer as \\(x_{i}^{(l)}\\), and such embedding after the perturbation is applied to the input as \\(\\hat{x}_i^{(l)}\\). We are seeking to set \\(x_{i}^{(2)} - \\hat{x}_{i}^{(2)} = r_{i}\\), where \\(r_i\\) is the desired adversarial direction that most influences the layers next to the first layer of the transformer. According to the self-attention:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.201, + 0.416, + 0.221 + ], + "angle": 0, + "content": "\\[\nx _ {i} ^ {(l + 1)} = V ^ {(l)}. \\text {s o f t m a x} \\left(q _ {i} ^ {(l) \\top} K / \\sqrt {d}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.224, + 0.484, + 0.388 + ], + "angle": 0, + "content": "where \\(q_{i}^{(l)} = W_{q}x_{i}^{(l)},v_{j}^{(l)}\\) the \\(j\\)-th column of \\(V^{(l)}\\) be \\(v_{j}^{(l)} = W_{v}x_{j}^{(l)},k_{j}^{(l)}\\) the \\(j\\)-th column of \\(\\mathbf{K}\\) be \\(k_{j}^{(l)} = W_{k}x_{j}^{(l)}\\), and \\(d\\) be the embedding dimension. For simplicity, we assume that the dimensions of embeddings of key, query, and values are all the same and are equal to the dimension of the original embeddings. Now, \\(W_{v}(X^{(1)}S - \\hat{X}^{(1)}\\hat{S}) = R\\), where the columns of \\(R\\) and \\(X^{(l)}\\) are represented by \\(r_i\\) and \\(x_{i}^{(l)}\\), and the columns in \\(S\\) and \\(\\hat{S}\\) correspond to the softmax values of each query token calculated for \\(X\\) and the perturbed inputs \\(\\hat{X}\\), respectively. Therefore," + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.394, + 0.308, + 0.412 + ], + "angle": 0, + "content": "\\[\nW _ {v} (X ^ {(1)} S - \\hat {X} ^ {(1)} S) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.415, + 0.433, + 0.434 + ], + "angle": 0, + "content": "\\[\nW _ {v} \\left(X ^ {(1)} S - \\hat {X} ^ {(1)} \\hat {S} + \\hat {X} ^ {(1)} \\hat {S} - \\hat {X} ^ {(1)} S\\right) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.437, + 0.427, + 0.455 + ], + "angle": 0, + "content": "\\[\nR + W _ {v} \\hat {X} ^ {(1)} (\\hat {S} - S) = R + W _ {v} \\hat {X} ^ {(1)} \\Delta S.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.351, + 0.481 + ], + "angle": 0, + "content": "where \\(\\Delta S\\coloneqq \\hat{S} -S\\) . Hence, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.487, + 0.434, + 0.523 + ], + "angle": 0, + "content": "\\[\nX ^ {(1)} - \\hat {X} ^ {(1)} = \\underbrace {W _ {v} ^ {+} (R + W _ {v} \\hat {X} ^ {(1)} \\Delta S)} _ {U} S ^ {+}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.485, + 0.833 + ], + "angle": 0, + "content": "Where \\( W_{v}^{+} \\) and \\( S^{+} \\) are the pseudo-inverses of the matrices \\( W_{v} \\) and \\( S \\), respectively. Note that for the \\( \\delta := X^{(1)} - \\hat{X}^{(1)} \\) perturbation in the input, in order to make the required perturbation in the output of the network, we get \\( \\| \\delta \\| \\leq \\lambda_{max} \\| U \\| \\), where \\( \\lambda_{max} \\) is the maximum eigenvalue of \\( S^{+} \\). But we note that for a stochastic matrix \\( S \\), the maximum eigenvalue is 1. Therefore, for the inverse matrix, \\( S^{+} \\), the maximum eigenvalue would be larger than 1. As a result, we would expect the \\( \\| \\delta \\| \\) to grow proportional to \\( \\lambda_{max} \\geq 1 \\). We note that the smaller \\( \\| \\delta \\| \\) would imply more vulnerability as it becomes easier to attack the network with a smaller perturbation length. On the other hand, for a purely localized \\( S \\), where each token only attends to itself \\( \\lambda_{max} = 1 \\), showing that the increase in \\( \\| \\delta \\| \\) would likely be small compared to the general case, leading to more vulnerability. Therefore, we conclude that a localized and overly sparse attention map would typically imply an easier attack on the network and hence higher vulnerability against adversarial attacks. Sparse attention values may also resemble what happens in convolutional networks that have small receptive fields." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.226, + 0.862 + ], + "angle": 0, + "content": "5. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Overview. Current AD and AL methods perform poorly when subjected to adversarial attacks. Previous studies have" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.243 + ], + "angle": 0, + "content": "demonstrated the effectiveness of OE in enhancing robustness against such attacks [3, 10, 11, 51]. Additionally, ViTs have shown promising results in standard (clean) AD and AL tasks [15, 37, 41, 64], and our findings suggest their potential for adversarial settings with targeted modifications of attention degrees. Building on these insights and our theoretical foundations, we propose PatchGuard, a novel method designed to enhance adversarial robustness in AD and AL. The following subsections detail the key components of PatchGuard." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.252, + 0.907, + 0.281 + ], + "angle": 0, + "content": "5.1. Foreground-Aware Pseudo-Anomaly Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.909, + 0.44 + ], + "angle": 0, + "content": "To generate anomalous data from normal samples, it is essential to apply targeted distortions that ensure the image becomes anomalous [8, 13, 14, 26]. Additionally, the generated pseudo-anomalies must satisfy two key requirements: (1) inclusion of localization maps to enhance the AL task, and (2) close alignment with the normal sample distribution [51, 53]. Meeting these requirements simultaneously necessitates precise but minimal alterations within the foreground. Thus, the proposed method must reliably identify a region of the image that we can confidently attribute to the foreground." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.909, + 0.651 + ], + "angle": 0, + "content": "To achieve this, we leverage insights from [55], which indicate that pretrained models benefit from style-agnostic processing to improve classification accuracy. To this end, we first introduce soft transformations, which are commonly used in self-supervised learning studies due to their ability to preserve semantic content [8, 13, 14, 26, 31]. Next, we apply Grad-CAM [60] to generate a saliency map using a standard pre-trained ResNet18 [30] model on a given input. Specifically, for a normal sample \\( x \\), we randomly select \\( k_{\\mathrm{soft}} \\) soft transformations \\( t_1^s,\\ldots ,t_{k_{\\mathrm{soft}}}^s \\). We then compute saliency maps for \\( x,t_1^s (x),\\ldots ,t_{k_{\\mathrm{soft}}}^s (x) \\) and take their element-wise product to produce a style-invariant saliency map, \\( G(x) \\). As shown in Figure 3 (part ①), \\( G(x) \\) effectively assigns higher values to the foreground region." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.911, + 0.849 + ], + "angle": 0, + "content": "To distort the obtained region, we first randomly sample an anchor point from it. We then sample \\( k_{\\mathrm{hard}} \\) hard transformations \\( t_1^h, \\ldots, t_{k_{\\mathrm{hard}}}^h \\), which are typically associated with altering semantic integrity, as highlighted in prior research [1, 9, 18, 23, 33, 38, 40, 62, 63, 66, 77]. We then design a rectangular mask centered on this chosen point, with its width \\( w \\) and height \\( h \\) satisfying \\( w \\sim U(0.05W, 0.3W) \\) and \\( h \\sim U(0.05H, 0.3H) \\), where \\( W \\) and \\( H \\) are the image's width and height. Next, we rotate the mask by an angle of \\( \\theta \\sim U(-45^\\circ, 45^\\circ) \\). Finally, we sequentially apply the hard transformations only to the region within this mask, leaving the remaining unmasked areas of the image unchanged. Mathematically, the final anomaly is obtained by applying" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.857, + 0.908, + 0.877 + ], + "angle": 0, + "content": "\\[\nx ^ {\\prime} = t _ {1} ^ {h} \\left(\\dots \\left(t _ {k _ {\\text {h a r d}}} ^ {h} \\left(x _ {\\text {m a s k e d}}\\right) \\dots\\right) + \\left(x - x _ {\\text {m a s k e d}}\\right). \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.909, + 0.902 + ], + "angle": 0, + "content": "An additional advantage of this pseudo-anomaly crafting ap-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20386" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.46, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.306, + 0.441, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.09, + 0.898, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Figure 3. Overview of the PatchGuard framework. (1) We use Grad-CAM to identify regions likely to belong to the foreground in an image. By applying \\( k_{\\mathrm{soft}} = 3 \\) augmentations to the image \\( x \\), we generate a combined saliency map \\( G(x) \\). (2) The input \\( x \\) and its saliency map \\( G(x) \\) are then passed to the anomaly generator to produce an anomaly sample \\( x' \\) along with its ground-truth mask \\( x_m' \\). (3) For each normal sample \\( x_i \\), the input batch includes the image, its corresponding anomaly version \\( x_i' \\), and their adversarially attacked variants, along with their ground-truth masks. Each batch sample is processed through a Vision Transformer (ViT), where the Attention Discriminator assigns an anomaly score \\( p_i \\) to each patch embedding. The final anomaly map is constructed from these scores and trained with a novel loss \\( \\mathcal{L} \\) to replicate the ground truth." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.592, + 0.909, + 0.625 + ], + "angle": 0, + "content": "Table 1. Detection performance of various AD methods, including PatchGuard, across benchmark datasets. Results are shown as “Clean / Adversarial”, representing performance on “Clean” and “Adversarial” data (PGD-1000, \\(\\epsilon = \\frac{8}{255}\\), measured by AUROC %), respectively." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.633, + 0.907, + 0.856 + ], + "angle": 0, + "content": "
DatasetMethod
DRAEMPatchCoreSimpleNetReContrastGeneralADGLADGLASSPrincipaLS*ZARND*RODEO*Ours
MVTec AD98.0 / 5.199.1 / 3.999.6 / 2.399.5 / 0.099.2 / 13.799.3 / 11.899.9 / 6.763.8 / 10.671.6 / 11.061.5 / 4.988.1 / 71.1
VisA88.7 / 3.895.1 / 1.796.2 / 0.297.5 / 6.795.9 / 14.599.5 / 8.698.8 / 7.258.9 / 9.367.7 / 9.363.1 / 3.888.5 / 74.3
BTAD87.1 / 4.996.9 / 2.495.1 / 4.695.7 / 2.197.0 / 10.998.5 / 7.697.4 / 8.360.1 / 7.869.7 / 8.360.0 / 5.885.3 / 82.1
MPDD94.3 / 2.891.3 / 6.896.6 / 7.094.5 / 4.598.0 / 12.897.5 / 9.799.6 / 9.171.9 / 15.765.7 / 12.268.8 / 7.485.4 / 68.6
WFDD94.0 / 7.196.3 / 3.998.8 / 5.197.6 / 3.399.0 / 10.799.1 / 12.5100 / 6.066.3 / 14.176.1 / 16.371.2 / 7.884.2 / 65.0
DTD-Synthetic97.1 / 3.665.8 / 4.096.8 / 2.498.2 / 9.197.9 / 16.398.6 / 10.399.1 / 6.177.9 / 9.370.3 / 13.772.8 / 12.091.5 / 60.5
Head-CT72.1 / 8.192.4 / 5.184.9 / 4.886.1 / 6.389.1 / 9.193.1 / 14.294.7 / 8.773.2 / 16.064.3 / 9.785.7 / 61.097.3 / 90.4
BraTS202162.3 / 1.491.6 / 6.382.5 / 6.182.4 / 3.188.3 / 13.092.3 / 12.095.7 / 7.275.3 / 8.766.7 / 14.089.1 / 64.794.3 / 81.0
Average86.7 / 4.691.0 / 4.293.8 / 4.093.9 / 4.395.5 / 12.697.2 / 10.898.1 / 7.468.4 / 11.469.0 / 11.871.5 / 20.989.3 / 74.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.096, + 0.857, + 0.469, + 0.87 + ], + "angle": 0, + "content": "*These works incorporated adversarial training into their proposed AD methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20387" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.121 + ], + "angle": 0, + "content": "Table 2. Localization performance of various AL methods, including PatchGuard, across benchmark datasets. Results are shown as \"Clean / Adversarial\", representing performance on \"Clean\" and \"Adversarial\" data (PGD-1000, \\( \\epsilon = \\frac{8}{255} \\), measured by AUROC %), respectively." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.131, + 0.907, + 0.359 + ], + "angle": 0, + "content": "
DatasetMethod
DRAEMPatchCoreSimpleNetReContrastGeneralADGLADGLASSMuScReConPatchRealNetOurs
MVTec AD97.3 / 3.998.1 / 4.298.1 / 6.898.4 / 4.897.7 / 12.398.7 / 15.399.3 / 8.197.3 / 4.699.2 / 2.899.0 / 8.792.7 / 73.8
VisA90.7 / 5.698.5 / 3.097.4 / 4.898.2 / 7.199.0 / 7.698.6 / 13.498.8 / 7.098.8 / 3.899.2 / 1.798.8 / 6.296.9 / 85.2
BTAD89.0 / 7.192.6 / 6.395.2 / 3.796.8 / 6.496.7 / 9.398.2 / 14.597.0 / 5.797.3 / 5.397.5 / 3.097.9 / 9.293.2 / 73.0
MPDD90.7 / 4.798.5 / 1.997.4 / 7.394.6 / 3.698.1 / 14.998.7 / 10.899.4 / 6.395.5 / 6.096.7 / 2.698.2 / 7.593.8 / 86.6
WFDD95.8 / 3.998.1 / 2.198.0 / 3.497.3 / 5.198.7 / 13.597.6 / 11.398.9 / 7.697.0 / 4.198.3 / 4.998.4 / 6.894.6 / 71.6
DTD-Synthetic96.1 / 4.895.6 / 6.797.0 / 2.996.8 / 7.697.9 / 10.897.3 / 13.898.1 / 7.097.8 / 5.796.3 / 3.998.0 / 8.395.9 / 74.8
Head-CT86.4 / 8.987.3 / 6.189.7 / 0.890.1 / 8.691.7 / 18.393.4 / 16.792.2 / 9.185.2 / 8.589.7 / 6.694.8 / 7.393.6 / 89.3
BraTS202182.2 / 6.196.9 / 4.594.7 / 6.795.3 / 11.196.8 / 14.795.1 / 13.594.0 / 6.367.5 / 6.397.5 / 5.996.9 / 10.297.7 / 94.5
Average91.0 / 5.695.7 / 4.395.9 / 4.595.9 / 6.797.0 / 12.697.2 / 13.697.2 / 7.192.0 / 5.596.8 / 3.997.7 / 8.094.8 / 81.1
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.486, + 0.571 + ], + "angle": 0, + "content": "proach is that the location of the distorted region is precisely known, allowing us to generate an accurate localization map \\( x_{m}^{\\prime} \\) for AL training, where each pixel's value equals 1 if it's anomalous, and 0 otherwise. By creating an anomalous counterpart \\( x^{\\prime} \\) and its corresponding localization map \\( x_{m}^{\\prime} \\) for each normal sample \\( x \\) in the training set, we provide the model with both the anomaly and its location, as shown in Figure 3 (part ②). This pairing equips the model with the information needed to improve its localization accuracy, as we will discuss in Section 5.2. For details on the soft and hard transforms, and ablations on the hypers used in this approach, refer to Table 6 and Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.608 + ], + "angle": 0, + "content": "5.2. Anomaly Detection and Localization with Vision Transformer" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.486, + 0.765 + ], + "angle": 0, + "content": "To enhance adversarial robustness through the effective use of pseudo-anomaly samples and their corresponding localization maps, we employ a Vision Transformer (ViT) model. This model generates output embeddings that are processed by an attention-based discriminator network, which uses a multi-head attention (MLH) mechanism followed by a shared-weight multi-layer perceptron (MLP). The discriminator independently assigns an anomaly score to each patch in the input, providing a patch-level anomaly prediction as shown in Figure 3 (part ③)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "The model is trained using a combination of patch-level cross-entropy (CE) and an additional regularization term \\( R \\), designed to encourage a higher degree of attention in a specified ViT layer. As illustrated in Figure 2 and the theory in Section 4, this regularizer is expected to further enhance adversarial robustness. The cross-entropy loss component aligns the predicted anomaly scores with ground truth labels from the localization map, \\( x_{m} \\), where a patch label equals 1 if more than \\( 5\\% \\) of the pixels in the corresponding patch in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.388, + 0.757, + 0.404 + ], + "angle": 0, + "content": "\\(x_{m}\\) are anomalous, and 0 otherwise:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.416, + 0.907, + 0.433 + ], + "angle": 0, + "content": "\\[\nM = \\operatorname {M L P} (\\operatorname {M L H} (x)) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.436, + 0.907, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} (x, x _ {m}) = \\sum_ {i, j \\in x _ {\\text {p a t c h e s}}} \\mathrm {C E} \\left(M _ {(i, j)}, x _ {m, (i, j)}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.909, + 0.618 + ], + "angle": 0, + "content": "Here, \\(M\\) denotes the predicted AL map, consisting of anomaly scores for each patch, \\(x\\) is any input sample (normal or anomaly), and \\(i,j\\) are the indices for individual patches. The cross-entropy loss term \\(\\mathcal{L}_{\\mathrm{CE}}\\) ensures alignment between the patch \\(M_{(i,j)}\\) and the ground truth anomaly map \\(x_{m,(i,j)}\\), promoting precise anomaly localization. The final AL map is obtained by upsampling \\(M\\) to the original image size, and we use the average of the top-\\(k\\) patch-level anomaly scores as the image-level anomaly score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.909, + 0.679 + ], + "angle": 0, + "content": "To further improve the robustness of the model, we introduce a regularization term \\( R \\), which encourages an increased attention degree in a specified layer \\( \\ell \\) of the ViT, chosen here as the final layer. This regularization term is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.69, + 0.907, + 0.725 + ], + "angle": 0, + "content": "\\[\nR (x) = \\frac {1}{\\sum_ {i} \\sum_ {j} \\sum_ {k} A _ {i j k} \\cdot \\mathbb {1} \\left(A _ {i j k} \\leq \\delta\\right)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Here, \\( A_{ijk} \\) represents the attention coefficient corresponding to the \\( i \\)-th head, the \\( j \\)-th output patch, and the \\( k \\)-th input patch. Regularization will increase the values of attention coefficients that are below the threshold \\( \\delta \\). For a specific output patch, since a softmax function is applied to its corresponding attention coefficients, increasing the lower values naturally decreases the higher ones. With \\( \\delta = \\frac{1}{(\\text{number of input patches})} \\), the coefficient distribution gradually shifts toward a uniform distribution at this threshold, representing the maximum attention degree a patch can achieve. The total loss function \\( \\mathcal{L} \\) used for training the model is the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20388" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.443, + 0.107 + ], + "angle": 0, + "content": "sum of the cross-entropy and the regularization term:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.119, + 0.484, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(x, x _ {m}\\right) = \\mathcal {L} _ {\\mathrm {C E}} \\left(x, x _ {m}\\right) + \\alpha R (x) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.148, + 0.485, + 0.269 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a scaling factor that controls the strength of the regularization. This combined objective encourages the model to not only accurately predict anomaly locations but also maintain an increased level of attention, enhancing both detection performance and adversarial robustness. It is then used in adversarial training, as explained in Section 5.3. Comprehensive ablation studies of hyperparameters including \\(\\delta\\) and \\(\\alpha\\) are provided in Appendix C.2 and Section 7." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.279, + 0.383, + 0.296 + ], + "angle": 0, + "content": "5.3. Adversarial Training and Testing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.484, + 0.346 + ], + "angle": 0, + "content": "Given an input sample \\( x \\), an adversarial example \\( x_{adv} \\) is generated by introducing a perturbation \\( \\delta^{*} \\), optimized to maximize the final loss:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.358, + 0.484, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\delta^ {*} = \\underset {\\| \\delta \\| _ {\\infty} \\leq \\epsilon} {\\operatorname {a r g m a x}} \\mathcal {L} (x + \\delta , x _ {m}), \\quad x _ {a d v} = x + \\delta^ {*} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.485, + 0.473 + ], + "angle": 0, + "content": "These adversarial examples are then used in the training process alongside the original examples. The adversarial training objective is framed as a min-max problem, optimizing the model parameters \\(\\theta\\) to minimize the expected loss over both clean and adversarial examples:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.483, + 0.484, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {(x, y) \\in \\mathcal {B}} \\left[ \\max _ {\\| \\delta \\| _ {\\infty} \\leq \\epsilon} \\mathcal {L} (x + \\delta , x _ {m}; \\theta) \\right]. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.528, + 0.224, + 0.545 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.75 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of PatchGuard, we conduct extensive experiments assessing its performance alongside several established anomaly detection methods on multiple benchmarks for anomaly detection and localization. Our evaluation includes both adversarially trained models and those trained under standard conditions. To ensure a fair comparison, we assess all methods under both \"Clean\" and \"Adversarial\" test conditions, as shown in Table 1 and 2. In the Clean setting, test samples are unaltered, whereas the Adversarial setting includes samples perturbed by the PGD attack [48]. Additional results on PatchGuard's performance against alternative attacks are provided in Section 7.2. Datasets are detailed in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Evaluation Details. In the adversarial setting, we evaluate each method using the \\(\\ell_{\\infty}\\) PGD-1000 attack with \\(\\epsilon = \\frac{8}{255}\\). Detection and localization results are shown in Tables 1 and 2, respectively. \\(\\ell_{2}\\) norm evaluations are in Appendix J." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Adapting State-of-the-Art AD and AL Methods to Adversarial Settings. Prior to developing our novel Patch-Guard framework, we first investigated whether existing state-of-the-art methods in AD and AL could be enhanced through adversarial training—a technique where models are" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.121 + ], + "angle": 0, + "content": "Table 3. Performance of prior methods adapted for adversarial settings, shown as \"Clean / Adversarial\" (PGD-1000, \\(\\epsilon = \\frac{8}{255}\\))." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.131, + 0.907, + 0.416 + ], + "angle": 0, + "content": "
MethodTaskDataset
MVTec ADVisABTADBraTS2021
PatchCoreAD92.3 / 14.388.4 / 10.887.1 / 12.788.6 / 16.1
AL89.1 / 10.387.6 / 8.885.0 / 11.586.7 / 13.6
ReContrastAD93.8 / 5.193.7 / 16.789.9 / 13.478.2 / 13.1
AL94.5 / 17.292.5 / 18.288.6 / 17.990.8 / 14.2
SimpleNetAD91.1 / 18.189.2 / 14.988.6 / 12.775.3 / 14.2
AL94.1 / 19.589.4 / 15.888.2 / 11.790.1 / 17.0
GeneralADAD84.6 / 24.781.2 / 22.182.9 / 20.680.7 / 19.9
AL83.7 / 21.883.1 / 17.384.7 / 18.379.6 / 21.0
DRAEMAD83.0 / 19.281.7 / 16.879.1 / 17.057.4 / 12.2
AL86.1 / 17.578.6 / 15.080.9 / 18.673.6 / 14.9
GLASSAD84.2 / 26.783.5 / 19.984.7 / 25.483.7 / 20.1
AL85.9 / 21.287.1 / 21.783.0 / 19.085.8 / 23.6
OursAD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 / 85.293.2 / 73.097.7 / 94.5
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.583 + ], + "angle": 0, + "content": "trained on adversarial examples to improve their robustness. In Appendix E, we provide a detailed methodology for generating adversarial examples for each method and describe our optimized approach for improving their resilience. The empirical results, presented in Table 3, reveal that even after incorporating adversarial training, current SOTA AD and AL methods remain vulnerable to adversarial attacks and exhibit suboptimal performance. These limitations motivated the development of PatchGuard, our proposed solution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Analyzing Results. As shown in Tables 1 and 2, prior SOTA methods such as Recontrast demonstrate substantial drops in performance under adversarial settings, despite excelling in clean conditions. Furthermore, as indicated in Table 3, even when adversarily trained, these methods struggle to achieve robust performance. By contrast, approaches specifically designed to address adversarial robustness, such as RODEO, still fall short when compared to PatchGuard. On average, PatchGuard improves robust detection across diverse datasets by up to \\(53.2\\%\\) percent, with a notable robust improvement of \\(68.5\\%\\) percent on the challenging MVtec dataset. Notably, we achieve a \\(71.8\\%\\) improvement in AL on the challenging VisA dataset, with details in Appendix F. As prior research indicates, a slight decrease in clean performance is generally viewed as an acceptable tradeoff for the critical gains in robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.904 + ], + "angle": 0, + "content": "Implementation Details. We leverage a from-scratch ViT model in our pipeline to showcase our setup's independence from pre-train datasets. However, we conduct an ablation study on this manner in Appendix C.4. Further, we leverage PGD-10 with \\(\\epsilon = \\frac{8}{255}\\) for adversarial training," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20389" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.131 + ], + "angle": 0, + "content": "Table 4. Evaluation of PatchGuard's performance against advanced classification and segmentation attacks, showcasing its robustness in diverse attack scenarios." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.144, + 0.485, + 0.304 + ], + "angle": 0, + "content": "
DatasetTaskDetection AttacksLocalization Attacks
PGD-1000CAAAutoAttackA3SegPGDSEA
MVTec ADAD71.174.870.869.871.969.2
AL73.876.772.970.074.671.3
VisAAD74.378.574.673.175.073.6
AL85.287.485.584.785.083.7
BTADAD82.183.482.080.781.980.8
AL73.074.772.371.572.971.1
BraTS2021AD81.084.781.980.382.081.1
AL94.595.094.293.894.893.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.348 + ], + "angle": 0, + "content": "Table 5. Ablation study on the scaling factor \\( \\alpha \\), illustrating the effectiveness of our regularization." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.36, + 0.485, + 0.53 + ], + "angle": 0, + "content": "
Scaling Factor (α)TaskDataset
MVTec-ADVisABTADBraTS2021
0AD89.3 / 60.790.1 / 62.988.6 / 74.395.2 / 72.0
AL93.5 / 65.395.8 / 72.194.0 / 59.798.2 / 83.4
1 (Ours)AD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 / 85.293.2 / 73.097.7 / 94.5
2AD86.2 / 70.187.6 / 75.184.7 / 83.492.8 / 81.2
AL91.0 / 68.994.8 / 84.191.9 / 73.995.8 / 94.6
5AD80.1 / 64.178.6 / 65.473.8 / 69.682.5 / 73.2
AL84.7 / 65.681.1 / 78.984.0 / 65.386.3 / 79.9
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.486, + 0.605 + ], + "angle": 0, + "content": "but the test results are generated with the PGD-1000 attack. For more details on the implementation parameters, refer to Appendix G." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.617, + 0.258, + 0.634 + ], + "angle": 0, + "content": "7. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.485, + 0.673 + ], + "angle": 0, + "content": "In this section, we provided further analysis of our choices of parameters and other components of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.245, + 0.699 + ], + "angle": 0, + "content": "7.1. Regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.486, + 0.84 + ], + "angle": 0, + "content": "To experimentally demonstrate the effectiveness of our regularization term on the model's robustness and the impact of the scaling factor \\(\\alpha\\) (set to 1 by default in our method), we conduct an ablation study on \\(\\alpha\\), as shown in Table 5. When \\(\\alpha = 0\\), the model's robustness against adversarial attacks decreases due to the lack of regularization. On the other hand, higher values of \\(\\alpha\\), such as \\(\\alpha = 5\\), cause the regularization term to dominate, leading to an unstable loss function and inappropriate results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.27, + 0.865 + ], + "angle": 0, + "content": "7.2. Advanced Attacks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In our primary experiments in Section 6, we utilized variations of the PGD attack [48] for both model training and" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.12 + ], + "angle": 0, + "content": "Table 6. An ablation study on our method's performance using various anomaly generation techniques that incorporate masks." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.131, + 0.907, + 0.33 + ], + "angle": 0, + "content": "
Pseudo-Anomaly GeneratorTaskDataset
MVTec ADVisABTADBraTS2021
CutPasteAD79.1 / 60.374.4 / 61.574.8 / 69.280.3 / 72.7
AL83.9 / 60.885.4 / 73.082.6 / 61.291.2 / 82.1
FPIAD78.9 / 62.175.1 / 64.372.9 / 71.084.1 / 74.1
AL84.3 / 63.882.9 / 71.886.3 / 60.593.3 / 84.8
NSAAD81.7 / 66.583.1 / 65.477.1 / 73.486.3 / 73.8
AL85.1 / 65.083.3 / 76.785.4 / 64.989.4 / 86.8
GLASSAD83.2 / 62.084.0 / 68.779.3 / 73.983.1 / 66.1
AL86.7 / 67.086.2 / 78.586.8 / 68.292.2 / 87.3
OursAD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 /85.293.2 /73.097.7 /94.5
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.909, + 0.571 + ], + "angle": 0, + "content": "testing. To further illustrate the adaptability and robustness of our proposed method under diverse adversarial conditions, we also evaluate its performance against a variety of other attack types including Detection attacks and Localization attacks. The Detection attacks consist of black-box attacks [28], FGSM [24], CAA [49], AutoAttack [16], and \\(\\mathrm{A}^3\\) (Adversarial Attack Automation) [43]. The localization attacks are adapted from the well-known segmentation attacks in literature, namely SegPGD [27] and SEA [17]. All results are available in Table 4, and the exact details on our adaptation approach for these attacks are available in Appendix H. It is important to note that the training process remains consistent, employing only the standard PGD-10 for simplicity and practicality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.581, + 0.842, + 0.599 + ], + "angle": 0, + "content": "7.3. Pseudo-Anomaly Generation Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.909, + 0.741 + ], + "angle": 0, + "content": "Since our method targets both AD and AL using pseudo-anomalies, it is crucial to evaluate our pseudo-anomaly generation against compatible methods that produce localization maps. Thus, methods like MIXUP [33], FakeIt [50], and Dream-OOD [20] are excluded. We compare instead with CutPaste [40], NSA [59], GLASS [12], and FPI [67], keeping other pipeline components fixed. As shown in Table 6, our method outperforms these approaches, highlighting its effectiveness for robust anomaly localization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.634, + 0.771 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We introduce PatchGuard, an adversarially robust method for AD and AL. It employs Foreground-Aware Pseudo-Anomaly Generation using segmentation masks. Additionally, we propose a regularization term, motivated by observations and theory, that strengthens robustness against adversarial attacks by increasing the attention degree. PatchGuard outperforms previous methods on AD and AL benchmarks under adversarial settings." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20390" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.172 + ], + "angle": 0, + "content": "We are very grateful to Mohammad Javad Maheronnaghsh for his collaboration in the initial stages of this project and to the anonymous reviewers for their helpful discussions and valuable feedback." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.203, + 0.188, + 0.218 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.227, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[1] M. Eren Akbiyik. Data augmentation in training cnns: Injecting noise to images. ArXiv, abs/2307.06855, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.258, + 0.484, + 0.298 + ], + "angle": 0, + "content": "[2] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. *IEEE Access*, 6:14410–14430, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[3] Mohammad Azizmalayeri, Arshia Soltani Moakhar, Arman Zarei, Reihaneh Zohrabi, Mohammad Manzuri, and Mohammad Hossein Rohhan. Your out-of-distribution detection method is not robust! Advances in Neural Information Processing Systems, 35:4887-4901, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[4] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.484, + 0.513 + ], + "angle": 0, + "content": "[5] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1893-1902, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.516, + 0.484, + 0.584 + ], + "angle": 0, + "content": "[6] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.484, + 0.656 + ], + "angle": 0, + "content": "[7] Louis Béthane, Paul Novello, Thibaut Boissin, Guillaume Coiffier, Mathieu Serrurier, Quentin Vincenot, and Andres Troya-Galvis. Robust one-class classification with signed distance function using 1-lipschitz neural networks. arXiv preprint arXiv:2303.01978, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.484, + 0.727 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.484, + 0.826 + ], + "angle": 0, + "content": "[9] Chengwei Chen, Yuan Xie, Shaohui Lin, Ruizhi Qiao, Jian Zhou, Xin Tan, Yi Zhang, and Lizhuang Ma. Novelty detection via contrastive learning with negative data augmentation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI 2021, Virtual Event / Montreal, Canada, 19-27 August 2021, pages 606-614. ijcai.org, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[10] Jiefeng Chen, Yixuan Li, Xi Wu, Yingyu Liang, and Somesh Jha. Robust out-of-distribution detection for neural networks. arXiv preprint arXiv:2003.09711, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Jiefeng Chen, Yixuan Li, Xi Wu, Yingyu Liang, and Somesh Jha. Atom: Robustifying out-of-distribution detection using" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.227, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.149 + ], + "angle": 0, + "content": "outlier mining. In Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13-17, 2021, Proceedings, Part III 21, pages 430-445. Springer, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.907, + 0.204 + ], + "angle": 0, + "content": "[12] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. arXiv preprint arXiv:2407.09359, 2024. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[15] Matan Jacob Cohen and Shai Avidan. Transformally-two (feature spaces) are better than one. arXiv preprint arXiv:2112.04185, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.363, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[16] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.907, + 0.46 + ], + "angle": 0, + "content": "[17] Francesco Croce, Naman D Singh, and Matthias Hein. Towards reliable evaluation and fast training of robust semantic segmentation models. In ECCV, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[18] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.506, + 0.907, + 0.586 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.907, + 0.63 + ], + "angle": 0, + "content": "[20] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. arXiv preprint arXiv:2309.13415, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[21] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[22] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[23] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. arXiv preprint arXiv:2012.07177, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[24] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[25] Adam Goodge, Bryan Hooi, See Kiong Ng, and Wee Siong Ng. Robustness of autoencoders for anomaly detection under adversarial impact. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 1244-1250, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.227, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20391" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[26] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[27] Jindong Gu, Hengshuang Zhao, Volker Tresp, and Philip HS Torr. Segpgd: An effective and efficient adversarial attack for evaluating and boosting segmentation robustness. In European Conference on Computer Vision, pages 308-325. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[28] Chuan Guo, Jacob Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Weinberger. Simple black-box adversarial attacks. In International Conference on Machine Learning, pages 2484-2493. PMLR, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[29] Jia Guo, Lize Jia, Weihang Zhang, Huiqi Li, et al. Recontrast: Domain-specific anomaly detection via contrastive reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.484, + 0.482 + ], + "angle": 0, + "content": "[31] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.484, + 0.525 + ], + "angle": 0, + "content": "[32] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.484, + 0.567 + ], + "angle": 0, + "content": "[33] Yann N. Dauphin Hongyi Zhang, Moustapha Cisse. mixup: Beyond empirical risk minimization. International Conference on Learning Representations, 2018. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.484, + 0.609 + ], + "angle": 0, + "content": "[34] Mohammad Hossein Rohban Hossein Mirzaei, ... Scanning trojaned models using out-of-distribution samples. In Advances in Neural Information Processing Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.61, + 0.484, + 0.664 + ], + "angle": 0, + "content": "[35] Jiande Huang, Xingxing Li, Xin Li, Jiaqi Wu, Keke Zhang, Yongqiang Yuan, and Wei Zhang. Real-time outlier detection of satellite orbit and clock products using reverse error estimation. GPS Solutions, 29(1):2, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.484, + 0.706 + ], + "angle": 0, + "content": "[36] Nicholas Jeffrey, Qing Tan, and José R Villar. A review of anomaly detection strategies to detect threats to cyberphysical systems. *Electronics*, 12(15):3283, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.484, + 0.776 + ], + "angle": 0, + "content": "[37] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.777, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[38] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[39] Jungbeom Lee, Eunj Kim, Jisoo Mok, and Sungroh Yoon. Anti-adversarily manipulated attributions for weakly supervised semantic segmentation and object localization. IEEE transactions on pattern analysis and machine intelligence, 46 (3):1618-1634, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[40] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9664-9674, 2021. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.908, + 0.218 + ], + "angle": 0, + "content": "[41] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. Musc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. arXiv preprint arXiv:2401.16753, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.908, + 0.273 + ], + "angle": 0, + "content": "[42] Jiaqi Liu, Guoyang Xie, Jinbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. Machine Intelligence Research, 21(1):104-135, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.908, + 0.316 + ], + "angle": 0, + "content": "[43] Ye Liu, Yaya Cheng, Lianli Gao, Xianglong Liu, Qilong Zhang, and Jingkuan Song. Practical evaluation of adversarial robustness via adaptive auto attack, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.908, + 0.384 + ], + "angle": 0, + "content": "[44] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.908, + 0.428 + ], + "angle": 0, + "content": "[45] Shao-Yuan Lo, Poojan Oza, and Vishal M Patel. Adversarily robust one-class novelty detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.908, + 0.469 + ], + "angle": 0, + "content": "[46] Pauline Luc, Camille Couprie, Soumith Chintala, and Jakob Verbeek. Semantic segmentation using adversarial networks. arXiv preprint arXiv:1611.08408, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.908, + 0.525 + ], + "angle": 0, + "content": "[47] Yuan Luo, Ya Xiao, Long Cheng, Guojun Peng, and Danfeng Yao. Deep learning-based anomaly detection in cyberphysical systems: Progress and opportunities. ACM Computing Surveys (CSUR), 54(5):1-36, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.908, + 0.58 + ], + "angle": 0, + "content": "[48] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.908, + 0.622 + ], + "angle": 0, + "content": "[49] Xiaofeng Mao, Yuefeng Chen, Shuhui Wang, Hang Su, Yuan He, and Hui Xue. Composite adversarial attacks. ArXiv, abs/2012.05434, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.908, + 0.692 + ], + "angle": 0, + "content": "[50] Hossein Mirzaei, Mohammadreza Salehi, Sajjad Shahabi, Efstratios Gavves, Cees GM Snoek, Mohammad Sabokrou, and Mohammad Hossein Rohban. Fake it till you make it: Near-distribution novelty detection by score-based generative models. arXiv preprint arXiv:2205.14297, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.908, + 0.789 + ], + "angle": 0, + "content": "[51] Hossein Mirzaei, Mohammad Jafari, Hamid Reza Dehbashi, Ali Ansari, Sepehr Ghobadi, Masoud Hadi, Arshia Soltani Moakhar, Mohammad Azizmalayeri, Mahdieh Soleymani Baghshah, and Mohammad Hossein Rohban. Rodeo: Robust outlier detection via exposing adaptive out-of-distribution samples. In Forty-first International Conference on Machine Learning, 2024. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.908, + 0.873 + ], + "angle": 0, + "content": "[52] Hossein Mirzaei, Mohammad Jafari, Hamid Reza Dehbashi, Zeinab Sadat Taghavi, Mohammad Sabokrou, and Mohammad Hossein Rohban. Killing it with zero-shot: Adversarially robust novelty detection. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7415-7419. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.908, + 0.902 + ], + "angle": 0, + "content": "[53] Hossein Mirzaei, Mojtaba Nafez, Mohammad Jafari, Mohammad Bagher Soltani, Mohammad Azizmalayeri, Jafar Habibi," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20392" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "Mohammad Sabokrou, and Mohammad Hossein Rohban. Universal novelty detection through adaptive contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22914-22923, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.486, + 0.204 + ], + "angle": 0, + "content": "[54] Arian Mousakhan, Thomas Brox, and Jawad Tayyyub. Anomaly detection with conditioned denoising diffusion models. arXiv preprint arXiv:2305.15956, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[55] Fahimeh Hosseini Noohdani, Parsa Hosseini, Aryan Yazdan Parast, Hamidreza Yaghoubi Araghi, and Mahdieh Soleymani Baghshah. Decompose-and-compose: A compositional approach to mitigating spurious correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27662-27671, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[56] Pramuditha Perera, Poojan Oza, and Vishal M Patel. One-class classification: A survey. arXiv preprint arXiv:2101.03064, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[57] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.484, + 0.446 + ], + "angle": 0, + "content": "[58] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.484, + 0.504 + ], + "angle": 0, + "content": "[59] Hannah M Schlüter, Jeremy Tan, Benjamin Hou, and Bernhard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In European Conference on Computer Vision, pages 474-489. Springer, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.484, + 0.574 + ], + "angle": 0, + "content": "[60] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[61] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16, pages 682-698. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.484, + 0.7 + ], + "angle": 0, + "content": "[62] Abhishek Sinha, Kumar Ayush, Jiaming Song, Burak Uzkent, Hongxia Jin, and Stefano Ermon. Negative data augmentation. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[63] Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister. Learning and evaluating representations for deep one-class classification. arXiv preprint arXiv:2011.02578, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[64] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Generalad: Anomaly detection across domains by attending to distorted features. arXiv preprint arXiv:2407.12427, 2024. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[65] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[66] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distribu" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "tionally shifted instances. Advances in neural information processing systems, 33:11839-11852, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[67] Jeremy Tan, Benjamin Hou, James Batten, Huaqi Qiu, and Bernhard Kainz. Detecting outliers with foreign patch interpolation. Machine Learning for Biomedical Imaging, 1(April 2022):1-27, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[68] Maximilian E Tschuchnig and Michael Gadermayr. Anomaly detection in medical imaging-a mini review. In Data Science-Analysics and Applications: Proceedings of the 4th International Data Science Conference-iDSC2021, pages 33-38. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.906, + 0.304 + ], + "angle": 0, + "content": "[69] Qi Wang, Junyu Gao, and Xuelong Li. Weakly supervised adversarial domain adaptation for semantic segmentation in urban scenes. IEEE Transactions on Image Processing, 28 (9):4376-4386, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.374 + ], + "angle": 0, + "content": "[70] Weiping Wang, Zhaorong Wang, Zhanfan Zhou, Haixia Deng, Weiliang Zhao, Chunyang Wang, and Yongzhen Guo. Anomaly detection of industrial control systems based on transfer learning. Tsinghua Science and Technology, 26(6): 821-832, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.907, + 0.418 + ], + "angle": 0, + "content": "[71] Yue Xing, Qifan Song, and Guang Cheng. Why do artificially generated data help adversarial robustness. Advances in Neural Information Processing Systems, 35:954-966, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.461 + ], + "angle": 0, + "content": "[72] Han Xu, Yao Ma, Haochen Liu, Debayan Deb, Hui Liu, Jiliang Tang, and Anil K. Jain. Adversarial attacks and defenses in images, graphs and text: A review, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.907, + 0.504 + ], + "angle": 0, + "content": "[73] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, pages 1-28, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.907, + 0.573 + ], + "angle": 0, + "content": "[74] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. arXiv preprint arXiv:2406.07487, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.907, + 0.63 + ], + "angle": 0, + "content": "[75] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.907, + 0.687 + ], + "angle": 0, + "content": "[76] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE transactions on neural networks and learning systems, 30(9): 2805-2824, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.907, + 0.744 + ], + "angle": 0, + "content": "[77] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.907, + 0.814 + ], + "angle": 0, + "content": "[78] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.907, + 0.857 + ], + "angle": 0, + "content": "[79] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Dsr – a dual subspace re-projection network for surface anomaly detection, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[80] Chaoning Zhang, Philipp Benz, Chenguo Lin, Adil Karjauv, Jing Wu, and In So Kweon. A survey on universal adversarial attack. arXiv preprint arXiv:2103.01498, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20393" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[81] Hui Zhang, Zheng Wang, Zuxuan Wu, and Yu-Gang Jiang. Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection. arXiv preprint arXiv:2303.08730, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[82] Lin Zhang and Yang Dai. A gan anomaly detection method based on multi-scale endogenous enhancement. In Blockchain and Web3 Technology Innovation and Application Exchange Conference, pages 269-281. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.274 + ], + "angle": 0, + "content": "[83] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "20394" + } + ] +] \ No newline at end of file diff --git a/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/6310723f-d72e-495d-87c7-6b59d06d7ca9_origin.pdf b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/6310723f-d72e-495d-87c7-6b59d06d7ca9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9100592a0ae8fd33e5511fd33bea84028b6aeab7 --- /dev/null +++ b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/6310723f-d72e-495d-87c7-6b59d06d7ca9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d51ae3f9db17b8552148218d48c672c00a76859168fc81791f0ecb880e911370 +size 4990575 diff --git a/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/full.md b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8e298a453edd95efec5740728c4c00767bc36cc4 --- /dev/null +++ b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/full.md @@ -0,0 +1,330 @@ +# PatchGuard: Adversarily Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies + +Mojtaba Nafez1 Amirhossein Koochakian1 * Arad Maleki1 * Jafar Habibi1 Mohammad Hossein Rohban1 1Sharif University of Technology, Iran + +{mojtaba.nafez77, amir.kochakian12, arad.maleki02, rohban}@sharif.edu + +# Abstract + +Anomaly Detection (AD) and Anomaly Localization (AL) are crucial in fields that demand high reliability, such as medical imaging and industrial monitoring. However, current AD and AL approaches are often susceptible to adversarial attacks due to limitations in training data, which typically include only normal, unlabeled samples. This study introduces PatchGuard, an adversarily robust AD and AL method that incorporates pseudo anomalies with localization masks within a Vision Transformer (ViT)-based architecture to address these vulnerabilities. We begin by examining the essential properties of pseudo anomalies, and follow it by providing theoretical insights into the attention mechanisms required to enhance the adversarial robustness of AD and AL systems. We then present our approach, which leverages Foreground-Aware Pseudo-Anomalies to overcome the deficiencies of previous anomaly-aware methods. Our method incorporates these crafted pseudo-anomaly samples into a ViT-based framework, with adversarial training guided by a novel loss function designed to improve model robustness, as supported by our theoretical analysis. Experimental results on well-established industrial and medical datasets demonstrate that PatchGuard significantly outperforms previous methods in adversarial settings, achieving performance gains of $53.2\%$ in AD and $68.5\%$ in AL, while also maintaining competitive accuracy in non-adversarial settings. The code repository is available at here. + +# 1. Introduction + +In recent years, Anomaly Detection (AD) and Anomaly Localization (AL) have become critical techniques across various fields, including medical imaging, fraud detection, and industrial monitoring [22, 36, 42, 47, 68, 70, 73]. While state-of-the-art (SOTA) AD and AL methods exhibit near-perfect performance on standard benchmarks [6, 83], they + +![](images/f77e3f787947d6371e0f60b757d73fc6f18f6d3ff42807f4c70ce0a0b193ecb8.jpg) +Figure 1. Impact of Adversarial Attacks on Anomaly Localization Methods: Localization maps for multiple methods are shown before and after a PGD-1000 attack, illustrating the vulnerability of existing methods to adversarial attacks, even when performing perfectly in clean conditions. Our proposed method demonstrates enhanced robustness in these adversarial scenarios. + +frequently face significant performance degradation when subjected to adversarial attacks, as illustrated in Figure 1 [12, 54, 64, 66, 74, 81]. These attacks involve introducing subtle, imperceptible perturbations into the input data, misleading models to misclassify normal1 samples as anomalies or to overlook actual anomalies [2, 24, 48, 65, 80]. To address this vulnerability, a variety of strategies have been proposed to strengthen the adversarial robustness of AD systems [3, 7, 10, 25, 45, 61]. However, despite these advancements, the resilience of these methods remains limited, with even mild adversarial attacks still causing noticeable performance drops, even on widely used anomaly benchmarks [6, 83]. Furthermore, adversarial robustness in AL remains an unexplored area, highlighting a critical gap in the current research. + +Adversarial training, a strategy designed to improve model robustness in classification tasks, incorporates adversarial samples into the training process [48, 76]. Despite its success in standard classification, this approach proves less effective in AD and AL scenarios [51]. The primary + +obstacle to achieving adversarial robustness in AD lies in the absence of anomalous samples during training [10, 11]. Without such samples, models cannot be trained on adversarial perturbations specific to anomalous data [5, 51, 56, 58], leaving them vulnerable to attacks on perturbed anomalies [3, 10, 11]. To address this issue in the AD setup, recent efforts have combined Outlier Exposure (OE) [32] with adversarial training [3, 10, 11]. However, this method struggles when applied to high-resolution, real-world datasets [51]. This challenge is even more pronounced in AL settings, as robust AL methods must perform pixel-level classification, which inherently requires pixel-level supervision, while at the same time operating within the same data constraints as the AD framework [35, 82]. Despite ongoing research, the adversarial robustness of AL methods remains an open problem, with no comprehensive solutions proposed to date. + +In this study, we identify two essential properties that OE techniques must satisfy for effective application in adversarially robust AD and AL. First, the outlier exposures must include localization maps to facilitate the training of robust AL models. This requirement aligns with prior research on adversarially robust segmentation models, which underscores the importance of generating localization maps to enhance model robustness [39, 46, 69]. Second, as noted by [51], OE samples should be "near-distributed" relative to normal data, meaning they should share similar semantic and stylistic attributes. This condition is consistent with the literature on adversarial robustness, which highlights the advantages of near-distribution decision boundaries samples for optimizing the model robustness [71]. + +Moreover, the recent success of Vision Transformer (ViT)-based architectures in clean AD and AL tasks [15, 37, 41, 64], attributed to their global contextual modeling [19], motivated us to explore their potential in adversarial settings. For an input image $x$ and a fixed attention layer $\ell$ , we define the Attention Degree of $\ell$ under $x$ as the average number of input embeddings each output embedding attends to. As shown in Figure 2, we observe that images that induce higher attention degrees in the model are more resistant to adversarial attacks. Furthermore, adversarial training appears to implicitly steer the model toward achieving a high attention degree. We provide a theoretical analysis and justification for these findings in Section 4. + +Building on these insights, we propose PatchGuard, an adversarially robust AD and AL method that addresses limitations in current OE-based methods and leverages attention-based insights. To address OE challenges, we generate pseudo-anomaly samples from normal data, ensuring they meet two essential criteria: availability of localization maps and near-distribution alignment with inliers. To create samples that exhibit anomalous characteristics while remaining close to the inlier distribution, we introduce Foreground-Aware Pseudo-Anomaly Generation. This approach iden + +tifies causal regions in normal samples using an enhanced Grad-CAM [60] feature attribution technique with tailored augmentations. These identified regions are then selectively distorted, drawing on insights from [55], to produce anomaly samples that retain stylistic and semantic coherence with the original data. By precisely controlling the distortion location, we can generate accurate localization maps for the pseudoanomalies, effectively meeting both criteria for robust AD and AL. + +In the next step, we leverage normal samples and pseudoanomalies to perform adversarial training on a ViT-based architecture. Building on the provided insights, we introduce a novel loss function designed to increase the model's last-layer attention degree. As our theoretical justifications suggest, this property significantly enhances the model's adversarial robustness in both AD and AL settings. + +Contribution. In this paper, we propose PatchGuard, an adversarially robust AD and AL method. To the best of our knowledge, we are the first to raise and effectively address the issue of adversarial robustness in AL. We provide insights into the shortcomings of previous OE-based methods and introduce the Foreground-Aware Pseudo-Anomaly Generation framework to address these limitations. Additionally, we provide insights and theoretical justifications on enhancing the adversarial robustness of ViT-based architectures, and propose a novel loss function aligned with these findings. We evaluate PatchGuard in both clean and adversarial settings, rigorously testing its robustness under several strong attacks, including PGD-1000 [48], $\mathrm{A}^3$ [43], and SEA [17]. Results demonstrate that PatchGuard significantly outperforms existing methods in adversarial contexts, achieving an AUROC improvement of $53.2\%$ in AD and $68.5\%$ in AL, while maintaining competitive performance in standard conditions. Our evaluations span various well-established industrial and medical datasets, including MVTec [6], VisA [83], and BraTS2021 [4], highlighting PatchGuard's applicability. + +# 2. Related Work + +From an adversarial robustness perspective, anomaly detection and localization methods can be mainly classified into three categories: 1) methods with anomaly-free training, 2) methods that leverage anomalies at the embedding level, and 3) methods that leverage anomalies at the input level. We experimentally adapted methods from each category to adversarial training processes, and the results reveal they still exhibit significant vulnerabilities even after adversarial training. The first category consists of methods such as Recontrast[29], Transformaly [15], and PatchCore [57], that operate on the hypothesis that models trained exclusively on normal samples can reconstruct normal images well but fail to do so with abnormal regions. These methods lack robustness, since the model does not encounter any anomaly + +data during adversarial training, meaning that when adversarial perturbations are added to anomaly data, the model still remains vulnerable. The second category, which includes methods such as SimpleNet [44], GeneralAD [64], UniAD [75], and DSR [79], utilize a feature extractor during training to obtain feature vectors of the normal set. By perturbing these features, they create anomaly features and primarily train a discriminator over these normal and anomaly features. However, these methods lack robustness due to their reliance on highly vulnerable feature extractors and the absence of anomaly data in the input space, which prevents proper training of a robust feature extractor. The third category consists of DRAEM [78], NSA [59], Cutpaste [40], and GLASS [12], generate pseudo-anomaly data during training. Nevertheless, some of these methods lack access to masks or do not use a foreground-aware generation process. Despite their astonishing success in clean settings, we empirically show in Table 3 that even when adversarily trained, they exhibit unsatisfactory results. Furthermore, while anomaly localization has not been explored in an adversarial setup, some other works have pursued anomaly detection in an adversarial setting, including PrincipaLS [45], ZARND [52], and Rodeo [51]. These approaches have achieved relatively better results by incorporating pseudo-anomaly generation processes and adversarial training. However, their performance falls short in real-world high-resolution datasets (as presented in Table 1). For more details about previous works, see Appendix A. + +# 3. Preliminaries + +Anomaly Detection (AD) and Localization (AL): AD involves identifying instances that deviate from expected patterns, while AL focuses on localizing abnormal subregions within an image. In AD, a model assigns an anomaly score $S(x; f)$ to each sample $x$ using model $f$ ; if this score exceeds a predefined threshold, the sample is classified as anomalous. Conversely, lower scores indicate normal samples. In AL, a model generates an anomaly map $M(x; f)$ using model $f$ , where each pixel of the map corresponds to an anomaly score for the respective pixel in the input image. + +Adversarial Robustness of Anomaly Detectors: An adversarial attack, primarily studied within classification tasks, involves the malicious alteration of a data sample $x$ with its label $y$ into a new sample $x^{*}$ that maximizes the loss function $\ell(x^{*};y)$ [34, 72, 76]. This new sample $x^{*}$ is known as an adversarial example of $x$ , and the difference $(x^{*} - x)$ is referred to as adversarial perturbation. To prevent significant semantic changes, the $l_{p}$ norm of the adversarial noise is restricted by an upper limit $\epsilon$ . Specifically, an adversarial example $x^{*}$ must satisfy the condition $x^{*} = \arg \max_{x^{\prime}:\|x - x^{\prime}\|_{p} \leq \epsilon} \ell(x^{\prime};y)$ . A widely used and effective attack method is the Projected Gradient Descent (PGD) [48] technique, which iteratively increases the loss function by moving in the direction of the gradient sign of + +$\ell(x^{*};y)$ with a specified step size $\alpha$ . When adapting adversarial attacks for AD, the objective shifts from maximizing the loss value to increasing $S(x,f)$ for normal samples and decreasing it for anomalies. The attack is formulated as $x_0^* = x$ , $x_{t+1}^* = x_t^* + y \cdot \alpha \cdot \mathrm{sign}(\nabla_x S(x_t^*;f_\theta))$ , and $x^* = x_k^*$ , where $y = +1$ for normal samples and $y = -1$ for anomaly samples. This approach is consistently applied to other attacks in the study [21, 76]. + +Adversarial Robustness of Anomaly Localization: Adversarial attacks aim to introduce perturbations that uniformly modify pixel scores in an anomaly map $M$ , flipping scores from 0 to 1 or vice versa. In the context of AL, the attack objective transitions from merely maximizing a loss function to specifically increasing the values of $M(x, f)$ for normal samples while decreasing them for anomalies. The attack is formalized as follows: starting with $x_0^* = x$ , the adversarial example at each step is updated as $x_{t+1}^* = x_t^* + Y \cdot \alpha \cdot \mathrm{sign}(\nabla_x M(x_t^*; f_\theta))$ , and the final adversarial example is $x^* = x_k^*$ . Here, $Y_{i,j} = +1$ for normal pixels and $Y_{i,j} = -1$ for anomaly pixels, where $i, j$ denote pixel coordinates. This strategy is extended to other advanced attack methods throughout the study [17]. Despite its significance, adversarial robustness in anomaly localization remains underexplored. Enhancing robustness is particularly critical for applications requiring spatial precision, such as medical diagnostics and industrial monitoring systems. + +# 4. Theoretical Insights + +![](images/79c34287f45507354ea2031fb5ff756f66282d5125e01d025f8cd07a61938e5a.jpg) +Figure 2. The figure demonstrates how increasing the average last-layer attention degree in the ViT-base architecture reduces vulnerability to adversarial attacks. Specifically, the BraTS dataset images were clustered into five groups based on their attention degree values (x-axis). The y-axis represents Vulnerability, measured as the absolute difference in the model's localization performance (AUROC%) for each cluster before and after a PGD attack. (a) illustrates the decreasing trend in a clean-trained model. (b) shows a similar decreasing trend in the adversarially trained model, where the attention degree is relatively higher than in the clean model. + +![](images/a0c1739ded82b78932cbf62868f2289f6788980eaa63019daa71bbb90600b27e.jpg) + +Here, we want to show why a self-attention layer, in which each token attends to a higher number of other tokens is more adversarially robust compared to the case that the attention is distributed over a smaller number of tokens. + +Let's consider the self-attention output embedding of the $i$ th token in the $l$ -th layer as $x_{i}^{(l)}$ , and such embedding after the perturbation is applied to the input as $\hat{x}_i^{(l)}$ . We are seeking to set $x_{i}^{(2)} - \hat{x}_{i}^{(2)} = r_{i}$ , where $r_i$ is the desired adversarial direction that most influences the layers next to the first layer of the transformer. According to the self-attention: + +$$ +x _ {i} ^ {(l + 1)} = V ^ {(l)}. \text {s o f t m a x} \left(q _ {i} ^ {(l) \top} K / \sqrt {d}\right), +$$ + +where $q_{i}^{(l)} = W_{q}x_{i}^{(l)},v_{j}^{(l)}$ the $j$ -th column of $V^{(l)}$ be $v_{j}^{(l)} = W_{v}x_{j}^{(l)},k_{j}^{(l)}$ the $j$ -th column of $\mathbf{K}$ be $k_{j}^{(l)} = W_{k}x_{j}^{(l)}$ , and $d$ be the embedding dimension. For simplicity, we assume that the dimensions of embeddings of key, query, and values are all the same and are equal to the dimension of the original embeddings. Now, $W_{v}(X^{(1)}S - \hat{X}^{(1)}\hat{S}) = R$ , where the columns of $R$ and $X^{(l)}$ are represented by $r_i$ and $x_{i}^{(l)}$ , and the columns in $S$ and $\hat{S}$ correspond to the softmax values of each query token calculated for $X$ and the perturbed inputs $\hat{X}$ , respectively. Therefore, + +$$ +W _ {v} (X ^ {(1)} S - \hat {X} ^ {(1)} S) = +$$ + +$$ +W _ {v} \left(X ^ {(1)} S - \hat {X} ^ {(1)} \hat {S} + \hat {X} ^ {(1)} \hat {S} - \hat {X} ^ {(1)} S\right) = +$$ + +$$ +R + W _ {v} \hat {X} ^ {(1)} (\hat {S} - S) = R + W _ {v} \hat {X} ^ {(1)} \Delta S. +$$ + +where $\Delta S\coloneqq \hat{S} -S$ . Hence, we have: + +$$ +X ^ {(1)} - \hat {X} ^ {(1)} = \underbrace {W _ {v} ^ {+} (R + W _ {v} \hat {X} ^ {(1)} \Delta S)} _ {U} S ^ {+}. +$$ + +Where $W_{v}^{+}$ and $S^{+}$ are the pseudo-inverses of the matrices $W_{v}$ and $S$ , respectively. Note that for the $\delta := X^{(1)} - \hat{X}^{(1)}$ perturbation in the input, in order to make the required perturbation in the output of the network, we get $\| \delta \| \leq \lambda_{max} \| U \|$ , where $\lambda_{max}$ is the maximum eigenvalue of $S^{+}$ . But we note that for a stochastic matrix $S$ , the maximum eigenvalue is 1. Therefore, for the inverse matrix, $S^{+}$ , the maximum eigenvalue would be larger than 1. As a result, we would expect the $\| \delta \|$ to grow proportional to $\lambda_{max} \geq 1$ . We note that the smaller $\| \delta \|$ would imply more vulnerability as it becomes easier to attack the network with a smaller perturbation length. On the other hand, for a purely localized $S$ , where each token only attends to itself $\lambda_{max} = 1$ , showing that the increase in $\| \delta \|$ would likely be small compared to the general case, leading to more vulnerability. Therefore, we conclude that a localized and overly sparse attention map would typically imply an easier attack on the network and hence higher vulnerability against adversarial attacks. Sparse attention values may also resemble what happens in convolutional networks that have small receptive fields. + +# 5. Methodology + +Overview. Current AD and AL methods perform poorly when subjected to adversarial attacks. Previous studies have + +demonstrated the effectiveness of OE in enhancing robustness against such attacks [3, 10, 11, 51]. Additionally, ViTs have shown promising results in standard (clean) AD and AL tasks [15, 37, 41, 64], and our findings suggest their potential for adversarial settings with targeted modifications of attention degrees. Building on these insights and our theoretical foundations, we propose PatchGuard, a novel method designed to enhance adversarial robustness in AD and AL. The following subsections detail the key components of PatchGuard. + +# 5.1. Foreground-Aware Pseudo-Anomaly Generation + +To generate anomalous data from normal samples, it is essential to apply targeted distortions that ensure the image becomes anomalous [8, 13, 14, 26]. Additionally, the generated pseudo-anomalies must satisfy two key requirements: (1) inclusion of localization maps to enhance the AL task, and (2) close alignment with the normal sample distribution [51, 53]. Meeting these requirements simultaneously necessitates precise but minimal alterations within the foreground. Thus, the proposed method must reliably identify a region of the image that we can confidently attribute to the foreground. + +To achieve this, we leverage insights from [55], which indicate that pretrained models benefit from style-agnostic processing to improve classification accuracy. To this end, we first introduce soft transformations, which are commonly used in self-supervised learning studies due to their ability to preserve semantic content [8, 13, 14, 26, 31]. Next, we apply Grad-CAM [60] to generate a saliency map using a standard pre-trained ResNet18 [30] model on a given input. Specifically, for a normal sample $x$ , we randomly select $k_{\mathrm{soft}}$ soft transformations $t_1^s,\ldots ,t_{k_{\mathrm{soft}}}^s$ . We then compute saliency maps for $x,t_1^s (x),\ldots ,t_{k_{\mathrm{soft}}}^s (x)$ and take their element-wise product to produce a style-invariant saliency map, $G(x)$ . As shown in Figure 3 (part ①), $G(x)$ effectively assigns higher values to the foreground region. + +To distort the obtained region, we first randomly sample an anchor point from it. We then sample $k_{\mathrm{hard}}$ hard transformations $t_1^h, \ldots, t_{k_{\mathrm{hard}}}^h$ , which are typically associated with altering semantic integrity, as highlighted in prior research [1, 9, 18, 23, 33, 38, 40, 62, 63, 66, 77]. We then design a rectangular mask centered on this chosen point, with its width $w$ and height $h$ satisfying $w \sim U(0.05W, 0.3W)$ and $h \sim U(0.05H, 0.3H)$ , where $W$ and $H$ are the image's width and height. Next, we rotate the mask by an angle of $\theta \sim U(-45^\circ, 45^\circ)$ . Finally, we sequentially apply the hard transformations only to the region within this mask, leaving the remaining unmasked areas of the image unchanged. Mathematically, the final anomaly is obtained by applying + +$$ +x ^ {\prime} = t _ {1} ^ {h} \left(\dots \left(t _ {k _ {\text {h a r d}}} ^ {h} \left(x _ {\text {m a s k e d}}\right) \dots\right) + \left(x - x _ {\text {m a s k e d}}\right). \right. \tag {1} +$$ + +An additional advantage of this pseudo-anomaly crafting ap- + +![](images/5c9deb95b88b8d22cd4d1f87d9d1a4fc393f0c62e6fcb33c765505e6e90afbe9.jpg) + +![](images/d5cb6698cfd038056731f52aa802055062cddea6a312e0abf28dd5f8bc324b6b.jpg) +Figure 3. Overview of the PatchGuard framework. (1) We use Grad-CAM to identify regions likely to belong to the foreground in an image. By applying $k_{\mathrm{soft}} = 3$ augmentations to the image $x$ , we generate a combined saliency map $G(x)$ . (2) The input $x$ and its saliency map $G(x)$ are then passed to the anomaly generator to produce an anomaly sample $x'$ along with its ground-truth mask $x_m'$ . (3) For each normal sample $x_i$ , the input batch includes the image, its corresponding anomaly version $x_i'$ , and their adversarially attacked variants, along with their ground-truth masks. Each batch sample is processed through a Vision Transformer (ViT), where the Attention Discriminator assigns an anomaly score $p_i$ to each patch embedding. The final anomaly map is constructed from these scores and trained with a novel loss $\mathcal{L}$ to replicate the ground truth. + +![](images/4d5ed744a1270fd6bb81971e28608eccaf9d6ed5241360a8cd2aa11bc6010e51.jpg) + +Table 1. Detection performance of various AD methods, including PatchGuard, across benchmark datasets. Results are shown as “Clean / Adversarial”, representing performance on “Clean” and “Adversarial” data (PGD-1000, $\epsilon = \frac{8}{255}$ , measured by AUROC %), respectively. + +
DatasetMethod
DRAEMPatchCoreSimpleNetReContrastGeneralADGLADGLASSPrincipaLS*ZARND*RODEO*Ours
MVTec AD98.0 / 5.199.1 / 3.999.6 / 2.399.5 / 0.099.2 / 13.799.3 / 11.899.9 / 6.763.8 / 10.671.6 / 11.061.5 / 4.988.1 / 71.1
VisA88.7 / 3.895.1 / 1.796.2 / 0.297.5 / 6.795.9 / 14.599.5 / 8.698.8 / 7.258.9 / 9.367.7 / 9.363.1 / 3.888.5 / 74.3
BTAD87.1 / 4.996.9 / 2.495.1 / 4.695.7 / 2.197.0 / 10.998.5 / 7.697.4 / 8.360.1 / 7.869.7 / 8.360.0 / 5.885.3 / 82.1
MPDD94.3 / 2.891.3 / 6.896.6 / 7.094.5 / 4.598.0 / 12.897.5 / 9.799.6 / 9.171.9 / 15.765.7 / 12.268.8 / 7.485.4 / 68.6
WFDD94.0 / 7.196.3 / 3.998.8 / 5.197.6 / 3.399.0 / 10.799.1 / 12.5100 / 6.066.3 / 14.176.1 / 16.371.2 / 7.884.2 / 65.0
DTD-Synthetic97.1 / 3.665.8 / 4.096.8 / 2.498.2 / 9.197.9 / 16.398.6 / 10.399.1 / 6.177.9 / 9.370.3 / 13.772.8 / 12.091.5 / 60.5
Head-CT72.1 / 8.192.4 / 5.184.9 / 4.886.1 / 6.389.1 / 9.193.1 / 14.294.7 / 8.773.2 / 16.064.3 / 9.785.7 / 61.097.3 / 90.4
BraTS202162.3 / 1.491.6 / 6.382.5 / 6.182.4 / 3.188.3 / 13.092.3 / 12.095.7 / 7.275.3 / 8.766.7 / 14.089.1 / 64.794.3 / 81.0
Average86.7 / 4.691.0 / 4.293.8 / 4.093.9 / 4.395.5 / 12.697.2 / 10.898.1 / 7.468.4 / 11.469.0 / 11.871.5 / 20.989.3 / 74.1
+ +*These works incorporated adversarial training into their proposed AD methods. + +Table 2. Localization performance of various AL methods, including PatchGuard, across benchmark datasets. Results are shown as "Clean / Adversarial", representing performance on "Clean" and "Adversarial" data (PGD-1000, $\epsilon = \frac{8}{255}$ , measured by AUROC %), respectively. + +
DatasetMethod
DRAEMPatchCoreSimpleNetReContrastGeneralADGLADGLASSMuScReConPatchRealNetOurs
MVTec AD97.3 / 3.998.1 / 4.298.1 / 6.898.4 / 4.897.7 / 12.398.7 / 15.399.3 / 8.197.3 / 4.699.2 / 2.899.0 / 8.792.7 / 73.8
VisA90.7 / 5.698.5 / 3.097.4 / 4.898.2 / 7.199.0 / 7.698.6 / 13.498.8 / 7.098.8 / 3.899.2 / 1.798.8 / 6.296.9 / 85.2
BTAD89.0 / 7.192.6 / 6.395.2 / 3.796.8 / 6.496.7 / 9.398.2 / 14.597.0 / 5.797.3 / 5.397.5 / 3.097.9 / 9.293.2 / 73.0
MPDD90.7 / 4.798.5 / 1.997.4 / 7.394.6 / 3.698.1 / 14.998.7 / 10.899.4 / 6.395.5 / 6.096.7 / 2.698.2 / 7.593.8 / 86.6
WFDD95.8 / 3.998.1 / 2.198.0 / 3.497.3 / 5.198.7 / 13.597.6 / 11.398.9 / 7.697.0 / 4.198.3 / 4.998.4 / 6.894.6 / 71.6
DTD-Synthetic96.1 / 4.895.6 / 6.797.0 / 2.996.8 / 7.697.9 / 10.897.3 / 13.898.1 / 7.097.8 / 5.796.3 / 3.998.0 / 8.395.9 / 74.8
Head-CT86.4 / 8.987.3 / 6.189.7 / 0.890.1 / 8.691.7 / 18.393.4 / 16.792.2 / 9.185.2 / 8.589.7 / 6.694.8 / 7.393.6 / 89.3
BraTS202182.2 / 6.196.9 / 4.594.7 / 6.795.3 / 11.196.8 / 14.795.1 / 13.594.0 / 6.367.5 / 6.397.5 / 5.996.9 / 10.297.7 / 94.5
Average91.0 / 5.695.7 / 4.395.9 / 4.595.9 / 6.797.0 / 12.697.2 / 13.697.2 / 7.192.0 / 5.596.8 / 3.997.7 / 8.094.8 / 81.1
+ +proach is that the location of the distorted region is precisely known, allowing us to generate an accurate localization map $x_{m}^{\prime}$ for AL training, where each pixel's value equals 1 if it's anomalous, and 0 otherwise. By creating an anomalous counterpart $x^{\prime}$ and its corresponding localization map $x_{m}^{\prime}$ for each normal sample $x$ in the training set, we provide the model with both the anomaly and its location, as shown in Figure 3 (part ②). This pairing equips the model with the information needed to improve its localization accuracy, as we will discuss in Section 5.2. For details on the soft and hard transforms, and ablations on the hypers used in this approach, refer to Table 6 and Appendix B. + +# 5.2. Anomaly Detection and Localization with Vision Transformer + +To enhance adversarial robustness through the effective use of pseudo-anomaly samples and their corresponding localization maps, we employ a Vision Transformer (ViT) model. This model generates output embeddings that are processed by an attention-based discriminator network, which uses a multi-head attention (MLH) mechanism followed by a shared-weight multi-layer perceptron (MLP). The discriminator independently assigns an anomaly score to each patch in the input, providing a patch-level anomaly prediction as shown in Figure 3 (part ③). + +The model is trained using a combination of patch-level cross-entropy (CE) and an additional regularization term $R$ , designed to encourage a higher degree of attention in a specified ViT layer. As illustrated in Figure 2 and the theory in Section 4, this regularizer is expected to further enhance adversarial robustness. The cross-entropy loss component aligns the predicted anomaly scores with ground truth labels from the localization map, $x_{m}$ , where a patch label equals 1 if more than $5\%$ of the pixels in the corresponding patch in + +$x_{m}$ are anomalous, and 0 otherwise: + +$$ +M = \operatorname {M L P} (\operatorname {M L H} (x)) \tag {2} +$$ + +$$ +\mathcal {L} _ {\mathrm {C E}} (x, x _ {m}) = \sum_ {i, j \in x _ {\text {p a t c h e s}}} \mathrm {C E} \left(M _ {(i, j)}, x _ {m, (i, j)}\right) \tag {3} +$$ + +Here, $M$ denotes the predicted AL map, consisting of anomaly scores for each patch, $x$ is any input sample (normal or anomaly), and $i,j$ are the indices for individual patches. The cross-entropy loss term $\mathcal{L}_{\mathrm{CE}}$ ensures alignment between the patch $M_{(i,j)}$ and the ground truth anomaly map $x_{m,(i,j)}$ , promoting precise anomaly localization. The final AL map is obtained by upsampling $M$ to the original image size, and we use the average of the top- $k$ patch-level anomaly scores as the image-level anomaly score. + +To further improve the robustness of the model, we introduce a regularization term $R$ , which encourages an increased attention degree in a specified layer $\ell$ of the ViT, chosen here as the final layer. This regularization term is formulated as: + +$$ +R (x) = \frac {1}{\sum_ {i} \sum_ {j} \sum_ {k} A _ {i j k} \cdot \mathbb {1} \left(A _ {i j k} \leq \delta\right)} \tag {4} +$$ + +Here, $A_{ijk}$ represents the attention coefficient corresponding to the $i$ -th head, the $j$ -th output patch, and the $k$ -th input patch. Regularization will increase the values of attention coefficients that are below the threshold $\delta$ . For a specific output patch, since a softmax function is applied to its corresponding attention coefficients, increasing the lower values naturally decreases the higher ones. With $\delta = \frac{1}{(\text{number of input patches})}$ , the coefficient distribution gradually shifts toward a uniform distribution at this threshold, representing the maximum attention degree a patch can achieve. The total loss function $\mathcal{L}$ used for training the model is the + +sum of the cross-entropy and the regularization term: + +$$ +\mathcal {L} \left(x, x _ {m}\right) = \mathcal {L} _ {\mathrm {C E}} \left(x, x _ {m}\right) + \alpha R (x) \tag {5} +$$ + +where $\alpha$ is a scaling factor that controls the strength of the regularization. This combined objective encourages the model to not only accurately predict anomaly locations but also maintain an increased level of attention, enhancing both detection performance and adversarial robustness. It is then used in adversarial training, as explained in Section 5.3. Comprehensive ablation studies of hyperparameters including $\delta$ and $\alpha$ are provided in Appendix C.2 and Section 7. + +# 5.3. Adversarial Training and Testing + +Given an input sample $x$ , an adversarial example $x_{adv}$ is generated by introducing a perturbation $\delta^{*}$ , optimized to maximize the final loss: + +$$ +\delta^ {*} = \underset {\| \delta \| _ {\infty} \leq \epsilon} {\operatorname {a r g m a x}} \mathcal {L} (x + \delta , x _ {m}), \quad x _ {a d v} = x + \delta^ {*} \tag {6} +$$ + +These adversarial examples are then used in the training process alongside the original examples. The adversarial training objective is framed as a min-max problem, optimizing the model parameters $\theta$ to minimize the expected loss over both clean and adversarial examples: + +$$ +\min _ {\theta} \mathbb {E} _ {(x, y) \in \mathcal {B}} \left[ \max _ {\| \delta \| _ {\infty} \leq \epsilon} \mathcal {L} (x + \delta , x _ {m}; \theta) \right]. \tag {7} +$$ + +# 6. Experiments + +To demonstrate the effectiveness of PatchGuard, we conduct extensive experiments assessing its performance alongside several established anomaly detection methods on multiple benchmarks for anomaly detection and localization. Our evaluation includes both adversarially trained models and those trained under standard conditions. To ensure a fair comparison, we assess all methods under both "Clean" and "Adversarial" test conditions, as shown in Table 1 and 2. In the Clean setting, test samples are unaltered, whereas the Adversarial setting includes samples perturbed by the PGD attack [48]. Additional results on PatchGuard's performance against alternative attacks are provided in Section 7.2. Datasets are detailed in Appendix D. + +Evaluation Details. In the adversarial setting, we evaluate each method using the $\ell_{\infty}$ PGD-1000 attack with $\epsilon = \frac{8}{255}$ . Detection and localization results are shown in Tables 1 and 2, respectively. $\ell_{2}$ norm evaluations are in Appendix J. + +Adapting State-of-the-Art AD and AL Methods to Adversarial Settings. Prior to developing our novel Patch-Guard framework, we first investigated whether existing state-of-the-art methods in AD and AL could be enhanced through adversarial training—a technique where models are + +Table 3. Performance of prior methods adapted for adversarial settings, shown as "Clean / Adversarial" (PGD-1000, $\epsilon = \frac{8}{255}$ ). + +
MethodTaskDataset
MVTec ADVisABTADBraTS2021
PatchCoreAD92.3 / 14.388.4 / 10.887.1 / 12.788.6 / 16.1
AL89.1 / 10.387.6 / 8.885.0 / 11.586.7 / 13.6
ReContrastAD93.8 / 5.193.7 / 16.789.9 / 13.478.2 / 13.1
AL94.5 / 17.292.5 / 18.288.6 / 17.990.8 / 14.2
SimpleNetAD91.1 / 18.189.2 / 14.988.6 / 12.775.3 / 14.2
AL94.1 / 19.589.4 / 15.888.2 / 11.790.1 / 17.0
GeneralADAD84.6 / 24.781.2 / 22.182.9 / 20.680.7 / 19.9
AL83.7 / 21.883.1 / 17.384.7 / 18.379.6 / 21.0
DRAEMAD83.0 / 19.281.7 / 16.879.1 / 17.057.4 / 12.2
AL86.1 / 17.578.6 / 15.080.9 / 18.673.6 / 14.9
GLASSAD84.2 / 26.783.5 / 19.984.7 / 25.483.7 / 20.1
AL85.9 / 21.287.1 / 21.783.0 / 19.085.8 / 23.6
OursAD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 / 85.293.2 / 73.097.7 / 94.5
+ +trained on adversarial examples to improve their robustness. In Appendix E, we provide a detailed methodology for generating adversarial examples for each method and describe our optimized approach for improving their resilience. The empirical results, presented in Table 3, reveal that even after incorporating adversarial training, current SOTA AD and AL methods remain vulnerable to adversarial attacks and exhibit suboptimal performance. These limitations motivated the development of PatchGuard, our proposed solution. + +Analyzing Results. As shown in Tables 1 and 2, prior SOTA methods such as Recontrast demonstrate substantial drops in performance under adversarial settings, despite excelling in clean conditions. Furthermore, as indicated in Table 3, even when adversarily trained, these methods struggle to achieve robust performance. By contrast, approaches specifically designed to address adversarial robustness, such as RODEO, still fall short when compared to PatchGuard. On average, PatchGuard improves robust detection across diverse datasets by up to $53.2\%$ percent, with a notable robust improvement of $68.5\%$ percent on the challenging MVtec dataset. Notably, we achieve a $71.8\%$ improvement in AL on the challenging VisA dataset, with details in Appendix F. As prior research indicates, a slight decrease in clean performance is generally viewed as an acceptable tradeoff for the critical gains in robustness. + +Implementation Details. We leverage a from-scratch ViT model in our pipeline to showcase our setup's independence from pre-train datasets. However, we conduct an ablation study on this manner in Appendix C.4. Further, we leverage PGD-10 with $\epsilon = \frac{8}{255}$ for adversarial training, + +Table 4. Evaluation of PatchGuard's performance against advanced classification and segmentation attacks, showcasing its robustness in diverse attack scenarios. + +
DatasetTaskDetection AttacksLocalization Attacks
PGD-1000CAAAutoAttackA3SegPGDSEA
MVTec ADAD71.174.870.869.871.969.2
AL73.876.772.970.074.671.3
VisAAD74.378.574.673.175.073.6
AL85.287.485.584.785.083.7
BTADAD82.183.482.080.781.980.8
AL73.074.772.371.572.971.1
BraTS2021AD81.084.781.980.382.081.1
AL94.595.094.293.894.893.4
+ +Table 5. Ablation study on the scaling factor $\alpha$ , illustrating the effectiveness of our regularization. + +
Scaling Factor (α)TaskDataset
MVTec-ADVisABTADBraTS2021
0AD89.3 / 60.790.1 / 62.988.6 / 74.395.2 / 72.0
AL93.5 / 65.395.8 / 72.194.0 / 59.798.2 / 83.4
1 (Ours)AD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 / 85.293.2 / 73.097.7 / 94.5
2AD86.2 / 70.187.6 / 75.184.7 / 83.492.8 / 81.2
AL91.0 / 68.994.8 / 84.191.9 / 73.995.8 / 94.6
5AD80.1 / 64.178.6 / 65.473.8 / 69.682.5 / 73.2
AL84.7 / 65.681.1 / 78.984.0 / 65.386.3 / 79.9
+ +but the test results are generated with the PGD-1000 attack. For more details on the implementation parameters, refer to Appendix G. + +# 7. Ablation Studies + +In this section, we provided further analysis of our choices of parameters and other components of our method. + +# 7.1. Regularization + +To experimentally demonstrate the effectiveness of our regularization term on the model's robustness and the impact of the scaling factor $\alpha$ (set to 1 by default in our method), we conduct an ablation study on $\alpha$ , as shown in Table 5. When $\alpha = 0$ , the model's robustness against adversarial attacks decreases due to the lack of regularization. On the other hand, higher values of $\alpha$ , such as $\alpha = 5$ , cause the regularization term to dominate, leading to an unstable loss function and inappropriate results. + +# 7.2. Advanced Attacks + +In our primary experiments in Section 6, we utilized variations of the PGD attack [48] for both model training and + +Table 6. An ablation study on our method's performance using various anomaly generation techniques that incorporate masks. + +
Pseudo-Anomaly GeneratorTaskDataset
MVTec ADVisABTADBraTS2021
CutPasteAD79.1 / 60.374.4 / 61.574.8 / 69.280.3 / 72.7
AL83.9 / 60.885.4 / 73.082.6 / 61.291.2 / 82.1
FPIAD78.9 / 62.175.1 / 64.372.9 / 71.084.1 / 74.1
AL84.3 / 63.882.9 / 71.886.3 / 60.593.3 / 84.8
NSAAD81.7 / 66.583.1 / 65.477.1 / 73.486.3 / 73.8
AL85.1 / 65.083.3 / 76.785.4 / 64.989.4 / 86.8
GLASSAD83.2 / 62.084.0 / 68.779.3 / 73.983.1 / 66.1
AL86.7 / 67.086.2 / 78.586.8 / 68.292.2 / 87.3
OursAD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 /85.293.2 /73.097.7 /94.5
+ +testing. To further illustrate the adaptability and robustness of our proposed method under diverse adversarial conditions, we also evaluate its performance against a variety of other attack types including Detection attacks and Localization attacks. The Detection attacks consist of black-box attacks [28], FGSM [24], CAA [49], AutoAttack [16], and $\mathrm{A}^3$ (Adversarial Attack Automation) [43]. The localization attacks are adapted from the well-known segmentation attacks in literature, namely SegPGD [27] and SEA [17]. All results are available in Table 4, and the exact details on our adaptation approach for these attacks are available in Appendix H. It is important to note that the training process remains consistent, employing only the standard PGD-10 for simplicity and practicality. + +# 7.3. Pseudo-Anomaly Generation Strategy + +Since our method targets both AD and AL using pseudo-anomalies, it is crucial to evaluate our pseudo-anomaly generation against compatible methods that produce localization maps. Thus, methods like MIXUP [33], FakeIt [50], and Dream-OOD [20] are excluded. We compare instead with CutPaste [40], NSA [59], GLASS [12], and FPI [67], keeping other pipeline components fixed. As shown in Table 6, our method outperforms these approaches, highlighting its effectiveness for robust anomaly localization. + +# 8. Conclusion + +We introduce PatchGuard, an adversarially robust method for AD and AL. It employs Foreground-Aware Pseudo-Anomaly Generation using segmentation masks. Additionally, we propose a regularization term, motivated by observations and theory, that strengthens robustness against adversarial attacks by increasing the attention degree. PatchGuard outperforms previous methods on AD and AL benchmarks under adversarial settings. + +# Acknowledgments + +We are very grateful to Mohammad Javad Maheronnaghsh for his collaboration in the initial stages of this project and to the anonymous reviewers for their helpful discussions and valuable feedback. + +# References + +[1] M. Eren Akbiyik. Data augmentation in training cnns: Injecting noise to images. ArXiv, abs/2307.06855, 2019. 4 +[2] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. *IEEE Access*, 6:14410–14430, 2018. 1 +[3] Mohammad Azizmalayeri, Arshia Soltani Moakhar, Arman Zarei, Reihaneh Zohrabi, Mohammad Manzuri, and Mohammad Hossein Rohhan. Your out-of-distribution detection method is not robust! Advances in Neural Information Processing Systems, 35:4887-4901, 2022. 1, 2, 4 +[4] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 2 +[5] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1893-1902, 2015. 2 +[6] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 2 +[7] Louis Béthane, Paul Novello, Thibaut Boissin, Guillaume Coiffier, Mathieu Serrurier, Quentin Vincenot, and Andres Troya-Galvis. Robust one-class classification with signed distance function using 1-lipschitz neural networks. arXiv preprint arXiv:2303.01978, 2023. 1 +[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 4 +[9] Chengwei Chen, Yuan Xie, Shaohui Lin, Ruizhi Qiao, Jian Zhou, Xin Tan, Yi Zhang, and Lizhuang Ma. Novelty detection via contrastive learning with negative data augmentation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI 2021, Virtual Event / Montreal, Canada, 19-27 August 2021, pages 606-614. ijcai.org, 2021. 4 +[10] Jiefeng Chen, Yixuan Li, Xi Wu, Yingyu Liang, and Somesh Jha. Robust out-of-distribution detection for neural networks. arXiv preprint arXiv:2003.09711, 2020. 1, 2, 4 +[11] Jiefeng Chen, Yixuan Li, Xi Wu, Yingyu Liang, and Somesh Jha. Atom: Robustifying out-of-distribution detection using + +outlier mining. In Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13-17, 2021, Proceedings, Part III 21, pages 430-445. Springer, 2021. 2, 4 + +[12] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. arXiv preprint arXiv:2407.09359, 2024. 1, 3, 8 + +[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4 + +[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 4 +[15] Matan Jacob Cohen and Shai Avidan. Transformally-two (feature spaces) are better than one. arXiv preprint arXiv:2112.04185, 2021. 2, 4 +[16] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 8 +[17] Francesco Croce, Naman D Singh, and Matthias Hein. Towards reliable evaluation and fast training of robust semantic segmentation models. In ECCV, 2024. 2, 3, 8 +[18] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 4 +[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[20] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. arXiv preprint arXiv:2309.13415, 2023. 8 +[21] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. 3 +[22] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1 +[23] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. arXiv preprint arXiv:2012.07177, 2020. 4 +[24] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 8 +[25] Adam Goodge, Bryan Hooi, See Kiong Ng, and Wee Siong Ng. Robustness of autoencoders for anomaly detection under adversarial impact. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 1244-1250, 2021. 1 + +[26] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4 +[27] Jindong Gu, Hengshuang Zhao, Volker Tresp, and Philip HS Torr. Segpgd: An effective and efficient adversarial attack for evaluating and boosting segmentation robustness. In European Conference on Computer Vision, pages 308-325. Springer, 2022. 8 +[28] Chuan Guo, Jacob Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Weinberger. Simple black-box adversarial attacks. In International Conference on Machine Learning, pages 2484-2493. PMLR, 2019. 8 +[29] Jia Guo, Lize Jia, Weihang Zhang, Huiqi Li, et al. Recontrast: Domain-specific anomaly detection via contrastive reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 2 +[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4 +[31] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4 +[32] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606, 2018. 2 +[33] Yann N. Dauphin Hongyi Zhang, Moustapha Cisse. mixup: Beyond empirical risk minimization. International Conference on Learning Representations, 2018. 4, 8 +[34] Mohammad Hossein Rohban Hossein Mirzaei, ... Scanning trojaned models using out-of-distribution samples. In Advances in Neural Information Processing Systems, 2024. 3 +[35] Jiande Huang, Xingxing Li, Xin Li, Jiaqi Wu, Keke Zhang, Yongqiang Yuan, and Wei Zhang. Real-time outlier detection of satellite orbit and clock products using reverse error estimation. GPS Solutions, 29(1):2, 2025. 2 +[36] Nicholas Jeffrey, Qing Tan, and José R Villar. A review of anomaly detection strategies to detect threats to cyberphysical systems. *Electronics*, 12(15):3283, 2023. 1 +[37] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2, 4 +[38] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020. 4 +[39] Jungbeom Lee, Eunj Kim, Jisoo Mok, and Sungroh Yoon. Anti-adversarily manipulated attributions for weakly supervised semantic segmentation and object localization. IEEE transactions on pattern analysis and machine intelligence, 46 (3):1618-1634, 2022. 2 + +[40] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9664-9674, 2021. 3, 4, 8 +[41] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. Musc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. arXiv preprint arXiv:2401.16753, 2024. 2, 4 +[42] Jiaqi Liu, Guoyang Xie, Jinbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. Machine Intelligence Research, 21(1):104-135, 2024. 1 +[43] Ye Liu, Yaya Cheng, Lianli Gao, Xianglong Liu, Qilong Zhang, and Jingkuan Song. Practical evaluation of adversarial robustness via adaptive auto attack, 2022. 2, 8 +[44] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 3 +[45] Shao-Yuan Lo, Poojan Oza, and Vishal M Patel. Adversarily robust one-class novelty detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3 +[46] Pauline Luc, Camille Couprie, Soumith Chintala, and Jakob Verbeek. Semantic segmentation using adversarial networks. arXiv preprint arXiv:1611.08408, 2016. 2 +[47] Yuan Luo, Ya Xiao, Long Cheng, Guojun Peng, and Danfeng Yao. Deep learning-based anomaly detection in cyberphysical systems: Progress and opportunities. ACM Computing Surveys (CSUR), 54(5):1-36, 2021. 1 +[48] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2, 3, 7, 8 +[49] Xiaofeng Mao, Yuefeng Chen, Shuhui Wang, Hang Su, Yuan He, and Hui Xue. Composite adversarial attacks. ArXiv, abs/2012.05434, 2020. 8 +[50] Hossein Mirzaei, Mohammadreza Salehi, Sajjad Shahabi, Efstratios Gavves, Cees GM Snoek, Mohammad Sabokrou, and Mohammad Hossein Rohban. Fake it till you make it: Near-distribution novelty detection by score-based generative models. arXiv preprint arXiv:2205.14297, 2022. 8 +[51] Hossein Mirzaei, Mohammad Jafari, Hamid Reza Dehbashi, Ali Ansari, Sepehr Ghobadi, Masoud Hadi, Arshia Soltani Moakhar, Mohammad Azizmalayeri, Mahdieh Soleymani Baghshah, and Mohammad Hossein Rohban. Rodeo: Robust outlier detection via exposing adaptive out-of-distribution samples. In Forty-first International Conference on Machine Learning, 2024. 1, 2, 3, 4 +[52] Hossein Mirzaei, Mohammad Jafari, Hamid Reza Dehbashi, Zeinab Sadat Taghavi, Mohammad Sabokrou, and Mohammad Hossein Rohban. Killing it with zero-shot: Adversarially robust novelty detection. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7415-7419. IEEE, 2024. 3 +[53] Hossein Mirzaei, Mojtaba Nafez, Mohammad Jafari, Mohammad Bagher Soltani, Mohammad Azizmalayeri, Jafar Habibi, + +Mohammad Sabokrou, and Mohammad Hossein Rohban. Universal novelty detection through adaptive contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22914-22923, 2024. 4 +[54] Arian Mousakhan, Thomas Brox, and Jawad Tayyyub. Anomaly detection with conditioned denoising diffusion models. arXiv preprint arXiv:2305.15956, 2023. 1 +[55] Fahimeh Hosseini Noohdani, Parsa Hosseini, Aryan Yazdan Parast, Hamidreza Yaghoubi Araghi, and Mahdieh Soleymani Baghshah. Decompose-and-compose: A compositional approach to mitigating spurious correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27662-27671, 2024. 2, 4 +[56] Pramuditha Perera, Poojan Oza, and Vishal M Patel. One-class classification: A survey. arXiv preprint arXiv:2101.03064, 2021. 2 +[57] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection, 2021. 2 +[58] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051, 2021. 2 +[59] Hannah M Schlüter, Jeremy Tan, Benjamin Hou, and Bernhard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In European Conference on Computer Vision, pages 474-489. Springer, 2022. 3, 8 +[60] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 4 +[61] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16, pages 682-698. Springer, 2020. 1 +[62] Abhishek Sinha, Kumar Ayush, Jiaming Song, Burak Uzkent, Hongxia Jin, and Stefano Ermon. Negative data augmentation. In International Conference on Learning Representations, 2021. 4 +[63] Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister. Learning and evaluating representations for deep one-class classification. arXiv preprint arXiv:2011.02578, 2020. 4 +[64] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Generalad: Anomaly detection across domains by attending to distorted features. arXiv preprint arXiv:2407.12427, 2024. 1, 2, 3, 4 +[65] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1 +[66] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distribu + +tionally shifted instances. Advances in neural information processing systems, 33:11839-11852, 2020. 1, 4 +[67] Jeremy Tan, Benjamin Hou, James Batten, Huaqi Qiu, and Bernhard Kainz. Detecting outliers with foreign patch interpolation. Machine Learning for Biomedical Imaging, 1(April 2022):1-27, 2022. 8 +[68] Maximilian E Tschuchnig and Michael Gadermayr. Anomaly detection in medical imaging-a mini review. In Data Science-Analysics and Applications: Proceedings of the 4th International Data Science Conference-iDSC2021, pages 33-38. Springer, 2022. 1 +[69] Qi Wang, Junyu Gao, and Xuelong Li. Weakly supervised adversarial domain adaptation for semantic segmentation in urban scenes. IEEE Transactions on Image Processing, 28 (9):4376-4386, 2019. 2 +[70] Weiping Wang, Zhaorong Wang, Zhanfan Zhou, Haixia Deng, Weiliang Zhao, Chunyang Wang, and Yongzhen Guo. Anomaly detection of industrial control systems based on transfer learning. Tsinghua Science and Technology, 26(6): 821-832, 2021. 1 +[71] Yue Xing, Qifan Song, and Guang Cheng. Why do artificially generated data help adversarial robustness. Advances in Neural Information Processing Systems, 35:954-966, 2022. 2 +[72] Han Xu, Yao Ma, Haochen Liu, Debayan Deb, Hui Liu, Jiliang Tang, and Anil K. Jain. Adversarial attacks and defenses in images, graphs and text: A review, 2019. 3 +[73] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, pages 1-28, 2024. 1 +[74] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. arXiv preprint arXiv:2406.07487, 2024. 1 +[75] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 3 +[76] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE transactions on neural networks and learning systems, 30(9): 2805-2824, 2019. 1, 3 +[77] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features, 2019. 4 +[78] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 3 +[79] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Dsr – a dual subspace re-projection network for surface anomaly detection, 2022. 3 +[80] Chaoning Zhang, Philipp Benz, Chenguo Lin, Adil Karjauv, Jing Wu, and In So Kweon. A survey on universal adversarial attack. arXiv preprint arXiv:2103.01498, 2021. 1 + +[81] Hui Zhang, Zheng Wang, Zuxuan Wu, and Yu-Gang Jiang. Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection. arXiv preprint arXiv:2303.08730, 2023. 1 +[82] Lin Zhang and Yang Dai. A gan anomaly detection method based on multi-scale endogenous enhancement. In Blockchain and Web3 Technology Innovation and Application Exchange Conference, pages 269-281. Springer, 2024. 2 +[83] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 1, 2 \ No newline at end of file diff --git a/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/images.zip b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3742d40ec318568997d2f0660fac0883b654910a --- /dev/null +++ b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f78abf640a444d37c7f41588f12784c747272e432e125214a3c9e6d217b0d431 +size 533522 diff --git a/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/layout.json b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f090e660a7127a932b9b3c1b97be3da7626e8b1a --- /dev/null +++ b/2025/PatchGuard_ Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies/layout.json @@ -0,0 +1,10395 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 138 + ], + "type": "text", + "content": "PatchGuard: Adversarily Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 160, + 462, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 160, + 462, + 204 + ], + "spans": [ + { + "bbox": [ + 151, + 160, + 462, + 204 + ], + "type": "text", + "content": "Mojtaba Nafez1 Amirhossein Koochakian1 * Arad Maleki1 * Jafar Habibi1 Mohammad Hossein Rohban1 1Sharif University of Technology, Iran" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 206, + 470, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 206, + 470, + 217 + ], + "spans": [ + { + "bbox": [ + 138, + 206, + 470, + 217 + ], + "type": "text", + "content": "{mojtaba.nafez77, amir.kochakian12, arad.maleki02, rohban}@sharif.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "text", + "content": "Anomaly Detection (AD) and Anomaly Localization (AL) are crucial in fields that demand high reliability, such as medical imaging and industrial monitoring. However, current AD and AL approaches are often susceptible to adversarial attacks due to limitations in training data, which typically include only normal, unlabeled samples. This study introduces PatchGuard, an adversarily robust AD and AL method that incorporates pseudo anomalies with localization masks within a Vision Transformer (ViT)-based architecture to address these vulnerabilities. We begin by examining the essential properties of pseudo anomalies, and follow it by providing theoretical insights into the attention mechanisms required to enhance the adversarial robustness of AD and AL systems. We then present our approach, which leverages Foreground-Aware Pseudo-Anomalies to overcome the deficiencies of previous anomaly-aware methods. Our method incorporates these crafted pseudo-anomaly samples into a ViT-based framework, with adversarial training guided by a novel loss function designed to improve model robustness, as supported by our theoretical analysis. Experimental results on well-established industrial and medical datasets demonstrate that PatchGuard significantly outperforms previous methods in adversarial settings, achieving performance gains of " + }, + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "inline_equation", + "content": "53.2\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "text", + "content": " in AD and " + }, + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "text", + "content": " in AL, while also maintaining competitive accuracy in non-adversarial settings. The code repository is available at here." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "content": "In recent years, Anomaly Detection (AD) and Anomaly Localization (AL) have become critical techniques across various fields, including medical imaging, fraud detection, and industrial monitoring [22, 36, 42, 47, 68, 70, 73]. While state-of-the-art (SOTA) AD and AL methods exhibit near-perfect performance on standard benchmarks [6, 83], they" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 243, + 555, + 343 + ], + "blocks": [ + { + "bbox": [ + 314, + 243, + 555, + 343 + ], + "lines": [ + { + "bbox": [ + 314, + 243, + 555, + 343 + ], + "spans": [ + { + "bbox": [ + 314, + 243, + 555, + 343 + ], + "type": "image", + "image_path": "f77e3f787947d6371e0f60b757d73fc6f18f6d3ff42807f4c70ce0a0b193ecb8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 352, + 556, + 419 + ], + "lines": [ + { + "bbox": [ + 313, + 352, + 556, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 556, + 419 + ], + "type": "text", + "content": "Figure 1. Impact of Adversarial Attacks on Anomaly Localization Methods: Localization maps for multiple methods are shown before and after a PGD-1000 attack, illustrating the vulnerability of existing methods to adversarial attacks, even when performing perfectly in clean conditions. Our proposed method demonstrates enhanced robustness in these adversarial scenarios." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 441, + 556, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 556, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 556, + 620 + ], + "type": "text", + "content": "frequently face significant performance degradation when subjected to adversarial attacks, as illustrated in Figure 1 [12, 54, 64, 66, 74, 81]. These attacks involve introducing subtle, imperceptible perturbations into the input data, misleading models to misclassify normal1 samples as anomalies or to overlook actual anomalies [2, 24, 48, 65, 80]. To address this vulnerability, a variety of strategies have been proposed to strengthen the adversarial robustness of AD systems [3, 7, 10, 25, 45, 61]. However, despite these advancements, the resilience of these methods remains limited, with even mild adversarial attacks still causing noticeable performance drops, even on widely used anomaly benchmarks [6, 83]. Furthermore, adversarial robustness in AL remains an unexplored area, highlighting a critical gap in the current research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 622, + 556, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 556, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 556, + 682 + ], + "type": "text", + "content": "Adversarial training, a strategy designed to improve model robustness in classification tasks, incorporates adversarial samples into the training process [48, 76]. Despite its success in standard classification, this approach proves less effective in AD and AL scenarios [51]. The primary" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "text", + "content": "1In this study, the term 'normal' refers to inlier samples, not to be mistaken with Gaussian distribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 135, + 712 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20383" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "type": "text", + "content": "obstacle to achieving adversarial robustness in AD lies in the absence of anomalous samples during training [10, 11]. Without such samples, models cannot be trained on adversarial perturbations specific to anomalous data [5, 51, 56, 58], leaving them vulnerable to attacks on perturbed anomalies [3, 10, 11]. To address this issue in the AD setup, recent efforts have combined Outlier Exposure (OE) [32] with adversarial training [3, 10, 11]. However, this method struggles when applied to high-resolution, real-world datasets [51]. This challenge is even more pronounced in AL settings, as robust AL methods must perform pixel-level classification, which inherently requires pixel-level supervision, while at the same time operating within the same data constraints as the AD framework [35, 82]. Despite ongoing research, the adversarial robustness of AL methods remains an open problem, with no comprehensive solutions proposed to date." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 266, + 297, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 297, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 297, + 434 + ], + "type": "text", + "content": "In this study, we identify two essential properties that OE techniques must satisfy for effective application in adversarially robust AD and AL. First, the outlier exposures must include localization maps to facilitate the training of robust AL models. This requirement aligns with prior research on adversarially robust segmentation models, which underscores the importance of generating localization maps to enhance model robustness [39, 46, 69]. Second, as noted by [51], OE samples should be \"near-distributed\" relative to normal data, meaning they should share similar semantic and stylistic attributes. This condition is consistent with the literature on adversarial robustness, which highlights the advantages of near-distribution decision boundaries samples for optimizing the model robustness [71]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 436, + 297, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 297, + 590 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 297, + 590 + ], + "type": "text", + "content": "Moreover, the recent success of Vision Transformer (ViT)-based architectures in clean AD and AL tasks [15, 37, 41, 64], attributed to their global contextual modeling [19], motivated us to explore their potential in adversarial settings. For an input image " + }, + { + "bbox": [ + 54, + 436, + 297, + 590 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 436, + 297, + 590 + ], + "type": "text", + "content": " and a fixed attention layer " + }, + { + "bbox": [ + 54, + 436, + 297, + 590 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 54, + 436, + 297, + 590 + ], + "type": "text", + "content": ", we define the Attention Degree of " + }, + { + "bbox": [ + 54, + 436, + 297, + 590 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 54, + 436, + 297, + 590 + ], + "type": "text", + "content": " under " + }, + { + "bbox": [ + 54, + 436, + 297, + 590 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 436, + 297, + 590 + ], + "type": "text", + "content": " as the average number of input embeddings each output embedding attends to. As shown in Figure 2, we observe that images that induce higher attention degrees in the model are more resistant to adversarial attacks. Furthermore, adversarial training appears to implicitly steer the model toward achieving a high attention degree. We provide a theoretical analysis and justification for these findings in Section 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": "Building on these insights, we propose PatchGuard, an adversarially robust AD and AL method that addresses limitations in current OE-based methods and leverages attention-based insights. To address OE challenges, we generate pseudo-anomaly samples from normal data, ensuring they meet two essential criteria: availability of localization maps and near-distribution alignment with inliers. To create samples that exhibit anomalous characteristics while remaining close to the inlier distribution, we introduce Foreground-Aware Pseudo-Anomaly Generation. This approach iden" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": "tifies causal regions in normal samples using an enhanced Grad-CAM [60] feature attribution technique with tailored augmentations. These identified regions are then selectively distorted, drawing on insights from [55], to produce anomaly samples that retain stylistic and semantic coherence with the original data. By precisely controlling the distortion location, we can generate accurate localization maps for the pseudoanomalies, effectively meeting both criteria for robust AD and AL." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 556, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 556, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 556, + 263 + ], + "type": "text", + "content": "In the next step, we leverage normal samples and pseudoanomalies to perform adversarial training on a ViT-based architecture. Building on the provided insights, we introduce a novel loss function designed to increase the model's last-layer attention degree. As our theoretical justifications suggest, this property significantly enhances the model's adversarial robustness in both AD and AL settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 264, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 556, + 515 + ], + "type": "text", + "content": "Contribution. In this paper, we propose PatchGuard, an adversarially robust AD and AL method. To the best of our knowledge, we are the first to raise and effectively address the issue of adversarial robustness in AL. We provide insights into the shortcomings of previous OE-based methods and introduce the Foreground-Aware Pseudo-Anomaly Generation framework to address these limitations. Additionally, we provide insights and theoretical justifications on enhancing the adversarial robustness of ViT-based architectures, and propose a novel loss function aligned with these findings. We evaluate PatchGuard in both clean and adversarial settings, rigorously testing its robustness under several strong attacks, including PGD-1000 [48], " + }, + { + "bbox": [ + 313, + 264, + 556, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^3" + }, + { + "bbox": [ + 313, + 264, + 556, + 515 + ], + "type": "text", + "content": " [43], and SEA [17]. Results demonstrate that PatchGuard significantly outperforms existing methods in adversarial contexts, achieving an AUROC improvement of " + }, + { + "bbox": [ + 313, + 264, + 556, + 515 + ], + "type": "inline_equation", + "content": "53.2\\%" + }, + { + "bbox": [ + 313, + 264, + 556, + 515 + ], + "type": "text", + "content": " in AD and " + }, + { + "bbox": [ + 313, + 264, + 556, + 515 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 313, + 264, + 556, + 515 + ], + "type": "text", + "content": " in AL, while maintaining competitive performance in standard conditions. Our evaluations span various well-established industrial and medical datasets, including MVTec [6], VisA [83], and BraTS2021 [4], highlighting PatchGuard's applicability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 526, + 402, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 526, + 402, + 539 + ], + "spans": [ + { + "bbox": [ + 314, + 526, + 402, + 539 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "From an adversarial robustness perspective, anomaly detection and localization methods can be mainly classified into three categories: 1) methods with anomaly-free training, 2) methods that leverage anomalies at the embedding level, and 3) methods that leverage anomalies at the input level. We experimentally adapted methods from each category to adversarial training processes, and the results reveal they still exhibit significant vulnerabilities even after adversarial training. The first category consists of methods such as Recontrast[29], Transformaly [15], and PatchCore [57], that operate on the hypothesis that models trained exclusively on normal samples can reconstruct normal images well but fail to do so with abnormal regions. These methods lack robustness, since the model does not encounter any anomaly" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20384" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 396 + ], + "type": "text", + "content": "data during adversarial training, meaning that when adversarial perturbations are added to anomaly data, the model still remains vulnerable. The second category, which includes methods such as SimpleNet [44], GeneralAD [64], UniAD [75], and DSR [79], utilize a feature extractor during training to obtain feature vectors of the normal set. By perturbing these features, they create anomaly features and primarily train a discriminator over these normal and anomaly features. However, these methods lack robustness due to their reliance on highly vulnerable feature extractors and the absence of anomaly data in the input space, which prevents proper training of a robust feature extractor. The third category consists of DRAEM [78], NSA [59], Cutpaste [40], and GLASS [12], generate pseudo-anomaly data during training. Nevertheless, some of these methods lack access to masks or do not use a foreground-aware generation process. Despite their astonishing success in clean settings, we empirically show in Table 3 that even when adversarily trained, they exhibit unsatisfactory results. Furthermore, while anomaly localization has not been explored in an adversarial setup, some other works have pursued anomaly detection in an adversarial setting, including PrincipaLS [45], ZARND [52], and Rodeo [51]. These approaches have achieved relatively better results by incorporating pseudo-anomaly generation processes and adversarial training. However, their performance falls short in real-world high-resolution datasets (as presented in Table 1). For more details about previous works, see Appendix A." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 406, + 140, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 140, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 140, + 418 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "text", + "content": "Anomaly Detection (AD) and Localization (AL): AD involves identifying instances that deviate from expected patterns, while AL focuses on localizing abnormal subregions within an image. In AD, a model assigns an anomaly score " + }, + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "inline_equation", + "content": "S(x; f)" + }, + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "text", + "content": " to each sample " + }, + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "text", + "content": " using model " + }, + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "text", + "content": "; if this score exceeds a predefined threshold, the sample is classified as anomalous. Conversely, lower scores indicate normal samples. In AL, a model generates an anomaly map " + }, + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "inline_equation", + "content": "M(x; f)" + }, + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "text", + "content": " using model " + }, + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 426, + 296, + 545 + ], + "type": "text", + "content": ", where each pixel of the map corresponds to an anomaly score for the respective pixel in the input image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": "Adversarial Robustness of Anomaly Detectors: An adversarial attack, primarily studied within classification tasks, involves the malicious alteration of a data sample " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": " with its label " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": " into a new sample " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": " that maximizes the loss function " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\ell(x^{*};y)" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": " [34, 72, 76]. This new sample " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": " is known as an adversarial example of " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": ", and the difference " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "(x^{*} - x)" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": " is referred to as adversarial perturbation. To prevent significant semantic changes, the " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": " norm of the adversarial noise is restricted by an upper limit " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": ". Specifically, an adversarial example " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": " must satisfy the condition " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "x^{*} = \\arg \\max_{x^{\\prime}:\\|x - x^{\\prime}\\|_{p} \\leq \\epsilon} \\ell(x^{\\prime};y)" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": ". A widely used and effective attack method is the Projected Gradient Descent (PGD) [48] technique, which iteratively increases the loss function by moving in the direction of the gradient sign of" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "\\ell(x^{*};y)" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " with a specified step size " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ". When adapting adversarial attacks for AD, the objective shifts from maximizing the loss value to increasing " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "S(x,f)" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " for normal samples and decreasing it for anomalies. The attack is formulated as " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "x_0^* = x" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "x_{t+1}^* = x_t^* + y \\cdot \\alpha \\cdot \\mathrm{sign}(\\nabla_x S(x_t^*;f_\\theta))" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "x^* = x_k^*" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "y = +1" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " for normal samples and " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "y = -1" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " for anomaly samples. This approach is consistently applied to other attacks in the study [21, 76]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "text", + "content": "Adversarial Robustness of Anomaly Localization: Adversarial attacks aim to introduce perturbations that uniformly modify pixel scores in an anomaly map " + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "text", + "content": ", flipping scores from 0 to 1 or vice versa. In the context of AL, the attack objective transitions from merely maximizing a loss function to specifically increasing the values of " + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "inline_equation", + "content": "M(x, f)" + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "text", + "content": " for normal samples while decreasing them for anomalies. The attack is formalized as follows: starting with " + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "inline_equation", + "content": "x_0^* = x" + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "text", + "content": ", the adversarial example at each step is updated as " + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "inline_equation", + "content": "x_{t+1}^* = x_t^* + Y \\cdot \\alpha \\cdot \\mathrm{sign}(\\nabla_x M(x_t^*; f_\\theta))" + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "text", + "content": ", and the final adversarial example is " + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "inline_equation", + "content": "x^* = x_k^*" + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "inline_equation", + "content": "Y_{i,j} = +1" + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "text", + "content": " for normal pixels and " + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "inline_equation", + "content": "Y_{i,j} = -1" + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "text", + "content": " for anomaly pixels, where " + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 313, + 168, + 556, + 384 + ], + "type": "text", + "content": " denote pixel coordinates. This strategy is extended to other advanced attack methods throughout the study [17]. Despite its significance, adversarial robustness in anomaly localization remains underexplored. Enhancing robustness is particularly critical for applications requiring spatial precision, such as medical diagnostics and industrial monitoring systems." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 396, + 432, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 432, + 411 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 432, + 411 + ], + "type": "text", + "content": "4. Theoretical Insights" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 430, + 432, + 521 + ], + "blocks": [ + { + "bbox": [ + 316, + 430, + 432, + 521 + ], + "lines": [ + { + "bbox": [ + 316, + 430, + 432, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 432, + 521 + ], + "type": "image", + "image_path": "79c34287f45507354ea2031fb5ff756f66282d5125e01d025f8cd07a61938e5a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 531, + 556, + 641 + ], + "lines": [ + { + "bbox": [ + 313, + 531, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 556, + 641 + ], + "type": "text", + "content": "Figure 2. The figure demonstrates how increasing the average last-layer attention degree in the ViT-base architecture reduces vulnerability to adversarial attacks. Specifically, the BraTS dataset images were clustered into five groups based on their attention degree values (x-axis). The y-axis represents Vulnerability, measured as the absolute difference in the model's localization performance (AUROC%) for each cluster before and after a PGD attack. (a) illustrates the decreasing trend in a clean-trained model. (b) shows a similar decreasing trend in the adversarially trained model, where the attention degree is relatively higher than in the clean model." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 437, + 429, + 553, + 521 + ], + "blocks": [ + { + "bbox": [ + 437, + 429, + 553, + 521 + ], + "lines": [ + { + "bbox": [ + 437, + 429, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 437, + 429, + 553, + 521 + ], + "type": "image", + "image_path": "a0c1739ded82b78932cbf62868f2289f6788980eaa63019daa71bbb90600b27e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Here, we want to show why a self-attention layer, in which each token attends to a higher number of other tokens is more adversarially robust compared to the case that the attention is distributed over a smaller number of tokens." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20385" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "text", + "content": "Let's consider the self-attention output embedding of the " + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "text", + "content": "th token in the " + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "text", + "content": "-th layer as " + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "inline_equation", + "content": "x_{i}^{(l)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "text", + "content": ", and such embedding after the perturbation is applied to the input as " + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^{(l)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "text", + "content": ". We are seeking to set " + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "inline_equation", + "content": "x_{i}^{(2)} - \\hat{x}_{i}^{(2)} = r_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 150 + ], + "type": "text", + "content": " is the desired adversarial direction that most influences the layers next to the first layer of the transformer. According to the self-attention:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 95, + 159, + 254, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 159, + 254, + 175 + ], + "spans": [ + { + "bbox": [ + 95, + 159, + 254, + 175 + ], + "type": "interline_equation", + "content": "x _ {i} ^ {(l + 1)} = V ^ {(l)}. \\text {s o f t m a x} \\left(q _ {i} ^ {(l) \\top} K / \\sqrt {d}\\right),", + "image_path": "77bd561197663f0cb7c81dbf7032621712d556d5a1fe77e228e36aa460848022.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "q_{i}^{(l)} = W_{q}x_{i}^{(l)},v_{j}^{(l)}" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": " the " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": "-th column of " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "V^{(l)}" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "v_{j}^{(l)} = W_{v}x_{j}^{(l)},k_{j}^{(l)}" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": " the " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": "-th column of " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": " be " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "k_{j}^{(l)} = W_{k}x_{j}^{(l)}" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": " be the embedding dimension. For simplicity, we assume that the dimensions of embeddings of key, query, and values are all the same and are equal to the dimension of the original embeddings. Now, " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "W_{v}(X^{(1)}S - \\hat{X}^{(1)}\\hat{S}) = R" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": ", where the columns of " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "X^{(l)}" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": " are represented by " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "x_{i}^{(l)}" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": ", and the columns in " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": " correspond to the softmax values of each query token calculated for " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": " and the perturbed inputs " + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 55, + 177, + 296, + 307 + ], + "type": "text", + "content": ", respectively. Therefore," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 312, + 188, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 312, + 188, + 326 + ], + "spans": [ + { + "bbox": [ + 85, + 312, + 188, + 326 + ], + "type": "interline_equation", + "content": "W _ {v} (X ^ {(1)} S - \\hat {X} ^ {(1)} S) =", + "image_path": "597e2f808d2ff5548d3acff1083639957622117f96dcd46e5e42dcd8c7926c4f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 328, + 264, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 328, + 264, + 343 + ], + "spans": [ + { + "bbox": [ + 85, + 328, + 264, + 343 + ], + "type": "interline_equation", + "content": "W _ {v} \\left(X ^ {(1)} S - \\hat {X} ^ {(1)} \\hat {S} + \\hat {X} ^ {(1)} \\hat {S} - \\hat {X} ^ {(1)} S\\right) =", + "image_path": "a9aa377de4adf5ce89abe0b5c82ead031951f6def638e57adbf4f4f4e52eb29a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 346, + 261, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 346, + 261, + 360 + ], + "spans": [ + { + "bbox": [ + 87, + 346, + 261, + 360 + ], + "type": "interline_equation", + "content": "R + W _ {v} \\hat {X} ^ {(1)} (\\hat {S} - S) = R + W _ {v} \\hat {X} ^ {(1)} \\Delta S.", + "image_path": "bbe7489e00b2120ba12c02bf5b257b857ef53fbf957c0f24ba0b3546933b72d6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 368, + 214, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 214, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 214, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 368, + 214, + 380 + ], + "type": "inline_equation", + "content": "\\Delta S\\coloneqq \\hat{S} -S" + }, + { + "bbox": [ + 55, + 368, + 214, + 380 + ], + "type": "text", + "content": " . Hence, we have:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 385, + 265, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 385, + 265, + 414 + ], + "spans": [ + { + "bbox": [ + 85, + 385, + 265, + 414 + ], + "type": "interline_equation", + "content": "X ^ {(1)} - \\hat {X} ^ {(1)} = \\underbrace {W _ {v} ^ {+} (R + W _ {v} \\hat {X} ^ {(1)} \\Delta S)} _ {U} S ^ {+}.", + "image_path": "49418daf42eae266d57ea2736a42f5b569e9dfe0b567d88cd57e3e5fb4f13766.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "W_{v}^{+}" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "S^{+}" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": " are the pseudo-inverses of the matrices " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": ", respectively. Note that for the " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\delta := X^{(1)} - \\hat{X}^{(1)}" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": " perturbation in the input, in order to make the required perturbation in the output of the network, we get " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\| \\delta \\| \\leq \\lambda_{max} \\| U \\|" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\lambda_{max}" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": " is the maximum eigenvalue of " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "S^{+}" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": ". But we note that for a stochastic matrix " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": ", the maximum eigenvalue is 1. Therefore, for the inverse matrix, " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "S^{+}" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": ", the maximum eigenvalue would be larger than 1. As a result, we would expect the " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\| \\delta \\|" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": " to grow proportional to " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\lambda_{max} \\geq 1" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": ". We note that the smaller " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\| \\delta \\|" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": " would imply more vulnerability as it becomes easier to attack the network with a smaller perturbation length. On the other hand, for a purely localized " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": ", where each token only attends to itself " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\lambda_{max} = 1" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": ", showing that the increase in " + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\| \\delta \\|" + }, + { + "bbox": [ + 55, + 420, + 296, + 659 + ], + "type": "text", + "content": " would likely be small compared to the general case, leading to more vulnerability. Therefore, we conclude that a localized and overly sparse attention map would typically imply an easier attack on the network and hence higher vulnerability against adversarial attacks. Sparse attention values may also resemble what happens in convolutional networks that have small receptive fields." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 669, + 138, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 138, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 138, + 682 + ], + "type": "text", + "content": "5. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Overview. Current AD and AL methods perform poorly when subjected to adversarial attacks. Previous studies have" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": "demonstrated the effectiveness of OE in enhancing robustness against such attacks [3, 10, 11, 51]. Additionally, ViTs have shown promising results in standard (clean) AD and AL tasks [15, 37, 41, 64], and our findings suggest their potential for adversarial settings with targeted modifications of attention degrees. Building on these insights and our theoretical foundations, we propose PatchGuard, a novel method designed to enhance adversarial robustness in AD and AL. The following subsections detail the key components of PatchGuard." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 199, + 555, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 222 + ], + "type": "text", + "content": "5.1. Foreground-Aware Pseudo-Anomaly Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 228, + 556, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 556, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 556, + 348 + ], + "type": "text", + "content": "To generate anomalous data from normal samples, it is essential to apply targeted distortions that ensure the image becomes anomalous [8, 13, 14, 26]. Additionally, the generated pseudo-anomalies must satisfy two key requirements: (1) inclusion of localization maps to enhance the AL task, and (2) close alignment with the normal sample distribution [51, 53]. Meeting these requirements simultaneously necessitates precise but minimal alterations within the foreground. Thus, the proposed method must reliably identify a region of the image that we can confidently attribute to the foreground." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "text", + "content": "To achieve this, we leverage insights from [55], which indicate that pretrained models benefit from style-agnostic processing to improve classification accuracy. To this end, we first introduce soft transformations, which are commonly used in self-supervised learning studies due to their ability to preserve semantic content [8, 13, 14, 26, 31]. Next, we apply Grad-CAM [60] to generate a saliency map using a standard pre-trained ResNet18 [30] model on a given input. Specifically, for a normal sample " + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "text", + "content": ", we randomly select " + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{soft}}" + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "text", + "content": " soft transformations " + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "inline_equation", + "content": "t_1^s,\\ldots ,t_{k_{\\mathrm{soft}}}^s" + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "text", + "content": ". We then compute saliency maps for " + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "inline_equation", + "content": "x,t_1^s (x),\\ldots ,t_{k_{\\mathrm{soft}}}^s (x)" + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "text", + "content": " and take their element-wise product to produce a style-invariant saliency map, " + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "text", + "content": ". As shown in Figure 3 (part ①), " + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "text", + "content": " effectively assigns higher values to the foreground region." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "text", + "content": "To distort the obtained region, we first randomly sample an anchor point from it. We then sample " + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{hard}}" + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "text", + "content": " hard transformations " + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "inline_equation", + "content": "t_1^h, \\ldots, t_{k_{\\mathrm{hard}}}^h" + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "text", + "content": ", which are typically associated with altering semantic integrity, as highlighted in prior research [1, 9, 18, 23, 33, 38, 40, 62, 63, 66, 77]. We then design a rectangular mask centered on this chosen point, with its width " + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "text", + "content": " and height " + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "text", + "content": " satisfying " + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "inline_equation", + "content": "w \\sim U(0.05W, 0.3W)" + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "inline_equation", + "content": "h \\sim U(0.05H, 0.3H)" + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "text", + "content": " are the image's width and height. Next, we rotate the mask by an angle of " + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "inline_equation", + "content": "\\theta \\sim U(-45^\\circ, 45^\\circ)" + }, + { + "bbox": [ + 313, + 515, + 557, + 672 + ], + "type": "text", + "content": ". Finally, we sequentially apply the hard transformations only to the region within this mask, leaving the remaining unmasked areas of the image unchanged. Mathematically, the final anomaly is obtained by applying" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 678, + 555, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 678, + 555, + 694 + ], + "spans": [ + { + "bbox": [ + 336, + 678, + 555, + 694 + ], + "type": "interline_equation", + "content": "x ^ {\\prime} = t _ {1} ^ {h} \\left(\\dots \\left(t _ {k _ {\\text {h a r d}}} ^ {h} \\left(x _ {\\text {m a s k e d}}\\right) \\dots\\right) + \\left(x - x _ {\\text {m a s k e d}}\\right). \\right. \\tag {1}", + "image_path": "54df5ef58c128e2f5c8af4688437413501596e3b0bb30a4b81ede48f3a73cf9c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 701, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 556, + 714 + ], + "type": "text", + "content": "An additional advantage of this pseudo-anomaly crafting ap-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20386" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 281, + 226 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 281, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 281, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 281, + 226 + ], + "type": "image", + "image_path": "5c9deb95b88b8d22cd4d1f87d9d1a4fc393f0c62e6fcb33c765505e6e90afbe9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 242, + 269, + 357 + ], + "blocks": [ + { + "bbox": [ + 62, + 242, + 269, + 357 + ], + "lines": [ + { + "bbox": [ + 62, + 242, + 269, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 242, + 269, + 357 + ], + "type": "image", + "image_path": "d5cb6698cfd038056731f52aa802055062cddea6a312e0abf28dd5f8bc324b6b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": "Figure 3. Overview of the PatchGuard framework. (1) We use Grad-CAM to identify regions likely to belong to the foreground in an image. By applying " + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "inline_equation", + "content": "k_{\\mathrm{soft}} = 3" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": " augmentations to the image " + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": ", we generate a combined saliency map " + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": ". (2) The input " + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": " and its saliency map " + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": " are then passed to the anomaly generator to produce an anomaly sample " + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": " along with its ground-truth mask " + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "inline_equation", + "content": "x_m'" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": ". (3) For each normal sample " + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": ", the input batch includes the image, its corresponding anomaly version " + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "inline_equation", + "content": "x_i'" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": ", and their adversarially attacked variants, along with their ground-truth masks. Each batch sample is processed through a Vision Transformer (ViT), where the Attention Discriminator assigns an anomaly score " + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": " to each patch embedding. The final anomaly map is constructed from these scores and trained with a novel loss " + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 369, + 555, + 449 + ], + "type": "text", + "content": " to replicate the ground truth." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 285, + 71, + 549, + 358 + ], + "blocks": [ + { + "bbox": [ + 285, + 71, + 549, + 358 + ], + "lines": [ + { + "bbox": [ + 285, + 71, + 549, + 358 + ], + "spans": [ + { + "bbox": [ + 285, + 71, + 549, + 358 + ], + "type": "image", + "image_path": "4d5ed744a1270fd6bb81971e28608eccaf9d6ed5241360a8cd2aa11bc6010e51.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 501, + 555, + 677 + ], + "blocks": [ + { + "bbox": [ + 55, + 468, + 556, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 556, + 495 + ], + "type": "text", + "content": "Table 1. Detection performance of various AD methods, including PatchGuard, across benchmark datasets. Results are shown as “Clean / Adversarial”, representing performance on “Clean” and “Adversarial” data (PGD-1000, " + }, + { + "bbox": [ + 55, + 468, + 556, + 495 + ], + "type": "inline_equation", + "content": "\\epsilon = \\frac{8}{255}" + }, + { + "bbox": [ + 55, + 468, + 556, + 495 + ], + "type": "text", + "content": ", measured by AUROC %), respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 501, + 555, + 677 + ], + "lines": [ + { + "bbox": [ + 56, + 501, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 555, + 677 + ], + "type": "table", + "html": "
DatasetMethod
DRAEMPatchCoreSimpleNetReContrastGeneralADGLADGLASSPrincipaLS*ZARND*RODEO*Ours
MVTec AD98.0 / 5.199.1 / 3.999.6 / 2.399.5 / 0.099.2 / 13.799.3 / 11.899.9 / 6.763.8 / 10.671.6 / 11.061.5 / 4.988.1 / 71.1
VisA88.7 / 3.895.1 / 1.796.2 / 0.297.5 / 6.795.9 / 14.599.5 / 8.698.8 / 7.258.9 / 9.367.7 / 9.363.1 / 3.888.5 / 74.3
BTAD87.1 / 4.996.9 / 2.495.1 / 4.695.7 / 2.197.0 / 10.998.5 / 7.697.4 / 8.360.1 / 7.869.7 / 8.360.0 / 5.885.3 / 82.1
MPDD94.3 / 2.891.3 / 6.896.6 / 7.094.5 / 4.598.0 / 12.897.5 / 9.799.6 / 9.171.9 / 15.765.7 / 12.268.8 / 7.485.4 / 68.6
WFDD94.0 / 7.196.3 / 3.998.8 / 5.197.6 / 3.399.0 / 10.799.1 / 12.5100 / 6.066.3 / 14.176.1 / 16.371.2 / 7.884.2 / 65.0
DTD-Synthetic97.1 / 3.665.8 / 4.096.8 / 2.498.2 / 9.197.9 / 16.398.6 / 10.399.1 / 6.177.9 / 9.370.3 / 13.772.8 / 12.091.5 / 60.5
Head-CT72.1 / 8.192.4 / 5.184.9 / 4.886.1 / 6.389.1 / 9.193.1 / 14.294.7 / 8.773.2 / 16.064.3 / 9.785.7 / 61.097.3 / 90.4
BraTS202162.3 / 1.491.6 / 6.382.5 / 6.182.4 / 3.188.3 / 13.092.3 / 12.095.7 / 7.275.3 / 8.766.7 / 14.089.1 / 64.794.3 / 81.0
Average86.7 / 4.691.0 / 4.293.8 / 4.093.9 / 4.395.5 / 12.697.2 / 10.898.1 / 7.468.4 / 11.469.0 / 11.871.5 / 20.989.3 / 74.1
", + "image_path": "827ad7a3dc58fc161fd52f1378f0b682782b1e7d5239446832ff9ea3c5e1bcd6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 58, + 678, + 287, + 689 + ], + "lines": [ + { + "bbox": [ + 58, + 678, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 58, + 678, + 287, + 689 + ], + "type": "text", + "content": "*These works incorporated adversarial training into their proposed AD methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20387" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 103, + 555, + 284 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 2. Localization performance of various AL methods, including PatchGuard, across benchmark datasets. Results are shown as \"Clean / Adversarial\", representing performance on \"Clean\" and \"Adversarial\" data (PGD-1000, " + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "\\epsilon = \\frac{8}{255}" + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": ", measured by AUROC %), respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 103, + 555, + 284 + ], + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 284 + ], + "type": "table", + "html": "
DatasetMethod
DRAEMPatchCoreSimpleNetReContrastGeneralADGLADGLASSMuScReConPatchRealNetOurs
MVTec AD97.3 / 3.998.1 / 4.298.1 / 6.898.4 / 4.897.7 / 12.398.7 / 15.399.3 / 8.197.3 / 4.699.2 / 2.899.0 / 8.792.7 / 73.8
VisA90.7 / 5.698.5 / 3.097.4 / 4.898.2 / 7.199.0 / 7.698.6 / 13.498.8 / 7.098.8 / 3.899.2 / 1.798.8 / 6.296.9 / 85.2
BTAD89.0 / 7.192.6 / 6.395.2 / 3.796.8 / 6.496.7 / 9.398.2 / 14.597.0 / 5.797.3 / 5.397.5 / 3.097.9 / 9.293.2 / 73.0
MPDD90.7 / 4.798.5 / 1.997.4 / 7.394.6 / 3.698.1 / 14.998.7 / 10.899.4 / 6.395.5 / 6.096.7 / 2.698.2 / 7.593.8 / 86.6
WFDD95.8 / 3.998.1 / 2.198.0 / 3.497.3 / 5.198.7 / 13.597.6 / 11.398.9 / 7.697.0 / 4.198.3 / 4.998.4 / 6.894.6 / 71.6
DTD-Synthetic96.1 / 4.895.6 / 6.797.0 / 2.996.8 / 7.697.9 / 10.897.3 / 13.898.1 / 7.097.8 / 5.796.3 / 3.998.0 / 8.395.9 / 74.8
Head-CT86.4 / 8.987.3 / 6.189.7 / 0.890.1 / 8.691.7 / 18.393.4 / 16.792.2 / 9.185.2 / 8.589.7 / 6.694.8 / 7.393.6 / 89.3
BraTS202182.2 / 6.196.9 / 4.594.7 / 6.795.3 / 11.196.8 / 14.795.1 / 13.594.0 / 6.367.5 / 6.397.5 / 5.996.9 / 10.297.7 / 94.5
Average91.0 / 5.695.7 / 4.395.9 / 4.595.9 / 6.797.0 / 12.697.2 / 13.697.2 / 7.192.0 / 5.596.8 / 3.997.7 / 8.094.8 / 81.1
", + "image_path": "532dd786bbcf664ffd4ca0b9427489b7c4063a650c736d7d4f9f891a4c6b1e4f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 307, + 297, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 297, + 452 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 297, + 452 + ], + "type": "text", + "content": "proach is that the location of the distorted region is precisely known, allowing us to generate an accurate localization map " + }, + { + "bbox": [ + 54, + 307, + 297, + 452 + ], + "type": "inline_equation", + "content": "x_{m}^{\\prime}" + }, + { + "bbox": [ + 54, + 307, + 297, + 452 + ], + "type": "text", + "content": " for AL training, where each pixel's value equals 1 if it's anomalous, and 0 otherwise. By creating an anomalous counterpart " + }, + { + "bbox": [ + 54, + 307, + 297, + 452 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 54, + 307, + 297, + 452 + ], + "type": "text", + "content": " and its corresponding localization map " + }, + { + "bbox": [ + 54, + 307, + 297, + 452 + ], + "type": "inline_equation", + "content": "x_{m}^{\\prime}" + }, + { + "bbox": [ + 54, + 307, + 297, + 452 + ], + "type": "text", + "content": " for each normal sample " + }, + { + "bbox": [ + 54, + 307, + 297, + 452 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 307, + 297, + 452 + ], + "type": "text", + "content": " in the training set, we provide the model with both the anomaly and its location, as shown in Figure 3 (part ②). This pairing equips the model with the information needed to improve its localization accuracy, as we will discuss in Section 5.2. For details on the soft and hard transforms, and ablations on the hypers used in this approach, refer to Table 6 and Appendix B." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 456, + 296, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 481 + ], + "type": "text", + "content": "5.2. Anomaly Detection and Localization with Vision Transformer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 486, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 297, + 605 + ], + "type": "text", + "content": "To enhance adversarial robustness through the effective use of pseudo-anomaly samples and their corresponding localization maps, we employ a Vision Transformer (ViT) model. This model generates output embeddings that are processed by an attention-based discriminator network, which uses a multi-head attention (MLH) mechanism followed by a shared-weight multi-layer perceptron (MLP). The discriminator independently assigns an anomaly score to each patch in the input, providing a patch-level anomaly prediction as shown in Figure 3 (part ③)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "content": "The model is trained using a combination of patch-level cross-entropy (CE) and an additional regularization term " + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "content": ", designed to encourage a higher degree of attention in a specified ViT layer. As illustrated in Figure 2 and the theory in Section 4, this regularizer is expected to further enhance adversarial robustness. The cross-entropy loss component aligns the predicted anomaly scores with ground truth labels from the localization map, " + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "content": ", where a patch label equals 1 if more than " + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "content": " of the pixels in the corresponding patch in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 307, + 463, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 463, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 463, + 319 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 313, + 307, + 463, + 319 + ], + "type": "text", + "content": " are anomalous, and 0 otherwise:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 380, + 329, + 555, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 329, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 380, + 329, + 555, + 342 + ], + "type": "interline_equation", + "content": "M = \\operatorname {M L P} (\\operatorname {M L H} (x)) \\tag {2}", + "image_path": "f546bb4d0414f454b64e40b9e1e974ff50ead9630e68cd6914e4c067a38352c3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 343, + 345, + 555, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 345, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 343, + 345, + 555, + 372 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} (x, x _ {m}) = \\sum_ {i, j \\in x _ {\\text {p a t c h e s}}} \\mathrm {C E} \\left(M _ {(i, j)}, x _ {m, (i, j)}\\right) \\tag {3}", + "image_path": "d0c28c1dd4d577139033df8a649014a0534ed9c1d6d05f73edf37c60bf97f780.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": " denotes the predicted AL map, consisting of anomaly scores for each patch, " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": " is any input sample (normal or anomaly), and " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": " are the indices for individual patches. The cross-entropy loss term " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": " ensures alignment between the patch " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "M_{(i,j)}" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": " and the ground truth anomaly map " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "x_{m,(i,j)}" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": ", promoting precise anomaly localization. The final AL map is obtained by upsampling " + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": " to the original image size, and we use the average of the top-" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 380, + 556, + 489 + ], + "type": "text", + "content": " patch-level anomaly scores as the image-level anomaly score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 489, + 556, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 556, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 556, + 537 + ], + "type": "text", + "content": "To further improve the robustness of the model, we introduce a regularization term " + }, + { + "bbox": [ + 313, + 489, + 556, + 537 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 489, + 556, + 537 + ], + "type": "text", + "content": ", which encourages an increased attention degree in a specified layer " + }, + { + "bbox": [ + 313, + 489, + 556, + 537 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 489, + 556, + 537 + ], + "type": "text", + "content": " of the ViT, chosen here as the final layer. This regularization term is formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 546, + 555, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 546, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 351, + 546, + 555, + 574 + ], + "type": "interline_equation", + "content": "R (x) = \\frac {1}{\\sum_ {i} \\sum_ {j} \\sum_ {k} A _ {i j k} \\cdot \\mathbb {1} \\left(A _ {i j k} \\leq \\delta\\right)} \\tag {4}", + "image_path": "830737a90d3d8aca6f09426390eec2c32b5267c06cc75209829dbfedb7cfa51c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "inline_equation", + "content": "A_{ijk}" + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "text", + "content": " represents the attention coefficient corresponding to the " + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "text", + "content": "-th head, the " + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "text", + "content": "-th output patch, and the " + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "text", + "content": "-th input patch. Regularization will increase the values of attention coefficients that are below the threshold " + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "text", + "content": ". For a specific output patch, since a softmax function is applied to its corresponding attention coefficients, increasing the lower values naturally decreases the higher ones. With " + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\delta = \\frac{1}{(\\text{number of input patches})}" + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "text", + "content": ", the coefficient distribution gradually shifts toward a uniform distribution at this threshold, representing the maximum attention degree a patch can achieve. The total loss function " + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 582, + 557, + 714 + ], + "type": "text", + "content": " used for training the model is the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20388" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 271, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 271, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 271, + 84 + ], + "type": "text", + "content": "sum of the cross-entropy and the regularization term:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 94, + 296, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 94, + 296, + 107 + ], + "spans": [ + { + "bbox": [ + 104, + 94, + 296, + 107 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(x, x _ {m}\\right) = \\mathcal {L} _ {\\mathrm {C E}} \\left(x, x _ {m}\\right) + \\alpha R (x) \\tag {5}", + "image_path": "66138f6ba27180f386ccf2e71727067eee8a717758fe432858e81efc8df699cb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 117, + 296, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 117, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 117, + 296, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 117, + 296, + 213 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 117, + 296, + 213 + ], + "type": "text", + "content": " is a scaling factor that controls the strength of the regularization. This combined objective encourages the model to not only accurately predict anomaly locations but also maintain an increased level of attention, enhancing both detection performance and adversarial robustness. It is then used in adversarial training, as explained in Section 5.3. Comprehensive ablation studies of hyperparameters including " + }, + { + "bbox": [ + 55, + 117, + 296, + 213 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 117, + 296, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 117, + 296, + 213 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 117, + 296, + 213 + ], + "type": "text", + "content": " are provided in Appendix C.2 and Section 7." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 220, + 234, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 234, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 234, + 234 + ], + "type": "text", + "content": "5.3. Adversarial Training and Testing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 238, + 296, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 274 + ], + "type": "text", + "content": "Given an input sample " + }, + { + "bbox": [ + 55, + 238, + 296, + 274 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 238, + 296, + 274 + ], + "type": "text", + "content": ", an adversarial example " + }, + { + "bbox": [ + 55, + 238, + 296, + 274 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 55, + 238, + 296, + 274 + ], + "type": "text", + "content": " is generated by introducing a perturbation " + }, + { + "bbox": [ + 55, + 238, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\delta^{*}" + }, + { + "bbox": [ + 55, + 238, + 296, + 274 + ], + "type": "text", + "content": ", optimized to maximize the final loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 81, + 283, + 296, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 283, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 81, + 283, + 296, + 306 + ], + "type": "interline_equation", + "content": "\\delta^ {*} = \\underset {\\| \\delta \\| _ {\\infty} \\leq \\epsilon} {\\operatorname {a r g m a x}} \\mathcal {L} (x + \\delta , x _ {m}), \\quad x _ {a d v} = x + \\delta^ {*} \\tag {6}", + "image_path": "a79ae873a59a118e93b42967a922c55b30a22a073b5e6fb911e7fc79a9709e6f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 314, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 374 + ], + "type": "text", + "content": "These adversarial examples are then used in the training process alongside the original examples. The adversarial training objective is framed as a min-max problem, optimizing the model parameters " + }, + { + "bbox": [ + 55, + 314, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 314, + 296, + 374 + ], + "type": "text", + "content": " to minimize the expected loss over both clean and adversarial examples:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 382, + 296, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 382, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 91, + 382, + 296, + 409 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {(x, y) \\in \\mathcal {B}} \\left[ \\max _ {\\| \\delta \\| _ {\\infty} \\leq \\epsilon} \\mathcal {L} (x + \\delta , x _ {m}; \\theta) \\right]. \\tag {7}", + "image_path": "107d086a2dec8e06eee4d757b90fef33afb547288c7af36d5b7f57843331f2dd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 418, + 137, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 137, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 137, + 431 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "content": "To demonstrate the effectiveness of PatchGuard, we conduct extensive experiments assessing its performance alongside several established anomaly detection methods on multiple benchmarks for anomaly detection and localization. Our evaluation includes both adversarially trained models and those trained under standard conditions. To ensure a fair comparison, we assess all methods under both \"Clean\" and \"Adversarial\" test conditions, as shown in Table 1 and 2. In the Clean setting, test samples are unaltered, whereas the Adversarial setting includes samples perturbed by the PGD attack [48]. Additional results on PatchGuard's performance against alternative attacks are provided in Section 7.2. Datasets are detailed in Appendix D." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": "Evaluation Details. In the adversarial setting, we evaluate each method using the " + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": " PGD-1000 attack with " + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\epsilon = \\frac{8}{255}" + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": ". Detection and localization results are shown in Tables 1 and 2, respectively. " + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": " norm evaluations are in Appendix J." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Adapting State-of-the-Art AD and AL Methods to Adversarial Settings. Prior to developing our novel Patch-Guard framework, we first investigated whether existing state-of-the-art methods in AD and AL could be enhanced through adversarial training—a technique where models are" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 103, + 555, + 329 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "type": "text", + "content": "Table 3. Performance of prior methods adapted for adversarial settings, shown as \"Clean / Adversarial\" (PGD-1000, " + }, + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "type": "inline_equation", + "content": "\\epsilon = \\frac{8}{255}" + }, + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 103, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 315, + 103, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 315, + 103, + 555, + 329 + ], + "type": "table", + "html": "
MethodTaskDataset
MVTec ADVisABTADBraTS2021
PatchCoreAD92.3 / 14.388.4 / 10.887.1 / 12.788.6 / 16.1
AL89.1 / 10.387.6 / 8.885.0 / 11.586.7 / 13.6
ReContrastAD93.8 / 5.193.7 / 16.789.9 / 13.478.2 / 13.1
AL94.5 / 17.292.5 / 18.288.6 / 17.990.8 / 14.2
SimpleNetAD91.1 / 18.189.2 / 14.988.6 / 12.775.3 / 14.2
AL94.1 / 19.589.4 / 15.888.2 / 11.790.1 / 17.0
GeneralADAD84.6 / 24.781.2 / 22.182.9 / 20.680.7 / 19.9
AL83.7 / 21.883.1 / 17.384.7 / 18.379.6 / 21.0
DRAEMAD83.0 / 19.281.7 / 16.879.1 / 17.057.4 / 12.2
AL86.1 / 17.578.6 / 15.080.9 / 18.673.6 / 14.9
GLASSAD84.2 / 26.783.5 / 19.984.7 / 25.483.7 / 20.1
AL85.9 / 21.287.1 / 21.783.0 / 19.085.8 / 23.6
OursAD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 / 85.293.2 / 73.097.7 / 94.5
", + "image_path": "1cb3e5a91fa83810ea064588b0d795aa98b7b12e149c353c957df7bb1c550fa5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 354, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 461 + ], + "type": "text", + "content": "trained on adversarial examples to improve their robustness. In Appendix E, we provide a detailed methodology for generating adversarial examples for each method and describe our optimized approach for improving their resilience. The empirical results, presented in Table 3, reveal that even after incorporating adversarial training, current SOTA AD and AL methods remain vulnerable to adversarial attacks and exhibit suboptimal performance. These limitations motivated the development of PatchGuard, our proposed solution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 462, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 653 + ], + "type": "text", + "content": "Analyzing Results. As shown in Tables 1 and 2, prior SOTA methods such as Recontrast demonstrate substantial drops in performance under adversarial settings, despite excelling in clean conditions. Furthermore, as indicated in Table 3, even when adversarily trained, these methods struggle to achieve robust performance. By contrast, approaches specifically designed to address adversarial robustness, such as RODEO, still fall short when compared to PatchGuard. On average, PatchGuard improves robust detection across diverse datasets by up to " + }, + { + "bbox": [ + 313, + 462, + 556, + 653 + ], + "type": "inline_equation", + "content": "53.2\\%" + }, + { + "bbox": [ + 313, + 462, + 556, + 653 + ], + "type": "text", + "content": " percent, with a notable robust improvement of " + }, + { + "bbox": [ + 313, + 462, + 556, + 653 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 313, + 462, + 556, + 653 + ], + "type": "text", + "content": " percent on the challenging MVtec dataset. Notably, we achieve a " + }, + { + "bbox": [ + 313, + 462, + 556, + 653 + ], + "type": "inline_equation", + "content": "71.8\\%" + }, + { + "bbox": [ + 313, + 462, + 556, + 653 + ], + "type": "text", + "content": " improvement in AL on the challenging VisA dataset, with details in Appendix F. As prior research indicates, a slight decrease in clean performance is generally viewed as an acceptable tradeoff for the critical gains in robustness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Implementation Details. We leverage a from-scratch ViT model in our pipeline to showcase our setup's independence from pre-train datasets. However, we conduct an ablation study on this manner in Appendix C.4. Further, we leverage PGD-10 with " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon = \\frac{8}{255}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " for adversarial training," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20389" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 114, + 296, + 240 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 103 + ], + "type": "text", + "content": "Table 4. Evaluation of PatchGuard's performance against advanced classification and segmentation attacks, showcasing its robustness in diverse attack scenarios." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 114, + 296, + 240 + ], + "lines": [ + { + "bbox": [ + 56, + 114, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 114, + 296, + 240 + ], + "type": "table", + "html": "
DatasetTaskDetection AttacksLocalization Attacks
PGD-1000CAAAutoAttackA3SegPGDSEA
MVTec ADAD71.174.870.869.871.969.2
AL73.876.772.970.074.671.3
VisAAD74.378.574.673.175.073.6
AL85.287.485.584.785.083.7
BTADAD82.183.482.080.781.980.8
AL73.074.772.371.572.971.1
BraTS2021AD81.084.781.980.382.081.1
AL94.595.094.293.894.893.4
", + "image_path": "41981ffa0a29b65e7ee394009613ce0690d64355fe0d75eb6f4ab8facf10400e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 285, + 296, + 419 + ], + "blocks": [ + { + "bbox": [ + 55, + 252, + 296, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 275 + ], + "type": "text", + "content": "Table 5. Ablation study on the scaling factor " + }, + { + "bbox": [ + 55, + 252, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 252, + 296, + 275 + ], + "type": "text", + "content": ", illustrating the effectiveness of our regularization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 285, + 296, + 419 + ], + "lines": [ + { + "bbox": [ + 56, + 285, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 296, + 419 + ], + "type": "table", + "html": "
Scaling Factor (α)TaskDataset
MVTec-ADVisABTADBraTS2021
0AD89.3 / 60.790.1 / 62.988.6 / 74.395.2 / 72.0
AL93.5 / 65.395.8 / 72.194.0 / 59.798.2 / 83.4
1 (Ours)AD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 / 85.293.2 / 73.097.7 / 94.5
2AD86.2 / 70.187.6 / 75.184.7 / 83.492.8 / 81.2
AL91.0 / 68.994.8 / 84.191.9 / 73.995.8 / 94.6
5AD80.1 / 64.178.6 / 65.473.8 / 69.682.5 / 73.2
AL84.7 / 65.681.1 / 78.984.0 / 65.386.3 / 79.9
", + "image_path": "7f53e183a163b55bb96e2123b52fd3f020b0494b93d323b8607df5868a9285c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 441, + 297, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 297, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 297, + 479 + ], + "type": "text", + "content": "but the test results are generated with the PGD-1000 attack. For more details on the implementation parameters, refer to Appendix G." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 488, + 157, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 157, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 157, + 502 + ], + "type": "text", + "content": "7. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 508, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 533 + ], + "type": "text", + "content": "In this section, we provided further analysis of our choices of parameters and other components of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 149, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 149, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 149, + 553 + ], + "type": "text", + "content": "7.1. Regularization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "text", + "content": "To experimentally demonstrate the effectiveness of our regularization term on the model's robustness and the impact of the scaling factor " + }, + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "text", + "content": " (set to 1 by default in our method), we conduct an ablation study on " + }, + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "text", + "content": ", as shown in Table 5. When " + }, + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "text", + "content": ", the model's robustness against adversarial attacks decreases due to the lack of regularization. On the other hand, higher values of " + }, + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "text", + "content": ", such as " + }, + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "inline_equation", + "content": "\\alpha = 5" + }, + { + "bbox": [ + 55, + 557, + 297, + 665 + ], + "type": "text", + "content": ", cause the regularization term to dominate, leading to an unstable loss function and inappropriate results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 165, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 165, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 165, + 685 + ], + "type": "text", + "content": "7.2. Advanced Attacks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "content": "In our primary experiments in Section 6, we utilized variations of the PGD attack [48] for both model training and" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 103, + 555, + 261 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "type": "text", + "content": "Table 6. An ablation study on our method's performance using various anomaly generation techniques that incorporate masks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 103, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 315, + 103, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 315, + 103, + 555, + 261 + ], + "type": "table", + "html": "
Pseudo-Anomaly GeneratorTaskDataset
MVTec ADVisABTADBraTS2021
CutPasteAD79.1 / 60.374.4 / 61.574.8 / 69.280.3 / 72.7
AL83.9 / 60.885.4 / 73.082.6 / 61.291.2 / 82.1
FPIAD78.9 / 62.175.1 / 64.372.9 / 71.084.1 / 74.1
AL84.3 / 63.882.9 / 71.886.3 / 60.593.3 / 84.8
NSAAD81.7 / 66.583.1 / 65.477.1 / 73.486.3 / 73.8
AL85.1 / 65.083.3 / 76.785.4 / 64.989.4 / 86.8
GLASSAD83.2 / 62.084.0 / 68.779.3 / 73.983.1 / 66.1
AL86.7 / 67.086.2 / 78.586.8 / 68.292.2 / 87.3
OursAD88.1 / 71.188.5 / 74.385.3 / 82.194.3 / 81.0
AL92.7 / 73.896.9 /85.293.2 /73.097.7 /94.5
", + "image_path": "8c54935fba97c20127d235a19f8a7f7253beedc5a1d1ea5e08a01a2e0f401a20.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 284, + 556, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 556, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 556, + 452 + ], + "type": "text", + "content": "testing. To further illustrate the adaptability and robustness of our proposed method under diverse adversarial conditions, we also evaluate its performance against a variety of other attack types including Detection attacks and Localization attacks. The Detection attacks consist of black-box attacks [28], FGSM [24], CAA [49], AutoAttack [16], and " + }, + { + "bbox": [ + 313, + 284, + 556, + 452 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^3" + }, + { + "bbox": [ + 313, + 284, + 556, + 452 + ], + "type": "text", + "content": " (Adversarial Attack Automation) [43]. The localization attacks are adapted from the well-known segmentation attacks in literature, namely SegPGD [27] and SEA [17]. All results are available in Table 4, and the exact details on our adaptation approach for these attacks are available in Appendix H. It is important to note that the training process remains consistent, employing only the standard PGD-10 for simplicity and practicality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 460, + 515, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 515, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 515, + 474 + ], + "type": "text", + "content": "7.3. Pseudo-Anomaly Generation Strategy" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 478, + 556, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 556, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 556, + 586 + ], + "type": "text", + "content": "Since our method targets both AD and AL using pseudo-anomalies, it is crucial to evaluate our pseudo-anomaly generation against compatible methods that produce localization maps. Thus, methods like MIXUP [33], FakeIt [50], and Dream-OOD [20] are excluded. We compare instead with CutPaste [40], NSA [59], GLASS [12], and FPI [67], keeping other pipeline components fixed. As shown in Table 6, our method outperforms these approaches, highlighting its effectiveness for robust anomaly localization." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 597, + 388, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 388, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 388, + 610 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "We introduce PatchGuard, an adversarially robust method for AD and AL. It employs Foreground-Aware Pseudo-Anomaly Generation using segmentation masks. Additionally, we propose a regularization term, motivated by observations and theory, that strengthens robustness against adversarial attacks by increasing the attention degree. PatchGuard outperforms previous methods on AD and AL benchmarks under adversarial settings." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20390" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 136 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 136 + ], + "type": "text", + "content": "We are very grateful to Mohammad Javad Maheronnaghsh for his collaboration in the initial stages of this project and to the anonymous reviewers for their helpful discussions and valuable feedback." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 160, + 115, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 160, + 115, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 160, + 115, + 172 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 179, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 179, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 179, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 179, + 296, + 201 + ], + "type": "text", + "content": "[1] M. Eren Akbiyik. Data augmentation in training cnns: Injecting noise to images. ArXiv, abs/2307.06855, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 204, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 204, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 296, + 236 + ], + "type": "text", + "content": "[2] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. *IEEE Access*, 6:14410–14430, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 296, + 293 + ], + "type": "text", + "content": "[3] Mohammad Azizmalayeri, Arshia Soltani Moakhar, Arman Zarei, Reihaneh Zohrabi, Mohammad Manzuri, and Mohammad Hossein Rohhan. Your out-of-distribution detection method is not robust! Advances in Neural Information Processing Systems, 35:4887-4901, 2022. 1, 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 295, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 296, + 360 + ], + "type": "text", + "content": "[4] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 296, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 296, + 406 + ], + "type": "text", + "content": "[5] Abhijit Bendale and Terrance Boult. Towards open world recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1893-1902, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 408, + 296, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 296, + 462 + ], + "type": "text", + "content": "[6] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 465, + 296, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 296, + 519 + ], + "type": "text", + "content": "[7] Louis Béthane, Paul Novello, Thibaut Boissin, Guillaume Coiffier, Mathieu Serrurier, Quentin Vincenot, and Andres Troya-Galvis. Robust one-class classification with signed distance function using 1-lipschitz neural networks. arXiv preprint arXiv:2303.01978, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 521, + 296, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 296, + 575 + ], + "type": "text", + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 578, + 296, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 296, + 654 + ], + "type": "text", + "content": "[9] Chengwei Chen, Yuan Xie, Shaohui Lin, Ruizhi Qiao, Jian Zhou, Xin Tan, Yi Zhang, and Lizhuang Ma. Novelty detection via contrastive learning with negative data augmentation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI 2021, Virtual Event / Montreal, Canada, 19-27 August 2021, pages 606-614. ijcai.org, 2021. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 657, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 296, + 689 + ], + "type": "text", + "content": "[10] Jiefeng Chen, Yixuan Li, Xi Wu, Yingyu Liang, and Somesh Jha. Robust out-of-distribution detection for neural networks. arXiv preprint arXiv:2003.09711, 2020. 1, 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "text", + "content": "[11] Jiefeng Chen, Yixuan Li, Xi Wu, Yingyu Liang, and Somesh Jha. Atom: Robustifying out-of-distribution detection using" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 333, + 73, + 555, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 118 + ], + "type": "text", + "content": "outlier mining. In Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain, September 13-17, 2021, Proceedings, Part III 21, pages 430-445. Springer, 2021. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 119, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 555, + 161 + ], + "type": "text", + "content": "[12] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. arXiv preprint arXiv:2407.09359, 2024. 1, 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "type": "text", + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 57, + 179, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 317, + 209, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 554, + 251 + ], + "type": "text", + "content": "[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 555, + 285 + ], + "type": "text", + "content": "[15] Matan Jacob Cohen and Shai Avidan. Transformally-two (feature spaces) are better than one. arXiv preprint arXiv:2112.04185, 2021. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 287, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 287, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 287, + 555, + 330 + ], + "type": "text", + "content": "[16] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 332, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 555, + 364 + ], + "type": "text", + "content": "[17] Francesco Croce, Naman D Singh, and Matthias Hein. Towards reliable evaluation and fast training of robust semantic segmentation models. In ECCV, 2024. 2, 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 366, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 554, + 398 + ], + "type": "text", + "content": "[18] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 400, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 400, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 400, + 555, + 464 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 467, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 555, + 498 + ], + "type": "text", + "content": "[20] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. arXiv preprint arXiv:2309.13415, 2023. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 501, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 555, + 533 + ], + "type": "text", + "content": "[21] Reuben Feinman, Ryan R Curtin, Saurabh Shintre, and Andrew B Gardner. Detecting adversarial samples from artifacts. arXiv preprint arXiv:1703.00410, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 555, + 578 + ], + "type": "text", + "content": "[22] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 580, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 555, + 623 + ], + "type": "text", + "content": "[23] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. arXiv preprint arXiv:2012.07177, 2020. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "type": "text", + "content": "[24] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "text", + "content": "[25] Adam Goodge, Bryan Hooi, See Kiong Ng, and Wee Siong Ng. Robustness of autoencoders for anomaly detection under adversarial impact. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 1244-1250, 2021. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20391" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[26] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "text", + "content": "[27] Jindong Gu, Hengshuang Zhao, Volker Tresp, and Philip HS Torr. Segpgd: An effective and efficient adversarial attack for evaluating and boosting segmentation robustness. In European Conference on Computer Vision, pages 308-325. Springer, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 195, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 239 + ], + "type": "text", + "content": "[28] Chuan Guo, Jacob Gardner, Yurong You, Andrew Gordon Wilson, and Kilian Weinberger. Simple black-box adversarial attacks. In International Conference on Machine Learning, pages 2484-2493. PMLR, 2019. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "type": "text", + "content": "[29] Jia Guo, Lize Jia, Weihang Zhang, Huiqi Li, et al. Recontrast: Domain-specific anomaly detection via contrastive reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 284, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 296, + 327 + ], + "type": "text", + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 296, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 381 + ], + "type": "text", + "content": "[31] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 296, + 415 + ], + "type": "text", + "content": "[32] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 296, + 449 + ], + "type": "text", + "content": "[33] Yann N. Dauphin Hongyi Zhang, Moustapha Cisse. mixup: Beyond empirical risk minimization. International Conference on Learning Representations, 2018. 4, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 296, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 482 + ], + "type": "text", + "content": "[34] Mohammad Hossein Rohban Hossein Mirzaei, ... Scanning trojaned models using out-of-distribution samples. In Advances in Neural Information Processing Systems, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 483, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 483, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 296, + 525 + ], + "type": "text", + "content": "[35] Jiande Huang, Xingxing Li, Xin Li, Jiaqi Wu, Keke Zhang, Yongqiang Yuan, and Wei Zhang. Real-time outlier detection of satellite orbit and clock products using reverse error estimation. GPS Solutions, 29(1):2, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 296, + 559 + ], + "type": "text", + "content": "[36] Nicholas Jeffrey, Qing Tan, and José R Villar. A review of anomaly detection strategies to detect threats to cyberphysical systems. *Electronics*, 12(15):3283, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 296, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 296, + 614 + ], + "type": "text", + "content": "[37] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 615, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 615, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 615, + 296, + 658 + ], + "type": "text", + "content": "[38] Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning. Advances in Neural Information Processing Systems, 33:21798-21809, 2020. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "text", + "content": "[39] Jungbeom Lee, Eunj Kim, Jisoo Mok, and Sungroh Yoon. Anti-adversarily manipulated attributions for weakly supervised semantic segmentation and object localization. IEEE transactions on pattern analysis and machine intelligence, 46 (3):1618-1634, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[40] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9664-9674, 2021. 3, 4, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "text", + "content": "[41] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. Musc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. arXiv preprint arXiv:2401.16753, 2024. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 555, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 555, + 216 + ], + "type": "text", + "content": "[42] Jiaqi Liu, Guoyang Xie, Jinbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. Machine Intelligence Research, 21(1):104-135, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 555, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 250 + ], + "type": "text", + "content": "[43] Ye Liu, Yaya Cheng, Lianli Gao, Xianglong Liu, Qilong Zhang, and Jingkuan Song. Practical evaluation of adversarial robustness via adaptive auto attack, 2022. 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 555, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 304 + ], + "type": "text", + "content": "[44] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 304, + 555, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 555, + 338 + ], + "type": "text", + "content": "[45] Shao-Yuan Lo, Poojan Oza, and Vishal M Patel. Adversarily robust one-class novelty detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 339, + 555, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 555, + 371 + ], + "type": "text", + "content": "[46] Pauline Luc, Camille Couprie, Soumith Chintala, and Jakob Verbeek. Semantic segmentation using adversarial networks. arXiv preprint arXiv:1611.08408, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 555, + 415 + ], + "type": "text", + "content": "[47] Yuan Luo, Ya Xiao, Long Cheng, Guojun Peng, and Danfeng Yao. Deep learning-based anomaly detection in cyberphysical systems: Progress and opportunities. ACM Computing Surveys (CSUR), 54(5):1-36, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 555, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 555, + 459 + ], + "type": "text", + "content": "[48] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 555, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 555, + 492 + ], + "type": "text", + "content": "[49] Xiaofeng Mao, Yuefeng Chen, Shuhui Wang, Hang Su, Yuan He, and Hui Xue. Composite adversarial attacks. ArXiv, abs/2012.05434, 2020. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 493, + 555, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 555, + 548 + ], + "type": "text", + "content": "[50] Hossein Mirzaei, Mohammadreza Salehi, Sajjad Shahabi, Efstratios Gavves, Cees GM Snoek, Mohammad Sabokrou, and Mohammad Hossein Rohban. Fake it till you make it: Near-distribution novelty detection by score-based generative models. arXiv preprint arXiv:2205.14297, 2022. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 548, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 624 + ], + "type": "text", + "content": "[51] Hossein Mirzaei, Mohammad Jafari, Hamid Reza Dehbashi, Ali Ansari, Sepehr Ghobadi, Masoud Hadi, Arshia Soltani Moakhar, Mohammad Azizmalayeri, Mahdieh Soleymani Baghshah, and Mohammad Hossein Rohban. Rodeo: Robust outlier detection via exposing adaptive out-of-distribution samples. In Forty-first International Conference on Machine Learning, 2024. 1, 2, 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 555, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 555, + 691 + ], + "type": "text", + "content": "[52] Hossein Mirzaei, Mohammad Jafari, Hamid Reza Dehbashi, Zeinab Sadat Taghavi, Mohammad Sabokrou, and Mohammad Hossein Rohban. Killing it with zero-shot: Adversarially robust novelty detection. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7415-7419. IEEE, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 714 + ], + "type": "text", + "content": "[53] Hossein Mirzaei, Mojtaba Nafez, Mohammad Jafari, Mohammad Bagher Soltani, Mohammad Azizmalayeri, Jafar Habibi," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20392" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 297, + 127 + ], + "type": "text", + "content": "Mohammad Sabokrou, and Mohammad Hossein Rohban. Universal novelty detection through adaptive contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22914-22923, 2024. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 297, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 297, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 297, + 161 + ], + "type": "text", + "content": "[54] Arian Mousakhan, Thomas Brox, and Jawad Tayyyub. Anomaly detection with conditioned denoising diffusion models. arXiv preprint arXiv:2305.15956, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 229 + ], + "type": "text", + "content": "[55] Fahimeh Hosseini Noohdani, Parsa Hosseini, Aryan Yazdan Parast, Hamidreza Yaghoubi Araghi, and Mahdieh Soleymani Baghshah. Decompose-and-compose: A compositional approach to mitigating spurious correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27662-27671, 2024. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 262 + ], + "type": "text", + "content": "[56] Pramuditha Perera, Poojan Oza, and Vishal M Patel. One-class classification: A survey. arXiv preprint arXiv:2101.03064, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "text", + "content": "[57] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 296, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 296, + 353 + ], + "type": "text", + "content": "[58] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. arXiv preprint arXiv:2110.14051, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 399 + ], + "type": "text", + "content": "[59] Hannah M Schlüter, Jeremy Tan, Benjamin Hou, and Bernhard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In European Conference on Computer Vision, pages 474-489. Springer, 2022. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 296, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 454 + ], + "type": "text", + "content": "[60] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "type": "text", + "content": "[61] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVII 16, pages 682-698. Springer, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 511, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 296, + 554 + ], + "type": "text", + "content": "[62] Abhishek Sinha, Kumar Ayush, Jiaming Song, Burak Uzkent, Hongxia Jin, and Stefano Ermon. Negative data augmentation. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 296, + 600 + ], + "type": "text", + "content": "[63] Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister. Learning and evaluating representations for deep one-class classification. arXiv preprint arXiv:2011.02578, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 296, + 645 + ], + "type": "text", + "content": "[64] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Generalad: Anomaly detection across domains by attending to distorted features. arXiv preprint arXiv:2407.12427, 2024. 1, 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "text", + "content": "[65] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[66] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distribu" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "tionally shifted instances. Advances in neural information processing systems, 33:11839-11852, 2020. 1, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "text", + "content": "[67] Jeremy Tan, Benjamin Hou, James Batten, Huaqi Qiu, and Bernhard Kainz. Detecting outliers with foreign patch interpolation. Machine Learning for Biomedical Imaging, 1(April 2022):1-27, 2022. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 196 + ], + "type": "text", + "content": "[68] Maximilian E Tschuchnig and Michael Gadermayr. Anomaly detection in medical imaging-a mini review. In Data Science-Analysics and Applications: Proceedings of the 4th International Data Science Conference-iDSC2021, pages 33-38. Springer, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 554, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 240 + ], + "type": "text", + "content": "[69] Qi Wang, Junyu Gao, and Xuelong Li. Weakly supervised adversarial domain adaptation for semantic segmentation in urban scenes. IEEE Transactions on Image Processing, 28 (9):4376-4386, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 296 + ], + "type": "text", + "content": "[70] Weiping Wang, Zhaorong Wang, Zhanfan Zhou, Haixia Deng, Weiliang Zhao, Chunyang Wang, and Yongzhen Guo. Anomaly detection of industrial control systems based on transfer learning. Tsinghua Science and Technology, 26(6): 821-832, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 331 + ], + "type": "text", + "content": "[71] Yue Xing, Qifan Song, and Guang Cheng. Why do artificially generated data help adversarial robustness. Advances in Neural Information Processing Systems, 35:954-966, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 365 + ], + "type": "text", + "content": "[72] Han Xu, Yao Ma, Haochen Liu, Debayan Deb, Hui Liu, Jiliang Tang, and Anil K. Jain. Adversarial attacks and defenses in images, graphs and text: A review, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 555, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 399 + ], + "type": "text", + "content": "[73] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, pages 1-28, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 453 + ], + "type": "text", + "content": "[74] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. arXiv preprint arXiv:2406.07487, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 555, + 498 + ], + "type": "text", + "content": "[75] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 500, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 555, + 544 + ], + "type": "text", + "content": "[76] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE transactions on neural networks and learning systems, 30(9): 2805-2824, 2019. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 545, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 555, + 589 + ], + "type": "text", + "content": "[77] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features, 2019. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 555, + 644 + ], + "type": "text", + "content": "[78] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 678 + ], + "type": "text", + "content": "[79] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Dsr – a dual subspace re-projection network for surface anomaly detection, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "text", + "content": "[80] Chaoning Zhang, Philipp Benz, Chenguo Lin, Adil Karjauv, Jing Wu, and In So Kweon. A survey on universal adversarial attack. arXiv preprint arXiv:2103.01498, 2021. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20393" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 115 + ], + "type": "text", + "content": "[81] Hui Zhang, Zheng Wang, Zuxuan Wu, and Yu-Gang Jiang. Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection. arXiv preprint arXiv:2303.08730, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[82] Lin Zhang and Yang Dai. A gan anomaly detection method based on multi-scale endogenous enhancement. In Blockchain and Web3 Technology Innovation and Application Exchange Conference, pages 269-281. Springer, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "text", + "content": "[83] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20394" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/7794bc7e-86d2-49a8-8adc-3430e39dd287_content_list.json b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/7794bc7e-86d2-49a8-8adc-3430e39dd287_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4c3d77a9aed8420a4054701f9fb741b9e232a14e --- /dev/null +++ b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/7794bc7e-86d2-49a8-8adc-3430e39dd287_content_list.json @@ -0,0 +1,1667 @@ +[ + { + "type": "text", + "text": "PatchVSR: Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution", + "text_level": 1, + "bbox": [ + 194, + 128, + 803, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shian Du $^{1*‡}$ , Menghan Xia $^{2*‡}$ , Chang Liu $^{1}$ , Xintao Wang $^{2}$ , Jing Wang $^{3}$ , Pengfei Wan $^{2}$ , Di Zhang $^{2}$ , Xiangyang Ji $^{1*}$", + "bbox": [ + 158, + 208, + 836, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University, $^{2}$ Kuaishou Technology, $^{3}$ Beijing Institute of Technology", + "bbox": [ + 187, + 248, + 805, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 315, + 326, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pre-trained video generation models hold great potential for generative video super-resolution (VSR). However, adapting them for full-size VSR, as most existing methods do, suffers from unnecessary intensive full-attention computation and fixed output resolution. To overcome these limitations, we make the first exploration into utilizing video diffusion priors for patch-wise VSR. This is non-trivial because pre-trained video diffusion models are not native for patch-level detail generation. To mitigate this challenge, we propose an innovative approach, called PatchVSR, which integrates a dual-stream adapter for conditional guidance. The patch branch extracts features from input patches to maintain content fidelity while the global branch extracts context features from the resized full video to bridge the generation gap caused by incomplete semantics of patches. Particularly, we also inject the patch's location information into the model to better contextualize patch synthesis within the global video frame. Experiments demonstrate that our method can synthesize high-fidelity, high-resolution details at the patch level. A tailor-made multi-patch joint modulation is proposed to ensure visual consistency across individually enhanced patches. Due to the flexibility of our patch-based paradigm, we can achieve highly competitive 4K VSR based on a $512 \\times 512$ resolution base model, with extremely high efficiency.", + "bbox": [ + 89, + 348, + 485, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 758, + 220, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video super-resolution (VSR) has long been a challenging research problem in the field of computer vision. With the advent of deep learning, significant advancements have been made in end-to-end VSR models trained on curated datasets. However, due to limited model capacity and data", + "bbox": [ + 89, + 785, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "coverage, these models typically achieve sharp object structures or coarse textures but struggle to synthesize faithful details. Recently, the emergence of diffusion generative models has enabled open-domain video generation of remarkable quality. The generative capabilities of these foundation models present new opportunities for video super-resolution techniques [16, 55]. Nevertheless, existing base models are generally constrained to generating videos of fixed sizes due to their inherent architecture. This limitation is inherited in their downstream super-resolution models, making it challenging for them to support arbitrary resolution outputs.", + "bbox": [ + 511, + 301, + 906, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This study is motivated by the insight that feature attention in super-resolution models is more locally focused compared to generation models. This localization is reasonable, as super-resolution tasks have a low-resolution video as a reference, allowing details to be generated based on neighborhood semantics rather than requiring intensive global coherence considerations. To leverage this insight, we propose a patch-based strategy for video superresolution. The concept is straightforward and intuitive: we split the input video into patches sized to be compatible with the pre-trained base model, enhance each patch individually, and then combine them to produce the final result. This approach enables the achievement of super-resolution results at arbitrary resolutions, independent of the base model's inherent resolution constraints.", + "bbox": [ + 511, + 488, + 908, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The key challenge lies in that video generation-based models are trained on complete video frames and face noticeable performance degradation for patch-level generation. To adapt a pre-trained video generation model for generating patch-level details in super-resolution, we propose an innovative approach, called PatchVSR, which integrates a dual-stream adapter for conditional guidance. The local branch extracts features from the input patches to guide the base model for detail synthesis adhering to the input, while the global branch captures contextual information from the entire input frame to bridge the generation gap caused by incomplete semantics of patches. To make the context guid-", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during an internship at KwaiVGI, Kuaishou Technology.", + "bbox": [ + 112, + 875, + 482, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Equal contribution. $\\dagger$ Corresponding author.", + "bbox": [ + 112, + 887, + 359, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17799", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ance more customized, the global branch also incorporates a binary mask indicating the target patch position as input. Additionally, we employ LoRA technology to fine-tune the base model, further improving its adaptability for generating patch-level details.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results demonstrate that our method effectively adapts the generative ability of the base model to produce faithful details at the patch level. Specifically, to alleviate the boundary inconsistency across patches, we propose a training-free multi-patch joint modulation scheme, where auxiliary bridging patches are created to allow feature interaction across patches during generation. Thanks to the flexibility of PatchVSR, our method, based on a pre-trained $512 \\times 512$ resolution video base model, supports super-resolution video generation up to 4K resolution, while significantly outperforming current state-of-the-art methods in terms of fidelity and detail realism. Notably, compared to existing methods that operate on full-frame videos, our approach processes relatively smaller patches, without full-range attention computation, demonstrating a significant efficiency advantage.", + "bbox": [ + 89, + 167, + 482, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 109, + 411, + 413, + 425 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce an innovative approach to overcome the resolution limitations of using pre-trained video diffusion models for VSR. Our method achieves any-resolution VSR output based on a fixed-resolution base model.", + "- We propose a novel dual-branch adapter paradigm that effectively harnesses the capabilities of pre-trained video models for patch-level detail generation.", + "- Through extensive experimentation, we demonstrate the significant superiority of our method in terms of faithful detail generation and computational efficiency." + ], + "bbox": [ + 99, + 426, + 482, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 594, + 232, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-video Generation", + "text_level": 1, + "bbox": [ + 89, + 619, + 318, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the widespread use of pre-trained text-to-image models [10, 13, 18, 32, 39, 52] and the preparation of large-scale text-video pairing datasets [9, 34], substantial efforts have been made to train a large-scale text-to-video model. Instead of training from scratch, some methods [2, 41, 46, 51] focus on inserting temporal modules and fine-tuning them or all parameters on video data. To avoid catastrophic forgetting and degradation on visual quality, they often adopt image-video joint training strategy [7, 8, 47]. Recently, with Diffusion Transformer (DiT) [36] being proposed, it has replaced the traditional 3D UNet architecture used by previous approaches with better generation quality and scalability, and has become the dominant framework for text-to-video generation [28, 30, 49]. Although achieving high visual quality, extending them to higher-resolution is non-trivial since it requires substantial training resource and high-quality high-resolution video dataset, which is un", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "available by this moment. To bypass this difficulty, current methods often adopt a cascade framework [15, 44, 54]: the low-resolution video is first generated by the base model using text prompt, and the refinement model outputs high-resolution video through a noisig-denoising process [31] by taking the low-resolution video as condition. However, existing methods are limited to generate full video at once, which still demands a high training overhead. At the same time, the low inference efficiency and high memory usage also hinder its downstream applications.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Video Super-Resolution", + "text_level": 1, + "bbox": [ + 511, + 257, + 732, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video super-resolution aims at restoring high-resolution (HR) video frames from its degraded low-resolution (LR) counterparts and enhancing its visual quality. Traditional methods often assume a fixed degradation process and their performance deteriorates significantly when handling AI-generated and real-world videos. Due to the lack of real-world LR-HR video data pair for training, RealVSR [48] collects LR-HR video data pair by utilizing the multicamera system of iPhone. Nevertheless, the sequences are captured by a single device, which only contains camera-specific degradations. To avoid burdensome data collection, other methods [6, 22, 55] use stochastic combinations of multi-order degradation processes to simulate real-world degradation processes. While performing more robust when handling real-world cases, due to the absence of deep generative prior, it is challenging for traditional CNN-based [4-6, 21, 43] and transformer-based [3, 14, 25, 40] methods to generate photo-realistic details and textures. By utilizing a pre-trained image diffusion model like Stable Diffusion (SD) $\\times 4$ Upscaler, Lavie-SR [44] and Upscale-A-Video [55] circumvent the need for exhaustive training from scratch and exhibit improved performance in fixed spatial scale within a short sequence. To improve temporal consistency for long sequence, I2VGen-XL [54] and VEnhancer [16] utilize a pre-trained low-resolution video generative model as the base model, and directly fine-tune the model parameters using high-resolution video dataset following noisy-denoising manner [31]. However, the aforementioned methods are all limited to generate full video at once, which requires expanding the training resolution of pre-trained model, bringing low training efficiency and slow convergence speed. Meanwhile, high resolution pretrained video generative models are currently unavailable to access, hindering the scalability of video super-resolution methods with deep generative prior. To tackle this problem, we make the first exploration of utilizing pre-trained video generative model for patch-wise video super-resolution. By maintaining the training resolution of the pre-trained model, our approach not only avoids significant training overhead, making it possible to generate 4K resolution videos using a $512 \\times 512$ resolution pre-trained model. At the same time,", + "bbox": [ + 511, + 281, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17800", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3a130280773855dd3be87b1e9a0c1c055c7a493d43d4caee1aa4059ad7141be2.jpg", + "image_caption": [ + "Figure 1. Flowchart of our PatchVSR. Building upon a pre-trained latent T2V model, we incorporate a patch condition branch and a global context branch. These branches extract features from partitioned video patches and the resized full video together with a binary mask that indicates the location of the ROI patch, respectively. Particularly, local patch features are added to the output of each block, while the global context feature is fused with the backbone feature through newly introduced cross-attention modules (G-CA). For simplicity, we have omitted other conditional inputs such as text prompts and time steps from this diagram. The processed patches are fused via a joint modulation scheme to produce a coherent super-resolution video." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 908, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "our method significantly improves inference efficiency and reduces memory usage, facilitating downstream applications of video super-resolution methods.", + "bbox": [ + 88, + 387, + 486, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 450, + 184, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to develop a video super-resolution (VSR) framework designed to restore degraded videos of arbitrary resolutions by generating visually coherent and structurally faithful details. Our primary innovation involves strategically adapting pre-trained text-to-video (T2V) diffusion models—powerful generative tools typically constrained to fixed output size (denoted as $s = h \\times w$ )—through a curated conditioning paradigm. To bypass their inherent token-length limitations, we introduce a patch-based processing strategy: the input video is decomposed into $s$ -sized patches, each independently enhanced using the diffusion model, and then reintegrated into a unified high-resolution output through optimized spatial-temporal fusion. This approach is grounded in the distinct characteristics of VSR compared to video generation. While video synthesis demands global coherence across full-frame contexts, superresolution prioritizes localized detail restoration atop existing structural content.", + "bbox": [ + 88, + 476, + 482, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we augment the base model with two conditional branches: a patch branch that extracts features from input patches, and a global branch that extracts context from the full video. This global context mainly aids in generating patch-level details from the T2V model that was originally trained on full-frame videos. An overview of the model architecture is illustrated in Fig. 1. Given a low-resolution image $\\mathbf{V}_l\\in \\mathbb{R}^{F\\times H\\times W\\times 3}$ and the target upscale factor $k$ , our process begins by upscaling $\\mathbf{V}_l$ to the desired resolution $\\mathbf{V}_l^\\uparrow \\in \\mathbb{R}^{F\\times [H\\cdot k]\\times [W\\cdot k]\\times 3}$ using bicubic", + "bbox": [ + 88, + 750, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "interpolation per frame. We then partition $\\mathbf{V}_l^\\dagger$ into patches $\\{\\mathbf{P}_i\\in \\mathbb{R}^{F\\times h\\times w\\times 3}\\}_{i = 1}^N$ , where the patch dimensions $(h,w)$ match the generation size of the pre-trained T2V model. Additionally, we resize the full video to $\\mathbf{G}\\in \\mathbb{R}^{F\\times h\\times w\\times 3}$ to serve as global guidance. Finally, both $\\mathbf{P}_i$ and $\\mathbf{G}$ are input into the VSR model for patch-level detail enhancement, and multi-patch joint modulation is involved to ensure coherence across patches. This formulation ensures compatibility with the pre-trained T2V model while allowing for flexible processing of videos of arbitrary sizes.", + "bbox": [ + 511, + 387, + 906, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Video Diffusion Base Model", + "text_level": 1, + "bbox": [ + 511, + 547, + 870, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model is implemented on top of a pre-trained T2V diffusion model, which adopts the Rectified Flow framework [11] for the noise schedule and denoising process. The forward process is defined as straight paths between data distribution and a standard normal distribution, i.e.", + "bbox": [ + 511, + 569, + 906, + 645 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = (1 - t) z _ {0} + t \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 656, + 906, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\in \\mathcal{N}(0,I)$ and $t$ denotes the iterative timestep. To solve the denoising processing, we define a mapping between samples $z_{1}$ from a noise distribution $p_1$ to samples $z_{0}$ from a data distribution $p_0$ in terms of an ordinary differential equation (ODE), namely:", + "bbox": [ + 511, + 681, + 908, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd z _ {t} = v _ {\\Theta} (z _ {t}, t) d t, \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 768, + 906, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the velocity $v$ is parameterized by the weights $\\Theta$ of a neural network. For inference, we employ Euler discretization for Eq. 2 and perform discretization over the timestep interval at [0,1], starting at $t = 1$ . We then processed with iterative sampling with: $z_{t_i} = z_{t_j} + v_\\Theta (z_{t_j},t_j)*(t_j - t_i)$ .", + "bbox": [ + 511, + 794, + 908, + 872 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For consideration of computing efficiency, the mapping is formulated in a latent space constructed with a", + "bbox": [ + 511, + 871, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17801", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Variational Auto-Encoder (VAE) [24]. The core is a Transformer-based diffusion model (DiT) [35], where each block is instantiated as a sequence of spatial attention, temporal attention, and cross-attention modules. Text prompts and other micro parameters (like time step, aspect ratio, etc) are used as generative conditions. Due to the full-attention property of Transformer, our model only supports the generation of a fixed number of tokens as it was trained with, i.e. $512 \\times 512$ of diverse aspect-ratio. This is a common characteristic of almost all existing diffusion models for image or video generation [7, 8, 37, 49], and extending pre-trained diffusion models for higher resolution generation remains an open research topic [15, 17]. For VSR as a downstream task, we propose a patch-wise processing strategy to break this resolution limit.", + "bbox": [ + 89, + 90, + 483, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Patch Video Diffusion with Global Guidance", + "text_level": 1, + "bbox": [ + 89, + 327, + 467, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To adapt a pre-trained T2V model for patch-level VSR, we must consider two key aspects: (i) enabling the input video to guide the base model's denoising process and (ii) bridging the gap between the base model's capabilities in generating full-frame content and patch-level content synthesis requirements. We achieve this through a dual-branch adapter, as detailed below.", + "bbox": [ + 89, + 349, + 483, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Patch Condition Branch. There are generally two approaches to incorporating visual conditions into pre-trained generative models: concatenating with noisy data in the noise space [39, 55], or fusing with the backbone feature through an adapter branch [16, 42]. The former involves modifying the first layers' channel count and necessitates base model fine-tuning, which is typically more data and computationally intensive. In contrast, the latter is more lightweight, as only a newly added branch requires training for feature adaptation. We adopt a hybrid solution that combines elements of both approaches. First, we employ an adapter comprising several Transformer blocks to extract features from the input video patch $\\mathbf{P}_i$ and inject them into the base model's blocks. Additionally, considering that the base model was trained for full video generation, we need to adapt its embedded knowledge for patch-level generation. To this end, we allow the base model to be fine-tuned using a Low-Rank Adaptation (LoRA) [19]. An ablation study in Sec. 4.3 demonstrates the advantages of this design.", + "bbox": [ + 89, + 473, + 483, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global Context Branch. To further guide content generation in a natural manner, we incorporate a global branch that extracts context from the full input video. This branch utilizes a Transformer-based encoder architecture that reduces the number of tokens to one quarter. The resulting context features are fused with each block of the base model through cross-attention $\\{\\mathbf{Q},\\mathbf{K}_g,\\mathbf{V}_g\\}$ , where the $\\mathbf{Q}$ is shared with the existing text-prompt cross-attention", + "bbox": [ + 89, + 780, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c616b517cf3fa13d021b21cb0439b0853bb1f1b36bf469ab38f991ff07c59f86.jpg", + "image_caption": [ + "Patch Partition" + ], + "image_footnote": [], + "bbox": [ + 517, + 107, + 651, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/96042419c5e573aa0a636f9f49339d24cfda762f719df1c5d8c7826ebcbfcd35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 122, + 689, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/be6d6bf131eb8eb88408e25c5fc7a2de9ad7920ce5dc9e69a7f19ba32477145c.jpg", + "image_caption": [ + "Created auxiliary patches", + "Figure 2. Patch Partition Visualization. The input video is divided into non-overlapping segments, as the solid blue boxes mark. For joint modulation, auxiliary patches are created, indicated by the red dashed boxes, resulting in an overlapping ratio of $50\\%$ ." + ], + "image_footnote": [], + "bbox": [ + 697, + 89, + 903, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\{\\mathbf{Q}, \\mathbf{K}_t, \\mathbf{V}_t\\}$ . Specifically, the input is constructed by concatenating the resized full video $\\mathbf{V}_g$ (to accommodate a fixed number of tokens) with a binary map $\\mathbf{M}_i$ indicating the target patch's $\\mathbf{P}_i$ location within it. Our experiments show that the global context branch not only enhances the vividness of generated patch-level details but also improves consistency across patches.", + "bbox": [ + 511, + 277, + 906, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-Patch Joint Modulation", + "text_level": 1, + "bbox": [ + 513, + 395, + 779, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the global guidance shared by patches is critical for maintaining color consistency, tonal uniformity, and semantic coherence of details, naive stitching of independently enhanced patches often introduces visible seams. These boundary artifacts primarily stem from the stochastic nature of the generative process when hallucinating textures on degraded or underdetermined regions. This is a fundamental challenge in ill-posed inverse problems where multiple high-resolution solutions may satisfy the constraints of a single low-resolution input.", + "bbox": [ + 511, + 417, + 906, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate this problem, we introduce inter-patch modulation during the detail generation process. Specifically, a training-free inference strategy called multi-diffusion [1] is adapted to solve our problem. This approach involves splitting the original video into overlapping patches and updating the overlapping regions by averaging contributions from all relevant patches at each denoising step. As illustrated in Fig. 2, instead of directly splitting the video into overlapping patches, we first partition it into non-overlapping segments, and then construct auxiliary patches by combining the halves of adjacent patches. This creates a new set of patch groups with an overlapping ratio of $50\\%$ . Our experiment shows that directly taking the average in overlap regions suffers from artifacts of black holes or seamlines, as evidenced in Fig. 5 (middle). Instead, when fusing base patches and its related auxiliary patches, we manage to gradually reduce its influence of auxiliary patches from its center splitting line to two sides by assigning spatial weight maps to them. This method effectively reduces visible seams between patches and promotes overall consistency. A comparative result demonstrating the effectiveness of this approach is presented in Fig. 5.", + "bbox": [ + 511, + 569, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17802", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/68ad9643dbcebd1718f9046f9a080576c2a1b0cf33ee8615283c380f45f13e45.jpg", + "image_caption": [ + "Video & prompt" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 310, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/41475bfb1abf29ce9fdaa15823cd234524938f83cd36b1bff415fd29e324becd.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 357, + 125, + 491, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/628c51588e7aac50f9a562daaf57daebb190c7608983ee483bb1200da80ed5e8.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 94, + 198, + 225, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e8f7cf94e7d802ec39a5b83809ced7a0e2df332049e63e37d1029f4174d6cdc0.jpg", + "image_caption": [ + "LaVie-SR" + ], + "image_footnote": [], + "bbox": [ + 227, + 198, + 357, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b2c429bd8f612d3ca6d489d5a21879a43a9bd6513f523e8d72452bbf2441801.jpg", + "image_caption": [ + "VEnhancer" + ], + "image_footnote": [], + "bbox": [ + 359, + 198, + 491, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7fbd124ab6af09870e0450fc95051178252c106723c79b6b0ce19f01e2853695.jpg", + "image_caption": [ + "RealBasicVSR" + ], + "image_footnote": [], + "bbox": [ + 94, + 273, + 225, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af4295e2bcd9827b1955e60443334d167e36615fee2c38ab5698b12b70747421.jpg", + "image_caption": [ + "Upscale-A-Video" + ], + "image_footnote": [], + "bbox": [ + 227, + 273, + 357, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2b6f68ca39a7a522a01831a947ed9aad379c1765b505f213c9f1984a0c71788e.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 359, + 273, + 491, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2521f2f0580ae6c73908a06cf11dbde1149d0d95bbddabc91e1232759a155ee2.jpg", + "image_caption": [ + "Video & prompt" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 717, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/948f1cbbaebca6221de677ac87e4f76343270e3c2c9bd01100748c2b08a8d279.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 769, + 125, + 903, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eae1ad88aa5e70662ef692def10460a768bcc062a6fe019d85330ec21fb08583.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 504, + 198, + 635, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0ad2c1c1ddcb8c0e5b27e6078c2731998b57e600f9373c1ce438d3b4be21f5b4.jpg", + "image_caption": [ + "LaVie-SR" + ], + "image_footnote": [], + "bbox": [ + 638, + 198, + 769, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dbff98e7545b6b10e99f21e6d2eddc771339dae2ca727d05f95a34a2b5545711.jpg", + "image_caption": [ + "VEnhancer" + ], + "image_footnote": [], + "bbox": [ + 771, + 198, + 901, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d3b18950f91bc25cf0fccef143ae28b91c938aa251353ae863d160c53929acf5.jpg", + "image_caption": [ + "RealBasicVSR" + ], + "image_footnote": [], + "bbox": [ + 504, + 273, + 635, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f85f09b0a3cb06cd5b871f369f2157d4abce0e2c60bcd299f61dbb6416293277.jpg", + "image_caption": [ + "Upscale-A-Video", + "Figure 3. Qualitative comparisons on SynVideo30 (test). These videos $\\times 4$ super-resolution results. Zoom in for best view." + ], + "image_footnote": [], + "bbox": [ + 638, + 273, + 769, + 332 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c9f434d578b5247fe5eb3b717b2bcbaf8329fbd7082a3df20a0db0863a7e5ecf.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 771, + 273, + 901, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 383, + 222, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To verify the superiority of PatchVSR, we compare it with SOTA methods, including VEnhancer [16], Upscale-A-Video [55], LaVie-SR [44] and RealBasicVSR [6]. We also make ablation study to analyze the effectiveness of our proposed designs. Besides, some extended analysis is conducted to demonstrate the capability of our method.", + "bbox": [ + 89, + 407, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 508, + 307, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Datasets We train our model using 460K self-collected high-quality video-text pairs, with each resolution from $1024 \\times 1024$ to 2K. The text prompts are all captioned using LAVA captioner [26], and encoded by T5 text encoder [38] with no more than 256 tokens.", + "bbox": [ + 89, + 529, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Testing Datasets. We evaluate the VSR methods on two kinds of dataset, i.e. downscale synthetic videos and AIGC videos. For synthetic videos, since our employed base model is fixed to generate 77-frames video, we utilize a public dataset: REDS30 [33] for fair comparisons, which contains 100 frames. To evaluate higher resolution, we self-collect 30 high-quality 2K resolution videos, each containing more than 150 frames. Additionally, we collect and test on an AIGC dataset, which comprises 30 AI-generated videos from public text-to-video generative models [50].", + "bbox": [ + 89, + 617, + 483, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. Our model is trained on 16 NVIDIA H800-80G GPUs with a total batch size of 64. AdamW [29] is used as the optimizer with a learning rate of $10^{-5}$ . The text prompt is randomly replaced by a null prompt with $10\\%$ probability. To obtain the LR-HR training data pair, the HR video is first down-sampled using bilinear interpolation with down scale $s$ and up-sampled back to the original resolution. To match the size of the pre-trained model,", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we randomly crop a $512 \\times 512$ area from the up-sampled LR video with aspect ratios from 0.5 to 2.0. To enhance the robustness of our model to different degradation scenarios, we utilize the noise augmentation technique by injecting noise into the input latent using a diffuse process. The noise timestep is randomly selected from 200 to 300, aiming at preserving the structure of the input video. We have additionally encoded noise timestep, downscale factor, and crop location as conditions for the model. We have replaced the text prompt with a fixed prompt for the patch condition branch, with the base model and global context branch using text prompts. To provide initialization weight, we followed the structure of the pre-trained model and sampled the layer index from it at equal intervals.", + "bbox": [ + 511, + 385, + 906, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference Details. We perform 50 PNDM [27] sampling steps with classifier-free guidance scale 5.0. We have also used timestep shift [12] with shift value 5.0.", + "bbox": [ + 511, + 611, + 905, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the fidelity of our method on synthesized videos, we utilize PSNR, SSIM, and LPIPS [53]. For visual quality, we conduct our evaluation of commonly used quality assessment metrics MUSIQ [23] and DOVER [45]. Furthermore, we also include aesthetic score from a video generation benchmark VBench [20] for more comprehensive comparisons.", + "bbox": [ + 511, + 671, + 906, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with SOTA methods", + "text_level": 1, + "bbox": [ + 511, + 787, + 803, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Results. The visual results of synthetic and AIGC videos are presented in Fig. 3 and Fig. 4 respectively. It is observed that our method significantly outperforms existing CNN-based and diffusion-based models in both details generation and temporal consistency (Readers are recommended to check our provided video results). Specif", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17803", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bc96866090653a2c3deb647fd4601848388ecac12aa9a33925484016c57aa71e.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons on VideoGen30 (test). These videos are $\\times 4$ super-resolution results. Zoom in for best view." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 906, + 358 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/307b9903a91931a39689d91719f7c38ff69adf4d204140d3c0397d62b6036f45.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsMetricsRealBasicVSR [6]LaVie-SR [44]Upscale-A-Video [55]VEnhancer [16]Ours
REDS30PSNR ↑29.69828.73531.25728.54430.111
SSIM ↑0.6560.6220.6610.6110.647
LPIPS ↓0.3340.4070.3280.4210.326
DOVER ↑0.3050.3180.3100.3440.323
MUSIQ ↑40.58743.57742.57544.53344.956
Aesthetics ↑0.4220.4860.4710.4820.493
SynVideo30PSNR ↑33.50731.43333.43228.85630.857
SSIM ↑0.7760.7160.7280.6970.732
LPIPS ↓0.1850.1940.2050.1990.183
DOVER ↑0.5840.6120.5290.5500.597
MUSIQ ↑49.55748.14649.83943.53850.695
Aesthetics ↑0.4960.5090.4940.5030.520
VideoGen30DOVER ↑0.5880.6090.5100.5120.590
MUSIQ ↑47.56448.32946.89946.93050.559
Aesthetics ↑0.5470.6000.5740.5510.602
", + "bbox": [ + 91, + 388, + 910, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons on different VSR benchmarks from synthetic (REDS30, SynVideo30) and AIGC (VideoGen30) data. Red and blue numbers indicate the best and second best results respectively.", + "bbox": [ + 89, + 654, + 906, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ically, our method not only successfully recovers the details of the red silk scarf in Fig. 3, but also adds more details to the \"Angkor Wat\" as shown in Fig. 4, which others tend to generate blurry or unnatural results. In Fig. 5, we also illustrate full VSR results achieved through the multi-patch joint modulation scheme. It can be seen that creating auxiliary patches greatly alleviates the seam problem and significantly reduces the color difference between different patches, while further improving the contrast of the results. To show the advantage of our multi-patch joint modulation, we compare with a baseline method: stitching the patches in latent space and decoding the big latent together, which suffers obvious seamline artifacts.", + "bbox": [ + 88, + 696, + 485, + 893 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. The quantitative comparison is shown in Tab. 1. It can be seen that the proposed method achieves the highest LPIPS score in both REDS30 and SynVideo30, indicating the high fidelity of our method for synthetic videos. However, our method show inferior performance in PSNR and SSIM. This can be explained by the fact that our method tends to generate rich and vivid details, which are not favorable to metrics that capture pixel-wise deviations (i.e. PSRN and SSIM). Instead, as a perceptual-level fidelity measure, LPIPS might be more informative to reflect the content fidelity for generative tasks. We evaluate the visual quality of generated high-resolution videos through comparisons with competing methods on MUSIQ", + "bbox": [ + 511, + 696, + 908, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17804", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5cd909b4dea0786c2a3bf31b0afc3166e930ba7fac770e6279b9dec80f8f5416.jpg", + "image_caption": [ + "Figure 5. Visual evaluation on multi-patch joint modulation. Here the super-resolved videos of 2K resolution are used for comparison, where $3 \\times 3 = 9$ patches are involved. The top, middle, bottom denote input video, result of stitching latent of patches, and result of ours respectively." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 478, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and DOVER metrics, demonstrating our approach's superiority. This performance reflects contributions from both the base model's capabilities and the effectiveness of our method in leveraging full-frame generation for patch-level detail synthesis. Interestingly, our approach even preserves subjective properties like aesthetic quality from the base model, as evidenced by the aesthetic scores. This highlights our framework's ability to inherit and maintain critical perceptual characteristics during the enhancement process.", + "bbox": [ + 89, + 553, + 482, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 699, + 243, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies on key designs. Considering the limited space, we provide a quantitative evaluation here and the typical visual results are compared in the supplementary material.", + "bbox": [ + 89, + 720, + 482, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Global Branch. To validate the effectiveness of the global branch, We build a baseline by removing the global branch from our method. As shown in Tab. 2, removing the global branch results in significant performance deterioration, leading to worse DOVER, MUSIQ, and aesthetic score. Besides, to check the necessity of using the patch location embeddings, we build a baseline by removing the", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9b2166f0e1976fcb023f6b3ff311d19b6ac7e767f03b8f8543a900d8a4e162a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ComponentDOVER ↑MUSIQ ↑Aesthetics ↑
w/o global branch0.50246.0740.589
w/o LoRA0.58250.4960.600
w/o location embed0.57450.0840.601
w/o fixed prompt0.56248.1330.597
Ours0.59050.5590.602
", + "bbox": [ + 517, + 88, + 903, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study of different model components on VideoGen30 test dataset. The **bold** numbers indicate the best.", + "bbox": [ + 511, + 199, + 906, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "patch position information from the global branch input. As shown in Tab. 2, after incorporating the location embedding, we obtain an increase of DOVER from 0.574 to 0.590, which demonstrates the effectiveness of utilizing the location mask as the patch position information.", + "bbox": [ + 511, + 234, + 906, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text Prompts. Text prompts are used in the base model and the conditional global branch. As shown in Tab. 2, we observe that replacing the text prompt with a fixed prompt leads to superior results. We conjecture that the use of text prompts in patch condition branch brings about a mismatch between the video patch and the global semantic information, which in turn destroys the dynamic visual prior in the pre-trained model aligned to the text during training.", + "bbox": [ + 511, + 316, + 906, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Base model finetuning. Since the data distribution of local patches deviates from the full view of the range of common videos, we propose to fine-tune the base model instead of freezing it. It can be seen in Tab. 2 that adding LoRA results in improved performance across all metrics, which demonstrates the effectiveness of utilizing LoRA to adapt to the distribution of local video patches.", + "bbox": [ + 511, + 446, + 908, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 511, + 560, + 735, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We would like to further study some characteristics of our proposed method, including detail generation magnitude, robustness to patch semantics, and computation efficiency.", + "bbox": [ + 511, + 582, + 906, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Detail generation magnitude. The input and output sizes of our VSR model are the same and our model can synthesize different levels of details. In our method, a scale factor is used to guide the model to generate reasonable details that should be presented on the input video if it is captured normally by the camera at that resolution/size. So, the scale factor actually plays the role of affecting the magnitude of details. Here we illustrate the effects of using different scale factors as conditions. In Fig. 6, increasing the scale factor progressively adds more high-resolution details to the results, which demonstrates the flexibility of our method to generate results with different detail magnitudes.", + "bbox": [ + 511, + 635, + 908, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to Patch Semantics. As stated above, our approach will partition the input video into patches in a content-agnostic manner. So, some patches can contain incomplete or even unclear semantics, which are more challenging for the model with the generative reasonable details.", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17805", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7da4b1b25bf002ffdc5a21b500eb30f3e4a0d2f77858e7a79e2fa0a14dc5122a.jpg", + "image_caption": [ + "Input patch" + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 285, + 174 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac762976b9219bbc2a58ad7f99c69cec6fc6148e28b3723e9a7397d266886087.jpg", + "image_caption": [ + "Upscale factor $= 1$" + ], + "image_footnote": [], + "bbox": [ + 289, + 90, + 480, + 174 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/975dd54e4f3806b1b1c33c798ed1ed7d4bd738fc82167a6398e3ce80381faa0c.jpg", + "image_caption": [ + "Upscale factor=2" + ], + "image_footnote": [], + "bbox": [ + 94, + 189, + 285, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7f87ea1be7a60375e5304f789d326828c09495ebd1fe020186351bdfe992437.jpg", + "image_caption": [ + "Upscale factor $= 3$" + ], + "image_footnote": [], + "bbox": [ + 289, + 189, + 480, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07fe9ccade066828e437fea0e2e95984c92167cede84fbf01c3bf4a5dac82076.jpg", + "image_caption": [ + "Upscale factor $= 4$", + "Figure 6. Comparison of different up-sampling scales. Note that the upscale factor is uniformly randomly sampled within the range of 1 to 4 during the training phase." + ], + "image_footnote": [], + "bbox": [ + 94, + 291, + 285, + 377 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4e91ad53575c5b33980bea0bff110d272880726f10c102731aa4c9f9b9987c8.jpg", + "image_caption": [ + "Upscale factor $= 5$" + ], + "image_footnote": [], + "bbox": [ + 289, + 291, + 480, + 377 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2c907d623cb3787ac94d812605ddb156c76074fc286c0e14c2f32f229a7e8707.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTime (s) ↓Memory (G) ↓
LaVie-SR226168
Upscale-A-Video274347
VEnhancer156262
Ours68040
", + "bbox": [ + 117, + 444, + 455, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Computation efficiency for 2K video generation.", + "bbox": [ + 114, + 541, + 455, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is interesting to study whether the performance varies across patches with diverse semantic quality. It can be seen in Fig. 7 that our model can successfully handle the video patches with incomplete semantics. As the global semantic tokens are extracted and input to the base model using global cross-attention modules, the model learns to utilize global semantic information to complete the semantics in the local video patch, which improves the robustness of our model against incomplete semantics.", + "bbox": [ + 88, + 575, + 482, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computation efficiency. Our method is more efficient than existing pre-trained T2V-based VSR methods because the self-attention is conducted within patches. Here presents a simple deduce of the computation complexity against patch size: given a video of $n$ tokens and each patch of $m$ tokens, the patch number is $k = \\frac{n}{m}$ and the inter-token computation times are $k * \\left( \\frac{n}{k} \\right)^2 = n * m$ . Since the base model we adopted uses fixed $512 \\times 512$ resolution, our patch size follows this size in our experiments. The inference time and memory usage of generating a 2K resolution video is illustrated in Tab. 3. It can be seen that, after dividing the input video into patches, the computation burden is signif", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d82803f04939eff4cb0b2aa507b73fe7c770ccdf0efaf0960017e6600ecdcbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 90, + 903, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c64a63248aeb32528b40723527be436a4915b6de7f92ba4edcf20ba1dc4fb7dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 162, + 903, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b8930ed5dc5789277d744bc994c28df541a6e9d6a948e6627a12876fbd6ddff.jpg", + "image_caption": [ + "Incomplete semantics", + "Complete semantics", + "Figure 7. Robustness against patch semantic incompleteness." + ], + "image_footnote": [], + "bbox": [ + 516, + 234, + 903, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "icantly reduced, further promoting the applications of generative VSR models. Thanks to the computation efficiency and the inference flexibility, our method even support 4K video generation, which are shown in the supplementary materials.", + "bbox": [ + 511, + 359, + 903, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "**Limitation.** Our method has limitations in two aspects. Firstly, it experiences noticeable performance degradation when applied to real-world low-resolution images, due to the lack of specialized data augmentation in our training pairs. Since the primary focus of this paper is to address the resolution limitations of using pre-trained base models for VSR, extending our approach to handle real-world noise remains an interesting direction for future work. Secondly, our approach inherits an iterative inference scheme from the base model, which introduces efficiency issues for practical applications. For VSR, the low-resolution input video provides significant opportunities for algorithm optimization to reduce the number of inference steps, which will be another area of future research.", + "bbox": [ + 511, + 454, + 906, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present the first exploration of utilizing a pre-trained T2V base model for patch-level video superresolution (VSR). To accomplish this, we propose an effective dual-branch adapter consisting of a patch condition branch and a global context branch. The global context branch is crucial for bridging the gap between full-video generative capabilities and patch-level detail synthesis. Additionally, our proposed multi-patch joint modulation scheme achieves consistent results across patches. Thanks to the flexibility of our PatchVSR approach, it outperforms existing VSR methods in generating faithful details while maintaining high computational efficiency. We hope this work will inspire further research in this area.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17806", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by National Natural Science Foundation of China (NSFC) under Grant No.62406167, U24B6012.", + "bbox": [ + 89, + 114, + 483, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning (ICML), 2023. 4", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2", + "[3] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2", + "[4] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 2", + "[5] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022.", + "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5962-5971, 2022. 2, 5, 6", + "[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 4", + "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 4", + "[9] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2", + "[10] Shian Du, Xiaotian Cheng, Qi Qian, Henglu Wei, Yi Xu, and Xiangyang Ji. Efficient personalized text-to-image generation by leveraging textual subspace. arXiv preprint arXiv:2407.00608, 2024. 2" + ], + "bbox": [ + 93, + 202, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 3", + "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 5", + "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[14] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17441-17451, 2022. 2", + "[15] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, Ying Shan, and Bihan Wen. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation. In European Conference on Computer Vision (ECCV), 2024. 2, 4", + "[16] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation. arXiv preprint arXiv:2407.07667, 2024. 1, 2, 4, 5, 6", + "[17] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In International Conference on Learning Representations (ICLR), 2024. 4", + "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[19] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 4", + "[20] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5", + "[21] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17807", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "structure-detail network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 645-660. Springer, 2020. 2", + "[22] Mehran Jeelani, Noshaba Cheema, Klaus Illgner-Fehns, Philipp Slusallek, Sunil Jaiswal, et al. Expanding synthetic real-world degradations for blind video super resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1199-1208, 2023. 2", + "[23] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5", + "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations (ICLR), 2014. 4", + "[25] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5687-5696, 2022. 2", + "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5", + "[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 5", + "[28] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2", + "[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[30] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2", + "[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2", + "[32] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2", + "[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 5", + "[34] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4", + "[36] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2", + "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 4", + "[38] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4", + "[40] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. Advances in Neural Information Processing Systems, 35:36081-36093, 2022. 2", + "[41] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2", + "[42] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C. K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 2024. 4", + "[43] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2", + "[44] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 5, 6", + "[45] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20144-20154, 2023. 5", + "[46] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2", + "[47] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying" + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17808", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2", + "[48] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4781-4790, 2021. 2", + "[49] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4", + "[50] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 5", + "[51] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1-15, 2024. 2", + "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2", + "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[54] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2", + "[55] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 1, 2, 4, 5, 6" + ], + "bbox": [ + 91, + 90, + 482, + 683 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17809", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/7794bc7e-86d2-49a8-8adc-3430e39dd287_model.json b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/7794bc7e-86d2-49a8-8adc-3430e39dd287_model.json new file mode 100644 index 0000000000000000000000000000000000000000..792881853c855d0cc9cd8625c4432e839bce0963 --- /dev/null +++ b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/7794bc7e-86d2-49a8-8adc-3430e39dd287_model.json @@ -0,0 +1,2499 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.13, + 0.804, + 0.176 + ], + "angle": 0, + "content": "PatchVSR: Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.209, + 0.838, + 0.248 + ], + "angle": 0, + "content": "Shian Du\\(^{1*‡}\\), Menghan Xia\\(^{2*‡}\\), Chang Liu\\(^{1}\\), Xintao Wang\\(^{2}\\), Jing Wang\\(^{3}\\), Pengfei Wan\\(^{2}\\), Di Zhang\\(^{2}\\), Xiangyang Ji\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.249, + 0.807, + 0.268 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University, \\(^{2}\\)Kuaishou Technology, \\(^{3}\\)Beijing Institute of Technology" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.316, + 0.327, + 0.332 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.486, + 0.729 + ], + "angle": 0, + "content": "Pre-trained video generation models hold great potential for generative video super-resolution (VSR). However, adapting them for full-size VSR, as most existing methods do, suffers from unnecessary intensive full-attention computation and fixed output resolution. To overcome these limitations, we make the first exploration into utilizing video diffusion priors for patch-wise VSR. This is non-trivial because pre-trained video diffusion models are not native for patch-level detail generation. To mitigate this challenge, we propose an innovative approach, called PatchVSR, which integrates a dual-stream adapter for conditional guidance. The patch branch extracts features from input patches to maintain content fidelity while the global branch extracts context features from the resized full video to bridge the generation gap caused by incomplete semantics of patches. Particularly, we also inject the patch's location information into the model to better contextualize patch synthesis within the global video frame. Experiments demonstrate that our method can synthesize high-fidelity, high-resolution details at the patch level. A tailor-made multi-patch joint modulation is proposed to ensure visual consistency across individually enhanced patches. Due to the flexibility of our patch-based paradigm, we can achieve highly competitive 4K VSR based on a \\(512 \\times 512\\) resolution base model, with extremely high efficiency." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.759, + 0.222, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.484, + 0.862 + ], + "angle": 0, + "content": "Video super-resolution (VSR) has long been a challenging research problem in the field of computer vision. With the advent of deep learning, significant advancements have been made in end-to-end VSR models trained on curated datasets. However, due to limited model capacity and data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.908, + 0.485 + ], + "angle": 0, + "content": "coverage, these models typically achieve sharp object structures or coarse textures but struggle to synthesize faithful details. Recently, the emergence of diffusion generative models has enabled open-domain video generation of remarkable quality. The generative capabilities of these foundation models present new opportunities for video super-resolution techniques [16, 55]. Nevertheless, existing base models are generally constrained to generating videos of fixed sizes due to their inherent architecture. This limitation is inherited in their downstream super-resolution models, making it challenging for them to support arbitrary resolution outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.716 + ], + "angle": 0, + "content": "This study is motivated by the insight that feature attention in super-resolution models is more locally focused compared to generation models. This localization is reasonable, as super-resolution tasks have a low-resolution video as a reference, allowing details to be generated based on neighborhood semantics rather than requiring intensive global coherence considerations. To leverage this insight, we propose a patch-based strategy for video superresolution. The concept is straightforward and intuitive: we split the input video into patches sized to be compatible with the pre-trained base model, enhance each patch individually, and then combine them to produce the final result. This approach enables the achievement of super-resolution results at arbitrary resolutions, independent of the base model's inherent resolution constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The key challenge lies in that video generation-based models are trained on complete video frames and face noticeable performance degradation for patch-level generation. To adapt a pre-trained video generation model for generating patch-level details in super-resolution, we propose an innovative approach, called PatchVSR, which integrates a dual-stream adapter for conditional guidance. The local branch extracts features from the input patches to guide the base model for detail synthesis adhering to the input, while the global branch captures contextual information from the entire input frame to bridge the generation gap caused by incomplete semantics of patches. To make the context guid-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.483, + 0.888 + ], + "angle": 0, + "content": "* Work done during an internship at KwaiVGI, Kuaishou Technology." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.36, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Equal contribution. \\(\\dagger\\) Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17799" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "ance more customized, the global branch also incorporates a binary mask indicating the target patch position as input. Additionally, we employ LoRA technology to fine-tune the base model, further improving its adaptability for generating patch-level details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.41 + ], + "angle": 0, + "content": "Experimental results demonstrate that our method effectively adapts the generative ability of the base model to produce faithful details at the patch level. Specifically, to alleviate the boundary inconsistency across patches, we propose a training-free multi-patch joint modulation scheme, where auxiliary bridging patches are created to allow feature interaction across patches during generation. Thanks to the flexibility of PatchVSR, our method, based on a pre-trained \\(512 \\times 512\\) resolution video base model, supports super-resolution video generation up to 4K resolution, while significantly outperforming current state-of-the-art methods in terms of fidelity and detail realism. Notably, compared to existing methods that operate on full-frame videos, our approach processes relatively smaller patches, without full-range attention computation, demonstrating a significant efficiency advantage." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.412, + 0.414, + 0.426 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.428, + 0.483, + 0.488 + ], + "angle": 0, + "content": "- We introduce an innovative approach to overcome the resolution limitations of using pre-trained video diffusion models for VSR. Our method achieves any-resolution VSR output based on a fixed-resolution base model." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.488, + 0.483, + 0.533 + ], + "angle": 0, + "content": "- We propose a novel dual-branch adapter paradigm that effectively harnesses the capabilities of pre-trained video models for patch-level detail generation." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.533, + 0.483, + 0.579 + ], + "angle": 0, + "content": "- Through extensive experimentation, we demonstrate the significant superiority of our method in terms of faithful detail generation and computational efficiency." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.428, + 0.483, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.595, + 0.233, + 0.611 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.32, + 0.636 + ], + "angle": 0, + "content": "2.1. Text-to-video Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "With the widespread use of pre-trained text-to-image models [10, 13, 18, 32, 39, 52] and the preparation of large-scale text-video pairing datasets [9, 34], substantial efforts have been made to train a large-scale text-to-video model. Instead of training from scratch, some methods [2, 41, 46, 51] focus on inserting temporal modules and fine-tuning them or all parameters on video data. To avoid catastrophic forgetting and degradation on visual quality, they often adopt image-video joint training strategy [7, 8, 47]. Recently, with Diffusion Transformer (DiT) [36] being proposed, it has replaced the traditional 3D UNet architecture used by previous approaches with better generation quality and scalability, and has become the dominant framework for text-to-video generation [28, 30, 49]. Although achieving high visual quality, extending them to higher-resolution is non-trivial since it requires substantial training resource and high-quality high-resolution video dataset, which is un" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "available by this moment. To bypass this difficulty, current methods often adopt a cascade framework [15, 44, 54]: the low-resolution video is first generated by the base model using text prompt, and the refinement model outputs high-resolution video through a noisig-denoising process [31] by taking the low-resolution video as condition. However, existing methods are limited to generate full video at once, which still demands a high training overhead. At the same time, the low inference efficiency and high memory usage also hinder its downstream applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.258, + 0.733, + 0.274 + ], + "angle": 0, + "content": "2.2. Video Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Video super-resolution aims at restoring high-resolution (HR) video frames from its degraded low-resolution (LR) counterparts and enhancing its visual quality. Traditional methods often assume a fixed degradation process and their performance deteriorates significantly when handling AI-generated and real-world videos. Due to the lack of real-world LR-HR video data pair for training, RealVSR [48] collects LR-HR video data pair by utilizing the multicamera system of iPhone. Nevertheless, the sequences are captured by a single device, which only contains camera-specific degradations. To avoid burdensome data collection, other methods [6, 22, 55] use stochastic combinations of multi-order degradation processes to simulate real-world degradation processes. While performing more robust when handling real-world cases, due to the absence of deep generative prior, it is challenging for traditional CNN-based [4-6, 21, 43] and transformer-based [3, 14, 25, 40] methods to generate photo-realistic details and textures. By utilizing a pre-trained image diffusion model like Stable Diffusion (SD) \\(\\times 4\\) Upscaler, Lavie-SR [44] and Upscale-A-Video [55] circumvent the need for exhaustive training from scratch and exhibit improved performance in fixed spatial scale within a short sequence. To improve temporal consistency for long sequence, I2VGen-XL [54] and VEnhancer [16] utilize a pre-trained low-resolution video generative model as the base model, and directly fine-tune the model parameters using high-resolution video dataset following noisy-denoising manner [31]. However, the aforementioned methods are all limited to generate full video at once, which requires expanding the training resolution of pre-trained model, bringing low training efficiency and slow convergence speed. Meanwhile, high resolution pretrained video generative models are currently unavailable to access, hindering the scalability of video super-resolution methods with deep generative prior. To tackle this problem, we make the first exploration of utilizing pre-trained video generative model for patch-wise video super-resolution. By maintaining the training resolution of the pre-trained model, our approach not only avoids significant training overhead, making it possible to generate 4K resolution videos using a \\(512 \\times 512\\) resolution pre-trained model. At the same time," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17800" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.91, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.291, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Figure 1. Flowchart of our PatchVSR. Building upon a pre-trained latent T2V model, we incorporate a patch condition branch and a global context branch. These branches extract features from partitioned video patches and the resized full video together with a binary mask that indicates the location of the ROI patch, respectively. Particularly, local patch features are added to the output of each block, while the global context feature is fused with the backbone feature through newly introduced cross-attention modules (G-CA). For simplicity, we have omitted other conditional inputs such as text prompts and time steps from this diagram. The processed patches are fused via a joint modulation scheme to produce a coherent super-resolution video." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.487, + 0.437 + ], + "angle": 0, + "content": "our method significantly improves inference efficiency and reduces memory usage, facilitating downstream applications of video super-resolution methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.451, + 0.185, + 0.467 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.749 + ], + "angle": 0, + "content": "We aim to develop a video super-resolution (VSR) framework designed to restore degraded videos of arbitrary resolutions by generating visually coherent and structurally faithful details. Our primary innovation involves strategically adapting pre-trained text-to-video (T2V) diffusion models—powerful generative tools typically constrained to fixed output size (denoted as \\( s = h \\times w \\))—through a curated conditioning paradigm. To bypass their inherent token-length limitations, we introduce a patch-based processing strategy: the input video is decomposed into \\( s \\)-sized patches, each independently enhanced using the diffusion model, and then reintegrated into a unified high-resolution output through optimized spatial-temporal fusion. This approach is grounded in the distinct characteristics of VSR compared to video generation. While video synthesis demands global coherence across full-frame contexts, superresolution prioritizes localized detail restoration atop existing structural content." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.904 + ], + "angle": 0, + "content": "Specifically, we augment the base model with two conditional branches: a patch branch that extracts features from input patches, and a global branch that extracts context from the full video. This global context mainly aids in generating patch-level details from the T2V model that was originally trained on full-frame videos. An overview of the model architecture is illustrated in Fig. 1. Given a low-resolution image \\(\\mathbf{V}_l\\in \\mathbb{R}^{F\\times H\\times W\\times 3}\\) and the target upscale factor \\(k\\) , our process begins by upscaling \\(\\mathbf{V}_l\\) to the desired resolution \\(\\mathbf{V}_l^\\uparrow \\in \\mathbb{R}^{F\\times [H\\cdot k]\\times [W\\cdot k]\\times 3}\\) using bicubic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.542 + ], + "angle": 0, + "content": "interpolation per frame. We then partition \\(\\mathbf{V}_l^\\dagger\\) into patches \\(\\{\\mathbf{P}_i\\in \\mathbb{R}^{F\\times h\\times w\\times 3}\\}_{i = 1}^N\\), where the patch dimensions \\((h,w)\\) match the generation size of the pre-trained T2V model. Additionally, we resize the full video to \\(\\mathbf{G}\\in \\mathbb{R}^{F\\times h\\times w\\times 3}\\) to serve as global guidance. Finally, both \\(\\mathbf{P}_i\\) and \\(\\mathbf{G}\\) are input into the VSR model for patch-level detail enhancement, and multi-patch joint modulation is involved to ensure coherence across patches. This formulation ensures compatibility with the pre-trained T2V model while allowing for flexible processing of videos of arbitrary sizes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.871, + 0.566 + ], + "angle": 0, + "content": "3.1. Preliminary: Video Diffusion Base Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.646 + ], + "angle": 0, + "content": "Our model is implemented on top of a pre-trained T2V diffusion model, which adopts the Rectified Flow framework [11] for the noise schedule and denoising process. The forward process is defined as straight paths between data distribution and a standard normal distribution, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.657, + 0.907, + 0.675 + ], + "angle": 0, + "content": "\\[\nz _ {t} = (1 - t) z _ {0} + t \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.909, + 0.76 + ], + "angle": 0, + "content": "where \\(\\epsilon \\in \\mathcal{N}(0,I)\\) and \\(t\\) denotes the iterative timestep. To solve the denoising processing, we define a mapping between samples \\(z_{1}\\) from a noise distribution \\(p_1\\) to samples \\(z_{0}\\) from a data distribution \\(p_0\\) in terms of an ordinary differential equation (ODE), namely:" + }, + { + "type": "equation", + "bbox": [ + 0.646, + 0.769, + 0.907, + 0.788 + ], + "angle": 0, + "content": "\\[\nd z _ {t} = v _ {\\Theta} (z _ {t}, t) d t, \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.873 + ], + "angle": 0, + "content": "where the velocity \\( v \\) is parameterized by the weights \\( \\Theta \\) of a neural network. For inference, we employ Euler discretization for Eq. 2 and perform discretization over the timestep interval at [0,1], starting at \\( t = 1 \\). We then processed with iterative sampling with: \\( z_{t_i} = z_{t_j} + v_\\Theta (z_{t_j},t_j)*(t_j - t_i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.872, + 0.908, + 0.902 + ], + "angle": 0, + "content": "For consideration of computing efficiency, the mapping is formulated in a latent space constructed with a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "17801" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.318 + ], + "angle": 0, + "content": "3D Variational Auto-Encoder (VAE) [24]. The core is a Transformer-based diffusion model (DiT) [35], where each block is instantiated as a sequence of spatial attention, temporal attention, and cross-attention modules. Text prompts and other micro parameters (like time step, aspect ratio, etc) are used as generative conditions. Due to the full-attention property of Transformer, our model only supports the generation of a fixed number of tokens as it was trained with, i.e. \\( 512 \\times 512 \\) of diverse aspect-ratio. This is a common characteristic of almost all existing diffusion models for image or video generation [7, 8, 37, 49], and extending pre-trained diffusion models for higher resolution generation remains an open research topic [15, 17]. For VSR as a downstream task, we propose a patch-wise processing strategy to break this resolution limit." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.328, + 0.468, + 0.343 + ], + "angle": 0, + "content": "3.2. Patch Video Diffusion with Global Guidance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.456 + ], + "angle": 0, + "content": "To adapt a pre-trained T2V model for patch-level VSR, we must consider two key aspects: (i) enabling the input video to guide the base model's denoising process and (ii) bridging the gap between the base model's capabilities in generating full-frame content and patch-level content synthesis requirements. We achieve this through a dual-branch adapter, as detailed below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Patch Condition Branch. There are generally two approaches to incorporating visual conditions into pre-trained generative models: concatenating with noisy data in the noise space [39, 55], or fusing with the backbone feature through an adapter branch [16, 42]. The former involves modifying the first layers' channel count and necessitates base model fine-tuning, which is typically more data and computationally intensive. In contrast, the latter is more lightweight, as only a newly added branch requires training for feature adaptation. We adopt a hybrid solution that combines elements of both approaches. First, we employ an adapter comprising several Transformer blocks to extract features from the input video patch \\(\\mathbf{P}_i\\) and inject them into the base model's blocks. Additionally, considering that the base model was trained for full video generation, we need to adapt its embedded knowledge for patch-level generation. To this end, we allow the base model to be fine-tuned using a Low-Rank Adaptation (LoRA) [19]. An ablation study in Sec. 4.3 demonstrates the advantages of this design." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Global Context Branch. To further guide content generation in a natural manner, we incorporate a global branch that extracts context from the full input video. This branch utilizes a Transformer-based encoder architecture that reduces the number of tokens to one quarter. The resulting context features are fused with each block of the base model through cross-attention \\(\\{\\mathbf{Q},\\mathbf{K}_g,\\mathbf{V}_g\\}\\), where the \\(\\mathbf{Q}\\) is shared with the existing text-prompt cross-attention" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.108, + 0.653, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.197, + 0.627, + 0.209 + ], + "angle": 0, + "content": "Patch Partition" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.123, + 0.691, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.09, + 0.905, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.197, + 0.869, + 0.21 + ], + "angle": 0, + "content": "Created auxiliary patches" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.215, + 0.907, + 0.271 + ], + "angle": 0, + "content": "Figure 2. Patch Partition Visualization. The input video is divided into non-overlapping segments, as the solid blue boxes mark. For joint modulation, auxiliary patches are created, indicated by the red dashed boxes, resulting in an overlapping ratio of \\(50\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.279, + 0.907, + 0.385 + ], + "angle": 0, + "content": "\\(\\{\\mathbf{Q}, \\mathbf{K}_t, \\mathbf{V}_t\\}\\). Specifically, the input is constructed by concatenating the resized full video \\(\\mathbf{V}_g\\) (to accommodate a fixed number of tokens) with a binary map \\(\\mathbf{M}_i\\) indicating the target patch's \\(\\mathbf{P}_i\\) location within it. Our experiments show that the global context branch not only enhances the vividness of generated patch-level details but also improves consistency across patches." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.396, + 0.781, + 0.41 + ], + "angle": 0, + "content": "3.3. Multi-Patch Joint Modulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.907, + 0.569 + ], + "angle": 0, + "content": "While the global guidance shared by patches is critical for maintaining color consistency, tonal uniformity, and semantic coherence of details, naive stitching of independently enhanced patches often introduces visible seams. These boundary artifacts primarily stem from the stochastic nature of the generative process when hallucinating textures on degraded or underdetermined regions. This is a fundamental challenge in ill-posed inverse problems where multiple high-resolution solutions may satisfy the constraints of a single low-resolution input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To mitigate this problem, we introduce inter-patch modulation during the detail generation process. Specifically, a training-free inference strategy called multi-diffusion [1] is adapted to solve our problem. This approach involves splitting the original video into overlapping patches and updating the overlapping regions by averaging contributions from all relevant patches at each denoising step. As illustrated in Fig. 2, instead of directly splitting the video into overlapping patches, we first partition it into non-overlapping segments, and then construct auxiliary patches by combining the halves of adjacent patches. This creates a new set of patch groups with an overlapping ratio of \\(50\\%\\). Our experiment shows that directly taking the average in overlap regions suffers from artifacts of black holes or seamlines, as evidenced in Fig. 5 (middle). Instead, when fusing base patches and its related auxiliary patches, we manage to gradually reduce its influence of auxiliary patches from its center splitting line to two sides by assigning spatial weight maps to them. This method effectively reduces visible seams between patches and promotes overall consistency. A comparative result demonstrating the effectiveness of this approach is presented in Fig. 5." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17802" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.312, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.185, + 0.243, + 0.197 + ], + "angle": 0, + "content": "Video & prompt" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.125, + 0.493, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.185, + 0.435, + 0.195 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.199, + 0.227, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.259, + 0.176, + 0.269 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.199, + 0.359, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.258, + 0.321, + 0.268 + ], + "angle": 0, + "content": "LaVie-SR" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.199, + 0.493, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.259, + 0.457, + 0.269 + ], + "angle": 0, + "content": "VEnhancer" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.274, + 0.227, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.334, + 0.202, + 0.344 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.274, + 0.359, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.334, + 0.342, + 0.345 + ], + "angle": 0, + "content": "Upscale-A-Video" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.274, + 0.493, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.334, + 0.441, + 0.344 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.091, + 0.718, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.185, + 0.655, + 0.197 + ], + "angle": 0, + "content": "Video & prompt" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.125, + 0.904, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.185, + 0.846, + 0.195 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.199, + 0.637, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.259, + 0.586, + 0.27 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.199, + 0.771, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.259, + 0.732, + 0.269 + ], + "angle": 0, + "content": "LaVie-SR" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.199, + 0.903, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.259, + 0.869, + 0.269 + ], + "angle": 0, + "content": "VEnhancer" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.274, + 0.637, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.334, + 0.611, + 0.344 + ], + "angle": 0, + "content": "RealBasicVSR" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.274, + 0.771, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.334, + 0.752, + 0.345 + ], + "angle": 0, + "content": "Upscale-A-Video" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.274, + 0.903, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.334, + 0.851, + 0.344 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.351, + 0.862, + 0.365 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons on SynVideo30 (test). These videos \\(\\times 4\\) super-resolution results. Zoom in for best view." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.385, + 0.223, + 0.401 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.501 + ], + "angle": 0, + "content": "To verify the superiority of PatchVSR, we compare it with SOTA methods, including VEnhancer [16], Upscale-A-Video [55], LaVie-SR [44] and RealBasicVSR [6]. We also make ablation study to analyze the effectiveness of our proposed designs. Besides, some extended analysis is conducted to demonstrate the capability of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.509, + 0.308, + 0.525 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Training Datasets We train our model using 460K self-collected high-quality video-text pairs, with each resolution from \\(1024 \\times 1024\\) to 2K. The text prompts are all captioned using LAVA captioner [26], and encoded by T5 text encoder [38] with no more than 256 tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.484, + 0.77 + ], + "angle": 0, + "content": "Testing Datasets. We evaluate the VSR methods on two kinds of dataset, i.e. downscale synthetic videos and AIGC videos. For synthetic videos, since our employed base model is fixed to generate 77-frames video, we utilize a public dataset: REDS30 [33] for fair comparisons, which contains 100 frames. To evaluate higher resolution, we self-collect 30 high-quality 2K resolution videos, each containing more than 150 frames. Additionally, we collect and test on an AIGC dataset, which comprises 30 AI-generated videos from public text-to-video generative models [50]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Training Details. Our model is trained on 16 NVIDIA H800-80G GPUs with a total batch size of 64. AdamW [29] is used as the optimizer with a learning rate of \\(10^{-5}\\). The text prompt is randomly replaced by a null prompt with \\(10\\%\\) probability. To obtain the LR-HR training data pair, the HR video is first down-sampled using bilinear interpolation with down scale \\(s\\) and up-sampled back to the original resolution. To match the size of the pre-trained model," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.907, + 0.598 + ], + "angle": 0, + "content": "we randomly crop a \\(512 \\times 512\\) area from the up-sampled LR video with aspect ratios from 0.5 to 2.0. To enhance the robustness of our model to different degradation scenarios, we utilize the noise augmentation technique by injecting noise into the input latent using a diffuse process. The noise timestep is randomly selected from 200 to 300, aiming at preserving the structure of the input video. We have additionally encoded noise timestep, downscale factor, and crop location as conditions for the model. We have replaced the text prompt with a fixed prompt for the patch condition branch, with the base model and global context branch using text prompts. To provide initialization weight, we followed the structure of the pre-trained model and sampled the layer index from it at equal intervals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.657 + ], + "angle": 0, + "content": "Inference Details. We perform 50 PNDM [27] sampling steps with classifier-free guidance scale 5.0. We have also used timestep shift [12] with shift value 5.0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the fidelity of our method on synthesized videos, we utilize PSNR, SSIM, and LPIPS [53]. For visual quality, we conduct our evaluation of commonly used quality assessment metrics MUSIQ [23] and DOVER [45]. Furthermore, we also include aesthetic score from a video generation benchmark VBench [20] for more comprehensive comparisons." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.805, + 0.804 + ], + "angle": 0, + "content": "4.2. Comparison with SOTA methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results. The visual results of synthetic and AIGC videos are presented in Fig. 3 and Fig. 4 respectively. It is observed that our method significantly outperforms existing CNN-based and diffusion-based models in both details generation and temporal consistency (Readers are recommended to check our provided video results). Specif" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17803" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.363, + 0.875, + 0.378 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons on VideoGen30 (test). These videos are \\(\\times 4\\) super-resolution results. Zoom in for best view." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.39, + 0.911, + 0.651 + ], + "angle": 0, + "content": "
DatasetsMetricsRealBasicVSR [6]LaVie-SR [44]Upscale-A-Video [55]VEnhancer [16]Ours
REDS30PSNR ↑29.69828.73531.25728.54430.111
SSIM ↑0.6560.6220.6610.6110.647
LPIPS ↓0.3340.4070.3280.4210.326
DOVER ↑0.3050.3180.3100.3440.323
MUSIQ ↑40.58743.57742.57544.53344.956
Aesthetics ↑0.4220.4860.4710.4820.493
SynVideo30PSNR ↑33.50731.43333.43228.85630.857
SSIM ↑0.7760.7160.7280.6970.732
LPIPS ↓0.1850.1940.2050.1990.183
DOVER ↑0.5840.6120.5290.5500.597
MUSIQ ↑49.55748.14649.83943.53850.695
Aesthetics ↑0.4960.5090.4940.5030.520
VideoGen30DOVER ↑0.5880.6090.5100.5120.590
MUSIQ ↑47.56448.32946.89946.93050.559
Aesthetics ↑0.5470.6000.5740.5510.602
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.655, + 0.907, + 0.684 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on different VSR benchmarks from synthetic (REDS30, SynVideo30) and AIGC (VideoGen30) data. Red and blue numbers indicate the best and second best results respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.697, + 0.486, + 0.894 + ], + "angle": 0, + "content": "ically, our method not only successfully recovers the details of the red silk scarf in Fig. 3, but also adds more details to the \"Angkor Wat\" as shown in Fig. 4, which others tend to generate blurry or unnatural results. In Fig. 5, we also illustrate full VSR results achieved through the multi-patch joint modulation scheme. It can be seen that creating auxiliary patches greatly alleviates the seam problem and significantly reduces the color difference between different patches, while further improving the contrast of the results. To show the advantage of our multi-patch joint modulation, we compare with a baseline method: stitching the patches in latent space and decoding the big latent together, which suffers obvious seamline artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.909, + 0.895 + ], + "angle": 0, + "content": "Quantitative Results. The quantitative comparison is shown in Tab. 1. It can be seen that the proposed method achieves the highest LPIPS score in both REDS30 and SynVideo30, indicating the high fidelity of our method for synthetic videos. However, our method show inferior performance in PSNR and SSIM. This can be explained by the fact that our method tends to generate rich and vivid details, which are not favorable to metrics that capture pixel-wise deviations (i.e. PSRN and SSIM). Instead, as a perceptual-level fidelity measure, LPIPS might be more informative to reflect the content fidelity for generative tasks. We evaluate the visual quality of generated high-resolution videos through comparisons with competing methods on MUSIQ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17804" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.092, + 0.48, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.542 + ], + "angle": 0, + "content": "Figure 5. Visual evaluation on multi-patch joint modulation. Here the super-resolved videos of 2K resolution are used for comparison, where \\(3 \\times 3 = 9\\) patches are involved. The top, middle, bottom denote input video, result of stitching latent of patches, and result of ours respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.692 + ], + "angle": 0, + "content": "and DOVER metrics, demonstrating our approach's superiority. This performance reflects contributions from both the base model's capabilities and the effectiveness of our method in leveraging full-frame generation for patch-level detail synthesis. Interestingly, our approach even preserves subjective properties like aesthetic quality from the base model, as evidenced by the aesthetic scores. This highlights our framework's ability to inherit and maintain critical perceptual characteristics during the enhancement process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.7, + 0.245, + 0.717 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.483, + 0.784 + ], + "angle": 0, + "content": "We conduct ablation studies on key designs. Considering the limited space, we provide a quantitative evaluation here and the typical visual results are compared in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Global Branch. To validate the effectiveness of the global branch, We build a baseline by removing the global branch from our method. As shown in Tab. 2, removing the global branch results in significant performance deterioration, leading to worse DOVER, MUSIQ, and aesthetic score. Besides, to check the necessity of using the patch location embeddings, we build a baseline by removing the" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.196 + ], + "angle": 0, + "content": "
ComponentDOVER ↑MUSIQ ↑Aesthetics ↑
w/o global branch0.50246.0740.589
w/o LoRA0.58250.4960.600
w/o location embed0.57450.0840.601
w/o fixed prompt0.56248.1330.597
Ours0.59050.5590.602
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.2, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Table 2. Ablation study of different model components on VideoGen30 test dataset. The **bold** numbers indicate the best." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.235, + 0.907, + 0.311 + ], + "angle": 0, + "content": "patch position information from the global branch input. As shown in Tab. 2, after incorporating the location embedding, we obtain an increase of DOVER from 0.574 to 0.590, which demonstrates the effectiveness of utilizing the location mask as the patch position information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Text Prompts. Text prompts are used in the base model and the conditional global branch. As shown in Tab. 2, we observe that replacing the text prompt with a fixed prompt leads to superior results. We conjecture that the use of text prompts in patch condition branch brings about a mismatch between the video patch and the global semantic information, which in turn destroys the dynamic visual prior in the pre-trained model aligned to the text during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.909, + 0.553 + ], + "angle": 0, + "content": "Base model finetuning. Since the data distribution of local patches deviates from the full view of the range of common videos, we propose to fine-tune the base model instead of freezing it. It can be seen in Tab. 2 that adding LoRA results in improved performance across all metrics, which demonstrates the effectiveness of utilizing LoRA to adapt to the distribution of local video patches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.736, + 0.577 + ], + "angle": 0, + "content": "4.4. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.63 + ], + "angle": 0, + "content": "We would like to further study some characteristics of our proposed method, including detail generation magnitude, robustness to patch semantics, and computation efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.909, + 0.818 + ], + "angle": 0, + "content": "Detail generation magnitude. The input and output sizes of our VSR model are the same and our model can synthesize different levels of details. In our method, a scale factor is used to guide the model to generate reasonable details that should be presented on the input video if it is captured normally by the camera at that resolution/size. So, the scale factor actually plays the role of affecting the magnitude of details. Here we illustrate the effects of using different scale factors as conditions. In Fig. 6, increasing the scale factor progressively adds more high-resolution details to the results, which demonstrates the flexibility of our method to generate results with different detail magnitudes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Robustness to Patch Semantics. As stated above, our approach will partition the input video into patches in a content-agnostic manner. So, some patches can contain incomplete or even unclear semantics, which are more challenging for the model with the generative reasonable details." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17805" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.287, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.176, + 0.221, + 0.188 + ], + "angle": 0, + "content": "Input patch" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.091, + 0.481, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.176, + 0.432, + 0.188 + ], + "angle": 0, + "content": "Upscale factor \\(= 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.19, + 0.286, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.277, + 0.236, + 0.289 + ], + "angle": 0, + "content": "Upscale factor=2" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.19, + 0.481, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.278, + 0.432, + 0.289 + ], + "angle": 0, + "content": "Upscale factor \\(= 3\\)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.292, + 0.286, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.379, + 0.238, + 0.39 + ], + "angle": 0, + "content": "Upscale factor \\(= 4\\)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.292, + 0.481, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.379, + 0.433, + 0.39 + ], + "angle": 0, + "content": "Upscale factor \\(= 5\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.438 + ], + "angle": 0, + "content": "Figure 6. Comparison of different up-sampling scales. Note that the upscale factor is uniformly randomly sampled within the range of 1 to 4 during the training phase." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.445, + 0.457, + 0.538 + ], + "angle": 0, + "content": "
MethodsTime (s) ↓Memory (G) ↓
LaVie-SR226168
Upscale-A-Video274347
VEnhancer156262
Ours68040
" + }, + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.542, + 0.457, + 0.556 + ], + "angle": 0, + "content": "Table 3. Computation efficiency for 2K video generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.576, + 0.483, + 0.711 + ], + "angle": 0, + "content": "It is interesting to study whether the performance varies across patches with diverse semantic quality. It can be seen in Fig. 7 that our model can successfully handle the video patches with incomplete semantics. As the global semantic tokens are extracted and input to the base model using global cross-attention modules, the model learns to utilize global semantic information to complete the semantics in the local video patch, which improves the robustness of our model against incomplete semantics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Computation efficiency. Our method is more efficient than existing pre-trained T2V-based VSR methods because the self-attention is conducted within patches. Here presents a simple deduce of the computation complexity against patch size: given a video of \\( n \\) tokens and each patch of \\( m \\) tokens, the patch number is \\( k = \\frac{n}{m} \\) and the inter-token computation times are \\( k * \\left( \\frac{n}{k} \\right)^2 = n * m \\). Since the base model we adopted uses fixed \\( 512 \\times 512 \\) resolution, our patch size follows this size in our experiments. The inference time and memory usage of generating a 2K resolution video is illustrated in Tab. 3. It can be seen that, after dividing the input video into patches, the computation burden is signif" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.905, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.31, + 0.696, + 0.324 + ], + "angle": 0, + "content": "Incomplete semantics" + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.31, + 0.875, + 0.324 + ], + "angle": 0, + "content": "Complete semantics" + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.33, + 0.89, + 0.344 + ], + "angle": 0, + "content": "Figure 7. Robustness against patch semantic incompleteness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.905, + 0.435 + ], + "angle": 0, + "content": "icantly reduced, further promoting the applications of generative VSR models. Thanks to the computation efficiency and the inference flexibility, our method even support 4K video generation, which are shown in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.907, + 0.665 + ], + "angle": 0, + "content": "**Limitation.** Our method has limitations in two aspects. Firstly, it experiences noticeable performance degradation when applied to real-world low-resolution images, due to the lack of specialized data augmentation in our training pairs. Since the primary focus of this paper is to address the resolution limitations of using pre-trained base models for VSR, extending our approach to handle real-world noise remains an interesting direction for future work. Secondly, our approach inherits an iterative inference scheme from the base model, which introduces efficiency issues for practical applications. For VSR, the low-resolution input video provides significant opportunities for algorithm optimization to reduce the number of inference steps, which will be another area of future research." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present the first exploration of utilizing a pre-trained T2V base model for patch-level video superresolution (VSR). To accomplish this, we propose an effective dual-branch adapter consisting of a patch condition branch and a global context branch. The global context branch is crucial for bridging the gap between full-video generative capabilities and patch-level detail synthesis. Additionally, our proposed multi-patch joint modulation scheme achieves consistent results across patches. Thanks to the flexibility of our PatchVSR approach, it outperforms existing VSR methods in generating faithful details while maintaining high computational efficiency. We hope this work will inspire further research in this area." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17806" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.161 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China (NSFC) under Grant No.62406167, U24B6012." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.188, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.203, + 0.484, + 0.258 + ], + "angle": 0, + "content": "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning (ICML), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.261, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.332, + 0.483, + 0.373 + ], + "angle": 0, + "content": "[3] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.483, + 0.445 + ], + "angle": 0, + "content": "[4] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.483, + 0.516 + ], + "angle": 0, + "content": "[5] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5962-5971, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[9] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[10] Shian Du, Xiaotian Cheng, Qi Qian, Henglu Wei, Yi Xu, and Xiangyang Ji. Efficient personalized text-to-image generation by leveraging textual subspace. arXiv preprint arXiv:2407.00608, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.203, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[11] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[14] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17441-17451, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[15] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, Ying Shan, and Bihan Wen. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation. In European Conference on Computer Vision (ECCV), 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[16] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation. arXiv preprint arXiv:2407.07667, 2024. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.656 + ], + "angle": 0, + "content": "[17] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In International Conference on Learning Representations (ICLR), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[19] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[20] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17807" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "structure-detail network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 645-660. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[22] Mehran Jeelani, Noshaba Cheema, Klaus Illgner-Fehns, Philipp Slusallek, Sunil Jaiswal, et al. Expanding synthetic real-world degradations for blind video super resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1199-1208, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[23] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations (ICLR), 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[25] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5687-5696, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[28] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[30] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[32] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[34] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[36] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.204, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[38] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.907, + 0.413 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[40] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. Advances in Neural Information Processing Systems, 35:36081-36093, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[41] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.907, + 0.566 + ], + "angle": 0, + "content": "[42] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C. K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.907, + 0.636 + ], + "angle": 0, + "content": "[43] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[44] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[45] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20144-20154, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[46] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[47] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17808" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[48] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4781-4790, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[49] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[50] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[51] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1-15, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[54] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[55] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 1, 2, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17809" + } + ] +] \ No newline at end of file diff --git a/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/7794bc7e-86d2-49a8-8adc-3430e39dd287_origin.pdf b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/7794bc7e-86d2-49a8-8adc-3430e39dd287_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9e2e90272177b34459f181ede978f17c39162efd --- /dev/null +++ b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/7794bc7e-86d2-49a8-8adc-3430e39dd287_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8dcdd19844d7fd3f3815f488918f93b23594620debba9dce23237c1c7ca5059 +size 3442994 diff --git a/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/full.md b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b8caee2af3edd6d0120d54db3ab9afcb6bd092d9 --- /dev/null +++ b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/full.md @@ -0,0 +1,322 @@ +# PatchVSR: Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution + +Shian Du $^{1*‡}$ , Menghan Xia $^{2*‡}$ , Chang Liu $^{1}$ , Xintao Wang $^{2}$ , Jing Wang $^{3}$ , Pengfei Wan $^{2}$ , Di Zhang $^{2}$ , Xiangyang Ji $^{1*}$ + +$^{1}$ Tsinghua University, $^{2}$ Kuaishou Technology, $^{3}$ Beijing Institute of Technology + +# Abstract + +Pre-trained video generation models hold great potential for generative video super-resolution (VSR). However, adapting them for full-size VSR, as most existing methods do, suffers from unnecessary intensive full-attention computation and fixed output resolution. To overcome these limitations, we make the first exploration into utilizing video diffusion priors for patch-wise VSR. This is non-trivial because pre-trained video diffusion models are not native for patch-level detail generation. To mitigate this challenge, we propose an innovative approach, called PatchVSR, which integrates a dual-stream adapter for conditional guidance. The patch branch extracts features from input patches to maintain content fidelity while the global branch extracts context features from the resized full video to bridge the generation gap caused by incomplete semantics of patches. Particularly, we also inject the patch's location information into the model to better contextualize patch synthesis within the global video frame. Experiments demonstrate that our method can synthesize high-fidelity, high-resolution details at the patch level. A tailor-made multi-patch joint modulation is proposed to ensure visual consistency across individually enhanced patches. Due to the flexibility of our patch-based paradigm, we can achieve highly competitive 4K VSR based on a $512 \times 512$ resolution base model, with extremely high efficiency. + +# 1. Introduction + +Video super-resolution (VSR) has long been a challenging research problem in the field of computer vision. With the advent of deep learning, significant advancements have been made in end-to-end VSR models trained on curated datasets. However, due to limited model capacity and data + +coverage, these models typically achieve sharp object structures or coarse textures but struggle to synthesize faithful details. Recently, the emergence of diffusion generative models has enabled open-domain video generation of remarkable quality. The generative capabilities of these foundation models present new opportunities for video super-resolution techniques [16, 55]. Nevertheless, existing base models are generally constrained to generating videos of fixed sizes due to their inherent architecture. This limitation is inherited in their downstream super-resolution models, making it challenging for them to support arbitrary resolution outputs. + +This study is motivated by the insight that feature attention in super-resolution models is more locally focused compared to generation models. This localization is reasonable, as super-resolution tasks have a low-resolution video as a reference, allowing details to be generated based on neighborhood semantics rather than requiring intensive global coherence considerations. To leverage this insight, we propose a patch-based strategy for video superresolution. The concept is straightforward and intuitive: we split the input video into patches sized to be compatible with the pre-trained base model, enhance each patch individually, and then combine them to produce the final result. This approach enables the achievement of super-resolution results at arbitrary resolutions, independent of the base model's inherent resolution constraints. + +The key challenge lies in that video generation-based models are trained on complete video frames and face noticeable performance degradation for patch-level generation. To adapt a pre-trained video generation model for generating patch-level details in super-resolution, we propose an innovative approach, called PatchVSR, which integrates a dual-stream adapter for conditional guidance. The local branch extracts features from the input patches to guide the base model for detail synthesis adhering to the input, while the global branch captures contextual information from the entire input frame to bridge the generation gap caused by incomplete semantics of patches. To make the context guid- + +ance more customized, the global branch also incorporates a binary mask indicating the target patch position as input. Additionally, we employ LoRA technology to fine-tune the base model, further improving its adaptability for generating patch-level details. + +Experimental results demonstrate that our method effectively adapts the generative ability of the base model to produce faithful details at the patch level. Specifically, to alleviate the boundary inconsistency across patches, we propose a training-free multi-patch joint modulation scheme, where auxiliary bridging patches are created to allow feature interaction across patches during generation. Thanks to the flexibility of PatchVSR, our method, based on a pre-trained $512 \times 512$ resolution video base model, supports super-resolution video generation up to 4K resolution, while significantly outperforming current state-of-the-art methods in terms of fidelity and detail realism. Notably, compared to existing methods that operate on full-frame videos, our approach processes relatively smaller patches, without full-range attention computation, demonstrating a significant efficiency advantage. + +Our contributions are summarized as follows: + +- We introduce an innovative approach to overcome the resolution limitations of using pre-trained video diffusion models for VSR. Our method achieves any-resolution VSR output based on a fixed-resolution base model. +- We propose a novel dual-branch adapter paradigm that effectively harnesses the capabilities of pre-trained video models for patch-level detail generation. +- Through extensive experimentation, we demonstrate the significant superiority of our method in terms of faithful detail generation and computational efficiency. + +# 2. Related Work + +# 2.1. Text-to-video Generation + +With the widespread use of pre-trained text-to-image models [10, 13, 18, 32, 39, 52] and the preparation of large-scale text-video pairing datasets [9, 34], substantial efforts have been made to train a large-scale text-to-video model. Instead of training from scratch, some methods [2, 41, 46, 51] focus on inserting temporal modules and fine-tuning them or all parameters on video data. To avoid catastrophic forgetting and degradation on visual quality, they often adopt image-video joint training strategy [7, 8, 47]. Recently, with Diffusion Transformer (DiT) [36] being proposed, it has replaced the traditional 3D UNet architecture used by previous approaches with better generation quality and scalability, and has become the dominant framework for text-to-video generation [28, 30, 49]. Although achieving high visual quality, extending them to higher-resolution is non-trivial since it requires substantial training resource and high-quality high-resolution video dataset, which is un + +available by this moment. To bypass this difficulty, current methods often adopt a cascade framework [15, 44, 54]: the low-resolution video is first generated by the base model using text prompt, and the refinement model outputs high-resolution video through a noisig-denoising process [31] by taking the low-resolution video as condition. However, existing methods are limited to generate full video at once, which still demands a high training overhead. At the same time, the low inference efficiency and high memory usage also hinder its downstream applications. + +# 2.2. Video Super-Resolution + +Video super-resolution aims at restoring high-resolution (HR) video frames from its degraded low-resolution (LR) counterparts and enhancing its visual quality. Traditional methods often assume a fixed degradation process and their performance deteriorates significantly when handling AI-generated and real-world videos. Due to the lack of real-world LR-HR video data pair for training, RealVSR [48] collects LR-HR video data pair by utilizing the multicamera system of iPhone. Nevertheless, the sequences are captured by a single device, which only contains camera-specific degradations. To avoid burdensome data collection, other methods [6, 22, 55] use stochastic combinations of multi-order degradation processes to simulate real-world degradation processes. While performing more robust when handling real-world cases, due to the absence of deep generative prior, it is challenging for traditional CNN-based [4-6, 21, 43] and transformer-based [3, 14, 25, 40] methods to generate photo-realistic details and textures. By utilizing a pre-trained image diffusion model like Stable Diffusion (SD) $\times 4$ Upscaler, Lavie-SR [44] and Upscale-A-Video [55] circumvent the need for exhaustive training from scratch and exhibit improved performance in fixed spatial scale within a short sequence. To improve temporal consistency for long sequence, I2VGen-XL [54] and VEnhancer [16] utilize a pre-trained low-resolution video generative model as the base model, and directly fine-tune the model parameters using high-resolution video dataset following noisy-denoising manner [31]. However, the aforementioned methods are all limited to generate full video at once, which requires expanding the training resolution of pre-trained model, bringing low training efficiency and slow convergence speed. Meanwhile, high resolution pretrained video generative models are currently unavailable to access, hindering the scalability of video super-resolution methods with deep generative prior. To tackle this problem, we make the first exploration of utilizing pre-trained video generative model for patch-wise video super-resolution. By maintaining the training resolution of the pre-trained model, our approach not only avoids significant training overhead, making it possible to generate 4K resolution videos using a $512 \times 512$ resolution pre-trained model. At the same time, + +![](images/3a130280773855dd3be87b1e9a0c1c055c7a493d43d4caee1aa4059ad7141be2.jpg) +Figure 1. Flowchart of our PatchVSR. Building upon a pre-trained latent T2V model, we incorporate a patch condition branch and a global context branch. These branches extract features from partitioned video patches and the resized full video together with a binary mask that indicates the location of the ROI patch, respectively. Particularly, local patch features are added to the output of each block, while the global context feature is fused with the backbone feature through newly introduced cross-attention modules (G-CA). For simplicity, we have omitted other conditional inputs such as text prompts and time steps from this diagram. The processed patches are fused via a joint modulation scheme to produce a coherent super-resolution video. + +our method significantly improves inference efficiency and reduces memory usage, facilitating downstream applications of video super-resolution methods. + +# 3. Method + +We aim to develop a video super-resolution (VSR) framework designed to restore degraded videos of arbitrary resolutions by generating visually coherent and structurally faithful details. Our primary innovation involves strategically adapting pre-trained text-to-video (T2V) diffusion models—powerful generative tools typically constrained to fixed output size (denoted as $s = h \times w$ )—through a curated conditioning paradigm. To bypass their inherent token-length limitations, we introduce a patch-based processing strategy: the input video is decomposed into $s$ -sized patches, each independently enhanced using the diffusion model, and then reintegrated into a unified high-resolution output through optimized spatial-temporal fusion. This approach is grounded in the distinct characteristics of VSR compared to video generation. While video synthesis demands global coherence across full-frame contexts, superresolution prioritizes localized detail restoration atop existing structural content. + +Specifically, we augment the base model with two conditional branches: a patch branch that extracts features from input patches, and a global branch that extracts context from the full video. This global context mainly aids in generating patch-level details from the T2V model that was originally trained on full-frame videos. An overview of the model architecture is illustrated in Fig. 1. Given a low-resolution image $\mathbf{V}_l\in \mathbb{R}^{F\times H\times W\times 3}$ and the target upscale factor $k$ , our process begins by upscaling $\mathbf{V}_l$ to the desired resolution $\mathbf{V}_l^\uparrow \in \mathbb{R}^{F\times [H\cdot k]\times [W\cdot k]\times 3}$ using bicubic + +interpolation per frame. We then partition $\mathbf{V}_l^\dagger$ into patches $\{\mathbf{P}_i\in \mathbb{R}^{F\times h\times w\times 3}\}_{i = 1}^N$ , where the patch dimensions $(h,w)$ match the generation size of the pre-trained T2V model. Additionally, we resize the full video to $\mathbf{G}\in \mathbb{R}^{F\times h\times w\times 3}$ to serve as global guidance. Finally, both $\mathbf{P}_i$ and $\mathbf{G}$ are input into the VSR model for patch-level detail enhancement, and multi-patch joint modulation is involved to ensure coherence across patches. This formulation ensures compatibility with the pre-trained T2V model while allowing for flexible processing of videos of arbitrary sizes. + +# 3.1. Preliminary: Video Diffusion Base Model + +Our model is implemented on top of a pre-trained T2V diffusion model, which adopts the Rectified Flow framework [11] for the noise schedule and denoising process. The forward process is defined as straight paths between data distribution and a standard normal distribution, i.e. + +$$ +z _ {t} = (1 - t) z _ {0} + t \epsilon , \tag {1} +$$ + +where $\epsilon \in \mathcal{N}(0,I)$ and $t$ denotes the iterative timestep. To solve the denoising processing, we define a mapping between samples $z_{1}$ from a noise distribution $p_1$ to samples $z_{0}$ from a data distribution $p_0$ in terms of an ordinary differential equation (ODE), namely: + +$$ +d z _ {t} = v _ {\Theta} (z _ {t}, t) d t, \qquad (2) +$$ + +where the velocity $v$ is parameterized by the weights $\Theta$ of a neural network. For inference, we employ Euler discretization for Eq. 2 and perform discretization over the timestep interval at [0,1], starting at $t = 1$ . We then processed with iterative sampling with: $z_{t_i} = z_{t_j} + v_\Theta (z_{t_j},t_j)*(t_j - t_i)$ . + +For consideration of computing efficiency, the mapping is formulated in a latent space constructed with a + +3D Variational Auto-Encoder (VAE) [24]. The core is a Transformer-based diffusion model (DiT) [35], where each block is instantiated as a sequence of spatial attention, temporal attention, and cross-attention modules. Text prompts and other micro parameters (like time step, aspect ratio, etc) are used as generative conditions. Due to the full-attention property of Transformer, our model only supports the generation of a fixed number of tokens as it was trained with, i.e. $512 \times 512$ of diverse aspect-ratio. This is a common characteristic of almost all existing diffusion models for image or video generation [7, 8, 37, 49], and extending pre-trained diffusion models for higher resolution generation remains an open research topic [15, 17]. For VSR as a downstream task, we propose a patch-wise processing strategy to break this resolution limit. + +# 3.2. Patch Video Diffusion with Global Guidance + +To adapt a pre-trained T2V model for patch-level VSR, we must consider two key aspects: (i) enabling the input video to guide the base model's denoising process and (ii) bridging the gap between the base model's capabilities in generating full-frame content and patch-level content synthesis requirements. We achieve this through a dual-branch adapter, as detailed below. + +Patch Condition Branch. There are generally two approaches to incorporating visual conditions into pre-trained generative models: concatenating with noisy data in the noise space [39, 55], or fusing with the backbone feature through an adapter branch [16, 42]. The former involves modifying the first layers' channel count and necessitates base model fine-tuning, which is typically more data and computationally intensive. In contrast, the latter is more lightweight, as only a newly added branch requires training for feature adaptation. We adopt a hybrid solution that combines elements of both approaches. First, we employ an adapter comprising several Transformer blocks to extract features from the input video patch $\mathbf{P}_i$ and inject them into the base model's blocks. Additionally, considering that the base model was trained for full video generation, we need to adapt its embedded knowledge for patch-level generation. To this end, we allow the base model to be fine-tuned using a Low-Rank Adaptation (LoRA) [19]. An ablation study in Sec. 4.3 demonstrates the advantages of this design. + +Global Context Branch. To further guide content generation in a natural manner, we incorporate a global branch that extracts context from the full input video. This branch utilizes a Transformer-based encoder architecture that reduces the number of tokens to one quarter. The resulting context features are fused with each block of the base model through cross-attention $\{\mathbf{Q},\mathbf{K}_g,\mathbf{V}_g\}$ , where the $\mathbf{Q}$ is shared with the existing text-prompt cross-attention + +![](images/c616b517cf3fa13d021b21cb0439b0853bb1f1b36bf469ab38f991ff07c59f86.jpg) +Patch Partition + +![](images/96042419c5e573aa0a636f9f49339d24cfda762f719df1c5d8c7826ebcbfcd35.jpg) + +![](images/be6d6bf131eb8eb88408e25c5fc7a2de9ad7920ce5dc9e69a7f19ba32477145c.jpg) +Created auxiliary patches +Figure 2. Patch Partition Visualization. The input video is divided into non-overlapping segments, as the solid blue boxes mark. For joint modulation, auxiliary patches are created, indicated by the red dashed boxes, resulting in an overlapping ratio of $50\%$ . + +$\{\mathbf{Q}, \mathbf{K}_t, \mathbf{V}_t\}$ . Specifically, the input is constructed by concatenating the resized full video $\mathbf{V}_g$ (to accommodate a fixed number of tokens) with a binary map $\mathbf{M}_i$ indicating the target patch's $\mathbf{P}_i$ location within it. Our experiments show that the global context branch not only enhances the vividness of generated patch-level details but also improves consistency across patches. + +# 3.3. Multi-Patch Joint Modulation + +While the global guidance shared by patches is critical for maintaining color consistency, tonal uniformity, and semantic coherence of details, naive stitching of independently enhanced patches often introduces visible seams. These boundary artifacts primarily stem from the stochastic nature of the generative process when hallucinating textures on degraded or underdetermined regions. This is a fundamental challenge in ill-posed inverse problems where multiple high-resolution solutions may satisfy the constraints of a single low-resolution input. + +To mitigate this problem, we introduce inter-patch modulation during the detail generation process. Specifically, a training-free inference strategy called multi-diffusion [1] is adapted to solve our problem. This approach involves splitting the original video into overlapping patches and updating the overlapping regions by averaging contributions from all relevant patches at each denoising step. As illustrated in Fig. 2, instead of directly splitting the video into overlapping patches, we first partition it into non-overlapping segments, and then construct auxiliary patches by combining the halves of adjacent patches. This creates a new set of patch groups with an overlapping ratio of $50\%$ . Our experiment shows that directly taking the average in overlap regions suffers from artifacts of black holes or seamlines, as evidenced in Fig. 5 (middle). Instead, when fusing base patches and its related auxiliary patches, we manage to gradually reduce its influence of auxiliary patches from its center splitting line to two sides by assigning spatial weight maps to them. This method effectively reduces visible seams between patches and promotes overall consistency. A comparative result demonstrating the effectiveness of this approach is presented in Fig. 5. + +![](images/68ad9643dbcebd1718f9046f9a080576c2a1b0cf33ee8615283c380f45f13e45.jpg) +Video & prompt + +![](images/41475bfb1abf29ce9fdaa15823cd234524938f83cd36b1bff415fd29e324becd.jpg) +GT + +![](images/628c51588e7aac50f9a562daaf57daebb190c7608983ee483bb1200da80ed5e8.jpg) +Input + +![](images/e8f7cf94e7d802ec39a5b83809ced7a0e2df332049e63e37d1029f4174d6cdc0.jpg) +LaVie-SR + +![](images/5b2c429bd8f612d3ca6d489d5a21879a43a9bd6513f523e8d72452bbf2441801.jpg) +VEnhancer + +![](images/7fbd124ab6af09870e0450fc95051178252c106723c79b6b0ce19f01e2853695.jpg) +RealBasicVSR + +![](images/af4295e2bcd9827b1955e60443334d167e36615fee2c38ab5698b12b70747421.jpg) +Upscale-A-Video + +![](images/2b6f68ca39a7a522a01831a947ed9aad379c1765b505f213c9f1984a0c71788e.jpg) +Ours + +![](images/2521f2f0580ae6c73908a06cf11dbde1149d0d95bbddabc91e1232759a155ee2.jpg) +Video & prompt + +![](images/948f1cbbaebca6221de677ac87e4f76343270e3c2c9bd01100748c2b08a8d279.jpg) +GT + +![](images/eae1ad88aa5e70662ef692def10460a768bcc062a6fe019d85330ec21fb08583.jpg) +Input + +![](images/0ad2c1c1ddcb8c0e5b27e6078c2731998b57e600f9373c1ce438d3b4be21f5b4.jpg) +LaVie-SR + +![](images/dbff98e7545b6b10e99f21e6d2eddc771339dae2ca727d05f95a34a2b5545711.jpg) +VEnhancer + +![](images/d3b18950f91bc25cf0fccef143ae28b91c938aa251353ae863d160c53929acf5.jpg) +RealBasicVSR + +![](images/f85f09b0a3cb06cd5b871f369f2157d4abce0e2c60bcd299f61dbb6416293277.jpg) +Upscale-A-Video +Figure 3. Qualitative comparisons on SynVideo30 (test). These videos $\times 4$ super-resolution results. Zoom in for best view. + +![](images/c9f434d578b5247fe5eb3b717b2bcbaf8329fbd7082a3df20a0db0863a7e5ecf.jpg) +Ours + +# 4. Experiments + +To verify the superiority of PatchVSR, we compare it with SOTA methods, including VEnhancer [16], Upscale-A-Video [55], LaVie-SR [44] and RealBasicVSR [6]. We also make ablation study to analyze the effectiveness of our proposed designs. Besides, some extended analysis is conducted to demonstrate the capability of our method. + +# 4.1. Implementation Details + +Training Datasets We train our model using 460K self-collected high-quality video-text pairs, with each resolution from $1024 \times 1024$ to 2K. The text prompts are all captioned using LAVA captioner [26], and encoded by T5 text encoder [38] with no more than 256 tokens. + +Testing Datasets. We evaluate the VSR methods on two kinds of dataset, i.e. downscale synthetic videos and AIGC videos. For synthetic videos, since our employed base model is fixed to generate 77-frames video, we utilize a public dataset: REDS30 [33] for fair comparisons, which contains 100 frames. To evaluate higher resolution, we self-collect 30 high-quality 2K resolution videos, each containing more than 150 frames. Additionally, we collect and test on an AIGC dataset, which comprises 30 AI-generated videos from public text-to-video generative models [50]. + +Training Details. Our model is trained on 16 NVIDIA H800-80G GPUs with a total batch size of 64. AdamW [29] is used as the optimizer with a learning rate of $10^{-5}$ . The text prompt is randomly replaced by a null prompt with $10\%$ probability. To obtain the LR-HR training data pair, the HR video is first down-sampled using bilinear interpolation with down scale $s$ and up-sampled back to the original resolution. To match the size of the pre-trained model, + +we randomly crop a $512 \times 512$ area from the up-sampled LR video with aspect ratios from 0.5 to 2.0. To enhance the robustness of our model to different degradation scenarios, we utilize the noise augmentation technique by injecting noise into the input latent using a diffuse process. The noise timestep is randomly selected from 200 to 300, aiming at preserving the structure of the input video. We have additionally encoded noise timestep, downscale factor, and crop location as conditions for the model. We have replaced the text prompt with a fixed prompt for the patch condition branch, with the base model and global context branch using text prompts. To provide initialization weight, we followed the structure of the pre-trained model and sampled the layer index from it at equal intervals. + +Inference Details. We perform 50 PNDM [27] sampling steps with classifier-free guidance scale 5.0. We have also used timestep shift [12] with shift value 5.0. + +Evaluation Metrics. To evaluate the fidelity of our method on synthesized videos, we utilize PSNR, SSIM, and LPIPS [53]. For visual quality, we conduct our evaluation of commonly used quality assessment metrics MUSIQ [23] and DOVER [45]. Furthermore, we also include aesthetic score from a video generation benchmark VBench [20] for more comprehensive comparisons. + +# 4.2. Comparison with SOTA methods + +Qualitative Results. The visual results of synthetic and AIGC videos are presented in Fig. 3 and Fig. 4 respectively. It is observed that our method significantly outperforms existing CNN-based and diffusion-based models in both details generation and temporal consistency (Readers are recommended to check our provided video results). Specif + +![](images/bc96866090653a2c3deb647fd4601848388ecac12aa9a33925484016c57aa71e.jpg) +Figure 4. Qualitative comparisons on VideoGen30 (test). These videos are $\times 4$ super-resolution results. Zoom in for best view. + +
DatasetsMetricsRealBasicVSR [6]LaVie-SR [44]Upscale-A-Video [55]VEnhancer [16]Ours
REDS30PSNR ↑29.69828.73531.25728.54430.111
SSIM ↑0.6560.6220.6610.6110.647
LPIPS ↓0.3340.4070.3280.4210.326
DOVER ↑0.3050.3180.3100.3440.323
MUSIQ ↑40.58743.57742.57544.53344.956
Aesthetics ↑0.4220.4860.4710.4820.493
SynVideo30PSNR ↑33.50731.43333.43228.85630.857
SSIM ↑0.7760.7160.7280.6970.732
LPIPS ↓0.1850.1940.2050.1990.183
DOVER ↑0.5840.6120.5290.5500.597
MUSIQ ↑49.55748.14649.83943.53850.695
Aesthetics ↑0.4960.5090.4940.5030.520
VideoGen30DOVER ↑0.5880.6090.5100.5120.590
MUSIQ ↑47.56448.32946.89946.93050.559
Aesthetics ↑0.5470.6000.5740.5510.602
+ +Table 1. Quantitative comparisons on different VSR benchmarks from synthetic (REDS30, SynVideo30) and AIGC (VideoGen30) data. Red and blue numbers indicate the best and second best results respectively. + +ically, our method not only successfully recovers the details of the red silk scarf in Fig. 3, but also adds more details to the "Angkor Wat" as shown in Fig. 4, which others tend to generate blurry or unnatural results. In Fig. 5, we also illustrate full VSR results achieved through the multi-patch joint modulation scheme. It can be seen that creating auxiliary patches greatly alleviates the seam problem and significantly reduces the color difference between different patches, while further improving the contrast of the results. To show the advantage of our multi-patch joint modulation, we compare with a baseline method: stitching the patches in latent space and decoding the big latent together, which suffers obvious seamline artifacts. + +Quantitative Results. The quantitative comparison is shown in Tab. 1. It can be seen that the proposed method achieves the highest LPIPS score in both REDS30 and SynVideo30, indicating the high fidelity of our method for synthetic videos. However, our method show inferior performance in PSNR and SSIM. This can be explained by the fact that our method tends to generate rich and vivid details, which are not favorable to metrics that capture pixel-wise deviations (i.e. PSRN and SSIM). Instead, as a perceptual-level fidelity measure, LPIPS might be more informative to reflect the content fidelity for generative tasks. We evaluate the visual quality of generated high-resolution videos through comparisons with competing methods on MUSIQ + +![](images/5cd909b4dea0786c2a3bf31b0afc3166e930ba7fac770e6279b9dec80f8f5416.jpg) +Figure 5. Visual evaluation on multi-patch joint modulation. Here the super-resolved videos of 2K resolution are used for comparison, where $3 \times 3 = 9$ patches are involved. The top, middle, bottom denote input video, result of stitching latent of patches, and result of ours respectively. + +and DOVER metrics, demonstrating our approach's superiority. This performance reflects contributions from both the base model's capabilities and the effectiveness of our method in leveraging full-frame generation for patch-level detail synthesis. Interestingly, our approach even preserves subjective properties like aesthetic quality from the base model, as evidenced by the aesthetic scores. This highlights our framework's ability to inherit and maintain critical perceptual characteristics during the enhancement process. + +# 4.3. Ablation Study + +We conduct ablation studies on key designs. Considering the limited space, we provide a quantitative evaluation here and the typical visual results are compared in the supplementary material. + +Global Branch. To validate the effectiveness of the global branch, We build a baseline by removing the global branch from our method. As shown in Tab. 2, removing the global branch results in significant performance deterioration, leading to worse DOVER, MUSIQ, and aesthetic score. Besides, to check the necessity of using the patch location embeddings, we build a baseline by removing the + +
ComponentDOVER ↑MUSIQ ↑Aesthetics ↑
w/o global branch0.50246.0740.589
w/o LoRA0.58250.4960.600
w/o location embed0.57450.0840.601
w/o fixed prompt0.56248.1330.597
Ours0.59050.5590.602
+ +Table 2. Ablation study of different model components on VideoGen30 test dataset. The **bold** numbers indicate the best. + +patch position information from the global branch input. As shown in Tab. 2, after incorporating the location embedding, we obtain an increase of DOVER from 0.574 to 0.590, which demonstrates the effectiveness of utilizing the location mask as the patch position information. + +Text Prompts. Text prompts are used in the base model and the conditional global branch. As shown in Tab. 2, we observe that replacing the text prompt with a fixed prompt leads to superior results. We conjecture that the use of text prompts in patch condition branch brings about a mismatch between the video patch and the global semantic information, which in turn destroys the dynamic visual prior in the pre-trained model aligned to the text during training. + +Base model finetuning. Since the data distribution of local patches deviates from the full view of the range of common videos, we propose to fine-tune the base model instead of freezing it. It can be seen in Tab. 2 that adding LoRA results in improved performance across all metrics, which demonstrates the effectiveness of utilizing LoRA to adapt to the distribution of local video patches. + +# 4.4. Analysis and Discussion + +We would like to further study some characteristics of our proposed method, including detail generation magnitude, robustness to patch semantics, and computation efficiency. + +Detail generation magnitude. The input and output sizes of our VSR model are the same and our model can synthesize different levels of details. In our method, a scale factor is used to guide the model to generate reasonable details that should be presented on the input video if it is captured normally by the camera at that resolution/size. So, the scale factor actually plays the role of affecting the magnitude of details. Here we illustrate the effects of using different scale factors as conditions. In Fig. 6, increasing the scale factor progressively adds more high-resolution details to the results, which demonstrates the flexibility of our method to generate results with different detail magnitudes. + +Robustness to Patch Semantics. As stated above, our approach will partition the input video into patches in a content-agnostic manner. So, some patches can contain incomplete or even unclear semantics, which are more challenging for the model with the generative reasonable details. + +![](images/7da4b1b25bf002ffdc5a21b500eb30f3e4a0d2f77858e7a79e2fa0a14dc5122a.jpg) +Input patch + +![](images/ac762976b9219bbc2a58ad7f99c69cec6fc6148e28b3723e9a7397d266886087.jpg) +Upscale factor $= 1$ + +![](images/975dd54e4f3806b1b1c33c798ed1ed7d4bd738fc82167a6398e3ce80381faa0c.jpg) +Upscale factor=2 + +![](images/b7f87ea1be7a60375e5304f789d326828c09495ebd1fe020186351bdfe992437.jpg) +Upscale factor $= 3$ + +![](images/07fe9ccade066828e437fea0e2e95984c92167cede84fbf01c3bf4a5dac82076.jpg) +Upscale factor $= 4$ +Figure 6. Comparison of different up-sampling scales. Note that the upscale factor is uniformly randomly sampled within the range of 1 to 4 during the training phase. + +![](images/f4e91ad53575c5b33980bea0bff110d272880726f10c102731aa4c9f9b9987c8.jpg) +Upscale factor $= 5$ + +
MethodsTime (s) ↓Memory (G) ↓
LaVie-SR226168
Upscale-A-Video274347
VEnhancer156262
Ours68040
+ +Table 3. Computation efficiency for 2K video generation. + +It is interesting to study whether the performance varies across patches with diverse semantic quality. It can be seen in Fig. 7 that our model can successfully handle the video patches with incomplete semantics. As the global semantic tokens are extracted and input to the base model using global cross-attention modules, the model learns to utilize global semantic information to complete the semantics in the local video patch, which improves the robustness of our model against incomplete semantics. + +Computation efficiency. Our method is more efficient than existing pre-trained T2V-based VSR methods because the self-attention is conducted within patches. Here presents a simple deduce of the computation complexity against patch size: given a video of $n$ tokens and each patch of $m$ tokens, the patch number is $k = \frac{n}{m}$ and the inter-token computation times are $k * \left( \frac{n}{k} \right)^2 = n * m$ . Since the base model we adopted uses fixed $512 \times 512$ resolution, our patch size follows this size in our experiments. The inference time and memory usage of generating a 2K resolution video is illustrated in Tab. 3. It can be seen that, after dividing the input video into patches, the computation burden is signif + +![](images/3d82803f04939eff4cb0b2aa507b73fe7c770ccdf0efaf0960017e6600ecdcbb.jpg) + +![](images/c64a63248aeb32528b40723527be436a4915b6de7f92ba4edcf20ba1dc4fb7dd.jpg) + +![](images/6b8930ed5dc5789277d744bc994c28df541a6e9d6a948e6627a12876fbd6ddff.jpg) +Incomplete semantics +Complete semantics +Figure 7. Robustness against patch semantic incompleteness. + +icantly reduced, further promoting the applications of generative VSR models. Thanks to the computation efficiency and the inference flexibility, our method even support 4K video generation, which are shown in the supplementary materials. + +**Limitation.** Our method has limitations in two aspects. Firstly, it experiences noticeable performance degradation when applied to real-world low-resolution images, due to the lack of specialized data augmentation in our training pairs. Since the primary focus of this paper is to address the resolution limitations of using pre-trained base models for VSR, extending our approach to handle real-world noise remains an interesting direction for future work. Secondly, our approach inherits an iterative inference scheme from the base model, which introduces efficiency issues for practical applications. For VSR, the low-resolution input video provides significant opportunities for algorithm optimization to reduce the number of inference steps, which will be another area of future research. + +# 5. Conclusion + +In this paper, we present the first exploration of utilizing a pre-trained T2V base model for patch-level video superresolution (VSR). To accomplish this, we propose an effective dual-branch adapter consisting of a patch condition branch and a global context branch. The global context branch is crucial for bridging the gap between full-video generative capabilities and patch-level detail synthesis. Additionally, our proposed multi-patch joint modulation scheme achieves consistent results across patches. Thanks to the flexibility of our PatchVSR approach, it outperforms existing VSR methods in generating faithful details while maintaining high computational efficiency. We hope this work will inspire further research in this area. + +# Acknowledgement + +This work was supported by National Natural Science Foundation of China (NSFC) under Grant No.62406167, U24B6012. + +# References + +[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning (ICML), 2023. 4 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2 +[3] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2 +[4] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 2 +[5] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. +[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5962-5971, 2022. 2, 5, 6 +[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 4 +[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 4 +[9] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2 +[10] Shian Du, Xiaotian Cheng, Qi Qian, Henglu Wei, Yi Xu, and Xiangyang Ji. Efficient personalized text-to-image generation by leveraging textual subspace. arXiv preprint arXiv:2407.00608, 2024. 2 + +[11] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 3 +[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 5 +[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[14] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17441-17451, 2022. 2 +[15] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, Ying Shan, and Bihan Wen. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation. In European Conference on Computer Vision (ECCV), 2024. 2, 4 +[16] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation. arXiv preprint arXiv:2407.07667, 2024. 1, 2, 4, 5, 6 +[17] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In International Conference on Learning Representations (ICLR), 2024. 4 +[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[19] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 4 +[20] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5 +[21] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent + +structure-detail network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 645-660. Springer, 2020. 2 +[22] Mehran Jeelani, Noshaba Cheema, Klaus Illgner-Fehns, Philipp Slusallek, Sunil Jaiswal, et al. Expanding synthetic real-world degradations for blind video super resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1199-1208, 2023. 2 +[23] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5 +[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations (ICLR), 2014. 4 +[25] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5687-5696, 2022. 2 +[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5 +[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 5 +[28] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2 +[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[30] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2 +[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2 +[32] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2 +[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 5 +[34] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 2 + +[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4 +[36] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2 +[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 4 +[38] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4 +[40] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. Advances in Neural Information Processing Systems, 35:36081-36093, 2022. 2 +[41] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2 +[42] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C. K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 2024. 4 +[43] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2 +[44] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 5, 6 +[45] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20144-20154, 2023. 5 +[46] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2 +[47] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying + +Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2 +[48] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4781-4790, 2021. 2 +[49] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4 +[50] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 5 +[51] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1-15, 2024. 2 +[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[54] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2 +[55] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 1, 2, 4, 5, 6 \ No newline at end of file diff --git a/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/images.zip b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5be9203633a5e51ba9850b3cc3b932cb3057687d --- /dev/null +++ b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a20896daeb8565da292ab70bf517585c4f3e8267f41accd916119a28becaead3 +size 725010 diff --git a/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/layout.json b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a20ecd5eecc51b806b148b2c13b4d947829b3a8c --- /dev/null +++ b/2025/PatchVSR_ Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution/layout.json @@ -0,0 +1,9058 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 102, + 492, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 102, + 492, + 139 + ], + "spans": [ + { + "bbox": [ + 119, + 102, + 492, + 139 + ], + "type": "text", + "content": "PatchVSR: Breaking Video Diffusion Resolution Limits with Patch-wise Video Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "spans": [ + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "text", + "content": "Shian Du" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{1*‡}" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "text", + "content": ", Menghan Xia" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{2*‡}" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "text", + "content": ", Chang Liu" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "text", + "content": ", Xintao Wang" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "text", + "content": ", Jing Wang" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "text", + "content": ", Pengfei Wan" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "text", + "content": ", Di Zhang" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "text", + "content": ", Xiangyang Ji" + }, + { + "bbox": [ + 97, + 165, + 512, + 196 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 197, + 493, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 197, + 493, + 212 + ], + "spans": [ + { + "bbox": [ + 115, + 197, + 493, + 212 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 197, + 493, + 212 + ], + "type": "text", + "content": "Tsinghua University, " + }, + { + "bbox": [ + 115, + 197, + 493, + 212 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 197, + 493, + 212 + ], + "type": "text", + "content": "Kuaishou Technology, " + }, + { + "bbox": [ + 115, + 197, + 493, + 212 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 197, + 493, + 212 + ], + "type": "text", + "content": "Beijing Institute of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 250, + 200, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 250, + 200, + 262 + ], + "spans": [ + { + "bbox": [ + 151, + 250, + 200, + 262 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 276, + 297, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 297, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 297, + 577 + ], + "type": "text", + "content": "Pre-trained video generation models hold great potential for generative video super-resolution (VSR). However, adapting them for full-size VSR, as most existing methods do, suffers from unnecessary intensive full-attention computation and fixed output resolution. To overcome these limitations, we make the first exploration into utilizing video diffusion priors for patch-wise VSR. This is non-trivial because pre-trained video diffusion models are not native for patch-level detail generation. To mitigate this challenge, we propose an innovative approach, called PatchVSR, which integrates a dual-stream adapter for conditional guidance. The patch branch extracts features from input patches to maintain content fidelity while the global branch extracts context features from the resized full video to bridge the generation gap caused by incomplete semantics of patches. Particularly, we also inject the patch's location information into the model to better contextualize patch synthesis within the global video frame. Experiments demonstrate that our method can synthesize high-fidelity, high-resolution details at the patch level. A tailor-made multi-patch joint modulation is proposed to ensure visual consistency across individually enhanced patches. Due to the flexibility of our patch-based paradigm, we can achieve highly competitive 4K VSR based on a " + }, + { + "bbox": [ + 55, + 276, + 297, + 577 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 276, + 297, + 577 + ], + "type": "text", + "content": " resolution base model, with extremely high efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 622, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 682 + ], + "type": "text", + "content": "Video super-resolution (VSR) has long been a challenging research problem in the field of computer vision. With the advent of deep learning, significant advancements have been made in end-to-end VSR models trained on curated datasets. However, due to limited model capacity and data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 384 + ], + "type": "text", + "content": "coverage, these models typically achieve sharp object structures or coarse textures but struggle to synthesize faithful details. Recently, the emergence of diffusion generative models has enabled open-domain video generation of remarkable quality. The generative capabilities of these foundation models present new opportunities for video super-resolution techniques [16, 55]. Nevertheless, existing base models are generally constrained to generating videos of fixed sizes due to their inherent architecture. This limitation is inherited in their downstream super-resolution models, making it challenging for them to support arbitrary resolution outputs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 387, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 567 + ], + "type": "text", + "content": "This study is motivated by the insight that feature attention in super-resolution models is more locally focused compared to generation models. This localization is reasonable, as super-resolution tasks have a low-resolution video as a reference, allowing details to be generated based on neighborhood semantics rather than requiring intensive global coherence considerations. To leverage this insight, we propose a patch-based strategy for video superresolution. The concept is straightforward and intuitive: we split the input video into patches sized to be compatible with the pre-trained base model, enhance each patch individually, and then combine them to produce the final result. This approach enables the achievement of super-resolution results at arbitrary resolutions, independent of the base model's inherent resolution constraints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "The key challenge lies in that video generation-based models are trained on complete video frames and face noticeable performance degradation for patch-level generation. To adapt a pre-trained video generation model for generating patch-level details in super-resolution, we propose an innovative approach, called PatchVSR, which integrates a dual-stream adapter for conditional guidance. The local branch extracts features from the input patches to guide the base model for detail synthesis adhering to the input, while the global branch captures contextual information from the entire input frame to bridge the generation gap caused by incomplete semantics of patches. To make the context guid-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 295, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 295, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 295, + 703 + ], + "type": "text", + "content": "* Work done during an internship at KwaiVGI, Kuaishou Technology." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 703, + 220, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 220, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 220, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 69, + 703, + 220, + 712 + ], + "type": "text", + "content": " Equal contribution. " + }, + { + "bbox": [ + 69, + 703, + 220, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 69, + 703, + 220, + 712 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17799" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "ance more customized, the global branch also incorporates a binary mask indicating the target patch position as input. Additionally, we employ LoRA technology to fine-tune the base model, further improving its adaptability for generating patch-level details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "type": "text", + "content": "Experimental results demonstrate that our method effectively adapts the generative ability of the base model to produce faithful details at the patch level. Specifically, to alleviate the boundary inconsistency across patches, we propose a training-free multi-patch joint modulation scheme, where auxiliary bridging patches are created to allow feature interaction across patches during generation. Thanks to the flexibility of PatchVSR, our method, based on a pre-trained " + }, + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "type": "text", + "content": " resolution video base model, supports super-resolution video generation up to 4K resolution, while significantly outperforming current state-of-the-art methods in terms of fidelity and detail realism. Notably, compared to existing methods that operate on full-frame videos, our approach processes relatively smaller patches, without full-range attention computation, demonstrating a significant efficiency advantage." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 326, + 253, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 326, + 253, + 337 + ], + "spans": [ + { + "bbox": [ + 67, + 326, + 253, + 337 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 338, + 295, + 458 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 61, + 338, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 338, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 61, + 338, + 295, + 386 + ], + "type": "text", + "content": "- We introduce an innovative approach to overcome the resolution limitations of using pre-trained video diffusion models for VSR. Our method achieves any-resolution VSR output based on a fixed-resolution base model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 386, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 386, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 61, + 386, + 295, + 422 + ], + "type": "text", + "content": "- We propose a novel dual-branch adapter paradigm that effectively harnesses the capabilities of pre-trained video models for patch-level detail generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 422, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 422, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 61, + 422, + 295, + 458 + ], + "type": "text", + "content": "- Through extensive experimentation, we demonstrate the significant superiority of our method in terms of faithful detail generation and computational efficiency." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 471, + 142, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 142, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 142, + 483 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 491, + 195, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 195, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 195, + 503 + ], + "type": "text", + "content": "2.1. Text-to-video Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "With the widespread use of pre-trained text-to-image models [10, 13, 18, 32, 39, 52] and the preparation of large-scale text-video pairing datasets [9, 34], substantial efforts have been made to train a large-scale text-to-video model. Instead of training from scratch, some methods [2, 41, 46, 51] focus on inserting temporal modules and fine-tuning them or all parameters on video data. To avoid catastrophic forgetting and degradation on visual quality, they often adopt image-video joint training strategy [7, 8, 47]. Recently, with Diffusion Transformer (DiT) [36] being proposed, it has replaced the traditional 3D UNet architecture used by previous approaches with better generation quality and scalability, and has become the dominant framework for text-to-video generation [28, 30, 49]. Although achieving high visual quality, extending them to higher-resolution is non-trivial since it requires substantial training resource and high-quality high-resolution video dataset, which is un" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "available by this moment. To bypass this difficulty, current methods often adopt a cascade framework [15, 44, 54]: the low-resolution video is first generated by the base model using text prompt, and the refinement model outputs high-resolution video through a noisig-denoising process [31] by taking the low-resolution video as condition. However, existing methods are limited to generate full video at once, which still demands a high training overhead. At the same time, the low inference efficiency and high memory usage also hinder its downstream applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 204, + 448, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 448, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 448, + 217 + ], + "type": "text", + "content": "2.2. Video Super-Resolution" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 223, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 715 + ], + "type": "text", + "content": "Video super-resolution aims at restoring high-resolution (HR) video frames from its degraded low-resolution (LR) counterparts and enhancing its visual quality. Traditional methods often assume a fixed degradation process and their performance deteriorates significantly when handling AI-generated and real-world videos. Due to the lack of real-world LR-HR video data pair for training, RealVSR [48] collects LR-HR video data pair by utilizing the multicamera system of iPhone. Nevertheless, the sequences are captured by a single device, which only contains camera-specific degradations. To avoid burdensome data collection, other methods [6, 22, 55] use stochastic combinations of multi-order degradation processes to simulate real-world degradation processes. While performing more robust when handling real-world cases, due to the absence of deep generative prior, it is challenging for traditional CNN-based [4-6, 21, 43] and transformer-based [3, 14, 25, 40] methods to generate photo-realistic details and textures. By utilizing a pre-trained image diffusion model like Stable Diffusion (SD) " + }, + { + "bbox": [ + 313, + 223, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 223, + 555, + 715 + ], + "type": "text", + "content": " Upscaler, Lavie-SR [44] and Upscale-A-Video [55] circumvent the need for exhaustive training from scratch and exhibit improved performance in fixed spatial scale within a short sequence. To improve temporal consistency for long sequence, I2VGen-XL [54] and VEnhancer [16] utilize a pre-trained low-resolution video generative model as the base model, and directly fine-tune the model parameters using high-resolution video dataset following noisy-denoising manner [31]. However, the aforementioned methods are all limited to generate full video at once, which requires expanding the training resolution of pre-trained model, bringing low training efficiency and slow convergence speed. Meanwhile, high resolution pretrained video generative models are currently unavailable to access, hindering the scalability of video super-resolution methods with deep generative prior. To tackle this problem, we make the first exploration of utilizing pre-trained video generative model for patch-wise video super-resolution. By maintaining the training resolution of the pre-trained model, our approach not only avoids significant training overhead, making it possible to generate 4K resolution videos using a " + }, + { + "bbox": [ + 313, + 223, + 555, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 223, + 555, + 715 + ], + "type": "text", + "content": " resolution pre-trained model. At the same time," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17800" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 556, + 222 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 556, + 222 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 556, + 222 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 556, + 222 + ], + "type": "image", + "image_path": "3a130280773855dd3be87b1e9a0c1c055c7a493d43d4caee1aa4059ad7141be2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 230, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 54, + 230, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 54, + 230, + 555, + 298 + ], + "type": "text", + "content": "Figure 1. Flowchart of our PatchVSR. Building upon a pre-trained latent T2V model, we incorporate a patch condition branch and a global context branch. These branches extract features from partitioned video patches and the resized full video together with a binary mask that indicates the location of the ROI patch, respectively. Particularly, local patch features are added to the output of each block, while the global context feature is fused with the backbone feature through newly introduced cross-attention modules (G-CA). For simplicity, we have omitted other conditional inputs such as text prompts and time steps from this diagram. The processed patches are fused via a joint modulation scheme to produce a coherent super-resolution video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 307, + 298, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 298, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 298, + 346 + ], + "type": "text", + "content": "our method significantly improves inference efficiency and reduces memory usage, facilitating downstream applications of video super-resolution methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 113, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 113, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 113, + 369 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 377, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 593 + ], + "type": "text", + "content": "We aim to develop a video super-resolution (VSR) framework designed to restore degraded videos of arbitrary resolutions by generating visually coherent and structurally faithful details. Our primary innovation involves strategically adapting pre-trained text-to-video (T2V) diffusion models—powerful generative tools typically constrained to fixed output size (denoted as " + }, + { + "bbox": [ + 54, + 377, + 295, + 593 + ], + "type": "inline_equation", + "content": "s = h \\times w" + }, + { + "bbox": [ + 54, + 377, + 295, + 593 + ], + "type": "text", + "content": ")—through a curated conditioning paradigm. To bypass their inherent token-length limitations, we introduce a patch-based processing strategy: the input video is decomposed into " + }, + { + "bbox": [ + 54, + 377, + 295, + 593 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 54, + 377, + 295, + 593 + ], + "type": "text", + "content": "-sized patches, each independently enhanced using the diffusion model, and then reintegrated into a unified high-resolution output through optimized spatial-temporal fusion. This approach is grounded in the distinct characteristics of VSR compared to video generation. While video synthesis demands global coherence across full-frame contexts, superresolution prioritizes localized detail restoration atop existing structural content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": "Specifically, we augment the base model with two conditional branches: a patch branch that extracts features from input patches, and a global branch that extracts context from the full video. This global context mainly aids in generating patch-level details from the T2V model that was originally trained on full-frame videos. An overview of the model architecture is illustrated in Fig. 1. Given a low-resolution image " + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_l\\in \\mathbb{R}^{F\\times H\\times W\\times 3}" + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": " and the target upscale factor " + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": " , our process begins by upscaling " + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_l" + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": " to the desired resolution " + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_l^\\uparrow \\in \\mathbb{R}^{F\\times [H\\cdot k]\\times [W\\cdot k]\\times 3}" + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": " using bicubic" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "text", + "content": "interpolation per frame. We then partition " + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_l^\\dagger" + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "text", + "content": " into patches " + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_i\\in \\mathbb{R}^{F\\times h\\times w\\times 3}\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "text", + "content": ", where the patch dimensions " + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "text", + "content": " match the generation size of the pre-trained T2V model. Additionally, we resize the full video to " + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{G}\\in \\mathbb{R}^{F\\times h\\times w\\times 3}" + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "text", + "content": " to serve as global guidance. Finally, both " + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 307, + 555, + 429 + ], + "type": "text", + "content": " are input into the VSR model for patch-level detail enhancement, and multi-patch joint modulation is involved to ensure coherence across patches. This formulation ensures compatibility with the pre-trained T2V model while allowing for flexible processing of videos of arbitrary sizes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 434, + 533, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 533, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 533, + 448 + ], + "type": "text", + "content": "3.1. Preliminary: Video Diffusion Base Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 451, + 555, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 511 + ], + "type": "text", + "content": "Our model is implemented on top of a pre-trained T2V diffusion model, which adopts the Rectified Flow framework [11] for the noise schedule and denoising process. The forward process is defined as straight paths between data distribution and a standard normal distribution, i.e." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 392, + 520, + 555, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 520, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 392, + 520, + 555, + 534 + ], + "type": "interline_equation", + "content": "z _ {t} = (1 - t) z _ {0} + t \\epsilon , \\tag {1}", + "image_path": "db89b84bccdf56d4d414c2de3dabd91acaa55bde831b47304e42417951fdd541.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "text", + "content": " denotes the iterative timestep. To solve the denoising processing, we define a mapping between samples " + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "inline_equation", + "content": "z_{1}" + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "text", + "content": " from a noise distribution " + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "text", + "content": " to samples " + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "text", + "content": " from a data distribution " + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 540, + 556, + 601 + ], + "type": "text", + "content": " in terms of an ordinary differential equation (ODE), namely:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 395, + 609, + 555, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 609, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 395, + 609, + 555, + 624 + ], + "type": "interline_equation", + "content": "d z _ {t} = v _ {\\Theta} (z _ {t}, t) d t, \\qquad (2)", + "image_path": "ea3c622a2ffbe9261a654e39648e6ce0c6abfd5d8ffdd85dafab5d3456ce1fac.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 556, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 691 + ], + "type": "text", + "content": "where the velocity " + }, + { + "bbox": [ + 313, + 629, + 556, + 691 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 629, + 556, + 691 + ], + "type": "text", + "content": " is parameterized by the weights " + }, + { + "bbox": [ + 313, + 629, + 556, + 691 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 629, + 556, + 691 + ], + "type": "text", + "content": " of a neural network. For inference, we employ Euler discretization for Eq. 2 and perform discretization over the timestep interval at [0,1], starting at " + }, + { + "bbox": [ + 313, + 629, + 556, + 691 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 629, + 556, + 691 + ], + "type": "text", + "content": ". We then processed with iterative sampling with: " + }, + { + "bbox": [ + 313, + 629, + 556, + 691 + ], + "type": "inline_equation", + "content": "z_{t_i} = z_{t_j} + v_\\Theta (z_{t_j},t_j)*(t_j - t_i)" + }, + { + "bbox": [ + 313, + 629, + 556, + 691 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "type": "text", + "content": "For consideration of computing efficiency, the mapping is formulated in a latent space constructed with a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "17801" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": "3D Variational Auto-Encoder (VAE) [24]. The core is a Transformer-based diffusion model (DiT) [35], where each block is instantiated as a sequence of spatial attention, temporal attention, and cross-attention modules. Text prompts and other micro parameters (like time step, aspect ratio, etc) are used as generative conditions. Due to the full-attention property of Transformer, our model only supports the generation of a fixed number of tokens as it was trained with, i.e. " + }, + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": " of diverse aspect-ratio. This is a common characteristic of almost all existing diffusion models for image or video generation [7, 8, 37, 49], and extending pre-trained diffusion models for higher resolution generation remains an open research topic [15, 17]. For VSR as a downstream task, we propose a patch-wise processing strategy to break this resolution limit." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 259, + 286, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 286, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 286, + 271 + ], + "type": "text", + "content": "3.2. Patch Video Diffusion with Global Guidance" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "text", + "content": "To adapt a pre-trained T2V model for patch-level VSR, we must consider two key aspects: (i) enabling the input video to guide the base model's denoising process and (ii) bridging the gap between the base model's capabilities in generating full-frame content and patch-level content synthesis requirements. We achieve this through a dual-branch adapter, as detailed below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 375, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 604 + ], + "type": "text", + "content": "Patch Condition Branch. There are generally two approaches to incorporating visual conditions into pre-trained generative models: concatenating with noisy data in the noise space [39, 55], or fusing with the backbone feature through an adapter branch [16, 42]. The former involves modifying the first layers' channel count and necessitates base model fine-tuning, which is typically more data and computationally intensive. In contrast, the latter is more lightweight, as only a newly added branch requires training for feature adaptation. We adopt a hybrid solution that combines elements of both approaches. First, we employ an adapter comprising several Transformer blocks to extract features from the input video patch " + }, + { + "bbox": [ + 55, + 375, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 55, + 375, + 296, + 604 + ], + "type": "text", + "content": " and inject them into the base model's blocks. Additionally, considering that the base model was trained for full video generation, we need to adapt its embedded knowledge for patch-level generation. To this end, we allow the base model to be fine-tuned using a Low-Rank Adaptation (LoRA) [19]. An ablation study in Sec. 4.3 demonstrates the advantages of this design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": "Global Context Branch. To further guide content generation in a natural manner, we incorporate a global branch that extracts context from the full input video. This branch utilizes a Transformer-based encoder architecture that reduces the number of tokens to one quarter. The resulting context features are fused with each block of the base model through cross-attention " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q},\\mathbf{K}_g,\\mathbf{V}_g\\}" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": " is shared with the existing text-prompt cross-attention" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 85, + 399, + 137 + ], + "blocks": [ + { + "bbox": [ + 317, + 85, + 399, + 137 + ], + "lines": [ + { + "bbox": [ + 317, + 85, + 399, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 85, + 399, + 137 + ], + "type": "image", + "image_path": "c616b517cf3fa13d021b21cb0439b0853bb1f1b36bf469ab38f991ff07c59f86.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 156, + 383, + 165 + ], + "lines": [ + { + "bbox": [ + 334, + 156, + 383, + 165 + ], + "spans": [ + { + "bbox": [ + 334, + 156, + 383, + 165 + ], + "type": "text", + "content": "Patch Partition" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 408, + 97, + 422, + 125 + ], + "blocks": [ + { + "bbox": [ + 408, + 97, + 422, + 125 + ], + "lines": [ + { + "bbox": [ + 408, + 97, + 422, + 125 + ], + "spans": [ + { + "bbox": [ + 408, + 97, + 422, + 125 + ], + "type": "image", + "image_path": "96042419c5e573aa0a636f9f49339d24cfda762f719df1c5d8c7826ebcbfcd35.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 71, + 553, + 151 + ], + "blocks": [ + { + "bbox": [ + 427, + 71, + 553, + 151 + ], + "lines": [ + { + "bbox": [ + 427, + 71, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 427, + 71, + 553, + 151 + ], + "type": "image", + "image_path": "be6d6bf131eb8eb88408e25c5fc7a2de9ad7920ce5dc9e69a7f19ba32477145c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 156, + 531, + 166 + ], + "lines": [ + { + "bbox": [ + 449, + 156, + 531, + 166 + ], + "spans": [ + { + "bbox": [ + 449, + 156, + 531, + 166 + ], + "type": "text", + "content": "Created auxiliary patches" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 170, + 555, + 214 + ], + "lines": [ + { + "bbox": [ + 313, + 170, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 555, + 214 + ], + "type": "text", + "content": "Figure 2. Patch Partition Visualization. The input video is divided into non-overlapping segments, as the solid blue boxes mark. For joint modulation, auxiliary patches are created, indicated by the red dashed boxes, resulting in an overlapping ratio of " + }, + { + "bbox": [ + 313, + 170, + 555, + 214 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 170, + 555, + 214 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Q}, \\mathbf{K}_t, \\mathbf{V}_t\\}" + }, + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "type": "text", + "content": ". Specifically, the input is constructed by concatenating the resized full video " + }, + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_g" + }, + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "type": "text", + "content": " (to accommodate a fixed number of tokens) with a binary map " + }, + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "type": "text", + "content": " indicating the target patch's " + }, + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 313, + 220, + 555, + 304 + ], + "type": "text", + "content": " location within it. Our experiments show that the global context branch not only enhances the vividness of generated patch-level details but also improves consistency across patches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 313, + 477, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 313, + 477, + 324 + ], + "spans": [ + { + "bbox": [ + 314, + 313, + 477, + 324 + ], + "type": "text", + "content": "3.3. Multi-Patch Joint Modulation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 331, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 450 + ], + "type": "text", + "content": "While the global guidance shared by patches is critical for maintaining color consistency, tonal uniformity, and semantic coherence of details, naive stitching of independently enhanced patches often introduces visible seams. These boundary artifacts primarily stem from the stochastic nature of the generative process when hallucinating textures on degraded or underdetermined regions. This is a fundamental challenge in ill-posed inverse problems where multiple high-resolution solutions may satisfy the constraints of a single low-resolution input." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "content": "To mitigate this problem, we introduce inter-patch modulation during the detail generation process. Specifically, a training-free inference strategy called multi-diffusion [1] is adapted to solve our problem. This approach involves splitting the original video into overlapping patches and updating the overlapping regions by averaging contributions from all relevant patches at each denoising step. As illustrated in Fig. 2, instead of directly splitting the video into overlapping patches, we first partition it into non-overlapping segments, and then construct auxiliary patches by combining the halves of adjacent patches. This creates a new set of patch groups with an overlapping ratio of " + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "content": ". Our experiment shows that directly taking the average in overlap regions suffers from artifacts of black holes or seamlines, as evidenced in Fig. 5 (middle). Instead, when fusing base patches and its related auxiliary patches, we manage to gradually reduce its influence of auxiliary patches from its center splitting line to two sides by assigning spatial weight maps to them. This method effectively reduces visible seams between patches and promotes overall consistency. A comparative result demonstrating the effectiveness of this approach is presented in Fig. 5." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17802" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 190, + 145 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 190, + 145 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 190, + 145 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 190, + 145 + ], + "type": "image", + "image_path": "68ad9643dbcebd1718f9046f9a080576c2a1b0cf33ee8615283c380f45f13e45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 146, + 148, + 156 + ], + "lines": [ + { + "bbox": [ + 94, + 146, + 148, + 156 + ], + "spans": [ + { + "bbox": [ + 94, + 146, + 148, + 156 + ], + "type": "text", + "content": "Video & prompt" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 99, + 301, + 145 + ], + "blocks": [ + { + "bbox": [ + 219, + 99, + 301, + 145 + ], + "lines": [ + { + "bbox": [ + 219, + 99, + 301, + 145 + ], + "spans": [ + { + "bbox": [ + 219, + 99, + 301, + 145 + ], + "type": "image", + "image_path": "41475bfb1abf29ce9fdaa15823cd234524938f83cd36b1bff415fd29e324becd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 146, + 266, + 154 + ], + "lines": [ + { + "bbox": [ + 254, + 146, + 266, + 154 + ], + "spans": [ + { + "bbox": [ + 254, + 146, + 266, + 154 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 157, + 138, + 204 + ], + "blocks": [ + { + "bbox": [ + 58, + 157, + 138, + 204 + ], + "lines": [ + { + "bbox": [ + 58, + 157, + 138, + 204 + ], + "spans": [ + { + "bbox": [ + 58, + 157, + 138, + 204 + ], + "type": "image", + "image_path": "628c51588e7aac50f9a562daaf57daebb190c7608983ee483bb1200da80ed5e8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 205, + 107, + 213 + ], + "lines": [ + { + "bbox": [ + 89, + 205, + 107, + 213 + ], + "spans": [ + { + "bbox": [ + 89, + 205, + 107, + 213 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 139, + 157, + 219, + 204 + ], + "blocks": [ + { + "bbox": [ + 139, + 157, + 219, + 204 + ], + "lines": [ + { + "bbox": [ + 139, + 157, + 219, + 204 + ], + "spans": [ + { + "bbox": [ + 139, + 157, + 219, + 204 + ], + "type": "image", + "image_path": "e8f7cf94e7d802ec39a5b83809ced7a0e2df332049e63e37d1029f4174d6cdc0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 204, + 196, + 212 + ], + "lines": [ + { + "bbox": [ + 162, + 204, + 196, + 212 + ], + "spans": [ + { + "bbox": [ + 162, + 204, + 196, + 212 + ], + "type": "text", + "content": "LaVie-SR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 157, + 301, + 204 + ], + "blocks": [ + { + "bbox": [ + 220, + 157, + 301, + 204 + ], + "lines": [ + { + "bbox": [ + 220, + 157, + 301, + 204 + ], + "spans": [ + { + "bbox": [ + 220, + 157, + 301, + 204 + ], + "type": "image", + "image_path": "5b2c429bd8f612d3ca6d489d5a21879a43a9bd6513f523e8d72452bbf2441801.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 205, + 279, + 213 + ], + "lines": [ + { + "bbox": [ + 242, + 205, + 279, + 213 + ], + "spans": [ + { + "bbox": [ + 242, + 205, + 279, + 213 + ], + "type": "text", + "content": "VEnhancer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 58, + 217, + 138, + 262 + ], + "blocks": [ + { + "bbox": [ + 58, + 217, + 138, + 262 + ], + "lines": [ + { + "bbox": [ + 58, + 217, + 138, + 262 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 138, + 262 + ], + "type": "image", + "image_path": "7fbd124ab6af09870e0450fc95051178252c106723c79b6b0ce19f01e2853695.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 264, + 123, + 272 + ], + "lines": [ + { + "bbox": [ + 74, + 264, + 123, + 272 + ], + "spans": [ + { + "bbox": [ + 74, + 264, + 123, + 272 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 139, + 217, + 219, + 262 + ], + "blocks": [ + { + "bbox": [ + 139, + 217, + 219, + 262 + ], + "lines": [ + { + "bbox": [ + 139, + 217, + 219, + 262 + ], + "spans": [ + { + "bbox": [ + 139, + 217, + 219, + 262 + ], + "type": "image", + "image_path": "af4295e2bcd9827b1955e60443334d167e36615fee2c38ab5698b12b70747421.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 264, + 209, + 273 + ], + "lines": [ + { + "bbox": [ + 150, + 264, + 209, + 273 + ], + "spans": [ + { + "bbox": [ + 150, + 264, + 209, + 273 + ], + "type": "text", + "content": "Upscale-A-Video" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 220, + 217, + 301, + 262 + ], + "blocks": [ + { + "bbox": [ + 220, + 217, + 301, + 262 + ], + "lines": [ + { + "bbox": [ + 220, + 217, + 301, + 262 + ], + "spans": [ + { + "bbox": [ + 220, + 217, + 301, + 262 + ], + "type": "image", + "image_path": "2b6f68ca39a7a522a01831a947ed9aad379c1765b505f213c9f1984a0c71788e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 264, + 269, + 272 + ], + "lines": [ + { + "bbox": [ + 252, + 264, + 269, + 272 + ], + "spans": [ + { + "bbox": [ + 252, + 264, + 269, + 272 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 439, + 145 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 439, + 145 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 439, + 145 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 439, + 145 + ], + "type": "image", + "image_path": "2521f2f0580ae6c73908a06cf11dbde1149d0d95bbddabc91e1232759a155ee2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 146, + 400, + 156 + ], + "lines": [ + { + "bbox": [ + 347, + 146, + 400, + 156 + ], + "spans": [ + { + "bbox": [ + 347, + 146, + 400, + 156 + ], + "type": "text", + "content": "Video & prompt" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 471, + 99, + 553, + 145 + ], + "blocks": [ + { + "bbox": [ + 471, + 99, + 553, + 145 + ], + "lines": [ + { + "bbox": [ + 471, + 99, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 471, + 99, + 553, + 145 + ], + "type": "image", + "image_path": "948f1cbbaebca6221de677ac87e4f76343270e3c2c9bd01100748c2b08a8d279.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 146, + 517, + 154 + ], + "lines": [ + { + "bbox": [ + 507, + 146, + 517, + 154 + ], + "spans": [ + { + "bbox": [ + 507, + 146, + 517, + 154 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 309, + 157, + 389, + 204 + ], + "blocks": [ + { + "bbox": [ + 309, + 157, + 389, + 204 + ], + "lines": [ + { + "bbox": [ + 309, + 157, + 389, + 204 + ], + "spans": [ + { + "bbox": [ + 309, + 157, + 389, + 204 + ], + "type": "image", + "image_path": "eae1ad88aa5e70662ef692def10460a768bcc062a6fe019d85330ec21fb08583.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 205, + 358, + 213 + ], + "lines": [ + { + "bbox": [ + 340, + 205, + 358, + 213 + ], + "spans": [ + { + "bbox": [ + 340, + 205, + 358, + 213 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 391, + 157, + 471, + 204 + ], + "blocks": [ + { + "bbox": [ + 391, + 157, + 471, + 204 + ], + "lines": [ + { + "bbox": [ + 391, + 157, + 471, + 204 + ], + "spans": [ + { + "bbox": [ + 391, + 157, + 471, + 204 + ], + "type": "image", + "image_path": "0ad2c1c1ddcb8c0e5b27e6078c2731998b57e600f9373c1ce438d3b4be21f5b4.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 205, + 447, + 213 + ], + "lines": [ + { + "bbox": [ + 414, + 205, + 447, + 213 + ], + "spans": [ + { + "bbox": [ + 414, + 205, + 447, + 213 + ], + "type": "text", + "content": "LaVie-SR" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 472, + 157, + 552, + 204 + ], + "blocks": [ + { + "bbox": [ + 472, + 157, + 552, + 204 + ], + "lines": [ + { + "bbox": [ + 472, + 157, + 552, + 204 + ], + "spans": [ + { + "bbox": [ + 472, + 157, + 552, + 204 + ], + "type": "image", + "image_path": "dbff98e7545b6b10e99f21e6d2eddc771339dae2ca727d05f95a34a2b5545711.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 205, + 531, + 213 + ], + "lines": [ + { + "bbox": [ + 493, + 205, + 531, + 213 + ], + "spans": [ + { + "bbox": [ + 493, + 205, + 531, + 213 + ], + "type": "text", + "content": "VEnhancer" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 309, + 217, + 389, + 262 + ], + "blocks": [ + { + "bbox": [ + 309, + 217, + 389, + 262 + ], + "lines": [ + { + "bbox": [ + 309, + 217, + 389, + 262 + ], + "spans": [ + { + "bbox": [ + 309, + 217, + 389, + 262 + ], + "type": "image", + "image_path": "d3b18950f91bc25cf0fccef143ae28b91c938aa251353ae863d160c53929acf5.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 264, + 373, + 272 + ], + "lines": [ + { + "bbox": [ + 324, + 264, + 373, + 272 + ], + "spans": [ + { + "bbox": [ + 324, + 264, + 373, + 272 + ], + "type": "text", + "content": "RealBasicVSR" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 391, + 217, + 471, + 263 + ], + "blocks": [ + { + "bbox": [ + 391, + 217, + 471, + 263 + ], + "lines": [ + { + "bbox": [ + 391, + 217, + 471, + 263 + ], + "spans": [ + { + "bbox": [ + 391, + 217, + 471, + 263 + ], + "type": "image", + "image_path": "f85f09b0a3cb06cd5b871f369f2157d4abce0e2c60bcd299f61dbb6416293277.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 264, + 460, + 273 + ], + "lines": [ + { + "bbox": [ + 401, + 264, + 460, + 273 + ], + "spans": [ + { + "bbox": [ + 401, + 264, + 460, + 273 + ], + "type": "text", + "content": "Upscale-A-Video" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 277, + 527, + 289 + ], + "lines": [ + { + "bbox": [ + 82, + 277, + 527, + 289 + ], + "spans": [ + { + "bbox": [ + 82, + 277, + 527, + 289 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons on SynVideo30 (test). These videos " + }, + { + "bbox": [ + 82, + 277, + 527, + 289 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 82, + 277, + 527, + 289 + ], + "type": "text", + "content": " super-resolution results. Zoom in for best view." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 472, + 217, + 552, + 263 + ], + "blocks": [ + { + "bbox": [ + 472, + 217, + 552, + 263 + ], + "lines": [ + { + "bbox": [ + 472, + 217, + 552, + 263 + ], + "spans": [ + { + "bbox": [ + 472, + 217, + 552, + 263 + ], + "type": "image", + "image_path": "c9f434d578b5247fe5eb3b717b2bcbaf8329fbd7082a3df20a0db0863a7e5ecf.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 264, + 520, + 272 + ], + "lines": [ + { + "bbox": [ + 504, + 264, + 520, + 272 + ], + "spans": [ + { + "bbox": [ + 504, + 264, + 520, + 272 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 304, + 136, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 136, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 136, + 317 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 323, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 396 + ], + "type": "text", + "content": "To verify the superiority of PatchVSR, we compare it with SOTA methods, including VEnhancer [16], Upscale-A-Video [55], LaVie-SR [44] and RealBasicVSR [6]. We also make ablation study to analyze the effectiveness of our proposed designs. Besides, some extended analysis is conducted to demonstrate the capability of our method." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 403, + 188, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 188, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 188, + 415 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": "Training Datasets We train our model using 460K self-collected high-quality video-text pairs, with each resolution from " + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": " to 2K. The text prompts are all captioned using LAVA captioner [26], and encoded by T5 text encoder [38] with no more than 256 tokens." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 489, + 296, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 296, + 609 + ], + "type": "text", + "content": "Testing Datasets. We evaluate the VSR methods on two kinds of dataset, i.e. downscale synthetic videos and AIGC videos. For synthetic videos, since our employed base model is fixed to generate 77-frames video, we utilize a public dataset: REDS30 [33] for fair comparisons, which contains 100 frames. To evaluate higher resolution, we self-collect 30 high-quality 2K resolution videos, each containing more than 150 frames. Additionally, we collect and test on an AIGC dataset, which comprises 30 AI-generated videos from public text-to-video generative models [50]." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Training Details. Our model is trained on 16 NVIDIA H800-80G GPUs with a total batch size of 64. AdamW [29] is used as the optimizer with a learning rate of " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ". The text prompt is randomly replaced by a null prompt with " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " probability. To obtain the LR-HR training data pair, the HR video is first down-sampled using bilinear interpolation with down scale " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " and up-sampled back to the original resolution. To match the size of the pre-trained model," + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 305, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 555, + 473 + ], + "type": "text", + "content": "we randomly crop a " + }, + { + "bbox": [ + 313, + 305, + 555, + 473 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 305, + 555, + 473 + ], + "type": "text", + "content": " area from the up-sampled LR video with aspect ratios from 0.5 to 2.0. To enhance the robustness of our model to different degradation scenarios, we utilize the noise augmentation technique by injecting noise into the input latent using a diffuse process. The noise timestep is randomly selected from 200 to 300, aiming at preserving the structure of the input video. We have additionally encoded noise timestep, downscale factor, and crop location as conditions for the model. We have replaced the text prompt with a fixed prompt for the patch condition branch, with the base model and global context branch using text prompts. To provide initialization weight, we followed the structure of the pre-trained model and sampled the layer index from it at equal intervals." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 484, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 520 + ], + "type": "text", + "content": "Inference Details. We perform 50 PNDM [27] sampling steps with classifier-free guidance scale 5.0. We have also used timestep shift [12] with shift value 5.0." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 616 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the fidelity of our method on synthesized videos, we utilize PSNR, SSIM, and LPIPS [53]. For visual quality, we conduct our evaluation of commonly used quality assessment metrics MUSIQ [23] and DOVER [45]. Furthermore, we also include aesthetic score from a video generation benchmark VBench [20] for more comprehensive comparisons." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 624, + 492, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 492, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 492, + 636 + ], + "type": "text", + "content": "4.2. Comparison with SOTA methods" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "Qualitative Results. The visual results of synthetic and AIGC videos are presented in Fig. 3 and Fig. 4 respectively. It is observed that our method significantly outperforms existing CNN-based and diffusion-based models in both details generation and temporal consistency (Readers are recommended to check our provided video results). Specif" + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17803" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 555, + 284 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 284 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 284 + ], + "type": "image", + "image_path": "bc96866090653a2c3deb647fd4601848388ecac12aa9a33925484016c57aa71e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 287, + 535, + 299 + ], + "lines": [ + { + "bbox": [ + 75, + 287, + 535, + 299 + ], + "spans": [ + { + "bbox": [ + 75, + 287, + 535, + 299 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons on VideoGen30 (test). These videos are " + }, + { + "bbox": [ + 75, + 287, + 535, + 299 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 75, + 287, + 535, + 299 + ], + "type": "text", + "content": " super-resolution results. Zoom in for best view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 308, + 557, + 515 + ], + "blocks": [ + { + "bbox": [ + 56, + 308, + 557, + 515 + ], + "lines": [ + { + "bbox": [ + 56, + 308, + 557, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 557, + 515 + ], + "type": "table", + "html": "
DatasetsMetricsRealBasicVSR [6]LaVie-SR [44]Upscale-A-Video [55]VEnhancer [16]Ours
REDS30PSNR ↑29.69828.73531.25728.54430.111
SSIM ↑0.6560.6220.6610.6110.647
LPIPS ↓0.3340.4070.3280.4210.326
DOVER ↑0.3050.3180.3100.3440.323
MUSIQ ↑40.58743.57742.57544.53344.956
Aesthetics ↑0.4220.4860.4710.4820.493
SynVideo30PSNR ↑33.50731.43333.43228.85630.857
SSIM ↑0.7760.7160.7280.6970.732
LPIPS ↓0.1850.1940.2050.1990.183
DOVER ↑0.5840.6120.5290.5500.597
MUSIQ ↑49.55748.14649.83943.53850.695
Aesthetics ↑0.4960.5090.4940.5030.520
VideoGen30DOVER ↑0.5880.6090.5100.5120.590
MUSIQ ↑47.56448.32946.89946.93050.559
Aesthetics ↑0.5470.6000.5740.5510.602
", + "image_path": "307b9903a91931a39689d91719f7c38ff69adf4d204140d3c0397d62b6036f45.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 518, + 555, + 541 + ], + "lines": [ + { + "bbox": [ + 55, + 518, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 555, + 541 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on different VSR benchmarks from synthetic (REDS30, SynVideo30) and AIGC (VideoGen30) data. Red and blue numbers indicate the best and second best results respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 552, + 297, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 552, + 297, + 708 + ], + "spans": [ + { + "bbox": [ + 54, + 552, + 297, + 708 + ], + "type": "text", + "content": "ically, our method not only successfully recovers the details of the red silk scarf in Fig. 3, but also adds more details to the \"Angkor Wat\" as shown in Fig. 4, which others tend to generate blurry or unnatural results. In Fig. 5, we also illustrate full VSR results achieved through the multi-patch joint modulation scheme. It can be seen that creating auxiliary patches greatly alleviates the seam problem and significantly reduces the color difference between different patches, while further improving the contrast of the results. To show the advantage of our multi-patch joint modulation, we compare with a baseline method: stitching the patches in latent space and decoding the big latent together, which suffers obvious seamline artifacts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "text", + "content": "Quantitative Results. The quantitative comparison is shown in Tab. 1. It can be seen that the proposed method achieves the highest LPIPS score in both REDS30 and SynVideo30, indicating the high fidelity of our method for synthetic videos. However, our method show inferior performance in PSNR and SSIM. This can be explained by the fact that our method tends to generate rich and vivid details, which are not favorable to metrics that capture pixel-wise deviations (i.e. PSRN and SSIM). Instead, as a perceptual-level fidelity measure, LPIPS might be more informative to reflect the content fidelity for generative tasks. We evaluate the visual quality of generated high-resolution videos through comparisons with competing methods on MUSIQ" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17804" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 293, + 367 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 293, + 367 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 293, + 367 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 293, + 367 + ], + "type": "image", + "image_path": "5cd909b4dea0786c2a3bf31b0afc3166e930ba7fac770e6279b9dec80f8f5416.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 372, + 295, + 429 + ], + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 429 + ], + "type": "text", + "content": "Figure 5. Visual evaluation on multi-patch joint modulation. Here the super-resolved videos of 2K resolution are used for comparison, where " + }, + { + "bbox": [ + 55, + 372, + 295, + 429 + ], + "type": "inline_equation", + "content": "3 \\times 3 = 9" + }, + { + "bbox": [ + 55, + 372, + 295, + 429 + ], + "type": "text", + "content": " patches are involved. The top, middle, bottom denote input video, result of stitching latent of patches, and result of ours respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 438, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 548 + ], + "type": "text", + "content": "and DOVER metrics, demonstrating our approach's superiority. This performance reflects contributions from both the base model's capabilities and the effectiveness of our method in leveraging full-frame generation for patch-level detail synthesis. Interestingly, our approach even preserves subjective properties like aesthetic quality from the base model, as evidenced by the aesthetic scores. This highlights our framework's ability to inherit and maintain critical perceptual characteristics during the enhancement process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 554, + 149, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 149, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 149, + 567 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 571, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 620 + ], + "type": "text", + "content": "We conduct ablation studies on key designs. Considering the limited space, we provide a quantitative evaluation here and the typical visual results are compared in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Global Branch. To validate the effectiveness of the global branch, We build a baseline by removing the global branch from our method. As shown in Tab. 2, removing the global branch results in significant performance deterioration, leading to worse DOVER, MUSIQ, and aesthetic score. Besides, to check the necessity of using the patch location embeddings, we build a baseline by removing the" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 155 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 155 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 155 + ], + "type": "table", + "html": "
ComponentDOVER ↑MUSIQ ↑Aesthetics ↑
w/o global branch0.50246.0740.589
w/o LoRA0.58250.4960.600
w/o location embed0.57450.0840.601
w/o fixed prompt0.56248.1330.597
Ours0.59050.5590.602
", + "image_path": "9b2166f0e1976fcb023f6b3ff311d19b6ac7e767f03b8f8543a900d8a4e162a5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 158, + 555, + 180 + ], + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 180 + ], + "type": "text", + "content": "Table 2. Ablation study of different model components on VideoGen30 test dataset. The **bold** numbers indicate the best." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "type": "text", + "content": "patch position information from the global branch input. As shown in Tab. 2, after incorporating the location embedding, we obtain an increase of DOVER from 0.574 to 0.590, which demonstrates the effectiveness of utilizing the location mask as the patch position information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "text", + "content": "Text Prompts. Text prompts are used in the base model and the conditional global branch. As shown in Tab. 2, we observe that replacing the text prompt with a fixed prompt leads to superior results. We conjecture that the use of text prompts in patch condition branch brings about a mismatch between the video patch and the global semantic information, which in turn destroys the dynamic visual prior in the pre-trained model aligned to the text during training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 354, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 437 + ], + "type": "text", + "content": "Base model finetuning. Since the data distribution of local patches deviates from the full view of the range of common videos, we propose to fine-tune the base model instead of freezing it. It can be seen in Tab. 2 that adding LoRA results in improved performance across all metrics, which demonstrates the effectiveness of utilizing LoRA to adapt to the distribution of local video patches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 444, + 450, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 450, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 450, + 456 + ], + "type": "text", + "content": "4.4. Analysis and Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 498 + ], + "type": "text", + "content": "We would like to further study some characteristics of our proposed method, including detail generation magnitude, robustness to patch semantics, and computation efficiency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 503, + 556, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 556, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 556, + 647 + ], + "type": "text", + "content": "Detail generation magnitude. The input and output sizes of our VSR model are the same and our model can synthesize different levels of details. In our method, a scale factor is used to guide the model to generate reasonable details that should be presented on the input video if it is captured normally by the camera at that resolution/size. So, the scale factor actually plays the role of affecting the magnitude of details. Here we illustrate the effects of using different scale factors as conditions. In Fig. 6, increasing the scale factor progressively adds more high-resolution details to the results, which demonstrates the flexibility of our method to generate results with different detail magnitudes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Robustness to Patch Semantics. As stated above, our approach will partition the input video into patches in a content-agnostic manner. So, some patches can contain incomplete or even unclear semantics, which are more challenging for the model with the generative reasonable details." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17805" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 175, + 138 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 175, + 138 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 175, + 138 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 175, + 138 + ], + "type": "image", + "image_path": "7da4b1b25bf002ffdc5a21b500eb30f3e4a0d2f77858e7a79e2fa0a14dc5122a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 139, + 135, + 148 + ], + "lines": [ + { + "bbox": [ + 97, + 139, + 135, + 148 + ], + "spans": [ + { + "bbox": [ + 97, + 139, + 135, + 148 + ], + "type": "text", + "content": "Input patch" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 72, + 294, + 138 + ], + "blocks": [ + { + "bbox": [ + 177, + 72, + 294, + 138 + ], + "lines": [ + { + "bbox": [ + 177, + 72, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 177, + 72, + 294, + 138 + ], + "type": "image", + "image_path": "ac762976b9219bbc2a58ad7f99c69cec6fc6148e28b3723e9a7397d266886087.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 139, + 264, + 148 + ], + "lines": [ + { + "bbox": [ + 206, + 139, + 264, + 148 + ], + "spans": [ + { + "bbox": [ + 206, + 139, + 264, + 148 + ], + "type": "text", + "content": "Upscale factor " + }, + { + "bbox": [ + 206, + 139, + 264, + 148 + ], + "type": "inline_equation", + "content": "= 1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 150, + 175, + 218 + ], + "blocks": [ + { + "bbox": [ + 58, + 150, + 175, + 218 + ], + "lines": [ + { + "bbox": [ + 58, + 150, + 175, + 218 + ], + "spans": [ + { + "bbox": [ + 58, + 150, + 175, + 218 + ], + "type": "image", + "image_path": "975dd54e4f3806b1b1c33c798ed1ed7d4bd738fc82167a6398e3ce80381faa0c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 219, + 144, + 228 + ], + "lines": [ + { + "bbox": [ + 88, + 219, + 144, + 228 + ], + "spans": [ + { + "bbox": [ + 88, + 219, + 144, + 228 + ], + "type": "text", + "content": "Upscale factor=2" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 177, + 150, + 294, + 219 + ], + "blocks": [ + { + "bbox": [ + 177, + 150, + 294, + 219 + ], + "lines": [ + { + "bbox": [ + 177, + 150, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 177, + 150, + 294, + 219 + ], + "type": "image", + "image_path": "b7f87ea1be7a60375e5304f789d326828c09495ebd1fe020186351bdfe992437.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 220, + 264, + 228 + ], + "lines": [ + { + "bbox": [ + 206, + 220, + 264, + 228 + ], + "spans": [ + { + "bbox": [ + 206, + 220, + 264, + 228 + ], + "type": "text", + "content": "Upscale factor " + }, + { + "bbox": [ + 206, + 220, + 264, + 228 + ], + "type": "inline_equation", + "content": "= 3" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 231, + 175, + 299 + ], + "blocks": [ + { + "bbox": [ + 58, + 231, + 175, + 299 + ], + "lines": [ + { + "bbox": [ + 58, + 231, + 175, + 299 + ], + "spans": [ + { + "bbox": [ + 58, + 231, + 175, + 299 + ], + "type": "image", + "image_path": "07fe9ccade066828e437fea0e2e95984c92167cede84fbf01c3bf4a5dac82076.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 300, + 145, + 308 + ], + "lines": [ + { + "bbox": [ + 87, + 300, + 145, + 308 + ], + "spans": [ + { + "bbox": [ + 87, + 300, + 145, + 308 + ], + "type": "text", + "content": "Upscale factor " + }, + { + "bbox": [ + 87, + 300, + 145, + 308 + ], + "type": "inline_equation", + "content": "= 4" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 313, + 295, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 346 + ], + "type": "text", + "content": "Figure 6. Comparison of different up-sampling scales. Note that the upscale factor is uniformly randomly sampled within the range of 1 to 4 during the training phase." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 177, + 231, + 294, + 299 + ], + "blocks": [ + { + "bbox": [ + 177, + 231, + 294, + 299 + ], + "lines": [ + { + "bbox": [ + 177, + 231, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 177, + 231, + 294, + 299 + ], + "type": "image", + "image_path": "f4e91ad53575c5b33980bea0bff110d272880726f10c102731aa4c9f9b9987c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 300, + 264, + 308 + ], + "lines": [ + { + "bbox": [ + 206, + 300, + 264, + 308 + ], + "spans": [ + { + "bbox": [ + 206, + 300, + 264, + 308 + ], + "type": "text", + "content": "Upscale factor " + }, + { + "bbox": [ + 206, + 300, + 264, + 308 + ], + "type": "inline_equation", + "content": "= 5" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 72, + 352, + 279, + 426 + ], + "blocks": [ + { + "bbox": [ + 72, + 352, + 279, + 426 + ], + "lines": [ + { + "bbox": [ + 72, + 352, + 279, + 426 + ], + "spans": [ + { + "bbox": [ + 72, + 352, + 279, + 426 + ], + "type": "table", + "html": "
MethodsTime (s) ↓Memory (G) ↓
LaVie-SR226168
Upscale-A-Video274347
VEnhancer156262
Ours68040
", + "image_path": "2c907d623cb3787ac94d812605ddb156c76074fc286c0e14c2f32f229a7e8707.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 70, + 429, + 279, + 440 + ], + "lines": [ + { + "bbox": [ + 70, + 429, + 279, + 440 + ], + "spans": [ + { + "bbox": [ + 70, + 429, + 279, + 440 + ], + "type": "text", + "content": "Table 3. Computation efficiency for 2K video generation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 456, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 456, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 456, + 295, + 563 + ], + "type": "text", + "content": "It is interesting to study whether the performance varies across patches with diverse semantic quality. It can be seen in Fig. 7 that our model can successfully handle the video patches with incomplete semantics. As the global semantic tokens are extracted and input to the base model using global cross-attention modules, the model learns to utilize global semantic information to complete the semantics in the local video patch, which improves the robustness of our model against incomplete semantics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Computation efficiency. Our method is more efficient than existing pre-trained T2V-based VSR methods because the self-attention is conducted within patches. Here presents a simple deduce of the computation complexity against patch size: given a video of " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " tokens and each patch of " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " tokens, the patch number is " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "k = \\frac{n}{m}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " and the inter-token computation times are " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "k * \\left( \\frac{n}{k} \\right)^2 = n * m" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ". Since the base model we adopted uses fixed " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " resolution, our patch size follows this size in our experiments. The inference time and memory usage of generating a 2K resolution video is illustrated in Tab. 3. It can be seen that, after dividing the input video into patches, the computation burden is signif" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 553, + 129 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 129 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 129 + ], + "type": "image", + "image_path": "3d82803f04939eff4cb0b2aa507b73fe7c770ccdf0efaf0960017e6600ecdcbb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 316, + 129, + 553, + 185 + ], + "blocks": [ + { + "bbox": [ + 316, + 129, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 185 + ], + "type": "image", + "image_path": "c64a63248aeb32528b40723527be436a4915b6de7f92ba4edcf20ba1dc4fb7dd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 316, + 186, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 316, + 186, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 243 + ], + "type": "image", + "image_path": "6b8930ed5dc5789277d744bc994c28df541a6e9d6a948e6627a12876fbd6ddff.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 245, + 425, + 256 + ], + "lines": [ + { + "bbox": [ + 343, + 245, + 425, + 256 + ], + "spans": [ + { + "bbox": [ + 343, + 245, + 425, + 256 + ], + "type": "text", + "content": "Incomplete semantics" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 245, + 535, + 256 + ], + "lines": [ + { + "bbox": [ + 459, + 245, + 535, + 256 + ], + "spans": [ + { + "bbox": [ + 459, + 245, + 535, + 256 + ], + "type": "text", + "content": "Complete semantics" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 261, + 544, + 272 + ], + "lines": [ + { + "bbox": [ + 323, + 261, + 544, + 272 + ], + "spans": [ + { + "bbox": [ + 323, + 261, + 544, + 272 + ], + "type": "text", + "content": "Figure 7. Robustness against patch semantic incompleteness." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 285, + 553, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 553, + 344 + ], + "type": "text", + "content": "icantly reduced, further promoting the applications of generative VSR models. Thanks to the computation efficiency and the inference flexibility, our method even support 4K video generation, which are shown in the supplementary materials." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 360, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 526 + ], + "type": "text", + "content": "**Limitation.** Our method has limitations in two aspects. Firstly, it experiences noticeable performance degradation when applied to real-world low-resolution images, due to the lack of specialized data augmentation in our training pairs. Since the primary focus of this paper is to address the resolution limitations of using pre-trained base models for VSR, extending our approach to handle real-world noise remains an interesting direction for future work. Secondly, our approach inherits an iterative inference scheme from the base model, which introduces efficiency issues for practical applications. For VSR, the low-resolution input video provides significant opportunities for algorithm optimization to reduce the number of inference steps, which will be another area of future research." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we present the first exploration of utilizing a pre-trained T2V base model for patch-level video superresolution (VSR). To accomplish this, we propose an effective dual-branch adapter consisting of a patch condition branch and a global context branch. The global context branch is crucial for bridging the gap between full-video generative capabilities and patch-level detail synthesis. Additionally, our proposed multi-patch joint modulation scheme achieves consistent results across patches. Thanks to the flexibility of our PatchVSR approach, it outperforms existing VSR methods in generating faithful details while maintaining high computational efficiency. We hope this work will inspire further research in this area." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17806" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China (NSFC) under Grant No.62406167, U24B6012." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 160, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 160, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 296, + 204 + ], + "type": "text", + "content": "[1] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning (ICML), 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 206, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 206, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 206, + 296, + 261 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 295, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 295 + ], + "type": "text", + "content": "[3] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 297, + 295, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 295, + 352 + ], + "type": "text", + "content": "[4] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 354, + 295, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 295, + 408 + ], + "type": "text", + "content": "[5] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 411, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 295, + 464 + ], + "type": "text", + "content": "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5962-5971, 2022. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 467, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 295, + 521 + ], + "type": "text", + "content": "[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 523, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 295, + 589 + ], + "type": "text", + "content": "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 591, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 295, + 667 + ], + "type": "text", + "content": "[9] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[10] Shian Du, Xiaotian Cheng, Qi Qian, Henglu Wei, Yi Xu, and Xiangyang Ji. Efficient personalized text-to-image generation by leveraging textual subspace. arXiv preprint arXiv:2407.00608, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "type": "text", + "content": "[11] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "text", + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "text", + "content": "[13] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 275, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 330 + ], + "type": "text", + "content": "[14] Zhicheng Geng, Luming Liang, Tianyu Ding, and Ilya Zharkov. Rstt: Real-time spatial temporal transformer for space-time video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17441-17451, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 332, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 396 + ], + "type": "text", + "content": "[15] Lanqing Guo, Yingqing He, Haoxin Chen, Menghan Xia, Xiaodong Cun, Yufei Wang, Siyu Huang, Yong Zhang, Xintao Wang, Qifeng Chen, Ying Shan, and Bihan Wen. Make a cheap scaling: A self-cascade diffusion model for higher-resolution adaptation. In European Conference on Computer Vision (ECCV), 2024. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[16] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation. arXiv preprint arXiv:2407.07667, 2024. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 455, + 553, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 519 + ], + "type": "text", + "content": "[17] Yingqing He, Shaoshu Yang, Haoxin Chen, Xiaodong Cun, Menghan Xia, Yong Zhang, Xintao Wang, Ran He, Qifeng Chen, and Ying Shan. Scalecrafter: Tuning-free higher-resolution visual generation with diffusion models. In International Conference on Learning Representations (ICLR), 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "type": "text", + "content": "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 567, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 621 + ], + "type": "text", + "content": "[19] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 624, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 689 + ], + "type": "text", + "content": "[20] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[21] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17807" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "text", + "content": "structure-detail network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 645-660. Springer, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "text", + "content": "[22] Mehran Jeelani, Noshaba Cheema, Klaus Illgner-Fehns, Philipp Slusallek, Sunil Jaiswal, et al. Expanding synthetic real-world degradations for blind video super resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1199-1208, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "text", + "content": "[23] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations (ICLR), 2014. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "type": "text", + "content": "[25] Chengxu Liu, Huan Yang, Jianlong Fu, and Xueming Qian. Learning trajectory-aware transformer for video superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5687-5696, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 338 + ], + "type": "text", + "content": "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "text", + "content": "[27] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "text", + "content": "[28] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 448 + ], + "type": "text", + "content": "[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "text", + "content": "[30] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 592 + ], + "type": "text", + "content": "[32] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 593, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 658 + ], + "type": "text", + "content": "[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[34] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[36] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 161, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 553, + 216 + ], + "type": "text", + "content": "[37] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[38] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 555, + 327 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "text", + "content": "[40] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. Advances in Neural Information Processing Systems, 35:36081-36093, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "text", + "content": "[41] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 404, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 555, + 448 + ], + "type": "text", + "content": "[42] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C. K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 2024. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 555, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 503 + ], + "type": "text", + "content": "[43] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "text", + "content": "[44] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "text", + "content": "[45] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20144-20154, 2023. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 690 + ], + "type": "text", + "content": "[46] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[47] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17808" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 541 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "type": "text", + "content": "Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[48] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4781-4790, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[49] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 294, + 262 + ], + "type": "text", + "content": "[50] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[51] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1-15, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "text", + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 418 + ], + "type": "text", + "content": "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "type": "text", + "content": "[54] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 541 + ], + "type": "text", + "content": "[55] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2545, 2024. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17809" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/f625c957-1581-449f-8cd5-463f4552272c_content_list.json b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/f625c957-1581-449f-8cd5-463f4552272c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8fba5316e0f41c62249b95a01f1e90c847a70091 --- /dev/null +++ b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/f625c957-1581-449f-8cd5-463f4552272c_content_list.json @@ -0,0 +1,1528 @@ +[ + { + "type": "text", + "text": "Pathways on the Image Manifold: Image Editing via Video Generation", + "text_level": 1, + "bbox": [ + 142, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Noam Rotstein", + "bbox": [ + 300, + 181, + 424, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Roy Velich", + "bbox": [ + 300, + 199, + 392, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gal Yona", + "bbox": [ + 464, + 181, + 542, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "David Bensaïd", + "bbox": [ + 431, + 199, + 550, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daniel Silver", + "bbox": [ + 581, + 181, + 691, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ron Kimmel", + "bbox": [ + 589, + 199, + 696, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Technion - Israel Institute of Technology", + "text_level": 1, + "bbox": [ + 333, + 233, + 658, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in image editing, driven by image diffusion models, have shown remarkable progress. However, significant challenges remain, as these models often struggle to follow complex edit instructions accurately and frequently compromise fidelity by altering key elements of the original image. Simultaneously, video generation has made remarkable strides, with models that effectively function as consistent and continuous world simulators. In this paper, we propose merging these two fields by utilizing image-to-video models for image editing. We reformulate image editing as a temporal process, using pretrained video models to create smooth transitions from the original image to the desired edit. This approach traverses the image manifold continuously, ensuring consistent edits while preserving the original image's key aspects. Our approach achieves state-of-the-art results on text-based image editing, demonstrating significant improvements in both edit accuracy and image preservation. Visit our project page.", + "bbox": [ + 89, + 316, + 483, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 603, + 220, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image editing has witnessed remarkable advancements through deep learning and text-guided diffusion networks. These developments have set a new benchmark for image manipulations, enhancing both control and quality. However, current approaches continue to face significant limitations in real-world scenarios. These methods often struggle with two key challenges: achieving precise edits that accurately reflect the intended modifications, and preserving the essential characteristics of the original image content.", + "bbox": [ + 89, + 628, + 482, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State-of-the-art techniques predominantly rely on text-guided diffusion models, which iteratively denoise random latent representations to generate edited images. Such methods condition the generation process on both the source image—using techniques such as latent inversion [32] or model fine-tuning [22]—and the target edit description. However, these approaches require the model to generate a single output image that preserves source image fidelity while implementing complex edits, often compromises", + "bbox": [ + 89, + 765, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d302793114c47e6db147aefc107c0e6e79428b955fa321790717cf9f8d7d48b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 285, + 617, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A photo of a tomato in a blue tennis court.\"", + "bbox": [ + 630, + 310, + 785, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/90c39c6cbe426a659ad8f14ed3c0a80814b528e1aec259b08b4e98dea56532a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 285, + 897, + 362 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b52cef901214128c3507daf9bff0aacf738c891e3d154c13f8f4ff57e74aa8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 402, + 619, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A vase of flowers filled with green water.\"", + "bbox": [ + 635, + 424, + 777, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e77637dc6d543fc646c2569bb9ab32c32478ad097b80948df973b214ee9b5cb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 404, + 898, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a28a86e44233a98e5a4d79f19da755da1d7fa9430d912cb8f87f5c973741b207.jpg", + "image_caption": [ + "Figure 1. Visualization of Frame2Frame's editing process. Temporal progression of our video-based approach. Starting from the source image (leftmost), frames illustrate the natural evolution toward the target edit (rightmost). Our method produces temporally coherent intermediate states while preserving fidelity to both the source content and the editing intent." + ], + "image_footnote": [], + "bbox": [ + 517, + 522, + 617, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A photo of a teddy bear doing pushups.\"", + "bbox": [ + 630, + 547, + 785, + 573 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8223c5762538660c06b6eeb47871af82f55253f16cf4607417c132e4005996e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 522, + 897, + 598 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80f547005fe8c15b9d426009781fb698c9b8f6c5bad44ec593b4dfe00f2bec40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 598, + 875, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing edit accuracy and content preservation.", + "bbox": [ + 511, + 763, + 797, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a paradigm shift in image editing by reformulating it as a video generation task. Rather than a single-state transition, our approach harnesses temporal coherence: the source image serves as the initial frame of a video that progressively and naturally transforms toward the target edit. This temporal evolution allows the editing process to unfold through physically plausible intermediate states, providing a continuous path between source", + "bbox": [ + 509, + 780, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7857", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and target images, as illustrated in Figure 1 and Figure S9 in the appendix. This temporal approach leverages the sophisticated world understanding embedded in recent video generation models, which have achieved breakthrough results in temporal coherence and visual quality through training on large-scale internet data [2, 48]. From a geometrical perspective, conventional editing approaches project initial noise onto the natural image manifold, targeting a single point where the image aligns with both the source and the edit request. In contrast, our approach generates a continuous path along the manifold between the original and edited image, producing a smooth realistic transition across different image states, as thoroughly discussed in Section 4.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We implement the proposed approach through a structured pipeline called Frame2Frame (F2F). First, we transform the edit instruction into a Temporal Editing Caption — a scenario describing how the edit should naturally evolve over time—using a pretrained Vision-Language Model (VLM). Next, a state-of-the-art image-to-video model generates a temporally coherent sequence guided by the temporal caption. Finally, we identify the frame that best realizes the desired edit with the assistance of a VLM. Extensive experiments demonstrate improvements over existing imaged-to-image approaches. We evaluate on TEdBench [22] and PosEdit, a newly curated dataset derived from UTD-MHAD [6], which focuses on human pose transformations. PosEdit pairs source images with ground-truth targets of the same subject in different poses, enabling rigorous evaluation of both edit accuracy and source fidelity. Beyond commonly defined editing tasks, our framework shows promising results in more classical computer vision problems such as de-blurring, de-noising, and relighting by recasting them as temporal progressions, suggesting broader applications for video-based image transformations.", + "bbox": [ + 91, + 289, + 483, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions include:", + "text_level": 1, + "bbox": [ + 109, + 607, + 339, + 619 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Reformulating image editing as a generative video task—leveraging temporal coherence to create edit paths on the natural image manifold, enabling high-fidelity manipulations while preserving source characteristics.", + "2. Frame2Frame: an end-to-end framework that realizes the reformulation through three key components: (1) temporal editing captions, (2) generated video-based editing, and (3) automated frame selection.", + "3. Comprehensive evaluation showing state-of-the-art performance on TEdBench and PosEdit, a new dataset for evaluating human pose edits." + ], + "bbox": [ + 91, + 623, + 482, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Efforts", + "text_level": 1, + "bbox": [ + 91, + 805, + 243, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Editing", + "text_level": 1, + "bbox": [ + 91, + 830, + 236, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-based image editing has advanced significantly with the success of generative diffusion models [19]. These models, in their text-to-image version, generate images through", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a denoising diffusion process conditioned on input text, relying on ground truth text-image pairs for training [37, 39]. In contrast, image editing often lacks predefined ground truth data for source and target images, posing unique challenges. This limitation has led researchers to explore diverse editing methodologies [20]. For example, SDEdit [30] injects noise into an image and then denoises it based on an editing target prompt. Imagic [22] fine-tunes a text-to-image model on a single image, subsequently interpolating between input and target text embeddings to produce edits. Other methods first invert the input image into a diffusion model's latent space [21, 32] and then generate the edited image from that latent representation using various techniques for structure preservation and manipulation [18, 34]. InstructPix2Pix [5] synthesizes an editing dataset using the approach in [18], filters it, and employs this dataset to train a diffusion model in a supervised fashion. Paint-by-Inpaint [47] further explores this supervised approach, generating a real-image dataset for object insertion.", + "bbox": [ + 511, + 90, + 906, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Generative Video Models", + "text_level": 1, + "bbox": [ + 513, + 391, + 743, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent years have seen remarkable advancements in video generation, evolving from domain-specific systems [15] and brief clips to models capable of generating diverse, high-fidelity content. This progress has stemmed from paradigm shifts and the large-scale expansion of datasets [7] and architectures. Approaches have evolved from recurrent networks [16, 42] and generative adversarial networks (GANs) [9, 43, 44] to latent diffusion models (LDMs) [2, 3, 11], which leverage large U-Net or transformer-based architectures alongside vast internet-sourced datasets. Notable recent efforts include Stable Video Diffusion [2], which trains an LDM on large curated datasets, and OpenAI's Sora [28], which achieves impressive results by integrating large architectures with extensive public and private datasets. These models, often termed \"world simulators\" due to their emergent understanding of physical dynamics and temporal coherence. Our work builds upon CogVideoX [48], a transformer-based latent diffusion model that employs a 3D Variational Autoencoder to compress videos across both spatial and temporal dimensions, enhancing coherence. To improve text-video alignment, CogVideoX also integrates an expert transformer with adaptive LayerNorm, enabling deep fusion of visual and textual modalities.", + "bbox": [ + 511, + 416, + 908, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Close to our work, several recent efforts have utilized the world simulation capabilities of video diffusion models for various computer vision tasks. Make-A-Video3D [41] temporally extends static NeRFs using Score Distillation Sampling from video models. ViVid-1-to-3 [25] generates images along a camera trajectory around objects to enable novel view synthesis. PhysDreamer [51] models rigid object properties through 3D Gaussians and material fields, trained via distillation from pretrained video generators.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7858", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ef0d71e99f13875a32f65951debb3c6a04bcf59a2fdcab9d4faab0bd663eb636.jpg", + "image_caption": [ + "Figure 2. Editing Manifold Pathway. Given an input image and target caption \"A happy person making a heart shape with their hands\", our method generates a continuous path on the natural image manifold. Each generated frame (indicated by black arrows) represents a plausible intermediate state between the source and target, maintaining temporal consistency throughout the transformation. As a result, in contrast to the competing approach, F2F achieves the desired edit while preserving the \"AI\" text on the person's shirt." + ], + "image_footnote": [], + "bbox": [ + 207, + 92, + 794, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Image Editing and Video", + "text_level": 1, + "bbox": [ + 89, + 464, + 318, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The intersection of image editing and video has received limited attention. Existing methods focus on sampling pairs of random frames from videos to build image pair datasets that capture the same subject under varying conditions. For example, AnyDoor [8] uses the paired frames as an augmentation method, segmenting foreground objects in each frame and assigning one masked object as the target edited appearance of the subject. MagicFixup [1] employs these frames to build a dataset focused on refining user-made subject coarse 2D edits. Recently, drag-based editing approaches [29, 40], used frames extracted from video and computed optical flow to collect a dataset aimed at editing by spatially dragging points within the image. In contrast to these approaches, which rely on frame sampling for image data collection, we are the first to directly perform image editing using generative video models.", + "bbox": [ + 89, + 483, + 483, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Frame2Frame", + "text_level": 1, + "bbox": [ + 89, + 734, + 236, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present Frame2Frame, a framework that reformulates image editing as a temporal transformation process. Our approach leverages video generation models to create natural transitions between source and target images, achieving consistent and realistic edits. The proposed method has three main steps, as illustrated in Figure 3.", + "bbox": [ + 89, + 753, + 483, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Temporal Editing Captions", + "text_level": 1, + "bbox": [ + 89, + 849, + 336, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-based image editing methods typically operate on two inputs: a source image $I_{s}$ and a target caption $c$ , where $c$", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "specifies the desired modifications to $I_{s}$ . Our approach differs fundamentally by modeling editing as a temporal process. This requires a novel type of prompt—the Temporal Editing Caption, denoted by $\\tilde{c}$ —that describes the sequential transformation from source to target image. We construct $\\tilde{c}$ by combining information from $I_{s}$ and $c$ to create a description of how the desired edit unfolds over time.", + "bbox": [ + 511, + 464, + 906, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To automate this process, we leverage recent advances in Vision-Language Models (VLMs) [10, 14, 38, 45, 46]. The VLM, specifically ChatGPT-4o [33], is instructed to produce a concise video scenario that highlights how elements within the image change or move over time. The generated caption captures the essential transformations while maintaining a static camera perspective unless movement is necessary. To improve generation quality, we utilize in-context learning (ICL) [12], providing the VLM with nine exemplar prompt-caption pairs. The complete prompt template, ICL examples, and an ablation study comparing this approach with directly using the target captions are included in Section B of the Appendix.", + "bbox": [ + 511, + 571, + 908, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Video Generation", + "text_level": 1, + "bbox": [ + 511, + 784, + 687, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ CogVideoX (I2V-5B) [48], a pretrained generative video latent diffusion model utilizing a transformer-based architecture. Specifically, we use its image-to-video variant, which has been fine-tuned to generate videos starting from an input image $I_{s}$ . During generation, $I_{s}$ is encoded and concatenated with noise in latent space, where", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7859", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the model applies a denoising process guided by the temporal caption $\\tilde{c}$ . As elaborated in Section 4, this conditioning allows generated videos to start from $I_{s}$ and evolve naturally along the image manifold, maintaining temporal coherence and consistency. Additionally, the model's transformer architecture enables effective fusion of visual and textual information, allowing precise control over the editing process through our temporal captions. Formally, given the video generator $G$ , we define the generation process as:", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} (I _ {s}, \\tilde {c}) = V = \\{f _ {1}, \\dots , f _ {T} \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 239, + 370, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V$ denotes the generated video with $T$ frames, and $f_{t}$ represents the frame at timestep $t$ .", + "bbox": [ + 89, + 271, + 483, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Frame Selection", + "text_level": 1, + "bbox": [ + 89, + 313, + 251, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observed that the optimal number of frames required for an edit can vary—small changes may be completed in fewer frames, while more extensive transformations often necessitate additional ones. Additionally, later frames tend to deviate further from the source image. Thus, even though $V$ serves as an editing path originating from $I_{s}$ , there is no guarantee that $f_{T}$ is the optimal edited image in $V$ . For instance, in Figure 3, the edited image of the cat might be expected to capture the midpoint of its jump, but the last frame shows the jump already completed. Therefore, we aim to identify the optimal edited frame, denoted $f_{t^{*}}$ , which corresponds to the earliest timestep $t$ that achieves the desired edit. The transition from the initial frame $f_{1}$ (or $I_{s}$ ) to the final edited frame $f_{t^{*}}$ motivates our method's name, Frame2Frame.", + "bbox": [ + 89, + 338, + 483, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To automate the selection of $t^*$ and avoid manual frame-by-frame review, we employ an automated approach. After generating the sequence $V$ , we sample every fourth frame, imprinting each with a unique identifier and assembling them into an image collage alongside $I_s$ . Inspired by [23], which introduces a novel approach to video comprehension by transforming videos into image grids, we use a VLM, specifically GPT-4o, to assist in selecting $t^*$ by providing it with the collage and the editing prompt $c$ . The VLM is tasked with identifying the frame that best fulfills the editing intent, evaluating each frame's alignment with $c$ and fidelity to $I_s$ . The model is instructed to select the optimal frame with the lowest index that completes the edit. An ablation study, detailed in Section C of the appendix, evaluates the effectiveness of our automated approach against the baseline of selecting the final frame of the video.", + "bbox": [ + 89, + 566, + 483, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that while we refer to an optimal $t^*$ , the definition of the required edit can be subjective and dependent on user preferences. For example, in Figure 3, different users may select frames of the cat jumping based on variations in its altitude. Thus, frame selection could serve as a flexible and customizable advantage in certain scenarios.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/779a5ec34b73187ef6abba799b1ec668c0bc9e424702f0ee9cb80ebcaf605dd7.jpg", + "image_caption": [ + "Figure 3. Frame2Frame Overview. Given a source image and editing prompt, our pipeline proceeds in three steps. First, a Vision-Language Model generates a temporal caption describing the transformation. Next, this caption guides a video generator to create a natural progression of the edit. Finally, our frame selection strategy identifies the optimal frame that best realizes the desired edit, producing the final image of the cat mid-leap." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 906, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Editing Manifold Pathway", + "text_level": 1, + "bbox": [ + 511, + 343, + 759, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To illustrate the advantages of our method over conventional image-to-image approaches, we aim to visualize the editing process within the natural image manifold, where realistic images reside. Given the high dimensionality of this space, we project it into a lower-dimensional representation suitable for interpretation. To achieve this, we first generate three sets of images, each containing 200 samples, using the text-to-image generator FLUX.1-dev [26]. The sets are based on the following prompts:", + "bbox": [ + 511, + 369, + 906, + 505 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. AI: \"Full-body portrait of a happy person wearing a shirt with the word 'AI' on it\".", + "2. AI + Heart Hands: \"Full-body portrait of a happy person wearing a shirt with the word 'AI' on it and making a heart shape with their hands\".", + "3. Heart Hands: \"Full-body portrait of a happy person making a heart shape with their hands\"." + ], + "bbox": [ + 513, + 507, + 905, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We manually filter the outcomes to ensure alignment with descriptions, with examples provided in the supplementary material in Section D. Then, we use a CLIP ViT-B/32 model [36] to extract a 512-dimensional feature vector for each image. Using these features, along with 25 feature vectors extracted from random noise images, we perform Principal Component Analysis (PCA) to reduce the dimensionality into a two-dimensional subspace. The resulting subspace, now suitable for visualization, is depicted in Figure 2.", + "bbox": [ + 511, + 613, + 905, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in the figure, the natural images form a smooth manifold, distinctly separated from the distribution of noise images. Within this manifold, there is a clear semantic progression: images of people with 'AI' shirts (green cluster) are close to images of people with 'AI' shirts making a heart shape (purple cluster), which are adjacent to images of people only making a heart shape (red cluster). Thus, transitioning smoothly along the manifold allows a person with an 'AI' shirt to perform a heart shape with their hands while preserving the shirt's text.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7860", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/96618022732d534bfd468ad7ceee43f6bb1e3ce0f5920fa40a49388dcf4df173.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSourceTarget
LPIPS↓CLIP-I↑CLIP↑
SDEdit0.300.850.60
Pix2Pix-ZERO0.290.840.62
Imagic0.520.860.63
LEDITS++0.230.870.63
FlowEdit0.220.890.61
F2F0.220.890.63
", + "bbox": [ + 120, + 88, + 452, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider an original image from the AI group that we wish to edit using the target prompt: \"A happy person making a heart shape with their hands\". Current editing methods generate a single image, which may cause an abrupt transition to the red cluster, effectively removing the 'AI' on the shirt. This behavior is illustrated in the figure, where the edit by LEDs++ is positioned near the red cluster, and the 'AI' text on the shirt disappears. In contrast, our method leverages video generation to perform the edit smoothly, moving along the manifold with incremental changes until reaching the required edit in the purple cluster, as indicated by black arrows in Figure 2. The temporal editing caption guiding this process is: \"A happy person very slowly raising their hands to form a heart shape\". This gradual progression lets us reach the purple cluster, resulting in an edited image where the person maintains the 'AI' on their shirt while making a heart shape—a faithful preservation of the original image's key attributes. This experiment illustrates that the proposed paradigm enables smooth traversal across the image manifold, enabling consistent edits while preserving the essential characteristics of the original image.", + "bbox": [ + 91, + 335, + 483, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 662, + 223, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate Frame2Frame against state-of-the-art image editing methods, including LEdits++ [4], SDEdit [31], Pix2Pix-Zero [35] Imagic [22], and FLUX-based FlowEdit [24]. Our evaluation spans two benchmarks: the established TEdBench [22] for general image editing, and our newly introduced PosEdit, specifically designed for human pose editing. Finally, we conduct a human evaluation to assess our method based on real user preferences.", + "bbox": [ + 89, + 685, + 482, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 89, + 816, + 279, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct our experiments following a consistent evaluation protocol across all methods and benchmarks. Following common practice [4, 22], for each method and source image, we manually select the best result from fifteen ran", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/090da99a4ebcec0f91b7163fea07b6b4a6424f87ea73d2873352e6aee0ca4c6c.jpg", + "table_caption": [ + "Table 1. TEdBench Results. Quantitative comparison on TEd-Bench benchmark. Source metrics (LPIPS and CLIP-I) measure content preservation, while Target metric (CLIP) measures edit accuracy. Our Frame2Frame (F2F) method achieves better or comparable performance across all metrics." + ], + "table_footnote": [], + "table_body": "
ModelSourceTarget
LPIPS↓CLIP-I↑LPIPS↓CLIP-I↑CLIP↑
GT (Reference)0.080.91010.61
SDEdit0.390.610.390.640.57
Pix2Pix-ZERO0.390.570.400.600.56
LEDITS++0.260.650.280.690.64
F2F0.140.820.150.840.64
", + "bbox": [ + 514, + 88, + 905, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. PosEdit Results. Quantitative evaluation on PosEdit. Source metrics assess similarity to the original image, while Target metrics include LPIPS and CLIP-I comparisons to the ground-truth target image, along with the CLIP score for edit accuracy. \"GT\" provides ground-truth target image metrics for context.", + "bbox": [ + 511, + 239, + 903, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dom seeds based on visual quality and edit accuracy, ensuring the same seed set is used across all methods. For all methods, we use the default hyperparameters and settings as provided in their official implementations or official Hugging Face repositories1.", + "bbox": [ + 511, + 324, + 905, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Preprocessing and Generation Details. We use CogVideoX [48] as our video generation backbone, which operates at a fixed resolution of $720 \\times 480$ pixels. Both TEd-Bench and PosEdit benchmarks consist of images with 1:1 aspect ratio. To accommodate CogVideoX's resolution requirements while preserving image content, we first resize all source images to $480 \\times 480$ pixels, then horizontally pad them with black pixels to reach the required $720 \\times 480$ resolution (adding 120 pixels on each side). After generating the video sequence, we crop the central $480 \\times 480$ region of the selected frame to remove the padding and resize it to the final evaluation resolution of $512 \\times 512$ pixels, matching the standard resolution used in prior work. For video generation, we adopt the default hyperparameters proposed by CogVideoX: a guidance scale of 6, 49 generated frames per sequence, and 50 denoising inference steps. This configuration generates videos approximately 6 seconds in duration at a frame rate of 8 frames per second.", + "bbox": [ + 511, + 412, + 906, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Method-Specific Requirements. Pix2Pix-Zero [35] requires additional source image descriptions along with the standard inputs. We generate these automatically using BLIP-2 [27], a state-of-the-art image captioning model, ensuring consistent source descriptions across all experiments. For our method, we transform the original editing prompts from both benchmarks into temporal editing captions, as described in Section 3.1.", + "bbox": [ + 511, + 702, + 905, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. TEDBench Evaluation Results", + "text_level": 1, + "bbox": [ + 511, + 828, + 782, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We quantitatively evaluate our method and baselines on the TEdBench benchmark using three complementary met", + "bbox": [ + 511, + 847, + 903, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1LEdits++,Pix2Pix-Zero,SDEdit", + "bbox": [ + 531, + 886, + 712, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7861", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1d6cff178fbd211640499817e1b9d3566be7a95b201fb664e76a553c22cd51c1.jpg", + "image_caption": [ + "Figure 4. Qualitative Results on TEdBench. Comparison with other methods across various editing tasks. Our approach consistently produces edits that better align with the target prompt while preserving the source image's content and structure. For instance, in the teddy bear example, our method uniquely achieves complex structural modifications while maintaining high visual quality." + ], + "image_footnote": [], + "bbox": [ + 145, + 97, + 854, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rics. For each method, we compute the metrics between the source image and its corresponding best edited version. LPIPS [50] measures perceptual similarity to assess how much the edit preserves the source image's content, with lower values indicating better preservation. CLIP-I evaluates the similarity between source and edited images in CLIP's [36] feature space, where higher values indicate better preservation of semantic content. Finally, CLIP score measures alignment between the edited image and the target prompt, with higher values indicating better adherence to the editing instruction. As shown in Table 1, our method achieves strong performance across all metrics, demonstrating effective balance between preserving source content and achieving the desired edit.", + "bbox": [ + 88, + 494, + 482, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The qualitative advantages of our method are visually evident in the comparisons presented in Figure 4. The figure demonstrates our method's superior performance across diverse editing scenarios, producing results that both faithfully execute the intended edits while maintaining strong alignment with the source image. Additional examples showcasing these capabilities are provided in the appendix in Section A.", + "bbox": [ + 89, + 705, + 482, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. PosEdit Benchmark", + "text_level": 1, + "bbox": [ + 89, + 835, + 282, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We introduce PosEdit, a benchmark for human pose editing derived from the UTD-MHAD dataset [6]. UTD-MHAD includes RGB videos of 8 subjects (4 females and 4 males)", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performing predefined actions in a controlled indoor environment. We carefully curated 58 editing tasks encompassing 8 distinct action categories, ranging from simple poses like a raised hand to complex athletic poses such as basketball shooting and lunging.", + "bbox": [ + 511, + 494, + 906, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed benchmark specifically focuses on human pose manipulation. Unlike TEdBench, each editing task in PosEdit includes a ground truth frame extracted from the same subject in the target pose. The availability of reference images allows for a more comprehensive evaluation of both editing accuracy and identity preservation. In the dataset, each editing task consists of two images: a source image showing the subject in a neutral standing pose with arms relaxed at their sides, and a ground-truth edited target image capturing the subject performing in a specific pose. The benchmark also provides a prompt for each task that describes the target pose the subject should achieve (e.g., \"A person in a basketball shooting posture\").", + "bbox": [ + 511, + 570, + 908, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark Evaluation. The evaluation on PosEdit follows the same protocol established for TEdBench: generating multiple variants using nine different seeds and manually selecting the best result for each method. However, PosEdit's ground-truth target images enable additional evaluation metrics. Beyond measuring preservation of source content (via LPIPS and CLIP-I between source and edited images) and edit accuracy (via CLIP score with the prompt),", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7862", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a16e8b9e7a7b6f540b789dded85b7b895e39e854787ded52815966272a019583.jpg", + "image_caption": [ + "Figure 5. Qualitative Results on PosEdit. Comparison between our Frame2Frame method and LEDITS++ on human motion editing tasks. For each example, we show the source image, edited results from both methods, and the ground-truth target image. Our method better preserves subject identity while achieving more natural pose transitions. The evaluation metrics for each image are provided in Section F of the appendix." + ], + "image_footnote": [], + "bbox": [ + 101, + 93, + 480, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we compute LPIPS and CLIP-I metrics between the edited image and its corresponding ground-truth target. This enhanced evaluation directly assesses both the accuracy of the transformation and preservation of identity features. Additionally, we include the evaluation metrics of the ground-truth images as a reference to the rest of the results.", + "bbox": [ + 89, + 411, + 483, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 demonstrates that our method consistently outperforms competitive methods across all metrics, with particularly strong performance in similarity measures with the target ground-truth image. This advantage highlights our method's superior ability to preserve key features while achieving the desired edit, as illustrated by the qualitative comparisons in Figure 5. The figure shows that our method generates more natural pose transitions while maintaining crucial identity attributes such as facial features, body proportions, and clothing details.", + "bbox": [ + 89, + 502, + 483, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two important implementation notes: First, for Pix2PixZero, which requires source image descriptions, we use the same static prompt across all tasks: \"A person standing naturally with his arms relaxed at his sides.\" Second, we exclude Imagic [22] from this comparison as their official implementation is not publicly available and their results are only reported on TEdBench.", + "bbox": [ + 89, + 654, + 483, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Human Evaluation Survey", + "text_level": 1, + "bbox": [ + 89, + 770, + 331, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further evaluate our method, we conducted a human survey. As indicated in Table 1, LEDs++ emerges as the most competitive method compared to ours, making it a natural choice for comparison. Each participant was presented with 20 randomly sampled TEdBench samples, including the source image, the target edit prompt, and the corresponding edited outputs from both methods. Inspired", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9b314e0e3f74bbc4a20c5434f58416446248a32c4ef086eaae8a587b889c00a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodEdit AccuracyEdit Quality
OverallPer-ImageOverallPer-Image
F2F54.1%53.0%65.6%67.0%
LEDITS++45.9%47.0%34.4%33.0%
", + "bbox": [ + 514, + 89, + 901, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Human Survey Results. Human evaluation on TEd-Bench shows that F2F surpasses LEDs++ in edit accuracy while offering a significant advantage in preserving the original image.", + "bbox": [ + 511, + 176, + 906, + 219 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by the methodology proposed by [49], participants evaluated each comparison based on two criteria: (1) edit accuracy relative to the prompt, and (2) edit quality—defined as the preservation of visual fidelity to the source image, seamless integration of edited elements, and the overall natural appearance of modifications. We collected responses from 59 randomly selected online evaluators. To ensure an unbiased assessment, evaluators were not informed of the study's objectives. The results were quantified using two metrics: (i) overall global preference, expressed as a percentage, and (ii) aggregated per-image preference, where a tie resulted in each method receiving 0.5 points.", + "bbox": [ + 511, + 233, + 906, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our results, summarized in Table 3, demonstrate that F2F outperforms LEdits++ on both metrics. Specifically: For edit quality, F2F achieved a global preference score of $53\\%$ , slightly surpassing the $47\\%$ obtained by LEdits++. In terms of edit accuracy, F2F achieved a $65.6\\%$ global preference compared to LEdits++'s $34.4\\%$ . The per-image preference results mirrored this trend, indicating robustness across individual examples and no significant influence of outliers. These findings reinforce our claim that smooth temporal editing—using video as a medium—preserves essential scene characteristics while successfully performing the edit. Our results suggest that the gap between F2F and LEdits++ in source preservation is larger than reflected by the LPIPS scores in Table 1. The full survey details are provided in the supplementary materials in Section G.", + "bbox": [ + 511, + 415, + 908, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Additional Vision Tasks", + "text_level": 1, + "bbox": [ + 511, + 650, + 733, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate our framework's applicability beyond traditional editing by applying it to fundamental image manipulation tasks: denoising, deblurring, outpainting, and relighting. For these experiments, we utilize Runway Gen-3 [13] as our video generation backbone, as it showed superior performance on these specific tasks compared to other generative video models. Since these tasks serve as extensions and potential applications of our method, we focus on qualitative results, leaving quantitative evaluation for future work. Our results, shown in Figure 6, demonstrate strong performance across all tasks. We associate this success to the natural alignment between these operations and common video scenarios: deblurring maps to camera focusing, outpainting to camera motion, and relighting to time-lapse lighting changes. This alignment enables our", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7863", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03566d17a72cee53af2ebdb84fd43fea620f3fb3ea5730971a110b7b521c8a7f.jpg", + "image_caption": [ + "Figure 6. Additional Vision Tasks. Qualitative results of our image-to-video-to-image editing approach on selected traditional tasks." + ], + "image_footnote": [], + "bbox": [ + 93, + 93, + 205, + 361 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aad66f8d257570f75d637ed2551ddda1a4b4f406f5bd3871d0faeff54939ad5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 93, + 444, + 361 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8e6b3b82f01a525d6ccef2ff60cb4396a3c6ebd4862ae9c70d130d66190a69d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 93, + 669, + 361 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b246a518221c2fd0623766d7892bd1bad54b0ddaa0a182296ff36d10a0b5190.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 93, + 787, + 361 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c0f05f0f41dadc09a1a172347381c19404933054d0fad9774def4a87c371865.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 93, + 903, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method to leverage the video model's learned temporal dynamics, achieving high-quality results without task-specific training. Full prompts are provided in Section I of the supplementary material.", + "bbox": [ + 89, + 414, + 482, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extended Results. Additional examples illustrating our method's editing performance across diverse backgrounds are provided in Section H of the Appendix.", + "bbox": [ + 89, + 484, + 483, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 89, + 544, + 212, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our approach addresses key shortcomings in current image editing methods, it also introduces unique challenges. For example, natural camera motion sometimes appears in video sequences, which, when replicated in generated content, can lead to unintended perspective shifts. As with all generative models, video models are trained on specific data domains, making it challenging to produce results that deviate significantly from the model's training data, which predominantly includes real-world transformations. Despite these challenges, F2F demonstrates success in several cases. For example, in the middle row of Figure 1, the vase is \"magically\" filled with green water, showcasing the model's ability to perform imaginative edits, with additional examples discussed in Section H of the supplementary. Additionally, our method is computationally intensive, as transforming an image into a video sequence is resource-heavy and often slower than other image editing methods. However, video generation efficiency is advancing rapidly, as demonstrated by models like Runway Turbo $^{2}$ and LTX-Video [17], which can execute the process in", + "bbox": [ + 89, + 571, + 482, + 872 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "seconds, making this approach increasingly less resource-intensive. Additionally, optimizing the number of frames per edit—currently fixed at 49 for CogVideoX—could enable faster editing possibilities.", + "bbox": [ + 511, + 414, + 905, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 484, + 638, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced Frame2Frame, a novel approach that reformulates image editing through video generation. By leveraging video models' inherent understanding of temporal transformations, our method achieves state-of-the-art editing results while maintaining high fidelity to source images. We demonstrated our framework's effectiveness on standard benchmarks, introduced PosEdit for human pose editing, and showed promising results on more classical vision tasks. As video generation technology advances, we expect our approach to enable increasingly sophisticated image manipulations while maintaining natural and physically plausible results.", + "bbox": [ + 511, + 508, + 906, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future Research. Our work opens several promising directions. First, fine-tuning video generators specifically for image editing presents an exciting opportunity, including straightforward solutions like enforcing static camera scenarios or using datasets curated for editing tasks, alongside more complex approaches yet to emerge. Secondly, a key direction could involve reducing the overhead of full video generation while preserving the benefits of gradual, temporally coherent transformations, potentially enhancing both the efficiency and speed of editing. The speed can be affected both in terms of per-frame average generation speed or the number of frames generated per edit, which is currently fixed at 49 for CogVideoX.", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "2Runway Turbo API", + "bbox": [ + 107, + 886, + 220, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7864", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hadi Alzayer, Zhihao Xia, Xuaner Zhang, Eli Shechtman, Jia-Bin Huang, and Michael Gharbi. Magic fixup: Streamlining photo editing by watching dynamic videos. arXiv preprint arXiv:2403.13044, 2024. 3", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2", + "[4] Manuel Brack, Felix Friedrich, Katharina Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. 2023. 5", + "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2", + "[6] Chen Chen, Roozbeh Jafari, and Nasser Kehtarnavaz. Utdmhad: A multimodal dataset for human action recognition utilizing a depth camera and a wearable inertial sensor. In 2015 IEEE International conference on image processing (ICIP), pages 168-172. IEEE, 2015. 2, 6", + "[7] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320–13331, 2024. 2", + "[8] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 3", + "[9] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. In ICCV, 2019. 2", + "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3", + "[11] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 2", + "[12] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 3", + "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models, 2023. 7", + "[14] Roy Ganz, Yair Kittenplon, Aviad Aberdam, Elad Ben Avraham, Oren Nuriel, Shai Mazor, and Ron Litman. Question aware vision transformer for multimodal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13861-13871, 2024. 3", + "[15] Yue Gao, Yuan Zhou, Jinglu Wang, Xiao Li, Xiang Ming, and Yan Lu. High-fidelity and freely controllable talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5609-5619, 2023. 2", + "[16] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 2", + "[17] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024.8", + "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[20] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024. 2", + "[21] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12469-12478, 2024. 2", + "[22] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1, 2, 5, 7", + "[23] Wonkyun Kim, Changin Choi, Wonseok Lee, and Wonjong Rhee. An image grid can be worth a video: Zero-shot video question answering using a vlm. arXiv preprint arXiv:2403.18406, 2024. 4, 12", + "[24] Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. arXiv preprint arXiv:2412.08629, 2024. 5", + "[25] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7865", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775-6785, 2024. 2", + "[26] Black Forest Labs. Flux.1-dev. https://github.com/black-forest-labs/FLUX.1-dev, 2024.4", + "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 5", + "[28] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2", + "[29] Grace Luo, Trevor Darrell, Oliver Wang, Dan B Goldman, and Aleksander Holynski. Readout guidance: Learning control from diffusion features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8227, 2024. 3", + "[30] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2", + "[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 5", + "[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2", + "[33] OpenAI. Chatgpt-4o (october 2024 version). https://chat.openai.com/chat, 2024.3", + "[34] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2", + "[35] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 5", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[38] Noam Rotstein, David Bensaïd, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models for enriched fused image captions. In Proceedings of the" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5689-5700, 2024. 3", + "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2", + "[40] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024. 3", + "[41] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 2", + "[42] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning. PMLR, 2015. 2", + "[43] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, 2018. 2", + "[44] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. In NeurIPS, 2016. 2", + "[45] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 3", + "[46] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2024. 3", + "[47] Navee Wasserman, Noam Rotstein, Roy Ganz, and Ron Kimmel. Paint by inpaint: Learning to add image objects by removing them first. arXiv preprint arXiv:2404.18212, 2024.2", + "[48] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 3, 5", + "[49] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 7, 14", + "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[51] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision. Springer, 2025. 2" + ], + "bbox": [ + 516, + 92, + 903, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7866", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/f625c957-1581-449f-8cd5-463f4552272c_model.json b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/f625c957-1581-449f-8cd5-463f4552272c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..df92a6f501bbd1c0ff6e4236c27eeca914e127b3 --- /dev/null +++ b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/f625c957-1581-449f-8cd5-463f4552272c_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.856, + 0.153 + ], + "angle": 0, + "content": "Pathways on the Image Manifold: Image Editing via Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.182, + 0.425, + 0.198 + ], + "angle": 0, + "content": "Noam Rotstein" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.2, + 0.393, + 0.217 + ], + "angle": 0, + "content": "Roy Velich" + }, + { + "type": "text", + "bbox": [ + 0.465, + 0.182, + 0.543, + 0.198 + ], + "angle": 0, + "content": "Gal Yona" + }, + { + "type": "text", + "bbox": [ + 0.433, + 0.2, + 0.551, + 0.216 + ], + "angle": 0, + "content": "David Bensaïd" + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.182, + 0.692, + 0.198 + ], + "angle": 0, + "content": "Daniel Silver" + }, + { + "type": "text", + "bbox": [ + 0.591, + 0.2, + 0.697, + 0.216 + ], + "angle": 0, + "content": "Ron Kimmel" + }, + { + "type": "title", + "bbox": [ + 0.334, + 0.234, + 0.66, + 0.253 + ], + "angle": 0, + "content": "Technion - Israel Institute of Technology" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.485, + 0.591 + ], + "angle": 0, + "content": "Recent advances in image editing, driven by image diffusion models, have shown remarkable progress. However, significant challenges remain, as these models often struggle to follow complex edit instructions accurately and frequently compromise fidelity by altering key elements of the original image. Simultaneously, video generation has made remarkable strides, with models that effectively function as consistent and continuous world simulators. In this paper, we propose merging these two fields by utilizing image-to-video models for image editing. We reformulate image editing as a temporal process, using pretrained video models to create smooth transitions from the original image to the desired edit. This approach traverses the image manifold continuously, ensuring consistent edits while preserving the original image's key aspects. Our approach achieves state-of-the-art results on text-based image editing, demonstrating significant improvements in both edit accuracy and image preservation. Visit our project page." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.604, + 0.222, + 0.619 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Image editing has witnessed remarkable advancements through deep learning and text-guided diffusion networks. These developments have set a new benchmark for image manipulations, enhancing both control and quality. However, current approaches continue to face significant limitations in real-world scenarios. These methods often struggle with two key challenges: achieving precise edits that accurately reflect the intended modifications, and preserving the essential characteristics of the original image content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.484, + 0.902 + ], + "angle": 0, + "content": "State-of-the-art techniques predominantly rely on text-guided diffusion models, which iteratively denoise random latent representations to generate edited images. Such methods condition the generation process on both the source image—using techniques such as latent inversion [32] or model fine-tuning [22]—and the target edit description. However, these approaches require the model to generate a single output image that preserves source image fidelity while implementing complex edits, often compromises" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.286, + 0.619, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.631, + 0.311, + 0.787, + 0.345 + ], + "angle": 0, + "content": "\"A photo of a tomato in a blue tennis court.\"" + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.286, + 0.898, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.404, + 0.62, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.636, + 0.425, + 0.779, + 0.461 + ], + "angle": 0, + "content": "\"A vase of flowers filled with green water.\"" + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.405, + 0.899, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.523, + 0.619, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.631, + 0.549, + 0.787, + 0.574 + ], + "angle": 0, + "content": "\"A photo of a teddy bear doing pushups.\"" + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.523, + 0.898, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.599, + 0.876, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.667, + 0.907, + 0.752 + ], + "angle": 0, + "content": "Figure 1. Visualization of Frame2Frame's editing process. Temporal progression of our video-based approach. Starting from the source image (leftmost), frames illustrate the natural evolution toward the target edit (rightmost). Our method produces temporally coherent intermediate states while preserving fidelity to both the source content and the editing intent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.798, + 0.779 + ], + "angle": 0, + "content": "ing edit accuracy and content preservation." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a paradigm shift in image editing by reformulating it as a video generation task. Rather than a single-state transition, our approach harnesses temporal coherence: the source image serves as the initial frame of a video that progressively and naturally transforms toward the target edit. This temporal evolution allows the editing process to unfold through physically plausible intermediate states, providing a continuous path between source" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7857" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "and target images, as illustrated in Figure 1 and Figure S9 in the appendix. This temporal approach leverages the sophisticated world understanding embedded in recent video generation models, which have achieved breakthrough results in temporal coherence and visual quality through training on large-scale internet data [2, 48]. From a geometrical perspective, conventional editing approaches project initial noise onto the natural image manifold, targeting a single point where the image aligns with both the source and the edit request. In contrast, our approach generates a continuous path along the manifold between the original and edited image, producing a smooth realistic transition across different image states, as thoroughly discussed in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.29, + 0.485, + 0.606 + ], + "angle": 0, + "content": "We implement the proposed approach through a structured pipeline called Frame2Frame (F2F). First, we transform the edit instruction into a Temporal Editing Caption — a scenario describing how the edit should naturally evolve over time—using a pretrained Vision-Language Model (VLM). Next, a state-of-the-art image-to-video model generates a temporally coherent sequence guided by the temporal caption. Finally, we identify the frame that best realizes the desired edit with the assistance of a VLM. Extensive experiments demonstrate improvements over existing imaged-to-image approaches. We evaluate on TEdBench [22] and PosEdit, a newly curated dataset derived from UTD-MHAD [6], which focuses on human pose transformations. PosEdit pairs source images with ground-truth targets of the same subject in different poses, enabling rigorous evaluation of both edit accuracy and source fidelity. Beyond commonly defined editing tasks, our framework shows promising results in more classical computer vision problems such as de-blurring, de-noising, and relighting by recasting them as temporal progressions, suggesting broader applications for video-based image transformations." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.608, + 0.34, + 0.621 + ], + "angle": 0, + "content": "Our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.624, + 0.483, + 0.681 + ], + "angle": 0, + "content": "1. Reformulating image editing as a generative video task—leveraging temporal coherence to create edit paths on the natural image manifold, enabling high-fidelity manipulations while preserving source characteristics." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.684, + 0.482, + 0.741 + ], + "angle": 0, + "content": "2. Frame2Frame: an end-to-end framework that realizes the reformulation through three key components: (1) temporal editing captions, (2) generated video-based editing, and (3) automated frame selection." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.744, + 0.482, + 0.787 + ], + "angle": 0, + "content": "3. Comprehensive evaluation showing state-of-the-art performance on TEdBench and PosEdit, a new dataset for evaluating human pose edits." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.624, + 0.483, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.806, + 0.245, + 0.821 + ], + "angle": 0, + "content": "2. Related Efforts" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.832, + 0.238, + 0.848 + ], + "angle": 0, + "content": "2.1. Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Text-based image editing has advanced significantly with the success of generative diffusion models [19]. These models, in their text-to-image version, generate images through" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.38 + ], + "angle": 0, + "content": "a denoising diffusion process conditioned on input text, relying on ground truth text-image pairs for training [37, 39]. In contrast, image editing often lacks predefined ground truth data for source and target images, posing unique challenges. This limitation has led researchers to explore diverse editing methodologies [20]. For example, SDEdit [30] injects noise into an image and then denoises it based on an editing target prompt. Imagic [22] fine-tunes a text-to-image model on a single image, subsequently interpolating between input and target text embeddings to produce edits. Other methods first invert the input image into a diffusion model's latent space [21, 32] and then generate the edited image from that latent representation using various techniques for structure preservation and manipulation [18, 34]. InstructPix2Pix [5] synthesizes an editing dataset using the approach in [18], filters it, and employs this dataset to train a diffusion model in a supervised fashion. Paint-by-Inpaint [47] further explores this supervised approach, generating a real-image dataset for object insertion." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.392, + 0.745, + 0.407 + ], + "angle": 0, + "content": "2.2. Generative Video Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.909, + 0.763 + ], + "angle": 0, + "content": "Recent years have seen remarkable advancements in video generation, evolving from domain-specific systems [15] and brief clips to models capable of generating diverse, high-fidelity content. This progress has stemmed from paradigm shifts and the large-scale expansion of datasets [7] and architectures. Approaches have evolved from recurrent networks [16, 42] and generative adversarial networks (GANs) [9, 43, 44] to latent diffusion models (LDMs) [2, 3, 11], which leverage large U-Net or transformer-based architectures alongside vast internet-sourced datasets. Notable recent efforts include Stable Video Diffusion [2], which trains an LDM on large curated datasets, and OpenAI's Sora [28], which achieves impressive results by integrating large architectures with extensive public and private datasets. These models, often termed \"world simulators\" due to their emergent understanding of physical dynamics and temporal coherence. Our work builds upon CogVideoX [48], a transformer-based latent diffusion model that employs a 3D Variational Autoencoder to compress videos across both spatial and temporal dimensions, enhancing coherence. To improve text-video alignment, CogVideoX also integrates an expert transformer with adaptive LayerNorm, enabling deep fusion of visual and textual modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Close to our work, several recent efforts have utilized the world simulation capabilities of video diffusion models for various computer vision tasks. Make-A-Video3D [41] temporally extends static NeRFs using Score Distillation Sampling from video models. ViVid-1-to-3 [25] generates images along a camera trajectory around objects to enable novel view synthesis. PhysDreamer [51] models rigid object properties through 3D Gaussians and material fields, trained via distillation from pretrained video generators." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7858" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.209, + 0.093, + 0.795, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.393, + 0.908, + 0.451 + ], + "angle": 0, + "content": "Figure 2. Editing Manifold Pathway. Given an input image and target caption \"A happy person making a heart shape with their hands\", our method generates a continuous path on the natural image manifold. Each generated frame (indicated by black arrows) represents a plausible intermediate state between the source and target, maintaining temporal consistency throughout the transformation. As a result, in contrast to the competing approach, F2F achieves the desired edit while preserving the \"AI\" text on the person's shirt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.465, + 0.32, + 0.48 + ], + "angle": 0, + "content": "2.3. Image Editing and Video" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.726 + ], + "angle": 0, + "content": "The intersection of image editing and video has received limited attention. Existing methods focus on sampling pairs of random frames from videos to build image pair datasets that capture the same subject under varying conditions. For example, AnyDoor [8] uses the paired frames as an augmentation method, segmenting foreground objects in each frame and assigning one masked object as the target edited appearance of the subject. MagicFixup [1] employs these frames to build a dataset focused on refining user-made subject coarse 2D edits. Recently, drag-based editing approaches [29, 40], used frames extracted from video and computed optical flow to collect a dataset aimed at editing by spatially dragging points within the image. In contrast to these approaches, which rely on frame sampling for image data collection, we are the first to directly perform image editing using generative video models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.735, + 0.238, + 0.749 + ], + "angle": 0, + "content": "3. Frame2Frame" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.754, + 0.484, + 0.846 + ], + "angle": 0, + "content": "We present Frame2Frame, a framework that reformulates image editing as a temporal transformation process. Our approach leverages video generation models to create natural transitions between source and target images, achieving consistent and realistic edits. The proposed method has three main steps, as illustrated in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.85, + 0.337, + 0.867 + ], + "angle": 0, + "content": "3.1. Temporal Editing Captions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Text-based image editing methods typically operate on two inputs: a source image \\( I_{s} \\) and a target caption \\( c \\), where \\( c \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.907, + 0.571 + ], + "angle": 0, + "content": "specifies the desired modifications to \\( I_{s} \\). Our approach differs fundamentally by modeling editing as a temporal process. This requires a novel type of prompt—the Temporal Editing Caption, denoted by \\( \\tilde{c} \\)—that describes the sequential transformation from source to target image. We construct \\( \\tilde{c} \\) by combining information from \\( I_{s} \\) and \\( c \\) to create a description of how the desired edit unfolds over time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.909, + 0.77 + ], + "angle": 0, + "content": "To automate this process, we leverage recent advances in Vision-Language Models (VLMs) [10, 14, 38, 45, 46]. The VLM, specifically ChatGPT-4o [33], is instructed to produce a concise video scenario that highlights how elements within the image change or move over time. The generated caption captures the essential transformations while maintaining a static camera perspective unless movement is necessary. To improve generation quality, we utilize in-context learning (ICL) [12], providing the VLM with nine exemplar prompt-caption pairs. The complete prompt template, ICL examples, and an ablation study comparing this approach with directly using the target captions are included in Section B of the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.688, + 0.8 + ], + "angle": 0, + "content": "3.2. Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We employ CogVideoX (I2V-5B) [48], a pretrained generative video latent diffusion model utilizing a transformer-based architecture. Specifically, we use its image-to-video variant, which has been fine-tuned to generate videos starting from an input image \\( I_{s} \\). During generation, \\( I_{s} \\) is encoded and concatenated with noise in latent space, where" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7859" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "the model applies a denoising process guided by the temporal caption \\(\\tilde{c}\\). As elaborated in Section 4, this conditioning allows generated videos to start from \\(I_{s}\\) and evolve naturally along the image manifold, maintaining temporal coherence and consistency. Additionally, the model's transformer architecture enables effective fusion of visual and textual information, allowing precise control over the editing process through our temporal captions. Formally, given the video generator \\(G\\), we define the generation process as:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.24, + 0.372, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} (I _ {s}, \\tilde {c}) = V = \\{f _ {1}, \\dots , f _ {T} \\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.272, + 0.484, + 0.3 + ], + "angle": 0, + "content": "where \\(V\\) denotes the generated video with \\(T\\) frames, and \\(f_{t}\\) represents the frame at timestep \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.314, + 0.253, + 0.329 + ], + "angle": 0, + "content": "3.3. Frame Selection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.484, + 0.564 + ], + "angle": 0, + "content": "We observed that the optimal number of frames required for an edit can vary—small changes may be completed in fewer frames, while more extensive transformations often necessitate additional ones. Additionally, later frames tend to deviate further from the source image. Thus, even though \\( V \\) serves as an editing path originating from \\( I_{s} \\), there is no guarantee that \\( f_{T} \\) is the optimal edited image in \\( V \\). For instance, in Figure 3, the edited image of the cat might be expected to capture the midpoint of its jump, but the last frame shows the jump already completed. Therefore, we aim to identify the optimal edited frame, denoted \\( f_{t^{*}} \\), which corresponds to the earliest timestep \\( t \\) that achieves the desired edit. The transition from the initial frame \\( f_{1} \\) (or \\( I_{s} \\)) to the final edited frame \\( f_{t^{*}} \\) motivates our method's name, Frame2Frame." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.809 + ], + "angle": 0, + "content": "To automate the selection of \\( t^* \\) and avoid manual frame-by-frame review, we employ an automated approach. After generating the sequence \\( V \\), we sample every fourth frame, imprinting each with a unique identifier and assembling them into an image collage alongside \\( I_s \\). Inspired by [23], which introduces a novel approach to video comprehension by transforming videos into image grids, we use a VLM, specifically GPT-4o, to assist in selecting \\( t^* \\) by providing it with the collage and the editing prompt \\( c \\). The VLM is tasked with identifying the frame that best fulfills the editing intent, evaluating each frame's alignment with \\( c \\) and fidelity to \\( I_s \\). The model is instructed to select the optimal frame with the lowest index that completes the edit. An ablation study, detailed in Section C of the appendix, evaluates the effectiveness of our automated approach against the baseline of selecting the final frame of the video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Note that while we refer to an optimal \\( t^* \\), the definition of the required edit can be subjective and dependent on user preferences. For example, in Figure 3, different users may select frames of the cat jumping based on variations in its altitude. Thus, frame selection could serve as a flexible and customizable advantage in certain scenarios." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.907, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.23, + 0.907, + 0.329 + ], + "angle": 0, + "content": "Figure 3. Frame2Frame Overview. Given a source image and editing prompt, our pipeline proceeds in three steps. First, a Vision-Language Model generates a temporal caption describing the transformation. Next, this caption guides a video generator to create a natural progression of the edit. Finally, our frame selection strategy identifies the optimal frame that best realizes the desired edit, producing the final image of the cat mid-leap." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.344, + 0.761, + 0.363 + ], + "angle": 0, + "content": "4. Editing Manifold Pathway" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.506 + ], + "angle": 0, + "content": "To illustrate the advantages of our method over conventional image-to-image approaches, we aim to visualize the editing process within the natural image manifold, where realistic images reside. Given the high dimensionality of this space, we project it into a lower-dimensional representation suitable for interpretation. To achieve this, we first generate three sets of images, each containing 200 samples, using the text-to-image generator FLUX.1-dev [26]. The sets are based on the following prompts:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.508, + 0.906, + 0.536 + ], + "angle": 0, + "content": "1. AI: \"Full-body portrait of a happy person wearing a shirt with the word 'AI' on it\"." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.538, + 0.906, + 0.581 + ], + "angle": 0, + "content": "2. AI + Heart Hands: \"Full-body portrait of a happy person wearing a shirt with the word 'AI' on it and making a heart shape with their hands\"." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.583, + 0.905, + 0.613 + ], + "angle": 0, + "content": "3. Heart Hands: \"Full-body portrait of a happy person making a heart shape with their hands\"." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.508, + 0.906, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.749 + ], + "angle": 0, + "content": "We manually filter the outcomes to ensure alignment with descriptions, with examples provided in the supplementary material in Section D. Then, we use a CLIP ViT-B/32 model [36] to extract a 512-dimensional feature vector for each image. Using these features, along with 25 feature vectors extracted from random noise images, we perform Principal Component Analysis (PCA) to reduce the dimensionality into a two-dimensional subspace. The resulting subspace, now suitable for visualization, is depicted in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As illustrated in the figure, the natural images form a smooth manifold, distinctly separated from the distribution of noise images. Within this manifold, there is a clear semantic progression: images of people with 'AI' shirts (green cluster) are close to images of people with 'AI' shirts making a heart shape (purple cluster), which are adjacent to images of people only making a heart shape (red cluster). Thus, transitioning smoothly along the manifold allows a person with an 'AI' shirt to perform a heart shape with their hands while preserving the shirt's text." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7860" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.089, + 0.454, + 0.238 + ], + "angle": 0, + "content": "
ModelSourceTarget
LPIPS↓CLIP-I↑CLIP↑
SDEdit0.300.850.60
Pix2Pix-ZERO0.290.840.62
Imagic0.520.860.63
LEDITS++0.230.870.63
FlowEdit0.220.890.61
F2F0.220.890.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.249, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Table 1. TEdBench Results. Quantitative comparison on TEd-Bench benchmark. Source metrics (LPIPS and CLIP-I) measure content preservation, while Target metric (CLIP) measures edit accuracy. Our Frame2Frame (F2F) method achieves better or comparable performance across all metrics." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.336, + 0.485, + 0.654 + ], + "angle": 0, + "content": "Consider an original image from the AI group that we wish to edit using the target prompt: \"A happy person making a heart shape with their hands\". Current editing methods generate a single image, which may cause an abrupt transition to the red cluster, effectively removing the 'AI' on the shirt. This behavior is illustrated in the figure, where the edit by LEDs++ is positioned near the red cluster, and the 'AI' text on the shirt disappears. In contrast, our method leverages video generation to perform the edit smoothly, moving along the manifold with incremental changes until reaching the required edit in the purple cluster, as indicated by black arrows in Figure 2. The temporal editing caption guiding this process is: \"A happy person very slowly raising their hands to form a heart shape\". This gradual progression lets us reach the purple cluster, resulting in an edited image where the person maintains the 'AI' on their shirt while making a heart shape—a faithful preservation of the original image's key attributes. This experiment illustrates that the proposed paradigm enables smooth traversal across the image manifold, enabling consistent edits while preserving the essential characteristics of the original image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.225, + 0.681 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.807 + ], + "angle": 0, + "content": "We evaluate Frame2Frame against state-of-the-art image editing methods, including LEdits++ [4], SDEdit [31], Pix2Pix-Zero [35] Imagic [22], and FLUX-based FlowEdit [24]. Our evaluation spans two benchmarks: the established TEdBench [22] for general image editing, and our newly introduced PosEdit, specifically designed for human pose editing. Finally, we conduct a human evaluation to assess our method based on real user preferences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.281, + 0.833 + ], + "angle": 0, + "content": "5.1. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conduct our experiments following a consistent evaluation protocol across all methods and benchmarks. Following common practice [4, 22], for each method and source image, we manually select the best result from fifteen ran" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.906, + 0.231 + ], + "angle": 0, + "content": "
ModelSourceTarget
LPIPS↓CLIP-I↑LPIPS↓CLIP-I↑CLIP↑
GT (Reference)0.080.91010.61
SDEdit0.390.610.390.640.57
Pix2Pix-ZERO0.390.570.400.600.56
LEDITS++0.260.650.280.690.64
F2F0.140.820.150.840.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.24, + 0.905, + 0.311 + ], + "angle": 0, + "content": "Table 2. PosEdit Results. Quantitative evaluation on PosEdit. Source metrics assess similarity to the original image, while Target metrics include LPIPS and CLIP-I comparisons to the ground-truth target image, along with the CLIP score for edit accuracy. \"GT\" provides ground-truth target image metrics for context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.906, + 0.402 + ], + "angle": 0, + "content": "dom seeds based on visual quality and edit accuracy, ensuring the same seed set is used across all methods. For all methods, we use the default hyperparameters and settings as provided in their official implementations or official Hugging Face repositories1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.686 + ], + "angle": 0, + "content": "Image Preprocessing and Generation Details. We use CogVideoX [48] as our video generation backbone, which operates at a fixed resolution of \\(720 \\times 480\\) pixels. Both TEd-Bench and PosEdit benchmarks consist of images with 1:1 aspect ratio. To accommodate CogVideoX's resolution requirements while preserving image content, we first resize all source images to \\(480 \\times 480\\) pixels, then horizontally pad them with black pixels to reach the required \\(720 \\times 480\\) resolution (adding 120 pixels on each side). After generating the video sequence, we crop the central \\(480 \\times 480\\) region of the selected frame to remove the padding and resize it to the final evaluation resolution of \\(512 \\times 512\\) pixels, matching the standard resolution used in prior work. For video generation, we adopt the default hyperparameters proposed by CogVideoX: a guidance scale of 6, 49 generated frames per sequence, and 50 denoising inference steps. This configuration generates videos approximately 6 seconds in duration at a frame rate of 8 frames per second." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Method-Specific Requirements. Pix2Pix-Zero [35] requires additional source image descriptions along with the standard inputs. We generate these automatically using BLIP-2 [27], a state-of-the-art image captioning model, ensuring consistent source descriptions across all experiments. For our method, we transform the original editing prompts from both benchmarks into temporal editing captions, as described in Section 3.1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.829, + 0.783, + 0.844 + ], + "angle": 0, + "content": "5.2. TEDBench Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.848, + 0.905, + 0.88 + ], + "angle": 0, + "content": "We quantitatively evaluate our method and baselines on the TEdBench benchmark using three complementary met" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.714, + 0.9 + ], + "angle": 0, + "content": "1LEdits++,Pix2Pix-Zero,SDEdit" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7861" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.147, + 0.098, + 0.856, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.437, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results on TEdBench. Comparison with other methods across various editing tasks. Our approach consistently produces edits that better align with the target prompt while preserving the source image's content and structure. For instance, in the teddy bear example, our method uniquely achieves complex structural modifications while maintaining high visual quality." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.495, + 0.483, + 0.706 + ], + "angle": 0, + "content": "rics. For each method, we compute the metrics between the source image and its corresponding best edited version. LPIPS [50] measures perceptual similarity to assess how much the edit preserves the source image's content, with lower values indicating better preservation. CLIP-I evaluates the similarity between source and edited images in CLIP's [36] feature space, where higher values indicate better preservation of semantic content. Finally, CLIP score measures alignment between the edited image and the target prompt, with higher values indicating better adherence to the editing instruction. As shown in Table 1, our method achieves strong performance across all metrics, demonstrating effective balance between preserving source content and achieving the desired edit." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.483, + 0.829 + ], + "angle": 0, + "content": "The qualitative advantages of our method are visually evident in the comparisons presented in Figure 4. The figure demonstrates our method's superior performance across diverse editing scenarios, producing results that both faithfully execute the intended edits while maintaining strong alignment with the source image. Additional examples showcasing these capabilities are provided in the appendix in Section A." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.836, + 0.283, + 0.851 + ], + "angle": 0, + "content": "5.3. PosEdit Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We introduce PosEdit, a benchmark for human pose editing derived from the UTD-MHAD dataset [6]. UTD-MHAD includes RGB videos of 8 subjects (4 females and 4 males)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.907, + 0.571 + ], + "angle": 0, + "content": "performing predefined actions in a controlled indoor environment. We carefully curated 58 editing tasks encompassing 8 distinct action categories, ranging from simple poses like a raised hand to complex athletic poses such as basketball shooting and lunging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.909, + 0.768 + ], + "angle": 0, + "content": "The proposed benchmark specifically focuses on human pose manipulation. Unlike TEdBench, each editing task in PosEdit includes a ground truth frame extracted from the same subject in the target pose. The availability of reference images allows for a more comprehensive evaluation of both editing accuracy and identity preservation. In the dataset, each editing task consists of two images: a source image showing the subject in a neutral standing pose with arms relaxed at their sides, and a ground-truth edited target image capturing the subject performing in a specific pose. The benchmark also provides a prompt for each task that describes the target pose the subject should achieve (e.g., \"A person in a basketball shooting posture\")." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Benchmark Evaluation. The evaluation on PosEdit follows the same protocol established for TEdBench: generating multiple variants using nine different seeds and manually selecting the best result for each method. However, PosEdit's ground-truth target images enable additional evaluation metrics. Beyond measuring preservation of source content (via LPIPS and CLIP-I between source and edited images) and edit accuracy (via CLIP score with the prompt)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7862" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.094, + 0.482, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.299, + 0.484, + 0.397 + ], + "angle": 0, + "content": "Figure 5. Qualitative Results on PosEdit. Comparison between our Frame2Frame method and LEDITS++ on human motion editing tasks. For each example, we show the source image, edited results from both methods, and the ground-truth target image. Our method better preserves subject identity while achieving more natural pose transitions. The evaluation metrics for each image are provided in Section F of the appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.501 + ], + "angle": 0, + "content": "we compute LPIPS and CLIP-I metrics between the edited image and its corresponding ground-truth target. This enhanced evaluation directly assesses both the accuracy of the transformation and preservation of identity features. Additionally, we include the evaluation metrics of the ground-truth images as a reference to the rest of the results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Table 2 demonstrates that our method consistently outperforms competitive methods across all metrics, with particularly strong performance in similarity measures with the target ground-truth image. This advantage highlights our method's superior ability to preserve key features while achieving the desired edit, as illustrated by the qualitative comparisons in Figure 5. The figure shows that our method generates more natural pose transitions while maintaining crucial identity attributes such as facial features, body proportions, and clothing details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.484, + 0.76 + ], + "angle": 0, + "content": "Two important implementation notes: First, for Pix2PixZero, which requires source image descriptions, we use the same static prompt across all tasks: \"A person standing naturally with his arms relaxed at his sides.\" Second, we exclude Imagic [22] from this comparison as their official implementation is not publicly available and their results are only reported on TEdBench." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.332, + 0.789 + ], + "angle": 0, + "content": "5.4. Human Evaluation Survey" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To further evaluate our method, we conducted a human survey. As indicated in Table 1, LEDs++ emerges as the most competitive method compared to ours, making it a natural choice for comparison. Each participant was presented with 20 randomly sampled TEdBench samples, including the source image, the target edit prompt, and the corresponding edited outputs from both methods. Inspired" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.09, + 0.903, + 0.169 + ], + "angle": 0, + "content": "
MethodEdit AccuracyEdit Quality
OverallPer-ImageOverallPer-Image
F2F54.1%53.0%65.6%67.0%
LEDITS++45.9%47.0%34.4%33.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.177, + 0.907, + 0.22 + ], + "angle": 0, + "content": "Table 3. Human Survey Results. Human evaluation on TEd-Bench shows that F2F surpasses LEDs++ in edit accuracy while offering a significant advantage in preserving the original image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.907, + 0.415 + ], + "angle": 0, + "content": "by the methodology proposed by [49], participants evaluated each comparison based on two criteria: (1) edit accuracy relative to the prompt, and (2) edit quality—defined as the preservation of visual fidelity to the source image, seamless integration of edited elements, and the overall natural appearance of modifications. We collected responses from 59 randomly selected online evaluators. To ensure an unbiased assessment, evaluators were not informed of the study's objectives. The results were quantified using two metrics: (i) overall global preference, expressed as a percentage, and (ii) aggregated per-image preference, where a tie resulted in each method receiving 0.5 points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.909, + 0.644 + ], + "angle": 0, + "content": "Our results, summarized in Table 3, demonstrate that F2F outperforms LEdits++ on both metrics. Specifically: For edit quality, F2F achieved a global preference score of \\(53\\%\\), slightly surpassing the \\(47\\%\\) obtained by LEdits++. In terms of edit accuracy, F2F achieved a \\(65.6\\%\\) global preference compared to LEdits++'s \\(34.4\\%\\). The per-image preference results mirrored this trend, indicating robustness across individual examples and no significant influence of outliers. These findings reinforce our claim that smooth temporal editing—using video as a medium—preserves essential scene characteristics while successfully performing the edit. Our results suggest that the gap between F2F and LEdits++ in source preservation is larger than reflected by the LPIPS scores in Table 1. The full survey details are provided in the supplementary materials in Section G." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.734, + 0.667 + ], + "angle": 0, + "content": "5.5. Additional Vision Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We demonstrate our framework's applicability beyond traditional editing by applying it to fundamental image manipulation tasks: denoising, deblurring, outpainting, and relighting. For these experiments, we utilize Runway Gen-3 [13] as our video generation backbone, as it showed superior performance on these specific tasks compared to other generative video models. Since these tasks serve as extensions and potential applications of our method, we focus on qualitative results, leaving quantitative evaluation for future work. Our results, shown in Figure 6, demonstrate strong performance across all tasks. We associate this success to the natural alignment between these operations and common video scenarios: deblurring maps to camera focusing, outpainting to camera motion, and relighting to time-lapse lighting changes. This alignment enables our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7863" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.094, + 0.207, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.094, + 0.446, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.094, + 0.671, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.094, + 0.788, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.094, + 0.905, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.373, + 0.893, + 0.387 + ], + "angle": 0, + "content": "Figure 6. Additional Vision Tasks. Qualitative results of our image-to-video-to-image editing approach on selected traditional tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.475 + ], + "angle": 0, + "content": "method to leverage the video model's learned temporal dynamics, achieving high-quality results without task-specific training. Full prompts are provided in Section I of the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.532 + ], + "angle": 0, + "content": "Extended Results. Additional examples illustrating our method's editing performance across diverse backgrounds are provided in Section H of the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.214, + 0.56 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.483, + 0.873 + ], + "angle": 0, + "content": "While our approach addresses key shortcomings in current image editing methods, it also introduces unique challenges. For example, natural camera motion sometimes appears in video sequences, which, when replicated in generated content, can lead to unintended perspective shifts. As with all generative models, video models are trained on specific data domains, making it challenging to produce results that deviate significantly from the model's training data, which predominantly includes real-world transformations. Despite these challenges, F2F demonstrates success in several cases. For example, in the middle row of Figure 1, the vase is \"magically\" filled with green water, showcasing the model's ability to perform imaginative edits, with additional examples discussed in Section H of the supplementary. Additionally, our method is computationally intensive, as transforming an image into a video sequence is resource-heavy and often slower than other image editing methods. However, video generation efficiency is advancing rapidly, as demonstrated by models like Runway Turbo\\(^{2}\\) and LTX-Video [17], which can execute the process in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.906, + 0.474 + ], + "angle": 0, + "content": "seconds, making this approach increasingly less resource-intensive. Additionally, optimizing the number of frames per edit—currently fixed at 49 for CogVideoX—could enable faster editing possibilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.486, + 0.64, + 0.501 + ], + "angle": 0, + "content": "7. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.69 + ], + "angle": 0, + "content": "We introduced Frame2Frame, a novel approach that reformulates image editing through video generation. By leveraging video models' inherent understanding of temporal transformations, our method achieves state-of-the-art editing results while maintaining high fidelity to source images. We demonstrated our framework's effectiveness on standard benchmarks, introduced PosEdit for human pose editing, and showed promising results on more classical vision tasks. As video generation technology advances, we expect our approach to enable increasingly sophisticated image manipulations while maintaining natural and physically plausible results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Future Research. Our work opens several promising directions. First, fine-tuning video generators specifically for image editing presents an exciting opportunity, including straightforward solutions like enforcing static camera scenarios or using datasets curated for editing tasks, alongside more complex approaches yet to emerge. Secondly, a key direction could involve reducing the overhead of full video generation while preserving the benefits of gradual, temporally coherent transformations, potentially enhancing both the efficiency and speed of editing. The speed can be affected both in terms of per-frame average generation speed or the number of frames generated per edit, which is currently fixed at 49 for CogVideoX." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "2Runway Turbo API" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7864" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Hadi Alzayer, Zhihao Xia, Xuaner Zhang, Eli Shechtman, Jia-Bin Huang, and Michael Gharbi. Magic fixup: Streamlining photo editing by watching dynamic videos. arXiv preprint arXiv:2403.13044, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[4] Manuel Brack, Felix Friedrich, Katharina Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.454 + ], + "angle": 0, + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[6] Chen Chen, Roozbeh Jafari, and Nasser Kehtarnavaz. Utdmhad: A multimodal dataset for human action recognition utilizing a depth camera and a wearable inertial sensor. In 2015 IEEE International conference on image processing (ICIP), pages 168-172. IEEE, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.627 + ], + "angle": 0, + "content": "[7] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320–13331, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.483, + 0.698 + ], + "angle": 0, + "content": "[8] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.701, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[9] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.744, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[14] Roy Ganz, Yair Kittenplon, Aviad Aberdam, Elad Ben Avraham, Oren Nuriel, Shai Mazor, and Ron Litman. Question aware vision transformer for multimodal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13861-13871, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[15] Yue Gao, Yuan Zhou, Jinglu Wang, Xiao Li, Xiang Ming, and Yan Lu. High-fidelity and freely controllable talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5609-5619, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[16] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.364, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[17] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[20] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[21] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12469-12478, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[22] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[23] Wonkyun Kim, Changin Choi, Wonseok Lee, and Wonjong Rhee. An image grid can be worth a video: Zero-shot video question answering using a vlm. arXiv preprint arXiv:2403.18406, 2024. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[24] Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. arXiv preprint arXiv:2412.08629, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7865" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775-6785, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Black Forest Labs. Flux.1-dev. https://github.com/black-forest-labs/FLUX.1-dev, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[28] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[29] Grace Luo, Trevor Darrell, Oliver Wang, Dan B Goldman, and Aleksander Holynski. Readout guidance: Learning control from diffusion features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8227, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[30] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.483, + 0.483 + ], + "angle": 0, + "content": "[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[33] OpenAI. Chatgpt-4o (october 2024 version). https://chat.openai.com/chat, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.636 + ], + "angle": 0, + "content": "[34] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.483, + 0.706 + ], + "angle": 0, + "content": "[35] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Noam Rotstein, David Bensaïd, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models for enriched fused image captions. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5689-5700, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[40] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[41] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.904, + 0.387 + ], + "angle": 0, + "content": "[42] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[43] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.904, + 0.472 + ], + "angle": 0, + "content": "[44] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. In NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[45] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[46] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[47] Navee Wasserman, Noam Rotstein, Roy Ganz, and Ron Kimmel. Paint by inpaint: Learning to add image objects by removing them first. arXiv preprint arXiv:2404.18212, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[48] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[49] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.829, + 0.905, + 0.898 + ], + "angle": 0, + "content": "[51] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision. Springer, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7866" + } + ] +] \ No newline at end of file diff --git a/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/f625c957-1581-449f-8cd5-463f4552272c_origin.pdf b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/f625c957-1581-449f-8cd5-463f4552272c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..66ace558f5bc6bd2399cd0defe3dacf591595004 --- /dev/null +++ b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/f625c957-1581-449f-8cd5-463f4552272c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab8189553a67ab10ba7b84ffc33bbd0f39f6b25dedfd528e477738f80ba8e1b3 +size 6711830 diff --git a/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/full.md b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae63cae3924743619b6498ca0ac1bb53f25f0284 --- /dev/null +++ b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/full.md @@ -0,0 +1,285 @@ +# Pathways on the Image Manifold: Image Editing via Video Generation + +Noam Rotstein + +Roy Velich + +Gal Yona + +David Bensaïd + +Daniel Silver + +Ron Kimmel + +# Technion - Israel Institute of Technology + +# Abstract + +Recent advances in image editing, driven by image diffusion models, have shown remarkable progress. However, significant challenges remain, as these models often struggle to follow complex edit instructions accurately and frequently compromise fidelity by altering key elements of the original image. Simultaneously, video generation has made remarkable strides, with models that effectively function as consistent and continuous world simulators. In this paper, we propose merging these two fields by utilizing image-to-video models for image editing. We reformulate image editing as a temporal process, using pretrained video models to create smooth transitions from the original image to the desired edit. This approach traverses the image manifold continuously, ensuring consistent edits while preserving the original image's key aspects. Our approach achieves state-of-the-art results on text-based image editing, demonstrating significant improvements in both edit accuracy and image preservation. Visit our project page. + +# 1. Introduction + +Image editing has witnessed remarkable advancements through deep learning and text-guided diffusion networks. These developments have set a new benchmark for image manipulations, enhancing both control and quality. However, current approaches continue to face significant limitations in real-world scenarios. These methods often struggle with two key challenges: achieving precise edits that accurately reflect the intended modifications, and preserving the essential characteristics of the original image content. + +State-of-the-art techniques predominantly rely on text-guided diffusion models, which iteratively denoise random latent representations to generate edited images. Such methods condition the generation process on both the source image—using techniques such as latent inversion [32] or model fine-tuning [22]—and the target edit description. However, these approaches require the model to generate a single output image that preserves source image fidelity while implementing complex edits, often compromises + +![](images/d302793114c47e6db147aefc107c0e6e79428b955fa321790717cf9f8d7d48b5.jpg) + +"A photo of a tomato in a blue tennis court." + +![](images/90c39c6cbe426a659ad8f14ed3c0a80814b528e1aec259b08b4e98dea56532a6.jpg) + +![](images/0b52cef901214128c3507daf9bff0aacf738c891e3d154c13f8f4ff57e74aa8e.jpg) + +"A vase of flowers filled with green water." + +![](images/e77637dc6d543fc646c2569bb9ab32c32478ad097b80948df973b214ee9b5cb6.jpg) + +![](images/a28a86e44233a98e5a4d79f19da755da1d7fa9430d912cb8f87f5c973741b207.jpg) +Figure 1. Visualization of Frame2Frame's editing process. Temporal progression of our video-based approach. Starting from the source image (leftmost), frames illustrate the natural evolution toward the target edit (rightmost). Our method produces temporally coherent intermediate states while preserving fidelity to both the source content and the editing intent. + +"A photo of a teddy bear doing pushups." + +![](images/8223c5762538660c06b6eeb47871af82f55253f16cf4607417c132e4005996e9.jpg) + +![](images/80f547005fe8c15b9d426009781fb698c9b8f6c5bad44ec593b4dfe00f2bec40.jpg) + +ing edit accuracy and content preservation. + +In this paper, we propose a paradigm shift in image editing by reformulating it as a video generation task. Rather than a single-state transition, our approach harnesses temporal coherence: the source image serves as the initial frame of a video that progressively and naturally transforms toward the target edit. This temporal evolution allows the editing process to unfold through physically plausible intermediate states, providing a continuous path between source + +and target images, as illustrated in Figure 1 and Figure S9 in the appendix. This temporal approach leverages the sophisticated world understanding embedded in recent video generation models, which have achieved breakthrough results in temporal coherence and visual quality through training on large-scale internet data [2, 48]. From a geometrical perspective, conventional editing approaches project initial noise onto the natural image manifold, targeting a single point where the image aligns with both the source and the edit request. In contrast, our approach generates a continuous path along the manifold between the original and edited image, producing a smooth realistic transition across different image states, as thoroughly discussed in Section 4. + +We implement the proposed approach through a structured pipeline called Frame2Frame (F2F). First, we transform the edit instruction into a Temporal Editing Caption — a scenario describing how the edit should naturally evolve over time—using a pretrained Vision-Language Model (VLM). Next, a state-of-the-art image-to-video model generates a temporally coherent sequence guided by the temporal caption. Finally, we identify the frame that best realizes the desired edit with the assistance of a VLM. Extensive experiments demonstrate improvements over existing imaged-to-image approaches. We evaluate on TEdBench [22] and PosEdit, a newly curated dataset derived from UTD-MHAD [6], which focuses on human pose transformations. PosEdit pairs source images with ground-truth targets of the same subject in different poses, enabling rigorous evaluation of both edit accuracy and source fidelity. Beyond commonly defined editing tasks, our framework shows promising results in more classical computer vision problems such as de-blurring, de-noising, and relighting by recasting them as temporal progressions, suggesting broader applications for video-based image transformations. + +# Our main contributions include: + +1. Reformulating image editing as a generative video task—leveraging temporal coherence to create edit paths on the natural image manifold, enabling high-fidelity manipulations while preserving source characteristics. +2. Frame2Frame: an end-to-end framework that realizes the reformulation through three key components: (1) temporal editing captions, (2) generated video-based editing, and (3) automated frame selection. +3. Comprehensive evaluation showing state-of-the-art performance on TEdBench and PosEdit, a new dataset for evaluating human pose edits. + +# 2. Related Efforts + +# 2.1. Image Editing + +Text-based image editing has advanced significantly with the success of generative diffusion models [19]. These models, in their text-to-image version, generate images through + +a denoising diffusion process conditioned on input text, relying on ground truth text-image pairs for training [37, 39]. In contrast, image editing often lacks predefined ground truth data for source and target images, posing unique challenges. This limitation has led researchers to explore diverse editing methodologies [20]. For example, SDEdit [30] injects noise into an image and then denoises it based on an editing target prompt. Imagic [22] fine-tunes a text-to-image model on a single image, subsequently interpolating between input and target text embeddings to produce edits. Other methods first invert the input image into a diffusion model's latent space [21, 32] and then generate the edited image from that latent representation using various techniques for structure preservation and manipulation [18, 34]. InstructPix2Pix [5] synthesizes an editing dataset using the approach in [18], filters it, and employs this dataset to train a diffusion model in a supervised fashion. Paint-by-Inpaint [47] further explores this supervised approach, generating a real-image dataset for object insertion. + +# 2.2. Generative Video Models + +Recent years have seen remarkable advancements in video generation, evolving from domain-specific systems [15] and brief clips to models capable of generating diverse, high-fidelity content. This progress has stemmed from paradigm shifts and the large-scale expansion of datasets [7] and architectures. Approaches have evolved from recurrent networks [16, 42] and generative adversarial networks (GANs) [9, 43, 44] to latent diffusion models (LDMs) [2, 3, 11], which leverage large U-Net or transformer-based architectures alongside vast internet-sourced datasets. Notable recent efforts include Stable Video Diffusion [2], which trains an LDM on large curated datasets, and OpenAI's Sora [28], which achieves impressive results by integrating large architectures with extensive public and private datasets. These models, often termed "world simulators" due to their emergent understanding of physical dynamics and temporal coherence. Our work builds upon CogVideoX [48], a transformer-based latent diffusion model that employs a 3D Variational Autoencoder to compress videos across both spatial and temporal dimensions, enhancing coherence. To improve text-video alignment, CogVideoX also integrates an expert transformer with adaptive LayerNorm, enabling deep fusion of visual and textual modalities. + +Close to our work, several recent efforts have utilized the world simulation capabilities of video diffusion models for various computer vision tasks. Make-A-Video3D [41] temporally extends static NeRFs using Score Distillation Sampling from video models. ViVid-1-to-3 [25] generates images along a camera trajectory around objects to enable novel view synthesis. PhysDreamer [51] models rigid object properties through 3D Gaussians and material fields, trained via distillation from pretrained video generators. + +![](images/ef0d71e99f13875a32f65951debb3c6a04bcf59a2fdcab9d4faab0bd663eb636.jpg) +Figure 2. Editing Manifold Pathway. Given an input image and target caption "A happy person making a heart shape with their hands", our method generates a continuous path on the natural image manifold. Each generated frame (indicated by black arrows) represents a plausible intermediate state between the source and target, maintaining temporal consistency throughout the transformation. As a result, in contrast to the competing approach, F2F achieves the desired edit while preserving the "AI" text on the person's shirt. + +# 2.3. Image Editing and Video + +The intersection of image editing and video has received limited attention. Existing methods focus on sampling pairs of random frames from videos to build image pair datasets that capture the same subject under varying conditions. For example, AnyDoor [8] uses the paired frames as an augmentation method, segmenting foreground objects in each frame and assigning one masked object as the target edited appearance of the subject. MagicFixup [1] employs these frames to build a dataset focused on refining user-made subject coarse 2D edits. Recently, drag-based editing approaches [29, 40], used frames extracted from video and computed optical flow to collect a dataset aimed at editing by spatially dragging points within the image. In contrast to these approaches, which rely on frame sampling for image data collection, we are the first to directly perform image editing using generative video models. + +# 3. Frame2Frame + +We present Frame2Frame, a framework that reformulates image editing as a temporal transformation process. Our approach leverages video generation models to create natural transitions between source and target images, achieving consistent and realistic edits. The proposed method has three main steps, as illustrated in Figure 3. + +# 3.1. Temporal Editing Captions + +Text-based image editing methods typically operate on two inputs: a source image $I_{s}$ and a target caption $c$ , where $c$ + +specifies the desired modifications to $I_{s}$ . Our approach differs fundamentally by modeling editing as a temporal process. This requires a novel type of prompt—the Temporal Editing Caption, denoted by $\tilde{c}$ —that describes the sequential transformation from source to target image. We construct $\tilde{c}$ by combining information from $I_{s}$ and $c$ to create a description of how the desired edit unfolds over time. + +To automate this process, we leverage recent advances in Vision-Language Models (VLMs) [10, 14, 38, 45, 46]. The VLM, specifically ChatGPT-4o [33], is instructed to produce a concise video scenario that highlights how elements within the image change or move over time. The generated caption captures the essential transformations while maintaining a static camera perspective unless movement is necessary. To improve generation quality, we utilize in-context learning (ICL) [12], providing the VLM with nine exemplar prompt-caption pairs. The complete prompt template, ICL examples, and an ablation study comparing this approach with directly using the target captions are included in Section B of the Appendix. + +# 3.2. Video Generation + +We employ CogVideoX (I2V-5B) [48], a pretrained generative video latent diffusion model utilizing a transformer-based architecture. Specifically, we use its image-to-video variant, which has been fine-tuned to generate videos starting from an input image $I_{s}$ . During generation, $I_{s}$ is encoded and concatenated with noise in latent space, where + +the model applies a denoising process guided by the temporal caption $\tilde{c}$ . As elaborated in Section 4, this conditioning allows generated videos to start from $I_{s}$ and evolve naturally along the image manifold, maintaining temporal coherence and consistency. Additionally, the model's transformer architecture enables effective fusion of visual and textual information, allowing precise control over the editing process through our temporal captions. Formally, given the video generator $G$ , we define the generation process as: + +$$ +\mathbf {G} (I _ {s}, \tilde {c}) = V = \{f _ {1}, \dots , f _ {T} \} +$$ + +where $V$ denotes the generated video with $T$ frames, and $f_{t}$ represents the frame at timestep $t$ . + +# 3.3. Frame Selection + +We observed that the optimal number of frames required for an edit can vary—small changes may be completed in fewer frames, while more extensive transformations often necessitate additional ones. Additionally, later frames tend to deviate further from the source image. Thus, even though $V$ serves as an editing path originating from $I_{s}$ , there is no guarantee that $f_{T}$ is the optimal edited image in $V$ . For instance, in Figure 3, the edited image of the cat might be expected to capture the midpoint of its jump, but the last frame shows the jump already completed. Therefore, we aim to identify the optimal edited frame, denoted $f_{t^{*}}$ , which corresponds to the earliest timestep $t$ that achieves the desired edit. The transition from the initial frame $f_{1}$ (or $I_{s}$ ) to the final edited frame $f_{t^{*}}$ motivates our method's name, Frame2Frame. + +To automate the selection of $t^*$ and avoid manual frame-by-frame review, we employ an automated approach. After generating the sequence $V$ , we sample every fourth frame, imprinting each with a unique identifier and assembling them into an image collage alongside $I_s$ . Inspired by [23], which introduces a novel approach to video comprehension by transforming videos into image grids, we use a VLM, specifically GPT-4o, to assist in selecting $t^*$ by providing it with the collage and the editing prompt $c$ . The VLM is tasked with identifying the frame that best fulfills the editing intent, evaluating each frame's alignment with $c$ and fidelity to $I_s$ . The model is instructed to select the optimal frame with the lowest index that completes the edit. An ablation study, detailed in Section C of the appendix, evaluates the effectiveness of our automated approach against the baseline of selecting the final frame of the video. + +Note that while we refer to an optimal $t^*$ , the definition of the required edit can be subjective and dependent on user preferences. For example, in Figure 3, different users may select frames of the cat jumping based on variations in its altitude. Thus, frame selection could serve as a flexible and customizable advantage in certain scenarios. + +![](images/779a5ec34b73187ef6abba799b1ec668c0bc9e424702f0ee9cb80ebcaf605dd7.jpg) +Figure 3. Frame2Frame Overview. Given a source image and editing prompt, our pipeline proceeds in three steps. First, a Vision-Language Model generates a temporal caption describing the transformation. Next, this caption guides a video generator to create a natural progression of the edit. Finally, our frame selection strategy identifies the optimal frame that best realizes the desired edit, producing the final image of the cat mid-leap. + +# 4. Editing Manifold Pathway + +To illustrate the advantages of our method over conventional image-to-image approaches, we aim to visualize the editing process within the natural image manifold, where realistic images reside. Given the high dimensionality of this space, we project it into a lower-dimensional representation suitable for interpretation. To achieve this, we first generate three sets of images, each containing 200 samples, using the text-to-image generator FLUX.1-dev [26]. The sets are based on the following prompts: + +1. AI: "Full-body portrait of a happy person wearing a shirt with the word 'AI' on it". +2. AI + Heart Hands: "Full-body portrait of a happy person wearing a shirt with the word 'AI' on it and making a heart shape with their hands". +3. Heart Hands: "Full-body portrait of a happy person making a heart shape with their hands". + +We manually filter the outcomes to ensure alignment with descriptions, with examples provided in the supplementary material in Section D. Then, we use a CLIP ViT-B/32 model [36] to extract a 512-dimensional feature vector for each image. Using these features, along with 25 feature vectors extracted from random noise images, we perform Principal Component Analysis (PCA) to reduce the dimensionality into a two-dimensional subspace. The resulting subspace, now suitable for visualization, is depicted in Figure 2. + +As illustrated in the figure, the natural images form a smooth manifold, distinctly separated from the distribution of noise images. Within this manifold, there is a clear semantic progression: images of people with 'AI' shirts (green cluster) are close to images of people with 'AI' shirts making a heart shape (purple cluster), which are adjacent to images of people only making a heart shape (red cluster). Thus, transitioning smoothly along the manifold allows a person with an 'AI' shirt to perform a heart shape with their hands while preserving the shirt's text. + +
ModelSourceTarget
LPIPS↓CLIP-I↑CLIP↑
SDEdit0.300.850.60
Pix2Pix-ZERO0.290.840.62
Imagic0.520.860.63
LEDITS++0.230.870.63
FlowEdit0.220.890.61
F2F0.220.890.63
+ +Consider an original image from the AI group that we wish to edit using the target prompt: "A happy person making a heart shape with their hands". Current editing methods generate a single image, which may cause an abrupt transition to the red cluster, effectively removing the 'AI' on the shirt. This behavior is illustrated in the figure, where the edit by LEDs++ is positioned near the red cluster, and the 'AI' text on the shirt disappears. In contrast, our method leverages video generation to perform the edit smoothly, moving along the manifold with incremental changes until reaching the required edit in the purple cluster, as indicated by black arrows in Figure 2. The temporal editing caption guiding this process is: "A happy person very slowly raising their hands to form a heart shape". This gradual progression lets us reach the purple cluster, resulting in an edited image where the person maintains the 'AI' on their shirt while making a heart shape—a faithful preservation of the original image's key attributes. This experiment illustrates that the proposed paradigm enables smooth traversal across the image manifold, enabling consistent edits while preserving the essential characteristics of the original image. + +# 5. Experiments + +We evaluate Frame2Frame against state-of-the-art image editing methods, including LEdits++ [4], SDEdit [31], Pix2Pix-Zero [35] Imagic [22], and FLUX-based FlowEdit [24]. Our evaluation spans two benchmarks: the established TEdBench [22] for general image editing, and our newly introduced PosEdit, specifically designed for human pose editing. Finally, we conduct a human evaluation to assess our method based on real user preferences. + +# 5.1. Evaluation Protocol + +We conduct our experiments following a consistent evaluation protocol across all methods and benchmarks. Following common practice [4, 22], for each method and source image, we manually select the best result from fifteen ran + +Table 1. TEdBench Results. Quantitative comparison on TEd-Bench benchmark. Source metrics (LPIPS and CLIP-I) measure content preservation, while Target metric (CLIP) measures edit accuracy. Our Frame2Frame (F2F) method achieves better or comparable performance across all metrics. + +
ModelSourceTarget
LPIPS↓CLIP-I↑LPIPS↓CLIP-I↑CLIP↑
GT (Reference)0.080.91010.61
SDEdit0.390.610.390.640.57
Pix2Pix-ZERO0.390.570.400.600.56
LEDITS++0.260.650.280.690.64
F2F0.140.820.150.840.64
+ +Table 2. PosEdit Results. Quantitative evaluation on PosEdit. Source metrics assess similarity to the original image, while Target metrics include LPIPS and CLIP-I comparisons to the ground-truth target image, along with the CLIP score for edit accuracy. "GT" provides ground-truth target image metrics for context. + +dom seeds based on visual quality and edit accuracy, ensuring the same seed set is used across all methods. For all methods, we use the default hyperparameters and settings as provided in their official implementations or official Hugging Face repositories1. + +Image Preprocessing and Generation Details. We use CogVideoX [48] as our video generation backbone, which operates at a fixed resolution of $720 \times 480$ pixels. Both TEd-Bench and PosEdit benchmarks consist of images with 1:1 aspect ratio. To accommodate CogVideoX's resolution requirements while preserving image content, we first resize all source images to $480 \times 480$ pixels, then horizontally pad them with black pixels to reach the required $720 \times 480$ resolution (adding 120 pixels on each side). After generating the video sequence, we crop the central $480 \times 480$ region of the selected frame to remove the padding and resize it to the final evaluation resolution of $512 \times 512$ pixels, matching the standard resolution used in prior work. For video generation, we adopt the default hyperparameters proposed by CogVideoX: a guidance scale of 6, 49 generated frames per sequence, and 50 denoising inference steps. This configuration generates videos approximately 6 seconds in duration at a frame rate of 8 frames per second. + +Method-Specific Requirements. Pix2Pix-Zero [35] requires additional source image descriptions along with the standard inputs. We generate these automatically using BLIP-2 [27], a state-of-the-art image captioning model, ensuring consistent source descriptions across all experiments. For our method, we transform the original editing prompts from both benchmarks into temporal editing captions, as described in Section 3.1. + +# 5.2. TEDBench Evaluation Results + +We quantitatively evaluate our method and baselines on the TEdBench benchmark using three complementary met + +![](images/1d6cff178fbd211640499817e1b9d3566be7a95b201fb664e76a553c22cd51c1.jpg) +Figure 4. Qualitative Results on TEdBench. Comparison with other methods across various editing tasks. Our approach consistently produces edits that better align with the target prompt while preserving the source image's content and structure. For instance, in the teddy bear example, our method uniquely achieves complex structural modifications while maintaining high visual quality. + +rics. For each method, we compute the metrics between the source image and its corresponding best edited version. LPIPS [50] measures perceptual similarity to assess how much the edit preserves the source image's content, with lower values indicating better preservation. CLIP-I evaluates the similarity between source and edited images in CLIP's [36] feature space, where higher values indicate better preservation of semantic content. Finally, CLIP score measures alignment between the edited image and the target prompt, with higher values indicating better adherence to the editing instruction. As shown in Table 1, our method achieves strong performance across all metrics, demonstrating effective balance between preserving source content and achieving the desired edit. + +The qualitative advantages of our method are visually evident in the comparisons presented in Figure 4. The figure demonstrates our method's superior performance across diverse editing scenarios, producing results that both faithfully execute the intended edits while maintaining strong alignment with the source image. Additional examples showcasing these capabilities are provided in the appendix in Section A. + +# 5.3. PosEdit Benchmark + +We introduce PosEdit, a benchmark for human pose editing derived from the UTD-MHAD dataset [6]. UTD-MHAD includes RGB videos of 8 subjects (4 females and 4 males) + +performing predefined actions in a controlled indoor environment. We carefully curated 58 editing tasks encompassing 8 distinct action categories, ranging from simple poses like a raised hand to complex athletic poses such as basketball shooting and lunging. + +The proposed benchmark specifically focuses on human pose manipulation. Unlike TEdBench, each editing task in PosEdit includes a ground truth frame extracted from the same subject in the target pose. The availability of reference images allows for a more comprehensive evaluation of both editing accuracy and identity preservation. In the dataset, each editing task consists of two images: a source image showing the subject in a neutral standing pose with arms relaxed at their sides, and a ground-truth edited target image capturing the subject performing in a specific pose. The benchmark also provides a prompt for each task that describes the target pose the subject should achieve (e.g., "A person in a basketball shooting posture"). + +Benchmark Evaluation. The evaluation on PosEdit follows the same protocol established for TEdBench: generating multiple variants using nine different seeds and manually selecting the best result for each method. However, PosEdit's ground-truth target images enable additional evaluation metrics. Beyond measuring preservation of source content (via LPIPS and CLIP-I between source and edited images) and edit accuracy (via CLIP score with the prompt), + +![](images/a16e8b9e7a7b6f540b789dded85b7b895e39e854787ded52815966272a019583.jpg) +Figure 5. Qualitative Results on PosEdit. Comparison between our Frame2Frame method and LEDITS++ on human motion editing tasks. For each example, we show the source image, edited results from both methods, and the ground-truth target image. Our method better preserves subject identity while achieving more natural pose transitions. The evaluation metrics for each image are provided in Section F of the appendix. + +we compute LPIPS and CLIP-I metrics between the edited image and its corresponding ground-truth target. This enhanced evaluation directly assesses both the accuracy of the transformation and preservation of identity features. Additionally, we include the evaluation metrics of the ground-truth images as a reference to the rest of the results. + +Table 2 demonstrates that our method consistently outperforms competitive methods across all metrics, with particularly strong performance in similarity measures with the target ground-truth image. This advantage highlights our method's superior ability to preserve key features while achieving the desired edit, as illustrated by the qualitative comparisons in Figure 5. The figure shows that our method generates more natural pose transitions while maintaining crucial identity attributes such as facial features, body proportions, and clothing details. + +Two important implementation notes: First, for Pix2PixZero, which requires source image descriptions, we use the same static prompt across all tasks: "A person standing naturally with his arms relaxed at his sides." Second, we exclude Imagic [22] from this comparison as their official implementation is not publicly available and their results are only reported on TEdBench. + +# 5.4. Human Evaluation Survey + +To further evaluate our method, we conducted a human survey. As indicated in Table 1, LEDs++ emerges as the most competitive method compared to ours, making it a natural choice for comparison. Each participant was presented with 20 randomly sampled TEdBench samples, including the source image, the target edit prompt, and the corresponding edited outputs from both methods. Inspired + +
MethodEdit AccuracyEdit Quality
OverallPer-ImageOverallPer-Image
F2F54.1%53.0%65.6%67.0%
LEDITS++45.9%47.0%34.4%33.0%
+ +Table 3. Human Survey Results. Human evaluation on TEd-Bench shows that F2F surpasses LEDs++ in edit accuracy while offering a significant advantage in preserving the original image. + +by the methodology proposed by [49], participants evaluated each comparison based on two criteria: (1) edit accuracy relative to the prompt, and (2) edit quality—defined as the preservation of visual fidelity to the source image, seamless integration of edited elements, and the overall natural appearance of modifications. We collected responses from 59 randomly selected online evaluators. To ensure an unbiased assessment, evaluators were not informed of the study's objectives. The results were quantified using two metrics: (i) overall global preference, expressed as a percentage, and (ii) aggregated per-image preference, where a tie resulted in each method receiving 0.5 points. + +Our results, summarized in Table 3, demonstrate that F2F outperforms LEdits++ on both metrics. Specifically: For edit quality, F2F achieved a global preference score of $53\%$ , slightly surpassing the $47\%$ obtained by LEdits++. In terms of edit accuracy, F2F achieved a $65.6\%$ global preference compared to LEdits++'s $34.4\%$ . The per-image preference results mirrored this trend, indicating robustness across individual examples and no significant influence of outliers. These findings reinforce our claim that smooth temporal editing—using video as a medium—preserves essential scene characteristics while successfully performing the edit. Our results suggest that the gap between F2F and LEdits++ in source preservation is larger than reflected by the LPIPS scores in Table 1. The full survey details are provided in the supplementary materials in Section G. + +# 5.5. Additional Vision Tasks + +We demonstrate our framework's applicability beyond traditional editing by applying it to fundamental image manipulation tasks: denoising, deblurring, outpainting, and relighting. For these experiments, we utilize Runway Gen-3 [13] as our video generation backbone, as it showed superior performance on these specific tasks compared to other generative video models. Since these tasks serve as extensions and potential applications of our method, we focus on qualitative results, leaving quantitative evaluation for future work. Our results, shown in Figure 6, demonstrate strong performance across all tasks. We associate this success to the natural alignment between these operations and common video scenarios: deblurring maps to camera focusing, outpainting to camera motion, and relighting to time-lapse lighting changes. This alignment enables our + +![](images/03566d17a72cee53af2ebdb84fd43fea620f3fb3ea5730971a110b7b521c8a7f.jpg) +Figure 6. Additional Vision Tasks. Qualitative results of our image-to-video-to-image editing approach on selected traditional tasks. + +![](images/aad66f8d257570f75d637ed2551ddda1a4b4f406f5bd3871d0faeff54939ad5f.jpg) + +![](images/8e6b3b82f01a525d6ccef2ff60cb4396a3c6ebd4862ae9c70d130d66190a69d3.jpg) + +![](images/7b246a518221c2fd0623766d7892bd1bad54b0ddaa0a182296ff36d10a0b5190.jpg) + +![](images/6c0f05f0f41dadc09a1a172347381c19404933054d0fad9774def4a87c371865.jpg) + +method to leverage the video model's learned temporal dynamics, achieving high-quality results without task-specific training. Full prompts are provided in Section I of the supplementary material. + +Extended Results. Additional examples illustrating our method's editing performance across diverse backgrounds are provided in Section H of the Appendix. + +# 6. Limitations + +While our approach addresses key shortcomings in current image editing methods, it also introduces unique challenges. For example, natural camera motion sometimes appears in video sequences, which, when replicated in generated content, can lead to unintended perspective shifts. As with all generative models, video models are trained on specific data domains, making it challenging to produce results that deviate significantly from the model's training data, which predominantly includes real-world transformations. Despite these challenges, F2F demonstrates success in several cases. For example, in the middle row of Figure 1, the vase is "magically" filled with green water, showcasing the model's ability to perform imaginative edits, with additional examples discussed in Section H of the supplementary. Additionally, our method is computationally intensive, as transforming an image into a video sequence is resource-heavy and often slower than other image editing methods. However, video generation efficiency is advancing rapidly, as demonstrated by models like Runway Turbo $^{2}$ and LTX-Video [17], which can execute the process in + +seconds, making this approach increasingly less resource-intensive. Additionally, optimizing the number of frames per edit—currently fixed at 49 for CogVideoX—could enable faster editing possibilities. + +# 7. Conclusions + +We introduced Frame2Frame, a novel approach that reformulates image editing through video generation. By leveraging video models' inherent understanding of temporal transformations, our method achieves state-of-the-art editing results while maintaining high fidelity to source images. We demonstrated our framework's effectiveness on standard benchmarks, introduced PosEdit for human pose editing, and showed promising results on more classical vision tasks. As video generation technology advances, we expect our approach to enable increasingly sophisticated image manipulations while maintaining natural and physically plausible results. + +Future Research. Our work opens several promising directions. First, fine-tuning video generators specifically for image editing presents an exciting opportunity, including straightforward solutions like enforcing static camera scenarios or using datasets curated for editing tasks, alongside more complex approaches yet to emerge. Secondly, a key direction could involve reducing the overhead of full video generation while preserving the benefits of gradual, temporally coherent transformations, potentially enhancing both the efficiency and speed of editing. The speed can be affected both in terms of per-frame average generation speed or the number of frames generated per edit, which is currently fixed at 49 for CogVideoX. + +# References + +[1] Hadi Alzayer, Zhihao Xia, Xuaner Zhang, Eli Shechtman, Jia-Bin Huang, and Michael Gharbi. Magic fixup: Streamlining photo editing by watching dynamic videos. arXiv preprint arXiv:2403.13044, 2024. 3 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2 +[4] Manuel Brack, Felix Friedrich, Katharina Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. 2023. 5 +[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2 +[6] Chen Chen, Roozbeh Jafari, and Nasser Kehtarnavaz. Utdmhad: A multimodal dataset for human action recognition utilizing a depth camera and a wearable inertial sensor. In 2015 IEEE International conference on image processing (ICIP), pages 168-172. IEEE, 2015. 2, 6 +[7] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320–13331, 2024. 2 +[8] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 3 +[9] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. In ICCV, 2019. 2 +[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3 +[11] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 2 +[12] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu + +Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 3 +[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models, 2023. 7 +[14] Roy Ganz, Yair Kittenplon, Aviad Aberdam, Elad Ben Avraham, Oren Nuriel, Shai Mazor, and Ron Litman. Question aware vision transformer for multimodal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13861-13871, 2024. 3 +[15] Yue Gao, Yuan Zhou, Jinglu Wang, Xiao Li, Xiang Ming, and Yan Lu. High-fidelity and freely controllable talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5609-5619, 2023. 2 +[16] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 2 +[17] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024.8 +[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[20] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024. 2 +[21] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12469-12478, 2024. 2 +[22] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1, 2, 5, 7 +[23] Wonkyun Kim, Changin Choi, Wonseok Lee, and Wonjong Rhee. An image grid can be worth a video: Zero-shot video question answering using a vlm. arXiv preprint arXiv:2403.18406, 2024. 4, 12 +[24] Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. arXiv preprint arXiv:2412.08629, 2024. 5 +[25] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775-6785, 2024. 2 +[26] Black Forest Labs. Flux.1-dev. https://github.com/black-forest-labs/FLUX.1-dev, 2024.4 +[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 5 +[28] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2 +[29] Grace Luo, Trevor Darrell, Oliver Wang, Dan B Goldman, and Aleksander Holynski. Readout guidance: Learning control from diffusion features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8227, 2024. 3 +[30] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2 +[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 5 +[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2 +[33] OpenAI. Chatgpt-4o (october 2024 version). https://chat.openai.com/chat, 2024.3 +[34] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2 +[35] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 5 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[38] Noam Rotstein, David Bensaïd, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models for enriched fused image captions. In Proceedings of the + +IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5689-5700, 2024. 3 +[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2 +[40] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024. 3 +[41] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 2 +[42] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning. PMLR, 2015. 2 +[43] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, 2018. 2 +[44] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. In NeurIPS, 2016. 2 +[45] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 3 +[46] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2024. 3 +[47] Navee Wasserman, Noam Rotstein, Roy Ganz, and Ron Kimmel. Paint by inpaint: Learning to add image objects by removing them first. arXiv preprint arXiv:2404.18212, 2024.2 +[48] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 3, 5 +[49] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 7, 14 +[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[51] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision. Springer, 2025. 2 \ No newline at end of file diff --git a/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/images.zip b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f7b6fbc87e82daf145efcc0bc135fa71c553d238 --- /dev/null +++ b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b72e3cc3702757d667a06cf1e937d342664cc6781258da72ff396a45275a3964 +size 568550 diff --git a/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/layout.json b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2178a6ac4621e73ec8c2d6a0046637c0d1b96532 --- /dev/null +++ b/2025/Pathways on the Image Manifold_ Image Editing via Video Generation/layout.json @@ -0,0 +1,7440 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "text", + "content": "Pathways on the Image Manifold: Image Editing via Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 184, + 144, + 260, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 144, + 260, + 156 + ], + "spans": [ + { + "bbox": [ + 184, + 144, + 260, + 156 + ], + "type": "text", + "content": "Noam Rotstein" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 158, + 240, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 158, + 240, + 171 + ], + "spans": [ + { + "bbox": [ + 184, + 158, + 240, + 171 + ], + "type": "text", + "content": "Roy Velich" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 284, + 144, + 332, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 144, + 332, + 156 + ], + "spans": [ + { + "bbox": [ + 284, + 144, + 332, + 156 + ], + "type": "text", + "content": "Gal Yona" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 264, + 158, + 337, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 158, + 337, + 171 + ], + "spans": [ + { + "bbox": [ + 264, + 158, + 337, + 171 + ], + "type": "text", + "content": "David Bensaïd" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 356, + 144, + 423, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 144, + 423, + 156 + ], + "spans": [ + { + "bbox": [ + 356, + 144, + 423, + 156 + ], + "type": "text", + "content": "Daniel Silver" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 361, + 158, + 426, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 158, + 426, + 171 + ], + "spans": [ + { + "bbox": [ + 361, + 158, + 426, + 171 + ], + "type": "text", + "content": "Ron Kimmel" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 204, + 185, + 403, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 185, + 403, + 200 + ], + "spans": [ + { + "bbox": [ + 204, + 185, + 403, + 200 + ], + "type": "text", + "content": "Technion - Israel Institute of Technology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 152, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 251, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 468 + ], + "type": "text", + "content": "Recent advances in image editing, driven by image diffusion models, have shown remarkable progress. However, significant challenges remain, as these models often struggle to follow complex edit instructions accurately and frequently compromise fidelity by altering key elements of the original image. Simultaneously, video generation has made remarkable strides, with models that effectively function as consistent and continuous world simulators. In this paper, we propose merging these two fields by utilizing image-to-video models for image editing. We reformulate image editing as a temporal process, using pretrained video models to create smooth transitions from the original image to the desired edit. This approach traverses the image manifold continuously, ensuring consistent edits while preserving the original image's key aspects. Our approach achieves state-of-the-art results on text-based image editing, demonstrating significant improvements in both edit accuracy and image preservation. Visit our project page." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 478, + 135, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 135, + 490 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 135, + 490 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "text", + "content": "Image editing has witnessed remarkable advancements through deep learning and text-guided diffusion networks. These developments have set a new benchmark for image manipulations, enhancing both control and quality. However, current approaches continue to face significant limitations in real-world scenarios. These methods often struggle with two key challenges: achieving precise edits that accurately reflect the intended modifications, and preserving the essential characteristics of the original image content." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": "State-of-the-art techniques predominantly rely on text-guided diffusion models, which iteratively denoise random latent representations to generate edited images. Such methods condition the generation process on both the source image—using techniques such as latent inversion [32] or model fine-tuning [22]—and the target edit description. However, these approaches require the model to generate a single output image that preserves source image fidelity while implementing complex edits, often compromises" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 226, + 378, + 288 + ], + "blocks": [ + { + "bbox": [ + 317, + 226, + 378, + 288 + ], + "lines": [ + { + "bbox": [ + 317, + 226, + 378, + 288 + ], + "spans": [ + { + "bbox": [ + 317, + 226, + 378, + 288 + ], + "type": "image", + "image_path": "d302793114c47e6db147aefc107c0e6e79428b955fa321790717cf9f8d7d48b5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 386, + 246, + 481, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 246, + 481, + 273 + ], + "spans": [ + { + "bbox": [ + 386, + 246, + 481, + 273 + ], + "type": "text", + "content": "\"A photo of a tomato in a blue tennis court.\"" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 488, + 226, + 549, + 287 + ], + "blocks": [ + { + "bbox": [ + 488, + 226, + 549, + 287 + ], + "lines": [ + { + "bbox": [ + 488, + 226, + 549, + 287 + ], + "spans": [ + { + "bbox": [ + 488, + 226, + 549, + 287 + ], + "type": "image", + "image_path": "90c39c6cbe426a659ad8f14ed3c0a80814b528e1aec259b08b4e98dea56532a6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 317, + 319, + 379, + 380 + ], + "blocks": [ + { + "bbox": [ + 317, + 319, + 379, + 380 + ], + "lines": [ + { + "bbox": [ + 317, + 319, + 379, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 319, + 379, + 380 + ], + "type": "image", + "image_path": "0b52cef901214128c3507daf9bff0aacf738c891e3d154c13f8f4ff57e74aa8e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 389, + 336, + 476, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 336, + 476, + 365 + ], + "spans": [ + { + "bbox": [ + 389, + 336, + 476, + 365 + ], + "type": "text", + "content": "\"A vase of flowers filled with green water.\"" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 488, + 320, + 550, + 380 + ], + "blocks": [ + { + "bbox": [ + 488, + 320, + 550, + 380 + ], + "lines": [ + { + "bbox": [ + 488, + 320, + 550, + 380 + ], + "spans": [ + { + "bbox": [ + 488, + 320, + 550, + 380 + ], + "type": "image", + "image_path": "e77637dc6d543fc646c2569bb9ab32c32478ad097b80948df973b214ee9b5cb6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 317, + 414, + 378, + 474 + ], + "blocks": [ + { + "bbox": [ + 317, + 414, + 378, + 474 + ], + "lines": [ + { + "bbox": [ + 317, + 414, + 378, + 474 + ], + "spans": [ + { + "bbox": [ + 317, + 414, + 378, + 474 + ], + "type": "image", + "image_path": "a28a86e44233a98e5a4d79f19da755da1d7fa9430d912cb8f87f5c973741b207.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 528, + 555, + 595 + ], + "lines": [ + { + "bbox": [ + 313, + 528, + 555, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 555, + 595 + ], + "type": "text", + "content": "Figure 1. Visualization of Frame2Frame's editing process. Temporal progression of our video-based approach. Starting from the source image (leftmost), frames illustrate the natural evolution toward the target edit (rightmost). Our method produces temporally coherent intermediate states while preserving fidelity to both the source content and the editing intent." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 386, + 434, + 481, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 434, + 481, + 454 + ], + "spans": [ + { + "bbox": [ + 386, + 434, + 481, + 454 + ], + "type": "text", + "content": "\"A photo of a teddy bear doing pushups.\"" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 488, + 414, + 549, + 474 + ], + "blocks": [ + { + "bbox": [ + 488, + 414, + 549, + 474 + ], + "lines": [ + { + "bbox": [ + 488, + 414, + 549, + 474 + ], + "spans": [ + { + "bbox": [ + 488, + 414, + 549, + 474 + ], + "type": "image", + "image_path": "8223c5762538660c06b6eeb47871af82f55253f16cf4607417c132e4005996e9.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 323, + 474, + 536, + 503 + ], + "blocks": [ + { + "bbox": [ + 323, + 474, + 536, + 503 + ], + "lines": [ + { + "bbox": [ + 323, + 474, + 536, + 503 + ], + "spans": [ + { + "bbox": [ + 323, + 474, + 536, + 503 + ], + "type": "image", + "image_path": "80f547005fe8c15b9d426009781fb698c9b8f6c5bad44ec593b4dfe00f2bec40.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 605, + 488, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 488, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 488, + 616 + ], + "type": "text", + "content": "ing edit accuracy and content preservation." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 618, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a paradigm shift in image editing by reformulating it as a video generation task. Rather than a single-state transition, our approach harnesses temporal coherence: the source image serves as the initial frame of a video that progressively and naturally transforms toward the target edit. This temporal evolution allows the editing process to unfold through physically plausible intermediate states, providing a continuous path between source" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7857" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "and target images, as illustrated in Figure 1 and Figure S9 in the appendix. This temporal approach leverages the sophisticated world understanding embedded in recent video generation models, which have achieved breakthrough results in temporal coherence and visual quality through training on large-scale internet data [2, 48]. From a geometrical perspective, conventional editing approaches project initial noise onto the natural image manifold, targeting a single point where the image aligns with both the source and the edit request. In contrast, our approach generates a continuous path along the manifold between the original and edited image, producing a smooth realistic transition across different image states, as thoroughly discussed in Section 4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 229, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 296, + 479 + ], + "type": "text", + "content": "We implement the proposed approach through a structured pipeline called Frame2Frame (F2F). First, we transform the edit instruction into a Temporal Editing Caption — a scenario describing how the edit should naturally evolve over time—using a pretrained Vision-Language Model (VLM). Next, a state-of-the-art image-to-video model generates a temporally coherent sequence guided by the temporal caption. Finally, we identify the frame that best realizes the desired edit with the assistance of a VLM. Extensive experiments demonstrate improvements over existing imaged-to-image approaches. We evaluate on TEdBench [22] and PosEdit, a newly curated dataset derived from UTD-MHAD [6], which focuses on human pose transformations. PosEdit pairs source images with ground-truth targets of the same subject in different poses, enabling rigorous evaluation of both edit accuracy and source fidelity. Beyond commonly defined editing tasks, our framework shows promising results in more classical computer vision problems such as de-blurring, de-noising, and relighting by recasting them as temporal progressions, suggesting broader applications for video-based image transformations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 481, + 208, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 481, + 208, + 491 + ], + "spans": [ + { + "bbox": [ + 67, + 481, + 208, + 491 + ], + "type": "text", + "content": "Our main contributions include:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 494, + 295, + 623 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 494, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 494, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 56, + 494, + 295, + 539 + ], + "type": "text", + "content": "1. Reformulating image editing as a generative video task—leveraging temporal coherence to create edit paths on the natural image manifold, enabling high-fidelity manipulations while preserving source characteristics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 541, + 294, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 541, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 541, + 294, + 586 + ], + "type": "text", + "content": "2. Frame2Frame: an end-to-end framework that realizes the reformulation through three key components: (1) temporal editing captions, (2) generated video-based editing, and (3) automated frame selection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 589, + 294, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 294, + 623 + ], + "type": "text", + "content": "3. Comprehensive evaluation showing state-of-the-art performance on TEdBench and PosEdit, a new dataset for evaluating human pose edits." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 56, + 638, + 149, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 638, + 149, + 650 + ], + "spans": [ + { + "bbox": [ + 56, + 638, + 149, + 650 + ], + "type": "text", + "content": "2. Related Efforts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 658, + 145, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 145, + 671 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 145, + 671 + ], + "type": "text", + "content": "2.1. Image Editing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Text-based image editing has advanced significantly with the success of generative diffusion models [19]. These models, in their text-to-image version, generate images through" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": "a denoising diffusion process conditioned on input text, relying on ground truth text-image pairs for training [37, 39]. In contrast, image editing often lacks predefined ground truth data for source and target images, posing unique challenges. This limitation has led researchers to explore diverse editing methodologies [20]. For example, SDEdit [30] injects noise into an image and then denoises it based on an editing target prompt. Imagic [22] fine-tunes a text-to-image model on a single image, subsequently interpolating between input and target text embeddings to produce edits. Other methods first invert the input image into a diffusion model's latent space [21, 32] and then generate the edited image from that latent representation using various techniques for structure preservation and manipulation [18, 34]. InstructPix2Pix [5] synthesizes an editing dataset using the approach in [18], filters it, and employs this dataset to train a diffusion model in a supervised fashion. Paint-by-Inpaint [47] further explores this supervised approach, generating a real-image dataset for object insertion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 310, + 455, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 310, + 455, + 322 + ], + "spans": [ + { + "bbox": [ + 314, + 310, + 455, + 322 + ], + "type": "text", + "content": "2.2. Generative Video Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 330, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 556, + 604 + ], + "type": "text", + "content": "Recent years have seen remarkable advancements in video generation, evolving from domain-specific systems [15] and brief clips to models capable of generating diverse, high-fidelity content. This progress has stemmed from paradigm shifts and the large-scale expansion of datasets [7] and architectures. Approaches have evolved from recurrent networks [16, 42] and generative adversarial networks (GANs) [9, 43, 44] to latent diffusion models (LDMs) [2, 3, 11], which leverage large U-Net or transformer-based architectures alongside vast internet-sourced datasets. Notable recent efforts include Stable Video Diffusion [2], which trains an LDM on large curated datasets, and OpenAI's Sora [28], which achieves impressive results by integrating large architectures with extensive public and private datasets. These models, often termed \"world simulators\" due to their emergent understanding of physical dynamics and temporal coherence. Our work builds upon CogVideoX [48], a transformer-based latent diffusion model that employs a 3D Variational Autoencoder to compress videos across both spatial and temporal dimensions, enhancing coherence. To improve text-video alignment, CogVideoX also integrates an expert transformer with adaptive LayerNorm, enabling deep fusion of visual and textual modalities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Close to our work, several recent efforts have utilized the world simulation capabilities of video diffusion models for various computer vision tasks. Make-A-Video3D [41] temporally extends static NeRFs using Score Distillation Sampling from video models. ViVid-1-to-3 [25] generates images along a camera trajectory around objects to enable novel view synthesis. PhysDreamer [51] models rigid object properties through 3D Gaussians and material fields, trained via distillation from pretrained video generators." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7858" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 127, + 73, + 486, + 305 + ], + "blocks": [ + { + "bbox": [ + 127, + 73, + 486, + 305 + ], + "lines": [ + { + "bbox": [ + 127, + 73, + 486, + 305 + ], + "spans": [ + { + "bbox": [ + 127, + 73, + 486, + 305 + ], + "type": "image", + "image_path": "ef0d71e99f13875a32f65951debb3c6a04bcf59a2fdcab9d4faab0bd663eb636.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 311, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 54, + 311, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 555, + 357 + ], + "type": "text", + "content": "Figure 2. Editing Manifold Pathway. Given an input image and target caption \"A happy person making a heart shape with their hands\", our method generates a continuous path on the natural image manifold. Each generated frame (indicated by black arrows) represents a plausible intermediate state between the source and target, maintaining temporal consistency throughout the transformation. As a result, in contrast to the competing approach, F2F achieves the desired edit while preserving the \"AI\" text on the person's shirt." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 368, + 195, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 195, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 195, + 380 + ], + "type": "text", + "content": "2.3. Image Editing and Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 383, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 574 + ], + "type": "text", + "content": "The intersection of image editing and video has received limited attention. Existing methods focus on sampling pairs of random frames from videos to build image pair datasets that capture the same subject under varying conditions. For example, AnyDoor [8] uses the paired frames as an augmentation method, segmenting foreground objects in each frame and assigning one masked object as the target edited appearance of the subject. MagicFixup [1] employs these frames to build a dataset focused on refining user-made subject coarse 2D edits. Recently, drag-based editing approaches [29, 40], used frames extracted from video and computed optical flow to collect a dataset aimed at editing by spatially dragging points within the image. In contrast to these approaches, which rely on frame sampling for image data collection, we are the first to directly perform image editing using generative video models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 145, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 145, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 145, + 593 + ], + "type": "text", + "content": "3. Frame2Frame" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 597, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 670 + ], + "type": "text", + "content": "We present Frame2Frame, a framework that reformulates image editing as a temporal transformation process. Our approach leverages video generation models to create natural transitions between source and target images, achieving consistent and realistic edits. The proposed method has three main steps, as illustrated in Figure 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 673, + 206, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 206, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 206, + 686 + ], + "type": "text", + "content": "3.1. Temporal Editing Captions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Text-based image editing methods typically operate on two inputs: a source image " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " and a target caption " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "c" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "text", + "content": "specifies the desired modifications to " + }, + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "text", + "content": ". Our approach differs fundamentally by modeling editing as a temporal process. This requires a novel type of prompt—the Temporal Editing Caption, denoted by " + }, + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\tilde{c}" + }, + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "text", + "content": "—that describes the sequential transformation from source to target image. We construct " + }, + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\tilde{c}" + }, + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "text", + "content": " by combining information from " + }, + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 368, + 555, + 452 + ], + "type": "text", + "content": " to create a description of how the desired edit unfolds over time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 556, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 556, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 556, + 609 + ], + "type": "text", + "content": "To automate this process, we leverage recent advances in Vision-Language Models (VLMs) [10, 14, 38, 45, 46]. The VLM, specifically ChatGPT-4o [33], is instructed to produce a concise video scenario that highlights how elements within the image change or move over time. The generated caption captures the essential transformations while maintaining a static camera perspective unless movement is necessary. To improve generation quality, we utilize in-context learning (ICL) [12], providing the VLM with nine exemplar prompt-caption pairs. The complete prompt template, ICL examples, and an ablation study comparing this approach with directly using the target captions are included in Section B of the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 621, + 421, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 421, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 421, + 633 + ], + "type": "text", + "content": "3.2. Video Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "We employ CogVideoX (I2V-5B) [48], a pretrained generative video latent diffusion model utilizing a transformer-based architecture. Specifically, we use its image-to-video variant, which has been fine-tuned to generate videos starting from an input image " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": ". During generation, " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": " is encoded and concatenated with noise in latent space, where" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7859" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "the model applies a denoising process guided by the temporal caption " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\tilde{c}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ". As elaborated in Section 4, this conditioning allows generated videos to start from " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and evolve naturally along the image manifold, maintaining temporal coherence and consistency. Additionally, the model's transformer architecture enables effective fusion of visual and textual information, allowing precise control over the editing process through our temporal captions. Formally, given the video generator " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", we define the generation process as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 103, + 190, + 227, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 190, + 227, + 204 + ], + "spans": [ + { + "bbox": [ + 103, + 190, + 227, + 204 + ], + "type": "interline_equation", + "content": "\\mathbf {G} (I _ {s}, \\tilde {c}) = V = \\{f _ {1}, \\dots , f _ {T} \\}", + "image_path": "3a6d75587b984394d4266967c3124ea79e0717faf07a6bbb3007f8c21e5cf57b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 215, + 296, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 215, + 296, + 237 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 215, + 296, + 237 + ], + "type": "text", + "content": " denotes the generated video with " + }, + { + "bbox": [ + 55, + 215, + 296, + 237 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 215, + 296, + 237 + ], + "type": "text", + "content": " frames, and " + }, + { + "bbox": [ + 55, + 215, + 296, + 237 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 215, + 296, + 237 + ], + "type": "text", + "content": " represents the frame at timestep " + }, + { + "bbox": [ + 55, + 215, + 296, + 237 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 215, + 296, + 237 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 248, + 154, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 154, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 154, + 260 + ], + "type": "text", + "content": "3.3. Frame Selection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "text", + "content": "We observed that the optimal number of frames required for an edit can vary—small changes may be completed in fewer frames, while more extensive transformations often necessitate additional ones. Additionally, later frames tend to deviate further from the source image. Thus, even though " + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "text", + "content": " serves as an editing path originating from " + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "text", + "content": ", there is no guarantee that " + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "text", + "content": " is the optimal edited image in " + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "text", + "content": ". For instance, in Figure 3, the edited image of the cat might be expected to capture the midpoint of its jump, but the last frame shows the jump already completed. Therefore, we aim to identify the optimal edited frame, denoted " + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "inline_equation", + "content": "f_{t^{*}}" + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "text", + "content": ", which corresponds to the earliest timestep " + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "text", + "content": " that achieves the desired edit. The transition from the initial frame " + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "text", + "content": ") to the final edited frame " + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "inline_equation", + "content": "f_{t^{*}}" + }, + { + "bbox": [ + 55, + 268, + 296, + 446 + ], + "type": "text", + "content": " motivates our method's name, Frame2Frame." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "text", + "content": "To automate the selection of " + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "text", + "content": " and avoid manual frame-by-frame review, we employ an automated approach. After generating the sequence " + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "text", + "content": ", we sample every fourth frame, imprinting each with a unique identifier and assembling them into an image collage alongside " + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "inline_equation", + "content": "I_s" + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "text", + "content": ". Inspired by [23], which introduces a novel approach to video comprehension by transforming videos into image grids, we use a VLM, specifically GPT-4o, to assist in selecting " + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "text", + "content": " by providing it with the collage and the editing prompt " + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "text", + "content": ". The VLM is tasked with identifying the frame that best fulfills the editing intent, evaluating each frame's alignment with " + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "text", + "content": " and fidelity to " + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "inline_equation", + "content": "I_s" + }, + { + "bbox": [ + 55, + 449, + 296, + 640 + ], + "type": "text", + "content": ". The model is instructed to select the optimal frame with the lowest index that completes the edit. An ablation study, detailed in Section C of the appendix, evaluates the effectiveness of our automated approach against the baseline of selecting the final frame of the video." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Note that while we refer to an optimal " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", the definition of the required edit can be subjective and dependent on user preferences. For example, in Figure 3, different users may select frames of the cat jumping based on variations in its altitude. Thus, frame selection could serve as a flexible and customizable advantage in certain scenarios." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 555, + 170 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 555, + 170 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 555, + 170 + ], + "type": "image", + "image_path": "779a5ec34b73187ef6abba799b1ec668c0bc9e424702f0ee9cb80ebcaf605dd7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 182, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 260 + ], + "type": "text", + "content": "Figure 3. Frame2Frame Overview. Given a source image and editing prompt, our pipeline proceeds in three steps. First, a Vision-Language Model generates a temporal caption describing the transformation. Next, this caption guides a video generator to create a natural progression of the edit. Finally, our frame selection strategy identifies the optimal frame that best realizes the desired edit, producing the final image of the cat mid-leap." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 272, + 465, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 465, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 465, + 287 + ], + "type": "text", + "content": "4. Editing Manifold Pathway" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 293, + 555, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 400 + ], + "type": "text", + "content": "To illustrate the advantages of our method over conventional image-to-image approaches, we aim to visualize the editing process within the natural image manifold, where realistic images reside. Given the high dimensionality of this space, we project it into a lower-dimensional representation suitable for interpretation. To achieve this, we first generate three sets of images, each containing 200 samples, using the text-to-image generator FLUX.1-dev [26]. The sets are based on the following prompts:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 402, + 554, + 485 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 315, + 402, + 554, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 402, + 554, + 424 + ], + "spans": [ + { + "bbox": [ + 315, + 402, + 554, + 424 + ], + "type": "text", + "content": "1. AI: \"Full-body portrait of a happy person wearing a shirt with the word 'AI' on it\"." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 426, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 554, + 460 + ], + "type": "text", + "content": "2. AI + Heart Hands: \"Full-body portrait of a happy person wearing a shirt with the word 'AI' on it and making a heart shape with their hands\"." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 461, + 553, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 461, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 314, + 461, + 553, + 485 + ], + "type": "text", + "content": "3. Heart Hands: \"Full-body portrait of a happy person making a heart shape with their hands\"." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 486, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 593 + ], + "type": "text", + "content": "We manually filter the outcomes to ensure alignment with descriptions, with examples provided in the supplementary material in Section D. Then, we use a CLIP ViT-B/32 model [36] to extract a 512-dimensional feature vector for each image. Using these features, along with 25 feature vectors extracted from random noise images, we perform Principal Component Analysis (PCA) to reduce the dimensionality into a two-dimensional subspace. The resulting subspace, now suitable for visualization, is depicted in Figure 2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "As illustrated in the figure, the natural images form a smooth manifold, distinctly separated from the distribution of noise images. Within this manifold, there is a clear semantic progression: images of people with 'AI' shirts (green cluster) are close to images of people with 'AI' shirts making a heart shape (purple cluster), which are adjacent to images of people only making a heart shape (red cluster). Thus, transitioning smoothly along the manifold allows a person with an 'AI' shirt to perform a heart shape with their hands while preserving the shirt's text." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7860" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 70, + 277, + 188 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 277, + 188 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 277, + 188 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 277, + 188 + ], + "type": "table", + "html": "
ModelSourceTarget
LPIPS↓CLIP-I↑CLIP↑
SDEdit0.300.850.60
Pix2Pix-ZERO0.290.840.62
Imagic0.520.860.63
LEDITS++0.230.870.63
FlowEdit0.220.890.61
F2F0.220.890.63
", + "image_path": "96618022732d534bfd468ad7ceee43f6bb1e3ce0f5920fa40a49388dcf4df173.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 266, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 296, + 517 + ], + "type": "text", + "content": "Consider an original image from the AI group that we wish to edit using the target prompt: \"A happy person making a heart shape with their hands\". Current editing methods generate a single image, which may cause an abrupt transition to the red cluster, effectively removing the 'AI' on the shirt. This behavior is illustrated in the figure, where the edit by LEDs++ is positioned near the red cluster, and the 'AI' text on the shirt disappears. In contrast, our method leverages video generation to perform the edit smoothly, moving along the manifold with incremental changes until reaching the required edit in the purple cluster, as indicated by black arrows in Figure 2. The temporal editing caption guiding this process is: \"A happy person very slowly raising their hands to form a heart shape\". This gradual progression lets us reach the purple cluster, resulting in an edited image where the person maintains the 'AI' on their shirt while making a heart shape—a faithful preservation of the original image's key attributes. This experiment illustrates that the proposed paradigm enables smooth traversal across the image manifold, enabling consistent edits while preserving the essential characteristics of the original image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 525, + 137, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 137, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 137, + 539 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 639 + ], + "type": "text", + "content": "We evaluate Frame2Frame against state-of-the-art image editing methods, including LEdits++ [4], SDEdit [31], Pix2Pix-Zero [35] Imagic [22], and FLUX-based FlowEdit [24]. Our evaluation spans two benchmarks: the established TEdBench [22] for general image editing, and our newly introduced PosEdit, specifically designed for human pose editing. Finally, we conduct a human evaluation to assess our method based on real user preferences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 171, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 171, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 171, + 659 + ], + "type": "text", + "content": "5.1. Evaluation Protocol" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "We conduct our experiments following a consistent evaluation protocol across all methods and benchmarks. Following common practice [4, 22], for each method and source image, we manually select the best result from fifteen ran" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 554, + 182 + ], + "blocks": [ + { + "bbox": [ + 55, + 197, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 295, + 251 + ], + "type": "text", + "content": "Table 1. TEdBench Results. Quantitative comparison on TEd-Bench benchmark. Source metrics (LPIPS and CLIP-I) measure content preservation, while Target metric (CLIP) measures edit accuracy. Our Frame2Frame (F2F) method achieves better or comparable performance across all metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 554, + 182 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 554, + 182 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 554, + 182 + ], + "type": "table", + "html": "
ModelSourceTarget
LPIPS↓CLIP-I↑LPIPS↓CLIP-I↑CLIP↑
GT (Reference)0.080.91010.61
SDEdit0.390.610.390.640.57
Pix2Pix-ZERO0.390.570.400.600.56
LEDITS++0.260.650.280.690.64
F2F0.140.820.150.840.64
", + "image_path": "090da99a4ebcec0f91b7163fea07b6b4a6424f87ea73d2873352e6aee0ca4c6c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 190, + 553, + 246 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 553, + 246 + ], + "type": "text", + "content": "Table 2. PosEdit Results. Quantitative evaluation on PosEdit. Source metrics assess similarity to the original image, while Target metrics include LPIPS and CLIP-I comparisons to the ground-truth target image, along with the CLIP score for edit accuracy. \"GT\" provides ground-truth target image metrics for context." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 257, + 554, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 554, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 554, + 318 + ], + "type": "text", + "content": "dom seeds based on visual quality and edit accuracy, ensuring the same seed set is used across all methods. For all methods, we use the default hyperparameters and settings as provided in their official implementations or official Hugging Face repositories1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "text", + "content": "Image Preprocessing and Generation Details. We use CogVideoX [48] as our video generation backbone, which operates at a fixed resolution of " + }, + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "inline_equation", + "content": "720 \\times 480" + }, + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "text", + "content": " pixels. Both TEd-Bench and PosEdit benchmarks consist of images with 1:1 aspect ratio. To accommodate CogVideoX's resolution requirements while preserving image content, we first resize all source images to " + }, + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "inline_equation", + "content": "480 \\times 480" + }, + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "text", + "content": " pixels, then horizontally pad them with black pixels to reach the required " + }, + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "inline_equation", + "content": "720 \\times 480" + }, + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "text", + "content": " resolution (adding 120 pixels on each side). After generating the video sequence, we crop the central " + }, + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "inline_equation", + "content": "480 \\times 480" + }, + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "text", + "content": " region of the selected frame to remove the padding and resize it to the final evaluation resolution of " + }, + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 327, + 555, + 543 + ], + "type": "text", + "content": " pixels, matching the standard resolution used in prior work. For video generation, we adopt the default hyperparameters proposed by CogVideoX: a guidance scale of 6, 49 generated frames per sequence, and 50 denoising inference steps. This configuration generates videos approximately 6 seconds in duration at a frame rate of 8 frames per second." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 556, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 652 + ], + "type": "text", + "content": "Method-Specific Requirements. Pix2Pix-Zero [35] requires additional source image descriptions along with the standard inputs. We generate these automatically using BLIP-2 [27], a state-of-the-art image captioning model, ensuring consistent source descriptions across all experiments. For our method, we transform the original editing prompts from both benchmarks into temporal editing captions, as described in Section 3.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 656, + 479, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 479, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 479, + 668 + ], + "type": "text", + "content": "5.2. TEDBench Evaluation Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 671, + 553, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 553, + 696 + ], + "type": "text", + "content": "We quantitatively evaluate our method and baselines on the TEdBench benchmark using three complementary met" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 436, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 436, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 436, + 712 + ], + "type": "text", + "content": "1LEdits++,Pix2Pix-Zero,SDEdit" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7861" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 77, + 523, + 339 + ], + "blocks": [ + { + "bbox": [ + 89, + 77, + 523, + 339 + ], + "lines": [ + { + "bbox": [ + 89, + 77, + 523, + 339 + ], + "spans": [ + { + "bbox": [ + 89, + 77, + 523, + 339 + ], + "type": "image", + "image_path": "1d6cff178fbd211640499817e1b9d3566be7a95b201fb664e76a553c22cd51c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 346, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 555, + 381 + ], + "type": "text", + "content": "Figure 4. Qualitative Results on TEdBench. Comparison with other methods across various editing tasks. Our approach consistently produces edits that better align with the target prompt while preserving the source image's content and structure. For instance, in the teddy bear example, our method uniquely achieves complex structural modifications while maintaining high visual quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 392, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 392, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 54, + 392, + 295, + 559 + ], + "type": "text", + "content": "rics. For each method, we compute the metrics between the source image and its corresponding best edited version. LPIPS [50] measures perceptual similarity to assess how much the edit preserves the source image's content, with lower values indicating better preservation. CLIP-I evaluates the similarity between source and edited images in CLIP's [36] feature space, where higher values indicate better preservation of semantic content. Finally, CLIP score measures alignment between the edited image and the target prompt, with higher values indicating better adherence to the editing instruction. As shown in Table 1, our method achieves strong performance across all metrics, demonstrating effective balance between preserving source content and achieving the desired edit." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 559, + 295, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 656 + ], + "type": "text", + "content": "The qualitative advantages of our method are visually evident in the comparisons presented in Figure 4. The figure demonstrates our method's superior performance across diverse editing scenarios, producing results that both faithfully execute the intended edits while maintaining strong alignment with the source image. Additional examples showcasing these capabilities are provided in the appendix in Section A." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 662, + 173, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 173, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 173, + 673 + ], + "type": "text", + "content": "5.3. PosEdit Benchmark" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "We introduce PosEdit, a benchmark for human pose editing derived from the UTD-MHAD dataset [6]. UTD-MHAD includes RGB videos of 8 subjects (4 females and 4 males)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 392, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 452 + ], + "type": "text", + "content": "performing predefined actions in a controlled indoor environment. We carefully curated 58 editing tasks encompassing 8 distinct action categories, ranging from simple poses like a raised hand to complex athletic poses such as basketball shooting and lunging." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 452, + 556, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 556, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 556, + 608 + ], + "type": "text", + "content": "The proposed benchmark specifically focuses on human pose manipulation. Unlike TEdBench, each editing task in PosEdit includes a ground truth frame extracted from the same subject in the target pose. The availability of reference images allows for a more comprehensive evaluation of both editing accuracy and identity preservation. In the dataset, each editing task consists of two images: a source image showing the subject in a neutral standing pose with arms relaxed at their sides, and a ground-truth edited target image capturing the subject performing in a specific pose. The benchmark also provides a prompt for each task that describes the target pose the subject should achieve (e.g., \"A person in a basketball shooting posture\")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Benchmark Evaluation. The evaluation on PosEdit follows the same protocol established for TEdBench: generating multiple variants using nine different seeds and manually selecting the best result for each method. However, PosEdit's ground-truth target images enable additional evaluation metrics. Beyond measuring preservation of source content (via LPIPS and CLIP-I between source and edited images) and edit accuracy (via CLIP score with the prompt)," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7862" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 74, + 294, + 223 + ], + "blocks": [ + { + "bbox": [ + 62, + 74, + 294, + 223 + ], + "lines": [ + { + "bbox": [ + 62, + 74, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 74, + 294, + 223 + ], + "type": "image", + "image_path": "a16e8b9e7a7b6f540b789dded85b7b895e39e854787ded52815966272a019583.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 296, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 314 + ], + "type": "text", + "content": "Figure 5. Qualitative Results on PosEdit. Comparison between our Frame2Frame method and LEDITS++ on human motion editing tasks. For each example, we show the source image, edited results from both methods, and the ground-truth target image. Our method better preserves subject identity while achieving more natural pose transitions. The evaluation metrics for each image are provided in Section F of the appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 396 + ], + "type": "text", + "content": "we compute LPIPS and CLIP-I metrics between the edited image and its corresponding ground-truth target. This enhanced evaluation directly assesses both the accuracy of the transformation and preservation of identity features. Additionally, we include the evaluation metrics of the ground-truth images as a reference to the rest of the results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 398, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 517 + ], + "type": "text", + "content": "Table 2 demonstrates that our method consistently outperforms competitive methods across all metrics, with particularly strong performance in similarity measures with the target ground-truth image. This advantage highlights our method's superior ability to preserve key features while achieving the desired edit, as illustrated by the qualitative comparisons in Figure 5. The figure shows that our method generates more natural pose transitions while maintaining crucial identity attributes such as facial features, body proportions, and clothing details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 518, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 601 + ], + "type": "text", + "content": "Two important implementation notes: First, for Pix2PixZero, which requires source image descriptions, we use the same static prompt across all tasks: \"A person standing naturally with his arms relaxed at his sides.\" Second, we exclude Imagic [22] from this comparison as their official implementation is not publicly available and their results are only reported on TEdBench." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 610, + 203, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 203, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 203, + 624 + ], + "type": "text", + "content": "5.4. Human Evaluation Survey" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "To further evaluate our method, we conducted a human survey. As indicated in Table 1, LEDs++ emerges as the most competitive method compared to ours, making it a natural choice for comparison. Each participant was presented with 20 randomly sampled TEdBench samples, including the source image, the target edit prompt, and the corresponding edited outputs from both methods. Inspired" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 552, + 133 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 552, + 133 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 552, + 133 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 552, + 133 + ], + "type": "table", + "html": "
MethodEdit AccuracyEdit Quality
OverallPer-ImageOverallPer-Image
F2F54.1%53.0%65.6%67.0%
LEDITS++45.9%47.0%34.4%33.0%
", + "image_path": "9b314e0e3f74bbc4a20c5434f58416446248a32c4ef086eaae8a587b889c00a1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 140, + 555, + 174 + ], + "lines": [ + { + "bbox": [ + 313, + 140, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 555, + 174 + ], + "type": "text", + "content": "Table 3. Human Survey Results. Human evaluation on TEd-Bench shows that F2F surpasses LEDs++ in edit accuracy while offering a significant advantage in preserving the original image." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 185, + 555, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 555, + 328 + ], + "type": "text", + "content": "by the methodology proposed by [49], participants evaluated each comparison based on two criteria: (1) edit accuracy relative to the prompt, and (2) edit quality—defined as the preservation of visual fidelity to the source image, seamless integration of edited elements, and the overall natural appearance of modifications. We collected responses from 59 randomly selected online evaluators. To ensure an unbiased assessment, evaluators were not informed of the study's objectives. The results were quantified using two metrics: (i) overall global preference, expressed as a percentage, and (ii) aggregated per-image preference, where a tie resulted in each method receiving 0.5 points." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 329, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 556, + 510 + ], + "type": "text", + "content": "Our results, summarized in Table 3, demonstrate that F2F outperforms LEdits++ on both metrics. Specifically: For edit quality, F2F achieved a global preference score of " + }, + { + "bbox": [ + 313, + 329, + 556, + 510 + ], + "type": "inline_equation", + "content": "53\\%" + }, + { + "bbox": [ + 313, + 329, + 556, + 510 + ], + "type": "text", + "content": ", slightly surpassing the " + }, + { + "bbox": [ + 313, + 329, + 556, + 510 + ], + "type": "inline_equation", + "content": "47\\%" + }, + { + "bbox": [ + 313, + 329, + 556, + 510 + ], + "type": "text", + "content": " obtained by LEdits++. In terms of edit accuracy, F2F achieved a " + }, + { + "bbox": [ + 313, + 329, + 556, + 510 + ], + "type": "inline_equation", + "content": "65.6\\%" + }, + { + "bbox": [ + 313, + 329, + 556, + 510 + ], + "type": "text", + "content": " global preference compared to LEdits++'s " + }, + { + "bbox": [ + 313, + 329, + 556, + 510 + ], + "type": "inline_equation", + "content": "34.4\\%" + }, + { + "bbox": [ + 313, + 329, + 556, + 510 + ], + "type": "text", + "content": ". The per-image preference results mirrored this trend, indicating robustness across individual examples and no significant influence of outliers. These findings reinforce our claim that smooth temporal editing—using video as a medium—preserves essential scene characteristics while successfully performing the edit. Our results suggest that the gap between F2F and LEdits++ in source preservation is larger than reflected by the LPIPS scores in Table 1. The full survey details are provided in the supplementary materials in Section G." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 515, + 449, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 449, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 449, + 528 + ], + "type": "text", + "content": "5.5. Additional Vision Tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "We demonstrate our framework's applicability beyond traditional editing by applying it to fundamental image manipulation tasks: denoising, deblurring, outpainting, and relighting. For these experiments, we utilize Runway Gen-3 [13] as our video generation backbone, as it showed superior performance on these specific tasks compared to other generative video models. Since these tasks serve as extensions and potential applications of our method, we focus on qualitative results, leaving quantitative evaluation for future work. Our results, shown in Figure 6, demonstrate strong performance across all tasks. We associate this success to the natural alignment between these operations and common video scenarios: deblurring maps to camera focusing, outpainting to camera motion, and relighting to time-lapse lighting changes. This alignment enables our" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7863" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 74, + 126, + 286 + ], + "blocks": [ + { + "bbox": [ + 57, + 74, + 126, + 286 + ], + "lines": [ + { + "bbox": [ + 57, + 74, + 126, + 286 + ], + "spans": [ + { + "bbox": [ + 57, + 74, + 126, + 286 + ], + "type": "image", + "image_path": "03566d17a72cee53af2ebdb84fd43fea620f3fb3ea5730971a110b7b521c8a7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 295, + 546, + 306 + ], + "lines": [ + { + "bbox": [ + 62, + 295, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 546, + 306 + ], + "type": "text", + "content": "Figure 6. Additional Vision Tasks. Qualitative results of our image-to-video-to-image editing approach on selected traditional tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 138, + 74, + 272, + 286 + ], + "blocks": [ + { + "bbox": [ + 138, + 74, + 272, + 286 + ], + "lines": [ + { + "bbox": [ + 138, + 74, + 272, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 74, + 272, + 286 + ], + "type": "image", + "image_path": "aad66f8d257570f75d637ed2551ddda1a4b4f406f5bd3871d0faeff54939ad5f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 276, + 74, + 410, + 286 + ], + "blocks": [ + { + "bbox": [ + 276, + 74, + 410, + 286 + ], + "lines": [ + { + "bbox": [ + 276, + 74, + 410, + 286 + ], + "spans": [ + { + "bbox": [ + 276, + 74, + 410, + 286 + ], + "type": "image", + "image_path": "8e6b3b82f01a525d6ccef2ff60cb4396a3c6ebd4862ae9c70d130d66190a69d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 413, + 74, + 482, + 286 + ], + "blocks": [ + { + "bbox": [ + 413, + 74, + 482, + 286 + ], + "lines": [ + { + "bbox": [ + 413, + 74, + 482, + 286 + ], + "spans": [ + { + "bbox": [ + 413, + 74, + 482, + 286 + ], + "type": "image", + "image_path": "7b246a518221c2fd0623766d7892bd1bad54b0ddaa0a182296ff36d10a0b5190.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 485, + 74, + 553, + 286 + ], + "blocks": [ + { + "bbox": [ + 485, + 74, + 553, + 286 + ], + "lines": [ + { + "bbox": [ + 485, + 74, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 485, + 74, + 553, + 286 + ], + "type": "image", + "image_path": "6c0f05f0f41dadc09a1a172347381c19404933054d0fad9774def4a87c371865.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 328, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 376 + ], + "type": "text", + "content": "method to leverage the video model's learned temporal dynamics, achieving high-quality results without task-specific training. Full prompts are provided in Section I of the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 384, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 421 + ], + "type": "text", + "content": "Extended Results. Additional examples illustrating our method's editing performance across diverse backgrounds are provided in Section H of the Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 431, + 130, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 130, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 130, + 443 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 453, + 295, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 691 + ], + "type": "text", + "content": "While our approach addresses key shortcomings in current image editing methods, it also introduces unique challenges. For example, natural camera motion sometimes appears in video sequences, which, when replicated in generated content, can lead to unintended perspective shifts. As with all generative models, video models are trained on specific data domains, making it challenging to produce results that deviate significantly from the model's training data, which predominantly includes real-world transformations. Despite these challenges, F2F demonstrates success in several cases. For example, in the middle row of Figure 1, the vase is \"magically\" filled with green water, showcasing the model's ability to perform imaginative edits, with additional examples discussed in Section H of the supplementary. Additionally, our method is computationally intensive, as transforming an image into a video sequence is resource-heavy and often slower than other image editing methods. However, video generation efficiency is advancing rapidly, as demonstrated by models like Runway Turbo" + }, + { + "bbox": [ + 55, + 453, + 295, + 691 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 55, + 453, + 295, + 691 + ], + "type": "text", + "content": " and LTX-Video [17], which can execute the process in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 328, + 554, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 554, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 554, + 375 + ], + "type": "text", + "content": "seconds, making this approach increasingly less resource-intensive. Additionally, optimizing the number of frames per edit—currently fixed at 49 for CogVideoX—could enable faster editing possibilities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 391, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 391, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 391, + 396 + ], + "type": "text", + "content": "7. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 403, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 546 + ], + "type": "text", + "content": "We introduced Frame2Frame, a novel approach that reformulates image editing through video generation. By leveraging video models' inherent understanding of temporal transformations, our method achieves state-of-the-art editing results while maintaining high fidelity to source images. We demonstrated our framework's effectiveness on standard benchmarks, introduced PosEdit for human pose editing, and showed promising results on more classical vision tasks. As video generation technology advances, we expect our approach to enable increasingly sophisticated image manipulations while maintaining natural and physically plausible results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Future Research. Our work opens several promising directions. First, fine-tuning video generators specifically for image editing presents an exciting opportunity, including straightforward solutions like enforcing static camera scenarios or using datasets curated for editing tasks, alongside more complex approaches yet to emerge. Secondly, a key direction could involve reducing the overhead of full video generation while preserving the benefits of gradual, temporally coherent transformations, potentially enhancing both the efficiency and speed of editing. The speed can be affected both in terms of per-frame average generation speed or the number of frames generated per edit, which is currently fixed at 49 for CogVideoX." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 135, + 712 + ], + "type": "text", + "content": "2Runway Turbo API" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7864" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Hadi Alzayer, Zhihao Xia, Xuaner Zhang, Eli Shechtman, Jia-Bin Huang, and Michael Gharbi. Magic fixup: Streamlining photo editing by watching dynamic videos. arXiv preprint arXiv:2403.13044, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 258 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 304 + ], + "type": "text", + "content": "[4] Manuel Brack, Felix Friedrich, Katharina Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 295, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 359 + ], + "type": "text", + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 363, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 295, + 418 + ], + "type": "text", + "content": "[6] Chen Chen, Roozbeh Jafari, and Nasser Kehtarnavaz. Utdmhad: A multimodal dataset for human action recognition utilizing a depth camera and a wearable inertial sensor. In 2015 IEEE International conference on image processing (ICIP), pages 168-172. IEEE, 2015. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 295, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 496 + ], + "type": "text", + "content": "[7] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320–13331, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 498, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 295, + 552 + ], + "type": "text", + "content": "[8] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 555, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 295, + 586 + ], + "type": "text", + "content": "[9] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. In ICCV, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 589, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 589, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 589, + 295, + 643 + ], + "type": "text", + "content": "[10] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "text", + "content": "[11] Duolikun Danier, Fan Zhang, and David Bull. Ldmvfi: Video frame interpolation with latent diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[12] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "type": "text", + "content": "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models, 2023. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "type": "text", + "content": "[14] Roy Ganz, Yair Kittenplon, Aviad Aberdam, Elad Ben Avraham, Oren Nuriel, Shai Mazor, and Ron Litman. Question aware vision transformer for multimodal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13861-13871, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[15] Yue Gao, Yuan Zhou, Jinglu Wang, Xiao Li, Xiang Ming, and Yan Lu. High-fidelity and freely controllable talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5609-5619, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "type": "text", + "content": "[16] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 288, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 288, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 288, + 553, + 341 + ], + "type": "text", + "content": "[17] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024.8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 387 + ], + "type": "text", + "content": "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 466 + ], + "type": "text", + "content": "[20] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "text", + "content": "[21] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12469-12478, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 523, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 588 + ], + "type": "text", + "content": "[22] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 1, 2, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "text", + "content": "[23] Wonkyun Kim, Changin Choi, Wonseok Lee, and Wonjong Rhee. An image grid can be worth a video: Zero-shot video question answering using a vlm. arXiv preprint arXiv:2403.18406, 2024. 4, 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "text", + "content": "[24] Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Flowedit: Inversion-free text-based editing using pre-trained flow models. arXiv preprint arXiv:2412.08629, 2024. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "type": "text", + "content": "[25] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7865" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775-6785, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 294, + 116 + ], + "type": "text", + "content": "[26] Black Forest Labs. Flux.1-dev. https://github.com/black-forest-labs/FLUX.1-dev, 2024.4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 172, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 295, + 228 + ], + "type": "text", + "content": "[28] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 282 + ], + "type": "text", + "content": "[29] Grace Luo, Trevor Darrell, Oliver Wang, Dan B Goldman, and Aleksander Holynski. Readout guidance: Learning control from diffusion features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8227, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 295, + 327 + ], + "type": "text", + "content": "[30] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 295, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 382 + ], + "type": "text", + "content": "[31] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In International Conference on Learning Representations, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 437 + ], + "type": "text", + "content": "[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "type": "text", + "content": "[33] OpenAI. Chatgpt-4o (october 2024 version). https://chat.openai.com/chat, 2024.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "type": "text", + "content": "[34] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 295, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 295, + 559 + ], + "type": "text", + "content": "[35] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 624 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[38] Noam Rotstein, David Bensaïd, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models for enriched fused image captions. In Proceedings of the" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5689-5700, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "text", + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[40] Yujun Shi, Jun Hao Liew, Hanshu Yan, Vincent YF Tan, and Jiashi Feng. Instadrag: Lightning fast and accurate drag-based image editing emerging from videos. arXiv preprint arXiv:2405.13722, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "type": "text", + "content": "[41] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 262, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 306 + ], + "type": "text", + "content": "[42] Nitish Srivastava, Elman Mansimov, and Ruslan Salakhudinov. Unsupervised learning of video representations using lstms. In International conference on machine learning. PMLR, 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 342 + ], + "type": "text", + "content": "[43] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "text", + "content": "[44] Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Generating videos with scene dynamics. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "text", + "content": "[45] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "text", + "content": "[46] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 476, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 520 + ], + "type": "text", + "content": "[47] Navee Wasserman, Noam Rotstein, Roy Ganz, and Ron Kimmel. Paint by inpaint: Learning to add image objects by removing them first. arXiv preprint arXiv:2404.18212, 2024.2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 521, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 576 + ], + "type": "text", + "content": "[48] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 577, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 621 + ], + "type": "text", + "content": "[49] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 7, 14" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 622, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 655 + ], + "type": "text", + "content": "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 656, + 553, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 656, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 316, + 656, + 553, + 711 + ], + "type": "text", + "content": "[51] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision. Springer, 2025. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7866" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/b8f485b4-2d7f-474b-acf2-8dc4a4e3047c_content_list.json b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/b8f485b4-2d7f-474b-acf2-8dc4a4e3047c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e7e39b22062428c8997680feed0f03db264e3dc1 --- /dev/null +++ b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/b8f485b4-2d7f-474b-acf2-8dc4a4e3047c_content_list.json @@ -0,0 +1,1591 @@ +[ + { + "type": "text", + "text": "Patient-Level Anatomy Meets Scanning-Level Physics: Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model", + "text_level": 1, + "bbox": [ + 102, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyuan Yang $^{1,5}$ , Yingyu Chen $^{1,5}$ , Zhiwen Wang $^{2}$ , Hongming Shan $^{3}$ , Yang Chen $^{4}$ , Yi Zhang $^{1,5,*}$ $^{1}$ Sichuan University $^{2}$ Southwest Petroleum University $^{3}$ Fudan University $^{4}$ Southeast University $^{5}$ Tianfu Jiangxi Laboratory \n{cziyuanyang, cyy262511, zwwang1228}@gmail.com, hmshan@fudan.edu.cn \nchenyang.list@seu.edu.cn, yzhang@scu.edu.cn", + "bbox": [ + 125, + 202, + 870, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 340, + 326, + 357 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reducing radiation doses benefits patients, but the resultant low-dose computed tomography (LDCT) images often suffer from clinically unacceptable noise and artifacts. While deep learning (DL) has shown promise in LDCT reconstruction, it requires large-scale data collection from multiple clients, raising privacy concerns. Federated learning (FL) has been introduced to mitigate these privacy concerns; however, current methods are typically tailored to specific scanning protocols, which limits their generalizability and makes them less effective for unseen protocols. To address these issues, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Since the noise distribution in LDCT data is closely tied to scanning protocols and anatomical structures, we propose a dual-level physics-informed way to address these challenges. Specifically, we incorporate physical and anatomical prompts into our physics-informed hypernetworks to capture scanning- and anatomy-specific information, enabling dual-level physics-driven personalization of imaging features. These prompts are derived from the scanning protocol and the radiology report generated by a medical large language model (MLLM). Subsequently, client-specific decoders project these dual-level personalized imaging features back into the image domain. Besides, to tackle the challenge of unseen data, we introduce a novel protocol vector-quantization strategy (PVQS), which ensures consistent performance across new clients by quantifying unseen scanning codes to the closest match in the scanning codebook. Extensive experimental results demonstrate the superior performance of SCAN-PhysFed on pub", + "bbox": [ + 88, + 385, + 485, + 856 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lic datasets1.", + "bbox": [ + 513, + 327, + 602, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 372, + 643, + 387 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computed tomography (CT) is a key diagnostic tool that non-invasively visualizes anatomical structures within the human body. Despite its clinical benefits, CT scans pose radiation concerns, which potentially lead to genetic, cancerous, and other health diseases. Lowering CT radiation dose—either by decreasing incident photons or limiting sampling views—has emerged as a promising solution to reduce radiation risks and accelerate scanning. However, reconstructed images under these conditions suffer from significant quality degradation, compromising clinical applicability [27].", + "bbox": [ + 511, + 397, + 906, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, deep learning (DL)-based methods have shown great promise for low-dose CT (LDCT) imaging [28]. However, most DL-based LDCT reconstruction methods are condition-specific and overlook privacy concerns. Recently, researchers have attempted to introduce federated learning (FL), a privacy-preserving distributed learning paradigm, for LDCT reconstruction to alleviate privacy concerns. Although these methods aim to enhance generalization, they still face significant challenges arising from data heterogeneity, caused by diverse noise distributions in CT imaging across different protocols and anatomical structures. We believe that the further advancement of FL in LDCT reconstruction urgently requires a new framework capable of leveraging comprehensive physical information from the scanning process, thereby breaking through current bottlenecks to achieve better results.", + "bbox": [ + 511, + 564, + 908, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this issue, we aim to leverage physical principles of the imaging process. During the projection process, the measurement is influenced by the anatomical structures the X-ray passed through, such as fat, bone, and fluid, each", + "bbox": [ + 511, + 805, + 908, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 120, + 887, + 250, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 Our code has been released at https://github.com/Zi-YuanYang/SCAN-PhysFed.", + "bbox": [ + 514, + 875, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5154", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/60200b64a90c355528f48b7d4824df9a108087adeb73fc4b6e5254a5d9bc00eb.jpg", + "image_caption": [ + "Figure 1. Illustration of different radiation reduction strategies. 'Sparse-View' refers to scanning with a reduced number of projection views, while 'Low-Dose' represents scanning with a lower number of incident photons." + ], + "image_footnote": [], + "bbox": [ + 101, + 95, + 470, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with distinct attenuation coefficients. Additionally, the system matrix is significantly affected by scanning protocols, including tube current, voltage settings, and scanning angles. These factors collectively impact the quality and characteristics of reconstructed CT images during the backprojection phase, resulting in varying levels of noise and artifacts. An example is provided in Figure 1 to illustrate this concept more clearly.", + "bbox": [ + 89, + 323, + 483, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By leveraging the physics-specific knowledge discussed earlier, we reimagine LDCT reconstruction within the federated learning framework. Our approach adopts a dual-level strategy to alleviate heterogeneity, addressing variations at both the scanning and patient levels. In practice, scanning protocols, which serve as scanner hyperparameters, are readily available. However, generating anatomical information has traditionally been challenging due to patient variability and the complexity of accurately modeling different tissue types. Recent advances in large language models (LLMs) offer new possibilities for automatically generating anatomical information, potentially improving the reconstruction.", + "bbox": [ + 89, + 446, + 483, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on this, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Specifically, we address the data heterogeneous using a dual-level strategy. A detailed radiology report is generated using a pretrained LLM. Then, a patient-level anatomy-informed hypernetwork is designed to generate a modulation map from the radiology report, which seamlessly integrates with imaging features and aligns them with each patient's unique anatomical structures. Simultaneously, scanning features are integrated via a scanning-informed hypernetwork to personalize imaging features based on the protocol. To ensure distinguishability of scanning feature vectors, we impose an orthogonal loss during training to reinforce personalization. This dual-modulation strategy facilitates personalized CT imaging at both patient and scanning levels, guided by the principles of physical imaging.", + "bbox": [ + 89, + 643, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, existing personalized FL CT imaging methods do not address the challenge posed by unseen clients with varying scanning protocols. In this paper, we propose a Protocol Vector Quantization Strategy (PVQS) to quantize unseen scanning protocol by matching it with the closest existing vectors, thereby selecting the corresponding client-specific decoder. The main contributions of this paper are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce LLM into LDCT reconstruction within the FL framework, proposing SCAN-PhysFed. To the best of our knowledge, this is the first work to apply LLM in this field.", + "- To address the challenge of data heterogeneity, we draw inspiration from physical imaging principles and propose a dual-level approach, achieving personalized CT imaging at both the scanning and patient levels.", + "- We introduce a novel strategy, PVQS, to ensure our method remains effective for unseen clients with diverse scanning protocols—a critical challenge often overlooked in previous works." + ], + "bbox": [ + 511, + 212, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 407, + 661, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LDCT Imaging. Traditional reconstruction methods primarily rely on image sparsity [25]. However, these methods are generally time-consuming. Recently, deep learning-based methods have shown faster processing and promising performance, gaining significant attention in this field. For example, Chen et al. [4] introduced residual structure into an autoencoder and proposed the encoder-decoder convolutional neural network (RED-CNN) for LDCT denoising. You et al. [39] extended RED-CNN by incorporating a generative adversarial network (GAN) for super resolution. Additionally, researchers also explored transformer-based architectures to enhance robustness, as seen in Uformer [31], TransCT [40], and RegFormer [34]. Recently, researchers have proposed diffusion-based CT reconstruction methods [9, 21, 33], which show promising results but incur substantial computational costs. Moreover, the methods mentioned above are tailored to specific scanning protocols and depend on data collection from multiple clients, raising significant privacy concerns.", + "bbox": [ + 511, + 431, + 906, + 719 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Federated Learning. In recent years, FL has gained increasing attention as it offers a decentralized approach to DL while preserving data privacy [38]. FedAvg [23], the earliest FL method, introduced a server-client framework to learn a global-shared model by averaging local models from different parties. However, this method suffers from model shift due to serious statistic heterogeneity [11]. To alleviate this concern, recent works have proposed methods to obtain a more robust global-shared model, such as FedProx [16], MOON [15], and FedNova [29]. For serious non-iid data distributions, a single global-shared model is insufficient to address the variability across different datasets. As", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5155", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c60c76674ed7c98a49b9ac68176ac7dad9a3cac2e9e5719c8a8f7e23316603f2.jpg", + "image_caption": [ + "Figure 2. The overall learning paradigm of our proposed SCAN-PhysFed." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 897, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a result, several personalized FL methods have been proposed [3, 7, 17, 35].", + "bbox": [ + 89, + 357, + 482, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FL has recently been introduced in healthcare applications due to its privacy-preserving nature [10, 32, 36]. Several studies have focused on its applications in medical imaging [8, 14]. For example, Chen et al. [5] proposed a cross-domain learning approach that combines individual-client sinogram learning and cross-client image reconstruction learning. Yang et al. [37] proposed a client-specific hypernetwork to modulate the global-shared data. Chen et al. [6] utilized separate networks to process high- and low-frequency components of CT data. While these methods have achieved promising performance, they still face significant challenges from data heterogeneity and fail to address the issue of unseen protocols.", + "bbox": [ + 88, + 386, + 482, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Language Model. The emergence of large language models (LLMs), such as GPT [1] and Llama [26], has ushered in a new era of artificial intelligence. Recently, largescale medical LLMs (MLLMs) have been developed specifically for healthcare applications [18, 20, 41]. These models demonstrate significant potential in understanding and reasoning over medical data, excelling in tasks such as visual question answering (VQA) and text generation [24]. For instance, Alkhaldi et al. [2] developed Minigpt-med, which supports multiple tasks, including VQA, medical report generation, and disease identification from medical images. Although MLLMs have been applied to various downstream tasks [13], they have not yet been explored for mitigating domain shift in CT reconstruction within FL settings.", + "bbox": [ + 89, + 584, + 482, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 821, + 225, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 847, + 200, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper aims to address generalization challenges in FL for LDCT reconstruction, with a focus on alleviating data", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "heterogeneity among existing clients and maintaining performance for new clients. Figure 2 illustrates the overall learning paradigm of the proposed SCAN-PhysFed framework. Specifically, robust imaging representations are obtained using a collaboratively learned shared encoder across clients. To tackle the heterogeneity issue, two physics-informed hypernetworks are designed to capture scanning-and anatomy-specific information, personalizing the imaging features via a specialized modulation block. Subsequently, client-specific decoders project the personalized imaging features back to the image domain. For the $i$ -th client, the reconstruction process is formulated as follows:", + "bbox": [ + 511, + 357, + 906, + 537 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Y}} = D _ {i} \\left(P \\left(f _ {\\mathbf {X}}, H _ {s} (g; \\theta_ {H _ {s}}), H _ {a} \\left(f _ {t}; \\theta_ {H _ {a}}\\right)\\right); \\theta_ {D _ {i}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 549, + 906, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{\\mathbf{X}}$ denotes the imaging feature of $\\mathbf{X}$ extracted from the shared encoder $E$ , and $\\hat{\\mathbf{Y}}$ is the corresponding normal-dose CT image. $\\theta_{H_s}$ and $\\theta_{H_a}$ are the scanning- and anatomy-informed hypernetworks, parameterized by $\\theta_s$ and $\\theta_{a}$ , respectively. $g$ and $f_{t}$ represent the scanning protocol vector and the radiology report feature derived from the MLLM. $P$ denotes the personalized modulation block, and $D_{i}$ denotes the $i$ -th client decoder, parameterized by $\\theta_{D_i}$ .", + "bbox": [ + 511, + 579, + 905, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, to handle unseen data, we propose a novel protocol vector-quantization strategy (PVQS), which ensures robust performance across previously unseen clients by quantizing the unseen scanning code to the closest match in the scanning codebook.", + "bbox": [ + 511, + 700, + 905, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Personalized Modulation Block", + "text_level": 1, + "bbox": [ + 511, + 786, + 790, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anatomy-Level Personalization. The patient's unique anatomy significantly affects noise distribution during the imaging process, making it an essential factor to consider. To address this issue, we leverage MLLM, which is trained on large-scale datasets and demonstrates strong performance in radiology report generation. Specifically,", + "bbox": [ + 511, + 810, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5156", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we adapt miniGPT-Med [2] to generate radiology reports from CT images, using the text encoder's feature output $f_{t} \\in \\mathbb{R}^{1 \\times d}$ as a prompt for our anatomy-informed hypernetwork, where $d$ denotes the feature dimension.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the high dimensionality of $f_{t}$ , we first apply a pooling operation to reduce its size. The process is formulated as follows:", + "bbox": [ + 89, + 152, + 482, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t l} = \\operatorname {L i n e a r} \\left(\\operatorname {A v g P o o l} \\left(f _ {t}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 210, + 482, + 228 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{AvgPool}$ denotes an average pooling operation that reduces the feature dimension by a factor of $1 / 4$ . A linear layer is then applied to map the feature to $f_{tl} \\in \\mathbb{R}^{1 \\times h}$ , where $h$ denotes the hidden dimension.", + "bbox": [ + 89, + 239, + 482, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we capture long-range dependencies within $f_{tl}$ using the multi-head attention mechanism, followed by a linear layer to extract the anatomy feature $f_{an}$ . This process is formulated as:", + "bbox": [ + 89, + 301, + 483, + 361 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {a n} = \\operatorname {L i n e a r} (\\operatorname {M u l t i - H e a d} (f _ {t l})). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 376, + 482, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After this stage, $f_{an}$ encapsulates anatomy-specific information tailored to different CT data. Since this feature encodes anatomy-specific details, we use it to spatially personalize the imaging feature $f_{\\mathbf{X}}$ . The process is formulated as:", + "bbox": [ + 89, + 405, + 483, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {a n a} = f _ {\\mathbf {X}} \\odot \\operatorname {R e s h a p e} \\left(f _ {a n}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 482, + 482, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\mathbf{X}}$ denotes the imaging feature, Reshape represents the reshape operation for dimensional alignment, and “ $\\odot$ ” denotes the Hadamard product. $f_{ana}$ is the anatomy-level personalized feature.", + "bbox": [ + 89, + 507, + 483, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scanning-Level Personalization. In addition to anatomical variations affecting noise distribution, the physical imaging process shows that noise distribution is primarily governed by the scanning protocol. This protocol comprises several key physical parameters in CT scanners, including the number of views, the number of detector bins, pixel length, detector bin length, the distance between the source and rotation center, the distance between the detector and rotation center, and the photon number of incident X-rays.", + "bbox": [ + 89, + 568, + 482, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we utilize $g$ , a vector encoding these parameters, to construct our scanning-level prompt, which captures the latent relationship between the scanning protocol and noise distribution for the imaging network. A detailed description of $g$ is provided in Appendix B.", + "bbox": [ + 89, + 705, + 482, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the low dimensionality of the physical parameters, we employ a Multi-Layer Perceptron (MLP) with three linear layers to capture and represent the underlying physical information. Two modulation parameter vectors are then derived to encode scanning-level personalization knowledge. This process can be formulated as follows:", + "bbox": [ + 89, + 781, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha , \\beta = H _ {s} (g; \\theta_ {s}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are the modulation parameter vectors, which are used to inject scanning-level information to personalize the imaging feature as:", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {p e r} = \\alpha \\otimes f _ {a n a} + \\beta , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 146, + 903, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{per}$ is the dual-level personalized imaging feature, and $\\otimes$ denotes the multiplication operation along the channel dimension. This approach applies the scanning prior to modulate the imaging feature.", + "bbox": [ + 511, + 170, + 903, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed personalization block is architecture-agnostic, allowing the personalization parameters to be finetuned accordingly.", + "bbox": [ + 511, + 231, + 903, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Protocol Vector-Quantization Strategy", + "text_level": 1, + "bbox": [ + 511, + 284, + 844, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing FL-based CT imaging methods have not addressed the challenge of unseen data, resulting in degraded performance in new domains. To overcome this limitation and ensure robust performance for unseen protocols, we introduce PVQS. While patient anatomy remains largely consistent, the primary source of noise variation between unseen and existing clients stems from the low-dose strategy. Therefore, PVQS is designed to enable SCAN-PhysFed to maintain consistent performance across diverse and unseen low-dose protocols.", + "bbox": [ + 511, + 306, + 903, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, in PVQS, protocol codes for different existing scanning protocols, generated by the MLP in $H_{s}$ , are stored in a protocol codebook. For an unseen protocol, we first quantize it to the closest code vector in the protocol codebook. The process is defined as follows:", + "bbox": [ + 511, + 458, + 905, + 534 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ni ^ {*} = \\arg \\min _ {i} \\left(1 - \\frac {\\mathbf {c} _ {u n} \\cdot \\mathbf {c} _ {i}}{\\| \\mathbf {c} _ {u n} \\| \\| \\mathbf {c} _ {i} \\|}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 542, + 903, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_{un}$ and $\\mathbf{c}_i$ denote the protocol code of the unseen protocol and that of the $i$ -th client, respectively. $i^*$ indicates the index of the closest code vector.", + "bbox": [ + 511, + 582, + 903, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of using the unseen protocol as input to $H_{s}$ , we directly replace it with $\\mathbf{c}_{i^{*}}$ to personalize the imaging feature and use the corresponding decoder $D_{i^{*}}$ to project the personalized features back to the image domain.", + "bbox": [ + 511, + 627, + 903, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, PVQS effectively avoids potential distribution shifts in unseen data and preserves a consistent feature space across domains, ensuring reliable and robust performance for unseen protocols.", + "bbox": [ + 511, + 686, + 903, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Learning Paradigm", + "text_level": 1, + "bbox": [ + 511, + 756, + 700, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the noise distribution varies across protocols, it is challenging to use a single shared imaging network to accommodate all variations. In FL-based LDCT imaging, as assumed in previous works [6, 37], the optimization challenge consists of two main components: imaging feature extraction and personalized projection. Specifically, for feature extraction, we propose aggregating client-side encoders to capture robust, comprehensive imaging features.", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5157", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/46d3c525fb99831c5d5bd3471eae49d1cb79dae1f9c0e543bc65e92a296101a5.jpg", + "table_caption": [ + "Table 1. Quantitative Results (PSNR (dB) and SSIM $(\\%)$ ) for CNN-based LDCT Imaging Method." + ], + "table_footnote": [], + "table_body": "
Client #1Client #2Client #3Client #4Client #5Client #6Client #7Client #8Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]29.9982.2032.8278.2334.6886.9932.3080.6336.8792.0637.3689.0232.7380.6635.4687.6534.0384.70
FedProx [16]29.9483.1232.6975.4636.9591.5832.3280.9138.9595.3437.5888.1332.9481.6235.8688.6534.6585.60
FedNova [29]29.9582.2932.5780.5737.7792.1632.0978.9834.4288.6936.0589.5732.8181.4135.8988.7133.9485.30
MOON [15]32.5075.6135.5585.4135.7280.3035.6982.1035.3878.4736.3184.3535.7081.1735.6280.4035.3180.98
FedDG [19]33.3781.4034.0182.3933.5676.3734.4381.4934.0676.7033.9780.1734.7580.8734.2579.2334.0579.83
FedKD [32]29.7483.1132.1771.6536.6590.9332.1079.1940.5896.7137.0782.5532.8881.3036.2889.0834.6884.31
Personalized FL Methods
FedPer [3]33.2388.2035.0684.8938.7992.7035.5886.1241.7696.9736.8589.9635.6986.3038.4891.6936.9389.60
FedBN [17]32.5584.5934.7684.1538.1890.3034.7783.3041.2496.1138.6691.1934.7483.3537.8990.3536.6087.92
FedMRI [8]33.2883.4334.5185.8138.6292.4735.8687.2141.4996.2738.2291.3235.8786.3438.0991.9436.9989.35
HyperFed [37]33.7488.5734.8385.0538.8792.7835.4485.4042.9497.4839.1592.9634.9985.1938.1291.2937.2689.84
FedFDD [6]36.1893.6738.4394.1939.7296.4738.5195.3045.5698.7342.7297.6140.7496.0138.5494.3940.0595.79
SCAN-PhysFed34.4996.4039.2196.0940.4597.9340.5197.6147.8199.1444.0898.5142.5497.8640.0297.3741.1497.62
", + "bbox": [ + 88, + 112, + 906, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e2ed308faa97a69215e437e2567f90cd171f3eab8374cbf1ce2efebaa5808718.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 303, + 160, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/df320a31fde33d8a04455b194af685367fc386bc034619a43ebe8f5635209b4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 303, + 223, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1621a7b22ab0b4585a96394a2ab048d94f71d1db0d177ace0beb4125f70648dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 303, + 285, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6e8dad575802e27a24f9978c05a0c34d869a60c162e7fff50c8e4573939ccf67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 303, + 348, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a4f9850e4c9c1987d39d1c95dc43b6ae5e12b6170e7a9e567150392d79c5b468.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 303, + 411, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/535f9ab309baccf653ff651bbb12a680f621de40c6344b9732d19432fbbecc52.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 413, + 303, + 473, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5051a10275f509569b4bd3291eda3da090e36a48139f7c6a76cd1660304489d1.jpg", + "image_caption": [ + "(a)", + "(b)", + "(g)", + "Figure 3. Simulated examples under different protocols. (a)-(h) show examples from training clients with known protocols, while (i)-(l) present examples from unseen protocols." + ], + "image_footnote": [], + "bbox": [ + 99, + 357, + 161, + 404 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a12e95d7688dc525aeb9f99fc69a05a986bb1b64df328f923766c068a791574c.jpg", + "image_caption": [ + "(h)" + ], + "image_footnote": [], + "bbox": [ + 163, + 357, + 223, + 404 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3dcab6b06dec54e402b18748666968c8e567a3ec44e8dd646d287029990b4cd2.jpg", + "image_caption": [ + "", + "(i)" + ], + "image_footnote": [], + "bbox": [ + 225, + 357, + 285, + 404 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c840ccd4dc1b67661d13e7561a544f1438871505d987bd0a5df82c7cffbec3ac.jpg", + "image_caption": [ + "(d)", + "(1)" + ], + "image_footnote": [], + "bbox": [ + 287, + 357, + 348, + 404 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0cf2042361aa975a2f1aef7bd6deda92d8d920493f4b81514cde0811f23b0475.jpg", + "image_caption": [ + "(e)", + "(k)" + ], + "image_footnote": [], + "bbox": [ + 349, + 357, + 411, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To fully leverage the diversity of client data, we globally share the hypernetworks in SCAN-PhysFed. However, projecting imaging features from different domains into a consistent, noise-free representation is challenging with a single shared decoder. Therefore, we propose protocol-specific decoders to effectively adapt to variations in scanning protocols across clients.", + "bbox": [ + 89, + 460, + 482, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In PVQS, the unseen protocol code is quantized based on the nearest entry in the protocol codebook. Therefore, it is essential to ensure the protocol code is discriminative and informative for personalization. To achieve this, we introduce an orthogonal loss $\\mathcal{L}_{\\mathrm{orth}}$ for the $i$ -th client, defined as follows:", + "bbox": [ + 89, + 565, + 483, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o r t h}} = \\sum_ {j = 1} ^ {K} \\left| \\mathbf {c} _ {i} \\cdot \\mathbf {c} _ {j} \\right| ^ {2} \\quad s. t. \\quad i \\neq j, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 666, + 482, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K$ is the number of clients (protocols), and $j$ denotes the client index.", + "bbox": [ + 89, + 720, + 482, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Mean Squared Error (MSE) is used as the imaging loss, and the total loss is formulated as:", + "bbox": [ + 89, + 751, + 482, + 780 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {M S E}} (\\mathbf {Y}, \\hat {\\mathbf {Y}}) + \\tau \\mathcal {L} _ {\\text {o r t h}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 792, + 482, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Y}$ represents the reconstructed image, and $\\tau$ denotes the temperature weight of $\\mathcal{L}_{\\mathrm{orth}}$ .", + "bbox": [ + 89, + 824, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To facilitate understanding of the implementation details, we provide the pseudocode of our method in Appendix A.", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 299, + 643, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Setting. The proposed method was implemented using PyTorch and optimized by Adam [12] at a learning rate of 0.001. The number of communication rounds was set to 200 and the batch size was 20. The experiments were conducted a system equipped with an AMD Ryzen 75800X CPU and a single NVIDIA GTX 3080 Ti GPU.", + "bbox": [ + 511, + 324, + 903, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. The \"2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge\" dataset [22], which contains 5,936 full-dose CT images from 10 patients, was used to evaluate the proposed method. In this study, 8 patients were used for training, while the remaining 2 were reserved for testing. To simulate a realistic scenario, we ensured that the patients for training did not overlap across different clients. The two patients for testing were simulated using different protocols to form the testing set.", + "bbox": [ + 511, + 430, + 905, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Simulated examples are shown in Figure 3, where noticeable noise heterogeneity among clients can be observed, primarily due to differences in anatomy and scanning protocols. Detailed dataset preparation steps and scanning protocols are provided in Appendix B.", + "bbox": [ + 511, + 566, + 905, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compared our SCAN-PhysFed approach with both generic and personalized FL methods. For generic FL methods, we include FedAvg [23], FedProx [16], FedNova [29], MOON [15], FedDG [19], and FedKD [32]. For personalized FL methods, we evaluated FedPer [3], FedBN [17], FedMRI [8], HyperFed [37], and FedFDD [6]. To ensure fairness, we kept the training settings consistent across all methods.", + "bbox": [ + 511, + 642, + 905, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use the peak signal-to-noise ratio (PSNR) to evaluate pixel-wise accuracy and the structural similarity (SSIM) to assess perceived visual quality [30]. For both metrics, higher values indicate better performance.", + "bbox": [ + 511, + 762, + 903, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparison with other methods", + "text_level": 1, + "bbox": [ + 511, + 847, + 797, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method alongside other FL methods. To ensure fairness, we use the same imaging network, RED", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5158", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/319310e54dc1ecebd8393ade11f4f345fd33b8283cecb1cfcdaace46c40bf822.jpg", + "image_caption": [ + "Figure 4. Qualitative results of six selected comparison methods and our method across different clients using the classical convolutional-based LDCT imaging network. Rows one to five correspond to Clients #2, #3, #5, #6, and #7, respectively. The display window for the first row is [-1024, 200] HU, while for the remaining rows, it is [-160, 240] HU." + ], + "image_footnote": [], + "bbox": [ + 102, + 95, + 890, + 470 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f533115db85bd21fcbf9159a0a753652c41cfaeef2f79c76c3c97094afbee6b3.jpg", + "image_caption": [ + "Figure 5. Boxplots of the average results across all clients. (a) and (b) represent PSNR and SSIM, respectively." + ], + "image_footnote": [], + "bbox": [ + 99, + 542, + 478, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CNN [4], a state-of-the-art model in CT imaging. Quantitative results are provided in Table 1. Our method achieves the best performance across most clients, outperforming other generic and personalized FL methods in terms of average PSNR and SSIM. Furthermore, unlike other personalized methods that show instability across clients, such as HyperFed in Client #2 / #7 and FedFDD in Client #4, our method maintains consistent performance across varying protocols. We also observe that sparse-view protocols (Clients #2, #4, and #6) pose challenges for both generic and personalized FL methods, resulting in significant performance gaps compared to other protocols. How-", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ever, by incorporating dual-level physics-informed prior, our method maintains robust performance.", + "bbox": [ + 511, + 541, + 906, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Representative slices produced by different methods are shown in Figure 4. We compared the top three methods from both generic and personalized FL approaches. Representative clients were selected based on different radiation reduction strategies: sparse-view (Clients #2 and #6) and low-dose (Clients #3, #5, and #7) reconstructions. It can be observed that other methods exhibit noticeable noise or artifacts in sparse-view protocols, and fail to preserve details in low-dose protocols. In contrast, our method effectively reduces noise and artifacts in sparse-view protocols while preserving fine details in low-dose protocols. SCAN-PhysFed achieves strong performance across slices from different anatomical regions, benefiting from the introduction of anatomy priors. Both quantitative and qualitative results demonstrate the effectiveness of incorporating physics-informed priors.", + "bbox": [ + 511, + 577, + 906, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we present boxplots of the results across all clients for different methods in Figure 5 to validate their stability. It can be seen that SCAN-PhysFed outperforms all other methods in both PSNR and SSIM scores. In both plots, SCAN-PhysFed demonstrates a higher median and", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5159", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/905cc7a89ec4f2eefc841616c0f91e011d3d44a7b7517b7c47b0d230d3418e4d.jpg", + "table_caption": [ + "Table 2. Quantitative Results of PSNR and SSIM for Transformer-based LDCT Imaging Methods." + ], + "table_footnote": [], + "table_body": "
Client #1Client #2Client #3Client #4Client #5Client #6Client #7Client #8Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]34.3187.8838.0690.5443.5197.5641.0595.4345.2498.4743.1597.0941.4595.9542.8896.1641.2194.89
FedProx [16]34.3287.0137.7488.8943.7397.9541.1293.9245.3998.6142.9997.0841.3894.3142.1094.5541.1094.04
FedNova [29]29.5578.4431.3868.1337.0782.4731.7973.4540.6188.8035.4679.7732.3873.6635.7782.2234.2578.37
MOON [15]34.2893.2038.0991.1743.6497.7741.0995.6645.2598.4342.8296.9141.4295.9742.2096.1841.1095.66
FedKD [32]35.9289.9338.7091.2144.2698.1342.3095.9945.9398.7543.5797.4142.5496.2243.5196.2342.0995.48
Personalized FL Methods
FedPer [3]35.3793.1839.0393.5542.7597.1340.8896.2046.5598.8342.8796.6841.8796.7040.7994.7841.2695.82
FedMRI [8]34.6093.6738.0494.5542.9497.4439.3495.7846.7998.9142.8297.3040.8195.3240.8595.7040.7896.01
HyperFed [37]34.6087.5938.6291.0844.1397.6142.2596.2445.9298.5943.7197.3742.4496.6643.5696.6341.9095.22
FedFDD [6]37.3596.2039.6795.1844.3198.2639.0095.2147.8999.2043.8198.0642.1797.4040.5095.4941.8496.89
SCAN-PhysFed38.5594.0641.7496.2045.5298.5242.4896.5549.5099.3945.3297.9343.3596.7342.7196.7943.6597.03
Ground TruthInputFedAvgMOONFedKDFedMRIHyperFedFedFDDSCAN-PhysFed
", + "bbox": [ + 86, + 112, + 906, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6. Qualitative results of the compared methods across different clients based on a transformer-based imaging network. Rows one to five correspond to Clients #2, #3, #5, #6, and #7, respectively. The display window is set to [-160, 240] HU.", + "bbox": [ + 89, + 648, + 906, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a narrower interquartile range, indicating strong stability in restoring LDCT data with varying noise distributions across different clients. Individual boxplots for each client are provided in Appendix C.", + "bbox": [ + 89, + 690, + 482, + 751 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Generalization Evaluation", + "text_level": 1, + "bbox": [ + 89, + 761, + 328, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Backbone", + "text_level": 1, + "bbox": [ + 89, + 785, + 215, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to using RED-CNN as the imaging network in previous experiments, we evaluate our method and other FL methods with an alternative imaging network, Uformer [31], a transformer-based model for LDCT imaging. This comparison further validates the generalization capability of different FL methods.", + "bbox": [ + 89, + 809, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The qualitative and quantitative results are shown in Figure 6 and Table 2, respectively. Consistent with previous experiments, our method achieves the best performance compared to other methods across most clients, with significantly higher average PSNR and SSIM values. This demonstrates the strong generalization capability of our method when integrated with different imaging networks. Among competing methods, FedFDD serves as a strong baseline. While FedFDD's SSIM is slightly higher than ours in some clients, our method achieves a significantly better PSNR, demonstrating superior detail preservation. Additionally, unlike FedFDD, which requires two imaging networks, our method achieves high performance with", + "bbox": [ + 511, + 690, + 906, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5160", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ec33a3a7abbf2f007e4be564c493578d72708db768c10d990743998f59d9e282.jpg", + "table_caption": [ + "Table 3. The Quantitative Results of PSNR and SSIM for the Unseen Clients." + ], + "table_footnote": [], + "table_body": "
Unseen Client #1Unseen Client #2Unseen Client #3Unseen Client #4Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]35.5788.6834.0085.2632.5276.7136.8691.4432.8981.35
FedProx [16]36.6991.1634.3086.3932.5872.1339.1394.8734.7485.52
FedNova [29]35.4588.8334.2386.2932.3969.2035.4089.9934.3883.58
MOON [15]35.2178.1935.3079.8134.5583.0735.5678.9435.1680.00
FedDG [19]34.8679.5434.7980.9133.1580.3134.4977.4234.3279.55
FedKD [32]36.8291.6034.2786.3732.4970.5440.7395.9036.0886.10
Personalized FL Methods
FedPer [3]36.8688.5835.4687.0231.9671.6438.5690.9035.7184.53
FedBN [17]37.8489.9536.1688.4633.6781.6340.2194.8936.9788.73
FedMRI [8]31.1282.4231.9579.9531.7271.9632.0481.6331.7178.99
HyperFed [37]36.1889.6033.4782.4931.9465.5241.7094.8535.9083.12
FedFDD [6]37.8092.3337.0691.7733.7180.2640.5395.1237.2789.87
SCAN-PhysFed40.7797.6938.7795.7534.3881.8040.2796.9238.5593.04
", + "bbox": [ + 156, + 113, + 828, + 296 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ece96145636ab9d272004ffef703672543f9ae00e3f4548dd7e52de314884a3c.jpg", + "table_caption": [ + "Table 4. Ablation on different components in SCAN-PhysFed." + ], + "table_footnote": [], + "table_body": "
ScanningComponentAverage
AnatomyParadigmLossPSNRSSIM
XXGenericLMSE34.0384.70
✓XPersonalizedLMSE36.8288.78
X✓PersonalizedLMSE38.0494.14
✓✓PersonalizedLMSE40.6797.27
✓✓PersonalizedLMSE & Lorth41.1497.62
", + "bbox": [ + 91, + 333, + 480, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "out relying on an additional network. Similar to the previous experiment, SCAN-PhysFed shows better qualitative performance and lower variation across different protocols, indicating more consistent results irrespective of scanning protocols or imaging networks.", + "bbox": [ + 89, + 433, + 482, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.2 Unseen Protocol", + "text_level": 1, + "bbox": [ + 89, + 527, + 259, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Existing FL methods in medical imaging struggle to maintain performance when encountering unseen noise distributions from new clients. Here, we evaluate the effectiveness of the proposed PVQS across four unseen protocols. Detailed protocol settings can be found in Appendix B. For generic FL methods, we directly evaluate the shared model on unseen clients. However, personalized methods mostly ignore how to operate in the unseen domain, so we average the performance of all local models to establish a vanilla baseline in this paper.", + "bbox": [ + 89, + 551, + 482, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results of different methods on unseen clients are presented in Table 3. We observe that the performance of generic FL methods shows only minor degradation on unseen clients, whereas personalized FL methods experience significant declines. This is because local optimization in personalized FL methods focuses heavily on each client's specific data distribution, underfitting other potential distributions and leading to greater performance drops on unseen clients. In PVQS, we prevent the network from directly mapping the unseen protocol to modulate the imaging network, as this could cause network collapse if the unseen protocol significantly differs from known ones. Instead, we quantize the unseen protocol using a protocol codebook, al", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lowing SCAN-PhysFed to leverage prior knowledge for better adaptation to unseen distributions.", + "bbox": [ + 511, + 309, + 903, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 351, + 666, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of various components in SCAN-PhysFed, and the results are presented in Table 4. Compared to the baseline results, which exclude scanning and anatomy information, introducing either significantly improves performance, highlighting the importance of integrating physical prior in LDCT imaging. This approach substantially reduces learning complexity and improves the overall model performance. Moreover, introducing both types of physical information simultaneously is complementary rather than redundant, leading to further performance gains. Building on this, we use $\\mathcal{L}_{\\mathrm{orth}}$ to enhance the discriminative ability of the scanning code, thereby personalizing imaging features and effectively mitigating data heterogeneity. In the above studies, \"Personalized\" indicates that the decoder is tailored to each client. An ablation study on the personalized components is provided in Appendix D.", + "bbox": [ + 511, + 376, + 906, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To preserve privacy and address heterogeneity in CT imaging, we propose a dual-level, physics-informed approach that utilizes both scanning- and anatomy-level prompts to enable personalized CT imaging with the assistance of an MLLM. This approach facilitates both patient- and scanning-level personalization, with extensive results demonstrating that incorporating physical prior significantly improves performance and effectively mitigates heterogeneity. Additionally, we propose a novel strategy, PVQS, to quantize protocol codes and maintain robust performance on unseen protocols. PVQS leverages known protocol codes to prevent latent collapse when encountering significantly different unseen protocols. In future work, we aim to develop a general FL framework for medical imaging that can support both CT and MRI denoising, which we believe holds promising potential.", + "bbox": [ + 511, + 657, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5161", + "bbox": [ + 482, + 944, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62271335, in part by the Sichuan Science and Technology Program under Grant 2025ZNSFSC0470, and in part by Sichuan University \"From 0 to 1\" Innovative Research Program under Grant 2022SCUH0016.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3", + "[2] Asma Alkhaldi, Raneem Alnajim, Layan Alabdullatef, Rawan Alyahya, Jun Chen, Deyao Zhu, Ahmed Alsinan, and Mohamed Elhoseiny. Minigpt-med: Large language model as a general interface for radiology diagnosis. arXiv preprint arXiv:2407.04106, 2024. 3, 4", + "[3] Manoj Ghuhan Arivazhagan, Vinay Aggarwal, Aaditya Kumar Singh, et al. Federated learning with personalization layers. arXiv preprint arXiv:1912.00818, 2019. 3, 5, 7, 8", + "[4] Hu Chen, Yi Zhang, Mannudeep K Kalra, et al. Low-dose ct with a residual encoder-decoder convolutional neural network. IEEE Trans. Med. Imaging, 36(12):2524-2535, 2017. 2, 6", + "[5] Shixuan Chen, Boxuan Cao, Yinda Du, Yaoduo Zhang, Ji He, Zhaoying Bian, Dong Zeng, and Jianhua Ma. Federated condition generalization on low-dose ct reconstruction via cross-domain learning. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 47-56. Springer, 2023. 3", + "[6] Xuhang Chen, Zeju Li, Zikun Xu, Cheng Ouyang, Chen Qin, et al. Fedfdd: Federated learning with frequency domain decomposition for low-dose ct denoising. In Medical Imaging with Deep Learning, 2024. 3, 4, 5, 7, 8", + "[7] Dongshang Deng, Xuangou Wu, Tao Zhang, Xiangyun Tang, Hongyang Du, Jiawen Kang, Jiqiang Liu, and Dusit Niyato. Fedasa: A personalized federated learning with adaptive model aggregation for heterogeneous mobile edge computing. IEEE Transactions on Mobile Computing, 2024. 3", + "[8] Chun-Mei Feng, Yunlu Yan, Shanshan Wang, Yong Xu, Ling Shao, and Huazhu Fu. Specificity-preserving federated learning for mr image reconstruction. IEEE Transactions on Medical Imaging, 42(7):2010-2021, 2022. 3, 5, 7, 8", + "[9] Qi Gao, Zilong Li, Junping Zhang, Yi Zhang, and Hongming Shan. Corediff: Contextual error-modulated generalized diffusion model for low-dose ct denoising and generalization. IEEE Transactions on Medical Imaging, 2023. 2", + "[10] Hao Guan, Pew-Thian Yap, Andrea Bozoki, and Mingxia Liu. Federated learning for medical image analysis: A survey. Pattern Recognition, page 110424, 2024. 3", + "[11] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for feder-" + ], + "bbox": [ + 93, + 246, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 2", + "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[13] Gurucharan Marthi Krishna Kumar, Aman Chadha, Janine Mendola, and Amir Shmuel. Medvisionllama: Leveraging pre-trained large language model layers to enhance medical image segmentation. arXiv preprint arXiv:2410.02458, 2024.3", + "[14] Danyang Li, Hao Wang, Jingyi Liao, Mingqiang Meng, Cuidie Zeng, Yuting Wang, Dong Zeng, and Jianhua Ma. Semi-centralized federated learning network for low-dose ct imaging. In Medical Imaging 2023: Physics of Medical Imaging, pages 1024-1028. SPIE, 2023. 3", + "[15] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 10713–10722, 2021. 2, 5, 7, 8", + "[16] Tian Li, Anit Kumar Sahu, Manzil Zaheer, et al. Federated optimization in heterogeneous networks. In Proc. Mach. Learn. Syst. (MLSys), pages 429-450, 2020. 2, 5, 7, 8", + "[17] Xiaoxiao Li, Meirui JIANG, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. In International Conference on Learning Representations, 2021. 3, 5, 8", + "[18] Fenglin Liu, Tingting Zhu, Xian Wu, Bang Yang, Chenyu You, Chenyang Wang, Lei Lu, Zhangdaihong Liu, Yefeng Zheng, Xu Sun, et al. A medical multimodal large language model for future pandemics. NPJ Digital Medicine, 6(1): 226, 2023. 3", + "[19] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1013–1023, 2021. 5, 8", + "[20] Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Derong Xu, Feng Tian, and Yefeng Zheng. When moe meets llms: Parameter efficient fine-tuning for multi-task medical applications. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1114, 2024. 3", + "[21] Zexin Lu, Qi Gao, Tao Wang, Ziyuan Yang, Zhiwen Wang, Hui Yu, Hu Chen, Jiliu Zhou, Hongming Shan, and Yi Zhang. Pridediff: Physics-regularized generalized diffusion model for ct reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences, 2024. 2", + "[22] C McCollough. Tu-fg-207a-04: overview of the low dose ct grand challenge. Med. Phys., 43(6):3759–3760, 2016. 5", + "[23] Brendan McMahan, Eider Moore, Daniel Ramage, et al. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 2, 5, 7, 8", + "[24] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medi" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5162", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cal artificial intelligence. Nature, 616(7956):259-265, 2023. 3", + "[25] Saiprasad Ravishankar, Jong Chul Ye, and Jeffrey A Fessler. Image reconstruction: From sparsity to data-adaptive methods and machine learning. Proceedings of the IEEE, 108(1): 86-109, 2019. 2", + "[26] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[27] Ge Wang, Hengyong Yu, and Bruno De Man. An outlook on x-ray ct research and development. Medical physics, 35(3): 1051-1064, 2008. 1", + "[28] Ge Wang, Jong Chul Ye, and Bruno De Man. Deep learning for tomographic image reconstruction. Nature machine intelligence, 2(12):737-748, 2020. 1", + "[29] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 2, 5, 7, 8", + "[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[31] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 2, 7", + "[32] Chuhan Wu, Fangzhao Wu, Lingjuan Lyu, Yongfeng Huang, and Xing Xie. Communication-efficient federated learning via knowledge distillation. Nature Communications, 13(1): 2032, 2022. 3, 5, 7, 8", + "[33] Wenjun Xia, Qing Lyu, and Ge Wang. Low-dose ct using denoising diffusion probabilistic model for 20x speedup. arXiv preprint arXiv:2209.15136, 2022. 2", + "[34] Wenjun Xia, Ziyuan Yang, Zexin Lu, Zhongxian Wang, and Yi Zhang. Regformer: A local-nonlocal regularization-based model for sparse-view ct reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences, 8(2): 184-194, 2023. 2", + "[35] Xiyuan Yang, Wenke Huang, and Mang Ye. Fedas: Bridging inconsistency in personalized federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11986-11995, 2024. 3", + "[36] Ziyuan Yang, Yingyu Chen, Huijie Huangfu, Maosong Ran, Hui Wang, Xiaoxiao Li, and Yi Zhang. Dynamic corrected split federated learning with homomorphic encryption for u-shaped medical image networks. IEEE Journal of Biomedical and Health Informatics, 27(12):5946-5957, 2023. 3", + "[37] Ziyuan Yang, Wenjun Xia, Zexin Lu, Yingyu Chen, Xiaoxiao Li, and Yi Zhang. Hypernetwork-based physics-driven personalized federated learning for ct imaging. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3, 4, 5, 7, 8" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Mang Ye, Xiuwen Fang, Bo Du, Pong C Yuen, and Dacheng Tao. Heterogeneous federated learning: State-of-the-art and research challenges. ACM Computing Surveys, 56(3):1-44, 2023. 2", + "[39] Chenyu You, Guang Li, Yi Zhang, et al. Ct super-resolution gan constrained by the identical, residual, and cycle learning ensemble (gan-circle). IEEE Trans. Med. Imaging, 39(1): 188-203, 2019. 2", + "[40] Zhicheng Zhang, Lequan Yu, Xiaokun Liang, Wei Zhao, and Lei Xing. Transc: dual-path transformer for low dose computed tomography. In Medical Image Computing and Computer Assisted Intervention, pages 55–64. Springer, 2021. 2", + "[41] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large language models meet few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3065-3075, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 330 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5163", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/b8f485b4-2d7f-474b-acf2-8dc4a4e3047c_model.json b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/b8f485b4-2d7f-474b-acf2-8dc4a4e3047c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e23f52bce6c9ee213ea07e56b060199b01ae6c0b --- /dev/null +++ b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/b8f485b4-2d7f-474b-acf2-8dc4a4e3047c_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.13, + 0.895, + 0.177 + ], + "angle": 0, + "content": "Patient-Level Anatomy Meets Scanning-Level Physics: Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.203, + 0.871, + 0.293 + ], + "angle": 0, + "content": "Ziyuan Yang\\(^{1,5}\\), Yingyu Chen\\(^{1,5}\\), Zhiwen Wang\\(^{2}\\), Hongming Shan\\(^{3}\\), Yang Chen\\(^{4}\\), Yi Zhang\\(^{1,5,*}\\) \n\\(^{1}\\)Sichuan University \\(^{2}\\)Southwest Petroleum University \\(^{3}\\)Fudan University \\(^{4}\\)Southeast University \\(^{5}\\)Tianfu Jiangxi Laboratory \n{cziyuanyang, cyy262511, zwwang1228}@gmail.com, hmshan@fudan.edu.cn \nchenyang.list@seu.edu.cn, yzhang@scu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.342, + 0.327, + 0.358 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.386, + 0.486, + 0.857 + ], + "angle": 0, + "content": "Reducing radiation doses benefits patients, but the resultant low-dose computed tomography (LDCT) images often suffer from clinically unacceptable noise and artifacts. While deep learning (DL) has shown promise in LDCT reconstruction, it requires large-scale data collection from multiple clients, raising privacy concerns. Federated learning (FL) has been introduced to mitigate these privacy concerns; however, current methods are typically tailored to specific scanning protocols, which limits their generalizability and makes them less effective for unseen protocols. To address these issues, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Since the noise distribution in LDCT data is closely tied to scanning protocols and anatomical structures, we propose a dual-level physics-informed way to address these challenges. Specifically, we incorporate physical and anatomical prompts into our physics-informed hypernetworks to capture scanning- and anatomy-specific information, enabling dual-level physics-driven personalization of imaging features. These prompts are derived from the scanning protocol and the radiology report generated by a medical large language model (MLLM). Subsequently, client-specific decoders project these dual-level personalized imaging features back into the image domain. Besides, to tackle the challenge of unseen data, we introduce a novel protocol vector-quantization strategy (PVQS), which ensures consistent performance across new clients by quantifying unseen scanning codes to the closest match in the scanning codebook. Extensive experimental results demonstrate the superior performance of SCAN-PhysFed on pub" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.328, + 0.603, + 0.343 + ], + "angle": 0, + "content": "lic datasets1." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.373, + 0.645, + 0.388 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.563 + ], + "angle": 0, + "content": "Computed tomography (CT) is a key diagnostic tool that non-invasively visualizes anatomical structures within the human body. Despite its clinical benefits, CT scans pose radiation concerns, which potentially lead to genetic, cancerous, and other health diseases. Lowering CT radiation dose—either by decreasing incident photons or limiting sampling views—has emerged as a promising solution to reduce radiation risks and accelerate scanning. However, reconstructed images under these conditions suffer from significant quality degradation, compromising clinical applicability [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.805 + ], + "angle": 0, + "content": "Recently, deep learning (DL)-based methods have shown great promise for low-dose CT (LDCT) imaging [28]. However, most DL-based LDCT reconstruction methods are condition-specific and overlook privacy concerns. Recently, researchers have attempted to introduce federated learning (FL), a privacy-preserving distributed learning paradigm, for LDCT reconstruction to alleviate privacy concerns. Although these methods aim to enhance generalization, they still face significant challenges arising from data heterogeneity, caused by diverse noise distributions in CT imaging across different protocols and anatomical structures. We believe that the further advancement of FL in LDCT reconstruction urgently requires a new framework capable of leveraging comprehensive physical information from the scanning process, thereby breaking through current bottlenecks to achieve better results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.806, + 0.909, + 0.868 + ], + "angle": 0, + "content": "To tackle this issue, we aim to leverage physical principles of the imaging process. During the projection process, the measurement is influenced by the anatomical structures the X-ray passed through, such as fat, bone, and fluid, each" + }, + { + "type": "page_footnote", + "bbox": [ + 0.122, + 0.888, + 0.251, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.515, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1 Our code has been released at https://github.com/Zi-YuanYang/SCAN-PhysFed." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5154" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.096, + 0.472, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.264, + 0.483, + 0.32 + ], + "angle": 0, + "content": "Figure 1. Illustration of different radiation reduction strategies. 'Sparse-View' refers to scanning with a reduced number of projection views, while 'Low-Dose' represents scanning with a lower number of incident photons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.446 + ], + "angle": 0, + "content": "with distinct attenuation coefficients. Additionally, the system matrix is significantly affected by scanning protocols, including tube current, voltage settings, and scanning angles. These factors collectively impact the quality and characteristics of reconstructed CT images during the backprojection phase, resulting in varying levels of noise and artifacts. An example is provided in Figure 1 to illustrate this concept more clearly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.642 + ], + "angle": 0, + "content": "By leveraging the physics-specific knowledge discussed earlier, we reimagine LDCT reconstruction within the federated learning framework. Our approach adopts a dual-level strategy to alleviate heterogeneity, addressing variations at both the scanning and patient levels. In practice, scanning protocols, which serve as scanner hyperparameters, are readily available. However, generating anatomical information has traditionally been challenging due to patient variability and the complexity of accurately modeling different tissue types. Recent advances in large language models (LLMs) offer new possibilities for automatically generating anatomical information, potentially improving the reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Building on this, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Specifically, we address the data heterogeneous using a dual-level strategy. A detailed radiology report is generated using a pretrained LLM. Then, a patient-level anatomy-informed hypernetwork is designed to generate a modulation map from the radiology report, which seamlessly integrates with imaging features and aligns them with each patient's unique anatomical structures. Simultaneously, scanning features are integrated via a scanning-informed hypernetwork to personalize imaging features based on the protocol. To ensure distinguishability of scanning feature vectors, we impose an orthogonal loss during training to reinforce personalization. This dual-modulation strategy facilitates personalized CT imaging at both patient and scanning levels, guided by the principles of physical imaging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "Additionally, existing personalized FL CT imaging methods do not address the challenge posed by unseen clients with varying scanning protocols. In this paper, we propose a Protocol Vector Quantization Strategy (PVQS) to quantize unseen scanning protocol by matching it with the closest existing vectors, thereby selecting the corresponding client-specific decoder. The main contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.213, + 0.905, + 0.271 + ], + "angle": 0, + "content": "- We introduce LLM into LDCT reconstruction within the FL framework, proposing SCAN-PhysFed. To the best of our knowledge, this is the first work to apply LLM in this field." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.332 + ], + "angle": 0, + "content": "- To address the challenge of data heterogeneity, we draw inspiration from physical imaging principles and propose a dual-level approach, achieving personalized CT imaging at both the scanning and patient levels." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.905, + 0.393 + ], + "angle": 0, + "content": "- We introduce a novel strategy, PVQS, to ensure our method remains effective for unseen clients with diverse scanning protocols—a critical challenge often overlooked in previous works." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.663, + 0.423 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.72 + ], + "angle": 0, + "content": "LDCT Imaging. Traditional reconstruction methods primarily rely on image sparsity [25]. However, these methods are generally time-consuming. Recently, deep learning-based methods have shown faster processing and promising performance, gaining significant attention in this field. For example, Chen et al. [4] introduced residual structure into an autoencoder and proposed the encoder-decoder convolutional neural network (RED-CNN) for LDCT denoising. You et al. [39] extended RED-CNN by incorporating a generative adversarial network (GAN) for super resolution. Additionally, researchers also explored transformer-based architectures to enhance robustness, as seen in Uformer [31], TransCT [40], and RegFormer [34]. Recently, researchers have proposed diffusion-based CT reconstruction methods [9, 21, 33], which show promising results but incur substantial computational costs. Moreover, the methods mentioned above are tailored to specific scanning protocols and depend on data collection from multiple clients, raising significant privacy concerns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Federated Learning. In recent years, FL has gained increasing attention as it offers a decentralized approach to DL while preserving data privacy [38]. FedAvg [23], the earliest FL method, introduced a server-client framework to learn a global-shared model by averaging local models from different parties. However, this method suffers from model shift due to serious statistic heterogeneity [11]. To alleviate this concern, recent works have proposed methods to obtain a more robust global-shared model, such as FedProx [16], MOON [15], and FedNova [29]. For serious non-iid data distributions, a single global-shared model is insufficient to address the variability across different datasets. As" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5155" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.898, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.335, + 0.719, + 0.35 + ], + "angle": 0, + "content": "Figure 2. The overall learning paradigm of our proposed SCAN-PhysFed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.483, + 0.387 + ], + "angle": 0, + "content": "a result, several personalized FL methods have been proposed [3, 7, 17, 35]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.387, + 0.483, + 0.584 + ], + "angle": 0, + "content": "FL has recently been introduced in healthcare applications due to its privacy-preserving nature [10, 32, 36]. Several studies have focused on its applications in medical imaging [8, 14]. For example, Chen et al. [5] proposed a cross-domain learning approach that combines individual-client sinogram learning and cross-client image reconstruction learning. Yang et al. [37] proposed a client-specific hypernetwork to modulate the global-shared data. Chen et al. [6] utilized separate networks to process high- and low-frequency components of CT data. While these methods have achieved promising performance, they still face significant challenges from data heterogeneity and fail to address the issue of unseen protocols." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.812 + ], + "angle": 0, + "content": "Large Language Model. The emergence of large language models (LLMs), such as GPT [1] and Llama [26], has ushered in a new era of artificial intelligence. Recently, largescale medical LLMs (MLLMs) have been developed specifically for healthcare applications [18, 20, 41]. These models demonstrate significant potential in understanding and reasoning over medical data, excelling in tasks such as visual question answering (VQA) and text generation [24]. For instance, Alkhaldi et al. [2] developed Minigpt-med, which supports multiple tasks, including VQA, medical report generation, and disease identification from medical images. Although MLLMs have been applied to various downstream tasks [13], they have not yet been explored for mitigating domain shift in CT reconstruction within FL settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.226, + 0.84 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.202, + 0.863 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "This paper aims to address generalization challenges in FL for LDCT reconstruction, with a focus on alleviating data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.538 + ], + "angle": 0, + "content": "heterogeneity among existing clients and maintaining performance for new clients. Figure 2 illustrates the overall learning paradigm of the proposed SCAN-PhysFed framework. Specifically, robust imaging representations are obtained using a collaboratively learned shared encoder across clients. To tackle the heterogeneity issue, two physics-informed hypernetworks are designed to capture scanning-and anatomy-specific information, personalizing the imaging features via a specialized modulation block. Subsequently, client-specific decoders project the personalized imaging features back to the image domain. For the \\(i\\)-th client, the reconstruction process is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.55, + 0.907, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Y}} = D _ {i} \\left(P \\left(f _ {\\mathbf {X}}, H _ {s} (g; \\theta_ {H _ {s}}), H _ {a} \\left(f _ {t}; \\theta_ {H _ {a}}\\right)\\right); \\theta_ {D _ {i}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.701 + ], + "angle": 0, + "content": "where \\(f_{\\mathbf{X}}\\) denotes the imaging feature of \\(\\mathbf{X}\\) extracted from the shared encoder \\(E\\), and \\(\\hat{\\mathbf{Y}}\\) is the corresponding normal-dose CT image. \\(\\theta_{H_s}\\) and \\(\\theta_{H_a}\\) are the scanning- and anatomy-informed hypernetworks, parameterized by \\(\\theta_s\\) and \\(\\theta_{a}\\), respectively. \\(g\\) and \\(f_{t}\\) represent the scanning protocol vector and the radiology report feature derived from the MLLM. \\(P\\) denotes the personalized modulation block, and \\(D_{i}\\) denotes the \\(i\\)-th client decoder, parameterized by \\(\\theta_{D_i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.906, + 0.777 + ], + "angle": 0, + "content": "Additionally, to handle unseen data, we propose a novel protocol vector-quantization strategy (PVQS), which ensures robust performance across previously unseen clients by quantizing the unseen scanning code to the closest match in the scanning codebook." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.791, + 0.802 + ], + "angle": 0, + "content": "3.2. Personalized Modulation Block" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Anatomy-Level Personalization. The patient's unique anatomy significantly affects noise distribution during the imaging process, making it an essential factor to consider. To address this issue, we leverage MLLM, which is trained on large-scale datasets and demonstrates strong performance in radiology report generation. Specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5156" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "we adapt miniGPT-Med [2] to generate radiology reports from CT images, using the text encoder's feature output \\( f_{t} \\in \\mathbb{R}^{1 \\times d} \\) as a prompt for our anatomy-informed hypernetwork, where \\( d \\) denotes the feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.197 + ], + "angle": 0, + "content": "Given the high dimensionality of \\( f_{t} \\), we first apply a pooling operation to reduce its size. The process is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.212, + 0.483, + 0.229 + ], + "angle": 0, + "content": "\\[\nf _ {t l} = \\operatorname {L i n e a r} \\left(\\operatorname {A v g P o o l} \\left(f _ {t}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.241, + 0.483, + 0.3 + ], + "angle": 0, + "content": "where \\(\\mathrm{AvgPool}\\) denotes an average pooling operation that reduces the feature dimension by a factor of \\(1 / 4\\). A linear layer is then applied to map the feature to \\(f_{tl} \\in \\mathbb{R}^{1 \\times h}\\), where \\(h\\) denotes the hidden dimension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.484, + 0.362 + ], + "angle": 0, + "content": "Next, we capture long-range dependencies within \\( f_{tl} \\) using the multi-head attention mechanism, followed by a linear layer to extract the anatomy feature \\( f_{an} \\). This process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.377, + 0.483, + 0.393 + ], + "angle": 0, + "content": "\\[\nf _ {a n} = \\operatorname {L i n e a r} (\\operatorname {M u l t i - H e a d} (f _ {t l})). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.48 + ], + "angle": 0, + "content": "After this stage, \\( f_{an} \\) encapsulates anatomy-specific information tailored to different CT data. Since this feature encodes anatomy-specific details, we use it to spatially personalize the imaging feature \\( f_{\\mathbf{X}} \\). The process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.483, + 0.483, + 0.5 + ], + "angle": 0, + "content": "\\[\nf _ {a n a} = f _ {\\mathbf {X}} \\odot \\operatorname {R e s h a p e} \\left(f _ {a n}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.568 + ], + "angle": 0, + "content": "where \\( f_{\\mathbf{X}} \\) denotes the imaging feature, Reshape represents the reshape operation for dimensional alignment, and “\\( \\odot \\)” denotes the Hadamard product. \\( f_{ana} \\) is the anatomy-level personalized feature." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.705 + ], + "angle": 0, + "content": "Scanning-Level Personalization. In addition to anatomical variations affecting noise distribution, the physical imaging process shows that noise distribution is primarily governed by the scanning protocol. This protocol comprises several key physical parameters in CT scanners, including the number of views, the number of detector bins, pixel length, detector bin length, the distance between the source and rotation center, the distance between the detector and rotation center, and the photon number of incident X-rays." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.483, + 0.781 + ], + "angle": 0, + "content": "Then, we utilize \\( g \\), a vector encoding these parameters, to construct our scanning-level prompt, which captures the latent relationship between the scanning protocol and noise distribution for the imaging network. A detailed description of \\( g \\) is provided in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.782, + 0.484, + 0.871 + ], + "angle": 0, + "content": "Given the low dimensionality of the physical parameters, we employ a Multi-Layer Perceptron (MLP) with three linear layers to capture and represent the underlying physical information. Two modulation parameter vectors are then derived to encode scanning-level personalization knowledge. This process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\alpha , \\beta = H _ {s} (g; \\theta_ {s}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are the modulation parameter vectors, which are used to inject scanning-level information to personalize the imaging feature as:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.147, + 0.905, + 0.163 + ], + "angle": 0, + "content": "\\[\nf _ {p e r} = \\alpha \\otimes f _ {a n a} + \\beta , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.171, + 0.905, + 0.232 + ], + "angle": 0, + "content": "where \\( f_{per} \\) is the dual-level personalized imaging feature, and \\( \\otimes \\) denotes the multiplication operation along the channel dimension. This approach applies the scanning prior to modulate the imaging feature." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.905, + 0.277 + ], + "angle": 0, + "content": "Our proposed personalization block is architecture-agnostic, allowing the personalization parameters to be finetuned accordingly." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.285, + 0.845, + 0.302 + ], + "angle": 0, + "content": "3.3. Protocol Vector-Quantization Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.905, + 0.459 + ], + "angle": 0, + "content": "Existing FL-based CT imaging methods have not addressed the challenge of unseen data, resulting in degraded performance in new domains. To overcome this limitation and ensure robust performance for unseen protocols, we introduce PVQS. While patient anatomy remains largely consistent, the primary source of noise variation between unseen and existing clients stems from the low-dose strategy. Therefore, PVQS is designed to enable SCAN-PhysFed to maintain consistent performance across diverse and unseen low-dose protocols." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.906, + 0.535 + ], + "angle": 0, + "content": "Specifically, in PVQS, protocol codes for different existing scanning protocols, generated by the MLP in \\( H_{s} \\), are stored in a protocol codebook. For an unseen protocol, we first quantize it to the closest code vector in the protocol codebook. The process is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.543, + 0.905, + 0.577 + ], + "angle": 0, + "content": "\\[\ni ^ {*} = \\arg \\min _ {i} \\left(1 - \\frac {\\mathbf {c} _ {u n} \\cdot \\mathbf {c} _ {i}}{\\| \\mathbf {c} _ {u n} \\| \\| \\mathbf {c} _ {i} \\|}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.627 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_{un}\\) and \\(\\mathbf{c}_i\\) denote the protocol code of the unseen protocol and that of the \\(i\\)-th client, respectively. \\(i^*\\) indicates the index of the closest code vector." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.905, + 0.688 + ], + "angle": 0, + "content": "Instead of using the unseen protocol as input to \\( H_{s} \\), we directly replace it with \\( \\mathbf{c}_{i^{*}} \\) to personalize the imaging feature and use the corresponding decoder \\( D_{i^{*}} \\) to project the personalized features back to the image domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.905, + 0.749 + ], + "angle": 0, + "content": "Thus, PVQS effectively avoids potential distribution shifts in unseen data and preserves a consistent feature space across domains, ensuring reliable and robust performance for unseen protocols." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.702, + 0.773 + ], + "angle": 0, + "content": "3.4. Learning Paradigm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Since the noise distribution varies across protocols, it is challenging to use a single shared imaging network to accommodate all variations. In FL-based LDCT imaging, as assumed in previous works [6, 37], the optimization challenge consists of two main components: imaging feature extraction and personalized projection. Specifically, for feature extraction, we propose aggregating client-side encoders to capture robust, comprehensive imaging features." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5157" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.206, + 0.089, + 0.792, + 0.104 + ], + "angle": 0, + "content": "Table 1. Quantitative Results (PSNR (dB) and SSIM \\((\\%)\\)) for CNN-based LDCT Imaging Method." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.113, + 0.907, + 0.289 + ], + "angle": 0, + "content": "
Client #1Client #2Client #3Client #4Client #5Client #6Client #7Client #8Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]29.9982.2032.8278.2334.6886.9932.3080.6336.8792.0637.3689.0232.7380.6635.4687.6534.0384.70
FedProx [16]29.9483.1232.6975.4636.9591.5832.3280.9138.9595.3437.5888.1332.9481.6235.8688.6534.6585.60
FedNova [29]29.9582.2932.5780.5737.7792.1632.0978.9834.4288.6936.0589.5732.8181.4135.8988.7133.9485.30
MOON [15]32.5075.6135.5585.4135.7280.3035.6982.1035.3878.4736.3184.3535.7081.1735.6280.4035.3180.98
FedDG [19]33.3781.4034.0182.3933.5676.3734.4381.4934.0676.7033.9780.1734.7580.8734.2579.2334.0579.83
FedKD [32]29.7483.1132.1771.6536.6590.9332.1079.1940.5896.7137.0782.5532.8881.3036.2889.0834.6884.31
Personalized FL Methods
FedPer [3]33.2388.2035.0684.8938.7992.7035.5886.1241.7696.9736.8589.9635.6986.3038.4891.6936.9389.60
FedBN [17]32.5584.5934.7684.1538.1890.3034.7783.3041.2496.1138.6691.1934.7483.3537.8990.3536.6087.92
FedMRI [8]33.2883.4334.5185.8138.6292.4735.8687.2141.4996.2738.2291.3235.8786.3438.0991.9436.9989.35
HyperFed [37]33.7488.5734.8385.0538.8792.7835.4485.4042.9497.4839.1592.9634.9985.1938.1291.2937.2689.84
FedFDD [6]36.1893.6738.4394.1939.7296.4738.5195.3045.5698.7342.7297.6140.7496.0138.5494.3940.0595.79
SCAN-PhysFed34.4996.4039.2196.0940.4597.9340.5197.6147.8199.1444.0898.5142.5497.8640.0297.3741.1497.62
" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.304, + 0.161, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.352, + 0.138, + 0.358 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.163, + 0.304, + 0.225, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.352, + 0.2, + 0.358 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.304, + 0.287, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.352, + 0.239, + 0.358 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.304, + 0.349, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.352, + 0.325, + 0.358 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.304, + 0.412, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.352, + 0.388, + 0.358 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.304, + 0.474, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.414, + 0.352, + 0.449, + 0.358 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.358, + 0.162, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.406, + 0.138, + 0.414 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.358, + 0.225, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.406, + 0.201, + 0.414 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.358, + 0.287, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.406, + 0.264, + 0.414 + ], + "angle": 0, + "content": "(i)" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.358, + 0.349, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.406, + 0.326, + 0.414 + ], + "angle": 0, + "content": "(1)" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.358, + 0.412, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.406, + 0.388, + 0.414 + ], + "angle": 0, + "content": "(k)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.421, + 0.483, + 0.462 + ], + "angle": 0, + "content": "Figure 3. Simulated examples under different protocols. (a)-(h) show examples from training clients with known protocols, while (i)-(l) present examples from unseen protocols." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.564 + ], + "angle": 0, + "content": "To fully leverage the diversity of client data, we globally share the hypernetworks in SCAN-PhysFed. However, projecting imaging features from different domains into a consistent, noise-free representation is challenging with a single shared decoder. Therefore, we propose protocol-specific decoders to effectively adapt to variations in scanning protocols across clients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.655 + ], + "angle": 0, + "content": "In PVQS, the unseen protocol code is quantized based on the nearest entry in the protocol codebook. Therefore, it is essential to ensure the protocol code is discriminative and informative for personalization. To achieve this, we introduce an orthogonal loss \\(\\mathcal{L}_{\\mathrm{orth}}\\) for the \\(i\\)-th client, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.667, + 0.483, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o r t h}} = \\sum_ {j = 1} ^ {K} \\left| \\mathbf {c} _ {i} \\cdot \\mathbf {c} _ {j} \\right| ^ {2} \\quad s. t. \\quad i \\neq j, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.483, + 0.75 + ], + "angle": 0, + "content": "where \\(K\\) is the number of clients (protocols), and \\(j\\) denotes the client index." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.483, + 0.781 + ], + "angle": 0, + "content": "The Mean Squared Error (MSE) is used as the imaging loss, and the total loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.794, + 0.483, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {M S E}} (\\mathbf {Y}, \\hat {\\mathbf {Y}}) + \\tau \\mathcal {L} _ {\\text {o r t h}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where \\(\\mathbf{Y}\\) represents the reconstructed image, and \\(\\tau\\) denotes the temperature weight of \\(\\mathcal{L}_{\\mathrm{orth}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To facilitate understanding of the implementation details, we provide the pseudocode of our method in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.3, + 0.645, + 0.317 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.905, + 0.429 + ], + "angle": 0, + "content": "Experimental Setting. The proposed method was implemented using PyTorch and optimized by Adam [12] at a learning rate of 0.001. The number of communication rounds was set to 200 and the batch size was 20. The experiments were conducted a system equipped with an AMD Ryzen 75800X CPU and a single NVIDIA GTX 3080 Ti GPU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.567 + ], + "angle": 0, + "content": "Dataset. The \"2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge\" dataset [22], which contains 5,936 full-dose CT images from 10 patients, was used to evaluate the proposed method. In this study, 8 patients were used for training, while the remaining 2 were reserved for testing. To simulate a realistic scenario, we ensured that the patients for training did not overlap across different clients. The two patients for testing were simulated using different protocols to form the testing set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.642 + ], + "angle": 0, + "content": "Simulated examples are shown in Figure 3, where noticeable noise heterogeneity among clients can be observed, primarily due to differences in anatomy and scanning protocols. Detailed dataset preparation steps and scanning protocols are provided in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.906, + 0.762 + ], + "angle": 0, + "content": "Baselines. We compared our SCAN-PhysFed approach with both generic and personalized FL methods. For generic FL methods, we include FedAvg [23], FedProx [16], FedNova [29], MOON [15], FedDG [19], and FedKD [32]. For personalized FL methods, we evaluated FedPer [3], FedBN [17], FedMRI [8], HyperFed [37], and FedFDD [6]. To ensure fairness, we kept the training settings consistent across all methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.905, + 0.837 + ], + "angle": 0, + "content": "Evaluation Metrics. We use the peak signal-to-noise ratio (PSNR) to evaluate pixel-wise accuracy and the structural similarity (SSIM) to assess perceived visual quality [30]. For both metrics, higher values indicate better performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.799, + 0.864 + ], + "angle": 0, + "content": "4.1. Comparison with other methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We evaluate our method alongside other FL methods. To ensure fairness, we use the same imaging network, RED" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5158" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.096, + 0.891, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.485, + 0.908, + 0.528 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of six selected comparison methods and our method across different clients using the classical convolutional-based LDCT imaging network. Rows one to five correspond to Clients #2, #3, #5, #6, and #7, respectively. The display window for the first row is [-1024, 200] HU, while for the remaining rows, it is [-160, 240] HU." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.543, + 0.48, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.674, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Figure 5. Boxplots of the average results across all clients. (a) and (b) represent PSNR and SSIM, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "CNN [4], a state-of-the-art model in CT imaging. Quantitative results are provided in Table 1. Our method achieves the best performance across most clients, outperforming other generic and personalized FL methods in terms of average PSNR and SSIM. Furthermore, unlike other personalized methods that show instability across clients, such as HyperFed in Client #2 / #7 and FedFDD in Client #4, our method maintains consistent performance across varying protocols. We also observe that sparse-view protocols (Clients #2, #4, and #6) pose challenges for both generic and personalized FL methods, resulting in significant performance gaps compared to other protocols. How-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.907, + 0.571 + ], + "angle": 0, + "content": "ever, by incorporating dual-level physics-informed prior, our method maintains robust performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.82 + ], + "angle": 0, + "content": "Representative slices produced by different methods are shown in Figure 4. We compared the top three methods from both generic and personalized FL approaches. Representative clients were selected based on different radiation reduction strategies: sparse-view (Clients #2 and #6) and low-dose (Clients #3, #5, and #7) reconstructions. It can be observed that other methods exhibit noticeable noise or artifacts in sparse-view protocols, and fail to preserve details in low-dose protocols. In contrast, our method effectively reduces noise and artifacts in sparse-view protocols while preserving fine details in low-dose protocols. SCAN-PhysFed achieves strong performance across slices from different anatomical regions, benefiting from the introduction of anatomy priors. Both quantitative and qualitative results demonstrate the effectiveness of incorporating physics-informed priors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Additionally, we present boxplots of the results across all clients for different methods in Figure 5 to validate their stability. It can be seen that SCAN-PhysFed outperforms all other methods in both PSNR and SSIM scores. In both plots, SCAN-PhysFed demonstrates a higher median and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5159" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.208, + 0.09, + 0.79, + 0.104 + ], + "angle": 0, + "content": "Table 2. Quantitative Results of PSNR and SSIM for Transformer-based LDCT Imaging Methods." + }, + { + "type": "table", + "bbox": [ + 0.087, + 0.113, + 0.907, + 0.643 + ], + "angle": 0, + "content": "
Client #1Client #2Client #3Client #4Client #5Client #6Client #7Client #8Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]34.3187.8838.0690.5443.5197.5641.0595.4345.2498.4743.1597.0941.4595.9542.8896.1641.2194.89
FedProx [16]34.3287.0137.7488.8943.7397.9541.1293.9245.3998.6142.9997.0841.3894.3142.1094.5541.1094.04
FedNova [29]29.5578.4431.3868.1337.0782.4731.7973.4540.6188.8035.4679.7732.3873.6635.7782.2234.2578.37
MOON [15]34.2893.2038.0991.1743.6497.7741.0995.6645.2598.4342.8296.9141.4295.9742.2096.1841.1095.66
FedKD [32]35.9289.9338.7091.2144.2698.1342.3095.9945.9398.7543.5797.4142.5496.2243.5196.2342.0995.48
Personalized FL Methods
FedPer [3]35.3793.1839.0393.5542.7597.1340.8896.2046.5598.8342.8796.6841.8796.7040.7994.7841.2695.82
FedMRI [8]34.6093.6738.0494.5542.9497.4439.3495.7846.7998.9142.8297.3040.8195.3240.8595.7040.7896.01
HyperFed [37]34.6087.5938.6291.0844.1397.6142.2596.2445.9298.5943.7197.3742.4496.6643.5696.6341.9095.22
FedFDD [6]37.3596.2039.6795.1844.3198.2639.0095.2147.8999.2043.8198.0642.1797.4040.5095.4941.8496.89
SCAN-PhysFed38.5594.0641.7496.2045.5298.5242.4896.5549.5099.3945.3297.9343.3596.7342.7196.7943.6597.03
Ground TruthInputFedAvgMOONFedKDFedMRIHyperFedFedFDDSCAN-PhysFed
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.649, + 0.907, + 0.678 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of the compared methods across different clients based on a transformer-based imaging network. Rows one to five correspond to Clients #2, #3, #5, #6, and #7, respectively. The display window is set to [-160, 240] HU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.483, + 0.752 + ], + "angle": 0, + "content": "a narrower interquartile range, indicating strong stability in restoring LDCT data with varying noise distributions across different clients. Individual boxplots for each client are provided in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.762, + 0.33, + 0.777 + ], + "angle": 0, + "content": "4.2. Generalization Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.786, + 0.216, + 0.8 + ], + "angle": 0, + "content": "4.2.1 Backbone" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.9 + ], + "angle": 0, + "content": "In addition to using RED-CNN as the imaging network in previous experiments, we evaluate our method and other FL methods with an alternative imaging network, Uformer [31], a transformer-based model for LDCT imaging. This comparison further validates the generalization capability of different FL methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.907, + 0.889 + ], + "angle": 0, + "content": "The qualitative and quantitative results are shown in Figure 6 and Table 2, respectively. Consistent with previous experiments, our method achieves the best performance compared to other methods across most clients, with significantly higher average PSNR and SSIM values. This demonstrates the strong generalization capability of our method when integrated with different imaging networks. Among competing methods, FedFDD serves as a strong baseline. While FedFDD's SSIM is slightly higher than ours in some clients, our method achieves a significantly better PSNR, demonstrating superior detail preservation. Additionally, unlike FedFDD, which requires two imaging networks, our method achieves high performance with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5160" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.27, + 0.089, + 0.726, + 0.103 + ], + "angle": 0, + "content": "Table 3. The Quantitative Results of PSNR and SSIM for the Unseen Clients." + }, + { + "type": "table", + "bbox": [ + 0.158, + 0.114, + 0.83, + 0.297 + ], + "angle": 0, + "content": "
Unseen Client #1Unseen Client #2Unseen Client #3Unseen Client #4Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]35.5788.6834.0085.2632.5276.7136.8691.4432.8981.35
FedProx [16]36.6991.1634.3086.3932.5872.1339.1394.8734.7485.52
FedNova [29]35.4588.8334.2386.2932.3969.2035.4089.9934.3883.58
MOON [15]35.2178.1935.3079.8134.5583.0735.5678.9435.1680.00
FedDG [19]34.8679.5434.7980.9133.1580.3134.4977.4234.3279.55
FedKD [32]36.8291.6034.2786.3732.4970.5440.7395.9036.0886.10
Personalized FL Methods
FedPer [3]36.8688.5835.4687.0231.9671.6438.5690.9035.7184.53
FedBN [17]37.8489.9536.1688.4633.6781.6340.2194.8936.9788.73
FedMRI [8]31.1282.4231.9579.9531.7271.9632.0481.6331.7178.99
HyperFed [37]36.1889.6033.4782.4931.9465.5241.7094.8535.9083.12
FedFDD [6]37.8092.3337.0691.7733.7180.2640.5395.1237.2789.87
SCAN-PhysFed40.7797.6938.7795.7534.3881.8040.2796.9238.5593.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.309, + 0.472, + 0.323 + ], + "angle": 0, + "content": "Table 4. Ablation on different components in SCAN-PhysFed." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.334, + 0.482, + 0.424 + ], + "angle": 0, + "content": "
ScanningComponentAverage
AnatomyParadigmLossPSNRSSIM
XXGenericLMSE34.0384.70
✓XPersonalizedLMSE36.8288.78
X✓PersonalizedLMSE38.0494.14
✓✓PersonalizedLMSE40.6797.27
✓✓PersonalizedLMSE & Lorth41.1497.62
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.483, + 0.51 + ], + "angle": 0, + "content": "out relying on an additional network. Similar to the previous experiment, SCAN-PhysFed shows better qualitative performance and lower variation across different protocols, indicating more consistent results irrespective of scanning protocols or imaging networks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.529, + 0.26, + 0.543 + ], + "angle": 0, + "content": "4.2.2 Unseen Protocol" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.705 + ], + "angle": 0, + "content": "Existing FL methods in medical imaging struggle to maintain performance when encountering unseen noise distributions from new clients. Here, we evaluate the effectiveness of the proposed PVQS across four unseen protocols. Detailed protocol settings can be found in Appendix B. For generic FL methods, we directly evaluate the shared model on unseen clients. However, personalized methods mostly ignore how to operate in the unseen domain, so we average the performance of all local models to establish a vanilla baseline in this paper." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The results of different methods on unseen clients are presented in Table 3. We observe that the performance of generic FL methods shows only minor degradation on unseen clients, whereas personalized FL methods experience significant declines. This is because local optimization in personalized FL methods focuses heavily on each client's specific data distribution, underfitting other potential distributions and leading to greater performance drops on unseen clients. In PVQS, we prevent the network from directly mapping the unseen protocol to modulate the imaging network, as this could cause network collapse if the unseen protocol significantly differs from known ones. Instead, we quantize the unseen protocol using a protocol codebook, al" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.905, + 0.34 + ], + "angle": 0, + "content": "lowing SCAN-PhysFed to leverage prior knowledge for better adaptation to unseen distributions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.352, + 0.667, + 0.369 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.619 + ], + "angle": 0, + "content": "We evaluate the effectiveness of various components in SCAN-PhysFed, and the results are presented in Table 4. Compared to the baseline results, which exclude scanning and anatomy information, introducing either significantly improves performance, highlighting the importance of integrating physical prior in LDCT imaging. This approach substantially reduces learning complexity and improves the overall model performance. Moreover, introducing both types of physical information simultaneously is complementary rather than redundant, leading to further performance gains. Building on this, we use \\(\\mathcal{L}_{\\mathrm{orth}}\\) to enhance the discriminative ability of the scanning code, thereby personalizing imaging features and effectively mitigating data heterogeneity. In the above studies, \"Personalized\" indicates that the decoder is tailored to each client. An ablation study on the personalized components is provided in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To preserve privacy and address heterogeneity in CT imaging, we propose a dual-level, physics-informed approach that utilizes both scanning- and anatomy-level prompts to enable personalized CT imaging with the assistance of an MLLM. This approach facilitates both patient- and scanning-level personalization, with extensive results demonstrating that incorporating physical prior significantly improves performance and effectively mitigates heterogeneity. Additionally, we propose a novel strategy, PVQS, to quantize protocol codes and maintain robust performance on unseen protocols. PVQS leverages known protocol codes to prevent latent collapse when encountering significantly different unseen protocols. In future work, we aim to develop a general FL framework for medical imaging that can support both CT and MRI denoising, which we believe holds promising potential." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.513, + 0.957 + ], + "angle": 0, + "content": "5161" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.206 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62271335, in part by the Sichuan Science and Technology Program under Grant 2025ZNSFSC0470, and in part by Sichuan University \"From 0 to 1\" Innovative Research Program under Grant 2022SCUH0016." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.222, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.247, + 0.484, + 0.316 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.318, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[2] Asma Alkhaldi, Raneem Alnajim, Layan Alabdullatef, Rawan Alyahya, Jun Chen, Deyao Zhu, Ahmed Alsinan, and Mohamed Elhoseiny. Minigpt-med: Large language model as a general interface for radiology diagnosis. arXiv preprint arXiv:2407.04106, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[3] Manoj Ghuhan Arivazhagan, Vinay Aggarwal, Aaditya Kumar Singh, et al. Federated learning with personalization layers. arXiv preprint arXiv:1912.00818, 2019. 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.483, + 0.486 + ], + "angle": 0, + "content": "[4] Hu Chen, Yi Zhang, Mannudeep K Kalra, et al. Low-dose ct with a residual encoder-decoder convolutional neural network. IEEE Trans. Med. Imaging, 36(12):2524-2535, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[5] Shixuan Chen, Boxuan Cao, Yinda Du, Yaoduo Zhang, Ji He, Zhaoying Bian, Dong Zeng, and Jianhua Ma. Federated condition generalization on low-dose ct reconstruction via cross-domain learning. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 47-56. Springer, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[6] Xuhang Chen, Zeju Li, Zikun Xu, Cheng Ouyang, Chen Qin, et al. Fedfdd: Federated learning with frequency domain decomposition for low-dose ct denoising. In Medical Imaging with Deep Learning, 2024. 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[7] Dongshang Deng, Xuangou Wu, Tao Zhang, Xiangyun Tang, Hongyang Du, Jiawen Kang, Jiqiang Liu, and Dusit Niyato. Fedasa: A personalized federated learning with adaptive model aggregation for heterogeneous mobile edge computing. IEEE Transactions on Mobile Computing, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[8] Chun-Mei Feng, Yunlu Yan, Shanshan Wang, Yong Xu, Ling Shao, and Huazhu Fu. Specificity-preserving federated learning for mr image reconstruction. IEEE Transactions on Medical Imaging, 42(7):2010-2021, 2022. 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[9] Qi Gao, Zilong Li, Junping Zhang, Yi Zhang, and Hongming Shan. Corediff: Contextual error-modulated generalized diffusion model for low-dose ct denoising and generalization. IEEE Transactions on Medical Imaging, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[10] Hao Guan, Pew-Thian Yap, Andrea Bozoki, and Mingxia Liu. Federated learning for medical image analysis: A survey. Pattern Recognition, page 110424, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for feder-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.247, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "ated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[13] Gurucharan Marthi Krishna Kumar, Aman Chadha, Janine Mendola, and Amir Shmuel. Medvisionllama: Leveraging pre-trained large language model layers to enhance medical image segmentation. arXiv preprint arXiv:2410.02458, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[14] Danyang Li, Hao Wang, Jingyi Liao, Mingqiang Meng, Cuidie Zeng, Yuting Wang, Dong Zeng, and Jianhua Ma. Semi-centralized federated learning network for low-dose ct imaging. In Medical Imaging 2023: Physics of Medical Imaging, pages 1024-1028. SPIE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[15] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 10713–10722, 2021. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.904, + 0.405 + ], + "angle": 0, + "content": "[16] Tian Li, Anit Kumar Sahu, Manzil Zaheer, et al. Federated optimization in heterogeneous networks. In Proc. Mach. Learn. Syst. (MLSys), pages 429-450, 2020. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[17] Xiaoxiao Li, Meirui JIANG, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. In International Conference on Learning Representations, 2021. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.464, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[18] Fenglin Liu, Tingting Zhu, Xian Wu, Bang Yang, Chenyu You, Chenyang Wang, Lei Lu, Zhangdaihong Liu, Yefeng Zheng, Xu Sun, et al. A medical multimodal large language model for future pandemics. NPJ Digital Medicine, 6(1): 226, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[19] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1013–1023, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[20] Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Derong Xu, Feng Tian, and Yefeng Zheng. When moe meets llms: Parameter efficient fine-tuning for multi-task medical applications. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1114, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[21] Zexin Lu, Qi Gao, Tao Wang, Ziyuan Yang, Zhiwen Wang, Hui Yu, Hu Chen, Jiliu Zhou, Hongming Shan, and Yi Zhang. Pridediff: Physics-regularized generalized diffusion model for ct reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[22] C McCollough. Tu-fg-207a-04: overview of the low dose ct grand challenge. Med. Phys., 43(6):3759–3760, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[23] Brendan McMahan, Eider Moore, Daniel Ramage, et al. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medi" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5162" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "cal artificial intelligence. Nature, 616(7956):259-265, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[25] Saiprasad Ravishankar, Jong Chul Ye, and Jeffrey A Fessler. Image reconstruction: From sparsity to data-adaptive methods and machine learning. Proceedings of the IEEE, 108(1): 86-109, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[26] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[27] Ge Wang, Hengyong Yu, and Bruno De Man. An outlook on x-ray ct research and development. Medical physics, 35(3): 1051-1064, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[28] Ge Wang, Jong Chul Ye, and Bruno De Man. Deep learning for tomographic image reconstruction. Nature machine intelligence, 2(12):737-748, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.336, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[29] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[31] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[32] Chuhan Wu, Fangzhao Wu, Lingjuan Lyu, Yongfeng Huang, and Xing Xie. Communication-efficient federated learning via knowledge distillation. Nature Communications, 13(1): 2032, 2022. 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.632 + ], + "angle": 0, + "content": "[33] Wenjun Xia, Qing Lyu, and Ge Wang. Low-dose ct using denoising diffusion probabilistic model for 20x speedup. arXiv preprint arXiv:2209.15136, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.634, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[34] Wenjun Xia, Ziyuan Yang, Zexin Lu, Zhongxian Wang, and Yi Zhang. Regformer: A local-nonlocal regularization-based model for sparse-view ct reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences, 8(2): 184-194, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[35] Xiyuan Yang, Wenke Huang, and Mang Ye. Fedas: Bridging inconsistency in personalized federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11986-11995, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[36] Ziyuan Yang, Yingyu Chen, Huijie Huangfu, Maosong Ran, Hui Wang, Xiaoxiao Li, and Yi Zhang. Dynamic corrected split federated learning with homomorphic encryption for u-shaped medical image networks. IEEE Journal of Biomedical and Health Informatics, 27(12):5946-5957, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[37] Ziyuan Yang, Wenjun Xia, Zexin Lu, Yingyu Chen, Xiaoxiao Li, and Yi Zhang. Hypernetwork-based physics-driven personalized federated learning for ct imaging. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3, 4, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[38] Mang Ye, Xiuwen Fang, Bo Du, Pong C Yuen, and Dacheng Tao. Heterogeneous federated learning: State-of-the-art and research challenges. ACM Computing Surveys, 56(3):1-44, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[39] Chenyu You, Guang Li, Yi Zhang, et al. Ct super-resolution gan constrained by the identical, residual, and cycle learning ensemble (gan-circle). IEEE Trans. Med. Imaging, 39(1): 188-203, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[40] Zhicheng Zhang, Lequan Yu, Xiaokun Liang, Wei Zhao, and Lei Xing. Transc: dual-path transformer for low dose computed tomography. In Medical Image Computing and Computer Assisted Intervention, pages 55–64. Springer, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[41] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large language models meet few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3065-3075, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5163" + } + ] +] \ No newline at end of file diff --git a/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/b8f485b4-2d7f-474b-acf2-8dc4a4e3047c_origin.pdf b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/b8f485b4-2d7f-474b-acf2-8dc4a4e3047c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8eaa0fbc06773fb47a665e87e4f2e0e212384c10 --- /dev/null +++ b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/b8f485b4-2d7f-474b-acf2-8dc4a4e3047c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab3c4a39320ad9e16f8457d54782515b6b1832cdeeaf90349e8f2cab1b664754 +size 2645139 diff --git a/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/full.md b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6f46108eae55308511366c03b517c32122fd8476 --- /dev/null +++ b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/full.md @@ -0,0 +1,316 @@ +# Patient-Level Anatomy Meets Scanning-Level Physics: Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model + +Ziyuan Yang $^{1,5}$ , Yingyu Chen $^{1,5}$ , Zhiwen Wang $^{2}$ , Hongming Shan $^{3}$ , Yang Chen $^{4}$ , Yi Zhang $^{1,5,*}$ $^{1}$ Sichuan University $^{2}$ Southwest Petroleum University $^{3}$ Fudan University $^{4}$ Southeast University $^{5}$ Tianfu Jiangxi Laboratory +{cziyuanyang, cyy262511, zwwang1228}@gmail.com, hmshan@fudan.edu.cn +chenyang.list@seu.edu.cn, yzhang@scu.edu.cn + +# Abstract + +Reducing radiation doses benefits patients, but the resultant low-dose computed tomography (LDCT) images often suffer from clinically unacceptable noise and artifacts. While deep learning (DL) has shown promise in LDCT reconstruction, it requires large-scale data collection from multiple clients, raising privacy concerns. Federated learning (FL) has been introduced to mitigate these privacy concerns; however, current methods are typically tailored to specific scanning protocols, which limits their generalizability and makes them less effective for unseen protocols. To address these issues, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Since the noise distribution in LDCT data is closely tied to scanning protocols and anatomical structures, we propose a dual-level physics-informed way to address these challenges. Specifically, we incorporate physical and anatomical prompts into our physics-informed hypernetworks to capture scanning- and anatomy-specific information, enabling dual-level physics-driven personalization of imaging features. These prompts are derived from the scanning protocol and the radiology report generated by a medical large language model (MLLM). Subsequently, client-specific decoders project these dual-level personalized imaging features back into the image domain. Besides, to tackle the challenge of unseen data, we introduce a novel protocol vector-quantization strategy (PVQS), which ensures consistent performance across new clients by quantifying unseen scanning codes to the closest match in the scanning codebook. Extensive experimental results demonstrate the superior performance of SCAN-PhysFed on pub + +lic datasets1. + +# 1. Introduction + +Computed tomography (CT) is a key diagnostic tool that non-invasively visualizes anatomical structures within the human body. Despite its clinical benefits, CT scans pose radiation concerns, which potentially lead to genetic, cancerous, and other health diseases. Lowering CT radiation dose—either by decreasing incident photons or limiting sampling views—has emerged as a promising solution to reduce radiation risks and accelerate scanning. However, reconstructed images under these conditions suffer from significant quality degradation, compromising clinical applicability [27]. + +Recently, deep learning (DL)-based methods have shown great promise for low-dose CT (LDCT) imaging [28]. However, most DL-based LDCT reconstruction methods are condition-specific and overlook privacy concerns. Recently, researchers have attempted to introduce federated learning (FL), a privacy-preserving distributed learning paradigm, for LDCT reconstruction to alleviate privacy concerns. Although these methods aim to enhance generalization, they still face significant challenges arising from data heterogeneity, caused by diverse noise distributions in CT imaging across different protocols and anatomical structures. We believe that the further advancement of FL in LDCT reconstruction urgently requires a new framework capable of leveraging comprehensive physical information from the scanning process, thereby breaking through current bottlenecks to achieve better results. + +To tackle this issue, we aim to leverage physical principles of the imaging process. During the projection process, the measurement is influenced by the anatomical structures the X-ray passed through, such as fat, bone, and fluid, each + +![](images/60200b64a90c355528f48b7d4824df9a108087adeb73fc4b6e5254a5d9bc00eb.jpg) +Figure 1. Illustration of different radiation reduction strategies. 'Sparse-View' refers to scanning with a reduced number of projection views, while 'Low-Dose' represents scanning with a lower number of incident photons. + +with distinct attenuation coefficients. Additionally, the system matrix is significantly affected by scanning protocols, including tube current, voltage settings, and scanning angles. These factors collectively impact the quality and characteristics of reconstructed CT images during the backprojection phase, resulting in varying levels of noise and artifacts. An example is provided in Figure 1 to illustrate this concept more clearly. + +By leveraging the physics-specific knowledge discussed earlier, we reimagine LDCT reconstruction within the federated learning framework. Our approach adopts a dual-level strategy to alleviate heterogeneity, addressing variations at both the scanning and patient levels. In practice, scanning protocols, which serve as scanner hyperparameters, are readily available. However, generating anatomical information has traditionally been challenging due to patient variability and the complexity of accurately modeling different tissue types. Recent advances in large language models (LLMs) offer new possibilities for automatically generating anatomical information, potentially improving the reconstruction. + +Building on this, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Specifically, we address the data heterogeneous using a dual-level strategy. A detailed radiology report is generated using a pretrained LLM. Then, a patient-level anatomy-informed hypernetwork is designed to generate a modulation map from the radiology report, which seamlessly integrates with imaging features and aligns them with each patient's unique anatomical structures. Simultaneously, scanning features are integrated via a scanning-informed hypernetwork to personalize imaging features based on the protocol. To ensure distinguishability of scanning feature vectors, we impose an orthogonal loss during training to reinforce personalization. This dual-modulation strategy facilitates personalized CT imaging at both patient and scanning levels, guided by the principles of physical imaging. + +Additionally, existing personalized FL CT imaging methods do not address the challenge posed by unseen clients with varying scanning protocols. In this paper, we propose a Protocol Vector Quantization Strategy (PVQS) to quantize unseen scanning protocol by matching it with the closest existing vectors, thereby selecting the corresponding client-specific decoder. The main contributions of this paper are summarized as follows: + +- We introduce LLM into LDCT reconstruction within the FL framework, proposing SCAN-PhysFed. To the best of our knowledge, this is the first work to apply LLM in this field. +- To address the challenge of data heterogeneity, we draw inspiration from physical imaging principles and propose a dual-level approach, achieving personalized CT imaging at both the scanning and patient levels. +- We introduce a novel strategy, PVQS, to ensure our method remains effective for unseen clients with diverse scanning protocols—a critical challenge often overlooked in previous works. + +# 2. Related Works + +LDCT Imaging. Traditional reconstruction methods primarily rely on image sparsity [25]. However, these methods are generally time-consuming. Recently, deep learning-based methods have shown faster processing and promising performance, gaining significant attention in this field. For example, Chen et al. [4] introduced residual structure into an autoencoder and proposed the encoder-decoder convolutional neural network (RED-CNN) for LDCT denoising. You et al. [39] extended RED-CNN by incorporating a generative adversarial network (GAN) for super resolution. Additionally, researchers also explored transformer-based architectures to enhance robustness, as seen in Uformer [31], TransCT [40], and RegFormer [34]. Recently, researchers have proposed diffusion-based CT reconstruction methods [9, 21, 33], which show promising results but incur substantial computational costs. Moreover, the methods mentioned above are tailored to specific scanning protocols and depend on data collection from multiple clients, raising significant privacy concerns. + +Federated Learning. In recent years, FL has gained increasing attention as it offers a decentralized approach to DL while preserving data privacy [38]. FedAvg [23], the earliest FL method, introduced a server-client framework to learn a global-shared model by averaging local models from different parties. However, this method suffers from model shift due to serious statistic heterogeneity [11]. To alleviate this concern, recent works have proposed methods to obtain a more robust global-shared model, such as FedProx [16], MOON [15], and FedNova [29]. For serious non-iid data distributions, a single global-shared model is insufficient to address the variability across different datasets. As + +![](images/c60c76674ed7c98a49b9ac68176ac7dad9a3cac2e9e5719c8a8f7e23316603f2.jpg) +Figure 2. The overall learning paradigm of our proposed SCAN-PhysFed. + +a result, several personalized FL methods have been proposed [3, 7, 17, 35]. + +FL has recently been introduced in healthcare applications due to its privacy-preserving nature [10, 32, 36]. Several studies have focused on its applications in medical imaging [8, 14]. For example, Chen et al. [5] proposed a cross-domain learning approach that combines individual-client sinogram learning and cross-client image reconstruction learning. Yang et al. [37] proposed a client-specific hypernetwork to modulate the global-shared data. Chen et al. [6] utilized separate networks to process high- and low-frequency components of CT data. While these methods have achieved promising performance, they still face significant challenges from data heterogeneity and fail to address the issue of unseen protocols. + +Large Language Model. The emergence of large language models (LLMs), such as GPT [1] and Llama [26], has ushered in a new era of artificial intelligence. Recently, largescale medical LLMs (MLLMs) have been developed specifically for healthcare applications [18, 20, 41]. These models demonstrate significant potential in understanding and reasoning over medical data, excelling in tasks such as visual question answering (VQA) and text generation [24]. For instance, Alkhaldi et al. [2] developed Minigpt-med, which supports multiple tasks, including VQA, medical report generation, and disease identification from medical images. Although MLLMs have been applied to various downstream tasks [13], they have not yet been explored for mitigating domain shift in CT reconstruction within FL settings. + +# 3. Methodology + +# 3.1. Overview + +This paper aims to address generalization challenges in FL for LDCT reconstruction, with a focus on alleviating data + +heterogeneity among existing clients and maintaining performance for new clients. Figure 2 illustrates the overall learning paradigm of the proposed SCAN-PhysFed framework. Specifically, robust imaging representations are obtained using a collaboratively learned shared encoder across clients. To tackle the heterogeneity issue, two physics-informed hypernetworks are designed to capture scanning-and anatomy-specific information, personalizing the imaging features via a specialized modulation block. Subsequently, client-specific decoders project the personalized imaging features back to the image domain. For the $i$ -th client, the reconstruction process is formulated as follows: + +$$ +\hat {\mathbf {Y}} = D _ {i} \left(P \left(f _ {\mathbf {X}}, H _ {s} (g; \theta_ {H _ {s}}), H _ {a} \left(f _ {t}; \theta_ {H _ {a}}\right)\right); \theta_ {D _ {i}}\right), \tag {1} +$$ + +where $f_{\mathbf{X}}$ denotes the imaging feature of $\mathbf{X}$ extracted from the shared encoder $E$ , and $\hat{\mathbf{Y}}$ is the corresponding normal-dose CT image. $\theta_{H_s}$ and $\theta_{H_a}$ are the scanning- and anatomy-informed hypernetworks, parameterized by $\theta_s$ and $\theta_{a}$ , respectively. $g$ and $f_{t}$ represent the scanning protocol vector and the radiology report feature derived from the MLLM. $P$ denotes the personalized modulation block, and $D_{i}$ denotes the $i$ -th client decoder, parameterized by $\theta_{D_i}$ . + +Additionally, to handle unseen data, we propose a novel protocol vector-quantization strategy (PVQS), which ensures robust performance across previously unseen clients by quantizing the unseen scanning code to the closest match in the scanning codebook. + +# 3.2. Personalized Modulation Block + +Anatomy-Level Personalization. The patient's unique anatomy significantly affects noise distribution during the imaging process, making it an essential factor to consider. To address this issue, we leverage MLLM, which is trained on large-scale datasets and demonstrates strong performance in radiology report generation. Specifically, + +we adapt miniGPT-Med [2] to generate radiology reports from CT images, using the text encoder's feature output $f_{t} \in \mathbb{R}^{1 \times d}$ as a prompt for our anatomy-informed hypernetwork, where $d$ denotes the feature dimension. + +Given the high dimensionality of $f_{t}$ , we first apply a pooling operation to reduce its size. The process is formulated as follows: + +$$ +f _ {t l} = \operatorname {L i n e a r} \left(\operatorname {A v g P o o l} \left(f _ {t}\right)\right), \tag {2} +$$ + +where $\mathrm{AvgPool}$ denotes an average pooling operation that reduces the feature dimension by a factor of $1 / 4$ . A linear layer is then applied to map the feature to $f_{tl} \in \mathbb{R}^{1 \times h}$ , where $h$ denotes the hidden dimension. + +Next, we capture long-range dependencies within $f_{tl}$ using the multi-head attention mechanism, followed by a linear layer to extract the anatomy feature $f_{an}$ . This process is formulated as: + +$$ +f _ {a n} = \operatorname {L i n e a r} (\operatorname {M u l t i - H e a d} (f _ {t l})). \tag {3} +$$ + +After this stage, $f_{an}$ encapsulates anatomy-specific information tailored to different CT data. Since this feature encodes anatomy-specific details, we use it to spatially personalize the imaging feature $f_{\mathbf{X}}$ . The process is formulated as: + +$$ +f _ {a n a} = f _ {\mathbf {X}} \odot \operatorname {R e s h a p e} \left(f _ {a n}\right), \tag {4} +$$ + +where $f_{\mathbf{X}}$ denotes the imaging feature, Reshape represents the reshape operation for dimensional alignment, and “ $\odot$ ” denotes the Hadamard product. $f_{ana}$ is the anatomy-level personalized feature. + +Scanning-Level Personalization. In addition to anatomical variations affecting noise distribution, the physical imaging process shows that noise distribution is primarily governed by the scanning protocol. This protocol comprises several key physical parameters in CT scanners, including the number of views, the number of detector bins, pixel length, detector bin length, the distance between the source and rotation center, the distance between the detector and rotation center, and the photon number of incident X-rays. + +Then, we utilize $g$ , a vector encoding these parameters, to construct our scanning-level prompt, which captures the latent relationship between the scanning protocol and noise distribution for the imaging network. A detailed description of $g$ is provided in Appendix B. + +Given the low dimensionality of the physical parameters, we employ a Multi-Layer Perceptron (MLP) with three linear layers to capture and represent the underlying physical information. Two modulation parameter vectors are then derived to encode scanning-level personalization knowledge. This process can be formulated as follows: + +$$ +\alpha , \beta = H _ {s} (g; \theta_ {s}), \tag {5} +$$ + +where $\alpha$ and $\beta$ are the modulation parameter vectors, which are used to inject scanning-level information to personalize the imaging feature as: + +$$ +f _ {p e r} = \alpha \otimes f _ {a n a} + \beta , \tag {6} +$$ + +where $f_{per}$ is the dual-level personalized imaging feature, and $\otimes$ denotes the multiplication operation along the channel dimension. This approach applies the scanning prior to modulate the imaging feature. + +Our proposed personalization block is architecture-agnostic, allowing the personalization parameters to be finetuned accordingly. + +# 3.3. Protocol Vector-Quantization Strategy + +Existing FL-based CT imaging methods have not addressed the challenge of unseen data, resulting in degraded performance in new domains. To overcome this limitation and ensure robust performance for unseen protocols, we introduce PVQS. While patient anatomy remains largely consistent, the primary source of noise variation between unseen and existing clients stems from the low-dose strategy. Therefore, PVQS is designed to enable SCAN-PhysFed to maintain consistent performance across diverse and unseen low-dose protocols. + +Specifically, in PVQS, protocol codes for different existing scanning protocols, generated by the MLP in $H_{s}$ , are stored in a protocol codebook. For an unseen protocol, we first quantize it to the closest code vector in the protocol codebook. The process is defined as follows: + +$$ +i ^ {*} = \arg \min _ {i} \left(1 - \frac {\mathbf {c} _ {u n} \cdot \mathbf {c} _ {i}}{\| \mathbf {c} _ {u n} \| \| \mathbf {c} _ {i} \|}\right), \tag {7} +$$ + +where $\mathbf{c}_{un}$ and $\mathbf{c}_i$ denote the protocol code of the unseen protocol and that of the $i$ -th client, respectively. $i^*$ indicates the index of the closest code vector. + +Instead of using the unseen protocol as input to $H_{s}$ , we directly replace it with $\mathbf{c}_{i^{*}}$ to personalize the imaging feature and use the corresponding decoder $D_{i^{*}}$ to project the personalized features back to the image domain. + +Thus, PVQS effectively avoids potential distribution shifts in unseen data and preserves a consistent feature space across domains, ensuring reliable and robust performance for unseen protocols. + +# 3.4. Learning Paradigm + +Since the noise distribution varies across protocols, it is challenging to use a single shared imaging network to accommodate all variations. In FL-based LDCT imaging, as assumed in previous works [6, 37], the optimization challenge consists of two main components: imaging feature extraction and personalized projection. Specifically, for feature extraction, we propose aggregating client-side encoders to capture robust, comprehensive imaging features. + +Table 1. Quantitative Results (PSNR (dB) and SSIM $(\%)$ ) for CNN-based LDCT Imaging Method. + +
Client #1Client #2Client #3Client #4Client #5Client #6Client #7Client #8Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]29.9982.2032.8278.2334.6886.9932.3080.6336.8792.0637.3689.0232.7380.6635.4687.6534.0384.70
FedProx [16]29.9483.1232.6975.4636.9591.5832.3280.9138.9595.3437.5888.1332.9481.6235.8688.6534.6585.60
FedNova [29]29.9582.2932.5780.5737.7792.1632.0978.9834.4288.6936.0589.5732.8181.4135.8988.7133.9485.30
MOON [15]32.5075.6135.5585.4135.7280.3035.6982.1035.3878.4736.3184.3535.7081.1735.6280.4035.3180.98
FedDG [19]33.3781.4034.0182.3933.5676.3734.4381.4934.0676.7033.9780.1734.7580.8734.2579.2334.0579.83
FedKD [32]29.7483.1132.1771.6536.6590.9332.1079.1940.5896.7137.0782.5532.8881.3036.2889.0834.6884.31
Personalized FL Methods
FedPer [3]33.2388.2035.0684.8938.7992.7035.5886.1241.7696.9736.8589.9635.6986.3038.4891.6936.9389.60
FedBN [17]32.5584.5934.7684.1538.1890.3034.7783.3041.2496.1138.6691.1934.7483.3537.8990.3536.6087.92
FedMRI [8]33.2883.4334.5185.8138.6292.4735.8687.2141.4996.2738.2291.3235.8786.3438.0991.9436.9989.35
HyperFed [37]33.7488.5734.8385.0538.8792.7835.4485.4042.9497.4839.1592.9634.9985.1938.1291.2937.2689.84
FedFDD [6]36.1893.6738.4394.1939.7296.4738.5195.3045.5698.7342.7297.6140.7496.0138.5494.3940.0595.79
SCAN-PhysFed34.4996.4039.2196.0940.4597.9340.5197.6147.8199.1444.0898.5142.5497.8640.0297.3741.1497.62
+ +![](images/e2ed308faa97a69215e437e2567f90cd171f3eab8374cbf1ce2efebaa5808718.jpg) + +![](images/df320a31fde33d8a04455b194af685367fc386bc034619a43ebe8f5635209b4d.jpg) + +![](images/1621a7b22ab0b4585a96394a2ab048d94f71d1db0d177ace0beb4125f70648dd.jpg) + +![](images/6e8dad575802e27a24f9978c05a0c34d869a60c162e7fff50c8e4573939ccf67.jpg) + +![](images/a4f9850e4c9c1987d39d1c95dc43b6ae5e12b6170e7a9e567150392d79c5b468.jpg) + +![](images/535f9ab309baccf653ff651bbb12a680f621de40c6344b9732d19432fbbecc52.jpg) +(f) + +![](images/5051a10275f509569b4bd3291eda3da090e36a48139f7c6a76cd1660304489d1.jpg) +(a) +(b) +(g) +Figure 3. Simulated examples under different protocols. (a)-(h) show examples from training clients with known protocols, while (i)-(l) present examples from unseen protocols. + +![](images/a12e95d7688dc525aeb9f99fc69a05a986bb1b64df328f923766c068a791574c.jpg) +(h) + +![](images/3dcab6b06dec54e402b18748666968c8e567a3ec44e8dd646d287029990b4cd2.jpg) + +(i) + +![](images/c840ccd4dc1b67661d13e7561a544f1438871505d987bd0a5df82c7cffbec3ac.jpg) +(d) +(1) + +![](images/0cf2042361aa975a2f1aef7bd6deda92d8d920493f4b81514cde0811f23b0475.jpg) +(e) +(k) + +To fully leverage the diversity of client data, we globally share the hypernetworks in SCAN-PhysFed. However, projecting imaging features from different domains into a consistent, noise-free representation is challenging with a single shared decoder. Therefore, we propose protocol-specific decoders to effectively adapt to variations in scanning protocols across clients. + +In PVQS, the unseen protocol code is quantized based on the nearest entry in the protocol codebook. Therefore, it is essential to ensure the protocol code is discriminative and informative for personalization. To achieve this, we introduce an orthogonal loss $\mathcal{L}_{\mathrm{orth}}$ for the $i$ -th client, defined as follows: + +$$ +\mathcal {L} _ {\text {o r t h}} = \sum_ {j = 1} ^ {K} \left| \mathbf {c} _ {i} \cdot \mathbf {c} _ {j} \right| ^ {2} \quad s. t. \quad i \neq j, \tag {8} +$$ + +where $K$ is the number of clients (protocols), and $j$ denotes the client index. + +The Mean Squared Error (MSE) is used as the imaging loss, and the total loss is formulated as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\mathrm {M S E}} (\mathbf {Y}, \hat {\mathbf {Y}}) + \tau \mathcal {L} _ {\text {o r t h}}, \tag {9} +$$ + +where $\mathbf{Y}$ represents the reconstructed image, and $\tau$ denotes the temperature weight of $\mathcal{L}_{\mathrm{orth}}$ . + +To facilitate understanding of the implementation details, we provide the pseudocode of our method in Appendix A. + +# 4. Experiments + +Experimental Setting. The proposed method was implemented using PyTorch and optimized by Adam [12] at a learning rate of 0.001. The number of communication rounds was set to 200 and the batch size was 20. The experiments were conducted a system equipped with an AMD Ryzen 75800X CPU and a single NVIDIA GTX 3080 Ti GPU. + +Dataset. The "2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge" dataset [22], which contains 5,936 full-dose CT images from 10 patients, was used to evaluate the proposed method. In this study, 8 patients were used for training, while the remaining 2 were reserved for testing. To simulate a realistic scenario, we ensured that the patients for training did not overlap across different clients. The two patients for testing were simulated using different protocols to form the testing set. + +Simulated examples are shown in Figure 3, where noticeable noise heterogeneity among clients can be observed, primarily due to differences in anatomy and scanning protocols. Detailed dataset preparation steps and scanning protocols are provided in Appendix B. + +Baselines. We compared our SCAN-PhysFed approach with both generic and personalized FL methods. For generic FL methods, we include FedAvg [23], FedProx [16], FedNova [29], MOON [15], FedDG [19], and FedKD [32]. For personalized FL methods, we evaluated FedPer [3], FedBN [17], FedMRI [8], HyperFed [37], and FedFDD [6]. To ensure fairness, we kept the training settings consistent across all methods. + +Evaluation Metrics. We use the peak signal-to-noise ratio (PSNR) to evaluate pixel-wise accuracy and the structural similarity (SSIM) to assess perceived visual quality [30]. For both metrics, higher values indicate better performance. + +# 4.1. Comparison with other methods + +We evaluate our method alongside other FL methods. To ensure fairness, we use the same imaging network, RED + +![](images/319310e54dc1ecebd8393ade11f4f345fd33b8283cecb1cfcdaace46c40bf822.jpg) +Figure 4. Qualitative results of six selected comparison methods and our method across different clients using the classical convolutional-based LDCT imaging network. Rows one to five correspond to Clients #2, #3, #5, #6, and #7, respectively. The display window for the first row is [-1024, 200] HU, while for the remaining rows, it is [-160, 240] HU. + +![](images/f533115db85bd21fcbf9159a0a753652c41cfaeef2f79c76c3c97094afbee6b3.jpg) +Figure 5. Boxplots of the average results across all clients. (a) and (b) represent PSNR and SSIM, respectively. + +CNN [4], a state-of-the-art model in CT imaging. Quantitative results are provided in Table 1. Our method achieves the best performance across most clients, outperforming other generic and personalized FL methods in terms of average PSNR and SSIM. Furthermore, unlike other personalized methods that show instability across clients, such as HyperFed in Client #2 / #7 and FedFDD in Client #4, our method maintains consistent performance across varying protocols. We also observe that sparse-view protocols (Clients #2, #4, and #6) pose challenges for both generic and personalized FL methods, resulting in significant performance gaps compared to other protocols. How- + +ever, by incorporating dual-level physics-informed prior, our method maintains robust performance. + +Representative slices produced by different methods are shown in Figure 4. We compared the top three methods from both generic and personalized FL approaches. Representative clients were selected based on different radiation reduction strategies: sparse-view (Clients #2 and #6) and low-dose (Clients #3, #5, and #7) reconstructions. It can be observed that other methods exhibit noticeable noise or artifacts in sparse-view protocols, and fail to preserve details in low-dose protocols. In contrast, our method effectively reduces noise and artifacts in sparse-view protocols while preserving fine details in low-dose protocols. SCAN-PhysFed achieves strong performance across slices from different anatomical regions, benefiting from the introduction of anatomy priors. Both quantitative and qualitative results demonstrate the effectiveness of incorporating physics-informed priors. + +Additionally, we present boxplots of the results across all clients for different methods in Figure 5 to validate their stability. It can be seen that SCAN-PhysFed outperforms all other methods in both PSNR and SSIM scores. In both plots, SCAN-PhysFed demonstrates a higher median and + +Table 2. Quantitative Results of PSNR and SSIM for Transformer-based LDCT Imaging Methods. + +
Client #1Client #2Client #3Client #4Client #5Client #6Client #7Client #8Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]34.3187.8838.0690.5443.5197.5641.0595.4345.2498.4743.1597.0941.4595.9542.8896.1641.2194.89
FedProx [16]34.3287.0137.7488.8943.7397.9541.1293.9245.3998.6142.9997.0841.3894.3142.1094.5541.1094.04
FedNova [29]29.5578.4431.3868.1337.0782.4731.7973.4540.6188.8035.4679.7732.3873.6635.7782.2234.2578.37
MOON [15]34.2893.2038.0991.1743.6497.7741.0995.6645.2598.4342.8296.9141.4295.9742.2096.1841.1095.66
FedKD [32]35.9289.9338.7091.2144.2698.1342.3095.9945.9398.7543.5797.4142.5496.2243.5196.2342.0995.48
Personalized FL Methods
FedPer [3]35.3793.1839.0393.5542.7597.1340.8896.2046.5598.8342.8796.6841.8796.7040.7994.7841.2695.82
FedMRI [8]34.6093.6738.0494.5542.9497.4439.3495.7846.7998.9142.8297.3040.8195.3240.8595.7040.7896.01
HyperFed [37]34.6087.5938.6291.0844.1397.6142.2596.2445.9298.5943.7197.3742.4496.6643.5696.6341.9095.22
FedFDD [6]37.3596.2039.6795.1844.3198.2639.0095.2147.8999.2043.8198.0642.1797.4040.5095.4941.8496.89
SCAN-PhysFed38.5594.0641.7496.2045.5298.5242.4896.5549.5099.3945.3297.9343.3596.7342.7196.7943.6597.03
Ground TruthInputFedAvgMOONFedKDFedMRIHyperFedFedFDDSCAN-PhysFed
+ +Figure 6. Qualitative results of the compared methods across different clients based on a transformer-based imaging network. Rows one to five correspond to Clients #2, #3, #5, #6, and #7, respectively. The display window is set to [-160, 240] HU. + +a narrower interquartile range, indicating strong stability in restoring LDCT data with varying noise distributions across different clients. Individual boxplots for each client are provided in Appendix C. + +# 4.2. Generalization Evaluation + +# 4.2.1 Backbone + +In addition to using RED-CNN as the imaging network in previous experiments, we evaluate our method and other FL methods with an alternative imaging network, Uformer [31], a transformer-based model for LDCT imaging. This comparison further validates the generalization capability of different FL methods. + +The qualitative and quantitative results are shown in Figure 6 and Table 2, respectively. Consistent with previous experiments, our method achieves the best performance compared to other methods across most clients, with significantly higher average PSNR and SSIM values. This demonstrates the strong generalization capability of our method when integrated with different imaging networks. Among competing methods, FedFDD serves as a strong baseline. While FedFDD's SSIM is slightly higher than ours in some clients, our method achieves a significantly better PSNR, demonstrating superior detail preservation. Additionally, unlike FedFDD, which requires two imaging networks, our method achieves high performance with + +Table 3. The Quantitative Results of PSNR and SSIM for the Unseen Clients. + +
Unseen Client #1Unseen Client #2Unseen Client #3Unseen Client #4Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]35.5788.6834.0085.2632.5276.7136.8691.4432.8981.35
FedProx [16]36.6991.1634.3086.3932.5872.1339.1394.8734.7485.52
FedNova [29]35.4588.8334.2386.2932.3969.2035.4089.9934.3883.58
MOON [15]35.2178.1935.3079.8134.5583.0735.5678.9435.1680.00
FedDG [19]34.8679.5434.7980.9133.1580.3134.4977.4234.3279.55
FedKD [32]36.8291.6034.2786.3732.4970.5440.7395.9036.0886.10
Personalized FL Methods
FedPer [3]36.8688.5835.4687.0231.9671.6438.5690.9035.7184.53
FedBN [17]37.8489.9536.1688.4633.6781.6340.2194.8936.9788.73
FedMRI [8]31.1282.4231.9579.9531.7271.9632.0481.6331.7178.99
HyperFed [37]36.1889.6033.4782.4931.9465.5241.7094.8535.9083.12
FedFDD [6]37.8092.3337.0691.7733.7180.2640.5395.1237.2789.87
SCAN-PhysFed40.7797.6938.7795.7534.3881.8040.2796.9238.5593.04
+ +Table 4. Ablation on different components in SCAN-PhysFed. + +
ScanningComponentAverage
AnatomyParadigmLossPSNRSSIM
XXGenericLMSE34.0384.70
✓XPersonalizedLMSE36.8288.78
X✓PersonalizedLMSE38.0494.14
✓✓PersonalizedLMSE40.6797.27
✓✓PersonalizedLMSE & Lorth41.1497.62
+ +out relying on an additional network. Similar to the previous experiment, SCAN-PhysFed shows better qualitative performance and lower variation across different protocols, indicating more consistent results irrespective of scanning protocols or imaging networks. + +# 4.2.2 Unseen Protocol + +Existing FL methods in medical imaging struggle to maintain performance when encountering unseen noise distributions from new clients. Here, we evaluate the effectiveness of the proposed PVQS across four unseen protocols. Detailed protocol settings can be found in Appendix B. For generic FL methods, we directly evaluate the shared model on unseen clients. However, personalized methods mostly ignore how to operate in the unseen domain, so we average the performance of all local models to establish a vanilla baseline in this paper. + +The results of different methods on unseen clients are presented in Table 3. We observe that the performance of generic FL methods shows only minor degradation on unseen clients, whereas personalized FL methods experience significant declines. This is because local optimization in personalized FL methods focuses heavily on each client's specific data distribution, underfitting other potential distributions and leading to greater performance drops on unseen clients. In PVQS, we prevent the network from directly mapping the unseen protocol to modulate the imaging network, as this could cause network collapse if the unseen protocol significantly differs from known ones. Instead, we quantize the unseen protocol using a protocol codebook, al + +lowing SCAN-PhysFed to leverage prior knowledge for better adaptation to unseen distributions. + +# 4.3. Ablation Study + +We evaluate the effectiveness of various components in SCAN-PhysFed, and the results are presented in Table 4. Compared to the baseline results, which exclude scanning and anatomy information, introducing either significantly improves performance, highlighting the importance of integrating physical prior in LDCT imaging. This approach substantially reduces learning complexity and improves the overall model performance. Moreover, introducing both types of physical information simultaneously is complementary rather than redundant, leading to further performance gains. Building on this, we use $\mathcal{L}_{\mathrm{orth}}$ to enhance the discriminative ability of the scanning code, thereby personalizing imaging features and effectively mitigating data heterogeneity. In the above studies, "Personalized" indicates that the decoder is tailored to each client. An ablation study on the personalized components is provided in Appendix D. + +# 5. Conclusion + +To preserve privacy and address heterogeneity in CT imaging, we propose a dual-level, physics-informed approach that utilizes both scanning- and anatomy-level prompts to enable personalized CT imaging with the assistance of an MLLM. This approach facilitates both patient- and scanning-level personalization, with extensive results demonstrating that incorporating physical prior significantly improves performance and effectively mitigates heterogeneity. Additionally, we propose a novel strategy, PVQS, to quantize protocol codes and maintain robust performance on unseen protocols. PVQS leverages known protocol codes to prevent latent collapse when encountering significantly different unseen protocols. In future work, we aim to develop a general FL framework for medical imaging that can support both CT and MRI denoising, which we believe holds promising potential. + +# Acknowledgment + +This work was supported in part by the National Natural Science Foundation of China under Grant 62271335, in part by the Sichuan Science and Technology Program under Grant 2025ZNSFSC0470, and in part by Sichuan University "From 0 to 1" Innovative Research Program under Grant 2022SCUH0016. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[2] Asma Alkhaldi, Raneem Alnajim, Layan Alabdullatef, Rawan Alyahya, Jun Chen, Deyao Zhu, Ahmed Alsinan, and Mohamed Elhoseiny. Minigpt-med: Large language model as a general interface for radiology diagnosis. arXiv preprint arXiv:2407.04106, 2024. 3, 4 +[3] Manoj Ghuhan Arivazhagan, Vinay Aggarwal, Aaditya Kumar Singh, et al. Federated learning with personalization layers. arXiv preprint arXiv:1912.00818, 2019. 3, 5, 7, 8 +[4] Hu Chen, Yi Zhang, Mannudeep K Kalra, et al. Low-dose ct with a residual encoder-decoder convolutional neural network. IEEE Trans. Med. Imaging, 36(12):2524-2535, 2017. 2, 6 +[5] Shixuan Chen, Boxuan Cao, Yinda Du, Yaoduo Zhang, Ji He, Zhaoying Bian, Dong Zeng, and Jianhua Ma. Federated condition generalization on low-dose ct reconstruction via cross-domain learning. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 47-56. Springer, 2023. 3 +[6] Xuhang Chen, Zeju Li, Zikun Xu, Cheng Ouyang, Chen Qin, et al. Fedfdd: Federated learning with frequency domain decomposition for low-dose ct denoising. In Medical Imaging with Deep Learning, 2024. 3, 4, 5, 7, 8 +[7] Dongshang Deng, Xuangou Wu, Tao Zhang, Xiangyun Tang, Hongyang Du, Jiawen Kang, Jiqiang Liu, and Dusit Niyato. Fedasa: A personalized federated learning with adaptive model aggregation for heterogeneous mobile edge computing. IEEE Transactions on Mobile Computing, 2024. 3 +[8] Chun-Mei Feng, Yunlu Yan, Shanshan Wang, Yong Xu, Ling Shao, and Huazhu Fu. Specificity-preserving federated learning for mr image reconstruction. IEEE Transactions on Medical Imaging, 42(7):2010-2021, 2022. 3, 5, 7, 8 +[9] Qi Gao, Zilong Li, Junping Zhang, Yi Zhang, and Hongming Shan. Corediff: Contextual error-modulated generalized diffusion model for low-dose ct denoising and generalization. IEEE Transactions on Medical Imaging, 2023. 2 +[10] Hao Guan, Pew-Thian Yap, Andrea Bozoki, and Mingxia Liu. Federated learning for medical image analysis: A survey. Pattern Recognition, page 110424, 2024. 3 +[11] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for feder- + +ated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 2 +[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[13] Gurucharan Marthi Krishna Kumar, Aman Chadha, Janine Mendola, and Amir Shmuel. Medvisionllama: Leveraging pre-trained large language model layers to enhance medical image segmentation. arXiv preprint arXiv:2410.02458, 2024.3 +[14] Danyang Li, Hao Wang, Jingyi Liao, Mingqiang Meng, Cuidie Zeng, Yuting Wang, Dong Zeng, and Jianhua Ma. Semi-centralized federated learning network for low-dose ct imaging. In Medical Imaging 2023: Physics of Medical Imaging, pages 1024-1028. SPIE, 2023. 3 +[15] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 10713–10722, 2021. 2, 5, 7, 8 +[16] Tian Li, Anit Kumar Sahu, Manzil Zaheer, et al. Federated optimization in heterogeneous networks. In Proc. Mach. Learn. Syst. (MLSys), pages 429-450, 2020. 2, 5, 7, 8 +[17] Xiaoxiao Li, Meirui JIANG, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. In International Conference on Learning Representations, 2021. 3, 5, 8 +[18] Fenglin Liu, Tingting Zhu, Xian Wu, Bang Yang, Chenyu You, Chenyang Wang, Lei Lu, Zhangdaihong Liu, Yefeng Zheng, Xu Sun, et al. A medical multimodal large language model for future pandemics. NPJ Digital Medicine, 6(1): 226, 2023. 3 +[19] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1013–1023, 2021. 5, 8 +[20] Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Derong Xu, Feng Tian, and Yefeng Zheng. When moe meets llms: Parameter efficient fine-tuning for multi-task medical applications. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1114, 2024. 3 +[21] Zexin Lu, Qi Gao, Tao Wang, Ziyuan Yang, Zhiwen Wang, Hui Yu, Hu Chen, Jiliu Zhou, Hongming Shan, and Yi Zhang. Pridediff: Physics-regularized generalized diffusion model for ct reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences, 2024. 2 +[22] C McCollough. Tu-fg-207a-04: overview of the low dose ct grand challenge. Med. Phys., 43(6):3759–3760, 2016. 5 +[23] Brendan McMahan, Eider Moore, Daniel Ramage, et al. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 2, 5, 7, 8 +[24] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medi + +cal artificial intelligence. Nature, 616(7956):259-265, 2023. 3 +[25] Saiprasad Ravishankar, Jong Chul Ye, and Jeffrey A Fessler. Image reconstruction: From sparsity to data-adaptive methods and machine learning. Proceedings of the IEEE, 108(1): 86-109, 2019. 2 +[26] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[27] Ge Wang, Hengyong Yu, and Bruno De Man. An outlook on x-ray ct research and development. Medical physics, 35(3): 1051-1064, 2008. 1 +[28] Ge Wang, Jong Chul Ye, and Bruno De Man. Deep learning for tomographic image reconstruction. Nature machine intelligence, 2(12):737-748, 2020. 1 +[29] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 2, 5, 7, 8 +[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[31] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 2, 7 +[32] Chuhan Wu, Fangzhao Wu, Lingjuan Lyu, Yongfeng Huang, and Xing Xie. Communication-efficient federated learning via knowledge distillation. Nature Communications, 13(1): 2032, 2022. 3, 5, 7, 8 +[33] Wenjun Xia, Qing Lyu, and Ge Wang. Low-dose ct using denoising diffusion probabilistic model for 20x speedup. arXiv preprint arXiv:2209.15136, 2022. 2 +[34] Wenjun Xia, Ziyuan Yang, Zexin Lu, Zhongxian Wang, and Yi Zhang. Regformer: A local-nonlocal regularization-based model for sparse-view ct reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences, 8(2): 184-194, 2023. 2 +[35] Xiyuan Yang, Wenke Huang, and Mang Ye. Fedas: Bridging inconsistency in personalized federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11986-11995, 2024. 3 +[36] Ziyuan Yang, Yingyu Chen, Huijie Huangfu, Maosong Ran, Hui Wang, Xiaoxiao Li, and Yi Zhang. Dynamic corrected split federated learning with homomorphic encryption for u-shaped medical image networks. IEEE Journal of Biomedical and Health Informatics, 27(12):5946-5957, 2023. 3 +[37] Ziyuan Yang, Wenjun Xia, Zexin Lu, Yingyu Chen, Xiaoxiao Li, and Yi Zhang. Hypernetwork-based physics-driven personalized federated learning for ct imaging. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3, 4, 5, 7, 8 + +[38] Mang Ye, Xiuwen Fang, Bo Du, Pong C Yuen, and Dacheng Tao. Heterogeneous federated learning: State-of-the-art and research challenges. ACM Computing Surveys, 56(3):1-44, 2023. 2 +[39] Chenyu You, Guang Li, Yi Zhang, et al. Ct super-resolution gan constrained by the identical, residual, and cycle learning ensemble (gan-circle). IEEE Trans. Med. Imaging, 39(1): 188-203, 2019. 2 +[40] Zhicheng Zhang, Lequan Yu, Xiaokun Liang, Wei Zhao, and Lei Xing. Transc: dual-path transformer for low dose computed tomography. In Medical Image Computing and Computer Assisted Intervention, pages 55–64. Springer, 2021. 2 +[41] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large language models meet few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3065-3075, 2024. 3 \ No newline at end of file diff --git a/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/images.zip b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d52d976e7db1c7fc920437a1bef36f74363696eb --- /dev/null +++ b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2a8aaa2bcafaaad2d6e9cd1e97b1743160ae7f23479c8e84c71e2996db0a03c +size 1101190 diff --git a/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/layout.json b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d09f2e870f3d2e1301881d580fc15f4096c3751d --- /dev/null +++ b/2025/Patient-Level Anatomy Meets Scanning-Level Physics_ Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model/layout.json @@ -0,0 +1,7980 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "text", + "content": "Patient-Level Anatomy Meets Scanning-Level Physics: Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "spans": [ + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "content": "Ziyuan Yang" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "content": ", Yingyu Chen" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "content": ", Zhiwen Wang" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "content": ", Hongming Shan" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "content": ", Yang Chen" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "content": ", Yi Zhang" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{1,5,*}" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "content": "Sichuan University " + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "content": "Southwest Petroleum University " + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "content": "Fudan University " + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "content": "Southeast University " + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 77, + 160, + 533, + 232 + ], + "type": "text", + "content": "Tianfu Jiangxi Laboratory \n{cziyuanyang, cyy262511, zwwang1228}@gmail.com, hmshan@fudan.edu.cn \nchenyang.list@seu.edu.cn, yzhang@scu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 270, + 200, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 270, + 200, + 283 + ], + "spans": [ + { + "bbox": [ + 151, + 270, + 200, + 283 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 305, + 297, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 305, + 297, + 678 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 297, + 678 + ], + "type": "text", + "content": "Reducing radiation doses benefits patients, but the resultant low-dose computed tomography (LDCT) images often suffer from clinically unacceptable noise and artifacts. While deep learning (DL) has shown promise in LDCT reconstruction, it requires large-scale data collection from multiple clients, raising privacy concerns. Federated learning (FL) has been introduced to mitigate these privacy concerns; however, current methods are typically tailored to specific scanning protocols, which limits their generalizability and makes them less effective for unseen protocols. To address these issues, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Since the noise distribution in LDCT data is closely tied to scanning protocols and anatomical structures, we propose a dual-level physics-informed way to address these challenges. Specifically, we incorporate physical and anatomical prompts into our physics-informed hypernetworks to capture scanning- and anatomy-specific information, enabling dual-level physics-driven personalization of imaging features. These prompts are derived from the scanning protocol and the radiology report generated by a medical large language model (MLLM). Subsequently, client-specific decoders project these dual-level personalized imaging features back into the image domain. Besides, to tackle the challenge of unseen data, we introduce a novel protocol vector-quantization strategy (PVQS), which ensures consistent performance across new clients by quantifying unseen scanning codes to the closest match in the scanning codebook. Extensive experimental results demonstrate the superior performance of SCAN-PhysFed on pub" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 259, + 369, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 259, + 369, + 271 + ], + "spans": [ + { + "bbox": [ + 314, + 259, + 369, + 271 + ], + "type": "text", + "content": "lic datasets1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 295, + 394, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 295, + 394, + 307 + ], + "spans": [ + { + "bbox": [ + 314, + 295, + 394, + 307 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 315, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 445 + ], + "type": "text", + "content": "Computed tomography (CT) is a key diagnostic tool that non-invasively visualizes anatomical structures within the human body. Despite its clinical benefits, CT scans pose radiation concerns, which potentially lead to genetic, cancerous, and other health diseases. Lowering CT radiation dose—either by decreasing incident photons or limiting sampling views—has emerged as a promising solution to reduce radiation risks and accelerate scanning. However, reconstructed images under these conditions suffer from significant quality degradation, compromising clinical applicability [27]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 447, + 556, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 637 + ], + "type": "text", + "content": "Recently, deep learning (DL)-based methods have shown great promise for low-dose CT (LDCT) imaging [28]. However, most DL-based LDCT reconstruction methods are condition-specific and overlook privacy concerns. Recently, researchers have attempted to introduce federated learning (FL), a privacy-preserving distributed learning paradigm, for LDCT reconstruction to alleviate privacy concerns. Although these methods aim to enhance generalization, they still face significant challenges arising from data heterogeneity, caused by diverse noise distributions in CT imaging across different protocols and anatomical structures. We believe that the further advancement of FL in LDCT reconstruction urgently requires a new framework capable of leveraging comprehensive physical information from the scanning process, thereby breaking through current bottlenecks to achieve better results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 638, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 556, + 687 + ], + "type": "text", + "content": "To tackle this issue, we aim to leverage physical principles of the imaging process. During the projection process, the measurement is influenced by the anatomical structures the X-ray passed through, such as fat, bone, and fluid, each" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 703, + 153, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 703, + 153, + 713 + ], + "spans": [ + { + "bbox": [ + 74, + 703, + 153, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 693, + 553, + 713 + ], + "type": "text", + "content": "1 Our code has been released at https://github.com/Zi-YuanYang/SCAN-PhysFed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5154" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 76, + 288, + 198 + ], + "blocks": [ + { + "bbox": [ + 62, + 76, + 288, + 198 + ], + "lines": [ + { + "bbox": [ + 62, + 76, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 62, + 76, + 288, + 198 + ], + "type": "image", + "image_path": "60200b64a90c355528f48b7d4824df9a108087adeb73fc4b6e5254a5d9bc00eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 295, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 253 + ], + "type": "text", + "content": "Figure 1. Illustration of different radiation reduction strategies. 'Sparse-View' refers to scanning with a reduced number of projection views, while 'Low-Dose' represents scanning with a lower number of incident photons." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 256, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 353 + ], + "type": "text", + "content": "with distinct attenuation coefficients. Additionally, the system matrix is significantly affected by scanning protocols, including tube current, voltage settings, and scanning angles. These factors collectively impact the quality and characteristics of reconstructed CT images during the backprojection phase, resulting in varying levels of noise and artifacts. An example is provided in Figure 1 to illustrate this concept more clearly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 508 + ], + "type": "text", + "content": "By leveraging the physics-specific knowledge discussed earlier, we reimagine LDCT reconstruction within the federated learning framework. Our approach adopts a dual-level strategy to alleviate heterogeneity, addressing variations at both the scanning and patient levels. In practice, scanning protocols, which serve as scanner hyperparameters, are readily available. However, generating anatomical information has traditionally been challenging due to patient variability and the complexity of accurately modeling different tissue types. Recent advances in large language models (LLMs) offer new possibilities for automatically generating anatomical information, potentially improving the reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": "Building on this, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Specifically, we address the data heterogeneous using a dual-level strategy. A detailed radiology report is generated using a pretrained LLM. Then, a patient-level anatomy-informed hypernetwork is designed to generate a modulation map from the radiology report, which seamlessly integrates with imaging features and aligns them with each patient's unique anatomical structures. Simultaneously, scanning features are integrated via a scanning-informed hypernetwork to personalize imaging features based on the protocol. To ensure distinguishability of scanning feature vectors, we impose an orthogonal loss during training to reinforce personalization. This dual-modulation strategy facilitates personalized CT imaging at both patient and scanning levels, guided by the principles of physical imaging." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "Additionally, existing personalized FL CT imaging methods do not address the challenge posed by unseen clients with varying scanning protocols. In this paper, we propose a Protocol Vector Quantization Strategy (PVQS) to quantize unseen scanning protocol by matching it with the closest existing vectors, thereby selecting the corresponding client-specific decoder. The main contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 168, + 553, + 311 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 168, + 553, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 553, + 214 + ], + "type": "text", + "content": "- We introduce LLM into LDCT reconstruction within the FL framework, proposing SCAN-PhysFed. To the best of our knowledge, this is the first work to apply LLM in this field." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 216, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 262 + ], + "type": "text", + "content": "- To address the challenge of data heterogeneity, we draw inspiration from physical imaging principles and propose a dual-level approach, achieving personalized CT imaging at both the scanning and patient levels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 264, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 311 + ], + "type": "text", + "content": "- We introduce a novel strategy, PVQS, to ensure our method remains effective for unseen clients with diverse scanning protocols—a critical challenge often overlooked in previous works." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 323, + 405, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 405, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 405, + 335 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 570 + ], + "type": "text", + "content": "LDCT Imaging. Traditional reconstruction methods primarily rely on image sparsity [25]. However, these methods are generally time-consuming. Recently, deep learning-based methods have shown faster processing and promising performance, gaining significant attention in this field. For example, Chen et al. [4] introduced residual structure into an autoencoder and proposed the encoder-decoder convolutional neural network (RED-CNN) for LDCT denoising. You et al. [39] extended RED-CNN by incorporating a generative adversarial network (GAN) for super resolution. Additionally, researchers also explored transformer-based architectures to enhance robustness, as seen in Uformer [31], TransCT [40], and RegFormer [34]. Recently, researchers have proposed diffusion-based CT reconstruction methods [9, 21, 33], which show promising results but incur substantial computational costs. Moreover, the methods mentioned above are tailored to specific scanning protocols and depend on data collection from multiple clients, raising significant privacy concerns." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Federated Learning. In recent years, FL has gained increasing attention as it offers a decentralized approach to DL while preserving data privacy [38]. FedAvg [23], the earliest FL method, introduced a server-client framework to learn a global-shared model by averaging local models from different parties. However, this method suffers from model shift due to serious statistic heterogeneity [11]. To alleviate this concern, recent works have proposed methods to obtain a more robust global-shared model, such as FedProx [16], MOON [15], and FedNova [29]. For serious non-iid data distributions, a single global-shared model is insufficient to address the variability across different datasets. As" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5155" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 549, + 264 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 549, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 549, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 549, + 264 + ], + "type": "image", + "image_path": "c60c76674ed7c98a49b9ac68176ac7dad9a3cac2e9e5719c8a8f7e23316603f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 265, + 440, + 277 + ], + "lines": [ + { + "bbox": [ + 170, + 265, + 440, + 277 + ], + "spans": [ + { + "bbox": [ + 170, + 265, + 440, + 277 + ], + "type": "text", + "content": "Figure 2. The overall learning paradigm of our proposed SCAN-PhysFed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 283, + 295, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 295, + 306 + ], + "type": "text", + "content": "a result, several personalized FL methods have been proposed [3, 7, 17, 35]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 306, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 306, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 295, + 462 + ], + "type": "text", + "content": "FL has recently been introduced in healthcare applications due to its privacy-preserving nature [10, 32, 36]. Several studies have focused on its applications in medical imaging [8, 14]. For example, Chen et al. [5] proposed a cross-domain learning approach that combines individual-client sinogram learning and cross-client image reconstruction learning. Yang et al. [37] proposed a client-specific hypernetwork to modulate the global-shared data. Chen et al. [6] utilized separate networks to process high- and low-frequency components of CT data. While these methods have achieved promising performance, they still face significant challenges from data heterogeneity and fail to address the issue of unseen protocols." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 463, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 643 + ], + "type": "text", + "content": "Large Language Model. The emergence of large language models (LLMs), such as GPT [1] and Llama [26], has ushered in a new era of artificial intelligence. Recently, largescale medical LLMs (MLLMs) have been developed specifically for healthcare applications [18, 20, 41]. These models demonstrate significant potential in understanding and reasoning over medical data, excelling in tasks such as visual question answering (VQA) and text generation [24]. For instance, Alkhaldi et al. [2] developed Minigpt-med, which supports multiple tasks, including VQA, medical report generation, and disease identification from medical images. Although MLLMs have been applied to various downstream tasks [13], they have not yet been explored for mitigating domain shift in CT reconstruction within FL settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 651, + 138, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 138, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 138, + 665 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 123, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 123, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 123, + 683 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "This paper aims to address generalization challenges in FL for LDCT reconstruction, with a focus on alleviating data" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 283, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 426 + ], + "type": "text", + "content": "heterogeneity among existing clients and maintaining performance for new clients. Figure 2 illustrates the overall learning paradigm of the proposed SCAN-PhysFed framework. Specifically, robust imaging representations are obtained using a collaboratively learned shared encoder across clients. To tackle the heterogeneity issue, two physics-informed hypernetworks are designed to capture scanning-and anatomy-specific information, personalizing the imaging features via a specialized modulation block. Subsequently, client-specific decoders project the personalized imaging features back to the image domain. For the " + }, + { + "bbox": [ + 313, + 283, + 555, + 426 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 283, + 555, + 426 + ], + "type": "text", + "content": "-th client, the reconstruction process is formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 435, + 555, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 435, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 328, + 435, + 555, + 449 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Y}} = D _ {i} \\left(P \\left(f _ {\\mathbf {X}}, H _ {s} (g; \\theta_ {H _ {s}}), H _ {a} \\left(f _ {t}; \\theta_ {H _ {a}}\\right)\\right); \\theta_ {D _ {i}}\\right), \\tag {1}", + "image_path": "63bdbe6ccd73b407d466fdef3186c529490ac215f6b446a4b4c5d7f47edf76e2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": " denotes the imaging feature of " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": " extracted from the shared encoder " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": " is the corresponding normal-dose CT image. " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\theta_{H_s}" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\theta_{H_a}" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": " are the scanning- and anatomy-informed hypernetworks, parameterized by " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\theta_s" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\theta_{a}" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": " represent the scanning protocol vector and the radiology report feature derived from the MLLM. " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": " denotes the personalized modulation block, and " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": "-th client decoder, parameterized by " + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\theta_{D_i}" + }, + { + "bbox": [ + 313, + 459, + 554, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 555, + 554, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 615 + ], + "type": "text", + "content": "Additionally, to handle unseen data, we propose a novel protocol vector-quantization strategy (PVQS), which ensures robust performance across previously unseen clients by quantizing the unseen scanning code to the closest match in the scanning codebook." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 623, + 484, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 484, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 484, + 635 + ], + "type": "text", + "content": "3.2. Personalized Modulation Block" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "content": "Anatomy-Level Personalization. The patient's unique anatomy significantly affects noise distribution during the imaging process, making it an essential factor to consider. To address this issue, we leverage MLLM, which is trained on large-scale datasets and demonstrates strong performance in radiology report generation. Specifically," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5156" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "we adapt miniGPT-Med [2] to generate radiology reports from CT images, using the text encoder's feature output " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "f_{t} \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " as a prompt for our anatomy-informed hypernetwork, where " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " denotes the feature dimension." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 156 + ], + "type": "text", + "content": "Given the high dimensionality of " + }, + { + "bbox": [ + 55, + 121, + 295, + 156 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 121, + 295, + 156 + ], + "type": "text", + "content": ", we first apply a pooling operation to reduce its size. The process is formulated as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 167, + 295, + 181 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 167, + 295, + 181 + ], + "spans": [ + { + "bbox": [ + 116, + 167, + 295, + 181 + ], + "type": "interline_equation", + "content": "f _ {t l} = \\operatorname {L i n e a r} \\left(\\operatorname {A v g P o o l} \\left(f _ {t}\\right)\\right), \\tag {2}", + "image_path": "2c143719b3591ef9308692b4ff355bd1f34ff66fd94754c3f1fd07a350e80ad9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "type": "inline_equation", + "content": "\\mathrm{AvgPool}" + }, + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "type": "text", + "content": " denotes an average pooling operation that reduces the feature dimension by a factor of " + }, + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "type": "inline_equation", + "content": "1 / 4" + }, + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "type": "text", + "content": ". A linear layer is then applied to map the feature to " + }, + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "type": "inline_equation", + "content": "f_{tl} \\in \\mathbb{R}^{1 \\times h}" + }, + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 190, + 295, + 237 + ], + "type": "text", + "content": " denotes the hidden dimension." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 239, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 286 + ], + "type": "text", + "content": "Next, we capture long-range dependencies within " + }, + { + "bbox": [ + 55, + 239, + 296, + 286 + ], + "type": "inline_equation", + "content": "f_{tl}" + }, + { + "bbox": [ + 55, + 239, + 296, + 286 + ], + "type": "text", + "content": " using the multi-head attention mechanism, followed by a linear layer to extract the anatomy feature " + }, + { + "bbox": [ + 55, + 239, + 296, + 286 + ], + "type": "inline_equation", + "content": "f_{an}" + }, + { + "bbox": [ + 55, + 239, + 296, + 286 + ], + "type": "text", + "content": ". This process is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 298, + 295, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 298, + 295, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 295, + 311 + ], + "type": "interline_equation", + "content": "f _ {a n} = \\operatorname {L i n e a r} (\\operatorname {M u l t i - H e a d} (f _ {t l})). \\tag {3}", + "image_path": "960faa7061db7c7a6f627a234c6022db6d48d6b897c7d82234a88fb004d31ef9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 321, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 380 + ], + "type": "text", + "content": "After this stage, " + }, + { + "bbox": [ + 55, + 321, + 296, + 380 + ], + "type": "inline_equation", + "content": "f_{an}" + }, + { + "bbox": [ + 55, + 321, + 296, + 380 + ], + "type": "text", + "content": " encapsulates anatomy-specific information tailored to different CT data. Since this feature encodes anatomy-specific details, we use it to spatially personalize the imaging feature " + }, + { + "bbox": [ + 55, + 321, + 296, + 380 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 321, + 296, + 380 + ], + "type": "text", + "content": ". The process is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 382, + 295, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 382, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 115, + 382, + 295, + 396 + ], + "type": "interline_equation", + "content": "f _ {a n a} = f _ {\\mathbf {X}} \\odot \\operatorname {R e s h a p e} \\left(f _ {a n}\\right), \\tag {4}", + "image_path": "ad828d7c05f788644000bf5d26264b03bc002b024de6995a954eeaca4fa59883.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 402, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 402, + 296, + 449 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 402, + 296, + 449 + ], + "type": "text", + "content": " denotes the imaging feature, Reshape represents the reshape operation for dimensional alignment, and “" + }, + { + "bbox": [ + 55, + 402, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 402, + 296, + 449 + ], + "type": "text", + "content": "” denotes the Hadamard product. " + }, + { + "bbox": [ + 55, + 402, + 296, + 449 + ], + "type": "inline_equation", + "content": "f_{ana}" + }, + { + "bbox": [ + 55, + 402, + 296, + 449 + ], + "type": "text", + "content": " is the anatomy-level personalized feature." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 450, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 558 + ], + "type": "text", + "content": "Scanning-Level Personalization. In addition to anatomical variations affecting noise distribution, the physical imaging process shows that noise distribution is primarily governed by the scanning protocol. This protocol comprises several key physical parameters in CT scanners, including the number of views, the number of detector bins, pixel length, detector bin length, the distance between the source and rotation center, the distance between the detector and rotation center, and the photon number of incident X-rays." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "text", + "content": "Then, we utilize " + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "text", + "content": ", a vector encoding these parameters, to construct our scanning-level prompt, which captures the latent relationship between the scanning protocol and noise distribution for the imaging network. A detailed description of " + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 559, + 295, + 618 + ], + "type": "text", + "content": " is provided in Appendix B." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 619, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 296, + 689 + ], + "type": "text", + "content": "Given the low dimensionality of the physical parameters, we employ a Multi-Layer Perceptron (MLP) with three linear layers to capture and represent the underlying physical information. Two modulation parameter vectors are then derived to encode scanning-level personalization knowledge. This process can be formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 137, + 701, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 137, + 701, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\alpha , \\beta = H _ {s} (g; \\theta_ {s}), \\tag {5}", + "image_path": "e6e4d0673fbe61d1ac595dd4705dfb7d9718b2298c4df801a9fd9f74117fce99.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " are the modulation parameter vectors, which are used to inject scanning-level information to personalize the imaging feature as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 388, + 116, + 553, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 116, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 388, + 116, + 553, + 129 + ], + "type": "interline_equation", + "content": "f _ {p e r} = \\alpha \\otimes f _ {a n a} + \\beta , \\tag {6}", + "image_path": "725937da88a184aa5bb910855ba92f102bb8618e515a0b4ab535a43ef0179b69.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 135, + 553, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 553, + 183 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 135, + 553, + 183 + ], + "type": "inline_equation", + "content": "f_{per}" + }, + { + "bbox": [ + 313, + 135, + 553, + 183 + ], + "type": "text", + "content": " is the dual-level personalized imaging feature, and " + }, + { + "bbox": [ + 313, + 135, + 553, + 183 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 135, + 553, + 183 + ], + "type": "text", + "content": " denotes the multiplication operation along the channel dimension. This approach applies the scanning prior to modulate the imaging feature." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 183, + 553, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 219 + ], + "type": "text", + "content": "Our proposed personalization block is architecture-agnostic, allowing the personalization parameters to be finetuned accordingly." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 225, + 517, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 517, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 517, + 239 + ], + "type": "text", + "content": "3.3. Protocol Vector-Quantization Strategy" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 243, + 553, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 363 + ], + "type": "text", + "content": "Existing FL-based CT imaging methods have not addressed the challenge of unseen data, resulting in degraded performance in new domains. To overcome this limitation and ensure robust performance for unseen protocols, we introduce PVQS. While patient anatomy remains largely consistent, the primary source of noise variation between unseen and existing clients stems from the low-dose strategy. Therefore, PVQS is designed to enable SCAN-PhysFed to maintain consistent performance across diverse and unseen low-dose protocols." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": "Specifically, in PVQS, protocol codes for different existing scanning protocols, generated by the MLP in " + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": ", are stored in a protocol codebook. For an unseen protocol, we first quantize it to the closest code vector in the protocol codebook. The process is defined as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 362, + 430, + 553, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 430, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 362, + 430, + 553, + 456 + ], + "type": "interline_equation", + "content": "i ^ {*} = \\arg \\min _ {i} \\left(1 - \\frac {\\mathbf {c} _ {u n} \\cdot \\mathbf {c} _ {i}}{\\| \\mathbf {c} _ {u n} \\| \\| \\mathbf {c} _ {i} \\|}\\right), \\tag {7}", + "image_path": "9d0c130b9c7088e21f5e0721171aaf1905e9a6020186c988fbc01f79ffbd8421.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 461, + 553, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 461, + 553, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{un}" + }, + { + "bbox": [ + 313, + 461, + 553, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 461, + 553, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 313, + 461, + 553, + 496 + ], + "type": "text", + "content": " denote the protocol code of the unseen protocol and that of the " + }, + { + "bbox": [ + 313, + 461, + 553, + 496 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 461, + 553, + 496 + ], + "type": "text", + "content": "-th client, respectively. " + }, + { + "bbox": [ + 313, + 461, + 553, + 496 + ], + "type": "inline_equation", + "content": "i^*" + }, + { + "bbox": [ + 313, + 461, + 553, + 496 + ], + "type": "text", + "content": " indicates the index of the closest code vector." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 497, + 553, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 553, + 544 + ], + "type": "text", + "content": "Instead of using the unseen protocol as input to " + }, + { + "bbox": [ + 313, + 497, + 553, + 544 + ], + "type": "inline_equation", + "content": "H_{s}" + }, + { + "bbox": [ + 313, + 497, + 553, + 544 + ], + "type": "text", + "content": ", we directly replace it with " + }, + { + "bbox": [ + 313, + 497, + 553, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{i^{*}}" + }, + { + "bbox": [ + 313, + 497, + 553, + 544 + ], + "type": "text", + "content": " to personalize the imaging feature and use the corresponding decoder " + }, + { + "bbox": [ + 313, + 497, + 553, + 544 + ], + "type": "inline_equation", + "content": "D_{i^{*}}" + }, + { + "bbox": [ + 313, + 497, + 553, + 544 + ], + "type": "text", + "content": " to project the personalized features back to the image domain." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 593 + ], + "type": "text", + "content": "Thus, PVQS effectively avoids potential distribution shifts in unseen data and preserves a consistent feature space across domains, ensuring reliable and robust performance for unseen protocols." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 599, + 429, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 429, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 429, + 612 + ], + "type": "text", + "content": "3.4. Learning Paradigm" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "Since the noise distribution varies across protocols, it is challenging to use a single shared imaging network to accommodate all variations. In FL-based LDCT imaging, as assumed in previous works [6, 37], the optimization challenge consists of two main components: imaging feature extraction and personalized projection. Specifically, for feature extraction, we propose aggregating client-side encoders to capture robust, comprehensive imaging features." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5157" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 89, + 555, + 228 + ], + "blocks": [ + { + "bbox": [ + 126, + 70, + 484, + 82 + ], + "lines": [ + { + "bbox": [ + 126, + 70, + 484, + 82 + ], + "spans": [ + { + "bbox": [ + 126, + 70, + 484, + 82 + ], + "type": "text", + "content": "Table 1. Quantitative Results (PSNR (dB) and SSIM " + }, + { + "bbox": [ + 126, + 70, + 484, + 82 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 126, + 70, + 484, + 82 + ], + "type": "text", + "content": ") for CNN-based LDCT Imaging Method." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 89, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 54, + 89, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 89, + 555, + 228 + ], + "type": "table", + "html": "
Client #1Client #2Client #3Client #4Client #5Client #6Client #7Client #8Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]29.9982.2032.8278.2334.6886.9932.3080.6336.8792.0637.3689.0232.7380.6635.4687.6534.0384.70
FedProx [16]29.9483.1232.6975.4636.9591.5832.3280.9138.9595.3437.5888.1332.9481.6235.8688.6534.6585.60
FedNova [29]29.9582.2932.5780.5737.7792.1632.0978.9834.4288.6936.0589.5732.8181.4135.8988.7133.9485.30
MOON [15]32.5075.6135.5585.4135.7280.3035.6982.1035.3878.4736.3184.3535.7081.1735.6280.4035.3180.98
FedDG [19]33.3781.4034.0182.3933.5676.3734.4381.4934.0676.7033.9780.1734.7580.8734.2579.2334.0579.83
FedKD [32]29.7483.1132.1771.6536.6590.9332.1079.1940.5896.7137.0782.5532.8881.3036.2889.0834.6884.31
Personalized FL Methods
FedPer [3]33.2388.2035.0684.8938.7992.7035.5886.1241.7696.9736.8589.9635.6986.3038.4891.6936.9389.60
FedBN [17]32.5584.5934.7684.1538.1890.3034.7783.3041.2496.1138.6691.1934.7483.3537.8990.3536.6087.92
FedMRI [8]33.2883.4334.5185.8138.6292.4735.8687.2141.4996.2738.2291.3235.8786.3438.0991.9436.9989.35
HyperFed [37]33.7488.5734.8385.0538.8792.7835.4485.4042.9497.4839.1592.9634.9985.1938.1291.2937.2689.84
FedFDD [6]36.1893.6738.4394.1939.7296.4738.5195.3045.5698.7342.7297.6140.7496.0138.5494.3940.0595.79
SCAN-PhysFed34.4996.4039.2196.0940.4597.9340.5197.6147.8199.1444.0898.5142.5497.8640.0297.3741.1497.62
", + "image_path": "46d3c525fb99831c5d5bd3471eae49d1cb79dae1f9c0e543bc65e92a296101a5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 240, + 98, + 277 + ], + "blocks": [ + { + "bbox": [ + 61, + 240, + 98, + 277 + ], + "lines": [ + { + "bbox": [ + 61, + 240, + 98, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 98, + 277 + ], + "type": "image", + "image_path": "e2ed308faa97a69215e437e2567f90cd171f3eab8374cbf1ce2efebaa5808718.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 99, + 240, + 137, + 277 + ], + "blocks": [ + { + "bbox": [ + 99, + 240, + 137, + 277 + ], + "lines": [ + { + "bbox": [ + 99, + 240, + 137, + 277 + ], + "spans": [ + { + "bbox": [ + 99, + 240, + 137, + 277 + ], + "type": "image", + "image_path": "df320a31fde33d8a04455b194af685367fc386bc034619a43ebe8f5635209b4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 138, + 240, + 175, + 277 + ], + "blocks": [ + { + "bbox": [ + 138, + 240, + 175, + 277 + ], + "lines": [ + { + "bbox": [ + 138, + 240, + 175, + 277 + ], + "spans": [ + { + "bbox": [ + 138, + 240, + 175, + 277 + ], + "type": "image", + "image_path": "1621a7b22ab0b4585a96394a2ab048d94f71d1db0d177ace0beb4125f70648dd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 176, + 240, + 213, + 277 + ], + "blocks": [ + { + "bbox": [ + 176, + 240, + 213, + 277 + ], + "lines": [ + { + "bbox": [ + 176, + 240, + 213, + 277 + ], + "spans": [ + { + "bbox": [ + 176, + 240, + 213, + 277 + ], + "type": "image", + "image_path": "6e8dad575802e27a24f9978c05a0c34d869a60c162e7fff50c8e4573939ccf67.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 214, + 240, + 252, + 277 + ], + "blocks": [ + { + "bbox": [ + 214, + 240, + 252, + 277 + ], + "lines": [ + { + "bbox": [ + 214, + 240, + 252, + 277 + ], + "spans": [ + { + "bbox": [ + 214, + 240, + 252, + 277 + ], + "type": "image", + "image_path": "a4f9850e4c9c1987d39d1c95dc43b6ae5e12b6170e7a9e567150392d79c5b468.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 253, + 240, + 290, + 277 + ], + "blocks": [ + { + "bbox": [ + 253, + 240, + 290, + 277 + ], + "lines": [ + { + "bbox": [ + 253, + 240, + 290, + 277 + ], + "spans": [ + { + "bbox": [ + 253, + 240, + 290, + 277 + ], + "type": "image", + "image_path": "535f9ab309baccf653ff651bbb12a680f621de40c6344b9732d19432fbbecc52.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 278, + 274, + 283 + ], + "lines": [ + { + "bbox": [ + 253, + 278, + 274, + 283 + ], + "spans": [ + { + "bbox": [ + 253, + 278, + 274, + 283 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 61, + 283, + 99, + 320 + ], + "blocks": [ + { + "bbox": [ + 78, + 278, + 84, + 283 + ], + "lines": [ + { + "bbox": [ + 78, + 278, + 84, + 283 + ], + "spans": [ + { + "bbox": [ + 78, + 278, + 84, + 283 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 278, + 122, + 283 + ], + "lines": [ + { + "bbox": [ + 99, + 278, + 122, + 283 + ], + "spans": [ + { + "bbox": [ + 99, + 278, + 122, + 283 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 283, + 99, + 320 + ], + "lines": [ + { + "bbox": [ + 61, + 283, + 99, + 320 + ], + "spans": [ + { + "bbox": [ + 61, + 283, + 99, + 320 + ], + "type": "image", + "image_path": "5051a10275f509569b4bd3291eda3da090e36a48139f7c6a76cd1660304489d1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 321, + 84, + 327 + ], + "lines": [ + { + "bbox": [ + 77, + 321, + 84, + 327 + ], + "spans": [ + { + "bbox": [ + 77, + 321, + 84, + 327 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 333, + 295, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 333, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 295, + 365 + ], + "type": "text", + "content": "Figure 3. Simulated examples under different protocols. (a)-(h) show examples from training clients with known protocols, while (i)-(l) present examples from unseen protocols." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 100, + 283, + 137, + 320 + ], + "blocks": [ + { + "bbox": [ + 100, + 283, + 137, + 320 + ], + "lines": [ + { + "bbox": [ + 100, + 283, + 137, + 320 + ], + "spans": [ + { + "bbox": [ + 100, + 283, + 137, + 320 + ], + "type": "image", + "image_path": "a12e95d7688dc525aeb9f99fc69a05a986bb1b64df328f923766c068a791574c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 321, + 123, + 327 + ], + "lines": [ + { + "bbox": [ + 115, + 321, + 123, + 327 + ], + "spans": [ + { + "bbox": [ + 115, + 321, + 123, + 327 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 138, + 283, + 175, + 320 + ], + "blocks": [ + { + "bbox": [ + 138, + 278, + 146, + 283 + ], + "lines": [ + { + "bbox": [ + 138, + 278, + 146, + 283 + ], + "spans": [ + { + "bbox": [ + 138, + 278, + 146, + 283 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 283, + 175, + 320 + ], + "lines": [ + { + "bbox": [ + 138, + 283, + 175, + 320 + ], + "spans": [ + { + "bbox": [ + 138, + 283, + 175, + 320 + ], + "type": "image", + "image_path": "3dcab6b06dec54e402b18748666968c8e567a3ec44e8dd646d287029990b4cd2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 321, + 161, + 327 + ], + "lines": [ + { + "bbox": [ + 154, + 321, + 161, + 327 + ], + "spans": [ + { + "bbox": [ + 154, + 321, + 161, + 327 + ], + "type": "text", + "content": "(i)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 176, + 283, + 213, + 320 + ], + "blocks": [ + { + "bbox": [ + 176, + 278, + 198, + 283 + ], + "lines": [ + { + "bbox": [ + 176, + 278, + 198, + 283 + ], + "spans": [ + { + "bbox": [ + 176, + 278, + 198, + 283 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 283, + 213, + 320 + ], + "lines": [ + { + "bbox": [ + 176, + 283, + 213, + 320 + ], + "spans": [ + { + "bbox": [ + 176, + 283, + 213, + 320 + ], + "type": "image", + "image_path": "c840ccd4dc1b67661d13e7561a544f1438871505d987bd0a5df82c7cffbec3ac.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 321, + 199, + 327 + ], + "lines": [ + { + "bbox": [ + 192, + 321, + 199, + 327 + ], + "spans": [ + { + "bbox": [ + 192, + 321, + 199, + 327 + ], + "type": "text", + "content": "(1)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 214, + 283, + 252, + 320 + ], + "blocks": [ + { + "bbox": [ + 214, + 278, + 237, + 283 + ], + "lines": [ + { + "bbox": [ + 214, + 278, + 237, + 283 + ], + "spans": [ + { + "bbox": [ + 214, + 278, + 237, + 283 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 214, + 283, + 252, + 320 + ], + "lines": [ + { + "bbox": [ + 214, + 283, + 252, + 320 + ], + "spans": [ + { + "bbox": [ + 214, + 283, + 252, + 320 + ], + "type": "image", + "image_path": "0cf2042361aa975a2f1aef7bd6deda92d8d920493f4b81514cde0811f23b0475.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 321, + 237, + 327 + ], + "lines": [ + { + "bbox": [ + 229, + 321, + 237, + 327 + ], + "spans": [ + { + "bbox": [ + 229, + 321, + 237, + 327 + ], + "type": "text", + "content": "(k)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 365, + 295, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 446 + ], + "type": "text", + "content": "To fully leverage the diversity of client data, we globally share the hypernetworks in SCAN-PhysFed. However, projecting imaging features from different domains into a consistent, noise-free representation is challenging with a single shared decoder. Therefore, we propose protocol-specific decoders to effectively adapt to variations in scanning protocols across clients." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 448, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 518 + ], + "type": "text", + "content": "In PVQS, the unseen protocol code is quantized based on the nearest entry in the protocol codebook. Therefore, it is essential to ensure the protocol code is discriminative and informative for personalization. To achieve this, we introduce an orthogonal loss " + }, + { + "bbox": [ + 55, + 448, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{orth}}" + }, + { + "bbox": [ + 55, + 448, + 296, + 518 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 448, + 296, + 518 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 448, + 296, + 518 + ], + "type": "text", + "content": "-th client, defined as follows:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 528, + 295, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 528, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 104, + 528, + 295, + 561 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o r t h}} = \\sum_ {j = 1} ^ {K} \\left| \\mathbf {c} _ {i} \\cdot \\mathbf {c} _ {j} \\right| ^ {2} \\quad s. t. \\quad i \\neq j, \\tag {8}", + "image_path": "d6902979f02dc6d08ecdf68ad0ac5b741bc40da9e2ce1a02e9608a34ba3c6fb8.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "text", + "content": " is the number of clients (protocols), and " + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "text", + "content": " denotes the client index." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "type": "text", + "content": "The Mean Squared Error (MSE) is used as the imaging loss, and the total loss is formulated as:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 111, + 628, + 295, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 628, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 111, + 628, + 295, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {M S E}} (\\mathbf {Y}, \\hat {\\mathbf {Y}}) + \\tau \\mathcal {L} _ {\\text {o r t h}}, \\tag {9}", + "image_path": "c6389fd1101a4bded90f9a763bf89b646f87b5f19b68929c98109945dccf7e17.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": " represents the reconstructed image, and " + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": " denotes the temperature weight of " + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{orth}}" + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "To facilitate understanding of the implementation details, we provide the pseudocode of our method in Appendix A." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 314, + 237, + 394, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 237, + 394, + 251 + ], + "spans": [ + { + "bbox": [ + 314, + 237, + 394, + 251 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 257, + 553, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 553, + 339 + ], + "type": "text", + "content": "Experimental Setting. The proposed method was implemented using PyTorch and optimized by Adam [12] at a learning rate of 0.001. The number of communication rounds was set to 200 and the batch size was 20. The experiments were conducted a system equipped with an AMD Ryzen 75800X CPU and a single NVIDIA GTX 3080 Ti GPU." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 449 + ], + "type": "text", + "content": "Dataset. The \"2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge\" dataset [22], which contains 5,936 full-dose CT images from 10 patients, was used to evaluate the proposed method. In this study, 8 patients were used for training, while the remaining 2 were reserved for testing. To simulate a realistic scenario, we ensured that the patients for training did not overlap across different clients. The two patients for testing were simulated using different protocols to form the testing set." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "text", + "content": "Simulated examples are shown in Figure 3, where noticeable noise heterogeneity among clients can be observed, primarily due to differences in anatomy and scanning protocols. Detailed dataset preparation steps and scanning protocols are provided in Appendix B." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 509, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 603 + ], + "type": "text", + "content": "Baselines. We compared our SCAN-PhysFed approach with both generic and personalized FL methods. For generic FL methods, we include FedAvg [23], FedProx [16], FedNova [29], MOON [15], FedDG [19], and FedKD [32]. For personalized FL methods, we evaluated FedPer [3], FedBN [17], FedMRI [8], HyperFed [37], and FedFDD [6]. To ensure fairness, we kept the training settings consistent across all methods." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 604, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 553, + 662 + ], + "type": "text", + "content": "Evaluation Metrics. We use the peak signal-to-noise ratio (PSNR) to evaluate pixel-wise accuracy and the structural similarity (SSIM) to assess perceived visual quality [30]. For both metrics, higher values indicate better performance." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 671, + 488, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 488, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 488, + 684 + ], + "type": "text", + "content": "4.1. Comparison with other methods" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "We evaluate our method alongside other FL methods. To ensure fairness, we use the same imaging network, RED" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5158" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 76, + 545, + 373 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 545, + 373 + ], + "type": "image", + "image_path": "319310e54dc1ecebd8393ade11f4f345fd33b8283cecb1cfcdaace46c40bf822.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 384, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 555, + 418 + ], + "type": "text", + "content": "Figure 4. Qualitative results of six selected comparison methods and our method across different clients using the classical convolutional-based LDCT imaging network. Rows one to five correspond to Clients #2, #3, #5, #6, and #7, respectively. The display window for the first row is [-1024, 200] HU, while for the remaining rows, it is [-160, 240] HU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 430, + 293, + 531 + ], + "blocks": [ + { + "bbox": [ + 61, + 430, + 293, + 531 + ], + "lines": [ + { + "bbox": [ + 61, + 430, + 293, + 531 + ], + "spans": [ + { + "bbox": [ + 61, + 430, + 293, + 531 + ], + "type": "image", + "image_path": "f533115db85bd21fcbf9159a0a753652c41cfaeef2f79c76c3c97094afbee6b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "content": "Figure 5. Boxplots of the average results across all clients. (a) and (b) represent PSNR and SSIM, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "CNN [4], a state-of-the-art model in CT imaging. Quantitative results are provided in Table 1. Our method achieves the best performance across most clients, outperforming other generic and personalized FL methods in terms of average PSNR and SSIM. Furthermore, unlike other personalized methods that show instability across clients, such as HyperFed in Client #2 / #7 and FedFDD in Client #4, our method maintains consistent performance across varying protocols. We also observe that sparse-view protocols (Clients #2, #4, and #6) pose challenges for both generic and personalized FL methods, resulting in significant performance gaps compared to other protocols. How-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 429, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 452 + ], + "type": "text", + "content": "ever, by incorporating dual-level physics-informed prior, our method maintains robust performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 457, + 555, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 649 + ], + "type": "text", + "content": "Representative slices produced by different methods are shown in Figure 4. We compared the top three methods from both generic and personalized FL approaches. Representative clients were selected based on different radiation reduction strategies: sparse-view (Clients #2 and #6) and low-dose (Clients #3, #5, and #7) reconstructions. It can be observed that other methods exhibit noticeable noise or artifacts in sparse-view protocols, and fail to preserve details in low-dose protocols. In contrast, our method effectively reduces noise and artifacts in sparse-view protocols while preserving fine details in low-dose protocols. SCAN-PhysFed achieves strong performance across slices from different anatomical regions, benefiting from the introduction of anatomy priors. Both quantitative and qualitative results demonstrate the effectiveness of incorporating physics-informed priors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Additionally, we present boxplots of the results across all clients for different methods in Figure 5 to validate their stability. It can be seen that SCAN-PhysFed outperforms all other methods in both PSNR and SSIM scores. In both plots, SCAN-PhysFed demonstrates a higher median and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5159" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 89, + 555, + 509 + ], + "blocks": [ + { + "bbox": [ + 127, + 71, + 483, + 82 + ], + "lines": [ + { + "bbox": [ + 127, + 71, + 483, + 82 + ], + "spans": [ + { + "bbox": [ + 127, + 71, + 483, + 82 + ], + "type": "text", + "content": "Table 2. Quantitative Results of PSNR and SSIM for Transformer-based LDCT Imaging Methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 89, + 555, + 509 + ], + "lines": [ + { + "bbox": [ + 53, + 89, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 89, + 555, + 509 + ], + "type": "table", + "html": "
Client #1Client #2Client #3Client #4Client #5Client #6Client #7Client #8Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]34.3187.8838.0690.5443.5197.5641.0595.4345.2498.4743.1597.0941.4595.9542.8896.1641.2194.89
FedProx [16]34.3287.0137.7488.8943.7397.9541.1293.9245.3998.6142.9997.0841.3894.3142.1094.5541.1094.04
FedNova [29]29.5578.4431.3868.1337.0782.4731.7973.4540.6188.8035.4679.7732.3873.6635.7782.2234.2578.37
MOON [15]34.2893.2038.0991.1743.6497.7741.0995.6645.2598.4342.8296.9141.4295.9742.2096.1841.1095.66
FedKD [32]35.9289.9338.7091.2144.2698.1342.3095.9945.9398.7543.5797.4142.5496.2243.5196.2342.0995.48
Personalized FL Methods
FedPer [3]35.3793.1839.0393.5542.7597.1340.8896.2046.5598.8342.8796.6841.8796.7040.7994.7841.2695.82
FedMRI [8]34.6093.6738.0494.5542.9497.4439.3495.7846.7998.9142.8297.3040.8195.3240.8595.7040.7896.01
HyperFed [37]34.6087.5938.6291.0844.1397.6142.2596.2445.9298.5943.7197.3742.4496.6643.5696.6341.9095.22
FedFDD [6]37.3596.2039.6795.1844.3198.2639.0095.2147.8999.2043.8198.0642.1797.4040.5095.4941.8496.89
SCAN-PhysFed38.5594.0641.7496.2045.5298.5242.4896.5549.5099.3945.3297.9343.3596.7342.7196.7943.6597.03
Ground TruthInputFedAvgMOONFedKDFedMRIHyperFedFedFDDSCAN-PhysFed
", + "image_path": "905cc7a89ec4f2eefc841616c0f91e011d3d44a7b7517b7c47b0d230d3418e4d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 514, + 555, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 514, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 555, + 536 + ], + "type": "text", + "content": "Figure 6. Qualitative results of the compared methods across different clients based on a transformer-based imaging network. Rows one to five correspond to Clients #2, #3, #5, #6, and #7, respectively. The display window is set to [-160, 240] HU." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 547, + 295, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 295, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 295, + 595 + ], + "type": "text", + "content": "a narrower interquartile range, indicating strong stability in restoring LDCT data with varying noise distributions across different clients. Individual boxplots for each client are provided in Appendix C." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 603, + 201, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 201, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 201, + 615 + ], + "type": "text", + "content": "4.2. Generalization Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 622, + 132, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 132, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 132, + 633 + ], + "type": "text", + "content": "4.2.1 Backbone" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 712 + ], + "type": "text", + "content": "In addition to using RED-CNN as the imaging network in previous experiments, we evaluate our method and other FL methods with an alternative imaging network, Uformer [31], a transformer-based model for LDCT imaging. This comparison further validates the generalization capability of different FL methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 547, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 555, + 704 + ], + "type": "text", + "content": "The qualitative and quantitative results are shown in Figure 6 and Table 2, respectively. Consistent with previous experiments, our method achieves the best performance compared to other methods across most clients, with significantly higher average PSNR and SSIM values. This demonstrates the strong generalization capability of our method when integrated with different imaging networks. Among competing methods, FedFDD serves as a strong baseline. While FedFDD's SSIM is slightly higher than ours in some clients, our method achieves a significantly better PSNR, demonstrating superior detail preservation. Additionally, unlike FedFDD, which requires two imaging networks, our method achieves high performance with" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5160" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 96, + 90, + 507, + 235 + ], + "blocks": [ + { + "bbox": [ + 165, + 70, + 444, + 81 + ], + "lines": [ + { + "bbox": [ + 165, + 70, + 444, + 81 + ], + "spans": [ + { + "bbox": [ + 165, + 70, + 444, + 81 + ], + "type": "text", + "content": "Table 3. The Quantitative Results of PSNR and SSIM for the Unseen Clients." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 90, + 507, + 235 + ], + "lines": [ + { + "bbox": [ + 96, + 90, + 507, + 235 + ], + "spans": [ + { + "bbox": [ + 96, + 90, + 507, + 235 + ], + "type": "table", + "html": "
Unseen Client #1Unseen Client #2Unseen Client #3Unseen Client #4Average
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
Generic FL Methods
FedAvg [23]35.5788.6834.0085.2632.5276.7136.8691.4432.8981.35
FedProx [16]36.6991.1634.3086.3932.5872.1339.1394.8734.7485.52
FedNova [29]35.4588.8334.2386.2932.3969.2035.4089.9934.3883.58
MOON [15]35.2178.1935.3079.8134.5583.0735.5678.9435.1680.00
FedDG [19]34.8679.5434.7980.9133.1580.3134.4977.4234.3279.55
FedKD [32]36.8291.6034.2786.3732.4970.5440.7395.9036.0886.10
Personalized FL Methods
FedPer [3]36.8688.5835.4687.0231.9671.6438.5690.9035.7184.53
FedBN [17]37.8489.9536.1688.4633.6781.6340.2194.8936.9788.73
FedMRI [8]31.1282.4231.9579.9531.7271.9632.0481.6331.7178.99
HyperFed [37]36.1889.6033.4782.4931.9465.5241.7094.8535.9083.12
FedFDD [6]37.8092.3337.0691.7733.7180.2640.5395.1237.2789.87
SCAN-PhysFed40.7797.6938.7795.7534.3881.8040.2796.9238.5593.04
", + "image_path": "ec33a3a7abbf2f007e4be564c493578d72708db768c10d990743998f59d9e282.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 264, + 294, + 335 + ], + "blocks": [ + { + "bbox": [ + 61, + 244, + 288, + 255 + ], + "lines": [ + { + "bbox": [ + 61, + 244, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 244, + 288, + 255 + ], + "type": "text", + "content": "Table 4. Ablation on different components in SCAN-PhysFed." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 264, + 294, + 335 + ], + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 335 + ], + "type": "table", + "html": "
ScanningComponentAverage
AnatomyParadigmLossPSNRSSIM
XXGenericLMSE34.0384.70
✓XPersonalizedLMSE36.8288.78
X✓PersonalizedLMSE38.0494.14
✓✓PersonalizedLMSE40.6797.27
✓✓PersonalizedLMSE & Lorth41.1497.62
", + "image_path": "ece96145636ab9d272004ffef703672543f9ae00e3f4548dd7e52de314884a3c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 343, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 295, + 403 + ], + "type": "text", + "content": "out relying on an additional network. Similar to the previous experiment, SCAN-PhysFed shows better qualitative performance and lower variation across different protocols, indicating more consistent results irrespective of scanning protocols or imaging networks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 159, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 159, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 159, + 430 + ], + "type": "text", + "content": "4.2.2 Unseen Protocol" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 558 + ], + "type": "text", + "content": "Existing FL methods in medical imaging struggle to maintain performance when encountering unseen noise distributions from new clients. Here, we evaluate the effectiveness of the proposed PVQS across four unseen protocols. Detailed protocol settings can be found in Appendix B. For generic FL methods, we directly evaluate the shared model on unseen clients. However, personalized methods mostly ignore how to operate in the unseen domain, so we average the performance of all local models to establish a vanilla baseline in this paper." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "The results of different methods on unseen clients are presented in Table 3. We observe that the performance of generic FL methods shows only minor degradation on unseen clients, whereas personalized FL methods experience significant declines. This is because local optimization in personalized FL methods focuses heavily on each client's specific data distribution, underfitting other potential distributions and leading to greater performance drops on unseen clients. In PVQS, we prevent the network from directly mapping the unseen protocol to modulate the imaging network, as this could cause network collapse if the unseen protocol significantly differs from known ones. Instead, we quantize the unseen protocol using a protocol codebook, al" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 245, + 553, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 269 + ], + "type": "text", + "content": "lowing SCAN-PhysFed to leverage prior knowledge for better adaptation to unseen distributions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 278, + 408, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 408, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 408, + 292 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 298, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 490 + ], + "type": "text", + "content": "We evaluate the effectiveness of various components in SCAN-PhysFed, and the results are presented in Table 4. Compared to the baseline results, which exclude scanning and anatomy information, introducing either significantly improves performance, highlighting the importance of integrating physical prior in LDCT imaging. This approach substantially reduces learning complexity and improves the overall model performance. Moreover, introducing both types of physical information simultaneously is complementary rather than redundant, leading to further performance gains. Building on this, we use " + }, + { + "bbox": [ + 313, + 298, + 555, + 490 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{orth}}" + }, + { + "bbox": [ + 313, + 298, + 555, + 490 + ], + "type": "text", + "content": " to enhance the discriminative ability of the scanning code, thereby personalizing imaging features and effectively mitigating data heterogeneity. In the above studies, \"Personalized\" indicates that the decoder is tailored to each client. An ablation study on the personalized components is provided in Appendix D." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": "To preserve privacy and address heterogeneity in CT imaging, we propose a dual-level, physics-informed approach that utilizes both scanning- and anatomy-level prompts to enable personalized CT imaging with the assistance of an MLLM. This approach facilitates both patient- and scanning-level personalization, with extensive results demonstrating that incorporating physical prior significantly improves performance and effectively mitigates heterogeneity. Additionally, we propose a novel strategy, PVQS, to quantize protocol codes and maintain robust performance on unseen protocols. PVQS leverages known protocol codes to prevent latent collapse when encountering significantly different unseen protocols. In future work, we aim to develop a general FL framework for medical imaging that can support both CT and MRI denoising, which we believe holds promising potential." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 313, + 757 + ], + "type": "text", + "content": "5161" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62271335, in part by the Sichuan Science and Technology Program under Grant 2025ZNSFSC0470, and in part by Sichuan University \"From 0 to 1\" Innovative Research Program under Grant 2022SCUH0016." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 195, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 195, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 195, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 296, + 250 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "type": "text", + "content": "[2] Asma Alkhaldi, Raneem Alnajim, Layan Alabdullatef, Rawan Alyahya, Jun Chen, Deyao Zhu, Ahmed Alsinan, and Mohamed Elhoseiny. Minigpt-med: Large language model as a general interface for radiology diagnosis. arXiv preprint arXiv:2407.04106, 2024. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 295, + 341 + ], + "type": "text", + "content": "[3] Manoj Ghuhan Arivazhagan, Vinay Aggarwal, Aaditya Kumar Singh, et al. Federated learning with personalization layers. arXiv preprint arXiv:1912.00818, 2019. 3, 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 384 + ], + "type": "text", + "content": "[4] Hu Chen, Yi Zhang, Mannudeep K Kalra, et al. Low-dose ct with a residual encoder-decoder convolutional neural network. IEEE Trans. Med. Imaging, 36(12):2524-2535, 2017. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 387, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 296, + 453 + ], + "type": "text", + "content": "[5] Shixuan Chen, Boxuan Cao, Yinda Du, Yaoduo Zhang, Ji He, Zhaoying Bian, Dong Zeng, and Jianhua Ma. Federated condition generalization on low-dose ct reconstruction via cross-domain learning. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 47-56. Springer, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 498 + ], + "type": "text", + "content": "[6] Xuhang Chen, Zeju Li, Zikun Xu, Cheng Ouyang, Chen Qin, et al. Fedfdd: Federated learning with frequency domain decomposition for low-dose ct denoising. In Medical Imaging with Deep Learning, 2024. 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 500, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 294, + 554 + ], + "type": "text", + "content": "[7] Dongshang Deng, Xuangou Wu, Tao Zhang, Xiangyun Tang, Hongyang Du, Jiawen Kang, Jiqiang Liu, and Dusit Niyato. Fedasa: A personalized federated learning with adaptive model aggregation for heterogeneous mobile edge computing. IEEE Transactions on Mobile Computing, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 556, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 294, + 600 + ], + "type": "text", + "content": "[8] Chun-Mei Feng, Yunlu Yan, Shanshan Wang, Yong Xu, Ling Shao, and Huazhu Fu. Specificity-preserving federated learning for mr image reconstruction. IEEE Transactions on Medical Imaging, 42(7):2010-2021, 2022. 3, 5, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 601, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 294, + 645 + ], + "type": "text", + "content": "[9] Qi Gao, Zilong Li, Junping Zhang, Yi Zhang, and Hongming Shan. Corediff: Contextual error-modulated generalized diffusion model for low-dose ct denoising and generalization. IEEE Transactions on Medical Imaging, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 58, + 647, + 295, + 679 + ], + "type": "text", + "content": "[10] Hao Guan, Pew-Thian Yap, Andrea Bozoki, and Mingxia Liu. Federated learning for medical image analysis: A survey. Pattern Recognition, page 110424, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[11] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for feder-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "text", + "content": "[12] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[13] Gurucharan Marthi Krishna Kumar, Aman Chadha, Janine Mendola, and Amir Shmuel. Medvisionllama: Leveraging pre-trained large language model layers to enhance medical image segmentation. arXiv preprint arXiv:2410.02458, 2024.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "text", + "content": "[14] Danyang Li, Hao Wang, Jingyi Liao, Mingqiang Meng, Cuidie Zeng, Yuting Wang, Dong Zeng, and Jianhua Ma. Semi-centralized federated learning network for low-dose ct imaging. In Medical Imaging 2023: Physics of Medical Imaging, pages 1024-1028. SPIE, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[15] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 10713–10722, 2021. 2, 5, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 288, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 320 + ], + "type": "text", + "content": "[16] Tian Li, Anit Kumar Sahu, Manzil Zaheer, et al. Federated optimization in heterogeneous networks. In Proc. Mach. Learn. Syst. (MLSys), pages 429-450, 2020. 2, 5, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 322, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 365 + ], + "type": "text", + "content": "[17] Xiaoxiao Li, Meirui JIANG, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fedbn: Federated learning on non-iid features via local batch normalization. In International Conference on Learning Representations, 2021. 3, 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 367, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 367, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 367, + 553, + 421 + ], + "type": "text", + "content": "[18] Fenglin Liu, Tingting Zhu, Xian Wu, Bang Yang, Chenyu You, Chenyang Wang, Lei Lu, Zhangdaihong Liu, Yefeng Zheng, Xu Sun, et al. A medical multimodal large language model for future pandemics. NPJ Digital Medicine, 6(1): 226, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 422, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 487 + ], + "type": "text", + "content": "[19] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1013–1023, 2021. 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 490, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 555 + ], + "type": "text", + "content": "[20] Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Derong Xu, Feng Tian, and Yefeng Zheng. When moe meets llms: Parameter efficient fine-tuning for multi-task medical applications. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1104-1114, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 556, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 610 + ], + "type": "text", + "content": "[21] Zexin Lu, Qi Gao, Tao Wang, Ziyuan Yang, Zhiwen Wang, Hui Yu, Hu Chen, Jiliu Zhou, Hongming Shan, and Yi Zhang. Pridediff: Physics-regularized generalized diffusion model for ct reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 634 + ], + "type": "text", + "content": "[22] C McCollough. Tu-fg-207a-04: overview of the low dose ct grand challenge. Med. Phys., 43(6):3759–3760, 2016. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "text", + "content": "[23] Brendan McMahan, Eider Moore, Daniel Ramage, et al. Communication-efficient learning of deep networks from decentralized data. In Artificial Intelligence and Statistics, pages 1273-1282. PMLR, 2017. 2, 5, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[24] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medi" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5162" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "cal artificial intelligence. Nature, 616(7956):259-265, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "text", + "content": "[25] Saiprasad Ravishankar, Jong Chul Ye, and Jeffrey A Fessler. Image reconstruction: From sparsity to data-adaptive methods and machine learning. Proceedings of the IEEE, 108(1): 86-109, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[26] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 229 + ], + "type": "text", + "content": "[27] Ge Wang, Hengyong Yu, and Bruno De Man. An outlook on x-ray ct research and development. Medical physics, 35(3): 1051-1064, 2008. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 263 + ], + "type": "text", + "content": "[28] Ge Wang, Jong Chul Ye, and Bruno De Man. Deep learning for tomographic image reconstruction. Nature machine intelligence, 2(12):737-748, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 266, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 294, + 319 + ], + "type": "text", + "content": "[29] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 2, 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 322, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 294, + 364 + ], + "type": "text", + "content": "[30] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "type": "text", + "content": "[31] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 423, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 465 + ], + "type": "text", + "content": "[32] Chuhan Wu, Fangzhao Wu, Lingjuan Lyu, Yongfeng Huang, and Xing Xie. Communication-efficient federated learning via knowledge distillation. Nature Communications, 13(1): 2032, 2022. 3, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 294, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 500 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 500 + ], + "type": "text", + "content": "[33] Wenjun Xia, Qing Lyu, and Ge Wang. Low-dose ct using denoising diffusion probabilistic model for 20x speedup. arXiv preprint arXiv:2209.15136, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 502, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 294, + 555 + ], + "type": "text", + "content": "[34] Wenjun Xia, Ziyuan Yang, Zexin Lu, Zhongxian Wang, and Yi Zhang. Regformer: A local-nonlocal regularization-based model for sparse-view ct reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences, 8(2): 184-194, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 558, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 294, + 601 + ], + "type": "text", + "content": "[35] Xiyuan Yang, Wenke Huang, and Mang Ye. Fedas: Bridging inconsistency in personalized federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11986-11995, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "text", + "content": "[36] Ziyuan Yang, Yingyu Chen, Huijie Huangfu, Maosong Ran, Hui Wang, Xiaoxiao Li, and Yi Zhang. Dynamic corrected split federated learning with homomorphic encryption for u-shaped medical image networks. IEEE Journal of Biomedical and Health Informatics, 27(12):5946-5957, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[37] Ziyuan Yang, Wenjun Xia, Zexin Lu, Yingyu Chen, Xiaoxiao Li, and Yi Zhang. Hypernetwork-based physics-driven personalized federated learning for ct imaging. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 262 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[38] Mang Ye, Xiuwen Fang, Bo Du, Pong C Yuen, and Dacheng Tao. Heterogeneous federated learning: State-of-the-art and research challenges. ACM Computing Surveys, 56(3):1-44, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[39] Chenyu You, Guang Li, Yi Zhang, et al. Ct super-resolution gan constrained by the identical, residual, and cycle learning ensemble (gan-circle). IEEE Trans. Med. Imaging, 39(1): 188-203, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[40] Zhicheng Zhang, Lequan Yu, Xiaokun Liang, Wei Zhao, and Lei Xing. Transc: dual-path transformer for low dose computed tomography. In Medical Image Computing and Computer Assisted Intervention, pages 55–64. Springer, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[41] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large language models meet few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3065-3075, 2024. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5163" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/86d3746e-1656-4dc5-ab96-b8f696f1d699_content_list.json b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/86d3746e-1656-4dc5-ab96-b8f696f1d699_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..01b70350d553058c9df555cd9ad83753b8e3bf87 --- /dev/null +++ b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/86d3746e-1656-4dc5-ab96-b8f696f1d699_content_list.json @@ -0,0 +1,1571 @@ +[ + { + "type": "text", + "text": "Pattern Analogies: Learning to Perform Programmatic Image Edits by Analogy", + "text_level": 1, + "bbox": [ + 83, + 130, + 887, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aditya Ganeshan* Brown University", + "bbox": [ + 142, + 181, + 290, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thibault Groueix \nAdobe Research", + "bbox": [ + 334, + 181, + 473, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Paul Guerrero \nAdobe Research", + "bbox": [ + 517, + 181, + 651, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Radomír Měch \nAdobe Research", + "bbox": [ + 694, + 181, + 826, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthew Fisher Adobe Research", + "bbox": [ + 310, + 241, + 442, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daniel Ritchie Brown University", + "bbox": [ + 516, + 241, + 660, + 277 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd6ae179c58cf2692114081c8e54597143d51ca15a2478607cc048ccce89449f.jpg", + "image_caption": [ + "Figure 1. Our system performs programmatic edits on pattern images without inferring their underlying programs. (Left) Desired edits, expressed with a pair of patterns $(A, A')$ , are executed on a target pattern $B$ by a generative model to produce $B'$ . (Right) Parametric changes $A \\to A'$ enabled by our domain-specific pattern language induce corresponding changes to the more complex pattern $B$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 304, + 890, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 568, + 313, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pattern images are everywhere in the digital and physical worlds, and tools to edit them are valuable. But editing pattern images is tricky: desired edits are often programmatic: structure-aware edits that alter the underlying program which generates the pattern. One could attempt to infer this underlying program, but current methods for doing so struggle with complex images and produce unorganized programs that make editing tedious. In this work, we introduce a novel approach to perform programmatic edits on pattern images. By using a pattern analogy—a pair of simple patterns to demonstrate the intended edit—and a learning-based generative model to execute these edits, our method allows users to intuitively edit patterns. To enable this paradigm, we introduce SPLITWEAVE, a domain-specific language that, combined with a framework for sampling synthetic pattern analogies, enables the creation of a large, high-quality synthetic training dataset. We also present TRIFUSER, a Latent Diffusion Model (LDM) designed to overcome critical issues that arise when naively", + "bbox": [ + 75, + 613, + 473, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "deploying LDMs to this task. Extensive experiments on real-world, artist-sourced patterns reveals that our method faithfully performs the demonstrated edit while also generalizing to related pattern styles beyond its training distribution.", + "bbox": [ + 496, + 570, + 890, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 636, + 632, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual pattern designs enhance digital media such as presentations, website themes, and user interfaces, and they are woven into the physical world through textiles, wallpapers, and product designs like hardware covers. Given the ubiquity of patterns, methods for editing them are essential: designers should be able to quickly experiment with variations, customize designs to meet specific needs, and adapt existing patterns to align with evolving trends.", + "bbox": [ + 496, + 662, + 890, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Editing pattern images is not straightforward, as patterns are inherently structured, defined by rules that govern their layout and composition: tiling patterns adhere to principles of alignment and repetition (see Figure 1: top left), while retro-style designs rely on spatial divisions and fills (see Figure 1: bottom center). The edits that designers desire", + "bbox": [ + 496, + 784, + 890, + 875 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work performed during an internship at Adobe Research", + "bbox": [ + 517, + 887, + 823, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28715", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "often aim to adjust these underlying organizational rules rather than make superficial, pixel-level changes. We refer to such edits as programmatic edits, requiring manipulation of the underlying program that defines a pattern's structure.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One strategy for enabling such programmatic edits is visual program inference (VPI) [5, 32, 46], where a program that replicates an image is automatically inferred, allowing users to modify the image by adjusting program parameters. However, applying VPI to patterns presents two obstacles. First, VPI attempts to infer a program that fully replicates a pattern, which can be challenging as patterns are often semiparametric, blending rule-based logic with non-parametric components. For instance, the layout of elements in a tiling pattern may be rule-based, but the elements themselves may not be. Second, editing with an inferred program can be cumbersome, as they are often poorly-structured, with many unlabeled parameters, making them difficult to interpret. Consequently, VPI not only solves a more complex problem than necessary but also makes editing more challenging.", + "bbox": [ + 75, + 154, + 470, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Can we perform programmatic edits without inferring the underlying program? Doing so requires the ability to express and execute the edit—both without direct access to the program's parameters. To express a programmatic edit, it's crucial to specify both which underlying parameter(s) to change and how to modify them. We draw inspiration from how humans communicate transformations: through analogies. By providing a pair of simple example patterns $(A, A')$ that illustrate the desired change, users can intuitively convey both aspects of the edit. To execute these edits, we employ a learning-based conditional generative model. Given a pair of simple patterns $(A, A')$ and a complex target pattern $B$ , our system generates $B'$ , an edited version of $B$ which performs the transformation demonstrated between $A$ and $A'$ while preserving $B$ 's other structural features. Crucially, $A$ does not need to replicate or even be similar to $B$ —it only needs to demonstrate which property to edit and how. Thus, specifying $A$ is a much easier task than solving VPI. While prior works [1, 47, 51] have applied analogical editing to image manipulation, they focus primarily on appearance modifications. In contrast, our approach is the first to use analogies for programmatic, structure-aware edits. Figure 1 (left) shows examples of analogical editing on complex, real-world patterns.", + "bbox": [ + 75, + 383, + 470, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To make our approach possible, we introduce SPLITWEAVE: a domain-specific language (DSL) for crafting visual patterns. SPLITWEAVE serves two purposes in our method. First, it enables parametric definition of input pairs $(A,A^{\\prime})$ , allowing users to guide transformations in $(B,B^{\\prime})$ as if the underlying program for $B$ were accessible. In Figure 1 (right), modifying the SPLITWEAVE program for $A^\\prime$ produces corresponding changes in $B^{\\prime}$ . Second, SPLITWEAVE supports the creation of large-scale synthetic training data. We develop program samplers", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that generate high-quality patterns in two common styles: tiling-based designs with repeating elements and color field patterns characterized by splitting the canvas into intricate colored regions. Training a model for analogical editing requires a dataset of quartets $(A, A', B, B')$ . By applying identical programmatic edits to the SPLITWEAVE programs for both $A$ and $B$ to produce $A'$ and $B'$ , we ensure that the transformation from $A$ to $A'$ mirrors that from $B$ to $B'$ . This approach allows us to generate a diverse dataset of analogical quartets. Models trained on this dataset generalize effectively to real-world patterns within these styles and even extend to related styles.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use this synthetic dataset to train a novel diffusion-based conditional generative model for executing analogical edits. Our model directly generates edited patterns $B'$ by conditioning on visual features extracted from input patterns $(A, A', B)$ . Existing image-conditioned diffusion models [43, 54] prove ineffective, as they fail to interpret the input analogies accurately and neglect fine details. To address these issues, we incorporate architectural enhancements that enable our model, TRIFUSER, to effectively perform analogical edits. With these improvements, TRIFUSER surpasses prior architectures for analogical editing when applied to pattern images.", + "bbox": [ + 496, + 284, + 890, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate our method, we curated a test set of 50 patterns from Adobe Stock spanning 7 distinct styles. A perceptual study on this dataset reveals that participants consistently prefer edits by TRIFUSER over those produced by recent training-free and training-based analogical editing methods. Despite training only on two of these styles, our model successfully generalizes to the remaining five, which were unseen during training. On a synthetic validation set with ground-truth analogical edits, our model achieves higher structural and perceptual similarity to the ground truth than prior methods. Finally, we showcase two compelling applications of our approach: mixing attributes of different patterns and transferring pattern animations.", + "bbox": [ + 496, + 479, + 890, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 519, + 688, + 818, + 702 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A novel framework for performing programmatic edits to pattern images without requiring program inference, leveraging analogies to specify and apply edits.", + "2. SPLITWEAVE, a DSL for crafting a diverse range of visual patterns, designed to support both parametric control and synthetic dataset generation.", + "3. A procedure for generating synthetic analogical quartets, enabling editing of in-the-wild patterns.", + "4. TRIFUSER, a diffusion-based conditional generative model that achieves high fidelity in analogical edits, surpassing prior techniques in both analogical fidelity and generation quality." + ], + "bbox": [ + 511, + 715, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28716", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 218, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We review three key areas: (1) Visual Program Inference (VPI) for programmatic editing of structured visual data and its limitations, (2) DSLs and synthetic data generation, specifically for visual patterns, and (3) analogical reasoning in computing, particularly for editing images.", + "bbox": [ + 75, + 114, + 468, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Program Inference for Editing: Visual Program Inference (VPI) enables programmatic edits of visual data by inferring executable programs from visual inputs. Prior works have achieved promising results in inferring material graphs [19, 27, 31, 46] and CAD programs for 2D [11, 28] and 3D [42, 53] inputs, using large annotated datasets [46, 52], differentiable program approximations [19, 41], or bootstrapped learning [9, 23, 25]. VPI is challenging to adapt to pattern editing due to the scarcity of high-quality annotated pattern data and the non-differentiability of most pattern programs. Also, VPI approaches often yield complex programs that are difficult to edit and interpret, making them impractical for editing. To address these challenges, recent work has aimed to simplify programmatic editing by inferring edit-specific controls [3, 10, 15] or a limited set of semantically meaningful parameters [22, 24, 26, 56]. Our approach shares this goal of enabling accessible control but extends it further: we transfer control from simple parametric objects to complex in-the-wild images via analogy, bypassing the need for VPI.", + "bbox": [ + 75, + 190, + 470, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DSL and Synthetic data Domain-Specific Languages (DSLs) enable concise descriptions of structured objects, facilitating their creation. Prior works have developed DSLs for Zentangle patterns [45], material graphs [46], semiparametric textures [14], and 3D models [21, 38]. Our DSL focuses on visual patterns constructed through partitioning and merging of canvas fragments. Closest to our work is ETD [30], which also uses canvas partitioning and merging operators, though it is limited to stationary patterns.", + "bbox": [ + 75, + 492, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analogical Reasoning Analogical reasoning is a foundational AI task: early work includes Evans' ANALOGY program [6], CopyCat [18], and Structure-Mapping Engine [7]. In visual computing, Image Analogies [16] pioneered the concept of analogy-driven editing. Recently, diffusion models have been adapted for analogical editing. DIA [51] introduced a training-free approach to analogical editing using pretrained diffusion models. Analogist [13] offers a complementary method, leveraging inpainting models alongside multimodal reasoning from large language models [36]. These training-free approaches are limited to images within the diffusion model's training domain, limiting their applicability to patterns. Other methods attempt to learn analogical editors by finetuning diffusion models on analogical pairs [1, 34, 47]. However, the focus of all these works remains largely on stylistic, appearance edits, often failing to perform programmatic edits. This limitation arises both from the models' architectures", + "bbox": [ + 75, + 628, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/913e64525e76d1571c4d45cbccb7a748fef3f921fd26c16dcd740aea4e3d2b42.jpg", + "image_caption": [ + "Figure 2. Overview: To create high-quality visual patterns, we introduce a custom DSL called SPLITWEAVE. Pairs of SPLITWEAVE programs $(A,B)$ are then jointly edited to create analogical quartets. This synthetic data is then used to train TRI-FUSER, a neural network for analogical pattern editing." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 692, + 339 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/48e6af05707f57f7bf544a13dbf34e0afb2c37d716440149fdcb58ff236fb24d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 90, + 890, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and from the lack of training pairs with programmatic edits. Our work addresses both these gaps, enabling structured, programmatic analogical edits for visual patterns.", + "bbox": [ + 496, + 436, + 890, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 494, + 589, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to enable programmatic edits of 2D visual patterns without inferring their underlying programs. Instead, we propose an alternative that uses analogies to express desired edits and a conditional generative model to execute them. Formally, given two source patterns $A$ and $A'$ that demonstrate a desired edit, along with a target pattern $B$ , our goal is to generate an edited target pattern $B'$ that applies this edit to $B$ . This task is defined as learning a mapping $f(A, A', B) \\to B'$ , where $A, A', B$ , and $B'$ are 2D RGB images ( $\\in \\mathbb{R}^{H \\times W \\times 3}$ ). To learn this mapping, we generate a large synthetic dataset of analogical pattern quartets ( $A, A', B, B'$ ).", + "bbox": [ + 496, + 520, + 890, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 provides a schematic overview of our approach. First, in Section 3.1 we introduce SPLITWEAVE, a Domain-Specific Language (DSL) that enables the creation and manipulation of various kinds of patterns. Section 3.2 describes our approach for sampling analogical quartets in SPLITWEAVE to create the synthetic training data. Finally, in Section 3.3, we present TRIFUSER, a conditional generative model that learns to execute analogical edits.", + "bbox": [ + 496, + 700, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. A Language for Visual Patterns", + "text_level": 1, + "bbox": [ + 500, + 832, + 777, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable programmatic edits without program inference, our approach requires two core capabilities: (a) generating a large, high-quality synthetic dataset for training models to", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28717", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/edcb66e9ed39bf509518c9da09cab9e3ab775863a271893b17d82084cccc2a62.jpg", + "image_caption": [ + "Figure 3. Custom program samplers for two pattern styles. Our samplers produce diverse and high-quality patterns, enabling generalization to real-world patterns." + ], + "image_footnote": [], + "bbox": [ + 86, + 90, + 462, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reliably execute analogical edits, and (b) the ability to create and parametrically control analogy inputs at test time to effectively express desired edits. Existing pattern generation tools are insufficient for these needs, as they are either limited to narrow pattern domains [45] or demand intense coding effort to produce diverse, high-quality patterns [30, 33]. To address these limitations, we introduce SPLITWEAVE, a DSL designed specifically to support analogical transformations in visual patterns. SPLITWEAVE combines abstractions for pattern synthesis with a node-based visual programming interface (see Supplementary), enabling efficient generation of high-quality synthetic patterns for training while allowing flexible, precise pattern manipulation to define analogy inputs at test time.", + "bbox": [ + 75, + 340, + 467, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SPLITWEAVE uses three types of operations for structured pattern creation: (1) Canvas Fragmentation, which allows structured divisions of the canvas, such as brick-like or voronoi splits; (2) Fragment ID-Aware Operations, enabling transformations that vary across fragments (e.g., scaling alternating rows or columns) to support spatial variability in non-stationary pattern designs; and (3) various SVG Operators for outlining, coloring, and compositing. Together, these operations enable efficient creation of patterns with complex structure and visual variety. Figure 2 (left) illustrates these capabilities in a SPLITWEAVE program for generating a tiling pattern design.", + "bbox": [ + 75, + 553, + 467, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to generate high-quality synthetic patterns using SPLITWEAVE that enable trained editing models to generalize well to real-world patterns. Naive sampling from the DSL grammar often leads to overly complex or incoherent patterns, limiting their effectiveness in model training. Instead, we draw inspiration from recent advances in fields such as geometric problem solving [50] and abstract reasoning [29], where tailored data generators have proven to be essential for tackling complex tasks. Following a similar approach, we design custom program samplers for two versatile and widely-used pattern styles. The first, Motif", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eb3ddf8bc4238274e4781425b7305a145a88a8aff3e59d7976c5ce23b9672deb.jpg", + "image_caption": [ + "Figure 4. We create synthetic analogical quartets $(A, A', B, B')$ with consistent edits between $A$ and $B$ pairs, providing data for training an analogical editing models." + ], + "image_footnote": [], + "bbox": [ + 501, + 92, + 856, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tiling Patterns (MTP), consists of compositions based on repeated Tile elements. These patterns exhibit controlled variations in tile properties across the canvas (e.g. orientation, color, and scale), creating visually cohesive yet richly diverse structures. The second, Split-Filling Patterns (SFP), are generated by dividing the canvas into ordered fragments, applying region-specific coloring and transformations based on fragment IDs. Both pattern styles are common in digital design and support a wide range of programmatic variations, making them particularly suited for analogical editing tasks. Example patterns generated by our program samplers are shown in Figure 3; additional implementation details are in the supplementary materials.", + "bbox": [ + 496, + 297, + 890, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Sampling Analogical Quartets", + "text_level": 1, + "bbox": [ + 500, + 503, + 767, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the ability to generate diverse synthetic patterns using SPLITWEAVE (Section 3.1), our goal is now to construct analogical pattern quartets $(A, A', B, B')$ . Each pattern image in a quartet is generated by a SPLITWEAVE program $z$ . These quartets serve as structured training data for editing models, allowing them to learn consistent transformations that can generalize across different pattern domains.", + "bbox": [ + 496, + 526, + 890, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analogies in our framework are grounded in Structure Mapping Theory [12], which defines analogies as mappings of relational structure from a base to a target domain. We designate $(A, A')$ as the base and $(B, B')$ as the target, with the requirement that the relationship $R$ between program pairs $(z_A, z_{A'})$ and $(z_B, z_{B'})$ remains consistent:", + "bbox": [ + 496, + 632, + 890, + 723 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR \\left(z _ {A}, z _ {A ^ {\\prime}}\\right) = R \\left(z _ {B}, z _ {B ^ {\\prime}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 736, + 890, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rather than focusing on visual similarity between the patterns $(A, A')$ themselves, this program-level analogy allows us to generate quartets with transformations that affect the underlying program, facilitating programmatic edits.", + "bbox": [ + 496, + 763, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To construct these analogical quartets, we use a program sampler along with a predefined set of editing operators $E$ . For each quartet, we begin by sampling an edit $e \\in E$ , followed by sampling initial programs $z_{A}$ and $z_{B}$ that are compatible with $e$ . Applying $e$ to both $z_{A}$ and $z_{B}$ yields trans", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28718", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e58a36e6a34b85b90780df24aa9f1019d13ffdbb202fceb99a55e70fdcabae2.jpg", + "image_caption": [ + "Figure 5. (Left) TRIFUSER is a latent diffusion model conditioned on patch-wise tokens of the input images $(A, A', B)$ to generate the analogically edited pattern $B'$ . (Right) To achieve high-quality edits, we enrich these tokens by fusing multi-level features from multiple encoders, followed by a 3D positional encoding: 2D to specify spatial locations and 1D to specify the token's source $(A, A'$ or $B$ )." + ], + "image_footnote": [], + "bbox": [ + 94, + 93, + 468, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9523dd832b579fce58bd2582e931c69429c263448d226604391f7aa1cf1311bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 93, + 890, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "formed programs $z_{A'}$ and $z_{B'}$ . By using identical transformations across domains, we ensure a consistent \"edit relation\" across the quartet, satisfying Equation 1 by construction. In Figure 4, we illustrate examples of synthetic analogical quartets generated using this method, demonstrating consistent transformations between $(A, A')$ and $(B, B')$ .", + "bbox": [ + 75, + 304, + 467, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Edit Operators $E$ . We focus on edits targeting specific sub-parts of the program. Specifically, we consider three types of edits: insertion, removal, and replacement of subprograms. For example, an edit operator might replace the sub-program responsible for splitting the canvas, while other edits may insert or remove tiles within the pattern. Please refer to the supplementary for more details.", + "bbox": [ + 75, + 395, + 467, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Learning an Analogical Editor", + "text_level": 1, + "bbox": [ + 76, + 513, + 346, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to train a model on the synthetic data that is capable of performing analogical edits on real, in-the-wild patterns. Specifically, we aim to generate the target pattern $B'$ from an input triplet $(A, A', B)$ . This approach allows users to demonstrate desired edits with a simple pattern pairs $(A, A')$ , which the model then applies to a complex pattern $B$ to produce $B'$ . Given the success of Latent Diffusion Models (LDMs) in various generative modeling tasks [43], we chose to adapt an LDM for our task as well. We propose TRIFUSER, a latent diffusion model (LDM) for analogical editing (Figure 5). We provide a brief overview of LDMs to provide context before detailing TRIFUSER's modifications for analogical editing.", + "bbox": [ + 75, + 537, + 467, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preliminaries: Denoising Diffusion Probabilistic Models (DDPMs) [17] transform random noise into structured data via reverse diffusion steps guided with a conditioning embedding $c(y)$ (often derived from text). Latent Diffusion Models (LDMs) extend DDPMs by mapping data to a lower-dimensional latent space via an encoder. During training, a UNet model [44] learns to remove noise introduced into the latents. During inference, a latent sampled from a normal distribution is iteratively denoised by the model to yield a clean latent. Finally, the clean latent is decoded to generate the output image. Please refer to [55] for", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a more thorough overview. For analogical editing we adapt an Image Variation (IM) model [54], which uses patch-wise image tokens extracted using a text-image encoder [39] as the conditioning embedding $c(y)$ .", + "bbox": [ + 496, + 304, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The simplest adaptation of an IM model to our task is to generate $B'$ conditioned on image tokens from all three input images, concatenated as $C = c(A) \\| c(A') \\| c(B)$ , where $\\| \\cdot$ denotes token-wise concatenation. This approach, however, suffers from three drawbacks: Token Entanglement, Semantic Bias, and Detail Erosion. We discuss each of these issues briefly, along with our solutions.", + "bbox": [ + 496, + 364, + 890, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detail Erosion: Despite using patch-wise tokens, the extracted features lack the fine-grained information needed to retain key aspects of $B$ in the generated pattern $B'$ . Consequently, the model often struggles to preserve elements like tile textures. To address this problem, we combine features from both the first and last layers of the feature encoder:", + "bbox": [ + 496, + 470, + 890, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC _ {h l} (P) = \\operatorname {L i n e a r} \\left(L N \\left(c _ {\\text {h i g h}} (P)\\right) \\cdot L N \\left(c _ {\\text {l o w}} (P)\\right), \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 574, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $LN$ is layer normalization, $\\cdot$ denotes channel-wise concatenation, $P$ is an input pattern, and Linear is a linear projection layer that fuses low- and high-level features.", + "bbox": [ + 496, + 602, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Bias: Image variation models typically use feature extractors such as CLIP [39], which are trained to align image embeddings with corresponding text embeddings. Such embeddings emphasize high-level semantics but lack spatial and fine-grained visual details. Combining these embeddings with features from text-free, self-supervised extractors, such as DiNO [2], has been shown to improve performance in various downstream tasks [20, 49]. For our task, a similar approach—combining features from both text-image $(m_{1})$ and self-supervised $(m_{2})$ feature extractors—significantly enhances generation quality. The extracted features are fused as follows:", + "bbox": [ + 496, + 648, + 890, + 828 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC _ {m i x} (P) = \\operatorname {M i x e r} \\left(C _ {h l} ^ {m _ {1}} (P) \\cdot C _ {h l} ^ {m _ {2}} (P)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 840, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\text{Mixer}$ is a two-layer MLP that integrates features from the two extractors.", + "bbox": [ + 496, + 869, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28719", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Token Entanglement: To successfully perform an analogical edit, for each patch-level feature token, the model must be able to identify to which source image $(A, A', \\text{or } B)$ that patch belongs as well as the 2D position of the patch within that image. Without these distinctions, the model often fails to identify the pattern to edit (i.e., $B$ ) and to recognize the desired edit from $(A, A')$ . To address this problem, we introduce 3D positional encodings: two dimensions for spatial location within each pattern and one dimension for the source image. These encodings are applied to the extracted embeddings, yielding:", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC ^ {\\Omega} = C _ {P E} (A) \\parallel C _ {P E} \\left(A ^ {\\prime}\\right) \\parallel C _ {P E} (B), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 270, + 468, + 287 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC _ {P E} (P) ^ {x y} = C _ {m i x} (P) ^ {x y} + P E \\left(t _ {P}, x, y\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 290, + 468, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $t_P$ is a one-hot vector encoding which input image a token comes from and $PE(t_P, x, y)$ positionally encodes both spatial and source information for each token.", + "bbox": [ + 75, + 321, + 468, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As we demonstrate in Section 4.4, conditioning on $C^\\Omega$ instead of $C$ significantly enhances the quality of patterns generated by TRIFUSER. Our adapted architecture, shown in Figure 5, integrates the modifications described above to effectively address the described drawbacks. Note that this architecture is designed for general analogical editing, not just pattern editing. To enhance generalizability to real-world patterns, we initialize TRIFUSER with an existing pretrained IM model [54], and fine-tune only the denoising UNet and the projection layers in our feature extractor.", + "bbox": [ + 75, + 367, + 472, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 535, + 202, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate our approach along three directions: (1) the effectiveness of TRIFUSER at performing analogical edits on complex, real-world patterns, emphasizing how our synthetic data enables editing of in-the-wild pattern images; (2) the ability of TRIFUSER to support programmatic, structure-preserving edits without explicit program inference; and (3) the impact of architectural modifications introduced in TRIFUSER on the quality of generated patterns. We conduct a human perceptual study, quantitative assessments, and qualitative comparisons to demonstrate our system's ability to perform high-quality analogical edits across a range of pattern types.", + "bbox": [ + 75, + 561, + 468, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Design", + "text_level": 1, + "bbox": [ + 76, + 755, + 261, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: We generate a large synthetic dataset of analogical quartets, i.e., pairs of analogical patterns $(A, A', B, B')$ , using the SPLITWEAVE program samplers introduced in Section 3.2. This synthetic dataset contains approximately 1 million samples covering two pattern styles, namely Split Filling Patterns (SFP) and Motif Tiling Patterns (MTP) (cf. Section 3.1). For MTP patterns, we synthesize $100k$ distinct tiles using the LayerDiffuse [58] model, guided by text", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "prompts derived from WordNet [35] noun synsets. Additionally, we construct a synthetic test set with 1000 analogical quartets to evaluate model performance on unseen synthetic data. Further details on dataset construction are provided in the supplementary material.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess TRIFUSER on real-world patterns, we curate a test dataset of 50 patterns created by professional artists and sourced from Adobe Stock. This dataset spans seven distinct sub-domains of 2D patterns, representing a range of pattern styles. These styles include MTP and SFP patterns as well as previously unseen pattern styles such as Memphis-style, geometric, and digital textile patterns. Each pattern is annotated with a desired edit, and we use SPLITWEAVE to generate a pair of simpler patterns $(A, A')$ demonstrating this edit. This test set provides a challenging benchmark to evaluate TRIFUSER's generalization to diverse, real-world editing tasks.", + "bbox": [ + 496, + 167, + 892, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details: We fine-tune a pre-trained diffusion model using our synthetic dataset of analogical quartets (cf. Section 3.3). We initialize our model with Versatile-Diffusion's Image Variation model [54]. We use SigLIP [57] as our text-image feature encoder and DiNOv2 [37] for self-supervised features. We fine-tune the model on 8 A100 GPUs using a batch size of 224 for $\\sim 65$ epochs over 7 days. During inference, $B^{\\prime}$ is inferred with typical diffusion parameter settings such as a classifier-free guidance weight of 7.5 and 50 denoising steps.", + "bbox": [ + 496, + 349, + 893, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Analogical Editing Baselines", + "text_level": 1, + "bbox": [ + 498, + 512, + 756, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the analogical editing capability of TRIFUSER, we compare it to three baseline methods, each representing a leading approach for analogical image editing.", + "bbox": [ + 496, + 536, + 890, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we consider training-free editors and latent arithmetic editors. Training-free editors repurpose pre-trained diffusion models to perform analogical edits without additional training [13, 51], leveraging the rich representations learned by diffusion models for editing. In this category, we compare against Analogist [13], the current state-of-the-art method. Latent arithmetic editors, on the other hand, rely on transformations in a learned latent space to infer analogical modifications [40, 48]. Note that these approaches only require samples from the target domain, not analogical training pairs. We implement a baseline for this method by fine-tuning a naive Image Variation model [54] on our synthetic dataset to learn a generative latent embedding space of patterns. At inference, analogical edits are generated using latent arithmetic: given patterns $A$ , $A'$ , and $B$ , we condition the generation of $B'$ on $E(B) + E(A') - E(A)$ . We refer to this baseline as LatentMod.", + "bbox": [ + 496, + 583, + 892, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we consider analogy-conditioned generative editors, where models are explicitly trained on analogical data to learn analogical transformations [47]. This category includes our proposed TRIFUSER as well. Image Brush,", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28720", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dadfce37205e3ab655e8f97912c11b153cf00a5e9bb36ef31fce493307676bde.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison between patterns generated by our model, TRIFUSER, and the baselines. TRIFUSER generates higher quality patterns with greater fidelity to the input analogy." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 313 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fcb1a5b85a5ce01412329e4e77f5656e3f25dede10773b21f0fe4b51a22c053f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Preference Rate
TRIFUSER vs. Analogist89.24%
TRIFUSER vs. LatentMod80.59%
TRIFUSER vs. Inpainter72.05%
", + "bbox": [ + 114, + 373, + 431, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Results of a two-alternative forced-choice perceptual study comparing our model (TRIFUSER) against three baselines. Ours is preferred in the overwhelming majority of judgments.", + "bbox": [ + 75, + 460, + 468, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the state-of-the-art method, fine-tunes a diffusion inpainting model for analogical editing with multi-modal conditioning. Since code for Image Brush is unavailable, we implement a similar baseline by fine-tuning a Stable Diffusion inpainting model. This model, which we term Inpainter, performs analogical editing by inpainting the lower-left quadrant of a 2x2 analogy grid containing $(A, A', B)$ and conditioned on a fixed text template.", + "bbox": [ + 75, + 518, + 468, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Editing Real-World Patterns", + "text_level": 1, + "bbox": [ + 76, + 648, + 333, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate TRIFUSER's real-world analogical editing capabilities, we conducted a human preference study on the curated test set of Adobe Stock patterns.", + "bbox": [ + 75, + 672, + 468, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We performed a two-alternative forced-choice perceptual study comparing TRIFUSER with baseline methods on all 50 entries in the test set. Each method generates $k = 9$ outputs for each input tuple, and we select the best one based on visual inspection. Participants were shown edited patterns generated by two different methods along with the input patterns $(A, A', B)$ and instructed to select the edit that best preserved the analogical relationship and exhibited higher image quality. We recruited 42 participants for the study, resulting in a total of 1720 total judgments.", + "bbox": [ + 75, + 719, + 468, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 presents the results, showing that TRIFUSER was preferred over all three baselines. Due to the domain gap", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac7c32ef2eda801bc4abff8e6c77245fbaa55a6678fb9bd1d1330fc83abc2c27.jpg", + "image_caption": [ + "Figure 7. TRIFUSER effectively edits patterns from novel pattern styles not present in the training dataset. TRIFUSER shows a noteworthy ability to generalize beyond its training distribution." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 888, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "between the training data of the underlying model [43] and pattern images, Analogist fails to interpret and edit pattern images. Meanwhile, LatentMod fails to perform reasonable edits as the embedding space lacks the low-level details necessary for programmatic edits. While these baselines perform adequately on stylistic edits, they are unsuitable for programmatic editing. When compared to Inpainter, TRI-FUSER was favored in $72.2\\%$ of comparisons. Both methods benefit from training on analogical quartets, yet Inpainter sacrifices pattern quality as it generates the edited pattern in only a quarter of the full canvas resolution.", + "bbox": [ + 496, + 397, + 890, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6 shows examples of pattern edits generated by TRIFUSER and the baselines, with our model consistently delivering superior results. In Figure 7, we show examples of TRIFUSER's edits on out-of-distribution pattern styles not present in the training set. These results suggest that our synthetic training data enables manipulation of real-world patterns, even extending to certain untrained pattern styles.", + "bbox": [ + 496, + 563, + 890, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Editing Synthetic Patterns", + "text_level": 1, + "bbox": [ + 500, + 679, + 738, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we evaluate TRIFUSER's ability to perform programmatic edits on the synthetic validation set, which contains ground truth patterns $B'$ . Ideally, this would involve verifying that the underlying program $z_{\\hat{B}'}$ of the generated pattern reflects the same transformation from $z_B$ as that between $z_A$ and $z_A'$ . However, this would require visual program inference on $\\hat{B}'$ , which is infeasible. Instead, we approximate this criterion by comparing the program outputs $\\hat{B}'$ and $B'$ to see if the visual results align with the intended transformation. To quantify this alignment, we use perceptual metrics—DSim [8], DIST [4] and LPIPS [59]—along with SSIM to capture pixel-level structural similarity.", + "bbox": [ + 496, + 702, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that analogies can have multiple valid interpreta", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28721", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0fb0af6ad33272bd0405ec9dd8fbdcce2e89e8f7db82239553ef0559cf4eb46d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DSIM (↓)DISTS (↓)LPIPS (↓)SSIM (↑)
Analogist0.4960.4320.6970.494
LatentMod0.2420.3200.6130.502
Inpainter0.0920.2560.3710.713
TRIFUSER0.0740.1840.3040.704
", + "bbox": [ + 83, + 88, + 464, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6e2cae2d974e1d2c2d3d9383921f0c1095318eea28afb9699bc25653903514cc.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation on the synthetic validation set shows that TRIFUSER generates patterns with higher perceptual similarity to the ground truth than the baselines." + ], + "table_footnote": [], + "table_body": "
DSIM (↓)DISTS (↓)LPIPS (↓)SSIM (↑)
TRIFUSER0.0740.1840.3040.704
- Pos. Enc.0.1470.2390.3830.659
- Lower0.0870.1960.3350.652
- Mix0.0980.2100.3450.682
Base [54]0.5850.4600.8150.435
", + "bbox": [ + 83, + 234, + 464, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Subtractive ablation study on TRIFUSER shows that removing any component (see Section 3.3) degrades performance, and that removing all components (Base) results in a sharp decline.", + "bbox": [ + 76, + 354, + 468, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions, and even a single interpretation may yield several visually-related variations. To account for this multiplicity, we generate $k = 5$ output patterns for each input set $(A,A^{\\prime},B)$ and select the one that maximizes each metric. In other words, we evaluate whether at least one generated output aligns with the intended target.", + "bbox": [ + 76, + 414, + 468, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 shows the results of this experiment. First, we note that TRIFUSER outperforms others across all perceptual metrics. These metrics capture different aspects of perceptual similarity and superior performance across all of them suggests a comprehensive improvement. Second, we observe that the analogy-conditioned generative editors (Inpainter & TRIFUSER) surpass both the training-free and latent modification editors. Interestingly, Inpainter achieves slightly higher SSIM scores than TRIFUSER, suggesting that integrating the two approaches could be beneficial.", + "bbox": [ + 75, + 506, + 468, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. TRIFUSER Ablation", + "text_level": 1, + "bbox": [ + 76, + 665, + 261, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the impact of each model component introduced in Section 3.3, we conduct a subtractive analysis on the synthetic validation set. For this ablation study, we remove each component one at a time and measure the resulting performance, as reported in Table 3. The results demonstrate that removing any component leads to a performance drop, with the removal of 3D positional encoding leading to severe degradation. This is understandable: without 3D positional encoding, the network often fails to accurately identify which pattern to edit. For comparison, we also include results from the original Image Variation model [54] trained without any modifications (Base). As expected, this model performs poorly, underscoring the importance of our modifications in achieving high-quality analogical edits.", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd914bb398cbb9513f0cfc65c9f944a86415e8ca5ec02facf4a13f538bdaf2aa.jpg", + "image_caption": [ + "Figure 8. Our model helps users mix elements of different real-world patterns together, accelerating design exploration." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 632, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12acb91595b8d52a5af3c4b292685fb67ea1d037271e448d46ab21fa92a96126.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 90, + 756, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5ee99dfa369a5563cedcc90b05c09a90532b549b91cede03b12bf456e588955.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 90, + 887, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Application", + "text_level": 1, + "bbox": [ + 500, + 239, + 622, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ability to edit patterns without requiring program inference unlocks new creative possibilities. We demonstrate two practical applications of analogical pattern editing:", + "bbox": [ + 498, + 267, + 890, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pattern Mixing: Figure 8 shows example of using our method to mix elements of two real-world patterns $X$ and $Y$ , allowing the user to create unique, hybrid designs. The Mix operator is implemented by using a synthetic pair $(A, A')$ to create a variant $X'$ of $X$ and then using the pair $(X, X')$ to specify an edit to $Y$ : $\\text{Mix}(X, Y) = f(X, X', Y)$ , where $X' = f(A, A', X)$ . See the supplementary material for more details.", + "bbox": [ + 498, + 313, + 892, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Animation Transfer: TRIFUSER can also be used to create animated sequences of edited patterns. By leveraging parametric SPLITWEAVE programs, users can generate animations for simple patterns and then apply these animations to complex patterns with no additional effort. See the video in the supplementary material for examples.", + "bbox": [ + 498, + 434, + 890, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 542, + 617, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced a novel approach for programmatic editing of visual patterns without inferring the underlying program. By using analogies to express desired edits and a learned conditional generative model to execute them, our method provides an intuitive solution for pattern manipulation. A key component of our approach is SPLITWEAVE, a domain-specific language for generating diverse, structured pattern data. Paired with our procedure for sampling analogical quartets, SPLITWEAVE enables the creation of a large, high-quality dataset for training. We also presented TRIFUSER, a Latent Diffusion Model (LDM) designed to overcome critical issues that emerge when LDMs are naively deployed for analogical pattern editing, enabling high-fidelity edits that capture user intentions. Our experiments demonstrate that TRIFUSER successfully edits real-world patterns and surpasses baseline methods, while also generalizing to novel pattern styles beyond its training distribution. We believe that our DSL, dataset, and model will help drive further research on in-the-wild pattern image editing. Looking forward, we aim to extend this analogical editing framework to other domains such as semi-parametric 3D modeling while continuing to improve synthetic data generation.", + "bbox": [ + 496, + 568, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28722", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei A. Efros. Visual prompting via image inpainting. arXiv preprint arXiv:2209.00647, 2022. 2, 3", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5", + "[3] Ta-Ying Cheng, Matheus Gadelha, Thibault Groueix, Matthew Fisher, Radomir Mech, Andrew Markham, and Niki Trigoni. Learning continuous 3d words for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6753-6762, 2024. 3", + "[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P. Simoncelli. Image quality assessment: Unifying structure and texture similarity. CoRR, abs/2004.07728, 2020. 7", + "[5] Kevin Ellis, Daniel Ritchie, Armando Solar-Lezama, and Josh Tenenbaum. Learning to infer graphics programs from hand-drawn images. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 2", + "[6] Thomas G. Evans. A heuristic program to solve geometric-analogy problems. In Proceedings of the April 21-23, 1964, Spring Joint Computer Conference, page 327-338, New York, NY, USA, 1964. Association for Computing Machinery. 3", + "[7] Brian Falkenhainer, Kenneth D. Forbus, and Dedre Gentner. The structure-mapping engine. In Proceedings of the Fifth AAAI National Conference on Artificial Intelligence, page 272-277. AAAI Press, 1986. 3", + "[8] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[9] Aditya Ganeshan, R. Kenny Jones, and Daniel Ritchie. Improving unsupervised visual program inference with code rewriting families. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3", + "[10] Aditya Ganeshan, Ryan Y. Huang, Xianghao Xu, R. Kenny Jones, and Daniel Ritchie. Parsel: Parameterized shape editing with language, 2024. 3", + "[11] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Advances in Neural Information Processing Systems, pages 5885–5897. Curran Associates, Inc., 2021. 3", + "[12] Dedre Gentner. Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2):155-170, 1983. 4", + "[13] Zheng GU, Shiyuan Yang, Jing Liao, Jing Huo, and Yang Gao. Analogist: Out-of-the-box visual in-context learning with image diffusion model. ACM Transactions on Graphics (TOG), 2024. 3, 6", + "[14] P. Guehl, R. Allegre, J.-M. Dischler, B. Benes, and E. Galin. Semi-procedural textures using point process texture basis functions. Computer Graphics Forum, 39(4):159-171, 2020. 3" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Julia Guerrero-Viu, Milos Hasan, Arthur Roullier, Midhun Harikumar, Yiwei Hu, Paul Guerrero, Diego Gutierrez, Bellen Masia, and Valentin Deschaintre. Texsliders: Diffusion-based texture editing in clip space. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 3", + "[16] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image Analogies. Association for Computing Machinery, New York, NY, USA, 1 edition, 2023. 3", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 5", + "[18] Douglas Hofstadter and Melanie Mitchell. The Copycat project: a model of mental fluidity and analogy-making, page 205-267. Basic Books, Inc., USA, 1995. 3", + "[19] Yiwei Hu, Paul Guerrero, Milos Hasan, Holly Rushmeier, and Valentin Deschaintre. Node graph optimization using differentiable proxies. In ACM SIGGRAPH Conference Proceedings, 2022. 3", + "[20] Dongsheng Jiang, Yuchen Liu, Songlin Liu, XIAOPENG ZHANG, Jin Li, Hongkai Xiong, and Qi Tian. From CLIP to DINO: Visual encoders shout in multi-modal large language models, 2024. 5", + "[21] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapeassembly: Learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics (TOG), Siggraph Asia 2020, 2020. 3", + "[22] R. Kenny Jones, David Charatan, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapemod: Macro operation discovery for 3d shape programs. ACM Transactions on Graphics (TOG), Siggraph 2021, 2021. 3", + "[23] R. Kenny Jones, Homer Walke, and Daniel Ritchie. Plad: Learning to infer shape programs with pseudo-labels and approximate distributions. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[24] R. Kenny Jones, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapecoder: Discovering abstractions for visual programs from unstructured primitives. ACM Transactions on Graphics (TOG), Siggraph 2023, 42(4), 2023. 3", + "[25] R. Kenny Jones, Renhao Zhang, Aditya Ganeshan, and Daniel Ritchie. Learning to edit visual programs with self-supervision. In Advances in Neural Information Processing Systems, 2024. 3", + "[26] Milin Kodnongbua, Benjamin Jones, Maaz Bin Safeer Ahmad, Vladimir Kim, and Adriana Schulz. Reparamcad: Zero-shot cad re-parameterization for interactive manipulation. In SIGGRAPH Asia 2023 Conference Papers, New York, NY, USA, 2023. Association for Computing Machinery. 3", + "[27] Beichen Li, Liang Shi, and Wojciech Matusik. End-to-end procedural material capture with proxy-free mixed-integer optimization. ACM Trans. Graph., 42(4), 2023. 3", + "[28] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J. Mitra. Sketch2cad: Sequential cad modeling by sketching in" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28723", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "context. ACM Trans. Graph. (Proceedings of SIGGRAPH Asia 2020), 39(6):164:1-164:14, 2020. 3", + "[29] Wen-Ding Li, Keya Hu, Carter Larsen, Yuqing Wu, Simon Alford, Caleb Woo, Spencer M. Dunn, Hao Tang, Michelangelo Naim, Dat Nguyen, Wei-Long Zheng, Zenna Tavares, Yewen Pu, and Kevin Ellis. Combining induction and transduction for abstract reasoning, 2024. 4", + "[30] Hugo Loi, Thomas Hurtut, Romain Vergne, and Joelle Thollot. Programmable 2d arrangements for element texture design. ACM Trans. Graph., 36(4), 2017. 3, 4", + "[31] Arman Maesumi, Dylan Hu, Krishi Saripalli, Vladimir Kim, Matthew Fisher, Soeren Pink, and Daniel Ritchie. One noise to rule them all: Learning a unified model of spatially-varying noise patterns. ACM Trans. Graph., 43(4), 2024. 3", + "[32] Jiayuan Mao, Xiuming Zhang, Yikai Li, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Program-Guided Image Manipulators. In International Conference on Computer Vision, 2019. 2", + "[33] L. McCarthy, C. Reas, and B. Fry. Getting Started with P5.js: Making Interactive Graphics in JavaScript and Processing. Maker Media, Incorporated, 2015. 4", + "[34] Zichong Meng, Changdi Yang, Jun Liu, Hao Tang, Pu Zhao, and Yanzhi Wang. Instructgie: Towards generalizable image editing. arXiv preprint arXiv:2403.05018, 2024. 3", + "[35] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 6", + "[36] OpenAI. Gpt-4 technical report, 2024. 3", + "[37] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024, 2024. 6", + "[38] Ofek Pearl, Itai Lang, Yuhua Hu, Raymond A. Yeh, and Rana Hanocka. Geocode: Interpretable shape programs. 2022. 3", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5", + "[40] Scott Reed, Yi Zhang, Yuting Zhang, and Honglak Lee. Deep visual analogy-making. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, page 1252-1260, Cambridge, MA, USA, 2015. MIT Press. 6", + "[41] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, Haiyong Jiang, Zhongang Cai, Junzhe Zhang, Liang Pan, Mingyuan Zhang, Haiyu Zhao, and Shuai Yi. Csg-stump: A learning friendly csg-like representation for interpretable" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shape parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3", + "[42] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, and Junzhe Zhang. Extrudenet: Unsupervised inverse sketch-and-extrude fornbsp; shape parsing. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, page 482-498, Berlin, Heidelberg, 2022. Springer-Verlag. 3", + "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 5, 7", + "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 5", + "[45] Christian Santoni and Fabio Pellacini. gtangle: a grammar for the procedural generation of tangle patterns. ACM Trans. Graph., 35(6), 2016. 3, 4", + "[46] Liang Shi, Beichen Li, Miloš Hasan, Kalyan Sunkavalli, Tamy Boubekeur, Radomir Mech, and Wojciech Matusik. Match: differentiable material graphs for procedural material capture. ACM Trans. Graph., 39(6), 2020. 2, 3", + "[47] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: learning visual in-context instructions for exemplar-based image manipulation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 2, 3, 6", + "[48] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 6", + "[49] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 5", + "[50] Trieu Trinh, Yuhuai Tony Wu, Quoc Le, He He, and Thang Luong. Solving olympiad geometry without human demonstrations. Nature, 625:476-482, 2024. 4", + "[51] Adéla Šubrtová, Michal Lukáč, Jan Čech, David Futschik, Eli Shechtman, and Daniel Sykora. Diffusion image analogies. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 3, 6", + "[52] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6772-6782, 2021. 3", + "[53] Xianghao Xu, Wenzhe Peng, Chin-Yi Cheng, Karl D. D. Willis, and Daniel Ritchie. Inferring cad modeling sequences using zone graphs. In CVPR, 2021. 3", + "[54] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7754-7765, 2023. 2, 5, 6, 8" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28724", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 5", + "[56] Mehmet Ersin Yumer, Siddhartha Chaudhuri, Jessica K. Hodgins, and Levent Burak Kara. Semantic shape editing using deformation handles. ACM Trans. Graph., 34(4), 2015. 3", + "[57] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 6", + "[58] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. ACM Trans. Graph., 43(4), 2024. 6", + "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + ], + "bbox": [ + 78, + 90, + 468, + 361 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28725", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/86d3746e-1656-4dc5-ab96-b8f696f1d699_model.json b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/86d3746e-1656-4dc5-ab96-b8f696f1d699_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4170c553847eaa01b6c0efe6c991557734f238eb --- /dev/null +++ b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/86d3746e-1656-4dc5-ab96-b8f696f1d699_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.084, + 0.131, + 0.888, + 0.155 + ], + "angle": 0, + "content": "Pattern Analogies: Learning to Perform Programmatic Image Edits by Analogy" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.182, + 0.291, + 0.218 + ], + "angle": 0, + "content": "Aditya Ganeshan* Brown University" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.182, + 0.474, + 0.216 + ], + "angle": 0, + "content": "Thibault Groueix \nAdobe Research" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.182, + 0.652, + 0.216 + ], + "angle": 0, + "content": "Paul Guerrero \nAdobe Research" + }, + { + "type": "text", + "bbox": [ + 0.695, + 0.182, + 0.828, + 0.216 + ], + "angle": 0, + "content": "Radomír Měch \nAdobe Research" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.242, + 0.444, + 0.276 + ], + "angle": 0, + "content": "Matthew Fisher Adobe Research" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.242, + 0.661, + 0.278 + ], + "angle": 0, + "content": "Daniel Ritchie Brown University" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.305, + 0.891, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.516, + 0.894, + 0.559 + ], + "angle": 0, + "content": "Figure 1. Our system performs programmatic edits on pattern images without inferring their underlying programs. (Left) Desired edits, expressed with a pair of patterns \\((A, A')\\), are executed on a target pattern \\(B\\) by a generative model to produce \\(B'\\). (Right) Parametric changes \\(A \\to A'\\) enabled by our domain-specific pattern language induce corresponding changes to the more complex pattern \\(B\\)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.569, + 0.314, + 0.585 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.474, + 0.902 + ], + "angle": 0, + "content": "Pattern images are everywhere in the digital and physical worlds, and tools to edit them are valuable. But editing pattern images is tricky: desired edits are often programmatic: structure-aware edits that alter the underlying program which generates the pattern. One could attempt to infer this underlying program, but current methods for doing so struggle with complex images and produce unorganized programs that make editing tedious. In this work, we introduce a novel approach to perform programmatic edits on pattern images. By using a pattern analogy—a pair of simple patterns to demonstrate the intended edit—and a learning-based generative model to execute these edits, our method allows users to intuitively edit patterns. To enable this paradigm, we introduce SPLITWEAVE, a domain-specific language that, combined with a framework for sampling synthetic pattern analogies, enables the creation of a large, high-quality synthetic training dataset. We also present TRIFUSER, a Latent Diffusion Model (LDM) designed to overcome critical issues that arise when naively" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.632 + ], + "angle": 0, + "content": "deploying LDMs to this task. Extensive experiments on real-world, artist-sourced patterns reveals that our method faithfully performs the demonstrated edit while also generalizing to related pattern styles beyond its training distribution." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.637, + 0.633, + 0.653 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.892, + 0.784 + ], + "angle": 0, + "content": "Visual pattern designs enhance digital media such as presentations, website themes, and user interfaces, and they are woven into the physical world through textiles, wallpapers, and product designs like hardware covers. Given the ubiquity of patterns, methods for editing them are essential: designers should be able to quickly experiment with variations, customize designs to meet specific needs, and adapt existing patterns to align with evolving trends." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.785, + 0.892, + 0.876 + ], + "angle": 0, + "content": "Editing pattern images is not straightforward, as patterns are inherently structured, defined by rules that govern their layout and composition: tiling patterns adhere to principles of alignment and repetition (see Figure 1: top left), while retro-style designs rely on spatial divisions and fills (see Figure 1: bottom center). The edits that designers desire" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.825, + 0.901 + ], + "angle": 0, + "content": "*Work performed during an internship at Adobe Research" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28715" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "often aim to adjust these underlying organizational rules rather than make superficial, pixel-level changes. We refer to such edits as programmatic edits, requiring manipulation of the underlying program that defines a pattern's structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.155, + 0.471, + 0.382 + ], + "angle": 0, + "content": "One strategy for enabling such programmatic edits is visual program inference (VPI) [5, 32, 46], where a program that replicates an image is automatically inferred, allowing users to modify the image by adjusting program parameters. However, applying VPI to patterns presents two obstacles. First, VPI attempts to infer a program that fully replicates a pattern, which can be challenging as patterns are often semiparametric, blending rule-based logic with non-parametric components. For instance, the layout of elements in a tiling pattern may be rule-based, but the elements themselves may not be. Second, editing with an inferred program can be cumbersome, as they are often poorly-structured, with many unlabeled parameters, making them difficult to interpret. Consequently, VPI not only solves a more complex problem than necessary but also makes editing more challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.471, + 0.747 + ], + "angle": 0, + "content": "Can we perform programmatic edits without inferring the underlying program? Doing so requires the ability to express and execute the edit—both without direct access to the program's parameters. To express a programmatic edit, it's crucial to specify both which underlying parameter(s) to change and how to modify them. We draw inspiration from how humans communicate transformations: through analogies. By providing a pair of simple example patterns \\((A, A')\\) that illustrate the desired change, users can intuitively convey both aspects of the edit. To execute these edits, we employ a learning-based conditional generative model. Given a pair of simple patterns \\((A, A')\\) and a complex target pattern \\(B\\), our system generates \\(B'\\), an edited version of \\(B\\) which performs the transformation demonstrated between \\(A\\) and \\(A'\\) while preserving \\(B\\)'s other structural features. Crucially, \\(A\\) does not need to replicate or even be similar to \\(B\\)—it only needs to demonstrate which property to edit and how. Thus, specifying \\(A\\) is a much easier task than solving VPI. While prior works [1, 47, 51] have applied analogical editing to image manipulation, they focus primarily on appearance modifications. In contrast, our approach is the first to use analogies for programmatic, structure-aware edits. Figure 1 (left) shows examples of analogical editing on complex, real-world patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To make our approach possible, we introduce SPLITWEAVE: a domain-specific language (DSL) for crafting visual patterns. SPLITWEAVE serves two purposes in our method. First, it enables parametric definition of input pairs \\((A,A^{\\prime})\\), allowing users to guide transformations in \\((B,B^{\\prime})\\) as if the underlying program for \\(B\\) were accessible. In Figure 1 (right), modifying the SPLITWEAVE program for \\(A^\\prime\\) produces corresponding changes in \\(B^{\\prime}\\). Second, SPLITWEAVE supports the creation of large-scale synthetic training data. We develop program samplers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "that generate high-quality patterns in two common styles: tiling-based designs with repeating elements and color field patterns characterized by splitting the canvas into intricate colored regions. Training a model for analogical editing requires a dataset of quartets \\((A, A', B, B')\\). By applying identical programmatic edits to the SPLITWEAVE programs for both \\(A\\) and \\(B\\) to produce \\(A'\\) and \\(B'\\), we ensure that the transformation from \\(A\\) to \\(A'\\) mirrors that from \\(B\\) to \\(B'\\). This approach allows us to generate a diverse dataset of analogical quartets. Models trained on this dataset generalize effectively to real-world patterns within these styles and even extend to related styles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.285, + 0.892, + 0.468 + ], + "angle": 0, + "content": "We use this synthetic dataset to train a novel diffusion-based conditional generative model for executing analogical edits. Our model directly generates edited patterns \\( B' \\) by conditioning on visual features extracted from input patterns \\( (A, A', B) \\). Existing image-conditioned diffusion models [43, 54] prove ineffective, as they fail to interpret the input analogies accurately and neglect fine details. To address these issues, we incorporate architectural enhancements that enable our model, TRIFUSER, to effectively perform analogical edits. With these improvements, TRIFUSER surpasses prior architectures for analogical editing when applied to pattern images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.676 + ], + "angle": 0, + "content": "To evaluate our method, we curated a test set of 50 patterns from Adobe Stock spanning 7 distinct styles. A perceptual study on this dataset reveals that participants consistently prefer edits by TRIFUSER over those produced by recent training-free and training-based analogical editing methods. Despite training only on two of these styles, our model successfully generalizes to the remaining five, which were unseen during training. On a synthetic validation set with ground-truth analogical edits, our model achieves higher structural and perceptual similarity to the ground truth than prior methods. Finally, we showcase two compelling applications of our approach: mixing attributes of different patterns and transferring pattern animations." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.689, + 0.82, + 0.703 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.716, + 0.891, + 0.761 + ], + "angle": 0, + "content": "1. A novel framework for performing programmatic edits to pattern images without requiring program inference, leveraging analogies to specify and apply edits." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.763, + 0.892, + 0.807 + ], + "angle": 0, + "content": "2. SPLITWEAVE, a DSL for crafting a diverse range of visual patterns, designed to support both parametric control and synthetic dataset generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.809, + 0.891, + 0.839 + ], + "angle": 0, + "content": "3. A procedure for generating synthetic analogical quartets, enabling editing of in-the-wild patterns." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "4. TRIFUSER, a diffusion-based conditional generative model that achieves high fidelity in analogical edits, surpassing prior techniques in both analogical fidelity and generation quality." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.716, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28716" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.22, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.191 + ], + "angle": 0, + "content": "We review three key areas: (1) Visual Program Inference (VPI) for programmatic editing of structured visual data and its limitations, (2) DSLs and synthetic data generation, specifically for visual patterns, and (3) analogical reasoning in computing, particularly for editing images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.191, + 0.471, + 0.493 + ], + "angle": 0, + "content": "Visual Program Inference for Editing: Visual Program Inference (VPI) enables programmatic edits of visual data by inferring executable programs from visual inputs. Prior works have achieved promising results in inferring material graphs [19, 27, 31, 46] and CAD programs for 2D [11, 28] and 3D [42, 53] inputs, using large annotated datasets [46, 52], differentiable program approximations [19, 41], or bootstrapped learning [9, 23, 25]. VPI is challenging to adapt to pattern editing due to the scarcity of high-quality annotated pattern data and the non-differentiability of most pattern programs. Also, VPI approaches often yield complex programs that are difficult to edit and interpret, making them impractical for editing. To address these challenges, recent work has aimed to simplify programmatic editing by inferring edit-specific controls [3, 10, 15] or a limited set of semantically meaningful parameters [22, 24, 26, 56]. Our approach shares this goal of enabling accessible control but extends it further: we transfer control from simple parametric objects to complex in-the-wild images via analogy, bypassing the need for VPI." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.63 + ], + "angle": 0, + "content": "DSL and Synthetic data Domain-Specific Languages (DSLs) enable concise descriptions of structured objects, facilitating their creation. Prior works have developed DSLs for Zentangle patterns [45], material graphs [46], semiparametric textures [14], and 3D models [21, 38]. Our DSL focuses on visual patterns constructed through partitioning and merging of canvas fragments. Closest to our work is ETD [30], which also uses canvas partitioning and merging operators, though it is limited to stationary patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Analogical Reasoning Analogical reasoning is a foundational AI task: early work includes Evans' ANALOGY program [6], CopyCat [18], and Structure-Mapping Engine [7]. In visual computing, Image Analogies [16] pioneered the concept of analogy-driven editing. Recently, diffusion models have been adapted for analogical editing. DIA [51] introduced a training-free approach to analogical editing using pretrained diffusion models. Analogist [13] offers a complementary method, leveraging inpainting models alongside multimodal reasoning from large language models [36]. These training-free approaches are limited to images within the diffusion model's training domain, limiting their applicability to patterns. Other methods attempt to learn analogical editors by finetuning diffusion models on analogical pairs [1, 34, 47]. However, the focus of all these works remains largely on stylistic, appearance edits, often failing to perform programmatic edits. This limitation arises both from the models' architectures" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.693, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.091, + 0.891, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.352, + 0.894, + 0.422 + ], + "angle": 0, + "content": "Figure 2. Overview: To create high-quality visual patterns, we introduce a custom DSL called SPLITWEAVE. Pairs of SPLITWEAVE programs \\((A,B)\\) are then jointly edited to create analogical quartets. This synthetic data is then used to train TRI-FUSER, a neural network for analogical pattern editing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.484 + ], + "angle": 0, + "content": "and from the lack of training pairs with programmatic edits. Our work addresses both these gaps, enabling structured, programmatic analogical edits for visual patterns." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.591, + 0.511 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Our objective is to enable programmatic edits of 2D visual patterns without inferring their underlying programs. Instead, we propose an alternative that uses analogies to express desired edits and a conditional generative model to execute them. Formally, given two source patterns \\( A \\) and \\( A' \\) that demonstrate a desired edit, along with a target pattern \\( B \\), our goal is to generate an edited target pattern \\( B' \\) that applies this edit to \\( B \\). This task is defined as learning a mapping \\( f(A, A', B) \\to B' \\), where \\( A, A', B \\), and \\( B' \\) are 2D RGB images (\\( \\in \\mathbb{R}^{H \\times W \\times 3} \\)). To learn this mapping, we generate a large synthetic dataset of analogical pattern quartets (\\( A, A', B, B' \\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.823 + ], + "angle": 0, + "content": "Figure 2 provides a schematic overview of our approach. First, in Section 3.1 we introduce SPLITWEAVE, a Domain-Specific Language (DSL) that enables the creation and manipulation of various kinds of patterns. Section 3.2 describes our approach for sampling analogical quartets in SPLITWEAVE to create the synthetic training data. Finally, in Section 3.3, we present TRIFUSER, a conditional generative model that learns to execute analogical edits." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.779, + 0.849 + ], + "angle": 0, + "content": "3.1. A Language for Visual Patterns" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To enable programmatic edits without program inference, our approach requires two core capabilities: (a) generating a large, high-quality synthetic dataset for training models to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28717" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.092, + 0.464, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.47, + 0.33 + ], + "angle": 0, + "content": "Figure 3. Custom program samplers for two pattern styles. Our samplers produce diverse and high-quality patterns, enabling generalization to real-world patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.468, + 0.552 + ], + "angle": 0, + "content": "reliably execute analogical edits, and (b) the ability to create and parametrically control analogy inputs at test time to effectively express desired edits. Existing pattern generation tools are insufficient for these needs, as they are either limited to narrow pattern domains [45] or demand intense coding effort to produce diverse, high-quality patterns [30, 33]. To address these limitations, we introduce SPLITWEAVE, a DSL designed specifically to support analogical transformations in visual patterns. SPLITWEAVE combines abstractions for pattern synthesis with a node-based visual programming interface (see Supplementary), enabling efficient generation of high-quality synthetic patterns for training while allowing flexible, precise pattern manipulation to define analogy inputs at test time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.468, + 0.734 + ], + "angle": 0, + "content": "SPLITWEAVE uses three types of operations for structured pattern creation: (1) Canvas Fragmentation, which allows structured divisions of the canvas, such as brick-like or voronoi splits; (2) Fragment ID-Aware Operations, enabling transformations that vary across fragments (e.g., scaling alternating rows or columns) to support spatial variability in non-stationary pattern designs; and (3) various SVG Operators for outlining, coloring, and compositing. Together, these operations enable efficient creation of patterns with complex structure and visual variety. Figure 2 (left) illustrates these capabilities in a SPLITWEAVE program for generating a tiling pattern design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our goal is to generate high-quality synthetic patterns using SPLITWEAVE that enable trained editing models to generalize well to real-world patterns. Naive sampling from the DSL grammar often leads to overly complex or incoherent patterns, limiting their effectiveness in model training. Instead, we draw inspiration from recent advances in fields such as geometric problem solving [50] and abstract reasoning [29], where tailored data generators have proven to be essential for tackling complex tasks. Following a similar approach, we design custom program samplers for two versatile and widely-used pattern styles. The first, Motif" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.093, + 0.857, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.242, + 0.892, + 0.284 + ], + "angle": 0, + "content": "Figure 4. We create synthetic analogical quartets \\((A, A', B, B')\\) with consistent edits between \\(A\\) and \\(B\\) pairs, providing data for training an analogical editing models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.892, + 0.493 + ], + "angle": 0, + "content": "Tiling Patterns (MTP), consists of compositions based on repeated Tile elements. These patterns exhibit controlled variations in tile properties across the canvas (e.g. orientation, color, and scale), creating visually cohesive yet richly diverse structures. The second, Split-Filling Patterns (SFP), are generated by dividing the canvas into ordered fragments, applying region-specific coloring and transformations based on fragment IDs. Both pattern styles are common in digital design and support a wide range of programmatic variations, making them particularly suited for analogical editing tasks. Example patterns generated by our program samplers are shown in Figure 3; additional implementation details are in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.504, + 0.768, + 0.52 + ], + "angle": 0, + "content": "3.2. Sampling Analogical Quartets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.633 + ], + "angle": 0, + "content": "With the ability to generate diverse synthetic patterns using SPLITWEAVE (Section 3.1), our goal is now to construct analogical pattern quartets \\((A, A', B, B')\\). Each pattern image in a quartet is generated by a SPLITWEAVE program \\(z\\). These quartets serve as structured training data for editing models, allowing them to learn consistent transformations that can generalize across different pattern domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.892, + 0.724 + ], + "angle": 0, + "content": "Analogies in our framework are grounded in Structure Mapping Theory [12], which defines analogies as mappings of relational structure from a base to a target domain. We designate \\((A, A')\\) as the base and \\((B, B')\\) as the target, with the requirement that the relationship \\(R\\) between program pairs \\((z_A, z_{A'})\\) and \\((z_B, z_{B'})\\) remains consistent:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.737, + 0.892, + 0.753 + ], + "angle": 0, + "content": "\\[\nR \\left(z _ {A}, z _ {A ^ {\\prime}}\\right) = R \\left(z _ {B}, z _ {B ^ {\\prime}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.891, + 0.825 + ], + "angle": 0, + "content": "Rather than focusing on visual similarity between the patterns \\((A, A')\\) themselves, this program-level analogy allows us to generate quartets with transformations that affect the underlying program, facilitating programmatic edits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To construct these analogical quartets, we use a program sampler along with a predefined set of editing operators \\( E \\). For each quartet, we begin by sampling an edit \\( e \\in E \\), followed by sampling initial programs \\( z_{A} \\) and \\( z_{B} \\) that are compatible with \\( e \\). Applying \\( e \\) to both \\( z_{A} \\) and \\( z_{B} \\) yields trans" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28718" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.094, + 0.47, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.094, + 0.892, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.247, + 0.894, + 0.29 + ], + "angle": 0, + "content": "Figure 5. (Left) TRIFUSER is a latent diffusion model conditioned on patch-wise tokens of the input images \\((A, A', B)\\) to generate the analogically edited pattern \\(B'\\). (Right) To achieve high-quality edits, we enrich these tokens by fusing multi-level features from multiple encoders, followed by a 3D positional encoding: 2D to specify spatial locations and 1D to specify the token's source \\((A, A'\\) or \\(B\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.468, + 0.395 + ], + "angle": 0, + "content": "formed programs \\( z_{A'} \\) and \\( z_{B'} \\). By using identical transformations across domains, we ensure a consistent \"edit relation\" across the quartet, satisfying Equation 1 by construction. In Figure 4, we illustrate examples of synthetic analogical quartets generated using this method, demonstrating consistent transformations between \\( (A, A') \\) and \\( (B, B') \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.468, + 0.501 + ], + "angle": 0, + "content": "Edit Operators \\( E \\). We focus on edits targeting specific sub-parts of the program. Specifically, we consider three types of edits: insertion, removal, and replacement of subprograms. For example, an edit operator might replace the sub-program responsible for splitting the canvas, while other edits may insert or remove tiles within the pattern. Please refer to the supplementary for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.348, + 0.53 + ], + "angle": 0, + "content": "3.3. Learning an Analogical Editor" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Our goal is to train a model on the synthetic data that is capable of performing analogical edits on real, in-the-wild patterns. Specifically, we aim to generate the target pattern \\( B' \\) from an input triplet \\( (A, A', B) \\). This approach allows users to demonstrate desired edits with a simple pattern pairs \\( (A, A') \\), which the model then applies to a complex pattern \\( B \\) to produce \\( B' \\). Given the success of Latent Diffusion Models (LDMs) in various generative modeling tasks [43], we chose to adapt an LDM for our task as well. We propose TRIFUSER, a latent diffusion model (LDM) for analogical editing (Figure 5). We provide a brief overview of LDMs to provide context before detailing TRIFUSER's modifications for analogical editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Preliminaries: Denoising Diffusion Probabilistic Models (DDPMs) [17] transform random noise into structured data via reverse diffusion steps guided with a conditioning embedding \\( c(y) \\) (often derived from text). Latent Diffusion Models (LDMs) extend DDPMs by mapping data to a lower-dimensional latent space via an encoder. During training, a UNet model [44] learns to remove noise introduced into the latents. During inference, a latent sampled from a normal distribution is iteratively denoised by the model to yield a clean latent. Finally, the clean latent is decoded to generate the output image. Please refer to [55] for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.892, + 0.365 + ], + "angle": 0, + "content": "a more thorough overview. For analogical editing we adapt an Image Variation (IM) model [54], which uses patch-wise image tokens extracted using a text-image encoder [39] as the conditioning embedding \\( c(y) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.47 + ], + "angle": 0, + "content": "The simplest adaptation of an IM model to our task is to generate \\( B' \\) conditioned on image tokens from all three input images, concatenated as \\( C = c(A) \\| c(A') \\| c(B) \\), where \\( \\| \\cdot \\) denotes token-wise concatenation. This approach, however, suffers from three drawbacks: Token Entanglement, Semantic Bias, and Detail Erosion. We discuss each of these issues briefly, along with our solutions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.562 + ], + "angle": 0, + "content": "Detail Erosion: Despite using patch-wise tokens, the extracted features lack the fine-grained information needed to retain key aspects of \\( B \\) in the generated pattern \\( B' \\). Consequently, the model often struggles to preserve elements like tile textures. To address this problem, we combine features from both the first and last layers of the feature encoder:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.575, + 0.892, + 0.591 + ], + "angle": 0, + "content": "\\[\nC _ {h l} (P) = \\operatorname {L i n e a r} \\left(L N \\left(c _ {\\text {h i g h}} (P)\\right) \\cdot L N \\left(c _ {\\text {l o w}} (P)\\right), \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.891, + 0.648 + ], + "angle": 0, + "content": "where \\( LN \\) is layer normalization, \\( \\cdot \\) denotes channel-wise concatenation, \\( P \\) is an input pattern, and Linear is a linear projection layer that fuses low- and high-level features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.829 + ], + "angle": 0, + "content": "Semantic Bias: Image variation models typically use feature extractors such as CLIP [39], which are trained to align image embeddings with corresponding text embeddings. Such embeddings emphasize high-level semantics but lack spatial and fine-grained visual details. Combining these embeddings with features from text-free, self-supervised extractors, such as DiNO [2], has been shown to improve performance in various downstream tasks [20, 49]. For our task, a similar approach—combining features from both text-image \\((m_{1})\\) and self-supervised \\((m_{2})\\) feature extractors—significantly enhances generation quality. The extracted features are fused as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.842, + 0.891, + 0.86 + ], + "angle": 0, + "content": "\\[\nC _ {m i x} (P) = \\operatorname {M i x e r} \\left(C _ {h l} ^ {m _ {1}} (P) \\cdot C _ {h l} ^ {m _ {2}} (P)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.891, + 0.9 + ], + "angle": 0, + "content": "where \\( \\text{Mixer} \\) is a two-layer MLP that integrates features from the two extractors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28719" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Token Entanglement: To successfully perform an analogical edit, for each patch-level feature token, the model must be able to identify to which source image \\((A, A', \\text{or } B)\\) that patch belongs as well as the 2D position of the patch within that image. Without these distinctions, the model often fails to identify the pattern to edit (i.e., \\(B\\)) and to recognize the desired edit from \\((A, A')\\). To address this problem, we introduce 3D positional encodings: two dimensions for spatial location within each pattern and one dimension for the source image. These encodings are applied to the extracted embeddings, yielding:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.271, + 0.469, + 0.289 + ], + "angle": 0, + "content": "\\[\nC ^ {\\Omega} = C _ {P E} (A) \\parallel C _ {P E} \\left(A ^ {\\prime}\\right) \\parallel C _ {P E} (B), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.291, + 0.469, + 0.308 + ], + "angle": 0, + "content": "\\[\nC _ {P E} (P) ^ {x y} = C _ {m i x} (P) ^ {x y} + P E \\left(t _ {P}, x, y\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.47, + 0.366 + ], + "angle": 0, + "content": "where \\( t_P \\) is a one-hot vector encoding which input image a token comes from and \\( PE(t_P, x, y) \\) positionally encodes both spatial and source information for each token." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.519 + ], + "angle": 0, + "content": "As we demonstrate in Section 4.4, conditioning on \\( C^\\Omega \\) instead of \\( C \\) significantly enhances the quality of patterns generated by TRIFUSER. Our adapted architecture, shown in Figure 5, integrates the modifications described above to effectively address the described drawbacks. Note that this architecture is designed for general analogical editing, not just pattern editing. To enhance generalizability to real-world patterns, we initialize TRIFUSER with an existing pretrained IM model [54], and fine-tune only the denoising UNet and the projection layers in our feature extractor." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.536, + 0.203, + 0.552 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.744 + ], + "angle": 0, + "content": "In this section, we evaluate our approach along three directions: (1) the effectiveness of TRIFUSER at performing analogical edits on complex, real-world patterns, emphasizing how our synthetic data enables editing of in-the-wild pattern images; (2) the ability of TRIFUSER to support programmatic, structure-preserving edits without explicit program inference; and (3) the impact of architectural modifications introduced in TRIFUSER on the quality of generated patterns. We conduct a human perceptual study, quantitative assessments, and qualitative comparisons to demonstrate our system's ability to perform high-quality analogical edits across a range of pattern types." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.262, + 0.772 + ], + "angle": 0, + "content": "4.1. Experiment Design" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Datasets: We generate a large synthetic dataset of analogical quartets, i.e., pairs of analogical patterns \\((A, A', B, B')\\), using the SPLITWEAVE program samplers introduced in Section 3.2. This synthetic dataset contains approximately 1 million samples covering two pattern styles, namely Split Filling Patterns (SFP) and Motif Tiling Patterns (MTP) (cf. Section 3.1). For MTP patterns, we synthesize \\(100k\\) distinct tiles using the LayerDiffuse [58] model, guided by text" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "prompts derived from WordNet [35] noun synsets. Additionally, we construct a synthetic test set with 1000 analogical quartets to evaluate model performance on unseen synthetic data. Further details on dataset construction are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.893, + 0.349 + ], + "angle": 0, + "content": "To assess TRIFUSER on real-world patterns, we curate a test dataset of 50 patterns created by professional artists and sourced from Adobe Stock. This dataset spans seven distinct sub-domains of 2D patterns, representing a range of pattern styles. These styles include MTP and SFP patterns as well as previously unseen pattern styles such as Memphis-style, geometric, and digital textile patterns. Each pattern is annotated with a desired edit, and we use SPLITWEAVE to generate a pair of simpler patterns \\((A, A')\\) demonstrating this edit. This test set provides a challenging benchmark to evaluate TRIFUSER's generalization to diverse, real-world editing tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.894, + 0.502 + ], + "angle": 0, + "content": "Training details: We fine-tune a pre-trained diffusion model using our synthetic dataset of analogical quartets (cf. Section 3.3). We initialize our model with Versatile-Diffusion's Image Variation model [54]. We use SigLIP [57] as our text-image feature encoder and DiNOv2 [37] for self-supervised features. We fine-tune the model on 8 A100 GPUs using a batch size of 224 for \\(\\sim 65\\) epochs over 7 days. During inference, \\(B^{\\prime}\\) is inferred with typical diffusion parameter settings such as a classifier-free guidance weight of 7.5 and 50 denoising steps." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.513, + 0.757, + 0.53 + ], + "angle": 0, + "content": "4.2. Analogical Editing Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.582 + ], + "angle": 0, + "content": "To evaluate the analogical editing capability of TRIFUSER, we compare it to three baseline methods, each representing a leading approach for analogical image editing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.839 + ], + "angle": 0, + "content": "First, we consider training-free editors and latent arithmetic editors. Training-free editors repurpose pre-trained diffusion models to perform analogical edits without additional training [13, 51], leveraging the rich representations learned by diffusion models for editing. In this category, we compare against Analogist [13], the current state-of-the-art method. Latent arithmetic editors, on the other hand, rely on transformations in a learned latent space to infer analogical modifications [40, 48]. Note that these approaches only require samples from the target domain, not analogical training pairs. We implement a baseline for this method by fine-tuning a naive Image Variation model [54] on our synthetic dataset to learn a generative latent embedding space of patterns. At inference, analogical edits are generated using latent arithmetic: given patterns \\( A \\), \\( A' \\), and \\( B \\), we condition the generation of \\( B' \\) on \\( E(B) + E(A') - E(A) \\). We refer to this baseline as LatentMod." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Finally, we consider analogy-conditioned generative editors, where models are explicitly trained on analogical data to learn analogical transformations [47]. This category includes our proposed TRIFUSER as well. Image Brush," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28720" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.09, + 0.468, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.329, + 0.47, + 0.371 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison between patterns generated by our model, TRIFUSER, and the baselines. TRIFUSER generates higher quality patterns with greater fidelity to the input analogy." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.374, + 0.432, + 0.452 + ], + "angle": 0, + "content": "
Preference Rate
TRIFUSER vs. Analogist89.24%
TRIFUSER vs. LatentMod80.59%
TRIFUSER vs. Inpainter72.05%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.504 + ], + "angle": 0, + "content": "Table 1. Results of a two-alternative forced-choice perceptual study comparing our model (TRIFUSER) against three baselines. Ours is preferred in the overwhelming majority of judgments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.469, + 0.64 + ], + "angle": 0, + "content": "the state-of-the-art method, fine-tunes a diffusion inpainting model for analogical editing with multi-modal conditioning. Since code for Image Brush is unavailable, we implement a similar baseline by fine-tuning a Stable Diffusion inpainting model. This model, which we term Inpainter, performs analogical editing by inpainting the lower-left quadrant of a 2x2 analogy grid containing \\((A, A', B)\\) and conditioned on a fixed text template." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.334, + 0.667 + ], + "angle": 0, + "content": "4.3. Editing Real-World Patterns" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.719 + ], + "angle": 0, + "content": "To evaluate TRIFUSER's real-world analogical editing capabilities, we conducted a human preference study on the curated test set of Adobe Stock patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.87 + ], + "angle": 0, + "content": "We performed a two-alternative forced-choice perceptual study comparing TRIFUSER with baseline methods on all 50 entries in the test set. Each method generates \\( k = 9 \\) outputs for each input tuple, and we select the best one based on visual inspection. Participants were shown edited patterns generated by two different methods along with the input patterns \\( (A, A', B) \\) and instructed to select the edit that best preserved the analogical relationship and exhibited higher image quality. We recruited 42 participants for the study, resulting in a total of 1720 total judgments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Table 1 presents the results, showing that TRIFUSER was preferred over all three baselines. Due to the domain gap" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.889, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.343, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Figure 7. TRIFUSER effectively edits patterns from novel pattern styles not present in the training dataset. TRIFUSER shows a noteworthy ability to generalize beyond its training distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.563 + ], + "angle": 0, + "content": "between the training data of the underlying model [43] and pattern images, Analogist fails to interpret and edit pattern images. Meanwhile, LatentMod fails to perform reasonable edits as the embedding space lacks the low-level details necessary for programmatic edits. While these baselines perform adequately on stylistic edits, they are unsuitable for programmatic editing. When compared to Inpainter, TRI-FUSER was favored in \\(72.2\\%\\) of comparisons. Both methods benefit from training on analogical quartets, yet Inpainter sacrifices pattern quality as it generates the edited pattern in only a quarter of the full canvas resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.67 + ], + "angle": 0, + "content": "Figure 6 shows examples of pattern edits generated by TRIFUSER and the baselines, with our model consistently delivering superior results. In Figure 7, we show examples of TRIFUSER's edits on out-of-distribution pattern styles not present in the training set. These results suggest that our synthetic training data enables manipulation of real-world patterns, even extending to certain untrained pattern styles." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.74, + 0.696 + ], + "angle": 0, + "content": "4.4. Editing Synthetic Patterns" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Next, we evaluate TRIFUSER's ability to perform programmatic edits on the synthetic validation set, which contains ground truth patterns \\( B' \\). Ideally, this would involve verifying that the underlying program \\( z_{\\hat{B}'} \\) of the generated pattern reflects the same transformation from \\( z_B \\) as that between \\( z_A \\) and \\( z_A' \\). However, this would require visual program inference on \\( \\hat{B}' \\), which is infeasible. Instead, we approximate this criterion by comparing the program outputs \\( \\hat{B}' \\) and \\( B' \\) to see if the visual results align with the intended transformation. To quantify this alignment, we use perceptual metrics—DSim [8], DIST [4] and LPIPS [59]—along with SSIM to capture pixel-level structural similarity." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Note that analogies can have multiple valid interpreta" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28721" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.465, + 0.182 + ], + "angle": 0, + "content": "
DSIM (↓)DISTS (↓)LPIPS (↓)SSIM (↑)
Analogist0.4960.4320.6970.494
LatentMod0.2420.3200.6130.502
Inpainter0.0920.2560.3710.713
TRIFUSER0.0740.1840.3040.704
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.19, + 0.47, + 0.232 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation on the synthetic validation set shows that TRIFUSER generates patterns with higher perceptual similarity to the ground truth than the baselines." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.235, + 0.465, + 0.348 + ], + "angle": 0, + "content": "
DSIM (↓)DISTS (↓)LPIPS (↓)SSIM (↑)
TRIFUSER0.0740.1840.3040.704
- Pos. Enc.0.1470.2390.3830.659
- Lower0.0870.1960.3350.652
- Mix0.0980.2100.3450.682
Base [54]0.5850.4600.8150.435
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.356, + 0.47, + 0.398 + ], + "angle": 0, + "content": "Table 3. Subtractive ablation study on TRIFUSER shows that removing any component (see Section 3.3) degrades performance, and that removing all components (Base) results in a sharp decline." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.415, + 0.469, + 0.506 + ], + "angle": 0, + "content": "tions, and even a single interpretation may yield several visually-related variations. To account for this multiplicity, we generate \\( k = 5 \\) output patterns for each input set \\( (A,A^{\\prime},B) \\) and select the one that maximizes each metric. In other words, we evaluate whether at least one generated output aligns with the intended target." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Table 2 shows the results of this experiment. First, we note that TRIFUSER outperforms others across all perceptual metrics. These metrics capture different aspects of perceptual similarity and superior performance across all of them suggests a comprehensive improvement. Second, we observe that the analogy-conditioned generative editors (Inpainter & TRIFUSER) surpass both the training-free and latent modification editors. Interestingly, Inpainter achieves slightly higher SSIM scores than TRIFUSER, suggesting that integrating the two approaches could be beneficial." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.262, + 0.681 + ], + "angle": 0, + "content": "4.5. TRIFUSER Ablation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To evaluate the impact of each model component introduced in Section 3.3, we conduct a subtractive analysis on the synthetic validation set. For this ablation study, we remove each component one at a time and measure the resulting performance, as reported in Table 3. The results demonstrate that removing any component leads to a performance drop, with the removal of 3D positional encoding leading to severe degradation. This is understandable: without 3D positional encoding, the network often fails to accurately identify which pattern to edit. For comparison, we also include results from the original Image Variation model [54] trained without any modifications (Base). As expected, this model performs poorly, underscoring the importance of our modifications in achieving high-quality analogical edits." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.633, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.092, + 0.757, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.091, + 0.888, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.201, + 0.892, + 0.231 + ], + "angle": 0, + "content": "Figure 8. Our model helps users mix elements of different real-world patterns together, accelerating design exploration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.241, + 0.623, + 0.259 + ], + "angle": 0, + "content": "5. Application" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.268, + 0.892, + 0.313 + ], + "angle": 0, + "content": "The ability to edit patterns without requiring program inference unlocks new creative possibilities. We demonstrate two practical applications of analogical pattern editing:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.314, + 0.893, + 0.434 + ], + "angle": 0, + "content": "Pattern Mixing: Figure 8 shows example of using our method to mix elements of two real-world patterns \\( X \\) and \\( Y \\), allowing the user to create unique, hybrid designs. The Mix operator is implemented by using a synthetic pair \\( (A, A') \\) to create a variant \\( X' \\) of \\( X \\) and then using the pair \\( (X, X') \\) to specify an edit to \\( Y \\): \\( \\text{Mix}(X, Y) = f(X, X', Y) \\), where \\( X' = f(A, A', X) \\). See the supplementary material for more details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.435, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Animation Transfer: TRIFUSER can also be used to create animated sequences of edited patterns. By leveraging parametric SPLITWEAVE programs, users can generate animations for simple patterns and then apply these animations to complex patterns with no additional effort. See the video in the supplementary material for examples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.619, + 0.558 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We introduced a novel approach for programmatic editing of visual patterns without inferring the underlying program. By using analogies to express desired edits and a learned conditional generative model to execute them, our method provides an intuitive solution for pattern manipulation. A key component of our approach is SPLITWEAVE, a domain-specific language for generating diverse, structured pattern data. Paired with our procedure for sampling analogical quartets, SPLITWEAVE enables the creation of a large, high-quality dataset for training. We also presented TRIFUSER, a Latent Diffusion Model (LDM) designed to overcome critical issues that emerge when LDMs are naively deployed for analogical pattern editing, enabling high-fidelity edits that capture user intentions. Our experiments demonstrate that TRIFUSER successfully edits real-world patterns and surpasses baseline methods, while also generalizing to novel pattern styles beyond its training distribution. We believe that our DSL, dataset, and model will help drive further research on in-the-wild pattern image editing. Looking forward, we aim to extend this analogical editing framework to other domains such as semi-parametric 3D modeling while continuing to improve synthetic data generation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28722" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.157 + ], + "angle": 0, + "content": "[1] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei A. Efros. Visual prompting via image inpainting. arXiv preprint arXiv:2209.00647, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.226 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.228, + 0.468, + 0.311 + ], + "angle": 0, + "content": "[3] Ta-Ying Cheng, Matheus Gadelha, Thibault Groueix, Matthew Fisher, Radomir Mech, Andrew Markham, and Niki Trigoni. Learning continuous 3d words for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6753-6762, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.313, + 0.468, + 0.352 + ], + "angle": 0, + "content": "[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P. Simoncelli. Image quality assessment: Unifying structure and texture similarity. CoRR, abs/2004.07728, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.354, + 0.468, + 0.409 + ], + "angle": 0, + "content": "[5] Kevin Ellis, Daniel Ritchie, Armando Solar-Lezama, and Josh Tenenbaum. Learning to infer graphics programs from hand-drawn images. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.411, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[6] Thomas G. Evans. A heuristic program to solve geometric-analogy problems. In Proceedings of the April 21-23, 1964, Spring Joint Computer Conference, page 327-338, New York, NY, USA, 1964. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.481, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[7] Brian Falkenhainer, Kenneth D. Forbus, and Dedre Gentner. The structure-mapping engine. In Proceedings of the Fifth AAAI National Conference on Artificial Intelligence, page 272-277. AAAI Press, 1986. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.537, + 0.468, + 0.605 + ], + "angle": 0, + "content": "[8] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.607, + 0.468, + 0.661 + ], + "angle": 0, + "content": "[9] Aditya Ganeshan, R. Kenny Jones, and Daniel Ritchie. Improving unsupervised visual program inference with code rewriting families. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[10] Aditya Ganeshan, Ryan Y. Huang, Xianghao Xu, R. Kenny Jones, and Daniel Ritchie. Parsel: Parameterized shape editing with language, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[11] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Advances in Neural Information Processing Systems, pages 5885–5897. Curran Associates, Inc., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.468, + 0.789 + ], + "angle": 0, + "content": "[12] Dedre Gentner. Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2):155-170, 1983. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[13] Zheng GU, Shiyuan Yang, Jing Liao, Jing Huo, and Yang Gao. Analogist: Out-of-the-box visual in-context learning with image diffusion model. ACM Transactions on Graphics (TOG), 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[14] P. Guehl, R. Allegre, J.-M. Dischler, B. Benes, and E. Galin. Semi-procedural textures using point process texture basis functions. Computer Graphics Forum, 39(4):159-171, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] Julia Guerrero-Viu, Milos Hasan, Arthur Roullier, Midhun Harikumar, Yiwei Hu, Paul Guerrero, Diego Gutierrez, Bellen Masia, and Valentin Deschaintre. Texsliders: Diffusion-based texture editing in clip space. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[16] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image Analogies. Association for Computing Machinery, New York, NY, USA, 1 edition, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[18] Douglas Hofstadter and Melanie Mitchell. The Copycat project: a model of mental fluidity and analogy-making, page 205-267. Basic Books, Inc., USA, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[19] Yiwei Hu, Paul Guerrero, Milos Hasan, Holly Rushmeier, and Valentin Deschaintre. Node graph optimization using differentiable proxies. In ACM SIGGRAPH Conference Proceedings, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[20] Dongsheng Jiang, Yuchen Liu, Songlin Liu, XIAOPENG ZHANG, Jin Li, Hongkai Xiong, and Qi Tian. From CLIP to DINO: Visual encoders shout in multi-modal large language models, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[21] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapeassembly: Learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics (TOG), Siggraph Asia 2020, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[22] R. Kenny Jones, David Charatan, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapemod: Macro operation discovery for 3d shape programs. ACM Transactions on Graphics (TOG), Siggraph 2021, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[23] R. Kenny Jones, Homer Walke, and Daniel Ritchie. Plad: Learning to infer shape programs with pseudo-labels and approximate distributions. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[24] R. Kenny Jones, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapecoder: Discovering abstractions for visual programs from unstructured primitives. ACM Transactions on Graphics (TOG), Siggraph 2023, 42(4), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[25] R. Kenny Jones, Renhao Zhang, Aditya Ganeshan, and Daniel Ritchie. Learning to edit visual programs with self-supervision. In Advances in Neural Information Processing Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[26] Milin Kodnongbua, Benjamin Jones, Maaz Bin Safeer Ahmad, Vladimir Kim, and Adriana Schulz. Reparamcad: Zero-shot cad re-parameterization for interactive manipulation. In SIGGRAPH Asia 2023 Conference Papers, New York, NY, USA, 2023. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[27] Beichen Li, Liang Shi, and Wojciech Matusik. End-to-end procedural material capture with proxy-free mixed-integer optimization. ACM Trans. Graph., 42(4), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J. Mitra. Sketch2cad: Sequential cad modeling by sketching in" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28723" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "context. ACM Trans. Graph. (Proceedings of SIGGRAPH Asia 2020), 39(6):164:1-164:14, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[29] Wen-Ding Li, Keya Hu, Carter Larsen, Yuqing Wu, Simon Alford, Caleb Woo, Spencer M. Dunn, Hao Tang, Michelangelo Naim, Dat Nguyen, Wei-Long Zheng, Zenna Tavares, Yewen Pu, and Kevin Ellis. Combining induction and transduction for abstract reasoning, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[30] Hugo Loi, Thomas Hurtut, Romain Vergne, and Joelle Thollot. Programmable 2d arrangements for element texture design. ACM Trans. Graph., 36(4), 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[31] Arman Maesumi, Dylan Hu, Krishi Saripalli, Vladimir Kim, Matthew Fisher, Soeren Pink, and Daniel Ritchie. One noise to rule them all: Learning a unified model of spatially-varying noise patterns. ACM Trans. Graph., 43(4), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.362 + ], + "angle": 0, + "content": "[32] Jiayuan Mao, Xiuming Zhang, Yikai Li, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Program-Guided Image Manipulators. In International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[33] L. McCarthy, C. Reas, and B. Fry. Getting Started with P5.js: Making Interactive Graphics in JavaScript and Processing. Maker Media, Incorporated, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[34] Zichong Meng, Changdi Yang, Jun Liu, Hao Tang, Pu Zhao, and Yanzhi Wang. Instructgie: Towards generalizable image editing. arXiv preprint arXiv:2403.05018, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[35] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.354, + 0.52 + ], + "angle": 0, + "content": "[36] OpenAI. Gpt-4 technical report, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.522, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[37] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[38] Ofek Pearl, Itai Lang, Yuhua Hu, Raymond A. Yeh, and Rana Hanocka. Geocode: Interpretable shape programs. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[40] Scott Reed, Yi Zhang, Yuting Zhang, and Honglak Lee. Deep visual analogy-making. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, page 1252-1260, Cambridge, MA, USA, 2015. MIT Press. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, Haiyong Jiang, Zhongang Cai, Junzhe Zhang, Liang Pan, Mingyuan Zhang, Haiyu Zhao, and Shuai Yi. Csg-stump: A learning friendly csg-like representation for interpretable" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "shape parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[42] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, and Junzhe Zhang. Extrudenet: Unsupervised inverse sketch-and-extrude fornbsp; shape parsing. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, page 482-498, Berlin, Heidelberg, 2022. Springer-Verlag. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.894, + 0.273 + ], + "angle": 0, + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.893, + 0.343 + ], + "angle": 0, + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.893, + 0.384 + ], + "angle": 0, + "content": "[45] Christian Santoni and Fabio Pellacini. gtangle: a grammar for the procedural generation of tangle patterns. ACM Trans. Graph., 35(6), 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[46] Liang Shi, Beichen Li, Miloš Hasan, Kalyan Sunkavalli, Tamy Boubekeur, Radomir Mech, and Wojciech Matusik. Match: differentiable material graphs for procedural material capture. ACM Trans. Graph., 39(6), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[47] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: learning visual in-context instructions for exemplar-based image manipulation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.538, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[48] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[49] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[50] Trieu Trinh, Yuhuai Tony Wu, Quoc Le, He He, and Thang Luong. Solving olympiad geometry without human demonstrations. Nature, 625:476-482, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[51] Adéla Šubrtová, Michal Lukáč, Jan Čech, David Futschik, Eli Shechtman, and Daniel Sykora. Diffusion image analogies. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[52] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6772-6782, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[53] Xianghao Xu, Wenzhe Peng, Chin-Yi Cheng, Karl D. D. Willis, and Daniel Ritchie. Inferring cad modeling sequences using zone graphs. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[54] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7754-7765, 2023. 2, 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28724" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[56] Mehmet Ersin Yumer, Siddhartha Chaudhuri, Jessica K. Hodgins, and Levent Burak Kara. Semantic shape editing using deformation handles. ACM Trans. Graph., 34(4), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[57] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[58] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. ACM Trans. Graph., 43(4), 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.362 + ], + "angle": 0, + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28725" + } + ] +] \ No newline at end of file diff --git a/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/86d3746e-1656-4dc5-ab96-b8f696f1d699_origin.pdf b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/86d3746e-1656-4dc5-ab96-b8f696f1d699_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..081844bc41099512cd997830b36896cd69c9883d --- /dev/null +++ b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/86d3746e-1656-4dc5-ab96-b8f696f1d699_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a3b6a44c499f86362205f6692b1608a60c0c169b4b67caac34addbfeda82f12 +size 8685984 diff --git a/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/full.md b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d31486cd0de855dc86cf6827e5c6578cf2d15e03 --- /dev/null +++ b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/full.md @@ -0,0 +1,311 @@ +# Pattern Analogies: Learning to Perform Programmatic Image Edits by Analogy + +Aditya Ganeshan* Brown University + +Thibault Groueix +Adobe Research + +Paul Guerrero +Adobe Research + +Radomír Měch +Adobe Research + +Matthew Fisher Adobe Research + +Daniel Ritchie Brown University + +![](images/fd6ae179c58cf2692114081c8e54597143d51ca15a2478607cc048ccce89449f.jpg) +Figure 1. Our system performs programmatic edits on pattern images without inferring their underlying programs. (Left) Desired edits, expressed with a pair of patterns $(A, A')$ , are executed on a target pattern $B$ by a generative model to produce $B'$ . (Right) Parametric changes $A \to A'$ enabled by our domain-specific pattern language induce corresponding changes to the more complex pattern $B$ . + +# Abstract + +Pattern images are everywhere in the digital and physical worlds, and tools to edit them are valuable. But editing pattern images is tricky: desired edits are often programmatic: structure-aware edits that alter the underlying program which generates the pattern. One could attempt to infer this underlying program, but current methods for doing so struggle with complex images and produce unorganized programs that make editing tedious. In this work, we introduce a novel approach to perform programmatic edits on pattern images. By using a pattern analogy—a pair of simple patterns to demonstrate the intended edit—and a learning-based generative model to execute these edits, our method allows users to intuitively edit patterns. To enable this paradigm, we introduce SPLITWEAVE, a domain-specific language that, combined with a framework for sampling synthetic pattern analogies, enables the creation of a large, high-quality synthetic training dataset. We also present TRIFUSER, a Latent Diffusion Model (LDM) designed to overcome critical issues that arise when naively + +deploying LDMs to this task. Extensive experiments on real-world, artist-sourced patterns reveals that our method faithfully performs the demonstrated edit while also generalizing to related pattern styles beyond its training distribution. + +# 1. Introduction + +Visual pattern designs enhance digital media such as presentations, website themes, and user interfaces, and they are woven into the physical world through textiles, wallpapers, and product designs like hardware covers. Given the ubiquity of patterns, methods for editing them are essential: designers should be able to quickly experiment with variations, customize designs to meet specific needs, and adapt existing patterns to align with evolving trends. + +Editing pattern images is not straightforward, as patterns are inherently structured, defined by rules that govern their layout and composition: tiling patterns adhere to principles of alignment and repetition (see Figure 1: top left), while retro-style designs rely on spatial divisions and fills (see Figure 1: bottom center). The edits that designers desire + +often aim to adjust these underlying organizational rules rather than make superficial, pixel-level changes. We refer to such edits as programmatic edits, requiring manipulation of the underlying program that defines a pattern's structure. + +One strategy for enabling such programmatic edits is visual program inference (VPI) [5, 32, 46], where a program that replicates an image is automatically inferred, allowing users to modify the image by adjusting program parameters. However, applying VPI to patterns presents two obstacles. First, VPI attempts to infer a program that fully replicates a pattern, which can be challenging as patterns are often semiparametric, blending rule-based logic with non-parametric components. For instance, the layout of elements in a tiling pattern may be rule-based, but the elements themselves may not be. Second, editing with an inferred program can be cumbersome, as they are often poorly-structured, with many unlabeled parameters, making them difficult to interpret. Consequently, VPI not only solves a more complex problem than necessary but also makes editing more challenging. + +Can we perform programmatic edits without inferring the underlying program? Doing so requires the ability to express and execute the edit—both without direct access to the program's parameters. To express a programmatic edit, it's crucial to specify both which underlying parameter(s) to change and how to modify them. We draw inspiration from how humans communicate transformations: through analogies. By providing a pair of simple example patterns $(A, A')$ that illustrate the desired change, users can intuitively convey both aspects of the edit. To execute these edits, we employ a learning-based conditional generative model. Given a pair of simple patterns $(A, A')$ and a complex target pattern $B$ , our system generates $B'$ , an edited version of $B$ which performs the transformation demonstrated between $A$ and $A'$ while preserving $B$ 's other structural features. Crucially, $A$ does not need to replicate or even be similar to $B$ —it only needs to demonstrate which property to edit and how. Thus, specifying $A$ is a much easier task than solving VPI. While prior works [1, 47, 51] have applied analogical editing to image manipulation, they focus primarily on appearance modifications. In contrast, our approach is the first to use analogies for programmatic, structure-aware edits. Figure 1 (left) shows examples of analogical editing on complex, real-world patterns. + +To make our approach possible, we introduce SPLITWEAVE: a domain-specific language (DSL) for crafting visual patterns. SPLITWEAVE serves two purposes in our method. First, it enables parametric definition of input pairs $(A,A^{\prime})$ , allowing users to guide transformations in $(B,B^{\prime})$ as if the underlying program for $B$ were accessible. In Figure 1 (right), modifying the SPLITWEAVE program for $A^\prime$ produces corresponding changes in $B^{\prime}$ . Second, SPLITWEAVE supports the creation of large-scale synthetic training data. We develop program samplers + +that generate high-quality patterns in two common styles: tiling-based designs with repeating elements and color field patterns characterized by splitting the canvas into intricate colored regions. Training a model for analogical editing requires a dataset of quartets $(A, A', B, B')$ . By applying identical programmatic edits to the SPLITWEAVE programs for both $A$ and $B$ to produce $A'$ and $B'$ , we ensure that the transformation from $A$ to $A'$ mirrors that from $B$ to $B'$ . This approach allows us to generate a diverse dataset of analogical quartets. Models trained on this dataset generalize effectively to real-world patterns within these styles and even extend to related styles. + +We use this synthetic dataset to train a novel diffusion-based conditional generative model for executing analogical edits. Our model directly generates edited patterns $B'$ by conditioning on visual features extracted from input patterns $(A, A', B)$ . Existing image-conditioned diffusion models [43, 54] prove ineffective, as they fail to interpret the input analogies accurately and neglect fine details. To address these issues, we incorporate architectural enhancements that enable our model, TRIFUSER, to effectively perform analogical edits. With these improvements, TRIFUSER surpasses prior architectures for analogical editing when applied to pattern images. + +To evaluate our method, we curated a test set of 50 patterns from Adobe Stock spanning 7 distinct styles. A perceptual study on this dataset reveals that participants consistently prefer edits by TRIFUSER over those produced by recent training-free and training-based analogical editing methods. Despite training only on two of these styles, our model successfully generalizes to the remaining five, which were unseen during training. On a synthetic validation set with ground-truth analogical edits, our model achieves higher structural and perceptual similarity to the ground truth than prior methods. Finally, we showcase two compelling applications of our approach: mixing attributes of different patterns and transferring pattern animations. + +In summary, our contributions are as follows: + +1. A novel framework for performing programmatic edits to pattern images without requiring program inference, leveraging analogies to specify and apply edits. +2. SPLITWEAVE, a DSL for crafting a diverse range of visual patterns, designed to support both parametric control and synthetic dataset generation. +3. A procedure for generating synthetic analogical quartets, enabling editing of in-the-wild patterns. +4. TRIFUSER, a diffusion-based conditional generative model that achieves high fidelity in analogical edits, surpassing prior techniques in both analogical fidelity and generation quality. + +# 2. Related Work + +We review three key areas: (1) Visual Program Inference (VPI) for programmatic editing of structured visual data and its limitations, (2) DSLs and synthetic data generation, specifically for visual patterns, and (3) analogical reasoning in computing, particularly for editing images. + +Visual Program Inference for Editing: Visual Program Inference (VPI) enables programmatic edits of visual data by inferring executable programs from visual inputs. Prior works have achieved promising results in inferring material graphs [19, 27, 31, 46] and CAD programs for 2D [11, 28] and 3D [42, 53] inputs, using large annotated datasets [46, 52], differentiable program approximations [19, 41], or bootstrapped learning [9, 23, 25]. VPI is challenging to adapt to pattern editing due to the scarcity of high-quality annotated pattern data and the non-differentiability of most pattern programs. Also, VPI approaches often yield complex programs that are difficult to edit and interpret, making them impractical for editing. To address these challenges, recent work has aimed to simplify programmatic editing by inferring edit-specific controls [3, 10, 15] or a limited set of semantically meaningful parameters [22, 24, 26, 56]. Our approach shares this goal of enabling accessible control but extends it further: we transfer control from simple parametric objects to complex in-the-wild images via analogy, bypassing the need for VPI. + +DSL and Synthetic data Domain-Specific Languages (DSLs) enable concise descriptions of structured objects, facilitating their creation. Prior works have developed DSLs for Zentangle patterns [45], material graphs [46], semiparametric textures [14], and 3D models [21, 38]. Our DSL focuses on visual patterns constructed through partitioning and merging of canvas fragments. Closest to our work is ETD [30], which also uses canvas partitioning and merging operators, though it is limited to stationary patterns. + +Analogical Reasoning Analogical reasoning is a foundational AI task: early work includes Evans' ANALOGY program [6], CopyCat [18], and Structure-Mapping Engine [7]. In visual computing, Image Analogies [16] pioneered the concept of analogy-driven editing. Recently, diffusion models have been adapted for analogical editing. DIA [51] introduced a training-free approach to analogical editing using pretrained diffusion models. Analogist [13] offers a complementary method, leveraging inpainting models alongside multimodal reasoning from large language models [36]. These training-free approaches are limited to images within the diffusion model's training domain, limiting their applicability to patterns. Other methods attempt to learn analogical editors by finetuning diffusion models on analogical pairs [1, 34, 47]. However, the focus of all these works remains largely on stylistic, appearance edits, often failing to perform programmatic edits. This limitation arises both from the models' architectures + +![](images/913e64525e76d1571c4d45cbccb7a748fef3f921fd26c16dcd740aea4e3d2b42.jpg) +Figure 2. Overview: To create high-quality visual patterns, we introduce a custom DSL called SPLITWEAVE. Pairs of SPLITWEAVE programs $(A,B)$ are then jointly edited to create analogical quartets. This synthetic data is then used to train TRI-FUSER, a neural network for analogical pattern editing. + +![](images/48e6af05707f57f7bf544a13dbf34e0afb2c37d716440149fdcb58ff236fb24d.jpg) + +and from the lack of training pairs with programmatic edits. Our work addresses both these gaps, enabling structured, programmatic analogical edits for visual patterns. + +# 3. Method + +Our objective is to enable programmatic edits of 2D visual patterns without inferring their underlying programs. Instead, we propose an alternative that uses analogies to express desired edits and a conditional generative model to execute them. Formally, given two source patterns $A$ and $A'$ that demonstrate a desired edit, along with a target pattern $B$ , our goal is to generate an edited target pattern $B'$ that applies this edit to $B$ . This task is defined as learning a mapping $f(A, A', B) \to B'$ , where $A, A', B$ , and $B'$ are 2D RGB images ( $\in \mathbb{R}^{H \times W \times 3}$ ). To learn this mapping, we generate a large synthetic dataset of analogical pattern quartets ( $A, A', B, B'$ ). + +Figure 2 provides a schematic overview of our approach. First, in Section 3.1 we introduce SPLITWEAVE, a Domain-Specific Language (DSL) that enables the creation and manipulation of various kinds of patterns. Section 3.2 describes our approach for sampling analogical quartets in SPLITWEAVE to create the synthetic training data. Finally, in Section 3.3, we present TRIFUSER, a conditional generative model that learns to execute analogical edits. + +# 3.1. A Language for Visual Patterns + +To enable programmatic edits without program inference, our approach requires two core capabilities: (a) generating a large, high-quality synthetic dataset for training models to + +![](images/edcb66e9ed39bf509518c9da09cab9e3ab775863a271893b17d82084cccc2a62.jpg) +Figure 3. Custom program samplers for two pattern styles. Our samplers produce diverse and high-quality patterns, enabling generalization to real-world patterns. + +reliably execute analogical edits, and (b) the ability to create and parametrically control analogy inputs at test time to effectively express desired edits. Existing pattern generation tools are insufficient for these needs, as they are either limited to narrow pattern domains [45] or demand intense coding effort to produce diverse, high-quality patterns [30, 33]. To address these limitations, we introduce SPLITWEAVE, a DSL designed specifically to support analogical transformations in visual patterns. SPLITWEAVE combines abstractions for pattern synthesis with a node-based visual programming interface (see Supplementary), enabling efficient generation of high-quality synthetic patterns for training while allowing flexible, precise pattern manipulation to define analogy inputs at test time. + +SPLITWEAVE uses three types of operations for structured pattern creation: (1) Canvas Fragmentation, which allows structured divisions of the canvas, such as brick-like or voronoi splits; (2) Fragment ID-Aware Operations, enabling transformations that vary across fragments (e.g., scaling alternating rows or columns) to support spatial variability in non-stationary pattern designs; and (3) various SVG Operators for outlining, coloring, and compositing. Together, these operations enable efficient creation of patterns with complex structure and visual variety. Figure 2 (left) illustrates these capabilities in a SPLITWEAVE program for generating a tiling pattern design. + +Our goal is to generate high-quality synthetic patterns using SPLITWEAVE that enable trained editing models to generalize well to real-world patterns. Naive sampling from the DSL grammar often leads to overly complex or incoherent patterns, limiting their effectiveness in model training. Instead, we draw inspiration from recent advances in fields such as geometric problem solving [50] and abstract reasoning [29], where tailored data generators have proven to be essential for tackling complex tasks. Following a similar approach, we design custom program samplers for two versatile and widely-used pattern styles. The first, Motif + +![](images/eb3ddf8bc4238274e4781425b7305a145a88a8aff3e59d7976c5ce23b9672deb.jpg) +Figure 4. We create synthetic analogical quartets $(A, A', B, B')$ with consistent edits between $A$ and $B$ pairs, providing data for training an analogical editing models. + +Tiling Patterns (MTP), consists of compositions based on repeated Tile elements. These patterns exhibit controlled variations in tile properties across the canvas (e.g. orientation, color, and scale), creating visually cohesive yet richly diverse structures. The second, Split-Filling Patterns (SFP), are generated by dividing the canvas into ordered fragments, applying region-specific coloring and transformations based on fragment IDs. Both pattern styles are common in digital design and support a wide range of programmatic variations, making them particularly suited for analogical editing tasks. Example patterns generated by our program samplers are shown in Figure 3; additional implementation details are in the supplementary materials. + +# 3.2. Sampling Analogical Quartets + +With the ability to generate diverse synthetic patterns using SPLITWEAVE (Section 3.1), our goal is now to construct analogical pattern quartets $(A, A', B, B')$ . Each pattern image in a quartet is generated by a SPLITWEAVE program $z$ . These quartets serve as structured training data for editing models, allowing them to learn consistent transformations that can generalize across different pattern domains. + +Analogies in our framework are grounded in Structure Mapping Theory [12], which defines analogies as mappings of relational structure from a base to a target domain. We designate $(A, A')$ as the base and $(B, B')$ as the target, with the requirement that the relationship $R$ between program pairs $(z_A, z_{A'})$ and $(z_B, z_{B'})$ remains consistent: + +$$ +R \left(z _ {A}, z _ {A ^ {\prime}}\right) = R \left(z _ {B}, z _ {B ^ {\prime}}\right). \tag {1} +$$ + +Rather than focusing on visual similarity between the patterns $(A, A')$ themselves, this program-level analogy allows us to generate quartets with transformations that affect the underlying program, facilitating programmatic edits. + +To construct these analogical quartets, we use a program sampler along with a predefined set of editing operators $E$ . For each quartet, we begin by sampling an edit $e \in E$ , followed by sampling initial programs $z_{A}$ and $z_{B}$ that are compatible with $e$ . Applying $e$ to both $z_{A}$ and $z_{B}$ yields trans + +![](images/1e58a36e6a34b85b90780df24aa9f1019d13ffdbb202fceb99a55e70fdcabae2.jpg) +Figure 5. (Left) TRIFUSER is a latent diffusion model conditioned on patch-wise tokens of the input images $(A, A', B)$ to generate the analogically edited pattern $B'$ . (Right) To achieve high-quality edits, we enrich these tokens by fusing multi-level features from multiple encoders, followed by a 3D positional encoding: 2D to specify spatial locations and 1D to specify the token's source $(A, A'$ or $B$ ). + +![](images/9523dd832b579fce58bd2582e931c69429c263448d226604391f7aa1cf1311bf.jpg) + +formed programs $z_{A'}$ and $z_{B'}$ . By using identical transformations across domains, we ensure a consistent "edit relation" across the quartet, satisfying Equation 1 by construction. In Figure 4, we illustrate examples of synthetic analogical quartets generated using this method, demonstrating consistent transformations between $(A, A')$ and $(B, B')$ . + +Edit Operators $E$ . We focus on edits targeting specific sub-parts of the program. Specifically, we consider three types of edits: insertion, removal, and replacement of subprograms. For example, an edit operator might replace the sub-program responsible for splitting the canvas, while other edits may insert or remove tiles within the pattern. Please refer to the supplementary for more details. + +# 3.3. Learning an Analogical Editor + +Our goal is to train a model on the synthetic data that is capable of performing analogical edits on real, in-the-wild patterns. Specifically, we aim to generate the target pattern $B'$ from an input triplet $(A, A', B)$ . This approach allows users to demonstrate desired edits with a simple pattern pairs $(A, A')$ , which the model then applies to a complex pattern $B$ to produce $B'$ . Given the success of Latent Diffusion Models (LDMs) in various generative modeling tasks [43], we chose to adapt an LDM for our task as well. We propose TRIFUSER, a latent diffusion model (LDM) for analogical editing (Figure 5). We provide a brief overview of LDMs to provide context before detailing TRIFUSER's modifications for analogical editing. + +Preliminaries: Denoising Diffusion Probabilistic Models (DDPMs) [17] transform random noise into structured data via reverse diffusion steps guided with a conditioning embedding $c(y)$ (often derived from text). Latent Diffusion Models (LDMs) extend DDPMs by mapping data to a lower-dimensional latent space via an encoder. During training, a UNet model [44] learns to remove noise introduced into the latents. During inference, a latent sampled from a normal distribution is iteratively denoised by the model to yield a clean latent. Finally, the clean latent is decoded to generate the output image. Please refer to [55] for + +a more thorough overview. For analogical editing we adapt an Image Variation (IM) model [54], which uses patch-wise image tokens extracted using a text-image encoder [39] as the conditioning embedding $c(y)$ . + +The simplest adaptation of an IM model to our task is to generate $B'$ conditioned on image tokens from all three input images, concatenated as $C = c(A) \| c(A') \| c(B)$ , where $\| \cdot$ denotes token-wise concatenation. This approach, however, suffers from three drawbacks: Token Entanglement, Semantic Bias, and Detail Erosion. We discuss each of these issues briefly, along with our solutions. + +Detail Erosion: Despite using patch-wise tokens, the extracted features lack the fine-grained information needed to retain key aspects of $B$ in the generated pattern $B'$ . Consequently, the model often struggles to preserve elements like tile textures. To address this problem, we combine features from both the first and last layers of the feature encoder: + +$$ +C _ {h l} (P) = \operatorname {L i n e a r} \left(L N \left(c _ {\text {h i g h}} (P)\right) \cdot L N \left(c _ {\text {l o w}} (P)\right), \right. \tag {2} +$$ + +where $LN$ is layer normalization, $\cdot$ denotes channel-wise concatenation, $P$ is an input pattern, and Linear is a linear projection layer that fuses low- and high-level features. + +Semantic Bias: Image variation models typically use feature extractors such as CLIP [39], which are trained to align image embeddings with corresponding text embeddings. Such embeddings emphasize high-level semantics but lack spatial and fine-grained visual details. Combining these embeddings with features from text-free, self-supervised extractors, such as DiNO [2], has been shown to improve performance in various downstream tasks [20, 49]. For our task, a similar approach—combining features from both text-image $(m_{1})$ and self-supervised $(m_{2})$ feature extractors—significantly enhances generation quality. The extracted features are fused as follows: + +$$ +C _ {m i x} (P) = \operatorname {M i x e r} \left(C _ {h l} ^ {m _ {1}} (P) \cdot C _ {h l} ^ {m _ {2}} (P)\right), \tag {3} +$$ + +where $\text{Mixer}$ is a two-layer MLP that integrates features from the two extractors. + +Token Entanglement: To successfully perform an analogical edit, for each patch-level feature token, the model must be able to identify to which source image $(A, A', \text{or } B)$ that patch belongs as well as the 2D position of the patch within that image. Without these distinctions, the model often fails to identify the pattern to edit (i.e., $B$ ) and to recognize the desired edit from $(A, A')$ . To address this problem, we introduce 3D positional encodings: two dimensions for spatial location within each pattern and one dimension for the source image. These encodings are applied to the extracted embeddings, yielding: + +$$ +C ^ {\Omega} = C _ {P E} (A) \parallel C _ {P E} \left(A ^ {\prime}\right) \parallel C _ {P E} (B), \tag {4} +$$ + +$$ +C _ {P E} (P) ^ {x y} = C _ {m i x} (P) ^ {x y} + P E \left(t _ {P}, x, y\right), \tag {5} +$$ + +where $t_P$ is a one-hot vector encoding which input image a token comes from and $PE(t_P, x, y)$ positionally encodes both spatial and source information for each token. + +As we demonstrate in Section 4.4, conditioning on $C^\Omega$ instead of $C$ significantly enhances the quality of patterns generated by TRIFUSER. Our adapted architecture, shown in Figure 5, integrates the modifications described above to effectively address the described drawbacks. Note that this architecture is designed for general analogical editing, not just pattern editing. To enhance generalizability to real-world patterns, we initialize TRIFUSER with an existing pretrained IM model [54], and fine-tune only the denoising UNet and the projection layers in our feature extractor. + +# 4. Experiment + +In this section, we evaluate our approach along three directions: (1) the effectiveness of TRIFUSER at performing analogical edits on complex, real-world patterns, emphasizing how our synthetic data enables editing of in-the-wild pattern images; (2) the ability of TRIFUSER to support programmatic, structure-preserving edits without explicit program inference; and (3) the impact of architectural modifications introduced in TRIFUSER on the quality of generated patterns. We conduct a human perceptual study, quantitative assessments, and qualitative comparisons to demonstrate our system's ability to perform high-quality analogical edits across a range of pattern types. + +# 4.1. Experiment Design + +Datasets: We generate a large synthetic dataset of analogical quartets, i.e., pairs of analogical patterns $(A, A', B, B')$ , using the SPLITWEAVE program samplers introduced in Section 3.2. This synthetic dataset contains approximately 1 million samples covering two pattern styles, namely Split Filling Patterns (SFP) and Motif Tiling Patterns (MTP) (cf. Section 3.1). For MTP patterns, we synthesize $100k$ distinct tiles using the LayerDiffuse [58] model, guided by text + +prompts derived from WordNet [35] noun synsets. Additionally, we construct a synthetic test set with 1000 analogical quartets to evaluate model performance on unseen synthetic data. Further details on dataset construction are provided in the supplementary material. + +To assess TRIFUSER on real-world patterns, we curate a test dataset of 50 patterns created by professional artists and sourced from Adobe Stock. This dataset spans seven distinct sub-domains of 2D patterns, representing a range of pattern styles. These styles include MTP and SFP patterns as well as previously unseen pattern styles such as Memphis-style, geometric, and digital textile patterns. Each pattern is annotated with a desired edit, and we use SPLITWEAVE to generate a pair of simpler patterns $(A, A')$ demonstrating this edit. This test set provides a challenging benchmark to evaluate TRIFUSER's generalization to diverse, real-world editing tasks. + +Training details: We fine-tune a pre-trained diffusion model using our synthetic dataset of analogical quartets (cf. Section 3.3). We initialize our model with Versatile-Diffusion's Image Variation model [54]. We use SigLIP [57] as our text-image feature encoder and DiNOv2 [37] for self-supervised features. We fine-tune the model on 8 A100 GPUs using a batch size of 224 for $\sim 65$ epochs over 7 days. During inference, $B^{\prime}$ is inferred with typical diffusion parameter settings such as a classifier-free guidance weight of 7.5 and 50 denoising steps. + +# 4.2. Analogical Editing Baselines + +To evaluate the analogical editing capability of TRIFUSER, we compare it to three baseline methods, each representing a leading approach for analogical image editing. + +First, we consider training-free editors and latent arithmetic editors. Training-free editors repurpose pre-trained diffusion models to perform analogical edits without additional training [13, 51], leveraging the rich representations learned by diffusion models for editing. In this category, we compare against Analogist [13], the current state-of-the-art method. Latent arithmetic editors, on the other hand, rely on transformations in a learned latent space to infer analogical modifications [40, 48]. Note that these approaches only require samples from the target domain, not analogical training pairs. We implement a baseline for this method by fine-tuning a naive Image Variation model [54] on our synthetic dataset to learn a generative latent embedding space of patterns. At inference, analogical edits are generated using latent arithmetic: given patterns $A$ , $A'$ , and $B$ , we condition the generation of $B'$ on $E(B) + E(A') - E(A)$ . We refer to this baseline as LatentMod. + +Finally, we consider analogy-conditioned generative editors, where models are explicitly trained on analogical data to learn analogical transformations [47]. This category includes our proposed TRIFUSER as well. Image Brush, + +![](images/dadfce37205e3ab655e8f97912c11b153cf00a5e9bb36ef31fce493307676bde.jpg) +Figure 6. Qualitative comparison between patterns generated by our model, TRIFUSER, and the baselines. TRIFUSER generates higher quality patterns with greater fidelity to the input analogy. + +
Preference Rate
TRIFUSER vs. Analogist89.24%
TRIFUSER vs. LatentMod80.59%
TRIFUSER vs. Inpainter72.05%
+ +Table 1. Results of a two-alternative forced-choice perceptual study comparing our model (TRIFUSER) against three baselines. Ours is preferred in the overwhelming majority of judgments. + +the state-of-the-art method, fine-tunes a diffusion inpainting model for analogical editing with multi-modal conditioning. Since code for Image Brush is unavailable, we implement a similar baseline by fine-tuning a Stable Diffusion inpainting model. This model, which we term Inpainter, performs analogical editing by inpainting the lower-left quadrant of a 2x2 analogy grid containing $(A, A', B)$ and conditioned on a fixed text template. + +# 4.3. Editing Real-World Patterns + +To evaluate TRIFUSER's real-world analogical editing capabilities, we conducted a human preference study on the curated test set of Adobe Stock patterns. + +We performed a two-alternative forced-choice perceptual study comparing TRIFUSER with baseline methods on all 50 entries in the test set. Each method generates $k = 9$ outputs for each input tuple, and we select the best one based on visual inspection. Participants were shown edited patterns generated by two different methods along with the input patterns $(A, A', B)$ and instructed to select the edit that best preserved the analogical relationship and exhibited higher image quality. We recruited 42 participants for the study, resulting in a total of 1720 total judgments. + +Table 1 presents the results, showing that TRIFUSER was preferred over all three baselines. Due to the domain gap + +![](images/ac7c32ef2eda801bc4abff8e6c77245fbaa55a6678fb9bd1d1330fc83abc2c27.jpg) +Figure 7. TRIFUSER effectively edits patterns from novel pattern styles not present in the training dataset. TRIFUSER shows a noteworthy ability to generalize beyond its training distribution. + +between the training data of the underlying model [43] and pattern images, Analogist fails to interpret and edit pattern images. Meanwhile, LatentMod fails to perform reasonable edits as the embedding space lacks the low-level details necessary for programmatic edits. While these baselines perform adequately on stylistic edits, they are unsuitable for programmatic editing. When compared to Inpainter, TRI-FUSER was favored in $72.2\%$ of comparisons. Both methods benefit from training on analogical quartets, yet Inpainter sacrifices pattern quality as it generates the edited pattern in only a quarter of the full canvas resolution. + +Figure 6 shows examples of pattern edits generated by TRIFUSER and the baselines, with our model consistently delivering superior results. In Figure 7, we show examples of TRIFUSER's edits on out-of-distribution pattern styles not present in the training set. These results suggest that our synthetic training data enables manipulation of real-world patterns, even extending to certain untrained pattern styles. + +# 4.4. Editing Synthetic Patterns + +Next, we evaluate TRIFUSER's ability to perform programmatic edits on the synthetic validation set, which contains ground truth patterns $B'$ . Ideally, this would involve verifying that the underlying program $z_{\hat{B}'}$ of the generated pattern reflects the same transformation from $z_B$ as that between $z_A$ and $z_A'$ . However, this would require visual program inference on $\hat{B}'$ , which is infeasible. Instead, we approximate this criterion by comparing the program outputs $\hat{B}'$ and $B'$ to see if the visual results align with the intended transformation. To quantify this alignment, we use perceptual metrics—DSim [8], DIST [4] and LPIPS [59]—along with SSIM to capture pixel-level structural similarity. + +Note that analogies can have multiple valid interpreta + +
DSIM (↓)DISTS (↓)LPIPS (↓)SSIM (↑)
Analogist0.4960.4320.6970.494
LatentMod0.2420.3200.6130.502
Inpainter0.0920.2560.3710.713
TRIFUSER0.0740.1840.3040.704
+ +Table 2. Quantitative evaluation on the synthetic validation set shows that TRIFUSER generates patterns with higher perceptual similarity to the ground truth than the baselines. + +
DSIM (↓)DISTS (↓)LPIPS (↓)SSIM (↑)
TRIFUSER0.0740.1840.3040.704
- Pos. Enc.0.1470.2390.3830.659
- Lower0.0870.1960.3350.652
- Mix0.0980.2100.3450.682
Base [54]0.5850.4600.8150.435
+ +Table 3. Subtractive ablation study on TRIFUSER shows that removing any component (see Section 3.3) degrades performance, and that removing all components (Base) results in a sharp decline. + +tions, and even a single interpretation may yield several visually-related variations. To account for this multiplicity, we generate $k = 5$ output patterns for each input set $(A,A^{\prime},B)$ and select the one that maximizes each metric. In other words, we evaluate whether at least one generated output aligns with the intended target. + +Table 2 shows the results of this experiment. First, we note that TRIFUSER outperforms others across all perceptual metrics. These metrics capture different aspects of perceptual similarity and superior performance across all of them suggests a comprehensive improvement. Second, we observe that the analogy-conditioned generative editors (Inpainter & TRIFUSER) surpass both the training-free and latent modification editors. Interestingly, Inpainter achieves slightly higher SSIM scores than TRIFUSER, suggesting that integrating the two approaches could be beneficial. + +# 4.5. TRIFUSER Ablation + +To evaluate the impact of each model component introduced in Section 3.3, we conduct a subtractive analysis on the synthetic validation set. For this ablation study, we remove each component one at a time and measure the resulting performance, as reported in Table 3. The results demonstrate that removing any component leads to a performance drop, with the removal of 3D positional encoding leading to severe degradation. This is understandable: without 3D positional encoding, the network often fails to accurately identify which pattern to edit. For comparison, we also include results from the original Image Variation model [54] trained without any modifications (Base). As expected, this model performs poorly, underscoring the importance of our modifications in achieving high-quality analogical edits. + +![](images/fd914bb398cbb9513f0cfc65c9f944a86415e8ca5ec02facf4a13f538bdaf2aa.jpg) +Figure 8. Our model helps users mix elements of different real-world patterns together, accelerating design exploration. + +![](images/12acb91595b8d52a5af3c4b292685fb67ea1d037271e448d46ab21fa92a96126.jpg) + +![](images/e5ee99dfa369a5563cedcc90b05c09a90532b549b91cede03b12bf456e588955.jpg) + +# 5. Application + +The ability to edit patterns without requiring program inference unlocks new creative possibilities. We demonstrate two practical applications of analogical pattern editing: + +Pattern Mixing: Figure 8 shows example of using our method to mix elements of two real-world patterns $X$ and $Y$ , allowing the user to create unique, hybrid designs. The Mix operator is implemented by using a synthetic pair $(A, A')$ to create a variant $X'$ of $X$ and then using the pair $(X, X')$ to specify an edit to $Y$ : $\text{Mix}(X, Y) = f(X, X', Y)$ , where $X' = f(A, A', X)$ . See the supplementary material for more details. + +Animation Transfer: TRIFUSER can also be used to create animated sequences of edited patterns. By leveraging parametric SPLITWEAVE programs, users can generate animations for simple patterns and then apply these animations to complex patterns with no additional effort. See the video in the supplementary material for examples. + +# 6. Conclusion + +We introduced a novel approach for programmatic editing of visual patterns without inferring the underlying program. By using analogies to express desired edits and a learned conditional generative model to execute them, our method provides an intuitive solution for pattern manipulation. A key component of our approach is SPLITWEAVE, a domain-specific language for generating diverse, structured pattern data. Paired with our procedure for sampling analogical quartets, SPLITWEAVE enables the creation of a large, high-quality dataset for training. We also presented TRIFUSER, a Latent Diffusion Model (LDM) designed to overcome critical issues that emerge when LDMs are naively deployed for analogical pattern editing, enabling high-fidelity edits that capture user intentions. Our experiments demonstrate that TRIFUSER successfully edits real-world patterns and surpasses baseline methods, while also generalizing to novel pattern styles beyond its training distribution. We believe that our DSL, dataset, and model will help drive further research on in-the-wild pattern image editing. Looking forward, we aim to extend this analogical editing framework to other domains such as semi-parametric 3D modeling while continuing to improve synthetic data generation. + +# References + +[1] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei A. Efros. Visual prompting via image inpainting. arXiv preprint arXiv:2209.00647, 2022. 2, 3 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5 +[3] Ta-Ying Cheng, Matheus Gadelha, Thibault Groueix, Matthew Fisher, Radomir Mech, Andrew Markham, and Niki Trigoni. Learning continuous 3d words for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6753-6762, 2024. 3 +[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P. Simoncelli. Image quality assessment: Unifying structure and texture similarity. CoRR, abs/2004.07728, 2020. 7 +[5] Kevin Ellis, Daniel Ritchie, Armando Solar-Lezama, and Josh Tenenbaum. Learning to infer graphics programs from hand-drawn images. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 2 +[6] Thomas G. Evans. A heuristic program to solve geometric-analogy problems. In Proceedings of the April 21-23, 1964, Spring Joint Computer Conference, page 327-338, New York, NY, USA, 1964. Association for Computing Machinery. 3 +[7] Brian Falkenhainer, Kenneth D. Forbus, and Dedre Gentner. The structure-mapping engine. In Proceedings of the Fifth AAAI National Conference on Artificial Intelligence, page 272-277. AAAI Press, 1986. 3 +[8] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 7 +[9] Aditya Ganeshan, R. Kenny Jones, and Daniel Ritchie. Improving unsupervised visual program inference with code rewriting families. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3 +[10] Aditya Ganeshan, Ryan Y. Huang, Xianghao Xu, R. Kenny Jones, and Daniel Ritchie. Parsel: Parameterized shape editing with language, 2024. 3 +[11] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Advances in Neural Information Processing Systems, pages 5885–5897. Curran Associates, Inc., 2021. 3 +[12] Dedre Gentner. Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2):155-170, 1983. 4 +[13] Zheng GU, Shiyuan Yang, Jing Liao, Jing Huo, and Yang Gao. Analogist: Out-of-the-box visual in-context learning with image diffusion model. ACM Transactions on Graphics (TOG), 2024. 3, 6 +[14] P. Guehl, R. Allegre, J.-M. Dischler, B. Benes, and E. Galin. Semi-procedural textures using point process texture basis functions. Computer Graphics Forum, 39(4):159-171, 2020. 3 + +[15] Julia Guerrero-Viu, Milos Hasan, Arthur Roullier, Midhun Harikumar, Yiwei Hu, Paul Guerrero, Diego Gutierrez, Bellen Masia, and Valentin Deschaintre. Texsliders: Diffusion-based texture editing in clip space. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 3 +[16] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image Analogies. Association for Computing Machinery, New York, NY, USA, 1 edition, 2023. 3 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 5 +[18] Douglas Hofstadter and Melanie Mitchell. The Copycat project: a model of mental fluidity and analogy-making, page 205-267. Basic Books, Inc., USA, 1995. 3 +[19] Yiwei Hu, Paul Guerrero, Milos Hasan, Holly Rushmeier, and Valentin Deschaintre. Node graph optimization using differentiable proxies. In ACM SIGGRAPH Conference Proceedings, 2022. 3 +[20] Dongsheng Jiang, Yuchen Liu, Songlin Liu, XIAOPENG ZHANG, Jin Li, Hongkai Xiong, and Qi Tian. From CLIP to DINO: Visual encoders shout in multi-modal large language models, 2024. 5 +[21] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapeassembly: Learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics (TOG), Siggraph Asia 2020, 2020. 3 +[22] R. Kenny Jones, David Charatan, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapemod: Macro operation discovery for 3d shape programs. ACM Transactions on Graphics (TOG), Siggraph 2021, 2021. 3 +[23] R. Kenny Jones, Homer Walke, and Daniel Ritchie. Plad: Learning to infer shape programs with pseudo-labels and approximate distributions. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[24] R. Kenny Jones, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapecoder: Discovering abstractions for visual programs from unstructured primitives. ACM Transactions on Graphics (TOG), Siggraph 2023, 42(4), 2023. 3 +[25] R. Kenny Jones, Renhao Zhang, Aditya Ganeshan, and Daniel Ritchie. Learning to edit visual programs with self-supervision. In Advances in Neural Information Processing Systems, 2024. 3 +[26] Milin Kodnongbua, Benjamin Jones, Maaz Bin Safeer Ahmad, Vladimir Kim, and Adriana Schulz. Reparamcad: Zero-shot cad re-parameterization for interactive manipulation. In SIGGRAPH Asia 2023 Conference Papers, New York, NY, USA, 2023. Association for Computing Machinery. 3 +[27] Beichen Li, Liang Shi, and Wojciech Matusik. End-to-end procedural material capture with proxy-free mixed-integer optimization. ACM Trans. Graph., 42(4), 2023. 3 +[28] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J. Mitra. Sketch2cad: Sequential cad modeling by sketching in + +context. ACM Trans. Graph. (Proceedings of SIGGRAPH Asia 2020), 39(6):164:1-164:14, 2020. 3 +[29] Wen-Ding Li, Keya Hu, Carter Larsen, Yuqing Wu, Simon Alford, Caleb Woo, Spencer M. Dunn, Hao Tang, Michelangelo Naim, Dat Nguyen, Wei-Long Zheng, Zenna Tavares, Yewen Pu, and Kevin Ellis. Combining induction and transduction for abstract reasoning, 2024. 4 +[30] Hugo Loi, Thomas Hurtut, Romain Vergne, and Joelle Thollot. Programmable 2d arrangements for element texture design. ACM Trans. Graph., 36(4), 2017. 3, 4 +[31] Arman Maesumi, Dylan Hu, Krishi Saripalli, Vladimir Kim, Matthew Fisher, Soeren Pink, and Daniel Ritchie. One noise to rule them all: Learning a unified model of spatially-varying noise patterns. ACM Trans. Graph., 43(4), 2024. 3 +[32] Jiayuan Mao, Xiuming Zhang, Yikai Li, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Program-Guided Image Manipulators. In International Conference on Computer Vision, 2019. 2 +[33] L. McCarthy, C. Reas, and B. Fry. Getting Started with P5.js: Making Interactive Graphics in JavaScript and Processing. Maker Media, Incorporated, 2015. 4 +[34] Zichong Meng, Changdi Yang, Jun Liu, Hao Tang, Pu Zhao, and Yanzhi Wang. Instructgie: Towards generalizable image editing. arXiv preprint arXiv:2403.05018, 2024. 3 +[35] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 6 +[36] OpenAI. Gpt-4 technical report, 2024. 3 +[37] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024, 2024. 6 +[38] Ofek Pearl, Itai Lang, Yuhua Hu, Raymond A. Yeh, and Rana Hanocka. Geocode: Interpretable shape programs. 2022. 3 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5 +[40] Scott Reed, Yi Zhang, Yuting Zhang, and Honglak Lee. Deep visual analogy-making. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, page 1252-1260, Cambridge, MA, USA, 2015. MIT Press. 6 +[41] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, Haiyong Jiang, Zhongang Cai, Junzhe Zhang, Liang Pan, Mingyuan Zhang, Haiyu Zhao, and Shuai Yi. Csg-stump: A learning friendly csg-like representation for interpretable + +shape parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3 +[42] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, and Junzhe Zhang. Extrudenet: Unsupervised inverse sketch-and-extrude fornbsp; shape parsing. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, page 482-498, Berlin, Heidelberg, 2022. Springer-Verlag. 3 +[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 5, 7 +[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 5 +[45] Christian Santoni and Fabio Pellacini. gtangle: a grammar for the procedural generation of tangle patterns. ACM Trans. Graph., 35(6), 2016. 3, 4 +[46] Liang Shi, Beichen Li, Miloš Hasan, Kalyan Sunkavalli, Tamy Boubekeur, Radomir Mech, and Wojciech Matusik. Match: differentiable material graphs for procedural material capture. ACM Trans. Graph., 39(6), 2020. 2, 3 +[47] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: learning visual in-context instructions for exemplar-based image manipulation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 2, 3, 6 +[48] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 6 +[49] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 5 +[50] Trieu Trinh, Yuhuai Tony Wu, Quoc Le, He He, and Thang Luong. Solving olympiad geometry without human demonstrations. Nature, 625:476-482, 2024. 4 +[51] Adéla Šubrtová, Michal Lukáč, Jan Čech, David Futschik, Eli Shechtman, and Daniel Sykora. Diffusion image analogies. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 3, 6 +[52] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6772-6782, 2021. 3 +[53] Xianghao Xu, Wenzhe Peng, Chin-Yi Cheng, Karl D. D. Willis, and Daniel Ritchie. Inferring cad modeling sequences using zone graphs. In CVPR, 2021. 3 +[54] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7754-7765, 2023. 2, 5, 6, 8 + +[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 5 +[56] Mehmet Ersin Yumer, Siddhartha Chaudhuri, Jessica K. Hodgins, and Levent Burak Kara. Semantic shape editing using deformation handles. ACM Trans. Graph., 34(4), 2015. 3 +[57] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 6 +[58] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. ACM Trans. Graph., 43(4), 2024. 6 +[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7 \ No newline at end of file diff --git a/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/images.zip b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..73bc75adfd228ccc504d90f2a86129d5cff066d2 --- /dev/null +++ b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e01df9701998ef919a8a9ff535de36633ff52457e98b7807b4b61d6f99be02b6 +size 599071 diff --git a/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/layout.json b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..45eb6d7122081ff4dbaf759012c8ff47c57b8452 --- /dev/null +++ b/2025/Pattern Analogies_ Learning to Perform Programmatic Image Edits by Analogy/layout.json @@ -0,0 +1,9480 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 51, + 103, + 543, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 103, + 543, + 122 + ], + "spans": [ + { + "bbox": [ + 51, + 103, + 543, + 122 + ], + "type": "text", + "content": "Pattern Analogies: Learning to Perform Programmatic Image Edits by Analogy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 144, + 178, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 144, + 178, + 172 + ], + "spans": [ + { + "bbox": [ + 87, + 144, + 178, + 172 + ], + "type": "text", + "content": "Aditya Ganeshan* Brown University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 144, + 290, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 144, + 290, + 171 + ], + "spans": [ + { + "bbox": [ + 205, + 144, + 290, + 171 + ], + "type": "text", + "content": "Thibault Groueix \nAdobe Research" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 144, + 399, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 144, + 399, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 144, + 399, + 171 + ], + "type": "text", + "content": "Paul Guerrero \nAdobe Research" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 425, + 144, + 506, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 144, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 425, + 144, + 506, + 171 + ], + "type": "text", + "content": "Radomír Měch \nAdobe Research" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 190, + 191, + 271, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 191, + 271, + 218 + ], + "spans": [ + { + "bbox": [ + 190, + 191, + 271, + 218 + ], + "type": "text", + "content": "Matthew Fisher Adobe Research" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 191, + 404, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 191, + 404, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 191, + 404, + 220 + ], + "type": "text", + "content": "Daniel Ritchie Brown University" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 241, + 545, + 405 + ], + "blocks": [ + { + "bbox": [ + 50, + 241, + 545, + 405 + ], + "lines": [ + { + "bbox": [ + 50, + 241, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 50, + 241, + 545, + 405 + ], + "type": "image", + "image_path": "fd6ae179c58cf2692114081c8e54597143d51ca15a2478607cc048ccce89449f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "lines": [ + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "type": "text", + "content": "Figure 1. Our system performs programmatic edits on pattern images without inferring their underlying programs. (Left) Desired edits, expressed with a pair of patterns " + }, + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "type": "inline_equation", + "content": "(A, A')" + }, + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "type": "text", + "content": ", are executed on a target pattern " + }, + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "type": "text", + "content": " by a generative model to produce " + }, + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "type": "text", + "content": ". (Right) Parametric changes " + }, + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "type": "inline_equation", + "content": "A \\to A'" + }, + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "type": "text", + "content": " enabled by our domain-specific pattern language induce corresponding changes to the more complex pattern " + }, + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 408, + 547, + 442 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 450, + 192, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 450, + 192, + 463 + ], + "spans": [ + { + "bbox": [ + 143, + 450, + 192, + 463 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": "Pattern images are everywhere in the digital and physical worlds, and tools to edit them are valuable. But editing pattern images is tricky: desired edits are often programmatic: structure-aware edits that alter the underlying program which generates the pattern. One could attempt to infer this underlying program, but current methods for doing so struggle with complex images and produce unorganized programs that make editing tedious. In this work, we introduce a novel approach to perform programmatic edits on pattern images. By using a pattern analogy—a pair of simple patterns to demonstrate the intended edit—and a learning-based generative model to execute these edits, our method allows users to intuitively edit patterns. To enable this paradigm, we introduce SPLITWEAVE, a domain-specific language that, combined with a framework for sampling synthetic pattern analogies, enables the creation of a large, high-quality synthetic training dataset. We also present TRIFUSER, a Latent Diffusion Model (LDM) designed to overcome critical issues that arise when naively" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 452, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 500 + ], + "type": "text", + "content": "deploying LDMs to this task. Extensive experiments on real-world, artist-sourced patterns reveals that our method faithfully performs the demonstrated edit while also generalizing to related pattern styles beyond its training distribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 504, + 387, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 387, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 387, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 525, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 620 + ], + "type": "text", + "content": "Visual pattern designs enhance digital media such as presentations, website themes, and user interfaces, and they are woven into the physical world through textiles, wallpapers, and product designs like hardware covers. Given the ubiquity of patterns, methods for editing them are essential: designers should be able to quickly experiment with variations, customize designs to meet specific needs, and adapt existing patterns to align with evolving trends." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 621, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 621, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 545, + 693 + ], + "type": "text", + "content": "Editing pattern images is not straightforward, as patterns are inherently structured, defined by rules that govern their layout and composition: tiling patterns adhere to principles of alignment and repetition (see Figure 1: top left), while retro-style designs rely on spatial divisions and fills (see Figure 1: bottom center). The edits that designers desire" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 317, + 703, + 504, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 504, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 504, + 713 + ], + "type": "text", + "content": "*Work performed during an internship at Adobe Research" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28715" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "often aim to adjust these underlying organizational rules rather than make superficial, pixel-level changes. We refer to such edits as programmatic edits, requiring manipulation of the underlying program that defines a pattern's structure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 288, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 122, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 288, + 302 + ], + "type": "text", + "content": "One strategy for enabling such programmatic edits is visual program inference (VPI) [5, 32, 46], where a program that replicates an image is automatically inferred, allowing users to modify the image by adjusting program parameters. However, applying VPI to patterns presents two obstacles. First, VPI attempts to infer a program that fully replicates a pattern, which can be challenging as patterns are often semiparametric, blending rule-based logic with non-parametric components. For instance, the layout of elements in a tiling pattern may be rule-based, but the elements themselves may not be. Second, editing with an inferred program can be cumbersome, as they are often poorly-structured, with many unlabeled parameters, making them difficult to interpret. Consequently, VPI not only solves a more complex problem than necessary but also makes editing more challenging." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": "Can we perform programmatic edits without inferring the underlying program? Doing so requires the ability to express and execute the edit—both without direct access to the program's parameters. To express a programmatic edit, it's crucial to specify both which underlying parameter(s) to change and how to modify them. We draw inspiration from how humans communicate transformations: through analogies. By providing a pair of simple example patterns " + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "inline_equation", + "content": "(A, A')" + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": " that illustrate the desired change, users can intuitively convey both aspects of the edit. To execute these edits, we employ a learning-based conditional generative model. Given a pair of simple patterns " + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "inline_equation", + "content": "(A, A')" + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": " and a complex target pattern " + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": ", our system generates " + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": ", an edited version of " + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": " which performs the transformation demonstrated between " + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": " while preserving " + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": "'s other structural features. Crucially, " + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": " does not need to replicate or even be similar to " + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": "—it only needs to demonstrate which property to edit and how. Thus, specifying " + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 304, + 288, + 591 + ], + "type": "text", + "content": " is a much easier task than solving VPI. While prior works [1, 47, 51] have applied analogical editing to image manipulation, they focus primarily on appearance modifications. In contrast, our approach is the first to use analogies for programmatic, structure-aware edits. Figure 1 (left) shows examples of analogical editing on complex, real-world patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "To make our approach possible, we introduce SPLITWEAVE: a domain-specific language (DSL) for crafting visual patterns. SPLITWEAVE serves two purposes in our method. First, it enables parametric definition of input pairs " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "(A,A^{\\prime})" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", allowing users to guide transformations in " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "(B,B^{\\prime})" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " as if the underlying program for " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " were accessible. In Figure 1 (right), modifying the SPLITWEAVE program for " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "A^\\prime" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " produces corresponding changes in " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "B^{\\prime}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ". Second, SPLITWEAVE supports the creation of large-scale synthetic training data. We develop program samplers" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "that generate high-quality patterns in two common styles: tiling-based designs with repeating elements and color field patterns characterized by splitting the canvas into intricate colored regions. Training a model for analogical editing requires a dataset of quartets " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "(A, A', B, B')" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ". By applying identical programmatic edits to the SPLITWEAVE programs for both " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ", we ensure that the transformation from " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " mirrors that from " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": ". This approach allows us to generate a diverse dataset of analogical quartets. Models trained on this dataset generalize effectively to real-world patterns within these styles and even extend to related styles." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 225, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 225, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 545, + 370 + ], + "type": "text", + "content": "We use this synthetic dataset to train a novel diffusion-based conditional generative model for executing analogical edits. Our model directly generates edited patterns " + }, + { + "bbox": [ + 304, + 225, + 545, + 370 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 304, + 225, + 545, + 370 + ], + "type": "text", + "content": " by conditioning on visual features extracted from input patterns " + }, + { + "bbox": [ + 304, + 225, + 545, + 370 + ], + "type": "inline_equation", + "content": "(A, A', B)" + }, + { + "bbox": [ + 304, + 225, + 545, + 370 + ], + "type": "text", + "content": ". Existing image-conditioned diffusion models [43, 54] prove ineffective, as they fail to interpret the input analogies accurately and neglect fine details. To address these issues, we incorporate architectural enhancements that enable our model, TRIFUSER, to effectively perform analogical edits. With these improvements, TRIFUSER surpasses prior architectures for analogical editing when applied to pattern images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 380, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 535 + ], + "type": "text", + "content": "To evaluate our method, we curated a test set of 50 patterns from Adobe Stock spanning 7 distinct styles. A perceptual study on this dataset reveals that participants consistently prefer edits by TRIFUSER over those produced by recent training-free and training-based analogical editing methods. Despite training only on two of these styles, our model successfully generalizes to the remaining five, which were unseen during training. On a synthetic validation set with ground-truth analogical edits, our model achieves higher structural and perceptual similarity to the ground truth than prior methods. Finally, we showcase two compelling applications of our approach: mixing attributes of different patterns and transferring pattern animations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 545, + 501, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 545, + 501, + 556 + ], + "spans": [ + { + "bbox": [ + 318, + 545, + 501, + 556 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 567, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 567, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 567, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 314, + 567, + 545, + 602 + ], + "type": "text", + "content": "1. A novel framework for performing programmatic edits to pattern images without requiring program inference, leveraging analogies to specify and apply edits." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 604, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 545, + 639 + ], + "type": "text", + "content": "2. SPLITWEAVE, a DSL for crafting a diverse range of visual patterns, designed to support both parametric control and synthetic dataset generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 640, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 545, + 664 + ], + "type": "text", + "content": "3. A procedure for generating synthetic analogical quartets, enabling editing of in-the-wild patterns." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 545, + 713 + ], + "type": "text", + "content": "4. TRIFUSER, a diffusion-based conditional generative model that achieves high fidelity in analogical edits, surpassing prior techniques in both analogical fidelity and generation quality." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28716" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 151 + ], + "type": "text", + "content": "We review three key areas: (1) Visual Program Inference (VPI) for programmatic editing of structured visual data and its limitations, (2) DSLs and synthetic data generation, specifically for visual patterns, and (3) analogical reasoning in computing, particularly for editing images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 151, + 288, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 151, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 151, + 288, + 390 + ], + "type": "text", + "content": "Visual Program Inference for Editing: Visual Program Inference (VPI) enables programmatic edits of visual data by inferring executable programs from visual inputs. Prior works have achieved promising results in inferring material graphs [19, 27, 31, 46] and CAD programs for 2D [11, 28] and 3D [42, 53] inputs, using large annotated datasets [46, 52], differentiable program approximations [19, 41], or bootstrapped learning [9, 23, 25]. VPI is challenging to adapt to pattern editing due to the scarcity of high-quality annotated pattern data and the non-differentiability of most pattern programs. Also, VPI approaches often yield complex programs that are difficult to edit and interpret, making them impractical for editing. To address these challenges, recent work has aimed to simplify programmatic editing by inferring edit-specific controls [3, 10, 15] or a limited set of semantically meaningful parameters [22, 24, 26, 56]. Our approach shares this goal of enabling accessible control but extends it further: we transfer control from simple parametric objects to complex in-the-wild images via analogy, bypassing the need for VPI." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 390, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 498 + ], + "type": "text", + "content": "DSL and Synthetic data Domain-Specific Languages (DSLs) enable concise descriptions of structured objects, facilitating their creation. Prior works have developed DSLs for Zentangle patterns [45], material graphs [46], semiparametric textures [14], and 3D models [21, 38]. Our DSL focuses on visual patterns constructed through partitioning and merging of canvas fragments. Closest to our work is ETD [30], which also uses canvas partitioning and merging operators, though it is limited to stationary patterns." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "content": "Analogical Reasoning Analogical reasoning is a foundational AI task: early work includes Evans' ANALOGY program [6], CopyCat [18], and Structure-Mapping Engine [7]. In visual computing, Image Analogies [16] pioneered the concept of analogy-driven editing. Recently, diffusion models have been adapted for analogical editing. DIA [51] introduced a training-free approach to analogical editing using pretrained diffusion models. Analogist [13] offers a complementary method, leveraging inpainting models alongside multimodal reasoning from large language models [36]. These training-free approaches are limited to images within the diffusion model's training domain, limiting their applicability to patterns. Other methods attempt to learn analogical editors by finetuning diffusion models on analogical pairs [1, 34, 47]. However, the focus of all these works remains largely on stylistic, appearance edits, often failing to perform programmatic edits. This limitation arises both from the models' architectures" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 424, + 269 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 424, + 269 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 424, + 269 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 424, + 269 + ], + "type": "image", + "image_path": "913e64525e76d1571c4d45cbccb7a748fef3f921fd26c16dcd740aea4e3d2b42.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 278, + 547, + 334 + ], + "lines": [ + { + "bbox": [ + 304, + 278, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 547, + 334 + ], + "type": "text", + "content": "Figure 2. Overview: To create high-quality visual patterns, we introduce a custom DSL called SPLITWEAVE. Pairs of SPLITWEAVE programs " + }, + { + "bbox": [ + 304, + 278, + 547, + 334 + ], + "type": "inline_equation", + "content": "(A,B)" + }, + { + "bbox": [ + 304, + 278, + 547, + 334 + ], + "type": "text", + "content": " are then jointly edited to create analogical quartets. This synthetic data is then used to train TRI-FUSER, a neural network for analogical pattern editing." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 425, + 72, + 545, + 269 + ], + "blocks": [ + { + "bbox": [ + 425, + 72, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 425, + 72, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 425, + 72, + 545, + 269 + ], + "type": "image", + "image_path": "48e6af05707f57f7bf544a13dbf34e0afb2c37d716440149fdcb58ff236fb24d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 346, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 383 + ], + "type": "text", + "content": "and from the lack of training pairs with programmatic edits. Our work addresses both these gaps, enabling structured, programmatic analogical edits for visual patterns." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 392, + 361, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 361, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 361, + 404 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "content": "Our objective is to enable programmatic edits of 2D visual patterns without inferring their underlying programs. Instead, we propose an alternative that uses analogies to express desired edits and a conditional generative model to execute them. Formally, given two source patterns " + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "content": " that demonstrate a desired edit, along with a target pattern " + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "content": ", our goal is to generate an edited target pattern " + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "content": " that applies this edit to " + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "content": ". This task is defined as learning a mapping " + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "inline_equation", + "content": "f(A, A', B) \\to B'" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "inline_equation", + "content": "A, A', B" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "content": " are 2D RGB images (" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "content": "). To learn this mapping, we generate a large synthetic dataset of analogical pattern quartets (" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "inline_equation", + "content": "A, A', B, B'" + }, + { + "bbox": [ + 304, + 412, + 545, + 555 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 555, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 651 + ], + "type": "text", + "content": "Figure 2 provides a schematic overview of our approach. First, in Section 3.1 we introduce SPLITWEAVE, a Domain-Specific Language (DSL) that enables the creation and manipulation of various kinds of patterns. Section 3.2 describes our approach for sampling analogical quartets in SPLITWEAVE to create the synthetic training data. Finally, in Section 3.3, we present TRIFUSER, a conditional generative model that learns to execute analogical edits." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 659, + 476, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 476, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 476, + 672 + ], + "type": "text", + "content": "3.1. A Language for Visual Patterns" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "To enable programmatic edits without program inference, our approach requires two core capabilities: (a) generating a large, high-quality synthetic dataset for training models to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28717" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 72, + 283, + 217 + ], + "blocks": [ + { + "bbox": [ + 53, + 72, + 283, + 217 + ], + "lines": [ + { + "bbox": [ + 53, + 72, + 283, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 283, + 217 + ], + "type": "image", + "image_path": "edcb66e9ed39bf509518c9da09cab9e3ab775863a271893b17d82084cccc2a62.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 287, + 261 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 287, + 261 + ], + "type": "text", + "content": "Figure 3. Custom program samplers for two pattern styles. Our samplers produce diverse and high-quality patterns, enabling generalization to real-world patterns." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 270, + 286, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 286, + 437 + ], + "type": "text", + "content": "reliably execute analogical edits, and (b) the ability to create and parametrically control analogy inputs at test time to effectively express desired edits. Existing pattern generation tools are insufficient for these needs, as they are either limited to narrow pattern domains [45] or demand intense coding effort to produce diverse, high-quality patterns [30, 33]. To address these limitations, we introduce SPLITWEAVE, a DSL designed specifically to support analogical transformations in visual patterns. SPLITWEAVE combines abstractions for pattern synthesis with a node-based visual programming interface (see Supplementary), enabling efficient generation of high-quality synthetic patterns for training while allowing flexible, precise pattern manipulation to define analogy inputs at test time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 438, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 286, + 581 + ], + "type": "text", + "content": "SPLITWEAVE uses three types of operations for structured pattern creation: (1) Canvas Fragmentation, which allows structured divisions of the canvas, such as brick-like or voronoi splits; (2) Fragment ID-Aware Operations, enabling transformations that vary across fragments (e.g., scaling alternating rows or columns) to support spatial variability in non-stationary pattern designs; and (3) various SVG Operators for outlining, coloring, and compositing. Together, these operations enable efficient creation of patterns with complex structure and visual variety. Figure 2 (left) illustrates these capabilities in a SPLITWEAVE program for generating a tiling pattern design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "Our goal is to generate high-quality synthetic patterns using SPLITWEAVE that enable trained editing models to generalize well to real-world patterns. Naive sampling from the DSL grammar often leads to overly complex or incoherent patterns, limiting their effectiveness in model training. Instead, we draw inspiration from recent advances in fields such as geometric problem solving [50] and abstract reasoning [29], where tailored data generators have proven to be essential for tackling complex tasks. Following a similar approach, we design custom program samplers for two versatile and widely-used pattern styles. The first, Motif" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 73, + 524, + 180 + ], + "blocks": [ + { + "bbox": [ + 307, + 73, + 524, + 180 + ], + "lines": [ + { + "bbox": [ + 307, + 73, + 524, + 180 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 524, + 180 + ], + "type": "image", + "image_path": "eb3ddf8bc4238274e4781425b7305a145a88a8aff3e59d7976c5ce23b9672deb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "type": "text", + "content": "Figure 4. We create synthetic analogical quartets " + }, + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "type": "inline_equation", + "content": "(A, A', B, B')" + }, + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "type": "text", + "content": " with consistent edits between " + }, + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "type": "text", + "content": " pairs, providing data for training an analogical editing models." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 236, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 545, + 390 + ], + "type": "text", + "content": "Tiling Patterns (MTP), consists of compositions based on repeated Tile elements. These patterns exhibit controlled variations in tile properties across the canvas (e.g. orientation, color, and scale), creating visually cohesive yet richly diverse structures. The second, Split-Filling Patterns (SFP), are generated by dividing the canvas into ordered fragments, applying region-specific coloring and transformations based on fragment IDs. Both pattern styles are common in digital design and support a wide range of programmatic variations, making them particularly suited for analogical editing tasks. Example patterns generated by our program samplers are shown in Figure 3; additional implementation details are in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 399, + 470, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 470, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 470, + 411 + ], + "type": "text", + "content": "3.2. Sampling Analogical Quartets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "text", + "content": "With the ability to generate diverse synthetic patterns using SPLITWEAVE (Section 3.1), our goal is now to construct analogical pattern quartets " + }, + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "inline_equation", + "content": "(A, A', B, B')" + }, + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "text", + "content": ". Each pattern image in a quartet is generated by a SPLITWEAVE program " + }, + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 417, + 545, + 501 + ], + "type": "text", + "content": ". These quartets serve as structured training data for editing models, allowing them to learn consistent transformations that can generalize across different pattern domains." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "text", + "content": "Analogies in our framework are grounded in Structure Mapping Theory [12], which defines analogies as mappings of relational structure from a base to a target domain. We designate " + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "inline_equation", + "content": "(A, A')" + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "text", + "content": " as the base and " + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "inline_equation", + "content": "(B, B')" + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "text", + "content": " as the target, with the requirement that the relationship " + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "text", + "content": " between program pairs " + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "inline_equation", + "content": "(z_A, z_{A'})" + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "inline_equation", + "content": "(z_B, z_{B'})" + }, + { + "bbox": [ + 304, + 501, + 545, + 573 + ], + "type": "text", + "content": " remains consistent:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 371, + 583, + 545, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 583, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 371, + 583, + 545, + 596 + ], + "type": "interline_equation", + "content": "R \\left(z _ {A}, z _ {A ^ {\\prime}}\\right) = R \\left(z _ {B}, z _ {B ^ {\\prime}}\\right). \\tag {1}", + "image_path": "4adf99173f368fef8d558af424cc9218d7f0159ddbee1ae3958d367d1faf1a58.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "content": "Rather than focusing on visual similarity between the patterns " + }, + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "inline_equation", + "content": "(A, A')" + }, + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "content": " themselves, this program-level analogy allows us to generate quartets with transformations that affect the underlying program, facilitating programmatic edits." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "To construct these analogical quartets, we use a program sampler along with a predefined set of editing operators " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". For each quartet, we begin by sampling an edit " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "e \\in E" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", followed by sampling initial programs " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{A}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{B}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " that are compatible with " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". Applying " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " to both " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{A}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "z_{B}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " yields trans" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28718" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 74, + 287, + 185 + ], + "blocks": [ + { + "bbox": [ + 58, + 74, + 287, + 185 + ], + "lines": [ + { + "bbox": [ + 58, + 74, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 58, + 74, + 287, + 185 + ], + "type": "image", + "image_path": "1e58a36e6a34b85b90780df24aa9f1019d13ffdbb202fceb99a55e70fdcabae2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 195, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 195, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 547, + 229 + ], + "type": "text", + "content": "Figure 5. (Left) TRIFUSER is a latent diffusion model conditioned on patch-wise tokens of the input images " + }, + { + "bbox": [ + 46, + 195, + 547, + 229 + ], + "type": "inline_equation", + "content": "(A, A', B)" + }, + { + "bbox": [ + 46, + 195, + 547, + 229 + ], + "type": "text", + "content": " to generate the analogically edited pattern " + }, + { + "bbox": [ + 46, + 195, + 547, + 229 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 46, + 195, + 547, + 229 + ], + "type": "text", + "content": ". (Right) To achieve high-quality edits, we enrich these tokens by fusing multi-level features from multiple encoders, followed by a 3D positional encoding: 2D to specify spatial locations and 1D to specify the token's source " + }, + { + "bbox": [ + 46, + 195, + 547, + 229 + ], + "type": "inline_equation", + "content": "(A, A'" + }, + { + "bbox": [ + 46, + 195, + 547, + 229 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 195, + 547, + 229 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 195, + 547, + 229 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 292, + 74, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 292, + 74, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 292, + 74, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 292, + 74, + 545, + 182 + ], + "type": "image", + "image_path": "9523dd832b579fce58bd2582e931c69429c263448d226604391f7aa1cf1311bf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 241, + 286, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 286, + 312 + ], + "type": "text", + "content": "formed programs " + }, + { + "bbox": [ + 46, + 241, + 286, + 312 + ], + "type": "inline_equation", + "content": "z_{A'}" + }, + { + "bbox": [ + 46, + 241, + 286, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 241, + 286, + 312 + ], + "type": "inline_equation", + "content": "z_{B'}" + }, + { + "bbox": [ + 46, + 241, + 286, + 312 + ], + "type": "text", + "content": ". By using identical transformations across domains, we ensure a consistent \"edit relation\" across the quartet, satisfying Equation 1 by construction. In Figure 4, we illustrate examples of synthetic analogical quartets generated using this method, demonstrating consistent transformations between " + }, + { + "bbox": [ + 46, + 241, + 286, + 312 + ], + "type": "inline_equation", + "content": "(A, A')" + }, + { + "bbox": [ + 46, + 241, + 286, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 241, + 286, + 312 + ], + "type": "inline_equation", + "content": "(B, B')" + }, + { + "bbox": [ + 46, + 241, + 286, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 313, + 286, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 286, + 396 + ], + "type": "text", + "content": "Edit Operators " + }, + { + "bbox": [ + 46, + 313, + 286, + 396 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 313, + 286, + 396 + ], + "type": "text", + "content": ". We focus on edits targeting specific sub-parts of the program. Specifically, we consider three types of edits: insertion, removal, and replacement of subprograms. For example, an edit operator might replace the sub-program responsible for splitting the canvas, while other edits may insert or remove tiles within the pattern. Please refer to the supplementary for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 407, + 212, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 212, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 212, + 419 + ], + "type": "text", + "content": "3.3. Learning an Analogical Editor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "text", + "content": "Our goal is to train a model on the synthetic data that is capable of performing analogical edits on real, in-the-wild patterns. Specifically, we aim to generate the target pattern " + }, + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "text", + "content": " from an input triplet " + }, + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "inline_equation", + "content": "(A, A', B)" + }, + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "text", + "content": ". This approach allows users to demonstrate desired edits with a simple pattern pairs " + }, + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "inline_equation", + "content": "(A, A')" + }, + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "text", + "content": ", which the model then applies to a complex pattern " + }, + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 46, + 426, + 286, + 581 + ], + "type": "text", + "content": ". Given the success of Latent Diffusion Models (LDMs) in various generative modeling tasks [43], we chose to adapt an LDM for our task as well. We propose TRIFUSER, a latent diffusion model (LDM) for analogical editing (Figure 5). We provide a brief overview of LDMs to provide context before detailing TRIFUSER's modifications for analogical editing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "Preliminaries: Denoising Diffusion Probabilistic Models (DDPMs) [17] transform random noise into structured data via reverse diffusion steps guided with a conditioning embedding " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "c(y)" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " (often derived from text). Latent Diffusion Models (LDMs) extend DDPMs by mapping data to a lower-dimensional latent space via an encoder. During training, a UNet model [44] learns to remove noise introduced into the latents. During inference, a latent sampled from a normal distribution is iteratively denoised by the model to yield a clean latent. Finally, the clean latent is decoded to generate the output image. Please refer to [55] for" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "text", + "content": "a more thorough overview. For analogical editing we adapt an Image Variation (IM) model [54], which uses patch-wise image tokens extracted using a text-image encoder [39] as the conditioning embedding " + }, + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "inline_equation", + "content": "c(y)" + }, + { + "bbox": [ + 304, + 241, + 545, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "content": "The simplest adaptation of an IM model to our task is to generate " + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "content": " conditioned on image tokens from all three input images, concatenated as " + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "inline_equation", + "content": "C = c(A) \\| c(A') \\| c(B)" + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\| \\cdot" + }, + { + "bbox": [ + 304, + 289, + 545, + 372 + ], + "type": "text", + "content": " denotes token-wise concatenation. This approach, however, suffers from three drawbacks: Token Entanglement, Semantic Bias, and Detail Erosion. We discuss each of these issues briefly, along with our solutions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 373, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 445 + ], + "type": "text", + "content": "Detail Erosion: Despite using patch-wise tokens, the extracted features lack the fine-grained information needed to retain key aspects of " + }, + { + "bbox": [ + 304, + 373, + 545, + 445 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 373, + 545, + 445 + ], + "type": "text", + "content": " in the generated pattern " + }, + { + "bbox": [ + 304, + 373, + 545, + 445 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 304, + 373, + 545, + 445 + ], + "type": "text", + "content": ". Consequently, the model often struggles to preserve elements like tile textures. To address this problem, we combine features from both the first and last layers of the feature encoder:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 455, + 545, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 455, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 331, + 455, + 545, + 468 + ], + "type": "interline_equation", + "content": "C _ {h l} (P) = \\operatorname {L i n e a r} \\left(L N \\left(c _ {\\text {h i g h}} (P)\\right) \\cdot L N \\left(c _ {\\text {l o w}} (P)\\right), \\right. \\tag {2}", + "image_path": "0980511f83ce605a17eabf1decc8ad05c53d3e431b556172c379f57a6142c625.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 477, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 477, + 545, + 513 + ], + "type": "inline_equation", + "content": "LN" + }, + { + "bbox": [ + 304, + 477, + 545, + 513 + ], + "type": "text", + "content": " is layer normalization, " + }, + { + "bbox": [ + 304, + 477, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 304, + 477, + 545, + 513 + ], + "type": "text", + "content": " denotes channel-wise concatenation, " + }, + { + "bbox": [ + 304, + 477, + 545, + 513 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 477, + 545, + 513 + ], + "type": "text", + "content": " is an input pattern, and Linear is a linear projection layer that fuses low- and high-level features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 514, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 656 + ], + "type": "text", + "content": "Semantic Bias: Image variation models typically use feature extractors such as CLIP [39], which are trained to align image embeddings with corresponding text embeddings. Such embeddings emphasize high-level semantics but lack spatial and fine-grained visual details. Combining these embeddings with features from text-free, self-supervised extractors, such as DiNO [2], has been shown to improve performance in various downstream tasks [20, 49]. For our task, a similar approach—combining features from both text-image " + }, + { + "bbox": [ + 304, + 514, + 545, + 656 + ], + "type": "inline_equation", + "content": "(m_{1})" + }, + { + "bbox": [ + 304, + 514, + 545, + 656 + ], + "type": "text", + "content": " and self-supervised " + }, + { + "bbox": [ + 304, + 514, + 545, + 656 + ], + "type": "inline_equation", + "content": "(m_{2})" + }, + { + "bbox": [ + 304, + 514, + 545, + 656 + ], + "type": "text", + "content": " feature extractors—significantly enhances generation quality. The extracted features are fused as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 345, + 666, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 666, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 345, + 666, + 545, + 681 + ], + "type": "interline_equation", + "content": "C _ {m i x} (P) = \\operatorname {M i x e r} \\left(C _ {h l} ^ {m _ {1}} (P) \\cdot C _ {h l} ^ {m _ {2}} (P)\\right), \\tag {3}", + "image_path": "1e7bd71d9dc32d94cee6bf3b385586f91aaa672aa57c4d45a190535d67a4c6eb.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\text{Mixer}" + }, + { + "bbox": [ + 304, + 689, + 545, + 712 + ], + "type": "text", + "content": " is a two-layer MLP that integrates features from the two extractors." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28719" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "Token Entanglement: To successfully perform an analogical edit, for each patch-level feature token, the model must be able to identify to which source image " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "(A, A', \\text{or } B)" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " that patch belongs as well as the 2D position of the patch within that image. Without these distinctions, the model often fails to identify the pattern to edit (i.e., " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ") and to recognize the desired edit from " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "(A, A')" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ". To address this problem, we introduce 3D positional encodings: two dimensions for spatial location within each pattern and one dimension for the source image. These encodings are applied to the extracted embeddings, yielding:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 214, + 287, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 214, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 287, + 228 + ], + "type": "interline_equation", + "content": "C ^ {\\Omega} = C _ {P E} (A) \\parallel C _ {P E} \\left(A ^ {\\prime}\\right) \\parallel C _ {P E} (B), \\tag {4}", + "image_path": "8dbbce0e4308f59e5416c1cad51c24b8066390c8dc843860c43dc85797991331.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 79, + 230, + 287, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 230, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 79, + 230, + 287, + 243 + ], + "type": "interline_equation", + "content": "C _ {P E} (P) ^ {x y} = C _ {m i x} (P) ^ {x y} + P E \\left(t _ {P}, x, y\\right), \\tag {5}", + "image_path": "65c54b31415e91fa2ab0d27e42f65c7c6f6dbd133960195111c50581e6c0cac7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 255, + 287, + 289 + ], + "type": "inline_equation", + "content": "t_P" + }, + { + "bbox": [ + 46, + 255, + 287, + 289 + ], + "type": "text", + "content": " is a one-hot vector encoding which input image a token comes from and " + }, + { + "bbox": [ + 46, + 255, + 287, + 289 + ], + "type": "inline_equation", + "content": "PE(t_P, x, y)" + }, + { + "bbox": [ + 46, + 255, + 287, + 289 + ], + "type": "text", + "content": " positionally encodes both spatial and source information for each token." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 291, + 289, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 411 + ], + "type": "text", + "content": "As we demonstrate in Section 4.4, conditioning on " + }, + { + "bbox": [ + 46, + 291, + 289, + 411 + ], + "type": "inline_equation", + "content": "C^\\Omega" + }, + { + "bbox": [ + 46, + 291, + 289, + 411 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 46, + 291, + 289, + 411 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 291, + 289, + 411 + ], + "type": "text", + "content": " significantly enhances the quality of patterns generated by TRIFUSER. Our adapted architecture, shown in Figure 5, integrates the modifications described above to effectively address the described drawbacks. Note that this architecture is designed for general analogical editing, not just pattern editing. To enhance generalizability to real-world patterns, we initialize TRIFUSER with an existing pretrained IM model [54], and fine-tune only the denoising UNet and the projection layers in our feature extractor." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 424, + 124, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 124, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 124, + 437 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 445, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 589 + ], + "type": "text", + "content": "In this section, we evaluate our approach along three directions: (1) the effectiveness of TRIFUSER at performing analogical edits on complex, real-world patterns, emphasizing how our synthetic data enables editing of in-the-wild pattern images; (2) the ability of TRIFUSER to support programmatic, structure-preserving edits without explicit program inference; and (3) the impact of architectural modifications introduced in TRIFUSER on the quality of generated patterns. We conduct a human perceptual study, quantitative assessments, and qualitative comparisons to demonstrate our system's ability to perform high-quality analogical edits across a range of pattern types." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 598, + 160, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 160, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 160, + 611 + ], + "type": "text", + "content": "4.1. Experiment Design" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "Datasets: We generate a large synthetic dataset of analogical quartets, i.e., pairs of analogical patterns " + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "inline_equation", + "content": "(A, A', B, B')" + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": ", using the SPLITWEAVE program samplers introduced in Section 3.2. This synthetic dataset contains approximately 1 million samples covering two pattern styles, namely Split Filling Patterns (SFP) and Motif Tiling Patterns (MTP) (cf. Section 3.1). For MTP patterns, we synthesize " + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "inline_equation", + "content": "100k" + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": " distinct tiles using the LayerDiffuse [58] model, guided by text" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "prompts derived from WordNet [35] noun synsets. Additionally, we construct a synthetic test set with 1000 analogical quartets to evaluate model performance on unseen synthetic data. Further details on dataset construction are provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 133, + 546, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 276 + ], + "type": "text", + "content": "To assess TRIFUSER on real-world patterns, we curate a test dataset of 50 patterns created by professional artists and sourced from Adobe Stock. This dataset spans seven distinct sub-domains of 2D patterns, representing a range of pattern styles. These styles include MTP and SFP patterns as well as previously unseen pattern styles such as Memphis-style, geometric, and digital textile patterns. Each pattern is annotated with a desired edit, and we use SPLITWEAVE to generate a pair of simpler patterns " + }, + { + "bbox": [ + 304, + 133, + 546, + 276 + ], + "type": "inline_equation", + "content": "(A, A')" + }, + { + "bbox": [ + 304, + 133, + 546, + 276 + ], + "type": "text", + "content": " demonstrating this edit. This test set provides a challenging benchmark to evaluate TRIFUSER's generalization to diverse, real-world editing tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 277, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 397 + ], + "type": "text", + "content": "Training details: We fine-tune a pre-trained diffusion model using our synthetic dataset of analogical quartets (cf. Section 3.3). We initialize our model with Versatile-Diffusion's Image Variation model [54]. We use SigLIP [57] as our text-image feature encoder and DiNOv2 [37] for self-supervised features. We fine-tune the model on 8 A100 GPUs using a batch size of 224 for " + }, + { + "bbox": [ + 304, + 277, + 547, + 397 + ], + "type": "inline_equation", + "content": "\\sim 65" + }, + { + "bbox": [ + 304, + 277, + 547, + 397 + ], + "type": "text", + "content": " epochs over 7 days. During inference, " + }, + { + "bbox": [ + 304, + 277, + 547, + 397 + ], + "type": "inline_equation", + "content": "B^{\\prime}" + }, + { + "bbox": [ + 304, + 277, + 547, + 397 + ], + "type": "text", + "content": " is inferred with typical diffusion parameter settings such as a classifier-free guidance weight of 7.5 and 50 denoising steps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 406, + 463, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 406, + 463, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 406, + 463, + 419 + ], + "type": "text", + "content": "4.2. Analogical Editing Baselines" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 425, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 460 + ], + "type": "text", + "content": "To evaluate the analogical editing capability of TRIFUSER, we compare it to three baseline methods, each representing a leading approach for analogical image editing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "text", + "content": "First, we consider training-free editors and latent arithmetic editors. Training-free editors repurpose pre-trained diffusion models to perform analogical edits without additional training [13, 51], leveraging the rich representations learned by diffusion models for editing. In this category, we compare against Analogist [13], the current state-of-the-art method. Latent arithmetic editors, on the other hand, rely on transformations in a learned latent space to infer analogical modifications [40, 48]. Note that these approaches only require samples from the target domain, not analogical training pairs. We implement a baseline for this method by fine-tuning a naive Image Variation model [54] on our synthetic dataset to learn a generative latent embedding space of patterns. At inference, analogical edits are generated using latent arithmetic: given patterns " + }, + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "text", + "content": ", we condition the generation of " + }, + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "inline_equation", + "content": "E(B) + E(A') - E(A)" + }, + { + "bbox": [ + 304, + 462, + 546, + 664 + ], + "type": "text", + "content": ". We refer to this baseline as LatentMod." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Finally, we consider analogy-conditioned generative editors, where models are explicitly trained on analogical data to learn analogical transformations [47]. This category includes our proposed TRIFUSER as well. Image Brush," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28720" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 248 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 248 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 248 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 248 + ], + "type": "image", + "image_path": "dadfce37205e3ab655e8f97912c11b153cf00a5e9bb36ef31fce493307676bde.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 260, + 287, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 287, + 293 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison between patterns generated by our model, TRIFUSER, and the baselines. TRIFUSER generates higher quality patterns with greater fidelity to the input analogy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 296, + 264, + 357 + ], + "blocks": [ + { + "bbox": [ + 70, + 296, + 264, + 357 + ], + "lines": [ + { + "bbox": [ + 70, + 296, + 264, + 357 + ], + "spans": [ + { + "bbox": [ + 70, + 296, + 264, + 357 + ], + "type": "table", + "html": "
Preference Rate
TRIFUSER vs. Analogist89.24%
TRIFUSER vs. LatentMod80.59%
TRIFUSER vs. Inpainter72.05%
", + "image_path": "fcb1a5b85a5ce01412329e4e77f5656e3f25dede10773b21f0fe4b51a22c053f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 287, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 399 + ], + "type": "text", + "content": "Table 1. Results of a two-alternative forced-choice perceptual study comparing our model (TRIFUSER) against three baselines. Ours is preferred in the overwhelming majority of judgments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 411, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 506 + ], + "type": "text", + "content": "the state-of-the-art method, fine-tunes a diffusion inpainting model for analogical editing with multi-modal conditioning. Since code for Image Brush is unavailable, we implement a similar baseline by fine-tuning a Stable Diffusion inpainting model. This model, which we term Inpainter, performs analogical editing by inpainting the lower-left quadrant of a 2x2 analogy grid containing " + }, + { + "bbox": [ + 46, + 411, + 287, + 506 + ], + "type": "inline_equation", + "content": "(A, A', B)" + }, + { + "bbox": [ + 46, + 411, + 287, + 506 + ], + "type": "text", + "content": " and conditioned on a fixed text template." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 514, + 204, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 204, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 204, + 528 + ], + "type": "text", + "content": "4.3. Editing Real-World Patterns" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 569 + ], + "type": "text", + "content": "To evaluate TRIFUSER's real-world analogical editing capabilities, we conducted a human preference study on the curated test set of Adobe Stock patterns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": "We performed a two-alternative forced-choice perceptual study comparing TRIFUSER with baseline methods on all 50 entries in the test set. Each method generates " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "k = 9" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": " outputs for each input tuple, and we select the best one based on visual inspection. Participants were shown edited patterns generated by two different methods along with the input patterns " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "(A, A', B)" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": " and instructed to select the edit that best preserved the analogical relationship and exhibited higher image quality. We recruited 42 participants for the study, resulting in a total of 1720 total judgments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Table 1 presents the results, showing that TRIFUSER was preferred over all three baselines. Due to the domain gap" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 544, + 259 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 544, + 259 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 544, + 259 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 544, + 259 + ], + "type": "image", + "image_path": "ac7c32ef2eda801bc4abff8e6c77245fbaa55a6678fb9bd1d1330fc83abc2c27.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 271, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 305, + 271, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 545, + 304 + ], + "type": "text", + "content": "Figure 7. TRIFUSER effectively edits patterns from novel pattern styles not present in the training dataset. TRIFUSER shows a noteworthy ability to generalize beyond its training distribution." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 315, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 445 + ], + "type": "text", + "content": "between the training data of the underlying model [43] and pattern images, Analogist fails to interpret and edit pattern images. Meanwhile, LatentMod fails to perform reasonable edits as the embedding space lacks the low-level details necessary for programmatic edits. While these baselines perform adequately on stylistic edits, they are unsuitable for programmatic editing. When compared to Inpainter, TRI-FUSER was favored in " + }, + { + "bbox": [ + 304, + 315, + 545, + 445 + ], + "type": "inline_equation", + "content": "72.2\\%" + }, + { + "bbox": [ + 304, + 315, + 545, + 445 + ], + "type": "text", + "content": " of comparisons. Both methods benefit from training on analogical quartets, yet Inpainter sacrifices pattern quality as it generates the edited pattern in only a quarter of the full canvas resolution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 446, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 530 + ], + "type": "text", + "content": "Figure 6 shows examples of pattern edits generated by TRIFUSER and the baselines, with our model consistently delivering superior results. In Figure 7, we show examples of TRIFUSER's edits on out-of-distribution pattern styles not present in the training set. These results suggest that our synthetic training data enables manipulation of real-world patterns, even extending to certain untrained pattern styles." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 538, + 452, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 452, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 452, + 551 + ], + "type": "text", + "content": "4.4. Editing Synthetic Patterns" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": "Next, we evaluate TRIFUSER's ability to perform programmatic edits on the synthetic validation set, which contains ground truth patterns " + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": ". Ideally, this would involve verifying that the underlying program " + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "inline_equation", + "content": "z_{\\hat{B}'}" + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": " of the generated pattern reflects the same transformation from " + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "inline_equation", + "content": "z_B" + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": " as that between " + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "inline_equation", + "content": "z_A" + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "inline_equation", + "content": "z_A'" + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": ". However, this would require visual program inference on " + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\hat{B}'" + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": ", which is infeasible. Instead, we approximate this criterion by comparing the program outputs " + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\hat{B}'" + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 304, + 556, + 545, + 700 + ], + "type": "text", + "content": " to see if the visual results align with the intended transformation. To quantify this alignment, we use perceptual metrics—DSim [8], DIST [4] and LPIPS [59]—along with SSIM to capture pixel-level structural similarity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Note that analogies can have multiple valid interpreta" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28721" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 284, + 144 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 284, + 144 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 284, + 144 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 284, + 144 + ], + "type": "table", + "html": "
DSIM (↓)DISTS (↓)LPIPS (↓)SSIM (↑)
Analogist0.4960.4320.6970.494
LatentMod0.2420.3200.6130.502
Inpainter0.0920.2560.3710.713
TRIFUSER0.0740.1840.3040.704
", + "image_path": "0fb0af6ad33272bd0405ec9dd8fbdcce2e89e8f7db82239553ef0559cf4eb46d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 186, + 284, + 275 + ], + "blocks": [ + { + "bbox": [ + 47, + 150, + 287, + 183 + ], + "lines": [ + { + "bbox": [ + 47, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 47, + 150, + 287, + 183 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation on the synthetic validation set shows that TRIFUSER generates patterns with higher perceptual similarity to the ground truth than the baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 186, + 284, + 275 + ], + "lines": [ + { + "bbox": [ + 51, + 186, + 284, + 275 + ], + "spans": [ + { + "bbox": [ + 51, + 186, + 284, + 275 + ], + "type": "table", + "html": "
DSIM (↓)DISTS (↓)LPIPS (↓)SSIM (↑)
TRIFUSER0.0740.1840.3040.704
- Pos. Enc.0.1470.2390.3830.659
- Lower0.0870.1960.3350.652
- Mix0.0980.2100.3450.682
Base [54]0.5850.4600.8150.435
", + "image_path": "6e2cae2d974e1d2c2d3d9383921f0c1095318eea28afb9699bc25653903514cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 281, + 287, + 315 + ], + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 315 + ], + "type": "text", + "content": "Table 3. Subtractive ablation study on TRIFUSER shows that removing any component (see Section 3.3) degrades performance, and that removing all components (Base) results in a sharp decline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 328, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 287, + 400 + ], + "type": "text", + "content": "tions, and even a single interpretation may yield several visually-related variations. To account for this multiplicity, we generate " + }, + { + "bbox": [ + 47, + 328, + 287, + 400 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 47, + 328, + 287, + 400 + ], + "type": "text", + "content": " output patterns for each input set " + }, + { + "bbox": [ + 47, + 328, + 287, + 400 + ], + "type": "inline_equation", + "content": "(A,A^{\\prime},B)" + }, + { + "bbox": [ + 47, + 328, + 287, + 400 + ], + "type": "text", + "content": " and select the one that maximizes each metric. In other words, we evaluate whether at least one generated output aligns with the intended target." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 521 + ], + "type": "text", + "content": "Table 2 shows the results of this experiment. First, we note that TRIFUSER outperforms others across all perceptual metrics. These metrics capture different aspects of perceptual similarity and superior performance across all of them suggests a comprehensive improvement. Second, we observe that the analogy-conditioned generative editors (Inpainter & TRIFUSER) surpass both the training-free and latent modification editors. Interestingly, Inpainter achieves slightly higher SSIM scores than TRIFUSER, suggesting that integrating the two approaches could be beneficial." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 527, + 160, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 160, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 160, + 539 + ], + "type": "text", + "content": "4.5. TRIFUSER Ablation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "To evaluate the impact of each model component introduced in Section 3.3, we conduct a subtractive analysis on the synthetic validation set. For this ablation study, we remove each component one at a time and measure the resulting performance, as reported in Table 3. The results demonstrate that removing any component leads to a performance drop, with the removal of 3D positional encoding leading to severe degradation. This is understandable: without 3D positional encoding, the network often fails to accurately identify which pattern to edit. For comparison, we also include results from the original Image Variation model [54] trained without any modifications (Base). As expected, this model performs poorly, underscoring the importance of our modifications in achieving high-quality analogical edits." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 387, + 148 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 387, + 148 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 387, + 148 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 387, + 148 + ], + "type": "image", + "image_path": "fd914bb398cbb9513f0cfc65c9f944a86415e8ca5ec02facf4a13f538bdaf2aa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 159, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 305, + 159, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 159, + 545, + 182 + ], + "type": "text", + "content": "Figure 8. Our model helps users mix elements of different real-world patterns together, accelerating design exploration." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 389, + 72, + 463, + 148 + ], + "blocks": [ + { + "bbox": [ + 389, + 72, + 463, + 148 + ], + "lines": [ + { + "bbox": [ + 389, + 72, + 463, + 148 + ], + "spans": [ + { + "bbox": [ + 389, + 72, + 463, + 148 + ], + "type": "image", + "image_path": "12acb91595b8d52a5af3c4b292685fb67ea1d037271e448d46ab21fa92a96126.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 466, + 72, + 543, + 148 + ], + "blocks": [ + { + "bbox": [ + 466, + 72, + 543, + 148 + ], + "lines": [ + { + "bbox": [ + 466, + 72, + 543, + 148 + ], + "spans": [ + { + "bbox": [ + 466, + 72, + 543, + 148 + ], + "type": "image", + "image_path": "e5ee99dfa369a5563cedcc90b05c09a90532b549b91cede03b12bf456e588955.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 190, + 381, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 381, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 381, + 205 + ], + "type": "text", + "content": "5. Application" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 212, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 212, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 545, + 247 + ], + "type": "text", + "content": "The ability to edit patterns without requiring program inference unlocks new creative possibilities. We demonstrate two practical applications of analogical pattern editing:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "spans": [ + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "text", + "content": "Pattern Mixing: Figure 8 shows example of using our method to mix elements of two real-world patterns " + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "text", + "content": ", allowing the user to create unique, hybrid designs. The Mix operator is implemented by using a synthetic pair " + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "inline_equation", + "content": "(A, A')" + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "text", + "content": " to create a variant " + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "text", + "content": " and then using the pair " + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "inline_equation", + "content": "(X, X')" + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "text", + "content": " to specify an edit to " + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "inline_equation", + "content": "\\text{Mix}(X, Y) = f(X, X', Y)" + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "inline_equation", + "content": "X' = f(A, A', X)" + }, + { + "bbox": [ + 305, + 248, + 546, + 343 + ], + "type": "text", + "content": ". See the supplementary material for more details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 344, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 344, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 344, + 545, + 417 + ], + "type": "text", + "content": "Animation Transfer: TRIFUSER can also be used to create animated sequences of edited patterns. By leveraging parametric SPLITWEAVE programs, users can generate animations for simple patterns and then apply these animations to complex patterns with no additional effort. See the video in the supplementary material for examples." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 378, + 441 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "content": "We introduced a novel approach for programmatic editing of visual patterns without inferring the underlying program. By using analogies to express desired edits and a learned conditional generative model to execute them, our method provides an intuitive solution for pattern manipulation. A key component of our approach is SPLITWEAVE, a domain-specific language for generating diverse, structured pattern data. Paired with our procedure for sampling analogical quartets, SPLITWEAVE enables the creation of a large, high-quality dataset for training. We also presented TRIFUSER, a Latent Diffusion Model (LDM) designed to overcome critical issues that emerge when LDMs are naively deployed for analogical pattern editing, enabling high-fidelity edits that capture user intentions. Our experiments demonstrate that TRIFUSER successfully edits real-world patterns and surpasses baseline methods, while also generalizing to novel pattern styles beyond its training distribution. We believe that our DSL, dataset, and model will help drive further research on in-the-wild pattern image editing. Looking forward, we aim to extend this analogical editing framework to other domains such as semi-parametric 3D modeling while continuing to improve synthetic data generation." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28722" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "type": "text", + "content": "[1] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei A. Efros. Visual prompting via image inpainting. arXiv preprint arXiv:2209.00647, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 178 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 180, + 286, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 180, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 54, + 180, + 286, + 246 + ], + "type": "text", + "content": "[3] Ta-Ying Cheng, Matheus Gadelha, Thibault Groueix, Matthew Fisher, Radomir Mech, Andrew Markham, and Niki Trigoni. Learning continuous 3d words for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6753-6762, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 247, + 286, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 286, + 278 + ], + "type": "text", + "content": "[4] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P. Simoncelli. Image quality assessment: Unifying structure and texture similarity. CoRR, abs/2004.07728, 2020. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 286, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 286, + 323 + ], + "type": "text", + "content": "[5] Kevin Ellis, Daniel Ritchie, Armando Solar-Lezama, and Josh Tenenbaum. Learning to infer graphics programs from hand-drawn images. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 286, + 379 + ], + "type": "text", + "content": "[6] Thomas G. Evans. A heuristic program to solve geometric-analogy problems. In Proceedings of the April 21-23, 1964, Spring Joint Computer Conference, page 327-338, New York, NY, USA, 1964. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 380, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 380, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 286, + 423 + ], + "type": "text", + "content": "[7] Brian Falkenhainer, Kenneth D. Forbus, and Dedre Gentner. The structure-mapping engine. In Proceedings of the Fifth AAAI National Conference on Artificial Intelligence, page 272-277. AAAI Press, 1986. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 425, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 286, + 479 + ], + "type": "text", + "content": "[8] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 480, + 286, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 480, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 54, + 480, + 286, + 523 + ], + "type": "text", + "content": "[9] Aditya Ganeshan, R. Kenny Jones, and Daniel Ritchie. Improving unsupervised visual program inference with code rewriting families. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "text", + "content": "[10] Aditya Ganeshan, Ryan Y. Huang, Xianghao Xu, R. Kenny Jones, and Daniel Ritchie. Parsel: Parameterized shape editing with language, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 601 + ], + "type": "text", + "content": "[11] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Advances in Neural Information Processing Systems, pages 5885–5897. Curran Associates, Inc., 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 624 + ], + "type": "text", + "content": "[12] Dedre Gentner. Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7(2):155-170, 1983. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 668 + ], + "type": "text", + "content": "[13] Zheng GU, Shiyuan Yang, Jing Liao, Jing Huo, and Yang Gao. Analogist: Out-of-the-box visual in-context learning with image diffusion model. ACM Transactions on Graphics (TOG), 2024. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[14] P. Guehl, R. Allegre, J.-M. Dischler, B. Benes, and E. Galin. Semi-procedural textures using point process texture basis functions. Computer Graphics Forum, 39(4):159-171, 2020. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "text", + "content": "[15] Julia Guerrero-Viu, Milos Hasan, Arthur Roullier, Midhun Harikumar, Yiwei Hu, Paul Guerrero, Diego Gutierrez, Bellen Masia, and Valentin Deschaintre. Texsliders: Diffusion-based texture editing in clip space. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 141, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 183 + ], + "type": "text", + "content": "[16] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image Analogies. Association for Computing Machinery, New York, NY, USA, 1 edition, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 262 + ], + "type": "text", + "content": "[18] Douglas Hofstadter and Melanie Mitchell. The Copycat project: a model of mental fluidity and analogy-making, page 205-267. Basic Books, Inc., USA, 1995. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[19] Yiwei Hu, Paul Guerrero, Milos Hasan, Holly Rushmeier, and Valentin Deschaintre. Node graph optimization using differentiable proxies. In ACM SIGGRAPH Conference Proceedings, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 352 + ], + "type": "text", + "content": "[20] Dongsheng Jiang, Yuchen Liu, Songlin Liu, XIAOPENG ZHANG, Jin Li, Hongkai Xiong, and Qi Tian. From CLIP to DINO: Visual encoders shout in multi-modal large language models, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 354, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 409 + ], + "type": "text", + "content": "[21] R. Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapeassembly: Learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics (TOG), Siggraph Asia 2020, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "text", + "content": "[22] R. Kenny Jones, David Charatan, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapemod: Macro operation discovery for 3d shape programs. ACM Transactions on Graphics (TOG), Siggraph 2021, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "text", + "content": "[23] R. Kenny Jones, Homer Walke, and Daniel Ritchie. Plad: Learning to infer shape programs with pseudo-labels and approximate distributions. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 500, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 544 + ], + "type": "text", + "content": "[24] R. Kenny Jones, Paul Guerrero, Niloy J. Mitra, and Daniel Ritchie. Shapecoder: Discovering abstractions for visual programs from unstructured primitives. ACM Transactions on Graphics (TOG), Siggraph 2023, 42(4), 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 589 + ], + "type": "text", + "content": "[25] R. Kenny Jones, Renhao Zhang, Aditya Ganeshan, and Daniel Ritchie. Learning to edit visual programs with self-supervision. In Advances in Neural Information Processing Systems, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 590, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 655 + ], + "type": "text", + "content": "[26] Milin Kodnongbua, Benjamin Jones, Maaz Bin Safeer Ahmad, Vladimir Kim, and Adriana Schulz. Reparamcad: Zero-shot cad re-parameterization for interactive manipulation. In SIGGRAPH Asia 2023 Conference Papers, New York, NY, USA, 2023. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[27] Beichen Li, Liang Shi, and Wojciech Matusik. End-to-end procedural material capture with proxy-free mixed-integer optimization. ACM Trans. Graph., 42(4), 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[28] Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J. Mitra. Sketch2cad: Sequential cad modeling by sketching in" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28723" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "context. ACM Trans. Graph. (Proceedings of SIGGRAPH Asia 2020), 39(6):164:1-164:14, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[29] Wen-Ding Li, Keya Hu, Carter Larsen, Yuqing Wu, Simon Alford, Caleb Woo, Spencer M. Dunn, Hao Tang, Michelangelo Naim, Dat Nguyen, Wei-Long Zheng, Zenna Tavares, Yewen Pu, and Kevin Ellis. Combining induction and transduction for abstract reasoning, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[30] Hugo Loi, Thomas Hurtut, Romain Vergne, and Joelle Thollot. Programmable 2d arrangements for element texture design. ACM Trans. Graph., 36(4), 2017. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "text", + "content": "[31] Arman Maesumi, Dylan Hu, Krishi Saripalli, Vladimir Kim, Matthew Fisher, Soeren Pink, and Daniel Ritchie. One noise to rule them all: Learning a unified model of spatially-varying noise patterns. ACM Trans. Graph., 43(4), 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 286 + ], + "type": "text", + "content": "[32] Jiayuan Mao, Xiuming Zhang, Yikai Li, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Program-Guided Image Manipulators. In International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "text", + "content": "[33] L. McCarthy, C. Reas, and B. Fry. Getting Started with P5.js: Making Interactive Graphics in JavaScript and Processing. Maker Media, Incorporated, 2015. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 354 + ], + "type": "text", + "content": "[34] Zichong Meng, Changdi Yang, Jun Liu, Hao Tang, Pu Zhao, and Yanzhi Wang. Instructgie: Towards generalizable image editing. arXiv preprint arXiv:2403.05018, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 399 + ], + "type": "text", + "content": "[35] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994, 1994. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 400, + 216, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 216, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 216, + 411 + ], + "type": "text", + "content": "[36] OpenAI. Gpt-4 technical report, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 413, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 413, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 413, + 287, + 511 + ], + "type": "text", + "content": "[37] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 534 + ], + "type": "text", + "content": "[38] Ofek Pearl, Itai Lang, Yuhua Hu, Raymond A. Yeh, and Rana Hanocka. Geocode: Interpretable shape programs. 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 612 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[40] Scott Reed, Yi Zhang, Yuting Zhang, and Honglak Lee. Deep visual analogy-making. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, page 1252-1260, Cambridge, MA, USA, 2015. MIT Press. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[41] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, Haiyong Jiang, Zhongang Cai, Junzhe Zhang, Liang Pan, Mingyuan Zhang, Haiyu Zhao, and Shuai Yi. Csg-stump: A learning friendly csg-like representation for interpretable" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "shape parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 161 + ], + "type": "text", + "content": "[42] Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, and Junzhe Zhang. Extrudenet: Unsupervised inverse sketch-and-extrude fornbsp; shape parsing. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, page 482-498, Berlin, Heidelberg, 2022. Springer-Verlag. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 162, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 547, + 216 + ], + "type": "text", + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 546, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 546, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 546, + 271 + ], + "type": "text", + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 272, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 546, + 304 + ], + "type": "text", + "content": "[45] Christian Santoni and Fabio Pellacini. gtangle: a grammar for the procedural generation of tangle patterns. ACM Trans. Graph., 35(6), 2016. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 545, + 349 + ], + "type": "text", + "content": "[46] Liang Shi, Beichen Li, Miloš Hasan, Kalyan Sunkavalli, Tamy Boubekeur, Radomir Mech, and Wojciech Matusik. Match: differentiable material graphs for procedural material capture. ACM Trans. Graph., 39(6), 2020. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 350, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 425 + ], + "type": "text", + "content": "[47] Yasheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: learning visual in-context instructions for exemplar-based image manipulation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "text", + "content": "[48] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 460, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 491 + ], + "type": "text", + "content": "[49] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 545, + 525 + ], + "type": "text", + "content": "[50] Trieu Trinh, Yuhuai Tony Wu, Quoc Le, He He, and Thang Luong. Solving olympiad geometry without human demonstrations. Nature, 625:476-482, 2024. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 525, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 581 + ], + "type": "text", + "content": "[51] Adéla Šubrtová, Michal Lukáč, Jan Čech, David Futschik, Eli Shechtman, and Daniel Sykora. Diffusion image analogies. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "text", + "content": "[52] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6772-6782, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 658 + ], + "type": "text", + "content": "[53] Xianghao Xu, Wenzhe Peng, Chin-Yi Cheng, Karl D. D. Willis, and Daniel Ritchie. Inferring cad modeling sequences using zone graphs. In CVPR, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "text", + "content": "[54] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7754-7765, 2023. 2, 5, 6, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28724" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 286 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[55] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[56] Mehmet Ersin Yumer, Siddhartha Chaudhuri, Jessica K. Hodgins, and Levent Burak Kara. Semantic shape editing using deformation handles. ACM Trans. Graph., 34(4), 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "type": "text", + "content": "[57] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 252 + ], + "type": "text", + "content": "[58] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. ACM Trans. Graph., 43(4), 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 286 + ], + "type": "text", + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28725" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Pay Attention to the Foreground in Object-Centric Learning/b4bf29e9-ae1b-4c69-a186-15cc519fb707_content_list.json b/2025/Pay Attention to the Foreground in Object-Centric Learning/b4bf29e9-ae1b-4c69-a186-15cc519fb707_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c149981deae81f1211afe627e08771b43d3b1e7a --- /dev/null +++ b/2025/Pay Attention to the Foreground in Object-Centric Learning/b4bf29e9-ae1b-4c69-a186-15cc519fb707_content_list.json @@ -0,0 +1,1419 @@ +[ + { + "type": "text", + "text": "Pay Attention to the Foreground in Object-Centric Learning", + "text_level": 1, + "bbox": [ + 192, + 130, + 805, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pinzhuo Tian $^{1}$ , Shengjie Yang $^{1}$ , Hang Yu $^{1*}$ , Alex Kot $^{2}$", + "bbox": [ + 281, + 179, + 712, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Engineering and Science, Shanghai University, China", + "bbox": [ + 200, + 198, + 795, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore", + "bbox": [ + 120, + 215, + 875, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{pinzhuo, sjyang, yuhang}@shu.edu.cn, eackot@ntu.edu.sg", + "bbox": [ + 254, + 237, + 743, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 303, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The slot attention-based method is widely used in unsupervised object-centric learning, aiming to decompose scenes into interpretable objects and associate them with slots. However, complex backgrounds in the real images can disrupt the model's focus, leading it to excessively segment background stuff into different regions based on low-level information such as color or texture variations. As a result, the detailed segmentation of foreground objects, which requires shape or geometric information, is often neglected. To address this issue, we introduce a contrastive learning-based indicator designed to differentiate between foreground and background. Integrating this indicator into the existing slot attention-based method enables the model to focus more on segmenting foreground objects while minimizing background distractions. During the testing phase, we utilize a spectral clustering mechanism to refine the results based on the similarity between the slots. Experimental results show that incorporating our method with various state-of-the-art models significantly improves their performance on both simulated data and real-world datasets. Furthermore, multiple sets of ablation experiments confirm the effectiveness of each proposed component. The source code is available at https://github.com/sjyjs09/FG-BG_Indicator.", + "bbox": [ + 86, + 335, + 483, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 727, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In perceiving the surrounding environment, humans typically divide it into distinct objects, subsequently understanding the scene through the relationships and connections among these objects [20, 25]. Thus, enabling our model to decompose the visual scene into different objects", + "bbox": [ + 89, + 753, + 482, + 829 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a6d541043eb5956d76897096cf7b62ad91251e038a6c2a5b2a2f47e7b91adfb.jpg", + "image_caption": [ + "Figure 1. The overview of our method. The background stuff in the complex scenes can influence the process of slot binding. Variations in color or texture within the background may cause the slot to incorrectly segment the background into different regions while neglecting the foreground (shown in the above branch). In our method, we integrate information from our indicator to enhance the distinction between foreground and background areas and mitigate changes in background region, enabling the slot attention method to focus more effectively on foreground object segmentation (shown in the below branch)." + ], + "image_footnote": [], + "bbox": [ + 514, + 301, + 906, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is highly meaningful and lays the foundation for reasoning [24]. Unsupervised object-centric learning [12] is an approach designed to address this challenge, aiming to represent a scene as a composition of distinct objects using only visual information.", + "bbox": [ + 511, + 611, + 905, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "After significant development, the slot attention-based method [23] has become the mainstream framework in the field. This kind of method can be divided into two primary stages: the binding stage and the reconstruction stage. In the binding stage, an attention-based mechanism is used to assign different areas of the image to distinct slots. The reconstruction stage then involves rebuilding the input based on information derived from these slots, ensuring that each slot accurately represents an individual object within the image. These two stages enable the slot attention-based method to achieve impressive results in various areas, such as [27, 38].", + "bbox": [ + 511, + 688, + 906, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, current slot attention-based methods do not distinguish between foreground objects and background stuff. When applied to complex scenes, variations in color", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 112, + 839, + 241, + 851 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work is supported by projects of the National Natural Science Foundation of China (Grant No.62206166, No.62302287), Shanghai Sailing Program (Grant No.23YF1413000), and Shanghai Committee of Science and Technology (Grant No.23ZR1423500).", + "bbox": [ + 89, + 851, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30281", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "or texture within the background can confuse the slots. This misdirection often causes the slots to divide the background into different regions based on low-level visual cues. As a result, the slots struggle to bind to foreground objects, which require a more nuanced recognition of shape and geometric details.", + "bbox": [ + 89, + 90, + 480, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we propose an unsupervised indicator that provides crucial information to differentiate between foreground and background. After training, it can be incorporated into various slot attention methods without requiring retraining. Figure 1 shows an overview of our method. Our unsupervised indicator is based on a contrastive learning framework [15] and consists of three components: a pixel-level contrastive loss that ensures consistency in pixel predictions across different views to promote meaningful representation learning; a stuff-level contrastive loss that minimizes differences within the background, reducing misguidance caused by color and texture variations in the background during slot binding; and an entropy-based regularization that enhances the distinction between foreground and background elements.", + "bbox": [ + 89, + 184, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By incorporating information from this indicator, the slot attention-based method shifts its focus toward more accurate object segmentation while minimizing background distractions. However, due to the absence of supervised ground truth, the indicator cannot perfectly differentiate between foreground and background regions, which may introduce noise or incorrect labels. To mitigate the impact of this noise during training, we first train the original slot attention-based method for several epochs, then incorporate the foreground and background information from the indicator. During the testing phase, recognizing that the similarity between slots can reflect the similarity between image regions, we employ a spectral clustering merging mechanism [33] to refine the results and reduce oversegmentation. In our experiments, we integrate our method with two prominent object-centric learning approaches, DINOSAUR [28] and SPOT [17]. The results demonstrate that our approach improves their performance on both simulated and real-world datasets.", + "bbox": [ + 89, + 412, + 482, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are threefold:", + "bbox": [ + 109, + 702, + 403, + 715 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We consider the issue that variations in background information, like color or texture, can mislead the slot allocation process and lead to the neglect of foreground object segmentation, which requires detailed shape and boundary information.", + "- We propose an unsupervised foreground and background indicator to solve this problem and employ a spectral clustering merging mechanism that utilizes the similarity between slots to refine the results.", + "- Our experimental results demonstrate that integrating our method with various state-of-the-art models significantly enhances their performance both on simulated data and" + ], + "bbox": [ + 91, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "real-world datasets.", + "bbox": [ + 527, + 90, + 658, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 119, + 651, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object-centric learning. Object-centric learning aims to represent scenes as compositions of distinct objects. MONet [2] and IODINE [11] are early studies in this area. However, they rely on multiple encode-decode steps. Slot Attention [23] achieves efficient scene decomposition by replacing this process with a single encoding step that iterates attention. Although Slot Attention has gradually become a major paradigm in object-centric learning, this approach still faces challenges when applied to real-world scenes. To address this issue, some studies focus on improving the encoder, like DINOSAUR [28]. Other studies focus on designing more effective decoders for scene decomposition, like SLATE [29], SPOT [17], and [16, 36].", + "bbox": [ + 511, + 146, + 903, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive learning. Contrastive learning is a mainstream paradigm in self-supervised learning (SSL), aiming to learn consistent representations by contrasting positive and negative sample pairs without the need for additional annotations [4, 7, 8]. Classical contrastive learning methods treat the image as a whole part and have shown strong potential in learning transferable visual representations [6, 14, 30, 39]. However, real-world scenes often contain complex and diverse information, making it difficult for global features to represent all the information. To address this issue, some works leverage local crops [3, 31, 37], while others gradually move toward pixel-level contrastive learning. Pixel-level contrastive learning extends the task of instance discrimination from image-level feature vectors to feature maps [22, 26, 34], which naturally aids in learning dense representations.", + "bbox": [ + 511, + 343, + 906, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 599, + 604, + 614 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach consists of three key procedures: 1) the foreground and background indicator, which supplies additional information about foreground objects and background elements, 2) the fusion module, responsible for integrating information from the indicator into the existing slot attention method, 3) the spectral clustering module, which merges regions to refine prediction results in the test stage.", + "bbox": [ + 511, + 625, + 903, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Foreground and background indicator", + "text_level": 1, + "bbox": [ + 511, + 742, + 844, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This part aims to achieve two goals: 1) it needs to be an unsupervised method, and 2) it should provide an accurate and robust representation to distinguish the foreground and background.", + "bbox": [ + 511, + 763, + 903, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hence, the implementation of our indicator leverages advancements in the contrastive learning framework, like teacher-student framework. Specifically, our approach involves a student network, denoted by $\\theta$ , and a teacher network, denoted by $\\xi$ . Both networks share the same architec", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30282", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5a7f13e025d4f274a7e2c89ce4aac2acaffbafef8174c72c522e2abd1826b08a.jpg", + "image_caption": [ + "Figure 2. Illustration of our pipeline. During the training stage, the foreground and background indicator are trained to provide highly distinguishable foreground and background features (the part of (a)). Then, the feature from the indicator is integrated into the original slot attention-based methods (the part of (b)). In the testing stage, we use region combination to refine the results." + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 883, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ture but use different sets of weights. The parameters of the teacher network, $\\xi$ , are updated as an exponential moving average of the student network, $\\theta$ . In each network, we use two distinct slots, $S = \\{s^f, s^b\\}$ , i.e., $S_{\\theta}$ and $S_{\\xi}$ for the student and teacher networks, respectively, to help recognize the foreground and background in each image. $S_{\\xi}$ is also optimized by EMA.", + "bbox": [ + 89, + 645, + 482, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike the classical slot attention method, $s_{\\theta \\text{ or } \\xi}^{f}$ and $s_{\\theta \\text{ or } \\xi}^{b}$ are given semantic meaning, enabling the classification of pixels into their corresponding semantic groups, such as foreground and background. Here, $s_{\\theta}^{f}$ and $s_{\\theta}^{b}$ represent the slots for foreground and background in the student network, respectively, while $s_{\\xi}^{f}$ and $s_{\\xi}^{b}$ represent the slots for foreground and background in the teacher network, respectively. However, we need an initialized boot mode to provide the initialized semantic cues for foreground objects.", + "bbox": [ + 89, + 761, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-trained models have proven effective in providing initial semantic information across various fields [19]. Therefore, we employ the pre-trained model as the backbone of our method instead of training from scratch.", + "bbox": [ + 511, + 645, + 906, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our indicator contains three aspects. The first is the pixel-level consistency. For different augmentations of the same image, we ensure that pixels in the same location receive similar assignment scores. Given an input image $x$ , two random augmentations are applied to produce two augmented views $v^{1}(x)$ and $v^{2}(x)$ . Each augmentation view is then encoded with an encoder $f$ into a hidden feature map and then transformed with a multilayer perceptron (MLP) $g$ to get the feature map $z$ , i.e., $z_{\\theta}(x) = g_{\\theta}(f_{\\theta}(v^{1}(x)))$ and $z_{\\xi}(x) = g_{\\xi}(f_{\\xi}(v^{2}(x)))$ . The fore-background representation vectors of the image $x$ in different networks can be obtained via slot attention [23], i.e., $\\hat{S}_{\\theta}(x) =$", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30283", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "slot attention $(z_{\\theta}(x),S_{\\theta})$ $\\hat{S}_{\\xi} =$ slot attention $(z_{\\xi}(x),S_{\\xi})$ Then, the pixel assignment matrix, defining the score fore or background each pixel belongs to, $\\mathcal{P}_{\\theta}$ and $\\mathcal{Q}_{\\xi}$ can obtain by $\\mathcal{P}_{\\theta}(x) = z_{\\theta}\\cdot \\hat{S}_{\\theta}^{\\top}$ . The assignment matrix $\\mathcal{Q}_{\\xi}(x)$ for the teacher network can be obtained in the same way.", + "bbox": [ + 89, + 89, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To keep the pixel level consistency, we enforce the prediction of the two views assignment $\\mathcal{P}_{\\theta}(x)$ and $\\mathcal{Q}_{\\xi}(x)$ match each other. However, the geometric image augmentations, such as random crop, scale, or flip, between different views will cause the scale or layout of the two feature maps to be inconsistent. So similar to [35], we use the inverse augmentation process on the predicted assignments to align their spatial locations, i.e., $\\tilde{\\mathcal{P}}_{\\theta}(x) = \\mathrm{invaug}(\\mathcal{P}_{\\theta}(x)), \\tilde{\\mathcal{Q}}_{\\xi}(x) = \\mathrm{invaug}(\\mathcal{Q}_{\\xi}(x))$", + "bbox": [ + 89, + 169, + 480, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the aligned assignments, we apply the cross-entropy loss to enforce the consistency in assignment scores between spatial-aligned pixels from different views. The cross-entropy loss is averaged over all spatial locations to keep the pixel prediction consistency", + "bbox": [ + 89, + 305, + 480, + 380 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {p i x e l}} (x) = \\frac {1}{H \\times W} \\sum_ {i, j} \\mathcal {L} ^ {\\mathrm {C E}} \\left(\\tilde {\\mathcal {Q}} _ {\\xi} (x) [ i, j ], \\tilde {\\mathcal {P}} _ {\\theta} (x) [ i, j ]\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 391, + 482, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$i,j$ means the position in the assignment matrix.", + "bbox": [ + 89, + 457, + 410, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the pixel-level contrastive loss, we incorporate a stuff-level consistency loss. This loss aims to alleviate variations in the background by treating the background in different images as positive samples, while considering all foreground samples as negative samples. According to the pixel assignment matrix $\\tilde{\\mathcal{P}}_{\\theta}(x)$ and $\\tilde{\\mathcal{Q}}_{\\xi}(x)$ , we can obtain the prototypes of the foreground and background by region average pooling to represent the whole foreground and background regions of the image $x$ . $p_{\\theta}^{f}(x), p_{\\theta}^{b}(x)$ represent the foreground and background prototypes obtained from the student network. $p_{\\xi}^{f}(x), p_{\\xi}^{b}(x)$ are from the teacher network. Then, the stuff-level loss can be written as follows:", + "bbox": [ + 89, + 473, + 482, + 670 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} ^ {\\text {s t u f f}} (B) = \\frac {1}{B (B - 1)} \\sum_ {i = 1} ^ {B} \\sum_ {j \\neq i} ^ {B} [ \\operatorname {s i m} \\left(p _ {\\theta} ^ {b} \\left(x _ {i}\\right), p _ {\\xi} ^ {b} \\left(x _ {j}\\right)\\right) - \\\\ \\left. \\left. \\operatorname {s i m} \\left(p _ {\\theta} ^ {b} \\left(x _ {i}\\right), p _ {\\xi} ^ {f} \\left(x _ {j}\\right)\\right) \\right]. \\right. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 680, + 480, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq. 2, we hope the similarity between the background across different views will be large, however, the similarity between foreground and background should be small. We adopt the cosine similarity in our method. $B$ means the batch of training data.", + "bbox": [ + 89, + 779, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To encourage the semantic centers to emphasize different visual contents and cover diverse semantics to avoid collapse, we apply a regularization to the assignment matrix of", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "different views, as follows:", + "bbox": [ + 513, + 90, + 694, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\operatorname {s e p}} (x) = \\operatorname {E n t r o y} \\left(\\mathcal {P} _ {\\theta} (x)\\right) + \\operatorname {E n t r o p y} \\left(\\mathcal {Q} _ {\\xi} (x)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 114, + 903, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The total loss to optimize the proposed indicator is", + "bbox": [ + 532, + 141, + 867, + 156 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} ^ {\\text {p i x e l}} (x) - \\lambda \\mathcal {L} ^ {\\text {s t u f f}} (B) - \\gamma \\mathcal {L} ^ {\\text {s e p}} (x). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 165, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Fusion of information into original slot attention method", + "text_level": 1, + "bbox": [ + 511, + 191, + 903, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The features provided by our indicator allow for clear distinction between the foreground object and background elements. Simultaneously, the background is processed uniformly, harmonizing its characteristics across various regions. Integrating the features obtained through our method with existing approaches significantly enhances the differentiation between the foreground and background. Furthermore, once trained, our indicator is frozen and only provides supplementary information to the existing slot attention-based methods.", + "bbox": [ + 511, + 228, + 903, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our fusion stage, we first train the slot attention-based method for several epochs. Once it has learned meaningful information, we next integrate the foreground and background information from our indicator into the slot attention-based method as supplementary information. Finally, the features from both the original slot attention-based method and our indicator are fused together and input into the next decoder for image reconstruction. The feature fusion stage can be written as follows:", + "bbox": [ + 511, + 378, + 905, + 515 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz ^ {\\text {f u s i o n}} (x) = \\operatorname {M L P} \\left(z ^ {\\text {s l o t}} (x) + z _ {\\xi} ^ {\\text {i n d i c a t o r}} (x)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 523, + 903, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z^{\\mathrm{slot}}(x)$ is the feature map from the encoder in the original slot attention method. $z_{\\xi}^{\\mathrm{indicator}}(x)$ is the feature map provided by our indicator. $x$ is the input image. These two feature maps are added together and then passed through a MLP layer to transform the combined feature map into a new space. The resulting fusion feature $z^{\\mathrm{fusion}}(x)$ replaces the feature map $z^{\\mathrm{slot}}(x)$ used in the original slot attention-based method to do the next stages, such as slot attention, and decoding, which are the same as the original slot attention framework.", + "bbox": [ + 511, + 551, + 905, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Region Combination", + "text_level": 1, + "bbox": [ + 511, + 712, + 709, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Defining the number of slots is a common challenge in recent slot attention methods [10] and can significantly impact their performance. Determining the appropriate number of slots is not straightforward, and recent approaches often opt for a higher number of slots, leading to slight oversegmentation, in order to achieve satisfactory results. In this section, we propose a spectral clustering-based method to group regions likely belonging to the same semantic object, which is used to refine the predictions during the testing phase.", + "bbox": [ + 511, + 734, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The similarity between slots represents the similarity between regions in the image. The similarity matrix between", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30284", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "different slots can be regarded as a reference to help recognize which slots possibly belong to the same semantic object. In our method, we directly use spectral clustering to divide the slots into different groups based on this matrix. Specifically, we calculate the cosine similarity between each pair of the slots to construct the similarity matrix $\\mathbf{M}$ . Then according to $\\mathbf{M}$ , the group results can be obtained by", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = \\text {s p e c t r a l c l u s t e r i n g} (\\mathbf {M}, N). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 210, + 482, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the group result $\\mathcal{G}$ , the corresponding areas of slots in the same group can be combined. To determining the group number $N$ , we primarily employ eigen-gap heuristic [32]. In this method, we use the Laplacian matrix of the graph and find the largest gap between consecutive eigenvalues, which suggests the optimal number of clusters.", + "bbox": [ + 89, + 243, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 351, + 222, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 376, + 171, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on one synthetic and two real-world datasets. For the synthetic dataset, we use MOVi-C [13], which contains approximately 1,000 real 3D-scanned objects, with each scene featuring between 3 and 10 objects. Although MOVi-C is originally a video dataset, we convert it into an image dataset by randomly selecting frames following [28]. For the real-world datasets, we utilize MS COCO 2017 [21] and PASCAL VOC 2012 [9], which are commonly used for object detection and segmentation. PASCAL VOC 2012 contains many images with a single salient object, while MS COCO 2017 features more complex scenes, offering a more challenging evaluation.", + "bbox": [ + 89, + 398, + 482, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. To assess object-centric learning, we used Mean Best Overlap (mBO) and Mean Intersection over Union (mIoU). The Mean Best Overlap (mBO) includes instance-level $\\mathrm{mBO}^i$ and category-level $\\mathrm{mBO}^c$ . $\\mathrm{mBO}^i$ and $\\mathrm{mBO}^c$ match the predicted masks to the instance and semantic masks, respectively, by finding the best overlap. Meanwhile, mIoU employs the Hungarian algorithm to match predicted and ground truth masks. [17] suggests that FG-ARI, as a clustering similarity metric, may yield unreliable segmentation quality assessments by overlooking the accuracy of predicted masks. Therefore, FG-ARI is not reported in this paper.", + "bbox": [ + 89, + 582, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. To optimize the proposed foreground and background indicator, we employ the SGD optimizer. The learning rate is 0.001. Batch size is 512. The learning rate follows a linear warm-up for 10,000 optimization steps and an exponential decay schedule. Additionally, we clip the gradient norm to 1 to stabilize the training process. ViT-B/16 [1], initialized with DINO [5], is used as the encoder in our indicator. During the fusion stage, the model is trained using the same configuration as the corresponding", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "original slot attention method you choose. All models are trained on a single A6000 GPU with 48 Gbytes.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Integrating into SOTA object-centric methods", + "text_level": 1, + "bbox": [ + 511, + 132, + 901, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We apply our approach to two slot attention methods, DINOSAUR [28] and SPOT [17]. DINOSAUR utilizes a Vision Transformer encoder for feature reconstruction and has demonstrated applicability to real-world datasets. SPOT is a recent SOTA object-centric model that employs a two-stage self-training scheme and a patch-order permutation strategy for the transformer decoder. Furthermore, we also list the results of several other object-centric methods, including Slot Attention [23], SLASH [18], SLATE [29], SlotDiffusion [36], and LSD [16]. A more detailed description of the baseline methods can be found in the Appendix.", + "bbox": [ + 511, + 155, + 903, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Main Results on Each Dataset. Tab. 1 presents the main results of our method on the COCO, PASCAL, and MOViC datasets. Our method improves the performance of DINOSAUR and SPOT across all datasets. The fusion stage yields consistent improvements, while the region combination stage leads to a slight decrease in the $\\mathrm{mBO}^i$ metric on the PASCAL and COCO datasets. This is due to the fact that, during clustering, slots corresponding to two instances of the same category may be merged into one, resulting in an increase in $\\mathrm{mBO}^c$ but a decrease in $\\mathrm{mBO}^i$ . Despite this, our method still outperforms the baseline models, DINOSAUR and SPOT, demonstrating a clear overall performance advantage. This highlights the crucial role of the fusion stage in driving performance improvements, while the region combination stage primarily serves to refine the predictions.", + "bbox": [ + 511, + 337, + 903, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. More discussions", + "text_level": 1, + "bbox": [ + 511, + 590, + 679, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we analyze various aspects of our approach: the significance of the foreground and background indicators, the impact of different loss functions within the indicator, the role of the fusion stage, and the consistency of performance improvement across different slot numbers.", + "bbox": [ + 511, + 613, + 903, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(A) Does the extra foreground and background information really improve the slot attention method? We evaluate the slot attention method in four cases. Tab. 2 shows the results. In this table, \"Random\" refers to using random noise as the input to the fusion stage, \"ViT\" refers to directly using the pre-trained DINO ViT as the indicator, and \"Ours\" refers to our foreground and background indicator. We find that when the indicator is random noise, the baseline method struggles to distinguish between foreground and background, leading to a significant performance drop. When ViT is used as the indicator, the pretrained ViT has some inherent ability to differentiate between foreground and background, resulting in a slight performance improvement. Our designed foreground", + "bbox": [ + 511, + 689, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30285", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b681b2de751304b7fdea6fb7b20d20dd509f95ee1a8a8469d8677adcb6a0d0c4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMOVi-CPASCALCOCO
mBOimIoUmBOimBOcmBOimBOc
Slot Attention [23] [NeurIPS20]26.3-24.624.917.219.2
SLASH [18] [CVPR23]-27.7----
SLATE [29] [ICLR22]39.437.835.941.529.133.6
SlotDiffusion [36] [NeurIPS23]--50.455.331.035.0
(Stable-)LSD [16] [NeurIPS23]45.644.2--30.4-
DINOSAUR [28] [ICLR23]42.441.844.051.231.138.9
DINOSAUR + FS (Ours)45.3 ↑ 2.944.5 ↑ 2.745.4 ↑ 1.452.5 ↑ 1.331.5 ↑ 0.439.9 ↑ 1.0
DINOSAUR + FS + RC (Ours)46.0 ↑ 3.644.8 ↑ 3.045.2 ↑ 1.252.9 ↑ 1.731.2 ↑ 0.140.1 ↑ 1.2
SPOT [17] [CVPR24]47.046.448.155.334.744.3
SPOT + FS (Ours)48.5 ↑ 1.547.7 ↑ 1.349.1 ↑ 1.056.3 ↑ 1.035.5 ↑ 0.845.0 ↑ 0.7
SPOT + FS + RC (Ours)49.0 ↑ 2.047.8 ↑ 1.449.3 ↑ 1.256.5 ↑ 1.235.7 ↑ 1.045.3 ↑ 1.0
", + "bbox": [ + 112, + 88, + 885, + 296 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b19bae34777309140efdac5a00adcd6265278ac9f39ea6b5aac72d4a5dfb88cf.jpg", + "table_caption": [ + "Table 1. Comparison with different object-centric methods on MOVi-C, PASCAL, and COCO datasets. DINOSAUR uses an autoregressive decoder and DINO [5] ViT encoder. The experimental results of the baseline are derived from the original corresponding papers. FS means using the fusion stage in our paper. RC means the trick of region combination." + ], + "table_footnote": [], + "table_body": "
SA MethodIndicatorIoUmBOcmBOi
DINOSAUR--51.244.0
Random-25.222.1
ViT30.251.444.7
Ours52.652.545.4
", + "bbox": [ + 114, + 372, + 460, + 465 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/17aad365b3b6afd5fcca0663a7454e289db85e3b4a8f39b41a7f77330b46aed0.jpg", + "table_caption": [ + "Table 2. The effect of extra information on the existing slot attention-based method. All experiments are conducted on PASCAL dataset. IoU is used to show the model's performance on segmenting the foreground and background. $\\mathrm{mBO}^c$ and $\\mathrm{mBO}^i$ are employed to assess the final performance after fusing the extra knowledge from the indicator." + ], + "table_footnote": [], + "table_body": "
MethodMOVi-CPASCALCOCO
DINOSAUR (2 slots)45.247.646.9
SPOT (2 slots)46.650.846.4
Ours47.452.648.4
", + "bbox": [ + 114, + 575, + 460, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. The performance of using object-centric methods as the foreground and background indicator. Fig. 3 shows some visualization results.", + "bbox": [ + 88, + 659, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and background indicator makes the foreground and background highly separable and significantly improves performance.", + "bbox": [ + 89, + 732, + 482, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(B) How about using the classical slot attention methods as the foreground and background indicator in our method? Tab. 3 evaluates the performance of different models when the number of slots is set to 2. We continue to use the Mean Best Overlap (mBO) as the evaluation metric. The difference is that, during metric computation, we convert the ground truth masks into binary masks containing only foreground (0) and background (1). Then, the pre", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dicted masks are matched with the ground truth masks using the best overlap matching method.", + "bbox": [ + 511, + 375, + 906, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From the data in Tab. 3, it can be observed that our method achieves the best performance with significant performance improvements on both COCO and PASCAL datasets. This is one of the reasons for using the foreground and background indicator. Another reason is that autoencoder-based methods like DINOSAUR and SPOT, while capable of dividing the scene into two parts, struggle to achieve accurate foreground-background segmentation. As shown in Fig. 3, the baseline methods tend to divide the scene into two parts, top and bottom, while our method focuses on separating the foreground and background. This is because our method employs both pixel-level and stufflevel contrastive learning, enabling the model to better distinguish foreground from background features, leading to a more effective foreground-background decomposition.", + "bbox": [ + 511, + 409, + 908, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(C) Impact of different losses. We conduct the ablation experiments to further analyze the role of different losses in the indicator, as shown in Tab. 4. We find that without using $\\mathcal{L}^{\\mathrm{sep}}$ , the indicator struggle to achieve a performance improvement. This is because when only $\\mathcal{L}^{\\mathrm{pixel}}$ and $\\mathcal{L}^{\\mathrm{stuff}}$ are used, the model tends to collapse all information into a single slot, where one slot binds to the entire image and the other slot remains empty. In this case, the model can achieve low $\\mathcal{L}^{\\mathrm{pixel}}$ and $\\mathcal{L}^{\\mathrm{stuff}}$ values, but it cannot effectively perform the foreground-background separation. Therefore, we introduce $\\mathcal{L}^{\\mathrm{sep}}$ to encourage semantic centers to emphasize different visual content and cover diverse semantics. However, relying on $\\mathcal{L}^{\\mathrm{sep}}$ and $\\mathcal{L}^{\\mathrm{pixel}}$ makes it difficult for the model to adapt to variations in the background. As a result, $\\mathcal{L}^{\\mathrm{sep}}$ , $\\mathcal{L}^{\\mathrm{pixel}}$ , and $\\mathcal{L}^{\\mathrm{stuff}}$ must work together in synergy to achieve effective foreground-background separation.", + "bbox": [ + 511, + 638, + 908, + 882 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the same time, it can be seen that the performance of", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30286", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/740561010922650beec7cb8d634406b8ea0711dc76fba8e4c20c5511fa668675.jpg", + "image_caption": [ + "Figure 3. Example results of the proposed indicator and SOTA slot attention method with 2 slots. Our method distinguishes foreground and background better." + ], + "image_footnote": [], + "bbox": [ + 101, + 92, + 477, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the indicator is positively correlated with the final results. This further confirms the conclusion in (A): the stronger the indicator's ability to distinguish between foreground and background, the more effectively it enhances the feature fusion process.", + "bbox": [ + 89, + 390, + 483, + 465 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/061e0a5f30940a954afa74d44bce4313b85e5ca57401861a62fcef4f8ed38d72.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LpixelLstuffLsepIoUmBOcmBOi
✓28.851.444.6
✓✓28.451.344.7
✓✓48.951.744.8
✓✓✓52.652.545.4
", + "bbox": [ + 114, + 477, + 459, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Analysis of different loss functions on PASCAL dataset. IoU represents the results of the indicator on segmenting foreground and background regions. $\\mathrm{mBO}^c$ and $\\mathrm{mBO}^i$ represent the final results of DINOSAUR after fusing knowledge from different indicators, which are optimized using different types of loss functions.", + "bbox": [ + 89, + 589, + 482, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(D) The importance of the fusion stage and region combination stage. We conduct a detailed analysis of the importance of the fusion stage and the region combination stage. The results are summarized in Table 5. We observe that directly applying the region combination to DINOSAUR and SPOT typically leads to a decline in performance metrics. We hypothesize that this is due to the limited ability of the baseline methods to differentiate between foreground and background. Without a clear distinction, the region combination stage inappropriately merges foreground and background. The visualization results are shown in Figure 4.", + "bbox": [ + 89, + 689, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the MOVi dataset, applying region combination directly to SPOT results in a performance improvement. This", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e184b3c267ddaebe0499b2e2e14785a29e8c205b79f2f15209ad3fb439d1bc6.jpg", + "image_caption": [ + "Figure 4. The visualization results in our different components. FS means fusion stage, and RC is the stage of region combination." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 900, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "may result from the relatively simple background in the MOVi dataset, where SPOT is already effective at distinguishing between foreground and background. As shown in Tab. 3, the metric for SPOT with 2 slots closely aligns with our indicator result, further supporting this observation.", + "bbox": [ + 511, + 392, + 905, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(E) Consistent improvement across different slot numbers. The quality of results is influenced by the number of slots used during model training. Therefore, understanding the model's robustness to variations in the number of slots is crucial. In this part, we compare our method with DINOSAUR to evaluate the stability of our method with different numbers of slot. We conduct experiments on two real-world datasets, COCO and PASCAL, using five slot numbers, with the results presented in Tab. 6. Notably, our method outperforms DINOSAUR across all slot numbers and exhibits lower variance. This suggests that, while improving baseline performance, our method also demonstrates better generalization ability when applied with different numbers of slots.", + "bbox": [ + 511, + 469, + 906, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative results", + "text_level": 1, + "bbox": [ + 511, + 695, + 691, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 shows example results on the COCO and PASCAL datasets for DINOSAUR, SPOT, and our method. Row (a) demonstrates that, compared to baseline methods, our approach correctly fuses the foreground and background. For example, in the PASCAL dataset, our method successfully merges the ground that was originally divided into two separate slots. Rows (b) and (c) illustrate that when the baseline method accurately segments the background (foreground), our method correctly fuses the foreground (background) regions. For instance, in row (b) of the PASCAL dataset, when the background is correctly segmented, our method successfully merges the train and streetlight, which were", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30287", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/37333a7b52abfe8328fc7dc4e7f434f9dcf5e178004fae5ddd56912817e9b0a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMOVi-CPASCALCOCO
mBOimIoUmBOimBOcmBOimBOc
DINOSAUR42.441.844.051.231.138.9
DINOSAUR+RC41.6 ↓0.839.5 ↓2.343.3 ↓0.750.7 ↓0.530.4 ↓0.739.4 ↑0.5
DINOSAUR+FS (Ours)45.3 ↑2.944.5 ↑2.745.4 ↑1.452.5 ↑1.331.5 ↑0.439.9 ↑1.0
DINOSAUR+FS+RC (Ours)46.0 ↑3.644.8 ↑3.045.2 ↑1.252.9 ↑1.731.2 ↑0.140.1 ↑1.2
SPOT47.046.448.155.334.744.3
SPOT+RC47.8 ↑0.847.1 ↑0.748.0 ↓0.154.8 ↓0.533.2 ↓1.543.5 ↓0.8
SPOT+FS (Ours)48.5 ↑1.547.7 ↑1.349.1 ↑1.056.3 ↑1.035.5 ↑0.845.0 ↑0.7
SPOT+FS+RC (Ours)49.0 ↑2.047.8 ↑1.449.3 ↑1.256.5 ↑1.235.7 ↑1.045.3 ↑1.0
", + "bbox": [ + 112, + 88, + 885, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/70293c6a398e4134211456b21d6ede2a6e1795090f14afad9a26ea04f1b791d5.jpg", + "image_caption": [ + "Figure 5. Qualitative results on COCO and PASCAL datasets. Ours: fusion stage + region combination." + ], + "image_footnote": [], + "bbox": [ + 98, + 303, + 500, + 488 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17765ede36fb5b66311932a5510a2a0ebd09151a5bf6c3f1262dcf05e5cc0e6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 303, + 893, + 488 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a9b89756d213e5d2e1bcc923b4b22dec72e3ef074f570689902bc02a18d8079e.jpg", + "table_caption": [ + "Table 5. The importance of the different components in our method. FS is the fusion stage and RS is the region combination stage. Using RS alone even hurt the performance. FS is crucial." + ], + "table_footnote": [], + "table_body": "
DatasetKDINOSAURDINOSAUR+FS
mBOimBOcmBOimBOc
COCO227.620.530.2 ↑ 2.622.2 ↑ 1.7
537.929.339.3 ↑ 1.430.3 ↑ 1.0
739.031.139.9 ↑ 0.931.5 ↑ 0.4
938.531.739.8 ↑ 1.332.0 ↑ 0.3
1138.031.639.5 ↑ 1.532.6 ↑ 1.0
PASCAL234.340.435.4 ↑ 1.141.6 ↑ 1.2
444.351.644.6 ↑ 0.352.0 ↑ 0.4
644.851.845.4 ↑ 0.652.5 ↑ 0.7
845.351.745.5 ↑ 0.252.3 ↑ 0.6
1045.451.345.6 ↑ 0.252.0 ↑ 0.7
", + "bbox": [ + 96, + 539, + 480, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. The performance of our method with different numbers of slot. For a fair comparison, only the fusion stage is adopted. $K$ represents the number of slots.", + "bbox": [ + 89, + 757, + 485, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "originally assigned to two separate slots.", + "bbox": [ + 89, + 825, + 359, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Training cost", + "text_level": 1, + "bbox": [ + 89, + 848, + 228, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The additional training cost of our method primarily consists of two components. First, training the indicator. Due", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to its simple structure, training on the PASCAL dataset using a single A6000 GPU takes approximately 3 hours. Once trained, the indicator can be directly applied to different slot attention methods on the same dataset without the need for retraining. Second, retraining the existing slot attention methods based on the fused features, where the training time depends on the chosen slot attention method.", + "bbox": [ + 511, + 542, + 906, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we consider the problem that complex backgrounds in images can disrupt the phase of slot binding. Hence, we introduce a contrastive learning-based indicator to provide extra information about the foreground and background region to the existing slot attention method. Integrating this indicator into the slot attention-based method allows the model to focus more effectively on segmenting foreground objects and minimizing background distractions. According to the similarity between the slots, we also use a spectral clustering mechanism to refine the results in the testing phase. In experiments, we integrate our method into two SOTA slot attention methods. All the results show that our method can consistently enhance the performance on different datasets.", + "bbox": [ + 511, + 689, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30288", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.5", + "[2] Christopher P Burgess, Loic Matthew, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. 2", + "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2", + "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 6", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2", + "[9] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 5", + "[10] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23062-23071, 2024. 4", + "[11] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In International conference on machine learning, pages 2424-2433. PMLR, 2019. 2", + "[12] Klaus Greff, Sjoerd Van Steenkiste, and Jürgen Schmidhuber. On the binding problem in artificial neural networks. arXiv preprint arXiv:2012.05208, 2020. 1", + "[13] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapra" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 5", + "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[15] A Jaiswal, A Ramesh Babu, M Zaki Zadeh, D Banerjee, and F Makedon. A survey on contrastive self-supervised learning, 2020. 2", + "[16] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 8563-8601, 2023. 2, 5, 6", + "[17] Ioannis Kakogeorgiou, Spyros Gidaris, Konstantinos Karantzalos, and Nikos Komodakis. Spot: Self-training with patch-order permutation for object-centric learning with autoregressive transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22776-22786, 2024. 2, 5, 6", + "[18] Jinwoo Kim, Janghyuk Choi, Ho-Jin Choi, and Seon Joo Kim. Shepherding slots to objects: Towards stable and robust object-centric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19198-19207, 2023. 5, 6", + "[19] Hyeokjun Kweon, Sung-Hoon Yoon, and Kuk-Jin Yoon. Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11329-11339, 2023. 3", + "[20] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 40:e253, 2017. 1", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[22] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection without imagenet. arXiv preprint arXiv:2011.13677, 2020. 2", + "[23] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 1, 2, 3, 5, 6", + "[24] Amin Mansouri, Jason Hartford, Yan Zhang, and Yoshua Bengio. Object-centric architectures enable efficient causal representation learning. arXiv preprint arXiv:2310.19054, 2023. 1", + "[25] Gary F Marcus. The algebraic mind: Integrating connectionism and cognitive science. MIT press, 2003. 1" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30289", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020. 2", + "[27] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16675-16687, 2023. 1", + "[28] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. In The 11th International Conference on Learning Representations, 2023. 2, 5, 6", + "[29] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dall-e learns to compose. In 10th International Conference on Learning Representations, ICLR 2022, 2022. 2, 5, 6", + "[30] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020. 2", + "[31] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. Advances in Neural Information Processing Systems, 34: 16238-16250, 2021. 2", + "[32] Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17:395-416, 2007. 5", + "[33] Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17:395-416, 2007. 2", + "[34] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 2", + "[35] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. Advances in neural information processing systems, 35:16423-16438, 2022. 4", + "[36] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotdiffusion: Object-centric generative modeling with diffusion models. Advances in Neural Information Processing Systems, 36:50932-50958, 2023. 2, 5, 6", + "[37] Ceyuan Yang, Zhirong Wu, Bolei Zhou, and Stephen Lin. Instance localization for self-supervised detection pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3987-3996, 2021. 2", + "[38] Ying Zheng, Lei Yao, Yuejiao Su, Yi Zhang, Yi Wang, Sicheng Zhao, Yiyi Zhang, and Lap-Pui Chau. A survey of embodied learning for object-centric robotic manipulation. arXiv preprint arXiv:2408.11537, 2024. 1", + "[39] Rui Zhu, Bingchen Zhao, Jingen Liu, Zhenglong Sun, and Chang Wen Chen. Improving contrastive learning by visualizing feature transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10306-10315, 2021. 2" + ], + "bbox": [ + 91, + 92, + 482, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30290", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Pay Attention to the Foreground in Object-Centric Learning/b4bf29e9-ae1b-4c69-a186-15cc519fb707_model.json b/2025/Pay Attention to the Foreground in Object-Centric Learning/b4bf29e9-ae1b-4c69-a186-15cc519fb707_model.json new file mode 100644 index 0000000000000000000000000000000000000000..683c1b8c939f9d09d5027f68ca5c7344d1e22350 --- /dev/null +++ b/2025/Pay Attention to the Foreground in Object-Centric Learning/b4bf29e9-ae1b-4c69-a186-15cc519fb707_model.json @@ -0,0 +1,1892 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.194, + 0.131, + 0.806, + 0.154 + ], + "angle": 0, + "content": "Pay Attention to the Foreground in Object-Centric Learning" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.18, + 0.714, + 0.199 + ], + "angle": 0, + "content": "Pinzhuo Tian\\(^{1}\\), Shengjie Yang\\(^{1}\\), Hang Yu\\(^{1*}\\), Alex Kot\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.199, + 0.796, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Engineering and Science, Shanghai University, China" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.217, + 0.877, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.256, + 0.238, + 0.744, + 0.252 + ], + "angle": 0, + "content": "{pinzhuo, sjyang, yuhang}@shu.edu.cn, eackot@ntu.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.304, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.336, + 0.485, + 0.699 + ], + "angle": 0, + "content": "The slot attention-based method is widely used in unsupervised object-centric learning, aiming to decompose scenes into interpretable objects and associate them with slots. However, complex backgrounds in the real images can disrupt the model's focus, leading it to excessively segment background stuff into different regions based on low-level information such as color or texture variations. As a result, the detailed segmentation of foreground objects, which requires shape or geometric information, is often neglected. To address this issue, we introduce a contrastive learning-based indicator designed to differentiate between foreground and background. Integrating this indicator into the existing slot attention-based method enables the model to focus more on segmenting foreground objects while minimizing background distractions. During the testing phase, we utilize a spectral clustering mechanism to refine the results based on the similarity between the slots. Experimental results show that incorporating our method with various state-of-the-art models significantly improves their performance on both simulated data and real-world datasets. Furthermore, multiple sets of ablation experiments confirm the effectiveness of each proposed component. The source code is available at https://github.com/sjyjs09/FG-BG_Indicator." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.728, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.754, + 0.483, + 0.83 + ], + "angle": 0, + "content": "In perceiving the surrounding environment, humans typically divide it into distinct objects, subsequently understanding the scene through the relationships and connections among these objects [20, 25]. Thus, enabling our model to decompose the visual scene into different objects" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.302, + 0.907, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.584 + ], + "angle": 0, + "content": "Figure 1. The overview of our method. The background stuff in the complex scenes can influence the process of slot binding. Variations in color or texture within the background may cause the slot to incorrectly segment the background into different regions while neglecting the foreground (shown in the above branch). In our method, we integrate information from our indicator to enhance the distinction between foreground and background areas and mitigate changes in background region, enabling the slot attention method to focus more effectively on foreground object segmentation (shown in the below branch)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.687 + ], + "angle": 0, + "content": "is highly meaningful and lays the foundation for reasoning [24]. Unsupervised object-centric learning [12] is an approach designed to address this challenge, aiming to represent a scene as a composition of distinct objects using only visual information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.854 + ], + "angle": 0, + "content": "After significant development, the slot attention-based method [23] has become the mainstream framework in the field. This kind of method can be divided into two primary stages: the binding stage and the reconstruction stage. In the binding stage, an attention-based mechanism is used to assign different areas of the image to distinct slots. The reconstruction stage then involves rebuilding the input based on information derived from these slots, ensuring that each slot accurately represents an individual object within the image. These two stages enable the slot attention-based method to achieve impressive results in various areas, such as [27, 38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "However, current slot attention-based methods do not distinguish between foreground objects and background stuff. When applied to complex scenes, variations in color" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.84, + 0.243, + 0.852 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.852, + 0.483, + 0.9 + ], + "angle": 0, + "content": "This work is supported by projects of the National Natural Science Foundation of China (Grant No.62206166, No.62302287), Shanghai Sailing Program (Grant No.23YF1413000), and Shanghai Committee of Science and Technology (Grant No.23ZR1423500)." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30281" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.181 + ], + "angle": 0, + "content": "or texture within the background can confuse the slots. This misdirection often causes the slots to divide the background into different regions based on low-level visual cues. As a result, the slots struggle to bind to foreground objects, which require a more nuanced recognition of shape and geometric details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.412 + ], + "angle": 0, + "content": "To address this issue, we propose an unsupervised indicator that provides crucial information to differentiate between foreground and background. After training, it can be incorporated into various slot attention methods without requiring retraining. Figure 1 shows an overview of our method. Our unsupervised indicator is based on a contrastive learning framework [15] and consists of three components: a pixel-level contrastive loss that ensures consistency in pixel predictions across different views to promote meaningful representation learning; a stuff-level contrastive loss that minimizes differences within the background, reducing misguidance caused by color and texture variations in the background during slot binding; and an entropy-based regularization that enhances the distinction between foreground and background elements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.483, + 0.7 + ], + "angle": 0, + "content": "By incorporating information from this indicator, the slot attention-based method shifts its focus toward more accurate object segmentation while minimizing background distractions. However, due to the absence of supervised ground truth, the indicator cannot perfectly differentiate between foreground and background regions, which may introduce noise or incorrect labels. To mitigate the impact of this noise during training, we first train the original slot attention-based method for several epochs, then incorporate the foreground and background information from the indicator. During the testing phase, recognizing that the similarity between slots can reflect the similarity between image regions, we employ a spectral clustering merging mechanism [33] to refine the results and reduce oversegmentation. In our experiments, we integrate our method with two prominent object-centric learning approaches, DINOSAUR [28] and SPOT [17]. The results demonstrate that our approach improves their performance on both simulated and real-world datasets." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.703, + 0.404, + 0.717 + ], + "angle": 0, + "content": "In summary, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.795 + ], + "angle": 0, + "content": "- We consider the issue that variations in background information, like color or texture, can mislead the slot allocation process and lead to the neglect of foreground object segmentation, which requires detailed shape and boundary information." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.796, + 0.483, + 0.855 + ], + "angle": 0, + "content": "- We propose an unsupervised foreground and background indicator to solve this problem and employ a spectral clustering merging mechanism that utilizes the similarity between slots to refine the results." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.857, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Our experimental results demonstrate that integrating our method with various state-of-the-art models significantly enhances their performance both on simulated data and" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.659, + 0.106 + ], + "angle": 0, + "content": "real-world datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.121, + 0.652, + 0.137 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.147, + 0.905, + 0.344 + ], + "angle": 0, + "content": "Object-centric learning. Object-centric learning aims to represent scenes as compositions of distinct objects. MONet [2] and IODINE [11] are early studies in this area. However, they rely on multiple encode-decode steps. Slot Attention [23] achieves efficient scene decomposition by replacing this process with a single encoding step that iterates attention. Although Slot Attention has gradually become a major paradigm in object-centric learning, this approach still faces challenges when applied to real-world scenes. To address this issue, some studies focus on improving the encoder, like DINOSAUR [28]. Other studies focus on designing more effective decoders for scene decomposition, like SLATE [29], SPOT [17], and [16, 36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.907, + 0.586 + ], + "angle": 0, + "content": "Contrastive learning. Contrastive learning is a mainstream paradigm in self-supervised learning (SSL), aiming to learn consistent representations by contrasting positive and negative sample pairs without the need for additional annotations [4, 7, 8]. Classical contrastive learning methods treat the image as a whole part and have shown strong potential in learning transferable visual representations [6, 14, 30, 39]. However, real-world scenes often contain complex and diverse information, making it difficult for global features to represent all the information. To address this issue, some works leverage local crops [3, 31, 37], while others gradually move toward pixel-level contrastive learning. Pixel-level contrastive learning extends the task of instance discrimination from image-level feature vectors to feature maps [22, 26, 34], which naturally aids in learning dense representations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.6, + 0.605, + 0.616 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.905, + 0.732 + ], + "angle": 0, + "content": "Our approach consists of three key procedures: 1) the foreground and background indicator, which supplies additional information about foreground objects and background elements, 2) the fusion module, responsible for integrating information from the indicator into the existing slot attention method, 3) the spectral clustering module, which merges regions to refine prediction results in the test stage." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.845, + 0.759 + ], + "angle": 0, + "content": "3.1. Foreground and background indicator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.825 + ], + "angle": 0, + "content": "This part aims to achieve two goals: 1) it needs to be an unsupervised method, and 2) it should provide an accurate and robust representation to distinguish the foreground and background." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Hence, the implementation of our indicator leverages advancements in the contrastive learning framework, like teacher-student framework. Specifically, our approach involves a student network, denoted by \\(\\theta\\), and a teacher network, denoted by \\(\\xi\\). Both networks share the same architec" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30282" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.09, + 0.885, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.578, + 0.908, + 0.622 + ], + "angle": 0, + "content": "Figure 2. Illustration of our pipeline. During the training stage, the foreground and background indicator are trained to provide highly distinguishable foreground and background features (the part of (a)). Then, the feature from the indicator is integrated into the original slot attention-based methods (the part of (b)). In the testing stage, we use region combination to refine the results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.752 + ], + "angle": 0, + "content": "ture but use different sets of weights. The parameters of the teacher network, \\(\\xi\\), are updated as an exponential moving average of the student network, \\(\\theta\\). In each network, we use two distinct slots, \\(S = \\{s^f, s^b\\}\\), i.e., \\(S_{\\theta}\\) and \\(S_{\\xi}\\) for the student and teacher networks, respectively, to help recognize the foreground and background in each image. \\(S_{\\xi}\\) is also optimized by EMA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Unlike the classical slot attention method, \\( s_{\\theta \\text{ or } \\xi}^{f} \\) and \\( s_{\\theta \\text{ or } \\xi}^{b} \\) are given semantic meaning, enabling the classification of pixels into their corresponding semantic groups, such as foreground and background. Here, \\( s_{\\theta}^{f} \\) and \\( s_{\\theta}^{b} \\) represent the slots for foreground and background in the student network, respectively, while \\( s_{\\xi}^{f} \\) and \\( s_{\\xi}^{b} \\) represent the slots for foreground and background in the teacher network, respectively. However, we need an initialized boot mode to provide the initialized semantic cues for foreground objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.907, + 0.708 + ], + "angle": 0, + "content": "Pre-trained models have proven effective in providing initial semantic information across various fields [19]. Therefore, we employ the pre-trained model as the backbone of our method instead of training from scratch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our indicator contains three aspects. The first is the pixel-level consistency. For different augmentations of the same image, we ensure that pixels in the same location receive similar assignment scores. Given an input image \\( x \\), two random augmentations are applied to produce two augmented views \\( v^{1}(x) \\) and \\( v^{2}(x) \\). Each augmentation view is then encoded with an encoder \\( f \\) into a hidden feature map and then transformed with a multilayer perceptron (MLP) \\( g \\) to get the feature map \\( z \\), i.e., \\( z_{\\theta}(x) = g_{\\theta}(f_{\\theta}(v^{1}(x))) \\) and \\( z_{\\xi}(x) = g_{\\xi}(f_{\\xi}(v^{2}(x))) \\). The fore-background representation vectors of the image \\( x \\) in different networks can be obtained via slot attention [23], i.e., \\( \\hat{S}_{\\theta}(x) = \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30283" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.168 + ], + "angle": 0, + "content": "slot attention \\((z_{\\theta}(x),S_{\\theta})\\) \\(\\hat{S}_{\\xi} =\\) slot attention \\((z_{\\xi}(x),S_{\\xi})\\) Then, the pixel assignment matrix, defining the score fore or background each pixel belongs to, \\(\\mathcal{P}_{\\theta}\\) and \\(\\mathcal{Q}_{\\xi}\\) can obtain by \\(\\mathcal{P}_{\\theta}(x) = z_{\\theta}\\cdot \\hat{S}_{\\theta}^{\\top}\\) . The assignment matrix \\(\\mathcal{Q}_{\\xi}(x)\\) for the teacher network can be obtained in the same way." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.482, + 0.306 + ], + "angle": 0, + "content": "To keep the pixel level consistency, we enforce the prediction of the two views assignment \\(\\mathcal{P}_{\\theta}(x)\\) and \\(\\mathcal{Q}_{\\xi}(x)\\) match each other. However, the geometric image augmentations, such as random crop, scale, or flip, between different views will cause the scale or layout of the two feature maps to be inconsistent. So similar to [35], we use the inverse augmentation process on the predicted assignments to align their spatial locations, i.e., \\(\\tilde{\\mathcal{P}}_{\\theta}(x) = \\mathrm{invaug}(\\mathcal{P}_{\\theta}(x)), \\tilde{\\mathcal{Q}}_{\\xi}(x) = \\mathrm{invaug}(\\mathcal{Q}_{\\xi}(x))\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.482, + 0.381 + ], + "angle": 0, + "content": "Based on the aligned assignments, we apply the cross-entropy loss to enforce the consistency in assignment scores between spatial-aligned pixels from different views. The cross-entropy loss is averaged over all spatial locations to keep the pixel prediction consistency" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.392, + 0.483, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {p i x e l}} (x) = \\frac {1}{H \\times W} \\sum_ {i, j} \\mathcal {L} ^ {\\mathrm {C E}} \\left(\\tilde {\\mathcal {Q}} _ {\\xi} (x) [ i, j ], \\tilde {\\mathcal {P}} _ {\\theta} (x) [ i, j ]\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.458, + 0.411, + 0.472 + ], + "angle": 0, + "content": "\\(i,j\\) means the position in the assignment matrix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.671 + ], + "angle": 0, + "content": "In addition to the pixel-level contrastive loss, we incorporate a stuff-level consistency loss. This loss aims to alleviate variations in the background by treating the background in different images as positive samples, while considering all foreground samples as negative samples. According to the pixel assignment matrix \\(\\tilde{\\mathcal{P}}_{\\theta}(x)\\) and \\(\\tilde{\\mathcal{Q}}_{\\xi}(x)\\), we can obtain the prototypes of the foreground and background by region average pooling to represent the whole foreground and background regions of the image \\(x\\). \\(p_{\\theta}^{f}(x), p_{\\theta}^{b}(x)\\) represent the foreground and background prototypes obtained from the student network. \\(p_{\\xi}^{f}(x), p_{\\xi}^{b}(x)\\) are from the teacher network. Then, the stuff-level loss can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.681, + 0.482, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} ^ {\\text {s t u f f}} (B) = \\frac {1}{B (B - 1)} \\sum_ {i = 1} ^ {B} \\sum_ {j \\neq i} ^ {B} [ \\operatorname {s i m} \\left(p _ {\\theta} ^ {b} \\left(x _ {i}\\right), p _ {\\xi} ^ {b} \\left(x _ {j}\\right)\\right) - \\\\ \\left. \\left. \\operatorname {s i m} \\left(p _ {\\theta} ^ {b} \\left(x _ {i}\\right), p _ {\\xi} ^ {f} \\left(x _ {j}\\right)\\right) \\right]. \\right. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.855 + ], + "angle": 0, + "content": "In Eq. 2, we hope the similarity between the background across different views will be large, however, the similarity between foreground and background should be small. We adopt the cosine similarity in our method. \\( B \\) means the batch of training data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To encourage the semantic centers to emphasize different visual contents and cover diverse semantics to avoid collapse, we apply a regularization to the assignment matrix of" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.696, + 0.106 + ], + "angle": 0, + "content": "different views, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.116, + 0.905, + 0.133 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\operatorname {s e p}} (x) = \\operatorname {E n t r o y} \\left(\\mathcal {P} _ {\\theta} (x)\\right) + \\operatorname {E n t r o p y} \\left(\\mathcal {Q} _ {\\xi} (x)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.142, + 0.869, + 0.157 + ], + "angle": 0, + "content": "The total loss to optimize the proposed indicator is" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.166, + 0.905, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} ^ {\\text {p i x e l}} (x) - \\lambda \\mathcal {L} ^ {\\text {s t u f f}} (B) - \\gamma \\mathcal {L} ^ {\\text {s e p}} (x). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.192, + 0.905, + 0.222 + ], + "angle": 0, + "content": "3.2. Fusion of information into original slot attention method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.905, + 0.379 + ], + "angle": 0, + "content": "The features provided by our indicator allow for clear distinction between the foreground object and background elements. Simultaneously, the background is processed uniformly, harmonizing its characteristics across various regions. Integrating the features obtained through our method with existing approaches significantly enhances the differentiation between the foreground and background. Furthermore, once trained, our indicator is frozen and only provides supplementary information to the existing slot attention-based methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.906, + 0.516 + ], + "angle": 0, + "content": "In our fusion stage, we first train the slot attention-based method for several epochs. Once it has learned meaningful information, we next integrate the foreground and background information from our indicator into the slot attention-based method as supplementary information. Finally, the features from both the original slot attention-based method and our indicator are fused together and input into the next decoder for image reconstruction. The feature fusion stage can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.524, + 0.905, + 0.544 + ], + "angle": 0, + "content": "\\[\nz ^ {\\text {f u s i o n}} (x) = \\operatorname {M L P} \\left(z ^ {\\text {s l o t}} (x) + z _ {\\xi} ^ {\\text {i n d i c a t o r}} (x)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.704 + ], + "angle": 0, + "content": "where \\( z^{\\mathrm{slot}}(x) \\) is the feature map from the encoder in the original slot attention method. \\( z_{\\xi}^{\\mathrm{indicator}}(x) \\) is the feature map provided by our indicator. \\( x \\) is the input image. These two feature maps are added together and then passed through a MLP layer to transform the combined feature map into a new space. The resulting fusion feature \\( z^{\\mathrm{fusion}}(x) \\) replaces the feature map \\( z^{\\mathrm{slot}}(x) \\) used in the original slot attention-based method to do the next stages, such as slot attention, and decoding, which are the same as the original slot attention framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.71, + 0.729 + ], + "angle": 0, + "content": "3.3. Region Combination" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.871 + ], + "angle": 0, + "content": "Defining the number of slots is a common challenge in recent slot attention methods [10] and can significantly impact their performance. Determining the appropriate number of slots is not straightforward, and recent approaches often opt for a higher number of slots, leading to slight oversegmentation, in order to achieve satisfactory results. In this section, we propose a spectral clustering-based method to group regions likely belonging to the same semantic object, which is used to refine the predictions during the testing phase." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The similarity between slots represents the similarity between regions in the image. The similarity matrix between" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30284" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "different slots can be regarded as a reference to help recognize which slots possibly belong to the same semantic object. In our method, we directly use spectral clustering to divide the slots into different groups based on this matrix. Specifically, we calculate the cosine similarity between each pair of the slots to construct the similarity matrix \\(\\mathbf{M}\\). Then according to \\(\\mathbf{M}\\), the group results can be obtained by" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.212, + 0.483, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = \\text {s p e c t r a l c l u s t e r i n g} (\\mathbf {M}, N). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.485, + 0.336 + ], + "angle": 0, + "content": "Based on the group result \\(\\mathcal{G}\\), the corresponding areas of slots in the same group can be combined. To determining the group number \\(N\\), we primarily employ eigen-gap heuristic [32]. In this method, we use the Laplacian matrix of the graph and find the largest gap between consecutive eigenvalues, which suggests the optimal number of clusters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.352, + 0.223, + 0.369 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.377, + 0.173, + 0.394 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.581 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on one synthetic and two real-world datasets. For the synthetic dataset, we use MOVi-C [13], which contains approximately 1,000 real 3D-scanned objects, with each scene featuring between 3 and 10 objects. Although MOVi-C is originally a video dataset, we convert it into an image dataset by randomly selecting frames following [28]. For the real-world datasets, we utilize MS COCO 2017 [21] and PASCAL VOC 2012 [9], which are commonly used for object detection and segmentation. PASCAL VOC 2012 contains many images with a single salient object, while MS COCO 2017 features more complex scenes, offering a more challenging evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Metrics. To assess object-centric learning, we used Mean Best Overlap (mBO) and Mean Intersection over Union (mIoU). The Mean Best Overlap (mBO) includes instance-level \\(\\mathrm{mBO}^i\\) and category-level \\(\\mathrm{mBO}^c\\). \\(\\mathrm{mBO}^i\\) and \\(\\mathrm{mBO}^c\\) match the predicted masks to the instance and semantic masks, respectively, by finding the best overlap. Meanwhile, mIoU employs the Hungarian algorithm to match predicted and ground truth masks. [17] suggests that FG-ARI, as a clustering similarity metric, may yield unreliable segmentation quality assessments by overlooking the accuracy of predicted masks. Therefore, FG-ARI is not reported in this paper." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. To optimize the proposed foreground and background indicator, we employ the SGD optimizer. The learning rate is 0.001. Batch size is 512. The learning rate follows a linear warm-up for 10,000 optimization steps and an exponential decay schedule. Additionally, we clip the gradient norm to 1 to stabilize the training process. ViT-B/16 [1], initialized with DINO [5], is used as the encoder in our indicator. During the fusion stage, the model is trained using the same configuration as the corresponding" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "original slot attention method you choose. All models are trained on a single A6000 GPU with 48 Gbytes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.133, + 0.902, + 0.15 + ], + "angle": 0, + "content": "4.2. Integrating into SOTA object-centric methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.156, + 0.905, + 0.337 + ], + "angle": 0, + "content": "Baselines. We apply our approach to two slot attention methods, DINOSAUR [28] and SPOT [17]. DINOSAUR utilizes a Vision Transformer encoder for feature reconstruction and has demonstrated applicability to real-world datasets. SPOT is a recent SOTA object-centric model that employs a two-stage self-training scheme and a patch-order permutation strategy for the transformer decoder. Furthermore, we also list the results of several other object-centric methods, including Slot Attention [23], SLASH [18], SLATE [29], SlotDiffusion [36], and LSD [16]. A more detailed description of the baseline methods can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.905, + 0.58 + ], + "angle": 0, + "content": "Main Results on Each Dataset. Tab. 1 presents the main results of our method on the COCO, PASCAL, and MOViC datasets. Our method improves the performance of DINOSAUR and SPOT across all datasets. The fusion stage yields consistent improvements, while the region combination stage leads to a slight decrease in the \\(\\mathrm{mBO}^i\\) metric on the PASCAL and COCO datasets. This is due to the fact that, during clustering, slots corresponding to two instances of the same category may be merged into one, resulting in an increase in \\(\\mathrm{mBO}^c\\) but a decrease in \\(\\mathrm{mBO}^i\\). Despite this, our method still outperforms the baseline models, DINOSAUR and SPOT, demonstrating a clear overall performance advantage. This highlights the crucial role of the fusion stage in driving performance improvements, while the region combination stage primarily serves to refine the predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.681, + 0.605 + ], + "angle": 0, + "content": "4.3. More discussions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.689 + ], + "angle": 0, + "content": "In this section, we analyze various aspects of our approach: the significance of the foreground and background indicators, the impact of different loss functions within the indicator, the role of the fusion stage, and the consistency of performance improvement across different slot numbers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.905, + 0.901 + ], + "angle": 0, + "content": "(A) Does the extra foreground and background information really improve the slot attention method? We evaluate the slot attention method in four cases. Tab. 2 shows the results. In this table, \"Random\" refers to using random noise as the input to the fusion stage, \"ViT\" refers to directly using the pre-trained DINO ViT as the indicator, and \"Ours\" refers to our foreground and background indicator. We find that when the indicator is random noise, the baseline method struggles to distinguish between foreground and background, leading to a significant performance drop. When ViT is used as the indicator, the pretrained ViT has some inherent ability to differentiate between foreground and background, resulting in a slight performance improvement. Our designed foreground" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30285" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.887, + 0.297 + ], + "angle": 0, + "content": "
MethodMOVi-CPASCALCOCO
mBOimIoUmBOimBOcmBOimBOc
Slot Attention [23] [NeurIPS20]26.3-24.624.917.219.2
SLASH [18] [CVPR23]-27.7----
SLATE [29] [ICLR22]39.437.835.941.529.133.6
SlotDiffusion [36] [NeurIPS23]--50.455.331.035.0
(Stable-)LSD [16] [NeurIPS23]45.644.2--30.4-
DINOSAUR [28] [ICLR23]42.441.844.051.231.138.9
DINOSAUR + FS (Ours)45.3 ↑ 2.944.5 ↑ 2.745.4 ↑ 1.452.5 ↑ 1.331.5 ↑ 0.439.9 ↑ 1.0
DINOSAUR + FS + RC (Ours)46.0 ↑ 3.644.8 ↑ 3.045.2 ↑ 1.252.9 ↑ 1.731.2 ↑ 0.140.1 ↑ 1.2
SPOT [17] [CVPR24]47.046.448.155.334.744.3
SPOT + FS (Ours)48.5 ↑ 1.547.7 ↑ 1.349.1 ↑ 1.056.3 ↑ 1.035.5 ↑ 0.845.0 ↑ 0.7
SPOT + FS + RC (Ours)49.0 ↑ 2.047.8 ↑ 1.449.3 ↑ 1.256.5 ↑ 1.235.7 ↑ 1.045.3 ↑ 1.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.307, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Table 1. Comparison with different object-centric methods on MOVi-C, PASCAL, and COCO datasets. DINOSAUR uses an autoregressive decoder and DINO [5] ViT encoder. The experimental results of the baseline are derived from the original corresponding papers. FS means using the fusion stage in our paper. RC means the trick of region combination." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.373, + 0.462, + 0.466 + ], + "angle": 0, + "content": "
SA MethodIndicatorIoUmBOcmBOi
DINOSAUR--51.244.0
Random-25.222.1
ViT30.251.444.7
Ours52.652.545.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.56 + ], + "angle": 0, + "content": "Table 2. The effect of extra information on the existing slot attention-based method. All experiments are conducted on PASCAL dataset. IoU is used to show the model's performance on segmenting the foreground and background. \\(\\mathrm{mBO}^c\\) and \\(\\mathrm{mBO}^i\\) are employed to assess the final performance after fusing the extra knowledge from the indicator." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.577, + 0.462, + 0.65 + ], + "angle": 0, + "content": "
MethodMOVi-CPASCALCOCO
DINOSAUR (2 slots)45.247.646.9
SPOT (2 slots)46.650.846.4
Ours47.452.648.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.66, + 0.483, + 0.702 + ], + "angle": 0, + "content": "Table 3. The performance of using object-centric methods as the foreground and background indicator. Fig. 3 shows some visualization results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.776 + ], + "angle": 0, + "content": "and background indicator makes the foreground and background highly separable and significantly improves performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "(B) How about using the classical slot attention methods as the foreground and background indicator in our method? Tab. 3 evaluates the performance of different models when the number of slots is set to 2. We continue to use the Mean Best Overlap (mBO) as the evaluation metric. The difference is that, during metric computation, we convert the ground truth masks into binary masks containing only foreground (0) and background (1). Then, the pre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.406 + ], + "angle": 0, + "content": "dicted masks are matched with the ground truth masks using the best overlap matching method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.909, + 0.637 + ], + "angle": 0, + "content": "From the data in Tab. 3, it can be observed that our method achieves the best performance with significant performance improvements on both COCO and PASCAL datasets. This is one of the reasons for using the foreground and background indicator. Another reason is that autoencoder-based methods like DINOSAUR and SPOT, while capable of dividing the scene into two parts, struggle to achieve accurate foreground-background segmentation. As shown in Fig. 3, the baseline methods tend to divide the scene into two parts, top and bottom, while our method focuses on separating the foreground and background. This is because our method employs both pixel-level and stufflevel contrastive learning, enabling the model to better distinguish foreground from background features, leading to a more effective foreground-background decomposition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.909, + 0.883 + ], + "angle": 0, + "content": "(C) Impact of different losses. We conduct the ablation experiments to further analyze the role of different losses in the indicator, as shown in Tab. 4. We find that without using \\(\\mathcal{L}^{\\mathrm{sep}}\\), the indicator struggle to achieve a performance improvement. This is because when only \\(\\mathcal{L}^{\\mathrm{pixel}}\\) and \\(\\mathcal{L}^{\\mathrm{stuff}}\\) are used, the model tends to collapse all information into a single slot, where one slot binds to the entire image and the other slot remains empty. In this case, the model can achieve low \\(\\mathcal{L}^{\\mathrm{pixel}}\\) and \\(\\mathcal{L}^{\\mathrm{stuff}}\\) values, but it cannot effectively perform the foreground-background separation. Therefore, we introduce \\(\\mathcal{L}^{\\mathrm{sep}}\\) to encourage semantic centers to emphasize different visual content and cover diverse semantics. However, relying on \\(\\mathcal{L}^{\\mathrm{sep}}\\) and \\(\\mathcal{L}^{\\mathrm{pixel}}\\) makes it difficult for the model to adapt to variations in the background. As a result, \\(\\mathcal{L}^{\\mathrm{sep}}\\), \\(\\mathcal{L}^{\\mathrm{pixel}}\\), and \\(\\mathcal{L}^{\\mathrm{stuff}}\\) must work together in synergy to achieve effective foreground-background separation." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.908, + 0.902 + ], + "angle": 0, + "content": "At the same time, it can be seen that the performance of" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30286" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.093, + 0.478, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.324, + 0.483, + 0.366 + ], + "angle": 0, + "content": "Figure 3. Example results of the proposed indicator and SOTA slot attention method with 2 slots. Our method distinguishes foreground and background better." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.484, + 0.466 + ], + "angle": 0, + "content": "the indicator is positively correlated with the final results. This further confirms the conclusion in (A): the stronger the indicator's ability to distinguish between foreground and background, the more effectively it enhances the feature fusion process." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.478, + 0.46, + 0.58 + ], + "angle": 0, + "content": "
LpixelLstuffLsepIoUmBOcmBOi
✓28.851.444.6
✓✓28.451.344.7
✓✓48.951.744.8
✓✓✓52.652.545.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.59, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Table 4. Analysis of different loss functions on PASCAL dataset. IoU represents the results of the indicator on segmenting foreground and background regions. \\(\\mathrm{mBO}^c\\) and \\(\\mathrm{mBO}^i\\) represent the final results of DINOSAUR after fusing knowledge from different indicators, which are optimized using different types of loss functions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.87 + ], + "angle": 0, + "content": "(D) The importance of the fusion stage and region combination stage. We conduct a detailed analysis of the importance of the fusion stage and the region combination stage. The results are summarized in Table 5. We observe that directly applying the region combination to DINOSAUR and SPOT typically leads to a decline in performance metrics. We hypothesize that this is due to the limited ability of the baseline methods to differentiate between foreground and background. Without a clear distinction, the region combination stage inappropriately merges foreground and background. The visualization results are shown in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "On the MOVi dataset, applying region combination directly to SPOT results in a performance improvement. This" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.901, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.335, + 0.906, + 0.363 + ], + "angle": 0, + "content": "Figure 4. The visualization results in our different components. FS means fusion stage, and RC is the stage of region combination." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.906, + 0.468 + ], + "angle": 0, + "content": "may result from the relatively simple background in the MOVi dataset, where SPOT is already effective at distinguishing between foreground and background. As shown in Tab. 3, the metric for SPOT with 2 slots closely aligns with our indicator result, further supporting this observation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.68 + ], + "angle": 0, + "content": "(E) Consistent improvement across different slot numbers. The quality of results is influenced by the number of slots used during model training. Therefore, understanding the model's robustness to variations in the number of slots is crucial. In this part, we compare our method with DINOSAUR to evaluate the stability of our method with different numbers of slot. We conduct experiments on two real-world datasets, COCO and PASCAL, using five slot numbers, with the results presented in Tab. 6. Notably, our method outperforms DINOSAUR across all slot numbers and exhibits lower variance. This suggests that, while improving baseline performance, our method also demonstrates better generalization ability when applied with different numbers of slots." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.692, + 0.712 + ], + "angle": 0, + "content": "4.4. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Fig. 5 shows example results on the COCO and PASCAL datasets for DINOSAUR, SPOT, and our method. Row (a) demonstrates that, compared to baseline methods, our approach correctly fuses the foreground and background. For example, in the PASCAL dataset, our method successfully merges the ground that was originally divided into two separate slots. Rows (b) and (c) illustrate that when the baseline method accurately segments the background (foreground), our method correctly fuses the foreground (background) regions. For instance, in row (b) of the PASCAL dataset, when the background is correctly segmented, our method successfully merges the train and streetlight, which were" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30287" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.887, + 0.251 + ], + "angle": 0, + "content": "
MethodMOVi-CPASCALCOCO
mBOimIoUmBOimBOcmBOimBOc
DINOSAUR42.441.844.051.231.138.9
DINOSAUR+RC41.6 ↓0.839.5 ↓2.343.3 ↓0.750.7 ↓0.530.4 ↓0.739.4 ↑0.5
DINOSAUR+FS (Ours)45.3 ↑2.944.5 ↑2.745.4 ↑1.452.5 ↑1.331.5 ↑0.439.9 ↑1.0
DINOSAUR+FS+RC (Ours)46.0 ↑3.644.8 ↑3.045.2 ↑1.252.9 ↑1.731.2 ↑0.140.1 ↑1.2
SPOT47.046.448.155.334.744.3
SPOT+RC47.8 ↑0.847.1 ↑0.748.0 ↓0.154.8 ↓0.533.2 ↓1.543.5 ↓0.8
SPOT+FS (Ours)48.5 ↑1.547.7 ↑1.349.1 ↑1.056.3 ↑1.035.5 ↑0.845.0 ↑0.7
SPOT+FS+RC (Ours)49.0 ↑2.047.8 ↑1.449.3 ↑1.256.5 ↑1.235.7 ↑1.045.3 ↑1.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.261, + 0.907, + 0.29 + ], + "angle": 0, + "content": "Table 5. The importance of the different components in our method. FS is the fusion stage and RS is the region combination stage. Using RS alone even hurt the performance. FS is crucial." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.304, + 0.501, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.304, + 0.895, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.502, + 0.807, + 0.516 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on COCO and PASCAL datasets. Ours: fusion stage + region combination." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.54, + 0.481, + 0.747 + ], + "angle": 0, + "content": "
DatasetKDINOSAURDINOSAUR+FS
mBOimBOcmBOimBOc
COCO227.620.530.2 ↑ 2.622.2 ↑ 1.7
537.929.339.3 ↑ 1.430.3 ↑ 1.0
739.031.139.9 ↑ 0.931.5 ↑ 0.4
938.531.739.8 ↑ 1.332.0 ↑ 0.3
1138.031.639.5 ↑ 1.532.6 ↑ 1.0
PASCAL234.340.435.4 ↑ 1.141.6 ↑ 1.2
444.351.644.6 ↑ 0.352.0 ↑ 0.4
644.851.845.4 ↑ 0.652.5 ↑ 0.7
845.351.745.5 ↑ 0.252.3 ↑ 0.6
1045.451.345.6 ↑ 0.252.0 ↑ 0.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.758, + 0.486, + 0.8 + ], + "angle": 0, + "content": "Table 6. The performance of our method with different numbers of slot. For a fair comparison, only the fusion stage is adopted. \\( K \\) represents the number of slots." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.361, + 0.84 + ], + "angle": 0, + "content": "originally assigned to two separate slots." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.229, + 0.866 + ], + "angle": 0, + "content": "4.5. Training cost" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The additional training cost of our method primarily consists of two components. First, training the indicator. Due" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.648 + ], + "angle": 0, + "content": "to its simple structure, training on the PASCAL dataset using a single A6000 GPU takes approximately 3 hours. Once trained, the indicator can be directly applied to different slot attention methods on the same dataset without the need for retraining. Second, retraining the existing slot attention methods based on the fused features, where the training time depends on the chosen slot attention method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this paper, we consider the problem that complex backgrounds in images can disrupt the phase of slot binding. Hence, we introduce a contrastive learning-based indicator to provide extra information about the foreground and background region to the existing slot attention method. Integrating this indicator into the slot attention-based method allows the model to focus more effectively on segmenting foreground objects and minimizing background distractions. According to the similarity between the slots, we also use a spectral clustering mechanism to refine the results in the testing phase. In experiments, we integrate our method into two SOTA slot attention methods. All the results show that our method can consistently enhance the performance on different datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30288" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.484, + 0.225 + ], + "angle": 0, + "content": "[2] Christopher P Burgess, Loic Matthew, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.296 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[9] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[10] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23062-23071, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[11] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In International conference on machine learning, pages 2424-2433. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[12] Klaus Greff, Sjoerd Van Steenkiste, and Jürgen Schmidhuber. On the binding problem in artificial neural networks. arXiv preprint arXiv:2012.05208, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapra" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "gasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[15] A Jaiswal, A Ramesh Babu, M Zaki Zadeh, D Banerjee, and F Makedon. A survey on contrastive self-supervised learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[16] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 8563-8601, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[17] Ioannis Kakogeorgiou, Spyros Gidaris, Konstantinos Karantzalos, and Nikos Komodakis. Spot: Self-training with patch-order permutation for object-centric learning with autoregressive transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22776-22786, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[18] Jinwoo Kim, Janghyuk Choi, Ho-Jin Choi, and Seon Joo Kim. Shepherding slots to objects: Towards stable and robust object-centric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19198-19207, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.907, + 0.546 + ], + "angle": 0, + "content": "[19] Hyeokjun Kweon, Sung-Hoon Yoon, and Kuk-Jin Yoon. Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11329-11339, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[20] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 40:e253, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[22] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection without imagenet. arXiv preprint arXiv:2011.13677, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[23] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Amin Mansouri, Jason Hartford, Yan Zhang, and Yoshua Bengio. Object-centric architectures enable efficient causal representation learning. arXiv preprint arXiv:2310.19054, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Gary F Marcus. The algebraic mind: Integrating connectionism and cognitive science. MIT press, 2003. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30289" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[27] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16675-16687, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[28] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. In The 11th International Conference on Learning Representations, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[29] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dall-e learns to compose. In 10th International Conference on Learning Representations, ICLR 2022, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[30] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[31] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. Advances in Neural Information Processing Systems, 34: 16238-16250, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.473, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[32] Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17:395-416, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[33] Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17:395-416, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[34] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[35] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. Advances in neural information processing systems, 35:16423-16438, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[36] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotdiffusion: Object-centric generative modeling with diffusion models. Advances in Neural Information Processing Systems, 36:50932-50958, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[37] Ceyuan Yang, Zhirong Wu, Bolei Zhou, and Stephen Lin. Instance localization for self-supervised detection pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3987-3996, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.772, + 0.482, + 0.827 + ], + "angle": 0, + "content": "[38] Ying Zheng, Lei Yao, Yuejiao Su, Yi Zhang, Yi Wang, Sicheng Zhao, Yiyi Zhang, and Lap-Pui Chau. A survey of embodied learning for object-centric robotic manipulation. arXiv preprint arXiv:2408.11537, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.828, + 0.482, + 0.898 + ], + "angle": 0, + "content": "[39] Rui Zhu, Bingchen Zhao, Jingen Liu, Zhenglong Sun, and Chang Wen Chen. Improving contrastive learning by visualizing feature transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10306-10315, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30290" + } + ] +] \ No newline at end of file diff --git a/2025/Pay Attention to the Foreground in Object-Centric Learning/b4bf29e9-ae1b-4c69-a186-15cc519fb707_origin.pdf b/2025/Pay Attention to the Foreground in Object-Centric Learning/b4bf29e9-ae1b-4c69-a186-15cc519fb707_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c9bb80a06bd33402134732b43c66ea345550649b --- /dev/null +++ b/2025/Pay Attention to the Foreground in Object-Centric Learning/b4bf29e9-ae1b-4c69-a186-15cc519fb707_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8f84c575ceacedadb9c9668ee58f2bf8ee24d5755b4967096cf86ce9e3558aa +size 2815089 diff --git a/2025/Pay Attention to the Foreground in Object-Centric Learning/full.md b/2025/Pay Attention to the Foreground in Object-Centric Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1b5d1233026124d322c3d95cefc4c9004b384ffa --- /dev/null +++ b/2025/Pay Attention to the Foreground in Object-Centric Learning/full.md @@ -0,0 +1,269 @@ +# Pay Attention to the Foreground in Object-Centric Learning + +Pinzhuo Tian $^{1}$ , Shengjie Yang $^{1}$ , Hang Yu $^{1*}$ , Alex Kot $^{2}$ + +$^{1}$ School of Computer Engineering and Science, Shanghai University, China + +$^{2}$ School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore + +{pinzhuo, sjyang, yuhang}@shu.edu.cn, eackot@ntu.edu.sg + +# Abstract + +The slot attention-based method is widely used in unsupervised object-centric learning, aiming to decompose scenes into interpretable objects and associate them with slots. However, complex backgrounds in the real images can disrupt the model's focus, leading it to excessively segment background stuff into different regions based on low-level information such as color or texture variations. As a result, the detailed segmentation of foreground objects, which requires shape or geometric information, is often neglected. To address this issue, we introduce a contrastive learning-based indicator designed to differentiate between foreground and background. Integrating this indicator into the existing slot attention-based method enables the model to focus more on segmenting foreground objects while minimizing background distractions. During the testing phase, we utilize a spectral clustering mechanism to refine the results based on the similarity between the slots. Experimental results show that incorporating our method with various state-of-the-art models significantly improves their performance on both simulated data and real-world datasets. Furthermore, multiple sets of ablation experiments confirm the effectiveness of each proposed component. The source code is available at https://github.com/sjyjs09/FG-BG_Indicator. + +# 1. Introduction + +In perceiving the surrounding environment, humans typically divide it into distinct objects, subsequently understanding the scene through the relationships and connections among these objects [20, 25]. Thus, enabling our model to decompose the visual scene into different objects + +![](images/7a6d541043eb5956d76897096cf7b62ad91251e038a6c2a5b2a2f47e7b91adfb.jpg) +Figure 1. The overview of our method. The background stuff in the complex scenes can influence the process of slot binding. Variations in color or texture within the background may cause the slot to incorrectly segment the background into different regions while neglecting the foreground (shown in the above branch). In our method, we integrate information from our indicator to enhance the distinction between foreground and background areas and mitigate changes in background region, enabling the slot attention method to focus more effectively on foreground object segmentation (shown in the below branch). + +is highly meaningful and lays the foundation for reasoning [24]. Unsupervised object-centric learning [12] is an approach designed to address this challenge, aiming to represent a scene as a composition of distinct objects using only visual information. + +After significant development, the slot attention-based method [23] has become the mainstream framework in the field. This kind of method can be divided into two primary stages: the binding stage and the reconstruction stage. In the binding stage, an attention-based mechanism is used to assign different areas of the image to distinct slots. The reconstruction stage then involves rebuilding the input based on information derived from these slots, ensuring that each slot accurately represents an individual object within the image. These two stages enable the slot attention-based method to achieve impressive results in various areas, such as [27, 38]. + +However, current slot attention-based methods do not distinguish between foreground objects and background stuff. When applied to complex scenes, variations in color + +or texture within the background can confuse the slots. This misdirection often causes the slots to divide the background into different regions based on low-level visual cues. As a result, the slots struggle to bind to foreground objects, which require a more nuanced recognition of shape and geometric details. + +To address this issue, we propose an unsupervised indicator that provides crucial information to differentiate between foreground and background. After training, it can be incorporated into various slot attention methods without requiring retraining. Figure 1 shows an overview of our method. Our unsupervised indicator is based on a contrastive learning framework [15] and consists of three components: a pixel-level contrastive loss that ensures consistency in pixel predictions across different views to promote meaningful representation learning; a stuff-level contrastive loss that minimizes differences within the background, reducing misguidance caused by color and texture variations in the background during slot binding; and an entropy-based regularization that enhances the distinction between foreground and background elements. + +By incorporating information from this indicator, the slot attention-based method shifts its focus toward more accurate object segmentation while minimizing background distractions. However, due to the absence of supervised ground truth, the indicator cannot perfectly differentiate between foreground and background regions, which may introduce noise or incorrect labels. To mitigate the impact of this noise during training, we first train the original slot attention-based method for several epochs, then incorporate the foreground and background information from the indicator. During the testing phase, recognizing that the similarity between slots can reflect the similarity between image regions, we employ a spectral clustering merging mechanism [33] to refine the results and reduce oversegmentation. In our experiments, we integrate our method with two prominent object-centric learning approaches, DINOSAUR [28] and SPOT [17]. The results demonstrate that our approach improves their performance on both simulated and real-world datasets. + +In summary, our contributions are threefold: + +- We consider the issue that variations in background information, like color or texture, can mislead the slot allocation process and lead to the neglect of foreground object segmentation, which requires detailed shape and boundary information. +- We propose an unsupervised foreground and background indicator to solve this problem and employ a spectral clustering merging mechanism that utilizes the similarity between slots to refine the results. +- Our experimental results demonstrate that integrating our method with various state-of-the-art models significantly enhances their performance both on simulated data and + +real-world datasets. + +# 2. Related work + +Object-centric learning. Object-centric learning aims to represent scenes as compositions of distinct objects. MONet [2] and IODINE [11] are early studies in this area. However, they rely on multiple encode-decode steps. Slot Attention [23] achieves efficient scene decomposition by replacing this process with a single encoding step that iterates attention. Although Slot Attention has gradually become a major paradigm in object-centric learning, this approach still faces challenges when applied to real-world scenes. To address this issue, some studies focus on improving the encoder, like DINOSAUR [28]. Other studies focus on designing more effective decoders for scene decomposition, like SLATE [29], SPOT [17], and [16, 36]. + +Contrastive learning. Contrastive learning is a mainstream paradigm in self-supervised learning (SSL), aiming to learn consistent representations by contrasting positive and negative sample pairs without the need for additional annotations [4, 7, 8]. Classical contrastive learning methods treat the image as a whole part and have shown strong potential in learning transferable visual representations [6, 14, 30, 39]. However, real-world scenes often contain complex and diverse information, making it difficult for global features to represent all the information. To address this issue, some works leverage local crops [3, 31, 37], while others gradually move toward pixel-level contrastive learning. Pixel-level contrastive learning extends the task of instance discrimination from image-level feature vectors to feature maps [22, 26, 34], which naturally aids in learning dense representations. + +# 3. Method + +Our approach consists of three key procedures: 1) the foreground and background indicator, which supplies additional information about foreground objects and background elements, 2) the fusion module, responsible for integrating information from the indicator into the existing slot attention method, 3) the spectral clustering module, which merges regions to refine prediction results in the test stage. + +# 3.1. Foreground and background indicator + +This part aims to achieve two goals: 1) it needs to be an unsupervised method, and 2) it should provide an accurate and robust representation to distinguish the foreground and background. + +Hence, the implementation of our indicator leverages advancements in the contrastive learning framework, like teacher-student framework. Specifically, our approach involves a student network, denoted by $\theta$ , and a teacher network, denoted by $\xi$ . Both networks share the same architec + +![](images/5a7f13e025d4f274a7e2c89ce4aac2acaffbafef8174c72c522e2abd1826b08a.jpg) +Figure 2. Illustration of our pipeline. During the training stage, the foreground and background indicator are trained to provide highly distinguishable foreground and background features (the part of (a)). Then, the feature from the indicator is integrated into the original slot attention-based methods (the part of (b)). In the testing stage, we use region combination to refine the results. + +ture but use different sets of weights. The parameters of the teacher network, $\xi$ , are updated as an exponential moving average of the student network, $\theta$ . In each network, we use two distinct slots, $S = \{s^f, s^b\}$ , i.e., $S_{\theta}$ and $S_{\xi}$ for the student and teacher networks, respectively, to help recognize the foreground and background in each image. $S_{\xi}$ is also optimized by EMA. + +Unlike the classical slot attention method, $s_{\theta \text{ or } \xi}^{f}$ and $s_{\theta \text{ or } \xi}^{b}$ are given semantic meaning, enabling the classification of pixels into their corresponding semantic groups, such as foreground and background. Here, $s_{\theta}^{f}$ and $s_{\theta}^{b}$ represent the slots for foreground and background in the student network, respectively, while $s_{\xi}^{f}$ and $s_{\xi}^{b}$ represent the slots for foreground and background in the teacher network, respectively. However, we need an initialized boot mode to provide the initialized semantic cues for foreground objects. + +Pre-trained models have proven effective in providing initial semantic information across various fields [19]. Therefore, we employ the pre-trained model as the backbone of our method instead of training from scratch. + +Our indicator contains three aspects. The first is the pixel-level consistency. For different augmentations of the same image, we ensure that pixels in the same location receive similar assignment scores. Given an input image $x$ , two random augmentations are applied to produce two augmented views $v^{1}(x)$ and $v^{2}(x)$ . Each augmentation view is then encoded with an encoder $f$ into a hidden feature map and then transformed with a multilayer perceptron (MLP) $g$ to get the feature map $z$ , i.e., $z_{\theta}(x) = g_{\theta}(f_{\theta}(v^{1}(x)))$ and $z_{\xi}(x) = g_{\xi}(f_{\xi}(v^{2}(x)))$ . The fore-background representation vectors of the image $x$ in different networks can be obtained via slot attention [23], i.e., $\hat{S}_{\theta}(x) =$ + +slot attention $(z_{\theta}(x),S_{\theta})$ $\hat{S}_{\xi} =$ slot attention $(z_{\xi}(x),S_{\xi})$ Then, the pixel assignment matrix, defining the score fore or background each pixel belongs to, $\mathcal{P}_{\theta}$ and $\mathcal{Q}_{\xi}$ can obtain by $\mathcal{P}_{\theta}(x) = z_{\theta}\cdot \hat{S}_{\theta}^{\top}$ . The assignment matrix $\mathcal{Q}_{\xi}(x)$ for the teacher network can be obtained in the same way. + +To keep the pixel level consistency, we enforce the prediction of the two views assignment $\mathcal{P}_{\theta}(x)$ and $\mathcal{Q}_{\xi}(x)$ match each other. However, the geometric image augmentations, such as random crop, scale, or flip, between different views will cause the scale or layout of the two feature maps to be inconsistent. So similar to [35], we use the inverse augmentation process on the predicted assignments to align their spatial locations, i.e., $\tilde{\mathcal{P}}_{\theta}(x) = \mathrm{invaug}(\mathcal{P}_{\theta}(x)), \tilde{\mathcal{Q}}_{\xi}(x) = \mathrm{invaug}(\mathcal{Q}_{\xi}(x))$ + +Based on the aligned assignments, we apply the cross-entropy loss to enforce the consistency in assignment scores between spatial-aligned pixels from different views. The cross-entropy loss is averaged over all spatial locations to keep the pixel prediction consistency + +$$ +\mathcal {L} ^ {\text {p i x e l}} (x) = \frac {1}{H \times W} \sum_ {i, j} \mathcal {L} ^ {\mathrm {C E}} \left(\tilde {\mathcal {Q}} _ {\xi} (x) [ i, j ], \tilde {\mathcal {P}} _ {\theta} (x) [ i, j ]\right). \tag {1} +$$ + +$i,j$ means the position in the assignment matrix. + +In addition to the pixel-level contrastive loss, we incorporate a stuff-level consistency loss. This loss aims to alleviate variations in the background by treating the background in different images as positive samples, while considering all foreground samples as negative samples. According to the pixel assignment matrix $\tilde{\mathcal{P}}_{\theta}(x)$ and $\tilde{\mathcal{Q}}_{\xi}(x)$ , we can obtain the prototypes of the foreground and background by region average pooling to represent the whole foreground and background regions of the image $x$ . $p_{\theta}^{f}(x), p_{\theta}^{b}(x)$ represent the foreground and background prototypes obtained from the student network. $p_{\xi}^{f}(x), p_{\xi}^{b}(x)$ are from the teacher network. Then, the stuff-level loss can be written as follows: + +$$ +\begin{array}{l} \mathcal {L} ^ {\text {s t u f f}} (B) = \frac {1}{B (B - 1)} \sum_ {i = 1} ^ {B} \sum_ {j \neq i} ^ {B} [ \operatorname {s i m} \left(p _ {\theta} ^ {b} \left(x _ {i}\right), p _ {\xi} ^ {b} \left(x _ {j}\right)\right) - \\ \left. \left. \operatorname {s i m} \left(p _ {\theta} ^ {b} \left(x _ {i}\right), p _ {\xi} ^ {f} \left(x _ {j}\right)\right) \right]. \right. \tag {2} \\ \end{array} +$$ + +In Eq. 2, we hope the similarity between the background across different views will be large, however, the similarity between foreground and background should be small. We adopt the cosine similarity in our method. $B$ means the batch of training data. + +To encourage the semantic centers to emphasize different visual contents and cover diverse semantics to avoid collapse, we apply a regularization to the assignment matrix of + +different views, as follows: + +$$ +\mathcal {L} ^ {\operatorname {s e p}} (x) = \operatorname {E n t r o y} \left(\mathcal {P} _ {\theta} (x)\right) + \operatorname {E n t r o p y} \left(\mathcal {Q} _ {\xi} (x)\right). \tag {3} +$$ + +The total loss to optimize the proposed indicator is + +$$ +\mathcal {L} = \mathcal {L} ^ {\text {p i x e l}} (x) - \lambda \mathcal {L} ^ {\text {s t u f f}} (B) - \gamma \mathcal {L} ^ {\text {s e p}} (x). \tag {4} +$$ + +# 3.2. Fusion of information into original slot attention method + +The features provided by our indicator allow for clear distinction between the foreground object and background elements. Simultaneously, the background is processed uniformly, harmonizing its characteristics across various regions. Integrating the features obtained through our method with existing approaches significantly enhances the differentiation between the foreground and background. Furthermore, once trained, our indicator is frozen and only provides supplementary information to the existing slot attention-based methods. + +In our fusion stage, we first train the slot attention-based method for several epochs. Once it has learned meaningful information, we next integrate the foreground and background information from our indicator into the slot attention-based method as supplementary information. Finally, the features from both the original slot attention-based method and our indicator are fused together and input into the next decoder for image reconstruction. The feature fusion stage can be written as follows: + +$$ +z ^ {\text {f u s i o n}} (x) = \operatorname {M L P} \left(z ^ {\text {s l o t}} (x) + z _ {\xi} ^ {\text {i n d i c a t o r}} (x)\right), \tag {5} +$$ + +where $z^{\mathrm{slot}}(x)$ is the feature map from the encoder in the original slot attention method. $z_{\xi}^{\mathrm{indicator}}(x)$ is the feature map provided by our indicator. $x$ is the input image. These two feature maps are added together and then passed through a MLP layer to transform the combined feature map into a new space. The resulting fusion feature $z^{\mathrm{fusion}}(x)$ replaces the feature map $z^{\mathrm{slot}}(x)$ used in the original slot attention-based method to do the next stages, such as slot attention, and decoding, which are the same as the original slot attention framework. + +# 3.3. Region Combination + +Defining the number of slots is a common challenge in recent slot attention methods [10] and can significantly impact their performance. Determining the appropriate number of slots is not straightforward, and recent approaches often opt for a higher number of slots, leading to slight oversegmentation, in order to achieve satisfactory results. In this section, we propose a spectral clustering-based method to group regions likely belonging to the same semantic object, which is used to refine the predictions during the testing phase. + +The similarity between slots represents the similarity between regions in the image. The similarity matrix between + +different slots can be regarded as a reference to help recognize which slots possibly belong to the same semantic object. In our method, we directly use spectral clustering to divide the slots into different groups based on this matrix. Specifically, we calculate the cosine similarity between each pair of the slots to construct the similarity matrix $\mathbf{M}$ . Then according to $\mathbf{M}$ , the group results can be obtained by + +$$ +\mathcal {G} = \text {s p e c t r a l c l u s t e r i n g} (\mathbf {M}, N). \tag {6} +$$ + +Based on the group result $\mathcal{G}$ , the corresponding areas of slots in the same group can be combined. To determining the group number $N$ , we primarily employ eigen-gap heuristic [32]. In this method, we use the Laplacian matrix of the graph and find the largest gap between consecutive eigenvalues, which suggests the optimal number of clusters. + +# 4. Experiments + +# 4.1. Setup + +Datasets. We evaluate our method on one synthetic and two real-world datasets. For the synthetic dataset, we use MOVi-C [13], which contains approximately 1,000 real 3D-scanned objects, with each scene featuring between 3 and 10 objects. Although MOVi-C is originally a video dataset, we convert it into an image dataset by randomly selecting frames following [28]. For the real-world datasets, we utilize MS COCO 2017 [21] and PASCAL VOC 2012 [9], which are commonly used for object detection and segmentation. PASCAL VOC 2012 contains many images with a single salient object, while MS COCO 2017 features more complex scenes, offering a more challenging evaluation. + +Metrics. To assess object-centric learning, we used Mean Best Overlap (mBO) and Mean Intersection over Union (mIoU). The Mean Best Overlap (mBO) includes instance-level $\mathrm{mBO}^i$ and category-level $\mathrm{mBO}^c$ . $\mathrm{mBO}^i$ and $\mathrm{mBO}^c$ match the predicted masks to the instance and semantic masks, respectively, by finding the best overlap. Meanwhile, mIoU employs the Hungarian algorithm to match predicted and ground truth masks. [17] suggests that FG-ARI, as a clustering similarity metric, may yield unreliable segmentation quality assessments by overlooking the accuracy of predicted masks. Therefore, FG-ARI is not reported in this paper. + +Implementation Details. To optimize the proposed foreground and background indicator, we employ the SGD optimizer. The learning rate is 0.001. Batch size is 512. The learning rate follows a linear warm-up for 10,000 optimization steps and an exponential decay schedule. Additionally, we clip the gradient norm to 1 to stabilize the training process. ViT-B/16 [1], initialized with DINO [5], is used as the encoder in our indicator. During the fusion stage, the model is trained using the same configuration as the corresponding + +original slot attention method you choose. All models are trained on a single A6000 GPU with 48 Gbytes. + +# 4.2. Integrating into SOTA object-centric methods + +Baselines. We apply our approach to two slot attention methods, DINOSAUR [28] and SPOT [17]. DINOSAUR utilizes a Vision Transformer encoder for feature reconstruction and has demonstrated applicability to real-world datasets. SPOT is a recent SOTA object-centric model that employs a two-stage self-training scheme and a patch-order permutation strategy for the transformer decoder. Furthermore, we also list the results of several other object-centric methods, including Slot Attention [23], SLASH [18], SLATE [29], SlotDiffusion [36], and LSD [16]. A more detailed description of the baseline methods can be found in the Appendix. + +Main Results on Each Dataset. Tab. 1 presents the main results of our method on the COCO, PASCAL, and MOViC datasets. Our method improves the performance of DINOSAUR and SPOT across all datasets. The fusion stage yields consistent improvements, while the region combination stage leads to a slight decrease in the $\mathrm{mBO}^i$ metric on the PASCAL and COCO datasets. This is due to the fact that, during clustering, slots corresponding to two instances of the same category may be merged into one, resulting in an increase in $\mathrm{mBO}^c$ but a decrease in $\mathrm{mBO}^i$ . Despite this, our method still outperforms the baseline models, DINOSAUR and SPOT, demonstrating a clear overall performance advantage. This highlights the crucial role of the fusion stage in driving performance improvements, while the region combination stage primarily serves to refine the predictions. + +# 4.3. More discussions + +In this section, we analyze various aspects of our approach: the significance of the foreground and background indicators, the impact of different loss functions within the indicator, the role of the fusion stage, and the consistency of performance improvement across different slot numbers. + +(A) Does the extra foreground and background information really improve the slot attention method? We evaluate the slot attention method in four cases. Tab. 2 shows the results. In this table, "Random" refers to using random noise as the input to the fusion stage, "ViT" refers to directly using the pre-trained DINO ViT as the indicator, and "Ours" refers to our foreground and background indicator. We find that when the indicator is random noise, the baseline method struggles to distinguish between foreground and background, leading to a significant performance drop. When ViT is used as the indicator, the pretrained ViT has some inherent ability to differentiate between foreground and background, resulting in a slight performance improvement. Our designed foreground + +
MethodMOVi-CPASCALCOCO
mBOimIoUmBOimBOcmBOimBOc
Slot Attention [23] [NeurIPS20]26.3-24.624.917.219.2
SLASH [18] [CVPR23]-27.7----
SLATE [29] [ICLR22]39.437.835.941.529.133.6
SlotDiffusion [36] [NeurIPS23]--50.455.331.035.0
(Stable-)LSD [16] [NeurIPS23]45.644.2--30.4-
DINOSAUR [28] [ICLR23]42.441.844.051.231.138.9
DINOSAUR + FS (Ours)45.3 ↑ 2.944.5 ↑ 2.745.4 ↑ 1.452.5 ↑ 1.331.5 ↑ 0.439.9 ↑ 1.0
DINOSAUR + FS + RC (Ours)46.0 ↑ 3.644.8 ↑ 3.045.2 ↑ 1.252.9 ↑ 1.731.2 ↑ 0.140.1 ↑ 1.2
SPOT [17] [CVPR24]47.046.448.155.334.744.3
SPOT + FS (Ours)48.5 ↑ 1.547.7 ↑ 1.349.1 ↑ 1.056.3 ↑ 1.035.5 ↑ 0.845.0 ↑ 0.7
SPOT + FS + RC (Ours)49.0 ↑ 2.047.8 ↑ 1.449.3 ↑ 1.256.5 ↑ 1.235.7 ↑ 1.045.3 ↑ 1.0
+ +Table 1. Comparison with different object-centric methods on MOVi-C, PASCAL, and COCO datasets. DINOSAUR uses an autoregressive decoder and DINO [5] ViT encoder. The experimental results of the baseline are derived from the original corresponding papers. FS means using the fusion stage in our paper. RC means the trick of region combination. + +
SA MethodIndicatorIoUmBOcmBOi
DINOSAUR--51.244.0
Random-25.222.1
ViT30.251.444.7
Ours52.652.545.4
+ +Table 2. The effect of extra information on the existing slot attention-based method. All experiments are conducted on PASCAL dataset. IoU is used to show the model's performance on segmenting the foreground and background. $\mathrm{mBO}^c$ and $\mathrm{mBO}^i$ are employed to assess the final performance after fusing the extra knowledge from the indicator. + +
MethodMOVi-CPASCALCOCO
DINOSAUR (2 slots)45.247.646.9
SPOT (2 slots)46.650.846.4
Ours47.452.648.4
+ +Table 3. The performance of using object-centric methods as the foreground and background indicator. Fig. 3 shows some visualization results. + +and background indicator makes the foreground and background highly separable and significantly improves performance. + +(B) How about using the classical slot attention methods as the foreground and background indicator in our method? Tab. 3 evaluates the performance of different models when the number of slots is set to 2. We continue to use the Mean Best Overlap (mBO) as the evaluation metric. The difference is that, during metric computation, we convert the ground truth masks into binary masks containing only foreground (0) and background (1). Then, the pre + +dicted masks are matched with the ground truth masks using the best overlap matching method. + +From the data in Tab. 3, it can be observed that our method achieves the best performance with significant performance improvements on both COCO and PASCAL datasets. This is one of the reasons for using the foreground and background indicator. Another reason is that autoencoder-based methods like DINOSAUR and SPOT, while capable of dividing the scene into two parts, struggle to achieve accurate foreground-background segmentation. As shown in Fig. 3, the baseline methods tend to divide the scene into two parts, top and bottom, while our method focuses on separating the foreground and background. This is because our method employs both pixel-level and stufflevel contrastive learning, enabling the model to better distinguish foreground from background features, leading to a more effective foreground-background decomposition. + +(C) Impact of different losses. We conduct the ablation experiments to further analyze the role of different losses in the indicator, as shown in Tab. 4. We find that without using $\mathcal{L}^{\mathrm{sep}}$ , the indicator struggle to achieve a performance improvement. This is because when only $\mathcal{L}^{\mathrm{pixel}}$ and $\mathcal{L}^{\mathrm{stuff}}$ are used, the model tends to collapse all information into a single slot, where one slot binds to the entire image and the other slot remains empty. In this case, the model can achieve low $\mathcal{L}^{\mathrm{pixel}}$ and $\mathcal{L}^{\mathrm{stuff}}$ values, but it cannot effectively perform the foreground-background separation. Therefore, we introduce $\mathcal{L}^{\mathrm{sep}}$ to encourage semantic centers to emphasize different visual content and cover diverse semantics. However, relying on $\mathcal{L}^{\mathrm{sep}}$ and $\mathcal{L}^{\mathrm{pixel}}$ makes it difficult for the model to adapt to variations in the background. As a result, $\mathcal{L}^{\mathrm{sep}}$ , $\mathcal{L}^{\mathrm{pixel}}$ , and $\mathcal{L}^{\mathrm{stuff}}$ must work together in synergy to achieve effective foreground-background separation. + +At the same time, it can be seen that the performance of + +![](images/740561010922650beec7cb8d634406b8ea0711dc76fba8e4c20c5511fa668675.jpg) +Figure 3. Example results of the proposed indicator and SOTA slot attention method with 2 slots. Our method distinguishes foreground and background better. + +the indicator is positively correlated with the final results. This further confirms the conclusion in (A): the stronger the indicator's ability to distinguish between foreground and background, the more effectively it enhances the feature fusion process. + +
LpixelLstuffLsepIoUmBOcmBOi
✓28.851.444.6
✓✓28.451.344.7
✓✓48.951.744.8
✓✓✓52.652.545.4
+ +Table 4. Analysis of different loss functions on PASCAL dataset. IoU represents the results of the indicator on segmenting foreground and background regions. $\mathrm{mBO}^c$ and $\mathrm{mBO}^i$ represent the final results of DINOSAUR after fusing knowledge from different indicators, which are optimized using different types of loss functions. + +(D) The importance of the fusion stage and region combination stage. We conduct a detailed analysis of the importance of the fusion stage and the region combination stage. The results are summarized in Table 5. We observe that directly applying the region combination to DINOSAUR and SPOT typically leads to a decline in performance metrics. We hypothesize that this is due to the limited ability of the baseline methods to differentiate between foreground and background. Without a clear distinction, the region combination stage inappropriately merges foreground and background. The visualization results are shown in Figure 4. + +On the MOVi dataset, applying region combination directly to SPOT results in a performance improvement. This + +![](images/2e184b3c267ddaebe0499b2e2e14785a29e8c205b79f2f15209ad3fb439d1bc6.jpg) +Figure 4. The visualization results in our different components. FS means fusion stage, and RC is the stage of region combination. + +may result from the relatively simple background in the MOVi dataset, where SPOT is already effective at distinguishing between foreground and background. As shown in Tab. 3, the metric for SPOT with 2 slots closely aligns with our indicator result, further supporting this observation. + +(E) Consistent improvement across different slot numbers. The quality of results is influenced by the number of slots used during model training. Therefore, understanding the model's robustness to variations in the number of slots is crucial. In this part, we compare our method with DINOSAUR to evaluate the stability of our method with different numbers of slot. We conduct experiments on two real-world datasets, COCO and PASCAL, using five slot numbers, with the results presented in Tab. 6. Notably, our method outperforms DINOSAUR across all slot numbers and exhibits lower variance. This suggests that, while improving baseline performance, our method also demonstrates better generalization ability when applied with different numbers of slots. + +# 4.4. Qualitative results + +Fig. 5 shows example results on the COCO and PASCAL datasets for DINOSAUR, SPOT, and our method. Row (a) demonstrates that, compared to baseline methods, our approach correctly fuses the foreground and background. For example, in the PASCAL dataset, our method successfully merges the ground that was originally divided into two separate slots. Rows (b) and (c) illustrate that when the baseline method accurately segments the background (foreground), our method correctly fuses the foreground (background) regions. For instance, in row (b) of the PASCAL dataset, when the background is correctly segmented, our method successfully merges the train and streetlight, which were + +
MethodMOVi-CPASCALCOCO
mBOimIoUmBOimBOcmBOimBOc
DINOSAUR42.441.844.051.231.138.9
DINOSAUR+RC41.6 ↓0.839.5 ↓2.343.3 ↓0.750.7 ↓0.530.4 ↓0.739.4 ↑0.5
DINOSAUR+FS (Ours)45.3 ↑2.944.5 ↑2.745.4 ↑1.452.5 ↑1.331.5 ↑0.439.9 ↑1.0
DINOSAUR+FS+RC (Ours)46.0 ↑3.644.8 ↑3.045.2 ↑1.252.9 ↑1.731.2 ↑0.140.1 ↑1.2
SPOT47.046.448.155.334.744.3
SPOT+RC47.8 ↑0.847.1 ↑0.748.0 ↓0.154.8 ↓0.533.2 ↓1.543.5 ↓0.8
SPOT+FS (Ours)48.5 ↑1.547.7 ↑1.349.1 ↑1.056.3 ↑1.035.5 ↑0.845.0 ↑0.7
SPOT+FS+RC (Ours)49.0 ↑2.047.8 ↑1.449.3 ↑1.256.5 ↑1.235.7 ↑1.045.3 ↑1.0
+ +![](images/70293c6a398e4134211456b21d6ede2a6e1795090f14afad9a26ea04f1b791d5.jpg) +Figure 5. Qualitative results on COCO and PASCAL datasets. Ours: fusion stage + region combination. + +![](images/17765ede36fb5b66311932a5510a2a0ebd09151a5bf6c3f1262dcf05e5cc0e6d.jpg) + +Table 5. The importance of the different components in our method. FS is the fusion stage and RS is the region combination stage. Using RS alone even hurt the performance. FS is crucial. + +
DatasetKDINOSAURDINOSAUR+FS
mBOimBOcmBOimBOc
COCO227.620.530.2 ↑ 2.622.2 ↑ 1.7
537.929.339.3 ↑ 1.430.3 ↑ 1.0
739.031.139.9 ↑ 0.931.5 ↑ 0.4
938.531.739.8 ↑ 1.332.0 ↑ 0.3
1138.031.639.5 ↑ 1.532.6 ↑ 1.0
PASCAL234.340.435.4 ↑ 1.141.6 ↑ 1.2
444.351.644.6 ↑ 0.352.0 ↑ 0.4
644.851.845.4 ↑ 0.652.5 ↑ 0.7
845.351.745.5 ↑ 0.252.3 ↑ 0.6
1045.451.345.6 ↑ 0.252.0 ↑ 0.7
+ +Table 6. The performance of our method with different numbers of slot. For a fair comparison, only the fusion stage is adopted. $K$ represents the number of slots. + +originally assigned to two separate slots. + +# 4.5. Training cost + +The additional training cost of our method primarily consists of two components. First, training the indicator. Due + +to its simple structure, training on the PASCAL dataset using a single A6000 GPU takes approximately 3 hours. Once trained, the indicator can be directly applied to different slot attention methods on the same dataset without the need for retraining. Second, retraining the existing slot attention methods based on the fused features, where the training time depends on the chosen slot attention method. + +# 5. Conclusion + +In this paper, we consider the problem that complex backgrounds in images can disrupt the phase of slot binding. Hence, we introduce a contrastive learning-based indicator to provide extra information about the foreground and background region to the existing slot attention method. Integrating this indicator into the slot attention-based method allows the model to focus more effectively on segmenting foreground objects and minimizing background distractions. According to the similarity between the slots, we also use a spectral clustering mechanism to refine the results in the testing phase. In experiments, we integrate our method into two SOTA slot attention methods. All the results show that our method can consistently enhance the performance on different datasets. + +# References + +[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.5 +[2] Christopher P Burgess, Loic Matthew, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. 2 +[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2 +[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 6 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2 +[9] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 5 +[10] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23062-23071, 2024. 4 +[11] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In International conference on machine learning, pages 2424-2433. PMLR, 2019. 2 +[12] Klaus Greff, Sjoerd Van Steenkiste, and Jürgen Schmidhuber. On the binding problem in artificial neural networks. arXiv preprint arXiv:2012.05208, 2020. 1 +[13] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapra + +gasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 5 +[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[15] A Jaiswal, A Ramesh Babu, M Zaki Zadeh, D Banerjee, and F Makedon. A survey on contrastive self-supervised learning, 2020. 2 +[16] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 8563-8601, 2023. 2, 5, 6 +[17] Ioannis Kakogeorgiou, Spyros Gidaris, Konstantinos Karantzalos, and Nikos Komodakis. Spot: Self-training with patch-order permutation for object-centric learning with autoregressive transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22776-22786, 2024. 2, 5, 6 +[18] Jinwoo Kim, Janghyuk Choi, Ho-Jin Choi, and Seon Joo Kim. Shepherding slots to objects: Towards stable and robust object-centric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19198-19207, 2023. 5, 6 +[19] Hyeokjun Kweon, Sung-Hoon Yoon, and Kuk-Jin Yoon. Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11329-11339, 2023. 3 +[20] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 40:e253, 2017. 1 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[22] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection without imagenet. arXiv preprint arXiv:2011.13677, 2020. 2 +[23] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 1, 2, 3, 5, 6 +[24] Amin Mansouri, Jason Hartford, Yan Zhang, and Yoshua Bengio. Object-centric architectures enable efficient causal representation learning. arXiv preprint arXiv:2310.19054, 2023. 1 +[25] Gary F Marcus. The algebraic mind: Integrating connectionism and cognitive science. MIT press, 2003. 1 + +[26] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020. 2 +[27] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16675-16687, 2023. 1 +[28] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. In The 11th International Conference on Learning Representations, 2023. 2, 5, 6 +[29] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dall-e learns to compose. In 10th International Conference on Learning Representations, ICLR 2022, 2022. 2, 5, 6 +[30] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020. 2 +[31] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. Advances in Neural Information Processing Systems, 34: 16238-16250, 2021. 2 +[32] Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17:395-416, 2007. 5 +[33] Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17:395-416, 2007. 2 +[34] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 2 +[35] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. Advances in neural information processing systems, 35:16423-16438, 2022. 4 +[36] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotdiffusion: Object-centric generative modeling with diffusion models. Advances in Neural Information Processing Systems, 36:50932-50958, 2023. 2, 5, 6 +[37] Ceyuan Yang, Zhirong Wu, Bolei Zhou, and Stephen Lin. Instance localization for self-supervised detection pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3987-3996, 2021. 2 +[38] Ying Zheng, Lei Yao, Yuejiao Su, Yi Zhang, Yi Wang, Sicheng Zhao, Yiyi Zhang, and Lap-Pui Chau. A survey of embodied learning for object-centric robotic manipulation. arXiv preprint arXiv:2408.11537, 2024. 1 +[39] Rui Zhu, Bingchen Zhao, Jingen Liu, Zhenglong Sun, and Chang Wen Chen. Improving contrastive learning by visualizing feature transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10306-10315, 2021. 2 \ No newline at end of file diff --git a/2025/Pay Attention to the Foreground in Object-Centric Learning/images.zip b/2025/Pay Attention to the Foreground in Object-Centric Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7b4b6a89b41252ad0170dc8d442fb7dafd4e038b --- /dev/null +++ b/2025/Pay Attention to the Foreground in Object-Centric Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc464e6b5a18e53c0800f5c850c9043af941954f92a7cc184809277508655037 +size 694924 diff --git a/2025/Pay Attention to the Foreground in Object-Centric Learning/layout.json b/2025/Pay Attention to the Foreground in Object-Centric Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38d84eeec8ee126a04f68ab29445c714c632230e --- /dev/null +++ b/2025/Pay Attention to the Foreground in Object-Centric Learning/layout.json @@ -0,0 +1,7283 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 493, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 493, + 121 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 493, + 121 + ], + "type": "text", + "content": "Pay Attention to the Foreground in Object-Centric Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 172, + 142, + 436, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 142, + 436, + 157 + ], + "spans": [ + { + "bbox": [ + 172, + 142, + 436, + 157 + ], + "type": "text", + "content": "Pinzhuo Tian" + }, + { + "bbox": [ + 172, + 142, + 436, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 172, + 142, + 436, + 157 + ], + "type": "text", + "content": ", Shengjie Yang" + }, + { + "bbox": [ + 172, + 142, + 436, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 172, + 142, + 436, + 157 + ], + "type": "text", + "content": ", Hang Yu" + }, + { + "bbox": [ + 172, + 142, + 436, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 172, + 142, + 436, + 157 + ], + "type": "text", + "content": ", Alex Kot" + }, + { + "bbox": [ + 172, + 142, + 436, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 157, + 487, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 157, + 487, + 171 + ], + "spans": [ + { + "bbox": [ + 123, + 157, + 487, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 157, + 487, + 171 + ], + "type": "text", + "content": "School of Computer Engineering and Science, Shanghai University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "spans": [ + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "type": "text", + "content": "School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 156, + 188, + 455, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 188, + 455, + 199 + ], + "spans": [ + { + "bbox": [ + 156, + 188, + 455, + 199 + ], + "type": "text", + "content": "{pinzhuo, sjyang, yuhang}@shu.edu.cn, eackot@ntu.edu.sg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 266, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 266, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 296, + 553 + ], + "type": "text", + "content": "The slot attention-based method is widely used in unsupervised object-centric learning, aiming to decompose scenes into interpretable objects and associate them with slots. However, complex backgrounds in the real images can disrupt the model's focus, leading it to excessively segment background stuff into different regions based on low-level information such as color or texture variations. As a result, the detailed segmentation of foreground objects, which requires shape or geometric information, is often neglected. To address this issue, we introduce a contrastive learning-based indicator designed to differentiate between foreground and background. Integrating this indicator into the existing slot attention-based method enables the model to focus more on segmenting foreground objects while minimizing background distractions. During the testing phase, we utilize a spectral clustering mechanism to refine the results based on the similarity between the slots. Experimental results show that incorporating our method with various state-of-the-art models significantly improves their performance on both simulated data and real-world datasets. Furthermore, multiple sets of ablation experiments confirm the effectiveness of each proposed component. The source code is available at https://github.com/sjyjs09/FG-BG_Indicator." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 597, + 295, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 657 + ], + "type": "text", + "content": "In perceiving the surrounding environment, humans typically divide it into distinct objects, subsequently understanding the scene through the relationships and connections among these objects [20, 25]. Thus, enabling our model to decompose the visual scene into different objects" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 239, + 555, + 344 + ], + "blocks": [ + { + "bbox": [ + 315, + 239, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 315, + 239, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 315, + 239, + 555, + 344 + ], + "type": "image", + "image_path": "7a6d541043eb5956d76897096cf7b62ad91251e038a6c2a5b2a2f47e7b91adfb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 352, + 555, + 462 + ], + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 462 + ], + "type": "text", + "content": "Figure 1. The overview of our method. The background stuff in the complex scenes can influence the process of slot binding. Variations in color or texture within the background may cause the slot to incorrectly segment the background into different regions while neglecting the foreground (shown in the above branch). In our method, we integrate information from our indicator to enhance the distinction between foreground and background areas and mitigate changes in background region, enabling the slot attention method to focus more effectively on foreground object segmentation (shown in the below branch)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 544 + ], + "type": "text", + "content": "is highly meaningful and lays the foundation for reasoning [24]. Unsupervised object-centric learning [12] is an approach designed to address this challenge, aiming to represent a scene as a composition of distinct objects using only visual information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 676 + ], + "type": "text", + "content": "After significant development, the slot attention-based method [23] has become the mainstream framework in the field. This kind of method can be divided into two primary stages: the binding stage and the reconstruction stage. In the binding stage, an attention-based mechanism is used to assign different areas of the image to distinct slots. The reconstruction stage then involves rebuilding the input based on information derived from these slots, ensuring that each slot accurately represents an individual object within the image. These two stages enable the slot attention-based method to achieve impressive results in various areas, such as [27, 38]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "However, current slot attention-based methods do not distinguish between foreground objects and background stuff. When applied to complex scenes, variations in color" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 665, + 148, + 674 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 665, + 148, + 674 + ], + "spans": [ + { + "bbox": [ + 69, + 665, + 148, + 674 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 674, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 295, + 712 + ], + "type": "text", + "content": "This work is supported by projects of the National Natural Science Foundation of China (Grant No.62206166, No.62302287), Shanghai Sailing Program (Grant No.23YF1413000), and Shanghai Committee of Science and Technology (Grant No.23ZR1423500)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30281" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "content": "or texture within the background can confuse the slots. This misdirection often causes the slots to divide the background into different regions based on low-level visual cues. As a result, the slots struggle to bind to foreground objects, which require a more nuanced recognition of shape and geometric details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 326 + ], + "type": "text", + "content": "To address this issue, we propose an unsupervised indicator that provides crucial information to differentiate between foreground and background. After training, it can be incorporated into various slot attention methods without requiring retraining. Figure 1 shows an overview of our method. Our unsupervised indicator is based on a contrastive learning framework [15] and consists of three components: a pixel-level contrastive loss that ensures consistency in pixel predictions across different views to promote meaningful representation learning; a stuff-level contrastive loss that minimizes differences within the background, reducing misguidance caused by color and texture variations in the background during slot binding; and an entropy-based regularization that enhances the distinction between foreground and background elements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 554 + ], + "type": "text", + "content": "By incorporating information from this indicator, the slot attention-based method shifts its focus toward more accurate object segmentation while minimizing background distractions. However, due to the absence of supervised ground truth, the indicator cannot perfectly differentiate between foreground and background regions, which may introduce noise or incorrect labels. To mitigate the impact of this noise during training, we first train the original slot attention-based method for several epochs, then incorporate the foreground and background information from the indicator. During the testing phase, recognizing that the similarity between slots can reflect the similarity between image regions, we employ a spectral clustering merging mechanism [33] to refine the results and reduce oversegmentation. In our experiments, we integrate our method with two prominent object-centric learning approaches, DINOSAUR [28] and SPOT [17]. The results demonstrate that our approach improves their performance on both simulated and real-world datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 556, + 247, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 556, + 247, + 567 + ], + "spans": [ + { + "bbox": [ + 67, + 556, + 247, + 567 + ], + "type": "text", + "content": "In summary, our contributions are threefold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 570, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 570, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 629 + ], + "type": "text", + "content": "- We consider the issue that variations in background information, like color or texture, can mislead the slot allocation process and lead to the neglect of foreground object segmentation, which requires detailed shape and boundary information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 630, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 295, + 677 + ], + "type": "text", + "content": "- We propose an unsupervised foreground and background indicator to solve this problem and employ a spectral clustering merging mechanism that utilizes the similarity between slots to refine the results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 678, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 678, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 678, + 295, + 713 + ], + "type": "text", + "content": "- Our experimental results demonstrate that integrating our method with various state-of-the-art models significantly enhances their performance both on simulated data and" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 403, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 403, + 83 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 403, + 83 + ], + "type": "text", + "content": "real-world datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 95, + 399, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 95, + 399, + 108 + ], + "spans": [ + { + "bbox": [ + 314, + 95, + 399, + 108 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 116, + 553, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 116, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 553, + 272 + ], + "type": "text", + "content": "Object-centric learning. Object-centric learning aims to represent scenes as compositions of distinct objects. MONet [2] and IODINE [11] are early studies in this area. However, they rely on multiple encode-decode steps. Slot Attention [23] achieves efficient scene decomposition by replacing this process with a single encoding step that iterates attention. Although Slot Attention has gradually become a major paradigm in object-centric learning, this approach still faces challenges when applied to real-world scenes. To address this issue, some studies focus on improving the encoder, like DINOSAUR [28]. Other studies focus on designing more effective decoders for scene decomposition, like SLATE [29], SPOT [17], and [16, 36]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 272, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 555, + 464 + ], + "type": "text", + "content": "Contrastive learning. Contrastive learning is a mainstream paradigm in self-supervised learning (SSL), aiming to learn consistent representations by contrasting positive and negative sample pairs without the need for additional annotations [4, 7, 8]. Classical contrastive learning methods treat the image as a whole part and have shown strong potential in learning transferable visual representations [6, 14, 30, 39]. However, real-world scenes often contain complex and diverse information, making it difficult for global features to represent all the information. To address this issue, some works leverage local crops [3, 31, 37], while others gradually move toward pixel-level contrastive learning. Pixel-level contrastive learning extends the task of instance discrimination from image-level feature vectors to feature maps [22, 26, 34], which naturally aids in learning dense representations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 475, + 370, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 475, + 370, + 487 + ], + "spans": [ + { + "bbox": [ + 314, + 475, + 370, + 487 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 495, + 553, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 553, + 579 + ], + "type": "text", + "content": "Our approach consists of three key procedures: 1) the foreground and background indicator, which supplies additional information about foreground objects and background elements, 2) the fusion module, responsible for integrating information from the indicator into the existing slot attention method, 3) the spectral clustering module, which merges regions to refine prediction results in the test stage." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 588, + 517, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 517, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 517, + 601 + ], + "type": "text", + "content": "3.1. Foreground and background indicator" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 653 + ], + "type": "text", + "content": "This part aims to achieve two goals: 1) it needs to be an unsupervised method, and 2) it should provide an accurate and robust representation to distinguish the foreground and background." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Hence, the implementation of our indicator leverages advancements in the contrastive learning framework, like teacher-student framework. Specifically, our approach involves a student network, denoted by " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ", and a teacher network, denoted by " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ". Both networks share the same architec" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30282" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 541, + 453 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 541, + 453 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 541, + 453 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 541, + 453 + ], + "type": "image", + "image_path": "5a7f13e025d4f274a7e2c89ce4aac2acaffbafef8174c72c522e2abd1826b08a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 457, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 55, + 457, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 555, + 492 + ], + "type": "text", + "content": "Figure 2. Illustration of our pipeline. During the training stage, the foreground and background indicator are trained to provide highly distinguishable foreground and background features (the part of (a)). Then, the feature from the indicator is integrated into the original slot attention-based methods (the part of (b)). In the testing stage, we use region combination to refine the results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "text", + "content": "ture but use different sets of weights. The parameters of the teacher network, " + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "text", + "content": ", are updated as an exponential moving average of the student network, " + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "text", + "content": ". In each network, we use two distinct slots, " + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "inline_equation", + "content": "S = \\{s^f, s^b\\}" + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "inline_equation", + "content": "S_{\\xi}" + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "text", + "content": " for the student and teacher networks, respectively, to help recognize the foreground and background in each image. " + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "inline_equation", + "content": "S_{\\xi}" + }, + { + "bbox": [ + 55, + 511, + 295, + 595 + ], + "type": "text", + "content": " is also optimized by EMA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": "Unlike the classical slot attention method, " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "s_{\\theta \\text{ or } \\xi}^{f}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "s_{\\theta \\text{ or } \\xi}^{b}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " are given semantic meaning, enabling the classification of pixels into their corresponding semantic groups, such as foreground and background. Here, " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "s_{\\theta}^{f}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "s_{\\theta}^{b}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " represent the slots for foreground and background in the student network, respectively, while " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "s_{\\xi}^{f}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "inline_equation", + "content": "s_{\\xi}^{b}" + }, + { + "bbox": [ + 55, + 603, + 296, + 715 + ], + "type": "text", + "content": " represent the slots for foreground and background in the teacher network, respectively. However, we need an initialized boot mode to provide the initialized semantic cues for foreground objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 511, + 555, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 555, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 555, + 560 + ], + "type": "text", + "content": "Pre-trained models have proven effective in providing initial semantic information across various fields [19]. Therefore, we employ the pre-trained model as the backbone of our method instead of training from scratch." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Our indicator contains three aspects. The first is the pixel-level consistency. For different augmentations of the same image, we ensure that pixels in the same location receive similar assignment scores. Given an input image " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": ", two random augmentations are applied to produce two augmented views " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "v^{1}(x)" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "v^{2}(x)" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": ". Each augmentation view is then encoded with an encoder " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " into a hidden feature map and then transformed with a multilayer perceptron (MLP) " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " to get the feature map " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "z_{\\theta}(x) = g_{\\theta}(f_{\\theta}(v^{1}(x)))" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "z_{\\xi}(x) = g_{\\xi}(f_{\\xi}(v^{2}(x)))" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": ". The fore-background representation vectors of the image " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " in different networks can be obtained via slot attention [23], i.e., " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\theta}(x) =" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30283" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": "slot attention " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "(z_{\\theta}(x),S_{\\theta})" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\hat{S}_{\\xi} =" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": " slot attention " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "(z_{\\xi}(x),S_{\\xi})" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": " Then, the pixel assignment matrix, defining the score fore or background each pixel belongs to, " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\xi}" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": " can obtain by " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}(x) = z_{\\theta}\\cdot \\hat{S}_{\\theta}^{\\top}" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": " . The assignment matrix " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\xi}(x)" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": " for the teacher network can be obtained in the same way." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 294, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 294, + 242 + ], + "type": "text", + "content": "To keep the pixel level consistency, we enforce the prediction of the two views assignment " + }, + { + "bbox": [ + 55, + 134, + 294, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}(x)" + }, + { + "bbox": [ + 55, + 134, + 294, + 242 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 134, + 294, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\xi}(x)" + }, + { + "bbox": [ + 55, + 134, + 294, + 242 + ], + "type": "text", + "content": " match each other. However, the geometric image augmentations, such as random crop, scale, or flip, between different views will cause the scale or layout of the two feature maps to be inconsistent. So similar to [35], we use the inverse augmentation process on the predicted assignments to align their spatial locations, i.e., " + }, + { + "bbox": [ + 55, + 134, + 294, + 242 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{P}}_{\\theta}(x) = \\mathrm{invaug}(\\mathcal{P}_{\\theta}(x)), \\tilde{\\mathcal{Q}}_{\\xi}(x) = \\mathrm{invaug}(\\mathcal{Q}_{\\xi}(x))" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 242, + 294, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 294, + 301 + ], + "type": "text", + "content": "Based on the aligned assignments, we apply the cross-entropy loss to enforce the consistency in assignment scores between spatial-aligned pixels from different views. The cross-entropy loss is averaged over all spatial locations to keep the pixel prediction consistency" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 310, + 295, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 310, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 61, + 310, + 295, + 352 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {p i x e l}} (x) = \\frac {1}{H \\times W} \\sum_ {i, j} \\mathcal {L} ^ {\\mathrm {C E}} \\left(\\tilde {\\mathcal {Q}} _ {\\xi} (x) [ i, j ], \\tilde {\\mathcal {P}} _ {\\theta} (x) [ i, j ]\\right). \\tag {1}", + "image_path": "fa253eb2f1dcd300ca42c742f69bb49706923348bb39401c1f007221d89bbef4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 362, + 251, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 251, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 251, + 373 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 55, + 362, + 251, + 373 + ], + "type": "text", + "content": " means the position in the assignment matrix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "text", + "content": "In addition to the pixel-level contrastive loss, we incorporate a stuff-level consistency loss. This loss aims to alleviate variations in the background by treating the background in different images as positive samples, while considering all foreground samples as negative samples. According to the pixel assignment matrix " + }, + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{P}}_{\\theta}(x)" + }, + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Q}}_{\\xi}(x)" + }, + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "text", + "content": ", we can obtain the prototypes of the foreground and background by region average pooling to represent the whole foreground and background regions of the image " + }, + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "inline_equation", + "content": "p_{\\theta}^{f}(x), p_{\\theta}^{b}(x)" + }, + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "text", + "content": " represent the foreground and background prototypes obtained from the student network. " + }, + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "inline_equation", + "content": "p_{\\xi}^{f}(x), p_{\\xi}^{b}(x)" + }, + { + "bbox": [ + 55, + 375, + 295, + 531 + ], + "type": "text", + "content": " are from the teacher network. Then, the stuff-level loss can be written as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 539, + 294, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 539, + 294, + 608 + ], + "spans": [ + { + "bbox": [ + 62, + 539, + 294, + 608 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} ^ {\\text {s t u f f}} (B) = \\frac {1}{B (B - 1)} \\sum_ {i = 1} ^ {B} \\sum_ {j \\neq i} ^ {B} [ \\operatorname {s i m} \\left(p _ {\\theta} ^ {b} \\left(x _ {i}\\right), p _ {\\xi} ^ {b} \\left(x _ {j}\\right)\\right) - \\\\ \\left. \\left. \\operatorname {s i m} \\left(p _ {\\theta} ^ {b} \\left(x _ {i}\\right), p _ {\\xi} ^ {f} \\left(x _ {j}\\right)\\right) \\right]. \\right. \\tag {2} \\\\ \\end{array}", + "image_path": "1a9d5b457542261e62ec97b42140056e6c2180a6e13e13524c261ea0ec89b95c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "In Eq. 2, we hope the similarity between the background across different views will be large, however, the similarity between foreground and background should be small. We adopt the cosine similarity in our method. " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": " means the batch of training data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "To encourage the semantic centers to emphasize different visual contents and cover diverse semantics to avoid collapse, we apply a regularization to the assignment matrix of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 425, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 425, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 425, + 83 + ], + "type": "text", + "content": "different views, as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 91, + 553, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 91, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 340, + 91, + 553, + 105 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\operatorname {s e p}} (x) = \\operatorname {E n t r o y} \\left(\\mathcal {P} _ {\\theta} (x)\\right) + \\operatorname {E n t r o p y} \\left(\\mathcal {Q} _ {\\xi} (x)\\right). \\tag {3}", + "image_path": "c4aaace754524f379358dd961b98b6e2dc0fc1a684bcfff9cfd7bf8823d1a17f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 112, + 531, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 112, + 531, + 124 + ], + "spans": [ + { + "bbox": [ + 326, + 112, + 531, + 124 + ], + "type": "text", + "content": "The total loss to optimize the proposed indicator is" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 131, + 553, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 131, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 353, + 131, + 553, + 144 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} ^ {\\text {p i x e l}} (x) - \\lambda \\mathcal {L} ^ {\\text {s t u f f}} (B) - \\gamma \\mathcal {L} ^ {\\text {s e p}} (x). \\tag {4}", + "image_path": "1d61d4d10719fc865ae208b3756cb86ccea68b8bc9b6e7c343806d8e40017393.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 152, + 553, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 175 + ], + "type": "text", + "content": "3.2. Fusion of information into original slot attention method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 181, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 300 + ], + "type": "text", + "content": "The features provided by our indicator allow for clear distinction between the foreground object and background elements. Simultaneously, the background is processed uniformly, harmonizing its characteristics across various regions. Integrating the features obtained through our method with existing approaches significantly enhances the differentiation between the foreground and background. Furthermore, once trained, our indicator is frozen and only provides supplementary information to the existing slot attention-based methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 300, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 408 + ], + "type": "text", + "content": "In our fusion stage, we first train the slot attention-based method for several epochs. Once it has learned meaningful information, we next integrate the foreground and background information from our indicator into the slot attention-based method as supplementary information. Finally, the features from both the original slot attention-based method and our indicator are fused together and input into the next decoder for image reconstruction. The feature fusion stage can be written as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 415, + 553, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 415, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 349, + 415, + 553, + 430 + ], + "type": "interline_equation", + "content": "z ^ {\\text {f u s i o n}} (x) = \\operatorname {M L P} \\left(z ^ {\\text {s l o t}} (x) + z _ {\\xi} ^ {\\text {i n d i c a t o r}} (x)\\right), \\tag {5}", + "image_path": "c89ed64d7228012c52179581805533c51567bba6e4314b8b5eb6552d49b1a04c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{slot}}(x)" + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "content": " is the feature map from the encoder in the original slot attention method. " + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "inline_equation", + "content": "z_{\\xi}^{\\mathrm{indicator}}(x)" + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "content": " is the feature map provided by our indicator. " + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "content": " is the input image. These two feature maps are added together and then passed through a MLP layer to transform the combined feature map into a new space. The resulting fusion feature " + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{fusion}}(x)" + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "content": " replaces the feature map " + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{slot}}(x)" + }, + { + "bbox": [ + 313, + 437, + 554, + 557 + ], + "type": "text", + "content": " used in the original slot attention-based method to do the next stages, such as slot attention, and decoding, which are the same as the original slot attention framework." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 564, + 434, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 434, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 434, + 577 + ], + "type": "text", + "content": "3.3. Region Combination" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": "Defining the number of slots is a common challenge in recent slot attention methods [10] and can significantly impact their performance. Determining the appropriate number of slots is not straightforward, and recent approaches often opt for a higher number of slots, leading to slight oversegmentation, in order to achieve satisfactory results. In this section, we propose a spectral clustering-based method to group regions likely belonging to the same semantic object, which is used to refine the predictions during the testing phase." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "The similarity between slots represents the similarity between regions in the image. The similarity matrix between" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30284" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "different slots can be regarded as a reference to help recognize which slots possibly belong to the same semantic object. In our method, we directly use spectral clustering to divide the slots into different groups based on this matrix. Specifically, we calculate the cosine similarity between each pair of the slots to construct the similarity matrix " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ". Then according to " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", the group results can be obtained by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 167, + 295, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 167, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 110, + 167, + 295, + 180 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = \\text {s p e c t r a l c l u s t e r i n g} (\\mathbf {M}, N). \\tag {6}", + "image_path": "7459bd911bfed1c61c6b91c3159c0caff76440dbbc72736f5271b60c48289b01.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 193, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 266 + ], + "type": "text", + "content": "Based on the group result " + }, + { + "bbox": [ + 55, + 193, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 193, + 296, + 266 + ], + "type": "text", + "content": ", the corresponding areas of slots in the same group can be combined. To determining the group number " + }, + { + "bbox": [ + 55, + 193, + 296, + 266 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 193, + 296, + 266 + ], + "type": "text", + "content": ", we primarily employ eigen-gap heuristic [32]. In this method, we use the Laplacian matrix of the graph and find the largest gap between consecutive eigenvalues, which suggests the optimal number of clusters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 278, + 136, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 136, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 136, + 292 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 298, + 105, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 105, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 105, + 312 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 316, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 460 + ], + "type": "text", + "content": "Datasets. We evaluate our method on one synthetic and two real-world datasets. For the synthetic dataset, we use MOVi-C [13], which contains approximately 1,000 real 3D-scanned objects, with each scene featuring between 3 and 10 objects. Although MOVi-C is originally a video dataset, we convert it into an image dataset by randomly selecting frames following [28]. For the real-world datasets, we utilize MS COCO 2017 [21] and PASCAL VOC 2012 [9], which are commonly used for object detection and segmentation. PASCAL VOC 2012 contains many images with a single salient object, while MS COCO 2017 features more complex scenes, offering a more challenging evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "content": "Metrics. To assess object-centric learning, we used Mean Best Overlap (mBO) and Mean Intersection over Union (mIoU). The Mean Best Overlap (mBO) includes instance-level " + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{mBO}^i" + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "content": " and category-level " + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{mBO}^c" + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{mBO}^i" + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{mBO}^c" + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "content": " match the predicted masks to the instance and semantic masks, respectively, by finding the best overlap. Meanwhile, mIoU employs the Hungarian algorithm to match predicted and ground truth masks. [17] suggests that FG-ARI, as a clustering similarity metric, may yield unreliable segmentation quality assessments by overlooking the accuracy of predicted masks. Therefore, FG-ARI is not reported in this paper." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Implementation Details. To optimize the proposed foreground and background indicator, we employ the SGD optimizer. The learning rate is 0.001. Batch size is 512. The learning rate follows a linear warm-up for 10,000 optimization steps and an exponential decay schedule. Additionally, we clip the gradient norm to 1 to stabilize the training process. ViT-B/16 [1], initialized with DINO [5], is used as the encoder in our indicator. During the fusion stage, the model is trained using the same configuration as the corresponding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "original slot attention method you choose. All models are trained on a single A6000 GPU with 48 Gbytes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 105, + 552, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 552, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 552, + 118 + ], + "type": "text", + "content": "4.2. Integrating into SOTA object-centric methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 123, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 123, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 553, + 266 + ], + "type": "text", + "content": "Baselines. We apply our approach to two slot attention methods, DINOSAUR [28] and SPOT [17]. DINOSAUR utilizes a Vision Transformer encoder for feature reconstruction and has demonstrated applicability to real-world datasets. SPOT is a recent SOTA object-centric model that employs a two-stage self-training scheme and a patch-order permutation strategy for the transformer decoder. Furthermore, we also list the results of several other object-centric methods, including Slot Attention [23], SLASH [18], SLATE [29], SlotDiffusion [36], and LSD [16]. A more detailed description of the baseline methods can be found in the Appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 267, + 553, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 553, + 459 + ], + "type": "text", + "content": "Main Results on Each Dataset. Tab. 1 presents the main results of our method on the COCO, PASCAL, and MOViC datasets. Our method improves the performance of DINOSAUR and SPOT across all datasets. The fusion stage yields consistent improvements, while the region combination stage leads to a slight decrease in the " + }, + { + "bbox": [ + 313, + 267, + 553, + 459 + ], + "type": "inline_equation", + "content": "\\mathrm{mBO}^i" + }, + { + "bbox": [ + 313, + 267, + 553, + 459 + ], + "type": "text", + "content": " metric on the PASCAL and COCO datasets. This is due to the fact that, during clustering, slots corresponding to two instances of the same category may be merged into one, resulting in an increase in " + }, + { + "bbox": [ + 313, + 267, + 553, + 459 + ], + "type": "inline_equation", + "content": "\\mathrm{mBO}^c" + }, + { + "bbox": [ + 313, + 267, + 553, + 459 + ], + "type": "text", + "content": " but a decrease in " + }, + { + "bbox": [ + 313, + 267, + 553, + 459 + ], + "type": "inline_equation", + "content": "\\mathrm{mBO}^i" + }, + { + "bbox": [ + 313, + 267, + 553, + 459 + ], + "type": "text", + "content": ". Despite this, our method still outperforms the baseline models, DINOSAUR and SPOT, demonstrating a clear overall performance advantage. This highlights the crucial role of the fusion stage in driving performance improvements, while the region combination stage primarily serves to refine the predictions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 468, + 416, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 416, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 416, + 479 + ], + "type": "text", + "content": "4.3. More discussions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 545 + ], + "type": "text", + "content": "In this section, we analyze various aspects of our approach: the significance of the foreground and background indicators, the impact of different loss functions within the indicator, the role of the fusion stage, and the consistency of performance improvement across different slot numbers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "content": "(A) Does the extra foreground and background information really improve the slot attention method? We evaluate the slot attention method in four cases. Tab. 2 shows the results. In this table, \"Random\" refers to using random noise as the input to the fusion stage, \"ViT\" refers to directly using the pre-trained DINO ViT as the indicator, and \"Ours\" refers to our foreground and background indicator. We find that when the indicator is random noise, the baseline method struggles to distinguish between foreground and background, leading to a significant performance drop. When ViT is used as the indicator, the pretrained ViT has some inherent ability to differentiate between foreground and background, resulting in a slight performance improvement. Our designed foreground" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30285" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 542, + 235 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 542, + 235 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 542, + 235 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 542, + 235 + ], + "type": "table", + "html": "
MethodMOVi-CPASCALCOCO
mBOimIoUmBOimBOcmBOimBOc
Slot Attention [23] [NeurIPS20]26.3-24.624.917.219.2
SLASH [18] [CVPR23]-27.7----
SLATE [29] [ICLR22]39.437.835.941.529.133.6
SlotDiffusion [36] [NeurIPS23]--50.455.331.035.0
(Stable-)LSD [16] [NeurIPS23]45.644.2--30.4-
DINOSAUR [28] [ICLR23]42.441.844.051.231.138.9
DINOSAUR + FS (Ours)45.3 ↑ 2.944.5 ↑ 2.745.4 ↑ 1.452.5 ↑ 1.331.5 ↑ 0.439.9 ↑ 1.0
DINOSAUR + FS + RC (Ours)46.0 ↑ 3.644.8 ↑ 3.045.2 ↑ 1.252.9 ↑ 1.731.2 ↑ 0.140.1 ↑ 1.2
SPOT [17] [CVPR24]47.046.448.155.334.744.3
SPOT + FS (Ours)48.5 ↑ 1.547.7 ↑ 1.349.1 ↑ 1.056.3 ↑ 1.035.5 ↑ 0.845.0 ↑ 0.7
SPOT + FS + RC (Ours)49.0 ↑ 2.047.8 ↑ 1.449.3 ↑ 1.256.5 ↑ 1.235.7 ↑ 1.045.3 ↑ 1.0
", + "image_path": "b681b2de751304b7fdea6fb7b20d20dd509f95ee1a8a8469d8677adcb6a0d0c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 295, + 282, + 369 + ], + "blocks": [ + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "text", + "content": "Table 1. Comparison with different object-centric methods on MOVi-C, PASCAL, and COCO datasets. DINOSAUR uses an autoregressive decoder and DINO [5] ViT encoder. The experimental results of the baseline are derived from the original corresponding papers. FS means using the fusion stage in our paper. RC means the trick of region combination." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 295, + 282, + 369 + ], + "lines": [ + { + "bbox": [ + 70, + 295, + 282, + 369 + ], + "spans": [ + { + "bbox": [ + 70, + 295, + 282, + 369 + ], + "type": "table", + "html": "
SA MethodIndicatorIoUmBOcmBOi
DINOSAUR--51.244.0
Random-25.222.1
ViT30.251.444.7
Ours52.652.545.4
", + "image_path": "b19bae34777309140efdac5a00adcd6265278ac9f39ea6b5aac72d4a5dfb88cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 70, + 456, + 282, + 514 + ], + "blocks": [ + { + "bbox": [ + 55, + 376, + 296, + 443 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 443 + ], + "type": "text", + "content": "Table 2. The effect of extra information on the existing slot attention-based method. All experiments are conducted on PASCAL dataset. IoU is used to show the model's performance on segmenting the foreground and background. " + }, + { + "bbox": [ + 55, + 376, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{mBO}^c" + }, + { + "bbox": [ + 55, + 376, + 296, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 376, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{mBO}^i" + }, + { + "bbox": [ + 55, + 376, + 296, + 443 + ], + "type": "text", + "content": " are employed to assess the final performance after fusing the extra knowledge from the indicator." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 456, + 282, + 514 + ], + "lines": [ + { + "bbox": [ + 70, + 456, + 282, + 514 + ], + "spans": [ + { + "bbox": [ + 70, + 456, + 282, + 514 + ], + "type": "table", + "html": "
MethodMOVi-CPASCALCOCO
DINOSAUR (2 slots)45.247.646.9
SPOT (2 slots)46.650.846.4
Ours47.452.648.4
", + "image_path": "17aad365b3b6afd5fcca0663a7454e289db85e3b4a8f39b41a7f77330b46aed0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 522, + 295, + 555 + ], + "lines": [ + { + "bbox": [ + 54, + 522, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 295, + 555 + ], + "type": "text", + "content": "Table 3. The performance of using object-centric methods as the foreground and background indicator. Fig. 3 shows some visualization results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 580, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 614 + ], + "type": "text", + "content": "and background indicator makes the foreground and background highly separable and significantly improves performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "(B) How about using the classical slot attention methods as the foreground and background indicator in our method? Tab. 3 evaluates the performance of different models when the number of slots is set to 2. We continue to use the Mean Best Overlap (mBO) as the evaluation metric. The difference is that, during metric computation, we convert the ground truth masks into binary masks containing only foreground (0) and background (1). Then, the pre" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 297, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 321 + ], + "type": "text", + "content": "dicted masks are matched with the ground truth masks using the best overlap matching method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 324, + 556, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 556, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 556, + 504 + ], + "type": "text", + "content": "From the data in Tab. 3, it can be observed that our method achieves the best performance with significant performance improvements on both COCO and PASCAL datasets. This is one of the reasons for using the foreground and background indicator. Another reason is that autoencoder-based methods like DINOSAUR and SPOT, while capable of dividing the scene into two parts, struggle to achieve accurate foreground-background segmentation. As shown in Fig. 3, the baseline methods tend to divide the scene into two parts, top and bottom, while our method focuses on separating the foreground and background. This is because our method employs both pixel-level and stufflevel contrastive learning, enabling the model to better distinguish foreground from background features, leading to a more effective foreground-background decomposition." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": "(C) Impact of different losses. We conduct the ablation experiments to further analyze the role of different losses in the indicator, as shown in Tab. 4. We find that without using " + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{sep}}" + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": ", the indicator struggle to achieve a performance improvement. This is because when only " + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{pixel}}" + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{stuff}}" + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": " are used, the model tends to collapse all information into a single slot, where one slot binds to the entire image and the other slot remains empty. In this case, the model can achieve low " + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{pixel}}" + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{stuff}}" + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": " values, but it cannot effectively perform the foreground-background separation. Therefore, we introduce " + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{sep}}" + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": " to encourage semantic centers to emphasize different visual content and cover diverse semantics. However, relying on " + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{sep}}" + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{pixel}}" + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": " makes it difficult for the model to adapt to variations in the background. As a result, " + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{sep}}" + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{pixel}}" + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{stuff}}" + }, + { + "bbox": [ + 313, + 506, + 556, + 699 + ], + "type": "text", + "content": " must work together in synergy to achieve effective foreground-background separation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "At the same time, it can be seen that the performance of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30286" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 73, + 292, + 252 + ], + "blocks": [ + { + "bbox": [ + 62, + 73, + 292, + 252 + ], + "lines": [ + { + "bbox": [ + 62, + 73, + 292, + 252 + ], + "spans": [ + { + "bbox": [ + 62, + 73, + 292, + 252 + ], + "type": "image", + "image_path": "740561010922650beec7cb8d634406b8ea0711dc76fba8e4c20c5511fa668675.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 256, + 295, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 256, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 295, + 289 + ], + "type": "text", + "content": "Figure 3. Example results of the proposed indicator and SOTA slot attention method with 2 slots. Our method distinguishes foreground and background better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 309, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 296, + 369 + ], + "type": "text", + "content": "the indicator is positively correlated with the final results. This further confirms the conclusion in (A): the stronger the indicator's ability to distinguish between foreground and background, the more effectively it enhances the feature fusion process." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 70, + 378, + 281, + 459 + ], + "blocks": [ + { + "bbox": [ + 70, + 378, + 281, + 459 + ], + "lines": [ + { + "bbox": [ + 70, + 378, + 281, + 459 + ], + "spans": [ + { + "bbox": [ + 70, + 378, + 281, + 459 + ], + "type": "table", + "html": "
LpixelLstuffLsepIoUmBOcmBOi
✓28.851.444.6
✓✓28.451.344.7
✓✓48.951.744.8
✓✓✓52.652.545.4
", + "image_path": "061e0a5f30940a954afa74d44bce4313b85e5ca57401861a62fcef4f8ed38d72.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 467, + 295, + 534 + ], + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 534 + ], + "type": "text", + "content": "Table 4. Analysis of different loss functions on PASCAL dataset. IoU represents the results of the indicator on segmenting foreground and background regions. " + }, + { + "bbox": [ + 55, + 467, + 295, + 534 + ], + "type": "inline_equation", + "content": "\\mathrm{mBO}^c" + }, + { + "bbox": [ + 55, + 467, + 295, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 467, + 295, + 534 + ], + "type": "inline_equation", + "content": "\\mathrm{mBO}^i" + }, + { + "bbox": [ + 55, + 467, + 295, + 534 + ], + "type": "text", + "content": " represent the final results of DINOSAUR after fusing knowledge from different indicators, which are optimized using different types of loss functions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "type": "text", + "content": "(D) The importance of the fusion stage and region combination stage. We conduct a detailed analysis of the importance of the fusion stage and the region combination stage. The results are summarized in Table 5. We observe that directly applying the region combination to DINOSAUR and SPOT typically leads to a decline in performance metrics. We hypothesize that this is due to the limited ability of the baseline methods to differentiate between foreground and background. Without a clear distinction, the region combination stage inappropriately merges foreground and background. The visualization results are shown in Figure 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "On the MOVi dataset, applying region combination directly to SPOT results in a performance improvement. This" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 551, + 262 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 551, + 262 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 551, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 551, + 262 + ], + "type": "image", + "image_path": "2e184b3c267ddaebe0499b2e2e14785a29e8c205b79f2f15209ad3fb439d1bc6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 265, + 554, + 287 + ], + "lines": [ + { + "bbox": [ + 313, + 265, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 554, + 287 + ], + "type": "text", + "content": "Figure 4. The visualization results in our different components. FS means fusion stage, and RC is the stage of region combination." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 311, + 554, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 554, + 370 + ], + "type": "text", + "content": "may result from the relatively simple background in the MOVi dataset, where SPOT is already effective at distinguishing between foreground and background. As shown in Tab. 3, the metric for SPOT with 2 slots closely aligns with our indicator result, further supporting this observation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 372, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 538 + ], + "type": "text", + "content": "(E) Consistent improvement across different slot numbers. The quality of results is influenced by the number of slots used during model training. Therefore, understanding the model's robustness to variations in the number of slots is crucial. In this part, we compare our method with DINOSAUR to evaluate the stability of our method with different numbers of slot. We conduct experiments on two real-world datasets, COCO and PASCAL, using five slot numbers, with the results presented in Tab. 6. Notably, our method outperforms DINOSAUR across all slot numbers and exhibits lower variance. This suggests that, while improving baseline performance, our method also demonstrates better generalization ability when applied with different numbers of slots." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 551, + 423, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 423, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 423, + 563 + ], + "type": "text", + "content": "4.4. Qualitative results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Fig. 5 shows example results on the COCO and PASCAL datasets for DINOSAUR, SPOT, and our method. Row (a) demonstrates that, compared to baseline methods, our approach correctly fuses the foreground and background. For example, in the PASCAL dataset, our method successfully merges the ground that was originally divided into two separate slots. Rows (b) and (c) illustrate that when the baseline method accurately segments the background (foreground), our method correctly fuses the foreground (background) regions. For instance, in row (b) of the PASCAL dataset, when the background is correctly segmented, our method successfully merges the train and streetlight, which were" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30287" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 542, + 198 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 542, + 198 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 542, + 198 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 542, + 198 + ], + "type": "table", + "html": "
MethodMOVi-CPASCALCOCO
mBOimIoUmBOimBOcmBOimBOc
DINOSAUR42.441.844.051.231.138.9
DINOSAUR+RC41.6 ↓0.839.5 ↓2.343.3 ↓0.750.7 ↓0.530.4 ↓0.739.4 ↑0.5
DINOSAUR+FS (Ours)45.3 ↑2.944.5 ↑2.745.4 ↑1.452.5 ↑1.331.5 ↑0.439.9 ↑1.0
DINOSAUR+FS+RC (Ours)46.0 ↑3.644.8 ↑3.045.2 ↑1.252.9 ↑1.731.2 ↑0.140.1 ↑1.2
SPOT47.046.448.155.334.744.3
SPOT+RC47.8 ↑0.847.1 ↑0.748.0 ↓0.154.8 ↓0.533.2 ↓1.543.5 ↓0.8
SPOT+FS (Ours)48.5 ↑1.547.7 ↑1.349.1 ↑1.056.3 ↑1.035.5 ↑0.845.0 ↑0.7
SPOT+FS+RC (Ours)49.0 ↑2.047.8 ↑1.449.3 ↑1.256.5 ↑1.235.7 ↑1.045.3 ↑1.0
", + "image_path": "37333a7b52abfe8328fc7dc4e7f434f9dcf5e178004fae5ddd56912817e9b0a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 240, + 306, + 387 + ], + "blocks": [ + { + "bbox": [ + 60, + 240, + 306, + 387 + ], + "lines": [ + { + "bbox": [ + 60, + 240, + 306, + 387 + ], + "spans": [ + { + "bbox": [ + 60, + 240, + 306, + 387 + ], + "type": "image", + "image_path": "70293c6a398e4134211456b21d6ede2a6e1795090f14afad9a26ea04f1b791d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 397, + 493, + 408 + ], + "lines": [ + { + "bbox": [ + 116, + 397, + 493, + 408 + ], + "spans": [ + { + "bbox": [ + 116, + 397, + 493, + 408 + ], + "type": "text", + "content": "Figure 5. Qualitative results on COCO and PASCAL datasets. Ours: fusion stage + region combination." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 240, + 547, + 387 + ], + "blocks": [ + { + "bbox": [ + 310, + 240, + 547, + 387 + ], + "lines": [ + { + "bbox": [ + 310, + 240, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 310, + 240, + 547, + 387 + ], + "type": "image", + "image_path": "17765ede36fb5b66311932a5510a2a0ebd09151a5bf6c3f1262dcf05e5cc0e6d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 59, + 427, + 294, + 591 + ], + "blocks": [ + { + "bbox": [ + 55, + 206, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 555, + 229 + ], + "type": "text", + "content": "Table 5. The importance of the different components in our method. FS is the fusion stage and RS is the region combination stage. Using RS alone even hurt the performance. FS is crucial." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 427, + 294, + 591 + ], + "lines": [ + { + "bbox": [ + 59, + 427, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 59, + 427, + 294, + 591 + ], + "type": "table", + "html": "
DatasetKDINOSAURDINOSAUR+FS
mBOimBOcmBOimBOc
COCO227.620.530.2 ↑ 2.622.2 ↑ 1.7
537.929.339.3 ↑ 1.430.3 ↑ 1.0
739.031.139.9 ↑ 0.931.5 ↑ 0.4
938.531.739.8 ↑ 1.332.0 ↑ 0.3
1138.031.639.5 ↑ 1.532.6 ↑ 1.0
PASCAL234.340.435.4 ↑ 1.141.6 ↑ 1.2
444.351.644.6 ↑ 0.352.0 ↑ 0.4
644.851.845.4 ↑ 0.652.5 ↑ 0.7
845.351.745.5 ↑ 0.252.3 ↑ 0.6
1045.451.345.6 ↑ 0.252.0 ↑ 0.7
", + "image_path": "a9b89756d213e5d2e1bcc923b4b22dec72e3ef074f570689902bc02a18d8079e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 600, + 297, + 633 + ], + "lines": [ + { + "bbox": [ + 55, + 600, + 297, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 297, + 633 + ], + "type": "text", + "content": "Table 6. The performance of our method with different numbers of slot. For a fair comparison, only the fusion stage is adopted. " + }, + { + "bbox": [ + 55, + 600, + 297, + 633 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 600, + 297, + 633 + ], + "type": "text", + "content": " represents the number of slots." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 654, + 220, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 220, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 220, + 665 + ], + "type": "text", + "content": "originally assigned to two separate slots." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 140, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 140, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 140, + 685 + ], + "type": "text", + "content": "4.5. Training cost" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "The additional training cost of our method primarily consists of two components. First, training the indicator. Due" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 430, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 513 + ], + "type": "text", + "content": "to its simple structure, training on the PASCAL dataset using a single A6000 GPU takes approximately 3 hours. Once trained, the indicator can be directly applied to different slot attention methods on the same dataset without the need for retraining. Second, retraining the existing slot attention methods based on the fused features, where the training time depends on the chosen slot attention method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 546, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 712 + ], + "type": "text", + "content": "In this paper, we consider the problem that complex backgrounds in images can disrupt the phase of slot binding. Hence, we introduce a contrastive learning-based indicator to provide extra information about the foreground and background region to the existing slot attention method. Integrating this indicator into the slot attention-based method allows the model to focus more effectively on segmenting foreground objects and minimizing background distractions. According to the similarity between the slots, we also use a spectral clustering mechanism to refine the results in the testing phase. In experiments, we integrate our method into two SOTA slot attention methods. All the results show that our method can consistently enhance the performance on different datasets." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30288" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 178 + ], + "type": "text", + "content": "[2] Christopher P Burgess, Loic Matthew, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "text", + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "type": "text", + "content": "[4] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "text", + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "text", + "content": "[9] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "type": "text", + "content": "[10] Ke Fan, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco Locatello, and Zheng Zhang. Adaptive slot attention: Object discovery with dynamic slot number. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23062-23071, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 656 + ], + "type": "text", + "content": "[11] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In International conference on machine learning, pages 2424-2433. PMLR, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "text", + "content": "[12] Klaus Greff, Sjoerd Van Steenkiste, and Jürgen Schmidhuber. On the binding problem in artificial neural networks. arXiv preprint arXiv:2012.05208, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[13] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapra" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "text", + "content": "gasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "text", + "content": "[14] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[15] A Jaiswal, A Ramesh Babu, M Zaki Zadeh, D Banerjee, and F Makedon. A survey on contrastive self-supervised learning, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[16] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 8563-8601, 2023. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 319 + ], + "type": "text", + "content": "[17] Ioannis Kakogeorgiou, Spyros Gidaris, Konstantinos Karantzalos, and Nikos Komodakis. Spot: Self-training with patch-order permutation for object-centric learning with autoregressive transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22776-22786, 2024. 2, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 321, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 375 + ], + "type": "text", + "content": "[18] Jinwoo Kim, Janghyuk Choi, Ho-Jin Choi, and Seon Joo Kim. Shepherding slots to objects: Towards stable and robust object-centric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19198-19207, 2023. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 376, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 555, + 432 + ], + "type": "text", + "content": "[19] Hyeokjun Kweon, Sung-Hoon Yoon, and Kuk-Jin Yoon. Weakly supervised semantic segmentation via adversarial learning of classifier and reconstructor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11329-11339, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 434, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 475 + ], + "type": "text", + "content": "[20] Brenden M Lake, Tomer D Ullman, Joshua B Tenenbaum, and Samuel J Gershman. Building machines that learn and think like people. Behavioral and brain sciences, 40:e253, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 478, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 544 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 545, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 577 + ], + "type": "text", + "content": "[22] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection without imagenet. arXiv preprint arXiv:2011.13677, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 579, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 644 + ], + "type": "text", + "content": "[23] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33:11525-11538, 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[24] Amin Mansouri, Jason Hartford, Yan Zhang, and Yoshua Bengio. Object-centric architectures enable efficient causal representation learning. arXiv preprint arXiv:2310.19054, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[25] Gary F Marcus. The algebraic mind: Integrating connectionism and cognitive science. MIT press, 2003. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30289" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "text", + "content": "[27] Rui Qian, Shuangrui Ding, Xian Liu, and Dahua Lin. Semantics meets temporal correspondence: Self-supervised object-centric learning in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16675-16687, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 239 + ], + "type": "text", + "content": "[28] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. In The 11th International Conference on Learning Representations, 2023. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 274 + ], + "type": "text", + "content": "[29] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dall-e learns to compose. In 10th International Conference on Learning Representations, ICLR 2022, 2022. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 318 + ], + "type": "text", + "content": "[30] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in neural information processing systems, 33:6827-6839, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 373 + ], + "type": "text", + "content": "[31] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc V Gool. Revisiting contrastive methods for unsupervised learning of visual representations. Advances in Neural Information Processing Systems, 34: 16238-16250, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 374, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 374, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 374, + 294, + 397 + ], + "type": "text", + "content": "[32] Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17:395-416, 2007. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 294, + 420 + ], + "type": "text", + "content": "[33] Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17:395-416, 2007. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "type": "text", + "content": "[34] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 475, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 520 + ], + "type": "text", + "content": "[35] Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. Advances in neural information processing systems, 35:16423-16438, 2022. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 521, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 294, + 565 + ], + "type": "text", + "content": "[36] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotdiffusion: Object-centric generative modeling with diffusion models. Advances in Neural Information Processing Systems, 36:50932-50958, 2023. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "text", + "content": "[37] Ceyuan Yang, Zhirong Wu, Bolei Zhou, and Stephen Lin. Instance localization for self-supervised detection pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3987-3996, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 611, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 611, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 611, + 294, + 654 + ], + "type": "text", + "content": "[38] Ying Zheng, Lei Yao, Yuejiao Su, Yi Zhang, Yi Wang, Sicheng Zhao, Yiyi Zhang, and Lap-Pui Chau. A survey of embodied learning for object-centric robotic manipulation. arXiv preprint arXiv:2408.11537, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 655, + 294, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 655, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 655, + 294, + 711 + ], + "type": "text", + "content": "[39] Rui Zhu, Bingchen Zhao, Jingen Liu, Zhenglong Sun, and Chang Wen Chen. Improving contrastive learning by visualizing feature transformation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10306-10315, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30290" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PerLA_ Perceptive 3D Language Assistant/54882f4a-04ec-4ab3-b66a-a66c8e83acf5_content_list.json b/2025/PerLA_ Perceptive 3D Language Assistant/54882f4a-04ec-4ab3-b66a-a66c8e83acf5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a3b2cff8863055291d964dac494b0b7a9799827e --- /dev/null +++ b/2025/PerLA_ Perceptive 3D Language Assistant/54882f4a-04ec-4ab3-b66a-a66c8e83acf5_content_list.json @@ -0,0 +1,1403 @@ +[ + { + "type": "text", + "text": "PerLA: Perceptive 3D language assistant", + "text_level": 1, + "bbox": [ + 269, + 127, + 727, + 157 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guofeng Mei $^{1}$ Wei Lin $^{2}$ Luigi Riz $^{1}$ Yujiao Wu $^{3}$ Fabio Poiesi $^{1}$ Yiming Wang $^{1}$ $^{1}$ Fondazione Bruno Kessler, Italy $^{2}$ JKU Linz, Austria $^{3}$ CSIRO, Australia", + "bbox": [ + 145, + 190, + 849, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gmei@fbk.eu", + "bbox": [ + 444, + 229, + 545, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f170598bfc50c4a30adbdff664e93b3edefe38b17009eb83b905c054e915b4da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 280, + 480, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac232aac1bab0d051b6c7f3e2a340086609463c42d91248458bb2b4a53e15546.jpg", + "image_caption": [ + "Figure 1. PerLA is a 3D language assistant that integrates local details with global context to learn informative representations of 3D scenes, whereas state-of-the-art (SOTA) 3DLAs focus solely on global context information. PerLA can provide more accurate responses, correctly distinguishing between objects such as a \"black computer monitor\" and a \"black suitcase,\" where SOTA models instead fail with hallucinated responses. Examples in figures show cases where capturing details from the point cloud matters for accurate output captions." + ], + "image_footnote": [], + "bbox": [ + 514, + 280, + 901, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 542, + 326, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Enabling Large Language Models (LLMs) to understand the 3D physical world is an emerging yet challenging research direction. Current strategies for processing point clouds typically downsample the scene or divide it into smaller parts for separate analysis. However, both approaches risk losing key local details or global contextual information. In this paper, we introduce PerLA, a 3D language assistant designed to be more perceptive to both details and context, making visual representations more informative for the LLM. PerLA captures high-resolution (local) details in parallel from different point cloud areas and integrates them with (global) context obtained from a lower-resolution whole point cloud. We present a novel algorithm that preserves point cloud locality through the Hilbert curve and effectively aggregates local-to-global information via cross-attention and a graph neural network. Lastly, we introduce a novel loss for local representation consensus to promote training stability. PerLA outperforms state-of-the-art 3D language assistants, with gains of up to $+1.34$ CiDEr on ScanQA for question answering, and $+4.22$ on ScanRefer and $+3.88$ on Nr3D for dense captioning. Project page: https://gfmei.github.io/PerLA", + "bbox": [ + 88, + 575, + 485, + 909 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 542, + 643, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D language assistants (3DLAs) jointly process natural language and 3D data to achieve 3D scene understanding, such as recognizing object categories, locations, appearances, and relationships, without requiring specialized models for each recognition task [9, 11, 21]. These capabilities are primarily powered by Large Language Models (LLMs) trained in large text corpora [54]. These approaches can aggregate multi-view features [21] or process point clouds [9] to generate 3D representations, which are in turn converted to tokens for the LLM [54, 66]. However, extracting multi-view representations is computationally costly and often fails to capture essential geometric properties [21]. Directly processing point clouds can yield more accurate results, yet it is even more computationally costly than handling multiview data, as point clouds typically have rather large cardinalities [60]. To address this, the point cloud cardinality can be reduced via downsampling [8, 9]. However, as with images [36], downsampling can compromise downstream task performance due to the reduced model's ability to perceive fine details of 3D scenes [47].", + "bbox": [ + 511, + 566, + 906, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fig. 1 illustrates two cases where our approach can accurately capture details and describe small objects within large", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14369", + "bbox": [ + 480, + 948, + 519, + 960 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scenes, while a state-of-the-art method hallucinates object details [9]. Although fine-grained attributes are critical for performance, directly extracting detailed information from high-resolution 3D data for 3DLAs remains underexplored.", + "bbox": [ + 89, + 90, + 485, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim to enhance 3DLAs' ability to perceive finer details in point clouds in order to execute downstream tasks more accurately. While increasing the number of visual tokens as the point cloud grows in size seems the straightforward solution, our empirical study (Tab. 3) shows that this solution has limited effectiveness in capturing scene details and it just increases computational burden. To address this, we propose PerLA, a novel 3DLA with a perceptive 3D scene encoder that captures detailed point cloud information, allowing the language model to generate more accurate responses without processing additional tokens. PerLA first divides the complete 3D scene into non-overlapping local parts to be processed in parallel, then integrates this local information with the global context obtained from a lower-resolution representation of the entire point cloud. Although dividing visual input has gained popularity in image processing [35], it has not yet been applied to point clouds. Processing point clouds presents unique challenges beyond those of images, as it requires handling an unordered set of points rather than rasterized pixels. To address these challenges, we serialize and partition the point cloud before encoding using a Hilbert curve approach [50], which efficiently preserves locality. We then combine local and global information through an efficient Hilbert curve-based $k$ -NN search and aggregate this information via a novel cross-attention module and Graph Convolutional Network (GCN) to generate highly informative point-level representations for the LLM. Lastly, we train PerLA with a novel loss function designed to promote consensus on local representation, addressing the issue of divergent representations during local-to-global aggregation. We validate our approach on the question answering benchmark ScanQA [2], and the 3D dense captioning benchmarks ScanRefer [6] and Nr3D [1] to demonstrate the effectiveness of PerLA. PerLA demonstrates high transferability on both tasks of 3D question answering and 3D dense captioning, achieving state-of-the-art performance. In summary, our contributions are:", + "bbox": [ + 91, + 161, + 485, + 738 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel perceptive 3D encoder to preserve local and global information for 3D language assistant.", + "- We introduce an efficient approach based on Hilbert curve $k$ -NN search and cross attention to aggregate local and global information at point level.", + "- We propose a novel algorithm based on Graph Neural Network to refine and enhance aggregated information.", + "- We introduce a novel loss objective to enable local representation consensus for local-to-global information aggregation." + ], + "bbox": [ + 89, + 747, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 514, + 89, + 651, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D language understanding involves understanding 3D scenes by describing or answering to scene-relevant questions in the format of natural language. Examples of typical downstream tasks include 3D Dense Captioning [8, 12, 56], and 3D Question Answering [2, 39, 65, 67]. 3D Dense Captioning consists of 3D localization of object instances, and textual descriptions of each object instance [4, 8, 10, 12]. 3D Question Answering requires the model equipped with a language decoder to answer questions regarding the visual context in the given 3D scene [2, 44]. Several works focus on addressing the problems of 3D-language pre-alignment [13, 26], or designing adapter layers [9, 21], or constructing 3D synthetic data [64]. In contrast, our work focuses on designing an encoding approach that can enhance 3D-language understanding by capturing fine-grained details.", + "bbox": [ + 511, + 119, + 906, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D language assistants (3DLAs) can be roughly categorized into two types: object-level 3DLAs [23, 46, 53, 61] and scene-level 3DLAs [9, 21, 59, 69]. Object-level 3DLAs learn from large 3D object datasets [15] to connect object-level 3D representation with language models. However, they underperform in compositional reasoning in complex 3D scenes with numerous objects. Scene-level 3DLAs, such as 3D-LLM [21], LLaVA-3D [69], Chat-3D [59], and LL3DA [9] enable scene understanding via the interaction with objects. Because 3D data is orders of magnitude less than 2D data, existing 3DLAs address such lack of data by leveraging pre-trained 2D Large Multimodal Models (LMMs) [21, 69], or data-efficient training recipes [9, 59]. 3D-LLM [21] leverages 2D pre-trained representations of rendered multi-view images to construct 3D representations and 2D VLMs as backbones. LLaVA-3D [69] adapts LLaVA [36] for 3D scene understanding by associating 2D patch representations with their positions in 3D space. Chat-Scene [22] improves the referencing and grounding capability and models scene representations as a sequence of object-level representations. LL3DA [9] extracts point-level representations from a downsampled 3D scene, and includes both interaction prompts and textual instructions to resemble human interactions with the 3D environment. Orthogonal to the efforts of aligning 3D-language with limited 3D data in existing 3DLAs, PerLA aims to improve the capability of 3DLAs in perceiving scene details.", + "bbox": [ + 511, + 367, + 908, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual perception enhancement on multimodal models. Multi-granularity representation learning has been explored in 2D multimodal models [5, 18, 20, 29, 49, 51, 63], showing that combining local and global views yields more informative representations than relying on a single global view. Mini-Gemini [32] introduces a dual vision encoder setup using CLIP ViT [48] and ConvNeXt [37] to process low- and high-resolution views of an input image. Mod-", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14370", + "bbox": [ + 480, + 948, + 519, + 960 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "els in the LLaVA-Next series [30, 35, 62] and InternLM-XComposer2 [17] use an additional branch to handle view partitions, increasing the number of visual tokens and preserving more fine-grained visual details compared to language-vision models with a single global view branch. In the 3D domain, Scene-LLM [19] and Segment3D [24] improve segmentation accuracy by transferring semantic details from multi-view images to point clouds. While extensive work has been done to enhance perception in 2D models and transfer multi-view image information into 3D, methods for preserving detailed representations from point clouds in 3DLAs remain underexplored. Our method aims to address this gap by preserving both global and local visual information to enhance 3D perception.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Perceptive 3D language assistant (PerLA)", + "text_level": 1, + "bbox": [ + 89, + 311, + 460, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PerLA takes as inputs $i$ a text prompt in natural language, $ii$ the 3D scene in the form of a point cloud, and $iii$ a visual prompt provided as either a user click or a bounding box. The text prompt is processed by a text prompt encoder to produce text representations, which are then input to both the Large Language Model (LLM) and the multimodal adapter (MMA). The text encoder is a transformer based on BLIP-2 [31]. The point cloud is processed by our perceptive scene encoder, which generates scene representations that feed into both the MMA and the subsequent encoder. We will detail the perceptive scene encoder in the next sections. The visual prompt is processed by the visual prompt encoder, which, by combining the perceptive scene encoder's representations, outputs scene representations that are further processed by the MMA. For more details on visual prompts, please refer to Supp. Mat. The MMA takes as input these multimodal representations and outputs tokens for the LLM. The MMA is implemented as a Q-former [31]. MMA's output is projected into the LLM's representation space through linear projector. Lastly, these projected representations are processed by the LLM to generate the output response. We train PerLA using data provided by [21] and finetune it for each downstream task. Fig. 2 illustrates our approach.", + "bbox": [ + 91, + 337, + 482, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, let $\\mathcal{P} = \\left\\{\\left(\\pmb {p}_i\\in \\mathbb{R}^3,\\pmb {f}_i\\in \\mathbb{R}^{d_0}\\right)\\mid i = 1,2,\\dots ,N\\right\\}$ denote PerLA's input point cloud, where $\\pmb{p_i}$ is a point coordinate, $\\pmb{f_{i}}$ is a $d_0$ -dimensional feature vector (e.g., color or normal vector) corresponding to $\\pmb {p}_i$ , and $N$ is the number of points (cardinality). Let $\\mathcal{I}^t$ and $\\mathcal{I}^v$ denote the input text prompt and visual prompt, respectively. Based on these inputs, PerLA (denoted as $\\Phi$ ) generates a free-form natural language response $\\mathcal{O}^t$ for various 3D-related tasks.", + "bbox": [ + 89, + 699, + 482, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Perceptive scene encoder", + "text_level": 1, + "bbox": [ + 89, + 832, + 316, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To compute point-level representations through the scene encoder, $\\mathcal{P}$ is typically downsampled into super-points using Farthest Point Sampling (FPS) [8, 9]. However, down", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sampling and using only $\\mathcal{P}$ may hinder the encoder to capture fine-grained details, thus affecting performance in downstream 3D scene understanding tasks. Our proposed perceptive scene encoder is designed to preserve such scene details without increasing the number of tokens or the representation dimensions of the 3D scene. We propose to split $\\mathcal{P}$ into parts, and employ a pre-trained 3D scene encoder to encode these parts and the whole point cloud separately. We then aggregate the output representations from these different pieces into a single (highly-informative) representation using a cross-attention-based module and Graph Convolutional Network (GCN).", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hilbert-based scene partitioning. We partition $\\mathcal{P}$ into $L$ equally-sized parts, each containing the same number of points, $\\left\\lfloor \\frac{N}{L} \\right\\rfloor$ , where $\\lfloor \\cdot \\rfloor$ denotes the greatest integer less than or equal to its argument. Enforcing equal cardinality across parts allows for spatially smaller parts in highly structured regions (areas with more semantic information) and spatially larger parts in less structured regions (areas with less semantic information). To achieve this, we adopt a Hilbert curve approach [50] to efficiently serialize the point cloud and partition it into parts [33, 57, 60].", + "bbox": [ + 511, + 275, + 903, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Global and partial scene encoding. Both full-scene point cloud and the partial-scene point clouds are encoded separately by the same pre-trained 3D scene encoder $\\phi$ . The 3D scene encoder $\\phi$ downsamples the input point cloud into super-points using Farthest Point Sampling (FPS) [41] and produces a representation for each super-point. The representation of the full scene, i.e., the global representation, encode the overall scene context, while the representations of the partial scenes, i.e., the local representations, encode scene details. Let $\\mathcal{P}^g = \\{\\pmb{p}_i^g \\in \\mathbb{R}^3\\}_{i=1}^M$ denote the downsampled full-scene point cloud composed of $M$ points, and $\\mathcal{F}^g = \\{\\pmb{f}_i^g \\in \\mathbb{R}^d\\}_{i=1}^M$ denote the associated global representations. Let $\\mathcal{P}^l = \\{\\pmb{p}_i^l \\in \\mathbb{R}^3\\}_{i=1}^{L \\cdot M}$ denote the set of downsampled partial-scene point clouds, and $\\mathcal{F}^l = \\{\\pmb{f}_i^l \\in \\mathbb{R}^d\\}_{i=1}^{L \\cdot M}$ the associated local representations. Note that, we aggregate all the points of the downsampled partial-scene point clouds into the single set $\\mathcal{P}^l$ , and both the full-scene and the partial-scenes are downsampled to the same number of super-points $M$ . The representations $\\mathcal{F}^l$ provide a more detailed view of the scene, as they are derived from the same number of points downsampled from smaller, localized regions, resulting in higher resolution compared to the global versions. Yet, their semantic visibility is focused within each local part. In the following steps, we aggregate local and global information, to enrich the representations with details within the scene context.", + "bbox": [ + 511, + 429, + 903, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hilbert-based nearest-neighbor search. In order to enhance global representations through local representations, we need to first find the correspondences among them, i.e., which super-points of partial-scene point clouds are spatially neighbors with super-points of the full-scene point", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14371", + "bbox": [ + 480, + 948, + 517, + 959 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/527958e90390560bc5710e6da3bc2c83bd2e7ea02ca04705e8d33e86669ce5c5.jpg", + "image_caption": [ + "Figure 2. Overview of PerLA. (Left): The overall pipeline of PerLA, which begins by extracting interaction-aware 3D scene representations. These representations are then projected onto the prefix of textual instructions via MMA, serving as input to a frozen language model (LLM). (Right): The detailed design of PerLA. First, the 3D scene is divided into spatially compact regions using Hilbert-based scene serialization [50]. Next, an efficient $k$ -NN algorithm associates each point-level global representation with its detail-enriched local representations, creating a comprehensive scene representation through a Graph Convolutional Network (GCN). Finally, smoothness and regularization losses are applied to promote stable learning for the proposed perceptive scene encoder." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 903, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cloud. This can be done via $k$ -NN search. However, traditional $k$ -NN searches can be computationally intensive on large-scale or high-resolution point clouds and cannot ensure that the identified neighbors maintain geometric consistency. To address these limitations, we approximate $k$ -NN with an efficient neighbor mapping technique using Hilbert serialized point clouds, which improves query speed while reducing computational complexity. With serialized point clouds, we can leverage point indices to perform $k$ -NN searches with $O(1)$ complexity [60].", + "bbox": [ + 88, + 373, + 483, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first serialize the union $\\mathcal{P}^g \\bigcup \\mathcal{P}^l$ using the Hilbert curve ordering. We then apply geometric partitioning [27] on the original point clouds to generate geometric labels $\\mathcal{Y}^g \\bigcup \\mathcal{Y}^l$ . The center points of these labels are incorporated as high-order bits (by bit shifting) in the serialized index, ensuring that points with different labels occupy non-overlapping index ranges. Next, for each $\\pmb{p}_i^g \\in \\mathcal{P}^g$ , we identify the $\\mathtt{k}$ nearest super-points within $\\mathcal{P}^l$ based on the serialized indexes. These geometric labels guarantee that $\\pmb{p}_i^g$ and its nearest local super-points all originate from the same instance. Let $\\mathcal{P}_{\\mathrm{k}_i}^l$ denote the set of $\\mathtt{k}$ nearest local super-points to the global super-point $\\pmb{p}_i^g$ . Consequently, $\\mathcal{F}_{\\mathrm{k}_i}^l$ are the representations of the super-points in $\\mathcal{P}_{\\mathrm{k}_i}^l$ .", + "bbox": [ + 89, + 529, + 483, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local-global representation aggregation. We make global representations more informative by combining them with information from local representations. To enable this, we present a novel two-step aggregation technique. In the first step, we employ a novel cross-attention algorithm between local and global representations using neighborhood information to update global representations, which we term as localized cross-attention. In the second step, we refine global representations through a Graph Convolutional Network based formulation based on message passing, which we term as GCN message passing (Fig. 2).", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Localized cross-attention operates with a learning-based weighting mechanism that is a function of the representations and relative positions of global super-points and their associated local neighbors. We constrain cross-attention to local neighborhood regions because $i$ ) limiting the number of points reduces computational complexity, and $ii$ ) neighbors of a given point are likely to belong to the same object. Specifically, for each point $\\pmb{p}_i^g$ and its nearest neighbors $\\mathcal{P}_{\\mathrm{k}_i}^l$ , their relative position embeddings $\\mathcal{R}_i$ are extracted as", + "bbox": [ + 511, + 373, + 906, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {i} = \\left\\{\\mathcal {R} _ {i j} = \\operatorname {p o s} \\left(\\left(\\boldsymbol {p} _ {i} ^ {g} - \\boldsymbol {p} _ {j} ^ {l}\\right) / \\sigma\\right) \\in \\mathbb {R} ^ {d} \\right\\} _ {j = 1} ^ {K ^ {l}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 510, + 903, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma > 0$ is a learnable parameter that controls the relative position scaling, and $\\mathrm{pos}(\\cdot)$ is the 3D Fourier positional embedding [52] operation, which satisfies:", + "bbox": [ + 511, + 537, + 906, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\operatorname {p o s} (\\boldsymbol {x}) = \\left[ \\sin (2 \\pi \\boldsymbol {x} \\cdot B); \\cos (2 \\pi \\boldsymbol {x} \\cdot B) \\right], \\quad \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 582, + 903, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B \\in \\mathbb{R}^{3 \\times (d/2)}$ is a learnable matrix. We update the global representations through cross-attention as follows", + "bbox": [ + 511, + 597, + 903, + 627 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {i j} = \\left(W _ {q} \\boldsymbol {f} _ {i} ^ {g}\\right) ^ {\\top} \\left(W _ {k} \\left(\\boldsymbol {f} _ {j} ^ {l} + W _ {r} \\mathcal {R} _ {i j}\\right)\\right) / \\sqrt {d},\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 628, + 854, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {i} = \\left\\{\\boldsymbol {s} _ {i j} \\right\\} _ {i = 1} ^ {K ^ {l}}, \\boldsymbol {w} _ {i} = \\text {s o f t m a x} \\left(\\boldsymbol {s} _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 648, + 903, + 670 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {f}} _ {i} ^ {g} = \\boldsymbol {f} _ {i} ^ {g} + \\boldsymbol {w} _ {i} \\left(W _ {v} \\left(\\mathcal {F} _ {\\mathcal {K} _ {i}} ^ {l} + \\mathcal {R} _ {i}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 674, + 797, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{f}}_i^g$ are the updated representations, and $W_{q},W_{k},W_{v},W_{r}$ are projection parameter matrices. Let $\\hat{\\mathcal{F}}^g = \\{\\hat{\\pmb{f}}_i^g\\}_{i = 1}^M$ denote the set of $\\hat{\\pmb{f}}_i^g$", + "bbox": [ + 511, + 700, + 903, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GCN message passing further refines $\\hat{\\mathcal{F}}^g$ by aggregating information from neighboring global super-points. Let $\\mathcal{G}^g = \\{\\mathcal{P}^g, \\mathcal{W}^g\\}$ be a $k$ -NN graph, where $\\mathcal{P}^g$ is the set of the vertices, i.e., global super-points, and $\\mathcal{W}^g \\in \\mathbb{R}^{N \\times N}$ is the adjacency matrix defining the edges. The construction of the adjacency matrix uses the similarity between superpoint representations $\\mathcal{F}^G$ . Note that we use the (original) global representations to define the adjacency matrix because we empirically experienced more stability in training as opposed to using the updated global representation,", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14372", + "bbox": [ + 478, + 948, + 519, + 960 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "likely due to the fact that they change value during training. The adjacency matrix $\\mathcal{W}^g$ with elements $\\mathcal{W}_{ij}^{g}$ is defined as", + "bbox": [ + 89, + 90, + 480, + 125 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} _ {i j} ^ {g} = \\left\\{ \\begin{array}{l l} \\frac {\\boldsymbol {f} _ {i} ^ {g \\top} \\boldsymbol {f} _ {j} ^ {g}}{\\| \\boldsymbol {f} _ {i} ^ {g} \\| \\| \\boldsymbol {f} _ {j} ^ {g} \\|} \\cdot \\left(\\boldsymbol {f} _ {i} ^ {g \\top} \\boldsymbol {f} _ {j} ^ {g} > 0\\right), & \\text {i f p _ {j} ^ {g} \\in k - N N (p _ {i} ^ {g})}, \\\\ 0, & \\text {o t h e r w i s e}. \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 126, + 467, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, we perform one-time GCN message passing (MP) on the graph $\\mathcal{G}^g$ to obtain an updated feature matrix $\\hat{\\mathcal{F}}^g$ .", + "bbox": [ + 89, + 175, + 482, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\tilde{\\mathcal{W}}^g$ be the symmetrically normalized adjacency matrix of $\\mathcal{W}^g$ by $\\tilde{\\mathcal{W}}^g = \\mathbf{D}^{-\\frac{1}{2}}\\mathcal{W}^g\\mathbf{D}^{-\\frac{1}{2}}\\in \\mathbb{R}^{M\\times M}$ where $\\mathbf{D}$ is the diagonal degree matrix of $\\mathcal{W}^g$ .", + "bbox": [ + 89, + 205, + 483, + 251 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {F}} ^ {g} \\leftarrow \\operatorname {R e L U} \\left(W _ {m} \\left(\\tilde {\\mathcal {W}} ^ {g} \\hat {\\mathcal {F}} ^ {g}\\right) + W _ {s} \\hat {\\mathcal {F}} ^ {g}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 253, + 482, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{m}$ and $W_{s}$ are trainable parameters. This MP layer structure allows each node to aggregate information from its neighbors while also preserving its own feature representation (via the skip connection).", + "bbox": [ + 89, + 280, + 483, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Learning with local representation consensus", + "text_level": 1, + "bbox": [ + 89, + 349, + 472, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The main learning objective encourages the model to generate outputs that are close to the target responses in the training dataset. This objective is typically expressed as a per-token cross-entropy loss $\\mathcal{L}_{pred}$ [9, 21] (see details in our Supp. Mat.). We found that adding a consensus term to $\\mathcal{L}_{pred}$ that encourages local regularization of representations improves training stability. Firstly, points belonging to the same object (i.e., local neighborhoods) should share similar aggregated representations. Secondly, aggregated super-point representations should remain close to their original point representations to preserve context knowledge. Our novel consensus loss $\\mathcal{L}_{\\mathrm{con}}$ is defined as", + "bbox": [ + 89, + 372, + 483, + 553 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}} = \\mathcal {L} _ {\\text {s m t}} + \\mu \\mathcal {L} _ {\\text {r e g}},\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 556, + 310, + 573 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s m t}} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\mathcal {W} _ {i j} ^ {g} \\left\\| \\frac {\\hat {\\boldsymbol {f}} _ {i} ^ {g}}{\\sqrt {d _ {i}}} - \\frac {\\hat {\\boldsymbol {f}} _ {j} ^ {g}}{\\sqrt {d _ {j}}} \\right\\|, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 575, + 482, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e g}} = \\sum_ {i = 1} ^ {N} \\left\\| \\hat {\\pmb {f}} _ {i} ^ {g} - \\pmb {f} _ {i} ^ {g} \\right\\|,\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 622, + 331, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_{k} = \\sum_{l}\\mathcal{W}_{kl}^{g}$ is the diagonal matrix representing row-wise sums of $\\mathcal{W}^g$ . The first term $\\mathcal{L}_{\\mathrm{smt}}$ enforces spatial connectivity by promoting similarity among fused representations of points within the same object, addressing the first attribute. $\\mathcal{L}_{\\mathrm{reg}}$ is a regularization term that promotes the learned representations stay close to the original global representations. $\\mu$ is a hyperparameter. The overall loss $\\mathcal{L}$ , is a weighted sum of $\\mathcal{L}_{pred}$ and $\\mathcal{L}_{con}$", + "bbox": [ + 89, + 666, + 483, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda \\mathcal {L} _ {\\text {c o n}} + \\mathcal {L} _ {\\text {p r e d}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 791, + 482, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the balance hyperparameter. By jointly training with these two loss terms, we encourage the model to learn more object-aware representations and enhance its ability to extract finer details. This approach also preserves global context information, which is essential for capturing spatial relationships between different objects.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, we use beam search to predict response $\\mathcal{O}^t$ that maximizes the following objective:", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} ^ {t} = \\arg \\max _ {\\mathcal {O}} \\Phi (\\mathcal {O} \\mid \\mathcal {P}, \\mathcal {I} ^ {t}, \\mathcal {I} ^ {v}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 133, + 906, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we set a beam size of 4.", + "bbox": [ + 513, + 162, + 718, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 193, + 645, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate PerLA on 3D Question Answering and 3D Dense Captioning downstream tasks, and compare its performance with state-of-the-art methods from the literature.", + "bbox": [ + 511, + 219, + 905, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments using the ScanNet dataset [14], which encompasses 1,201 training and 312 validation scenes, featuring diverse and complex indoor 3D environments. Language annotations are from ScanQA [2], ScanRefer [6], Nr3D [1], and the ScanNet subset of 3D-LLM [21], collectively supporting a range of tasks including instance and scene descriptions, conversations, embodied planning, and question answering. For additional data statistics, please refer to the supplementary materials.", + "bbox": [ + 511, + 267, + 906, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We follow LL3DA's evaluation protocol [9] to evaluate the quality of output responses. We use the abbreviations C, B4, M and R for CiDEr [55], BLEU-4 [43], METEOR [3], and Rouge-L [34], respectively.", + "bbox": [ + 511, + 406, + 905, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. As in [9, 12], we input 40,000 randomly sampled points from each 3D scene. We divide the point cloud into partitions $L = 6$ and choose $k = 4L = 24$ neighbors to enhance global representations. We set $\\lambda = \\mu = 0.1$ in the loss. We use the pre-trained OPT-1.3B [66] language model, kept frozen and loaded in float16 precision. We use the AdamW [38] optimizer with a weight decay of 0.1, applying a cosine annealing scheduler that decays the learning rate from $10^{-4}$ to $10^{-6}$ over approximately 100,000 iterations. All tasks are trained with a total batch size of 16. Each training process is completed within two days using up to two NVIDIA H100 P0 (96GB) GPUs. For each evaluation, we fine-tune the model's parameters on the respective task for about 30,000 iterations.", + "bbox": [ + 511, + 470, + 906, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Results", + "text_level": 1, + "bbox": [ + 511, + 694, + 607, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D question answering is a task that involves answering questions about a 3D scene. It allows a model to provide information about the objects, relationships, and attributes within a 3D environment based on a given question. Tab. 1 shows the results of ScanQA's validation and test sets. Classification-based methods (CLS) select responses from a predefined answer set. Generation-based approaches (GEN) generate the entire textual response. PerLA consistently outperforms existing approaches across all evaluation sets and metrics, in particular with $+1.34$ CiDEr score over LL3DA. We also compare against our reproduced version of LL3DA, where PerLA scores $+3.76$ CiDEr.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14373", + "bbox": [ + 480, + 948, + 517, + 960 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/754915ddb53c7db43e840ab5796eb80b07d491119b72c41f20bbb52ca0d04dce.jpg", + "table_caption": [ + "Table 1. Comparative results for 3D Question Answering on ScanQA [2] benchmark. CLS and GEN denote classification-based and generation-based methods, respectively. LL3DA (repr.) is results of LL3DA we reproduced. PerLA outperforms all the other methods." + ], + "table_footnote": [], + "table_body": "
MethodTypeValidationTest w/ objectTest w/o object
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
ScanQA[2]CLS64.8610.0813.1433.3367.2912.0413.5534.3460.2410.7512.5931.09
Clip-Guided[44]-----69.5314.6413.9435.1562.8311.7313.2832.41
Multi-CLIP[16]CLS----68.7012.6513.9735.4663.2012.8713.3632.61
3D-VLP[26]CLS66.9711.1513.5334.5170.1811.2314.1635.9763.4015.8413.1331.79
3D-VisTA[70]-----68.6010.5013.8035.5055.708.7011.6929.60
3D-LLM [21]GEN69.4012.0014.5035.7069.6011.6014.9035.30----
LL3DA[9]GEN76.7913.5315.8837.3178.1613.9716.3838.1570.2912.1914.8535.17
LL3DA (repr.)GEN74.3713.5015.0936.31--------
PerLAGEN78.1314.4917.4439.6080.9117.2116.4940.7174.8214.9715.2338.18
Δ w.r.t. LL3DA[9]-+1.34+0.96+1.56+2.29+2.75+3.24+0.11+2.56+4.53+2.78+0.38+3.01
", + "bbox": [ + 91, + 121, + 903, + 303 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e7dfa24369b410d084bdfa2c2d2c3307b13db995d359a8216dd4518dcf3c2a5d.jpg", + "image_caption": [ + "Figure 3. The qualitative comparison between our method, PerLA, and LL3DA [9] on the ScanQA [2] dataset shows that our approach achieves higher accuracy in responding to \"what\"-related questions." + ], + "image_footnote": [], + "bbox": [ + 91, + 313, + 509, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/230abff78840d6d3ab5edf3cf836a2879436956dada63378a847bf8077ebed90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 311, + 903, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 provides qualitative comparisons between LL3DA and PerLA on the ScanQA benchmark [2]. It highlights PerLA's accuracy in answering questions regarding object attributes and spatial relationships within a 3D scene. PerLA provides precise answers, correctly identifying objects such as a \"picture\", \"trash bin\", and \"TV\" along with their specific colors and types, while LL3DA often yields incorrect or less specific responses. In particular, PerLA accurately identifies the colors of objects within the scene, such as the door, the monitor, and the instrument case, correctly answering the questions about these attributes. In contrast, LL3DA frequently misidentifies colors, labeling objects as \"White\" instead of \"Black,\" for example. This illustrates PerLA's superior ability to capture fine-grained details and contextual information, delivering more accurate answers in complex environments.", + "bbox": [ + 88, + 502, + 485, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D dense captioning involves localizing and describing each 3D instance within complex 3D environments. Tab. 2 shows the results on ScanRefer [6] and $\\mathrm{Nr3D}$ [1] benchmarks. Following previous works [9, 21], we use the m@kIoU metric, where $m\\in \\{\\mathbf{C},\\mathbf{B4},\\mathbf{M},\\mathbf{R}\\}$ and $k$ denote the IoU threshold. As in [12, 25], we report C@0.25 and C@0.5 for ScanRefer, and C@0.5 for $\\mathrm{Nr3D}$ . UniT3D [13], 3DJCG [4], and 3D-VLP [26] are pre-trained on multiple 3D vision and language tasks from ScanNet scenes. UniT3D uses image caption models and multi-view im", + "bbox": [ + 88, + 750, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ages to generate extra instance captions for pre-training. For fair comparison, we report results from models trained using standard per-word cross-entropy loss without additional 3D scenes. Box annotations are estimated with Vote2CapDETR and used as visual prompts. PerLA consistently outperforms comparison methods on both benchmarks. PerLA significantly outperforms LL3DA by scoring +3.75 and +4.22 CiDEr on ScanRefer, and +3.88 CiDEr on Nr3D. These results highlight the effectiveness of PerLA in 3D dense captioning tasks.", + "bbox": [ + 511, + 502, + 906, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 4 provides examples of qualitative results of LL3DA and PerLA on Nr3D and ScanRefer. PerLA produces more accurate and detailed descriptions, effectively capturing spatial relationships and fine-grained attributes, such as the positioning of objects relative to surrounding elements.", + "bbox": [ + 511, + 652, + 908, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 739, + 663, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Local, global, and GCN representations. To investigate the impact of local and global representations, we design two variants of PerLA: one using only local representations and the other using only global representations. We also report an extended version of LL3DA, termed $\\mathrm{LL3DA}^{\\dagger}$ , in which we increase the number of query tokens used to encode local partitions. Representations are first extracted from each partition, and FPS is applied to sample 1,024 points with their corresponding representations from the", + "bbox": [ + 509, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14374", + "bbox": [ + 480, + 948, + 519, + 960 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/40a84db6c8fc9b74152688c1617e144cb94c6ff3705172b9d0a94a3cfeb0f838.jpg", + "table_caption": [ + "Table 2. Comparative results for 3D Dense Captioning on ScanRefer [6] and Nr3D [1] benchmarks. LL3DA (repr.) is results of LL3DA we reproduced. Generally, PerLA outperforms all other methods on both benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodScanRefer@0.25ScanRefer@0.5Nr3D@0.5
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
Scan2Cap[12]56.8234.1826.2955.2739.0823.3221.9744.7827.4717.2421.8049.06
MORE[25]62.9136.2526.7556.3340.9422.9321.6644.42----
SpaCap3D[56]-----44.0225.2622.3333.7119.9222.6150.50
REMAN[40]62.0136.3726.7656.2545.0026.3123.1346.9634.8120.3722.7150.90
D3Net[7]-----51.67--35.2620.4222.7753.38
Contextual[68]-----46.0723.4023.95----
UniT3D[13]----46.6927.5221.9145.98----
3DJCG[4]64.7040.1727.6359.2349.4831.6324.3650.8038.0622.8223.7752.99
3D-VLP[26]70.7341.0328.1459.7254.9432.3124.8351.51----
3D-VisTA*[70]----61.6034.1026.8055.00----
Vote2CapDETR[8]71.4539.3428.2559.6361.8134.4626.2254.4043.8426.6825.4154.43
LL3DA[9]74.1741.4127.7659.5365.1936.7925.9755.0651.1828.7525.9156.61
LL3DA (repr.)71.8639.5727.2958.3763.7935.6725.9454.5648.3828.3625.7255.66
PerLA77.9243.4128.9759.6969.4138.0229.0756.8055.0631.2428.5259.13
Δ w.r.t. LL3DA[9]+3.75+2.00+1.21+0.16+4.22+1.23+2.27+1.74+3.88+2.49+2.61+2.52
Describe this [bbox] object.
PerLA The toilet paper is on the wall next to the toilet.
LL3DA The towel closest to the toilet.
", + "bbox": [ + 91, + 119, + 903, + 481 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b03d220d4cb43b9a38d2231b9af24499221c9a326eba35d98be6b2852e225ac.jpg", + "table_caption": [ + "(a) 3D visual dense captioning on $\\mathrm{Nr3D}$ [1]" + ], + "table_footnote": [], + "table_body": "
Describe this [bbox] object.Describe this [bbox] object.
PerLA\nThere is a rectangular bathroom vanity. It is to the right of the toilet.PerLA\nThis is a white sink.\nIt is to the left of the toilet.
LL3DA\nThis is a bathroom sink. The sink is on the left side of the room.LL3DA\nThis is a white sink.\nIt is on the left side of the sink.
(b) 3D visual dense captioning on ScanRefer [6]
", + "bbox": [ + 94, + 500, + 903, + 614 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/065e45acfb73d7b974ca00b474647fbd5138615ef82db58b1053cfad138932de.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Describe this [bbox] object.Describe this [bbox] object.
PerLA\nThis is a black computer tower. It is under a desk.\nLL3DA\nThe computer is on the left side of the desk. The computer is on the right side.PerLA\nThis is a black computer tower. It is under a desk.\nLL3DA\nThere is a black desk. It is on the left side of the room.
", + "bbox": [ + 94, + 625, + 903, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4. Qualitative comparisons on the dense captioning task across the Nr3D [1] and ScanRefer [6]. We compare the results of our PerLA with LL3DA [9]. PerLA generates accurate descriptions, effectively capturing fine-grained object attributes and spatial relationships.", + "bbox": [ + 89, + 733, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "union of these partitions. These representations are then passed through a multimodal adapter to produce 32 global tokens. LL3DA generates an additional 32 global tokens from the point cloud of the entire scene, resulting in a total of 64 tokens, which are concatenated and fed into the LLM for response generation. (For details of $\\mathrm{LL3DA}^{\\dagger}$ please refer to the Supp. Mat.). Tab. 3 shows that the combination of local and global representations yields a much better", + "bbox": [ + 88, + 776, + 485, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance than the two independently used. Tab. 3 also shows that $\\mathrm{LL3DA}^{\\dagger}$ does not reach PerLA performance, underlying the importance of information exchange between local and global representations. To evaluate the effectiveness of GCN, we also report the results without using GCN (w/o GCN). Compared to our PerLA, which incorporates GCN, it consistently improves performance across the three datasets, confirming its effectiveness.", + "bbox": [ + 511, + 776, + 908, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14375", + "bbox": [ + 478, + 948, + 519, + 960 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/65fd3ce53c58d80042bb22df80896381ccf1a89c12ff30caece34afc08e90e66.jpg", + "table_caption": [ + "Table 3. Ablation study of our local-to-global, GCN representation aggregation algorithms on ScanQA, ScanRefer[6] and Nr3D[1] benchmarks. LL3DA† denotes an extended version of LL3DA with an increased number of query tokens." + ], + "table_footnote": [], + "table_body": "
MethodScanQAScanRefer@0.5Nr3D@0.5
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
\\( LL3DA^{\\dagger} \\)74.5412.8915.1136.9662.2534.5025.5553.8448.7028.2425.7255.59
Global74.4913.5015.1636.5563.7935.6725.9454.5648.3828.3625.7255.66
Local73.5512.9814.9536.0762.7934.7825.9454.1649.0928.2025.8055.89
w/o GCN77.6113.8715.9339.2869.0437.4328.7155.5954.3029.7827.4158.32
PerLA78.1314.4917.4439.6069.4138.0229.0756.8055.0631.2428.5259.13
", + "bbox": [ + 91, + 119, + 901, + 234 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2aea2dbcd627916650a43912a0f67331b78ce2b02c70c4a7eaaf61b7782412a0.jpg", + "table_caption": [ + "Table 4. Ablation study on the impact of constrained attention on ScanQA validation dataset." + ], + "table_footnote": [], + "table_body": "
cross-attentionmean poolingmax pooling
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
78.114.517.439.674.412.515.236.574.013.014.835.7
", + "bbox": [ + 91, + 270, + 478, + 325 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f994f076557d2df002db75c207ca367adad30eaf5b180967587f40f0ac105437.jpg", + "table_caption": [ + "Table 5. Ablation study on the impact of joint learning by different loss combination on ScanQA validation dataset." + ], + "table_footnote": [], + "table_body": "
LossScanQA (Validation)
LpreLsmtLregC↑B4↑M↑R↑
✓75.3113.8315.9037.48
✓✓76.6213.9616.9938.14
✓✓76.4714.0716.7238.79
✓✓✓78.1314.4917.4439.60
", + "bbox": [ + 91, + 369, + 478, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Localized cross-attention. To evaluate the effectiveness of our novel localized cross-attention module, we tested max pooling and mean pooling as alternative methods. Tab. 4 shows that localized cross-attention significantly outperforms both mean pooling and max pooling across all metrics. These results indicate that localized cross-attention enhances ability of model to generate more informative scene representations, thereby improving performance on 3D question answering tasks. This improvement comes from the incorporation of positional information and semantic similarity in cross-attention, which helps to exclude neighboring points that do not belong to the same object.", + "bbox": [ + 88, + 486, + 482, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss function. To evaluate the effectiveness of our novel loss term $\\mathcal{L}_{\\mathrm{con}}$ , we train PerLA using different combinations of $\\mathcal{L}_{\\mathrm{con}}$ ’s components (Eq. 5). Since our task is intended for downstream applications, each combination also includes the task-specific loss $\\mathcal{L}_{\\mathrm{pre}}$ . Tab. 5 shows that adding $\\mathcal{L}_{\\mathrm{smt}}$ to $\\mathcal{L}_{\\mathrm{pre}}$ improves all metrics (C, B4, M, and R). Incorporating the loss of regularization $\\mathcal{L}_{\\mathrm{reg}}$ further enhances performance, particularly in metrics C, M, and R. These results show the effectiveness of learning jointly with both semantic awareness and regularization losses.", + "bbox": [ + 88, + 670, + 482, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of partitions. To evaluate the effectiveness of different number of partitions, we test PerLA by splitting the point cloud into 4, 6, and 8 partitions. Tab. 6 shows that increasing the number of partitions generally improves performance, with 6 partitions yielding the best results. Al-", + "bbox": [ + 88, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f5be9ad6dfc37c6900dfc5d774b923bd3bd4fffb1a93f7e4f61caf85a8d67406.jpg", + "table_caption": [ + "Table 6. Ablation study on the impact of number of parts on ScanQA validation dataset." + ], + "table_footnote": [], + "table_body": "
468
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
76.913.616.137.378.114.517.439.678.014.517.239.4
", + "bbox": [ + 514, + 270, + 901, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "though 8 partitions outperform 4, they show slightly lower performance than 6. This suggests that 6 partitions provide a better balance between granularity and effectiveness, capturing relevant scene details with lower computational complexity. More ablation studies, including an analysis of performance in different training strategies, refer to Supp. Mat..", + "bbox": [ + 511, + 340, + 906, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 446, + 640, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented PerLA, a perceptive 3D language assistant capable of capturing both detailed and contextual information to enhance visual representations for LLMs. PerLA features a dual-branch architecture: the global branch processes superpoints from the whole point cloud via downsampling, while the local branch focuses on partitioned regions. We demonstrated that by integrating representations from both branches, PerLA effectively captures scene details, reducing hallucinations. Moreover, we employ a Graph Convolutional Network to facilitate information exchange among neighboring local and global superpoints, and introduce a novel loss term for local representation consensus to promote training stability. Experiments on the ScanQA, ScanRefer, and Nr3D benchmarks highlight the effectiveness of our approach, setting a new state-of-the-art performance in 3D question answering and dense captioning.", + "bbox": [ + 511, + 470, + 906, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations & Future Work. PerLA primarily focuses on enhancing performance using point cloud input, but integrating it with optimized modules, such as token merging, presents a promising direction to extend its capabilities. Although PerLA shows strong performance on standard benchmarks, future work could explore its robustness and generalizability in more complex 3D scenarios, broadening its applicability in diverse real-world settings.", + "bbox": [ + 511, + 715, + 903, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by PNRR FAIR - Future AI Research (PE00000013) and ICSC National Research Centre for HPC, Big Data and Quantum Computing (CN00000013), funded by NextGeneration EU.", + "bbox": [ + 511, + 835, + 903, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14376", + "bbox": [ + 478, + 948, + 519, + 960 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 5, 6, 7, 8, 14, 15, 16, 17, 18", + "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, 2022. 2, 5, 6, 14, 15, 16, 17, 18", + "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 5, 16", + "[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2, 6, 7", + "[5] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, pages 357–366, 2021. 2", + "[6] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 2, 5, 6, 7, 8, 14, 15, 16, 17, 18", + "[7] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In ECCV, pages 487-505. Springer, 2022. 7", + "[8] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In CVPR, pages 11124-11133, 2023. 1, 2, 3, 7, 12", + "[9] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In CVPR, 2024. 1, 2, 3, 5, 6, 7, 12, 14, 16, 17, 18", + "[10] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, YU Gang, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE TPAMI, 2024. 2", + "[11] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 1", + "[12] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, pages 3193-3203, 2021. 2, 5, 6, 7" + ], + "bbox": [ + 91, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2, 6, 7", + "[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5, 14, 15", + "[15] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. NeurIPS, 36, 2024. 2", + "[16] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329, 2023. 6", + "[17] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 3", + "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 2", + "[19] Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Henan Xiong. Scene-llm: Extending language model for 3d visual understanding and reasoning. arXiv preprint arXiv:2403.11401, 2024. 3", + "[20] Ritwik Gupta, Shufan Li, Tyler Zhu, Jitendra Malik, Trevor Darrell, and Karttikeya Mangalam. xt: Nested tokenization for larger context in large images. arXiv preprint arXiv:2403.01915, 2024. 2", + "[21] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. NeurIPS, 2023. 1, 2, 3, 5, 6, 14, 16, 17", + "[22] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. NeurIPS, 2024. 2", + "[23] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In ICML, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14377", + "bbox": [ + 480, + 948, + 519, + 960 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3d: Learning fine-grained class-agnostic 3d segmentation without manual labels. In ECCV, pages 278–295. Springer, 2025. 3", + "[25] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In ECCV, pages 528-545. Springer, 2022. 6, 7", + "[26] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In CVPR, pages 10984-10994, 2023. 2, 6, 7", + "[27] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, pages 4558-4567, 2018. 4, 14", + "[28] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2024. 15", + "[29] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In CVPR, pages 7287-7296, 2022. 2", + "[30] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3", + "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 3, 12, 13", + "[32] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2", + "[33] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaoqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In NeurIPS, 2024. 3", + "[34] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5, 16", + "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llavanext: Improved reasoning,OCR, and world knowledge, 2024.2,3", + "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 1, 2", + "[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 2", + "[38] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017. 5", + "[39] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2", + "[40] Aihua Mao, Zhi Yang, Wanxin Chen, Ran Yi, and Yong-jin Liu. Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE TVCG, 2023. 7", + "[41] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, ECCV Laboratory, 2003. 3", + "[42] Bongki Moon, Hosagrahar V Jagadish, Christos Faloutsos, and Joel H. Saltz. Analysis of the clustering properties of the hilbert space-filling curve. IEEE TKDE, 13(1):124-141, 2001. 13", + "[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 5, 16", + "[44] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In CVPR, pages 5607-5612, 2023. 2, 6", + "[45] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 12", + "[46] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. Shapellm: Universal 3d object understanding for embodied interaction. In ECCV, pages 214-238, 2025. 2", + "[47] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 1", + "[48] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2", + "[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14378", + "bbox": [ + 480, + 948, + 519, + 960 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In ICML, pages 29441-29454. PMLR, 2023. 2", + "[50] Hans Sagan. Hilbert's space-filling curve. In Space-filling curves, pages 9-30. Springer, 1994. 2, 3, 4", + "[51] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In ECCV, pages 444-462. Springer, 2025. 2", + "[52] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 33: 7537-7547, 2020. 4", + "[53] Yuan Tang, Xu Han, Xianzhi Li, Qiao Yu, Yixue Hao, Long Hu, and Min Chen. Minigpt-3d: Efficiently aligning 3d point clouds with large language models using 2d priors. ACM'MM, 2024. 2", + "[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.1, 17", + "[55] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 5, 16", + "[56] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022.2, 7", + "[57] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. ACM TOG, 42(4):1-11, 2023. 3", + "[58] Qing Wang et al. Divergence-augmented policy optimization. NeurIPS, 32, 2019. 15", + "[59] Zehan Wang, Haifeng Huang, Yang Zhao, Zhang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 2", + "[60] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 1, 3, 4", + "[61] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 2", + "[62] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 3", + "[63] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. NeurIPS, 34:30008-30022, 2021. 2", + "[64] Jianing Yang, Xuweiyi Chen, Nikhil Madaan, Madhavan Iyengar, Shengyi Qian, David F. Fouhey, and Joyce Chai. 3d-grand: A million-scale dataset for 3d-llms with better grounding and less hallucination. arXiv preprint 2406.05132, 2024. 2", + "[65] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. arXiv preprint arXiv:2112.08359, 2021. 2", + "[66] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 5, 17", + "[67] Lichen Zhao, Daigang Cai, Jing Zhang, Lu Sheng, Dong Xu, Rui Zheng, Yinjie Zhao, Lipeng Wang, and Xibo Fan. Toward explainable 3d grounded visual question answering: A new benchmark and strong baseline. IEEE TCSVT, 33(6):2935-2949, 2022. 2", + "[68] Yufeng Zhong, Long Xu, Jiebo Luo, and Lin Ma. Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925, 2022. 7", + "[69] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 2", + "[70] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 6, 7" + ], + "bbox": [ + 516, + 90, + 903, + 628 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14379", + "bbox": [ + 480, + 948, + 519, + 960 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/PerLA_ Perceptive 3D Language Assistant/54882f4a-04ec-4ab3-b66a-a66c8e83acf5_model.json b/2025/PerLA_ Perceptive 3D Language Assistant/54882f4a-04ec-4ab3-b66a-a66c8e83acf5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a6bbac8f91e35ee452fa468764fac61153ce09ee --- /dev/null +++ b/2025/PerLA_ Perceptive 3D Language Assistant/54882f4a-04ec-4ab3-b66a-a66c8e83acf5_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.271, + 0.128, + 0.728, + 0.159 + ], + "angle": 0, + "content": "PerLA: Perceptive 3D language assistant" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.191, + 0.851, + 0.227 + ], + "angle": 0, + "content": "Guofeng Mei\\(^{1}\\) Wei Lin\\(^{2}\\) Luigi Riz\\(^{1}\\) Yujiao Wu\\(^{3}\\) Fabio Poiesi\\(^{1}\\) Yiming Wang\\(^{1}\\) \n\\(^{1}\\)Fondazione Bruno Kessler, Italy \\(^{2}\\)JKU Linz, Austria \\(^{3}\\)CSIRO, Australia" + }, + { + "type": "text", + "bbox": [ + 0.446, + 0.231, + 0.547, + 0.245 + ], + "angle": 0, + "content": "gmei@fbk.eu" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.281, + 0.482, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.281, + 0.903, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.475, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Figure 1. PerLA is a 3D language assistant that integrates local details with global context to learn informative representations of 3D scenes, whereas state-of-the-art (SOTA) 3DLAs focus solely on global context information. PerLA can provide more accurate responses, correctly distinguishing between objects such as a \"black computer monitor\" and a \"black suitcase,\" where SOTA models instead fail with hallucinated responses. Examples in figures show cases where capturing details from the point cloud matters for accurate output captions." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.543, + 0.327, + 0.558 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.577, + 0.486, + 0.91 + ], + "angle": 0, + "content": "Enabling Large Language Models (LLMs) to understand the 3D physical world is an emerging yet challenging research direction. Current strategies for processing point clouds typically downsample the scene or divide it into smaller parts for separate analysis. However, both approaches risk losing key local details or global contextual information. In this paper, we introduce PerLA, a 3D language assistant designed to be more perceptive to both details and context, making visual representations more informative for the LLM. PerLA captures high-resolution (local) details in parallel from different point cloud areas and integrates them with (global) context obtained from a lower-resolution whole point cloud. We present a novel algorithm that preserves point cloud locality through the Hilbert curve and effectively aggregates local-to-global information via cross-attention and a graph neural network. Lastly, we introduce a novel loss for local representation consensus to promote training stability. PerLA outperforms state-of-the-art 3D language assistants, with gains of up to \\(+1.34\\) CiDEr on ScanQA for question answering, and \\(+4.22\\) on ScanRefer and \\(+3.88\\) on Nr3D for dense captioning. Project page: https://gfmei.github.io/PerLA" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.543, + 0.645, + 0.558 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.87 + ], + "angle": 0, + "content": "3D language assistants (3DLAs) jointly process natural language and 3D data to achieve 3D scene understanding, such as recognizing object categories, locations, appearances, and relationships, without requiring specialized models for each recognition task [9, 11, 21]. These capabilities are primarily powered by Large Language Models (LLMs) trained in large text corpora [54]. These approaches can aggregate multi-view features [21] or process point clouds [9] to generate 3D representations, which are in turn converted to tokens for the LLM [54, 66]. However, extracting multi-view representations is computationally costly and often fails to capture essential geometric properties [21]. Directly processing point clouds can yield more accurate results, yet it is even more computationally costly than handling multiview data, as point clouds typically have rather large cardinalities [60]. To address this, the point cloud cardinality can be reduced via downsampling [8, 9]. However, as with images [36], downsampling can compromise downstream task performance due to the reduced model's ability to perceive fine details of 3D scenes [47]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Fig. 1 illustrates two cases where our approach can accurately capture details and describe small objects within large" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.949, + 0.52, + 0.961 + ], + "angle": 0, + "content": "14369" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.155 + ], + "angle": 0, + "content": "scenes, while a state-of-the-art method hallucinates object details [9]. Although fine-grained attributes are critical for performance, directly extracting detailed information from high-resolution 3D data for 3DLAs remains underexplored." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.162, + 0.486, + 0.739 + ], + "angle": 0, + "content": "In this work, we aim to enhance 3DLAs' ability to perceive finer details in point clouds in order to execute downstream tasks more accurately. While increasing the number of visual tokens as the point cloud grows in size seems the straightforward solution, our empirical study (Tab. 3) shows that this solution has limited effectiveness in capturing scene details and it just increases computational burden. To address this, we propose PerLA, a novel 3DLA with a perceptive 3D scene encoder that captures detailed point cloud information, allowing the language model to generate more accurate responses without processing additional tokens. PerLA first divides the complete 3D scene into non-overlapping local parts to be processed in parallel, then integrates this local information with the global context obtained from a lower-resolution representation of the entire point cloud. Although dividing visual input has gained popularity in image processing [35], it has not yet been applied to point clouds. Processing point clouds presents unique challenges beyond those of images, as it requires handling an unordered set of points rather than rasterized pixels. To address these challenges, we serialize and partition the point cloud before encoding using a Hilbert curve approach [50], which efficiently preserves locality. We then combine local and global information through an efficient Hilbert curve-based \\( k \\)-NN search and aggregate this information via a novel cross-attention module and Graph Convolutional Network (GCN) to generate highly informative point-level representations for the LLM. Lastly, we train PerLA with a novel loss function designed to promote consensus on local representation, addressing the issue of divergent representations during local-to-global aggregation. We validate our approach on the question answering benchmark ScanQA [2], and the 3D dense captioning benchmarks ScanRefer [6] and Nr3D [1] to demonstrate the effectiveness of PerLA. PerLA demonstrates high transferability on both tasks of 3D question answering and 3D dense captioning, achieving state-of-the-art performance. In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.748, + 0.483, + 0.777 + ], + "angle": 0, + "content": "- We present a novel perceptive 3D encoder to preserve local and global information for 3D language assistant." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.779, + 0.483, + 0.823 + ], + "angle": 0, + "content": "- We introduce an efficient approach based on Hilbert curve \\( k \\)-NN search and cross attention to aggregate local and global information at point level." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.823, + 0.483, + 0.853 + ], + "angle": 0, + "content": "- We propose a novel algorithm based on Graph Neural Network to refine and enhance aggregated information." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.854, + 0.483, + 0.899 + ], + "angle": 0, + "content": "- We introduce a novel loss objective to enable local representation consensus for local-to-global information aggregation." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.748, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.09, + 0.652, + 0.107 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.12, + 0.907, + 0.363 + ], + "angle": 0, + "content": "3D language understanding involves understanding 3D scenes by describing or answering to scene-relevant questions in the format of natural language. Examples of typical downstream tasks include 3D Dense Captioning [8, 12, 56], and 3D Question Answering [2, 39, 65, 67]. 3D Dense Captioning consists of 3D localization of object instances, and textual descriptions of each object instance [4, 8, 10, 12]. 3D Question Answering requires the model equipped with a language decoder to answer questions regarding the visual context in the given 3D scene [2, 44]. Several works focus on addressing the problems of 3D-language pre-alignment [13, 26], or designing adapter layers [9, 21], or constructing 3D synthetic data [64]. In contrast, our work focuses on designing an encoding approach that can enhance 3D-language understanding by capturing fine-grained details." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.368, + 0.909, + 0.777 + ], + "angle": 0, + "content": "3D language assistants (3DLAs) can be roughly categorized into two types: object-level 3DLAs [23, 46, 53, 61] and scene-level 3DLAs [9, 21, 59, 69]. Object-level 3DLAs learn from large 3D object datasets [15] to connect object-level 3D representation with language models. However, they underperform in compositional reasoning in complex 3D scenes with numerous objects. Scene-level 3DLAs, such as 3D-LLM [21], LLaVA-3D [69], Chat-3D [59], and LL3DA [9] enable scene understanding via the interaction with objects. Because 3D data is orders of magnitude less than 2D data, existing 3DLAs address such lack of data by leveraging pre-trained 2D Large Multimodal Models (LMMs) [21, 69], or data-efficient training recipes [9, 59]. 3D-LLM [21] leverages 2D pre-trained representations of rendered multi-view images to construct 3D representations and 2D VLMs as backbones. LLaVA-3D [69] adapts LLaVA [36] for 3D scene understanding by associating 2D patch representations with their positions in 3D space. Chat-Scene [22] improves the referencing and grounding capability and models scene representations as a sequence of object-level representations. LL3DA [9] extracts point-level representations from a downsampled 3D scene, and includes both interaction prompts and textual instructions to resemble human interactions with the 3D environment. Orthogonal to the efforts of aligning 3D-language with limited 3D data in existing 3DLAs, PerLA aims to improve the capability of 3DLAs in perceiving scene details." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Visual perception enhancement on multimodal models. Multi-granularity representation learning has been explored in 2D multimodal models [5, 18, 20, 29, 49, 51, 63], showing that combining local and global views yields more informative representations than relying on a single global view. Mini-Gemini [32] introduces a dual vision encoder setup using CLIP ViT [48] and ConvNeXt [37] to process low- and high-resolution views of an input image. Mod-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.949, + 0.521, + 0.961 + ], + "angle": 0, + "content": "14370" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "els in the LLaVA-Next series [30, 35, 62] and InternLM-XComposer2 [17] use an additional branch to handle view partitions, increasing the number of visual tokens and preserving more fine-grained visual details compared to language-vision models with a single global view branch. In the 3D domain, Scene-LLM [19] and Segment3D [24] improve segmentation accuracy by transferring semantic details from multi-view images to point clouds. While extensive work has been done to enhance perception in 2D models and transfer multi-view image information into 3D, methods for preserving detailed representations from point clouds in 3DLAs remain underexplored. Our method aims to address this gap by preserving both global and local visual information to enhance 3D perception." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.312, + 0.462, + 0.329 + ], + "angle": 0, + "content": "3. Perceptive 3D language assistant (PerLA)" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.338, + 0.483, + 0.699 + ], + "angle": 0, + "content": "PerLA takes as inputs \\( i \\) a text prompt in natural language, \\( ii \\) the 3D scene in the form of a point cloud, and \\( iii \\) a visual prompt provided as either a user click or a bounding box. The text prompt is processed by a text prompt encoder to produce text representations, which are then input to both the Large Language Model (LLM) and the multimodal adapter (MMA). The text encoder is a transformer based on BLIP-2 [31]. The point cloud is processed by our perceptive scene encoder, which generates scene representations that feed into both the MMA and the subsequent encoder. We will detail the perceptive scene encoder in the next sections. The visual prompt is processed by the visual prompt encoder, which, by combining the perceptive scene encoder's representations, outputs scene representations that are further processed by the MMA. For more details on visual prompts, please refer to Supp. Mat. The MMA takes as input these multimodal representations and outputs tokens for the LLM. The MMA is implemented as a Q-former [31]. MMA's output is projected into the LLM's representation space through linear projector. Lastly, these projected representations are processed by the LLM to generate the output response. We train PerLA using data provided by [21] and finetune it for each downstream task. Fig. 2 illustrates our approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Formally, let \\(\\mathcal{P} = \\left\\{\\left(\\pmb {p}_i\\in \\mathbb{R}^3,\\pmb {f}_i\\in \\mathbb{R}^{d_0}\\right)\\mid i = 1,2,\\dots ,N\\right\\}\\) denote PerLA's input point cloud, where \\(\\pmb{p_i}\\) is a point coordinate, \\(\\pmb{f_{i}}\\) is a \\(d_0\\) -dimensional feature vector (e.g., color or normal vector) corresponding to \\(\\pmb {p}_i\\), and \\(N\\) is the number of points (cardinality). Let \\(\\mathcal{I}^t\\) and \\(\\mathcal{I}^v\\) denote the input text prompt and visual prompt, respectively. Based on these inputs, PerLA (denoted as \\(\\Phi\\)) generates a free-form natural language response \\(\\mathcal{O}^t\\) for various 3D-related tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.318, + 0.849 + ], + "angle": 0, + "content": "3.1. Perceptive scene encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To compute point-level representations through the scene encoder, \\(\\mathcal{P}\\) is typically downsampled into super-points using Farthest Point Sampling (FPS) [8, 9]. However, down" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "sampling and using only \\(\\mathcal{P}\\) may hinder the encoder to capture fine-grained details, thus affecting performance in downstream 3D scene understanding tasks. Our proposed perceptive scene encoder is designed to preserve such scene details without increasing the number of tokens or the representation dimensions of the 3D scene. We propose to split \\(\\mathcal{P}\\) into parts, and employ a pre-trained 3D scene encoder to encode these parts and the whole point cloud separately. We then aggregate the output representations from these different pieces into a single (highly-informative) representation using a cross-attention-based module and Graph Convolutional Network (GCN)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.905, + 0.427 + ], + "angle": 0, + "content": "Hilbert-based scene partitioning. We partition \\(\\mathcal{P}\\) into \\(L\\) equally-sized parts, each containing the same number of points, \\(\\left\\lfloor \\frac{N}{L} \\right\\rfloor\\), where \\(\\lfloor \\cdot \\rfloor\\) denotes the greatest integer less than or equal to its argument. Enforcing equal cardinality across parts allows for spatially smaller parts in highly structured regions (areas with more semantic information) and spatially larger parts in less structured regions (areas with less semantic information). To achieve this, we adopt a Hilbert curve approach [50] to efficiently serialize the point cloud and partition it into parts [33, 57, 60]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.905, + 0.822 + ], + "angle": 0, + "content": "Global and partial scene encoding. Both full-scene point cloud and the partial-scene point clouds are encoded separately by the same pre-trained 3D scene encoder \\(\\phi\\). The 3D scene encoder \\(\\phi\\) downsamples the input point cloud into super-points using Farthest Point Sampling (FPS) [41] and produces a representation for each super-point. The representation of the full scene, i.e., the global representation, encode the overall scene context, while the representations of the partial scenes, i.e., the local representations, encode scene details. Let \\(\\mathcal{P}^g = \\{\\pmb{p}_i^g \\in \\mathbb{R}^3\\}_{i=1}^M\\) denote the downsampled full-scene point cloud composed of \\(M\\) points, and \\(\\mathcal{F}^g = \\{\\pmb{f}_i^g \\in \\mathbb{R}^d\\}_{i=1}^M\\) denote the associated global representations. Let \\(\\mathcal{P}^l = \\{\\pmb{p}_i^l \\in \\mathbb{R}^3\\}_{i=1}^{L \\cdot M}\\) denote the set of downsampled partial-scene point clouds, and \\(\\mathcal{F}^l = \\{\\pmb{f}_i^l \\in \\mathbb{R}^d\\}_{i=1}^{L \\cdot M}\\) the associated local representations. Note that, we aggregate all the points of the downsampled partial-scene point clouds into the single set \\(\\mathcal{P}^l\\), and both the full-scene and the partial-scenes are downsampled to the same number of super-points \\(M\\). The representations \\(\\mathcal{F}^l\\) provide a more detailed view of the scene, as they are derived from the same number of points downsampled from smaller, localized regions, resulting in higher resolution compared to the global versions. Yet, their semantic visibility is focused within each local part. In the following steps, we aggregate local and global information, to enrich the representations with details within the scene context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Hilbert-based nearest-neighbor search. In order to enhance global representations through local representations, we need to first find the correspondences among them, i.e., which super-points of partial-scene point clouds are spatially neighbors with super-points of the full-scene point" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.949, + 0.518, + 0.96 + ], + "angle": 0, + "content": "14371" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.905, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.275, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Overview of PerLA. (Left): The overall pipeline of PerLA, which begins by extracting interaction-aware 3D scene representations. These representations are then projected onto the prefix of textual instructions via MMA, serving as input to a frozen language model (LLM). (Right): The detailed design of PerLA. First, the 3D scene is divided into spatially compact regions using Hilbert-based scene serialization [50]. Next, an efficient \\( k \\)-NN algorithm associates each point-level global representation with its detail-enriched local representations, creating a comprehensive scene representation through a Graph Convolutional Network (GCN). Finally, smoothness and regularization losses are applied to promote stable learning for the proposed perceptive scene encoder." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.374, + 0.484, + 0.526 + ], + "angle": 0, + "content": "cloud. This can be done via \\( k \\)-NN search. However, traditional \\( k \\)-NN searches can be computationally intensive on large-scale or high-resolution point clouds and cannot ensure that the identified neighbors maintain geometric consistency. To address these limitations, we approximate \\( k \\)-NN with an efficient neighbor mapping technique using Hilbert serialized point clouds, which improves query speed while reducing computational complexity. With serialized point clouds, we can leverage point indices to perform \\( k \\)-NN searches with \\( O(1) \\) complexity [60]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.73 + ], + "angle": 0, + "content": "Specifically, we first serialize the union \\(\\mathcal{P}^g \\bigcup \\mathcal{P}^l\\) using the Hilbert curve ordering. We then apply geometric partitioning [27] on the original point clouds to generate geometric labels \\(\\mathcal{Y}^g \\bigcup \\mathcal{Y}^l\\). The center points of these labels are incorporated as high-order bits (by bit shifting) in the serialized index, ensuring that points with different labels occupy non-overlapping index ranges. Next, for each \\(\\pmb{p}_i^g \\in \\mathcal{P}^g\\), we identify the \\(\\mathtt{k}\\) nearest super-points within \\(\\mathcal{P}^l\\) based on the serialized indexes. These geometric labels guarantee that \\(\\pmb{p}_i^g\\) and its nearest local super-points all originate from the same instance. Let \\(\\mathcal{P}_{\\mathrm{k}_i}^l\\) denote the set of \\(\\mathtt{k}\\) nearest local super-points to the global super-point \\(\\pmb{p}_i^g\\). Consequently, \\(\\mathcal{F}_{\\mathrm{k}_i}^l\\) are the representations of the super-points in \\(\\mathcal{P}_{\\mathrm{k}_i}^l\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Local-global representation aggregation. We make global representations more informative by combining them with information from local representations. To enable this, we present a novel two-step aggregation technique. In the first step, we employ a novel cross-attention algorithm between local and global representations using neighborhood information to update global representations, which we term as localized cross-attention. In the second step, we refine global representations through a Graph Convolutional Network based formulation based on message passing, which we term as GCN message passing (Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.907, + 0.51 + ], + "angle": 0, + "content": "Localized cross-attention operates with a learning-based weighting mechanism that is a function of the representations and relative positions of global super-points and their associated local neighbors. We constrain cross-attention to local neighborhood regions because \\(i\\)) limiting the number of points reduces computational complexity, and \\(ii\\)) neighbors of a given point are likely to belong to the same object. Specifically, for each point \\(\\pmb{p}_i^g\\) and its nearest neighbors \\(\\mathcal{P}_{\\mathrm{k}_i}^l\\), their relative position embeddings \\(\\mathcal{R}_i\\) are extracted as" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.511, + 0.905, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {i} = \\left\\{\\mathcal {R} _ {i j} = \\operatorname {p o s} \\left(\\left(\\boldsymbol {p} _ {i} ^ {g} - \\boldsymbol {p} _ {j} ^ {l}\\right) / \\sigma\\right) \\in \\mathbb {R} ^ {d} \\right\\} _ {j = 1} ^ {K ^ {l}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.582 + ], + "angle": 0, + "content": "where \\(\\sigma > 0\\) is a learnable parameter that controls the relative position scaling, and \\(\\mathrm{pos}(\\cdot)\\) is the 3D Fourier positional embedding [52] operation, which satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.583, + 0.905, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\left. \\operatorname {p o s} (\\boldsymbol {x}) = \\left[ \\sin (2 \\pi \\boldsymbol {x} \\cdot B); \\cos (2 \\pi \\boldsymbol {x} \\cdot B) \\right], \\quad \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.905, + 0.628 + ], + "angle": 0, + "content": "where \\(B \\in \\mathbb{R}^{3 \\times (d/2)}\\) is a learnable matrix. We update the global representations through cross-attention as follows" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.629, + 0.856, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {i j} = \\left(W _ {q} \\boldsymbol {f} _ {i} ^ {g}\\right) ^ {\\top} \\left(W _ {k} \\left(\\boldsymbol {f} _ {j} ^ {l} + W _ {r} \\mathcal {R} _ {i j}\\right)\\right) / \\sqrt {d},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.65, + 0.905, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {i} = \\left\\{\\boldsymbol {s} _ {i j} \\right\\} _ {i = 1} ^ {K ^ {l}}, \\boldsymbol {w} _ {i} = \\text {s o f t m a x} \\left(\\boldsymbol {s} _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.675, + 0.799, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {f}} _ {i} ^ {g} = \\boldsymbol {f} _ {i} ^ {g} + \\boldsymbol {w} _ {i} \\left(W _ {v} \\left(\\mathcal {F} _ {\\mathcal {K} _ {i}} ^ {l} + \\mathcal {R} _ {i}\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.905, + 0.75 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{f}}_i^g\\) are the updated representations, and \\(W_{q},W_{k},W_{v},W_{r}\\) are projection parameter matrices. Let \\(\\hat{\\mathcal{F}}^g = \\{\\hat{\\pmb{f}}_i^g\\}_{i = 1}^M\\) denote the set of \\(\\hat{\\pmb{f}}_i^g\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "GCN message passing further refines \\(\\hat{\\mathcal{F}}^g\\) by aggregating information from neighboring global super-points. Let \\(\\mathcal{G}^g = \\{\\mathcal{P}^g, \\mathcal{W}^g\\}\\) be a \\(k\\)-NN graph, where \\(\\mathcal{P}^g\\) is the set of the vertices, i.e., global super-points, and \\(\\mathcal{W}^g \\in \\mathbb{R}^{N \\times N}\\) is the adjacency matrix defining the edges. The construction of the adjacency matrix uses the similarity between superpoint representations \\(\\mathcal{F}^G\\). Note that we use the (original) global representations to define the adjacency matrix because we empirically experienced more stability in training as opposed to using the updated global representation," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.949, + 0.52, + 0.961 + ], + "angle": 0, + "content": "14372" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.125 + ], + "angle": 0, + "content": "likely due to the fact that they change value during training. The adjacency matrix \\(\\mathcal{W}^g\\) with elements \\(\\mathcal{W}_{ij}^{g}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.127, + 0.468, + 0.174 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} _ {i j} ^ {g} = \\left\\{ \\begin{array}{l l} \\frac {\\boldsymbol {f} _ {i} ^ {g \\top} \\boldsymbol {f} _ {j} ^ {g}}{\\| \\boldsymbol {f} _ {i} ^ {g} \\| \\| \\boldsymbol {f} _ {j} ^ {g} \\|} \\cdot \\left(\\boldsymbol {f} _ {i} ^ {g \\top} \\boldsymbol {f} _ {j} ^ {g} > 0\\right), & \\text {i f p _ {j} ^ {g} \\in k - N N (p _ {i} ^ {g})}, \\\\ 0, & \\text {o t h e r w i s e}. \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.176, + 0.483, + 0.207 + ], + "angle": 0, + "content": "Lastly, we perform one-time GCN message passing (MP) on the graph \\(\\mathcal{G}^g\\) to obtain an updated feature matrix \\(\\hat{\\mathcal{F}}^g\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.207, + 0.484, + 0.252 + ], + "angle": 0, + "content": "Let \\(\\tilde{\\mathcal{W}}^g\\) be the symmetrically normalized adjacency matrix of \\(\\mathcal{W}^g\\) by \\(\\tilde{\\mathcal{W}}^g = \\mathbf{D}^{-\\frac{1}{2}}\\mathcal{W}^g\\mathbf{D}^{-\\frac{1}{2}}\\in \\mathbb{R}^{M\\times M}\\) where \\(\\mathbf{D}\\) is the diagonal degree matrix of \\(\\mathcal{W}^g\\)." + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.255, + 0.483, + 0.28 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {F}} ^ {g} \\leftarrow \\operatorname {R e L U} \\left(W _ {m} \\left(\\tilde {\\mathcal {W}} ^ {g} \\hat {\\mathcal {F}} ^ {g}\\right) + W _ {s} \\hat {\\mathcal {F}} ^ {g}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.484, + 0.342 + ], + "angle": 0, + "content": "where \\( W_{m} \\) and \\( W_{s} \\) are trainable parameters. This MP layer structure allows each node to aggregate information from its neighbors while also preserving its own feature representation (via the skip connection)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.351, + 0.473, + 0.367 + ], + "angle": 0, + "content": "3.2. Learning with local representation consensus" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.554 + ], + "angle": 0, + "content": "The main learning objective encourages the model to generate outputs that are close to the target responses in the training dataset. This objective is typically expressed as a per-token cross-entropy loss \\(\\mathcal{L}_{pred}\\) [9, 21] (see details in our Supp. Mat.). We found that adding a consensus term to \\(\\mathcal{L}_{pred}\\) that encourages local regularization of representations improves training stability. Firstly, points belonging to the same object (i.e., local neighborhoods) should share similar aggregated representations. Secondly, aggregated super-point representations should remain close to their original point representations to preserve context knowledge. Our novel consensus loss \\(\\mathcal{L}_{\\mathrm{con}}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.557, + 0.312, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n}} = \\mathcal {L} _ {\\text {s m t}} + \\mu \\mathcal {L} _ {\\text {r e g}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.576, + 0.483, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s m t}} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\mathcal {W} _ {i j} ^ {g} \\left\\| \\frac {\\hat {\\boldsymbol {f}} _ {i} ^ {g}}{\\sqrt {d _ {i}}} - \\frac {\\hat {\\boldsymbol {f}} _ {j} ^ {g}}{\\sqrt {d _ {j}}} \\right\\|, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.623, + 0.332, + 0.664 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e g}} = \\sum_ {i = 1} ^ {N} \\left\\| \\hat {\\pmb {f}} _ {i} ^ {g} - \\pmb {f} _ {i} ^ {g} \\right\\|,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.484, + 0.789 + ], + "angle": 0, + "content": "where \\(d_{k} = \\sum_{l}\\mathcal{W}_{kl}^{g}\\) is the diagonal matrix representing row-wise sums of \\(\\mathcal{W}^g\\). The first term \\(\\mathcal{L}_{\\mathrm{smt}}\\) enforces spatial connectivity by promoting similarity among fused representations of points within the same object, addressing the first attribute. \\(\\mathcal{L}_{\\mathrm{reg}}\\) is a regularization term that promotes the learned representations stay close to the original global representations. \\(\\mu\\) is a hyperparameter. The overall loss \\(\\mathcal{L}\\), is a weighted sum of \\(\\mathcal{L}_{pred}\\) and \\(\\mathcal{L}_{con}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.792, + 0.483, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda \\mathcal {L} _ {\\text {c o n}} + \\mathcal {L} _ {\\text {p r e d}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the balance hyperparameter. By jointly training with these two loss terms, we encourage the model to learn more object-aware representations and enhance its ability to extract finer details. This approach also preserves global context information, which is essential for capturing spatial relationships between different objects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "During inference, we use beam search to predict response \\(\\mathcal{O}^t\\) that maximizes the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.134, + 0.907, + 0.157 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} ^ {t} = \\arg \\max _ {\\mathcal {O}} \\Phi (\\mathcal {O} \\mid \\mathcal {P}, \\mathcal {I} ^ {t}, \\mathcal {I} ^ {v}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.163, + 0.719, + 0.177 + ], + "angle": 0, + "content": "where we set a beam size of 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.194, + 0.647, + 0.211 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.906, + 0.265 + ], + "angle": 0, + "content": "We evaluate PerLA on 3D Question Answering and 3D Dense Captioning downstream tasks, and compare its performance with state-of-the-art methods from the literature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.907, + 0.405 + ], + "angle": 0, + "content": "Datasets. We conduct experiments using the ScanNet dataset [14], which encompasses 1,201 training and 312 validation scenes, featuring diverse and complex indoor 3D environments. Language annotations are from ScanQA [2], ScanRefer [6], Nr3D [1], and the ScanNet subset of 3D-LLM [21], collectively supporting a range of tasks including instance and scene descriptions, conversations, embodied planning, and question answering. For additional data statistics, please refer to the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.906, + 0.469 + ], + "angle": 0, + "content": "Metrics. We follow LL3DA's evaluation protocol [9] to evaluate the quality of output responses. We use the abbreviations C, B4, M and R for CiDEr [55], BLEU-4 [43], METEOR [3], and Rouge-L [34], respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.683 + ], + "angle": 0, + "content": "Implementation details. As in [9, 12], we input 40,000 randomly sampled points from each 3D scene. We divide the point cloud into partitions \\( L = 6 \\) and choose \\( k = 4L = 24 \\) neighbors to enhance global representations. We set \\( \\lambda = \\mu = 0.1 \\) in the loss. We use the pre-trained OPT-1.3B [66] language model, kept frozen and loaded in float16 precision. We use the AdamW [38] optimizer with a weight decay of 0.1, applying a cosine annealing scheduler that decays the learning rate from \\( 10^{-4} \\) to \\( 10^{-6} \\) over approximately 100,000 iterations. All tasks are trained with a total batch size of 16. Each training process is completed within two days using up to two NVIDIA H100 P0 (96GB) GPUs. For each evaluation, we fine-tune the model's parameters on the respective task for about 30,000 iterations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.608, + 0.709 + ], + "angle": 0, + "content": "4.1. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "3D question answering is a task that involves answering questions about a 3D scene. It allows a model to provide information about the objects, relationships, and attributes within a 3D environment based on a given question. Tab. 1 shows the results of ScanQA's validation and test sets. Classification-based methods (CLS) select responses from a predefined answer set. Generation-based approaches (GEN) generate the entire textual response. PerLA consistently outperforms existing approaches across all evaluation sets and metrics, in particular with \\(+1.34\\) CiDEr score over LL3DA. We also compare against our reproduced version of LL3DA, where PerLA scores \\(+3.76\\) CiDEr." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.949, + 0.519, + 0.961 + ], + "angle": 0, + "content": "14373" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparative results for 3D Question Answering on ScanQA [2] benchmark. CLS and GEN denote classification-based and generation-based methods, respectively. LL3DA (repr.) is results of LL3DA we reproduced. PerLA outperforms all the other methods." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.122, + 0.905, + 0.304 + ], + "angle": 0, + "content": "
MethodTypeValidationTest w/ objectTest w/o object
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
ScanQA[2]CLS64.8610.0813.1433.3367.2912.0413.5534.3460.2410.7512.5931.09
Clip-Guided[44]-----69.5314.6413.9435.1562.8311.7313.2832.41
Multi-CLIP[16]CLS----68.7012.6513.9735.4663.2012.8713.3632.61
3D-VLP[26]CLS66.9711.1513.5334.5170.1811.2314.1635.9763.4015.8413.1331.79
3D-VisTA[70]-----68.6010.5013.8035.5055.708.7011.6929.60
3D-LLM [21]GEN69.4012.0014.5035.7069.6011.6014.9035.30----
LL3DA[9]GEN76.7913.5315.8837.3178.1613.9716.3838.1570.2912.1914.8535.17
LL3DA (repr.)GEN74.3713.5015.0936.31--------
PerLAGEN78.1314.4917.4439.6080.9117.2116.4940.7174.8214.9715.2338.18
Δ w.r.t. LL3DA[9]-+1.34+0.96+1.56+2.29+2.75+3.24+0.11+2.56+4.53+2.78+0.38+3.01
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.314, + 0.511, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.313, + 0.905, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.466, + 0.908, + 0.494 + ], + "angle": 0, + "content": "Figure 3. The qualitative comparison between our method, PerLA, and LL3DA [9] on the ScanQA [2] dataset shows that our approach achieves higher accuracy in responding to \"what\"-related questions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.503, + 0.486, + 0.744 + ], + "angle": 0, + "content": "Fig. 3 provides qualitative comparisons between LL3DA and PerLA on the ScanQA benchmark [2]. It highlights PerLA's accuracy in answering questions regarding object attributes and spatial relationships within a 3D scene. PerLA provides precise answers, correctly identifying objects such as a \"picture\", \"trash bin\", and \"TV\" along with their specific colors and types, while LL3DA often yields incorrect or less specific responses. In particular, PerLA accurately identifies the colors of objects within the scene, such as the door, the monitor, and the instrument case, correctly answering the questions about these attributes. In contrast, LL3DA frequently misidentifies colors, labeling objects as \"White\" instead of \"Black,\" for example. This illustrates PerLA's superior ability to capture fine-grained details and contextual information, delivering more accurate answers in complex environments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "3D dense captioning involves localizing and describing each 3D instance within complex 3D environments. Tab. 2 shows the results on ScanRefer [6] and \\(\\mathrm{Nr3D}\\) [1] benchmarks. Following previous works [9, 21], we use the m@kIoU metric, where \\(m\\in \\{\\mathbf{C},\\mathbf{B4},\\mathbf{M},\\mathbf{R}\\}\\) and \\(k\\) denote the IoU threshold. As in [12, 25], we report C@0.25 and C@0.5 for ScanRefer, and C@0.5 for \\(\\mathrm{Nr3D}\\). UniT3D [13], 3DJCG [4], and 3D-VLP [26] are pre-trained on multiple 3D vision and language tasks from ScanNet scenes. UniT3D uses image caption models and multi-view im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.653 + ], + "angle": 0, + "content": "ages to generate extra instance captions for pre-training. For fair comparison, we report results from models trained using standard per-word cross-entropy loss without additional 3D scenes. Box annotations are estimated with Vote2CapDETR and used as visual prompts. PerLA consistently outperforms comparison methods on both benchmarks. PerLA significantly outperforms LL3DA by scoring +3.75 and +4.22 CiDEr on ScanRefer, and +3.88 CiDEr on Nr3D. These results highlight the effectiveness of PerLA in 3D dense captioning tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.909, + 0.731 + ], + "angle": 0, + "content": "Fig. 4 provides examples of qualitative results of LL3DA and PerLA on Nr3D and ScanRefer. PerLA produces more accurate and detailed descriptions, effectively capturing spatial relationships and fine-grained attributes, such as the positioning of objects relative to surrounding elements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.665, + 0.758 + ], + "angle": 0, + "content": "4.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Local, global, and GCN representations. To investigate the impact of local and global representations, we design two variants of PerLA: one using only local representations and the other using only global representations. We also report an extended version of LL3DA, termed \\(\\mathrm{LL3DA}^{\\dagger}\\), in which we increase the number of query tokens used to encode local partitions. Representations are first extracted from each partition, and FPS is applied to sample 1,024 points with their corresponding representations from the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.949, + 0.52, + 0.961 + ], + "angle": 0, + "content": "14374" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.117 + ], + "angle": 0, + "content": "Table 2. Comparative results for 3D Dense Captioning on ScanRefer [6] and Nr3D [1] benchmarks. LL3DA (repr.) is results of LL3DA we reproduced. Generally, PerLA outperforms all other methods on both benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.12, + 0.905, + 0.482 + ], + "angle": 0, + "content": "
MethodScanRefer@0.25ScanRefer@0.5Nr3D@0.5
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
Scan2Cap[12]56.8234.1826.2955.2739.0823.3221.9744.7827.4717.2421.8049.06
MORE[25]62.9136.2526.7556.3340.9422.9321.6644.42----
SpaCap3D[56]-----44.0225.2622.3333.7119.9222.6150.50
REMAN[40]62.0136.3726.7656.2545.0026.3123.1346.9634.8120.3722.7150.90
D3Net[7]-----51.67--35.2620.4222.7753.38
Contextual[68]-----46.0723.4023.95----
UniT3D[13]----46.6927.5221.9145.98----
3DJCG[4]64.7040.1727.6359.2349.4831.6324.3650.8038.0622.8223.7752.99
3D-VLP[26]70.7341.0328.1459.7254.9432.3124.8351.51----
3D-VisTA*[70]----61.6034.1026.8055.00----
Vote2CapDETR[8]71.4539.3428.2559.6361.8134.4626.2254.4043.8426.6825.4154.43
LL3DA[9]74.1741.4127.7659.5365.1936.7925.9755.0651.1828.7525.9156.61
LL3DA (repr.)71.8639.5727.2958.3763.7935.6725.9454.5648.3828.3625.7255.66
PerLA77.9243.4128.9759.6969.4138.0229.0756.8055.0631.2428.5259.13
Δ w.r.t. LL3DA[9]+3.75+2.00+1.21+0.16+4.22+1.23+2.27+1.74+3.88+2.49+2.61+2.52
Describe this [bbox] object.
PerLA The toilet paper is on the wall next to the toilet.
LL3DA The towel closest to the toilet.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.415, + 0.482, + 0.644, + 0.494 + ], + "angle": 0, + "content": "(a) 3D visual dense captioning on \\(\\mathrm{Nr3D}\\) [1]" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.501, + 0.905, + 0.615 + ], + "angle": 0, + "content": "
Describe this [bbox] object.Describe this [bbox] object.
PerLA\nThere is a rectangular bathroom vanity. It is to the right of the toilet.PerLA\nThis is a white sink.\nIt is to the left of the toilet.
LL3DA\nThis is a bathroom sink. The sink is on the left side of the room.LL3DA\nThis is a white sink.\nIt is on the left side of the sink.
(b) 3D visual dense captioning on ScanRefer [6]
" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.625, + 0.905, + 0.733 + ], + "angle": 0, + "content": "
Describe this [bbox] object.Describe this [bbox] object.
PerLA\nThis is a black computer tower. It is under a desk.\nLL3DA\nThe computer is on the left side of the desk. The computer is on the right side.PerLA\nThis is a black computer tower. It is under a desk.\nLL3DA\nThere is a black desk. It is on the left side of the room.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.734, + 0.908, + 0.763 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons on the dense captioning task across the Nr3D [1] and ScanRefer [6]. We compare the results of our PerLA with LL3DA [9]. PerLA generates accurate descriptions, effectively capturing fine-grained object attributes and spatial relationships." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.777, + 0.486, + 0.899 + ], + "angle": 0, + "content": "union of these partitions. These representations are then passed through a multimodal adapter to produce 32 global tokens. LL3DA generates an additional 32 global tokens from the point cloud of the entire scene, resulting in a total of 64 tokens, which are concatenated and fed into the LLM for response generation. (For details of \\(\\mathrm{LL3DA}^{\\dagger}\\) please refer to the Supp. Mat.). Tab. 3 shows that the combination of local and global representations yields a much better" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.909, + 0.899 + ], + "angle": 0, + "content": "performance than the two independently used. Tab. 3 also shows that \\(\\mathrm{LL3DA}^{\\dagger}\\) does not reach PerLA performance, underlying the importance of information exchange between local and global representations. To evaluate the effectiveness of GCN, we also report the results without using GCN (w/o GCN). Compared to our PerLA, which incorporates GCN, it consistently improves performance across the three datasets, confirming its effectiveness." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.949, + 0.52, + 0.961 + ], + "angle": 0, + "content": "14375" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.905, + 0.118 + ], + "angle": 0, + "content": "Table 3. Ablation study of our local-to-global, GCN representation aggregation algorithms on ScanQA, ScanRefer[6] and Nr3D[1] benchmarks. LL3DA† denotes an extended version of LL3DA with an increased number of query tokens." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.12, + 0.903, + 0.235 + ], + "angle": 0, + "content": "
MethodScanQAScanRefer@0.5Nr3D@0.5
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
\\( LL3DA^{\\dagger} \\)74.5412.8915.1136.9662.2534.5025.5553.8448.7028.2425.7255.59
Global74.4913.5015.1636.5563.7935.6725.9454.5648.3828.3625.7255.66
Local73.5512.9814.9536.0762.7934.7825.9454.1649.0928.2025.8055.89
w/o GCN77.6113.8715.9339.2869.0437.4328.7155.5954.3029.7827.4158.32
PerLA78.1314.4917.4439.6069.4138.0229.0756.8055.0631.2428.5259.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.244, + 0.483, + 0.271 + ], + "angle": 0, + "content": "Table 4. Ablation study on the impact of constrained attention on ScanQA validation dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.271, + 0.48, + 0.327 + ], + "angle": 0, + "content": "
cross-attentionmean poolingmax pooling
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
78.114.517.439.674.412.515.236.574.013.014.835.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.341, + 0.483, + 0.369 + ], + "angle": 0, + "content": "Table 5. Ablation study on the impact of joint learning by different loss combination on ScanQA validation dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.37, + 0.48, + 0.471 + ], + "angle": 0, + "content": "
LossScanQA (Validation)
LpreLsmtLregC↑B4↑M↑R↑
✓75.3113.8315.9037.48
✓✓76.6213.9616.9938.14
✓✓76.4714.0716.7238.79
✓✓✓78.1314.4917.4439.60
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.487, + 0.483, + 0.669 + ], + "angle": 0, + "content": "Localized cross-attention. To evaluate the effectiveness of our novel localized cross-attention module, we tested max pooling and mean pooling as alternative methods. Tab. 4 shows that localized cross-attention significantly outperforms both mean pooling and max pooling across all metrics. These results indicate that localized cross-attention enhances ability of model to generate more informative scene representations, thereby improving performance on 3D question answering tasks. This improvement comes from the incorporation of positional information and semantic similarity in cross-attention, which helps to exclude neighboring points that do not belong to the same object." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.671, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Loss function. To evaluate the effectiveness of our novel loss term \\(\\mathcal{L}_{\\mathrm{con}}\\), we train PerLA using different combinations of \\(\\mathcal{L}_{\\mathrm{con}}\\)’s components (Eq. 5). Since our task is intended for downstream applications, each combination also includes the task-specific loss \\(\\mathcal{L}_{\\mathrm{pre}}\\). Tab. 5 shows that adding \\(\\mathcal{L}_{\\mathrm{smt}}\\) to \\(\\mathcal{L}_{\\mathrm{pre}}\\) improves all metrics (C, B4, M, and R). Incorporating the loss of regularization \\(\\mathcal{L}_{\\mathrm{reg}}\\) further enhances performance, particularly in metrics C, M, and R. These results show the effectiveness of learning jointly with both semantic awareness and regularization losses." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Number of partitions. To evaluate the effectiveness of different number of partitions, we test PerLA by splitting the point cloud into 4, 6, and 8 partitions. Tab. 6 shows that increasing the number of partitions generally improves performance, with 6 partitions yielding the best results. Al-" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.244, + 0.905, + 0.271 + ], + "angle": 0, + "content": "Table 6. Ablation study on the impact of number of parts on ScanQA validation dataset." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.271, + 0.903, + 0.327 + ], + "angle": 0, + "content": "
468
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
76.913.616.137.378.114.517.439.678.014.517.239.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.435 + ], + "angle": 0, + "content": "though 8 partitions outperform 4, they show slightly lower performance than 6. This suggests that 6 partitions provide a better balance between granularity and effectiveness, capturing relevant scene details with lower computational complexity. More ablation studies, including an analysis of performance in different training strategies, refer to Supp. Mat.." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.447, + 0.642, + 0.463 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.713 + ], + "angle": 0, + "content": "We presented PerLA, a perceptive 3D language assistant capable of capturing both detailed and contextual information to enhance visual representations for LLMs. PerLA features a dual-branch architecture: the global branch processes superpoints from the whole point cloud via downsampling, while the local branch focuses on partitioned regions. We demonstrated that by integrating representations from both branches, PerLA effectively captures scene details, reducing hallucinations. Moreover, we employ a Graph Convolutional Network to facilitate information exchange among neighboring local and global superpoints, and introduce a novel loss term for local representation consensus to promote training stability. Experiments on the ScanQA, ScanRefer, and Nr3D benchmarks highlight the effectiveness of our approach, setting a new state-of-the-art performance in 3D question answering and dense captioning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.905, + 0.837 + ], + "angle": 0, + "content": "Limitations & Future Work. PerLA primarily focuses on enhancing performance using point cloud input, but integrating it with optimized modules, such as token merging, presents a promising direction to extend its capabilities. Although PerLA shows strong performance on standard benchmarks, future work could explore its robustness and generalizability in more complex 3D scenarios, broadening its applicability in diverse real-world settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.837, + 0.905, + 0.898 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by PNRR FAIR - Future AI Research (PE00000013) and ICSC National Research Centre for HPC, Big Data and Quantum Computing (CN00000013), funded by NextGeneration EU." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.949, + 0.521, + 0.961 + ], + "angle": 0, + "content": "14376" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.116, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 5, 6, 7, 8, 14, 15, 16, 17, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.193, + 0.484, + 0.252 + ], + "angle": 0, + "content": "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, 2022. 2, 5, 6, 14, 15, 16, 17, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.255, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 5, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, pages 357–366, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.44, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[6] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 2, 5, 6, 7, 8, 14, 15, 16, 17, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.562 + ], + "angle": 0, + "content": "[7] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In ECCV, pages 487-505. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.483, + 0.623 + ], + "angle": 0, + "content": "[8] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In CVPR, pages 11124-11133, 2023. 1, 2, 3, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.626, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[9] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In CVPR, 2024. 1, 2, 3, 5, 6, 7, 12, 14, 16, 17, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.483, + 0.776 + ], + "angle": 0, + "content": "[10] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, YU Gang, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE TPAMI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.779, + 0.483, + 0.839 + ], + "angle": 0, + "content": "[11] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.841, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, pages 3193-3203, 2021. 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.116, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "[13] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.155, + 0.905, + 0.214 + ], + "angle": 0, + "content": "[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.292 + ], + "angle": 0, + "content": "[15] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. NeurIPS, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[16] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.905, + 0.492 + ], + "angle": 0, + "content": "[17] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.619 + ], + "angle": 0, + "content": "[19] Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Henan Xiong. Scene-llm: Extending language model for 3d visual understanding and reasoning. arXiv preprint arXiv:2403.11401, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.682 + ], + "angle": 0, + "content": "[20] Ritwik Gupta, Shufan Li, Tyler Zhu, Jitendra Malik, Trevor Darrell, and Karttikeya Mangalam. xt: Nested tokenization for larger context in large images. arXiv preprint arXiv:2403.01915, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[21] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. NeurIPS, 2023. 1, 2, 3, 5, 6, 14, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.823 + ], + "angle": 0, + "content": "[22] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.826, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[23] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In ICML, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.949, + 0.52, + 0.961 + ], + "angle": 0, + "content": "14377" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.091, + 0.486, + 0.181 + ], + "angle": 0, + "content": "[24] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3d: Learning fine-grained class-agnostic 3d segmentation without manual labels. In ECCV, pages 278–295. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.184, + 0.485, + 0.245 + ], + "angle": 0, + "content": "[25] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In ECCV, pages 528-545. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[26] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In CVPR, pages 10984-10994, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.309, + 0.484, + 0.355 + ], + "angle": 0, + "content": "[27] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, pages 4558-4567, 2018. 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[28] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2024. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.484, + 0.478 + ], + "angle": 0, + "content": "[29] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In CVPR, pages 7287-7296, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.481, + 0.484, + 0.556 + ], + "angle": 0, + "content": "[30] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 3, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.484, + 0.712 + ], + "angle": 0, + "content": "[32] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.714, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[33] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaoqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.822 + ], + "angle": 0, + "content": "[34] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.824, + 0.484, + 0.883 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llavanext: Improved reasoning,OCR, and world knowledge, 2024.2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.091, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.55, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.181 + ], + "angle": 0, + "content": "[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.184, + 0.905, + 0.228 + ], + "angle": 0, + "content": "[38] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.23, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[39] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.35 + ], + "angle": 0, + "content": "[40] Aihua Mao, Zhi Yang, Wanxin Chen, Ran Yi, and Yong-jin Liu. Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE TVCG, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.352, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[41] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, ECCV Laboratory, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[42] Bongki Moon, Hosagrahar V Jagadish, Christos Faloutsos, and Joel H. Saltz. Analysis of the clustering properties of the hilbert space-filling curve. IEEE TKDE, 13(1):124-141, 2001. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 5, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[44] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In CVPR, pages 5607-5612, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.597, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[45] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[46] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. Shapellm: Universal 3d object understanding for embodied interaction. In ECCV, pages 214-238, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.792 + ], + "angle": 0, + "content": "[47] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.795, + 0.905, + 0.855 + ], + "angle": 0, + "content": "[48] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.857, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.949, + 0.52, + 0.961 + ], + "angle": 0, + "content": "14378" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.128, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In ICML, pages 29441-29454. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.138, + 0.482, + 0.168 + ], + "angle": 0, + "content": "[50] Hans Sagan. Hilbert's space-filling curve. In Space-filling curves, pages 9-30. Springer, 1994. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.17, + 0.482, + 0.214 + ], + "angle": 0, + "content": "[51] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In ECCV, pages 444-462. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.215, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[52] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 33: 7537-7547, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.366 + ], + "angle": 0, + "content": "[53] Yuan Tang, Xu Han, Xianzhi Li, Qiao Yu, Yixue Hao, Long Hu, and Min Chen. Minigpt-3d: Efficiently aligning 3d point clouds with large language models using 2d priors. ACM'MM, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.368, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.1, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.459, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[55] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 5, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[56] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022.2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.624 + ], + "angle": 0, + "content": "[57] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. ACM TOG, 42(4):1-11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.627, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[58] Qing Wang et al. Divergence-augmented policy optimization. NeurIPS, 32, 2019. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[59] Zehan Wang, Haifeng Huang, Yang Zhao, Zhang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.793 + ], + "angle": 0, + "content": "[60] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.796, + 0.482, + 0.855 + ], + "angle": 0, + "content": "[61] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.857, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[62] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.55, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.182 + ], + "angle": 0, + "content": "[63] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. NeurIPS, 34:30008-30022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.185, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[64] Jianing Yang, Xuweiyi Chen, Nikhil Madaan, Madhavan Iyengar, Shengyi Qian, David F. Fouhey, and Joyce Chai. 3d-grand: A million-scale dataset for 3d-llms with better grounding and less hallucination. arXiv preprint 2406.05132, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.307 + ], + "angle": 0, + "content": "[65] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. arXiv preprint arXiv:2112.08359, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.309, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[66] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 5, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[67] Lichen Zhao, Daigang Cai, Jing Zhang, Lu Sheng, Dong Xu, Rui Zheng, Yinjie Zhao, Lipeng Wang, and Xibo Fan. Toward explainable 3d grounded visual question answering: A new benchmark and strong baseline. IEEE TCSVT, 33(6):2935-2949, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.507 + ], + "angle": 0, + "content": "[68] Yufeng Zhong, Long Xu, Jiebo Luo, and Lin Ma. Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.509, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[69] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.57, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[70] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.949, + 0.52, + 0.961 + ], + "angle": 0, + "content": "14379" + } + ] +] \ No newline at end of file diff --git a/2025/PerLA_ Perceptive 3D Language Assistant/54882f4a-04ec-4ab3-b66a-a66c8e83acf5_origin.pdf b/2025/PerLA_ Perceptive 3D Language Assistant/54882f4a-04ec-4ab3-b66a-a66c8e83acf5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..84af781003f66142813c2adea9aa4e2b8a3e0143 --- /dev/null +++ b/2025/PerLA_ Perceptive 3D Language Assistant/54882f4a-04ec-4ab3-b66a-a66c8e83acf5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:496d0a003ce68144f02cd794da5d66043de364c8308b1ca5473167d811e0d45b +size 5985883 diff --git a/2025/PerLA_ Perceptive 3D Language Assistant/full.md b/2025/PerLA_ Perceptive 3D Language Assistant/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e5166612df0cf1992ecc8d49462865f3da902d4c --- /dev/null +++ b/2025/PerLA_ Perceptive 3D Language Assistant/full.md @@ -0,0 +1,320 @@ +# PerLA: Perceptive 3D language assistant + +Guofeng Mei $^{1}$ Wei Lin $^{2}$ Luigi Riz $^{1}$ Yujiao Wu $^{3}$ Fabio Poiesi $^{1}$ Yiming Wang $^{1}$ $^{1}$ Fondazione Bruno Kessler, Italy $^{2}$ JKU Linz, Austria $^{3}$ CSIRO, Australia + +gmei@fbk.eu + +![](images/f170598bfc50c4a30adbdff664e93b3edefe38b17009eb83b905c054e915b4da.jpg) + +![](images/ac232aac1bab0d051b6c7f3e2a340086609463c42d91248458bb2b4a53e15546.jpg) +Figure 1. PerLA is a 3D language assistant that integrates local details with global context to learn informative representations of 3D scenes, whereas state-of-the-art (SOTA) 3DLAs focus solely on global context information. PerLA can provide more accurate responses, correctly distinguishing between objects such as a "black computer monitor" and a "black suitcase," where SOTA models instead fail with hallucinated responses. Examples in figures show cases where capturing details from the point cloud matters for accurate output captions. + +# Abstract + +Enabling Large Language Models (LLMs) to understand the 3D physical world is an emerging yet challenging research direction. Current strategies for processing point clouds typically downsample the scene or divide it into smaller parts for separate analysis. However, both approaches risk losing key local details or global contextual information. In this paper, we introduce PerLA, a 3D language assistant designed to be more perceptive to both details and context, making visual representations more informative for the LLM. PerLA captures high-resolution (local) details in parallel from different point cloud areas and integrates them with (global) context obtained from a lower-resolution whole point cloud. We present a novel algorithm that preserves point cloud locality through the Hilbert curve and effectively aggregates local-to-global information via cross-attention and a graph neural network. Lastly, we introduce a novel loss for local representation consensus to promote training stability. PerLA outperforms state-of-the-art 3D language assistants, with gains of up to $+1.34$ CiDEr on ScanQA for question answering, and $+4.22$ on ScanRefer and $+3.88$ on Nr3D for dense captioning. Project page: https://gfmei.github.io/PerLA + +# 1. Introduction + +3D language assistants (3DLAs) jointly process natural language and 3D data to achieve 3D scene understanding, such as recognizing object categories, locations, appearances, and relationships, without requiring specialized models for each recognition task [9, 11, 21]. These capabilities are primarily powered by Large Language Models (LLMs) trained in large text corpora [54]. These approaches can aggregate multi-view features [21] or process point clouds [9] to generate 3D representations, which are in turn converted to tokens for the LLM [54, 66]. However, extracting multi-view representations is computationally costly and often fails to capture essential geometric properties [21]. Directly processing point clouds can yield more accurate results, yet it is even more computationally costly than handling multiview data, as point clouds typically have rather large cardinalities [60]. To address this, the point cloud cardinality can be reduced via downsampling [8, 9]. However, as with images [36], downsampling can compromise downstream task performance due to the reduced model's ability to perceive fine details of 3D scenes [47]. + +Fig. 1 illustrates two cases where our approach can accurately capture details and describe small objects within large + +scenes, while a state-of-the-art method hallucinates object details [9]. Although fine-grained attributes are critical for performance, directly extracting detailed information from high-resolution 3D data for 3DLAs remains underexplored. + +In this work, we aim to enhance 3DLAs' ability to perceive finer details in point clouds in order to execute downstream tasks more accurately. While increasing the number of visual tokens as the point cloud grows in size seems the straightforward solution, our empirical study (Tab. 3) shows that this solution has limited effectiveness in capturing scene details and it just increases computational burden. To address this, we propose PerLA, a novel 3DLA with a perceptive 3D scene encoder that captures detailed point cloud information, allowing the language model to generate more accurate responses without processing additional tokens. PerLA first divides the complete 3D scene into non-overlapping local parts to be processed in parallel, then integrates this local information with the global context obtained from a lower-resolution representation of the entire point cloud. Although dividing visual input has gained popularity in image processing [35], it has not yet been applied to point clouds. Processing point clouds presents unique challenges beyond those of images, as it requires handling an unordered set of points rather than rasterized pixels. To address these challenges, we serialize and partition the point cloud before encoding using a Hilbert curve approach [50], which efficiently preserves locality. We then combine local and global information through an efficient Hilbert curve-based $k$ -NN search and aggregate this information via a novel cross-attention module and Graph Convolutional Network (GCN) to generate highly informative point-level representations for the LLM. Lastly, we train PerLA with a novel loss function designed to promote consensus on local representation, addressing the issue of divergent representations during local-to-global aggregation. We validate our approach on the question answering benchmark ScanQA [2], and the 3D dense captioning benchmarks ScanRefer [6] and Nr3D [1] to demonstrate the effectiveness of PerLA. PerLA demonstrates high transferability on both tasks of 3D question answering and 3D dense captioning, achieving state-of-the-art performance. In summary, our contributions are: + +- We present a novel perceptive 3D encoder to preserve local and global information for 3D language assistant. +- We introduce an efficient approach based on Hilbert curve $k$ -NN search and cross attention to aggregate local and global information at point level. +- We propose a novel algorithm based on Graph Neural Network to refine and enhance aggregated information. +- We introduce a novel loss objective to enable local representation consensus for local-to-global information aggregation. + +# 2. Related work + +3D language understanding involves understanding 3D scenes by describing or answering to scene-relevant questions in the format of natural language. Examples of typical downstream tasks include 3D Dense Captioning [8, 12, 56], and 3D Question Answering [2, 39, 65, 67]. 3D Dense Captioning consists of 3D localization of object instances, and textual descriptions of each object instance [4, 8, 10, 12]. 3D Question Answering requires the model equipped with a language decoder to answer questions regarding the visual context in the given 3D scene [2, 44]. Several works focus on addressing the problems of 3D-language pre-alignment [13, 26], or designing adapter layers [9, 21], or constructing 3D synthetic data [64]. In contrast, our work focuses on designing an encoding approach that can enhance 3D-language understanding by capturing fine-grained details. + +3D language assistants (3DLAs) can be roughly categorized into two types: object-level 3DLAs [23, 46, 53, 61] and scene-level 3DLAs [9, 21, 59, 69]. Object-level 3DLAs learn from large 3D object datasets [15] to connect object-level 3D representation with language models. However, they underperform in compositional reasoning in complex 3D scenes with numerous objects. Scene-level 3DLAs, such as 3D-LLM [21], LLaVA-3D [69], Chat-3D [59], and LL3DA [9] enable scene understanding via the interaction with objects. Because 3D data is orders of magnitude less than 2D data, existing 3DLAs address such lack of data by leveraging pre-trained 2D Large Multimodal Models (LMMs) [21, 69], or data-efficient training recipes [9, 59]. 3D-LLM [21] leverages 2D pre-trained representations of rendered multi-view images to construct 3D representations and 2D VLMs as backbones. LLaVA-3D [69] adapts LLaVA [36] for 3D scene understanding by associating 2D patch representations with their positions in 3D space. Chat-Scene [22] improves the referencing and grounding capability and models scene representations as a sequence of object-level representations. LL3DA [9] extracts point-level representations from a downsampled 3D scene, and includes both interaction prompts and textual instructions to resemble human interactions with the 3D environment. Orthogonal to the efforts of aligning 3D-language with limited 3D data in existing 3DLAs, PerLA aims to improve the capability of 3DLAs in perceiving scene details. + +Visual perception enhancement on multimodal models. Multi-granularity representation learning has been explored in 2D multimodal models [5, 18, 20, 29, 49, 51, 63], showing that combining local and global views yields more informative representations than relying on a single global view. Mini-Gemini [32] introduces a dual vision encoder setup using CLIP ViT [48] and ConvNeXt [37] to process low- and high-resolution views of an input image. Mod- + +els in the LLaVA-Next series [30, 35, 62] and InternLM-XComposer2 [17] use an additional branch to handle view partitions, increasing the number of visual tokens and preserving more fine-grained visual details compared to language-vision models with a single global view branch. In the 3D domain, Scene-LLM [19] and Segment3D [24] improve segmentation accuracy by transferring semantic details from multi-view images to point clouds. While extensive work has been done to enhance perception in 2D models and transfer multi-view image information into 3D, methods for preserving detailed representations from point clouds in 3DLAs remain underexplored. Our method aims to address this gap by preserving both global and local visual information to enhance 3D perception. + +# 3. Perceptive 3D language assistant (PerLA) + +PerLA takes as inputs $i$ a text prompt in natural language, $ii$ the 3D scene in the form of a point cloud, and $iii$ a visual prompt provided as either a user click or a bounding box. The text prompt is processed by a text prompt encoder to produce text representations, which are then input to both the Large Language Model (LLM) and the multimodal adapter (MMA). The text encoder is a transformer based on BLIP-2 [31]. The point cloud is processed by our perceptive scene encoder, which generates scene representations that feed into both the MMA and the subsequent encoder. We will detail the perceptive scene encoder in the next sections. The visual prompt is processed by the visual prompt encoder, which, by combining the perceptive scene encoder's representations, outputs scene representations that are further processed by the MMA. For more details on visual prompts, please refer to Supp. Mat. The MMA takes as input these multimodal representations and outputs tokens for the LLM. The MMA is implemented as a Q-former [31]. MMA's output is projected into the LLM's representation space through linear projector. Lastly, these projected representations are processed by the LLM to generate the output response. We train PerLA using data provided by [21] and finetune it for each downstream task. Fig. 2 illustrates our approach. + +Formally, let $\mathcal{P} = \left\{\left(\pmb {p}_i\in \mathbb{R}^3,\pmb {f}_i\in \mathbb{R}^{d_0}\right)\mid i = 1,2,\dots ,N\right\}$ denote PerLA's input point cloud, where $\pmb{p_i}$ is a point coordinate, $\pmb{f_{i}}$ is a $d_0$ -dimensional feature vector (e.g., color or normal vector) corresponding to $\pmb {p}_i$ , and $N$ is the number of points (cardinality). Let $\mathcal{I}^t$ and $\mathcal{I}^v$ denote the input text prompt and visual prompt, respectively. Based on these inputs, PerLA (denoted as $\Phi$ ) generates a free-form natural language response $\mathcal{O}^t$ for various 3D-related tasks. + +# 3.1. Perceptive scene encoder + +To compute point-level representations through the scene encoder, $\mathcal{P}$ is typically downsampled into super-points using Farthest Point Sampling (FPS) [8, 9]. However, down + +sampling and using only $\mathcal{P}$ may hinder the encoder to capture fine-grained details, thus affecting performance in downstream 3D scene understanding tasks. Our proposed perceptive scene encoder is designed to preserve such scene details without increasing the number of tokens or the representation dimensions of the 3D scene. We propose to split $\mathcal{P}$ into parts, and employ a pre-trained 3D scene encoder to encode these parts and the whole point cloud separately. We then aggregate the output representations from these different pieces into a single (highly-informative) representation using a cross-attention-based module and Graph Convolutional Network (GCN). + +Hilbert-based scene partitioning. We partition $\mathcal{P}$ into $L$ equally-sized parts, each containing the same number of points, $\left\lfloor \frac{N}{L} \right\rfloor$ , where $\lfloor \cdot \rfloor$ denotes the greatest integer less than or equal to its argument. Enforcing equal cardinality across parts allows for spatially smaller parts in highly structured regions (areas with more semantic information) and spatially larger parts in less structured regions (areas with less semantic information). To achieve this, we adopt a Hilbert curve approach [50] to efficiently serialize the point cloud and partition it into parts [33, 57, 60]. + +Global and partial scene encoding. Both full-scene point cloud and the partial-scene point clouds are encoded separately by the same pre-trained 3D scene encoder $\phi$ . The 3D scene encoder $\phi$ downsamples the input point cloud into super-points using Farthest Point Sampling (FPS) [41] and produces a representation for each super-point. The representation of the full scene, i.e., the global representation, encode the overall scene context, while the representations of the partial scenes, i.e., the local representations, encode scene details. Let $\mathcal{P}^g = \{\pmb{p}_i^g \in \mathbb{R}^3\}_{i=1}^M$ denote the downsampled full-scene point cloud composed of $M$ points, and $\mathcal{F}^g = \{\pmb{f}_i^g \in \mathbb{R}^d\}_{i=1}^M$ denote the associated global representations. Let $\mathcal{P}^l = \{\pmb{p}_i^l \in \mathbb{R}^3\}_{i=1}^{L \cdot M}$ denote the set of downsampled partial-scene point clouds, and $\mathcal{F}^l = \{\pmb{f}_i^l \in \mathbb{R}^d\}_{i=1}^{L \cdot M}$ the associated local representations. Note that, we aggregate all the points of the downsampled partial-scene point clouds into the single set $\mathcal{P}^l$ , and both the full-scene and the partial-scenes are downsampled to the same number of super-points $M$ . The representations $\mathcal{F}^l$ provide a more detailed view of the scene, as they are derived from the same number of points downsampled from smaller, localized regions, resulting in higher resolution compared to the global versions. Yet, their semantic visibility is focused within each local part. In the following steps, we aggregate local and global information, to enrich the representations with details within the scene context. + +Hilbert-based nearest-neighbor search. In order to enhance global representations through local representations, we need to first find the correspondences among them, i.e., which super-points of partial-scene point clouds are spatially neighbors with super-points of the full-scene point + +![](images/527958e90390560bc5710e6da3bc2c83bd2e7ea02ca04705e8d33e86669ce5c5.jpg) +Figure 2. Overview of PerLA. (Left): The overall pipeline of PerLA, which begins by extracting interaction-aware 3D scene representations. These representations are then projected onto the prefix of textual instructions via MMA, serving as input to a frozen language model (LLM). (Right): The detailed design of PerLA. First, the 3D scene is divided into spatially compact regions using Hilbert-based scene serialization [50]. Next, an efficient $k$ -NN algorithm associates each point-level global representation with its detail-enriched local representations, creating a comprehensive scene representation through a Graph Convolutional Network (GCN). Finally, smoothness and regularization losses are applied to promote stable learning for the proposed perceptive scene encoder. + +cloud. This can be done via $k$ -NN search. However, traditional $k$ -NN searches can be computationally intensive on large-scale or high-resolution point clouds and cannot ensure that the identified neighbors maintain geometric consistency. To address these limitations, we approximate $k$ -NN with an efficient neighbor mapping technique using Hilbert serialized point clouds, which improves query speed while reducing computational complexity. With serialized point clouds, we can leverage point indices to perform $k$ -NN searches with $O(1)$ complexity [60]. + +Specifically, we first serialize the union $\mathcal{P}^g \bigcup \mathcal{P}^l$ using the Hilbert curve ordering. We then apply geometric partitioning [27] on the original point clouds to generate geometric labels $\mathcal{Y}^g \bigcup \mathcal{Y}^l$ . The center points of these labels are incorporated as high-order bits (by bit shifting) in the serialized index, ensuring that points with different labels occupy non-overlapping index ranges. Next, for each $\pmb{p}_i^g \in \mathcal{P}^g$ , we identify the $\mathtt{k}$ nearest super-points within $\mathcal{P}^l$ based on the serialized indexes. These geometric labels guarantee that $\pmb{p}_i^g$ and its nearest local super-points all originate from the same instance. Let $\mathcal{P}_{\mathrm{k}_i}^l$ denote the set of $\mathtt{k}$ nearest local super-points to the global super-point $\pmb{p}_i^g$ . Consequently, $\mathcal{F}_{\mathrm{k}_i}^l$ are the representations of the super-points in $\mathcal{P}_{\mathrm{k}_i}^l$ . + +Local-global representation aggregation. We make global representations more informative by combining them with information from local representations. To enable this, we present a novel two-step aggregation technique. In the first step, we employ a novel cross-attention algorithm between local and global representations using neighborhood information to update global representations, which we term as localized cross-attention. In the second step, we refine global representations through a Graph Convolutional Network based formulation based on message passing, which we term as GCN message passing (Fig. 2). + +Localized cross-attention operates with a learning-based weighting mechanism that is a function of the representations and relative positions of global super-points and their associated local neighbors. We constrain cross-attention to local neighborhood regions because $i$ ) limiting the number of points reduces computational complexity, and $ii$ ) neighbors of a given point are likely to belong to the same object. Specifically, for each point $\pmb{p}_i^g$ and its nearest neighbors $\mathcal{P}_{\mathrm{k}_i}^l$ , their relative position embeddings $\mathcal{R}_i$ are extracted as + +$$ +\mathcal {R} _ {i} = \left\{\mathcal {R} _ {i j} = \operatorname {p o s} \left(\left(\boldsymbol {p} _ {i} ^ {g} - \boldsymbol {p} _ {j} ^ {l}\right) / \sigma\right) \in \mathbb {R} ^ {d} \right\} _ {j = 1} ^ {K ^ {l}}, \tag {1} +$$ + +where $\sigma > 0$ is a learnable parameter that controls the relative position scaling, and $\mathrm{pos}(\cdot)$ is the 3D Fourier positional embedding [52] operation, which satisfies: + +$$ +\left. \operatorname {p o s} (\boldsymbol {x}) = \left[ \sin (2 \pi \boldsymbol {x} \cdot B); \cos (2 \pi \boldsymbol {x} \cdot B) \right], \quad \right. \tag {2} +$$ + +where $B \in \mathbb{R}^{3 \times (d/2)}$ is a learnable matrix. We update the global representations through cross-attention as follows + +$$ +\boldsymbol {s} _ {i j} = \left(W _ {q} \boldsymbol {f} _ {i} ^ {g}\right) ^ {\top} \left(W _ {k} \left(\boldsymbol {f} _ {j} ^ {l} + W _ {r} \mathcal {R} _ {i j}\right)\right) / \sqrt {d}, +$$ + +$$ +\boldsymbol {s} _ {i} = \left\{\boldsymbol {s} _ {i j} \right\} _ {i = 1} ^ {K ^ {l}}, \boldsymbol {w} _ {i} = \text {s o f t m a x} \left(\boldsymbol {s} _ {i}\right), \tag {3} +$$ + +$$ +\hat {\boldsymbol {f}} _ {i} ^ {g} = \boldsymbol {f} _ {i} ^ {g} + \boldsymbol {w} _ {i} \left(W _ {v} \left(\mathcal {F} _ {\mathcal {K} _ {i}} ^ {l} + \mathcal {R} _ {i}\right)\right), +$$ + +where $\hat{\pmb{f}}_i^g$ are the updated representations, and $W_{q},W_{k},W_{v},W_{r}$ are projection parameter matrices. Let $\hat{\mathcal{F}}^g = \{\hat{\pmb{f}}_i^g\}_{i = 1}^M$ denote the set of $\hat{\pmb{f}}_i^g$ + +GCN message passing further refines $\hat{\mathcal{F}}^g$ by aggregating information from neighboring global super-points. Let $\mathcal{G}^g = \{\mathcal{P}^g, \mathcal{W}^g\}$ be a $k$ -NN graph, where $\mathcal{P}^g$ is the set of the vertices, i.e., global super-points, and $\mathcal{W}^g \in \mathbb{R}^{N \times N}$ is the adjacency matrix defining the edges. The construction of the adjacency matrix uses the similarity between superpoint representations $\mathcal{F}^G$ . Note that we use the (original) global representations to define the adjacency matrix because we empirically experienced more stability in training as opposed to using the updated global representation, + +likely due to the fact that they change value during training. The adjacency matrix $\mathcal{W}^g$ with elements $\mathcal{W}_{ij}^{g}$ is defined as + +$$ +\mathcal {W} _ {i j} ^ {g} = \left\{ \begin{array}{l l} \frac {\boldsymbol {f} _ {i} ^ {g \top} \boldsymbol {f} _ {j} ^ {g}}{\| \boldsymbol {f} _ {i} ^ {g} \| \| \boldsymbol {f} _ {j} ^ {g} \|} \cdot \left(\boldsymbol {f} _ {i} ^ {g \top} \boldsymbol {f} _ {j} ^ {g} > 0\right), & \text {i f p _ {j} ^ {g} \in k - N N (p _ {i} ^ {g})}, \\ 0, & \text {o t h e r w i s e}. \end{array} \right. +$$ + +Lastly, we perform one-time GCN message passing (MP) on the graph $\mathcal{G}^g$ to obtain an updated feature matrix $\hat{\mathcal{F}}^g$ . + +Let $\tilde{\mathcal{W}}^g$ be the symmetrically normalized adjacency matrix of $\mathcal{W}^g$ by $\tilde{\mathcal{W}}^g = \mathbf{D}^{-\frac{1}{2}}\mathcal{W}^g\mathbf{D}^{-\frac{1}{2}}\in \mathbb{R}^{M\times M}$ where $\mathbf{D}$ is the diagonal degree matrix of $\mathcal{W}^g$ . + +$$ +\hat {\mathcal {F}} ^ {g} \leftarrow \operatorname {R e L U} \left(W _ {m} \left(\tilde {\mathcal {W}} ^ {g} \hat {\mathcal {F}} ^ {g}\right) + W _ {s} \hat {\mathcal {F}} ^ {g}\right), \tag {4} +$$ + +where $W_{m}$ and $W_{s}$ are trainable parameters. This MP layer structure allows each node to aggregate information from its neighbors while also preserving its own feature representation (via the skip connection). + +# 3.2. Learning with local representation consensus + +The main learning objective encourages the model to generate outputs that are close to the target responses in the training dataset. This objective is typically expressed as a per-token cross-entropy loss $\mathcal{L}_{pred}$ [9, 21] (see details in our Supp. Mat.). We found that adding a consensus term to $\mathcal{L}_{pred}$ that encourages local regularization of representations improves training stability. Firstly, points belonging to the same object (i.e., local neighborhoods) should share similar aggregated representations. Secondly, aggregated super-point representations should remain close to their original point representations to preserve context knowledge. Our novel consensus loss $\mathcal{L}_{\mathrm{con}}$ is defined as + +$$ +\mathcal {L} _ {\text {c o n}} = \mathcal {L} _ {\text {s m t}} + \mu \mathcal {L} _ {\text {r e g}}, +$$ + +$$ +\mathcal {L} _ {\mathrm {s m t}} = \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N} \mathcal {W} _ {i j} ^ {g} \left\| \frac {\hat {\boldsymbol {f}} _ {i} ^ {g}}{\sqrt {d _ {i}}} - \frac {\hat {\boldsymbol {f}} _ {j} ^ {g}}{\sqrt {d _ {j}}} \right\|, \tag {5} +$$ + +$$ +\mathcal {L} _ {\mathrm {r e g}} = \sum_ {i = 1} ^ {N} \left\| \hat {\pmb {f}} _ {i} ^ {g} - \pmb {f} _ {i} ^ {g} \right\|, +$$ + +where $d_{k} = \sum_{l}\mathcal{W}_{kl}^{g}$ is the diagonal matrix representing row-wise sums of $\mathcal{W}^g$ . The first term $\mathcal{L}_{\mathrm{smt}}$ enforces spatial connectivity by promoting similarity among fused representations of points within the same object, addressing the first attribute. $\mathcal{L}_{\mathrm{reg}}$ is a regularization term that promotes the learned representations stay close to the original global representations. $\mu$ is a hyperparameter. The overall loss $\mathcal{L}$ , is a weighted sum of $\mathcal{L}_{pred}$ and $\mathcal{L}_{con}$ + +$$ +\mathcal {L} = \lambda \mathcal {L} _ {\text {c o n}} + \mathcal {L} _ {\text {p r e d}}, \tag {6} +$$ + +where $\lambda$ is the balance hyperparameter. By jointly training with these two loss terms, we encourage the model to learn more object-aware representations and enhance its ability to extract finer details. This approach also preserves global context information, which is essential for capturing spatial relationships between different objects. + +During inference, we use beam search to predict response $\mathcal{O}^t$ that maximizes the following objective: + +$$ +\mathcal {O} ^ {t} = \arg \max _ {\mathcal {O}} \Phi (\mathcal {O} \mid \mathcal {P}, \mathcal {I} ^ {t}, \mathcal {I} ^ {v}). \tag {7} +$$ + +where we set a beam size of 4. + +# 4. Experiments + +We evaluate PerLA on 3D Question Answering and 3D Dense Captioning downstream tasks, and compare its performance with state-of-the-art methods from the literature. + +Datasets. We conduct experiments using the ScanNet dataset [14], which encompasses 1,201 training and 312 validation scenes, featuring diverse and complex indoor 3D environments. Language annotations are from ScanQA [2], ScanRefer [6], Nr3D [1], and the ScanNet subset of 3D-LLM [21], collectively supporting a range of tasks including instance and scene descriptions, conversations, embodied planning, and question answering. For additional data statistics, please refer to the supplementary materials. + +Metrics. We follow LL3DA's evaluation protocol [9] to evaluate the quality of output responses. We use the abbreviations C, B4, M and R for CiDEr [55], BLEU-4 [43], METEOR [3], and Rouge-L [34], respectively. + +Implementation details. As in [9, 12], we input 40,000 randomly sampled points from each 3D scene. We divide the point cloud into partitions $L = 6$ and choose $k = 4L = 24$ neighbors to enhance global representations. We set $\lambda = \mu = 0.1$ in the loss. We use the pre-trained OPT-1.3B [66] language model, kept frozen and loaded in float16 precision. We use the AdamW [38] optimizer with a weight decay of 0.1, applying a cosine annealing scheduler that decays the learning rate from $10^{-4}$ to $10^{-6}$ over approximately 100,000 iterations. All tasks are trained with a total batch size of 16. Each training process is completed within two days using up to two NVIDIA H100 P0 (96GB) GPUs. For each evaluation, we fine-tune the model's parameters on the respective task for about 30,000 iterations. + +# 4.1. Results + +3D question answering is a task that involves answering questions about a 3D scene. It allows a model to provide information about the objects, relationships, and attributes within a 3D environment based on a given question. Tab. 1 shows the results of ScanQA's validation and test sets. Classification-based methods (CLS) select responses from a predefined answer set. Generation-based approaches (GEN) generate the entire textual response. PerLA consistently outperforms existing approaches across all evaluation sets and metrics, in particular with $+1.34$ CiDEr score over LL3DA. We also compare against our reproduced version of LL3DA, where PerLA scores $+3.76$ CiDEr. + +Table 1. Comparative results for 3D Question Answering on ScanQA [2] benchmark. CLS and GEN denote classification-based and generation-based methods, respectively. LL3DA (repr.) is results of LL3DA we reproduced. PerLA outperforms all the other methods. + +
MethodTypeValidationTest w/ objectTest w/o object
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
ScanQA[2]CLS64.8610.0813.1433.3367.2912.0413.5534.3460.2410.7512.5931.09
Clip-Guided[44]-----69.5314.6413.9435.1562.8311.7313.2832.41
Multi-CLIP[16]CLS----68.7012.6513.9735.4663.2012.8713.3632.61
3D-VLP[26]CLS66.9711.1513.5334.5170.1811.2314.1635.9763.4015.8413.1331.79
3D-VisTA[70]-----68.6010.5013.8035.5055.708.7011.6929.60
3D-LLM [21]GEN69.4012.0014.5035.7069.6011.6014.9035.30----
LL3DA[9]GEN76.7913.5315.8837.3178.1613.9716.3838.1570.2912.1914.8535.17
LL3DA (repr.)GEN74.3713.5015.0936.31--------
PerLAGEN78.1314.4917.4439.6080.9117.2116.4940.7174.8214.9715.2338.18
Δ w.r.t. LL3DA[9]-+1.34+0.96+1.56+2.29+2.75+3.24+0.11+2.56+4.53+2.78+0.38+3.01
+ +![](images/e7dfa24369b410d084bdfa2c2d2c3307b13db995d359a8216dd4518dcf3c2a5d.jpg) +Figure 3. The qualitative comparison between our method, PerLA, and LL3DA [9] on the ScanQA [2] dataset shows that our approach achieves higher accuracy in responding to "what"-related questions. + +![](images/230abff78840d6d3ab5edf3cf836a2879436956dada63378a847bf8077ebed90.jpg) + +Fig. 3 provides qualitative comparisons between LL3DA and PerLA on the ScanQA benchmark [2]. It highlights PerLA's accuracy in answering questions regarding object attributes and spatial relationships within a 3D scene. PerLA provides precise answers, correctly identifying objects such as a "picture", "trash bin", and "TV" along with their specific colors and types, while LL3DA often yields incorrect or less specific responses. In particular, PerLA accurately identifies the colors of objects within the scene, such as the door, the monitor, and the instrument case, correctly answering the questions about these attributes. In contrast, LL3DA frequently misidentifies colors, labeling objects as "White" instead of "Black," for example. This illustrates PerLA's superior ability to capture fine-grained details and contextual information, delivering more accurate answers in complex environments. + +3D dense captioning involves localizing and describing each 3D instance within complex 3D environments. Tab. 2 shows the results on ScanRefer [6] and $\mathrm{Nr3D}$ [1] benchmarks. Following previous works [9, 21], we use the m@kIoU metric, where $m\in \{\mathbf{C},\mathbf{B4},\mathbf{M},\mathbf{R}\}$ and $k$ denote the IoU threshold. As in [12, 25], we report C@0.25 and C@0.5 for ScanRefer, and C@0.5 for $\mathrm{Nr3D}$ . UniT3D [13], 3DJCG [4], and 3D-VLP [26] are pre-trained on multiple 3D vision and language tasks from ScanNet scenes. UniT3D uses image caption models and multi-view im + +ages to generate extra instance captions for pre-training. For fair comparison, we report results from models trained using standard per-word cross-entropy loss without additional 3D scenes. Box annotations are estimated with Vote2CapDETR and used as visual prompts. PerLA consistently outperforms comparison methods on both benchmarks. PerLA significantly outperforms LL3DA by scoring +3.75 and +4.22 CiDEr on ScanRefer, and +3.88 CiDEr on Nr3D. These results highlight the effectiveness of PerLA in 3D dense captioning tasks. + +Fig. 4 provides examples of qualitative results of LL3DA and PerLA on Nr3D and ScanRefer. PerLA produces more accurate and detailed descriptions, effectively capturing spatial relationships and fine-grained attributes, such as the positioning of objects relative to surrounding elements. + +# 4.2. Ablation study + +Local, global, and GCN representations. To investigate the impact of local and global representations, we design two variants of PerLA: one using only local representations and the other using only global representations. We also report an extended version of LL3DA, termed $\mathrm{LL3DA}^{\dagger}$ , in which we increase the number of query tokens used to encode local partitions. Representations are first extracted from each partition, and FPS is applied to sample 1,024 points with their corresponding representations from the + +Table 2. Comparative results for 3D Dense Captioning on ScanRefer [6] and Nr3D [1] benchmarks. LL3DA (repr.) is results of LL3DA we reproduced. Generally, PerLA outperforms all other methods on both benchmarks. + +
MethodScanRefer@0.25ScanRefer@0.5Nr3D@0.5
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
Scan2Cap[12]56.8234.1826.2955.2739.0823.3221.9744.7827.4717.2421.8049.06
MORE[25]62.9136.2526.7556.3340.9422.9321.6644.42----
SpaCap3D[56]-----44.0225.2622.3333.7119.9222.6150.50
REMAN[40]62.0136.3726.7656.2545.0026.3123.1346.9634.8120.3722.7150.90
D3Net[7]-----51.67--35.2620.4222.7753.38
Contextual[68]-----46.0723.4023.95----
UniT3D[13]----46.6927.5221.9145.98----
3DJCG[4]64.7040.1727.6359.2349.4831.6324.3650.8038.0622.8223.7752.99
3D-VLP[26]70.7341.0328.1459.7254.9432.3124.8351.51----
3D-VisTA*[70]----61.6034.1026.8055.00----
Vote2CapDETR[8]71.4539.3428.2559.6361.8134.4626.2254.4043.8426.6825.4154.43
LL3DA[9]74.1741.4127.7659.5365.1936.7925.9755.0651.1828.7525.9156.61
LL3DA (repr.)71.8639.5727.2958.3763.7935.6725.9454.5648.3828.3625.7255.66
PerLA77.9243.4128.9759.6969.4138.0229.0756.8055.0631.2428.5259.13
Δ w.r.t. LL3DA[9]+3.75+2.00+1.21+0.16+4.22+1.23+2.27+1.74+3.88+2.49+2.61+2.52
Describe this [bbox] object.
PerLA The toilet paper is on the wall next to the toilet.
LL3DA The towel closest to the toilet.
+ +(a) 3D visual dense captioning on $\mathrm{Nr3D}$ [1] + +
Describe this [bbox] object.Describe this [bbox] object.
PerLA +There is a rectangular bathroom vanity. It is to the right of the toilet.PerLA +This is a white sink. +It is to the left of the toilet.
LL3DA +This is a bathroom sink. The sink is on the left side of the room.LL3DA +This is a white sink. +It is on the left side of the sink.
(b) 3D visual dense captioning on ScanRefer [6]
+ +
Describe this [bbox] object.Describe this [bbox] object.
PerLA +This is a black computer tower. It is under a desk. +LL3DA +The computer is on the left side of the desk. The computer is on the right side.PerLA +This is a black computer tower. It is under a desk. +LL3DA +There is a black desk. It is on the left side of the room.
+ +Figure 4. Qualitative comparisons on the dense captioning task across the Nr3D [1] and ScanRefer [6]. We compare the results of our PerLA with LL3DA [9]. PerLA generates accurate descriptions, effectively capturing fine-grained object attributes and spatial relationships. + +union of these partitions. These representations are then passed through a multimodal adapter to produce 32 global tokens. LL3DA generates an additional 32 global tokens from the point cloud of the entire scene, resulting in a total of 64 tokens, which are concatenated and fed into the LLM for response generation. (For details of $\mathrm{LL3DA}^{\dagger}$ please refer to the Supp. Mat.). Tab. 3 shows that the combination of local and global representations yields a much better + +performance than the two independently used. Tab. 3 also shows that $\mathrm{LL3DA}^{\dagger}$ does not reach PerLA performance, underlying the importance of information exchange between local and global representations. To evaluate the effectiveness of GCN, we also report the results without using GCN (w/o GCN). Compared to our PerLA, which incorporates GCN, it consistently improves performance across the three datasets, confirming its effectiveness. + +Table 3. Ablation study of our local-to-global, GCN representation aggregation algorithms on ScanQA, ScanRefer[6] and Nr3D[1] benchmarks. LL3DA† denotes an extended version of LL3DA with an increased number of query tokens. + +
MethodScanQAScanRefer@0.5Nr3D@0.5
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
\( LL3DA^{\dagger} \)74.5412.8915.1136.9662.2534.5025.5553.8448.7028.2425.7255.59
Global74.4913.5015.1636.5563.7935.6725.9454.5648.3828.3625.7255.66
Local73.5512.9814.9536.0762.7934.7825.9454.1649.0928.2025.8055.89
w/o GCN77.6113.8715.9339.2869.0437.4328.7155.5954.3029.7827.4158.32
PerLA78.1314.4917.4439.6069.4138.0229.0756.8055.0631.2428.5259.13
+ +Table 4. Ablation study on the impact of constrained attention on ScanQA validation dataset. + +
cross-attentionmean poolingmax pooling
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
78.114.517.439.674.412.515.236.574.013.014.835.7
+ +Table 5. Ablation study on the impact of joint learning by different loss combination on ScanQA validation dataset. + +
LossScanQA (Validation)
LpreLsmtLregC↑B4↑M↑R↑
✓75.3113.8315.9037.48
✓✓76.6213.9616.9938.14
✓✓76.4714.0716.7238.79
✓✓✓78.1314.4917.4439.60
+ +Localized cross-attention. To evaluate the effectiveness of our novel localized cross-attention module, we tested max pooling and mean pooling as alternative methods. Tab. 4 shows that localized cross-attention significantly outperforms both mean pooling and max pooling across all metrics. These results indicate that localized cross-attention enhances ability of model to generate more informative scene representations, thereby improving performance on 3D question answering tasks. This improvement comes from the incorporation of positional information and semantic similarity in cross-attention, which helps to exclude neighboring points that do not belong to the same object. + +Loss function. To evaluate the effectiveness of our novel loss term $\mathcal{L}_{\mathrm{con}}$ , we train PerLA using different combinations of $\mathcal{L}_{\mathrm{con}}$ ’s components (Eq. 5). Since our task is intended for downstream applications, each combination also includes the task-specific loss $\mathcal{L}_{\mathrm{pre}}$ . Tab. 5 shows that adding $\mathcal{L}_{\mathrm{smt}}$ to $\mathcal{L}_{\mathrm{pre}}$ improves all metrics (C, B4, M, and R). Incorporating the loss of regularization $\mathcal{L}_{\mathrm{reg}}$ further enhances performance, particularly in metrics C, M, and R. These results show the effectiveness of learning jointly with both semantic awareness and regularization losses. + +Number of partitions. To evaluate the effectiveness of different number of partitions, we test PerLA by splitting the point cloud into 4, 6, and 8 partitions. Tab. 6 shows that increasing the number of partitions generally improves performance, with 6 partitions yielding the best results. Al- + +Table 6. Ablation study on the impact of number of parts on ScanQA validation dataset. + +
468
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
76.913.616.137.378.114.517.439.678.014.517.239.4
+ +though 8 partitions outperform 4, they show slightly lower performance than 6. This suggests that 6 partitions provide a better balance between granularity and effectiveness, capturing relevant scene details with lower computational complexity. More ablation studies, including an analysis of performance in different training strategies, refer to Supp. Mat.. + +# 5. Conclusions + +We presented PerLA, a perceptive 3D language assistant capable of capturing both detailed and contextual information to enhance visual representations for LLMs. PerLA features a dual-branch architecture: the global branch processes superpoints from the whole point cloud via downsampling, while the local branch focuses on partitioned regions. We demonstrated that by integrating representations from both branches, PerLA effectively captures scene details, reducing hallucinations. Moreover, we employ a Graph Convolutional Network to facilitate information exchange among neighboring local and global superpoints, and introduce a novel loss term for local representation consensus to promote training stability. Experiments on the ScanQA, ScanRefer, and Nr3D benchmarks highlight the effectiveness of our approach, setting a new state-of-the-art performance in 3D question answering and dense captioning. + +Limitations & Future Work. PerLA primarily focuses on enhancing performance using point cloud input, but integrating it with optimized modules, such as token merging, presents a promising direction to extend its capabilities. Although PerLA shows strong performance on standard benchmarks, future work could explore its robustness and generalizability in more complex 3D scenarios, broadening its applicability in diverse real-world settings. + +Acknowledgement. This work was supported by PNRR FAIR - Future AI Research (PE00000013) and ICSC National Research Centre for HPC, Big Data and Quantum Computing (CN00000013), funded by NextGeneration EU. + +# References + +[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 5, 6, 7, 8, 14, 15, 16, 17, 18 +[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, 2022. 2, 5, 6, 14, 15, 16, 17, 18 +[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 5, 16 +[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2, 6, 7 +[5] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, pages 357–366, 2021. 2 +[6] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 2, 5, 6, 7, 8, 14, 15, 16, 17, 18 +[7] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In ECCV, pages 487-505. Springer, 2022. 7 +[8] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In CVPR, pages 11124-11133, 2023. 1, 2, 3, 7, 12 +[9] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In CVPR, 2024. 1, 2, 3, 5, 6, 7, 12, 14, 16, 17, 18 +[10] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, YU Gang, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE TPAMI, 2024. 2 +[11] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 1 +[12] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, pages 3193-3203, 2021. 2, 5, 6, 7 + +[13] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2, 6, 7 +[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5, 14, 15 +[15] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. NeurIPS, 36, 2024. 2 +[16] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329, 2023. 6 +[17] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 3 +[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 2 +[19] Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Henan Xiong. Scene-llm: Extending language model for 3d visual understanding and reasoning. arXiv preprint arXiv:2403.11401, 2024. 3 +[20] Ritwik Gupta, Shufan Li, Tyler Zhu, Jitendra Malik, Trevor Darrell, and Karttikeya Mangalam. xt: Nested tokenization for larger context in large images. arXiv preprint arXiv:2403.01915, 2024. 2 +[21] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. NeurIPS, 2023. 1, 2, 3, 5, 6, 14, 16, 17 +[22] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. NeurIPS, 2024. 2 +[23] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In ICML, 2024. 2 + +[24] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3d: Learning fine-grained class-agnostic 3d segmentation without manual labels. In ECCV, pages 278–295. Springer, 2025. 3 +[25] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In ECCV, pages 528-545. Springer, 2022. 6, 7 +[26] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In CVPR, pages 10984-10994, 2023. 2, 6, 7 +[27] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, pages 4558-4567, 2018. 4, 14 +[28] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2024. 15 +[29] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In CVPR, pages 7287-7296, 2022. 2 +[30] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3 +[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 3, 12, 13 +[32] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2 +[33] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaoqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In NeurIPS, 2024. 3 +[34] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5, 16 +[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llavanext: Improved reasoning,OCR, and world knowledge, 2024.2,3 +[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and + +Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 1, 2 +[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 2 +[38] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017. 5 +[39] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2 +[40] Aihua Mao, Zhi Yang, Wanxin Chen, Ran Yi, and Yong-jin Liu. Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE TVCG, 2023. 7 +[41] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, ECCV Laboratory, 2003. 3 +[42] Bongki Moon, Hosagrahar V Jagadish, Christos Faloutsos, and Joel H. Saltz. Analysis of the clustering properties of the hilbert space-filling curve. IEEE TKDE, 13(1):124-141, 2001. 13 +[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 5, 16 +[44] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In CVPR, pages 5607-5612, 2023. 2, 6 +[45] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 12 +[46] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. Shapellm: Universal 3d object understanding for embodied interaction. In ECCV, pages 214-238, 2025. 2 +[47] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 1 +[48] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2 +[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy + +Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In ICML, pages 29441-29454. PMLR, 2023. 2 +[50] Hans Sagan. Hilbert's space-filling curve. In Space-filling curves, pages 9-30. Springer, 1994. 2, 3, 4 +[51] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In ECCV, pages 444-462. Springer, 2025. 2 +[52] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 33: 7537-7547, 2020. 4 +[53] Yuan Tang, Xu Han, Xianzhi Li, Qiao Yu, Yixue Hao, Long Hu, and Min Chen. Minigpt-3d: Efficiently aligning 3d point clouds with large language models using 2d priors. ACM'MM, 2024. 2 +[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.1, 17 +[55] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 5, 16 +[56] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022.2, 7 +[57] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. ACM TOG, 42(4):1-11, 2023. 3 +[58] Qing Wang et al. Divergence-augmented policy optimization. NeurIPS, 32, 2019. 15 +[59] Zehan Wang, Haifeng Huang, Yang Zhao, Zhang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 2 +[60] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 1, 3, 4 +[61] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 2 +[62] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm + +perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 3 +[63] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. NeurIPS, 34:30008-30022, 2021. 2 +[64] Jianing Yang, Xuweiyi Chen, Nikhil Madaan, Madhavan Iyengar, Shengyi Qian, David F. Fouhey, and Joyce Chai. 3d-grand: A million-scale dataset for 3d-llms with better grounding and less hallucination. arXiv preprint 2406.05132, 2024. 2 +[65] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. arXiv preprint arXiv:2112.08359, 2021. 2 +[66] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 5, 17 +[67] Lichen Zhao, Daigang Cai, Jing Zhang, Lu Sheng, Dong Xu, Rui Zheng, Yinjie Zhao, Lipeng Wang, and Xibo Fan. Toward explainable 3d grounded visual question answering: A new benchmark and strong baseline. IEEE TCSVT, 33(6):2935-2949, 2022. 2 +[68] Yufeng Zhong, Long Xu, Jiebo Luo, and Lin Ma. Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925, 2022. 7 +[69] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 2 +[70] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 6, 7 \ No newline at end of file diff --git a/2025/PerLA_ Perceptive 3D Language Assistant/images.zip b/2025/PerLA_ Perceptive 3D Language Assistant/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e4d07311260dde07c6c2f1df7aea7a159cb3a4fd --- /dev/null +++ b/2025/PerLA_ Perceptive 3D Language Assistant/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b51a9c9965e81c98471578928f852f6a4a25f4fd6b1558a012c06c697ed7cf46 +size 785808 diff --git a/2025/PerLA_ Perceptive 3D Language Assistant/layout.json b/2025/PerLA_ Perceptive 3D Language Assistant/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..60378ec663d2af62d0886259460860e583cdee1b --- /dev/null +++ b/2025/PerLA_ Perceptive 3D Language Assistant/layout.json @@ -0,0 +1,9207 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 101, + 445, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 101, + 445, + 125 + ], + "spans": [ + { + "bbox": [ + 165, + 101, + 445, + 125 + ], + "type": "text", + "content": "PerLA: Perceptive 3D language assistant" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "spans": [ + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "text", + "content": "Guofeng Mei" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "text", + "content": " Wei Lin" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "text", + "content": " Luigi Riz" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "text", + "content": " Yujiao Wu" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "text", + "content": " Fabio Poiesi" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "text", + "content": " Yiming Wang" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "text", + "content": "Fondazione Bruno Kessler, Italy " + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "text", + "content": "JKU Linz, Austria " + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 89, + 151, + 520, + 179 + ], + "type": "text", + "content": "CSIRO, Australia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 272, + 182, + 334, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 182, + 334, + 194 + ], + "spans": [ + { + "bbox": [ + 272, + 182, + 334, + 194 + ], + "type": "text", + "content": "gmei@fbk.eu" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 222, + 294, + 373 + ], + "blocks": [ + { + "bbox": [ + 59, + 222, + 294, + 373 + ], + "lines": [ + { + "bbox": [ + 59, + 222, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 59, + 222, + 294, + 373 + ], + "type": "image", + "image_path": "f170598bfc50c4a30adbdff664e93b3edefe38b17009eb83b905c054e915b4da.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 222, + 552, + 375 + ], + "blocks": [ + { + "bbox": [ + 315, + 222, + 552, + 375 + ], + "lines": [ + { + "bbox": [ + 315, + 222, + 552, + 375 + ], + "spans": [ + { + "bbox": [ + 315, + 222, + 552, + 375 + ], + "type": "image", + "image_path": "ac232aac1bab0d051b6c7f3e2a340086609463c42d91248458bb2b4a53e15546.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 376, + 555, + 421 + ], + "lines": [ + { + "bbox": [ + 54, + 376, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 555, + 421 + ], + "type": "text", + "content": "Figure 1. PerLA is a 3D language assistant that integrates local details with global context to learn informative representations of 3D scenes, whereas state-of-the-art (SOTA) 3DLAs focus solely on global context information. PerLA can provide more accurate responses, correctly distinguishing between objects such as a \"black computer monitor\" and a \"black suitcase,\" where SOTA models instead fail with hallucinated responses. Examples in figures show cases where capturing details from the point cloud matters for accurate output captions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 430, + 200, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 430, + 200, + 441 + ], + "spans": [ + { + "bbox": [ + 152, + 430, + 200, + 441 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 456, + 297, + 720 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 456, + 297, + 720 + ], + "spans": [ + { + "bbox": [ + 54, + 456, + 297, + 720 + ], + "type": "text", + "content": "Enabling Large Language Models (LLMs) to understand the 3D physical world is an emerging yet challenging research direction. Current strategies for processing point clouds typically downsample the scene or divide it into smaller parts for separate analysis. However, both approaches risk losing key local details or global contextual information. In this paper, we introduce PerLA, a 3D language assistant designed to be more perceptive to both details and context, making visual representations more informative for the LLM. PerLA captures high-resolution (local) details in parallel from different point cloud areas and integrates them with (global) context obtained from a lower-resolution whole point cloud. We present a novel algorithm that preserves point cloud locality through the Hilbert curve and effectively aggregates local-to-global information via cross-attention and a graph neural network. Lastly, we introduce a novel loss for local representation consensus to promote training stability. PerLA outperforms state-of-the-art 3D language assistants, with gains of up to " + }, + { + "bbox": [ + 54, + 456, + 297, + 720 + ], + "type": "inline_equation", + "content": "+1.34" + }, + { + "bbox": [ + 54, + 456, + 297, + 720 + ], + "type": "text", + "content": " CiDEr on ScanQA for question answering, and " + }, + { + "bbox": [ + 54, + 456, + 297, + 720 + ], + "type": "inline_equation", + "content": "+4.22" + }, + { + "bbox": [ + 54, + 456, + 297, + 720 + ], + "type": "text", + "content": " on ScanRefer and " + }, + { + "bbox": [ + 54, + 456, + 297, + 720 + ], + "type": "inline_equation", + "content": "+3.88" + }, + { + "bbox": [ + 54, + 456, + 297, + 720 + ], + "type": "text", + "content": " on Nr3D for dense captioning. Project page: https://gfmei.github.io/PerLA" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 430, + 394, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 430, + 394, + 441 + ], + "spans": [ + { + "bbox": [ + 315, + 430, + 394, + 441 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 689 + ], + "type": "text", + "content": "3D language assistants (3DLAs) jointly process natural language and 3D data to achieve 3D scene understanding, such as recognizing object categories, locations, appearances, and relationships, without requiring specialized models for each recognition task [9, 11, 21]. These capabilities are primarily powered by Large Language Models (LLMs) trained in large text corpora [54]. These approaches can aggregate multi-view features [21] or process point clouds [9] to generate 3D representations, which are in turn converted to tokens for the LLM [54, 66]. However, extracting multi-view representations is computationally costly and often fails to capture essential geometric properties [21]. Directly processing point clouds can yield more accurate results, yet it is even more computationally costly than handling multiview data, as point clouds typically have rather large cardinalities [60]. To address this, the point cloud cardinality can be reduced via downsampling [8, 9]. However, as with images [36], downsampling can compromise downstream task performance due to the reduced model's ability to perceive fine details of 3D scenes [47]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Fig. 1 illustrates two cases where our approach can accurately capture details and describe small objects within large" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "text", + "content": "14369" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "content": "scenes, while a state-of-the-art method hallucinates object details [9]. Although fine-grained attributes are critical for performance, directly extracting detailed information from high-resolution 3D data for 3DLAs remains underexplored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 297, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 297, + 585 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 297, + 585 + ], + "type": "text", + "content": "In this work, we aim to enhance 3DLAs' ability to perceive finer details in point clouds in order to execute downstream tasks more accurately. While increasing the number of visual tokens as the point cloud grows in size seems the straightforward solution, our empirical study (Tab. 3) shows that this solution has limited effectiveness in capturing scene details and it just increases computational burden. To address this, we propose PerLA, a novel 3DLA with a perceptive 3D scene encoder that captures detailed point cloud information, allowing the language model to generate more accurate responses without processing additional tokens. PerLA first divides the complete 3D scene into non-overlapping local parts to be processed in parallel, then integrates this local information with the global context obtained from a lower-resolution representation of the entire point cloud. Although dividing visual input has gained popularity in image processing [35], it has not yet been applied to point clouds. Processing point clouds presents unique challenges beyond those of images, as it requires handling an unordered set of points rather than rasterized pixels. To address these challenges, we serialize and partition the point cloud before encoding using a Hilbert curve approach [50], which efficiently preserves locality. We then combine local and global information through an efficient Hilbert curve-based " + }, + { + "bbox": [ + 56, + 128, + 297, + 585 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 128, + 297, + 585 + ], + "type": "text", + "content": "-NN search and aggregate this information via a novel cross-attention module and Graph Convolutional Network (GCN) to generate highly informative point-level representations for the LLM. Lastly, we train PerLA with a novel loss function designed to promote consensus on local representation, addressing the issue of divergent representations during local-to-global aggregation. We validate our approach on the question answering benchmark ScanQA [2], and the 3D dense captioning benchmarks ScanRefer [6] and Nr3D [1] to demonstrate the effectiveness of PerLA. PerLA demonstrates high transferability on both tasks of 3D question answering and 3D dense captioning, achieving state-of-the-art performance. In summary, our contributions are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 592, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 592, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 615 + ], + "type": "text", + "content": "- We present a novel perceptive 3D encoder to preserve local and global information for 3D language assistant." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 616, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 295, + 651 + ], + "type": "text", + "content": "- We introduce an efficient approach based on Hilbert curve " + }, + { + "bbox": [ + 56, + 616, + 295, + 651 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 616, + 295, + 651 + ], + "type": "text", + "content": "-NN search and cross attention to aggregate local and global information at point level." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 651, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 295, + 675 + ], + "type": "text", + "content": "- We propose a novel algorithm based on Graph Neural Network to refine and enhance aggregated information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 676, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 676, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 676, + 295, + 712 + ], + "type": "text", + "content": "- We introduce a novel loss objective to enable local representation consensus for local-to-global information aggregation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 315, + 71, + 399, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 71, + 399, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 399, + 84 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 95, + 555, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 95, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 95, + 555, + 287 + ], + "type": "text", + "content": "3D language understanding involves understanding 3D scenes by describing or answering to scene-relevant questions in the format of natural language. Examples of typical downstream tasks include 3D Dense Captioning [8, 12, 56], and 3D Question Answering [2, 39, 65, 67]. 3D Dense Captioning consists of 3D localization of object instances, and textual descriptions of each object instance [4, 8, 10, 12]. 3D Question Answering requires the model equipped with a language decoder to answer questions regarding the visual context in the given 3D scene [2, 44]. Several works focus on addressing the problems of 3D-language pre-alignment [13, 26], or designing adapter layers [9, 21], or constructing 3D synthetic data [64]. In contrast, our work focuses on designing an encoding approach that can enhance 3D-language understanding by capturing fine-grained details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 291, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 556, + 615 + ], + "type": "text", + "content": "3D language assistants (3DLAs) can be roughly categorized into two types: object-level 3DLAs [23, 46, 53, 61] and scene-level 3DLAs [9, 21, 59, 69]. Object-level 3DLAs learn from large 3D object datasets [15] to connect object-level 3D representation with language models. However, they underperform in compositional reasoning in complex 3D scenes with numerous objects. Scene-level 3DLAs, such as 3D-LLM [21], LLaVA-3D [69], Chat-3D [59], and LL3DA [9] enable scene understanding via the interaction with objects. Because 3D data is orders of magnitude less than 2D data, existing 3DLAs address such lack of data by leveraging pre-trained 2D Large Multimodal Models (LMMs) [21, 69], or data-efficient training recipes [9, 59]. 3D-LLM [21] leverages 2D pre-trained representations of rendered multi-view images to construct 3D representations and 2D VLMs as backbones. LLaVA-3D [69] adapts LLaVA [36] for 3D scene understanding by associating 2D patch representations with their positions in 3D space. Chat-Scene [22] improves the referencing and grounding capability and models scene representations as a sequence of object-level representations. LL3DA [9] extracts point-level representations from a downsampled 3D scene, and includes both interaction prompts and textual instructions to resemble human interactions with the 3D environment. Orthogonal to the efforts of aligning 3D-language with limited 3D data in existing 3DLAs, PerLA aims to improve the capability of 3DLAs in perceiving scene details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "Visual perception enhancement on multimodal models. Multi-granularity representation learning has been explored in 2D multimodal models [5, 18, 20, 29, 49, 51, 63], showing that combining local and global views yields more informative representations than relying on a single global view. Mini-Gemini [32] introduces a dual vision encoder setup using CLIP ViT [48] and ConvNeXt [37] to process low- and high-resolution views of an input image. Mod-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "text", + "content": "14370" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "els in the LLaVA-Next series [30, 35, 62] and InternLM-XComposer2 [17] use an additional branch to handle view partitions, increasing the number of visual tokens and preserving more fine-grained visual details compared to language-vision models with a single global view branch. In the 3D domain, Scene-LLM [19] and Segment3D [24] improve segmentation accuracy by transferring semantic details from multi-view images to point clouds. While extensive work has been done to enhance perception in 2D models and transfer multi-view image information into 3D, methods for preserving detailed representations from point clouds in 3DLAs remain underexplored. Our method aims to address this gap by preserving both global and local visual information to enhance 3D perception." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 282, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 282, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 282, + 260 + ], + "type": "text", + "content": "3. Perceptive 3D language assistant (PerLA)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 267, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 295, + 553 + ], + "type": "text", + "content": "PerLA takes as inputs " + }, + { + "bbox": [ + 56, + 267, + 295, + 553 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 267, + 295, + 553 + ], + "type": "text", + "content": " a text prompt in natural language, " + }, + { + "bbox": [ + 56, + 267, + 295, + 553 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 56, + 267, + 295, + 553 + ], + "type": "text", + "content": " the 3D scene in the form of a point cloud, and " + }, + { + "bbox": [ + 56, + 267, + 295, + 553 + ], + "type": "inline_equation", + "content": "iii" + }, + { + "bbox": [ + 56, + 267, + 295, + 553 + ], + "type": "text", + "content": " a visual prompt provided as either a user click or a bounding box. The text prompt is processed by a text prompt encoder to produce text representations, which are then input to both the Large Language Model (LLM) and the multimodal adapter (MMA). The text encoder is a transformer based on BLIP-2 [31]. The point cloud is processed by our perceptive scene encoder, which generates scene representations that feed into both the MMA and the subsequent encoder. We will detail the perceptive scene encoder in the next sections. The visual prompt is processed by the visual prompt encoder, which, by combining the perceptive scene encoder's representations, outputs scene representations that are further processed by the MMA. For more details on visual prompts, please refer to Supp. Mat. The MMA takes as input these multimodal representations and outputs tokens for the LLM. The MMA is implemented as a Q-former [31]. MMA's output is projected into the LLM's representation space through linear projector. Lastly, these projected representations are processed by the LLM to generate the output response. We train PerLA using data provided by [21] and finetune it for each downstream task. Fig. 2 illustrates our approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "content": "Formally, let " + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\left\\{\\left(\\pmb {p}_i\\in \\mathbb{R}^3,\\pmb {f}_i\\in \\mathbb{R}^{d_0}\\right)\\mid i = 1,2,\\dots ,N\\right\\}" + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "content": " denote PerLA's input point cloud, where " + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\pmb{p_i}" + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "content": " is a point coordinate, " + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\pmb{f_{i}}" + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "inline_equation", + "content": "d_0" + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "content": " -dimensional feature vector (e.g., color or normal vector) corresponding to " + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\pmb {p}_i" + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "content": " is the number of points (cardinality). Let " + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^t" + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^v" + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "content": " denote the input text prompt and visual prompt, respectively. Based on these inputs, PerLA (denoted as " + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "content": ") generates a free-form natural language response " + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{O}^t" + }, + { + "bbox": [ + 55, + 554, + 295, + 651 + ], + "type": "text", + "content": " for various 3D-related tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 659, + 194, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 194, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 194, + 672 + ], + "type": "text", + "content": "3.1. Perceptive scene encoder" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "To compute point-level representations through the scene encoder, " + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": " is typically downsampled into super-points using Farthest Point Sampling (FPS) [8, 9]. However, down" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "sampling and using only " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " may hinder the encoder to capture fine-grained details, thus affecting performance in downstream 3D scene understanding tasks. Our proposed perceptive scene encoder is designed to preserve such scene details without increasing the number of tokens or the representation dimensions of the 3D scene. We propose to split " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " into parts, and employ a pre-trained 3D scene encoder to encode these parts and the whole point cloud separately. We then aggregate the output representations from these different pieces into a single (highly-informative) representation using a cross-attention-based module and Graph Convolutional Network (GCN)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 218, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 338 + ], + "type": "text", + "content": "Hilbert-based scene partitioning. We partition " + }, + { + "bbox": [ + 313, + 218, + 553, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 218, + 553, + 338 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 218, + 553, + 338 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 218, + 553, + 338 + ], + "type": "text", + "content": " equally-sized parts, each containing the same number of points, " + }, + { + "bbox": [ + 313, + 218, + 553, + 338 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\frac{N}{L} \\right\\rfloor" + }, + { + "bbox": [ + 313, + 218, + 553, + 338 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 218, + 553, + 338 + ], + "type": "inline_equation", + "content": "\\lfloor \\cdot \\rfloor" + }, + { + "bbox": [ + 313, + 218, + 553, + 338 + ], + "type": "text", + "content": " denotes the greatest integer less than or equal to its argument. Enforcing equal cardinality across parts allows for spatially smaller parts in highly structured regions (areas with more semantic information) and spatially larger parts in less structured regions (areas with less semantic information). To achieve this, we adopt a Hilbert curve approach [50] to efficiently serialize the point cloud and partition it into parts [33, 57, 60]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "content": "Global and partial scene encoding. Both full-scene point cloud and the partial-scene point clouds are encoded separately by the same pre-trained 3D scene encoder " + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "content": ". The 3D scene encoder " + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "content": " downsamples the input point cloud into super-points using Farthest Point Sampling (FPS) [41] and produces a representation for each super-point. The representation of the full scene, i.e., the global representation, encode the overall scene context, while the representations of the partial scenes, i.e., the local representations, encode scene details. Let " + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^g = \\{\\pmb{p}_i^g \\in \\mathbb{R}^3\\}_{i=1}^M" + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "content": " denote the downsampled full-scene point cloud composed of " + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "content": " points, and " + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^g = \\{\\pmb{f}_i^g \\in \\mathbb{R}^d\\}_{i=1}^M" + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "content": " denote the associated global representations. Let " + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^l = \\{\\pmb{p}_i^l \\in \\mathbb{R}^3\\}_{i=1}^{L \\cdot M}" + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "content": " denote the set of downsampled partial-scene point clouds, and " + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^l = \\{\\pmb{f}_i^l \\in \\mathbb{R}^d\\}_{i=1}^{L \\cdot M}" + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "content": " the associated local representations. Note that, we aggregate all the points of the downsampled partial-scene point clouds into the single set " + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^l" + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "content": ", and both the full-scene and the partial-scenes are downsampled to the same number of super-points " + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "content": ". The representations " + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^l" + }, + { + "bbox": [ + 313, + 340, + 553, + 651 + ], + "type": "text", + "content": " provide a more detailed view of the scene, as they are derived from the same number of points downsampled from smaller, localized regions, resulting in higher resolution compared to the global versions. Yet, their semantic visibility is focused within each local part. In the following steps, we aggregate local and global information, to enrich the representations with details within the scene context." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Hilbert-based nearest-neighbor search. In order to enhance global representations through local representations, we need to first find the correspondences among them, i.e., which super-points of partial-scene point clouds are spatially neighbors with super-points of the full-scene point" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 751, + 317, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 751, + 317, + 760 + ], + "spans": [ + { + "bbox": [ + 294, + 751, + 317, + 760 + ], + "type": "text", + "content": "14371" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 553, + 208 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 553, + 208 + ], + "type": "image", + "image_path": "527958e90390560bc5710e6da3bc2c83bd2e7ea02ca04705e8d33e86669ce5c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 217, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 54, + 217, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 555, + 285 + ], + "type": "text", + "content": "Figure 2. Overview of PerLA. (Left): The overall pipeline of PerLA, which begins by extracting interaction-aware 3D scene representations. These representations are then projected onto the prefix of textual instructions via MMA, serving as input to a frozen language model (LLM). (Right): The detailed design of PerLA. First, the 3D scene is divided into spatially compact regions using Hilbert-based scene serialization [50]. Next, an efficient " + }, + { + "bbox": [ + 54, + 217, + 555, + 285 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 217, + 555, + 285 + ], + "type": "text", + "content": "-NN algorithm associates each point-level global representation with its detail-enriched local representations, creating a comprehensive scene representation through a Graph Convolutional Network (GCN). Finally, smoothness and regularization losses are applied to promote stable learning for the proposed perceptive scene encoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "text", + "content": "cloud. This can be done via " + }, + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "text", + "content": "-NN search. However, traditional " + }, + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "text", + "content": "-NN searches can be computationally intensive on large-scale or high-resolution point clouds and cannot ensure that the identified neighbors maintain geometric consistency. To address these limitations, we approximate " + }, + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "text", + "content": "-NN with an efficient neighbor mapping technique using Hilbert serialized point clouds, which improves query speed while reducing computational complexity. With serialized point clouds, we can leverage point indices to perform " + }, + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "text", + "content": "-NN searches with " + }, + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "inline_equation", + "content": "O(1)" + }, + { + "bbox": [ + 54, + 296, + 296, + 416 + ], + "type": "text", + "content": " complexity [60]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": "Specifically, we first serialize the union " + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^g \\bigcup \\mathcal{P}^l" + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": " using the Hilbert curve ordering. We then apply geometric partitioning [27] on the original point clouds to generate geometric labels " + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^g \\bigcup \\mathcal{Y}^l" + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": ". The center points of these labels are incorporated as high-order bits (by bit shifting) in the serialized index, ensuring that points with different labels occupy non-overlapping index ranges. Next, for each " + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^g \\in \\mathcal{P}^g" + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": ", we identify the " + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathtt{k}" + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": " nearest super-points within " + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^l" + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": " based on the serialized indexes. These geometric labels guarantee that " + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^g" + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": " and its nearest local super-points all originate from the same instance. Let " + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{k}_i}^l" + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": " denote the set of " + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathtt{k}" + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": " nearest local super-points to the global super-point " + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^g" + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": ". Consequently, " + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{k}_i}^l" + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": " are the representations of the super-points in " + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{k}_i}^l" + }, + { + "bbox": [ + 55, + 419, + 296, + 578 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Local-global representation aggregation. We make global representations more informative by combining them with information from local representations. To enable this, we present a novel two-step aggregation technique. In the first step, we employ a novel cross-attention algorithm between local and global representations using neighborhood information to update global representations, which we term as localized cross-attention. In the second step, we refine global representations through a Graph Convolutional Network based formulation based on message passing, which we term as GCN message passing (Fig. 2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "text", + "content": "Localized cross-attention operates with a learning-based weighting mechanism that is a function of the representations and relative positions of global super-points and their associated local neighbors. We constrain cross-attention to local neighborhood regions because " + }, + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "text", + "content": ") limiting the number of points reduces computational complexity, and " + }, + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "text", + "content": ") neighbors of a given point are likely to belong to the same object. Specifically, for each point " + }, + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i^g" + }, + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "text", + "content": " and its nearest neighbors " + }, + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{k}_i}^l" + }, + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "text", + "content": ", their relative position embeddings " + }, + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i" + }, + { + "bbox": [ + 313, + 296, + 555, + 403 + ], + "type": "text", + "content": " are extracted as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 338, + 404, + 553, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 404, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 338, + 404, + 553, + 424 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {i} = \\left\\{\\mathcal {R} _ {i j} = \\operatorname {p o s} \\left(\\left(\\boldsymbol {p} _ {i} ^ {g} - \\boldsymbol {p} _ {j} ^ {l}\\right) / \\sigma\\right) \\in \\mathbb {R} ^ {d} \\right\\} _ {j = 1} ^ {K ^ {l}}, \\tag {1}", + "image_path": "73080081b07aa661be1c9f683da7638244220e5479ed4a6b9d89711edb988bab.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 426, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 426, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\sigma > 0" + }, + { + "bbox": [ + 313, + 426, + 555, + 460 + ], + "type": "text", + "content": " is a learnable parameter that controls the relative position scaling, and " + }, + { + "bbox": [ + 313, + 426, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{pos}(\\cdot)" + }, + { + "bbox": [ + 313, + 426, + 555, + 460 + ], + "type": "text", + "content": " is the 3D Fourier positional embedding [52] operation, which satisfies:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 348, + 461, + 553, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 461, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 348, + 461, + 553, + 473 + ], + "type": "interline_equation", + "content": "\\left. \\operatorname {p o s} (\\boldsymbol {x}) = \\left[ \\sin (2 \\pi \\boldsymbol {x} \\cdot B); \\cos (2 \\pi \\boldsymbol {x} \\cdot B) \\right], \\quad \\right. \\tag {2}", + "image_path": "8011debe0135c95e2bfaf1eafcd4277419dd9cfe53b43b0c3c33f8faa7e4d237.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 473, + 553, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 553, + 497 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 473, + 553, + 497 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{3 \\times (d/2)}" + }, + { + "bbox": [ + 313, + 473, + 553, + 497 + ], + "type": "text", + "content": " is a learnable matrix. We update the global representations through cross-attention as follows" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 498, + 523, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 498, + 523, + 513 + ], + "spans": [ + { + "bbox": [ + 344, + 498, + 523, + 513 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {i j} = \\left(W _ {q} \\boldsymbol {f} _ {i} ^ {g}\\right) ^ {\\top} \\left(W _ {k} \\left(\\boldsymbol {f} _ {j} ^ {l} + W _ {r} \\mathcal {R} _ {i j}\\right)\\right) / \\sqrt {d},", + "image_path": "776b74426bc1a45dd6da6f51c07967c813a3a5625991dfae2acb2aace0636f0b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 344, + 514, + 553, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 514, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 344, + 514, + 553, + 531 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {i} = \\left\\{\\boldsymbol {s} _ {i j} \\right\\} _ {i = 1} ^ {K ^ {l}}, \\boldsymbol {w} _ {i} = \\text {s o f t m a x} \\left(\\boldsymbol {s} _ {i}\\right), \\tag {3}", + "image_path": "67b024eceddd8983624813913c13cc17226cc57a593321033bf2c948a0b5b214.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 344, + 534, + 488, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 534, + 488, + 548 + ], + "spans": [ + { + "bbox": [ + 344, + 534, + 488, + 548 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {f}} _ {i} ^ {g} = \\boldsymbol {f} _ {i} ^ {g} + \\boldsymbol {w} _ {i} \\left(W _ {v} \\left(\\mathcal {F} _ {\\mathcal {K} _ {i}} ^ {l} + \\mathcal {R} _ {i}\\right)\\right),", + "image_path": "557492da78a07d10db25c244a8daf157ffdbd24ffeb9f8df30883accfe4e5ec1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 555, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 553, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 555, + 553, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_i^g" + }, + { + "bbox": [ + 313, + 555, + 553, + 594 + ], + "type": "text", + "content": " are the updated representations, and " + }, + { + "bbox": [ + 313, + 555, + 553, + 594 + ], + "type": "inline_equation", + "content": "W_{q},W_{k},W_{v},W_{r}" + }, + { + "bbox": [ + 313, + 555, + 553, + 594 + ], + "type": "text", + "content": " are projection parameter matrices. Let " + }, + { + "bbox": [ + 313, + 555, + 553, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}^g = \\{\\hat{\\pmb{f}}_i^g\\}_{i = 1}^M" + }, + { + "bbox": [ + 313, + 555, + 553, + 594 + ], + "type": "text", + "content": " denote the set of " + }, + { + "bbox": [ + 313, + 555, + 553, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_i^g" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "GCN message passing further refines " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}^g" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " by aggregating information from neighboring global super-points. Let " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^g = \\{\\mathcal{P}^g, \\mathcal{W}^g\\}" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " be a " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "-NN graph, where " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^g" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " is the set of the vertices, i.e., global super-points, and " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^g \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " is the adjacency matrix defining the edges. The construction of the adjacency matrix uses the similarity between superpoint representations " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^G" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ". Note that we use the (original) global representations to define the adjacency matrix because we empirically experienced more stability in training as opposed to using the updated global representation," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 751, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 751, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 751, + 318, + 761 + ], + "type": "text", + "content": "14372" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 99 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 99 + ], + "type": "text", + "content": "likely due to the fact that they change value during training. The adjacency matrix " + }, + { + "bbox": [ + 55, + 72, + 294, + 99 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^g" + }, + { + "bbox": [ + 55, + 72, + 294, + 99 + ], + "type": "text", + "content": " with elements " + }, + { + "bbox": [ + 55, + 72, + 294, + 99 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{ij}^{g}" + }, + { + "bbox": [ + 55, + 72, + 294, + 99 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 100, + 286, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 100, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 64, + 100, + 286, + 137 + ], + "type": "interline_equation", + "content": "\\mathcal {W} _ {i j} ^ {g} = \\left\\{ \\begin{array}{l l} \\frac {\\boldsymbol {f} _ {i} ^ {g \\top} \\boldsymbol {f} _ {j} ^ {g}}{\\| \\boldsymbol {f} _ {i} ^ {g} \\| \\| \\boldsymbol {f} _ {j} ^ {g} \\|} \\cdot \\left(\\boldsymbol {f} _ {i} ^ {g \\top} \\boldsymbol {f} _ {j} ^ {g} > 0\\right), & \\text {i f p _ {j} ^ {g} \\in k - N N (p _ {i} ^ {g})}, \\\\ 0, & \\text {o t h e r w i s e}. \\end{array} \\right.", + "image_path": "698b5ff87faf93449983e17aaa9f464cfc0e2080f5b327c69da12ba2d230839b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 139, + 295, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 295, + 163 + ], + "type": "text", + "content": "Lastly, we perform one-time GCN message passing (MP) on the graph " + }, + { + "bbox": [ + 55, + 139, + 295, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^g" + }, + { + "bbox": [ + 55, + 139, + 295, + 163 + ], + "type": "text", + "content": " to obtain an updated feature matrix " + }, + { + "bbox": [ + 55, + 139, + 295, + 163 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}^g" + }, + { + "bbox": [ + 55, + 139, + 295, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{W}}^g" + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "text", + "content": " be the symmetrically normalized adjacency matrix of " + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^g" + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{W}}^g = \\mathbf{D}^{-\\frac{1}{2}}\\mathcal{W}^g\\mathbf{D}^{-\\frac{1}{2}}\\in \\mathbb{R}^{M\\times M}" + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "text", + "content": " is the diagonal degree matrix of " + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^g" + }, + { + "bbox": [ + 55, + 163, + 296, + 199 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 201, + 295, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 201, + 295, + 221 + ], + "spans": [ + { + "bbox": [ + 89, + 201, + 295, + 221 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {F}} ^ {g} \\leftarrow \\operatorname {R e L U} \\left(W _ {m} \\left(\\tilde {\\mathcal {W}} ^ {g} \\hat {\\mathcal {F}} ^ {g}\\right) + W _ {s} \\hat {\\mathcal {F}} ^ {g}\\right), \\tag {4}", + "image_path": "0590ee5d0c3aab22b52c29e292156d4ede8e06cb1a84f445d0b09e4a7a06fb26.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "type": "inline_equation", + "content": "W_{m}" + }, + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "type": "inline_equation", + "content": "W_{s}" + }, + { + "bbox": [ + 55, + 222, + 296, + 270 + ], + "type": "text", + "content": " are trainable parameters. This MP layer structure allows each node to aggregate information from its neighbors while also preserving its own feature representation (via the skip connection)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 277, + 289, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 289, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 289, + 290 + ], + "type": "text", + "content": "3.2. Learning with local representation consensus" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 295, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 438 + ], + "type": "text", + "content": "The main learning objective encourages the model to generate outputs that are close to the target responses in the training dataset. This objective is typically expressed as a per-token cross-entropy loss " + }, + { + "bbox": [ + 55, + 295, + 296, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pred}" + }, + { + "bbox": [ + 55, + 295, + 296, + 438 + ], + "type": "text", + "content": " [9, 21] (see details in our Supp. Mat.). We found that adding a consensus term to " + }, + { + "bbox": [ + 55, + 295, + 296, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pred}" + }, + { + "bbox": [ + 55, + 295, + 296, + 438 + ], + "type": "text", + "content": " that encourages local regularization of representations improves training stability. Firstly, points belonging to the same object (i.e., local neighborhoods) should share similar aggregated representations. Secondly, aggregated super-point representations should remain close to their original point representations to preserve context knowledge. Our novel consensus loss " + }, + { + "bbox": [ + 55, + 295, + 296, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{con}}" + }, + { + "bbox": [ + 55, + 295, + 296, + 438 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 101, + 441, + 190, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 441, + 190, + 454 + ], + "spans": [ + { + "bbox": [ + 101, + 441, + 190, + 454 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n}} = \\mathcal {L} _ {\\text {s m t}} + \\mu \\mathcal {L} _ {\\text {r e g}},", + "image_path": "90c7ad621a6888aa9eb0f174f5bdf9f045529e1e41e78223790caeb7813e5fe8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 101, + 456, + 295, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 456, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 101, + 456, + 295, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s m t}} = \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} \\mathcal {W} _ {i j} ^ {g} \\left\\| \\frac {\\hat {\\boldsymbol {f}} _ {i} ^ {g}}{\\sqrt {d _ {i}}} - \\frac {\\hat {\\boldsymbol {f}} _ {j} ^ {g}}{\\sqrt {d _ {j}}} \\right\\|, \\tag {5}", + "image_path": "adb40d66290e7eb579899156fed7d20361249bd468081775a6de38780bcffc74.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 493, + 203, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 493, + 203, + 525 + ], + "spans": [ + { + "bbox": [ + 102, + 493, + 203, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e g}} = \\sum_ {i = 1} ^ {N} \\left\\| \\hat {\\pmb {f}} _ {i} ^ {g} - \\pmb {f} _ {i} ^ {g} \\right\\|,", + "image_path": "2b705da4bb19a153bcb6d2ff241c236a4e43882e1f3492e6289d713dae945290.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "inline_equation", + "content": "d_{k} = \\sum_{l}\\mathcal{W}_{kl}^{g}" + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "text", + "content": " is the diagonal matrix representing row-wise sums of " + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^g" + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "text", + "content": ". The first term " + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{smt}}" + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "text", + "content": " enforces spatial connectivity by promoting similarity among fused representations of points within the same object, addressing the first attribute. " + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "text", + "content": " is a regularization term that promotes the learned representations stay close to the original global representations. " + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "text", + "content": " is a hyperparameter. The overall loss " + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "text", + "content": ", is a weighted sum of " + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pred}" + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 528, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 627, + 295, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 627, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 132, + 627, + 295, + 639 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda \\mathcal {L} _ {\\text {c o n}} + \\mathcal {L} _ {\\text {p r e d}}, \\tag {6}", + "image_path": "18c4ef6537508c4741e5e9d49e35f22d98348c3021c0869a799eb8763d752691.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " is the balance hyperparameter. By jointly training with these two loss terms, we encourage the model to learn more object-aware representations and enhance its ability to extract finer details. This approach also preserves global context information, which is essential for capturing spatial relationships between different objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "During inference, we use beam search to predict response " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "\\mathcal{O}^t" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": " that maximizes the following objective:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 106, + 555, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 106, + 555, + 124 + ], + "spans": [ + { + "bbox": [ + 364, + 106, + 555, + 124 + ], + "type": "interline_equation", + "content": "\\mathcal {O} ^ {t} = \\arg \\max _ {\\mathcal {O}} \\Phi (\\mathcal {O} \\mid \\mathcal {P}, \\mathcal {I} ^ {t}, \\mathcal {I} ^ {v}). \\tag {7}", + "image_path": "7c86b0910f006281706b0a2e07e4b778dd93f0033e8de74ab0a5c1610a59b0f8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 129, + 440, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 129, + 440, + 140 + ], + "spans": [ + { + "bbox": [ + 314, + 129, + 440, + 140 + ], + "type": "text", + "content": "where we set a beam size of 4." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 153, + 395, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 395, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 395, + 167 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 174, + 554, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 554, + 209 + ], + "type": "text", + "content": "We evaluate PerLA on 3D Question Answering and 3D Dense Captioning downstream tasks, and compare its performance with state-of-the-art methods from the literature." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 212, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 320 + ], + "type": "text", + "content": "Datasets. We conduct experiments using the ScanNet dataset [14], which encompasses 1,201 training and 312 validation scenes, featuring diverse and complex indoor 3D environments. Language annotations are from ScanQA [2], ScanRefer [6], Nr3D [1], and the ScanNet subset of 3D-LLM [21], collectively supporting a range of tasks including instance and scene descriptions, conversations, embodied planning, and question answering. For additional data statistics, please refer to the supplementary materials." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 322, + 554, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 554, + 371 + ], + "type": "text", + "content": "Metrics. We follow LL3DA's evaluation protocol [9] to evaluate the quality of output responses. We use the abbreviations C, B4, M and R for CiDEr [55], BLEU-4 [43], METEOR [3], and Rouge-L [34], respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "text", + "content": "Implementation details. As in [9, 12], we input 40,000 randomly sampled points from each 3D scene. We divide the point cloud into partitions " + }, + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "inline_equation", + "content": "L = 6" + }, + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "text", + "content": " and choose " + }, + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "inline_equation", + "content": "k = 4L = 24" + }, + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "text", + "content": " neighbors to enhance global representations. We set " + }, + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\lambda = \\mu = 0.1" + }, + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "text", + "content": " in the loss. We use the pre-trained OPT-1.3B [66] language model, kept frozen and loaded in float16 precision. We use the AdamW [38] optimizer with a weight decay of 0.1, applying a cosine annealing scheduler that decays the learning rate from " + }, + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 313, + 373, + 555, + 540 + ], + "type": "text", + "content": " over approximately 100,000 iterations. All tasks are trained with a total batch size of 16. Each training process is completed within two days using up to two NVIDIA H100 P0 (96GB) GPUs. For each evaluation, we fine-tune the model's parameters on the respective task for about 30,000 iterations." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 550, + 372, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 372, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 372, + 561 + ], + "type": "text", + "content": "4.1. Results" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "3D question answering is a task that involves answering questions about a 3D scene. It allows a model to provide information about the objects, relationships, and attributes within a 3D environment based on a given question. Tab. 1 shows the results of ScanQA's validation and test sets. Classification-based methods (CLS) select responses from a predefined answer set. Generation-based approaches (GEN) generate the entire textual response. PerLA consistently outperforms existing approaches across all evaluation sets and metrics, in particular with " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "+1.34" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " CiDEr score over LL3DA. We also compare against our reproduced version of LL3DA, where PerLA scores " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "+3.76" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " CiDEr." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 751, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 751, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 751, + 317, + 761 + ], + "type": "text", + "content": "14373" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 96, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparative results for 3D Question Answering on ScanQA [2] benchmark. CLS and GEN denote classification-based and generation-based methods, respectively. LL3DA (repr.) is results of LL3DA we reproduced. PerLA outperforms all the other methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 96, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 56, + 96, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 553, + 240 + ], + "type": "table", + "html": "
MethodTypeValidationTest w/ objectTest w/o object
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
ScanQA[2]CLS64.8610.0813.1433.3367.2912.0413.5534.3460.2410.7512.5931.09
Clip-Guided[44]-----69.5314.6413.9435.1562.8311.7313.2832.41
Multi-CLIP[16]CLS----68.7012.6513.9735.4663.2012.8713.3632.61
3D-VLP[26]CLS66.9711.1513.5334.5170.1811.2314.1635.9763.4015.8413.1331.79
3D-VisTA[70]-----68.6010.5013.8035.5055.708.7011.6929.60
3D-LLM [21]GEN69.4012.0014.5035.7069.6011.6014.9035.30----
LL3DA[9]GEN76.7913.5315.8837.3178.1613.9716.3838.1570.2912.1914.8535.17
LL3DA (repr.)GEN74.3713.5015.0936.31--------
PerLAGEN78.1314.4917.4439.6080.9117.2116.4940.7174.8214.9715.2338.18
Δ w.r.t. LL3DA[9]-+1.34+0.96+1.56+2.29+2.75+3.24+0.11+2.56+4.53+2.78+0.38+3.01
", + "image_path": "754915ddb53c7db43e840ab5796eb80b07d491119b72c41f20bbb52ca0d04dce.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 248, + 312, + 365 + ], + "blocks": [ + { + "bbox": [ + 56, + 248, + 312, + 365 + ], + "lines": [ + { + "bbox": [ + 56, + 248, + 312, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 248, + 312, + 365 + ], + "type": "image", + "image_path": "e7dfa24369b410d084bdfa2c2d2c3307b13db995d359a8216dd4518dcf3c2a5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 369, + 555, + 391 + ], + "lines": [ + { + "bbox": [ + 54, + 369, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 555, + 391 + ], + "type": "text", + "content": "Figure 3. The qualitative comparison between our method, PerLA, and LL3DA [9] on the ScanQA [2] dataset shows that our approach achieves higher accuracy in responding to \"what\"-related questions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 318, + 247, + 553, + 365 + ], + "blocks": [ + { + "bbox": [ + 318, + 247, + 553, + 365 + ], + "lines": [ + { + "bbox": [ + 318, + 247, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 318, + 247, + 553, + 365 + ], + "type": "image", + "image_path": "230abff78840d6d3ab5edf3cf836a2879436956dada63378a847bf8077ebed90.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 398, + 297, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 398, + 297, + 589 + ], + "spans": [ + { + "bbox": [ + 54, + 398, + 297, + 589 + ], + "type": "text", + "content": "Fig. 3 provides qualitative comparisons between LL3DA and PerLA on the ScanQA benchmark [2]. It highlights PerLA's accuracy in answering questions regarding object attributes and spatial relationships within a 3D scene. PerLA provides precise answers, correctly identifying objects such as a \"picture\", \"trash bin\", and \"TV\" along with their specific colors and types, while LL3DA often yields incorrect or less specific responses. In particular, PerLA accurately identifies the colors of objects within the scene, such as the door, the monitor, and the instrument case, correctly answering the questions about these attributes. In contrast, LL3DA frequently misidentifies colors, labeling objects as \"White\" instead of \"Black,\" for example. This illustrates PerLA's superior ability to capture fine-grained details and contextual information, delivering more accurate answers in complex environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": "3D dense captioning involves localizing and describing each 3D instance within complex 3D environments. Tab. 2 shows the results on ScanRefer [6] and " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": " [1] benchmarks. Following previous works [9, 21], we use the m@kIoU metric, where " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "m\\in \\{\\mathbf{C},\\mathbf{B4},\\mathbf{M},\\mathbf{R}\\}" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": " denote the IoU threshold. As in [12, 25], we report C@0.25 and C@0.5 for ScanRefer, and C@0.5 for " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": ". UniT3D [13], 3DJCG [4], and 3D-VLP [26] are pre-trained on multiple 3D vision and language tasks from ScanNet scenes. UniT3D uses image caption models and multi-view im" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 398, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 517 + ], + "type": "text", + "content": "ages to generate extra instance captions for pre-training. For fair comparison, we report results from models trained using standard per-word cross-entropy loss without additional 3D scenes. Box annotations are estimated with Vote2CapDETR and used as visual prompts. PerLA consistently outperforms comparison methods on both benchmarks. PerLA significantly outperforms LL3DA by scoring +3.75 and +4.22 CiDEr on ScanRefer, and +3.88 CiDEr on Nr3D. These results highlight the effectiveness of PerLA in 3D dense captioning tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 517, + 556, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 556, + 578 + ], + "type": "text", + "content": "Fig. 4 provides examples of qualitative results of LL3DA and PerLA on Nr3D and ScanRefer. PerLA produces more accurate and detailed descriptions, effectively capturing spatial relationships and fine-grained attributes, such as the positioning of objects relative to surrounding elements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 586, + 406, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 406, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 406, + 600 + ], + "type": "text", + "content": "4.2. Ablation study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "content": "Local, global, and GCN representations. To investigate the impact of local and global representations, we design two variants of PerLA: one using only local representations and the other using only global representations. We also report an extended version of LL3DA, termed " + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{LL3DA}^{\\dagger}" + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "content": ", in which we increase the number of query tokens used to encode local partitions. Representations are first extracted from each partition, and FPS is applied to sample 1,024 points with their corresponding representations from the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "text", + "content": "14374" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 95, + 553, + 381 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "text", + "content": "Table 2. Comparative results for 3D Dense Captioning on ScanRefer [6] and Nr3D [1] benchmarks. LL3DA (repr.) is results of LL3DA we reproduced. Generally, PerLA outperforms all other methods on both benchmarks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 95, + 553, + 381 + ], + "lines": [ + { + "bbox": [ + 56, + 95, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 553, + 381 + ], + "type": "table", + "html": "
MethodScanRefer@0.25ScanRefer@0.5Nr3D@0.5
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
Scan2Cap[12]56.8234.1826.2955.2739.0823.3221.9744.7827.4717.2421.8049.06
MORE[25]62.9136.2526.7556.3340.9422.9321.6644.42----
SpaCap3D[56]-----44.0225.2622.3333.7119.9222.6150.50
REMAN[40]62.0136.3726.7656.2545.0026.3123.1346.9634.8120.3722.7150.90
D3Net[7]-----51.67--35.2620.4222.7753.38
Contextual[68]-----46.0723.4023.95----
UniT3D[13]----46.6927.5221.9145.98----
3DJCG[4]64.7040.1727.6359.2349.4831.6324.3650.8038.0622.8223.7752.99
3D-VLP[26]70.7341.0328.1459.7254.9432.3124.8351.51----
3D-VisTA*[70]----61.6034.1026.8055.00----
Vote2CapDETR[8]71.4539.3428.2559.6361.8134.4626.2254.4043.8426.6825.4154.43
LL3DA[9]74.1741.4127.7659.5365.1936.7925.9755.0651.1828.7525.9156.61
LL3DA (repr.)71.8639.5727.2958.3763.7935.6725.9454.5648.3828.3625.7255.66
PerLA77.9243.4128.9759.6969.4138.0229.0756.8055.0631.2428.5259.13
Δ w.r.t. LL3DA[9]+3.75+2.00+1.21+0.16+4.22+1.23+2.27+1.74+3.88+2.49+2.61+2.52
Describe this [bbox] object.
PerLA The toilet paper is on the wall next to the toilet.
LL3DA The towel closest to the toilet.
", + "image_path": "40a84db6c8fc9b74152688c1617e144cb94c6ff3705172b9d0a94a3cfeb0f838.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 396, + 553, + 487 + ], + "blocks": [ + { + "bbox": [ + 253, + 381, + 394, + 391 + ], + "lines": [ + { + "bbox": [ + 253, + 381, + 394, + 391 + ], + "spans": [ + { + "bbox": [ + 253, + 381, + 394, + 391 + ], + "type": "text", + "content": "(a) 3D visual dense captioning on " + }, + { + "bbox": [ + 253, + 381, + 394, + 391 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 253, + 381, + 394, + 391 + ], + "type": "text", + "content": " [1]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 396, + 553, + 487 + ], + "lines": [ + { + "bbox": [ + 58, + 396, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 58, + 396, + 553, + 487 + ], + "type": "table", + "html": "
Describe this [bbox] object.Describe this [bbox] object.
PerLA\nThere is a rectangular bathroom vanity. It is to the right of the toilet.PerLA\nThis is a white sink.\nIt is to the left of the toilet.
LL3DA\nThis is a bathroom sink. The sink is on the left side of the room.LL3DA\nThis is a white sink.\nIt is on the left side of the sink.
(b) 3D visual dense captioning on ScanRefer [6]
", + "image_path": "6b03d220d4cb43b9a38d2231b9af24499221c9a326eba35d98be6b2852e225ac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 495, + 553, + 580 + ], + "blocks": [ + { + "bbox": [ + 58, + 495, + 553, + 580 + ], + "lines": [ + { + "bbox": [ + 58, + 495, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 58, + 495, + 553, + 580 + ], + "type": "table", + "html": "
Describe this [bbox] object.Describe this [bbox] object.
PerLA\nThis is a black computer tower. It is under a desk.\nLL3DA\nThe computer is on the left side of the desk. The computer is on the right side.PerLA\nThis is a black computer tower. It is under a desk.\nLL3DA\nThere is a black desk. It is on the left side of the room.
", + "image_path": "065e45acfb73d7b974ca00b474647fbd5138615ef82db58b1053cfad138932de.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 581, + 555, + 604 + ], + "lines": [ + { + "bbox": [ + 55, + 581, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 555, + 604 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons on the dense captioning task across the Nr3D [1] and ScanRefer [6]. We compare the results of our PerLA with LL3DA [9]. PerLA generates accurate descriptions, effectively capturing fine-grained object attributes and spatial relationships." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 615, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 615, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 615, + 297, + 712 + ], + "type": "text", + "content": "union of these partitions. These representations are then passed through a multimodal adapter to produce 32 global tokens. LL3DA generates an additional 32 global tokens from the point cloud of the entire scene, resulting in a total of 64 tokens, which are concatenated and fed into the LLM for response generation. (For details of " + }, + { + "bbox": [ + 54, + 615, + 297, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{LL3DA}^{\\dagger}" + }, + { + "bbox": [ + 54, + 615, + 297, + 712 + ], + "type": "text", + "content": " please refer to the Supp. Mat.). Tab. 3 shows that the combination of local and global representations yields a much better" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 615, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 556, + 712 + ], + "type": "text", + "content": "performance than the two independently used. Tab. 3 also shows that " + }, + { + "bbox": [ + 313, + 615, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{LL3DA}^{\\dagger}" + }, + { + "bbox": [ + 313, + 615, + 556, + 712 + ], + "type": "text", + "content": " does not reach PerLA performance, underlying the importance of information exchange between local and global representations. To evaluate the effectiveness of GCN, we also report the results without using GCN (w/o GCN). Compared to our PerLA, which incorporates GCN, it consistently improves performance across the three datasets, confirming its effectiveness." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 751, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 751, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 751, + 318, + 761 + ], + "type": "text", + "content": "14375" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 95, + 552, + 186 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 553, + 93 + ], + "type": "text", + "content": "Table 3. Ablation study of our local-to-global, GCN representation aggregation algorithms on ScanQA, ScanRefer[6] and Nr3D[1] benchmarks. LL3DA† denotes an extended version of LL3DA with an increased number of query tokens." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 95, + 552, + 186 + ], + "lines": [ + { + "bbox": [ + 56, + 95, + 552, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 552, + 186 + ], + "type": "table", + "html": "
MethodScanQAScanRefer@0.5Nr3D@0.5
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
\\( LL3DA^{\\dagger} \\)74.5412.8915.1136.9662.2534.5025.5553.8448.7028.2425.7255.59
Global74.4913.5015.1636.5563.7935.6725.9454.5648.3828.3625.7255.66
Local73.5512.9814.9536.0762.7934.7825.9454.1649.0928.2025.8055.89
w/o GCN77.6113.8715.9339.2869.0437.4328.7155.5954.3029.7827.4158.32
PerLA78.1314.4917.4439.6069.4138.0229.0756.8055.0631.2428.5259.13
", + "image_path": "65fd3ce53c58d80042bb22df80896381ccf1a89c12ff30caece34afc08e90e66.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 214, + 293, + 258 + ], + "blocks": [ + { + "bbox": [ + 56, + 193, + 295, + 214 + ], + "lines": [ + { + "bbox": [ + 56, + 193, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 295, + 214 + ], + "type": "text", + "content": "Table 4. Ablation study on the impact of constrained attention on ScanQA validation dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 214, + 293, + 258 + ], + "lines": [ + { + "bbox": [ + 56, + 214, + 293, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 214, + 293, + 258 + ], + "type": "table", + "html": "
cross-attentionmean poolingmax pooling
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
78.114.517.439.674.412.515.236.574.013.014.835.7
", + "image_path": "2aea2dbcd627916650a43912a0f67331b78ce2b02c70c4a7eaaf61b7782412a0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 293, + 293, + 373 + ], + "blocks": [ + { + "bbox": [ + 56, + 270, + 295, + 292 + ], + "lines": [ + { + "bbox": [ + 56, + 270, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 295, + 292 + ], + "type": "text", + "content": "Table 5. Ablation study on the impact of joint learning by different loss combination on ScanQA validation dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 293, + 293, + 373 + ], + "lines": [ + { + "bbox": [ + 56, + 293, + 293, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 293, + 373 + ], + "type": "table", + "html": "
LossScanQA (Validation)
LpreLsmtLregC↑B4↑M↑R↑
✓75.3113.8315.9037.48
✓✓76.6213.9616.9938.14
✓✓76.4714.0716.7238.79
✓✓✓78.1314.4917.4439.60
", + "image_path": "f994f076557d2df002db75c207ca367adad30eaf5b180967587f40f0ac105437.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 385, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 385, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 385, + 295, + 529 + ], + "type": "text", + "content": "Localized cross-attention. To evaluate the effectiveness of our novel localized cross-attention module, we tested max pooling and mean pooling as alternative methods. Tab. 4 shows that localized cross-attention significantly outperforms both mean pooling and max pooling across all metrics. These results indicate that localized cross-attention enhances ability of model to generate more informative scene representations, thereby improving performance on 3D question answering tasks. This improvement comes from the incorporation of positional information and semantic similarity in cross-attention, which helps to exclude neighboring points that do not belong to the same object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "text", + "content": "Loss function. To evaluate the effectiveness of our novel loss term " + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{con}}" + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "text", + "content": ", we train PerLA using different combinations of " + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{con}}" + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "text", + "content": "’s components (Eq. 5). Since our task is intended for downstream applications, each combination also includes the task-specific loss " + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pre}}" + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "text", + "content": ". Tab. 5 shows that adding " + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{smt}}" + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pre}}" + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "text", + "content": " improves all metrics (C, B4, M, and R). Incorporating the loss of regularization " + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "text", + "content": " further enhances performance, particularly in metrics C, M, and R. These results show the effectiveness of learning jointly with both semantic awareness and regularization losses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": "Number of partitions. To evaluate the effectiveness of different number of partitions, we test PerLA by splitting the point cloud into 4, 6, and 8 partitions. Tab. 6 shows that increasing the number of partitions generally improves performance, with 6 partitions yielding the best results. Al-" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 214, + 552, + 258 + ], + "blocks": [ + { + "bbox": [ + 313, + 193, + 553, + 214 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 214 + ], + "type": "text", + "content": "Table 6. Ablation study on the impact of number of parts on ScanQA validation dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 214, + 552, + 258 + ], + "lines": [ + { + "bbox": [ + 315, + 214, + 552, + 258 + ], + "spans": [ + { + "bbox": [ + 315, + 214, + 552, + 258 + ], + "type": "table", + "html": "
468
C↑B4↑M↑R↑C↑B4↑M↑R↑C↑B4↑M↑R↑
76.913.616.137.378.114.517.439.678.014.517.239.4
", + "image_path": "f5be9ad6dfc37c6900dfc5d774b923bd3bd4fffb1a93f7e4f61caf85a8d67406.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 270, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 344 + ], + "type": "text", + "content": "though 8 partitions outperform 4, they show slightly lower performance than 6. This suggests that 6 partitions provide a better balance between granularity and effectiveness, capturing relevant scene details with lower computational complexity. More ablation studies, including an analysis of performance in different training strategies, refer to Supp. Mat.." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 354, + 392, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 392, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 392, + 366 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 564 + ], + "type": "text", + "content": "We presented PerLA, a perceptive 3D language assistant capable of capturing both detailed and contextual information to enhance visual representations for LLMs. PerLA features a dual-branch architecture: the global branch processes superpoints from the whole point cloud via downsampling, while the local branch focuses on partitioned regions. We demonstrated that by integrating representations from both branches, PerLA effectively captures scene details, reducing hallucinations. Moreover, we employ a Graph Convolutional Network to facilitate information exchange among neighboring local and global superpoints, and introduce a novel loss term for local representation consensus to promote training stability. Experiments on the ScanQA, ScanRefer, and Nr3D benchmarks highlight the effectiveness of our approach, setting a new state-of-the-art performance in 3D question answering and dense captioning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 567, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 553, + 662 + ], + "type": "text", + "content": "Limitations & Future Work. PerLA primarily focuses on enhancing performance using point cloud input, but integrating it with optimized modules, such as token merging, presents a promising direction to extend its capabilities. Although PerLA shows strong performance on standard benchmarks, future work could explore its robustness and generalizability in more complex 3D scenarios, broadening its applicability in diverse real-world settings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 662, + 553, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 553, + 711 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by PNRR FAIR - Future AI Research (PE00000013) and ICSC National Research Centre for HPC, Big Data and Quantum Computing (CN00000013), funded by NextGeneration EU." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 751, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 751, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 751, + 318, + 761 + ], + "type": "text", + "content": "14376" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 62, + 91, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 91, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 62, + 91, + 296, + 151 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 2, 5, 6, 7, 8, 14, 15, 16, 17, 18" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 152, + 296, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 152, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 62, + 152, + 296, + 199 + ], + "type": "text", + "content": "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, 2022. 2, 5, 6, 14, 15, 16, 17, 18" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 201, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 201, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 201, + 295, + 248 + ], + "type": "text", + "content": "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 5, 16" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 297 + ], + "type": "text", + "content": "[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In CVPR, pages 16464-16473, 2022. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 300, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 295, + 346 + ], + "type": "text", + "content": "[5] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In ICCV, pages 357–366, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 348, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 295, + 396 + ], + "type": "text", + "content": "[6] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 2, 5, 6, 7, 8, 14, 15, 16, 17, 18" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 445 + ], + "type": "text", + "content": "[7] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In ECCV, pages 487-505. Springer, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 446, + 295, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 295, + 493 + ], + "type": "text", + "content": "[8] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In CVPR, pages 11124-11133, 2023. 1, 2, 3, 7, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 495, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 495, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 495, + 295, + 554 + ], + "type": "text", + "content": "[9] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In CVPR, 2024. 1, 2, 3, 5, 6, 7, 12, 14, 16, 17, 18" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 295, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 295, + 614 + ], + "type": "text", + "content": "[10] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, YU Gang, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE TPAMI, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 616, + 295, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 295, + 664 + ], + "type": "text", + "content": "[11] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 666, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 666, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 666, + 295, + 712 + ], + "type": "text", + "content": "[12] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In CVPR, pages 3193-3203, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 120 + ], + "type": "text", + "content": "[13] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV, pages 18109-18119, 2023. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 122, + 553, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 122, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 316, + 122, + 553, + 169 + ], + "type": "text", + "content": "[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5, 14, 15" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 553, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 231 + ], + "type": "text", + "content": "[15] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 316, + 172, + 553, + 231 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 316, + 172, + 553, + 231 + ], + "type": "text", + "content": " objects. NeurIPS, 36, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 234, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 553, + 304 + ], + "type": "text", + "content": "[16] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 389 + ], + "type": "text", + "content": "[17] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 393, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 441 + ], + "type": "text", + "content": "[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, pages 6824-6835, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 443, + 553, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 490 + ], + "type": "text", + "content": "[19] Rao Fu, Jingyu Liu, Xilun Chen, Yixin Nie, and Henan Xiong. Scene-llm: Extending language model for 3d visual understanding and reasoning. arXiv preprint arXiv:2403.11401, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 492, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 540 + ], + "type": "text", + "content": "[20] Ritwik Gupta, Shufan Li, Tyler Zhu, Jitendra Malik, Trevor Darrell, and Karttikeya Mangalam. xt: Nested tokenization for larger context in large images. arXiv preprint arXiv:2403.01915, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 543, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 590 + ], + "type": "text", + "content": "[21] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. NeurIPS, 2023. 1, 2, 3, 5, 6, 14, 16, 17" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 592, + 553, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 651 + ], + "type": "text", + "content": "[22] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 654, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 553, + 712 + ], + "type": "text", + "content": "[23] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In ICML, 2024. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "text", + "content": "14377" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 143 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 143 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 143 + ], + "type": "text", + "content": "[24] Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis Engelmann. Segment3d: Learning fine-grained class-agnostic 3d segmentation without manual labels. In ECCV, pages 278–295. Springer, 2025. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 145, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 145, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 296, + 194 + ], + "type": "text", + "content": "[25] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In ECCV, pages 528-545. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 195, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 243 + ], + "type": "text", + "content": "[26] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In CVPR, pages 10984-10994, 2023. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 244, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 244, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 244, + 296, + 281 + ], + "type": "text", + "content": "[27] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, pages 4558-4567, 2018. 4, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 296, + 330 + ], + "type": "text", + "content": "[28] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2024. 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 296, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 296, + 378 + ], + "type": "text", + "content": "[29] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In CVPR, pages 7287-7296, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 380, + 296, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 380, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 56, + 380, + 296, + 440 + ], + "type": "text", + "content": "[30] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 296, + 501 + ], + "type": "text", + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 3, 12, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 503, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 296, + 563 + ], + "type": "text", + "content": "[32] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 565, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 296, + 612 + ], + "type": "text", + "content": "[33] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaoqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 296, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 651 + ], + "type": "text", + "content": "[34] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5, 16" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 652, + 296, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 652, + 296, + 699 + ], + "spans": [ + { + "bbox": [ + 56, + 652, + 296, + 699 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llavanext: Improved reasoning,OCR, and world knowledge, 2024.2,3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 701, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 701, + 296, + 714 + ], + "type": "text", + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 336, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 336, + 72, + 553, + 95 + ], + "type": "text", + "content": "Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 98, + 553, + 143 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 143 + ], + "type": "text", + "content": "[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 145, + 553, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 145, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 316, + 145, + 553, + 180 + ], + "type": "text", + "content": "[38] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 182, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 182, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 182, + 553, + 229 + ], + "type": "text", + "content": "[39] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 277 + ], + "type": "text", + "content": "[40] Aihua Mao, Zhi Yang, Wanxin Chen, Ran Yi, and Yong-jin Liu. Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE TVCG, 2023. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 278, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 278, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 278, + 553, + 314 + ], + "type": "text", + "content": "[41] Carsten Moenning and Neil A Dodgson. Fast marching farthest point sampling. Technical report, University of Cambridge, ECCV Laboratory, 2003. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 361 + ], + "type": "text", + "content": "[42] Bongki Moon, Hosagrahar V Jagadish, Christos Faloutsos, and Joel H. Saltz. Analysis of the clustering properties of the hilbert space-filling curve. IEEE TKDE, 13(1):124-141, 2001. 13" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 363, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 553, + 410 + ], + "type": "text", + "content": "[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 5, 16" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 412, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 471 + ], + "type": "text", + "content": "[44] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In CVPR, pages 5607-5612, 2023. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 472, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 518 + ], + "type": "text", + "content": "[45] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 12" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 520, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 578 + ], + "type": "text", + "content": "[46] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. Shapellm: Universal 3d object understanding for embodied interaction. In ECCV, pages 214-238, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 581, + 553, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 627 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 627 + ], + "type": "text", + "content": "[47] Shi Qiu, Saeed Anwar, and Nick Barnes. Semantic segmentation for real point cloud scenes via bilateral augmentation and adaptive fusion. In CVPR, pages 1757-1767, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 629, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 629, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 629, + 553, + 677 + ], + "type": "text", + "content": "[48] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 678, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 678, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 678, + 553, + 714 + ], + "type": "text", + "content": "[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "text", + "content": "14378" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 78, + 72, + 294, + 107 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 294, + 107 + ], + "type": "text", + "content": "Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In ICML, pages 29441-29454. PMLR, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 109, + 294, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 109, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 109, + 294, + 133 + ], + "type": "text", + "content": "[50] Hans Sagan. Hilbert's space-filling curve. In Space-filling curves, pages 9-30. Springer, 1994. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 134, + 294, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 134, + 294, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 294, + 169 + ], + "type": "text", + "content": "[51] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In ECCV, pages 444-462. Springer, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 170, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 294, + 240 + ], + "type": "text", + "content": "[52] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. NeurIPS, 33: 7537-7547, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 289 + ], + "type": "text", + "content": "[53] Yuan Tang, Xu Han, Xianzhi Li, Qiao Yu, Yixue Hao, Long Hu, and Min Chen. Minigpt-3d: Efficiently aligning 3d point clouds with large language models using 2d priors. ACM'MM, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 291, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 294, + 361 + ], + "type": "text", + "content": "[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.1, 17" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 363, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 294, + 409 + ], + "type": "text", + "content": "[55] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 5, 16" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 458 + ], + "type": "text", + "content": "[56] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022.2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 494 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 494 + ], + "type": "text", + "content": "[57] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. ACM TOG, 42(4):1-11, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 496, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 496, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 496, + 294, + 518 + ], + "type": "text", + "content": "[58] Qing Wang et al. Divergence-augmented policy optimization. NeurIPS, 32, 2019. 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 520, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 568 + ], + "type": "text", + "content": "[59] Zehan Wang, Haifeng Huang, Yang Zhao, Zhang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 628 + ], + "type": "text", + "content": "[60] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 1, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 630, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 294, + 677 + ], + "type": "text", + "content": "[61] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 678, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 678, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 678, + 294, + 713 + ], + "type": "text", + "content": "[62] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 498 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 336, + 72, + 553, + 96 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 336, + 72, + 553, + 96 + ], + "type": "text", + "content": "perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 98, + 553, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 144 + ], + "type": "text", + "content": "[63] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal attention for long-range interactions in vision transformers. NeurIPS, 34:30008-30022, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 146, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 146, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 146, + 553, + 205 + ], + "type": "text", + "content": "[64] Jianing Yang, Xuweiyi Chen, Nikhil Madaan, Madhavan Iyengar, Shengyi Qian, David F. Fouhey, and Joyce Chai. 3d-grand: A million-scale dataset for 3d-llms with better grounding and less hallucination. arXiv preprint 2406.05132, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 207, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 243 + ], + "type": "text", + "content": "[65] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. arXiv preprint arXiv:2112.08359, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 244, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 244, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 244, + 553, + 303 + ], + "type": "text", + "content": "[66] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 5, 17" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 304, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 363 + ], + "type": "text", + "content": "[67] Lichen Zhao, Daigang Cai, Jing Zhang, Lu Sheng, Dong Xu, Rui Zheng, Yinjie Zhao, Lipeng Wang, and Xibo Fan. Toward explainable 3d grounded visual question answering: A new benchmark and strong baseline. IEEE TCSVT, 33(6):2935-2949, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 401 + ], + "type": "text", + "content": "[68] Yufeng Zhong, Long Xu, Jiebo Luo, and Lin Ma. Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925, 2022. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 403, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 403, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 403, + 553, + 449 + ], + "type": "text", + "content": "[69] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 451, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 451, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 451, + 553, + 498 + ], + "type": "text", + "content": "[70] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In ICCV, pages 2911-2921, 2023. 6, 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 294, + 751, + 318, + 761 + ], + "type": "text", + "content": "14379" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/b34e32e0-849d-4c9e-8711-6ce6de1dc89c_content_list.json b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/b34e32e0-849d-4c9e-8711-6ce6de1dc89c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ff3c37cd6208e108d7e4f346278885836350ee49 --- /dev/null +++ b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/b34e32e0-849d-4c9e-8711-6ce6de1dc89c_content_list.json @@ -0,0 +1,1861 @@ +[ + { + "type": "text", + "text": "Percept, Memory, and Imagine: World Feature Simulating for Open-Domain Unknown Object Detection", + "text_level": 1, + "bbox": [ + 186, + 130, + 816, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Among Wu, Cheng Deng*", + "bbox": [ + 392, + 204, + 611, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Electronic Engineering, Xidian University, Xi'an, China", + "bbox": [ + 235, + 220, + 759, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "amwu@xidian.edu.cn, chdeng@mail.xidian.edu.cn", + "bbox": [ + 295, + 241, + 696, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To accelerate the safe deployment of object detectors, we focus on reducing the impact of both covariate and semantic shifts. And we consider a realistic yet challenging scenario, namely Open-Domain Unknown Object Detection (ODU-OD), which aims to detect unknown objects in unseen target domains without accessing any auxiliary data. Towards ODU-OD, it is feasible to learn a robust discriminative boundary by synthesizing virtual features. Generally, perception, memory, and imagination are three essential capacities for human beings. Through multi-level perception and rich memory about known objects, the characteristics of unknown objects can be imagined sufficiently, enhancing the ability of discriminating known from unknown objects. Inspired by this idea, an approach of World Feature Simulation (WFS) is proposed, mainly consisting of a multi-level perception, memory recorder, and unknown-feature generator. Specifically, after extracting the features of the input, we separately employ a Mamba and Graph Network to obtain the global-level and connective-level representations. Next, a codebook containing multiple learnable codewords is defined to preserve fragmented memory of known objects. Meanwhile, we perform a modulated operation on the memory to form the imagination bank involving unknown characteristics. Finally, to alleviate the impact of lacking supervision data, based on the multi-level representation and imagination bank, a dedicated unknown-feature generator is designed to recurrently synthesize outlier features deviating from in-distribution (ID) objects. The significant performance gains on four different detection tasks demonstrate the superiorities of our method. The code will be released at https://github.com/AmingWu/WFS.", + "bbox": [ + 91, + 323, + 483, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 820, + 220, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rejuvenation of deep neural networks, object detection has achieved many advances [1, 14, 31-33] recently.", + "bbox": [ + 89, + 845, + 482, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Though these advances have obtained impressive performance, in real-world scenarios, they inevitably encounter both covariate shifts (e.g., test-time changes in image styles) and semantic shifts (e.g., test-time unknown classes), which limits their applications.", + "bbox": [ + 511, + 292, + 906, + 368 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since there is no auxiliary data available, to alleviate the covariate-shift impact, existing researches [25, 41, 43, 48] mainly pay attention to enhancing the generalization ability by learning domain-invariant features. As for the semantic-shift challenge, most existing approaches [4, 6, 26, 42, 44] explore synthesizing diverse virtual features to strengthen the capability of discriminating unknown objects. Although these methods have been demonstrated to be effective, they are only concerned with a single-shift case and could not cope with a more difficult scenario involving both covariate and semantic shifts, e.g., for an autonomous car, the detector may encounter an unknown object in unseen adverse weather, resulting in a safety risk.", + "bbox": [ + 509, + 369, + 906, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To promote the safe deployment of detectors, we propose a realistic yet challenging task, namely Open-Domain Unknown Object Detection (ODU-OD). As shown in Fig. 1, given one source dataset including multiple pre-defined categories for training, the goal of ODU-OD is to accurately detect unknown out-of-distribution (OOD) objects in multiple unseen target domains with different styles. Since ODU-OD contains both covariate and semantic shifts and fails to access any auxiliary data for training, existing single-shift methods can not be directly utilized to tackle this task.", + "bbox": [ + 511, + 566, + 908, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In essence, towards covariate and semantic shifts, it is critical to learn the features that lead to a robust discriminative boundary. Particularly, enhancing the robustness could attenuate the covariate-shift impact and improve the generalization ability. And strengthening the discrimination is instrumental in accurately recognizing known and unknown objects. Moreover, recent researches [8-10, 35, 36] indicate that based on the current observations and formed memory, human beings could produce rich imaginations, improving the adaptation ability for unknown situations. Therefore, we explore simulating humans to leverage plentiful perception and fragmented memory to imagine unknowns for training", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 886, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4682", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f4352d79a6183dde2dbda3c7fff3f56d4ad25c0021c7392702346c3524c001bf.jpg", + "image_caption": [ + "Figure 1. World Feature Simulation for Open-Domain Unknown Object Detection, which imitates humans to synthesize virtual unknown features and consists of a multi-level perception, memory recorder, and dedicated unknown-feature generator. We first calculate the multi-level representation from the local, global, and connective levels. Then, a memory codebook is constructed to preserve the local object information (as shown in pink boxes). Next, a modulated operation is performed on the fragmented memory to form the imagination bank involving unknown characteristics. Finally, we design an unknown-feature generator to recurrently synthesize virtual unknown features deviating from known objects, enhancing the ability of detecting unknown objects within unseen domains (as shown in green boxes)." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 416, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c11afc1c9c3d4294cc580a1ed381ba73e832eaddc4238b1e1cc4be4311e5f1a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 88, + 637, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/515fd2a86094b91420fcf94427c11f7dd34407f231242e8de3f80d4eabcf53c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 88, + 905, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(as shown in Fig. 1), enhancing the ability of distinguishing unknown objects within unseen domains.", + "bbox": [ + 89, + 356, + 482, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, as shown in Fig. 1, a method of World Feature Simulation (WFS) is proposed, mainly consisting of a multi-level perception, memory recorder, and unknown-feature generator. For a specific object, in order to deepen its perception, humans usually observe the object from multiple different perspectives, e.g., local and global levels. To this end, given an ID input, we first utilize a convolutional backbone [13] to extract its local features. Then, a Mamba [7, 12, 49] and a Graph network [19, 38] are separately performed on the local features to obtain the global and connective features. And the local, global, and connective features are concatenated to form the multi-level representation. Next, we define a codebook containing multiple learnable codewords to preserve fragmented memory of the known objects. By fusing the memory into the current multi-level representation, the robustness against covariateshift risk could be enhanced effectively.", + "bbox": [ + 88, + 386, + 482, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Fig. 1, to achieve the ability of imagining unknowns, a modulated operation is performed on the known codebook to obtain the imagination bank. And we design a contrastive constraint to promote the bank to involve plentiful unknown characteristics. Finally, a dedicated unknown-feature generator is proposed to leverage the multi-level representation and imagined unknown characteristics to recurrently synthesize virtual unknown features deviating from ID inputs, which is conducive to learning a robust discriminative boundary between known and unknown objects within unseen domains. Experimental results on four different detection tasks demonstrate the superiorities of our method.", + "bbox": [ + 89, + 643, + 482, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions are summarized as follows:", + "bbox": [ + 109, + 825, + 413, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) To promote the deployment of detectors, we propose a new task, namely Open-Domain Unknown Object Detection (ODU-OD), aiming to detect unknown objects in unseen target domains without accessing any auxiliary data.", + "bbox": [ + 89, + 840, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) To tackle ODU-OD involving both covariate and semantic shifts, we imitate humans to explore a new framework of World Feature Simulation, which aims to synthesize virtual unknown features for training and contains a multi-level perception, fragmented memory recorder, and unknown-feature generator.", + "(3) We build an Open-Domain Dataset for unknown object detection to verify our method WFS, consisting of three different styles [20], i.e., 'Acrylic', 'Purple-Brush', and 'Sketch'. Meanwhile, our method is further evaluated on three single-shift tasks, i.e., OOD-OD [6], open-set object detection [11], and Single-DGOD [41]. The significant performance gains over multiple baselines demonstrate the effectiveness of our method. Meanwhile, visualization analyses also show that the built memory and reasoned imagination could indeed capture fragmented visual content." + ], + "bbox": [ + 511, + 356, + 908, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 616, + 653, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Covariate-shift Detection. To improve the generalization of detectors, a task of Single-DGOD [41] is proposed, aiming to learn an object detector that performs well on multiple unseen target domains with only one source domain for training. Some works [23, 39, 41] utilize feature disentanglement and large-scale models to extract domain-invariant representations. Moreover, Liu et al. [26] further consider the biased problem and propose a causal-based method to learn unbiased generalized representations. And Danish et al. [3] explore diversifying the source domain and aligning detections based on prediction confidence. These methods have been verified on multiple benchmarks. However, they only consider the covariate-shift scenario and could not detect unknown objects.", + "bbox": [ + 511, + 642, + 906, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic-shift Detection. Recently, a task of unsupervised OOD-OD [6] is proposed, whose goal is to detect unknown objects without relying on any OOD data. To alle", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4683", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fcd8e9bf4f90cde9792c69866df3a77d2619c4df5ccb5257398cac02da8df11c.jpg", + "image_caption": [ + "Figure 2. The details of World Feature Simulation for ODU-OD. To enhance the robustness against covariate shift, a multi-level perception is used to extract the representation $F_{v}$ from the local, global, and connective levels. Meanwhile, we design a memory recorder to preserve rich content about known objects. By calculating a memory-enhanced map, the robustness could be further strengthened. Next, through a modulated operation, we can reason the imagination bank involving plentiful unknown characteristics. Finally, a dedicated unknown-feature generator is proposed to utilize the built memory and imagination characteristics to recurrently synthesize virtual unknown features, improving the ability of detecting unknown objects within unseen domains." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 906, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "viate the impact of lacking OOD data for supervision, VOS [6] and DFDD [44] explore synthesizing virtual OOD features that slightly deviate from the distribution of ID data. Besides, the work [5] further learns unknown-aware knowledge from external videos, which does not match the setting of unsupervised OOD-OD. Finally, Du et al. [4] present to use a distance-based mechanism to shape the learned representations. Meanwhile, Wilson et al. [40] attempt to enhance the sensitivity of the learned features, which is beneficial for improving the performance of detecting OOD objects. The work [26] struggles to leverage generative models trained on large-scale open-set data to synthesize OOD samples. Although these methods have been demonstrated to be effective, they only consider the semantic-shift problem and ignore the covariate-shift impact, which limits their applications to unknown objects within unseen domains.", + "bbox": [ + 88, + 426, + 480, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from the above methods, in this paper, we explore a more difficult problem involving both covariate and semantic shifts, i.e., Open-Domain Unknown Object Detection. And a new approach of World Feature Simulation is proposed for synthesizing unknown features for training, consisting of a multi-level perception, fragmented memory recorder, and unknown-feature generator. Experimental results on multiple datasets indicate its superiorities.", + "bbox": [ + 88, + 670, + 482, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. World Feature Simulation", + "text_level": 1, + "bbox": [ + 89, + 811, + 334, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For ODU-OD, during training, we can only access single source-domain data with some pre-defined categories. During test, the detector is required to accurately localize and recognize unknown objects within multiple unseen do", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mains. To reduce the entangled impact of covariate and semantic shifts, as shown in Fig. 2, we design a World Feature Simulation for synthesizing unknown features.", + "bbox": [ + 511, + 426, + 906, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Observing Objects from Multiple Levels", + "text_level": 1, + "bbox": [ + 511, + 479, + 859, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generally, in order to deepen the perception of a specific object, human beings usually observe the object from multiple different perspectives, which is beneficial for constructing a complete impression. Therefore, in Fig. 2, we explore imitating this capability based on local, global, and connective levels. Specifically, we exploit the widely used object detector, i.e., Faster R-CNN [14, 33], as the basic model.", + "bbox": [ + 511, + 501, + 905, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an ID input image, a convolutional network, e.g., ResNet [13], is taken as the backbone to extract the local feature map $F_{l} \\in \\mathbb{R}^{w \\times h \\times c}$ , where $w$ , $h$ , and $c$ separately represent width, height, and the number of channels. Each element in $F_{l}$ describes a local region of the input. Besides the local level, human beings usually intensify their impression by global observation and discovering the relations between local regions. Hence, we first unfold $F_{l}$ to a sequence of visual elements, i.e., $D = \\{d_{i} \\in \\mathbb{R}^{c}, i = 1, \\dots, wh\\}$ . Then, one-layer vision mamba [49] is used to process $D$ , which outputs $H = \\{h_{i} \\in \\mathbb{R}^{c}, i = 1, \\dots, wh\\}$ . Finally, $H$ is reshaped to the original size to obtain the global representation $F_{g} \\in \\mathbb{R}^{w \\times h \\times c}$ [12]. Next, we use one-layer graph convolution [19] to further capture the semantic connections between local regions. Concretely, we first reshape $F_{l}$ to $\\tilde{F} \\in \\mathbb{R}^{wh \\times c}$ . Then, an adjacency matrix for an undirected graph is defined as $\\mathbf{A} = \\text{softmax}(\\tilde{F}\\tilde{F}^{\\top}) + I_{wh}$ , where $I_{wh}$ indicates the identity matrix:", + "bbox": [ + 511, + 607, + 906, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} = \\mathbf {A} \\tilde {F}, \\quad F _ {c} = r e l u (W _ {c} * \\tilde {\\mathcal {R}} + b _ {c}), \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 883, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4684", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathcal{R}}\\in \\mathbb{R}^{w\\times h\\times c}$ denotes the reshaped result of $\\mathcal{R}$ . $W_{c}\\in \\mathbb{R}^{3\\times 3\\times c\\times c}$ and $b_{c}\\in \\mathbb{R}^{c}$ indicate the trainable parameters. $F_{c}\\in \\mathbb{R}^{w\\times h\\times c}$ represents the connective feature. Finally, $F_{l},F_{g}$ , and $F_{c}$ are concatenated as the multi-level representation $F_{v}\\in \\mathbb{R}^{w\\times h\\times c}$ , where $F_{v} = \\phi ([F_{l},F_{g},F_{c}])$ and $\\phi (\\cdot)\\in \\mathbb{R}^{1\\times 1\\times 3c\\times c}$ is utilized to transform the number of channels. Compared with $F_{l},F_{v}$ involves sufficient object-related information from three levels, which strengthens the robustness against the covariate shift.", + "bbox": [ + 89, + 89, + 483, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Mining Fragmented Memory and Imagination", + "text_level": 1, + "bbox": [ + 89, + 242, + 482, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, human beings could exploit familiar fragmented memory to improve the robustness of recognizing known objects. Then, based on the memory, human beings could produce rich imagination, enhancing the ability of discriminating unknowns. To simulate this capability, we present a memory recorder [46] to preserve the information of known objects. Particularly, a learnable memory codebook is denoted as $M = \\{m_j \\in \\mathbb{R}^c, j = 1, \\dots, Q\\}$ that contains $Q$ codewords. Meanwhile, by means of a weight $a_{ij}$ , each element $f_i \\in \\mathbb{R}^c$ of the multi-level representation $F_v$ can be assigned to each codeword $m_j$ . And the residual vector is represented as $r_{ij} = f_i - m_j$ , where $i = 1, \\dots, wh$ . Here, the residual operation is conducive to promoting the memory codewords to effectively capture the corresponding content of local visual regions. The processes of capturing memory are shown as follows:", + "bbox": [ + 89, + 267, + 483, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {j} = \\sum_ {i = 1} ^ {w h} a _ {i j} r _ {i j}, \\quad a _ {i j} = \\frac {\\exp \\left(- k _ {j} \\| r _ {i j} \\| ^ {2}\\right)}{\\sum_ {j = 1} ^ {Q} \\exp \\left(- k _ {j} \\| r _ {i j} \\| ^ {2}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 525, + 482, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k_{j}$ represents the learnable smoothing factor for the codeword $m_j$ . Finally, $\\mathcal{V} \\in \\mathbb{R}^{Q \\times c}$ is the output result. The concatenation of $F_v$ and $\\mathcal{V}$ is used to calculate the memory-enhanced map $E \\in \\mathbb{R}^{w \\times h \\times c}$ , i.e., $E = \\psi([F_v, \\tilde{\\mathcal{V}}W_0 + b_0])$ , where $\\tilde{\\mathcal{V}} \\in \\mathbb{R}^{1 \\times Qc}$ is the reshaped result of $\\mathcal{V}$ . Meanwhile, $W_0 \\in \\mathbb{R}^{Qc \\times c}$ and $b_0 \\in \\mathbb{R}^c$ are parameters of a fully-connected layer. $\\psi(\\cdot)$ includes two convolutional layers to transform the concatenated result. Through fusing memory, the information capacity of $E$ could be increased, improving the robustness of object localization.", + "bbox": [ + 89, + 580, + 483, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we further explore how to leverage the memory to reason the imagination that guides the synthesis of virtual unknown features. Particularly, we assume that the imagination should satisfy two conditions: (1) Imagination and memory should belong to different spaces. (2) Imagination mainly describes unknowns. It could not well represent known visual content. To this end, based on the memory $M \\in \\mathbb{R}^{Q \\times c}$ , a modulated operation is performed on $M$ to obtain the imagination bank $\\mathcal{I} \\in \\mathbb{R}^{Q \\times c}$ , i.e., $\\mathcal{I} = \\gamma \\odot M + \\beta$ , where $\\gamma$ and $\\beta$ are learnable parameters for channel-wisely scaling and shifting $M$ . Next, we separately utilize $M$ and", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/19462839cd050a9b5595f959dfc073517ed97763d6518ff4176c912636cbd353.jpg", + "image_caption": [ + "Figure 3. Unknown-feature generator. Based on $F_{m} \\in \\mathbb{R}^{w \\times h \\times c}$ , a variational operation is performed on $F_{m}$ to obtain $Z \\in \\mathbb{R}^{w \\times h \\times c}$ . Then, we exploit the reasoned imagination bank to re-represent $Z$ , which outputs $G \\in \\mathbb{R}^{w \\times h \\times c}$ . Next, $Z$ and $G$ are unfolded into a sequence with $wh$ elements. Finally, a dedicated recurrent generator including three inputs and one output is presented to synthesize virtual unknown features, improving the discrimination." + ], + "image_footnote": [], + "bbox": [ + 524, + 88, + 893, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{I}$ to re-represent $F_{v}$ :", + "bbox": [ + 513, + 340, + 653, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {m} = \\operatorname {s o f t} x \\left(\\tilde {F} _ {v} M ^ {\\top}\\right) M, F _ {i} = \\operatorname {s o f t} x \\left(\\tilde {F} _ {v} \\mathcal {I} ^ {\\top}\\right) \\mathcal {I}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 361, + 906, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where 'softx' denotes the softmax operation. $\\tilde{F}_v\\in \\mathbb{R}^{wh\\times c}$ is the reshaped result of $F_{v}$ . $F_{m}\\in \\mathbb{R}^{w\\times h\\times c}$ and $F_{i}\\in \\mathbb{R}^{w\\times h\\times c}$ indicate the re-represented results based on the fragmented memory and imagination. Finally, a contrastive loss is exploited to enlarge the gap between the memory recorder and imagination bank:", + "bbox": [ + 511, + 386, + 905, + 477 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i m} = - \\log \\frac {\\exp \\left(\\sin \\left(F _ {m} , F _ {v}\\right) / \\tau\\right)}{\\sum_ {j = 0} ^ {1} \\exp \\left(\\sin \\left(B [ j ] , F _ {v}\\right) / \\tau\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 483, + 906, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B = [F_m, F_i]$ . $\\mathrm{sim}(F_m, F_v)$ denotes the average of the cosine similarity between all corresponding elements of feature map $F_m$ and $F_v$ . $\\tau$ is a hyper-parameter, which is set to 1.0 in the experiments. Obviously, by minimizing the contrastive loss, the distance between $F_m$ and $F_v$ could be narrowed, promoting the memory recorder to capture sufficient content about known objects. Instead, the distance between $F_i$ and $F_v$ is enlarged, forcing the reasoned imagination bank to contain plentiful information deviating from knowns. Next, the imagination bank is utilized to generate expected unknown features.", + "bbox": [ + 511, + 526, + 906, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Generating Unknown Features for Training", + "text_level": 1, + "bbox": [ + 511, + 700, + 883, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For ODU-OD, to attenuate the impact of lacking unknown data for supervision, we design an unknown-feature generator to synthesize outlier features involving unknown characteristics. Specifically, two convolutional networks $W_{\\mu}$ and $W_{V}$ are performed on $F_{m}$ to estimate the corresponding means $\\mu \\in \\mathbb{R}^{w\\times h\\times c}$ and variances $V \\in \\mathbb{R}^{w\\times h\\times c}$ . Then, an encoding operation is performed to obtain $Z \\in \\mathbb{R}^{w\\times h\\times c}$ . Meanwhile, we exploit the reasoned imagination bank $\\mathcal{I}$ to re-represent $Z$ . The processes are shown as follows:", + "bbox": [ + 511, + 722, + 905, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu = W _ {\\mu} * F _ {m}, \\quad V = W _ {V} * F _ {m}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 864, + 903, + 890 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ = \\mu + \\epsilon \\odot V, \\quad G = \\text {s o f t m a x} (\\tilde {Z} \\mathcal {I} ^ {\\top}) \\mathcal {I},\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 885, + 851, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4685", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{Z} \\in \\mathbb{R}^{wh \\times c}$ is the reshaped result of $Z$ . $\\epsilon$ represents Gaussian noise sampled from $\\mathcal{N}(0, I)$ . $G \\in \\mathbb{R}^{w \\times h \\times c}$ is the re-represented result based on the imagination bank $\\mathcal{I}$ . Since the imagination could not well express the knowns, $G$ contains plentiful information deviating from known visual content, which is instrumental in promoting the generation of virtual unknown features.", + "bbox": [ + 89, + 89, + 480, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we present a dedicated generator to recurrently synthesize expected unknown features. As shown in Fig. 3, the generator contains three inputs and one output. Particularly, at each time, the generator takes $Z_{t}$ , $G_{t}$ , and the previous output $S_{t-1}$ as the input, and synthesizes the current output $S_{t}$ . The generating processes are shown as follows:", + "bbox": [ + 89, + 196, + 480, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {t} = \\sigma \\left(W _ {z z} * Z _ {t} + W _ {o z} * G _ {t} + W _ {s z} * S _ {t - 1} + b _ {z}\\right), \\\\ g _ {t} = \\sigma \\left(W _ {z g} * Z _ {t} + W _ {o g} * G _ {t} + W _ {s g} * S _ {t - 1} + b _ {g}\\right), \\\\ \\tilde {S} _ {t} = \\operatorname {T a n h} \\left(W _ {s} * Z _ {t} + W _ {h} * \\left(g _ {t} \\odot S _ {t - 1} + (1 - g _ {t}) \\odot G _ {t}\\right)\\right), \\\\ S _ {t} = \\left(1 - z _ {t}\\right) \\odot S _ {t - 1} + z _ {t} \\odot \\hat {S} _ {t}, \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 299, + 490, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Z_{t}\\in \\mathbb{R}^{c}$ and $G_{t}\\in \\mathbb{R}^{c}$ are the current element of $Z$ and $G$ , $t = 1,\\dots ,wh$ . And $W_{zz}$ , $W_{oz}$ , $W_{sz}$ , $b_{z}$ , $W_{zg}$ , $W_{og}$ , $W_{sg}$ , $b_{g}$ , $W_{s}$ , and $W_{h}$ represent the trainable parameters. Through this generating mechanism, the synthesized features could sufficiently retain the information in previous steps, improving the content continuity of the generated features. Finally, the generated sequence $\\{S_1,S_2,\\dots S_{wh}\\}$ are concatenated to obtain unknown feature $S_{u}\\in \\mathbb{R}^{w\\times h\\times c}$ . We still utilize a contrastive loss to expand the difference between known and the synthesized unknown features:", + "bbox": [ + 89, + 402, + 483, + 554 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u n k} = - \\log \\frac {\\exp \\left(\\sin \\left(F _ {m} , F _ {v}\\right) / \\tau\\right)}{\\sum_ {j = 0} ^ {1} \\exp \\left(\\sin \\left(A [ j ] , F _ {v}\\right) / \\tau\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 565, + 480, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A = [F_m, S_u]$ . $\\tau$ is set to 1.0 in the experiments. Since the generation process involves perception, memory, and imagination, using the contrastive loss could not only promote $S_u$ to contain plentiful unknown information but also facilitate $M$ and $\\mathcal{I}$ to separately capture rich known and unknown characteristics.", + "bbox": [ + 89, + 612, + 480, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. WFS-Driven Unknown-Object Detection", + "text_level": 1, + "bbox": [ + 89, + 712, + 437, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, the memory-enhanced map $E$ is taken as the input of Region-Proposal Network (RPN) [14, 33] to obtain a series of object proposals $\\mathcal{O}$ . Meanwhile, based on $\\mathcal{O}$ , RoI-Alignment followed by RoI-Feature extraction [14] is performed on $E$ to obtain $O_{\\mathrm{in}} \\in \\mathbb{R}^{p \\times n}$ , where $p$ and $n$ denote the number of proposals and channels, respectively. Then, $O_{\\mathrm{in}}$ is input to the object classifier and regressor to calculate the classification loss $\\mathcal{L}_{\\mathrm{cls}}$ and localization loss $\\mathcal{L}_{\\mathrm{loc}}$ , i.e., $\\mathcal{L}_{\\mathrm{det}} = \\mathcal{L}_{\\mathrm{cls}} + \\mathcal{L}_{\\mathrm{loc}}$ . By optimizing $\\mathcal{L}_{\\mathrm{det}}$ , the memory recorder could be promoted to capture plentiful information about known objects, enhancing the robustness.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we further enlarge the gap between the known and synthesized unknown features from the instance perspective. Concretely, based on $\\mathcal{O}$ , RoI-Alignment followed by RoI-Feature extraction is performed on $S_{u}$ to obtain $O_{\\mathrm{unk}} \\in \\mathbb{R}^{p \\times n}$ . Meanwhile, we define a loss function $\\mathcal{L}_{\\mathrm{dis}}$ to maximize the $KL$ -divergence between $O_{\\mathrm{unk}}$ and $O_{\\mathrm{in}}$ , i.e., $\\mathcal{L}_{\\mathrm{dis}} = \\mathrm{KL}[q(O_{\\mathrm{unk}}), q(O_{\\mathrm{in}})]$ , where $q(\\cdot)$ represents the probability distribution. Besides, to attain the goal of distinguishing unknown objects from known ID objects, $O_{\\mathrm{unk}}$ and $O_{\\mathrm{in}}$ are used to compute an uncertainty loss [6], regularizing the detector to produce a low score for the ID object features, and a high score for the virtual unknown features:", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {u n t y}} = \\mathbb {E} _ {x \\sim O _ {\\text {i n}}} [ - \\log \\frac {\\exp^ {- \\mathbf {E} (x)}}{1 + \\exp^ {- \\mathbf {E} (x)}} ] + \\tag {8} \\\\ \\mathbb {E} _ {v \\sim O _ {\\mathrm {u n k}}} [ - \\log \\frac {1}{1 + \\exp^ {- \\mathbf {E} (v)} ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 282, + 903, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{E}(\\cdot)$ is the object-level energy score [6, 27, 44]. The overall objective is shown as follows:", + "bbox": [ + 511, + 358, + 903, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t r a s t}} = \\mathcal {L} _ {i m} + \\mathcal {L} _ {u n k} - \\mathcal {L} _ {\\text {d i s}}, \\\\ \\mathcal {L} _ {i m} = \\mathcal {L} _ {u n k} + \\mathcal {L} _ {u n k} + \\lambda_ {i} \\mathcal {L} _ {i m}. \\end{array} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 401, + 901, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e t}} + \\alpha \\cdot \\mathcal {L} _ {\\text {c o n t r a s t}} + \\lambda \\cdot \\mathcal {L} _ {\\text {u n t y}},\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 420, + 834, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\lambda$ are two hyper-parameters, which are set to 0.001 and 0.1 in the experiments.", + "bbox": [ + 511, + 444, + 903, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Inference for Unknown Object Detection", + "text_level": 1, + "bbox": [ + 511, + 483, + 861, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During test, we only calculate the uncertainty score for unknown object detection [6, 44]. Specifically, for a predicted bounding box $\\mathbf{b}$ , the processes of discriminating unknown objects are shown as follows:", + "bbox": [ + 511, + 506, + 903, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} = \\frac {\\exp^ {- \\mathbf {E} (\\mathbf {b})}}{1 + \\exp^ {- \\mathbf {E} (\\mathbf {b})}}, \\quad \\mathcal {C} (\\mathbf {b}) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} \\mathcal {E} < \\delta , \\\\ 1 & \\text {i f} \\mathcal {E} \\geq \\delta . \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 577, + 903, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the output of the classifier $\\mathcal{C}(\\cdot)$ , we utilize the threshold mechanism [6] to distinguish known objects (the result is 1) from unknown objects (the result is 0). The threshold $\\delta$ is commonly set to 0.95 so that a high fraction of known ID objects are correctly classified. In supplementary material, we will show more details of the proposed WFS algorithm and give more discussion.", + "bbox": [ + 511, + 619, + 906, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 739, + 643, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the experiments, our method is evaluated on four different detection tasks, i.e., ODU-OD, OOD-OD [6], open-set object detection (OSOD) [11], and Single-DGOD [41]. Particularly, we construct an Open-Domain dataset for ODU-OD, which utilizes Clipstyler [20] to transform testing data of OOD-OD [6] to three different styles, i.e., 'Acrylic', 'Purple-Brush', and 'Sketch'.", + "bbox": [ + 511, + 763, + 903, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We utilize Faster R-CNN [33] with RoI-Alignment layer [14] as the basic detection model.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4686", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/74050db151c9a4a4dfb833105c194156f09060df50c8ff369162b728ab436b6d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Source DataMethodStyle: Acrylic\nMS-COCO/OpenImagesStyle: Purple-Brush\nMS-COCO/OpenImagesStyle: Sketch\nMS-COCO/OpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
VOCSR-VAE [42]88.25/84.6371.02/69.6487.76/69.9367.42/79.1889.67/74.8266.47/77.04
DFDD [44]87.68/85.1270.85/69.2187.24/70.0867.23/79.4689.83/74.9266.13/76.82
VOS [6]88.73/84.7171.13/69.5688.66/70.1767.06/79.3389.76/74.6166.50/77.63
VOS + WFS80.16/77.5274.31/73.8580.07/60.6870.86/83.8782.16/67.2369.87/81.28
SAFE [40]86.62/47.8363.47/78.5186.31/39.6261.28/86.8287.56/55.2860.71/83.48
SAFE + WFS79.38/40.1667.28/81.6979.23/33.8566.47/88.7380.49/48.6165.82/84.17
BDDSR-VAE [42]64.72/53.9872.34/78.1166.43/57.2870.69/76.2469.36/59.2769.17/75.25
DFDD [44]64.08/53.2672.48/78.3665.91/56.7771.03/76.4868.74/58.8969.22/75.44
VOS [6]65.71/54.7371.96/77.8068.50/58.7970.18/75.8370.80/60.7169.05/74.97
VOS + WFS56.48/45.4275.87/82.1859.64/47.8874.34/79.8761.01/51.9473.78/79.36
SAFE [40]54.82/40.1873.71/82.6757.63/43.9772.34/80.8359.44/46.2871.27/79.53
SAFE + WFS45.91/32.0378.35/85.1447.78/34.4176.93/84.2650.26/37.1875.54/82.75
", + "bbox": [ + 155, + 88, + 841, + 335 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2104bfb640219412c1187adf19fa6ca44e7061d6ed9c49b0e70e2bd1d56118db.jpg", + "table_caption": [ + "Table 1. Unknown-object detection results under unseen open domains. $\\uparrow$ denotes larger values are better and $\\downarrow$ represents smaller values are better. 'VOS+WFS' and 'SAFE+WFS' indicate our method is directly plugged into VOS [6] and SAFE [40]. We can see that our method could reduce the impact of both covariate and semantic shifts." + ], + "table_footnote": [], + "table_body": "
SourceMethodFPR95 ↓AUROC ↑
OOD: MS-COCO / OpenImages
VOCMSP [15]70.99 / 73.1383.45 / 81.91
ODIN [24]59.82 / 63.1482.20 / 82.59
Mahalanobis [22]67.73 / 65.4181.45 / 81.48
Gram matrices [34]62.75 / 67.4279.88 / 77.62
Energy score [27]56.89 / 58.6983.69 / 82.98
Generalized ODIN [16]59.57 / 70.2883.12 / 79.23
CSI [37]59.91 / 57.4181.83 / 82.95
GAN-synthesis [21]60.93 / 59.9783.67 / 82.67
SIREN-vMF [4]64.68 / 68.5385.36 / 82.78
SIREN-KNN [4]47.45 / 50.3889.67 / 88.80
SR-VAE [42]42.17 / 46.2690.28 / 87.89
DFDD [44]41.34 / 44.5290.79 / 88.65
VOS (Baseline) [6]47.53 / 51.3388.70 / 85.23
VOS + WFS37.91 / 41.3691.47 / 89.78
SAFE (Baseline) [40]47.40 / 20.0680.30 / 92.28
SAFE + WFS37.48 / 11.2785.62 / 94.91
BDDMSP [15]80.94 / 79.0475.87 / 77.38
ODIN [24]62.85 / 58.9274.44 / 76.61
Mahalanobis [22]55.74 / 47.6985.71 / 88.05
Gram matrices [34]60.93 / 77.5574.93 / 59.38
Energy score [27]60.06 / 54.9777.48 / 79.60
Generalized ODIN [16]57.27 / 50.1785.22 / 87.18
CSI [37]47.10 / 37.0684.09 / 87.99
GAN-synthesis [21]57.03 / 50.6178.82 / 81.25
SR-VAE [42]32.23 / 21.8190.69 / 93.55
DFDD [44]30.71 / 22.6790.74 / 92.48
VOS (Baseline) [6]44.27 / 35.5486.87 / 88.52
VOS + WFS28.66 / 20.3992.37 / 94.13
SAFE (Baseline) [40]32.56 / 16.0488.96 / 94.64
SAFE + WFS21.84 / 7.8393.41 / 96.85
", + "bbox": [ + 94, + 388, + 483, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The performance of unsupervised OOD-OD [6] that assumes the training and test data are from the same domain. We can see that our method can reduce the impact of semantic shift.", + "bbox": [ + 89, + 780, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ResNet-50 [13] is taken as the backbone. For the memory recorder, the number $Q$ of the codewords is set to 24. All the experiments are trained using the standard SGD optimizer with a learning rate of 0.02. In supplementary material, we", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "will give the details of the used datasets and metrics.", + "bbox": [ + 511, + 393, + 861, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Analysis of ODU-OD and OOD-OD", + "text_level": 1, + "bbox": [ + 511, + 419, + 823, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 and 2 separately illustrate the ODU-OD and OOD-OD performance. Here, we only compare with state-of-the-art methods. We can observe that though existing methods [6, 40, 42, 44] could effectively attenuate the semantic-shift impact, their performance drops severely when encountering the covariate shift (e.g., changes in image styles). Plugging our method into existing methods [6, 40] significantly improves their performance whether objects are within open or closed domains. This demonstrates that for the entangled case of covariate and semantic shifts, imitating humans to construct perception, memory, and imagination and further generating unknown features for training could simultaneously strengthen the robustness and discrimination.", + "bbox": [ + 511, + 441, + 906, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 4, we show some detection examples with different styles. We can observe that plugging our method could indeed attenuate both covariate and semantic shifts, improving the detection accuracy under open domains.", + "bbox": [ + 511, + 638, + 906, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Analysis of OSOD and Single-DGOD", + "text_level": 1, + "bbox": [ + 511, + 710, + 834, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further demonstrate that our method could attenuate the impact of two-shift entanglement, we verify our method on two dedicated tasks, i.e., OSOD [11] and Single-DGOD [41]. We directly plug our method into three existing methods [11, 39, 41] and do not change the training processes. Table 3 and 4 show the corresponding results. We can observe that plugging our method could improve the detection performance effectively. This further demonstrates that our method WFS could synthesize expected unknown features that are beneficial for alleviating the impact of covariate and semantic shifts, which promotes the detector to learn a ro", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4687", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7b2ff19fd9c9c63f255a5ed7337cb2a7593e25ac2bddc451c4cee1245b322c1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 90, + 297, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03ce489de9365809201dee26c311131efe519a14d8bad60b6af0b47afa397bd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 90, + 480, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f571e54a2fe6baebe687bc822f36b02e781405d37165c19195ca79a7d18a65d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 90, + 653, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d632e65771615928e9fdf297c954f38cb86f536d81123137e4962b804e91bba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 90, + 776, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14095d82bd384930a08417bccb037f9a52c615af9c56c5581a1d8dbe72e0db64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 90, + 898, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6df42ef3a9e1e6682df6dad44219fb311a9c850fe290af09d2e5a3effdbe5399.jpg", + "image_caption": [ + "(a) Acrylic" + ], + "image_footnote": [], + "bbox": [ + 96, + 189, + 297, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e0c16a0db04979a2994d5fa137b4eb25c74b4298a570a3c1c2461fd60c81be5.jpg", + "image_caption": [ + "(b) Purple-Brush" + ], + "image_footnote": [], + "bbox": [ + 303, + 189, + 480, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d95fdf15a5be65475dbf5e31bd3b476704c76723c1cc16a4607b1cdbd2f346d1.jpg", + "image_caption": [ + "(c) Sketch", + "Figure 4. Detection results on target images with multiple different styles. The first and second rows respectively indicate results based on VOS [6] and our method. The in-distribution source dataset is BDD-100k [45]. 'Arcylic', 'Purple-Brush', and 'Sketch' represent the rendered target styles [20]. 'Nature' is the same style as the source dataset." + ], + "image_footnote": [], + "bbox": [ + 483, + 189, + 653, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e075021f3e9080f07d6de26d0a6dc9f7571f718ecaeb057d3cf40e4b2948b008.jpg", + "image_caption": [ + "(d) Nature" + ], + "image_footnote": [], + "bbox": [ + 658, + 189, + 776, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fcaf5632b835e4236bd53303173211647c73603af2a8130f389be4f0c2f4fc18.jpg", + "image_caption": [ + "(e) Nature" + ], + "image_footnote": [], + "bbox": [ + 781, + 189, + 898, + 281 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5e317d68234065002d01ba2d47b37b8d1329ccc5c7c785bbeacf81875d8e9a88.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVOCVOC-COCO-20VOC-COCO-40VOC-COCO-60
\\( \\mathrm{mAP}_{\\mathcal{K}\\uparrow} \\)\\( WI_{\\downarrow} \\)\\( AOSE_{\\downarrow} \\)\\( mAP_{\\mathcal{K}\\uparrow} \\)\\( AP_{\\mathcal{U}\\uparrow} \\)\\( WI_{\\downarrow} \\)\\( AOSE_{\\downarrow} \\)\\( mAP_{\\mathcal{K}\\uparrow} \\)\\( AP_{\\mathcal{U}\\uparrow} \\)\\( WI_{\\downarrow} \\)\\( AOSE_{\\downarrow} \\)\\( mAP_{\\mathcal{K}\\uparrow} \\)\\( AP_{u\\uparrow} \\)
FR-CNN [33]80.1018.391511858.45022.742339155.26018.492547255.830
\\( FR-CNN^{\\dagger} \\)[33]80.0118.831194157.91023.241825754.77018.721956655.340
PROSER [47]79.6819.161303557.6610.9224.151983154.667.6219.642132255.203.25
ORE [18]79.8018.181281158.252.6022.401975255.301.7018.352141555.470.53
DS [28]80.0416.981286858.355.1320.861977555.313.3917.222192155.771.25
DFDD [44]80.2612.731072760.2116.2214.921363857.9812.0912.571666857.515.12
OpenDet [11]80.0214.951128658.7514.9318.231680055.8310.5814.241825056.374.36
OpenDet + Ours80.5311.981004360.5317.1914.181267058.6212.4711.831623158.025.63
", + "bbox": [ + 91, + 349, + 906, + 503 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4f3dd6f0ff87db80d8ca4155daf475c36cb03f7d803c6a95b6a4d2e1c2076023.jpg", + "table_caption": [ + "Table 3. Performance analysis of OSOD. We report close-set performance $(\\mathrm{mAP}_{\\mathcal{K}})$ on VOC, and both close-set $(\\mathrm{mAP}_{\\mathcal{K}})$ and open-set (WI, AOSE, $\\mathrm{AP}_{\\mathcal{U}}$ ) performance of different methods on VOC-COCO-{20, 40, 60}. Here, '20', '40', and '60' indicate that the testing COCO images separately contain 20, 40, and 60 non-VOC classes. '†' means a higher score threshold (i.e., 0.1) for testing." + ], + "table_footnote": [], + "table_body": "
MethodNSDRNRDF
SW [30]33.426.313.730.8
IBN-Net [29]32.126.114.329.6
IterNorm [17]29.622.812.628.4
ISW [2]33.225.914.131.8
S-DGOD [41]36.628.216.633.5
S-DGOD + Ours38.230.418.335.7
Clip-gap [39]36.932.318.738.5
Clip-gap + Ours38.534.120.240.1
", + "bbox": [ + 130, + 556, + 439, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bust discriminative boundary and accelerates the safe deployment of the model.", + "bbox": [ + 89, + 758, + 482, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation and Visualization Analysis", + "text_level": 1, + "bbox": [ + 89, + 801, + 401, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We utilize PASCAL VOC as the source data and MS-COCO with 'Acrylic' style as the target data to perform an ablation analysis. Our method is directly plugged into VOS [6].", + "bbox": [ + 89, + 823, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of WFS. As shown in Fig. 2, our method aims to emulate humans and mainly consists of three modules,", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5cb734242cb6161ff91ec1d11131c5e683ee0bb2f7d4db016483ae43fecbbf21.jpg", + "table_caption": [ + "Table 4. Single-DGOD analysis. 'NS', 'DR', 'NR', and 'DF' separately represent Night-Sunny, Daytime-Rainy, Night-Rainy, and Daytime-Foggy weather. mAP is taken as the metric." + ], + "table_footnote": [], + "table_body": "
PerceptMemoryImagineFPR95 ↓AUROC ↑
✓84.68%72.23%
✓83.13%72.75%
✓✓82.85%73.06%
✓✓81.07%73.94%
✓✓✓80.16%74.31%
", + "bbox": [ + 526, + 556, + 888, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation analysis of WFS for ODU-OD. Here, 'Percept', 'Memory', and 'Imagine' separately indicate multi-level perception, memory recorder, and unknown-feature generator.", + "bbox": [ + 511, + 657, + 906, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "i.e., multi-level perception, fragmented memory recorder, and imagination-based unknown-feature generator. In Table 5, we make an ablation analysis of the three modules. We can observe that employing dedicated perception, memory, and imagination-based generator could effectively reduce the impact of two-shift entanglement and improve the detection performance significantly. Particularly, when using the generator for synthesizing unknown features, the performance could be further enhanced. This shows that imitating humans to leverage perception and memory to imagine unknowns is an effective solution for addressing two-shift entanglement, promoting learning a robust and discriminative boundary between known and unknown objects.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4688", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7816d702ba4f333d4007aecd00c08c141b0deb97405e86bc8dda6eb1ef482a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 88, + 243, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6af2c9e2e71696338163f5ef456cd442351d7150684119eb230452b80fedd699.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 88, + 374, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/840c243ca0081aedf46179da98ceb1ee24400bc515418db81e44cd1978c8c367.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 88, + 504, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb78ab8bb3d46871eae72cfd6885f5e939b2380c9a97d9ce5cb2a66c31c06a7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 88, + 633, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/381157da76156e8862d8aee7b5fc77e89d9c657115fdad0cc52995a0e785afdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 88, + 763, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02b4052fedae00e0369823b996a2f6e3f3b63cc6280a4a4433bf4ed210c579a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 89, + 895, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29ecbfaaeec0562be98a0915ebdf26bd6ed1f4249086c4d04d15220343e03eca.jpg", + "image_caption": [ + "Figure 5. Memory vs. Imagination. The first and second rows respectively show the captured memory and reasoned imagination. Here, we directly project the memory codebook and imagination bank into the corresponding visual regions with three different styles. We can observe that for the same input, memory and imagination separately pay attention to different fragmented content." + ], + "image_footnote": [], + "bbox": [ + 114, + 162, + 243, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d913f5daa46a8a7f2bd447dd8a8bafa459082df9b17a14e83c4690c56a49eec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 162, + 374, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c4e6355119b3e276a67ba4fe924170d467cf0b28060198c68268a7d219b2be7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 162, + 504, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7cd235b86860394974936228a54e67824ff760baa295725eb1ae932edc25f4e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 162, + 633, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67fb31a0290ba2146bcbc9cdf5d60aeced81e4635235efbd2d1fdc7b6ce63087.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 162, + 763, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f28a8d5250108df04dc8a8c8fa37b59826f26e33a2f7047968339d09705ba33c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 162, + 895, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6da988c23ba6f49c6ce8fcf255d1d31da9393ceba003f9ee115ce84767830037.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 286, + 230, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/536bdda7e54ae6801b7b06982362501cd946a4f352cdfa36068d22764da80f84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 286, + 352, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9226f751195059ca222a8ff48cbbbde0750b063618b0c1778e649df26626205d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 286, + 475, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33cbb57dc76ac45e4af8fab885015ebd12158ba8f6e2e38353b7436091f72de3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 362, + 230, + 433 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ee62376f41a930d22c0443dad774b5ba4488222aae85f3c989c5be96318c36b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 362, + 352, + 433 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c7b84bd9511e4fe03ac570c643b76688ebe0346dbb87beab0b9a6418f948b657.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 362, + 475, + 433 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19d3dbd4b28d1a3c1b77270d7acc3f98cb09352ed9358e089919011f8f429862.jpg", + "image_caption": [ + "(a) Style Image", + "Figure 6. Visualization of the memory-enhanced map $E$ and generated unknown map $S_{u}$ . For each feature map, the channels corresponding to the maximum value are selected for visualization." + ], + "image_footnote": [], + "bbox": [ + 109, + 438, + 228, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e12ec459e8eed967cc1a57b37f87ab1c48e87ac1ef27da8e75600495629c2582.jpg", + "image_caption": [ + "(b) Enhanced $E$" + ], + "image_footnote": [], + "bbox": [ + 232, + 438, + 352, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39d599cf85465a8b287e8a84bba31e5c9011b414f081e8e0d1a894c1486602ba.jpg", + "image_caption": [ + "(c) Unknown $S_{u}$" + ], + "image_footnote": [], + "bbox": [ + 354, + 438, + 475, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Multi-level Perception. As shown in Fig. 1 and 2, the goal of multi-level perception is to construct a complete impression of objects. Here, we make an ablation analysis of three perception levels. We only change the perception level. Other operations and settings are kept unchanged. Table 6 shows the results. We can observe that employing different perception level affects the corresponding detection performance. Particularly, since the connective level considers the relationships between visual regions, using this level could enhance the performance effectively. When combining these three levels, the performance could be further improved. This indicates that multi-level perception is indeed conducive to building a complete observation of objects, strengthening the robustness.", + "bbox": [ + 89, + 579, + 482, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Codeword Number $Q$ . For the fragmented memory recorder, we utilize $Q$ codewords to preserve the information of known objects. Here, we analyze the $Q$ impact. We observe that when $Q$ is set to 16 and 32, the FPR95 value increases slightly, i.e., $80.54\\%$ and $80.33\\%$ . The reason may be that utilizing fewer codewords could not sufficiently capture known characteristics. Utilizing more code", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d431760264ff824eb3c18e7895deb3efd735c5acefece166dd316b42a35c15b3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LocalGlobalConnectiveFPR95 ↓AUROC ↑
✓81.19%73.04%
✓81.62%73.18%
✓81.37%73.21%
✓✓80.85%73.96%
✓✓80.63%74.12%
✓✓80.58%74.09%
✓✓✓80.16%74.31%
", + "bbox": [ + 527, + 287, + 888, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation analysis of local, global, and connective perception perspectives. Other operations are kept unchanged.", + "bbox": [ + 511, + 416, + 903, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "words increases parameters, resulting in overfitting.", + "bbox": [ + 511, + 453, + 852, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 5, we intuitively show the corresponding content captured by the built memory and reasoned imagination. Obviously, for the images with different styles, the memory recorder can effectively focus on fragmented object-related content, enhancing the robustness against the covariate-shift impact. Furthermore, the reasoned imagination mainly pays attention to background regions that involve plentiful visual information, which is instrumental in promoting the following generator to synthesize expected unknown features and attenuates the impact of two-shift entanglement.", + "bbox": [ + 511, + 469, + 903, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization Analysis. Fig. 6 shows three examples of memory-enhanced map and virtual unknown map. We can further observe that fusing the known memory could indeed strengthen object-related information. The synthesized unknown features mainly pay attention to background regions involving plentiful visual characteristics, enlarging the gap with the known features significantly. In supplementary material, we will give more analysis about our method.", + "bbox": [ + 511, + 619, + 905, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 753, + 633, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper explores a challenging problem, i.e., ODU-OD, which aims to detect unknown objects within unseen target domains. To weaken the entangled impact of covariate and semantic shifts, we propose a new WFS method and imitate humans to leverage multi-level perception, fragmented memory, and reasoned imagination to recurrently generate virtual unknown features. Extensive experimental results on four tasks demonstrate the superiorities of our method.", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4689", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the National Key R&D Program of China (No. 2023YFC3305600), Joint Fund of Ministry of Education of China (8091B022149, 8091B02072404), National Natural Science Foundation of China (62472333, 62132016, 62171343), and Natural Science Basic Research Program of Shaanxi (2020JC-23).", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ECCV, 2020. 1", + "[2] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, pages 11580-11590, 2021. 7", + "[3] Muhammad Sohail Danish, Muhammad Haris Khan, Muhammad Akhtar Munir, M Saqib Sarfraz, and Mohsen Ali. Improving single domain-generalized object detection: A focus on diversification and alignment. In CVPR, pages 17732-17742, 2024. 2", + "[4] Xuefeng Du, Gabriel Gozum, Yifei Ming, and Yixuan Li. Siren: Shaping representations for detecting out-of-distribution objects. In NeurIPS, 2022. 1, 3, 6", + "[5] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware object detection: Learning what you don't know from videos in the wild. In CVPR, pages 13678-13688, 2022. 3", + "[6] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. ICLR, 2022. 1, 2, 3, 5, 6, 7", + "[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2", + "[8] Yanchen Guan, Haicheng Liao, Zhenning Li, Guohui Zhang, and Chengzhong Xu. World models for autonomous driving: An initial survey. arXiv preprint arXiv:2403.02622, 2024. 1", + "[9] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018.", + "[10] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. *ICLR*, 2019. 1", + "[11] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In CVPR, pages 9591-9600, 2022. 2, 5, 6, 7", + "[12] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 2, 3", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2, 3, 6" + ], + "bbox": [ + 93, + 244, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3, 5", + "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. *ICLR*, 2017. 6", + "[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, pages 10951-10960, 2020. 6", + "[17] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, pages 4874-4883, 2019. 7", + "[18] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In CVPR, pages 5830-5840, 2021. 7", + "[19] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2, 3", + "[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. arXiv preprint arXiv:2112.00374, 2021. 2, 5, 7", + "[21] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. *ICLR*, 2018. 6", + "[22] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. NeurIPS, 31, 2018. 6", + "[23] Deng Li, Among Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In CVPR, pages 17606-17615, 2024. 2", + "[24] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. *ICLR*, 2017. 6", + "[25] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In ICCV, pages 8771-8780, 2021. 1", + "[26] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation models? ECCV, 2024. 1, 2, 3", + "[27] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. NeurIPS, 33: 21464-21475, 2020. 5, 6", + "[28] Dimity Miller, Lachlan Nicholson, Feras Dayoub, and Niko Sünderhauf. Dropout sampling for robust object detection in open-set conditions. In ICRA, pages 3243-3249, 2018. 7", + "[29] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, pages 464-479, 2018. 7", + "[30] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, pages 1863-1871, 2019. 7", + "[31] Bimsara Pathiraja, Malitha Gunawardhana, and Muhammad Haris Khan. Multiclass confidence and localization calibration for object detection. In CVPR, pages 19734-19743, 2023. 1" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4690", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, pages 7263-7271, 2017.", + "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurlPS, pages 91-99, 2015. 1, 3, 5, 7", + "[34] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with gram matrices. In ICML, pages 8491-8501. PMLR, 2020. 6", + "[35] Eleanor Spens and Neil Burgess. A generative model of memory construction and consolidation. Nature Human Behaviour, pages 1-18, 2024. 1", + "[36] Weinan Sun, Madhu Advani, Nelson Spruston, Andrew Saxe, and James E Fitzgerald. Organizing memories for generalization in complementary learning systems. Nature neuroscience, 26(8):1438-1448, 2023. 1", + "[37] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. NeurIPS, 33:11839-11852, 2020. 6", + "[38] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. *ICLR*, 2018. 2", + "[39] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, pages 3219-3229, 2023. 2, 6, 7", + "[40] Samuel Wilson, Tobias Fischer, Feras Dayoub, Dimity Miller, and Niko Sunderhauf. Safe: Sensitivity-aware features for out-of-distribution object detection. In ICCV, pages 23565-23576, 2023. 3, 6", + "[41] Among Wu and Cheng Deng. Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation. In CVPR, pages 847-856, 2022. 1, 2, 5, 6, 7", + "[42] Among Wu and Cheng Deng. Discriminating known from unknown objects via structure-enhanced recurrent variational autoencoder. In CVPR, pages 23956-23965, 2023. 1, 6", + "[43] Among Wu, Yahong Han, Linchao Zhu, and Yi Yang. Instance-invariant domain adaptive object detection via progressive disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4178-4193, 2021. 1", + "[44] Among Wu, Da Chen, and Cheng Deng. Deep feature deblurring diffusion for detecting out-of-distribution objects. In ICCV, pages 13381-13391, 2023. 1, 3, 5, 6, 7", + "[45] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 7", + "[46] Hang Zhang, Jia Xue, and Kristin Dana. Deep ten: Texture encoding network. In CVPR, pages 708-717, 2017. 4", + "[47] Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Learning placeholders for open-set recognition. In CVPR, pages 4401-4410, 2021. 7", + "[48] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 1" + ], + "bbox": [ + 89, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[49] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 3", + "bbox": [ + 516, + 92, + 903, + 147 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4691", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/b34e32e0-849d-4c9e-8711-6ce6de1dc89c_model.json b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/b34e32e0-849d-4c9e-8711-6ce6de1dc89c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a7a41408763466e8a185e3a22e90e0e44a9b542d --- /dev/null +++ b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/b34e32e0-849d-4c9e-8711-6ce6de1dc89c_model.json @@ -0,0 +1,2409 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.817, + 0.177 + ], + "angle": 0, + "content": "Percept, Memory, and Imagine: World Feature Simulating for Open-Domain Unknown Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.205, + 0.612, + 0.222 + ], + "angle": 0, + "content": "Among Wu, Cheng Deng*" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.222, + 0.761, + 0.24 + ], + "angle": 0, + "content": "School of Electronic Engineering, Xidian University, Xi'an, China" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.242, + 0.697, + 0.256 + ], + "angle": 0, + "content": "amwu@xidian.edu.cn, chdeng@mail.xidian.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.324, + 0.485, + 0.792 + ], + "angle": 0, + "content": "To accelerate the safe deployment of object detectors, we focus on reducing the impact of both covariate and semantic shifts. And we consider a realistic yet challenging scenario, namely Open-Domain Unknown Object Detection (ODU-OD), which aims to detect unknown objects in unseen target domains without accessing any auxiliary data. Towards ODU-OD, it is feasible to learn a robust discriminative boundary by synthesizing virtual features. Generally, perception, memory, and imagination are three essential capacities for human beings. Through multi-level perception and rich memory about known objects, the characteristics of unknown objects can be imagined sufficiently, enhancing the ability of discriminating known from unknown objects. Inspired by this idea, an approach of World Feature Simulation (WFS) is proposed, mainly consisting of a multi-level perception, memory recorder, and unknown-feature generator. Specifically, after extracting the features of the input, we separately employ a Mamba and Graph Network to obtain the global-level and connective-level representations. Next, a codebook containing multiple learnable codewords is defined to preserve fragmented memory of known objects. Meanwhile, we perform a modulated operation on the memory to form the imagination bank involving unknown characteristics. Finally, to alleviate the impact of lacking supervision data, based on the multi-level representation and imagination bank, a dedicated unknown-feature generator is designed to recurrently synthesize outlier features deviating from in-distribution (ID) objects. The significant performance gains on four different detection tasks demonstrate the superiorities of our method. The code will be released at https://github.com/AmingWu/WFS." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.821, + 0.222, + 0.837 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.847, + 0.483, + 0.878 + ], + "angle": 0, + "content": "With the rejuvenation of deep neural networks, object detection has achieved many advances [1, 14, 31-33] recently." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.293, + 0.907, + 0.369 + ], + "angle": 0, + "content": "Though these advances have obtained impressive performance, in real-world scenarios, they inevitably encounter both covariate shifts (e.g., test-time changes in image styles) and semantic shifts (e.g., test-time unknown classes), which limits their applications." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.37, + 0.907, + 0.567 + ], + "angle": 0, + "content": "Since there is no auxiliary data available, to alleviate the covariate-shift impact, existing researches [25, 41, 43, 48] mainly pay attention to enhancing the generalization ability by learning domain-invariant features. As for the semantic-shift challenge, most existing approaches [4, 6, 26, 42, 44] explore synthesizing diverse virtual features to strengthen the capability of discriminating unknown objects. Although these methods have been demonstrated to be effective, they are only concerned with a single-shift case and could not cope with a more difficult scenario involving both covariate and semantic shifts, e.g., for an autonomous car, the detector may encounter an unknown object in unseen adverse weather, resulting in a safety risk." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.718 + ], + "angle": 0, + "content": "To promote the safe deployment of detectors, we propose a realistic yet challenging task, namely Open-Domain Unknown Object Detection (ODU-OD). As shown in Fig. 1, given one source dataset including multiple pre-defined categories for training, the goal of ODU-OD is to accurately detect unknown out-of-distribution (OOD) objects in multiple unseen target domains with different styles. Since ODU-OD contains both covariate and semantic shifts and fails to access any auxiliary data for training, existing single-shift methods can not be directly utilized to tackle this task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In essence, towards covariate and semantic shifts, it is critical to learn the features that lead to a robust discriminative boundary. Particularly, enhancing the robustness could attenuate the covariate-shift impact and improve the generalization ability. And strengthening the discrimination is instrumental in accurately recognizing known and unknown objects. Moreover, recent researches [8-10, 35, 36] indicate that based on the current observations and formed memory, human beings could produce rich imaginations, improving the adaptation ability for unknown situations. Therefore, we explore simulating humans to leverage plentiful perception and fragmented memory to imagine unknowns for training" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4682" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.418, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.089, + 0.638, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.089, + 0.906, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.264, + 0.907, + 0.349 + ], + "angle": 0, + "content": "Figure 1. World Feature Simulation for Open-Domain Unknown Object Detection, which imitates humans to synthesize virtual unknown features and consists of a multi-level perception, memory recorder, and dedicated unknown-feature generator. We first calculate the multi-level representation from the local, global, and connective levels. Then, a memory codebook is constructed to preserve the local object information (as shown in pink boxes). Next, a modulated operation is performed on the fragmented memory to form the imagination bank involving unknown characteristics. Finally, we design an unknown-feature generator to recurrently synthesize virtual unknown features deviating from known objects, enhancing the ability of detecting unknown objects within unseen domains (as shown in green boxes)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.483, + 0.387 + ], + "angle": 0, + "content": "(as shown in Fig. 1), enhancing the ability of distinguishing unknown objects within unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.387, + 0.483, + 0.644 + ], + "angle": 0, + "content": "Specifically, as shown in Fig. 1, a method of World Feature Simulation (WFS) is proposed, mainly consisting of a multi-level perception, memory recorder, and unknown-feature generator. For a specific object, in order to deepen its perception, humans usually observe the object from multiple different perspectives, e.g., local and global levels. To this end, given an ID input, we first utilize a convolutional backbone [13] to extract its local features. Then, a Mamba [7, 12, 49] and a Graph network [19, 38] are separately performed on the local features to obtain the global and connective features. And the local, global, and connective features are concatenated to form the multi-level representation. Next, we define a codebook containing multiple learnable codewords to preserve fragmented memory of the known objects. By fusing the memory into the current multi-level representation, the robustness against covariateshift risk could be enhanced effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.825 + ], + "angle": 0, + "content": "In Fig. 1, to achieve the ability of imagining unknowns, a modulated operation is performed on the known codebook to obtain the imagination bank. And we design a contrastive constraint to promote the bank to involve plentiful unknown characteristics. Finally, a dedicated unknown-feature generator is proposed to leverage the multi-level representation and imagined unknown characteristics to recurrently synthesize virtual unknown features deviating from ID inputs, which is conducive to learning a robust discriminative boundary between known and unknown objects within unseen domains. Experimental results on four different detection tasks demonstrate the superiorities of our method." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.826, + 0.414, + 0.839 + ], + "angle": 0, + "content": "The contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.902 + ], + "angle": 0, + "content": "(1) To promote the deployment of detectors, we propose a new task, namely Open-Domain Unknown Object Detection (ODU-OD), aiming to detect unknown objects in unseen target domains without accessing any auxiliary data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.447 + ], + "angle": 0, + "content": "(2) To tackle ODU-OD involving both covariate and semantic shifts, we imitate humans to explore a new framework of World Feature Simulation, which aims to synthesize virtual unknown features for training and contains a multi-level perception, fragmented memory recorder, and unknown-feature generator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.909, + 0.6 + ], + "angle": 0, + "content": "(3) We build an Open-Domain Dataset for unknown object detection to verify our method WFS, consisting of three different styles [20], i.e., 'Acrylic', 'Purple-Brush', and 'Sketch'. Meanwhile, our method is further evaluated on three single-shift tasks, i.e., OOD-OD [6], open-set object detection [11], and Single-DGOD [41]. The significant performance gains over multiple baselines demonstrate the effectiveness of our method. Meanwhile, visualization analyses also show that the built memory and reasoned imagination could indeed capture fragmented visual content." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.357, + 0.909, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.617, + 0.655, + 0.632 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Covariate-shift Detection. To improve the generalization of detectors, a task of Single-DGOD [41] is proposed, aiming to learn an object detector that performs well on multiple unseen target domains with only one source domain for training. Some works [23, 39, 41] utilize feature disentanglement and large-scale models to extract domain-invariant representations. Moreover, Liu et al. [26] further consider the biased problem and propose a causal-based method to learn unbiased generalized representations. And Danish et al. [3] explore diversifying the source domain and aligning detections based on prediction confidence. These methods have been verified on multiple benchmarks. However, they only consider the covariate-shift scenario and could not detect unknown objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Semantic-shift Detection. Recently, a task of unsupervised OOD-OD [6] is proposed, whose goal is to detect unknown objects without relying on any OOD data. To alle" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4683" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.907, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.336, + 0.907, + 0.421 + ], + "angle": 0, + "content": "Figure 2. The details of World Feature Simulation for ODU-OD. To enhance the robustness against covariate shift, a multi-level perception is used to extract the representation \\( F_{v} \\) from the local, global, and connective levels. Meanwhile, we design a memory recorder to preserve rich content about known objects. By calculating a memory-enhanced map, the robustness could be further strengthened. Next, through a modulated operation, we can reason the imagination bank involving plentiful unknown characteristics. Finally, a dedicated unknown-feature generator is proposed to utilize the built memory and imagination characteristics to recurrently synthesize virtual unknown features, improving the ability of detecting unknown objects within unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.428, + 0.482, + 0.669 + ], + "angle": 0, + "content": "viate the impact of lacking OOD data for supervision, VOS [6] and DFDD [44] explore synthesizing virtual OOD features that slightly deviate from the distribution of ID data. Besides, the work [5] further learns unknown-aware knowledge from external videos, which does not match the setting of unsupervised OOD-OD. Finally, Du et al. [4] present to use a distance-based mechanism to shape the learned representations. Meanwhile, Wilson et al. [40] attempt to enhance the sensitivity of the learned features, which is beneficial for improving the performance of detecting OOD objects. The work [26] struggles to leverage generative models trained on large-scale open-set data to synthesize OOD samples. Although these methods have been demonstrated to be effective, they only consider the semantic-shift problem and ignore the covariate-shift impact, which limits their applications to unknown objects within unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.671, + 0.483, + 0.792 + ], + "angle": 0, + "content": "Different from the above methods, in this paper, we explore a more difficult problem involving both covariate and semantic shifts, i.e., Open-Domain Unknown Object Detection. And a new approach of World Feature Simulation is proposed for synthesizing unknown features for training, consisting of a multi-level perception, fragmented memory recorder, and unknown-feature generator. Experimental results on multiple datasets indicate its superiorities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.813, + 0.336, + 0.829 + ], + "angle": 0, + "content": "3. World Feature Simulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "For ODU-OD, during training, we can only access single source-domain data with some pre-defined categories. During test, the detector is required to accurately localize and recognize unknown objects within multiple unseen do" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.473 + ], + "angle": 0, + "content": "mains. To reduce the entangled impact of covariate and semantic shifts, as shown in Fig. 2, we design a World Feature Simulation for synthesizing unknown features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.48, + 0.86, + 0.496 + ], + "angle": 0, + "content": "3.1. Observing Objects from Multiple Levels" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.906, + 0.607 + ], + "angle": 0, + "content": "Generally, in order to deepen the perception of a specific object, human beings usually observe the object from multiple different perspectives, which is beneficial for constructing a complete impression. Therefore, in Fig. 2, we explore imitating this capability based on local, global, and connective levels. Specifically, we exploit the widely used object detector, i.e., Faster R-CNN [14, 33], as the basic model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Given an ID input image, a convolutional network, e.g., ResNet [13], is taken as the backbone to extract the local feature map \\( F_{l} \\in \\mathbb{R}^{w \\times h \\times c} \\), where \\( w \\), \\( h \\), and \\( c \\) separately represent width, height, and the number of channels. Each element in \\( F_{l} \\) describes a local region of the input. Besides the local level, human beings usually intensify their impression by global observation and discovering the relations between local regions. Hence, we first unfold \\( F_{l} \\) to a sequence of visual elements, i.e., \\( D = \\{d_{i} \\in \\mathbb{R}^{c}, i = 1, \\dots, wh\\} \\). Then, one-layer vision mamba [49] is used to process \\( D \\), which outputs \\( H = \\{h_{i} \\in \\mathbb{R}^{c}, i = 1, \\dots, wh\\} \\). Finally, \\( H \\) is reshaped to the original size to obtain the global representation \\( F_{g} \\in \\mathbb{R}^{w \\times h \\times c} \\) [12]. Next, we use one-layer graph convolution [19] to further capture the semantic connections between local regions. Concretely, we first reshape \\( F_{l} \\) to \\( \\tilde{F} \\in \\mathbb{R}^{wh \\times c} \\). Then, an adjacency matrix for an undirected graph is defined as \\( \\mathbf{A} = \\text{softmax}(\\tilde{F}\\tilde{F}^{\\top}) + I_{wh} \\), where \\( I_{wh} \\) indicates the identity matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.884, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} = \\mathbf {A} \\tilde {F}, \\quad F _ {c} = r e l u (W _ {c} * \\tilde {\\mathcal {R}} + b _ {c}), \\qquad (1)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4684" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.485, + 0.228 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathcal{R}}\\in \\mathbb{R}^{w\\times h\\times c}\\) denotes the reshaped result of \\(\\mathcal{R}\\). \\(W_{c}\\in \\mathbb{R}^{3\\times 3\\times c\\times c}\\) and \\(b_{c}\\in \\mathbb{R}^{c}\\) indicate the trainable parameters. \\(F_{c}\\in \\mathbb{R}^{w\\times h\\times c}\\) represents the connective feature. Finally, \\(F_{l},F_{g}\\), and \\(F_{c}\\) are concatenated as the multi-level representation \\(F_{v}\\in \\mathbb{R}^{w\\times h\\times c}\\), where \\(F_{v} = \\phi ([F_{l},F_{g},F_{c}])\\) and \\(\\phi (\\cdot)\\in \\mathbb{R}^{1\\times 1\\times 3c\\times c}\\) is utilized to transform the number of channels. Compared with \\(F_{l},F_{v}\\) involves sufficient object-related information from three levels, which strengthens the robustness against the covariate shift." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.26 + ], + "angle": 0, + "content": "3.2. Mining Fragmented Memory and Imagination" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.485, + 0.51 + ], + "angle": 0, + "content": "In general, human beings could exploit familiar fragmented memory to improve the robustness of recognizing known objects. Then, based on the memory, human beings could produce rich imagination, enhancing the ability of discriminating unknowns. To simulate this capability, we present a memory recorder [46] to preserve the information of known objects. Particularly, a learnable memory codebook is denoted as \\( M = \\{m_j \\in \\mathbb{R}^c, j = 1, \\dots, Q\\} \\) that contains \\( Q \\) codewords. Meanwhile, by means of a weight \\( a_{ij} \\), each element \\( f_i \\in \\mathbb{R}^c \\) of the multi-level representation \\( F_v \\) can be assigned to each codeword \\( m_j \\). And the residual vector is represented as \\( r_{ij} = f_i - m_j \\), where \\( i = 1, \\dots, wh \\). Here, the residual operation is conducive to promoting the memory codewords to effectively capture the corresponding content of local visual regions. The processes of capturing memory are shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.526, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\nv _ {j} = \\sum_ {i = 1} ^ {w h} a _ {i j} r _ {i j}, \\quad a _ {i j} = \\frac {\\exp \\left(- k _ {j} \\| r _ {i j} \\| ^ {2}\\right)}{\\sum_ {j = 1} ^ {Q} \\exp \\left(- k _ {j} \\| r _ {i j} \\| ^ {2}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.734 + ], + "angle": 0, + "content": "where \\(k_{j}\\) represents the learnable smoothing factor for the codeword \\(m_j\\). Finally, \\(\\mathcal{V} \\in \\mathbb{R}^{Q \\times c}\\) is the output result. The concatenation of \\(F_v\\) and \\(\\mathcal{V}\\) is used to calculate the memory-enhanced map \\(E \\in \\mathbb{R}^{w \\times h \\times c}\\), i.e., \\(E = \\psi([F_v, \\tilde{\\mathcal{V}}W_0 + b_0])\\), where \\(\\tilde{\\mathcal{V}} \\in \\mathbb{R}^{1 \\times Qc}\\) is the reshaped result of \\(\\mathcal{V}\\). Meanwhile, \\(W_0 \\in \\mathbb{R}^{Qc \\times c}\\) and \\(b_0 \\in \\mathbb{R}^c\\) are parameters of a fully-connected layer. \\(\\psi(\\cdot)\\) includes two convolutional layers to transform the concatenated result. Through fusing memory, the information capacity of \\(E\\) could be increased, improving the robustness of object localization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Next, we further explore how to leverage the memory to reason the imagination that guides the synthesis of virtual unknown features. Particularly, we assume that the imagination should satisfy two conditions: (1) Imagination and memory should belong to different spaces. (2) Imagination mainly describes unknowns. It could not well represent known visual content. To this end, based on the memory \\( M \\in \\mathbb{R}^{Q \\times c} \\), a modulated operation is performed on \\( M \\) to obtain the imagination bank \\( \\mathcal{I} \\in \\mathbb{R}^{Q \\times c} \\), i.e., \\( \\mathcal{I} = \\gamma \\odot M + \\beta \\), where \\( \\gamma \\) and \\( \\beta \\) are learnable parameters for channel-wisely scaling and shifting \\( M \\). Next, we separately utilize \\( M \\) and" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.089, + 0.895, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.239, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Figure 3. Unknown-feature generator. Based on \\( F_{m} \\in \\mathbb{R}^{w \\times h \\times c} \\), a variational operation is performed on \\( F_{m} \\) to obtain \\( Z \\in \\mathbb{R}^{w \\times h \\times c} \\). Then, we exploit the reasoned imagination bank to re-represent \\( Z \\), which outputs \\( G \\in \\mathbb{R}^{w \\times h \\times c} \\). Next, \\( Z \\) and \\( G \\) are unfolded into a sequence with \\( wh \\) elements. Finally, a dedicated recurrent generator including three inputs and one output is presented to synthesize virtual unknown features, improving the discrimination." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.342, + 0.655, + 0.356 + ], + "angle": 0, + "content": "\\(\\mathcal{I}\\) to re-represent \\(F_{v}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.362, + 0.907, + 0.381 + ], + "angle": 0, + "content": "\\[\nF _ {m} = \\operatorname {s o f t} x \\left(\\tilde {F} _ {v} M ^ {\\top}\\right) M, F _ {i} = \\operatorname {s o f t} x \\left(\\tilde {F} _ {v} \\mathcal {I} ^ {\\top}\\right) \\mathcal {I}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.387, + 0.906, + 0.478 + ], + "angle": 0, + "content": "where 'softx' denotes the softmax operation. \\(\\tilde{F}_v\\in \\mathbb{R}^{wh\\times c}\\) is the reshaped result of \\(F_{v}\\). \\(F_{m}\\in \\mathbb{R}^{w\\times h\\times c}\\) and \\(F_{i}\\in \\mathbb{R}^{w\\times h\\times c}\\) indicate the re-represented results based on the fragmented memory and imagination. Finally, a contrastive loss is exploited to enlarge the gap between the memory recorder and imagination bank:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.484, + 0.907, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i m} = - \\log \\frac {\\exp \\left(\\sin \\left(F _ {m} , F _ {v}\\right) / \\tau\\right)}{\\sum_ {j = 0} ^ {1} \\exp \\left(\\sin \\left(B [ j ] , F _ {v}\\right) / \\tau\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.694 + ], + "angle": 0, + "content": "where \\( B = [F_m, F_i] \\). \\( \\mathrm{sim}(F_m, F_v) \\) denotes the average of the cosine similarity between all corresponding elements of feature map \\( F_m \\) and \\( F_v \\). \\( \\tau \\) is a hyper-parameter, which is set to 1.0 in the experiments. Obviously, by minimizing the contrastive loss, the distance between \\( F_m \\) and \\( F_v \\) could be narrowed, promoting the memory recorder to capture sufficient content about known objects. Instead, the distance between \\( F_i \\) and \\( F_v \\) is enlarged, forcing the reasoned imagination bank to contain plentiful information deviating from knowns. Next, the imagination bank is utilized to generate expected unknown features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.701, + 0.885, + 0.718 + ], + "angle": 0, + "content": "3.3. Generating Unknown Features for Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.723, + 0.906, + 0.86 + ], + "angle": 0, + "content": "For ODU-OD, to attenuate the impact of lacking unknown data for supervision, we design an unknown-feature generator to synthesize outlier features involving unknown characteristics. Specifically, two convolutional networks \\( W_{\\mu} \\) and \\( W_{V} \\) are performed on \\( F_{m} \\) to estimate the corresponding means \\( \\mu \\in \\mathbb{R}^{w\\times h\\times c} \\) and variances \\( V \\in \\mathbb{R}^{w\\times h\\times c} \\). Then, an encoding operation is performed to obtain \\( Z \\in \\mathbb{R}^{w\\times h\\times c} \\). Meanwhile, we exploit the reasoned imagination bank \\( \\mathcal{I} \\) to re-represent \\( Z \\). The processes are shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.866, + 0.905, + 0.891 + ], + "angle": 0, + "content": "\\[\n\\mu = W _ {\\mu} * F _ {m}, \\quad V = W _ {V} * F _ {m}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.886, + 0.852, + 0.904 + ], + "angle": 0, + "content": "\\[\nZ = \\mu + \\epsilon \\odot V, \\quad G = \\text {s o f t m a x} (\\tilde {Z} \\mathcal {I} ^ {\\top}) \\mathcal {I},\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4685" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.196 + ], + "angle": 0, + "content": "where \\(\\tilde{Z} \\in \\mathbb{R}^{wh \\times c}\\) is the reshaped result of \\(Z\\). \\(\\epsilon\\) represents Gaussian noise sampled from \\(\\mathcal{N}(0, I)\\). \\(G \\in \\mathbb{R}^{w \\times h \\times c}\\) is the re-represented result based on the imagination bank \\(\\mathcal{I}\\). Since the imagination could not well express the knowns, \\(G\\) contains plentiful information deviating from known visual content, which is instrumental in promoting the generation of virtual unknown features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.197, + 0.482, + 0.289 + ], + "angle": 0, + "content": "Finally, we present a dedicated generator to recurrently synthesize expected unknown features. As shown in Fig. 3, the generator contains three inputs and one output. Particularly, at each time, the generator takes \\( Z_{t} \\), \\( G_{t} \\), and the previous output \\( S_{t-1} \\) as the input, and synthesizes the current output \\( S_{t} \\). The generating processes are shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.3, + 0.491, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {t} = \\sigma \\left(W _ {z z} * Z _ {t} + W _ {o z} * G _ {t} + W _ {s z} * S _ {t - 1} + b _ {z}\\right), \\\\ g _ {t} = \\sigma \\left(W _ {z g} * Z _ {t} + W _ {o g} * G _ {t} + W _ {s g} * S _ {t - 1} + b _ {g}\\right), \\\\ \\tilde {S} _ {t} = \\operatorname {T a n h} \\left(W _ {s} * Z _ {t} + W _ {h} * \\left(g _ {t} \\odot S _ {t - 1} + (1 - g _ {t}) \\odot G _ {t}\\right)\\right), \\\\ S _ {t} = \\left(1 - z _ {t}\\right) \\odot S _ {t - 1} + z _ {t} \\odot \\hat {S} _ {t}, \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.555 + ], + "angle": 0, + "content": "where \\(Z_{t}\\in \\mathbb{R}^{c}\\) and \\(G_{t}\\in \\mathbb{R}^{c}\\) are the current element of \\(Z\\) and \\(G\\), \\(t = 1,\\dots ,wh\\). And \\(W_{zz}\\), \\(W_{oz}\\), \\(W_{sz}\\), \\(b_{z}\\), \\(W_{zg}\\), \\(W_{og}\\), \\(W_{sg}\\), \\(b_{g}\\), \\(W_{s}\\), and \\(W_{h}\\) represent the trainable parameters. Through this generating mechanism, the synthesized features could sufficiently retain the information in previous steps, improving the content continuity of the generated features. Finally, the generated sequence \\(\\{S_1,S_2,\\dots S_{wh}\\}\\) are concatenated to obtain unknown feature \\(S_{u}\\in \\mathbb{R}^{w\\times h\\times c}\\). We still utilize a contrastive loss to expand the difference between known and the synthesized unknown features:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.566, + 0.482, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u n k} = - \\log \\frac {\\exp \\left(\\sin \\left(F _ {m} , F _ {v}\\right) / \\tau\\right)}{\\sum_ {j = 0} ^ {1} \\exp \\left(\\sin \\left(A [ j ] , F _ {v}\\right) / \\tau\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.482, + 0.703 + ], + "angle": 0, + "content": "where \\(A = [F_m, S_u]\\). \\(\\tau\\) is set to 1.0 in the experiments. Since the generation process involves perception, memory, and imagination, using the contrastive loss could not only promote \\(S_u\\) to contain plentiful unknown information but also facilitate \\(M\\) and \\(\\mathcal{I}\\) to separately capture rich known and unknown characteristics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.439, + 0.729 + ], + "angle": 0, + "content": "3.4. WFS-Driven Unknown-Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Lastly, the memory-enhanced map \\( E \\) is taken as the input of Region-Proposal Network (RPN) [14, 33] to obtain a series of object proposals \\( \\mathcal{O} \\). Meanwhile, based on \\( \\mathcal{O} \\), RoI-Alignment followed by RoI-Feature extraction [14] is performed on \\( E \\) to obtain \\( O_{\\mathrm{in}} \\in \\mathbb{R}^{p \\times n} \\), where \\( p \\) and \\( n \\) denote the number of proposals and channels, respectively. Then, \\( O_{\\mathrm{in}} \\) is input to the object classifier and regressor to calculate the classification loss \\( \\mathcal{L}_{\\mathrm{cls}} \\) and localization loss \\( \\mathcal{L}_{\\mathrm{loc}} \\), i.e., \\( \\mathcal{L}_{\\mathrm{det}} = \\mathcal{L}_{\\mathrm{cls}} + \\mathcal{L}_{\\mathrm{loc}} \\). By optimizing \\( \\mathcal{L}_{\\mathrm{det}} \\), the memory recorder could be promoted to capture plentiful information about known objects, enhancing the robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Next, we further enlarge the gap between the known and synthesized unknown features from the instance perspective. Concretely, based on \\(\\mathcal{O}\\), RoI-Alignment followed by RoI-Feature extraction is performed on \\(S_{u}\\) to obtain \\(O_{\\mathrm{unk}} \\in \\mathbb{R}^{p \\times n}\\). Meanwhile, we define a loss function \\(\\mathcal{L}_{\\mathrm{dis}}\\) to maximize the \\(KL\\)-divergence between \\(O_{\\mathrm{unk}}\\) and \\(O_{\\mathrm{in}}\\), i.e., \\(\\mathcal{L}_{\\mathrm{dis}} = \\mathrm{KL}[q(O_{\\mathrm{unk}}), q(O_{\\mathrm{in}})]\\), where \\(q(\\cdot)\\) represents the probability distribution. Besides, to attain the goal of distinguishing unknown objects from known ID objects, \\(O_{\\mathrm{unk}}\\) and \\(O_{\\mathrm{in}}\\) are used to compute an uncertainty loss [6], regularizing the detector to produce a low score for the ID object features, and a high score for the virtual unknown features:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.284, + 0.905, + 0.349 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {u n t y}} = \\mathbb {E} _ {x \\sim O _ {\\text {i n}}} [ - \\log \\frac {\\exp^ {- \\mathbf {E} (x)}}{1 + \\exp^ {- \\mathbf {E} (x)}} ] + \\tag {8} \\\\ \\mathbb {E} _ {v \\sim O _ {\\mathrm {u n k}}} [ - \\log \\frac {1}{1 + \\exp^ {- \\mathbf {E} (v)} ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.359, + 0.905, + 0.39 + ], + "angle": 0, + "content": "where \\(\\mathbf{E}(\\cdot)\\) is the object-level energy score [6, 27, 44]. The overall objective is shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.402, + 0.903, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t r a s t}} = \\mathcal {L} _ {i m} + \\mathcal {L} _ {u n k} - \\mathcal {L} _ {\\text {d i s}}, \\\\ \\mathcal {L} _ {i m} = \\mathcal {L} _ {u n k} + \\mathcal {L} _ {u n k} + \\lambda_ {i} \\mathcal {L} _ {i m}. \\end{array} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.421, + 0.836, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e t}} + \\alpha \\cdot \\mathcal {L} _ {\\text {c o n t r a s t}} + \\lambda \\cdot \\mathcal {L} _ {\\text {u n t y}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.905, + 0.476 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\lambda\\) are two hyper-parameters, which are set to 0.001 and 0.1 in the experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.862, + 0.5 + ], + "angle": 0, + "content": "3.5. Inference for Unknown Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.905, + 0.567 + ], + "angle": 0, + "content": "During test, we only calculate the uncertainty score for unknown object detection [6, 44]. Specifically, for a predicted bounding box \\(\\mathbf{b}\\), the processes of discriminating unknown objects are shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.578, + 0.905, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} = \\frac {\\exp^ {- \\mathbf {E} (\\mathbf {b})}}{1 + \\exp^ {- \\mathbf {E} (\\mathbf {b})}}, \\quad \\mathcal {C} (\\mathbf {b}) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} \\mathcal {E} < \\delta , \\\\ 1 & \\text {i f} \\mathcal {E} \\geq \\delta . \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.907, + 0.727 + ], + "angle": 0, + "content": "For the output of the classifier \\(\\mathcal{C}(\\cdot)\\), we utilize the threshold mechanism [6] to distinguish known objects (the result is 1) from unknown objects (the result is 0). The threshold \\(\\delta\\) is commonly set to 0.95 so that a high fraction of known ID objects are correctly classified. In supplementary material, we will show more details of the proposed WFS algorithm and give more discussion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.645, + 0.757 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.87 + ], + "angle": 0, + "content": "In the experiments, our method is evaluated on four different detection tasks, i.e., ODU-OD, OOD-OD [6], open-set object detection (OSOD) [11], and Single-DGOD [41]. Particularly, we construct an Open-Domain dataset for ODU-OD, which utilizes Clipstyler [20] to transform testing data of OOD-OD [6] to three different styles, i.e., 'Acrylic', 'Purple-Brush', and 'Sketch'." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We utilize Faster R-CNN [33] with RoI-Alignment layer [14] as the basic detection model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4686" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.089, + 0.842, + 0.336 + ], + "angle": 0, + "content": "
Source DataMethodStyle: Acrylic\nMS-COCO/OpenImagesStyle: Purple-Brush\nMS-COCO/OpenImagesStyle: Sketch\nMS-COCO/OpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
VOCSR-VAE [42]88.25/84.6371.02/69.6487.76/69.9367.42/79.1889.67/74.8266.47/77.04
DFDD [44]87.68/85.1270.85/69.2187.24/70.0867.23/79.4689.83/74.9266.13/76.82
VOS [6]88.73/84.7171.13/69.5688.66/70.1767.06/79.3389.76/74.6166.50/77.63
VOS + WFS80.16/77.5274.31/73.8580.07/60.6870.86/83.8782.16/67.2369.87/81.28
SAFE [40]86.62/47.8363.47/78.5186.31/39.6261.28/86.8287.56/55.2860.71/83.48
SAFE + WFS79.38/40.1667.28/81.6979.23/33.8566.47/88.7380.49/48.6165.82/84.17
BDDSR-VAE [42]64.72/53.9872.34/78.1166.43/57.2870.69/76.2469.36/59.2769.17/75.25
DFDD [44]64.08/53.2672.48/78.3665.91/56.7771.03/76.4868.74/58.8969.22/75.44
VOS [6]65.71/54.7371.96/77.8068.50/58.7970.18/75.8370.80/60.7169.05/74.97
VOS + WFS56.48/45.4275.87/82.1859.64/47.8874.34/79.8761.01/51.9473.78/79.36
SAFE [40]54.82/40.1873.71/82.6757.63/43.9772.34/80.8359.44/46.2871.27/79.53
SAFE + WFS45.91/32.0378.35/85.1447.78/34.4176.93/84.2650.26/37.1875.54/82.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.339, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Table 1. Unknown-object detection results under unseen open domains. \\(\\uparrow\\) denotes larger values are better and \\(\\downarrow\\) represents smaller values are better. 'VOS+WFS' and 'SAFE+WFS' indicate our method is directly plugged into VOS [6] and SAFE [40]. We can see that our method could reduce the impact of both covariate and semantic shifts." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.39, + 0.485, + 0.779 + ], + "angle": 0, + "content": "
SourceMethodFPR95 ↓AUROC ↑
OOD: MS-COCO / OpenImages
VOCMSP [15]70.99 / 73.1383.45 / 81.91
ODIN [24]59.82 / 63.1482.20 / 82.59
Mahalanobis [22]67.73 / 65.4181.45 / 81.48
Gram matrices [34]62.75 / 67.4279.88 / 77.62
Energy score [27]56.89 / 58.6983.69 / 82.98
Generalized ODIN [16]59.57 / 70.2883.12 / 79.23
CSI [37]59.91 / 57.4181.83 / 82.95
GAN-synthesis [21]60.93 / 59.9783.67 / 82.67
SIREN-vMF [4]64.68 / 68.5385.36 / 82.78
SIREN-KNN [4]47.45 / 50.3889.67 / 88.80
SR-VAE [42]42.17 / 46.2690.28 / 87.89
DFDD [44]41.34 / 44.5290.79 / 88.65
VOS (Baseline) [6]47.53 / 51.3388.70 / 85.23
VOS + WFS37.91 / 41.3691.47 / 89.78
SAFE (Baseline) [40]47.40 / 20.0680.30 / 92.28
SAFE + WFS37.48 / 11.2785.62 / 94.91
BDDMSP [15]80.94 / 79.0475.87 / 77.38
ODIN [24]62.85 / 58.9274.44 / 76.61
Mahalanobis [22]55.74 / 47.6985.71 / 88.05
Gram matrices [34]60.93 / 77.5574.93 / 59.38
Energy score [27]60.06 / 54.9777.48 / 79.60
Generalized ODIN [16]57.27 / 50.1785.22 / 87.18
CSI [37]47.10 / 37.0684.09 / 87.99
GAN-synthesis [21]57.03 / 50.6178.82 / 81.25
SR-VAE [42]32.23 / 21.8190.69 / 93.55
DFDD [44]30.71 / 22.6790.74 / 92.48
VOS (Baseline) [6]44.27 / 35.5486.87 / 88.52
VOS + WFS28.66 / 20.3992.37 / 94.13
SAFE (Baseline) [40]32.56 / 16.0488.96 / 94.64
SAFE + WFS21.84 / 7.8393.41 / 96.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Table 2. The performance of unsupervised OOD-OD [6] that assumes the training and test data are from the same domain. We can see that our method can reduce the impact of semantic shift." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "ResNet-50 [13] is taken as the backbone. For the memory recorder, the number \\( Q \\) of the codewords is set to 24. All the experiments are trained using the standard SGD optimizer with a learning rate of 0.02. In supplementary material, we" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.862, + 0.41 + ], + "angle": 0, + "content": "will give the details of the used datasets and metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.42, + 0.825, + 0.436 + ], + "angle": 0, + "content": "4.1. Analysis of ODU-OD and OOD-OD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.639 + ], + "angle": 0, + "content": "Table 1 and 2 separately illustrate the ODU-OD and OOD-OD performance. Here, we only compare with state-of-the-art methods. We can observe that though existing methods [6, 40, 42, 44] could effectively attenuate the semantic-shift impact, their performance drops severely when encountering the covariate shift (e.g., changes in image styles). Plugging our method into existing methods [6, 40] significantly improves their performance whether objects are within open or closed domains. This demonstrates that for the entangled case of covariate and semantic shifts, imitating humans to construct perception, memory, and imagination and further generating unknown features for training could simultaneously strengthen the robustness and discrimination." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.702 + ], + "angle": 0, + "content": "In Fig. 4, we show some detection examples with different styles. We can observe that plugging our method could indeed attenuate both covariate and semantic shifts, improving the detection accuracy under open domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.836, + 0.729 + ], + "angle": 0, + "content": "4.2. Analysis of OSOD and Single-DGOD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To further demonstrate that our method could attenuate the impact of two-shift entanglement, we verify our method on two dedicated tasks, i.e., OSOD [11] and Single-DGOD [41]. We directly plug our method into three existing methods [11, 39, 41] and do not change the training processes. Table 3 and 4 show the corresponding results. We can observe that plugging our method could improve the detection performance effectively. This further demonstrates that our method WFS could synthesize expected unknown features that are beneficial for alleviating the impact of covariate and semantic shifts, which promotes the detector to learn a ro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4687" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.298, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.092, + 0.481, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.092, + 0.654, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.092, + 0.777, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.092, + 0.899, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.19, + 0.298, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.283, + 0.234, + 0.297 + ], + "angle": 0, + "content": "(a) Acrylic" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.19, + 0.481, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.322, + 0.283, + 0.448, + 0.297 + ], + "angle": 0, + "content": "(b) Purple-Brush" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.19, + 0.654, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.283, + 0.606, + 0.296 + ], + "angle": 0, + "content": "(c) Sketch" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.19, + 0.777, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.283, + 0.752, + 0.296 + ], + "angle": 0, + "content": "(d) Nature" + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.19, + 0.899, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.283, + 0.872, + 0.296 + ], + "angle": 0, + "content": "(e) Nature" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.302, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Figure 4. Detection results on target images with multiple different styles. The first and second rows respectively indicate results based on VOS [6] and our method. The in-distribution source dataset is BDD-100k [45]. 'Arcylic', 'Purple-Brush', and 'Sketch' represent the rendered target styles [20]. 'Nature' is the same style as the source dataset." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.351, + 0.907, + 0.504 + ], + "angle": 0, + "content": "
MethodVOCVOC-COCO-20VOC-COCO-40VOC-COCO-60
\\( \\mathrm{mAP}_{\\mathcal{K}\\uparrow} \\)\\( WI_{\\downarrow} \\)\\( AOSE_{\\downarrow} \\)\\( mAP_{\\mathcal{K}\\uparrow} \\)\\( AP_{\\mathcal{U}\\uparrow} \\)\\( WI_{\\downarrow} \\)\\( AOSE_{\\downarrow} \\)\\( mAP_{\\mathcal{K}\\uparrow} \\)\\( AP_{\\mathcal{U}\\uparrow} \\)\\( WI_{\\downarrow} \\)\\( AOSE_{\\downarrow} \\)\\( mAP_{\\mathcal{K}\\uparrow} \\)\\( AP_{u\\uparrow} \\)
FR-CNN [33]80.1018.391511858.45022.742339155.26018.492547255.830
\\( FR-CNN^{\\dagger} \\)[33]80.0118.831194157.91023.241825754.77018.721956655.340
PROSER [47]79.6819.161303557.6610.9224.151983154.667.6219.642132255.203.25
ORE [18]79.8018.181281158.252.6022.401975255.301.7018.352141555.470.53
DS [28]80.0416.981286858.355.1320.861977555.313.3917.222192155.771.25
DFDD [44]80.2612.731072760.2116.2214.921363857.9812.0912.571666857.515.12
OpenDet [11]80.0214.951128658.7514.9318.231680055.8310.5814.241825056.374.36
OpenDet + Ours80.5311.981004360.5317.1914.181267058.6212.4711.831623158.025.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.51, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Table 3. Performance analysis of OSOD. We report close-set performance \\((\\mathrm{mAP}_{\\mathcal{K}})\\) on VOC, and both close-set \\((\\mathrm{mAP}_{\\mathcal{K}})\\) and open-set (WI, AOSE, \\(\\mathrm{AP}_{\\mathcal{U}}\\)) performance of different methods on VOC-COCO-{20, 40, 60}. Here, '20', '40', and '60' indicate that the testing COCO images separately contain 20, 40, and 60 non-VOC classes. '†' means a higher score threshold (i.e., 0.1) for testing." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.558, + 0.441, + 0.704 + ], + "angle": 0, + "content": "
MethodNSDRNRDF
SW [30]33.426.313.730.8
IBN-Net [29]32.126.114.329.6
IterNorm [17]29.622.812.628.4
ISW [2]33.225.914.131.8
S-DGOD [41]36.628.216.633.5
S-DGOD + Ours38.230.418.335.7
Clip-gap [39]36.932.318.738.5
Clip-gap + Ours38.534.120.240.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.707, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Table 4. Single-DGOD analysis. 'NS', 'DR', 'NR', and 'DF' separately represent Night-Sunny, Daytime-Rainy, Night-Rainy, and Daytime-Foggy weather. mAP is taken as the metric." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.483, + 0.79 + ], + "angle": 0, + "content": "bust discriminative boundary and accelerates the safe deployment of the model." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.802, + 0.402, + 0.817 + ], + "angle": 0, + "content": "4.3. Ablation and Visualization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.824, + 0.483, + 0.87 + ], + "angle": 0, + "content": "We utilize PASCAL VOC as the source data and MS-COCO with 'Acrylic' style as the target data to perform an ablation analysis. Our method is directly plugged into VOS [6]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Analysis of WFS. As shown in Fig. 2, our method aims to emulate humans and mainly consists of three modules," + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.558, + 0.89, + 0.657 + ], + "angle": 0, + "content": "
PerceptMemoryImagineFPR95 ↓AUROC ↑
✓84.68%72.23%
✓83.13%72.75%
✓✓82.85%73.06%
✓✓81.07%73.94%
✓✓✓80.16%74.31%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.659, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Table 5. Ablation analysis of WFS for ODU-OD. Here, 'Percept', 'Memory', and 'Imagine' separately indicate multi-level perception, memory recorder, and unknown-feature generator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "i.e., multi-level perception, fragmented memory recorder, and imagination-based unknown-feature generator. In Table 5, we make an ablation analysis of the three modules. We can observe that employing dedicated perception, memory, and imagination-based generator could effectively reduce the impact of two-shift entanglement and improve the detection performance significantly. Particularly, when using the generator for synthesizing unknown features, the performance could be further enhanced. This shows that imitating humans to leverage perception and memory to imagine unknowns is an effective solution for addressing two-shift entanglement, promoting learning a robust and discriminative boundary between known and unknown objects." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4688" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.245, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.089, + 0.375, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.089, + 0.505, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.089, + 0.635, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.764, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.09, + 0.896, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.163, + 0.245, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.163, + 0.375, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.163, + 0.505, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.163, + 0.635, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.163, + 0.764, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.163, + 0.896, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.236, + 0.905, + 0.278 + ], + "angle": 0, + "content": "Figure 5. Memory vs. Imagination. The first and second rows respectively show the captured memory and reasoned imagination. Here, we directly project the memory codebook and imagination bank into the corresponding visual regions with three different styles. We can observe that for the same input, memory and imagination separately pay attention to different fragmented content." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.287, + 0.231, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.287, + 0.354, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.287, + 0.477, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.363, + 0.231, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.363, + 0.354, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.363, + 0.477, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.439, + 0.23, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.511, + 0.211, + 0.524 + ], + "angle": 0, + "content": "(a) Style Image" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.439, + 0.354, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.511, + 0.335, + 0.523 + ], + "angle": 0, + "content": "(b) Enhanced \\(E\\)" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.439, + 0.477, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.512, + 0.459, + 0.524 + ], + "angle": 0, + "content": "(c) Unknown \\(S_{u}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.526, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Figure 6. Visualization of the memory-enhanced map \\( E \\) and generated unknown map \\( S_{u} \\). For each feature map, the channels corresponding to the maximum value are selected for visualization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Analysis of Multi-level Perception. As shown in Fig. 1 and 2, the goal of multi-level perception is to construct a complete impression of objects. Here, we make an ablation analysis of three perception levels. We only change the perception level. Other operations and settings are kept unchanged. Table 6 shows the results. We can observe that employing different perception level affects the corresponding detection performance. Particularly, since the connective level considers the relationships between visual regions, using this level could enhance the performance effectively. When combining these three levels, the performance could be further improved. This indicates that multi-level perception is indeed conducive to building a complete observation of objects, strengthening the robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Analysis of Codeword Number \\( Q \\). For the fragmented memory recorder, we utilize \\( Q \\) codewords to preserve the information of known objects. Here, we analyze the \\( Q \\) impact. We observe that when \\( Q \\) is set to 16 and 32, the FPR95 value increases slightly, i.e., \\( 80.54\\% \\) and \\( 80.33\\% \\). The reason may be that utilizing fewer codewords could not sufficiently capture known characteristics. Utilizing more code" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.288, + 0.889, + 0.415 + ], + "angle": 0, + "content": "
LocalGlobalConnectiveFPR95 ↓AUROC ↑
✓81.19%73.04%
✓81.62%73.18%
✓81.37%73.21%
✓✓80.85%73.96%
✓✓80.63%74.12%
✓✓80.58%74.09%
✓✓✓80.16%74.31%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.417, + 0.905, + 0.446 + ], + "angle": 0, + "content": "Table 6. Ablation analysis of local, global, and connective perception perspectives. Other operations are kept unchanged." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.454, + 0.854, + 0.469 + ], + "angle": 0, + "content": "words increases parameters, resulting in overfitting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.905, + 0.62 + ], + "angle": 0, + "content": "In Fig. 5, we intuitively show the corresponding content captured by the built memory and reasoned imagination. Obviously, for the images with different styles, the memory recorder can effectively focus on fragmented object-related content, enhancing the robustness against the covariate-shift impact. Furthermore, the reasoned imagination mainly pays attention to background regions that involve plentiful visual information, which is instrumental in promoting the following generator to synthesize expected unknown features and attenuates the impact of two-shift entanglement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.906, + 0.741 + ], + "angle": 0, + "content": "Visualization Analysis. Fig. 6 shows three examples of memory-enhanced map and virtual unknown map. We can further observe that fusing the known memory could indeed strengthen object-related information. The synthesized unknown features mainly pay attention to background regions involving plentiful visual characteristics, enlarging the gap with the known features significantly. In supplementary material, we will give more analysis about our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.755, + 0.634, + 0.77 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This paper explores a challenging problem, i.e., ODU-OD, which aims to detect unknown objects within unseen target domains. To weaken the entangled impact of covariate and semantic shifts, we propose a new WFS method and imitate humans to leverage multi-level perception, fragmented memory, and reasoned imagination to recurrently generate virtual unknown features. Extensive experimental results on four tasks demonstrate the superiorities of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4689" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work is supported in part by the National Key R&D Program of China (No. 2023YFC3305600), Joint Fund of Ministry of Education of China (8091B022149, 8091B02072404), National Natural Science Foundation of China (62472333, 62132016, 62171343), and Natural Science Basic Research Program of Shaanxi (2020JC-23)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.221, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.289, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[2] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, pages 11580-11590, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.483, + 0.428 + ], + "angle": 0, + "content": "[3] Muhammad Sohail Danish, Muhammad Haris Khan, Muhammad Akhtar Munir, M Saqib Sarfraz, and Mohsen Ali. Improving single domain-generalized object detection: A focus on diversification and alignment. In CVPR, pages 17732-17742, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.483, + 0.472 + ], + "angle": 0, + "content": "[4] Xuefeng Du, Gabriel Gozum, Yifei Ming, and Yixuan Li. Siren: Shaping representations for detecting out-of-distribution objects. In NeurIPS, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.484, + 0.528 + ], + "angle": 0, + "content": "[5] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware object detection: Learning what you don't know from videos in the wild. In CVPR, pages 13678-13688, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[6] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. ICLR, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[8] Yanchen Guan, Haicheng Liao, Zhenning Li, Guohui Zhang, and Chengzhong Xu. World models for autonomous driving: An initial survey. arXiv preprint arXiv:2403.02622, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[9] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[10] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. *ICLR*, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[11] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In CVPR, pages 9591-9600, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[12] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.246, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.133 + ], + "angle": 0, + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. *ICLR*, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, pages 10951-10960, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[17] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, pages 4874-4883, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[18] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In CVPR, pages 5830-5840, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[19] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. arXiv preprint arXiv:2112.00374, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[21] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. *ICLR*, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[22] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. NeurIPS, 31, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[23] Deng Li, Among Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In CVPR, pages 17606-17615, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[24] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. *ICLR*, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[25] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In ICCV, pages 8771-8780, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[26] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation models? ECCV, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[27] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. NeurIPS, 33: 21464-21475, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[28] Dimity Miller, Lachlan Nicholson, Feras Dayoub, and Niko Sünderhauf. Dropout sampling for robust object detection in open-set conditions. In ICRA, pages 3243-3249, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[29] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, pages 464-479, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[30] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, pages 1863-1871, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[31] Bimsara Pathiraja, Malitha Gunawardhana, and Muhammad Haris Khan. Multiclass confidence and localization calibration for object detection. In CVPR, pages 19734-19743, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4690" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[32] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, pages 7263-7271, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.121, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurlPS, pages 91-99, 2015. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.177, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[34] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with gram matrices. In ICML, pages 8491-8501. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.219, + 0.483, + 0.26 + ], + "angle": 0, + "content": "[35] Eleanor Spens and Neil Burgess. A generative model of memory construction and consolidation. Nature Human Behaviour, pages 1-18, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.261, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[36] Weinan Sun, Madhu Advani, Nelson Spruston, Andrew Saxe, and James E Fitzgerald. Organizing memories for generalization in complementary learning systems. Nature neuroscience, 26(8):1438-1448, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.317, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[37] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. NeurIPS, 33:11839-11852, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.372, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[38] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. *ICLR*, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.414, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[39] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, pages 3219-3229, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.457, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[40] Samuel Wilson, Tobias Fischer, Feras Dayoub, Dimity Miller, and Niko Sunderhauf. Safe: Sensitivity-aware features for out-of-distribution object detection. In ICCV, pages 23565-23576, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.512, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[41] Among Wu and Cheng Deng. Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation. In CVPR, pages 847-856, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.554, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[42] Among Wu and Cheng Deng. Discriminating known from unknown objects via structure-enhanced recurrent variational autoencoder. In CVPR, pages 23956-23965, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.609, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[43] Among Wu, Yahong Han, Linchao Zhu, and Yi Yang. Instance-invariant domain adaptive object detection via progressive disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4178-4193, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.679, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[44] Among Wu, Da Chen, and Cheng Deng. Deep feature deblurring diffusion for detecting out-of-distribution objects. In ICCV, pages 13381-13391, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.721, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[45] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.777, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[46] Hang Zhang, Jia Xue, and Kristin Dana. Deep ten: Texture encoding network. In CVPR, pages 708-717, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[47] Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Learning placeholders for open-set recognition. In CVPR, pages 4401-4410, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[48] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[49] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 3" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "4691" + } + ] +] \ No newline at end of file diff --git a/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/b34e32e0-849d-4c9e-8711-6ce6de1dc89c_origin.pdf b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/b34e32e0-849d-4c9e-8711-6ce6de1dc89c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..63f29084812d61deffa6e79703d7968f4166b1f7 --- /dev/null +++ b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/b34e32e0-849d-4c9e-8711-6ce6de1dc89c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86f41c157851de9960fcde9a27ab9ce0144692bb68b3dcd7f6bc19b65ba0695c +size 3978706 diff --git a/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/full.md b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a0dff26990809ad6bd05089e6e2272c500b8f919 --- /dev/null +++ b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/full.md @@ -0,0 +1,366 @@ +# Percept, Memory, and Imagine: World Feature Simulating for Open-Domain Unknown Object Detection + +Among Wu, Cheng Deng* + +School of Electronic Engineering, Xidian University, Xi'an, China + +amwu@xidian.edu.cn, chdeng@mail.xidian.edu.cn + +# Abstract + +To accelerate the safe deployment of object detectors, we focus on reducing the impact of both covariate and semantic shifts. And we consider a realistic yet challenging scenario, namely Open-Domain Unknown Object Detection (ODU-OD), which aims to detect unknown objects in unseen target domains without accessing any auxiliary data. Towards ODU-OD, it is feasible to learn a robust discriminative boundary by synthesizing virtual features. Generally, perception, memory, and imagination are three essential capacities for human beings. Through multi-level perception and rich memory about known objects, the characteristics of unknown objects can be imagined sufficiently, enhancing the ability of discriminating known from unknown objects. Inspired by this idea, an approach of World Feature Simulation (WFS) is proposed, mainly consisting of a multi-level perception, memory recorder, and unknown-feature generator. Specifically, after extracting the features of the input, we separately employ a Mamba and Graph Network to obtain the global-level and connective-level representations. Next, a codebook containing multiple learnable codewords is defined to preserve fragmented memory of known objects. Meanwhile, we perform a modulated operation on the memory to form the imagination bank involving unknown characteristics. Finally, to alleviate the impact of lacking supervision data, based on the multi-level representation and imagination bank, a dedicated unknown-feature generator is designed to recurrently synthesize outlier features deviating from in-distribution (ID) objects. The significant performance gains on four different detection tasks demonstrate the superiorities of our method. The code will be released at https://github.com/AmingWu/WFS. + +# 1. Introduction + +With the rejuvenation of deep neural networks, object detection has achieved many advances [1, 14, 31-33] recently. + +Though these advances have obtained impressive performance, in real-world scenarios, they inevitably encounter both covariate shifts (e.g., test-time changes in image styles) and semantic shifts (e.g., test-time unknown classes), which limits their applications. + +Since there is no auxiliary data available, to alleviate the covariate-shift impact, existing researches [25, 41, 43, 48] mainly pay attention to enhancing the generalization ability by learning domain-invariant features. As for the semantic-shift challenge, most existing approaches [4, 6, 26, 42, 44] explore synthesizing diverse virtual features to strengthen the capability of discriminating unknown objects. Although these methods have been demonstrated to be effective, they are only concerned with a single-shift case and could not cope with a more difficult scenario involving both covariate and semantic shifts, e.g., for an autonomous car, the detector may encounter an unknown object in unseen adverse weather, resulting in a safety risk. + +To promote the safe deployment of detectors, we propose a realistic yet challenging task, namely Open-Domain Unknown Object Detection (ODU-OD). As shown in Fig. 1, given one source dataset including multiple pre-defined categories for training, the goal of ODU-OD is to accurately detect unknown out-of-distribution (OOD) objects in multiple unseen target domains with different styles. Since ODU-OD contains both covariate and semantic shifts and fails to access any auxiliary data for training, existing single-shift methods can not be directly utilized to tackle this task. + +In essence, towards covariate and semantic shifts, it is critical to learn the features that lead to a robust discriminative boundary. Particularly, enhancing the robustness could attenuate the covariate-shift impact and improve the generalization ability. And strengthening the discrimination is instrumental in accurately recognizing known and unknown objects. Moreover, recent researches [8-10, 35, 36] indicate that based on the current observations and formed memory, human beings could produce rich imaginations, improving the adaptation ability for unknown situations. Therefore, we explore simulating humans to leverage plentiful perception and fragmented memory to imagine unknowns for training + +![](images/f4352d79a6183dde2dbda3c7fff3f56d4ad25c0021c7392702346c3524c001bf.jpg) +Figure 1. World Feature Simulation for Open-Domain Unknown Object Detection, which imitates humans to synthesize virtual unknown features and consists of a multi-level perception, memory recorder, and dedicated unknown-feature generator. We first calculate the multi-level representation from the local, global, and connective levels. Then, a memory codebook is constructed to preserve the local object information (as shown in pink boxes). Next, a modulated operation is performed on the fragmented memory to form the imagination bank involving unknown characteristics. Finally, we design an unknown-feature generator to recurrently synthesize virtual unknown features deviating from known objects, enhancing the ability of detecting unknown objects within unseen domains (as shown in green boxes). + +![](images/c11afc1c9c3d4294cc580a1ed381ba73e832eaddc4238b1e1cc4be4311e5f1a4.jpg) + +![](images/515fd2a86094b91420fcf94427c11f7dd34407f231242e8de3f80d4eabcf53c8.jpg) + +(as shown in Fig. 1), enhancing the ability of distinguishing unknown objects within unseen domains. + +Specifically, as shown in Fig. 1, a method of World Feature Simulation (WFS) is proposed, mainly consisting of a multi-level perception, memory recorder, and unknown-feature generator. For a specific object, in order to deepen its perception, humans usually observe the object from multiple different perspectives, e.g., local and global levels. To this end, given an ID input, we first utilize a convolutional backbone [13] to extract its local features. Then, a Mamba [7, 12, 49] and a Graph network [19, 38] are separately performed on the local features to obtain the global and connective features. And the local, global, and connective features are concatenated to form the multi-level representation. Next, we define a codebook containing multiple learnable codewords to preserve fragmented memory of the known objects. By fusing the memory into the current multi-level representation, the robustness against covariateshift risk could be enhanced effectively. + +In Fig. 1, to achieve the ability of imagining unknowns, a modulated operation is performed on the known codebook to obtain the imagination bank. And we design a contrastive constraint to promote the bank to involve plentiful unknown characteristics. Finally, a dedicated unknown-feature generator is proposed to leverage the multi-level representation and imagined unknown characteristics to recurrently synthesize virtual unknown features deviating from ID inputs, which is conducive to learning a robust discriminative boundary between known and unknown objects within unseen domains. Experimental results on four different detection tasks demonstrate the superiorities of our method. + +The contributions are summarized as follows: + +(1) To promote the deployment of detectors, we propose a new task, namely Open-Domain Unknown Object Detection (ODU-OD), aiming to detect unknown objects in unseen target domains without accessing any auxiliary data. + +(2) To tackle ODU-OD involving both covariate and semantic shifts, we imitate humans to explore a new framework of World Feature Simulation, which aims to synthesize virtual unknown features for training and contains a multi-level perception, fragmented memory recorder, and unknown-feature generator. +(3) We build an Open-Domain Dataset for unknown object detection to verify our method WFS, consisting of three different styles [20], i.e., 'Acrylic', 'Purple-Brush', and 'Sketch'. Meanwhile, our method is further evaluated on three single-shift tasks, i.e., OOD-OD [6], open-set object detection [11], and Single-DGOD [41]. The significant performance gains over multiple baselines demonstrate the effectiveness of our method. Meanwhile, visualization analyses also show that the built memory and reasoned imagination could indeed capture fragmented visual content. + +# 2. Related Work + +Covariate-shift Detection. To improve the generalization of detectors, a task of Single-DGOD [41] is proposed, aiming to learn an object detector that performs well on multiple unseen target domains with only one source domain for training. Some works [23, 39, 41] utilize feature disentanglement and large-scale models to extract domain-invariant representations. Moreover, Liu et al. [26] further consider the biased problem and propose a causal-based method to learn unbiased generalized representations. And Danish et al. [3] explore diversifying the source domain and aligning detections based on prediction confidence. These methods have been verified on multiple benchmarks. However, they only consider the covariate-shift scenario and could not detect unknown objects. + +Semantic-shift Detection. Recently, a task of unsupervised OOD-OD [6] is proposed, whose goal is to detect unknown objects without relying on any OOD data. To alle + +![](images/fcd8e9bf4f90cde9792c69866df3a77d2619c4df5ccb5257398cac02da8df11c.jpg) +Figure 2. The details of World Feature Simulation for ODU-OD. To enhance the robustness against covariate shift, a multi-level perception is used to extract the representation $F_{v}$ from the local, global, and connective levels. Meanwhile, we design a memory recorder to preserve rich content about known objects. By calculating a memory-enhanced map, the robustness could be further strengthened. Next, through a modulated operation, we can reason the imagination bank involving plentiful unknown characteristics. Finally, a dedicated unknown-feature generator is proposed to utilize the built memory and imagination characteristics to recurrently synthesize virtual unknown features, improving the ability of detecting unknown objects within unseen domains. + +viate the impact of lacking OOD data for supervision, VOS [6] and DFDD [44] explore synthesizing virtual OOD features that slightly deviate from the distribution of ID data. Besides, the work [5] further learns unknown-aware knowledge from external videos, which does not match the setting of unsupervised OOD-OD. Finally, Du et al. [4] present to use a distance-based mechanism to shape the learned representations. Meanwhile, Wilson et al. [40] attempt to enhance the sensitivity of the learned features, which is beneficial for improving the performance of detecting OOD objects. The work [26] struggles to leverage generative models trained on large-scale open-set data to synthesize OOD samples. Although these methods have been demonstrated to be effective, they only consider the semantic-shift problem and ignore the covariate-shift impact, which limits their applications to unknown objects within unseen domains. + +Different from the above methods, in this paper, we explore a more difficult problem involving both covariate and semantic shifts, i.e., Open-Domain Unknown Object Detection. And a new approach of World Feature Simulation is proposed for synthesizing unknown features for training, consisting of a multi-level perception, fragmented memory recorder, and unknown-feature generator. Experimental results on multiple datasets indicate its superiorities. + +# 3. World Feature Simulation + +For ODU-OD, during training, we can only access single source-domain data with some pre-defined categories. During test, the detector is required to accurately localize and recognize unknown objects within multiple unseen do + +mains. To reduce the entangled impact of covariate and semantic shifts, as shown in Fig. 2, we design a World Feature Simulation for synthesizing unknown features. + +# 3.1. Observing Objects from Multiple Levels + +Generally, in order to deepen the perception of a specific object, human beings usually observe the object from multiple different perspectives, which is beneficial for constructing a complete impression. Therefore, in Fig. 2, we explore imitating this capability based on local, global, and connective levels. Specifically, we exploit the widely used object detector, i.e., Faster R-CNN [14, 33], as the basic model. + +Given an ID input image, a convolutional network, e.g., ResNet [13], is taken as the backbone to extract the local feature map $F_{l} \in \mathbb{R}^{w \times h \times c}$ , where $w$ , $h$ , and $c$ separately represent width, height, and the number of channels. Each element in $F_{l}$ describes a local region of the input. Besides the local level, human beings usually intensify their impression by global observation and discovering the relations between local regions. Hence, we first unfold $F_{l}$ to a sequence of visual elements, i.e., $D = \{d_{i} \in \mathbb{R}^{c}, i = 1, \dots, wh\}$ . Then, one-layer vision mamba [49] is used to process $D$ , which outputs $H = \{h_{i} \in \mathbb{R}^{c}, i = 1, \dots, wh\}$ . Finally, $H$ is reshaped to the original size to obtain the global representation $F_{g} \in \mathbb{R}^{w \times h \times c}$ [12]. Next, we use one-layer graph convolution [19] to further capture the semantic connections between local regions. Concretely, we first reshape $F_{l}$ to $\tilde{F} \in \mathbb{R}^{wh \times c}$ . Then, an adjacency matrix for an undirected graph is defined as $\mathbf{A} = \text{softmax}(\tilde{F}\tilde{F}^{\top}) + I_{wh}$ , where $I_{wh}$ indicates the identity matrix: + +$$ +\mathcal {R} = \mathbf {A} \tilde {F}, \quad F _ {c} = r e l u (W _ {c} * \tilde {\mathcal {R}} + b _ {c}), \qquad (1) +$$ + +where $\tilde{\mathcal{R}}\in \mathbb{R}^{w\times h\times c}$ denotes the reshaped result of $\mathcal{R}$ . $W_{c}\in \mathbb{R}^{3\times 3\times c\times c}$ and $b_{c}\in \mathbb{R}^{c}$ indicate the trainable parameters. $F_{c}\in \mathbb{R}^{w\times h\times c}$ represents the connective feature. Finally, $F_{l},F_{g}$ , and $F_{c}$ are concatenated as the multi-level representation $F_{v}\in \mathbb{R}^{w\times h\times c}$ , where $F_{v} = \phi ([F_{l},F_{g},F_{c}])$ and $\phi (\cdot)\in \mathbb{R}^{1\times 1\times 3c\times c}$ is utilized to transform the number of channels. Compared with $F_{l},F_{v}$ involves sufficient object-related information from three levels, which strengthens the robustness against the covariate shift. + +# 3.2. Mining Fragmented Memory and Imagination + +In general, human beings could exploit familiar fragmented memory to improve the robustness of recognizing known objects. Then, based on the memory, human beings could produce rich imagination, enhancing the ability of discriminating unknowns. To simulate this capability, we present a memory recorder [46] to preserve the information of known objects. Particularly, a learnable memory codebook is denoted as $M = \{m_j \in \mathbb{R}^c, j = 1, \dots, Q\}$ that contains $Q$ codewords. Meanwhile, by means of a weight $a_{ij}$ , each element $f_i \in \mathbb{R}^c$ of the multi-level representation $F_v$ can be assigned to each codeword $m_j$ . And the residual vector is represented as $r_{ij} = f_i - m_j$ , where $i = 1, \dots, wh$ . Here, the residual operation is conducive to promoting the memory codewords to effectively capture the corresponding content of local visual regions. The processes of capturing memory are shown as follows: + +$$ +v _ {j} = \sum_ {i = 1} ^ {w h} a _ {i j} r _ {i j}, \quad a _ {i j} = \frac {\exp \left(- k _ {j} \| r _ {i j} \| ^ {2}\right)}{\sum_ {j = 1} ^ {Q} \exp \left(- k _ {j} \| r _ {i j} \| ^ {2}\right)}, \tag {2} +$$ + +where $k_{j}$ represents the learnable smoothing factor for the codeword $m_j$ . Finally, $\mathcal{V} \in \mathbb{R}^{Q \times c}$ is the output result. The concatenation of $F_v$ and $\mathcal{V}$ is used to calculate the memory-enhanced map $E \in \mathbb{R}^{w \times h \times c}$ , i.e., $E = \psi([F_v, \tilde{\mathcal{V}}W_0 + b_0])$ , where $\tilde{\mathcal{V}} \in \mathbb{R}^{1 \times Qc}$ is the reshaped result of $\mathcal{V}$ . Meanwhile, $W_0 \in \mathbb{R}^{Qc \times c}$ and $b_0 \in \mathbb{R}^c$ are parameters of a fully-connected layer. $\psi(\cdot)$ includes two convolutional layers to transform the concatenated result. Through fusing memory, the information capacity of $E$ could be increased, improving the robustness of object localization. + +Next, we further explore how to leverage the memory to reason the imagination that guides the synthesis of virtual unknown features. Particularly, we assume that the imagination should satisfy two conditions: (1) Imagination and memory should belong to different spaces. (2) Imagination mainly describes unknowns. It could not well represent known visual content. To this end, based on the memory $M \in \mathbb{R}^{Q \times c}$ , a modulated operation is performed on $M$ to obtain the imagination bank $\mathcal{I} \in \mathbb{R}^{Q \times c}$ , i.e., $\mathcal{I} = \gamma \odot M + \beta$ , where $\gamma$ and $\beta$ are learnable parameters for channel-wisely scaling and shifting $M$ . Next, we separately utilize $M$ and + +![](images/19462839cd050a9b5595f959dfc073517ed97763d6518ff4176c912636cbd353.jpg) +Figure 3. Unknown-feature generator. Based on $F_{m} \in \mathbb{R}^{w \times h \times c}$ , a variational operation is performed on $F_{m}$ to obtain $Z \in \mathbb{R}^{w \times h \times c}$ . Then, we exploit the reasoned imagination bank to re-represent $Z$ , which outputs $G \in \mathbb{R}^{w \times h \times c}$ . Next, $Z$ and $G$ are unfolded into a sequence with $wh$ elements. Finally, a dedicated recurrent generator including three inputs and one output is presented to synthesize virtual unknown features, improving the discrimination. + +$\mathcal{I}$ to re-represent $F_{v}$ : + +$$ +F _ {m} = \operatorname {s o f t} x \left(\tilde {F} _ {v} M ^ {\top}\right) M, F _ {i} = \operatorname {s o f t} x \left(\tilde {F} _ {v} \mathcal {I} ^ {\top}\right) \mathcal {I}, \tag {3} +$$ + +where 'softx' denotes the softmax operation. $\tilde{F}_v\in \mathbb{R}^{wh\times c}$ is the reshaped result of $F_{v}$ . $F_{m}\in \mathbb{R}^{w\times h\times c}$ and $F_{i}\in \mathbb{R}^{w\times h\times c}$ indicate the re-represented results based on the fragmented memory and imagination. Finally, a contrastive loss is exploited to enlarge the gap between the memory recorder and imagination bank: + +$$ +\mathcal {L} _ {i m} = - \log \frac {\exp \left(\sin \left(F _ {m} , F _ {v}\right) / \tau\right)}{\sum_ {j = 0} ^ {1} \exp \left(\sin \left(B [ j ] , F _ {v}\right) / \tau\right)}, \tag {4} +$$ + +where $B = [F_m, F_i]$ . $\mathrm{sim}(F_m, F_v)$ denotes the average of the cosine similarity between all corresponding elements of feature map $F_m$ and $F_v$ . $\tau$ is a hyper-parameter, which is set to 1.0 in the experiments. Obviously, by minimizing the contrastive loss, the distance between $F_m$ and $F_v$ could be narrowed, promoting the memory recorder to capture sufficient content about known objects. Instead, the distance between $F_i$ and $F_v$ is enlarged, forcing the reasoned imagination bank to contain plentiful information deviating from knowns. Next, the imagination bank is utilized to generate expected unknown features. + +# 3.3. Generating Unknown Features for Training + +For ODU-OD, to attenuate the impact of lacking unknown data for supervision, we design an unknown-feature generator to synthesize outlier features involving unknown characteristics. Specifically, two convolutional networks $W_{\mu}$ and $W_{V}$ are performed on $F_{m}$ to estimate the corresponding means $\mu \in \mathbb{R}^{w\times h\times c}$ and variances $V \in \mathbb{R}^{w\times h\times c}$ . Then, an encoding operation is performed to obtain $Z \in \mathbb{R}^{w\times h\times c}$ . Meanwhile, we exploit the reasoned imagination bank $\mathcal{I}$ to re-represent $Z$ . The processes are shown as follows: + +$$ +\mu = W _ {\mu} * F _ {m}, \quad V = W _ {V} * F _ {m}, \tag {5} +$$ + +$$ +Z = \mu + \epsilon \odot V, \quad G = \text {s o f t m a x} (\tilde {Z} \mathcal {I} ^ {\top}) \mathcal {I}, +$$ + +where $\tilde{Z} \in \mathbb{R}^{wh \times c}$ is the reshaped result of $Z$ . $\epsilon$ represents Gaussian noise sampled from $\mathcal{N}(0, I)$ . $G \in \mathbb{R}^{w \times h \times c}$ is the re-represented result based on the imagination bank $\mathcal{I}$ . Since the imagination could not well express the knowns, $G$ contains plentiful information deviating from known visual content, which is instrumental in promoting the generation of virtual unknown features. + +Finally, we present a dedicated generator to recurrently synthesize expected unknown features. As shown in Fig. 3, the generator contains three inputs and one output. Particularly, at each time, the generator takes $Z_{t}$ , $G_{t}$ , and the previous output $S_{t-1}$ as the input, and synthesizes the current output $S_{t}$ . The generating processes are shown as follows: + +$$ +\begin{array}{l} z _ {t} = \sigma \left(W _ {z z} * Z _ {t} + W _ {o z} * G _ {t} + W _ {s z} * S _ {t - 1} + b _ {z}\right), \\ g _ {t} = \sigma \left(W _ {z g} * Z _ {t} + W _ {o g} * G _ {t} + W _ {s g} * S _ {t - 1} + b _ {g}\right), \\ \tilde {S} _ {t} = \operatorname {T a n h} \left(W _ {s} * Z _ {t} + W _ {h} * \left(g _ {t} \odot S _ {t - 1} + (1 - g _ {t}) \odot G _ {t}\right)\right), \\ S _ {t} = \left(1 - z _ {t}\right) \odot S _ {t - 1} + z _ {t} \odot \hat {S} _ {t}, \tag {6} \\ \end{array} +$$ + +where $Z_{t}\in \mathbb{R}^{c}$ and $G_{t}\in \mathbb{R}^{c}$ are the current element of $Z$ and $G$ , $t = 1,\dots ,wh$ . And $W_{zz}$ , $W_{oz}$ , $W_{sz}$ , $b_{z}$ , $W_{zg}$ , $W_{og}$ , $W_{sg}$ , $b_{g}$ , $W_{s}$ , and $W_{h}$ represent the trainable parameters. Through this generating mechanism, the synthesized features could sufficiently retain the information in previous steps, improving the content continuity of the generated features. Finally, the generated sequence $\{S_1,S_2,\dots S_{wh}\}$ are concatenated to obtain unknown feature $S_{u}\in \mathbb{R}^{w\times h\times c}$ . We still utilize a contrastive loss to expand the difference between known and the synthesized unknown features: + +$$ +\mathcal {L} _ {u n k} = - \log \frac {\exp \left(\sin \left(F _ {m} , F _ {v}\right) / \tau\right)}{\sum_ {j = 0} ^ {1} \exp \left(\sin \left(A [ j ] , F _ {v}\right) / \tau\right)}, \tag {7} +$$ + +where $A = [F_m, S_u]$ . $\tau$ is set to 1.0 in the experiments. Since the generation process involves perception, memory, and imagination, using the contrastive loss could not only promote $S_u$ to contain plentiful unknown information but also facilitate $M$ and $\mathcal{I}$ to separately capture rich known and unknown characteristics. + +# 3.4. WFS-Driven Unknown-Object Detection + +Lastly, the memory-enhanced map $E$ is taken as the input of Region-Proposal Network (RPN) [14, 33] to obtain a series of object proposals $\mathcal{O}$ . Meanwhile, based on $\mathcal{O}$ , RoI-Alignment followed by RoI-Feature extraction [14] is performed on $E$ to obtain $O_{\mathrm{in}} \in \mathbb{R}^{p \times n}$ , where $p$ and $n$ denote the number of proposals and channels, respectively. Then, $O_{\mathrm{in}}$ is input to the object classifier and regressor to calculate the classification loss $\mathcal{L}_{\mathrm{cls}}$ and localization loss $\mathcal{L}_{\mathrm{loc}}$ , i.e., $\mathcal{L}_{\mathrm{det}} = \mathcal{L}_{\mathrm{cls}} + \mathcal{L}_{\mathrm{loc}}$ . By optimizing $\mathcal{L}_{\mathrm{det}}$ , the memory recorder could be promoted to capture plentiful information about known objects, enhancing the robustness. + +Next, we further enlarge the gap between the known and synthesized unknown features from the instance perspective. Concretely, based on $\mathcal{O}$ , RoI-Alignment followed by RoI-Feature extraction is performed on $S_{u}$ to obtain $O_{\mathrm{unk}} \in \mathbb{R}^{p \times n}$ . Meanwhile, we define a loss function $\mathcal{L}_{\mathrm{dis}}$ to maximize the $KL$ -divergence between $O_{\mathrm{unk}}$ and $O_{\mathrm{in}}$ , i.e., $\mathcal{L}_{\mathrm{dis}} = \mathrm{KL}[q(O_{\mathrm{unk}}), q(O_{\mathrm{in}})]$ , where $q(\cdot)$ represents the probability distribution. Besides, to attain the goal of distinguishing unknown objects from known ID objects, $O_{\mathrm{unk}}$ and $O_{\mathrm{in}}$ are used to compute an uncertainty loss [6], regularizing the detector to produce a low score for the ID object features, and a high score for the virtual unknown features: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {u n t y}} = \mathbb {E} _ {x \sim O _ {\text {i n}}} [ - \log \frac {\exp^ {- \mathbf {E} (x)}}{1 + \exp^ {- \mathbf {E} (x)}} ] + \tag {8} \\ \mathbb {E} _ {v \sim O _ {\mathrm {u n k}}} [ - \log \frac {1}{1 + \exp^ {- \mathbf {E} (v)} ], \\ \end{array} +$$ + +where $\mathbf{E}(\cdot)$ is the object-level energy score [6, 27, 44]. The overall objective is shown as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c o n t r a s t}} = \mathcal {L} _ {i m} + \mathcal {L} _ {u n k} - \mathcal {L} _ {\text {d i s}}, \\ \mathcal {L} _ {i m} = \mathcal {L} _ {u n k} + \mathcal {L} _ {u n k} + \lambda_ {i} \mathcal {L} _ {i m}. \end{array} \tag {9} +$$ + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d e t}} + \alpha \cdot \mathcal {L} _ {\text {c o n t r a s t}} + \lambda \cdot \mathcal {L} _ {\text {u n t y}}, +$$ + +where $\alpha$ and $\lambda$ are two hyper-parameters, which are set to 0.001 and 0.1 in the experiments. + +# 3.5. Inference for Unknown Object Detection + +During test, we only calculate the uncertainty score for unknown object detection [6, 44]. Specifically, for a predicted bounding box $\mathbf{b}$ , the processes of discriminating unknown objects are shown as follows: + +$$ +\mathcal {E} = \frac {\exp^ {- \mathbf {E} (\mathbf {b})}}{1 + \exp^ {- \mathbf {E} (\mathbf {b})}}, \quad \mathcal {C} (\mathbf {b}) = \left\{ \begin{array}{l l} 0 & \text {i f} \mathcal {E} < \delta , \\ 1 & \text {i f} \mathcal {E} \geq \delta . \end{array} \right. \tag {10} +$$ + +For the output of the classifier $\mathcal{C}(\cdot)$ , we utilize the threshold mechanism [6] to distinguish known objects (the result is 1) from unknown objects (the result is 0). The threshold $\delta$ is commonly set to 0.95 so that a high fraction of known ID objects are correctly classified. In supplementary material, we will show more details of the proposed WFS algorithm and give more discussion. + +# 4. Experiments + +In the experiments, our method is evaluated on four different detection tasks, i.e., ODU-OD, OOD-OD [6], open-set object detection (OSOD) [11], and Single-DGOD [41]. Particularly, we construct an Open-Domain dataset for ODU-OD, which utilizes Clipstyler [20] to transform testing data of OOD-OD [6] to three different styles, i.e., 'Acrylic', 'Purple-Brush', and 'Sketch'. + +Implementation Details. We utilize Faster R-CNN [33] with RoI-Alignment layer [14] as the basic detection model. + +
Source DataMethodStyle: Acrylic +MS-COCO/OpenImagesStyle: Purple-Brush +MS-COCO/OpenImagesStyle: Sketch +MS-COCO/OpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
VOCSR-VAE [42]88.25/84.6371.02/69.6487.76/69.9367.42/79.1889.67/74.8266.47/77.04
DFDD [44]87.68/85.1270.85/69.2187.24/70.0867.23/79.4689.83/74.9266.13/76.82
VOS [6]88.73/84.7171.13/69.5688.66/70.1767.06/79.3389.76/74.6166.50/77.63
VOS + WFS80.16/77.5274.31/73.8580.07/60.6870.86/83.8782.16/67.2369.87/81.28
SAFE [40]86.62/47.8363.47/78.5186.31/39.6261.28/86.8287.56/55.2860.71/83.48
SAFE + WFS79.38/40.1667.28/81.6979.23/33.8566.47/88.7380.49/48.6165.82/84.17
BDDSR-VAE [42]64.72/53.9872.34/78.1166.43/57.2870.69/76.2469.36/59.2769.17/75.25
DFDD [44]64.08/53.2672.48/78.3665.91/56.7771.03/76.4868.74/58.8969.22/75.44
VOS [6]65.71/54.7371.96/77.8068.50/58.7970.18/75.8370.80/60.7169.05/74.97
VOS + WFS56.48/45.4275.87/82.1859.64/47.8874.34/79.8761.01/51.9473.78/79.36
SAFE [40]54.82/40.1873.71/82.6757.63/43.9772.34/80.8359.44/46.2871.27/79.53
SAFE + WFS45.91/32.0378.35/85.1447.78/34.4176.93/84.2650.26/37.1875.54/82.75
+ +Table 1. Unknown-object detection results under unseen open domains. $\uparrow$ denotes larger values are better and $\downarrow$ represents smaller values are better. 'VOS+WFS' and 'SAFE+WFS' indicate our method is directly plugged into VOS [6] and SAFE [40]. We can see that our method could reduce the impact of both covariate and semantic shifts. + +
SourceMethodFPR95 ↓AUROC ↑
OOD: MS-COCO / OpenImages
VOCMSP [15]70.99 / 73.1383.45 / 81.91
ODIN [24]59.82 / 63.1482.20 / 82.59
Mahalanobis [22]67.73 / 65.4181.45 / 81.48
Gram matrices [34]62.75 / 67.4279.88 / 77.62
Energy score [27]56.89 / 58.6983.69 / 82.98
Generalized ODIN [16]59.57 / 70.2883.12 / 79.23
CSI [37]59.91 / 57.4181.83 / 82.95
GAN-synthesis [21]60.93 / 59.9783.67 / 82.67
SIREN-vMF [4]64.68 / 68.5385.36 / 82.78
SIREN-KNN [4]47.45 / 50.3889.67 / 88.80
SR-VAE [42]42.17 / 46.2690.28 / 87.89
DFDD [44]41.34 / 44.5290.79 / 88.65
VOS (Baseline) [6]47.53 / 51.3388.70 / 85.23
VOS + WFS37.91 / 41.3691.47 / 89.78
SAFE (Baseline) [40]47.40 / 20.0680.30 / 92.28
SAFE + WFS37.48 / 11.2785.62 / 94.91
BDDMSP [15]80.94 / 79.0475.87 / 77.38
ODIN [24]62.85 / 58.9274.44 / 76.61
Mahalanobis [22]55.74 / 47.6985.71 / 88.05
Gram matrices [34]60.93 / 77.5574.93 / 59.38
Energy score [27]60.06 / 54.9777.48 / 79.60
Generalized ODIN [16]57.27 / 50.1785.22 / 87.18
CSI [37]47.10 / 37.0684.09 / 87.99
GAN-synthesis [21]57.03 / 50.6178.82 / 81.25
SR-VAE [42]32.23 / 21.8190.69 / 93.55
DFDD [44]30.71 / 22.6790.74 / 92.48
VOS (Baseline) [6]44.27 / 35.5486.87 / 88.52
VOS + WFS28.66 / 20.3992.37 / 94.13
SAFE (Baseline) [40]32.56 / 16.0488.96 / 94.64
SAFE + WFS21.84 / 7.8393.41 / 96.85
+ +Table 2. The performance of unsupervised OOD-OD [6] that assumes the training and test data are from the same domain. We can see that our method can reduce the impact of semantic shift. + +ResNet-50 [13] is taken as the backbone. For the memory recorder, the number $Q$ of the codewords is set to 24. All the experiments are trained using the standard SGD optimizer with a learning rate of 0.02. In supplementary material, we + +will give the details of the used datasets and metrics. + +# 4.1. Analysis of ODU-OD and OOD-OD + +Table 1 and 2 separately illustrate the ODU-OD and OOD-OD performance. Here, we only compare with state-of-the-art methods. We can observe that though existing methods [6, 40, 42, 44] could effectively attenuate the semantic-shift impact, their performance drops severely when encountering the covariate shift (e.g., changes in image styles). Plugging our method into existing methods [6, 40] significantly improves their performance whether objects are within open or closed domains. This demonstrates that for the entangled case of covariate and semantic shifts, imitating humans to construct perception, memory, and imagination and further generating unknown features for training could simultaneously strengthen the robustness and discrimination. + +In Fig. 4, we show some detection examples with different styles. We can observe that plugging our method could indeed attenuate both covariate and semantic shifts, improving the detection accuracy under open domains. + +# 4.2. Analysis of OSOD and Single-DGOD + +To further demonstrate that our method could attenuate the impact of two-shift entanglement, we verify our method on two dedicated tasks, i.e., OSOD [11] and Single-DGOD [41]. We directly plug our method into three existing methods [11, 39, 41] and do not change the training processes. Table 3 and 4 show the corresponding results. We can observe that plugging our method could improve the detection performance effectively. This further demonstrates that our method WFS could synthesize expected unknown features that are beneficial for alleviating the impact of covariate and semantic shifts, which promotes the detector to learn a ro + +![](images/7b2ff19fd9c9c63f255a5ed7337cb2a7593e25ac2bddc451c4cee1245b322c1b.jpg) + +![](images/03ce489de9365809201dee26c311131efe519a14d8bad60b6af0b47afa397bd9.jpg) + +![](images/0f571e54a2fe6baebe687bc822f36b02e781405d37165c19195ca79a7d18a65d.jpg) + +![](images/3d632e65771615928e9fdf297c954f38cb86f536d81123137e4962b804e91bba.jpg) + +![](images/14095d82bd384930a08417bccb037f9a52c615af9c56c5581a1d8dbe72e0db64.jpg) + +![](images/6df42ef3a9e1e6682df6dad44219fb311a9c850fe290af09d2e5a3effdbe5399.jpg) +(a) Acrylic + +![](images/4e0c16a0db04979a2994d5fa137b4eb25c74b4298a570a3c1c2461fd60c81be5.jpg) +(b) Purple-Brush + +![](images/d95fdf15a5be65475dbf5e31bd3b476704c76723c1cc16a4607b1cdbd2f346d1.jpg) +(c) Sketch +Figure 4. Detection results on target images with multiple different styles. The first and second rows respectively indicate results based on VOS [6] and our method. The in-distribution source dataset is BDD-100k [45]. 'Arcylic', 'Purple-Brush', and 'Sketch' represent the rendered target styles [20]. 'Nature' is the same style as the source dataset. + +![](images/e075021f3e9080f07d6de26d0a6dc9f7571f718ecaeb057d3cf40e4b2948b008.jpg) +(d) Nature + +![](images/fcaf5632b835e4236bd53303173211647c73603af2a8130f389be4f0c2f4fc18.jpg) +(e) Nature + +
MethodVOCVOC-COCO-20VOC-COCO-40VOC-COCO-60
\( \mathrm{mAP}_{\mathcal{K}\uparrow} \)\( WI_{\downarrow} \)\( AOSE_{\downarrow} \)\( mAP_{\mathcal{K}\uparrow} \)\( AP_{\mathcal{U}\uparrow} \)\( WI_{\downarrow} \)\( AOSE_{\downarrow} \)\( mAP_{\mathcal{K}\uparrow} \)\( AP_{\mathcal{U}\uparrow} \)\( WI_{\downarrow} \)\( AOSE_{\downarrow} \)\( mAP_{\mathcal{K}\uparrow} \)\( AP_{u\uparrow} \)
FR-CNN [33]80.1018.391511858.45022.742339155.26018.492547255.830
\( FR-CNN^{\dagger} \)[33]80.0118.831194157.91023.241825754.77018.721956655.340
PROSER [47]79.6819.161303557.6610.9224.151983154.667.6219.642132255.203.25
ORE [18]79.8018.181281158.252.6022.401975255.301.7018.352141555.470.53
DS [28]80.0416.981286858.355.1320.861977555.313.3917.222192155.771.25
DFDD [44]80.2612.731072760.2116.2214.921363857.9812.0912.571666857.515.12
OpenDet [11]80.0214.951128658.7514.9318.231680055.8310.5814.241825056.374.36
OpenDet + Ours80.5311.981004360.5317.1914.181267058.6212.4711.831623158.025.63
+ +Table 3. Performance analysis of OSOD. We report close-set performance $(\mathrm{mAP}_{\mathcal{K}})$ on VOC, and both close-set $(\mathrm{mAP}_{\mathcal{K}})$ and open-set (WI, AOSE, $\mathrm{AP}_{\mathcal{U}}$ ) performance of different methods on VOC-COCO-{20, 40, 60}. Here, '20', '40', and '60' indicate that the testing COCO images separately contain 20, 40, and 60 non-VOC classes. '†' means a higher score threshold (i.e., 0.1) for testing. + +
MethodNSDRNRDF
SW [30]33.426.313.730.8
IBN-Net [29]32.126.114.329.6
IterNorm [17]29.622.812.628.4
ISW [2]33.225.914.131.8
S-DGOD [41]36.628.216.633.5
S-DGOD + Ours38.230.418.335.7
Clip-gap [39]36.932.318.738.5
Clip-gap + Ours38.534.120.240.1
+ +bust discriminative boundary and accelerates the safe deployment of the model. + +# 4.3. Ablation and Visualization Analysis + +We utilize PASCAL VOC as the source data and MS-COCO with 'Acrylic' style as the target data to perform an ablation analysis. Our method is directly plugged into VOS [6]. + +Analysis of WFS. As shown in Fig. 2, our method aims to emulate humans and mainly consists of three modules, + +Table 4. Single-DGOD analysis. 'NS', 'DR', 'NR', and 'DF' separately represent Night-Sunny, Daytime-Rainy, Night-Rainy, and Daytime-Foggy weather. mAP is taken as the metric. + +
PerceptMemoryImagineFPR95 ↓AUROC ↑
✓84.68%72.23%
✓83.13%72.75%
✓✓82.85%73.06%
✓✓81.07%73.94%
✓✓✓80.16%74.31%
+ +Table 5. Ablation analysis of WFS for ODU-OD. Here, 'Percept', 'Memory', and 'Imagine' separately indicate multi-level perception, memory recorder, and unknown-feature generator. + +i.e., multi-level perception, fragmented memory recorder, and imagination-based unknown-feature generator. In Table 5, we make an ablation analysis of the three modules. We can observe that employing dedicated perception, memory, and imagination-based generator could effectively reduce the impact of two-shift entanglement and improve the detection performance significantly. Particularly, when using the generator for synthesizing unknown features, the performance could be further enhanced. This shows that imitating humans to leverage perception and memory to imagine unknowns is an effective solution for addressing two-shift entanglement, promoting learning a robust and discriminative boundary between known and unknown objects. + +![](images/a7816d702ba4f333d4007aecd00c08c141b0deb97405e86bc8dda6eb1ef482a3.jpg) + +![](images/6af2c9e2e71696338163f5ef456cd442351d7150684119eb230452b80fedd699.jpg) + +![](images/840c243ca0081aedf46179da98ceb1ee24400bc515418db81e44cd1978c8c367.jpg) + +![](images/fb78ab8bb3d46871eae72cfd6885f5e939b2380c9a97d9ce5cb2a66c31c06a7b.jpg) + +![](images/381157da76156e8862d8aee7b5fc77e89d9c657115fdad0cc52995a0e785afdd.jpg) + +![](images/02b4052fedae00e0369823b996a2f6e3f3b63cc6280a4a4433bf4ed210c579a7.jpg) + +![](images/29ecbfaaeec0562be98a0915ebdf26bd6ed1f4249086c4d04d15220343e03eca.jpg) +Figure 5. Memory vs. Imagination. The first and second rows respectively show the captured memory and reasoned imagination. Here, we directly project the memory codebook and imagination bank into the corresponding visual regions with three different styles. We can observe that for the same input, memory and imagination separately pay attention to different fragmented content. + +![](images/3d913f5daa46a8a7f2bd447dd8a8bafa459082df9b17a14e83c4690c56a49eec.jpg) + +![](images/4c4e6355119b3e276a67ba4fe924170d467cf0b28060198c68268a7d219b2be7.jpg) + +![](images/7cd235b86860394974936228a54e67824ff760baa295725eb1ae932edc25f4e6.jpg) + +![](images/67fb31a0290ba2146bcbc9cdf5d60aeced81e4635235efbd2d1fdc7b6ce63087.jpg) + +![](images/f28a8d5250108df04dc8a8c8fa37b59826f26e33a2f7047968339d09705ba33c.jpg) + +![](images/6da988c23ba6f49c6ce8fcf255d1d31da9393ceba003f9ee115ce84767830037.jpg) + +![](images/536bdda7e54ae6801b7b06982362501cd946a4f352cdfa36068d22764da80f84.jpg) + +![](images/9226f751195059ca222a8ff48cbbbde0750b063618b0c1778e649df26626205d.jpg) + +![](images/33cbb57dc76ac45e4af8fab885015ebd12158ba8f6e2e38353b7436091f72de3.jpg) + +![](images/0ee62376f41a930d22c0443dad774b5ba4488222aae85f3c989c5be96318c36b.jpg) + +![](images/c7b84bd9511e4fe03ac570c643b76688ebe0346dbb87beab0b9a6418f948b657.jpg) + +![](images/19d3dbd4b28d1a3c1b77270d7acc3f98cb09352ed9358e089919011f8f429862.jpg) +(a) Style Image +Figure 6. Visualization of the memory-enhanced map $E$ and generated unknown map $S_{u}$ . For each feature map, the channels corresponding to the maximum value are selected for visualization. + +![](images/e12ec459e8eed967cc1a57b37f87ab1c48e87ac1ef27da8e75600495629c2582.jpg) +(b) Enhanced $E$ + +![](images/39d599cf85465a8b287e8a84bba31e5c9011b414f081e8e0d1a894c1486602ba.jpg) +(c) Unknown $S_{u}$ + +Analysis of Multi-level Perception. As shown in Fig. 1 and 2, the goal of multi-level perception is to construct a complete impression of objects. Here, we make an ablation analysis of three perception levels. We only change the perception level. Other operations and settings are kept unchanged. Table 6 shows the results. We can observe that employing different perception level affects the corresponding detection performance. Particularly, since the connective level considers the relationships between visual regions, using this level could enhance the performance effectively. When combining these three levels, the performance could be further improved. This indicates that multi-level perception is indeed conducive to building a complete observation of objects, strengthening the robustness. + +Analysis of Codeword Number $Q$ . For the fragmented memory recorder, we utilize $Q$ codewords to preserve the information of known objects. Here, we analyze the $Q$ impact. We observe that when $Q$ is set to 16 and 32, the FPR95 value increases slightly, i.e., $80.54\%$ and $80.33\%$ . The reason may be that utilizing fewer codewords could not sufficiently capture known characteristics. Utilizing more code + +
LocalGlobalConnectiveFPR95 ↓AUROC ↑
✓81.19%73.04%
✓81.62%73.18%
✓81.37%73.21%
✓✓80.85%73.96%
✓✓80.63%74.12%
✓✓80.58%74.09%
✓✓✓80.16%74.31%
+ +Table 6. Ablation analysis of local, global, and connective perception perspectives. Other operations are kept unchanged. + +words increases parameters, resulting in overfitting. + +In Fig. 5, we intuitively show the corresponding content captured by the built memory and reasoned imagination. Obviously, for the images with different styles, the memory recorder can effectively focus on fragmented object-related content, enhancing the robustness against the covariate-shift impact. Furthermore, the reasoned imagination mainly pays attention to background regions that involve plentiful visual information, which is instrumental in promoting the following generator to synthesize expected unknown features and attenuates the impact of two-shift entanglement. + +Visualization Analysis. Fig. 6 shows three examples of memory-enhanced map and virtual unknown map. We can further observe that fusing the known memory could indeed strengthen object-related information. The synthesized unknown features mainly pay attention to background regions involving plentiful visual characteristics, enlarging the gap with the known features significantly. In supplementary material, we will give more analysis about our method. + +# 5. Conclusion + +This paper explores a challenging problem, i.e., ODU-OD, which aims to detect unknown objects within unseen target domains. To weaken the entangled impact of covariate and semantic shifts, we propose a new WFS method and imitate humans to leverage multi-level perception, fragmented memory, and reasoned imagination to recurrently generate virtual unknown features. Extensive experimental results on four tasks demonstrate the superiorities of our method. + +# Acknowledgments + +This work is supported in part by the National Key R&D Program of China (No. 2023YFC3305600), Joint Fund of Ministry of Education of China (8091B022149, 8091B02072404), National Natural Science Foundation of China (62472333, 62132016, 62171343), and Natural Science Basic Research Program of Shaanxi (2020JC-23). + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ECCV, 2020. 1 +[2] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, pages 11580-11590, 2021. 7 +[3] Muhammad Sohail Danish, Muhammad Haris Khan, Muhammad Akhtar Munir, M Saqib Sarfraz, and Mohsen Ali. Improving single domain-generalized object detection: A focus on diversification and alignment. In CVPR, pages 17732-17742, 2024. 2 +[4] Xuefeng Du, Gabriel Gozum, Yifei Ming, and Yixuan Li. Siren: Shaping representations for detecting out-of-distribution objects. In NeurIPS, 2022. 1, 3, 6 +[5] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware object detection: Learning what you don't know from videos in the wild. In CVPR, pages 13678-13688, 2022. 3 +[6] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. ICLR, 2022. 1, 2, 3, 5, 6, 7 +[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2 +[8] Yanchen Guan, Haicheng Liao, Zhenning Li, Guohui Zhang, and Chengzhong Xu. World models for autonomous driving: An initial survey. arXiv preprint arXiv:2403.02622, 2024. 1 +[9] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. +[10] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. *ICLR*, 2019. 1 +[11] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In CVPR, pages 9591-9600, 2022. 2, 5, 6, 7 +[12] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 2, 3 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2, 3, 6 + +[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3, 5 +[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. *ICLR*, 2017. 6 +[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, pages 10951-10960, 2020. 6 +[17] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, pages 4874-4883, 2019. 7 +[18] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In CVPR, pages 5830-5840, 2021. 7 +[19] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2, 3 +[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. arXiv preprint arXiv:2112.00374, 2021. 2, 5, 7 +[21] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. *ICLR*, 2018. 6 +[22] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. NeurIPS, 31, 2018. 6 +[23] Deng Li, Among Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In CVPR, pages 17606-17615, 2024. 2 +[24] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. *ICLR*, 2017. 6 +[25] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In ICCV, pages 8771-8780, 2021. 1 +[26] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation models? ECCV, 2024. 1, 2, 3 +[27] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. NeurIPS, 33: 21464-21475, 2020. 5, 6 +[28] Dimity Miller, Lachlan Nicholson, Feras Dayoub, and Niko Sünderhauf. Dropout sampling for robust object detection in open-set conditions. In ICRA, pages 3243-3249, 2018. 7 +[29] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, pages 464-479, 2018. 7 +[30] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, pages 1863-1871, 2019. 7 +[31] Bimsara Pathiraja, Malitha Gunawardhana, and Muhammad Haris Khan. Multiclass confidence and localization calibration for object detection. In CVPR, pages 19734-19743, 2023. 1 + +[32] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, pages 7263-7271, 2017. +[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurlPS, pages 91-99, 2015. 1, 3, 5, 7 +[34] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with gram matrices. In ICML, pages 8491-8501. PMLR, 2020. 6 +[35] Eleanor Spens and Neil Burgess. A generative model of memory construction and consolidation. Nature Human Behaviour, pages 1-18, 2024. 1 +[36] Weinan Sun, Madhu Advani, Nelson Spruston, Andrew Saxe, and James E Fitzgerald. Organizing memories for generalization in complementary learning systems. Nature neuroscience, 26(8):1438-1448, 2023. 1 +[37] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. NeurIPS, 33:11839-11852, 2020. 6 +[38] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. *ICLR*, 2018. 2 +[39] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, pages 3219-3229, 2023. 2, 6, 7 +[40] Samuel Wilson, Tobias Fischer, Feras Dayoub, Dimity Miller, and Niko Sunderhauf. Safe: Sensitivity-aware features for out-of-distribution object detection. In ICCV, pages 23565-23576, 2023. 3, 6 +[41] Among Wu and Cheng Deng. Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation. In CVPR, pages 847-856, 2022. 1, 2, 5, 6, 7 +[42] Among Wu and Cheng Deng. Discriminating known from unknown objects via structure-enhanced recurrent variational autoencoder. In CVPR, pages 23956-23965, 2023. 1, 6 +[43] Among Wu, Yahong Han, Linchao Zhu, and Yi Yang. Instance-invariant domain adaptive object detection via progressive disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4178-4193, 2021. 1 +[44] Among Wu, Da Chen, and Cheng Deng. Deep feature deblurring diffusion for detecting out-of-distribution objects. In ICCV, pages 13381-13391, 2023. 1, 3, 5, 6, 7 +[45] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 7 +[46] Hang Zhang, Jia Xue, and Kristin Dana. Deep ten: Texture encoding network. In CVPR, pages 708-717, 2017. 4 +[47] Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Learning placeholders for open-set recognition. In CVPR, pages 4401-4410, 2021. 7 +[48] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 1 + +[49] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 3 \ No newline at end of file diff --git a/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/images.zip b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0876ea1592f8ddca7c7a54b3d57295eea30ce90a --- /dev/null +++ b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae48ae0f8a4a0cbbef3ddfa60aaa2825bedf8c1718194b3fb236de994e7543a1 +size 1085474 diff --git a/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/layout.json b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cc8909dc2f1db05b9e91e55c4497e4fa4845f7a5 --- /dev/null +++ b/2025/Percept, Memory, and Imagine_ World Feature Simulating for Open-Domain Unknown Object Detection/layout.json @@ -0,0 +1,10993 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 500, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 500, + 140 + ], + "type": "text", + "content": "Percept, Memory, and Imagine: World Feature Simulating for Open-Domain Unknown Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 240, + 162, + 374, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 162, + 374, + 175 + ], + "spans": [ + { + "bbox": [ + 240, + 162, + 374, + 175 + ], + "type": "text", + "content": "Among Wu, Cheng Deng*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 175, + 465, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 175, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 144, + 175, + 465, + 190 + ], + "type": "text", + "content": "School of Electronic Engineering, Xidian University, Xi'an, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 181, + 191, + 426, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 191, + 426, + 202 + ], + "spans": [ + { + "bbox": [ + 181, + 191, + 426, + 202 + ], + "type": "text", + "content": "amwu@xidian.edu.cn, chdeng@mail.xidian.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 256, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 296, + 627 + ], + "type": "text", + "content": "To accelerate the safe deployment of object detectors, we focus on reducing the impact of both covariate and semantic shifts. And we consider a realistic yet challenging scenario, namely Open-Domain Unknown Object Detection (ODU-OD), which aims to detect unknown objects in unseen target domains without accessing any auxiliary data. Towards ODU-OD, it is feasible to learn a robust discriminative boundary by synthesizing virtual features. Generally, perception, memory, and imagination are three essential capacities for human beings. Through multi-level perception and rich memory about known objects, the characteristics of unknown objects can be imagined sufficiently, enhancing the ability of discriminating known from unknown objects. Inspired by this idea, an approach of World Feature Simulation (WFS) is proposed, mainly consisting of a multi-level perception, memory recorder, and unknown-feature generator. Specifically, after extracting the features of the input, we separately employ a Mamba and Graph Network to obtain the global-level and connective-level representations. Next, a codebook containing multiple learnable codewords is defined to preserve fragmented memory of known objects. Meanwhile, we perform a modulated operation on the memory to form the imagination bank involving unknown characteristics. Finally, to alleviate the impact of lacking supervision data, based on the multi-level representation and imagination bank, a dedicated unknown-feature generator is designed to recurrently synthesize outlier features deviating from in-distribution (ID) objects. The significant performance gains on four different detection tasks demonstrate the superiorities of our method. The code will be released at https://github.com/AmingWu/WFS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 650, + 135, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 650, + 135, + 662 + ], + "spans": [ + { + "bbox": [ + 56, + 650, + 135, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 670, + 295, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 695 + ], + "type": "text", + "content": "With the rejuvenation of deep neural networks, object detection has achieved many advances [1, 14, 31-33] recently." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 555, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 292 + ], + "type": "text", + "content": "Though these advances have obtained impressive performance, in real-world scenarios, they inevitably encounter both covariate shifts (e.g., test-time changes in image styles) and semantic shifts (e.g., test-time unknown classes), which limits their applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 293, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 293, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 312, + 293, + 555, + 449 + ], + "type": "text", + "content": "Since there is no auxiliary data available, to alleviate the covariate-shift impact, existing researches [25, 41, 43, 48] mainly pay attention to enhancing the generalization ability by learning domain-invariant features. As for the semantic-shift challenge, most existing approaches [4, 6, 26, 42, 44] explore synthesizing diverse virtual features to strengthen the capability of discriminating unknown objects. Although these methods have been demonstrated to be effective, they are only concerned with a single-shift case and could not cope with a more difficult scenario involving both covariate and semantic shifts, e.g., for an autonomous car, the detector may encounter an unknown object in unseen adverse weather, resulting in a safety risk." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "type": "text", + "content": "To promote the safe deployment of detectors, we propose a realistic yet challenging task, namely Open-Domain Unknown Object Detection (ODU-OD). As shown in Fig. 1, given one source dataset including multiple pre-defined categories for training, the goal of ODU-OD is to accurately detect unknown out-of-distribution (OOD) objects in multiple unseen target domains with different styles. Since ODU-OD contains both covariate and semantic shifts and fails to access any auxiliary data for training, existing single-shift methods can not be directly utilized to tackle this task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "In essence, towards covariate and semantic shifts, it is critical to learn the features that lead to a robust discriminative boundary. Particularly, enhancing the robustness could attenuate the covariate-shift impact and improve the generalization ability. And strengthening the discrimination is instrumental in accurately recognizing known and unknown objects. Moreover, recent researches [8-10, 35, 36] indicate that based on the current observations and formed memory, human beings could produce rich imaginations, improving the adaptation ability for unknown situations. Therefore, we explore simulating humans to leverage plentiful perception and fragmented memory to imagine unknowns for training" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4682" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 255, + 205 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 255, + 205 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 255, + 205 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 255, + 205 + ], + "type": "image", + "image_path": "f4352d79a6183dde2dbda3c7fff3f56d4ad25c0021c7392702346c3524c001bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 555, + 276 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 555, + 276 + ], + "type": "text", + "content": "Figure 1. World Feature Simulation for Open-Domain Unknown Object Detection, which imitates humans to synthesize virtual unknown features and consists of a multi-level perception, memory recorder, and dedicated unknown-feature generator. We first calculate the multi-level representation from the local, global, and connective levels. Then, a memory codebook is constructed to preserve the local object information (as shown in pink boxes). Next, a modulated operation is performed on the fragmented memory to form the imagination bank involving unknown characteristics. Finally, we design an unknown-feature generator to recurrently synthesize virtual unknown features deviating from known objects, enhancing the ability of detecting unknown objects within unseen domains (as shown in green boxes)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 273, + 70, + 390, + 205 + ], + "blocks": [ + { + "bbox": [ + 273, + 70, + 390, + 205 + ], + "lines": [ + { + "bbox": [ + 273, + 70, + 390, + 205 + ], + "spans": [ + { + "bbox": [ + 273, + 70, + 390, + 205 + ], + "type": "image", + "image_path": "c11afc1c9c3d4294cc580a1ed381ba73e832eaddc4238b1e1cc4be4311e5f1a4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 406, + 70, + 554, + 205 + ], + "blocks": [ + { + "bbox": [ + 406, + 70, + 554, + 205 + ], + "lines": [ + { + "bbox": [ + 406, + 70, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 406, + 70, + 554, + 205 + ], + "type": "image", + "image_path": "515fd2a86094b91420fcf94427c11f7dd34407f231242e8de3f80d4eabcf53c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 282, + 295, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 295, + 306 + ], + "type": "text", + "content": "(as shown in Fig. 1), enhancing the ability of distinguishing unknown objects within unseen domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 306, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 306, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 295, + 510 + ], + "type": "text", + "content": "Specifically, as shown in Fig. 1, a method of World Feature Simulation (WFS) is proposed, mainly consisting of a multi-level perception, memory recorder, and unknown-feature generator. For a specific object, in order to deepen its perception, humans usually observe the object from multiple different perspectives, e.g., local and global levels. To this end, given an ID input, we first utilize a convolutional backbone [13] to extract its local features. Then, a Mamba [7, 12, 49] and a Graph network [19, 38] are separately performed on the local features to obtain the global and connective features. And the local, global, and connective features are concatenated to form the multi-level representation. Next, we define a codebook containing multiple learnable codewords to preserve fragmented memory of the known objects. By fusing the memory into the current multi-level representation, the robustness against covariateshift risk could be enhanced effectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 653 + ], + "type": "text", + "content": "In Fig. 1, to achieve the ability of imagining unknowns, a modulated operation is performed on the known codebook to obtain the imagination bank. And we design a contrastive constraint to promote the bank to involve plentiful unknown characteristics. Finally, a dedicated unknown-feature generator is proposed to leverage the multi-level representation and imagined unknown characteristics to recurrently synthesize virtual unknown features deviating from ID inputs, which is conducive to learning a robust discriminative boundary between known and unknown objects within unseen domains. Experimental results on four different detection tasks demonstrate the superiorities of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 654, + 253, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 654, + 253, + 664 + ], + "spans": [ + { + "bbox": [ + 67, + 654, + 253, + 664 + ], + "type": "text", + "content": "The contributions are summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": "(1) To promote the deployment of detectors, we propose a new task, namely Open-Domain Unknown Object Detection (ODU-OD), aiming to detect unknown objects in unseen target domains without accessing any auxiliary data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 282, + 556, + 475 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 282, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 354 + ], + "type": "text", + "content": "(2) To tackle ODU-OD involving both covariate and semantic shifts, we imitate humans to explore a new framework of World Feature Simulation, which aims to synthesize virtual unknown features for training and contains a multi-level perception, fragmented memory recorder, and unknown-feature generator." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 355, + 556, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 556, + 475 + ], + "type": "text", + "content": "(3) We build an Open-Domain Dataset for unknown object detection to verify our method WFS, consisting of three different styles [20], i.e., 'Acrylic', 'Purple-Brush', and 'Sketch'. Meanwhile, our method is further evaluated on three single-shift tasks, i.e., OOD-OD [6], open-set object detection [11], and Single-DGOD [41]. The significant performance gains over multiple baselines demonstrate the effectiveness of our method. Meanwhile, visualization analyses also show that the built memory and reasoned imagination could indeed capture fragmented visual content." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 488, + 400, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 488, + 400, + 500 + ], + "spans": [ + { + "bbox": [ + 314, + 488, + 400, + 500 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 509, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 676 + ], + "type": "text", + "content": "Covariate-shift Detection. To improve the generalization of detectors, a task of Single-DGOD [41] is proposed, aiming to learn an object detector that performs well on multiple unseen target domains with only one source domain for training. Some works [23, 39, 41] utilize feature disentanglement and large-scale models to extract domain-invariant representations. Moreover, Liu et al. [26] further consider the biased problem and propose a causal-based method to learn unbiased generalized representations. And Danish et al. [3] explore diversifying the source domain and aligning detections based on prediction confidence. These methods have been verified on multiple benchmarks. However, they only consider the covariate-shift scenario and could not detect unknown objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Semantic-shift Detection. Recently, a task of unsupervised OOD-OD [6] is proposed, whose goal is to detect unknown objects without relying on any OOD data. To alle" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4683" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 555, + 263 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 555, + 263 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 555, + 263 + ], + "type": "image", + "image_path": "fcd8e9bf4f90cde9792c69866df3a77d2619c4df5ccb5257398cac02da8df11c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 333 + ], + "type": "text", + "content": "Figure 2. The details of World Feature Simulation for ODU-OD. To enhance the robustness against covariate shift, a multi-level perception is used to extract the representation " + }, + { + "bbox": [ + 54, + 266, + 555, + 333 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 54, + 266, + 555, + 333 + ], + "type": "text", + "content": " from the local, global, and connective levels. Meanwhile, we design a memory recorder to preserve rich content about known objects. By calculating a memory-enhanced map, the robustness could be further strengthened. Next, through a modulated operation, we can reason the imagination bank involving plentiful unknown characteristics. Finally, a dedicated unknown-feature generator is proposed to utilize the built memory and imagination characteristics to recurrently synthesize virtual unknown features, improving the ability of detecting unknown objects within unseen domains." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 338, + 294, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 338, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 294, + 529 + ], + "type": "text", + "content": "viate the impact of lacking OOD data for supervision, VOS [6] and DFDD [44] explore synthesizing virtual OOD features that slightly deviate from the distribution of ID data. Besides, the work [5] further learns unknown-aware knowledge from external videos, which does not match the setting of unsupervised OOD-OD. Finally, Du et al. [4] present to use a distance-based mechanism to shape the learned representations. Meanwhile, Wilson et al. [40] attempt to enhance the sensitivity of the learned features, which is beneficial for improving the performance of detecting OOD objects. The work [26] struggles to leverage generative models trained on large-scale open-set data to synthesize OOD samples. Although these methods have been demonstrated to be effective, they only consider the semantic-shift problem and ignore the covariate-shift impact, which limits their applications to unknown objects within unseen domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 531, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 531, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 54, + 531, + 295, + 627 + ], + "type": "text", + "content": "Different from the above methods, in this paper, we explore a more difficult problem involving both covariate and semantic shifts, i.e., Open-Domain Unknown Object Detection. And a new approach of World Feature Simulation is proposed for synthesizing unknown features for training, consisting of a multi-level perception, fragmented memory recorder, and unknown-feature generator. Experimental results on multiple datasets indicate its superiorities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 643, + 205, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 205, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 205, + 656 + ], + "type": "text", + "content": "3. World Feature Simulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "For ODU-OD, during training, we can only access single source-domain data with some pre-defined categories. During test, the detector is required to accurately localize and recognize unknown objects within multiple unseen do" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 338, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 374 + ], + "type": "text", + "content": "mains. To reduce the entangled impact of covariate and semantic shifts, as shown in Fig. 2, we design a World Feature Simulation for synthesizing unknown features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 380, + 526, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 526, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 526, + 392 + ], + "type": "text", + "content": "3.1. Observing Objects from Multiple Levels" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 397, + 554, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 554, + 480 + ], + "type": "text", + "content": "Generally, in order to deepen the perception of a specific object, human beings usually observe the object from multiple different perspectives, which is beneficial for constructing a complete impression. Therefore, in Fig. 2, we explore imitating this capability based on local, global, and connective levels. Specifically, we exploit the widely used object detector, i.e., Faster R-CNN [14, 33], as the basic model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": "Given an ID input image, a convolutional network, e.g., ResNet [13], is taken as the backbone to extract the local feature map " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "F_{l} \\in \\mathbb{R}^{w \\times h \\times c}" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": " separately represent width, height, and the number of channels. Each element in " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": " describes a local region of the input. Besides the local level, human beings usually intensify their impression by global observation and discovering the relations between local regions. Hence, we first unfold " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": " to a sequence of visual elements, i.e., " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "D = \\{d_{i} \\in \\mathbb{R}^{c}, i = 1, \\dots, wh\\}" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": ". Then, one-layer vision mamba [49] is used to process " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": ", which outputs " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "H = \\{h_{i} \\in \\mathbb{R}^{c}, i = 1, \\dots, wh\\}" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": " is reshaped to the original size to obtain the global representation " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "F_{g} \\in \\mathbb{R}^{w \\times h \\times c}" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": " [12]. Next, we use one-layer graph convolution [19] to further capture the semantic connections between local regions. Concretely, we first reshape " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\tilde{F} \\in \\mathbb{R}^{wh \\times c}" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": ". Then, an adjacency matrix for an undirected graph is defined as " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = \\text{softmax}(\\tilde{F}\\tilde{F}^{\\top}) + I_{wh}" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "inline_equation", + "content": "I_{wh}" + }, + { + "bbox": [ + 313, + 481, + 555, + 696 + ], + "type": "text", + "content": " indicates the identity matrix:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 353, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 353, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {R} = \\mathbf {A} \\tilde {F}, \\quad F _ {c} = r e l u (W _ {c} * \\tilde {\\mathcal {R}} + b _ {c}), \\qquad (1)", + "image_path": "469e60b9bb47dea8a239561a5140627608ce87e53c088baeab7a9c83f957b93f.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4684" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{R}}\\in \\mathbb{R}^{w\\times h\\times c}" + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": " denotes the reshaped result of " + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "inline_equation", + "content": "W_{c}\\in \\mathbb{R}^{3\\times 3\\times c\\times c}" + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "inline_equation", + "content": "b_{c}\\in \\mathbb{R}^{c}" + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": " indicate the trainable parameters. " + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "inline_equation", + "content": "F_{c}\\in \\mathbb{R}^{w\\times h\\times c}" + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": " represents the connective feature. Finally, " + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "inline_equation", + "content": "F_{l},F_{g}" + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": " are concatenated as the multi-level representation " + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "inline_equation", + "content": "F_{v}\\in \\mathbb{R}^{w\\times h\\times c}" + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "inline_equation", + "content": "F_{v} = \\phi ([F_{l},F_{g},F_{c}])" + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)\\in \\mathbb{R}^{1\\times 1\\times 3c\\times c}" + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": " is utilized to transform the number of channels. Compared with " + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "inline_equation", + "content": "F_{l},F_{v}" + }, + { + "bbox": [ + 55, + 71, + 296, + 180 + ], + "type": "text", + "content": " involves sufficient object-related information from three levels, which strengthens the robustness against the covariate shift." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 295, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 205 + ], + "type": "text", + "content": "3.2. Mining Fragmented Memory and Imagination" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "text", + "content": "In general, human beings could exploit familiar fragmented memory to improve the robustness of recognizing known objects. Then, based on the memory, human beings could produce rich imagination, enhancing the ability of discriminating unknowns. To simulate this capability, we present a memory recorder [46] to preserve the information of known objects. Particularly, a learnable memory codebook is denoted as " + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "inline_equation", + "content": "M = \\{m_j \\in \\mathbb{R}^c, j = 1, \\dots, Q\\}" + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "text", + "content": " that contains " + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "text", + "content": " codewords. Meanwhile, by means of a weight " + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "inline_equation", + "content": "a_{ij}" + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "text", + "content": ", each element " + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "inline_equation", + "content": "f_i \\in \\mathbb{R}^c" + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "text", + "content": " of the multi-level representation " + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "inline_equation", + "content": "F_v" + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "text", + "content": " can be assigned to each codeword " + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "text", + "content": ". And the residual vector is represented as " + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "inline_equation", + "content": "r_{ij} = f_i - m_j" + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, wh" + }, + { + "bbox": [ + 55, + 212, + 296, + 403 + ], + "type": "text", + "content": ". Here, the residual operation is conducive to promoting the memory codewords to effectively capture the corresponding content of local visual regions. The processes of capturing memory are shown as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 416, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 416, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 67, + 416, + 295, + 449 + ], + "type": "interline_equation", + "content": "v _ {j} = \\sum_ {i = 1} ^ {w h} a _ {i j} r _ {i j}, \\quad a _ {i j} = \\frac {\\exp \\left(- k _ {j} \\| r _ {i j} \\| ^ {2}\\right)}{\\sum_ {j = 1} ^ {Q} \\exp \\left(- k _ {j} \\| r _ {i j} \\| ^ {2}\\right)}, \\tag {2}", + "image_path": "ebed914eb9261f2bfbb6c9112fdf418abb1b96ac182a827e6cb97a140113eac0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "k_{j}" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " represents the learnable smoothing factor for the codeword " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathbb{R}^{Q \\times c}" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " is the output result. The concatenation of " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "F_v" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " is used to calculate the memory-enhanced map " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "E \\in \\mathbb{R}^{w \\times h \\times c}" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "E = \\psi([F_v, \\tilde{\\mathcal{V}}W_0 + b_0])" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{V}} \\in \\mathbb{R}^{1 \\times Qc}" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " is the reshaped result of " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": ". Meanwhile, " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "W_0 \\in \\mathbb{R}^{Qc \\times c}" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "b_0 \\in \\mathbb{R}^c" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " are parameters of a fully-connected layer. " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\psi(\\cdot)" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " includes two convolutional layers to transform the concatenated result. Through fusing memory, the information capacity of " + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 460, + 296, + 581 + ], + "type": "text", + "content": " could be increased, improving the robustness of object localization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Next, we further explore how to leverage the memory to reason the imagination that guides the synthesis of virtual unknown features. Particularly, we assume that the imagination should satisfy two conditions: (1) Imagination and memory should belong to different spaces. (2) Imagination mainly describes unknowns. It could not well represent known visual content. To this end, based on the memory " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{Q \\times c}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", a modulated operation is performed on " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " to obtain the imagination bank " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\in \\mathbb{R}^{Q \\times c}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\gamma \\odot M + \\beta" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " are learnable parameters for channel-wisely scaling and shifting " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". Next, we separately utilize " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 321, + 70, + 547, + 186 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 547, + 186 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 547, + 186 + ], + "type": "image", + "image_path": "19462839cd050a9b5595f959dfc073517ed97763d6518ff4176c912636cbd353.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "text", + "content": "Figure 3. Unknown-feature generator. Based on " + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "inline_equation", + "content": "F_{m} \\in \\mathbb{R}^{w \\times h \\times c}" + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "text", + "content": ", a variational operation is performed on " + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "inline_equation", + "content": "F_{m}" + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{w \\times h \\times c}" + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "text", + "content": ". Then, we exploit the reasoned imagination bank to re-represent " + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "text", + "content": ", which outputs " + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{w \\times h \\times c}" + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "text", + "content": ". Next, " + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "text", + "content": " are unfolded into a sequence with " + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "inline_equation", + "content": "wh" + }, + { + "bbox": [ + 313, + 189, + 555, + 266 + ], + "type": "text", + "content": " elements. Finally, a dedicated recurrent generator including three inputs and one output is presented to synthesize virtual unknown features, improving the discrimination." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 270, + 400, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 270, + 400, + 281 + ], + "spans": [ + { + "bbox": [ + 314, + 270, + 400, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 314, + 270, + 400, + 281 + ], + "type": "text", + "content": " to re-represent " + }, + { + "bbox": [ + 314, + 270, + 400, + 281 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 314, + 270, + 400, + 281 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 286, + 555, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 286, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 328, + 286, + 555, + 301 + ], + "type": "interline_equation", + "content": "F _ {m} = \\operatorname {s o f t} x \\left(\\tilde {F} _ {v} M ^ {\\top}\\right) M, F _ {i} = \\operatorname {s o f t} x \\left(\\tilde {F} _ {v} \\mathcal {I} ^ {\\top}\\right) \\mathcal {I}, \\tag {3}", + "image_path": "2e17f35ddca75add722302dc9e44417961f1bf771750c1cf6fa6845ccd7f01f4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 306, + 554, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 378 + ], + "type": "text", + "content": "where 'softx' denotes the softmax operation. " + }, + { + "bbox": [ + 313, + 306, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\tilde{F}_v\\in \\mathbb{R}^{wh\\times c}" + }, + { + "bbox": [ + 313, + 306, + 554, + 378 + ], + "type": "text", + "content": " is the reshaped result of " + }, + { + "bbox": [ + 313, + 306, + 554, + 378 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 313, + 306, + 554, + 378 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 306, + 554, + 378 + ], + "type": "inline_equation", + "content": "F_{m}\\in \\mathbb{R}^{w\\times h\\times c}" + }, + { + "bbox": [ + 313, + 306, + 554, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 306, + 554, + 378 + ], + "type": "inline_equation", + "content": "F_{i}\\in \\mathbb{R}^{w\\times h\\times c}" + }, + { + "bbox": [ + 313, + 306, + 554, + 378 + ], + "type": "text", + "content": " indicate the re-represented results based on the fragmented memory and imagination. Finally, a contrastive loss is exploited to enlarge the gap between the memory recorder and imagination bank:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 345, + 383, + 555, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 383, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 345, + 383, + 555, + 412 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i m} = - \\log \\frac {\\exp \\left(\\sin \\left(F _ {m} , F _ {v}\\right) / \\tau\\right)}{\\sum_ {j = 0} ^ {1} \\exp \\left(\\sin \\left(B [ j ] , F _ {v}\\right) / \\tau\\right)}, \\tag {4}", + "image_path": "347708cf387780bf41d5f650f04721d8e04b0009940306128f2b48f1ce0d4013.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "inline_equation", + "content": "B = [F_m, F_i]" + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(F_m, F_v)" + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "text", + "content": " denotes the average of the cosine similarity between all corresponding elements of feature map " + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "inline_equation", + "content": "F_m" + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "inline_equation", + "content": "F_v" + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "text", + "content": " is a hyper-parameter, which is set to 1.0 in the experiments. Obviously, by minimizing the contrastive loss, the distance between " + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "inline_equation", + "content": "F_m" + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "inline_equation", + "content": "F_v" + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "text", + "content": " could be narrowed, promoting the memory recorder to capture sufficient content about known objects. Instead, the distance between " + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "inline_equation", + "content": "F_i" + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "inline_equation", + "content": "F_v" + }, + { + "bbox": [ + 313, + 417, + 555, + 549 + ], + "type": "text", + "content": " is enlarged, forcing the reasoned imagination bank to contain plentiful information deviating from knowns. Next, the imagination bank is utilized to generate expected unknown features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 555, + 541, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 541, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 541, + 568 + ], + "type": "text", + "content": "3.3. Generating Unknown Features for Training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "text", + "content": "For ODU-OD, to attenuate the impact of lacking unknown data for supervision, we design an unknown-feature generator to synthesize outlier features involving unknown characteristics. Specifically, two convolutional networks " + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "inline_equation", + "content": "W_{\\mu}" + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "inline_equation", + "content": "W_{V}" + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "text", + "content": " are performed on " + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "inline_equation", + "content": "F_{m}" + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "text", + "content": " to estimate the corresponding means " + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^{w\\times h\\times c}" + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "text", + "content": " and variances " + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{w\\times h\\times c}" + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "text", + "content": ". Then, an encoding operation is performed to obtain " + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{w\\times h\\times c}" + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "text", + "content": ". Meanwhile, we exploit the reasoned imagination bank " + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "text", + "content": " to re-represent " + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 572, + 554, + 681 + ], + "type": "text", + "content": ". The processes are shown as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 345, + 685, + 553, + 705 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 685, + 553, + 705 + ], + "spans": [ + { + "bbox": [ + 345, + 685, + 553, + 705 + ], + "type": "interline_equation", + "content": "\\mu = W _ {\\mu} * F _ {m}, \\quad V = W _ {V} * F _ {m}, \\tag {5}", + "image_path": "99b17f503b691554b47d4c4acdb775b9525346b497bc69fe73a5fb5c71c826e2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 701, + 521, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 701, + 521, + 715 + ], + "spans": [ + { + "bbox": [ + 347, + 701, + 521, + 715 + ], + "type": "interline_equation", + "content": "Z = \\mu + \\epsilon \\odot V, \\quad G = \\text {s o f t m a x} (\\tilde {Z} \\mathcal {I} ^ {\\top}) \\mathcal {I},", + "image_path": "fbe48022b0b6e00506bc8a17a7200790ad4b25a4a4df51e9a3782f74c4d63aa9.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4685" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "inline_equation", + "content": "\\tilde{Z} \\in \\mathbb{R}^{wh \\times c}" + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "text", + "content": " is the reshaped result of " + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "text", + "content": " represents Gaussian noise sampled from " + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, I)" + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{w \\times h \\times c}" + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "text", + "content": " is the re-represented result based on the imagination bank " + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "text", + "content": ". Since the imagination could not well express the knowns, " + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 71, + 294, + 155 + ], + "type": "text", + "content": " contains plentiful information deviating from known visual content, which is instrumental in promoting the generation of virtual unknown features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "text", + "content": "Finally, we present a dedicated generator to recurrently synthesize expected unknown features. As shown in Fig. 3, the generator contains three inputs and one output. Particularly, at each time, the generator takes " + }, + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "inline_equation", + "content": "G_{t}" + }, + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "text", + "content": ", and the previous output " + }, + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "inline_equation", + "content": "S_{t-1}" + }, + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "text", + "content": " as the input, and synthesizes the current output " + }, + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 55, + 156, + 294, + 228 + ], + "type": "text", + "content": ". The generating processes are shown as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 237, + 300, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 237, + 300, + 311 + ], + "spans": [ + { + "bbox": [ + 58, + 237, + 300, + 311 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {t} = \\sigma \\left(W _ {z z} * Z _ {t} + W _ {o z} * G _ {t} + W _ {s z} * S _ {t - 1} + b _ {z}\\right), \\\\ g _ {t} = \\sigma \\left(W _ {z g} * Z _ {t} + W _ {o g} * G _ {t} + W _ {s g} * S _ {t - 1} + b _ {g}\\right), \\\\ \\tilde {S} _ {t} = \\operatorname {T a n h} \\left(W _ {s} * Z _ {t} + W _ {h} * \\left(g _ {t} \\odot S _ {t - 1} + (1 - g _ {t}) \\odot G _ {t}\\right)\\right), \\\\ S _ {t} = \\left(1 - z _ {t}\\right) \\odot S _ {t - 1} + z _ {t} \\odot \\hat {S} _ {t}, \\tag {6} \\\\ \\end{array}", + "image_path": "65a2a82a1968b57f6e2ee48a6b3c4c45090ac011589515d7f2935a675e69f08c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "Z_{t}\\in \\mathbb{R}^{c}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "G_{t}\\in \\mathbb{R}^{c}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": " are the current element of " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "t = 1,\\dots ,wh" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ". And " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "W_{zz}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "W_{oz}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "W_{sz}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "b_{z}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "W_{zg}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "W_{og}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "W_{sg}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "b_{g}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "W_{s}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "W_{h}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": " represent the trainable parameters. Through this generating mechanism, the synthesized features could sufficiently retain the information in previous steps, improving the content continuity of the generated features. Finally, the generated sequence " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\{S_1,S_2,\\dots S_{wh}\\}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": " are concatenated to obtain unknown feature " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "S_{u}\\in \\mathbb{R}^{w\\times h\\times c}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ". We still utilize a contrastive loss to expand the difference between known and the synthesized unknown features:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 448, + 294, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 448, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 87, + 448, + 294, + 477 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u n k} = - \\log \\frac {\\exp \\left(\\sin \\left(F _ {m} , F _ {v}\\right) / \\tau\\right)}{\\sum_ {j = 0} ^ {1} \\exp \\left(\\sin \\left(A [ j ] , F _ {v}\\right) / \\tau\\right)}, \\tag {7}", + "image_path": "dda5c713cafc63fa0aff2df46f4e46b3401d1264d3b91938d878567e77b09087.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "inline_equation", + "content": "A = [F_m, S_u]" + }, + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "text", + "content": " is set to 1.0 in the experiments. Since the generation process involves perception, memory, and imagination, using the contrastive loss could not only promote " + }, + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "inline_equation", + "content": "S_u" + }, + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "text", + "content": " to contain plentiful unknown information but also facilitate " + }, + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 485, + 294, + 556 + ], + "type": "text", + "content": " to separately capture rich known and unknown characteristics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 564, + 268, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 268, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 268, + 577 + ], + "type": "text", + "content": "3.4. WFS-Driven Unknown-Object Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Lastly, the memory-enhanced map " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " is taken as the input of Region-Proposal Network (RPN) [14, 33] to obtain a series of object proposals " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ". Meanwhile, based on " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ", RoI-Alignment followed by RoI-Feature extraction [14] is performed on " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{in}} \\in \\mathbb{R}^{p \\times n}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " denote the number of proposals and channels, respectively. Then, " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{in}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " is input to the object classifier and regressor to calculate the classification loss " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " and localization loss " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{loc}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{det}} = \\mathcal{L}_{\\mathrm{cls}} + \\mathcal{L}_{\\mathrm{loc}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ". By optimizing " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{det}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ", the memory recorder could be promoted to capture plentiful information about known objects, enhancing the robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "Next, we further enlarge the gap between the known and synthesized unknown features from the instance perspective. Concretely, based on " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ", RoI-Alignment followed by RoI-Feature extraction is performed on " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "S_{u}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{unk}} \\in \\mathbb{R}^{p \\times n}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ". Meanwhile, we define a loss function " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dis}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " to maximize the " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "KL" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "-divergence between " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{unk}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{in}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dis}} = \\mathrm{KL}[q(O_{\\mathrm{unk}}), q(O_{\\mathrm{in}})]" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "q(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " represents the probability distribution. Besides, to attain the goal of distinguishing unknown objects from known ID objects, " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{unk}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{in}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " are used to compute an uncertainty loss [6], regularizing the detector to produce a low score for the ID object features, and a high score for the virtual unknown features:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 355, + 224, + 553, + 276 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 224, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 355, + 224, + 553, + 276 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {u n t y}} = \\mathbb {E} _ {x \\sim O _ {\\text {i n}}} [ - \\log \\frac {\\exp^ {- \\mathbf {E} (x)}}{1 + \\exp^ {- \\mathbf {E} (x)}} ] + \\tag {8} \\\\ \\mathbb {E} _ {v \\sim O _ {\\mathrm {u n k}}} [ - \\log \\frac {1}{1 + \\exp^ {- \\mathbf {E} (v)} ], \\\\ \\end{array}", + "image_path": "510440d9face23a898e6cd44209e47490071e8223b7edea0d33b9d49d2a8019c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 284, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 553, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 284, + 553, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(\\cdot)" + }, + { + "bbox": [ + 313, + 284, + 553, + 308 + ], + "type": "text", + "content": " is the object-level energy score [6, 27, 44]. The overall objective is shown as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 318, + 552, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 318, + 552, + 336 + ], + "spans": [ + { + "bbox": [ + 358, + 318, + 552, + 336 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {c o n t r a s t}} = \\mathcal {L} _ {i m} + \\mathcal {L} _ {u n k} - \\mathcal {L} _ {\\text {d i s}}, \\\\ \\mathcal {L} _ {i m} = \\mathcal {L} _ {u n k} + \\mathcal {L} _ {u n k} + \\lambda_ {i} \\mathcal {L} _ {i m}. \\end{array} \\tag {9}", + "image_path": "5f8014ab52cb6cc6b64e4afcba5ef1dddd913e9957521bcc15490171fad96989.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 333, + 511, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 333, + 511, + 344 + ], + "spans": [ + { + "bbox": [ + 358, + 333, + 511, + 344 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {d e t}} + \\alpha \\cdot \\mathcal {L} _ {\\text {c o n t r a s t}} + \\lambda \\cdot \\mathcal {L} _ {\\text {u n t y}},", + "image_path": "63a82ddcefad4da306c325f8cefc5531d38a64dd49af9f910aab7fa7d8375d55.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 352, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 553, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 352, + 553, + 376 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 352, + 553, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 352, + 553, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 352, + 553, + 376 + ], + "type": "text", + "content": " are two hyper-parameters, which are set to 0.001 and 0.1 in the experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 383, + 527, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 527, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 527, + 396 + ], + "type": "text", + "content": "3.5. Inference for Unknown Object Detection" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 401, + 553, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 449 + ], + "type": "text", + "content": "During test, we only calculate the uncertainty score for unknown object detection [6, 44]. Specifically, for a predicted bounding box " + }, + { + "bbox": [ + 313, + 401, + 553, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 313, + 401, + 553, + 449 + ], + "type": "text", + "content": ", the processes of discriminating unknown objects are shown as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 322, + 457, + 553, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 457, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 322, + 457, + 553, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {E} = \\frac {\\exp^ {- \\mathbf {E} (\\mathbf {b})}}{1 + \\exp^ {- \\mathbf {E} (\\mathbf {b})}}, \\quad \\mathcal {C} (\\mathbf {b}) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} \\mathcal {E} < \\delta , \\\\ 1 & \\text {i f} \\mathcal {E} \\geq \\delta . \\end{array} \\right. \\tag {10}", + "image_path": "2841e0a93c577de33be21aa60eb4eaffd9373a7144290d53730aa5049b75d3a0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "text", + "content": "For the output of the classifier " + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\cdot)" + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "text", + "content": ", we utilize the threshold mechanism [6] to distinguish known objects (the result is 1) from unknown objects (the result is 0). The threshold " + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "text", + "content": " is commonly set to 0.95 so that a high fraction of known ID objects are correctly classified. In supplementary material, we will show more details of the proposed WFS algorithm and give more discussion." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 586, + 394, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 394, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 394, + 599 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "content": "In the experiments, our method is evaluated on four different detection tasks, i.e., ODU-OD, OOD-OD [6], open-set object detection (OSOD) [11], and Single-DGOD [41]. Particularly, we construct an Open-Domain dataset for ODU-OD, which utilizes Clipstyler [20] to transform testing data of OOD-OD [6] to three different styles, i.e., 'Acrylic', 'Purple-Brush', and 'Sketch'." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Implementation Details. We utilize Faster R-CNN [33] with RoI-Alignment layer [14] as the basic detection model." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4686" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 70, + 515, + 266 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 515, + 266 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 515, + 266 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 515, + 266 + ], + "type": "table", + "html": "
Source DataMethodStyle: Acrylic\nMS-COCO/OpenImagesStyle: Purple-Brush\nMS-COCO/OpenImagesStyle: Sketch\nMS-COCO/OpenImages
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
VOCSR-VAE [42]88.25/84.6371.02/69.6487.76/69.9367.42/79.1889.67/74.8266.47/77.04
DFDD [44]87.68/85.1270.85/69.2187.24/70.0867.23/79.4689.83/74.9266.13/76.82
VOS [6]88.73/84.7171.13/69.5688.66/70.1767.06/79.3389.76/74.6166.50/77.63
VOS + WFS80.16/77.5274.31/73.8580.07/60.6870.86/83.8782.16/67.2369.87/81.28
SAFE [40]86.62/47.8363.47/78.5186.31/39.6261.28/86.8287.56/55.2860.71/83.48
SAFE + WFS79.38/40.1667.28/81.6979.23/33.8566.47/88.7380.49/48.6165.82/84.17
BDDSR-VAE [42]64.72/53.9872.34/78.1166.43/57.2870.69/76.2469.36/59.2769.17/75.25
DFDD [44]64.08/53.2672.48/78.3665.91/56.7771.03/76.4868.74/58.8969.22/75.44
VOS [6]65.71/54.7371.96/77.8068.50/58.7970.18/75.8370.80/60.7169.05/74.97
VOS + WFS56.48/45.4275.87/82.1859.64/47.8874.34/79.8761.01/51.9473.78/79.36
SAFE [40]54.82/40.1873.71/82.6757.63/43.9772.34/80.8359.44/46.2871.27/79.53
SAFE + WFS45.91/32.0378.35/85.1447.78/34.4176.93/84.2650.26/37.1875.54/82.75
", + "image_path": "74050db151c9a4a4dfb833105c194156f09060df50c8ff369162b728ab436b6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 308, + 296, + 616 + ], + "blocks": [ + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "type": "text", + "content": "Table 1. Unknown-object detection results under unseen open domains. " + }, + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "type": "text", + "content": " denotes larger values are better and " + }, + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 268, + 555, + 301 + ], + "type": "text", + "content": " represents smaller values are better. 'VOS+WFS' and 'SAFE+WFS' indicate our method is directly plugged into VOS [6] and SAFE [40]. We can see that our method could reduce the impact of both covariate and semantic shifts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 308, + 296, + 616 + ], + "lines": [ + { + "bbox": [ + 58, + 308, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 58, + 308, + 296, + 616 + ], + "type": "table", + "html": "
SourceMethodFPR95 ↓AUROC ↑
OOD: MS-COCO / OpenImages
VOCMSP [15]70.99 / 73.1383.45 / 81.91
ODIN [24]59.82 / 63.1482.20 / 82.59
Mahalanobis [22]67.73 / 65.4181.45 / 81.48
Gram matrices [34]62.75 / 67.4279.88 / 77.62
Energy score [27]56.89 / 58.6983.69 / 82.98
Generalized ODIN [16]59.57 / 70.2883.12 / 79.23
CSI [37]59.91 / 57.4181.83 / 82.95
GAN-synthesis [21]60.93 / 59.9783.67 / 82.67
SIREN-vMF [4]64.68 / 68.5385.36 / 82.78
SIREN-KNN [4]47.45 / 50.3889.67 / 88.80
SR-VAE [42]42.17 / 46.2690.28 / 87.89
DFDD [44]41.34 / 44.5290.79 / 88.65
VOS (Baseline) [6]47.53 / 51.3388.70 / 85.23
VOS + WFS37.91 / 41.3691.47 / 89.78
SAFE (Baseline) [40]47.40 / 20.0680.30 / 92.28
SAFE + WFS37.48 / 11.2785.62 / 94.91
BDDMSP [15]80.94 / 79.0475.87 / 77.38
ODIN [24]62.85 / 58.9274.44 / 76.61
Mahalanobis [22]55.74 / 47.6985.71 / 88.05
Gram matrices [34]60.93 / 77.5574.93 / 59.38
Energy score [27]60.06 / 54.9777.48 / 79.60
Generalized ODIN [16]57.27 / 50.1785.22 / 87.18
CSI [37]47.10 / 37.0684.09 / 87.99
GAN-synthesis [21]57.03 / 50.6178.82 / 81.25
SR-VAE [42]32.23 / 21.8190.69 / 93.55
DFDD [44]30.71 / 22.6790.74 / 92.48
VOS (Baseline) [6]44.27 / 35.5486.87 / 88.52
VOS + WFS28.66 / 20.3992.37 / 94.13
SAFE (Baseline) [40]32.56 / 16.0488.96 / 94.64
SAFE + WFS21.84 / 7.8393.41 / 96.85
", + "image_path": "2104bfb640219412c1187adf19fa6ca44e7061d6ed9c49b0e70e2bd1d56118db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "type": "text", + "content": "Table 2. The performance of unsupervised OOD-OD [6] that assumes the training and test data are from the same domain. We can see that our method can reduce the impact of semantic shift." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "ResNet-50 [13] is taken as the backbone. For the memory recorder, the number " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " of the codewords is set to 24. All the experiments are trained using the standard SGD optimizer with a learning rate of 0.02. In supplementary material, we" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 312, + 527, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 527, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 527, + 324 + ], + "type": "text", + "content": "will give the details of the used datasets and metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 332, + 504, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 504, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 504, + 345 + ], + "type": "text", + "content": "4.1. Analysis of ODU-OD and OOD-OD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 506 + ], + "type": "text", + "content": "Table 1 and 2 separately illustrate the ODU-OD and OOD-OD performance. Here, we only compare with state-of-the-art methods. We can observe that though existing methods [6, 40, 42, 44] could effectively attenuate the semantic-shift impact, their performance drops severely when encountering the covariate shift (e.g., changes in image styles). Plugging our method into existing methods [6, 40] significantly improves their performance whether objects are within open or closed domains. This demonstrates that for the entangled case of covariate and semantic shifts, imitating humans to construct perception, memory, and imagination and further generating unknown features for training could simultaneously strengthen the robustness and discrimination." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 506, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 555 + ], + "type": "text", + "content": "In Fig. 4, we show some detection examples with different styles. We can observe that plugging our method could indeed attenuate both covariate and semantic shifts, improving the detection accuracy under open domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 563, + 511, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 511, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 511, + 577 + ], + "type": "text", + "content": "4.2. Analysis of OSOD and Single-DGOD" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "To further demonstrate that our method could attenuate the impact of two-shift entanglement, we verify our method on two dedicated tasks, i.e., OSOD [11] and Single-DGOD [41]. We directly plug our method into three existing methods [11, 39, 41] and do not change the training processes. Table 3 and 4 show the corresponding results. We can observe that plugging our method could improve the detection performance effectively. This further demonstrates that our method WFS could synthesize expected unknown features that are beneficial for alleviating the impact of covariate and semantic shifts, which promotes the detector to learn a ro" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4687" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 182, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 182, + 148 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 182, + 148 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 182, + 148 + ], + "type": "image", + "image_path": "7b2ff19fd9c9c63f255a5ed7337cb2a7593e25ac2bddc451c4cee1245b322c1b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 72, + 294, + 147 + ], + "blocks": [ + { + "bbox": [ + 186, + 72, + 294, + 147 + ], + "lines": [ + { + "bbox": [ + 186, + 72, + 294, + 147 + ], + "spans": [ + { + "bbox": [ + 186, + 72, + 294, + 147 + ], + "type": "image", + "image_path": "03ce489de9365809201dee26c311131efe519a14d8bad60b6af0b47afa397bd9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 296, + 72, + 400, + 147 + ], + "blocks": [ + { + "bbox": [ + 296, + 72, + 400, + 147 + ], + "lines": [ + { + "bbox": [ + 296, + 72, + 400, + 147 + ], + "spans": [ + { + "bbox": [ + 296, + 72, + 400, + 147 + ], + "type": "image", + "image_path": "0f571e54a2fe6baebe687bc822f36b02e781405d37165c19195ca79a7d18a65d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 403, + 72, + 475, + 147 + ], + "blocks": [ + { + "bbox": [ + 403, + 72, + 475, + 147 + ], + "lines": [ + { + "bbox": [ + 403, + 72, + 475, + 147 + ], + "spans": [ + { + "bbox": [ + 403, + 72, + 475, + 147 + ], + "type": "image", + "image_path": "3d632e65771615928e9fdf297c954f38cb86f536d81123137e4962b804e91bba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 478, + 72, + 550, + 147 + ], + "blocks": [ + { + "bbox": [ + 478, + 72, + 550, + 147 + ], + "lines": [ + { + "bbox": [ + 478, + 72, + 550, + 147 + ], + "spans": [ + { + "bbox": [ + 478, + 72, + 550, + 147 + ], + "type": "image", + "image_path": "14095d82bd384930a08417bccb037f9a52c615af9c56c5581a1d8dbe72e0db64.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 150, + 182, + 223 + ], + "blocks": [ + { + "bbox": [ + 59, + 150, + 182, + 223 + ], + "lines": [ + { + "bbox": [ + 59, + 150, + 182, + 223 + ], + "spans": [ + { + "bbox": [ + 59, + 150, + 182, + 223 + ], + "type": "image", + "image_path": "6df42ef3a9e1e6682df6dad44219fb311a9c850fe290af09d2e5a3effdbe5399.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 224, + 143, + 235 + ], + "lines": [ + { + "bbox": [ + 91, + 224, + 143, + 235 + ], + "spans": [ + { + "bbox": [ + 91, + 224, + 143, + 235 + ], + "type": "text", + "content": "(a) Acrylic" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 186, + 150, + 294, + 223 + ], + "blocks": [ + { + "bbox": [ + 186, + 150, + 294, + 223 + ], + "lines": [ + { + "bbox": [ + 186, + 150, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 186, + 150, + 294, + 223 + ], + "type": "image", + "image_path": "4e0c16a0db04979a2994d5fa137b4eb25c74b4298a570a3c1c2461fd60c81be5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 224, + 274, + 235 + ], + "lines": [ + { + "bbox": [ + 197, + 224, + 274, + 235 + ], + "spans": [ + { + "bbox": [ + 197, + 224, + 274, + 235 + ], + "type": "text", + "content": "(b) Purple-Brush" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 296, + 150, + 400, + 223 + ], + "blocks": [ + { + "bbox": [ + 296, + 150, + 400, + 223 + ], + "lines": [ + { + "bbox": [ + 296, + 150, + 400, + 223 + ], + "spans": [ + { + "bbox": [ + 296, + 150, + 400, + 223 + ], + "type": "image", + "image_path": "d95fdf15a5be65475dbf5e31bd3b476704c76723c1cc16a4607b1cdbd2f346d1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 224, + 370, + 234 + ], + "lines": [ + { + "bbox": [ + 323, + 224, + 370, + 234 + ], + "spans": [ + { + "bbox": [ + 323, + 224, + 370, + 234 + ], + "type": "text", + "content": "(c) Sketch" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 239, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 271 + ], + "type": "text", + "content": "Figure 4. Detection results on target images with multiple different styles. The first and second rows respectively indicate results based on VOS [6] and our method. The in-distribution source dataset is BDD-100k [45]. 'Arcylic', 'Purple-Brush', and 'Sketch' represent the rendered target styles [20]. 'Nature' is the same style as the source dataset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 403, + 150, + 475, + 223 + ], + "blocks": [ + { + "bbox": [ + 403, + 150, + 475, + 223 + ], + "lines": [ + { + "bbox": [ + 403, + 150, + 475, + 223 + ], + "spans": [ + { + "bbox": [ + 403, + 150, + 475, + 223 + ], + "type": "image", + "image_path": "e075021f3e9080f07d6de26d0a6dc9f7571f718ecaeb057d3cf40e4b2948b008.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 224, + 460, + 234 + ], + "lines": [ + { + "bbox": [ + 411, + 224, + 460, + 234 + ], + "spans": [ + { + "bbox": [ + 411, + 224, + 460, + 234 + ], + "type": "text", + "content": "(d) Nature" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 478, + 150, + 550, + 223 + ], + "blocks": [ + { + "bbox": [ + 478, + 150, + 550, + 223 + ], + "lines": [ + { + "bbox": [ + 478, + 150, + 550, + 223 + ], + "spans": [ + { + "bbox": [ + 478, + 150, + 550, + 223 + ], + "type": "image", + "image_path": "fcaf5632b835e4236bd53303173211647c73603af2a8130f389be4f0c2f4fc18.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 224, + 533, + 234 + ], + "lines": [ + { + "bbox": [ + 486, + 224, + 533, + 234 + ], + "spans": [ + { + "bbox": [ + 486, + 224, + 533, + 234 + ], + "type": "text", + "content": "(e) Nature" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 56, + 277, + 555, + 399 + ], + "blocks": [ + { + "bbox": [ + 56, + 277, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 56, + 277, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 555, + 399 + ], + "type": "table", + "html": "
MethodVOCVOC-COCO-20VOC-COCO-40VOC-COCO-60
\\( \\mathrm{mAP}_{\\mathcal{K}\\uparrow} \\)\\( WI_{\\downarrow} \\)\\( AOSE_{\\downarrow} \\)\\( mAP_{\\mathcal{K}\\uparrow} \\)\\( AP_{\\mathcal{U}\\uparrow} \\)\\( WI_{\\downarrow} \\)\\( AOSE_{\\downarrow} \\)\\( mAP_{\\mathcal{K}\\uparrow} \\)\\( AP_{\\mathcal{U}\\uparrow} \\)\\( WI_{\\downarrow} \\)\\( AOSE_{\\downarrow} \\)\\( mAP_{\\mathcal{K}\\uparrow} \\)\\( AP_{u\\uparrow} \\)
FR-CNN [33]80.1018.391511858.45022.742339155.26018.492547255.830
\\( FR-CNN^{\\dagger} \\)[33]80.0118.831194157.91023.241825754.77018.721956655.340
PROSER [47]79.6819.161303557.6610.9224.151983154.667.6219.642132255.203.25
ORE [18]79.8018.181281158.252.6022.401975255.301.7018.352141555.470.53
DS [28]80.0416.981286858.355.1320.861977555.313.3917.222192155.771.25
DFDD [44]80.2612.731072760.2116.2214.921363857.9812.0912.571666857.515.12
OpenDet [11]80.0214.951128658.7514.9318.231680055.8310.5814.241825056.374.36
OpenDet + Ours80.5311.981004360.5317.1914.181267058.6212.4711.831623158.025.63
", + "image_path": "5e317d68234065002d01ba2d47b37b8d1329ccc5c7c785bbeacf81875d8e9a88.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 80, + 441, + 269, + 557 + ], + "blocks": [ + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "type": "text", + "content": "Table 3. Performance analysis of OSOD. We report close-set performance " + }, + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "type": "inline_equation", + "content": "(\\mathrm{mAP}_{\\mathcal{K}})" + }, + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "type": "text", + "content": " on VOC, and both close-set " + }, + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "type": "inline_equation", + "content": "(\\mathrm{mAP}_{\\mathcal{K}})" + }, + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "type": "text", + "content": " and open-set (WI, AOSE, " + }, + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathcal{U}}" + }, + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "type": "text", + "content": ") performance of different methods on VOC-COCO-{20, 40, 60}. Here, '20', '40', and '60' indicate that the testing COCO images separately contain 20, 40, and 60 non-VOC classes. '†' means a higher score threshold (i.e., 0.1) for testing." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 441, + 269, + 557 + ], + "lines": [ + { + "bbox": [ + 80, + 441, + 269, + 557 + ], + "spans": [ + { + "bbox": [ + 80, + 441, + 269, + 557 + ], + "type": "table", + "html": "
MethodNSDRNRDF
SW [30]33.426.313.730.8
IBN-Net [29]32.126.114.329.6
IterNorm [17]29.622.812.628.4
ISW [2]33.225.914.131.8
S-DGOD [41]36.628.216.633.5
S-DGOD + Ours38.230.418.335.7
Clip-gap [39]36.932.318.738.5
Clip-gap + Ours38.534.120.240.1
", + "image_path": "4f3dd6f0ff87db80d8ca4155daf475c36cb03f7d803c6a95b6a4d2e1c2076023.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 601, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 625 + ], + "type": "text", + "content": "bust discriminative boundary and accelerates the safe deployment of the model." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 635, + 246, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 246, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 246, + 647 + ], + "type": "text", + "content": "4.3. Ablation and Visualization Analysis" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 652, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 295, + 689 + ], + "type": "text", + "content": "We utilize PASCAL VOC as the source data and MS-COCO with 'Acrylic' style as the target data to perform an ablation analysis. Our method is directly plugged into VOS [6]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Analysis of WFS. As shown in Fig. 2, our method aims to emulate humans and mainly consists of three modules," + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 322, + 441, + 544, + 520 + ], + "blocks": [ + { + "bbox": [ + 55, + 559, + 295, + 593 + ], + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 593 + ], + "type": "text", + "content": "Table 4. Single-DGOD analysis. 'NS', 'DR', 'NR', and 'DF' separately represent Night-Sunny, Daytime-Rainy, Night-Rainy, and Daytime-Foggy weather. mAP is taken as the metric." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 441, + 544, + 520 + ], + "lines": [ + { + "bbox": [ + 322, + 441, + 544, + 520 + ], + "spans": [ + { + "bbox": [ + 322, + 441, + 544, + 520 + ], + "type": "table", + "html": "
PerceptMemoryImagineFPR95 ↓AUROC ↑
✓84.68%72.23%
✓83.13%72.75%
✓✓82.85%73.06%
✓✓81.07%73.94%
✓✓✓80.16%74.31%
", + "image_path": "5cb734242cb6161ff91ec1d11131c5e683ee0bb2f7d4db016483ae43fecbbf21.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 521, + 555, + 555 + ], + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 555 + ], + "type": "text", + "content": "Table 5. Ablation analysis of WFS for ODU-OD. Here, 'Percept', 'Memory', and 'Imagine' separately indicate multi-level perception, memory recorder, and unknown-feature generator." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "i.e., multi-level perception, fragmented memory recorder, and imagination-based unknown-feature generator. In Table 5, we make an ablation analysis of the three modules. We can observe that employing dedicated perception, memory, and imagination-based generator could effectively reduce the impact of two-shift entanglement and improve the detection performance significantly. Particularly, when using the generator for synthesizing unknown features, the performance could be further enhanced. This shows that imitating humans to leverage perception and memory to imagine unknowns is an effective solution for addressing two-shift entanglement, promoting learning a robust and discriminative boundary between known and unknown objects." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4688" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 149, + 126 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 149, + 126 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 149, + 126 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 149, + 126 + ], + "type": "image", + "image_path": "a7816d702ba4f333d4007aecd00c08c141b0deb97405e86bc8dda6eb1ef482a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 151, + 70, + 229, + 126 + ], + "blocks": [ + { + "bbox": [ + 151, + 70, + 229, + 126 + ], + "lines": [ + { + "bbox": [ + 151, + 70, + 229, + 126 + ], + "spans": [ + { + "bbox": [ + 151, + 70, + 229, + 126 + ], + "type": "image", + "image_path": "6af2c9e2e71696338163f5ef456cd442351d7150684119eb230452b80fedd699.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 230, + 70, + 309, + 126 + ], + "blocks": [ + { + "bbox": [ + 230, + 70, + 309, + 126 + ], + "lines": [ + { + "bbox": [ + 230, + 70, + 309, + 126 + ], + "spans": [ + { + "bbox": [ + 230, + 70, + 309, + 126 + ], + "type": "image", + "image_path": "840c243ca0081aedf46179da98ceb1ee24400bc515418db81e44cd1978c8c367.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 388, + 126 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 388, + 126 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 388, + 126 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 388, + 126 + ], + "type": "image", + "image_path": "fb78ab8bb3d46871eae72cfd6885f5e939b2380c9a97d9ce5cb2a66c31c06a7b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 467, + 126 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 467, + 126 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 467, + 126 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 467, + 126 + ], + "type": "image", + "image_path": "381157da76156e8862d8aee7b5fc77e89d9c657115fdad0cc52995a0e785afdd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 468, + 71, + 548, + 126 + ], + "blocks": [ + { + "bbox": [ + 468, + 71, + 548, + 126 + ], + "lines": [ + { + "bbox": [ + 468, + 71, + 548, + 126 + ], + "spans": [ + { + "bbox": [ + 468, + 71, + 548, + 126 + ], + "type": "image", + "image_path": "02b4052fedae00e0369823b996a2f6e3f3b63cc6280a4a4433bf4ed210c579a7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 70, + 129, + 149, + 185 + ], + "blocks": [ + { + "bbox": [ + 70, + 129, + 149, + 185 + ], + "lines": [ + { + "bbox": [ + 70, + 129, + 149, + 185 + ], + "spans": [ + { + "bbox": [ + 70, + 129, + 149, + 185 + ], + "type": "image", + "image_path": "29ecbfaaeec0562be98a0915ebdf26bd6ed1f4249086c4d04d15220343e03eca.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 553, + 220 + ], + "type": "text", + "content": "Figure 5. Memory vs. Imagination. The first and second rows respectively show the captured memory and reasoned imagination. Here, we directly project the memory codebook and imagination bank into the corresponding visual regions with three different styles. We can observe that for the same input, memory and imagination separately pay attention to different fragmented content." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 151, + 129, + 229, + 185 + ], + "blocks": [ + { + "bbox": [ + 151, + 129, + 229, + 185 + ], + "lines": [ + { + "bbox": [ + 151, + 129, + 229, + 185 + ], + "spans": [ + { + "bbox": [ + 151, + 129, + 229, + 185 + ], + "type": "image", + "image_path": "3d913f5daa46a8a7f2bd447dd8a8bafa459082df9b17a14e83c4690c56a49eec.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 230, + 129, + 309, + 185 + ], + "blocks": [ + { + "bbox": [ + 230, + 129, + 309, + 185 + ], + "lines": [ + { + "bbox": [ + 230, + 129, + 309, + 185 + ], + "spans": [ + { + "bbox": [ + 230, + 129, + 309, + 185 + ], + "type": "image", + "image_path": "4c4e6355119b3e276a67ba4fe924170d467cf0b28060198c68268a7d219b2be7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 129, + 388, + 185 + ], + "blocks": [ + { + "bbox": [ + 309, + 129, + 388, + 185 + ], + "lines": [ + { + "bbox": [ + 309, + 129, + 388, + 185 + ], + "spans": [ + { + "bbox": [ + 309, + 129, + 388, + 185 + ], + "type": "image", + "image_path": "7cd235b86860394974936228a54e67824ff760baa295725eb1ae932edc25f4e6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 389, + 129, + 467, + 185 + ], + "blocks": [ + { + "bbox": [ + 389, + 129, + 467, + 185 + ], + "lines": [ + { + "bbox": [ + 389, + 129, + 467, + 185 + ], + "spans": [ + { + "bbox": [ + 389, + 129, + 467, + 185 + ], + "type": "image", + "image_path": "67fb31a0290ba2146bcbc9cdf5d60aeced81e4635235efbd2d1fdc7b6ce63087.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 468, + 129, + 548, + 185 + ], + "blocks": [ + { + "bbox": [ + 468, + 129, + 548, + 185 + ], + "lines": [ + { + "bbox": [ + 468, + 129, + 548, + 185 + ], + "spans": [ + { + "bbox": [ + 468, + 129, + 548, + 185 + ], + "type": "image", + "image_path": "f28a8d5250108df04dc8a8c8fa37b59826f26e33a2f7047968339d09705ba33c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 67, + 227, + 141, + 285 + ], + "blocks": [ + { + "bbox": [ + 67, + 227, + 141, + 285 + ], + "lines": [ + { + "bbox": [ + 67, + 227, + 141, + 285 + ], + "spans": [ + { + "bbox": [ + 67, + 227, + 141, + 285 + ], + "type": "image", + "image_path": "6da988c23ba6f49c6ce8fcf255d1d31da9393ceba003f9ee115ce84767830037.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 142, + 227, + 216, + 284 + ], + "blocks": [ + { + "bbox": [ + 142, + 227, + 216, + 284 + ], + "lines": [ + { + "bbox": [ + 142, + 227, + 216, + 284 + ], + "spans": [ + { + "bbox": [ + 142, + 227, + 216, + 284 + ], + "type": "image", + "image_path": "536bdda7e54ae6801b7b06982362501cd946a4f352cdfa36068d22764da80f84.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 217, + 227, + 291, + 284 + ], + "blocks": [ + { + "bbox": [ + 217, + 227, + 291, + 284 + ], + "lines": [ + { + "bbox": [ + 217, + 227, + 291, + 284 + ], + "spans": [ + { + "bbox": [ + 217, + 227, + 291, + 284 + ], + "type": "image", + "image_path": "9226f751195059ca222a8ff48cbbbde0750b063618b0c1778e649df26626205d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 67, + 287, + 141, + 343 + ], + "blocks": [ + { + "bbox": [ + 67, + 287, + 141, + 343 + ], + "lines": [ + { + "bbox": [ + 67, + 287, + 141, + 343 + ], + "spans": [ + { + "bbox": [ + 67, + 287, + 141, + 343 + ], + "type": "image", + "image_path": "33cbb57dc76ac45e4af8fab885015ebd12158ba8f6e2e38353b7436091f72de3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 142, + 287, + 216, + 343 + ], + "blocks": [ + { + "bbox": [ + 142, + 287, + 216, + 343 + ], + "lines": [ + { + "bbox": [ + 142, + 287, + 216, + 343 + ], + "spans": [ + { + "bbox": [ + 142, + 287, + 216, + 343 + ], + "type": "image", + "image_path": "0ee62376f41a930d22c0443dad774b5ba4488222aae85f3c989c5be96318c36b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 217, + 287, + 291, + 343 + ], + "blocks": [ + { + "bbox": [ + 217, + 287, + 291, + 343 + ], + "lines": [ + { + "bbox": [ + 217, + 287, + 291, + 343 + ], + "spans": [ + { + "bbox": [ + 217, + 287, + 291, + 343 + ], + "type": "image", + "image_path": "c7b84bd9511e4fe03ac570c643b76688ebe0346dbb87beab0b9a6418f948b657.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 67, + 347, + 140, + 403 + ], + "blocks": [ + { + "bbox": [ + 67, + 347, + 140, + 403 + ], + "lines": [ + { + "bbox": [ + 67, + 347, + 140, + 403 + ], + "spans": [ + { + "bbox": [ + 67, + 347, + 140, + 403 + ], + "type": "image", + "image_path": "19d3dbd4b28d1a3c1b77270d7acc3f98cb09352ed9358e089919011f8f429862.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 404, + 129, + 415 + ], + "lines": [ + { + "bbox": [ + 73, + 404, + 129, + 415 + ], + "spans": [ + { + "bbox": [ + 73, + 404, + 129, + 415 + ], + "type": "text", + "content": "(a) Style Image" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 416, + 295, + 450 + ], + "lines": [ + { + "bbox": [ + 55, + 416, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 295, + 450 + ], + "type": "text", + "content": "Figure 6. Visualization of the memory-enhanced map " + }, + { + "bbox": [ + 55, + 416, + 295, + 450 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 416, + 295, + 450 + ], + "type": "text", + "content": " and generated unknown map " + }, + { + "bbox": [ + 55, + 416, + 295, + 450 + ], + "type": "inline_equation", + "content": "S_{u}" + }, + { + "bbox": [ + 55, + 416, + 295, + 450 + ], + "type": "text", + "content": ". For each feature map, the channels corresponding to the maximum value are selected for visualization." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 142, + 347, + 216, + 403 + ], + "blocks": [ + { + "bbox": [ + 142, + 347, + 216, + 403 + ], + "lines": [ + { + "bbox": [ + 142, + 347, + 216, + 403 + ], + "spans": [ + { + "bbox": [ + 142, + 347, + 216, + 403 + ], + "type": "image", + "image_path": "e12ec459e8eed967cc1a57b37f87ab1c48e87ac1ef27da8e75600495629c2582.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 404, + 205, + 414 + ], + "lines": [ + { + "bbox": [ + 147, + 404, + 205, + 414 + ], + "spans": [ + { + "bbox": [ + 147, + 404, + 205, + 414 + ], + "type": "text", + "content": "(b) Enhanced " + }, + { + "bbox": [ + 147, + 404, + 205, + 414 + ], + "type": "inline_equation", + "content": "E" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 217, + 347, + 291, + 403 + ], + "blocks": [ + { + "bbox": [ + 217, + 347, + 291, + 403 + ], + "lines": [ + { + "bbox": [ + 217, + 347, + 291, + 403 + ], + "spans": [ + { + "bbox": [ + 217, + 347, + 291, + 403 + ], + "type": "image", + "image_path": "39d599cf85465a8b287e8a84bba31e5c9011b414f081e8e0d1a894c1486602ba.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 405, + 280, + 415 + ], + "lines": [ + { + "bbox": [ + 220, + 405, + 280, + 415 + ], + "spans": [ + { + "bbox": [ + 220, + 405, + 280, + 415 + ], + "type": "text", + "content": "(c) Unknown " + }, + { + "bbox": [ + 220, + 405, + 280, + 415 + ], + "type": "inline_equation", + "content": "S_{u}" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 459, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 626 + ], + "type": "text", + "content": "Analysis of Multi-level Perception. As shown in Fig. 1 and 2, the goal of multi-level perception is to construct a complete impression of objects. Here, we make an ablation analysis of three perception levels. We only change the perception level. Other operations and settings are kept unchanged. Table 6 shows the results. We can observe that employing different perception level affects the corresponding detection performance. Particularly, since the connective level considers the relationships between visual regions, using this level could enhance the performance effectively. When combining these three levels, the performance could be further improved. This indicates that multi-level perception is indeed conducive to building a complete observation of objects, strengthening the robustness." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "Analysis of Codeword Number " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": ". For the fragmented memory recorder, we utilize " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " codewords to preserve the information of known objects. Here, we analyze the " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " impact. We observe that when " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " is set to 16 and 32, the FPR95 value increases slightly, i.e., " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "80.54\\%" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "80.33\\%" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": ". The reason may be that utilizing fewer codewords could not sufficiently capture known characteristics. Utilizing more code" + } + ] + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 323, + 228, + 544, + 328 + ], + "blocks": [ + { + "bbox": [ + 323, + 228, + 544, + 328 + ], + "lines": [ + { + "bbox": [ + 323, + 228, + 544, + 328 + ], + "spans": [ + { + "bbox": [ + 323, + 228, + 544, + 328 + ], + "type": "table", + "html": "
LocalGlobalConnectiveFPR95 ↓AUROC ↑
✓81.19%73.04%
✓81.62%73.18%
✓81.37%73.21%
✓✓80.85%73.96%
✓✓80.63%74.12%
✓✓80.58%74.09%
✓✓✓80.16%74.31%
", + "image_path": "d431760264ff824eb3c18e7895deb3efd735c5acefece166dd316b42a35c15b3.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 330, + 553, + 353 + ], + "lines": [ + { + "bbox": [ + 313, + 330, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 553, + 353 + ], + "type": "text", + "content": "Table 6. Ablation analysis of local, global, and connective perception perspectives. Other operations are kept unchanged." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 359, + 522, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 522, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 522, + 371 + ], + "type": "text", + "content": "words increases parameters, resulting in overfitting." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 372, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 553, + 491 + ], + "type": "text", + "content": "In Fig. 5, we intuitively show the corresponding content captured by the built memory and reasoned imagination. Obviously, for the images with different styles, the memory recorder can effectively focus on fragmented object-related content, enhancing the robustness against the covariate-shift impact. Furthermore, the reasoned imagination mainly pays attention to background regions that involve plentiful visual information, which is instrumental in promoting the following generator to synthesize expected unknown features and attenuates the impact of two-shift entanglement." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "text", + "content": "Visualization Analysis. Fig. 6 shows three examples of memory-enhanced map and virtual unknown map. We can further observe that fusing the known memory could indeed strengthen object-related information. The synthesized unknown features mainly pay attention to background regions involving plentiful visual characteristics, enlarging the gap with the known features significantly. In supplementary material, we will give more analysis about our method." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 314, + 597, + 388, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 388, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 388, + 609 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "This paper explores a challenging problem, i.e., ODU-OD, which aims to detect unknown objects within unseen target domains. To weaken the entangled impact of covariate and semantic shifts, we propose a new WFS method and imitate humans to leverage multi-level perception, fragmented memory, and reasoned imagination to recurrently generate virtual unknown features. Extensive experimental results on four tasks demonstrate the superiorities of our method." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4689" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work is supported in part by the National Key R&D Program of China (No. 2023YFC3305600), Joint Fund of Ministry of Education of China (8091B022149, 8091B02072404), National Natural Science Foundation of China (62472333, 62132016, 62171343), and Natural Science Basic Research Program of Shaanxi (2020JC-23)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 194, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 194, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 295, + 228 + ], + "type": "text", + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ECCV, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 282 + ], + "type": "text", + "content": "[2] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, pages 11580-11590, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 338 + ], + "type": "text", + "content": "[3] Muhammad Sohail Danish, Muhammad Haris Khan, Muhammad Akhtar Munir, M Saqib Sarfraz, and Mohsen Ali. Improving single domain-generalized object detection: A focus on diversification and alignment. In CVPR, pages 17732-17742, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 295, + 373 + ], + "type": "text", + "content": "[4] Xuefeng Du, Gabriel Gozum, Yifei Ming, and Yixuan Li. Siren: Shaping representations for detecting out-of-distribution objects. In NeurIPS, 2022. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 375, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 296, + 418 + ], + "type": "text", + "content": "[5] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware object detection: Learning what you don't know from videos in the wild. In CVPR, pages 13678-13688, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 420, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 452 + ], + "type": "text", + "content": "[6] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. ICLR, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 454, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 295, + 486 + ], + "type": "text", + "content": "[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 488, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 295, + 521 + ], + "type": "text", + "content": "[8] Yanchen Guan, Haicheng Liao, Zhenning Li, Guohui Zhang, and Chengzhong Xu. World models for autonomous driving: An initial survey. arXiv preprint arXiv:2403.02622, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 522, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 295, + 544 + ], + "type": "text", + "content": "[9] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 545, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 295, + 578 + ], + "type": "text", + "content": "[10] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. *ICLR*, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 579, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 623 + ], + "type": "text", + "content": "[11] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In CVPR, pages 9591-9600, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "text", + "content": "[12] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 2, 3, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "text", + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 139 + ], + "type": "text", + "content": "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. *ICLR*, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 554, + 184 + ], + "type": "text", + "content": "[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, pages 10951-10960, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 186, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 554, + 217 + ], + "type": "text", + "content": "[17] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, pages 4874-4883, 2019. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "text", + "content": "[18] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In CVPR, pages 5830-5840, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 254, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 285 + ], + "type": "text", + "content": "[19] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[20] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. arXiv preprint arXiv:2112.00374, 2021. 2, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "text", + "content": "[21] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. *ICLR*, 2018. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 355, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 387 + ], + "type": "text", + "content": "[22] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. NeurIPS, 31, 2018. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "text", + "content": "[23] Deng Li, Among Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In CVPR, pages 17606-17615, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "text", + "content": "[24] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. *ICLR*, 2017. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "text", + "content": "[25] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In ICCV, pages 8771-8780, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "text", + "content": "[26] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation models? ECCV, 2024. 1, 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 535, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 566 + ], + "type": "text", + "content": "[27] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. NeurIPS, 33: 21464-21475, 2020. 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "text", + "content": "[28] Dimity Miller, Lachlan Nicholson, Feras Dayoub, and Niko Sünderhauf. Dropout sampling for robust object detection in open-set conditions. In ICRA, pages 3243-3249, 2018. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "type": "text", + "content": "[29] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, pages 464-479, 2018. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[30] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, pages 1863-1871, 2019. 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[31] Bimsara Pathiraja, Malitha Gunawardhana, and Muhammad Haris Khan. Multiclass confidence and localization calibration for object detection. In CVPR, pages 19734-19743, 2023. 1" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4690" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[32] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, pages 7263-7271, 2017." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 95, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 95, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 95, + 295, + 138 + ], + "type": "text", + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurlPS, pages 91-99, 2015. 1, 3, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 140, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 140, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 57, + 140, + 294, + 172 + ], + "type": "text", + "content": "[34] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with gram matrices. In ICML, pages 8491-8501. PMLR, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 173, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 173, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 173, + 295, + 205 + ], + "type": "text", + "content": "[35] Eleanor Spens and Neil Burgess. A generative model of memory construction and consolidation. Nature Human Behaviour, pages 1-18, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 206, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 206, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 57, + 206, + 294, + 249 + ], + "type": "text", + "content": "[36] Weinan Sun, Madhu Advani, Nelson Spruston, Andrew Saxe, and James E Fitzgerald. Organizing memories for generalization in complementary learning systems. Nature neuroscience, 26(8):1438-1448, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 251, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 251, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 251, + 294, + 293 + ], + "type": "text", + "content": "[37] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. NeurIPS, 33:11839-11852, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 294, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 294, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 57, + 294, + 294, + 326 + ], + "type": "text", + "content": "[38] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. *ICLR*, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 327, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 327, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 57, + 327, + 294, + 360 + ], + "type": "text", + "content": "[39] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, pages 3219-3229, 2023. 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 361, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 361, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 57, + 361, + 294, + 403 + ], + "type": "text", + "content": "[40] Samuel Wilson, Tobias Fischer, Feras Dayoub, Dimity Miller, and Niko Sunderhauf. Safe: Sensitivity-aware features for out-of-distribution object detection. In ICCV, pages 23565-23576, 2023. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 405, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 405, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 57, + 405, + 294, + 437 + ], + "type": "text", + "content": "[41] Among Wu and Cheng Deng. Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation. In CVPR, pages 847-856, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 438, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 438, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 57, + 438, + 294, + 480 + ], + "type": "text", + "content": "[42] Among Wu and Cheng Deng. Discriminating known from unknown objects via structure-enhanced recurrent variational autoencoder. In CVPR, pages 23956-23965, 2023. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 482, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 482, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 482, + 294, + 535 + ], + "type": "text", + "content": "[43] Among Wu, Yahong Han, Linchao Zhu, and Yi Yang. Instance-invariant domain adaptive object detection via progressive disentanglement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4178-4193, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 537, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 537, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 57, + 537, + 294, + 569 + ], + "type": "text", + "content": "[44] Among Wu, Da Chen, and Cheng Deng. Deep feature deblurring diffusion for detecting out-of-distribution objects. In ICCV, pages 13381-13391, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 571, + 294, + 613 + ], + "type": "text", + "content": "[45] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, pages 2636-2645, 2020. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 615, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 615, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 615, + 294, + 635 + ], + "type": "text", + "content": "[46] Hang Zhang, Jia Xue, and Kristin Dana. Deep ten: Texture encoding network. In CVPR, pages 708-717, 2017. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 637, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 294, + 668 + ], + "type": "text", + "content": "[47] Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Learning placeholders for open-set recognition. In CVPR, pages 4401-4410, 2021. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[48] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4396-4415, 2022. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[49] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 3" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4691" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/a1c25909-0376-4a01-a3cb-0e20c3229024_content_list.json b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/a1c25909-0376-4a01-a3cb-0e20c3229024_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..60035d909fc12ec2fb44988751421c532d3c7958 --- /dev/null +++ b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/a1c25909-0376-4a01-a3cb-0e20c3229024_content_list.json @@ -0,0 +1,1521 @@ +[ + { + "type": "text", + "text": "Perception Tokens Enhance Visual Reasoning in Multimodal Language Models", + "text_level": 1, + "bbox": [ + 101, + 130, + 897, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mahtab Bigverdi1 Zelun Luo2 Cheng-Yu Hsieh1 \nEthan Shen1 Dongping Chen1 Linda G. Shapiro1 Ranjay Krishna1 \n1University of Washington, 2Google Research", + "bbox": [ + 186, + 179, + 808, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mahtab, cydhsieh, ethans03, shapiro, ranjay}@cs.washington.edu, alanzluo@gmail.com, cdp0612@uw.edu", + "bbox": [ + 147, + 234, + 844, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal language models (MLMs) still face challenges in fundamental visual perception tasks where specialized models excel. Tasks requiring reasoning about 3D structures benefit from depth estimation, and reasoning about 2D object instances benefits from object detection. Yet, MLMs can not produce intermediate depth or boxes to reason over. Fine-tuning MLMs on relevant data doesn't generalize well and outsourcing computation to specialized vision tools is too comput-intensive and memory-inefficient. To address this, we introduce Perception Tokens, intrinsic image representations designed to assist reasoning tasks where language is insufficient. Perception tokens act as auxiliary reasoning tokens, akin to chain-of-thought prompts in language models. For example, in a depth-related task, an MLM augmented with perception tokens can reason by generating a depth map as tokens, enabling it to solve the problem effectively. We propose AURORA, a training method that augments MLMs with perception tokens for improved reasoning over visual inputs. AURORA leverages a VQVAE to transform intermediate image representations, such as depth maps into a tokenized format and bounding box tokens, which are then used in a multi-task training framework. AURORA achieves notable improvements across counting benchmarks: $+10.8\\%$ on BLINK, $+11.3\\%$ on CVBench, and $+8.3\\%$ on SEED-Bench, outperforming fine-tuning approaches in generalization across datasets. It also improves on relative depth: over $+6\\%$ on BLINK. With perception tokens, AURORA expands the scope of MLMs beyond language-based reasoning, paving the way for more effective visual reasoning capabilities. Code and data will be released at the project page.", + "bbox": [ + 88, + 335, + 483, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 844, + 220, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to the growing emphasis on building multimodal language models (MLMs), computer vision was originally", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85b4a617c42d599b1615a6f5b14654d479f7d1b763f68b4de3c3b9787b89dabf.jpg", + "image_caption": [ + "Which point is closer to the camera?" + ], + "image_footnote": [], + "bbox": [ + 516, + 382, + 653, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LLaVA-Aurora", + "text_level": 1, + "bbox": [ + 735, + 311, + 830, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Step 1:", + "bbox": [ + 666, + 335, + 709, + 348 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point A is at (250, 200), B is at (240, 300), and C is at (480, 210)", + "bbox": [ + 725, + 338, + 888, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Step 2:", + "bbox": [ + 668, + 369, + 709, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "I need depth estimation to determine which point is closer", + "bbox": [ + 725, + 371, + 887, + 392 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Step 3:", + "bbox": [ + 668, + 402, + 709, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The depth map of the input image is: ... ", + "bbox": [ + 725, + 402, + 883, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/476fccd735a2ec8a8ed2a13c82625b56855dd87f6eb661f488c3301988b45edf.jpg", + "image_caption": [ + "Figure 1. We introduce Perception Tokens, intermediate reasoning tokens that allow MLMs to go beyond using language in reasoning. With it, we develop AURORA, a framework that trains multimodal language models to leverage visual perception tokens, allowing them to use depth estimation and bounding box predictions while reasoning." + ], + "image_footnote": [], + "bbox": [ + 725, + 434, + 781, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Step 4:", + "bbox": [ + 668, + 483, + 709, + 493 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point B has higher pixel value on depth map, so it is closer", + "bbox": [ + 725, + 484, + 879, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "attempting to interpret images as projections of indescribable 3D intrinsics, not just processing 2D arrays of language patterns [9, 34, 35]. Towards this endeavor, early vision research developed a series of intermediate image representations—enabling geometric reasoning through depth estimation [42] and instance reasoning through bounding box grounding [14]. As pointed out by recent work, we have focused less on such perceptual representations and instead tackled reasoning problems that require limited visual involvement [11, 40, 41]. This is likely because many traditional vision tasks remain ambiguous through natural language. Consider the task of identifying which of a set of N points is furthest away from the camera. While language doesn't lend itself to reason over this problem, a depth estimation would provide the appropriate abstraction to reason over.", + "bbox": [ + 511, + 625, + 906, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Numerous attempts have been made to enable MLMs to reason over intrinsic image representations. The default ap-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3836", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86e7ba82d877c5ffe4ff24d7d3e8c68df57535da89c7e4f1fbe07fcbbfb256d5.jpg", + "image_caption": [ + "Figure 2. We demonstrate relative depth estimation and counting questions where LLaVA fails. In contrast, by learning to utilize visual perception tokens as intermediate reasoning steps, AURORA successfully complete these tasks requiring perceptual understanding." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proach is to fine-tune MLMs on data tailored to the specific perception task of interest, hoping that the model implicitly learns the required intrinsic representations [4]. Another option is to outsource the computation to external tools: the MLM can invoke a depth estimator or object detector to produce the appropriate intrinsic [17]. Unfortunately, relying on external models makes the task more computationally expensive and requires loading additional models with more memory. Similarly, vanilla fine-tuning (even with advancements like LoRA [16]) has shown marginal improvements.", + "bbox": [ + 88, + 297, + 480, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conceptually, we introduce Perception Tokens, intrinsic image representations that aid in reasoning where language is insufficient. To solve the aforementioned task, an MLM augmented with perception tokens can solve the task similar to how language models use chain-of-thought. They will produce a response like the following: \"The depth map is . Therefore, point $D$ is closest to the camera.\" Here, is a set of tokens that implicitly estimate the depth of the image. Similarly, for a counting task, the model can first generate perception tokens that represents the location of the relevant bounding boxes of the desired object, and count the number of boxes to support its final answer.", + "bbox": [ + 88, + 453, + 482, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate the utility of perception tokens, we introduce AURORA, a training algorithm to augment MLMs with the ability to use perception tokens as intermediate reasoning steps. For certain intermediate representations (e.g. depth maps), we train a VQVAE to transform them into a set of tokens, treating the learned VQVAE codebook indices as a collection of perception tokens. while for others, such as bounding boxes, we use directly encoded structured tokens. Next, we follow a multi-task training approach [15, 18] to train MLMs to use perception tokens as chain-of-thought tokens (see Fig. 1). Additionally, we adopt a curriculum learning approach to avoid catastrophic forgetting.", + "bbox": [ + 88, + 654, + 482, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We apply the AURORA training algorithm to the LLaVA [27] model, resulting in our LLaVA-AURORA variant. Our LLaVA-AURORA model significantly outperforms standard fine-tuning approaches across multiple perception", + "bbox": [ + 88, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "demanding tasks, demonstrating the generality and effectiveness of our method. LLaVA-AURORA achieves state-of-the-art results on both relative depth estimation and object counting tasks. For instance, on BLINK relative depth estimation, LLaVA-AURORA delivers a performance boost of $6.4\\%$ compared to the fine-tuning baseline. Similarly, on counting tasks, LLaVA-AURORA drives improvements of $10.8\\%$ on BLINK, $11.3\\%$ on CVBench, and $8.3\\%$ on SEED-Bench. Fig. 2 illustrates examples from these tasks. Perception tokens open up a whole new modality through which MLMs can begin to reason, tackling tasks beyond just language reasoning.", + "bbox": [ + 511, + 297, + 906, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 501, + 651, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal language models (MLMs). MLMs aim to solve a variety of tasks (e.g., Visual Question Answering (VQA) and captioning) based on vision and language inputs. Most modern architectures accomplish this by relying on either cross-attention [1, 2] or visual instruction tuning [26-28, 51] to interleave multimodal information. Cross-attention architectures operate by independently encoding images and then cross-attending to a language model backbone. On the other hand, visual instruction tuning produces token embeddings from image representations that can be interleaved with language tokens to ground generation. While these techniques work well for high level tasks, MLMs still struggle with mid-level and low-level tasks such as counting, depth reasoning, and segmentation. Most MLMs can be classified as either end-to-end MLMs or tool-using MLMs.", + "bbox": [ + 511, + 529, + 906, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "End-to-end MLMs. End-to-end MLMs use a single, unified architecture [1, 8, 24, 26, 27] that can be repeatedly used for different tasks without requiring any architectural changes. While training end-to-end MLMs can be costly due to the need for large amounts of multi-task data, the use of diverse datasets allows end-to-end MLMs to generalize effectively and learn nuanced visual representations.", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3837", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tool-using MLMs. Tool-using MLMs enable LLMs to perform vision tasks by attaching specialized vision modules such as segmentation or captioning networks [30, 33, 46, 52]. Tool-using MLMs use a router to select the optimal vision module to use for a given input. The use of specialized modules allows MLMs to achieve higher performances on many tasks. However, tool-using MLMs are sensitive to errors because they link together several networks, which are each a potential point of failure.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Any-to-any networks. To address this, recent works [19, 20, 38, 45] such as Unified-IO [31, 32] have experimented with shared embedding spaces and visual decoders for vision and language tasks, training MLMs to generate segmentation masks, keypoints, and depth maps. Similarly, [47] generates special task tokens to route language model representations to diffusion heads for image, video, and audio generation. To handle complex segmentation scenarios, LISA [22] and GSVA [48] train language models to generate an additional segmentation token that can be used to produce segmentation masks, grounding the mask in language reasoning.", + "bbox": [ + 89, + 237, + 482, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While these approaches can generate visual outputs, they cannot reason over their own generations to solve related visual perception tasks. In contrast, augmenting MLMs with perception tokens as chain-of-thought tokens enables them to perform visual reasoning directly, providing significant gains in detail-oriented question-answering tasks such as depth estimation and counting.", + "bbox": [ + 89, + 417, + 480, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Perception Tokens & Aurora", + "text_level": 1, + "bbox": [ + 89, + 539, + 356, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce Perception Tokens and our Aurora training algorithm (see Fig. 3), which augment multimodal language models with perception tokens, enabling the MLM to leverage these tokens effectively during training and incorporate them into its chain-of-thought reasoning process for enhanced visual reasoning.", + "bbox": [ + 89, + 564, + 482, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem formulation", + "text_level": 1, + "bbox": [ + 89, + 666, + 287, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In autoregressive large language models, chain-of-thought (CoT) reasoning can be formulated as a multi-step inferential process in which the model iteratively generates intermediate reasoning steps to arrive at a final answer. Given a task input $x$ , the model generates a response $y$ conditioned on the input and a sequence of $m$ intermediate reasoning steps $\\{s_i\\}_{i=1}^m$ , where $x, y$ , and each $s_i$ are sequences of tokens from the model's vocabulary $V$ .", + "bbox": [ + 89, + 688, + 482, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing multimodal language models often rely on limited vocabulary tokens derived from text or pre-trained image embeddings like CLIP, restricting their capacity to interpret other representations crucial for reasoning. Mid-level and low-level visual features such as depth maps, instance segmentation masks, human poses and bounding", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "boxes which could substantially improve visual reasoning, are currently incompatible with the model and cannot be integrated during training or inference. Our key insight is to introduce auxiliary perception tokens for these intermediate steps with an expanded vocabulary $V' = V \\cup V_{\\mathrm{aux}}$ , bridging this gap by allowing the model to integrate richer visual representations into its reasoning process. Conditioning the final output on these tokens enhances the model's accuracy and interpretability across multimodal tasks.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Perception token prediction and reasoning", + "text_level": 1, + "bbox": [ + 511, + 237, + 874, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Introducing an expanded vocabulary to enhance multimodal reasoning presents two main challenges. The first challenge is enabling the model to generate tokens from the new auxiliary vocabulary $V_{\\mathrm{aux}}$ , which includes specialized tokens for low- and mid-level visual features, such as depth maps and bounding boxes. The second challenge is ensuring that the model can effectively condition on these auxiliary tokens to improve reasoning, particularly for multi-step inference tasks.", + "bbox": [ + 511, + 258, + 903, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Perception token prediction. To address the first challenge, we employ a specialist-to-generalist distillation approach, using pre-trained specialist models (e.g., depth estimation or instance segmentation) to guide auxiliary token generation through cross-entropy loss. For each input $x$ , the specialist model provides a target probability distribution $q_{i}$ over its tokens. Let $M: V_{\\mathrm{spec}} \\to V_{\\mathrm{aux}}$ denote a one-to-one mapping from the specialist model's vocabulary $V_{\\mathrm{spec}}$ to the auxiliary token vocabulary $V_{\\mathrm{aux}}$ . We define the distillation loss as:", + "bbox": [ + 511, + 402, + 903, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {d i s t} = \\min _ {M} \\left(- \\sum_ {i} q _ {i} \\log p _ {M (i)}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 551, + 903, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_{M(i)}$ is the probability assigned by our model to the auxiliary token corresponding to the mapped token $M(i)$ . This consistent mapping allows the model to effectively align its predictions with the specialist model's output distribution, enhancing the relevance and accuracy of its auxiliary token predictions.", + "bbox": [ + 511, + 587, + 903, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to distillation, we incorporate a reconstruction loss to enhance the token prediction and the interpretability of our model. Each set of auxiliary tokens corresponds to a specific representation, such as a depth map or a bounding box vector, and is trained to directly predict this feature. To achieve this, we introduced a lightweight decoder $g$ that maps the tokens into the feature space, allowing for efficient and interpretable transformations. Formally, for a token set $T \\subset V_{\\mathrm{aux}}$ , decoder $g$ , and its target feature $f$ , the reconstruction loss is defined as:", + "bbox": [ + 511, + 679, + 905, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {r e c} = \\left\\| g (T) - f \\right\\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 840, + 903, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $g(T)$ is the decoded representation for the token set $T$ in the feature space. This reconstruction process not only", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3838", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c15e4918570fa52681db82a6387e166441e0238746739e6b88dd82855da39622.jpg", + "image_caption": [ + "Figure 3. The overall AURORA training framework. We first learn visual perception tokens using VQVAE. We then finetune MLMs with a multi-task training approach where we distill intrinsic image representations (e.g., depth map) into MLMs by training them to decode the visual tokens as intermediate reasoning steps towards completing the tasks." + ], + "image_footnote": [], + "bbox": [ + 155, + 89, + 843, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "aligns each token with a meaningful feature, improving interpretability, but also reinforces the accuracy of auxiliary token predictions through direct feature supervision. In practice, while the reconstruction objective improves performance and interpretability, it is optional, as models can be effectively trained with the distillation objective alone to reduce computational overhead.", + "bbox": [ + 88, + 369, + 480, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reasoning with perception tokens. The second challenge is enabling the model to condition on tokens from $V_{\\mathrm{aux}}$ effectively when generating each subsequent reasoning step, thereby enhancing its reasoning capabilities. To achieve this, we progressively introduce chain-of-thought reasoning during training, starting with simpler, single-step reasoning examples and gradually incorporating more complex, multi-step reasoning data. The model begins by learning single-step reasoning, predicting an initial reasoning step $s_1$ based on the input $x$ . It then progresses to multistep reasoning, predicting sequences $s_1, \\dots, s_m$ and effectively utilizing auxiliary tokens to support extended chains of thought. To facilitate this, we propose a multi-task data synthesis approach combined with curriculum learning across a variety of synthetic tasks. Additional implementation details can be found in Section 3.4. Finally, we reinforce this process at inference time using constrained decoding, wherein sampling is selectively restricted to auxiliary tokens at certain reasoning steps, guiding the model to generate coherent intermediate steps from $V_{\\mathrm{aux}}$ .", + "bbox": [ + 89, + 488, + 483, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tokenization", + "text_level": 1, + "bbox": [ + 89, + 801, + 225, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A unified tokenization space is crucial for multimodal models as it creates a consistent framework through which varied visual tasks can be represented, processed, and interpreted. Inspired by [36], we establish a unified tokenization space which enables the model to learn varied visual fea", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tures within a shared representation seamlessly. In our experiments, we implement two tailored tokenization schemes for commonly used visual representations. Importantly, our framework is designed with flexibility, allowing it to generalize to a broad range of visual representations beyond those presented here.", + "bbox": [ + 511, + 369, + 906, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pixel-level representation. This tokenization scheme captures fine-grained spatial information, such as depth maps and segmentation masks, providing the model with detailed pixel-level data essential for accurate visual processing. For these types of tokens, we leverage visual tokenizers like VQVAE and VQGAN, which take in the ground truth masks or depth maps and return discrete target tokens [10, 31, 32, 44].", + "bbox": [ + 511, + 468, + 908, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structured representation. This scheme encodes structured yet abstracted visual features, such as human poses, bounding boxes, and coordinates, allowing the model to reason with higher-level spatial relationships and object hierarchies. For these tokens, we define the domain of the tokens based on specific properties; for example, the domain for coordinates can range from 0 to the maximum number of pixels in the image's height or width [5].", + "bbox": [ + 511, + 597, + 908, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Curriculum learning with progressive CoTs", + "text_level": 1, + "bbox": [ + 511, + 727, + 883, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The objective of training the model is to develop a data and computationally efficient method for learning to predict novel, fine-grained visual tokens and using them to complete complex visual reasoning tasks. We observed that the standard approach, which relies on a fixed mixture data, encounters a trade-off between the accuracy of novel tokens predictions and the model's reasoning capability, primarily due to catastrophic forgetting and challenges in reasoning with new tokens. Conversely, fine-tuning the model with the original training mixture significantly raises computational", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3839", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/0c5bc260074474996613d20051b24dc291da2d4ea7c7c780cf71898cec683e39.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTrainingBLINK [11]2 PointsHardBLINK3 PointsHardBLINK4 PointsHardBLINK5 PointsAverage
DirectLa-labelingDataDepthGener-ationDataCoTData
LLaVA OneVision 72BXXX67.751.650.044.353.4
LLaVA 1.5 13BXXX54.035.537.92939.1
Fine-tuned LLaVA✓XX68.558.952.441.155.2
LLaVA-AURORA (Ours)✓✓✓64.566.960.554.861.6
GPT-4oXXX53.258.95036.349.6
GPT-4 TurboXXX58.154.841.932.246.7
GPT-4 Turbo + ToolXXX70.257.244.326.649.6
", + "bbox": [ + 107, + 88, + 888, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Performance comparison between our fine-tuned model, the baseline (without depth tokens), and the original base model on the relative depth accuracy (\\%) task. Results demonstrate that our approach, utilizing depth tokens and intermediate reasoning steps, significantly outperforms both the baseline and the base model, particularly on more challenging configurations with 3, 4, and 5 points sampled from the image's mid-height region. The gray rows represent closed-source models, providing additional baselines for comparison.", + "bbox": [ + 88, + 273, + 906, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "costs and may be impractical if the original data, particularly proprietary datasets, is unavailable, making this approach less scalable for incorporating new tokens in the future.", + "bbox": [ + 88, + 357, + 480, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a curriculum learning-inspired training scheme that begins with atomic tasks and gradually advances to more complex ones requiring sophisticated, multi-hop reasoning. Let $d_{t}$ represent the difficulty of task $t$ , with difficulties $d_{1} < d_{2} < \\dots d_{T}$ across $T$ tasks, and let $p(t,s)$ denote the probability of sampling data points from task $t$ at training step $s$ . We define $p(d_{t},s)$ using a temperature-scaled Softmax formulation as follows:", + "bbox": [ + 88, + 417, + 482, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (d _ {t}, s) = \\frac {\\exp \\left(- d _ {t} / \\tau (s)\\right)}{\\sum_ {i = 1} ^ {m} \\exp \\left(- d _ {i} / \\tau (s)\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 549, + 482, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau(s)$ modulates the task difficulty over time, allowing a smooth shift in the probability distribution toward harder tasks. This temperature function is defined as:", + "bbox": [ + 89, + 593, + 482, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau (s) = \\frac {\\tau_ {0}}{1 + \\lambda \\cdot s / S}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 648, + 482, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\tau_0$ is the initial temperature, $\\lambda$ is the annealing rate, and $S$ is the total number of training steps.", + "bbox": [ + 89, + 689, + 482, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach for defining $d_{t}$ values is based on the inherent complexity of each task, which corresponds to the depth of reasoning involved. Specifically, we assign $d_{1}$ to the most atomic task, involving the prediction of newly introduced tokens. At the other end of the spectrum, $d_{m}$ represents the final task, requiring the full chain-of-thought (CoT) reasoning steps. Between them, intermediate tasks $\\{d_{t}\\}_{t = 1}^{T}$ , serves to bridge the gap between the atomic tasks and the comprehensive chain-of-thought responses.", + "bbox": [ + 88, + 719, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this project, we introduce three types of data subsets for each downstream task, organized in increasing levels of difficulty. The first and most atomic task involves teaching", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the model to generate tokens from the new auxiliary vocabulary. For instance, in depth-related tasks, we train the model to learn depth maps; in segmentation-related tasks, we teach the model to generate masks; and so on.", + "bbox": [ + 511, + 357, + 906, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The other two data subsets involve chain-of-thought (CoT) prompts and direct labeling which help with reasoning with auxiliary tokens. In the CoT subset, we use the new intermediate visual perception tokens to answer downstream-specific questions, encouraging the model to reason step by step. In the direct labeling subset, we pose the same questions but instruct the model to provide direct answers without step-by-step reasoning.", + "bbox": [ + 511, + 421, + 908, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [15], we employ a multitasking approach for these two data subsets. For each image, we sequentially present both the CoT and direct labeling questions, allowing the model to tackle each image with both reasoning styles in sequence. We use a sequential sampler rather than a random sampler, shuffling the images beforehand. This strategy enables the model to learn from both types of reasoning tasks effectively, enhancing its ability to perform complex visual reasoning. More details in Sec. 9.", + "bbox": [ + 511, + 545, + 908, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 705, + 645, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We base our work on LLaVA 1.5 13B as the foundation for our model, which we refer to as LLaVA-AURORA. Our approach augments the MLM with perception tokens to enhance reasoning and improve performance across both 3D and 2D visual tasks. We evaluate our approach on relative depth estimation (3D) using pixel-level depth map tokens for fine-grained depth capture, and on object counting (2D) with mid-level bounding box tokens for precise localization. These tokens not only enhance task-specific results but also highlight our framework's potential to generalize effectively across a broad spectrum of visual reasoning tasks.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3840", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. 3D reasoning task", + "text_level": 1, + "bbox": [ + 89, + 90, + 266, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We choose relative depth estimation as our 3D task because it enables the model to determine spatial relationships within a scene by identifying which points are closer to or farther from the camera. This foundational skill is crucial for scene understanding and applications requiring spatial awareness, such as robotics and autonomous systems. Specifically, this task involves identifying the point closest to the camera among multiple marked points in an image. To support the model's reasoning, we use discrete depth map tokens that capture spatial depth information, enhancing the model's understanding of proximity.", + "bbox": [ + 89, + 112, + 485, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tokenization. To capture fine-grained spatial details, we tokenize depth maps into sequences of discrete tokens. Inspired by the approach in AiT [36], we use a Vector Quantized Variational Autoencoder (VQVAE) [44] with a codebook size of 128. In this setup, each depth map is encoded as a grid of embeddings, with each embedding matched to the nearest entry in the codebook, yielding a compact depth representation. The VQVAE decoder reconstructs the depth map from this sequence of latent codes, and the entire model is optimized with a reconstruction loss to ensure precise encoding. Each $320 \\times 320$ depth map is compressed into a $10 \\times 10$ grid of code indices, resulting in a 100-token sequence where each token corresponds to one of the 128 codebook entries. We introduce 128 unique depth tokens, labeled $DEPTH\\_0$ through $DEPTH\\_127$ , and enclose each sequence with special tokens $DEPTH\\_START$ and $DEPTH\\_END$ , expanding the model's vocabulary by 130 depth-related tokens in total.", + "bbox": [ + 89, + 287, + 483, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training data. We train the VQVAE model on pseudodepth maps generated from the ADE20k dataset [54, 55] using the Depth Anything model [49, 50]. This dataset provides a diverse range of scenes, enhancing the model's ability to generalize.", + "bbox": [ + 89, + 566, + 483, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For fine-tuning, we prepare three types of data tailored for relative depth estimation (as detailed in Sec. 3.4). First, we generate depth maps for $20\\mathrm{k}$ ADE20k images, tokenize them with the pre-trained VQVAE, and format each sample in a Q&A structure, prompting the model with a depth estimation question and an answer sequence of depth tokens. Additionally, we construct a dataset of 500 ADE20k images for chain-of-thought (CoT) reasoning, with 2-5 markers in each image. Here, the prompt guides the model to generate the coordinates of the markers and then the depth map as intermediate steps, then identify the marker closest to the camera. This CoT training improves sequential reasoning and relative depth estimation accuracy. Finally, we use the same 500 images for direct labeling, prompting the model to directly identify and label the marker closest to the camera, see Sec. 9.", + "bbox": [ + 89, + 643, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This fine-tuning setup enables our model to effectively", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "handle step-by-step reasoning for the relative depth estimation task.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. 2D reasoning task", + "text_level": 1, + "bbox": [ + 511, + 132, + 687, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We select counting as a critical 2D visual task. For object counting, we incorporate bounding box predictions as an intermediate reasoning step to improve accuracy in answering counting queries. Given an image and a question about the number of specific objects, the model first identifies and predicts bounding boxes for each instance of the target object. These bounding box tokens serve as structured, intermediate representations, enabling the model to understand spatial arrangements and accurately count object instances.", + "bbox": [ + 511, + 154, + 906, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tokenization. To represent bounding boxes as discrete tokens, we resize all input images to a fixed resolution of 336x336 pixels. This preprocessing step allows us to add 336 unique tokens to the model's vocabulary, each representing a specific pixel position within the resized image. These tokens, labeled Pixel_0 to Pixel_335, enable the model to uniquely reference each pixel location. Bounding boxes are encoded as tuples of four tokens, formatted as (PIXEL_i, Pixel_j, Pixel_k, Pixel_m), where Pixel_i and Pixel_j denote the coordinates of the top-left corner and Pixel_k and Pixel_m represent the bottom-right corner (i.e., $(x_{1},y_{1},x_{2},y_{2})$ ). This discrete representation allows the model to interpret and use bounding box locations effectively, providing the spatial structure needed for accurate object counting.", + "bbox": [ + 511, + 297, + 908, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training data. To fine-tune the model for object counting, we draw on three types of data tailored to this task (as detailed in Sec. 3.4). First, we use task-specific data from the LVIS dataset [13], selecting 5k images with objects whose counts range from 0 to 15. For each selected image, we specify an object type (e.g., \"beds\") and structure the fine-tuning samples to prompt the model for bounding box predictions of the specified objects within the image.", + "bbox": [ + 511, + 534, + 908, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enhance the model's reasoning ability, we include a small subset of 250 LVIS images for chain-of-thought (CoT) training. Here, each question prompt encourages step-by-step reasoning by instructing the model to first generate bounding boxes for the target object, followed by providing the final count. Additionally, we create a direct labeling subset using the same 250 images. In this subset, we prompt the model to directly identify and label the total count of the specified objects without the intermediate step of bounding box generation.", + "bbox": [ + 511, + 655, + 908, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Benchmarks", + "text_level": 1, + "bbox": [ + 511, + 816, + 648, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relative depth. A recent benchmark, BLINK [11], introduces tasks designed to be intuitive for humans yet challenging for multimodal models, with relative depth estimation as one of its tasks. BLINK provides 124 images, each", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3841", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/17ff087b4f382ecfc45ede1a42871bf92c0a735c664934a2f2d63ed0d82109d0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTrainingCV-Bench CountingSEED-Bench CountingBLINK Counting
Direct Labeling DataBounding Box DataCoT Data
LLaVA One VisionXXX34.431.735.8
LLaVA 1.5 13BXXX40.952.235.0
Fine-tuned LlaVA✓XX44.746.30.2
LLaVA-AURORA (Ours)✓✓✓56.054.645.8
GPT-4oXXX70.1864.647.5
GPT-4 TurboXXX61.364.857.5
GPT-4 Turbo + ToolXXX48.629.926.7
", + "bbox": [ + 166, + 88, + 831, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of object counting accuracy (%) across three benchmarks (CV-Bench, SEED-Bench, and BLINK). Our Aurora model, using auxiliary tokens to encode bounding box information for intermediate reasoning, demonstrates superior performance compared to the baseline models and the original base model. The gray rows represent closed-source models, providing additional baselines for comparison.", + "bbox": [ + 88, + 250, + 906, + 295 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "containing two marked points labeled as $A$ and $B$ , and asks which point is closer to the camera. To reduce biases that language models have toward answering multiple-choice questions [39, 53], we modify the original BLINK questions by removing the answer choices. To further evaluate the model's reasoning and 3D understanding in relative depth, we curated a series of more challenging benchmark sets, collectively called HardBLINK. We progressively increase task difficulty by altering the prompts and image configurations as follows:", + "bbox": [ + 88, + 318, + 480, + 468 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Prompt Modification: In the prompts, we exclude the number of markers and their labels, requiring the model to infer these details solely from the image.", + "2. Increased Point Complexity: We generate four variations of BLINK by adding more markers to each image, using Depth Anything to produce pseudo-depth maps for precise placement. These curated sets—HardBLINK 3points, HardBLINK 4points, and HardBLINK 5points—contain the same images as BLINK but with 3, 4, and 5 randomly placed markers, respectively, each with reasonable depth and distance differences. This setup tests the model's depth reasoning across more complex spatial configurations.", + "3. Mitigating Height-Based Bias: To prevent the model from assuming that higher points are farther from the camera, we place markers at mid-height within the image. This approach encourages reliance on depth information rather than positional cues [6]." + ], + "bbox": [ + 89, + 470, + 482, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Counting. For counting, we evaluate the model on CV-Bench [40], SEED-Bench [23], and BLINK's counting [11] subtask. To better capture the model's capabilities, we also remove multiple-choice options, requiring it to generate an exact count.", + "bbox": [ + 89, + 750, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Baselines", + "text_level": 1, + "bbox": [ + 89, + 833, + 199, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate a diverse range of models, including closed-source models like GPT-4o [37] and GPT-4 Turbo [37], as well as state-of-the-art open-source models such as LLaVA", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d0f4997cca7b4f96551273234bc3befc64ac3a0623eedebeff7bae07b89c5c5.jpg", + "image_caption": [ + "Figure 4. Depth maps generated by Aurora are imperfect but resemble the ground-truths from Depth Anything [49]." + ], + "image_footnote": [], + "bbox": [ + 516, + 316, + 905, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OneVision [25]. Another key baseline for our work involves fine-tuning the base model AURORA is applied on, solely on the direct labeling portion of the training data for each task, omitting the newly introduced tokens. This approach allows us to isolate the impact of our token-based enhancements. Additionally, we evaluate the base model for AURORA, LLaVA 1.5 13B, to assess its performance without task-specific adaptations.", + "bbox": [ + 511, + 530, + 906, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also add tool-augmented baselines. For relative depth estimation, we employ GPT-4 Turbo, providing it with the ground truth depth maps generated by Depth Anything for each image, allowing it to use this information in its responses [17]. For counting, we also use GPT-4 Turbo in a tool-augmented setup. In this process, GPT-4 Turbo first identifies the object specified in the question, then uses Grounding DINO [29] to locate the bounding boxes for each instance of the object and finally counts them.", + "bbox": [ + 511, + 652, + 908, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. How do new tokens improve 3D reasoning?", + "text_level": 1, + "bbox": [ + 511, + 801, + 880, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experiments demonstrate that incorporating new visual tokens significantly enhances the model's 3D reasoning abilities, specifically in the relative depth estimation task. Results in Tab. 1 show that our model outperforms both the primary baseline, which is fine-tuned solely on direct label-", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3842", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a502d8ea345d881c5862aaa42cce7338e8cd453c0114c0ba27fa3ae8201b57fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBLINK [11] 2 pointsHardBLINK 3 PointsHardBLINK 4 PointsHardBLINK 5 PointsAverage
VQGAN [10] (16384 codes)82.266.1533759.6
Unified-IO70.275.875.875.874.4
Unified-IO25437.92127.435.1
AURORA (Ours)91.978.271.775.879.2
Our VQVAE (128 codes)96.794.395.296.795.7
", + "bbox": [ + 168, + 88, + 826, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. While not the main aim of our work, we report the depth generation performance across benchmarks with 2, 3, 4, and 5 marked points using BLINK [11]'s relative depth subtask images. We report relative depth estimation accuracy $(\\%)$ , calculated by programmatically extracting depth values at specific coordinates from model-generated depth maps. Our model consistently outperforms other multimodal models, including Unified-IO [31] and Unified-IO 2 [32]. The gray row represents our VQVAE model, which serves as the upper bound for performance.", + "bbox": [ + 89, + 209, + 906, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing data, and the original base model, indicating that the added tokens contribute meaningfully to task accuracy.", + "bbox": [ + 88, + 306, + 480, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As task complexity increases—for instance, when additional markers are introduced in the HardBLINK benchmarks—our model's performance advantage becomes even more pronounced. Not only does it maintain high accuracy in distinguishing depth relationships, but it also surpasses both advanced closed-source models, such as GPT-4o, GPT-4 Turbo, and GPT-4 Turbo + Tool that use ground truth depth maps. This suggests that the new tokens enable our model to develop a more nuanced understanding of depth cues, even in cases where complex spatial reasoning is required.", + "bbox": [ + 88, + 337, + 482, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, these results highlight the effectiveness of our token-based approach in enhancing 3D reasoning, allowing the model to handle increasingly difficult tasks with robust performance and accuracy.", + "bbox": [ + 89, + 503, + 482, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. How do new tokens improve 2D reasoning?", + "text_level": 1, + "bbox": [ + 89, + 573, + 457, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the 2D task of object counting, incorporating new visual tokens provides a significant advantage over the primary baselines. As shown in Tab. 2, our model outperforms both the baseline fine-tuned solely on direct labeling data and the original base model, as well as the state-of-the-art open-source model LLaVA-OneVision and GPT-4 Turbo + Tool, underscoring the value of these tokens in enhancing counting accuracy.", + "bbox": [ + 89, + 595, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our model does not yet surpass advanced closed-source models, the results demonstrate that the new tokens yield a meaningful improvement in 2D reasoning, enabling more reliable and accurate object counting compared to standard fine-tuning approaches and open-source alternatives.", + "bbox": [ + 89, + 717, + 482, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Perception token decoding", + "text_level": 1, + "bbox": [ + 89, + 816, + 331, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our approach enables the decoding of learned perception tokens into specialist features, such as depth maps and object bounding boxes, to assess their fidelity and utility. We, in particular, evaluate the accuracy and correctness", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of the depth maps generated by LLaVA-AURORA, specifically focusing on its ability to represent spatial relationships in visual scenes. For this evaluation, we use the relative depth images from the BLINK benchmark [11], ensuring consistency with our relative depth assessments. LLaVA-AURORA generates depth tokens from BLINK images, which are then reconstructed into full depth maps via the decoder of our pre-trained VQVAE model. Notably, the depth maps output by our pre-trained VQVAE provide an upper bound on the quality of the depth maps generated by LLaVA-AURORA. We use programmatic relative depth accuracy as our metric, which measures how well the reconstructed depth maps capture the relative depth of marked points. This evaluation spans several configurations in the benchmark, including images with 2, 3, 4, and 5 labeled points. By comparing the model's predicted depth with the ground-truth marker coordinates, we calculate relative depth accuracy, allowing us to assess the precision of depth map generation in reflecting spatial depth relationships. As shown in Tab. 3, LLaVA-AURORA outperforms Unified-IO [31, 32] in this task. Furthermore, qualitative analysis in Fig. 4 reveals that LLaVA-AURORA's depth maps capture spatial details effectively highlighting its capacity to interpret and represent fine-grained depth information.", + "bbox": [ + 511, + 306, + 906, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 681, + 633, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our algorithm enables the lightweight and scalable integration of perception tokens, such as depth maps and bounding box coordinates, into MLMs, allowing them to perform intermediate reasoning steps akin to chain-of-thought processes. Our method achieves state-of-the-art results on challenging tasks like 2D object counting and 3D relative depth estimation. It also enhances model generalization and interpretability without relying on external tools or task-specific fine-tuning. The framework is inherently adaptable, incorporating new perception tokens as they emerge, making it a future-proof solution for advancing multimodal reasoning.", + "bbox": [ + 511, + 707, + 906, + 875 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3843", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning, 2022.", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023.", + "[3] Luca Beurer-Kellner, Marc Fischer, and Martin Vechev. Guiding llms the right way: Fast, non-invasive constrained generation. arXiv preprint arXiv:2403.06988, 2024.", + "[4] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024.", + "[5] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021.", + "[6] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016.", + "[7] Aidan Cooper. A guide to structured outputs using constrained decoding, 2024.", + "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023.", + "[9] PHILOSO EPHY DO CT OR OF. MACHINE PERCEPTION OF THREE-DIMENSIONAL, SO LIDS. PhD thesis, MASSACHUSETTS INSTITUTE OF TECHNOLOGY, 1961.", + "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021.", + "[11] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024.", + "[12] Saibo Geng, Martin Josifoski, Maxime Peyrard, and Robert West. Grammar-constrained decoding for structured nlp tasks without finetuning. arXiv preprint arXiv:2305.13971, 2023.", + "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Pro" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019.", + "[14] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, pages 10-5244. CiteSeer, 1988.", + "[15] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. arXiv preprint arXiv:2305.02301, 2023.", + "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.", + "[17] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024.", + "[18] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9601, 2024.", + "[19] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu, Chengru Song, Xiaoqiang Lei, Di Zhang, Wenwu Ou, Kun Gai, and Yadong Mu. Unified language-vision pretraining in llm with dynamic discrete visual tokenization, 2024.", + "[20] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models, 2023.", + "[21] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017.", + "[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model, 2024.", + "[23] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023.", + "[24] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning, 2023.", + "[25] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024.", + "[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023.", + "[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023." + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3844", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024.", + "[29] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023.", + "[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024.", + "[31] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks, 2022.", + "[32] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action, 2023.", + "[33] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models, 2023.", + "[34] David Marr. Vision: A computational investigation into the human representation and processing of visual information. MIT press, 2010.", + "[35] Marvin Minsky and Seymour Papert. An introduction to computational geometry. Cambridge, 1947: 104, 1969.", + "[36] Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Unifying output space of visual tasks via soft token. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19900-19910, 2023.", + "[37] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.", + "[38] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models, 2024.", + "[39] Pouya Pezeshkpour and Estevam Hruschka. Large language models sensitivity to the order of options in multiple-choice questions. arXiv preprint arXiv:2308.11483, 2023.", + "[40] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024.", + "[41] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024.", + "[42] Antonio Torralba and Aude Oliva. Depth estimation from image structure. IEEE Transactions on pattern analysis and machine intelligence, 24(9):1226-1238, 2002." + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Vivien Tran-Thien. Fast, high-fidelity llm decoding with regex constraints, 2024.", + "[44] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017.", + "[45] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework, 2022.", + "[46] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models, 2023.", + "[47] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm, 2024.", + "[48] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models, 2024.", + "[49] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024.", + "[50] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024.", + "[51] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl: Modularization empowers large language models with multimodality, 2024.", + "[52] Haoxuan You, Rui Sun, Zhecan Wang, Long Chen, Gengyu Wang, Hammad A. Ayyubi, Kai-Wei Chang, and Shih-Fu Chang. Idealgpt: Iteratively decomposing vision and language reasoning via large language models, 2023.", + "[53] Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. Large language models are not robust multiple choice selectors. In The Twelfth International Conference on Learning Representations, 2023.", + "[54] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017.", + "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019." + ], + "bbox": [ + 516, + 92, + 906, + 784 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3845", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/a1c25909-0376-4a01-a3cb-0e20c3229024_model.json b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/a1c25909-0376-4a01-a3cb-0e20c3229024_model.json new file mode 100644 index 0000000000000000000000000000000000000000..200bbda8bd9f1498223a22e2af3fa9b35924aefb --- /dev/null +++ b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/a1c25909-0376-4a01-a3cb-0e20c3229024_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.898, + 0.155 + ], + "angle": 0, + "content": "Perception Tokens Enhance Visual Reasoning in Multimodal Language Models" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.18, + 0.81, + 0.234 + ], + "angle": 0, + "content": "Mahtab Bigverdi1 Zelun Luo2 Cheng-Yu Hsieh1 \nEthan Shen1 Dongping Chen1 Linda G. Shapiro1 Ranjay Krishna1 \n1University of Washington, 2Google Research" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.235, + 0.846, + 0.27 + ], + "angle": 0, + "content": "{mahtab, cydhsieh, ethans03, shapiro, ranjay}@cs.washington.edu, alanzluo@gmail.com, cdp0612@uw.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.336, + 0.485, + 0.82 + ], + "angle": 0, + "content": "Multimodal language models (MLMs) still face challenges in fundamental visual perception tasks where specialized models excel. Tasks requiring reasoning about 3D structures benefit from depth estimation, and reasoning about 2D object instances benefits from object detection. Yet, MLMs can not produce intermediate depth or boxes to reason over. Fine-tuning MLMs on relevant data doesn't generalize well and outsourcing computation to specialized vision tools is too comput-intensive and memory-inefficient. To address this, we introduce Perception Tokens, intrinsic image representations designed to assist reasoning tasks where language is insufficient. Perception tokens act as auxiliary reasoning tokens, akin to chain-of-thought prompts in language models. For example, in a depth-related task, an MLM augmented with perception tokens can reason by generating a depth map as tokens, enabling it to solve the problem effectively. We propose AURORA, a training method that augments MLMs with perception tokens for improved reasoning over visual inputs. AURORA leverages a VQVAE to transform intermediate image representations, such as depth maps into a tokenized format and bounding box tokens, which are then used in a multi-task training framework. AURORA achieves notable improvements across counting benchmarks: \\(+10.8\\%\\) on BLINK, \\(+11.3\\%\\) on CVBench, and \\(+8.3\\%\\) on SEED-Bench, outperforming fine-tuning approaches in generalization across datasets. It also improves on relative depth: over \\(+6\\%\\) on BLINK. With perception tokens, AURORA expands the scope of MLMs beyond language-based reasoning, paving the way for more effective visual reasoning capabilities. Code and data will be released at the project page." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.845, + 0.222, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In contrast to the growing emphasis on building multimodal language models (MLMs), computer vision was originally" + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.357, + 0.645, + 0.38 + ], + "angle": 0, + "content": "Which point is closer to the camera?" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.383, + 0.655, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.736, + 0.312, + 0.831, + 0.324 + ], + "angle": 0, + "content": "LLaVA-Aurora" + }, + { + "type": "text", + "bbox": [ + 0.668, + 0.337, + 0.71, + 0.349 + ], + "angle": 0, + "content": "Step 1:" + }, + { + "type": "text", + "bbox": [ + 0.727, + 0.339, + 0.89, + 0.361 + ], + "angle": 0, + "content": "Point A is at (250, 200), B is at (240, 300), and C is at (480, 210)" + }, + { + "type": "text", + "bbox": [ + 0.669, + 0.37, + 0.71, + 0.381 + ], + "angle": 0, + "content": "Step 2:" + }, + { + "type": "text", + "bbox": [ + 0.727, + 0.372, + 0.888, + 0.393 + ], + "angle": 0, + "content": "I need depth estimation to determine which point is closer" + }, + { + "type": "text", + "bbox": [ + 0.669, + 0.403, + 0.71, + 0.413 + ], + "angle": 0, + "content": "Step 3:" + }, + { + "type": "text", + "bbox": [ + 0.726, + 0.403, + 0.884, + 0.434 + ], + "angle": 0, + "content": "The depth map of the input image is: ... " + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.435, + 0.782, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.669, + 0.484, + 0.71, + 0.494 + ], + "angle": 0, + "content": "Step 4:" + }, + { + "type": "text", + "bbox": [ + 0.727, + 0.485, + 0.88, + 0.507 + ], + "angle": 0, + "content": "Point B has higher pixel value on depth map, so it is closer" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.605 + ], + "angle": 0, + "content": "Figure 1. We introduce Perception Tokens, intermediate reasoning tokens that allow MLMs to go beyond using language in reasoning. With it, we develop AURORA, a framework that trains multimodal language models to leverage visual perception tokens, allowing them to use depth estimation and bounding box predictions while reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.866 + ], + "angle": 0, + "content": "attempting to interpret images as projections of indescribable 3D intrinsics, not just processing 2D arrays of language patterns [9, 34, 35]. Towards this endeavor, early vision research developed a series of intermediate image representations—enabling geometric reasoning through depth estimation [42] and instance reasoning through bounding box grounding [14]. As pointed out by recent work, we have focused less on such perceptual representations and instead tackled reasoning problems that require limited visual involvement [11, 40, 41]. This is likely because many traditional vision tasks remain ambiguous through natural language. Consider the task of identifying which of a set of N points is furthest away from the camera. While language doesn't lend itself to reason over this problem, a depth estimation would provide the appropriate abstraction to reason over." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Numerous attempts have been made to enable MLMs to reason over intrinsic image representations. The default ap-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3836" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.255, + 0.908, + 0.285 + ], + "angle": 0, + "content": "Figure 2. We demonstrate relative depth estimation and counting questions where LLaVA fails. In contrast, by learning to utilize visual perception tokens as intermediate reasoning steps, AURORA successfully complete these tasks requiring perceptual understanding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.299, + 0.482, + 0.45 + ], + "angle": 0, + "content": "proach is to fine-tune MLMs on data tailored to the specific perception task of interest, hoping that the model implicitly learns the required intrinsic representations [4]. Another option is to outsource the computation to external tools: the MLM can invoke a depth estimator or object detector to produce the appropriate intrinsic [17]. Unfortunately, relying on external models makes the task more computationally expensive and requires loading additional models with more memory. Similarly, vanilla fine-tuning (even with advancements like LoRA [16]) has shown marginal improvements." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.454, + 0.483, + 0.651 + ], + "angle": 0, + "content": "Conceptually, we introduce Perception Tokens, intrinsic image representations that aid in reasoning where language is insufficient. To solve the aforementioned task, an MLM augmented with perception tokens can solve the task similar to how language models use chain-of-thought. They will produce a response like the following: \"The depth map is . Therefore, point \\( D \\) is closest to the camera.\" Here, is a set of tokens that implicitly estimate the depth of the image. Similarly, for a counting task, the model can first generate perception tokens that represents the location of the relevant bounding boxes of the desired object, and count the number of boxes to support its final answer." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.655, + 0.483, + 0.838 + ], + "angle": 0, + "content": "To demonstrate the utility of perception tokens, we introduce AURORA, a training algorithm to augment MLMs with the ability to use perception tokens as intermediate reasoning steps. For certain intermediate representations (e.g. depth maps), we train a VQVAE to transform them into a set of tokens, treating the learned VQVAE codebook indices as a collection of perception tokens. while for others, such as bounding boxes, we use directly encoded structured tokens. Next, we follow a multi-task training approach [15, 18] to train MLMs to use perception tokens as chain-of-thought tokens (see Fig. 1). Additionally, we adopt a curriculum learning approach to avoid catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We apply the AURORA training algorithm to the LLaVA [27] model, resulting in our LLaVA-AURORA variant. Our LLaVA-AURORA model significantly outperforms standard fine-tuning approaches across multiple perception" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.908, + 0.481 + ], + "angle": 0, + "content": "demanding tasks, demonstrating the generality and effectiveness of our method. LLaVA-AURORA achieves state-of-the-art results on both relative depth estimation and object counting tasks. For instance, on BLINK relative depth estimation, LLaVA-AURORA delivers a performance boost of \\(6.4\\%\\) compared to the fine-tuning baseline. Similarly, on counting tasks, LLaVA-AURORA drives improvements of \\(10.8\\%\\) on BLINK, \\(11.3\\%\\) on CVBench, and \\(8.3\\%\\) on SEED-Bench. Fig. 2 illustrates examples from these tasks. Perception tokens open up a whole new modality through which MLMs can begin to reason, tackling tasks beyond just language reasoning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.502, + 0.652, + 0.518 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.773 + ], + "angle": 0, + "content": "Multimodal language models (MLMs). MLMs aim to solve a variety of tasks (e.g., Visual Question Answering (VQA) and captioning) based on vision and language inputs. Most modern architectures accomplish this by relying on either cross-attention [1, 2] or visual instruction tuning [26-28, 51] to interleave multimodal information. Cross-attention architectures operate by independently encoding images and then cross-attending to a language model backbone. On the other hand, visual instruction tuning produces token embeddings from image representations that can be interleaved with language tokens to ground generation. While these techniques work well for high level tasks, MLMs still struggle with mid-level and low-level tasks such as counting, depth reasoning, and segmentation. Most MLMs can be classified as either end-to-end MLMs or tool-using MLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "End-to-end MLMs. End-to-end MLMs use a single, unified architecture [1, 8, 24, 26, 27] that can be repeatedly used for different tasks without requiring any architectural changes. While training end-to-end MLMs can be costly due to the need for large amounts of multi-task data, the use of diverse datasets allows end-to-end MLMs to generalize effectively and learn nuanced visual representations." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3837" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.228 + ], + "angle": 0, + "content": "Tool-using MLMs. Tool-using MLMs enable LLMs to perform vision tasks by attaching specialized vision modules such as segmentation or captioning networks [30, 33, 46, 52]. Tool-using MLMs use a router to select the optimal vision module to use for a given input. The use of specialized modules allows MLMs to achieve higher performances on many tasks. However, tool-using MLMs are sensitive to errors because they link together several networks, which are each a potential point of failure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.483, + 0.418 + ], + "angle": 0, + "content": "Any-to-any networks. To address this, recent works [19, 20, 38, 45] such as Unified-IO [31, 32] have experimented with shared embedding spaces and visual decoders for vision and language tasks, training MLMs to generate segmentation masks, keypoints, and depth maps. Similarly, [47] generates special task tokens to route language model representations to diffusion heads for image, video, and audio generation. To handle complex segmentation scenarios, LISA [22] and GSVA [48] train language models to generate an additional segmentation token that can be used to produce segmentation masks, grounding the mask in language reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.482, + 0.525 + ], + "angle": 0, + "content": "While these approaches can generate visual outputs, they cannot reason over their own generations to solve related visual perception tasks. In contrast, augmenting MLMs with perception tokens as chain-of-thought tokens enables them to perform visual reasoning directly, providing significant gains in detail-oriented question-answering tasks such as depth estimation and counting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.54, + 0.357, + 0.557 + ], + "angle": 0, + "content": "3. Perception Tokens & Aurora" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.656 + ], + "angle": 0, + "content": "We introduce Perception Tokens and our Aurora training algorithm (see Fig. 3), which augment multimodal language models with perception tokens, enabling the MLM to leverage these tokens effectively during training and incorporate them into its chain-of-thought reasoning process for enhanced visual reasoning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.289, + 0.682 + ], + "angle": 0, + "content": "3.1. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.81 + ], + "angle": 0, + "content": "In autoregressive large language models, chain-of-thought (CoT) reasoning can be formulated as a multi-step inferential process in which the model iteratively generates intermediate reasoning steps to arrive at a final answer. Given a task input \\( x \\), the model generates a response \\( y \\) conditioned on the input and a sequence of \\( m \\) intermediate reasoning steps \\( \\{s_i\\}_{i=1}^m \\), where \\( x, y \\), and each \\( s_i \\) are sequences of tokens from the model's vocabulary \\( V \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Existing multimodal language models often rely on limited vocabulary tokens derived from text or pre-trained image embeddings like CLIP, restricting their capacity to interpret other representations crucial for reasoning. Mid-level and low-level visual features such as depth maps, instance segmentation masks, human poses and bounding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "boxes which could substantially improve visual reasoning, are currently incompatible with the model and cannot be integrated during training or inference. Our key insight is to introduce auxiliary perception tokens for these intermediate steps with an expanded vocabulary \\( V' = V \\cup V_{\\mathrm{aux}} \\), bridging this gap by allowing the model to integrate richer visual representations into its reasoning process. Conditioning the final output on these tokens enhances the model's accuracy and interpretability across multimodal tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.238, + 0.875, + 0.254 + ], + "angle": 0, + "content": "3.2. Perception token prediction and reasoning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.395 + ], + "angle": 0, + "content": "Introducing an expanded vocabulary to enhance multimodal reasoning presents two main challenges. The first challenge is enabling the model to generate tokens from the new auxiliary vocabulary \\( V_{\\mathrm{aux}} \\), which includes specialized tokens for low- and mid-level visual features, such as depth maps and bounding boxes. The second challenge is ensuring that the model can effectively condition on these auxiliary tokens to improve reasoning, particularly for multi-step inference tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.905, + 0.552 + ], + "angle": 0, + "content": "Perception token prediction. To address the first challenge, we employ a specialist-to-generalist distillation approach, using pre-trained specialist models (e.g., depth estimation or instance segmentation) to guide auxiliary token generation through cross-entropy loss. For each input \\( x \\), the specialist model provides a target probability distribution \\( q_{i} \\) over its tokens. Let \\( M: V_{\\mathrm{spec}} \\to V_{\\mathrm{aux}} \\) denote a one-to-one mapping from the specialist model's vocabulary \\( V_{\\mathrm{spec}} \\) to the auxiliary token vocabulary \\( V_{\\mathrm{aux}} \\). We define the distillation loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.552, + 0.905, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\ell_ {d i s t} = \\min _ {M} \\left(- \\sum_ {i} q _ {i} \\log p _ {M (i)}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.905, + 0.679 + ], + "angle": 0, + "content": "where \\( p_{M(i)} \\) is the probability assigned by our model to the auxiliary token corresponding to the mapped token \\( M(i) \\). This consistent mapping allows the model to effectively align its predictions with the specialist model's output distribution, enhancing the relevance and accuracy of its auxiliary token predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.906, + 0.83 + ], + "angle": 0, + "content": "In addition to distillation, we incorporate a reconstruction loss to enhance the token prediction and the interpretability of our model. Each set of auxiliary tokens corresponds to a specific representation, such as a depth map or a bounding box vector, and is trained to directly predict this feature. To achieve this, we introduced a lightweight decoder \\( g \\) that maps the tokens into the feature space, allowing for efficient and interpretable transformations. Formally, for a token set \\( T \\subset V_{\\mathrm{aux}} \\), decoder \\( g \\), and its target feature \\( f \\), the reconstruction loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.841, + 0.905, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\ell_ {r e c} = \\left\\| g (T) - f \\right\\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\( g(T) \\) is the decoded representation for the token set \\( T \\) in the feature space. This reconstruction process not only" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3838" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.09, + 0.844, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.317, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 3. The overall AURORA training framework. We first learn visual perception tokens using VQVAE. We then finetune MLMs with a multi-task training approach where we distill intrinsic image representations (e.g., depth map) into MLMs by training them to decode the visual tokens as intermediate reasoning steps towards completing the tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.482, + 0.477 + ], + "angle": 0, + "content": "aligns each token with a meaningful feature, improving interpretability, but also reinforces the accuracy of auxiliary token predictions through direct feature supervision. In practice, while the reconstruction objective improves performance and interpretability, it is optional, as models can be effectively trained with the distillation objective alone to reduce computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.791 + ], + "angle": 0, + "content": "Reasoning with perception tokens. The second challenge is enabling the model to condition on tokens from \\( V_{\\mathrm{aux}} \\) effectively when generating each subsequent reasoning step, thereby enhancing its reasoning capabilities. To achieve this, we progressively introduce chain-of-thought reasoning during training, starting with simpler, single-step reasoning examples and gradually incorporating more complex, multi-step reasoning data. The model begins by learning single-step reasoning, predicting an initial reasoning step \\( s_1 \\) based on the input \\( x \\). It then progresses to multistep reasoning, predicting sequences \\( s_1, \\dots, s_m \\) and effectively utilizing auxiliary tokens to support extended chains of thought. To facilitate this, we propose a multi-task data synthesis approach combined with curriculum learning across a variety of synthetic tasks. Additional implementation details can be found in Section 3.4. Finally, we reinforce this process at inference time using constrained decoding, wherein sampling is selectively restricted to auxiliary tokens at certain reasoning steps, guiding the model to generate coherent intermediate steps from \\( V_{\\mathrm{aux}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.227, + 0.817 + ], + "angle": 0, + "content": "3.3. Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "A unified tokenization space is crucial for multimodal models as it creates a consistent framework through which varied visual tasks can be represented, processed, and interpreted. Inspired by [36], we establish a unified tokenization space which enables the model to learn varied visual fea" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.462 + ], + "angle": 0, + "content": "tures within a shared representation seamlessly. In our experiments, we implement two tailored tokenization schemes for commonly used visual representations. Importantly, our framework is designed with flexibility, allowing it to generalize to a broad range of visual representations beyond those presented here." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.909, + 0.59 + ], + "angle": 0, + "content": "Pixel-level representation. This tokenization scheme captures fine-grained spatial information, such as depth maps and segmentation masks, providing the model with detailed pixel-level data essential for accurate visual processing. For these types of tokens, we leverage visual tokenizers like VQVAE and VQGAN, which take in the ground truth masks or depth maps and return discrete target tokens [10, 31, 32, 44]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Structured representation. This scheme encodes structured yet abstracted visual features, such as human poses, bounding boxes, and coordinates, allowing the model to reason with higher-level spatial relationships and object hierarchies. For these tokens, we define the domain of the tokens based on specific properties; for example, the domain for coordinates can range from 0 to the maximum number of pixels in the image's height or width [5]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.885, + 0.744 + ], + "angle": 0, + "content": "3.4. Curriculum learning with progressive CoTs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The objective of training the model is to develop a data and computationally efficient method for learning to predict novel, fine-grained visual tokens and using them to complete complex visual reasoning tasks. We observed that the standard approach, which relies on a fixed mixture data, encounters a trade-off between the accuracy of novel tokens predictions and the model's reasoning capability, primarily due to catastrophic forgetting and challenges in reasoning with new tokens. Conversely, fine-tuning the model with the original training mixture significantly raises computational" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3839" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.889, + 0.266 + ], + "angle": 0, + "content": "
ModelTrainingBLINK [11]2 PointsHardBLINK3 PointsHardBLINK4 PointsHardBLINK5 PointsAverage
DirectLa-labelingDataDepthGener-ationDataCoTData
LLaVA OneVision 72BXXX67.751.650.044.353.4
LLaVA 1.5 13BXXX54.035.537.92939.1
Fine-tuned LLaVA✓XX68.558.952.441.155.2
LLaVA-AURORA (Ours)✓✓✓64.566.960.554.861.6
GPT-4oXXX53.258.95036.349.6
GPT-4 TurboXXX58.154.841.932.246.7
GPT-4 Turbo + ToolXXX70.257.244.326.649.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.275, + 0.908, + 0.334 + ], + "angle": 0, + "content": "Table 1. Performance comparison between our fine-tuned model, the baseline (without depth tokens), and the original base model on the relative depth accuracy (\\%) task. Results demonstrate that our approach, utilizing depth tokens and intermediate reasoning steps, significantly outperforms both the baseline and the base model, particularly on more challenging configurations with 3, 4, and 5 points sampled from the image's mid-height region. The gray rows represent closed-source models, providing additional baselines for comparison." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.358, + 0.482, + 0.417 + ], + "angle": 0, + "content": "costs and may be impractical if the original data, particularly proprietary datasets, is unavailable, making this approach less scalable for incorporating new tokens in the future." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.419, + 0.483, + 0.54 + ], + "angle": 0, + "content": "We propose a curriculum learning-inspired training scheme that begins with atomic tasks and gradually advances to more complex ones requiring sophisticated, multi-hop reasoning. Let \\( d_{t} \\) represent the difficulty of task \\( t \\), with difficulties \\( d_{1} < d_{2} < \\dots d_{T} \\) across \\( T \\) tasks, and let \\( p(t,s) \\) denote the probability of sampling data points from task \\( t \\) at training step \\( s \\). We define \\( p(d_{t},s) \\) using a temperature-scaled Softmax formulation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.55, + 0.483, + 0.587 + ], + "angle": 0, + "content": "\\[\np (d _ {t}, s) = \\frac {\\exp \\left(- d _ {t} / \\tau (s)\\right)}{\\sum_ {i = 1} ^ {m} \\exp \\left(- d _ {i} / \\tau (s)\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.641 + ], + "angle": 0, + "content": "where \\(\\tau(s)\\) modulates the task difficulty over time, allowing a smooth shift in the probability distribution toward harder tasks. This temperature function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.65, + 0.483, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\tau (s) = \\frac {\\tau_ {0}}{1 + \\lambda \\cdot s / S}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Here, \\(\\tau_0\\) is the initial temperature, \\(\\lambda\\) is the annealing rate, and \\(S\\) is the total number of training steps." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Our approach for defining \\( d_{t} \\) values is based on the inherent complexity of each task, which corresponds to the depth of reasoning involved. Specifically, we assign \\( d_{1} \\) to the most atomic task, involving the prediction of newly introduced tokens. At the other end of the spectrum, \\( d_{m} \\) represents the final task, requiring the full chain-of-thought (CoT) reasoning steps. Between them, intermediate tasks \\( \\{d_{t}\\}_{t = 1}^{T} \\), serves to bridge the gap between the atomic tasks and the comprehensive chain-of-thought responses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this project, we introduce three types of data subsets for each downstream task, organized in increasing levels of difficulty. The first and most atomic task involves teaching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.42 + ], + "angle": 0, + "content": "the model to generate tokens from the new auxiliary vocabulary. For instance, in depth-related tasks, we train the model to learn depth maps; in segmentation-related tasks, we teach the model to generate masks; and so on." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.909, + 0.544 + ], + "angle": 0, + "content": "The other two data subsets involve chain-of-thought (CoT) prompts and direct labeling which help with reasoning with auxiliary tokens. In the CoT subset, we use the new intermediate visual perception tokens to answer downstream-specific questions, encouraging the model to reason step by step. In the direct labeling subset, we pose the same questions but instruct the model to provide direct answers without step-by-step reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.909, + 0.683 + ], + "angle": 0, + "content": "Inspired by [15], we employ a multitasking approach for these two data subsets. For each image, we sequentially present both the CoT and direct labeling questions, allowing the model to tackle each image with both reasoning styles in sequence. We use a sequential sampler rather than a random sampler, shuffling the images beforehand. This strategy enables the model to learn from both types of reasoning tasks effectively, enhancing its ability to perform complex visual reasoning. More details in Sec. 9." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.707, + 0.646, + 0.723 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We base our work on LLaVA 1.5 13B as the foundation for our model, which we refer to as LLaVA-AURORA. Our approach augments the MLM with perception tokens to enhance reasoning and improve performance across both 3D and 2D visual tasks. We evaluate our approach on relative depth estimation (3D) using pixel-level depth map tokens for fine-grained depth capture, and on object counting (2D) with mid-level bounding box tokens for precise localization. These tokens not only enhance task-specific results but also highlight our framework's potential to generalize effectively across a broad spectrum of visual reasoning tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3840" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.267, + 0.108 + ], + "angle": 0, + "content": "4.1. 3D reasoning task" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.281 + ], + "angle": 0, + "content": "We choose relative depth estimation as our 3D task because it enables the model to determine spatial relationships within a scene by identifying which points are closer to or farther from the camera. This foundational skill is crucial for scene understanding and applications requiring spatial awareness, such as robotics and autonomous systems. Specifically, this task involves identifying the point closest to the camera among multiple marked points in an image. To support the model's reasoning, we use discrete depth map tokens that capture spatial depth information, enhancing the model's understanding of proximity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.485, + 0.56 + ], + "angle": 0, + "content": "Tokenization. To capture fine-grained spatial details, we tokenize depth maps into sequences of discrete tokens. Inspired by the approach in AiT [36], we use a Vector Quantized Variational Autoencoder (VQVAE) [44] with a codebook size of 128. In this setup, each depth map is encoded as a grid of embeddings, with each embedding matched to the nearest entry in the codebook, yielding a compact depth representation. The VQVAE decoder reconstructs the depth map from this sequence of latent codes, and the entire model is optimized with a reconstruction loss to ensure precise encoding. Each \\(320 \\times 320\\) depth map is compressed into a \\(10 \\times 10\\) grid of code indices, resulting in a 100-token sequence where each token corresponds to one of the 128 codebook entries. We introduce 128 unique depth tokens, labeled \\(DEPTH\\_0\\) through \\(DEPTH\\_127\\), and enclose each sequence with special tokens \\(DEPTH\\_START\\) and \\(DEPTH\\_END\\), expanding the model's vocabulary by 130 depth-related tokens in total." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.644 + ], + "angle": 0, + "content": "Training data. We train the VQVAE model on pseudodepth maps generated from the ADE20k dataset [54, 55] using the Depth Anything model [49, 50]. This dataset provides a diverse range of scenes, enhancing the model's ability to generalize." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.884 + ], + "angle": 0, + "content": "For fine-tuning, we prepare three types of data tailored for relative depth estimation (as detailed in Sec. 3.4). First, we generate depth maps for \\(20\\mathrm{k}\\) ADE20k images, tokenize them with the pre-trained VQVAE, and format each sample in a Q&A structure, prompting the model with a depth estimation question and an answer sequence of depth tokens. Additionally, we construct a dataset of 500 ADE20k images for chain-of-thought (CoT) reasoning, with 2-5 markers in each image. Here, the prompt guides the model to generate the coordinates of the markers and then the depth map as intermediate steps, then identify the marker closest to the camera. This CoT training improves sequential reasoning and relative depth estimation accuracy. Finally, we use the same 500 images for direct labeling, prompting the model to directly identify and label the marker closest to the camera, see Sec. 9." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "This fine-tuning setup enables our model to effectively" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "handle step-by-step reasoning for the relative depth estimation task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.133, + 0.689, + 0.149 + ], + "angle": 0, + "content": "4.2. 2D reasoning task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.907, + 0.293 + ], + "angle": 0, + "content": "We select counting as a critical 2D visual task. For object counting, we incorporate bounding box predictions as an intermediate reasoning step to improve accuracy in answering counting queries. Given an image and a question about the number of specific objects, the model first identifies and predicts bounding boxes for each instance of the target object. These bounding box tokens serve as structured, intermediate representations, enabling the model to understand spatial arrangements and accurately count object instances." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.298, + 0.909, + 0.526 + ], + "angle": 0, + "content": "Tokenization. To represent bounding boxes as discrete tokens, we resize all input images to a fixed resolution of 336x336 pixels. This preprocessing step allows us to add 336 unique tokens to the model's vocabulary, each representing a specific pixel position within the resized image. These tokens, labeled Pixel_0 to Pixel_335, enable the model to uniquely reference each pixel location. Bounding boxes are encoded as tuples of four tokens, formatted as (PIXEL_i, Pixel_j, Pixel_k, Pixel_m), where Pixel_i and Pixel_j denote the coordinates of the top-left corner and Pixel_k and Pixel_m represent the bottom-right corner (i.e., \\((x_{1},y_{1},x_{2},y_{2})\\)). This discrete representation allows the model to interpret and use bounding box locations effectively, providing the spatial structure needed for accurate object counting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.909, + 0.656 + ], + "angle": 0, + "content": "Training data. To fine-tune the model for object counting, we draw on three types of data tailored to this task (as detailed in Sec. 3.4). First, we use task-specific data from the LVIS dataset [13], selecting 5k images with objects whose counts range from 0 to 15. For each selected image, we specify an object type (e.g., \"beds\") and structure the fine-tuning samples to prompt the model for bounding box predictions of the specified objects within the image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.909, + 0.808 + ], + "angle": 0, + "content": "To enhance the model's reasoning ability, we include a small subset of 250 LVIS images for chain-of-thought (CoT) training. Here, each question prompt encourages step-by-step reasoning by instructing the model to first generate bounding boxes for the target object, followed by providing the final count. Additionally, we create a direct labeling subset using the same 250 images. In this subset, we prompt the model to directly identify and label the total count of the specified objects without the intermediate step of bounding box generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.649, + 0.833 + ], + "angle": 0, + "content": "4.3. Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Relative depth. A recent benchmark, BLINK [11], introduces tasks designed to be intuitive for humans yet challenging for multimodal models, with relative depth estimation as one of its tasks. BLINK provides 124 images, each" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.958 + ], + "angle": 0, + "content": "3841" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.167, + 0.089, + 0.833, + 0.242 + ], + "angle": 0, + "content": "
ModelTrainingCV-Bench CountingSEED-Bench CountingBLINK Counting
Direct Labeling DataBounding Box DataCoT Data
LLaVA One VisionXXX34.431.735.8
LLaVA 1.5 13BXXX40.952.235.0
Fine-tuned LlaVA✓XX44.746.30.2
LLaVA-AURORA (Ours)✓✓✓56.054.645.8
GPT-4oXXX70.1864.647.5
GPT-4 TurboXXX61.364.857.5
GPT-4 Turbo + ToolXXX48.629.926.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.251, + 0.908, + 0.296 + ], + "angle": 0, + "content": "Table 2. Comparison of object counting accuracy (%) across three benchmarks (CV-Bench, SEED-Bench, and BLINK). Our Aurora model, using auxiliary tokens to encode bounding box information for intermediate reasoning, demonstrates superior performance compared to the baseline models and the original base model. The gray rows represent closed-source models, providing additional baselines for comparison." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.482, + 0.469 + ], + "angle": 0, + "content": "containing two marked points labeled as \\( A \\) and \\( B \\), and asks which point is closer to the camera. To reduce biases that language models have toward answering multiple-choice questions [39, 53], we modify the original BLINK questions by removing the answer choices. To further evaluate the model's reasoning and 3D understanding in relative depth, we curated a series of more challenging benchmark sets, collectively called HardBLINK. We progressively increase task difficulty by altering the prompts and image configurations as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.471, + 0.483, + 0.516 + ], + "angle": 0, + "content": "1. Prompt Modification: In the prompts, we exclude the number of markers and their labels, requiring the model to infer these details solely from the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.483, + 0.667 + ], + "angle": 0, + "content": "2. Increased Point Complexity: We generate four variations of BLINK by adding more markers to each image, using Depth Anything to produce pseudo-depth maps for precise placement. These curated sets—HardBLINK 3points, HardBLINK 4points, and HardBLINK 5points—contain the same images as BLINK but with 3, 4, and 5 randomly placed markers, respectively, each with reasonable depth and distance differences. This setup tests the model's depth reasoning across more complex spatial configurations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.743 + ], + "angle": 0, + "content": "3. Mitigating Height-Based Bias: To prevent the model from assuming that higher points are farther from the camera, we place markers at mid-height within the image. This approach encourages reliance on depth information rather than positional cues [6]." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.471, + 0.483, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Counting. For counting, we evaluate the model on CV-Bench [40], SEED-Bench [23], and BLINK's counting [11] subtask. To better capture the model's capabilities, we also remove multiple-choice options, requiring it to generate an exact count." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.2, + 0.849 + ], + "angle": 0, + "content": "4.4. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate a diverse range of models, including closed-source models like GPT-4o [37] and GPT-4 Turbo [37], as well as state-of-the-art open-source models such as LLaVA" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.317, + 0.906, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.474, + 0.907, + 0.503 + ], + "angle": 0, + "content": "Figure 4. Depth maps generated by Aurora are imperfect but resemble the ground-truths from Depth Anything [49]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.907, + 0.652 + ], + "angle": 0, + "content": "OneVision [25]. Another key baseline for our work involves fine-tuning the base model AURORA is applied on, solely on the direct labeling portion of the training data for each task, omitting the newly introduced tokens. This approach allows us to isolate the impact of our token-based enhancements. Additionally, we evaluate the base model for AURORA, LLaVA 1.5 13B, to assess its performance without task-specific adaptations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.909, + 0.79 + ], + "angle": 0, + "content": "We also add tool-augmented baselines. For relative depth estimation, we employ GPT-4 Turbo, providing it with the ground truth depth maps generated by Depth Anything for each image, allowing it to use this information in its responses [17]. For counting, we also use GPT-4 Turbo in a tool-augmented setup. In this process, GPT-4 Turbo first identifies the object specified in the question, then uses Grounding DINO [29] to locate the bounding boxes for each instance of the object and finally counts them." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.881, + 0.818 + ], + "angle": 0, + "content": "4.5. How do new tokens improve 3D reasoning?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Our experiments demonstrate that incorporating new visual tokens significantly enhances the model's 3D reasoning abilities, specifically in the relative depth estimation task. Results in Tab. 1 show that our model outperforms both the primary baseline, which is fine-tuned solely on direct label-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3842" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.169, + 0.089, + 0.827, + 0.201 + ], + "angle": 0, + "content": "
ModelBLINK [11] 2 pointsHardBLINK 3 PointsHardBLINK 4 PointsHardBLINK 5 PointsAverage
VQGAN [10] (16384 codes)82.266.1533759.6
Unified-IO70.275.875.875.874.4
Unified-IO25437.92127.435.1
AURORA (Ours)91.978.271.775.879.2
Our VQVAE (128 codes)96.794.395.296.795.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.21, + 0.907, + 0.281 + ], + "angle": 0, + "content": "Table 3. While not the main aim of our work, we report the depth generation performance across benchmarks with 2, 3, 4, and 5 marked points using BLINK [11]'s relative depth subtask images. We report relative depth estimation accuracy \\((\\%)\\), calculated by programmatically extracting depth values at specific coordinates from model-generated depth maps. Our model consistently outperforms other multimodal models, including Unified-IO [31] and Unified-IO 2 [32]. The gray row represents our VQVAE model, which serves as the upper bound for performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.307, + 0.482, + 0.337 + ], + "angle": 0, + "content": "ing data, and the original base model, indicating that the added tokens contribute meaningfully to task accuracy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.483, + 0.502 + ], + "angle": 0, + "content": "As task complexity increases—for instance, when additional markers are introduced in the HardBLINK benchmarks—our model's performance advantage becomes even more pronounced. Not only does it maintain high accuracy in distinguishing depth relationships, but it also surpasses both advanced closed-source models, such as GPT-4o, GPT-4 Turbo, and GPT-4 Turbo + Tool that use ground truth depth maps. This suggests that the new tokens enable our model to develop a more nuanced understanding of depth cues, even in cases where complex spatial reasoning is required." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.483, + 0.564 + ], + "angle": 0, + "content": "Overall, these results highlight the effectiveness of our token-based approach in enhancing 3D reasoning, allowing the model to handle increasingly difficult tasks with robust performance and accuracy." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.574, + 0.458, + 0.591 + ], + "angle": 0, + "content": "4.6. How do new tokens improve 2D reasoning?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.717 + ], + "angle": 0, + "content": "In the 2D task of object counting, incorporating new visual tokens provides a significant advantage over the primary baselines. As shown in Tab. 2, our model outperforms both the baseline fine-tuned solely on direct labeling data and the original base model, as well as the state-of-the-art open-source model LLaVA-OneVision and GPT-4 Turbo + Tool, underscoring the value of these tokens in enhancing counting accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Although our model does not yet surpass advanced closed-source models, the results demonstrate that the new tokens yield a meaningful improvement in 2D reasoning, enabling more reliable and accurate object counting compared to standard fine-tuning approaches and open-source alternatives." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.332, + 0.835 + ], + "angle": 0, + "content": "4.7. Perception token decoding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our approach enables the decoding of learned perception tokens into specialist features, such as depth maps and object bounding boxes, to assess their fidelity and utility. We, in particular, evaluate the accuracy and correctness" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.908, + 0.67 + ], + "angle": 0, + "content": "of the depth maps generated by LLaVA-AURORA, specifically focusing on its ability to represent spatial relationships in visual scenes. For this evaluation, we use the relative depth images from the BLINK benchmark [11], ensuring consistency with our relative depth assessments. LLaVA-AURORA generates depth tokens from BLINK images, which are then reconstructed into full depth maps via the decoder of our pre-trained VQVAE model. Notably, the depth maps output by our pre-trained VQVAE provide an upper bound on the quality of the depth maps generated by LLaVA-AURORA. We use programmatic relative depth accuracy as our metric, which measures how well the reconstructed depth maps capture the relative depth of marked points. This evaluation spans several configurations in the benchmark, including images with 2, 3, 4, and 5 labeled points. By comparing the model's predicted depth with the ground-truth marker coordinates, we calculate relative depth accuracy, allowing us to assess the precision of depth map generation in reflecting spatial depth relationships. As shown in Tab. 3, LLaVA-AURORA outperforms Unified-IO [31, 32] in this task. Furthermore, qualitative analysis in Fig. 4 reveals that LLaVA-AURORA's depth maps capture spatial details effectively highlighting its capacity to interpret and represent fine-grained depth information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.634, + 0.699 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.907, + 0.875 + ], + "angle": 0, + "content": "Our algorithm enables the lightweight and scalable integration of perception tokens, such as depth maps and bounding box coordinates, into MLMs, allowing them to perform intermediate reasoning steps akin to chain-of-thought processes. Our method achieves state-of-the-art results on challenging tasks like 2D object counting and 3D relative depth estimation. It also enhances model generalization and interpretability without relying on external tools or task-specific fine-tuning. The framework is inherently adaptable, incorporating new perception tokens as they emerge, making it a future-proof solution for advancing multimodal reasoning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3843" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.252 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.256, + 0.482, + 0.311 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.314, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[3] Luca Beurer-Kellner, Marc Fischer, and Martin Vechev. Guiding llms the right way: Fast, non-invasive constrained generation. arXiv preprint arXiv:2403.06988, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.357, + 0.482, + 0.438 + ], + "angle": 0, + "content": "[4] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[5] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.482, + 0.527 + ], + "angle": 0, + "content": "[6] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.482, + 0.556 + ], + "angle": 0, + "content": "[7] Aidan Cooper. A guide to structured outputs using constrained decoding, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[9] PHILOSO EPHY DO CT OR OF. MACHINE PERCEPTION OF THREE-DIMENSIONAL, SO LIDS. PhD thesis, MASSACHUSETTS INSTITUTE OF TECHNOLOGY, 1961." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[11] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[12] Saibo Geng, Martin Josifoski, Maxime Peyrard, and Robert West. Grammar-constrained decoding for structured nlp tasks without finetuning. arXiv preprint arXiv:2305.13971, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[14] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, pages 10-5244. CiteSeer, 1988." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[15] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. arXiv preprint arXiv:2305.02301, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.303, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[17] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[18] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9601, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[19] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu, Chengru Song, Xiaoqiang Lei, Di Zhang, Wenwu Ou, Kun Gai, and Yadong Mu. Unified language-vision pretraining in llm with dynamic discrete visual tokenization, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.525, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[20] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.634 + ], + "angle": 0, + "content": "[21] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[23] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.734, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[24] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[25] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3844" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[29] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[31] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[32] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[33] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[34] David Marr. Vision: A computational investigation into the human representation and processing of visual information. MIT press, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[35] Marvin Minsky and Seymour Papert. An introduction to computational geometry. Cambridge, 1947: 104, 1969." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[36] Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Unifying output space of visual tasks via soft token. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19900-19910, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[37] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[38] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[39] Pouya Pezeshkpour and Estevam Hruschka. Large language models sensitivity to the order of options in multiple-choice questions. arXiv preprint arXiv:2308.11483, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[40] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[41] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[42] Antonio Torralba and Aude Oliva. Depth estimation from image structure. IEEE Transactions on pattern analysis and machine intelligence, 24(9):1226-1238, 2002." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[43] Vivien Tran-Thien. Fast, high-fidelity llm decoding with regex constraints, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[44] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[45] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[46] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[47] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[48] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.907, + 0.418 + ], + "angle": 0, + "content": "[49] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[50] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[51] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl: Modularization empowers large language models with multimodality, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[52] Haoxuan You, Rui Sun, Zhecan Wang, Long Chen, Gengyu Wang, Hammad A. Ayyubi, Kai-Wei Chang, and Shih-Fu Chang. Idealgpt: Iteratively decomposing vision and language reasoning via large language models, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[53] Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. Large language models are not robust multiple choice selectors. In The Twelfth International Conference on Learning Representations, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[54] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3845" + } + ] +] \ No newline at end of file diff --git a/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/a1c25909-0376-4a01-a3cb-0e20c3229024_origin.pdf b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/a1c25909-0376-4a01-a3cb-0e20c3229024_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2bb42ed80852fcd8a2b2615930eb98f369481d11 --- /dev/null +++ b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/a1c25909-0376-4a01-a3cb-0e20c3229024_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60e10b8c3746164266dbe643fd1c61dde3d9d84ac46cc143eb26c5e10b9b04e6 +size 5221800 diff --git a/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/full.md b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2230214b979020215c34f111fdb19d5a23002afd --- /dev/null +++ b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/full.md @@ -0,0 +1,298 @@ +# Perception Tokens Enhance Visual Reasoning in Multimodal Language Models + +Mahtab Bigverdi1 Zelun Luo2 Cheng-Yu Hsieh1 +Ethan Shen1 Dongping Chen1 Linda G. Shapiro1 Ranjay Krishna1 +1University of Washington, 2Google Research + +{mahtab, cydhsieh, ethans03, shapiro, ranjay}@cs.washington.edu, alanzluo@gmail.com, cdp0612@uw.edu + +# Abstract + +Multimodal language models (MLMs) still face challenges in fundamental visual perception tasks where specialized models excel. Tasks requiring reasoning about 3D structures benefit from depth estimation, and reasoning about 2D object instances benefits from object detection. Yet, MLMs can not produce intermediate depth or boxes to reason over. Fine-tuning MLMs on relevant data doesn't generalize well and outsourcing computation to specialized vision tools is too comput-intensive and memory-inefficient. To address this, we introduce Perception Tokens, intrinsic image representations designed to assist reasoning tasks where language is insufficient. Perception tokens act as auxiliary reasoning tokens, akin to chain-of-thought prompts in language models. For example, in a depth-related task, an MLM augmented with perception tokens can reason by generating a depth map as tokens, enabling it to solve the problem effectively. We propose AURORA, a training method that augments MLMs with perception tokens for improved reasoning over visual inputs. AURORA leverages a VQVAE to transform intermediate image representations, such as depth maps into a tokenized format and bounding box tokens, which are then used in a multi-task training framework. AURORA achieves notable improvements across counting benchmarks: $+10.8\%$ on BLINK, $+11.3\%$ on CVBench, and $+8.3\%$ on SEED-Bench, outperforming fine-tuning approaches in generalization across datasets. It also improves on relative depth: over $+6\%$ on BLINK. With perception tokens, AURORA expands the scope of MLMs beyond language-based reasoning, paving the way for more effective visual reasoning capabilities. Code and data will be released at the project page. + +# 1. Introduction + +In contrast to the growing emphasis on building multimodal language models (MLMs), computer vision was originally + +![](images/85b4a617c42d599b1615a6f5b14654d479f7d1b763f68b4de3c3b9787b89dabf.jpg) +Which point is closer to the camera? + +# LLaVA-Aurora + +Step 1: + +Point A is at (250, 200), B is at (240, 300), and C is at (480, 210) + +Step 2: + +I need depth estimation to determine which point is closer + +Step 3: + +The depth map of the input image is: ... + +![](images/476fccd735a2ec8a8ed2a13c82625b56855dd87f6eb661f488c3301988b45edf.jpg) +Figure 1. We introduce Perception Tokens, intermediate reasoning tokens that allow MLMs to go beyond using language in reasoning. With it, we develop AURORA, a framework that trains multimodal language models to leverage visual perception tokens, allowing them to use depth estimation and bounding box predictions while reasoning. + +Step 4: + +Point B has higher pixel value on depth map, so it is closer + +attempting to interpret images as projections of indescribable 3D intrinsics, not just processing 2D arrays of language patterns [9, 34, 35]. Towards this endeavor, early vision research developed a series of intermediate image representations—enabling geometric reasoning through depth estimation [42] and instance reasoning through bounding box grounding [14]. As pointed out by recent work, we have focused less on such perceptual representations and instead tackled reasoning problems that require limited visual involvement [11, 40, 41]. This is likely because many traditional vision tasks remain ambiguous through natural language. Consider the task of identifying which of a set of N points is furthest away from the camera. While language doesn't lend itself to reason over this problem, a depth estimation would provide the appropriate abstraction to reason over. + +Numerous attempts have been made to enable MLMs to reason over intrinsic image representations. The default ap- + +![](images/86e7ba82d877c5ffe4ff24d7d3e8c68df57535da89c7e4f1fbe07fcbbfb256d5.jpg) +Figure 2. We demonstrate relative depth estimation and counting questions where LLaVA fails. In contrast, by learning to utilize visual perception tokens as intermediate reasoning steps, AURORA successfully complete these tasks requiring perceptual understanding. + +proach is to fine-tune MLMs on data tailored to the specific perception task of interest, hoping that the model implicitly learns the required intrinsic representations [4]. Another option is to outsource the computation to external tools: the MLM can invoke a depth estimator or object detector to produce the appropriate intrinsic [17]. Unfortunately, relying on external models makes the task more computationally expensive and requires loading additional models with more memory. Similarly, vanilla fine-tuning (even with advancements like LoRA [16]) has shown marginal improvements. + +Conceptually, we introduce Perception Tokens, intrinsic image representations that aid in reasoning where language is insufficient. To solve the aforementioned task, an MLM augmented with perception tokens can solve the task similar to how language models use chain-of-thought. They will produce a response like the following: "The depth map is . Therefore, point $D$ is closest to the camera." Here, is a set of tokens that implicitly estimate the depth of the image. Similarly, for a counting task, the model can first generate perception tokens that represents the location of the relevant bounding boxes of the desired object, and count the number of boxes to support its final answer. + +To demonstrate the utility of perception tokens, we introduce AURORA, a training algorithm to augment MLMs with the ability to use perception tokens as intermediate reasoning steps. For certain intermediate representations (e.g. depth maps), we train a VQVAE to transform them into a set of tokens, treating the learned VQVAE codebook indices as a collection of perception tokens. while for others, such as bounding boxes, we use directly encoded structured tokens. Next, we follow a multi-task training approach [15, 18] to train MLMs to use perception tokens as chain-of-thought tokens (see Fig. 1). Additionally, we adopt a curriculum learning approach to avoid catastrophic forgetting. + +We apply the AURORA training algorithm to the LLaVA [27] model, resulting in our LLaVA-AURORA variant. Our LLaVA-AURORA model significantly outperforms standard fine-tuning approaches across multiple perception + +demanding tasks, demonstrating the generality and effectiveness of our method. LLaVA-AURORA achieves state-of-the-art results on both relative depth estimation and object counting tasks. For instance, on BLINK relative depth estimation, LLaVA-AURORA delivers a performance boost of $6.4\%$ compared to the fine-tuning baseline. Similarly, on counting tasks, LLaVA-AURORA drives improvements of $10.8\%$ on BLINK, $11.3\%$ on CVBench, and $8.3\%$ on SEED-Bench. Fig. 2 illustrates examples from these tasks. Perception tokens open up a whole new modality through which MLMs can begin to reason, tackling tasks beyond just language reasoning. + +# 2. Related work + +Multimodal language models (MLMs). MLMs aim to solve a variety of tasks (e.g., Visual Question Answering (VQA) and captioning) based on vision and language inputs. Most modern architectures accomplish this by relying on either cross-attention [1, 2] or visual instruction tuning [26-28, 51] to interleave multimodal information. Cross-attention architectures operate by independently encoding images and then cross-attending to a language model backbone. On the other hand, visual instruction tuning produces token embeddings from image representations that can be interleaved with language tokens to ground generation. While these techniques work well for high level tasks, MLMs still struggle with mid-level and low-level tasks such as counting, depth reasoning, and segmentation. Most MLMs can be classified as either end-to-end MLMs or tool-using MLMs. + +End-to-end MLMs. End-to-end MLMs use a single, unified architecture [1, 8, 24, 26, 27] that can be repeatedly used for different tasks without requiring any architectural changes. While training end-to-end MLMs can be costly due to the need for large amounts of multi-task data, the use of diverse datasets allows end-to-end MLMs to generalize effectively and learn nuanced visual representations. + +Tool-using MLMs. Tool-using MLMs enable LLMs to perform vision tasks by attaching specialized vision modules such as segmentation or captioning networks [30, 33, 46, 52]. Tool-using MLMs use a router to select the optimal vision module to use for a given input. The use of specialized modules allows MLMs to achieve higher performances on many tasks. However, tool-using MLMs are sensitive to errors because they link together several networks, which are each a potential point of failure. + +Any-to-any networks. To address this, recent works [19, 20, 38, 45] such as Unified-IO [31, 32] have experimented with shared embedding spaces and visual decoders for vision and language tasks, training MLMs to generate segmentation masks, keypoints, and depth maps. Similarly, [47] generates special task tokens to route language model representations to diffusion heads for image, video, and audio generation. To handle complex segmentation scenarios, LISA [22] and GSVA [48] train language models to generate an additional segmentation token that can be used to produce segmentation masks, grounding the mask in language reasoning. + +While these approaches can generate visual outputs, they cannot reason over their own generations to solve related visual perception tasks. In contrast, augmenting MLMs with perception tokens as chain-of-thought tokens enables them to perform visual reasoning directly, providing significant gains in detail-oriented question-answering tasks such as depth estimation and counting. + +# 3. Perception Tokens & Aurora + +We introduce Perception Tokens and our Aurora training algorithm (see Fig. 3), which augment multimodal language models with perception tokens, enabling the MLM to leverage these tokens effectively during training and incorporate them into its chain-of-thought reasoning process for enhanced visual reasoning. + +# 3.1. Problem formulation + +In autoregressive large language models, chain-of-thought (CoT) reasoning can be formulated as a multi-step inferential process in which the model iteratively generates intermediate reasoning steps to arrive at a final answer. Given a task input $x$ , the model generates a response $y$ conditioned on the input and a sequence of $m$ intermediate reasoning steps $\{s_i\}_{i=1}^m$ , where $x, y$ , and each $s_i$ are sequences of tokens from the model's vocabulary $V$ . + +Existing multimodal language models often rely on limited vocabulary tokens derived from text or pre-trained image embeddings like CLIP, restricting their capacity to interpret other representations crucial for reasoning. Mid-level and low-level visual features such as depth maps, instance segmentation masks, human poses and bounding + +boxes which could substantially improve visual reasoning, are currently incompatible with the model and cannot be integrated during training or inference. Our key insight is to introduce auxiliary perception tokens for these intermediate steps with an expanded vocabulary $V' = V \cup V_{\mathrm{aux}}$ , bridging this gap by allowing the model to integrate richer visual representations into its reasoning process. Conditioning the final output on these tokens enhances the model's accuracy and interpretability across multimodal tasks. + +# 3.2. Perception token prediction and reasoning + +Introducing an expanded vocabulary to enhance multimodal reasoning presents two main challenges. The first challenge is enabling the model to generate tokens from the new auxiliary vocabulary $V_{\mathrm{aux}}$ , which includes specialized tokens for low- and mid-level visual features, such as depth maps and bounding boxes. The second challenge is ensuring that the model can effectively condition on these auxiliary tokens to improve reasoning, particularly for multi-step inference tasks. + +Perception token prediction. To address the first challenge, we employ a specialist-to-generalist distillation approach, using pre-trained specialist models (e.g., depth estimation or instance segmentation) to guide auxiliary token generation through cross-entropy loss. For each input $x$ , the specialist model provides a target probability distribution $q_{i}$ over its tokens. Let $M: V_{\mathrm{spec}} \to V_{\mathrm{aux}}$ denote a one-to-one mapping from the specialist model's vocabulary $V_{\mathrm{spec}}$ to the auxiliary token vocabulary $V_{\mathrm{aux}}$ . We define the distillation loss as: + +$$ +\ell_ {d i s t} = \min _ {M} \left(- \sum_ {i} q _ {i} \log p _ {M (i)}\right), \tag {1} +$$ + +where $p_{M(i)}$ is the probability assigned by our model to the auxiliary token corresponding to the mapped token $M(i)$ . This consistent mapping allows the model to effectively align its predictions with the specialist model's output distribution, enhancing the relevance and accuracy of its auxiliary token predictions. + +In addition to distillation, we incorporate a reconstruction loss to enhance the token prediction and the interpretability of our model. Each set of auxiliary tokens corresponds to a specific representation, such as a depth map or a bounding box vector, and is trained to directly predict this feature. To achieve this, we introduced a lightweight decoder $g$ that maps the tokens into the feature space, allowing for efficient and interpretable transformations. Formally, for a token set $T \subset V_{\mathrm{aux}}$ , decoder $g$ , and its target feature $f$ , the reconstruction loss is defined as: + +$$ +\ell_ {r e c} = \left\| g (T) - f \right\| _ {2} ^ {2}, \tag {2} +$$ + +where $g(T)$ is the decoded representation for the token set $T$ in the feature space. This reconstruction process not only + +![](images/c15e4918570fa52681db82a6387e166441e0238746739e6b88dd82855da39622.jpg) +Figure 3. The overall AURORA training framework. We first learn visual perception tokens using VQVAE. We then finetune MLMs with a multi-task training approach where we distill intrinsic image representations (e.g., depth map) into MLMs by training them to decode the visual tokens as intermediate reasoning steps towards completing the tasks. + +aligns each token with a meaningful feature, improving interpretability, but also reinforces the accuracy of auxiliary token predictions through direct feature supervision. In practice, while the reconstruction objective improves performance and interpretability, it is optional, as models can be effectively trained with the distillation objective alone to reduce computational overhead. + +Reasoning with perception tokens. The second challenge is enabling the model to condition on tokens from $V_{\mathrm{aux}}$ effectively when generating each subsequent reasoning step, thereby enhancing its reasoning capabilities. To achieve this, we progressively introduce chain-of-thought reasoning during training, starting with simpler, single-step reasoning examples and gradually incorporating more complex, multi-step reasoning data. The model begins by learning single-step reasoning, predicting an initial reasoning step $s_1$ based on the input $x$ . It then progresses to multistep reasoning, predicting sequences $s_1, \dots, s_m$ and effectively utilizing auxiliary tokens to support extended chains of thought. To facilitate this, we propose a multi-task data synthesis approach combined with curriculum learning across a variety of synthetic tasks. Additional implementation details can be found in Section 3.4. Finally, we reinforce this process at inference time using constrained decoding, wherein sampling is selectively restricted to auxiliary tokens at certain reasoning steps, guiding the model to generate coherent intermediate steps from $V_{\mathrm{aux}}$ . + +# 3.3. Tokenization + +A unified tokenization space is crucial for multimodal models as it creates a consistent framework through which varied visual tasks can be represented, processed, and interpreted. Inspired by [36], we establish a unified tokenization space which enables the model to learn varied visual fea + +tures within a shared representation seamlessly. In our experiments, we implement two tailored tokenization schemes for commonly used visual representations. Importantly, our framework is designed with flexibility, allowing it to generalize to a broad range of visual representations beyond those presented here. + +Pixel-level representation. This tokenization scheme captures fine-grained spatial information, such as depth maps and segmentation masks, providing the model with detailed pixel-level data essential for accurate visual processing. For these types of tokens, we leverage visual tokenizers like VQVAE and VQGAN, which take in the ground truth masks or depth maps and return discrete target tokens [10, 31, 32, 44]. + +Structured representation. This scheme encodes structured yet abstracted visual features, such as human poses, bounding boxes, and coordinates, allowing the model to reason with higher-level spatial relationships and object hierarchies. For these tokens, we define the domain of the tokens based on specific properties; for example, the domain for coordinates can range from 0 to the maximum number of pixels in the image's height or width [5]. + +# 3.4. Curriculum learning with progressive CoTs + +The objective of training the model is to develop a data and computationally efficient method for learning to predict novel, fine-grained visual tokens and using them to complete complex visual reasoning tasks. We observed that the standard approach, which relies on a fixed mixture data, encounters a trade-off between the accuracy of novel tokens predictions and the model's reasoning capability, primarily due to catastrophic forgetting and challenges in reasoning with new tokens. Conversely, fine-tuning the model with the original training mixture significantly raises computational + +
ModelTrainingBLINK [11]2 PointsHardBLINK3 PointsHardBLINK4 PointsHardBLINK5 PointsAverage
DirectLa-labelingDataDepthGener-ationDataCoTData
LLaVA OneVision 72BXXX67.751.650.044.353.4
LLaVA 1.5 13BXXX54.035.537.92939.1
Fine-tuned LLaVA✓XX68.558.952.441.155.2
LLaVA-AURORA (Ours)✓✓✓64.566.960.554.861.6
GPT-4oXXX53.258.95036.349.6
GPT-4 TurboXXX58.154.841.932.246.7
GPT-4 Turbo + ToolXXX70.257.244.326.649.6
+ +Table 1. Performance comparison between our fine-tuned model, the baseline (without depth tokens), and the original base model on the relative depth accuracy (\%) task. Results demonstrate that our approach, utilizing depth tokens and intermediate reasoning steps, significantly outperforms both the baseline and the base model, particularly on more challenging configurations with 3, 4, and 5 points sampled from the image's mid-height region. The gray rows represent closed-source models, providing additional baselines for comparison. + +costs and may be impractical if the original data, particularly proprietary datasets, is unavailable, making this approach less scalable for incorporating new tokens in the future. + +We propose a curriculum learning-inspired training scheme that begins with atomic tasks and gradually advances to more complex ones requiring sophisticated, multi-hop reasoning. Let $d_{t}$ represent the difficulty of task $t$ , with difficulties $d_{1} < d_{2} < \dots d_{T}$ across $T$ tasks, and let $p(t,s)$ denote the probability of sampling data points from task $t$ at training step $s$ . We define $p(d_{t},s)$ using a temperature-scaled Softmax formulation as follows: + +$$ +p (d _ {t}, s) = \frac {\exp \left(- d _ {t} / \tau (s)\right)}{\sum_ {i = 1} ^ {m} \exp \left(- d _ {i} / \tau (s)\right)}, \tag {3} +$$ + +where $\tau(s)$ modulates the task difficulty over time, allowing a smooth shift in the probability distribution toward harder tasks. This temperature function is defined as: + +$$ +\tau (s) = \frac {\tau_ {0}}{1 + \lambda \cdot s / S}. \tag {4} +$$ + +Here, $\tau_0$ is the initial temperature, $\lambda$ is the annealing rate, and $S$ is the total number of training steps. + +Our approach for defining $d_{t}$ values is based on the inherent complexity of each task, which corresponds to the depth of reasoning involved. Specifically, we assign $d_{1}$ to the most atomic task, involving the prediction of newly introduced tokens. At the other end of the spectrum, $d_{m}$ represents the final task, requiring the full chain-of-thought (CoT) reasoning steps. Between them, intermediate tasks $\{d_{t}\}_{t = 1}^{T}$ , serves to bridge the gap between the atomic tasks and the comprehensive chain-of-thought responses. + +In this project, we introduce three types of data subsets for each downstream task, organized in increasing levels of difficulty. The first and most atomic task involves teaching + +the model to generate tokens from the new auxiliary vocabulary. For instance, in depth-related tasks, we train the model to learn depth maps; in segmentation-related tasks, we teach the model to generate masks; and so on. + +The other two data subsets involve chain-of-thought (CoT) prompts and direct labeling which help with reasoning with auxiliary tokens. In the CoT subset, we use the new intermediate visual perception tokens to answer downstream-specific questions, encouraging the model to reason step by step. In the direct labeling subset, we pose the same questions but instruct the model to provide direct answers without step-by-step reasoning. + +Inspired by [15], we employ a multitasking approach for these two data subsets. For each image, we sequentially present both the CoT and direct labeling questions, allowing the model to tackle each image with both reasoning styles in sequence. We use a sequential sampler rather than a random sampler, shuffling the images beforehand. This strategy enables the model to learn from both types of reasoning tasks effectively, enhancing its ability to perform complex visual reasoning. More details in Sec. 9. + +# 4. Experiments + +We base our work on LLaVA 1.5 13B as the foundation for our model, which we refer to as LLaVA-AURORA. Our approach augments the MLM with perception tokens to enhance reasoning and improve performance across both 3D and 2D visual tasks. We evaluate our approach on relative depth estimation (3D) using pixel-level depth map tokens for fine-grained depth capture, and on object counting (2D) with mid-level bounding box tokens for precise localization. These tokens not only enhance task-specific results but also highlight our framework's potential to generalize effectively across a broad spectrum of visual reasoning tasks. + +# 4.1. 3D reasoning task + +We choose relative depth estimation as our 3D task because it enables the model to determine spatial relationships within a scene by identifying which points are closer to or farther from the camera. This foundational skill is crucial for scene understanding and applications requiring spatial awareness, such as robotics and autonomous systems. Specifically, this task involves identifying the point closest to the camera among multiple marked points in an image. To support the model's reasoning, we use discrete depth map tokens that capture spatial depth information, enhancing the model's understanding of proximity. + +Tokenization. To capture fine-grained spatial details, we tokenize depth maps into sequences of discrete tokens. Inspired by the approach in AiT [36], we use a Vector Quantized Variational Autoencoder (VQVAE) [44] with a codebook size of 128. In this setup, each depth map is encoded as a grid of embeddings, with each embedding matched to the nearest entry in the codebook, yielding a compact depth representation. The VQVAE decoder reconstructs the depth map from this sequence of latent codes, and the entire model is optimized with a reconstruction loss to ensure precise encoding. Each $320 \times 320$ depth map is compressed into a $10 \times 10$ grid of code indices, resulting in a 100-token sequence where each token corresponds to one of the 128 codebook entries. We introduce 128 unique depth tokens, labeled $DEPTH\_0$ through $DEPTH\_127$ , and enclose each sequence with special tokens $DEPTH\_START$ and $DEPTH\_END$ , expanding the model's vocabulary by 130 depth-related tokens in total. + +Training data. We train the VQVAE model on pseudodepth maps generated from the ADE20k dataset [54, 55] using the Depth Anything model [49, 50]. This dataset provides a diverse range of scenes, enhancing the model's ability to generalize. + +For fine-tuning, we prepare three types of data tailored for relative depth estimation (as detailed in Sec. 3.4). First, we generate depth maps for $20\mathrm{k}$ ADE20k images, tokenize them with the pre-trained VQVAE, and format each sample in a Q&A structure, prompting the model with a depth estimation question and an answer sequence of depth tokens. Additionally, we construct a dataset of 500 ADE20k images for chain-of-thought (CoT) reasoning, with 2-5 markers in each image. Here, the prompt guides the model to generate the coordinates of the markers and then the depth map as intermediate steps, then identify the marker closest to the camera. This CoT training improves sequential reasoning and relative depth estimation accuracy. Finally, we use the same 500 images for direct labeling, prompting the model to directly identify and label the marker closest to the camera, see Sec. 9. + +This fine-tuning setup enables our model to effectively + +handle step-by-step reasoning for the relative depth estimation task. + +# 4.2. 2D reasoning task + +We select counting as a critical 2D visual task. For object counting, we incorporate bounding box predictions as an intermediate reasoning step to improve accuracy in answering counting queries. Given an image and a question about the number of specific objects, the model first identifies and predicts bounding boxes for each instance of the target object. These bounding box tokens serve as structured, intermediate representations, enabling the model to understand spatial arrangements and accurately count object instances. + +Tokenization. To represent bounding boxes as discrete tokens, we resize all input images to a fixed resolution of 336x336 pixels. This preprocessing step allows us to add 336 unique tokens to the model's vocabulary, each representing a specific pixel position within the resized image. These tokens, labeled Pixel_0 to Pixel_335, enable the model to uniquely reference each pixel location. Bounding boxes are encoded as tuples of four tokens, formatted as (PIXEL_i, Pixel_j, Pixel_k, Pixel_m), where Pixel_i and Pixel_j denote the coordinates of the top-left corner and Pixel_k and Pixel_m represent the bottom-right corner (i.e., $(x_{1},y_{1},x_{2},y_{2})$ ). This discrete representation allows the model to interpret and use bounding box locations effectively, providing the spatial structure needed for accurate object counting. + +Training data. To fine-tune the model for object counting, we draw on three types of data tailored to this task (as detailed in Sec. 3.4). First, we use task-specific data from the LVIS dataset [13], selecting 5k images with objects whose counts range from 0 to 15. For each selected image, we specify an object type (e.g., "beds") and structure the fine-tuning samples to prompt the model for bounding box predictions of the specified objects within the image. + +To enhance the model's reasoning ability, we include a small subset of 250 LVIS images for chain-of-thought (CoT) training. Here, each question prompt encourages step-by-step reasoning by instructing the model to first generate bounding boxes for the target object, followed by providing the final count. Additionally, we create a direct labeling subset using the same 250 images. In this subset, we prompt the model to directly identify and label the total count of the specified objects without the intermediate step of bounding box generation. + +# 4.3. Benchmarks + +Relative depth. A recent benchmark, BLINK [11], introduces tasks designed to be intuitive for humans yet challenging for multimodal models, with relative depth estimation as one of its tasks. BLINK provides 124 images, each + +
ModelTrainingCV-Bench CountingSEED-Bench CountingBLINK Counting
Direct Labeling DataBounding Box DataCoT Data
LLaVA One VisionXXX34.431.735.8
LLaVA 1.5 13BXXX40.952.235.0
Fine-tuned LlaVA✓XX44.746.30.2
LLaVA-AURORA (Ours)✓✓✓56.054.645.8
GPT-4oXXX70.1864.647.5
GPT-4 TurboXXX61.364.857.5
GPT-4 Turbo + ToolXXX48.629.926.7
+ +Table 2. Comparison of object counting accuracy (%) across three benchmarks (CV-Bench, SEED-Bench, and BLINK). Our Aurora model, using auxiliary tokens to encode bounding box information for intermediate reasoning, demonstrates superior performance compared to the baseline models and the original base model. The gray rows represent closed-source models, providing additional baselines for comparison. + +containing two marked points labeled as $A$ and $B$ , and asks which point is closer to the camera. To reduce biases that language models have toward answering multiple-choice questions [39, 53], we modify the original BLINK questions by removing the answer choices. To further evaluate the model's reasoning and 3D understanding in relative depth, we curated a series of more challenging benchmark sets, collectively called HardBLINK. We progressively increase task difficulty by altering the prompts and image configurations as follows: + +1. Prompt Modification: In the prompts, we exclude the number of markers and their labels, requiring the model to infer these details solely from the image. +2. Increased Point Complexity: We generate four variations of BLINK by adding more markers to each image, using Depth Anything to produce pseudo-depth maps for precise placement. These curated sets—HardBLINK 3points, HardBLINK 4points, and HardBLINK 5points—contain the same images as BLINK but with 3, 4, and 5 randomly placed markers, respectively, each with reasonable depth and distance differences. This setup tests the model's depth reasoning across more complex spatial configurations. +3. Mitigating Height-Based Bias: To prevent the model from assuming that higher points are farther from the camera, we place markers at mid-height within the image. This approach encourages reliance on depth information rather than positional cues [6]. + +Counting. For counting, we evaluate the model on CV-Bench [40], SEED-Bench [23], and BLINK's counting [11] subtask. To better capture the model's capabilities, we also remove multiple-choice options, requiring it to generate an exact count. + +# 4.4. Baselines + +We evaluate a diverse range of models, including closed-source models like GPT-4o [37] and GPT-4 Turbo [37], as well as state-of-the-art open-source models such as LLaVA + +![](images/1d0f4997cca7b4f96551273234bc3befc64ac3a0623eedebeff7bae07b89c5c5.jpg) +Figure 4. Depth maps generated by Aurora are imperfect but resemble the ground-truths from Depth Anything [49]. + +OneVision [25]. Another key baseline for our work involves fine-tuning the base model AURORA is applied on, solely on the direct labeling portion of the training data for each task, omitting the newly introduced tokens. This approach allows us to isolate the impact of our token-based enhancements. Additionally, we evaluate the base model for AURORA, LLaVA 1.5 13B, to assess its performance without task-specific adaptations. + +We also add tool-augmented baselines. For relative depth estimation, we employ GPT-4 Turbo, providing it with the ground truth depth maps generated by Depth Anything for each image, allowing it to use this information in its responses [17]. For counting, we also use GPT-4 Turbo in a tool-augmented setup. In this process, GPT-4 Turbo first identifies the object specified in the question, then uses Grounding DINO [29] to locate the bounding boxes for each instance of the object and finally counts them. + +# 4.5. How do new tokens improve 3D reasoning? + +Our experiments demonstrate that incorporating new visual tokens significantly enhances the model's 3D reasoning abilities, specifically in the relative depth estimation task. Results in Tab. 1 show that our model outperforms both the primary baseline, which is fine-tuned solely on direct label- + +
ModelBLINK [11] 2 pointsHardBLINK 3 PointsHardBLINK 4 PointsHardBLINK 5 PointsAverage
VQGAN [10] (16384 codes)82.266.1533759.6
Unified-IO70.275.875.875.874.4
Unified-IO25437.92127.435.1
AURORA (Ours)91.978.271.775.879.2
Our VQVAE (128 codes)96.794.395.296.795.7
+ +Table 3. While not the main aim of our work, we report the depth generation performance across benchmarks with 2, 3, 4, and 5 marked points using BLINK [11]'s relative depth subtask images. We report relative depth estimation accuracy $(\%)$ , calculated by programmatically extracting depth values at specific coordinates from model-generated depth maps. Our model consistently outperforms other multimodal models, including Unified-IO [31] and Unified-IO 2 [32]. The gray row represents our VQVAE model, which serves as the upper bound for performance. + +ing data, and the original base model, indicating that the added tokens contribute meaningfully to task accuracy. + +As task complexity increases—for instance, when additional markers are introduced in the HardBLINK benchmarks—our model's performance advantage becomes even more pronounced. Not only does it maintain high accuracy in distinguishing depth relationships, but it also surpasses both advanced closed-source models, such as GPT-4o, GPT-4 Turbo, and GPT-4 Turbo + Tool that use ground truth depth maps. This suggests that the new tokens enable our model to develop a more nuanced understanding of depth cues, even in cases where complex spatial reasoning is required. + +Overall, these results highlight the effectiveness of our token-based approach in enhancing 3D reasoning, allowing the model to handle increasingly difficult tasks with robust performance and accuracy. + +# 4.6. How do new tokens improve 2D reasoning? + +In the 2D task of object counting, incorporating new visual tokens provides a significant advantage over the primary baselines. As shown in Tab. 2, our model outperforms both the baseline fine-tuned solely on direct labeling data and the original base model, as well as the state-of-the-art open-source model LLaVA-OneVision and GPT-4 Turbo + Tool, underscoring the value of these tokens in enhancing counting accuracy. + +Although our model does not yet surpass advanced closed-source models, the results demonstrate that the new tokens yield a meaningful improvement in 2D reasoning, enabling more reliable and accurate object counting compared to standard fine-tuning approaches and open-source alternatives. + +# 4.7. Perception token decoding + +Our approach enables the decoding of learned perception tokens into specialist features, such as depth maps and object bounding boxes, to assess their fidelity and utility. We, in particular, evaluate the accuracy and correctness + +of the depth maps generated by LLaVA-AURORA, specifically focusing on its ability to represent spatial relationships in visual scenes. For this evaluation, we use the relative depth images from the BLINK benchmark [11], ensuring consistency with our relative depth assessments. LLaVA-AURORA generates depth tokens from BLINK images, which are then reconstructed into full depth maps via the decoder of our pre-trained VQVAE model. Notably, the depth maps output by our pre-trained VQVAE provide an upper bound on the quality of the depth maps generated by LLaVA-AURORA. We use programmatic relative depth accuracy as our metric, which measures how well the reconstructed depth maps capture the relative depth of marked points. This evaluation spans several configurations in the benchmark, including images with 2, 3, 4, and 5 labeled points. By comparing the model's predicted depth with the ground-truth marker coordinates, we calculate relative depth accuracy, allowing us to assess the precision of depth map generation in reflecting spatial depth relationships. As shown in Tab. 3, LLaVA-AURORA outperforms Unified-IO [31, 32] in this task. Furthermore, qualitative analysis in Fig. 4 reveals that LLaVA-AURORA's depth maps capture spatial details effectively highlighting its capacity to interpret and represent fine-grained depth information. + +# 5. Conclusion + +Our algorithm enables the lightweight and scalable integration of perception tokens, such as depth maps and bounding box coordinates, into MLMs, allowing them to perform intermediate reasoning steps akin to chain-of-thought processes. Our method achieves state-of-the-art results on challenging tasks like 2D object counting and 3D relative depth estimation. It also enhances model generalization and interpretability without relying on external tools or task-specific fine-tuning. The framework is inherently adaptable, incorporating new perception tokens as they emerge, making it a future-proof solution for advancing multimodal reasoning. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning, 2022. +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. +[3] Luca Beurer-Kellner, Marc Fischer, and Martin Vechev. Guiding llms the right way: Fast, non-invasive constrained generation. arXiv preprint arXiv:2403.06988, 2024. +[4] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. +[5] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021. +[6] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. +[7] Aidan Cooper. A guide to structured outputs using constrained decoding, 2024. +[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. +[9] PHILOSO EPHY DO CT OR OF. MACHINE PERCEPTION OF THREE-DIMENSIONAL, SO LIDS. PhD thesis, MASSACHUSETTS INSTITUTE OF TECHNOLOGY, 1961. +[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. +[11] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024. +[12] Saibo Geng, Martin Josifoski, Maxime Peyrard, and Robert West. Grammar-constrained decoding for structured nlp tasks without finetuning. arXiv preprint arXiv:2305.13971, 2023. +[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Pro + +ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. +[14] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, pages 10-5244. CiteSeer, 1988. +[15] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. arXiv preprint arXiv:2305.02301, 2023. +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. +[17] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. +[18] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9601, 2024. +[19] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu, Chengru Song, Xiaoqiang Lei, Di Zhang, Wenwu Ou, Kun Gai, and Yadong Mu. Unified language-vision pretraining in llm with dynamic discrete visual tokenization, 2024. +[20] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models, 2023. +[21] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. +[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model, 2024. +[23] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. +[24] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning, 2023. +[25] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. +[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. +[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. + +[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. +[29] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. +[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024. +[31] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks, 2022. +[32] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action, 2023. +[33] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models, 2023. +[34] David Marr. Vision: A computational investigation into the human representation and processing of visual information. MIT press, 2010. +[35] Marvin Minsky and Seymour Papert. An introduction to computational geometry. Cambridge, 1947: 104, 1969. +[36] Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Unifying output space of visual tasks via soft token. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19900-19910, 2023. +[37] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. +[38] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models, 2024. +[39] Pouya Pezeshkpour and Estevam Hruschka. Large language models sensitivity to the order of options in multiple-choice questions. arXiv preprint arXiv:2308.11483, 2023. +[40] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. +[41] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. +[42] Antonio Torralba and Aude Oliva. Depth estimation from image structure. IEEE Transactions on pattern analysis and machine intelligence, 24(9):1226-1238, 2002. + +[43] Vivien Tran-Thien. Fast, high-fidelity llm decoding with regex constraints, 2024. +[44] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. +[45] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework, 2022. +[46] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models, 2023. +[47] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm, 2024. +[48] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models, 2024. +[49] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. +[50] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. +[51] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl: Modularization empowers large language models with multimodality, 2024. +[52] Haoxuan You, Rui Sun, Zhecan Wang, Long Chen, Gengyu Wang, Hammad A. Ayyubi, Kai-Wei Chang, and Shih-Fu Chang. Idealgpt: Iteratively decomposing vision and language reasoning via large language models, 2023. +[53] Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. Large language models are not robust multiple choice selectors. In The Twelfth International Conference on Learning Representations, 2023. +[54] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. +[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. \ No newline at end of file diff --git a/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/images.zip b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5209237bb92d7bdbe9ac0ddc0fef7fad14387d06 --- /dev/null +++ b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80df19f8856940f847f6f78345492038f8618d9212e871693c3677ba12d9dafb +size 325121 diff --git a/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/layout.json b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3a19615d51c9ea3a4fe99c1e50bf114aad2319a1 --- /dev/null +++ b/2025/Perception Tokens Enhance Visual Reasoning in Multimodal Language Models/layout.json @@ -0,0 +1,7687 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 549, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 549, + 122 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 549, + 122 + ], + "type": "text", + "content": "Perception Tokens Enhance Visual Reasoning in Multimodal Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 142, + 495, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 142, + 495, + 185 + ], + "spans": [ + { + "bbox": [ + 114, + 142, + 495, + 185 + ], + "type": "text", + "content": "Mahtab Bigverdi1 Zelun Luo2 Cheng-Yu Hsieh1 \nEthan Shen1 Dongping Chen1 Linda G. Shapiro1 Ranjay Krishna1 \n1University of Washington, 2Google Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 186, + 517, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 186, + 517, + 213 + ], + "spans": [ + { + "bbox": [ + 90, + 186, + 517, + 213 + ], + "type": "text", + "content": "{mahtab, cydhsieh, ethans03, shapiro, ranjay}@cs.washington.edu, alanzluo@gmail.com, cdp0612@uw.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 266, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 296, + 649 + ], + "type": "text", + "content": "Multimodal language models (MLMs) still face challenges in fundamental visual perception tasks where specialized models excel. Tasks requiring reasoning about 3D structures benefit from depth estimation, and reasoning about 2D object instances benefits from object detection. Yet, MLMs can not produce intermediate depth or boxes to reason over. Fine-tuning MLMs on relevant data doesn't generalize well and outsourcing computation to specialized vision tools is too comput-intensive and memory-inefficient. To address this, we introduce Perception Tokens, intrinsic image representations designed to assist reasoning tasks where language is insufficient. Perception tokens act as auxiliary reasoning tokens, akin to chain-of-thought prompts in language models. For example, in a depth-related task, an MLM augmented with perception tokens can reason by generating a depth map as tokens, enabling it to solve the problem effectively. We propose AURORA, a training method that augments MLMs with perception tokens for improved reasoning over visual inputs. AURORA leverages a VQVAE to transform intermediate image representations, such as depth maps into a tokenized format and bounding box tokens, which are then used in a multi-task training framework. AURORA achieves notable improvements across counting benchmarks: " + }, + { + "bbox": [ + 54, + 266, + 296, + 649 + ], + "type": "inline_equation", + "content": "+10.8\\%" + }, + { + "bbox": [ + 54, + 266, + 296, + 649 + ], + "type": "text", + "content": " on BLINK, " + }, + { + "bbox": [ + 54, + 266, + 296, + 649 + ], + "type": "inline_equation", + "content": "+11.3\\%" + }, + { + "bbox": [ + 54, + 266, + 296, + 649 + ], + "type": "text", + "content": " on CVBench, and " + }, + { + "bbox": [ + 54, + 266, + 296, + 649 + ], + "type": "inline_equation", + "content": "+8.3\\%" + }, + { + "bbox": [ + 54, + 266, + 296, + 649 + ], + "type": "text", + "content": " on SEED-Bench, outperforming fine-tuning approaches in generalization across datasets. It also improves on relative depth: over " + }, + { + "bbox": [ + 54, + 266, + 296, + 649 + ], + "type": "inline_equation", + "content": "+6\\%" + }, + { + "bbox": [ + 54, + 266, + 296, + 649 + ], + "type": "text", + "content": " on BLINK. With perception tokens, AURORA expands the scope of MLMs beyond language-based reasoning, paving the way for more effective visual reasoning capabilities. Code and data will be released at the project page." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 669, + 135, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 135, + 681 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 135, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "In contrast to the growing emphasis on building multimodal language models (MLMs), computer vision was originally" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 303, + 400, + 387 + ], + "blocks": [ + { + "bbox": [ + 316, + 282, + 394, + 300 + ], + "lines": [ + { + "bbox": [ + 316, + 282, + 394, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 394, + 300 + ], + "type": "text", + "content": "Which point is closer to the camera?" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 303, + 400, + 387 + ], + "lines": [ + { + "bbox": [ + 316, + 303, + 400, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 400, + 387 + ], + "type": "image", + "image_path": "85b4a617c42d599b1615a6f5b14654d479f7d1b763f68b4de3c3b9787b89dabf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 450, + 247, + 508, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 450, + 247, + 508, + 256 + ], + "spans": [ + { + "bbox": [ + 450, + 247, + 508, + 256 + ], + "type": "text", + "content": "LLaVA-Aurora" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 408, + 266, + 434, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 266, + 434, + 276 + ], + "spans": [ + { + "bbox": [ + 408, + 266, + 434, + 276 + ], + "type": "text", + "content": "Step 1:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 444, + 268, + 544, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 444, + 268, + 544, + 285 + ], + "spans": [ + { + "bbox": [ + 444, + 268, + 544, + 285 + ], + "type": "text", + "content": "Point A is at (250, 200), B is at (240, 300), and C is at (480, 210)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 409, + 293, + 434, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 293, + 434, + 301 + ], + "spans": [ + { + "bbox": [ + 409, + 293, + 434, + 301 + ], + "type": "text", + "content": "Step 2:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 444, + 294, + 543, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 444, + 294, + 543, + 311 + ], + "spans": [ + { + "bbox": [ + 444, + 294, + 543, + 311 + ], + "type": "text", + "content": "I need depth estimation to determine which point is closer" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 409, + 319, + 434, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 319, + 434, + 327 + ], + "spans": [ + { + "bbox": [ + 409, + 319, + 434, + 327 + ], + "type": "text", + "content": "Step 3:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 444, + 319, + 541, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 444, + 319, + 541, + 343 + ], + "spans": [ + { + "bbox": [ + 444, + 319, + 541, + 343 + ], + "type": "text", + "content": "The depth map of the input image is: ... " + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 444, + 344, + 478, + 376 + ], + "blocks": [ + { + "bbox": [ + 444, + 344, + 478, + 376 + ], + "lines": [ + { + "bbox": [ + 444, + 344, + 478, + 376 + ], + "spans": [ + { + "bbox": [ + 444, + 344, + 478, + 376 + ], + "type": "image", + "image_path": "476fccd735a2ec8a8ed2a13c82625b56855dd87f6eb661f488c3301988b45edf.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 413, + 555, + 479 + ], + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 479 + ], + "type": "text", + "content": "Figure 1. We introduce Perception Tokens, intermediate reasoning tokens that allow MLMs to go beyond using language in reasoning. With it, we develop AURORA, a framework that trains multimodal language models to leverage visual perception tokens, allowing them to use depth estimation and bounding box predictions while reasoning." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 409, + 383, + 434, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 383, + 434, + 391 + ], + "spans": [ + { + "bbox": [ + 409, + 383, + 434, + 391 + ], + "type": "text", + "content": "Step 4:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 444, + 384, + 538, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 444, + 384, + 538, + 401 + ], + "spans": [ + { + "bbox": [ + 444, + 384, + 538, + 401 + ], + "type": "text", + "content": "Point B has higher pixel value on depth map, so it is closer" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 495, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 685 + ], + "type": "text", + "content": "attempting to interpret images as projections of indescribable 3D intrinsics, not just processing 2D arrays of language patterns [9, 34, 35]. Towards this endeavor, early vision research developed a series of intermediate image representations—enabling geometric reasoning through depth estimation [42] and instance reasoning through bounding box grounding [14]. As pointed out by recent work, we have focused less on such perceptual representations and instead tackled reasoning problems that require limited visual involvement [11, 40, 41]. This is likely because many traditional vision tasks remain ambiguous through natural language. Consider the task of identifying which of a set of N points is furthest away from the camera. While language doesn't lend itself to reason over this problem, a depth estimation would provide the appropriate abstraction to reason over." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Numerous attempts have been made to enable MLMs to reason over intrinsic image representations. The default ap-" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3836" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 202 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 202 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 202 + ], + "type": "image", + "image_path": "86e7ba82d877c5ffe4ff24d7d3e8c68df57535da89c7e4f1fbe07fcbbfb256d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 555, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 225 + ], + "type": "text", + "content": "Figure 2. We demonstrate relative depth estimation and counting questions where LLaVA fails. In contrast, by learning to utilize visual perception tokens as intermediate reasoning steps, AURORA successfully complete these tasks requiring perceptual understanding." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 236, + 294, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 236, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 294, + 356 + ], + "type": "text", + "content": "proach is to fine-tune MLMs on data tailored to the specific perception task of interest, hoping that the model implicitly learns the required intrinsic representations [4]. Another option is to outsource the computation to external tools: the MLM can invoke a depth estimator or object detector to produce the appropriate intrinsic [17]. Unfortunately, relying on external models makes the task more computationally expensive and requires loading additional models with more memory. Similarly, vanilla fine-tuning (even with advancements like LoRA [16]) has shown marginal improvements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 359, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 359, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 359, + 295, + 515 + ], + "type": "text", + "content": "Conceptually, we introduce Perception Tokens, intrinsic image representations that aid in reasoning where language is insufficient. To solve the aforementioned task, an MLM augmented with perception tokens can solve the task similar to how language models use chain-of-thought. They will produce a response like the following: \"The depth map is . Therefore, point " + }, + { + "bbox": [ + 54, + 359, + 295, + 515 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 359, + 295, + 515 + ], + "type": "text", + "content": " is closest to the camera.\" Here, is a set of tokens that implicitly estimate the depth of the image. Similarly, for a counting task, the model can first generate perception tokens that represents the location of the relevant bounding boxes of the desired object, and count the number of boxes to support its final answer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 518, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 518, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 54, + 518, + 295, + 663 + ], + "type": "text", + "content": "To demonstrate the utility of perception tokens, we introduce AURORA, a training algorithm to augment MLMs with the ability to use perception tokens as intermediate reasoning steps. For certain intermediate representations (e.g. depth maps), we train a VQVAE to transform them into a set of tokens, treating the learned VQVAE codebook indices as a collection of perception tokens. while for others, such as bounding boxes, we use directly encoded structured tokens. Next, we follow a multi-task training approach [15, 18] to train MLMs to use perception tokens as chain-of-thought tokens (see Fig. 1). Additionally, we adopt a curriculum learning approach to avoid catastrophic forgetting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 295, + 714 + ], + "type": "text", + "content": "We apply the AURORA training algorithm to the LLaVA [27] model, resulting in our LLaVA-AURORA variant. Our LLaVA-AURORA model significantly outperforms standard fine-tuning approaches across multiple perception" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "text", + "content": "demanding tasks, demonstrating the generality and effectiveness of our method. LLaVA-AURORA achieves state-of-the-art results on both relative depth estimation and object counting tasks. For instance, on BLINK relative depth estimation, LLaVA-AURORA delivers a performance boost of " + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "text", + "content": " compared to the fine-tuning baseline. Similarly, on counting tasks, LLaVA-AURORA drives improvements of " + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "inline_equation", + "content": "10.8\\%" + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "text", + "content": " on BLINK, " + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "inline_equation", + "content": "11.3\\%" + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "text", + "content": " on CVBench, and " + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "inline_equation", + "content": "8.3\\%" + }, + { + "bbox": [ + 313, + 236, + 555, + 380 + ], + "type": "text", + "content": " on SEED-Bench. Fig. 2 illustrates examples from these tasks. Perception tokens open up a whole new modality through which MLMs can begin to reason, tackling tasks beyond just language reasoning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 397, + 399, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 397, + 399, + 410 + ], + "spans": [ + { + "bbox": [ + 314, + 397, + 399, + 410 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 419, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 612 + ], + "type": "text", + "content": "Multimodal language models (MLMs). MLMs aim to solve a variety of tasks (e.g., Visual Question Answering (VQA) and captioning) based on vision and language inputs. Most modern architectures accomplish this by relying on either cross-attention [1, 2] or visual instruction tuning [26-28, 51] to interleave multimodal information. Cross-attention architectures operate by independently encoding images and then cross-attending to a language model backbone. On the other hand, visual instruction tuning produces token embeddings from image representations that can be interleaved with language tokens to ground generation. While these techniques work well for high level tasks, MLMs still struggle with mid-level and low-level tasks such as counting, depth reasoning, and segmentation. Most MLMs can be classified as either end-to-end MLMs or tool-using MLMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "End-to-end MLMs. End-to-end MLMs use a single, unified architecture [1, 8, 24, 26, 27] that can be repeatedly used for different tasks without requiring any architectural changes. While training end-to-end MLMs can be costly due to the need for large amounts of multi-task data, the use of diverse datasets allows end-to-end MLMs to generalize effectively and learn nuanced visual representations." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3837" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "Tool-using MLMs. Tool-using MLMs enable LLMs to perform vision tasks by attaching specialized vision modules such as segmentation or captioning networks [30, 33, 46, 52]. Tool-using MLMs use a router to select the optimal vision module to use for a given input. The use of specialized modules allows MLMs to achieve higher performances on many tasks. However, tool-using MLMs are sensitive to errors because they link together several networks, which are each a potential point of failure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 188, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 295, + 331 + ], + "type": "text", + "content": "Any-to-any networks. To address this, recent works [19, 20, 38, 45] such as Unified-IO [31, 32] have experimented with shared embedding spaces and visual decoders for vision and language tasks, training MLMs to generate segmentation masks, keypoints, and depth maps. Similarly, [47] generates special task tokens to route language model representations to diffusion heads for image, video, and audio generation. To handle complex segmentation scenarios, LISA [22] and GSVA [48] train language models to generate an additional segmentation token that can be used to produce segmentation masks, grounding the mask in language reasoning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 331, + 294, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 294, + 415 + ], + "type": "text", + "content": "While these approaches can generate visual outputs, they cannot reason over their own generations to solve related visual perception tasks. In contrast, augmenting MLMs with perception tokens as chain-of-thought tokens enables them to perform visual reasoning directly, providing significant gains in detail-oriented question-answering tasks such as depth estimation and counting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 427, + 218, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 218, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 218, + 441 + ], + "type": "text", + "content": "3. Perception Tokens & Aurora" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 447, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 519 + ], + "type": "text", + "content": "We introduce Perception Tokens and our Aurora training algorithm (see Fig. 3), which augment multimodal language models with perception tokens, enabling the MLM to leverage these tokens effectively during training and incorporate them into its chain-of-thought reasoning process for enhanced visual reasoning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 528, + 176, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 176, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 176, + 540 + ], + "type": "text", + "content": "3.1. Problem formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": "In autoregressive large language models, chain-of-thought (CoT) reasoning can be formulated as a multi-step inferential process in which the model iteratively generates intermediate reasoning steps to arrive at a final answer. Given a task input " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": ", the model generates a response " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " conditioned on the input and a sequence of " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " intermediate reasoning steps " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^m" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": ", and each " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " are sequences of tokens from the model's vocabulary " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Existing multimodal language models often rely on limited vocabulary tokens derived from text or pre-trained image embeddings like CLIP, restricting their capacity to interpret other representations crucial for reasoning. Mid-level and low-level visual features such as depth maps, instance segmentation masks, human poses and bounding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "boxes which could substantially improve visual reasoning, are currently incompatible with the model and cannot be integrated during training or inference. Our key insight is to introduce auxiliary perception tokens for these intermediate steps with an expanded vocabulary " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "V' = V \\cup V_{\\mathrm{aux}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ", bridging this gap by allowing the model to integrate richer visual representations into its reasoning process. Conditioning the final output on these tokens enhances the model's accuracy and interpretability across multimodal tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 188, + 535, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 535, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 535, + 201 + ], + "type": "text", + "content": "3.2. Perception token prediction and reasoning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 205, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 312 + ], + "type": "text", + "content": "Introducing an expanded vocabulary to enhance multimodal reasoning presents two main challenges. The first challenge is enabling the model to generate tokens from the new auxiliary vocabulary " + }, + { + "bbox": [ + 313, + 205, + 553, + 312 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{aux}}" + }, + { + "bbox": [ + 313, + 205, + 553, + 312 + ], + "type": "text", + "content": ", which includes specialized tokens for low- and mid-level visual features, such as depth maps and bounding boxes. The second challenge is ensuring that the model can effectively condition on these auxiliary tokens to improve reasoning, particularly for multi-step inference tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "text", + "content": "Perception token prediction. To address the first challenge, we employ a specialist-to-generalist distillation approach, using pre-trained specialist models (e.g., depth estimation or instance segmentation) to guide auxiliary token generation through cross-entropy loss. For each input " + }, + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "text", + "content": ", the specialist model provides a target probability distribution " + }, + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "text", + "content": " over its tokens. Let " + }, + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "inline_equation", + "content": "M: V_{\\mathrm{spec}} \\to V_{\\mathrm{aux}}" + }, + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "text", + "content": " denote a one-to-one mapping from the specialist model's vocabulary " + }, + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{spec}}" + }, + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "text", + "content": " to the auxiliary token vocabulary " + }, + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{aux}}" + }, + { + "bbox": [ + 313, + 319, + 553, + 437 + ], + "type": "text", + "content": ". We define the distillation loss as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 364, + 437, + 553, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 437, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 364, + 437, + 553, + 460 + ], + "type": "interline_equation", + "content": "\\ell_ {d i s t} = \\min _ {M} \\left(- \\sum_ {i} q _ {i} \\log p _ {M (i)}\\right), \\tag {1}", + "image_path": "8025a51d93fb4fa99b871abb79d68fe6a2fddd77594629b157facbc6d05d1024.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 465, + 553, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 553, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 553, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 465, + 553, + 537 + ], + "type": "inline_equation", + "content": "p_{M(i)}" + }, + { + "bbox": [ + 313, + 465, + 553, + 537 + ], + "type": "text", + "content": " is the probability assigned by our model to the auxiliary token corresponding to the mapped token " + }, + { + "bbox": [ + 313, + 465, + 553, + 537 + ], + "type": "inline_equation", + "content": "M(i)" + }, + { + "bbox": [ + 313, + 465, + 553, + 537 + ], + "type": "text", + "content": ". This consistent mapping allows the model to effectively align its predictions with the specialist model's output distribution, enhancing the relevance and accuracy of its auxiliary token predictions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 538, + 554, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 554, + 657 + ], + "type": "text", + "content": "In addition to distillation, we incorporate a reconstruction loss to enhance the token prediction and the interpretability of our model. Each set of auxiliary tokens corresponds to a specific representation, such as a depth map or a bounding box vector, and is trained to directly predict this feature. To achieve this, we introduced a lightweight decoder " + }, + { + "bbox": [ + 313, + 538, + 554, + 657 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 538, + 554, + 657 + ], + "type": "text", + "content": " that maps the tokens into the feature space, allowing for efficient and interpretable transformations. Formally, for a token set " + }, + { + "bbox": [ + 313, + 538, + 554, + 657 + ], + "type": "inline_equation", + "content": "T \\subset V_{\\mathrm{aux}}" + }, + { + "bbox": [ + 313, + 538, + 554, + 657 + ], + "type": "text", + "content": ", decoder " + }, + { + "bbox": [ + 313, + 538, + 554, + 657 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 538, + 554, + 657 + ], + "type": "text", + "content": ", and its target feature " + }, + { + "bbox": [ + 313, + 538, + 554, + 657 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 538, + 554, + 657 + ], + "type": "text", + "content": ", the reconstruction loss is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 390, + 666, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 666, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 390, + 666, + 553, + 681 + ], + "type": "interline_equation", + "content": "\\ell_ {r e c} = \\left\\| g (T) - f \\right\\| _ {2} ^ {2}, \\tag {2}", + "image_path": "7b6be7c84c84d99b58a33351af6160dccd650de42dea9429f4e5915d01658c55.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "g(T)" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " is the decoded representation for the token set " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " in the feature space. This reconstruction process not only" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3838" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 71, + 516, + 244 + ], + "blocks": [ + { + "bbox": [ + 95, + 71, + 516, + 244 + ], + "lines": [ + { + "bbox": [ + 95, + 71, + 516, + 244 + ], + "spans": [ + { + "bbox": [ + 95, + 71, + 516, + 244 + ], + "type": "image", + "image_path": "c15e4918570fa52681db82a6387e166441e0238746739e6b88dd82855da39622.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 251, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 555, + 285 + ], + "type": "text", + "content": "Figure 3. The overall AURORA training framework. We first learn visual perception tokens using VQVAE. We then finetune MLMs with a multi-task training approach where we distill intrinsic image representations (e.g., depth map) into MLMs by training them to decode the visual tokens as intermediate reasoning steps towards completing the tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 293, + 294, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 294, + 377 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 294, + 377 + ], + "type": "text", + "content": "aligns each token with a meaningful feature, improving interpretability, but also reinforces the accuracy of auxiliary token predictions through direct feature supervision. In practice, while the reconstruction objective improves performance and interpretability, it is optional, as models can be effectively trained with the distillation objective alone to reduce computational overhead." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "text", + "content": "Reasoning with perception tokens. The second challenge is enabling the model to condition on tokens from " + }, + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{aux}}" + }, + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "text", + "content": " effectively when generating each subsequent reasoning step, thereby enhancing its reasoning capabilities. To achieve this, we progressively introduce chain-of-thought reasoning during training, starting with simpler, single-step reasoning examples and gradually incorporating more complex, multi-step reasoning data. The model begins by learning single-step reasoning, predicting an initial reasoning step " + }, + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "text", + "content": " based on the input " + }, + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "text", + "content": ". It then progresses to multistep reasoning, predicting sequences " + }, + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "inline_equation", + "content": "s_1, \\dots, s_m" + }, + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "text", + "content": " and effectively utilizing auxiliary tokens to support extended chains of thought. To facilitate this, we propose a multi-task data synthesis approach combined with curriculum learning across a variety of synthetic tasks. Additional implementation details can be found in Section 3.4. Finally, we reinforce this process at inference time using constrained decoding, wherein sampling is selectively restricted to auxiliary tokens at certain reasoning steps, guiding the model to generate coherent intermediate steps from " + }, + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{aux}}" + }, + { + "bbox": [ + 55, + 387, + 296, + 626 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 635, + 138, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 138, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 138, + 647 + ], + "type": "text", + "content": "3.3. Tokenization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "A unified tokenization space is crucial for multimodal models as it creates a consistent framework through which varied visual tasks can be represented, processed, and interpreted. Inspired by [36], we establish a unified tokenization space which enables the model to learn varied visual fea" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 293, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 365 + ], + "type": "text", + "content": "tures within a shared representation seamlessly. In our experiments, we implement two tailored tokenization schemes for commonly used visual representations. Importantly, our framework is designed with flexibility, allowing it to generalize to a broad range of visual representations beyond those presented here." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 371, + 556, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 556, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 556, + 467 + ], + "type": "text", + "content": "Pixel-level representation. This tokenization scheme captures fine-grained spatial information, such as depth maps and segmentation masks, providing the model with detailed pixel-level data essential for accurate visual processing. For these types of tokens, we leverage visual tokenizers like VQVAE and VQGAN, which take in the ground truth masks or depth maps and return discrete target tokens [10, 31, 32, 44]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 473, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 568 + ], + "type": "text", + "content": "Structured representation. This scheme encodes structured yet abstracted visual features, such as human poses, bounding boxes, and coordinates, allowing the model to reason with higher-level spatial relationships and object hierarchies. For these tokens, we define the domain of the tokens based on specific properties; for example, the domain for coordinates can range from 0 to the maximum number of pixels in the image's height or width [5]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 576, + 541, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 541, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 541, + 589 + ], + "type": "text", + "content": "3.4. Curriculum learning with progressive CoTs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "The objective of training the model is to develop a data and computationally efficient method for learning to predict novel, fine-grained visual tokens and using them to complete complex visual reasoning tasks. We observed that the standard approach, which relies on a fixed mixture data, encounters a trade-off between the accuracy of novel tokens predictions and the model's reasoning capability, primarily due to catastrophic forgetting and challenges in reasoning with new tokens. Conversely, fine-tuning the model with the original training mixture significantly raises computational" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3839" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 544, + 210 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 544, + 210 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 544, + 210 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 544, + 210 + ], + "type": "table", + "html": "
ModelTrainingBLINK [11]2 PointsHardBLINK3 PointsHardBLINK4 PointsHardBLINK5 PointsAverage
DirectLa-labelingDataDepthGener-ationDataCoTData
LLaVA OneVision 72BXXX67.751.650.044.353.4
LLaVA 1.5 13BXXX54.035.537.92939.1
Fine-tuned LLaVA✓XX68.558.952.441.155.2
LLaVA-AURORA (Ours)✓✓✓64.566.960.554.861.6
GPT-4oXXX53.258.95036.349.6
GPT-4 TurboXXX58.154.841.932.246.7
GPT-4 Turbo + ToolXXX70.257.244.326.649.6
", + "image_path": "0c5bc260074474996613d20051b24dc291da2d4ea7c7c780cf71898cec683e39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 217, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 54, + 217, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 555, + 264 + ], + "type": "text", + "content": "Table 1. Performance comparison between our fine-tuned model, the baseline (without depth tokens), and the original base model on the relative depth accuracy (\\%) task. Results demonstrate that our approach, utilizing depth tokens and intermediate reasoning steps, significantly outperforms both the baseline and the base model, particularly on more challenging configurations with 3, 4, and 5 points sampled from the image's mid-height region. The gray rows represent closed-source models, providing additional baselines for comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 283, + 294, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 283, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 294, + 330 + ], + "type": "text", + "content": "costs and may be impractical if the original data, particularly proprietary datasets, is unavailable, making this approach less scalable for incorporating new tokens in the future." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "text", + "content": "We propose a curriculum learning-inspired training scheme that begins with atomic tasks and gradually advances to more complex ones requiring sophisticated, multi-hop reasoning. Let " + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "text", + "content": " represent the difficulty of task " + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "text", + "content": ", with difficulties " + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "inline_equation", + "content": "d_{1} < d_{2} < \\dots d_{T}" + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "text", + "content": " tasks, and let " + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "inline_equation", + "content": "p(t,s)" + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "text", + "content": " denote the probability of sampling data points from task " + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "text", + "content": " at training step " + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "inline_equation", + "content": "p(d_{t},s)" + }, + { + "bbox": [ + 54, + 331, + 295, + 427 + ], + "type": "text", + "content": " using a temperature-scaled Softmax formulation as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 103, + 435, + 295, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 435, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 103, + 435, + 295, + 464 + ], + "type": "interline_equation", + "content": "p (d _ {t}, s) = \\frac {\\exp \\left(- d _ {t} / \\tau (s)\\right)}{\\sum_ {i = 1} ^ {m} \\exp \\left(- d _ {i} / \\tau (s)\\right)}, \\tag {3}", + "image_path": "1bba3431374037d99f3b38b62b4dc26e4f875964d1f0777d552db3d1ad7fd25b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 470, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 507 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 470, + 295, + 507 + ], + "type": "inline_equation", + "content": "\\tau(s)" + }, + { + "bbox": [ + 55, + 470, + 295, + 507 + ], + "type": "text", + "content": " modulates the task difficulty over time, allowing a smooth shift in the probability distribution toward harder tasks. This temperature function is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 514, + 295, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 514, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 514, + 295, + 538 + ], + "type": "interline_equation", + "content": "\\tau (s) = \\frac {\\tau_ {0}}{1 + \\lambda \\cdot s / S}. \\tag {4}", + "image_path": "acd2cd245b0651bb169e330de81c86080d538f97cf34ee67606538625364ca99.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 546, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 570 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 546, + 295, + 570 + ], + "type": "inline_equation", + "content": "\\tau_0" + }, + { + "bbox": [ + 55, + 546, + 295, + 570 + ], + "type": "text", + "content": " is the initial temperature, " + }, + { + "bbox": [ + 55, + 546, + 295, + 570 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 546, + 295, + 570 + ], + "type": "text", + "content": " is the annealing rate, and " + }, + { + "bbox": [ + 55, + 546, + 295, + 570 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 546, + 295, + 570 + ], + "type": "text", + "content": " is the total number of training steps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "text", + "content": "Our approach for defining " + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "text", + "content": " values is based on the inherent complexity of each task, which corresponds to the depth of reasoning involved. Specifically, we assign " + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "d_{1}" + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "text", + "content": " to the most atomic task, involving the prediction of newly introduced tokens. At the other end of the spectrum, " + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "d_{m}" + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "text", + "content": " represents the final task, requiring the full chain-of-thought (CoT) reasoning steps. Between them, intermediate tasks " + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\{d_{t}\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "text", + "content": ", serves to bridge the gap between the atomic tasks and the comprehensive chain-of-thought responses." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "In this project, we introduce three types of data subsets for each downstream task, organized in increasing levels of difficulty. The first and most atomic task involves teaching" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 283, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 332 + ], + "type": "text", + "content": "the model to generate tokens from the new auxiliary vocabulary. For instance, in depth-related tasks, we train the model to learn depth maps; in segmentation-related tasks, we teach the model to generate masks; and so on." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 334, + 556, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 556, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 556, + 430 + ], + "type": "text", + "content": "The other two data subsets involve chain-of-thought (CoT) prompts and direct labeling which help with reasoning with auxiliary tokens. In the CoT subset, we use the new intermediate visual perception tokens to answer downstream-specific questions, encouraging the model to reason step by step. In the direct labeling subset, we pose the same questions but instruct the model to provide direct answers without step-by-step reasoning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 432, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 556, + 540 + ], + "type": "text", + "content": "Inspired by [15], we employ a multitasking approach for these two data subsets. For each image, we sequentially present both the CoT and direct labeling questions, allowing the model to tackle each image with both reasoning styles in sequence. We use a sequential sampler rather than a random sampler, shuffling the images beforehand. This strategy enables the model to learn from both types of reasoning tasks effectively, enhancing its ability to perform complex visual reasoning. More details in Sec. 9." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 559, + 395, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 395, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 395, + 572 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "We base our work on LLaVA 1.5 13B as the foundation for our model, which we refer to as LLaVA-AURORA. Our approach augments the MLM with perception tokens to enhance reasoning and improve performance across both 3D and 2D visual tasks. We evaluate our approach on relative depth estimation (3D) using pixel-level depth map tokens for fine-grained depth capture, and on object counting (2D) with mid-level bounding box tokens for precise localization. These tokens not only enhance task-specific results but also highlight our framework's potential to generalize effectively across a broad spectrum of visual reasoning tasks." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3840" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 163, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 163, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 163, + 85 + ], + "type": "text", + "content": "4.1. 3D reasoning task" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 222 + ], + "type": "text", + "content": "We choose relative depth estimation as our 3D task because it enables the model to determine spatial relationships within a scene by identifying which points are closer to or farther from the camera. This foundational skill is crucial for scene understanding and applications requiring spatial awareness, such as robotics and autonomous systems. Specifically, this task involves identifying the point closest to the camera among multiple marked points in an image. To support the model's reasoning, we use discrete depth map tokens that capture spatial depth information, enhancing the model's understanding of proximity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "content": "Tokenization. To capture fine-grained spatial details, we tokenize depth maps into sequences of discrete tokens. Inspired by the approach in AiT [36], we use a Vector Quantized Variational Autoencoder (VQVAE) [44] with a codebook size of 128. In this setup, each depth map is encoded as a grid of embeddings, with each embedding matched to the nearest entry in the codebook, yielding a compact depth representation. The VQVAE decoder reconstructs the depth map from this sequence of latent codes, and the entire model is optimized with a reconstruction loss to ensure precise encoding. Each " + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "inline_equation", + "content": "320 \\times 320" + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "content": " depth map is compressed into a " + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "inline_equation", + "content": "10 \\times 10" + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "content": " grid of code indices, resulting in a 100-token sequence where each token corresponds to one of the 128 codebook entries. We introduce 128 unique depth tokens, labeled " + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "inline_equation", + "content": "DEPTH\\_0" + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "inline_equation", + "content": "DEPTH\\_127" + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "content": ", and enclose each sequence with special tokens " + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "inline_equation", + "content": "DEPTH\\_START" + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "inline_equation", + "content": "DEPTH\\_END" + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "content": ", expanding the model's vocabulary by 130 depth-related tokens in total." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": "Training data. We train the VQVAE model on pseudodepth maps generated from the ADE20k dataset [54, 55] using the Depth Anything model [49, 50]. This dataset provides a diverse range of scenes, enhancing the model's ability to generalize." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "content": "For fine-tuning, we prepare three types of data tailored for relative depth estimation (as detailed in Sec. 3.4). First, we generate depth maps for " + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "inline_equation", + "content": "20\\mathrm{k}" + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "content": " ADE20k images, tokenize them with the pre-trained VQVAE, and format each sample in a Q&A structure, prompting the model with a depth estimation question and an answer sequence of depth tokens. Additionally, we construct a dataset of 500 ADE20k images for chain-of-thought (CoT) reasoning, with 2-5 markers in each image. Here, the prompt guides the model to generate the coordinates of the markers and then the depth map as intermediate steps, then identify the marker closest to the camera. This CoT training improves sequential reasoning and relative depth estimation accuracy. Finally, we use the same 500 images for direct labeling, prompting the model to directly identify and label the marker closest to the camera, see Sec. 9." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "This fine-tuning setup enables our model to effectively" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "handle step-by-step reasoning for the relative depth estimation task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 105, + 421, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 421, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 421, + 118 + ], + "type": "text", + "content": "4.2. 2D reasoning task" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 122, + 555, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 555, + 232 + ], + "type": "text", + "content": "We select counting as a critical 2D visual task. For object counting, we incorporate bounding box predictions as an intermediate reasoning step to improve accuracy in answering counting queries. Given an image and a question about the number of specific objects, the model first identifies and predicts bounding boxes for each instance of the target object. These bounding box tokens serve as structured, intermediate representations, enabling the model to understand spatial arrangements and accurately count object instances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 236, + 556, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 556, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 556, + 416 + ], + "type": "text", + "content": "Tokenization. To represent bounding boxes as discrete tokens, we resize all input images to a fixed resolution of 336x336 pixels. This preprocessing step allows us to add 336 unique tokens to the model's vocabulary, each representing a specific pixel position within the resized image. These tokens, labeled Pixel_0 to Pixel_335, enable the model to uniquely reference each pixel location. Bounding boxes are encoded as tuples of four tokens, formatted as (PIXEL_i, Pixel_j, Pixel_k, Pixel_m), where Pixel_i and Pixel_j denote the coordinates of the top-left corner and Pixel_k and Pixel_m represent the bottom-right corner (i.e., " + }, + { + "bbox": [ + 313, + 236, + 556, + 416 + ], + "type": "inline_equation", + "content": "(x_{1},y_{1},x_{2},y_{2})" + }, + { + "bbox": [ + 313, + 236, + 556, + 416 + ], + "type": "text", + "content": "). This discrete representation allows the model to interpret and use bounding box locations effectively, providing the spatial structure needed for accurate object counting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "text", + "content": "Training data. To fine-tune the model for object counting, we draw on three types of data tailored to this task (as detailed in Sec. 3.4). First, we use task-specific data from the LVIS dataset [13], selecting 5k images with objects whose counts range from 0 to 15. For each selected image, we specify an object type (e.g., \"beds\") and structure the fine-tuning samples to prompt the model for bounding box predictions of the specified objects within the image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 519, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 556, + 639 + ], + "type": "text", + "content": "To enhance the model's reasoning ability, we include a small subset of 250 LVIS images for chain-of-thought (CoT) training. Here, each question prompt encourages step-by-step reasoning by instructing the model to first generate bounding boxes for the target object, followed by providing the final count. Additionally, we create a direct labeling subset using the same 250 images. In this subset, we prompt the model to directly identify and label the total count of the specified objects without the intermediate step of bounding box generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 647, + 397, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 397, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 397, + 659 + ], + "type": "text", + "content": "4.3. Benchmarks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Relative depth. A recent benchmark, BLINK [11], introduces tasks designed to be intuitive for humans yet challenging for multimodal models, with relative depth estimation as one of its tasks. BLINK provides 124 images, each" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "3841" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 102, + 70, + 509, + 191 + ], + "blocks": [ + { + "bbox": [ + 102, + 70, + 509, + 191 + ], + "lines": [ + { + "bbox": [ + 102, + 70, + 509, + 191 + ], + "spans": [ + { + "bbox": [ + 102, + 70, + 509, + 191 + ], + "type": "table", + "html": "
ModelTrainingCV-Bench CountingSEED-Bench CountingBLINK Counting
Direct Labeling DataBounding Box DataCoT Data
LLaVA One VisionXXX34.431.735.8
LLaVA 1.5 13BXXX40.952.235.0
Fine-tuned LlaVA✓XX44.746.30.2
LLaVA-AURORA (Ours)✓✓✓56.054.645.8
GPT-4oXXX70.1864.647.5
GPT-4 TurboXXX61.364.857.5
GPT-4 Turbo + ToolXXX48.629.926.7
", + "image_path": "17ff087b4f382ecfc45ede1a42871bf92c0a735c664934a2f2d63ed0d82109d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 198, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 54, + 198, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 555, + 234 + ], + "type": "text", + "content": "Table 2. Comparison of object counting accuracy (%) across three benchmarks (CV-Bench, SEED-Bench, and BLINK). Our Aurora model, using auxiliary tokens to encode bounding box information for intermediate reasoning, demonstrates superior performance compared to the baseline models and the original base model. The gray rows represent closed-source models, providing additional baselines for comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 252, + 294, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 294, + 371 + ], + "type": "text", + "content": "containing two marked points labeled as " + }, + { + "bbox": [ + 54, + 252, + 294, + 371 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 252, + 294, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 252, + 294, + 371 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 54, + 252, + 294, + 371 + ], + "type": "text", + "content": ", and asks which point is closer to the camera. To reduce biases that language models have toward answering multiple-choice questions [39, 53], we modify the original BLINK questions by removing the answer choices. To further evaluate the model's reasoning and 3D understanding in relative depth, we curated a series of more challenging benchmark sets, collectively called HardBLINK. We progressively increase task difficulty by altering the prompts and image configurations as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 373, + 295, + 588 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 373, + 295, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 408 + ], + "type": "text", + "content": "1. Prompt Modification: In the prompts, we exclude the number of markers and their labels, requiring the model to infer these details solely from the image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "text", + "content": "2. Increased Point Complexity: We generate four variations of BLINK by adding more markers to each image, using Depth Anything to produce pseudo-depth maps for precise placement. These curated sets—HardBLINK 3points, HardBLINK 4points, and HardBLINK 5points—contain the same images as BLINK but with 3, 4, and 5 randomly placed markers, respectively, each with reasonable depth and distance differences. This setup tests the model's depth reasoning across more complex spatial configurations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 528, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 588 + ], + "type": "text", + "content": "3. Mitigating Height-Based Bias: To prevent the model from assuming that higher points are farther from the camera, we place markers at mid-height within the image. This approach encourages reliance on depth information rather than positional cues [6]." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": "Counting. For counting, we evaluate the model on CV-Bench [40], SEED-Bench [23], and BLINK's counting [11] subtask. To better capture the model's capabilities, we also remove multiple-choice options, requiring it to generate an exact count." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 660, + 122, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 122, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 122, + 672 + ], + "type": "text", + "content": "4.4. Baselines" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "We evaluate a diverse range of models, including closed-source models like GPT-4o [37] and GPT-4 Turbo [37], as well as state-of-the-art open-source models such as LLaVA" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 251, + 554, + 368 + ], + "blocks": [ + { + "bbox": [ + 316, + 251, + 554, + 368 + ], + "lines": [ + { + "bbox": [ + 316, + 251, + 554, + 368 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 554, + 368 + ], + "type": "image", + "image_path": "1d0f4997cca7b4f96551273234bc3befc64ac3a0623eedebeff7bae07b89c5c5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 375, + 555, + 398 + ], + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 398 + ], + "type": "text", + "content": "Figure 4. Depth maps generated by Aurora are imperfect but resemble the ground-truths from Depth Anything [49]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 420, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 555, + 516 + ], + "type": "text", + "content": "OneVision [25]. Another key baseline for our work involves fine-tuning the base model AURORA is applied on, solely on the direct labeling portion of the training data for each task, omitting the newly introduced tokens. This approach allows us to isolate the impact of our token-based enhancements. Additionally, we evaluate the base model for AURORA, LLaVA 1.5 13B, to assess its performance without task-specific adaptations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 517, + 556, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 556, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 556, + 625 + ], + "type": "text", + "content": "We also add tool-augmented baselines. For relative depth estimation, we employ GPT-4 Turbo, providing it with the ground truth depth maps generated by Depth Anything for each image, allowing it to use this information in its responses [17]. For counting, we also use GPT-4 Turbo in a tool-augmented setup. In this process, GPT-4 Turbo first identifies the object specified in the question, then uses Grounding DINO [29] to locate the bounding boxes for each instance of the object and finally counts them." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 635, + 539, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 539, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 539, + 647 + ], + "type": "text", + "content": "4.5. How do new tokens improve 3D reasoning?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "Our experiments demonstrate that incorporating new visual tokens significantly enhances the model's 3D reasoning abilities, specifically in the relative depth estimation task. Results in Tab. 1 show that our model outperforms both the primary baseline, which is fine-tuned solely on direct label-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3842" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 103, + 70, + 506, + 159 + ], + "blocks": [ + { + "bbox": [ + 103, + 70, + 506, + 159 + ], + "lines": [ + { + "bbox": [ + 103, + 70, + 506, + 159 + ], + "spans": [ + { + "bbox": [ + 103, + 70, + 506, + 159 + ], + "type": "table", + "html": "
ModelBLINK [11] 2 pointsHardBLINK 3 PointsHardBLINK 4 PointsHardBLINK 5 PointsAverage
VQGAN [10] (16384 codes)82.266.1533759.6
Unified-IO70.275.875.875.874.4
Unified-IO25437.92127.435.1
AURORA (Ours)91.978.271.775.879.2
Our VQVAE (128 codes)96.794.395.296.795.7
", + "image_path": "a502d8ea345d881c5862aaa42cce7338e8cd453c0114c0ba27fa3ae8201b57fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 166, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 166, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 555, + 222 + ], + "type": "text", + "content": "Table 3. While not the main aim of our work, we report the depth generation performance across benchmarks with 2, 3, 4, and 5 marked points using BLINK [11]'s relative depth subtask images. We report relative depth estimation accuracy " + }, + { + "bbox": [ + 55, + 166, + 555, + 222 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 166, + 555, + 222 + ], + "type": "text", + "content": ", calculated by programmatically extracting depth values at specific coordinates from model-generated depth maps. Our model consistently outperforms other multimodal models, including Unified-IO [31] and Unified-IO 2 [32]. The gray row represents our VQVAE model, which serves as the upper bound for performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 243, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 243, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 294, + 266 + ], + "type": "text", + "content": "ing data, and the original base model, indicating that the added tokens contribute meaningfully to task accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 267, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 295, + 397 + ], + "type": "text", + "content": "As task complexity increases—for instance, when additional markers are introduced in the HardBLINK benchmarks—our model's performance advantage becomes even more pronounced. Not only does it maintain high accuracy in distinguishing depth relationships, but it also surpasses both advanced closed-source models, such as GPT-4o, GPT-4 Turbo, and GPT-4 Turbo + Tool that use ground truth depth maps. This suggests that the new tokens enable our model to develop a more nuanced understanding of depth cues, even in cases where complex spatial reasoning is required." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 295, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 446 + ], + "type": "text", + "content": "Overall, these results highlight the effectiveness of our token-based approach in enhancing 3D reasoning, allowing the model to handle increasingly difficult tasks with robust performance and accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 454, + 280, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 280, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 280, + 468 + ], + "type": "text", + "content": "4.6. How do new tokens improve 2D reasoning?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 567 + ], + "type": "text", + "content": "In the 2D task of object counting, incorporating new visual tokens provides a significant advantage over the primary baselines. As shown in Tab. 2, our model outperforms both the baseline fine-tuned solely on direct labeling data and the original base model, as well as the state-of-the-art open-source model LLaVA-OneVision and GPT-4 Turbo + Tool, underscoring the value of these tokens in enhancing counting accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 568, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 639 + ], + "type": "text", + "content": "Although our model does not yet surpass advanced closed-source models, the results demonstrate that the new tokens yield a meaningful improvement in 2D reasoning, enabling more reliable and accurate object counting compared to standard fine-tuning approaches and open-source alternatives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 203, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 203, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 203, + 661 + ], + "type": "text", + "content": "4.7. Perception token decoding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Our approach enables the decoding of learned perception tokens into specialist features, such as depth maps and object bounding boxes, to assess their fidelity and utility. We, in particular, evaluate the accuracy and correctness" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 243, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 530 + ], + "type": "text", + "content": "of the depth maps generated by LLaVA-AURORA, specifically focusing on its ability to represent spatial relationships in visual scenes. For this evaluation, we use the relative depth images from the BLINK benchmark [11], ensuring consistency with our relative depth assessments. LLaVA-AURORA generates depth tokens from BLINK images, which are then reconstructed into full depth maps via the decoder of our pre-trained VQVAE model. Notably, the depth maps output by our pre-trained VQVAE provide an upper bound on the quality of the depth maps generated by LLaVA-AURORA. We use programmatic relative depth accuracy as our metric, which measures how well the reconstructed depth maps capture the relative depth of marked points. This evaluation spans several configurations in the benchmark, including images with 2, 3, 4, and 5 labeled points. By comparing the model's predicted depth with the ground-truth marker coordinates, we calculate relative depth accuracy, allowing us to assess the precision of depth map generation in reflecting spatial depth relationships. As shown in Tab. 3, LLaVA-AURORA outperforms Unified-IO [31, 32] in this task. Furthermore, qualitative analysis in Fig. 4 reveals that LLaVA-AURORA's depth maps capture spatial details effectively highlighting its capacity to interpret and represent fine-grained depth information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 540, + 388, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 388, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 388, + 553 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 560, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 555, + 693 + ], + "type": "text", + "content": "Our algorithm enables the lightweight and scalable integration of perception tokens, such as depth maps and bounding box coordinates, into MLMs, allowing them to perform intermediate reasoning steps akin to chain-of-thought processes. Our method achieves state-of-the-art results on challenging tasks like 2D object counting and 3D relative depth estimation. It also enhances model generalization and interpretability without relying on external tools or task-specific fine-tuning. The framework is inherently adaptable, incorporating new perception tokens as they emerge, making it a future-proof solution for advancing multimodal reasoning." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3843" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 199 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 202, + 294, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 202, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 61, + 202, + 294, + 246 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 248, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 248, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 248, + 294, + 280 + ], + "type": "text", + "content": "[3] Luca Beurer-Kellner, Marc Fischer, and Martin Vechev. Guiding llms the right way: Fast, non-invasive constrained generation. arXiv preprint arXiv:2403.06988, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 282, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 282, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 294, + 346 + ], + "type": "text", + "content": "[4] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 350, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 294, + 382 + ], + "type": "text", + "content": "[5] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 384, + 294, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 294, + 417 + ], + "type": "text", + "content": "[6] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 419, + 294, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 294, + 440 + ], + "type": "text", + "content": "[7] Aidan Cooper. A guide to structured outputs using constrained decoding, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 294, + 495 + ], + "type": "text", + "content": "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 498, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 294, + 541 + ], + "type": "text", + "content": "[9] PHILOSO EPHY DO CT OR OF. MACHINE PERCEPTION OF THREE-DIMENSIONAL, SO LIDS. PhD thesis, MASSACHUSETTS INSTITUTE OF TECHNOLOGY, 1961." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "type": "text", + "content": "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 643 + ], + "type": "text", + "content": "[11] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 646, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 689 + ], + "type": "text", + "content": "[12] Saibo Geng, Martin Josifoski, Maxime Peyrard, and Robert West. Grammar-constrained decoding for structured nlp tasks without finetuning. arXiv preprint arXiv:2305.13971, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Pro" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "type": "text", + "content": "[14] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In *Alvey vision conference*, pages 10-5244. CiteSeer, 1988." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "type": "text", + "content": "[15] Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. arXiv preprint arXiv:2305.02301, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 195, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 238 + ], + "type": "text", + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 553, + 293 + ], + "type": "text", + "content": "[17] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 294, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 360 + ], + "type": "text", + "content": "[18] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9601, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "text", + "content": "[19] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu, Chengru Song, Xiaoqiang Lei, Di Zhang, Wenwu Ou, Kun Gai, and Yadong Mu. Unified language-vision pretraining in llm with dynamic discrete visual tokenization, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 415, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 553, + 437 + ], + "type": "text", + "content": "[20] Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language models, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 438, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 502 + ], + "type": "text", + "content": "[21] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 504, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 536 + ], + "type": "text", + "content": "[22] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 537, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 579 + ], + "type": "text", + "content": "[23] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 581, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 581, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 581, + 553, + 613 + ], + "type": "text", + "content": "[24] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimic-it: Multi-modal in-context instruction tuning, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "text", + "content": "[25] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "type": "text", + "content": "[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "text", + "content": "[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3844" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[29] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "text", + "content": "[31] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "text", + "content": "[32] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision, language, audio, and action, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[33] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 374 + ], + "type": "text", + "content": "[34] David Marr. Vision: A computational investigation into the human representation and processing of visual information. MIT press, 2010." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "text", + "content": "[35] Marvin Minsky and Seymour Papert. An introduction to computational geometry. Cambridge, 1947: 104, 1969." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[36] Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Unifying output space of visual tasks via soft token. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19900-19910, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 487 + ], + "type": "text", + "content": "[37] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 522 + ], + "type": "text", + "content": "[38] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-g: Generating images in context with multimodal large language models, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 524, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 294, + 556 + ], + "type": "text", + "content": "[39] Pouya Pezeshkpour and Estevam Hruschka. Large language models sensitivity to the order of options in multiple-choice questions. arXiv preprint arXiv:2308.11483, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 558, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 294, + 621 + ], + "type": "text", + "content": "[40] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "text", + "content": "[41] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[42] Antonio Torralba and Aude Oliva. Depth estimation from image structure. IEEE Transactions on pattern analysis and machine intelligence, 24(9):1226-1238, 2002." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 621 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[43] Vivien Tran-Thien. Fast, high-fidelity llm decoding with regex constraints, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[44] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[45] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "text", + "content": "[46] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 241 + ], + "type": "text", + "content": "[47] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "type": "text", + "content": "[48] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 277, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 555, + 331 + ], + "type": "text", + "content": "[49] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 332, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 365 + ], + "type": "text", + "content": "[50] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 366, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 431 + ], + "type": "text", + "content": "[51] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl: Modularization empowers large language models with multimodality, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 477 + ], + "type": "text", + "content": "[52] Haoxuan You, Rui Sun, Zhecan Wang, Long Chen, Gengyu Wang, Hammad A. Ayyubi, Kai-Wei Chang, and Shih-Fu Chang. Idealgpt: Iteratively decomposing vision and language reasoning via large language models, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[53] Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. Large language models are not robust multiple choice selectors. In The Twelfth International Conference on Learning Representations, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 523, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 576 + ], + "type": "text", + "content": "[54] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 578, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 621 + ], + "type": "text", + "content": "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3845" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/cdaa1dd9-bb97-42b0-9c82-a5fe76697ec3_content_list.json b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/cdaa1dd9-bb97-42b0-9c82-a5fe76697ec3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..39cfbb47c01f3a770e13b282e42a72ef09dd70d3 --- /dev/null +++ b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/cdaa1dd9-bb97-42b0-9c82-a5fe76697ec3_content_list.json @@ -0,0 +1,1198 @@ +[ + { + "type": "text", + "text": "Perceptual Inductive Bias Is What You Need Before Contrastive Learning", + "text_level": 1, + "bbox": [ + 127, + 130, + 870, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junru Zhao Tianqin Li Dunhan Jiang Shenghao Wu Alan Ramirez Tai Sing Lee Carnegie Mellon University", + "bbox": [ + 135, + 180, + 861, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{junruz, tianqinl, dunhanj, shenghaw, alanrami, taislee}@andrew.cmu.edu", + "bbox": [ + 181, + 224, + 810, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 275, + 326, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "David Marr's seminal theory of human perception stipulates that visual processing is a multi-stage process, prioritizing the derivation of boundary and surface properties before forming semantic object representations. In contrast, contrastive representation learning frameworks typically bypass this explicit multi-stage approach, defining their objective as the direct learning of a semantic representation space for objects. While effective in general contexts, this approach sacrifices the inductive biases of vision, leading to slower convergence speed and learning shortcut resulting in texture bias. In this work, we demonstrate that leveraging Marr's multi-stage theory—by first constructing boundary and surface-level representations using perceptual constructs from early visual processing stages and subsequently training for object semantics—leads to $2x$ faster convergence on ResNet18, improved final representations on semantic segmentation, depth estimation, and object recognition, and enhanced robustness and out-of-distribution capability. Together, we propose a pretraining stage before the general contrastive representation pretraining to further enhance the final representation quality and reduce the overall convergence time via inductive bias from human vision systems.", + "bbox": [ + 89, + 311, + 483, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 691, + 220, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human visual information processing is organized in stages, progressing from contrast and edge detection to perceptual organization and figure-ground segmentation, then to computing surfaces and shapes, and ultimately to object recognition. End-to-end supervised learning in neural networks for object classification in images often relies on the simplest features or shortcuts, bypassing essential intermediate representations, resulting in texture bias. Self-supervised contrastive learning, on the other hand, leverages the inductive bias of semantic segregation by grouping semantically similar scenes and separating dissimilar scenes in latent space. This also leads to representations that ac", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41f7a0da2d88bc88904bedce87effa10ba803a5d7fd8f9f2fb3bd5e71c91eda3.jpg", + "image_caption": [ + "Figure 1. We propose to leverage perceptual constructs from early vision processing stages to improve the contrastive learning pipeline. By introducing a pre-pretraining stage before general contrastive learning (CL), we inject inductive bias into the contrastive representation learning. We demonstrate that such inductive bias is beneficial to the representation quality and provides significantly convergence speed-up the contrastive learning." + ], + "image_footnote": [], + "bbox": [ + 544, + 273, + 877, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cess rich semantic information without building intermediate representations sensitive to surfaces and boundaries first.", + "bbox": [ + 511, + 566, + 906, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we evaluate whether perceptual constructs from early visual processing, such as figure-ground segmentation and intrinsic image decomposition, can serve as additional inductive biases for contrastive learning. Evidence of figure-ground organization has been found in the primary visual cortex and secondary visual cortex in primate visual systems [19, 22, 34, 35]. This early perceptual construct distinguishes figures against the ground, assigns border ownership to surfaces, and enables the construction of a neural representation of local shapes and global shapes along the ventral visual hierarchy. The perception of shapes could be innate or at least develop very fast [8, 29]. Furthermore, categorization and identification of unseen objects are considered to rely on the similarity of the overall shape, conceptualized through shape prototypes [6]. The recognition of shapes was found to be parallel to the development of vocabularies in language in infancy. Infants can learn new nouns faster when nonce objects are organized in shape, implicating a linked development between labeling objects linguistically and shape prototypes in the visual", + "bbox": [ + 511, + 599, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9621", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "system [14]. We conjecture that shape prototypes learned as cluster means of shape silhouettes obtained from figure-ground segregation can allow contrastive learning to learn better intermediate representations for a variety of visual tasks, such as object recognition, segmentation as well as depth estimation in the context of this paper.", + "bbox": [ + 93, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another set of early perceptual constructs is called intrinsic images [2]. Intrinsic image decomposition [1, 15] is a generalization of Retinex developed by Edwin Land [20] to explain color constancy by discounting illumination effects to infer the reflectance or color of the objects. Intrinsic image decomposition factorizes an image as a product of a reflectance image and a shading image. Reflectance images reflect surface material properties and govern our perception of color and lightness. Shading images reflect the interaction of illumination and the 3D surface shape and are thus rich in implicit 2.5D information and affect our depth perception. What kinds of visual tasks would benefit from intermediate representations learned under the guidance of the inductive bias of these intrinsic images?", + "bbox": [ + 93, + 181, + 480, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work shows that global shapes of figures, particularly through a set of remembered prototypes, provide an inductive bias that enhances contrastive learning, resulting in learned representations that improve object recognition, segmentation, and depth estimation, with an increase in the shape bias of the model. We find that the inductive bias from the shape prototypes increases the speed of representational learning at the beginning, as in the early development of human infants, but ultimately has to be disabled to allow the learning of more refined representations to accommodate a greater variety of semantics. Thus, we propose a hybrid coarse-to-fine inductive bias that allows contrastive learning to learn intermediate representations first with the help of the prototypes and then refine the representation based on semantics. This results in a 2x faster representation learning. On the other hand, intrinsic reflectance images enhance the learned representations for segmentation and object recognition but do not benefit depth estimation. Conversely, intrinsic shading images improve depth estimation capabilities but with little impact on segmentation and object recognition. Finally, we demonstrate that integrating these three perceptual constructs provides additional performance benefits. Thus, we demonstrate that inductive biases derived from early perceptual constructs can differentially or collectively promote the contrastive learning of intermediate representations with greater sensitivity to surface depths and object shapes and better performance in downstream visual tasks.", + "bbox": [ + 93, + 393, + 480, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 94, + 830, + 236, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive learning is one of the most commonly used frameworks in discriminative self-supervised learning. It inherently aligns with the Infomax principle [25] by aim-", + "bbox": [ + 94, + 857, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing to maximize the mutual information between semantically similar scenes, which are usually different views of the same image. This aim is realized by maximizing InfoNCE loss function [28, 32]. InfoNCE is designed to maximize a lower bound on the mutual information between representations of positive view pairs while distinguishing them from negative view pairs. The views are usually generated by different data augmentations. Many studies [4, 17] emphasize the importance of data augmentations as contrastive learning requires intensive data augmentation to learn good semantic representations. Some works combine contrastive learning with clustering [3, 23]. In this kind of framework, the concept of prototype is introduced to organize the representation space in order to have more compact semantic clusters.", + "bbox": [ + 516, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After self-supervised pretraining, the representations are commonly evaluated by downstream tasks such as classification, segmentation and depth estimation. Many works [27, 30, 33] point out that there exists a trade-off between image-level classification (semantics) and pixel-wise classification or segmentation (perceptual grouping) in discriminative models. In other words, representations learned by state-of-the-art discriminative models often fail to perform effectively across both high-level semantic tasks and intermediate-level perceptual tasks.", + "bbox": [ + 516, + 318, + 903, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While humans rely heavily on intermediate perceptual constructs like object shapes in object recognition, texture bias is a well-known misalignment between CNN-based models and humans [10, 11, 13]. Many works endeavor to solve this misalignment by different approaches including 1) using data augmentation methods to vary texture, e.g. texture synthesis, style transfer [12, 13, 31], 2) changing to vision transformer [26], or 3) using more complex multimodal generative frameworks [9, 18]. Improved shape bias results in improvement of recognition robustness and domain generalization [13, 24]. However, a recent study [18] argues that most discriminative models are still strongly texture-biased and less robust to tremendous style changes. We attribute this phenomenon to the focus of discriminative models on learning semantic representations while overlooking intermediate-level representations.", + "bbox": [ + 516, + 469, + 903, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 516, + 722, + 619, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this part, we will introduce how we incorporate global shape of figures in a prototypical manner and the intrinsic images as views into contrastive learning.", + "bbox": [ + 516, + 750, + 903, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Shape prototypical contrastive learning (S-PCL) Inspired by the fact that humans rely on the similarity of the overall shape envelope for object recognition, we design a contrastive learning framework that is capable of learning the concept of a group of similar shapes, i.e. shape prototypes. To represent the global shape of an object,", + "bbox": [ + 516, + 811, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9622", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c5affb8a114f46e6a133aa6e0d18d5b1b20d813b35cdce9b51de366d171f4f91.jpg", + "image_caption": [ + "Figure 2. Illustration of Shape Prototypical Contrastive Learning (S-PCL) training framework The original images are input to the encoder and the offline-generated shape silhouettes generated are input to the momentum encoder. The K-Means is performed on the embedding of shape silhouettes. Similar shapes are clustered together and the centroid of each cluster is considered to be the shape prototype. The objective function, ShapeProtoNCE, is calculated on the representation of the original image and its corresponding shape prototype." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 467, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we use the figure-ground shape silhouette as presented in Figure.2. Our framework is built upon prototypical contrastive learning (PCL)[23], which involves an online k-means clustering on the representation of the entire training data and considers the centroids of clusters as prototypes. In shape prototypical contrastive learning, the clustering is performed on the representation of shape silhouettes, as illustrated in Figure.2. The representation learning is guided by the clusters of shapes. With a set of images $X = \\{x_{1}, x_{2}, \\dots, x_{n}\\}$ and matched shape silhouettes $M = \\{m_{1}, m_{2}, \\dots, m_{n}\\}$ , they are encoded by the same encoder to get their representations $V = \\{v_{1}, v_{2}, \\dots, v_{n}\\}$ and $U = \\{u_{1}, u_{2}, \\dots, u_{n}\\}$ , respectively. K-Means clustering is employed on the set $U$ , producing $K$ clusters and a set of centroids, $S = \\{s_{1}, s_{2}, \\dots, s_{K}\\}$ . The centroid of the cluster is representative of elements in the cluster, and thus is", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4ab995d317b15ef71b93f419b696001117691500262389085bda80f9530984c4.jpg", + "image_caption": [ + "Figure 3. Illustration of the training framework that incorporates the intrinsic image as a view The original image is input to the encoder and one intrinsic image decomposed by Retinex algorithm is input to the momentum encoder. The InfoNCE between the representation of the original image and that of the intrinsic image is computed." + ], + "image_footnote": [], + "bbox": [ + 544, + 90, + 885, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "considered as the shape prototype. The mutual information between the representation of the original image $v_{i}$ and the representation of the shape prototype $s_{i}$ is maximized by optimizing the InfoNCE objective function:", + "bbox": [ + 511, + 401, + 906, + 462 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S h a p e P r o t o N C E}} = \\sum_ {i = 1} ^ {n} - \\log \\frac {\\exp \\left(v _ {i} \\cdot s _ {p}\\right) / \\phi_ {p})}{\\sum_ {j = 0} ^ {r} \\exp \\left(v _ {i} \\cdot s _ {j} / \\phi_ {j}\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 473, + 906, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s_p$ is the prototype of the cluster $p$ to which $u_i$ belongs, $s_j, j \\neq p$ are shape prototypes representing other shape clusters, and $\\phi_j$ is the concentration estimation of cluster $j$ . Similar to PCL, we perform clustering for $N$ times with different $K_i, i = 1,2,\\dots,N$ and average the losses. The final loss function used is the combination of InfoNCE between the normally augmented view and the shape loss:", + "bbox": [ + 511, + 523, + 908, + 645 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {I n f o N C E}} + \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\text {S h a p e P r o t o N C E}, K _ {i}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 657, + 906, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intrinsic image as a view in contrastive learning Unlike shape prototypes, we treat the intrinsic image, either reflectance or shading, as an augmented view. We maximize the mutual information between the representations of normally augmented views and those of reflectance or shading images by optimizing the InfoNCE objective function, as illustrated in Figure 3. The final loss function is the sum of the standard InfoNCE loss between representations of two augmented views and the InfoNCE loss between representations of one augmented view and one intrinsic image. The approaches that incorporate reflectance or shading intrinsic images are termed Reflectance Contrastive Learn", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9623", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing (RefCL) and Shading Contrastive Learning (ShadCL), respectively.", + "bbox": [ + 89, + 90, + 480, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combination of shape and intrinsic images We also integrate the shape and the intrinsic image together according to equation (3). We call the combined approach mid-level vision contrastive learning (MidVCL).", + "bbox": [ + 89, + 138, + 482, + 202 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {I n f o N C E}} + \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\text {S h a p e P r o t o N C E}, K _ {i}} + \\alpha \\mathcal {L} _ {\\text {S h a d}} + \\beta \\mathcal {L} _ {\\text {R e f l}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 212, + 482, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ is the weight of the shading and reflectance losses, respectively.", + "bbox": [ + 89, + 266, + 482, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation details The experiments are conducted on the dataset ImageNet-100 $^{1}$ , a subset of ImageNet-1K [7], and STL-10 [5]. The silhouette mask of the object is generated offline using TRACER [21]. We decompose a full-detail image to a reflectance intrinsic image and a shading image by Retinex algorithm [1, 15]. The ResNet18 [16] is adopted as the encoder and the output dimension of the last fully-connected layer is 256-D. We follow the data augmentation methods in previous works [4, 17, 23]. These data augmentations are applied to only the original image and the preprocessing for the shape silhouette and the intrinsic image only contains resizing and normalization. More details can be found in the supplementary material.", + "bbox": [ + 89, + 311, + 483, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 89, + 523, + 176, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Representation Evaluation", + "text_level": 1, + "bbox": [ + 89, + 547, + 331, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First and foremost, we evaluate representation learned under the guidance of intermediate perceptual constructs through multiple downstream visual tasks, including linear classification, segmentation, and depth estimation.", + "bbox": [ + 89, + 569, + 483, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Linear classification and transfer learning to ImageNet-1k We trained a linear classifier on frozen representations using the entire labeled training set of ImageNet-100 and STL-10, and here we report the Top-1 and Top-5 accuracies of ImageNet-1k on the validation sets. The results on STL-10 are included in the supplementary material. As illustrated in Table.1, S-PCL significantly outperforms the baselines during the early training stages (e.g., at 100 epochs) on both datasets. However, when training is extended to 400 epochs, the performance gains of S-PCL diminish, and it is even surpassed by the baseline approaches significantly on ImageNet-100, indicating a potential faster saturation of S-PCL compared to others. In contrast, RefCL maintains consistent performance improvements between 100 and 400 epochs, suggesting that reflectance information effectively", + "bbox": [ + 89, + 638, + 483, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "facilitates object recognition. Conversely, ShadCL, with inductive bias from shading, provides only a tiny gain during the early stages of training and may hinder object recognition when trained for a large number of epochs. MidVCL is similar to S-PCL on the classification task, being the best at epoch 100 while plateauing when the number of epoch increases to 400. We also test generalizability by transferring to a different dataset. This is performed by the linear classification on ImageNet-1k which has 10x more classes than ImageNet-100. The transfer learning results reported in Table.1 display the similar trend as the linear classification.", + "bbox": [ + 511, + 90, + 906, + 270 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/801c8ccf1f96c52278bd2d3aa523bc28dc2bd941a38b95fb0992972c45d73d56.jpg", + "table_caption": [ + "Table 1. Linear classification and ImageNet-1k transfer learning accuracies (\\%) of different approaches trained for 100 and 400 epochs on the ImageNet-100. For hybrid methods (denoted as \"X+Y\"), training consists of 100 epochs using method X followed by 300 epochs using method Y. The best-performing results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
EpochsImageNet-100ImageNet-1k
Top-1 Acc (%)↑Top-5 Acc (%)↑Top-1 Acc (%)↑Top-5 Acc (%)↑
SimCLR10065.187.333.158.2
MoCoV210061.786.031.155.1
PCL10062.186.330.554.7
SwAV10064.888.333.457.8
BYOL10065.888.533.055.4
DINO10065.388.333.255.5
S-PCL(Ours)10070.290.337.261.9
RefCL(Ours)10065.588.135.260.1
ShadCL(Ours)10062.586.532.155.9
MidVCL(Ours)10070.590.437.062.7
SimCLR40077.293.840.866.9
MoCoV240077.093.841.668.1
PCL40074.091.937.863.3
SwAV40076.893.638.364.2
BYOL40075.892.642.967.6
DINO40077.193.638.062.9
S-PCL(Ours)40071.890.937.063.1
RefCL(Ours)40077.293.440.065.4
ShadCL(Ours)40071.890.934.960.6
MidVCL(Ours)40072.490.937.161.7
S-PCL+MoCoV2(Ours)40078.094.243.969.4
MidVCL+MoCoV2(Ours)40077.894.243.869.4
", + "bbox": [ + 516, + 393, + 931, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Segmentaion and Depth Estimation Besides linear classification, which is a semantic and image-level task, it is also essential to evaluate representations on pixel-level tasks such as segmentation and depth estimation. Therefore, we conduct experiments on various benchmark datasets to assess performance in those tasks. The models evaluated in this part are pretrained for 400 epochs.", + "bbox": [ + 511, + 718, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For segmentation tasks, we fine-tune the pretrained models with a segmentation head on the ADE20K and Cityscapes datasets. The mean Intersection over Union (mIoU) scores are reported in Table 2. Our MidVCL approach outperforms all other methods on both datasets, and", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1The dataset used here is reorganized by adjusting the train-validation split to a 7:3 ratio to intensify the challenge", + "bbox": [ + 89, + 875, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9624", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0458a01f5516eaf52afccaf0fe1517ecb0fc644f32038ec808a96a4b5b167404.jpg", + "image_caption": [ + "Figure 4. Visualization of elements in clusters produced by S-PCL after trained on STL10 for 400 epochs. Each block represents a cluster and 10 randomly selected images (Top) and shape silhouettes (bottom) are presented. Images in each cluster have similar shape irrespective to their ground truth category, indicating the S-PCL's capability of grouping objects with similar shape." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 897, + 348 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/168d5ff2fd24f43901d3c626fa726d9125c97bb1d16c8b955c33113880ca187d.jpg", + "image_caption": [ + "Figure 5. A. AMI between cluster assignment of S-PCL trained on ImageNet-100 and ground truth label. The AMI increases at the early stage and decreases later on. B. AMI between cluster assignment of PCL and S-PCl+PCL trained on ImageNet-100 and ground truth label. The AMI of PCL keeps increasing. S-PCL warm-up moves the curve upward. C and D. Linear classification accuracy on two datasets for PCL, S-PCL and S-PCL+PCL. S-PCL suffers rapid saturation but is able to serve as a warm-up strategy to accelerate training." + ], + "image_footnote": [], + "bbox": [ + 91, + 416, + 295, + 545 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/02f3f563a7ec45b971cb8a2dd1e61e872cdf5869f05c9420eba463736d0b22c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 416, + 500, + 545 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/07cf7d1ea6bcc49da8d94e76a5576333b6d0190c7dba9d98aecc618641ff3602.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 416, + 699, + 545 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6946beee34b2397b72c01591990984866518198502855e7f1b229e3432e8258e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 416, + 903, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the S-PCL is on the second place. This superior performance is attributable to the incorporation of shape silhouette masks in both MidVCL and S-PCL, which provide figure-ground information critical for segmentation tasks. Additionally, we observe that ReflCL ranks the third, presumably because the reflectance intrinsic image segregates instances based on albedo, thereby facilitating more accurate boundary detection. Also, it turns out that ReflCL performs poorly on depth estimation, suggesting that solely reflectance is insufficient to distinguish foreground and background although the boundaries of instances are clearer. In contrast, ShadCL ends up with only nuanced improvements in segmentation performance.", + "bbox": [ + 88, + 638, + 485, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For depth estimation, we fine-tune the pretrained models with an additional head using the NYU Depth V2 dataset. The relative errors (REL) and root mean errors (RME) are presented in Table 3. MidVCL and S-PCL continue", + "bbox": [ + 88, + 839, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to achieve the highest performance among all evaluated approaches. Notably, ShadCL ranks third, indicating that shading information is beneficial for depth estimation tasks. However, the advantages of incorporating reflectance information in this context appear to be limited.", + "bbox": [ + 511, + 638, + 906, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall Our evaluations on these downstream tasks demonstrate that shape prototypes, reflectance intrinsic image, and shading intrinsic image facilitate representation learning in different ways. The shape prototype helps all three downstream tasks, especially superior in pixel-level tasks (segmentation and depth estimation). However, it is only beneficial in the early stage for the image-level classification task and saturates rapidly. Additionally, the reflectance intrinsic image helps classification and segmentation but not depth estimation. On the other hand, the shading intrinsic image enhances depth estimation thanks to its implicit information of 3D shape and illuminance, but not classification and seg-", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mentation, where spatial information is not that important. Among three perceptual constructs, we find the shape prototype yields the most interesting results, thus we conduct further studies to understand the underlying mechanism of the shape prototype.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/04e6f368761e6a6dff7912c8888c24d29f2cdfa505449a06e6efd273f5563ae0.jpg", + "table_caption": [ + "Table 2. Segmentation results. The mIoU(%) values on ADE20K and Cityscapes are reported. The best performances are in bold." + ], + "table_footnote": [], + "table_body": "
EpochsADE20K mIoU(%)↑Cityscapes mIoU(%)↑
SimCLR40030.463.4
MoCoV240030.463.1
PCL40030.564.1
SwAV40030.261.9
BYOL40030.961.6
DINO40030.261.0
S-PCL(Ours)40031.868.3
RefCL(Ours)40031.865.5
ShadCL(Ours)40031.064.4
MidVCL(Ours)40032.168.3
", + "bbox": [ + 129, + 233, + 446, + 402 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dda5e8da77f4a31fb1a659ba73c04fc6d216ca428b2462b736b7eaa3871b5ec6.jpg", + "table_caption": [ + "Table 3. Depth estimation results. The relative error and root mean error on NYU Depth V2 dataset are reported. The best performances are in bold." + ], + "table_footnote": [], + "table_body": "
EpochsREL↓RME↓
SimCLR4000.25450.1420
MoCoV24000.25500.1434
PCL4000.25870.1437
SwAV4000.25860.1459
BYOL4000.25430.1458
DINO4000.25950.1451
S-PCL(Ours)4000.24280.1381
RefCL(Ours)4000.25330.1420
ShadCL(Ours)4000.24500.1399
MidVCL(Ours)4000.24060.1354
", + "bbox": [ + 137, + 501, + 437, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Cluster Analysis", + "text_level": 1, + "bbox": [ + 89, + 686, + 253, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Clusters of similar shapes To gain a deeper understanding of shape prototypes influence the representation learning, we first analyze the clusters produced by S-PCL. As illustrated in Figure.4, we visualize STL-10 images randomly selected from high-density clusters after 400 training epochs. It is evident that the elements within each cluster exhibit similar global shapes, demonstrating the S-PCL's ability to capture shape features and utilize shape as a structural regularity to organize the representation space.", + "bbox": [ + 89, + 709, + 482, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cluster assignment diverges from semantic category in late stages To study the early superior performance and fast saturation of S-PCL observed in the linear classifica", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Results of linear classification and segmentation of hybrid models. For hybrid methods (denoted as \"X+Y\"), training consists of 100 epochs using method X followed by 300 epochs using method Y. We can observe that constructing representation with perceptual inductive bias first will boost the performance of contrastive representation learning on both segmentation and classification tasks. The best two performances are in bold.", + "bbox": [ + 511, + 89, + 906, + 186 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8eb439dc35acdabc7ab6fcef04181932b8d70e6ca002b950f7630e9fa658e8e6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EpochsImageNet-100 Top-1 Acc(%)↑ADE20K mIoU(%)↑Depth RME(%)↓
PCL40074.030.50.1437
MoCoV240077.030.40.1434
S-PCL(Ours)40071.831.80.1381
S-PCL+PCL(Ours)40076.031.90.1373
S-PCL+MoCoV2(Ours)40078.031.90.1398
MidVCL(Ours)40074.032.10.1354
MidVCL+MoCoV2(Ours)40077.932.00.1409
", + "bbox": [ + 514, + 212, + 911, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion evaluation, we analyze the Adjusted Mutual Information (AMI) between K-Means cluster assignments of S-PCL and the categorical ground truth. A higher AMI indicates greater alignment between cluster assignments and semantic categories. As illustrated in Figure.5A, the AMI increases during the initial training stages but progressively decreases as the number of epochs increases. This trend suggests that clustering based solely on shape leads to a divergence from the semantic categories. In other words, it suggests that relying exclusively on shape is insufficient for effective object classification, which greatly requires semantic information. However, the shape shows enormous advantages at the early stage. These findings motivate the development of a hybrid framework that employs S-PCL and MidVCL as a warm-up strategy in the early training stages, which is detailed in §4.4.", + "bbox": [ + 511, + 364, + 906, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Perceptual inductive bias is beneficial for both image-level task and pixel-level task", + "text_level": 1, + "bbox": [ + 511, + 619, + 903, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Accelerated representation learning for classification In the results of linear classification (§4.1) and cluster analysis (§4.2), we observe tremendous benefits of incorporating shape prototypes. We attribute this phenomenon to an analogy with the accelerated learning of new nouns grouped by shapes during human infancy. Consequently, we investigate a hybrid framework that combines perceptual inductive bias and semantic representation training. Specifically, we propose training with S-PCL or MidVCL for the first 100 epochs and then switching to traditional contrastive learning frameworks (denoted as S-PCL+X and MidVCL+X respectively), allowing the model to discover additional structural regularities and develop a better representation space. Figure.5B shows the AMI results of PCL and S-PCL+PCL. It can be observed that the S-PCL warm-up moves the AMI curve upward. Figure.5C and D show the linear classifi", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9626", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cation accuracies for PCL and S-PCL+PCL on ImageNet-100 and STL10 datasets versus the number of pretraining epochs. For a fair comparison, all frameworks are trained using identical parameter settings and learning rate schedules. The results demonstrate that the S-PCL+PCL method achieves accuracy levels comparable to PCL while requiring less than half the number of training epochs on both ImageNet-100 and STL-10 datasets. The curve of hybrid frameworks of MidVCL is included in supplementary material and it demonstrates the same trend. Overall, our experimental results demonstrate that the sequential learning of intermediate shape representations followed by high-level semantic representations is beneficial for semantic tasks, particularly in accelerating the learning process.", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sequential training resolves the trade-off between image-level and pixel-level tasks Beyond the classification task, hybrid frameworks demonstrate their superiority also in segmentation task. As reported in Table.4, hybrid frameworks achieve strong performance in both classification and segmentation tasks, unlike methods relying only on either semantic or perceptual representations. Although hybrid frameworks are not as good as S-PCL and MidVCL in depth estimation, they still significantly outperform baseline methods. These results indicate that the additional learning with perceptual inductive bias before semantic contrastive learning provides a solution to the tradeoff mentioned in related works $\\S 2$", + "bbox": [ + 89, + 325, + 483, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Concentration on Object Contour and Shape Bias", + "text_level": 1, + "bbox": [ + 89, + 540, + 482, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shape silhouette forces model to focus on object contours Next, we study how shape representations vary the model's prediction decision-making process. To this end, we employed the SmoothGrad method to generate sensitivity maps, where highlighted regions contribute significantly to the model's inferences. We show some example images in the validation set of ImageNet-100 in Figure.6. Notably, comparing the sensitivity maps of S-PCL and others, we find that the semantic representations lead to a stronger concentration on local features of the object, which is a shortcut for classification. While the shape representations guide the model to focus on the contour of the object.", + "bbox": [ + 89, + 580, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "One may argue that although MoCoV2 and PCL focus more on texture, the silhouettes of objects can still be observed to some degree. To address this, we also analyze sensitivity maps of style-transferred images, that is, preserving object shape while altering texture, in addition to the original images, see Figure.7. After enormous style changes, MoCov2 and PCL are detracted by those irrelevant local features while S-PCL still has a strong concentration on the contours of the objects.", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0f4466150da650095c0deb057d5bb8cbe66c8726a3303603e8de7966812c7cd.jpg", + "image_caption": [ + "Figure 6. Smoothgrad sensitivity on different images." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6537bc662e66ecc7bfaacbb4e3451f9e8ef9404c472b89e945f8116b95c9c41a.jpg", + "image_caption": [ + "Figure 7. Smoothgrad sensitivity map on the image in different styles." + ], + "image_footnote": [], + "bbox": [ + 516, + 316, + 903, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shape bias score So far, we have found that the shape prototype can increase the shape bias of the model, alleviating a well-known machine's misalignment with humans. To quantitatively evaluate the shape bias of the model, we test our model on the model-vs-human benchmark. The model-vs-human benchmark is a protocol that is able to evaluate the 2D shape bias of a model. With a set of images with conflict cues, e.g. a cat shape with elephant skin texture, we can know which is the predominant visual cue for object recognition between shape and texture. Figure.8 demonstrates shape score across different frameworks. Importantly, the proposed SPCL is the only model that uses shape as the predominant cue, i.e. shape bias score greater than 0.5. RelfCL shows improvements compared to the rest of the approaches but it is still biased towards texture (fraction of shape decisions less than 0.5).", + "bbox": [ + 511, + 569, + 906, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Out-of-distribution robustness Many studies reveal that the model with high shape bias will also have higher robustness to out-of-distribution (OOD) data. To this end, we test models on the 17 OOD datasets included in the model-vs-human benchmarks. Figure.9 presents average classifica", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9627", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88fbd229946d84f976f1d6f67e6ad37eab9daed2c410ffc75edde96e40c279f4.jpg", + "image_caption": [ + "Figure 8. The shape bias score. The value on the top represents the fraction of shape decisions and the value at the bottom represents the fraction of texture decisions. Two values sum to 1 and 0.5 means neutrality, indicating no bias towards either shape or texture. Markers are the fraction of certain shape categories and vertical lines are the mean fraction across shape categories. The star markers represent our proposed approaches." + ], + "image_footnote": [], + "bbox": [ + 101, + 93, + 446, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion accuracies across all 17 OOD datasets in the model-vs-human benchmark. The full results are included in the supplementary material. S-PCL demonstrates statistically significant improvements in OOD accuracy compared to baseline models. This suggests that leveraging shape information not only aids in better generalization on seen data but also enhances the model's performance on unfamiliar inputs.", + "bbox": [ + 89, + 525, + 483, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 89, + 662, + 204, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we aim to integrate inductive biases from early and intermediate perceptual constructs into contrastive learning rather than relying solely on high-level semantic information. To achieve this, we maximize the mutual information between full-detail images and one or more perceptual constructs. Our findings reveal that inductive bias from global shape prototypes not only accelerates early-stage classification but also improves segmentation and depth estimation while enhancing the model's overall shape bias. Reflectance intrinsic images facilitate representation learning for classification and segmentation, although they do not benefit depth estimation, whereas shading intrinsic images enhance depth estimation without aiding classification or segmentation. Combining all three perceptual con", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e2d2c1a6dd1ac62a8f4ff54c40dcc615f64ebe9da05f5cee1625504e4cd6601.jpg", + "image_caption": [ + "Figure 9. Mean accuracy across 17 OOD dataset in model- vs-human benchmark. Bars are positioned in the descending order. The error bar is the standard error of the mean (SEM). A paired T-test is conducted. \\*\\* denotes $p < 0.01$ , $*$ denotes $p < 0.05$ , n.s. denotes \"not significant\" $p > 0.05$ ." + ], + "image_footnote": [], + "bbox": [ + 519, + 92, + 898, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "structures enables us to leverage the unique advantages of each. Moreover, employing sequential training—where the model first harnesses perceptual inductive biases before applying semantic contrastive learning—effectively mitigates the trade-off between image-level and pixel-level tasks. Overall, our results indicate that incorporating perceptual constructs generated by early and intermediate visual processing into self-supervised learning can help networks develop distinct intermediate representations that improve downstream visual tasks, offering a promising avenue for future research in contrastive learning.", + "bbox": [ + 511, + 429, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitation", + "text_level": 1, + "bbox": [ + 513, + 619, + 627, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We developed distinct approaches for specific perceptual constructs, focusing on shape, figure-ground, and intrinsic images. These constructs were chosen due to their close association with fundamental visual tasks like classification, segmentation, and depth estimation. However, additional mid-level perceptual constructs merit further exploration. Moreover, our experiments were conducted on a small model and limited datasets because of computing power constraints, underscoring the need to address scalability in future research.", + "bbox": [ + 511, + 648, + 906, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 825, + 692, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by NSF CISE RI 1816568 and NIH R01 EY030226-01A1. This work was also supported by the CMU HURAY program.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9628", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Edward H Adelson and Alex P Pentland. The perception of shading and reflectance. Perception as Bayesian inference, 409:423, 1996.", + "[2] Harry Barrow, J Tenenbaum, A Hanson, and E Riseman. Recovering intrinsic scene characteristics. Comput. vis. syst, 2 (3-26):2, 1978.", + "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments, 2021.", + "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[5] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223, Fort Lauderdale, FL, USA, 2011. PMLR.", + "[6] Hans P Op de Beeck, Katrien Torfs, and Johan Wagemans. Perceived shape similarity among unfamiliar objects and the organization of the human object vision pathway. Journal of Neuroscience, 28(40):10111-10123, 2008.", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009.", + "[8] James H Elder and Ljiljana Velisavljevic. Cue dynamics underlying rapid detection of animals in natural scenes. J Vis, 9(7):7, 2009.", + "[9] Paul Gavrikov, Jovita Lukasik, Steffen Jung, Robert Geirhos, Bianca Lamm, Muhammad Jehanzeb Mirza, Margret Keuper, and Janis Keuper. Are vision language models texture or shape biased and can we steer them?, 2024.", + "[10] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. On the surprising similarities between supervised and self-supervised models, 2020.", + "[11] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. In Advances in Neural Information Processing Systems, pages 23885-23899. Curran Associates, Inc., 2021.", + "[12] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. Advances in Neural Information Processing Systems, 34:23885-23899, 2021.", + "[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, 2022.", + "[14] Lisa Gershkoff-Stowe and Linda B Smith. Shape and the first hundred nouns. Child Dev, 75(4):1098-1114, 2004." + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In 2009 IEEE 12th International Conference on Computer Vision, pages 2335-2342. IEEE, 2009.", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.", + "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020.", + "[18] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers, 2024.", + "[19] V A Lamme. The neurophysiology of figure-ground segregation in primary visual cortex. J Neurosci, 15(2):1605-1615, 1995.", + "[20] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977.", + "[21] Min Seok Lee, Wooseok Shin, and Sung Won Han. Tracer: Extreme attention guided salient object tracing network, 2022.", + "[22] Tai Sing Lee, David Mumford, Richard Romero, and Victor A.F. Lamme. The role of the primary visual cortex in higher level vision. Vision Research, 38(15):2429-2454, 1998.", + "[23] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021.", + "[24] Tianqin Li, Ziqi Wen, Yangfan Li, and Tai Sing Lee. Emergence of shape bias in convolutional neural networks through activation sparsity, 2023.", + "[25] Ralph Linsker. Self-organization in a perceptual network. Computer, 21(3):105-117, 1988.", + "[26] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers, 2021.", + "[27] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020.", + "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019.", + "[29] Johan Wagemans, Joeri De Winter, Hans Op de Beeck, Annemie Ploeger, Tom Beckers, and Peter Vanroose. Identification of everyday objects on the basis of silhouette and outline versions. Perception, 37(2):207-244, 2008.", + "[30] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021.", + "[31] Ziqi Wen, Tianqin Li, and Tai Sing Lee. Does resistance to style-transfer equal shape bias? evaluating shape bias by distorted shape, 2023." + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9629", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Zhirong Wu, Yuanjun Xiong, Stella Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance-level discrimination, 2018.", + "[33] Yike Yuan, Xinghe Fu, Yunlong Yu, and Xi Li. Densedino: boosting dense self-supervised learning with token-based point-level consistency. arXiv preprint arXiv:2306.04654, 2023.", + "[34] Hong Zhou, Howard S Friedman, and Rüdiger Von Der Heydt. Coding of border ownership in monkey visual cortex. Journal of Neuroscience, 20(17):6594-6611, 2000.", + "[35] Karl Zipser, Victor A. F. Lamme, and Peter H. Schiller. Contextual modulation in primary visual cortex. Journal of Neuroscience, 16(22):7376-7389, 1996." + ], + "bbox": [ + 89, + 92, + 482, + 273 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9630", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/cdaa1dd9-bb97-42b0-9c82-a5fe76697ec3_model.json b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/cdaa1dd9-bb97-42b0-9c82-a5fe76697ec3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..753e33c4197f17d5416d289e939665de161646cc --- /dev/null +++ b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/cdaa1dd9-bb97-42b0-9c82-a5fe76697ec3_model.json @@ -0,0 +1,1617 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.872, + 0.155 + ], + "angle": 0, + "content": "Perceptual Inductive Bias Is What You Need Before Contrastive Learning" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.181, + 0.862, + 0.224 + ], + "angle": 0, + "content": "Junru Zhao Tianqin Li Dunhan Jiang Shenghao Wu Alan Ramirez Tai Sing Lee Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.226, + 0.812, + 0.241 + ], + "angle": 0, + "content": "{junruz, tianqinl, dunhanj, shenghaw, alanrami, taislee}@andrew.cmu.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.276, + 0.327, + 0.292 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.485, + 0.659 + ], + "angle": 0, + "content": "David Marr's seminal theory of human perception stipulates that visual processing is a multi-stage process, prioritizing the derivation of boundary and surface properties before forming semantic object representations. In contrast, contrastive representation learning frameworks typically bypass this explicit multi-stage approach, defining their objective as the direct learning of a semantic representation space for objects. While effective in general contexts, this approach sacrifices the inductive biases of vision, leading to slower convergence speed and learning shortcut resulting in texture bias. In this work, we demonstrate that leveraging Marr's multi-stage theory—by first constructing boundary and surface-level representations using perceptual constructs from early visual processing stages and subsequently training for object semantics—leads to \\(2x\\) faster convergence on ResNet18, improved final representations on semantic segmentation, depth estimation, and object recognition, and enhanced robustness and out-of-distribution capability. Together, we propose a pretraining stage before the general contrastive representation pretraining to further enhance the final representation quality and reduce the overall convergence time via inductive bias from human vision systems." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.692, + 0.222, + 0.709 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Human visual information processing is organized in stages, progressing from contrast and edge detection to perceptual organization and figure-ground segmentation, then to computing surfaces and shapes, and ultimately to object recognition. End-to-end supervised learning in neural networks for object classification in images often relies on the simplest features or shortcuts, bypassing essential intermediate representations, resulting in texture bias. Self-supervised contrastive learning, on the other hand, leverages the inductive bias of semantic segregation by grouping semantically similar scenes and separating dissimilar scenes in latent space. This also leads to representations that ac" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.274, + 0.878, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.428, + 0.908, + 0.526 + ], + "angle": 0, + "content": "Figure 1. We propose to leverage perceptual constructs from early vision processing stages to improve the contrastive learning pipeline. By introducing a pre-pretraining stage before general contrastive learning (CL), we inject inductive bias into the contrastive representation learning. We demonstrate that such inductive bias is beneficial to the representation quality and provides significantly convergence speed-up the contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.597 + ], + "angle": 0, + "content": "cess rich semantic information without building intermediate representations sensitive to surfaces and boundaries first." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this work, we evaluate whether perceptual constructs from early visual processing, such as figure-ground segmentation and intrinsic image decomposition, can serve as additional inductive biases for contrastive learning. Evidence of figure-ground organization has been found in the primary visual cortex and secondary visual cortex in primate visual systems [19, 22, 34, 35]. This early perceptual construct distinguishes figures against the ground, assigns border ownership to surfaces, and enables the construction of a neural representation of local shapes and global shapes along the ventral visual hierarchy. The perception of shapes could be innate or at least develop very fast [8, 29]. Furthermore, categorization and identification of unseen objects are considered to rely on the similarity of the overall shape, conceptualized through shape prototypes [6]. The recognition of shapes was found to be parallel to the development of vocabularies in language in infancy. Infants can learn new nouns faster when nonce objects are organized in shape, implicating a linked development between labeling objects linguistically and shape prototypes in the visual" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9621" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.481, + 0.182 + ], + "angle": 0, + "content": "system [14]. We conjecture that shape prototypes learned as cluster means of shape silhouettes obtained from figure-ground segregation can allow contrastive learning to learn better intermediate representations for a variety of visual tasks, such as object recognition, segmentation as well as depth estimation in the context of this paper." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.183, + 0.482, + 0.393 + ], + "angle": 0, + "content": "Another set of early perceptual constructs is called intrinsic images [2]. Intrinsic image decomposition [1, 15] is a generalization of Retinex developed by Edwin Land [20] to explain color constancy by discounting illumination effects to infer the reflectance or color of the objects. Intrinsic image decomposition factorizes an image as a product of a reflectance image and a shading image. Reflectance images reflect surface material properties and govern our perception of color and lightness. Shading images reflect the interaction of illumination and the 3D surface shape and are thus rich in implicit 2.5D information and affect our depth perception. What kinds of visual tasks would benefit from intermediate representations learned under the guidance of the inductive bias of these intrinsic images?" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.395, + 0.482, + 0.815 + ], + "angle": 0, + "content": "This work shows that global shapes of figures, particularly through a set of remembered prototypes, provide an inductive bias that enhances contrastive learning, resulting in learned representations that improve object recognition, segmentation, and depth estimation, with an increase in the shape bias of the model. We find that the inductive bias from the shape prototypes increases the speed of representational learning at the beginning, as in the early development of human infants, but ultimately has to be disabled to allow the learning of more refined representations to accommodate a greater variety of semantics. Thus, we propose a hybrid coarse-to-fine inductive bias that allows contrastive learning to learn intermediate representations first with the help of the prototypes and then refine the representation based on semantics. This results in a 2x faster representation learning. On the other hand, intrinsic reflectance images enhance the learned representations for segmentation and object recognition but do not benefit depth estimation. Conversely, intrinsic shading images improve depth estimation capabilities but with little impact on segmentation and object recognition. Finally, we demonstrate that integrating these three perceptual constructs provides additional performance benefits. Thus, we demonstrate that inductive biases derived from early perceptual constructs can differentially or collectively promote the contrastive learning of intermediate representations with greater sensitivity to surface depths and object shapes and better performance in downstream visual tasks." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.832, + 0.238, + 0.845 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.858, + 0.481, + 0.9 + ], + "angle": 0, + "content": "Contrastive learning is one of the most commonly used frameworks in discriminative self-supervised learning. It inherently aligns with the Infomax principle [25] by aim-" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.317 + ], + "angle": 0, + "content": "ing to maximize the mutual information between semantically similar scenes, which are usually different views of the same image. This aim is realized by maximizing InfoNCE loss function [28, 32]. InfoNCE is designed to maximize a lower bound on the mutual information between representations of positive view pairs while distinguishing them from negative view pairs. The views are usually generated by different data augmentations. Many studies [4, 17] emphasize the importance of data augmentations as contrastive learning requires intensive data augmentation to learn good semantic representations. Some works combine contrastive learning with clustering [3, 23]. In this kind of framework, the concept of prototype is introduced to organize the representation space in order to have more compact semantic clusters." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.469 + ], + "angle": 0, + "content": "After self-supervised pretraining, the representations are commonly evaluated by downstream tasks such as classification, segmentation and depth estimation. Many works [27, 30, 33] point out that there exists a trade-off between image-level classification (semantics) and pixel-wise classification or segmentation (perceptual grouping) in discriminative models. In other words, representations learned by state-of-the-art discriminative models often fail to perform effectively across both high-level semantic tasks and intermediate-level perceptual tasks." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.71 + ], + "angle": 0, + "content": "While humans rely heavily on intermediate perceptual constructs like object shapes in object recognition, texture bias is a well-known misalignment between CNN-based models and humans [10, 11, 13]. Many works endeavor to solve this misalignment by different approaches including 1) using data augmentation methods to vary texture, e.g. texture synthesis, style transfer [12, 13, 31], 2) changing to vision transformer [26], or 3) using more complex multimodal generative frameworks [9, 18]. Improved shape bias results in improvement of recognition robustness and domain generalization [13, 24]. However, a recent study [18] argues that most discriminative models are still strongly texture-biased and less robust to tremendous style changes. We attribute this phenomenon to the focus of discriminative models on learning semantic representations while overlooking intermediate-level representations." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.723, + 0.62, + 0.739 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.75, + 0.904, + 0.792 + ], + "angle": 0, + "content": "In this part, we will introduce how we incorporate global shape of figures in a prototypical manner and the intrinsic images as views into contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.812, + 0.904, + 0.9 + ], + "angle": 0, + "content": "Shape prototypical contrastive learning (S-PCL) Inspired by the fact that humans rely on the similarity of the overall shape envelope for object recognition, we design a contrastive learning framework that is capable of learning the concept of a group of similar shapes, i.e. shape prototypes. To represent the global shape of an object," + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9622" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.088, + 0.468, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.616 + ], + "angle": 0, + "content": "Figure 2. Illustration of Shape Prototypical Contrastive Learning (S-PCL) training framework The original images are input to the encoder and the offline-generated shape silhouettes generated are input to the momentum encoder. The K-Means is performed on the embedding of shape silhouettes. Similar shapes are clustered together and the centroid of each cluster is considered to be the shape prototype. The objective function, ShapeProtoNCE, is calculated on the representation of the original image and its corresponding shape prototype." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "we use the figure-ground shape silhouette as presented in Figure.2. Our framework is built upon prototypical contrastive learning (PCL)[23], which involves an online k-means clustering on the representation of the entire training data and considers the centroids of clusters as prototypes. In shape prototypical contrastive learning, the clustering is performed on the representation of shape silhouettes, as illustrated in Figure.2. The representation learning is guided by the clusters of shapes. With a set of images \\( X = \\{x_{1}, x_{2}, \\dots, x_{n}\\} \\) and matched shape silhouettes \\( M = \\{m_{1}, m_{2}, \\dots, m_{n}\\} \\), they are encoded by the same encoder to get their representations \\( V = \\{v_{1}, v_{2}, \\dots, v_{n}\\} \\) and \\( U = \\{u_{1}, u_{2}, \\dots, u_{n}\\} \\), respectively. K-Means clustering is employed on the set \\( U \\), producing \\( K \\) clusters and a set of centroids, \\( S = \\{s_{1}, s_{2}, \\dots, s_{K}\\} \\). The centroid of the cluster is representative of elements in the cluster, and thus is" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.092, + 0.887, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.277, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Illustration of the training framework that incorporates the intrinsic image as a view The original image is input to the encoder and one intrinsic image decomposed by Retinex algorithm is input to the momentum encoder. The InfoNCE between the representation of the original image and that of the intrinsic image is computed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.463 + ], + "angle": 0, + "content": "considered as the shape prototype. The mutual information between the representation of the original image \\( v_{i} \\) and the representation of the shape prototype \\( s_{i} \\) is maximized by optimizing the InfoNCE objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.474, + 0.907, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S h a p e P r o t o N C E}} = \\sum_ {i = 1} ^ {n} - \\log \\frac {\\exp \\left(v _ {i} \\cdot s _ {p}\\right) / \\phi_ {p})}{\\sum_ {j = 0} ^ {r} \\exp \\left(v _ {i} \\cdot s _ {j} / \\phi_ {j}\\right)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.909, + 0.646 + ], + "angle": 0, + "content": "where \\( s_p \\) is the prototype of the cluster \\( p \\) to which \\( u_i \\) belongs, \\( s_j, j \\neq p \\) are shape prototypes representing other shape clusters, and \\( \\phi_j \\) is the concentration estimation of cluster \\( j \\). Similar to PCL, we perform clustering for \\( N \\) times with different \\( K_i, i = 1,2,\\dots,N \\) and average the losses. The final loss function used is the combination of InfoNCE between the normally augmented view and the shape loss:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.659, + 0.907, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {I n f o N C E}} + \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\text {S h a p e P r o t o N C E}, K _ {i}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Intrinsic image as a view in contrastive learning Unlike shape prototypes, we treat the intrinsic image, either reflectance or shading, as an augmented view. We maximize the mutual information between the representations of normally augmented views and those of reflectance or shading images by optimizing the InfoNCE objective function, as illustrated in Figure 3. The final loss function is the sum of the standard InfoNCE loss between representations of two augmented views and the InfoNCE loss between representations of one augmented view and one intrinsic image. The approaches that incorporate reflectance or shading intrinsic images are termed Reflectance Contrastive Learn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9623" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.123 + ], + "angle": 0, + "content": "ing (RefCL) and Shading Contrastive Learning (ShadCL), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.14, + 0.483, + 0.203 + ], + "angle": 0, + "content": "Combination of shape and intrinsic images We also integrate the shape and the intrinsic image together according to equation (3). We call the combined approach mid-level vision contrastive learning (MidVCL)." + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.213, + 0.483, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {I n f o N C E}} + \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\text {S h a p e P r o t o N C E}, K _ {i}} + \\alpha \\mathcal {L} _ {\\text {S h a d}} + \\beta \\mathcal {L} _ {\\text {R e f l}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.483, + 0.299 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) is the weight of the shading and reflectance losses, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.484, + 0.51 + ], + "angle": 0, + "content": "Implementation details The experiments are conducted on the dataset ImageNet-100\\(^{1}\\), a subset of ImageNet-1K [7], and STL-10 [5]. The silhouette mask of the object is generated offline using TRACER [21]. We decompose a full-detail image to a reflectance intrinsic image and a shading image by Retinex algorithm [1, 15]. The ResNet18 [16] is adopted as the encoder and the output dimension of the last fully-connected layer is 256-D. We follow the data augmentation methods in previous works [4, 17, 23]. These data augmentations are applied to only the original image and the preprocessing for the shape silhouette and the intrinsic image only contains resizing and normalization. More details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.524, + 0.178, + 0.539 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.548, + 0.332, + 0.565 + ], + "angle": 0, + "content": "4.1. Representation Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.632 + ], + "angle": 0, + "content": "First and foremost, we evaluate representation learned under the guidance of intermediate perceptual constructs through multiple downstream visual tasks, including linear classification, segmentation, and depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Linear classification and transfer learning to ImageNet-1k We trained a linear classifier on frozen representations using the entire labeled training set of ImageNet-100 and STL-10, and here we report the Top-1 and Top-5 accuracies of ImageNet-1k on the validation sets. The results on STL-10 are included in the supplementary material. As illustrated in Table.1, S-PCL significantly outperforms the baselines during the early training stages (e.g., at 100 epochs) on both datasets. However, when training is extended to 400 epochs, the performance gains of S-PCL diminish, and it is even surpassed by the baseline approaches significantly on ImageNet-100, indicating a potential faster saturation of S-PCL compared to others. In contrast, RefCL maintains consistent performance improvements between 100 and 400 epochs, suggesting that reflectance information effectively" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.907, + 0.271 + ], + "angle": 0, + "content": "facilitates object recognition. Conversely, ShadCL, with inductive bias from shading, provides only a tiny gain during the early stages of training and may hinder object recognition when trained for a large number of epochs. MidVCL is similar to S-PCL on the classification task, being the best at epoch 100 while plateauing when the number of epoch increases to 400. We also test generalizability by transferring to a different dataset. This is performed by the linear classification on ImageNet-1k which has 10x more classes than ImageNet-100. The transfer learning results reported in Table.1 display the similar trend as the linear classification." + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.285, + 0.907, + 0.369 + ], + "angle": 0, + "content": "Table 1. Linear classification and ImageNet-1k transfer learning accuracies (\\%) of different approaches trained for 100 and 400 epochs on the ImageNet-100. For hybrid methods (denoted as \"X+Y\"), training consists of 100 epochs using method X followed by 300 epochs using method Y. The best-performing results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.395, + 0.933, + 0.675 + ], + "angle": 0, + "content": "
EpochsImageNet-100ImageNet-1k
Top-1 Acc (%)↑Top-5 Acc (%)↑Top-1 Acc (%)↑Top-5 Acc (%)↑
SimCLR10065.187.333.158.2
MoCoV210061.786.031.155.1
PCL10062.186.330.554.7
SwAV10064.888.333.457.8
BYOL10065.888.533.055.4
DINO10065.388.333.255.5
S-PCL(Ours)10070.290.337.261.9
RefCL(Ours)10065.588.135.260.1
ShadCL(Ours)10062.586.532.155.9
MidVCL(Ours)10070.590.437.062.7
SimCLR40077.293.840.866.9
MoCoV240077.093.841.668.1
PCL40074.091.937.863.3
SwAV40076.893.638.364.2
BYOL40075.892.642.967.6
DINO40077.193.638.062.9
S-PCL(Ours)40071.890.937.063.1
RefCL(Ours)40077.293.440.065.4
ShadCL(Ours)40071.890.934.960.6
MidVCL(Ours)40072.490.937.161.7
S-PCL+MoCoV2(Ours)40078.094.243.969.4
MidVCL+MoCoV2(Ours)40077.894.243.869.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Segmentaion and Depth Estimation Besides linear classification, which is a semantic and image-level task, it is also essential to evaluate representations on pixel-level tasks such as segmentation and depth estimation. Therefore, we conduct experiments on various benchmark datasets to assess performance in those tasks. The models evaluated in this part are pretrained for 400 epochs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "For segmentation tasks, we fine-tune the pretrained models with a segmentation head on the ADE20K and Cityscapes datasets. The mean Intersection over Union (mIoU) scores are reported in Table 2. Our MidVCL approach outperforms all other methods on both datasets, and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.902 + ], + "angle": 0, + "content": "1The dataset used here is reorganized by adjusting the train-validation split to a 7:3 ratio to intensify the challenge" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9624" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.898, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.361, + 0.908, + 0.406 + ], + "angle": 0, + "content": "Figure 4. Visualization of elements in clusters produced by S-PCL after trained on STL10 for 400 epochs. Each block represents a cluster and 10 randomly selected images (Top) and shape silhouettes (bottom) are presented. Images in each cluster have similar shape irrespective to their ground truth category, indicating the S-PCL's capability of grouping objects with similar shape." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.417, + 0.297, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.417, + 0.5, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.417, + 0.7, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.417, + 0.905, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.557, + 0.908, + 0.613 + ], + "angle": 0, + "content": "Figure 5. A. AMI between cluster assignment of S-PCL trained on ImageNet-100 and ground truth label. The AMI increases at the early stage and decreases later on. B. AMI between cluster assignment of PCL and S-PCl+PCL trained on ImageNet-100 and ground truth label. The AMI of PCL keeps increasing. S-PCL warm-up moves the curve upward. C and D. Linear classification accuracy on two datasets for PCL, S-PCL and S-PCL+PCL. S-PCL suffers rapid saturation but is able to serve as a warm-up strategy to accelerate training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.639, + 0.486, + 0.837 + ], + "angle": 0, + "content": "the S-PCL is on the second place. This superior performance is attributable to the incorporation of shape silhouette masks in both MidVCL and S-PCL, which provide figure-ground information critical for segmentation tasks. Additionally, we observe that ReflCL ranks the third, presumably because the reflectance intrinsic image segregates instances based on albedo, thereby facilitating more accurate boundary detection. Also, it turns out that ReflCL performs poorly on depth estimation, suggesting that solely reflectance is insufficient to distinguish foreground and background although the boundaries of instances are clearer. In contrast, ShadCL ends up with only nuanced improvements in segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.486, + 0.903 + ], + "angle": 0, + "content": "For depth estimation, we fine-tune the pretrained models with an additional head using the NYU Depth V2 dataset. The relative errors (REL) and root mean errors (RME) are presented in Table 3. MidVCL and S-PCL continue" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.716 + ], + "angle": 0, + "content": "to achieve the highest performance among all evaluated approaches. Notably, ShadCL ranks third, indicating that shading information is beneficial for depth estimation tasks. However, the advantages of incorporating reflectance information in this context appear to be limited." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Overall Our evaluations on these downstream tasks demonstrate that shape prototypes, reflectance intrinsic image, and shading intrinsic image facilitate representation learning in different ways. The shape prototype helps all three downstream tasks, especially superior in pixel-level tasks (segmentation and depth estimation). However, it is only beneficial in the early stage for the image-level classification task and saturates rapidly. Additionally, the reflectance intrinsic image helps classification and segmentation but not depth estimation. On the other hand, the shading intrinsic image enhances depth estimation thanks to its implicit information of 3D shape and illuminance, but not classification and seg-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9625" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "mentation, where spatial information is not that important. Among three perceptual constructs, we find the shape prototype yields the most interesting results, thus we conduct further studies to understand the underlying mechanism of the shape prototype." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.181, + 0.483, + 0.21 + ], + "angle": 0, + "content": "Table 2. Segmentation results. The mIoU(%) values on ADE20K and Cityscapes are reported. The best performances are in bold." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.234, + 0.447, + 0.403 + ], + "angle": 0, + "content": "
EpochsADE20K mIoU(%)↑Cityscapes mIoU(%)↑
SimCLR40030.463.4
MoCoV240030.463.1
PCL40030.564.1
SwAV40030.261.9
BYOL40030.961.6
DINO40030.261.0
S-PCL(Ours)40031.868.3
RefCL(Ours)40031.865.5
ShadCL(Ours)40031.064.4
MidVCL(Ours)40032.168.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.435, + 0.483, + 0.477 + ], + "angle": 0, + "content": "Table 3. Depth estimation results. The relative error and root mean error on NYU Depth V2 dataset are reported. The best performances are in bold." + }, + { + "type": "table", + "bbox": [ + 0.138, + 0.502, + 0.439, + 0.658 + ], + "angle": 0, + "content": "
EpochsREL↓RME↓
SimCLR4000.25450.1420
MoCoV24000.25500.1434
PCL4000.25870.1437
SwAV4000.25860.1459
BYOL4000.25430.1458
DINO4000.25950.1451
S-PCL(Ours)4000.24280.1381
RefCL(Ours)4000.25330.1420
ShadCL(Ours)4000.24500.1399
MidVCL(Ours)4000.24060.1354
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.687, + 0.254, + 0.704 + ], + "angle": 0, + "content": "4.2. Cluster Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.71, + 0.483, + 0.846 + ], + "angle": 0, + "content": "Clusters of similar shapes To gain a deeper understanding of shape prototypes influence the representation learning, we first analyze the clusters produced by S-PCL. As illustrated in Figure.4, we visualize STL-10 images randomly selected from high-density clusters after 400 training epochs. It is evident that the elements within each cluster exhibit similar global shapes, demonstrating the S-PCL's ability to capture shape features and utilize shape as a structural regularity to organize the representation space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Cluster assignment diverges from semantic category in late stages To study the early superior performance and fast saturation of S-PCL observed in the linear classifica" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.907, + 0.187 + ], + "angle": 0, + "content": "Table 4. Results of linear classification and segmentation of hybrid models. For hybrid methods (denoted as \"X+Y\"), training consists of 100 epochs using method X followed by 300 epochs using method Y. We can observe that constructing representation with perceptual inductive bias first will boost the performance of contrastive representation learning on both segmentation and classification tasks. The best two performances are in bold." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.213, + 0.912, + 0.336 + ], + "angle": 0, + "content": "
EpochsImageNet-100 Top-1 Acc(%)↑ADE20K mIoU(%)↑Depth RME(%)↓
PCL40074.030.50.1437
MoCoV240077.030.40.1434
S-PCL(Ours)40071.831.80.1381
S-PCL+PCL(Ours)40076.031.90.1373
S-PCL+MoCoV2(Ours)40078.031.90.1398
MidVCL(Ours)40074.032.10.1354
MidVCL+MoCoV2(Ours)40077.932.00.1409
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.608 + ], + "angle": 0, + "content": "tion evaluation, we analyze the Adjusted Mutual Information (AMI) between K-Means cluster assignments of S-PCL and the categorical ground truth. A higher AMI indicates greater alignment between cluster assignments and semantic categories. As illustrated in Figure.5A, the AMI increases during the initial training stages but progressively decreases as the number of epochs increases. This trend suggests that clustering based solely on shape leads to a divergence from the semantic categories. In other words, it suggests that relying exclusively on shape is insufficient for effective object classification, which greatly requires semantic information. However, the shape shows enormous advantages at the early stage. These findings motivate the development of a hybrid framework that employs S-PCL and MidVCL as a warm-up strategy in the early training stages, which is detailed in §4.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.905, + 0.653 + ], + "angle": 0, + "content": "4.3. Perceptual inductive bias is beneficial for both image-level task and pixel-level task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Accelerated representation learning for classification In the results of linear classification (§4.1) and cluster analysis (§4.2), we observe tremendous benefits of incorporating shape prototypes. We attribute this phenomenon to an analogy with the accelerated learning of new nouns grouped by shapes during human infancy. Consequently, we investigate a hybrid framework that combines perceptual inductive bias and semantic representation training. Specifically, we propose training with S-PCL or MidVCL for the first 100 epochs and then switching to traditional contrastive learning frameworks (denoted as S-PCL+X and MidVCL+X respectively), allowing the model to discover additional structural regularities and develop a better representation space. Figure.5B shows the AMI results of PCL and S-PCL+PCL. It can be observed that the S-PCL warm-up moves the AMI curve upward. Figure.5C and D show the linear classifi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9626" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "cation accuracies for PCL and S-PCL+PCL on ImageNet-100 and STL10 datasets versus the number of pretraining epochs. For a fair comparison, all frameworks are trained using identical parameter settings and learning rate schedules. The results demonstrate that the S-PCL+PCL method achieves accuracy levels comparable to PCL while requiring less than half the number of training epochs on both ImageNet-100 and STL-10 datasets. The curve of hybrid frameworks of MidVCL is included in supplementary material and it demonstrates the same trend. Overall, our experimental results demonstrate that the sequential learning of intermediate shape representations followed by high-level semantic representations is beneficial for semantic tasks, particularly in accelerating the learning process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Sequential training resolves the trade-off between image-level and pixel-level tasks Beyond the classification task, hybrid frameworks demonstrate their superiority also in segmentation task. As reported in Table.4, hybrid frameworks achieve strong performance in both classification and segmentation tasks, unlike methods relying only on either semantic or perceptual representations. Although hybrid frameworks are not as good as S-PCL and MidVCL in depth estimation, they still significantly outperform baseline methods. These results indicate that the additional learning with perceptual inductive bias before semantic contrastive learning provides a solution to the tradeoff mentioned in related works \\(\\S 2\\)" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.541, + 0.483, + 0.571 + ], + "angle": 0, + "content": "4.4. Concentration on Object Contour and Shape Bias" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Shape silhouette forces model to focus on object contours Next, we study how shape representations vary the model's prediction decision-making process. To this end, we employed the SmoothGrad method to generate sensitivity maps, where highlighted regions contribute significantly to the model's inferences. We show some example images in the validation set of ImageNet-100 in Figure.6. Notably, comparing the sensitivity maps of S-PCL and others, we find that the semantic representations lead to a stronger concentration on local features of the object, which is a shortcut for classification. While the shape representations guide the model to focus on the contour of the object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "One may argue that although MoCoV2 and PCL focus more on texture, the silhouettes of objects can still be observed to some degree. To address this, we also analyze sensitivity maps of style-transferred images, that is, preserving object shape while altering texture, in addition to the original images, see Figure.7. After enormous style changes, MoCov2 and PCL are detracted by those irrelevant local features while S-PCL still has a strong concentration on the contours of the objects." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.286, + 0.868, + 0.301 + ], + "angle": 0, + "content": "Figure 6. Smoothgrad sensitivity on different images." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.513, + 0.907, + 0.541 + ], + "angle": 0, + "content": "Figure 7. Smoothgrad sensitivity map on the image in different styles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.812 + ], + "angle": 0, + "content": "Shape bias score So far, we have found that the shape prototype can increase the shape bias of the model, alleviating a well-known machine's misalignment with humans. To quantitatively evaluate the shape bias of the model, we test our model on the model-vs-human benchmark. The model-vs-human benchmark is a protocol that is able to evaluate the 2D shape bias of a model. With a set of images with conflict cues, e.g. a cat shape with elephant skin texture, we can know which is the predominant visual cue for object recognition between shape and texture. Figure.8 demonstrates shape score across different frameworks. Importantly, the proposed SPCL is the only model that uses shape as the predominant cue, i.e. shape bias score greater than 0.5. RelfCL shows improvements compared to the rest of the approaches but it is still biased towards texture (fraction of shape decisions less than 0.5)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Out-of-distribution robustness Many studies reveal that the model with high shape bias will also have higher robustness to out-of-distribution (OOD) data. To this end, we test models on the 17 OOD datasets included in the model-vs-human benchmarks. Figure.9 presents average classifica" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9627" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.094, + 0.447, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.499 + ], + "angle": 0, + "content": "Figure 8. The shape bias score. The value on the top represents the fraction of shape decisions and the value at the bottom represents the fraction of texture decisions. Two values sum to 1 and 0.5 means neutrality, indicating no bias towards either shape or texture. Markers are the fraction of certain shape categories and vertical lines are the mean fraction across shape categories. The star markers represent our proposed approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.648 + ], + "angle": 0, + "content": "tion accuracies across all 17 OOD datasets in the model-vs-human benchmark. The full results are included in the supplementary material. S-PCL demonstrates statistically significant improvements in OOD accuracy compared to baseline models. This suggests that leveraging shape information not only aids in better generalization on seen data but also enhances the model's performance on unfamiliar inputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.663, + 0.205, + 0.68 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this work, we aim to integrate inductive biases from early and intermediate perceptual constructs into contrastive learning rather than relying solely on high-level semantic information. To achieve this, we maximize the mutual information between full-detail images and one or more perceptual constructs. Our findings reveal that inductive bias from global shape prototypes not only accelerates early-stage classification but also improves segmentation and depth estimation while enhancing the model's overall shape bias. Reflectance intrinsic images facilitate representation learning for classification and segmentation, although they do not benefit depth estimation, whereas shading intrinsic images enhance depth estimation without aiding classification or segmentation. Combining all three perceptual con" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.093, + 0.9, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.326, + 0.907, + 0.396 + ], + "angle": 0, + "content": "Figure 9. Mean accuracy across 17 OOD dataset in model- vs-human benchmark. Bars are positioned in the descending order. The error bar is the standard error of the mean (SEM). A paired T-test is conducted. \\*\\* denotes \\(p < 0.01\\), \\(*\\) denotes \\(p < 0.05\\), n.s. denotes \"not significant\" \\(p > 0.05\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.596 + ], + "angle": 0, + "content": "structures enables us to leverage the unique advantages of each. Moreover, employing sequential training—where the model first harnesses perceptual inductive biases before applying semantic contrastive learning—effectively mitigates the trade-off between image-level and pixel-level tasks. Overall, our results indicate that incorporating perceptual constructs generated by early and intermediate visual processing into self-supervised learning can help networks develop distinct intermediate representations that improve downstream visual tasks, offering a promising avenue for future research in contrastive learning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.628, + 0.637 + ], + "angle": 0, + "content": "6. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.801 + ], + "angle": 0, + "content": "We developed distinct approaches for specific perceptual constructs, focusing on shape, figure-ground, and intrinsic images. These constructs were chosen due to their close association with fundamental visual tasks like classification, segmentation, and depth estimation. However, additional mid-level perceptual constructs merit further exploration. Moreover, our experiments were conducted on a small model and limited datasets because of computing power constraints, underscoring the need to address scalability in future research." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.827, + 0.694, + 0.844 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work was supported by NSF CISE RI 1816568 and NIH R01 EY030226-01A1. This work was also supported by the CMU HURAY program." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9628" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Edward H Adelson and Alex P Pentland. The perception of shading and reflectance. Perception as Bayesian inference, 409:423, 1996." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.2 + ], + "angle": 0, + "content": "[2] Harry Barrow, J Tenenbaum, A Hanson, and E Riseman. Recovering intrinsic scene characteristics. Comput. vis. syst, 2 (3-26):2, 1978." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.483, + 0.244 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.484, + 0.372 + ], + "angle": 0, + "content": "[5] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223, Fort Lauderdale, FL, USA, 2011. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[6] Hans P Op de Beeck, Katrien Torfs, and Johan Wagemans. Perceived shape similarity among unfamiliar objects and the organization of the human object vision pathway. Journal of Neuroscience, 28(40):10111-10123, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[8] James H Elder and Ljiljana Velisavljevic. Cue dynamics underlying rapid detection of animals in natural scenes. J Vis, 9(7):7, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[9] Paul Gavrikov, Jovita Lukasik, Steffen Jung, Robert Geirhos, Bianca Lamm, Muhammad Jehanzeb Mirza, Margret Keuper, and Janis Keuper. Are vision language models texture or shape biased and can we steer them?, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.589, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[10] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. On the surprising similarities between supervised and self-supervised models, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.647, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[11] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. In Advances in Neural Information Processing Systems, pages 23885-23899. Curran Associates, Inc., 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[12] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. Advances in Neural Information Processing Systems, 34:23885-23899, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Lisa Gershkoff-Stowe and Linda B Smith. Shape and the first hundred nouns. Child Dev, 75(4):1098-1114, 2004." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[15] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In 2009 IEEE 12th International Conference on Computer Vision, pages 2335-2342. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[18] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[19] V A Lamme. The neurophysiology of figure-ground segregation in primary visual cortex. J Neurosci, 15(2):1605-1615, 1995." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[20] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[21] Min Seok Lee, Wooseok Shin, and Sung Won Han. Tracer: Extreme attention guided salient object tracing network, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[22] Tai Sing Lee, David Mumford, Richard Romero, and Victor A.F. Lamme. The role of the primary visual cortex in higher level vision. Vision Research, 38(15):2429-2454, 1998." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[23] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[24] Tianqin Li, Ziqi Wen, Yangfan Li, and Tai Sing Lee. Emergence of shape bias in convolutional neural networks through activation sparsity, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.904, + 0.603 + ], + "angle": 0, + "content": "[25] Ralph Linsker. Self-organization in a perceptual network. Computer, 21(3):105-117, 1988." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[26] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[27] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[29] Johan Wagemans, Joeri De Winter, Hans Op de Beeck, Annemie Ploeger, Tom Beckers, and Peter Vanroose. Identification of everyday objects on the basis of silhouette and outline versions. Perception, 37(2):207-244, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[30] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Ziqi Wen, Tianqin Li, and Tai Sing Lee. Does resistance to style-transfer equal shape bias? evaluating shape bias by distorted shape, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9629" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[32] Zhirong Wu, Yuanjun Xiong, Stella Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance-level discrimination, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[33] Yike Yuan, Xinghe Fu, Yunlong Yu, and Xi Li. Densedino: boosting dense self-supervised learning with token-based point-level consistency. arXiv preprint arXiv:2306.04654, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[34] Hong Zhou, Howard S Friedman, and Rüdiger Von Der Heydt. Coding of border ownership in monkey visual cortex. Journal of Neuroscience, 20(17):6594-6611, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.235, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[35] Karl Zipser, Victor A. F. Lamme, and Peter H. Schiller. Contextual modulation in primary visual cortex. Journal of Neuroscience, 16(22):7376-7389, 1996." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9630" + } + ] +] \ No newline at end of file diff --git a/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/cdaa1dd9-bb97-42b0-9c82-a5fe76697ec3_origin.pdf b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/cdaa1dd9-bb97-42b0-9c82-a5fe76697ec3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8b405db6cf060b67b8dbaecf92d26dd1ecba9eb7 --- /dev/null +++ b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/cdaa1dd9-bb97-42b0-9c82-a5fe76697ec3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0f26e3f144a04ea843a61e002ed2a59430c42361ab7a19d0e811c78353c4129 +size 6007826 diff --git a/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/full.md b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..00295c53830c4a71615199a7f188c2393ba003ca --- /dev/null +++ b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/full.md @@ -0,0 +1,222 @@ +# Perceptual Inductive Bias Is What You Need Before Contrastive Learning + +Junru Zhao Tianqin Li Dunhan Jiang Shenghao Wu Alan Ramirez Tai Sing Lee Carnegie Mellon University + +{junruz, tianqinl, dunhanj, shenghaw, alanrami, taislee}@andrew.cmu.edu + +# Abstract + +David Marr's seminal theory of human perception stipulates that visual processing is a multi-stage process, prioritizing the derivation of boundary and surface properties before forming semantic object representations. In contrast, contrastive representation learning frameworks typically bypass this explicit multi-stage approach, defining their objective as the direct learning of a semantic representation space for objects. While effective in general contexts, this approach sacrifices the inductive biases of vision, leading to slower convergence speed and learning shortcut resulting in texture bias. In this work, we demonstrate that leveraging Marr's multi-stage theory—by first constructing boundary and surface-level representations using perceptual constructs from early visual processing stages and subsequently training for object semantics—leads to $2x$ faster convergence on ResNet18, improved final representations on semantic segmentation, depth estimation, and object recognition, and enhanced robustness and out-of-distribution capability. Together, we propose a pretraining stage before the general contrastive representation pretraining to further enhance the final representation quality and reduce the overall convergence time via inductive bias from human vision systems. + +# 1. Introduction + +Human visual information processing is organized in stages, progressing from contrast and edge detection to perceptual organization and figure-ground segmentation, then to computing surfaces and shapes, and ultimately to object recognition. End-to-end supervised learning in neural networks for object classification in images often relies on the simplest features or shortcuts, bypassing essential intermediate representations, resulting in texture bias. Self-supervised contrastive learning, on the other hand, leverages the inductive bias of semantic segregation by grouping semantically similar scenes and separating dissimilar scenes in latent space. This also leads to representations that ac + +![](images/41f7a0da2d88bc88904bedce87effa10ba803a5d7fd8f9f2fb3bd5e71c91eda3.jpg) +Figure 1. We propose to leverage perceptual constructs from early vision processing stages to improve the contrastive learning pipeline. By introducing a pre-pretraining stage before general contrastive learning (CL), we inject inductive bias into the contrastive representation learning. We demonstrate that such inductive bias is beneficial to the representation quality and provides significantly convergence speed-up the contrastive learning. + +cess rich semantic information without building intermediate representations sensitive to surfaces and boundaries first. + +In this work, we evaluate whether perceptual constructs from early visual processing, such as figure-ground segmentation and intrinsic image decomposition, can serve as additional inductive biases for contrastive learning. Evidence of figure-ground organization has been found in the primary visual cortex and secondary visual cortex in primate visual systems [19, 22, 34, 35]. This early perceptual construct distinguishes figures against the ground, assigns border ownership to surfaces, and enables the construction of a neural representation of local shapes and global shapes along the ventral visual hierarchy. The perception of shapes could be innate or at least develop very fast [8, 29]. Furthermore, categorization and identification of unseen objects are considered to rely on the similarity of the overall shape, conceptualized through shape prototypes [6]. The recognition of shapes was found to be parallel to the development of vocabularies in language in infancy. Infants can learn new nouns faster when nonce objects are organized in shape, implicating a linked development between labeling objects linguistically and shape prototypes in the visual + +system [14]. We conjecture that shape prototypes learned as cluster means of shape silhouettes obtained from figure-ground segregation can allow contrastive learning to learn better intermediate representations for a variety of visual tasks, such as object recognition, segmentation as well as depth estimation in the context of this paper. + +Another set of early perceptual constructs is called intrinsic images [2]. Intrinsic image decomposition [1, 15] is a generalization of Retinex developed by Edwin Land [20] to explain color constancy by discounting illumination effects to infer the reflectance or color of the objects. Intrinsic image decomposition factorizes an image as a product of a reflectance image and a shading image. Reflectance images reflect surface material properties and govern our perception of color and lightness. Shading images reflect the interaction of illumination and the 3D surface shape and are thus rich in implicit 2.5D information and affect our depth perception. What kinds of visual tasks would benefit from intermediate representations learned under the guidance of the inductive bias of these intrinsic images? + +This work shows that global shapes of figures, particularly through a set of remembered prototypes, provide an inductive bias that enhances contrastive learning, resulting in learned representations that improve object recognition, segmentation, and depth estimation, with an increase in the shape bias of the model. We find that the inductive bias from the shape prototypes increases the speed of representational learning at the beginning, as in the early development of human infants, but ultimately has to be disabled to allow the learning of more refined representations to accommodate a greater variety of semantics. Thus, we propose a hybrid coarse-to-fine inductive bias that allows contrastive learning to learn intermediate representations first with the help of the prototypes and then refine the representation based on semantics. This results in a 2x faster representation learning. On the other hand, intrinsic reflectance images enhance the learned representations for segmentation and object recognition but do not benefit depth estimation. Conversely, intrinsic shading images improve depth estimation capabilities but with little impact on segmentation and object recognition. Finally, we demonstrate that integrating these three perceptual constructs provides additional performance benefits. Thus, we demonstrate that inductive biases derived from early perceptual constructs can differentially or collectively promote the contrastive learning of intermediate representations with greater sensitivity to surface depths and object shapes and better performance in downstream visual tasks. + +# 2. Related Works + +Contrastive learning is one of the most commonly used frameworks in discriminative self-supervised learning. It inherently aligns with the Infomax principle [25] by aim- + +ing to maximize the mutual information between semantically similar scenes, which are usually different views of the same image. This aim is realized by maximizing InfoNCE loss function [28, 32]. InfoNCE is designed to maximize a lower bound on the mutual information between representations of positive view pairs while distinguishing them from negative view pairs. The views are usually generated by different data augmentations. Many studies [4, 17] emphasize the importance of data augmentations as contrastive learning requires intensive data augmentation to learn good semantic representations. Some works combine contrastive learning with clustering [3, 23]. In this kind of framework, the concept of prototype is introduced to organize the representation space in order to have more compact semantic clusters. + +After self-supervised pretraining, the representations are commonly evaluated by downstream tasks such as classification, segmentation and depth estimation. Many works [27, 30, 33] point out that there exists a trade-off between image-level classification (semantics) and pixel-wise classification or segmentation (perceptual grouping) in discriminative models. In other words, representations learned by state-of-the-art discriminative models often fail to perform effectively across both high-level semantic tasks and intermediate-level perceptual tasks. + +While humans rely heavily on intermediate perceptual constructs like object shapes in object recognition, texture bias is a well-known misalignment between CNN-based models and humans [10, 11, 13]. Many works endeavor to solve this misalignment by different approaches including 1) using data augmentation methods to vary texture, e.g. texture synthesis, style transfer [12, 13, 31], 2) changing to vision transformer [26], or 3) using more complex multimodal generative frameworks [9, 18]. Improved shape bias results in improvement of recognition robustness and domain generalization [13, 24]. However, a recent study [18] argues that most discriminative models are still strongly texture-biased and less robust to tremendous style changes. We attribute this phenomenon to the focus of discriminative models on learning semantic representations while overlooking intermediate-level representations. + +# 3. Approach + +In this part, we will introduce how we incorporate global shape of figures in a prototypical manner and the intrinsic images as views into contrastive learning. + +Shape prototypical contrastive learning (S-PCL) Inspired by the fact that humans rely on the similarity of the overall shape envelope for object recognition, we design a contrastive learning framework that is capable of learning the concept of a group of similar shapes, i.e. shape prototypes. To represent the global shape of an object, + +![](images/c5affb8a114f46e6a133aa6e0d18d5b1b20d813b35cdce9b51de366d171f4f91.jpg) +Figure 2. Illustration of Shape Prototypical Contrastive Learning (S-PCL) training framework The original images are input to the encoder and the offline-generated shape silhouettes generated are input to the momentum encoder. The K-Means is performed on the embedding of shape silhouettes. Similar shapes are clustered together and the centroid of each cluster is considered to be the shape prototype. The objective function, ShapeProtoNCE, is calculated on the representation of the original image and its corresponding shape prototype. + +we use the figure-ground shape silhouette as presented in Figure.2. Our framework is built upon prototypical contrastive learning (PCL)[23], which involves an online k-means clustering on the representation of the entire training data and considers the centroids of clusters as prototypes. In shape prototypical contrastive learning, the clustering is performed on the representation of shape silhouettes, as illustrated in Figure.2. The representation learning is guided by the clusters of shapes. With a set of images $X = \{x_{1}, x_{2}, \dots, x_{n}\}$ and matched shape silhouettes $M = \{m_{1}, m_{2}, \dots, m_{n}\}$ , they are encoded by the same encoder to get their representations $V = \{v_{1}, v_{2}, \dots, v_{n}\}$ and $U = \{u_{1}, u_{2}, \dots, u_{n}\}$ , respectively. K-Means clustering is employed on the set $U$ , producing $K$ clusters and a set of centroids, $S = \{s_{1}, s_{2}, \dots, s_{K}\}$ . The centroid of the cluster is representative of elements in the cluster, and thus is + +![](images/4ab995d317b15ef71b93f419b696001117691500262389085bda80f9530984c4.jpg) +Figure 3. Illustration of the training framework that incorporates the intrinsic image as a view The original image is input to the encoder and one intrinsic image decomposed by Retinex algorithm is input to the momentum encoder. The InfoNCE between the representation of the original image and that of the intrinsic image is computed. + +considered as the shape prototype. The mutual information between the representation of the original image $v_{i}$ and the representation of the shape prototype $s_{i}$ is maximized by optimizing the InfoNCE objective function: + +$$ +\mathcal {L} _ {\text {S h a p e P r o t o N C E}} = \sum_ {i = 1} ^ {n} - \log \frac {\exp \left(v _ {i} \cdot s _ {p}\right) / \phi_ {p})}{\sum_ {j = 0} ^ {r} \exp \left(v _ {i} \cdot s _ {j} / \phi_ {j}\right)} \tag {1} +$$ + +where $s_p$ is the prototype of the cluster $p$ to which $u_i$ belongs, $s_j, j \neq p$ are shape prototypes representing other shape clusters, and $\phi_j$ is the concentration estimation of cluster $j$ . Similar to PCL, we perform clustering for $N$ times with different $K_i, i = 1,2,\dots,N$ and average the losses. The final loss function used is the combination of InfoNCE between the normally augmented view and the shape loss: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {I n f o N C E}} + \frac {1}{N} \sum_ {i = 1} ^ {N} \mathcal {L} _ {\text {S h a p e P r o t o N C E}, K _ {i}} \tag {2} +$$ + +Intrinsic image as a view in contrastive learning Unlike shape prototypes, we treat the intrinsic image, either reflectance or shading, as an augmented view. We maximize the mutual information between the representations of normally augmented views and those of reflectance or shading images by optimizing the InfoNCE objective function, as illustrated in Figure 3. The final loss function is the sum of the standard InfoNCE loss between representations of two augmented views and the InfoNCE loss between representations of one augmented view and one intrinsic image. The approaches that incorporate reflectance or shading intrinsic images are termed Reflectance Contrastive Learn + +ing (RefCL) and Shading Contrastive Learning (ShadCL), respectively. + +Combination of shape and intrinsic images We also integrate the shape and the intrinsic image together according to equation (3). We call the combined approach mid-level vision contrastive learning (MidVCL). + +$$ +\mathcal {L} = \mathcal {L} _ {\text {I n f o N C E}} + \frac {1}{N} \sum_ {i = 1} ^ {N} \mathcal {L} _ {\text {S h a p e P r o t o N C E}, K _ {i}} + \alpha \mathcal {L} _ {\text {S h a d}} + \beta \mathcal {L} _ {\text {R e f l}} \tag {3} +$$ + +where $\alpha$ and $\beta$ is the weight of the shading and reflectance losses, respectively. + +Implementation details The experiments are conducted on the dataset ImageNet-100 $^{1}$ , a subset of ImageNet-1K [7], and STL-10 [5]. The silhouette mask of the object is generated offline using TRACER [21]. We decompose a full-detail image to a reflectance intrinsic image and a shading image by Retinex algorithm [1, 15]. The ResNet18 [16] is adopted as the encoder and the output dimension of the last fully-connected layer is 256-D. We follow the data augmentation methods in previous works [4, 17, 23]. These data augmentations are applied to only the original image and the preprocessing for the shape silhouette and the intrinsic image only contains resizing and normalization. More details can be found in the supplementary material. + +# 4. Results + +# 4.1. Representation Evaluation + +First and foremost, we evaluate representation learned under the guidance of intermediate perceptual constructs through multiple downstream visual tasks, including linear classification, segmentation, and depth estimation. + +Linear classification and transfer learning to ImageNet-1k We trained a linear classifier on frozen representations using the entire labeled training set of ImageNet-100 and STL-10, and here we report the Top-1 and Top-5 accuracies of ImageNet-1k on the validation sets. The results on STL-10 are included in the supplementary material. As illustrated in Table.1, S-PCL significantly outperforms the baselines during the early training stages (e.g., at 100 epochs) on both datasets. However, when training is extended to 400 epochs, the performance gains of S-PCL diminish, and it is even surpassed by the baseline approaches significantly on ImageNet-100, indicating a potential faster saturation of S-PCL compared to others. In contrast, RefCL maintains consistent performance improvements between 100 and 400 epochs, suggesting that reflectance information effectively + +facilitates object recognition. Conversely, ShadCL, with inductive bias from shading, provides only a tiny gain during the early stages of training and may hinder object recognition when trained for a large number of epochs. MidVCL is similar to S-PCL on the classification task, being the best at epoch 100 while plateauing when the number of epoch increases to 400. We also test generalizability by transferring to a different dataset. This is performed by the linear classification on ImageNet-1k which has 10x more classes than ImageNet-100. The transfer learning results reported in Table.1 display the similar trend as the linear classification. + +Table 1. Linear classification and ImageNet-1k transfer learning accuracies (\%) of different approaches trained for 100 and 400 epochs on the ImageNet-100. For hybrid methods (denoted as "X+Y"), training consists of 100 epochs using method X followed by 300 epochs using method Y. The best-performing results are highlighted in bold. + +
EpochsImageNet-100ImageNet-1k
Top-1 Acc (%)↑Top-5 Acc (%)↑Top-1 Acc (%)↑Top-5 Acc (%)↑
SimCLR10065.187.333.158.2
MoCoV210061.786.031.155.1
PCL10062.186.330.554.7
SwAV10064.888.333.457.8
BYOL10065.888.533.055.4
DINO10065.388.333.255.5
S-PCL(Ours)10070.290.337.261.9
RefCL(Ours)10065.588.135.260.1
ShadCL(Ours)10062.586.532.155.9
MidVCL(Ours)10070.590.437.062.7
SimCLR40077.293.840.866.9
MoCoV240077.093.841.668.1
PCL40074.091.937.863.3
SwAV40076.893.638.364.2
BYOL40075.892.642.967.6
DINO40077.193.638.062.9
S-PCL(Ours)40071.890.937.063.1
RefCL(Ours)40077.293.440.065.4
ShadCL(Ours)40071.890.934.960.6
MidVCL(Ours)40072.490.937.161.7
S-PCL+MoCoV2(Ours)40078.094.243.969.4
MidVCL+MoCoV2(Ours)40077.894.243.869.4
+ +Segmentaion and Depth Estimation Besides linear classification, which is a semantic and image-level task, it is also essential to evaluate representations on pixel-level tasks such as segmentation and depth estimation. Therefore, we conduct experiments on various benchmark datasets to assess performance in those tasks. The models evaluated in this part are pretrained for 400 epochs. + +For segmentation tasks, we fine-tune the pretrained models with a segmentation head on the ADE20K and Cityscapes datasets. The mean Intersection over Union (mIoU) scores are reported in Table 2. Our MidVCL approach outperforms all other methods on both datasets, and + +![](images/0458a01f5516eaf52afccaf0fe1517ecb0fc644f32038ec808a96a4b5b167404.jpg) +Figure 4. Visualization of elements in clusters produced by S-PCL after trained on STL10 for 400 epochs. Each block represents a cluster and 10 randomly selected images (Top) and shape silhouettes (bottom) are presented. Images in each cluster have similar shape irrespective to their ground truth category, indicating the S-PCL's capability of grouping objects with similar shape. + +![](images/168d5ff2fd24f43901d3c626fa726d9125c97bb1d16c8b955c33113880ca187d.jpg) +Figure 5. A. AMI between cluster assignment of S-PCL trained on ImageNet-100 and ground truth label. The AMI increases at the early stage and decreases later on. B. AMI between cluster assignment of PCL and S-PCl+PCL trained on ImageNet-100 and ground truth label. The AMI of PCL keeps increasing. S-PCL warm-up moves the curve upward. C and D. Linear classification accuracy on two datasets for PCL, S-PCL and S-PCL+PCL. S-PCL suffers rapid saturation but is able to serve as a warm-up strategy to accelerate training. + +![](images/02f3f563a7ec45b971cb8a2dd1e61e872cdf5869f05c9420eba463736d0b22c6.jpg) + +![](images/07cf7d1ea6bcc49da8d94e76a5576333b6d0190c7dba9d98aecc618641ff3602.jpg) + +![](images/6946beee34b2397b72c01591990984866518198502855e7f1b229e3432e8258e.jpg) + +the S-PCL is on the second place. This superior performance is attributable to the incorporation of shape silhouette masks in both MidVCL and S-PCL, which provide figure-ground information critical for segmentation tasks. Additionally, we observe that ReflCL ranks the third, presumably because the reflectance intrinsic image segregates instances based on albedo, thereby facilitating more accurate boundary detection. Also, it turns out that ReflCL performs poorly on depth estimation, suggesting that solely reflectance is insufficient to distinguish foreground and background although the boundaries of instances are clearer. In contrast, ShadCL ends up with only nuanced improvements in segmentation performance. + +For depth estimation, we fine-tune the pretrained models with an additional head using the NYU Depth V2 dataset. The relative errors (REL) and root mean errors (RME) are presented in Table 3. MidVCL and S-PCL continue + +to achieve the highest performance among all evaluated approaches. Notably, ShadCL ranks third, indicating that shading information is beneficial for depth estimation tasks. However, the advantages of incorporating reflectance information in this context appear to be limited. + +Overall Our evaluations on these downstream tasks demonstrate that shape prototypes, reflectance intrinsic image, and shading intrinsic image facilitate representation learning in different ways. The shape prototype helps all three downstream tasks, especially superior in pixel-level tasks (segmentation and depth estimation). However, it is only beneficial in the early stage for the image-level classification task and saturates rapidly. Additionally, the reflectance intrinsic image helps classification and segmentation but not depth estimation. On the other hand, the shading intrinsic image enhances depth estimation thanks to its implicit information of 3D shape and illuminance, but not classification and seg- + +mentation, where spatial information is not that important. Among three perceptual constructs, we find the shape prototype yields the most interesting results, thus we conduct further studies to understand the underlying mechanism of the shape prototype. + +Table 2. Segmentation results. The mIoU(%) values on ADE20K and Cityscapes are reported. The best performances are in bold. + +
EpochsADE20K mIoU(%)↑Cityscapes mIoU(%)↑
SimCLR40030.463.4
MoCoV240030.463.1
PCL40030.564.1
SwAV40030.261.9
BYOL40030.961.6
DINO40030.261.0
S-PCL(Ours)40031.868.3
RefCL(Ours)40031.865.5
ShadCL(Ours)40031.064.4
MidVCL(Ours)40032.168.3
+ +Table 3. Depth estimation results. The relative error and root mean error on NYU Depth V2 dataset are reported. The best performances are in bold. + +
EpochsREL↓RME↓
SimCLR4000.25450.1420
MoCoV24000.25500.1434
PCL4000.25870.1437
SwAV4000.25860.1459
BYOL4000.25430.1458
DINO4000.25950.1451
S-PCL(Ours)4000.24280.1381
RefCL(Ours)4000.25330.1420
ShadCL(Ours)4000.24500.1399
MidVCL(Ours)4000.24060.1354
+ +# 4.2. Cluster Analysis + +Clusters of similar shapes To gain a deeper understanding of shape prototypes influence the representation learning, we first analyze the clusters produced by S-PCL. As illustrated in Figure.4, we visualize STL-10 images randomly selected from high-density clusters after 400 training epochs. It is evident that the elements within each cluster exhibit similar global shapes, demonstrating the S-PCL's ability to capture shape features and utilize shape as a structural regularity to organize the representation space. + +Cluster assignment diverges from semantic category in late stages To study the early superior performance and fast saturation of S-PCL observed in the linear classifica + +Table 4. Results of linear classification and segmentation of hybrid models. For hybrid methods (denoted as "X+Y"), training consists of 100 epochs using method X followed by 300 epochs using method Y. We can observe that constructing representation with perceptual inductive bias first will boost the performance of contrastive representation learning on both segmentation and classification tasks. The best two performances are in bold. + +
EpochsImageNet-100 Top-1 Acc(%)↑ADE20K mIoU(%)↑Depth RME(%)↓
PCL40074.030.50.1437
MoCoV240077.030.40.1434
S-PCL(Ours)40071.831.80.1381
S-PCL+PCL(Ours)40076.031.90.1373
S-PCL+MoCoV2(Ours)40078.031.90.1398
MidVCL(Ours)40074.032.10.1354
MidVCL+MoCoV2(Ours)40077.932.00.1409
+ +tion evaluation, we analyze the Adjusted Mutual Information (AMI) between K-Means cluster assignments of S-PCL and the categorical ground truth. A higher AMI indicates greater alignment between cluster assignments and semantic categories. As illustrated in Figure.5A, the AMI increases during the initial training stages but progressively decreases as the number of epochs increases. This trend suggests that clustering based solely on shape leads to a divergence from the semantic categories. In other words, it suggests that relying exclusively on shape is insufficient for effective object classification, which greatly requires semantic information. However, the shape shows enormous advantages at the early stage. These findings motivate the development of a hybrid framework that employs S-PCL and MidVCL as a warm-up strategy in the early training stages, which is detailed in §4.4. + +# 4.3. Perceptual inductive bias is beneficial for both image-level task and pixel-level task + +Accelerated representation learning for classification In the results of linear classification (§4.1) and cluster analysis (§4.2), we observe tremendous benefits of incorporating shape prototypes. We attribute this phenomenon to an analogy with the accelerated learning of new nouns grouped by shapes during human infancy. Consequently, we investigate a hybrid framework that combines perceptual inductive bias and semantic representation training. Specifically, we propose training with S-PCL or MidVCL for the first 100 epochs and then switching to traditional contrastive learning frameworks (denoted as S-PCL+X and MidVCL+X respectively), allowing the model to discover additional structural regularities and develop a better representation space. Figure.5B shows the AMI results of PCL and S-PCL+PCL. It can be observed that the S-PCL warm-up moves the AMI curve upward. Figure.5C and D show the linear classifi + +cation accuracies for PCL and S-PCL+PCL on ImageNet-100 and STL10 datasets versus the number of pretraining epochs. For a fair comparison, all frameworks are trained using identical parameter settings and learning rate schedules. The results demonstrate that the S-PCL+PCL method achieves accuracy levels comparable to PCL while requiring less than half the number of training epochs on both ImageNet-100 and STL-10 datasets. The curve of hybrid frameworks of MidVCL is included in supplementary material and it demonstrates the same trend. Overall, our experimental results demonstrate that the sequential learning of intermediate shape representations followed by high-level semantic representations is beneficial for semantic tasks, particularly in accelerating the learning process. + +Sequential training resolves the trade-off between image-level and pixel-level tasks Beyond the classification task, hybrid frameworks demonstrate their superiority also in segmentation task. As reported in Table.4, hybrid frameworks achieve strong performance in both classification and segmentation tasks, unlike methods relying only on either semantic or perceptual representations. Although hybrid frameworks are not as good as S-PCL and MidVCL in depth estimation, they still significantly outperform baseline methods. These results indicate that the additional learning with perceptual inductive bias before semantic contrastive learning provides a solution to the tradeoff mentioned in related works $\S 2$ + +# 4.4. Concentration on Object Contour and Shape Bias + +Shape silhouette forces model to focus on object contours Next, we study how shape representations vary the model's prediction decision-making process. To this end, we employed the SmoothGrad method to generate sensitivity maps, where highlighted regions contribute significantly to the model's inferences. We show some example images in the validation set of ImageNet-100 in Figure.6. Notably, comparing the sensitivity maps of S-PCL and others, we find that the semantic representations lead to a stronger concentration on local features of the object, which is a shortcut for classification. While the shape representations guide the model to focus on the contour of the object. + +One may argue that although MoCoV2 and PCL focus more on texture, the silhouettes of objects can still be observed to some degree. To address this, we also analyze sensitivity maps of style-transferred images, that is, preserving object shape while altering texture, in addition to the original images, see Figure.7. After enormous style changes, MoCov2 and PCL are detracted by those irrelevant local features while S-PCL still has a strong concentration on the contours of the objects. + +![](images/f0f4466150da650095c0deb057d5bb8cbe66c8726a3303603e8de7966812c7cd.jpg) +Figure 6. Smoothgrad sensitivity on different images. + +![](images/6537bc662e66ecc7bfaacbb4e3451f9e8ef9404c472b89e945f8116b95c9c41a.jpg) +Figure 7. Smoothgrad sensitivity map on the image in different styles. + +Shape bias score So far, we have found that the shape prototype can increase the shape bias of the model, alleviating a well-known machine's misalignment with humans. To quantitatively evaluate the shape bias of the model, we test our model on the model-vs-human benchmark. The model-vs-human benchmark is a protocol that is able to evaluate the 2D shape bias of a model. With a set of images with conflict cues, e.g. a cat shape with elephant skin texture, we can know which is the predominant visual cue for object recognition between shape and texture. Figure.8 demonstrates shape score across different frameworks. Importantly, the proposed SPCL is the only model that uses shape as the predominant cue, i.e. shape bias score greater than 0.5. RelfCL shows improvements compared to the rest of the approaches but it is still biased towards texture (fraction of shape decisions less than 0.5). + +Out-of-distribution robustness Many studies reveal that the model with high shape bias will also have higher robustness to out-of-distribution (OOD) data. To this end, we test models on the 17 OOD datasets included in the model-vs-human benchmarks. Figure.9 presents average classifica + +![](images/88fbd229946d84f976f1d6f67e6ad37eab9daed2c410ffc75edde96e40c279f4.jpg) +Figure 8. The shape bias score. The value on the top represents the fraction of shape decisions and the value at the bottom represents the fraction of texture decisions. Two values sum to 1 and 0.5 means neutrality, indicating no bias towards either shape or texture. Markers are the fraction of certain shape categories and vertical lines are the mean fraction across shape categories. The star markers represent our proposed approaches. + +tion accuracies across all 17 OOD datasets in the model-vs-human benchmark. The full results are included in the supplementary material. S-PCL demonstrates statistically significant improvements in OOD accuracy compared to baseline models. This suggests that leveraging shape information not only aids in better generalization on seen data but also enhances the model's performance on unfamiliar inputs. + +# 5. Discussion + +In this work, we aim to integrate inductive biases from early and intermediate perceptual constructs into contrastive learning rather than relying solely on high-level semantic information. To achieve this, we maximize the mutual information between full-detail images and one or more perceptual constructs. Our findings reveal that inductive bias from global shape prototypes not only accelerates early-stage classification but also improves segmentation and depth estimation while enhancing the model's overall shape bias. Reflectance intrinsic images facilitate representation learning for classification and segmentation, although they do not benefit depth estimation, whereas shading intrinsic images enhance depth estimation without aiding classification or segmentation. Combining all three perceptual con + +![](images/6e2d2c1a6dd1ac62a8f4ff54c40dcc615f64ebe9da05f5cee1625504e4cd6601.jpg) +Figure 9. Mean accuracy across 17 OOD dataset in model- vs-human benchmark. Bars are positioned in the descending order. The error bar is the standard error of the mean (SEM). A paired T-test is conducted. \*\* denotes $p < 0.01$ , $*$ denotes $p < 0.05$ , n.s. denotes "not significant" $p > 0.05$ . + +structures enables us to leverage the unique advantages of each. Moreover, employing sequential training—where the model first harnesses perceptual inductive biases before applying semantic contrastive learning—effectively mitigates the trade-off between image-level and pixel-level tasks. Overall, our results indicate that incorporating perceptual constructs generated by early and intermediate visual processing into self-supervised learning can help networks develop distinct intermediate representations that improve downstream visual tasks, offering a promising avenue for future research in contrastive learning. + +# 6. Limitation + +We developed distinct approaches for specific perceptual constructs, focusing on shape, figure-ground, and intrinsic images. These constructs were chosen due to their close association with fundamental visual tasks like classification, segmentation, and depth estimation. However, additional mid-level perceptual constructs merit further exploration. Moreover, our experiments were conducted on a small model and limited datasets because of computing power constraints, underscoring the need to address scalability in future research. + +# 7. Acknowledgement + +This work was supported by NSF CISE RI 1816568 and NIH R01 EY030226-01A1. This work was also supported by the CMU HURAY program. + +# References + +[1] Edward H Adelson and Alex P Pentland. The perception of shading and reflectance. Perception as Bayesian inference, 409:423, 1996. +[2] Harry Barrow, J Tenenbaum, A Hanson, and E Riseman. Recovering intrinsic scene characteristics. Comput. vis. syst, 2 (3-26):2, 1978. +[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments, 2021. +[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[5] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223, Fort Lauderdale, FL, USA, 2011. PMLR. +[6] Hans P Op de Beeck, Katrien Torfs, and Johan Wagemans. Perceived shape similarity among unfamiliar objects and the organization of the human object vision pathway. Journal of Neuroscience, 28(40):10111-10123, 2008. +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. +[8] James H Elder and Ljiljana Velisavljevic. Cue dynamics underlying rapid detection of animals in natural scenes. J Vis, 9(7):7, 2009. +[9] Paul Gavrikov, Jovita Lukasik, Steffen Jung, Robert Geirhos, Bianca Lamm, Muhammad Jehanzeb Mirza, Margret Keuper, and Janis Keuper. Are vision language models texture or shape biased and can we steer them?, 2024. +[10] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. On the surprising similarities between supervised and self-supervised models, 2020. +[11] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. In Advances in Neural Information Processing Systems, pages 23885-23899. Curran Associates, Inc., 2021. +[12] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. Advances in Neural Information Processing Systems, 34:23885-23899, 2021. +[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, 2022. +[14] Lisa Gershkoff-Stowe and Linda B Smith. Shape and the first hundred nouns. Child Dev, 75(4):1098-1114, 2004. + +[15] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In 2009 IEEE 12th International Conference on Computer Vision, pages 2335-2342. IEEE, 2009. +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. +[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. +[18] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers, 2024. +[19] V A Lamme. The neurophysiology of figure-ground segregation in primary visual cortex. J Neurosci, 15(2):1605-1615, 1995. +[20] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. +[21] Min Seok Lee, Wooseok Shin, and Sung Won Han. Tracer: Extreme attention guided salient object tracing network, 2022. +[22] Tai Sing Lee, David Mumford, Richard Romero, and Victor A.F. Lamme. The role of the primary visual cortex in higher level vision. Vision Research, 38(15):2429-2454, 1998. +[23] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. +[24] Tianqin Li, Ziqi Wen, Yangfan Li, and Tai Sing Lee. Emergence of shape bias in convolutional neural networks through activation sparsity, 2023. +[25] Ralph Linsker. Self-organization in a perceptual network. Computer, 21(3):105-117, 1988. +[26] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers, 2021. +[27] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020. +[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. +[29] Johan Wagemans, Joeri De Winter, Hans Op de Beeck, Annemie Ploeger, Tom Beckers, and Peter Vanroose. Identification of everyday objects on the basis of silhouette and outline versions. Perception, 37(2):207-244, 2008. +[30] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. +[31] Ziqi Wen, Tianqin Li, and Tai Sing Lee. Does resistance to style-transfer equal shape bias? evaluating shape bias by distorted shape, 2023. + +[32] Zhirong Wu, Yuanjun Xiong, Stella Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance-level discrimination, 2018. +[33] Yike Yuan, Xinghe Fu, Yunlong Yu, and Xi Li. Densedino: boosting dense self-supervised learning with token-based point-level consistency. arXiv preprint arXiv:2306.04654, 2023. +[34] Hong Zhou, Howard S Friedman, and Rüdiger Von Der Heydt. Coding of border ownership in monkey visual cortex. Journal of Neuroscience, 20(17):6594-6611, 2000. +[35] Karl Zipser, Victor A. F. Lamme, and Peter H. Schiller. Contextual modulation in primary visual cortex. Journal of Neuroscience, 16(22):7376-7389, 1996. \ No newline at end of file diff --git a/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/images.zip b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ffa1896e4d6ab665ab17e5798b139b5a70fdb500 --- /dev/null +++ b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61d62fc4415e6321699953e76803fa7700b3af75646ef85b64566f7aec6f7f53 +size 667399 diff --git a/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/layout.json b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..119db9fec8d93f12dcc926af44bd23eb27d4c929 --- /dev/null +++ b/2025/Perceptual Inductive Bias Is What You Need Before Contrastive Learning/layout.json @@ -0,0 +1,5553 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 533, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 533, + 122 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 533, + 122 + ], + "type": "text", + "content": "Perceptual Inductive Bias Is What You Need Before Contrastive Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 143, + 527, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 143, + 527, + 177 + ], + "spans": [ + { + "bbox": [ + 83, + 143, + 527, + 177 + ], + "type": "text", + "content": "Junru Zhao Tianqin Li Dunhan Jiang Shenghao Wu Alan Ramirez Tai Sing Lee Carnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 178, + 496, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 178, + 496, + 190 + ], + "spans": [ + { + "bbox": [ + 111, + 178, + 496, + 190 + ], + "type": "text", + "content": "{junruz, tianqinl, dunhanj, shenghaw, alanrami, taislee}@andrew.cmu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 218, + 200, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 218, + 200, + 231 + ], + "spans": [ + { + "bbox": [ + 151, + 218, + 200, + 231 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 247, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 521 + ], + "type": "text", + "content": "David Marr's seminal theory of human perception stipulates that visual processing is a multi-stage process, prioritizing the derivation of boundary and surface properties before forming semantic object representations. In contrast, contrastive representation learning frameworks typically bypass this explicit multi-stage approach, defining their objective as the direct learning of a semantic representation space for objects. While effective in general contexts, this approach sacrifices the inductive biases of vision, leading to slower convergence speed and learning shortcut resulting in texture bias. In this work, we demonstrate that leveraging Marr's multi-stage theory—by first constructing boundary and surface-level representations using perceptual constructs from early visual processing stages and subsequently training for object semantics—leads to " + }, + { + "bbox": [ + 55, + 247, + 296, + 521 + ], + "type": "inline_equation", + "content": "2x" + }, + { + "bbox": [ + 55, + 247, + 296, + 521 + ], + "type": "text", + "content": " faster convergence on ResNet18, improved final representations on semantic segmentation, depth estimation, and object recognition, and enhanced robustness and out-of-distribution capability. Together, we propose a pretraining stage before the general contrastive representation pretraining to further enhance the final representation quality and reduce the overall convergence time via inductive bias from human vision systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 548, + 135, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 135, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 135, + 561 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Human visual information processing is organized in stages, progressing from contrast and edge detection to perceptual organization and figure-ground segmentation, then to computing surfaces and shapes, and ultimately to object recognition. End-to-end supervised learning in neural networks for object classification in images often relies on the simplest features or shortcuts, bypassing essential intermediate representations, resulting in texture bias. Self-supervised contrastive learning, on the other hand, leverages the inductive bias of semantic segregation by grouping semantically similar scenes and separating dissimilar scenes in latent space. This also leads to representations that ac" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 333, + 217, + 537, + 331 + ], + "blocks": [ + { + "bbox": [ + 333, + 217, + 537, + 331 + ], + "lines": [ + { + "bbox": [ + 333, + 217, + 537, + 331 + ], + "spans": [ + { + "bbox": [ + 333, + 217, + 537, + 331 + ], + "type": "image", + "image_path": "41f7a0da2d88bc88904bedce87effa10ba803a5d7fd8f9f2fb3bd5e71c91eda3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 338, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 416 + ], + "type": "text", + "content": "Figure 1. We propose to leverage perceptual constructs from early vision processing stages to improve the contrastive learning pipeline. By introducing a pre-pretraining stage before general contrastive learning (CL), we inject inductive bias into the contrastive representation learning. We demonstrate that such inductive bias is beneficial to the representation quality and provides significantly convergence speed-up the contrastive learning." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 472 + ], + "type": "text", + "content": "cess rich semantic information without building intermediate representations sensitive to surfaces and boundaries first." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "content": "In this work, we evaluate whether perceptual constructs from early visual processing, such as figure-ground segmentation and intrinsic image decomposition, can serve as additional inductive biases for contrastive learning. Evidence of figure-ground organization has been found in the primary visual cortex and secondary visual cortex in primate visual systems [19, 22, 34, 35]. This early perceptual construct distinguishes figures against the ground, assigns border ownership to surfaces, and enables the construction of a neural representation of local shapes and global shapes along the ventral visual hierarchy. The perception of shapes could be innate or at least develop very fast [8, 29]. Furthermore, categorization and identification of unseen objects are considered to rely on the similarity of the overall shape, conceptualized through shape prototypes [6]. The recognition of shapes was found to be parallel to the development of vocabularies in language in infancy. Infants can learn new nouns faster when nonce objects are organized in shape, implicating a linked development between labeling objects linguistically and shape prototypes in the visual" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9621" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 144 + ], + "type": "text", + "content": "system [14]. We conjecture that shape prototypes learned as cluster means of shape silhouettes obtained from figure-ground segregation can allow contrastive learning to learn better intermediate representations for a variety of visual tasks, such as object recognition, segmentation as well as depth estimation in the context of this paper." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 144, + 294, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 144, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 57, + 144, + 294, + 311 + ], + "type": "text", + "content": "Another set of early perceptual constructs is called intrinsic images [2]. Intrinsic image decomposition [1, 15] is a generalization of Retinex developed by Edwin Land [20] to explain color constancy by discounting illumination effects to infer the reflectance or color of the objects. Intrinsic image decomposition factorizes an image as a product of a reflectance image and a shading image. Reflectance images reflect surface material properties and govern our perception of color and lightness. Shading images reflect the interaction of illumination and the 3D surface shape and are thus rich in implicit 2.5D information and affect our depth perception. What kinds of visual tasks would benefit from intermediate representations learned under the guidance of the inductive bias of these intrinsic images?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 312, + 294, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 312, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 312, + 294, + 645 + ], + "type": "text", + "content": "This work shows that global shapes of figures, particularly through a set of remembered prototypes, provide an inductive bias that enhances contrastive learning, resulting in learned representations that improve object recognition, segmentation, and depth estimation, with an increase in the shape bias of the model. We find that the inductive bias from the shape prototypes increases the speed of representational learning at the beginning, as in the early development of human infants, but ultimately has to be disabled to allow the learning of more refined representations to accommodate a greater variety of semantics. Thus, we propose a hybrid coarse-to-fine inductive bias that allows contrastive learning to learn intermediate representations first with the help of the prototypes and then refine the representation based on semantics. This results in a 2x faster representation learning. On the other hand, intrinsic reflectance images enhance the learned representations for segmentation and object recognition but do not benefit depth estimation. Conversely, intrinsic shading images improve depth estimation capabilities but with little impact on segmentation and object recognition. Finally, we demonstrate that integrating these three perceptual constructs provides additional performance benefits. Thus, we demonstrate that inductive biases derived from early perceptual constructs can differentially or collectively promote the contrastive learning of intermediate representations with greater sensitivity to surface depths and object shapes and better performance in downstream visual tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 658, + 145, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 658, + 145, + 669 + ], + "spans": [ + { + "bbox": [ + 58, + 658, + 145, + 669 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 679, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 679, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 679, + 294, + 712 + ], + "type": "text", + "content": "Contrastive learning is one of the most commonly used frameworks in discriminative self-supervised learning. It inherently aligns with the Infomax principle [25] by aim-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 251 + ], + "type": "text", + "content": "ing to maximize the mutual information between semantically similar scenes, which are usually different views of the same image. This aim is realized by maximizing InfoNCE loss function [28, 32]. InfoNCE is designed to maximize a lower bound on the mutual information between representations of positive view pairs while distinguishing them from negative view pairs. The views are usually generated by different data augmentations. Many studies [4, 17] emphasize the importance of data augmentations as contrastive learning requires intensive data augmentation to learn good semantic representations. Some works combine contrastive learning with clustering [3, 23]. In this kind of framework, the concept of prototype is introduced to organize the representation space in order to have more compact semantic clusters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 252, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 371 + ], + "type": "text", + "content": "After self-supervised pretraining, the representations are commonly evaluated by downstream tasks such as classification, segmentation and depth estimation. Many works [27, 30, 33] point out that there exists a trade-off between image-level classification (semantics) and pixel-wise classification or segmentation (perceptual grouping) in discriminative models. In other words, representations learned by state-of-the-art discriminative models often fail to perform effectively across both high-level semantic tasks and intermediate-level perceptual tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 372, + 553, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 562 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 562 + ], + "type": "text", + "content": "While humans rely heavily on intermediate perceptual constructs like object shapes in object recognition, texture bias is a well-known misalignment between CNN-based models and humans [10, 11, 13]. Many works endeavor to solve this misalignment by different approaches including 1) using data augmentation methods to vary texture, e.g. texture synthesis, style transfer [12, 13, 31], 2) changing to vision transformer [26], or 3) using more complex multimodal generative frameworks [9, 18]. Improved shape bias results in improvement of recognition robustness and domain generalization [13, 24]. However, a recent study [18] argues that most discriminative models are still strongly texture-biased and less robust to tremendous style changes. We attribute this phenomenon to the focus of discriminative models on learning semantic representations while overlooking intermediate-level representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 572, + 379, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 572, + 379, + 585 + ], + "spans": [ + { + "bbox": [ + 316, + 572, + 379, + 585 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 594, + 553, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 594, + 553, + 627 + ], + "spans": [ + { + "bbox": [ + 316, + 594, + 553, + 627 + ], + "type": "text", + "content": "In this part, we will introduce how we incorporate global shape of figures in a prototypical manner and the intrinsic images as views into contrastive learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 643, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 643, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 643, + 553, + 712 + ], + "type": "text", + "content": "Shape prototypical contrastive learning (S-PCL) Inspired by the fact that humans rely on the similarity of the overall shape envelope for object recognition, we design a contrastive learning framework that is capable of learning the concept of a group of similar shapes, i.e. shape prototypes. To represent the global shape of an object," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "9622" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 286, + 376 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 286, + 376 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 286, + 376 + ], + "type": "image", + "image_path": "c5affb8a114f46e6a133aa6e0d18d5b1b20d813b35cdce9b51de366d171f4f91.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 387, + 296, + 487 + ], + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 487 + ], + "type": "text", + "content": "Figure 2. Illustration of Shape Prototypical Contrastive Learning (S-PCL) training framework The original images are input to the encoder and the offline-generated shape silhouettes generated are input to the momentum encoder. The K-Means is performed on the embedding of shape silhouettes. Similar shapes are clustered together and the centroid of each cluster is considered to be the shape prototype. The objective function, ShapeProtoNCE, is calculated on the representation of the original image and its corresponding shape prototype." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "we use the figure-ground shape silhouette as presented in Figure.2. Our framework is built upon prototypical contrastive learning (PCL)[23], which involves an online k-means clustering on the representation of the entire training data and considers the centroids of clusters as prototypes. In shape prototypical contrastive learning, the clustering is performed on the representation of shape silhouettes, as illustrated in Figure.2. The representation learning is guided by the clusters of shapes. With a set of images " + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "inline_equation", + "content": "X = \\{x_{1}, x_{2}, \\dots, x_{n}\\}" + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": " and matched shape silhouettes " + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "inline_equation", + "content": "M = \\{m_{1}, m_{2}, \\dots, m_{n}\\}" + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": ", they are encoded by the same encoder to get their representations " + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "inline_equation", + "content": "V = \\{v_{1}, v_{2}, \\dots, v_{n}\\}" + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "inline_equation", + "content": "U = \\{u_{1}, u_{2}, \\dots, u_{n}\\}" + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": ", respectively. K-Means clustering is employed on the set " + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": ", producing " + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": " clusters and a set of centroids, " + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "inline_equation", + "content": "S = \\{s_{1}, s_{2}, \\dots, s_{K}\\}" + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": ". The centroid of the cluster is representative of elements in the cluster, and thus is" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 333, + 72, + 542, + 211 + ], + "blocks": [ + { + "bbox": [ + 333, + 72, + 542, + 211 + ], + "lines": [ + { + "bbox": [ + 333, + 72, + 542, + 211 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 542, + 211 + ], + "type": "image", + "image_path": "4ab995d317b15ef71b93f419b696001117691500262389085bda80f9530984c4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 219, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 219, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 555, + 285 + ], + "type": "text", + "content": "Figure 3. Illustration of the training framework that incorporates the intrinsic image as a view The original image is input to the encoder and one intrinsic image decomposed by Retinex algorithm is input to the momentum encoder. The InfoNCE between the representation of the original image and that of the intrinsic image is computed." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "type": "text", + "content": "considered as the shape prototype. The mutual information between the representation of the original image " + }, + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "type": "text", + "content": " and the representation of the shape prototype " + }, + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "type": "inline_equation", + "content": "s_{i}" + }, + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "type": "text", + "content": " is maximized by optimizing the InfoNCE objective function:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 329, + 375, + 555, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 375, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 329, + 375, + 555, + 407 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S h a p e P r o t o N C E}} = \\sum_ {i = 1} ^ {n} - \\log \\frac {\\exp \\left(v _ {i} \\cdot s _ {p}\\right) / \\phi_ {p})}{\\sum_ {j = 0} ^ {r} \\exp \\left(v _ {i} \\cdot s _ {j} / \\phi_ {j}\\right)} \\tag {1}", + "image_path": "0d0a104c2b38f0dd53201319b055080ebd2abd5c89f4fac887c5ea32fe454163.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "inline_equation", + "content": "s_p" + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "text", + "content": " is the prototype of the cluster " + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "text", + "content": " to which " + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "text", + "content": " belongs, " + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "inline_equation", + "content": "s_j, j \\neq p" + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "text", + "content": " are shape prototypes representing other shape clusters, and " + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "inline_equation", + "content": "\\phi_j" + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "text", + "content": " is the concentration estimation of cluster " + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "text", + "content": ". Similar to PCL, we perform clustering for " + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "text", + "content": " times with different " + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "inline_equation", + "content": "K_i, i = 1,2,\\dots,N" + }, + { + "bbox": [ + 313, + 415, + 556, + 511 + ], + "type": "text", + "content": " and average the losses. The final loss function used is the combination of InfoNCE between the normally augmented view and the shape loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 352, + 521, + 555, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 521, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 352, + 521, + 555, + 554 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {I n f o N C E}} + \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\text {S h a p e P r o t o N C E}, K _ {i}} \\tag {2}", + "image_path": "96b01a01885bd904569c021e12e1f2678e487ea78c45dc11a05a61d6ac99c206.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Intrinsic image as a view in contrastive learning Unlike shape prototypes, we treat the intrinsic image, either reflectance or shading, as an augmented view. We maximize the mutual information between the representations of normally augmented views and those of reflectance or shading images by optimizing the InfoNCE objective function, as illustrated in Figure 3. The final loss function is the sum of the standard InfoNCE loss between representations of two augmented views and the InfoNCE loss between representations of one augmented view and one intrinsic image. The approaches that incorporate reflectance or shading intrinsic images are termed Reflectance Contrastive Learn" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9623" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "content": "ing (RefCL) and Shading Contrastive Learning (ShadCL), respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 295, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 160 + ], + "type": "text", + "content": "Combination of shape and intrinsic images We also integrate the shape and the intrinsic image together according to equation (3). We call the combined approach mid-level vision contrastive learning (MidVCL)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 168, + 295, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 211 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {I n f o N C E}} + \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {\\text {S h a p e P r o t o N C E}, K _ {i}} + \\alpha \\mathcal {L} _ {\\text {S h a d}} + \\beta \\mathcal {L} _ {\\text {R e f l}} \\tag {3}", + "image_path": "560c304f18f573c0cc0fcb82ecc5e916651a365b0e02528618a9847c09a08d6a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 211, + 295, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 295, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 211, + 295, + 236 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 211, + 295, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 211, + 295, + 236 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 211, + 295, + 236 + ], + "type": "text", + "content": " is the weight of the shading and reflectance losses, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "text", + "content": "Implementation details The experiments are conducted on the dataset ImageNet-100" + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "text", + "content": ", a subset of ImageNet-1K [7], and STL-10 [5]. The silhouette mask of the object is generated offline using TRACER [21]. We decompose a full-detail image to a reflectance intrinsic image and a shading image by Retinex algorithm [1, 15]. The ResNet18 [16] is adopted as the encoder and the output dimension of the last fully-connected layer is 256-D. We follow the data augmentation methods in previous works [4, 17, 23]. These data augmentations are applied to only the original image and the preprocessing for the shape silhouette and the intrinsic image only contains resizing and normalization. More details can be found in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 108, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 108, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 108, + 426 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 434, + 203, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 203, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 203, + 447 + ], + "type": "text", + "content": "4.1. Representation Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 451, + 296, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 500 + ], + "type": "text", + "content": "First and foremost, we evaluate representation learned under the guidance of intermediate perceptual constructs through multiple downstream visual tasks, including linear classification, segmentation, and depth estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 506, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 686 + ], + "type": "text", + "content": "Linear classification and transfer learning to ImageNet-1k We trained a linear classifier on frozen representations using the entire labeled training set of ImageNet-100 and STL-10, and here we report the Top-1 and Top-5 accuracies of ImageNet-1k on the validation sets. The results on STL-10 are included in the supplementary material. As illustrated in Table.1, S-PCL significantly outperforms the baselines during the early training stages (e.g., at 100 epochs) on both datasets. However, when training is extended to 400 epochs, the performance gains of S-PCL diminish, and it is even surpassed by the baseline approaches significantly on ImageNet-100, indicating a potential faster saturation of S-PCL compared to others. In contrast, RefCL maintains consistent performance improvements between 100 and 400 epochs, suggesting that reflectance information effectively" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 214 + ], + "type": "text", + "content": "facilitates object recognition. Conversely, ShadCL, with inductive bias from shading, provides only a tiny gain during the early stages of training and may hinder object recognition when trained for a large number of epochs. MidVCL is similar to S-PCL on the classification task, being the best at epoch 100 while plateauing when the number of epoch increases to 400. We also test generalizability by transferring to a different dataset. This is performed by the linear classification on ImageNet-1k which has 10x more classes than ImageNet-100. The transfer learning results reported in Table.1 display the similar trend as the linear classification." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 312, + 570, + 534 + ], + "blocks": [ + { + "bbox": [ + 313, + 225, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 292 + ], + "type": "text", + "content": "Table 1. Linear classification and ImageNet-1k transfer learning accuracies (\\%) of different approaches trained for 100 and 400 epochs on the ImageNet-100. For hybrid methods (denoted as \"X+Y\"), training consists of 100 epochs using method X followed by 300 epochs using method Y. The best-performing results are highlighted in bold." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 312, + 570, + 534 + ], + "lines": [ + { + "bbox": [ + 316, + 312, + 570, + 534 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 570, + 534 + ], + "type": "table", + "html": "
EpochsImageNet-100ImageNet-1k
Top-1 Acc (%)↑Top-5 Acc (%)↑Top-1 Acc (%)↑Top-5 Acc (%)↑
SimCLR10065.187.333.158.2
MoCoV210061.786.031.155.1
PCL10062.186.330.554.7
SwAV10064.888.333.457.8
BYOL10065.888.533.055.4
DINO10065.388.333.255.5
S-PCL(Ours)10070.290.337.261.9
RefCL(Ours)10065.588.135.260.1
ShadCL(Ours)10062.586.532.155.9
MidVCL(Ours)10070.590.437.062.7
SimCLR40077.293.840.866.9
MoCoV240077.093.841.668.1
PCL40074.091.937.863.3
SwAV40076.893.638.364.2
BYOL40075.892.642.967.6
DINO40077.193.638.062.9
S-PCL(Ours)40071.890.937.063.1
RefCL(Ours)40077.293.440.065.4
ShadCL(Ours)40071.890.934.960.6
MidVCL(Ours)40072.490.937.161.7
S-PCL+MoCoV2(Ours)40078.094.243.969.4
MidVCL+MoCoV2(Ours)40077.894.243.869.4
", + "image_path": "801c8ccf1f96c52278bd2d3aa523bc28dc2bd941a38b95fb0992972c45d73d56.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 653 + ], + "type": "text", + "content": "Segmentaion and Depth Estimation Besides linear classification, which is a semantic and image-level task, it is also essential to evaluate representations on pixel-level tasks such as segmentation and depth estimation. Therefore, we conduct experiments on various benchmark datasets to assess performance in those tasks. The models evaluated in this part are pretrained for 400 epochs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "For segmentation tasks, we fine-tune the pretrained models with a segmentation head on the ADE20K and Cityscapes datasets. The mean Intersection over Union (mIoU) scores are reported in Table 2. Our MidVCL approach outperforms all other methods on both datasets, and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "text", + "content": "1The dataset used here is reorganized by adjusting the train-validation split to a 7:3 ratio to intensify the challenge" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9624" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 549, + 276 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 549, + 276 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 549, + 276 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 549, + 276 + ], + "type": "image", + "image_path": "0458a01f5516eaf52afccaf0fe1517ecb0fc644f32038ec808a96a4b5b167404.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 285, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 54, + 285, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 555, + 321 + ], + "type": "text", + "content": "Figure 4. Visualization of elements in clusters produced by S-PCL after trained on STL10 for 400 epochs. Each block represents a cluster and 10 randomly selected images (Top) and shape silhouettes (bottom) are presented. Images in each cluster have similar shape irrespective to their ground truth category, indicating the S-PCL's capability of grouping objects with similar shape." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 330, + 181, + 432 + ], + "blocks": [ + { + "bbox": [ + 56, + 330, + 181, + 432 + ], + "lines": [ + { + "bbox": [ + 56, + 330, + 181, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 330, + 181, + 432 + ], + "type": "image", + "image_path": "168d5ff2fd24f43901d3c626fa726d9125c97bb1d16c8b955c33113880ca187d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 441, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 54, + 441, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 441, + 555, + 485 + ], + "type": "text", + "content": "Figure 5. A. AMI between cluster assignment of S-PCL trained on ImageNet-100 and ground truth label. The AMI increases at the early stage and decreases later on. B. AMI between cluster assignment of PCL and S-PCl+PCL trained on ImageNet-100 and ground truth label. The AMI of PCL keeps increasing. S-PCL warm-up moves the curve upward. C and D. Linear classification accuracy on two datasets for PCL, S-PCL and S-PCL+PCL. S-PCL suffers rapid saturation but is able to serve as a warm-up strategy to accelerate training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 182, + 330, + 306, + 432 + ], + "blocks": [ + { + "bbox": [ + 182, + 330, + 306, + 432 + ], + "lines": [ + { + "bbox": [ + 182, + 330, + 306, + 432 + ], + "spans": [ + { + "bbox": [ + 182, + 330, + 306, + 432 + ], + "type": "image", + "image_path": "02f3f563a7ec45b971cb8a2dd1e61e872cdf5869f05c9420eba463736d0b22c6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 330, + 428, + 432 + ], + "blocks": [ + { + "bbox": [ + 307, + 330, + 428, + 432 + ], + "lines": [ + { + "bbox": [ + 307, + 330, + 428, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 330, + 428, + 432 + ], + "type": "image", + "image_path": "07cf7d1ea6bcc49da8d94e76a5576333b6d0190c7dba9d98aecc618641ff3602.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 429, + 330, + 553, + 432 + ], + "blocks": [ + { + "bbox": [ + 429, + 330, + 553, + 432 + ], + "lines": [ + { + "bbox": [ + 429, + 330, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 429, + 330, + 553, + 432 + ], + "type": "image", + "image_path": "6946beee34b2397b72c01591990984866518198502855e7f1b229e3432e8258e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 506, + 297, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 506, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 54, + 506, + 297, + 662 + ], + "type": "text", + "content": "the S-PCL is on the second place. This superior performance is attributable to the incorporation of shape silhouette masks in both MidVCL and S-PCL, which provide figure-ground information critical for segmentation tasks. Additionally, we observe that ReflCL ranks the third, presumably because the reflectance intrinsic image segregates instances based on albedo, thereby facilitating more accurate boundary detection. Also, it turns out that ReflCL performs poorly on depth estimation, suggesting that solely reflectance is insufficient to distinguish foreground and background although the boundaries of instances are clearer. In contrast, ShadCL ends up with only nuanced improvements in segmentation performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 665, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 297, + 715 + ], + "type": "text", + "content": "For depth estimation, we fine-tune the pretrained models with an additional head using the NYU Depth V2 dataset. The relative errors (REL) and root mean errors (RME) are presented in Table 3. MidVCL and S-PCL continue" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "text", + "content": "to achieve the highest performance among all evaluated approaches. Notably, ShadCL ranks third, indicating that shading information is beneficial for depth estimation tasks. However, the advantages of incorporating reflectance information in this context appear to be limited." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Overall Our evaluations on these downstream tasks demonstrate that shape prototypes, reflectance intrinsic image, and shading intrinsic image facilitate representation learning in different ways. The shape prototype helps all three downstream tasks, especially superior in pixel-level tasks (segmentation and depth estimation). However, it is only beneficial in the early stage for the image-level classification task and saturates rapidly. Additionally, the reflectance intrinsic image helps classification and segmentation but not depth estimation. On the other hand, the shading intrinsic image enhances depth estimation thanks to its implicit information of 3D shape and illuminance, but not classification and seg-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9625" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "mentation, where spatial information is not that important. Among three perceptual constructs, we find the shape prototype yields the most interesting results, thus we conduct further studies to understand the underlying mechanism of the shape prototype." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 79, + 185, + 273, + 319 + ], + "blocks": [ + { + "bbox": [ + 55, + 143, + 295, + 166 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 295, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 295, + 166 + ], + "type": "text", + "content": "Table 2. Segmentation results. The mIoU(%) values on ADE20K and Cityscapes are reported. The best performances are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 185, + 273, + 319 + ], + "lines": [ + { + "bbox": [ + 79, + 185, + 273, + 319 + ], + "spans": [ + { + "bbox": [ + 79, + 185, + 273, + 319 + ], + "type": "table", + "html": "
EpochsADE20K mIoU(%)↑Cityscapes mIoU(%)↑
SimCLR40030.463.4
MoCoV240030.463.1
PCL40030.564.1
SwAV40030.261.9
BYOL40030.961.6
DINO40030.261.0
S-PCL(Ours)40031.868.3
RefCL(Ours)40031.865.5
ShadCL(Ours)40031.064.4
MidVCL(Ours)40032.168.3
", + "image_path": "04e6f368761e6a6dff7912c8888c24d29f2cdfa505449a06e6efd273f5563ae0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 84, + 397, + 268, + 521 + ], + "blocks": [ + { + "bbox": [ + 55, + 344, + 295, + 377 + ], + "lines": [ + { + "bbox": [ + 55, + 344, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 295, + 377 + ], + "type": "text", + "content": "Table 3. Depth estimation results. The relative error and root mean error on NYU Depth V2 dataset are reported. The best performances are in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 84, + 397, + 268, + 521 + ], + "lines": [ + { + "bbox": [ + 84, + 397, + 268, + 521 + ], + "spans": [ + { + "bbox": [ + 84, + 397, + 268, + 521 + ], + "type": "table", + "html": "
EpochsREL↓RME↓
SimCLR4000.25450.1420
MoCoV24000.25500.1434
PCL4000.25870.1437
SwAV4000.25860.1459
BYOL4000.25430.1458
DINO4000.25950.1451
S-PCL(Ours)4000.24280.1381
RefCL(Ours)4000.25330.1420
ShadCL(Ours)4000.24500.1399
MidVCL(Ours)4000.24060.1354
", + "image_path": "dda5e8da77f4a31fb1a659ba73c04fc6d216ca428b2462b736b7eaa3871b5ec6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 155, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 155, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 155, + 557 + ], + "type": "text", + "content": "4.2. Cluster Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 562, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 295, + 670 + ], + "type": "text", + "content": "Clusters of similar shapes To gain a deeper understanding of shape prototypes influence the representation learning, we first analyze the clusters produced by S-PCL. As illustrated in Figure.4, we visualize STL-10 images randomly selected from high-density clusters after 400 training epochs. It is evident that the elements within each cluster exhibit similar global shapes, demonstrating the S-PCL's ability to capture shape features and utilize shape as a structural regularity to organize the representation space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Cluster assignment diverges from semantic category in late stages To study the early superior performance and fast saturation of S-PCL observed in the linear classifica" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 71, + 555, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 148 + ], + "type": "text", + "content": "Table 4. Results of linear classification and segmentation of hybrid models. For hybrid methods (denoted as \"X+Y\"), training consists of 100 epochs using method X followed by 300 epochs using method Y. We can observe that constructing representation with perceptual inductive bias first will boost the performance of contrastive representation learning on both segmentation and classification tasks. The best two performances are in bold." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 168, + 558, + 266 + ], + "blocks": [ + { + "bbox": [ + 315, + 168, + 558, + 266 + ], + "lines": [ + { + "bbox": [ + 315, + 168, + 558, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 168, + 558, + 266 + ], + "type": "table", + "html": "
EpochsImageNet-100 Top-1 Acc(%)↑ADE20K mIoU(%)↑Depth RME(%)↓
PCL40074.030.50.1437
MoCoV240077.030.40.1434
S-PCL(Ours)40071.831.80.1381
S-PCL+PCL(Ours)40076.031.90.1373
S-PCL+MoCoV2(Ours)40078.031.90.1398
MidVCL(Ours)40074.032.10.1354
MidVCL+MoCoV2(Ours)40077.932.00.1409
", + "image_path": "8eb439dc35acdabc7ab6fcef04181932b8d70e6ca002b950f7630e9fa658e8e6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 289, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 481 + ], + "type": "text", + "content": "tion evaluation, we analyze the Adjusted Mutual Information (AMI) between K-Means cluster assignments of S-PCL and the categorical ground truth. A higher AMI indicates greater alignment between cluster assignments and semantic categories. As illustrated in Figure.5A, the AMI increases during the initial training stages but progressively decreases as the number of epochs increases. This trend suggests that clustering based solely on shape leads to a divergence from the semantic categories. In other words, it suggests that relying exclusively on shape is insufficient for effective object classification, which greatly requires semantic information. However, the shape shows enormous advantages at the early stage. These findings motivate the development of a hybrid framework that employs S-PCL and MidVCL as a warm-up strategy in the early training stages, which is detailed in §4.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 491, + 553, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 553, + 517 + ], + "type": "text", + "content": "4.3. Perceptual inductive bias is beneficial for both image-level task and pixel-level task" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "Accelerated representation learning for classification In the results of linear classification (§4.1) and cluster analysis (§4.2), we observe tremendous benefits of incorporating shape prototypes. We attribute this phenomenon to an analogy with the accelerated learning of new nouns grouped by shapes during human infancy. Consequently, we investigate a hybrid framework that combines perceptual inductive bias and semantic representation training. Specifically, we propose training with S-PCL or MidVCL for the first 100 epochs and then switching to traditional contrastive learning frameworks (denoted as S-PCL+X and MidVCL+X respectively), allowing the model to discover additional structural regularities and develop a better representation space. Figure.5B shows the AMI results of PCL and S-PCL+PCL. It can be observed that the S-PCL warm-up moves the AMI curve upward. Figure.5C and D show the linear classifi" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9626" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "cation accuracies for PCL and S-PCL+PCL on ImageNet-100 and STL10 datasets versus the number of pretraining epochs. For a fair comparison, all frameworks are trained using identical parameter settings and learning rate schedules. The results demonstrate that the S-PCL+PCL method achieves accuracy levels comparable to PCL while requiring less than half the number of training epochs on both ImageNet-100 and STL-10 datasets. The curve of hybrid frameworks of MidVCL is included in supplementary material and it demonstrates the same trend. Overall, our experimental results demonstrate that the sequential learning of intermediate shape representations followed by high-level semantic representations is beneficial for semantic tasks, particularly in accelerating the learning process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": "Sequential training resolves the trade-off between image-level and pixel-level tasks Beyond the classification task, hybrid frameworks demonstrate their superiority also in segmentation task. As reported in Table.4, hybrid frameworks achieve strong performance in both classification and segmentation tasks, unlike methods relying only on either semantic or perceptual representations. Although hybrid frameworks are not as good as S-PCL and MidVCL in depth estimation, they still significantly outperform baseline methods. These results indicate that the additional learning with perceptual inductive bias before semantic contrastive learning provides a solution to the tradeoff mentioned in related works " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\S 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 428, + 295, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 295, + 452 + ], + "type": "text", + "content": "4.4. Concentration on Object Contour and Shape Bias" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 604 + ], + "type": "text", + "content": "Shape silhouette forces model to focus on object contours Next, we study how shape representations vary the model's prediction decision-making process. To this end, we employed the SmoothGrad method to generate sensitivity maps, where highlighted regions contribute significantly to the model's inferences. We show some example images in the validation set of ImageNet-100 in Figure.6. Notably, comparing the sensitivity maps of S-PCL and others, we find that the semantic representations lead to a stronger concentration on local features of the object, which is a shortcut for classification. While the shape representations guide the model to focus on the contour of the object." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "One may argue that although MoCoV2 and PCL focus more on texture, the silhouettes of objects can still be observed to some degree. To address this, we also analyze sensitivity maps of style-transferred images, that is, preserving object shape while altering texture, in addition to the original images, see Figure.7. After enormous style changes, MoCov2 and PCL are detracted by those irrelevant local features while S-PCL still has a strong concentration on the contours of the objects." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 216 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 216 + ], + "type": "image", + "image_path": "f0f4466150da650095c0deb057d5bb8cbe66c8726a3303603e8de7966812c7cd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 226, + 531, + 238 + ], + "lines": [ + { + "bbox": [ + 336, + 226, + 531, + 238 + ], + "spans": [ + { + "bbox": [ + 336, + 226, + 531, + 238 + ], + "type": "text", + "content": "Figure 6. Smoothgrad sensitivity on different images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 251, + 553, + 396 + ], + "blocks": [ + { + "bbox": [ + 316, + 251, + 553, + 396 + ], + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 396 + ], + "type": "image", + "image_path": "6537bc662e66ecc7bfaacbb4e3451f9e8ef9404c472b89e945f8116b95c9c41a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 406, + 555, + 428 + ], + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 428 + ], + "type": "text", + "content": "Figure 7. Smoothgrad sensitivity map on the image in different styles." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 451, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 643 + ], + "type": "text", + "content": "Shape bias score So far, we have found that the shape prototype can increase the shape bias of the model, alleviating a well-known machine's misalignment with humans. To quantitatively evaluate the shape bias of the model, we test our model on the model-vs-human benchmark. The model-vs-human benchmark is a protocol that is able to evaluate the 2D shape bias of a model. With a set of images with conflict cues, e.g. a cat shape with elephant skin texture, we can know which is the predominant visual cue for object recognition between shape and texture. Figure.8 demonstrates shape score across different frameworks. Importantly, the proposed SPCL is the only model that uses shape as the predominant cue, i.e. shape bias score greater than 0.5. RelfCL shows improvements compared to the rest of the approaches but it is still biased towards texture (fraction of shape decisions less than 0.5)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Out-of-distribution robustness Many studies reveal that the model with high shape bias will also have higher robustness to out-of-distribution (OOD) data. To this end, we test models on the 17 OOD datasets included in the model-vs-human benchmarks. Figure.9 presents average classifica" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9627" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 74, + 273, + 305 + ], + "blocks": [ + { + "bbox": [ + 62, + 74, + 273, + 305 + ], + "lines": [ + { + "bbox": [ + 62, + 74, + 273, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 74, + 273, + 305 + ], + "type": "image", + "image_path": "88fbd229946d84f976f1d6f67e6ad37eab9daed2c410ffc75edde96e40c279f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 316, + 296, + 395 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 395 + ], + "type": "text", + "content": "Figure 8. The shape bias score. The value on the top represents the fraction of shape decisions and the value at the bottom represents the fraction of texture decisions. Two values sum to 1 and 0.5 means neutrality, indicating no bias towards either shape or texture. Markers are the fraction of certain shape categories and vertical lines are the mean fraction across shape categories. The star markers represent our proposed approaches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "content": "tion accuracies across all 17 OOD datasets in the model-vs-human benchmark. The full results are included in the supplementary material. S-PCL demonstrates statistically significant improvements in OOD accuracy compared to baseline models. This suggests that leveraging shape information not only aids in better generalization on seen data but also enhances the model's performance on unfamiliar inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 525, + 125, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 125, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 125, + 538 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "In this work, we aim to integrate inductive biases from early and intermediate perceptual constructs into contrastive learning rather than relying solely on high-level semantic information. To achieve this, we maximize the mutual information between full-detail images and one or more perceptual constructs. Our findings reveal that inductive bias from global shape prototypes not only accelerates early-stage classification but also improves segmentation and depth estimation while enhancing the model's overall shape bias. Reflectance intrinsic images facilitate representation learning for classification and segmentation, although they do not benefit depth estimation, whereas shading intrinsic images enhance depth estimation without aiding classification or segmentation. Combining all three perceptual con" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 318, + 73, + 550, + 245 + ], + "blocks": [ + { + "bbox": [ + 318, + 73, + 550, + 245 + ], + "lines": [ + { + "bbox": [ + 318, + 73, + 550, + 245 + ], + "spans": [ + { + "bbox": [ + 318, + 73, + 550, + 245 + ], + "type": "image", + "image_path": "6e2d2c1a6dd1ac62a8f4ff54c40dcc615f64ebe9da05f5cee1625504e4cd6601.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 258, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 313 + ], + "type": "text", + "content": "Figure 9. Mean accuracy across 17 OOD dataset in model- vs-human benchmark. Bars are positioned in the descending order. The error bar is the standard error of the mean (SEM). A paired T-test is conducted. \\*\\* denotes " + }, + { + "bbox": [ + 313, + 258, + 555, + 313 + ], + "type": "inline_equation", + "content": "p < 0.01" + }, + { + "bbox": [ + 313, + 258, + 555, + 313 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 258, + 555, + 313 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 313, + 258, + 555, + 313 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 313, + 258, + 555, + 313 + ], + "type": "inline_equation", + "content": "p < 0.05" + }, + { + "bbox": [ + 313, + 258, + 555, + 313 + ], + "type": "text", + "content": ", n.s. denotes \"not significant\" " + }, + { + "bbox": [ + 313, + 258, + 555, + 313 + ], + "type": "inline_equation", + "content": "p > 0.05" + }, + { + "bbox": [ + 313, + 258, + 555, + 313 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 340, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 472 + ], + "type": "text", + "content": "structures enables us to leverage the unique advantages of each. Moreover, employing sequential training—where the model first harnesses perceptual inductive biases before applying semantic contrastive learning—effectively mitigates the trade-off between image-level and pixel-level tasks. Overall, our results indicate that incorporating perceptual constructs generated by early and intermediate visual processing into self-supervised learning can help networks develop distinct intermediate representations that improve downstream visual tasks, offering a promising avenue for future research in contrastive learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 491, + 384, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 384, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 384, + 504 + ], + "type": "text", + "content": "6. Limitation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "text", + "content": "We developed distinct approaches for specific perceptual constructs, focusing on shape, figure-ground, and intrinsic images. These constructs were chosen due to their close association with fundamental visual tasks like classification, segmentation, and depth estimation. However, additional mid-level perceptual constructs merit further exploration. Moreover, our experiments were conducted on a small model and limited datasets because of computing power constraints, underscoring the need to address scalability in future research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 424, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 424, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 424, + 668 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "This work was supported by NSF CISE RI 1816568 and NIH R01 EY030226-01A1. This work was also supported by the CMU HURAY program." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9628" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Edward H Adelson and Alex P Pentland. The perception of shading and reflectance. Perception as Bayesian inference, 409:423, 1996." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "type": "text", + "content": "[2] Harry Barrow, J Tenenbaum, A Hanson, and E Riseman. Recovering intrinsic scene characteristics. Comput. vis. syst, 2 (3-26):2, 1978." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 295, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 295, + 193 + ], + "type": "text", + "content": "[3] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 239 + ], + "type": "text", + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 240, + 296, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 296, + 294 + ], + "type": "text", + "content": "[5] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223, Fort Lauderdale, FL, USA, 2011. PMLR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 296, + 339 + ], + "type": "text", + "content": "[6] Hans P Op de Beeck, Katrien Torfs, and Johan Wagemans. Perceived shape similarity among unfamiliar objects and the organization of the human object vision pathway. Journal of Neuroscience, 28(40):10111-10123, 2008." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 341, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 295, + 385 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 387, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 295, + 419 + ], + "type": "text", + "content": "[8] James H Elder and Ljiljana Velisavljevic. Cue dynamics underlying rapid detection of animals in natural scenes. J Vis, 9(7):7, 2009." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 421, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 295, + 464 + ], + "type": "text", + "content": "[9] Paul Gavrikov, Jovita Lukasik, Steffen Jung, Robert Geirhos, Bianca Lamm, Muhammad Jehanzeb Mirza, Margret Keuper, and Janis Keuper. Are vision language models texture or shape biased and can we steer them?, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 466, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 466, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 57, + 466, + 295, + 510 + ], + "type": "text", + "content": "[10] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. On the surprising similarities between supervised and self-supervised models, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 512, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 512, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 295, + 577 + ], + "type": "text", + "content": "[11] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. In Advances in Neural Information Processing Systems, pages 23885-23899. Curran Associates, Inc., 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 643 + ], + "type": "text", + "content": "[12] Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision. Advances in Neural Information Processing Systems, 34:23885-23899, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "text", + "content": "[13] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[14] Lisa Gershkoff-Stowe and Linda B Smith. Shape and the first hundred nouns. Child Dev, 75(4):1098-1114, 2004." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[15] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In 2009 IEEE 12th International Conference on Computer Vision, pages 2335-2342. IEEE, 2009." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 173 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 175, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 228 + ], + "type": "text", + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 230, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 251 + ], + "type": "text", + "content": "[18] Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 253, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 284 + ], + "type": "text", + "content": "[19] V A Lamme. The neurophysiology of figure-ground segregation in primary visual cortex. J Neurosci, 15(2):1605-1615, 1995." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 308 + ], + "type": "text", + "content": "[20] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 341 + ], + "type": "text", + "content": "[21] Min Seok Lee, Wooseok Shin, and Sung Won Han. Tracer: Extreme attention guided salient object tracing network, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "type": "text", + "content": "[22] Tai Sing Lee, David Mumford, Richard Romero, and Victor A.F. Lamme. The role of the primary visual cortex in higher level vision. Vision Research, 38(15):2429-2454, 1998." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "text", + "content": "[23] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "text", + "content": "[24] Tianqin Li, Ziqi Wen, Yangfan Li, and Tai Sing Lee. Emergence of shape bias in convolutional neural networks through activation sparsity, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 456, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 477 + ], + "type": "text", + "content": "[25] Ralph Linsker. Self-organization in a perceptual network. Computer, 21(3):105-117, 1988." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "type": "text", + "content": "[26] Muzammal Naseer, Kanchana Ranasinghe, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 513, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 554, + 555 + ], + "type": "text", + "content": "[27] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 557, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 579 + ], + "type": "text", + "content": "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "text", + "content": "[29] Johan Wagemans, Joeri De Winter, Hans Op de Beeck, Annemie Ploeger, Tom Beckers, and Peter Vanroose. Identification of everyday objects on the basis of silhouette and outline versions. Perception, 37(2):207-244, 2008." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 625, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 678 + ], + "type": "text", + "content": "[30] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[31] Ziqi Wen, Tianqin Li, and Tai Sing Lee. Does resistance to style-transfer equal shape bias? evaluating shape bias by distorted shape, 2023." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9629" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 217 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[32] Zhirong Wu, Yuanjun Xiong, Stella Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance-level discrimination, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 295, + 149 + ], + "type": "text", + "content": "[33] Yike Yuan, Xinghe Fu, Yunlong Yu, and Xi Li. Densedino: boosting dense self-supervised learning with token-based point-level consistency. arXiv preprint arXiv:2306.04654, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 294, + 184 + ], + "type": "text", + "content": "[34] Hong Zhou, Howard S Friedman, and Rüdiger Von Der Heydt. Coding of border ownership in monkey visual cortex. Journal of Neuroscience, 20(17):6594-6611, 2000." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 186, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 186, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 186, + 294, + 217 + ], + "type": "text", + "content": "[35] Karl Zipser, Victor A. F. Lamme, and Peter H. Schiller. Contextual modulation in primary visual cortex. Journal of Neuroscience, 16(22):7376-7389, 1996." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9630" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Perceptual Video Compression with Neural Wrapping/4f847378-cfb2-4455-91dd-a70a0227cf83_content_list.json b/2025/Perceptual Video Compression with Neural Wrapping/4f847378-cfb2-4455-91dd-a70a0227cf83_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..000cd66317b8b7bd5835d9d4de214d83d90784e5 --- /dev/null +++ b/2025/Perceptual Video Compression with Neural Wrapping/4f847378-cfb2-4455-91dd-a70a0227cf83_content_list.json @@ -0,0 +1,1875 @@ +[ + { + "type": "text", + "text": "Perceptual Video Compression with Neural Wrapping", + "text_level": 1, + "bbox": [ + 210, + 130, + 758, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Muhammad Umar Karim Khan, Aaron Chadha, Mohammad Ashraful Anam, Yiannis Andreopoulos, Sony Interactive Entertainment", + "text_level": 1, + "bbox": [ + 86, + 180, + 883, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{Umar . Khan, Aaron . Chadha, Russell . Anam, Yiannis . Andreopoulos}@sony.com", + "bbox": [ + 153, + 219, + 810, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 310, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Standard video codecs are rate-distortion optimization machines, where distortion is typically quantified using PSNR versus the source. However, it is now widely accepted that increasing PSNR does not necessarily translate to better visual quality. In this paper, a better balance between perception and fidelity is targeted, in order to provide for significant rate savings over state-of-the-art standards-based video codecs. Specifically, pre- and post-processing neural networks are proposed that enhance the coding efficiency of standard video codecs when benchmarked with an array of well-established perceptual quality scores. These \"neural wrapper\" elements are end-to-end trained with a neural codec module serving as a differentiable proxy for standard video codecs. The codec proxy is jointly optimized with the pre- and post components via a novel two-phase pretraining strategy and end-to-end iterative refinement with stop-gradient. This allows the neural pre- and postprocessor to learn to embed, remove and recover information in a codec-aware manner, thus improving its rate-quality performance. A single neural-wrapper model is thereby established and used for the entire rate-quality curve without needing any downscaling or upscaling. The trained model is tested with the AV1 and VVC standard codecs via an array of well-established objective quality scores (SSIM, MS-SSIM, VMAF, AVQT), as well as mean opinion scores (MOS) derived from ITU-T P.910 subjective testing. Experimental results show that the proposed approach improves all quality scores, with $-18.5\\%$ average Bjontegaard Delta-rate (BD-rate) saving over all objective scores and MOS improvement over both standard codecs. This illustrates the significant potential of neural wrapper components over standards-based video coding.", + "bbox": [ + 75, + 287, + 472, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 784, + 207, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is now widely accepted that that signal-to-noise (SNR) ratio is a poor indicator of visual quality in video coding [6, 19, 26, 42]. Instead, quality scores that include elements of human perception [42], perceptual modelling of encoding artifacts [51], as well as viewing setup awareness [4, 65] have emerged as strong contenders for the best means to", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "objectively characterize visual quality. This has also led to the research community now moving away from SNR-optimization [26] in favor of structural similarity (SSIM [65]), video multimethod assessment fusion (VMAF) and Apple's advanced video quality tool (AVQT) optimization [4, 42]. However, all current perceptual optimization approaches in standards-based video encoders have one or more of the following detriments:", + "bbox": [ + 496, + 268, + 890, + 391 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- they require multiple encoding passes or in-loop implementation within a specific encoder;", + "- they only optimize for a single quality scoring method like VMAF or SSIM and are shown to be detrimental in other quality scores;", + "- they comprise hand-crafted (shallow) models of low-level human perception and fail to encapsulate several characteristics of more advanced quality scoring methods like VMAF or AVQT in a data-driven and learnable manner;", + "their Bjontegaard Delta-rate (BD-rate) improvement on well-established quality scores like SSIM, VMAF and AVQT is in the order of a few percentile points." + ], + "bbox": [ + 500, + 395, + 890, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A neural codec \"wrapper\" is proposed in this paper as the means to augment standard video codecs and optimize them for perceptual quality. Including neural pre- and postprocessing in conventional encoding pipelines is shown to clearly outperform the equivalent codecs without these components. This allows for conventional streaming ecosystems to continue operating \"as-is\" and does not break standards. That is, when neural components cannot be supported before encoding or after decoding, the system can just disable them with appropriate adaptation, without jeopardizing video streaming. The key contributions of the paper are listed as follows.", + "bbox": [ + 509, + 577, + 890, + 758 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel pretraining regime for modelling a conventional codec with a neural codec model is proposed, which uses the implicit encoder-decoder structure of the neural codec to align both video quality and rate.", + "- An end-to-end training regime for jointly training pre and postprocessors is proposed. Starting with a pretrained neural codec model, it is proposed to iteratively update the codec model. This approach allows for more accurate gradients to be backpropagated to the prepro" + ], + "bbox": [ + 511, + 763, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17743", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2259fbf37c5d80f7edf44b7b7d2f9ae61208435af9ec137f09cd4559ece28bb.jpg", + "image_caption": [ + "Figure 1. Left: Functional block diagram of the multi-stage training data flow. Right: Components of codec model $M$ that are used as a proxy for the target codec $C$ . $AE/AD$ denotes arithmetic encoding followed by arithmetic decoding and $Q$ denotes quantization. See Sec. 3 for details." + ], + "image_footnote": [], + "bbox": [ + 138, + 92, + 831, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cessor.", + "bbox": [ + 106, + 330, + 156, + 342 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- The trained pipeline is evaluated on gaming sequences and significant BD-rate gains versus AV1 and VVC encoding are demonstrated with just a single pre- and postprocessor model. In addition, the proposed neural wrapper is shown to outperform a state-of-the-art neural encoder [35], as well as the state-of-the-art in deep perceptual preprocessing proposed recently [11].", + "bbox": [ + 89, + 343, + 472, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A functional block diagram for the proposed system is shown in Fig. 1. Example visuals produced by the proposed method are shown in Fig. 4, demonstrating that the proposed neural wrapper framework allows for better detail preservation than the underlying standard codec.", + "bbox": [ + 89, + 454, + 470, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 550, + 233, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Neural Video Codes", + "text_level": 1, + "bbox": [ + 89, + 575, + 289, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural video codecs optimize neural networks for rate and distortion by exploiting redundancies in video data. Most notable are residual coding methods; Lu et al. [46] introduce deep video codec, which adopts neural networks within the residual coding pipeline of a traditional codec. Optical flow is employed as a proxy for motion vectors in order to warp the reference frame and generate a residual frame for encoding. Numerous other methods have been proposed based on residual coding [44, 49, 53]. Scale-space flow estimation is proposed as an improvement over traditional optical flow by Agustsson et al. [2], namely for better handling of complex motion in the scene. Beyond residual coding methods, conditional coding methods learn feature contexts implicitly from the encoded context [23, 33, 34, 45, 47, 56]. Li et al. [35] propose using diverse contexts in spatial and temporal directions to perform video compression and outperform standard video codecs when the latter are used in low-delay mode. They improve the performance and bitrate range in [36]. Recent work has additionally focused on", + "bbox": [ + 89, + 598, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "reducing the complexity of neural CODECs. For example, Van Rozendaal et al. [64] deploy an architecture similar to Agustsson et al. [2] on a smart phone and achieve real time performance at 1080p. Hu et al. [25] adopt slimmable networks for variable complexity.", + "bbox": [ + 511, + 328, + 893, + 405 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Preprocessors", + "text_level": 1, + "bbox": [ + 513, + 417, + 658, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While postprocessing is a more mature line of research, there has been a prevalence of work in preprocessing, both on image and video modalities. In the image domain, Strumpler et al. [58] propose to modify the JPEG encoder with an attention network and a learnable quantization table, in order to modulate the source frame and provide for additional rate savings over the JPEG encoder alone. Conversely, Talebi et al. [59] avoid adapting the JPEG encoder by prepended the standard encoder with a preprocessor for deployment with a standard JPEG encoder and decoder. Klopp et al. [31] train a preprocessor with a surrogate neural codec model, which is trained to model a target image codec. In the video domain, Boutssoulatze et al. [9] propose a precoding network for downscaling with a online mode selection algorithm, in order to alter the precoding based on scene characteristics. Conversely, Chadha et al. [11] propose an entirely offline training mechanism by first modelling the core components of a video codec in a differentiable manner. A preprocessor is optimized on the codec model, in order to provide further rate-quality savings over the target codec.", + "bbox": [ + 511, + 440, + 893, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Postprocessors", + "text_level": 1, + "bbox": [ + 513, + 771, + 663, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Numerous deep learning methods have been proposed for compressed video denoising, video super-resolution or both tasks. Single frame methods for denoising compressed videos have been proposed in [71, 72]. Multiframe methods have provided superior gains in metrics compared to single frame methods. MFQE [73] detects frames with higher quality and uses it as a reference to", + "bbox": [ + 511, + 794, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17744", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "improve neighboring frames. TSAN [69] uses a temporal deformable alignment and multi-scale fusion for the task of video restoration. STDF [16] uses spatio-temporal deformable convolution to aggregate temporal context rather than using optical flow for frame warping. A common approach for video super-resolution is to align neighboring frames and transfer information across frames [60, 61, 66, 70]. On the contrary, other methods choose non-local modules or 3D convolutions to gather temporal context [29, 30, 38, 74]. Some methods recurrently propagate history in the forward direction [17, 28, 55]. Recent methods such as BasicVSR [12], GOVSR [75] and BasicVSR++ [13] propagate history in forward and backward directions, such that both past and future frames are used to super-resolve the current frame. Architectures developed for video super-resolution can easily be adapted for video denoising. Few authors have focused on the task of jointly denoising and upscaling videos. FTVSR [50] uses a frequency transformer which performs self-attention in space, time and frequency to recover details. RealBasicVSR [14] considers sensor noise, motion blur and compression artifacts with super-resolution. CIAF [76] uses motion-vectors from the video to approximate optical flow. CAVSR [67] adapts the super-resolution to the compression level of the input video. COMISR [39] uses an architecture based on BasicVSR [12] to perform joint denoising and upscaling.", + "bbox": [ + 93, + 90, + 467, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Pre- and Postprocessors with Traditional Codes", + "text_level": 1, + "bbox": [ + 93, + 513, + 467, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent methods have also explored joint training of pre and postprocessors as neural wrappers. In the image domain, recent work [20, 21] emulates the core blocks of the JPEG encoder and decoder, and uses it to train pre and postprocessors for rate-distortion savings. Similar sandwich approach for videos has been proposed in [24]. Son et al. [57] propose distinct neural networks to model the distortion and rate behavior of a standard codec, which is thus sandwiched by pre- and postprocessor networks for rate-distortion optimization. However, these methods have no target codec awareness beyond the implicit intra/inter coding structure. Either the codec proxy is not learned (e.g. derived from differentiable JPEG/DCT as in [24]), only partially learned via rate-distortion optimization (e.g. the loop filter as in [27]) or only operates in intra mode (as in [57]). In this paper, the proposed framework is directly conditioned on the target conventional codec. Tian et al. [62] target additional rate savings with learnable rescaling. However, this method has no rate control or alignment outside of the scaling factor. Conversely, the framework proposed in this paper is jointly conditioned on both the target codec rate and distortion and is thus able to offer rate savings even when encoding at full res", + "bbox": [ + 93, + 553, + 467, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "olution. Additionally, whereas recent work restricts evaluation to a limited range of the rate-distortion space, at high bitrates [61] or substantially lower bitrates and resolutions [22] and on older encoding standards (AVC or HEVC), the proposed method is evaluated against state-of-the-art AV1 and VVC encoding, i.e.: $(i)$ the entire rate-quality region of practical significance is considered (VMAF values between 40-96 and bitrates from 100kbps to 15mbps); $(ii)$ slow encoding presets are used (preset=4 for SVT-AV1 and preset=slow for VVenC), without limiting the encoders to low-delay mode.", + "bbox": [ + 516, + 90, + 890, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 516, + 270, + 645, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The top-level functional block diagram of the proposed training flow is shown in Fig. 1 and includes a preprocessor $P$ , codec model $M$ , target codec $C$ and postprocessor $O$ . Notably, the target codec is a traditional video codec, such as SVT-AV1 [32]. First, a source frame $I_{\\mathrm{in}}$ is processed by $P$ to generate a corresponding preprocessed frame $I_P$ . The preprocessed frame is simultaneously encoded by $M$ and $C$ , which generate encoded frames $I_M$ and $I_C$ respectively. Finally, encoded frames are postprocessed with $O$ to generate the output frame $I_{\\mathrm{out}}$ . End-to-end distortion $D$ is computed between $I_{\\mathrm{in}}$ and $I_{\\mathrm{out}}$ and the overall training loss is the weighted sum of $D$ and the bitrate $R_M$ . Notably, at inference, the codec model $M$ is removed from the pipeline and only the target codec $C$ is deployed for encoding.", + "bbox": [ + 516, + 295, + 890, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Codec Model Pretraining", + "text_level": 1, + "bbox": [ + 516, + 530, + 743, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the target codec $C$ in Fig. 1 is not differentiable, a mechanism for adapting the proposed pipeline to the target codec behavior is required. One option is black-box or derivative-free optimization methods [3, 7, 18]. However these methods typically rely on an approximation of the unknown function to be optimized, and are computationally infeasible for training at scale. Another option is to replicate the core prediction and transform blocks of the target codec in a differentiable manner [15, 27, 54]. Such approaches can model older codecs such as JPEG, or faster presets of AVC or HEVC successfully to an extent but fail to accurately model the operation of more sophisticated codecs such as AV1 or VVC.", + "bbox": [ + 516, + 553, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent work [31, 57] has instead leveraged on neural image and video CODEs for predicting the target codec behavior. In this work, a similar approach is followed, and a neural video codec $M$ is adapted to model the ratedistortion behavior of a target codec $C$ . Specifically, the scale-space flow architecture [2] is used to generate lossy frames, and the implicit encoder-decoder structure and associated loss functions are leveraged to align to the target codec behavior in a spatially localized manner. As per Agustsson et al. [2], for each of intra, flow and residual", + "bbox": [ + 516, + 750, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17745", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "encoding, the hyperprior based architecture of Balle et al. [5] is used, without any autoregressive components.", + "bbox": [ + 89, + 90, + 467, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, the hyperprior architecture comprises analysis and synthesis transforms, $g_{a}$ and $g_{s}$ , for encoding to and decoding from the quantized latent space $\\tilde{y}$ respectively. A distortion loss $D(:,g_a,g_s)$ can thus be computed between source frame $I_{in}$ (input to $g_{a}$ ) and reconstructed frame $I_M$ (output by $g_{s}$ ). A hyper-analysis and hypersynthesis transform $(h_a$ and $h_s$ respectively) constitute the hyperprior, which is used to model scale dependencies $\\tilde{\\sigma}$ within the latent space. The latent space $\\tilde{y}$ is thus modelled as a Gaussian scale mixture on which the rate $R(:,h_a,h_s)$ can be optimized (as Shannon entropy). Traditionally, the architecture would then be trained with an aggregate loss function $L = D + \\lambda R$ , where $\\lambda$ represents a Lagrangian parameter for balancing between the rate and distortion components.", + "bbox": [ + 89, + 121, + 467, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When performing neural codec alignment to a target codec, there are two key issues: (i) the model encoded frame must align to the target encoded frame (and not the source); (ii) selecting the best value for $\\lambda$ when trying to replicate target codec behavior is a non-trivial problem. Crucially, it is noted that the rate model is exclusively parameterized by $h_a$ and $h_s$ , and distortion exclusively by $g_a$ and $g_s$ . Subsequently, a novel two phase training methodology is proposed for aligning both rate and distortion to the target codec:", + "bbox": [ + 89, + 347, + 467, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Phase 1: The first issue of target codec alignment is addressed via the distortion loss. To this end, the reference for the distortion loss is changed from the source frame $I_{in}$ to the target codec frame $I_C$ and the mean squared error is optimized:", + "bbox": [ + 89, + 513, + 467, + 588 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {M} ^ {\\mathrm {h y p}} \\left(I _ {M}, I _ {C}; \\theta_ {g}\\right) = \\mathbb {E} _ {I _ {C}} \\left[ \\left\\| I _ {M} - I _ {C} \\right\\| ^ {2} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 594, + 467, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{g}$ represent the analysis and synthesis transform parameters.", + "bbox": [ + 89, + 621, + 467, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Phase 2: Following alignment of $g_{a}$ and $g_{s}$ in Phase 1, the latent space $\\tilde{y}$ is now conditioned on the target codec $C$ . To circumvent selection of an optimal $\\lambda$ for rate, the implicit encoder-decoder structural prior of the aligned neural codec is leveraged in order to learn an unsupervised rate alignment to the target codec. Specifically, the maximum likelihood on the hyperprior model parameters is optimized, in order to learn the latent space distribution under target codec distortion alignment. The rate is thus optimized as:", + "bbox": [ + 89, + 652, + 467, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {M} ^ {\\mathrm {h y p}} \\left(I _ {M}; \\theta_ {h}\\right) = - \\mathbb {E} _ {\\tilde {y}, \\tilde {z}} \\left[ \\log p _ {\\tilde {y} | \\tilde {z}, I _ {C}} \\left(\\tilde {y} | \\tilde {z}, I _ {C}\\right) + \\log p _ {\\tilde {z} | I _ {C}} \\left(\\tilde {z} | I _ {C}\\right) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 806, + 501, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_h$ represents the hyperprior model parameters and $\\tilde{z}$ is the quantized hyperprior latent space (which would also be encoded and transmitted as side information in a neural codec pipeline). For intra coding, the intra", + "bbox": [ + 89, + 840, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hyperprior model is optimized on intra encoded source frames using the two-phase strategy. For inter coding, the residual and flow hyperprior models are optimized jointly by summing the rate and distortion components in each phase.", + "bbox": [ + 511, + 90, + 890, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed two-phase pretraining ensures that $R_{M}^{\\mathrm{hyp}}$ is close to the target codec $R_{C}$ by performing an affine transform on $R_{M}^{\\mathrm{hyp}}$ , i.e., $R_{M} = \\alpha R_{M}^{\\mathrm{hyp}} + \\beta$ , where $\\alpha$ and $\\beta$ are scalar learnable parameters optimized with MSE:", + "bbox": [ + 511, + 167, + 890, + 242 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M} ^ {(\\alpha , \\beta)} \\left(R _ {M}, R _ {C}; \\alpha , \\beta\\right) = \\mathbb {E} _ {R _ {C}} \\left[ \\left\\| R _ {M} - R _ {C} \\right\\| ^ {2} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 244, + 890, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The similarity of $R_{M}^{\\mathrm{hyp}}$ and $R_{C}$ is required to properly balance the rate and distortion losses in end-to-end training. The aggregate rate and distortion for a sequence can thus be measured by computing the sum over all frames.", + "bbox": [ + 511, + 277, + 890, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. End-to-End Training", + "text_level": 1, + "bbox": [ + 513, + 348, + 712, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the pretrained neural codec model, an end-to-end trainable pipeline with both the pre and postprocessor is developed, as illustrated in Fig. 1. In this case, the training loss is very similar to conventional neural CODECs, but is now exclusively used to train the pre and postprocessors, i.e.:", + "bbox": [ + 511, + 371, + 890, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL \\left(I _ {\\text {o u t}}, I _ {\\text {i n}}; \\theta_ {P}, \\theta_ {O}\\right) = \\mathbb {E} _ {I _ {\\text {i n}}} \\left[ D \\left(I _ {\\text {i n}}, I _ {\\text {o u t}}\\right) + \\lambda R _ {M} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 465, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{P}$ and $\\theta_{O}$ are the pre and postprocessor parameters respectively, $R_{M}$ is the codec model rate estimation (proportional to $R_{M}^{\\mathrm{flow}} + R_{M}^{\\mathrm{res}}$ for inter coding the flow and residual from motion compensation, and $R_{M}^{\\mathrm{intra}}$ for intra coding) and $D$ is the distortion loss component. Following subjective assessment tests, a weighted sum of: mean absolute error (MAE), SSIM, MS-SSIM and detail loss metric (DLM) * [37] is proposed for the distortion component $D$ in Eq. (4). Specifically, $D(I_{\\mathrm{in}},I_{\\mathrm{out}}) = \\lambda_{\\mathrm{L1}}||I_{\\mathrm{in}} - I_{\\mathrm{out}}|| + \\lambda_{\\mathrm{SSIM}}(1 - D_{\\mathrm{SSIM}}(I_{\\mathrm{in}},I_{\\mathrm{out}})) + \\lambda_{\\mathrm{MSSSIM}}(1 - D_{\\mathrm{MSSSIM}}(I_{\\mathrm{in}},I_{\\mathrm{out}})) + \\lambda_{\\mathrm{DLM}}D_{\\mathrm{DLM}}(I_{\\mathrm{in}},I_{\\mathrm{out}})$ .", + "bbox": [ + 511, + 494, + 890, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept drift or domain shift for the codec model is an expected consequence of end-to-end training. The codec model is pretrained with natural videos whereas the preprocessor is continuously updated as training proceeds; thus, the data observed by the codec model has a non-stationary distribution. Since neural networks generally do not perform well on out-of-distribution data, an iterative refinement mechanism is proposed, where the codec model parameters $\\theta_{M}$ are independently updated after each iteration of end-to-end training using the two-phase approach described previously. Notably, in this case the input to the target codec is now the output of the", + "bbox": [ + 511, + 661, + 890, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "* DLM is one of the core components of VMAF [41] and is found to better preserve details under visual degradation. Given that it is primarily wavelet based, it can be implemented in a differentiable manner and used for model optimization.", + "bbox": [ + 500, + 851, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17746", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "preprocessor $I_{P}$ and not the source frame, and $I_{C}$ and $R_{C}$ represent the frame and rate output by the target codec.", + "bbox": [ + 89, + 90, + 467, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, the gradients backpropagated to the preprocessor can be better aligned towards the target codec via the stop-gradient operator. First, it is assumed for the remaining discussion in the section that the tensors representing the frames are flattened, i.e., $\\{I_P,I_M,I_C\\} \\in \\mathbb{R}^{N\\times 1}$ , where $N$ is the number of pixels in the tensor. If the target codec $C$ was differentiable then the accuracy of the gradients backpropagated will increase by decreasing:", + "bbox": [ + 89, + 121, + 468, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG = \\left| J _ {M} \\left(I _ {P}\\right) \\nabla_ {I _ {M}} L - J _ {C} \\left(I _ {P}\\right) \\nabla_ {I _ {C}} L \\right|, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 250, + 468, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $J_{M}(I_{P})\\in \\mathbb{R}^{N\\times N}$ is the transposed Jacobian of $M$ w.r.t. $I_P$ and $\\nabla_{I_M}L$ is the gradient of the objective w.r.t. $I_M$ . Assuming that the frames $I_C$ generated by the target codec $C$ are input to the postprocessor $O$ then Eq. (5) becomes", + "bbox": [ + 89, + 270, + 468, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG = \\left| \\left(J _ {M} \\left(I _ {P}\\right) - J _ {C} \\left(I _ {P}\\right)\\right) \\nabla_ {I _ {C}} L \\right|, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 349, + 468, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since $\\nabla_{I_C}^{(j)}L$ is necessary for the networks to train, $J_M^{(i,j)}(I_P) - J_C^{(i,j)}(I_P)$ should be close to zero. Thus,", + "bbox": [ + 89, + 376, + 468, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\forall i, j \\in [ 1, N ], \\left| \\frac {\\partial I _ {M} ^ {(i)}}{\\partial I _ {P} ^ {(j)}} - \\frac {\\partial I _ {C} ^ {(i)}}{\\partial I _ {P} ^ {(j)}} \\right| = \\left| \\frac {\\partial}{\\partial I _ {P} ^ {(j)}} \\left(I _ {M} ^ {(i)} - I _ {C} ^ {(i)}\\right) \\right| \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 425, + 468, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "should be minimized. The above analysis can be extended to the rate component. Eq. (7) shows that accurate gradients can be backpropagated to the preprocessor if the codec model's output is close to the target codec's output and the input to the postprocessor is from the target video codec. In the proposed training regime, the input to the postprocessor is therefore set as", + "bbox": [ + 89, + 479, + 468, + 584 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {O _ {\\text {i n}}} = \\overline {{\\left(I _ {C} - I _ {M}\\right)}} + I _ {M}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 585, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\overline{a}$ indicates the stop-gradient operator [63] applied to the tensor $a$ and $I_{C}$ is the output of the non-differentiable video codec. Eq. (8) allows frames generated by the target codec to be passed to the postprocessor while using the codec model for backpropagation.", + "bbox": [ + 89, + 613, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the proposed training procedure has two main steps; iterative update of the codec model and the end-to-end update of the pre- and postprocessors. For the end-to-end update, a stop-gradient-based approach is proposed. Simple CNNs are used for pre and postprocessing. Details of the pre and postprocessor architecture are in the supplementary. The end-to-end training pipeline is shown in Algorithm 1.", + "bbox": [ + 89, + 689, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 821, + 295, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Training", + "text_level": 1, + "bbox": [ + 89, + 847, + 194, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Vimeo-90k dataset [70] was used to train the codec model and perform end-to-end training. The training", + "bbox": [ + 89, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 End-to-end training pipeline of the pre and postprocessor", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: $I_{\\mathrm{in}}$ , batch of video frames", + "bbox": [ + 500, + 125, + 743, + 140 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: $\\theta_{M}$ , parameters of the codec model", + "bbox": [ + 501, + 141, + 803, + 155 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: $\\theta_{P}$ , parameters of the preprocessor", + "bbox": [ + 501, + 156, + 800, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: $\\theta_{O}$ , parameters of the postprocessor", + "bbox": [ + 501, + 171, + 807, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: $\\alpha_{L}$ , learning rate for training", + "bbox": [ + 501, + 186, + 761, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for all $I_{\\mathrm{in}}$ do", + "bbox": [ + 517, + 202, + 607, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {P} \\leftarrow P (\\boldsymbol {I} _ {\\text {i n}})\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 215, + 633, + 231 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\boldsymbol {I} _ {C}, \\boldsymbol {R} _ {C}\\right) \\leftarrow C \\left(\\boldsymbol {I} _ {P}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 232, + 678, + 246 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\boldsymbol {I} _ {M}, \\boldsymbol {R} _ {M}\\right) \\leftarrow M \\left(\\boldsymbol {I} _ {P}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 247, + 689, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Get $D_M^{hyp}$ and $\\theta_g\\gets \\theta_g - \\alpha_L\\nabla_{\\theta_g}D_M^{hyp}$", + "bbox": [ + 542, + 262, + 799, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Get $R_M^{hyp}$ and $\\theta_h\\gets \\theta_h - \\alpha_L\\nabla_{\\theta_h}R_M^{hyp}$", + "bbox": [ + 542, + 280, + 797, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Get $L_{M}^{(\\alpha ,\\beta)}$ and $(\\alpha ,\\beta)\\gets (\\alpha ,\\beta) - \\alpha_{L}\\nabla_{(\\alpha ,\\beta)}L_{M}^{(\\alpha ,\\beta)}$", + "bbox": [ + 542, + 296, + 877, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generate $I_{O_{\\mathrm{in}}}$ using Eq. (8)", + "bbox": [ + 542, + 315, + 725, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\pmb{I}_{\\mathrm{out}}\\gets O(\\pmb{I}_{O_{\\mathrm{in}}})$", + "bbox": [ + 542, + 329, + 651, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Get $L$ using Eq. (4)", + "bbox": [ + 542, + 345, + 671, + 358 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {P} \\leftarrow \\theta_ {P} - \\alpha_ {L} \\nabla_ {\\theta_ {P}} L\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 359, + 687, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {O} \\leftarrow \\theta_ {O} - \\alpha_ {L} \\nabla_ {\\theta_ {O}} L\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 375, + 687, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "end for", + "text_level": 1, + "bbox": [ + 517, + 391, + 570, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pipeline and the pre- and postprocessor are implemented in the YUV colorspace. A fixed learning rate of $10^{-4}$ is used for each network with the rest of hyper-parameters set as: $\\lambda = 0.1$ ; $\\lambda_{\\mathrm{L1}} = 0.1$ , $\\lambda_{\\mathrm{SSIM}} = 0.25$ ; $\\lambda_{\\mathrm{MSSSIM}} = 0.25$ ; $\\lambda_{\\mathrm{DLM}} = 0.2$ . Batch size, GOP and maximum iterations are set to 4, 4 and 500000, respectively. Both Phase 1 and 2 of pretraining are performed for 500000 iterations. The model is trained on source frames only in an open-loop configuration; i.e., no reconstructed frames are fed into the preprocessor as input. A low-delay recipe of SVT-AV1 (described in supplementary) is used for codec model alignment for pre-training and end-to-end training. The codec model encoding is aligned to the GOP of the target codec, such that intra and inter coding are also aligned. For experiments on VVC, it was found that codec model alignment to a low-delay recipe of VVenC provided slight BD-rate gains over SVT-AV1 alignment, albeit at high computational overhead during training. Therefore, the presented results are carried out with a single neural wrapper model (one encode) developed based on codec proxy model alignment to SVT-AV1. This is practically more feasible than [24], which requires 9 combinations of models (9 encodes) + convex hull optimization and is thus far more computationally demanding.", + "bbox": [ + 511, + 430, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 513, + 801, + 702, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Four 1080p video datasets were used for evaluation: (i) 16 gaming sequences captured in a commercial setup (Gaming); (ii) The 16 XIPH natural sequences used by recent work [11] (XIPH); (iii) The 1080p sequences of the UVG [48] dataset (UVG); (iv) HEVC-B [8] dataset", + "bbox": [ + 511, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17747", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(HEVC-B). The proposed method was applied with SVT-AV1 [32] for AV1 and VVenC [10] for VVC. All experiments were performed with YUV 4:2:0 8-bit sequences. In all cases, the protocol of Li et al. [35] is followed and 96 frames are evaluated. However, unlike Li [35], both standard encoders are used without their low-delay limitation, in order to attain their best performance (see supplementary for details on sequences and encoding recipes). Ablation results are with the SVT-AV1 lowdelay recipe (see supplementary). Results in Tab. 3 and 4 are on the Gaming dataset.", + "bbox": [ + 89, + 90, + 472, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to cover a wide range of well-established objective quality scoring methods, the following scores are used: SSIM, MS-SSIM, VMAF [42, 52] and the AVQT score from Apple [4]. In order to showcase the generality of the proposed approach, the same pre- and postprocessor models are used for: (i) all the points on the RD curves, (ii) all quality scores and (iii) both standard codecs. Perceptual models are used for competing methods when available. Comparison is provided against a competitive video preprocessor (DPP [11]), SOTA neural codecs (DCVC-DC [35] and DCVC-FM [36]), and perceptual modes of standard codecs. The comparisons include results of end-to-end training with the codec model of Hu et al. [24] with the proposed network architectures and losses used to illustrate the effectiveness of the proposed codec model. For DCVC-DC and DCVC-FM, the best available pretrained models (publicly provided by the authors) are used.", + "bbox": [ + 91, + 257, + 472, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Objective Quality Assessment", + "text_level": 1, + "bbox": [ + 89, + 539, + 356, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative results of the proposed method are shown in† Fig. 2, and Tab. 1 and Tab. 2. These results show that the proposed approach provides for BD-rates between $-5.4\\%$ to $-30\\%$ in comparison to the underlying standard encoders. BD-rate savings tend to be higher for quality scores that are more perceptually-oriented. Compared to competing methods, the proposed approach is shown to offer significantly more consistent BD-rate improvements over all quality scoring methods (ALL columns of Tab. 1 and Tab. 2). Importantly, as shown in Fig. 2, unlike DCVC-DC that is only covering a very limited bitrate-quality range at the very top end, the proposed approach outperforms all other methods on the entire \"active\" region of bitrate-quality, i.e., VMAF of 40-93 [32], which encapsulates the entirety of commercially-relevant bitrates for 1080p video streaming. Given that the neural wrapper is optimizing perceptually, BD-rates for PSNR tend to show regression on PSNR (see supplementary) but visual inspection and an array of perceptual metrics indicate that the proposed approach offers significant benefits", + "bbox": [ + 89, + 561, + 472, + 864 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/65c01de4bcda9262aea18f8548bb9701ee71161eac47d04e9f811fff78595cb2.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 514, + 98, + 684, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/825fde789f275700e39dad52a643630271375a0f8adf42751298eb814e2a7bc2.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 714, + 98, + 877, + 183 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43c24bb04a97d86178a09b4081705d4911f79f0bee25609b94c8a00cfcb76007.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 517, + 218, + 684, + 301 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f8d5f43a991db678e7df6cf5bb05450de3f53f824f28b7d5d3882680939bad64.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 718, + 220, + 877, + 300 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f0e2c9981762f202e9c52ed2a4c25a5271e4cd7ffa32b0f54375e40c94ceeec.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 514, + 337, + 681, + 421 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/233c71f97da0b0a3afc2113377f60647550034c1e7e0e20ed986a0a6772857fd.jpg", + "image_caption": [ + "(f)" + ], + "image_footnote": [], + "bbox": [ + 712, + 337, + 875, + 421 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/96cce1064cfc1d05ecaf8517cad37f29915ae71b8eea34d9ad0113df58a1ebda.jpg", + "image_caption": [ + "(g)" + ], + "image_footnote": [], + "bbox": [ + 514, + 455, + 681, + 541 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f30839fc121e9fefc8419dcd74925d00fceb6b91eee5dba5dbe9d677692fc19.jpg", + "image_caption": [ + "(h)", + "Figure 2. Combined rate-quality plots over Gaming ((a), (b)), XIPH ((c), (d)), UVG ((e), (f)) and HEVC-B ((g), (h)) datasets. See supplementary for more results." + ], + "image_footnote": [], + "bbox": [ + 712, + 455, + 875, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in visual quality. Nonetheless, the results of DCVC-DC and DCVC-FM are remarkably good given that they are fully neural approaches.", + "bbox": [ + 511, + 651, + 890, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Subjective Quality Assessment", + "text_level": 1, + "bbox": [ + 511, + 710, + 784, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the visual quality of the proposed approach, a 5-scale absolute categorical rating (ACR) test with hidden reference was conducted under the ITU-T P.910 test conditions and SUREAL post-processing [40, 43] (see supplementary for details). Raters were shown videos generated by the codec and proposed method. Separate tests were conducted for AV1 and VVC with each test comprising 72 sequences. The combined rate-MOS plots are shown in Fig. 3. The BD-rates reported in the figure for the proposed approach are mostly in line with AVQT and VMAF BD-rates reported for Gaming in Tab. 1 and", + "bbox": [ + 511, + 734, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ All combined plots of Fig. 2 and Fig. 3 are produced with the slope-based integration method of Wu et al. [68] and are shown in log-scale.", + "bbox": [ + 76, + 875, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17748", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/db6c88e3ee2e7646b74d56d07076f76b4cd6b9d4f2d41e50e3288b7a554cb0c3.jpg", + "table_caption": [ + "Table 1. BD-rates vs. AV1 (more negative is better). AV1+SSIM denotes codec tuned for SSIM (-tune ssim). ALL indicates average BD-rate over all quality scores." + ], + "table_footnote": [], + "table_body": "
DatasetMethodSSIMMSSSIMAVQTVMAFALL
GamingAV1+SSIM-4.5-5.4-5.12.1-3.2
AV1+DPP [11]-5.2-6.1-17.6-11.3-10.0
AV1+[24]14.96.312.2-1.48.0
DCVC-DC [35]-15.4-36.1-35.621.84-16.3
DCVC-FM [36]14.86.04.2-2.95.5
AV1+Proposed-15.5-13.8-19.6-30.0-19.7
XIPHAV1+SSIM-3.5-4.0-5.43.2-2.4
AV1+DPP [11]-4.3-3.9-9.0-16.7-8.5
AV1+[24]12.96.14.51.76.3
DCVC-DC [35]-18.4-25.512.032.20.3
DCVC-FM [36]17.69.08.23.59.6
AV1+Proposed-10.9-7.1-28.8-30.1-19.2
UVGAV1+SSIM-2.8-3.1-2.82.9-1.4
AV1+DPP [11]-3.8-3.0-9.2-21.7-9.4
AV1+[24]9.14.43.33.25.0
DCVC-DC [35]-20.8-10.24.218.2-2.1
DCVC-FM [36]20.49.69.89.412.3
AV1+Proposed-16.2-9.7-22.2-31.9-20.0
HEVC-BAV1+SSIM-5.4-5.3-4.72.9-3.1
AV1+DPP [11]-8.7-6.9-5.7-20.1-10.3
AV1+[24]2.61.8-3.11.10.6
DCVC-DC [35]-35.3-29.1-11.637.2-9.7
DCVC-FM [36]-6.5-11.7-7.7-5.8-7.9
AV1+Proposed-14.5-13.4-15.1-41.8-21.2
", + "bbox": [ + 83, + 141, + 464, + 433 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/47f51655552e5af1cd24c295ab9acaa94689ca477f77e2db7231438d503f6e1e.jpg", + "table_caption": [ + "Table 2. BD-rates vs. VVC (more negative is better). $\\mathrm{VVC + QPA}$ indicates quantization parameter adaptation is enabled. ALL indicates average BD-rate over all quality scores." + ], + "table_footnote": [], + "table_body": "
DatasetMethodSSIMMSSSIMAVQTVMAFALL
GamingVVC+QPA0.44.613.41.54.9
VVC+DPP [11]-1.3-1.6-9.2-9.0-5.3
VVC+[24]39.532.311.47.822.7
DCVC-DC [35]9.1-11.5-21.338.93.8
DCVC-FM [36]4.918.410.918.413.1
VVC+Proposed-16.6-8.7-19.0-14.3-14.6
XIPHVVC+QPA-0.92.21.12.71.3
VVC+DPP [11]-2.3-1.6-10.5-10.4-6.2
VVC+[24]29.424.412.29.518.9
DCVC-DC [35]-8.1-12.09.241.17.5
DCVC-FM [36]4.313.10.516.88.7
VVC+Proposed-11.4-5.4-16.5-18.0-12.8
UVGVVC+QPA-0.11.67.71.52.7
VVC+DPP [11]-1.4-1.1-7.0-11.4-5.2
VVC+[24]26.823.57.96.516.2
DCVC-DC [35]-8.216.0-1.429.18.9
DCVC-FM [36]-2.714.112.514.79.65
VVC+Proposed-11.8-5.4-13.9-13.8-11.2
HEVC-BVVC+QPA0.91.21.12.11.3
VVC+DPP [11]-19.2-16.7-17.2-16.7-17.4
VVC+[24]0.50.50.1-2.3-0.3
DCVC-DC [35]-17.0-5.56.349.98.4
DCVC-FM [36]18.815.814.412.115.3
VVC+Proposed-31.8-27.4-25.2-33.3-29.4
", + "bbox": [ + 84, + 503, + 464, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2. Sample visual results of the proposed method are shown in Fig. 4. The results show that the proposed approach is temporally coherent, and better capable of preserving structure, retaining textures and denoising codec artifacts, also reflected in the MOS scores of Fig. 3.", + "bbox": [ + 89, + 824, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/788833425de26784a381ec26c10bb7cbc270fdb557b8716396bd3923cdd4c25c.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 514, + 99, + 679, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/327424f0839c08d2ce220f686a7a9981316a3a4606eb3f64d034798b920adb30.jpg", + "image_caption": [ + "(b)", + "Figure 3. Combined rate-quality plots with P.910 MOS." + ], + "image_footnote": [], + "bbox": [ + 702, + 99, + 867, + 195 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4dafdb6f78637341ba9d77004af6180748e8c668f80abc79f0fbf7627f74c1e0.jpg", + "table_caption": [ + "Table 3. BD-rates of different component configurations." + ], + "table_footnote": [], + "table_body": "
PrePostSSIMMSSSIMAVQTVMAF
✓✓-25.6-20.8-21.4-20.5
✓✗17.913.229.8-34.1
✗✓-2.90.44.1-33.3
", + "bbox": [ + 560, + 277, + 834, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation", + "text_level": 1, + "bbox": [ + 513, + 363, + 617, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two-phase pretraining: The rate-quality alignment between target codec $C$ and model $M$ after two-phase pretraining is validated by comparing their performance on a dataset ov dieo sequences. The Pearson correlation coefficient (PCC) is improved from 0.56 and 0.60 to 0.88 and 0.85 for bitrate and PSNR, respectively (see supplementary for more results). Note that the PCC for rate is computed between $R_{C}$ and $R_{M}^{hyp}$ . This indicates that both distortion and bitrate of the codec model and target codec correlate better after the two-phase pretraining, which is a sufficient condition for gradient-based optimization [31]. Tab. 4 shows the effect of pretraining on wrappers.", + "bbox": [ + 511, + 387, + 890, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pre/post processing: Tab. 3 shows the effect of removing either the pre or the postprocessor from the overall pipeline under the same training pipeline. Removing the preprocessor is equivalent to training a video denoiser. While BD-rate gains for VMAF are maintained with only the pre- and postprocessor, this can be attributed to sharpening effects over the source content, which VMAF has susceptibility to [1]. Indeed, with more naive content enhancement such as sharpening, fidelity oriented methods such as SSIM and MS-SSIM will experience BD-rate losses, which can be seen when training both pre and postprocessors alone. Only jointly training both pre- and postprocessors is able to achieve substantial BD-rate gains across all quality scores. Fig. 5 provides a generic example where the preprocessor embeds information into the source frame which does not significantly increase the bitrate and can survive the codec compression. This information is used by the postprocessor for reconstruction.", + "bbox": [ + 511, + 569, + 892, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "End-to-end training: For end-to-end training, an ablation is performed over the proposed iterative updates of the codec model and stop-gradient during training. The results are provided in Tab. 4 under different training con", + "bbox": [ + 511, + 840, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17749", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d889a979e5edd61335cb6f939923f6830fd09395e6f79308859539dfa3a01874.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 89, + 385, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84ae54ff76a05b71383ad094cbc2ec1f3ce537f2c7e4b5a029f958edc50a3fdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 90, + 589, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d92019e321f35b83c55cee1a9eb93b8a6a93a8a00b2e463ecd6012006977ff9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 90, + 795, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ec950dadcabe2aadc1bc52cbd9f2ff865457c28823754c4e623fb111d436114.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of codecs with and without the proposed method. Zoom in for better viewing." + ], + "image_footnote": [], + "bbox": [ + 168, + 273, + 385, + 441 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d20947a0105eec2526383e64108f069158ea431b02dbefcb8a441c80d8282472.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 273, + 589, + 441 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e7ee760f98682c083c332657d678ccc8a76d86523edf90af550732cbd5d36ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 273, + 795, + 441 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e67df1c3a24f8795c9c7c565baab1cee6becd7f67fa8d4e7e59ae7ef962946a0.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 120, + 489, + 300, + 570 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef51856a557b7a713cbbb07c1619de990484ba21683b93548de4e6da2deb29a7.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 303, + 491, + 482, + 570 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/efe18297648ee5ee11480c7b745083ecb554794f7cb5cddb6353dd0d71cc0115.jpg", + "image_caption": [ + "(c)", + "Figure 5. Results of components of the pipeline where (a), (b), (c) and (d) are the source frame, preprocessor output, codec output and postprocessor output, respectively." + ], + "image_footnote": [], + "bbox": [ + 485, + 491, + 663, + 570 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8cf22dddd057e91f0608f2464b208c62c31737ff5e39e74e999214feb11268d.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 666, + 491, + 846, + 570 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/af374c48eb3a8112a5ab6f971ca58d1f77f9e0810c01ce209f8600d6c7d4fc87.jpg", + "table_caption": [ + "Table 4. BD-rates with different training configurations." + ], + "table_footnote": [], + "table_body": "
PretrainIter. UpdateStop Grad.SSIMMSSSIMAVQTVMAF
✓✓✓-25.6-20.8-21.4-20.5
X✓✓-23.1-14.4-17.3-18.7
✓✓X16.40.0110.5-15.3
✓X✓-11.2-6.5-7.7-13.1
", + "bbox": [ + 86, + 672, + 459, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figurations. Notably, when training the pre- and postprocessor with only iterative updates, there are BD-rate losses in SSIM and MS-SSIM. This is a by-product of the domain shift between the source and preprocessor output that increases as end-to-end training progresses. When performing end-to-end training with both iterative updates and stop gradient, there is a sufficient alignment constraint on the codec model, which leads to demonstrable BD-rate gains across quality scoring methods.", + "bbox": [ + 89, + 768, + 472, + 906 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 647, + 633, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work proposes a new approach to jointly train neural pre and postprocessors to improve the rate-quality performance of conventional video CODECs. It is proposed to use a neural codec to model the standard codec's sophisticated rate-quality characteristics, combined with: (i) a novel approach to pretrain the codec model; (ii) a new method to train the neural wrapper model jointly with the codec model in the middle. Extensive evaluations with state-of-the-art AV1 and VVC CODECs show that the proposed neural wrapper approach provides for consistent BD-rate improvement over standard CODECs. Unlike neural methods for preprocessing and encoding, the proposed method allows for consistent improvement over all quality scores (SSIM, MSSSIM, AVQT, VMAF, P.910 MOS) over the entire rate-quality region of modern standards.", + "bbox": [ + 511, + 672, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17750", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 93, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] On vmaf's property in the presence of image enhancement operations. https : // docs . google . com / document / d / 1dJczEhX00MZjBSNyKmd3ARiCTdFVMNPBykH4 _ HMPoyY/edit#heading=h.oaikhnw46pw5. Last accessed: 2020-11-16. 7", + "[2] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2020. 2, 3", + "[3] Stéphane Alarie, Charles Audet, Aïmen E Gheribi, Michael Kokkolaras, and Sébastien Le Digabel. Two decades of blackbox optimization applications. EURO Journal on Computational Optimization, 9:100011, 2021. 3", + "[4] Apple. \"Apple avqt presentation\", 2021. https://developer.apple.com/videos/play/wwdc2021/10145.1,6", + "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 4", + "[6] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6228-6237, 2018. 1", + "[7] Mohammad Reza Bonyadi and Zbigniew Michalewicz. Particle swarm optimization for single objective continuous space problems: a review. Evolutionary computation, 25(1):1-54, 2017. 3", + "[8] Frank Bossen et al. Common test conditions and software reference configurations. JCTVC-L1100, 12(7):1, 2013. 5", + "[9] Eirina Bourtsovlatze, Aaron Chadha, Ilya Fadeev, Vasileios Giotasas, and Yiannis Andreopoulos. Deep video precoding. IEEE Transactions on Circuits and Systems for Video Technology, 2019. 2", + "[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 6", + "[11] Aaron Chadha and Yiannis Andreopoulos. Deep perceptual preprocessing for video coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14852-14861, 2021. 2, 5, 6, 7", + "[12] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 3", + "[13] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In" + ], + "bbox": [ + 93, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 3", + "[14] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5962-5971, 2022. 3", + "[15] Jinyoung Choi and Bohyung Han. Task-aware quantization network forJPEG image compression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 309-324. Springer, 2020. 3", + "[16] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. In Proceedings of the AAAI conference on artificial intelligence, pages 10696-10703, 2020. 3", + "[17] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3476-3485. IEEE, 2019. 3", + "[18] Ryosuke Furuta, Naoto Inoue, and Toshihiko Yamasaki. Pixelrl: Fully convolutional network with reinforcement learning for image processing. IEEE Transactions on Multimedia, 22(7):1704-1719, 2019. 3", + "[19] Vasileios Giotsas, Nikos Deligiannis, Pam Fisher, and Yiannis Andreopoulos. Perceptual video quality estimation by regression with myopic experts. In 2015 Seventh International Workshop on Quality of Multimedia Experience (QoMEX), pages 1-6. IEEE, 2015. 1", + "[20] Onur G Guleryuz, Philip A Chou, Hugues Hoppe, Danhang Tang, Ruofei Du, Philip Davidson, and Sean Fanello. Sandwiched image compression: wrapping neural networks around a standard codec. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3757-3761. IEEE, 2021. 3", + "[21] Onur G Guleryuz, Philip A Chou, Hugues Hoppe, Danhang Tang, Ruofei Du, Philip Davidson, and Sean Fanello. Sandwiched image compression: Increasing the resolution and dynamic range of standard codec. In 2022 Picture Coding Symposium (PCS), pages 175-179. IEEE, 2022. 3", + "[22] Mengxi Guo, Shijie Zhao, Hao Jiang, Junlin Li, and Li Zhang. Video compression with arbitrary rescaling network. arXiv preprint arXiv:2306.04202, 2023. 3", + "[23] Yung-Han Ho, Chih-Peng Chang, Peng-Yu Chen, Alessandro Gnutti, and Wen-Hsiao Peng. Canf-vc: Conditional augmented normalizing flows for video compression. In European Conference on Computer Vision, pages 207-223. Springer, 2022. 2", + "[24] Yueyu Hu, Chenhao Zhang, Onur G Guleryuz, Debargha Mukherjee, and Yao Wang. Standard compliant video coding using low complexity, switchable neural wrappers. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1922-1928. IEEE, 2024. 3, 5, 6, 7", + "[25] Zhihao Hu and Dong Xu. Complexity-guided slimmable decoder for efficient deep video compression. In Proceed-" + ], + "bbox": [ + 517, + 93, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17751", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14358-14367, 2023. 2", + "[26] Quan Huynh-Thu and Mohammed Ghanbari. The accuracy of psnr in predicting video quality for different video scenes and frame rates. Telecommunication Systems, 49: 35-48, 2012. 1", + "[27] Berivan Isik, Onur G Guleryuz, Danhang Tang, Jonathan Taylor, and Philip A Chou. Sandwiched video compression: Efficiently extending the reach of standard codecswith neural wrappers. arXiv preprint arXiv:2303.11473, 2023. 3", + "[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 645-660. Springer, 2020. 3", + "[29] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8008-8017, 2020. 3", + "[30] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3224-3232, 2018. 3", + "[31] Jan P Klopp, Keng-Chi Liu, Liang-Gee Chen, and Shao-Yi Chien. How to exploit the transferability of learned image compression to conventional codescs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16165-16174, 2021. 2, 3, 7", + "[32] Faouzi Kossentini, Hassen Guermazi, Nader Mahdi, Chekib Nouira, Amir Naghdinezhad, Hassene Tmar, Omar Khelif, Phoenix Worth, and Foued Ben Amara. The svt-av1 encoder: overview, features and speed-quality tradeoffs. In Applications of Digital Image Processing XLIII, page 1151021. International Society for Optics and Photonics, 2020. 3, 6", + "[33] Théo Ladune, Pierrick Philippe, Wassim Hamidouche, Lu Zhang, and Olivier Déforges. Conditional coding for flexible learned video compression. arXiv preprint arXiv:2104.07930, 2021. 2", + "[34] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022. 2", + "[35] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 2, 6, 7", + "[36] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with feature modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26099-26108, 2024. 2, 6, 7" + ], + "bbox": [ + 91, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Songnan Li, Fan Zhang, Lin Ma, and King Ngi Ngan. Image quality assessment by separately evaluating detail losses and additive impairments. IEEE Transactions on Multimedia, 13(5):935-949, 2011. 4", + "[38] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. Mucan: Multi-correspondence aggregation network for video super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 335-351. Springer, 2020. 3", + "[39] Yinxiao Li, Pengchong Jin, Feng Yang, Ce Liu, Ming-Hsuan Yang, and Peyman Milanfar. Comisr: Compression-informed video super-resolution. in 2021 IEEE. In CVF International Conference on Computer Vision (ICCV), pages 2523-2532, 2021. 3", + "[40] Zhi Li and Christos G Bampis. Recover subjective quality scores from noisy measurements. In 2017 Data compression conference (DCC), pages 52-61. IEEE, 2017. 6", + "[41] Zhi Li, Anne Aaron, Ioannis Katsavounidis, Anush Moorthy, and Megha Manohara. Toward a practical perceptual video quality metric. The Netflix Tech Blog, 6, 2016. 4", + "[42] Zhi Li, Christos Bampis, Julie Novak, Anne Aaron, Kyle Swanson, Anush Moorthy, and J Cock. Vmaf: The journey continues. *Netflix Technology Blog*, 2018. 1, 6", + "[43] Zhi Li, Christos G Bampis, Lukáš Krasula, Lucjan Janowski, and Ioannis Katsavounidis. A simple model for subject behavior in subjective experiments. arXiv preprint arXiv:2004.02067, 2020. 6", + "[44] Jianping Lin, Dong Liu, Houqiang Li, and Feng Wu. MLvc: Multiple frames prediction for learned video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3546-3554, 2020. 2", + "[45] Jerry Liu, Shenlong Wang, Wei-Chiu Ma, Meet Shah, Rui Hu, Pranaab Dhawan, and Raquel Urtasun. Conditional entropy coding for efficient video compression. In European Conference on Computer Vision, pages 453-468. Springer, 2020. 2", + "[46] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. Dvc: An end-to-end deep video compression framework. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 2", + "[47] Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, and Eirikur Agustsson. Vct: A video compression transformer. arXiv preprint arXiv:2206.07307, 2022. 2", + "[48] Alexandre Mercat, Marko Viitanen, and Jarno Vanne. Uvg dataset: 50/120fps 4k sequences for video codec analysis and development. In Proceedings of the 11th ACM Multimedia Systems Conference, pages 297-302, 2020. 5", + "[49] Reza Pourreza, Hoang Le, Amir Said, Guillaume Sautiere, and Auke Wiggers. Boosting neural video codecs by exploiting hierarchical redundancy. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5355-5364, 2023. 2" + ], + "bbox": [ + 516, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17752", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Zhongwei Qiu, Huan Yang, Jianlong Fu, and Dongmei Fu. Learning spatiotemporal frequency-transformer for compressed video super-resolution. In European Conference on Computer Vision, pages 257-273. Springer, 2022. 3", + "[51] Alexander Raake, Silvio Borer, Shahid M Satti, Jörgen Gustafsson, Rakesh Rao Ramachandra Rao, Stefano Medagli, Peter List, Steve Göring, David Lindero, Werner Robitza, et al. Multi-model standard for bitstream-, pixel-based and hybrid video quality assessment of uhd/4k: Itu-t p. 1204. IEEE Access, 8:193020-193049, 2020. 1", + "[52] Reza Rassool. Vmaf reproducibility: Validating a perceptual practical video quality metric. In 2017 IEEE international symposium on broadband multimedia systems and broadcasting (BMSB), pages 1-2. IEEE, 2017. 6", + "[53] Oren Rippel, Alexander G Anderson, Kedar Tatwawadi, Sanjay Nair, Craig Lytle, and Lubomir Bourdev. Elf-vc: Efficient learned flexible-rate video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14479-14488, 2021. 2", + "[54] Amir Said, Manish Kumar Singh, and Reza Pourreza. Differentiable bit-rate estimation for neural-based video codec enhancement. In 2022 Picture Coding Symposium (PCS), pages 379-383. IEEE, 2022. 3", + "[55] Mehdi SM Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6626-6634, 2018. 3", + "[56] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 2", + "[57] Hanbin Son, Taeoh Kim, Hyeongmin Lee, and Sangyoun Lee. Enhanced standard compatible image compression framework based on auxiliary codec networks. IEEE Transactions on Image Processing, 31:664-677, 2021. 3", + "[58] Yannick Strümpler, Ren Yang, and Radu Timofte. Learning to improve image compression without changing the standard decoder. In European Conference on Computer Vision, pages 200–216. Springer, 2020. 2", + "[59] Hossein Talebi and Peyman Milanfar. Learned perceptual image enhancement. In 2018 IEEE International Conference on Computational Photography (ICCP), pages 1-13. IEEE, 2018. 2", + "[60] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In Proceedings of the IEEE international conference on computer vision, pages 4472-4480, 2017. 3", + "[61] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu Tdan. temporally-deformable alignment network for video super-resolution. in 2020 ieee. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3357-3366, 2020. 3", + "[62] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 3" + ], + "bbox": [ + 91, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5", + "[64] Ties van Rozendaal, Tushar Singhal, Hoang Le, Guillaume Sautiere, Amir Said, Krishna Buska, Anjuman Raha, Dimitris Kalatzis, Hitarth Mehta, Frank Mayer, et al. Mobilenvc: Real-time 1080p neural video compression on a mobile device. arXiv preprint arXiv:2310.01258, 2023. 2", + "[65] Abhinau K Venkataramanan, Chengyang Wu, Alan C Bovik, Ioannis Katsavounidis, and Zafar Shahid. A hitchhiker's guide to structural similarity. IEEE Access, 9: 28872-28896, 2021. 1", + "[66] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 3", + "[67] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Huchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2012-2021, 2023. 3", + "[68] Ping-Hao Wu, Ioannis Katsavounidis, Zhijun Lei, David Ronca, Hassene Tmar, Omran Abdelkafi, Colton Cheung, Foued Ben Amara, and Faouzi Kossentini. Towards much better svt-av1 quality-cycles tradeoffs for vod applications. In Applications of Digital Image Processing XLIV, pages 236–256. SPIE, 2021. 6", + "[69] Li Xu, Gang He, Jinjia Zhou, Jie Lei, Weiying Xie, Yun-song Li, and Yu-Wing Tai. Transcoded video restoration by temporal spatial auxiliary network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2875-2883, 2022. 3", + "[70] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8):1106-1125, 2019. 3, 5", + "[71] Ren Yang, Mai Xu, and Zulin Wang. Decoder-side hevc quality enhancement with scalable convolutional neural network. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 817-822. IEEE, 2017. 2", + "[72] Ren Yang, Mai Xu, Tie Liu, Zulin Wang, and Zhenyu Guan. Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology, 29(7):2039-2054, 2018. 2", + "[73] Ren Yang, Mai Xu, Zulin Wang, and Tianyi Li. Multi-frame quality enhancement for compressed video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6664-6673, 2018. 2", + "[74] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3106-3115, 2019. 3", + "[75] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video superresolution. In Proceedings of the IEEE/CVF International" + ], + "bbox": [ + 516, + 90, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17753", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision, pages 4429-4438, 2021. 3", + "[76] Hengsheng Zhang, Xueyi Zou, Jiaming Guo, Youliang Yan, Rong Xie, and Li Song. A codec information assisted framework for efficient compressed video superresolution. In European Conference on Computer Vision, pages 220-235. Springer, 2022. 3" + ], + "bbox": [ + 91, + 90, + 468, + 191 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "17754", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Perceptual Video Compression with Neural Wrapping/4f847378-cfb2-4455-91dd-a70a0227cf83_model.json b/2025/Perceptual Video Compression with Neural Wrapping/4f847378-cfb2-4455-91dd-a70a0227cf83_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0ec0f56b1c6c2ff29f20f3b45f5a387610d955bc --- /dev/null +++ b/2025/Perceptual Video Compression with Neural Wrapping/4f847378-cfb2-4455-91dd-a70a0227cf83_model.json @@ -0,0 +1,2853 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.212, + 0.131, + 0.759, + 0.154 + ], + "angle": 0, + "content": "Perceptual Video Compression with Neural Wrapping" + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.181, + 0.885, + 0.217 + ], + "angle": 0, + "content": "Muhammad Umar Karim Khan, Aaron Chadha, Mohammad Ashraful Anam, Yiannis Andreopoulos, Sony Interactive Entertainment" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.22, + 0.812, + 0.234 + ], + "angle": 0, + "content": "{Umar . Khan, Aaron . Chadha, Russell . Anam, Yiannis . Andreopoulos}@sony.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.312, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.473, + 0.773 + ], + "angle": 0, + "content": "Standard video codecs are rate-distortion optimization machines, where distortion is typically quantified using PSNR versus the source. However, it is now widely accepted that increasing PSNR does not necessarily translate to better visual quality. In this paper, a better balance between perception and fidelity is targeted, in order to provide for significant rate savings over state-of-the-art standards-based video codecs. Specifically, pre- and post-processing neural networks are proposed that enhance the coding efficiency of standard video codecs when benchmarked with an array of well-established perceptual quality scores. These \"neural wrapper\" elements are end-to-end trained with a neural codec module serving as a differentiable proxy for standard video codecs. The codec proxy is jointly optimized with the pre- and post components via a novel two-phase pretraining strategy and end-to-end iterative refinement with stop-gradient. This allows the neural pre- and postprocessor to learn to embed, remove and recover information in a codec-aware manner, thus improving its rate-quality performance. A single neural-wrapper model is thereby established and used for the entire rate-quality curve without needing any downscaling or upscaling. The trained model is tested with the AV1 and VVC standard codecs via an array of well-established objective quality scores (SSIM, MS-SSIM, VMAF, AVQT), as well as mean opinion scores (MOS) derived from ITU-T P.910 subjective testing. Experimental results show that the proposed approach improves all quality scores, with \\(-18.5\\%\\) average Bjontegaard Delta-rate (BD-rate) saving over all objective scores and MOS improvement over both standard codecs. This illustrates the significant potential of neural wrapper components over standards-based video coding." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.785, + 0.208, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "It is now widely accepted that that signal-to-noise (SNR) ratio is a poor indicator of visual quality in video coding [6, 19, 26, 42]. Instead, quality scores that include elements of human perception [42], perceptual modelling of encoding artifacts [51], as well as viewing setup awareness [4, 65] have emerged as strong contenders for the best means to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.892, + 0.392 + ], + "angle": 0, + "content": "objectively characterize visual quality. This has also led to the research community now moving away from SNR-optimization [26] in favor of structural similarity (SSIM [65]), video multimethod assessment fusion (VMAF) and Apple's advanced video quality tool (AVQT) optimization [4, 42]. However, all current perceptual optimization approaches in standards-based video encoders have one or more of the following detriments:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.396, + 0.891, + 0.425 + ], + "angle": 0, + "content": "- they require multiple encoding passes or in-loop implementation within a specific encoder;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.426, + 0.891, + 0.47 + ], + "angle": 0, + "content": "- they only optimize for a single quality scoring method like VMAF or SSIM and are shown to be detrimental in other quality scores;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.471, + 0.891, + 0.531 + ], + "angle": 0, + "content": "- they comprise hand-crafted (shallow) models of low-level human perception and fail to encapsulate several characteristics of more advanced quality scoring methods like VMAF or AVQT in a data-driven and learnable manner;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.532, + 0.891, + 0.577 + ], + "angle": 0, + "content": "their Bjontegaard Delta-rate (BD-rate) improvement on well-established quality scores like SSIM, VMAF and AVQT is in the order of a few percentile points." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.396, + 0.891, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.578, + 0.892, + 0.759 + ], + "angle": 0, + "content": "A neural codec \"wrapper\" is proposed in this paper as the means to augment standard video codecs and optimize them for perceptual quality. Including neural pre- and postprocessing in conventional encoding pipelines is shown to clearly outperform the equivalent codecs without these components. This allows for conventional streaming ecosystems to continue operating \"as-is\" and does not break standards. That is, when neural components cannot be supported before encoding or after decoding, the system can just disable them with appropriate adaptation, without jeopardizing video streaming. The key contributions of the paper are listed as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.892, + 0.826 + ], + "angle": 0, + "content": "- A novel pretraining regime for modelling a conventional codec with a neural codec model is proposed, which uses the implicit encoder-decoder structure of the neural codec to align both video quality and rate." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- An end-to-end training regime for jointly training pre and postprocessors is proposed. Starting with a pretrained neural codec model, it is proposed to iteratively update the codec model. This approach allows for more accurate gradients to be backpropagated to the prepro" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17743" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.093, + 0.833, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.26, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Figure 1. Left: Functional block diagram of the multi-stage training data flow. Right: Components of codec model \\( M \\) that are used as a proxy for the target codec \\( C \\). \\( AE/AD \\) denotes arithmetic encoding followed by arithmetic decoding and \\( Q \\) denotes quantization. See Sec. 3 for details." + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.331, + 0.158, + 0.343 + ], + "angle": 0, + "content": "cessor." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.344, + 0.473, + 0.452 + ], + "angle": 0, + "content": "- The trained pipeline is evaluated on gaming sequences and significant BD-rate gains versus AV1 and VVC encoding are demonstrated with just a single pre- and postprocessor model. In addition, the proposed neural wrapper is shown to outperform a state-of-the-art neural encoder [35], as well as the state-of-the-art in deep perceptual preprocessing proposed recently [11]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.471, + 0.532 + ], + "angle": 0, + "content": "A functional block diagram for the proposed system is shown in Fig. 1. Example visuals produced by the proposed method are shown in Fig. 4, demonstrating that the proposed neural wrapper framework allows for better detail preservation than the underlying standard codec." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.235, + 0.567 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.576, + 0.29, + 0.591 + ], + "angle": 0, + "content": "2.1. Neural Video Codes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Neural video codecs optimize neural networks for rate and distortion by exploiting redundancies in video data. Most notable are residual coding methods; Lu et al. [46] introduce deep video codec, which adopts neural networks within the residual coding pipeline of a traditional codec. Optical flow is employed as a proxy for motion vectors in order to warp the reference frame and generate a residual frame for encoding. Numerous other methods have been proposed based on residual coding [44, 49, 53]. Scale-space flow estimation is proposed as an improvement over traditional optical flow by Agustsson et al. [2], namely for better handling of complex motion in the scene. Beyond residual coding methods, conditional coding methods learn feature contexts implicitly from the encoded context [23, 33, 34, 45, 47, 56]. Li et al. [35] propose using diverse contexts in spatial and temporal directions to perform video compression and outperform standard video codecs when the latter are used in low-delay mode. They improve the performance and bitrate range in [36]. Recent work has additionally focused on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.895, + 0.406 + ], + "angle": 0, + "content": "reducing the complexity of neural CODECs. For example, Van Rozendaal et al. [64] deploy an architecture similar to Agustsson et al. [2] on a smart phone and achieve real time performance at 1080p. Hu et al. [25] adopt slimmable networks for variable complexity." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.418, + 0.659, + 0.435 + ], + "angle": 0, + "content": "2.2. Preprocessors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.895, + 0.76 + ], + "angle": 0, + "content": "While postprocessing is a more mature line of research, there has been a prevalence of work in preprocessing, both on image and video modalities. In the image domain, Strumpler et al. [58] propose to modify the JPEG encoder with an attention network and a learnable quantization table, in order to modulate the source frame and provide for additional rate savings over the JPEG encoder alone. Conversely, Talebi et al. [59] avoid adapting the JPEG encoder by prepended the standard encoder with a preprocessor for deployment with a standard JPEG encoder and decoder. Klopp et al. [31] train a preprocessor with a surrogate neural codec model, which is trained to model a target image codec. In the video domain, Boutssoulatze et al. [9] propose a precoding network for downscaling with a online mode selection algorithm, in order to alter the precoding based on scene characteristics. Conversely, Chadha et al. [11] propose an entirely offline training mechanism by first modelling the core components of a video codec in a differentiable manner. A preprocessor is optimized on the codec model, in order to provide further rate-quality savings over the target codec." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.772, + 0.665, + 0.789 + ], + "angle": 0, + "content": "2.3. Postprocessors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Numerous deep learning methods have been proposed for compressed video denoising, video super-resolution or both tasks. Single frame methods for denoising compressed videos have been proposed in [71, 72]. Multiframe methods have provided superior gains in metrics compared to single frame methods. MFQE [73] detects frames with higher quality and uses it as a reference to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17744" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.468, + 0.5 + ], + "angle": 0, + "content": "improve neighboring frames. TSAN [69] uses a temporal deformable alignment and multi-scale fusion for the task of video restoration. STDF [16] uses spatio-temporal deformable convolution to aggregate temporal context rather than using optical flow for frame warping. A common approach for video super-resolution is to align neighboring frames and transfer information across frames [60, 61, 66, 70]. On the contrary, other methods choose non-local modules or 3D convolutions to gather temporal context [29, 30, 38, 74]. Some methods recurrently propagate history in the forward direction [17, 28, 55]. Recent methods such as BasicVSR [12], GOVSR [75] and BasicVSR++ [13] propagate history in forward and backward directions, such that both past and future frames are used to super-resolve the current frame. Architectures developed for video super-resolution can easily be adapted for video denoising. Few authors have focused on the task of jointly denoising and upscaling videos. FTVSR [50] uses a frequency transformer which performs self-attention in space, time and frequency to recover details. RealBasicVSR [14] considers sensor noise, motion blur and compression artifacts with super-resolution. CIAF [76] uses motion-vectors from the video to approximate optical flow. CAVSR [67] adapts the super-resolution to the compression level of the input video. COMISR [39] uses an architecture based on BasicVSR [12] to perform joint denoising and upscaling." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.515, + 0.468, + 0.545 + ], + "angle": 0, + "content": "2.4. Pre- and Postprocessors with Traditional Codes" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.554, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Recent methods have also explored joint training of pre and postprocessors as neural wrappers. In the image domain, recent work [20, 21] emulates the core blocks of the JPEG encoder and decoder, and uses it to train pre and postprocessors for rate-distortion savings. Similar sandwich approach for videos has been proposed in [24]. Son et al. [57] propose distinct neural networks to model the distortion and rate behavior of a standard codec, which is thus sandwiched by pre- and postprocessor networks for rate-distortion optimization. However, these methods have no target codec awareness beyond the implicit intra/inter coding structure. Either the codec proxy is not learned (e.g. derived from differentiable JPEG/DCT as in [24]), only partially learned via rate-distortion optimization (e.g. the loop filter as in [27]) or only operates in intra mode (as in [57]). In this paper, the proposed framework is directly conditioned on the target conventional codec. Tian et al. [62] target additional rate savings with learnable rescaling. However, this method has no rate control or alignment outside of the scaling factor. Conversely, the framework proposed in this paper is jointly conditioned on both the target codec rate and distortion and is thus able to offer rate savings even when encoding at full res" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "olution. Additionally, whereas recent work restricts evaluation to a limited range of the rate-distortion space, at high bitrates [61] or substantially lower bitrates and resolutions [22] and on older encoding standards (AVC or HEVC), the proposed method is evaluated against state-of-the-art AV1 and VVC encoding, i.e.: \\((i)\\) the entire rate-quality region of practical significance is considered (VMAF values between 40-96 and bitrates from 100kbps to 15mbps); \\((ii)\\) slow encoding presets are used (preset=4 for SVT-AV1 and preset=slow for VVenC), without limiting the encoders to low-delay mode." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.271, + 0.647, + 0.287 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.296, + 0.892, + 0.522 + ], + "angle": 0, + "content": "The top-level functional block diagram of the proposed training flow is shown in Fig. 1 and includes a preprocessor \\( P \\), codec model \\( M \\), target codec \\( C \\) and postprocessor \\( O \\). Notably, the target codec is a traditional video codec, such as SVT-AV1 [32]. First, a source frame \\( I_{\\mathrm{in}} \\) is processed by \\( P \\) to generate a corresponding preprocessed frame \\( I_P \\). The preprocessed frame is simultaneously encoded by \\( M \\) and \\( C \\), which generate encoded frames \\( I_M \\) and \\( I_C \\) respectively. Finally, encoded frames are postprocessed with \\( O \\) to generate the output frame \\( I_{\\mathrm{out}} \\). End-to-end distortion \\( D \\) is computed between \\( I_{\\mathrm{in}} \\) and \\( I_{\\mathrm{out}} \\) and the overall training loss is the weighted sum of \\( D \\) and the bitrate \\( R_M \\). Notably, at inference, the codec model \\( M \\) is removed from the pipeline and only the target codec \\( C \\) is deployed for encoding." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.531, + 0.744, + 0.547 + ], + "angle": 0, + "content": "3.1. Codec Model Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.554, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Given the target codec \\( C \\) in Fig. 1 is not differentiable, a mechanism for adapting the proposed pipeline to the target codec behavior is required. One option is black-box or derivative-free optimization methods [3, 7, 18]. However these methods typically rely on an approximation of the unknown function to be optimized, and are computationally infeasible for training at scale. Another option is to replicate the core prediction and transform blocks of the target codec in a differentiable manner [15, 27, 54]. Such approaches can model older codecs such as JPEG, or faster presets of AVC or HEVC successfully to an extent but fail to accurately model the operation of more sophisticated codecs such as AV1 or VVC." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.751, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Recent work [31, 57] has instead leveraged on neural image and video CODEs for predicting the target codec behavior. In this work, a similar approach is followed, and a neural video codec \\(M\\) is adapted to model the ratedistortion behavior of a target codec \\(C\\). Specifically, the scale-space flow architecture [2] is used to generate lossy frames, and the implicit encoder-decoder structure and associated loss functions are leveraged to align to the target codec behavior in a spatially localized manner. As per Agustsson et al. [2], for each of intra, flow and residual" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17745" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "encoding, the hyperprior based architecture of Balle et al. [5] is used, without any autoregressive components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.468, + 0.348 + ], + "angle": 0, + "content": "Notably, the hyperprior architecture comprises analysis and synthesis transforms, \\( g_{a} \\) and \\( g_{s} \\), for encoding to and decoding from the quantized latent space \\( \\tilde{y} \\) respectively. A distortion loss \\( D(:,g_a,g_s) \\) can thus be computed between source frame \\( I_{in} \\) (input to \\( g_{a} \\)) and reconstructed frame \\( I_M \\) (output by \\( g_{s} \\)). A hyper-analysis and hypersynthesis transform \\( (h_a \\) and \\( h_s \\) respectively) constitute the hyperprior, which is used to model scale dependencies \\( \\tilde{\\sigma} \\) within the latent space. The latent space \\( \\tilde{y} \\) is thus modelled as a Gaussian scale mixture on which the rate \\( R(:,h_a,h_s) \\) can be optimized (as Shannon entropy). Traditionally, the architecture would then be trained with an aggregate loss function \\( L = D + \\lambda R \\), where \\( \\lambda \\) represents a Lagrangian parameter for balancing between the rate and distortion components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.468, + 0.499 + ], + "angle": 0, + "content": "When performing neural codec alignment to a target codec, there are two key issues: (i) the model encoded frame must align to the target encoded frame (and not the source); (ii) selecting the best value for \\(\\lambda\\) when trying to replicate target codec behavior is a non-trivial problem. Crucially, it is noted that the rate model is exclusively parameterized by \\(h_a\\) and \\(h_s\\), and distortion exclusively by \\(g_a\\) and \\(g_s\\). Subsequently, a novel two phase training methodology is proposed for aligning both rate and distortion to the target codec:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.468, + 0.589 + ], + "angle": 0, + "content": "Phase 1: The first issue of target codec alignment is addressed via the distortion loss. To this end, the reference for the distortion loss is changed from the source frame \\( I_{in} \\) to the target codec frame \\( I_C \\) and the mean squared error is optimized:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.595, + 0.468, + 0.613 + ], + "angle": 0, + "content": "\\[\nD _ {M} ^ {\\mathrm {h y p}} \\left(I _ {M}, I _ {C}; \\theta_ {g}\\right) = \\mathbb {E} _ {I _ {C}} \\left[ \\left\\| I _ {M} - I _ {C} \\right\\| ^ {2} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.468, + 0.651 + ], + "angle": 0, + "content": "where \\(\\theta_{g}\\) represent the analysis and synthesis transform parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.468, + 0.803 + ], + "angle": 0, + "content": "Phase 2: Following alignment of \\( g_{a} \\) and \\( g_{s} \\) in Phase 1, the latent space \\( \\tilde{y} \\) is now conditioned on the target codec \\( C \\). To circumvent selection of an optimal \\( \\lambda \\) for rate, the implicit encoder-decoder structural prior of the aligned neural codec is leveraged in order to learn an unsupervised rate alignment to the target codec. Specifically, the maximum likelihood on the hyperprior model parameters is optimized, in order to learn the latent space distribution under target codec distortion alignment. The rate is thus optimized as:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.808, + 0.503, + 0.84 + ], + "angle": 0, + "content": "\\[\nR _ {M} ^ {\\mathrm {h y p}} \\left(I _ {M}; \\theta_ {h}\\right) = - \\mathbb {E} _ {\\tilde {y}, \\tilde {z}} \\left[ \\log p _ {\\tilde {y} | \\tilde {z}, I _ {C}} \\left(\\tilde {y} | \\tilde {z}, I _ {C}\\right) + \\log p _ {\\tilde {z} | I _ {C}} \\left(\\tilde {z} | I _ {C}\\right) \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.468, + 0.901 + ], + "angle": 0, + "content": "where \\(\\theta_h\\) represents the hyperprior model parameters and \\(\\tilde{z}\\) is the quantized hyperprior latent space (which would also be encoded and transmitted as side information in a neural codec pipeline). For intra coding, the intra" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "hyperprior model is optimized on intra encoded source frames using the two-phase strategy. For inter coding, the residual and flow hyperprior models are optimized jointly by summing the rate and distortion components in each phase." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.892, + 0.243 + ], + "angle": 0, + "content": "The proposed two-phase pretraining ensures that \\( R_{M}^{\\mathrm{hyp}} \\) is close to the target codec \\( R_{C} \\) by performing an affine transform on \\( R_{M}^{\\mathrm{hyp}} \\), i.e., \\( R_{M} = \\alpha R_{M}^{\\mathrm{hyp}} + \\beta \\), where \\( \\alpha \\) and \\( \\beta \\) are scalar learnable parameters optimized with MSE:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.246, + 0.892, + 0.267 + ], + "angle": 0, + "content": "\\[\nL _ {M} ^ {(\\alpha , \\beta)} \\left(R _ {M}, R _ {C}; \\alpha , \\beta\\right) = \\mathbb {E} _ {R _ {C}} \\left[ \\left\\| R _ {M} - R _ {C} \\right\\| ^ {2} \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.892, + 0.34 + ], + "angle": 0, + "content": "The similarity of \\( R_{M}^{\\mathrm{hyp}} \\) and \\( R_{C} \\) is required to properly balance the rate and distortion losses in end-to-end training. The aggregate rate and distortion for a sequence can thus be measured by computing the sum over all frames." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.349, + 0.713, + 0.365 + ], + "angle": 0, + "content": "3.2. End-to-End Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.892, + 0.462 + ], + "angle": 0, + "content": "With the pretrained neural codec model, an end-to-end trainable pipeline with both the pre and postprocessor is developed, as illustrated in Fig. 1. In this case, the training loss is very similar to conventional neural CODECs, but is now exclusively used to train the pre and postprocessors, i.e.:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.467, + 0.892, + 0.485 + ], + "angle": 0, + "content": "\\[\nL \\left(I _ {\\text {o u t}}, I _ {\\text {i n}}; \\theta_ {P}, \\theta_ {O}\\right) = \\mathbb {E} _ {I _ {\\text {i n}}} \\left[ D \\left(I _ {\\text {i n}}, I _ {\\text {o u t}}\\right) + \\lambda R _ {M} \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.495, + 0.892, + 0.662 + ], + "angle": 0, + "content": "where \\(\\theta_{P}\\) and \\(\\theta_{O}\\) are the pre and postprocessor parameters respectively, \\(R_{M}\\) is the codec model rate estimation (proportional to \\(R_{M}^{\\mathrm{flow}} + R_{M}^{\\mathrm{res}}\\) for inter coding the flow and residual from motion compensation, and \\(R_{M}^{\\mathrm{intra}}\\) for intra coding) and \\(D\\) is the distortion loss component. Following subjective assessment tests, a weighted sum of: mean absolute error (MAE), SSIM, MS-SSIM and detail loss metric (DLM) * [37] is proposed for the distortion component \\(D\\) in Eq. (4). Specifically, \\(D(I_{\\mathrm{in}},I_{\\mathrm{out}}) = \\lambda_{\\mathrm{L1}}||I_{\\mathrm{in}} - I_{\\mathrm{out}}|| + \\lambda_{\\mathrm{SSIM}}(1 - D_{\\mathrm{SSIM}}(I_{\\mathrm{in}},I_{\\mathrm{out}})) + \\lambda_{\\mathrm{MSSSIM}}(1 - D_{\\mathrm{MSSSIM}}(I_{\\mathrm{in}},I_{\\mathrm{out}})) + \\lambda_{\\mathrm{DLM}}D_{\\mathrm{DLM}}(I_{\\mathrm{in}},I_{\\mathrm{out}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.662, + 0.892, + 0.844 + ], + "angle": 0, + "content": "Concept drift or domain shift for the codec model is an expected consequence of end-to-end training. The codec model is pretrained with natural videos whereas the preprocessor is continuously updated as training proceeds; thus, the data observed by the codec model has a non-stationary distribution. Since neural networks generally do not perform well on out-of-distribution data, an iterative refinement mechanism is proposed, where the codec model parameters \\(\\theta_{M}\\) are independently updated after each iteration of end-to-end training using the two-phase approach described previously. Notably, in this case the input to the target codec is now the output of the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.852, + 0.892, + 0.901 + ], + "angle": 0, + "content": "* DLM is one of the core components of VMAF [41] and is found to better preserve details under visual degradation. Given that it is primarily wavelet based, it can be implemented in a differentiable manner and used for model optimization." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17746" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "preprocessor \\(I_{P}\\) and not the source frame, and \\(I_{C}\\) and \\(R_{C}\\) represent the frame and rate output by the target codec." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.47, + 0.243 + ], + "angle": 0, + "content": "Additionally, the gradients backpropagated to the preprocessor can be better aligned towards the target codec via the stop-gradient operator. First, it is assumed for the remaining discussion in the section that the tensors representing the frames are flattened, i.e., \\(\\{I_P,I_M,I_C\\} \\in \\mathbb{R}^{N\\times 1}\\), where \\(N\\) is the number of pixels in the tensor. If the target codec \\(C\\) was differentiable then the accuracy of the gradients backpropagated will increase by decreasing:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.25, + 0.469, + 0.267 + ], + "angle": 0, + "content": "\\[\nG = \\left| J _ {M} \\left(I _ {P}\\right) \\nabla_ {I _ {M}} L - J _ {C} \\left(I _ {P}\\right) \\nabla_ {I _ {C}} L \\right|, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.271, + 0.469, + 0.346 + ], + "angle": 0, + "content": "where \\(J_{M}(I_{P})\\in \\mathbb{R}^{N\\times N}\\) is the transposed Jacobian of \\(M\\) w.r.t. \\(I_P\\) and \\(\\nabla_{I_M}L\\) is the gradient of the objective w.r.t. \\(I_M\\). Assuming that the frames \\(I_C\\) generated by the target codec \\(C\\) are input to the postprocessor \\(O\\) then Eq. (5) becomes" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.35, + 0.469, + 0.367 + ], + "angle": 0, + "content": "\\[\nG = \\left| \\left(J _ {M} \\left(I _ {P}\\right) - J _ {C} \\left(I _ {P}\\right)\\right) \\nabla_ {I _ {C}} L \\right|, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.469, + 0.414 + ], + "angle": 0, + "content": "Since \\(\\nabla_{I_C}^{(j)}L\\) is necessary for the networks to train, \\(J_M^{(i,j)}(I_P) - J_C^{(i,j)}(I_P)\\) should be close to zero. Thus," + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.426, + 0.469, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\forall i, j \\in [ 1, N ], \\left| \\frac {\\partial I _ {M} ^ {(i)}}{\\partial I _ {P} ^ {(j)}} - \\frac {\\partial I _ {C} ^ {(i)}}{\\partial I _ {P} ^ {(j)}} \\right| = \\left| \\frac {\\partial}{\\partial I _ {P} ^ {(j)}} \\left(I _ {M} ^ {(i)} - I _ {C} ^ {(i)}\\right) \\right| \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.469, + 0.585 + ], + "angle": 0, + "content": "should be minimized. The above analysis can be extended to the rate component. Eq. (7) shows that accurate gradients can be backpropagated to the preprocessor if the codec model's output is close to the target codec's output and the input to the postprocessor is from the target video codec. In the proposed training regime, the input to the postprocessor is therefore set as" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.587, + 0.469, + 0.605 + ], + "angle": 0, + "content": "\\[\nI _ {O _ {\\text {i n}}} = \\overline {{\\left(I _ {C} - I _ {M}\\right)}} + I _ {M}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.469, + 0.689 + ], + "angle": 0, + "content": "where \\(\\overline{a}\\) indicates the stop-gradient operator [63] applied to the tensor \\(a\\) and \\(I_{C}\\) is the output of the non-differentiable video codec. Eq. (8) allows frames generated by the target codec to be passed to the postprocessor while using the codec model for backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.469, + 0.81 + ], + "angle": 0, + "content": "In summary, the proposed training procedure has two main steps; iterative update of the codec model and the end-to-end update of the pre- and postprocessors. For the end-to-end update, a stop-gradient-based approach is proposed. Simple CNNs are used for pre and postprocessing. Details of the pre and postprocessor architecture are in the supplementary. The end-to-end training pipeline is shown in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.296, + 0.839 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.195, + 0.864 + ], + "angle": 0, + "content": "4.1. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The Vimeo-90k dataset [70] was used to train the codec model and perform end-to-end training. The training" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "Algorithm 1 End-to-end training pipeline of the pre and postprocessor" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.126, + 0.744, + 0.141 + ], + "angle": 0, + "content": "Require: \\(I_{\\mathrm{in}}\\) , batch of video frames" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.142, + 0.805, + 0.156 + ], + "angle": 0, + "content": "Require: \\(\\theta_{M}\\), parameters of the codec model" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.157, + 0.802, + 0.171 + ], + "angle": 0, + "content": "Require: \\(\\theta_{P}\\), parameters of the preprocessor" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.172, + 0.808, + 0.186 + ], + "angle": 0, + "content": "Require: \\(\\theta_{O}\\), parameters of the postprocessor" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.187, + 0.763, + 0.202 + ], + "angle": 0, + "content": "Require: \\(\\alpha_{L}\\), learning rate for training" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.203, + 0.609, + 0.216 + ], + "angle": 0, + "content": "for all \\(I_{\\mathrm{in}}\\) do" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.217, + 0.635, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {P} \\leftarrow P (\\boldsymbol {I} _ {\\text {i n}})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.233, + 0.679, + 0.247 + ], + "angle": 0, + "content": "\\[\n\\left(\\boldsymbol {I} _ {C}, \\boldsymbol {R} _ {C}\\right) \\leftarrow C \\left(\\boldsymbol {I} _ {P}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.248, + 0.69, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\left(\\boldsymbol {I} _ {M}, \\boldsymbol {R} _ {M}\\right) \\leftarrow M \\left(\\boldsymbol {I} _ {P}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.263, + 0.8, + 0.28 + ], + "angle": 0, + "content": "Get \\(D_M^{hyp}\\) and \\(\\theta_g\\gets \\theta_g - \\alpha_L\\nabla_{\\theta_g}D_M^{hyp}\\)" + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.281, + 0.799, + 0.297 + ], + "angle": 0, + "content": "Get \\(R_M^{hyp}\\) and \\(\\theta_h\\gets \\theta_h - \\alpha_L\\nabla_{\\theta_h}R_M^{hyp}\\)" + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.297, + 0.879, + 0.315 + ], + "angle": 0, + "content": "Get \\(L_{M}^{(\\alpha ,\\beta)}\\) and \\((\\alpha ,\\beta)\\gets (\\alpha ,\\beta) - \\alpha_{L}\\nabla_{(\\alpha ,\\beta)}L_{M}^{(\\alpha ,\\beta)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.316, + 0.726, + 0.329 + ], + "angle": 0, + "content": "Generate \\(I_{O_{\\mathrm{in}}}\\) using Eq. (8)" + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.33, + 0.652, + 0.345 + ], + "angle": 0, + "content": "\\(\\pmb{I}_{\\mathrm{out}}\\gets O(\\pmb{I}_{O_{\\mathrm{in}}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.346, + 0.672, + 0.359 + ], + "angle": 0, + "content": "Get \\(L\\) using Eq. (4)" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.36, + 0.689, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\theta_ {P} \\leftarrow \\theta_ {P} - \\alpha_ {L} \\nabla_ {\\theta_ {P}} L\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.376, + 0.689, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\theta_ {O} \\leftarrow \\theta_ {O} - \\alpha_ {L} \\nabla_ {\\theta_ {O}} L\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.519, + 0.392, + 0.571, + 0.403 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.892, + 0.794 + ], + "angle": 0, + "content": "pipeline and the pre- and postprocessor are implemented in the YUV colorspace. A fixed learning rate of \\(10^{-4}\\) is used for each network with the rest of hyper-parameters set as: \\(\\lambda = 0.1\\); \\(\\lambda_{\\mathrm{L1}} = 0.1\\), \\(\\lambda_{\\mathrm{SSIM}} = 0.25\\); \\(\\lambda_{\\mathrm{MSSSIM}} = 0.25\\); \\(\\lambda_{\\mathrm{DLM}} = 0.2\\). Batch size, GOP and maximum iterations are set to 4, 4 and 500000, respectively. Both Phase 1 and 2 of pretraining are performed for 500000 iterations. The model is trained on source frames only in an open-loop configuration; i.e., no reconstructed frames are fed into the preprocessor as input. A low-delay recipe of SVT-AV1 (described in supplementary) is used for codec model alignment for pre-training and end-to-end training. The codec model encoding is aligned to the GOP of the target codec, such that intra and inter coding are also aligned. For experiments on VVC, it was found that codec model alignment to a low-delay recipe of VVenC provided slight BD-rate gains over SVT-AV1 alignment, albeit at high computational overhead during training. Therefore, the presented results are carried out with a single neural wrapper model (one encode) developed based on codec proxy model alignment to SVT-AV1. This is practically more feasible than [24], which requires 9 combinations of models (9 encodes) + convex hull optimization and is thus far more computationally demanding." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.802, + 0.703, + 0.816 + ], + "angle": 0, + "content": "4.2. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Four 1080p video datasets were used for evaluation: (i) 16 gaming sequences captured in a commercial setup (Gaming); (ii) The 16 XIPH natural sequences used by recent work [11] (XIPH); (iii) The 1080p sequences of the UVG [48] dataset (UVG); (iv) HEVC-B [8] dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17747" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "(HEVC-B). The proposed method was applied with SVT-AV1 [32] for AV1 and VVenC [10] for VVC. All experiments were performed with YUV 4:2:0 8-bit sequences. In all cases, the protocol of Li et al. [35] is followed and 96 frames are evaluated. However, unlike Li [35], both standard encoders are used without their low-delay limitation, in order to attain their best performance (see supplementary for details on sequences and encoding recipes). Ablation results are with the SVT-AV1 lowdelay recipe (see supplementary). Results in Tab. 3 and 4 are on the Gaming dataset." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.258, + 0.473, + 0.53 + ], + "angle": 0, + "content": "In order to cover a wide range of well-established objective quality scoring methods, the following scores are used: SSIM, MS-SSIM, VMAF [42, 52] and the AVQT score from Apple [4]. In order to showcase the generality of the proposed approach, the same pre- and postprocessor models are used for: (i) all the points on the RD curves, (ii) all quality scores and (iii) both standard codecs. Perceptual models are used for competing methods when available. Comparison is provided against a competitive video preprocessor (DPP [11]), SOTA neural codecs (DCVC-DC [35] and DCVC-FM [36]), and perceptual modes of standard codecs. The comparisons include results of end-to-end training with the codec model of Hu et al. [24] with the proposed network architectures and losses used to illustrate the effectiveness of the proposed codec model. For DCVC-DC and DCVC-FM, the best available pretrained models (publicly provided by the authors) are used." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.54, + 0.357, + 0.556 + ], + "angle": 0, + "content": "4.3. Objective Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.473, + 0.866 + ], + "angle": 0, + "content": "The quantitative results of the proposed method are shown in† Fig. 2, and Tab. 1 and Tab. 2. These results show that the proposed approach provides for BD-rates between \\(-5.4\\%\\) to \\(-30\\%\\) in comparison to the underlying standard encoders. BD-rate savings tend to be higher for quality scores that are more perceptually-oriented. Compared to competing methods, the proposed approach is shown to offer significantly more consistent BD-rate improvements over all quality scoring methods (ALL columns of Tab. 1 and Tab. 2). Importantly, as shown in Fig. 2, unlike DCVC-DC that is only covering a very limited bitrate-quality range at the very top end, the proposed approach outperforms all other methods on the entire \"active\" region of bitrate-quality, i.e., VMAF of 40-93 [32], which encapsulates the entirety of commercially-relevant bitrates for 1080p video streaming. Given that the neural wrapper is optimizing perceptually, BD-rates for PSNR tend to show regression on PSNR (see supplementary) but visual inspection and an array of perceptual metrics indicate that the proposed approach offers significant benefits" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.099, + 0.685, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.197, + 0.608, + 0.208 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.099, + 0.879, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.196, + 0.804, + 0.208 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.219, + 0.685, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.316, + 0.608, + 0.327 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.221, + 0.879, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.315, + 0.805, + 0.327 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.338, + 0.682, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.436, + 0.607, + 0.446 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.338, + 0.877, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.435, + 0.804, + 0.447 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.457, + 0.682, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.554, + 0.608, + 0.566 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.457, + 0.877, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.554, + 0.804, + 0.565 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.578, + 0.892, + 0.621 + ], + "angle": 0, + "content": "Figure 2. Combined rate-quality plots over Gaming ((a), (b)), XIPH ((c), (d)), UVG ((e), (f)) and HEVC-B ((g), (h)) datasets. See supplementary for more results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.892, + 0.698 + ], + "angle": 0, + "content": "in visual quality. Nonetheless, the results of DCVC-DC and DCVC-FM are remarkably good given that they are fully neural approaches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.785, + 0.728 + ], + "angle": 0, + "content": "4.4. Subjective Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To assess the visual quality of the proposed approach, a 5-scale absolute categorical rating (ACR) test with hidden reference was conducted under the ITU-T P.910 test conditions and SUREAL post-processing [40, 43] (see supplementary for details). Raters were shown videos generated by the codec and proposed method. Separate tests were conducted for AV1 and VVC with each test comprising 72 sequences. The combined rate-MOS plots are shown in Fig. 3. The BD-rates reported in the figure for the proposed approach are mostly in line with AVQT and VMAF BD-rates reported for Gaming in Tab. 1 and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.471, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) All combined plots of Fig. 2 and Fig. 3 are produced with the slope-based integration method of Wu et al. [68] and are shown in log-scale." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17748" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.132 + ], + "angle": 0, + "content": "Table 1. BD-rates vs. AV1 (more negative is better). AV1+SSIM denotes codec tuned for SSIM (-tune ssim). ALL indicates average BD-rate over all quality scores." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.142, + 0.465, + 0.434 + ], + "angle": 0, + "content": "
DatasetMethodSSIMMSSSIMAVQTVMAFALL
GamingAV1+SSIM-4.5-5.4-5.12.1-3.2
AV1+DPP [11]-5.2-6.1-17.6-11.3-10.0
AV1+[24]14.96.312.2-1.48.0
DCVC-DC [35]-15.4-36.1-35.621.84-16.3
DCVC-FM [36]14.86.04.2-2.95.5
AV1+Proposed-15.5-13.8-19.6-30.0-19.7
XIPHAV1+SSIM-3.5-4.0-5.43.2-2.4
AV1+DPP [11]-4.3-3.9-9.0-16.7-8.5
AV1+[24]12.96.14.51.76.3
DCVC-DC [35]-18.4-25.512.032.20.3
DCVC-FM [36]17.69.08.23.59.6
AV1+Proposed-10.9-7.1-28.8-30.1-19.2
UVGAV1+SSIM-2.8-3.1-2.82.9-1.4
AV1+DPP [11]-3.8-3.0-9.2-21.7-9.4
AV1+[24]9.14.43.33.25.0
DCVC-DC [35]-20.8-10.24.218.2-2.1
DCVC-FM [36]20.49.69.89.412.3
AV1+Proposed-16.2-9.7-22.2-31.9-20.0
HEVC-BAV1+SSIM-5.4-5.3-4.72.9-3.1
AV1+DPP [11]-8.7-6.9-5.7-20.1-10.3
AV1+[24]2.61.8-3.11.10.6
DCVC-DC [35]-35.3-29.1-11.637.2-9.7
DCVC-FM [36]-6.5-11.7-7.7-5.8-7.9
AV1+Proposed-14.5-13.4-15.1-41.8-21.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.451, + 0.47, + 0.495 + ], + "angle": 0, + "content": "Table 2. BD-rates vs. VVC (more negative is better). \\(\\mathrm{VVC + QPA}\\) indicates quantization parameter adaptation is enabled. ALL indicates average BD-rate over all quality scores." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.504, + 0.465, + 0.795 + ], + "angle": 0, + "content": "
DatasetMethodSSIMMSSSIMAVQTVMAFALL
GamingVVC+QPA0.44.613.41.54.9
VVC+DPP [11]-1.3-1.6-9.2-9.0-5.3
VVC+[24]39.532.311.47.822.7
DCVC-DC [35]9.1-11.5-21.338.93.8
DCVC-FM [36]4.918.410.918.413.1
VVC+Proposed-16.6-8.7-19.0-14.3-14.6
XIPHVVC+QPA-0.92.21.12.71.3
VVC+DPP [11]-2.3-1.6-10.5-10.4-6.2
VVC+[24]29.424.412.29.518.9
DCVC-DC [35]-8.1-12.09.241.17.5
DCVC-FM [36]4.313.10.516.88.7
VVC+Proposed-11.4-5.4-16.5-18.0-12.8
UVGVVC+QPA-0.11.67.71.52.7
VVC+DPP [11]-1.4-1.1-7.0-11.4-5.2
VVC+[24]26.823.57.96.516.2
DCVC-DC [35]-8.216.0-1.429.18.9
DCVC-FM [36]-2.714.112.514.79.65
VVC+Proposed-11.8-5.4-13.9-13.8-11.2
HEVC-BVVC+QPA0.91.21.12.11.3
VVC+DPP [11]-19.2-16.7-17.2-16.7-17.4
VVC+[24]0.50.50.1-2.3-0.3
DCVC-DC [35]-17.0-5.56.349.98.4
DCVC-FM [36]18.815.814.412.115.3
VVC+Proposed-31.8-27.4-25.2-33.3-29.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Tab. 2. Sample visual results of the proposed method are shown in Fig. 4. The results show that the proposed approach is temporally coherent, and better capable of preserving structure, retaining textures and denoising codec artifacts, also reflected in the MOS scores of Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.1, + 0.681, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.201, + 0.612, + 0.212 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.1, + 0.868, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.2, + 0.801, + 0.212 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.224, + 0.862, + 0.239 + ], + "angle": 0, + "content": "Figure 3. Combined rate-quality plots with P.910 MOS." + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.254, + 0.865, + 0.269 + ], + "angle": 0, + "content": "Table 3. BD-rates of different component configurations." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.279, + 0.835, + 0.34 + ], + "angle": 0, + "content": "
PrePostSSIMMSSSIMAVQTVMAF
✓✓-25.6-20.8-21.4-20.5
✓✗17.913.229.8-34.1
✗✓-2.90.44.1-33.3
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.364, + 0.618, + 0.379 + ], + "angle": 0, + "content": "4.5. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.388, + 0.892, + 0.569 + ], + "angle": 0, + "content": "Two-phase pretraining: The rate-quality alignment between target codec \\( C \\) and model \\( M \\) after two-phase pretraining is validated by comparing their performance on a dataset ov dieo sequences. The Pearson correlation coefficient (PCC) is improved from 0.56 and 0.60 to 0.88 and 0.85 for bitrate and PSNR, respectively (see supplementary for more results). Note that the PCC for rate is computed between \\( R_{C} \\) and \\( R_{M}^{hyp} \\). This indicates that both distortion and bitrate of the codec model and target codec correlate better after the two-phase pretraining, which is a sufficient condition for gradient-based optimization [31]. Tab. 4 shows the effect of pretraining on wrappers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.57, + 0.893, + 0.841 + ], + "angle": 0, + "content": "Pre/post processing: Tab. 3 shows the effect of removing either the pre or the postprocessor from the overall pipeline under the same training pipeline. Removing the preprocessor is equivalent to training a video denoiser. While BD-rate gains for VMAF are maintained with only the pre- and postprocessor, this can be attributed to sharpening effects over the source content, which VMAF has susceptibility to [1]. Indeed, with more naive content enhancement such as sharpening, fidelity oriented methods such as SSIM and MS-SSIM will experience BD-rate losses, which can be seen when training both pre and postprocessors alone. Only jointly training both pre- and postprocessors is able to achieve substantial BD-rate gains across all quality scores. Fig. 5 provides a generic example where the preprocessor embeds information into the source frame which does not significantly increase the bitrate and can survive the codec compression. This information is used by the postprocessor for reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.894, + 0.902 + ], + "angle": 0, + "content": "End-to-end training: For end-to-end training, an ablation is performed over the proposed iterative updates of the codec model and stop-gradient during training. The results are provided in Tab. 4 under different training con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17749" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.09, + 0.386, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.091, + 0.591, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.092, + 0.797, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.169, + 0.274, + 0.386, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.274, + 0.591, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.274, + 0.797, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.457, + 0.816, + 0.473 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of codecs with and without the proposed method. Zoom in for better viewing." + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.491, + 0.302, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.574, + 0.214, + 0.581 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.492, + 0.483, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.574, + 0.396, + 0.581 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.492, + 0.665, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.574, + 0.579, + 0.581 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.492, + 0.847, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.574, + 0.76, + 0.581 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.595, + 0.895, + 0.625 + ], + "angle": 0, + "content": "Figure 5. Results of components of the pipeline where (a), (b), (c) and (d) are the source frame, preprocessor output, codec output and postprocessor output, respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.648, + 0.441, + 0.663 + ], + "angle": 0, + "content": "Table 4. BD-rates with different training configurations." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.673, + 0.46, + 0.74 + ], + "angle": 0, + "content": "
PretrainIter. UpdateStop Grad.SSIMMSSSIMAVQTVMAF
✓✓✓-25.6-20.8-21.4-20.5
X✓✓-23.1-14.4-17.3-18.7
✓✓X16.40.0110.5-15.3
✓X✓-11.2-6.5-7.7-13.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.473, + 0.907 + ], + "angle": 0, + "content": "Figurations. Notably, when training the pre- and postprocessor with only iterative updates, there are BD-rate losses in SSIM and MS-SSIM. This is a by-product of the domain shift between the source and preprocessor output that increases as end-to-end training progresses. When performing end-to-end training with both iterative updates and stop gradient, there is a sufficient alignment constraint on the codec model, which leads to demonstrable BD-rate gains across quality scoring methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.648, + 0.634, + 0.664 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.895, + 0.901 + ], + "angle": 0, + "content": "This work proposes a new approach to jointly train neural pre and postprocessors to improve the rate-quality performance of conventional video CODECs. It is proposed to use a neural codec to model the standard codec's sophisticated rate-quality characteristics, combined with: (i) a novel approach to pretrain the codec model; (ii) a new method to train the neural wrapper model jointly with the codec model in the middle. Extensive evaluations with state-of-the-art AV1 and VVC CODECs show that the proposed neural wrapper approach provides for consistent BD-rate improvement over standard CODECs. Unlike neural methods for preprocessing and encoding, the proposed method allows for consistent improvement over all quality scores (SSIM, MSSSIM, AVQT, VMAF, P.910 MOS) over the entire rate-quality region of modern standards." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17750" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.094, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] On vmaf's property in the presence of image enhancement operations. https : // docs . google . com / document / d / 1dJczEhX00MZjBSNyKmd3ARiCTdFVMNPBykH4 _ HMPoyY/edit#heading=h.oaikhnw46pw5. Last accessed: 2020-11-16. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.472, + 0.284 + ], + "angle": 0, + "content": "[2] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.286, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[3] Stéphane Alarie, Charles Audet, Aïmen E Gheribi, Michael Kokkolaras, and Sébastien Le Digabel. Two decades of blackbox optimization applications. EURO Journal on Computational Optimization, 9:100011, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.357, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[4] Apple. \"Apple avqt presentation\", 2021. https://developer.apple.com/videos/play/wwdc2021/10145.1,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.469, + 0.513 + ], + "angle": 0, + "content": "[6] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6228-6237, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.516, + 0.469, + 0.57 + ], + "angle": 0, + "content": "[7] Mohammad Reza Bonyadi and Zbigniew Michalewicz. Particle swarm optimization for single objective continuous space problems: a review. Evolutionary computation, 25(1):1-54, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.573, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[8] Frank Bossen et al. Common test conditions and software reference configurations. JCTVC-L1100, 12(7):1, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[9] Eirina Bourtsovlatze, Aaron Chadha, Ilya Fadeev, Vasileios Giotasas, and Yiannis Andreopoulos. Deep video precoding. IEEE Transactions on Circuits and Systems for Video Technology, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[11] Aaron Chadha and Yiannis Andreopoulos. Deep perceptual preprocessing for video coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14852-14861, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[12] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[14] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5962-5971, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[15] Jinyoung Choi and Bohyung Han. Task-aware quantization network forJPEG image compression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 309-324. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[16] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. In Proceedings of the AAAI conference on artificial intelligence, pages 10696-10703, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.335, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[17] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3476-3485. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[18] Ryosuke Furuta, Naoto Inoue, and Toshihiko Yamasaki. Pixelrl: Fully convolutional network with reinforcement learning for image processing. IEEE Transactions on Multimedia, 22(7):1704-1719, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[19] Vasileios Giotsas, Nikos Deligiannis, Pam Fisher, and Yiannis Andreopoulos. Perceptual video quality estimation by regression with myopic experts. In 2015 Seventh International Workshop on Quality of Multimedia Experience (QoMEX), pages 1-6. IEEE, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[20] Onur G Guleryuz, Philip A Chou, Hugues Hoppe, Danhang Tang, Ruofei Du, Philip Davidson, and Sean Fanello. Sandwiched image compression: wrapping neural networks around a standard codec. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3757-3761. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[21] Onur G Guleryuz, Philip A Chou, Hugues Hoppe, Danhang Tang, Ruofei Du, Philip Davidson, and Sean Fanello. Sandwiched image compression: Increasing the resolution and dynamic range of standard codec. In 2022 Picture Coding Symposium (PCS), pages 175-179. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[22] Mengxi Guo, Shijie Zhao, Hao Jiang, Junlin Li, and Li Zhang. Video compression with arbitrary rescaling network. arXiv preprint arXiv:2306.04202, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[23] Yung-Han Ho, Chih-Peng Chang, Peng-Yu Chen, Alessandro Gnutti, and Wen-Hsiao Peng. Canf-vc: Conditional augmented normalizing flows for video compression. In European Conference on Computer Vision, pages 207-223. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Yueyu Hu, Chenhao Zhang, Onur G Guleryuz, Debargha Mukherjee, and Yao Wang. Standard compliant video coding using low complexity, switchable neural wrappers. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1922-1928. IEEE, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Zhihao Hu and Dong Xu. Complexity-guided slimmable decoder for efficient deep video compression. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17751" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14358-14367, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.472, + 0.178 + ], + "angle": 0, + "content": "[26] Quan Huynh-Thu and Mohammed Ghanbari. The accuracy of psnr in predicting video quality for different video scenes and frame rates. Telecommunication Systems, 49: 35-48, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[27] Berivan Isik, Onur G Guleryuz, Danhang Tang, Jonathan Taylor, and Philip A Chou. Sandwiched video compression: Efficiently extending the reach of standard codecswith neural wrappers. arXiv preprint arXiv:2303.11473, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.471, + 0.333 + ], + "angle": 0, + "content": "[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 645-660. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[29] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8008-8017, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.471, + 0.502 + ], + "angle": 0, + "content": "[30] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3224-3232, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.471, + 0.574 + ], + "angle": 0, + "content": "[31] Jan P Klopp, Keng-Chi Liu, Liang-Gee Chen, and Shao-Yi Chien. How to exploit the transferability of learned image compression to conventional codescs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16165-16174, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.471, + 0.672 + ], + "angle": 0, + "content": "[32] Faouzi Kossentini, Hassen Guermazi, Nader Mahdi, Chekib Nouira, Amir Naghdinezhad, Hassene Tmar, Omar Khelif, Phoenix Worth, and Foued Ben Amara. The svt-av1 encoder: overview, features and speed-quality tradeoffs. In Applications of Digital Image Processing XLIII, page 1151021. International Society for Optics and Photonics, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.471, + 0.729 + ], + "angle": 0, + "content": "[33] Théo Ladune, Pierrick Philippe, Wassim Hamidouche, Lu Zhang, and Olivier Déforges. Conditional coding for flexible learned video compression. arXiv preprint arXiv:2104.07930, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[34] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[35] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[36] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with feature modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26099-26108, 2024. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[37] Songnan Li, Fan Zhang, Lin Ma, and King Ngi Ngan. Image quality assessment by separately evaluating detail losses and additive impairments. IEEE Transactions on Multimedia, 13(5):935-949, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[38] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. Mucan: Multi-correspondence aggregation network for video super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 335-351. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[39] Yinxiao Li, Pengchong Jin, Feng Yang, Ce Liu, Ming-Hsuan Yang, and Peyman Milanfar. Comisr: Compression-informed video super-resolution. in 2021 IEEE. In CVF International Conference on Computer Vision (ICCV), pages 2523-2532, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.309, + 0.892, + 0.351 + ], + "angle": 0, + "content": "[40] Zhi Li and Christos G Bampis. Recover subjective quality scores from noisy measurements. In 2017 Data compression conference (DCC), pages 52-61. IEEE, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.353, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[41] Zhi Li, Anne Aaron, Ioannis Katsavounidis, Anush Moorthy, and Megha Manohara. Toward a practical perceptual video quality metric. The Netflix Tech Blog, 6, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.398, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[42] Zhi Li, Christos Bampis, Julie Novak, Anne Aaron, Kyle Swanson, Anush Moorthy, and J Cock. Vmaf: The journey continues. *Netflix Technology Blog*, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[43] Zhi Li, Christos G Bampis, Lukáš Krasula, Lucjan Janowski, and Ioannis Katsavounidis. A simple model for subject behavior in subjective experiments. arXiv preprint arXiv:2004.02067, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.5, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[44] Jianping Lin, Dong Liu, Houqiang Li, and Feng Wu. MLvc: Multiple frames prediction for learned video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3546-3554, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[45] Jerry Liu, Shenlong Wang, Wei-Chiu Ma, Meet Shah, Rui Hu, Pranaab Dhawan, and Raquel Urtasun. Conditional entropy coding for efficient video compression. In European Conference on Computer Vision, pages 453-468. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[46] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. Dvc: An end-to-end deep video compression framework. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[47] Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, and Eirikur Agustsson. Vct: A video compression transformer. arXiv preprint arXiv:2206.07307, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[48] Alexandre Mercat, Marko Viitanen, and Jarno Vanne. Uvg dataset: 50/120fps 4k sequences for video codec analysis and development. In Proceedings of the 11th ACM Multimedia Systems Conference, pages 297-302, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[49] Reza Pourreza, Hoang Le, Amir Said, Guillaume Sautiere, and Auke Wiggers. Boosting neural video codecs by exploiting hierarchical redundancy. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5355-5364, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17752" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[50] Zhongwei Qiu, Huan Yang, Jianlong Fu, and Dongmei Fu. Learning spatiotemporal frequency-transformer for compressed video super-resolution. In European Conference on Computer Vision, pages 257-273. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[51] Alexander Raake, Silvio Borer, Shahid M Satti, Jörgen Gustafsson, Rakesh Rao Ramachandra Rao, Stefano Medagli, Peter List, Steve Göring, David Lindero, Werner Robitza, et al. Multi-model standard for bitstream-, pixel-based and hybrid video quality assessment of uhd/4k: Itu-t p. 1204. IEEE Access, 8:193020-193049, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.471, + 0.292 + ], + "angle": 0, + "content": "[52] Reza Rassool. Vmaf reproducibility: Validating a perceptual practical video quality metric. In 2017 IEEE international symposium on broadband multimedia systems and broadcasting (BMSB), pages 1-2. IEEE, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.47, + 0.364 + ], + "angle": 0, + "content": "[53] Oren Rippel, Alexander G Anderson, Kedar Tatwawadi, Sanjay Nair, Craig Lytle, and Lubomir Bourdev. Elf-vc: Efficient learned flexible-rate video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14479-14488, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.366, + 0.47, + 0.422 + ], + "angle": 0, + "content": "[54] Amir Said, Manish Kumar Singh, and Reza Pourreza. Differentiable bit-rate estimation for neural-based video codec enhancement. In 2022 Picture Coding Symposium (PCS), pages 379-383. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.424, + 0.47, + 0.48 + ], + "angle": 0, + "content": "[55] Mehdi SM Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6626-6634, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.47, + 0.524 + ], + "angle": 0, + "content": "[56] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.47, + 0.582 + ], + "angle": 0, + "content": "[57] Hanbin Son, Taeoh Kim, Hyeongmin Lee, and Sangyoun Lee. Enhanced standard compatible image compression framework based on auxiliary codec networks. IEEE Transactions on Image Processing, 31:664-677, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.47, + 0.641 + ], + "angle": 0, + "content": "[58] Yannick Strümpler, Ren Yang, and Radu Timofte. Learning to improve image compression without changing the standard decoder. In European Conference on Computer Vision, pages 200–216. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.47, + 0.698 + ], + "angle": 0, + "content": "[59] Hossein Talebi and Peyman Milanfar. Learned perceptual image enhancement. In 2018 IEEE International Conference on Computational Photography (ICCP), pages 1-13. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[60] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In Proceedings of the IEEE international conference on computer vision, pages 4472-4480, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[61] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu Tdan. temporally-deformable alignment network for video super-resolution. in 2020 ieee. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3357-3366, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[62] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[64] Ties van Rozendaal, Tushar Singhal, Hoang Le, Guillaume Sautiere, Amir Said, Krishna Buska, Anjuman Raha, Dimitris Kalatzis, Hitarth Mehta, Frank Mayer, et al. Mobilenvc: Real-time 1080p neural video compression on a mobile device. arXiv preprint arXiv:2310.01258, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[65] Abhinau K Venkataramanan, Chengyang Wu, Alan C Bovik, Ioannis Katsavounidis, and Zafar Shahid. A hitchhiker's guide to structural similarity. IEEE Access, 9: 28872-28896, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.895, + 0.329 + ], + "angle": 0, + "content": "[66] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[67] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Huchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2012-2021, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[68] Ping-Hao Wu, Ioannis Katsavounidis, Zhijun Lei, David Ronca, Hassene Tmar, Omran Abdelkafi, Colton Cheung, Foued Ben Amara, and Faouzi Kossentini. Towards much better svt-av1 quality-cycles tradeoffs for vod applications. In Applications of Digital Image Processing XLIV, pages 236–256. SPIE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[69] Li Xu, Gang He, Jinjia Zhou, Jie Lei, Weiying Xie, Yun-song Li, and Yu-Wing Tai. Transcoded video restoration by temporal spatial auxiliary network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2875-2883, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[70] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8):1106-1125, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[71] Ren Yang, Mai Xu, and Zulin Wang. Decoder-side hevc quality enhancement with scalable convolutional neural network. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 817-822. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[72] Ren Yang, Mai Xu, Tie Liu, Zulin Wang, and Zhenyu Guan. Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology, 29(7):2039-2054, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[73] Ren Yang, Mai Xu, Zulin Wang, and Tianyi Li. Multi-frame quality enhancement for compressed video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6664-6673, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[74] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3106-3115, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[75] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video superresolution. In Proceedings of the IEEE/CVF International" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17753" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision, pages 4429-4438, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[76] Hengsheng Zhang, Xueyi Zou, Jiaming Guo, Youliang Yan, Rong Xie, and Li Song. A codec information assisted framework for efficient compressed video superresolution. In European Conference on Computer Vision, pages 220-235. Springer, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.47, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17754" + } + ] +] \ No newline at end of file diff --git a/2025/Perceptual Video Compression with Neural Wrapping/4f847378-cfb2-4455-91dd-a70a0227cf83_origin.pdf b/2025/Perceptual Video Compression with Neural Wrapping/4f847378-cfb2-4455-91dd-a70a0227cf83_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c1c5ca33f7c94b313d98e48570f1a042d395c6af --- /dev/null +++ b/2025/Perceptual Video Compression with Neural Wrapping/4f847378-cfb2-4455-91dd-a70a0227cf83_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f516771f788d9fb6155c901b827aea9602256d93e6df447430adf737a59f1be4 +size 7424935 diff --git a/2025/Perceptual Video Compression with Neural Wrapping/full.md b/2025/Perceptual Video Compression with Neural Wrapping/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6d69ee990d36c6be3d79858845ede2343ff70a76 --- /dev/null +++ b/2025/Perceptual Video Compression with Neural Wrapping/full.md @@ -0,0 +1,394 @@ +# Perceptual Video Compression with Neural Wrapping + +# Muhammad Umar Karim Khan, Aaron Chadha, Mohammad Ashraful Anam, Yiannis Andreopoulos, Sony Interactive Entertainment + +{Umar . Khan, Aaron . Chadha, Russell . Anam, Yiannis . Andreopoulos}@sony.com + +# Abstract + +Standard video codecs are rate-distortion optimization machines, where distortion is typically quantified using PSNR versus the source. However, it is now widely accepted that increasing PSNR does not necessarily translate to better visual quality. In this paper, a better balance between perception and fidelity is targeted, in order to provide for significant rate savings over state-of-the-art standards-based video codecs. Specifically, pre- and post-processing neural networks are proposed that enhance the coding efficiency of standard video codecs when benchmarked with an array of well-established perceptual quality scores. These "neural wrapper" elements are end-to-end trained with a neural codec module serving as a differentiable proxy for standard video codecs. The codec proxy is jointly optimized with the pre- and post components via a novel two-phase pretraining strategy and end-to-end iterative refinement with stop-gradient. This allows the neural pre- and postprocessor to learn to embed, remove and recover information in a codec-aware manner, thus improving its rate-quality performance. A single neural-wrapper model is thereby established and used for the entire rate-quality curve without needing any downscaling or upscaling. The trained model is tested with the AV1 and VVC standard codecs via an array of well-established objective quality scores (SSIM, MS-SSIM, VMAF, AVQT), as well as mean opinion scores (MOS) derived from ITU-T P.910 subjective testing. Experimental results show that the proposed approach improves all quality scores, with $-18.5\%$ average Bjontegaard Delta-rate (BD-rate) saving over all objective scores and MOS improvement over both standard codecs. This illustrates the significant potential of neural wrapper components over standards-based video coding. + +# 1. Introduction + +It is now widely accepted that that signal-to-noise (SNR) ratio is a poor indicator of visual quality in video coding [6, 19, 26, 42]. Instead, quality scores that include elements of human perception [42], perceptual modelling of encoding artifacts [51], as well as viewing setup awareness [4, 65] have emerged as strong contenders for the best means to + +objectively characterize visual quality. This has also led to the research community now moving away from SNR-optimization [26] in favor of structural similarity (SSIM [65]), video multimethod assessment fusion (VMAF) and Apple's advanced video quality tool (AVQT) optimization [4, 42]. However, all current perceptual optimization approaches in standards-based video encoders have one or more of the following detriments: + +- they require multiple encoding passes or in-loop implementation within a specific encoder; +- they only optimize for a single quality scoring method like VMAF or SSIM and are shown to be detrimental in other quality scores; +- they comprise hand-crafted (shallow) models of low-level human perception and fail to encapsulate several characteristics of more advanced quality scoring methods like VMAF or AVQT in a data-driven and learnable manner; +their Bjontegaard Delta-rate (BD-rate) improvement on well-established quality scores like SSIM, VMAF and AVQT is in the order of a few percentile points. + +A neural codec "wrapper" is proposed in this paper as the means to augment standard video codecs and optimize them for perceptual quality. Including neural pre- and postprocessing in conventional encoding pipelines is shown to clearly outperform the equivalent codecs without these components. This allows for conventional streaming ecosystems to continue operating "as-is" and does not break standards. That is, when neural components cannot be supported before encoding or after decoding, the system can just disable them with appropriate adaptation, without jeopardizing video streaming. The key contributions of the paper are listed as follows. + +- A novel pretraining regime for modelling a conventional codec with a neural codec model is proposed, which uses the implicit encoder-decoder structure of the neural codec to align both video quality and rate. +- An end-to-end training regime for jointly training pre and postprocessors is proposed. Starting with a pretrained neural codec model, it is proposed to iteratively update the codec model. This approach allows for more accurate gradients to be backpropagated to the prepro + +![](images/d2259fbf37c5d80f7edf44b7b7d2f9ae61208435af9ec137f09cd4559ece28bb.jpg) +Figure 1. Left: Functional block diagram of the multi-stage training data flow. Right: Components of codec model $M$ that are used as a proxy for the target codec $C$ . $AE/AD$ denotes arithmetic encoding followed by arithmetic decoding and $Q$ denotes quantization. See Sec. 3 for details. + +cessor. + +- The trained pipeline is evaluated on gaming sequences and significant BD-rate gains versus AV1 and VVC encoding are demonstrated with just a single pre- and postprocessor model. In addition, the proposed neural wrapper is shown to outperform a state-of-the-art neural encoder [35], as well as the state-of-the-art in deep perceptual preprocessing proposed recently [11]. + +A functional block diagram for the proposed system is shown in Fig. 1. Example visuals produced by the proposed method are shown in Fig. 4, demonstrating that the proposed neural wrapper framework allows for better detail preservation than the underlying standard codec. + +# 2. Related Work + +# 2.1. Neural Video Codes + +Neural video codecs optimize neural networks for rate and distortion by exploiting redundancies in video data. Most notable are residual coding methods; Lu et al. [46] introduce deep video codec, which adopts neural networks within the residual coding pipeline of a traditional codec. Optical flow is employed as a proxy for motion vectors in order to warp the reference frame and generate a residual frame for encoding. Numerous other methods have been proposed based on residual coding [44, 49, 53]. Scale-space flow estimation is proposed as an improvement over traditional optical flow by Agustsson et al. [2], namely for better handling of complex motion in the scene. Beyond residual coding methods, conditional coding methods learn feature contexts implicitly from the encoded context [23, 33, 34, 45, 47, 56]. Li et al. [35] propose using diverse contexts in spatial and temporal directions to perform video compression and outperform standard video codecs when the latter are used in low-delay mode. They improve the performance and bitrate range in [36]. Recent work has additionally focused on + +reducing the complexity of neural CODECs. For example, Van Rozendaal et al. [64] deploy an architecture similar to Agustsson et al. [2] on a smart phone and achieve real time performance at 1080p. Hu et al. [25] adopt slimmable networks for variable complexity. + +# 2.2. Preprocessors + +While postprocessing is a more mature line of research, there has been a prevalence of work in preprocessing, both on image and video modalities. In the image domain, Strumpler et al. [58] propose to modify the JPEG encoder with an attention network and a learnable quantization table, in order to modulate the source frame and provide for additional rate savings over the JPEG encoder alone. Conversely, Talebi et al. [59] avoid adapting the JPEG encoder by prepended the standard encoder with a preprocessor for deployment with a standard JPEG encoder and decoder. Klopp et al. [31] train a preprocessor with a surrogate neural codec model, which is trained to model a target image codec. In the video domain, Boutssoulatze et al. [9] propose a precoding network for downscaling with a online mode selection algorithm, in order to alter the precoding based on scene characteristics. Conversely, Chadha et al. [11] propose an entirely offline training mechanism by first modelling the core components of a video codec in a differentiable manner. A preprocessor is optimized on the codec model, in order to provide further rate-quality savings over the target codec. + +# 2.3. Postprocessors + +Numerous deep learning methods have been proposed for compressed video denoising, video super-resolution or both tasks. Single frame methods for denoising compressed videos have been proposed in [71, 72]. Multiframe methods have provided superior gains in metrics compared to single frame methods. MFQE [73] detects frames with higher quality and uses it as a reference to + +improve neighboring frames. TSAN [69] uses a temporal deformable alignment and multi-scale fusion for the task of video restoration. STDF [16] uses spatio-temporal deformable convolution to aggregate temporal context rather than using optical flow for frame warping. A common approach for video super-resolution is to align neighboring frames and transfer information across frames [60, 61, 66, 70]. On the contrary, other methods choose non-local modules or 3D convolutions to gather temporal context [29, 30, 38, 74]. Some methods recurrently propagate history in the forward direction [17, 28, 55]. Recent methods such as BasicVSR [12], GOVSR [75] and BasicVSR++ [13] propagate history in forward and backward directions, such that both past and future frames are used to super-resolve the current frame. Architectures developed for video super-resolution can easily be adapted for video denoising. Few authors have focused on the task of jointly denoising and upscaling videos. FTVSR [50] uses a frequency transformer which performs self-attention in space, time and frequency to recover details. RealBasicVSR [14] considers sensor noise, motion blur and compression artifacts with super-resolution. CIAF [76] uses motion-vectors from the video to approximate optical flow. CAVSR [67] adapts the super-resolution to the compression level of the input video. COMISR [39] uses an architecture based on BasicVSR [12] to perform joint denoising and upscaling. + +# 2.4. Pre- and Postprocessors with Traditional Codes + +Recent methods have also explored joint training of pre and postprocessors as neural wrappers. In the image domain, recent work [20, 21] emulates the core blocks of the JPEG encoder and decoder, and uses it to train pre and postprocessors for rate-distortion savings. Similar sandwich approach for videos has been proposed in [24]. Son et al. [57] propose distinct neural networks to model the distortion and rate behavior of a standard codec, which is thus sandwiched by pre- and postprocessor networks for rate-distortion optimization. However, these methods have no target codec awareness beyond the implicit intra/inter coding structure. Either the codec proxy is not learned (e.g. derived from differentiable JPEG/DCT as in [24]), only partially learned via rate-distortion optimization (e.g. the loop filter as in [27]) or only operates in intra mode (as in [57]). In this paper, the proposed framework is directly conditioned on the target conventional codec. Tian et al. [62] target additional rate savings with learnable rescaling. However, this method has no rate control or alignment outside of the scaling factor. Conversely, the framework proposed in this paper is jointly conditioned on both the target codec rate and distortion and is thus able to offer rate savings even when encoding at full res + +olution. Additionally, whereas recent work restricts evaluation to a limited range of the rate-distortion space, at high bitrates [61] or substantially lower bitrates and resolutions [22] and on older encoding standards (AVC or HEVC), the proposed method is evaluated against state-of-the-art AV1 and VVC encoding, i.e.: $(i)$ the entire rate-quality region of practical significance is considered (VMAF values between 40-96 and bitrates from 100kbps to 15mbps); $(ii)$ slow encoding presets are used (preset=4 for SVT-AV1 and preset=slow for VVenC), without limiting the encoders to low-delay mode. + +# 3. Methodology + +The top-level functional block diagram of the proposed training flow is shown in Fig. 1 and includes a preprocessor $P$ , codec model $M$ , target codec $C$ and postprocessor $O$ . Notably, the target codec is a traditional video codec, such as SVT-AV1 [32]. First, a source frame $I_{\mathrm{in}}$ is processed by $P$ to generate a corresponding preprocessed frame $I_P$ . The preprocessed frame is simultaneously encoded by $M$ and $C$ , which generate encoded frames $I_M$ and $I_C$ respectively. Finally, encoded frames are postprocessed with $O$ to generate the output frame $I_{\mathrm{out}}$ . End-to-end distortion $D$ is computed between $I_{\mathrm{in}}$ and $I_{\mathrm{out}}$ and the overall training loss is the weighted sum of $D$ and the bitrate $R_M$ . Notably, at inference, the codec model $M$ is removed from the pipeline and only the target codec $C$ is deployed for encoding. + +# 3.1. Codec Model Pretraining + +Given the target codec $C$ in Fig. 1 is not differentiable, a mechanism for adapting the proposed pipeline to the target codec behavior is required. One option is black-box or derivative-free optimization methods [3, 7, 18]. However these methods typically rely on an approximation of the unknown function to be optimized, and are computationally infeasible for training at scale. Another option is to replicate the core prediction and transform blocks of the target codec in a differentiable manner [15, 27, 54]. Such approaches can model older codecs such as JPEG, or faster presets of AVC or HEVC successfully to an extent but fail to accurately model the operation of more sophisticated codecs such as AV1 or VVC. + +Recent work [31, 57] has instead leveraged on neural image and video CODEs for predicting the target codec behavior. In this work, a similar approach is followed, and a neural video codec $M$ is adapted to model the ratedistortion behavior of a target codec $C$ . Specifically, the scale-space flow architecture [2] is used to generate lossy frames, and the implicit encoder-decoder structure and associated loss functions are leveraged to align to the target codec behavior in a spatially localized manner. As per Agustsson et al. [2], for each of intra, flow and residual + +encoding, the hyperprior based architecture of Balle et al. [5] is used, without any autoregressive components. + +Notably, the hyperprior architecture comprises analysis and synthesis transforms, $g_{a}$ and $g_{s}$ , for encoding to and decoding from the quantized latent space $\tilde{y}$ respectively. A distortion loss $D(:,g_a,g_s)$ can thus be computed between source frame $I_{in}$ (input to $g_{a}$ ) and reconstructed frame $I_M$ (output by $g_{s}$ ). A hyper-analysis and hypersynthesis transform $(h_a$ and $h_s$ respectively) constitute the hyperprior, which is used to model scale dependencies $\tilde{\sigma}$ within the latent space. The latent space $\tilde{y}$ is thus modelled as a Gaussian scale mixture on which the rate $R(:,h_a,h_s)$ can be optimized (as Shannon entropy). Traditionally, the architecture would then be trained with an aggregate loss function $L = D + \lambda R$ , where $\lambda$ represents a Lagrangian parameter for balancing between the rate and distortion components. + +When performing neural codec alignment to a target codec, there are two key issues: (i) the model encoded frame must align to the target encoded frame (and not the source); (ii) selecting the best value for $\lambda$ when trying to replicate target codec behavior is a non-trivial problem. Crucially, it is noted that the rate model is exclusively parameterized by $h_a$ and $h_s$ , and distortion exclusively by $g_a$ and $g_s$ . Subsequently, a novel two phase training methodology is proposed for aligning both rate and distortion to the target codec: + +Phase 1: The first issue of target codec alignment is addressed via the distortion loss. To this end, the reference for the distortion loss is changed from the source frame $I_{in}$ to the target codec frame $I_C$ and the mean squared error is optimized: + +$$ +D _ {M} ^ {\mathrm {h y p}} \left(I _ {M}, I _ {C}; \theta_ {g}\right) = \mathbb {E} _ {I _ {C}} \left[ \left\| I _ {M} - I _ {C} \right\| ^ {2} \right] \tag {1} +$$ + +where $\theta_{g}$ represent the analysis and synthesis transform parameters. + +Phase 2: Following alignment of $g_{a}$ and $g_{s}$ in Phase 1, the latent space $\tilde{y}$ is now conditioned on the target codec $C$ . To circumvent selection of an optimal $\lambda$ for rate, the implicit encoder-decoder structural prior of the aligned neural codec is leveraged in order to learn an unsupervised rate alignment to the target codec. Specifically, the maximum likelihood on the hyperprior model parameters is optimized, in order to learn the latent space distribution under target codec distortion alignment. The rate is thus optimized as: + +$$ +R _ {M} ^ {\mathrm {h y p}} \left(I _ {M}; \theta_ {h}\right) = - \mathbb {E} _ {\tilde {y}, \tilde {z}} \left[ \log p _ {\tilde {y} | \tilde {z}, I _ {C}} \left(\tilde {y} | \tilde {z}, I _ {C}\right) + \log p _ {\tilde {z} | I _ {C}} \left(\tilde {z} | I _ {C}\right) \right] \tag {2} +$$ + +where $\theta_h$ represents the hyperprior model parameters and $\tilde{z}$ is the quantized hyperprior latent space (which would also be encoded and transmitted as side information in a neural codec pipeline). For intra coding, the intra + +hyperprior model is optimized on intra encoded source frames using the two-phase strategy. For inter coding, the residual and flow hyperprior models are optimized jointly by summing the rate and distortion components in each phase. + +The proposed two-phase pretraining ensures that $R_{M}^{\mathrm{hyp}}$ is close to the target codec $R_{C}$ by performing an affine transform on $R_{M}^{\mathrm{hyp}}$ , i.e., $R_{M} = \alpha R_{M}^{\mathrm{hyp}} + \beta$ , where $\alpha$ and $\beta$ are scalar learnable parameters optimized with MSE: + +$$ +L _ {M} ^ {(\alpha , \beta)} \left(R _ {M}, R _ {C}; \alpha , \beta\right) = \mathbb {E} _ {R _ {C}} \left[ \left\| R _ {M} - R _ {C} \right\| ^ {2} \right]. \tag {3} +$$ + +The similarity of $R_{M}^{\mathrm{hyp}}$ and $R_{C}$ is required to properly balance the rate and distortion losses in end-to-end training. The aggregate rate and distortion for a sequence can thus be measured by computing the sum over all frames. + +# 3.2. End-to-End Training + +With the pretrained neural codec model, an end-to-end trainable pipeline with both the pre and postprocessor is developed, as illustrated in Fig. 1. In this case, the training loss is very similar to conventional neural CODECs, but is now exclusively used to train the pre and postprocessors, i.e.: + +$$ +L \left(I _ {\text {o u t}}, I _ {\text {i n}}; \theta_ {P}, \theta_ {O}\right) = \mathbb {E} _ {I _ {\text {i n}}} \left[ D \left(I _ {\text {i n}}, I _ {\text {o u t}}\right) + \lambda R _ {M} \right] \tag {4} +$$ + +where $\theta_{P}$ and $\theta_{O}$ are the pre and postprocessor parameters respectively, $R_{M}$ is the codec model rate estimation (proportional to $R_{M}^{\mathrm{flow}} + R_{M}^{\mathrm{res}}$ for inter coding the flow and residual from motion compensation, and $R_{M}^{\mathrm{intra}}$ for intra coding) and $D$ is the distortion loss component. Following subjective assessment tests, a weighted sum of: mean absolute error (MAE), SSIM, MS-SSIM and detail loss metric (DLM) * [37] is proposed for the distortion component $D$ in Eq. (4). Specifically, $D(I_{\mathrm{in}},I_{\mathrm{out}}) = \lambda_{\mathrm{L1}}||I_{\mathrm{in}} - I_{\mathrm{out}}|| + \lambda_{\mathrm{SSIM}}(1 - D_{\mathrm{SSIM}}(I_{\mathrm{in}},I_{\mathrm{out}})) + \lambda_{\mathrm{MSSSIM}}(1 - D_{\mathrm{MSSSIM}}(I_{\mathrm{in}},I_{\mathrm{out}})) + \lambda_{\mathrm{DLM}}D_{\mathrm{DLM}}(I_{\mathrm{in}},I_{\mathrm{out}})$ . + +Concept drift or domain shift for the codec model is an expected consequence of end-to-end training. The codec model is pretrained with natural videos whereas the preprocessor is continuously updated as training proceeds; thus, the data observed by the codec model has a non-stationary distribution. Since neural networks generally do not perform well on out-of-distribution data, an iterative refinement mechanism is proposed, where the codec model parameters $\theta_{M}$ are independently updated after each iteration of end-to-end training using the two-phase approach described previously. Notably, in this case the input to the target codec is now the output of the + +preprocessor $I_{P}$ and not the source frame, and $I_{C}$ and $R_{C}$ represent the frame and rate output by the target codec. + +Additionally, the gradients backpropagated to the preprocessor can be better aligned towards the target codec via the stop-gradient operator. First, it is assumed for the remaining discussion in the section that the tensors representing the frames are flattened, i.e., $\{I_P,I_M,I_C\} \in \mathbb{R}^{N\times 1}$ , where $N$ is the number of pixels in the tensor. If the target codec $C$ was differentiable then the accuracy of the gradients backpropagated will increase by decreasing: + +$$ +G = \left| J _ {M} \left(I _ {P}\right) \nabla_ {I _ {M}} L - J _ {C} \left(I _ {P}\right) \nabla_ {I _ {C}} L \right|, \tag {5} +$$ + +where $J_{M}(I_{P})\in \mathbb{R}^{N\times N}$ is the transposed Jacobian of $M$ w.r.t. $I_P$ and $\nabla_{I_M}L$ is the gradient of the objective w.r.t. $I_M$ . Assuming that the frames $I_C$ generated by the target codec $C$ are input to the postprocessor $O$ then Eq. (5) becomes + +$$ +G = \left| \left(J _ {M} \left(I _ {P}\right) - J _ {C} \left(I _ {P}\right)\right) \nabla_ {I _ {C}} L \right|, \tag {6} +$$ + +Since $\nabla_{I_C}^{(j)}L$ is necessary for the networks to train, $J_M^{(i,j)}(I_P) - J_C^{(i,j)}(I_P)$ should be close to zero. Thus, + +$$ +\forall i, j \in [ 1, N ], \left| \frac {\partial I _ {M} ^ {(i)}}{\partial I _ {P} ^ {(j)}} - \frac {\partial I _ {C} ^ {(i)}}{\partial I _ {P} ^ {(j)}} \right| = \left| \frac {\partial}{\partial I _ {P} ^ {(j)}} \left(I _ {M} ^ {(i)} - I _ {C} ^ {(i)}\right) \right| \tag {7} +$$ + +should be minimized. The above analysis can be extended to the rate component. Eq. (7) shows that accurate gradients can be backpropagated to the preprocessor if the codec model's output is close to the target codec's output and the input to the postprocessor is from the target video codec. In the proposed training regime, the input to the postprocessor is therefore set as + +$$ +I _ {O _ {\text {i n}}} = \overline {{\left(I _ {C} - I _ {M}\right)}} + I _ {M}, \tag {8} +$$ + +where $\overline{a}$ indicates the stop-gradient operator [63] applied to the tensor $a$ and $I_{C}$ is the output of the non-differentiable video codec. Eq. (8) allows frames generated by the target codec to be passed to the postprocessor while using the codec model for backpropagation. + +In summary, the proposed training procedure has two main steps; iterative update of the codec model and the end-to-end update of the pre- and postprocessors. For the end-to-end update, a stop-gradient-based approach is proposed. Simple CNNs are used for pre and postprocessing. Details of the pre and postprocessor architecture are in the supplementary. The end-to-end training pipeline is shown in Algorithm 1. + +# 4. Experimental Results + +# 4.1. Training + +The Vimeo-90k dataset [70] was used to train the codec model and perform end-to-end training. The training + +Algorithm 1 End-to-end training pipeline of the pre and postprocessor + +Require: $I_{\mathrm{in}}$ , batch of video frames + +Require: $\theta_{M}$ , parameters of the codec model + +Require: $\theta_{P}$ , parameters of the preprocessor + +Require: $\theta_{O}$ , parameters of the postprocessor + +Require: $\alpha_{L}$ , learning rate for training + +for all $I_{\mathrm{in}}$ do + +$$ +\boldsymbol {I} _ {P} \leftarrow P (\boldsymbol {I} _ {\text {i n}}) +$$ + +$$ +\left(\boldsymbol {I} _ {C}, \boldsymbol {R} _ {C}\right) \leftarrow C \left(\boldsymbol {I} _ {P}\right) +$$ + +$$ +\left(\boldsymbol {I} _ {M}, \boldsymbol {R} _ {M}\right) \leftarrow M \left(\boldsymbol {I} _ {P}\right) +$$ + +Get $D_M^{hyp}$ and $\theta_g\gets \theta_g - \alpha_L\nabla_{\theta_g}D_M^{hyp}$ + +Get $R_M^{hyp}$ and $\theta_h\gets \theta_h - \alpha_L\nabla_{\theta_h}R_M^{hyp}$ + +Get $L_{M}^{(\alpha ,\beta)}$ and $(\alpha ,\beta)\gets (\alpha ,\beta) - \alpha_{L}\nabla_{(\alpha ,\beta)}L_{M}^{(\alpha ,\beta)}$ + +Generate $I_{O_{\mathrm{in}}}$ using Eq. (8) + +$\pmb{I}_{\mathrm{out}}\gets O(\pmb{I}_{O_{\mathrm{in}}})$ + +Get $L$ using Eq. (4) + +$$ +\theta_ {P} \leftarrow \theta_ {P} - \alpha_ {L} \nabla_ {\theta_ {P}} L +$$ + +$$ +\theta_ {O} \leftarrow \theta_ {O} - \alpha_ {L} \nabla_ {\theta_ {O}} L +$$ + +# end for + +pipeline and the pre- and postprocessor are implemented in the YUV colorspace. A fixed learning rate of $10^{-4}$ is used for each network with the rest of hyper-parameters set as: $\lambda = 0.1$ ; $\lambda_{\mathrm{L1}} = 0.1$ , $\lambda_{\mathrm{SSIM}} = 0.25$ ; $\lambda_{\mathrm{MSSSIM}} = 0.25$ ; $\lambda_{\mathrm{DLM}} = 0.2$ . Batch size, GOP and maximum iterations are set to 4, 4 and 500000, respectively. Both Phase 1 and 2 of pretraining are performed for 500000 iterations. The model is trained on source frames only in an open-loop configuration; i.e., no reconstructed frames are fed into the preprocessor as input. A low-delay recipe of SVT-AV1 (described in supplementary) is used for codec model alignment for pre-training and end-to-end training. The codec model encoding is aligned to the GOP of the target codec, such that intra and inter coding are also aligned. For experiments on VVC, it was found that codec model alignment to a low-delay recipe of VVenC provided slight BD-rate gains over SVT-AV1 alignment, albeit at high computational overhead during training. Therefore, the presented results are carried out with a single neural wrapper model (one encode) developed based on codec proxy model alignment to SVT-AV1. This is practically more feasible than [24], which requires 9 combinations of models (9 encodes) + convex hull optimization and is thus far more computationally demanding. + +# 4.2. Evaluation Protocol + +Four 1080p video datasets were used for evaluation: (i) 16 gaming sequences captured in a commercial setup (Gaming); (ii) The 16 XIPH natural sequences used by recent work [11] (XIPH); (iii) The 1080p sequences of the UVG [48] dataset (UVG); (iv) HEVC-B [8] dataset + +(HEVC-B). The proposed method was applied with SVT-AV1 [32] for AV1 and VVenC [10] for VVC. All experiments were performed with YUV 4:2:0 8-bit sequences. In all cases, the protocol of Li et al. [35] is followed and 96 frames are evaluated. However, unlike Li [35], both standard encoders are used without their low-delay limitation, in order to attain their best performance (see supplementary for details on sequences and encoding recipes). Ablation results are with the SVT-AV1 lowdelay recipe (see supplementary). Results in Tab. 3 and 4 are on the Gaming dataset. + +In order to cover a wide range of well-established objective quality scoring methods, the following scores are used: SSIM, MS-SSIM, VMAF [42, 52] and the AVQT score from Apple [4]. In order to showcase the generality of the proposed approach, the same pre- and postprocessor models are used for: (i) all the points on the RD curves, (ii) all quality scores and (iii) both standard codecs. Perceptual models are used for competing methods when available. Comparison is provided against a competitive video preprocessor (DPP [11]), SOTA neural codecs (DCVC-DC [35] and DCVC-FM [36]), and perceptual modes of standard codecs. The comparisons include results of end-to-end training with the codec model of Hu et al. [24] with the proposed network architectures and losses used to illustrate the effectiveness of the proposed codec model. For DCVC-DC and DCVC-FM, the best available pretrained models (publicly provided by the authors) are used. + +# 4.3. Objective Quality Assessment + +The quantitative results of the proposed method are shown in† Fig. 2, and Tab. 1 and Tab. 2. These results show that the proposed approach provides for BD-rates between $-5.4\%$ to $-30\%$ in comparison to the underlying standard encoders. BD-rate savings tend to be higher for quality scores that are more perceptually-oriented. Compared to competing methods, the proposed approach is shown to offer significantly more consistent BD-rate improvements over all quality scoring methods (ALL columns of Tab. 1 and Tab. 2). Importantly, as shown in Fig. 2, unlike DCVC-DC that is only covering a very limited bitrate-quality range at the very top end, the proposed approach outperforms all other methods on the entire "active" region of bitrate-quality, i.e., VMAF of 40-93 [32], which encapsulates the entirety of commercially-relevant bitrates for 1080p video streaming. Given that the neural wrapper is optimizing perceptually, BD-rates for PSNR tend to show regression on PSNR (see supplementary) but visual inspection and an array of perceptual metrics indicate that the proposed approach offers significant benefits + +![](images/65c01de4bcda9262aea18f8548bb9701ee71161eac47d04e9f811fff78595cb2.jpg) +(a) + +![](images/825fde789f275700e39dad52a643630271375a0f8adf42751298eb814e2a7bc2.jpg) +(b) + +![](images/43c24bb04a97d86178a09b4081705d4911f79f0bee25609b94c8a00cfcb76007.jpg) +(c) + +![](images/f8d5f43a991db678e7df6cf5bb05450de3f53f824f28b7d5d3882680939bad64.jpg) +(d) + +![](images/4f0e2c9981762f202e9c52ed2a4c25a5271e4cd7ffa32b0f54375e40c94ceeec.jpg) +(e) + +![](images/233c71f97da0b0a3afc2113377f60647550034c1e7e0e20ed986a0a6772857fd.jpg) +(f) + +![](images/96cce1064cfc1d05ecaf8517cad37f29915ae71b8eea34d9ad0113df58a1ebda.jpg) +(g) + +![](images/0f30839fc121e9fefc8419dcd74925d00fceb6b91eee5dba5dbe9d677692fc19.jpg) +(h) +Figure 2. Combined rate-quality plots over Gaming ((a), (b)), XIPH ((c), (d)), UVG ((e), (f)) and HEVC-B ((g), (h)) datasets. See supplementary for more results. + +in visual quality. Nonetheless, the results of DCVC-DC and DCVC-FM are remarkably good given that they are fully neural approaches. + +# 4.4. Subjective Quality Assessment + +To assess the visual quality of the proposed approach, a 5-scale absolute categorical rating (ACR) test with hidden reference was conducted under the ITU-T P.910 test conditions and SUREAL post-processing [40, 43] (see supplementary for details). Raters were shown videos generated by the codec and proposed method. Separate tests were conducted for AV1 and VVC with each test comprising 72 sequences. The combined rate-MOS plots are shown in Fig. 3. The BD-rates reported in the figure for the proposed approach are mostly in line with AVQT and VMAF BD-rates reported for Gaming in Tab. 1 and + +Table 1. BD-rates vs. AV1 (more negative is better). AV1+SSIM denotes codec tuned for SSIM (-tune ssim). ALL indicates average BD-rate over all quality scores. + +
DatasetMethodSSIMMSSSIMAVQTVMAFALL
GamingAV1+SSIM-4.5-5.4-5.12.1-3.2
AV1+DPP [11]-5.2-6.1-17.6-11.3-10.0
AV1+[24]14.96.312.2-1.48.0
DCVC-DC [35]-15.4-36.1-35.621.84-16.3
DCVC-FM [36]14.86.04.2-2.95.5
AV1+Proposed-15.5-13.8-19.6-30.0-19.7
XIPHAV1+SSIM-3.5-4.0-5.43.2-2.4
AV1+DPP [11]-4.3-3.9-9.0-16.7-8.5
AV1+[24]12.96.14.51.76.3
DCVC-DC [35]-18.4-25.512.032.20.3
DCVC-FM [36]17.69.08.23.59.6
AV1+Proposed-10.9-7.1-28.8-30.1-19.2
UVGAV1+SSIM-2.8-3.1-2.82.9-1.4
AV1+DPP [11]-3.8-3.0-9.2-21.7-9.4
AV1+[24]9.14.43.33.25.0
DCVC-DC [35]-20.8-10.24.218.2-2.1
DCVC-FM [36]20.49.69.89.412.3
AV1+Proposed-16.2-9.7-22.2-31.9-20.0
HEVC-BAV1+SSIM-5.4-5.3-4.72.9-3.1
AV1+DPP [11]-8.7-6.9-5.7-20.1-10.3
AV1+[24]2.61.8-3.11.10.6
DCVC-DC [35]-35.3-29.1-11.637.2-9.7
DCVC-FM [36]-6.5-11.7-7.7-5.8-7.9
AV1+Proposed-14.5-13.4-15.1-41.8-21.2
+ +Table 2. BD-rates vs. VVC (more negative is better). $\mathrm{VVC + QPA}$ indicates quantization parameter adaptation is enabled. ALL indicates average BD-rate over all quality scores. + +
DatasetMethodSSIMMSSSIMAVQTVMAFALL
GamingVVC+QPA0.44.613.41.54.9
VVC+DPP [11]-1.3-1.6-9.2-9.0-5.3
VVC+[24]39.532.311.47.822.7
DCVC-DC [35]9.1-11.5-21.338.93.8
DCVC-FM [36]4.918.410.918.413.1
VVC+Proposed-16.6-8.7-19.0-14.3-14.6
XIPHVVC+QPA-0.92.21.12.71.3
VVC+DPP [11]-2.3-1.6-10.5-10.4-6.2
VVC+[24]29.424.412.29.518.9
DCVC-DC [35]-8.1-12.09.241.17.5
DCVC-FM [36]4.313.10.516.88.7
VVC+Proposed-11.4-5.4-16.5-18.0-12.8
UVGVVC+QPA-0.11.67.71.52.7
VVC+DPP [11]-1.4-1.1-7.0-11.4-5.2
VVC+[24]26.823.57.96.516.2
DCVC-DC [35]-8.216.0-1.429.18.9
DCVC-FM [36]-2.714.112.514.79.65
VVC+Proposed-11.8-5.4-13.9-13.8-11.2
HEVC-BVVC+QPA0.91.21.12.11.3
VVC+DPP [11]-19.2-16.7-17.2-16.7-17.4
VVC+[24]0.50.50.1-2.3-0.3
DCVC-DC [35]-17.0-5.56.349.98.4
DCVC-FM [36]18.815.814.412.115.3
VVC+Proposed-31.8-27.4-25.2-33.3-29.4
+ +Tab. 2. Sample visual results of the proposed method are shown in Fig. 4. The results show that the proposed approach is temporally coherent, and better capable of preserving structure, retaining textures and denoising codec artifacts, also reflected in the MOS scores of Fig. 3. + +![](images/788833425de26784a381ec26c10bb7cbc270fdb557b8716396bd3923cdd4c25c.jpg) +(a) + +![](images/327424f0839c08d2ce220f686a7a9981316a3a4606eb3f64d034798b920adb30.jpg) +(b) +Figure 3. Combined rate-quality plots with P.910 MOS. + +Table 3. BD-rates of different component configurations. + +
PrePostSSIMMSSSIMAVQTVMAF
✓✓-25.6-20.8-21.4-20.5
✓✗17.913.229.8-34.1
✗✓-2.90.44.1-33.3
+ +# 4.5. Ablation + +Two-phase pretraining: The rate-quality alignment between target codec $C$ and model $M$ after two-phase pretraining is validated by comparing their performance on a dataset ov dieo sequences. The Pearson correlation coefficient (PCC) is improved from 0.56 and 0.60 to 0.88 and 0.85 for bitrate and PSNR, respectively (see supplementary for more results). Note that the PCC for rate is computed between $R_{C}$ and $R_{M}^{hyp}$ . This indicates that both distortion and bitrate of the codec model and target codec correlate better after the two-phase pretraining, which is a sufficient condition for gradient-based optimization [31]. Tab. 4 shows the effect of pretraining on wrappers. + +Pre/post processing: Tab. 3 shows the effect of removing either the pre or the postprocessor from the overall pipeline under the same training pipeline. Removing the preprocessor is equivalent to training a video denoiser. While BD-rate gains for VMAF are maintained with only the pre- and postprocessor, this can be attributed to sharpening effects over the source content, which VMAF has susceptibility to [1]. Indeed, with more naive content enhancement such as sharpening, fidelity oriented methods such as SSIM and MS-SSIM will experience BD-rate losses, which can be seen when training both pre and postprocessors alone. Only jointly training both pre- and postprocessors is able to achieve substantial BD-rate gains across all quality scores. Fig. 5 provides a generic example where the preprocessor embeds information into the source frame which does not significantly increase the bitrate and can survive the codec compression. This information is used by the postprocessor for reconstruction. + +End-to-end training: For end-to-end training, an ablation is performed over the proposed iterative updates of the codec model and stop-gradient during training. The results are provided in Tab. 4 under different training con + +![](images/d889a979e5edd61335cb6f939923f6830fd09395e6f79308859539dfa3a01874.jpg) + +![](images/84ae54ff76a05b71383ad094cbc2ec1f3ce537f2c7e4b5a029f958edc50a3fdf.jpg) + +![](images/d92019e321f35b83c55cee1a9eb93b8a6a93a8a00b2e463ecd6012006977ff9b.jpg) + +![](images/3ec950dadcabe2aadc1bc52cbd9f2ff865457c28823754c4e623fb111d436114.jpg) +Figure 4. Qualitative comparison of codecs with and without the proposed method. Zoom in for better viewing. + +![](images/d20947a0105eec2526383e64108f069158ea431b02dbefcb8a441c80d8282472.jpg) + +![](images/5e7ee760f98682c083c332657d678ccc8a76d86523edf90af550732cbd5d36ff.jpg) + +![](images/e67df1c3a24f8795c9c7c565baab1cee6becd7f67fa8d4e7e59ae7ef962946a0.jpg) +(a) + +![](images/ef51856a557b7a713cbbb07c1619de990484ba21683b93548de4e6da2deb29a7.jpg) +(b) + +![](images/efe18297648ee5ee11480c7b745083ecb554794f7cb5cddb6353dd0d71cc0115.jpg) +(c) +Figure 5. Results of components of the pipeline where (a), (b), (c) and (d) are the source frame, preprocessor output, codec output and postprocessor output, respectively. + +![](images/a8cf22dddd057e91f0608f2464b208c62c31737ff5e39e74e999214feb11268d.jpg) +(d) + +Table 4. BD-rates with different training configurations. + +
PretrainIter. UpdateStop Grad.SSIMMSSSIMAVQTVMAF
✓✓✓-25.6-20.8-21.4-20.5
X✓✓-23.1-14.4-17.3-18.7
✓✓X16.40.0110.5-15.3
✓X✓-11.2-6.5-7.7-13.1
+ +Figurations. Notably, when training the pre- and postprocessor with only iterative updates, there are BD-rate losses in SSIM and MS-SSIM. This is a by-product of the domain shift between the source and preprocessor output that increases as end-to-end training progresses. When performing end-to-end training with both iterative updates and stop gradient, there is a sufficient alignment constraint on the codec model, which leads to demonstrable BD-rate gains across quality scoring methods. + +# 5. Conclusion + +This work proposes a new approach to jointly train neural pre and postprocessors to improve the rate-quality performance of conventional video CODECs. It is proposed to use a neural codec to model the standard codec's sophisticated rate-quality characteristics, combined with: (i) a novel approach to pretrain the codec model; (ii) a new method to train the neural wrapper model jointly with the codec model in the middle. Extensive evaluations with state-of-the-art AV1 and VVC CODECs show that the proposed neural wrapper approach provides for consistent BD-rate improvement over standard CODECs. Unlike neural methods for preprocessing and encoding, the proposed method allows for consistent improvement over all quality scores (SSIM, MSSSIM, AVQT, VMAF, P.910 MOS) over the entire rate-quality region of modern standards. + +# References + +[1] On vmaf's property in the presence of image enhancement operations. https : // docs . google . com / document / d / 1dJczEhX00MZjBSNyKmd3ARiCTdFVMNPBykH4 _ HMPoyY/edit#heading=h.oaikhnw46pw5. Last accessed: 2020-11-16. 7 +[2] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2020. 2, 3 +[3] Stéphane Alarie, Charles Audet, Aïmen E Gheribi, Michael Kokkolaras, and Sébastien Le Digabel. Two decades of blackbox optimization applications. EURO Journal on Computational Optimization, 9:100011, 2021. 3 +[4] Apple. "Apple avqt presentation", 2021. https://developer.apple.com/videos/play/wwdc2021/10145.1,6 +[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 4 +[6] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6228-6237, 2018. 1 +[7] Mohammad Reza Bonyadi and Zbigniew Michalewicz. Particle swarm optimization for single objective continuous space problems: a review. Evolutionary computation, 25(1):1-54, 2017. 3 +[8] Frank Bossen et al. Common test conditions and software reference configurations. JCTVC-L1100, 12(7):1, 2013. 5 +[9] Eirina Bourtsovlatze, Aaron Chadha, Ilya Fadeev, Vasileios Giotasas, and Yiannis Andreopoulos. Deep video precoding. IEEE Transactions on Circuits and Systems for Video Technology, 2019. 2 +[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 6 +[11] Aaron Chadha and Yiannis Andreopoulos. Deep perceptual preprocessing for video coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14852-14861, 2021. 2, 5, 6, 7 +[12] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 3 +[13] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In + +Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 3 +[14] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5962-5971, 2022. 3 +[15] Jinyoung Choi and Bohyung Han. Task-aware quantization network forJPEG image compression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 309-324. Springer, 2020. 3 +[16] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. In Proceedings of the AAAI conference on artificial intelligence, pages 10696-10703, 2020. 3 +[17] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3476-3485. IEEE, 2019. 3 +[18] Ryosuke Furuta, Naoto Inoue, and Toshihiko Yamasaki. Pixelrl: Fully convolutional network with reinforcement learning for image processing. IEEE Transactions on Multimedia, 22(7):1704-1719, 2019. 3 +[19] Vasileios Giotsas, Nikos Deligiannis, Pam Fisher, and Yiannis Andreopoulos. Perceptual video quality estimation by regression with myopic experts. In 2015 Seventh International Workshop on Quality of Multimedia Experience (QoMEX), pages 1-6. IEEE, 2015. 1 +[20] Onur G Guleryuz, Philip A Chou, Hugues Hoppe, Danhang Tang, Ruofei Du, Philip Davidson, and Sean Fanello. Sandwiched image compression: wrapping neural networks around a standard codec. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3757-3761. IEEE, 2021. 3 +[21] Onur G Guleryuz, Philip A Chou, Hugues Hoppe, Danhang Tang, Ruofei Du, Philip Davidson, and Sean Fanello. Sandwiched image compression: Increasing the resolution and dynamic range of standard codec. In 2022 Picture Coding Symposium (PCS), pages 175-179. IEEE, 2022. 3 +[22] Mengxi Guo, Shijie Zhao, Hao Jiang, Junlin Li, and Li Zhang. Video compression with arbitrary rescaling network. arXiv preprint arXiv:2306.04202, 2023. 3 +[23] Yung-Han Ho, Chih-Peng Chang, Peng-Yu Chen, Alessandro Gnutti, and Wen-Hsiao Peng. Canf-vc: Conditional augmented normalizing flows for video compression. In European Conference on Computer Vision, pages 207-223. Springer, 2022. 2 +[24] Yueyu Hu, Chenhao Zhang, Onur G Guleryuz, Debargha Mukherjee, and Yao Wang. Standard compliant video coding using low complexity, switchable neural wrappers. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1922-1928. IEEE, 2024. 3, 5, 6, 7 +[25] Zhihao Hu and Dong Xu. Complexity-guided slimmable decoder for efficient deep video compression. In Proceed- + +ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14358-14367, 2023. 2 +[26] Quan Huynh-Thu and Mohammed Ghanbari. The accuracy of psnr in predicting video quality for different video scenes and frame rates. Telecommunication Systems, 49: 35-48, 2012. 1 +[27] Berivan Isik, Onur G Guleryuz, Danhang Tang, Jonathan Taylor, and Philip A Chou. Sandwiched video compression: Efficiently extending the reach of standard codecswith neural wrappers. arXiv preprint arXiv:2303.11473, 2023. 3 +[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 645-660. Springer, 2020. 3 +[29] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8008-8017, 2020. 3 +[30] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3224-3232, 2018. 3 +[31] Jan P Klopp, Keng-Chi Liu, Liang-Gee Chen, and Shao-Yi Chien. How to exploit the transferability of learned image compression to conventional codescs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16165-16174, 2021. 2, 3, 7 +[32] Faouzi Kossentini, Hassen Guermazi, Nader Mahdi, Chekib Nouira, Amir Naghdinezhad, Hassene Tmar, Omar Khelif, Phoenix Worth, and Foued Ben Amara. The svt-av1 encoder: overview, features and speed-quality tradeoffs. In Applications of Digital Image Processing XLIII, page 1151021. International Society for Optics and Photonics, 2020. 3, 6 +[33] Théo Ladune, Pierrick Philippe, Wassim Hamidouche, Lu Zhang, and Olivier Déforges. Conditional coding for flexible learned video compression. arXiv preprint arXiv:2104.07930, 2021. 2 +[34] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022. 2 +[35] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 2, 6, 7 +[36] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with feature modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26099-26108, 2024. 2, 6, 7 + +[37] Songnan Li, Fan Zhang, Lin Ma, and King Ngi Ngan. Image quality assessment by separately evaluating detail losses and additive impairments. IEEE Transactions on Multimedia, 13(5):935-949, 2011. 4 +[38] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. Mucan: Multi-correspondence aggregation network for video super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 335-351. Springer, 2020. 3 +[39] Yinxiao Li, Pengchong Jin, Feng Yang, Ce Liu, Ming-Hsuan Yang, and Peyman Milanfar. Comisr: Compression-informed video super-resolution. in 2021 IEEE. In CVF International Conference on Computer Vision (ICCV), pages 2523-2532, 2021. 3 +[40] Zhi Li and Christos G Bampis. Recover subjective quality scores from noisy measurements. In 2017 Data compression conference (DCC), pages 52-61. IEEE, 2017. 6 +[41] Zhi Li, Anne Aaron, Ioannis Katsavounidis, Anush Moorthy, and Megha Manohara. Toward a practical perceptual video quality metric. The Netflix Tech Blog, 6, 2016. 4 +[42] Zhi Li, Christos Bampis, Julie Novak, Anne Aaron, Kyle Swanson, Anush Moorthy, and J Cock. Vmaf: The journey continues. *Netflix Technology Blog*, 2018. 1, 6 +[43] Zhi Li, Christos G Bampis, Lukáš Krasula, Lucjan Janowski, and Ioannis Katsavounidis. A simple model for subject behavior in subjective experiments. arXiv preprint arXiv:2004.02067, 2020. 6 +[44] Jianping Lin, Dong Liu, Houqiang Li, and Feng Wu. MLvc: Multiple frames prediction for learned video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3546-3554, 2020. 2 +[45] Jerry Liu, Shenlong Wang, Wei-Chiu Ma, Meet Shah, Rui Hu, Pranaab Dhawan, and Raquel Urtasun. Conditional entropy coding for efficient video compression. In European Conference on Computer Vision, pages 453-468. Springer, 2020. 2 +[46] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. Dvc: An end-to-end deep video compression framework. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 2 +[47] Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, and Eirikur Agustsson. Vct: A video compression transformer. arXiv preprint arXiv:2206.07307, 2022. 2 +[48] Alexandre Mercat, Marko Viitanen, and Jarno Vanne. Uvg dataset: 50/120fps 4k sequences for video codec analysis and development. In Proceedings of the 11th ACM Multimedia Systems Conference, pages 297-302, 2020. 5 +[49] Reza Pourreza, Hoang Le, Amir Said, Guillaume Sautiere, and Auke Wiggers. Boosting neural video codecs by exploiting hierarchical redundancy. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5355-5364, 2023. 2 + +[50] Zhongwei Qiu, Huan Yang, Jianlong Fu, and Dongmei Fu. Learning spatiotemporal frequency-transformer for compressed video super-resolution. In European Conference on Computer Vision, pages 257-273. Springer, 2022. 3 +[51] Alexander Raake, Silvio Borer, Shahid M Satti, Jörgen Gustafsson, Rakesh Rao Ramachandra Rao, Stefano Medagli, Peter List, Steve Göring, David Lindero, Werner Robitza, et al. Multi-model standard for bitstream-, pixel-based and hybrid video quality assessment of uhd/4k: Itu-t p. 1204. IEEE Access, 8:193020-193049, 2020. 1 +[52] Reza Rassool. Vmaf reproducibility: Validating a perceptual practical video quality metric. In 2017 IEEE international symposium on broadband multimedia systems and broadcasting (BMSB), pages 1-2. IEEE, 2017. 6 +[53] Oren Rippel, Alexander G Anderson, Kedar Tatwawadi, Sanjay Nair, Craig Lytle, and Lubomir Bourdev. Elf-vc: Efficient learned flexible-rate video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14479-14488, 2021. 2 +[54] Amir Said, Manish Kumar Singh, and Reza Pourreza. Differentiable bit-rate estimation for neural-based video codec enhancement. In 2022 Picture Coding Symposium (PCS), pages 379-383. IEEE, 2022. 3 +[55] Mehdi SM Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6626-6634, 2018. 3 +[56] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 2 +[57] Hanbin Son, Taeoh Kim, Hyeongmin Lee, and Sangyoun Lee. Enhanced standard compatible image compression framework based on auxiliary codec networks. IEEE Transactions on Image Processing, 31:664-677, 2021. 3 +[58] Yannick Strümpler, Ren Yang, and Radu Timofte. Learning to improve image compression without changing the standard decoder. In European Conference on Computer Vision, pages 200–216. Springer, 2020. 2 +[59] Hossein Talebi and Peyman Milanfar. Learned perceptual image enhancement. In 2018 IEEE International Conference on Computational Photography (ICCP), pages 1-13. IEEE, 2018. 2 +[60] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In Proceedings of the IEEE international conference on computer vision, pages 4472-4480, 2017. 3 +[61] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu Tdan. temporally-deformable alignment network for video super-resolution. in 2020 ieee. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3357-3366, 2020. 3 +[62] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 3 + +[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5 +[64] Ties van Rozendaal, Tushar Singhal, Hoang Le, Guillaume Sautiere, Amir Said, Krishna Buska, Anjuman Raha, Dimitris Kalatzis, Hitarth Mehta, Frank Mayer, et al. Mobilenvc: Real-time 1080p neural video compression on a mobile device. arXiv preprint arXiv:2310.01258, 2023. 2 +[65] Abhinau K Venkataramanan, Chengyang Wu, Alan C Bovik, Ioannis Katsavounidis, and Zafar Shahid. A hitchhiker's guide to structural similarity. IEEE Access, 9: 28872-28896, 2021. 1 +[66] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 3 +[67] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Huchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2012-2021, 2023. 3 +[68] Ping-Hao Wu, Ioannis Katsavounidis, Zhijun Lei, David Ronca, Hassene Tmar, Omran Abdelkafi, Colton Cheung, Foued Ben Amara, and Faouzi Kossentini. Towards much better svt-av1 quality-cycles tradeoffs for vod applications. In Applications of Digital Image Processing XLIV, pages 236–256. SPIE, 2021. 6 +[69] Li Xu, Gang He, Jinjia Zhou, Jie Lei, Weiying Xie, Yun-song Li, and Yu-Wing Tai. Transcoded video restoration by temporal spatial auxiliary network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2875-2883, 2022. 3 +[70] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8):1106-1125, 2019. 3, 5 +[71] Ren Yang, Mai Xu, and Zulin Wang. Decoder-side hevc quality enhancement with scalable convolutional neural network. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 817-822. IEEE, 2017. 2 +[72] Ren Yang, Mai Xu, Tie Liu, Zulin Wang, and Zhenyu Guan. Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology, 29(7):2039-2054, 2018. 2 +[73] Ren Yang, Mai Xu, Zulin Wang, and Tianyi Li. Multi-frame quality enhancement for compressed video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6664-6673, 2018. 2 +[74] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3106-3115, 2019. 3 +[75] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video superresolution. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision, pages 4429-4438, 2021. 3 +[76] Hengsheng Zhang, Xueyi Zou, Jiaming Guo, Youliang Yan, Rong Xie, and Li Song. A codec information assisted framework for efficient compressed video superresolution. In European Conference on Computer Vision, pages 220-235. Springer, 2022. 3 \ No newline at end of file diff --git a/2025/Perceptual Video Compression with Neural Wrapping/images.zip b/2025/Perceptual Video Compression with Neural Wrapping/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..21707d18e132e3c4e49ad9cd6851cb7970fedb4f --- /dev/null +++ b/2025/Perceptual Video Compression with Neural Wrapping/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94318730fdaf58898dc48eee9a1653a59d636e2bac80379e19642e3af99a5453 +size 554624 diff --git a/2025/Perceptual Video Compression with Neural Wrapping/layout.json b/2025/Perceptual Video Compression with Neural Wrapping/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a6673c35de949864d7ab359ee35059545f25a6fd --- /dev/null +++ b/2025/Perceptual Video Compression with Neural Wrapping/layout.json @@ -0,0 +1,11205 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 464, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 464, + 121 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 464, + 121 + ], + "type": "text", + "content": "Perceptual Video Compression with Neural Wrapping" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 143, + 541, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 143, + 541, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 541, + 171 + ], + "type": "text", + "content": "Muhammad Umar Karim Khan, Aaron Chadha, Mohammad Ashraful Anam, Yiannis Andreopoulos, Sony Interactive Entertainment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 174, + 496, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 174, + 496, + 185 + ], + "spans": [ + { + "bbox": [ + 94, + 174, + 496, + 185 + ], + "type": "text", + "content": "{Umar . Khan, Aaron . Chadha, Russell . Anam, Yiannis . Andreopoulos}@sony.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 228, + 289, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 612 + ], + "type": "text", + "content": "Standard video codecs are rate-distortion optimization machines, where distortion is typically quantified using PSNR versus the source. However, it is now widely accepted that increasing PSNR does not necessarily translate to better visual quality. In this paper, a better balance between perception and fidelity is targeted, in order to provide for significant rate savings over state-of-the-art standards-based video codecs. Specifically, pre- and post-processing neural networks are proposed that enhance the coding efficiency of standard video codecs when benchmarked with an array of well-established perceptual quality scores. These \"neural wrapper\" elements are end-to-end trained with a neural codec module serving as a differentiable proxy for standard video codecs. The codec proxy is jointly optimized with the pre- and post components via a novel two-phase pretraining strategy and end-to-end iterative refinement with stop-gradient. This allows the neural pre- and postprocessor to learn to embed, remove and recover information in a codec-aware manner, thus improving its rate-quality performance. A single neural-wrapper model is thereby established and used for the entire rate-quality curve without needing any downscaling or upscaling. The trained model is tested with the AV1 and VVC standard codecs via an array of well-established objective quality scores (SSIM, MS-SSIM, VMAF, AVQT), as well as mean opinion scores (MOS) derived from ITU-T P.910 subjective testing. Experimental results show that the proposed approach improves all quality scores, with " + }, + { + "bbox": [ + 46, + 228, + 289, + 612 + ], + "type": "inline_equation", + "content": "-18.5\\%" + }, + { + "bbox": [ + 46, + 228, + 289, + 612 + ], + "type": "text", + "content": " average Bjontegaard Delta-rate (BD-rate) saving over all objective scores and MOS improvement over both standard codecs. This illustrates the significant potential of neural wrapper components over standards-based video coding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 621, + 127, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 127, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 127, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "It is now widely accepted that that signal-to-noise (SNR) ratio is a poor indicator of visual quality in video coding [6, 19, 26, 42]. Instead, quality scores that include elements of human perception [42], perceptual modelling of encoding artifacts [51], as well as viewing setup awareness [4, 65] have emerged as strong contenders for the best means to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 310 + ], + "type": "text", + "content": "objectively characterize visual quality. This has also led to the research community now moving away from SNR-optimization [26] in favor of structural similarity (SSIM [65]), video multimethod assessment fusion (VMAF) and Apple's advanced video quality tool (AVQT) optimization [4, 42]. However, all current perceptual optimization approaches in standards-based video encoders have one or more of the following detriments:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 313, + 545, + 456 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 306, + 313, + 545, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 313, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 306, + 313, + 545, + 336 + ], + "type": "text", + "content": "- they require multiple encoding passes or in-loop implementation within a specific encoder;" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 337, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 337, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 337, + 545, + 372 + ], + "type": "text", + "content": "- they only optimize for a single quality scoring method like VMAF or SSIM and are shown to be detrimental in other quality scores;" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 373, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 545, + 420 + ], + "type": "text", + "content": "- they comprise hand-crafted (shallow) models of low-level human perception and fail to encapsulate several characteristics of more advanced quality scoring methods like VMAF or AVQT in a data-driven and learnable manner;" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 421, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 421, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 421, + 545, + 456 + ], + "type": "text", + "content": "their Bjontegaard Delta-rate (BD-rate) improvement on well-established quality scores like SSIM, VMAF and AVQT is in the order of a few percentile points." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 457, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 457, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 312, + 457, + 545, + 601 + ], + "type": "text", + "content": "A neural codec \"wrapper\" is proposed in this paper as the means to augment standard video codecs and optimize them for perceptual quality. Including neural pre- and postprocessing in conventional encoding pipelines is shown to clearly outperform the equivalent codecs without these components. This allows for conventional streaming ecosystems to continue operating \"as-is\" and does not break standards. That is, when neural components cannot be supported before encoding or after decoding, the system can just disable them with appropriate adaptation, without jeopardizing video streaming. The key contributions of the paper are listed as follows." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 605, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 545, + 654 + ], + "type": "text", + "content": "- A novel pretraining regime for modelling a conventional codec with a neural codec model is proposed, which uses the implicit encoder-decoder structure of the neural codec to align both video quality and rate." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 545, + 713 + ], + "type": "text", + "content": "- An end-to-end training regime for jointly training pre and postprocessors is proposed. Starting with a pretrained neural codec model, it is proposed to iteratively update the codec model. This approach allows for more accurate gradients to be backpropagated to the prepro" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17743" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 73, + 509, + 196 + ], + "blocks": [ + { + "bbox": [ + 85, + 73, + 509, + 196 + ], + "lines": [ + { + "bbox": [ + 85, + 73, + 509, + 196 + ], + "spans": [ + { + "bbox": [ + 85, + 73, + 509, + 196 + ], + "type": "image", + "image_path": "d2259fbf37c5d80f7edf44b7b7d2f9ae61208435af9ec137f09cd4559ece28bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "type": "text", + "content": "Figure 1. Left: Functional block diagram of the multi-stage training data flow. Right: Components of codec model " + }, + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "type": "text", + "content": " that are used as a proxy for the target codec " + }, + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "type": "inline_equation", + "content": "AE/AD" + }, + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "type": "text", + "content": " denotes arithmetic encoding followed by arithmetic decoding and " + }, + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "type": "text", + "content": " denotes quantization. See Sec. 3 for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 262, + 96, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 262, + 96, + 271 + ], + "spans": [ + { + "bbox": [ + 65, + 262, + 96, + 271 + ], + "type": "text", + "content": "cessor." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 272, + 289, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 289, + 357 + ], + "type": "text", + "content": "- The trained pipeline is evaluated on gaming sequences and significant BD-rate gains versus AV1 and VVC encoding are demonstrated with just a single pre- and postprocessor model. In addition, the proposed neural wrapper is shown to outperform a state-of-the-art neural encoder [35], as well as the state-of-the-art in deep perceptual preprocessing proposed recently [11]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 288, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 288, + 421 + ], + "type": "text", + "content": "A functional block diagram for the proposed system is shown in Fig. 1. Example visuals produced by the proposed method are shown in Fig. 4, demonstrating that the proposed neural wrapper framework allows for better detail preservation than the underlying standard codec." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 436, + 143, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 143, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 143, + 449 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 177, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 177, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 177, + 468 + ], + "type": "text", + "content": "2.1. Neural Video Codes" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 289, + 715 + ], + "type": "text", + "content": "Neural video codecs optimize neural networks for rate and distortion by exploiting redundancies in video data. Most notable are residual coding methods; Lu et al. [46] introduce deep video codec, which adopts neural networks within the residual coding pipeline of a traditional codec. Optical flow is employed as a proxy for motion vectors in order to warp the reference frame and generate a residual frame for encoding. Numerous other methods have been proposed based on residual coding [44, 49, 53]. Scale-space flow estimation is proposed as an improvement over traditional optical flow by Agustsson et al. [2], namely for better handling of complex motion in the scene. Beyond residual coding methods, conditional coding methods learn feature contexts implicitly from the encoded context [23, 33, 34, 45, 47, 56]. Li et al. [35] propose using diverse contexts in spatial and temporal directions to perform video compression and outperform standard video codecs when the latter are used in low-delay mode. They improve the performance and bitrate range in [36]. Recent work has additionally focused on" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 260, + 547, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 547, + 321 + ], + "type": "text", + "content": "reducing the complexity of neural CODECs. For example, Van Rozendaal et al. [64] deploy an architecture similar to Agustsson et al. [2] on a smart phone and achieve real time performance at 1080p. Hu et al. [25] adopt slimmable networks for variable complexity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 331, + 403, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 331, + 403, + 344 + ], + "spans": [ + { + "bbox": [ + 314, + 331, + 403, + 344 + ], + "type": "text", + "content": "2.2. Preprocessors" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 349, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 547, + 601 + ], + "type": "text", + "content": "While postprocessing is a more mature line of research, there has been a prevalence of work in preprocessing, both on image and video modalities. In the image domain, Strumpler et al. [58] propose to modify the JPEG encoder with an attention network and a learnable quantization table, in order to modulate the source frame and provide for additional rate savings over the JPEG encoder alone. Conversely, Talebi et al. [59] avoid adapting the JPEG encoder by prepended the standard encoder with a preprocessor for deployment with a standard JPEG encoder and decoder. Klopp et al. [31] train a preprocessor with a surrogate neural codec model, which is trained to model a target image codec. In the video domain, Boutssoulatze et al. [9] propose a precoding network for downscaling with a online mode selection algorithm, in order to alter the precoding based on scene characteristics. Conversely, Chadha et al. [11] propose an entirely offline training mechanism by first modelling the core components of a video codec in a differentiable manner. A preprocessor is optimized on the codec model, in order to provide further rate-quality savings over the target codec." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 611, + 406, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 611, + 406, + 624 + ], + "spans": [ + { + "bbox": [ + 314, + 611, + 406, + 624 + ], + "type": "text", + "content": "2.3. Postprocessors" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 547, + 715 + ], + "type": "text", + "content": "Numerous deep learning methods have been proposed for compressed video denoising, video super-resolution or both tasks. Single frame methods for denoising compressed videos have been proposed in [71, 72]. Multiframe methods have provided superior gains in metrics compared to single frame methods. MFQE [73] detects frames with higher quality and uses it as a reference to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17744" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 286, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 286, + 396 + ], + "type": "text", + "content": "improve neighboring frames. TSAN [69] uses a temporal deformable alignment and multi-scale fusion for the task of video restoration. STDF [16] uses spatio-temporal deformable convolution to aggregate temporal context rather than using optical flow for frame warping. A common approach for video super-resolution is to align neighboring frames and transfer information across frames [60, 61, 66, 70]. On the contrary, other methods choose non-local modules or 3D convolutions to gather temporal context [29, 30, 38, 74]. Some methods recurrently propagate history in the forward direction [17, 28, 55]. Recent methods such as BasicVSR [12], GOVSR [75] and BasicVSR++ [13] propagate history in forward and backward directions, such that both past and future frames are used to super-resolve the current frame. Architectures developed for video super-resolution can easily be adapted for video denoising. Few authors have focused on the task of jointly denoising and upscaling videos. FTVSR [50] uses a frequency transformer which performs self-attention in space, time and frequency to recover details. RealBasicVSR [14] considers sensor noise, motion blur and compression artifacts with super-resolution. CIAF [76] uses motion-vectors from the video to approximate optical flow. CAVSR [67] adapts the super-resolution to the compression level of the input video. COMISR [39] uses an architecture based on BasicVSR [12] to perform joint denoising and upscaling." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 407, + 286, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 407, + 286, + 431 + ], + "spans": [ + { + "bbox": [ + 57, + 407, + 286, + 431 + ], + "type": "text", + "content": "2.4. Pre- and Postprocessors with Traditional Codes" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 438, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 438, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 438, + 286, + 712 + ], + "type": "text", + "content": "Recent methods have also explored joint training of pre and postprocessors as neural wrappers. In the image domain, recent work [20, 21] emulates the core blocks of the JPEG encoder and decoder, and uses it to train pre and postprocessors for rate-distortion savings. Similar sandwich approach for videos has been proposed in [24]. Son et al. [57] propose distinct neural networks to model the distortion and rate behavior of a standard codec, which is thus sandwiched by pre- and postprocessor networks for rate-distortion optimization. However, these methods have no target codec awareness beyond the implicit intra/inter coding structure. Either the codec proxy is not learned (e.g. derived from differentiable JPEG/DCT as in [24]), only partially learned via rate-distortion optimization (e.g. the loop filter as in [27]) or only operates in intra mode (as in [57]). In this paper, the proposed framework is directly conditioned on the target conventional codec. Tian et al. [62] target additional rate savings with learnable rescaling. However, this method has no rate control or alignment outside of the scaling factor. Conversely, the framework proposed in this paper is jointly conditioned on both the target codec rate and distortion and is thus able to offer rate savings even when encoding at full res" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 316, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 545, + 204 + ], + "type": "text", + "content": "olution. Additionally, whereas recent work restricts evaluation to a limited range of the rate-distortion space, at high bitrates [61] or substantially lower bitrates and resolutions [22] and on older encoding standards (AVC or HEVC), the proposed method is evaluated against state-of-the-art AV1 and VVC encoding, i.e.: " + }, + { + "bbox": [ + 316, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 316, + 72, + 545, + 204 + ], + "type": "text", + "content": " the entire rate-quality region of practical significance is considered (VMAF values between 40-96 and bitrates from 100kbps to 15mbps); " + }, + { + "bbox": [ + 316, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 316, + 72, + 545, + 204 + ], + "type": "text", + "content": " slow encoding presets are used (preset=4 for SVT-AV1 and preset=slow for VVenC), without limiting the encoders to low-delay mode." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 214, + 395, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 214, + 395, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 214, + 395, + 227 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": "The top-level functional block diagram of the proposed training flow is shown in Fig. 1 and includes a preprocessor " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": ", codec model " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": ", target codec " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " and postprocessor " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": ". Notably, the target codec is a traditional video codec, such as SVT-AV1 [32]. First, a source frame " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{in}}" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " is processed by " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " to generate a corresponding preprocessed frame " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "I_P" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": ". The preprocessed frame is simultaneously encoded by " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": ", which generate encoded frames " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "I_M" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "I_C" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " respectively. Finally, encoded frames are postprocessed with " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " to generate the output frame " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{out}}" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": ". End-to-end distortion " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " is computed between " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{in}}" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{out}}" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " and the overall training loss is the weighted sum of " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " and the bitrate " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "R_M" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": ". Notably, at inference, the codec model " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " is removed from the pipeline and only the target codec " + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 316, + 234, + 545, + 413 + ], + "type": "text", + "content": " is deployed for encoding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 420, + 455, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 455, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 455, + 433 + ], + "type": "text", + "content": "3.1. Codec Model Pretraining" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 438, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 545, + 593 + ], + "type": "text", + "content": "Given the target codec " + }, + { + "bbox": [ + 316, + 438, + 545, + 593 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 316, + 438, + 545, + 593 + ], + "type": "text", + "content": " in Fig. 1 is not differentiable, a mechanism for adapting the proposed pipeline to the target codec behavior is required. One option is black-box or derivative-free optimization methods [3, 7, 18]. However these methods typically rely on an approximation of the unknown function to be optimized, and are computationally infeasible for training at scale. Another option is to replicate the core prediction and transform blocks of the target codec in a differentiable manner [15, 27, 54]. Such approaches can model older codecs such as JPEG, or faster presets of AVC or HEVC successfully to an extent but fail to accurately model the operation of more sophisticated codecs such as AV1 or VVC." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 594, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 594, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 594, + 545, + 712 + ], + "type": "text", + "content": "Recent work [31, 57] has instead leveraged on neural image and video CODEs for predicting the target codec behavior. In this work, a similar approach is followed, and a neural video codec " + }, + { + "bbox": [ + 316, + 594, + 545, + 712 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 316, + 594, + 545, + 712 + ], + "type": "text", + "content": " is adapted to model the ratedistortion behavior of a target codec " + }, + { + "bbox": [ + 316, + 594, + 545, + 712 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 316, + 594, + 545, + 712 + ], + "type": "text", + "content": ". Specifically, the scale-space flow architecture [2] is used to generate lossy frames, and the implicit encoder-decoder structure and associated loss functions are leveraged to align to the target codec behavior in a spatially localized manner. As per Agustsson et al. [2], for each of intra, flow and residual" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17745" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "text", + "content": "encoding, the hyperprior based architecture of Balle et al. [5] is used, without any autoregressive components." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": "Notably, the hyperprior architecture comprises analysis and synthesis transforms, " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": ", for encoding to and decoding from the quantized latent space " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": " respectively. A distortion loss " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "D(:,g_a,g_s)" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": " can thus be computed between source frame " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": " (input to " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": ") and reconstructed frame " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "I_M" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": " (output by " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": "). A hyper-analysis and hypersynthesis transform " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "(h_a" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": " respectively) constitute the hyperprior, which is used to model scale dependencies " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "\\tilde{\\sigma}" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": " within the latent space. The latent space " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": " is thus modelled as a Gaussian scale mixture on which the rate " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "R(:,h_a,h_s)" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": " can be optimized (as Shannon entropy). Traditionally, the architecture would then be trained with an aggregate loss function " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "L = D + \\lambda R" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 96, + 286, + 275 + ], + "type": "text", + "content": " represents a Lagrangian parameter for balancing between the rate and distortion components." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "text", + "content": "When performing neural codec alignment to a target codec, there are two key issues: (i) the model encoded frame must align to the target encoded frame (and not the source); (ii) selecting the best value for " + }, + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "text", + "content": " when trying to replicate target codec behavior is a non-trivial problem. Crucially, it is noted that the rate model is exclusively parameterized by " + }, + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "inline_equation", + "content": "h_a" + }, + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "text", + "content": ", and distortion exclusively by " + }, + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "inline_equation", + "content": "g_a" + }, + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "inline_equation", + "content": "g_s" + }, + { + "bbox": [ + 55, + 275, + 286, + 395 + ], + "type": "text", + "content": ". Subsequently, a novel two phase training methodology is proposed for aligning both rate and distortion to the target codec:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 286, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 286, + 466 + ], + "type": "text", + "content": "Phase 1: The first issue of target codec alignment is addressed via the distortion loss. To this end, the reference for the distortion loss is changed from the source frame " + }, + { + "bbox": [ + 55, + 407, + 286, + 466 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 55, + 407, + 286, + 466 + ], + "type": "text", + "content": " to the target codec frame " + }, + { + "bbox": [ + 55, + 407, + 286, + 466 + ], + "type": "inline_equation", + "content": "I_C" + }, + { + "bbox": [ + 55, + 407, + 286, + 466 + ], + "type": "text", + "content": " and the mean squared error is optimized:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 471, + 286, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 471, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 90, + 471, + 286, + 485 + ], + "type": "interline_equation", + "content": "D _ {M} ^ {\\mathrm {h y p}} \\left(I _ {M}, I _ {C}; \\theta_ {g}\\right) = \\mathbb {E} _ {I _ {C}} \\left[ \\left\\| I _ {M} - I _ {C} \\right\\| ^ {2} \\right] \\tag {1}", + "image_path": "57381f05067ba5d0dd49dff6f852749e2725477ae1ba04e7a0d2eea703dcabb0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 492, + 286, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 286, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 286, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 492, + 286, + 515 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 55, + 492, + 286, + 515 + ], + "type": "text", + "content": " represent the analysis and synthesis transform parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "text", + "content": "Phase 2: Following alignment of " + }, + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "inline_equation", + "content": "g_{a}" + }, + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "text", + "content": " in Phase 1, the latent space " + }, + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "text", + "content": " is now conditioned on the target codec " + }, + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "text", + "content": ". To circumvent selection of an optimal " + }, + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 517, + 286, + 635 + ], + "type": "text", + "content": " for rate, the implicit encoder-decoder structural prior of the aligned neural codec is leveraged in order to learn an unsupervised rate alignment to the target codec. Specifically, the maximum likelihood on the hyperprior model parameters is optimized, in order to learn the latent space distribution under target codec distortion alignment. The rate is thus optimized as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 639, + 307, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 307, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 307, + 665 + ], + "type": "interline_equation", + "content": "R _ {M} ^ {\\mathrm {h y p}} \\left(I _ {M}; \\theta_ {h}\\right) = - \\mathbb {E} _ {\\tilde {y}, \\tilde {z}} \\left[ \\log p _ {\\tilde {y} | \\tilde {z}, I _ {C}} \\left(\\tilde {y} | \\tilde {z}, I _ {C}\\right) + \\log p _ {\\tilde {z} | I _ {C}} \\left(\\tilde {z} | I _ {C}\\right) \\right] \\tag {2}", + "image_path": "02102e14e3f082bae207d0db67e987e11702ef96d559a91fd5ccb1d6c911d10b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 666, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 286, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 666, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\theta_h" + }, + { + "bbox": [ + 55, + 666, + 286, + 713 + ], + "type": "text", + "content": " represents the hyperprior model parameters and " + }, + { + "bbox": [ + 55, + 666, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 55, + 666, + 286, + 713 + ], + "type": "text", + "content": " is the quantized hyperprior latent space (which would also be encoded and transmitted as side information in a neural codec pipeline). For intra coding, the intra" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 131 + ], + "type": "text", + "content": "hyperprior model is optimized on intra encoded source frames using the two-phase strategy. For inter coding, the residual and flow hyperprior models are optimized jointly by summing the rate and distortion components in each phase." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "text", + "content": "The proposed two-phase pretraining ensures that " + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "inline_equation", + "content": "R_{M}^{\\mathrm{hyp}}" + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "text", + "content": " is close to the target codec " + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "inline_equation", + "content": "R_{C}" + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "text", + "content": " by performing an affine transform on " + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "inline_equation", + "content": "R_{M}^{\\mathrm{hyp}}" + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "inline_equation", + "content": "R_{M} = \\alpha R_{M}^{\\mathrm{hyp}} + \\beta" + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 133, + 545, + 192 + ], + "type": "text", + "content": " are scalar learnable parameters optimized with MSE:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 194, + 545, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 194, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 328, + 194, + 545, + 211 + ], + "type": "interline_equation", + "content": "L _ {M} ^ {(\\alpha , \\beta)} \\left(R _ {M}, R _ {C}; \\alpha , \\beta\\right) = \\mathbb {E} _ {R _ {C}} \\left[ \\left\\| R _ {M} - R _ {C} \\right\\| ^ {2} \\right]. \\tag {3}", + "image_path": "c26daa717faab8e963685b1d544347afb25a0c3b4965502647361a6c474e7369.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 220, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 545, + 269 + ], + "type": "text", + "content": "The similarity of " + }, + { + "bbox": [ + 313, + 220, + 545, + 269 + ], + "type": "inline_equation", + "content": "R_{M}^{\\mathrm{hyp}}" + }, + { + "bbox": [ + 313, + 220, + 545, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 220, + 545, + 269 + ], + "type": "inline_equation", + "content": "R_{C}" + }, + { + "bbox": [ + 313, + 220, + 545, + 269 + ], + "type": "text", + "content": " is required to properly balance the rate and distortion losses in end-to-end training. The aggregate rate and distortion for a sequence can thus be measured by computing the sum over all frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 276, + 436, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 276, + 436, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 276, + 436, + 289 + ], + "type": "text", + "content": "3.2. End-to-End Training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 294, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 545, + 365 + ], + "type": "text", + "content": "With the pretrained neural codec model, an end-to-end trainable pipeline with both the pre and postprocessor is developed, as illustrated in Fig. 1. In this case, the training loss is very similar to conventional neural CODECs, but is now exclusively used to train the pre and postprocessors, i.e.:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 369, + 545, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 369, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 324, + 369, + 545, + 384 + ], + "type": "interline_equation", + "content": "L \\left(I _ {\\text {o u t}}, I _ {\\text {i n}}; \\theta_ {P}, \\theta_ {O}\\right) = \\mathbb {E} _ {I _ {\\text {i n}}} \\left[ D \\left(I _ {\\text {i n}}, I _ {\\text {o u t}}\\right) + \\lambda R _ {M} \\right] \\tag {4}", + "image_path": "8806a0c6dc1e98282fcd1c465c55277c397afca9fab573c63f0362bfe9f2fbab.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "inline_equation", + "content": "\\theta_{P}" + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "inline_equation", + "content": "\\theta_{O}" + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "text", + "content": " are the pre and postprocessor parameters respectively, " + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "inline_equation", + "content": "R_{M}" + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "text", + "content": " is the codec model rate estimation (proportional to " + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "inline_equation", + "content": "R_{M}^{\\mathrm{flow}} + R_{M}^{\\mathrm{res}}" + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "text", + "content": " for inter coding the flow and residual from motion compensation, and " + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "inline_equation", + "content": "R_{M}^{\\mathrm{intra}}" + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "text", + "content": " for intra coding) and " + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "text", + "content": " is the distortion loss component. Following subjective assessment tests, a weighted sum of: mean absolute error (MAE), SSIM, MS-SSIM and detail loss metric (DLM) * [37] is proposed for the distortion component " + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "text", + "content": " in Eq. (4). Specifically, " + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "inline_equation", + "content": "D(I_{\\mathrm{in}},I_{\\mathrm{out}}) = \\lambda_{\\mathrm{L1}}||I_{\\mathrm{in}} - I_{\\mathrm{out}}|| + \\lambda_{\\mathrm{SSIM}}(1 - D_{\\mathrm{SSIM}}(I_{\\mathrm{in}},I_{\\mathrm{out}})) + \\lambda_{\\mathrm{MSSSIM}}(1 - D_{\\mathrm{MSSSIM}}(I_{\\mathrm{in}},I_{\\mathrm{out}})) + \\lambda_{\\mathrm{DLM}}D_{\\mathrm{DLM}}(I_{\\mathrm{in}},I_{\\mathrm{out}})" + }, + { + "bbox": [ + 313, + 392, + 545, + 524 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 524, + 545, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 545, + 668 + ], + "type": "text", + "content": "Concept drift or domain shift for the codec model is an expected consequence of end-to-end training. The codec model is pretrained with natural videos whereas the preprocessor is continuously updated as training proceeds; thus, the data observed by the codec model has a non-stationary distribution. Since neural networks generally do not perform well on out-of-distribution data, an iterative refinement mechanism is proposed, where the codec model parameters " + }, + { + "bbox": [ + 313, + 524, + 545, + 668 + ], + "type": "inline_equation", + "content": "\\theta_{M}" + }, + { + "bbox": [ + 313, + 524, + 545, + 668 + ], + "type": "text", + "content": " are independently updated after each iteration of end-to-end training using the two-phase approach described previously. Notably, in this case the input to the target codec is now the output of the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 674, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 674, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 674, + 545, + 713 + ], + "type": "text", + "content": "* DLM is one of the core components of VMAF [41] and is found to better preserve details under visual degradation. Given that it is primarily wavelet based, it can be implemented in a differentiable manner and used for model optimization." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17746" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "text", + "content": "preprocessor " + }, + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "inline_equation", + "content": "I_{P}" + }, + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "text", + "content": " and not the source frame, and " + }, + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "inline_equation", + "content": "I_{C}" + }, + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "inline_equation", + "content": "R_{C}" + }, + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "text", + "content": " represent the frame and rate output by the target codec." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 287, + 192 + ], + "type": "text", + "content": "Additionally, the gradients backpropagated to the preprocessor can be better aligned towards the target codec via the stop-gradient operator. First, it is assumed for the remaining discussion in the section that the tensors representing the frames are flattened, i.e., " + }, + { + "bbox": [ + 55, + 96, + 287, + 192 + ], + "type": "inline_equation", + "content": "\\{I_P,I_M,I_C\\} \\in \\mathbb{R}^{N\\times 1}" + }, + { + "bbox": [ + 55, + 96, + 287, + 192 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 96, + 287, + 192 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 96, + 287, + 192 + ], + "type": "text", + "content": " is the number of pixels in the tensor. If the target codec " + }, + { + "bbox": [ + 55, + 96, + 287, + 192 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 96, + 287, + 192 + ], + "type": "text", + "content": " was differentiable then the accuracy of the gradients backpropagated will increase by decreasing:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 91, + 198, + 287, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 198, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 91, + 198, + 287, + 211 + ], + "type": "interline_equation", + "content": "G = \\left| J _ {M} \\left(I _ {P}\\right) \\nabla_ {I _ {M}} L - J _ {C} \\left(I _ {P}\\right) \\nabla_ {I _ {C}} L \\right|, \\tag {5}", + "image_path": "0c61b707f19f6966376b1d4ff8476a45a2cc68b801e2ddb4933eaf708b21d4ef.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "inline_equation", + "content": "J_{M}(I_{P})\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "text", + "content": " is the transposed Jacobian of " + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "text", + "content": " w.r.t. " + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "inline_equation", + "content": "I_P" + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "inline_equation", + "content": "\\nabla_{I_M}L" + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "text", + "content": " is the gradient of the objective w.r.t. " + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "inline_equation", + "content": "I_M" + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "text", + "content": ". Assuming that the frames " + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "inline_equation", + "content": "I_C" + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "text", + "content": " generated by the target codec " + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "text", + "content": " are input to the postprocessor " + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 214, + 287, + 274 + ], + "type": "text", + "content": " then Eq. (5) becomes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 277, + 287, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 277, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 100, + 277, + 287, + 290 + ], + "type": "interline_equation", + "content": "G = \\left| \\left(J _ {M} \\left(I _ {P}\\right) - J _ {C} \\left(I _ {P}\\right)\\right) \\nabla_ {I _ {C}} L \\right|, \\tag {6}", + "image_path": "d97321bd2dbdf5650cabe8c9ddb87b7bafd210ad3db4ac15564c1fed5f886209.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 298, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 287, + 327 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 55, + 298, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\nabla_{I_C}^{(j)}L" + }, + { + "bbox": [ + 55, + 298, + 287, + 327 + ], + "type": "text", + "content": " is necessary for the networks to train, " + }, + { + "bbox": [ + 55, + 298, + 287, + 327 + ], + "type": "inline_equation", + "content": "J_M^{(i,j)}(I_P) - J_C^{(i,j)}(I_P)" + }, + { + "bbox": [ + 55, + 298, + 287, + 327 + ], + "type": "text", + "content": " should be close to zero. Thus," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 337, + 287, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 337, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 59, + 337, + 287, + 378 + ], + "type": "interline_equation", + "content": "\\forall i, j \\in [ 1, N ], \\left| \\frac {\\partial I _ {M} ^ {(i)}}{\\partial I _ {P} ^ {(j)}} - \\frac {\\partial I _ {C} ^ {(i)}}{\\partial I _ {P} ^ {(j)}} \\right| = \\left| \\frac {\\partial}{\\partial I _ {P} ^ {(j)}} \\left(I _ {M} ^ {(i)} - I _ {C} ^ {(i)}\\right) \\right| \\tag {7}", + "image_path": "dc01146695a9b6fecf0af3d4ed37f782002716e0bc797b8423996a9dc63fb9e1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 380, + 287, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 287, + 463 + ], + "type": "text", + "content": "should be minimized. The above analysis can be extended to the rate component. Eq. (7) shows that accurate gradients can be backpropagated to the preprocessor if the codec model's output is close to the target codec's output and the input to the postprocessor is from the target video codec. In the proposed training regime, the input to the postprocessor is therefore set as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 464, + 287, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 464, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 118, + 464, + 287, + 479 + ], + "type": "interline_equation", + "content": "I _ {O _ {\\text {i n}}} = \\overline {{\\left(I _ {C} - I _ {M}\\right)}} + I _ {M}, \\tag {8}", + "image_path": "58a607ffc54ecdac2af4c10b994cd14eb196fe1898d1c5e2afb24ae531ef73fe.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 486, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 287, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\overline{a}" + }, + { + "bbox": [ + 55, + 486, + 287, + 545 + ], + "type": "text", + "content": " indicates the stop-gradient operator [63] applied to the tensor " + }, + { + "bbox": [ + 55, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 486, + 287, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 486, + 287, + 545 + ], + "type": "inline_equation", + "content": "I_{C}" + }, + { + "bbox": [ + 55, + 486, + 287, + 545 + ], + "type": "text", + "content": " is the output of the non-differentiable video codec. Eq. (8) allows frames generated by the target codec to be passed to the postprocessor while using the codec model for backpropagation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 546, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 287, + 641 + ], + "type": "text", + "content": "In summary, the proposed training procedure has two main steps; iterative update of the codec model and the end-to-end update of the pre- and postprocessors. For the end-to-end update, a stop-gradient-based approach is proposed. Simple CNNs are used for pre and postprocessing. Details of the pre and postprocessor architecture are in the supplementary. The end-to-end training pipeline is shown in Algorithm 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 651, + 181, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 181, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 181, + 664 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 671, + 119, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 119, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 119, + 684 + ], + "type": "text", + "content": "4.1. Training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 287, + 713 + ], + "type": "text", + "content": "The Vimeo-90k dataset [70] was used to train the codec model and perform end-to-end training. The training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "Algorithm 1 End-to-end training pipeline of the pre and postprocessor" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 99, + 455, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 99, + 455, + 111 + ], + "spans": [ + { + "bbox": [ + 306, + 99, + 455, + 111 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 306, + 99, + 455, + 111 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{in}}" + }, + { + "bbox": [ + 306, + 99, + 455, + 111 + ], + "type": "text", + "content": " , batch of video frames" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 112, + 492, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 112, + 492, + 123 + ], + "spans": [ + { + "bbox": [ + 307, + 112, + 492, + 123 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 307, + 112, + 492, + 123 + ], + "type": "inline_equation", + "content": "\\theta_{M}" + }, + { + "bbox": [ + 307, + 112, + 492, + 123 + ], + "type": "text", + "content": ", parameters of the codec model" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 124, + 490, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 124, + 490, + 135 + ], + "spans": [ + { + "bbox": [ + 307, + 124, + 490, + 135 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 307, + 124, + 490, + 135 + ], + "type": "inline_equation", + "content": "\\theta_{P}" + }, + { + "bbox": [ + 307, + 124, + 490, + 135 + ], + "type": "text", + "content": ", parameters of the preprocessor" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 136, + 494, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 136, + 494, + 147 + ], + "spans": [ + { + "bbox": [ + 307, + 136, + 494, + 147 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 307, + 136, + 494, + 147 + ], + "type": "inline_equation", + "content": "\\theta_{O}" + }, + { + "bbox": [ + 307, + 136, + 494, + 147 + ], + "type": "text", + "content": ", parameters of the postprocessor" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 148, + 466, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 466, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 466, + 159 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 307, + 148, + 466, + 159 + ], + "type": "inline_equation", + "content": "\\alpha_{L}" + }, + { + "bbox": [ + 307, + 148, + 466, + 159 + ], + "type": "text", + "content": ", learning rate for training" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 160, + 372, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 160, + 372, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 160, + 372, + 171 + ], + "type": "text", + "content": "for all " + }, + { + "bbox": [ + 317, + 160, + 372, + 171 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{in}}" + }, + { + "bbox": [ + 317, + 160, + 372, + 171 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 332, + 171, + 388, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 171, + 388, + 183 + ], + "spans": [ + { + "bbox": [ + 332, + 171, + 388, + 183 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {P} \\leftarrow P (\\boldsymbol {I} _ {\\text {i n}})", + "image_path": "e73f5227869c2c6b13ca5f9cbacceb5ab50fd133ddf44e05c85b970cc159d598.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 332, + 184, + 415, + 195 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 184, + 415, + 195 + ], + "spans": [ + { + "bbox": [ + 332, + 184, + 415, + 195 + ], + "type": "interline_equation", + "content": "\\left(\\boldsymbol {I} _ {C}, \\boldsymbol {R} _ {C}\\right) \\leftarrow C \\left(\\boldsymbol {I} _ {P}\\right)", + "image_path": "b6ba7db16665e9d4f8daf753bc159fe56252af2ce3e993f964bfedbd1128919b.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 332, + 196, + 422, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 196, + 422, + 207 + ], + "spans": [ + { + "bbox": [ + 332, + 196, + 422, + 207 + ], + "type": "interline_equation", + "content": "\\left(\\boldsymbol {I} _ {M}, \\boldsymbol {R} _ {M}\\right) \\leftarrow M \\left(\\boldsymbol {I} _ {P}\\right)", + "image_path": "541439eeefd1f45202f2e91f55c340645af9bdcef1a28d1f72f5c8a6cd516cb9.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 332, + 208, + 489, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 208, + 489, + 221 + ], + "spans": [ + { + "bbox": [ + 332, + 208, + 489, + 221 + ], + "type": "text", + "content": "Get " + }, + { + "bbox": [ + 332, + 208, + 489, + 221 + ], + "type": "inline_equation", + "content": "D_M^{hyp}" + }, + { + "bbox": [ + 332, + 208, + 489, + 221 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 332, + 208, + 489, + 221 + ], + "type": "inline_equation", + "content": "\\theta_g\\gets \\theta_g - \\alpha_L\\nabla_{\\theta_g}D_M^{hyp}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 332, + 222, + 488, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 222, + 488, + 235 + ], + "spans": [ + { + "bbox": [ + 332, + 222, + 488, + 235 + ], + "type": "text", + "content": "Get " + }, + { + "bbox": [ + 332, + 222, + 488, + 235 + ], + "type": "inline_equation", + "content": "R_M^{hyp}" + }, + { + "bbox": [ + 332, + 222, + 488, + 235 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 332, + 222, + 488, + 235 + ], + "type": "inline_equation", + "content": "\\theta_h\\gets \\theta_h - \\alpha_L\\nabla_{\\theta_h}R_M^{hyp}" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 332, + 235, + 537, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 235, + 537, + 249 + ], + "spans": [ + { + "bbox": [ + 332, + 235, + 537, + 249 + ], + "type": "text", + "content": "Get " + }, + { + "bbox": [ + 332, + 235, + 537, + 249 + ], + "type": "inline_equation", + "content": "L_{M}^{(\\alpha ,\\beta)}" + }, + { + "bbox": [ + 332, + 235, + 537, + 249 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 332, + 235, + 537, + 249 + ], + "type": "inline_equation", + "content": "(\\alpha ,\\beta)\\gets (\\alpha ,\\beta) - \\alpha_{L}\\nabla_{(\\alpha ,\\beta)}L_{M}^{(\\alpha ,\\beta)}" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 332, + 250, + 444, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 250, + 444, + 260 + ], + "spans": [ + { + "bbox": [ + 332, + 250, + 444, + 260 + ], + "type": "text", + "content": "Generate " + }, + { + "bbox": [ + 332, + 250, + 444, + 260 + ], + "type": "inline_equation", + "content": "I_{O_{\\mathrm{in}}}" + }, + { + "bbox": [ + 332, + 250, + 444, + 260 + ], + "type": "text", + "content": " using Eq. (8)" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 332, + 261, + 399, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 261, + 399, + 273 + ], + "spans": [ + { + "bbox": [ + 332, + 261, + 399, + 273 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{\\mathrm{out}}\\gets O(\\pmb{I}_{O_{\\mathrm{in}}})" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 332, + 274, + 411, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 274, + 411, + 284 + ], + "spans": [ + { + "bbox": [ + 332, + 274, + 411, + 284 + ], + "type": "text", + "content": "Get " + }, + { + "bbox": [ + 332, + 274, + 411, + 284 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 332, + 274, + 411, + 284 + ], + "type": "text", + "content": " using Eq. (4)" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 332, + 285, + 421, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 285, + 421, + 297 + ], + "spans": [ + { + "bbox": [ + 332, + 285, + 421, + 297 + ], + "type": "interline_equation", + "content": "\\theta_ {P} \\leftarrow \\theta_ {P} - \\alpha_ {L} \\nabla_ {\\theta_ {P}} L", + "image_path": "03c7c57959a1e3c5d6c2b28ced54cab92bc2ec3c242576129ded11c057576278.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 332, + 297, + 421, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 297, + 421, + 309 + ], + "spans": [ + { + "bbox": [ + 332, + 297, + 421, + 309 + ], + "type": "interline_equation", + "content": "\\theta_ {O} \\leftarrow \\theta_ {O} - \\alpha_ {L} \\nabla_ {\\theta_ {O}} L", + "image_path": "5b7ad7c71526068c0542d93ae39246604f96af2982a6ff4fc62c7723f1b58e07.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 310, + 349, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 349, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 349, + 319 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "text", + "content": "pipeline and the pre- and postprocessor are implemented in the YUV colorspace. A fixed learning rate of " + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "text", + "content": " is used for each network with the rest of hyper-parameters set as: " + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{L1}} = 0.1" + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SSIM}} = 0.25" + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MSSSIM}} = 0.25" + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{DLM}} = 0.2" + }, + { + "bbox": [ + 313, + 341, + 545, + 628 + ], + "type": "text", + "content": ". Batch size, GOP and maximum iterations are set to 4, 4 and 500000, respectively. Both Phase 1 and 2 of pretraining are performed for 500000 iterations. The model is trained on source frames only in an open-loop configuration; i.e., no reconstructed frames are fed into the preprocessor as input. A low-delay recipe of SVT-AV1 (described in supplementary) is used for codec model alignment for pre-training and end-to-end training. The codec model encoding is aligned to the GOP of the target codec, such that intra and inter coding are also aligned. For experiments on VVC, it was found that codec model alignment to a low-delay recipe of VVenC provided slight BD-rate gains over SVT-AV1 alignment, albeit at high computational overhead during training. Therefore, the presented results are carried out with a single neural wrapper model (one encode) developed based on codec proxy model alignment to SVT-AV1. This is practically more feasible than [24], which requires 9 combinations of models (9 encodes) + convex hull optimization and is thus far more computationally demanding." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 314, + 635, + 430, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 635, + 430, + 646 + ], + "spans": [ + { + "bbox": [ + 314, + 635, + 430, + 646 + ], + "type": "text", + "content": "4.2. Evaluation Protocol" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 545, + 713 + ], + "type": "text", + "content": "Four 1080p video datasets were used for evaluation: (i) 16 gaming sequences captured in a commercial setup (Gaming); (ii) The 16 XIPH natural sequences used by recent work [11] (XIPH); (iii) The 1080p sequences of the UVG [48] dataset (UVG); (iv) HEVC-B [8] dataset" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17747" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 289, + 203 + ], + "type": "text", + "content": "(HEVC-B). The proposed method was applied with SVT-AV1 [32] for AV1 and VVenC [10] for VVC. All experiments were performed with YUV 4:2:0 8-bit sequences. In all cases, the protocol of Li et al. [35] is followed and 96 frames are evaluated. However, unlike Li [35], both standard encoders are used without their low-delay limitation, in order to attain their best performance (see supplementary for details on sequences and encoding recipes). Ablation results are with the SVT-AV1 lowdelay recipe (see supplementary). Results in Tab. 3 and 4 are on the Gaming dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 204, + 289, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 204, + 289, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 289, + 419 + ], + "type": "text", + "content": "In order to cover a wide range of well-established objective quality scoring methods, the following scores are used: SSIM, MS-SSIM, VMAF [42, 52] and the AVQT score from Apple [4]. In order to showcase the generality of the proposed approach, the same pre- and postprocessor models are used for: (i) all the points on the RD curves, (ii) all quality scores and (iii) both standard codecs. Perceptual models are used for competing methods when available. Comparison is provided against a competitive video preprocessor (DPP [11]), SOTA neural codecs (DCVC-DC [35] and DCVC-FM [36]), and perceptual modes of standard codecs. The comparisons include results of end-to-end training with the codec model of Hu et al. [24] with the proposed network architectures and losses used to illustrate the effectiveness of the proposed codec model. For DCVC-DC and DCVC-FM, the best available pretrained models (publicly provided by the authors) are used." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 427, + 218, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 218, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 218, + 440 + ], + "type": "text", + "content": "4.3. Objective Quality Assessment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 289, + 685 + ], + "type": "text", + "content": "The quantitative results of the proposed method are shown in† Fig. 2, and Tab. 1 and Tab. 2. These results show that the proposed approach provides for BD-rates between " + }, + { + "bbox": [ + 55, + 445, + 289, + 685 + ], + "type": "inline_equation", + "content": "-5.4\\%" + }, + { + "bbox": [ + 55, + 445, + 289, + 685 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 445, + 289, + 685 + ], + "type": "inline_equation", + "content": "-30\\%" + }, + { + "bbox": [ + 55, + 445, + 289, + 685 + ], + "type": "text", + "content": " in comparison to the underlying standard encoders. BD-rate savings tend to be higher for quality scores that are more perceptually-oriented. Compared to competing methods, the proposed approach is shown to offer significantly more consistent BD-rate improvements over all quality scoring methods (ALL columns of Tab. 1 and Tab. 2). Importantly, as shown in Fig. 2, unlike DCVC-DC that is only covering a very limited bitrate-quality range at the very top end, the proposed approach outperforms all other methods on the entire \"active\" region of bitrate-quality, i.e., VMAF of 40-93 [32], which encapsulates the entirety of commercially-relevant bitrates for 1080p video streaming. Given that the neural wrapper is optimizing perceptually, BD-rates for PSNR tend to show regression on PSNR (see supplementary) but visual inspection and an array of perceptual metrics indicate that the proposed approach offers significant benefits" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 315, + 78, + 419, + 148 + ], + "blocks": [ + { + "bbox": [ + 315, + 78, + 419, + 148 + ], + "lines": [ + { + "bbox": [ + 315, + 78, + 419, + 148 + ], + "spans": [ + { + "bbox": [ + 315, + 78, + 419, + 148 + ], + "type": "image", + "image_path": "65c01de4bcda9262aea18f8548bb9701ee71161eac47d04e9f811fff78595cb2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 156, + 372, + 164 + ], + "lines": [ + { + "bbox": [ + 361, + 156, + 372, + 164 + ], + "spans": [ + { + "bbox": [ + 361, + 156, + 372, + 164 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 437, + 78, + 537, + 145 + ], + "blocks": [ + { + "bbox": [ + 437, + 78, + 537, + 145 + ], + "lines": [ + { + "bbox": [ + 437, + 78, + 537, + 145 + ], + "spans": [ + { + "bbox": [ + 437, + 78, + 537, + 145 + ], + "type": "image", + "image_path": "825fde789f275700e39dad52a643630271375a0f8adf42751298eb814e2a7bc2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 155, + 492, + 164 + ], + "lines": [ + { + "bbox": [ + 481, + 155, + 492, + 164 + ], + "spans": [ + { + "bbox": [ + 481, + 155, + 492, + 164 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 173, + 419, + 239 + ], + "blocks": [ + { + "bbox": [ + 317, + 173, + 419, + 239 + ], + "lines": [ + { + "bbox": [ + 317, + 173, + 419, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 419, + 239 + ], + "type": "image", + "image_path": "43c24bb04a97d86178a09b4081705d4911f79f0bee25609b94c8a00cfcb76007.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 250, + 372, + 258 + ], + "lines": [ + { + "bbox": [ + 361, + 250, + 372, + 258 + ], + "spans": [ + { + "bbox": [ + 361, + 250, + 372, + 258 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 440, + 175, + 537, + 238 + ], + "blocks": [ + { + "bbox": [ + 440, + 175, + 537, + 238 + ], + "lines": [ + { + "bbox": [ + 440, + 175, + 537, + 238 + ], + "spans": [ + { + "bbox": [ + 440, + 175, + 537, + 238 + ], + "type": "image", + "image_path": "f8d5f43a991db678e7df6cf5bb05450de3f53f824f28b7d5d3882680939bad64.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 249, + 492, + 258 + ], + "lines": [ + { + "bbox": [ + 481, + 249, + 492, + 258 + ], + "spans": [ + { + "bbox": [ + 481, + 249, + 492, + 258 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 267, + 417, + 334 + ], + "blocks": [ + { + "bbox": [ + 315, + 267, + 417, + 334 + ], + "lines": [ + { + "bbox": [ + 315, + 267, + 417, + 334 + ], + "spans": [ + { + "bbox": [ + 315, + 267, + 417, + 334 + ], + "type": "image", + "image_path": "4f0e2c9981762f202e9c52ed2a4c25a5271e4cd7ffa32b0f54375e40c94ceeec.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 345, + 371, + 353 + ], + "lines": [ + { + "bbox": [ + 361, + 345, + 371, + 353 + ], + "spans": [ + { + "bbox": [ + 361, + 345, + 371, + 353 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 436, + 267, + 536, + 334 + ], + "blocks": [ + { + "bbox": [ + 436, + 267, + 536, + 334 + ], + "lines": [ + { + "bbox": [ + 436, + 267, + 536, + 334 + ], + "spans": [ + { + "bbox": [ + 436, + 267, + 536, + 334 + ], + "type": "image", + "image_path": "233c71f97da0b0a3afc2113377f60647550034c1e7e0e20ed986a0a6772857fd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 344, + 492, + 354 + ], + "lines": [ + { + "bbox": [ + 481, + 344, + 492, + 354 + ], + "spans": [ + { + "bbox": [ + 481, + 344, + 492, + 354 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 315, + 361, + 417, + 429 + ], + "blocks": [ + { + "bbox": [ + 315, + 361, + 417, + 429 + ], + "lines": [ + { + "bbox": [ + 315, + 361, + 417, + 429 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 417, + 429 + ], + "type": "image", + "image_path": "96cce1064cfc1d05ecaf8517cad37f29915ae71b8eea34d9ad0113df58a1ebda.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 438, + 372, + 448 + ], + "lines": [ + { + "bbox": [ + 361, + 438, + 372, + 448 + ], + "spans": [ + { + "bbox": [ + 361, + 438, + 372, + 448 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 436, + 361, + 536, + 429 + ], + "blocks": [ + { + "bbox": [ + 436, + 361, + 536, + 429 + ], + "lines": [ + { + "bbox": [ + 436, + 361, + 536, + 429 + ], + "spans": [ + { + "bbox": [ + 436, + 361, + 536, + 429 + ], + "type": "image", + "image_path": "0f30839fc121e9fefc8419dcd74925d00fceb6b91eee5dba5dbe9d677692fc19.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 438, + 492, + 447 + ], + "lines": [ + { + "bbox": [ + 481, + 438, + 492, + 447 + ], + "spans": [ + { + "bbox": [ + 481, + 438, + 492, + 447 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 457, + 545, + 491 + ], + "lines": [ + { + "bbox": [ + 306, + 457, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 545, + 491 + ], + "type": "text", + "content": "Figure 2. Combined rate-quality plots over Gaming ((a), (b)), XIPH ((c), (d)), UVG ((e), (f)) and HEVC-B ((g), (h)) datasets. See supplementary for more results." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 516, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 545, + 552 + ], + "type": "text", + "content": "in visual quality. Nonetheless, the results of DCVC-DC and DCVC-FM are remarkably good given that they are fully neural approaches." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 563, + 480, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 480, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 480, + 576 + ], + "type": "text", + "content": "4.4. Subjective Quality Assessment" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 547, + 713 + ], + "type": "text", + "content": "To assess the visual quality of the proposed approach, a 5-scale absolute categorical rating (ACR) test with hidden reference was conducted under the ITU-T P.910 test conditions and SUREAL post-processing [40, 43] (see supplementary for details). Raters were shown videos generated by the codec and proposed method. Separate tests were conducted for AV1 and VVC with each test comprising 72 sequences. The combined rate-MOS plots are shown in Fig. 3. The BD-rates reported in the figure for the proposed approach are mostly in line with AVQT and VMAF BD-rates reported for Gaming in Tab. 1 and" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "type": "text", + "content": " All combined plots of Fig. 2 and Fig. 3 are produced with the slope-based integration method of Wu et al. [68] and are shown in log-scale." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17748" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 112, + 284, + 343 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "type": "text", + "content": "Table 1. BD-rates vs. AV1 (more negative is better). AV1+SSIM denotes codec tuned for SSIM (-tune ssim). ALL indicates average BD-rate over all quality scores." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 112, + 284, + 343 + ], + "lines": [ + { + "bbox": [ + 51, + 112, + 284, + 343 + ], + "spans": [ + { + "bbox": [ + 51, + 112, + 284, + 343 + ], + "type": "table", + "html": "
DatasetMethodSSIMMSSSIMAVQTVMAFALL
GamingAV1+SSIM-4.5-5.4-5.12.1-3.2
AV1+DPP [11]-5.2-6.1-17.6-11.3-10.0
AV1+[24]14.96.312.2-1.48.0
DCVC-DC [35]-15.4-36.1-35.621.84-16.3
DCVC-FM [36]14.86.04.2-2.95.5
AV1+Proposed-15.5-13.8-19.6-30.0-19.7
XIPHAV1+SSIM-3.5-4.0-5.43.2-2.4
AV1+DPP [11]-4.3-3.9-9.0-16.7-8.5
AV1+[24]12.96.14.51.76.3
DCVC-DC [35]-18.4-25.512.032.20.3
DCVC-FM [36]17.69.08.23.59.6
AV1+Proposed-10.9-7.1-28.8-30.1-19.2
UVGAV1+SSIM-2.8-3.1-2.82.9-1.4
AV1+DPP [11]-3.8-3.0-9.2-21.7-9.4
AV1+[24]9.14.43.33.25.0
DCVC-DC [35]-20.8-10.24.218.2-2.1
DCVC-FM [36]20.49.69.89.412.3
AV1+Proposed-16.2-9.7-22.2-31.9-20.0
HEVC-BAV1+SSIM-5.4-5.3-4.72.9-3.1
AV1+DPP [11]-8.7-6.9-5.7-20.1-10.3
AV1+[24]2.61.8-3.11.10.6
DCVC-DC [35]-35.3-29.1-11.637.2-9.7
DCVC-FM [36]-6.5-11.7-7.7-5.8-7.9
AV1+Proposed-14.5-13.4-15.1-41.8-21.2
", + "image_path": "db6c88e3ee2e7646b74d56d07076f76b4cd6b9d4f2d41e50e3288b7a554cb0c3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 52, + 399, + 284, + 629 + ], + "blocks": [ + { + "bbox": [ + 47, + 357, + 287, + 392 + ], + "lines": [ + { + "bbox": [ + 47, + 357, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 287, + 392 + ], + "type": "text", + "content": "Table 2. BD-rates vs. VVC (more negative is better). " + }, + { + "bbox": [ + 47, + 357, + 287, + 392 + ], + "type": "inline_equation", + "content": "\\mathrm{VVC + QPA}" + }, + { + "bbox": [ + 47, + 357, + 287, + 392 + ], + "type": "text", + "content": " indicates quantization parameter adaptation is enabled. ALL indicates average BD-rate over all quality scores." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 399, + 284, + 629 + ], + "lines": [ + { + "bbox": [ + 52, + 399, + 284, + 629 + ], + "spans": [ + { + "bbox": [ + 52, + 399, + 284, + 629 + ], + "type": "table", + "html": "
DatasetMethodSSIMMSSSIMAVQTVMAFALL
GamingVVC+QPA0.44.613.41.54.9
VVC+DPP [11]-1.3-1.6-9.2-9.0-5.3
VVC+[24]39.532.311.47.822.7
DCVC-DC [35]9.1-11.5-21.338.93.8
DCVC-FM [36]4.918.410.918.413.1
VVC+Proposed-16.6-8.7-19.0-14.3-14.6
XIPHVVC+QPA-0.92.21.12.71.3
VVC+DPP [11]-2.3-1.6-10.5-10.4-6.2
VVC+[24]29.424.412.29.518.9
DCVC-DC [35]-8.1-12.09.241.17.5
DCVC-FM [36]4.313.10.516.88.7
VVC+Proposed-11.4-5.4-16.5-18.0-12.8
UVGVVC+QPA-0.11.67.71.52.7
VVC+DPP [11]-1.4-1.1-7.0-11.4-5.2
VVC+[24]26.823.57.96.516.2
DCVC-DC [35]-8.216.0-1.429.18.9
DCVC-FM [36]-2.714.112.514.79.65
VVC+Proposed-11.8-5.4-13.9-13.8-11.2
HEVC-BVVC+QPA0.91.21.12.11.3
VVC+DPP [11]-19.2-16.7-17.2-16.7-17.4
VVC+[24]0.50.50.1-2.3-0.3
DCVC-DC [35]-17.0-5.56.349.98.4
DCVC-FM [36]18.815.814.412.115.3
VVC+Proposed-31.8-27.4-25.2-33.3-29.4
", + "image_path": "47f51655552e5af1cd24c295ab9acaa94689ca477f77e2db7231438d503f6e1e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 287, + 715 + ], + "lines": [ + { + "bbox": [ + 55, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 287, + 715 + ], + "type": "text", + "content": "Tab. 2. Sample visual results of the proposed method are shown in Fig. 4. The results show that the proposed approach is temporally coherent, and better capable of preserving structure, retaining textures and denoising codec artifacts, also reflected in the MOS scores of Fig. 3." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 79, + 416, + 155 + ], + "blocks": [ + { + "bbox": [ + 315, + 79, + 416, + 155 + ], + "lines": [ + { + "bbox": [ + 315, + 79, + 416, + 155 + ], + "spans": [ + { + "bbox": [ + 315, + 79, + 416, + 155 + ], + "type": "image", + "image_path": "788833425de26784a381ec26c10bb7cbc270fdb557b8716396bd3923cdd4c25c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 159, + 374, + 167 + ], + "lines": [ + { + "bbox": [ + 362, + 159, + 374, + 167 + ], + "spans": [ + { + "bbox": [ + 362, + 159, + 374, + 167 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 430, + 79, + 531, + 155 + ], + "blocks": [ + { + "bbox": [ + 430, + 79, + 531, + 155 + ], + "lines": [ + { + "bbox": [ + 430, + 79, + 531, + 155 + ], + "spans": [ + { + "bbox": [ + 430, + 79, + 531, + 155 + ], + "type": "image", + "image_path": "327424f0839c08d2ce220f686a7a9981316a3a4606eb3f64d034798b920adb30.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 158, + 490, + 167 + ], + "lines": [ + { + "bbox": [ + 479, + 158, + 490, + 167 + ], + "spans": [ + { + "bbox": [ + 479, + 158, + 490, + 167 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 177, + 527, + 189 + ], + "lines": [ + { + "bbox": [ + 323, + 177, + 527, + 189 + ], + "spans": [ + { + "bbox": [ + 323, + 177, + 527, + 189 + ], + "type": "text", + "content": "Figure 3. Combined rate-quality plots with P.910 MOS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 343, + 220, + 511, + 269 + ], + "blocks": [ + { + "bbox": [ + 322, + 201, + 529, + 213 + ], + "lines": [ + { + "bbox": [ + 322, + 201, + 529, + 213 + ], + "spans": [ + { + "bbox": [ + 322, + 201, + 529, + 213 + ], + "type": "text", + "content": "Table 3. BD-rates of different component configurations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 220, + 511, + 269 + ], + "lines": [ + { + "bbox": [ + 343, + 220, + 511, + 269 + ], + "spans": [ + { + "bbox": [ + 343, + 220, + 511, + 269 + ], + "type": "table", + "html": "
PrePostSSIMMSSSIMAVQTVMAF
✓✓-25.6-20.8-21.4-20.5
✓✗17.913.229.8-34.1
✗✓-2.90.44.1-33.3
", + "image_path": "4dafdb6f78637341ba9d77004af6180748e8c668f80abc79f0fbf7627f74c1e0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 288, + 378, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 288, + 378, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 288, + 378, + 300 + ], + "type": "text", + "content": "4.5. Ablation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 307, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 545, + 450 + ], + "type": "text", + "content": "Two-phase pretraining: The rate-quality alignment between target codec " + }, + { + "bbox": [ + 313, + 307, + 545, + 450 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 307, + 545, + 450 + ], + "type": "text", + "content": " and model " + }, + { + "bbox": [ + 313, + 307, + 545, + 450 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 307, + 545, + 450 + ], + "type": "text", + "content": " after two-phase pretraining is validated by comparing their performance on a dataset ov dieo sequences. The Pearson correlation coefficient (PCC) is improved from 0.56 and 0.60 to 0.88 and 0.85 for bitrate and PSNR, respectively (see supplementary for more results). Note that the PCC for rate is computed between " + }, + { + "bbox": [ + 313, + 307, + 545, + 450 + ], + "type": "inline_equation", + "content": "R_{C}" + }, + { + "bbox": [ + 313, + 307, + 545, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 307, + 545, + 450 + ], + "type": "inline_equation", + "content": "R_{M}^{hyp}" + }, + { + "bbox": [ + 313, + 307, + 545, + 450 + ], + "type": "text", + "content": ". This indicates that both distortion and bitrate of the codec model and target codec correlate better after the two-phase pretraining, which is a sufficient condition for gradient-based optimization [31]. Tab. 4 shows the effect of pretraining on wrappers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 451, + 546, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 546, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 546, + 666 + ], + "type": "text", + "content": "Pre/post processing: Tab. 3 shows the effect of removing either the pre or the postprocessor from the overall pipeline under the same training pipeline. Removing the preprocessor is equivalent to training a video denoiser. While BD-rate gains for VMAF are maintained with only the pre- and postprocessor, this can be attributed to sharpening effects over the source content, which VMAF has susceptibility to [1]. Indeed, with more naive content enhancement such as sharpening, fidelity oriented methods such as SSIM and MS-SSIM will experience BD-rate losses, which can be seen when training both pre and postprocessors alone. Only jointly training both pre- and postprocessors is able to achieve substantial BD-rate gains across all quality scores. Fig. 5 provides a generic example where the preprocessor embeds information into the source frame which does not significantly increase the bitrate and can survive the codec compression. This information is used by the postprocessor for reconstruction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 666, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 547, + 714 + ], + "type": "text", + "content": "End-to-end training: For end-to-end training, an ablation is performed over the proposed iterative updates of the codec model and stop-gradient during training. The results are provided in Tab. 4 under different training con" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17749" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 71, + 236, + 215 + ], + "blocks": [ + { + "bbox": [ + 104, + 71, + 236, + 215 + ], + "lines": [ + { + "bbox": [ + 104, + 71, + 236, + 215 + ], + "spans": [ + { + "bbox": [ + 104, + 71, + 236, + 215 + ], + "type": "image", + "image_path": "d889a979e5edd61335cb6f939923f6830fd09395e6f79308859539dfa3a01874.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 72, + 361, + 215 + ], + "blocks": [ + { + "bbox": [ + 238, + 72, + 361, + 215 + ], + "lines": [ + { + "bbox": [ + 238, + 72, + 361, + 215 + ], + "spans": [ + { + "bbox": [ + 238, + 72, + 361, + 215 + ], + "type": "image", + "image_path": "84ae54ff76a05b71383ad094cbc2ec1f3ce537f2c7e4b5a029f958edc50a3fdf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 362, + 72, + 487, + 215 + ], + "blocks": [ + { + "bbox": [ + 362, + 72, + 487, + 215 + ], + "lines": [ + { + "bbox": [ + 362, + 72, + 487, + 215 + ], + "spans": [ + { + "bbox": [ + 362, + 72, + 487, + 215 + ], + "type": "image", + "image_path": "d92019e321f35b83c55cee1a9eb93b8a6a93a8a00b2e463ecd6012006977ff9b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 103, + 217, + 236, + 350 + ], + "blocks": [ + { + "bbox": [ + 103, + 217, + 236, + 350 + ], + "lines": [ + { + "bbox": [ + 103, + 217, + 236, + 350 + ], + "spans": [ + { + "bbox": [ + 103, + 217, + 236, + 350 + ], + "type": "image", + "image_path": "3ec950dadcabe2aadc1bc52cbd9f2ff865457c28823754c4e623fb111d436114.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 361, + 499, + 374 + ], + "lines": [ + { + "bbox": [ + 94, + 361, + 499, + 374 + ], + "spans": [ + { + "bbox": [ + 94, + 361, + 499, + 374 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of codecs with and without the proposed method. Zoom in for better viewing." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 238, + 217, + 361, + 350 + ], + "blocks": [ + { + "bbox": [ + 238, + 217, + 361, + 350 + ], + "lines": [ + { + "bbox": [ + 238, + 217, + 361, + 350 + ], + "spans": [ + { + "bbox": [ + 238, + 217, + 361, + 350 + ], + "type": "image", + "image_path": "d20947a0105eec2526383e64108f069158ea431b02dbefcb8a441c80d8282472.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 217, + 487, + 350 + ], + "blocks": [ + { + "bbox": [ + 362, + 217, + 487, + 350 + ], + "lines": [ + { + "bbox": [ + 362, + 217, + 487, + 350 + ], + "spans": [ + { + "bbox": [ + 362, + 217, + 487, + 350 + ], + "type": "image", + "image_path": "5e7ee760f98682c083c332657d678ccc8a76d86523edf90af550732cbd5d36ff.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 74, + 388, + 184, + 452 + ], + "blocks": [ + { + "bbox": [ + 74, + 388, + 184, + 452 + ], + "lines": [ + { + "bbox": [ + 74, + 388, + 184, + 452 + ], + "spans": [ + { + "bbox": [ + 74, + 388, + 184, + 452 + ], + "type": "image", + "image_path": "e67df1c3a24f8795c9c7c565baab1cee6becd7f67fa8d4e7e59ae7ef962946a0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 454, + 130, + 460 + ], + "lines": [ + { + "bbox": [ + 124, + 454, + 130, + 460 + ], + "spans": [ + { + "bbox": [ + 124, + 454, + 130, + 460 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 186, + 389, + 295, + 452 + ], + "blocks": [ + { + "bbox": [ + 186, + 389, + 295, + 452 + ], + "lines": [ + { + "bbox": [ + 186, + 389, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 186, + 389, + 295, + 452 + ], + "type": "image", + "image_path": "ef51856a557b7a713cbbb07c1619de990484ba21683b93548de4e6da2deb29a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 454, + 242, + 460 + ], + "lines": [ + { + "bbox": [ + 235, + 454, + 242, + 460 + ], + "spans": [ + { + "bbox": [ + 235, + 454, + 242, + 460 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 297, + 389, + 406, + 452 + ], + "blocks": [ + { + "bbox": [ + 297, + 389, + 406, + 452 + ], + "lines": [ + { + "bbox": [ + 297, + 389, + 406, + 452 + ], + "spans": [ + { + "bbox": [ + 297, + 389, + 406, + 452 + ], + "type": "image", + "image_path": "efe18297648ee5ee11480c7b745083ecb554794f7cb5cddb6353dd0d71cc0115.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 454, + 354, + 460 + ], + "lines": [ + { + "bbox": [ + 347, + 454, + 354, + 460 + ], + "spans": [ + { + "bbox": [ + 347, + 454, + 354, + 460 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 471, + 547, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 471, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 547, + 495 + ], + "type": "text", + "content": "Figure 5. Results of components of the pipeline where (a), (b), (c) and (d) are the source frame, preprocessor output, codec output and postprocessor output, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 408, + 389, + 518, + 452 + ], + "blocks": [ + { + "bbox": [ + 408, + 389, + 518, + 452 + ], + "lines": [ + { + "bbox": [ + 408, + 389, + 518, + 452 + ], + "spans": [ + { + "bbox": [ + 408, + 389, + 518, + 452 + ], + "type": "image", + "image_path": "a8cf22dddd057e91f0608f2464b208c62c31737ff5e39e74e999214feb11268d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 454, + 465, + 460 + ], + "lines": [ + { + "bbox": [ + 458, + 454, + 465, + 460 + ], + "spans": [ + { + "bbox": [ + 458, + 454, + 465, + 460 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 53, + 533, + 281, + 586 + ], + "blocks": [ + { + "bbox": [ + 64, + 513, + 269, + 525 + ], + "lines": [ + { + "bbox": [ + 64, + 513, + 269, + 525 + ], + "spans": [ + { + "bbox": [ + 64, + 513, + 269, + 525 + ], + "type": "text", + "content": "Table 4. BD-rates with different training configurations." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 533, + 281, + 586 + ], + "lines": [ + { + "bbox": [ + 53, + 533, + 281, + 586 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 281, + 586 + ], + "type": "table", + "html": "
PretrainIter. UpdateStop Grad.SSIMMSSSIMAVQTVMAF
✓✓✓-25.6-20.8-21.4-20.5
X✓✓-23.1-14.4-17.3-18.7
✓✓X16.40.0110.5-15.3
✓X✓-11.2-6.5-7.7-13.1
", + "image_path": "af374c48eb3a8112a5ab6f971ca58d1f77f9e0810c01ce209f8600d6c7d4fc87.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 609, + 289, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 289, + 718 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 289, + 718 + ], + "type": "text", + "content": "Figurations. Notably, when training the pre- and postprocessor with only iterative updates, there are BD-rate losses in SSIM and MS-SSIM. This is a by-product of the domain shift between the source and preprocessor output that increases as end-to-end training progresses. When performing end-to-end training with both iterative updates and stop gradient, there is a sufficient alignment constraint on the codec model, which leads to demonstrable BD-rate gains across quality scoring methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 513, + 388, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 513, + 388, + 525 + ], + "spans": [ + { + "bbox": [ + 314, + 513, + 388, + 525 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 533, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 547, + 713 + ], + "type": "text", + "content": "This work proposes a new approach to jointly train neural pre and postprocessors to improve the rate-quality performance of conventional video CODECs. It is proposed to use a neural codec to model the standard codec's sophisticated rate-quality characteristics, combined with: (i) a novel approach to pretrain the codec model; (ii) a new method to train the neural wrapper model jointly with the codec model in the middle. Extensive evaluations with state-of-the-art AV1 and VVC CODECs show that the proposed neural wrapper approach provides for consistent BD-rate improvement over standard CODECs. Unlike neural methods for preprocessing and encoding, the proposed method allows for consistent improvement over all quality scores (SSIM, MSSSIM, AVQT, VMAF, P.910 MOS) over the entire rate-quality region of modern standards." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17750" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] On vmaf's property in the presence of image enhancement operations. https : // docs . google . com / document / d / 1dJczEhX00MZjBSNyKmd3ARiCTdFVMNPBykH4 _ HMPoyY/edit#heading=h.oaikhnw46pw5. Last accessed: 2020-11-16. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 159, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 288, + 224 + ], + "type": "text", + "content": "[2] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8503-8512, 2020. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 226, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 226, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 226, + 288, + 280 + ], + "type": "text", + "content": "[3] Stéphane Alarie, Charles Audet, Aïmen E Gheribi, Michael Kokkolaras, and Sébastien Le Digabel. Two decades of blackbox optimization applications. EURO Journal on Computational Optimization, 9:100011, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 282, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 282, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 287, + 315 + ], + "type": "text", + "content": "[4] Apple. \"Apple avqt presentation\", 2021. https://developer.apple.com/videos/play/wwdc2021/10145.1,6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 317, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 287, + 360 + ], + "type": "text", + "content": "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 362, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 287, + 406 + ], + "type": "text", + "content": "[6] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6228-6237, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 408, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 287, + 451 + ], + "type": "text", + "content": "[7] Mohammad Reza Bonyadi and Zbigniew Michalewicz. Particle swarm optimization for single objective continuous space problems: a review. Evolutionary computation, 25(1):1-54, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 453, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 287, + 475 + ], + "type": "text", + "content": "[8] Frank Bossen et al. Common test conditions and software reference configurations. JCTVC-L1100, 12(7):1, 2013. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 477, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 287, + 520 + ], + "type": "text", + "content": "[9] Eirina Bourtsovlatze, Aaron Chadha, Ilya Fadeev, Vasileios Giotasas, and Yiannis Andreopoulos. Deep video precoding. IEEE Transactions on Circuits and Systems for Video Technology, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 522, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 287, + 576 + ], + "type": "text", + "content": "[10] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 287, + 622 + ], + "type": "text", + "content": "[11] Aaron Chadha and Yiannis Andreopoulos. Deep perceptual preprocessing for video coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14852-14861, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 287, + 678 + ], + "type": "text", + "content": "[12] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 287, + 713 + ], + "type": "text", + "content": "[13] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 97, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 545, + 150 + ], + "type": "text", + "content": "[14] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5962-5971, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 153, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 545, + 207 + ], + "type": "text", + "content": "[15] Jinyoung Choi and Bohyung Han. Task-aware quantization network forJPEG image compression. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 309-324. Springer, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 545, + 262 + ], + "type": "text", + "content": "[16] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. In Proceedings of the AAAI conference on artificial intelligence, pages 10696-10703, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 265, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 265, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 265, + 545, + 319 + ], + "type": "text", + "content": "[17] Dario Fuoli, Shuhang Gu, and Radu Timofte. Efficient video super-resolution through recurrent latent space propagation. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3476-3485. IEEE, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 545, + 364 + ], + "type": "text", + "content": "[18] Ryosuke Furuta, Naoto Inoue, and Toshihiko Yamasaki. Pixelrl: Fully convolutional network with reinforcement learning for image processing. IEEE Transactions on Multimedia, 22(7):1704-1719, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 366, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 545, + 421 + ], + "type": "text", + "content": "[19] Vasileios Giotsas, Nikos Deligiannis, Pam Fisher, and Yiannis Andreopoulos. Perceptual video quality estimation by regression with myopic experts. In 2015 Seventh International Workshop on Quality of Multimedia Experience (QoMEX), pages 1-6. IEEE, 2015. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 545, + 487 + ], + "type": "text", + "content": "[20] Onur G Guleryuz, Philip A Chou, Hugues Hoppe, Danhang Tang, Ruofei Du, Philip Davidson, and Sean Fanello. Sandwiched image compression: wrapping neural networks around a standard codec. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3757-3761. IEEE, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 489, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 545, + 544 + ], + "type": "text", + "content": "[21] Onur G Guleryuz, Philip A Chou, Hugues Hoppe, Danhang Tang, Ruofei Du, Philip Davidson, and Sean Fanello. Sandwiched image compression: Increasing the resolution and dynamic range of standard codec. In 2022 Picture Coding Symposium (PCS), pages 175-179. IEEE, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 545, + 578 + ], + "type": "text", + "content": "[22] Mengxi Guo, Shijie Zhao, Hao Jiang, Junlin Li, and Li Zhang. Video compression with arbitrary rescaling network. arXiv preprint arXiv:2306.04202, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 580, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 545, + 633 + ], + "type": "text", + "content": "[23] Yung-Han Ho, Chih-Peng Chang, Peng-Yu Chen, Alessandro Gnutti, and Wen-Hsiao Peng. Canf-vc: Conditional augmented normalizing flows for video compression. In European Conference on Computer Vision, pages 207-223. Springer, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 545, + 689 + ], + "type": "text", + "content": "[24] Yueyu Hu, Chenhao Zhang, Onur G Guleryuz, Debargha Mukherjee, and Yao Wang. Standard compliant video coding using low complexity, switchable neural wrappers. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1922-1928. IEEE, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 545, + 713 + ], + "type": "text", + "content": "[25] Zhihao Hu and Dong Xu. Complexity-guided slimmable decoder for efficient deep video compression. In Proceed-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17751" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 287, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14358-14367, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 288, + 140 + ], + "type": "text", + "content": "[26] Quan Huynh-Thu and Mohammed Ghanbari. The accuracy of psnr in predicting video quality for different video scenes and frame rates. Telecommunication Systems, 49: 35-48, 2012. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 288, + 196 + ], + "type": "text", + "content": "[27] Berivan Isik, Onur G Guleryuz, Danhang Tang, Jonathan Taylor, and Philip A Chou. Sandwiched video compression: Efficiently extending the reach of standard codecswith neural wrappers. arXiv preprint arXiv:2303.11473, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 288, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 288, + 263 + ], + "type": "text", + "content": "[28] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video super-resolution with recurrent structure-detail network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 645-660. Springer, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 288, + 330 + ], + "type": "text", + "content": "[29] Takashi Isobe, Songjiang Li, Xu Jia, Shanxin Yuan, Gregory Slabaugh, Chunjing Xu, Ya-Li Li, Shengjin Wang, and Qi Tian. Video super-resolution with temporal group attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8008-8017, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 288, + 397 + ], + "type": "text", + "content": "[30] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, and Seon Joo Kim. Deep video super-resolution network using dynamic upsampling filters without explicit motion compensation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3224-3232, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 288, + 454 + ], + "type": "text", + "content": "[31] Jan P Klopp, Keng-Chi Liu, Liang-Gee Chen, and Shao-Yi Chien. How to exploit the transferability of learned image compression to conventional codescs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16165-16174, 2021. 2, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 288, + 532 + ], + "type": "text", + "content": "[32] Faouzi Kossentini, Hassen Guermazi, Nader Mahdi, Chekib Nouira, Amir Naghdinezhad, Hassene Tmar, Omar Khelif, Phoenix Worth, and Foued Ben Amara. The svt-av1 encoder: overview, features and speed-quality tradeoffs. In Applications of Digital Image Processing XLIII, page 1151021. International Society for Optics and Photonics, 2020. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 534, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 288, + 577 + ], + "type": "text", + "content": "[33] Théo Ladune, Pierrick Philippe, Wassim Hamidouche, Lu Zhang, and Olivier Déforges. Conditional coding for flexible learned video compression. arXiv preprint arXiv:2104.07930, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 578, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 288, + 624 + ], + "type": "text", + "content": "[34] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 288, + 668 + ], + "type": "text", + "content": "[35] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with diverse contexts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22616-22626, 2023. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 288, + 714 + ], + "type": "text", + "content": "[36] Jiahao Li, Bin Li, and Yan Lu. Neural video compression with feature modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26099-26108, 2024. 2, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 545, + 117 + ], + "type": "text", + "content": "[37] Songnan Li, Fan Zhang, Lin Ma, and King Ngi Ngan. Image quality assessment by separately evaluating detail losses and additive impairments. IEEE Transactions on Multimedia, 13(5):935-949, 2011. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 119, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 547, + 185 + ], + "type": "text", + "content": "[38] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. Mucan: Multi-correspondence aggregation network for video super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 335-351. Springer, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 186, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 545, + 242 + ], + "type": "text", + "content": "[39] Yinxiao Li, Pengchong Jin, Feng Yang, Ce Liu, Ming-Hsuan Yang, and Peyman Milanfar. Comisr: Compression-informed video super-resolution. in 2021 IEEE. In CVF International Conference on Computer Vision (ICCV), pages 2523-2532, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 244, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 244, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 244, + 545, + 277 + ], + "type": "text", + "content": "[40] Zhi Li and Christos G Bampis. Recover subjective quality scores from noisy measurements. In 2017 Data compression conference (DCC), pages 52-61. IEEE, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 279, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 279, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 279, + 545, + 312 + ], + "type": "text", + "content": "[41] Zhi Li, Anne Aaron, Ioannis Katsavounidis, Anush Moorthy, and Megha Manohara. Toward a practical perceptual video quality metric. The Netflix Tech Blog, 6, 2016. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 315, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 545, + 347 + ], + "type": "text", + "content": "[42] Zhi Li, Christos Bampis, Julie Novak, Anne Aaron, Kyle Swanson, Anush Moorthy, and J Cock. Vmaf: The journey continues. *Netflix Technology Blog*, 2018. 1, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 545, + 392 + ], + "type": "text", + "content": "[43] Zhi Li, Christos G Bampis, Lukáš Krasula, Lucjan Janowski, and Ioannis Katsavounidis. A simple model for subject behavior in subjective experiments. arXiv preprint arXiv:2004.02067, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 396, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 545, + 449 + ], + "type": "text", + "content": "[44] Jianping Lin, Dong Liu, Houqiang Li, and Feng Wu. MLvc: Multiple frames prediction for learned video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3546-3554, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 453, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 545, + 507 + ], + "type": "text", + "content": "[45] Jerry Liu, Shenlong Wang, Wei-Chiu Ma, Meet Shah, Rui Hu, Pranaab Dhawan, and Raquel Urtasun. Conditional entropy coding for efficient video compression. In European Conference on Computer Vision, pages 453-468. Springer, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 510, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 545, + 564 + ], + "type": "text", + "content": "[46] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. Dvc: An end-to-end deep video compression framework. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 567, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 545, + 610 + ], + "type": "text", + "content": "[47] Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, and Eirikur Agustsson. Vct: A video compression transformer. arXiv preprint arXiv:2206.07307, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 545, + 657 + ], + "type": "text", + "content": "[48] Alexandre Mercat, Marko Viitanen, and Jarno Vanne. Uvg dataset: 50/120fps 4k sequences for video codec analysis and development. In Proceedings of the 11th ACM Multimedia Systems Conference, pages 297-302, 2020. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 545, + 714 + ], + "type": "text", + "content": "[49] Reza Pourreza, Hoang Le, Amir Said, Guillaume Sautiere, and Auke Wiggers. Boosting neural video codecs by exploiting hierarchical redundancy. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5355-5364, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17752" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 287, + 117 + ], + "type": "text", + "content": "[50] Zhongwei Qiu, Huan Yang, Jianlong Fu, and Dongmei Fu. Learning spatiotemporal frequency-transformer for compressed video super-resolution. In European Conference on Computer Vision, pages 257-273. Springer, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 288, + 185 + ], + "type": "text", + "content": "[51] Alexander Raake, Silvio Borer, Shahid M Satti, Jörgen Gustafsson, Rakesh Rao Ramachandra Rao, Stefano Medagli, Peter List, Steve Göring, David Lindero, Werner Robitza, et al. Multi-model standard for bitstream-, pixel-based and hybrid video quality assessment of uhd/4k: Itu-t p. 1204. IEEE Access, 8:193020-193049, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 288, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 288, + 231 + ], + "type": "text", + "content": "[52] Reza Rassool. Vmaf reproducibility: Validating a perceptual practical video quality metric. In 2017 IEEE international symposium on broadband multimedia systems and broadcasting (BMSB), pages 1-2. IEEE, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 233, + 287, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 287, + 288 + ], + "type": "text", + "content": "[53] Oren Rippel, Alexander G Anderson, Kedar Tatwawadi, Sanjay Nair, Craig Lytle, and Lubomir Bourdev. Elf-vc: Efficient learned flexible-rate video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14479-14488, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 289, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 287, + 334 + ], + "type": "text", + "content": "[54] Amir Said, Manish Kumar Singh, and Reza Pourreza. Differentiable bit-rate estimation for neural-based video codec enhancement. In 2022 Picture Coding Symposium (PCS), pages 379-383. IEEE, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 335, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 287, + 380 + ], + "type": "text", + "content": "[55] Mehdi SM Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6626-6634, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 287, + 415 + ], + "type": "text", + "content": "[56] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal context mining for learned video compression. IEEE Transactions on Multimedia, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 417, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 287, + 460 + ], + "type": "text", + "content": "[57] Hanbin Son, Taeoh Kim, Hyeongmin Lee, and Sangyoun Lee. Enhanced standard compatible image compression framework based on auxiliary codec networks. IEEE Transactions on Image Processing, 31:664-677, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 463, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 287, + 507 + ], + "type": "text", + "content": "[58] Yannick Strümpler, Ren Yang, and Radu Timofte. Learning to improve image compression without changing the standard decoder. In European Conference on Computer Vision, pages 200–216. Springer, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 509, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 287, + 552 + ], + "type": "text", + "content": "[59] Hossein Talebi and Peyman Milanfar. Learned perceptual image enhancement. In 2018 IEEE International Conference on Computational Photography (ICCP), pages 1-13. IEEE, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 287, + 599 + ], + "type": "text", + "content": "[60] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and Jiaya Jia. Detail-revealing deep video super-resolution. In Proceedings of the IEEE international conference on computer vision, pages 4472-4480, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 287, + 655 + ], + "type": "text", + "content": "[61] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu Tdan. temporally-deformable alignment network for video super-resolution. in 2020 ieee. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3357-3366, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 287, + 713 + ], + "type": "text", + "content": "[62] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 545, + 106 + ], + "type": "text", + "content": "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 545, + 161 + ], + "type": "text", + "content": "[64] Ties van Rozendaal, Tushar Singhal, Hoang Le, Guillaume Sautiere, Amir Said, Krishna Buska, Anjuman Raha, Dimitris Kalatzis, Hitarth Mehta, Frank Mayer, et al. Mobilenvc: Real-time 1080p neural video compression on a mobile device. arXiv preprint arXiv:2310.01258, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 545, + 205 + ], + "type": "text", + "content": "[65] Abhinau K Venkataramanan, Chengyang Wu, Alan C Bovik, Ioannis Katsavounidis, and Zafar Shahid. A hitchhiker's guide to structural similarity. IEEE Access, 9: 28872-28896, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 547, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 547, + 260 + ], + "type": "text", + "content": "[66] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 545, + 315 + ], + "type": "text", + "content": "[67] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Huchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2012-2021, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 545, + 381 + ], + "type": "text", + "content": "[68] Ping-Hao Wu, Ioannis Katsavounidis, Zhijun Lei, David Ronca, Hassene Tmar, Omran Abdelkafi, Colton Cheung, Foued Ben Amara, and Faouzi Kossentini. Towards much better svt-av1 quality-cycles tradeoffs for vod applications. In Applications of Digital Image Processing XLIV, pages 236–256. SPIE, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 545, + 436 + ], + "type": "text", + "content": "[69] Li Xu, Gang He, Jinjia Zhou, Jie Lei, Weiying Xie, Yun-song Li, and Yu-Wing Tai. Transcoded video restoration by temporal spatial auxiliary network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2875-2883, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 545, + 480 + ], + "type": "text", + "content": "[70] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8):1106-1125, 2019. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 482, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 545, + 535 + ], + "type": "text", + "content": "[71] Ren Yang, Mai Xu, and Zulin Wang. Decoder-side hevc quality enhancement with scalable convolutional neural network. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 817-822. IEEE, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 536, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 545, + 581 + ], + "type": "text", + "content": "[72] Ren Yang, Mai Xu, Tie Liu, Zulin Wang, and Zhenyu Guan. Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology, 29(7):2039-2054, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 545, + 624 + ], + "type": "text", + "content": "[73] Ren Yang, Mai Xu, Zulin Wang, and Tianyi Li. Multi-frame quality enhancement for compressed video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6664-6673, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 545, + 680 + ], + "type": "text", + "content": "[74] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive fusion video super-resolution network via exploiting non-local spatio-temporal correlations. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3106-3115, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 545, + 713 + ], + "type": "text", + "content": "[75] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient video superresolution. In Proceedings of the IEEE/CVF International" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17753" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 287, + 152 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 76, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 286, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision, pages 4429-4438, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 287, + 152 + ], + "type": "text", + "content": "[76] Hengsheng Zhang, Xueyi Zou, Jiaming Guo, Youliang Yan, Rong Xie, and Li Song. A codec information assisted framework for efficient compressed video superresolution. In European Conference on Computer Vision, pages 220-235. Springer, 2022. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17754" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/f0ca3209-a68f-41a0-a209-e41298a3de25_content_list.json b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/f0ca3209-a68f-41a0-a209-e41298a3de25_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a80159ad5682f7a4408ff19b5329ed53297a65dc --- /dev/null +++ b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/f0ca3209-a68f-41a0-a209-e41298a3de25_content_list.json @@ -0,0 +1,1409 @@ +[ + { + "type": "text", + "text": "Perceptually Accurate 3D Talking Head Generation: New Definitions, Speech-Mesh Representation, and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 140, + 130, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lee Chae-Yeon $^{1*}$ Oh Hyun-Bin $^{2*}$ Han EunGi $^{1}$ Kim Sung-Bin $^{2}$ Suekyeong Nam $^{3}$ Tae-Hyun Oh $^{1,2,4}$", + "bbox": [ + 99, + 202, + 903, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Grad. School of AI, POSTECH $^{2}$ Dept. of Electrical Engineering, POSTECH $^{3}$ KRAFTON $^{4}$ School of Computing, KAIST", + "bbox": [ + 119, + 233, + 885, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb6b0c40d41458ddd681a07cd73458212aa53cb72f5c6642970e29e1aa81b450.jpg", + "image_caption": [ + "(a) Essential Criteria for Perceptually Accurate 3D Talking Head" + ], + "image_footnote": [], + "bbox": [ + 93, + 284, + 493, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c154fed79a46522ce17c99ae75dba0cec284ca566baa14e802bea69d8a08887a.jpg", + "image_caption": [ + "(b) Conceptual Diagram of Desired Representation Space", + "Figure 1. What defines perceptually accurate lip movement for a speech signal? In this work, we define three criteria to assess perceptual alignment between speech and lip movements of 3D talking heads: Temporal Synchronization, Lip Readability, and Expressiveness (a). The motivational hypothesis is the existence of a desirable representation space that models and complies well with the three criteria between diverse speech characteristics and 3D facial movements, as illustrated in (b); where representations with the same phonemes are clustered, are sensitive to temporal synchronization, and follow a certain pattern as the speech intensity increases. Consequently, we build a rich speech-mesh synchronized representation space that exhibits the desirable properties." + ], + "image_footnote": [], + "bbox": [ + 516, + 286, + 893, + 410 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 541, + 326, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in speech-driven 3D talking head generation have made significant progress in lip synchronization. However, existing models still struggle to capture the perceptual alignment between varying speech characteristics and corresponding lip movements. In this work, we claim that three criteria—Temporal Synchronization, Lip Readability, and Expressiveness—are crucial for achieving perceptually accurate lip movements. Motivated by our hypothesis that a desirable representation space exists to meet these three criteria, we introduce a speech-mesh synchronized representation that captures intricate correspondences between speech signals and 3D face meshes. We found that our learned representation exhibits desirable characteristics, and we plug it into existing models as a perceptual loss to better align lip movements to the given speech. In addition, we utilize this representation as a perceptual metric and introduce two other physically grounded lip synchronization metrics to assess how well the generated 3D talking heads align with these three criteria. Experiments show that training 3D talking head generation models with our perceptual loss", + "bbox": [ + 89, + 573, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "significantly improve all three aspects of perceptually accurate lip synchronization. Codes and datasets are available at https://perceptual-3d-talking-head.github.io/.", + "bbox": [ + 511, + 542, + 908, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 614, + 643, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Speech-driven 3D talking head generation focuses on generating 3D facial movements synchronized with input speech signals. This plays a key role in enhancing communication within multimedia applications, such as virtual reality, entertainment, and education [39]. To provide users with a more realistic and immersive experience, it is crucial that the facial and lip movements of 3D avatars are synchronized with the various aspects of speech. This synchronization should be perceptually accurate from a human perspective, ensuring that the avatars' expressions are both natural and convincing.", + "bbox": [ + 509, + 642, + 908, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While recent works in learning-based 3D talking head generation [8, 9, 21, 27, 33, 46] aim to enhance the lip synchronization capabilities, they commonly rely on minimizing the Mean Squared Error (MSE) loss between generated 3D facial motion and ground truth motion as a learning objective. This approach is practical, as it directly contributes to minimizing the Lip Vertex Error (LVE), a commonly used", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally.", + "bbox": [ + 93, + 886, + 279, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21065", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "metric that measures the MSE between the lip vertices of generated 3D facial motions and the ground truth.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the improvements in the LVE metric, existing models still struggle to correlate lip movements with some speech characteristics, such as wider mouth openings as speech volume increases. These characteristics are not adequately captured by existing datasets [8, 12], which have limited ranges of facial motion patterns due to the small dataset scale and restricted intensity range. Moreover, relying on MSE and LVE is insufficient for learning or assessing perceptually plausible lip motions [21, 37], as it focuses solely on vertex-wise geometric differences and overlooks the true correspondence between the speech signals and lip movements. A lower MSE and LVE do not necessarily correspond to a more perceptually accurate lip movement.", + "bbox": [ + 88, + 125, + 485, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These observations raise critical questions: What defines perceptually accurate lip movement in response to a speech signal, and how can we enhance this accuracy? We draw inspiration from findings on human audio-visual perception: 1) Humans are sensitive to temporal asynchrony between speech and lip movements; slight discrepancies can disrupt the perception of natural synchronization [43]. 2) Humans rely on accurate viseme-phoneme correspondence when assessing lip-sync accuracy, expecting visual lip movements to match the spoken phonemes [3]. 3) There is a proportional increase in jaw and lip movements as speech intensity increases, contributing to the expressiveness perceived in natural speech [18, 28, 34, 40]. Through our human study, we reveal an intriguing finding: participants favor lip movements with intensity that corresponds to speech—even if they exceed the established maximum acceptable asynchrony [43] by twice—over those that are perfectly synchronized but lack expressive alignment (Table 1-[Right]). This reveals that humans are more sensitive to expressiveness than temporal synchronization when perceiving, highlighting the importance of expressiveness. Building upon these insights, we define three criteria that significantly impact the perceptual lip synchronization of 3D talking heads: Temporal Synchronization, Lip Readability, and Expressiveness (Fig. 1 (a)).", + "bbox": [ + 91, + 324, + 486, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by our hypothesis that a desirable representation space exists to meet these three criteria, we propose a speech-mesh synchronized and rich representation that captures the intricate correspondence between speech and 3D face mesh. We design a transformer-based architecture that maps time-sequenced speech and mesh inputs to a shared representation space. To effectively train this system, we employ a two-stage training method: we first develop a robust audio-visual speech representation using a large-scale 2D video dataset [1], which then serves as an anchor for learning a speech-mesh representation. We found that the first step is important and leads to emergent desirable properties (illustrated in Fig. 1 (b)) in the final representation. This sequential approach ensures that the model first establishes", + "bbox": [ + 89, + 688, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a space that captures a wide range of speech characteristics, and then extends to explore the relationships between speech and 3D face mesh across diverse speech intensities and facial movements. Adopting this representation, we introduce a plug-and-play perceptual loss adaptable to any existing 3D talking head generation models [11, 21, 46], enhancing the perceptual quality of 3D talking heads.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, to assess the three criteria, we introduce three metrics for each aspect. We leverage our learned representation as a perceptual metric, Perceptual Lip Readability Score (PLRS), to evaluate the perceptual lip readability of lip movements. Also, we propose two physically grounded lip synchronization metrics: Mean Temporal Misalignment (MTM) for temporal synchronization and Speech-Lip Intensity Correlation Coefficient (SLCC) for expressiveness.", + "bbox": [ + 511, + 198, + 908, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate that our perceptual loss significantly enhances all three aspects: Temporal Synchronization, Lip Readability, and Expressiveness, which are demonstrated across various metrics: existing metric, our newly proposed metrics, and human evaluations. We also find that incorporating an additional pseudo-dataset [45], which captures diverse ranges of speech and lip movement intensities, can further improve expressiveness. Our main contributions are summarized as follows:", + "bbox": [ + 511, + 319, + 908, + 454 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Defining three aspects—Temporal Synchronization, Lip Readability, and Expressiveness—that affect the perceptual quality of 3D talking heads and proposing three evaluation metrics for these aspects.", + "- Constructing a speech-mesh representation space that captures rich and diverse correspondences between speech and lip movements.", + "- Proposing a plug-in perceptual loss using the constructed speech-mesh representation and demonstrating improvements on existing metric, our newly proposed metrics, and human evaluations." + ], + "bbox": [ + 511, + 457, + 908, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 648, + 656, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Speech-driven 3D talking head generation. Speech-driven 3D talking head generation aims to generate realistic 3D facial movements aligned with given speech. Among recent data-driven methods [8, 9, 11, 21, 22, 33, 46, 47], FaceFormer [11] introduces a transformer-based autoregressive model and leverages a pre-trained speech model to capture long-term audio context and past facial movements. CodeTalker [46] employs a VQ-VAE to construct a discrete facial motion space, addressing the over-smoothing problem. Diffusion models [17, 31] have also been demonstrated to be effective for 3D talking head generation [33, 36]. In addition to synthesizing neutral facial motions, several works extend the 3D talking head to express specific aspects, such as emotional expressions [9, 22], multilingual capabilities [37], or laughter [38]. Despite these advances, existing methods rely", + "bbox": [ + 511, + 674, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21066", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on minimizing MSE loss without a clear definition of perceptually accurate lip movement, overlooking the multifaceted nature of lip synchronization. To address this, we define three critical aspects of lip synchronization and propose rich speech-mesh representation, along with its application as a perceptual loss in a plug-and-play manner, enhancing all three aspects of lip synchronization quality in existing 3D talking head generation models [11, 21, 46].", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Speech-face representation learning. Well-aligned representation spaces learned from large-scale datasets such as CLIP [24] and ImageBind [14] are valued for their scalability and versatility. These spaces enable a wide range of applications, including auxiliary loss [42], intermediate representation [41], and evaluation metrics [16]. With this context, audio-visual representation spaces trained specifically on speech and 2D face videos have been proposed [6, 15, 29, 35]. For instance, SyncNet [6], a CNN-based model learns to detect audio-visual temporal synchronization and has been applied to several tasks, such as active speaker detection [1, 7] and 2D talking head generation [23, 44, 48]. Similarly, a transformer-based AV-HuBERT [29] has demonstrated remarkable effectiveness in various tasks, including lip reading [30], audio-visual translation [5], and 2D talking head generation [44]. While there has been significant progress in the 2D domain, advancements in the 3D domain remain under-explored. Yang et al. [47] extends the SyncNet architecture to accommodate speech and 3D face meshes; however, its application is limited to evaluating 3D talking heads. In this work, we demonstrate the versatility of our representation space as a plug-in module to improve the perceptual accuracy of the existing speech-driven 3D talking head generation models [11, 21, 46], as well as to assess their performance.", + "bbox": [ + 91, + 217, + 486, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation metrics for speech-driven 3D talking head. The prevalent evaluation metric, Lip Vertex Error (LVE) [27], measures the L2 distance between predicted lip vertices and ground truth. Additional metrics, such as Upper Face Dynamics Deviation (FDD) [46] and Lip Readability Percentage (LRP) [21], consider different regions of facial motion. These metrics, however, focus on vertex-wise geometric differences between the ground-truth 3D facial motions and neglect speech-related information. To incorporate both speech and 3D face mesh data for evaluation, MultiTalk [37] introduces an Audio-Visual Lip Readability (AVLR) metric, which assesses perceptual accuracy of lip readability using a pre-trained Audio-Visual Speech Recognition model [2]. Yet, AVLR relies on speech and 2D face video rendered from the 3D face mesh, which may not align with the 3D talking head domain. To address these limitations, we introduce novel and comprehensive evaluation metrics that focus on diverse aspects of lip synchronization: Mean Temporal Misalignment (MTM), Perceptual Lip Readability Score (PLRS), and Speech-Lip Intensity Correlation Coefficient (SLCC).", + "bbox": [ + 89, + 598, + 486, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Essential Criteria for Perceptually Accurate 3D Talking Head", + "text_level": 1, + "bbox": [ + 513, + 89, + 906, + 125 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generating a 3D talking head with lip movements that are perceptually accurate to human observers requires a clear understanding of the components that influence perceptual quality. Although existing works have focused on improving partial aspects of these talking head generation models, a comprehensive exploration for establishing a perceptually accurate 3D talking head model has barely been undertaken. Drawing inspirations from extensive research in existing works [10, 18, 21, 28, 34, 40, 43, 44] and our studies, we identify three fundamental criteria essential for achieving perceptually accurate lip movements in 3D talking heads.", + "bbox": [ + 511, + 132, + 906, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Synchronization. This alignment is particularly crucial in media involving human speech, where viewers expect lip movements to precisely match the corresponding speech in time. Temporal misalignment between speech and lip movements indeed distracts viewers, reduces user experiences, and negatively impacts audience perception [25]. Vatakis et al. [43] find that viewers are particularly sensitive to speech-lip asynchrony compared to other audio-visual asynchrony, such as music. They note that mismatches become noticeable when speech precedes lip movements by more than $50~\\mathrm{ms}$ or follows them by more than $220~\\mathrm{ms}$ . These findings may hold the same in 3D talking head generation, where any misalignment between speech and lip movements can break the illusion of realism, leading to a diminished user experience and reducing the perceived authenticity of the virtual character. Thus, we define the temporal synchronization of the talking head as an important aspect of lip synchronization quality.", + "bbox": [ + 511, + 303, + 908, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lip Readability. Visemes (or lip movements) must correspond accurately to the speech phonemes to ensure the spoken words are visually intelligible [3]. This aspect is widely acknowledged as important in existing literature in both 2D [44] and 3D [10, 21] speech-driven talking head models, which leverage lip reading experts as an auxiliary guidance to improve the visual intelligibility of the spoken word. However, the mapping between speech and lip movements is not one-to-one, making lip readability challenging to define. For example, the size and shape of the opening mouth and the dynamic movement patterns of the lips in response to specific utterance differ at every moment [47] or per individual [32]. To capture this complexity, we define the lip readability within speech-mesh synchronized representation space learned from a large-scale dataset, capturing the comprehensive and nuanced correspondence between speech and lip movements.", + "bbox": [ + 511, + 579, + 908, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Expressiveness. Speech conveys not only linguistic content but also varies in intensity. For instance, a speaker may express the same text softly or loudly, with jaw and lip movements proportionally increasing as speech intensity rises,", + "bbox": [ + 511, + 839, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21067", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bfc3b81fc34afd47d5a65a4cbf9cb533638836942a950c99d37a36af7b69549e.jpg", + "image_caption": [ + "Table 1. Human studies on alignment criteria. [Left] Preference scores (1-3) for 3D talking heads with varying lip movement intensities paired with different speech intensities. [Right] Human preference between (A) samples with precise timing but low expressiveness, and (B) samples with high expressiveness but 100ms asynchrony—twice the commonly accepted 50ms threshold [43]." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 292, + 215 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/167a30a7caf733ca565415124d904efd54814e973ba34d453d55d137b1c8ed05.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Samples
AB
Temp.✓×
Exp.×✓
Prefer (%)17.482.6
", + "bbox": [ + 321, + 104, + 478, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which contributes to perceived expressiveness in real-world face recognition [18, 28, 40]. To demonstrate that the positive correlation of human preference between the intensity of speech and lip movements also exists in 3D talking head field, we conduct a human study by presenting 3D talking heads with varied lip movement intensities paired with different speech intensities in Table 1-[Left]. Participants prefer the lip movements with the intensity that match the intensity of speech. Despite this distinct human preference for synchronization between the intensity of corresponding speech and lip movements, this aspect has barely been explored in the talking head generation field. We thus define the expressiveness in 3D talking head as the correlation between speech and lip movement intensity, which is crucial for establishing genuine talking heads and is expected to guide future research aimed at improving perceptual lip synchronization.", + "bbox": [ + 88, + 328, + 483, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Focus of this work. Among the three criteria for perceptually accurate 3D talking head, we find that recent 3D talking head generation models achieve reasonable temporal alignment between speech and lip movements (see supplementary for a visualization of temporal synchronization). Furthermore, we design an A vs. B test, prompting participants to choose between two samples: (A) temporally synchronized one while lacking expressive synchronization, and (B) the other with expressive synchronization but with $100\\mathrm{ms}$ asynchrony. Table 1-[Right] shows that users prefer sample B, suggesting that humans may prioritize expressive synchronization over strict temporal alignment. This insight directs our focus toward enhancing all three aspects to better capture perceptual realism in 3D talking head. Details of the human study are in the supplementary material. Before introducing metrics to assess the criteria, we present our synchronized representation for designing our perceptual loss.", + "bbox": [ + 89, + 573, + 483, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Speech-Mesh Synchronized Representation", + "text_level": 1, + "bbox": [ + 89, + 844, + 478, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We hypothesize that a desirable representation space exists to meet the three criteria defined in Sec. 3. Motivated by this", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hypothesis, we develop a rich speech-mesh synchronized representation which captures the intricate correspondence between speech and the 3D face mesh. We found that our learned representation exhibits desirable properties, and we adopt it as a perceptual loss to improve the perceptual accuracy of existing models with respect to these three criteria.", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. Directly learning a synchronized speech-mesh representation presents challenges due to the scarcity of speech-3D face mesh datasets. One potential solution is to construct pseudo-GT 3D face meshes using reconstruction models [13], although relying solely on pseudo-GT may not suffice for building a robust representation. To overcome this, we leverage the extensive knowledge from the speech-2D lip video representations and adapt this to accommodate 3D face meshes. Specifically, we propose a two-stage training process: in stage 1, we learn an audio-visual speech representation that accurately reflects lip movements from the unlabeled in-the-wild 2D synchronized talking face video dataset, LRS3 [1]. Subsequently, we leverage the pre-trained speech representation from stage 1, using it as the anchor space to learn a synchronized speech-mesh representation.", + "bbox": [ + 511, + 185, + 908, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 1. Learning audio-visual speech representation. This stage aims to learn a rich speech-2D lip video representation that effectively captures the correlation between varying speech characteristics and lip dynamics. Motivated by prior works [15, 35], we extend the integration of masked autoencoder (MAE) and cross-modal contrastive learning to learn a synchronized speech-2D lip video representation using 2D videos. The architecture for stage 1 consists of two modality-specific encoders, a cross-modal fusion encoder, and two modality-specific decoders. (see Fig. 2-[Stage 1]).", + "bbox": [ + 511, + 417, + 908, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a speech and 2D lip video pair, $(\\mathbf{X}_{\\mathbf{s}},\\mathbf{X}_{\\mathbf{v}})\\in D$ we begin by patchifying and tokenizing speech spectrograms and video frames into speech and video tokens as $\\mathbf{S} = (\\mathbf{s}_1,\\dots ,\\mathbf{s}_N)$ and $\\mathbf{V} = (\\mathbf{v}_1,\\dots ,\\mathbf{v}_M)$ , where $\\mathbf{s}_i,\\mathbf{v}_j\\in$ $\\mathbb{R}^H$ . We then randomly mask out $P\\%$ portion of tokens. The remaining unmasked speech tokens $\\mathbf{S}^{unmask}$ and video tokens $\\mathbf{V}^{unmask}$ are respectively fed into the speech encoder $E_{s}$ and video encoder $E_{v}$ , each consisting of $N_{e}$ transformer layers. Each encoder extracts uni-modal embeddings, $\\mathbf{Z}_{\\mathbf{s}}^{l} = E_{s}^{l}(\\mathbf{S}^{unmask})$ and $\\mathbf{Z}_{\\mathbf{v}}^{l} = E_{v}^{l}(\\mathbf{V}^{unmask})$ , where $l\\in (1,\\ldots ,N_e)$ denote the layer indices. Also, the mean pooled speech and video embeddings, $\\mathbf{c_s^l} = \\mathrm{MeanPool}(\\mathbf{Z_s^l})$ and $\\mathbf{c_v^l} = \\mathrm{MeanPool}(\\mathbf{Z_v^l})$ , are derived from the corresponding uni-modal embeddings. Following the uni-modal encoders, we introduce a multi-modal fusion encoder $E_{sv}$ to exploit complementary information from each extracted uni-modal embedding. The speech and video embeddings are jointly processed through this encoder, resulting in multi-modal fusion embeddings $\\mathbf{F_s}$ and $\\mathbf{F_v}$ , i.e., $[\\mathbf{F_s},\\mathbf{F_v}] = E_{sv}([\\mathbf{Z_s},\\mathbf{Z_v}])$ where $[,\\cdot ]$ denotes concatenation. With the fusion embeddings $\\mathbf{F_s}$ and $\\mathbf{F_v}$ , we utilize each modality's decoder to reconstruct the original signals. Specifically, we employ a", + "bbox": [ + 511, + 568, + 910, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21068", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d974d05e76c6c8e9a4c6306fb0e15cd76881e5ffe7df953523490cbeac87cc92.jpg", + "image_caption": [ + "Figure 2. Pipeline of speech-mesh synchronized representation learning. We train our speech-mesh representation space in a two-stage manner. In the first stage, we learn a rich audio-visual representation in 2D domain to capture the synchronization between lip movement and speech. In the second stage, we train the 3D mesh encoder to align the 3D mesh space with the frozen speech space. As an application of our speech-mesh representation space, we propose a plug-in perceptual loss to 3D talking head models to enhance the quality of lip movements." + ], + "image_footnote": [], + "bbox": [ + 143, + 84, + 495, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ba9880e77762d0057f882c1cd0af5b67e3498063f827cac49f1012342d34662.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 83, + 852, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "speech decoder $D_{s}$ and a video decoder $D_{v}$ , each consisting of $N_{d}$ transformer layers. We pad each multi-modal fusion embedding with trainable masked tokens at their original positions, resulting in $\\mathbf{F}_{\\mathbf{s}}^{\\prime}$ and $\\mathbf{F}_{\\mathbf{v}}^{\\prime}$ . Each decoder then reconstructs the respective signals; the reconstructed speech spectrogram and video tokens as $\\hat{\\mathbf{S}}^{mask} = D_s(\\mathbf{F}_{\\mathbf{s}}^{\\prime})$ and $\\hat{\\mathbf{V}}^{mask} = D_v(\\mathbf{F}_{\\mathbf{v}}^{\\prime})$ .", + "bbox": [ + 88, + 359, + 483, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training the model involves two objectives: learning the cross-modal alignment between the speech-2D lip video signals, while reconstructing their original signals. Given $B$ speech-2D lip video token pairs in a batch, $\\{\\mathbf{S}_i,\\mathbf{V}_i\\}_{i = 1}^B$ , we treat the true speech-2D lip video pair as positive samples, while the others in the batch are considered negative samples. To encourage alignment between temporally synchronized speech-2D lip video instances, we utilize a cross-modal contrastive learning strategy, InfoNCE [20]. We maximize the cosine similarity between positive sample of the mean pooled speech embedding $\\mathbf{c}_{\\mathbf{s},i}^{l} = \\mathrm{MeanPool}(E_{s}^{l}(\\mathbf{S}_{i}^{\\mathrm{unmask}}))$ and the corresponding synchronized mean pooled video embedding $\\mathbf{c}_{\\mathbf{v},i}^{l} = \\mathrm{MeanPool}(E_{v}^{l}(\\mathbf{V}_{i}^{\\mathrm{unmask}}))$ . We first define the speech-centric loss as:", + "bbox": [ + 88, + 467, + 483, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {V}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {v} , i} ^ {l} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {v} , j} ^ {l} / \\tau\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 685, + 483, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is a temperature hyperparameter. Also, we make the objective symmetric by defining a video-centric loss as $\\mathcal{L}_{\\mathrm{V}\\rightarrow \\mathrm{S}}$ . We sum $\\mathcal{L}_{\\mathrm{S}\\rightarrow \\mathrm{V}}$ and $\\mathcal{L}_{\\mathrm{V}\\rightarrow \\mathrm{S}}$ across $L$ selected encoder layers to obtain our final contrastive learning objective:", + "bbox": [ + 89, + 731, + 483, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {I n f o N C E}} = \\sum_ {l \\in L} \\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {V}} + \\mathcal {L} _ {\\mathrm {V} \\rightarrow \\mathrm {S}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 797, + 483, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For reconstruction, the model is trained in a self-supervised manner by minimizing the reconstruction loss $\\mathcal{L}_{\\mathrm{MAE}}$ as:", + "bbox": [ + 89, + 825, + 483, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M A E}} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left[ \\frac {\\sum \\| \\hat {\\mathbf {S}} _ {i} ^ {m a s k} - \\mathbf {S} _ {i} ^ {m a s k} \\| _ {F} ^ {2}}{| \\mathbf {S} _ {i} ^ {m a s k} |} + \\frac {\\sum \\| \\hat {\\mathbf {V}} _ {i} ^ {m a s k} - \\mathbf {V} _ {i} ^ {m a s k} \\| _ {F} ^ {2}}{| \\mathbf {V} _ {i} ^ {m a s k} |} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 861, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|\\mathbf{S}_i^{mask}|$ and $|\\mathbf{V}_i^{mask}|$ denote the number of masked speech and video tokens, respectively. Overall, our objective function is defined as:", + "bbox": [ + 511, + 359, + 906, + 404 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M A E}} + \\lambda \\mathcal {L} _ {\\text {I n f o N C E}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 410, + 906, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the weight factor for cross-modal contrastive loss.", + "bbox": [ + 511, + 428, + 906, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We observe that the representation space trained with the transformer architecture and a rich and large-scale 2D face video dataset in this way already possesses the desired properties we pursue, illustrated in Fig. 1 (see supplementary for visualizations of pre-trained speech representation). Motivated by this, we transfer these emergent properties to the speech-mesh representation space as follows.", + "bbox": [ + 509, + 443, + 906, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2. Learning speech-mesh representation. In this stage, we design a 3D mesh encoder that maps 3D face mesh to the speech representation pre-trained in stage 1. This pretrained speech representation, derived from rich speech-2D face video data, serves as a robust anchor space for learning the correlation between diverse speech characteristics and 3D facial motions. We use contrastive learning loss to align 3D facial motion embeddings with the anchored speech representations, as shown in Fig. 2-[Stage 2].", + "bbox": [ + 511, + 553, + 908, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a speech and 3D face mesh pair $(\\mathbf{X}_{\\mathbf{s}},\\mathbf{X}_{\\mathbf{m}})\\in D$ we patchify and tokenize speech spectrograms into speech tokens S, and map these into pre-trained uni-modal mean pooled speech embeddings $\\mathbf{c_s^l} = \\mathrm{MeanPool}(E_s^l (\\mathbf{S}))$ . Similarly, we patchify 3D face mesh into mesh tokens M, and feed them into the 3D mesh encoder $E_{m}$ consisting of $N_{e}$ transformer layers to extract mesh embeddings $\\mathbf{Z}_{\\mathbf{m}}^{l} = E_{m}^{l}(\\mathbf{M})$ and the corresponding mean-pooled mesh embedding $\\mathbf{c_m^l}$ . For the learning objective, given $B$ speech-3D face mesh token pairs in a batch, $\\{\\mathbf{S}_i,\\mathbf{M}_i\\}_{i = 1}^B$ , we first define speech-anchored loss as:", + "bbox": [ + 511, + 688, + 910, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {M}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {m} , i} ^ {l} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {m} , j} ^ {l} / \\tau\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 862, + 906, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21069", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as well as $\\mathcal{L}_{\\mathrm{M}\\to \\mathrm{S}}$ which is the mesh-anchored loss. Similar to Eq. (2), we sum $\\mathcal{L}_{\\mathrm{S}\\rightarrow \\mathrm{M}}$ and $\\mathcal{L}_{\\mathrm{M}\\rightarrow \\mathrm{S}}$ across $L$ selected encoder layers to train the 3D mesh encoder $E_{m}$ while the speech encoder fixed. As a result, this two-stage training approach yields a robust speech-mesh representation, which significantly enhances the alignment between the speech and 3D face mesh modalities.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adopting it as a perceptual loss. A key application of the speech-mesh representation learned through the two-stage training process is its use as a perceptual loss to enhance the perceptual accuracy of the 3D talking head model (see Fig. 2). Leveraging this representation as a perceptual loss ensures that the generated lip movements are perceptually accurate and aligned well with the speech.", + "bbox": [ + 89, + 200, + 480, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given $B$ speech and generated 3D face mesh pairs in a batch, $\\{\\mathbf{S}_i,\\hat{\\mathbf{M}}_i\\}_{i = 1}^B$ , we define our perceptual loss with the symmetric InfoNCE loss (Eq. (2)) as", + "bbox": [ + 89, + 306, + 480, + 352 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p e r c p}} = \\sum_ {l \\in L} \\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {M}} + \\mathcal {L} _ {\\mathrm {M} \\rightarrow \\mathrm {S}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 361, + 482, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our perceptual loss encourages synchronized speech and mesh embeddings to pull closer together, while unsynchronized ones to push apart. The effectiveness and analysis of our perceptual loss will be discussed in later sections.", + "bbox": [ + 89, + 402, + 483, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 474, + 274, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we describe our proposed evaluation metrics that assess each criterion impacting the quality of 3D lip accuracy, as discussed in Sec. 3. Here, we outline the high-level concepts of our proposed metrics. Additional details and experiments are provided in the supplementary material.", + "bbox": [ + 89, + 500, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mean Temporal Misalignment (MTM). To measure the temporal discrepancy between speech and corresponding lip movements, temporal correspondence annotations, such as onset times for each modality, would typically be required. As a proxy, we determine temporal correspondence and measure temporal discrepancies between the ground truth and predicted lip vertex displacement sequences by using Derivative Dynamic Time Warping (DDTW) [19], which robustly identifies local structural similarities compared to standard Dynamic Time Warping (DTW) [4]. For simplicity, we focus on the central vertices of the upper and lower lips when extracting displacement sequences, and use local extrema in the DDTW process to measure temporal misalignment by pinpointing precise time steps for mouth opening and closing events. Consequently, Mean Temporal Misalignment (MTM) $\\overline{\\Delta t}$ is defined as $\\overline{\\Delta t} = \\frac{1}{K}\\sum_{k=1}^{K}\\Delta t_k$ , where $K$ is the total number of video clips and $\\Delta t_k$ is the averaged temporal misalignment of the $k$ -th video clip.", + "bbox": [ + 89, + 579, + 483, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Perceptual Lip Readability Score (PLRS). While Lip Vertex Error (LVE) measures the accuracy of generated lip articulations against the ground truth, it does not fully assess", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "whether lip movements are perceptually aligned with the given speech. To address this, we leverage our speech-mesh representation in Sec. 4 as a perceptual lip readability evaluator. We compute the perceptual alignment using the cosine similarity of the mean pooled speech and mesh embeddings. Since this representation has learned a rich distribution of speech correspondences across various facial movements, our metric correlates highly with human perception, measuring perceptual lip movement alignment more accurately than LVE (see supplementary for the human study on metrics).", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Speech and Lip Intensity Correlation Coefficient (SLCC). As shown in Table 1-[Left], humans prefer aligned intensity between speech and lip movement. Thus, the intensity of generated lip movements should positively correlate with the corresponding input speech. To quantify this, we define the Speech and Lip Correlation Coefficient $r_{SL}$ as:", + "bbox": [ + 511, + 244, + 906, + 335 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nr _ {S L} = \\frac {\\sum_ {k = 1} ^ {K} (\\mathrm {S I} _ {k} - \\bar {\\mathrm {S I}}) (\\mathrm {L I} _ {k} - \\bar {\\mathrm {L I}})}{\\sqrt {\\sum_ {k = 1} ^ {K} (\\mathrm {S I} _ {k} - \\bar {\\mathrm {S I}}) ^ {2}} \\sqrt {\\sum_ {k = 1} ^ {K} (\\mathrm {L I} _ {k} - \\bar {\\mathrm {L I}}) ^ {2}}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 347, + 906, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{SI}_k$ and $\\mathrm{LI}_k$ denote Speech (SI) and Lip Intensity (LI), respectively, $\\bar{\\mathrm{SI}} = \\frac{1}{K}\\sum_{k = 1}^{K}\\mathrm{SI}_k$ and $\\bar{\\mathrm{LI}} = \\frac{1}{K}\\sum_{k = 1}^{K}\\mathrm{LI}_k$ . We define SI using speech loudness, specifically the z-normalized Root Mean Square (RMS) value, which is a widely accepted measure of speech intensity in signal processing. To define LI, we measure the averaged lip displacement value of a video clip, followed by the z-normalization.", + "bbox": [ + 511, + 386, + 906, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 513, + 506, + 645, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first outline the evaluation setup, and then present thorough analyses of the experimental results. Due to the space limitation, we present more implementation details and additional experiments in the supplementary material.", + "bbox": [ + 511, + 532, + 906, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 603, + 720, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Most of the existing speech-driven 3D talking head generation methods rely on VOCASET [8] and BIWI [12] to train and test the models. However, these datasets have limited ranges of facial motion patterns due to the small dataset scale and restricted intensity range, which restricts their ability to fully capture the intricate relationship between speech and 3D face mesh. To address the lack of training dataset, we construct two large-scale speech-3d face mesh benchmark datasets, LRS3-3D and MEAD-3D, by processing LRS3 and MEAD videos using two monocular face reconstruction methods: SPECTRE [13] for LRS3 [1], which ensures accurate lip movements, and SMIRK [26] for MEAD [45], which captures diverse speech and lip movement intensities. We use LRS3 in the first stage and LRS3-3D in the second stage to train our speech-mesh synchronized representation. Then, for base model training and evaluations, we adopt two configurations: (1) training and testing with VOCASET, in line with existing work, and (2)", + "bbox": [ + 511, + 628, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21070", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/89dd1f89657db398b2a19e5e5d89659d172252aa77fa2bd056c1b7c32a58490c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)SLCC / Δ (↓)
VOCASET--0.4090.34 / -
FaceFormer [11]53.63.3570.3680.26 / 0.08
+ Ours rep.52.23.0910.4630.37 / 0.03
CodeTalker [46]61.83.7000.3810.38 / 0.04
+ Ours rep.60.93.5790.3880.35 / 0.01
SelfTalk [21]50.12.9710.4140.41 / 0.07
+ Ours rep.49.22.9240.4180.35 / 0.01
", + "bbox": [ + 94, + 88, + 478, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8ee7c40b0f81bf6cdc552ff6b15fdbbdc18de6958c846f2a6e059060076b58ae.jpg", + "image_caption": [ + "Figure 3. Qualitative results of the effectiveness of our perceptual loss for lip readability. Our perceptual loss guides baselines [11, 21, 46] to generate perceptually accurate lip movements." + ], + "image_footnote": [], + "bbox": [ + 91, + 347, + 483, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "combining MEAD-3D with VOCASET during training to endow expressiveness and testing on MEAD-3D.", + "bbox": [ + 89, + 523, + 482, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Base methods. We use three state-of-the-art 3D talking head generation models [11, 21, 46] to evaluate the effectiveness of our perceptual loss.", + "bbox": [ + 89, + 558, + 483, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. To comprehensively evaluate the three aspects of lip synchronization, we assess MTM, PLRS, and SLCC, corresponding to temporal synchronization, lip readability, and expressiveness, respectively. Additionally, we compute the level-wise SLCC for the MEAD-3D test set, which includes three distinct emotional intensity levels, to evaluate the expressive capability of 3D talking head generation models. We also measure LVE as part of the lip readability evaluation.", + "bbox": [ + 89, + 608, + 483, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Experimental Results and Analysis", + "text_level": 1, + "bbox": [ + 89, + 741, + 395, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct evaluations to assess the effectiveness of our speech-mesh synchronized representation and the incorporation of an expressive speech-3D face mesh paired dataset (i.e., MEAD-3D) in enhancing the three criteria.", + "bbox": [ + 89, + 763, + 483, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How well do existing 3D talking head models achieve lip synchronization in all three aspects? The results are summarized in Table 2. For temporal synchronization, most base models achieve MTM values between 50 and 60ms, indicating performance close to the acceptable asynchrony", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8a127620c967bce4a0245591598dba9296611634e054bc221485281ab64b2018.jpg", + "table_caption": [ + "Table 2. Quantitative results of lip synchronization on VO-CASET [8] test set. We evaluate the base models on our proposed lip synchronization metrics. We denote $\\Delta$ as the difference in SLCC between the model and those measured on the data distribution. A lower $\\Delta$ indicates the model more closely represents the intensity correlation of the dataset. We demonstrate the effectiveness of our representation in consistently enhancing all three aspects of lip synchronization." + ], + "table_footnote": [], + "table_body": "
MethodPerceptual LossTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)SLCC / Δ (↓)
VOCASET---0.4090.34 / -
FaceFormer [11]X53.63.3570.3680.26 / 0.08
3D SyncNet55.63.3160.4350.38 / 0.04
Ours w/o 2D prior55.33.2780.4000.42 / 0.08
Ours w/ 2D prior52.23.0910.4630.37 / 0.03
CodeTalker [46]X61.83.7000.3810.38 / 0.04
3D SyncNet59.64.3190.3790.14 / 0.20
Ours w/o 2D prior55.93.5790.3740.23 / 0.11
Ours w/ 2D prior60.93.5790.3880.35 / 0.01
SelfTalk [21]X50.12.9710.4140.41 / 0.07
3D SyncNet49.52.9410.4050.35 / 0.01
Ours w/o 2D prior54.43.1490.4170.39 / 0.05
Ours w/ 2D prior49.22.9240.4180.35 / 0.01
", + "bbox": [ + 524, + 88, + 893, + 251 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study on architectural choice and 2D prior knowledge. We validate the effectiveness of the transformer-based architecture and curriculum learning with a pre-trained 2D speech representation by ablating them from our proposed representation.", + "bbox": [ + 511, + 252, + 903, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b22ebb0cf30f6c03f5d97d37a25d0c1543b1b5a51ac6a66c0b9ce71c975fa98.jpg", + "image_caption": [ + "Figure 4. t-SNE plot of ablation study. We plot the t-SNE graph for each perceptual critic model. We represent the features with same phoneme as same color. Squared and circled points denote mesh and speech features from each representation, respectively." + ], + "image_footnote": [], + "bbox": [ + 555, + 320, + 864, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/32f8920b1801e2fdbc270e8c0250e147e773508a0f3115cb9b5c6126d637af33.jpg", + "image_caption": [ + "Figure 5. Behaviors of our representation in temporal and expressiveness sensitivity. We demonstrate the effectiveness of our representation in temporal synchronization and expressiveness using a cosine similarity graph and speech feature plots, respectively. We color the point as low, medium, and high intensity." + ], + "image_footnote": [], + "bbox": [ + 555, + 511, + 859, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "threshold. Regarding lip readability, SelfTalk [21] achieves the best performance, while FaceFormer [11] has the lowest PLRS score. Furthermore, CodeTalker [46] demonstrates the closest SLCC values to the ground truth VOCASET mesh, while FaceFormer exhibits the highest SLCC discrepancy.", + "bbox": [ + 511, + 731, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Does our speech-mesh representation improve lip synchronization? Yes. Table 2 and Fig. 3 show consistent improvements in the three aspects with our perceptual loss.", + "bbox": [ + 511, + 809, + 908, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What makes our speech-mesh representation have lip synchronization ability? We hypothesize that the transformer-based architecture and curriculum learning with a pre-trained", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21071", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4e0c0f81521d8659774a5cab61ca4c5dfb8aec956e971a7eb01ceb6bdf53076.jpg", + "image_caption": [ + "Figure 6. Qualitative results for the expressiveness. Given high and low intensity levels of speech, models trained on both MEAD-3D and VOCASET show more expressive lip movements compared to those trained on VOCASET alone, and even better with our perceptual loss." + ], + "image_footnote": [], + "bbox": [ + 150, + 87, + 844, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "audio-visual speech representation contribute significantly to improved lip synchronization. To validate this, we conduct ablation studies, summarized in Table 3, examining the roles of pre-trained speech representation and architectural design. Without the pre-trained speech representation, base models show lower performance across metrics, and CNN-based architectures, 3D SyncNet, do not clearly show effectiveness, while ours with 2D prior consistently show improvement. The t-SNE plots of perceptual critic models in Fig. 4 show that, in (a) 3D SyncNet, speech and mesh features corresponding to the same phoneme group are separated. In (b) ours without 2D priors, speech and mesh features lack separation and appear scattered without clustering. Our representation (c), notably, tends to form more distinct clusters according to phonemes, with vowels and consonants grouped closely, potentially contributing to enhancing lip readability. We also observe a directional progression in the feature space, shifting from phonemes with mouth opening $(e.g., /aj/$ to those with mouth closing $(e.g., /f/$ .", + "bbox": [ + 88, + 271, + 485, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also examine our representation's performance in other aspects, temporal synchronization and expressiveness. Figure 5-(a) demonstrates temporal sensitivity, as cosine similarity drops when temporal misalignment is introduced between input speech and 3D face mesh. In Fig. 5-(b), we plot speech features at varying speech intensities, showing a directional trend as intensity increases from lowest to highest. Figures 4 and 5 imply that our representation holds favorable properties discussed in Fig. 1 for the three criteria.", + "bbox": [ + 89, + 563, + 483, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Can we unlock the expressive power of 3D talking heads?", + "text_level": 1, + "bbox": [ + 89, + 704, + 482, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Likely. Since VOCASET [8] lacks the range of diverse speech and lip movement intensities, evaluating expressive power requires testing on a dataset with a broader range of intensities. To study this, we examine the expressiveness of the base models on the MEAD-3D dataset, which includes three intensity levels. We assess SLCC at each intensity level to evaluate expressiveness and identify any expressiveness bounds as the level of intensity increases. We first test the base models trained on VOCASET [8] against the MEAD-3D test set. Table 4 shows these models demonstrate limited expressiveness, with SLCC values showing minimal increase across intensity levels. We hypothesize that", + "bbox": [ + 88, + 719, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0ec0f429a3450ed7473f3dc589663e6129187108f86021318a921cd449759d16.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)Lv1SLCC / Δ (↓)
Lv2Lv3Avg
MEAD-3D--0.2300.24 / -0.30 / -0.39 / -0.42 / -
FaceFormer [11]59.63.2070.2990.08 / 0.160.07 / 0.230.07 / 0.320.06 / 0.36
+ MEAD-3D59.51.1390.1760.26 / 0.020.30 / 0.000.34 / 0.050.35 / 0.07
+ Ours rep.55.81.1140.3060.27 / 0.030.27 / 0.030.32 / 0.070.33 / 0.09
CodeTalker [46]60.73.2360.2940.02 / 0.220.03 / 0.270.03 / 0.360.02 / 0.40
+ MEAD-3D60.92.9540.1540.09 / 0.150.12 / 0.180.06 / 0.330.11 / 0.31
+ Ours rep.58.62.7050.2210.18 / 0.060.29 / 0.010.31 / 0.080.31 / 0.11
SelfTalk [21]53.43.3960.2940.14 / 0.100.14 / 0.160.17 / 0.220.15 / 0.27
+ MEAD-3D54.21.2380.2160.16 / 0.080.28 / 0.020.32 / 0.070.31 / 0.11
+ Ours rep.52.71.1920.2300.17 / 0.070.29 / 0.010.34 / 0.050.33 / 0.09
", + "bbox": [ + 514, + 268, + 903, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative results of lip synchronization on MEAD-3D test set. We evaluate the base models on the MEAD-3D test set. We also compute the level-wise SLCC to evaluate the expressive capability of the models.", + "bbox": [ + 511, + 402, + 906, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "this limitation arises, because VOCASET's smaller scale and intensity range restrict the model's ability to learn relationships between speech and lip intensity. To address this, we integrate MEAD-3D with VOCASET for training, aiming to boost expressiveness in the base models. This approach consistently improves SLCC across all intensity levels. However, simply adding MEAD-3D degrades lip synchronization metrics, except for expressiveness, i.e., nontrivial. To counterbalance this, we leverage our perceptual loss, which effectively mitigates the degradation introduced by MEAD-3D while improving expressiveness (see Fig. 6).", + "bbox": [ + 511, + 468, + 908, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 648, + 633, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper addresses challenges in existing 3D talking head generation models, which often overlook the true correspondence between speech and lip movements, making it difficult to accurately link lip movements with varying speech characteristics. To overcome this issue, we identify three essential aspects—Temporal Synchronization, Lip Readability, and Expressiveness—that influence the perceptual quality of lip movements, and develop specific metrics for each aspect. We introduce a speech-mesh synchronized representation that exhibits these emergent properties and adopt it as a perceptual loss. Our extensive analyses demonstrate that our perceptual loss consistently enhances models across three aspects. We believe that our defined aspects will guide future research in generating more realistic 3D talking heads, and our representation will serve as a key component.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21072", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This research was supported by a grant from KRAFTON AI, and was also partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub; No.RS-2023-00225630, Development of Artificial Intelligence for Text-based 3D Movie Generation; No. RS-2024-00457882, National AI Research Lab Project) and Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2024 (Project Name: Development of barrier-free experiential XR contents technology to improve accessibility to online activities for the physically disabled, Project Number: RS-2024-00396700, Contribution Rate: $25\\%$ ).", + "bbox": [ + 89, + 90, + 485, + 316 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 332, + 187, + 348 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 2, 3, 4, 6", + "[2] Mohamed Anwar, Bowen Shi, Vedanuj Goswami, Wei-Ning Hsu, Juan Pino, and Changhan Wang. Muavic: A multilingual audio-visual corpus for robust speech recognition and robust speech-to-text translation. In INTERSPEECH 2023, 2023. 3", + "[3] Helen L Bear and Richard Harvey. Phoneme-to-viseme mappings: the good, the bad, and the ugly. Speech Communication, 2017. 2, 3", + "[4] Donald J Berndt and James Clifford. Using dynamic time warping to find patterns in time series. In Proceedings of the 3rd international conference on knowledge discovery and data mining, 1994. 6", + "[5] Jeongsoo Choi, Se Jin Park, Minsu Kim, and Yong Man Ro. Av2av: Direct audio-visual speech to audio-visual speech translation with unified audio-visual speech representation. In CVPR, 2024. 3", + "[6] J. S. Chung and A. Zisserman. Out of time: automated lip sync in the wild. In Workshop on Multi-view Lip-reading, ACCV, 2016. 3", + "[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 3", + "[8] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In CVPR, 2019. 1, 2, 6, 7, 8", + "[9] Radek Daněček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. In SIGGRAPH Asia 2023 Conference Papers, pages 1-13, 2023. 1, 2", + "[10] Han EunGi, Oh Hyun-Bin, Kim Sung-Bin, Corentin Nivelet Etcheberry, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Enhancing speech-driven 3d facial animation with audio-visual guidance from lip reading expert. In Interspeech 2024, pages 2940–2944, 2024. 3" + ], + "bbox": [ + 93, + 358, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In CVPR, 2022. 2, 3, 7, 8", + "[12] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE TMM, 2010. 2, 6", + "[13] Panagiotis P. Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Spectre: Visual speech-informed perceptual 3d facial expression reconstruction from videos. In CVPRW, pages 5745-5755, 2023. 4, 6", + "[14] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 3", + "[15] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In The Eleventh International Conference on Learning Representations, 2023. 3, 4", + "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: a reference-free evaluation metric for image captioning. In EMNLP, 2021. 3", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2", + "[18] Jessica E Huber and Bharath Chandrasekaran. Effects of increasing sound pressure level on lip and jaw movement parameters and consistency in young adults. 2006. 2, 3, 4", + "[19] Eamonn J. Keogh and M. Pazzani. Derivative dynamic time warping. In In First SIAM International Conference on Data Mining, 2001. 6", + "[20] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5", + "[21] Ziqiao Peng, Yihao Luo, Yue Shi, Hao Xu, Xiangyu Zhu, Hongyan Liu, Jun He, and Zhaoxin Fan. Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In ACM MM, 2023. 1, 2, 3, 7, 8", + "[22] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[23] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, page 484-492. Association for Computing Machinery, 2020. 3", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. on Mach. Learn., pages 8748-8763. PMLR, 2021. 3", + "[25] Byron Reeves and David Voelker. Effects of audio-video asynchrony on viewer's memory, evaluation of content and detec" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21073", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion ability. In Research Report, Stanford University, 1993. 3", + "[26] George Retsinas, Panagiotis P Filntisis, Radek Danecek, Victoria F Abrevaya, Anastasios Roussos, Timo Bolkart, and Petros Maragos. 3d facial expressions through analysis-by-neural-synthesis. In CVPR, 2024. 6", + "[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In ICCV, 2021. 1, 3", + "[28] Richard Schulman. Articulatory dynamics of loud and normal speech. The Journal of the Acoustical Society of America, 85 (1):295-312, 1989. 2, 3, 4", + "[29] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR, 2022. 3", + "[30] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. arXiv preprint arXiv:2201.01763, 2022. 3", + "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2", + "[32] Wenfeng Song, Xuan Wang, Shi Zheng, Shuai Li, Aimin Hao, and Xia Hou. Talkingstyle: Personalized speech-driven 3d facial animation with style preservation. IEEE Transactions on Visualization and Computer Graphics, 2024. 3", + "[33] Stefan Stan, Kazi Injamamul Haque, and Zerrin Yumak. Facediffuser: Speech-driven 3d facial animation synthesis using diffusion. In Proceedings of the 16th ACM SIGGRAPH Conference on Motion, Interaction and Games, pages 1-11, 2023. 1, 2", + "[34] Trinity Suma, Birate Sonia, Kwame Agyemang Baffour, and Oyewole Oyekoya. The effects of avatar voice and facial expression intensity on emotional recognition and user perception. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 2, 3", + "[35] Licai Sun, Zheng Lian, Bin Liu, and Jianhua Tao. Hicmae: Hierarchical contrastive masked autoencoder for self-supervised audio-visual emotion recognition. Information Fusion, 108: 102382, 2024. 3, 4", + "[36] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-Jin Liu. Diffposetalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM TOG, 43(4), 2024. 2", + "[37] Kim Sung-Bin, Lee Chae-Yeon, Gihun Son, Oh Hyun-Bin, Janghoon Ju, Suekyeong Nam, and Tae-Hyun Oh. Multitalk: Enhancing 3d talking head generation across languages with multilingual video dataset. In Interspeech 2024, pages 1380-1384, 2024. 2, 3", + "[38] Kim Sung-Bin, Lee Hyun, Da Hye Hong, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Laughtalk: Expressive 3d talking head generation with laughter. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6404-6413, 2024. 2", + "[39] Hiroki Tanaka, Satoshi Nakamura, et al. The acceptability of virtual characters as social skills trainers: usability study. *JMIR human factors*, 9(1):e35358, 2022. 1" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Stephen M Tasko and Michael D McClean. Variations in articulatory movement with changes in speech task. 2004. 2, 3, 4", + "[41] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 3", + "[42] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-cap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918–17928, 2022. 3", + "[43] Argiro Vatakis and Charles Spence. Audiovisual synchrony perception for speech and music assessed using a temporal order judgment task. *Neuroscience letters*, 393(1):40-44, 2006. 2, 3, 4", + "[44] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In CVPR, pages 14653-14662, 2023. 3", + "[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, 2020. 2, 6", + "[46] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In CVPR, 2023. 1, 2, 3, 7, 8", + "[47] Karren Yang, Anurag Ranjan, Rick Chang, Raviteja Vemulapalli, and Oncel Tuzel. Probabilistic speech-driven 3d facial motion synthesis: New benchmarks, methods, and applications. In CVPR, 2024. 2, 3", + "[48] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In CVPR, pages 8652-8661, 2023. 3" + ], + "bbox": [ + 516, + 90, + 906, + 599 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21074", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/f0ca3209-a68f-41a0-a209-e41298a3de25_model.json b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/f0ca3209-a68f-41a0-a209-e41298a3de25_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f612fa220a723c3d6a6dc76a0b4d3b716a85c85a --- /dev/null +++ b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/f0ca3209-a68f-41a0-a209-e41298a3de25_model.json @@ -0,0 +1,1980 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.142, + 0.131, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Perceptually Accurate 3D Talking Head Generation: New Definitions, Speech-Mesh Representation, and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.203, + 0.905, + 0.223 + ], + "angle": 0, + "content": "Lee Chae-Yeon\\(^{1*}\\) Oh Hyun-Bin\\(^{2*}\\) Han EunGi\\(^{1}\\) Kim Sung-Bin\\(^{2}\\) Suekyeong Nam\\(^{3}\\) Tae-Hyun Oh\\(^{1,2,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.234, + 0.886, + 0.25 + ], + "angle": 0, + "content": "\\(^{1}\\)Grad. School of AI, POSTECH \\(^{2}\\)Dept. of Electrical Engineering, POSTECH \\(^{3}\\)KRAFTON \\(^{4}\\)School of Computing, KAIST" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.285, + 0.494, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.418, + 0.461, + 0.431 + ], + "angle": 0, + "content": "(a) Essential Criteria for Perceptually Accurate 3D Talking Head" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.287, + 0.895, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.418, + 0.859, + 0.432 + ], + "angle": 0, + "content": "(b) Conceptual Diagram of Desired Representation Space" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.437, + 0.908, + 0.522 + ], + "angle": 0, + "content": "Figure 1. What defines perceptually accurate lip movement for a speech signal? In this work, we define three criteria to assess perceptual alignment between speech and lip movements of 3D talking heads: Temporal Synchronization, Lip Readability, and Expressiveness (a). The motivational hypothesis is the existence of a desirable representation space that models and complies well with the three criteria between diverse speech characteristics and 3D facial movements, as illustrated in (b); where representations with the same phonemes are clustered, are sensitive to temporal synchronization, and follow a certain pattern as the speech intensity increases. Consequently, we build a rich speech-mesh synchronized representation space that exhibits the desirable properties." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.542, + 0.327, + 0.558 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.486, + 0.877 + ], + "angle": 0, + "content": "Recent advancements in speech-driven 3D talking head generation have made significant progress in lip synchronization. However, existing models still struggle to capture the perceptual alignment between varying speech characteristics and corresponding lip movements. In this work, we claim that three criteria—Temporal Synchronization, Lip Readability, and Expressiveness—are crucial for achieving perceptually accurate lip movements. Motivated by our hypothesis that a desirable representation space exists to meet these three criteria, we introduce a speech-mesh synchronized representation that captures intricate correspondences between speech signals and 3D face meshes. We found that our learned representation exhibits desirable characteristics, and we plug it into existing models as a perceptual loss to better align lip movements to the given speech. In addition, we utilize this representation as a perceptual metric and introduce two other physically grounded lip synchronization metrics to assess how well the generated 3D talking heads align with these three criteria. Experiments show that training 3D talking head generation models with our perceptual loss" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.544, + 0.909, + 0.59 + ], + "angle": 0, + "content": "significantly improve all three aspects of perceptually accurate lip synchronization. Codes and datasets are available at https://perceptual-3d-talking-head.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.615, + 0.645, + 0.631 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.643, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Speech-driven 3D talking head generation focuses on generating 3D facial movements synchronized with input speech signals. This plays a key role in enhancing communication within multimedia applications, such as virtual reality, entertainment, and education [39]. To provide users with a more realistic and immersive experience, it is crucial that the facial and lip movements of 3D avatars are synchronized with the various aspects of speech. This synchronization should be perceptually accurate from a human perspective, ensuring that the avatars' expressions are both natural and convincing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "While recent works in learning-based 3D talking head generation [8, 9, 21, 27, 33, 46] aim to enhance the lip synchronization capabilities, they commonly rely on minimizing the Mean Squared Error (MSE) loss between generated 3D facial motion and ground truth motion as a learning objective. This approach is practical, as it directly contributes to minimizing the Lip Vertex Error (LVE), a commonly used" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.887, + 0.28, + 0.901 + ], + "angle": 0, + "content": "*These authors contributed equally." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21065" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "metric that measures the MSE between the lip vertices of generated 3D facial motions and the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.125, + 0.486, + 0.322 + ], + "angle": 0, + "content": "Despite the improvements in the LVE metric, existing models still struggle to correlate lip movements with some speech characteristics, such as wider mouth openings as speech volume increases. These characteristics are not adequately captured by existing datasets [8, 12], which have limited ranges of facial motion patterns due to the small dataset scale and restricted intensity range. Moreover, relying on MSE and LVE is insufficient for learning or assessing perceptually plausible lip motions [21, 37], as it focuses solely on vertex-wise geometric differences and overlooks the true correspondence between the speech signals and lip movements. A lower MSE and LVE do not necessarily correspond to a more perceptually accurate lip movement." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.325, + 0.487, + 0.688 + ], + "angle": 0, + "content": "These observations raise critical questions: What defines perceptually accurate lip movement in response to a speech signal, and how can we enhance this accuracy? We draw inspiration from findings on human audio-visual perception: 1) Humans are sensitive to temporal asynchrony between speech and lip movements; slight discrepancies can disrupt the perception of natural synchronization [43]. 2) Humans rely on accurate viseme-phoneme correspondence when assessing lip-sync accuracy, expecting visual lip movements to match the spoken phonemes [3]. 3) There is a proportional increase in jaw and lip movements as speech intensity increases, contributing to the expressiveness perceived in natural speech [18, 28, 34, 40]. Through our human study, we reveal an intriguing finding: participants favor lip movements with intensity that corresponds to speech—even if they exceed the established maximum acceptable asynchrony [43] by twice—over those that are perfectly synchronized but lack expressive alignment (Table 1-[Right]). This reveals that humans are more sensitive to expressiveness than temporal synchronization when perceiving, highlighting the importance of expressiveness. Building upon these insights, we define three criteria that significantly impact the perceptual lip synchronization of 3D talking heads: Temporal Synchronization, Lip Readability, and Expressiveness (Fig. 1 (a))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Motivated by our hypothesis that a desirable representation space exists to meet these three criteria, we propose a speech-mesh synchronized and rich representation that captures the intricate correspondence between speech and 3D face mesh. We design a transformer-based architecture that maps time-sequenced speech and mesh inputs to a shared representation space. To effectively train this system, we employ a two-stage training method: we first develop a robust audio-visual speech representation using a large-scale 2D video dataset [1], which then serves as an anchor for learning a speech-mesh representation. We found that the first step is important and leads to emergent desirable properties (illustrated in Fig. 1 (b)) in the final representation. This sequential approach ensures that the model first establishes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.198 + ], + "angle": 0, + "content": "a space that captures a wide range of speech characteristics, and then extends to explore the relationships between speech and 3D face mesh across diverse speech intensities and facial movements. Adopting this representation, we introduce a plug-and-play perceptual loss adaptable to any existing 3D talking head generation models [11, 21, 46], enhancing the perceptual quality of 3D talking heads." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.909, + 0.319 + ], + "angle": 0, + "content": "Furthermore, to assess the three criteria, we introduce three metrics for each aspect. We leverage our learned representation as a perceptual metric, Perceptual Lip Readability Score (PLRS), to evaluate the perceptual lip readability of lip movements. Also, we propose two physically grounded lip synchronization metrics: Mean Temporal Misalignment (MTM) for temporal synchronization and Speech-Lip Intensity Correlation Coefficient (SLCC) for expressiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.909, + 0.455 + ], + "angle": 0, + "content": "Extensive experiments demonstrate that our perceptual loss significantly enhances all three aspects: Temporal Synchronization, Lip Readability, and Expressiveness, which are demonstrated across various metrics: existing metric, our newly proposed metrics, and human evaluations. We also find that incorporating an additional pseudo-dataset [45], which captures diverse ranges of speech and lip movement intensities, can further improve expressiveness. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.458, + 0.909, + 0.52 + ], + "angle": 0, + "content": "- Defining three aspects—Temporal Synchronization, Lip Readability, and Expressiveness—that affect the perceptual quality of 3D talking heads and proposing three evaluation metrics for these aspects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.909, + 0.57 + ], + "angle": 0, + "content": "- Constructing a speech-mesh representation space that captures rich and diverse correspondences between speech and lip movements." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.575, + 0.909, + 0.635 + ], + "angle": 0, + "content": "- Proposing a plug-in perceptual loss using the constructed speech-mesh representation and demonstrating improvements on existing metric, our newly proposed metrics, and human evaluations." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.458, + 0.909, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.657, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Speech-driven 3D talking head generation. Speech-driven 3D talking head generation aims to generate realistic 3D facial movements aligned with given speech. Among recent data-driven methods [8, 9, 11, 21, 22, 33, 46, 47], FaceFormer [11] introduces a transformer-based autoregressive model and leverages a pre-trained speech model to capture long-term audio context and past facial movements. CodeTalker [46] employs a VQ-VAE to construct a discrete facial motion space, addressing the over-smoothing problem. Diffusion models [17, 31] have also been demonstrated to be effective for 3D talking head generation [33, 36]. In addition to synthesizing neutral facial motions, several works extend the 3D talking head to express specific aspects, such as emotional expressions [9, 22], multilingual capabilities [37], or laughter [38]. Despite these advances, existing methods rely" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21066" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "on minimizing MSE loss without a clear definition of perceptually accurate lip movement, overlooking the multifaceted nature of lip synchronization. To address this, we define three critical aspects of lip synchronization and propose rich speech-mesh representation, along with its application as a perceptual loss in a plug-and-play manner, enhancing all three aspects of lip synchronization quality in existing 3D talking head generation models [11, 21, 46]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.218, + 0.487, + 0.595 + ], + "angle": 0, + "content": "Speech-face representation learning. Well-aligned representation spaces learned from large-scale datasets such as CLIP [24] and ImageBind [14] are valued for their scalability and versatility. These spaces enable a wide range of applications, including auxiliary loss [42], intermediate representation [41], and evaluation metrics [16]. With this context, audio-visual representation spaces trained specifically on speech and 2D face videos have been proposed [6, 15, 29, 35]. For instance, SyncNet [6], a CNN-based model learns to detect audio-visual temporal synchronization and has been applied to several tasks, such as active speaker detection [1, 7] and 2D talking head generation [23, 44, 48]. Similarly, a transformer-based AV-HuBERT [29] has demonstrated remarkable effectiveness in various tasks, including lip reading [30], audio-visual translation [5], and 2D talking head generation [44]. While there has been significant progress in the 2D domain, advancements in the 3D domain remain under-explored. Yang et al. [47] extends the SyncNet architecture to accommodate speech and 3D face meshes; however, its application is limited to evaluating 3D talking heads. In this work, we demonstrate the versatility of our representation space as a plug-in module to improve the perceptual accuracy of the existing speech-driven 3D talking head generation models [11, 21, 46], as well as to assess their performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Evaluation metrics for speech-driven 3D talking head. The prevalent evaluation metric, Lip Vertex Error (LVE) [27], measures the L2 distance between predicted lip vertices and ground truth. Additional metrics, such as Upper Face Dynamics Deviation (FDD) [46] and Lip Readability Percentage (LRP) [21], consider different regions of facial motion. These metrics, however, focus on vertex-wise geometric differences between the ground-truth 3D facial motions and neglect speech-related information. To incorporate both speech and 3D face mesh data for evaluation, MultiTalk [37] introduces an Audio-Visual Lip Readability (AVLR) metric, which assesses perceptual accuracy of lip readability using a pre-trained Audio-Visual Speech Recognition model [2]. Yet, AVLR relies on speech and 2D face video rendered from the 3D face mesh, which may not align with the 3D talking head domain. To address these limitations, we introduce novel and comprehensive evaluation metrics that focus on diverse aspects of lip synchronization: Mean Temporal Misalignment (MTM), Perceptual Lip Readability Score (PLRS), and Speech-Lip Intensity Correlation Coefficient (SLCC)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.907, + 0.126 + ], + "angle": 0, + "content": "3. Essential Criteria for Perceptually Accurate 3D Talking Head" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.133, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Generating a 3D talking head with lip movements that are perceptually accurate to human observers requires a clear understanding of the components that influence perceptual quality. Although existing works have focused on improving partial aspects of these talking head generation models, a comprehensive exploration for establishing a perceptually accurate 3D talking head model has barely been undertaken. Drawing inspirations from extensive research in existing works [10, 18, 21, 28, 34, 40, 43, 44] and our studies, we identify three fundamental criteria essential for achieving perceptually accurate lip movements in 3D talking heads." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.304, + 0.909, + 0.576 + ], + "angle": 0, + "content": "Temporal Synchronization. This alignment is particularly crucial in media involving human speech, where viewers expect lip movements to precisely match the corresponding speech in time. Temporal misalignment between speech and lip movements indeed distracts viewers, reduces user experiences, and negatively impacts audience perception [25]. Vatakis et al. [43] find that viewers are particularly sensitive to speech-lip asynchrony compared to other audio-visual asynchrony, such as music. They note that mismatches become noticeable when speech precedes lip movements by more than \\(50~\\mathrm{ms}\\) or follows them by more than \\(220~\\mathrm{ms}\\). These findings may hold the same in 3D talking head generation, where any misalignment between speech and lip movements can break the illusion of realism, leading to a diminished user experience and reducing the perceived authenticity of the virtual character. Thus, we define the temporal synchronization of the talking head as an important aspect of lip synchronization quality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.58, + 0.909, + 0.837 + ], + "angle": 0, + "content": "Lip Readability. Visemes (or lip movements) must correspond accurately to the speech phonemes to ensure the spoken words are visually intelligible [3]. This aspect is widely acknowledged as important in existing literature in both 2D [44] and 3D [10, 21] speech-driven talking head models, which leverage lip reading experts as an auxiliary guidance to improve the visual intelligibility of the spoken word. However, the mapping between speech and lip movements is not one-to-one, making lip readability challenging to define. For example, the size and shape of the opening mouth and the dynamic movement patterns of the lips in response to specific utterance differ at every moment [47] or per individual [32]. To capture this complexity, we define the lip readability within speech-mesh synchronized representation space learned from a large-scale dataset, capturing the comprehensive and nuanced correspondence between speech and lip movements." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Expressiveness. Speech conveys not only linguistic content but also varies in intensity. For instance, a speaker may express the same text softly or loudly, with jaw and lip movements proportionally increasing as speech intensity rises," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21067" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.293, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.322, + 0.105, + 0.48, + 0.201 + ], + "angle": 0, + "content": "
Samples
AB
Temp.✓×
Exp.×✓
Prefer (%)17.482.6
" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.22, + 0.485, + 0.304 + ], + "angle": 0, + "content": "Table 1. Human studies on alignment criteria. [Left] Preference scores (1-3) for 3D talking heads with varying lip movement intensities paired with different speech intensities. [Right] Human preference between (A) samples with precise timing but low expressiveness, and (B) samples with high expressiveness but 100ms asynchrony—twice the commonly accepted 50ms threshold [43]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.329, + 0.485, + 0.571 + ], + "angle": 0, + "content": "which contributes to perceived expressiveness in real-world face recognition [18, 28, 40]. To demonstrate that the positive correlation of human preference between the intensity of speech and lip movements also exists in 3D talking head field, we conduct a human study by presenting 3D talking heads with varied lip movement intensities paired with different speech intensities in Table 1-[Left]. Participants prefer the lip movements with the intensity that match the intensity of speech. Despite this distinct human preference for synchronization between the intensity of corresponding speech and lip movements, this aspect has barely been explored in the talking head generation field. We thus define the expressiveness in 3D talking head as the correlation between speech and lip movement intensity, which is crucial for establishing genuine talking heads and is expected to guide future research aimed at improving perceptual lip synchronization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.832 + ], + "angle": 0, + "content": "Focus of this work. Among the three criteria for perceptually accurate 3D talking head, we find that recent 3D talking head generation models achieve reasonable temporal alignment between speech and lip movements (see supplementary for a visualization of temporal synchronization). Furthermore, we design an A vs. B test, prompting participants to choose between two samples: (A) temporally synchronized one while lacking expressive synchronization, and (B) the other with expressive synchronization but with \\(100\\mathrm{ms}\\) asynchrony. Table 1-[Right] shows that users prefer sample B, suggesting that humans may prioritize expressive synchronization over strict temporal alignment. This insight directs our focus toward enhancing all three aspects to better capture perceptual realism in 3D talking head. Details of the human study are in the supplementary material. Before introducing metrics to assess the criteria, we present our synchronized representation for designing our perceptual loss." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.845, + 0.48, + 0.865 + ], + "angle": 0, + "content": "4. Speech-Mesh Synchronized Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We hypothesize that a desirable representation space exists to meet the three criteria defined in Sec. 3. Motivated by this" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.183 + ], + "angle": 0, + "content": "hypothesis, we develop a rich speech-mesh synchronized representation which captures the intricate correspondence between speech and the 3D face mesh. We found that our learned representation exhibits desirable properties, and we adopt it as a perceptual loss to improve the perceptual accuracy of existing models with respect to these three criteria." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.186, + 0.909, + 0.414 + ], + "angle": 0, + "content": "Overview. Directly learning a synchronized speech-mesh representation presents challenges due to the scarcity of speech-3D face mesh datasets. One potential solution is to construct pseudo-GT 3D face meshes using reconstruction models [13], although relying solely on pseudo-GT may not suffice for building a robust representation. To overcome this, we leverage the extensive knowledge from the speech-2D lip video representations and adapt this to accommodate 3D face meshes. Specifically, we propose a two-stage training process: in stage 1, we learn an audio-visual speech representation that accurately reflects lip movements from the unlabeled in-the-wild 2D synchronized talking face video dataset, LRS3 [1]. Subsequently, we leverage the pre-trained speech representation from stage 1, using it as the anchor space to learn a synchronized speech-mesh representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.909, + 0.569 + ], + "angle": 0, + "content": "Stage 1. Learning audio-visual speech representation. This stage aims to learn a rich speech-2D lip video representation that effectively captures the correlation between varying speech characteristics and lip dynamics. Motivated by prior works [15, 35], we extend the integration of masked autoencoder (MAE) and cross-modal contrastive learning to learn a synchronized speech-2D lip video representation using 2D videos. The architecture for stage 1 consists of two modality-specific encoders, a cross-modal fusion encoder, and two modality-specific decoders. (see Fig. 2-[Stage 1])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Given a speech and 2D lip video pair, \\((\\mathbf{X}_{\\mathbf{s}},\\mathbf{X}_{\\mathbf{v}})\\in D\\) we begin by patchifying and tokenizing speech spectrograms and video frames into speech and video tokens as \\(\\mathbf{S} = (\\mathbf{s}_1,\\dots ,\\mathbf{s}_N)\\) and \\(\\mathbf{V} = (\\mathbf{v}_1,\\dots ,\\mathbf{v}_M)\\), where \\(\\mathbf{s}_i,\\mathbf{v}_j\\in\\) \\(\\mathbb{R}^H\\). We then randomly mask out \\(P\\%\\) portion of tokens. The remaining unmasked speech tokens \\(\\mathbf{S}^{unmask}\\) and video tokens \\(\\mathbf{V}^{unmask}\\) are respectively fed into the speech encoder \\(E_{s}\\) and video encoder \\(E_{v}\\), each consisting of \\(N_{e}\\) transformer layers. Each encoder extracts uni-modal embeddings, \\(\\mathbf{Z}_{\\mathbf{s}}^{l} = E_{s}^{l}(\\mathbf{S}^{unmask})\\) and \\(\\mathbf{Z}_{\\mathbf{v}}^{l} = E_{v}^{l}(\\mathbf{V}^{unmask})\\), where \\(l\\in (1,\\ldots ,N_e)\\) denote the layer indices. Also, the mean pooled speech and video embeddings, \\(\\mathbf{c_s^l} = \\mathrm{MeanPool}(\\mathbf{Z_s^l})\\) and \\(\\mathbf{c_v^l} = \\mathrm{MeanPool}(\\mathbf{Z_v^l})\\), are derived from the corresponding uni-modal embeddings. Following the uni-modal encoders, we introduce a multi-modal fusion encoder \\(E_{sv}\\) to exploit complementary information from each extracted uni-modal embedding. The speech and video embeddings are jointly processed through this encoder, resulting in multi-modal fusion embeddings \\(\\mathbf{F_s}\\) and \\(\\mathbf{F_v}\\), i.e., \\([\\mathbf{F_s},\\mathbf{F_v}] = E_{sv}([\\mathbf{Z_s},\\mathbf{Z_v}])\\) where \\([,\\cdot ]\\) denotes concatenation. With the fusion embeddings \\(\\mathbf{F_s}\\) and \\(\\mathbf{F_v}\\), we utilize each modality's decoder to reconstruct the original signals. Specifically, we employ a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21068" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.085, + 0.496, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.084, + 0.853, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.293, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Pipeline of speech-mesh synchronized representation learning. We train our speech-mesh representation space in a two-stage manner. In the first stage, we learn a rich audio-visual representation in 2D domain to capture the synchronization between lip movement and speech. In the second stage, we train the 3D mesh encoder to align the 3D mesh space with the frozen speech space. As an application of our speech-mesh representation space, we propose a plug-in perceptual loss to 3D talking head models to enhance the quality of lip movements." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.361, + 0.484, + 0.468 + ], + "angle": 0, + "content": "speech decoder \\(D_{s}\\) and a video decoder \\(D_{v}\\), each consisting of \\(N_{d}\\) transformer layers. We pad each multi-modal fusion embedding with trainable masked tokens at their original positions, resulting in \\(\\mathbf{F}_{\\mathbf{s}}^{\\prime}\\) and \\(\\mathbf{F}_{\\mathbf{v}}^{\\prime}\\). Each decoder then reconstructs the respective signals; the reconstructed speech spectrogram and video tokens as \\(\\hat{\\mathbf{S}}^{mask} = D_s(\\mathbf{F}_{\\mathbf{s}}^{\\prime})\\) and \\(\\hat{\\mathbf{V}}^{mask} = D_v(\\mathbf{F}_{\\mathbf{v}}^{\\prime})\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.468, + 0.484, + 0.68 + ], + "angle": 0, + "content": "Training the model involves two objectives: learning the cross-modal alignment between the speech-2D lip video signals, while reconstructing their original signals. Given \\(B\\) speech-2D lip video token pairs in a batch, \\(\\{\\mathbf{S}_i,\\mathbf{V}_i\\}_{i = 1}^B\\), we treat the true speech-2D lip video pair as positive samples, while the others in the batch are considered negative samples. To encourage alignment between temporally synchronized speech-2D lip video instances, we utilize a cross-modal contrastive learning strategy, InfoNCE [20]. We maximize the cosine similarity between positive sample of the mean pooled speech embedding \\(\\mathbf{c}_{\\mathbf{s},i}^{l} = \\mathrm{MeanPool}(E_{s}^{l}(\\mathbf{S}_{i}^{\\mathrm{unmask}}))\\) and the corresponding synchronized mean pooled video embedding \\(\\mathbf{c}_{\\mathbf{v},i}^{l} = \\mathrm{MeanPool}(E_{v}^{l}(\\mathbf{V}_{i}^{\\mathrm{unmask}}))\\). We first define the speech-centric loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.686, + 0.484, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {V}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {v} , i} ^ {l} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {v} , j} ^ {l} / \\tau\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.484, + 0.793 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a temperature hyperparameter. Also, we make the objective symmetric by defining a video-centric loss as \\(\\mathcal{L}_{\\mathrm{V}\\rightarrow \\mathrm{S}}\\). We sum \\(\\mathcal{L}_{\\mathrm{S}\\rightarrow \\mathrm{V}}\\) and \\(\\mathcal{L}_{\\mathrm{V}\\rightarrow \\mathrm{S}}\\) across \\(L\\) selected encoder layers to obtain our final contrastive learning objective:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.799, + 0.484, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {I n f o N C E}} = \\sum_ {l \\in L} \\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {V}} + \\mathcal {L} _ {\\mathrm {V} \\rightarrow \\mathrm {S}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.827, + 0.484, + 0.858 + ], + "angle": 0, + "content": "For reconstruction, the model is trained in a self-supervised manner by minimizing the reconstruction loss \\(\\mathcal{L}_{\\mathrm{MAE}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.862, + 0.485, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M A E}} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left[ \\frac {\\sum \\| \\hat {\\mathbf {S}} _ {i} ^ {m a s k} - \\mathbf {S} _ {i} ^ {m a s k} \\| _ {F} ^ {2}}{| \\mathbf {S} _ {i} ^ {m a s k} |} + \\frac {\\sum \\| \\hat {\\mathbf {V}} _ {i} ^ {m a s k} - \\mathbf {V} _ {i} ^ {m a s k} \\| _ {F} ^ {2}}{| \\mathbf {V} _ {i} ^ {m a s k} |} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.907, + 0.405 + ], + "angle": 0, + "content": "where \\( |\\mathbf{S}_i^{mask}| \\) and \\( |\\mathbf{V}_i^{mask}| \\) denote the number of masked speech and video tokens, respectively. Overall, our objective function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.411, + 0.907, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M A E}} + \\lambda \\mathcal {L} _ {\\text {I n f o N C E}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.443 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the weight factor for cross-modal contrastive loss." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.444, + 0.908, + 0.55 + ], + "angle": 0, + "content": "We observe that the representation space trained with the transformer architecture and a rich and large-scale 2D face video dataset in this way already possesses the desired properties we pursue, illustrated in Fig. 1 (see supplementary for visualizations of pre-trained speech representation). Motivated by this, we transfer these emergent properties to the speech-mesh representation space as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.689 + ], + "angle": 0, + "content": "Stage 2. Learning speech-mesh representation. In this stage, we design a 3D mesh encoder that maps 3D face mesh to the speech representation pre-trained in stage 1. This pretrained speech representation, derived from rich speech-2D face video data, serves as a robust anchor space for learning the correlation between diverse speech characteristics and 3D facial motions. We use contrastive learning loss to align 3D facial motion embeddings with the anchored speech representations, as shown in Fig. 2-[Stage 2]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.911, + 0.858 + ], + "angle": 0, + "content": "Given a speech and 3D face mesh pair \\((\\mathbf{X}_{\\mathbf{s}},\\mathbf{X}_{\\mathbf{m}})\\in D\\) we patchify and tokenize speech spectrograms into speech tokens S, and map these into pre-trained uni-modal mean pooled speech embeddings \\(\\mathbf{c_s^l} = \\mathrm{MeanPool}(E_s^l (\\mathbf{S}))\\) . Similarly, we patchify 3D face mesh into mesh tokens M, and feed them into the 3D mesh encoder \\(E_{m}\\) consisting of \\(N_{e}\\) transformer layers to extract mesh embeddings \\(\\mathbf{Z}_{\\mathbf{m}}^{l} = E_{m}^{l}(\\mathbf{M})\\) and the corresponding mean-pooled mesh embedding \\(\\mathbf{c_m^l}\\) . For the learning objective, given \\(B\\) speech-3D face mesh token pairs in a batch, \\(\\{\\mathbf{S}_i,\\mathbf{M}_i\\}_{i = 1}^B\\) , we first define speech-anchored loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.863, + 0.907, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {M}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {m} , i} ^ {l} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {m} , j} ^ {l} / \\tau\\right)}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21069" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "as well as \\(\\mathcal{L}_{\\mathrm{M}\\to \\mathrm{S}}\\) which is the mesh-anchored loss. Similar to Eq. (2), we sum \\(\\mathcal{L}_{\\mathrm{S}\\rightarrow \\mathrm{M}}\\) and \\(\\mathcal{L}_{\\mathrm{M}\\rightarrow \\mathrm{S}}\\) across \\(L\\) selected encoder layers to train the 3D mesh encoder \\(E_{m}\\) while the speech encoder fixed. As a result, this two-stage training approach yields a robust speech-mesh representation, which significantly enhances the alignment between the speech and 3D face mesh modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.201, + 0.482, + 0.306 + ], + "angle": 0, + "content": "Adopting it as a perceptual loss. A key application of the speech-mesh representation learned through the two-stage training process is its use as a perceptual loss to enhance the perceptual accuracy of the 3D talking head model (see Fig. 2). Leveraging this representation as a perceptual loss ensures that the generated lip movements are perceptually accurate and aligned well with the speech." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.482, + 0.353 + ], + "angle": 0, + "content": "Given \\(B\\) speech and generated 3D face mesh pairs in a batch, \\(\\{\\mathbf{S}_i,\\hat{\\mathbf{M}}_i\\}_{i = 1}^B\\), we define our perceptual loss with the symmetric InfoNCE loss (Eq. (2)) as" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.362, + 0.483, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p e r c p}} = \\sum_ {l \\in L} \\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {M}} + \\mathcal {L} _ {\\mathrm {M} \\rightarrow \\mathrm {S}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.485, + 0.464 + ], + "angle": 0, + "content": "Our perceptual loss encourages synchronized speech and mesh embeddings to pull closer together, while unsynchronized ones to push apart. The effectiveness and analysis of our perceptual loss will be discussed in later sections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.476, + 0.276, + 0.491 + ], + "angle": 0, + "content": "5. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.577 + ], + "angle": 0, + "content": "In this section, we describe our proposed evaluation metrics that assess each criterion impacting the quality of 3D lip accuracy, as discussed in Sec. 3. Here, we outline the high-level concepts of our proposed metrics. Additional details and experiments are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.852 + ], + "angle": 0, + "content": "Mean Temporal Misalignment (MTM). To measure the temporal discrepancy between speech and corresponding lip movements, temporal correspondence annotations, such as onset times for each modality, would typically be required. As a proxy, we determine temporal correspondence and measure temporal discrepancies between the ground truth and predicted lip vertex displacement sequences by using Derivative Dynamic Time Warping (DDTW) [19], which robustly identifies local structural similarities compared to standard Dynamic Time Warping (DTW) [4]. For simplicity, we focus on the central vertices of the upper and lower lips when extracting displacement sequences, and use local extrema in the DDTW process to measure temporal misalignment by pinpointing precise time steps for mouth opening and closing events. Consequently, Mean Temporal Misalignment (MTM) \\(\\overline{\\Delta t}\\) is defined as \\(\\overline{\\Delta t} = \\frac{1}{K}\\sum_{k=1}^{K}\\Delta t_k\\), where \\(K\\) is the total number of video clips and \\(\\Delta t_k\\) is the averaged temporal misalignment of the \\(k\\)-th video clip." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Perceptual Lip Readability Score (PLRS). While Lip Vertex Error (LVE) measures the accuracy of generated lip articulations against the ground truth, it does not fully assess" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "whether lip movements are perceptually aligned with the given speech. To address this, we leverage our speech-mesh representation in Sec. 4 as a perceptual lip readability evaluator. We compute the perceptual alignment using the cosine similarity of the mean pooled speech and mesh embeddings. Since this representation has learned a rich distribution of speech correspondences across various facial movements, our metric correlates highly with human perception, measuring perceptual lip movement alignment more accurately than LVE (see supplementary for the human study on metrics)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.246, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Speech and Lip Intensity Correlation Coefficient (SLCC). As shown in Table 1-[Left], humans prefer aligned intensity between speech and lip movement. Thus, the intensity of generated lip movements should positively correlate with the corresponding input speech. To quantify this, we define the Speech and Lip Correlation Coefficient \\( r_{SL} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.348, + 0.907, + 0.378 + ], + "angle": 0, + "content": "\\[\nr _ {S L} = \\frac {\\sum_ {k = 1} ^ {K} (\\mathrm {S I} _ {k} - \\bar {\\mathrm {S I}}) (\\mathrm {L I} _ {k} - \\bar {\\mathrm {L I}})}{\\sqrt {\\sum_ {k = 1} ^ {K} (\\mathrm {S I} _ {k} - \\bar {\\mathrm {S I}}) ^ {2}} \\sqrt {\\sum_ {k = 1} ^ {K} (\\mathrm {L I} _ {k} - \\bar {\\mathrm {L I}}) ^ {2}}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.387, + 0.907, + 0.493 + ], + "angle": 0, + "content": "where \\(\\mathrm{SI}_k\\) and \\(\\mathrm{LI}_k\\) denote Speech (SI) and Lip Intensity (LI), respectively, \\(\\bar{\\mathrm{SI}} = \\frac{1}{K}\\sum_{k = 1}^{K}\\mathrm{SI}_k\\) and \\(\\bar{\\mathrm{LI}} = \\frac{1}{K}\\sum_{k = 1}^{K}\\mathrm{LI}_k\\). We define SI using speech loudness, specifically the z-normalized Root Mean Square (RMS) value, which is a widely accepted measure of speech intensity in signal processing. To define LI, we measure the averaged lip displacement value of a video clip, followed by the z-normalization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.507, + 0.646, + 0.525 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.907, + 0.594 + ], + "angle": 0, + "content": "We first outline the evaluation setup, and then present thorough analyses of the experimental results. Due to the space limitation, we present more implementation details and additional experiments in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.722, + 0.62 + ], + "angle": 0, + "content": "6.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Datasets. Most of the existing speech-driven 3D talking head generation methods rely on VOCASET [8] and BIWI [12] to train and test the models. However, these datasets have limited ranges of facial motion patterns due to the small dataset scale and restricted intensity range, which restricts their ability to fully capture the intricate relationship between speech and 3D face mesh. To address the lack of training dataset, we construct two large-scale speech-3d face mesh benchmark datasets, LRS3-3D and MEAD-3D, by processing LRS3 and MEAD videos using two monocular face reconstruction methods: SPECTRE [13] for LRS3 [1], which ensures accurate lip movements, and SMIRK [26] for MEAD [45], which captures diverse speech and lip movement intensities. We use LRS3 in the first stage and LRS3-3D in the second stage to train our speech-mesh synchronized representation. Then, for base model training and evaluations, we adopt two configurations: (1) training and testing with VOCASET, in line with existing work, and (2)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21070" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.227 + ], + "angle": 0, + "content": "
MethodTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)SLCC / Δ (↓)
VOCASET--0.4090.34 / -
FaceFormer [11]53.63.3570.3680.26 / 0.08
+ Ours rep.52.23.0910.4630.37 / 0.03
CodeTalker [46]61.83.7000.3810.38 / 0.04
+ Ours rep.60.93.5790.3880.35 / 0.01
SelfTalk [21]50.12.9710.4140.41 / 0.07
+ Ours rep.49.22.9240.4180.35 / 0.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.228, + 0.485, + 0.339 + ], + "angle": 0, + "content": "Table 2. Quantitative results of lip synchronization on VO-CASET [8] test set. We evaluate the base models on our proposed lip synchronization metrics. We denote \\(\\Delta\\) as the difference in SLCC between the model and those measured on the data distribution. A lower \\(\\Delta\\) indicates the model more closely represents the intensity correlation of the dataset. We demonstrate the effectiveness of our representation in consistently enhancing all three aspects of lip synchronization." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.348, + 0.484, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.511 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of the effectiveness of our perceptual loss for lip readability. Our perceptual loss guides baselines [11, 21, 46] to generate perceptually accurate lip movements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.556 + ], + "angle": 0, + "content": "combining MEAD-3D with VOCASET during training to endow expressiveness and testing on MEAD-3D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.559, + 0.484, + 0.605 + ], + "angle": 0, + "content": "Base methods. We use three state-of-the-art 3D talking head generation models [11, 21, 46] to evaluate the effectiveness of our perceptual loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.485, + 0.73 + ], + "angle": 0, + "content": "Metrics. To comprehensively evaluate the three aspects of lip synchronization, we assess MTM, PLRS, and SLCC, corresponding to temporal synchronization, lip readability, and expressiveness, respectively. Additionally, we compute the level-wise SLCC for the MEAD-3D test set, which includes three distinct emotional intensity levels, to evaluate the expressive capability of 3D talking head generation models. We also measure LVE as part of the lip readability evaluation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.742, + 0.396, + 0.759 + ], + "angle": 0, + "content": "6.2. Experimental Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.825 + ], + "angle": 0, + "content": "We conduct evaluations to assess the effectiveness of our speech-mesh synchronized representation and the incorporation of an expressive speech-3D face mesh paired dataset (i.e., MEAD-3D) in enhancing the three criteria." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "How well do existing 3D talking head models achieve lip synchronization in all three aspects? The results are summarized in Table 2. For temporal synchronization, most base models achieve MTM values between 50 and 60ms, indicating performance close to the acceptable asynchrony" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.089, + 0.895, + 0.252 + ], + "angle": 0, + "content": "
MethodPerceptual LossTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)SLCC / Δ (↓)
VOCASET---0.4090.34 / -
FaceFormer [11]X53.63.3570.3680.26 / 0.08
3D SyncNet55.63.3160.4350.38 / 0.04
Ours w/o 2D prior55.33.2780.4000.42 / 0.08
Ours w/ 2D prior52.23.0910.4630.37 / 0.03
CodeTalker [46]X61.83.7000.3810.38 / 0.04
3D SyncNet59.64.3190.3790.14 / 0.20
Ours w/o 2D prior55.93.5790.3740.23 / 0.11
Ours w/ 2D prior60.93.5790.3880.35 / 0.01
SelfTalk [21]X50.12.9710.4140.41 / 0.07
3D SyncNet49.52.9410.4050.35 / 0.01
Ours w/o 2D prior54.43.1490.4170.39 / 0.05
Ours w/ 2D prior49.22.9240.4180.35 / 0.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.253, + 0.905, + 0.308 + ], + "angle": 0, + "content": "Table 3. Ablation study on architectural choice and 2D prior knowledge. We validate the effectiveness of the transformer-based architecture and curriculum learning with a pre-trained 2D speech representation by ablating them from our proposed representation." + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.321, + 0.866, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.451, + 0.906, + 0.507 + ], + "angle": 0, + "content": "Figure 4. t-SNE plot of ablation study. We plot the t-SNE graph for each perceptual critic model. We represent the features with same phoneme as same color. Squared and circled points denote mesh and speech features from each representation, respectively." + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.512, + 0.861, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.649, + 0.908, + 0.718 + ], + "angle": 0, + "content": "Figure 5. Behaviors of our representation in temporal and expressiveness sensitivity. We demonstrate the effectiveness of our representation in temporal synchronization and expressiveness using a cosine similarity graph and speech feature plots, respectively. We color the point as low, medium, and high intensity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.809 + ], + "angle": 0, + "content": "threshold. Regarding lip readability, SelfTalk [21] achieves the best performance, while FaceFormer [11] has the lowest PLRS score. Furthermore, CodeTalker [46] demonstrates the closest SLCC values to the ground truth VOCASET mesh, while FaceFormer exhibits the highest SLCC discrepancy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.856 + ], + "angle": 0, + "content": "Does our speech-mesh representation improve lip synchronization? Yes. Table 2 and Fig. 3 show consistent improvements in the three aspects with our perceptual loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "What makes our speech-mesh representation have lip synchronization ability? We hypothesize that the transformer-based architecture and curriculum learning with a pre-trained" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21071" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.088, + 0.845, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.232, + 0.907, + 0.261 + ], + "angle": 0, + "content": "Figure 6. Qualitative results for the expressiveness. Given high and low intensity levels of speech, models trained on both MEAD-3D and VOCASET show more expressive lip movements compared to those trained on VOCASET alone, and even better with our perceptual loss." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.272, + 0.486, + 0.56 + ], + "angle": 0, + "content": "audio-visual speech representation contribute significantly to improved lip synchronization. To validate this, we conduct ablation studies, summarized in Table 3, examining the roles of pre-trained speech representation and architectural design. Without the pre-trained speech representation, base models show lower performance across metrics, and CNN-based architectures, 3D SyncNet, do not clearly show effectiveness, while ours with 2D prior consistently show improvement. The t-SNE plots of perceptual critic models in Fig. 4 show that, in (a) 3D SyncNet, speech and mesh features corresponding to the same phoneme group are separated. In (b) ours without 2D priors, speech and mesh features lack separation and appear scattered without clustering. Our representation (c), notably, tends to form more distinct clusters according to phonemes, with vowels and consonants grouped closely, potentially contributing to enhancing lip readability. We also observe a directional progression in the feature space, shifting from phonemes with mouth opening \\((e.g., /aj/\\) to those with mouth closing \\((e.g., /f/\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.484, + 0.701 + ], + "angle": 0, + "content": "We also examine our representation's performance in other aspects, temporal synchronization and expressiveness. Figure 5-(a) demonstrates temporal sensitivity, as cosine similarity drops when temporal misalignment is introduced between input speech and 3D face mesh. In Fig. 5-(b), we plot speech features at varying speech intensities, showing a directional trend as intensity increases from lowest to highest. Figures 4 and 5 imply that our representation holds favorable properties discussed in Fig. 1 for the three criteria." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Can we unlock the expressive power of 3D talking heads?" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Likely. Since VOCASET [8] lacks the range of diverse speech and lip movement intensities, evaluating expressive power requires testing on a dataset with a broader range of intensities. To study this, we examine the expressiveness of the base models on the MEAD-3D dataset, which includes three intensity levels. We assess SLCC at each intensity level to evaluate expressiveness and identify any expressiveness bounds as the level of intensity increases. We first test the base models trained on VOCASET [8] against the MEAD-3D test set. Table 4 shows these models demonstrate limited expressiveness, with SLCC values showing minimal increase across intensity levels. We hypothesize that" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.27, + 0.905, + 0.404 + ], + "angle": 0, + "content": "
MethodTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)Lv1SLCC / Δ (↓)
Lv2Lv3Avg
MEAD-3D--0.2300.24 / -0.30 / -0.39 / -0.42 / -
FaceFormer [11]59.63.2070.2990.08 / 0.160.07 / 0.230.07 / 0.320.06 / 0.36
+ MEAD-3D59.51.1390.1760.26 / 0.020.30 / 0.000.34 / 0.050.35 / 0.07
+ Ours rep.55.81.1140.3060.27 / 0.030.27 / 0.030.32 / 0.070.33 / 0.09
CodeTalker [46]60.73.2360.2940.02 / 0.220.03 / 0.270.03 / 0.360.02 / 0.40
+ MEAD-3D60.92.9540.1540.09 / 0.150.12 / 0.180.06 / 0.330.11 / 0.31
+ Ours rep.58.62.7050.2210.18 / 0.060.29 / 0.010.31 / 0.080.31 / 0.11
SelfTalk [21]53.43.3960.2940.14 / 0.100.14 / 0.160.17 / 0.220.15 / 0.27
+ MEAD-3D54.21.2380.2160.16 / 0.080.28 / 0.020.32 / 0.070.31 / 0.11
+ Ours rep.52.71.1920.2300.17 / 0.070.29 / 0.010.34 / 0.050.33 / 0.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.404, + 0.907, + 0.46 + ], + "angle": 0, + "content": "Table 4. Quantitative results of lip synchronization on MEAD-3D test set. We evaluate the base models on the MEAD-3D test set. We also compute the level-wise SLCC to evaluate the expressive capability of the models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.909, + 0.637 + ], + "angle": 0, + "content": "this limitation arises, because VOCASET's smaller scale and intensity range restrict the model's ability to learn relationships between speech and lip intensity. To address this, we integrate MEAD-3D with VOCASET for training, aiming to boost expressiveness in the base models. This approach consistently improves SLCC across all intensity levels. However, simply adding MEAD-3D degrades lip synchronization metrics, except for expressiveness, i.e., nontrivial. To counterbalance this, we leverage our perceptual loss, which effectively mitigates the degradation introduced by MEAD-3D while improving expressiveness (see Fig. 6)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.634, + 0.665 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This paper addresses challenges in existing 3D talking head generation models, which often overlook the true correspondence between speech and lip movements, making it difficult to accurately link lip movements with varying speech characteristics. To overcome this issue, we identify three essential aspects—Temporal Synchronization, Lip Readability, and Expressiveness—that influence the perceptual quality of lip movements, and develop specific metrics for each aspect. We introduce a speech-mesh synchronized representation that exhibits these emergent properties and adopt it as a perceptual loss. Our extensive analyses demonstrate that our perceptual loss consistently enhances models across three aspects. We believe that our defined aspects will guide future research in generating more realistic 3D talking heads, and our representation will serve as a key component." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21072" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.318 + ], + "angle": 0, + "content": "Acknowledgments. This research was supported by a grant from KRAFTON AI, and was also partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub; No.RS-2023-00225630, Development of Artificial Intelligence for Text-based 3D Movie Generation; No. RS-2024-00457882, National AI Research Lab Project) and Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2024 (Project Name: Development of barrier-free experiential XR contents technology to improve accessibility to online activities for the physically disabled, Project Number: RS-2024-00396700, Contribution Rate: \\(25\\%\\))." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.333, + 0.188, + 0.349 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.359, + 0.486, + 0.412 + ], + "angle": 0, + "content": "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.416, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[2] Mohamed Anwar, Bowen Shi, Vedanuj Goswami, Wei-Ning Hsu, Juan Pino, and Changhan Wang. Muavic: A multilingual audio-visual corpus for robust speech recognition and robust speech-to-text translation. In INTERSPEECH 2023, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[3] Helen L Bear and Richard Harvey. Phoneme-to-viseme mappings: the good, the bad, and the ugly. Speech Communication, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[4] Donald J Berndt and James Clifford. Using dynamic time warping to find patterns in time series. In Proceedings of the 3rd international conference on knowledge discovery and data mining, 1994. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.484, + 0.628 + ], + "angle": 0, + "content": "[5] Jeongsoo Choi, Se Jin Park, Minsu Kim, and Yong Man Ro. Av2av: Direct audio-visual speech to audio-visual speech translation with unified audio-visual speech representation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[6] J. S. Chung and A. Zisserman. Out of time: automated lip sync in the wild. In Workshop on Multi-view Lip-reading, ACCV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[8] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In CVPR, 2019. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[9] Radek Daněček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. In SIGGRAPH Asia 2023 Conference Papers, pages 1-13, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Han EunGi, Oh Hyun-Bin, Kim Sung-Bin, Corentin Nivelet Etcheberry, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Enhancing speech-driven 3d facial animation with audio-visual guidance from lip reading expert. In Interspeech 2024, pages 2940–2944, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.359, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[11] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In CVPR, 2022. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[12] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE TMM, 2010. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[13] Panagiotis P. Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Spectre: Visual speech-informed perceptual 3d facial expression reconstruction from videos. In CVPRW, pages 5745-5755, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[14] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.907, + 0.374 + ], + "angle": 0, + "content": "[15] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In The Eleventh International Conference on Learning Representations, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.907, + 0.418 + ], + "angle": 0, + "content": "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: a reference-free evaluation metric for image captioning. In EMNLP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.459 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[18] Jessica E Huber and Bharath Chandrasekaran. Effects of increasing sound pressure level on lip and jaw movement parameters and consistency in young adults. 2006. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[19] Eamonn J. Keogh and M. Pazzani. Derivative dynamic time warping. In In First SIAM International Conference on Data Mining, 2001. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.549, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[20] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.907, + 0.646 + ], + "angle": 0, + "content": "[21] Ziqiao Peng, Yihao Luo, Yue Shi, Hao Xu, Xiangyu Zhu, Hongyan Liu, Jun He, and Zhaoxin Fan. Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In ACM MM, 2023. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.649, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[22] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[23] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, page 484-492. Association for Computing Machinery, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.907, + 0.87 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. on Mach. Learn., pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[25] Byron Reeves and David Voelker. Effects of audio-video asynchrony on viewer's memory, evaluation of content and detec" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21073" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "tion ability. In Research Report, Stanford University, 1993. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[26] George Retsinas, Panagiotis P Filntisis, Radek Danecek, Victoria F Abrevaya, Anastasios Roussos, Timo Bolkart, and Petros Maragos. 3d facial expressions through analysis-by-neural-synthesis. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.484, + 0.235 + ], + "angle": 0, + "content": "[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In ICCV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.483, + 0.279 + ], + "angle": 0, + "content": "[28] Richard Schulman. Articulatory dynamics of loud and normal speech. The Journal of the Acoustical Society of America, 85 (1):295-312, 1989. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.282, + 0.483, + 0.324 + ], + "angle": 0, + "content": "[29] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.327, + 0.483, + 0.368 + ], + "angle": 0, + "content": "[30] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. arXiv preprint arXiv:2201.01763, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.37, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[32] Wenfeng Song, Xuan Wang, Shi Zheng, Shuai Li, Aimin Hao, and Xia Hou. Talkingstyle: Personalized speech-driven 3d facial animation with style preservation. IEEE Transactions on Visualization and Computer Graphics, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.458, + 0.483, + 0.526 + ], + "angle": 0, + "content": "[33] Stefan Stan, Kazi Injamamul Haque, and Zerrin Yumak. Facediffuser: Speech-driven 3d facial animation synthesis using diffusion. In Proceedings of the 16th ACM SIGGRAPH Conference on Motion, Interaction and Games, pages 1-11, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.529, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[34] Trinity Suma, Birate Sonia, Kwame Agyemang Baffour, and Oyewole Oyekoya. The effects of avatar voice and facial expression intensity on emotional recognition and user perception. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[35] Licai Sun, Zheng Lian, Bin Liu, and Jianhua Tao. Hicmae: Hierarchical contrastive masked autoencoder for self-supervised audio-visual emotion recognition. Information Fusion, 108: 102382, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.483, + 0.714 + ], + "angle": 0, + "content": "[36] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-Jin Liu. Diffposetalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM TOG, 43(4), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[37] Kim Sung-Bin, Lee Chae-Yeon, Gihun Son, Oh Hyun-Bin, Janghoon Ju, Suekyeong Nam, and Tae-Hyun Oh. Multitalk: Enhancing 3d talking head generation across languages with multilingual video dataset. In Interspeech 2024, pages 1380-1384, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[38] Kim Sung-Bin, Lee Hyun, Da Hye Hong, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Laughtalk: Expressive 3d talking head generation with laughter. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6404-6413, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39] Hiroki Tanaka, Satoshi Nakamura, et al. The acceptability of virtual characters as social skills trainers: usability study. *JMIR human factors*, 9(1):e35358, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.133 + ], + "angle": 0, + "content": "[40] Stephen M Tasko and Michael D McClean. Variations in articulatory movement with changes in speech task. 2004. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[41] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.246 + ], + "angle": 0, + "content": "[42] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-cap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918–17928, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[43] Argiro Vatakis and Charles Spence. Audiovisual synchrony perception for speech and music assessed using a temporal order judgment task. *Neuroscience letters*, 393(1):40-44, 2006. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.906, + 0.359 + ], + "angle": 0, + "content": "[44] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In CVPR, pages 14653-14662, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.907, + 0.473 + ], + "angle": 0, + "content": "[46] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In CVPR, 2023. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.907, + 0.529 + ], + "angle": 0, + "content": "[47] Karren Yang, Anurag Ranjan, Rick Chang, Raviteja Vemulapalli, and Oncel Tuzel. Probabilistic speech-driven 3d facial motion synthesis: New benchmarks, methods, and applications. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.907, + 0.6 + ], + "angle": 0, + "content": "[48] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In CVPR, pages 8652-8661, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21074" + } + ] +] \ No newline at end of file diff --git a/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/f0ca3209-a68f-41a0-a209-e41298a3de25_origin.pdf b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/f0ca3209-a68f-41a0-a209-e41298a3de25_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f147f76032eac3c730a5e2ae30f61c6c2a3b08e6 --- /dev/null +++ b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/f0ca3209-a68f-41a0-a209-e41298a3de25_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:854a85e4e079fa3b4457c6d6aad28fb960287e659fe84b3b932e146faf225a9a +size 4877540 diff --git a/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/full.md b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8ae116388edaf7cd879e99ff7fe5407e69d43f22 --- /dev/null +++ b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/full.md @@ -0,0 +1,277 @@ +# Perceptually Accurate 3D Talking Head Generation: New Definitions, Speech-Mesh Representation, and Evaluation Metrics + +Lee Chae-Yeon $^{1*}$ Oh Hyun-Bin $^{2*}$ Han EunGi $^{1}$ Kim Sung-Bin $^{2}$ Suekyeong Nam $^{3}$ Tae-Hyun Oh $^{1,2,4}$ + +$^{1}$ Grad. School of AI, POSTECH $^{2}$ Dept. of Electrical Engineering, POSTECH $^{3}$ KRAFTON $^{4}$ School of Computing, KAIST + +![](images/eb6b0c40d41458ddd681a07cd73458212aa53cb72f5c6642970e29e1aa81b450.jpg) +(a) Essential Criteria for Perceptually Accurate 3D Talking Head + +![](images/c154fed79a46522ce17c99ae75dba0cec284ca566baa14e802bea69d8a08887a.jpg) +(b) Conceptual Diagram of Desired Representation Space +Figure 1. What defines perceptually accurate lip movement for a speech signal? In this work, we define three criteria to assess perceptual alignment between speech and lip movements of 3D talking heads: Temporal Synchronization, Lip Readability, and Expressiveness (a). The motivational hypothesis is the existence of a desirable representation space that models and complies well with the three criteria between diverse speech characteristics and 3D facial movements, as illustrated in (b); where representations with the same phonemes are clustered, are sensitive to temporal synchronization, and follow a certain pattern as the speech intensity increases. Consequently, we build a rich speech-mesh synchronized representation space that exhibits the desirable properties. + +# Abstract + +Recent advancements in speech-driven 3D talking head generation have made significant progress in lip synchronization. However, existing models still struggle to capture the perceptual alignment between varying speech characteristics and corresponding lip movements. In this work, we claim that three criteria—Temporal Synchronization, Lip Readability, and Expressiveness—are crucial for achieving perceptually accurate lip movements. Motivated by our hypothesis that a desirable representation space exists to meet these three criteria, we introduce a speech-mesh synchronized representation that captures intricate correspondences between speech signals and 3D face meshes. We found that our learned representation exhibits desirable characteristics, and we plug it into existing models as a perceptual loss to better align lip movements to the given speech. In addition, we utilize this representation as a perceptual metric and introduce two other physically grounded lip synchronization metrics to assess how well the generated 3D talking heads align with these three criteria. Experiments show that training 3D talking head generation models with our perceptual loss + +significantly improve all three aspects of perceptually accurate lip synchronization. Codes and datasets are available at https://perceptual-3d-talking-head.github.io/. + +# 1. Introduction + +Speech-driven 3D talking head generation focuses on generating 3D facial movements synchronized with input speech signals. This plays a key role in enhancing communication within multimedia applications, such as virtual reality, entertainment, and education [39]. To provide users with a more realistic and immersive experience, it is crucial that the facial and lip movements of 3D avatars are synchronized with the various aspects of speech. This synchronization should be perceptually accurate from a human perspective, ensuring that the avatars' expressions are both natural and convincing. + +While recent works in learning-based 3D talking head generation [8, 9, 21, 27, 33, 46] aim to enhance the lip synchronization capabilities, they commonly rely on minimizing the Mean Squared Error (MSE) loss between generated 3D facial motion and ground truth motion as a learning objective. This approach is practical, as it directly contributes to minimizing the Lip Vertex Error (LVE), a commonly used + +metric that measures the MSE between the lip vertices of generated 3D facial motions and the ground truth. + +Despite the improvements in the LVE metric, existing models still struggle to correlate lip movements with some speech characteristics, such as wider mouth openings as speech volume increases. These characteristics are not adequately captured by existing datasets [8, 12], which have limited ranges of facial motion patterns due to the small dataset scale and restricted intensity range. Moreover, relying on MSE and LVE is insufficient for learning or assessing perceptually plausible lip motions [21, 37], as it focuses solely on vertex-wise geometric differences and overlooks the true correspondence between the speech signals and lip movements. A lower MSE and LVE do not necessarily correspond to a more perceptually accurate lip movement. + +These observations raise critical questions: What defines perceptually accurate lip movement in response to a speech signal, and how can we enhance this accuracy? We draw inspiration from findings on human audio-visual perception: 1) Humans are sensitive to temporal asynchrony between speech and lip movements; slight discrepancies can disrupt the perception of natural synchronization [43]. 2) Humans rely on accurate viseme-phoneme correspondence when assessing lip-sync accuracy, expecting visual lip movements to match the spoken phonemes [3]. 3) There is a proportional increase in jaw and lip movements as speech intensity increases, contributing to the expressiveness perceived in natural speech [18, 28, 34, 40]. Through our human study, we reveal an intriguing finding: participants favor lip movements with intensity that corresponds to speech—even if they exceed the established maximum acceptable asynchrony [43] by twice—over those that are perfectly synchronized but lack expressive alignment (Table 1-[Right]). This reveals that humans are more sensitive to expressiveness than temporal synchronization when perceiving, highlighting the importance of expressiveness. Building upon these insights, we define three criteria that significantly impact the perceptual lip synchronization of 3D talking heads: Temporal Synchronization, Lip Readability, and Expressiveness (Fig. 1 (a)). + +Motivated by our hypothesis that a desirable representation space exists to meet these three criteria, we propose a speech-mesh synchronized and rich representation that captures the intricate correspondence between speech and 3D face mesh. We design a transformer-based architecture that maps time-sequenced speech and mesh inputs to a shared representation space. To effectively train this system, we employ a two-stage training method: we first develop a robust audio-visual speech representation using a large-scale 2D video dataset [1], which then serves as an anchor for learning a speech-mesh representation. We found that the first step is important and leads to emergent desirable properties (illustrated in Fig. 1 (b)) in the final representation. This sequential approach ensures that the model first establishes + +a space that captures a wide range of speech characteristics, and then extends to explore the relationships between speech and 3D face mesh across diverse speech intensities and facial movements. Adopting this representation, we introduce a plug-and-play perceptual loss adaptable to any existing 3D talking head generation models [11, 21, 46], enhancing the perceptual quality of 3D talking heads. + +Furthermore, to assess the three criteria, we introduce three metrics for each aspect. We leverage our learned representation as a perceptual metric, Perceptual Lip Readability Score (PLRS), to evaluate the perceptual lip readability of lip movements. Also, we propose two physically grounded lip synchronization metrics: Mean Temporal Misalignment (MTM) for temporal synchronization and Speech-Lip Intensity Correlation Coefficient (SLCC) for expressiveness. + +Extensive experiments demonstrate that our perceptual loss significantly enhances all three aspects: Temporal Synchronization, Lip Readability, and Expressiveness, which are demonstrated across various metrics: existing metric, our newly proposed metrics, and human evaluations. We also find that incorporating an additional pseudo-dataset [45], which captures diverse ranges of speech and lip movement intensities, can further improve expressiveness. Our main contributions are summarized as follows: + +- Defining three aspects—Temporal Synchronization, Lip Readability, and Expressiveness—that affect the perceptual quality of 3D talking heads and proposing three evaluation metrics for these aspects. +- Constructing a speech-mesh representation space that captures rich and diverse correspondences between speech and lip movements. +- Proposing a plug-in perceptual loss using the constructed speech-mesh representation and demonstrating improvements on existing metric, our newly proposed metrics, and human evaluations. + +# 2. Related Work + +Speech-driven 3D talking head generation. Speech-driven 3D talking head generation aims to generate realistic 3D facial movements aligned with given speech. Among recent data-driven methods [8, 9, 11, 21, 22, 33, 46, 47], FaceFormer [11] introduces a transformer-based autoregressive model and leverages a pre-trained speech model to capture long-term audio context and past facial movements. CodeTalker [46] employs a VQ-VAE to construct a discrete facial motion space, addressing the over-smoothing problem. Diffusion models [17, 31] have also been demonstrated to be effective for 3D talking head generation [33, 36]. In addition to synthesizing neutral facial motions, several works extend the 3D talking head to express specific aspects, such as emotional expressions [9, 22], multilingual capabilities [37], or laughter [38]. Despite these advances, existing methods rely + +on minimizing MSE loss without a clear definition of perceptually accurate lip movement, overlooking the multifaceted nature of lip synchronization. To address this, we define three critical aspects of lip synchronization and propose rich speech-mesh representation, along with its application as a perceptual loss in a plug-and-play manner, enhancing all three aspects of lip synchronization quality in existing 3D talking head generation models [11, 21, 46]. + +Speech-face representation learning. Well-aligned representation spaces learned from large-scale datasets such as CLIP [24] and ImageBind [14] are valued for their scalability and versatility. These spaces enable a wide range of applications, including auxiliary loss [42], intermediate representation [41], and evaluation metrics [16]. With this context, audio-visual representation spaces trained specifically on speech and 2D face videos have been proposed [6, 15, 29, 35]. For instance, SyncNet [6], a CNN-based model learns to detect audio-visual temporal synchronization and has been applied to several tasks, such as active speaker detection [1, 7] and 2D talking head generation [23, 44, 48]. Similarly, a transformer-based AV-HuBERT [29] has demonstrated remarkable effectiveness in various tasks, including lip reading [30], audio-visual translation [5], and 2D talking head generation [44]. While there has been significant progress in the 2D domain, advancements in the 3D domain remain under-explored. Yang et al. [47] extends the SyncNet architecture to accommodate speech and 3D face meshes; however, its application is limited to evaluating 3D talking heads. In this work, we demonstrate the versatility of our representation space as a plug-in module to improve the perceptual accuracy of the existing speech-driven 3D talking head generation models [11, 21, 46], as well as to assess their performance. + +Evaluation metrics for speech-driven 3D talking head. The prevalent evaluation metric, Lip Vertex Error (LVE) [27], measures the L2 distance between predicted lip vertices and ground truth. Additional metrics, such as Upper Face Dynamics Deviation (FDD) [46] and Lip Readability Percentage (LRP) [21], consider different regions of facial motion. These metrics, however, focus on vertex-wise geometric differences between the ground-truth 3D facial motions and neglect speech-related information. To incorporate both speech and 3D face mesh data for evaluation, MultiTalk [37] introduces an Audio-Visual Lip Readability (AVLR) metric, which assesses perceptual accuracy of lip readability using a pre-trained Audio-Visual Speech Recognition model [2]. Yet, AVLR relies on speech and 2D face video rendered from the 3D face mesh, which may not align with the 3D talking head domain. To address these limitations, we introduce novel and comprehensive evaluation metrics that focus on diverse aspects of lip synchronization: Mean Temporal Misalignment (MTM), Perceptual Lip Readability Score (PLRS), and Speech-Lip Intensity Correlation Coefficient (SLCC). + +# 3. Essential Criteria for Perceptually Accurate 3D Talking Head + +Generating a 3D talking head with lip movements that are perceptually accurate to human observers requires a clear understanding of the components that influence perceptual quality. Although existing works have focused on improving partial aspects of these talking head generation models, a comprehensive exploration for establishing a perceptually accurate 3D talking head model has barely been undertaken. Drawing inspirations from extensive research in existing works [10, 18, 21, 28, 34, 40, 43, 44] and our studies, we identify three fundamental criteria essential for achieving perceptually accurate lip movements in 3D talking heads. + +Temporal Synchronization. This alignment is particularly crucial in media involving human speech, where viewers expect lip movements to precisely match the corresponding speech in time. Temporal misalignment between speech and lip movements indeed distracts viewers, reduces user experiences, and negatively impacts audience perception [25]. Vatakis et al. [43] find that viewers are particularly sensitive to speech-lip asynchrony compared to other audio-visual asynchrony, such as music. They note that mismatches become noticeable when speech precedes lip movements by more than $50~\mathrm{ms}$ or follows them by more than $220~\mathrm{ms}$ . These findings may hold the same in 3D talking head generation, where any misalignment between speech and lip movements can break the illusion of realism, leading to a diminished user experience and reducing the perceived authenticity of the virtual character. Thus, we define the temporal synchronization of the talking head as an important aspect of lip synchronization quality. + +Lip Readability. Visemes (or lip movements) must correspond accurately to the speech phonemes to ensure the spoken words are visually intelligible [3]. This aspect is widely acknowledged as important in existing literature in both 2D [44] and 3D [10, 21] speech-driven talking head models, which leverage lip reading experts as an auxiliary guidance to improve the visual intelligibility of the spoken word. However, the mapping between speech and lip movements is not one-to-one, making lip readability challenging to define. For example, the size and shape of the opening mouth and the dynamic movement patterns of the lips in response to specific utterance differ at every moment [47] or per individual [32]. To capture this complexity, we define the lip readability within speech-mesh synchronized representation space learned from a large-scale dataset, capturing the comprehensive and nuanced correspondence between speech and lip movements. + +Expressiveness. Speech conveys not only linguistic content but also varies in intensity. For instance, a speaker may express the same text softly or loudly, with jaw and lip movements proportionally increasing as speech intensity rises, + +![](images/bfc3b81fc34afd47d5a65a4cbf9cb533638836942a950c99d37a36af7b69549e.jpg) +Table 1. Human studies on alignment criteria. [Left] Preference scores (1-3) for 3D talking heads with varying lip movement intensities paired with different speech intensities. [Right] Human preference between (A) samples with precise timing but low expressiveness, and (B) samples with high expressiveness but 100ms asynchrony—twice the commonly accepted 50ms threshold [43]. + +
Samples
AB
Temp.✓×
Exp.×✓
Prefer (%)17.482.6
+ +which contributes to perceived expressiveness in real-world face recognition [18, 28, 40]. To demonstrate that the positive correlation of human preference between the intensity of speech and lip movements also exists in 3D talking head field, we conduct a human study by presenting 3D talking heads with varied lip movement intensities paired with different speech intensities in Table 1-[Left]. Participants prefer the lip movements with the intensity that match the intensity of speech. Despite this distinct human preference for synchronization between the intensity of corresponding speech and lip movements, this aspect has barely been explored in the talking head generation field. We thus define the expressiveness in 3D talking head as the correlation between speech and lip movement intensity, which is crucial for establishing genuine talking heads and is expected to guide future research aimed at improving perceptual lip synchronization. + +Focus of this work. Among the three criteria for perceptually accurate 3D talking head, we find that recent 3D talking head generation models achieve reasonable temporal alignment between speech and lip movements (see supplementary for a visualization of temporal synchronization). Furthermore, we design an A vs. B test, prompting participants to choose between two samples: (A) temporally synchronized one while lacking expressive synchronization, and (B) the other with expressive synchronization but with $100\mathrm{ms}$ asynchrony. Table 1-[Right] shows that users prefer sample B, suggesting that humans may prioritize expressive synchronization over strict temporal alignment. This insight directs our focus toward enhancing all three aspects to better capture perceptual realism in 3D talking head. Details of the human study are in the supplementary material. Before introducing metrics to assess the criteria, we present our synchronized representation for designing our perceptual loss. + +# 4. Speech-Mesh Synchronized Representation + +We hypothesize that a desirable representation space exists to meet the three criteria defined in Sec. 3. Motivated by this + +hypothesis, we develop a rich speech-mesh synchronized representation which captures the intricate correspondence between speech and the 3D face mesh. We found that our learned representation exhibits desirable properties, and we adopt it as a perceptual loss to improve the perceptual accuracy of existing models with respect to these three criteria. + +Overview. Directly learning a synchronized speech-mesh representation presents challenges due to the scarcity of speech-3D face mesh datasets. One potential solution is to construct pseudo-GT 3D face meshes using reconstruction models [13], although relying solely on pseudo-GT may not suffice for building a robust representation. To overcome this, we leverage the extensive knowledge from the speech-2D lip video representations and adapt this to accommodate 3D face meshes. Specifically, we propose a two-stage training process: in stage 1, we learn an audio-visual speech representation that accurately reflects lip movements from the unlabeled in-the-wild 2D synchronized talking face video dataset, LRS3 [1]. Subsequently, we leverage the pre-trained speech representation from stage 1, using it as the anchor space to learn a synchronized speech-mesh representation. + +Stage 1. Learning audio-visual speech representation. This stage aims to learn a rich speech-2D lip video representation that effectively captures the correlation between varying speech characteristics and lip dynamics. Motivated by prior works [15, 35], we extend the integration of masked autoencoder (MAE) and cross-modal contrastive learning to learn a synchronized speech-2D lip video representation using 2D videos. The architecture for stage 1 consists of two modality-specific encoders, a cross-modal fusion encoder, and two modality-specific decoders. (see Fig. 2-[Stage 1]). + +Given a speech and 2D lip video pair, $(\mathbf{X}_{\mathbf{s}},\mathbf{X}_{\mathbf{v}})\in D$ we begin by patchifying and tokenizing speech spectrograms and video frames into speech and video tokens as $\mathbf{S} = (\mathbf{s}_1,\dots ,\mathbf{s}_N)$ and $\mathbf{V} = (\mathbf{v}_1,\dots ,\mathbf{v}_M)$ , where $\mathbf{s}_i,\mathbf{v}_j\in$ $\mathbb{R}^H$ . We then randomly mask out $P\%$ portion of tokens. The remaining unmasked speech tokens $\mathbf{S}^{unmask}$ and video tokens $\mathbf{V}^{unmask}$ are respectively fed into the speech encoder $E_{s}$ and video encoder $E_{v}$ , each consisting of $N_{e}$ transformer layers. Each encoder extracts uni-modal embeddings, $\mathbf{Z}_{\mathbf{s}}^{l} = E_{s}^{l}(\mathbf{S}^{unmask})$ and $\mathbf{Z}_{\mathbf{v}}^{l} = E_{v}^{l}(\mathbf{V}^{unmask})$ , where $l\in (1,\ldots ,N_e)$ denote the layer indices. Also, the mean pooled speech and video embeddings, $\mathbf{c_s^l} = \mathrm{MeanPool}(\mathbf{Z_s^l})$ and $\mathbf{c_v^l} = \mathrm{MeanPool}(\mathbf{Z_v^l})$ , are derived from the corresponding uni-modal embeddings. Following the uni-modal encoders, we introduce a multi-modal fusion encoder $E_{sv}$ to exploit complementary information from each extracted uni-modal embedding. The speech and video embeddings are jointly processed through this encoder, resulting in multi-modal fusion embeddings $\mathbf{F_s}$ and $\mathbf{F_v}$ , i.e., $[\mathbf{F_s},\mathbf{F_v}] = E_{sv}([\mathbf{Z_s},\mathbf{Z_v}])$ where $[,\cdot ]$ denotes concatenation. With the fusion embeddings $\mathbf{F_s}$ and $\mathbf{F_v}$ , we utilize each modality's decoder to reconstruct the original signals. Specifically, we employ a + +![](images/d974d05e76c6c8e9a4c6306fb0e15cd76881e5ffe7df953523490cbeac87cc92.jpg) +Figure 2. Pipeline of speech-mesh synchronized representation learning. We train our speech-mesh representation space in a two-stage manner. In the first stage, we learn a rich audio-visual representation in 2D domain to capture the synchronization between lip movement and speech. In the second stage, we train the 3D mesh encoder to align the 3D mesh space with the frozen speech space. As an application of our speech-mesh representation space, we propose a plug-in perceptual loss to 3D talking head models to enhance the quality of lip movements. + +![](images/8ba9880e77762d0057f882c1cd0af5b67e3498063f827cac49f1012342d34662.jpg) + +speech decoder $D_{s}$ and a video decoder $D_{v}$ , each consisting of $N_{d}$ transformer layers. We pad each multi-modal fusion embedding with trainable masked tokens at their original positions, resulting in $\mathbf{F}_{\mathbf{s}}^{\prime}$ and $\mathbf{F}_{\mathbf{v}}^{\prime}$ . Each decoder then reconstructs the respective signals; the reconstructed speech spectrogram and video tokens as $\hat{\mathbf{S}}^{mask} = D_s(\mathbf{F}_{\mathbf{s}}^{\prime})$ and $\hat{\mathbf{V}}^{mask} = D_v(\mathbf{F}_{\mathbf{v}}^{\prime})$ . + +Training the model involves two objectives: learning the cross-modal alignment between the speech-2D lip video signals, while reconstructing their original signals. Given $B$ speech-2D lip video token pairs in a batch, $\{\mathbf{S}_i,\mathbf{V}_i\}_{i = 1}^B$ , we treat the true speech-2D lip video pair as positive samples, while the others in the batch are considered negative samples. To encourage alignment between temporally synchronized speech-2D lip video instances, we utilize a cross-modal contrastive learning strategy, InfoNCE [20]. We maximize the cosine similarity between positive sample of the mean pooled speech embedding $\mathbf{c}_{\mathbf{s},i}^{l} = \mathrm{MeanPool}(E_{s}^{l}(\mathbf{S}_{i}^{\mathrm{unmask}}))$ and the corresponding synchronized mean pooled video embedding $\mathbf{c}_{\mathbf{v},i}^{l} = \mathrm{MeanPool}(E_{v}^{l}(\mathbf{V}_{i}^{\mathrm{unmask}}))$ . We first define the speech-centric loss as: + +$$ +\mathcal {L} _ {\mathrm {S} \rightarrow \mathrm {V}} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\exp \left(\mathbf {c} _ {\mathbf {s} , i} ^ {l} \cdot \mathbf {c} _ {\mathbf {v} , i} ^ {l} / \tau\right)}{\sum_ {j = 1} ^ {B} \exp \left(\mathbf {c} _ {\mathbf {s} , i} ^ {l} \cdot \mathbf {c} _ {\mathbf {v} , j} ^ {l} / \tau\right)}, \tag {1} +$$ + +where $\tau$ is a temperature hyperparameter. Also, we make the objective symmetric by defining a video-centric loss as $\mathcal{L}_{\mathrm{V}\rightarrow \mathrm{S}}$ . We sum $\mathcal{L}_{\mathrm{S}\rightarrow \mathrm{V}}$ and $\mathcal{L}_{\mathrm{V}\rightarrow \mathrm{S}}$ across $L$ selected encoder layers to obtain our final contrastive learning objective: + +$$ +\mathcal {L} _ {\text {I n f o N C E}} = \sum_ {l \in L} \mathcal {L} _ {\mathrm {S} \rightarrow \mathrm {V}} + \mathcal {L} _ {\mathrm {V} \rightarrow \mathrm {S}}. \tag {2} +$$ + +For reconstruction, the model is trained in a self-supervised manner by minimizing the reconstruction loss $\mathcal{L}_{\mathrm{MAE}}$ as: + +$$ +\mathcal {L} _ {\mathrm {M A E}} = \frac {1}{B} \sum_ {i = 1} ^ {B} \left[ \frac {\sum \| \hat {\mathbf {S}} _ {i} ^ {m a s k} - \mathbf {S} _ {i} ^ {m a s k} \| _ {F} ^ {2}}{| \mathbf {S} _ {i} ^ {m a s k} |} + \frac {\sum \| \hat {\mathbf {V}} _ {i} ^ {m a s k} - \mathbf {V} _ {i} ^ {m a s k} \| _ {F} ^ {2}}{| \mathbf {V} _ {i} ^ {m a s k} |} \right], \tag {3} +$$ + +where $|\mathbf{S}_i^{mask}|$ and $|\mathbf{V}_i^{mask}|$ denote the number of masked speech and video tokens, respectively. Overall, our objective function is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {M A E}} + \lambda \mathcal {L} _ {\text {I n f o N C E}}, \tag {4} +$$ + +where $\lambda$ is the weight factor for cross-modal contrastive loss. + +We observe that the representation space trained with the transformer architecture and a rich and large-scale 2D face video dataset in this way already possesses the desired properties we pursue, illustrated in Fig. 1 (see supplementary for visualizations of pre-trained speech representation). Motivated by this, we transfer these emergent properties to the speech-mesh representation space as follows. + +Stage 2. Learning speech-mesh representation. In this stage, we design a 3D mesh encoder that maps 3D face mesh to the speech representation pre-trained in stage 1. This pretrained speech representation, derived from rich speech-2D face video data, serves as a robust anchor space for learning the correlation between diverse speech characteristics and 3D facial motions. We use contrastive learning loss to align 3D facial motion embeddings with the anchored speech representations, as shown in Fig. 2-[Stage 2]. + +Given a speech and 3D face mesh pair $(\mathbf{X}_{\mathbf{s}},\mathbf{X}_{\mathbf{m}})\in D$ we patchify and tokenize speech spectrograms into speech tokens S, and map these into pre-trained uni-modal mean pooled speech embeddings $\mathbf{c_s^l} = \mathrm{MeanPool}(E_s^l (\mathbf{S}))$ . Similarly, we patchify 3D face mesh into mesh tokens M, and feed them into the 3D mesh encoder $E_{m}$ consisting of $N_{e}$ transformer layers to extract mesh embeddings $\mathbf{Z}_{\mathbf{m}}^{l} = E_{m}^{l}(\mathbf{M})$ and the corresponding mean-pooled mesh embedding $\mathbf{c_m^l}$ . For the learning objective, given $B$ speech-3D face mesh token pairs in a batch, $\{\mathbf{S}_i,\mathbf{M}_i\}_{i = 1}^B$ , we first define speech-anchored loss as: + +$$ +\mathcal {L} _ {\mathrm {S} \rightarrow \mathrm {M}} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\exp \left(\mathbf {c} _ {\mathbf {s} , i} ^ {l} \cdot \mathbf {c} _ {\mathbf {m} , i} ^ {l} / \tau\right)}{\sum_ {j = 1} ^ {B} \exp \left(\mathbf {c} _ {\mathbf {s} , i} ^ {l} \cdot \mathbf {c} _ {\mathbf {m} , j} ^ {l} / \tau\right)}, \tag {5} +$$ + +as well as $\mathcal{L}_{\mathrm{M}\to \mathrm{S}}$ which is the mesh-anchored loss. Similar to Eq. (2), we sum $\mathcal{L}_{\mathrm{S}\rightarrow \mathrm{M}}$ and $\mathcal{L}_{\mathrm{M}\rightarrow \mathrm{S}}$ across $L$ selected encoder layers to train the 3D mesh encoder $E_{m}$ while the speech encoder fixed. As a result, this two-stage training approach yields a robust speech-mesh representation, which significantly enhances the alignment between the speech and 3D face mesh modalities. + +Adopting it as a perceptual loss. A key application of the speech-mesh representation learned through the two-stage training process is its use as a perceptual loss to enhance the perceptual accuracy of the 3D talking head model (see Fig. 2). Leveraging this representation as a perceptual loss ensures that the generated lip movements are perceptually accurate and aligned well with the speech. + +Given $B$ speech and generated 3D face mesh pairs in a batch, $\{\mathbf{S}_i,\hat{\mathbf{M}}_i\}_{i = 1}^B$ , we define our perceptual loss with the symmetric InfoNCE loss (Eq. (2)) as + +$$ +\mathcal {L} _ {\text {p e r c p}} = \sum_ {l \in L} \mathcal {L} _ {\mathrm {S} \rightarrow \mathrm {M}} + \mathcal {L} _ {\mathrm {M} \rightarrow \mathrm {S}}. \tag {6} +$$ + +Our perceptual loss encourages synchronized speech and mesh embeddings to pull closer together, while unsynchronized ones to push apart. The effectiveness and analysis of our perceptual loss will be discussed in later sections. + +# 5. Evaluation Metrics + +In this section, we describe our proposed evaluation metrics that assess each criterion impacting the quality of 3D lip accuracy, as discussed in Sec. 3. Here, we outline the high-level concepts of our proposed metrics. Additional details and experiments are provided in the supplementary material. + +Mean Temporal Misalignment (MTM). To measure the temporal discrepancy between speech and corresponding lip movements, temporal correspondence annotations, such as onset times for each modality, would typically be required. As a proxy, we determine temporal correspondence and measure temporal discrepancies between the ground truth and predicted lip vertex displacement sequences by using Derivative Dynamic Time Warping (DDTW) [19], which robustly identifies local structural similarities compared to standard Dynamic Time Warping (DTW) [4]. For simplicity, we focus on the central vertices of the upper and lower lips when extracting displacement sequences, and use local extrema in the DDTW process to measure temporal misalignment by pinpointing precise time steps for mouth opening and closing events. Consequently, Mean Temporal Misalignment (MTM) $\overline{\Delta t}$ is defined as $\overline{\Delta t} = \frac{1}{K}\sum_{k=1}^{K}\Delta t_k$ , where $K$ is the total number of video clips and $\Delta t_k$ is the averaged temporal misalignment of the $k$ -th video clip. + +Perceptual Lip Readability Score (PLRS). While Lip Vertex Error (LVE) measures the accuracy of generated lip articulations against the ground truth, it does not fully assess + +whether lip movements are perceptually aligned with the given speech. To address this, we leverage our speech-mesh representation in Sec. 4 as a perceptual lip readability evaluator. We compute the perceptual alignment using the cosine similarity of the mean pooled speech and mesh embeddings. Since this representation has learned a rich distribution of speech correspondences across various facial movements, our metric correlates highly with human perception, measuring perceptual lip movement alignment more accurately than LVE (see supplementary for the human study on metrics). + +Speech and Lip Intensity Correlation Coefficient (SLCC). As shown in Table 1-[Left], humans prefer aligned intensity between speech and lip movement. Thus, the intensity of generated lip movements should positively correlate with the corresponding input speech. To quantify this, we define the Speech and Lip Correlation Coefficient $r_{SL}$ as: + +$$ +r _ {S L} = \frac {\sum_ {k = 1} ^ {K} (\mathrm {S I} _ {k} - \bar {\mathrm {S I}}) (\mathrm {L I} _ {k} - \bar {\mathrm {L I}})}{\sqrt {\sum_ {k = 1} ^ {K} (\mathrm {S I} _ {k} - \bar {\mathrm {S I}}) ^ {2}} \sqrt {\sum_ {k = 1} ^ {K} (\mathrm {L I} _ {k} - \bar {\mathrm {L I}}) ^ {2}}}, \tag {7} +$$ + +where $\mathrm{SI}_k$ and $\mathrm{LI}_k$ denote Speech (SI) and Lip Intensity (LI), respectively, $\bar{\mathrm{SI}} = \frac{1}{K}\sum_{k = 1}^{K}\mathrm{SI}_k$ and $\bar{\mathrm{LI}} = \frac{1}{K}\sum_{k = 1}^{K}\mathrm{LI}_k$ . We define SI using speech loudness, specifically the z-normalized Root Mean Square (RMS) value, which is a widely accepted measure of speech intensity in signal processing. To define LI, we measure the averaged lip displacement value of a video clip, followed by the z-normalization. + +# 6. Experiments + +We first outline the evaluation setup, and then present thorough analyses of the experimental results. Due to the space limitation, we present more implementation details and additional experiments in the supplementary material. + +# 6.1. Experimental Settings + +Datasets. Most of the existing speech-driven 3D talking head generation methods rely on VOCASET [8] and BIWI [12] to train and test the models. However, these datasets have limited ranges of facial motion patterns due to the small dataset scale and restricted intensity range, which restricts their ability to fully capture the intricate relationship between speech and 3D face mesh. To address the lack of training dataset, we construct two large-scale speech-3d face mesh benchmark datasets, LRS3-3D and MEAD-3D, by processing LRS3 and MEAD videos using two monocular face reconstruction methods: SPECTRE [13] for LRS3 [1], which ensures accurate lip movements, and SMIRK [26] for MEAD [45], which captures diverse speech and lip movement intensities. We use LRS3 in the first stage and LRS3-3D in the second stage to train our speech-mesh synchronized representation. Then, for base model training and evaluations, we adopt two configurations: (1) training and testing with VOCASET, in line with existing work, and (2) + +
MethodTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)SLCC / Δ (↓)
VOCASET--0.4090.34 / -
FaceFormer [11]53.63.3570.3680.26 / 0.08
+ Ours rep.52.23.0910.4630.37 / 0.03
CodeTalker [46]61.83.7000.3810.38 / 0.04
+ Ours rep.60.93.5790.3880.35 / 0.01
SelfTalk [21]50.12.9710.4140.41 / 0.07
+ Ours rep.49.22.9240.4180.35 / 0.01
+ +![](images/8ee7c40b0f81bf6cdc552ff6b15fdbbdc18de6958c846f2a6e059060076b58ae.jpg) +Figure 3. Qualitative results of the effectiveness of our perceptual loss for lip readability. Our perceptual loss guides baselines [11, 21, 46] to generate perceptually accurate lip movements. + +combining MEAD-3D with VOCASET during training to endow expressiveness and testing on MEAD-3D. + +Base methods. We use three state-of-the-art 3D talking head generation models [11, 21, 46] to evaluate the effectiveness of our perceptual loss. + +Metrics. To comprehensively evaluate the three aspects of lip synchronization, we assess MTM, PLRS, and SLCC, corresponding to temporal synchronization, lip readability, and expressiveness, respectively. Additionally, we compute the level-wise SLCC for the MEAD-3D test set, which includes three distinct emotional intensity levels, to evaluate the expressive capability of 3D talking head generation models. We also measure LVE as part of the lip readability evaluation. + +# 6.2. Experimental Results and Analysis + +We conduct evaluations to assess the effectiveness of our speech-mesh synchronized representation and the incorporation of an expressive speech-3D face mesh paired dataset (i.e., MEAD-3D) in enhancing the three criteria. + +How well do existing 3D talking head models achieve lip synchronization in all three aspects? The results are summarized in Table 2. For temporal synchronization, most base models achieve MTM values between 50 and 60ms, indicating performance close to the acceptable asynchrony + +Table 2. Quantitative results of lip synchronization on VO-CASET [8] test set. We evaluate the base models on our proposed lip synchronization metrics. We denote $\Delta$ as the difference in SLCC between the model and those measured on the data distribution. A lower $\Delta$ indicates the model more closely represents the intensity correlation of the dataset. We demonstrate the effectiveness of our representation in consistently enhancing all three aspects of lip synchronization. + +
MethodPerceptual LossTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)SLCC / Δ (↓)
VOCASET---0.4090.34 / -
FaceFormer [11]X53.63.3570.3680.26 / 0.08
3D SyncNet55.63.3160.4350.38 / 0.04
Ours w/o 2D prior55.33.2780.4000.42 / 0.08
Ours w/ 2D prior52.23.0910.4630.37 / 0.03
CodeTalker [46]X61.83.7000.3810.38 / 0.04
3D SyncNet59.64.3190.3790.14 / 0.20
Ours w/o 2D prior55.93.5790.3740.23 / 0.11
Ours w/ 2D prior60.93.5790.3880.35 / 0.01
SelfTalk [21]X50.12.9710.4140.41 / 0.07
3D SyncNet49.52.9410.4050.35 / 0.01
Ours w/o 2D prior54.43.1490.4170.39 / 0.05
Ours w/ 2D prior49.22.9240.4180.35 / 0.01
+ +Table 3. Ablation study on architectural choice and 2D prior knowledge. We validate the effectiveness of the transformer-based architecture and curriculum learning with a pre-trained 2D speech representation by ablating them from our proposed representation. + +![](images/6b22ebb0cf30f6c03f5d97d37a25d0c1543b1b5a51ac6a66c0b9ce71c975fa98.jpg) +Figure 4. t-SNE plot of ablation study. We plot the t-SNE graph for each perceptual critic model. We represent the features with same phoneme as same color. Squared and circled points denote mesh and speech features from each representation, respectively. + +![](images/32f8920b1801e2fdbc270e8c0250e147e773508a0f3115cb9b5c6126d637af33.jpg) +Figure 5. Behaviors of our representation in temporal and expressiveness sensitivity. We demonstrate the effectiveness of our representation in temporal synchronization and expressiveness using a cosine similarity graph and speech feature plots, respectively. We color the point as low, medium, and high intensity. + +threshold. Regarding lip readability, SelfTalk [21] achieves the best performance, while FaceFormer [11] has the lowest PLRS score. Furthermore, CodeTalker [46] demonstrates the closest SLCC values to the ground truth VOCASET mesh, while FaceFormer exhibits the highest SLCC discrepancy. + +Does our speech-mesh representation improve lip synchronization? Yes. Table 2 and Fig. 3 show consistent improvements in the three aspects with our perceptual loss. + +What makes our speech-mesh representation have lip synchronization ability? We hypothesize that the transformer-based architecture and curriculum learning with a pre-trained + +![](images/c4e0c0f81521d8659774a5cab61ca4c5dfb8aec956e971a7eb01ceb6bdf53076.jpg) +Figure 6. Qualitative results for the expressiveness. Given high and low intensity levels of speech, models trained on both MEAD-3D and VOCASET show more expressive lip movements compared to those trained on VOCASET alone, and even better with our perceptual loss. + +audio-visual speech representation contribute significantly to improved lip synchronization. To validate this, we conduct ablation studies, summarized in Table 3, examining the roles of pre-trained speech representation and architectural design. Without the pre-trained speech representation, base models show lower performance across metrics, and CNN-based architectures, 3D SyncNet, do not clearly show effectiveness, while ours with 2D prior consistently show improvement. The t-SNE plots of perceptual critic models in Fig. 4 show that, in (a) 3D SyncNet, speech and mesh features corresponding to the same phoneme group are separated. In (b) ours without 2D priors, speech and mesh features lack separation and appear scattered without clustering. Our representation (c), notably, tends to form more distinct clusters according to phonemes, with vowels and consonants grouped closely, potentially contributing to enhancing lip readability. We also observe a directional progression in the feature space, shifting from phonemes with mouth opening $(e.g., /aj/$ to those with mouth closing $(e.g., /f/$ . + +We also examine our representation's performance in other aspects, temporal synchronization and expressiveness. Figure 5-(a) demonstrates temporal sensitivity, as cosine similarity drops when temporal misalignment is introduced between input speech and 3D face mesh. In Fig. 5-(b), we plot speech features at varying speech intensities, showing a directional trend as intensity increases from lowest to highest. Figures 4 and 5 imply that our representation holds favorable properties discussed in Fig. 1 for the three criteria. + +# Can we unlock the expressive power of 3D talking heads? + +Likely. Since VOCASET [8] lacks the range of diverse speech and lip movement intensities, evaluating expressive power requires testing on a dataset with a broader range of intensities. To study this, we examine the expressiveness of the base models on the MEAD-3D dataset, which includes three intensity levels. We assess SLCC at each intensity level to evaluate expressiveness and identify any expressiveness bounds as the level of intensity increases. We first test the base models trained on VOCASET [8] against the MEAD-3D test set. Table 4 shows these models demonstrate limited expressiveness, with SLCC values showing minimal increase across intensity levels. We hypothesize that + +
MethodTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)Lv1SLCC / Δ (↓)
Lv2Lv3Avg
MEAD-3D--0.2300.24 / -0.30 / -0.39 / -0.42 / -
FaceFormer [11]59.63.2070.2990.08 / 0.160.07 / 0.230.07 / 0.320.06 / 0.36
+ MEAD-3D59.51.1390.1760.26 / 0.020.30 / 0.000.34 / 0.050.35 / 0.07
+ Ours rep.55.81.1140.3060.27 / 0.030.27 / 0.030.32 / 0.070.33 / 0.09
CodeTalker [46]60.73.2360.2940.02 / 0.220.03 / 0.270.03 / 0.360.02 / 0.40
+ MEAD-3D60.92.9540.1540.09 / 0.150.12 / 0.180.06 / 0.330.11 / 0.31
+ Ours rep.58.62.7050.2210.18 / 0.060.29 / 0.010.31 / 0.080.31 / 0.11
SelfTalk [21]53.43.3960.2940.14 / 0.100.14 / 0.160.17 / 0.220.15 / 0.27
+ MEAD-3D54.21.2380.2160.16 / 0.080.28 / 0.020.32 / 0.070.31 / 0.11
+ Ours rep.52.71.1920.2300.17 / 0.070.29 / 0.010.34 / 0.050.33 / 0.09
+ +Table 4. Quantitative results of lip synchronization on MEAD-3D test set. We evaluate the base models on the MEAD-3D test set. We also compute the level-wise SLCC to evaluate the expressive capability of the models. + +this limitation arises, because VOCASET's smaller scale and intensity range restrict the model's ability to learn relationships between speech and lip intensity. To address this, we integrate MEAD-3D with VOCASET for training, aiming to boost expressiveness in the base models. This approach consistently improves SLCC across all intensity levels. However, simply adding MEAD-3D degrades lip synchronization metrics, except for expressiveness, i.e., nontrivial. To counterbalance this, we leverage our perceptual loss, which effectively mitigates the degradation introduced by MEAD-3D while improving expressiveness (see Fig. 6). + +# 7. Conclusion + +This paper addresses challenges in existing 3D talking head generation models, which often overlook the true correspondence between speech and lip movements, making it difficult to accurately link lip movements with varying speech characteristics. To overcome this issue, we identify three essential aspects—Temporal Synchronization, Lip Readability, and Expressiveness—that influence the perceptual quality of lip movements, and develop specific metrics for each aspect. We introduce a speech-mesh synchronized representation that exhibits these emergent properties and adopt it as a perceptual loss. Our extensive analyses demonstrate that our perceptual loss consistently enhances models across three aspects. We believe that our defined aspects will guide future research in generating more realistic 3D talking heads, and our representation will serve as a key component. + +Acknowledgments. This research was supported by a grant from KRAFTON AI, and was also partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub; No.RS-2023-00225630, Development of Artificial Intelligence for Text-based 3D Movie Generation; No. RS-2024-00457882, National AI Research Lab Project) and Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2024 (Project Name: Development of barrier-free experiential XR contents technology to improve accessibility to online activities for the physically disabled, Project Number: RS-2024-00396700, Contribution Rate: $25\%$ ). + +# References + +[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 2, 3, 4, 6 +[2] Mohamed Anwar, Bowen Shi, Vedanuj Goswami, Wei-Ning Hsu, Juan Pino, and Changhan Wang. Muavic: A multilingual audio-visual corpus for robust speech recognition and robust speech-to-text translation. In INTERSPEECH 2023, 2023. 3 +[3] Helen L Bear and Richard Harvey. Phoneme-to-viseme mappings: the good, the bad, and the ugly. Speech Communication, 2017. 2, 3 +[4] Donald J Berndt and James Clifford. Using dynamic time warping to find patterns in time series. In Proceedings of the 3rd international conference on knowledge discovery and data mining, 1994. 6 +[5] Jeongsoo Choi, Se Jin Park, Minsu Kim, and Yong Man Ro. Av2av: Direct audio-visual speech to audio-visual speech translation with unified audio-visual speech representation. In CVPR, 2024. 3 +[6] J. S. Chung and A. Zisserman. Out of time: automated lip sync in the wild. In Workshop on Multi-view Lip-reading, ACCV, 2016. 3 +[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 3 +[8] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In CVPR, 2019. 1, 2, 6, 7, 8 +[9] Radek Daněček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. In SIGGRAPH Asia 2023 Conference Papers, pages 1-13, 2023. 1, 2 +[10] Han EunGi, Oh Hyun-Bin, Kim Sung-Bin, Corentin Nivelet Etcheberry, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Enhancing speech-driven 3d facial animation with audio-visual guidance from lip reading expert. In Interspeech 2024, pages 2940–2944, 2024. 3 + +[11] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In CVPR, 2022. 2, 3, 7, 8 +[12] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE TMM, 2010. 2, 6 +[13] Panagiotis P. Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Spectre: Visual speech-informed perceptual 3d facial expression reconstruction from videos. In CVPRW, pages 5745-5755, 2023. 4, 6 +[14] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 3 +[15] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In The Eleventh International Conference on Learning Representations, 2023. 3, 4 +[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: a reference-free evaluation metric for image captioning. In EMNLP, 2021. 3 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2 +[18] Jessica E Huber and Bharath Chandrasekaran. Effects of increasing sound pressure level on lip and jaw movement parameters and consistency in young adults. 2006. 2, 3, 4 +[19] Eamonn J. Keogh and M. Pazzani. Derivative dynamic time warping. In In First SIAM International Conference on Data Mining, 2001. 6 +[20] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5 +[21] Ziqiao Peng, Yihao Luo, Yue Shi, Hao Xu, Xiangyu Zhu, Hongyan Liu, Jun He, and Zhaoxin Fan. Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In ACM MM, 2023. 1, 2, 3, 7, 8 +[22] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[23] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, page 484-492. Association for Computing Machinery, 2020. 3 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. on Mach. Learn., pages 8748-8763. PMLR, 2021. 3 +[25] Byron Reeves and David Voelker. Effects of audio-video asynchrony on viewer's memory, evaluation of content and detec + +tion ability. In Research Report, Stanford University, 1993. 3 +[26] George Retsinas, Panagiotis P Filntisis, Radek Danecek, Victoria F Abrevaya, Anastasios Roussos, Timo Bolkart, and Petros Maragos. 3d facial expressions through analysis-by-neural-synthesis. In CVPR, 2024. 6 +[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In ICCV, 2021. 1, 3 +[28] Richard Schulman. Articulatory dynamics of loud and normal speech. The Journal of the Acoustical Society of America, 85 (1):295-312, 1989. 2, 3, 4 +[29] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR, 2022. 3 +[30] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. arXiv preprint arXiv:2201.01763, 2022. 3 +[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2 +[32] Wenfeng Song, Xuan Wang, Shi Zheng, Shuai Li, Aimin Hao, and Xia Hou. Talkingstyle: Personalized speech-driven 3d facial animation with style preservation. IEEE Transactions on Visualization and Computer Graphics, 2024. 3 +[33] Stefan Stan, Kazi Injamamul Haque, and Zerrin Yumak. Facediffuser: Speech-driven 3d facial animation synthesis using diffusion. In Proceedings of the 16th ACM SIGGRAPH Conference on Motion, Interaction and Games, pages 1-11, 2023. 1, 2 +[34] Trinity Suma, Birate Sonia, Kwame Agyemang Baffour, and Oyewole Oyekoya. The effects of avatar voice and facial expression intensity on emotional recognition and user perception. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 2, 3 +[35] Licai Sun, Zheng Lian, Bin Liu, and Jianhua Tao. Hicmae: Hierarchical contrastive masked autoencoder for self-supervised audio-visual emotion recognition. Information Fusion, 108: 102382, 2024. 3, 4 +[36] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-Jin Liu. Diffposetalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM TOG, 43(4), 2024. 2 +[37] Kim Sung-Bin, Lee Chae-Yeon, Gihun Son, Oh Hyun-Bin, Janghoon Ju, Suekyeong Nam, and Tae-Hyun Oh. Multitalk: Enhancing 3d talking head generation across languages with multilingual video dataset. In Interspeech 2024, pages 1380-1384, 2024. 2, 3 +[38] Kim Sung-Bin, Lee Hyun, Da Hye Hong, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Laughtalk: Expressive 3d talking head generation with laughter. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6404-6413, 2024. 2 +[39] Hiroki Tanaka, Satoshi Nakamura, et al. The acceptability of virtual characters as social skills trainers: usability study. *JMIR human factors*, 9(1):e35358, 2022. 1 + +[40] Stephen M Tasko and Michael D McClean. Variations in articulatory movement with changes in speech task. 2004. 2, 3, 4 +[41] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 3 +[42] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-cap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918–17928, 2022. 3 +[43] Argiro Vatakis and Charles Spence. Audiovisual synchrony perception for speech and music assessed using a temporal order judgment task. *Neuroscience letters*, 393(1):40-44, 2006. 2, 3, 4 +[44] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In CVPR, pages 14653-14662, 2023. 3 +[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, 2020. 2, 6 +[46] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In CVPR, 2023. 1, 2, 3, 7, 8 +[47] Karren Yang, Anurag Ranjan, Rick Chang, Raviteja Vemulapalli, and Oncel Tuzel. Probabilistic speech-driven 3d facial motion synthesis: New benchmarks, methods, and applications. In CVPR, 2024. 2, 3 +[48] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In CVPR, pages 8652-8661, 2023. 3 \ No newline at end of file diff --git a/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/images.zip b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a7fc29d5e547538a856b74083c9506f34240f0c5 --- /dev/null +++ b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13dc91a7cacc59f14b46b1117f1597cfcb7e10a4fd381272a1be13b29ab2bf4c +size 432827 diff --git a/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/layout.json b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a769651a547fdecf1d48bf3eee9dcc05a5ac92c1 --- /dev/null +++ b/2025/Perceptually Accurate 3D Talking Head Generation_ New Definitions, Speech-Mesh Representation, and Evaluation Metrics/layout.json @@ -0,0 +1,7752 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 523, + 140 + ], + "type": "text", + "content": "Perceptually Accurate 3D Talking Head Generation: New Definitions, Speech-Mesh Representation, and Evaluation Metrics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "text", + "content": "Lee Chae-Yeon" + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "text", + "content": " Oh Hyun-Bin" + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "text", + "content": " Han EunGi" + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "text", + "content": " Kim Sung-Bin" + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "text", + "content": " Suekyeong Nam" + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "text", + "content": " Tae-Hyun Oh" + }, + { + "bbox": [ + 61, + 160, + 553, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 185, + 542, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 185, + 542, + 198 + ], + "spans": [ + { + "bbox": [ + 73, + 185, + 542, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 185, + 542, + 198 + ], + "type": "text", + "content": "Grad. School of AI, POSTECH " + }, + { + "bbox": [ + 73, + 185, + 542, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 185, + 542, + 198 + ], + "type": "text", + "content": "Dept. of Electrical Engineering, POSTECH " + }, + { + "bbox": [ + 73, + 185, + 542, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 185, + 542, + 198 + ], + "type": "text", + "content": "KRAFTON " + }, + { + "bbox": [ + 73, + 185, + 542, + 198 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 73, + 185, + 542, + 198 + ], + "type": "text", + "content": "School of Computing, KAIST" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 225, + 302, + 327 + ], + "blocks": [ + { + "bbox": [ + 57, + 225, + 302, + 327 + ], + "lines": [ + { + "bbox": [ + 57, + 225, + 302, + 327 + ], + "spans": [ + { + "bbox": [ + 57, + 225, + 302, + 327 + ], + "type": "image", + "image_path": "eb6b0c40d41458ddd681a07cd73458212aa53cb72f5c6642970e29e1aa81b450.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 331, + 282, + 341 + ], + "lines": [ + { + "bbox": [ + 66, + 331, + 282, + 341 + ], + "spans": [ + { + "bbox": [ + 66, + 331, + 282, + 341 + ], + "type": "text", + "content": "(a) Essential Criteria for Perceptually Accurate 3D Talking Head" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 227, + 547, + 325 + ], + "blocks": [ + { + "bbox": [ + 316, + 227, + 547, + 325 + ], + "lines": [ + { + "bbox": [ + 316, + 227, + 547, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 547, + 325 + ], + "type": "image", + "image_path": "c154fed79a46522ce17c99ae75dba0cec284ca566baa14e802bea69d8a08887a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 331, + 525, + 342 + ], + "lines": [ + { + "bbox": [ + 331, + 331, + 525, + 342 + ], + "spans": [ + { + "bbox": [ + 331, + 331, + 525, + 342 + ], + "type": "text", + "content": "(b) Conceptual Diagram of Desired Representation Space" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 346, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 54, + 346, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 346, + 555, + 413 + ], + "type": "text", + "content": "Figure 1. What defines perceptually accurate lip movement for a speech signal? In this work, we define three criteria to assess perceptual alignment between speech and lip movements of 3D talking heads: Temporal Synchronization, Lip Readability, and Expressiveness (a). The motivational hypothesis is the existence of a desirable representation space that models and complies well with the three criteria between diverse speech characteristics and 3D facial movements, as illustrated in (b); where representations with the same phonemes are clustered, are sensitive to temporal synchronization, and follow a certain pattern as the speech intensity increases. Consequently, we build a rich speech-mesh synchronized representation space that exhibits the desirable properties." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 429, + 200, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 429, + 200, + 441 + ], + "spans": [ + { + "bbox": [ + 151, + 429, + 200, + 441 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 454, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 297, + 694 + ], + "type": "text", + "content": "Recent advancements in speech-driven 3D talking head generation have made significant progress in lip synchronization. However, existing models still struggle to capture the perceptual alignment between varying speech characteristics and corresponding lip movements. In this work, we claim that three criteria—Temporal Synchronization, Lip Readability, and Expressiveness—are crucial for achieving perceptually accurate lip movements. Motivated by our hypothesis that a desirable representation space exists to meet these three criteria, we introduce a speech-mesh synchronized representation that captures intricate correspondences between speech signals and 3D face meshes. We found that our learned representation exhibits desirable characteristics, and we plug it into existing models as a perceptual loss to better align lip movements to the given speech. In addition, we utilize this representation as a perceptual metric and introduce two other physically grounded lip synchronization metrics to assess how well the generated 3D talking heads align with these three criteria. Experiments show that training 3D talking head generation models with our perceptual loss" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 430, + 556, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 467 + ], + "type": "text", + "content": "significantly improve all three aspects of perceptually accurate lip synchronization. Codes and datasets are available at https://perceptual-3d-talking-head.github.io/." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 487, + 394, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 394, + 499 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 394, + 499 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 509, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 556, + 628 + ], + "type": "text", + "content": "Speech-driven 3D talking head generation focuses on generating 3D facial movements synchronized with input speech signals. This plays a key role in enhancing communication within multimedia applications, such as virtual reality, entertainment, and education [39]. To provide users with a more realistic and immersive experience, it is crucial that the facial and lip movements of 3D avatars are synchronized with the various aspects of speech. This synchronization should be perceptually accurate from a human perspective, ensuring that the avatars' expressions are both natural and convincing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "While recent works in learning-based 3D talking head generation [8, 9, 21, 27, 33, 46] aim to enhance the lip synchronization capabilities, they commonly rely on minimizing the Mean Squared Error (MSE) loss between generated 3D facial motion and ground truth motion as a learning objective. This approach is practical, as it directly contributes to minimizing the Lip Vertex Error (LVE), a commonly used" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 702, + 171, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 171, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 171, + 713 + ], + "type": "text", + "content": "*These authors contributed equally." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21065" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "metric that measures the MSE between the lip vertices of generated 3D facial motions and the ground truth." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 99, + 297, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 99, + 297, + 255 + ], + "spans": [ + { + "bbox": [ + 54, + 99, + 297, + 255 + ], + "type": "text", + "content": "Despite the improvements in the LVE metric, existing models still struggle to correlate lip movements with some speech characteristics, such as wider mouth openings as speech volume increases. These characteristics are not adequately captured by existing datasets [8, 12], which have limited ranges of facial motion patterns due to the small dataset scale and restricted intensity range. Moreover, relying on MSE and LVE is insufficient for learning or assessing perceptually plausible lip motions [21, 37], as it focuses solely on vertex-wise geometric differences and overlooks the true correspondence between the speech signals and lip movements. A lower MSE and LVE do not necessarily correspond to a more perceptually accurate lip movement." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 257, + 298, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 298, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 298, + 544 + ], + "type": "text", + "content": "These observations raise critical questions: What defines perceptually accurate lip movement in response to a speech signal, and how can we enhance this accuracy? We draw inspiration from findings on human audio-visual perception: 1) Humans are sensitive to temporal asynchrony between speech and lip movements; slight discrepancies can disrupt the perception of natural synchronization [43]. 2) Humans rely on accurate viseme-phoneme correspondence when assessing lip-sync accuracy, expecting visual lip movements to match the spoken phonemes [3]. 3) There is a proportional increase in jaw and lip movements as speech intensity increases, contributing to the expressiveness perceived in natural speech [18, 28, 34, 40]. Through our human study, we reveal an intriguing finding: participants favor lip movements with intensity that corresponds to speech—even if they exceed the established maximum acceptable asynchrony [43] by twice—over those that are perfectly synchronized but lack expressive alignment (Table 1-[Right]). This reveals that humans are more sensitive to expressiveness than temporal synchronization when perceiving, highlighting the importance of expressiveness. Building upon these insights, we define three criteria that significantly impact the perceptual lip synchronization of 3D talking heads: Temporal Synchronization, Lip Readability, and Expressiveness (Fig. 1 (a))." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": "Motivated by our hypothesis that a desirable representation space exists to meet these three criteria, we propose a speech-mesh synchronized and rich representation that captures the intricate correspondence between speech and 3D face mesh. We design a transformer-based architecture that maps time-sequenced speech and mesh inputs to a shared representation space. To effectively train this system, we employ a two-stage training method: we first develop a robust audio-visual speech representation using a large-scale 2D video dataset [1], which then serves as an anchor for learning a speech-mesh representation. We found that the first step is important and leads to emergent desirable properties (illustrated in Fig. 1 (b)) in the final representation. This sequential approach ensures that the model first establishes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "a space that captures a wide range of speech characteristics, and then extends to explore the relationships between speech and 3D face mesh across diverse speech intensities and facial movements. Adopting this representation, we introduce a plug-and-play perceptual loss adaptable to any existing 3D talking head generation models [11, 21, 46], enhancing the perceptual quality of 3D talking heads." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 157, + 556, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 556, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 556, + 252 + ], + "type": "text", + "content": "Furthermore, to assess the three criteria, we introduce three metrics for each aspect. We leverage our learned representation as a perceptual metric, Perceptual Lip Readability Score (PLRS), to evaluate the perceptual lip readability of lip movements. Also, we propose two physically grounded lip synchronization metrics: Mean Temporal Misalignment (MTM) for temporal synchronization and Speech-Lip Intensity Correlation Coefficient (SLCC) for expressiveness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 253, + 556, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 556, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 556, + 360 + ], + "type": "text", + "content": "Extensive experiments demonstrate that our perceptual loss significantly enhances all three aspects: Temporal Synchronization, Lip Readability, and Expressiveness, which are demonstrated across various metrics: existing metric, our newly proposed metrics, and human evaluations. We also find that incorporating an additional pseudo-dataset [45], which captures diverse ranges of speech and lip movement intensities, can further improve expressiveness. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 362, + 556, + 502 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 362, + 556, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 556, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 556, + 411 + ], + "type": "text", + "content": "- Defining three aspects—Temporal Synchronization, Lip Readability, and Expressiveness—that affect the perceptual quality of 3D talking heads and proposing three evaluation metrics for these aspects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 415, + 556, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 451 + ], + "type": "text", + "content": "- Constructing a speech-mesh representation space that captures rich and diverse correspondences between speech and lip movements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 455, + 556, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 556, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 556, + 502 + ], + "type": "text", + "content": "- Proposing a plug-in perceptual loss using the constructed speech-mesh representation and demonstrating improvements on existing metric, our newly proposed metrics, and human evaluations." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 514, + 402, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 402, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 402, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": "Speech-driven 3D talking head generation. Speech-driven 3D talking head generation aims to generate realistic 3D facial movements aligned with given speech. Among recent data-driven methods [8, 9, 11, 21, 22, 33, 46, 47], FaceFormer [11] introduces a transformer-based autoregressive model and leverages a pre-trained speech model to capture long-term audio context and past facial movements. CodeTalker [46] employs a VQ-VAE to construct a discrete facial motion space, addressing the over-smoothing problem. Diffusion models [17, 31] have also been demonstrated to be effective for 3D talking head generation [33, 36]. In addition to synthesizing neutral facial motions, several works extend the 3D talking head to express specific aspects, such as emotional expressions [9, 22], multilingual capabilities [37], or laughter [38]. Despite these advances, existing methods rely" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21066" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "on minimizing MSE loss without a clear definition of perceptually accurate lip movement, overlooking the multifaceted nature of lip synchronization. To address this, we define three critical aspects of lip synchronization and propose rich speech-mesh representation, along with its application as a perceptual loss in a plug-and-play manner, enhancing all three aspects of lip synchronization quality in existing 3D talking head generation models [11, 21, 46]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 172, + 298, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 298, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 298, + 471 + ], + "type": "text", + "content": "Speech-face representation learning. Well-aligned representation spaces learned from large-scale datasets such as CLIP [24] and ImageBind [14] are valued for their scalability and versatility. These spaces enable a wide range of applications, including auxiliary loss [42], intermediate representation [41], and evaluation metrics [16]. With this context, audio-visual representation spaces trained specifically on speech and 2D face videos have been proposed [6, 15, 29, 35]. For instance, SyncNet [6], a CNN-based model learns to detect audio-visual temporal synchronization and has been applied to several tasks, such as active speaker detection [1, 7] and 2D talking head generation [23, 44, 48]. Similarly, a transformer-based AV-HuBERT [29] has demonstrated remarkable effectiveness in various tasks, including lip reading [30], audio-visual translation [5], and 2D talking head generation [44]. While there has been significant progress in the 2D domain, advancements in the 3D domain remain under-explored. Yang et al. [47] extends the SyncNet architecture to accommodate speech and 3D face meshes; however, its application is limited to evaluating 3D talking heads. In this work, we demonstrate the versatility of our representation space as a plug-in module to improve the perceptual accuracy of the existing speech-driven 3D talking head generation models [11, 21, 46], as well as to assess their performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 474, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 298, + 714 + ], + "type": "text", + "content": "Evaluation metrics for speech-driven 3D talking head. The prevalent evaluation metric, Lip Vertex Error (LVE) [27], measures the L2 distance between predicted lip vertices and ground truth. Additional metrics, such as Upper Face Dynamics Deviation (FDD) [46] and Lip Readability Percentage (LRP) [21], consider different regions of facial motion. These metrics, however, focus on vertex-wise geometric differences between the ground-truth 3D facial motions and neglect speech-related information. To incorporate both speech and 3D face mesh data for evaluation, MultiTalk [37] introduces an Audio-Visual Lip Readability (AVLR) metric, which assesses perceptual accuracy of lip readability using a pre-trained Audio-Visual Speech Recognition model [2]. Yet, AVLR relies on speech and 2D face video rendered from the 3D face mesh, which may not align with the 3D talking head domain. To address these limitations, we introduce novel and comprehensive evaluation metrics that focus on diverse aspects of lip synchronization: Mean Temporal Misalignment (MTM), Perceptual Lip Readability Score (PLRS), and Speech-Lip Intensity Correlation Coefficient (SLCC)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 71, + 555, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 555, + 99 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 555, + 99 + ], + "type": "text", + "content": "3. Essential Criteria for Perceptually Accurate 3D Talking Head" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 105, + 555, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 555, + 237 + ], + "type": "text", + "content": "Generating a 3D talking head with lip movements that are perceptually accurate to human observers requires a clear understanding of the components that influence perceptual quality. Although existing works have focused on improving partial aspects of these talking head generation models, a comprehensive exploration for establishing a perceptually accurate 3D talking head model has barely been undertaken. Drawing inspirations from extensive research in existing works [10, 18, 21, 28, 34, 40, 43, 44] and our studies, we identify three fundamental criteria essential for achieving perceptually accurate lip movements in 3D talking heads." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 240, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 556, + 456 + ], + "type": "text", + "content": "Temporal Synchronization. This alignment is particularly crucial in media involving human speech, where viewers expect lip movements to precisely match the corresponding speech in time. Temporal misalignment between speech and lip movements indeed distracts viewers, reduces user experiences, and negatively impacts audience perception [25]. Vatakis et al. [43] find that viewers are particularly sensitive to speech-lip asynchrony compared to other audio-visual asynchrony, such as music. They note that mismatches become noticeable when speech precedes lip movements by more than " + }, + { + "bbox": [ + 313, + 240, + 556, + 456 + ], + "type": "inline_equation", + "content": "50~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 240, + 556, + 456 + ], + "type": "text", + "content": " or follows them by more than " + }, + { + "bbox": [ + 313, + 240, + 556, + 456 + ], + "type": "inline_equation", + "content": "220~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 240, + 556, + 456 + ], + "type": "text", + "content": ". These findings may hold the same in 3D talking head generation, where any misalignment between speech and lip movements can break the illusion of realism, leading to a diminished user experience and reducing the perceived authenticity of the virtual character. Thus, we define the temporal synchronization of the talking head as an important aspect of lip synchronization quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 459, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 662 + ], + "type": "text", + "content": "Lip Readability. Visemes (or lip movements) must correspond accurately to the speech phonemes to ensure the spoken words are visually intelligible [3]. This aspect is widely acknowledged as important in existing literature in both 2D [44] and 3D [10, 21] speech-driven talking head models, which leverage lip reading experts as an auxiliary guidance to improve the visual intelligibility of the spoken word. However, the mapping between speech and lip movements is not one-to-one, making lip readability challenging to define. For example, the size and shape of the opening mouth and the dynamic movement patterns of the lips in response to specific utterance differ at every moment [47] or per individual [32]. To capture this complexity, we define the lip readability within speech-mesh synchronized representation space learned from a large-scale dataset, capturing the comprehensive and nuanced correspondence between speech and lip movements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 665, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 557, + 715 + ], + "type": "text", + "content": "Expressiveness. Speech conveys not only linguistic content but also varies in intensity. For instance, a speaker may express the same text softly or loudly, with jaw and lip movements proportionally increasing as speech intensity rises," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21067" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 179, + 171 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 179, + 171 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 179, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 179, + 171 + ], + "type": "image", + "image_path": "bfc3b81fc34afd47d5a65a4cbf9cb533638836942a950c99d37a36af7b69549e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 174, + 296, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 174, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 296, + 240 + ], + "type": "text", + "content": "Table 1. Human studies on alignment criteria. [Left] Preference scores (1-3) for 3D talking heads with varying lip movement intensities paired with different speech intensities. [Right] Human preference between (A) samples with precise timing but low expressiveness, and (B) samples with high expressiveness but 100ms asynchrony—twice the commonly accepted 50ms threshold [43]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 197, + 83, + 293, + 159 + ], + "blocks": [ + { + "bbox": [ + 197, + 83, + 293, + 159 + ], + "lines": [ + { + "bbox": [ + 197, + 83, + 293, + 159 + ], + "spans": [ + { + "bbox": [ + 197, + 83, + 293, + 159 + ], + "type": "table", + "html": "
Samples
AB
Temp.✓×
Exp.×✓
Prefer (%)17.482.6
", + "image_path": "167a30a7caf733ca565415124d904efd54814e973ba34d453d55d137b1c8ed05.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 260, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 260, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 54, + 260, + 296, + 452 + ], + "type": "text", + "content": "which contributes to perceived expressiveness in real-world face recognition [18, 28, 40]. To demonstrate that the positive correlation of human preference between the intensity of speech and lip movements also exists in 3D talking head field, we conduct a human study by presenting 3D talking heads with varied lip movement intensities paired with different speech intensities in Table 1-[Left]. Participants prefer the lip movements with the intensity that match the intensity of speech. Despite this distinct human preference for synchronization between the intensity of corresponding speech and lip movements, this aspect has barely been explored in the talking head generation field. We thus define the expressiveness in 3D talking head as the correlation between speech and lip movement intensity, which is crucial for establishing genuine talking heads and is expected to guide future research aimed at improving perceptual lip synchronization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 454, + 296, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 658 + ], + "type": "text", + "content": "Focus of this work. Among the three criteria for perceptually accurate 3D talking head, we find that recent 3D talking head generation models achieve reasonable temporal alignment between speech and lip movements (see supplementary for a visualization of temporal synchronization). Furthermore, we design an A vs. B test, prompting participants to choose between two samples: (A) temporally synchronized one while lacking expressive synchronization, and (B) the other with expressive synchronization but with " + }, + { + "bbox": [ + 55, + 454, + 296, + 658 + ], + "type": "inline_equation", + "content": "100\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 454, + 296, + 658 + ], + "type": "text", + "content": " asynchrony. Table 1-[Right] shows that users prefer sample B, suggesting that humans may prioritize expressive synchronization over strict temporal alignment. This insight directs our focus toward enhancing all three aspects to better capture perceptual realism in 3D talking head. Details of the human study are in the supplementary material. Before introducing metrics to assess the criteria, we present our synchronized representation for designing our perceptual loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 669, + 293, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 293, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 293, + 685 + ], + "type": "text", + "content": "4. Speech-Mesh Synchronized Representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "We hypothesize that a desirable representation space exists to meet the three criteria defined in Sec. 3. Motivated by this" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "hypothesis, we develop a rich speech-mesh synchronized representation which captures the intricate correspondence between speech and the 3D face mesh. We found that our learned representation exhibits desirable properties, and we adopt it as a perceptual loss to improve the perceptual accuracy of existing models with respect to these three criteria." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 147, + 556, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 556, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 556, + 327 + ], + "type": "text", + "content": "Overview. Directly learning a synchronized speech-mesh representation presents challenges due to the scarcity of speech-3D face mesh datasets. One potential solution is to construct pseudo-GT 3D face meshes using reconstruction models [13], although relying solely on pseudo-GT may not suffice for building a robust representation. To overcome this, we leverage the extensive knowledge from the speech-2D lip video representations and adapt this to accommodate 3D face meshes. Specifically, we propose a two-stage training process: in stage 1, we learn an audio-visual speech representation that accurately reflects lip movements from the unlabeled in-the-wild 2D synchronized talking face video dataset, LRS3 [1]. Subsequently, we leverage the pre-trained speech representation from stage 1, using it as the anchor space to learn a synchronized speech-mesh representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 556, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 450 + ], + "type": "text", + "content": "Stage 1. Learning audio-visual speech representation. This stage aims to learn a rich speech-2D lip video representation that effectively captures the correlation between varying speech characteristics and lip dynamics. Motivated by prior works [15, 35], we extend the integration of masked autoencoder (MAE) and cross-modal contrastive learning to learn a synchronized speech-2D lip video representation using 2D videos. The architecture for stage 1 consists of two modality-specific encoders, a cross-modal fusion encoder, and two modality-specific decoders. (see Fig. 2-[Stage 1])." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": "Given a speech and 2D lip video pair, " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}_{\\mathbf{s}},\\mathbf{X}_{\\mathbf{v}})\\in D" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " we begin by patchifying and tokenizing speech spectrograms and video frames into speech and video tokens as " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = (\\mathbf{s}_1,\\dots ,\\mathbf{s}_N)" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = (\\mathbf{v}_1,\\dots ,\\mathbf{v}_M)" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i,\\mathbf{v}_j\\in" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^H" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": ". We then randomly mask out " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "P\\%" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " portion of tokens. The remaining unmasked speech tokens " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{unmask}" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " and video tokens " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{unmask}" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " are respectively fed into the speech encoder " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "E_{s}" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " and video encoder " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": ", each consisting of " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " transformer layers. Each encoder extracts uni-modal embeddings, " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathbf{s}}^{l} = E_{s}^{l}(\\mathbf{S}^{unmask})" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathbf{v}}^{l} = E_{v}^{l}(\\mathbf{V}^{unmask})" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "l\\in (1,\\ldots ,N_e)" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " denote the layer indices. Also, the mean pooled speech and video embeddings, " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c_s^l} = \\mathrm{MeanPool}(\\mathbf{Z_s^l})" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c_v^l} = \\mathrm{MeanPool}(\\mathbf{Z_v^l})" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": ", are derived from the corresponding uni-modal embeddings. Following the uni-modal encoders, we introduce a multi-modal fusion encoder " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "E_{sv}" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " to exploit complementary information from each extracted uni-modal embedding. The speech and video embeddings are jointly processed through this encoder, resulting in multi-modal fusion embeddings " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F_s}" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F_v}" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "[\\mathbf{F_s},\\mathbf{F_v}] = E_{sv}([\\mathbf{Z_s},\\mathbf{Z_v}])" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "[,\\cdot ]" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " denotes concatenation. With the fusion embeddings " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F_s}" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F_v}" + }, + { + "bbox": [ + 313, + 450, + 557, + 715 + ], + "type": "text", + "content": ", we utilize each modality's decoder to reconstruct the original signals. Specifically, we employ a" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21068" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 67, + 303, + 227 + ], + "blocks": [ + { + "bbox": [ + 88, + 67, + 303, + 227 + ], + "lines": [ + { + "bbox": [ + 88, + 67, + 303, + 227 + ], + "spans": [ + { + "bbox": [ + 88, + 67, + 303, + 227 + ], + "type": "image", + "image_path": "d974d05e76c6c8e9a4c6306fb0e15cd76881e5ffe7df953523490cbeac87cc92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "type": "text", + "content": "Figure 2. Pipeline of speech-mesh synchronized representation learning. We train our speech-mesh representation space in a two-stage manner. In the first stage, we learn a rich audio-visual representation in 2D domain to capture the synchronization between lip movement and speech. In the second stage, we train the 3D mesh encoder to align the 3D mesh space with the frozen speech space. As an application of our speech-mesh representation space, we propose a plug-in perceptual loss to 3D talking head models to enhance the quality of lip movements." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 305, + 66, + 522, + 228 + ], + "blocks": [ + { + "bbox": [ + 305, + 66, + 522, + 228 + ], + "lines": [ + { + "bbox": [ + 305, + 66, + 522, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 66, + 522, + 228 + ], + "type": "image", + "image_path": "8ba9880e77762d0057f882c1cd0af5b67e3498063f827cac49f1012342d34662.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "text", + "content": "speech decoder " + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "text", + "content": " and a video decoder " + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "inline_equation", + "content": "D_{v}" + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "text", + "content": ", each consisting of " + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "inline_equation", + "content": "N_{d}" + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "text", + "content": " transformer layers. We pad each multi-modal fusion embedding with trainable masked tokens at their original positions, resulting in " + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{s}}^{\\prime}" + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{v}}^{\\prime}" + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "text", + "content": ". Each decoder then reconstructs the respective signals; the reconstructed speech spectrogram and video tokens as " + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}^{mask} = D_s(\\mathbf{F}_{\\mathbf{s}}^{\\prime})" + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}^{mask} = D_v(\\mathbf{F}_{\\mathbf{v}}^{\\prime})" + }, + { + "bbox": [ + 54, + 285, + 296, + 370 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 370, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 370, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 54, + 370, + 296, + 538 + ], + "type": "text", + "content": "Training the model involves two objectives: learning the cross-modal alignment between the speech-2D lip video signals, while reconstructing their original signals. Given " + }, + { + "bbox": [ + 54, + 370, + 296, + 538 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 54, + 370, + 296, + 538 + ], + "type": "text", + "content": " speech-2D lip video token pairs in a batch, " + }, + { + "bbox": [ + 54, + 370, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_i,\\mathbf{V}_i\\}_{i = 1}^B" + }, + { + "bbox": [ + 54, + 370, + 296, + 538 + ], + "type": "text", + "content": ", we treat the true speech-2D lip video pair as positive samples, while the others in the batch are considered negative samples. To encourage alignment between temporally synchronized speech-2D lip video instances, we utilize a cross-modal contrastive learning strategy, InfoNCE [20]. We maximize the cosine similarity between positive sample of the mean pooled speech embedding " + }, + { + "bbox": [ + 54, + 370, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathbf{s},i}^{l} = \\mathrm{MeanPool}(E_{s}^{l}(\\mathbf{S}_{i}^{\\mathrm{unmask}}))" + }, + { + "bbox": [ + 54, + 370, + 296, + 538 + ], + "type": "text", + "content": " and the corresponding synchronized mean pooled video embedding " + }, + { + "bbox": [ + 54, + 370, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathbf{v},i}^{l} = \\mathrm{MeanPool}(E_{v}^{l}(\\mathbf{V}_{i}^{\\mathrm{unmask}}))" + }, + { + "bbox": [ + 54, + 370, + 296, + 538 + ], + "type": "text", + "content": ". We first define the speech-centric loss as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 543, + 296, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 543, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 88, + 543, + 296, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {V}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {v} , i} ^ {l} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {v} , j} ^ {l} / \\tau\\right)}, \\tag {1}", + "image_path": "70537add4551166e0512d4029784acce55a537dd7a2efdfd534f0d54950bd86c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "text", + "content": " is a temperature hyperparameter. Also, we make the objective symmetric by defining a video-centric loss as " + }, + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{V}\\rightarrow \\mathrm{S}}" + }, + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "text", + "content": ". We sum " + }, + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{S}\\rightarrow \\mathrm{V}}" + }, + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{V}\\rightarrow \\mathrm{S}}" + }, + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 579, + 296, + 628 + ], + "type": "text", + "content": " selected encoder layers to obtain our final contrastive learning objective:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 632, + 296, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 632, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 104, + 632, + 296, + 651 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {I n f o N C E}} = \\sum_ {l \\in L} \\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {V}} + \\mathcal {L} _ {\\mathrm {V} \\rightarrow \\mathrm {S}}. \\tag {2}", + "image_path": "881f9f519f749cc80fdd54cad0dcb77d8a025e7bf3391e717825be204106d605.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 679 + ], + "type": "text", + "content": "For reconstruction, the model is trained in a self-supervised manner by minimizing the reconstruction loss " + }, + { + "bbox": [ + 55, + 654, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MAE}}" + }, + { + "bbox": [ + 55, + 654, + 296, + 679 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 682, + 296, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 682, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 682, + 296, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M A E}} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left[ \\frac {\\sum \\| \\hat {\\mathbf {S}} _ {i} ^ {m a s k} - \\mathbf {S} _ {i} ^ {m a s k} \\| _ {F} ^ {2}}{| \\mathbf {S} _ {i} ^ {m a s k} |} + \\frac {\\sum \\| \\hat {\\mathbf {V}} _ {i} ^ {m a s k} - \\mathbf {V} _ {i} ^ {m a s k} \\| _ {F} ^ {2}}{| \\mathbf {V} _ {i} ^ {m a s k} |} \\right], \\tag {3}", + "image_path": "14fd620994c1dfbbf6346e7712b49891010f884eadd236bea70f9b801a25621a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 285, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 285, + 555, + 320 + ], + "type": "inline_equation", + "content": "|\\mathbf{S}_i^{mask}|" + }, + { + "bbox": [ + 313, + 285, + 555, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 285, + 555, + 320 + ], + "type": "inline_equation", + "content": "|\\mathbf{V}_i^{mask}|" + }, + { + "bbox": [ + 313, + 285, + 555, + 320 + ], + "type": "text", + "content": " denote the number of masked speech and video tokens, respectively. Overall, our objective function is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 384, + 325, + 555, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 325, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 384, + 325, + 555, + 337 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {M A E}} + \\lambda \\mathcal {L} _ {\\text {I n f o N C E}}, \\tag {4}", + "image_path": "156146369b360889dc704f2499a5a71754748e68b90baf97df1e31534bc62242.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 339, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 339, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 339, + 555, + 350 + ], + "type": "text", + "content": " is the weight factor for cross-modal contrastive loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 351, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 351, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 312, + 351, + 555, + 435 + ], + "type": "text", + "content": "We observe that the representation space trained with the transformer architecture and a rich and large-scale 2D face video dataset in this way already possesses the desired properties we pursue, illustrated in Fig. 1 (see supplementary for visualizations of pre-trained speech representation). Motivated by this, we transfer these emergent properties to the speech-mesh representation space as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 438, + 556, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 545 + ], + "type": "text", + "content": "Stage 2. Learning speech-mesh representation. In this stage, we design a 3D mesh encoder that maps 3D face mesh to the speech representation pre-trained in stage 1. This pretrained speech representation, derived from rich speech-2D face video data, serves as a robust anchor space for learning the correlation between diverse speech characteristics and 3D facial motions. We use contrastive learning loss to align 3D facial motion embeddings with the anchored speech representations, as shown in Fig. 2-[Stage 2]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "text", + "content": "Given a speech and 3D face mesh pair " + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}_{\\mathbf{s}},\\mathbf{X}_{\\mathbf{m}})\\in D" + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "text", + "content": " we patchify and tokenize speech spectrograms into speech tokens S, and map these into pre-trained uni-modal mean pooled speech embeddings " + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{c_s^l} = \\mathrm{MeanPool}(E_s^l (\\mathbf{S}))" + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "text", + "content": " . Similarly, we patchify 3D face mesh into mesh tokens M, and feed them into the 3D mesh encoder " + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "text", + "content": " transformer layers to extract mesh embeddings " + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathbf{m}}^{l} = E_{m}^{l}(\\mathbf{M})" + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "text", + "content": " and the corresponding mean-pooled mesh embedding " + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{c_m^l}" + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "text", + "content": " . For the learning objective, given " + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "text", + "content": " speech-3D face mesh token pairs in a batch, " + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_i,\\mathbf{M}_i\\}_{i = 1}^B" + }, + { + "bbox": [ + 313, + 545, + 557, + 679 + ], + "type": "text", + "content": " , we first define speech-anchored loss as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 345, + 683, + 555, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 683, + 555, + 716 + ], + "spans": [ + { + "bbox": [ + 345, + 683, + 555, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {M}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {m} , i} ^ {l} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\mathbf {c} _ {\\mathbf {s} , i} ^ {l} \\cdot \\mathbf {c} _ {\\mathbf {m} , j} ^ {l} / \\tau\\right)}, \\tag {5}", + "image_path": "1ed1cf162a8cc30a73c352da03b2b74fbaec7063144a68a5a5b6e706f9de9a4c.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21069" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "as well as " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{M}\\to \\mathrm{S}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " which is the mesh-anchored loss. Similar to Eq. (2), we sum " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{S}\\rightarrow \\mathrm{M}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{M}\\rightarrow \\mathrm{S}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " selected encoder layers to train the 3D mesh encoder " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " while the speech encoder fixed. As a result, this two-stage training approach yields a robust speech-mesh representation, which significantly enhances the alignment between the speech and 3D face mesh modalities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 159, + 294, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 294, + 242 + ], + "type": "text", + "content": "Adopting it as a perceptual loss. A key application of the speech-mesh representation learned through the two-stage training process is its use as a perceptual loss to enhance the perceptual accuracy of the 3D talking head model (see Fig. 2). Leveraging this representation as a perceptual loss ensures that the generated lip movements are perceptually accurate and aligned well with the speech." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 294, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 294, + 279 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 294, + 279 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 55, + 243, + 294, + 279 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 243, + 294, + 279 + ], + "type": "text", + "content": " speech and generated 3D face mesh pairs in a batch, " + }, + { + "bbox": [ + 55, + 243, + 294, + 279 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_i,\\hat{\\mathbf{M}}_i\\}_{i = 1}^B" + }, + { + "bbox": [ + 55, + 243, + 294, + 279 + ], + "type": "text", + "content": ", we define our perceptual loss with the symmetric InfoNCE loss (Eq. (2)) as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 286, + 295, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 286, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 112, + 286, + 295, + 312 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p e r c p}} = \\sum_ {l \\in L} \\mathcal {L} _ {\\mathrm {S} \\rightarrow \\mathrm {M}} + \\mathcal {L} _ {\\mathrm {M} \\rightarrow \\mathrm {S}}. \\tag {6}", + "image_path": "14eef7968400abd26bcfb96412626bd4c71dd2e3bc3dcf486d6aa0a701e473ce.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 319, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 367 + ], + "type": "text", + "content": "Our perceptual loss encourages synchronized speech and mesh embeddings to pull closer together, while unsynchronized ones to push apart. The effectiveness and analysis of our perceptual loss will be discussed in later sections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 376, + 168, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 168, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 168, + 388 + ], + "type": "text", + "content": "5. Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "text", + "content": "In this section, we describe our proposed evaluation metrics that assess each criterion impacting the quality of 3D lip accuracy, as discussed in Sec. 3. Here, we outline the high-level concepts of our proposed metrics. Additional details and experiments are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "text", + "content": "Mean Temporal Misalignment (MTM). To measure the temporal discrepancy between speech and corresponding lip movements, temporal correspondence annotations, such as onset times for each modality, would typically be required. As a proxy, we determine temporal correspondence and measure temporal discrepancies between the ground truth and predicted lip vertex displacement sequences by using Derivative Dynamic Time Warping (DDTW) [19], which robustly identifies local structural similarities compared to standard Dynamic Time Warping (DTW) [4]. For simplicity, we focus on the central vertices of the upper and lower lips when extracting displacement sequences, and use local extrema in the DDTW process to measure temporal misalignment by pinpointing precise time steps for mouth opening and closing events. Consequently, Mean Temporal Misalignment (MTM) " + }, + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\overline{\\Delta t}" + }, + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\overline{\\Delta t} = \\frac{1}{K}\\sum_{k=1}^{K}\\Delta t_k" + }, + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "text", + "content": " is the total number of video clips and " + }, + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\Delta t_k" + }, + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "text", + "content": " is the averaged temporal misalignment of the " + }, + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 459, + 296, + 674 + ], + "type": "text", + "content": "-th video clip." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Perceptual Lip Readability Score (PLRS). While Lip Vertex Error (LVE) measures the accuracy of generated lip articulations against the ground truth, it does not fully assess" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "whether lip movements are perceptually aligned with the given speech. To address this, we leverage our speech-mesh representation in Sec. 4 as a perceptual lip readability evaluator. We compute the perceptual alignment using the cosine similarity of the mean pooled speech and mesh embeddings. Since this representation has learned a rich distribution of speech correspondences across various facial movements, our metric correlates highly with human perception, measuring perceptual lip movement alignment more accurately than LVE (see supplementary for the human study on metrics)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 194, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 266 + ], + "type": "text", + "content": "Speech and Lip Intensity Correlation Coefficient (SLCC). As shown in Table 1-[Left], humans prefer aligned intensity between speech and lip movement. Thus, the intensity of generated lip movements should positively correlate with the corresponding input speech. To quantify this, we define the Speech and Lip Correlation Coefficient " + }, + { + "bbox": [ + 313, + 194, + 555, + 266 + ], + "type": "inline_equation", + "content": "r_{SL}" + }, + { + "bbox": [ + 313, + 194, + 555, + 266 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 275, + 555, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 275, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 351, + 275, + 555, + 299 + ], + "type": "interline_equation", + "content": "r _ {S L} = \\frac {\\sum_ {k = 1} ^ {K} (\\mathrm {S I} _ {k} - \\bar {\\mathrm {S I}}) (\\mathrm {L I} _ {k} - \\bar {\\mathrm {L I}})}{\\sqrt {\\sum_ {k = 1} ^ {K} (\\mathrm {S I} _ {k} - \\bar {\\mathrm {S I}}) ^ {2}} \\sqrt {\\sum_ {k = 1} ^ {K} (\\mathrm {L I} _ {k} - \\bar {\\mathrm {L I}}) ^ {2}}}, \\tag {7}", + "image_path": "2035aa1b1e92b9ccde38748c436b28072df7847279fccbfb52ffcec3928186b6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 306, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 306, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\mathrm{SI}_k" + }, + { + "bbox": [ + 313, + 306, + 555, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 306, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\mathrm{LI}_k" + }, + { + "bbox": [ + 313, + 306, + 555, + 390 + ], + "type": "text", + "content": " denote Speech (SI) and Lip Intensity (LI), respectively, " + }, + { + "bbox": [ + 313, + 306, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\bar{\\mathrm{SI}} = \\frac{1}{K}\\sum_{k = 1}^{K}\\mathrm{SI}_k" + }, + { + "bbox": [ + 313, + 306, + 555, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 306, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\bar{\\mathrm{LI}} = \\frac{1}{K}\\sum_{k = 1}^{K}\\mathrm{LI}_k" + }, + { + "bbox": [ + 313, + 306, + 555, + 390 + ], + "type": "text", + "content": ". We define SI using speech loudness, specifically the z-normalized Root Mean Square (RMS) value, which is a widely accepted measure of speech intensity in signal processing. To define LI, we measure the averaged lip displacement value of a video clip, followed by the z-normalization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 401, + 395, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 401, + 395, + 415 + ], + "spans": [ + { + "bbox": [ + 314, + 401, + 395, + 415 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 422, + 555, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 470 + ], + "type": "text", + "content": "We first outline the evaluation setup, and then present thorough analyses of the experimental results. Due to the space limitation, we present more implementation details and additional experiments in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 478, + 441, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 441, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 441, + 491 + ], + "type": "text", + "content": "6.1. Experimental Settings" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": "Datasets. Most of the existing speech-driven 3D talking head generation methods rely on VOCASET [8] and BIWI [12] to train and test the models. However, these datasets have limited ranges of facial motion patterns due to the small dataset scale and restricted intensity range, which restricts their ability to fully capture the intricate relationship between speech and 3D face mesh. To address the lack of training dataset, we construct two large-scale speech-3d face mesh benchmark datasets, LRS3-3D and MEAD-3D, by processing LRS3 and MEAD videos using two monocular face reconstruction methods: SPECTRE [13] for LRS3 [1], which ensures accurate lip movements, and SMIRK [26] for MEAD [45], which captures diverse speech and lip movement intensities. We use LRS3 in the first stage and LRS3-3D in the second stage to train our speech-mesh synchronized representation. Then, for base model training and evaluations, we adopt two configurations: (1) training and testing with VOCASET, in line with existing work, and (2)" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21070" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 179 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 179 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 179 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 179 + ], + "type": "table", + "html": "
MethodTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)SLCC / Δ (↓)
VOCASET--0.4090.34 / -
FaceFormer [11]53.63.3570.3680.26 / 0.08
+ Ours rep.52.23.0910.4630.37 / 0.03
CodeTalker [46]61.83.7000.3810.38 / 0.04
+ Ours rep.60.93.5790.3880.35 / 0.01
SelfTalk [21]50.12.9710.4140.41 / 0.07
+ Ours rep.49.22.9240.4180.35 / 0.01
", + "image_path": "89dd1f89657db398b2a19e5e5d89659d172252aa77fa2bd056c1b7c32a58490c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 275, + 296, + 369 + ], + "blocks": [ + { + "bbox": [ + 56, + 275, + 296, + 369 + ], + "lines": [ + { + "bbox": [ + 56, + 275, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 296, + 369 + ], + "type": "image", + "image_path": "8ee7c40b0f81bf6cdc552ff6b15fdbbdc18de6958c846f2a6e059060076b58ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 371, + 296, + 404 + ], + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 404 + ], + "type": "text", + "content": "Figure 3. Qualitative results of the effectiveness of our perceptual loss for lip readability. Our perceptual loss guides baselines [11, 21, 46] to generate perceptually accurate lip movements." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 440 + ], + "type": "text", + "content": "combining MEAD-3D with VOCASET during training to endow expressiveness and testing on MEAD-3D." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 442, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 296, + 479 + ], + "type": "text", + "content": "Base methods. We use three state-of-the-art 3D talking head generation models [11, 21, 46] to evaluate the effectiveness of our perceptual loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 482, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 296, + 578 + ], + "type": "text", + "content": "Metrics. To comprehensively evaluate the three aspects of lip synchronization, we assess MTM, PLRS, and SLCC, corresponding to temporal synchronization, lip readability, and expressiveness, respectively. Additionally, we compute the level-wise SLCC for the MEAD-3D test set, which includes three distinct emotional intensity levels, to evaluate the expressive capability of 3D talking head generation models. We also measure LVE as part of the lip readability evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 587, + 242, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 242, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 242, + 601 + ], + "type": "text", + "content": "6.2. Experimental Results and Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 653 + ], + "type": "text", + "content": "We conduct evaluations to assess the effectiveness of our speech-mesh synchronized representation and the incorporation of an expressive speech-3D face mesh paired dataset (i.e., MEAD-3D) in enhancing the three criteria." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "How well do existing 3D talking head models achieve lip synchronization in all three aspects? The results are summarized in Table 2. For temporal synchronization, most base models achieve MTM values between 50 and 60ms, indicating performance close to the acceptable asynchrony" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 321, + 70, + 547, + 199 + ], + "blocks": [ + { + "bbox": [ + 55, + 180, + 296, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 268 + ], + "type": "text", + "content": "Table 2. Quantitative results of lip synchronization on VO-CASET [8] test set. We evaluate the base models on our proposed lip synchronization metrics. We denote " + }, + { + "bbox": [ + 55, + 180, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 180, + 296, + 268 + ], + "type": "text", + "content": " as the difference in SLCC between the model and those measured on the data distribution. A lower " + }, + { + "bbox": [ + 55, + 180, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 180, + 296, + 268 + ], + "type": "text", + "content": " indicates the model more closely represents the intensity correlation of the dataset. We demonstrate the effectiveness of our representation in consistently enhancing all three aspects of lip synchronization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 70, + 547, + 199 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 547, + 199 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 547, + 199 + ], + "type": "table", + "html": "
MethodPerceptual LossTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)SLCC / Δ (↓)
VOCASET---0.4090.34 / -
FaceFormer [11]X53.63.3570.3680.26 / 0.08
3D SyncNet55.63.3160.4350.38 / 0.04
Ours w/o 2D prior55.33.2780.4000.42 / 0.08
Ours w/ 2D prior52.23.0910.4630.37 / 0.03
CodeTalker [46]X61.83.7000.3810.38 / 0.04
3D SyncNet59.64.3190.3790.14 / 0.20
Ours w/o 2D prior55.93.5790.3740.23 / 0.11
Ours w/ 2D prior60.93.5790.3880.35 / 0.01
SelfTalk [21]X50.12.9710.4140.41 / 0.07
3D SyncNet49.52.9410.4050.35 / 0.01
Ours w/o 2D prior54.43.1490.4170.39 / 0.05
Ours w/ 2D prior49.22.9240.4180.35 / 0.01
", + "image_path": "8a127620c967bce4a0245591598dba9296611634e054bc221485281ab64b2018.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 200, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 200, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 553, + 243 + ], + "type": "text", + "content": "Table 3. Ablation study on architectural choice and 2D prior knowledge. We validate the effectiveness of the transformer-based architecture and curriculum learning with a pre-trained 2D speech representation by ablating them from our proposed representation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 340, + 254, + 529, + 354 + ], + "blocks": [ + { + "bbox": [ + 340, + 254, + 529, + 354 + ], + "lines": [ + { + "bbox": [ + 340, + 254, + 529, + 354 + ], + "spans": [ + { + "bbox": [ + 340, + 254, + 529, + 354 + ], + "type": "image", + "image_path": "6b22ebb0cf30f6c03f5d97d37a25d0c1543b1b5a51ac6a66c0b9ce71c975fa98.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 357, + 554, + 401 + ], + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 401 + ], + "type": "text", + "content": "Figure 4. t-SNE plot of ablation study. We plot the t-SNE graph for each perceptual critic model. We represent the features with same phoneme as same color. Squared and circled points denote mesh and speech features from each representation, respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 340, + 405, + 526, + 510 + ], + "blocks": [ + { + "bbox": [ + 340, + 405, + 526, + 510 + ], + "lines": [ + { + "bbox": [ + 340, + 405, + 526, + 510 + ], + "spans": [ + { + "bbox": [ + 340, + 405, + 526, + 510 + ], + "type": "image", + "image_path": "32f8920b1801e2fdbc270e8c0250e147e773508a0f3115cb9b5c6126d637af33.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 514, + 555, + 568 + ], + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 568 + ], + "type": "text", + "content": "Figure 5. Behaviors of our representation in temporal and expressiveness sensitivity. We demonstrate the effectiveness of our representation in temporal synchronization and expressiveness using a cosine similarity graph and speech feature plots, respectively. We color the point as low, medium, and high intensity." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 579, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 640 + ], + "type": "text", + "content": "threshold. Regarding lip readability, SelfTalk [21] achieves the best performance, while FaceFormer [11] has the lowest PLRS score. Furthermore, CodeTalker [46] demonstrates the closest SLCC values to the ground truth VOCASET mesh, while FaceFormer exhibits the highest SLCC discrepancy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 641, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 677 + ], + "type": "text", + "content": "Does our speech-mesh representation improve lip synchronization? Yes. Table 2 and Fig. 3 show consistent improvements in the three aspects with our perceptual loss." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "What makes our speech-mesh representation have lip synchronization ability? We hypothesize that the transformer-based architecture and curriculum learning with a pre-trained" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21071" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 69, + 517, + 180 + ], + "blocks": [ + { + "bbox": [ + 92, + 69, + 517, + 180 + ], + "lines": [ + { + "bbox": [ + 92, + 69, + 517, + 180 + ], + "spans": [ + { + "bbox": [ + 92, + 69, + 517, + 180 + ], + "type": "image", + "image_path": "c4e0c0f81521d8659774a5cab61ca4c5dfb8aec956e971a7eb01ceb6bdf53076.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 555, + 206 + ], + "type": "text", + "content": "Figure 6. Qualitative results for the expressiveness. Given high and low intensity levels of speech, models trained on both MEAD-3D and VOCASET show more expressive lip movements compared to those trained on VOCASET alone, and even better with our perceptual loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 215, + 297, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 215, + 297, + 443 + ], + "spans": [ + { + "bbox": [ + 54, + 215, + 297, + 443 + ], + "type": "text", + "content": "audio-visual speech representation contribute significantly to improved lip synchronization. To validate this, we conduct ablation studies, summarized in Table 3, examining the roles of pre-trained speech representation and architectural design. Without the pre-trained speech representation, base models show lower performance across metrics, and CNN-based architectures, 3D SyncNet, do not clearly show effectiveness, while ours with 2D prior consistently show improvement. The t-SNE plots of perceptual critic models in Fig. 4 show that, in (a) 3D SyncNet, speech and mesh features corresponding to the same phoneme group are separated. In (b) ours without 2D priors, speech and mesh features lack separation and appear scattered without clustering. Our representation (c), notably, tends to form more distinct clusters according to phonemes, with vowels and consonants grouped closely, potentially contributing to enhancing lip readability. We also observe a directional progression in the feature space, shifting from phonemes with mouth opening " + }, + { + "bbox": [ + 54, + 215, + 297, + 443 + ], + "type": "inline_equation", + "content": "(e.g., /aj/" + }, + { + "bbox": [ + 54, + 215, + 297, + 443 + ], + "type": "text", + "content": " to those with mouth closing " + }, + { + "bbox": [ + 54, + 215, + 297, + 443 + ], + "type": "inline_equation", + "content": "(e.g., /f/" + }, + { + "bbox": [ + 54, + 215, + 297, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 446, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 555 + ], + "type": "text", + "content": "We also examine our representation's performance in other aspects, temporal synchronization and expressiveness. Figure 5-(a) demonstrates temporal sensitivity, as cosine similarity drops when temporal misalignment is introduced between input speech and 3D face mesh. In Fig. 5-(b), we plot speech features at varying speech intensities, showing a directional trend as intensity increases from lowest to highest. Figures 4 and 5 imply that our representation holds favorable properties discussed in Fig. 1 for the three criteria." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 295, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 569 + ], + "type": "text", + "content": "Can we unlock the expressive power of 3D talking heads?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": "Likely. Since VOCASET [8] lacks the range of diverse speech and lip movement intensities, evaluating expressive power requires testing on a dataset with a broader range of intensities. To study this, we examine the expressiveness of the base models on the MEAD-3D dataset, which includes three intensity levels. We assess SLCC at each intensity level to evaluate expressiveness and identify any expressiveness bounds as the level of intensity increases. We first test the base models trained on VOCASET [8] against the MEAD-3D test set. Table 4 shows these models demonstrate limited expressiveness, with SLCC values showing minimal increase across intensity levels. We hypothesize that" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 213, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 315, + 213, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 315, + 213, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 213, + 553, + 319 + ], + "type": "table", + "html": "
MethodTemporal SynchronizationLip ReadabilityExpressiveness
MTM (↓)LVE (↓)PLRS (↑)Lv1SLCC / Δ (↓)
Lv2Lv3Avg
MEAD-3D--0.2300.24 / -0.30 / -0.39 / -0.42 / -
FaceFormer [11]59.63.2070.2990.08 / 0.160.07 / 0.230.07 / 0.320.06 / 0.36
+ MEAD-3D59.51.1390.1760.26 / 0.020.30 / 0.000.34 / 0.050.35 / 0.07
+ Ours rep.55.81.1140.3060.27 / 0.030.27 / 0.030.32 / 0.070.33 / 0.09
CodeTalker [46]60.73.2360.2940.02 / 0.220.03 / 0.270.03 / 0.360.02 / 0.40
+ MEAD-3D60.92.9540.1540.09 / 0.150.12 / 0.180.06 / 0.330.11 / 0.31
+ Ours rep.58.62.7050.2210.18 / 0.060.29 / 0.010.31 / 0.080.31 / 0.11
SelfTalk [21]53.43.3960.2940.14 / 0.100.14 / 0.160.17 / 0.220.15 / 0.27
+ MEAD-3D54.21.2380.2160.16 / 0.080.28 / 0.020.32 / 0.070.31 / 0.11
+ Ours rep.52.71.1920.2300.17 / 0.070.29 / 0.010.34 / 0.050.33 / 0.09
", + "image_path": "0ec0f429a3450ed7473f3dc589663e6129187108f86021318a921cd449759d16.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 319, + 555, + 364 + ], + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 364 + ], + "type": "text", + "content": "Table 4. Quantitative results of lip synchronization on MEAD-3D test set. We evaluate the base models on the MEAD-3D test set. We also compute the level-wise SLCC to evaluate the expressive capability of the models." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 371, + 556, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 556, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 556, + 504 + ], + "type": "text", + "content": "this limitation arises, because VOCASET's smaller scale and intensity range restrict the model's ability to learn relationships between speech and lip intensity. To address this, we integrate MEAD-3D with VOCASET for training, aiming to boost expressiveness in the base models. This approach consistently improves SLCC across all intensity levels. However, simply adding MEAD-3D degrades lip synchronization metrics, except for expressiveness, i.e., nontrivial. To counterbalance this, we leverage our perceptual loss, which effectively mitigates the degradation introduced by MEAD-3D while improving expressiveness (see Fig. 6)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "This paper addresses challenges in existing 3D talking head generation models, which often overlook the true correspondence between speech and lip movements, making it difficult to accurately link lip movements with varying speech characteristics. To overcome this issue, we identify three essential aspects—Temporal Synchronization, Lip Readability, and Expressiveness—that influence the perceptual quality of lip movements, and develop specific metrics for each aspect. We introduce a speech-mesh synchronized representation that exhibits these emergent properties and adopt it as a perceptual loss. Our extensive analyses demonstrate that our perceptual loss consistently enhances models across three aspects. We believe that our defined aspects will guide future research in generating more realistic 3D talking heads, and our representation will serve as a key component." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21072" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 251 + ], + "type": "text", + "content": "Acknowledgments. This research was supported by a grant from KRAFTON AI, and was also partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2021-II212068, Artificial Intelligence Innovation Hub; No.RS-2023-00225630, Development of Artificial Intelligence for Text-based 3D Movie Generation; No. RS-2024-00457882, National AI Research Lab Project) and Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2024 (Project Name: Development of barrier-free experiential XR contents technology to improve accessibility to online activities for the physically disabled, Project Number: RS-2024-00396700, Contribution Rate: " + }, + { + "bbox": [ + 55, + 72, + 297, + 251 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 72, + 297, + 251 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 263, + 115, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 115, + 276 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 115, + 276 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 284, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 284, + 297, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 284, + 297, + 326 + ], + "spans": [ + { + "bbox": [ + 61, + 284, + 297, + 326 + ], + "type": "text", + "content": "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 2, 3, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 329, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 329, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 61, + 329, + 296, + 373 + ], + "type": "text", + "content": "[2] Mohamed Anwar, Bowen Shi, Vedanuj Goswami, Wei-Ning Hsu, Juan Pino, and Changhan Wang. Muavic: A multilingual audio-visual corpus for robust speech recognition and robust speech-to-text translation. In INTERSPEECH 2023, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 375, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 296, + 407 + ], + "type": "text", + "content": "[3] Helen L Bear and Richard Harvey. Phoneme-to-viseme mappings: the good, the bad, and the ugly. Speech Communication, 2017. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 409, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 453 + ], + "type": "text", + "content": "[4] Donald J Berndt and James Clifford. Using dynamic time warping to find patterns in time series. In Proceedings of the 3rd international conference on knowledge discovery and data mining, 1994. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 454, + 296, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 296, + 497 + ], + "type": "text", + "content": "[5] Jeongsoo Choi, Se Jin Park, Minsu Kim, and Yong Man Ro. Av2av: Direct audio-visual speech to audio-visual speech translation with unified audio-visual speech representation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 499, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 296, + 531 + ], + "type": "text", + "content": "[6] J. S. Chung and A. Zisserman. Out of time: automated lip sync in the wild. In Workshop on Multi-view Lip-reading, ACCV, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "type": "text", + "content": "[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 568, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 601 + ], + "type": "text", + "content": "[8] Daniel Cudeiro, Timo Bolkart, Cassidy Laidlaw, Anurag Ranjan, and Michael J Black. Capture, learning, and synthesis of 3d speaking styles. In CVPR, 2019. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "text", + "content": "[9] Radek Daněček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. In SIGGRAPH Asia 2023 Conference Papers, pages 1-13, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[10] Han EunGi, Oh Hyun-Bin, Kim Sung-Bin, Corentin Nivelet Etcheberry, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Enhancing speech-driven 3d facial animation with audio-visual guidance from lip reading expert. In Interspeech 2024, pages 2940–2944, 2024. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[11] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In CVPR, 2022. 2, 3, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "type": "text", + "content": "[12] Gabriele Fanelli, Juergen Gall, Harald Romsdorfer, Thibaut Weise, and Luc Van Gool. A 3-d audio-visual corpus of affective communication. IEEE TMM, 2010. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "text", + "content": "[13] Panagiotis P. Filntisis, George Retsinas, Foivos Paraperas-Papantoniou, Athanasios Katsamanis, Anastasios Roussos, and Petros Maragos. Spectre: Visual speech-informed perceptual 3d facial expression reconstruction from videos. In CVPRW, pages 5745-5755, 2023. 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "text", + "content": "[14] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 242, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 555, + 296 + ], + "type": "text", + "content": "[15] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In The Eleventh International Conference on Learning Representations, 2023. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 331 + ], + "type": "text", + "content": "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: a reference-free evaluation metric for image captioning. In EMNLP, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 363 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 399 + ], + "type": "text", + "content": "[18] Jessica E Huber and Bharath Chandrasekaran. Effects of increasing sound pressure level on lip and jaw movement parameters and consistency in young adults. 2006. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 400, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 432 + ], + "type": "text", + "content": "[19] Eamonn J. Keogh and M. Pazzani. Derivative dynamic time warping. In In First SIAM International Conference on Data Mining, 2001. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "text", + "content": "[20] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 468, + 555, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 555, + 511 + ], + "type": "text", + "content": "[21] Ziqiao Peng, Yihao Luo, Yue Shi, Hao Xu, Xiangyu Zhu, Hongyan Liu, Jun He, and Zhaoxin Fan. Selftalk: A self-supervised commutative training diagram to comprehend 3d talking faces. In ACM MM, 2023. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 514, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 514, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 514, + 555, + 567 + ], + "type": "text", + "content": "[22] Ziqiao Peng, Haoyu Wu, Zhenbo Song, Hao Xu, Xiangyu Zhu, Jun He, Hongyan Liu, and Zhaoxin Fan. Emotalk: Speech-driven emotional disentanglement for 3d face animation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 569, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 555, + 623 + ], + "type": "text", + "content": "[23] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, page 484-492. Association for Computing Machinery, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 624, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 555, + 689 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. on Mach. Learn., pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[25] Byron Reeves and David Voelker. Effects of audio-video asynchrony on viewer's memory, evaluation of content and detec" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21073" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "text", + "content": "tion ability. In Research Report, Stanford University, 1993. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[26] George Retsinas, Panagiotis P Filntisis, Radek Danecek, Victoria F Abrevaya, Anastasios Roussos, Timo Bolkart, and Petros Maragos. 3d facial expressions through analysis-by-neural-synthesis. In CVPR, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 296, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 296, + 186 + ], + "type": "text", + "content": "[27] Alexander Richard, Michael Zollhöfer, Yandong Wen, Fernando De la Torre, and Yaser Sheikh. Meshtalk: 3d face animation from speech using cross-modality disentanglement. In ICCV, 2021. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 188, + 295, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 295, + 220 + ], + "type": "text", + "content": "[28] Richard Schulman. Articulatory dynamics of loud and normal speech. The Journal of the Acoustical Society of America, 85 (1):295-312, 1989. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 223, + 295, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 295, + 256 + ], + "type": "text", + "content": "[29] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 258, + 295, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 295, + 291 + ], + "type": "text", + "content": "[30] Bowen Shi, Wei-Ning Hsu, and Abdelrahman Mohamed. Robust self-supervised audio-visual speech recognition. arXiv preprint arXiv:2201.01763, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 293, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 295, + 315 + ], + "type": "text", + "content": "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "type": "text", + "content": "[32] Wenfeng Song, Xuan Wang, Shi Zheng, Shuai Li, Aimin Hao, and Xia Hou. Talkingstyle: Personalized speech-driven 3d facial animation with style preservation. IEEE Transactions on Visualization and Computer Graphics, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 362, + 295, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 362, + 295, + 416 + ], + "spans": [ + { + "bbox": [ + 56, + 362, + 295, + 416 + ], + "type": "text", + "content": "[33] Stefan Stan, Kazi Injamamul Haque, and Zerrin Yumak. Facediffuser: Speech-driven 3d facial animation synthesis using diffusion. In Proceedings of the 16th ACM SIGGRAPH Conference on Motion, Interaction and Games, pages 1-11, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 418, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 295, + 474 + ], + "type": "text", + "content": "[34] Trinity Suma, Birate Sonia, Kwame Agyemang Baffour, and Oyewole Oyekoya. The effects of avatar voice and facial expression intensity on emotional recognition and user perception. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 475, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 295, + 518 + ], + "type": "text", + "content": "[35] Licai Sun, Zheng Lian, Bin Liu, and Jianhua Tao. Hicmae: Hierarchical contrastive masked autoencoder for self-supervised audio-visual emotion recognition. Information Fusion, 108: 102382, 2024. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 521, + 295, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 295, + 565 + ], + "type": "text", + "content": "[36] Zhiyao Sun, Tian Lv, Sheng Ye, Matthieu Lin, Jenny Sheng, Yu-Hui Wen, Minjing Yu, and Yong-Jin Liu. Diffposetalk: Speech-driven stylistic 3d facial animation and head pose generation via diffusion models. ACM TOG, 43(4), 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 567, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 295, + 621 + ], + "type": "text", + "content": "[37] Kim Sung-Bin, Lee Chae-Yeon, Gihun Son, Oh Hyun-Bin, Janghoon Ju, Suekyeong Nam, and Tae-Hyun Oh. Multitalk: Enhancing 3d talking head generation across languages with multilingual video dataset. In Interspeech 2024, pages 1380-1384, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "text", + "content": "[38] Kim Sung-Bin, Lee Hyun, Da Hye Hong, Suekyeong Nam, Janghoon Ju, and Tae-Hyun Oh. Laughtalk: Expressive 3d talking head generation with laughter. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6404-6413, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "text", + "content": "[39] Hiroki Tanaka, Satoshi Nakamura, et al. The acceptability of virtual characters as social skills trainers: usability study. *JMIR human factors*, 9(1):e35358, 2022. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 475 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 105 + ], + "type": "text", + "content": "[40] Stephen M Tasko and Michael D McClean. Variations in articulatory movement with changes in speech task. 2004. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "type": "text", + "content": "[41] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "type": "text", + "content": "[42] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-cap: Zero-shot image-to-text generation for visual-semantic arithmetic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17918–17928, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "text", + "content": "[43] Argiro Vatakis and Charles Spence. Audiovisual synchrony perception for speech and music assessed using a temporal order judgment task. *Neuroscience letters*, 393(1):40-44, 2006. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 554, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 554, + 284 + ], + "type": "text", + "content": "[44] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In CVPR, pages 14653-14662, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "type": "text", + "content": "[45] Kaisiyuan Wang, Qianyi Wu, Linsen Song, Zhuoqian Yang, Wayne Wu, Chen Qian, Ran He, Yu Qiao, and Chen Change Loy. Mead: A large-scale audio-visual dataset for emotional talking-face generation. In ECCV, 2020. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "type": "text", + "content": "[46] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In CVPR, 2023. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 555, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 555, + 418 + ], + "type": "text", + "content": "[47] Karren Yang, Anurag Ranjan, Rick Chang, Raviteja Vemulapalli, and Oncel Tuzel. Probabilistic speech-driven 3d facial motion synthesis: New benchmarks, methods, and applications. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 421, + 555, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 555, + 475 + ], + "type": "text", + "content": "[48] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In CVPR, pages 8652-8661, 2023. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21074" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/96cc4bcd-798e-44b6-b645-30ba105eaf18_content_list.json b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/96cc4bcd-798e-44b6-b645-30ba105eaf18_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..93480856c13f3978605e2cbb992ab99ecf25c8d9 --- /dev/null +++ b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/96cc4bcd-798e-44b6-b645-30ba105eaf18_content_list.json @@ -0,0 +1,1660 @@ +[ + { + "type": "text", + "text": "Period-LLM: Extending the Periodic Capability of Multimodal Large Language Model", + "text_level": 1, + "bbox": [ + 243, + 130, + 756, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuting Zhang $^{1*}$ , Hao Lu $^{1,2*}$ , Qingyong Hu $^{2}$ , Yin Wang $^{3}$ , Kaishen Yuan $^{1}$ , Xin Liu $^{4}$ , Kaishun Wu $^{1\\ddagger}$ $^{1}$ The Hong Kong University of Science & Technology (Guangzhou),", + "bbox": [ + 114, + 202, + 880, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2The Hong Kong University of Science & Technology, 3Zhejiang University,", + "bbox": [ + 197, + 239, + 800, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Lappeenranta-Lahti University of Technology", + "bbox": [ + 313, + 257, + 683, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yzhang430, hlu585} @connect.hkust-gz.edu.cn, wuks@hkust-gz.edu.cn", + "bbox": [ + 214, + 273, + 785, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "* Equal contribution, ‡ Corresponding author", + "bbox": [ + 318, + 292, + 673, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a9a47b60acf3fb554029b50e781f9180309e7821008cb4b94e099bc7e07ab12.jpg", + "image_caption": [ + "Periodic Tasks", + "Motion Counting" + ], + "image_footnote": [], + "bbox": [ + 127, + 364, + 238, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a24ad281ec5bc2e477d77288337603d0d2aacfb7c243962205e1c775b155a0f7.jpg", + "image_caption": [ + "Traffic Flow" + ], + "image_footnote": [], + "bbox": [ + 240, + 364, + 344, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ea43dcacdd53b0da69920837766f8c6a28da1340f60335648bc25422d6e1c13e.jpg", + "image_caption": [ + "Weather Forecast" + ], + "image_footnote": [], + "bbox": [ + 344, + 364, + 450, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7f10825e9bf79df9af5768f8c7685ff90f8023a6823233f934edeebe96162db.jpg", + "image_caption": [ + "Remote Health Care" + ], + "image_footnote": [], + "bbox": [ + 450, + 364, + 555, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9a314f038d8ce41ec97dcd706564b9c6f2b70557687ee0dc65ce2b2c83f7101a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 316, + 838, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How many times did the person in the video repeat running on the treadmill?", + "bbox": [ + 620, + 358, + 836, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/428f4be7642d557252d21d318953c655a9c869e02cc4801bb4fedfb2cff19827.jpg", + "image_caption": [ + "GPT-4" + ], + "image_footnote": [], + "bbox": [ + 570, + 390, + 607, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The analysis did not detect any clear repetitions in the video based on the motion patterns. Hard to analyse", + "bbox": [ + 632, + 392, + 844, + 430 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/03d370ec24e3d364cb62d7e1765572687dd8dd85b5173f914ca792531aff53f2.jpg", + "image_caption": [ + "Period-LLM", + "Figure 1. Existing multimodal models may fail to analyze periodic tasks, such as motion counting, traffic flow, weather forecasting, and remote health care. For example, GPT-4 fails to detect clear repetitions due to the difficulty in analyzing motion patterns, and VideoLLaMA incorrectly counts the repetitions. In contrast, Period-LLM accurately analyzes the actions and provides correct repetition count." + ], + "image_footnote": [], + "bbox": [ + 570, + 436, + 607, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The person in the video repeated running 5 times. $\\times$ Wrong number", + "bbox": [ + 632, + 441, + 862, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The repetitive motion in the video is running on the treadmill 14 times.", + "bbox": [ + 630, + 476, + 848, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 563, + 326, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Periodic or quasi-periodic phenomena reveal intrinsic characteristics in various natural processes, such as weather patterns, movement behaviors, traffic flows, and biological signals. Given that these phenomena span multiple modalities, the capabilities of Multimodal Large Language Models (MLLMs) offer promising potential to effectively capture and understand their complex nature. However, current MLLMs struggle with periodic tasks due to limitations in: 1) lack of temporal modelling and 2) conflict between short and long periods. This paper introduces Period-LLM, a multimodal large language model designed to enhance the performance of periodic tasks across various modalities, and constructs a benchmark of various difficulty for evaluating the cross-modal periodic capabilities of large models. Specially, We adopt an \"Easy to Hard Generalization\" paradigm, starting with relatively simple text-based tasks and progressing to more complex visual and multimodal tasks, ensuring that the model gradually builds robust periodic reasoning capabilities. Additionally, we propose a \"Resisting Logical Oblivion\" optimization strategy to maintain periodic reasoning abilities dur", + "bbox": [ + 86, + 583, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing semantic alignment. Extensive experiments demonstrate the superiority of the proposed Period-LLM over existing MLLMs in periodic tasks. The code is available at https: //github.com/keke-nice/Period-LLM.", + "bbox": [ + 511, + 565, + 906, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 630, + 645, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Periodic tasks refer to activities performed at fixed or predictable intervals, encompassing a wide range of applications. At a macro level, this includes repetitive motions in human daily life, such as step counts and the number of times one jumps rope [7, 10, 50], as well as the periodic analysis of meteorological data, where scientists can identify various weather patterns and trends, forecast future weather conditions [1, 35, 37], and issue timely warnings. At a micro level, in health monitoring, by capturing and analyzing subtle chromatic changes in facial videos, physiological signals such as heart rate, blood pressure, and respiratory rate can be extracted [22, 25, 45, 51]. The periodic fluctuations of these signals reflect the physiological state of human. Effectively handling periodic tasks can significantly enhance various aspects of human life, from predicting weather patterns to monitoring health conditions.", + "bbox": [ + 509, + 657, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29237", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given that periodic phenomena occur across various modalities, it becomes essential to utilize multimodal models for period analysis instead of single modality. In recent years, multimodal large language models [3, 16, 18, 48] have advanced the progress of visual and linguistic learning by incorporating visual encoders into various pre-trained large language models. Furthermore, some studies have extended vision to other modalities such as audio [9, 48], depth imagery [8], and near-infrared [8]. With the promising progress of LLM-enabled multimodal models, a natural question arises: Can multimodal large language models understand periodic tasks?", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unfortunately, existing large models have shown inadequacies in counting capabilities or understanding \"repetitive\" temporal sequences [15], creating obstacles for the application of periodic tasks in multimodal large language models, as shown in Fig. 1. These shortcomings mainly stem from the following challenges: 1) Interference of spatial pseudo-temporal information. The leakage of numerical information in visuals may lead models to take shortcuts rather than learning the true periodic information. When specific numbers appear in videos, models may tend to directly answer with these numbers, or be misled by them, rather than deriving answers through the analysis of periodic information [32]. This limits the models' generalization capabilities, rendering them unable to handle periodic tasks without direct numerical information. 2) Conflict between long-time periodic reasoning ability and short-time semantic alignment. The recognition of a periodic task lies not only in the ability to deduce the number of periodic repeats, but also in the accurate recognition of the semantic information in the video. Both require the model to have the ability to model the timing, but with different focus. Too much focus on the short-time semantic understanding may lead to ignorance of periodic reasoning ability. 3) Absence of captions is a significant issue. In the training corpora of MLLMs, most sentences are descriptive and rarely contain specific descriptions of the number of actions. This poses difficulties for models when dealing with periodic tasks that require counting the number of actions. For example, a model might respond with \"The boy in the video is doing pull-ups\" or \"The boy in the video does multiple pull-ups\", without a specific counting supervision.", + "bbox": [ + 91, + 276, + 483, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To avoid overfitting to modality-specific features while preserving temporal modeling capabilities, we propose a MLLM with periodic capabilities, termed Period-LLM. Specifically, we employ an easy-to-hard generalization paradigm. Initially, we define text-based, macro, and micro periodic tasks, ranging from easy to hard. We anticipate starting with text and progressively advancing, enabling Large Language Models (LLMs) to learn periodicity across different modalities. The rationale for this definition is twofold: 1) Despite modality alignment, LLMs still ex", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "hibit superior text processing capabilities compared to other modalities; 2) Macro periodic tasks contain clear semantic features, making their periodic nature easier to capture than that of micro tasks. Accordingly, we selected periodic tasks corresponding to different levels of difficulty. At the textual level, we construct a \"repeated word question-answer\" dataset to pre-train LLMs, deepening their understanding of \"repetitiveness\" from a textual perspective. At the macro level, we select the repetitive motion dataset Countix [7], which aims to predict the number of repetitive movements in videos. At the micro level, we choose the rPPG task, which involves capturing subtle periodic or quasi-periodic physiological signals from facial videos or radio-frequency signals. Furthermore, it is believed that a good periodic model should have both the ability of periodic reasoning and the ability of semantic understanding. Focusing on the optimization of semantic alignment may, to some extent, lead to forgetting the logical reasoning abilities learned from simple tasks. Therefore, we propose a Resisting Logical Oblivion optimization strategy. Specifically, by manually designing a method to dynamically update the backpropagation gradients, the model can adjust previously overlooked feature dimensions without interfering with existing ones, providing a guided approach to model optimization. A series of experiments demonstrate the superiority of the proposed Period-LLM over existing MLLMs in periodic tasks. This advantage is not limited to text and video; we further explore the potential of other modalities based on this foundation.", + "bbox": [ + 511, + 90, + 906, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 532, + 529, + 836, + 542 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we propose the first MLLM-based periodic model (Period-LLM) and construct a benchmark featuring periodic tasks of varying difficulty, including text, video, and other modalities.", + "- We propose an easy-to-hard training paradigm that progresses from simple to complex tasks, enabling the model to sequentially acquire the ability to perceive periodic information across different modalities, starting with text, then vision, and extending to other modalities.", + "- We propose a Resisting Logical Oblivion (RLO) optimization strategy to resolve the conflict between periodic reasoning and semantic alignment.", + "- Extensive experiments demonstrate that the proposed Period-LLM outperforms existing MLLMs in periodic tasks across various modalities." + ], + "bbox": [ + 511, + 547, + 903, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 777, + 653, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multimodal Large Language Model", + "text_level": 1, + "bbox": [ + 511, + 803, + 823, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent studies, large models have exhibited reliable capabilities in the field of video understanding. In addition to the powerful vision-language large models [12, 17, 26, 52], there is an increasing focus on exploring more modalities in recent research [13, 27-29, 46]. Bain et al. [5] intro", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29238", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1eb310121921262fee121006cd41059657bf8ffc434340acfbb717fb61019be5.jpg", + "image_caption": [ + "Figure 2. The process of the generation of Question-and-answer datasets. Samples of text and video are taken as examples. In the text sample, questions and the occurrences of specific words are randomly generated, while in the video sample, the questions are randomly generated and the keywords represent labels within the dataset. These elements are then incorporated into a GPT-4 prompt to construct answers that include key information. The right side of the image shows examples of the final Q&A pairs generated, where the answers are complete sentences containing the critical information from the key elements." + ], + "image_footnote": [], + "bbox": [ + 119, + 116, + 877, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "duce a large-scale dataset, which provides general descriptions associated with video content. Several LLM-based works [13, 27, 29, 46] are proposed to comprehend the visual content of the video. Besides, Video-LLaMa [47] extends the ability to comprehend both auditory and visual information, while Su et al. [41] leverages the power of multimodal encoders to understand across six modalities. Recently, Lv et al. [28] proposed video-based large language models for the task of VAD in a weakly supervised framework. Different from them, our work focuses on the periodic analysis of MLLMs.", + "bbox": [ + 89, + 352, + 483, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Periodic Tasks", + "text_level": 1, + "bbox": [ + 89, + 527, + 238, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Examples of periodic learning include recovering and magnifying physiological signals (e.g., heart rate or breathing) [25, 31], predicting weather and environmental changes (e.g., nowcasting of precipitation or land surface temperatures) [38], counting motions that are repetitious (e.g., exercises or therapies) [4, 7], and analyzing human behavior (e.g., gait) [23]. Prior work has focused on designing customized neural architectures [7, 20], loss functions [39], and leveraging relevant learning paradigms including transfer learning [24] and meta-learning [21] for periodic learning. This paper attempts to study the capability of periodic tasks in multimodal models.", + "bbox": [ + 89, + 550, + 483, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Setting and Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 739, + 405, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In nature, periodic phenomena contain both periodic and semantic information, which can be modeled by $x = K \\cdot p(\\omega t) + N \\cdot s(t)$ , where $p(\\omega t)$ is a periodic function with angular frequency $\\omega = 2\\pi / T$ and period $T$ . Here, $K$ represents the amplitude of the periodic component, $N$ denotes the amplitude of the semantic component, and $s(t)$ is a semantic function related to time $t$ . When discretizing continuous sequences as model inputs $X$ , we adopt a sampling interval $\\Delta t = T / n$ , where $n$ is a positive integer multiple", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of the sampling rate $(n = 1,2,3,\\ldots)$ . Ideally, when $n = 1$ and semantics are absent, the value at each sampling point equals the constant $K$ , resulting in $x = K$ . This can be mathematically expressed as:", + "bbox": [ + 511, + 352, + 906, + 412 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nX = \\left\\{ \\begin{array}{l l} K \\cdot p (\\omega t) + N \\cdot s (t), & \\Delta t = \\frac {T}{n}, n = 2, 3, 4, \\dots \\\\ K, & \\Delta t = T \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 421, + 905, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, in practical applications, data is often high-dimensional and noisy. For instance, after processing a text segment through a tokenizer, we obtain a matrix with dimensions $[L, \\mathrm{Dim}]$ , where $L$ represents the length of the text (i.e., the number of tokens), and Dim denotes the embedding dimension of each token. In cases of repetitive text such as \"period period period\", certain dimensions of the encoded feature matrix may exhibit constant values similar to $x = K$ , where $K$ is a vector representing the embedding values of these tokens in specific dimensions.", + "bbox": [ + 511, + 474, + 906, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inputs from other modalities, such as video, present more complex scenarios. Periodic movements may be obscured by a large amount of semantic or redundant information. In macro-periodic tasks, periodic movement is more prominent, allowing models to detect periodicity more easily; however, in micro-periodic tasks, the amplitude of the periodic signal $(K)$ may be small and masked by noise, making it challenging for models to capture these subtle periodic variations. Despite these complexities, research suggests that models can transfer knowledge from simpler tasks to more complex ones. In this context, repetitive text can be seen as a basic periodic task, while macro-periodic and micro-periodic video tasks represent more complex periodic challenges. Although \"repetition\" is the common thread across these tasks, the semantic content and amplitude of the periodic signal vary.", + "bbox": [ + 511, + 627, + 906, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given that large language models (LLMs) have been trained on extensive and diverse text corpora, they are par", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29239", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/291a3997c6228d68dc03a276bd5fe6ca42178f63429407d9785edc71208665d8.jpg", + "image_caption": [ + "Figure 3. The pipeline of training paradigm from easy to hard. The model begins training with simpler periodic tasks and gradually transitions to more complex ones. Throughout this process, Period-LLM's understanding of periodic tasks progressively strengthens and becomes more profound." + ], + "image_footnote": [], + "bbox": [ + 155, + 114, + 834, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ticularly well-suited for processing text tasks. We propose the following framework:", + "bbox": [ + 89, + 272, + 482, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA = F (X, Q) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 303, + 480, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This framework accommodates inputs $X$ from various modalities, where $X$ represents periodic data from any modality, $Q$ represents a given question, and the model computes the answer $A$ . We suggest beginning model training with simple repetitive text tasks and gradually progressing to macro-periodic and micro-periodic video tasks. This progressive training approach helps the model develop a comprehensive understanding of periodic phenomena, moving from simple to complex tasks, thereby enhancing its ability to handle periodic challenges effectively.", + "bbox": [ + 89, + 333, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 497, + 181, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we aim to enhance the understanding of large models regarding periodic tasks. Given the current absence of specialized question-and-answer datasets for periodic tasks and corresponding testing standards, we will undertake pioneering work in the content that follows. Specifically, we firstly elucidate the instruction generation for periodic tasks. Then, we explore the generalization process from simpler tasks to more complex ones. Additionally, we propose an optimization strategy to address the challenge of preserving periodic reasoning abilities during the semantic alignment phase of Multi-Modal Generalization.", + "bbox": [ + 89, + 522, + 483, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Instruction Generation for Periodic Tasks", + "text_level": 1, + "bbox": [ + 89, + 693, + 447, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. Text-Only Instructions", + "text_level": 1, + "bbox": [ + 89, + 715, + 294, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance our dataset's diversity, we use GPT-4 to help create a repetitive text dataset through several key steps: First, we define the core question: \"How many times is the word $\\{my\\text{word}\\}$ repeated in the string $\\{my\\text{string}\\} ?$ We then use GPT-4 to generate ten semantically similar questions, $q_{i}\\in \\{q_{1},q_{2},\\ldots ,q_{10}\\}$ . For word diversity, we randomly select a word from the GPT-4 technical report [2] as $\\{my\\text{word}\\}$ . To further diversify, we set a repetition count $n$ within $\\{2,3,\\dots,20\\}$ , constructing $\\{my\\text{string}\\}$ as \" $\\{\\mathrm{selected~word}\\} \\times n$ . Since GPT-4 can't directly calculate word repetitions, we guide it with precise prompts. We", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "instruct GPT-4: \"Construct an answer based on the question format, selected word, and repetition count.\" During this process, we randomly choose a question $q_{i}$ and provide the necessary answer elements, {selected word, $n$ times}, enabling GPT-4 to generate a fitting response. By following these steps, we effectively use GPT-4's generative capabilities to build a diverse and accurate repetitive text dataset, which supports robust model training and evaluation.", + "bbox": [ + 511, + 272, + 906, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.2. Multi-Modal Instructions", + "text_level": 1, + "bbox": [ + 511, + 405, + 736, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For tasks involving periodic data across various modalities like video, traffic streams, and health monitoring, we employ a method that integrates annotations, raw captions, and frequency information to generate comprehensive representations. This approach is applicable to diverse datasets, addressing challenges such as dynamic environments, variations in periodicity and counts, and action speeds. We transform data into a question-and-answer format inspired by repetitive text dataset construction. We generate unified captions for each segment by combining existing annotations and frequency data. The main question posed is: \"What is the total number of repetitive actions in this data?\" Using GPT-4's capabilities, we craft multiple semantically similar questions from this core inquiry. Explicit instructions and answer elements are provided to generate question-and-answer pairs, enhancing dataset representation, the process as shown in Fig. 2. This method enriches model understanding and processing of periodic tasks across modalities. We establish a link between repetitive text and periodic data, enabling generalization from simple to complex tasks. The process first involves self-optimization in relatively simple individual tasks, followed by fine-tuning on more complex tasks, as illustrated in Fig. 3.", + "bbox": [ + 511, + 425, + 906, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Easy-to-Hard Generalization", + "text_level": 1, + "bbox": [ + 511, + 801, + 772, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we explore the generalization process from simpler tasks to more complex ones, structured around the problem definition $A = F(f, Q)$ , where $A$ represents the answer, $f$ represents the features derived from the input data, and $Q$ represents the question.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29240", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/467885a30330a0b59d0a855988a9252b72c8f8554f1a2bb1673d47b5cbb68811.jpg", + "image_caption": [ + "Figure 4. The Period-LLM framework, using video as an example, integrates both text and video inputs. The video is processed through a visual encoder and a visual projector, then concatenated with text features derived from a word embedding layer. This combined input is fed into a large language model (LLM). Gradient optimization strategies are applied to the output feature channels of the LLM, specifically targeting those channels that failed to capture useful information, thereby enriching the semantic content in these channels." + ], + "image_footnote": [], + "bbox": [ + 133, + 90, + 861, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Text-Only Generalization", + "text_level": 1, + "bbox": [ + 89, + 372, + 313, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first stage of generalization focuses on text-based tasks, where the goal is to learn logical reasoning without needing to align other modalities with text. This can be mathematically represented as:", + "bbox": [ + 89, + 390, + 482, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA = F \\left(T _ {f}, Q\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 449, + 480, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "here, $T_{f}$ represents the features derived from textual input. The emphasis is on developing the model's logical reasoning capabilities, as the input modality is solely text, which simplifies the task by eliminating the need to handle multimodal alignment. The model learns to process and understand the logical structure of questions and generate appropriate answers purely based on textual information.", + "bbox": [ + 89, + 472, + 483, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2. Multi-Modal Generalization", + "text_level": 1, + "bbox": [ + 89, + 585, + 334, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The next stage involves more complex tasks that require the integration of multiple modalities, such as video and text. This stage not only demands logical reasoning but also the alignment of semantic information across different modalities. This process can be described by:", + "bbox": [ + 89, + 604, + 483, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA = F \\left(M _ {f}, Q\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 679, + 480, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in this case, $M_{f}$ represents features derived from a combination of inputs, including video and text. The challenge is twofold: the model must align semantic content across different modalities while simultaneously developing robust periodic logical reasoning capabilities. This complex task requires the model to integrate semantic features with textual information and establish accurate connections between them in order to answer questions correctly.", + "bbox": [ + 89, + 704, + 483, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Resisting Logical Oblivion", + "text_level": 1, + "bbox": [ + 89, + 833, + 330, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the significant differences between periodic logical reasoning abilities and the ability to understand semantic information in videos, we design a Resisting Logical Oblivion", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(RLO) optimization strategy to prevent the new knowledge learned in a new round of optimization from overwriting the reasoning knowledge acquired from simpler tasks. Vanilla gradient descent leads to an undifferentiated update of the entire feature space. If $\\theta = (\\theta_{1},\\theta_{2},\\dots,\\theta_{J})\\in \\mathbb{R}^{C\\times W}$ is the parameter of the model, where $W$ is related to the network structure, $\\frac{\\partial L(\\theta)}{\\partial\\theta_j}$ $(j = 1,2,\\ldots ,J)$ represents the partial derivative of $\\theta$ . The process of model updating is as follows:", + "bbox": [ + 511, + 372, + 906, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\theta_ {j} = - \\alpha \\frac {\\partial L (\\theta)}{\\partial \\theta_ {j}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 508, + 903, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {j} ^ {\\prime} \\leftarrow \\theta_ {j} + \\nabla \\theta_ {j}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 544, + 903, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is the learning rate, $\\nabla \\theta_{j}$ is the updating vector, and $\\theta_j^\\prime$ is the new weight for the next iteration.", + "bbox": [ + 511, + 574, + 905, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traditional gradient update strategies adjust model parameters based on global errors, which, although effective, can lead to knowledge interference between different types of tasks. Our new strategy introduces a feature channel weight allocation mechanism, dynamically adjusting the gradient update weights of different feature channels. Specifically, we introduce the weight function $\\Omega (k,c)$ implementation:", + "bbox": [ + 511, + 606, + 905, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega \\left(c _ {i}\\right) = \\left\\{ \\begin{array}{l l} 1 + \\beta \\cdot e ^ {\\frac {\\text {i t e r n u m}}{\\text {m a x i t e r}}}, & \\bar {c} _ {i} < \\bar {c}, \\\\ 1, & \\bar {c} _ {i} > \\bar {c}, \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 726, + 903, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is a hyper-parameter, $iter_{num}$ denotes the current iteration count, and $max_{iter}$ represents the maximum number of iterations. $\\bar{c}$ is the average across all feature channels, while $\\bar{c}_i$ refers to the average of the $i$ th feature channel. By dynamically weighting feature channels, we can exert control over the real-time updates of specific feature channels. This ensures that redundant feature channels can learn newly acquired semantic information without", + "bbox": [ + 511, + 780, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29241", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e431d41866c94feb883cb1d08d332b2731d099067594816b04e1c42d30587032.jpg", + "image_caption": [ + "Figure 5. Details of the function of the gradient weight function. The left side of the image shows that some channels have lower backpropagation gradients, while others have higher ones. Higher backpropagation gradients indicate that these feature channels are being updated more significantly." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 472, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "compromising existing reasoning abilities, thereby balancing the enhancement of reasoning abilities and video semantic understanding. As shown in Fig. 5, our gradient weighting function is applied to the backpropagation gradients, boosting those originally lower gradients, thereby substantially increasing the update of channels that were previously updated little or not at all. This process ensures that newly learned semantic information is directed to these newly emphasized feature channels for learning. Therefore, the model parameter update process with the introduction of the RLO optimization strategy is as follows, showing the changes from Eq. 5 to Eq. 8:", + "bbox": [ + 89, + 279, + 483, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\theta_ {j} ^ {*} = \\Omega \\left(c _ {i}\\right) \\cdot \\nabla \\theta_ {j}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 463, + 480, + 479 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {j} ^ {\\prime} \\leftarrow \\theta_ {j} + \\nabla \\theta_ {j} ^ {*}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 483, + 480, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the training process, the input data $\\mathcal{Z}$ consists of input videos and their corresponding questions $x_{i}$ , correct answers $y_{i}$ , where $i = 1,2,3,\\ldots ,K$ , and each $x_{i},y_{i}$ is a sequence of tokens. Large language models are denoted by $P_{\\phi}(y\\mid x)$ , where $\\phi$ represents model parameters. Typically, when fine-tuning large models, only the question-answer pairs $(x_{i},y_{i})$ are input. The goal is to adjust the parameters $\\phi$ to maximize the following expression:", + "bbox": [ + 89, + 512, + 483, + 633 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\phi} \\sum_ {(x, y) \\in \\mathcal {Z}} \\sum_ {t = 1} ^ {| y |} \\log \\left(P _ {\\phi} \\left(y _ {t} \\mid x, y _ {< t}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 638, + 482, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $|y|$ represents the length of the input vector $y$ , $P_{\\phi}(y_t|x, y_{MethodLLMZero-shotFine-tuneCountix-QArPPG-QACountix-QArPPG-QAMAERMSEBleu1CIDErMETEORMAERMSEMAERMSEBleu1CIDErMETEORMAERMSEFrozenBiLM [44]DeBERTa-V27.2211.530.1450.5500.11220.4321.565.229.440.2330.6100.20119.2120.46VideoLLaMA [48]Vicuna-7B6.1810.820.1870.3600.24518.8919.434.988.890.3890.5700.35818.2919.01LLaMA-Adapter [49]LLaMA-7B7.4410.510.2440.6830.31221.7823.345.7810.020.3780.7890.39720.5422.84VideoChat [14]Vicuna-7B6.7311.230.1650.5490.27719.5420.675.039.530.2750.6790.38818.3210.77Video-ChatGPT [30]Vicuna-7B6.4410.750.2030.5890.29718.1220.454.649.150.3560.6430.36717.5419.45BT-Adapter [19]Vicuna-7B7.7311.700.2760.6740.34119.3722.686.129.700.2960.6890.35318.6721.49LLaMA-VID [16]Vicuna-7B6.9711.210.2670.6800.30118.3118.565.349.850.3880.7830.41117.5117.60Period-LLMLLaMA-7B5.518.330.2880.7020.30815.0817.133.777.140.3970.8100.41513.7816.83", + "bbox": [ + 94, + 88, + 903, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3964fe6c76cffa056001065f27e32cab3bece6b76da324ed28f907f3562e3813.jpg", + "table_caption": [ + "Table 1. Comparison of our method with other leading methods on two newly introduced video question-and-answer (QA) datasets, with all methods evaluated at a resolution of ${224} \\times {224}$ . The results for Period-LLM in the zero-shot setting are derived from a model pretrained on text-QA dataset. For rPPG-QA, the V4V dataset is used as the validation set." + ], + "table_footnote": [], + "table_body": "
MethodLLMZero-shotFine-tune
RotNIST-QADrive-QARadar-QARotNIST-QADrive-QARadar-QA
MAERMSEMAERMSEMAERMSEMAERMSEMAERMSEMAERMSE
FrozenBiLM [44]DeBERTa-V24.264.7830.88120.5622.1730.422.883.2529.44110.5616.6628.55
VideoLLaMA [48]Vicuna-7B4.085.3235.45130.6723.4218.852.513.8931.22121.5517.0320.34
LLaMA-Adapter [49]LLaMA-7B3.954.4231.26125.4224.1231.792.663.2330.28118.7920.3428.56
VideoChat [14]Vicuna-7B3.564.2132.67126.8821.7129.582.233.0630.12115.2320.5427.38
Video-ChatGPT [30]Vicuna-7B3.444.8738.18118.2923.7732.132.012.2933.28112.3521.6130.45
BT-Adapter [19]Vicuna-7B4.035.4634.21110.2423.4231.472.783.5631.31107.6522.1230.19
LLaMA-VID [16]Vicuna-7B3.874.5734.98114.3318.2129.322.433.2532.45109.8718.2129.32
Period-LLMLLaMA-7B3.203.9630.88108.7417.5626.451.502.7528.71105.5414.2424.12
", + "bbox": [ + 133, + 277, + 859, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of our method with other leading methods on the RotNIST-QA, Drive-QA, and Radar-QA datasets. Drive-QA is a dataset of taxi trajectory data, while Radar-QA is a radar modality dataset capturing physiological signals. The zero-shot results for Period-LLM are based on a model pretrained on text-QA, Countix-QA, and rPPG-QA.", + "bbox": [ + 89, + 424, + 903, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "periodic reasoning abilities. The results show significant improvements over zero-shot, thanks to the inclusion of explicit periodic information in the training data. Period-LLM surpasses existing MLLMs by over 1 MAE on the Countix-QA dataset, while also showing significant improvements in generated text quality, as measured by Bleu-1, CIDEr, and METEOR scores. This improvement is driven by the easy-to-hard fine-tuning paradigm and the Resisting-Logical-Oblivion strategy, which together help the model preserve and enhance its periodic reasoning abilities, while also maintaining strong semantic understanding, resulting in superior performance.", + "bbox": [ + 88, + 470, + 482, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the visual periodic tasks, Table 2 presents the performance of these models on other modalities, specifically the RotNIST-QA, Drive-QA, and Radar-QA datasets. The zero-shot and fine-tune protocols follow the same approach as in Table 1, with the key difference being that experiments on these three datasets were conducted in parallel. The pretraining models used in these experiments were fine-tuned on the two visual periodic tasks from Table 1. On the RotNIST-QA dataset, Period-LLM achieves the lowest MAE (1.50) and RMSE (2.75), outperforming models such as FrozenBiLM, VideoLLaMA, and LLaMA-Adapter. For the Drive-QA dataset, Period-LLM demonstrates superior performance with the lowest MAE (28.71) and competitive RMSE (105.54). On the Radar-QA dataset, Period-LLM also performs strongly, further validating the effectiveness of our model across different modali", + "bbox": [ + 88, + 657, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ties. These results underscore Period-LLM's enhanced accuracy in periodic reasoning tasks and its robustness in handling diverse datasets, establishing it as a versatile solution for periodic question-answering across a range of domains, including health, weather, and traffic.", + "bbox": [ + 511, + 470, + 906, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 560, + 694, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "QA Examples. Fig. 6 showcases examples from the countix-QA and rPPG-QA datasets. It can be seen that Period-LLM is capable of understanding the semantic information contained within the video as well as recognizing periodicity within the video. For instance, in the countix-QA dataset, Period-LLM accurately provides the number of repetitive movements, while in the rPPG-QA dataset, it can determine the heart rate of the man in the video. These capabilities are beyond what other language models can achieve. According to Fig. 6, in the countix-QA dataset, Period-LLM identifies that the man in the video performs two pull-ups. In the rPPG-QA dataset, Period-LLM is able to detect that the man's heart rate is 75 beats per minute. This demonstrates Period-LLM's exceptional performance in handling video content and periodic information. More QA examples are provided in the appendix.", + "bbox": [ + 511, + 584, + 903, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of Loss Reduction. The Fig. 7 illustrates the loss reduction comparison between the Baseline (Vanilla Fine-Tuning on Countix-QA) and the proposed Period-LLM approach (with text-QA pretraining and RLO) on the Countix-QA dataset. As shown, our pro", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29243", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. The MAE value on Countix-QA with different parameter or setting choices in $\\Omega \\left( {c}_{i}\\right)$ of RLO optimization strategy, and ablation study on Countix-QA showing the impact of easy tasks (period-text-QA) and RLO (Resisting Logical Oblivion). Default settings are marked in gray .", + "bbox": [ + 89, + 88, + 906, + 132 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c50fc6abd5e23ef3c4c7d477248606193620c30ed3ced1707f75f767c78d4542.jpg", + "table_caption": [ + "(a) Impact of $\\beta$" + ], + "table_footnote": [], + "table_body": "
βMAE
0.013.85
0.053.77
0.13.90
0.54.05
", + "bbox": [ + 117, + 160, + 202, + 214 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8aa98fdd75ae35db08f954c76661fc0076c53e102278e81f978d1f21c8f7b9bf.jpg", + "table_caption": [ + "(b) Type of Threshold." + ], + "table_footnote": [], + "table_body": "
ThresholdMAE
Mean3.77
Median3.85
Learnable4.20
", + "bbox": [ + 236, + 160, + 339, + 207 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7c60bc2be8effcead6c4e502997e393c91dbda86fae73fddb4b56fbfb535b9fe.jpg", + "table_caption": [ + "(c) Results on Countix-QA w./w.o Easy Task and RLO." + ], + "table_footnote": [], + "table_body": "
Easy to HardRLOCountix-QA
Text-QACountix-QAMAECIDEr
✗✓✗4.300.661
✓✓✗3.890.782
✓✓✓3.770.810
", + "bbox": [ + 357, + 160, + 602, + 220 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6d089f81b8f085344037ec239700f7b61adc7b98d5d7a2a57e7cc2dbb5d1f729.jpg", + "table_caption": [ + "(d) Results on rPPG-QA w./w.o Easy Tasks and RLO." + ], + "table_footnote": [], + "table_body": "
Easy to HardRLOrPPG-QA
Text-QACountix-QArPPG-QAMAE
XX✓X15.45
✓X✓X14.46
✓✓✓X14.05
✓✓✓✓13.78
", + "bbox": [ + 614, + 160, + 895, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bad20fc015a3ae26c5ecfa6dd9ea548082722e2a8b53c41bc4c4ee52faf51f09.jpg", + "image_caption": [ + "Figure 6. Examples from Countix-QA and rPPG-QA." + ], + "image_footnote": [], + "bbox": [ + 135, + 222, + 434, + 334 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5778448ea4e77748d3c7a1285ccf9f5e5ebf498cce1f6de4d92976a53225d743.jpg", + "image_caption": [ + "Figure 7. Comparison of Loss Reduction Between Baseline (Vanilla Fine-Tuning on Countix-QA) and Proposed Method (Period-LLM with Text-QA Pretraining and RLO, Fine-Tuned on Countix-QA)." + ], + "image_footnote": [], + "bbox": [ + 127, + 354, + 439, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "posed method demonstrates a significantly faster decline in the loss values compared to the baseline. The Period-LLM model achieves a steeper descent in the initial training stages, indicating a more efficient learning process. Furthermore, it converges to a lower loss value in fewer epochs, suggesting improved optimization and stability. This faster convergence implies that our method better captures the underlying data patterns, leading to more effective fine-tuning on the Countix-QA.", + "bbox": [ + 89, + 527, + 483, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 666, + 243, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Setting of Weight Function $\\Omega(c_i)$ of RLO Optimization Strategy. As seen in Table 3a, $\\beta = 0.05$ yields the lowest MAE (3.77), indicating an effective balance between retaining prior knowledge and integrating new information. When $\\beta$ is too small (e.g., 0.01), the model's anti-forgetting ability is weak, causing higher error as new knowledge overwhelms older knowledge. Conversely, a high $\\beta$ (e.g., 0.5) overly prioritizes old knowledge, limiting adaptation. Thus, $\\beta = 0.05$ achieves the best balance for minimizing error. In Table 3b, setting the threshold to the mean results in the lowest MAE (3.77), effectively balancing old and new knowledge contributions. The median threshold performs slightly worse (MAE 3.85) due to its insensitivity to outliers, while a learnable threshold increases MAE", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(4.20), likely due to instability. Hence, the mean threshold is the most stable and effective choice.", + "bbox": [ + 511, + 247, + 906, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results of various combinations of Preiod-LLM Components. We evaluate our approach on the Countix-QA and rPPG-QA datasets, as shown in Tables 3c and 3d. For Countix-QA, our method—combining pre-training on simpler tasks with the Resisting Logical Oblivion (RLO) strategy—significantly improves performance, reducing the Mean Absolute Error (MAE) from 4.30 to 3.77 and increasing the CIDEr score from 0.661 to 0.810. Similarly, on the rPPG-QA dataset, this approach lowers the MAE from 15.45 to 13.78. These improvements reflect the model's enhanced capacity to capture periodic patterns effectively while also understanding semantic content more deeply. The decrease in MAE shows better accuracy in numerical predictions, while the CIDEr increase suggests improved semantic alignment in question-answering tasks. Together, these results confirm that the integration of easy-task pretraining with RLO enables the model to balance periodicity and semantic understanding, leading to faster convergence and more robust performance across QA tasks.", + "bbox": [ + 511, + 281, + 906, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 583, + 633, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce Period-LLM, a large multimodal model specifically designed to enhance performance on periodic tasks across various modalities. The model follows a progressive training paradigm, starting with simpler text-based tasks and advancing to more complex visual and multimodal tasks. Additionally, we propose the \"Resisting Logical Oblivion\" optimization strategy to preserve the model's periodic reasoning abilities during semantic alignment. The Our experimental results demonstrate that Period-LLM significantly outperforms existing models in handling periodic tasks, especially in challenging domains such as repetitive motion analysis and physiological signal estimation. Period-LLM not only achieves superior accuracy but also highlights its potential for future advancements in managing complex periodic phenomena across different modalities. Future work will focus on further refining model's capabilities, particularly in addressing more nuanced and less structured periodic tasks, and expanding its application to additional domains where periodicity plays a crucial role.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29244", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Guangdong Provincial Key Laboratory of Integrated Communication, Sensing and Computation for Ubiquitous Internet of Things (No. 2023B1212010007), China NSFC Grant 62472366, the Project of DEGP (No.2024GCZX003, 2023KCXTD042), 111 Center (No.D25008) and Shenzhen Science and Technology Foundation (ZDSYS20190902092853047).", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 233, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kumar Abhishek, Maheshwari Prasad Singh, Saswata Ghosh, and Abhishek Anand. Weather forecasting model using artificial neural network. Procedia Technology, 4:311-318, 2012. 1", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4", + "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2", + "[4] Mohammad Rafayet Ali, Javier Hernandez, E Ray Dorsey, Ehsan Hoque, and Daniel McDuff. Spatio-temporal attention and magnification for classification of parkinson's disease from videos collected via the internet. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 207-214. IEEE, 2020. 3", + "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2", + "[6] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 6", + "[7] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Counting out time: Class agnostic video repetition counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10387-10396, 2020. 1, 2, 3, 6", + "[8] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 2", + "[9] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xiangyu Yue. Onellm: One framework to align all modalities with language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26584-26595, 2024. 2" + ], + "bbox": [ + 99, + 258, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Huazhang Hu, Sixun Dong, Yiqun Zhao, Dongze Lian, Zhengxin Li, and Shenghua Gao. Transrac: Encoding multiscale temporal correlation with transformers for repetitive action counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19013-19022, 2022. 1", + "[11] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6", + "[12] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2", + "[13] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 3", + "[14] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6, 7", + "[15] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2", + "[16] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 2, 6, 7", + "[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 6", + "[19] Ruyang Liu, Chen Li, Yixiao Ge, Ying Shan, Thomas H Li, and Ge Li. One for all: Video conversation is feasible without video instruction tuning. arXiv preprint arXiv:2309.15785, 2023. 6, 7", + "[20] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems, 33:19400-19411, 2020. 3", + "[21] Xin Liu, Ziheng Jiang, Josh Fromm, Xuhai Xu, Shwetak Patel, and Daniel McDuff. Metaphys: few-shot adaptation for non-contact physiological measurement. In Proceedings of the conference on health, inference, and learning, pages 154-163, 2021. 3", + "[22] Xin Liu, Yuting Zhang, Zitong Yu, Hao Lu, Huanjing Yue, and Jingyu Yang. rppg-mae: Self-supervised pretraining with masked autoencoders for remote physiological measurements. IEEE Transactions on Multimedia, 2024. 1", + "[23] Yingcheng Liu, Guo Zhang, Christopher G Tarolli, Rumen Hristov, Stella Jensen-Roberts, Emma M Waddell, Taylor L Myers, Meghan E Pawlik, Julia M Soto, Renee M Wilson," + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29245", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. Monitoring gait at home with radio waves in parkinson's disease: A marker of severity, progression, and medication response. Science Translational Medicine, 14(663): eadc9669, 2022. 3", + "[24] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 669–684. Springer, 2019. 3", + "[25] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12404-12413, 2021. 1, 3", + "[26] Hao Lu, Xuesong Niu, Jiyao Wang, Yin Wang, Qingyong Hu, Jiaqi Tang, Yuting Zhang, Kaishen Yuan, Bin Huang, Zitong Yu, et al. Gpt as psychologist? preliminary evaluations for gpt-4v on visual affective computing. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) workshop, 2024. 2", + "[27] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Minghui Qiu, Pengcheng Lu, Tao Wang, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023. 2, 3", + "[28] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 3", + "[29] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2, 3", + "[30] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 6, 7", + "[31] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2019. 3, 6", + "[32] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3170-3180, 2023. 2", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[34] Ambareesh Revanur, Zhihua Li, Umur A Ciftci, Lijun Yin, and László A Jeni. The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2760-2767, 2021. 6", + "[35] Eduardo Rocha Rodrigues, Igor Oliveira, Renato Cunha, and Marco Netto. Deepdownscale: A deep learning strategy for high-resolution weather forecast. In 2018 IEEE 14th Interna-" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional Conference on e-Science (e-Science), pages 415-422. IEEE, 2018. 1", + "[36] Sven Schellenberger, Kilin Shi, Tobias Steigleder, Anke Malessa, Fabian Michler, Laura Hameyer, Nina Neumann, Fabian Lurz, Robert Weigel, Christoph Ostgathe, et al. A dataset of clinically recorded radar vital signs with synchronised reference sensor signals. Scientific data, 7(1):291, 2020. 6", + "[37] Sebastian Scher and Gabriele Messori. Predicting weather forecast uncertainty with machine learning. Quarterly Journal of the Royal Meteorological Society, 144(717):2830-2841, 2018. 1", + "[38] Casper Kaae Sønderby, Lasse Espeholt, Jonathan Heek, Mostafa Dehghani, Avital Oliver, Tim Salimans, Shreya Agrawal, Jason Hickey, and Nal Kalchbrenner. Metnet: A neural weather model for precipitation forecasting. arXiv preprint arXiv:2003.12140, 2020. 3", + "[39] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3", + "[40] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In The 23rd IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062. IEEE, 2014. 6", + "[41] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 3", + "[42] Jingyuan Wang, Jiawei Jiang, Wenjun Jiang, Chengkai Han, and Wayne Xin Zhao. Towards efficient and comprehensive urban spatial-temporal prediction: A unified library and performance benchmark. arXiv preprint arXiv:2304.14343, 2023.6", + "[43] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 1-7. IEEE, 2020. 6", + "[44] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 6, 7", + "[45] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. Simper: Simple self-supervised learning of periodic targets. arXiv preprint arXiv:2210.03115, 2022. 1", + "[46] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2, 3", + "[47] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3", + "[48] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29246", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 6, 7", + "[50] Yunhua Zhang, Ling Shao, and Cees GM Snoek. Repetitive activity counting by sight and sound. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14070-14079, 2021. 1", + "[51] Yuting Zhang, Hao Lu, Xin Liu, Yingcong Chen, and Kaishun Wu. Advancing generalizable remote physiological measurement through the integration of explicit and implicit prior knowledge. arXiv preprint arXiv:2403.06947, 2024. 1", + "[52] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29247", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/96cc4bcd-798e-44b6-b645-30ba105eaf18_model.json b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/96cc4bcd-798e-44b6-b645-30ba105eaf18_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5151339daecfb90dcc075bb355ddaee87aff8561 --- /dev/null +++ b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/96cc4bcd-798e-44b6-b645-30ba105eaf18_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.244, + 0.131, + 0.758, + 0.177 + ], + "angle": 0, + "content": "Period-LLM: Extending the Periodic Capability of Multimodal Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.203, + 0.882, + 0.24 + ], + "angle": 0, + "content": "Yuting Zhang\\(^{1*}\\), Hao Lu\\(^{1,2*}\\), Qingyong Hu\\(^{2}\\), Yin Wang\\(^{3}\\), Kaishen Yuan\\(^{1}\\), Xin Liu\\(^{4}\\), Kaishun Wu\\(^{1\\ddagger}\\) \n\\(^{1}\\)The Hong Kong University of Science & Technology (Guangzhou)," + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.24, + 0.802, + 0.258 + ], + "angle": 0, + "content": "2The Hong Kong University of Science & Technology, 3Zhejiang University," + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.258, + 0.684, + 0.276 + ], + "angle": 0, + "content": "4Lappeenranta-Lahti University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.275, + 0.786, + 0.293 + ], + "angle": 0, + "content": "{yzhang430, hlu585} @connect.hkust-gz.edu.cn, wuks@hkust-gz.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.293, + 0.674, + 0.311 + ], + "angle": 0, + "content": "* Equal contribution, ‡ Corresponding author" + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.342, + 0.262, + 0.357 + ], + "angle": 0, + "content": "Periodic Tasks" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.365, + 0.239, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.461, + 0.208, + 0.485 + ], + "angle": 0, + "content": "Motion Counting" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.365, + 0.345, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.46, + 0.311, + 0.484 + ], + "angle": 0, + "content": "Traffic Flow" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.365, + 0.451, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.46, + 0.427, + 0.48 + ], + "angle": 0, + "content": "Weather Forecast" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.365, + 0.557, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.46, + 0.552, + 0.48 + ], + "angle": 0, + "content": "Remote Health Care" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.318, + 0.839, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.621, + 0.359, + 0.837, + 0.385 + ], + "angle": 0, + "content": "How many times did the person in the video repeat running on the treadmill?" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.391, + 0.609, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.415, + 0.601, + 0.421 + ], + "angle": 0, + "content": "GPT-4" + }, + { + "type": "text", + "bbox": [ + 0.633, + 0.393, + 0.846, + 0.431 + ], + "angle": 0, + "content": "The analysis did not detect any clear repetitions in the video based on the motion patterns. Hard to analyse" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.438, + 0.609, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.469, + 0.61, + 0.502 + ], + "angle": 0, + "content": "Period-LLM" + }, + { + "type": "text", + "bbox": [ + 0.633, + 0.442, + 0.863, + 0.468 + ], + "angle": 0, + "content": "The person in the video repeated running 5 times. \\(\\times\\) Wrong number" + }, + { + "type": "text", + "bbox": [ + 0.631, + 0.477, + 0.849, + 0.502 + ], + "angle": 0, + "content": "The repetitive motion in the video is running on the treadmill 14 times." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.51, + 0.907, + 0.552 + ], + "angle": 0, + "content": "Figure 1. Existing multimodal models may fail to analyze periodic tasks, such as motion counting, traffic flow, weather forecasting, and remote health care. For example, GPT-4 fails to detect clear repetitions due to the difficulty in analyzing motion patterns, and VideoLLaMA incorrectly counts the repetitions. In contrast, Period-LLM accurately analyzes the actions and provides correct repetition count." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.564, + 0.327, + 0.58 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.584, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Periodic or quasi-periodic phenomena reveal intrinsic characteristics in various natural processes, such as weather patterns, movement behaviors, traffic flows, and biological signals. Given that these phenomena span multiple modalities, the capabilities of Multimodal Large Language Models (MLLMs) offer promising potential to effectively capture and understand their complex nature. However, current MLLMs struggle with periodic tasks due to limitations in: 1) lack of temporal modelling and 2) conflict between short and long periods. This paper introduces Period-LLM, a multimodal large language model designed to enhance the performance of periodic tasks across various modalities, and constructs a benchmark of various difficulty for evaluating the cross-modal periodic capabilities of large models. Specially, We adopt an \"Easy to Hard Generalization\" paradigm, starting with relatively simple text-based tasks and progressing to more complex visual and multimodal tasks, ensuring that the model gradually builds robust periodic reasoning capabilities. Additionally, we propose a \"Resisting Logical Oblivion\" optimization strategy to maintain periodic reasoning abilities dur" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.626 + ], + "angle": 0, + "content": "ing semantic alignment. Extensive experiments demonstrate the superiority of the proposed Period-LLM over existing MLLMs in periodic tasks. The code is available at https: //github.com/keke-nice/Period-LLM." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.631, + 0.646, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Periodic tasks refer to activities performed at fixed or predictable intervals, encompassing a wide range of applications. At a macro level, this includes repetitive motions in human daily life, such as step counts and the number of times one jumps rope [7, 10, 50], as well as the periodic analysis of meteorological data, where scientists can identify various weather patterns and trends, forecast future weather conditions [1, 35, 37], and issue timely warnings. At a micro level, in health monitoring, by capturing and analyzing subtle chromatic changes in facial videos, physiological signals such as heart rate, blood pressure, and respiratory rate can be extracted [22, 25, 45, 51]. The periodic fluctuations of these signals reflect the physiological state of human. Effectively handling periodic tasks can significantly enhance various aspects of human life, from predicting weather patterns to monitoring health conditions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "29237" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "Given that periodic phenomena occur across various modalities, it becomes essential to utilize multimodal models for period analysis instead of single modality. In recent years, multimodal large language models [3, 16, 18, 48] have advanced the progress of visual and linguistic learning by incorporating visual encoders into various pre-trained large language models. Furthermore, some studies have extended vision to other modalities such as audio [9, 48], depth imagery [8], and near-infrared [8]. With the promising progress of LLM-enabled multimodal models, a natural question arises: Can multimodal large language models understand periodic tasks?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.277, + 0.485, + 0.746 + ], + "angle": 0, + "content": "Unfortunately, existing large models have shown inadequacies in counting capabilities or understanding \"repetitive\" temporal sequences [15], creating obstacles for the application of periodic tasks in multimodal large language models, as shown in Fig. 1. These shortcomings mainly stem from the following challenges: 1) Interference of spatial pseudo-temporal information. The leakage of numerical information in visuals may lead models to take shortcuts rather than learning the true periodic information. When specific numbers appear in videos, models may tend to directly answer with these numbers, or be misled by them, rather than deriving answers through the analysis of periodic information [32]. This limits the models' generalization capabilities, rendering them unable to handle periodic tasks without direct numerical information. 2) Conflict between long-time periodic reasoning ability and short-time semantic alignment. The recognition of a periodic task lies not only in the ability to deduce the number of periodic repeats, but also in the accurate recognition of the semantic information in the video. Both require the model to have the ability to model the timing, but with different focus. Too much focus on the short-time semantic understanding may lead to ignorance of periodic reasoning ability. 3) Absence of captions is a significant issue. In the training corpora of MLLMs, most sentences are descriptive and rarely contain specific descriptions of the number of actions. This poses difficulties for models when dealing with periodic tasks that require counting the number of actions. For example, a model might respond with \"The boy in the video is doing pull-ups\" or \"The boy in the video does multiple pull-ups\", without a specific counting supervision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To avoid overfitting to modality-specific features while preserving temporal modeling capabilities, we propose a MLLM with periodic capabilities, termed Period-LLM. Specifically, we employ an easy-to-hard generalization paradigm. Initially, we define text-based, macro, and micro periodic tasks, ranging from easy to hard. We anticipate starting with text and progressively advancing, enabling Large Language Models (LLMs) to learn periodicity across different modalities. The rationale for this definition is twofold: 1) Despite modality alignment, LLMs still ex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.529 + ], + "angle": 0, + "content": "hibit superior text processing capabilities compared to other modalities; 2) Macro periodic tasks contain clear semantic features, making their periodic nature easier to capture than that of micro tasks. Accordingly, we selected periodic tasks corresponding to different levels of difficulty. At the textual level, we construct a \"repeated word question-answer\" dataset to pre-train LLMs, deepening their understanding of \"repetitiveness\" from a textual perspective. At the macro level, we select the repetitive motion dataset Countix [7], which aims to predict the number of repetitive movements in videos. At the micro level, we choose the rPPG task, which involves capturing subtle periodic or quasi-periodic physiological signals from facial videos or radio-frequency signals. Furthermore, it is believed that a good periodic model should have both the ability of periodic reasoning and the ability of semantic understanding. Focusing on the optimization of semantic alignment may, to some extent, lead to forgetting the logical reasoning abilities learned from simple tasks. Therefore, we propose a Resisting Logical Oblivion optimization strategy. Specifically, by manually designing a method to dynamically update the backpropagation gradients, the model can adjust previously overlooked feature dimensions without interfering with existing ones, providing a guided approach to model optimization. A series of experiments demonstrate the superiority of the proposed Period-LLM over existing MLLMs in periodic tasks. This advantage is not limited to text and video; we further explore the potential of other modalities based on this foundation." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.53, + 0.837, + 0.544 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.548, + 0.905, + 0.607 + ], + "angle": 0, + "content": "- To the best of our knowledge, we propose the first MLLM-based periodic model (Period-LLM) and construct a benchmark featuring periodic tasks of varying difficulty, including text, video, and other modalities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.609, + 0.905, + 0.683 + ], + "angle": 0, + "content": "- We propose an easy-to-hard training paradigm that progresses from simple to complex tasks, enabling the model to sequentially acquire the ability to perceive periodic information across different modalities, starting with text, then vision, and extending to other modalities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.684, + 0.905, + 0.729 + ], + "angle": 0, + "content": "- We propose a Resisting Logical Oblivion (RLO) optimization strategy to resolve the conflict between periodic reasoning and semantic alignment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.729, + 0.905, + 0.773 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that the proposed Period-LLM outperforms existing MLLMs in periodic tasks across various modalities." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.548, + 0.905, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.779, + 0.655, + 0.794 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.825, + 0.82 + ], + "angle": 0, + "content": "2.1. Multimodal Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In recent studies, large models have exhibited reliable capabilities in the field of video understanding. In addition to the powerful vision-language large models [12, 17, 26, 52], there is an increasing focus on exploring more modalities in recent research [13, 27-29, 46]. Bain et al. [5] intro" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29238" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.117, + 0.878, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.274, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Figure 2. The process of the generation of Question-and-answer datasets. Samples of text and video are taken as examples. In the text sample, questions and the occurrences of specific words are randomly generated, while in the video sample, the questions are randomly generated and the keywords represent labels within the dataset. These elements are then incorporated into a GPT-4 prompt to construct answers that include key information. The right side of the image shows examples of the final Q&A pairs generated, where the answers are complete sentences containing the critical information from the key elements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.519 + ], + "angle": 0, + "content": "duce a large-scale dataset, which provides general descriptions associated with video content. Several LLM-based works [13, 27, 29, 46] are proposed to comprehend the visual content of the video. Besides, Video-LLaMa [47] extends the ability to comprehend both auditory and visual information, while Su et al. [41] leverages the power of multimodal encoders to understand across six modalities. Recently, Lv et al. [28] proposed video-based large language models for the task of VAD in a weakly supervised framework. Different from them, our work focuses on the periodic analysis of MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.529, + 0.24, + 0.543 + ], + "angle": 0, + "content": "2.2. Periodic Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Examples of periodic learning include recovering and magnifying physiological signals (e.g., heart rate or breathing) [25, 31], predicting weather and environmental changes (e.g., nowcasting of precipitation or land surface temperatures) [38], counting motions that are repetitious (e.g., exercises or therapies) [4, 7], and analyzing human behavior (e.g., gait) [23]. Prior work has focused on designing customized neural architectures [7, 20], loss functions [39], and leveraging relevant learning paradigms including transfer learning [24] and meta-learning [21] for periodic learning. This paper attempts to study the capability of periodic tasks in multimodal models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.406, + 0.757 + ], + "angle": 0, + "content": "3. Problem Setting and Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In nature, periodic phenomena contain both periodic and semantic information, which can be modeled by \\( x = K \\cdot p(\\omega t) + N \\cdot s(t) \\), where \\( p(\\omega t) \\) is a periodic function with angular frequency \\( \\omega = 2\\pi / T \\) and period \\( T \\). Here, \\( K \\) represents the amplitude of the periodic component, \\( N \\) denotes the amplitude of the semantic component, and \\( s(t) \\) is a semantic function related to time \\( t \\). When discretizing continuous sequences as model inputs \\( X \\), we adopt a sampling interval \\( \\Delta t = T / n \\), where \\( n \\) is a positive integer multiple" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.413 + ], + "angle": 0, + "content": "of the sampling rate \\((n = 1,2,3,\\ldots)\\). Ideally, when \\(n = 1\\) and semantics are absent, the value at each sampling point equals the constant \\(K\\), resulting in \\(x = K\\). This can be mathematically expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.422, + 0.906, + 0.475 + ], + "angle": 0, + "content": "\\[\nX = \\left\\{ \\begin{array}{l l} K \\cdot p (\\omega t) + N \\cdot s (t), & \\Delta t = \\frac {T}{n}, n = 2, 3, 4, \\dots \\\\ K, & \\Delta t = T \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.627 + ], + "angle": 0, + "content": "However, in practical applications, data is often high-dimensional and noisy. For instance, after processing a text segment through a tokenizer, we obtain a matrix with dimensions \\([L, \\mathrm{Dim}]\\), where \\(L\\) represents the length of the text (i.e., the number of tokens), and Dim denotes the embedding dimension of each token. In cases of repetitive text such as \"period period period\", certain dimensions of the encoded feature matrix may exhibit constant values similar to \\(x = K\\), where \\(K\\) is a vector representing the embedding values of these tokens in specific dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Inputs from other modalities, such as video, present more complex scenarios. Periodic movements may be obscured by a large amount of semantic or redundant information. In macro-periodic tasks, periodic movement is more prominent, allowing models to detect periodicity more easily; however, in micro-periodic tasks, the amplitude of the periodic signal \\((K)\\) may be small and masked by noise, making it challenging for models to capture these subtle periodic variations. Despite these complexities, research suggests that models can transfer knowledge from simpler tasks to more complex ones. In this context, repetitive text can be seen as a basic periodic task, while macro-periodic and micro-periodic video tasks represent more complex periodic challenges. Although \"repetition\" is the common thread across these tasks, the semantic content and amplitude of the periodic signal vary." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Given that large language models (LLMs) have been trained on extensive and diverse text corpora, they are par" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29239" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.116, + 0.835, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.225, + 0.908, + 0.267 + ], + "angle": 0, + "content": "Figure 3. The pipeline of training paradigm from easy to hard. The model begins training with simpler periodic tasks and gradually transitions to more complex ones. Throughout this process, Period-LLM's understanding of periodic tasks progressively strengthens and becomes more profound." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.483, + 0.304 + ], + "angle": 0, + "content": "ticularly well-suited for processing text tasks. We propose the following framework:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.304, + 0.482, + 0.32 + ], + "angle": 0, + "content": "\\[\nA = F (X, Q) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.485, + 0.485 + ], + "angle": 0, + "content": "This framework accommodates inputs \\( X \\) from various modalities, where \\( X \\) represents periodic data from any modality, \\( Q \\) represents a given question, and the model computes the answer \\( A \\). We suggest beginning model training with simple repetitive text tasks and gradually progressing to macro-periodic and micro-periodic video tasks. This progressive training approach helps the model develop a comprehensive understanding of periodic phenomena, moving from simple to complex tasks, thereby enhancing its ability to handle periodic challenges effectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.182, + 0.513 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.689 + ], + "angle": 0, + "content": "In this paper, we aim to enhance the understanding of large models regarding periodic tasks. Given the current absence of specialized question-and-answer datasets for periodic tasks and corresponding testing standards, we will undertake pioneering work in the content that follows. Specifically, we firstly elucidate the instruction generation for periodic tasks. Then, we explore the generalization process from simpler tasks to more complex ones. Additionally, we propose an optimization strategy to address the challenge of preserving periodic reasoning abilities during the semantic alignment phase of Multi-Modal Generalization." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.694, + 0.449, + 0.709 + ], + "angle": 0, + "content": "4.1. Instruction Generation for Periodic Tasks" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.716, + 0.295, + 0.73 + ], + "angle": 0, + "content": "4.1.1. Text-Only Instructions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To enhance our dataset's diversity, we use GPT-4 to help create a repetitive text dataset through several key steps: First, we define the core question: \"How many times is the word \\(\\{my\\text{word}\\}\\) repeated in the string \\(\\{my\\text{string}\\} ?\\) We then use GPT-4 to generate ten semantically similar questions, \\(q_{i}\\in \\{q_{1},q_{2},\\ldots ,q_{10}\\}\\) . For word diversity, we randomly select a word from the GPT-4 technical report [2] as \\(\\{my\\text{word}\\}\\) . To further diversify, we set a repetition count \\(n\\) within \\(\\{2,3,\\dots,20\\}\\) , constructing \\(\\{my\\text{string}\\}\\) as \" \\(\\{\\mathrm{selected~word}\\} \\times n\\) . Since GPT-4 can't directly calculate word repetitions, we guide it with precise prompts. We" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.395 + ], + "angle": 0, + "content": "instruct GPT-4: \"Construct an answer based on the question format, selected word, and repetition count.\" During this process, we randomly choose a question \\( q_{i} \\) and provide the necessary answer elements, {selected word, \\( n \\) times}, enabling GPT-4 to generate a fitting response. By following these steps, we effectively use GPT-4's generative capabilities to build a diverse and accurate repetitive text dataset, which supports robust model training and evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.406, + 0.738, + 0.42 + ], + "angle": 0, + "content": "4.1.2. Multi-Modal Instructions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.907, + 0.789 + ], + "angle": 0, + "content": "For tasks involving periodic data across various modalities like video, traffic streams, and health monitoring, we employ a method that integrates annotations, raw captions, and frequency information to generate comprehensive representations. This approach is applicable to diverse datasets, addressing challenges such as dynamic environments, variations in periodicity and counts, and action speeds. We transform data into a question-and-answer format inspired by repetitive text dataset construction. We generate unified captions for each segment by combining existing annotations and frequency data. The main question posed is: \"What is the total number of repetitive actions in this data?\" Using GPT-4's capabilities, we craft multiple semantically similar questions from this core inquiry. Explicit instructions and answer elements are provided to generate question-and-answer pairs, enhancing dataset representation, the process as shown in Fig. 2. This method enriches model understanding and processing of periodic tasks across modalities. We establish a link between repetitive text and periodic data, enabling generalization from simple to complex tasks. The process first involves self-optimization in relatively simple individual tasks, followed by fine-tuning on more complex tasks, as illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.774, + 0.817 + ], + "angle": 0, + "content": "4.2. Easy-to-Hard Generalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this section, we explore the generalization process from simpler tasks to more complex ones, structured around the problem definition \\( A = F(f, Q) \\), where \\( A \\) represents the answer, \\( f \\) represents the features derived from the input data, and \\( Q \\) represents the question." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29240" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.092, + 0.862, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.309, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Figure 4. The Period-LLM framework, using video as an example, integrates both text and video inputs. The video is processed through a visual encoder and a visual projector, then concatenated with text features derived from a word embedding layer. This combined input is fed into a large language model (LLM). Gradient optimization strategies are applied to the output feature channels of the LLM, specifically targeting those channels that failed to capture useful information, thereby enriching the semantic content in these channels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.373, + 0.315, + 0.387 + ], + "angle": 0, + "content": "4.2.1. Text-Only Generalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.483, + 0.451 + ], + "angle": 0, + "content": "The first stage of generalization focuses on text-based tasks, where the goal is to learn logical reasoning without needing to align other modalities with text. This can be mathematically represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.45, + 0.482, + 0.465 + ], + "angle": 0, + "content": "\\[\nA = F \\left(T _ {f}, Q\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.581 + ], + "angle": 0, + "content": "here, \\( T_{f} \\) represents the features derived from textual input. The emphasis is on developing the model's logical reasoning capabilities, as the input modality is solely text, which simplifies the task by eliminating the need to handle multimodal alignment. The model learns to process and understand the logical structure of questions and generate appropriate answers purely based on textual information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.586, + 0.336, + 0.6 + ], + "angle": 0, + "content": "4.2.2. Multi-Modal Generalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.68 + ], + "angle": 0, + "content": "The next stage involves more complex tasks that require the integration of multiple modalities, such as video and text. This stage not only demands logical reasoning but also the alignment of semantic information across different modalities. This process can be described by:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.68, + 0.482, + 0.696 + ], + "angle": 0, + "content": "\\[\nA = F \\left(M _ {f}, Q\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.827 + ], + "angle": 0, + "content": "in this case, \\( M_{f} \\) represents features derived from a combination of inputs, including video and text. The challenge is twofold: the model must align semantic content across different modalities while simultaneously developing robust periodic logical reasoning capabilities. This complex task requires the model to integrate semantic features with textual information and establish accurate connections between them in order to answer questions correctly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.331, + 0.85 + ], + "angle": 0, + "content": "4.3. Resisting Logical Oblivion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Due to the significant differences between periodic logical reasoning abilities and the ability to understand semantic information in videos, we design a Resisting Logical Oblivion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.907, + 0.51 + ], + "angle": 0, + "content": "(RLO) optimization strategy to prevent the new knowledge learned in a new round of optimization from overwriting the reasoning knowledge acquired from simpler tasks. Vanilla gradient descent leads to an undifferentiated update of the entire feature space. If \\(\\theta = (\\theta_{1},\\theta_{2},\\dots,\\theta_{J})\\in \\mathbb{R}^{C\\times W}\\) is the parameter of the model, where \\(W\\) is related to the network structure, \\(\\frac{\\partial L(\\theta)}{\\partial\\theta_j}\\) \\((j = 1,2,\\ldots ,J)\\) represents the partial derivative of \\(\\theta\\). The process of model updating is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.51, + 0.905, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\nabla \\theta_ {j} = - \\alpha \\frac {\\partial L (\\theta)}{\\partial \\theta_ {j}}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.545, + 0.905, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\theta_ {j} ^ {\\prime} \\leftarrow \\theta_ {j} + \\nabla \\theta_ {j}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.575, + 0.906, + 0.606 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the learning rate, \\(\\nabla \\theta_{j}\\) is the updating vector, and \\(\\theta_j^\\prime\\) is the new weight for the next iteration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.727 + ], + "angle": 0, + "content": "Traditional gradient update strategies adjust model parameters based on global errors, which, although effective, can lead to knowledge interference between different types of tasks. Our new strategy introduces a feature channel weight allocation mechanism, dynamically adjusting the gradient update weights of different feature channels. Specifically, we introduce the weight function \\(\\Omega (k,c)\\) implementation:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.727, + 0.905, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\Omega \\left(c _ {i}\\right) = \\left\\{ \\begin{array}{l l} 1 + \\beta \\cdot e ^ {\\frac {\\text {i t e r n u m}}{\\text {m a x i t e r}}}, & \\bar {c} _ {i} < \\bar {c}, \\\\ 1, & \\bar {c} _ {i} > \\bar {c}, \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a hyper-parameter, \\(iter_{num}\\) denotes the current iteration count, and \\(max_{iter}\\) represents the maximum number of iterations. \\(\\bar{c}\\) is the average across all feature channels, while \\(\\bar{c}_i\\) refers to the average of the \\(i\\)th feature channel. By dynamically weighting feature channels, we can exert control over the real-time updates of specific feature channels. This ensures that redundant feature channels can learn newly acquired semantic information without" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29241" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.473, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.209, + 0.483, + 0.277 + ], + "angle": 0, + "content": "Figure 5. Details of the function of the gradient weight function. The left side of the image shows that some channels have lower backpropagation gradients, while others have higher ones. Higher backpropagation gradients indicate that these feature channels are being updated more significantly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.484, + 0.461 + ], + "angle": 0, + "content": "compromising existing reasoning abilities, thereby balancing the enhancement of reasoning abilities and video semantic understanding. As shown in Fig. 5, our gradient weighting function is applied to the backpropagation gradients, boosting those originally lower gradients, thereby substantially increasing the update of channels that were previously updated little or not at all. This process ensures that newly learned semantic information is directed to these newly emphasized feature channels for learning. Therefore, the model parameter update process with the introduction of the RLO optimization strategy is as follows, showing the changes from Eq. 5 to Eq. 8:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.464, + 0.482, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\nabla \\theta_ {j} ^ {*} = \\Omega \\left(c _ {i}\\right) \\cdot \\nabla \\theta_ {j}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.234, + 0.484, + 0.482, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\theta_ {j} ^ {\\prime} \\leftarrow \\theta_ {j} + \\nabla \\theta_ {j} ^ {*}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.484, + 0.634 + ], + "angle": 0, + "content": "During the training process, the input data \\(\\mathcal{Z}\\) consists of input videos and their corresponding questions \\(x_{i}\\), correct answers \\(y_{i}\\), where \\(i = 1,2,3,\\ldots ,K\\), and each \\(x_{i},y_{i}\\) is a sequence of tokens. Large language models are denoted by \\(P_{\\phi}(y\\mid x)\\), where \\(\\phi\\) represents model parameters. Typically, when fine-tuning large models, only the question-answer pairs \\((x_{i},y_{i})\\) are input. The goal is to adjust the parameters \\(\\phi\\) to maximize the following expression:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.64, + 0.483, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\phi} \\sum_ {(x, y) \\in \\mathcal {Z}} \\sum_ {t = 1} ^ {| y |} \\log \\left(P _ {\\phi} \\left(y _ {t} \\mid x, y _ {< t}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.484, + 0.788 + ], + "angle": 0, + "content": "where \\( |y| \\) represents the length of the input vector \\( y \\), \\( P_{\\phi}(y_t|x, y_{MethodLLMZero-shotFine-tuneCountix-QArPPG-QACountix-QArPPG-QAMAERMSEBleu1CIDErMETEORMAERMSEMAERMSEBleu1CIDErMETEORMAERMSEFrozenBiLM [44]DeBERTa-V27.2211.530.1450.5500.11220.4321.565.229.440.2330.6100.20119.2120.46VideoLLaMA [48]Vicuna-7B6.1810.820.1870.3600.24518.8919.434.988.890.3890.5700.35818.2919.01LLaMA-Adapter [49]LLaMA-7B7.4410.510.2440.6830.31221.7823.345.7810.020.3780.7890.39720.5422.84VideoChat [14]Vicuna-7B6.7311.230.1650.5490.27719.5420.675.039.530.2750.6790.38818.3210.77Video-ChatGPT [30]Vicuna-7B6.4410.750.2030.5890.29718.1220.454.649.150.3560.6430.36717.5419.45BT-Adapter [19]Vicuna-7B7.7311.700.2760.6740.34119.3722.686.129.700.2960.6890.35318.6721.49LLaMA-VID [16]Vicuna-7B6.9711.210.2670.6800.30118.3118.565.349.850.3880.7830.41117.5117.60Period-LLMLLaMA-7B5.518.330.2880.7020.30815.0817.133.777.140.3970.8100.41513.7816.83" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.227, + 0.905, + 0.269 + ], + "angle": 0, + "content": "Table 1. Comparison of our method with other leading methods on two newly introduced video question-and-answer (QA) datasets, with all methods evaluated at a resolution of \\( {224} \\times {224} \\) . The results for Period-LLM in the zero-shot setting are derived from a model pretrained on text-QA dataset. For rPPG-QA, the V4V dataset is used as the validation set." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.279, + 0.861, + 0.424 + ], + "angle": 0, + "content": "
MethodLLMZero-shotFine-tune
RotNIST-QADrive-QARadar-QARotNIST-QADrive-QARadar-QA
MAERMSEMAERMSEMAERMSEMAERMSEMAERMSEMAERMSE
FrozenBiLM [44]DeBERTa-V24.264.7830.88120.5622.1730.422.883.2529.44110.5616.6628.55
VideoLLaMA [48]Vicuna-7B4.085.3235.45130.6723.4218.852.513.8931.22121.5517.0320.34
LLaMA-Adapter [49]LLaMA-7B3.954.4231.26125.4224.1231.792.663.2330.28118.7920.3428.56
VideoChat [14]Vicuna-7B3.564.2132.67126.8821.7129.582.233.0630.12115.2320.5427.38
Video-ChatGPT [30]Vicuna-7B3.444.8738.18118.2923.7732.132.012.2933.28112.3521.6130.45
BT-Adapter [19]Vicuna-7B4.035.4634.21110.2423.4231.472.783.5631.31107.6522.1230.19
LLaMA-VID [16]Vicuna-7B3.874.5734.98114.3318.2129.322.433.2532.45109.8718.2129.32
Period-LLMLLaMA-7B3.203.9630.88108.7417.5626.451.502.7528.71105.5414.2424.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.425, + 0.905, + 0.466 + ], + "angle": 0, + "content": "Table 2. Comparison of our method with other leading methods on the RotNIST-QA, Drive-QA, and Radar-QA datasets. Drive-QA is a dataset of taxi trajectory data, while Radar-QA is a radar modality dataset capturing physiological signals. The zero-shot results for Period-LLM are based on a model pretrained on text-QA, Countix-QA, and rPPG-QA." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.483, + 0.652 + ], + "angle": 0, + "content": "periodic reasoning abilities. The results show significant improvements over zero-shot, thanks to the inclusion of explicit periodic information in the training data. Period-LLM surpasses existing MLLMs by over 1 MAE on the Countix-QA dataset, while also showing significant improvements in generated text quality, as measured by Bleu-1, CIDEr, and METEOR scores. This improvement is driven by the easy-to-hard fine-tuning paradigm and the Resisting-Logical-Oblivion strategy, which together help the model preserve and enhance its periodic reasoning abilities, while also maintaining strong semantic understanding, resulting in superior performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In addition to the visual periodic tasks, Table 2 presents the performance of these models on other modalities, specifically the RotNIST-QA, Drive-QA, and Radar-QA datasets. The zero-shot and fine-tune protocols follow the same approach as in Table 1, with the key difference being that experiments on these three datasets were conducted in parallel. The pretraining models used in these experiments were fine-tuned on the two visual periodic tasks from Table 1. On the RotNIST-QA dataset, Period-LLM achieves the lowest MAE (1.50) and RMSE (2.75), outperforming models such as FrozenBiLM, VideoLLaMA, and LLaMA-Adapter. For the Drive-QA dataset, Period-LLM demonstrates superior performance with the lowest MAE (28.71) and competitive RMSE (105.54). On the Radar-QA dataset, Period-LLM also performs strongly, further validating the effectiveness of our model across different modali" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.547 + ], + "angle": 0, + "content": "ties. These results underscore Period-LLM's enhanced accuracy in periodic reasoning tasks and its robustness in handling diverse datasets, establishing it as a versatile solution for periodic question-answering across a range of domains, including health, weather, and traffic." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.696, + 0.577 + ], + "angle": 0, + "content": "5.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.905, + 0.827 + ], + "angle": 0, + "content": "QA Examples. Fig. 6 showcases examples from the countix-QA and rPPG-QA datasets. It can be seen that Period-LLM is capable of understanding the semantic information contained within the video as well as recognizing periodicity within the video. For instance, in the countix-QA dataset, Period-LLM accurately provides the number of repetitive movements, while in the rPPG-QA dataset, it can determine the heart rate of the man in the video. These capabilities are beyond what other language models can achieve. According to Fig. 6, in the countix-QA dataset, Period-LLM identifies that the man in the video performs two pull-ups. In the rPPG-QA dataset, Period-LLM is able to detect that the man's heart rate is 75 beats per minute. This demonstrates Period-LLM's exceptional performance in handling video content and periodic information. More QA examples are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Comparison of Loss Reduction. The Fig. 7 illustrates the loss reduction comparison between the Baseline (Vanilla Fine-Tuning on Countix-QA) and the proposed Period-LLM approach (with text-QA pretraining and RLO) on the Countix-QA dataset. As shown, our pro" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29243" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.133 + ], + "angle": 0, + "content": "Table 3. The MAE value on Countix-QA with different parameter or setting choices in \\( \\Omega \\left( {c}_{i}\\right) \\) of RLO optimization strategy, and ablation study on Countix-QA showing the impact of easy tasks (period-text-QA) and RLO (Resisting Logical Oblivion). Default settings are marked in gray ." + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.145, + 0.203, + 0.158 + ], + "angle": 0, + "content": "(a) Impact of \\(\\beta\\)" + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.161, + 0.203, + 0.215 + ], + "angle": 0, + "content": "
βMAE
0.013.85
0.053.77
0.13.90
0.54.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.145, + 0.347, + 0.158 + ], + "angle": 0, + "content": "(b) Type of Threshold." + }, + { + "type": "table", + "bbox": [ + 0.238, + 0.161, + 0.34, + 0.208 + ], + "angle": 0, + "content": "
ThresholdMAE
Mean3.77
Median3.85
Learnable4.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.352, + 0.133, + 0.604, + 0.157 + ], + "angle": 0, + "content": "(c) Results on Countix-QA w./w.o Easy Task and RLO." + }, + { + "type": "table", + "bbox": [ + 0.358, + 0.161, + 0.603, + 0.221 + ], + "angle": 0, + "content": "
Easy to HardRLOCountix-QA
Text-QACountix-QAMAECIDEr
✗✓✗4.300.661
✓✓✗3.890.782
✓✓✓3.770.810
" + }, + { + "type": "table_caption", + "bbox": [ + 0.61, + 0.145, + 0.895, + 0.158 + ], + "angle": 0, + "content": "(d) Results on rPPG-QA w./w.o Easy Tasks and RLO." + }, + { + "type": "table", + "bbox": [ + 0.615, + 0.161, + 0.897, + 0.228 + ], + "angle": 0, + "content": "
Easy to HardRLOrPPG-QA
Text-QACountix-QArPPG-QAMAE
XX✓X15.45
✓X✓X14.46
✓✓✓X14.05
✓✓✓✓13.78
" + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.223, + 0.435, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.339, + 0.443, + 0.353 + ], + "angle": 0, + "content": "Figure 6. Examples from Countix-QA and rPPG-QA." + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.355, + 0.441, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.517 + ], + "angle": 0, + "content": "Figure 7. Comparison of Loss Reduction Between Baseline (Vanilla Fine-Tuning on Countix-QA) and Proposed Method (Period-LLM with Text-QA Pretraining and RLO, Fine-Tuned on Countix-QA)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.664 + ], + "angle": 0, + "content": "posed method demonstrates a significantly faster decline in the loss values compared to the baseline. The Period-LLM model achieves a steeper descent in the initial training stages, indicating a more efficient learning process. Furthermore, it converges to a lower loss value in fewer epochs, suggesting improved optimization and stability. This faster convergence implies that our method better captures the underlying data patterns, leading to more effective fine-tuning on the Countix-QA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.245, + 0.683 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The Setting of Weight Function \\(\\Omega(c_i)\\) of RLO Optimization Strategy. As seen in Table 3a, \\(\\beta = 0.05\\) yields the lowest MAE (3.77), indicating an effective balance between retaining prior knowledge and integrating new information. When \\(\\beta\\) is too small (e.g., 0.01), the model's anti-forgetting ability is weak, causing higher error as new knowledge overwhelms older knowledge. Conversely, a high \\(\\beta\\) (e.g., 0.5) overly prioritizes old knowledge, limiting adaptation. Thus, \\(\\beta = 0.05\\) achieves the best balance for minimizing error. In Table 3b, setting the threshold to the mean results in the lowest MAE (3.77), effectively balancing old and new knowledge contributions. The median threshold performs slightly worse (MAE 3.85) due to its insensitivity to outliers, while a learnable threshold increases MAE" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.248, + 0.907, + 0.277 + ], + "angle": 0, + "content": "(4.20), likely due to instability. Hence, the mean threshold is the most stable and effective choice." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.907, + 0.57 + ], + "angle": 0, + "content": "Results of various combinations of Preiod-LLM Components. We evaluate our approach on the Countix-QA and rPPG-QA datasets, as shown in Tables 3c and 3d. For Countix-QA, our method—combining pre-training on simpler tasks with the Resisting Logical Oblivion (RLO) strategy—significantly improves performance, reducing the Mean Absolute Error (MAE) from 4.30 to 3.77 and increasing the CIDEr score from 0.661 to 0.810. Similarly, on the rPPG-QA dataset, this approach lowers the MAE from 15.45 to 13.78. These improvements reflect the model's enhanced capacity to capture periodic patterns effectively while also understanding semantic content more deeply. The decrease in MAE shows better accuracy in numerical predictions, while the CIDEr increase suggests improved semantic alignment in question-answering tasks. Together, these results confirm that the integration of easy-task pretraining with RLO enables the model to balance periodicity and semantic understanding, leading to faster convergence and more robust performance across QA tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.584, + 0.634, + 0.6 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce Period-LLM, a large multimodal model specifically designed to enhance performance on periodic tasks across various modalities. The model follows a progressive training paradigm, starting with simpler text-based tasks and advancing to more complex visual and multimodal tasks. Additionally, we propose the \"Resisting Logical Oblivion\" optimization strategy to preserve the model's periodic reasoning abilities during semantic alignment. The Our experimental results demonstrate that Period-LLM significantly outperforms existing models in handling periodic tasks, especially in challenging domains such as repetitive motion analysis and physiological signal estimation. Period-LLM not only achieves superior accuracy but also highlights its potential for future advancements in managing complex periodic phenomena across different modalities. Future work will focus on further refining model's capabilities, particularly in addressing more nuanced and less structured periodic tasks, and expanding its application to additional domains where periodicity plays a crucial role." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29244" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This work was supported by the Guangdong Provincial Key Laboratory of Integrated Communication, Sensing and Computation for Ubiquitous Internet of Things (No. 2023B1212010007), China NSFC Grant 62472366, the Project of DEGP (No.2024GCZX003, 2023KCXTD042), 111 Center (No.D25008) and Shenzhen Science and Technology Foundation (ZDSYS20190902092853047)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.234, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[1] Kumar Abhishek, Maheshwari Prasad Singh, Saswata Ghosh, and Abhishek Anand. Weather forecasting model using artificial neural network. Procedia Technology, 4:311-318, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.316, + 0.484, + 0.383 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.385, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.484, + 0.551 + ], + "angle": 0, + "content": "[4] Mohammad Rafayet Ali, Javier Hernandez, E Ray Dorsey, Ehsan Hoque, and Daniel McDuff. Spatio-temporal attention and magnification for classification of parkinson's disease from videos collected via the internet. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 207-214. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.553, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[6] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[7] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Counting out time: Class agnostic video repetition counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10387-10396, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.735, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[8] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xiangyu Yue. Onellm: One framework to align all modalities with language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26584-26595, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.259, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.175 + ], + "angle": 0, + "content": "[10] Huazhang Hu, Sixun Dong, Yiqun Zhao, Dongze Lian, Zhengxin Li, and Shenghua Gao. Transrac: Encoding multiscale temporal correlation with transformers for repetitive action counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19013-19022, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[11] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[12] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.906, + 0.345 + ], + "angle": 0, + "content": "[13] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.348, + 0.906, + 0.402 + ], + "angle": 0, + "content": "[14] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.906, + 0.487 + ], + "angle": 0, + "content": "[15] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[16] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.574 + ], + "angle": 0, + "content": "[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.906, + 0.617 + ], + "angle": 0, + "content": "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[19] Ruyang Liu, Chen Li, Yixiao Ge, Ying Shan, Thomas H Li, and Ge Li. One for all: Video conversation is feasible without video instruction tuning. arXiv preprint arXiv:2309.15785, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[20] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems, 33:19400-19411, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[21] Xin Liu, Ziheng Jiang, Josh Fromm, Xuhai Xu, Shwetak Patel, and Daniel McDuff. Metaphys: few-shot adaptation for non-contact physiological measurement. In Proceedings of the conference on health, inference, and learning, pages 154-163, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[22] Xin Liu, Yuting Zhang, Zitong Yu, Hao Lu, Huanjing Yue, and Jingyu Yang. rppg-mae: Self-supervised pretraining with masked autoencoders for remote physiological measurements. IEEE Transactions on Multimedia, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Yingcheng Liu, Guo Zhang, Christopher G Tarolli, Rumen Hristov, Stella Jensen-Roberts, Emma M Waddell, Taylor L Myers, Meghan E Pawlik, Julia M Soto, Renee M Wilson," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29245" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "et al. Monitoring gait at home with radio waves in parkinson's disease: A marker of severity, progression, and medication response. Science Translational Medicine, 14(663): eadc9669, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[24] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 669–684. Springer, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[25] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12404-12413, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[26] Hao Lu, Xuesong Niu, Jiyao Wang, Yin Wang, Qingyong Hu, Jiaqi Tang, Yuting Zhang, Kaishen Yuan, Bin Huang, Zitong Yu, et al. Gpt as psychologist? preliminary evaluations for gpt-4v on visual affective computing. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) workshop, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[27] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Minghui Qiu, Pengcheng Lu, Tao Wang, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[28] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[29] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[30] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[31] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[32] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3170-3180, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[34] Ambareesh Revanur, Zhihua Li, Umur A Ciftci, Lijun Yin, and László A Jeni. The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2760-2767, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[35] Eduardo Rocha Rodrigues, Igor Oliveira, Renato Cunha, and Marco Netto. Deepdownscale: A deep learning strategy for high-resolution weather forecast. In 2018 IEEE 14th Interna-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "tional Conference on e-Science (e-Science), pages 415-422. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[36] Sven Schellenberger, Kilin Shi, Tobias Steigleder, Anke Malessa, Fabian Michler, Laura Hameyer, Nina Neumann, Fabian Lurz, Robert Weigel, Christoph Ostgathe, et al. A dataset of clinically recorded radar vital signs with synchronised reference sensor signals. Scientific data, 7(1):291, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[37] Sebastian Scher and Gabriele Messori. Predicting weather forecast uncertainty with machine learning. Quarterly Journal of the Royal Meteorological Society, 144(717):2830-2841, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[38] Casper Kaae Sønderby, Lasse Espeholt, Jonathan Heek, Mostafa Dehghani, Avital Oliver, Tim Salimans, Shreya Agrawal, Jason Hickey, and Nal Kalchbrenner. Metnet: A neural weather model for precipitation forecasting. arXiv preprint arXiv:2003.12140, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[39] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[40] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In The 23rd IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062. IEEE, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[41] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[42] Jingyuan Wang, Jiawei Jiang, Wenjun Jiang, Chengkai Han, and Wayne Xin Zhao. Towards efficient and comprehensive urban spatial-temporal prediction: A unified library and performance benchmark. arXiv preprint arXiv:2304.14343, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[43] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 1-7. IEEE, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[44] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[45] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. Simper: Simple self-supervised learning of periodic targets. arXiv preprint arXiv:2210.03115, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[46] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[47] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[48] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29246" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[49] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[50] Yunhua Zhang, Ling Shao, and Cees GM Snoek. Repetitive activity counting by sight and sound. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14070-14079, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[51] Yuting Zhang, Hao Lu, Xin Liu, Yingcong Chen, and Kaishun Wu. Advancing generalizable remote physiological measurement through the integration of explicit and implicit prior knowledge. arXiv preprint arXiv:2403.06947, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[52] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "29247" + } + ] +] \ No newline at end of file diff --git a/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/96cc4bcd-798e-44b6-b645-30ba105eaf18_origin.pdf b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/96cc4bcd-798e-44b6-b645-30ba105eaf18_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c2b1d134cbfc12a597343e1033561c3ca480aea7 --- /dev/null +++ b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/96cc4bcd-798e-44b6-b645-30ba105eaf18_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8765c3e7bd77e0b0294290483d65f43819461955c4343de39193b8ece3193d86 +size 731339 diff --git a/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/full.md b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..41c5f89c3baf1370d096354d45ee4c563a33275d --- /dev/null +++ b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/full.md @@ -0,0 +1,337 @@ +# Period-LLM: Extending the Periodic Capability of Multimodal Large Language Model + +Yuting Zhang $^{1*}$ , Hao Lu $^{1,2*}$ , Qingyong Hu $^{2}$ , Yin Wang $^{3}$ , Kaishen Yuan $^{1}$ , Xin Liu $^{4}$ , Kaishun Wu $^{1\ddagger}$ $^{1}$ The Hong Kong University of Science & Technology (Guangzhou), + +2The Hong Kong University of Science & Technology, 3Zhejiang University, + +4Lappeenranta-Lahti University of Technology + +{yzhang430, hlu585} @connect.hkust-gz.edu.cn, wuks@hkust-gz.edu.cn + +* Equal contribution, ‡ Corresponding author + +![](images/0a9a47b60acf3fb554029b50e781f9180309e7821008cb4b94e099bc7e07ab12.jpg) +Periodic Tasks +Motion Counting + +![](images/a24ad281ec5bc2e477d77288337603d0d2aacfb7c243962205e1c775b155a0f7.jpg) +Traffic Flow + +![](images/ea43dcacdd53b0da69920837766f8c6a28da1340f60335648bc25422d6e1c13e.jpg) +Weather Forecast + +![](images/a7f10825e9bf79df9af5768f8c7685ff90f8023a6823233f934edeebe96162db.jpg) +Remote Health Care + +![](images/9a314f038d8ce41ec97dcd706564b9c6f2b70557687ee0dc65ce2b2c83f7101a.jpg) + +How many times did the person in the video repeat running on the treadmill? + +![](images/428f4be7642d557252d21d318953c655a9c869e02cc4801bb4fedfb2cff19827.jpg) +GPT-4 + +The analysis did not detect any clear repetitions in the video based on the motion patterns. Hard to analyse + +![](images/03d370ec24e3d364cb62d7e1765572687dd8dd85b5173f914ca792531aff53f2.jpg) +Period-LLM +Figure 1. Existing multimodal models may fail to analyze periodic tasks, such as motion counting, traffic flow, weather forecasting, and remote health care. For example, GPT-4 fails to detect clear repetitions due to the difficulty in analyzing motion patterns, and VideoLLaMA incorrectly counts the repetitions. In contrast, Period-LLM accurately analyzes the actions and provides correct repetition count. + +The person in the video repeated running 5 times. $\times$ Wrong number + +The repetitive motion in the video is running on the treadmill 14 times. + +# Abstract + +Periodic or quasi-periodic phenomena reveal intrinsic characteristics in various natural processes, such as weather patterns, movement behaviors, traffic flows, and biological signals. Given that these phenomena span multiple modalities, the capabilities of Multimodal Large Language Models (MLLMs) offer promising potential to effectively capture and understand their complex nature. However, current MLLMs struggle with periodic tasks due to limitations in: 1) lack of temporal modelling and 2) conflict between short and long periods. This paper introduces Period-LLM, a multimodal large language model designed to enhance the performance of periodic tasks across various modalities, and constructs a benchmark of various difficulty for evaluating the cross-modal periodic capabilities of large models. Specially, We adopt an "Easy to Hard Generalization" paradigm, starting with relatively simple text-based tasks and progressing to more complex visual and multimodal tasks, ensuring that the model gradually builds robust periodic reasoning capabilities. Additionally, we propose a "Resisting Logical Oblivion" optimization strategy to maintain periodic reasoning abilities dur + +ing semantic alignment. Extensive experiments demonstrate the superiority of the proposed Period-LLM over existing MLLMs in periodic tasks. The code is available at https: //github.com/keke-nice/Period-LLM. + +# 1. Introduction + +Periodic tasks refer to activities performed at fixed or predictable intervals, encompassing a wide range of applications. At a macro level, this includes repetitive motions in human daily life, such as step counts and the number of times one jumps rope [7, 10, 50], as well as the periodic analysis of meteorological data, where scientists can identify various weather patterns and trends, forecast future weather conditions [1, 35, 37], and issue timely warnings. At a micro level, in health monitoring, by capturing and analyzing subtle chromatic changes in facial videos, physiological signals such as heart rate, blood pressure, and respiratory rate can be extracted [22, 25, 45, 51]. The periodic fluctuations of these signals reflect the physiological state of human. Effectively handling periodic tasks can significantly enhance various aspects of human life, from predicting weather patterns to monitoring health conditions. + +Given that periodic phenomena occur across various modalities, it becomes essential to utilize multimodal models for period analysis instead of single modality. In recent years, multimodal large language models [3, 16, 18, 48] have advanced the progress of visual and linguistic learning by incorporating visual encoders into various pre-trained large language models. Furthermore, some studies have extended vision to other modalities such as audio [9, 48], depth imagery [8], and near-infrared [8]. With the promising progress of LLM-enabled multimodal models, a natural question arises: Can multimodal large language models understand periodic tasks? + +Unfortunately, existing large models have shown inadequacies in counting capabilities or understanding "repetitive" temporal sequences [15], creating obstacles for the application of periodic tasks in multimodal large language models, as shown in Fig. 1. These shortcomings mainly stem from the following challenges: 1) Interference of spatial pseudo-temporal information. The leakage of numerical information in visuals may lead models to take shortcuts rather than learning the true periodic information. When specific numbers appear in videos, models may tend to directly answer with these numbers, or be misled by them, rather than deriving answers through the analysis of periodic information [32]. This limits the models' generalization capabilities, rendering them unable to handle periodic tasks without direct numerical information. 2) Conflict between long-time periodic reasoning ability and short-time semantic alignment. The recognition of a periodic task lies not only in the ability to deduce the number of periodic repeats, but also in the accurate recognition of the semantic information in the video. Both require the model to have the ability to model the timing, but with different focus. Too much focus on the short-time semantic understanding may lead to ignorance of periodic reasoning ability. 3) Absence of captions is a significant issue. In the training corpora of MLLMs, most sentences are descriptive and rarely contain specific descriptions of the number of actions. This poses difficulties for models when dealing with periodic tasks that require counting the number of actions. For example, a model might respond with "The boy in the video is doing pull-ups" or "The boy in the video does multiple pull-ups", without a specific counting supervision. + +To avoid overfitting to modality-specific features while preserving temporal modeling capabilities, we propose a MLLM with periodic capabilities, termed Period-LLM. Specifically, we employ an easy-to-hard generalization paradigm. Initially, we define text-based, macro, and micro periodic tasks, ranging from easy to hard. We anticipate starting with text and progressively advancing, enabling Large Language Models (LLMs) to learn periodicity across different modalities. The rationale for this definition is twofold: 1) Despite modality alignment, LLMs still ex + +hibit superior text processing capabilities compared to other modalities; 2) Macro periodic tasks contain clear semantic features, making their periodic nature easier to capture than that of micro tasks. Accordingly, we selected periodic tasks corresponding to different levels of difficulty. At the textual level, we construct a "repeated word question-answer" dataset to pre-train LLMs, deepening their understanding of "repetitiveness" from a textual perspective. At the macro level, we select the repetitive motion dataset Countix [7], which aims to predict the number of repetitive movements in videos. At the micro level, we choose the rPPG task, which involves capturing subtle periodic or quasi-periodic physiological signals from facial videos or radio-frequency signals. Furthermore, it is believed that a good periodic model should have both the ability of periodic reasoning and the ability of semantic understanding. Focusing on the optimization of semantic alignment may, to some extent, lead to forgetting the logical reasoning abilities learned from simple tasks. Therefore, we propose a Resisting Logical Oblivion optimization strategy. Specifically, by manually designing a method to dynamically update the backpropagation gradients, the model can adjust previously overlooked feature dimensions without interfering with existing ones, providing a guided approach to model optimization. A series of experiments demonstrate the superiority of the proposed Period-LLM over existing MLLMs in periodic tasks. This advantage is not limited to text and video; we further explore the potential of other modalities based on this foundation. + +Our contributions are summarized as follows: + +- To the best of our knowledge, we propose the first MLLM-based periodic model (Period-LLM) and construct a benchmark featuring periodic tasks of varying difficulty, including text, video, and other modalities. +- We propose an easy-to-hard training paradigm that progresses from simple to complex tasks, enabling the model to sequentially acquire the ability to perceive periodic information across different modalities, starting with text, then vision, and extending to other modalities. +- We propose a Resisting Logical Oblivion (RLO) optimization strategy to resolve the conflict between periodic reasoning and semantic alignment. +- Extensive experiments demonstrate that the proposed Period-LLM outperforms existing MLLMs in periodic tasks across various modalities. + +# 2. Related Work + +# 2.1. Multimodal Large Language Model + +In recent studies, large models have exhibited reliable capabilities in the field of video understanding. In addition to the powerful vision-language large models [12, 17, 26, 52], there is an increasing focus on exploring more modalities in recent research [13, 27-29, 46]. Bain et al. [5] intro + +![](images/1eb310121921262fee121006cd41059657bf8ffc434340acfbb717fb61019be5.jpg) +Figure 2. The process of the generation of Question-and-answer datasets. Samples of text and video are taken as examples. In the text sample, questions and the occurrences of specific words are randomly generated, while in the video sample, the questions are randomly generated and the keywords represent labels within the dataset. These elements are then incorporated into a GPT-4 prompt to construct answers that include key information. The right side of the image shows examples of the final Q&A pairs generated, where the answers are complete sentences containing the critical information from the key elements. + +duce a large-scale dataset, which provides general descriptions associated with video content. Several LLM-based works [13, 27, 29, 46] are proposed to comprehend the visual content of the video. Besides, Video-LLaMa [47] extends the ability to comprehend both auditory and visual information, while Su et al. [41] leverages the power of multimodal encoders to understand across six modalities. Recently, Lv et al. [28] proposed video-based large language models for the task of VAD in a weakly supervised framework. Different from them, our work focuses on the periodic analysis of MLLMs. + +# 2.2. Periodic Tasks + +Examples of periodic learning include recovering and magnifying physiological signals (e.g., heart rate or breathing) [25, 31], predicting weather and environmental changes (e.g., nowcasting of precipitation or land surface temperatures) [38], counting motions that are repetitious (e.g., exercises or therapies) [4, 7], and analyzing human behavior (e.g., gait) [23]. Prior work has focused on designing customized neural architectures [7, 20], loss functions [39], and leveraging relevant learning paradigms including transfer learning [24] and meta-learning [21] for periodic learning. This paper attempts to study the capability of periodic tasks in multimodal models. + +# 3. Problem Setting and Preliminaries + +In nature, periodic phenomena contain both periodic and semantic information, which can be modeled by $x = K \cdot p(\omega t) + N \cdot s(t)$ , where $p(\omega t)$ is a periodic function with angular frequency $\omega = 2\pi / T$ and period $T$ . Here, $K$ represents the amplitude of the periodic component, $N$ denotes the amplitude of the semantic component, and $s(t)$ is a semantic function related to time $t$ . When discretizing continuous sequences as model inputs $X$ , we adopt a sampling interval $\Delta t = T / n$ , where $n$ is a positive integer multiple + +of the sampling rate $(n = 1,2,3,\ldots)$ . Ideally, when $n = 1$ and semantics are absent, the value at each sampling point equals the constant $K$ , resulting in $x = K$ . This can be mathematically expressed as: + +$$ +X = \left\{ \begin{array}{l l} K \cdot p (\omega t) + N \cdot s (t), & \Delta t = \frac {T}{n}, n = 2, 3, 4, \dots \\ K, & \Delta t = T \end{array} \right. \tag {1} +$$ + +However, in practical applications, data is often high-dimensional and noisy. For instance, after processing a text segment through a tokenizer, we obtain a matrix with dimensions $[L, \mathrm{Dim}]$ , where $L$ represents the length of the text (i.e., the number of tokens), and Dim denotes the embedding dimension of each token. In cases of repetitive text such as "period period period", certain dimensions of the encoded feature matrix may exhibit constant values similar to $x = K$ , where $K$ is a vector representing the embedding values of these tokens in specific dimensions. + +Inputs from other modalities, such as video, present more complex scenarios. Periodic movements may be obscured by a large amount of semantic or redundant information. In macro-periodic tasks, periodic movement is more prominent, allowing models to detect periodicity more easily; however, in micro-periodic tasks, the amplitude of the periodic signal $(K)$ may be small and masked by noise, making it challenging for models to capture these subtle periodic variations. Despite these complexities, research suggests that models can transfer knowledge from simpler tasks to more complex ones. In this context, repetitive text can be seen as a basic periodic task, while macro-periodic and micro-periodic video tasks represent more complex periodic challenges. Although "repetition" is the common thread across these tasks, the semantic content and amplitude of the periodic signal vary. + +Given that large language models (LLMs) have been trained on extensive and diverse text corpora, they are par + +![](images/291a3997c6228d68dc03a276bd5fe6ca42178f63429407d9785edc71208665d8.jpg) +Figure 3. The pipeline of training paradigm from easy to hard. The model begins training with simpler periodic tasks and gradually transitions to more complex ones. Throughout this process, Period-LLM's understanding of periodic tasks progressively strengthens and becomes more profound. + +ticularly well-suited for processing text tasks. We propose the following framework: + +$$ +A = F (X, Q) \tag {2} +$$ + +This framework accommodates inputs $X$ from various modalities, where $X$ represents periodic data from any modality, $Q$ represents a given question, and the model computes the answer $A$ . We suggest beginning model training with simple repetitive text tasks and gradually progressing to macro-periodic and micro-periodic video tasks. This progressive training approach helps the model develop a comprehensive understanding of periodic phenomena, moving from simple to complex tasks, thereby enhancing its ability to handle periodic challenges effectively. + +# 4. Method + +In this paper, we aim to enhance the understanding of large models regarding periodic tasks. Given the current absence of specialized question-and-answer datasets for periodic tasks and corresponding testing standards, we will undertake pioneering work in the content that follows. Specifically, we firstly elucidate the instruction generation for periodic tasks. Then, we explore the generalization process from simpler tasks to more complex ones. Additionally, we propose an optimization strategy to address the challenge of preserving periodic reasoning abilities during the semantic alignment phase of Multi-Modal Generalization. + +# 4.1. Instruction Generation for Periodic Tasks + +# 4.1.1. Text-Only Instructions + +To enhance our dataset's diversity, we use GPT-4 to help create a repetitive text dataset through several key steps: First, we define the core question: "How many times is the word $\{my\text{word}\}$ repeated in the string $\{my\text{string}\} ?$ We then use GPT-4 to generate ten semantically similar questions, $q_{i}\in \{q_{1},q_{2},\ldots ,q_{10}\}$ . For word diversity, we randomly select a word from the GPT-4 technical report [2] as $\{my\text{word}\}$ . To further diversify, we set a repetition count $n$ within $\{2,3,\dots,20\}$ , constructing $\{my\text{string}\}$ as " $\{\mathrm{selected~word}\} \times n$ . Since GPT-4 can't directly calculate word repetitions, we guide it with precise prompts. We + +instruct GPT-4: "Construct an answer based on the question format, selected word, and repetition count." During this process, we randomly choose a question $q_{i}$ and provide the necessary answer elements, {selected word, $n$ times}, enabling GPT-4 to generate a fitting response. By following these steps, we effectively use GPT-4's generative capabilities to build a diverse and accurate repetitive text dataset, which supports robust model training and evaluation. + +# 4.1.2. Multi-Modal Instructions + +For tasks involving periodic data across various modalities like video, traffic streams, and health monitoring, we employ a method that integrates annotations, raw captions, and frequency information to generate comprehensive representations. This approach is applicable to diverse datasets, addressing challenges such as dynamic environments, variations in periodicity and counts, and action speeds. We transform data into a question-and-answer format inspired by repetitive text dataset construction. We generate unified captions for each segment by combining existing annotations and frequency data. The main question posed is: "What is the total number of repetitive actions in this data?" Using GPT-4's capabilities, we craft multiple semantically similar questions from this core inquiry. Explicit instructions and answer elements are provided to generate question-and-answer pairs, enhancing dataset representation, the process as shown in Fig. 2. This method enriches model understanding and processing of periodic tasks across modalities. We establish a link between repetitive text and periodic data, enabling generalization from simple to complex tasks. The process first involves self-optimization in relatively simple individual tasks, followed by fine-tuning on more complex tasks, as illustrated in Fig. 3. + +# 4.2. Easy-to-Hard Generalization + +In this section, we explore the generalization process from simpler tasks to more complex ones, structured around the problem definition $A = F(f, Q)$ , where $A$ represents the answer, $f$ represents the features derived from the input data, and $Q$ represents the question. + +![](images/467885a30330a0b59d0a855988a9252b72c8f8554f1a2bb1673d47b5cbb68811.jpg) +Figure 4. The Period-LLM framework, using video as an example, integrates both text and video inputs. The video is processed through a visual encoder and a visual projector, then concatenated with text features derived from a word embedding layer. This combined input is fed into a large language model (LLM). Gradient optimization strategies are applied to the output feature channels of the LLM, specifically targeting those channels that failed to capture useful information, thereby enriching the semantic content in these channels. + +# 4.2.1. Text-Only Generalization + +The first stage of generalization focuses on text-based tasks, where the goal is to learn logical reasoning without needing to align other modalities with text. This can be mathematically represented as: + +$$ +A = F \left(T _ {f}, Q\right), \tag {3} +$$ + +here, $T_{f}$ represents the features derived from textual input. The emphasis is on developing the model's logical reasoning capabilities, as the input modality is solely text, which simplifies the task by eliminating the need to handle multimodal alignment. The model learns to process and understand the logical structure of questions and generate appropriate answers purely based on textual information. + +# 4.2.2. Multi-Modal Generalization + +The next stage involves more complex tasks that require the integration of multiple modalities, such as video and text. This stage not only demands logical reasoning but also the alignment of semantic information across different modalities. This process can be described by: + +$$ +A = F \left(M _ {f}, Q\right), \tag {4} +$$ + +in this case, $M_{f}$ represents features derived from a combination of inputs, including video and text. The challenge is twofold: the model must align semantic content across different modalities while simultaneously developing robust periodic logical reasoning capabilities. This complex task requires the model to integrate semantic features with textual information and establish accurate connections between them in order to answer questions correctly. + +# 4.3. Resisting Logical Oblivion + +Due to the significant differences between periodic logical reasoning abilities and the ability to understand semantic information in videos, we design a Resisting Logical Oblivion + +(RLO) optimization strategy to prevent the new knowledge learned in a new round of optimization from overwriting the reasoning knowledge acquired from simpler tasks. Vanilla gradient descent leads to an undifferentiated update of the entire feature space. If $\theta = (\theta_{1},\theta_{2},\dots,\theta_{J})\in \mathbb{R}^{C\times W}$ is the parameter of the model, where $W$ is related to the network structure, $\frac{\partial L(\theta)}{\partial\theta_j}$ $(j = 1,2,\ldots ,J)$ represents the partial derivative of $\theta$ . The process of model updating is as follows: + +$$ +\nabla \theta_ {j} = - \alpha \frac {\partial L (\theta)}{\partial \theta_ {j}}, \tag {5} +$$ + +$$ +\theta_ {j} ^ {\prime} \leftarrow \theta_ {j} + \nabla \theta_ {j}, \tag {6} +$$ + +where $\alpha$ is the learning rate, $\nabla \theta_{j}$ is the updating vector, and $\theta_j^\prime$ is the new weight for the next iteration. + +Traditional gradient update strategies adjust model parameters based on global errors, which, although effective, can lead to knowledge interference between different types of tasks. Our new strategy introduces a feature channel weight allocation mechanism, dynamically adjusting the gradient update weights of different feature channels. Specifically, we introduce the weight function $\Omega (k,c)$ implementation: + +$$ +\Omega \left(c _ {i}\right) = \left\{ \begin{array}{l l} 1 + \beta \cdot e ^ {\frac {\text {i t e r n u m}}{\text {m a x i t e r}}}, & \bar {c} _ {i} < \bar {c}, \\ 1, & \bar {c} _ {i} > \bar {c}, \end{array} \right. \tag {7} +$$ + +where $\beta$ is a hyper-parameter, $iter_{num}$ denotes the current iteration count, and $max_{iter}$ represents the maximum number of iterations. $\bar{c}$ is the average across all feature channels, while $\bar{c}_i$ refers to the average of the $i$ th feature channel. By dynamically weighting feature channels, we can exert control over the real-time updates of specific feature channels. This ensures that redundant feature channels can learn newly acquired semantic information without + +![](images/e431d41866c94feb883cb1d08d332b2731d099067594816b04e1c42d30587032.jpg) +Figure 5. Details of the function of the gradient weight function. The left side of the image shows that some channels have lower backpropagation gradients, while others have higher ones. Higher backpropagation gradients indicate that these feature channels are being updated more significantly. + +compromising existing reasoning abilities, thereby balancing the enhancement of reasoning abilities and video semantic understanding. As shown in Fig. 5, our gradient weighting function is applied to the backpropagation gradients, boosting those originally lower gradients, thereby substantially increasing the update of channels that were previously updated little or not at all. This process ensures that newly learned semantic information is directed to these newly emphasized feature channels for learning. Therefore, the model parameter update process with the introduction of the RLO optimization strategy is as follows, showing the changes from Eq. 5 to Eq. 8: + +$$ +\nabla \theta_ {j} ^ {*} = \Omega \left(c _ {i}\right) \cdot \nabla \theta_ {j}, \tag {8} +$$ + +$$ +\theta_ {j} ^ {\prime} \leftarrow \theta_ {j} + \nabla \theta_ {j} ^ {*}, \tag {9} +$$ + +During the training process, the input data $\mathcal{Z}$ consists of input videos and their corresponding questions $x_{i}$ , correct answers $y_{i}$ , where $i = 1,2,3,\ldots ,K$ , and each $x_{i},y_{i}$ is a sequence of tokens. Large language models are denoted by $P_{\phi}(y\mid x)$ , where $\phi$ represents model parameters. Typically, when fine-tuning large models, only the question-answer pairs $(x_{i},y_{i})$ are input. The goal is to adjust the parameters $\phi$ to maximize the following expression: + +$$ +\max _ {\phi} \sum_ {(x, y) \in \mathcal {Z}} \sum_ {t = 1} ^ {| y |} \log \left(P _ {\phi} \left(y _ {t} \mid x, y _ {< t}\right)\right), \tag {10} +$$ + +where $|y|$ represents the length of the input vector $y$ , $P_{\phi}(y_t|x, y_{MethodLLMZero-shotFine-tuneCountix-QArPPG-QACountix-QArPPG-QAMAERMSEBleu1CIDErMETEORMAERMSEMAERMSEBleu1CIDErMETEORMAERMSEFrozenBiLM [44]DeBERTa-V27.2211.530.1450.5500.11220.4321.565.229.440.2330.6100.20119.2120.46VideoLLaMA [48]Vicuna-7B6.1810.820.1870.3600.24518.8919.434.988.890.3890.5700.35818.2919.01LLaMA-Adapter [49]LLaMA-7B7.4410.510.2440.6830.31221.7823.345.7810.020.3780.7890.39720.5422.84VideoChat [14]Vicuna-7B6.7311.230.1650.5490.27719.5420.675.039.530.2750.6790.38818.3210.77Video-ChatGPT [30]Vicuna-7B6.4410.750.2030.5890.29718.1220.454.649.150.3560.6430.36717.5419.45BT-Adapter [19]Vicuna-7B7.7311.700.2760.6740.34119.3722.686.129.700.2960.6890.35318.6721.49LLaMA-VID [16]Vicuna-7B6.9711.210.2670.6800.30118.3118.565.349.850.3880.7830.41117.5117.60Period-LLMLLaMA-7B5.518.330.2880.7020.30815.0817.133.777.140.3970.8100.41513.7816.83 + +Table 1. Comparison of our method with other leading methods on two newly introduced video question-and-answer (QA) datasets, with all methods evaluated at a resolution of ${224} \times {224}$ . The results for Period-LLM in the zero-shot setting are derived from a model pretrained on text-QA dataset. For rPPG-QA, the V4V dataset is used as the validation set. + +
MethodLLMZero-shotFine-tune
RotNIST-QADrive-QARadar-QARotNIST-QADrive-QARadar-QA
MAERMSEMAERMSEMAERMSEMAERMSEMAERMSEMAERMSE
FrozenBiLM [44]DeBERTa-V24.264.7830.88120.5622.1730.422.883.2529.44110.5616.6628.55
VideoLLaMA [48]Vicuna-7B4.085.3235.45130.6723.4218.852.513.8931.22121.5517.0320.34
LLaMA-Adapter [49]LLaMA-7B3.954.4231.26125.4224.1231.792.663.2330.28118.7920.3428.56
VideoChat [14]Vicuna-7B3.564.2132.67126.8821.7129.582.233.0630.12115.2320.5427.38
Video-ChatGPT [30]Vicuna-7B3.444.8738.18118.2923.7732.132.012.2933.28112.3521.6130.45
BT-Adapter [19]Vicuna-7B4.035.4634.21110.2423.4231.472.783.5631.31107.6522.1230.19
LLaMA-VID [16]Vicuna-7B3.874.5734.98114.3318.2129.322.433.2532.45109.8718.2129.32
Period-LLMLLaMA-7B3.203.9630.88108.7417.5626.451.502.7528.71105.5414.2424.12
+ +Table 2. Comparison of our method with other leading methods on the RotNIST-QA, Drive-QA, and Radar-QA datasets. Drive-QA is a dataset of taxi trajectory data, while Radar-QA is a radar modality dataset capturing physiological signals. The zero-shot results for Period-LLM are based on a model pretrained on text-QA, Countix-QA, and rPPG-QA. + +periodic reasoning abilities. The results show significant improvements over zero-shot, thanks to the inclusion of explicit periodic information in the training data. Period-LLM surpasses existing MLLMs by over 1 MAE on the Countix-QA dataset, while also showing significant improvements in generated text quality, as measured by Bleu-1, CIDEr, and METEOR scores. This improvement is driven by the easy-to-hard fine-tuning paradigm and the Resisting-Logical-Oblivion strategy, which together help the model preserve and enhance its periodic reasoning abilities, while also maintaining strong semantic understanding, resulting in superior performance. + +In addition to the visual periodic tasks, Table 2 presents the performance of these models on other modalities, specifically the RotNIST-QA, Drive-QA, and Radar-QA datasets. The zero-shot and fine-tune protocols follow the same approach as in Table 1, with the key difference being that experiments on these three datasets were conducted in parallel. The pretraining models used in these experiments were fine-tuned on the two visual periodic tasks from Table 1. On the RotNIST-QA dataset, Period-LLM achieves the lowest MAE (1.50) and RMSE (2.75), outperforming models such as FrozenBiLM, VideoLLaMA, and LLaMA-Adapter. For the Drive-QA dataset, Period-LLM demonstrates superior performance with the lowest MAE (28.71) and competitive RMSE (105.54). On the Radar-QA dataset, Period-LLM also performs strongly, further validating the effectiveness of our model across different modali + +ties. These results underscore Period-LLM's enhanced accuracy in periodic reasoning tasks and its robustness in handling diverse datasets, establishing it as a versatile solution for periodic question-answering across a range of domains, including health, weather, and traffic. + +# 5.4. Qualitative Results + +QA Examples. Fig. 6 showcases examples from the countix-QA and rPPG-QA datasets. It can be seen that Period-LLM is capable of understanding the semantic information contained within the video as well as recognizing periodicity within the video. For instance, in the countix-QA dataset, Period-LLM accurately provides the number of repetitive movements, while in the rPPG-QA dataset, it can determine the heart rate of the man in the video. These capabilities are beyond what other language models can achieve. According to Fig. 6, in the countix-QA dataset, Period-LLM identifies that the man in the video performs two pull-ups. In the rPPG-QA dataset, Period-LLM is able to detect that the man's heart rate is 75 beats per minute. This demonstrates Period-LLM's exceptional performance in handling video content and periodic information. More QA examples are provided in the appendix. + +Comparison of Loss Reduction. The Fig. 7 illustrates the loss reduction comparison between the Baseline (Vanilla Fine-Tuning on Countix-QA) and the proposed Period-LLM approach (with text-QA pretraining and RLO) on the Countix-QA dataset. As shown, our pro + +Table 3. The MAE value on Countix-QA with different parameter or setting choices in $\Omega \left( {c}_{i}\right)$ of RLO optimization strategy, and ablation study on Countix-QA showing the impact of easy tasks (period-text-QA) and RLO (Resisting Logical Oblivion). Default settings are marked in gray . + +(a) Impact of $\beta$ + +
βMAE
0.013.85
0.053.77
0.13.90
0.54.05
+ +(b) Type of Threshold. + +
ThresholdMAE
Mean3.77
Median3.85
Learnable4.20
+ +(c) Results on Countix-QA w./w.o Easy Task and RLO. + +
Easy to HardRLOCountix-QA
Text-QACountix-QAMAECIDEr
✗✓✗4.300.661
✓✓✗3.890.782
✓✓✓3.770.810
+ +(d) Results on rPPG-QA w./w.o Easy Tasks and RLO. + +
Easy to HardRLOrPPG-QA
Text-QACountix-QArPPG-QAMAE
XX✓X15.45
✓X✓X14.46
✓✓✓X14.05
✓✓✓✓13.78
+ +![](images/bad20fc015a3ae26c5ecfa6dd9ea548082722e2a8b53c41bc4c4ee52faf51f09.jpg) +Figure 6. Examples from Countix-QA and rPPG-QA. + +![](images/5778448ea4e77748d3c7a1285ccf9f5e5ebf498cce1f6de4d92976a53225d743.jpg) +Figure 7. Comparison of Loss Reduction Between Baseline (Vanilla Fine-Tuning on Countix-QA) and Proposed Method (Period-LLM with Text-QA Pretraining and RLO, Fine-Tuned on Countix-QA). + +posed method demonstrates a significantly faster decline in the loss values compared to the baseline. The Period-LLM model achieves a steeper descent in the initial training stages, indicating a more efficient learning process. Furthermore, it converges to a lower loss value in fewer epochs, suggesting improved optimization and stability. This faster convergence implies that our method better captures the underlying data patterns, leading to more effective fine-tuning on the Countix-QA. + +# 5.5. Ablation Study + +The Setting of Weight Function $\Omega(c_i)$ of RLO Optimization Strategy. As seen in Table 3a, $\beta = 0.05$ yields the lowest MAE (3.77), indicating an effective balance between retaining prior knowledge and integrating new information. When $\beta$ is too small (e.g., 0.01), the model's anti-forgetting ability is weak, causing higher error as new knowledge overwhelms older knowledge. Conversely, a high $\beta$ (e.g., 0.5) overly prioritizes old knowledge, limiting adaptation. Thus, $\beta = 0.05$ achieves the best balance for minimizing error. In Table 3b, setting the threshold to the mean results in the lowest MAE (3.77), effectively balancing old and new knowledge contributions. The median threshold performs slightly worse (MAE 3.85) due to its insensitivity to outliers, while a learnable threshold increases MAE + +(4.20), likely due to instability. Hence, the mean threshold is the most stable and effective choice. + +Results of various combinations of Preiod-LLM Components. We evaluate our approach on the Countix-QA and rPPG-QA datasets, as shown in Tables 3c and 3d. For Countix-QA, our method—combining pre-training on simpler tasks with the Resisting Logical Oblivion (RLO) strategy—significantly improves performance, reducing the Mean Absolute Error (MAE) from 4.30 to 3.77 and increasing the CIDEr score from 0.661 to 0.810. Similarly, on the rPPG-QA dataset, this approach lowers the MAE from 15.45 to 13.78. These improvements reflect the model's enhanced capacity to capture periodic patterns effectively while also understanding semantic content more deeply. The decrease in MAE shows better accuracy in numerical predictions, while the CIDEr increase suggests improved semantic alignment in question-answering tasks. Together, these results confirm that the integration of easy-task pretraining with RLO enables the model to balance periodicity and semantic understanding, leading to faster convergence and more robust performance across QA tasks. + +# 6. Conclusion + +In this paper, we introduce Period-LLM, a large multimodal model specifically designed to enhance performance on periodic tasks across various modalities. The model follows a progressive training paradigm, starting with simpler text-based tasks and advancing to more complex visual and multimodal tasks. Additionally, we propose the "Resisting Logical Oblivion" optimization strategy to preserve the model's periodic reasoning abilities during semantic alignment. The Our experimental results demonstrate that Period-LLM significantly outperforms existing models in handling periodic tasks, especially in challenging domains such as repetitive motion analysis and physiological signal estimation. Period-LLM not only achieves superior accuracy but also highlights its potential for future advancements in managing complex periodic phenomena across different modalities. Future work will focus on further refining model's capabilities, particularly in addressing more nuanced and less structured periodic tasks, and expanding its application to additional domains where periodicity plays a crucial role. + +# 7. Acknowledgment + +This work was supported by the Guangdong Provincial Key Laboratory of Integrated Communication, Sensing and Computation for Ubiquitous Internet of Things (No. 2023B1212010007), China NSFC Grant 62472366, the Project of DEGP (No.2024GCZX003, 2023KCXTD042), 111 Center (No.D25008) and Shenzhen Science and Technology Foundation (ZDSYS20190902092853047). + +# References + +[1] Kumar Abhishek, Maheshwari Prasad Singh, Saswata Ghosh, and Abhishek Anand. Weather forecasting model using artificial neural network. Procedia Technology, 4:311-318, 2012. 1 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2 +[4] Mohammad Rafayet Ali, Javier Hernandez, E Ray Dorsey, Ehsan Hoque, and Daniel McDuff. Spatio-temporal attention and magnification for classification of parkinson's disease from videos collected via the internet. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 207-214. IEEE, 2020. 3 +[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2 +[6] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 6 +[7] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Counting out time: Class agnostic video repetition counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10387-10396, 2020. 1, 2, 3, 6 +[8] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 2 +[9] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xiangyu Yue. Onellm: One framework to align all modalities with language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26584-26595, 2024. 2 + +[10] Huazhang Hu, Sixun Dong, Yiqun Zhao, Dongze Lian, Zhengxin Li, and Shenghua Gao. Transrac: Encoding multiscale temporal correlation with transformers for repetitive action counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19013-19022, 2022. 1 +[11] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6 +[12] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2 +[13] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 3 +[14] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6, 7 +[15] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2 +[16] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 2, 6, 7 +[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 6 +[19] Ruyang Liu, Chen Li, Yixiao Ge, Ying Shan, Thomas H Li, and Ge Li. One for all: Video conversation is feasible without video instruction tuning. arXiv preprint arXiv:2309.15785, 2023. 6, 7 +[20] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems, 33:19400-19411, 2020. 3 +[21] Xin Liu, Ziheng Jiang, Josh Fromm, Xuhai Xu, Shwetak Patel, and Daniel McDuff. Metaphys: few-shot adaptation for non-contact physiological measurement. In Proceedings of the conference on health, inference, and learning, pages 154-163, 2021. 3 +[22] Xin Liu, Yuting Zhang, Zitong Yu, Hao Lu, Huanjing Yue, and Jingyu Yang. rppg-mae: Self-supervised pretraining with masked autoencoders for remote physiological measurements. IEEE Transactions on Multimedia, 2024. 1 +[23] Yingcheng Liu, Guo Zhang, Christopher G Tarolli, Rumen Hristov, Stella Jensen-Roberts, Emma M Waddell, Taylor L Myers, Meghan E Pawlik, Julia M Soto, Renee M Wilson, + +et al. Monitoring gait at home with radio waves in parkinson's disease: A marker of severity, progression, and medication response. Science Translational Medicine, 14(663): eadc9669, 2022. 3 +[24] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 669–684. Springer, 2019. 3 +[25] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12404-12413, 2021. 1, 3 +[26] Hao Lu, Xuesong Niu, Jiyao Wang, Yin Wang, Qingyong Hu, Jiaqi Tang, Yuting Zhang, Kaishen Yuan, Bin Huang, Zitong Yu, et al. Gpt as psychologist? preliminary evaluations for gpt-4v on visual affective computing. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) workshop, 2024. 2 +[27] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Minghui Qiu, Pengcheng Lu, Tao Wang, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023. 2, 3 +[28] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 3 +[29] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2, 3 +[30] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 6, 7 +[31] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2019. 3, 6 +[32] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3170-3180, 2023. 2 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[34] Ambareesh Revanur, Zhihua Li, Umur A Ciftci, Lijun Yin, and László A Jeni. The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2760-2767, 2021. 6 +[35] Eduardo Rocha Rodrigues, Igor Oliveira, Renato Cunha, and Marco Netto. Deepdownscale: A deep learning strategy for high-resolution weather forecast. In 2018 IEEE 14th Interna- + +tional Conference on e-Science (e-Science), pages 415-422. IEEE, 2018. 1 +[36] Sven Schellenberger, Kilin Shi, Tobias Steigleder, Anke Malessa, Fabian Michler, Laura Hameyer, Nina Neumann, Fabian Lurz, Robert Weigel, Christoph Ostgathe, et al. A dataset of clinically recorded radar vital signs with synchronised reference sensor signals. Scientific data, 7(1):291, 2020. 6 +[37] Sebastian Scher and Gabriele Messori. Predicting weather forecast uncertainty with machine learning. Quarterly Journal of the Royal Meteorological Society, 144(717):2830-2841, 2018. 1 +[38] Casper Kaae Sønderby, Lasse Espeholt, Jonathan Heek, Mostafa Dehghani, Avital Oliver, Tim Salimans, Shreya Agrawal, Jason Hickey, and Nal Kalchbrenner. Metnet: A neural weather model for precipitation forecasting. arXiv preprint arXiv:2003.12140, 2020. 3 +[39] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3 +[40] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In The 23rd IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062. IEEE, 2014. 6 +[41] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 3 +[42] Jingyuan Wang, Jiawei Jiang, Wenjun Jiang, Chengkai Han, and Wayne Xin Zhao. Towards efficient and comprehensive urban spatial-temporal prediction: A unified library and performance benchmark. arXiv preprint arXiv:2304.14343, 2023.6 +[43] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 1-7. IEEE, 2020. 6 +[44] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 6, 7 +[45] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. Simper: Simple self-supervised learning of periodic targets. arXiv preprint arXiv:2210.03115, 2022. 1 +[46] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2, 3 +[47] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3 +[48] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 6, 7 + +[49] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 6, 7 +[50] Yunhua Zhang, Ling Shao, and Cees GM Snoek. Repetitive activity counting by sight and sound. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14070-14079, 2021. 1 +[51] Yuting Zhang, Hao Lu, Xin Liu, Yingcong Chen, and Kaishun Wu. Advancing generalizable remote physiological measurement through the integration of explicit and implicit prior knowledge. arXiv preprint arXiv:2403.06947, 2024. 1 +[52] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 \ No newline at end of file diff --git a/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/images.zip b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3820120e599b179a82a2e73e2c6ebe02abc58f19 --- /dev/null +++ b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e5233a9900aea4e6e00ee6ca962b1ebdf27e7c6cbcff6ee9b6904b4fef0b411 +size 499248 diff --git a/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/layout.json b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8f4c68ff2782d000c9de9b8c829ea4b5b2db7aec --- /dev/null +++ b/2025/Period-LLM_ Extending the Periodic Capability of Multimodal Large Language Model/layout.json @@ -0,0 +1,9130 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 103, + 463, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 103, + 463, + 140 + ], + "spans": [ + { + "bbox": [ + 149, + 103, + 463, + 140 + ], + "type": "text", + "content": "Period-LLM: Extending the Periodic Capability of Multimodal Large Language Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "spans": [ + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "text", + "content": "Yuting Zhang" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "text", + "content": ", Hao Lu" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "text", + "content": ", Qingyong Hu" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "text", + "content": ", Yin Wang" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "text", + "content": ", Kaishen Yuan" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "text", + "content": ", Xin Liu" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "text", + "content": ", Kaishun Wu" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1\\ddagger}" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 160, + 539, + 190 + ], + "type": "text", + "content": "The Hong Kong University of Science & Technology (Guangzhou)," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 190, + 490, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 190, + 490, + 204 + ], + "spans": [ + { + "bbox": [ + 121, + 190, + 490, + 204 + ], + "type": "text", + "content": "2The Hong Kong University of Science & Technology, 3Zhejiang University," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 204, + 418, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 204, + 418, + 218 + ], + "spans": [ + { + "bbox": [ + 192, + 204, + 418, + 218 + ], + "type": "text", + "content": "4Lappeenranta-Lahti University of Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 217, + 481, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 217, + 481, + 232 + ], + "spans": [ + { + "bbox": [ + 131, + 217, + 481, + 232 + ], + "type": "text", + "content": "{yzhang430, hlu585} @connect.hkust-gz.edu.cn, wuks@hkust-gz.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 195, + 232, + 412, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 232, + 412, + 246 + ], + "spans": [ + { + "bbox": [ + 195, + 232, + 412, + 246 + ], + "type": "text", + "content": "* Equal contribution, ‡ Corresponding author" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 78, + 289, + 146, + 363 + ], + "blocks": [ + { + "bbox": [ + 88, + 270, + 160, + 282 + ], + "lines": [ + { + "bbox": [ + 88, + 270, + 160, + 282 + ], + "spans": [ + { + "bbox": [ + 88, + 270, + 160, + 282 + ], + "type": "text", + "content": "Periodic Tasks" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 289, + 146, + 363 + ], + "lines": [ + { + "bbox": [ + 78, + 289, + 146, + 363 + ], + "spans": [ + { + "bbox": [ + 78, + 289, + 146, + 363 + ], + "type": "image", + "image_path": "0a9a47b60acf3fb554029b50e781f9180309e7821008cb4b94e099bc7e07ab12.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 365, + 127, + 384 + ], + "lines": [ + { + "bbox": [ + 99, + 365, + 127, + 384 + ], + "spans": [ + { + "bbox": [ + 99, + 365, + 127, + 384 + ], + "type": "text", + "content": "Motion Counting" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 147, + 289, + 211, + 363 + ], + "blocks": [ + { + "bbox": [ + 147, + 289, + 211, + 363 + ], + "lines": [ + { + "bbox": [ + 147, + 289, + 211, + 363 + ], + "spans": [ + { + "bbox": [ + 147, + 289, + 211, + 363 + ], + "type": "image", + "image_path": "a24ad281ec5bc2e477d77288337603d0d2aacfb7c243962205e1c775b155a0f7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 364, + 190, + 383 + ], + "lines": [ + { + "bbox": [ + 166, + 364, + 190, + 383 + ], + "spans": [ + { + "bbox": [ + 166, + 364, + 190, + 383 + ], + "type": "text", + "content": "Traffic Flow" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 211, + 289, + 276, + 363 + ], + "blocks": [ + { + "bbox": [ + 211, + 289, + 276, + 363 + ], + "lines": [ + { + "bbox": [ + 211, + 289, + 276, + 363 + ], + "spans": [ + { + "bbox": [ + 211, + 289, + 276, + 363 + ], + "type": "image", + "image_path": "ea43dcacdd53b0da69920837766f8c6a28da1340f60335648bc25422d6e1c13e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 364, + 261, + 380 + ], + "lines": [ + { + "bbox": [ + 228, + 364, + 261, + 380 + ], + "spans": [ + { + "bbox": [ + 228, + 364, + 261, + 380 + ], + "type": "text", + "content": "Weather Forecast" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 276, + 289, + 340, + 363 + ], + "blocks": [ + { + "bbox": [ + 276, + 289, + 340, + 363 + ], + "lines": [ + { + "bbox": [ + 276, + 289, + 340, + 363 + ], + "spans": [ + { + "bbox": [ + 276, + 289, + 340, + 363 + ], + "type": "image", + "image_path": "a7f10825e9bf79df9af5768f8c7685ff90f8023a6823233f934edeebe96162db.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 364, + 337, + 380 + ], + "lines": [ + { + "bbox": [ + 283, + 364, + 337, + 380 + ], + "spans": [ + { + "bbox": [ + 283, + 364, + 337, + 380 + ], + "type": "text", + "content": "Remote Health Care" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 374, + 251, + 513, + 280 + ], + "blocks": [ + { + "bbox": [ + 374, + 251, + 513, + 280 + ], + "lines": [ + { + "bbox": [ + 374, + 251, + 513, + 280 + ], + "spans": [ + { + "bbox": [ + 374, + 251, + 513, + 280 + ], + "type": "image", + "image_path": "9a314f038d8ce41ec97dcd706564b9c6f2b70557687ee0dc65ce2b2c83f7101a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 380, + 284, + 512, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 284, + 512, + 304 + ], + "spans": [ + { + "bbox": [ + 380, + 284, + 512, + 304 + ], + "type": "text", + "content": "How many times did the person in the video repeat running on the treadmill?" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 349, + 309, + 372, + 331 + ], + "blocks": [ + { + "bbox": [ + 349, + 309, + 372, + 331 + ], + "lines": [ + { + "bbox": [ + 349, + 309, + 372, + 331 + ], + "spans": [ + { + "bbox": [ + 349, + 309, + 372, + 331 + ], + "type": "image", + "image_path": "428f4be7642d557252d21d318953c655a9c869e02cc4801bb4fedfb2cff19827.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 328, + 367, + 333 + ], + "lines": [ + { + "bbox": [ + 353, + 328, + 367, + 333 + ], + "spans": [ + { + "bbox": [ + 353, + 328, + 367, + 333 + ], + "type": "text", + "content": "GPT-4" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 387, + 311, + 517, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 311, + 517, + 341 + ], + "spans": [ + { + "bbox": [ + 387, + 311, + 517, + 341 + ], + "type": "text", + "content": "The analysis did not detect any clear repetitions in the video based on the motion patterns. Hard to analyse" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 349, + 346, + 372, + 369 + ], + "blocks": [ + { + "bbox": [ + 349, + 346, + 372, + 369 + ], + "lines": [ + { + "bbox": [ + 349, + 346, + 372, + 369 + ], + "spans": [ + { + "bbox": [ + 349, + 346, + 372, + 369 + ], + "type": "image", + "image_path": "03d370ec24e3d364cb62d7e1765572687dd8dd85b5173f914ca792531aff53f2.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 371, + 373, + 397 + ], + "lines": [ + { + "bbox": [ + 345, + 371, + 373, + 397 + ], + "spans": [ + { + "bbox": [ + 345, + 371, + 373, + 397 + ], + "type": "text", + "content": "Period-LLM" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "type": "text", + "content": "Figure 1. Existing multimodal models may fail to analyze periodic tasks, such as motion counting, traffic flow, weather forecasting, and remote health care. For example, GPT-4 fails to detect clear repetitions due to the difficulty in analyzing motion patterns, and VideoLLaMA incorrectly counts the repetitions. In contrast, Period-LLM accurately analyzes the actions and provides correct repetition count." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 387, + 350, + 528, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 350, + 528, + 370 + ], + "spans": [ + { + "bbox": [ + 387, + 350, + 528, + 370 + ], + "type": "text", + "content": "The person in the video repeated running 5 times. " + }, + { + "bbox": [ + 387, + 350, + 528, + 370 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 387, + 350, + 528, + 370 + ], + "type": "text", + "content": " Wrong number" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 386, + 377, + 519, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 377, + 519, + 397 + ], + "spans": [ + { + "bbox": [ + 386, + 377, + 519, + 397 + ], + "type": "text", + "content": "The repetitive motion in the video is running on the treadmill 14 times." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 151, + 446, + 200, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 446, + 200, + 459 + ], + "spans": [ + { + "bbox": [ + 151, + 446, + 200, + 459 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 53, + 462, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 462, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 462, + 297, + 714 + ], + "type": "text", + "content": "Periodic or quasi-periodic phenomena reveal intrinsic characteristics in various natural processes, such as weather patterns, movement behaviors, traffic flows, and biological signals. Given that these phenomena span multiple modalities, the capabilities of Multimodal Large Language Models (MLLMs) offer promising potential to effectively capture and understand their complex nature. However, current MLLMs struggle with periodic tasks due to limitations in: 1) lack of temporal modelling and 2) conflict between short and long periods. This paper introduces Period-LLM, a multimodal large language model designed to enhance the performance of periodic tasks across various modalities, and constructs a benchmark of various difficulty for evaluating the cross-modal periodic capabilities of large models. Specially, We adopt an \"Easy to Hard Generalization\" paradigm, starting with relatively simple text-based tasks and progressing to more complex visual and multimodal tasks, ensuring that the model gradually builds robust periodic reasoning capabilities. Additionally, we propose a \"Resisting Logical Oblivion\" optimization strategy to maintain periodic reasoning abilities dur" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 448, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 495 + ], + "type": "text", + "content": "ing semantic alignment. Extensive experiments demonstrate the superiority of the proposed Period-LLM over existing MLLMs in periodic tasks. The code is available at https: //github.com/keke-nice/Period-LLM." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 499, + 395, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 499, + 395, + 511 + ], + "spans": [ + { + "bbox": [ + 314, + 499, + 395, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 521, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 714 + ], + "type": "text", + "content": "Periodic tasks refer to activities performed at fixed or predictable intervals, encompassing a wide range of applications. At a macro level, this includes repetitive motions in human daily life, such as step counts and the number of times one jumps rope [7, 10, 50], as well as the periodic analysis of meteorological data, where scientists can identify various weather patterns and trends, forecast future weather conditions [1, 35, 37], and issue timely warnings. At a micro level, in health monitoring, by capturing and analyzing subtle chromatic changes in facial videos, physiological signals such as heart rate, blood pressure, and respiratory rate can be extracted [22, 25, 45, 51]. The periodic fluctuations of these signals reflect the physiological state of human. Effectively handling periodic tasks can significantly enhance various aspects of human life, from predicting weather patterns to monitoring health conditions." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "29237" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "Given that periodic phenomena occur across various modalities, it becomes essential to utilize multimodal models for period analysis instead of single modality. In recent years, multimodal large language models [3, 16, 18, 48] have advanced the progress of visual and linguistic learning by incorporating visual encoders into various pre-trained large language models. Furthermore, some studies have extended vision to other modalities such as audio [9, 48], depth imagery [8], and near-infrared [8]. With the promising progress of LLM-enabled multimodal models, a natural question arises: Can multimodal large language models understand periodic tasks?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 219, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 590 + ], + "type": "text", + "content": "Unfortunately, existing large models have shown inadequacies in counting capabilities or understanding \"repetitive\" temporal sequences [15], creating obstacles for the application of periodic tasks in multimodal large language models, as shown in Fig. 1. These shortcomings mainly stem from the following challenges: 1) Interference of spatial pseudo-temporal information. The leakage of numerical information in visuals may lead models to take shortcuts rather than learning the true periodic information. When specific numbers appear in videos, models may tend to directly answer with these numbers, or be misled by them, rather than deriving answers through the analysis of periodic information [32]. This limits the models' generalization capabilities, rendering them unable to handle periodic tasks without direct numerical information. 2) Conflict between long-time periodic reasoning ability and short-time semantic alignment. The recognition of a periodic task lies not only in the ability to deduce the number of periodic repeats, but also in the accurate recognition of the semantic information in the video. Both require the model to have the ability to model the timing, but with different focus. Too much focus on the short-time semantic understanding may lead to ignorance of periodic reasoning ability. 3) Absence of captions is a significant issue. In the training corpora of MLLMs, most sentences are descriptive and rarely contain specific descriptions of the number of actions. This poses difficulties for models when dealing with periodic tasks that require counting the number of actions. For example, a model might respond with \"The boy in the video is doing pull-ups\" or \"The boy in the video does multiple pull-ups\", without a specific counting supervision." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "To avoid overfitting to modality-specific features while preserving temporal modeling capabilities, we propose a MLLM with periodic capabilities, termed Period-LLM. Specifically, we employ an easy-to-hard generalization paradigm. Initially, we define text-based, macro, and micro periodic tasks, ranging from easy to hard. We anticipate starting with text and progressively advancing, enabling Large Language Models (LLMs) to learn periodicity across different modalities. The rationale for this definition is twofold: 1) Despite modality alignment, LLMs still ex" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 418 + ], + "type": "text", + "content": "hibit superior text processing capabilities compared to other modalities; 2) Macro periodic tasks contain clear semantic features, making their periodic nature easier to capture than that of micro tasks. Accordingly, we selected periodic tasks corresponding to different levels of difficulty. At the textual level, we construct a \"repeated word question-answer\" dataset to pre-train LLMs, deepening their understanding of \"repetitiveness\" from a textual perspective. At the macro level, we select the repetitive motion dataset Countix [7], which aims to predict the number of repetitive movements in videos. At the micro level, we choose the rPPG task, which involves capturing subtle periodic or quasi-periodic physiological signals from facial videos or radio-frequency signals. Furthermore, it is believed that a good periodic model should have both the ability of periodic reasoning and the ability of semantic understanding. Focusing on the optimization of semantic alignment may, to some extent, lead to forgetting the logical reasoning abilities learned from simple tasks. Therefore, we propose a Resisting Logical Oblivion optimization strategy. Specifically, by manually designing a method to dynamically update the backpropagation gradients, the model can adjust previously overlooked feature dimensions without interfering with existing ones, providing a guided approach to model optimization. A series of experiments demonstrate the superiority of the proposed Period-LLM over existing MLLMs in periodic tasks. This advantage is not limited to text and video; we further explore the potential of other modalities based on this foundation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 326, + 419, + 512, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 419, + 512, + 430 + ], + "spans": [ + { + "bbox": [ + 326, + 419, + 512, + 430 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 434, + 553, + 612 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 434, + 553, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 434, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 434, + 553, + 480 + ], + "type": "text", + "content": "- To the best of our knowledge, we propose the first MLLM-based periodic model (Period-LLM) and construct a benchmark featuring periodic tasks of varying difficulty, including text, video, and other modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 482, + 553, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 553, + 540 + ], + "type": "text", + "content": "- We propose an easy-to-hard training paradigm that progresses from simple to complex tasks, enabling the model to sequentially acquire the ability to perceive periodic information across different modalities, starting with text, then vision, and extending to other modalities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 541, + 553, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 553, + 577 + ], + "type": "text", + "content": "- We propose a Resisting Logical Oblivion (RLO) optimization strategy to resolve the conflict between periodic reasoning and semantic alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 577, + 553, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 553, + 612 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that the proposed Period-LLM outperforms existing MLLMs in periodic tasks across various modalities." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 616, + 400, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 616, + 400, + 628 + ], + "spans": [ + { + "bbox": [ + 314, + 616, + 400, + 628 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 636, + 504, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 504, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 504, + 649 + ], + "type": "text", + "content": "2.1. Multimodal Large Language Model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "In recent studies, large models have exhibited reliable capabilities in the field of video understanding. In addition to the powerful vision-language large models [12, 17, 26, 52], there is an increasing focus on exploring more modalities in recent research [13, 27-29, 46]. Bain et al. [5] intro" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29238" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 92, + 537, + 211 + ], + "blocks": [ + { + "bbox": [ + 73, + 92, + 537, + 211 + ], + "lines": [ + { + "bbox": [ + 73, + 92, + 537, + 211 + ], + "spans": [ + { + "bbox": [ + 73, + 92, + 537, + 211 + ], + "type": "image", + "image_path": "1eb310121921262fee121006cd41059657bf8ffc434340acfbb717fb61019be5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 555, + 273 + ], + "type": "text", + "content": "Figure 2. The process of the generation of Question-and-answer datasets. Samples of text and video are taken as examples. In the text sample, questions and the occurrences of specific words are randomly generated, while in the video sample, the questions are randomly generated and the keywords represent labels within the dataset. These elements are then incorporated into a GPT-4 prompt to construct answers that include key information. The right side of the image shows examples of the final Q&A pairs generated, where the answers are complete sentences containing the critical information from the key elements." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 279, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 411 + ], + "type": "text", + "content": "duce a large-scale dataset, which provides general descriptions associated with video content. Several LLM-based works [13, 27, 29, 46] are proposed to comprehend the visual content of the video. Besides, Video-LLaMa [47] extends the ability to comprehend both auditory and visual information, while Su et al. [41] leverages the power of multimodal encoders to understand across six modalities. Recently, Lv et al. [28] proposed video-based large language models for the task of VAD in a weakly supervised framework. Different from them, our work focuses on the periodic analysis of MLLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 418, + 146, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 146, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 146, + 430 + ], + "type": "text", + "content": "2.2. Periodic Tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 436, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 579 + ], + "type": "text", + "content": "Examples of periodic learning include recovering and magnifying physiological signals (e.g., heart rate or breathing) [25, 31], predicting weather and environmental changes (e.g., nowcasting of precipitation or land surface temperatures) [38], counting motions that are repetitious (e.g., exercises or therapies) [4, 7], and analyzing human behavior (e.g., gait) [23]. Prior work has focused on designing customized neural architectures [7, 20], loss functions [39], and leveraging relevant learning paradigms including transfer learning [24] and meta-learning [21] for periodic learning. This paper attempts to study the capability of periodic tasks in multimodal models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 586, + 248, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 248, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 248, + 599 + ], + "type": "text", + "content": "3. Problem Setting and Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "In nature, periodic phenomena contain both periodic and semantic information, which can be modeled by " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "x = K \\cdot p(\\omega t) + N \\cdot s(t)" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "p(\\omega t)" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " is a periodic function with angular frequency " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\omega = 2\\pi / T" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " and period " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " represents the amplitude of the periodic component, " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " denotes the amplitude of the semantic component, and " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "s(t)" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " is a semantic function related to time " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ". When discretizing continuous sequences as model inputs " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ", we adopt a sampling interval " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta t = T / n" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " is a positive integer multiple" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 279, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 327 + ], + "type": "text", + "content": "of the sampling rate " + }, + { + "bbox": [ + 313, + 279, + 555, + 327 + ], + "type": "inline_equation", + "content": "(n = 1,2,3,\\ldots)" + }, + { + "bbox": [ + 313, + 279, + 555, + 327 + ], + "type": "text", + "content": ". Ideally, when " + }, + { + "bbox": [ + 313, + 279, + 555, + 327 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 313, + 279, + 555, + 327 + ], + "type": "text", + "content": " and semantics are absent, the value at each sampling point equals the constant " + }, + { + "bbox": [ + 313, + 279, + 555, + 327 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 279, + 555, + 327 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 313, + 279, + 555, + 327 + ], + "type": "inline_equation", + "content": "x = K" + }, + { + "bbox": [ + 313, + 279, + 555, + 327 + ], + "type": "text", + "content": ". This can be mathematically expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 334, + 554, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 334, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 318, + 334, + 554, + 376 + ], + "type": "interline_equation", + "content": "X = \\left\\{ \\begin{array}{l l} K \\cdot p (\\omega t) + N \\cdot s (t), & \\Delta t = \\frac {T}{n}, n = 2, 3, 4, \\dots \\\\ K, & \\Delta t = T \\end{array} \\right. \\tag {1}", + "image_path": "f33a97cb737dcc2f9e748d46631785f65c63be3b698cf2d28b0f349b9492a86a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "text", + "content": "However, in practical applications, data is often high-dimensional and noisy. For instance, after processing a text segment through a tokenizer, we obtain a matrix with dimensions " + }, + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "inline_equation", + "content": "[L, \\mathrm{Dim}]" + }, + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "text", + "content": " represents the length of the text (i.e., the number of tokens), and Dim denotes the embedding dimension of each token. In cases of repetitive text such as \"period period period\", certain dimensions of the encoded feature matrix may exhibit constant values similar to " + }, + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "inline_equation", + "content": "x = K" + }, + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 376, + 555, + 496 + ], + "type": "text", + "content": " is a vector representing the embedding values of these tokens in specific dimensions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 497, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 689 + ], + "type": "text", + "content": "Inputs from other modalities, such as video, present more complex scenarios. Periodic movements may be obscured by a large amount of semantic or redundant information. In macro-periodic tasks, periodic movement is more prominent, allowing models to detect periodicity more easily; however, in micro-periodic tasks, the amplitude of the periodic signal " + }, + { + "bbox": [ + 313, + 497, + 555, + 689 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 313, + 497, + 555, + 689 + ], + "type": "text", + "content": " may be small and masked by noise, making it challenging for models to capture these subtle periodic variations. Despite these complexities, research suggests that models can transfer knowledge from simpler tasks to more complex ones. In this context, repetitive text can be seen as a basic periodic task, while macro-periodic and micro-periodic video tasks represent more complex periodic challenges. Although \"repetition\" is the common thread across these tasks, the semantic content and amplitude of the periodic signal vary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Given that large language models (LLMs) have been trained on extensive and diverse text corpora, they are par" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29239" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 91, + 511, + 171 + ], + "blocks": [ + { + "bbox": [ + 95, + 91, + 511, + 171 + ], + "lines": [ + { + "bbox": [ + 95, + 91, + 511, + 171 + ], + "spans": [ + { + "bbox": [ + 95, + 91, + 511, + 171 + ], + "type": "image", + "image_path": "291a3997c6228d68dc03a276bd5fe6ca42178f63429407d9785edc71208665d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 178, + 555, + 211 + ], + "lines": [ + { + "bbox": [ + 54, + 178, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 54, + 178, + 555, + 211 + ], + "type": "text", + "content": "Figure 3. The pipeline of training paradigm from easy to hard. The model begins training with simpler periodic tasks and gradually transitions to more complex ones. Throughout this process, Period-LLM's understanding of periodic tasks progressively strengthens and becomes more profound." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 295, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 295, + 240 + ], + "type": "text", + "content": "ticularly well-suited for processing text tasks. We propose the following framework:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 240, + 294, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 240, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 145, + 240, + 294, + 253 + ], + "type": "interline_equation", + "content": "A = F (X, Q) \\tag {2}", + "image_path": "0282bd31a4079c86aa55c33b20798c9adefb62413a68796e1c251f33691e4cee.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": "This framework accommodates inputs " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": " from various modalities, where " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": " represents periodic data from any modality, " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": " represents a given question, and the model computes the answer " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": ". We suggest beginning model training with simple repetitive text tasks and gradually progressing to macro-periodic and micro-periodic video tasks. This progressive training approach helps the model develop a comprehensive understanding of periodic phenomena, moving from simple to complex tasks, thereby enhancing its ability to handle periodic challenges effectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 394, + 111, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 111, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 111, + 406 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 414, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 545 + ], + "type": "text", + "content": "In this paper, we aim to enhance the understanding of large models regarding periodic tasks. Given the current absence of specialized question-and-answer datasets for periodic tasks and corresponding testing standards, we will undertake pioneering work in the content that follows. Specifically, we firstly elucidate the instruction generation for periodic tasks. Then, we explore the generalization process from simpler tasks to more complex ones. Additionally, we propose an optimization strategy to address the challenge of preserving periodic reasoning abilities during the semantic alignment phase of Multi-Modal Generalization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 549, + 274, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 274, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 274, + 561 + ], + "type": "text", + "content": "4.1. Instruction Generation for Periodic Tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 180, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 180, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 180, + 578 + ], + "type": "text", + "content": "4.1.1. Text-Only Instructions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "To enhance our dataset's diversity, we use GPT-4 to help create a repetitive text dataset through several key steps: First, we define the core question: \"How many times is the word " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{my\\text{word}\\}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " repeated in the string " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{my\\text{string}\\} ?" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " We then use GPT-4 to generate ten semantically similar questions, " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "q_{i}\\in \\{q_{1},q_{2},\\ldots ,q_{10}\\}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " . For word diversity, we randomly select a word from the GPT-4 technical report [2] as " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{my\\text{word}\\}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " . To further diversify, we set a repetition count " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{2,3,\\dots,20\\}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " , constructing " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{my\\text{string}\\}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " as \" " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{selected~word}\\} \\times n" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " . Since GPT-4 can't directly calculate word repetitions, we guide it with precise prompts. We" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "type": "text", + "content": "instruct GPT-4: \"Construct an answer based on the question format, selected word, and repetition count.\" During this process, we randomly choose a question " + }, + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "type": "text", + "content": " and provide the necessary answer elements, {selected word, " + }, + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 216, + 555, + 312 + ], + "type": "text", + "content": " times}, enabling GPT-4 to generate a fitting response. By following these steps, we effectively use GPT-4's generative capabilities to build a diverse and accurate repetitive text dataset, which supports robust model training and evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 321, + 451, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 451, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 451, + 332 + ], + "type": "text", + "content": "4.1.2. Multi-Modal Instructions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 337, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 624 + ], + "type": "text", + "content": "For tasks involving periodic data across various modalities like video, traffic streams, and health monitoring, we employ a method that integrates annotations, raw captions, and frequency information to generate comprehensive representations. This approach is applicable to diverse datasets, addressing challenges such as dynamic environments, variations in periodicity and counts, and action speeds. We transform data into a question-and-answer format inspired by repetitive text dataset construction. We generate unified captions for each segment by combining existing annotations and frequency data. The main question posed is: \"What is the total number of repetitive actions in this data?\" Using GPT-4's capabilities, we craft multiple semantically similar questions from this core inquiry. Explicit instructions and answer elements are provided to generate question-and-answer pairs, enhancing dataset representation, the process as shown in Fig. 2. This method enriches model understanding and processing of periodic tasks across modalities. We establish a link between repetitive text and periodic data, enabling generalization from simple to complex tasks. The process first involves self-optimization in relatively simple individual tasks, followed by fine-tuning on more complex tasks, as illustrated in Fig. 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 635, + 473, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 473, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 473, + 647 + ], + "type": "text", + "content": "4.2. Easy-to-Hard Generalization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "In this section, we explore the generalization process from simpler tasks to more complex ones, structured around the problem definition " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "A = F(f, Q)" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " represents the answer, " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " represents the features derived from the input data, and " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " represents the question." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29240" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 72, + 527, + 242 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 527, + 242 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 527, + 242 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 527, + 242 + ], + "type": "image", + "image_path": "467885a30330a0b59d0a855988a9252b72c8f8554f1a2bb1673d47b5cbb68811.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 244, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 54, + 244, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 244, + 555, + 289 + ], + "type": "text", + "content": "Figure 4. The Period-LLM framework, using video as an example, integrates both text and video inputs. The video is processed through a visual encoder and a visual projector, then concatenated with text features derived from a word embedding layer. This combined input is fed into a large language model (LLM). Gradient optimization strategies are applied to the output feature channels of the LLM, specifically targeting those channels that failed to capture useful information, thereby enriching the semantic content in these channels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 295, + 192, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 192, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 192, + 306 + ], + "type": "text", + "content": "4.2.1. Text-Only Generalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "type": "text", + "content": "The first stage of generalization focuses on text-based tasks, where the goal is to learn logical reasoning without needing to align other modalities with text. This can be mathematically represented as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 356, + 294, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 356, + 294, + 368 + ], + "spans": [ + { + "bbox": [ + 143, + 356, + 294, + 368 + ], + "type": "interline_equation", + "content": "A = F \\left(T _ {f}, Q\\right), \\tag {3}", + "image_path": "1b95dfc4fe0c42a35ee83dc865b5ad7ce67ef465e8f58ce8d6938dfa734b5007.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "type": "text", + "content": "here, " + }, + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "type": "inline_equation", + "content": "T_{f}" + }, + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "type": "text", + "content": " represents the features derived from textual input. The emphasis is on developing the model's logical reasoning capabilities, as the input modality is solely text, which simplifies the task by eliminating the need to handle multimodal alignment. The model learns to process and understand the logical structure of questions and generate appropriate answers purely based on textual information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 464, + 205, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 205, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 205, + 475 + ], + "type": "text", + "content": "4.2.2. Multi-Modal Generalization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 479, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 538 + ], + "type": "text", + "content": "The next stage involves more complex tasks that require the integration of multiple modalities, such as video and text. This stage not only demands logical reasoning but also the alignment of semantic information across different modalities. This process can be described by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 538, + 294, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 538, + 294, + 551 + ], + "spans": [ + { + "bbox": [ + 141, + 538, + 294, + 551 + ], + "type": "interline_equation", + "content": "A = F \\left(M _ {f}, Q\\right), \\tag {4}", + "image_path": "f3ed2f37457b996d369afbe68e590af6f56a194076394e923669201388024cf7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 654 + ], + "type": "text", + "content": "in this case, " + }, + { + "bbox": [ + 55, + 558, + 296, + 654 + ], + "type": "inline_equation", + "content": "M_{f}" + }, + { + "bbox": [ + 55, + 558, + 296, + 654 + ], + "type": "text", + "content": " represents features derived from a combination of inputs, including video and text. The challenge is twofold: the model must align semantic content across different modalities while simultaneously developing robust periodic logical reasoning capabilities. This complex task requires the model to integrate semantic features with textual information and establish accurate connections between them in order to answer questions correctly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 660, + 202, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 202, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 202, + 673 + ], + "type": "text", + "content": "4.3. Resisting Logical Oblivion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Due to the significant differences between periodic logical reasoning abilities and the ability to understand semantic information in videos, we design a Resisting Logical Oblivion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "text", + "content": "(RLO) optimization strategy to prevent the new knowledge learned in a new round of optimization from overwriting the reasoning knowledge acquired from simpler tasks. Vanilla gradient descent leads to an undifferentiated update of the entire feature space. If " + }, + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\theta = (\\theta_{1},\\theta_{2},\\dots,\\theta_{J})\\in \\mathbb{R}^{C\\times W}" + }, + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "text", + "content": " is the parameter of the model, where " + }, + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "text", + "content": " is related to the network structure, " + }, + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L(\\theta)}{\\partial\\theta_j}" + }, + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "inline_equation", + "content": "(j = 1,2,\\ldots ,J)" + }, + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "text", + "content": " represents the partial derivative of " + }, + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "text", + "content": ". The process of model updating is as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 394, + 403, + 553, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 403, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 394, + 403, + 553, + 430 + ], + "type": "interline_equation", + "content": "\\nabla \\theta_ {j} = - \\alpha \\frac {\\partial L (\\theta)}{\\partial \\theta_ {j}}, \\tag {5}", + "image_path": "54c928ae214a85964c34f446ca11852c43300af7b33c42349a4df5eb753721b5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 403, + 431, + 553, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 431, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 403, + 431, + 553, + 445 + ], + "type": "interline_equation", + "content": "\\theta_ {j} ^ {\\prime} \\leftarrow \\theta_ {j} + \\nabla \\theta_ {j}, \\tag {6}", + "image_path": "29f6fb007cd8d5e5a5b11dd56196dab419e2f4a6f7a6422566b604ea234a1adc.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 455, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 455, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 455, + 554, + 479 + ], + "type": "text", + "content": " is the learning rate, " + }, + { + "bbox": [ + 313, + 455, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\nabla \\theta_{j}" + }, + { + "bbox": [ + 313, + 455, + 554, + 479 + ], + "type": "text", + "content": " is the updating vector, and " + }, + { + "bbox": [ + 313, + 455, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\theta_j^\\prime" + }, + { + "bbox": [ + 313, + 455, + 554, + 479 + ], + "type": "text", + "content": " is the new weight for the next iteration." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 575 + ], + "type": "text", + "content": "Traditional gradient update strategies adjust model parameters based on global errors, which, although effective, can lead to knowledge interference between different types of tasks. Our new strategy introduces a feature channel weight allocation mechanism, dynamically adjusting the gradient update weights of different feature channels. Specifically, we introduce the weight function " + }, + { + "bbox": [ + 313, + 480, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\Omega (k,c)" + }, + { + "bbox": [ + 313, + 480, + 554, + 575 + ], + "type": "text", + "content": " implementation:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 575, + 553, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 575, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 356, + 575, + 553, + 608 + ], + "type": "interline_equation", + "content": "\\Omega \\left(c _ {i}\\right) = \\left\\{ \\begin{array}{l l} 1 + \\beta \\cdot e ^ {\\frac {\\text {i t e r n u m}}{\\text {m a x i t e r}}}, & \\bar {c} _ {i} < \\bar {c}, \\\\ 1, & \\bar {c} _ {i} > \\bar {c}, \\end{array} \\right. \\tag {7}", + "image_path": "760a2573ce59d00f3379090a28fc11a06e1622a2d3960dd7f0c40dd4f41ff1be.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " is a hyper-parameter, " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "iter_{num}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " denotes the current iteration count, and " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "max_{iter}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " represents the maximum number of iterations. " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\bar{c}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " is the average across all feature channels, while " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\bar{c}_i" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " refers to the average of the " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "th feature channel. By dynamically weighting feature channels, we can exert control over the real-time updates of specific feature channels. This ensures that redundant feature channels can learn newly acquired semantic information without" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29241" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 289, + 165 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 289, + 165 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 289, + 165 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 289, + 165 + ], + "type": "image", + "image_path": "e431d41866c94feb883cb1d08d332b2731d099067594816b04e1c42d30587032.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 165, + 295, + 219 + ], + "lines": [ + { + "bbox": [ + 55, + 165, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 295, + 219 + ], + "type": "text", + "content": "Figure 5. Details of the function of the gradient weight function. The left side of the image shows that some channels have lower backpropagation gradients, while others have higher ones. Higher backpropagation gradients indicate that these feature channels are being updated more significantly." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 221, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 365 + ], + "type": "text", + "content": "compromising existing reasoning abilities, thereby balancing the enhancement of reasoning abilities and video semantic understanding. As shown in Fig. 5, our gradient weighting function is applied to the backpropagation gradients, boosting those originally lower gradients, thereby substantially increasing the update of channels that were previously updated little or not at all. This process ensures that newly learned semantic information is directed to these newly emphasized feature channels for learning. Therefore, the model parameter update process with the introduction of the RLO optimization strategy is as follows, showing the changes from Eq. 5 to Eq. 8:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 367, + 294, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 367, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 133, + 367, + 294, + 380 + ], + "type": "interline_equation", + "content": "\\nabla \\theta_ {j} ^ {*} = \\Omega \\left(c _ {i}\\right) \\cdot \\nabla \\theta_ {j}, \\tag {8}", + "image_path": "bd2dd58d4fa5e38545731503f38b14b6255173a4a6b69a9c414d415b5f7857be.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 383, + 294, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 383, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 143, + 383, + 294, + 397 + ], + "type": "interline_equation", + "content": "\\theta_ {j} ^ {\\prime} \\leftarrow \\theta_ {j} + \\nabla \\theta_ {j} ^ {*}, \\tag {9}", + "image_path": "ea1a16805c761586351f773b60183cffacb34c45140e2c1d3a3f9276670827fd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "content": "During the training process, the input data " + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "content": " consists of input videos and their corresponding questions " + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "content": ", correct answers " + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "inline_equation", + "content": "i = 1,2,3,\\ldots ,K" + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "content": ", and each " + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "inline_equation", + "content": "x_{i},y_{i}" + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "content": " is a sequence of tokens. Large language models are denoted by " + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "inline_equation", + "content": "P_{\\phi}(y\\mid x)" + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "content": " represents model parameters. Typically, when fine-tuning large models, only the question-answer pairs " + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i})" + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "content": " are input. The goal is to adjust the parameters " + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 406, + 296, + 502 + ], + "type": "text", + "content": " to maximize the following expression:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 506, + 295, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 506, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 100, + 506, + 295, + 541 + ], + "type": "interline_equation", + "content": "\\max _ {\\phi} \\sum_ {(x, y) \\in \\mathcal {Z}} \\sum_ {t = 1} ^ {| y |} \\log \\left(P _ {\\phi} \\left(y _ {t} \\mid x, y _ {< t}\\right)\\right), \\tag {10}", + "image_path": "9c3829f94d394b43d27f8aa2a7138bd332e675a97ccf0d461b972d99f545fad0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 551, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 296, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 551, + 296, + 624 + ], + "type": "inline_equation", + "content": "|y|" + }, + { + "bbox": [ + 55, + 551, + 296, + 624 + ], + "type": "text", + "content": " represents the length of the input vector " + }, + { + "bbox": [ + 55, + 551, + 296, + 624 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 551, + 296, + 624 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 551, + 296, + 624 + ], + "type": "inline_equation", + "content": "P_{\\phi}(y_t|x, y_{MethodLLMZero-shotFine-tuneCountix-QArPPG-QACountix-QArPPG-QAMAERMSEBleu1CIDErMETEORMAERMSEMAERMSEBleu1CIDErMETEORMAERMSEFrozenBiLM [44]DeBERTa-V27.2211.530.1450.5500.11220.4321.565.229.440.2330.6100.20119.2120.46VideoLLaMA [48]Vicuna-7B6.1810.820.1870.3600.24518.8919.434.988.890.3890.5700.35818.2919.01LLaMA-Adapter [49]LLaMA-7B7.4410.510.2440.6830.31221.7823.345.7810.020.3780.7890.39720.5422.84VideoChat [14]Vicuna-7B6.7311.230.1650.5490.27719.5420.675.039.530.2750.6790.38818.3210.77Video-ChatGPT [30]Vicuna-7B6.4410.750.2030.5890.29718.1220.454.649.150.3560.6430.36717.5419.45BT-Adapter [19]Vicuna-7B7.7311.700.2760.6740.34119.3722.686.129.700.2960.6890.35318.6721.49LLaMA-VID [16]Vicuna-7B6.9711.210.2670.6800.30118.3118.565.349.850.3880.7830.41117.5117.60Period-LLMLLaMA-7B5.518.330.2880.7020.30815.0817.133.777.140.3970.8100.41513.7816.83", + "image_path": "19adaec8d139b3791e7c5701ddd7cc0622d77c0122fd347abdaedd33224af11a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 82, + 220, + 526, + 335 + ], + "blocks": [ + { + "bbox": [ + 55, + 179, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 179, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 553, + 213 + ], + "type": "text", + "content": "Table 1. Comparison of our method with other leading methods on two newly introduced video question-and-answer (QA) datasets, with all methods evaluated at a resolution of " + }, + { + "bbox": [ + 55, + 179, + 553, + 213 + ], + "type": "inline_equation", + "content": "{224} \\times {224}" + }, + { + "bbox": [ + 55, + 179, + 553, + 213 + ], + "type": "text", + "content": " . The results for Period-LLM in the zero-shot setting are derived from a model pretrained on text-QA dataset. For rPPG-QA, the V4V dataset is used as the validation set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 220, + 526, + 335 + ], + "lines": [ + { + "bbox": [ + 82, + 220, + 526, + 335 + ], + "spans": [ + { + "bbox": [ + 82, + 220, + 526, + 335 + ], + "type": "table", + "html": "
MethodLLMZero-shotFine-tune
RotNIST-QADrive-QARadar-QARotNIST-QADrive-QARadar-QA
MAERMSEMAERMSEMAERMSEMAERMSEMAERMSEMAERMSE
FrozenBiLM [44]DeBERTa-V24.264.7830.88120.5622.1730.422.883.2529.44110.5616.6628.55
VideoLLaMA [48]Vicuna-7B4.085.3235.45130.6723.4218.852.513.8931.22121.5517.0320.34
LLaMA-Adapter [49]LLaMA-7B3.954.4231.26125.4224.1231.792.663.2330.28118.7920.3428.56
VideoChat [14]Vicuna-7B3.564.2132.67126.8821.7129.582.233.0630.12115.2320.5427.38
Video-ChatGPT [30]Vicuna-7B3.444.8738.18118.2923.7732.132.012.2933.28112.3521.6130.45
BT-Adapter [19]Vicuna-7B4.035.4634.21110.2423.4231.472.783.5631.31107.6522.1230.19
LLaMA-VID [16]Vicuna-7B3.874.5734.98114.3318.2129.322.433.2532.45109.8718.2129.32
Period-LLMLLaMA-7B3.203.9630.88108.7417.5626.451.502.7528.71105.5414.2424.12
", + "image_path": "3964fe6c76cffa056001065f27e32cab3bece6b76da324ed28f907f3562e3813.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 336, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 336, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 553, + 369 + ], + "type": "text", + "content": "Table 2. Comparison of our method with other leading methods on the RotNIST-QA, Drive-QA, and Radar-QA datasets. Drive-QA is a dataset of taxi trajectory data, while Radar-QA is a radar modality dataset capturing physiological signals. The zero-shot results for Period-LLM are based on a model pretrained on text-QA, Countix-QA, and rPPG-QA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 516 + ], + "type": "text", + "content": "periodic reasoning abilities. The results show significant improvements over zero-shot, thanks to the inclusion of explicit periodic information in the training data. Period-LLM surpasses existing MLLMs by over 1 MAE on the Countix-QA dataset, while also showing significant improvements in generated text quality, as measured by Bleu-1, CIDEr, and METEOR scores. This improvement is driven by the easy-to-hard fine-tuning paradigm and the Resisting-Logical-Oblivion strategy, which together help the model preserve and enhance its periodic reasoning abilities, while also maintaining strong semantic understanding, resulting in superior performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 713 + ], + "type": "text", + "content": "In addition to the visual periodic tasks, Table 2 presents the performance of these models on other modalities, specifically the RotNIST-QA, Drive-QA, and Radar-QA datasets. The zero-shot and fine-tune protocols follow the same approach as in Table 1, with the key difference being that experiments on these three datasets were conducted in parallel. The pretraining models used in these experiments were fine-tuned on the two visual periodic tasks from Table 1. On the RotNIST-QA dataset, Period-LLM achieves the lowest MAE (1.50) and RMSE (2.75), outperforming models such as FrozenBiLM, VideoLLaMA, and LLaMA-Adapter. For the Drive-QA dataset, Period-LLM demonstrates superior performance with the lowest MAE (28.71) and competitive RMSE (105.54). On the Radar-QA dataset, Period-LLM also performs strongly, further validating the effectiveness of our model across different modali" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 373, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 433 + ], + "type": "text", + "content": "ties. These results underscore Period-LLM's enhanced accuracy in periodic reasoning tasks and its robustness in handling diverse datasets, establishing it as a versatile solution for periodic question-answering across a range of domains, including health, weather, and traffic." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 444, + 425, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 425, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 425, + 456 + ], + "type": "text", + "content": "5.4. Qualitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 463, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 553, + 654 + ], + "type": "text", + "content": "QA Examples. Fig. 6 showcases examples from the countix-QA and rPPG-QA datasets. It can be seen that Period-LLM is capable of understanding the semantic information contained within the video as well as recognizing periodicity within the video. For instance, in the countix-QA dataset, Period-LLM accurately provides the number of repetitive movements, while in the rPPG-QA dataset, it can determine the heart rate of the man in the video. These capabilities are beyond what other language models can achieve. According to Fig. 6, in the countix-QA dataset, Period-LLM identifies that the man in the video performs two pull-ups. In the rPPG-QA dataset, Period-LLM is able to detect that the man's heart rate is 75 beats per minute. This demonstrates Period-LLM's exceptional performance in handling video content and periodic information. More QA examples are provided in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "Comparison of Loss Reduction. The Fig. 7 illustrates the loss reduction comparison between the Baseline (Vanilla Fine-Tuning on Countix-QA) and the proposed Period-LLM approach (with text-QA pretraining and RLO) on the Countix-QA dataset. As shown, our pro" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29243" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 3. The MAE value on Countix-QA with different parameter or setting choices in " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "\\Omega \\left( {c}_{i}\\right)" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " of RLO optimization strategy, and ablation study on Countix-QA showing the impact of easy tasks (period-text-QA) and RLO (Resisting Logical Oblivion). Default settings are marked in gray ." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 127, + 124, + 170 + ], + "blocks": [ + { + "bbox": [ + 69, + 114, + 124, + 125 + ], + "lines": [ + { + "bbox": [ + 69, + 114, + 124, + 125 + ], + "spans": [ + { + "bbox": [ + 69, + 114, + 124, + 125 + ], + "type": "text", + "content": "(a) Impact of " + }, + { + "bbox": [ + 69, + 114, + 124, + 125 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 127, + 124, + 170 + ], + "lines": [ + { + "bbox": [ + 72, + 127, + 124, + 170 + ], + "spans": [ + { + "bbox": [ + 72, + 127, + 124, + 170 + ], + "type": "table", + "html": "
βMAE
0.013.85
0.053.77
0.13.90
0.54.05
", + "image_path": "c50fc6abd5e23ef3c4c7d477248606193620c30ed3ced1707f75f767c78d4542.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 145, + 127, + 208, + 164 + ], + "blocks": [ + { + "bbox": [ + 138, + 114, + 212, + 125 + ], + "lines": [ + { + "bbox": [ + 138, + 114, + 212, + 125 + ], + "spans": [ + { + "bbox": [ + 138, + 114, + 212, + 125 + ], + "type": "text", + "content": "(b) Type of Threshold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 127, + 208, + 164 + ], + "lines": [ + { + "bbox": [ + 145, + 127, + 208, + 164 + ], + "spans": [ + { + "bbox": [ + 145, + 127, + 208, + 164 + ], + "type": "table", + "html": "
ThresholdMAE
Mean3.77
Median3.85
Learnable4.20
", + "image_path": "8aa98fdd75ae35db08f954c76661fc0076c53e102278e81f978d1f21c8f7b9bf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 219, + 127, + 369, + 175 + ], + "blocks": [ + { + "bbox": [ + 215, + 105, + 369, + 124 + ], + "lines": [ + { + "bbox": [ + 215, + 105, + 369, + 124 + ], + "spans": [ + { + "bbox": [ + 215, + 105, + 369, + 124 + ], + "type": "text", + "content": "(c) Results on Countix-QA w./w.o Easy Task and RLO." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 219, + 127, + 369, + 175 + ], + "lines": [ + { + "bbox": [ + 219, + 127, + 369, + 175 + ], + "spans": [ + { + "bbox": [ + 219, + 127, + 369, + 175 + ], + "type": "table", + "html": "
Easy to HardRLOCountix-QA
Text-QACountix-QAMAECIDEr
✗✓✗4.300.661
✓✓✗3.890.782
✓✓✓3.770.810
", + "image_path": "7c60bc2be8effcead6c4e502997e393c91dbda86fae73fddb4b56fbfb535b9fe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 376, + 127, + 548, + 180 + ], + "blocks": [ + { + "bbox": [ + 373, + 114, + 547, + 125 + ], + "lines": [ + { + "bbox": [ + 373, + 114, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 373, + 114, + 547, + 125 + ], + "type": "text", + "content": "(d) Results on rPPG-QA w./w.o Easy Tasks and RLO." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 376, + 127, + 548, + 180 + ], + "lines": [ + { + "bbox": [ + 376, + 127, + 548, + 180 + ], + "spans": [ + { + "bbox": [ + 376, + 127, + 548, + 180 + ], + "type": "table", + "html": "
Easy to HardRLOrPPG-QA
Text-QACountix-QArPPG-QAMAE
XX✓X15.45
✓X✓X14.46
✓✓✓X14.05
✓✓✓✓13.78
", + "image_path": "6d089f81b8f085344037ec239700f7b61adc7b98d5d7a2a57e7cc2dbb5d1f729.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 83, + 176, + 266, + 265 + ], + "blocks": [ + { + "bbox": [ + 83, + 176, + 266, + 265 + ], + "lines": [ + { + "bbox": [ + 83, + 176, + 266, + 265 + ], + "spans": [ + { + "bbox": [ + 83, + 176, + 266, + 265 + ], + "type": "image", + "image_path": "bad20fc015a3ae26c5ecfa6dd9ea548082722e2a8b53c41bc4c4ee52faf51f09.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 268, + 271, + 279 + ], + "lines": [ + { + "bbox": [ + 77, + 268, + 271, + 279 + ], + "spans": [ + { + "bbox": [ + 77, + 268, + 271, + 279 + ], + "type": "text", + "content": "Figure 6. Examples from Countix-QA and rPPG-QA." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 78, + 281, + 269, + 365 + ], + "blocks": [ + { + "bbox": [ + 78, + 281, + 269, + 365 + ], + "lines": [ + { + "bbox": [ + 78, + 281, + 269, + 365 + ], + "spans": [ + { + "bbox": [ + 78, + 281, + 269, + 365 + ], + "type": "image", + "image_path": "5778448ea4e77748d3c7a1285ccf9f5e5ebf498cce1f6de4d92976a53225d743.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 366, + 295, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 409 + ], + "type": "text", + "content": "Figure 7. Comparison of Loss Reduction Between Baseline (Vanilla Fine-Tuning on Countix-QA) and Proposed Method (Period-LLM with Text-QA Pretraining and RLO, Fine-Tuned on Countix-QA)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 418, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 525 + ], + "type": "text", + "content": "posed method demonstrates a significantly faster decline in the loss values compared to the baseline. The Period-LLM model achieves a steeper descent in the initial training stages, indicating a more efficient learning process. Furthermore, it converges to a lower loss value in fewer epochs, suggesting improved optimization and stability. This faster convergence implies that our method better captures the underlying data patterns, leading to more effective fine-tuning on the Countix-QA." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 528, + 149, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 149, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 149, + 540 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "The Setting of Weight Function " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Omega(c_i)" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " of RLO Optimization Strategy. As seen in Table 3a, " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\beta = 0.05" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " yields the lowest MAE (3.77), indicating an effective balance between retaining prior knowledge and integrating new information. When " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " is too small (e.g., 0.01), the model's anti-forgetting ability is weak, causing higher error as new knowledge overwhelms older knowledge. Conversely, a high " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " (e.g., 0.5) overly prioritizes old knowledge, limiting adaptation. Thus, " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\beta = 0.05" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " achieves the best balance for minimizing error. In Table 3b, setting the threshold to the mean results in the lowest MAE (3.77), effectively balancing old and new knowledge contributions. The median threshold performs slightly worse (MAE 3.85) due to its insensitivity to outliers, while a learnable threshold increases MAE" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 196, + 555, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 555, + 219 + ], + "type": "text", + "content": "(4.20), likely due to instability. Hence, the mean threshold is the most stable and effective choice." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 223, + 555, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 451 + ], + "type": "text", + "content": "Results of various combinations of Preiod-LLM Components. We evaluate our approach on the Countix-QA and rPPG-QA datasets, as shown in Tables 3c and 3d. For Countix-QA, our method—combining pre-training on simpler tasks with the Resisting Logical Oblivion (RLO) strategy—significantly improves performance, reducing the Mean Absolute Error (MAE) from 4.30 to 3.77 and increasing the CIDEr score from 0.661 to 0.810. Similarly, on the rPPG-QA dataset, this approach lowers the MAE from 15.45 to 13.78. These improvements reflect the model's enhanced capacity to capture periodic patterns effectively while also understanding semantic content more deeply. The decrease in MAE shows better accuracy in numerical predictions, while the CIDEr increase suggests improved semantic alignment in question-answering tasks. Together, these results confirm that the integration of easy-task pretraining with RLO enables the model to balance periodicity and semantic understanding, leading to faster convergence and more robust performance across QA tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 462, + 388, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 462, + 388, + 475 + ], + "spans": [ + { + "bbox": [ + 314, + 462, + 388, + 475 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce Period-LLM, a large multimodal model specifically designed to enhance performance on periodic tasks across various modalities. The model follows a progressive training paradigm, starting with simpler text-based tasks and advancing to more complex visual and multimodal tasks. Additionally, we propose the \"Resisting Logical Oblivion\" optimization strategy to preserve the model's periodic reasoning abilities during semantic alignment. The Our experimental results demonstrate that Period-LLM significantly outperforms existing models in handling periodic tasks, especially in challenging domains such as repetitive motion analysis and physiological signal estimation. Period-LLM not only achieves superior accuracy but also highlights its potential for future advancements in managing complex periodic phenomena across different modalities. Future work will focus on further refining model's capabilities, particularly in addressing more nuanced and less structured periodic tasks, and expanding its application to additional domains where periodicity plays a crucial role." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29244" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work was supported by the Guangdong Provincial Key Laboratory of Integrated Communication, Sensing and Computation for Ubiquitous Internet of Things (No. 2023B1212010007), China NSFC Grant 62472366, the Project of DEGP (No.2024GCZX003, 2023KCXTD042), 111 Center (No.D25008) and Shenzhen Science and Technology Foundation (ZDSYS20190902092853047)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 205, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 205, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 248 + ], + "type": "text", + "content": "[1] Kumar Abhishek, Maheshwari Prasad Singh, Saswata Ghosh, and Abhishek Anand. Weather forecasting model using artificial neural network. Procedia Technology, 4:311-318, 2012. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 250, + 296, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 250, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 61, + 250, + 296, + 303 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 296, + 369 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 371, + 296, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 296, + 436 + ], + "type": "text", + "content": "[4] Mohammad Rafayet Ali, Javier Hernandez, E Ray Dorsey, Ehsan Hoque, and Daniel McDuff. Spatio-temporal attention and magnification for classification of parkinson's disease from videos collected via the internet. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 207-214. IEEE, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 437, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 296, + 480 + ], + "type": "text", + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 482, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 296, + 525 + ], + "type": "text", + "content": "[6] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 526, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 296, + 581 + ], + "type": "text", + "content": "[7] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Counting out time: Class agnostic video repetition counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10387-10396, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 582, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 582, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 582, + 296, + 645 + ], + "type": "text", + "content": "[8] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "text", + "content": "[9] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xiangyu Yue. Onellm: One framework to align all modalities with language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26584-26595, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "text", + "content": "[10] Huazhang Hu, Sixun Dong, Yiqun Zhao, Dongze Lian, Zhengxin Li, and Shenghua Gao. Transrac: Encoding multiscale temporal correlation with transformers for repetitive action counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19013-19022, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 140, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 554, + 183 + ], + "type": "text", + "content": "[11] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 186, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 554, + 228 + ], + "type": "text", + "content": "[12] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 230, + 554, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 554, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 554, + 273 + ], + "type": "text", + "content": "[13] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 275, + 554, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 275, + 554, + 318 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 554, + 318 + ], + "type": "text", + "content": "[14] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 320, + 554, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 554, + 385 + ], + "type": "text", + "content": "[15] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 388, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 420 + ], + "type": "text", + "content": "[16] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 422, + 554, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 454 + ], + "type": "text", + "content": "[17] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 456, + 554, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 488 + ], + "type": "text", + "content": "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 490, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 554, + 533 + ], + "type": "text", + "content": "[19] Ruyang Liu, Chen Li, Yixiao Ge, Ying Shan, Thomas H Li, and Ge Li. One for all: Video conversation is feasible without video instruction tuning. arXiv preprint arXiv:2309.15785, 2023. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "type": "text", + "content": "[20] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems, 33:19400-19411, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 580, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 554, + 633 + ], + "type": "text", + "content": "[21] Xin Liu, Ziheng Jiang, Josh Fromm, Xuhai Xu, Shwetak Patel, and Daniel McDuff. Metaphys: few-shot adaptation for non-contact physiological measurement. In Proceedings of the conference on health, inference, and learning, pages 154-163, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 635, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 678 + ], + "type": "text", + "content": "[22] Xin Liu, Yuting Zhang, Zitong Yu, Hao Lu, Huanjing Yue, and Jingyu Yang. rppg-mae: Self-supervised pretraining with masked autoencoders for remote physiological measurements. IEEE Transactions on Multimedia, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[23] Yingcheng Liu, Guo Zhang, Christopher G Tarolli, Rumen Hristov, Stella Jensen-Roberts, Emma M Waddell, Taylor L Myers, Meghan E Pawlik, Julia M Soto, Renee M Wilson," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29245" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "text", + "content": "et al. Monitoring gait at home with radio waves in parkinson's disease: A marker of severity, progression, and medication response. Science Translational Medicine, 14(663): eadc9669, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[24] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 669–684. Springer, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[25] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12404-12413, 2021. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 296 + ], + "type": "text", + "content": "[26] Hao Lu, Xuesong Niu, Jiyao Wang, Yin Wang, Qingyong Hu, Jiaqi Tang, Yuting Zhang, Kaishen Yuan, Bin Huang, Zitong Yu, et al. Gpt as psychologist? preliminary evaluations for gpt-4v on visual affective computing. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) workshop, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "text", + "content": "[27] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Minghui Qiu, Pengcheng Lu, Tao Wang, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "type": "text", + "content": "[28] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "text", + "content": "[29] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "type": "text", + "content": "[30] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[31] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2019. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "text", + "content": "[32] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3170-3180, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 622 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[34] Ambareesh Revanur, Zhihua Li, Umur A Ciftci, Lijun Yin, and László A Jeni. The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2760-2767, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[35] Eduardo Rocha Rodrigues, Igor Oliveira, Renato Cunha, and Marco Netto. Deepdownscale: A deep learning strategy for high-resolution weather forecast. In 2018 IEEE 14th Interna-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "tional Conference on e-Science (e-Science), pages 415-422. IEEE, 2018. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 160 + ], + "type": "text", + "content": "[36] Sven Schellenberger, Kilin Shi, Tobias Steigleder, Anke Malessa, Fabian Michler, Laura Hameyer, Nina Neumann, Fabian Lurz, Robert Weigel, Christoph Ostgathe, et al. A dataset of clinically recorded radar vital signs with synchronised reference sensor signals. Scientific data, 7(1):291, 2020. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 204 + ], + "type": "text", + "content": "[37] Sebastian Scher and Gabriele Messori. Predicting weather forecast uncertainty with machine learning. Quarterly Journal of the Royal Meteorological Society, 144(717):2830-2841, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 205, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 259 + ], + "type": "text", + "content": "[38] Casper Kaae Sønderby, Lasse Espeholt, Jonathan Heek, Mostafa Dehghani, Avital Oliver, Tim Salimans, Shreya Agrawal, Jason Hickey, and Nal Kalchbrenner. Metnet: A neural weather model for precipitation forecasting. arXiv preprint arXiv:2003.12140, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "text", + "content": "[39] Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[40] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In The 23rd IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062. IEEE, 2014. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[41] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 435 + ], + "type": "text", + "content": "[42] Jingyuan Wang, Jiawei Jiang, Wenjun Jiang, Chengkai Han, and Wayne Xin Zhao. Towards efficient and comprehensive urban spatial-temporal prediction: A unified library and performance benchmark. arXiv preprint arXiv:2304.14343, 2023.6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "text", + "content": "[43] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 1-7. IEEE, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "text", + "content": "[44] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems, 35:124-141, 2022. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "text", + "content": "[45] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. Simper: Simple self-supervised learning of periodic targets. arXiv preprint arXiv:2210.03115, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "type": "text", + "content": "[46] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[47] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[48] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 6, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29246" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 262 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[49] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[50] Yunhua Zhang, Ling Shao, and Cees GM Snoek. Repetitive activity counting by sight and sound. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14070-14079, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[51] Yuting Zhang, Hao Lu, Xin Liu, Yingcong Chen, and Kaishun Wu. Advancing generalizable remote physiological measurement through the integration of explicit and implicit prior knowledge. arXiv preprint arXiv:2403.06947, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[52] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29247" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/1180818e-c94d-44fa-b5df-b1fc8919ac06_content_list.json b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/1180818e-c94d-44fa-b5df-b1fc8919ac06_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..47134e50c23b4aea8e927b9108b6d89424c3a663 --- /dev/null +++ b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/1180818e-c94d-44fa-b5df-b1fc8919ac06_content_list.json @@ -0,0 +1,1739 @@ +[ + { + "type": "text", + "text": "Person De-reidentification: A Variation-guided Identity Shift Modeling", + "text_level": 1, + "bbox": [ + 140, + 130, + 856, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi-Xing Peng $^{1,2,4}$ Yu-Ming Tang $^{1,4}$ Kun-Yu Lin $^{1,4}$ Qize Yang $^{3}$ Jingke Meng $^{1,4*}$ Xihan Wei $^{3}$ Wei-Shi Zheng $^{1,2,4*}$", + "bbox": [ + 192, + 178, + 803, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1School of Computer Science and Engineering, Sun Yat-sen University, China;", + "bbox": [ + 187, + 215, + 808, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Peng Cheng Laboratory, Shenzhen, China; $^{3}$ Tongyi Lab, Alibaba Group;", + "bbox": [ + 207, + 233, + 789, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China.", + "bbox": [ + 107, + 251, + 887, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{pengyx23,tangym9,linky5}@mail2.sysu.edu.cn wszheng@ieee.org", + "bbox": [ + 212, + 272, + 779, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person re-identification (ReID) is to associate images of individuals from different camera views against cross-view variations. Like other surveillance technologies, Re-ID faces serious privacy challenges, particularly the potential for unauthorized tracking. Although various tasks (e.g., face recognition) have developed machine unlearning techniques to address privacy concerns, such methods have not yet been explored within the Re-ID field. In this work, we pioneer the exploration of the person de-reidentification (De-ReID) problem and present its inherent challenges. In the context of ReID, De-ReID is to unlock the knowledge about accurately matching specific persons so that these \"unlearned persons\" cannot be re-identified across cameras for privacy guarantee. The primary challenge is to achieve the unlearning without degrading the identity-discriminative feature embeddings to ensure the model's utility. To address this, we formulate a De-ReID framework that utilizes a labeled dataset of unlearned persons for unlearning and an unlabeled dataset of accessible persons for knowledge preservation. Instead of unlearning based on (pseudo) identity labels, we introduce a variation-guided identity shift mechanism that unlearns the specific persons by fitting the variations in their images while preserving ReID ability on other persons by overcoming the variations in images of accessible persons. As a result, the model shifts the unlearned persons to a feature space that is vulnerable to cross-view variations. Extensive experiments on benchmarks demonstrate the superiority of our method.", + "bbox": [ + 89, + 353, + 485, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 790, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person re-identification (ReID) aims to match the images of the same person across different camera views based on feature similarity. While deep learning [9, 35, 43] has significantly enhanced the performance of the ReID mod", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a3a019974be3f120f13baeb520587213302d3b239815ca7788817d7c063accb2.jpg", + "image_caption": [ + "Figure 1. Person de-reidentification (De-ReID) prevents intelligent surveillance from tracking unlearned persons. For example, when using surveillance models for security (track and record the person in the red box), the residents in the community and the important customers in business (in the green box) should not be tracked for privacy and commercial secrets." + ], + "image_footnote": [], + "bbox": [ + 516, + 321, + 898, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els [10, 19, 45], there are growing public concerns about potential misuse, leading to ethical and societal implications. To avoid misuse and mitigate the potential privacy concerns, the research in surveillance has begun exploring machine unlearning techniques that limit the operational scope of these models. For instance, in face recognition, models are adjusted to \"unlearn\" individuals who raise privacy concerns, as mandated by regulations like the EU's General Data Protection Regulation (GDPR) [3, 5, 24, 53].", + "bbox": [ + 511, + 489, + 908, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, similar constraints on the ReID models have not yet to be explored, despite its significance for real-world applications. For instance, the residents in the neighborhood prefer to constrain the surveillance in the community to only focus on external people and not track themselves; in working buildings, security cameras should only be set for monitoring and preventing theft or accidents, not for tracking or identifying specific employees during their routine work; and homeowners expect smart home devices to track only unknown individuals for security purposes while avoiding any continuous surveillance of themselves; To meet these requirements, there is an urgent need to develop methods that limit the scope of ReID models, enabling them to \"ignore\" specific individuals during deployment.", + "bbox": [ + 511, + 628, + 908, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we pioneer the exploration of person deidentification (De-ReID) problem: limiting the ReID scope by unlearning specific persons (referred to as \"unlearned persons\") in the ReID model while maintaining the ReID", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29331", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "performance on other persons (referred to as \"accessible persons\"). In the context of image matching, a person is unlearned by a model means the model can hardly associate the images of the person under different camera views, as shown in Fig. 1. Forming a De-ReID model is challenging because only the knowledge strictly related to matching the unlearned persons should be eliminated, without destroying the identity-discriminative feature embeddings. A trivial solution is to distinguish the images of unlearned persons under different camera views for unlearning while associating the images of accessible persons based on their identity labels to preserve ReID knowledge. However, due to the large network capacity, the model may learn spurious cues or overfit the images of unlearned persons in the training set, leading to poor generalization during evaluation. Additionally, collecting identity labels for accessible persons to preserve ReID knowledge is both costly and difficult to scale, especially when the number of cameras is large [29, 31, 36].", + "bbox": [ + 89, + 90, + 485, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To handle the De-ReID, we formulate a weakly-supervised unlearning framework where a labeled image set of unlearned persons and an unlabeled dataset of accessible persons are available to adapt a well-trained ReID model to De-ReID. Given that the core of ReID is to overcome cross-view variations, such as different camera viewing angles and lighting conditions, we propose variation-guided identity shift that learns De-ReID based on the variations in images. The core of our method is to guide the ReID model bias towards environmental information (e.g., lighting) when encountering unlearned persons, which shifts the unlearned persons out of the view-invariant feature space. At the same time, we keep the model remains robust to cross-view variations when querying accessible persons.", + "bbox": [ + 89, + 364, + 485, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, for unlearned persons, we guide the model to distinguish their images and the images augmented by variations. Compared with distinguishing images under different camera views, learning to distinguish image and its augmentation results is more effective for unlearning since the augmented images are highly similar to the original image (e.g., body shape) but varies in style information (e.g., brightness and color tone). Therefore, the model is biased to encode the style information to push away these images. In contrast, for accessible persons, we pull closer their images and augmented versions in the feature space, expecting the model to overcome the variations when matching them. Moreover, we introduce relation regularization to characterize the desired properties of De-ReID in terms of relations.", + "bbox": [ + 89, + 579, + 485, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 109, + 792, + 447, + 806 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to investigate the challenging De-ReID task and conduct a new benchmark.", + "- We formulate a De-ReID framework and propose variation-guided identity shift that achieves De-ReID by fitting the variations in images of interest while against the variations in others to preserve ReID knowledge." + ], + "bbox": [ + 91, + 810, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- To further utilize the identity annotations in unlearned person images for learning, we propose relation regularization that characterizes the desired properties of ReID and De-ReID. Both qualitative and quantitative results demonstrate the effectiveness of our method.", + "bbox": [ + 513, + 90, + 908, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 180, + 663, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Person re-identification. Based on deep learning, advanced ReID models automatically learn a discriminative feature embedding from a large amount of data [20]. Various network architectures [19, 21, 22, 30, 39, 42, 44, 45] and loss functions [10, 40, 41, 48] are developed to enhance feature representation learning. Besides, the challenges in ReID including cross-modality [25, 31, 51, 55], cloth-changing [23, 26], low-resolution [62], occlusion [52], and unsupervised ReID [7, 50] are also widely studied.", + "bbox": [ + 511, + 205, + 906, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Privacy protection for person images. Apart from pursuing high performance, protecting human privacy in human-centric tasks is also important [54, 59, 60]. A stream of research focuses on generating anonymized images by removing or obscuring identity-related information, such as faces and silhouettes, while preserving the image's utility for subsequent tasks, such as action recognition [1, 4, 15].", + "bbox": [ + 511, + 342, + 908, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the field of ReID, Ye et al. [54, 57] proposed a generative model to anonymize the person images to avoid privacy leakage while keeping the ReID utility. Wang et al. [46] add imperceptible perturbations to the images of in the database to prevent malicious use of the images. Particularly, one stream of works explores person de-identification [1, 4, 15], which obscures the identity information in images or videos, including face and silhouette. The new sensors including event cameras and LiDAR are also investigated for privacy-preserving ReID [2, 6, 14, 17].", + "bbox": [ + 511, + 448, + 908, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their progress, previous works mainly focus on preventing the malicious use of images while we aim to prevent the malicious use of models. Technically, previous works [54, 57] learn to anonymize person images and jointly adapt the ReID model to ensure ReID utility. Their methods cannot prevent the ReID model from retrieving specific persons, as well as the methods that explore new sensors such as event cameras for ReID [2, 6, 14, 17]. Besides, although PRIDE [46] protects the images in the database from being maliciously used, the attacker can also collect images that are not in the database to track specific persons. Moreover, some works in person de-identification only concern the identity-unrelated utility [1, 4, 15], such as action recognition, while De-ReID requires identifying accessible persons.", + "bbox": [ + 511, + 599, + 908, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Machine unlearning. De-ReID has a similar goal to machine unlearning. Machine unlearning is first proposed to eliminate the influence of specific data on the deep model as if the model never uses these data for training, which enables users to erase their personal data [3, 5, 16, 18, 37].", + "bbox": [ + 511, + 810, + 908, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Kurmanji et al. [24] further apply machine unlearning for", + "bbox": [ + 532, + 886, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29332", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "various purposes such as removing the bias from the model or eliminating the negative effects of mislabeled samples. To maximize the error in forgetting classes, they maximize the difference between the outputs of the learned model and the original model in the forgetting samples. Choi et al. [11] propose to jointly maximize the loss on the forgetting data and learn on the other data. Chen et al. [8] unlearn specific classes by adjusting their decision boundary. The forgetting samples are assigned a nearest but incorrect label for boundary shrink. Zhao et al. [58] propose GS LoRA to maintain performance on remaining classes when maximizing the loss on the forgetting data. Ye et al. [53] enforce that the model should be different from the original model in terms of the attention map of intermediate features to forget classes.", + "bbox": [ + 89, + 90, + 483, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work is different from existing works. Firstly, even if the model does not see the unlearned person in training, the ReID model can effectively match the cross-view images of unlearned persons. Hence, the methods eliminating the influence of samples on network weights are not applicable to De-ReID. For the methods that maximize the errors in the forgetting classes, they mainly focus on classification tasks. Differently, ReID pursues discriminative features for matching. Even if the unlearned persons cannot be correctly classified as their methods desired, the model could retrieve the correct images as long as the feature distances between the query and the correct images are smaller than others.", + "bbox": [ + 89, + 301, + 483, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. De-ReID Learning", + "text_level": 1, + "bbox": [ + 89, + 497, + 272, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Problem formulation. Given a well-trained ReID model denoted as $f_{p}(\\cdot)$ and a set of unlearned persons $\\mathcal{P}_u$ , our goal is to learn a model $f(\\cdot)$ that the model cannot match the images of unlearned persons while keeping the ReID performance for other persons that are accessible to model. $f(\\cdot)$ is initialized as $f_{p}(\\cdot)$ , and $f_{p}(\\cdot)$ is kept unchanged. An ideal De-ReID model $f(\\cdot)$ is expected to satisfy the conditions:", + "bbox": [ + 88, + 523, + 482, + 630 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\phi_ {f} (x, x _ {p}) < \\phi_ {f} (x, x _ {n}), \\forall x \\in \\mathcal {P} _ {u}, \\forall x _ {n} \\text {s . t .} y _ {n} \\neq y, \\\\ \\phi_ {f} (x, x _ {p}) > \\phi_ {f} (x, x _ {n}), \\forall x \\notin \\mathcal {P} _ {u}, \\forall x _ {n} \\text {s . t .} y _ {n} \\neq y, \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 641, + 483, + 693 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi_f(\\cdot, \\cdot)$ is the similarity metric in the feature space of $f(\\cdot)$ . $x$ is a query image with label $y$ , and likewise $y_p$ and $y_n$ are labels of $x_p$ and $x_n$ . $x_p$ is a different image containing the same person with $x$ , but $x_n$ contains a different person.", + "bbox": [ + 89, + 694, + 482, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To learn $f(\\cdot)$ for approximating Eq. 1, we assume a labeled image set of $M_T$ unlearned persons denoted as $S_T = \\{(x_i^t,y_i^t)\\}_{i = 1}^{|S_T|}$ and an unlabeled image set of accessible persons denoted as $S_O = \\{x_j^o\\}_{j = 1}^{|S_O|}$ are available for fine-tuning. $x_{i}^{t}$ is the $i$ -th image in $S_T$ with corresponding label $y_{i}^{t}\\in \\{1,2,\\dots,M_{T}\\}$ . $x_{j}^{o}$ is the $j$ -th image in $S_O$ . The persons in $S_O$ are different from those in $S_T$ and the persons in testing. While unlearning the specific persons based on the labeled data $S_T$ seems straightforward (e.g., by max", + "bbox": [ + 89, + 753, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "imizing the classification error), it is hard to ensure only the knowledge strictly subject to matching the unlearned persons is eliminated, rather than simply destroy identity discriminative feature embedding.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the ubiquitous existence of cross-view variations, we propose variation-guided identity shift based on variations in images to unlearn the specific person in view-invariant feature space. Besides, we formulate the goals of De-ReID in the form of images relations and present relation regularization. An overview of our method is in Fig. 2.", + "bbox": [ + 511, + 151, + 906, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Variation-guided Identity Shift", + "text_level": 1, + "bbox": [ + 511, + 251, + 785, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intuitively, person ReID requires a view-invariant feature space, where image features are identity discriminative regardless of camera views. Hence, we propose to shift the unlearned persons out of the view-invariant feature space, and therefore their images from different camera views cannot be properly associated. To this end, we guide the model to fit the variations in images of unlearned persons and overcome the variations for accessible persons, forming asymmetric learning objectives. In this way, the images of unlearned persons will be encoded into distinct features affected by the variations in images while the images of accessible persons are encoded as view-invariant features.", + "bbox": [ + 511, + 272, + 906, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To introduce abundant variations, we employ augmentation function $\\mathcal{T}(\\cdot)$ to obtain an augmented view $\\hat{x}_i^t = \\mathcal{T}(x_i^t)$ for each image $x_i^t$ . To simulate the natural cross-view variations, $\\mathcal{T}(\\cdot)$ should not change the identity of $x_i^t$ but properly adjust the images' color, brightness, and so forth, ensuring that $\\hat{x}_i^t$ contains the same person with $x_i^t$ but is different in image style. We ablate the design of $\\mathcal{T}(\\cdot)$ in experiments in our Appendix. To shift the unlearned persons out of the view-invariant feature space, the model $f(\\cdot)$ should prioritize rich environmental information unrelated to identity when extracting features from the unlearned person's image $x_i^t$ . In other words, the feature $f(x_i^t)$ should encode rich environmental information (e.g., background, image style) that cannot be applied for matching unlearned person in other cameras. To this end, we guide $f(\\cdot)$ to push away $x_i^t$ and its augmented view $\\hat{x}_i^t$ in the feature space, which encourages the model to fit the variations for images of the unlearned person. Formally, the learning objective for shifting the unlearned persons out of view-invariant feature space is:", + "bbox": [ + 511, + 454, + 908, + 741 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {V I S} ^ {p} = \\left[ \\sigma_ {c} - \\| f \\left(x _ {i} ^ {t}\\right) - f \\left(\\hat {x} _ {i} ^ {t}\\right) \\| _ {2} ^ {2} \\right] _ {+}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 751, + 906, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $[a]_{+} = \\max(a, 0.0)$ and $\\sigma_{c}$ is the margin to bound the loss. Through optimization, the features of unlearned person images become image-specific and cannot be used for cross-view matching since the features are severely affected by the variations.", + "bbox": [ + 511, + 779, + 906, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Simply learning with Eq. 2 alone will cause the model to forget the ReID knowledge, and fail to extract features robust to variations for the accessible persons. It will cause", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29333", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2abb36322f7c62f6c6849e2d3af626cdfbd4f4a5fb6e947db5b4bc86bceefd47.jpg", + "image_caption": [ + "Figure 2. An overview of our method. In variation-guided identity shift, the model is guided to adapt to the variations when encountering images of unlearned persons, thereby shifting them out of the view-invariant feature space. Meanwhile, the view-invariant feature space is maintained for other accessible persons by guiding the model to overcome the variations when encountering images of these individuals. Relation regularization formulates the desired properties of De-ReID in terms of image relations and ensures the feature discriminability." + ], + "image_footnote": [], + "bbox": [ + 174, + 88, + 821, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "feature space collapse instead of the specific identity shift. To maintain a view-invariant feature space for re-identifying accessible persons, we guide the model to overcome the variations for images of these persons. For image $x_{j}^{o}$ of an accessible person, we obtain its augmented view $\\hat{x}_{j}^{o} = \\mathcal{T}(x_{j}^{o})$ . We encourage the model to pull closer $x_{j}^{o}$ and $\\hat{x}_{j}^{o}$ to remind the model of extracting robust features for accessible persons. Formally, the learning function for accessible person is:", + "bbox": [ + 89, + 349, + 483, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {V I S} ^ {o} = \\left\\| f \\left(x _ {j} ^ {o}\\right) - f \\left(\\hat {x} _ {j} ^ {o}\\right) \\right\\| _ {2} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 479, + 483, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining Eq. 3 and Eq. 2, we form an asymmetric objective to shift the identities of unlearned persons while preserving a view-invariant feature space for accessible persons:", + "bbox": [ + 89, + 511, + 485, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {V I S} = L _ {V I S} ^ {p} + L _ {V I S} ^ {o}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 571, + 483, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Discussion on asymmetric learning scheme. $L_{VIS}$ alters knowledge within the ReID model by treating the variations in images asymmetrically. The two asymmetric parts in the $L_{VIS}$ are collaborative, not isolated. Intuitively, one may think that guiding the ReID network against the variations $(L_{VIS}^{o})$ is merely for keeping ReID knowledge for accessible persons. However, $L_{VIS}^{o}$ is also helpful for unlearning because it maintains a view-invariant feature space, which in turn gives the direction for shifting the identities of unlearned persons $(L_{VIS}^{p})$ : out of the view-invariant space. As a result, more unlabeled images of accessible persons can improve the unlearning effect, which is not shown in other methods.", + "bbox": [ + 89, + 595, + 483, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Relation Regularization", + "text_level": 1, + "bbox": [ + 89, + 787, + 310, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The relations between pairwise images are critical in the context of person ReID, directly reflecting the discriminative capability of the features. As formulated in Eq. 1, there should be some accessible persons spread around an unlearned person. Hence, we introduce relation regularization to facilitate De-ReID and ReID learning. Specifically, we", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "constraints that, in the feature space, the distance between the images of the same unlearned person should be greater than the distance to the accessible persons. This TRiplet Constraint facilitates retrieving images of different persons when querying an unlearned person's image.", + "bbox": [ + 511, + 349, + 906, + 425 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {T R C} = \\left[ d _ {p} - d _ {n} + \\sigma_ {r} \\right] _ {+},\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 430, + 803, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} d _ {p} = \\left\\| x _ {i} ^ {t} - x _ {i, K} ^ {o} \\right\\| _ {2} ^ {2}, d _ {n} = \\left\\| x _ {i} ^ {t} - x _ {h} ^ {t} \\right\\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 443, + 905, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $x_{i,K}^{o}(x_{i,K}^{o}\\in S_{O})$ represents the $K$ -th nearest image to $x_{i}^{t}$ , and $\\sigma_r$ is the margin. $x_{h}^{t}$ is an image containing the same person as $x_{i}^{t}\\left(y_{h}^{t} = y_{i}^{t}\\right)$ . $L_{TRC}$ constrains that the distance between pairwise images of the same unlearned person should be at least larger than $d_p + \\sigma_r$ , where $d_p$ is dynamically determined based on the distance to the $K$ -th nearest accessible person. When $K$ is larger, $L_{TRC}$ encourages more accessible persons to spread around $x_{i}^{t}$ . Through optimizing $L_{TRC}$ , the top retrieval results mainly contain different persons when retrieving unlearned person images from feature space.", + "bbox": [ + 511, + 472, + 906, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Complementary to variation-guided identity shift that changes the cross-view invariance in features, we employ Relation Consistent Regularization on dataset $S_O$ to maintain the feature discriminability during De-ReID fine-tuning. The regularization is to ensure that the relations between accessible individuals remain as discriminative as those in the original model $f_p(\\cdot)$ during the unlearning process:", + "bbox": [ + 511, + 640, + 908, + 746 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {R C R} = \\left\\| \\operatorname {s i m} \\left(f _ {p} \\left(x _ {j} ^ {o}\\right), f _ {p} \\left(x _ {h} ^ {o}\\right)\\right) - \\operatorname {s i m} \\left(f \\left(x _ {j} ^ {o}\\right), f \\left(x _ {h} ^ {o}\\right)\\right) \\right\\| _ {2} ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 750, + 906, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By cooperating with $L_{VIS}$ , the model can extract discriminative and view-invariant features for accessible persons when unlearning the specific persons.", + "bbox": [ + 511, + 782, + 906, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Overall Learning objective. The model $f(.)$ is trained end-to-end with an overall learning objective as:", + "bbox": [ + 511, + 827, + 905, + 858 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {1} * L _ {V I S} + \\lambda_ {2} * L _ {T R C} + L _ {R C R}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 864, + 906, + 880 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are trade-off parameters.", + "bbox": [ + 511, + 885, + 792, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29334", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 223, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 123, + 382, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Datasets. We conduct experiments in Market-1501 [61], MSMT17 [47], Occ-Duke [34] and SYSU-MM01 [49] datasets. Due to the limited space, we leave the results in Occ-Duke and SYSU-MM01 datasets in our Appendix. Market-1501 (abbreviated as Market) consists of 1,501 persons, and all the images are captured from 6 camera views. In the conventional ReID protocol, there are 750 persons for training and 751 persons for testing. MSMT17 is a large-scale ReID dataset, containing images of 4101 persons from 15 disjoint camera views. Conventionally, 1,041 persons are for training and 3,060 persons for testing. For evaluating De-ReID, we slightly modify the conventional protocol by referring to the experimental setups of other machine unlearning works [8, 24, 53, 58].", + "bbox": [ + 88, + 154, + 485, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, in each dataset, we take $M_T$ persons from the testing set as the unlearned persons. Considering the scale of the datasets, we set $M_T \\in \\{25, 50, 75, 100\\}$ in MSMT17 and $M_T \\in \\{25, 50\\}$ in Market1501. For each unlearned person, we split his images into two image sets from disjoint cameras. One set is added to the training set and the other is kept in the testing set. After adding the unlearned persons to the training set, we divide the training set into two subsets. One pre-training subset consists of accessible persons with identity annotations and is used for pretraining the ReID model. There are 900 accessible persons in the pre-training subset of MSMT17 and 600 accessible persons in the pre-training subset of Market-1501.", + "bbox": [ + 89, + 375, + 483, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The fine-tuning subset consists of unlabeled images of accessible persons and labeled images of unlearned persons, simulating the practical scenario where we need to efficiently fine-tune the ReID model without collecting additional labeled images of accessible persons. The dataset statistics are shown in Table 1. We emphasize that the images of the unlearned persons in the testing set under the cameras are different from those of the training set. The identities of accessible persons in the testing set are different from those in the training set. More details are in Appendix.", + "bbox": [ + 89, + 580, + 485, + 733 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Experimental setup. In each dataset, the pre-training subset is used to learn a ReID model in a supervised manner as $f_{p}(\\cdot)$ . After obtaining the well-trained ReID model, the fine-tuning subset is for learning De-ReID model $f(\\cdot)$ in our main experiments.", + "- Evaluation metrics. We adopt the widely used Cumulative Matching Characteristic (CMC) and mean Average Precision (mAP) to evaluate our performance. Following previous works [38, 58], we also adopt H-Means to comprehensively evaluate the model's ability of ReID in accessible persons" + ], + "bbox": [ + 88, + 739, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e8b62b255fd692321cb2179166104c310a3914f8903d695a0cc49dfc143d94a6.jpg", + "table_caption": [ + "Table 1. Statistics in the case of $M_T = 100$ in MSMT17 and $M_T = 50$ in Market1501. $M_T$ is the number of unlearned persons for De-ReID. Notably, the accessible persons in training are different from those in testing. \"Train\" refers to the fine-tuning subset. For the unlearned persons, the images in training and those in testing are from different cameras. More details are in Appendix." + ], + "table_footnote": [], + "table_body": "
MSMT17Market-1501
MT|ST|MO|SO|MT|ST|MO|SO|
Train10042951415061509311512284
Query100533296010706501507003068
Gallery10041362960741655091870018033
", + "bbox": [ + 522, + 174, + 901, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and De-ReID in unlearned persons. Formally,", + "bbox": [ + 511, + 253, + 815, + 268 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {H} - \\text {M e a n} = \\frac {2 * \\mathrm {R} - 1 _ {O} (f) * \\Delta \\mathrm {R} - 1 _ {T}}{\\mathrm {R} - 1 _ {O} (f) + \\Delta \\mathrm {R} - 1 _ {T}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 279, + 906, + 313 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw h e r e \\Delta R - 1 _ {T} = R - 1 _ {T} (f _ {p}) - R - 1 _ {T} (f).\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 314, + 834, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{R - 1}_O(f)$ is model $f$ 's Rank-1 accuracy in accessible persons, and $\\Delta \\mathrm{R - 1}_T$ is the changes in the Rank-1 accuracy in unlearned persons after De-ReID fine-tuning using different methods. In the following, we only discuss the performance of $f(\\cdot)$ and therefore $\\mathrm{R - 1}_O$ means $\\mathrm{R - 1}_O(f)$ by default.", + "bbox": [ + 511, + 340, + 906, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 425, + 730, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By default, we use ViT-B [13] as the ReID backbone $f_{p}(\\cdot), f(\\cdot)$ . The images are resized to $256 \\times 128$ , and the data augmentation includes random crop and flip. For introducing variations in variation-guided identity shift, we use stronger augmentation derived from RandAugment [12]. We detailed the augmentation and conducted corresponding ablation studies in the Appendix. The AdamW [32] optimizer is adopted for De-ReID learning with an initial learning of 3e-4. The weight decay is set to 0. We empirically set trade-off parameters $\\lambda_{1} = 1.0$ and $\\lambda_{2} = 1.0$ . The batch size for accessible persons is 48, and that for unlearned persons is 32. The above hyper-parameters are the same in all experiments. $K$ in Eq. 5 is set to 20.", + "bbox": [ + 511, + 446, + 908, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison with Related Methods", + "text_level": 1, + "bbox": [ + 511, + 652, + 816, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Compared methods. Since this is the first work for De-ReID, we compare our method with existing methods potentially applicable to ReID. We pretrain a vanilla ViT-B in the pretraining subset of each dataset and then apply these methods to the ViT-B in the fine-tuning subset for De-ReID. We report the performance of the machine unlearning methods, including GS-LoRA [58], LIRF [53], SCRUB [24], and Boundary Shrink [8]. Besides, we implement a naive solution: label augmentation based on cameras, which assign different labels for images containing the same person but under different cameras. This naive solution is denoted as \"LabelAug\", and the triplet loss and cross-entropy loss [33] are applied for training LabelAug.", + "bbox": [ + 511, + 674, + 908, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notably, previous methods usually assume that the labeled data of accessible classes are available. Since the data", + "bbox": [ + 511, + 869, + 906, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29335", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/38a46ca88d855a74b2221e5a048f8196d554aaa27bb07c49524e610cf34fff33.jpg", + "table_caption": [ + "Table 2. Comparisons in the MSMT17 dataset. \"BS\" is the Boundary Shrink [8]. \"H\" denotes \"H-Mean\". R-1 $_T$ is the Rank-1 accuracy on the unlearned persons, and R-1 $_O$ is the Rank-1 accuracy on the accessible persons. R-1 $_T$ is expected to be low for unlearning specific persons, while a higher R-1 $_O$ and H is preferred. Since we apply additional augmentation for asymmetric contrastive learning, we further apply our augmentation to LIRF, denoted as LIRF*. \"RR\" is the Re-Ranking [63]. The method with the best/second-best H-Mean is marked in red/ blue. Blue background means the H-Mean are improved by Re-Ranking." + ], + "table_footnote": [], + "table_body": "
MethodMT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug54.673.838.156.873.532.4
+RR62.376.929.265.377.221.2
BS63.878.727.354.674.935.2
+RR70.082.218.362.578.725.3
SCRUB43.875.049.145.274.445.1
+RR52.378.441.353.778.236.6
LIRF10.062.566.218.554.656.8
+RR6.965.469.220.557.357.2
LIRF*18.568.865.135.169.752.8
GS-LoRA56.977.535.960.280.728.6
+RR56.979.736.264.182.923.2
Ours4.677.076.310.872.969.7
+RR4.679.077.310.475.170.9
", + "bbox": [ + 96, + 232, + 478, + 421 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ca81e20311d89bd7e516bce90d370f70b10123838f858a38758d1e995bcbbdac.jpg", + "table_caption": [ + "Table 3. Comparison results in the MSMT17 dataset when $M_T = 75$ and $M_T = 100$ . The notations are consistent with Table 2" + ], + "table_footnote": [], + "table_body": "
MethodMT=75MT=100
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug61.472.229.262.771.927.2
+ RR69.175.818.666.875.721.8
BS57.269.133.954.868.336.3
+ RR65.674.323.761.472.728.9
SCRUB54.774.837.550.567.840.6
+ RR62.979.027.755.772.835.9
LIRF18.651.956.120.849.753.8
+ RR19.655.157.524.852.453.5
LIRF*38.465.050.515.650.956.7
GS-LoRA62.479.528.462.176.128.3
+ RR65.381.824.563.878.626.2
Ours12.469.868.513.167.166.7
+ RR13.171.669.015.368.866.4
", + "bbox": [ + 96, + 467, + 478, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of accessible persons is unlabeled in our experiments, we apply our $L_{RCR}$ as a regularization for previous methods. We further apply our methods to advanced ReID models, including DCFormaler [27], PAT [28], and PHA [56].", + "bbox": [ + 88, + 686, + 482, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Results. Table 2 and Table 4 shows the comparison results. From the tables, we can observe that our method achieves the best performance in terms of H-Mean. Since the post-processing methods like Re-Ranking [63] can improve the results, we investigate whether these methods can reduce the unlearning effect when retrieving unlearned persons. The results show that existing unlearning methods usually get worse results after post-processing. For example, in Table 2, the Rank-1 accuracy of the SCRUB on unlearned persons is clearly increased from $43.8\\%$ to $52.3\\%$ when $M_T = 25$ ,", + "bbox": [ + 88, + 750, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e8d40e81a451f6a18177ea8e450de1154c318ff9386acafb18c5607d2a0c4d6f.jpg", + "table_caption": [ + "Table 4. Comparison results in the Market-1501 dataset." + ], + "table_footnote": [], + "table_body": "
MethodMT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug50.776.551.364.777.442.5
+RR49.377.652.866.078.941.3
BS57.373.144.562.072.244.3
+RR54.775.347.465.374.041.4
SCRUB50.776.051.254.770.250.4
+RR56.077.446.664.771.841.6
LIRF25.384.872.931.375.968.7
+RR26.784.471.932.775.767.7
LIRF*37.386.064.539.379.064.6
GS-LoRA56.091.748.965.391.443.7
+RR54.791.350.267.391.341.3
Ours10.791.184.412.084.483.2
+RR5.391.087.410.083.583.8
", + "bbox": [ + 517, + 104, + 903, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "demonstrating that the model does not effectively unlearn the persons and there are potential cues to re-identify them.", + "bbox": [ + 511, + 323, + 903, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While other methods achieve a worse H-mean after Re-ranking, our method can benefit from Re-Ranking [63] since the overall performance H-Mean is improved in most cases. For example, in the case of $M_T = 25$ in Market dataset, the H-Means of our method is increased to $87.4\\%$ from $84.4\\%$ .", + "bbox": [ + 511, + 354, + 906, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared with the other methods, Boundary Shrink changes the classification decision boundary of the unlearning classes, i.e., the unlearned persons in this work. However, the feature representation is the core for person image matching instead of the classifier. In contrast, our method directly optimizes the feature representation for De-ReID. As a result, our method outperforms Boundary Shrink by $30.4\\%$ H-Mean in MSMT17 with $M_T = 100$ .", + "bbox": [ + 511, + 431, + 908, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LIRF unlearns classes by constraining the attention map of intermediate features to be significantly different before and after unlearning. Since it directly constrains the features, it obtains the second-best results. However, the attention map has lots of knowledge about extracting identity information from images. Hence, their constraints largely hurt the ReID ability on accessible persons, and our method outperforms LIRF by $12.9\\%$ at H-Mean in MSMT17 when $M_T = 100$ . Besides, since we apply additional augmentation in variation-guided identity shift, we further validate the effectiveness of the additional augmentation by applying the augmentation in LIRF, denoted as $\\mathrm{LIRF^{*}}$ . However, the H-Mean is improved only in the case of $M_T = 100$ in MSMT17 and decreases in other cases, demonstrating that additional augmentation does not necessarily lead to higher performance.", + "bbox": [ + 511, + 551, + 908, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GS-LoRA effectively protects the ReID ability on accessible persons but falls behind in unlearning specific persons. LabelAug directly guides the model in distinguishing the images of unlearned persons under different camera views, but it has an unsatisfactory performance. We speculate it is because the model overfits the images in training and fails to unlearn the specific persons under cameras unseen in training. Differently, our method employs variation-guided", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29336", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/04b16c205a7d832acadd6a8880e67caa05d5902b9ae8d1456f996d4747ef3f80.jpg", + "table_caption": [ + "Table 5. Evaluation with different ReID models. Regardless of the ReID models, our method performs the best in terms of erasing the knowledge about unlearned persons (quantified by $\\mathrm{R - 1}_T$ ) and keeping ReID knowledge for accessible persons ( $\\mathrm{R - 1}_O$ ). \"Null\" indicates that no De-ReID method is applied." + ], + "table_footnote": [], + "table_body": "
ReID modelDe-ReID methodR-1T↓R-1O↑H↑
DCFormerNull81.284.4—
LabelAug44.154.046.2
GS-LORA75.381.916.4
LIRF18.051.057.7
Ours17.866.166.3
PATNull72.875.6—
LabelAug38.343.340.1
GS-LORA52.560.933.5
LIRF16.742.649.4
Ours12.055.059.0
PHANull70.278.7—
LabelAug44.351.441.2
GS-LORA56.468.333.6
LIRF23.650.552.7
Ours3.951.260.8
", + "bbox": [ + 96, + 160, + 460, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "identity shift $(L_{VIS}^{p}, L_{VIS}^{o})$ and guides the model to handle the variations adaptively. Fitting the variations in unlearned images helps the model bias to the environmental information when querying the unlearned persons. Simultaneously, the model is guided to be robust to the variations when querying accessible persons. As a result, the unlearned persons are shifted out of the view-invariant feature space while the accessible persons are retained.", + "bbox": [ + 88, + 383, + 483, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Evaluations on other ReID methods. Apart from the comparison on the vanilla ViT backbone, we also compared our method with related methods on the advanced ReID models [27, 28, 56]. The experimental results in Table 5 demonstrate that our method is superior to others regarding unlearning different ReID models.", + "bbox": [ + 88, + 505, + 483, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 604, + 243, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct experiments to show the effectiveness of each component in our design. Due to the space limit, we report the experimental results in Market and $M_T \\in \\{25, 100\\}$ in MSMT17 in Table 6, and the rest in Appendix. Regardless of $M_T$ , our designs clearly improve H-Mean that comprehensively evaluates the effectiveness of unlearning specific persons and ReID performance on accessible persons.", + "bbox": [ + 88, + 628, + 485, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- The effectiveness of the relation regularization. From Table 6, we first observe that the model suffers from catastrophic forgetting in ReID knowledge and collapses in the absence of $L_{RCR}$ . Although our variation-guided identity shift pulls closer the images and the augmented images for accessible persons, it mainly preserves the view-invariant feature embedding but it fails to preserve the feature discriminability, leading to model collapses. $L_{RCR}$ preserves the discriminability of features by preserving the relations between accessible persons.", + "bbox": [ + 88, + 734, + 485, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another constraint $L_{TRC}$ attempts to push away the im", + "bbox": [ + 109, + 885, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cd035e41188ebc4cd84a9291da29e332beb50061846b9431c2fd50bb51047064.jpg", + "table_caption": [ + "Table 6. Ablation study in MSMT17 dataset. \"W/o\" means \"without\", and other notations are the same as Table 2. $f_{p}(\\cdot)$ is the initial model for De-ReID learning. The full model achieves the best H-Mean. Notably, without $L_{RCR}$ , the model forgets ReID knowledge and collapses. \"SD\" denotes the \"self-augmented discrimination\"." + ], + "table_footnote": [], + "table_body": "
MethodMSMT17
MT=25MT=100
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
fp(·)80.385.4—79.585.4—
Components in Relation regularization
W/o LTRC5.573.474.118.365.363.2
W/o LRCR——————
Components in Variation-guided Identity Shift
W/o SD4.674.575.110.558.763.4
W/o LVIS10.075.872.918.063.262.3
W/o LVIS20.577.467.541.370.049.4
Full model4.677.076.313.167.166.7
", + "bbox": [ + 517, + 160, + 903, + 314 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/29ec98d02d41136121b9c06c21930e9552b97e14eaef565a79d5b7a95ce691a2.jpg", + "table_caption": [ + "Table 7. Ablation study in Market-1501." + ], + "table_footnote": [], + "table_body": "
MethodMarket-1501
MT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
fp(·)89.396.3—94.096.3—
Components in Relation regularization
W/o LTRC20.089.778.224.078.373.9
W/o LRCR——————
Components in Variation-guided Identity Shift
W/o SD24.090.375.811.379.781.2
W/o LVIS10.784.581.412.777.379.2
W/o LVIS46.793.958.655.389.954.1
Full model10.791.184.412.084.483.2
", + "bbox": [ + 517, + 342, + 901, + 497 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/551864891eb9ef9188b260dba7e6011342b35d93a9366987e5ab409eb8f282aa.jpg", + "image_caption": [ + "Figure 3. Evaluation results under various $\\lambda_{1}$ and $\\lambda_{2}$ ." + ], + "image_footnote": [], + "bbox": [ + 521, + 510, + 702, + 614 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e08e91c4565349eaf9d804561e1689c5da0f9716d202fa83d9f13f6c858ed308.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 508, + 893, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ages between unlearned persons to prevent them from being cross-view retrieved, referring to the distance between accessible persons and unlearned persons. We observe a degraded H-Mean in the absence of $L_{TRC}$ in all cases. For example, when deprecating $L_{TRC}$ in the case of $M_T = 50$ in Market, the Rank-1 accuracy of the unlearned persons is increased to $24.0\\%$ from $12.0\\%$ , meaning that more unlearned person images are correctly matched and is not expected. As a result, the H-Mean drops by $9.8\\%$ compared with the full model. In the case of $M_T = 100$ in MSMT17, the Rank-1 accuracy of the unlearned persons increases to $18.3\\%$ from $13.1\\%$ without $L_{TRC}$ , and the H-Mean drops by $3.3\\%$ .", + "bbox": [ + 511, + 643, + 906, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The hyper-parameter $\\lambda_{2}$ controls the weight of $L_{TRC}$ . We evaluate our model under different $\\lambda_{2}$ and show the performance in Fig. 3. When $\\lambda_{2} = 0$ , $L_{TRC}$ is not employed and the performance is inferior. As $\\lambda_{2}$ increases, the performance is improved. However, $L_{TRC}$ should not be set to", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29337", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "too large to avoid overwhelming other learning objectives.", + "bbox": [ + 89, + 90, + 477, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- The effectiveness of the variation-guided identity shift. The variation-guided identity shift (VIS) emphasizes the style difference between images of unlearned persons in the feature space while keeping the other images robust to the style variations. Overall, comparing the results of \"W/o $L_{VIS}$ \" and the full model in Table 6, we observe that the ReID model struggles to unlock the specific persons without $L_{VIS}$ . For example, in MSMT17 dataset, the Rank-1 accuracy on unlearned persons is $20.5\\% / 41.3\\%$ when $M_T = 25 / 100$ , which is $15.9\\% / 28.2\\%$ larger than the full model. This is because ReID requires view-invariant features and our model correspondingly guides the model to be biased to encode environmental information for images of unlearned persons. As a result, we effectively prevent the images of unlearned persons from being re-identified. The hyperparameter $\\lambda_1$ controls the weight of $L_{VIS}$ in optimization. We evaluate our model under different $\\lambda_1$ and show the performance in Fig. 3.", + "bbox": [ + 88, + 108, + 483, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also studied the effectiveness of the designs in VIS. As discussed in the introduction, we argue that simply pushing away the images under different camera views may learn the spurious cues for De-ReID, and therefore we push away the image and the corresponding augmented image in our VIS. We refer to this as \"self-augmented discrimination\". To verify its effectiveness, we conduct an experiment that pushes away the image and the augmented image from another image, denoted as \"W/o SD\", and observe a significant performance drop. For example, in the case of $M_T = 25$ in Market dataset, the Rank-1 accuracy on the unlearned person increases to $24.0\\%$ from $10.7\\%$ without SD, and the H-Mean drops by $8.6\\%$ . In other cases, the H-Mean drops without using the SD, demonstrating the effectiveness of SD.", + "bbox": [ + 89, + 381, + 483, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, we observe that guiding the ReID to overcome the variations for accessible persons is important for balancing the ReID ability and De-ReID of unlearned persons. Deprecating the $L_{VIS}^{o}$ leads to $3\\% \\sim 4\\%$ performance drops in H-Mean in different cases, and reduce the protection effect since the rank-1 accuracy of unlearned persons increases.", + "bbox": [ + 89, + 595, + 483, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- The impact of the unlabeled sample. We evaluate our method in MSMT17 under different ratios between unlabeled images of accessible persons and labeled images of unlearned persons in the fine-tuning subset. The results are in Table 8.", + "bbox": [ + 86, + 686, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the results, we obtain two conclusions. (i) More unlabeled data is helpful for both unlearning specific persons and keeping ReID performance on the accessible persons. When the number of accessible persons decreases, the Rank-1 accuracy of the unlearned persons (denoted as $\\mathrm{R - 1}_T$ ) increases, which means the unlearned persons are easier to retrieve by the ReID model. Simultaneously, the Rank-1 accuracy of the accessible persons $(\\mathrm{R - 1}_O)$ decreases, which means the ReID model cannot properly re-identify", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Evaluations under different numbers of accessible persons in MSMT17 when $M_T = 100$ . $M_O$ is the number of accessible persons in the fine-tuning subset. 'With LUP' means using the images in LUPerson dataset as auxiliary unlabeled data and the number in parentheses denotes the performance gain. By default, we do not use LUPerson dataset.", + "bbox": [ + 511, + 89, + 906, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c3518ce90fde570a59130552ede8203843210cd068cb4d5c9e9cd1c8ae677b4d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
\\(M_O\\)Without LUPWith LUP
R-1T ↓R-1O ↑H ↑R-1T ↓R-1O ↑H ↑
14113.167.166.7---
10015.861.362.514.168.266.8 (+4.3%)
8015.854.358.615.467.765.9 (+7.3%)
6017.451.457.615.465.664.8 (+7.2%)
", + "bbox": [ + 517, + 174, + 901, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4977e97c7dbfa613471b79774d592254c08161398b9f7df3bf13ebfe2706ea62.jpg", + "image_caption": [ + "Images of unlearned persons", + "Figure 4. Attention map on images of the unlearned persons and the images of accessible persons." + ], + "image_footnote": [], + "bbox": [ + 514, + 265, + 570, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6adc81d0df95302d4d74e97fe7c9b2f8002987382ff8cf6e4b05fb59428fbb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 265, + 633, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of unlearn", + "bbox": [ + 581, + 347, + 633, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1a2ae0c4b8e43b6d2d5290d6024c57328cc196e4fe257c82da6c0cbe9beeed59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 265, + 699, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "d persons", + "bbox": [ + 645, + 347, + 699, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/455732697bc2e7e6a864634da6bd8bcaf18ead8c9d8a3e86953cb41141430d0f.jpg", + "image_caption": [ + "Images of other persons" + ], + "image_footnote": [], + "bbox": [ + 720, + 265, + 772, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/610f1c468e76285bcf022bb8c83b364dabb4dab798967780278a4e75d7097c33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 265, + 864, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rsons", + "bbox": [ + 852, + 347, + 883, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the accessible persons. (ii) Although the performance of our method decreases when the unlabeled images become fewer, our method can effectively utilize the publicly available unlabeled person images for unlearning, demonstrating the scalability of our method. When using additional unlabeled images from LUPerson dataset, our method can achieve similar performance even though the number of accessible persons in MSMT17 decreases from 141 to 60.", + "bbox": [ + 511, + 400, + 906, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 511, + 532, + 650, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize the attention map on different persons in the testing set in Fig. 4. From the figure, for the images of unlearned persons, we can observe that the model's attention mainly focus on the background which contains abundant environmental information. At the same time, the model pays diverse attention to the informative parts of images for extracting discriminative features for accessible persons. These results demonstrate the effectiveness of our method.", + "bbox": [ + 511, + 556, + 906, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 693, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address the privacy concerns in person ReID, we pioneer the exploration of the De-ReID problem that guides the ReID model to forget the knowledge about matching specific persons. We formulate a framework and propose a novel variation-guided identity shift method for De-ReID. Our method unlearns the specific person shifting them out of the view-invariant feature space and simultaneously keeps the ReID knowledge by overcoming the cross-view variations for accessible persons. We further introduce relation regularization to characterize the desired properties of De-ReID. Extensive experiments conducted on a new benchmark show the superiority of our method.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29338", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 89, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported partially by NSFC(U21A20471, 92470202), by the National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025)", + "bbox": [ + 89, + 114, + 485, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Prachi Agrawal and P. J. Narayanan. Person de-identification in videos. IEEE Transactions on Circuits and Systems for Video Technology, 2011. 2", + "[2] Shafiq Ahmad, Pietro Morerio, and Alessio Del Bue. Person re-identification without identification via event anonymization. In 2023 IEEE/CVF International Conference on Computer Vision, 2023. 2", + "[3] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In IEEE Symposium on Security and Privacy, 2021. 1, 2", + "[4] Karla Brkic, Ivan Sikiric, Tomislav Hrkac, and Zoran Kalafatic. I know that person: Generative full body and face de-identification of people in images. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2", + "[5] Jonathan Brophy and Daniel Lowd. Machine unlearning for random forests. In International Conference on Machine Learning, 2021. 1, 2", + "[6] Chengzhi Cao, Xueyang Fu, Hongjian Liu, Yukun Huang, Kunyu Wang, Jiebo Luo, and Zheng-Jun Zha. Event-guided person re-identification via sparse-dense complementary learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[7] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Learning invariance from generated variance for unsupervised person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[8] Min Chen, Weizhuo Gao, Gaoyang Liu, Kai Peng, and Chen Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 5, 6", + "[9] Tianshui Chen, Tao Pu, Lingbo Liu, Yukai Shi, Zhijing Yang, and Liang Lin. Heterogeneous semantic transfer for multi-label recognition with partial labels. International Journal of Computer Vision, 2024. 1", + "[10] Weihua Chen, Xiaotang Chen, Jianguo Zhang, and Kaiqi Huang. Beyond triplet loss: A deep quadruplet network for person re-identification. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, 2017. 1, 2", + "[11] Dasol Choi and Dongbin Na. Towards machine unlearning benchmarks: Forgetting the personal identities in facial recognition systems. In AAAI 2024 workshop on Privacy-Preserving Artificial Intelligence, 2023. 3", + "[12] Ekin D. Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le. Randaugment: Practical automated data augmentation" + ], + "bbox": [ + 93, + 213, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with a reduced search space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 5", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5", + "[14] Lijie Fan, Tianhong Li, Rongyao Fang, Rumen Hristov, Yuan Yuan, and Dina Katabi. Learning longterm representations for person re-identification using radio signals. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[15] Ivan Filković, Zoran Kalafatić, and Tomislav Hrkač. Deep metric learning for person re-identification and de-identification. In 2016 39th International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO), 2016. 2", + "[16] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[17] Wenxuan Guo, Zhiyu Pan, Yingping Liang, Ziheng Xi, Zhi Chen Zhong, Jianjiang Feng, and Jie Zhou. Lidar-based person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[18] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. Advances in Neural Information Processing Systems, 2021. 2", + "[19] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In IEEE/CVF International Conference on Computer Vision, 2021. 1, 2", + "[20] Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qing-song Xie, Yizhou Wnag, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, Donglian Qi, and Yunfeng Yan. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[21] Bingliang Jiao, Lingqiao Liu, Liying Gao, Guosheng Lin, Lu Yang, Shizhou Zhang, Peng Wang, and Yanning Zhang. Dynamically transformed instance normalization network for generalizable person re-identification. In European Conference on Computer Vision, 2022. 2", + "[22] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Andy J Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multiscale dilated transformer for visual recognition. IEEE Transactions on Multimedia, 25:8906-8919, 2023. 2", + "[23] Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Cloth-changing person re-identification from a single image with gait prediction and regularization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29339", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Triantafillou. Towards unbounded machine unlearning. In Conference on Neural Information Processing Systems, 2023. 1, 2, 5", + "[25] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: Exploiting vision-language model for image re-identification without concrete text labels. Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2", + "[26] Weijia Li, Saihui Hou, Chunjie Zhang, Chunshui Cao, Xu Liu, Yongzhen Huang, and Yao Zhao. An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[27] Wen Li, Cheng Zou, Meng Wang, Furong Xu, Jianan Zhao, Ruobing Zheng, Yuan Cheng, and Wei Chu. Dc-former: Diverse and compact transformer for person re-identification. In AAAI Conference on Artificial Intelligence, 2023. 6, 7", + "[28] Yulin Li, Jianfeng He, Tianzhu Zhang, Xiang Liu, Yongdong Zhang, and Feng Wu. Diverse part discovery: Occluded person re-identification with part-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 6, 7", + "[29] Kun-Yu Lin, Jiaming Zhou, Yukun Qiu, and Wei-Shi Zheng. Adversarial partial domain adaptation by cycle inconsistency. In European Conference on Computer Vision, pages 530-548. Springer Nature Switzerland Cham, 2022. 2", + "[30] Kun-Yu Lin, Jia-Run Du, Yipeng Gao, Jiaming Zhou, and Wei-Shi Zheng. Diversifying spatial-temporal perception for video domain generalization. Advances in Neural Information Processing Systems, 36:56012–56026, 2023. 2", + "[31] Kun-Yu Lin, Henghui Ding, Jiaming Zhou, Yu-Ming Tang, Yi-Xing Peng, Zhilin Zhao, Chen Change Loy, and Wei-Shi Zheng. Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560, 2024. 2", + "[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5", + "[33] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019. 5", + "[34] Jiaxu Miao, Yu Wu, Ping Liu, Yuhang Ding, and Yi Yang. Pose-guided feature alignment for occluded person re-identification. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 542–551, 2019. 5", + "[35] Tao Pu, Tianshui Chen, Hefeng Wu, Yukai Shi, Zhijing Yang, and Liang Lin. Dual-perspective semantic-aware representation blending for multi-label image recognition with partial labels. Expert Systems with Applications, page 123526, 2024. 1", + "[36] Tao Pu, Qianru Lao, Hefeng Wu, Tianshui Chen, Ling Tian, Jie Liu, and Liang Lin. Category-adaptive label discovery and noise rejection for multi-label recognition with partial positive labels. IEEE Transactions on Multimedia, pages 1-12, 2024. 2" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 2021. 2", + "[38] Takashi Shibata, Go Irie, Daiki Ikami, and Yu Mitsuzumi. Learning with selective forgetting. In International Joint Conference on Artificial Intelligence, 2021. 5", + "[39] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and A strong convolutional baseline). In European Conference on Computer Vision, 2018. 2", + "[40] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[41] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. Learning to imagine: Diversify memory for incremental learning using unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9549-9558, 2022. 2", + "[42] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. When prompt-based incremental learning does not meet strong pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1706-1716, 2023. 2", + "[43] Yu-Ming Tang, Yi-Xing Peng, Jingke Meng, and Wei-Shi Zheng. Rethinking few-shot class-incremental learning: Learning from yourself. In European Conference on Computer Vision, pages 108-128. Springer, 2024. 1", + "[44] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACM International Conference on Multimedia, 2018. 2", + "[45] Haochen Wang, Jiayi Shen, Yongtuo Liu, Yan Gao, and Efstratos Gavves. Nformer: Robust person re-identification with neighbor transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2", + "[46] Lin Wang, Wanqian Zhang, Dayan Wu, Fei Zhu, and Bo Li. Attack is the best defense: Towards preemptive-protection person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2", + "[47] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 79–88, 2018. 5", + "[48] Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. A discriminative feature learning approach for deep face recognition. In European Conference on Computer Vision, 2016. 2", + "[49] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 5", + "[50] Bin Yang, Jun Chen, and Mang Ye. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In IEEE/CVF International Conference on Computer Vision, 2023. 2" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29340", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Bin Yang, Jun Chen, and Mang Ye. Shallow-deep collaborative learning for unsupervised visible-infrared person re-identification. In IEEE/CVF International Conference on Computer Vision, 2024. 2", + "[52] Jinrui Yang, Jiawei Zhang, Fufu Yu, Xinyang Jiang, Mengdan Zhang, Xing Sun, Ying-Cong Chen, and Wei-Shi Zheng. Learning to know where to see: A visibility-aware approach for occluded person re-identification. In IEEE/CVF international conference on computer vision, 2021. 2", + "[53] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In European Conference on Computer Vision, 2022. 1, 3, 5", + "[54] Mang Ye, Wei Shen, Junwu Zhang, Yao Yang, and Bo Du. Securereid: Privacy-preserving anonymization for person re-identification. IEEE Transactions on Information Forensics and Security, 2024. 2", + "[55] Mang Ye, Zesen Wu, Cuiqun Chen, and Bo Du. Channel augmentation for visible-infrared re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[56] Guiwei Zhang, Yongfei Zhang, Tianyu Zhang, Bo Li, and Shiliang Pu. Pha: Patch-wise high-frequency augmentation for transformer-based person re-identification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6, 7", + "[57] Junwu Zhang, Mang Ye, and Yao Yang. Learnable privacy-preserving anonymization for pedestrian images. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2", + "[58] Hongbo Zhao, Bolin Ni, Haochen Wang, Junsong Fan, Fei Zhu, Yuxi Wang, Yuntao Chen, Gaofeng Meng, and Zhaoxiang Zhang. Continual forgetting for pre-trained vision models. 2024. 3, 5", + "[59] Zhilin Zhao, Longbing Cao, and Kun-Yu Lin. Revealing the distributional vulnerability of discriminators by implicit generators. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8888-8901, 2022. 2", + "[60] Zhilin Zhao, Longbing Cao, and Kun-Yu Lin. Out-of-distribution detection by cross-class vicinity distribution of in-distribution data. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2", + "[61] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Computer Vision, IEEE International Conference on, 2015. 5", + "[62] Wei-Shi Zheng, Jincheng Hong, Jiening Jiao, Ancong Wu, Xiatian Zhu, Shaogang Gong, Jiayin Qin, and Jianhuang Lai. Joint bilateral-resolution identity modeling for crossresolution person re-identification. International Journal of Computer Vision, 2022. 2", + "[63] Zhun Zhong, Liang Zheng, Donglin Cao, and Shaozi Li. Re-ranking person re-identification with k-reciprocal encoding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 6" + ], + "bbox": [ + 91, + 92, + 485, + 868 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29341", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/1180818e-c94d-44fa-b5df-b1fc8919ac06_model.json b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/1180818e-c94d-44fa-b5df-b1fc8919ac06_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fe01699cb5bc7c6256fcf4a8f0aba89d1ba76d7d --- /dev/null +++ b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/1180818e-c94d-44fa-b5df-b1fc8919ac06_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.857, + 0.154 + ], + "angle": 0, + "content": "Person De-reidentification: A Variation-guided Identity Shift Modeling" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.179, + 0.805, + 0.217 + ], + "angle": 0, + "content": "Yi-Xing Peng\\(^{1,2,4}\\) Yu-Ming Tang\\(^{1,4}\\) Kun-Yu Lin\\(^{1,4}\\) Qize Yang\\(^{3}\\) Jingke Meng\\(^{1,4*}\\) Xihan Wei\\(^{3}\\) Wei-Shi Zheng\\(^{1,2,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.217, + 0.81, + 0.234 + ], + "angle": 0, + "content": "1School of Computer Science and Engineering, Sun Yat-sen University, China;" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.234, + 0.79, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)Peng Cheng Laboratory, Shenzhen, China; \\(^{3}\\)Tongyi Lab, Alibaba Group;" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.252, + 0.888, + 0.27 + ], + "angle": 0, + "content": "4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China." + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.273, + 0.781, + 0.287 + ], + "angle": 0, + "content": "{pengyx23,tangym9,linky5}@mail2.sysu.edu.cn wszheng@ieee.org" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.486, + 0.763 + ], + "angle": 0, + "content": "Person re-identification (ReID) is to associate images of individuals from different camera views against cross-view variations. Like other surveillance technologies, Re-ID faces serious privacy challenges, particularly the potential for unauthorized tracking. Although various tasks (e.g., face recognition) have developed machine unlearning techniques to address privacy concerns, such methods have not yet been explored within the Re-ID field. In this work, we pioneer the exploration of the person de-reidentification (De-ReID) problem and present its inherent challenges. In the context of ReID, De-ReID is to unlock the knowledge about accurately matching specific persons so that these \"unlearned persons\" cannot be re-identified across cameras for privacy guarantee. The primary challenge is to achieve the unlearning without degrading the identity-discriminative feature embeddings to ensure the model's utility. To address this, we formulate a De-ReID framework that utilizes a labeled dataset of unlearned persons for unlearning and an unlabeled dataset of accessible persons for knowledge preservation. Instead of unlearning based on (pseudo) identity labels, we introduce a variation-guided identity shift mechanism that unlearns the specific persons by fitting the variations in their images while preserving ReID ability on other persons by overcoming the variations in images of accessible persons. As a result, the model shifts the unlearned persons to a feature space that is vulnerable to cross-view variations. Extensive experiments on benchmarks demonstrate the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.791, + 0.222, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Person re-identification (ReID) aims to match the images of the same person across different camera views based on feature similarity. While deep learning [9, 35, 43] has significantly enhanced the performance of the ReID mod" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.322, + 0.9, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.488 + ], + "angle": 0, + "content": "Figure 1. Person de-reidentification (De-ReID) prevents intelligent surveillance from tracking unlearned persons. For example, when using surveillance models for security (track and record the person in the red box), the residents in the community and the important customers in business (in the green box) should not be tracked for privacy and commercial secrets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.909, + 0.627 + ], + "angle": 0, + "content": "els [10, 19, 45], there are growing public concerns about potential misuse, leading to ethical and societal implications. To avoid misuse and mitigate the potential privacy concerns, the research in surveillance has begun exploring machine unlearning techniques that limit the operational scope of these models. For instance, in face recognition, models are adjusted to \"unlearn\" individuals who raise privacy concerns, as mandated by regulations like the EU's General Data Protection Regulation (GDPR) [3, 5, 24, 53]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.84 + ], + "angle": 0, + "content": "However, similar constraints on the ReID models have not yet to be explored, despite its significance for real-world applications. For instance, the residents in the neighborhood prefer to constrain the surveillance in the community to only focus on external people and not track themselves; in working buildings, security cameras should only be set for monitoring and preventing theft or accidents, not for tracking or identifying specific employees during their routine work; and homeowners expect smart home devices to track only unknown individuals for security purposes while avoiding any continuous surveillance of themselves; To meet these requirements, there is an urgent need to develop methods that limit the scope of ReID models, enabling them to \"ignore\" specific individuals during deployment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this work, we pioneer the exploration of person deidentification (De-ReID) problem: limiting the ReID scope by unlearning specific persons (referred to as \"unlearned persons\") in the ReID model while maintaining the ReID" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29331" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.364 + ], + "angle": 0, + "content": "performance on other persons (referred to as \"accessible persons\"). In the context of image matching, a person is unlearned by a model means the model can hardly associate the images of the person under different camera views, as shown in Fig. 1. Forming a De-ReID model is challenging because only the knowledge strictly related to matching the unlearned persons should be eliminated, without destroying the identity-discriminative feature embeddings. A trivial solution is to distinguish the images of unlearned persons under different camera views for unlearning while associating the images of accessible persons based on their identity labels to preserve ReID knowledge. However, due to the large network capacity, the model may learn spurious cues or overfit the images of unlearned persons in the training set, leading to poor generalization during evaluation. Additionally, collecting identity labels for accessible persons to preserve ReID knowledge is both costly and difficult to scale, especially when the number of cameras is large [29, 31, 36]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.486, + 0.577 + ], + "angle": 0, + "content": "To handle the De-ReID, we formulate a weakly-supervised unlearning framework where a labeled image set of unlearned persons and an unlabeled dataset of accessible persons are available to adapt a well-trained ReID model to De-ReID. Given that the core of ReID is to overcome cross-view variations, such as different camera viewing angles and lighting conditions, we propose variation-guided identity shift that learns De-ReID based on the variations in images. The core of our method is to guide the ReID model bias towards environmental information (e.g., lighting) when encountering unlearned persons, which shifts the unlearned persons out of the view-invariant feature space. At the same time, we keep the model remains robust to cross-view variations when querying accessible persons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.486, + 0.791 + ], + "angle": 0, + "content": "Specifically, for unlearned persons, we guide the model to distinguish their images and the images augmented by variations. Compared with distinguishing images under different camera views, learning to distinguish image and its augmentation results is more effective for unlearning since the augmented images are highly similar to the original image (e.g., body shape) but varies in style information (e.g., brightness and color tone). Therefore, the model is biased to encode the style information to push away these images. In contrast, for accessible persons, we pull closer their images and augmented versions in the feature space, expecting the model to overcome the variations when matching them. Moreover, we introduce relation regularization to characterize the desired properties of De-ReID in terms of relations." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.794, + 0.449, + 0.808 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.811, + 0.482, + 0.839 + ], + "angle": 0, + "content": "- We are the first to investigate the challenging De-ReID task and conduct a new benchmark." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.841, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We formulate a De-ReID framework and propose variation-guided identity shift that achieves De-ReID by fitting the variations in images of interest while against the variations in others to preserve ReID knowledge." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.909, + 0.168 + ], + "angle": 0, + "content": "- To further utilize the identity annotations in unlearned person images for learning, we propose relation regularization that characterizes the desired properties of ReID and De-ReID. Both qualitative and quantitative results demonstrate the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.181, + 0.665, + 0.197 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.908, + 0.342 + ], + "angle": 0, + "content": "- Person re-identification. Based on deep learning, advanced ReID models automatically learn a discriminative feature embedding from a large amount of data [20]. Various network architectures [19, 21, 22, 30, 39, 42, 44, 45] and loss functions [10, 40, 41, 48] are developed to enhance feature representation learning. Besides, the challenges in ReID including cross-modality [25, 31, 51, 55], cloth-changing [23, 26], low-resolution [62], occlusion [52], and unsupervised ReID [7, 50] are also widely studied." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.909, + 0.448 + ], + "angle": 0, + "content": "- Privacy protection for person images. Apart from pursuing high performance, protecting human privacy in human-centric tasks is also important [54, 59, 60]. A stream of research focuses on generating anonymized images by removing or obscuring identity-related information, such as faces and silhouettes, while preserving the image's utility for subsequent tasks, such as action recognition [1, 4, 15]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.909, + 0.599 + ], + "angle": 0, + "content": "In the field of ReID, Ye et al. [54, 57] proposed a generative model to anonymize the person images to avoid privacy leakage while keeping the ReID utility. Wang et al. [46] add imperceptible perturbations to the images of in the database to prevent malicious use of the images. Particularly, one stream of works explores person de-identification [1, 4, 15], which obscures the identity information in images or videos, including face and silhouette. The new sensors including event cameras and LiDAR are also investigated for privacy-preserving ReID [2, 6, 14, 17]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.811 + ], + "angle": 0, + "content": "Despite their progress, previous works mainly focus on preventing the malicious use of images while we aim to prevent the malicious use of models. Technically, previous works [54, 57] learn to anonymize person images and jointly adapt the ReID model to ensure ReID utility. Their methods cannot prevent the ReID model from retrieving specific persons, as well as the methods that explore new sensors such as event cameras for ReID [2, 6, 14, 17]. Besides, although PRIDE [46] protects the images in the database from being maliciously used, the attacker can also collect images that are not in the database to track specific persons. Moreover, some works in person de-identification only concern the identity-unrelated utility [1, 4, 15], such as action recognition, while De-ReID requires identifying accessible persons." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.886 + ], + "angle": 0, + "content": "- Machine unlearning. De-ReID has a similar goal to machine unlearning. Machine unlearning is first proposed to eliminate the influence of specific data on the deep model as if the model never uses these data for training, which enables users to erase their personal data [3, 5, 16, 18, 37]." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.887, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Kurmanji et al. [24] further apply machine unlearning for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29332" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.302 + ], + "angle": 0, + "content": "various purposes such as removing the bias from the model or eliminating the negative effects of mislabeled samples. To maximize the error in forgetting classes, they maximize the difference between the outputs of the learned model and the original model in the forgetting samples. Choi et al. [11] propose to jointly maximize the loss on the forgetting data and learn on the other data. Chen et al. [8] unlearn specific classes by adjusting their decision boundary. The forgetting samples are assigned a nearest but incorrect label for boundary shrink. Zhao et al. [58] propose GS LoRA to maintain performance on remaining classes when maximizing the loss on the forgetting data. Ye et al. [53] enforce that the model should be different from the original model in terms of the attention map of intermediate features to forget classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.485 + ], + "angle": 0, + "content": "Our work is different from existing works. Firstly, even if the model does not see the unlearned person in training, the ReID model can effectively match the cross-view images of unlearned persons. Hence, the methods eliminating the influence of samples on network weights are not applicable to De-ReID. For the methods that maximize the errors in the forgetting classes, they mainly focus on classification tasks. Differently, ReID pursues discriminative features for matching. Even if the unlearned persons cannot be correctly classified as their methods desired, the model could retrieve the correct images as long as the feature distances between the query and the correct images are smaller than others." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.273, + 0.516 + ], + "angle": 0, + "content": "3. De-ReID Learning" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.483, + 0.631 + ], + "angle": 0, + "content": "- Problem formulation. Given a well-trained ReID model denoted as \\( f_{p}(\\cdot) \\) and a set of unlearned persons \\( \\mathcal{P}_u \\), our goal is to learn a model \\( f(\\cdot) \\) that the model cannot match the images of unlearned persons while keeping the ReID performance for other persons that are accessible to model. \\( f(\\cdot) \\) is initialized as \\( f_{p}(\\cdot) \\), and \\( f_{p}(\\cdot) \\) is kept unchanged. An ideal De-ReID model \\( f(\\cdot) \\) is expected to satisfy the conditions:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.642, + 0.484, + 0.694 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\phi_ {f} (x, x _ {p}) < \\phi_ {f} (x, x _ {n}), \\forall x \\in \\mathcal {P} _ {u}, \\forall x _ {n} \\text {s . t .} y _ {n} \\neq y, \\\\ \\phi_ {f} (x, x _ {p}) > \\phi_ {f} (x, x _ {n}), \\forall x \\notin \\mathcal {P} _ {u}, \\forall x _ {n} \\text {s . t .} y _ {n} \\neq y, \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.483, + 0.755 + ], + "angle": 0, + "content": "where \\(\\phi_f(\\cdot, \\cdot)\\) is the similarity metric in the feature space of \\(f(\\cdot)\\). \\(x\\) is a query image with label \\(y\\), and likewise \\(y_p\\) and \\(y_n\\) are labels of \\(x_p\\) and \\(x_n\\). \\(x_p\\) is a different image containing the same person with \\(x\\), but \\(x_n\\) contains a different person." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To learn \\( f(\\cdot) \\) for approximating Eq. 1, we assume a labeled image set of \\( M_T \\) unlearned persons denoted as \\( S_T = \\{(x_i^t,y_i^t)\\}_{i = 1}^{|S_T|} \\) and an unlabeled image set of accessible persons denoted as \\( S_O = \\{x_j^o\\}_{j = 1}^{|S_O|} \\) are available for fine-tuning. \\( x_{i}^{t} \\) is the \\( i \\)-th image in \\( S_T \\) with corresponding label \\( y_{i}^{t}\\in \\{1,2,\\dots,M_{T}\\} \\). \\( x_{j}^{o} \\) is the \\( j \\)-th image in \\( S_O \\). The persons in \\( S_O \\) are different from those in \\( S_T \\) and the persons in testing. While unlearning the specific persons based on the labeled data \\( S_T \\) seems straightforward (e.g., by max" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "imizing the classification error), it is hard to ensure only the knowledge strictly subject to matching the unlearned persons is eliminated, rather than simply destroy identity discriminative feature embedding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.908, + 0.243 + ], + "angle": 0, + "content": "Considering the ubiquitous existence of cross-view variations, we propose variation-guided identity shift based on variations in images to unlearn the specific person in view-invariant feature space. Besides, we formulate the goals of De-ReID in the form of images relations and present relation regularization. An overview of our method is in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.252, + 0.787, + 0.268 + ], + "angle": 0, + "content": "3.1. Variation-guided Identity Shift" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.453 + ], + "angle": 0, + "content": "Intuitively, person ReID requires a view-invariant feature space, where image features are identity discriminative regardless of camera views. Hence, we propose to shift the unlearned persons out of the view-invariant feature space, and therefore their images from different camera views cannot be properly associated. To this end, we guide the model to fit the variations in images of unlearned persons and overcome the variations for accessible persons, forming asymmetric learning objectives. In this way, the images of unlearned persons will be encoded into distinct features affected by the variations in images while the images of accessible persons are encoded as view-invariant features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.909, + 0.742 + ], + "angle": 0, + "content": "To introduce abundant variations, we employ augmentation function \\(\\mathcal{T}(\\cdot)\\) to obtain an augmented view \\(\\hat{x}_i^t = \\mathcal{T}(x_i^t)\\) for each image \\(x_i^t\\). To simulate the natural cross-view variations, \\(\\mathcal{T}(\\cdot)\\) should not change the identity of \\(x_i^t\\) but properly adjust the images' color, brightness, and so forth, ensuring that \\(\\hat{x}_i^t\\) contains the same person with \\(x_i^t\\) but is different in image style. We ablate the design of \\(\\mathcal{T}(\\cdot)\\) in experiments in our Appendix. To shift the unlearned persons out of the view-invariant feature space, the model \\(f(\\cdot)\\) should prioritize rich environmental information unrelated to identity when extracting features from the unlearned person's image \\(x_i^t\\). In other words, the feature \\(f(x_i^t)\\) should encode rich environmental information (e.g., background, image style) that cannot be applied for matching unlearned person in other cameras. To this end, we guide \\(f(\\cdot)\\) to push away \\(x_i^t\\) and its augmented view \\(\\hat{x}_i^t\\) in the feature space, which encourages the model to fit the variations for images of the unlearned person. Formally, the learning objective for shifting the unlearned persons out of view-invariant feature space is:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.752, + 0.907, + 0.771 + ], + "angle": 0, + "content": "\\[\nL _ {V I S} ^ {p} = \\left[ \\sigma_ {c} - \\| f \\left(x _ {i} ^ {t}\\right) - f \\left(\\hat {x} _ {i} ^ {t}\\right) \\| _ {2} ^ {2} \\right] _ {+}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.854 + ], + "angle": 0, + "content": "where \\([a]_{+} = \\max(a, 0.0)\\) and \\(\\sigma_{c}\\) is the margin to bound the loss. Through optimization, the features of unlearned person images become image-specific and cannot be used for cross-view matching since the features are severely affected by the variations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Simply learning with Eq. 2 alone will cause the model to forget the ReID knowledge, and fail to extract features robust to variations for the accessible persons. It will cause" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29333" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.089, + 0.822, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.277, + 0.908, + 0.335 + ], + "angle": 0, + "content": "Figure 2. An overview of our method. In variation-guided identity shift, the model is guided to adapt to the variations when encountering images of unlearned persons, thereby shifting them out of the view-invariant feature space. Meanwhile, the view-invariant feature space is maintained for other accessible persons by guiding the model to overcome the variations when encountering images of these individuals. Relation regularization formulates the desired properties of De-ReID in terms of image relations and ensures the feature discriminability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.472 + ], + "angle": 0, + "content": "feature space collapse instead of the specific identity shift. To maintain a view-invariant feature space for re-identifying accessible persons, we guide the model to overcome the variations for images of these persons. For image \\( x_{j}^{o} \\) of an accessible person, we obtain its augmented view \\( \\hat{x}_{j}^{o} = \\mathcal{T}(x_{j}^{o}) \\). We encourage the model to pull closer \\( x_{j}^{o} \\) and \\( \\hat{x}_{j}^{o} \\) to remind the model of extracting robust features for accessible persons. Formally, the learning function for accessible person is:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.481, + 0.484, + 0.502 + ], + "angle": 0, + "content": "\\[\nL _ {V I S} ^ {o} = \\left\\| f \\left(x _ {j} ^ {o}\\right) - f \\left(\\hat {x} _ {j} ^ {o}\\right) \\right\\| _ {2} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.486, + 0.558 + ], + "angle": 0, + "content": "Combining Eq. 3 and Eq. 2, we form an asymmetric objective to shift the identities of unlearned persons while preserving a view-invariant feature space for accessible persons:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.573, + 0.484, + 0.59 + ], + "angle": 0, + "content": "\\[\nL _ {V I S} = L _ {V I S} ^ {p} + L _ {V I S} ^ {o}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.779 + ], + "angle": 0, + "content": "- Discussion on asymmetric learning scheme. \\( L_{VIS} \\) alters knowledge within the ReID model by treating the variations in images asymmetrically. The two asymmetric parts in the \\( L_{VIS} \\) are collaborative, not isolated. Intuitively, one may think that guiding the ReID network against the variations \\( (L_{VIS}^{o}) \\) is merely for keeping ReID knowledge for accessible persons. However, \\( L_{VIS}^{o} \\) is also helpful for unlearning because it maintains a view-invariant feature space, which in turn gives the direction for shifting the identities of unlearned persons \\( (L_{VIS}^{p}) \\): out of the view-invariant space. As a result, more unlabeled images of accessible persons can improve the unlearning effect, which is not shown in other methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.312, + 0.804 + ], + "angle": 0, + "content": "3.2. Relation Regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.902 + ], + "angle": 0, + "content": "The relations between pairwise images are critical in the context of person ReID, directly reflecting the discriminative capability of the features. As formulated in Eq. 1, there should be some accessible persons spread around an unlearned person. Hence, we introduce relation regularization to facilitate De-ReID and ReID learning. Specifically, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.426 + ], + "angle": 0, + "content": "constraints that, in the feature space, the distance between the images of the same unlearned person should be greater than the distance to the accessible persons. This TRiplet Constraint facilitates retrieving images of different persons when querying an unlearned person's image." + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.431, + 0.804, + 0.448 + ], + "angle": 0, + "content": "\\[\nL _ {T R C} = \\left[ d _ {p} - d _ {n} + \\sigma_ {r} \\right] _ {+},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.444, + 0.906, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} d _ {p} = \\left\\| x _ {i} ^ {t} - x _ {i, K} ^ {o} \\right\\| _ {2} ^ {2}, d _ {n} = \\left\\| x _ {i} ^ {t} - x _ {h} ^ {t} \\right\\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.64 + ], + "angle": 0, + "content": "Here, \\( x_{i,K}^{o}(x_{i,K}^{o}\\in S_{O}) \\) represents the \\( K \\)-th nearest image to \\( x_{i}^{t} \\), and \\( \\sigma_r \\) is the margin. \\( x_{h}^{t} \\) is an image containing the same person as \\( x_{i}^{t}\\left(y_{h}^{t} = y_{i}^{t}\\right) \\). \\( L_{TRC} \\) constrains that the distance between pairwise images of the same unlearned person should be at least larger than \\( d_p + \\sigma_r \\), where \\( d_p \\) is dynamically determined based on the distance to the \\( K \\)-th nearest accessible person. When \\( K \\) is larger, \\( L_{TRC} \\) encourages more accessible persons to spread around \\( x_{i}^{t} \\). Through optimizing \\( L_{TRC} \\), the top retrieval results mainly contain different persons when retrieving unlearned person images from feature space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.909, + 0.747 + ], + "angle": 0, + "content": "Complementary to variation-guided identity shift that changes the cross-view invariance in features, we employ Relation Consistent Regularization on dataset \\( S_O \\) to maintain the feature discriminability during De-ReID fine-tuning. The regularization is to ensure that the relations between accessible individuals remain as discriminative as those in the original model \\( f_p(\\cdot) \\) during the unlearning process:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.751, + 0.907, + 0.782 + ], + "angle": 0, + "content": "\\[\nL _ {R C R} = \\left\\| \\operatorname {s i m} \\left(f _ {p} \\left(x _ {j} ^ {o}\\right), f _ {p} \\left(x _ {h} ^ {o}\\right)\\right) - \\operatorname {s i m} \\left(f \\left(x _ {j} ^ {o}\\right), f \\left(x _ {h} ^ {o}\\right)\\right) \\right\\| _ {2} ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.783, + 0.908, + 0.828 + ], + "angle": 0, + "content": "By cooperating with \\( L_{VIS} \\), the model can extract discriminative and view-invariant features for accessible persons when unlearning the specific persons." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.828, + 0.906, + 0.859 + ], + "angle": 0, + "content": "- Overall Learning objective. The model \\( f(.) \\) is trained end-to-end with an overall learning objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.865, + 0.907, + 0.881 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {1} * L _ {V I S} + \\lambda_ {2} * L _ {T R C} + L _ {R C R}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.794, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are trade-off parameters." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29334" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.224, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.124, + 0.383, + 0.14 + ], + "angle": 0, + "content": "4.1. Datasets and Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.155, + 0.486, + 0.368 + ], + "angle": 0, + "content": "- Datasets. We conduct experiments in Market-1501 [61], MSMT17 [47], Occ-Duke [34] and SYSU-MM01 [49] datasets. Due to the limited space, we leave the results in Occ-Duke and SYSU-MM01 datasets in our Appendix. Market-1501 (abbreviated as Market) consists of 1,501 persons, and all the images are captured from 6 camera views. In the conventional ReID protocol, there are 750 persons for training and 751 persons for testing. MSMT17 is a large-scale ReID dataset, containing images of 4101 persons from 15 disjoint camera views. Conventionally, 1,041 persons are for training and 3,060 persons for testing. For evaluating De-ReID, we slightly modify the conventional protocol by referring to the experimental setups of other machine unlearning works [8, 24, 53, 58]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.485, + 0.572 + ], + "angle": 0, + "content": "Specifically, in each dataset, we take \\( M_T \\) persons from the testing set as the unlearned persons. Considering the scale of the datasets, we set \\( M_T \\in \\{25, 50, 75, 100\\} \\) in MSMT17 and \\( M_T \\in \\{25, 50\\} \\) in Market1501. For each unlearned person, we split his images into two image sets from disjoint cameras. One set is added to the training set and the other is kept in the testing set. After adding the unlearned persons to the training set, we divide the training set into two subsets. One pre-training subset consists of accessible persons with identity annotations and is used for pretraining the ReID model. There are 900 accessible persons in the pre-training subset of MSMT17 and 600 accessible persons in the pre-training subset of Market-1501." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.486, + 0.734 + ], + "angle": 0, + "content": "The fine-tuning subset consists of unlabeled images of accessible persons and labeled images of unlearned persons, simulating the practical scenario where we need to efficiently fine-tune the ReID model without collecting additional labeled images of accessible persons. The dataset statistics are shown in Table 1. We emphasize that the images of the unlearned persons in the testing set under the cameras are different from those of the training set. The identities of accessible persons in the testing set are different from those in the training set. More details are in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.741, + 0.483, + 0.817 + ], + "angle": 0, + "content": "- Experimental setup. In each dataset, the pre-training subset is used to learn a ReID model in a supervised manner as \\( f_{p}(\\cdot) \\). After obtaining the well-trained ReID model, the fine-tuning subset is for learning De-ReID model \\( f(\\cdot) \\) in our main experiments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Evaluation metrics. We adopt the widely used Cumulative Matching Characteristic (CMC) and mean Average Precision (mAP) to evaluate our performance. Following previous works [38, 58], we also adopt H-Means to comprehensively evaluate the model's ability of ReID in accessible persons" + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.741, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.174 + ], + "angle": 0, + "content": "Table 1. Statistics in the case of \\( M_T = 100 \\) in MSMT17 and \\( M_T = 50 \\) in Market1501. \\( M_T \\) is the number of unlearned persons for De-ReID. Notably, the accessible persons in training are different from those in testing. \"Train\" refers to the fine-tuning subset. For the unlearned persons, the images in training and those in testing are from different cameras. More details are in Appendix." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.175, + 0.902, + 0.251 + ], + "angle": 0, + "content": "
MSMT17Market-1501
MT|ST|MO|SO|MT|ST|MO|SO|
Train10042951415061509311512284
Query100533296010706501507003068
Gallery10041362960741655091870018033
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.255, + 0.816, + 0.27 + ], + "angle": 0, + "content": "and De-ReID in unlearned persons. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.28, + 0.907, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\mathrm {H} - \\text {M e a n} = \\frac {2 * \\mathrm {R} - 1 _ {O} (f) * \\Delta \\mathrm {R} - 1 _ {T}}{\\mathrm {R} - 1 _ {O} (f) + \\Delta \\mathrm {R} - 1 _ {T}}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.315, + 0.836, + 0.332 + ], + "angle": 0, + "content": "\\[\nw h e r e \\Delta R - 1 _ {T} = R - 1 _ {T} (f _ {p}) - R - 1 _ {T} (f).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.907, + 0.419 + ], + "angle": 0, + "content": "where \\( \\mathrm{R - 1}_O(f) \\) is model \\( f \\)'s Rank-1 accuracy in accessible persons, and \\( \\Delta \\mathrm{R - 1}_T \\) is the changes in the Rank-1 accuracy in unlearned persons after De-ReID fine-tuning using different methods. In the following, we only discuss the performance of \\( f(\\cdot) \\) and therefore \\( \\mathrm{R - 1}_O \\) means \\( \\mathrm{R - 1}_O(f) \\) by default." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.426, + 0.731, + 0.441 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.645 + ], + "angle": 0, + "content": "By default, we use ViT-B [13] as the ReID backbone \\( f_{p}(\\cdot), f(\\cdot) \\). The images are resized to \\( 256 \\times 128 \\), and the data augmentation includes random crop and flip. For introducing variations in variation-guided identity shift, we use stronger augmentation derived from RandAugment [12]. We detailed the augmentation and conducted corresponding ablation studies in the Appendix. The AdamW [32] optimizer is adopted for De-ReID learning with an initial learning of 3e-4. The weight decay is set to 0. We empirically set trade-off parameters \\( \\lambda_{1} = 1.0 \\) and \\( \\lambda_{2} = 1.0 \\). The batch size for accessible persons is 48, and that for unlearned persons is 32. The above hyper-parameters are the same in all experiments. \\( K \\) in Eq. 5 is set to 20." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.818, + 0.669 + ], + "angle": 0, + "content": "4.3. Comparison with Related Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.87 + ], + "angle": 0, + "content": "- Compared methods. Since this is the first work for De-ReID, we compare our method with existing methods potentially applicable to ReID. We pretrain a vanilla ViT-B in the pretraining subset of each dataset and then apply these methods to the ViT-B in the fine-tuning subset for De-ReID. We report the performance of the machine unlearning methods, including GS-LoRA [58], LIRF [53], SCRUB [24], and Boundary Shrink [8]. Besides, we implement a naive solution: label augmentation based on cameras, which assign different labels for images containing the same person but under different cameras. This naive solution is denoted as \"LabelAug\", and the triplet loss and cross-entropy loss [33] are applied for training LabelAug." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.9 + ], + "angle": 0, + "content": "Notably, previous methods usually assume that the labeled data of accessible classes are available. Since the data" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29335" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.487, + 0.231 + ], + "angle": 0, + "content": "Table 2. Comparisons in the MSMT17 dataset. \"BS\" is the Boundary Shrink [8]. \"H\" denotes \"H-Mean\". R-1\\(_T\\) is the Rank-1 accuracy on the unlearned persons, and R-1\\(_O\\) is the Rank-1 accuracy on the accessible persons. R-1\\(_T\\) is expected to be low for unlearning specific persons, while a higher R-1\\(_O\\) and H is preferred. Since we apply additional augmentation for asymmetric contrastive learning, we further apply our augmentation to LIRF, denoted as LIRF*. \"RR\" is the Re-Ranking [63]. The method with the best/second-best H-Mean is marked in red/ blue. Blue background means the H-Mean are improved by Re-Ranking." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.233, + 0.48, + 0.422 + ], + "angle": 0, + "content": "
MethodMT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug54.673.838.156.873.532.4
+RR62.376.929.265.377.221.2
BS63.878.727.354.674.935.2
+RR70.082.218.362.578.725.3
SCRUB43.875.049.145.274.445.1
+RR52.378.441.353.778.236.6
LIRF10.062.566.218.554.656.8
+RR6.965.469.220.557.357.2
LIRF*18.568.865.135.169.752.8
GS-LoRA56.977.535.960.280.728.6
+RR56.979.736.264.182.923.2
Ours4.677.076.310.872.969.7
+RR4.679.077.310.475.170.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.439, + 0.483, + 0.466 + ], + "angle": 0, + "content": "Table 3. Comparison results in the MSMT17 dataset when \\( M_T = 75 \\) and \\( M_T = 100 \\). The notations are consistent with Table 2" + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.468, + 0.48, + 0.659 + ], + "angle": 0, + "content": "
MethodMT=75MT=100
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug61.472.229.262.771.927.2
+ RR69.175.818.666.875.721.8
BS57.269.133.954.868.336.3
+ RR65.674.323.761.472.728.9
SCRUB54.774.837.550.567.840.6
+ RR62.979.027.755.772.835.9
LIRF18.651.956.120.849.753.8
+ RR19.655.157.524.852.453.5
LIRF*38.465.050.515.650.956.7
GS-LoRA62.479.528.462.176.128.3
+ RR65.381.824.563.878.626.2
Ours12.469.868.513.167.166.7
+ RR13.171.669.015.368.866.4
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.688, + 0.483, + 0.748 + ], + "angle": 0, + "content": "of accessible persons is unlabeled in our experiments, we apply our \\( L_{RCR} \\) as a regularization for previous methods. We further apply our methods to advanced ReID models, including DCFormaler [27], PAT [28], and PHA [56]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "- Results. Table 2 and Table 4 shows the comparison results. From the tables, we can observe that our method achieves the best performance in terms of H-Mean. Since the post-processing methods like Re-Ranking [63] can improve the results, we investigate whether these methods can reduce the unlearning effect when retrieving unlearned persons. The results show that existing unlearning methods usually get worse results after post-processing. For example, in Table 2, the Rank-1 accuracy of the SCRUB on unlearned persons is clearly increased from \\(43.8\\%\\) to \\(52.3\\%\\) when \\(M_T = 25\\)," + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.09, + 0.877, + 0.103 + ], + "angle": 0, + "content": "Table 4. Comparison results in the Market-1501 dataset." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.105, + 0.905, + 0.297 + ], + "angle": 0, + "content": "
MethodMT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug50.776.551.364.777.442.5
+RR49.377.652.866.078.941.3
BS57.373.144.562.072.244.3
+RR54.775.347.465.374.041.4
SCRUB50.776.051.254.770.250.4
+RR56.077.446.664.771.841.6
LIRF25.384.872.931.375.968.7
+RR26.784.471.932.775.767.7
LIRF*37.386.064.539.379.064.6
GS-LoRA56.091.748.965.391.443.7
+RR54.791.350.267.391.341.3
Ours10.791.184.412.084.483.2
+RR5.391.087.410.083.583.8
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.905, + 0.354 + ], + "angle": 0, + "content": "demonstrating that the model does not effectively unlearn the persons and there are potential cues to re-identify them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.43 + ], + "angle": 0, + "content": "While other methods achieve a worse H-mean after Re-ranking, our method can benefit from Re-Ranking [63] since the overall performance H-Mean is improved in most cases. For example, in the case of \\( M_T = 25 \\) in Market dataset, the H-Means of our method is increased to \\( 87.4\\% \\) from \\( 84.4\\% \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.909, + 0.552 + ], + "angle": 0, + "content": "Compared with the other methods, Boundary Shrink changes the classification decision boundary of the unlearning classes, i.e., the unlearned persons in this work. However, the feature representation is the core for person image matching instead of the classifier. In contrast, our method directly optimizes the feature representation for De-ReID. As a result, our method outperforms Boundary Shrink by \\(30.4\\%\\) H-Mean in MSMT17 with \\(M_T = 100\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.78 + ], + "angle": 0, + "content": "LIRF unlearns classes by constraining the attention map of intermediate features to be significantly different before and after unlearning. Since it directly constrains the features, it obtains the second-best results. However, the attention map has lots of knowledge about extracting identity information from images. Hence, their constraints largely hurt the ReID ability on accessible persons, and our method outperforms LIRF by \\(12.9\\%\\) at H-Mean in MSMT17 when \\(M_T = 100\\). Besides, since we apply additional augmentation in variation-guided identity shift, we further validate the effectiveness of the additional augmentation by applying the augmentation in LIRF, denoted as \\(\\mathrm{LIRF^{*}}\\). However, the H-Mean is improved only in the case of \\(M_T = 100\\) in MSMT17 and decreases in other cases, demonstrating that additional augmentation does not necessarily lead to higher performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "GS-LoRA effectively protects the ReID ability on accessible persons but falls behind in unlearning specific persons. LabelAug directly guides the model in distinguishing the images of unlearned persons under different camera views, but it has an unsatisfactory performance. We speculate it is because the model overfits the images in training and fails to unlearn the specific persons under cameras unseen in training. Differently, our method employs variation-guided" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29336" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.16 + ], + "angle": 0, + "content": "Table 5. Evaluation with different ReID models. Regardless of the ReID models, our method performs the best in terms of erasing the knowledge about unlearned persons (quantified by \\( \\mathrm{R - 1}_T \\)) and keeping ReID knowledge for accessible persons (\\( \\mathrm{R - 1}_O \\)). \"Null\" indicates that no De-ReID method is applied." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.161, + 0.462, + 0.372 + ], + "angle": 0, + "content": "
ReID modelDe-ReID methodR-1T↓R-1O↑H↑
DCFormerNull81.284.4—
LabelAug44.154.046.2
GS-LORA75.381.916.4
LIRF18.051.057.7
Ours17.866.166.3
PATNull72.875.6—
LabelAug38.343.340.1
GS-LORA52.560.933.5
LIRF16.742.649.4
Ours12.055.059.0
PHANull70.278.7—
LabelAug44.351.441.2
GS-LORA56.468.333.6
LIRF23.650.552.7
Ours3.951.260.8
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.384, + 0.485, + 0.505 + ], + "angle": 0, + "content": "identity shift \\((L_{VIS}^{p}, L_{VIS}^{o})\\) and guides the model to handle the variations adaptively. Fitting the variations in unlearned images helps the model bias to the environmental information when querying the unlearned persons. Simultaneously, the model is guided to be robust to the variations when querying accessible persons. As a result, the unlearned persons are shifted out of the view-invariant feature space while the accessible persons are retained." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.506, + 0.485, + 0.597 + ], + "angle": 0, + "content": "- Evaluations on other ReID methods. Apart from the comparison on the vanilla ViT backbone, we also compared our method with related methods on the advanced ReID models [27, 28, 56]. The experimental results in Table 5 demonstrate that our method is superior to others regarding unlearning different ReID models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.245, + 0.623 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.486, + 0.734 + ], + "angle": 0, + "content": "We conduct experiments to show the effectiveness of each component in our design. Due to the space limit, we report the experimental results in Market and \\( M_T \\in \\{25, 100\\} \\) in MSMT17 in Table 6, and the rest in Appendix. Regardless of \\( M_T \\), our designs clearly improve H-Mean that comprehensively evaluates the effectiveness of unlearning specific persons and ReID performance on accessible persons." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.486, + 0.886 + ], + "angle": 0, + "content": "- The effectiveness of the relation regularization. From Table 6, we first observe that the model suffers from catastrophic forgetting in ReID knowledge and collapses in the absence of \\( L_{RCR} \\). Although our variation-guided identity shift pulls closer the images and the augmented images for accessible persons, it mainly preserves the view-invariant feature embedding but it fails to preserve the feature discriminability, leading to model collapses. \\( L_{RCR} \\) preserves the discriminability of features by preserving the relations between accessible persons." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Another constraint \\(L_{TRC}\\) attempts to push away the im" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.16 + ], + "angle": 0, + "content": "Table 6. Ablation study in MSMT17 dataset. \"W/o\" means \"without\", and other notations are the same as Table 2. \\( f_{p}(\\cdot) \\) is the initial model for De-ReID learning. The full model achieves the best H-Mean. Notably, without \\( L_{RCR} \\), the model forgets ReID knowledge and collapses. \"SD\" denotes the \"self-augmented discrimination\"." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.161, + 0.905, + 0.315 + ], + "angle": 0, + "content": "
MethodMSMT17
MT=25MT=100
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
fp(·)80.385.4—79.585.4—
Components in Relation regularization
W/o LTRC5.573.474.118.365.363.2
W/o LRCR——————
Components in Variation-guided Identity Shift
W/o SD4.674.575.110.558.763.4
W/o LVIS10.075.872.918.063.262.3
W/o LVIS20.577.467.541.370.049.4
Full model4.677.076.313.167.166.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.589, + 0.329, + 0.831, + 0.342 + ], + "angle": 0, + "content": "Table 7. Ablation study in Market-1501." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.343, + 0.903, + 0.498 + ], + "angle": 0, + "content": "
MethodMarket-1501
MT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
fp(·)89.396.3—94.096.3—
Components in Relation regularization
W/o LTRC20.089.778.224.078.373.9
W/o LRCR——————
Components in Variation-guided Identity Shift
W/o SD24.090.375.811.379.781.2
W/o LVIS10.784.581.412.777.379.2
W/o LVIS46.793.958.655.389.954.1
Full model10.791.184.412.084.483.2
" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.511, + 0.703, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.51, + 0.895, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.615, + 0.868, + 0.628 + ], + "angle": 0, + "content": "Figure 3. Evaluation results under various \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.825 + ], + "angle": 0, + "content": "ages between unlearned persons to prevent them from being cross-view retrieved, referring to the distance between accessible persons and unlearned persons. We observe a degraded H-Mean in the absence of \\( L_{TRC} \\) in all cases. For example, when deprecating \\( L_{TRC} \\) in the case of \\( M_T = 50 \\) in Market, the Rank-1 accuracy of the unlearned persons is increased to \\( 24.0\\% \\) from \\( 12.0\\% \\), meaning that more unlearned person images are correctly matched and is not expected. As a result, the H-Mean drops by \\( 9.8\\% \\) compared with the full model. In the case of \\( M_T = 100 \\) in MSMT17, the Rank-1 accuracy of the unlearned persons increases to \\( 18.3\\% \\) from \\( 13.1\\% \\) without \\( L_{TRC} \\), and the H-Mean drops by \\( 3.3\\% \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The hyper-parameter \\(\\lambda_{2}\\) controls the weight of \\(L_{TRC}\\). We evaluate our model under different \\(\\lambda_{2}\\) and show the performance in Fig. 3. When \\(\\lambda_{2} = 0\\), \\(L_{TRC}\\) is not employed and the performance is inferior. As \\(\\lambda_{2}\\) increases, the performance is improved. However, \\(L_{TRC}\\) should not be set to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29337" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.478, + 0.108 + ], + "angle": 0, + "content": "too large to avoid overwhelming other learning objectives." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.109, + 0.485, + 0.38 + ], + "angle": 0, + "content": "- The effectiveness of the variation-guided identity shift. The variation-guided identity shift (VIS) emphasizes the style difference between images of unlearned persons in the feature space while keeping the other images robust to the style variations. Overall, comparing the results of \"W/o \\( L_{VIS} \\)\" and the full model in Table 6, we observe that the ReID model struggles to unlock the specific persons without \\( L_{VIS} \\). For example, in MSMT17 dataset, the Rank-1 accuracy on unlearned persons is \\( 20.5\\% / 41.3\\% \\) when \\( M_T = 25 / 100 \\), which is \\( 15.9\\% / 28.2\\% \\) larger than the full model. This is because ReID requires view-invariant features and our model correspondingly guides the model to be biased to encode environmental information for images of unlearned persons. As a result, we effectively prevent the images of unlearned persons from being re-identified. The hyperparameter \\( \\lambda_1 \\) controls the weight of \\( L_{VIS} \\) in optimization. We evaluate our model under different \\( \\lambda_1 \\) and show the performance in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.484, + 0.594 + ], + "angle": 0, + "content": "We also studied the effectiveness of the designs in VIS. As discussed in the introduction, we argue that simply pushing away the images under different camera views may learn the spurious cues for De-ReID, and therefore we push away the image and the corresponding augmented image in our VIS. We refer to this as \"self-augmented discrimination\". To verify its effectiveness, we conduct an experiment that pushes away the image and the augmented image from another image, denoted as \"W/o SD\", and observe a significant performance drop. For example, in the case of \\( M_T = 25 \\) in Market dataset, the Rank-1 accuracy on the unlearned person increases to \\( 24.0\\% \\) from \\( 10.7\\% \\) without SD, and the H-Mean drops by \\( 8.6\\% \\). In other cases, the H-Mean drops without using the SD, demonstrating the effectiveness of SD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.687 + ], + "angle": 0, + "content": "Moreover, we observe that guiding the ReID to overcome the variations for accessible persons is important for balancing the ReID ability and De-ReID of unlearned persons. Deprecating the \\( L_{VIS}^{o} \\) leads to \\( 3\\% \\sim 4\\% \\) performance drops in H-Mean in different cases, and reduce the protection effect since the rank-1 accuracy of unlearned persons increases." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.688, + 0.483, + 0.763 + ], + "angle": 0, + "content": "- The impact of the unlabeled sample. We evaluate our method in MSMT17 under different ratios between unlabeled images of accessible persons and labeled images of unlearned persons in the fine-tuning subset. The results are in Table 8." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "From the results, we obtain two conclusions. (i) More unlabeled data is helpful for both unlearning specific persons and keeping ReID performance on the accessible persons. When the number of accessible persons decreases, the Rank-1 accuracy of the unlearned persons (denoted as \\( \\mathrm{R - 1}_T \\)) increases, which means the unlearned persons are easier to retrieve by the ReID model. Simultaneously, the Rank-1 accuracy of the accessible persons \\( (\\mathrm{R - 1}_O) \\) decreases, which means the ReID model cannot properly re-identify" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.173 + ], + "angle": 0, + "content": "Table 8. Evaluations under different numbers of accessible persons in MSMT17 when \\( M_T = 100 \\). \\( M_O \\) is the number of accessible persons in the fine-tuning subset. 'With LUP' means using the images in LUPerson dataset as auxiliary unlabeled data and the number in parentheses denotes the performance gain. By default, we do not use LUPerson dataset." + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.175, + 0.903, + 0.25 + ], + "angle": 0, + "content": "
\\(M_O\\)Without LUPWith LUP
R-1T ↓R-1O ↑H ↑R-1T ↓R-1O ↑H ↑
14113.167.166.7---
10015.861.362.514.168.266.8 (+4.3%)
8015.854.358.615.467.765.9 (+7.3%)
6017.451.457.615.465.664.8 (+7.2%)
" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.266, + 0.571, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.348, + 0.703, + 0.361 + ], + "angle": 0, + "content": "Images of unlearned persons" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.266, + 0.635, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.348, + 0.635, + 0.359 + ], + "angle": 0, + "content": "of unlearn" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.266, + 0.7, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.348, + 0.7, + 0.359 + ], + "angle": 0, + "content": "d persons" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.266, + 0.774, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.348, + 0.885, + 0.361 + ], + "angle": 0, + "content": "Images of other persons" + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.266, + 0.866, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.853, + 0.348, + 0.885, + 0.359 + ], + "angle": 0, + "content": "rsons" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.364, + 0.907, + 0.392 + ], + "angle": 0, + "content": "Figure 4. Attention map on images of the unlearned persons and the images of accessible persons." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.907, + 0.522 + ], + "angle": 0, + "content": "the accessible persons. (ii) Although the performance of our method decreases when the unlabeled images become fewer, our method can effectively utilize the publicly available unlabeled person images for unlearning, demonstrating the scalability of our method. When using additional unlabeled images from LUPerson dataset, our method can achieve similar performance even though the number of accessible persons in MSMT17 decreases from 141 to 60." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.534, + 0.651, + 0.549 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.907, + 0.678 + ], + "angle": 0, + "content": "We visualize the attention map on different persons in the testing set in Fig. 4. From the figure, for the images of unlearned persons, we can observe that the model's attention mainly focus on the background which contains abundant environmental information. At the same time, the model pays diverse attention to the informative parts of images for extracting discriminative features for accessible persons. These results demonstrate the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To address the privacy concerns in person ReID, we pioneer the exploration of the De-ReID problem that guides the ReID model to forget the knowledge about matching specific persons. We formulate a framework and propose a novel variation-guided identity shift method for De-ReID. Our method unlearns the specific person shifting them out of the view-invariant feature space and simultaneously keeps the ReID knowledge by overcoming the cross-view variations for accessible persons. We further introduce relation regularization to characterize the desired properties of De-ReID. Extensive experiments conducted on a new benchmark show the superiority of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29338" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.177 + ], + "angle": 0, + "content": "This work was supported partially by NSFC(U21A20471, 92470202), by the National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025)" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[1] Prachi Agrawal and P. J. Narayanan. Person de-identification in videos. IEEE Transactions on Circuits and Systems for Video Technology, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.484, + 0.311 + ], + "angle": 0, + "content": "[2] Shafiq Ahmad, Pietro Morerio, and Alessio Del Bue. Person re-identification without identification via event anonymization. In 2023 IEEE/CVF International Conference on Computer Vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.484, + 0.368 + ], + "angle": 0, + "content": "[3] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In IEEE Symposium on Security and Privacy, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.484, + 0.436 + ], + "angle": 0, + "content": "[4] Karla Brkic, Ivan Sikiric, Tomislav Hrkac, and Zoran Kalafatic. I know that person: Generative full body and face de-identification of people in images. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.484, + 0.479 + ], + "angle": 0, + "content": "[5] Jonathan Brophy and Daniel Lowd. Machine unlearning for random forests. In International Conference on Machine Learning, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.549 + ], + "angle": 0, + "content": "[6] Chengzhi Cao, Xueyang Fu, Hongjian Liu, Yukun Huang, Kunyu Wang, Jiebo Luo, and Zheng-Jun Zha. Event-guided person re-identification via sparse-dense complementary learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.619 + ], + "angle": 0, + "content": "[7] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Learning invariance from generated variance for unsupervised person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[8] Min Chen, Weizhuo Gao, Gaoyang Liu, Kai Peng, and Chen Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[9] Tianshui Chen, Tao Pu, Lingbo Liu, Yukai Shi, Zhijing Yang, and Liang Lin. Heterogeneous semantic transfer for multi-label recognition with partial labels. International Journal of Computer Vision, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[10] Weihua Chen, Xiaotang Chen, Jianguo Zhang, and Kaiqi Huang. Beyond triplet loss: A deep quadruplet network for person re-identification. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[11] Dasol Choi and Dongbin Na. Towards machine unlearning benchmarks: Forgetting the personal identities in facial recognition systems. In AAAI 2024 workshop on Privacy-Preserving Artificial Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Ekin D. Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le. Randaugment: Practical automated data augmentation" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.908, + 0.133 + ], + "angle": 0, + "content": "with a reduced search space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.233 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[14] Lijie Fan, Tianhong Li, Rongyao Fang, Rumen Hristov, Yuan Yuan, and Dina Katabi. Learning longterm representations for person re-identification using radio signals. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.908, + 0.375 + ], + "angle": 0, + "content": "[15] Ivan Filković, Zoran Kalafatić, and Tomislav Hrkač. Deep metric learning for person re-identification and de-identification. In 2016 39th International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[16] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.907, + 0.489 + ], + "angle": 0, + "content": "[17] Wenxuan Guo, Zhiyu Pan, Yingping Liang, Ziheng Xi, Zhi Chen Zhong, Jianjiang Feng, and Jie Zhou. Lidar-based person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.907, + 0.545 + ], + "angle": 0, + "content": "[18] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.907, + 0.603 + ], + "angle": 0, + "content": "[19] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In IEEE/CVF International Conference on Computer Vision, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.908, + 0.688 + ], + "angle": 0, + "content": "[20] Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qing-song Xie, Yizhou Wnag, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, Donglian Qi, and Yunfeng Yan. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.908, + 0.759 + ], + "angle": 0, + "content": "[21] Bingliang Jiao, Lingqiao Liu, Liying Gao, Guosheng Lin, Lu Yang, Shizhou Zhang, Peng Wang, and Yanning Zhang. Dynamically transformed instance normalization network for generalizable person re-identification. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[22] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Andy J Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multiscale dilated transformer for visual recognition. IEEE Transactions on Multimedia, 25:8906-8919, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[23] Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Cloth-changing person re-identification from a single image with gait prediction and regularization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29339" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[24] Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Triantafillou. Towards unbounded machine unlearning. In Conference on Neural Information Processing Systems, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.485, + 0.205 + ], + "angle": 0, + "content": "[25] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: Exploiting vision-language model for image re-identification without concrete text labels. Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[26] Weijia Li, Saihui Hou, Chunjie Zhang, Chunshui Cao, Xu Liu, Yongzhen Huang, and Yao Zhao. An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[27] Wen Li, Cheng Zou, Meng Wang, Furong Xu, Jianan Zhao, Ruobing Zheng, Yuan Cheng, and Wei Chu. Dc-former: Diverse and compact transformer for person re-identification. In AAAI Conference on Artificial Intelligence, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[28] Yulin Li, Jianfeng He, Tianzhu Zhang, Xiang Liu, Yongdong Zhang, and Feng Wu. Diverse part discovery: Occluded person re-identification with part-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[29] Kun-Yu Lin, Jiaming Zhou, Yukun Qiu, and Wei-Shi Zheng. Adversarial partial domain adaptation by cycle inconsistency. In European Conference on Computer Vision, pages 530-548. Springer Nature Switzerland Cham, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[30] Kun-Yu Lin, Jia-Run Du, Yipeng Gao, Jiaming Zhou, and Wei-Shi Zheng. Diversifying spatial-temporal perception for video domain generalization. Advances in Neural Information Processing Systems, 36:56012–56026, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[31] Kun-Yu Lin, Henghui Ding, Jiaming Zhou, Yu-Ming Tang, Yi-Xing Peng, Zhilin Zhao, Chen Change Loy, and Wei-Shi Zheng. Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.7 + ], + "angle": 0, + "content": "[33] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[34] Jiaxu Miao, Yu Wu, Ping Liu, Yuhang Ding, and Yi Yang. Pose-guided feature alignment for occluded person re-identification. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 542–551, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[35] Tao Pu, Tianshui Chen, Hefeng Wu, Yukai Shi, Zhijing Yang, and Liang Lin. Dual-perspective semantic-aware representation blending for multi-label image recognition with partial labels. Expert Systems with Applications, page 123526, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[36] Tao Pu, Qianru Lao, Hefeng Wu, Tianshui Chen, Ling Tian, Jie Liu, and Liang Lin. Category-adaptive label discovery and noise rejection for multi-label recognition with partial positive labels. IEEE Transactions on Multimedia, pages 1-12, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[37] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.907, + 0.191 + ], + "angle": 0, + "content": "[38] Takashi Shibata, Go Irie, Daiki Ikami, and Yu Mitsuzumi. Learning with selective forgetting. In International Joint Conference on Artificial Intelligence, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[39] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and A strong convolutional baseline). In European Conference on Computer Vision, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.907, + 0.318 + ], + "angle": 0, + "content": "[40] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.907, + 0.388 + ], + "angle": 0, + "content": "[41] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. Learning to imagine: Diversify memory for incremental learning using unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9549-9558, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.907, + 0.446 + ], + "angle": 0, + "content": "[42] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. When prompt-based incremental learning does not meet strong pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1706-1716, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[43] Yu-Ming Tang, Yi-Xing Peng, Jingke Meng, and Wei-Shi Zheng. Rethinking few-shot class-incremental learning: Learning from yourself. In European Conference on Computer Vision, pages 108-128. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.907, + 0.559 + ], + "angle": 0, + "content": "[44] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACM International Conference on Multimedia, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.907, + 0.616 + ], + "angle": 0, + "content": "[45] Haochen Wang, Jiayi Shen, Yongtuo Liu, Yan Gao, and Efstratos Gavves. Nformer: Robust person re-identification with neighbor transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[46] Lin Wang, Wanqian Zhang, Dayan Wu, Fei Zhu, and Bo Li. Attack is the best defense: Towards preemptive-protection person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[47] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 79–88, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.907, + 0.785 + ], + "angle": 0, + "content": "[48] Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. A discriminative feature learning approach for deep face recognition. In European Conference on Computer Vision, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[49] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[50] Bin Yang, Jun Chen, and Mang Ye. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In IEEE/CVF International Conference on Computer Vision, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29340" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[51] Bin Yang, Jun Chen, and Mang Ye. Shallow-deep collaborative learning for unsupervised visible-infrared person re-identification. In IEEE/CVF International Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.485, + 0.219 + ], + "angle": 0, + "content": "[52] Jinrui Yang, Jiawei Zhang, Fufu Yu, Xinyang Jiang, Mengdan Zhang, Xing Sun, Ying-Cong Chen, and Wei-Shi Zheng. Learning to know where to see: A visibility-aware approach for occluded person re-identification. In IEEE/CVF international conference on computer vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.483, + 0.274 + ], + "angle": 0, + "content": "[53] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In European Conference on Computer Vision, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.276, + 0.483, + 0.331 + ], + "angle": 0, + "content": "[54] Mang Ye, Wei Shen, Junwu Zhang, Yao Yang, and Bo Du. Securereid: Privacy-preserving anonymization for person re-identification. IEEE Transactions on Information Forensics and Security, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.485, + 0.386 + ], + "angle": 0, + "content": "[55] Mang Ye, Zesen Wu, Cuiqun Chen, and Bo Du. Channel augmentation for visible-infrared re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[56] Guiwei Zhang, Yongfei Zhang, Tianyu Zhang, Bo Li, and Shiliang Pu. Pha: Patch-wise high-frequency augmentation for transformer-based person re-identification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.485, + 0.514 + ], + "angle": 0, + "content": "[57] Junwu Zhang, Mang Ye, and Yao Yang. Learnable privacy-preserving anonymization for pedestrian images. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[58] Hongbo Zhao, Bolin Ni, Haochen Wang, Junsong Fan, Fei Zhu, Yuxi Wang, Yuntao Chen, Gaofeng Meng, and Zhaoxiang Zhang. Continual forgetting for pre-trained vision models. 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[59] Zhilin Zhao, Longbing Cao, and Kun-Yu Lin. Revealing the distributional vulnerability of discriminators by implicit generators. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8888-8901, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[60] Zhilin Zhao, Longbing Cao, and Kun-Yu Lin. Out-of-distribution detection by cross-class vicinity distribution of in-distribution data. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[61] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Computer Vision, IEEE International Conference on, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.743, + 0.483, + 0.811 + ], + "angle": 0, + "content": "[62] Wei-Shi Zheng, Jincheng Hong, Jiening Jiao, Ancong Wu, Xiatian Zhu, Shaogang Gong, Jiayin Qin, and Jianhuang Lai. Joint bilateral-resolution identity modeling for crossresolution person re-identification. International Journal of Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.814, + 0.483, + 0.869 + ], + "angle": 0, + "content": "[63] Zhun Zhong, Liang Zheng, Donglin Cao, and Shaozi Li. Re-ranking person re-identification with k-reciprocal encoding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29341" + } + ] +] \ No newline at end of file diff --git a/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/1180818e-c94d-44fa-b5df-b1fc8919ac06_origin.pdf b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/1180818e-c94d-44fa-b5df-b1fc8919ac06_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0f1a93ab055cbe3ea69288ad8f618b13bcda4d43 --- /dev/null +++ b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/1180818e-c94d-44fa-b5df-b1fc8919ac06_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62abdfccefa6c91e2594495d11e7d0919eb27c068ef0eb9b3fe6ce5d1a610850 +size 1095665 diff --git a/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/full.md b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5414f2067ace9bab38803f67e70d5c9f13546eab --- /dev/null +++ b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/full.md @@ -0,0 +1,356 @@ +# Person De-reidentification: A Variation-guided Identity Shift Modeling + +Yi-Xing Peng $^{1,2,4}$ Yu-Ming Tang $^{1,4}$ Kun-Yu Lin $^{1,4}$ Qize Yang $^{3}$ Jingke Meng $^{1,4*}$ Xihan Wei $^{3}$ Wei-Shi Zheng $^{1,2,4*}$ + +1School of Computer Science and Engineering, Sun Yat-sen University, China; + +$^{2}$ Peng Cheng Laboratory, Shenzhen, China; $^{3}$ Tongyi Lab, Alibaba Group; + +4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China. + +{pengyx23,tangym9,linky5}@mail2.sysu.edu.cn wszheng@ieee.org + +# Abstract + +Person re-identification (ReID) is to associate images of individuals from different camera views against cross-view variations. Like other surveillance technologies, Re-ID faces serious privacy challenges, particularly the potential for unauthorized tracking. Although various tasks (e.g., face recognition) have developed machine unlearning techniques to address privacy concerns, such methods have not yet been explored within the Re-ID field. In this work, we pioneer the exploration of the person de-reidentification (De-ReID) problem and present its inherent challenges. In the context of ReID, De-ReID is to unlock the knowledge about accurately matching specific persons so that these "unlearned persons" cannot be re-identified across cameras for privacy guarantee. The primary challenge is to achieve the unlearning without degrading the identity-discriminative feature embeddings to ensure the model's utility. To address this, we formulate a De-ReID framework that utilizes a labeled dataset of unlearned persons for unlearning and an unlabeled dataset of accessible persons for knowledge preservation. Instead of unlearning based on (pseudo) identity labels, we introduce a variation-guided identity shift mechanism that unlearns the specific persons by fitting the variations in their images while preserving ReID ability on other persons by overcoming the variations in images of accessible persons. As a result, the model shifts the unlearned persons to a feature space that is vulnerable to cross-view variations. Extensive experiments on benchmarks demonstrate the superiority of our method. + +# 1. Introduction + +Person re-identification (ReID) aims to match the images of the same person across different camera views based on feature similarity. While deep learning [9, 35, 43] has significantly enhanced the performance of the ReID mod + +![](images/a3a019974be3f120f13baeb520587213302d3b239815ca7788817d7c063accb2.jpg) +Figure 1. Person de-reidentification (De-ReID) prevents intelligent surveillance from tracking unlearned persons. For example, when using surveillance models for security (track and record the person in the red box), the residents in the community and the important customers in business (in the green box) should not be tracked for privacy and commercial secrets. + +els [10, 19, 45], there are growing public concerns about potential misuse, leading to ethical and societal implications. To avoid misuse and mitigate the potential privacy concerns, the research in surveillance has begun exploring machine unlearning techniques that limit the operational scope of these models. For instance, in face recognition, models are adjusted to "unlearn" individuals who raise privacy concerns, as mandated by regulations like the EU's General Data Protection Regulation (GDPR) [3, 5, 24, 53]. + +However, similar constraints on the ReID models have not yet to be explored, despite its significance for real-world applications. For instance, the residents in the neighborhood prefer to constrain the surveillance in the community to only focus on external people and not track themselves; in working buildings, security cameras should only be set for monitoring and preventing theft or accidents, not for tracking or identifying specific employees during their routine work; and homeowners expect smart home devices to track only unknown individuals for security purposes while avoiding any continuous surveillance of themselves; To meet these requirements, there is an urgent need to develop methods that limit the scope of ReID models, enabling them to "ignore" specific individuals during deployment. + +In this work, we pioneer the exploration of person deidentification (De-ReID) problem: limiting the ReID scope by unlearning specific persons (referred to as "unlearned persons") in the ReID model while maintaining the ReID + +performance on other persons (referred to as "accessible persons"). In the context of image matching, a person is unlearned by a model means the model can hardly associate the images of the person under different camera views, as shown in Fig. 1. Forming a De-ReID model is challenging because only the knowledge strictly related to matching the unlearned persons should be eliminated, without destroying the identity-discriminative feature embeddings. A trivial solution is to distinguish the images of unlearned persons under different camera views for unlearning while associating the images of accessible persons based on their identity labels to preserve ReID knowledge. However, due to the large network capacity, the model may learn spurious cues or overfit the images of unlearned persons in the training set, leading to poor generalization during evaluation. Additionally, collecting identity labels for accessible persons to preserve ReID knowledge is both costly and difficult to scale, especially when the number of cameras is large [29, 31, 36]. + +To handle the De-ReID, we formulate a weakly-supervised unlearning framework where a labeled image set of unlearned persons and an unlabeled dataset of accessible persons are available to adapt a well-trained ReID model to De-ReID. Given that the core of ReID is to overcome cross-view variations, such as different camera viewing angles and lighting conditions, we propose variation-guided identity shift that learns De-ReID based on the variations in images. The core of our method is to guide the ReID model bias towards environmental information (e.g., lighting) when encountering unlearned persons, which shifts the unlearned persons out of the view-invariant feature space. At the same time, we keep the model remains robust to cross-view variations when querying accessible persons. + +Specifically, for unlearned persons, we guide the model to distinguish their images and the images augmented by variations. Compared with distinguishing images under different camera views, learning to distinguish image and its augmentation results is more effective for unlearning since the augmented images are highly similar to the original image (e.g., body shape) but varies in style information (e.g., brightness and color tone). Therefore, the model is biased to encode the style information to push away these images. In contrast, for accessible persons, we pull closer their images and augmented versions in the feature space, expecting the model to overcome the variations when matching them. Moreover, we introduce relation regularization to characterize the desired properties of De-ReID in terms of relations. + +In summary, our main contributions are as follows: + +- We are the first to investigate the challenging De-ReID task and conduct a new benchmark. +- We formulate a De-ReID framework and propose variation-guided identity shift that achieves De-ReID by fitting the variations in images of interest while against the variations in others to preserve ReID knowledge. + +- To further utilize the identity annotations in unlearned person images for learning, we propose relation regularization that characterizes the desired properties of ReID and De-ReID. Both qualitative and quantitative results demonstrate the effectiveness of our method. + +# 2. Related Works + +- Person re-identification. Based on deep learning, advanced ReID models automatically learn a discriminative feature embedding from a large amount of data [20]. Various network architectures [19, 21, 22, 30, 39, 42, 44, 45] and loss functions [10, 40, 41, 48] are developed to enhance feature representation learning. Besides, the challenges in ReID including cross-modality [25, 31, 51, 55], cloth-changing [23, 26], low-resolution [62], occlusion [52], and unsupervised ReID [7, 50] are also widely studied. + +- Privacy protection for person images. Apart from pursuing high performance, protecting human privacy in human-centric tasks is also important [54, 59, 60]. A stream of research focuses on generating anonymized images by removing or obscuring identity-related information, such as faces and silhouettes, while preserving the image's utility for subsequent tasks, such as action recognition [1, 4, 15]. + +In the field of ReID, Ye et al. [54, 57] proposed a generative model to anonymize the person images to avoid privacy leakage while keeping the ReID utility. Wang et al. [46] add imperceptible perturbations to the images of in the database to prevent malicious use of the images. Particularly, one stream of works explores person de-identification [1, 4, 15], which obscures the identity information in images or videos, including face and silhouette. The new sensors including event cameras and LiDAR are also investigated for privacy-preserving ReID [2, 6, 14, 17]. + +Despite their progress, previous works mainly focus on preventing the malicious use of images while we aim to prevent the malicious use of models. Technically, previous works [54, 57] learn to anonymize person images and jointly adapt the ReID model to ensure ReID utility. Their methods cannot prevent the ReID model from retrieving specific persons, as well as the methods that explore new sensors such as event cameras for ReID [2, 6, 14, 17]. Besides, although PRIDE [46] protects the images in the database from being maliciously used, the attacker can also collect images that are not in the database to track specific persons. Moreover, some works in person de-identification only concern the identity-unrelated utility [1, 4, 15], such as action recognition, while De-ReID requires identifying accessible persons. + +- Machine unlearning. De-ReID has a similar goal to machine unlearning. Machine unlearning is first proposed to eliminate the influence of specific data on the deep model as if the model never uses these data for training, which enables users to erase their personal data [3, 5, 16, 18, 37]. + +Kurmanji et al. [24] further apply machine unlearning for + +various purposes such as removing the bias from the model or eliminating the negative effects of mislabeled samples. To maximize the error in forgetting classes, they maximize the difference between the outputs of the learned model and the original model in the forgetting samples. Choi et al. [11] propose to jointly maximize the loss on the forgetting data and learn on the other data. Chen et al. [8] unlearn specific classes by adjusting their decision boundary. The forgetting samples are assigned a nearest but incorrect label for boundary shrink. Zhao et al. [58] propose GS LoRA to maintain performance on remaining classes when maximizing the loss on the forgetting data. Ye et al. [53] enforce that the model should be different from the original model in terms of the attention map of intermediate features to forget classes. + +Our work is different from existing works. Firstly, even if the model does not see the unlearned person in training, the ReID model can effectively match the cross-view images of unlearned persons. Hence, the methods eliminating the influence of samples on network weights are not applicable to De-ReID. For the methods that maximize the errors in the forgetting classes, they mainly focus on classification tasks. Differently, ReID pursues discriminative features for matching. Even if the unlearned persons cannot be correctly classified as their methods desired, the model could retrieve the correct images as long as the feature distances between the query and the correct images are smaller than others. + +# 3. De-ReID Learning + +- Problem formulation. Given a well-trained ReID model denoted as $f_{p}(\cdot)$ and a set of unlearned persons $\mathcal{P}_u$ , our goal is to learn a model $f(\cdot)$ that the model cannot match the images of unlearned persons while keeping the ReID performance for other persons that are accessible to model. $f(\cdot)$ is initialized as $f_{p}(\cdot)$ , and $f_{p}(\cdot)$ is kept unchanged. An ideal De-ReID model $f(\cdot)$ is expected to satisfy the conditions: + +$$ +\left\{ \begin{array}{l} \phi_ {f} (x, x _ {p}) < \phi_ {f} (x, x _ {n}), \forall x \in \mathcal {P} _ {u}, \forall x _ {n} \text {s . t .} y _ {n} \neq y, \\ \phi_ {f} (x, x _ {p}) > \phi_ {f} (x, x _ {n}), \forall x \notin \mathcal {P} _ {u}, \forall x _ {n} \text {s . t .} y _ {n} \neq y, \end{array} \right. \tag {1} +$$ + +where $\phi_f(\cdot, \cdot)$ is the similarity metric in the feature space of $f(\cdot)$ . $x$ is a query image with label $y$ , and likewise $y_p$ and $y_n$ are labels of $x_p$ and $x_n$ . $x_p$ is a different image containing the same person with $x$ , but $x_n$ contains a different person. + +To learn $f(\cdot)$ for approximating Eq. 1, we assume a labeled image set of $M_T$ unlearned persons denoted as $S_T = \{(x_i^t,y_i^t)\}_{i = 1}^{|S_T|}$ and an unlabeled image set of accessible persons denoted as $S_O = \{x_j^o\}_{j = 1}^{|S_O|}$ are available for fine-tuning. $x_{i}^{t}$ is the $i$ -th image in $S_T$ with corresponding label $y_{i}^{t}\in \{1,2,\dots,M_{T}\}$ . $x_{j}^{o}$ is the $j$ -th image in $S_O$ . The persons in $S_O$ are different from those in $S_T$ and the persons in testing. While unlearning the specific persons based on the labeled data $S_T$ seems straightforward (e.g., by max + +imizing the classification error), it is hard to ensure only the knowledge strictly subject to matching the unlearned persons is eliminated, rather than simply destroy identity discriminative feature embedding. + +Considering the ubiquitous existence of cross-view variations, we propose variation-guided identity shift based on variations in images to unlearn the specific person in view-invariant feature space. Besides, we formulate the goals of De-ReID in the form of images relations and present relation regularization. An overview of our method is in Fig. 2. + +# 3.1. Variation-guided Identity Shift + +Intuitively, person ReID requires a view-invariant feature space, where image features are identity discriminative regardless of camera views. Hence, we propose to shift the unlearned persons out of the view-invariant feature space, and therefore their images from different camera views cannot be properly associated. To this end, we guide the model to fit the variations in images of unlearned persons and overcome the variations for accessible persons, forming asymmetric learning objectives. In this way, the images of unlearned persons will be encoded into distinct features affected by the variations in images while the images of accessible persons are encoded as view-invariant features. + +To introduce abundant variations, we employ augmentation function $\mathcal{T}(\cdot)$ to obtain an augmented view $\hat{x}_i^t = \mathcal{T}(x_i^t)$ for each image $x_i^t$ . To simulate the natural cross-view variations, $\mathcal{T}(\cdot)$ should not change the identity of $x_i^t$ but properly adjust the images' color, brightness, and so forth, ensuring that $\hat{x}_i^t$ contains the same person with $x_i^t$ but is different in image style. We ablate the design of $\mathcal{T}(\cdot)$ in experiments in our Appendix. To shift the unlearned persons out of the view-invariant feature space, the model $f(\cdot)$ should prioritize rich environmental information unrelated to identity when extracting features from the unlearned person's image $x_i^t$ . In other words, the feature $f(x_i^t)$ should encode rich environmental information (e.g., background, image style) that cannot be applied for matching unlearned person in other cameras. To this end, we guide $f(\cdot)$ to push away $x_i^t$ and its augmented view $\hat{x}_i^t$ in the feature space, which encourages the model to fit the variations for images of the unlearned person. Formally, the learning objective for shifting the unlearned persons out of view-invariant feature space is: + +$$ +L _ {V I S} ^ {p} = \left[ \sigma_ {c} - \| f \left(x _ {i} ^ {t}\right) - f \left(\hat {x} _ {i} ^ {t}\right) \| _ {2} ^ {2} \right] _ {+}, \tag {2} +$$ + +where $[a]_{+} = \max(a, 0.0)$ and $\sigma_{c}$ is the margin to bound the loss. Through optimization, the features of unlearned person images become image-specific and cannot be used for cross-view matching since the features are severely affected by the variations. + +Simply learning with Eq. 2 alone will cause the model to forget the ReID knowledge, and fail to extract features robust to variations for the accessible persons. It will cause + +![](images/2abb36322f7c62f6c6849e2d3af626cdfbd4f4a5fb6e947db5b4bc86bceefd47.jpg) +Figure 2. An overview of our method. In variation-guided identity shift, the model is guided to adapt to the variations when encountering images of unlearned persons, thereby shifting them out of the view-invariant feature space. Meanwhile, the view-invariant feature space is maintained for other accessible persons by guiding the model to overcome the variations when encountering images of these individuals. Relation regularization formulates the desired properties of De-ReID in terms of image relations and ensures the feature discriminability. + +feature space collapse instead of the specific identity shift. To maintain a view-invariant feature space for re-identifying accessible persons, we guide the model to overcome the variations for images of these persons. For image $x_{j}^{o}$ of an accessible person, we obtain its augmented view $\hat{x}_{j}^{o} = \mathcal{T}(x_{j}^{o})$ . We encourage the model to pull closer $x_{j}^{o}$ and $\hat{x}_{j}^{o}$ to remind the model of extracting robust features for accessible persons. Formally, the learning function for accessible person is: + +$$ +L _ {V I S} ^ {o} = \left\| f \left(x _ {j} ^ {o}\right) - f \left(\hat {x} _ {j} ^ {o}\right) \right\| _ {2} ^ {2}. \tag {3} +$$ + +Combining Eq. 3 and Eq. 2, we form an asymmetric objective to shift the identities of unlearned persons while preserving a view-invariant feature space for accessible persons: + +$$ +L _ {V I S} = L _ {V I S} ^ {p} + L _ {V I S} ^ {o}. \tag {4} +$$ + +- Discussion on asymmetric learning scheme. $L_{VIS}$ alters knowledge within the ReID model by treating the variations in images asymmetrically. The two asymmetric parts in the $L_{VIS}$ are collaborative, not isolated. Intuitively, one may think that guiding the ReID network against the variations $(L_{VIS}^{o})$ is merely for keeping ReID knowledge for accessible persons. However, $L_{VIS}^{o}$ is also helpful for unlearning because it maintains a view-invariant feature space, which in turn gives the direction for shifting the identities of unlearned persons $(L_{VIS}^{p})$ : out of the view-invariant space. As a result, more unlabeled images of accessible persons can improve the unlearning effect, which is not shown in other methods. + +# 3.2. Relation Regularization + +The relations between pairwise images are critical in the context of person ReID, directly reflecting the discriminative capability of the features. As formulated in Eq. 1, there should be some accessible persons spread around an unlearned person. Hence, we introduce relation regularization to facilitate De-ReID and ReID learning. Specifically, we + +constraints that, in the feature space, the distance between the images of the same unlearned person should be greater than the distance to the accessible persons. This TRiplet Constraint facilitates retrieving images of different persons when querying an unlearned person's image. + +$$ +L _ {T R C} = \left[ d _ {p} - d _ {n} + \sigma_ {r} \right] _ {+}, +$$ + +$$ +\text {w h e r e} d _ {p} = \left\| x _ {i} ^ {t} - x _ {i, K} ^ {o} \right\| _ {2} ^ {2}, d _ {n} = \left\| x _ {i} ^ {t} - x _ {h} ^ {t} \right\| _ {2} ^ {2}. \tag {5} +$$ + +Here, $x_{i,K}^{o}(x_{i,K}^{o}\in S_{O})$ represents the $K$ -th nearest image to $x_{i}^{t}$ , and $\sigma_r$ is the margin. $x_{h}^{t}$ is an image containing the same person as $x_{i}^{t}\left(y_{h}^{t} = y_{i}^{t}\right)$ . $L_{TRC}$ constrains that the distance between pairwise images of the same unlearned person should be at least larger than $d_p + \sigma_r$ , where $d_p$ is dynamically determined based on the distance to the $K$ -th nearest accessible person. When $K$ is larger, $L_{TRC}$ encourages more accessible persons to spread around $x_{i}^{t}$ . Through optimizing $L_{TRC}$ , the top retrieval results mainly contain different persons when retrieving unlearned person images from feature space. + +Complementary to variation-guided identity shift that changes the cross-view invariance in features, we employ Relation Consistent Regularization on dataset $S_O$ to maintain the feature discriminability during De-ReID fine-tuning. The regularization is to ensure that the relations between accessible individuals remain as discriminative as those in the original model $f_p(\cdot)$ during the unlearning process: + +$$ +L _ {R C R} = \left\| \operatorname {s i m} \left(f _ {p} \left(x _ {j} ^ {o}\right), f _ {p} \left(x _ {h} ^ {o}\right)\right) - \operatorname {s i m} \left(f \left(x _ {j} ^ {o}\right), f \left(x _ {h} ^ {o}\right)\right) \right\| _ {2} ^ {2}. \tag {6} +$$ + +By cooperating with $L_{VIS}$ , the model can extract discriminative and view-invariant features for accessible persons when unlearning the specific persons. + +- Overall Learning objective. The model $f(.)$ is trained end-to-end with an overall learning objective as: + +$$ +L = \lambda_ {1} * L _ {V I S} + \lambda_ {2} * L _ {T R C} + L _ {R C R}, \tag {7} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are trade-off parameters. + +# 4. Experiments + +# 4.1. Datasets and Experimental Setup + +- Datasets. We conduct experiments in Market-1501 [61], MSMT17 [47], Occ-Duke [34] and SYSU-MM01 [49] datasets. Due to the limited space, we leave the results in Occ-Duke and SYSU-MM01 datasets in our Appendix. Market-1501 (abbreviated as Market) consists of 1,501 persons, and all the images are captured from 6 camera views. In the conventional ReID protocol, there are 750 persons for training and 751 persons for testing. MSMT17 is a large-scale ReID dataset, containing images of 4101 persons from 15 disjoint camera views. Conventionally, 1,041 persons are for training and 3,060 persons for testing. For evaluating De-ReID, we slightly modify the conventional protocol by referring to the experimental setups of other machine unlearning works [8, 24, 53, 58]. + +Specifically, in each dataset, we take $M_T$ persons from the testing set as the unlearned persons. Considering the scale of the datasets, we set $M_T \in \{25, 50, 75, 100\}$ in MSMT17 and $M_T \in \{25, 50\}$ in Market1501. For each unlearned person, we split his images into two image sets from disjoint cameras. One set is added to the training set and the other is kept in the testing set. After adding the unlearned persons to the training set, we divide the training set into two subsets. One pre-training subset consists of accessible persons with identity annotations and is used for pretraining the ReID model. There are 900 accessible persons in the pre-training subset of MSMT17 and 600 accessible persons in the pre-training subset of Market-1501. + +The fine-tuning subset consists of unlabeled images of accessible persons and labeled images of unlearned persons, simulating the practical scenario where we need to efficiently fine-tune the ReID model without collecting additional labeled images of accessible persons. The dataset statistics are shown in Table 1. We emphasize that the images of the unlearned persons in the testing set under the cameras are different from those of the training set. The identities of accessible persons in the testing set are different from those in the training set. More details are in Appendix. + +- Experimental setup. In each dataset, the pre-training subset is used to learn a ReID model in a supervised manner as $f_{p}(\cdot)$ . After obtaining the well-trained ReID model, the fine-tuning subset is for learning De-ReID model $f(\cdot)$ in our main experiments. +- Evaluation metrics. We adopt the widely used Cumulative Matching Characteristic (CMC) and mean Average Precision (mAP) to evaluate our performance. Following previous works [38, 58], we also adopt H-Means to comprehensively evaluate the model's ability of ReID in accessible persons + +Table 1. Statistics in the case of $M_T = 100$ in MSMT17 and $M_T = 50$ in Market1501. $M_T$ is the number of unlearned persons for De-ReID. Notably, the accessible persons in training are different from those in testing. "Train" refers to the fine-tuning subset. For the unlearned persons, the images in training and those in testing are from different cameras. More details are in Appendix. + +
MSMT17Market-1501
MT|ST|MO|SO|MT|ST|MO|SO|
Train10042951415061509311512284
Query100533296010706501507003068
Gallery10041362960741655091870018033
+ +and De-ReID in unlearned persons. Formally, + +$$ +\mathrm {H} - \text {M e a n} = \frac {2 * \mathrm {R} - 1 _ {O} (f) * \Delta \mathrm {R} - 1 _ {T}}{\mathrm {R} - 1 _ {O} (f) + \Delta \mathrm {R} - 1 _ {T}}, \tag {8} +$$ + +$$ +w h e r e \Delta R - 1 _ {T} = R - 1 _ {T} (f _ {p}) - R - 1 _ {T} (f). +$$ + +where $\mathrm{R - 1}_O(f)$ is model $f$ 's Rank-1 accuracy in accessible persons, and $\Delta \mathrm{R - 1}_T$ is the changes in the Rank-1 accuracy in unlearned persons after De-ReID fine-tuning using different methods. In the following, we only discuss the performance of $f(\cdot)$ and therefore $\mathrm{R - 1}_O$ means $\mathrm{R - 1}_O(f)$ by default. + +# 4.2. Implementation Details + +By default, we use ViT-B [13] as the ReID backbone $f_{p}(\cdot), f(\cdot)$ . The images are resized to $256 \times 128$ , and the data augmentation includes random crop and flip. For introducing variations in variation-guided identity shift, we use stronger augmentation derived from RandAugment [12]. We detailed the augmentation and conducted corresponding ablation studies in the Appendix. The AdamW [32] optimizer is adopted for De-ReID learning with an initial learning of 3e-4. The weight decay is set to 0. We empirically set trade-off parameters $\lambda_{1} = 1.0$ and $\lambda_{2} = 1.0$ . The batch size for accessible persons is 48, and that for unlearned persons is 32. The above hyper-parameters are the same in all experiments. $K$ in Eq. 5 is set to 20. + +# 4.3. Comparison with Related Methods + +- Compared methods. Since this is the first work for De-ReID, we compare our method with existing methods potentially applicable to ReID. We pretrain a vanilla ViT-B in the pretraining subset of each dataset and then apply these methods to the ViT-B in the fine-tuning subset for De-ReID. We report the performance of the machine unlearning methods, including GS-LoRA [58], LIRF [53], SCRUB [24], and Boundary Shrink [8]. Besides, we implement a naive solution: label augmentation based on cameras, which assign different labels for images containing the same person but under different cameras. This naive solution is denoted as "LabelAug", and the triplet loss and cross-entropy loss [33] are applied for training LabelAug. + +Notably, previous methods usually assume that the labeled data of accessible classes are available. Since the data + +Table 2. Comparisons in the MSMT17 dataset. "BS" is the Boundary Shrink [8]. "H" denotes "H-Mean". R-1 $_T$ is the Rank-1 accuracy on the unlearned persons, and R-1 $_O$ is the Rank-1 accuracy on the accessible persons. R-1 $_T$ is expected to be low for unlearning specific persons, while a higher R-1 $_O$ and H is preferred. Since we apply additional augmentation for asymmetric contrastive learning, we further apply our augmentation to LIRF, denoted as LIRF*. "RR" is the Re-Ranking [63]. The method with the best/second-best H-Mean is marked in red/ blue. Blue background means the H-Mean are improved by Re-Ranking. + +
MethodMT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug54.673.838.156.873.532.4
+RR62.376.929.265.377.221.2
BS63.878.727.354.674.935.2
+RR70.082.218.362.578.725.3
SCRUB43.875.049.145.274.445.1
+RR52.378.441.353.778.236.6
LIRF10.062.566.218.554.656.8
+RR6.965.469.220.557.357.2
LIRF*18.568.865.135.169.752.8
GS-LoRA56.977.535.960.280.728.6
+RR56.979.736.264.182.923.2
Ours4.677.076.310.872.969.7
+RR4.679.077.310.475.170.9
+ +Table 3. Comparison results in the MSMT17 dataset when $M_T = 75$ and $M_T = 100$ . The notations are consistent with Table 2 + +
MethodMT=75MT=100
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug61.472.229.262.771.927.2
+ RR69.175.818.666.875.721.8
BS57.269.133.954.868.336.3
+ RR65.674.323.761.472.728.9
SCRUB54.774.837.550.567.840.6
+ RR62.979.027.755.772.835.9
LIRF18.651.956.120.849.753.8
+ RR19.655.157.524.852.453.5
LIRF*38.465.050.515.650.956.7
GS-LoRA62.479.528.462.176.128.3
+ RR65.381.824.563.878.626.2
Ours12.469.868.513.167.166.7
+ RR13.171.669.015.368.866.4
+ +of accessible persons is unlabeled in our experiments, we apply our $L_{RCR}$ as a regularization for previous methods. We further apply our methods to advanced ReID models, including DCFormaler [27], PAT [28], and PHA [56]. + +- Results. Table 2 and Table 4 shows the comparison results. From the tables, we can observe that our method achieves the best performance in terms of H-Mean. Since the post-processing methods like Re-Ranking [63] can improve the results, we investigate whether these methods can reduce the unlearning effect when retrieving unlearned persons. The results show that existing unlearning methods usually get worse results after post-processing. For example, in Table 2, the Rank-1 accuracy of the SCRUB on unlearned persons is clearly increased from $43.8\%$ to $52.3\%$ when $M_T = 25$ , + +Table 4. Comparison results in the Market-1501 dataset. + +
MethodMT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug50.776.551.364.777.442.5
+RR49.377.652.866.078.941.3
BS57.373.144.562.072.244.3
+RR54.775.347.465.374.041.4
SCRUB50.776.051.254.770.250.4
+RR56.077.446.664.771.841.6
LIRF25.384.872.931.375.968.7
+RR26.784.471.932.775.767.7
LIRF*37.386.064.539.379.064.6
GS-LoRA56.091.748.965.391.443.7
+RR54.791.350.267.391.341.3
Ours10.791.184.412.084.483.2
+RR5.391.087.410.083.583.8
+ +demonstrating that the model does not effectively unlearn the persons and there are potential cues to re-identify them. + +While other methods achieve a worse H-mean after Re-ranking, our method can benefit from Re-Ranking [63] since the overall performance H-Mean is improved in most cases. For example, in the case of $M_T = 25$ in Market dataset, the H-Means of our method is increased to $87.4\%$ from $84.4\%$ . + +Compared with the other methods, Boundary Shrink changes the classification decision boundary of the unlearning classes, i.e., the unlearned persons in this work. However, the feature representation is the core for person image matching instead of the classifier. In contrast, our method directly optimizes the feature representation for De-ReID. As a result, our method outperforms Boundary Shrink by $30.4\%$ H-Mean in MSMT17 with $M_T = 100$ . + +LIRF unlearns classes by constraining the attention map of intermediate features to be significantly different before and after unlearning. Since it directly constrains the features, it obtains the second-best results. However, the attention map has lots of knowledge about extracting identity information from images. Hence, their constraints largely hurt the ReID ability on accessible persons, and our method outperforms LIRF by $12.9\%$ at H-Mean in MSMT17 when $M_T = 100$ . Besides, since we apply additional augmentation in variation-guided identity shift, we further validate the effectiveness of the additional augmentation by applying the augmentation in LIRF, denoted as $\mathrm{LIRF^{*}}$ . However, the H-Mean is improved only in the case of $M_T = 100$ in MSMT17 and decreases in other cases, demonstrating that additional augmentation does not necessarily lead to higher performance. + +GS-LoRA effectively protects the ReID ability on accessible persons but falls behind in unlearning specific persons. LabelAug directly guides the model in distinguishing the images of unlearned persons under different camera views, but it has an unsatisfactory performance. We speculate it is because the model overfits the images in training and fails to unlearn the specific persons under cameras unseen in training. Differently, our method employs variation-guided + +Table 5. Evaluation with different ReID models. Regardless of the ReID models, our method performs the best in terms of erasing the knowledge about unlearned persons (quantified by $\mathrm{R - 1}_T$ ) and keeping ReID knowledge for accessible persons ( $\mathrm{R - 1}_O$ ). "Null" indicates that no De-ReID method is applied. + +
ReID modelDe-ReID methodR-1T↓R-1O↑H↑
DCFormerNull81.284.4—
LabelAug44.154.046.2
GS-LORA75.381.916.4
LIRF18.051.057.7
Ours17.866.166.3
PATNull72.875.6—
LabelAug38.343.340.1
GS-LORA52.560.933.5
LIRF16.742.649.4
Ours12.055.059.0
PHANull70.278.7—
LabelAug44.351.441.2
GS-LORA56.468.333.6
LIRF23.650.552.7
Ours3.951.260.8
+ +identity shift $(L_{VIS}^{p}, L_{VIS}^{o})$ and guides the model to handle the variations adaptively. Fitting the variations in unlearned images helps the model bias to the environmental information when querying the unlearned persons. Simultaneously, the model is guided to be robust to the variations when querying accessible persons. As a result, the unlearned persons are shifted out of the view-invariant feature space while the accessible persons are retained. + +- Evaluations on other ReID methods. Apart from the comparison on the vanilla ViT backbone, we also compared our method with related methods on the advanced ReID models [27, 28, 56]. The experimental results in Table 5 demonstrate that our method is superior to others regarding unlearning different ReID models. + +# 4.4. Ablation Study + +We conduct experiments to show the effectiveness of each component in our design. Due to the space limit, we report the experimental results in Market and $M_T \in \{25, 100\}$ in MSMT17 in Table 6, and the rest in Appendix. Regardless of $M_T$ , our designs clearly improve H-Mean that comprehensively evaluates the effectiveness of unlearning specific persons and ReID performance on accessible persons. + +- The effectiveness of the relation regularization. From Table 6, we first observe that the model suffers from catastrophic forgetting in ReID knowledge and collapses in the absence of $L_{RCR}$ . Although our variation-guided identity shift pulls closer the images and the augmented images for accessible persons, it mainly preserves the view-invariant feature embedding but it fails to preserve the feature discriminability, leading to model collapses. $L_{RCR}$ preserves the discriminability of features by preserving the relations between accessible persons. + +Another constraint $L_{TRC}$ attempts to push away the im + +Table 6. Ablation study in MSMT17 dataset. "W/o" means "without", and other notations are the same as Table 2. $f_{p}(\cdot)$ is the initial model for De-ReID learning. The full model achieves the best H-Mean. Notably, without $L_{RCR}$ , the model forgets ReID knowledge and collapses. "SD" denotes the "self-augmented discrimination". + +
MethodMSMT17
MT=25MT=100
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
fp(·)80.385.4—79.585.4—
Components in Relation regularization
W/o LTRC5.573.474.118.365.363.2
W/o LRCR——————
Components in Variation-guided Identity Shift
W/o SD4.674.575.110.558.763.4
W/o LVIS10.075.872.918.063.262.3
W/o LVIS20.577.467.541.370.049.4
Full model4.677.076.313.167.166.7
+ +Table 7. Ablation study in Market-1501. + +
MethodMarket-1501
MT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
fp(·)89.396.3—94.096.3—
Components in Relation regularization
W/o LTRC20.089.778.224.078.373.9
W/o LRCR——————
Components in Variation-guided Identity Shift
W/o SD24.090.375.811.379.781.2
W/o LVIS10.784.581.412.777.379.2
W/o LVIS46.793.958.655.389.954.1
Full model10.791.184.412.084.483.2
+ +![](images/551864891eb9ef9188b260dba7e6011342b35d93a9366987e5ab409eb8f282aa.jpg) +Figure 3. Evaluation results under various $\lambda_{1}$ and $\lambda_{2}$ . + +![](images/e08e91c4565349eaf9d804561e1689c5da0f9716d202fa83d9f13f6c858ed308.jpg) + +ages between unlearned persons to prevent them from being cross-view retrieved, referring to the distance between accessible persons and unlearned persons. We observe a degraded H-Mean in the absence of $L_{TRC}$ in all cases. For example, when deprecating $L_{TRC}$ in the case of $M_T = 50$ in Market, the Rank-1 accuracy of the unlearned persons is increased to $24.0\%$ from $12.0\%$ , meaning that more unlearned person images are correctly matched and is not expected. As a result, the H-Mean drops by $9.8\%$ compared with the full model. In the case of $M_T = 100$ in MSMT17, the Rank-1 accuracy of the unlearned persons increases to $18.3\%$ from $13.1\%$ without $L_{TRC}$ , and the H-Mean drops by $3.3\%$ . + +The hyper-parameter $\lambda_{2}$ controls the weight of $L_{TRC}$ . We evaluate our model under different $\lambda_{2}$ and show the performance in Fig. 3. When $\lambda_{2} = 0$ , $L_{TRC}$ is not employed and the performance is inferior. As $\lambda_{2}$ increases, the performance is improved. However, $L_{TRC}$ should not be set to + +too large to avoid overwhelming other learning objectives. + +- The effectiveness of the variation-guided identity shift. The variation-guided identity shift (VIS) emphasizes the style difference between images of unlearned persons in the feature space while keeping the other images robust to the style variations. Overall, comparing the results of "W/o $L_{VIS}$ " and the full model in Table 6, we observe that the ReID model struggles to unlock the specific persons without $L_{VIS}$ . For example, in MSMT17 dataset, the Rank-1 accuracy on unlearned persons is $20.5\% / 41.3\%$ when $M_T = 25 / 100$ , which is $15.9\% / 28.2\%$ larger than the full model. This is because ReID requires view-invariant features and our model correspondingly guides the model to be biased to encode environmental information for images of unlearned persons. As a result, we effectively prevent the images of unlearned persons from being re-identified. The hyperparameter $\lambda_1$ controls the weight of $L_{VIS}$ in optimization. We evaluate our model under different $\lambda_1$ and show the performance in Fig. 3. + +We also studied the effectiveness of the designs in VIS. As discussed in the introduction, we argue that simply pushing away the images under different camera views may learn the spurious cues for De-ReID, and therefore we push away the image and the corresponding augmented image in our VIS. We refer to this as "self-augmented discrimination". To verify its effectiveness, we conduct an experiment that pushes away the image and the augmented image from another image, denoted as "W/o SD", and observe a significant performance drop. For example, in the case of $M_T = 25$ in Market dataset, the Rank-1 accuracy on the unlearned person increases to $24.0\%$ from $10.7\%$ without SD, and the H-Mean drops by $8.6\%$ . In other cases, the H-Mean drops without using the SD, demonstrating the effectiveness of SD. + +Moreover, we observe that guiding the ReID to overcome the variations for accessible persons is important for balancing the ReID ability and De-ReID of unlearned persons. Deprecating the $L_{VIS}^{o}$ leads to $3\% \sim 4\%$ performance drops in H-Mean in different cases, and reduce the protection effect since the rank-1 accuracy of unlearned persons increases. + +- The impact of the unlabeled sample. We evaluate our method in MSMT17 under different ratios between unlabeled images of accessible persons and labeled images of unlearned persons in the fine-tuning subset. The results are in Table 8. + +From the results, we obtain two conclusions. (i) More unlabeled data is helpful for both unlearning specific persons and keeping ReID performance on the accessible persons. When the number of accessible persons decreases, the Rank-1 accuracy of the unlearned persons (denoted as $\mathrm{R - 1}_T$ ) increases, which means the unlearned persons are easier to retrieve by the ReID model. Simultaneously, the Rank-1 accuracy of the accessible persons $(\mathrm{R - 1}_O)$ decreases, which means the ReID model cannot properly re-identify + +Table 8. Evaluations under different numbers of accessible persons in MSMT17 when $M_T = 100$ . $M_O$ is the number of accessible persons in the fine-tuning subset. 'With LUP' means using the images in LUPerson dataset as auxiliary unlabeled data and the number in parentheses denotes the performance gain. By default, we do not use LUPerson dataset. + +
\(M_O\)Without LUPWith LUP
R-1T ↓R-1O ↑H ↑R-1T ↓R-1O ↑H ↑
14113.167.166.7---
10015.861.362.514.168.266.8 (+4.3%)
8015.854.358.615.467.765.9 (+7.3%)
6017.451.457.615.465.664.8 (+7.2%)
+ +![](images/4977e97c7dbfa613471b79774d592254c08161398b9f7df3bf13ebfe2706ea62.jpg) +Images of unlearned persons +Figure 4. Attention map on images of the unlearned persons and the images of accessible persons. + +![](images/e6adc81d0df95302d4d74e97fe7c9b2f8002987382ff8cf6e4b05fb59428fbb3.jpg) + +of unlearn + +![](images/1a2ae0c4b8e43b6d2d5290d6024c57328cc196e4fe257c82da6c0cbe9beeed59.jpg) + +d persons + +![](images/455732697bc2e7e6a864634da6bd8bcaf18ead8c9d8a3e86953cb41141430d0f.jpg) +Images of other persons + +![](images/610f1c468e76285bcf022bb8c83b364dabb4dab798967780278a4e75d7097c33.jpg) + +rsons + +the accessible persons. (ii) Although the performance of our method decreases when the unlabeled images become fewer, our method can effectively utilize the publicly available unlabeled person images for unlearning, demonstrating the scalability of our method. When using additional unlabeled images from LUPerson dataset, our method can achieve similar performance even though the number of accessible persons in MSMT17 decreases from 141 to 60. + +# 4.5. Visualization + +We visualize the attention map on different persons in the testing set in Fig. 4. From the figure, for the images of unlearned persons, we can observe that the model's attention mainly focus on the background which contains abundant environmental information. At the same time, the model pays diverse attention to the informative parts of images for extracting discriminative features for accessible persons. These results demonstrate the effectiveness of our method. + +# 5. Conclusion + +To address the privacy concerns in person ReID, we pioneer the exploration of the De-ReID problem that guides the ReID model to forget the knowledge about matching specific persons. We formulate a framework and propose a novel variation-guided identity shift method for De-ReID. Our method unlearns the specific person shifting them out of the view-invariant feature space and simultaneously keeps the ReID knowledge by overcoming the cross-view variations for accessible persons. We further introduce relation regularization to characterize the desired properties of De-ReID. Extensive experiments conducted on a new benchmark show the superiority of our method. + +# 6. Acknowledgment + +This work was supported partially by NSFC(U21A20471, 92470202), by the National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025) + +# References + +[1] Prachi Agrawal and P. J. Narayanan. Person de-identification in videos. IEEE Transactions on Circuits and Systems for Video Technology, 2011. 2 +[2] Shafiq Ahmad, Pietro Morerio, and Alessio Del Bue. Person re-identification without identification via event anonymization. In 2023 IEEE/CVF International Conference on Computer Vision, 2023. 2 +[3] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In IEEE Symposium on Security and Privacy, 2021. 1, 2 +[4] Karla Brkic, Ivan Sikiric, Tomislav Hrkac, and Zoran Kalafatic. I know that person: Generative full body and face de-identification of people in images. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2 +[5] Jonathan Brophy and Daniel Lowd. Machine unlearning for random forests. In International Conference on Machine Learning, 2021. 1, 2 +[6] Chengzhi Cao, Xueyang Fu, Hongjian Liu, Yukun Huang, Kunyu Wang, Jiebo Luo, and Zheng-Jun Zha. Event-guided person re-identification via sparse-dense complementary learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[7] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Learning invariance from generated variance for unsupervised person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[8] Min Chen, Weizhuo Gao, Gaoyang Liu, Kai Peng, and Chen Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 5, 6 +[9] Tianshui Chen, Tao Pu, Lingbo Liu, Yukai Shi, Zhijing Yang, and Liang Lin. Heterogeneous semantic transfer for multi-label recognition with partial labels. International Journal of Computer Vision, 2024. 1 +[10] Weihua Chen, Xiaotang Chen, Jianguo Zhang, and Kaiqi Huang. Beyond triplet loss: A deep quadruplet network for person re-identification. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, 2017. 1, 2 +[11] Dasol Choi and Dongbin Na. Towards machine unlearning benchmarks: Forgetting the personal identities in facial recognition systems. In AAAI 2024 workshop on Privacy-Preserving Artificial Intelligence, 2023. 3 +[12] Ekin D. Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le. Randaugment: Practical automated data augmentation + +with a reduced search space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 5 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5 +[14] Lijie Fan, Tianhong Li, Rongyao Fang, Rumen Hristov, Yuan Yuan, and Dina Katabi. Learning longterm representations for person re-identification using radio signals. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2 +[15] Ivan Filković, Zoran Kalafatić, and Tomislav Hrkač. Deep metric learning for person re-identification and de-identification. In 2016 39th International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO), 2016. 2 +[16] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2 +[17] Wenxuan Guo, Zhiyu Pan, Yingping Liang, Ziheng Xi, Zhi Chen Zhong, Jianjiang Feng, and Jie Zhou. Lidar-based person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[18] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. Advances in Neural Information Processing Systems, 2021. 2 +[19] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In IEEE/CVF International Conference on Computer Vision, 2021. 1, 2 +[20] Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qing-song Xie, Yizhou Wnag, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, Donglian Qi, and Yunfeng Yan. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[21] Bingliang Jiao, Lingqiao Liu, Liying Gao, Guosheng Lin, Lu Yang, Shizhou Zhang, Peng Wang, and Yanning Zhang. Dynamically transformed instance normalization network for generalizable person re-identification. In European Conference on Computer Vision, 2022. 2 +[22] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Andy J Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multiscale dilated transformer for visual recognition. IEEE Transactions on Multimedia, 25:8906-8919, 2023. 2 +[23] Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Cloth-changing person re-identification from a single image with gait prediction and regularization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 + +[24] Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Triantafillou. Towards unbounded machine unlearning. In Conference on Neural Information Processing Systems, 2023. 1, 2, 5 +[25] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: Exploiting vision-language model for image re-identification without concrete text labels. Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2 +[26] Weijia Li, Saihui Hou, Chunjie Zhang, Chunshui Cao, Xu Liu, Yongzhen Huang, and Yao Zhao. An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[27] Wen Li, Cheng Zou, Meng Wang, Furong Xu, Jianan Zhao, Ruobing Zheng, Yuan Cheng, and Wei Chu. Dc-former: Diverse and compact transformer for person re-identification. In AAAI Conference on Artificial Intelligence, 2023. 6, 7 +[28] Yulin Li, Jianfeng He, Tianzhu Zhang, Xiang Liu, Yongdong Zhang, and Feng Wu. Diverse part discovery: Occluded person re-identification with part-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 6, 7 +[29] Kun-Yu Lin, Jiaming Zhou, Yukun Qiu, and Wei-Shi Zheng. Adversarial partial domain adaptation by cycle inconsistency. In European Conference on Computer Vision, pages 530-548. Springer Nature Switzerland Cham, 2022. 2 +[30] Kun-Yu Lin, Jia-Run Du, Yipeng Gao, Jiaming Zhou, and Wei-Shi Zheng. Diversifying spatial-temporal perception for video domain generalization. Advances in Neural Information Processing Systems, 36:56012–56026, 2023. 2 +[31] Kun-Yu Lin, Henghui Ding, Jiaming Zhou, Yu-Ming Tang, Yi-Xing Peng, Zhilin Zhao, Chen Change Loy, and Wei-Shi Zheng. Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560, 2024. 2 +[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5 +[33] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019. 5 +[34] Jiaxu Miao, Yu Wu, Ping Liu, Yuhang Ding, and Yi Yang. Pose-guided feature alignment for occluded person re-identification. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 542–551, 2019. 5 +[35] Tao Pu, Tianshui Chen, Hefeng Wu, Yukai Shi, Zhijing Yang, and Liang Lin. Dual-perspective semantic-aware representation blending for multi-label image recognition with partial labels. Expert Systems with Applications, page 123526, 2024. 1 +[36] Tao Pu, Qianru Lao, Hefeng Wu, Tianshui Chen, Ling Tian, Jie Liu, and Liang Lin. Category-adaptive label discovery and noise rejection for multi-label recognition with partial positive labels. IEEE Transactions on Multimedia, pages 1-12, 2024. 2 + +[37] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 2021. 2 +[38] Takashi Shibata, Go Irie, Daiki Ikami, and Yu Mitsuzumi. Learning with selective forgetting. In International Joint Conference on Artificial Intelligence, 2021. 5 +[39] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and A strong convolutional baseline). In European Conference on Computer Vision, 2018. 2 +[40] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2 +[41] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. Learning to imagine: Diversify memory for incremental learning using unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9549-9558, 2022. 2 +[42] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. When prompt-based incremental learning does not meet strong pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1706-1716, 2023. 2 +[43] Yu-Ming Tang, Yi-Xing Peng, Jingke Meng, and Wei-Shi Zheng. Rethinking few-shot class-incremental learning: Learning from yourself. In European Conference on Computer Vision, pages 108-128. Springer, 2024. 1 +[44] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACM International Conference on Multimedia, 2018. 2 +[45] Haochen Wang, Jiayi Shen, Yongtuo Liu, Yan Gao, and Efstratos Gavves. Nformer: Robust person re-identification with neighbor transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2 +[46] Lin Wang, Wanqian Zhang, Dayan Wu, Fei Zhu, and Bo Li. Attack is the best defense: Towards preemptive-protection person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2 +[47] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 79–88, 2018. 5 +[48] Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. A discriminative feature learning approach for deep face recognition. In European Conference on Computer Vision, 2016. 2 +[49] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 5 +[50] Bin Yang, Jun Chen, and Mang Ye. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In IEEE/CVF International Conference on Computer Vision, 2023. 2 + +[51] Bin Yang, Jun Chen, and Mang Ye. Shallow-deep collaborative learning for unsupervised visible-infrared person re-identification. In IEEE/CVF International Conference on Computer Vision, 2024. 2 +[52] Jinrui Yang, Jiawei Zhang, Fufu Yu, Xinyang Jiang, Mengdan Zhang, Xing Sun, Ying-Cong Chen, and Wei-Shi Zheng. Learning to know where to see: A visibility-aware approach for occluded person re-identification. In IEEE/CVF international conference on computer vision, 2021. 2 +[53] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In European Conference on Computer Vision, 2022. 1, 3, 5 +[54] Mang Ye, Wei Shen, Junwu Zhang, Yao Yang, and Bo Du. Securereid: Privacy-preserving anonymization for person re-identification. IEEE Transactions on Information Forensics and Security, 2024. 2 +[55] Mang Ye, Zesen Wu, Cuiqun Chen, and Bo Du. Channel augmentation for visible-infrared re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[56] Guiwei Zhang, Yongfei Zhang, Tianyu Zhang, Bo Li, and Shiliang Pu. Pha: Patch-wise high-frequency augmentation for transformer-based person re-identification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6, 7 +[57] Junwu Zhang, Mang Ye, and Yao Yang. Learnable privacy-preserving anonymization for pedestrian images. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2 +[58] Hongbo Zhao, Bolin Ni, Haochen Wang, Junsong Fan, Fei Zhu, Yuxi Wang, Yuntao Chen, Gaofeng Meng, and Zhaoxiang Zhang. Continual forgetting for pre-trained vision models. 2024. 3, 5 +[59] Zhilin Zhao, Longbing Cao, and Kun-Yu Lin. Revealing the distributional vulnerability of discriminators by implicit generators. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8888-8901, 2022. 2 +[60] Zhilin Zhao, Longbing Cao, and Kun-Yu Lin. Out-of-distribution detection by cross-class vicinity distribution of in-distribution data. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2 +[61] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Computer Vision, IEEE International Conference on, 2015. 5 +[62] Wei-Shi Zheng, Jincheng Hong, Jiening Jiao, Ancong Wu, Xiatian Zhu, Shaogang Gong, Jiayin Qin, and Jianhuang Lai. Joint bilateral-resolution identity modeling for crossresolution person re-identification. International Journal of Computer Vision, 2022. 2 +[63] Zhun Zhong, Liang Zheng, Donglin Cao, and Shaozi Li. Re-ranking person re-identification with k-reciprocal encoding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 6 \ No newline at end of file diff --git a/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/images.zip b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..36e831bb246b52f03f182f18f580764584fb047c --- /dev/null +++ b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d90addee237f823c894cd6c8e51b86f755751086b4fd8856787dd6254cb6772c +size 539844 diff --git a/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/layout.json b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c23777f7eff35911b5da5eb2eb21ab568ecae9d8 --- /dev/null +++ b/2025/Person De-reidentification_ A Variation-guided Identity Shift Modeling/layout.json @@ -0,0 +1,10995 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 524, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 524, + 121 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 524, + 121 + ], + "type": "text", + "content": "Person De-reidentification: A Variation-guided Identity Shift Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "spans": [ + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "text", + "content": "Yi-Xing Peng" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "text", + "content": " Yu-Ming Tang" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "text", + "content": " Kun-Yu Lin" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "text", + "content": " Qize Yang" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "text", + "content": " Jingke Meng" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1,4*}" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "text", + "content": " Xihan Wei" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "text", + "content": " Wei-Shi Zheng" + }, + { + "bbox": [ + 118, + 141, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,4*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 171, + 495, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 171, + 495, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 171, + 495, + 185 + ], + "type": "text", + "content": "1School of Computer Science and Engineering, Sun Yat-sen University, China;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 185, + 483, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 185, + 483, + 199 + ], + "spans": [ + { + "bbox": [ + 127, + 185, + 483, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 185, + 483, + 199 + ], + "type": "text", + "content": "Peng Cheng Laboratory, Shenzhen, China; " + }, + { + "bbox": [ + 127, + 185, + 483, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 185, + 483, + 199 + ], + "type": "text", + "content": "Tongyi Lab, Alibaba Group;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 199, + 543, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 199, + 543, + 213 + ], + "spans": [ + { + "bbox": [ + 66, + 199, + 543, + 213 + ], + "type": "text", + "content": "4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 216, + 477, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 216, + 477, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 216, + 477, + 227 + ], + "type": "text", + "content": "{pengyx23,tangym9,linky5}@mail2.sysu.edu.cn wszheng@ieee.org" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 280, + 297, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 297, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 297, + 604 + ], + "type": "text", + "content": "Person re-identification (ReID) is to associate images of individuals from different camera views against cross-view variations. Like other surveillance technologies, Re-ID faces serious privacy challenges, particularly the potential for unauthorized tracking. Although various tasks (e.g., face recognition) have developed machine unlearning techniques to address privacy concerns, such methods have not yet been explored within the Re-ID field. In this work, we pioneer the exploration of the person de-reidentification (De-ReID) problem and present its inherent challenges. In the context of ReID, De-ReID is to unlock the knowledge about accurately matching specific persons so that these \"unlearned persons\" cannot be re-identified across cameras for privacy guarantee. The primary challenge is to achieve the unlearning without degrading the identity-discriminative feature embeddings to ensure the model's utility. To address this, we formulate a De-ReID framework that utilizes a labeled dataset of unlearned persons for unlearning and an unlabeled dataset of accessible persons for knowledge preservation. Instead of unlearning based on (pseudo) identity labels, we introduce a variation-guided identity shift mechanism that unlearns the specific persons by fitting the variations in their images while preserving ReID ability on other persons by overcoming the variations in images of accessible persons. As a result, the model shifts the unlearned persons to a feature space that is vulnerable to cross-view variations. Extensive experiments on benchmarks demonstrate the superiority of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 626, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 626, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 626, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "Person re-identification (ReID) aims to match the images of the same person across different camera views based on feature similarity. While deep learning [9, 35, 43] has significantly enhanced the performance of the ReID mod" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 255, + 550, + 315 + ], + "blocks": [ + { + "bbox": [ + 316, + 255, + 550, + 315 + ], + "lines": [ + { + "bbox": [ + 316, + 255, + 550, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 550, + 315 + ], + "type": "image", + "image_path": "a3a019974be3f120f13baeb520587213302d3b239815ca7788817d7c063accb2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 320, + 555, + 386 + ], + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 386 + ], + "type": "text", + "content": "Figure 1. Person de-reidentification (De-ReID) prevents intelligent surveillance from tracking unlearned persons. For example, when using surveillance models for security (track and record the person in the red box), the residents in the community and the important customers in business (in the green box) should not be tracked for privacy and commercial secrets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 388, + 556, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 496 + ], + "type": "text", + "content": "els [10, 19, 45], there are growing public concerns about potential misuse, leading to ethical and societal implications. To avoid misuse and mitigate the potential privacy concerns, the research in surveillance has begun exploring machine unlearning techniques that limit the operational scope of these models. For instance, in face recognition, models are adjusted to \"unlearn\" individuals who raise privacy concerns, as mandated by regulations like the EU's General Data Protection Regulation (GDPR) [3, 5, 24, 53]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 498, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 665 + ], + "type": "text", + "content": "However, similar constraints on the ReID models have not yet to be explored, despite its significance for real-world applications. For instance, the residents in the neighborhood prefer to constrain the surveillance in the community to only focus on external people and not track themselves; in working buildings, security cameras should only be set for monitoring and preventing theft or accidents, not for tracking or identifying specific employees during their routine work; and homeowners expect smart home devices to track only unknown individuals for security purposes while avoiding any continuous surveillance of themselves; To meet these requirements, there is an urgent need to develop methods that limit the scope of ReID models, enabling them to \"ignore\" specific individuals during deployment." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "In this work, we pioneer the exploration of person deidentification (De-ReID) problem: limiting the ReID scope by unlearning specific persons (referred to as \"unlearned persons\") in the ReID model while maintaining the ReID" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29331" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 288 + ], + "type": "text", + "content": "performance on other persons (referred to as \"accessible persons\"). In the context of image matching, a person is unlearned by a model means the model can hardly associate the images of the person under different camera views, as shown in Fig. 1. Forming a De-ReID model is challenging because only the knowledge strictly related to matching the unlearned persons should be eliminated, without destroying the identity-discriminative feature embeddings. A trivial solution is to distinguish the images of unlearned persons under different camera views for unlearning while associating the images of accessible persons based on their identity labels to preserve ReID knowledge. However, due to the large network capacity, the model may learn spurious cues or overfit the images of unlearned persons in the training set, leading to poor generalization during evaluation. Additionally, collecting identity labels for accessible persons to preserve ReID knowledge is both costly and difficult to scale, especially when the number of cameras is large [29, 31, 36]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 297, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 297, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 297, + 456 + ], + "type": "text", + "content": "To handle the De-ReID, we formulate a weakly-supervised unlearning framework where a labeled image set of unlearned persons and an unlabeled dataset of accessible persons are available to adapt a well-trained ReID model to De-ReID. Given that the core of ReID is to overcome cross-view variations, such as different camera viewing angles and lighting conditions, we propose variation-guided identity shift that learns De-ReID based on the variations in images. The core of our method is to guide the ReID model bias towards environmental information (e.g., lighting) when encountering unlearned persons, which shifts the unlearned persons out of the view-invariant feature space. At the same time, we keep the model remains robust to cross-view variations when querying accessible persons." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 459, + 297, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 297, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 297, + 626 + ], + "type": "text", + "content": "Specifically, for unlearned persons, we guide the model to distinguish their images and the images augmented by variations. Compared with distinguishing images under different camera views, learning to distinguish image and its augmentation results is more effective for unlearning since the augmented images are highly similar to the original image (e.g., body shape) but varies in style information (e.g., brightness and color tone). Therefore, the model is biased to encode the style information to push away these images. In contrast, for accessible persons, we pull closer their images and augmented versions in the feature space, expecting the model to overcome the variations when matching them. Moreover, we introduce relation regularization to characterize the desired properties of De-ReID in terms of relations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 628, + 274, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 628, + 274, + 639 + ], + "spans": [ + { + "bbox": [ + 67, + 628, + 274, + 639 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 642, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 642, + 294, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 294, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 294, + 664 + ], + "type": "text", + "content": "- We are the first to investigate the challenging De-ReID task and conduct a new benchmark." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 666, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 666, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 666, + 295, + 713 + ], + "type": "text", + "content": "- We formulate a De-ReID framework and propose variation-guided identity shift that achieves De-ReID by fitting the variations in images of interest while against the variations in others to preserve ReID knowledge." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 556, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 556, + 133 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 556, + 133 + ], + "type": "text", + "content": "- To further utilize the identity annotations in unlearned person images for learning, we propose relation regularization that characterizes the desired properties of ReID and De-ReID. Both qualitative and quantitative results demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 143, + 406, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 143, + 406, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 143, + 406, + 156 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 163, + 555, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 270 + ], + "type": "text", + "content": "- Person re-identification. Based on deep learning, advanced ReID models automatically learn a discriminative feature embedding from a large amount of data [20]. Various network architectures [19, 21, 22, 30, 39, 42, 44, 45] and loss functions [10, 40, 41, 48] are developed to enhance feature representation learning. Besides, the challenges in ReID including cross-modality [25, 31, 51, 55], cloth-changing [23, 26], low-resolution [62], occlusion [52], and unsupervised ReID [7, 50] are also widely studied." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 271, + 556, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 556, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 556, + 354 + ], + "type": "text", + "content": "- Privacy protection for person images. Apart from pursuing high performance, protecting human privacy in human-centric tasks is also important [54, 59, 60]. A stream of research focuses on generating anonymized images by removing or obscuring identity-related information, such as faces and silhouettes, while preserving the image's utility for subsequent tasks, such as action recognition [1, 4, 15]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 355, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 556, + 474 + ], + "type": "text", + "content": "In the field of ReID, Ye et al. [54, 57] proposed a generative model to anonymize the person images to avoid privacy leakage while keeping the ReID utility. Wang et al. [46] add imperceptible perturbations to the images of in the database to prevent malicious use of the images. Particularly, one stream of works explores person de-identification [1, 4, 15], which obscures the identity information in images or videos, including face and silhouette. The new sensors including event cameras and LiDAR are also investigated for privacy-preserving ReID [2, 6, 14, 17]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 556, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 642 + ], + "type": "text", + "content": "Despite their progress, previous works mainly focus on preventing the malicious use of images while we aim to prevent the malicious use of models. Technically, previous works [54, 57] learn to anonymize person images and jointly adapt the ReID model to ensure ReID utility. Their methods cannot prevent the ReID model from retrieving specific persons, as well as the methods that explore new sensors such as event cameras for ReID [2, 6, 14, 17]. Besides, although PRIDE [46] protects the images in the database from being maliciously used, the attacker can also collect images that are not in the database to track specific persons. Moreover, some works in person de-identification only concern the identity-unrelated utility [1, 4, 15], such as action recognition, while De-ReID requires identifying accessible persons." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 701 + ], + "type": "text", + "content": "- Machine unlearning. De-ReID has a similar goal to machine unlearning. Machine unlearning is first proposed to eliminate the influence of specific data on the deep model as if the model never uses these data for training, which enables users to erase their personal data [3, 5, 16, 18, 37]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 702, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 702, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 702, + 555, + 714 + ], + "type": "text", + "content": "Kurmanji et al. [24] further apply machine unlearning for" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29332" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "type": "text", + "content": "various purposes such as removing the bias from the model or eliminating the negative effects of mislabeled samples. To maximize the error in forgetting classes, they maximize the difference between the outputs of the learned model and the original model in the forgetting samples. Choi et al. [11] propose to jointly maximize the loss on the forgetting data and learn on the other data. Chen et al. [8] unlearn specific classes by adjusting their decision boundary. The forgetting samples are assigned a nearest but incorrect label for boundary shrink. Zhao et al. [58] propose GS LoRA to maintain performance on remaining classes when maximizing the loss on the forgetting data. Ye et al. [53] enforce that the model should be different from the original model in terms of the attention map of intermediate features to forget classes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 384 + ], + "type": "text", + "content": "Our work is different from existing works. Firstly, even if the model does not see the unlearned person in training, the ReID model can effectively match the cross-view images of unlearned persons. Hence, the methods eliminating the influence of samples on network weights are not applicable to De-ReID. For the methods that maximize the errors in the forgetting classes, they mainly focus on classification tasks. Differently, ReID pursues discriminative features for matching. Even if the unlearned persons cannot be correctly classified as their methods desired, the model could retrieve the correct images as long as the feature distances between the query and the correct images are smaller than others." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 394, + 167, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 167, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 167, + 408 + ], + "type": "text", + "content": "3. De-ReID Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "text", + "content": "- Problem formulation. Given a well-trained ReID model denoted as " + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "text", + "content": " and a set of unlearned persons " + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_u" + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "text", + "content": ", our goal is to learn a model " + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "text", + "content": " that the model cannot match the images of unlearned persons while keeping the ReID performance for other persons that are accessible to model. " + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "text", + "content": " is initialized as " + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "text", + "content": " is kept unchanged. An ideal De-ReID model " + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 54, + 415, + 295, + 499 + ], + "type": "text", + "content": " is expected to satisfy the conditions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 508, + 296, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 508, + 296, + 549 + ], + "spans": [ + { + "bbox": [ + 63, + 508, + 296, + 549 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\phi_ {f} (x, x _ {p}) < \\phi_ {f} (x, x _ {n}), \\forall x \\in \\mathcal {P} _ {u}, \\forall x _ {n} \\text {s . t .} y _ {n} \\neq y, \\\\ \\phi_ {f} (x, x _ {p}) > \\phi_ {f} (x, x _ {n}), \\forall x \\notin \\mathcal {P} _ {u}, \\forall x _ {n} \\text {s . t .} y _ {n} \\neq y, \\end{array} \\right. \\tag {1}", + "image_path": "7593076fb173f70a9479ca289f01f778c098ad842d12d70298fa2b4e4d15a2d3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "inline_equation", + "content": "\\phi_f(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": " is the similarity metric in the feature space of " + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": " is a query image with label " + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": ", and likewise " + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "inline_equation", + "content": "y_p" + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "inline_equation", + "content": "y_n" + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": " are labels of " + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "inline_equation", + "content": "x_p" + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "inline_equation", + "content": "x_n" + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "inline_equation", + "content": "x_p" + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": " is a different image containing the same person with " + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": ", but " + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "inline_equation", + "content": "x_n" + }, + { + "bbox": [ + 55, + 550, + 295, + 597 + ], + "type": "text", + "content": " contains a different person." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": "To learn " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": " for approximating Eq. 1, we assume a labeled image set of " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": " unlearned persons denoted as " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_T = \\{(x_i^t,y_i^t)\\}_{i = 1}^{|S_T|}" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": " and an unlabeled image set of accessible persons denoted as " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_O = \\{x_j^o\\}_{j = 1}^{|S_O|}" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": " are available for fine-tuning. " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{i}^{t}" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": "-th image in " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_T" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": " with corresponding label " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "y_{i}^{t}\\in \\{1,2,\\dots,M_{T}\\}" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{j}^{o}" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": "-th image in " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_O" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": ". The persons in " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_O" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": " are different from those in " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_T" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": " and the persons in testing. While unlearning the specific persons based on the labeled data " + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_T" + }, + { + "bbox": [ + 55, + 597, + 296, + 714 + ], + "type": "text", + "content": " seems straightforward (e.g., by max" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "imizing the classification error), it is hard to ensure only the knowledge strictly subject to matching the unlearned persons is eliminated, rather than simply destroy identity discriminative feature embedding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 120, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 555, + 192 + ], + "type": "text", + "content": "Considering the ubiquitous existence of cross-view variations, we propose variation-guided identity shift based on variations in images to unlearn the specific person in view-invariant feature space. Besides, we formulate the goals of De-ReID in the form of images relations and present relation regularization. An overview of our method is in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 199, + 481, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 481, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 481, + 212 + ], + "type": "text", + "content": "3.1. Variation-guided Identity Shift" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 216, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 358 + ], + "type": "text", + "content": "Intuitively, person ReID requires a view-invariant feature space, where image features are identity discriminative regardless of camera views. Hence, we propose to shift the unlearned persons out of the view-invariant feature space, and therefore their images from different camera views cannot be properly associated. To this end, we guide the model to fit the variations in images of unlearned persons and overcome the variations for accessible persons, forming asymmetric learning objectives. In this way, the images of unlearned persons will be encoded into distinct features affected by the variations in images while the images of accessible persons are encoded as view-invariant features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": "To introduce abundant variations, we employ augmentation function " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " to obtain an augmented view " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^t = \\mathcal{T}(x_i^t)" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " for each image " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "x_i^t" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": ". To simulate the natural cross-view variations, " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " should not change the identity of " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "x_i^t" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " but properly adjust the images' color, brightness, and so forth, ensuring that " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^t" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " contains the same person with " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "x_i^t" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " but is different in image style. We ablate the design of " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " in experiments in our Appendix. To shift the unlearned persons out of the view-invariant feature space, the model " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " should prioritize rich environmental information unrelated to identity when extracting features from the unlearned person's image " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "x_i^t" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": ". In other words, the feature " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "f(x_i^t)" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " should encode rich environmental information (e.g., background, image style) that cannot be applied for matching unlearned person in other cameras. To this end, we guide " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " to push away " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "x_i^t" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " and its augmented view " + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^t" + }, + { + "bbox": [ + 313, + 360, + 556, + 587 + ], + "type": "text", + "content": " in the feature space, which encourages the model to fit the variations for images of the unlearned person. Formally, the learning objective for shifting the unlearned persons out of view-invariant feature space is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 595, + 555, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 595, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 359, + 595, + 555, + 610 + ], + "type": "interline_equation", + "content": "L _ {V I S} ^ {p} = \\left[ \\sigma_ {c} - \\| f \\left(x _ {i} ^ {t}\\right) - f \\left(\\hat {x} _ {i} ^ {t}\\right) \\| _ {2} ^ {2} \\right] _ {+}, \\tag {2}", + "image_path": "fc8028a867d095299426eecfd65fea7e4d0e9de7769439f21b14498993fcf465.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 555, + 676 + ], + "type": "inline_equation", + "content": "[a]_{+} = \\max(a, 0.0)" + }, + { + "bbox": [ + 313, + 617, + 555, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\sigma_{c}" + }, + { + "bbox": [ + 313, + 617, + 555, + 676 + ], + "type": "text", + "content": " is the margin to bound the loss. Through optimization, the features of unlearned person images become image-specific and cannot be used for cross-view matching since the features are severely affected by the variations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Simply learning with Eq. 2 alone will cause the model to forget the ReID knowledge, and fail to extract features robust to variations for the accessible persons. It will cause" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29333" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 70, + 503, + 216 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 503, + 216 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 503, + 216 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 503, + 216 + ], + "type": "image", + "image_path": "2abb36322f7c62f6c6849e2d3af626cdfbd4f4a5fb6e947db5b4bc86bceefd47.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 219, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 54, + 219, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 54, + 219, + 555, + 265 + ], + "type": "text", + "content": "Figure 2. An overview of our method. In variation-guided identity shift, the model is guided to adapt to the variations when encountering images of unlearned persons, thereby shifting them out of the view-invariant feature space. Meanwhile, the view-invariant feature space is maintained for other accessible persons by guiding the model to overcome the variations when encountering images of these individuals. Relation regularization formulates the desired properties of De-ReID in terms of image relations and ensures the feature discriminability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": "feature space collapse instead of the specific identity shift. To maintain a view-invariant feature space for re-identifying accessible persons, we guide the model to overcome the variations for images of these persons. For image " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "x_{j}^{o}" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " of an accessible person, we obtain its augmented view " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\hat{x}_{j}^{o} = \\mathcal{T}(x_{j}^{o})" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": ". We encourage the model to pull closer " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "x_{j}^{o}" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\hat{x}_{j}^{o}" + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": " to remind the model of extracting robust features for accessible persons. Formally, the learning function for accessible person is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 380, + 296, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 380, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 117, + 380, + 296, + 397 + ], + "type": "interline_equation", + "content": "L _ {V I S} ^ {o} = \\left\\| f \\left(x _ {j} ^ {o}\\right) - f \\left(\\hat {x} _ {j} ^ {o}\\right) \\right\\| _ {2} ^ {2}. \\tag {3}", + "image_path": "791418149dd4e831de972685c8c965a5be2fdb5e8f7fa81d068d2795661e7f40.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 297, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 297, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 297, + 441 + ], + "type": "text", + "content": "Combining Eq. 3 and Eq. 2, we form an asymmetric objective to shift the identities of unlearned persons while preserving a view-invariant feature space for accessible persons:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 453, + 296, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 453, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 125, + 453, + 296, + 467 + ], + "type": "interline_equation", + "content": "L _ {V I S} = L _ {V I S} ^ {p} + L _ {V I S} ^ {o}. \\tag {4}", + "image_path": "aad679be0f9121f1b2e625a5756c9d7be9cf70687e598086e3808af5664c5506.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "text", + "content": "- Discussion on asymmetric learning scheme. " + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "inline_equation", + "content": "L_{VIS}" + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "text", + "content": " alters knowledge within the ReID model by treating the variations in images asymmetrically. The two asymmetric parts in the " + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "inline_equation", + "content": "L_{VIS}" + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "text", + "content": " are collaborative, not isolated. Intuitively, one may think that guiding the ReID network against the variations " + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "inline_equation", + "content": "(L_{VIS}^{o})" + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "text", + "content": " is merely for keeping ReID knowledge for accessible persons. However, " + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "inline_equation", + "content": "L_{VIS}^{o}" + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "text", + "content": " is also helpful for unlearning because it maintains a view-invariant feature space, which in turn gives the direction for shifting the identities of unlearned persons " + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "inline_equation", + "content": "(L_{VIS}^{p})" + }, + { + "bbox": [ + 55, + 472, + 296, + 616 + ], + "type": "text", + "content": ": out of the view-invariant space. As a result, more unlabeled images of accessible persons can improve the unlearning effect, which is not shown in other methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 190, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 190, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 190, + 636 + ], + "type": "text", + "content": "3.2. Relation Regularization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "The relations between pairwise images are critical in the context of person ReID, directly reflecting the discriminative capability of the features. As formulated in Eq. 1, there should be some accessible persons spread around an unlearned person. Hence, we introduce relation regularization to facilitate De-ReID and ReID learning. Specifically, we" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 277, + 555, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 337 + ], + "type": "text", + "content": "constraints that, in the feature space, the distance between the images of the same unlearned person should be greater than the distance to the accessible persons. This TRiplet Constraint facilitates retrieving images of different persons when querying an unlearned person's image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 380, + 341, + 492, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 341, + 492, + 354 + ], + "spans": [ + { + "bbox": [ + 380, + 341, + 492, + 354 + ], + "type": "interline_equation", + "content": "L _ {T R C} = \\left[ d _ {p} - d _ {n} + \\sigma_ {r} \\right] _ {+},", + "image_path": "8c35919c7c90f13c4bc817122682021e43816f7bfd724556c4c94cffc7a5fd5d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 351, + 554, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 351, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 343, + 351, + 554, + 370 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} d _ {p} = \\left\\| x _ {i} ^ {t} - x _ {i, K} ^ {o} \\right\\| _ {2} ^ {2}, d _ {n} = \\left\\| x _ {i} ^ {t} - x _ {h} ^ {t} \\right\\| _ {2} ^ {2}. \\tag {5}", + "image_path": "ed460e7e346583c828db465b953e850905177efe648fbea38b7cf32e939933f0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "x_{i,K}^{o}(x_{i,K}^{o}\\in S_{O})" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": "-th nearest image to " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "x_{i}^{t}" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\sigma_r" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": " is the margin. " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "x_{h}^{t}" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": " is an image containing the same person as " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "x_{i}^{t}\\left(y_{h}^{t} = y_{i}^{t}\\right)" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "L_{TRC}" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": " constrains that the distance between pairwise images of the same unlearned person should be at least larger than " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "d_p + \\sigma_r" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "d_p" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": " is dynamically determined based on the distance to the " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": "-th nearest accessible person. When " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": " is larger, " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "L_{TRC}" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": " encourages more accessible persons to spread around " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "x_{i}^{t}" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": ". Through optimizing " + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "inline_equation", + "content": "L_{TRC}" + }, + { + "bbox": [ + 313, + 374, + 555, + 506 + ], + "type": "text", + "content": ", the top retrieval results mainly contain different persons when retrieving unlearned person images from feature space." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 507, + 556, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 556, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 556, + 591 + ], + "type": "text", + "content": "Complementary to variation-guided identity shift that changes the cross-view invariance in features, we employ Relation Consistent Regularization on dataset " + }, + { + "bbox": [ + 313, + 507, + 556, + 591 + ], + "type": "inline_equation", + "content": "S_O" + }, + { + "bbox": [ + 313, + 507, + 556, + 591 + ], + "type": "text", + "content": " to maintain the feature discriminability during De-ReID fine-tuning. The regularization is to ensure that the relations between accessible individuals remain as discriminative as those in the original model " + }, + { + "bbox": [ + 313, + 507, + 556, + 591 + ], + "type": "inline_equation", + "content": "f_p(\\cdot)" + }, + { + "bbox": [ + 313, + 507, + 556, + 591 + ], + "type": "text", + "content": " during the unlearning process:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 594, + 555, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 594, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 319, + 594, + 555, + 619 + ], + "type": "interline_equation", + "content": "L _ {R C R} = \\left\\| \\operatorname {s i m} \\left(f _ {p} \\left(x _ {j} ^ {o}\\right), f _ {p} \\left(x _ {h} ^ {o}\\right)\\right) - \\operatorname {s i m} \\left(f \\left(x _ {j} ^ {o}\\right), f \\left(x _ {h} ^ {o}\\right)\\right) \\right\\| _ {2} ^ {2}. \\tag {6}", + "image_path": "f7a336f6ecfd723620323f8044b940dfe1429a93cefc7a202054e0b813503940.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 620, + 555, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 555, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 555, + 655 + ], + "type": "text", + "content": "By cooperating with " + }, + { + "bbox": [ + 313, + 620, + 555, + 655 + ], + "type": "inline_equation", + "content": "L_{VIS}" + }, + { + "bbox": [ + 313, + 620, + 555, + 655 + ], + "type": "text", + "content": ", the model can extract discriminative and view-invariant features for accessible persons when unlearning the specific persons." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 655, + 554, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 655, + 554, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 655, + 554, + 680 + ], + "type": "text", + "content": "- Overall Learning objective. The model " + }, + { + "bbox": [ + 313, + 655, + 554, + 680 + ], + "type": "inline_equation", + "content": "f(.)" + }, + { + "bbox": [ + 313, + 655, + 554, + 680 + ], + "type": "text", + "content": " is trained end-to-end with an overall learning objective as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 352, + 685, + 555, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 685, + 555, + 697 + ], + "spans": [ + { + "bbox": [ + 352, + 685, + 555, + 697 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {1} * L _ {V I S} + \\lambda_ {2} * L _ {T R C} + L _ {R C R}, \\tag {7}", + "image_path": "3fe048ad4a9b75e6a18983fe517b5a4b8055f95ce387cc903c0ac96332bb2ebd.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 701, + 485, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 485, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 485, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 485, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 701, + 485, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 701, + 485, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 701, + 485, + 714 + ], + "type": "text", + "content": " are trade-off parameters." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29334" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 98, + 234, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 98, + 234, + 110 + ], + "spans": [ + { + "bbox": [ + 55, + 98, + 234, + 110 + ], + "type": "text", + "content": "4.1. Datasets and Experimental Setup" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 122, + 297, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 122, + 297, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 122, + 297, + 291 + ], + "type": "text", + "content": "- Datasets. We conduct experiments in Market-1501 [61], MSMT17 [47], Occ-Duke [34] and SYSU-MM01 [49] datasets. Due to the limited space, we leave the results in Occ-Duke and SYSU-MM01 datasets in our Appendix. Market-1501 (abbreviated as Market) consists of 1,501 persons, and all the images are captured from 6 camera views. In the conventional ReID protocol, there are 750 persons for training and 751 persons for testing. MSMT17 is a large-scale ReID dataset, containing images of 4101 persons from 15 disjoint camera views. Conventionally, 1,041 persons are for training and 3,060 persons for testing. For evaluating De-ReID, we slightly modify the conventional protocol by referring to the experimental setups of other machine unlearning works [8, 24, 53, 58]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 453 + ], + "type": "text", + "content": "Specifically, in each dataset, we take " + }, + { + "bbox": [ + 55, + 297, + 296, + 453 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 55, + 297, + 296, + 453 + ], + "type": "text", + "content": " persons from the testing set as the unlearned persons. Considering the scale of the datasets, we set " + }, + { + "bbox": [ + 55, + 297, + 296, + 453 + ], + "type": "inline_equation", + "content": "M_T \\in \\{25, 50, 75, 100\\}" + }, + { + "bbox": [ + 55, + 297, + 296, + 453 + ], + "type": "text", + "content": " in MSMT17 and " + }, + { + "bbox": [ + 55, + 297, + 296, + 453 + ], + "type": "inline_equation", + "content": "M_T \\in \\{25, 50\\}" + }, + { + "bbox": [ + 55, + 297, + 296, + 453 + ], + "type": "text", + "content": " in Market1501. For each unlearned person, we split his images into two image sets from disjoint cameras. One set is added to the training set and the other is kept in the testing set. After adding the unlearned persons to the training set, we divide the training set into two subsets. One pre-training subset consists of accessible persons with identity annotations and is used for pretraining the ReID model. There are 900 accessible persons in the pre-training subset of MSMT17 and 600 accessible persons in the pre-training subset of Market-1501." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 297, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 297, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 297, + 581 + ], + "type": "text", + "content": "The fine-tuning subset consists of unlabeled images of accessible persons and labeled images of unlearned persons, simulating the practical scenario where we need to efficiently fine-tune the ReID model without collecting additional labeled images of accessible persons. The dataset statistics are shown in Table 1. We emphasize that the images of the unlearned persons in the testing set under the cameras are different from those of the training set. The identities of accessible persons in the testing set are different from those in the training set. More details are in Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 586, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 54, + 586, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 586, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 54, + 586, + 295, + 647 + ], + "type": "text", + "content": "- Experimental setup. In each dataset, the pre-training subset is used to learn a ReID model in a supervised manner as " + }, + { + "bbox": [ + 54, + 586, + 295, + 647 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 54, + 586, + 295, + 647 + ], + "type": "text", + "content": ". After obtaining the well-trained ReID model, the fine-tuning subset is for learning De-ReID model " + }, + { + "bbox": [ + 54, + 586, + 295, + 647 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 54, + 586, + 295, + 647 + ], + "type": "text", + "content": " in our main experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "content": "- Evaluation metrics. We adopt the widely used Cumulative Matching Characteristic (CMC) and mean Average Precision (mAP) to evaluate our performance. Following previous works [38, 58], we also adopt H-Means to comprehensively evaluate the model's ability of ReID in accessible persons" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 320, + 138, + 552, + 198 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 137 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 137 + ], + "type": "text", + "content": "Table 1. Statistics in the case of " + }, + { + "bbox": [ + 313, + 70, + 555, + 137 + ], + "type": "inline_equation", + "content": "M_T = 100" + }, + { + "bbox": [ + 313, + 70, + 555, + 137 + ], + "type": "text", + "content": " in MSMT17 and " + }, + { + "bbox": [ + 313, + 70, + 555, + 137 + ], + "type": "inline_equation", + "content": "M_T = 50" + }, + { + "bbox": [ + 313, + 70, + 555, + 137 + ], + "type": "text", + "content": " in Market1501. " + }, + { + "bbox": [ + 313, + 70, + 555, + 137 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 313, + 70, + 555, + 137 + ], + "type": "text", + "content": " is the number of unlearned persons for De-ReID. Notably, the accessible persons in training are different from those in testing. \"Train\" refers to the fine-tuning subset. For the unlearned persons, the images in training and those in testing are from different cameras. More details are in Appendix." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 138, + 552, + 198 + ], + "lines": [ + { + "bbox": [ + 320, + 138, + 552, + 198 + ], + "spans": [ + { + "bbox": [ + 320, + 138, + 552, + 198 + ], + "type": "table", + "html": "
MSMT17Market-1501
MT|ST|MO|SO|MT|ST|MO|SO|
Train10042951415061509311512284
Query100533296010706501507003068
Gallery10041362960741655091870018033
", + "image_path": "e8b62b255fd692321cb2179166104c310a3914f8903d695a0cc49dfc143d94a6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 201, + 499, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 499, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 499, + 213 + ], + "type": "text", + "content": "and De-ReID in unlearned persons. Formally," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 373, + 221, + 555, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 221, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 373, + 221, + 555, + 248 + ], + "type": "interline_equation", + "content": "\\mathrm {H} - \\text {M e a n} = \\frac {2 * \\mathrm {R} - 1 _ {O} (f) * \\Delta \\mathrm {R} - 1 _ {T}}{\\mathrm {R} - 1 _ {O} (f) + \\Delta \\mathrm {R} - 1 _ {T}}, \\tag {8}", + "image_path": "53a2dd553b67f6e6fe0d7f23659c607589e7ba60d0ad710ed9d6c1649796a776.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 249, + 511, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 249, + 511, + 262 + ], + "spans": [ + { + "bbox": [ + 351, + 249, + 511, + 262 + ], + "type": "interline_equation", + "content": "w h e r e \\Delta R - 1 _ {T} = R - 1 _ {T} (f _ {p}) - R - 1 _ {T} (f).", + "image_path": "81146ba2ad227a406bd2be3e781596a8ca20eb02553fc4657c8d61cbe3ceede1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{R - 1}_O(f)" + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "text", + "content": " is model " + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "text", + "content": "'s Rank-1 accuracy in accessible persons, and " + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{R - 1}_T" + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "text", + "content": " is the changes in the Rank-1 accuracy in unlearned persons after De-ReID fine-tuning using different methods. In the following, we only discuss the performance of " + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "text", + "content": " and therefore " + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{R - 1}_O" + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "text", + "content": " means " + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{R - 1}_O(f)" + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 337, + 447, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 447, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 447, + 349 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "content": "By default, we use ViT-B [13] as the ReID backbone " + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot), f(\\cdot)" + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "content": ". The images are resized to " + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "inline_equation", + "content": "256 \\times 128" + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "content": ", and the data augmentation includes random crop and flip. For introducing variations in variation-guided identity shift, we use stronger augmentation derived from RandAugment [12]. We detailed the augmentation and conducted corresponding ablation studies in the Appendix. The AdamW [32] optimizer is adopted for De-ReID learning with an initial learning of 3e-4. The weight decay is set to 0. We empirically set trade-off parameters " + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 1.0" + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 1.0" + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "content": ". The batch size for accessible persons is 48, and that for unlearned persons is 32. The above hyper-parameters are the same in all experiments. " + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "content": " in Eq. 5 is set to 20." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 517, + 500, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 500, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 500, + 529 + ], + "type": "text", + "content": "4.3. Comparison with Related Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 689 + ], + "type": "text", + "content": "- Compared methods. Since this is the first work for De-ReID, we compare our method with existing methods potentially applicable to ReID. We pretrain a vanilla ViT-B in the pretraining subset of each dataset and then apply these methods to the ViT-B in the fine-tuning subset for De-ReID. We report the performance of the machine unlearning methods, including GS-LoRA [58], LIRF [53], SCRUB [24], and Boundary Shrink [8]. Besides, we implement a naive solution: label augmentation based on cameras, which assign different labels for images containing the same person but under different cameras. This naive solution is denoted as \"LabelAug\", and the triplet loss and cross-entropy loss [33] are applied for training LabelAug." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "content": "Notably, previous methods usually assume that the labeled data of accessible classes are available. Since the data" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29335" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 184, + 293, + 334 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 298, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 298, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 298, + 182 + ], + "type": "text", + "content": "Table 2. Comparisons in the MSMT17 dataset. \"BS\" is the Boundary Shrink [8]. \"H\" denotes \"H-Mean\". R-1" + }, + { + "bbox": [ + 55, + 70, + 298, + 182 + ], + "type": "inline_equation", + "content": "_T" + }, + { + "bbox": [ + 55, + 70, + 298, + 182 + ], + "type": "text", + "content": " is the Rank-1 accuracy on the unlearned persons, and R-1" + }, + { + "bbox": [ + 55, + 70, + 298, + 182 + ], + "type": "inline_equation", + "content": "_O" + }, + { + "bbox": [ + 55, + 70, + 298, + 182 + ], + "type": "text", + "content": " is the Rank-1 accuracy on the accessible persons. R-1" + }, + { + "bbox": [ + 55, + 70, + 298, + 182 + ], + "type": "inline_equation", + "content": "_T" + }, + { + "bbox": [ + 55, + 70, + 298, + 182 + ], + "type": "text", + "content": " is expected to be low for unlearning specific persons, while a higher R-1" + }, + { + "bbox": [ + 55, + 70, + 298, + 182 + ], + "type": "inline_equation", + "content": "_O" + }, + { + "bbox": [ + 55, + 70, + 298, + 182 + ], + "type": "text", + "content": " and H is preferred. Since we apply additional augmentation for asymmetric contrastive learning, we further apply our augmentation to LIRF, denoted as LIRF*. \"RR\" is the Re-Ranking [63]. The method with the best/second-best H-Mean is marked in red/ blue. Blue background means the H-Mean are improved by Re-Ranking." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 184, + 293, + 334 + ], + "lines": [ + { + "bbox": [ + 59, + 184, + 293, + 334 + ], + "spans": [ + { + "bbox": [ + 59, + 184, + 293, + 334 + ], + "type": "table", + "html": "
MethodMT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug54.673.838.156.873.532.4
+RR62.376.929.265.377.221.2
BS63.878.727.354.674.935.2
+RR70.082.218.362.578.725.3
SCRUB43.875.049.145.274.445.1
+RR52.378.441.353.778.236.6
LIRF10.062.566.218.554.656.8
+RR6.965.469.220.557.357.2
LIRF*18.568.865.135.169.752.8
GS-LoRA56.977.535.960.280.728.6
+RR56.979.736.264.182.923.2
Ours4.677.076.310.872.969.7
+RR4.679.077.310.475.170.9
", + "image_path": "38a46ca88d855a74b2221e5a048f8196d554aaa27bb07c49524e610cf34fff33.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 370, + 293, + 521 + ], + "blocks": [ + { + "bbox": [ + 55, + 347, + 295, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 347, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 295, + 369 + ], + "type": "text", + "content": "Table 3. Comparison results in the MSMT17 dataset when " + }, + { + "bbox": [ + 55, + 347, + 295, + 369 + ], + "type": "inline_equation", + "content": "M_T = 75" + }, + { + "bbox": [ + 55, + 347, + 295, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 347, + 295, + 369 + ], + "type": "inline_equation", + "content": "M_T = 100" + }, + { + "bbox": [ + 55, + 347, + 295, + 369 + ], + "type": "text", + "content": ". The notations are consistent with Table 2" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 370, + 293, + 521 + ], + "lines": [ + { + "bbox": [ + 59, + 370, + 293, + 521 + ], + "spans": [ + { + "bbox": [ + 59, + 370, + 293, + 521 + ], + "type": "table", + "html": "
MethodMT=75MT=100
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug61.472.229.262.771.927.2
+ RR69.175.818.666.875.721.8
BS57.269.133.954.868.336.3
+ RR65.674.323.761.472.728.9
SCRUB54.774.837.550.567.840.6
+ RR62.979.027.755.772.835.9
LIRF18.651.956.120.849.753.8
+ RR19.655.157.524.852.453.5
LIRF*38.465.050.515.650.956.7
GS-LoRA62.479.528.462.176.128.3
+ RR65.381.824.563.878.626.2
Ours12.469.868.513.167.166.7
+ RR13.171.669.015.368.866.4
", + "image_path": "ca81e20311d89bd7e516bce90d370f70b10123838f858a38758d1e995bcbbdac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 544, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 544, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 54, + 544, + 295, + 592 + ], + "type": "text", + "content": "of accessible persons is unlabeled in our experiments, we apply our " + }, + { + "bbox": [ + 54, + 544, + 295, + 592 + ], + "type": "inline_equation", + "content": "L_{RCR}" + }, + { + "bbox": [ + 54, + 544, + 295, + 592 + ], + "type": "text", + "content": " as a regularization for previous methods. We further apply our methods to advanced ReID models, including DCFormaler [27], PAT [28], and PHA [56]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": "- Results. Table 2 and Table 4 shows the comparison results. From the tables, we can observe that our method achieves the best performance in terms of H-Mean. Since the post-processing methods like Re-Ranking [63] can improve the results, we investigate whether these methods can reduce the unlearning effect when retrieving unlearned persons. The results show that existing unlearning methods usually get worse results after post-processing. For example, in Table 2, the Rank-1 accuracy of the SCRUB on unlearned persons is clearly increased from " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "43.8\\%" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "52.3\\%" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "M_T = 25" + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 83, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 331, + 71, + 536, + 81 + ], + "lines": [ + { + "bbox": [ + 331, + 71, + 536, + 81 + ], + "spans": [ + { + "bbox": [ + 331, + 71, + 536, + 81 + ], + "type": "text", + "content": "Table 4. Comparison results in the Market-1501 dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 83, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 317, + 83, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 317, + 83, + 553, + 235 + ], + "type": "table", + "html": "
MethodMT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
LabelAug50.776.551.364.777.442.5
+RR49.377.652.866.078.941.3
BS57.373.144.562.072.244.3
+RR54.775.347.465.374.041.4
SCRUB50.776.051.254.770.250.4
+RR56.077.446.664.771.841.6
LIRF25.384.872.931.375.968.7
+RR26.784.471.932.775.767.7
LIRF*37.386.064.539.379.064.6
GS-LoRA56.091.748.965.391.443.7
+RR54.791.350.267.391.341.3
Ours10.791.184.412.084.483.2
+RR5.391.087.410.083.583.8
", + "image_path": "e8d40e81a451f6a18177ea8e450de1154c318ff9386acafb18c5607d2a0c4d6f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 256, + 553, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 553, + 280 + ], + "type": "text", + "content": "demonstrating that the model does not effectively unlearn the persons and there are potential cues to re-identify them." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 281, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 340 + ], + "type": "text", + "content": "While other methods achieve a worse H-mean after Re-ranking, our method can benefit from Re-Ranking [63] since the overall performance H-Mean is improved in most cases. For example, in the case of " + }, + { + "bbox": [ + 313, + 281, + 555, + 340 + ], + "type": "inline_equation", + "content": "M_T = 25" + }, + { + "bbox": [ + 313, + 281, + 555, + 340 + ], + "type": "text", + "content": " in Market dataset, the H-Means of our method is increased to " + }, + { + "bbox": [ + 313, + 281, + 555, + 340 + ], + "type": "inline_equation", + "content": "87.4\\%" + }, + { + "bbox": [ + 313, + 281, + 555, + 340 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 281, + 555, + 340 + ], + "type": "inline_equation", + "content": "84.4\\%" + }, + { + "bbox": [ + 313, + 281, + 555, + 340 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 437 + ], + "type": "text", + "content": "Compared with the other methods, Boundary Shrink changes the classification decision boundary of the unlearning classes, i.e., the unlearned persons in this work. However, the feature representation is the core for person image matching instead of the classifier. In contrast, our method directly optimizes the feature representation for De-ReID. As a result, our method outperforms Boundary Shrink by " + }, + { + "bbox": [ + 313, + 342, + 556, + 437 + ], + "type": "inline_equation", + "content": "30.4\\%" + }, + { + "bbox": [ + 313, + 342, + 556, + 437 + ], + "type": "text", + "content": " H-Mean in MSMT17 with " + }, + { + "bbox": [ + 313, + 342, + 556, + 437 + ], + "type": "inline_equation", + "content": "M_T = 100" + }, + { + "bbox": [ + 313, + 342, + 556, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "text", + "content": "LIRF unlearns classes by constraining the attention map of intermediate features to be significantly different before and after unlearning. Since it directly constrains the features, it obtains the second-best results. However, the attention map has lots of knowledge about extracting identity information from images. Hence, their constraints largely hurt the ReID ability on accessible persons, and our method outperforms LIRF by " + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "inline_equation", + "content": "12.9\\%" + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "text", + "content": " at H-Mean in MSMT17 when " + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "inline_equation", + "content": "M_T = 100" + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "text", + "content": ". Besides, since we apply additional augmentation in variation-guided identity shift, we further validate the effectiveness of the additional augmentation by applying the augmentation in LIRF, denoted as " + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{LIRF^{*}}" + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "text", + "content": ". However, the H-Mean is improved only in the case of " + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "inline_equation", + "content": "M_T = 100" + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "text", + "content": " in MSMT17 and decreases in other cases, demonstrating that additional augmentation does not necessarily lead to higher performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "GS-LoRA effectively protects the ReID ability on accessible persons but falls behind in unlearning specific persons. LabelAug directly guides the model in distinguishing the images of unlearned persons under different camera views, but it has an unsatisfactory performance. We speculate it is because the model overfits the images in training and fails to unlearn the specific persons under cameras unseen in training. Differently, our method employs variation-guided" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29336" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 127, + 282, + 294 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": "Table 5. Evaluation with different ReID models. Regardless of the ReID models, our method performs the best in terms of erasing the knowledge about unlearned persons (quantified by " + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "inline_equation", + "content": "\\mathrm{R - 1}_T" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": ") and keeping ReID knowledge for accessible persons (" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "inline_equation", + "content": "\\mathrm{R - 1}_O" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": "). \"Null\" indicates that no De-ReID method is applied." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 127, + 282, + 294 + ], + "lines": [ + { + "bbox": [ + 59, + 127, + 282, + 294 + ], + "spans": [ + { + "bbox": [ + 59, + 127, + 282, + 294 + ], + "type": "table", + "html": "
ReID modelDe-ReID methodR-1T↓R-1O↑H↑
DCFormerNull81.284.4—
LabelAug44.154.046.2
GS-LORA75.381.916.4
LIRF18.051.057.7
Ours17.866.166.3
PATNull72.875.6—
LabelAug38.343.340.1
GS-LORA52.560.933.5
LIRF16.742.649.4
Ours12.055.059.0
PHANull70.278.7—
LabelAug44.351.441.2
GS-LORA56.468.333.6
LIRF23.650.552.7
Ours3.951.260.8
", + "image_path": "04b16c205a7d832acadd6a8880e67caa05d5902b9ae8d1456f996d4747ef3f80.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 304, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 296, + 399 + ], + "type": "text", + "content": "identity shift " + }, + { + "bbox": [ + 54, + 304, + 296, + 399 + ], + "type": "inline_equation", + "content": "(L_{VIS}^{p}, L_{VIS}^{o})" + }, + { + "bbox": [ + 54, + 304, + 296, + 399 + ], + "type": "text", + "content": " and guides the model to handle the variations adaptively. Fitting the variations in unlearned images helps the model bias to the environmental information when querying the unlearned persons. Simultaneously, the model is guided to be robust to the variations when querying accessible persons. As a result, the unlearned persons are shifted out of the view-invariant feature space while the accessible persons are retained." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 400, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 400, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 296, + 472 + ], + "type": "text", + "content": "- Evaluations on other ReID methods. Apart from the comparison on the vanilla ViT backbone, we also compared our method with related methods on the advanced ReID models [27, 28, 56]. The experimental results in Table 5 demonstrate that our method is superior to others regarding unlearning different ReID models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 149, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 149, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 149, + 493 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 498, + 297, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 297, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 297, + 581 + ], + "type": "text", + "content": "We conduct experiments to show the effectiveness of each component in our design. Due to the space limit, we report the experimental results in Market and " + }, + { + "bbox": [ + 54, + 498, + 297, + 581 + ], + "type": "inline_equation", + "content": "M_T \\in \\{25, 100\\}" + }, + { + "bbox": [ + 54, + 498, + 297, + 581 + ], + "type": "text", + "content": " in MSMT17 in Table 6, and the rest in Appendix. Regardless of " + }, + { + "bbox": [ + 54, + 498, + 297, + 581 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 54, + 498, + 297, + 581 + ], + "type": "text", + "content": ", our designs clearly improve H-Mean that comprehensively evaluates the effectiveness of unlearning specific persons and ReID performance on accessible persons." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 582, + 297, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 297, + 701 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 297, + 701 + ], + "type": "text", + "content": "- The effectiveness of the relation regularization. From Table 6, we first observe that the model suffers from catastrophic forgetting in ReID knowledge and collapses in the absence of " + }, + { + "bbox": [ + 54, + 582, + 297, + 701 + ], + "type": "inline_equation", + "content": "L_{RCR}" + }, + { + "bbox": [ + 54, + 582, + 297, + 701 + ], + "type": "text", + "content": ". Although our variation-guided identity shift pulls closer the images and the augmented images for accessible persons, it mainly preserves the view-invariant feature embedding but it fails to preserve the feature discriminability, leading to model collapses. " + }, + { + "bbox": [ + 54, + 582, + 297, + 701 + ], + "type": "inline_equation", + "content": "L_{RCR}" + }, + { + "bbox": [ + 54, + 582, + 297, + 701 + ], + "type": "text", + "content": " preserves the discriminability of features by preserving the relations between accessible persons." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "text", + "content": "Another constraint " + }, + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{TRC}" + }, + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "text", + "content": " attempts to push away the im" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 127, + 553, + 249 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 126 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 126 + ], + "type": "text", + "content": "Table 6. Ablation study in MSMT17 dataset. \"W/o\" means \"without\", and other notations are the same as Table 2. " + }, + { + "bbox": [ + 313, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 313, + 70, + 555, + 126 + ], + "type": "text", + "content": " is the initial model for De-ReID learning. The full model achieves the best H-Mean. Notably, without " + }, + { + "bbox": [ + 313, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "L_{RCR}" + }, + { + "bbox": [ + 313, + 70, + 555, + 126 + ], + "type": "text", + "content": ", the model forgets ReID knowledge and collapses. \"SD\" denotes the \"self-augmented discrimination\"." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 127, + 553, + 249 + ], + "lines": [ + { + "bbox": [ + 317, + 127, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 127, + 553, + 249 + ], + "type": "table", + "html": "
MethodMSMT17
MT=25MT=100
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
fp(·)80.385.4—79.585.4—
Components in Relation regularization
W/o LTRC5.573.474.118.365.363.2
W/o LRCR——————
Components in Variation-guided Identity Shift
W/o SD4.674.575.110.558.763.4
W/o LVIS10.075.872.918.063.262.3
W/o LVIS20.577.467.541.370.049.4
Full model4.677.076.313.167.166.7
", + "image_path": "cd035e41188ebc4cd84a9291da29e332beb50061846b9431c2fd50bb51047064.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 271, + 552, + 394 + ], + "blocks": [ + { + "bbox": [ + 360, + 260, + 508, + 270 + ], + "lines": [ + { + "bbox": [ + 360, + 260, + 508, + 270 + ], + "spans": [ + { + "bbox": [ + 360, + 260, + 508, + 270 + ], + "type": "text", + "content": "Table 7. Ablation study in Market-1501." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 271, + 552, + 394 + ], + "lines": [ + { + "bbox": [ + 317, + 271, + 552, + 394 + ], + "spans": [ + { + "bbox": [ + 317, + 271, + 552, + 394 + ], + "type": "table", + "html": "
MethodMarket-1501
MT=25MT=50
R-1T↓R-1O↑H↑R-1T↓R-1O↑H↑
fp(·)89.396.3—94.096.3—
Components in Relation regularization
W/o LTRC20.089.778.224.078.373.9
W/o LRCR——————
Components in Variation-guided Identity Shift
W/o SD24.090.375.811.379.781.2
W/o LVIS10.784.581.412.777.379.2
W/o LVIS46.793.958.655.389.954.1
Full model10.791.184.412.084.483.2
", + "image_path": "29ec98d02d41136121b9c06c21930e9552b97e14eaef565a79d5b7a95ce691a2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 319, + 404, + 430, + 487 + ], + "blocks": [ + { + "bbox": [ + 319, + 404, + 430, + 487 + ], + "lines": [ + { + "bbox": [ + 319, + 404, + 430, + 487 + ], + "spans": [ + { + "bbox": [ + 319, + 404, + 430, + 487 + ], + "type": "image", + "image_path": "551864891eb9ef9188b260dba7e6011342b35d93a9366987e5ab409eb8f282aa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 487, + 531, + 497 + ], + "lines": [ + { + "bbox": [ + 336, + 487, + 531, + 497 + ], + "spans": [ + { + "bbox": [ + 336, + 487, + 531, + 497 + ], + "type": "text", + "content": "Figure 3. Evaluation results under various " + }, + { + "bbox": [ + 336, + 487, + 531, + 497 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 336, + 487, + 531, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 336, + 487, + 531, + 497 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 336, + 487, + 531, + 497 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 437, + 403, + 547, + 487 + ], + "blocks": [ + { + "bbox": [ + 437, + 403, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 437, + 403, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 437, + 403, + 547, + 487 + ], + "type": "image", + "image_path": "e08e91c4565349eaf9d804561e1689c5da0f9716d202fa83d9f13f6c858ed308.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": "ages between unlearned persons to prevent them from being cross-view retrieved, referring to the distance between accessible persons and unlearned persons. We observe a degraded H-Mean in the absence of " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "L_{TRC}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " in all cases. For example, when deprecating " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "L_{TRC}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " in the case of " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "M_T = 50" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " in Market, the Rank-1 accuracy of the unlearned persons is increased to " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "24.0\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "12.0\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": ", meaning that more unlearned person images are correctly matched and is not expected. As a result, the H-Mean drops by " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " compared with the full model. In the case of " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "M_T = 100" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " in MSMT17, the Rank-1 accuracy of the unlearned persons increases to " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "18.3\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " without " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "L_{TRC}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": ", and the H-Mean drops by " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "The hyper-parameter " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " controls the weight of " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "L_{TRC}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": ". We evaluate our model under different " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " and show the performance in Fig. 3. When " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "L_{TRC}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " is not employed and the performance is inferior. As " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " increases, the performance is improved. However, " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "L_{TRC}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " should not be set to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29337" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 292, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 292, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 292, + 85 + ], + "type": "text", + "content": "too large to avoid overwhelming other learning objectives." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "text", + "content": "- The effectiveness of the variation-guided identity shift. The variation-guided identity shift (VIS) emphasizes the style difference between images of unlearned persons in the feature space while keeping the other images robust to the style variations. Overall, comparing the results of \"W/o " + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "inline_equation", + "content": "L_{VIS}" + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "text", + "content": "\" and the full model in Table 6, we observe that the ReID model struggles to unlock the specific persons without " + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "inline_equation", + "content": "L_{VIS}" + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "text", + "content": ". For example, in MSMT17 dataset, the Rank-1 accuracy on unlearned persons is " + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "inline_equation", + "content": "20.5\\% / 41.3\\%" + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "inline_equation", + "content": "M_T = 25 / 100" + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "inline_equation", + "content": "15.9\\% / 28.2\\%" + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "text", + "content": " larger than the full model. This is because ReID requires view-invariant features and our model correspondingly guides the model to be biased to encode environmental information for images of unlearned persons. As a result, we effectively prevent the images of unlearned persons from being re-identified. The hyperparameter " + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "text", + "content": " controls the weight of " + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "inline_equation", + "content": "L_{VIS}" + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "text", + "content": " in optimization. We evaluate our model under different " + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 54, + 86, + 296, + 300 + ], + "type": "text", + "content": " and show the performance in Fig. 3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 302, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 296, + 470 + ], + "type": "text", + "content": "We also studied the effectiveness of the designs in VIS. As discussed in the introduction, we argue that simply pushing away the images under different camera views may learn the spurious cues for De-ReID, and therefore we push away the image and the corresponding augmented image in our VIS. We refer to this as \"self-augmented discrimination\". To verify its effectiveness, we conduct an experiment that pushes away the image and the augmented image from another image, denoted as \"W/o SD\", and observe a significant performance drop. For example, in the case of " + }, + { + "bbox": [ + 55, + 302, + 296, + 470 + ], + "type": "inline_equation", + "content": "M_T = 25" + }, + { + "bbox": [ + 55, + 302, + 296, + 470 + ], + "type": "text", + "content": " in Market dataset, the Rank-1 accuracy on the unlearned person increases to " + }, + { + "bbox": [ + 55, + 302, + 296, + 470 + ], + "type": "inline_equation", + "content": "24.0\\%" + }, + { + "bbox": [ + 55, + 302, + 296, + 470 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 302, + 296, + 470 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 55, + 302, + 296, + 470 + ], + "type": "text", + "content": " without SD, and the H-Mean drops by " + }, + { + "bbox": [ + 55, + 302, + 296, + 470 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 55, + 302, + 296, + 470 + ], + "type": "text", + "content": ". In other cases, the H-Mean drops without using the SD, demonstrating the effectiveness of SD." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "content": "Moreover, we observe that guiding the ReID to overcome the variations for accessible persons is important for balancing the ReID ability and De-ReID of unlearned persons. Deprecating the " + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "inline_equation", + "content": "L_{VIS}^{o}" + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "content": " leads to " + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "inline_equation", + "content": "3\\% \\sim 4\\%" + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "content": " performance drops in H-Mean in different cases, and reduce the protection effect since the rank-1 accuracy of unlearned persons increases." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 544, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 295, + 604 + ], + "type": "text", + "content": "- The impact of the unlabeled sample. We evaluate our method in MSMT17 under different ratios between unlabeled images of accessible persons and labeled images of unlearned persons in the fine-tuning subset. The results are in Table 8." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "From the results, we obtain two conclusions. (i) More unlabeled data is helpful for both unlearning specific persons and keeping ReID performance on the accessible persons. When the number of accessible persons decreases, the Rank-1 accuracy of the unlearned persons (denoted as " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{R - 1}_T" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ") increases, which means the unlearned persons are easier to retrieve by the ReID model. Simultaneously, the Rank-1 accuracy of the accessible persons " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{R - 1}_O)" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " decreases, which means the ReID model cannot properly re-identify" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 71, + 555, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 137 + ], + "type": "text", + "content": "Table 8. Evaluations under different numbers of accessible persons in MSMT17 when " + }, + { + "bbox": [ + 313, + 71, + 555, + 137 + ], + "type": "inline_equation", + "content": "M_T = 100" + }, + { + "bbox": [ + 313, + 71, + 555, + 137 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 71, + 555, + 137 + ], + "type": "inline_equation", + "content": "M_O" + }, + { + "bbox": [ + 313, + 71, + 555, + 137 + ], + "type": "text", + "content": " is the number of accessible persons in the fine-tuning subset. 'With LUP' means using the images in LUPerson dataset as auxiliary unlabeled data and the number in parentheses denotes the performance gain. By default, we do not use LUPerson dataset." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 138, + 552, + 198 + ], + "blocks": [ + { + "bbox": [ + 317, + 138, + 552, + 198 + ], + "lines": [ + { + "bbox": [ + 317, + 138, + 552, + 198 + ], + "spans": [ + { + "bbox": [ + 317, + 138, + 552, + 198 + ], + "type": "table", + "html": "
\\(M_O\\)Without LUPWith LUP
R-1T ↓R-1O ↑H ↑R-1T ↓R-1O ↑H ↑
14113.167.166.7---
10015.861.362.514.168.266.8 (+4.3%)
8015.854.358.615.467.765.9 (+7.3%)
6017.451.457.615.465.664.8 (+7.2%)
", + "image_path": "c3518ce90fde570a59130552ede8203843210cd068cb4d5c9e9cd1c8ae677b4d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 210, + 349, + 275 + ], + "blocks": [ + { + "bbox": [ + 315, + 210, + 349, + 275 + ], + "lines": [ + { + "bbox": [ + 315, + 210, + 349, + 275 + ], + "spans": [ + { + "bbox": [ + 315, + 210, + 349, + 275 + ], + "type": "image", + "image_path": "4977e97c7dbfa613471b79774d592254c08161398b9f7df3bf13ebfe2706ea62.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 275, + 430, + 285 + ], + "lines": [ + { + "bbox": [ + 322, + 275, + 430, + 285 + ], + "spans": [ + { + "bbox": [ + 322, + 275, + 430, + 285 + ], + "type": "text", + "content": "Images of unlearned persons" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 288, + 555, + 310 + ], + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 310 + ], + "type": "text", + "content": "Figure 4. Attention map on images of the unlearned persons and the images of accessible persons." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 356, + 210, + 388, + 275 + ], + "blocks": [ + { + "bbox": [ + 356, + 210, + 388, + 275 + ], + "lines": [ + { + "bbox": [ + 356, + 210, + 388, + 275 + ], + "spans": [ + { + "bbox": [ + 356, + 210, + 388, + 275 + ], + "type": "image", + "image_path": "e6adc81d0df95302d4d74e97fe7c9b2f8002987382ff8cf6e4b05fb59428fbb3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 356, + 275, + 388, + 284 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 275, + 388, + 284 + ], + "spans": [ + { + "bbox": [ + 356, + 275, + 388, + 284 + ], + "type": "text", + "content": "of unlearn" + } + ] + } + ], + "index": 11, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 395, + 210, + 428, + 275 + ], + "blocks": [ + { + "bbox": [ + 395, + 210, + 428, + 275 + ], + "lines": [ + { + "bbox": [ + 395, + 210, + 428, + 275 + ], + "spans": [ + { + "bbox": [ + 395, + 210, + 428, + 275 + ], + "type": "image", + "image_path": "1a2ae0c4b8e43b6d2d5290d6024c57328cc196e4fe257c82da6c0cbe9beeed59.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 395, + 275, + 428, + 284 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 275, + 428, + 284 + ], + "spans": [ + { + "bbox": [ + 395, + 275, + 428, + 284 + ], + "type": "text", + "content": "d persons" + } + ] + } + ], + "index": 13, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 441, + 210, + 473, + 275 + ], + "blocks": [ + { + "bbox": [ + 441, + 210, + 473, + 275 + ], + "lines": [ + { + "bbox": [ + 441, + 210, + 473, + 275 + ], + "spans": [ + { + "bbox": [ + 441, + 210, + 473, + 275 + ], + "type": "image", + "image_path": "455732697bc2e7e6a864634da6bd8bcaf18ead8c9d8a3e86953cb41141430d0f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 275, + 541, + 285 + ], + "lines": [ + { + "bbox": [ + 451, + 275, + 541, + 285 + ], + "spans": [ + { + "bbox": [ + 451, + 275, + 541, + 285 + ], + "type": "text", + "content": "Images of other persons" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 495, + 210, + 529, + 275 + ], + "blocks": [ + { + "bbox": [ + 495, + 210, + 529, + 275 + ], + "lines": [ + { + "bbox": [ + 495, + 210, + 529, + 275 + ], + "spans": [ + { + "bbox": [ + 495, + 210, + 529, + 275 + ], + "type": "image", + "image_path": "610f1c468e76285bcf022bb8c83b364dabb4dab798967780278a4e75d7097c33.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 522, + 275, + 541, + 284 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 522, + 275, + 541, + 284 + ], + "spans": [ + { + "bbox": [ + 522, + 275, + 541, + 284 + ], + "type": "text", + "content": "rsons" + } + ] + } + ], + "index": 17, + "type": "text" + }, + { + "bbox": [ + 313, + 317, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 413 + ], + "type": "text", + "content": "the accessible persons. (ii) Although the performance of our method decreases when the unlabeled images become fewer, our method can effectively utilize the publicly available unlabeled person images for unlearning, demonstrating the scalability of our method. When using additional unlabeled images from LUPerson dataset, our method can achieve similar performance even though the number of accessible persons in MSMT17 decreases from 141 to 60." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 422, + 398, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 398, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 398, + 434 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 441, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 536 + ], + "type": "text", + "content": "We visualize the attention map on different persons in the testing set in Fig. 4. From the figure, for the images of unlearned persons, we can observe that the model's attention mainly focus on the background which contains abundant environmental information. At the same time, the model pays diverse attention to the informative parts of images for extracting discriminative features for accessible persons. These results demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "To address the privacy concerns in person ReID, we pioneer the exploration of the De-ReID problem that guides the ReID model to forget the knowledge about matching specific persons. We formulate a framework and propose a novel variation-guided identity shift method for De-ReID. Our method unlearns the specific person shifting them out of the view-invariant feature space and simultaneously keeps the ReID knowledge by overcoming the cross-view variations for accessible persons. We further introduce relation regularization to characterize the desired properties of De-ReID. Extensive experiments conducted on a new benchmark show the superiority of our method." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29338" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "content": "This work was supported partially by NSFC(U21A20471, 92470202), by the National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "type": "text", + "content": "[1] Prachi Agrawal and P. J. Narayanan. Person de-identification in videos. IEEE Transactions on Circuits and Systems for Video Technology, 2011. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 203, + 296, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 296, + 246 + ], + "type": "text", + "content": "[2] Shafiq Ahmad, Pietro Morerio, and Alessio Del Bue. Person re-identification without identification via event anonymization. In 2023 IEEE/CVF International Conference on Computer Vision, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 296, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 291 + ], + "type": "text", + "content": "[3] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In IEEE Symposium on Security and Privacy, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 292, + 296, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 296, + 345 + ], + "type": "text", + "content": "[4] Karla Brkic, Ivan Sikiric, Tomislav Hrkac, and Zoran Kalafatic. I know that person: Generative full body and face de-identification of people in images. In 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 296, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 296, + 379 + ], + "type": "text", + "content": "[5] Jonathan Brophy and Daniel Lowd. Machine unlearning for random forests. In International Conference on Machine Learning, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "type": "text", + "content": "[6] Chengzhi Cao, Xueyang Fu, Hongjian Liu, Yukun Huang, Kunyu Wang, Jiebo Luo, and Zheng-Jun Zha. Event-guided person re-identification via sparse-dense complementary learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 296, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 490 + ], + "type": "text", + "content": "[7] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Learning invariance from generated variance for unsupervised person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 491, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 545 + ], + "type": "text", + "content": "[8] Min Chen, Weizhuo Gao, Gaoyang Liu, Kai Peng, and Chen Wang. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision boundary. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 547, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 296, + 590 + ], + "type": "text", + "content": "[9] Tianshui Chen, Tao Pu, Lingbo Liu, Yukai Shi, Zhijing Yang, and Liang Lin. Heterogeneous semantic transfer for multi-label recognition with partial labels. International Journal of Computer Vision, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 296, + 645 + ], + "type": "text", + "content": "[10] Weihua Chen, Xiaotang Chen, Jianguo Zhang, and Kaiqi Huang. Beyond triplet loss: A deep quadruplet network for person re-identification. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, 2017. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 690 + ], + "type": "text", + "content": "[11] Dasol Choi and Dongbin Na. Towards machine unlearning benchmarks: Forgetting the personal identities in facial recognition systems. In AAAI 2024 workshop on Privacy-Preserving Artificial Intelligence, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[12] Ekin D. Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le. Randaugment: Practical automated data augmentation" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 105 + ], + "type": "text", + "content": "with a reduced search space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 184 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "type": "text", + "content": "[14] Lijie Fan, Tianhong Li, Rongyao Fang, Rumen Hristov, Yuan Yuan, and Dina Katabi. Learning longterm representations for person re-identification using radio signals. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 555, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 297 + ], + "type": "text", + "content": "[15] Ivan Filković, Zoran Kalafatić, and Tomislav Hrkač. Deep metric learning for person re-identification and de-identification. In 2016 39th International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO), 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 342 + ], + "type": "text", + "content": "[16] Aditya Golatkar, Alessandro Achille, and Stefano Soatto. Eternal sunshine of the spotless net: Selective forgetting in deep networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 555, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 387 + ], + "type": "text", + "content": "[17] Wenxuan Guo, Zhiyu Pan, Yingping Liang, Ziheng Xi, Zhi Chen Zhong, Jianjiang Feng, and Jie Zhou. Lidar-based person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 555, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 431 + ], + "type": "text", + "content": "[18] Varun Gupta, Christopher Jung, Seth Neel, Aaron Roth, Saeed Sharifi-Malvajerdi, and Chris Waites. Adaptive machine unlearning. Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "type": "text", + "content": "[19] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In IEEE/CVF International Conference on Computer Vision, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 479, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 555, + 544 + ], + "type": "text", + "content": "[20] Weizhen He, Yiheng Deng, Shixiang Tang, Qihao Chen, Qing-song Xie, Yizhou Wnag, Lei Bai, Feng Zhu, Rui Zhao, Wanli Ouyang, Donglian Qi, and Yunfeng Yan. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 546, + 555, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 555, + 601 + ], + "type": "text", + "content": "[21] Bingliang Jiao, Lingqiao Liu, Liying Gao, Guosheng Lin, Lu Yang, Shizhou Zhang, Peng Wang, and Yanning Zhang. Dynamically transformed instance normalization network for generalizable person re-identification. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 602, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 555, + 645 + ], + "type": "text", + "content": "[22] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Andy J Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multiscale dilated transformer for visual recognition. IEEE Transactions on Multimedia, 25:8906-8919, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "type": "text", + "content": "[23] Xin Jin, Tianyu He, Kecheng Zheng, Zhiheng Yin, Xu Shen, Zhen Huang, Ruoyu Feng, Jianqiang Huang, Zhibo Chen, and Xian-Sheng Hua. Cloth-changing person re-identification from a single image with gait prediction and regularization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29339" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[24] Meghdad Kurmanji, Peter Triantafillou, Jamie Hayes, and Eleni Triantafillou. Towards unbounded machine unlearning. In Conference on Neural Information Processing Systems, 2023. 1, 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[25] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: Exploiting vision-language model for image re-identification without concrete text labels. Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "text", + "content": "[26] Weijia Li, Saihui Hou, Chunjie Zhang, Chunshui Cao, Xu Liu, Yongzhen Huang, and Yao Zhao. An in-depth exploration of person re-identification and gait recognition in cloth-changing conditions. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "text", + "content": "[27] Wen Li, Cheng Zou, Meng Wang, Furong Xu, Jianan Zhao, Ruobing Zheng, Yuan Cheng, and Wei Chu. Dc-former: Diverse and compact transformer for person re-identification. In AAAI Conference on Artificial Intelligence, 2023. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "text", + "content": "[28] Yulin Li, Jianfeng He, Tianzhu Zhang, Xiang Liu, Yongdong Zhang, and Feng Wu. Diverse part discovery: Occluded person re-identification with part-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 364 + ], + "type": "text", + "content": "[29] Kun-Yu Lin, Jiaming Zhou, Yukun Qiu, and Wei-Shi Zheng. Adversarial partial domain adaptation by cycle inconsistency. In European Conference on Computer Vision, pages 530-548. Springer Nature Switzerland Cham, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 409 + ], + "type": "text", + "content": "[30] Kun-Yu Lin, Jia-Run Du, Yipeng Gao, Jiaming Zhou, and Wei-Shi Zheng. Diversifying spatial-temporal perception for video domain generalization. Advances in Neural Information Processing Systems, 36:56012–56026, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 464 + ], + "type": "text", + "content": "[31] Kun-Yu Lin, Henghui Ding, Jiaming Zhou, Yu-Ming Tang, Yi-Xing Peng, Zhilin Zhao, Chen Change Loy, and Wei-Shi Zheng. Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "type": "text", + "content": "[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 554 + ], + "type": "text", + "content": "[33] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 296, + 601 + ], + "type": "text", + "content": "[34] Jiaxu Miao, Yu Wu, Ping Liu, Yuhang Ding, and Yi Yang. Pose-guided feature alignment for occluded person re-identification. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 542–551, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 656 + ], + "type": "text", + "content": "[35] Tao Pu, Tianshui Chen, Hefeng Wu, Yukai Shi, Zhijing Yang, and Liang Lin. Dual-perspective semantic-aware representation blending for multi-label image recognition with partial labels. Expert Systems with Applications, page 123526, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "text", + "content": "[36] Tao Pu, Qianru Lao, Hefeng Wu, Tianshui Chen, Ling Tian, Jie Liu, and Liang Lin. Category-adaptive label discovery and noise rejection for multi-label recognition with partial positive labels. IEEE Transactions on Multimedia, pages 1-12, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "type": "text", + "content": "[37] Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember what you want to forget: Algorithms for machine unlearning. Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 151 + ], + "type": "text", + "content": "[38] Takashi Shibata, Go Irie, Daiki Ikami, and Yu Mitsuzumi. Learning with selective forgetting. In International Joint Conference on Artificial Intelligence, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 196 + ], + "type": "text", + "content": "[39] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and A strong convolutional baseline). In European Conference on Computer Vision, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 555, + 251 + ], + "type": "text", + "content": "[40] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 555, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 555, + 307 + ], + "type": "text", + "content": "[41] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. Learning to imagine: Diversify memory for incremental learning using unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9549-9558, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 555, + 353 + ], + "type": "text", + "content": "[42] Yu-Ming Tang, Yi-Xing Peng, and Wei-Shi Zheng. When prompt-based incremental learning does not meet strong pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1706-1716, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 398 + ], + "type": "text", + "content": "[43] Yu-Ming Tang, Yi-Xing Peng, Jingke Meng, and Wei-Shi Zheng. Rethinking few-shot class-incremental learning: Learning from yourself. In European Conference on Computer Vision, pages 108-128. Springer, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "type": "text", + "content": "[44] Guanshuo Wang, Yufeng Yuan, Xiong Chen, Jiwei Li, and Xi Zhou. Learning discriminative features with multiple granularities for person re-identification. In ACM International Conference on Multimedia, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 487 + ], + "type": "text", + "content": "[45] Haochen Wang, Jiayi Shen, Yongtuo Liu, Yan Gao, and Efstratos Gavves. Nformer: Robust person re-identification with neighbor transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 490, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 555, + 533 + ], + "type": "text", + "content": "[46] Lin Wang, Wanqian Zhang, Dayan Wu, Fei Zhu, and Bo Li. Attack is the best defense: Towards preemptive-protection person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 535, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 555, + 578 + ], + "type": "text", + "content": "[47] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 79–88, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 580, + 555, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 621 + ], + "type": "text", + "content": "[48] Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. A discriminative feature learning approach for deep face recognition. In European Conference on Computer Vision, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "type": "text", + "content": "[49] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[50] Bin Yang, Jun Chen, and Mang Ye. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In IEEE/CVF International Conference on Computer Vision, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29340" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 688 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "text", + "content": "[51] Bin Yang, Jun Chen, and Mang Ye. Shallow-deep collaborative learning for unsupervised visible-infrared person re-identification. In IEEE/CVF International Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "text", + "content": "[52] Jinrui Yang, Jiawei Zhang, Fufu Yu, Xinyang Jiang, Mengdan Zhang, Xing Sun, Ying-Cong Chen, and Wei-Shi Zheng. Learning to know where to see: A visibility-aware approach for occluded person re-identification. In IEEE/CVF international conference on computer vision, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 295, + 217 + ], + "type": "text", + "content": "[53] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In European Conference on Computer Vision, 2022. 1, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 218, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 295, + 262 + ], + "type": "text", + "content": "[54] Mang Ye, Wei Shen, Junwu Zhang, Yao Yang, and Bo Du. Securereid: Privacy-preserving anonymization for person re-identification. IEEE Transactions on Information Forensics and Security, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 296, + 305 + ], + "type": "text", + "content": "[55] Mang Ye, Zesen Wu, Cuiqun Chen, and Bo Du. Channel augmentation for visible-infrared re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "type": "text", + "content": "[56] Guiwei Zhang, Yongfei Zhang, Tianyu Zhang, Bo Li, and Shiliang Pu. Pha: Patch-wise high-frequency augmentation for transformer-based person re-identification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 407 + ], + "type": "text", + "content": "[57] Junwu Zhang, Mang Ye, and Yao Yang. Learnable privacy-preserving anonymization for pedestrian images. In Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 295, + 452 + ], + "type": "text", + "content": "[58] Hongbo Zhao, Bolin Ni, Haochen Wang, Junsong Fan, Fei Zhu, Yuxi Wang, Yuntao Chen, Gaofeng Meng, and Zhaoxiang Zhang. Continual forgetting for pre-trained vision models. 2024. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 295, + 497 + ], + "type": "text", + "content": "[59] Zhilin Zhao, Longbing Cao, and Kun-Yu Lin. Revealing the distributional vulnerability of discriminators by implicit generators. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8888-8901, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 499, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 295, + 542 + ], + "type": "text", + "content": "[60] Zhilin Zhao, Longbing Cao, and Kun-Yu Lin. Out-of-distribution detection by cross-class vicinity distribution of in-distribution data. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 586 + ], + "type": "text", + "content": "[61] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Computer Vision, IEEE International Conference on, 2015. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 588, + 295, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 295, + 642 + ], + "type": "text", + "content": "[62] Wei-Shi Zheng, Jincheng Hong, Jiening Jiao, Ancong Wu, Xiatian Zhu, Shaogang Gong, Jiayin Qin, and Jianhuang Lai. Joint bilateral-resolution identity modeling for crossresolution person re-identification. International Journal of Computer Vision, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 644, + 295, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 644, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 644, + 295, + 688 + ], + "type": "text", + "content": "[63] Zhun Zhong, Liang Zheng, Donglin Cao, and Shaozi Li. Re-ranking person re-identification with k-reciprocal encoding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29341" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PersonaBooth_ Personalized Text-to-Motion Generation/62946df9-a553-4225-bd84-84b1acdb1a5f_content_list.json b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/62946df9-a553-4225-bd84-84b1acdb1a5f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..274759a615d1847b033dbdf6ae4e82fdef60bdbb --- /dev/null +++ b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/62946df9-a553-4225-bd84-84b1acdb1a5f_content_list.json @@ -0,0 +1,1574 @@ +[ + { + "type": "text", + "text": "PersonaBooth: Personalized Text-to-Motion Generation", + "text_level": 1, + "bbox": [ + 217, + 130, + 779, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boeun Kim1,2,3 Hea In Jeong2 JungHoon Sung3 Yihua Cheng1 Jeongmin Lee2 \nJu Yong Chang4 Sang-II Choi3 Younggeun Choi3 Saim Shin2 Jungho Kim2 Hyung Jin Chang1 \n1University of Birmingham 2Korea Electronics Technology Institute 3Dankook University 4Kwangwoon University \nhttp://boeun-kim.github.io/page-PersonaBooth", + "bbox": [ + 91, + 191, + 903, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 310, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces Motion Personalization, a new task that generates personalized motions aligned with text descriptions using several basic motions containing Persona. To support this novel task, we introduce a new large-scale motion dataset called PerMo (PersonaMotion), which captures the unique personas of multiple actors. We also propose a multi-modal finetuning method of a pretrained motion diffusion model called PersonaBooth. PersonaBooth addresses two main challenges: i) A significant distribution gap between the persona-focused PerMo dataset and the pretraining datasets, which lack persona-specific data, and ii) the difficulty of capturing a consistent persona from the motions vary in content (action type). To tackle the dataset distribution gap, we introduce a persona token to accept new persona features and perform multi-modal adaptation for both text and visuals during finetuning. To capture a consistent persona, we incorporate a contrastive learning technique to enhance intra-cohesion among samples with the same persona. Furthermore, we introduce a context-aware fusion mechanism to maximize the integration of persona cues from multiple input motions. PersonaBooth outperforms state-of-the-art motion style transfer methods, establishing a new benchmark for motion personalization.", + "bbox": [ + 88, + 342, + 485, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 722, + 220, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imagine a scenario where, by recording just a few basic movements of yourself, an avatar in a virtual world can mirror your personal traits and unique motion style. This kind of 'motion personalization' enables realistic interactions in virtual spaces such as games and the metaverse [14]. Moreover, if avatar motions could be directed through text, creating video content would become easy without the need for real-life stunt actors [2, 24].", + "bbox": [ + 89, + 748, + 482, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a task called Motion Personalization, which aims to generate text-based motions that", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0bd3c3c0f3a60c18c3243529198756082868ca673432a91f682450903be322da.jpg", + "image_caption": [ + "Figure 1. Motion Personalization generates text-driven, personalized motions based on persona embedded in atomic input motions. We propose a new framework, PersonaBooth, along with a new benchmark dataset, PerMo, for Motion Personalization" + ], + "image_footnote": [], + "bbox": [ + 535, + 309, + 888, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reflect individual personas using a few key atomic movements, such as jumping, punching, or walking. We define a persona as the unique style expression of an individual. As illustrated in Fig. 1, this task aims to generate a broad range of realistic motions robustly. To facilitate this, we introduce a model called PersonaBooth and a new benchmark, PerMo (PersonaMotion), specifically designed for motion personalization. Table 1 provides a comparison of similar motion generation tasks. Text-to-Motion Generation approaches [6, 21] generate motion using only text as input, without any reference motion, and numerous diffusion-based methods [3, 33, 35] have been proposed for this task. Motion Style Transfer (MST) focuses on transferring style", + "bbox": [ + 511, + 704, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22756", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/2612e3f631052b9f61a518029d8cee03797c1dff46b435d7c9604cb42c56dc2b.jpg", + "table_caption": [ + "Table 1. Comparison of motion generation tasks" + ], + "table_footnote": [], + "table_body": "
Motion TaskExisting DatasetsSource MotionsMulti-modal Finetuning
T2M Generation [3, 33, 35][6] [23]NoneX
Style Transfer [25, 32, 38][34][1][19]SingleX
Personalization (Ours)NoneMultiple✓
", + "bbox": [ + 91, + 103, + 480, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from a single source motion. While several datasets and methods are available for MST [25, 32, 34, 38], the use of a single source motion restricts the ability to generate the broad and diverse motions seen in real life [12]. The proposed Motion Personalization has no existing prior approaches or datasets. Furthermore, our proposed method introduces multi-modal finetuning—a process neglected in existing diffusion-based MST methods.", + "bbox": [ + 89, + 205, + 482, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PersonaBooth is a multi-modal finetuning method for pretrained text-to-motion diffusion models. The challenge in finetuning diffusion models for the motion domain, compared to the image domain [4, 15, 17, 30, 31, 37], lies in the relatively lower diversity of pretraining data. The most commonly used pretraining dataset, HumanML3D [6], contains only 15K samples, which is significantly smaller than the large-scale image datasets with 400M-650M samples [27, 28]. Additionally, HumanML3D includes few data reflecting personality traits, creating a substantial distribution gap when finetuning with the persona-focused PerMo dataset. MoMo [25] found that only 55 samples include styles for 'run,' 'walk,' 'jump,' and 'dance' locomotions in HumanML3D, and we could not find any persona-related samples. Existing diffusion-based MST methods [25, 38] consider only the visual features of the input motion and rely on a fixed text feature, which limits their ability to incorporate new data into the textual component. This inflexibility can lead to a phenomenon known as 'forgetting', where the pretrained model loses its initial capabilities when finetuned with new data [38]. To address this, we introduce learnable persona tokens to capture persona features from new data and propose an adaptation scheme for both text and visuals.", + "bbox": [ + 91, + 327, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another challenge is extracting a consistent persona across different atomic motions, which manifests in diverse ways depending on the action content. For example, the elegance of a ballerina is expressed in her feet as she walks and in her hands as she waves. MST methods have faced similar challenges in disrupting the content of the target motion when transferring styles from the source motion with different content. Approaches like those of Park et al. [20], Jang et al. [9], and MoMo [25] cite this as a limitation. In contrast, MoST [12] addresses the issue with a style disentanglement loss, while MCM-LDM [32] uses the target motion's trajectory to preserve content. However, MoST's loss function requires running the model twice for two different input motions, which is computationally expensive,", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "especially for diffusion models with numerous steps. Additionally, since our goal is to generate motion from text rather than from target motions, the method used by MCM-LDM is not applicable. Therefore, we introduce a novel contrastive learning-based loss specifically designed for text-to-motion diffusion, called persona cohesion loss. This loss facilitates cohesion across motion features with different content but the same persona.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, while MST uses a single input motion, we propose a new fusion method for multiple motions. In the image domain, InstantBooth [31] simply averages persona features, but this leads to unnatural blending in the motion. To address this, we introduce Context-Aware Fusion (CAF), which assigns weights to the fusion of persona features based on the similarity between the prompt and input motions. The summarized contributions of this paper are:", + "bbox": [ + 511, + 212, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the Motion Personalization task and present PerMo, a new large-scale persona dataset captured by multiple actors.", + "- We propose PersonaBooth, a multi-modal finetuning method for diffusion models that introduces text adaptation through a persona token. PersonaBooth applies contrastive learning to effectively capture personas across various content types.", + "- PersonaBooth achieves strong performance in both Motion Personalization and MST tasks and maximizes performance through a CAF for multiple inputs." + ], + "bbox": [ + 531, + 333, + 903, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 512, + 653, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Motion Style Transfer (MST) aims to transfer the style from a source motion to a target motion. The pioneering work by Aberman et al. [1] introduced the AdaIN layer, which simply replaces style statistics in specific layers during training. MotionPuzzle [9] proposes a Graph Convolution Network-based framework that transfers manually specified motion to a desired body part, enabling the combination of multiple source motions. MoST [12] addresses the issue of unwanted blending between different motion contents by introducing a transformer-based model [11] that more clearly separates style and content.", + "bbox": [ + 511, + 537, + 903, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent MST algorithms have seen notable performance gains through the use of diffusion models. MoMo [25] proposes a zero-shot motion transfer method using a pretrained diffusion model, where it mixes two input motions within the attention module of the diffusion model. However, MoMo identifies a limitation: when the source motion content differs significantly from the target motion or text, the generated motion may not follow the target accurately. For example, applying the style of a stationary source motion to a walking target motion may still result in a stationary output. SMooDi [38] proposes a finetuning method for a pretrained diffusion model. However, due to the forgetting issue, it retrained the model with both the original pretrain-", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22757", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aff20c64e70c7a75e15820d1a4cba40d49ffbd82f81d4d88601719985bee9235.jpg", + "image_caption": [ + "Figure 2. The overall framework of PersonaBooth. PersonaBooth has two adaptation paths—visual and text—for finetuning the Motion Diffusion model $(\\mathcal{D})$ . The Persona Extractor extracts both a visual persona feature $(V^{*})$ and a persona token $(P^{*})$ from the input motions. $V^{*}$ is input into the adaptive layer of $\\mathcal{D}$ , while $P^{*}$ is processed together with the input prompt through a Personalized Text Encoder, generating a personalized text feature, which is then input to $\\mathcal{D}$ . The entire model is trained with a classifier-free approach, incorporating a Persona Cohesion Loss. During inference, Context-Aware Fusion is applied for multiple input cases." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 911, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing dataset and the style dataset. Additionally, the use of classifier-based training and the adaptation approach of creating a trainable copy of the entire model led to a training speed 10 times slower than the base diffusion model. MCM-LDM [32] addresses the challenges regarding two different contents of input motions by explicitly extracting the trajectory from the target motion and using it as a condition for the diffusion model. However, it faces a limitation: only trajectories similar to the target motion are generated, which reduces diversity. This contradicts our goal of creating diverse motions based on text prompts.", + "bbox": [ + 88, + 393, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 571, + 225, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a method for multi-modal finetuning of a text-to-motion diffusion model, named PersonaBooth. The overall framework is shown in Fig. 2, and we aim to finetune the pretrained motion diffusion model, $\\mathcal{D}$ . The input includes multiple atomic motions $\\{M_i\\}$ and a text description $T_{in}$ . PersonaBooth generates personalized motion that reflects the description. We use the same motion representation as HumanML3D [6], $M_i \\in \\mathbb{R}^{f \\cdot 263}$ , where $f$ is the number of frames, and 263 includes joint rotations, positions, velocities, and foot contact. First, the Persona Extractor $\\mathcal{E}$ processes each input motion to extract a visual persona feature $V^*$ and a persona token $P^*$ . Built on a pretrained motion clip model [22] ( $\\mathcal{E}_{clip}$ ), which captures the general characteristics of motion sequences, $\\mathcal{E}$ includes a transformer structure ( $\\mathcal{E}_P$ ) specially designed to extract persona features. We applied a contrastive learning scheme to enhance the persona extraction capability of $\\mathcal{E}_P$ , which will be detailed in Sec. 3.1. The extracted $V^*$ is then fed into a newly added adaptive layer in $\\mathcal{D}$ . Meanwhile, the Personalized Text Encoder $\\mathcal{X}$ adaptively incorporates $P^*$ to pro", + "bbox": [ + 91, + 598, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "duce a personalized text feature $T^{*}$ . $T^{*}$ is input into $\\mathcal{D}$ , where $\\mathcal{D}$ ultimately generates the personalized motion. During training, a single motion-text pair is provided to enable reconstruction. During inference, given motion sets $\\{M_i\\}$ , the Context-Aware Fusion module generates weights to perform a weighted combination of the elements within $\\{V_{i}^{*}\\}$ and $\\{P_i^*\\}$ . PersonaBooth explores efficient training by using a classifier-free guidance approach. During training, $\\mathcal{E}_{clip}$ and $\\mathcal{D}$ , excluding the newly added adaptive layer, are kept frozen. Sec.3.1 to Sec.3.3 describe the process based on the training of single motion, $M$ . The inference process for multiple input motions $\\{M_i\\}$ is explained in Sec.3.4.", + "bbox": [ + 511, + 393, + 906, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Persona Extractor", + "text_level": 1, + "bbox": [ + 511, + 584, + 691, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Persona Extractor $\\mathcal{E}$ extracts the persona token $P^{*}$ along with the visual persona feature $V^{*}$ . We used the TMR [22] structure for $\\mathcal{E}_{clip}$ , replacing their text encoder with the CLIP text encoder [26] and retraining it. The CLIP text encoder is the same one used in the Personalized Text Encoder $\\mathcal{X}$ , and therefore, this replacement enables $P^{*}$ to share the same token embedding space as the input prompt. For input motion $M$ , persona features $V^{*}$ and $P^{*}$ are extracted as", + "bbox": [ + 511, + 606, + 906, + 727 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV ^ {*} = \\mathcal {E} ([ c l s ], M ]), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 734, + 903, + 753 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP ^ {*} = \\operatorname {M L P} (Y), \\text {w h e r e} Y = V ^ {*} [ 0 ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 753, + 903, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $[cls]$ denotes a class token.", + "bbox": [ + 511, + 779, + 736, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure the coherence of persona features across different motion contents of the same persona, we introduce the persona cohesion loss $L_{pc}$ . We utilize a supervised contrastive learning scheme [10], where motion samples from the same persona are encouraged to be closer in the persona feature space, while motion samples from different personas are pushed apart. This helps to guide the persona fea", + "bbox": [ + 511, + 795, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22758", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fbe5dc67339bbcb345b260df5695cfb0ec480fbf90ccae02e195487ad7200c6d.jpg", + "image_caption": [ + "Figure 3. Textual and visual adaptation. (a) Personalized Text Encoder, $\\mathcal{X}$ . (b) $t$ -th step of the Motion Diffusion, $\\mathcal{D}$ . Learnable parameters are denoted by the fire icon" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 483, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tures to form well-defined clusters for each persona. For a batch containing $N$ input motion data, $L_{pc}$ is applied to a total of $2N$ data points, with one positive sample for each input motion. The positive sample is randomly selected from the same persona group as the input motion. $L_{pc}$ for a positive pair of index $(i,j)$ is defined as", + "bbox": [ + 89, + 305, + 483, + 397 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p c} = - \\log \\frac {\\exp (\\mathrm {s i m} (h (Y _ {i}) , h (Y _ {j})) / \\tau)}{\\sum_ {k = 1} ^ {2 N} \\mathbb {1} _ {[ k \\neq i ]} \\exp (\\mathrm {s i m} (h (Y _ {i}) , h (Y _ {k})) / \\tau)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 406, + 482, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h(\\cdot)$ denotes projection head. $Y_{j}$ is extracted from $M_{j}$ which belongs to the same persona group as $M$ . $\\mathrm{sim}(u,v)$ indicates the cosine similarity between two vectors $u$ and $v$ . $\\tau$ denotes the temperature parameter.", + "bbox": [ + 89, + 452, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Textual and Visual Adaptation", + "text_level": 1, + "bbox": [ + 89, + 523, + 362, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The discrepancy between our PerMo dataset and the non-persona-specific pretraining dataset, HumanML3D [6] necessitates a carefully designed adaptation mechanism for effective finetuning. To address this, we propose two adaptation paths—textual and visual—to seamlessly integrate the two persona features, $P^{*}$ and $V^{*}$ .", + "bbox": [ + 89, + 544, + 483, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The structure of the Personalized Text Encoder $(\\mathcal{X})$ is shown in Fig. 3 (a). Given an input sentence like \"A person is dancing\" $(T_{in})$ , we introduce a personalized sentence \"[P*] person is dancing\" $(\\tilde{T}_{in})$ , where the word $[P^*]$ precedes the subject. Subsequently, the vector $P^*$ , derived from $\\mathcal{E}$ , replaces the token embedding corresponding to $[P^*]$ . Then, the modified sentence embedding, $\\mathcal{X}_{clip}(T_{in})$ , is adaptively merged with the original sentence embedding, $\\mathcal{X}_{clip}(T_{in})$ . The descriptions in HumanML3D are generally structured with subjects like 'A Person,' or 'Someone' and lack adjectives to characterize the subject. Thus, a smoother adaptation is necessary to enable the pretrained model to handle variations in sentence structure effectively. Specifically, personalized text feature $T^*$ is derived as", + "bbox": [ + 89, + 636, + 483, + 847 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} T ^ {*} = \\mathcal {X} \\left(T _ {i n}, P ^ {*}\\right) \\tag {4} \\\\ = \\mathcal {X} _ {c l i p} (T _ {i n}) + s _ {t} \\cdot \\mathrm {t a n h} (\\gamma_ {t}) \\cdot \\mathcal {X} _ {c l i p} (\\tilde {T} _ {i n}, P ^ {*}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 858, + 482, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma_{t}$ is a learnable parameter initialized as zero which is for zero gating [36]. $s_t$ is a scaling factor [31], a constant used to balance the adaptive layer during inference. CLIP text encoder $\\mathcal{X}_{clip}$ is frozen during training.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For visual adaptation, we introduce a single adaptive layer into the transformer structure, following the approach used in [31] for the UNet. As illustrated in Fig. 3 (b), for $b$ -th transformer block in $t$ -th step of the diffusion, the adaptive layer is placed between the original self-attention layer and the feedforward layer as", + "bbox": [ + 511, + 152, + 906, + 242 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz ^ {\\prime} = z + s _ {v} \\cdot \\tanh (\\gamma_ {v}) \\cdot \\mathrm {A d a p t} ([ z, V ^ {*} ]), \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 253, + 906, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Adapt indicates the adaptive layer and $z$ and $z'$ denote the input and output features of the adaptive layer, respectively. $\\gamma_v$ and $s_v$ serve as the gating parameter and scaling factor. We evaluated several adaptive layer structures, including self-attention, cross-attention, and the AdaIN layer, which is widely used in style transfer [1, 12]. Among these, self-attention was found to be the most effective.", + "bbox": [ + 511, + 280, + 905, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 513, + 395, + 617, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We used a pretrained 50-step MDM [33] for $\\mathcal{D}$ . Diffusion is modeled as a Markov noising process $\\{M^t\\}_{t=0}^T$ , where $M^0$ is sampled from the training motion, and $t$ denotes the noise time-step. The diffusion loss for finetuning is given by", + "bbox": [ + 511, + 417, + 905, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {D} := \\mathbb {E} _ {M ^ {0}, t, T} \\left[ \\left\\| M ^ {0} - \\mathcal {D} \\left(M ^ {t}, t, V ^ {*}, T ^ {*}\\right) \\right\\| _ {2} ^ {2} \\right] + L _ {\\text {g e o}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 487, + 906, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{geo}$ is the geometric loss used during the MDM pretraining [33]. The final training loss is expressed as", + "bbox": [ + 511, + 529, + 905, + 560 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {D} + \\lambda L _ {p c}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 571, + 903, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a weight for the $L_{pc}$ .", + "bbox": [ + 511, + 598, + 725, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize Classifier-Free Guidance (CFG) [8], which is a technique for adjusting the trade-off between the diversity and fidelity of generated samples. Similarly, we applied CFG during finetuning to enable control over the trade-off between the pretrained diffusion model's capability for generating diverse motions and fidelity to the newly provided conditions, $V^{*}$ and $T^{*}$ . During training, $V^{*}$ and $T^{*}$ are randomly dropped with a $10\\%$ probability, respectively. During sampling, the output of $t$ -th step is extrapolated as", + "bbox": [ + 511, + 613, + 905, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {D}} \\left(M ^ {t}, t, V ^ {*}, T ^ {*}\\right) = b \\mathcal {D} _ {T} + (1 - b) \\mathcal {D} _ {V}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 758, + 906, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {T} = \\mathcal {D} (M ^ {t}, t, V ^ {*}, \\emptyset) + g _ {t} (\\mathcal {D} (M ^ {t}, t, V ^ {*}, T ^ {*}) - \\mathcal {D} (M ^ {t}, t, V ^ {*}, \\emptyset)),\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 779, + 905, + 795 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {V} = \\mathcal {D} (M ^ {t}, t, \\emptyset , T ^ {*}) + g _ {v} (\\mathcal {D} (M ^ {t}, t, V ^ {*}, T ^ {*}) - \\mathcal {D} (M ^ {t}, t, \\emptyset , T ^ {*})),\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 797, + 905, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varnothing$ indicates the absence of an input. $g_{t}$ and $g_{v}$ represent guidance scales, while $b$ serves as the balancing factor between the modalities. As $g_{v}$ increases, the model puts more focus on the persona inherent in the visual feature. However, if $g_{v}$ is set too high, inconsistencies with the", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22759", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3dd2d8527a6233f29249a741e201a0e87c9fef8fe012f8a5f2c2d6431de15ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 88, + 354, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/595e37677116b5058e08ec9259b6e5215caa422031f7ce05bdc3b1efddc9ddf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 356, + 88, + 418, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/88bed377e986fbedf982d65103336f9ae2faa6432c336042c94e7b7143806aee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 88, + 480, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cbc0a3f448144278e960eb0bd12eac8ef2db027a1b35196f3a0ea6672bc6f96e.jpg", + "image_caption": [ + "(a)", + "(b)", + "(c)", + "Figure 4. (a) Motion capture studio and examples of data formats: skeleton, markers, and mesh. (b) Unique persona expressions of each actor in the 'Childish' category. (c) Rendered mesh for each actor in the 'Fearful' category" + ], + "image_footnote": [], + "bbox": [ + 93, + 229, + 480, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "text prompt may arise. On the other hand, $g_{t}$ controls the model's adherence to the prompt, with the intensity of $P^{*}$ in the text being adjusted by $s_t$ in Eq. (4). The hyperparameter settings are specified in Sec. 5, and an ablation study for these is provided in the supplementary material.", + "bbox": [ + 89, + 392, + 483, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Context-Aware Fusion for Multiple Inputs", + "text_level": 1, + "bbox": [ + 89, + 479, + 450, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, if multiple motions are input, these additional cues can help with generating motion. However, it's also crucial to choose the right cues, as transferring persona from motions with very different content types, even with $L_{pc}$ , can result in unnatural movements. A straightforward approach might be to take the mean of all features, as suggested in [31]. In the motion domain, however, taking all motions can result in blended motions that appear implausible. To overcome this, we introduce a Context-Aware Fusion (CAF) method that prioritizes input motions based on their contextual relevance to the input prompt. To assess this relevance, we use the motion encoder and text encoder of the motion clip ( $\\mathcal{E}$ clip and $\\mathcal{X}$ clip), which align the text and motion feature spaces. The CAF is formulated as", + "bbox": [ + 89, + 503, + 483, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\text {T o p - k}} = \\operatorname {a r g m a x} _ {i} \\left(S _ {i}\\right), \\quad S _ {i} = \\sin \\left(\\mathcal {E} _ {\\text {c l i p}} \\left(M _ {i}\\right), \\bar {T} ^ {*}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 728, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} = \\left\\{ \\begin{array}{l l} \\frac {\\exp (S _ {i})}{\\sum_ {n} \\exp (S _ {n})}, & \\text {i f} i \\in I _ {\\mathrm {T o p - k}}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 748, + 480, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{T}^*$ represents the personalized text feature, using the average value of $\\{P^{*}\\}$ as a persona token. $\\mathrm{sim}(\\cdot)$ denotes cosine similarity. Consequently,", + "bbox": [ + 89, + 813, + 483, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV ^ {*} = \\sum_ {i} w _ {i} V _ {i} ^ {*}, \\quad P ^ {*} = \\sum_ {i} w _ {i} P _ {i} ^ {*}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 872, + 482, + 904 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1528411b0a332144d35fd2c18fc59d36a499c6f3ce951715c9132c43a43bab0a.jpg", + "table_caption": [ + "Table 2. Comparison with existing motion style datasets. The note for \\* is provided in the main text. - indicates no information" + ], + "table_footnote": [], + "table_body": "
DatasetActorsStylesContentsClipsMeshText
Xia [34]-86572XX
BFA [1]116932XX
BN-1 [13]2*1517175XX
BN-2 [13]2*7102,902XX
100Style [19]1100*8*810XX
PerMo(ours)534106,610✓✓
", + "bbox": [ + 519, + 128, + 895, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. PerMo Dataset", + "text_level": 1, + "bbox": [ + 513, + 266, + 663, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We collected a large-scale PerMo dataset, capturing personas from multiple actors. To ensure variety, we hired five professional motion capture actors of diverse genders and body types. Each actor is assigned to perform 34 styles, categorized into Age, Character, Condition, Emotion, Traits, and Surroundings, resulting in a total of 170 personas. Each actor performed 10 distinct contents for every style, carefully selected to engage different body parts.", + "bbox": [ + 511, + 292, + 905, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The dataset was captured in a studio equipped with 33 OptiTrack cameras, using 41 optical markers per person. For compatibility with other motion datasets, we converted the marker data to SMPL-H [29] format using $\\mathrm{Mosh}++$ [18]. All data underwent expert cleaning and verification, with the entire process adhering to strict quality guidelines. Fig. 4 (a) illustrates the motion capture studio and the provided data formats, while (b) highlights the distinct personas expressed by each actor. Additionally, we constructed 20-30 varied descriptions for each content type, following research indicating that high-quality finetuning descriptions improve prompt-based control [7]. Starting with one detailed description, we instructed ChatGPT to generate a range from high-level descriptions (e.g., \"A person is jumping ahead\") to low-level descriptions (e.g., \"A person pushes off the ground and jumps forward several times\"). Additional details on data categories, capture process, data structure, and examples can be found in the supplementary.", + "bbox": [ + 511, + 415, + 905, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2 shows a comparison with existing motion style datasets. Notably, PerMo is the first dataset to collect data from multiple actors. Our dataset also offers the highest number of total clips and content categories among trainable datasets and is the only one that includes mesh, marker data, and descriptions. The BN datasets [13] involve two actors; however, for the same category, only one actor is used (BN-1), or there are no tags identifying the actors (BN-2). BN-1 is unsuitable for training, as it contains only a single motion sequence per category. The 100Style [19] offers the largest variety of styles, but 58 of its categories focus on content, rather than on style [38]. Additionally, the eight content categories in this dataset mainly consist of walking or running, offering limited variation.", + "bbox": [ + 511, + 689, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22760", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/80d672849ac141d6afeb7cbc97f39e81639c75892c3c876643daaeb0c585f4db.jpg", + "image_caption": [ + "Figure 5. Example of the ablation study. The input motions are from the 'Uppity' of Actor 1. The input prompt is \"A person walks in a circle.\" In (a) and (b), only $M_1$ is provided for the input, while both $M_1$ and $M_2$ are provided for (c) and (d). $L_{pc}$ encourages the generated motion to closely follow the prompt, while CAF prevents the motion from blending. We set $k = 1$ for CAF" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 910, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 316, + 225, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We conduct experiments using the proposed PerMo and 100Style [19] datasets. During training, a single motion is cropped differently to serve as both the input and the GT motion for generation. We set $\\lambda$ to $10^{-2}$ and use the AdamW [16] optimizer with a learning rate of $10^{-4}$ . Training runs for 500 epochs with a batch size of 64. We set $s_t$ and $s_v$ to 1 during training and to 0.3 during inference. $g_t$ is set to 10, and $g_v$ to 15 and 10 for PerMo and 100Style, respectively. We set $b$ to 0.7 for single-input and 0.5 for multi-input settings, and $k$ to 5 for CAF.", + "bbox": [ + 89, + 342, + 483, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Settings. For quantitative evaluation, we use descriptions from the HumanML3D test set as prompts and motion samples from the PerMo and 100Style datasets as inputs. For each description, $[P^*]$ is logically placed as a modifier for the subject. We used two evaluation settings: Single Input (SI) and Multiple Input (MI). In SI, one motion is sampled from the entire dataset for comparison with MST methods. In MI, a persona set is first sampled, and then $|M_i|$ motions are drawn from it.", + "bbox": [ + 89, + 494, + 483, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance is evaluated using Frechet Inception Distance (FID), R-Precision, and Diversity—metrics commonly used in motion generation tasks [3, 5, 6, 33]—as well as Persona Recognition Accuracy (PRA), a metric adopted by Style Recognition Accuracy (SRA) in MST [9, 32]. FID assesses the overall quality and realism of generated motions and is our primary metric [5]. R-Precision evaluates text-to-motion alignment accuracy [6], and Diversity [5] measures how well the generated motions reflect a broad distribution of personas. PRA uses a pretrained persona classifier to evaluate the persona consistency of generated motions. The details are in the supplementary material.", + "bbox": [ + 89, + 630, + 483, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 818, + 243, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 demonstrates the effects of our key contributions. The baseline represents results obtained using only visual adaptation. When we introduce $P^*$ for textual adaptation, the FID score drops significantly, and both PRA and Diver", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/636604fe04ea667393ce6e45e83726b8e449122a09f5285b64a2ce9fad8cb03b.jpg", + "table_caption": [ + "Table 3. Ablation study of the proposed components. $|M_{i}| = \\max$ indicates that all motions in the chosen persona set are input" + ], + "table_footnote": [], + "table_body": "
MethodsFID ↓R Precision ↑PRA avg.↑Diversity ↑
Top 1Top 2Top 3
Single Input (SI) Setting
baseline7.450.060.110.1617.997.48
+ P*5.060.050.100.1518.268.01
+Lpc3.180.150.260.3318.057.74
Multiple Input (MI) Setting
Mi|=14.270.080.150.2016.207.71
Mi|=max, [31]3.520.190.300.3819.247.88
Mi|=max, CAF2.950.190.300.3918.138.12
", + "bbox": [ + 513, + 356, + 906, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sity increase. This indicates that multimodal integration of persona information not only strengthens the reflection of persona but also supports the generation of plausible motion. Furthermore, adding $L_{pc}$ leads to an additional reduction in FID and a notable increase in R-precision. This implies that $L_{pc}$ helps the Persona Extractor effectively capture persona essence disentangled from the input content, resulting in generated motion that aligns well with the prompt without compromising content.", + "bbox": [ + 509, + 534, + 906, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, the effectiveness of CAF is validated in the MI setting. When multiple input motions are used $(|M_i| = \\max)$ , all metrics improve compared to using a single input $(|M_i| = 1)$ due to the higher likelihood of finding motion candidates that align with the prompt. Additionally, even in cases with multiple inputs, selecting the contextually most relevant Top-k motions—as done by CAF—is shown to be crucial. For $|M_i| = \\max$ , [31] presents a baseline approach that uses the average feature of all inputs. Introducing the proposed CAF reduces FID while increasing Diversity, indicating that the generated motions are sufficiently diverse and plausible rather than oversimplified by averaging. Although the PRA score decreases, we found through qualitative results that generating plausible motion is a more crucial factor in improving motion quality.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22761", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a9f0b197b8b28ee9feabe82438f4d8c33d972774853cab37262920349a83d5f6.jpg", + "image_caption": [ + "Figure 6. Example reflecting the different personas of Actors 1-5 for the same 'Childish' category. The input prompt is \"A person is running and bending to pick something.\" Red arrows highlight the frames that clearly show distinct movements. The input motion of each actor can be referenced in Fig. 4 (b)" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2464820ce61f3f86075dc4f8c4381f464a3b1d5df70fcaaebd1684b35ba10893.jpg", + "table_caption": [ + "Table 4. Ablation study regarding adaptation. PersonaBooth (SI) indicates our complete model for single inputs" + ], + "table_footnote": [], + "table_body": "
MethodsFID ↓R Precision ↑PRA avg.↑Diversity ↑
Top 1Top 2Top 3
AdaIN8.770.040.070.1015.107.33
Cross-Attn8.960.040.070.1016.377.62
w/o T adapt.3.610.140.230.3117.777.88
PersonaBooth (SI)3.180.150.260.3318.057.74
", + "bbox": [ + 91, + 345, + 483, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 provides an example. When using only $M_{1}$ as the input motion, if $L_{pc}$ is not applied, the generated motion fails to follow the path provided in the prompt, as shown in (a). However, with $L_{pc}$ , the motion better aligns with the prompt's trajectory, though it still includes unintended hand-waving motions unrelated to the prompt. When using multiple input motions, including $M_{2}$ (a walking motion with arms crossed), taking the mean of features results in a blended motion, as shown in (c). This blending causes one arm to bend while the other hangs low, creating unnatural poses. In contrast, with CAF, $M_{2}$ receives a higher weight, resulting in a more natural walking motion with arms crossed, as shown in (d).", + "bbox": [ + 88, + 460, + 482, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 presents various ablation results related to adaptation. For the visual adaptation layer, Self-Attention proved to be the most effective compared to AdaIN and CrossAttn. Additionally, introducing text adaptation at Personalized Text Encoder improves the PRA, FID, and R-precision metrics. Fig. 6 demonstrates how PersonaBooth effectively captures and represents each actor's unique action. The red arrows specifically highlight the distinct arm movements of each persona. Additional details, including variations in body tilt and movement speed, are more clearly observable in the supplementary video materials.", + "bbox": [ + 88, + 656, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison with State-of-the-Art methods", + "text_level": 1, + "bbox": [ + 89, + 832, + 457, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our PersonaBooth with the state-of-the-art diffusion-based MST approaches, MoMo [25] and MCM-LDM [32], on the PerMo dataset. Additionally, we evalu", + "bbox": [ + 88, + 854, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ate it on an existing style dataset, 100Style [19]. As MCM-LDM does not support prompt-based generation, we used MDM-generated motions [33] as target motions for evaluation. MCM-LDM was evaluated using the original model trained on HumanML3D, as well as finetuned models, referred to as MCM-LDM, which were trained on the PerMo and 100Style datasets.", + "bbox": [ + 511, + 308, + 906, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.1. PerMo Dataset", + "text_level": 1, + "bbox": [ + 511, + 420, + 663, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The comparative results for the PerMo dataset are shown in Table 5, with all subcategory results displayed for PRA. The single-input PersonaBooth significantly reduces FID compared to existing methods, indicating that while previous studies struggled with PerMo—a challenging dataset containing more diverse content than 100Style—PersonaBooth performs robust in generating high-quality motion. Additionally, PersonaBooth achieves the highest performance in R-Precision, PRA, and Diversity compared to existing methods. Applying multiple inputs and CAF further enhances performance. It's noteworthy that while MoMo and MCM-LDM utilized a 100-step and 1000-step diffusion model, respectively, we used a 50-step diffusion model.", + "bbox": [ + 511, + 438, + 906, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 7 shows qualitative results. MoMo [25] notes that their generated motions often follow the input motion's direction instead of the target's, which they attempted to copy the root rotation from the input motion (shown in (a) before and (b) after adjustment). The prompts for \"turns in the air\" were not met in (a), (b), or (c). Particularly, (c) does not reflect the style of ballerina. In contrast, PersonaBooth accurately generates the motion by first performing a hop that reflects the style, followed by a precise turn in the air. Example videos are provided in the supplementary material.", + "bbox": [ + 511, + 635, + 908, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.2. 100Style Dataset", + "text_level": 1, + "bbox": [ + 511, + 791, + 674, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6 presents the comparative results on the 100Style dataset. MoMo demonstrates high PRA and Diversity, indicating effective style reflection. However, it shows very low R-precision, suggesting poor alignment between generated motions and prompts. Finetuned MCM-LDM improved performance on nearly all metrics compared to its", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22762", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d84afe707c8fa765c6116804b0871950538a73e935b650e244d333da447c3528.jpg", + "table_caption": [ + "Table 5. Comparison with the state-of-the-art methods on the PerMo dataset. The comparison is made in the Single Input (SI) setting as the existing methods do not support multiple inputs. MCM-LDM* indicates the model is finetuned on the PerMo dataset." + ], + "table_footnote": [], + "table_body": "
MethodsStepsFID ↓R Precision ↑PRA ↑Diversity ↑
Top 1Top 2Top 3AgeChar1Char2Cond1Cond2Emo1Emo2TraitSurAvg.
MoMo [25]10013.910.050.090.1328.2310.2311.308.7315.979.9714.5712.607.9313.476.74
MCM-LDM [32]10009.160.130.230.3047.3811.4410.347.6318.9011.6917.2015.719.2217.006.70
MCM-LDM* [32]10009.700.130.230.3046.4912.129.227.9519.7711.7518.0515.187.3316.766.76
PersonaBooth (SI)503.180.150.260.3348.0013.6711.976.6918.6714.3419.7517.068.8018.057.74
PersonaBooth (MI)502.950.190.300.3951.9514.8912.005.1622.959.4418.1415.4510.0918.138.12
", + "bbox": [ + 91, + 130, + 906, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c4fc08e65ce11eb7aaf80941e558b4b5fea3416c29c54c283c1aaa8a666d9c2.jpg", + "image_caption": [ + "Input Motion (Actor1 Ballerina)" + ], + "image_footnote": [], + "bbox": [ + 91, + 243, + 264, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/548ca00ac43886abd51863ab34976cf0a44b554855f2719fbc77784102f8a7e6.jpg", + "image_caption": [ + "(a) MoMo (w/o copy root rot.)" + ], + "image_footnote": [], + "bbox": [ + 261, + 247, + 424, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1acb58de0e6575b6d5c1d341037d3df68d2967cf0543955c1439f2bc276698ae.jpg", + "image_caption": [ + "(b) MoMo" + ], + "image_footnote": [], + "bbox": [ + 428, + 250, + 596, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d08c0965aadb5a54a987083b2d3816180587f79c269b926e90f53247e25d00d2.jpg", + "image_caption": [ + "(c) MCM-LDM*" + ], + "image_footnote": [], + "bbox": [ + 596, + 247, + 756, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8dcf2acd08cef812a80589d4592ae4800a31e954e2cf3169b5650e953d17d1b4.jpg", + "image_caption": [ + "(d) PersonaBooth (SI)" + ], + "image_footnote": [], + "bbox": [ + 761, + 247, + 895, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57fa59f08fcc0be1c4672dcf6993f44d71ad5490f7a17f6e8c14fa043448f51b.jpg", + "image_caption": [ + "Input Motion (Airplane)" + ], + "image_footnote": [], + "bbox": [ + 91, + 393, + 254, + 489 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e9a864749f97a8a4e6adee2b53f0f5ec8174ecc96f67afd0da4278d6d8c21b5.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison on PerMo dataset. The input prompt is \"A person hops forward and turns in the air.\"", + "(a) MoMo (w/o copy root rot.)" + ], + "image_footnote": [], + "bbox": [ + 254, + 391, + 428, + 493 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/560be798c05ca2f8bfea0febbdd79a5ae77151fa009a7b8955dfcf2dae5678a3.jpg", + "image_caption": [ + "(b) MoMo", + "Figure 8. Qualitative comparison on 100Style dataset. The input prompt is \"A person walks backward and sits down on the chair.\"" + ], + "image_footnote": [], + "bbox": [ + 429, + 405, + 596, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1ae6bf02847599e9ad486805e27337755b1e45d693702dbad38bc86847aba36e.jpg", + "image_caption": [ + "(c) MCM-LDM*" + ], + "image_footnote": [], + "bbox": [ + 594, + 392, + 733, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f827cad047c1c818a0e36d6e109082b3c130348292c75fe270b731676448446d.jpg", + "image_caption": [ + "(d) PersonaBooth (SI)" + ], + "image_footnote": [], + "bbox": [ + 733, + 392, + 905, + 494 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2f0ed0d874ea2558dbd54490fef16d39212112133d975950eb3ea1db42e593a2.jpg", + "table_caption": [ + "Table 6. Comparison with the state-of-the-art methods on the 100Style dataset. MCM-LDM* indicates the model is finetuned on the 100Style dataset" + ], + "table_footnote": [], + "table_body": "
MethodsFID ↓R Precision ↑SRA ↑Diversity ↑
Top 1Top 2Top 3
MoMo [25]5.970.070.100.1460.577.82
MCM-LDM [32]7.200.180.290.3754.446.76
MCM-LDM* [32]6.530.170.280.3754.617.12
PersonaBooth (SI)3.270.200.310.4064.537.90
", + "bbox": [ + 91, + 603, + 483, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "original model. Although finetuned MCM-LDM achieved high R-precision, it also had a high FID, resulting in unnatural motion generation, with low PRA and Diversity indicating weaker style reflection. In contrast, our single-input PersonaBooth significantly reduces FID and achieves top performance across all metrics, including R-precision, PRA, and Diversity.", + "bbox": [ + 88, + 717, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 8 (a) and (b), MoMo produces sliding motions that don't align with the prompt. MCM-LDM shows sideward movements. In contrast, PersonaBooth accurately reflects both the input motion's style, like arm extension and leaning, and the sitting motion in line with the prompt.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 550, + 633, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose the PerMo dataset and the PersonaBooth framework for Motion Personalization. Through contrastive learning, we effectively capture the essence of a persona. Furthermore, we enhance the integration of persona traits into diffusion models using a multi-modal finetuning method. We also introduce the CAF, which effectively handles multiple inputs, ensuring cohesive representation of personalized motion characteristics.", + "bbox": [ + 511, + 575, + 906, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Works. The PerMo contains relatively short motion sequences, which sometimes lead to stationary poses at the end of generated motions of the same length. However, this issue does not occur when prompted with longer actions, suggesting that automatic adjustment of motion length remains a task for future work. In addition, we plan to investigate approaches for applying CAF to each sequential action within a prompt individually. For instance, applying CAF separately to 'run' and 'jump' in the prompt \"A person runs and then jumps.\"", + "bbox": [ + 511, + 696, + 908, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement.", + "text_level": 1, + "bbox": [ + 511, + 851, + 651, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by MSIT, Korea, under [RS-2022-00164860] and [RS-2024-00418641] supervised by IITP.", + "bbox": [ + 511, + 866, + 905, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22763", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. ACM Transactions on Graphics (TOG), 39(4):64-1, 2020. 2, 4, 5", + "[2] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1", + "[3] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 1, 2, 6", + "[4] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[5] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3D human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 6", + "[6] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 1, 2, 3, 4, 6", + "[7] Xingzhe He, Zhiwen Cao, Nicholas Kolkin, Lantao Yu, Kun Wan, Helge Rhodin, and Ratheesh Kalarot. A data perspective on enhanced identity preservation for diffusion personalization. arXiv preprint arXiv:2311.04315, 2023. 5", + "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4", + "[9] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2, 6", + "[10] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 3", + "[11] Boeun Kim, Hyung Jin Chang, Jungho Kim, and Jin Young Choi. Global-local motion transformer for unsupervised skeleton-based action learning. In European Conference on Computer Vision, pages 209-225. Springer, 2022. 2", + "[12] Boeun Kim, Jungho Kim, Hyung Jin Chang, and Jin Young Choi. Most: Motion style transformer between diverse action contents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1705-1714, 2024. 2, 4", + "[13] Makito Kobayashi, Chen-Chieh Liao, Keito Inoue, Sentaro Yojima, and Masafumi Takahashi. Motion capture dataset" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for practical use of ai-based motion editing and stylization. arXiv preprint arXiv:2306.08861, 2023.5", + "[14] Jehee Lee, Jinxiang Chai, Paul SA Reitsma, Jessica K Hodgins, and Nancy S Pollard. Interactive control of avatars animated with human motion data. In Proceedings of the 29th annual Conference on Computer Graphics and Interactive Techniques, pages 491-500, 2002. 1", + "[15] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[16] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[17] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2", + "[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 5", + "[19] Ian Mason, Sebastian Starke, and Taku Komura. Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 5(1):1-18, 2022. 2, 5, 6, 7", + "[20] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on Computer Graphics and Interactive Techniques, 4(3):1-17, 2021. 2", + "[21] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 1", + "[22] Mathis Petrovich, Michael J Black, and Gül Varol. Tmr: Text-to-motion retrieval using contrastive 3D human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9488-9497, 2023. 3", + "[23] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2", + "[24] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1", + "[25] Sigal Raab, Inbar Gat, Nathan Sala, Guy Tevet, Rotem Shalev-Arkushin, Ohad Fried, Amit H Bermano, and Daniel Cohen-Or. Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. arXiv preprint arXiv:2406.06508, 2024. 2, 7, 8", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22764", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2", + "[29] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 5", + "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2", + "[31] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 2, 4, 5, 6", + "[32] Wenfeng Song, Xingliang Jin, Shuai Li, Chenglizhao Chen, Aimin Hao, Xia Hou, Ning Li, and Hong Qin. Arbitrary motion style transfer with multi-condition motion latent diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821-830, 2024. 2, 3, 6, 7, 8", + "[33] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 4, 6, 7", + "[34] Shihong Xia, Congyi Wang, Jinxiang Chai, and Jessica Hodgins. Realtime style transfer for unlabeled heterogeneous human motion. ACM Transactions on Graphics (TOG), 34 (4):1-10, 2015. 2, 5", + "[35] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 2", + "[36] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 4", + "[37] Xulu Zhang, Xiao-Yong Wei, Wengyu Zhang, Jinlin Wu, Zhaoxiang Zhang, Zhen Lei, and Qing Li. A survey on personalized content synthesis with diffusion models. arXiv preprint arXiv:2405.05538, 2024. 2", + "[38] Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, and Huaizu Jiang. Smoodi: Stylized motion diffusion model. arXiv preprint arXiv:2407.12783, 2024. 2, 5" + ], + "bbox": [ + 91, + 92, + 482, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22765", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PersonaBooth_ Personalized Text-to-Motion Generation/62946df9-a553-4225-bd84-84b1acdb1a5f_model.json b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/62946df9-a553-4225-bd84-84b1acdb1a5f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e194e23161639598b0df98305765cc6c45dbd26c --- /dev/null +++ b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/62946df9-a553-4225-bd84-84b1acdb1a5f_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.131, + 0.78, + 0.151 + ], + "angle": 0, + "content": "PersonaBooth: Personalized Text-to-Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.193, + 0.905, + 0.266 + ], + "angle": 0, + "content": "Boeun Kim1,2,3 Hea In Jeong2 JungHoon Sung3 Yihua Cheng1 Jeongmin Lee2 \nJu Yong Chang4 Sang-II Choi3 Younggeun Choi3 Saim Shin2 Jungho Kim2 Hyung Jin Chang1 \n1University of Birmingham 2Korea Electronics Technology Institute 3Dankook University 4Kwangwoon University \nhttp://boeun-kim.github.io/page-PersonaBooth" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.311, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.486, + 0.691 + ], + "angle": 0, + "content": "This paper introduces Motion Personalization, a new task that generates personalized motions aligned with text descriptions using several basic motions containing Persona. To support this novel task, we introduce a new large-scale motion dataset called PerMo (PersonaMotion), which captures the unique personas of multiple actors. We also propose a multi-modal finetuning method of a pretrained motion diffusion model called PersonaBooth. PersonaBooth addresses two main challenges: i) A significant distribution gap between the persona-focused PerMo dataset and the pretraining datasets, which lack persona-specific data, and ii) the difficulty of capturing a consistent persona from the motions vary in content (action type). To tackle the dataset distribution gap, we introduce a persona token to accept new persona features and perform multi-modal adaptation for both text and visuals during finetuning. To capture a consistent persona, we incorporate a contrastive learning technique to enhance intra-cohesion among samples with the same persona. Furthermore, we introduce a context-aware fusion mechanism to maximize the integration of persona cues from multiple input motions. PersonaBooth outperforms state-of-the-art motion style transfer methods, establishing a new benchmark for motion personalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.723, + 0.222, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Imagine a scenario where, by recording just a few basic movements of yourself, an avatar in a virtual world can mirror your personal traits and unique motion style. This kind of 'motion personalization' enables realistic interactions in virtual spaces such as games and the metaverse [14]. Moreover, if avatar motions could be directed through text, creating video content would become easy without the need for real-life stunt actors [2, 24]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a task called Motion Personalization, which aims to generate text-based motions that" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.31, + 0.889, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.627, + 0.907, + 0.684 + ], + "angle": 0, + "content": "Figure 1. Motion Personalization generates text-driven, personalized motions based on persona embedded in atomic input motions. We propose a new framework, PersonaBooth, along with a new benchmark dataset, PerMo, for Motion Personalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.903 + ], + "angle": 0, + "content": "reflect individual personas using a few key atomic movements, such as jumping, punching, or walking. We define a persona as the unique style expression of an individual. As illustrated in Fig. 1, this task aims to generate a broad range of realistic motions robustly. To facilitate this, we introduce a model called PersonaBooth and a new benchmark, PerMo (PersonaMotion), specifically designed for motion personalization. Table 1 provides a comparison of similar motion generation tasks. Text-to-Motion Generation approaches [6, 21] generate motion using only text as input, without any reference motion, and numerous diffusion-based methods [3, 33, 35] have been proposed for this task. Motion Style Transfer (MST) focuses on transferring style" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22756" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.09, + 0.43, + 0.103 + ], + "angle": 0, + "content": "Table 1. Comparison of motion generation tasks" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.104, + 0.481, + 0.19 + ], + "angle": 0, + "content": "
Motion TaskExisting DatasetsSource MotionsMulti-modal Finetuning
T2M Generation [3, 33, 35][6] [23]NoneX
Style Transfer [25, 32, 38][34][1][19]SingleX
Personalization (Ours)NoneMultiple✓
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.206, + 0.483, + 0.327 + ], + "angle": 0, + "content": "from a single source motion. While several datasets and methods are available for MST [25, 32, 34, 38], the use of a single source motion restricts the ability to generate the broad and diverse motions seen in real life [12]. The proposed Motion Personalization has no existing prior approaches or datasets. Furthermore, our proposed method introduces multi-modal finetuning—a process neglected in existing diffusion-based MST methods." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.328, + 0.483, + 0.688 + ], + "angle": 0, + "content": "PersonaBooth is a multi-modal finetuning method for pretrained text-to-motion diffusion models. The challenge in finetuning diffusion models for the motion domain, compared to the image domain [4, 15, 17, 30, 31, 37], lies in the relatively lower diversity of pretraining data. The most commonly used pretraining dataset, HumanML3D [6], contains only 15K samples, which is significantly smaller than the large-scale image datasets with 400M-650M samples [27, 28]. Additionally, HumanML3D includes few data reflecting personality traits, creating a substantial distribution gap when finetuning with the persona-focused PerMo dataset. MoMo [25] found that only 55 samples include styles for 'run,' 'walk,' 'jump,' and 'dance' locomotions in HumanML3D, and we could not find any persona-related samples. Existing diffusion-based MST methods [25, 38] consider only the visual features of the input motion and rely on a fixed text feature, which limits their ability to incorporate new data into the textual component. This inflexibility can lead to a phenomenon known as 'forgetting', where the pretrained model loses its initial capabilities when finetuned with new data [38]. To address this, we introduce learnable persona tokens to capture persona features from new data and propose an adaptation scheme for both text and visuals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Another challenge is extracting a consistent persona across different atomic motions, which manifests in diverse ways depending on the action content. For example, the elegance of a ballerina is expressed in her feet as she walks and in her hands as she waves. MST methods have faced similar challenges in disrupting the content of the target motion when transferring styles from the source motion with different content. Approaches like those of Park et al. [20], Jang et al. [9], and MoMo [25] cite this as a limitation. In contrast, MoST [12] addresses the issue with a style disentanglement loss, while MCM-LDM [32] uses the target motion's trajectory to preserve content. However, MoST's loss function requires running the model twice for two different input motions, which is computationally expensive," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "especially for diffusion models with numerous steps. Additionally, since our goal is to generate motion from text rather than from target motions, the method used by MCM-LDM is not applicable. Therefore, we introduce a novel contrastive learning-based loss specifically designed for text-to-motion diffusion, called persona cohesion loss. This loss facilitates cohesion across motion features with different content but the same persona." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.333 + ], + "angle": 0, + "content": "In addition, while MST uses a single input motion, we propose a new fusion method for multiple motions. In the image domain, InstantBooth [31] simply averages persona features, but this leads to unnatural blending in the motion. To address this, we introduce Context-Aware Fusion (CAF), which assigns weights to the fusion of persona features based on the similarity between the prompt and input motions. The summarized contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.334, + 0.904, + 0.378 + ], + "angle": 0, + "content": "- We introduce the Motion Personalization task and present PerMo, a new large-scale persona dataset captured by multiple actors." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.379, + 0.905, + 0.453 + ], + "angle": 0, + "content": "- We propose PersonaBooth, a multi-modal finetuning method for diffusion models that introduces text adaptation through a persona token. PersonaBooth applies contrastive learning to effectively capture personas across various content types." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.455, + 0.905, + 0.499 + ], + "angle": 0, + "content": "- PersonaBooth achieves strong performance in both Motion Personalization and MST tasks and maximizes performance through a CAF for multiple inputs." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.334, + 0.905, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.513, + 0.655, + 0.529 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.905, + 0.704 + ], + "angle": 0, + "content": "The Motion Style Transfer (MST) aims to transfer the style from a source motion to a target motion. The pioneering work by Aberman et al. [1] introduced the AdaIN layer, which simply replaces style statistics in specific layers during training. MotionPuzzle [9] proposes a Graph Convolution Network-based framework that transfers manually specified motion to a desired body part, enabling the combination of multiple source motions. MoST [12] addresses the issue of unwanted blending between different motion contents by introducing a transformer-based model [11] that more clearly separates style and content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Recent MST algorithms have seen notable performance gains through the use of diffusion models. MoMo [25] proposes a zero-shot motion transfer method using a pretrained diffusion model, where it mixes two input motions within the attention module of the diffusion model. However, MoMo identifies a limitation: when the source motion content differs significantly from the target motion or text, the generated motion may not follow the target accurately. For example, applying the style of a stationary source motion to a walking target motion may still result in a stationary output. SMooDi [38] proposes a finetuning method for a pretrained diffusion model. However, due to the forgetting issue, it retrained the model with both the original pretrain-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22757" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.912, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.309, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 2. The overall framework of PersonaBooth. PersonaBooth has two adaptation paths—visual and text—for finetuning the Motion Diffusion model \\((\\mathcal{D})\\). The Persona Extractor extracts both a visual persona feature \\((V^{*})\\) and a persona token \\((P^{*})\\) from the input motions. \\(V^{*}\\) is input into the adaptive layer of \\(\\mathcal{D}\\), while \\(P^{*}\\) is processed together with the input prompt through a Personalized Text Encoder, generating a personalized text feature, which is then input to \\(\\mathcal{D}\\). The entire model is trained with a classifier-free approach, incorporating a Persona Cohesion Loss. During inference, Context-Aware Fusion is applied for multiple input cases." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.395, + 0.484, + 0.562 + ], + "angle": 0, + "content": "ing dataset and the style dataset. Additionally, the use of classifier-based training and the adaptation approach of creating a trainable copy of the entire model led to a training speed 10 times slower than the base diffusion model. MCM-LDM [32] addresses the challenges regarding two different contents of input motions by explicitly extracting the trajectory from the target motion and using it as a condition for the diffusion model. However, it faces a limitation: only trajectories similar to the target motion are generated, which reduces diversity. This contradicts our goal of creating diverse motions based on text prompts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.573, + 0.226, + 0.59 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.599, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We propose a method for multi-modal finetuning of a text-to-motion diffusion model, named PersonaBooth. The overall framework is shown in Fig. 2, and we aim to finetune the pretrained motion diffusion model, \\(\\mathcal{D}\\). The input includes multiple atomic motions \\(\\{M_i\\}\\) and a text description \\(T_{in}\\). PersonaBooth generates personalized motion that reflects the description. We use the same motion representation as HumanML3D [6], \\(M_i \\in \\mathbb{R}^{f \\cdot 263}\\), where \\(f\\) is the number of frames, and 263 includes joint rotations, positions, velocities, and foot contact. First, the Persona Extractor \\(\\mathcal{E}\\) processes each input motion to extract a visual persona feature \\(V^*\\) and a persona token \\(P^*\\). Built on a pretrained motion clip model [22] (\\(\\mathcal{E}_{clip}\\)), which captures the general characteristics of motion sequences, \\(\\mathcal{E}\\) includes a transformer structure (\\(\\mathcal{E}_P\\)) specially designed to extract persona features. We applied a contrastive learning scheme to enhance the persona extraction capability of \\(\\mathcal{E}_P\\), which will be detailed in Sec. 3.1. The extracted \\(V^*\\) is then fed into a newly added adaptive layer in \\(\\mathcal{D}\\). Meanwhile, the Personalized Text Encoder \\(\\mathcal{X}\\) adaptively incorporates \\(P^*\\) to pro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.577 + ], + "angle": 0, + "content": "duce a personalized text feature \\( T^{*} \\). \\( T^{*} \\) is input into \\( \\mathcal{D} \\), where \\( \\mathcal{D} \\) ultimately generates the personalized motion. During training, a single motion-text pair is provided to enable reconstruction. During inference, given motion sets \\( \\{M_i\\} \\), the Context-Aware Fusion module generates weights to perform a weighted combination of the elements within \\( \\{V_{i}^{*}\\} \\) and \\( \\{P_i^*\\} \\). PersonaBooth explores efficient training by using a classifier-free guidance approach. During training, \\( \\mathcal{E}_{clip} \\) and \\( \\mathcal{D} \\), excluding the newly added adaptive layer, are kept frozen. Sec.3.1 to Sec.3.3 describe the process based on the training of single motion, \\( M \\). The inference process for multiple input motions \\( \\{M_i\\} \\) is explained in Sec.3.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.585, + 0.692, + 0.6 + ], + "angle": 0, + "content": "3.1. Persona Extractor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.907, + 0.728 + ], + "angle": 0, + "content": "Persona Extractor \\(\\mathcal{E}\\) extracts the persona token \\(P^{*}\\) along with the visual persona feature \\(V^{*}\\). We used the TMR [22] structure for \\(\\mathcal{E}_{clip}\\), replacing their text encoder with the CLIP text encoder [26] and retraining it. The CLIP text encoder is the same one used in the Personalized Text Encoder \\(\\mathcal{X}\\), and therefore, this replacement enables \\(P^{*}\\) to share the same token embedding space as the input prompt. For input motion \\(M\\), persona features \\(V^{*}\\) and \\(P^{*}\\) are extracted as" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.736, + 0.905, + 0.754 + ], + "angle": 0, + "content": "\\[\nV ^ {*} = \\mathcal {E} ([ c l s ], M ]), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.755, + 0.905, + 0.772 + ], + "angle": 0, + "content": "\\[\nP ^ {*} = \\operatorname {M L P} (Y), \\text {w h e r e} Y = V ^ {*} [ 0 ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.78, + 0.738, + 0.796 + ], + "angle": 0, + "content": "where \\([cls]\\) denotes a class token." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To ensure the coherence of persona features across different motion contents of the same persona, we introduce the persona cohesion loss \\( L_{pc} \\). We utilize a supervised contrastive learning scheme [10], where motion samples from the same persona are encouraged to be closer in the persona feature space, while motion samples from different personas are pushed apart. This helps to guide the persona fea" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22758" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.485, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.255, + 0.483, + 0.297 + ], + "angle": 0, + "content": "Figure 3. Textual and visual adaptation. (a) Personalized Text Encoder, \\(\\mathcal{X}\\). (b) \\(t\\)-th step of the Motion Diffusion, \\(\\mathcal{D}\\). Learnable parameters are denoted by the fire icon" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.484, + 0.398 + ], + "angle": 0, + "content": "tures to form well-defined clusters for each persona. For a batch containing \\(N\\) input motion data, \\(L_{pc}\\) is applied to a total of \\(2N\\) data points, with one positive sample for each input motion. The positive sample is randomly selected from the same persona group as the input motion. \\(L_{pc}\\) for a positive pair of index \\((i,j)\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.407, + 0.483, + 0.445 + ], + "angle": 0, + "content": "\\[\nL _ {p c} = - \\log \\frac {\\exp (\\mathrm {s i m} (h (Y _ {i}) , h (Y _ {j})) / \\tau)}{\\sum_ {k = 1} ^ {2 N} \\mathbb {1} _ {[ k \\neq i ]} \\exp (\\mathrm {s i m} (h (Y _ {i}) , h (Y _ {k})) / \\tau)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.515 + ], + "angle": 0, + "content": "where \\( h(\\cdot) \\) denotes projection head. \\( Y_{j} \\) is extracted from \\( M_{j} \\) which belongs to the same persona group as \\( M \\). \\( \\mathrm{sim}(u,v) \\) indicates the cosine similarity between two vectors \\( u \\) and \\( v \\). \\( \\tau \\) denotes the temperature parameter." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.364, + 0.54 + ], + "angle": 0, + "content": "3.2. Textual and Visual Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.484, + 0.636 + ], + "angle": 0, + "content": "The discrepancy between our PerMo dataset and the non-persona-specific pretraining dataset, HumanML3D [6] necessitates a carefully designed adaptation mechanism for effective finetuning. To address this, we propose two adaptation paths—textual and visual—to seamlessly integrate the two persona features, \\( P^{*} \\) and \\( V^{*} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.484, + 0.848 + ], + "angle": 0, + "content": "The structure of the Personalized Text Encoder \\((\\mathcal{X})\\) is shown in Fig. 3 (a). Given an input sentence like \"A person is dancing\" \\((T_{in})\\), we introduce a personalized sentence \"[P*] person is dancing\" \\((\\tilde{T}_{in})\\), where the word \\([P^*]\\) precedes the subject. Subsequently, the vector \\(P^*\\), derived from \\(\\mathcal{E}\\), replaces the token embedding corresponding to \\([P^*]\\). Then, the modified sentence embedding, \\(\\mathcal{X}_{clip}(T_{in})\\), is adaptively merged with the original sentence embedding, \\(\\mathcal{X}_{clip}(T_{in})\\). The descriptions in HumanML3D are generally structured with subjects like 'A Person,' or 'Someone' and lack adjectives to characterize the subject. Thus, a smoother adaptation is necessary to enable the pretrained model to handle variations in sentence structure effectively. Specifically, personalized text feature \\(T^*\\) is derived as" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.859, + 0.483, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} T ^ {*} = \\mathcal {X} \\left(T _ {i n}, P ^ {*}\\right) \\tag {4} \\\\ = \\mathcal {X} _ {c l i p} (T _ {i n}) + s _ {t} \\cdot \\mathrm {t a n h} (\\gamma_ {t}) \\cdot \\mathcal {X} _ {c l i p} (\\tilde {T} _ {i n}, P ^ {*}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "where \\(\\gamma_{t}\\) is a learnable parameter initialized as zero which is for zero gating [36]. \\(s_t\\) is a scaling factor [31], a constant used to balance the adaptive layer during inference. CLIP text encoder \\(\\mathcal{X}_{clip}\\) is frozen during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.243 + ], + "angle": 0, + "content": "For visual adaptation, we introduce a single adaptive layer into the transformer structure, following the approach used in [31] for the UNet. As illustrated in Fig. 3 (b), for \\( b \\)-th transformer block in \\( t \\)-th step of the diffusion, the adaptive layer is placed between the original self-attention layer and the feedforward layer as" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.254, + 0.907, + 0.273 + ], + "angle": 0, + "content": "\\[\nz ^ {\\prime} = z + s _ {v} \\cdot \\tanh (\\gamma_ {v}) \\cdot \\mathrm {A d a p t} ([ z, V ^ {*} ]), \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.281, + 0.906, + 0.387 + ], + "angle": 0, + "content": "where Adapt indicates the adaptive layer and \\( z \\) and \\( z' \\) denote the input and output features of the adaptive layer, respectively. \\( \\gamma_v \\) and \\( s_v \\) serve as the gating parameter and scaling factor. We evaluated several adaptive layer structures, including self-attention, cross-attention, and the AdaIN layer, which is widely used in style transfer [1, 12]. Among these, self-attention was found to be the most effective." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.396, + 0.618, + 0.412 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.906, + 0.48 + ], + "angle": 0, + "content": "We used a pretrained 50-step MDM [33] for \\(\\mathcal{D}\\). Diffusion is modeled as a Markov noising process \\(\\{M^t\\}_{t=0}^T\\), where \\(M^0\\) is sampled from the training motion, and \\(t\\) denotes the noise time-step. The diffusion loss for finetuning is given by" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.488, + 0.907, + 0.521 + ], + "angle": 0, + "content": "\\[\nL _ {D} := \\mathbb {E} _ {M ^ {0}, t, T} \\left[ \\left\\| M ^ {0} - \\mathcal {D} \\left(M ^ {t}, t, V ^ {*}, T ^ {*}\\right) \\right\\| _ {2} ^ {2} \\right] + L _ {\\text {g e o}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.53, + 0.906, + 0.561 + ], + "angle": 0, + "content": "where \\(L_{geo}\\) is the geometric loss used during the MDM pretraining [33]. The final training loss is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.572, + 0.905, + 0.589 + ], + "angle": 0, + "content": "\\[\nL = L _ {D} + \\lambda L _ {p c}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.599, + 0.726, + 0.614 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a weight for the \\(L_{pc}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.75 + ], + "angle": 0, + "content": "We utilize Classifier-Free Guidance (CFG) [8], which is a technique for adjusting the trade-off between the diversity and fidelity of generated samples. Similarly, we applied CFG during finetuning to enable control over the trade-off between the pretrained diffusion model's capability for generating diverse motions and fidelity to the newly provided conditions, \\( V^{*} \\) and \\( T^{*} \\). During training, \\( V^{*} \\) and \\( T^{*} \\) are randomly dropped with a \\( 10\\% \\) probability, respectively. During sampling, the output of \\( t \\)-th step is extrapolated as" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.759, + 0.907, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {D}} \\left(M ^ {t}, t, V ^ {*}, T ^ {*}\\right) = b \\mathcal {D} _ {T} + (1 - b) \\mathcal {D} _ {V}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.78, + 0.906, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {T} = \\mathcal {D} (M ^ {t}, t, V ^ {*}, \\emptyset) + g _ {t} (\\mathcal {D} (M ^ {t}, t, V ^ {*}, T ^ {*}) - \\mathcal {D} (M ^ {t}, t, V ^ {*}, \\emptyset)),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.799, + 0.906, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {V} = \\mathcal {D} (M ^ {t}, t, \\emptyset , T ^ {*}) + g _ {v} (\\mathcal {D} (M ^ {t}, t, V ^ {*}, T ^ {*}) - \\mathcal {D} (M ^ {t}, t, \\emptyset , T ^ {*})),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(\\varnothing\\) indicates the absence of an input. \\(g_{t}\\) and \\(g_{v}\\) represent guidance scales, while \\(b\\) serves as the balancing factor between the modalities. As \\(g_{v}\\) increases, the model puts more focus on the persona inherent in the visual feature. However, if \\(g_{v}\\) is set too high, inconsistencies with the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22759" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.356, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.22, + 0.234, + 0.23 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.089, + 0.419, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.089, + 0.482, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.22, + 0.43, + 0.23 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.231, + 0.482, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.314, + 0.297, + 0.325 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.382 + ], + "angle": 0, + "content": "Figure 4. (a) Motion capture studio and examples of data formats: skeleton, markers, and mesh. (b) Unique persona expressions of each actor in the 'Childish' category. (c) Rendered mesh for each actor in the 'Fearful' category" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.469 + ], + "angle": 0, + "content": "text prompt may arise. On the other hand, \\( g_{t} \\) controls the model's adherence to the prompt, with the intensity of \\( P^{*} \\) in the text being adjusted by \\( s_t \\) in Eq. (4). The hyperparameter settings are specified in Sec. 5, and an ablation study for these is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.481, + 0.451, + 0.497 + ], + "angle": 0, + "content": "3.4. Context-Aware Fusion for Multiple Inputs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Intuitively, if multiple motions are input, these additional cues can help with generating motion. However, it's also crucial to choose the right cues, as transferring persona from motions with very different content types, even with \\( L_{pc} \\), can result in unnatural movements. A straightforward approach might be to take the mean of all features, as suggested in [31]. In the motion domain, however, taking all motions can result in blended motions that appear implausible. To overcome this, we introduce a Context-Aware Fusion (CAF) method that prioritizes input motions based on their contextual relevance to the input prompt. To assess this relevance, we use the motion encoder and text encoder of the motion clip (\\(\\mathcal{E}\\)clip and \\(\\mathcal{X}\\)clip), which align the text and motion feature spaces. The CAF is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.729, + 0.483, + 0.746 + ], + "angle": 0, + "content": "\\[\nI _ {\\text {T o p - k}} = \\operatorname {a r g m a x} _ {i} \\left(S _ {i}\\right), \\quad S _ {i} = \\sin \\left(\\mathcal {E} _ {\\text {c l i p}} \\left(M _ {i}\\right), \\bar {T} ^ {*}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.749, + 0.482, + 0.799 + ], + "angle": 0, + "content": "\\[\nw _ {i} = \\left\\{ \\begin{array}{l l} \\frac {\\exp (S _ {i})}{\\sum_ {n} \\exp (S _ {n})}, & \\text {i f} i \\in I _ {\\mathrm {T o p - k}}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.814, + 0.484, + 0.86 + ], + "angle": 0, + "content": "where \\(\\bar{T}^*\\) represents the personalized text feature, using the average value of \\(\\{P^{*}\\}\\) as a persona token. \\(\\mathrm{sim}(\\cdot)\\) denotes cosine similarity. Consequently," + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.873, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\nV ^ {*} = \\sum_ {i} w _ {i} V _ {i} ^ {*}, \\quad P ^ {*} = \\sum_ {i} w _ {i} P _ {i} ^ {*}. \\tag {11}\n\\]" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.906, + 0.117 + ], + "angle": 0, + "content": "Table 2. Comparison with existing motion style datasets. The note for \\* is provided in the main text. - indicates no information" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.13, + 0.897, + 0.25 + ], + "angle": 0, + "content": "
DatasetActorsStylesContentsClipsMeshText
Xia [34]-86572XX
BFA [1]116932XX
BN-1 [13]2*1517175XX
BN-2 [13]2*7102,902XX
100Style [19]1100*8*810XX
PerMo(ours)534106,610✓✓
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.267, + 0.665, + 0.282 + ], + "angle": 0, + "content": "4. PerMo Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.906, + 0.414 + ], + "angle": 0, + "content": "We collected a large-scale PerMo dataset, capturing personas from multiple actors. To ensure variety, we hired five professional motion capture actors of diverse genders and body types. Each actor is assigned to perform 34 styles, categorized into Age, Character, Condition, Emotion, Traits, and Surroundings, resulting in a total of 170 personas. Each actor performed 10 distinct contents for every style, carefully selected to engage different body parts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.906, + 0.688 + ], + "angle": 0, + "content": "The dataset was captured in a studio equipped with 33 OptiTrack cameras, using 41 optical markers per person. For compatibility with other motion datasets, we converted the marker data to SMPL-H [29] format using \\(\\mathrm{Mosh}++\\) [18]. All data underwent expert cleaning and verification, with the entire process adhering to strict quality guidelines. Fig. 4 (a) illustrates the motion capture studio and the provided data formats, while (b) highlights the distinct personas expressed by each actor. Additionally, we constructed 20-30 varied descriptions for each content type, following research indicating that high-quality finetuning descriptions improve prompt-based control [7]. Starting with one detailed description, we instructed ChatGPT to generate a range from high-level descriptions (e.g., \"A person is jumping ahead\") to low-level descriptions (e.g., \"A person pushes off the ground and jumps forward several times\"). Additional details on data categories, capture process, data structure, and examples can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Table 2 shows a comparison with existing motion style datasets. Notably, PerMo is the first dataset to collect data from multiple actors. Our dataset also offers the highest number of total clips and content categories among trainable datasets and is the only one that includes mesh, marker data, and descriptions. The BN datasets [13] involve two actors; however, for the same category, only one actor is used (BN-1), or there are no tags identifying the actors (BN-2). BN-1 is unsuitable for training, as it contains only a single motion sequence per category. The 100Style [19] offers the largest variety of styles, but 58 of its categories focus on content, rather than on style [38]. Additionally, the eight content categories in this dataset mainly consist of walking or running, offering limited variation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22760" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.911, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.251, + 0.908, + 0.294 + ], + "angle": 0, + "content": "Figure 5. Example of the ablation study. The input motions are from the 'Uppity' of Actor 1. The input prompt is \"A person walks in a circle.\" In (a) and (b), only \\( M_1 \\) is provided for the input, while both \\( M_1 \\) and \\( M_2 \\) are provided for (c) and (d). \\( L_{pc} \\) encourages the generated motion to closely follow the prompt, while CAF prevents the motion from blending. We set \\( k = 1 \\) for CAF" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.318, + 0.226, + 0.336 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.494 + ], + "angle": 0, + "content": "Implementation Details. We conduct experiments using the proposed PerMo and 100Style [19] datasets. During training, a single motion is cropped differently to serve as both the input and the GT motion for generation. We set \\(\\lambda\\) to \\(10^{-2}\\) and use the AdamW [16] optimizer with a learning rate of \\(10^{-4}\\). Training runs for 500 epochs with a batch size of 64. We set \\(s_t\\) and \\(s_v\\) to 1 during training and to 0.3 during inference. \\(g_t\\) is set to 10, and \\(g_v\\) to 15 and 10 for PerMo and 100Style, respectively. We set \\(b\\) to 0.7 for single-input and 0.5 for multi-input settings, and \\(k\\) to 5 for CAF." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.631 + ], + "angle": 0, + "content": "Evaluation Settings. For quantitative evaluation, we use descriptions from the HumanML3D test set as prompts and motion samples from the PerMo and 100Style datasets as inputs. For each description, \\([P^*]\\) is logically placed as a modifier for the subject. We used two evaluation settings: Single Input (SI) and Multiple Input (MI). In SI, one motion is sampled from the entire dataset for comparison with MST methods. In MI, a persona set is first sampled, and then \\(|M_i|\\) motions are drawn from it." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.484, + 0.812 + ], + "angle": 0, + "content": "Performance is evaluated using Frechet Inception Distance (FID), R-Precision, and Diversity—metrics commonly used in motion generation tasks [3, 5, 6, 33]—as well as Persona Recognition Accuracy (PRA), a metric adopted by Style Recognition Accuracy (SRA) in MST [9, 32]. FID assesses the overall quality and realism of generated motions and is our primary metric [5]. R-Precision evaluates text-to-motion alignment accuracy [6], and Diversity [5] measures how well the generated motions reflect a broad distribution of personas. PRA uses a pretrained persona classifier to evaluate the persona consistency of generated motions. The details are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.245, + 0.835 + ], + "angle": 0, + "content": "5.1. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Table 3 demonstrates the effects of our key contributions. The baseline represents results obtained using only visual adaptation. When we introduce \\( P^* \\) for textual adaptation, the FID score drops significantly, and both PRA and Diver" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.318, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Table 3. Ablation study of the proposed components. \\( |M_{i}| = \\max \\) indicates that all motions in the chosen persona set are input" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.357, + 0.907, + 0.503 + ], + "angle": 0, + "content": "
MethodsFID ↓R Precision ↑PRA avg.↑Diversity ↑
Top 1Top 2Top 3
Single Input (SI) Setting
baseline7.450.060.110.1617.997.48
+ P*5.060.050.100.1518.268.01
+Lpc3.180.150.260.3318.057.74
Multiple Input (MI) Setting
Mi|=14.270.080.150.2016.207.71
Mi|=max, [31]3.520.190.300.3819.247.88
Mi|=max, CAF2.950.190.300.3918.138.12
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.535, + 0.907, + 0.672 + ], + "angle": 0, + "content": "sity increase. This indicates that multimodal integration of persona information not only strengthens the reflection of persona but also supports the generation of plausible motion. Furthermore, adding \\( L_{pc} \\) leads to an additional reduction in FID and a notable increase in R-precision. This implies that \\( L_{pc} \\) helps the Persona Extractor effectively capture persona essence disentangled from the input content, resulting in generated motion that aligns well with the prompt without compromising content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Next, the effectiveness of CAF is validated in the MI setting. When multiple input motions are used \\( (|M_i| = \\max) \\), all metrics improve compared to using a single input \\( (|M_i| = 1) \\) due to the higher likelihood of finding motion candidates that align with the prompt. Additionally, even in cases with multiple inputs, selecting the contextually most relevant Top-k motions—as done by CAF—is shown to be crucial. For \\( |M_i| = \\max \\), [31] presents a baseline approach that uses the average feature of all inputs. Introducing the proposed CAF reduces FID while increasing Diversity, indicating that the generated motions are sufficiently diverse and plausible rather than oversimplified by averaging. Although the PRA score decreases, we found through qualitative results that generating plausible motion is a more crucial factor in improving motion quality." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22761" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.24, + 0.908, + 0.285 + ], + "angle": 0, + "content": "Figure 6. Example reflecting the different personas of Actors 1-5 for the same 'Childish' category. The input prompt is \"A person is running and bending to pick something.\" Red arrows highlight the frames that clearly show distinct movements. The input motion of each actor can be referenced in Fig. 4 (b)" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.307, + 0.482, + 0.336 + ], + "angle": 0, + "content": "Table 4. Ablation study regarding adaptation. PersonaBooth (SI) indicates our complete model for single inputs" + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.346, + 0.485, + 0.436 + ], + "angle": 0, + "content": "
MethodsFID ↓R Precision ↑PRA avg.↑Diversity ↑
Top 1Top 2Top 3
AdaIN8.770.040.070.1015.107.33
Cross-Attn8.960.040.070.1016.377.62
w/o T adapt.3.610.140.230.3117.777.88
PersonaBooth (SI)3.180.150.260.3318.057.74
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Fig. 5 provides an example. When using only \\( M_{1} \\) as the input motion, if \\( L_{pc} \\) is not applied, the generated motion fails to follow the path provided in the prompt, as shown in (a). However, with \\( L_{pc} \\), the motion better aligns with the prompt's trajectory, though it still includes unintended hand-waving motions unrelated to the prompt. When using multiple input motions, including \\( M_{2} \\) (a walking motion with arms crossed), taking the mean of features results in a blended motion, as shown in (c). This blending causes one arm to bend while the other hangs low, creating unnatural poses. In contrast, with CAF, \\( M_{2} \\) receives a higher weight, resulting in a more natural walking motion with arms crossed, as shown in (d)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.657, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Table 4 presents various ablation results related to adaptation. For the visual adaptation layer, Self-Attention proved to be the most effective compared to AdaIN and CrossAttn. Additionally, introducing text adaptation at Personalized Text Encoder improves the PRA, FID, and R-precision metrics. Fig. 6 demonstrates how PersonaBooth effectively captures and represents each actor's unique action. The red arrows specifically highlight the distinct arm movements of each persona. Additional details, including variations in body tilt and movement speed, are more clearly observable in the supplementary video materials." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.833, + 0.458, + 0.85 + ], + "angle": 0, + "content": "5.2. Comparison with State-of-the-Art methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We compare our PersonaBooth with the state-of-the-art diffusion-based MST approaches, MoMo [25] and MCM-LDM [32], on the PerMo dataset. Additionally, we evalu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.907, + 0.415 + ], + "angle": 0, + "content": "ate it on an existing style dataset, 100Style [19]. As MCM-LDM does not support prompt-based generation, we used MDM-generated motions [33] as target motions for evaluation. MCM-LDM was evaluated using the original model trained on HumanML3D, as well as finetuned models, referred to as MCM-LDM, which were trained on the PerMo and 100Style datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.421, + 0.665, + 0.434 + ], + "angle": 0, + "content": "5.2.1. PerMo Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.635 + ], + "angle": 0, + "content": "The comparative results for the PerMo dataset are shown in Table 5, with all subcategory results displayed for PRA. The single-input PersonaBooth significantly reduces FID compared to existing methods, indicating that while previous studies struggled with PerMo—a challenging dataset containing more diverse content than 100Style—PersonaBooth performs robust in generating high-quality motion. Additionally, PersonaBooth achieves the highest performance in R-Precision, PRA, and Diversity compared to existing methods. Applying multiple inputs and CAF further enhances performance. It's noteworthy that while MoMo and MCM-LDM utilized a 100-step and 1000-step diffusion model, respectively, we used a 50-step diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.909, + 0.788 + ], + "angle": 0, + "content": "Fig. 7 shows qualitative results. MoMo [25] notes that their generated motions often follow the input motion's direction instead of the target's, which they attempted to copy the root rotation from the input motion (shown in (a) before and (b) after adjustment). The prompts for \"turns in the air\" were not met in (a), (b), or (c). Particularly, (c) does not reflect the style of ballerina. In contrast, PersonaBooth accurately generates the motion by first performing a hop that reflects the style, followed by a precise turn in the air. Example videos are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.792, + 0.676, + 0.806 + ], + "angle": 0, + "content": "5.2.2. 100Style Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Table 6 presents the comparative results on the 100Style dataset. MoMo demonstrates high PRA and Diversity, indicating effective style reflection. However, it shows very low R-precision, suggesting poor alignment between generated motions and prompts. Finetuned MCM-LDM improved performance on nearly all metrics compared to its" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22762" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 5. Comparison with the state-of-the-art methods on the PerMo dataset. The comparison is made in the Single Input (SI) setting as the existing methods do not support multiple inputs. MCM-LDM* indicates the model is finetuned on the PerMo dataset." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.131, + 0.907, + 0.239 + ], + "angle": 0, + "content": "
MethodsStepsFID ↓R Precision ↑PRA ↑Diversity ↑
Top 1Top 2Top 3AgeChar1Char2Cond1Cond2Emo1Emo2TraitSurAvg.
MoMo [25]10013.910.050.090.1328.2310.2311.308.7315.979.9714.5712.607.9313.476.74
MCM-LDM [32]10009.160.130.230.3047.3811.4410.347.6318.9011.6917.2015.719.2217.006.70
MCM-LDM* [32]10009.700.130.230.3046.4912.129.227.9519.7711.7518.0515.187.3316.766.76
PersonaBooth (SI)503.180.150.260.3348.0013.6711.976.6918.6714.3419.7517.068.8018.057.74
PersonaBooth (MI)502.950.190.300.3951.9514.8912.005.1622.959.4418.1415.4510.0918.138.12
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.244, + 0.265, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.347, + 0.216, + 0.367 + ], + "angle": 0, + "content": "Input Motion (Actor1 Ballerina)" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.248, + 0.426, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.356, + 0.417, + 0.367 + ], + "angle": 0, + "content": "(a) MoMo (w/o copy root rot.)" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.251, + 0.597, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.356, + 0.537, + 0.367 + ], + "angle": 0, + "content": "(b) MoMo" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.248, + 0.758, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.356, + 0.712, + 0.367 + ], + "angle": 0, + "content": "(c) MCM-LDM*" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.248, + 0.897, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.356, + 0.89, + 0.367 + ], + "angle": 0, + "content": "(d) PersonaBooth (SI)" + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.375, + 0.845, + 0.39 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison on PerMo dataset. The input prompt is \"A person hops forward and turns in the air.\"" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.394, + 0.255, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.491, + 0.209, + 0.511 + ], + "angle": 0, + "content": "Input Motion (Airplane)" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.392, + 0.429, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.499, + 0.417, + 0.511 + ], + "angle": 0, + "content": "(a) MoMo (w/o copy root rot.)" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.406, + 0.598, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.499, + 0.537, + 0.51 + ], + "angle": 0, + "content": "(b) MoMo" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.393, + 0.734, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.499, + 0.711, + 0.51 + ], + "angle": 0, + "content": "(c) MCM-LDM*" + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.393, + 0.906, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.499, + 0.89, + 0.511 + ], + "angle": 0, + "content": "(d) PersonaBooth (SI)" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.519, + 0.882, + 0.534 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison on 100Style dataset. The input prompt is \"A person walks backward and sits down on the chair.\"" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.592 + ], + "angle": 0, + "content": "Table 6. Comparison with the state-of-the-art methods on the 100Style dataset. MCM-LDM* indicates the model is finetuned on the 100Style dataset" + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.604, + 0.484, + 0.698 + ], + "angle": 0, + "content": "
MethodsFID ↓R Precision ↑SRA ↑Diversity ↑
Top 1Top 2Top 3
MoMo [25]5.970.070.100.1460.577.82
MCM-LDM [32]7.200.180.290.3754.446.76
MCM-LDM* [32]6.530.170.280.3754.617.12
PersonaBooth (SI)3.270.200.310.4064.537.90
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.718, + 0.483, + 0.824 + ], + "angle": 0, + "content": "original model. Although finetuned MCM-LDM achieved high R-precision, it also had a high FID, resulting in unnatural motion generation, with low PRA and Diversity indicating weaker style reflection. In contrast, our single-input PersonaBooth significantly reduces FID and achieves top performance across all metrics, including R-precision, PRA, and Diversity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In Fig. 8 (a) and (b), MoMo produces sliding motions that don't align with the prompt. MCM-LDM shows sideward movements. In contrast, PersonaBooth accurately reflects both the input motion's style, like arm extension and leaning, and the sitting motion in line with the prompt." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.551, + 0.634, + 0.567 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.907, + 0.697 + ], + "angle": 0, + "content": "We propose the PerMo dataset and the PersonaBooth framework for Motion Personalization. Through contrastive learning, we effectively capture the essence of a persona. Furthermore, we enhance the integration of persona traits into diffusion models using a multi-modal finetuning method. We also introduce the CAF, which effectively handles multiple inputs, ensuring cohesive representation of personalized motion characteristics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.909, + 0.849 + ], + "angle": 0, + "content": "Limitations and Future Works. The PerMo contains relatively short motion sequences, which sometimes lead to stationary poses at the end of generated motions of the same length. However, this issue does not occur when prompted with longer actions, suggesting that automatic adjustment of motion length remains a task for future work. In addition, we plan to investigate approaches for applying CAF to each sequential action within a prompt individually. For instance, applying CAF separately to 'run' and 'jump' in the prompt \"A person runs and then jumps.\"" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.652, + 0.866 + ], + "angle": 0, + "content": "Acknowledgement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.867, + 0.906, + 0.898 + ], + "angle": 0, + "content": "This work was supported by MSIT, Korea, under [RS-2022-00164860] and [RS-2024-00418641] supervised by IITP." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22763" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. ACM Transactions on Graphics (TOG), 39(4):64-1, 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.313 + ], + "angle": 0, + "content": "[3] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.484, + 0.383 + ], + "angle": 0, + "content": "[4] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[5] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3D human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.484, + 0.527 + ], + "angle": 0, + "content": "[6] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.529, + 0.484, + 0.584 + ], + "angle": 0, + "content": "[7] Xingzhe He, Zhiwen Cao, Nicholas Kolkin, Lantao Yu, Kun Wan, Helge Rhodin, and Ratheesh Kalarot. A data perspective on enhanced identity preservation for diffusion personalization. arXiv preprint arXiv:2311.04315, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.67 + ], + "angle": 0, + "content": "[9] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[10] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[11] Boeun Kim, Hyung Jin Chang, Jungho Kim, and Jin Young Choi. Global-local motion transformer for unsupervised skeleton-based action learning. In European Conference on Computer Vision, pages 209-225. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Boeun Kim, Jungho Kim, Hyung Jin Chang, and Jin Young Choi. Most: Motion style transformer between diverse action contents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1705-1714, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Makito Kobayashi, Chen-Chieh Liao, Keito Inoue, Sentaro Yojima, and Masafumi Takahashi. Motion capture dataset" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "for practical use of ai-based motion editing and stylization. arXiv preprint arXiv:2306.08861, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[14] Jehee Lee, Jinxiang Chai, Paul SA Reitsma, Jessica K Hodgins, and Nancy S Pollard. Interactive control of avatars animated with human motion data. In Proceedings of the 29th annual Conference on Computer Graphics and Interactive Techniques, pages 491-500, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[16] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[17] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[19] Ian Mason, Sebastian Starke, and Taku Komura. Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 5(1):1-18, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[20] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on Computer Graphics and Interactive Techniques, 4(3):1-17, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[21] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[22] Mathis Petrovich, Michael J Black, and Gül Varol. Tmr: Text-to-motion retrieval using contrastive 3D human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9488-9497, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[23] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[24] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[25] Sigal Raab, Inbar Gat, Nathan Sala, Guy Tevet, Rotem Shalev-Arkushin, Ohad Fried, Amit H Bermano, and Daniel Cohen-Or. Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. arXiv preprint arXiv:2406.06508, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22764" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[29] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[31] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[32] Wenfeng Song, Xingliang Jin, Shuai Li, Chenglizhao Chen, Aimin Hao, Xia Hou, Ning Li, and Hong Qin. Arbitrary motion style transfer with multi-condition motion latent diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821-830, 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.502, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[33] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[34] Shihong Xia, Congyi Wang, Jinxiang Chai, and Jessica Hodgins. Realtime style transfer for unlabeled heterogeneous human motion. ACM Transactions on Graphics (TOG), 34 (4):1-10, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[35] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.482, + 0.739 + ], + "angle": 0, + "content": "[36] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.742, + 0.482, + 0.797 + ], + "angle": 0, + "content": "[37] Xulu Zhang, Xiao-Yong Wei, Wengyu Zhang, Jinlin Wu, Zhaoxiang Zhang, Zhen Lei, and Qing Li. A survey on personalized content synthesis with diffusion models. arXiv preprint arXiv:2405.05538, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.799, + 0.482, + 0.84 + ], + "angle": 0, + "content": "[38] Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, and Huaizu Jiang. Smoodi: Stylized motion diffusion model. arXiv preprint arXiv:2407.12783, 2024. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22765" + } + ] +] \ No newline at end of file diff --git a/2025/PersonaBooth_ Personalized Text-to-Motion Generation/62946df9-a553-4225-bd84-84b1acdb1a5f_origin.pdf b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/62946df9-a553-4225-bd84-84b1acdb1a5f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a3fb4e2af4f7ecf63c05e0d850763ffe53164c0a --- /dev/null +++ b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/62946df9-a553-4225-bd84-84b1acdb1a5f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cbe6889da9b404b95ab3eceba33fa82a3fbdb978f0948b07c016df331b2d8c3 +size 9517908 diff --git a/2025/PersonaBooth_ Personalized Text-to-Motion Generation/full.md b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..367445db47570422c21da563fd0f86a55fa53f6f --- /dev/null +++ b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/full.md @@ -0,0 +1,327 @@ +# PersonaBooth: Personalized Text-to-Motion Generation + +Boeun Kim1,2,3 Hea In Jeong2 JungHoon Sung3 Yihua Cheng1 Jeongmin Lee2 +Ju Yong Chang4 Sang-II Choi3 Younggeun Choi3 Saim Shin2 Jungho Kim2 Hyung Jin Chang1 +1University of Birmingham 2Korea Electronics Technology Institute 3Dankook University 4Kwangwoon University +http://boeun-kim.github.io/page-PersonaBooth + +# Abstract + +This paper introduces Motion Personalization, a new task that generates personalized motions aligned with text descriptions using several basic motions containing Persona. To support this novel task, we introduce a new large-scale motion dataset called PerMo (PersonaMotion), which captures the unique personas of multiple actors. We also propose a multi-modal finetuning method of a pretrained motion diffusion model called PersonaBooth. PersonaBooth addresses two main challenges: i) A significant distribution gap between the persona-focused PerMo dataset and the pretraining datasets, which lack persona-specific data, and ii) the difficulty of capturing a consistent persona from the motions vary in content (action type). To tackle the dataset distribution gap, we introduce a persona token to accept new persona features and perform multi-modal adaptation for both text and visuals during finetuning. To capture a consistent persona, we incorporate a contrastive learning technique to enhance intra-cohesion among samples with the same persona. Furthermore, we introduce a context-aware fusion mechanism to maximize the integration of persona cues from multiple input motions. PersonaBooth outperforms state-of-the-art motion style transfer methods, establishing a new benchmark for motion personalization. + +# 1. Introduction + +Imagine a scenario where, by recording just a few basic movements of yourself, an avatar in a virtual world can mirror your personal traits and unique motion style. This kind of 'motion personalization' enables realistic interactions in virtual spaces such as games and the metaverse [14]. Moreover, if avatar motions could be directed through text, creating video content would become easy without the need for real-life stunt actors [2, 24]. + +In this paper, we propose a task called Motion Personalization, which aims to generate text-based motions that + +![](images/0bd3c3c0f3a60c18c3243529198756082868ca673432a91f682450903be322da.jpg) +Figure 1. Motion Personalization generates text-driven, personalized motions based on persona embedded in atomic input motions. We propose a new framework, PersonaBooth, along with a new benchmark dataset, PerMo, for Motion Personalization + +reflect individual personas using a few key atomic movements, such as jumping, punching, or walking. We define a persona as the unique style expression of an individual. As illustrated in Fig. 1, this task aims to generate a broad range of realistic motions robustly. To facilitate this, we introduce a model called PersonaBooth and a new benchmark, PerMo (PersonaMotion), specifically designed for motion personalization. Table 1 provides a comparison of similar motion generation tasks. Text-to-Motion Generation approaches [6, 21] generate motion using only text as input, without any reference motion, and numerous diffusion-based methods [3, 33, 35] have been proposed for this task. Motion Style Transfer (MST) focuses on transferring style + +Table 1. Comparison of motion generation tasks + +
Motion TaskExisting DatasetsSource MotionsMulti-modal Finetuning
T2M Generation [3, 33, 35][6] [23]NoneX
Style Transfer [25, 32, 38][34][1][19]SingleX
Personalization (Ours)NoneMultiple✓
+ +from a single source motion. While several datasets and methods are available for MST [25, 32, 34, 38], the use of a single source motion restricts the ability to generate the broad and diverse motions seen in real life [12]. The proposed Motion Personalization has no existing prior approaches or datasets. Furthermore, our proposed method introduces multi-modal finetuning—a process neglected in existing diffusion-based MST methods. + +PersonaBooth is a multi-modal finetuning method for pretrained text-to-motion diffusion models. The challenge in finetuning diffusion models for the motion domain, compared to the image domain [4, 15, 17, 30, 31, 37], lies in the relatively lower diversity of pretraining data. The most commonly used pretraining dataset, HumanML3D [6], contains only 15K samples, which is significantly smaller than the large-scale image datasets with 400M-650M samples [27, 28]. Additionally, HumanML3D includes few data reflecting personality traits, creating a substantial distribution gap when finetuning with the persona-focused PerMo dataset. MoMo [25] found that only 55 samples include styles for 'run,' 'walk,' 'jump,' and 'dance' locomotions in HumanML3D, and we could not find any persona-related samples. Existing diffusion-based MST methods [25, 38] consider only the visual features of the input motion and rely on a fixed text feature, which limits their ability to incorporate new data into the textual component. This inflexibility can lead to a phenomenon known as 'forgetting', where the pretrained model loses its initial capabilities when finetuned with new data [38]. To address this, we introduce learnable persona tokens to capture persona features from new data and propose an adaptation scheme for both text and visuals. + +Another challenge is extracting a consistent persona across different atomic motions, which manifests in diverse ways depending on the action content. For example, the elegance of a ballerina is expressed in her feet as she walks and in her hands as she waves. MST methods have faced similar challenges in disrupting the content of the target motion when transferring styles from the source motion with different content. Approaches like those of Park et al. [20], Jang et al. [9], and MoMo [25] cite this as a limitation. In contrast, MoST [12] addresses the issue with a style disentanglement loss, while MCM-LDM [32] uses the target motion's trajectory to preserve content. However, MoST's loss function requires running the model twice for two different input motions, which is computationally expensive, + +especially for diffusion models with numerous steps. Additionally, since our goal is to generate motion from text rather than from target motions, the method used by MCM-LDM is not applicable. Therefore, we introduce a novel contrastive learning-based loss specifically designed for text-to-motion diffusion, called persona cohesion loss. This loss facilitates cohesion across motion features with different content but the same persona. + +In addition, while MST uses a single input motion, we propose a new fusion method for multiple motions. In the image domain, InstantBooth [31] simply averages persona features, but this leads to unnatural blending in the motion. To address this, we introduce Context-Aware Fusion (CAF), which assigns weights to the fusion of persona features based on the similarity between the prompt and input motions. The summarized contributions of this paper are: + +- We introduce the Motion Personalization task and present PerMo, a new large-scale persona dataset captured by multiple actors. +- We propose PersonaBooth, a multi-modal finetuning method for diffusion models that introduces text adaptation through a persona token. PersonaBooth applies contrastive learning to effectively capture personas across various content types. +- PersonaBooth achieves strong performance in both Motion Personalization and MST tasks and maximizes performance through a CAF for multiple inputs. + +# 2. Related Work + +The Motion Style Transfer (MST) aims to transfer the style from a source motion to a target motion. The pioneering work by Aberman et al. [1] introduced the AdaIN layer, which simply replaces style statistics in specific layers during training. MotionPuzzle [9] proposes a Graph Convolution Network-based framework that transfers manually specified motion to a desired body part, enabling the combination of multiple source motions. MoST [12] addresses the issue of unwanted blending between different motion contents by introducing a transformer-based model [11] that more clearly separates style and content. + +Recent MST algorithms have seen notable performance gains through the use of diffusion models. MoMo [25] proposes a zero-shot motion transfer method using a pretrained diffusion model, where it mixes two input motions within the attention module of the diffusion model. However, MoMo identifies a limitation: when the source motion content differs significantly from the target motion or text, the generated motion may not follow the target accurately. For example, applying the style of a stationary source motion to a walking target motion may still result in a stationary output. SMooDi [38] proposes a finetuning method for a pretrained diffusion model. However, due to the forgetting issue, it retrained the model with both the original pretrain- + +![](images/aff20c64e70c7a75e15820d1a4cba40d49ffbd82f81d4d88601719985bee9235.jpg) +Figure 2. The overall framework of PersonaBooth. PersonaBooth has two adaptation paths—visual and text—for finetuning the Motion Diffusion model $(\mathcal{D})$ . The Persona Extractor extracts both a visual persona feature $(V^{*})$ and a persona token $(P^{*})$ from the input motions. $V^{*}$ is input into the adaptive layer of $\mathcal{D}$ , while $P^{*}$ is processed together with the input prompt through a Personalized Text Encoder, generating a personalized text feature, which is then input to $\mathcal{D}$ . The entire model is trained with a classifier-free approach, incorporating a Persona Cohesion Loss. During inference, Context-Aware Fusion is applied for multiple input cases. + +ing dataset and the style dataset. Additionally, the use of classifier-based training and the adaptation approach of creating a trainable copy of the entire model led to a training speed 10 times slower than the base diffusion model. MCM-LDM [32] addresses the challenges regarding two different contents of input motions by explicitly extracting the trajectory from the target motion and using it as a condition for the diffusion model. However, it faces a limitation: only trajectories similar to the target motion are generated, which reduces diversity. This contradicts our goal of creating diverse motions based on text prompts. + +# 3. Methodology + +We propose a method for multi-modal finetuning of a text-to-motion diffusion model, named PersonaBooth. The overall framework is shown in Fig. 2, and we aim to finetune the pretrained motion diffusion model, $\mathcal{D}$ . The input includes multiple atomic motions $\{M_i\}$ and a text description $T_{in}$ . PersonaBooth generates personalized motion that reflects the description. We use the same motion representation as HumanML3D [6], $M_i \in \mathbb{R}^{f \cdot 263}$ , where $f$ is the number of frames, and 263 includes joint rotations, positions, velocities, and foot contact. First, the Persona Extractor $\mathcal{E}$ processes each input motion to extract a visual persona feature $V^*$ and a persona token $P^*$ . Built on a pretrained motion clip model [22] ( $\mathcal{E}_{clip}$ ), which captures the general characteristics of motion sequences, $\mathcal{E}$ includes a transformer structure ( $\mathcal{E}_P$ ) specially designed to extract persona features. We applied a contrastive learning scheme to enhance the persona extraction capability of $\mathcal{E}_P$ , which will be detailed in Sec. 3.1. The extracted $V^*$ is then fed into a newly added adaptive layer in $\mathcal{D}$ . Meanwhile, the Personalized Text Encoder $\mathcal{X}$ adaptively incorporates $P^*$ to pro + +duce a personalized text feature $T^{*}$ . $T^{*}$ is input into $\mathcal{D}$ , where $\mathcal{D}$ ultimately generates the personalized motion. During training, a single motion-text pair is provided to enable reconstruction. During inference, given motion sets $\{M_i\}$ , the Context-Aware Fusion module generates weights to perform a weighted combination of the elements within $\{V_{i}^{*}\}$ and $\{P_i^*\}$ . PersonaBooth explores efficient training by using a classifier-free guidance approach. During training, $\mathcal{E}_{clip}$ and $\mathcal{D}$ , excluding the newly added adaptive layer, are kept frozen. Sec.3.1 to Sec.3.3 describe the process based on the training of single motion, $M$ . The inference process for multiple input motions $\{M_i\}$ is explained in Sec.3.4. + +# 3.1. Persona Extractor + +Persona Extractor $\mathcal{E}$ extracts the persona token $P^{*}$ along with the visual persona feature $V^{*}$ . We used the TMR [22] structure for $\mathcal{E}_{clip}$ , replacing their text encoder with the CLIP text encoder [26] and retraining it. The CLIP text encoder is the same one used in the Personalized Text Encoder $\mathcal{X}$ , and therefore, this replacement enables $P^{*}$ to share the same token embedding space as the input prompt. For input motion $M$ , persona features $V^{*}$ and $P^{*}$ are extracted as + +$$ +V ^ {*} = \mathcal {E} ([ c l s ], M ]), \tag {1} +$$ + +$$ +P ^ {*} = \operatorname {M L P} (Y), \text {w h e r e} Y = V ^ {*} [ 0 ], \tag {2} +$$ + +where $[cls]$ denotes a class token. + +To ensure the coherence of persona features across different motion contents of the same persona, we introduce the persona cohesion loss $L_{pc}$ . We utilize a supervised contrastive learning scheme [10], where motion samples from the same persona are encouraged to be closer in the persona feature space, while motion samples from different personas are pushed apart. This helps to guide the persona fea + +![](images/fbe5dc67339bbcb345b260df5695cfb0ec480fbf90ccae02e195487ad7200c6d.jpg) +Figure 3. Textual and visual adaptation. (a) Personalized Text Encoder, $\mathcal{X}$ . (b) $t$ -th step of the Motion Diffusion, $\mathcal{D}$ . Learnable parameters are denoted by the fire icon + +tures to form well-defined clusters for each persona. For a batch containing $N$ input motion data, $L_{pc}$ is applied to a total of $2N$ data points, with one positive sample for each input motion. The positive sample is randomly selected from the same persona group as the input motion. $L_{pc}$ for a positive pair of index $(i,j)$ is defined as + +$$ +L _ {p c} = - \log \frac {\exp (\mathrm {s i m} (h (Y _ {i}) , h (Y _ {j})) / \tau)}{\sum_ {k = 1} ^ {2 N} \mathbb {1} _ {[ k \neq i ]} \exp (\mathrm {s i m} (h (Y _ {i}) , h (Y _ {k})) / \tau)}, \tag {3} +$$ + +where $h(\cdot)$ denotes projection head. $Y_{j}$ is extracted from $M_{j}$ which belongs to the same persona group as $M$ . $\mathrm{sim}(u,v)$ indicates the cosine similarity between two vectors $u$ and $v$ . $\tau$ denotes the temperature parameter. + +# 3.2. Textual and Visual Adaptation + +The discrepancy between our PerMo dataset and the non-persona-specific pretraining dataset, HumanML3D [6] necessitates a carefully designed adaptation mechanism for effective finetuning. To address this, we propose two adaptation paths—textual and visual—to seamlessly integrate the two persona features, $P^{*}$ and $V^{*}$ . + +The structure of the Personalized Text Encoder $(\mathcal{X})$ is shown in Fig. 3 (a). Given an input sentence like "A person is dancing" $(T_{in})$ , we introduce a personalized sentence "[P*] person is dancing" $(\tilde{T}_{in})$ , where the word $[P^*]$ precedes the subject. Subsequently, the vector $P^*$ , derived from $\mathcal{E}$ , replaces the token embedding corresponding to $[P^*]$ . Then, the modified sentence embedding, $\mathcal{X}_{clip}(T_{in})$ , is adaptively merged with the original sentence embedding, $\mathcal{X}_{clip}(T_{in})$ . The descriptions in HumanML3D are generally structured with subjects like 'A Person,' or 'Someone' and lack adjectives to characterize the subject. Thus, a smoother adaptation is necessary to enable the pretrained model to handle variations in sentence structure effectively. Specifically, personalized text feature $T^*$ is derived as + +$$ +\begin{array}{l} T ^ {*} = \mathcal {X} \left(T _ {i n}, P ^ {*}\right) \tag {4} \\ = \mathcal {X} _ {c l i p} (T _ {i n}) + s _ {t} \cdot \mathrm {t a n h} (\gamma_ {t}) \cdot \mathcal {X} _ {c l i p} (\tilde {T} _ {i n}, P ^ {*}), \\ \end{array} +$$ + +where $\gamma_{t}$ is a learnable parameter initialized as zero which is for zero gating [36]. $s_t$ is a scaling factor [31], a constant used to balance the adaptive layer during inference. CLIP text encoder $\mathcal{X}_{clip}$ is frozen during training. + +For visual adaptation, we introduce a single adaptive layer into the transformer structure, following the approach used in [31] for the UNet. As illustrated in Fig. 3 (b), for $b$ -th transformer block in $t$ -th step of the diffusion, the adaptive layer is placed between the original self-attention layer and the feedforward layer as + +$$ +z ^ {\prime} = z + s _ {v} \cdot \tanh (\gamma_ {v}) \cdot \mathrm {A d a p t} ([ z, V ^ {*} ]), \qquad (5) +$$ + +where Adapt indicates the adaptive layer and $z$ and $z'$ denote the input and output features of the adaptive layer, respectively. $\gamma_v$ and $s_v$ serve as the gating parameter and scaling factor. We evaluated several adaptive layer structures, including self-attention, cross-attention, and the AdaIN layer, which is widely used in style transfer [1, 12]. Among these, self-attention was found to be the most effective. + +# 3.3. Training + +We used a pretrained 50-step MDM [33] for $\mathcal{D}$ . Diffusion is modeled as a Markov noising process $\{M^t\}_{t=0}^T$ , where $M^0$ is sampled from the training motion, and $t$ denotes the noise time-step. The diffusion loss for finetuning is given by + +$$ +L _ {D} := \mathbb {E} _ {M ^ {0}, t, T} \left[ \left\| M ^ {0} - \mathcal {D} \left(M ^ {t}, t, V ^ {*}, T ^ {*}\right) \right\| _ {2} ^ {2} \right] + L _ {\text {g e o}}, \tag {6} +$$ + +where $L_{geo}$ is the geometric loss used during the MDM pretraining [33]. The final training loss is expressed as + +$$ +L = L _ {D} + \lambda L _ {p c}, \tag {7} +$$ + +where $\lambda$ is a weight for the $L_{pc}$ . + +We utilize Classifier-Free Guidance (CFG) [8], which is a technique for adjusting the trade-off between the diversity and fidelity of generated samples. Similarly, we applied CFG during finetuning to enable control over the trade-off between the pretrained diffusion model's capability for generating diverse motions and fidelity to the newly provided conditions, $V^{*}$ and $T^{*}$ . During training, $V^{*}$ and $T^{*}$ are randomly dropped with a $10\%$ probability, respectively. During sampling, the output of $t$ -th step is extrapolated as + +$$ +\hat {\mathcal {D}} \left(M ^ {t}, t, V ^ {*}, T ^ {*}\right) = b \mathcal {D} _ {T} + (1 - b) \mathcal {D} _ {V}, \tag {8} +$$ + +$$ +\mathcal {D} _ {T} = \mathcal {D} (M ^ {t}, t, V ^ {*}, \emptyset) + g _ {t} (\mathcal {D} (M ^ {t}, t, V ^ {*}, T ^ {*}) - \mathcal {D} (M ^ {t}, t, V ^ {*}, \emptyset)), +$$ + +$$ +\mathcal {D} _ {V} = \mathcal {D} (M ^ {t}, t, \emptyset , T ^ {*}) + g _ {v} (\mathcal {D} (M ^ {t}, t, V ^ {*}, T ^ {*}) - \mathcal {D} (M ^ {t}, t, \emptyset , T ^ {*})), +$$ + +where $\varnothing$ indicates the absence of an input. $g_{t}$ and $g_{v}$ represent guidance scales, while $b$ serves as the balancing factor between the modalities. As $g_{v}$ increases, the model puts more focus on the persona inherent in the visual feature. However, if $g_{v}$ is set too high, inconsistencies with the + +![](images/b3dd2d8527a6233f29249a741e201a0e87c9fef8fe012f8a5f2c2d6431de15ab.jpg) + +![](images/595e37677116b5058e08ec9259b6e5215caa422031f7ce05bdc3b1efddc9ddf8.jpg) + +![](images/88bed377e986fbedf982d65103336f9ae2faa6432c336042c94e7b7143806aee.jpg) + +![](images/cbc0a3f448144278e960eb0bd12eac8ef2db027a1b35196f3a0ea6672bc6f96e.jpg) +(a) +(b) +(c) +Figure 4. (a) Motion capture studio and examples of data formats: skeleton, markers, and mesh. (b) Unique persona expressions of each actor in the 'Childish' category. (c) Rendered mesh for each actor in the 'Fearful' category + +text prompt may arise. On the other hand, $g_{t}$ controls the model's adherence to the prompt, with the intensity of $P^{*}$ in the text being adjusted by $s_t$ in Eq. (4). The hyperparameter settings are specified in Sec. 5, and an ablation study for these is provided in the supplementary material. + +# 3.4. Context-Aware Fusion for Multiple Inputs + +Intuitively, if multiple motions are input, these additional cues can help with generating motion. However, it's also crucial to choose the right cues, as transferring persona from motions with very different content types, even with $L_{pc}$ , can result in unnatural movements. A straightforward approach might be to take the mean of all features, as suggested in [31]. In the motion domain, however, taking all motions can result in blended motions that appear implausible. To overcome this, we introduce a Context-Aware Fusion (CAF) method that prioritizes input motions based on their contextual relevance to the input prompt. To assess this relevance, we use the motion encoder and text encoder of the motion clip ( $\mathcal{E}$ clip and $\mathcal{X}$ clip), which align the text and motion feature spaces. The CAF is formulated as + +$$ +I _ {\text {T o p - k}} = \operatorname {a r g m a x} _ {i} \left(S _ {i}\right), \quad S _ {i} = \sin \left(\mathcal {E} _ {\text {c l i p}} \left(M _ {i}\right), \bar {T} ^ {*}\right), \tag {9} +$$ + +$$ +w _ {i} = \left\{ \begin{array}{l l} \frac {\exp (S _ {i})}{\sum_ {n} \exp (S _ {n})}, & \text {i f} i \in I _ {\mathrm {T o p - k}}, \\ 0, & \text {o t h e r w i s e}, \end{array} \right. \tag {10} +$$ + +where $\bar{T}^*$ represents the personalized text feature, using the average value of $\{P^{*}\}$ as a persona token. $\mathrm{sim}(\cdot)$ denotes cosine similarity. Consequently, + +$$ +V ^ {*} = \sum_ {i} w _ {i} V _ {i} ^ {*}, \quad P ^ {*} = \sum_ {i} w _ {i} P _ {i} ^ {*}. \tag {11} +$$ + +Table 2. Comparison with existing motion style datasets. The note for \* is provided in the main text. - indicates no information + +
DatasetActorsStylesContentsClipsMeshText
Xia [34]-86572XX
BFA [1]116932XX
BN-1 [13]2*1517175XX
BN-2 [13]2*7102,902XX
100Style [19]1100*8*810XX
PerMo(ours)534106,610✓✓
+ +# 4. PerMo Dataset + +We collected a large-scale PerMo dataset, capturing personas from multiple actors. To ensure variety, we hired five professional motion capture actors of diverse genders and body types. Each actor is assigned to perform 34 styles, categorized into Age, Character, Condition, Emotion, Traits, and Surroundings, resulting in a total of 170 personas. Each actor performed 10 distinct contents for every style, carefully selected to engage different body parts. + +The dataset was captured in a studio equipped with 33 OptiTrack cameras, using 41 optical markers per person. For compatibility with other motion datasets, we converted the marker data to SMPL-H [29] format using $\mathrm{Mosh}++$ [18]. All data underwent expert cleaning and verification, with the entire process adhering to strict quality guidelines. Fig. 4 (a) illustrates the motion capture studio and the provided data formats, while (b) highlights the distinct personas expressed by each actor. Additionally, we constructed 20-30 varied descriptions for each content type, following research indicating that high-quality finetuning descriptions improve prompt-based control [7]. Starting with one detailed description, we instructed ChatGPT to generate a range from high-level descriptions (e.g., "A person is jumping ahead") to low-level descriptions (e.g., "A person pushes off the ground and jumps forward several times"). Additional details on data categories, capture process, data structure, and examples can be found in the supplementary. + +Table 2 shows a comparison with existing motion style datasets. Notably, PerMo is the first dataset to collect data from multiple actors. Our dataset also offers the highest number of total clips and content categories among trainable datasets and is the only one that includes mesh, marker data, and descriptions. The BN datasets [13] involve two actors; however, for the same category, only one actor is used (BN-1), or there are no tags identifying the actors (BN-2). BN-1 is unsuitable for training, as it contains only a single motion sequence per category. The 100Style [19] offers the largest variety of styles, but 58 of its categories focus on content, rather than on style [38]. Additionally, the eight content categories in this dataset mainly consist of walking or running, offering limited variation. + +![](images/80d672849ac141d6afeb7cbc97f39e81639c75892c3c876643daaeb0c585f4db.jpg) +Figure 5. Example of the ablation study. The input motions are from the 'Uppity' of Actor 1. The input prompt is "A person walks in a circle." In (a) and (b), only $M_1$ is provided for the input, while both $M_1$ and $M_2$ are provided for (c) and (d). $L_{pc}$ encourages the generated motion to closely follow the prompt, while CAF prevents the motion from blending. We set $k = 1$ for CAF + +# 5. Experiments + +Implementation Details. We conduct experiments using the proposed PerMo and 100Style [19] datasets. During training, a single motion is cropped differently to serve as both the input and the GT motion for generation. We set $\lambda$ to $10^{-2}$ and use the AdamW [16] optimizer with a learning rate of $10^{-4}$ . Training runs for 500 epochs with a batch size of 64. We set $s_t$ and $s_v$ to 1 during training and to 0.3 during inference. $g_t$ is set to 10, and $g_v$ to 15 and 10 for PerMo and 100Style, respectively. We set $b$ to 0.7 for single-input and 0.5 for multi-input settings, and $k$ to 5 for CAF. + +Evaluation Settings. For quantitative evaluation, we use descriptions from the HumanML3D test set as prompts and motion samples from the PerMo and 100Style datasets as inputs. For each description, $[P^*]$ is logically placed as a modifier for the subject. We used two evaluation settings: Single Input (SI) and Multiple Input (MI). In SI, one motion is sampled from the entire dataset for comparison with MST methods. In MI, a persona set is first sampled, and then $|M_i|$ motions are drawn from it. + +Performance is evaluated using Frechet Inception Distance (FID), R-Precision, and Diversity—metrics commonly used in motion generation tasks [3, 5, 6, 33]—as well as Persona Recognition Accuracy (PRA), a metric adopted by Style Recognition Accuracy (SRA) in MST [9, 32]. FID assesses the overall quality and realism of generated motions and is our primary metric [5]. R-Precision evaluates text-to-motion alignment accuracy [6], and Diversity [5] measures how well the generated motions reflect a broad distribution of personas. PRA uses a pretrained persona classifier to evaluate the persona consistency of generated motions. The details are in the supplementary material. + +# 5.1. Ablation Study + +Table 3 demonstrates the effects of our key contributions. The baseline represents results obtained using only visual adaptation. When we introduce $P^*$ for textual adaptation, the FID score drops significantly, and both PRA and Diver + +Table 3. Ablation study of the proposed components. $|M_{i}| = \max$ indicates that all motions in the chosen persona set are input + +
MethodsFID ↓R Precision ↑PRA avg.↑Diversity ↑
Top 1Top 2Top 3
Single Input (SI) Setting
baseline7.450.060.110.1617.997.48
+ P*5.060.050.100.1518.268.01
+Lpc3.180.150.260.3318.057.74
Multiple Input (MI) Setting
Mi|=14.270.080.150.2016.207.71
Mi|=max, [31]3.520.190.300.3819.247.88
Mi|=max, CAF2.950.190.300.3918.138.12
+ +sity increase. This indicates that multimodal integration of persona information not only strengthens the reflection of persona but also supports the generation of plausible motion. Furthermore, adding $L_{pc}$ leads to an additional reduction in FID and a notable increase in R-precision. This implies that $L_{pc}$ helps the Persona Extractor effectively capture persona essence disentangled from the input content, resulting in generated motion that aligns well with the prompt without compromising content. + +Next, the effectiveness of CAF is validated in the MI setting. When multiple input motions are used $(|M_i| = \max)$ , all metrics improve compared to using a single input $(|M_i| = 1)$ due to the higher likelihood of finding motion candidates that align with the prompt. Additionally, even in cases with multiple inputs, selecting the contextually most relevant Top-k motions—as done by CAF—is shown to be crucial. For $|M_i| = \max$ , [31] presents a baseline approach that uses the average feature of all inputs. Introducing the proposed CAF reduces FID while increasing Diversity, indicating that the generated motions are sufficiently diverse and plausible rather than oversimplified by averaging. Although the PRA score decreases, we found through qualitative results that generating plausible motion is a more crucial factor in improving motion quality. + +![](images/a9f0b197b8b28ee9feabe82438f4d8c33d972774853cab37262920349a83d5f6.jpg) +Figure 6. Example reflecting the different personas of Actors 1-5 for the same 'Childish' category. The input prompt is "A person is running and bending to pick something." Red arrows highlight the frames that clearly show distinct movements. The input motion of each actor can be referenced in Fig. 4 (b) + +Table 4. Ablation study regarding adaptation. PersonaBooth (SI) indicates our complete model for single inputs + +
MethodsFID ↓R Precision ↑PRA avg.↑Diversity ↑
Top 1Top 2Top 3
AdaIN8.770.040.070.1015.107.33
Cross-Attn8.960.040.070.1016.377.62
w/o T adapt.3.610.140.230.3117.777.88
PersonaBooth (SI)3.180.150.260.3318.057.74
+ +Fig. 5 provides an example. When using only $M_{1}$ as the input motion, if $L_{pc}$ is not applied, the generated motion fails to follow the path provided in the prompt, as shown in (a). However, with $L_{pc}$ , the motion better aligns with the prompt's trajectory, though it still includes unintended hand-waving motions unrelated to the prompt. When using multiple input motions, including $M_{2}$ (a walking motion with arms crossed), taking the mean of features results in a blended motion, as shown in (c). This blending causes one arm to bend while the other hangs low, creating unnatural poses. In contrast, with CAF, $M_{2}$ receives a higher weight, resulting in a more natural walking motion with arms crossed, as shown in (d). + +Table 4 presents various ablation results related to adaptation. For the visual adaptation layer, Self-Attention proved to be the most effective compared to AdaIN and CrossAttn. Additionally, introducing text adaptation at Personalized Text Encoder improves the PRA, FID, and R-precision metrics. Fig. 6 demonstrates how PersonaBooth effectively captures and represents each actor's unique action. The red arrows specifically highlight the distinct arm movements of each persona. Additional details, including variations in body tilt and movement speed, are more clearly observable in the supplementary video materials. + +# 5.2. Comparison with State-of-the-Art methods + +We compare our PersonaBooth with the state-of-the-art diffusion-based MST approaches, MoMo [25] and MCM-LDM [32], on the PerMo dataset. Additionally, we evalu + +ate it on an existing style dataset, 100Style [19]. As MCM-LDM does not support prompt-based generation, we used MDM-generated motions [33] as target motions for evaluation. MCM-LDM was evaluated using the original model trained on HumanML3D, as well as finetuned models, referred to as MCM-LDM, which were trained on the PerMo and 100Style datasets. + +# 5.2.1. PerMo Dataset + +The comparative results for the PerMo dataset are shown in Table 5, with all subcategory results displayed for PRA. The single-input PersonaBooth significantly reduces FID compared to existing methods, indicating that while previous studies struggled with PerMo—a challenging dataset containing more diverse content than 100Style—PersonaBooth performs robust in generating high-quality motion. Additionally, PersonaBooth achieves the highest performance in R-Precision, PRA, and Diversity compared to existing methods. Applying multiple inputs and CAF further enhances performance. It's noteworthy that while MoMo and MCM-LDM utilized a 100-step and 1000-step diffusion model, respectively, we used a 50-step diffusion model. + +Fig. 7 shows qualitative results. MoMo [25] notes that their generated motions often follow the input motion's direction instead of the target's, which they attempted to copy the root rotation from the input motion (shown in (a) before and (b) after adjustment). The prompts for "turns in the air" were not met in (a), (b), or (c). Particularly, (c) does not reflect the style of ballerina. In contrast, PersonaBooth accurately generates the motion by first performing a hop that reflects the style, followed by a precise turn in the air. Example videos are provided in the supplementary material. + +# 5.2.2. 100Style Dataset + +Table 6 presents the comparative results on the 100Style dataset. MoMo demonstrates high PRA and Diversity, indicating effective style reflection. However, it shows very low R-precision, suggesting poor alignment between generated motions and prompts. Finetuned MCM-LDM improved performance on nearly all metrics compared to its + +Table 5. Comparison with the state-of-the-art methods on the PerMo dataset. The comparison is made in the Single Input (SI) setting as the existing methods do not support multiple inputs. MCM-LDM* indicates the model is finetuned on the PerMo dataset. + +
MethodsStepsFID ↓R Precision ↑PRA ↑Diversity ↑
Top 1Top 2Top 3AgeChar1Char2Cond1Cond2Emo1Emo2TraitSurAvg.
MoMo [25]10013.910.050.090.1328.2310.2311.308.7315.979.9714.5712.607.9313.476.74
MCM-LDM [32]10009.160.130.230.3047.3811.4410.347.6318.9011.6917.2015.719.2217.006.70
MCM-LDM* [32]10009.700.130.230.3046.4912.129.227.9519.7711.7518.0515.187.3316.766.76
PersonaBooth (SI)503.180.150.260.3348.0013.6711.976.6918.6714.3419.7517.068.8018.057.74
PersonaBooth (MI)502.950.190.300.3951.9514.8912.005.1622.959.4418.1415.4510.0918.138.12
+ +![](images/6c4fc08e65ce11eb7aaf80941e558b4b5fea3416c29c54c283c1aaa8a666d9c2.jpg) +Input Motion (Actor1 Ballerina) + +![](images/548ca00ac43886abd51863ab34976cf0a44b554855f2719fbc77784102f8a7e6.jpg) +(a) MoMo (w/o copy root rot.) + +![](images/1acb58de0e6575b6d5c1d341037d3df68d2967cf0543955c1439f2bc276698ae.jpg) +(b) MoMo + +![](images/d08c0965aadb5a54a987083b2d3816180587f79c269b926e90f53247e25d00d2.jpg) +(c) MCM-LDM* + +![](images/8dcf2acd08cef812a80589d4592ae4800a31e954e2cf3169b5650e953d17d1b4.jpg) +(d) PersonaBooth (SI) + +![](images/57fa59f08fcc0be1c4672dcf6993f44d71ad5490f7a17f6e8c14fa043448f51b.jpg) +Input Motion (Airplane) + +![](images/5e9a864749f97a8a4e6adee2b53f0f5ec8174ecc96f67afd0da4278d6d8c21b5.jpg) +Figure 7. Qualitative comparison on PerMo dataset. The input prompt is "A person hops forward and turns in the air." +(a) MoMo (w/o copy root rot.) + +![](images/560be798c05ca2f8bfea0febbdd79a5ae77151fa009a7b8955dfcf2dae5678a3.jpg) +(b) MoMo +Figure 8. Qualitative comparison on 100Style dataset. The input prompt is "A person walks backward and sits down on the chair." + +![](images/1ae6bf02847599e9ad486805e27337755b1e45d693702dbad38bc86847aba36e.jpg) +(c) MCM-LDM* + +![](images/f827cad047c1c818a0e36d6e109082b3c130348292c75fe270b731676448446d.jpg) +(d) PersonaBooth (SI) + +Table 6. Comparison with the state-of-the-art methods on the 100Style dataset. MCM-LDM* indicates the model is finetuned on the 100Style dataset + +
MethodsFID ↓R Precision ↑SRA ↑Diversity ↑
Top 1Top 2Top 3
MoMo [25]5.970.070.100.1460.577.82
MCM-LDM [32]7.200.180.290.3754.446.76
MCM-LDM* [32]6.530.170.280.3754.617.12
PersonaBooth (SI)3.270.200.310.4064.537.90
+ +original model. Although finetuned MCM-LDM achieved high R-precision, it also had a high FID, resulting in unnatural motion generation, with low PRA and Diversity indicating weaker style reflection. In contrast, our single-input PersonaBooth significantly reduces FID and achieves top performance across all metrics, including R-precision, PRA, and Diversity. + +In Fig. 8 (a) and (b), MoMo produces sliding motions that don't align with the prompt. MCM-LDM shows sideward movements. In contrast, PersonaBooth accurately reflects both the input motion's style, like arm extension and leaning, and the sitting motion in line with the prompt. + +# 6. Conclusion + +We propose the PerMo dataset and the PersonaBooth framework for Motion Personalization. Through contrastive learning, we effectively capture the essence of a persona. Furthermore, we enhance the integration of persona traits into diffusion models using a multi-modal finetuning method. We also introduce the CAF, which effectively handles multiple inputs, ensuring cohesive representation of personalized motion characteristics. + +Limitations and Future Works. The PerMo contains relatively short motion sequences, which sometimes lead to stationary poses at the end of generated motions of the same length. However, this issue does not occur when prompted with longer actions, suggesting that automatic adjustment of motion length remains a task for future work. In addition, we plan to investigate approaches for applying CAF to each sequential action within a prompt individually. For instance, applying CAF separately to 'run' and 'jump' in the prompt "A person runs and then jumps." + +# Acknowledgement. + +This work was supported by MSIT, Korea, under [RS-2022-00164860] and [RS-2024-00418641] supervised by IITP. + +# References + +[1] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. ACM Transactions on Graphics (TOG), 39(4):64-1, 2020. 2, 4, 5 +[2] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1 +[3] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 1, 2, 6 +[4] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[5] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3D human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 6 +[6] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 1, 2, 3, 4, 6 +[7] Xingzhe He, Zhiwen Cao, Nicholas Kolkin, Lantao Yu, Kun Wan, Helge Rhodin, and Ratheesh Kalarot. A data perspective on enhanced identity preservation for diffusion personalization. arXiv preprint arXiv:2311.04315, 2023. 5 +[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4 +[9] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2, 6 +[10] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 3 +[11] Boeun Kim, Hyung Jin Chang, Jungho Kim, and Jin Young Choi. Global-local motion transformer for unsupervised skeleton-based action learning. In European Conference on Computer Vision, pages 209-225. Springer, 2022. 2 +[12] Boeun Kim, Jungho Kim, Hyung Jin Chang, and Jin Young Choi. Most: Motion style transformer between diverse action contents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1705-1714, 2024. 2, 4 +[13] Makito Kobayashi, Chen-Chieh Liao, Keito Inoue, Sentaro Yojima, and Masafumi Takahashi. Motion capture dataset + +for practical use of ai-based motion editing and stylization. arXiv preprint arXiv:2306.08861, 2023.5 +[14] Jehee Lee, Jinxiang Chai, Paul SA Reitsma, Jessica K Hodgins, and Nancy S Pollard. Interactive control of avatars animated with human motion data. In Proceedings of the 29th annual Conference on Computer Graphics and Interactive Techniques, pages 491-500, 2002. 1 +[15] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2 +[16] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[17] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2 +[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 5 +[19] Ian Mason, Sebastian Starke, and Taku Komura. Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 5(1):1-18, 2022. 2, 5, 6, 7 +[20] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on Computer Graphics and Interactive Techniques, 4(3):1-17, 2021. 2 +[21] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 1 +[22] Mathis Petrovich, Michael J Black, and Gül Varol. Tmr: Text-to-motion retrieval using contrastive 3D human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9488-9497, 2023. 3 +[23] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2 +[24] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1 +[25] Sigal Raab, Inbar Gat, Nathan Sala, Guy Tevet, Rotem Shalev-Arkushin, Ohad Fried, Amit H Bermano, and Daniel Cohen-Or. Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. arXiv preprint arXiv:2406.06508, 2024. 2, 7, 8 +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3 + +[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2 +[29] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 5 +[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2 +[31] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 2, 4, 5, 6 +[32] Wenfeng Song, Xingliang Jin, Shuai Li, Chenglizhao Chen, Aimin Hao, Xia Hou, Ning Li, and Hong Qin. Arbitrary motion style transfer with multi-condition motion latent diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821-830, 2024. 2, 3, 6, 7, 8 +[33] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 4, 6, 7 +[34] Shihong Xia, Congyi Wang, Jinxiang Chai, and Jessica Hodgins. Realtime style transfer for unlabeled heterogeneous human motion. ACM Transactions on Graphics (TOG), 34 (4):1-10, 2015. 2, 5 +[35] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 2 +[36] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 4 +[37] Xulu Zhang, Xiao-Yong Wei, Wengyu Zhang, Jinlin Wu, Zhaoxiang Zhang, Zhen Lei, and Qing Li. A survey on personalized content synthesis with diffusion models. arXiv preprint arXiv:2405.05538, 2024. 2 +[38] Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, and Huaizu Jiang. Smoodi: Stylized motion diffusion model. arXiv preprint arXiv:2407.12783, 2024. 2, 5 \ No newline at end of file diff --git a/2025/PersonaBooth_ Personalized Text-to-Motion Generation/images.zip b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..469de4153267be1501421fb162cd5ea255c36435 --- /dev/null +++ b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e11f11e3b846f2020b474fcb7c4aa588e83d0f98f2135e0e0fd8e242f2b4cece +size 600025 diff --git a/2025/PersonaBooth_ Personalized Text-to-Motion Generation/layout.json b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..13be95af616b0ba057e93b07acc0252d3a817dbf --- /dev/null +++ b/2025/PersonaBooth_ Personalized Text-to-Motion Generation/layout.json @@ -0,0 +1,9574 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 103, + 477, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 103, + 477, + 119 + ], + "spans": [ + { + "bbox": [ + 133, + 103, + 477, + 119 + ], + "type": "text", + "content": "PersonaBooth: Personalized Text-to-Motion Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 152, + 553, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 553, + 210 + ], + "type": "text", + "content": "Boeun Kim1,2,3 Hea In Jeong2 JungHoon Sung3 Yihua Cheng1 Jeongmin Lee2 \nJu Yong Chang4 Sang-II Choi3 Younggeun Choi3 Saim Shin2 Jungho Kim2 Hyung Jin Chang1 \n1University of Birmingham 2Korea Electronics Technology Institute 3Dankook University 4Kwangwoon University \nhttp://boeun-kim.github.io/page-PersonaBooth" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 246, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 246, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 152, + 246, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 271, + 297, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 297, + 547 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 297, + 547 + ], + "type": "text", + "content": "This paper introduces Motion Personalization, a new task that generates personalized motions aligned with text descriptions using several basic motions containing Persona. To support this novel task, we introduce a new large-scale motion dataset called PerMo (PersonaMotion), which captures the unique personas of multiple actors. We also propose a multi-modal finetuning method of a pretrained motion diffusion model called PersonaBooth. PersonaBooth addresses two main challenges: i) A significant distribution gap between the persona-focused PerMo dataset and the pretraining datasets, which lack persona-specific data, and ii) the difficulty of capturing a consistent persona from the motions vary in content (action type). To tackle the dataset distribution gap, we introduce a persona token to accept new persona features and perform multi-modal adaptation for both text and visuals during finetuning. To capture a consistent persona, we incorporate a contrastive learning technique to enhance intra-cohesion among samples with the same persona. Furthermore, we introduce a context-aware fusion mechanism to maximize the integration of persona cues from multiple input motions. PersonaBooth outperforms state-of-the-art motion style transfer methods, establishing a new benchmark for motion personalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": "Imagine a scenario where, by recording just a few basic movements of yourself, an avatar in a virtual world can mirror your personal traits and unique motion style. This kind of 'motion personalization' enables realistic interactions in virtual spaces such as games and the metaverse [14]. Moreover, if avatar motions could be directed through text, creating video content would become easy without the need for real-life stunt actors [2, 24]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "In this paper, we propose a task called Motion Personalization, which aims to generate text-based motions that" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 245, + 544, + 488 + ], + "blocks": [ + { + "bbox": [ + 328, + 245, + 544, + 488 + ], + "lines": [ + { + "bbox": [ + 328, + 245, + 544, + 488 + ], + "spans": [ + { + "bbox": [ + 328, + 245, + 544, + 488 + ], + "type": "image", + "image_path": "0bd3c3c0f3a60c18c3243529198756082868ca673432a91f682450903be322da.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 496, + 555, + 541 + ], + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 541 + ], + "type": "text", + "content": "Figure 1. Motion Personalization generates text-driven, personalized motions based on persona embedded in atomic input motions. We propose a new framework, PersonaBooth, along with a new benchmark dataset, PerMo, for Motion Personalization" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": "reflect individual personas using a few key atomic movements, such as jumping, punching, or walking. We define a persona as the unique style expression of an individual. As illustrated in Fig. 1, this task aims to generate a broad range of realistic motions robustly. To facilitate this, we introduce a model called PersonaBooth and a new benchmark, PerMo (PersonaMotion), specifically designed for motion personalization. Table 1 provides a comparison of similar motion generation tasks. Text-to-Motion Generation approaches [6, 21] generate motion using only text as input, without any reference motion, and numerous diffusion-based methods [3, 33, 35] have been proposed for this task. Motion Style Transfer (MST) focuses on transferring style" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22756" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 82, + 294, + 150 + ], + "blocks": [ + { + "bbox": [ + 88, + 71, + 263, + 81 + ], + "lines": [ + { + "bbox": [ + 88, + 71, + 263, + 81 + ], + "spans": [ + { + "bbox": [ + 88, + 71, + 263, + 81 + ], + "type": "text", + "content": "Table 1. Comparison of motion generation tasks" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 82, + 294, + 150 + ], + "lines": [ + { + "bbox": [ + 56, + 82, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 82, + 294, + 150 + ], + "type": "table", + "html": "
Motion TaskExisting DatasetsSource MotionsMulti-modal Finetuning
T2M Generation [3, 33, 35][6] [23]NoneX
Style Transfer [25, 32, 38][34][1][19]SingleX
Personalization (Ours)NoneMultiple✓
", + "image_path": "2612e3f631052b9f61a518029d8cee03797c1dff46b435d7c9604cb42c56dc2b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 295, + 258 + ], + "type": "text", + "content": "from a single source motion. While several datasets and methods are available for MST [25, 32, 34, 38], the use of a single source motion restricts the ability to generate the broad and diverse motions seen in real life [12]. The proposed Motion Personalization has no existing prior approaches or datasets. Furthermore, our proposed method introduces multi-modal finetuning—a process neglected in existing diffusion-based MST methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 259, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 259, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 295, + 544 + ], + "type": "text", + "content": "PersonaBooth is a multi-modal finetuning method for pretrained text-to-motion diffusion models. The challenge in finetuning diffusion models for the motion domain, compared to the image domain [4, 15, 17, 30, 31, 37], lies in the relatively lower diversity of pretraining data. The most commonly used pretraining dataset, HumanML3D [6], contains only 15K samples, which is significantly smaller than the large-scale image datasets with 400M-650M samples [27, 28]. Additionally, HumanML3D includes few data reflecting personality traits, creating a substantial distribution gap when finetuning with the persona-focused PerMo dataset. MoMo [25] found that only 55 samples include styles for 'run,' 'walk,' 'jump,' and 'dance' locomotions in HumanML3D, and we could not find any persona-related samples. Existing diffusion-based MST methods [25, 38] consider only the visual features of the input motion and rely on a fixed text feature, which limits their ability to incorporate new data into the textual component. This inflexibility can lead to a phenomenon known as 'forgetting', where the pretrained model loses its initial capabilities when finetuned with new data [38]. To address this, we introduce learnable persona tokens to capture persona features from new data and propose an adaptation scheme for both text and visuals." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Another challenge is extracting a consistent persona across different atomic motions, which manifests in diverse ways depending on the action content. For example, the elegance of a ballerina is expressed in her feet as she walks and in her hands as she waves. MST methods have faced similar challenges in disrupting the content of the target motion when transferring styles from the source motion with different content. Approaches like those of Park et al. [20], Jang et al. [9], and MoMo [25] cite this as a limitation. In contrast, MoST [12] addresses the issue with a style disentanglement loss, while MCM-LDM [32] uses the target motion's trajectory to preserve content. However, MoST's loss function requires running the model twice for two different input motions, which is computationally expensive," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "especially for diffusion models with numerous steps. Additionally, since our goal is to generate motion from text rather than from target motions, the method used by MCM-LDM is not applicable. Therefore, we introduce a novel contrastive learning-based loss specifically designed for text-to-motion diffusion, called persona cohesion loss. This loss facilitates cohesion across motion features with different content but the same persona." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 168, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 263 + ], + "type": "text", + "content": "In addition, while MST uses a single input motion, we propose a new fusion method for multiple motions. In the image domain, InstantBooth [31] simply averages persona features, but this leads to unnatural blending in the motion. To address this, we introduce Context-Aware Fusion (CAF), which assigns weights to the fusion of persona features based on the similarity between the prompt and input motions. The summarized contributions of this paper are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 325, + 264, + 553, + 395 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 325, + 264, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 264, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 325, + 264, + 553, + 299 + ], + "type": "text", + "content": "- We introduce the Motion Personalization task and present PerMo, a new large-scale persona dataset captured by multiple actors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 325, + 300, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 300, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 325, + 300, + 553, + 358 + ], + "type": "text", + "content": "- We propose PersonaBooth, a multi-modal finetuning method for diffusion models that introduces text adaptation through a persona token. PersonaBooth applies contrastive learning to effectively capture personas across various content types." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 360, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 360, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 325, + 360, + 553, + 395 + ], + "type": "text", + "content": "- PersonaBooth achieves strong performance in both Motion Personalization and MST tasks and maximizes performance through a CAF for multiple inputs." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 406, + 400, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 406, + 400, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 406, + 400, + 418 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 553, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 557 + ], + "type": "text", + "content": "The Motion Style Transfer (MST) aims to transfer the style from a source motion to a target motion. The pioneering work by Aberman et al. [1] introduced the AdaIN layer, which simply replaces style statistics in specific layers during training. MotionPuzzle [9] proposes a Graph Convolution Network-based framework that transfers manually specified motion to a desired body part, enabling the combination of multiple source motions. MoST [12] addresses the issue of unwanted blending between different motion contents by introducing a transformer-based model [11] that more clearly separates style and content." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "Recent MST algorithms have seen notable performance gains through the use of diffusion models. MoMo [25] proposes a zero-shot motion transfer method using a pretrained diffusion model, where it mixes two input motions within the attention module of the diffusion model. However, MoMo identifies a limitation: when the source motion content differs significantly from the target motion or text, the generated motion may not follow the target accurately. For example, applying the style of a stationary source motion to a walking target motion may still result in a stationary output. SMooDi [38] proposes a finetuning method for a pretrained diffusion model. However, due to the forgetting issue, it retrained the model with both the original pretrain-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22757" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 558, + 238 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 558, + 238 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 558, + 238 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 558, + 238 + ], + "type": "image", + "image_path": "aff20c64e70c7a75e15820d1a4cba40d49ffbd82f81d4d88601719985bee9235.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "text", + "content": "Figure 2. The overall framework of PersonaBooth. PersonaBooth has two adaptation paths—visual and text—for finetuning the Motion Diffusion model " + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "inline_equation", + "content": "(\\mathcal{D})" + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "text", + "content": ". The Persona Extractor extracts both a visual persona feature " + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "inline_equation", + "content": "(V^{*})" + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "text", + "content": " and a persona token " + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "inline_equation", + "content": "(P^{*})" + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "text", + "content": " from the input motions. " + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "text", + "content": " is input into the adaptive layer of " + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "inline_equation", + "content": "P^{*}" + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "text", + "content": " is processed together with the input prompt through a Personalized Text Encoder, generating a personalized text feature, which is then input to " + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 54, + 244, + 555, + 301 + ], + "type": "text", + "content": ". The entire model is trained with a classifier-free approach, incorporating a Persona Cohesion Loss. During inference, Context-Aware Fusion is applied for multiple input cases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 312, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 296, + 445 + ], + "type": "text", + "content": "ing dataset and the style dataset. Additionally, the use of classifier-based training and the adaptation approach of creating a trainable copy of the entire model led to a training speed 10 times slower than the base diffusion model. MCM-LDM [32] addresses the challenges regarding two different contents of input motions by explicitly extracting the trajectory from the target motion and using it as a condition for the diffusion model. However, it faces a limitation: only trajectories similar to the target motion are generated, which reduces diversity. This contradicts our goal of creating diverse motions based on text prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 138, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 138, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 138, + 467 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": "We propose a method for multi-modal finetuning of a text-to-motion diffusion model, named PersonaBooth. The overall framework is shown in Fig. 2, and we aim to finetune the pretrained motion diffusion model, " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ". The input includes multiple atomic motions " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{M_i\\}" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " and a text description " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "T_{in}" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ". PersonaBooth generates personalized motion that reflects the description. We use the same motion representation as HumanML3D [6], " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_i \\in \\mathbb{R}^{f \\cdot 263}" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " is the number of frames, and 263 includes joint rotations, positions, velocities, and foot contact. First, the Persona Extractor " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " processes each input motion to extract a visual persona feature " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "V^*" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " and a persona token " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "P^*" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ". Built on a pretrained motion clip model [22] (" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{clip}" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": "), which captures the general characteristics of motion sequences, " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " includes a transformer structure (" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_P" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ") specially designed to extract persona features. We applied a contrastive learning scheme to enhance the persona extraction capability of " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_P" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ", which will be detailed in Sec. 3.1. The extracted " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "V^*" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " is then fed into a newly added adaptive layer in " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": ". Meanwhile, the Personalized Text Encoder " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " adaptively incorporates " + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "P^*" + }, + { + "bbox": [ + 56, + 474, + 296, + 714 + ], + "type": "text", + "content": " to pro" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": "duce a personalized text feature " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": " is input into " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": " ultimately generates the personalized motion. During training, a single motion-text pair is provided to enable reconstruction. During inference, given motion sets " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\{M_i\\}" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": ", the Context-Aware Fusion module generates weights to perform a weighted combination of the elements within " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\{V_{i}^{*}\\}" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\{P_i^*\\}" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": ". PersonaBooth explores efficient training by using a classifier-free guidance approach. During training, " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{clip}" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": ", excluding the newly added adaptive layer, are kept frozen. Sec.3.1 to Sec.3.3 describe the process based on the training of single motion, " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": ". The inference process for multiple input motions " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\{M_i\\}" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": " is explained in Sec.3.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 463, + 423, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 423, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 423, + 475 + ], + "type": "text", + "content": "3.1. Persona Extractor" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "text", + "content": "Persona Extractor " + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "text", + "content": " extracts the persona token " + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "inline_equation", + "content": "P^{*}" + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "text", + "content": " along with the visual persona feature " + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "text", + "content": ". We used the TMR [22] structure for " + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{clip}" + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "text", + "content": ", replacing their text encoder with the CLIP text encoder [26] and retraining it. The CLIP text encoder is the same one used in the Personalized Text Encoder " + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "text", + "content": ", and therefore, this replacement enables " + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "inline_equation", + "content": "P^{*}" + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "text", + "content": " to share the same token embedding space as the input prompt. For input motion " + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "text", + "content": ", persona features " + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "inline_equation", + "content": "P^{*}" + }, + { + "bbox": [ + 313, + 480, + 555, + 576 + ], + "type": "text", + "content": " are extracted as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 582, + 553, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 582, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 358, + 582, + 553, + 597 + ], + "type": "interline_equation", + "content": "V ^ {*} = \\mathcal {E} ([ c l s ], M ]), \\tag {1}", + "image_path": "b0bbd13d84dc112e4b991c6236fb39f02ba0684ff19ac7cb5be1cadd271088f0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 597, + 553, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 597, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 359, + 597, + 553, + 611 + ], + "type": "interline_equation", + "content": "P ^ {*} = \\operatorname {M L P} (Y), \\text {w h e r e} Y = V ^ {*} [ 0 ], \\tag {2}", + "image_path": "861112f28b15cf39794ca2f78b3ec6de5639c701a2c5e63c3f8f1931f7171e4d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 451, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 451, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 451, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 451, + 630 + ], + "type": "inline_equation", + "content": "[cls]" + }, + { + "bbox": [ + 313, + 617, + 451, + 630 + ], + "type": "text", + "content": " denotes a class token." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": "To ensure the coherence of persona features across different motion contents of the same persona, we introduce the persona cohesion loss " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "L_{pc}" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": ". We utilize a supervised contrastive learning scheme [10], where motion samples from the same persona are encouraged to be closer in the persona feature space, while motion samples from different personas are pushed apart. This helps to guide the persona fea" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22758" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 296, + 199 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 199 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 199 + ], + "type": "image", + "image_path": "fbe5dc67339bbcb345b260df5695cfb0ec480fbf90ccae02e195487ad7200c6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "type": "text", + "content": "Figure 3. Textual and visual adaptation. (a) Personalized Text Encoder, " + }, + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "type": "text", + "content": ". (b) " + }, + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "type": "text", + "content": "-th step of the Motion Diffusion, " + }, + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "type": "text", + "content": ". Learnable parameters are denoted by the fire icon" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "text", + "content": "tures to form well-defined clusters for each persona. For a batch containing " + }, + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "text", + "content": " input motion data, " + }, + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "inline_equation", + "content": "L_{pc}" + }, + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "text", + "content": " is applied to a total of " + }, + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "inline_equation", + "content": "2N" + }, + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "text", + "content": " data points, with one positive sample for each input motion. The positive sample is randomly selected from the same persona group as the input motion. " + }, + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "inline_equation", + "content": "L_{pc}" + }, + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "text", + "content": " for a positive pair of index " + }, + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 242, + 296, + 315 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 322, + 295, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 322, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 65, + 322, + 295, + 352 + ], + "type": "interline_equation", + "content": "L _ {p c} = - \\log \\frac {\\exp (\\mathrm {s i m} (h (Y _ {i}) , h (Y _ {j})) / \\tau)}{\\sum_ {k = 1} ^ {2 N} \\mathbb {1} _ {[ k \\neq i ]} \\exp (\\mathrm {s i m} (h (Y _ {i}) , h (Y _ {k})) / \\tau)}, \\tag {3}", + "image_path": "8dcb34caa69a32ab006b6a89b1218642e9dd39fe67de55828ff0a0493037024b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " denotes projection head. " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "Y_{j}" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " is extracted from " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "M_{j}" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " which belongs to the same persona group as " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(u,v)" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " indicates the cosine similarity between two vectors " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 358, + 296, + 407 + ], + "type": "text", + "content": " denotes the temperature parameter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 222, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 222, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 222, + 427 + ], + "type": "text", + "content": "3.2. Textual and Visual Adaptation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 431, + 296, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 296, + 503 + ], + "type": "text", + "content": "The discrepancy between our PerMo dataset and the non-persona-specific pretraining dataset, HumanML3D [6] necessitates a carefully designed adaptation mechanism for effective finetuning. To address this, we propose two adaptation paths—textual and visual—to seamlessly integrate the two persona features, " + }, + { + "bbox": [ + 55, + 431, + 296, + 503 + ], + "type": "inline_equation", + "content": "P^{*}" + }, + { + "bbox": [ + 55, + 431, + 296, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 431, + 296, + 503 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 55, + 431, + 296, + 503 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "content": "The structure of the Personalized Text Encoder " + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "inline_equation", + "content": "(\\mathcal{X})" + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "content": " is shown in Fig. 3 (a). Given an input sentence like \"A person is dancing\" " + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "inline_equation", + "content": "(T_{in})" + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "content": ", we introduce a personalized sentence \"[P*] person is dancing\" " + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "inline_equation", + "content": "(\\tilde{T}_{in})" + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "content": ", where the word " + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "inline_equation", + "content": "[P^*]" + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "content": " precedes the subject. Subsequently, the vector " + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "inline_equation", + "content": "P^*" + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "content": ", derived from " + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "content": ", replaces the token embedding corresponding to " + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "inline_equation", + "content": "[P^*]" + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "content": ". Then, the modified sentence embedding, " + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{clip}(T_{in})" + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "content": ", is adaptively merged with the original sentence embedding, " + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{clip}(T_{in})" + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "content": ". The descriptions in HumanML3D are generally structured with subjects like 'A Person,' or 'Someone' and lack adjectives to characterize the subject. Thus, a smoother adaptation is necessary to enable the pretrained model to handle variations in sentence structure effectively. Specifically, personalized text feature " + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "inline_equation", + "content": "T^*" + }, + { + "bbox": [ + 55, + 504, + 296, + 671 + ], + "type": "text", + "content": " is derived as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 73, + 680, + 295, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 73, + 680, + 295, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} T ^ {*} = \\mathcal {X} \\left(T _ {i n}, P ^ {*}\\right) \\tag {4} \\\\ = \\mathcal {X} _ {c l i p} (T _ {i n}) + s _ {t} \\cdot \\mathrm {t a n h} (\\gamma_ {t}) \\cdot \\mathcal {X} _ {c l i p} (\\tilde {T} _ {i n}, P ^ {*}), \\\\ \\end{array}", + "image_path": "7786260ce129c495710a703abfdfc7a3defb4338d652993e267b184abe8665e3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " is a learnable parameter initialized as zero which is for zero gating [36]. " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " is a scaling factor [31], a constant used to balance the adaptive layer during inference. CLIP text encoder " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{clip}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " is frozen during training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 121, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 192 + ], + "type": "text", + "content": "For visual adaptation, we introduce a single adaptive layer into the transformer structure, following the approach used in [31] for the UNet. As illustrated in Fig. 3 (b), for " + }, + { + "bbox": [ + 313, + 121, + 555, + 192 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 121, + 555, + 192 + ], + "type": "text", + "content": "-th transformer block in " + }, + { + "bbox": [ + 313, + 121, + 555, + 192 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 121, + 555, + 192 + ], + "type": "text", + "content": "-th step of the diffusion, the adaptive layer is placed between the original self-attention layer and the feedforward layer as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 350, + 201, + 555, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 201, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 350, + 201, + 555, + 216 + ], + "type": "interline_equation", + "content": "z ^ {\\prime} = z + s _ {v} \\cdot \\tanh (\\gamma_ {v}) \\cdot \\mathrm {A d a p t} ([ z, V ^ {*} ]), \\qquad (5)", + "image_path": "9de2b741e79e77f5a4e895b2a5e4d7b58c413b10298c5f91205b3daffa125c91.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 222, + 554, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 554, + 306 + ], + "type": "text", + "content": "where Adapt indicates the adaptive layer and " + }, + { + "bbox": [ + 313, + 222, + 554, + 306 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 222, + 554, + 306 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 222, + 554, + 306 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 313, + 222, + 554, + 306 + ], + "type": "text", + "content": " denote the input and output features of the adaptive layer, respectively. " + }, + { + "bbox": [ + 313, + 222, + 554, + 306 + ], + "type": "inline_equation", + "content": "\\gamma_v" + }, + { + "bbox": [ + 313, + 222, + 554, + 306 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 222, + 554, + 306 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 313, + 222, + 554, + 306 + ], + "type": "text", + "content": " serve as the gating parameter and scaling factor. We evaluated several adaptive layer structures, including self-attention, cross-attention, and the AdaIN layer, which is widely used in style transfer [1, 12]. Among these, self-attention was found to be the most effective." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 313, + 378, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 313, + 378, + 326 + ], + "spans": [ + { + "bbox": [ + 314, + 313, + 378, + 326 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "text", + "content": "We used a pretrained 50-step MDM [33] for " + }, + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "text", + "content": ". Diffusion is modeled as a Markov noising process " + }, + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "inline_equation", + "content": "\\{M^t\\}_{t=0}^T" + }, + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "inline_equation", + "content": "M^0" + }, + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "text", + "content": " is sampled from the training motion, and " + }, + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 331, + 554, + 380 + ], + "type": "text", + "content": " denotes the noise time-step. The diffusion loss for finetuning is given by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 386, + 555, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 386, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 318, + 386, + 555, + 412 + ], + "type": "interline_equation", + "content": "L _ {D} := \\mathbb {E} _ {M ^ {0}, t, T} \\left[ \\left\\| M ^ {0} - \\mathcal {D} \\left(M ^ {t}, t, V ^ {*}, T ^ {*}\\right) \\right\\| _ {2} ^ {2} \\right] + L _ {\\text {g e o}}, \\tag {6}", + "image_path": "a3cc5631f19f890949340c2af43c60b780d559aafac16c4283ee969cd804eb26.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 419, + 554, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 419, + 554, + 444 + ], + "type": "inline_equation", + "content": "L_{geo}" + }, + { + "bbox": [ + 313, + 419, + 554, + 444 + ], + "type": "text", + "content": " is the geometric loss used during the MDM pretraining [33]. The final training loss is expressed as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 398, + 453, + 553, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 453, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 398, + 453, + 553, + 466 + ], + "type": "interline_equation", + "content": "L = L _ {D} + \\lambda L _ {p c}, \\tag {7}", + "image_path": "94e0603a86c1bdad74d489d594155151635354c975f9c215c8b2d9b4433b9bbb.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 474, + 444, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 444, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 444, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 474, + 444, + 486 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 474, + 444, + 486 + ], + "type": "text", + "content": " is a weight for the " + }, + { + "bbox": [ + 313, + 474, + 444, + 486 + ], + "type": "inline_equation", + "content": "L_{pc}" + }, + { + "bbox": [ + 313, + 474, + 444, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "text", + "content": "We utilize Classifier-Free Guidance (CFG) [8], which is a technique for adjusting the trade-off between the diversity and fidelity of generated samples. Similarly, we applied CFG during finetuning to enable control over the trade-off between the pretrained diffusion model's capability for generating diverse motions and fidelity to the newly provided conditions, " + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "text", + "content": ". During training, " + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "text", + "content": " are randomly dropped with a " + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "text", + "content": " probability, respectively. During sampling, the output of " + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "text", + "content": "-th step is extrapolated as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 601, + 555, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 555, + 615 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {D}} \\left(M ^ {t}, t, V ^ {*}, T ^ {*}\\right) = b \\mathcal {D} _ {T} + (1 - b) \\mathcal {D} _ {V}, \\tag {8}", + "image_path": "32f171e80ffa9a50918638819f0a5af4a39c5fb4f03abb77cc389c00b4bf35cb.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 617, + 554, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 630 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {T} = \\mathcal {D} (M ^ {t}, t, V ^ {*}, \\emptyset) + g _ {t} (\\mathcal {D} (M ^ {t}, t, V ^ {*}, T ^ {*}) - \\mathcal {D} (M ^ {t}, t, V ^ {*}, \\emptyset)),", + "image_path": "6892c9f6517d3d10a61d43b20155fccb50bdfa4e595da4b6ce82345a0dc66080.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 632, + 554, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 645 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {V} = \\mathcal {D} (M ^ {t}, t, \\emptyset , T ^ {*}) + g _ {v} (\\mathcal {D} (M ^ {t}, t, V ^ {*}, T ^ {*}) - \\mathcal {D} (M ^ {t}, t, \\emptyset , T ^ {*})),", + "image_path": "c02f7f43407a35209663c155f248aab45ae41ceaefdd7496a184b92f9ebf4300.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " indicates the absence of an input. " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "g_{v}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " represent guidance scales, while " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " serves as the balancing factor between the modalities. As " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "g_{v}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " increases, the model puts more focus on the persona inherent in the visual feature. However, if " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "g_{v}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " is set too high, inconsistencies with the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22759" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 217, + 173 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 217, + 173 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 217, + 173 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 217, + 173 + ], + "type": "image", + "image_path": "b3dd2d8527a6233f29249a741e201a0e87c9fef8fe012f8a5f2c2d6431de15ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 218, + 70, + 256, + 173 + ], + "blocks": [ + { + "bbox": [ + 218, + 70, + 256, + 173 + ], + "lines": [ + { + "bbox": [ + 218, + 70, + 256, + 173 + ], + "spans": [ + { + "bbox": [ + 218, + 70, + 256, + 173 + ], + "type": "image", + "image_path": "595e37677116b5058e08ec9259b6e5215caa422031f7ce05bdc3b1efddc9ddf8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 257, + 70, + 294, + 173 + ], + "blocks": [ + { + "bbox": [ + 257, + 70, + 294, + 173 + ], + "lines": [ + { + "bbox": [ + 257, + 70, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 257, + 70, + 294, + 173 + ], + "type": "image", + "image_path": "88bed377e986fbedf982d65103336f9ae2faa6432c336042c94e7b7143806aee.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 182, + 294, + 247 + ], + "blocks": [ + { + "bbox": [ + 131, + 174, + 143, + 182 + ], + "lines": [ + { + "bbox": [ + 131, + 174, + 143, + 182 + ], + "spans": [ + { + "bbox": [ + 131, + 174, + 143, + 182 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 174, + 263, + 182 + ], + "lines": [ + { + "bbox": [ + 251, + 174, + 263, + 182 + ], + "spans": [ + { + "bbox": [ + 251, + 174, + 263, + 182 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 182, + 294, + 247 + ], + "lines": [ + { + "bbox": [ + 57, + 182, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 57, + 182, + 294, + 247 + ], + "type": "image", + "image_path": "cbc0a3f448144278e960eb0bd12eac8ef2db027a1b35196f3a0ea6672bc6f96e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 248, + 181, + 257 + ], + "lines": [ + { + "bbox": [ + 170, + 248, + 181, + 257 + ], + "spans": [ + { + "bbox": [ + 170, + 248, + 181, + 257 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 258, + 296, + 302 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 302 + ], + "type": "text", + "content": "Figure 4. (a) Motion capture studio and examples of data formats: skeleton, markers, and mesh. (b) Unique persona expressions of each actor in the 'Childish' category. (c) Rendered mesh for each actor in the 'Fearful' category" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "content": "text prompt may arise. On the other hand, " + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "content": " controls the model's adherence to the prompt, with the intensity of " + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "inline_equation", + "content": "P^{*}" + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "content": " in the text being adjusted by " + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "content": " in Eq. (4). The hyperparameter settings are specified in Sec. 5, and an ablation study for these is provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 380, + 276, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 276, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 276, + 393 + ], + "type": "text", + "content": "3.4. Context-Aware Fusion for Multiple Inputs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "content": "Intuitively, if multiple motions are input, these additional cues can help with generating motion. However, it's also crucial to choose the right cues, as transferring persona from motions with very different content types, even with " + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "inline_equation", + "content": "L_{pc}" + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "content": ", can result in unnatural movements. A straightforward approach might be to take the mean of all features, as suggested in [31]. In the motion domain, however, taking all motions can result in blended motions that appear implausible. To overcome this, we introduce a Context-Aware Fusion (CAF) method that prioritizes input motions based on their contextual relevance to the input prompt. To assess this relevance, we use the motion encoder and text encoder of the motion clip (" + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "content": "clip and " + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "content": "clip), which align the text and motion feature spaces. The CAF is formulated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 577, + 295, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 577, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 67, + 577, + 295, + 590 + ], + "type": "interline_equation", + "content": "I _ {\\text {T o p - k}} = \\operatorname {a r g m a x} _ {i} \\left(S _ {i}\\right), \\quad S _ {i} = \\sin \\left(\\mathcal {E} _ {\\text {c l i p}} \\left(M _ {i}\\right), \\bar {T} ^ {*}\\right), \\tag {9}", + "image_path": "7198fa7ebe81c944615929a2b4a94e80796824581029dcae5111961d5c34272e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 78, + 593, + 294, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 593, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 78, + 593, + 294, + 632 + ], + "type": "interline_equation", + "content": "w _ {i} = \\left\\{ \\begin{array}{l l} \\frac {\\exp (S _ {i})}{\\sum_ {n} \\exp (S _ {n})}, & \\text {i f} i \\in I _ {\\mathrm {T o p - k}}, \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {10}", + "image_path": "a6e1a469c9cf137b4de0edd341173bc2bf13249a330a0a525f66b32f13665f7d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\bar{T}^*" + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "text", + "content": " represents the personalized text feature, using the average value of " + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\{P^{*}\\}" + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "text", + "content": " as a persona token. " + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot)" + }, + { + "bbox": [ + 55, + 644, + 296, + 681 + ], + "type": "text", + "content": " denotes cosine similarity. Consequently," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 100, + 691, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 691, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 100, + 691, + 295, + 716 + ], + "type": "interline_equation", + "content": "V ^ {*} = \\sum_ {i} w _ {i} V _ {i} ^ {*}, \\quad P ^ {*} = \\sum_ {i} w _ {i} P _ {i} ^ {*}. \\tag {11}", + "image_path": "8238f945be1157910d842993669f6fcca6586c7bc92dc0c7aec77ec060518c87.jpg" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 318, + 102, + 548, + 198 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 554, + 92 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 554, + 92 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 554, + 92 + ], + "type": "text", + "content": "Table 2. Comparison with existing motion style datasets. The note for \\* is provided in the main text. - indicates no information" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 102, + 548, + 198 + ], + "lines": [ + { + "bbox": [ + 318, + 102, + 548, + 198 + ], + "spans": [ + { + "bbox": [ + 318, + 102, + 548, + 198 + ], + "type": "table", + "html": "
DatasetActorsStylesContentsClipsMeshText
Xia [34]-86572XX
BFA [1]116932XX
BN-1 [13]2*1517175XX
BN-2 [13]2*7102,902XX
100Style [19]1100*8*810XX
PerMo(ours)534106,610✓✓
", + "image_path": "1528411b0a332144d35fd2c18fc59d36a499c6f3ce951715c9132c43a43bab0a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 211, + 406, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 211, + 406, + 223 + ], + "spans": [ + { + "bbox": [ + 314, + 211, + 406, + 223 + ], + "type": "text", + "content": "4. PerMo Dataset" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 232, + 554, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 554, + 327 + ], + "type": "text", + "content": "We collected a large-scale PerMo dataset, capturing personas from multiple actors. To ensure variety, we hired five professional motion capture actors of diverse genders and body types. Each actor is assigned to perform 34 styles, categorized into Age, Character, Condition, Emotion, Traits, and Surroundings, resulting in a total of 170 personas. Each actor performed 10 distinct contents for every style, carefully selected to engage different body parts." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 329, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 544 + ], + "type": "text", + "content": "The dataset was captured in a studio equipped with 33 OptiTrack cameras, using 41 optical markers per person. For compatibility with other motion datasets, we converted the marker data to SMPL-H [29] format using " + }, + { + "bbox": [ + 313, + 329, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{Mosh}++" + }, + { + "bbox": [ + 313, + 329, + 554, + 544 + ], + "type": "text", + "content": " [18]. All data underwent expert cleaning and verification, with the entire process adhering to strict quality guidelines. Fig. 4 (a) illustrates the motion capture studio and the provided data formats, while (b) highlights the distinct personas expressed by each actor. Additionally, we constructed 20-30 varied descriptions for each content type, following research indicating that high-quality finetuning descriptions improve prompt-based control [7]. Starting with one detailed description, we instructed ChatGPT to generate a range from high-level descriptions (e.g., \"A person is jumping ahead\") to low-level descriptions (e.g., \"A person pushes off the ground and jumps forward several times\"). Additional details on data categories, capture process, data structure, and examples can be found in the supplementary." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": "Table 2 shows a comparison with existing motion style datasets. Notably, PerMo is the first dataset to collect data from multiple actors. Our dataset also offers the highest number of total clips and content categories among trainable datasets and is the only one that includes mesh, marker data, and descriptions. The BN datasets [13] involve two actors; however, for the same category, only one actor is used (BN-1), or there are no tags identifying the actors (BN-2). BN-1 is unsuitable for training, as it contains only a single motion sequence per category. The 100Style [19] offers the largest variety of styles, but 58 of its categories focus on content, rather than on style [38]. Additionally, the eight content categories in this dataset mainly consist of walking or running, offering limited variation." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22760" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 557, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 557, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 557, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 557, + 194 + ], + "type": "image", + "image_path": "80d672849ac141d6afeb7cbc97f39e81639c75892c3c876643daaeb0c585f4db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "text", + "content": "Figure 5. Example of the ablation study. The input motions are from the 'Uppity' of Actor 1. The input prompt is \"A person walks in a circle.\" In (a) and (b), only " + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "inline_equation", + "content": "M_1" + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "text", + "content": " is provided for the input, while both " + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "inline_equation", + "content": "M_1" + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "inline_equation", + "content": "M_2" + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "text", + "content": " are provided for (c) and (d). " + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "inline_equation", + "content": "L_{pc}" + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "text", + "content": " encourages the generated motion to closely follow the prompt, while CAF prevents the motion from blending. We set " + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "text", + "content": " for CAF" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 138, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 138, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 138, + 266 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "text", + "content": "Implementation Details. We conduct experiments using the proposed PerMo and 100Style [19] datasets. During training, a single motion is cropped differently to serve as both the input and the GT motion for generation. We set " + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "text", + "content": " and use the AdamW [16] optimizer with a learning rate of " + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "text", + "content": ". Training runs for 500 epochs with a batch size of 64. We set " + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "text", + "content": " to 1 during training and to 0.3 during inference. " + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "text", + "content": " is set to 10, and " + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "inline_equation", + "content": "g_v" + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "text", + "content": " to 15 and 10 for PerMo and 100Style, respectively. We set " + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "text", + "content": " to 0.7 for single-input and 0.5 for multi-input settings, and " + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 271, + 296, + 391 + ], + "type": "text", + "content": " to 5 for CAF." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "text", + "content": "Evaluation Settings. For quantitative evaluation, we use descriptions from the HumanML3D test set as prompts and motion samples from the PerMo and 100Style datasets as inputs. For each description, " + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "inline_equation", + "content": "[P^*]" + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "text", + "content": " is logically placed as a modifier for the subject. We used two evaluation settings: Single Input (SI) and Multiple Input (MI). In SI, one motion is sampled from the entire dataset for comparison with MST methods. In MI, a persona set is first sampled, and then " + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "inline_equation", + "content": "|M_i|" + }, + { + "bbox": [ + 55, + 392, + 296, + 499 + ], + "type": "text", + "content": " motions are drawn from it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 499, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 296, + 643 + ], + "type": "text", + "content": "Performance is evaluated using Frechet Inception Distance (FID), R-Precision, and Diversity—metrics commonly used in motion generation tasks [3, 5, 6, 33]—as well as Persona Recognition Accuracy (PRA), a metric adopted by Style Recognition Accuracy (SRA) in MST [9, 32]. FID assesses the overall quality and realism of generated motions and is our primary metric [5]. R-Precision evaluates text-to-motion alignment accuracy [6], and Diversity [5] measures how well the generated motions reflect a broad distribution of personas. PRA uses a pretrained persona classifier to evaluate the persona consistency of generated motions. The details are in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 648, + 149, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 149, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 149, + 661 + ], + "type": "text", + "content": "5.1. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Table 3 demonstrates the effects of our key contributions. The baseline represents results obtained using only visual adaptation. When we introduce " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "P^*" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " for textual adaptation, the FID score drops significantly, and both PRA and Diver" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 314, + 282, + 555, + 398 + ], + "blocks": [ + { + "bbox": [ + 313, + 251, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 275 + ], + "type": "text", + "content": "Table 3. Ablation study of the proposed components. " + }, + { + "bbox": [ + 313, + 251, + 555, + 275 + ], + "type": "inline_equation", + "content": "|M_{i}| = \\max" + }, + { + "bbox": [ + 313, + 251, + 555, + 275 + ], + "type": "text", + "content": " indicates that all motions in the chosen persona set are input" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 282, + 555, + 398 + ], + "lines": [ + { + "bbox": [ + 314, + 282, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 314, + 282, + 555, + 398 + ], + "type": "table", + "html": "
MethodsFID ↓R Precision ↑PRA avg.↑Diversity ↑
Top 1Top 2Top 3
Single Input (SI) Setting
baseline7.450.060.110.1617.997.48
+ P*5.060.050.100.1518.268.01
+Lpc3.180.150.260.3318.057.74
Multiple Input (MI) Setting
Mi|=14.270.080.150.2016.207.71
Mi|=max, [31]3.520.190.300.3819.247.88
Mi|=max, CAF2.950.190.300.3918.138.12
", + "image_path": "636604fe04ea667393ce6e45e83726b8e449122a09f5285b64a2ce9fad8cb03b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 423, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 423, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 312, + 423, + 555, + 532 + ], + "type": "text", + "content": "sity increase. This indicates that multimodal integration of persona information not only strengthens the reflection of persona but also supports the generation of plausible motion. Furthermore, adding " + }, + { + "bbox": [ + 312, + 423, + 555, + 532 + ], + "type": "inline_equation", + "content": "L_{pc}" + }, + { + "bbox": [ + 312, + 423, + 555, + 532 + ], + "type": "text", + "content": " leads to an additional reduction in FID and a notable increase in R-precision. This implies that " + }, + { + "bbox": [ + 312, + 423, + 555, + 532 + ], + "type": "inline_equation", + "content": "L_{pc}" + }, + { + "bbox": [ + 312, + 423, + 555, + 532 + ], + "type": "text", + "content": " helps the Persona Extractor effectively capture persona essence disentangled from the input content, resulting in generated motion that aligns well with the prompt without compromising content." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Next, the effectiveness of CAF is validated in the MI setting. When multiple input motions are used " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "(|M_i| = \\max)" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": ", all metrics improve compared to using a single input " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "(|M_i| = 1)" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " due to the higher likelihood of finding motion candidates that align with the prompt. Additionally, even in cases with multiple inputs, selecting the contextually most relevant Top-k motions—as done by CAF—is shown to be crucial. For " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "|M_i| = \\max" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": ", [31] presents a baseline approach that uses the average feature of all inputs. Introducing the proposed CAF reduces FID while increasing Diversity, indicating that the generated motions are sufficiently diverse and plausible rather than oversimplified by averaging. Although the PRA score decreases, we found through qualitative results that generating plausible motion is a more crucial factor in improving motion quality." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22761" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 182 + ], + "type": "image", + "image_path": "a9f0b197b8b28ee9feabe82438f4d8c33d972774853cab37262920349a83d5f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 190, + 555, + 225 + ], + "lines": [ + { + "bbox": [ + 54, + 190, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 555, + 225 + ], + "type": "text", + "content": "Figure 6. Example reflecting the different personas of Actors 1-5 for the same 'Childish' category. The input prompt is \"A person is running and bending to pick something.\" Red arrows highlight the frames that clearly show distinct movements. The input motion of each actor can be referenced in Fig. 4 (b)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 274, + 296, + 345 + ], + "blocks": [ + { + "bbox": [ + 55, + 243, + 294, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 294, + 266 + ], + "type": "text", + "content": "Table 4. Ablation study regarding adaptation. PersonaBooth (SI) indicates our complete model for single inputs" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 274, + 296, + 345 + ], + "lines": [ + { + "bbox": [ + 56, + 274, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 296, + 345 + ], + "type": "table", + "html": "
MethodsFID ↓R Precision ↑PRA avg.↑Diversity ↑
Top 1Top 2Top 3
AdaIN8.770.040.070.1015.107.33
Cross-Attn8.960.040.070.1016.377.62
w/o T adapt.3.610.140.230.3117.777.88
PersonaBooth (SI)3.180.150.260.3318.057.74
", + "image_path": "2464820ce61f3f86075dc4f8c4381f464a3b1d5df70fcaaebd1684b35ba10893.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "text", + "content": "Fig. 5 provides an example. When using only " + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "text", + "content": " as the input motion, if " + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "inline_equation", + "content": "L_{pc}" + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "text", + "content": " is not applied, the generated motion fails to follow the path provided in the prompt, as shown in (a). However, with " + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "inline_equation", + "content": "L_{pc}" + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "text", + "content": ", the motion better aligns with the prompt's trajectory, though it still includes unintended hand-waving motions unrelated to the prompt. When using multiple input motions, including " + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "inline_equation", + "content": "M_{2}" + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "text", + "content": " (a walking motion with arms crossed), taking the mean of features results in a blended motion, as shown in (c). This blending causes one arm to bend while the other hangs low, creating unnatural poses. In contrast, with CAF, " + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "inline_equation", + "content": "M_{2}" + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "text", + "content": " receives a higher weight, resulting in a more natural walking motion with arms crossed, as shown in (d)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 520, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 520, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 520, + 295, + 652 + ], + "type": "text", + "content": "Table 4 presents various ablation results related to adaptation. For the visual adaptation layer, Self-Attention proved to be the most effective compared to AdaIN and CrossAttn. Additionally, introducing text adaptation at Personalized Text Encoder improves the PRA, FID, and R-precision metrics. Fig. 6 demonstrates how PersonaBooth effectively captures and represents each actor's unique action. The red arrows specifically highlight the distinct arm movements of each persona. Additional details, including variations in body tilt and movement speed, are more clearly observable in the supplementary video materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 280, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 280, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 280, + 673 + ], + "type": "text", + "content": "5.2. Comparison with State-of-the-Art methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 677, + 295, + 714 + ], + "type": "text", + "content": "We compare our PersonaBooth with the state-of-the-art diffusion-based MST approaches, MoMo [25] and MCM-LDM [32], on the PerMo dataset. Additionally, we evalu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 244, + 555, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 555, + 328 + ], + "type": "text", + "content": "ate it on an existing style dataset, 100Style [19]. As MCM-LDM does not support prompt-based generation, we used MDM-generated motions [33] as target motions for evaluation. MCM-LDM was evaluated using the original model trained on HumanML3D, as well as finetuned models, referred to as MCM-LDM, which were trained on the PerMo and 100Style datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 333, + 406, + 343 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 406, + 343 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 406, + 343 + ], + "type": "text", + "content": "5.2.1. PerMo Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 347, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 502 + ], + "type": "text", + "content": "The comparative results for the PerMo dataset are shown in Table 5, with all subcategory results displayed for PRA. The single-input PersonaBooth significantly reduces FID compared to existing methods, indicating that while previous studies struggled with PerMo—a challenging dataset containing more diverse content than 100Style—PersonaBooth performs robust in generating high-quality motion. Additionally, PersonaBooth achieves the highest performance in R-Precision, PRA, and Diversity compared to existing methods. Applying multiple inputs and CAF further enhances performance. It's noteworthy that while MoMo and MCM-LDM utilized a 100-step and 1000-step diffusion model, respectively, we used a 50-step diffusion model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 503, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 556, + 624 + ], + "type": "text", + "content": "Fig. 7 shows qualitative results. MoMo [25] notes that their generated motions often follow the input motion's direction instead of the target's, which they attempted to copy the root rotation from the input motion (shown in (a) before and (b) after adjustment). The prompts for \"turns in the air\" were not met in (a), (b), or (c). Particularly, (c) does not reflect the style of ballerina. In contrast, PersonaBooth accurately generates the motion by first performing a hop that reflects the style, followed by a precise turn in the air. Example videos are provided in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 627, + 413, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 413, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 413, + 638 + ], + "type": "text", + "content": "5.2.2. 100Style Dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "Table 6 presents the comparative results on the 100Style dataset. MoMo demonstrates high PRA and Diversity, indicating effective style reflection. However, it shows very low R-precision, suggesting poor alignment between generated motions and prompts. Finetuned MCM-LDM improved performance on nearly all metrics compared to its" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22762" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 103, + 555, + 189 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 5. Comparison with the state-of-the-art methods on the PerMo dataset. The comparison is made in the Single Input (SI) setting as the existing methods do not support multiple inputs. MCM-LDM* indicates the model is finetuned on the PerMo dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 103, + 555, + 189 + ], + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 189 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 189 + ], + "type": "table", + "html": "
MethodsStepsFID ↓R Precision ↑PRA ↑Diversity ↑
Top 1Top 2Top 3AgeChar1Char2Cond1Cond2Emo1Emo2TraitSurAvg.
MoMo [25]10013.910.050.090.1328.2310.2311.308.7315.979.9714.5712.607.9313.476.74
MCM-LDM [32]10009.160.130.230.3047.3811.4410.347.6318.9011.6917.2015.719.2217.006.70
MCM-LDM* [32]10009.700.130.230.3046.4912.129.227.9519.7711.7518.0515.187.3316.766.76
PersonaBooth (SI)503.180.150.260.3348.0013.6711.976.6918.6714.3419.7517.068.8018.057.74
PersonaBooth (MI)502.950.190.300.3951.9514.8912.005.1622.959.4418.1415.4510.0918.138.12
", + "image_path": "d84afe707c8fa765c6116804b0871950538a73e935b650e244d333da447c3528.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 193, + 162, + 274 + ], + "blocks": [ + { + "bbox": [ + 56, + 193, + 162, + 274 + ], + "lines": [ + { + "bbox": [ + 56, + 193, + 162, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 162, + 274 + ], + "type": "image", + "image_path": "6c4fc08e65ce11eb7aaf80941e558b4b5fea3416c29c54c283c1aaa8a666d9c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 274, + 132, + 290 + ], + "lines": [ + { + "bbox": [ + 84, + 274, + 132, + 290 + ], + "spans": [ + { + "bbox": [ + 84, + 274, + 132, + 290 + ], + "type": "text", + "content": "Input Motion (Actor1 Ballerina)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 160, + 196, + 260, + 281 + ], + "blocks": [ + { + "bbox": [ + 160, + 196, + 260, + 281 + ], + "lines": [ + { + "bbox": [ + 160, + 196, + 260, + 281 + ], + "spans": [ + { + "bbox": [ + 160, + 196, + 260, + 281 + ], + "type": "image", + "image_path": "548ca00ac43886abd51863ab34976cf0a44b554855f2719fbc77784102f8a7e6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 281, + 255, + 290 + ], + "lines": [ + { + "bbox": [ + 160, + 281, + 255, + 290 + ], + "spans": [ + { + "bbox": [ + 160, + 281, + 255, + 290 + ], + "type": "text", + "content": "(a) MoMo (w/o copy root rot.)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 262, + 198, + 365, + 281 + ], + "blocks": [ + { + "bbox": [ + 262, + 198, + 365, + 281 + ], + "lines": [ + { + "bbox": [ + 262, + 198, + 365, + 281 + ], + "spans": [ + { + "bbox": [ + 262, + 198, + 365, + 281 + ], + "type": "image", + "image_path": "1acb58de0e6575b6d5c1d341037d3df68d2967cf0543955c1439f2bc276698ae.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 281, + 328, + 290 + ], + "lines": [ + { + "bbox": [ + 296, + 281, + 328, + 290 + ], + "spans": [ + { + "bbox": [ + 296, + 281, + 328, + 290 + ], + "type": "text", + "content": "(b) MoMo" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 365, + 196, + 463, + 281 + ], + "blocks": [ + { + "bbox": [ + 365, + 196, + 463, + 281 + ], + "lines": [ + { + "bbox": [ + 365, + 196, + 463, + 281 + ], + "spans": [ + { + "bbox": [ + 365, + 196, + 463, + 281 + ], + "type": "image", + "image_path": "d08c0965aadb5a54a987083b2d3816180587f79c269b926e90f53247e25d00d2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 281, + 435, + 290 + ], + "lines": [ + { + "bbox": [ + 385, + 281, + 435, + 290 + ], + "spans": [ + { + "bbox": [ + 385, + 281, + 435, + 290 + ], + "type": "text", + "content": "(c) MCM-LDM*" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 466, + 196, + 548, + 281 + ], + "blocks": [ + { + "bbox": [ + 466, + 196, + 548, + 281 + ], + "lines": [ + { + "bbox": [ + 466, + 196, + 548, + 281 + ], + "spans": [ + { + "bbox": [ + 466, + 196, + 548, + 281 + ], + "type": "image", + "image_path": "8dcf2acd08cef812a80589d4592ae4800a31e954e2cf3169b5650e953d17d1b4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 281, + 544, + 290 + ], + "lines": [ + { + "bbox": [ + 473, + 281, + 544, + 290 + ], + "spans": [ + { + "bbox": [ + 473, + 281, + 544, + 290 + ], + "type": "text", + "content": "(d) PersonaBooth (SI)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 56, + 312, + 156, + 388 + ], + "blocks": [ + { + "bbox": [ + 56, + 312, + 156, + 388 + ], + "lines": [ + { + "bbox": [ + 56, + 312, + 156, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 156, + 388 + ], + "type": "image", + "image_path": "57fa59f08fcc0be1c4672dcf6993f44d71ad5490f7a17f6e8c14fa043448f51b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 388, + 127, + 404 + ], + "lines": [ + { + "bbox": [ + 86, + 388, + 127, + 404 + ], + "spans": [ + { + "bbox": [ + 86, + 388, + 127, + 404 + ], + "type": "text", + "content": "Input Motion (Airplane)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 156, + 310, + 262, + 391 + ], + "blocks": [ + { + "bbox": [ + 91, + 297, + 517, + 308 + ], + "lines": [ + { + "bbox": [ + 91, + 297, + 517, + 308 + ], + "spans": [ + { + "bbox": [ + 91, + 297, + 517, + 308 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison on PerMo dataset. The input prompt is \"A person hops forward and turns in the air.\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 310, + 262, + 391 + ], + "lines": [ + { + "bbox": [ + 156, + 310, + 262, + 391 + ], + "spans": [ + { + "bbox": [ + 156, + 310, + 262, + 391 + ], + "type": "image", + "image_path": "5e9a864749f97a8a4e6adee2b53f0f5ec8174ecc96f67afd0da4278d6d8c21b5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 395, + 255, + 404 + ], + "lines": [ + { + "bbox": [ + 160, + 395, + 255, + 404 + ], + "spans": [ + { + "bbox": [ + 160, + 395, + 255, + 404 + ], + "type": "text", + "content": "(a) MoMo (w/o copy root rot.)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 263, + 321, + 365, + 392 + ], + "blocks": [ + { + "bbox": [ + 263, + 321, + 365, + 392 + ], + "lines": [ + { + "bbox": [ + 263, + 321, + 365, + 392 + ], + "spans": [ + { + "bbox": [ + 263, + 321, + 365, + 392 + ], + "type": "image", + "image_path": "560be798c05ca2f8bfea0febbdd79a5ae77151fa009a7b8955dfcf2dae5678a3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 395, + 328, + 403 + ], + "lines": [ + { + "bbox": [ + 296, + 395, + 328, + 403 + ], + "spans": [ + { + "bbox": [ + 296, + 395, + 328, + 403 + ], + "type": "text", + "content": "(b) MoMo" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 411, + 539, + 422 + ], + "lines": [ + { + "bbox": [ + 69, + 411, + 539, + 422 + ], + "spans": [ + { + "bbox": [ + 69, + 411, + 539, + 422 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison on 100Style dataset. The input prompt is \"A person walks backward and sits down on the chair.\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 364, + 311, + 449, + 392 + ], + "blocks": [ + { + "bbox": [ + 364, + 311, + 449, + 392 + ], + "lines": [ + { + "bbox": [ + 364, + 311, + 449, + 392 + ], + "spans": [ + { + "bbox": [ + 364, + 311, + 449, + 392 + ], + "type": "image", + "image_path": "1ae6bf02847599e9ad486805e27337755b1e45d693702dbad38bc86847aba36e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 395, + 435, + 403 + ], + "lines": [ + { + "bbox": [ + 386, + 395, + 435, + 403 + ], + "spans": [ + { + "bbox": [ + 386, + 395, + 435, + 403 + ], + "type": "text", + "content": "(c) MCM-LDM*" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 449, + 311, + 554, + 392 + ], + "blocks": [ + { + "bbox": [ + 449, + 311, + 554, + 392 + ], + "lines": [ + { + "bbox": [ + 449, + 311, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 449, + 311, + 554, + 392 + ], + "type": "image", + "image_path": "f827cad047c1c818a0e36d6e109082b3c130348292c75fe270b731676448446d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 395, + 544, + 404 + ], + "lines": [ + { + "bbox": [ + 473, + 395, + 544, + 404 + ], + "spans": [ + { + "bbox": [ + 473, + 395, + 544, + 404 + ], + "type": "text", + "content": "(d) PersonaBooth (SI)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 56, + 478, + 296, + 552 + ], + "blocks": [ + { + "bbox": [ + 55, + 436, + 295, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 468 + ], + "type": "text", + "content": "Table 6. Comparison with the state-of-the-art methods on the 100Style dataset. MCM-LDM* indicates the model is finetuned on the 100Style dataset" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 478, + 296, + 552 + ], + "lines": [ + { + "bbox": [ + 56, + 478, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 296, + 552 + ], + "type": "table", + "html": "
MethodsFID ↓R Precision ↑SRA ↑Diversity ↑
Top 1Top 2Top 3
MoMo [25]5.970.070.100.1460.577.82
MCM-LDM [32]7.200.180.290.3754.446.76
MCM-LDM* [32]6.530.170.280.3754.617.12
PersonaBooth (SI)3.270.200.310.4064.537.90
", + "image_path": "2f0ed0d874ea2558dbd54490fef16d39212112133d975950eb3ea1db42e593a2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 54, + 568, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 568, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 568, + 295, + 652 + ], + "type": "text", + "content": "original model. Although finetuned MCM-LDM achieved high R-precision, it also had a high FID, resulting in unnatural motion generation, with low PRA and Diversity indicating weaker style reflection. In contrast, our single-input PersonaBooth significantly reduces FID and achieves top performance across all metrics, including R-precision, PRA, and Diversity." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "In Fig. 8 (a) and (b), MoMo produces sliding motions that don't align with the prompt. MCM-LDM shows sideward movements. In contrast, PersonaBooth accurately reflects both the input motion's style, like arm extension and leaning, and the sitting motion in line with the prompt." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 436, + 388, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 436, + 388, + 449 + ], + "spans": [ + { + "bbox": [ + 314, + 436, + 388, + 449 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "text", + "content": "We propose the PerMo dataset and the PersonaBooth framework for Motion Personalization. Through contrastive learning, we effectively capture the essence of a persona. Furthermore, we enhance the integration of persona traits into diffusion models using a multi-modal finetuning method. We also introduce the CAF, which effectively handles multiple inputs, ensuring cohesive representation of personalized motion characteristics." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 552, + 556, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 556, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 556, + 672 + ], + "type": "text", + "content": "Limitations and Future Works. The PerMo contains relatively short motion sequences, which sometimes lead to stationary poses at the end of generated motions of the same length. However, this issue does not occur when prompted with longer actions, suggesting that automatic adjustment of motion length remains a task for future work. In addition, we plan to investigate approaches for applying CAF to each sequential action within a prompt individually. For instance, applying CAF separately to 'run' and 'jump' in the prompt \"A person runs and then jumps.\"" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 674, + 399, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 399, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 399, + 685 + ], + "type": "text", + "content": "Acknowledgement." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 686, + 554, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 686, + 554, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 686, + 554, + 711 + ], + "type": "text", + "content": "This work was supported by MSIT, Korea, under [RS-2022-00164860] and [RS-2024-00418641] supervised by IITP." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22763" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. ACM Transactions on Graphics (TOG), 39(4):64-1, 2020. 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "text", + "content": "[2] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 247 + ], + "type": "text", + "content": "[3] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 296, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 296, + 303 + ], + "type": "text", + "content": "[4] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 296, + 360 + ], + "type": "text", + "content": "[5] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3D human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 362, + 296, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 296, + 417 + ], + "type": "text", + "content": "[6] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 418, + 296, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 296, + 462 + ], + "type": "text", + "content": "[7] Xingzhe He, Zhiwen Cao, Nicholas Kolkin, Lantao Yu, Kun Wan, Helge Rhodin, and Ratheesh Kalarot. A data perspective on enhanced identity preservation for diffusion personalization. arXiv preprint arXiv:2311.04315, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 464, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 296, + 486 + ], + "type": "text", + "content": "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 296, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 530 + ], + "type": "text", + "content": "[9] Deok-Kyeong Jang, Soomin Park, and Sung-Hee Lee. Motion puzzle: Arbitrary motion style transfer by body part. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 586 + ], + "type": "text", + "content": "[10] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "text", + "content": "[11] Boeun Kim, Hyung Jin Chang, Jungho Kim, and Jin Young Choi. Global-local motion transformer for unsupervised skeleton-based action learning. In European Conference on Computer Vision, pages 209-225. Springer, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[12] Boeun Kim, Jungho Kim, Hyung Jin Chang, and Jin Young Choi. Most: Motion style transformer between diverse action contents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1705-1714, 2024. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Makito Kobayashi, Chen-Chieh Liao, Keito Inoue, Sentaro Yojima, and Masafumi Takahashi. Motion capture dataset" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "for practical use of ai-based motion editing and stylization. arXiv preprint arXiv:2306.08861, 2023.5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[14] Jehee Lee, Jinxiang Chai, Paul SA Reitsma, Jessica K Hodgins, and Nancy S Pollard. Interactive control of avatars animated with human motion data. In Proceedings of the 29th annual Conference on Computer Graphics and Interactive Techniques, pages 491-500, 2002. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[15] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 217 + ], + "type": "text", + "content": "[16] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 260 + ], + "type": "text", + "content": "[17] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 262, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 303 + ], + "type": "text", + "content": "[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, pages 5442-5451, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 306, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 358 + ], + "type": "text", + "content": "[19] Ian Mason, Sebastian Starke, and Taku Komura. Real-time style modelling of human locomotion via feature-wise transformations and local motion phases. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 5(1):1-18, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "text", + "content": "[20] Soomin Park, Deok-Kyeong Jang, and Sung-Hee Lee. *Diverse motion stylization for multiple style domains via spatial-temporal graph-based generative model.* Proceedings of the ACM on Computer Graphics and Interactive Techniques, 4(3):1-17, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[21] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "text", + "content": "[22] Mathis Petrovich, Michael J Black, and Gül Varol. Tmr: Text-to-motion retrieval using contrastive 3D human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9488-9497, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 505, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 553, + 536 + ], + "type": "text", + "content": "[23] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "text", + "content": "[24] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "type": "text", + "content": "[25] Sigal Raab, Inbar Gat, Nathan Sala, Guy Tevet, Rotem Shalev-Arkushin, Ohad Fried, Amit H Bermano, and Daniel Cohen-Or. Monkey see, monkey do: Harnessing self-attention in motion diffusion for zero-shot motion transfer. arXiv preprint arXiv:2406.06508, 2024. 2, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22764" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "text", + "content": "[29] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 207, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 294, + 272 + ], + "type": "text", + "content": "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "text", + "content": "[31] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 2, 4, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 396 + ], + "type": "text", + "content": "[32] Wenfeng Song, Xingliang Jin, Shuai Li, Chenglizhao Chen, Aimin Hao, Xia Hou, Ning Li, and Hong Qin. Arbitrary motion style transfer with multi-condition motion latent diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821-830, 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 397, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 294, + 441 + ], + "type": "text", + "content": "[33] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "text", + "content": "[34] Shihong Xia, Congyi Wang, Jinxiang Chai, and Jessica Hodgins. Realtime style transfer for unlabeled heterogeneous human motion. ACM Transactions on Graphics (TOG), 34 (4):1-10, 2015. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "text", + "content": "[35] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 532, + 294, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 294, + 585 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 294, + 585 + ], + "type": "text", + "content": "[36] Renrui Zhang, Jiaming Han, Chris Liu, Peng Gao, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 587, + 294, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 294, + 631 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 294, + 631 + ], + "type": "text", + "content": "[37] Xulu Zhang, Xiao-Yong Wei, Wengyu Zhang, Jinlin Wu, Zhaoxiang Zhang, Zhen Lei, and Qing Li. A survey on personalized content synthesis with diffusion models. arXiv preprint arXiv:2405.05538, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 632, + 294, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 632, + 294, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 632, + 294, + 665 + ], + "type": "text", + "content": "[38] Lei Zhong, Yiming Xie, Varun Jampani, Deqing Sun, and Huaizu Jiang. Smoodi: Stylized motion diffusion model. arXiv preprint arXiv:2407.12783, 2024. 2, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22765" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/00e10d7f-1169-4801-84c3-c1efaaab378c_content_list.json b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/00e10d7f-1169-4801-84c3-c1efaaab378c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c1125022cc1cbb3114c2db90d2ba9b0adad0e4f9 --- /dev/null +++ b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/00e10d7f-1169-4801-84c3-c1efaaab378c_content_list.json @@ -0,0 +1,1997 @@ +[ + { + "type": "text", + "text": "PersonaHOI: Effortlessly Improving Face Personalization in Human-Object Interaction Generation", + "text_level": 1, + "bbox": [ + 207, + 130, + 790, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinting Hu*", + "bbox": [ + 158, + 204, + 259, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xhu@mpi-inf.mpg.de", + "bbox": [ + 101, + 224, + 264, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoran Wang*", + "bbox": [ + 326, + 204, + 444, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hawang@mpi-inf.mpg.de", + "bbox": [ + 282, + 224, + 470, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jan Eric Lenssen", + "bbox": [ + 511, + 204, + 648, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jlenssen@mpi-inf.mpg.de", + "bbox": [ + 493, + 224, + 697, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bernt Schiele", + "bbox": [ + 746, + 204, + 857, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "schiele@mpi-inf.mpg.de", + "bbox": [ + 718, + 224, + 913, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Max Planck Institute for Informatics, Saarland Informatics Campus, Germany", + "bbox": [ + 199, + 246, + 815, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Context + HOI", + "text_level": 1, + "bbox": [ + 250, + 284, + 403, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A man riding a horse in front of a medieval castle", + "bbox": [ + 161, + 301, + 455, + 314 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b157e54b54f7e59e72c4458fd80cbff15435b7be97458c9d9108f2e545625445.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 316, + 254, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eee15c6a3d7f2f66d0fe64bb09f85357a7a22b1765f09202b4ca9899a2e8dc7c.jpg", + "image_caption": [ + "FastComposer" + ], + "image_footnote": [], + "bbox": [ + 282, + 315, + 377, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7b41c0e0d4124d394e79e4ca3536a1d6653346a3418912f9d207e1de6ed09cc.jpg", + "image_caption": [ + "+Ours" + ], + "image_footnote": [], + "bbox": [ + 387, + 315, + 485, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Style + HOI", + "text_level": 1, + "bbox": [ + 611, + 281, + 745, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A woman holding a cat in the style of Vincent Van Gogh", + "bbox": [ + 524, + 303, + 839, + 315 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7435031c37e19ec29770cf86db560bcce83ba502dd54db71fd0c9938e201e473.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 316, + 607, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/52fd20b7f4bdff5b304a657d125772808fe19d54d661aa9f3d435952bf704d63.jpg", + "image_caption": [ + "FastComposer" + ], + "image_footnote": [], + "bbox": [ + 627, + 316, + 723, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/320948f929912f5425969af98881dfbdb9780f8f9d4101acbe1558403e04cd82.jpg", + "image_caption": [ + "+Ours" + ], + "image_footnote": [], + "bbox": [ + 733, + 316, + 830, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/247ba23b3aa2cf8aedd490ce8aa68908e1ba25b335540164fe2bcc36869980d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 173, + 416, + 230, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab7c40c3cf4033effea10b05a9a3a6eefde85db862f49281f0c995aef0658d10.jpg", + "image_caption": [ + "Various man-Object action (HOI) descriptions" + ], + "image_footnote": [], + "bbox": [ + 243, + 453, + 269, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95dd086b65595a579d3d425c6609f565a3eda7f96aba4b06315864382c6834dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 412, + 346, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1cfa0df5c025adff299d9824ef2eb90e13a8db98b4b547113d3d8783195ebc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 412, + 415, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71f314f3842b0aa7dc66879f5b8a05ec49b4b5049469ab393b694a608d44ac77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 414, + 483, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ea63ada345a13ba311df7292efa55d208d946c220d7f6951aa595a6e0a2438c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 463, + 346, + 512 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a9f97ef19f0729a0e4aac6e809f40bd6a55abf839ae1a6219b9e08b6f8a40aae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 463, + 415, + 512 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3dedd8517a0228a615656366ae664bc76fd2c3aa79f5b8aed21058ddc60439f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 463, + 483, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Personalized", + "text_level": 1, + "bbox": [ + 526, + 417, + 702, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face with", + "bbox": [ + 539, + 445, + 686, + 464 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human-Object", + "bbox": [ + 540, + 468, + 684, + 487 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interaction", + "bbox": [ + 560, + 487, + 692, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training", + "bbox": [ + 738, + 426, + 810, + 443 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "&", + "bbox": [ + 764, + 445, + 782, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tuning", + "bbox": [ + 743, + 463, + 805, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Free!", + "bbox": [ + 750, + 481, + 797, + 497 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accessory + HOI", + "text_level": 1, + "bbox": [ + 238, + 527, + 411, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A woman wearing a Santa hat and taking selfie with phone", + "bbox": [ + 156, + 549, + 493, + 559 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc0f5f8077b75ab9c225f3f0cb46c7ed55d42813cb81362f2beca721fbafc31c.jpg", + "image_caption": [ + "Figure 1. Examples of Personalized Face with Human-Object Interaction (HOI) Generation. We present PersonaHOI, a training- and tuning-free framework built on existing diffusion models. Using a single reference image and diverse HOI prompts, PersonaHOI generates identity-consistent human-object interactions compared to FastComposer [35]. PersonaHOI can further seamlessly integrate varied contexts, styles, accessories, and multi-person scenarios, ensuring scalability and practicality for real-world applications." + ], + "image_footnote": [], + "bbox": [ + 158, + 561, + 256, + 636 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/88782716e9e95372ba6ba9459feb421f25b10a7b26fdf4d4fc7dd3761727640f.jpg", + "image_caption": [ + "FastComposer" + ], + "image_footnote": [], + "bbox": [ + 292, + 561, + 387, + 635 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4185e7907c4f3d2a7e93cdbb4b654b9ddd9d1743380b26001350fefc76257d0b.jpg", + "image_caption": [ + "+Ours" + ], + "image_footnote": [], + "bbox": [ + 397, + 561, + 493, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mult-Person + HOI", + "text_level": 1, + "bbox": [ + 596, + 527, + 772, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A man and a man are playing guitar in a bar", + "bbox": [ + 537, + 544, + 800, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db2c67893599da98510a6773350c766abb93adb5aead9340d04282c538714726.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 534, + 556, + 586, + 633 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/11fdfc71f8f8d066b87ca4a669ea495afb8320c9778279662b2a875a9f49a960.jpg", + "image_caption": [ + "FastComposer" + ], + "image_footnote": [], + "bbox": [ + 614, + 556, + 720, + 635 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a667851efcf6131abd92aa22122c5ed3199961961de8a62edb9ab49dff6607a7.jpg", + "image_caption": [ + "+Ours" + ], + "image_footnote": [], + "bbox": [ + 730, + 556, + 834, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 728, + 325, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce PersonaHOI, a training- and tuning-free framework that fuses a general StableDiffusion model with a personalized face diffusion (PFD) model to generate identity-consistent human-object interaction (HOI) images. While existing PFD models have advanced significantly, they often overemphasize facial features at the expense of full-body coherence, PersonaHOI introduces an additional", + "bbox": [ + 86, + 762, + 485, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "StableDiffusion (SD) branch guided by HOI-oriented text inputs. By incorporating cross-attention constraints in the PFD branch and spatial merging at both latent and residual levels, PersonaHOI preserves personalized facial details while ensuring interactive non-facial regions. Experiments, validated by a novel interaction alignment metric, demonstrate the superior realism and scalability of PersonaHOI, establishing a new standard for practical personalized face with HOI generation. Code is available at https://github.com/JoyHuYY1412/PersonaHOI.", + "bbox": [ + 511, + 729, + 908, + 881 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 107, + 887, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23775", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd865b7a270e59debd3e386ac1a78d3ee6ace7beac4b9887245343fa0cc04954.jpg", + "image_caption": [ + "Figure 2. (a) The spatial layout of StableDiffusion guides PersonaHOI to generate personalized content with coherent human-object interactions (HOI). (b) Analysis of identity injection timing in PFD models. We use FastComposer [35] for diffusion model generation. Injecting face representation at the start of image generation preserves facial details but lacks coherent HOI, while delayed injection continuously deviates from the original identity, resulting in random human features and meaningless human-object interactions." + ], + "image_footnote": [], + "bbox": [ + 102, + 92, + 897, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 377, + 223, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Personalized face generation has seen increasing public interest and demand in user-specified digital content creation. Current learning-based personalized face diffusion (PFD) models [18, 21, 29, 35], trained on large-scale face-centric datasets, can incorporate a single user-provided reference facial image into a text-to-image generation process, enabling the rapid creation of images that depict specific persons in diverse scenes, outfits, and styles within seconds.", + "bbox": [ + 89, + 407, + 483, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although these methods perform well in simpler scenarios, they often struggle with generating full-body depictions involving detailed human-object interactions (HOI). As illustrated in Figure 1, the generated images of FastComposer [35] show missing objects or body parts, resulting in portrait-focused outputs without HOI information. This limitation compromises the overall realism and practical utility of the generated content, especially in immersive or interactive applications.", + "bbox": [ + 89, + 532, + 483, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To diagnose the limitations of PFD models, we examine the results of PFD models by varying the timing of identity injection during the generation process. As seen in the first and last columns of Figure 2(b), early injection effectively preserves facial details from the reference image, but relying solely on text input (last column) results in incoherent HOI outputs. This analysis highlights that the core challenge lies not in maintaining identity fidelity but in generating coherent body movements and interactions driven by text prompts. Notably, pre-trained models used for PFD, such as StableDiffusion [25], can generate satisfactory HOI images due to their extensive and diverse training data. This observation implies that the fine-tuning process on face-centric datasets of PFD models diminishes their ability to follow complex HOI text prompts. Therefore, to advance", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "current PFD methods in HOI tasks, reintroducing the capability to leverage text prompts for natural full-body interactions is essential.", + "bbox": [ + 511, + 378, + 906, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by this need, we propose a straightforward yet powerful approach: leveraging the generative capabilities of pre-trained diffusion models, such as StableDiffusion (SD), to augment existing PFD frameworks. Our method, PersonaHOI, integrates the strengths of SD without requiring additional training or fine-tuning, thereby restoring the ability to generate realistic, text-driven full-body HOI while preserving identity-specific details from the reference image.", + "bbox": [ + 511, + 426, + 908, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Central to this integration, we incorporate an additional SD branch that aligns identity-specific details from PFD models with the HOI layouts generated by SD. A head segmentation mask derived from the SD output guides the merge of non-facial components from the SD branch with the facial details from the PFD branch. Specifically, in the PFD model branch, we introduce a Cross-Attention Constraint to prevent the overemphasis of identity features across the entire image, ensuring these details are confined to the facial region. To integrate HOI representations from SD, we implement Latent Merge and Residual Merge, merging identity and contextual features at both the latent representation level and via skip connections within the U-Net architecture. This multi-level merging strategy ensures that the generated images maintain realistic human-object interactions while preserving personalized facial details, resulting in cohesive, interaction-rich outputs.", + "bbox": [ + 511, + 550, + 908, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our method in HOI scenarios, we propose a novel \"interaction alignment\" metric. These metrics leverage HOI detectors to measure alignment between generated images and text prompts, objectively assessing interaction realism. Our contributions can be summarized as follows:", + "bbox": [ + 511, + 810, + 908, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23776", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present PersonaHOI, a dual-path architecture that integrates general and personalized diffusion models. Our framework generates realistic human-object interactions with specific identities, without requiring additional training or test-time tuning.", + "- Our method introduces key components, including crossattention constraints and a combination of latent and residual merge strategies. These innovations enable effective feature integration, ensuring that generated images maintain both interaction coherence and identity fidelity.", + "- We design a novel evaluation metric for human-object interaction. Our experiments show significant improvements over state-of-the-art face personalization techniques in terms of interaction realism across models, emphasizing the scalability and robustness of our approach." + ], + "bbox": [ + 88, + 90, + 480, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 330, + 240, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Personalized Face Generation Diffusion Models. Personalized face generation aims to create identity-consistent images of individuals from limited reference data [32, 36, 43]. Given pretrained general image generation models [25], traditional optimization-based approaches [5, 26, 27] refine model parameters during inference for each individual but are computationally expensive, making them unsuitable for real-time applications. In contrast, learning-based approaches leverage human-centric datasets [1, 15, 19, 20] to train robust models capable of varying identity inputs without requiring test-time fine-tuning. Facial representations, extracted by models like CLIP [23] or other face recognition models [3, 16, 33], are often fused with [18, 21, 35] or replace [30] textual embeddings of specific words, such as \"person\", to provide identity conditions for diffusion models and enhance contextual relevance [3, 29]. In IP-Adapter [37] and other works [7, 9, 34], facial features are processed separately through decoupled cross-attention modules. Incorporating human masks as prior, methods can clean the background during data processing; alternatively, masks can be leveraged to construct [2, 18, 21, 35] and adjust [12] loss functions during training. In our work, we extend face personalization to complex human-object interactions (HOI), achieving a higher level of personalization and adaptability for diverse generative tasks.", + "bbox": [ + 91, + 357, + 483, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human-Object Interaction in Diffusion Models. HOI detection aims to locate human-object pairs and categorize their interactions as triplets, such as (person, playing, football) [4, 38-40], while HOI image synthesis remains relatively under-explored. InteractGAN [6] generates HOI images using human pose templates and reference images, but its reliance on pose-template pools and object references limits flexibility. Recent diffusion-based approaches tackle HOI generation by incorporating layout controls such as bounding boxes [10, 14, 17] or human poses [17, 42], or by leveraging in-context samples with similar interac", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions [11, 44]. Additionally, ReCorD [14] integrates Latent Diffusion Models with Visual Language Models, introducing modules for reasoning and correcting interactions to enhance cross-modal alignment. In this work, we rely on the standard text-to-image diffusion model (i.e., StableDiffusion [25]) for HOI generation, and fuse the generation process with personalized face diffusion models to generation identity-preserving HOI images.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 224, + 638, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Understanding the fundamentals of StableDiffusion (SD) is essential to leverage its generative capabilities within our PFG framework. SD is a latent diffusion model designed to generate high-quality images based on text or image prompts. It operates in a latent space, where an input image is encoded in a latent representation $\\mathbf{z}_0$ using a pretrained variational autoencoder (VAE) encoder, reducing the dimensionality of the image for efficient processing.", + "bbox": [ + 511, + 250, + 903, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the generation phase, the SD model begins with an initial noisy latent representation $\\mathbf{z}_T$ and progressively denoises it over $T$ timesteps to reconstruct $\\mathbf{z}_0$ . This denoised latent is then decoded by the VAE decoder to produce the final image. At each timestep $t$ , the U-Net architecture $\\theta$ predicts noise $\\epsilon (\\mathbf{z}_t,t,\\mathbf{C})$ , where $\\mathbf{C}$ represents the conditioning inputs (text or image embeddings). The iterative denoising process follows:", + "bbox": [ + 511, + 371, + 903, + 492 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t - 1} = \\operatorname {D e n o i s e} \\left(\\mathbf {z} _ {t}, \\epsilon \\left(\\mathbf {z} _ {t}, t, \\mathbf {C}\\right); \\theta\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 503, + 903, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The U-Net architecture comprises an encoder-decoder structure with skip connections that transfer features between the encoder and the decoder. Within this architecture, SD employs a cross-attention mechanism to incorporate conditioning inputs effectively during denoising. This mechanism projects latent representations and conditioning inputs into query, key, and value spaces:", + "bbox": [ + 511, + 532, + 903, + 637 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathbf {Q} = \\mathbf {W} _ {q} \\mathbf {z}, \\mathbf {K} = \\mathbf {W} _ {k} \\mathbf {C}, \\mathbf {V} = \\mathbf {W} _ {v} \\mathbf {C}; \\\\ \\mathbf {A} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right), \\mathbf {z} _ {\\text {a t t n}} = \\mathbf {A V}. \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 646, + 903, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\mathbf{A}$ is the attention map that ensures that relevant conditioning tokens guide denoising at each generation step. This cross-attention mechanism is critical for aligning the generated image with the input prompts, enabling consistency and fidelity in the final output.", + "bbox": [ + 511, + 704, + 903, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 792, + 604, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 511, + 816, + 709, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 3, PersonaHOI augments the PFD model by integrating an additional StableDiffusion (SD) branch, enhancing its capacity to generate personalized images that capture complex human-object interactions (HOI)", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23777", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0bfaab5951f489661d45979cb536bf2ced2fa25a52626df45a24eec4946d2160.jpg", + "image_caption": [ + "Figure 3. Overview of Our Proposed Framework, PersonaHOI. The architecture integrates a personalized face diffusion (PFD) model with an additional StableDiffusion (SD) branch. First, SD generates an image $(I_{SD})$ from a text prompt and noisy latent representation $(z_{T})$ , which is decoded and segmented to produce a head mask. Next, SD and PFD run in parallel from the same $z_{T}$ . At every timestep $t$ , the head mask guides the Cross-Attention Constraint in PFD and merging modules (Latent Merge and Residual Merge) to merge interaction-relevant features from SD with identity-specific details from PFD. Iteratively, this process introduces HOI context to personalized face generation in a training&tuning-free manner." + ], + "image_footnote": [], + "bbox": [ + 158, + 85, + 848, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in a training-free manner. Given an input comprising a user-provided reference image and a text prompt specifying the interaction (e.g., \"a person kicking football\"), our framework produces a cohesive output image that retains the subject's identity while accurately depicting the specified HOI.", + "bbox": [ + 88, + 498, + 480, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To effectively merge identity from PFD and interaction features from SD, we introduce three core strategies: Cross-Attention Constraint, which regulates attention to reference image features within the PFD model (detailed in Sec. 4.2); Latent Merge, which combines features in the latent space at each generation timestep (Sec. 4.3); and Residual Merge, which integrates identity details through skip connections in the U-Net architecture (Sec. 4.4). Together, these strategies enable PersonaHOI to align generated content with interaction-driven text prompts while maintaining identity fidelity. Furthermore, this design allows PersonaHOI adaptable to various PFD models without the need for model retraining or test-time tuning.", + "bbox": [ + 88, + 575, + 482, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Cross-Attention Constraint (CAC)", + "text_level": 1, + "bbox": [ + 89, + 786, + 390, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the cross-attention layers of the PFD model, conditioning inputs $\\mathbf{C} \\in \\mathbb{R}^{N \\times D}$ include both text and image embeddings to guide the generation process, where $N$ is the number of tokens, and $D$ is the embedding dimension. Following established PFD model approaches [18, 21, 35], the $n_{img}$ -th token in $\\mathbf{C}$ incorporates the image embedding, encapsulat", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing identity-specific details essential for facial preservation, while the remaining $N - 1$ tokens correspond to text embeddings. In standard PFD models, attention to the $n_{img}$ -th token often spreads across the image, overemphasizing facial features thus decreasing spatial capacity for representing body movements and object interactions.", + "bbox": [ + 511, + 498, + 903, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we introduce Cross-Attention Constraint (CAC) within the cross-attention layers of the PFD model. CAC restricts identity features to specific facial regions, ensuring sufficient spatial capacity for body and object interactions. Different from the attention localization loss used in training [21, 35], our CAC utilizes fine-grained head mask derived from the SD-generated image $I_{SD}$ in image generation process, ensuring the attention constraint in PFD is consistent with the HOI layout generated by SD. The head mask is segmented by an off-the-shelf head segmentor [24], then resized to match the spatial dimensions $(H \\times W)$ of the latent representation $\\mathbf{z}$ , denoted as $\\mathbf{M}^{\\mathrm{head}} \\in [0,1]^{H \\times W}$ (1 for the head region and 0 for other regions). We apply $\\mathbf{M}^{\\mathrm{head}}$ to the attention map $\\mathbf{A} \\in \\mathbb{R}^{(H \\times W) \\times N}$ as follows:", + "bbox": [ + 511, + 589, + 906, + 815 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i \\in [ 0, 1, \\dots , N - 1 ]} ^ {C A C} = \\left\\{ \\begin{array}{l l} \\text {f l a t t e n} (\\mathbf {M} ^ {\\text {h e a d}}), & \\text {i f} i = i m g, \\\\ 1, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 816, + 903, + 858 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} ^ {C A C} = \\mathbf {A} \\odot (\\mathbf {M} ^ {C A C}) ^ {\\intercal}.\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 859, + 769, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{M}^{CAC}$ modifies the attention weights for the $n_{img}$ token", + "bbox": [ + 511, + 883, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23778", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/91e0629e18e3c345132d2a0ab10c395fffbb083ed48a7dc74dd08cdbb647fb69.jpg", + "image_caption": [ + "Figure 4. Illustration of Residual Merge. In each residual layer, Residual Merge operates within the U-Net skip connections, utilizing a head mask to guide the integration of high-frequency identity details from PFD residuals and low-frequency interaction layouts from SD residuals. The merged residuals are then concatenated to the corresponding bottleneck features from PFD." + ], + "image_footnote": [], + "bbox": [ + 109, + 90, + 472, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by setting the weights to 0 in regions outside the head, preventing the influence of human identity on non-facial areas. Since the head mask is derived from SD-generated outputs, CAC maintains facial layout coherence with SD. When further combined with Latent and Residual Merge, CAC helps blend facial features from PFD and non-facial interaction features from SD without mutual interference.", + "bbox": [ + 89, + 362, + 483, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Latent Merge (LM)", + "text_level": 1, + "bbox": [ + 89, + 479, + 277, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce Latent Merge, which directly merges the latent representation within U-Net from PFD and SD models. After applying Cross-Attention Constraint (CAC) in Sec. 4.2, the facial region generated by the PFD model becomes spatially aligned with the corresponding region in the broader HOI layout produced by the SD model. This alignment facilitates spatial merge, blending facial identity details from PFD with interaction contexts from SD in latent space.", + "bbox": [ + 89, + 503, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At each diffusion timestep $t$ of PFD and SD, we implement the latent space merging with $\\mathbf{M}^{head}$ as follows:", + "bbox": [ + 89, + 638, + 482, + 671 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t} = \\mathbf {M} ^ {h e a d} \\odot \\mathbf {z} _ {t} ^ {\\mathrm {P F D}} + \\left(1 - \\mathbf {M} ^ {h e a d}\\right) \\odot \\mathbf {z} _ {t} ^ {\\mathrm {S D}}. \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 681, + 482, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{z}_t^{\\mathrm{PFD}}$ and $\\mathbf{z}_t^{\\mathrm{SD}}$ represents the latent features from PFD and SD at timestep $t$ , respectively. The merged latent $\\mathbf{z}_t$ then serves as input for both SD and PFD models in the next denoising timestep $t - 1$ . This merging strategy allows identity details retained within facial regions, while non-facial regions maintain the HOI context generated by SD.", + "bbox": [ + 89, + 713, + 483, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Residual Merge (RM)", + "text_level": 1, + "bbox": [ + 89, + 832, + 294, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While Latent Merge enables interactions between the SD and PFD at the latent level, it lacks integration at the intermediate feature stage within the U-Net. To address this,", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we introduce Residual Merge, which enhances feature integration within the skip connections. In the U-Net, skip connections transmit residual features containing rich detail from the encoder to the decoder, significantly impacting the content and quality of the generated images [13, 31].", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our Residual Merge is applied to each skip connection layer $l$ ( $l = 1, \\dots, L$ ), as shown in Figure 4. First, residual features are extracted from both the PFD and SD U-Net backbones. A low-pass filter and high-pass filter is applied to the SD and PFD residuals, respectively. Low-pass and high-pass filters are applied to these residuals to balance global coherence and local precision: low-pass filtering SD residuals suppresses high-frequency noise to retain the broader scene layout, while high-pass filtering PFD residuals emphasizes fine-grained identity-specific details. To achieve precise integration between facial and non-facial regions, we apply the head segmentation mask $\\mathbf{M}^{\\mathrm{head}}$ , resized to match the residual resolution at layer $l$ , denoted as $\\mathbf{M}_{\\mathrm{R}}^{l}$ . The fusion of residual features from both paths is then implemented for each layer $l$ as follows:", + "bbox": [ + 511, + 167, + 906, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {\\text {m e r g e d}} ^ {l} = \\mathbf {M} _ {R} ^ {l} \\odot \\mathrm {H P} \\left(\\mathbf {R} _ {\\text {P F D}} ^ {l}\\right) + \\left(1 - \\mathbf {M} _ {R} ^ {l}\\right) \\odot \\mathrm {L P} \\left(\\mathbf {R} _ {\\text {S D}} ^ {l}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 405, + 906, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{HP}(\\cdot)$ and $\\mathrm{LP}(\\cdot)$ denote the high pass and low pass filter, $\\mathbf{R}_{\\mathrm{PFD}}$ and $\\mathbf{R}_{\\mathrm{SD}}$ denote the residual features from the PFD and SD paths, respectively. The merged residual $\\mathbf{R}_{\\mathrm{merged}}$ is then concatenated with the bottom-up features in the PFD U-Net for further processing.", + "bbox": [ + 511, + 435, + 905, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Residual Merge strategy enables fine-grained feature integration within the U-Net. Through this, facial regions retain high-frequency identity details from PFD while non-facial regions incorporate interaction layouts from SD, ensuring cohesive and contextually aligned outputs.", + "bbox": [ + 511, + 513, + 905, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 603, + 645, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 511, + 628, + 594, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Data. We use two test sets for evaluation:", + "bbox": [ + 511, + 654, + 870, + 667 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) HOI-Focused Personalized Face Generation: To test complex human-object interactions (HOI), we use interaction labels from the widely-used V-COCO [8] dataset. V-COCO provides diverse HOI scenarios for robust evaluation of interaction fidelity. We construct 30 prompts in the \"subject + interaction + object\" format (e.g., \"man kicking a ball\") with the same 15 reference subjects as in 1).", + "2) General Personalized Face Generation: Following prior studies [21, 35], we compile a benchmark with 15 reference subjects, each paired with 40 prompts spanning diverse scenarios like context, stylization, accessory, and actions. These prompts aim to assess both identity retention and the model's ability to adapt to various text-guided settings." + ], + "bbox": [ + 511, + 669, + 905, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Full prompt details are provided in the Appendix ??", + "bbox": [ + 511, + 868, + 859, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We employ the following evaluation metrics:", + "bbox": [ + 511, + 885, + 872, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23779", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/402827fbd901cebc655b5e14a31dd6590f49b959dacdde163a9b2a7ad51df308.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodArchitectureIdentity Preservation (%)Prompt Consistency (%)HOI Alignment (%)
StableDiffusion v1.5 [25]SD1.5NA24.9660.23
StableDiffusion XL [22]SDXLNA25.3270.78
FastComposer [35]SD1.553.5721.3035.96
+ OursSD1.555.28(+1.71)23.16(+1.86)56.65(+20.69)
IP-Adapter [37]SDXL62.8622.7349.83
+ OursSDXL55.74(-7.11)24.20(+1.47)68.30(+18.47)
PhotoMaker [18]SDXL50.5423.9656.77
+ OursSDXL51.34(+0.80)24.34(+0.38)76.01(+19.24)
", + "bbox": [ + 91, + 88, + 906, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of Our Method with Baseline Approaches on HOI-Specific Personalized Face Generation. StableDiffusion serves as the text-only baseline without subject conditioning. PersonaHOI seamlessly incorporates existing Personalized Face Diffusion models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) with their corresponding StableDiffusion architectures. We bold the higher number for each pair of comparison.", + "bbox": [ + 89, + 253, + 906, + 310 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c25e15c29661e3c6175bbf9f426a5082b5f09d14764dc55613f6f9362e9d0866.jpg", + "image_caption": [ + "Figure 5. Qualitative Examples of PersonaHOI and Baseline Models. Comparison of baseline models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) and their PersonaHOI-enhanced results for diverse human-object interaction prompts." + ], + "image_footnote": [], + "bbox": [ + 107, + 328, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Identity Preservation: Measures how well the generated images retain the identity of the reference image. Face detection is performed using MTCNN [41], and identity similarity is calculated with FaceNet [28], following [21, 35]. 2) Prompt Consistency: Evaluates text-image alignment using CLIP-based [23] scores, where higher values indicate better prompt adherence.", + "3) HOI Alignment (Ours): Assesses how well the gen" + ], + "bbox": [ + 89, + 780, + 480, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "erated images depict specified human-object interactions. We adopt UPT [40], an HOI detector pre-trained on V-COCO dataset, to identify the interaction specified in text prompts for generated images with HOI alignment score $S_{\\mathrm{HOI}}$ . Specifically, UPT provides confidence scores for detected human and object classes ( $S_{\\mathrm{Human}}$ , $S_{\\mathrm{Object}}$ ) from its object detection head, as well as interaction logits ( $L_{\\mathrm{Inter}}$ ) from its interaction head. With these outputs, HOI align-", + "bbox": [ + 511, + 780, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23780", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ment score $S_{\\mathrm{HOI}}$ is defined as:", + "bbox": [ + 89, + 90, + 292, + 104 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\mathrm {H O I}} = \\left(S _ {\\text {H u m a n}} \\cdot S _ {\\text {O b j e c t}}\\right) ^ {\\lambda} \\cdot \\sigma \\left(L _ {\\text {I n t e r}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 109, + 482, + 128 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\sigma$ is the sigmoid activation function and $\\lambda = 2.8$ . We validate the human alignment and effectiveness for our proposed HOI Alignment metric in Appendix ??", + "bbox": [ + 89, + 137, + 482, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared Methods and Implementation Details. We compare our method with state-of-the-art learning-based personalized face diffusion models, including FastComposer [35], IP-Adapter [37], and PhotoMaker [18]. Our PeronaHOI used their pre-trained diffusion backbones as the additional branch, i.e., StableDiffusion v1.5 for FastComposer and StableDiffusion XL for PhotoMaker, with output resolutions set to $512 \\times 512$ and $1024 \\times 1024$ , respectively. Head segmentation for SD generated image is performed with DensePose [24] due to its robust segmentation accuracy. In our generation process, no additional training or test-time tuning is applied to any compared methods.", + "bbox": [ + 89, + 185, + 483, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Personalized Face with HOI Generation", + "text_level": 1, + "bbox": [ + 89, + 375, + 433, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. Figure 5 illustrates qualitative comparisons between baseline models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) and their PersonaHOI-enhanced counterparts for diverse human-object interaction prompts. Baseline methods often fail to produce meaningful interactions, leading to missing or poorly integrated objects and unnatural human-object dynamics. For instance, in the prompt \"A man skateboarding with skateboard\" (first row of Figure 5), FastComposer [35] and PhotoMaker omit the skateboard entirely and IP-Adapter fails to generate the accurate interaction (\"holding\" the skateboard instead), resulting in inconsistent outputs with the text prompt. In contrast, PersonaHOI-enhanced models generate realistic skateboard placement and natural human-object dynamics. Moreover, PersonaHOI also ensures high-fidelity preservation of facial identity given in reference images.", + "bbox": [ + 88, + 398, + 482, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. Table ?? compares the performance of PersonaHOI across various baseline models. For FastComposer [35] and PhotoMaker [18], PersonaHOI achieves consistent improvements across all metrics. Notably, Interaction Alignment improves by $20.69\\%$ for FastComposer [35] and $19.24\\%$ for PhotoMaker [18], demonstrating superior human-object interaction coherence. Gains in Identity Preservation and Prompt Consistency further validate PersonaHOI's ability to retain identity while enhancing text-image alignment. For IP-Adapter [37], a decrease in Identity Preservation (from $62.86\\%$ to $55.74\\%$ ) is observed. This drop is attributed to IP-Adapter's inherent bias toward generating face-centric images, which inflates identity scores at the cost of interaction fidelity. PersonaHOI mitigates this trade-off by significantly improving Interaction Alignment $(+18.47\\%)$ and Prompt Consistency $(+1.47\\%)$ , balancing identity preser", + "bbox": [ + 89, + 643, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bdbe366dc80e66c0d1f34dec8e05c7e98318476de53bf0c1cb2ba4579b6c4bb.jpg", + "image_caption": [ + "SD v1.5" + ], + "image_footnote": [], + "bbox": [ + 521, + 107, + 598, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c7ab5efea08c132040a60736f44be0a123040bc8a3306fa06d4f8ab0c2d6aba.jpg", + "image_caption": [ + "Reference Images" + ], + "image_footnote": [], + "bbox": [ + 606, + 107, + 679, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d1c715e371185d08bb0e98d4eab87bfd1ba5dfb9b19e4e9ffb9eed581844a28.jpg", + "image_caption": [ + "ce Images" + ], + "image_footnote": [], + "bbox": [ + 684, + 107, + 759, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01f50b740e50ab61d8587932a7128dac0de71e3945069cfaa5133e81453d5728.jpg", + "image_caption": [ + "FastComposer" + ], + "image_footnote": [], + "bbox": [ + 766, + 107, + 843, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d188e5707d7e956c4753e1241c1af0dd4bd77e961969e79aca207cc8b0539f4a.jpg", + "image_caption": [ + "+Ours" + ], + "image_footnote": [], + "bbox": [ + 846, + 107, + 921, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75441ba04fcf1bbe5df8df5022fd18409f1cacd8e9aa466b7c6399187666c300.jpg", + "image_caption": [ + "Caption: A man and a man planting a tree" + ], + "image_footnote": [], + "bbox": [ + 521, + 179, + 596, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3b100927aa3fd3b79fffd8c1a4a7e09c25b936a4e4083643a8e9028659e81ed.jpg", + "image_caption": [ + "Caption: A man and a man cycling together" + ], + "image_footnote": [], + "bbox": [ + 609, + 179, + 676, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c18d9f62666ffd918d33e408363154d6e2ecefc1885cffb797d9e2e159453a39.jpg", + "image_caption": [ + "Figure 6. Qualitative results for multi-subject HOI generation. We compare generation outputs from SD v1.5 [25], FastComposer [35], and our PersonaHOI based on FastComposer [35] with different multi-subject interaction prompts. PersonaHOI not only preserves distinct identities but also generates coherent human-object interactions aligned with HOI layouts produced by SD v1.5." + ], + "image_footnote": [], + "bbox": [ + 684, + 179, + 759, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/584d47e0e45e91815ee42c6f9b5991aa3e4908462f72b80c26f98653174ccaf3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 179, + 844, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4905f95b6f1e79482f208088e00690b63521c3393c87359f17f69a4d238f1e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 846, + 179, + 921, + 237 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vation with realistic interaction synthesis. Despite the decrease, the identity preservation score of $55.74\\%$ remains competitive for personalized face generation.", + "bbox": [ + 511, + 362, + 903, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, PersonaHOI bridges the gap between personalized face generation and interaction realism. Its training and tuning-free design enables seamless integration into different SD-based personalized face diffusion models, providing scalability and adaptability across diverse architectures and application scenarios.", + "bbox": [ + 511, + 409, + 905, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Personalized Multi-Subject HOI Generation", + "text_level": 1, + "bbox": [ + 511, + 508, + 888, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generating images with multiple subjects engaged in human-object interactions (HOI) presents significant challenges in preserving distinct identities and accurately depicting interactions. As shown in Figure 6, our method enhances FastComposer [35] using its pre-trained SD v1.5 model, maintaining identity fidelity while generating coherent interactions. For instance, in the challenging \"cycling together\" (second line) scenario, PersonaHOI successfully preserves the unique identities of both subjects and generates two bicycles alongside a natural background. In contrast, FastComposer [35] produces two adjacent faces with minimal interaction details. These results highlight the potential of PersonaHOI in handling complex multi-subject HOI scenarios, ensuring both identity preservation and interaction realism.", + "bbox": [ + 511, + 532, + 906, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. More Results", + "text_level": 1, + "bbox": [ + 511, + 768, + 653, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "General Personalized Face Generation Performance. We evaluate PersonaHOI on the task of general personalized face generation [21, 35] across four categories of text prompts: Accessory, Style, Action, and Context. As shown in Table 2, the scores across all categories for the PersonaHOI-enhanced method are consistently comparable to or slightly better than the baseline FastComposer [35].", + "bbox": [ + 511, + 794, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23781", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b046d87726dcdf498f5e0c912e1a3b88dc4672bcf75ea083efaa5ee130d3f4ee.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccessory (%)Style (%)Action (%)Context (%)Mean (%)
StableDiffusion v1.5 [25]NA / 26.70NA / 27.21NA / 23.66NA / 25.86NA / 25.86
FastComposer [35]54.65 / 24.2241.13 / 24.0155.35 / 21.3052.70 / 22.3150.95 / 22.96
+ Ours56.43 / 24.2546.07 / 23.9755.09 / 22.2153.77 / 22.5952.84 / 23.26
", + "bbox": [ + 133, + 88, + 866, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These results demonstrate that while our method is specifically designed for personalized HOI scenarios, it effectively preserves the generative capabilities of the base models and does not degrade their performance on general text prompts, showcasing its adaptability and robustness.", + "bbox": [ + 88, + 244, + 480, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Individual Component. We conducted an ablation study on FastComposer [35] to evaluate the contributions of the Cross-Attention Constraint (CAC), Latent Merge (LM), and Residual Merge (RM) in PersonaHOI. As shown in Table ??, the full model, integrating all three components, achieves the best performance across all metrics, underscoring the complementary roles of our proposed modules. Removing LM or RM results in significant declines in both Identity Preservation and Interaction Alignment, highlighting the critical role of these merging strategies. The absence of CAC has a relatively smaller effect on Interaction Alignment, but it substantially impacts Identity Preservation. This is because without CAC, faces generated by PFD fail to align with the spatial layout provided by SD, disrupting the coherence between PFD and SD branches and leading to diminished alignment and identity fidelity in facial regions. In summary, the ablation results confirm the distinct and critical functionalities of each module. CAC ensures precise spatial alignment of faces between PFD and SD, while LM and RM are indispensable for consistently merging non-facial HOI-relevant features from SD with facial identity details from PFD.", + "bbox": [ + 91, + 324, + 483, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on Additional Baselines. To further validate the efficacy of our approach, we conduct additional evaluations using InstantID [34] and PuLID [7]. Results in Table 3 and Figure 7 indicate PersonaHOI can benefit both PFD methods, notably improving HOI alignment and image-text consistency. InstantID shares design similarities with IP-Adapter, prioritizing portrait generation, which improves identity preservation but compromises interaction fidelity. The fusion mechanism of PersonaHOI significantly enhances HOI alignment and overall image-text consistency. PuLID, on the other hand, balances personalization against text-guided image synthesis capability during training. However, it can struggle to correctly place the object in some cases (second row in Figure 7). By incorporating PersonaHOI, we achieve consistent improvements across identity preservation, image-prompt consistency, and", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "interaction alignment, demonstrating the versatility and robustness of our method.", + "bbox": [ + 511, + 244, + 903, + 275 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/deb446a2edefb06b1e589b58dab64f082fbd0d14ecfb5f717ec1b40ccede0dfc.jpg", + "table_caption": [ + "Table 2. Comparison of Our Method with FastComposer [35] on General Personalized Face Generation. We compare across four categories of text prompts including Accessory, Style, Action, and Context, following [21, 35]. Results are formatted as \"Identity Preservation (\\%) / Prompt Consistency (\\%)\" and we bold the higher number for each pair of comparison." + ], + "table_footnote": [], + "table_body": "
MethodIdentity (%)Prompt (%)HOI (%)
InstantID76.2122.9345.73
+ Ours69.93 (-9.28)23.80 (+0.87)61.82 (+16.09)
PuLID55.9324.5072.10
+ Ours62.56 (+6.63)25.21 (+0.71)72.18 (+0.08)
", + "bbox": [ + 519, + 284, + 898, + 366 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c063ad6db310890186d59dafd9df446fdc172068c3fa15666d7cde1d58204723.jpg", + "table_caption": [ + "Table 3. Comparison of our method with InstantID and PuLID on HOI-Specific Personalized Face Generation." + ], + "table_footnote": [], + "table_body": "
Ref. ImageInstantID+OursRef. ImagePuLD+Ours
Caption: A man surfboarding with a surfboardCaption: A woman catching a frisbee
Caption: A woman riding a horseCaption: A woman snowboarding with a snowboard
", + "bbox": [ + 521, + 422, + 901, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7. Visual comparison of InstantID [34] and PuLID [7] with our method.", + "bbox": [ + 511, + 556, + 905, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 603, + 638, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a training&tuning-free framework that combines the strengths of Personalized Face Diffusion (PFD) models and StableDiffusion (SD) models to generate personalized images featuring realistic and complex human-object interactions (HOI). By introducing Cross-Attention Constraint, Latent Merge, and Residual Merge, our approach achieves seamless integration of identity-specific facial features from PFD with contextual interaction details from SD, ensuring both spatial alignment and contextual coherence. The flexibility of our framework allows adaptation to various PFD and SD architectures, demonstrating its robustness in handling diverse interaction scenarios. Without requiring additional fine-tuning or task-specific datasets, our method significantly improves identity retention and interaction realism compared to state-of-the-art approaches. This highlights the potential of leveraging pre-trained models to enable high-quality, personalized content generation across diverse applications.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23782", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 3", + "[2] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovise: Tuning-free image customization with text-to-image diffusion models, 2023. 3", + "[3] Zhuowei Chen, Shancheng Fang, Wei Liu, Qian He, Mengqi Huang, Yongdong Zhang, and Zhendong Mao. Dreamidentity: Improved editability for efficient face-identity preserved image generation. arXiv preprint arXiv:2307.00300, 2023. 3", + "[4] Dylan Campbell Frederic Z. Zhang and Stephen Gould. Spatially Conditioned Graphs for Detecting Human-Object Interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13319–13327, 2021. 3", + "[5] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel CohenOr. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 3", + "[6] Chen Gao, Si Liu, Defa Zhu, Quan Liu, Jie Cao, Haoqian He, Ran He, and Shuicheng Yan. InteractGAN: Learning to Generate Human-Object Interaction. In Association for Computing Machinery, 2020. 3", + "[7] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. PuLID: Pure and Lightning ID Customization via Contrastive Alignment. In Advances in Neural Information Processing Systems, 2024. 3, 8", + "[8] Saurabh Gupta and Jitendra Malik. Visual Semantic Role Labeling. arXiv preprint arXiv:1505.04474, 2015. 5", + "[9] Junjie He, Yifeng Geng, and Liefeng Bo. UniPortrait: A Unified Framework for Identity-Preserving Single- and Multi-Human Image Personalization. arXiv preprint arXiv:2408.05939, 2024. 3", + "[10] Jiu Tian Hoe, Xudong Jiang, Chee Seng Chan, Yap-Peng Tan, and Weipeng Hu. InteractDiffusion: Interaction Control in Text-to-Image Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[11] Siteng Huang, Biao Gong, Yutong Feng, Xi Chen, Yuqian Fu, Yu Liu, and Donglin Wang. Learning disentangled identifiers for action-customized text-to-image generation. arXiv preprint arXiv:2311.15841, 2023. 3", + "[12] Junha Hyung, Jaeyo Shin, and Jaegul Choo. Magicapture: High-resolution multi-concept portrait customization. arXiv preprint arXiv:2309.06895, 2023. 3", + "[13] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing. arXiv preprint arXiv:2312.11392, 2023. 5", + "[14] Jian-Yu Jiang-Lin, Kang-Yang Huang, Ling Lo, Yi-Ning Huang, Terence Lin, Jhih-Ciang Wu, Hong-Han Shuai, and" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Wen-Huang Cheng. ReCorD: Reasoning and Correcting Diffusion for HOI Generation. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2024. 3", + "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[16] Xiaoming Li, Xinyu Hou, and Chen Change Loy. When stylegan meets stable diffusion: a w+ adapter for personalized image generation. arXiv preprint arXiv:2311.17461, 2023. 3", + "[17] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-Set Grounded Text-to-Image Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[18] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 6, 7", + "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep Learning Face Attributes in the Wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 3", + "[20] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 3", + "[21] Xu Peng, Junwei Zhu, Boyuan Jiang, Ying Tai, Donghao Luo, Jiangning Zhang, Wei Lin, Taisong Jin, Chengjie Wang, and Rongrong Ji. PortraitBooth: A Versatile Portrait Model for Fast Identity-preserved Personalization. arXiv preprint arXiv:2312.06354, 2023. 2, 3, 4, 5, 6, 7, 8", + "[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 6", + "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 6", + "[24] Iasonas Kokkinos Riza Alp Güler, Natalia Neverova. Dense-Pose: Dense Human Pose Estimation In The Wild. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 7", + "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8", + "[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23783", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3", + "[27] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 3", + "[28] Florian Schroff, Dmitry Kalenichenko, and James Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6", + "[29] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2, 3", + "[30] Kaede Shiohara and Toshihiko Yamasaki. Face2diffusion for fast and editable face personalization. arXiv preprint arXiv:2403.05094, 2024. 3", + "[31] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. FreeU: Free Lunch in Diffusion U-Net. In CVPR, 2024. 5", + "[32] Yu-Chuan Su, Kelvin CK Chan, Yandong Li, Yang Zhao, Han Zhang, Boqing Gong, Huisheng Wang, and Xuhui Jia. Identity encoder for personalized diffusion. arXiv preprint arXiv:2304.07429, 2023. 3", + "[33] Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. In SIGGRAPH Asia 2023 Conference Papers, pages 1-10, 2023. 3", + "[34] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. InstantID: Zero-shot Identity-Preserving Generation in Seconds. arXiv preprint arXiv:2401.07519, 2024. 3, 8", + "[35] Guangxuan Xiao, Tianwei Yin, William T Freeman, Fredo Durand, and Song Han. FastComposer: Tuning-Free Multi-Subject Image Generation with Localized Attention. In International Journal of Computer Vision, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[36] Yuxuan Yan, Chi Zhang, Rui Wang, Yichao Zhou, Gege Zhang, Pei Cheng, Gang Yu, and Bin Fu. Facestudio: Put your face everywhere in seconds. arXiv preprint arXiv:2312.02663, 2023. 3", + "[37] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IpAdapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 6, 7", + "[38] Hangjie Yuan, Jianwen Jiang, Samuel Albanie, Tao Feng, Ziyuan Huang, Dong Ni, and Mingqian Tang. RLIP: Relational Language-Image Pre-training for Human-Object Interaction Detection. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3", + "[39] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Samuel Albanie, Yining Pan, Tao Feng, Jianwen Jiang, Dong Ni, Yingya Zhang, and Deli Zhao. RLIPv2: Fast Scaling of Relational Language-Image Pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023." + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Frederic Z. Zhang, Dylan Campbell, and Stephen Gould. Efficient Two-Stage Detection of Human-Object Interactions with a Novel Unary-Pairwise Transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20104-20112, 2022. 3, 6", + "[41] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2016. 6", + "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In IEEE International Conference on Computer Vision (ICCV), 2023. 3", + "[43] Xulu Zhang, Xiao-Yong Wei, Wengyu Zhang, Jinlin Wu, Zhaoxiang Zhang, Zhen Lei, and Qing Li. A Survey on Personalized Content Synthesis with Diffusion Models. arXiv preprint arXiv:2405.05538, 2024. https://arxiv.org/abs/2405.05538.3", + "[44] Yuxin Zhang, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. MotionCrafter: One-Shot Motion Customization of Diffusion Models. arXiv preprint arXiv:2312.05288, 2023. 3" + ], + "bbox": [ + 516, + 90, + 905, + 415 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23784", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/00e10d7f-1169-4801-84c3-c1efaaab378c_model.json b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/00e10d7f-1169-4801-84c3-c1efaaab378c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a95be6c51237876091bfc8068083a42d101b19c9 --- /dev/null +++ b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/00e10d7f-1169-4801-84c3-c1efaaab378c_model.json @@ -0,0 +1,2640 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.208, + 0.131, + 0.791, + 0.177 + ], + "angle": 0, + "content": "PersonaHOI: Effortlessly Improving Face Personalization in Human-Object Interaction Generation" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.205, + 0.26, + 0.222 + ], + "angle": 0, + "content": "Xinting Hu*" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.226, + 0.266, + 0.24 + ], + "angle": 0, + "content": "xhu@mpi-inf.mpg.de" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.205, + 0.445, + 0.222 + ], + "angle": 0, + "content": "Haoran Wang*" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.226, + 0.472, + 0.24 + ], + "angle": 0, + "content": "hawang@mpi-inf.mpg.de" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.205, + 0.65, + 0.221 + ], + "angle": 0, + "content": "Jan Eric Lenssen" + }, + { + "type": "text", + "bbox": [ + 0.494, + 0.226, + 0.698, + 0.24 + ], + "angle": 0, + "content": "jlenssen@mpi-inf.mpg.de" + }, + { + "type": "text", + "bbox": [ + 0.747, + 0.205, + 0.858, + 0.221 + ], + "angle": 0, + "content": "Bernt Schiele" + }, + { + "type": "text", + "bbox": [ + 0.719, + 0.226, + 0.915, + 0.24 + ], + "angle": 0, + "content": "schiele@mpi-inf.mpg.de" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.247, + 0.816, + 0.265 + ], + "angle": 0, + "content": "Max Planck Institute for Informatics, Saarland Informatics Campus, Germany" + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.285, + 0.404, + 0.302 + ], + "angle": 0, + "content": "Context + HOI" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.303, + 0.456, + 0.315 + ], + "angle": 0, + "content": "A man riding a horse in front of a medieval castle" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.317, + 0.256, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.316, + 0.379, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.39, + 0.365, + 0.399 + ], + "angle": 0, + "content": "FastComposer" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.316, + 0.486, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.39, + 0.453, + 0.398 + ], + "angle": 0, + "content": "+Ours" + }, + { + "type": "title", + "bbox": [ + 0.612, + 0.282, + 0.746, + 0.304 + ], + "angle": 0, + "content": "Style + HOI" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.304, + 0.841, + 0.316 + ], + "angle": 0, + "content": "A woman holding a cat in the style of Vincent Van Gogh" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.317, + 0.608, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.318, + 0.724, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.391, + 0.711, + 0.4 + ], + "angle": 0, + "content": "FastComposer" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.317, + 0.831, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.391, + 0.8, + 0.4 + ], + "angle": 0, + "content": "+Ours" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.417, + 0.231, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.468, + 0.248, + 0.512 + ], + "angle": 0, + "content": "Various man-Object action (HOI) descriptions" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.454, + 0.27, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.414, + 0.348, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.414, + 0.416, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.415, + 0.485, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.464, + 0.348, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.464, + 0.416, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.464, + 0.485, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.527, + 0.418, + 0.703, + 0.446 + ], + "angle": 0, + "content": "Personalized" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.446, + 0.687, + 0.465 + ], + "angle": 0, + "content": "Face with" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.469, + 0.686, + 0.488 + ], + "angle": 0, + "content": "Human-Object" + }, + { + "type": "text", + "bbox": [ + 0.561, + 0.488, + 0.693, + 0.505 + ], + "angle": 0, + "content": "Interaction" + }, + { + "type": "text", + "bbox": [ + 0.739, + 0.428, + 0.812, + 0.444 + ], + "angle": 0, + "content": "Training" + }, + { + "type": "text", + "bbox": [ + 0.765, + 0.446, + 0.784, + 0.461 + ], + "angle": 0, + "content": "&" + }, + { + "type": "text", + "bbox": [ + 0.744, + 0.464, + 0.806, + 0.481 + ], + "angle": 0, + "content": "Tuning" + }, + { + "type": "text", + "bbox": [ + 0.751, + 0.482, + 0.799, + 0.498 + ], + "angle": 0, + "content": "Free!" + }, + { + "type": "title", + "bbox": [ + 0.239, + 0.529, + 0.413, + 0.549 + ], + "angle": 0, + "content": "Accessory + HOI" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.55, + 0.495, + 0.56 + ], + "angle": 0, + "content": "A woman wearing a Santa hat and taking selfie with phone" + }, + { + "type": "image", + "bbox": [ + 0.159, + 0.563, + 0.257, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.563, + 0.388, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.637, + 0.378, + 0.646 + ], + "angle": 0, + "content": "FastComposer" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.563, + 0.495, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.433, + 0.638, + 0.465, + 0.646 + ], + "angle": 0, + "content": "+Ours" + }, + { + "type": "title", + "bbox": [ + 0.597, + 0.529, + 0.774, + 0.544 + ], + "angle": 0, + "content": "Mult-Person + HOI" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.545, + 0.801, + 0.557 + ], + "angle": 0, + "content": "A man and a man are playing guitar in a bar" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.558, + 0.587, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.558, + 0.721, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.637, + 0.706, + 0.646 + ], + "angle": 0, + "content": "FastComposer" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.558, + 0.835, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.637, + 0.798, + 0.646 + ], + "angle": 0, + "content": "+Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.662, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Figure 1. Examples of Personalized Face with Human-Object Interaction (HOI) Generation. We present PersonaHOI, a training- and tuning-free framework built on existing diffusion models. Using a single reference image and diverse HOI prompts, PersonaHOI generates identity-consistent human-object interactions compared to FastComposer [35]. PersonaHOI can further seamlessly integrate varied contexts, styles, accessories, and multi-person scenarios, ensuring scalability and practicality for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.729, + 0.326, + 0.745 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.763, + 0.486, + 0.869 + ], + "angle": 0, + "content": "We introduce PersonaHOI, a training- and tuning-free framework that fuses a general StableDiffusion model with a personalized face diffusion (PFD) model to generate identity-consistent human-object interaction (HOI) images. While existing PFD models have advanced significantly, they often overemphasize facial features at the expense of full-body coherence, PersonaHOI introduces an additional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.909, + 0.882 + ], + "angle": 0, + "content": "StableDiffusion (SD) branch guided by HOI-oriented text inputs. By incorporating cross-attention constraints in the PFD branch and spatial merging at both latent and residual levels, PersonaHOI preserves personalized facial details while ensuring interactive non-facial regions. Experiments, validated by a novel interaction alignment metric, demonstrate the superior realism and scalability of PersonaHOI, establishing a new standard for practical personalized face with HOI generation. Code is available at https://github.com/JoyHuYY1412/PersonaHOI." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.223, + 0.9 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23775" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.093, + 0.898, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.297, + 0.908, + 0.355 + ], + "angle": 0, + "content": "Figure 2. (a) The spatial layout of StableDiffusion guides PersonaHOI to generate personalized content with coherent human-object interactions (HOI). (b) Analysis of identity injection timing in PFD models. We use FastComposer [35] for diffusion model generation. Injecting face representation at the start of image generation preserves facial details but lacks coherent HOI, while delayed injection continuously deviates from the original identity, resulting in random human features and meaningless human-object interactions." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.378, + 0.224, + 0.394 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.528 + ], + "angle": 0, + "content": "Personalized face generation has seen increasing public interest and demand in user-specified digital content creation. Current learning-based personalized face diffusion (PFD) models [18, 21, 29, 35], trained on large-scale face-centric datasets, can incorporate a single user-provided reference facial image into a text-to-image generation process, enabling the rapid creation of images that depict specific persons in diverse scenes, outfits, and styles within seconds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.671 + ], + "angle": 0, + "content": "Although these methods perform well in simpler scenarios, they often struggle with generating full-body depictions involving detailed human-object interactions (HOI). As illustrated in Figure 1, the generated images of FastComposer [35] show missing objects or body parts, resulting in portrait-focused outputs without HOI information. This limitation compromises the overall realism and practical utility of the generated content, especially in immersive or interactive applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To diagnose the limitations of PFD models, we examine the results of PFD models by varying the timing of identity injection during the generation process. As seen in the first and last columns of Figure 2(b), early injection effectively preserves facial details from the reference image, but relying solely on text input (last column) results in incoherent HOI outputs. This analysis highlights that the core challenge lies not in maintaining identity fidelity but in generating coherent body movements and interactions driven by text prompts. Notably, pre-trained models used for PFD, such as StableDiffusion [25], can generate satisfactory HOI images due to their extensive and diverse training data. This observation implies that the fine-tuning process on face-centric datasets of PFD models diminishes their ability to follow complex HOI text prompts. Therefore, to advance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.424 + ], + "angle": 0, + "content": "current PFD methods in HOI tasks, reintroducing the capability to leverage text prompts for natural full-body interactions is essential." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.909, + 0.548 + ], + "angle": 0, + "content": "Motivated by this need, we propose a straightforward yet powerful approach: leveraging the generative capabilities of pre-trained diffusion models, such as StableDiffusion (SD), to augment existing PFD frameworks. Our method, PersonaHOI, integrates the strengths of SD without requiring additional training or fine-tuning, thereby restoring the ability to generate realistic, text-driven full-body HOI while preserving identity-specific details from the reference image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Central to this integration, we incorporate an additional SD branch that aligns identity-specific details from PFD models with the HOI layouts generated by SD. A head segmentation mask derived from the SD output guides the merge of non-facial components from the SD branch with the facial details from the PFD branch. Specifically, in the PFD model branch, we introduce a Cross-Attention Constraint to prevent the overemphasis of identity features across the entire image, ensuring these details are confined to the facial region. To integrate HOI representations from SD, we implement Latent Merge and Residual Merge, merging identity and contextual features at both the latent representation level and via skip connections within the U-Net architecture. This multi-level merging strategy ensures that the generated images maintain realistic human-object interactions while preserving personalized facial details, resulting in cohesive, interaction-rich outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.9 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our method in HOI scenarios, we propose a novel \"interaction alignment\" metric. These metrics leverage HOI detectors to measure alignment between generated images and text prompts, objectively assessing interaction realism. Our contributions can be summarized as follows:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23776" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.167 + ], + "angle": 0, + "content": "- We present PersonaHOI, a dual-path architecture that integrates general and personalized diffusion models. Our framework generates realistic human-object interactions with specific identities, without requiring additional training or test-time tuning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.168, + 0.482, + 0.243 + ], + "angle": 0, + "content": "- Our method introduces key components, including crossattention constraints and a combination of latent and residual merge strategies. These innovations enable effective feature integration, ensuring that generated images maintain both interaction coherence and identity fidelity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.243, + 0.482, + 0.318 + ], + "angle": 0, + "content": "- We design a novel evaluation metric for human-object interaction. Our experiments show significant improvements over state-of-the-art face personalization techniques in terms of interaction realism across models, emphasizing the scalability and robustness of our approach." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.332, + 0.241, + 0.348 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.358, + 0.485, + 0.734 + ], + "angle": 0, + "content": "Personalized Face Generation Diffusion Models. Personalized face generation aims to create identity-consistent images of individuals from limited reference data [32, 36, 43]. Given pretrained general image generation models [25], traditional optimization-based approaches [5, 26, 27] refine model parameters during inference for each individual but are computationally expensive, making them unsuitable for real-time applications. In contrast, learning-based approaches leverage human-centric datasets [1, 15, 19, 20] to train robust models capable of varying identity inputs without requiring test-time fine-tuning. Facial representations, extracted by models like CLIP [23] or other face recognition models [3, 16, 33], are often fused with [18, 21, 35] or replace [30] textual embeddings of specific words, such as \"person\", to provide identity conditions for diffusion models and enhance contextual relevance [3, 29]. In IP-Adapter [37] and other works [7, 9, 34], facial features are processed separately through decoupled cross-attention modules. Incorporating human masks as prior, methods can clean the background during data processing; alternatively, masks can be leveraged to construct [2, 18, 21, 35] and adjust [12] loss functions during training. In our work, we extend face personalization to complex human-object interactions (HOI), achieving a higher level of personalization and adaptability for diverse generative tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Human-Object Interaction in Diffusion Models. HOI detection aims to locate human-object pairs and categorize their interactions as triplets, such as (person, playing, football) [4, 38-40], while HOI image synthesis remains relatively under-explored. InteractGAN [6] generates HOI images using human pose templates and reference images, but its reliance on pose-template pools and object references limits flexibility. Recent diffusion-based approaches tackle HOI generation by incorporating layout controls such as bounding boxes [10, 14, 17] or human poses [17, 42], or by leveraging in-context samples with similar interac" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "tions [11, 44]. Additionally, ReCorD [14] integrates Latent Diffusion Models with Visual Language Models, introducing modules for reasoning and correcting interactions to enhance cross-modal alignment. In this work, we rely on the standard text-to-image diffusion model (i.e., StableDiffusion [25]) for HOI generation, and fuse the generation process with personalized face diffusion models to generation identity-preserving HOI images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.226, + 0.64, + 0.243 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.905, + 0.372 + ], + "angle": 0, + "content": "Understanding the fundamentals of StableDiffusion (SD) is essential to leverage its generative capabilities within our PFG framework. SD is a latent diffusion model designed to generate high-quality images based on text or image prompts. It operates in a latent space, where an input image is encoded in a latent representation \\(\\mathbf{z}_0\\) using a pretrained variational autoencoder (VAE) encoder, reducing the dimensionality of the image for efficient processing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.905, + 0.493 + ], + "angle": 0, + "content": "During the generation phase, the SD model begins with an initial noisy latent representation \\(\\mathbf{z}_T\\) and progressively denoises it over \\(T\\) timesteps to reconstruct \\(\\mathbf{z}_0\\). This denoised latent is then decoded by the VAE decoder to produce the final image. At each timestep \\(t\\), the U-Net architecture \\(\\theta\\) predicts noise \\(\\epsilon (\\mathbf{z}_t,t,\\mathbf{C})\\), where \\(\\mathbf{C}\\) represents the conditioning inputs (text or image embeddings). The iterative denoising process follows:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.505, + 0.905, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t - 1} = \\operatorname {D e n o i s e} \\left(\\mathbf {z} _ {t}, \\epsilon \\left(\\mathbf {z} _ {t}, t, \\mathbf {C}\\right); \\theta\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.905, + 0.638 + ], + "angle": 0, + "content": "The U-Net architecture comprises an encoder-decoder structure with skip connections that transfer features between the encoder and the decoder. Within this architecture, SD employs a cross-attention mechanism to incorporate conditioning inputs effectively during denoising. This mechanism projects latent representations and conditioning inputs into query, key, and value spaces:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.647, + 0.905, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathbf {Q} = \\mathbf {W} _ {q} \\mathbf {z}, \\mathbf {K} = \\mathbf {W} _ {k} \\mathbf {C}, \\mathbf {V} = \\mathbf {W} _ {v} \\mathbf {C}; \\\\ \\mathbf {A} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right), \\mathbf {z} _ {\\text {a t t n}} = \\mathbf {A V}. \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.781 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{A}\\) is the attention map that ensures that relevant conditioning tokens guide denoising at each generation step. This cross-attention mechanism is critical for aligning the generated image with the input prompts, enabling consistency and fidelity in the final output." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.794, + 0.605, + 0.809 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.71, + 0.833 + ], + "angle": 0, + "content": "4.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Figure 3, PersonaHOI augments the PFD model by integrating an additional StableDiffusion (SD) branch, enhancing its capacity to generate personalized images that capture complex human-object interactions (HOI)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23777" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.087, + 0.849, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.396, + 0.908, + 0.48 + ], + "angle": 0, + "content": "Figure 3. Overview of Our Proposed Framework, PersonaHOI. The architecture integrates a personalized face diffusion (PFD) model with an additional StableDiffusion (SD) branch. First, SD generates an image \\((I_{SD})\\) from a text prompt and noisy latent representation \\((z_{T})\\), which is decoded and segmented to produce a head mask. Next, SD and PFD run in parallel from the same \\(z_{T}\\). At every timestep \\(t\\), the head mask guides the Cross-Attention Constraint in PFD and merging modules (Latent Merge and Residual Merge) to merge interaction-relevant features from SD with identity-specific details from PFD. Iteratively, this process introduces HOI context to personalized face generation in a training&tuning-free manner." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.499, + 0.482, + 0.575 + ], + "angle": 0, + "content": "in a training-free manner. Given an input comprising a user-provided reference image and a text prompt specifying the interaction (e.g., \"a person kicking football\"), our framework produces a cohesive output image that retains the subject's identity while accurately depicting the specified HOI." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.576, + 0.483, + 0.773 + ], + "angle": 0, + "content": "To effectively merge identity from PFD and interaction features from SD, we introduce three core strategies: Cross-Attention Constraint, which regulates attention to reference image features within the PFD model (detailed in Sec. 4.2); Latent Merge, which combines features in the latent space at each generation timestep (Sec. 4.3); and Residual Merge, which integrates identity details through skip connections in the U-Net architecture (Sec. 4.4). Together, these strategies enable PersonaHOI to align generated content with interaction-driven text prompts while maintaining identity fidelity. Furthermore, this design allows PersonaHOI adaptable to various PFD models without the need for model retraining or test-time tuning." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.787, + 0.392, + 0.803 + ], + "angle": 0, + "content": "4.2. Cross-Attention Constraint (CAC)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In the cross-attention layers of the PFD model, conditioning inputs \\(\\mathbf{C} \\in \\mathbb{R}^{N \\times D}\\) include both text and image embeddings to guide the generation process, where \\(N\\) is the number of tokens, and \\(D\\) is the embedding dimension. Following established PFD model approaches [18, 21, 35], the \\(n_{img}\\)-th token in \\(\\mathbf{C}\\) incorporates the image embedding, encapsulat" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.905, + 0.589 + ], + "angle": 0, + "content": "ing identity-specific details essential for facial preservation, while the remaining \\(N - 1\\) tokens correspond to text embeddings. In standard PFD models, attention to the \\(n_{img}\\)-th token often spreads across the image, overemphasizing facial features thus decreasing spatial capacity for representing body movements and object interactions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.816 + ], + "angle": 0, + "content": "To address this, we introduce Cross-Attention Constraint (CAC) within the cross-attention layers of the PFD model. CAC restricts identity features to specific facial regions, ensuring sufficient spatial capacity for body and object interactions. Different from the attention localization loss used in training [21, 35], our CAC utilizes fine-grained head mask derived from the SD-generated image \\( I_{SD} \\) in image generation process, ensuring the attention constraint in PFD is consistent with the HOI layout generated by SD. The head mask is segmented by an off-the-shelf head segmentor [24], then resized to match the spatial dimensions \\( (H \\times W) \\) of the latent representation \\( \\mathbf{z} \\), denoted as \\( \\mathbf{M}^{\\mathrm{head}} \\in [0,1]^{H \\times W} \\) (1 for the head region and 0 for other regions). We apply \\( \\mathbf{M}^{\\mathrm{head}} \\) to the attention map \\( \\mathbf{A} \\in \\mathbb{R}^{(H \\times W) \\times N} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.818, + 0.905, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i \\in [ 0, 1, \\dots , N - 1 ]} ^ {C A C} = \\left\\{ \\begin{array}{l l} \\text {f l a t t e n} (\\mathbf {M} ^ {\\text {h e a d}}), & \\text {i f} i = i m g, \\\\ 1, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.861, + 0.771, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} ^ {C A C} = \\mathbf {A} \\odot (\\mathbf {M} ^ {C A C}) ^ {\\intercal}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.885, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\(\\mathbf{M}^{CAC}\\) modifies the attention weights for the \\(n_{img}\\) token" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23778" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.092, + 0.473, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.251, + 0.485, + 0.336 + ], + "angle": 0, + "content": "Figure 4. Illustration of Residual Merge. In each residual layer, Residual Merge operates within the U-Net skip connections, utilizing a head mask to guide the integration of high-frequency identity details from PFD residuals and low-frequency interaction layouts from SD residuals. The merged residuals are then concatenated to the corresponding bottleneck features from PFD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.469 + ], + "angle": 0, + "content": "by setting the weights to 0 in regions outside the head, preventing the influence of human identity on non-facial areas. Since the head mask is derived from SD-generated outputs, CAC maintains facial layout coherence with SD. When further combined with Latent and Residual Merge, CAC helps blend facial features from PFD and non-facial interaction features from SD without mutual interference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.481, + 0.279, + 0.497 + ], + "angle": 0, + "content": "4.3. Latent Merge (LM)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.639 + ], + "angle": 0, + "content": "In this section, we introduce Latent Merge, which directly merges the latent representation within U-Net from PFD and SD models. After applying Cross-Attention Constraint (CAC) in Sec. 4.2, the facial region generated by the PFD model becomes spatially aligned with the corresponding region in the broader HOI layout produced by the SD model. This alignment facilitates spatial merge, blending facial identity details from PFD with interaction contexts from SD in latent space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.672 + ], + "angle": 0, + "content": "At each diffusion timestep \\(t\\) of PFD and SD, we implement the latent space merging with \\(\\mathbf{M}^{head}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.683, + 0.483, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t} = \\mathbf {M} ^ {h e a d} \\odot \\mathbf {z} _ {t} ^ {\\mathrm {P F D}} + \\left(1 - \\mathbf {M} ^ {h e a d}\\right) \\odot \\mathbf {z} _ {t} ^ {\\mathrm {S D}}. \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.484, + 0.807 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{z}_t^{\\mathrm{PFD}}\\) and \\(\\mathbf{z}_t^{\\mathrm{SD}}\\) represents the latent features from PFD and SD at timestep \\(t\\), respectively. The merged latent \\(\\mathbf{z}_t\\) then serves as input for both SD and PFD models in the next denoising timestep \\(t - 1\\). This merging strategy allows identity details retained within facial regions, while non-facial regions maintain the HOI context generated by SD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.295, + 0.849 + ], + "angle": 0, + "content": "4.4. Residual Merge (RM)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "While Latent Merge enables interactions between the SD and PFD at the latent level, it lacks integration at the intermediate feature stage within the U-Net. To address this," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "we introduce Residual Merge, which enhances feature integration within the skip connections. In the U-Net, skip connections transmit residual features containing rich detail from the encoder to the decoder, significantly impacting the content and quality of the generated images [13, 31]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Our Residual Merge is applied to each skip connection layer \\(l\\) (\\(l = 1, \\dots, L\\)), as shown in Figure 4. First, residual features are extracted from both the PFD and SD U-Net backbones. A low-pass filter and high-pass filter is applied to the SD and PFD residuals, respectively. Low-pass and high-pass filters are applied to these residuals to balance global coherence and local precision: low-pass filtering SD residuals suppresses high-frequency noise to retain the broader scene layout, while high-pass filtering PFD residuals emphasizes fine-grained identity-specific details. To achieve precise integration between facial and non-facial regions, we apply the head segmentation mask \\(\\mathbf{M}^{\\mathrm{head}}\\), resized to match the residual resolution at layer \\(l\\), denoted as \\(\\mathbf{M}_{\\mathrm{R}}^{l}\\). The fusion of residual features from both paths is then implemented for each layer \\(l\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.406, + 0.907, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {\\text {m e r g e d}} ^ {l} = \\mathbf {M} _ {R} ^ {l} \\odot \\mathrm {H P} \\left(\\mathbf {R} _ {\\text {P F D}} ^ {l}\\right) + \\left(1 - \\mathbf {M} _ {R} ^ {l}\\right) \\odot \\mathrm {L P} \\left(\\mathbf {R} _ {\\text {S D}} ^ {l}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.436, + 0.906, + 0.514 + ], + "angle": 0, + "content": "where \\(\\mathrm{HP}(\\cdot)\\) and \\(\\mathrm{LP}(\\cdot)\\) denote the high pass and low pass filter, \\(\\mathbf{R}_{\\mathrm{PFD}}\\) and \\(\\mathbf{R}_{\\mathrm{SD}}\\) denote the residual features from the PFD and SD paths, respectively. The merged residual \\(\\mathbf{R}_{\\mathrm{merged}}\\) is then concatenated with the bottom-up features in the PFD U-Net for further processing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.906, + 0.589 + ], + "angle": 0, + "content": "The Residual Merge strategy enables fine-grained feature integration within the U-Net. Through this, facial regions retain high-frequency identity details from PFD while non-facial regions incorporate interaction layouts from SD, ensuring cohesive and contextually aligned outputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.646, + 0.621 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.63, + 0.596, + 0.645 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.655, + 0.871, + 0.668 + ], + "angle": 0, + "content": "Evaluation Data. We use two test sets for evaluation:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.67, + 0.906, + 0.776 + ], + "angle": 0, + "content": "1) HOI-Focused Personalized Face Generation: To test complex human-object interactions (HOI), we use interaction labels from the widely-used V-COCO [8] dataset. V-COCO provides diverse HOI scenarios for robust evaluation of interaction fidelity. We construct 30 prompts in the \"subject + interaction + object\" format (e.g., \"man kicking a ball\") with the same 15 reference subjects as in 1)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.777, + 0.906, + 0.868 + ], + "angle": 0, + "content": "2) General Personalized Face Generation: Following prior studies [21, 35], we compile a benchmark with 15 reference subjects, each paired with 40 prompts spanning diverse scenarios like context, stylization, accessory, and actions. These prompts aim to assess both identity retention and the model's ability to adapt to various text-guided settings." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.67, + 0.906, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.869, + 0.861, + 0.883 + ], + "angle": 0, + "content": "Full prompt details are provided in the Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.874, + 0.901 + ], + "angle": 0, + "content": "Metrics. We employ the following evaluation metrics:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23779" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.245 + ], + "angle": 0, + "content": "
MethodArchitectureIdentity Preservation (%)Prompt Consistency (%)HOI Alignment (%)
StableDiffusion v1.5 [25]SD1.5NA24.9660.23
StableDiffusion XL [22]SDXLNA25.3270.78
FastComposer [35]SD1.553.5721.3035.96
+ OursSD1.555.28(+1.71)23.16(+1.86)56.65(+20.69)
IP-Adapter [37]SDXL62.8622.7349.83
+ OursSDXL55.74(-7.11)24.20(+1.47)68.30(+18.47)
PhotoMaker [18]SDXL50.5423.9656.77
+ OursSDXL51.34(+0.80)24.34(+0.38)76.01(+19.24)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.255, + 0.907, + 0.311 + ], + "angle": 0, + "content": "Table 1. Comparison of Our Method with Baseline Approaches on HOI-Specific Personalized Face Generation. StableDiffusion serves as the text-only baseline without subject conditioning. PersonaHOI seamlessly incorporates existing Personalized Face Diffusion models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) with their corresponding StableDiffusion architectures. We bold the higher number for each pair of comparison." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.329, + 0.892, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.725, + 0.907, + 0.755 + ], + "angle": 0, + "content": "Figure 5. Qualitative Examples of PersonaHOI and Baseline Models. Comparison of baseline models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) and their PersonaHOI-enhanced results for diverse human-object interaction prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.482, + 0.885 + ], + "angle": 0, + "content": "1) Identity Preservation: Measures how well the generated images retain the identity of the reference image. Face detection is performed using MTCNN [41], and identity similarity is calculated with FaceNet [28], following [21, 35]. 2) Prompt Consistency: Evaluates text-image alignment using CLIP-based [23] scores, where higher values indicate better prompt adherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.886, + 0.482, + 0.902 + ], + "angle": 0, + "content": "3) HOI Alignment (Ours): Assesses how well the gen" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.781, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.902 + ], + "angle": 0, + "content": "erated images depict specified human-object interactions. We adopt UPT [40], an HOI detector pre-trained on V-COCO dataset, to identify the interaction specified in text prompts for generated images with HOI alignment score \\( S_{\\mathrm{HOI}} \\). Specifically, UPT provides confidence scores for detected human and object classes (\\( S_{\\mathrm{Human}} \\), \\( S_{\\mathrm{Object}} \\)) from its object detection head, as well as interaction logits (\\( L_{\\mathrm{Inter}} \\)) from its interaction head. With these outputs, HOI align-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23780" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.294, + 0.106 + ], + "angle": 0, + "content": "ment score \\(S_{\\mathrm{HOI}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.111, + 0.483, + 0.13 + ], + "angle": 0, + "content": "\\[\nS _ {\\mathrm {H O I}} = \\left(S _ {\\text {H u m a n}} \\cdot S _ {\\text {O b j e c t}}\\right) ^ {\\lambda} \\cdot \\sigma \\left(L _ {\\text {I n t e r}}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.483, + 0.183 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the sigmoid activation function and \\(\\lambda = 2.8\\). We validate the human alignment and effectiveness for our proposed HOI Alignment metric in Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.484, + 0.368 + ], + "angle": 0, + "content": "Compared Methods and Implementation Details. We compare our method with state-of-the-art learning-based personalized face diffusion models, including FastComposer [35], IP-Adapter [37], and PhotoMaker [18]. Our PeronaHOI used their pre-trained diffusion backbones as the additional branch, i.e., StableDiffusion v1.5 for FastComposer and StableDiffusion XL for PhotoMaker, with output resolutions set to \\(512 \\times 512\\) and \\(1024 \\times 1024\\), respectively. Head segmentation for SD generated image is performed with DensePose [24] due to its robust segmentation accuracy. In our generation process, no additional training or test-time tuning is applied to any compared methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.376, + 0.434, + 0.391 + ], + "angle": 0, + "content": "5.2. Personalized Face with HOI Generation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Qualitative Comparison. Figure 5 illustrates qualitative comparisons between baseline models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) and their PersonaHOI-enhanced counterparts for diverse human-object interaction prompts. Baseline methods often fail to produce meaningful interactions, leading to missing or poorly integrated objects and unnatural human-object dynamics. For instance, in the prompt \"A man skateboarding with skateboard\" (first row of Figure 5), FastComposer [35] and PhotoMaker omit the skateboard entirely and IP-Adapter fails to generate the accurate interaction (\"holding\" the skateboard instead), resulting in inconsistent outputs with the text prompt. In contrast, PersonaHOI-enhanced models generate realistic skateboard placement and natural human-object dynamics. Moreover, PersonaHOI also ensures high-fidelity preservation of facial identity given in reference images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Quantitative Comparison. Table ?? compares the performance of PersonaHOI across various baseline models. For FastComposer [35] and PhotoMaker [18], PersonaHOI achieves consistent improvements across all metrics. Notably, Interaction Alignment improves by \\(20.69\\%\\) for FastComposer [35] and \\(19.24\\%\\) for PhotoMaker [18], demonstrating superior human-object interaction coherence. Gains in Identity Preservation and Prompt Consistency further validate PersonaHOI's ability to retain identity while enhancing text-image alignment. For IP-Adapter [37], a decrease in Identity Preservation (from \\(62.86\\%\\) to \\(55.74\\%\\)) is observed. This drop is attributed to IP-Adapter's inherent bias toward generating face-centric images, which inflates identity scores at the cost of interaction fidelity. PersonaHOI mitigates this trade-off by significantly improving Interaction Alignment \\((+18.47\\%)\\) and Prompt Consistency \\((+1.47\\%)\\), balancing identity preser" + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.095, + 0.582, + 0.106 + ], + "angle": 0, + "content": "SD v1.5" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.108, + 0.599, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.095, + 0.736, + 0.107 + ], + "angle": 0, + "content": "Reference Images" + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.108, + 0.681, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.095, + 0.76, + 0.107 + ], + "angle": 0, + "content": "ce Images" + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.108, + 0.76, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.095, + 0.848, + 0.107 + ], + "angle": 0, + "content": "FastComposer" + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.108, + 0.844, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.865, + 0.095, + 0.906, + 0.107 + ], + "angle": 0, + "content": "+Ours" + }, + { + "type": "image", + "bbox": [ + 0.848, + 0.108, + 0.923, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.167, + 0.761, + 0.179 + ], + "angle": 0, + "content": "Caption: A man and a man planting a tree" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.18, + 0.598, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.18, + 0.677, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.18, + 0.76, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.18, + 0.845, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.848, + 0.18, + 0.923, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.239, + 0.769, + 0.251 + ], + "angle": 0, + "content": "Caption: A man and a man cycling together" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.263, + 0.905, + 0.347 + ], + "angle": 0, + "content": "Figure 6. Qualitative results for multi-subject HOI generation. We compare generation outputs from SD v1.5 [25], FastComposer [35], and our PersonaHOI based on FastComposer [35] with different multi-subject interaction prompts. PersonaHOI not only preserves distinct identities but also generates coherent human-object interactions aligned with HOI layouts produced by SD v1.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.905, + 0.408 + ], + "angle": 0, + "content": "vation with realistic interaction synthesis. Despite the decrease, the identity preservation score of \\(55.74\\%\\) remains competitive for personalized face generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.5 + ], + "angle": 0, + "content": "Overall, PersonaHOI bridges the gap between personalized face generation and interaction realism. Its training and tuning-free design enables seamless integration into different SD-based personalized face diffusion models, providing scalability and adaptability across diverse architectures and application scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.51, + 0.89, + 0.527 + ], + "angle": 0, + "content": "5.3. Personalized Multi-Subject HOI Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.907, + 0.76 + ], + "angle": 0, + "content": "Generating images with multiple subjects engaged in human-object interactions (HOI) presents significant challenges in preserving distinct identities and accurately depicting interactions. As shown in Figure 6, our method enhances FastComposer [35] using its pre-trained SD v1.5 model, maintaining identity fidelity while generating coherent interactions. For instance, in the challenging \"cycling together\" (second line) scenario, PersonaHOI successfully preserves the unique identities of both subjects and generates two bicycles alongside a natural background. In contrast, FastComposer [35] produces two adjacent faces with minimal interaction details. These results highlight the potential of PersonaHOI in handling complex multi-subject HOI scenarios, ensuring both identity preservation and interaction realism." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.655, + 0.785 + ], + "angle": 0, + "content": "5.4. More Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.902 + ], + "angle": 0, + "content": "General Personalized Face Generation Performance. We evaluate PersonaHOI on the task of general personalized face generation [21, 35] across four categories of text prompts: Accessory, Style, Action, and Context. As shown in Table 2, the scores across all categories for the PersonaHOI-enhanced method are consistently comparable to or slightly better than the baseline FastComposer [35]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23781" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.867, + 0.168 + ], + "angle": 0, + "content": "
MethodAccessory (%)Style (%)Action (%)Context (%)Mean (%)
StableDiffusion v1.5 [25]NA / 26.70NA / 27.21NA / 23.66NA / 25.86NA / 25.86
FastComposer [35]54.65 / 24.2241.13 / 24.0155.35 / 21.3052.70 / 22.3150.95 / 22.96
+ Ours56.43 / 24.2546.07 / 23.9755.09 / 22.2153.77 / 22.5952.84 / 23.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.177, + 0.908, + 0.22 + ], + "angle": 0, + "content": "Table 2. Comparison of Our Method with FastComposer [35] on General Personalized Face Generation. We compare across four categories of text prompts including Accessory, Style, Action, and Context, following [21, 35]. Results are formatted as \"Identity Preservation (\\%) / Prompt Consistency (\\%)\" and we bold the higher number for each pair of comparison." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.246, + 0.482, + 0.321 + ], + "angle": 0, + "content": "These results demonstrate that while our method is specifically designed for personalized HOI scenarios, it effectively preserves the generative capabilities of the base models and does not degrade their performance on general text prompts, showcasing its adaptability and robustness." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.325, + 0.485, + 0.656 + ], + "angle": 0, + "content": "Effectiveness of Individual Component. We conducted an ablation study on FastComposer [35] to evaluate the contributions of the Cross-Attention Constraint (CAC), Latent Merge (LM), and Residual Merge (RM) in PersonaHOI. As shown in Table ??, the full model, integrating all three components, achieves the best performance across all metrics, underscoring the complementary roles of our proposed modules. Removing LM or RM results in significant declines in both Identity Preservation and Interaction Alignment, highlighting the critical role of these merging strategies. The absence of CAC has a relatively smaller effect on Interaction Alignment, but it substantially impacts Identity Preservation. This is because without CAC, faces generated by PFD fail to align with the spatial layout provided by SD, disrupting the coherence between PFD and SD branches and leading to diminished alignment and identity fidelity in facial regions. In summary, the ablation results confirm the distinct and critical functionalities of each module. CAC ensures precise spatial alignment of faces between PFD and SD, while LM and RM are indispensable for consistently merging non-facial HOI-relevant features from SD with facial identity details from PFD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results on Additional Baselines. To further validate the efficacy of our approach, we conduct additional evaluations using InstantID [34] and PuLID [7]. Results in Table 3 and Figure 7 indicate PersonaHOI can benefit both PFD methods, notably improving HOI alignment and image-text consistency. InstantID shares design similarities with IP-Adapter, prioritizing portrait generation, which improves identity preservation but compromises interaction fidelity. The fusion mechanism of PersonaHOI significantly enhances HOI alignment and overall image-text consistency. PuLID, on the other hand, balances personalization against text-guided image synthesis capability during training. However, it can struggle to correctly place the object in some cases (second row in Figure 7). By incorporating PersonaHOI, we achieve consistent improvements across identity preservation, image-prompt consistency, and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.246, + 0.905, + 0.276 + ], + "angle": 0, + "content": "interaction alignment, demonstrating the versatility and robustness of our method." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.285, + 0.9, + 0.367 + ], + "angle": 0, + "content": "
MethodIdentity (%)Prompt (%)HOI (%)
InstantID76.2122.9345.73
+ Ours69.93 (-9.28)23.80 (+0.87)61.82 (+16.09)
PuLID55.9324.5072.10
+ Ours62.56 (+6.63)25.21 (+0.71)72.18 (+0.08)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.377, + 0.906, + 0.404 + ], + "angle": 0, + "content": "Table 3. Comparison of our method with InstantID and PuLID on HOI-Specific Personalized Face Generation." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.424, + 0.902, + 0.541 + ], + "angle": 0, + "content": "
Ref. ImageInstantID+OursRef. ImagePuLD+Ours
Caption: A man surfboarding with a surfboardCaption: A woman catching a frisbee
Caption: A woman riding a horseCaption: A woman snowboarding with a snowboard
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.557, + 0.906, + 0.584 + ], + "angle": 0, + "content": "Figure 7. Visual comparison of InstantID [34] and PuLID [7] with our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.64, + 0.62 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a training&tuning-free framework that combines the strengths of Personalized Face Diffusion (PFD) models and StableDiffusion (SD) models to generate personalized images featuring realistic and complex human-object interactions (HOI). By introducing Cross-Attention Constraint, Latent Merge, and Residual Merge, our approach achieves seamless integration of identity-specific facial features from PFD with contextual interaction details from SD, ensuring both spatial alignment and contextual coherence. The flexibility of our framework allows adaptation to various PFD and SD architectures, demonstrating its robustness in handling diverse interaction scenarios. Without requiring additional fine-tuning or task-specific datasets, our method significantly improves identity retention and interaction realism compared to state-of-the-art approaches. This highlights the potential of leveraging pre-trained models to enable high-quality, personalized content generation across diverse applications." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23782" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovise: Tuning-free image customization with text-to-image diffusion models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.244, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[3] Zhuowei Chen, Shancheng Fang, Wei Liu, Qian He, Mengqi Huang, Yongdong Zhang, and Zhendong Mao. Dreamidentity: Improved editability for efficient face-identity preserved image generation. arXiv preprint arXiv:2307.00300, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.369 + ], + "angle": 0, + "content": "[4] Dylan Campbell Frederic Z. Zhang and Stephen Gould. Spatially Conditioned Graphs for Detecting Human-Object Interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13319–13327, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.372, + 0.483, + 0.441 + ], + "angle": 0, + "content": "[5] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel CohenOr. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.443, + 0.483, + 0.498 + ], + "angle": 0, + "content": "[6] Chen Gao, Si Liu, Defa Zhu, Quan Liu, Jie Cao, Haoqian He, Ran He, and Shuicheng Yan. InteractGAN: Learning to Generate Human-Object Interaction. In Association for Computing Machinery, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.483, + 0.556 + ], + "angle": 0, + "content": "[7] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. PuLID: Pure and Lightning ID Customization via Contrastive Alignment. In Advances in Neural Information Processing Systems, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.558, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[8] Saurabh Gupta and Jitendra Malik. Visual Semantic Role Labeling. arXiv preprint arXiv:1505.04474, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[9] Junjie He, Yifeng Geng, and Liefeng Bo. UniPortrait: A Unified Framework for Identity-Preserving Single- and Multi-Human Image Personalization. arXiv preprint arXiv:2408.05939, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[10] Jiu Tian Hoe, Xudong Jiang, Chee Seng Chan, Yap-Peng Tan, and Weipeng Hu. InteractDiffusion: Interaction Control in Text-to-Image Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.483, + 0.77 + ], + "angle": 0, + "content": "[11] Siteng Huang, Biao Gong, Yutong Feng, Xi Chen, Yuqian Fu, Yu Liu, and Donglin Wang. Learning disentangled identifiers for action-customized text-to-image generation. arXiv preprint arXiv:2311.15841, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[12] Junha Hyung, Jaeyo Shin, and Jaegul Choo. Magicapture: High-resolution multi-concept portrait customization. arXiv preprint arXiv:2309.06895, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[13] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing. arXiv preprint arXiv:2312.11392, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Jian-Yu Jiang-Lin, Kang-Yang Huang, Ling Lo, Yi-Ning Huang, Terence Lin, Jhih-Ciang Wu, Hong-Han Shuai, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Wen-Huang Cheng. ReCorD: Reasoning and Correcting Diffusion for HOI Generation. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[16] Xiaoming Li, Xinyu Hou, and Chen Change Loy. When stylegan meets stable diffusion: a w+ adapter for personalized image generation. arXiv preprint arXiv:2311.17461, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[17] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-Set Grounded Text-to-Image Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[18] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.907, + 0.445 + ], + "angle": 0, + "content": "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep Learning Face Attributes in the Wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[20] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[21] Xu Peng, Junwei Zhu, Boyuan Jiang, Ying Tai, Donghao Luo, Jiangning Zhang, Wei Lin, Taisong Jin, Chengjie Wang, and Rongrong Ji. PortraitBooth: A Versatile Portrait Model for Fast Identity-preserved Personalization. arXiv preprint arXiv:2312.06354, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[24] Iasonas Kokkinos Riza Alp Güler, Natalia Neverova. Dense-Pose: Dense Human Pose Estimation In The Wild. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23783" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.217 + ], + "angle": 0, + "content": "[27] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[28] Florian Schroff, Dmitry Kalenichenko, and James Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[29] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[30] Kaede Shiohara and Toshihiko Yamasaki. Face2diffusion for fast and editable face personalization. arXiv preprint arXiv:2403.05094, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[31] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. FreeU: Free Lunch in Diffusion U-Net. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[32] Yu-Chuan Su, Kelvin CK Chan, Yandong Li, Yang Zhao, Han Zhang, Boqing Gong, Huisheng Wang, and Xuhui Jia. Identity encoder for personalized diffusion. arXiv preprint arXiv:2304.07429, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.483, + 0.511 + ], + "angle": 0, + "content": "[33] Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. In SIGGRAPH Asia 2023 Conference Papers, pages 1-10, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[34] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. InstantID: Zero-shot Identity-Preserving Generation in Seconds. arXiv preprint arXiv:2401.07519, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[35] Guangxuan Xiao, Tianwei Yin, William T Freeman, Fredo Durand, and Song Han. FastComposer: Tuning-Free Multi-Subject Image Generation with Localized Attention. In International Journal of Computer Vision, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[36] Yuxuan Yan, Chi Zhang, Rui Wang, Yichao Zhou, Gege Zhang, Pei Cheng, Gang Yu, and Bin Fu. Facestudio: Put your face everywhere in seconds. arXiv preprint arXiv:2312.02663, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[37] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IpAdapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[38] Hangjie Yuan, Jianwen Jiang, Samuel Albanie, Tao Feng, Ziyuan Huang, Dong Ni, and Mingqian Tang. RLIP: Relational Language-Image Pre-training for Human-Object Interaction Detection. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[39] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Samuel Albanie, Yining Pan, Tao Feng, Jianwen Jiang, Dong Ni, Yingya Zhang, and Deli Zhao. RLIPv2: Fast Scaling of Relational Language-Image Pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[40] Frederic Z. Zhang, Dylan Campbell, and Stephen Gould. Efficient Two-Stage Detection of Human-Object Interactions with a Novel Unary-Pairwise Transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20104-20112, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[41] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In IEEE International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.906, + 0.359 + ], + "angle": 0, + "content": "[43] Xulu Zhang, Xiao-Yong Wei, Wengyu Zhang, Jinlin Wu, Zhaoxiang Zhang, Zhen Lei, and Qing Li. A Survey on Personalized Content Synthesis with Diffusion Models. arXiv preprint arXiv:2405.05538, 2024. https://arxiv.org/abs/2405.05538.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[44] Yuxin Zhang, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. MotionCrafter: One-Shot Motion Customization of Diffusion Models. arXiv preprint arXiv:2312.05288, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23784" + } + ] +] \ No newline at end of file diff --git a/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/00e10d7f-1169-4801-84c3-c1efaaab378c_origin.pdf b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/00e10d7f-1169-4801-84c3-c1efaaab378c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a376898524797d54aa1149affeee9c50149c4770 --- /dev/null +++ b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/00e10d7f-1169-4801-84c3-c1efaaab378c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51d04214fca654a1b293256dce4a806cebd264340c26b28793a7832a75d6c929 +size 9810027 diff --git a/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/full.md b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1469ccd475e04067b8687a9a2ed304ae828c52f7 --- /dev/null +++ b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/full.md @@ -0,0 +1,377 @@ +# PersonaHOI: Effortlessly Improving Face Personalization in Human-Object Interaction Generation + +Xinting Hu* + +xhu@mpi-inf.mpg.de + +Haoran Wang* + +hawang@mpi-inf.mpg.de + +Jan Eric Lenssen + +jlenssen@mpi-inf.mpg.de + +Bernt Schiele + +schiele@mpi-inf.mpg.de + +Max Planck Institute for Informatics, Saarland Informatics Campus, Germany + +# Context + HOI + +A man riding a horse in front of a medieval castle + +![](images/b157e54b54f7e59e72c4458fd80cbff15435b7be97458c9d9108f2e545625445.jpg) + +![](images/eee15c6a3d7f2f66d0fe64bb09f85357a7a22b1765f09202b4ca9899a2e8dc7c.jpg) +FastComposer + +![](images/e7b41c0e0d4124d394e79e4ca3536a1d6653346a3418912f9d207e1de6ed09cc.jpg) ++Ours + +# Style + HOI + +A woman holding a cat in the style of Vincent Van Gogh + +![](images/7435031c37e19ec29770cf86db560bcce83ba502dd54db71fd0c9938e201e473.jpg) + +![](images/52fd20b7f4bdff5b304a657d125772808fe19d54d661aa9f3d435952bf704d63.jpg) +FastComposer + +![](images/320948f929912f5425969af98881dfbdb9780f8f9d4101acbe1558403e04cd82.jpg) ++Ours + +![](images/247ba23b3aa2cf8aedd490ce8aa68908e1ba25b335540164fe2bcc36869980d7.jpg) + +![](images/ab7c40c3cf4033effea10b05a9a3a6eefde85db862f49281f0c995aef0658d10.jpg) +Various man-Object action (HOI) descriptions + +![](images/95dd086b65595a579d3d425c6609f565a3eda7f96aba4b06315864382c6834dd.jpg) + +![](images/a1cfa0df5c025adff299d9824ef2eb90e13a8db98b4b547113d3d8783195ebc8.jpg) + +![](images/71f314f3842b0aa7dc66879f5b8a05ec49b4b5049469ab393b694a608d44ac77.jpg) + +![](images/8ea63ada345a13ba311df7292efa55d208d946c220d7f6951aa595a6e0a2438c.jpg) + +![](images/a9f97ef19f0729a0e4aac6e809f40bd6a55abf839ae1a6219b9e08b6f8a40aae.jpg) + +![](images/3dedd8517a0228a615656366ae664bc76fd2c3aa79f5b8aed21058ddc60439f8.jpg) + +# Personalized + +Face with + +Human-Object + +Interaction + +Training + +& + +Tuning + +Free! + +# Accessory + HOI + +A woman wearing a Santa hat and taking selfie with phone + +![](images/cc0f5f8077b75ab9c225f3f0cb46c7ed55d42813cb81362f2beca721fbafc31c.jpg) +Figure 1. Examples of Personalized Face with Human-Object Interaction (HOI) Generation. We present PersonaHOI, a training- and tuning-free framework built on existing diffusion models. Using a single reference image and diverse HOI prompts, PersonaHOI generates identity-consistent human-object interactions compared to FastComposer [35]. PersonaHOI can further seamlessly integrate varied contexts, styles, accessories, and multi-person scenarios, ensuring scalability and practicality for real-world applications. + +![](images/88782716e9e95372ba6ba9459feb421f25b10a7b26fdf4d4fc7dd3761727640f.jpg) +FastComposer + +![](images/4185e7907c4f3d2a7e93cdbb4b654b9ddd9d1743380b26001350fefc76257d0b.jpg) ++Ours + +# Mult-Person + HOI + +A man and a man are playing guitar in a bar + +![](images/db2c67893599da98510a6773350c766abb93adb5aead9340d04282c538714726.jpg) + +![](images/11fdfc71f8f8d066b87ca4a669ea495afb8320c9778279662b2a875a9f49a960.jpg) +FastComposer + +![](images/a667851efcf6131abd92aa22122c5ed3199961961de8a62edb9ab49dff6607a7.jpg) ++Ours + +# Abstract + +We introduce PersonaHOI, a training- and tuning-free framework that fuses a general StableDiffusion model with a personalized face diffusion (PFD) model to generate identity-consistent human-object interaction (HOI) images. While existing PFD models have advanced significantly, they often overemphasize facial features at the expense of full-body coherence, PersonaHOI introduces an additional + +StableDiffusion (SD) branch guided by HOI-oriented text inputs. By incorporating cross-attention constraints in the PFD branch and spatial merging at both latent and residual levels, PersonaHOI preserves personalized facial details while ensuring interactive non-facial regions. Experiments, validated by a novel interaction alignment metric, demonstrate the superior realism and scalability of PersonaHOI, establishing a new standard for practical personalized face with HOI generation. Code is available at https://github.com/JoyHuYY1412/PersonaHOI. + +![](images/dd865b7a270e59debd3e386ac1a78d3ee6ace7beac4b9887245343fa0cc04954.jpg) +Figure 2. (a) The spatial layout of StableDiffusion guides PersonaHOI to generate personalized content with coherent human-object interactions (HOI). (b) Analysis of identity injection timing in PFD models. We use FastComposer [35] for diffusion model generation. Injecting face representation at the start of image generation preserves facial details but lacks coherent HOI, while delayed injection continuously deviates from the original identity, resulting in random human features and meaningless human-object interactions. + +# 1. Introduction + +Personalized face generation has seen increasing public interest and demand in user-specified digital content creation. Current learning-based personalized face diffusion (PFD) models [18, 21, 29, 35], trained on large-scale face-centric datasets, can incorporate a single user-provided reference facial image into a text-to-image generation process, enabling the rapid creation of images that depict specific persons in diverse scenes, outfits, and styles within seconds. + +Although these methods perform well in simpler scenarios, they often struggle with generating full-body depictions involving detailed human-object interactions (HOI). As illustrated in Figure 1, the generated images of FastComposer [35] show missing objects or body parts, resulting in portrait-focused outputs without HOI information. This limitation compromises the overall realism and practical utility of the generated content, especially in immersive or interactive applications. + +To diagnose the limitations of PFD models, we examine the results of PFD models by varying the timing of identity injection during the generation process. As seen in the first and last columns of Figure 2(b), early injection effectively preserves facial details from the reference image, but relying solely on text input (last column) results in incoherent HOI outputs. This analysis highlights that the core challenge lies not in maintaining identity fidelity but in generating coherent body movements and interactions driven by text prompts. Notably, pre-trained models used for PFD, such as StableDiffusion [25], can generate satisfactory HOI images due to their extensive and diverse training data. This observation implies that the fine-tuning process on face-centric datasets of PFD models diminishes their ability to follow complex HOI text prompts. Therefore, to advance + +current PFD methods in HOI tasks, reintroducing the capability to leverage text prompts for natural full-body interactions is essential. + +Motivated by this need, we propose a straightforward yet powerful approach: leveraging the generative capabilities of pre-trained diffusion models, such as StableDiffusion (SD), to augment existing PFD frameworks. Our method, PersonaHOI, integrates the strengths of SD without requiring additional training or fine-tuning, thereby restoring the ability to generate realistic, text-driven full-body HOI while preserving identity-specific details from the reference image. + +Central to this integration, we incorporate an additional SD branch that aligns identity-specific details from PFD models with the HOI layouts generated by SD. A head segmentation mask derived from the SD output guides the merge of non-facial components from the SD branch with the facial details from the PFD branch. Specifically, in the PFD model branch, we introduce a Cross-Attention Constraint to prevent the overemphasis of identity features across the entire image, ensuring these details are confined to the facial region. To integrate HOI representations from SD, we implement Latent Merge and Residual Merge, merging identity and contextual features at both the latent representation level and via skip connections within the U-Net architecture. This multi-level merging strategy ensures that the generated images maintain realistic human-object interactions while preserving personalized facial details, resulting in cohesive, interaction-rich outputs. + +To evaluate the effectiveness of our method in HOI scenarios, we propose a novel "interaction alignment" metric. These metrics leverage HOI detectors to measure alignment between generated images and text prompts, objectively assessing interaction realism. Our contributions can be summarized as follows: + +- We present PersonaHOI, a dual-path architecture that integrates general and personalized diffusion models. Our framework generates realistic human-object interactions with specific identities, without requiring additional training or test-time tuning. +- Our method introduces key components, including crossattention constraints and a combination of latent and residual merge strategies. These innovations enable effective feature integration, ensuring that generated images maintain both interaction coherence and identity fidelity. +- We design a novel evaluation metric for human-object interaction. Our experiments show significant improvements over state-of-the-art face personalization techniques in terms of interaction realism across models, emphasizing the scalability and robustness of our approach. + +# 2. Related Works + +Personalized Face Generation Diffusion Models. Personalized face generation aims to create identity-consistent images of individuals from limited reference data [32, 36, 43]. Given pretrained general image generation models [25], traditional optimization-based approaches [5, 26, 27] refine model parameters during inference for each individual but are computationally expensive, making them unsuitable for real-time applications. In contrast, learning-based approaches leverage human-centric datasets [1, 15, 19, 20] to train robust models capable of varying identity inputs without requiring test-time fine-tuning. Facial representations, extracted by models like CLIP [23] or other face recognition models [3, 16, 33], are often fused with [18, 21, 35] or replace [30] textual embeddings of specific words, such as "person", to provide identity conditions for diffusion models and enhance contextual relevance [3, 29]. In IP-Adapter [37] and other works [7, 9, 34], facial features are processed separately through decoupled cross-attention modules. Incorporating human masks as prior, methods can clean the background during data processing; alternatively, masks can be leveraged to construct [2, 18, 21, 35] and adjust [12] loss functions during training. In our work, we extend face personalization to complex human-object interactions (HOI), achieving a higher level of personalization and adaptability for diverse generative tasks. + +Human-Object Interaction in Diffusion Models. HOI detection aims to locate human-object pairs and categorize their interactions as triplets, such as (person, playing, football) [4, 38-40], while HOI image synthesis remains relatively under-explored. InteractGAN [6] generates HOI images using human pose templates and reference images, but its reliance on pose-template pools and object references limits flexibility. Recent diffusion-based approaches tackle HOI generation by incorporating layout controls such as bounding boxes [10, 14, 17] or human poses [17, 42], or by leveraging in-context samples with similar interac + +tions [11, 44]. Additionally, ReCorD [14] integrates Latent Diffusion Models with Visual Language Models, introducing modules for reasoning and correcting interactions to enhance cross-modal alignment. In this work, we rely on the standard text-to-image diffusion model (i.e., StableDiffusion [25]) for HOI generation, and fuse the generation process with personalized face diffusion models to generation identity-preserving HOI images. + +# 3. Preliminary + +Understanding the fundamentals of StableDiffusion (SD) is essential to leverage its generative capabilities within our PFG framework. SD is a latent diffusion model designed to generate high-quality images based on text or image prompts. It operates in a latent space, where an input image is encoded in a latent representation $\mathbf{z}_0$ using a pretrained variational autoencoder (VAE) encoder, reducing the dimensionality of the image for efficient processing. + +During the generation phase, the SD model begins with an initial noisy latent representation $\mathbf{z}_T$ and progressively denoises it over $T$ timesteps to reconstruct $\mathbf{z}_0$ . This denoised latent is then decoded by the VAE decoder to produce the final image. At each timestep $t$ , the U-Net architecture $\theta$ predicts noise $\epsilon (\mathbf{z}_t,t,\mathbf{C})$ , where $\mathbf{C}$ represents the conditioning inputs (text or image embeddings). The iterative denoising process follows: + +$$ +\mathbf {z} _ {t - 1} = \operatorname {D e n o i s e} \left(\mathbf {z} _ {t}, \epsilon \left(\mathbf {z} _ {t}, t, \mathbf {C}\right); \theta\right) \tag {1} +$$ + +The U-Net architecture comprises an encoder-decoder structure with skip connections that transfer features between the encoder and the decoder. Within this architecture, SD employs a cross-attention mechanism to incorporate conditioning inputs effectively during denoising. This mechanism projects latent representations and conditioning inputs into query, key, and value spaces: + +$$ +\left\{ \begin{array}{l} \mathbf {Q} = \mathbf {W} _ {q} \mathbf {z}, \mathbf {K} = \mathbf {W} _ {k} \mathbf {C}, \mathbf {V} = \mathbf {W} _ {v} \mathbf {C}; \\ \mathbf {A} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {\top}}{\sqrt {d}}\right), \mathbf {z} _ {\text {a t t n}} = \mathbf {A V}. \end{array} \right. \tag {2} +$$ + +Here, $\mathbf{A}$ is the attention map that ensures that relevant conditioning tokens guide denoising at each generation step. This cross-attention mechanism is critical for aligning the generated image with the input prompts, enabling consistency and fidelity in the final output. + +# 4. Method + +# 4.1. Overall Architecture + +As illustrated in Figure 3, PersonaHOI augments the PFD model by integrating an additional StableDiffusion (SD) branch, enhancing its capacity to generate personalized images that capture complex human-object interactions (HOI) + +![](images/0bfaab5951f489661d45979cb536bf2ced2fa25a52626df45a24eec4946d2160.jpg) +Figure 3. Overview of Our Proposed Framework, PersonaHOI. The architecture integrates a personalized face diffusion (PFD) model with an additional StableDiffusion (SD) branch. First, SD generates an image $(I_{SD})$ from a text prompt and noisy latent representation $(z_{T})$ , which is decoded and segmented to produce a head mask. Next, SD and PFD run in parallel from the same $z_{T}$ . At every timestep $t$ , the head mask guides the Cross-Attention Constraint in PFD and merging modules (Latent Merge and Residual Merge) to merge interaction-relevant features from SD with identity-specific details from PFD. Iteratively, this process introduces HOI context to personalized face generation in a training&tuning-free manner. + +in a training-free manner. Given an input comprising a user-provided reference image and a text prompt specifying the interaction (e.g., "a person kicking football"), our framework produces a cohesive output image that retains the subject's identity while accurately depicting the specified HOI. + +To effectively merge identity from PFD and interaction features from SD, we introduce three core strategies: Cross-Attention Constraint, which regulates attention to reference image features within the PFD model (detailed in Sec. 4.2); Latent Merge, which combines features in the latent space at each generation timestep (Sec. 4.3); and Residual Merge, which integrates identity details through skip connections in the U-Net architecture (Sec. 4.4). Together, these strategies enable PersonaHOI to align generated content with interaction-driven text prompts while maintaining identity fidelity. Furthermore, this design allows PersonaHOI adaptable to various PFD models without the need for model retraining or test-time tuning. + +# 4.2. Cross-Attention Constraint (CAC) + +In the cross-attention layers of the PFD model, conditioning inputs $\mathbf{C} \in \mathbb{R}^{N \times D}$ include both text and image embeddings to guide the generation process, where $N$ is the number of tokens, and $D$ is the embedding dimension. Following established PFD model approaches [18, 21, 35], the $n_{img}$ -th token in $\mathbf{C}$ incorporates the image embedding, encapsulat + +ing identity-specific details essential for facial preservation, while the remaining $N - 1$ tokens correspond to text embeddings. In standard PFD models, attention to the $n_{img}$ -th token often spreads across the image, overemphasizing facial features thus decreasing spatial capacity for representing body movements and object interactions. + +To address this, we introduce Cross-Attention Constraint (CAC) within the cross-attention layers of the PFD model. CAC restricts identity features to specific facial regions, ensuring sufficient spatial capacity for body and object interactions. Different from the attention localization loss used in training [21, 35], our CAC utilizes fine-grained head mask derived from the SD-generated image $I_{SD}$ in image generation process, ensuring the attention constraint in PFD is consistent with the HOI layout generated by SD. The head mask is segmented by an off-the-shelf head segmentor [24], then resized to match the spatial dimensions $(H \times W)$ of the latent representation $\mathbf{z}$ , denoted as $\mathbf{M}^{\mathrm{head}} \in [0,1]^{H \times W}$ (1 for the head region and 0 for other regions). We apply $\mathbf{M}^{\mathrm{head}}$ to the attention map $\mathbf{A} \in \mathbb{R}^{(H \times W) \times N}$ as follows: + +$$ +\mathbf {M} _ {i \in [ 0, 1, \dots , N - 1 ]} ^ {C A C} = \left\{ \begin{array}{l l} \text {f l a t t e n} (\mathbf {M} ^ {\text {h e a d}}), & \text {i f} i = i m g, \\ 1, & \text {o t h e r w i s e} \end{array} \right. \tag {3} +$$ + +$$ +\mathbf {A} ^ {C A C} = \mathbf {A} \odot (\mathbf {M} ^ {C A C}) ^ {\intercal}. +$$ + +$\mathbf{M}^{CAC}$ modifies the attention weights for the $n_{img}$ token + +![](images/91e0629e18e3c345132d2a0ab10c395fffbb083ed48a7dc74dd08cdbb647fb69.jpg) +Figure 4. Illustration of Residual Merge. In each residual layer, Residual Merge operates within the U-Net skip connections, utilizing a head mask to guide the integration of high-frequency identity details from PFD residuals and low-frequency interaction layouts from SD residuals. The merged residuals are then concatenated to the corresponding bottleneck features from PFD. + +by setting the weights to 0 in regions outside the head, preventing the influence of human identity on non-facial areas. Since the head mask is derived from SD-generated outputs, CAC maintains facial layout coherence with SD. When further combined with Latent and Residual Merge, CAC helps blend facial features from PFD and non-facial interaction features from SD without mutual interference. + +# 4.3. Latent Merge (LM) + +In this section, we introduce Latent Merge, which directly merges the latent representation within U-Net from PFD and SD models. After applying Cross-Attention Constraint (CAC) in Sec. 4.2, the facial region generated by the PFD model becomes spatially aligned with the corresponding region in the broader HOI layout produced by the SD model. This alignment facilitates spatial merge, blending facial identity details from PFD with interaction contexts from SD in latent space. + +At each diffusion timestep $t$ of PFD and SD, we implement the latent space merging with $\mathbf{M}^{head}$ as follows: + +$$ +\mathbf {z} _ {t} = \mathbf {M} ^ {h e a d} \odot \mathbf {z} _ {t} ^ {\mathrm {P F D}} + \left(1 - \mathbf {M} ^ {h e a d}\right) \odot \mathbf {z} _ {t} ^ {\mathrm {S D}}. \qquad (4) +$$ + +Here, $\mathbf{z}_t^{\mathrm{PFD}}$ and $\mathbf{z}_t^{\mathrm{SD}}$ represents the latent features from PFD and SD at timestep $t$ , respectively. The merged latent $\mathbf{z}_t$ then serves as input for both SD and PFD models in the next denoising timestep $t - 1$ . This merging strategy allows identity details retained within facial regions, while non-facial regions maintain the HOI context generated by SD. + +# 4.4. Residual Merge (RM) + +While Latent Merge enables interactions between the SD and PFD at the latent level, it lacks integration at the intermediate feature stage within the U-Net. To address this, + +we introduce Residual Merge, which enhances feature integration within the skip connections. In the U-Net, skip connections transmit residual features containing rich detail from the encoder to the decoder, significantly impacting the content and quality of the generated images [13, 31]. + +Our Residual Merge is applied to each skip connection layer $l$ ( $l = 1, \dots, L$ ), as shown in Figure 4. First, residual features are extracted from both the PFD and SD U-Net backbones. A low-pass filter and high-pass filter is applied to the SD and PFD residuals, respectively. Low-pass and high-pass filters are applied to these residuals to balance global coherence and local precision: low-pass filtering SD residuals suppresses high-frequency noise to retain the broader scene layout, while high-pass filtering PFD residuals emphasizes fine-grained identity-specific details. To achieve precise integration between facial and non-facial regions, we apply the head segmentation mask $\mathbf{M}^{\mathrm{head}}$ , resized to match the residual resolution at layer $l$ , denoted as $\mathbf{M}_{\mathrm{R}}^{l}$ . The fusion of residual features from both paths is then implemented for each layer $l$ as follows: + +$$ +\mathbf {R} _ {\text {m e r g e d}} ^ {l} = \mathbf {M} _ {R} ^ {l} \odot \mathrm {H P} \left(\mathbf {R} _ {\text {P F D}} ^ {l}\right) + \left(1 - \mathbf {M} _ {R} ^ {l}\right) \odot \mathrm {L P} \left(\mathbf {R} _ {\text {S D}} ^ {l}\right), \tag {5} +$$ + +where $\mathrm{HP}(\cdot)$ and $\mathrm{LP}(\cdot)$ denote the high pass and low pass filter, $\mathbf{R}_{\mathrm{PFD}}$ and $\mathbf{R}_{\mathrm{SD}}$ denote the residual features from the PFD and SD paths, respectively. The merged residual $\mathbf{R}_{\mathrm{merged}}$ is then concatenated with the bottom-up features in the PFD U-Net for further processing. + +The Residual Merge strategy enables fine-grained feature integration within the U-Net. Through this, facial regions retain high-frequency identity details from PFD while non-facial regions incorporate interaction layouts from SD, ensuring cohesive and contextually aligned outputs. + +# 5. Experiments + +# 5.1. Setup + +Evaluation Data. We use two test sets for evaluation: + +1) HOI-Focused Personalized Face Generation: To test complex human-object interactions (HOI), we use interaction labels from the widely-used V-COCO [8] dataset. V-COCO provides diverse HOI scenarios for robust evaluation of interaction fidelity. We construct 30 prompts in the "subject + interaction + object" format (e.g., "man kicking a ball") with the same 15 reference subjects as in 1). +2) General Personalized Face Generation: Following prior studies [21, 35], we compile a benchmark with 15 reference subjects, each paired with 40 prompts spanning diverse scenarios like context, stylization, accessory, and actions. These prompts aim to assess both identity retention and the model's ability to adapt to various text-guided settings. + +Full prompt details are provided in the Appendix ?? + +Metrics. We employ the following evaluation metrics: + +
MethodArchitectureIdentity Preservation (%)Prompt Consistency (%)HOI Alignment (%)
StableDiffusion v1.5 [25]SD1.5NA24.9660.23
StableDiffusion XL [22]SDXLNA25.3270.78
FastComposer [35]SD1.553.5721.3035.96
+ OursSD1.555.28(+1.71)23.16(+1.86)56.65(+20.69)
IP-Adapter [37]SDXL62.8622.7349.83
+ OursSDXL55.74(-7.11)24.20(+1.47)68.30(+18.47)
PhotoMaker [18]SDXL50.5423.9656.77
+ OursSDXL51.34(+0.80)24.34(+0.38)76.01(+19.24)
+ +Table 1. Comparison of Our Method with Baseline Approaches on HOI-Specific Personalized Face Generation. StableDiffusion serves as the text-only baseline without subject conditioning. PersonaHOI seamlessly incorporates existing Personalized Face Diffusion models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) with their corresponding StableDiffusion architectures. We bold the higher number for each pair of comparison. + +![](images/c25e15c29661e3c6175bbf9f426a5082b5f09d14764dc55613f6f9362e9d0866.jpg) +Figure 5. Qualitative Examples of PersonaHOI and Baseline Models. Comparison of baseline models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) and their PersonaHOI-enhanced results for diverse human-object interaction prompts. + +1) Identity Preservation: Measures how well the generated images retain the identity of the reference image. Face detection is performed using MTCNN [41], and identity similarity is calculated with FaceNet [28], following [21, 35]. 2) Prompt Consistency: Evaluates text-image alignment using CLIP-based [23] scores, where higher values indicate better prompt adherence. +3) HOI Alignment (Ours): Assesses how well the gen + +erated images depict specified human-object interactions. We adopt UPT [40], an HOI detector pre-trained on V-COCO dataset, to identify the interaction specified in text prompts for generated images with HOI alignment score $S_{\mathrm{HOI}}$ . Specifically, UPT provides confidence scores for detected human and object classes ( $S_{\mathrm{Human}}$ , $S_{\mathrm{Object}}$ ) from its object detection head, as well as interaction logits ( $L_{\mathrm{Inter}}$ ) from its interaction head. With these outputs, HOI align- + +ment score $S_{\mathrm{HOI}}$ is defined as: + +$$ +S _ {\mathrm {H O I}} = \left(S _ {\text {H u m a n}} \cdot S _ {\text {O b j e c t}}\right) ^ {\lambda} \cdot \sigma \left(L _ {\text {I n t e r}}\right) \tag {6} +$$ + +where $\sigma$ is the sigmoid activation function and $\lambda = 2.8$ . We validate the human alignment and effectiveness for our proposed HOI Alignment metric in Appendix ?? + +Compared Methods and Implementation Details. We compare our method with state-of-the-art learning-based personalized face diffusion models, including FastComposer [35], IP-Adapter [37], and PhotoMaker [18]. Our PeronaHOI used their pre-trained diffusion backbones as the additional branch, i.e., StableDiffusion v1.5 for FastComposer and StableDiffusion XL for PhotoMaker, with output resolutions set to $512 \times 512$ and $1024 \times 1024$ , respectively. Head segmentation for SD generated image is performed with DensePose [24] due to its robust segmentation accuracy. In our generation process, no additional training or test-time tuning is applied to any compared methods. + +# 5.2. Personalized Face with HOI Generation + +Qualitative Comparison. Figure 5 illustrates qualitative comparisons between baseline models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) and their PersonaHOI-enhanced counterparts for diverse human-object interaction prompts. Baseline methods often fail to produce meaningful interactions, leading to missing or poorly integrated objects and unnatural human-object dynamics. For instance, in the prompt "A man skateboarding with skateboard" (first row of Figure 5), FastComposer [35] and PhotoMaker omit the skateboard entirely and IP-Adapter fails to generate the accurate interaction ("holding" the skateboard instead), resulting in inconsistent outputs with the text prompt. In contrast, PersonaHOI-enhanced models generate realistic skateboard placement and natural human-object dynamics. Moreover, PersonaHOI also ensures high-fidelity preservation of facial identity given in reference images. + +Quantitative Comparison. Table ?? compares the performance of PersonaHOI across various baseline models. For FastComposer [35] and PhotoMaker [18], PersonaHOI achieves consistent improvements across all metrics. Notably, Interaction Alignment improves by $20.69\%$ for FastComposer [35] and $19.24\%$ for PhotoMaker [18], demonstrating superior human-object interaction coherence. Gains in Identity Preservation and Prompt Consistency further validate PersonaHOI's ability to retain identity while enhancing text-image alignment. For IP-Adapter [37], a decrease in Identity Preservation (from $62.86\%$ to $55.74\%$ ) is observed. This drop is attributed to IP-Adapter's inherent bias toward generating face-centric images, which inflates identity scores at the cost of interaction fidelity. PersonaHOI mitigates this trade-off by significantly improving Interaction Alignment $(+18.47\%)$ and Prompt Consistency $(+1.47\%)$ , balancing identity preser + +![](images/2bdbe366dc80e66c0d1f34dec8e05c7e98318476de53bf0c1cb2ba4579b6c4bb.jpg) +SD v1.5 + +![](images/9c7ab5efea08c132040a60736f44be0a123040bc8a3306fa06d4f8ab0c2d6aba.jpg) +Reference Images + +![](images/2d1c715e371185d08bb0e98d4eab87bfd1ba5dfb9b19e4e9ffb9eed581844a28.jpg) +ce Images + +![](images/01f50b740e50ab61d8587932a7128dac0de71e3945069cfaa5133e81453d5728.jpg) +FastComposer + +![](images/d188e5707d7e956c4753e1241c1af0dd4bd77e961969e79aca207cc8b0539f4a.jpg) ++Ours + +![](images/75441ba04fcf1bbe5df8df5022fd18409f1cacd8e9aa466b7c6399187666c300.jpg) +Caption: A man and a man planting a tree + +![](images/e3b100927aa3fd3b79fffd8c1a4a7e09c25b936a4e4083643a8e9028659e81ed.jpg) +Caption: A man and a man cycling together + +![](images/c18d9f62666ffd918d33e408363154d6e2ecefc1885cffb797d9e2e159453a39.jpg) +Figure 6. Qualitative results for multi-subject HOI generation. We compare generation outputs from SD v1.5 [25], FastComposer [35], and our PersonaHOI based on FastComposer [35] with different multi-subject interaction prompts. PersonaHOI not only preserves distinct identities but also generates coherent human-object interactions aligned with HOI layouts produced by SD v1.5. + +![](images/584d47e0e45e91815ee42c6f9b5991aa3e4908462f72b80c26f98653174ccaf3.jpg) + +![](images/c4905f95b6f1e79482f208088e00690b63521c3393c87359f17f69a4d238f1e6.jpg) + +vation with realistic interaction synthesis. Despite the decrease, the identity preservation score of $55.74\%$ remains competitive for personalized face generation. + +Overall, PersonaHOI bridges the gap between personalized face generation and interaction realism. Its training and tuning-free design enables seamless integration into different SD-based personalized face diffusion models, providing scalability and adaptability across diverse architectures and application scenarios. + +# 5.3. Personalized Multi-Subject HOI Generation + +Generating images with multiple subjects engaged in human-object interactions (HOI) presents significant challenges in preserving distinct identities and accurately depicting interactions. As shown in Figure 6, our method enhances FastComposer [35] using its pre-trained SD v1.5 model, maintaining identity fidelity while generating coherent interactions. For instance, in the challenging "cycling together" (second line) scenario, PersonaHOI successfully preserves the unique identities of both subjects and generates two bicycles alongside a natural background. In contrast, FastComposer [35] produces two adjacent faces with minimal interaction details. These results highlight the potential of PersonaHOI in handling complex multi-subject HOI scenarios, ensuring both identity preservation and interaction realism. + +# 5.4. More Results + +General Personalized Face Generation Performance. We evaluate PersonaHOI on the task of general personalized face generation [21, 35] across four categories of text prompts: Accessory, Style, Action, and Context. As shown in Table 2, the scores across all categories for the PersonaHOI-enhanced method are consistently comparable to or slightly better than the baseline FastComposer [35]. + +
MethodAccessory (%)Style (%)Action (%)Context (%)Mean (%)
StableDiffusion v1.5 [25]NA / 26.70NA / 27.21NA / 23.66NA / 25.86NA / 25.86
FastComposer [35]54.65 / 24.2241.13 / 24.0155.35 / 21.3052.70 / 22.3150.95 / 22.96
+ Ours56.43 / 24.2546.07 / 23.9755.09 / 22.2153.77 / 22.5952.84 / 23.26
+ +These results demonstrate that while our method is specifically designed for personalized HOI scenarios, it effectively preserves the generative capabilities of the base models and does not degrade their performance on general text prompts, showcasing its adaptability and robustness. + +Effectiveness of Individual Component. We conducted an ablation study on FastComposer [35] to evaluate the contributions of the Cross-Attention Constraint (CAC), Latent Merge (LM), and Residual Merge (RM) in PersonaHOI. As shown in Table ??, the full model, integrating all three components, achieves the best performance across all metrics, underscoring the complementary roles of our proposed modules. Removing LM or RM results in significant declines in both Identity Preservation and Interaction Alignment, highlighting the critical role of these merging strategies. The absence of CAC has a relatively smaller effect on Interaction Alignment, but it substantially impacts Identity Preservation. This is because without CAC, faces generated by PFD fail to align with the spatial layout provided by SD, disrupting the coherence between PFD and SD branches and leading to diminished alignment and identity fidelity in facial regions. In summary, the ablation results confirm the distinct and critical functionalities of each module. CAC ensures precise spatial alignment of faces between PFD and SD, while LM and RM are indispensable for consistently merging non-facial HOI-relevant features from SD with facial identity details from PFD. + +Results on Additional Baselines. To further validate the efficacy of our approach, we conduct additional evaluations using InstantID [34] and PuLID [7]. Results in Table 3 and Figure 7 indicate PersonaHOI can benefit both PFD methods, notably improving HOI alignment and image-text consistency. InstantID shares design similarities with IP-Adapter, prioritizing portrait generation, which improves identity preservation but compromises interaction fidelity. The fusion mechanism of PersonaHOI significantly enhances HOI alignment and overall image-text consistency. PuLID, on the other hand, balances personalization against text-guided image synthesis capability during training. However, it can struggle to correctly place the object in some cases (second row in Figure 7). By incorporating PersonaHOI, we achieve consistent improvements across identity preservation, image-prompt consistency, and + +interaction alignment, demonstrating the versatility and robustness of our method. + +Table 2. Comparison of Our Method with FastComposer [35] on General Personalized Face Generation. We compare across four categories of text prompts including Accessory, Style, Action, and Context, following [21, 35]. Results are formatted as "Identity Preservation (\%) / Prompt Consistency (\%)" and we bold the higher number for each pair of comparison. + +
MethodIdentity (%)Prompt (%)HOI (%)
InstantID76.2122.9345.73
+ Ours69.93 (-9.28)23.80 (+0.87)61.82 (+16.09)
PuLID55.9324.5072.10
+ Ours62.56 (+6.63)25.21 (+0.71)72.18 (+0.08)
+ +Table 3. Comparison of our method with InstantID and PuLID on HOI-Specific Personalized Face Generation. + +
Ref. ImageInstantID+OursRef. ImagePuLD+Ours
Caption: A man surfboarding with a surfboardCaption: A woman catching a frisbee
Caption: A woman riding a horseCaption: A woman snowboarding with a snowboard
+ +Figure 7. Visual comparison of InstantID [34] and PuLID [7] with our method. + +# 6. Conclusions + +In this work, we propose a training&tuning-free framework that combines the strengths of Personalized Face Diffusion (PFD) models and StableDiffusion (SD) models to generate personalized images featuring realistic and complex human-object interactions (HOI). By introducing Cross-Attention Constraint, Latent Merge, and Residual Merge, our approach achieves seamless integration of identity-specific facial features from PFD with contextual interaction details from SD, ensuring both spatial alignment and contextual coherence. The flexibility of our framework allows adaptation to various PFD and SD architectures, demonstrating its robustness in handling diverse interaction scenarios. Without requiring additional fine-tuning or task-specific datasets, our method significantly improves identity retention and interaction realism compared to state-of-the-art approaches. This highlights the potential of leveraging pre-trained models to enable high-quality, personalized content generation across diverse applications. + +# References + +[1] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 3 +[2] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovise: Tuning-free image customization with text-to-image diffusion models, 2023. 3 +[3] Zhuowei Chen, Shancheng Fang, Wei Liu, Qian He, Mengqi Huang, Yongdong Zhang, and Zhendong Mao. Dreamidentity: Improved editability for efficient face-identity preserved image generation. arXiv preprint arXiv:2307.00300, 2023. 3 +[4] Dylan Campbell Frederic Z. Zhang and Stephen Gould. Spatially Conditioned Graphs for Detecting Human-Object Interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13319–13327, 2021. 3 +[5] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel CohenOr. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 3 +[6] Chen Gao, Si Liu, Defa Zhu, Quan Liu, Jie Cao, Haoqian He, Ran He, and Shuicheng Yan. InteractGAN: Learning to Generate Human-Object Interaction. In Association for Computing Machinery, 2020. 3 +[7] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. PuLID: Pure and Lightning ID Customization via Contrastive Alignment. In Advances in Neural Information Processing Systems, 2024. 3, 8 +[8] Saurabh Gupta and Jitendra Malik. Visual Semantic Role Labeling. arXiv preprint arXiv:1505.04474, 2015. 5 +[9] Junjie He, Yifeng Geng, and Liefeng Bo. UniPortrait: A Unified Framework for Identity-Preserving Single- and Multi-Human Image Personalization. arXiv preprint arXiv:2408.05939, 2024. 3 +[10] Jiu Tian Hoe, Xudong Jiang, Chee Seng Chan, Yap-Peng Tan, and Weipeng Hu. InteractDiffusion: Interaction Control in Text-to-Image Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[11] Siteng Huang, Biao Gong, Yutong Feng, Xi Chen, Yuqian Fu, Yu Liu, and Donglin Wang. Learning disentangled identifiers for action-customized text-to-image generation. arXiv preprint arXiv:2311.15841, 2023. 3 +[12] Junha Hyung, Jaeyo Shin, and Jaegul Choo. Magicapture: High-resolution multi-concept portrait customization. arXiv preprint arXiv:2309.06895, 2023. 3 +[13] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing. arXiv preprint arXiv:2312.11392, 2023. 5 +[14] Jian-Yu Jiang-Lin, Kang-Yang Huang, Ling Lo, Yi-Ning Huang, Terence Lin, Jhih-Ciang Wu, Hong-Han Shuai, and + +Wen-Huang Cheng. ReCorD: Reasoning and Correcting Diffusion for HOI Generation. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2024. 3 +[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[16] Xiaoming Li, Xinyu Hou, and Chen Change Loy. When stylegan meets stable diffusion: a w+ adapter for personalized image generation. arXiv preprint arXiv:2311.17461, 2023. 3 +[17] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-Set Grounded Text-to-Image Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[18] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 6, 7 +[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep Learning Face Attributes in the Wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 3 +[20] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 3 +[21] Xu Peng, Junwei Zhu, Boyuan Jiang, Ying Tai, Donghao Luo, Jiangning Zhang, Wei Lin, Taisong Jin, Chengjie Wang, and Rongrong Ji. PortraitBooth: A Versatile Portrait Model for Fast Identity-preserved Personalization. arXiv preprint arXiv:2312.06354, 2023. 2, 3, 4, 5, 6, 7, 8 +[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 6 +[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 6 +[24] Iasonas Kokkinos Riza Alp Güler, Natalia Neverova. Dense-Pose: Dense Human Pose Estimation In The Wild. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 7 +[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8 +[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine + +tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3 +[27] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 3 +[28] Florian Schroff, Dmitry Kalenichenko, and James Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6 +[29] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2, 3 +[30] Kaede Shiohara and Toshihiko Yamasaki. Face2diffusion for fast and editable face personalization. arXiv preprint arXiv:2403.05094, 2024. 3 +[31] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. FreeU: Free Lunch in Diffusion U-Net. In CVPR, 2024. 5 +[32] Yu-Chuan Su, Kelvin CK Chan, Yandong Li, Yang Zhao, Han Zhang, Boqing Gong, Huisheng Wang, and Xuhui Jia. Identity encoder for personalized diffusion. arXiv preprint arXiv:2304.07429, 2023. 3 +[33] Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. In SIGGRAPH Asia 2023 Conference Papers, pages 1-10, 2023. 3 +[34] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. InstantID: Zero-shot Identity-Preserving Generation in Seconds. arXiv preprint arXiv:2401.07519, 2024. 3, 8 +[35] Guangxuan Xiao, Tianwei Yin, William T Freeman, Fredo Durand, and Song Han. FastComposer: Tuning-Free Multi-Subject Image Generation with Localized Attention. In International Journal of Computer Vision, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[36] Yuxuan Yan, Chi Zhang, Rui Wang, Yichao Zhou, Gege Zhang, Pei Cheng, Gang Yu, and Bin Fu. Facestudio: Put your face everywhere in seconds. arXiv preprint arXiv:2312.02663, 2023. 3 +[37] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IpAdapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 6, 7 +[38] Hangjie Yuan, Jianwen Jiang, Samuel Albanie, Tao Feng, Ziyuan Huang, Dong Ni, and Mingqian Tang. RLIP: Relational Language-Image Pre-training for Human-Object Interaction Detection. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3 +[39] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Samuel Albanie, Yining Pan, Tao Feng, Jianwen Jiang, Dong Ni, Yingya Zhang, and Deli Zhao. RLIPv2: Fast Scaling of Relational Language-Image Pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. + +[40] Frederic Z. Zhang, Dylan Campbell, and Stephen Gould. Efficient Two-Stage Detection of Human-Object Interactions with a Novel Unary-Pairwise Transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20104-20112, 2022. 3, 6 +[41] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2016. 6 +[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In IEEE International Conference on Computer Vision (ICCV), 2023. 3 +[43] Xulu Zhang, Xiao-Yong Wei, Wengyu Zhang, Jinlin Wu, Zhaoxiang Zhang, Zhen Lei, and Qing Li. A Survey on Personalized Content Synthesis with Diffusion Models. arXiv preprint arXiv:2405.05538, 2024. https://arxiv.org/abs/2405.05538.3 +[44] Yuxin Zhang, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. MotionCrafter: One-Shot Motion Customization of Diffusion Models. arXiv preprint arXiv:2312.05288, 2023. 3 \ No newline at end of file diff --git a/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/images.zip b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cee99864c20532c783dcb5907de38d20751f6012 --- /dev/null +++ b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:145f34551cf78449d01797ebaf2479f1d5daeeef29cc9756407c05e597ce3102 +size 847638 diff --git a/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/layout.json b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b695a8594059daea7b235527403c1b743d9a745c --- /dev/null +++ b/2025/PersonaHOI_ Effortlessly Improving Face Personalization in Human-Object Interaction Generation/layout.json @@ -0,0 +1,9534 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 484, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 484, + 140 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 484, + 140 + ], + "type": "text", + "content": "PersonaHOI: Effortlessly Improving Face Personalization in Human-Object Interaction Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 162, + 159, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 162, + 159, + 175 + ], + "spans": [ + { + "bbox": [ + 97, + 162, + 159, + 175 + ], + "type": "text", + "content": "Xinting Hu*" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 178, + 162, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 178, + 162, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 178, + 162, + 190 + ], + "type": "text", + "content": "xhu@mpi-inf.mpg.de" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 200, + 162, + 272, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 162, + 272, + 175 + ], + "spans": [ + { + "bbox": [ + 200, + 162, + 272, + 175 + ], + "type": "text", + "content": "Haoran Wang*" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 173, + 178, + 288, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 178, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 173, + 178, + 288, + 190 + ], + "type": "text", + "content": "hawang@mpi-inf.mpg.de" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 162, + 397, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 397, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 397, + 175 + ], + "type": "text", + "content": "Jan Eric Lenssen" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 178, + 427, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 178, + 427, + 190 + ], + "spans": [ + { + "bbox": [ + 302, + 178, + 427, + 190 + ], + "type": "text", + "content": "jlenssen@mpi-inf.mpg.de" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 457, + 162, + 525, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 457, + 162, + 525, + 175 + ], + "spans": [ + { + "bbox": [ + 457, + 162, + 525, + 175 + ], + "type": "text", + "content": "Bernt Schiele" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 440, + 178, + 559, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 440, + 178, + 559, + 190 + ], + "spans": [ + { + "bbox": [ + 440, + 178, + 559, + 190 + ], + "type": "text", + "content": "schiele@mpi-inf.mpg.de" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 122, + 195, + 499, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 195, + 499, + 209 + ], + "spans": [ + { + "bbox": [ + 122, + 195, + 499, + 209 + ], + "type": "text", + "content": "Max Planck Institute for Informatics, Saarland Informatics Campus, Germany" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 153, + 225, + 247, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 225, + 247, + 239 + ], + "spans": [ + { + "bbox": [ + 153, + 225, + 247, + 239 + ], + "type": "text", + "content": "Context + HOI" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 99, + 239, + 279, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 239, + 279, + 249 + ], + "spans": [ + { + "bbox": [ + 99, + 239, + 279, + 249 + ], + "type": "text", + "content": "A man riding a horse in front of a medieval castle" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 99, + 251, + 156, + 308 + ], + "blocks": [ + { + "bbox": [ + 99, + 251, + 156, + 308 + ], + "lines": [ + { + "bbox": [ + 99, + 251, + 156, + 308 + ], + "spans": [ + { + "bbox": [ + 99, + 251, + 156, + 308 + ], + "type": "image", + "image_path": "b157e54b54f7e59e72c4458fd80cbff15435b7be97458c9d9108f2e545625445.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 173, + 250, + 231, + 308 + ], + "blocks": [ + { + "bbox": [ + 173, + 250, + 231, + 308 + ], + "lines": [ + { + "bbox": [ + 173, + 250, + 231, + 308 + ], + "spans": [ + { + "bbox": [ + 173, + 250, + 231, + 308 + ], + "type": "image", + "image_path": "eee15c6a3d7f2f66d0fe64bb09f85357a7a22b1765f09202b4ca9899a2e8dc7c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 308, + 223, + 316 + ], + "lines": [ + { + "bbox": [ + 183, + 308, + 223, + 316 + ], + "spans": [ + { + "bbox": [ + 183, + 308, + 223, + 316 + ], + "type": "text", + "content": "FastComposer" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 237, + 250, + 297, + 308 + ], + "blocks": [ + { + "bbox": [ + 237, + 250, + 297, + 308 + ], + "lines": [ + { + "bbox": [ + 237, + 250, + 297, + 308 + ], + "spans": [ + { + "bbox": [ + 237, + 250, + 297, + 308 + ], + "type": "image", + "image_path": "e7b41c0e0d4124d394e79e4ca3536a1d6653346a3418912f9d207e1de6ed09cc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 308, + 277, + 315 + ], + "lines": [ + { + "bbox": [ + 257, + 308, + 277, + 315 + ], + "spans": [ + { + "bbox": [ + 257, + 308, + 277, + 315 + ], + "type": "text", + "content": "+Ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 374, + 223, + 456, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 223, + 456, + 240 + ], + "spans": [ + { + "bbox": [ + 374, + 223, + 456, + 240 + ], + "type": "text", + "content": "Style + HOI" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 240, + 514, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 240, + 514, + 250 + ], + "spans": [ + { + "bbox": [ + 321, + 240, + 514, + 250 + ], + "type": "text", + "content": "A woman holding a cat in the style of Vincent Van Gogh" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 326, + 251, + 372, + 309 + ], + "blocks": [ + { + "bbox": [ + 326, + 251, + 372, + 309 + ], + "lines": [ + { + "bbox": [ + 326, + 251, + 372, + 309 + ], + "spans": [ + { + "bbox": [ + 326, + 251, + 372, + 309 + ], + "type": "image", + "image_path": "7435031c37e19ec29770cf86db560bcce83ba502dd54db71fd0c9938e201e473.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 384, + 251, + 443, + 308 + ], + "blocks": [ + { + "bbox": [ + 384, + 251, + 443, + 308 + ], + "lines": [ + { + "bbox": [ + 384, + 251, + 443, + 308 + ], + "spans": [ + { + "bbox": [ + 384, + 251, + 443, + 308 + ], + "type": "image", + "image_path": "52fd20b7f4bdff5b304a657d125772808fe19d54d661aa9f3d435952bf704d63.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 309, + 435, + 316 + ], + "lines": [ + { + "bbox": [ + 395, + 309, + 435, + 316 + ], + "spans": [ + { + "bbox": [ + 395, + 309, + 435, + 316 + ], + "type": "text", + "content": "FastComposer" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 449, + 251, + 508, + 308 + ], + "blocks": [ + { + "bbox": [ + 449, + 251, + 508, + 308 + ], + "lines": [ + { + "bbox": [ + 449, + 251, + 508, + 308 + ], + "spans": [ + { + "bbox": [ + 449, + 251, + 508, + 308 + ], + "type": "image", + "image_path": "320948f929912f5425969af98881dfbdb9780f8f9d4101acbe1558403e04cd82.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 309, + 489, + 316 + ], + "lines": [ + { + "bbox": [ + 469, + 309, + 489, + 316 + ], + "spans": [ + { + "bbox": [ + 469, + 309, + 489, + 316 + ], + "type": "text", + "content": "+Ours" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 106, + 330, + 141, + 369 + ], + "blocks": [ + { + "bbox": [ + 106, + 330, + 141, + 369 + ], + "lines": [ + { + "bbox": [ + 106, + 330, + 141, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 141, + 369 + ], + "type": "image", + "image_path": "247ba23b3aa2cf8aedd490ce8aa68908e1ba25b335540164fe2bcc36869980d7.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 149, + 359, + 165, + 376 + ], + "blocks": [ + { + "bbox": [ + 113, + 370, + 151, + 405 + ], + "lines": [ + { + "bbox": [ + 113, + 370, + 151, + 405 + ], + "spans": [ + { + "bbox": [ + 113, + 370, + 151, + 405 + ], + "type": "text", + "content": "Various man-Object action (HOI) descriptions" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 359, + 165, + 376 + ], + "lines": [ + { + "bbox": [ + 149, + 359, + 165, + 376 + ], + "spans": [ + { + "bbox": [ + 149, + 359, + 165, + 376 + ], + "type": "image", + "image_path": "ab7c40c3cf4033effea10b05a9a3a6eefde85db862f49281f0c995aef0658d10.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 175, + 327, + 212, + 367 + ], + "blocks": [ + { + "bbox": [ + 175, + 327, + 212, + 367 + ], + "lines": [ + { + "bbox": [ + 175, + 327, + 212, + 367 + ], + "spans": [ + { + "bbox": [ + 175, + 327, + 212, + 367 + ], + "type": "image", + "image_path": "95dd086b65595a579d3d425c6609f565a3eda7f96aba4b06315864382c6834dd.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 216, + 327, + 254, + 366 + ], + "blocks": [ + { + "bbox": [ + 216, + 327, + 254, + 366 + ], + "lines": [ + { + "bbox": [ + 216, + 327, + 254, + 366 + ], + "spans": [ + { + "bbox": [ + 216, + 327, + 254, + 366 + ], + "type": "image", + "image_path": "a1cfa0df5c025adff299d9824ef2eb90e13a8db98b4b547113d3d8783195ebc8.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 258, + 328, + 296, + 366 + ], + "blocks": [ + { + "bbox": [ + 258, + 328, + 296, + 366 + ], + "lines": [ + { + "bbox": [ + 258, + 328, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 258, + 328, + 296, + 366 + ], + "type": "image", + "image_path": "71f314f3842b0aa7dc66879f5b8a05ec49b4b5049469ab393b694a608d44ac77.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 174, + 367, + 212, + 406 + ], + "blocks": [ + { + "bbox": [ + 174, + 367, + 212, + 406 + ], + "lines": [ + { + "bbox": [ + 174, + 367, + 212, + 406 + ], + "spans": [ + { + "bbox": [ + 174, + 367, + 212, + 406 + ], + "type": "image", + "image_path": "8ea63ada345a13ba311df7292efa55d208d946c220d7f6951aa595a6e0a2438c.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 216, + 367, + 254, + 406 + ], + "blocks": [ + { + "bbox": [ + 216, + 367, + 254, + 406 + ], + "lines": [ + { + "bbox": [ + 216, + 367, + 254, + 406 + ], + "spans": [ + { + "bbox": [ + 216, + 367, + 254, + 406 + ], + "type": "image", + "image_path": "a9f97ef19f0729a0e4aac6e809f40bd6a55abf839ae1a6219b9e08b6f8a40aae.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 258, + 367, + 296, + 406 + ], + "blocks": [ + { + "bbox": [ + 258, + 367, + 296, + 406 + ], + "lines": [ + { + "bbox": [ + 258, + 367, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 258, + 367, + 296, + 406 + ], + "type": "image", + "image_path": "3dedd8517a0228a615656366ae664bc76fd2c3aa79f5b8aed21058ddc60439f8.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 322, + 331, + 430, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 331, + 430, + 353 + ], + "spans": [ + { + "bbox": [ + 322, + 331, + 430, + 353 + ], + "type": "text", + "content": "Personalized" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 330, + 353, + 420, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 353, + 420, + 368 + ], + "spans": [ + { + "bbox": [ + 330, + 353, + 420, + 368 + ], + "type": "text", + "content": "Face with" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 331, + 371, + 419, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 371, + 419, + 386 + ], + "spans": [ + { + "bbox": [ + 331, + 371, + 419, + 386 + ], + "type": "text", + "content": "Human-Object" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 343, + 386, + 424, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 386, + 424, + 399 + ], + "spans": [ + { + "bbox": [ + 343, + 386, + 424, + 399 + ], + "type": "text", + "content": "Interaction" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 452, + 338, + 496, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 452, + 338, + 496, + 351 + ], + "spans": [ + { + "bbox": [ + 452, + 338, + 496, + 351 + ], + "type": "text", + "content": "Training" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 468, + 353, + 479, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 353, + 479, + 365 + ], + "spans": [ + { + "bbox": [ + 468, + 353, + 479, + 365 + ], + "type": "text", + "content": "&" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 455, + 367, + 493, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 455, + 367, + 493, + 380 + ], + "spans": [ + { + "bbox": [ + 455, + 367, + 493, + 380 + ], + "type": "text", + "content": "Tuning" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 459, + 381, + 488, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 459, + 381, + 488, + 394 + ], + "spans": [ + { + "bbox": [ + 459, + 381, + 488, + 394 + ], + "type": "text", + "content": "Free!" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 146, + 418, + 252, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 418, + 252, + 434 + ], + "spans": [ + { + "bbox": [ + 146, + 418, + 252, + 434 + ], + "type": "text", + "content": "Accessory + HOI" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 96, + 435, + 302, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 435, + 302, + 443 + ], + "spans": [ + { + "bbox": [ + 96, + 435, + 302, + 443 + ], + "type": "text", + "content": "A woman wearing a Santa hat and taking selfie with phone" + } + ] + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 97, + 445, + 157, + 504 + ], + "blocks": [ + { + "bbox": [ + 97, + 445, + 157, + 504 + ], + "lines": [ + { + "bbox": [ + 97, + 445, + 157, + 504 + ], + "spans": [ + { + "bbox": [ + 97, + 445, + 157, + 504 + ], + "type": "image", + "image_path": "cc0f5f8077b75ab9c225f3f0cb46c7ed55d42813cb81362f2beca721fbafc31c.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 524, + 555, + 568 + ], + "lines": [ + { + "bbox": [ + 54, + 524, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 524, + 555, + 568 + ], + "type": "text", + "content": "Figure 1. Examples of Personalized Face with Human-Object Interaction (HOI) Generation. We present PersonaHOI, a training- and tuning-free framework built on existing diffusion models. Using a single reference image and diverse HOI prompts, PersonaHOI generates identity-consistent human-object interactions compared to FastComposer [35]. PersonaHOI can further seamlessly integrate varied contexts, styles, accessories, and multi-person scenarios, ensuring scalability and practicality for real-world applications." + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 179, + 445, + 237, + 503 + ], + "blocks": [ + { + "bbox": [ + 179, + 445, + 237, + 503 + ], + "lines": [ + { + "bbox": [ + 179, + 445, + 237, + 503 + ], + "spans": [ + { + "bbox": [ + 179, + 445, + 237, + 503 + ], + "type": "image", + "image_path": "88782716e9e95372ba6ba9459feb421f25b10a7b26fdf4d4fc7dd3761727640f.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 504, + 231, + 511 + ], + "lines": [ + { + "bbox": [ + 190, + 504, + 231, + 511 + ], + "spans": [ + { + "bbox": [ + 190, + 504, + 231, + 511 + ], + "type": "text", + "content": "FastComposer" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 243, + 445, + 302, + 503 + ], + "blocks": [ + { + "bbox": [ + 243, + 445, + 302, + 503 + ], + "lines": [ + { + "bbox": [ + 243, + 445, + 302, + 503 + ], + "spans": [ + { + "bbox": [ + 243, + 445, + 302, + 503 + ], + "type": "image", + "image_path": "4185e7907c4f3d2a7e93cdbb4b654b9ddd9d1743380b26001350fefc76257d0b.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 505, + 284, + 511 + ], + "lines": [ + { + "bbox": [ + 264, + 505, + 284, + 511 + ], + "spans": [ + { + "bbox": [ + 264, + 505, + 284, + 511 + ], + "type": "text", + "content": "+Ours" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "bbox": [ + 365, + 418, + 473, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 418, + 473, + 430 + ], + "spans": [ + { + "bbox": [ + 365, + 418, + 473, + 430 + ], + "type": "text", + "content": "Mult-Person + HOI" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 329, + 431, + 490, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 431, + 490, + 441 + ], + "spans": [ + { + "bbox": [ + 329, + 431, + 490, + 441 + ], + "type": "text", + "content": "A man and a man are playing guitar in a bar" + } + ] + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 327, + 441, + 359, + 502 + ], + "blocks": [ + { + "bbox": [ + 327, + 441, + 359, + 502 + ], + "lines": [ + { + "bbox": [ + 327, + 441, + 359, + 502 + ], + "spans": [ + { + "bbox": [ + 327, + 441, + 359, + 502 + ], + "type": "image", + "image_path": "db2c67893599da98510a6773350c766abb93adb5aead9340d04282c538714726.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 376, + 441, + 441, + 503 + ], + "blocks": [ + { + "bbox": [ + 376, + 441, + 441, + 503 + ], + "lines": [ + { + "bbox": [ + 376, + 441, + 441, + 503 + ], + "spans": [ + { + "bbox": [ + 376, + 441, + 441, + 503 + ], + "type": "image", + "image_path": "11fdfc71f8f8d066b87ca4a669ea495afb8320c9778279662b2a875a9f49a960.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 504, + 432, + 511 + ], + "lines": [ + { + "bbox": [ + 392, + 504, + 432, + 511 + ], + "spans": [ + { + "bbox": [ + 392, + 504, + 432, + 511 + ], + "type": "text", + "content": "FastComposer" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 447, + 441, + 511, + 502 + ], + "blocks": [ + { + "bbox": [ + 447, + 441, + 511, + 502 + ], + "lines": [ + { + "bbox": [ + 447, + 441, + 511, + 502 + ], + "spans": [ + { + "bbox": [ + 447, + 441, + 511, + 502 + ], + "type": "image", + "image_path": "a667851efcf6131abd92aa22122c5ed3199961961de8a62edb9ab49dff6607a7.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 504, + 488, + 511 + ], + "lines": [ + { + "bbox": [ + 468, + 504, + 488, + 511 + ], + "spans": [ + { + "bbox": [ + 468, + 504, + 488, + 511 + ], + "type": "text", + "content": "+Ours" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 57 + }, + { + "bbox": [ + 152, + 577, + 199, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 577, + 199, + 590 + ], + "spans": [ + { + "bbox": [ + 152, + 577, + 199, + 590 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 53, + 604, + 297, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 604, + 297, + 688 + ], + "spans": [ + { + "bbox": [ + 53, + 604, + 297, + 688 + ], + "type": "text", + "content": "We introduce PersonaHOI, a training- and tuning-free framework that fuses a general StableDiffusion model with a personalized face diffusion (PFD) model to generate identity-consistent human-object interaction (HOI) images. While existing PFD models have advanced significantly, they often overemphasize facial features at the expense of full-body coherence, PersonaHOI introduces an additional" + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 313, + 578, + 556, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 556, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 556, + 698 + ], + "type": "text", + "content": "StableDiffusion (SD) branch guided by HOI-oriented text inputs. By incorporating cross-attention constraints in the PFD branch and spatial merging at both latent and residual levels, PersonaHOI preserves personalized facial details while ensuring interactive non-facial regions. Experiments, validated by a novel interaction alignment metric, demonstrate the superior realism and scalability of PersonaHOI, establishing a new standard for practical personalized face with HOI generation. Code is available at https://github.com/JoyHuYY1412/PersonaHOI." + } + ] + } + ], + "index": 62 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 703, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 136, + 712 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23775" + } + ] + } + ], + "index": 64 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 73, + 549, + 224 + ], + "blocks": [ + { + "bbox": [ + 63, + 73, + 549, + 224 + ], + "lines": [ + { + "bbox": [ + 63, + 73, + 549, + 224 + ], + "spans": [ + { + "bbox": [ + 63, + 73, + 549, + 224 + ], + "type": "image", + "image_path": "dd865b7a270e59debd3e386ac1a78d3ee6ace7beac4b9887245343fa0cc04954.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 235, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 555, + 281 + ], + "type": "text", + "content": "Figure 2. (a) The spatial layout of StableDiffusion guides PersonaHOI to generate personalized content with coherent human-object interactions (HOI). (b) Analysis of identity injection timing in PFD models. We use FastComposer [35] for diffusion model generation. Injecting face representation at the start of image generation preserves facial details but lacks coherent HOI, while delayed injection continuously deviates from the original identity, resulting in random human features and meaningless human-object interactions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 299, + 137, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 137, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 137, + 312 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 418 + ], + "type": "text", + "content": "Personalized face generation has seen increasing public interest and demand in user-specified digital content creation. Current learning-based personalized face diffusion (PFD) models [18, 21, 29, 35], trained on large-scale face-centric datasets, can incorporate a single user-provided reference facial image into a text-to-image generation process, enabling the rapid creation of images that depict specific persons in diverse scenes, outfits, and styles within seconds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 422, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 531 + ], + "type": "text", + "content": "Although these methods perform well in simpler scenarios, they often struggle with generating full-body depictions involving detailed human-object interactions (HOI). As illustrated in Figure 1, the generated images of FastComposer [35] show missing objects or body parts, resulting in portrait-focused outputs without HOI information. This limitation compromises the overall realism and practical utility of the generated content, especially in immersive or interactive applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "To diagnose the limitations of PFD models, we examine the results of PFD models by varying the timing of identity injection during the generation process. As seen in the first and last columns of Figure 2(b), early injection effectively preserves facial details from the reference image, but relying solely on text input (last column) results in incoherent HOI outputs. This analysis highlights that the core challenge lies not in maintaining identity fidelity but in generating coherent body movements and interactions driven by text prompts. Notably, pre-trained models used for PFD, such as StableDiffusion [25], can generate satisfactory HOI images due to their extensive and diverse training data. This observation implies that the fine-tuning process on face-centric datasets of PFD models diminishes their ability to follow complex HOI text prompts. Therefore, to advance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 300, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 335 + ], + "type": "text", + "content": "current PFD methods in HOI tasks, reintroducing the capability to leverage text prompts for natural full-body interactions is essential." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 338, + 556, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 556, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 556, + 434 + ], + "type": "text", + "content": "Motivated by this need, we propose a straightforward yet powerful approach: leveraging the generative capabilities of pre-trained diffusion models, such as StableDiffusion (SD), to augment existing PFD frameworks. Our method, PersonaHOI, integrates the strengths of SD without requiring additional training or fine-tuning, thereby restoring the ability to generate realistic, text-driven full-body HOI while preserving identity-specific details from the reference image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 436, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 640 + ], + "type": "text", + "content": "Central to this integration, we incorporate an additional SD branch that aligns identity-specific details from PFD models with the HOI layouts generated by SD. A head segmentation mask derived from the SD output guides the merge of non-facial components from the SD branch with the facial details from the PFD branch. Specifically, in the PFD model branch, we introduce a Cross-Attention Constraint to prevent the overemphasis of identity features across the entire image, ensuring these details are confined to the facial region. To integrate HOI representations from SD, we implement Latent Merge and Residual Merge, merging identity and contextual features at both the latent representation level and via skip connections within the U-Net architecture. This multi-level merging strategy ensures that the generated images maintain realistic human-object interactions while preserving personalized facial details, resulting in cohesive, interaction-rich outputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": "To evaluate the effectiveness of our method in HOI scenarios, we propose a novel \"interaction alignment\" metric. These metrics leverage HOI detectors to measure alignment between generated images and text prompts, objectively assessing interaction realism. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23776" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 251 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 132 + ], + "type": "text", + "content": "- We present PersonaHOI, a dual-path architecture that integrates general and personalized diffusion models. Our framework generates realistic human-object interactions with specific identities, without requiring additional training or test-time tuning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 133, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 133, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 54, + 133, + 294, + 192 + ], + "type": "text", + "content": "- Our method introduces key components, including crossattention constraints and a combination of latent and residual merge strategies. These innovations enable effective feature integration, ensuring that generated images maintain both interaction coherence and identity fidelity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 192, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 192, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 192, + 294, + 251 + ], + "type": "text", + "content": "- We design a novel evaluation metric for human-object interaction. Our experiments show significant improvements over state-of-the-art face personalization techniques in terms of interaction realism across models, emphasizing the scalability and robustness of our approach." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 262, + 147, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 147, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 147, + 275 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 581 + ], + "type": "text", + "content": "Personalized Face Generation Diffusion Models. Personalized face generation aims to create identity-consistent images of individuals from limited reference data [32, 36, 43]. Given pretrained general image generation models [25], traditional optimization-based approaches [5, 26, 27] refine model parameters during inference for each individual but are computationally expensive, making them unsuitable for real-time applications. In contrast, learning-based approaches leverage human-centric datasets [1, 15, 19, 20] to train robust models capable of varying identity inputs without requiring test-time fine-tuning. Facial representations, extracted by models like CLIP [23] or other face recognition models [3, 16, 33], are often fused with [18, 21, 35] or replace [30] textual embeddings of specific words, such as \"person\", to provide identity conditions for diffusion models and enhance contextual relevance [3, 29]. In IP-Adapter [37] and other works [7, 9, 34], facial features are processed separately through decoupled cross-attention modules. Incorporating human masks as prior, methods can clean the background during data processing; alternatively, masks can be leveraged to construct [2, 18, 21, 35] and adjust [12] loss functions during training. In our work, we extend face personalization to complex human-object interactions (HOI), achieving a higher level of personalization and adaptability for diverse generative tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Human-Object Interaction in Diffusion Models. HOI detection aims to locate human-object pairs and categorize their interactions as triplets, such as (person, playing, football) [4, 38-40], while HOI image synthesis remains relatively under-explored. InteractGAN [6] generates HOI images using human pose templates and reference images, but its reliance on pose-template pools and object references limits flexibility. Recent diffusion-based approaches tackle HOI generation by incorporating layout controls such as bounding boxes [10, 14, 17] or human poses [17, 42], or by leveraging in-context samples with similar interac" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "tions [11, 44]. Additionally, ReCorD [14] integrates Latent Diffusion Models with Visual Language Models, introducing modules for reasoning and correcting interactions to enhance cross-modal alignment. In this work, we rely on the standard text-to-image diffusion model (i.e., StableDiffusion [25]) for HOI generation, and fuse the generation process with personalized face diffusion models to generation identity-preserving HOI images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 178, + 391, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 391, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 391, + 192 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 198, + 553, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 294 + ], + "type": "text", + "content": "Understanding the fundamentals of StableDiffusion (SD) is essential to leverage its generative capabilities within our PFG framework. SD is a latent diffusion model designed to generate high-quality images based on text or image prompts. It operates in a latent space, where an input image is encoded in a latent representation " + }, + { + "bbox": [ + 313, + 198, + 553, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 313, + 198, + 553, + 294 + ], + "type": "text", + "content": " using a pretrained variational autoencoder (VAE) encoder, reducing the dimensionality of the image for efficient processing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "text", + "content": "During the generation phase, the SD model begins with an initial noisy latent representation " + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T" + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "text", + "content": " and progressively denoises it over " + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "text", + "content": " timesteps to reconstruct " + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "text", + "content": ". This denoised latent is then decoded by the VAE decoder to produce the final image. At each timestep " + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "text", + "content": ", the U-Net architecture " + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "text", + "content": " predicts noise " + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "inline_equation", + "content": "\\epsilon (\\mathbf{z}_t,t,\\mathbf{C})" + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 294, + 553, + 390 + ], + "type": "text", + "content": " represents the conditioning inputs (text or image embeddings). The iterative denoising process follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 364, + 399, + 553, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 399, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 364, + 399, + 553, + 412 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t - 1} = \\operatorname {D e n o i s e} \\left(\\mathbf {z} _ {t}, \\epsilon \\left(\\mathbf {z} _ {t}, t, \\mathbf {C}\\right); \\theta\\right) \\tag {1}", + "image_path": "04996605cb8ed61d0ef9906aeda9480c25ec69222de5dd927628796d67b73295.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 422, + 553, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 553, + 505 + ], + "type": "text", + "content": "The U-Net architecture comprises an encoder-decoder structure with skip connections that transfer features between the encoder and the decoder. Within this architecture, SD employs a cross-attention mechanism to incorporate conditioning inputs effectively during denoising. This mechanism projects latent representations and conditioning inputs into query, key, and value spaces:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 512, + 553, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 512, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 358, + 512, + 553, + 552 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathbf {Q} = \\mathbf {W} _ {q} \\mathbf {z}, \\mathbf {K} = \\mathbf {W} _ {k} \\mathbf {C}, \\mathbf {V} = \\mathbf {W} _ {v} \\mathbf {C}; \\\\ \\mathbf {A} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {\\top}}{\\sqrt {d}}\\right), \\mathbf {z} _ {\\text {a t t n}} = \\mathbf {A V}. \\end{array} \\right. \\tag {2}", + "image_path": "e473f2ca72746486c545c839a8c4e6d06b0f3e200066e17b51a93acc3b73fb2a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 553, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 618 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 558, + 553, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 558, + 553, + 618 + ], + "type": "text", + "content": " is the attention map that ensures that relevant conditioning tokens guide denoising at each generation step. This cross-attention mechanism is critical for aligning the generated image with the input prompts, enabling consistency and fidelity in the final output." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 628, + 370, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 370, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 370, + 640 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 647, + 434, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 434, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 434, + 659 + ], + "type": "text", + "content": "4.1. Overall Architecture" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "As illustrated in Figure 3, PersonaHOI augments the PFD model by integrating an additional StableDiffusion (SD) branch, enhancing its capacity to generate personalized images that capture complex human-object interactions (HOI)" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23777" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 68, + 519, + 304 + ], + "blocks": [ + { + "bbox": [ + 97, + 68, + 519, + 304 + ], + "lines": [ + { + "bbox": [ + 97, + 68, + 519, + 304 + ], + "spans": [ + { + "bbox": [ + 97, + 68, + 519, + 304 + ], + "type": "image", + "image_path": "0bfaab5951f489661d45979cb536bf2ced2fa25a52626df45a24eec4946d2160.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 313, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 54, + 313, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 54, + 313, + 555, + 380 + ], + "type": "text", + "content": "Figure 3. Overview of Our Proposed Framework, PersonaHOI. The architecture integrates a personalized face diffusion (PFD) model with an additional StableDiffusion (SD) branch. First, SD generates an image " + }, + { + "bbox": [ + 54, + 313, + 555, + 380 + ], + "type": "inline_equation", + "content": "(I_{SD})" + }, + { + "bbox": [ + 54, + 313, + 555, + 380 + ], + "type": "text", + "content": " from a text prompt and noisy latent representation " + }, + { + "bbox": [ + 54, + 313, + 555, + 380 + ], + "type": "inline_equation", + "content": "(z_{T})" + }, + { + "bbox": [ + 54, + 313, + 555, + 380 + ], + "type": "text", + "content": ", which is decoded and segmented to produce a head mask. Next, SD and PFD run in parallel from the same " + }, + { + "bbox": [ + 54, + 313, + 555, + 380 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 54, + 313, + 555, + 380 + ], + "type": "text", + "content": ". At every timestep " + }, + { + "bbox": [ + 54, + 313, + 555, + 380 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 313, + 555, + 380 + ], + "type": "text", + "content": ", the head mask guides the Cross-Attention Constraint in PFD and merging modules (Latent Merge and Residual Merge) to merge interaction-relevant features from SD with identity-specific details from PFD. Iteratively, this process introduces HOI context to personalized face generation in a training&tuning-free manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 395, + 294, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 395, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 395, + 294, + 455 + ], + "type": "text", + "content": "in a training-free manner. Given an input comprising a user-provided reference image and a text prompt specifying the interaction (e.g., \"a person kicking football\"), our framework produces a cohesive output image that retains the subject's identity while accurately depicting the specified HOI." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 456, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 456, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 54, + 456, + 295, + 612 + ], + "type": "text", + "content": "To effectively merge identity from PFD and interaction features from SD, we introduce three core strategies: Cross-Attention Constraint, which regulates attention to reference image features within the PFD model (detailed in Sec. 4.2); Latent Merge, which combines features in the latent space at each generation timestep (Sec. 4.3); and Residual Merge, which integrates identity details through skip connections in the U-Net architecture (Sec. 4.4). Together, these strategies enable PersonaHOI to align generated content with interaction-driven text prompts while maintaining identity fidelity. Furthermore, this design allows PersonaHOI adaptable to various PFD models without the need for model retraining or test-time tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 623, + 239, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 239, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 239, + 635 + ], + "type": "text", + "content": "4.2. Cross-Attention Constraint (CAC)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "In the cross-attention layers of the PFD model, conditioning inputs " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " include both text and image embeddings to guide the generation process, where " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is the number of tokens, and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is the embedding dimension. Following established PFD model approaches [18, 21, 35], the " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "n_{img}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "-th token in " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " incorporates the image embedding, encapsulat" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 395, + 553, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 553, + 466 + ], + "type": "text", + "content": "ing identity-specific details essential for facial preservation, while the remaining " + }, + { + "bbox": [ + 313, + 395, + 553, + 466 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 313, + 395, + 553, + 466 + ], + "type": "text", + "content": " tokens correspond to text embeddings. In standard PFD models, attention to the " + }, + { + "bbox": [ + 313, + 395, + 553, + 466 + ], + "type": "inline_equation", + "content": "n_{img}" + }, + { + "bbox": [ + 313, + 395, + 553, + 466 + ], + "type": "text", + "content": "-th token often spreads across the image, overemphasizing facial features thus decreasing spatial capacity for representing body movements and object interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "text", + "content": "To address this, we introduce Cross-Attention Constraint (CAC) within the cross-attention layers of the PFD model. CAC restricts identity features to specific facial regions, ensuring sufficient spatial capacity for body and object interactions. Different from the attention localization loss used in training [21, 35], our CAC utilizes fine-grained head mask derived from the SD-generated image " + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "inline_equation", + "content": "I_{SD}" + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "text", + "content": " in image generation process, ensuring the attention constraint in PFD is consistent with the HOI layout generated by SD. The head mask is segmented by an off-the-shelf head segmentor [24], then resized to match the spatial dimensions " + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "inline_equation", + "content": "(H \\times W)" + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "text", + "content": " of the latent representation " + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{\\mathrm{head}} \\in [0,1]^{H \\times W}" + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "text", + "content": " (1 for the head region and 0 for other regions). We apply " + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{\\mathrm{head}}" + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "text", + "content": " to the attention map " + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{(H \\times W) \\times N}" + }, + { + "bbox": [ + 313, + 467, + 555, + 646 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 325, + 647, + 553, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 325, + 647, + 553, + 680 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i \\in [ 0, 1, \\dots , N - 1 ]} ^ {C A C} = \\left\\{ \\begin{array}{l l} \\text {f l a t t e n} (\\mathbf {M} ^ {\\text {h e a d}}), & \\text {i f} i = i m g, \\\\ 1, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}", + "image_path": "90c7b0f9ac09335f3a6d8f83cc4b302e979b1b8e83ba0a9139b700865bd1306a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 681, + 471, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 681, + 471, + 696 + ], + "spans": [ + { + "bbox": [ + 361, + 681, + 471, + 696 + ], + "type": "interline_equation", + "content": "\\mathbf {A} ^ {C A C} = \\mathbf {A} \\odot (\\mathbf {M} ^ {C A C}) ^ {\\intercal}.", + "image_path": "b9d2b20c805ab8d844160f7a51896f226e98bf4d6258e54dea7b426b0a527fd1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{CAC}" + }, + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "type": "text", + "content": " modifies the attention weights for the " + }, + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "type": "inline_equation", + "content": "n_{img}" + }, + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "type": "text", + "content": " token" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23778" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 72, + 289, + 194 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 289, + 194 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 289, + 194 + ], + "type": "image", + "image_path": "91e0629e18e3c345132d2a0ab10c395fffbb083ed48a7dc74dd08cdbb647fb69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 266 + ], + "type": "text", + "content": "Figure 4. Illustration of Residual Merge. In each residual layer, Residual Merge operates within the U-Net skip connections, utilizing a head mask to guide the integration of high-frequency identity details from PFD residuals and low-frequency interaction layouts from SD residuals. The merged residuals are then concatenated to the corresponding bottleneck features from PFD." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 287, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 371 + ], + "type": "text", + "content": "by setting the weights to 0 in regions outside the head, preventing the influence of human identity on non-facial areas. Since the head mask is derived from SD-generated outputs, CAC maintains facial layout coherence with SD. When further combined with Latent and Residual Merge, CAC helps blend facial features from PFD and non-facial interaction features from SD without mutual interference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 380, + 170, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 170, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 170, + 393 + ], + "type": "text", + "content": "4.3. Latent Merge (LM)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 506 + ], + "type": "text", + "content": "In this section, we introduce Latent Merge, which directly merges the latent representation within U-Net from PFD and SD models. After applying Cross-Attention Constraint (CAC) in Sec. 4.2, the facial region generated by the PFD model becomes spatially aligned with the corresponding region in the broader HOI layout produced by the SD model. This alignment facilitates spatial merge, blending facial identity details from PFD with interaction contexts from SD in latent space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "text", + "content": "At each diffusion timestep " + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "text", + "content": " of PFD and SD, we implement the latent space merging with " + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{head}" + }, + { + "bbox": [ + 55, + 506, + 295, + 532 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 540, + 295, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 540, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 85, + 540, + 295, + 556 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t} = \\mathbf {M} ^ {h e a d} \\odot \\mathbf {z} _ {t} ^ {\\mathrm {P F D}} + \\left(1 - \\mathbf {M} ^ {h e a d}\\right) \\odot \\mathbf {z} _ {t} ^ {\\mathrm {S D}}. \\qquad (4)", + "image_path": "2855fa54dbfb08d8decc6b9429381673cae5301a94f055039d53a4bd2cf20172.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{PFD}}" + }, + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{\\mathrm{SD}}" + }, + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "text", + "content": " represents the latent features from PFD and SD at timestep " + }, + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "text", + "content": ", respectively. The merged latent " + }, + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "text", + "content": " then serves as input for both SD and PFD models in the next denoising timestep " + }, + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 55, + 565, + 296, + 639 + ], + "type": "text", + "content": ". This merging strategy allows identity details retained within facial regions, while non-facial regions maintain the HOI context generated by SD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 180, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 180, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 180, + 672 + ], + "type": "text", + "content": "4.4. Residual Merge (RM)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "While Latent Merge enables interactions between the SD and PFD at the latent level, it lacks integration at the intermediate feature stage within the U-Net. To address this," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "we introduce Residual Merge, which enhances feature integration within the skip connections. In the U-Net, skip connections transmit residual features containing rich detail from the encoder to the decoder, significantly impacting the content and quality of the generated images [13, 31]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "text", + "content": "Our Residual Merge is applied to each skip connection layer " + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "inline_equation", + "content": "l = 1, \\dots, L" + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "text", + "content": "), as shown in Figure 4. First, residual features are extracted from both the PFD and SD U-Net backbones. A low-pass filter and high-pass filter is applied to the SD and PFD residuals, respectively. Low-pass and high-pass filters are applied to these residuals to balance global coherence and local precision: low-pass filtering SD residuals suppresses high-frequency noise to retain the broader scene layout, while high-pass filtering PFD residuals emphasizes fine-grained identity-specific details. To achieve precise integration between facial and non-facial regions, we apply the head segmentation mask " + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{\\mathrm{head}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "text", + "content": ", resized to match the residual resolution at layer " + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{R}}^{l}" + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "text", + "content": ". The fusion of residual features from both paths is then implemented for each layer " + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 133, + 555, + 312 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 321, + 555, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 321, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 319, + 321, + 555, + 338 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {\\text {m e r g e d}} ^ {l} = \\mathbf {M} _ {R} ^ {l} \\odot \\mathrm {H P} \\left(\\mathbf {R} _ {\\text {P F D}} ^ {l}\\right) + \\left(1 - \\mathbf {M} _ {R} ^ {l}\\right) \\odot \\mathrm {L P} \\left(\\mathbf {R} _ {\\text {S D}} ^ {l}\\right), \\tag {5}", + "image_path": "48f2a048a9bef934513f5ef77bacf85d3b375f808b8586f9ca1b626078dee6bd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{HP}(\\cdot)" + }, + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}(\\cdot)" + }, + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "text", + "content": " denote the high pass and low pass filter, " + }, + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{PFD}}" + }, + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "text", + "content": " denote the residual features from the PFD and SD paths, respectively. The merged residual " + }, + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{merged}}" + }, + { + "bbox": [ + 313, + 345, + 554, + 407 + ], + "type": "text", + "content": " is then concatenated with the bottom-up features in the PFD U-Net for further processing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 407, + 554, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 466 + ], + "type": "text", + "content": "The Residual Merge strategy enables fine-grained feature integration within the U-Net. Through this, facial regions retain high-frequency identity details from PFD while non-facial regions incorporate interaction layouts from SD, ensuring cohesive and contextually aligned outputs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 478, + 395, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 395, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 395, + 491 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 498, + 364, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 364, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 364, + 510 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 518, + 533, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 533, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 533, + 529 + ], + "type": "text", + "content": "Evaluation Data. We use two test sets for evaluation:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 530, + 554, + 687 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 313, + 530, + 554, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 554, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 554, + 614 + ], + "type": "text", + "content": "1) HOI-Focused Personalized Face Generation: To test complex human-object interactions (HOI), we use interaction labels from the widely-used V-COCO [8] dataset. V-COCO provides diverse HOI scenarios for robust evaluation of interaction fidelity. We construct 30 prompts in the \"subject + interaction + object\" format (e.g., \"man kicking a ball\") with the same 15 reference subjects as in 1)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 554, + 687 + ], + "type": "text", + "content": "2) General Personalized Face Generation: Following prior studies [21, 35], we compile a benchmark with 15 reference subjects, each paired with 40 prompts spanning diverse scenarios like context, stylization, accessory, and actions. These prompts aim to assess both identity retention and the model's ability to adapt to various text-guided settings." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 688, + 526, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 688, + 526, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 526, + 699 + ], + "type": "text", + "content": "Full prompt details are provided in the Appendix ??" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 701, + 534, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 534, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 534, + 713 + ], + "type": "text", + "content": "Metrics. We employ the following evaluation metrics:" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23779" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 194 + ], + "type": "table", + "html": "
MethodArchitectureIdentity Preservation (%)Prompt Consistency (%)HOI Alignment (%)
StableDiffusion v1.5 [25]SD1.5NA24.9660.23
StableDiffusion XL [22]SDXLNA25.3270.78
FastComposer [35]SD1.553.5721.3035.96
+ OursSD1.555.28(+1.71)23.16(+1.86)56.65(+20.69)
IP-Adapter [37]SDXL62.8622.7349.83
+ OursSDXL55.74(-7.11)24.20(+1.47)68.30(+18.47)
PhotoMaker [18]SDXL50.5423.9656.77
+ OursSDXL51.34(+0.80)24.34(+0.38)76.01(+19.24)
", + "image_path": "402827fbd901cebc655b5e14a31dd6590f49b959dacdde163a9b2a7ad51df308.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 246 + ], + "type": "text", + "content": "Table 1. Comparison of Our Method with Baseline Approaches on HOI-Specific Personalized Face Generation. StableDiffusion serves as the text-only baseline without subject conditioning. PersonaHOI seamlessly incorporates existing Personalized Face Diffusion models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) with their corresponding StableDiffusion architectures. We bold the higher number for each pair of comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 66, + 260, + 545, + 567 + ], + "blocks": [ + { + "bbox": [ + 66, + 260, + 545, + 567 + ], + "lines": [ + { + "bbox": [ + 66, + 260, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 66, + 260, + 545, + 567 + ], + "type": "image", + "image_path": "c25e15c29661e3c6175bbf9f426a5082b5f09d14764dc55613f6f9362e9d0866.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 574, + 555, + 597 + ], + "lines": [ + { + "bbox": [ + 55, + 574, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 555, + 597 + ], + "type": "text", + "content": "Figure 5. Qualitative Examples of PersonaHOI and Baseline Models. Comparison of baseline models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) and their PersonaHOI-enhanced results for diverse human-object interaction prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 618, + 294, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 294, + 700 + ], + "type": "text", + "content": "1) Identity Preservation: Measures how well the generated images retain the identity of the reference image. Face detection is performed using MTCNN [41], and identity similarity is calculated with FaceNet [28], following [21, 35]. 2) Prompt Consistency: Evaluates text-image alignment using CLIP-based [23] scores, where higher values indicate better prompt adherence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 294, + 714 + ], + "type": "text", + "content": "3) HOI Alignment (Ours): Assesses how well the gen" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": "erated images depict specified human-object interactions. We adopt UPT [40], an HOI detector pre-trained on V-COCO dataset, to identify the interaction specified in text prompts for generated images with HOI alignment score " + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{HOI}}" + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": ". Specifically, UPT provides confidence scores for detected human and object classes (" + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{Human}}" + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{Object}}" + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": ") from its object detection head, as well as interaction logits (" + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{Inter}}" + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": ") from its interaction head. With these outputs, HOI align-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23780" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 179, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 179, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 179, + 83 + ], + "type": "text", + "content": "ment score " + }, + { + "bbox": [ + 55, + 72, + 179, + 83 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{HOI}}" + }, + { + "bbox": [ + 55, + 72, + 179, + 83 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 100, + 87, + 295, + 102 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 87, + 295, + 102 + ], + "spans": [ + { + "bbox": [ + 100, + 87, + 295, + 102 + ], + "type": "interline_equation", + "content": "S _ {\\mathrm {H O I}} = \\left(S _ {\\text {H u m a n}} \\cdot S _ {\\text {O b j e c t}}\\right) ^ {\\lambda} \\cdot \\sigma \\left(L _ {\\text {I n t e r}}\\right) \\tag {6}", + "image_path": "72b9288c145caf50cd5936a5ee89aac649019f21c13879f4366b878fea66f2cc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 109, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 295, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 109, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 109, + 295, + 144 + ], + "type": "text", + "content": " is the sigmoid activation function and " + }, + { + "bbox": [ + 55, + 109, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\lambda = 2.8" + }, + { + "bbox": [ + 55, + 109, + 295, + 144 + ], + "type": "text", + "content": ". We validate the human alignment and effectiveness for our proposed HOI Alignment metric in Appendix ??" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 147, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 296, + 291 + ], + "type": "text", + "content": "Compared Methods and Implementation Details. We compare our method with state-of-the-art learning-based personalized face diffusion models, including FastComposer [35], IP-Adapter [37], and PhotoMaker [18]. Our PeronaHOI used their pre-trained diffusion backbones as the additional branch, i.e., StableDiffusion v1.5 for FastComposer and StableDiffusion XL for PhotoMaker, with output resolutions set to " + }, + { + "bbox": [ + 55, + 147, + 296, + 291 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 147, + 296, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 147, + 296, + 291 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 147, + 296, + 291 + ], + "type": "text", + "content": ", respectively. Head segmentation for SD generated image is performed with DensePose [24] due to its robust segmentation accuracy. In our generation process, no additional training or test-time tuning is applied to any compared methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 297, + 265, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 265, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 265, + 309 + ], + "type": "text", + "content": "5.2. Personalized Face with HOI Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 316, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 295, + 508 + ], + "type": "text", + "content": "Qualitative Comparison. Figure 5 illustrates qualitative comparisons between baseline models (FastComposer [35], IP-Adapter [37], PhotoMaker [18]) and their PersonaHOI-enhanced counterparts for diverse human-object interaction prompts. Baseline methods often fail to produce meaningful interactions, leading to missing or poorly integrated objects and unnatural human-object dynamics. For instance, in the prompt \"A man skateboarding with skateboard\" (first row of Figure 5), FastComposer [35] and PhotoMaker omit the skateboard entirely and IP-Adapter fails to generate the accurate interaction (\"holding\" the skateboard instead), resulting in inconsistent outputs with the text prompt. In contrast, PersonaHOI-enhanced models generate realistic skateboard placement and natural human-object dynamics. Moreover, PersonaHOI also ensures high-fidelity preservation of facial identity given in reference images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": "Quantitative Comparison. Table ?? compares the performance of PersonaHOI across various baseline models. For FastComposer [35] and PhotoMaker [18], PersonaHOI achieves consistent improvements across all metrics. Notably, Interaction Alignment improves by " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "20.69\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " for FastComposer [35] and " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "19.24\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " for PhotoMaker [18], demonstrating superior human-object interaction coherence. Gains in Identity Preservation and Prompt Consistency further validate PersonaHOI's ability to retain identity while enhancing text-image alignment. For IP-Adapter [37], a decrease in Identity Preservation (from " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "62.86\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "55.74\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": ") is observed. This drop is attributed to IP-Adapter's inherent bias toward generating face-centric images, which inflates identity scores at the cost of interaction fidelity. PersonaHOI mitigates this trade-off by significantly improving Interaction Alignment " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "(+18.47\\%)" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " and Prompt Consistency " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "(+1.47\\%)" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": ", balancing identity preser" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 85, + 366, + 130 + ], + "blocks": [ + { + "bbox": [ + 326, + 75, + 356, + 83 + ], + "lines": [ + { + "bbox": [ + 326, + 75, + 356, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 75, + 356, + 83 + ], + "type": "text", + "content": "SD v1.5" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 85, + 366, + 130 + ], + "lines": [ + { + "bbox": [ + 319, + 85, + 366, + 130 + ], + "spans": [ + { + "bbox": [ + 319, + 85, + 366, + 130 + ], + "type": "image", + "image_path": "2bdbe366dc80e66c0d1f34dec8e05c7e98318476de53bf0c1cb2ba4579b6c4bb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 371, + 85, + 416, + 130 + ], + "blocks": [ + { + "bbox": [ + 388, + 75, + 450, + 84 + ], + "lines": [ + { + "bbox": [ + 388, + 75, + 450, + 84 + ], + "spans": [ + { + "bbox": [ + 388, + 75, + 450, + 84 + ], + "type": "text", + "content": "Reference Images" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 85, + 416, + 130 + ], + "lines": [ + { + "bbox": [ + 371, + 85, + 416, + 130 + ], + "spans": [ + { + "bbox": [ + 371, + 85, + 416, + 130 + ], + "type": "image", + "image_path": "9c7ab5efea08c132040a60736f44be0a123040bc8a3306fa06d4f8ab0c2d6aba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 419, + 85, + 465, + 130 + ], + "blocks": [ + { + "bbox": [ + 419, + 75, + 465, + 84 + ], + "lines": [ + { + "bbox": [ + 419, + 75, + 465, + 84 + ], + "spans": [ + { + "bbox": [ + 419, + 75, + 465, + 84 + ], + "type": "text", + "content": "ce Images" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 419, + 85, + 465, + 130 + ], + "lines": [ + { + "bbox": [ + 419, + 85, + 465, + 130 + ], + "spans": [ + { + "bbox": [ + 419, + 85, + 465, + 130 + ], + "type": "image", + "image_path": "2d1c715e371185d08bb0e98d4eab87bfd1ba5dfb9b19e4e9ffb9eed581844a28.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 469, + 85, + 516, + 130 + ], + "blocks": [ + { + "bbox": [ + 468, + 75, + 518, + 84 + ], + "lines": [ + { + "bbox": [ + 468, + 75, + 518, + 84 + ], + "spans": [ + { + "bbox": [ + 468, + 75, + 518, + 84 + ], + "type": "text", + "content": "FastComposer" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 85, + 516, + 130 + ], + "lines": [ + { + "bbox": [ + 469, + 85, + 516, + 130 + ], + "spans": [ + { + "bbox": [ + 469, + 85, + 516, + 130 + ], + "type": "image", + "image_path": "01f50b740e50ab61d8587932a7128dac0de71e3945069cfaa5133e81453d5728.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 518, + 85, + 564, + 130 + ], + "blocks": [ + { + "bbox": [ + 529, + 75, + 554, + 84 + ], + "lines": [ + { + "bbox": [ + 529, + 75, + 554, + 84 + ], + "spans": [ + { + "bbox": [ + 529, + 75, + 554, + 84 + ], + "type": "text", + "content": "+Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 518, + 85, + 564, + 130 + ], + "lines": [ + { + "bbox": [ + 518, + 85, + 564, + 130 + ], + "spans": [ + { + "bbox": [ + 518, + 85, + 564, + 130 + ], + "type": "image", + "image_path": "d188e5707d7e956c4753e1241c1af0dd4bd77e961969e79aca207cc8b0539f4a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 319, + 142, + 365, + 188 + ], + "blocks": [ + { + "bbox": [ + 321, + 132, + 465, + 141 + ], + "lines": [ + { + "bbox": [ + 321, + 132, + 465, + 141 + ], + "spans": [ + { + "bbox": [ + 321, + 132, + 465, + 141 + ], + "type": "text", + "content": "Caption: A man and a man planting a tree" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 142, + 365, + 188 + ], + "lines": [ + { + "bbox": [ + 319, + 142, + 365, + 188 + ], + "spans": [ + { + "bbox": [ + 319, + 142, + 365, + 188 + ], + "type": "image", + "image_path": "75441ba04fcf1bbe5df8df5022fd18409f1cacd8e9aa466b7c6399187666c300.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 373, + 142, + 414, + 188 + ], + "blocks": [ + { + "bbox": [ + 373, + 142, + 414, + 188 + ], + "lines": [ + { + "bbox": [ + 373, + 142, + 414, + 188 + ], + "spans": [ + { + "bbox": [ + 373, + 142, + 414, + 188 + ], + "type": "image", + "image_path": "e3b100927aa3fd3b79fffd8c1a4a7e09c25b936a4e4083643a8e9028659e81ed.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 189, + 470, + 198 + ], + "lines": [ + { + "bbox": [ + 321, + 189, + 470, + 198 + ], + "spans": [ + { + "bbox": [ + 321, + 189, + 470, + 198 + ], + "type": "text", + "content": "Caption: A man and a man cycling together" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 419, + 142, + 465, + 188 + ], + "blocks": [ + { + "bbox": [ + 419, + 142, + 465, + 188 + ], + "lines": [ + { + "bbox": [ + 419, + 142, + 465, + 188 + ], + "spans": [ + { + "bbox": [ + 419, + 142, + 465, + 188 + ], + "type": "image", + "image_path": "c18d9f62666ffd918d33e408363154d6e2ecefc1885cffb797d9e2e159453a39.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 208, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 208, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 553, + 274 + ], + "type": "text", + "content": "Figure 6. Qualitative results for multi-subject HOI generation. We compare generation outputs from SD v1.5 [25], FastComposer [35], and our PersonaHOI based on FastComposer [35] with different multi-subject interaction prompts. PersonaHOI not only preserves distinct identities but also generates coherent human-object interactions aligned with HOI layouts produced by SD v1.5." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 470, + 142, + 517, + 188 + ], + "blocks": [ + { + "bbox": [ + 470, + 142, + 517, + 188 + ], + "lines": [ + { + "bbox": [ + 470, + 142, + 517, + 188 + ], + "spans": [ + { + "bbox": [ + 470, + 142, + 517, + 188 + ], + "type": "image", + "image_path": "584d47e0e45e91815ee42c6f9b5991aa3e4908462f72b80c26f98653174ccaf3.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 518, + 142, + 564, + 188 + ], + "blocks": [ + { + "bbox": [ + 518, + 142, + 564, + 188 + ], + "lines": [ + { + "bbox": [ + 518, + 142, + 564, + 188 + ], + "spans": [ + { + "bbox": [ + 518, + 142, + 564, + 188 + ], + "type": "image", + "image_path": "c4905f95b6f1e79482f208088e00690b63521c3393c87359f17f69a4d238f1e6.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 287, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 553, + 323 + ], + "type": "text", + "content": "vation with realistic interaction synthesis. Despite the decrease, the identity preservation score of " + }, + { + "bbox": [ + 313, + 287, + 553, + 323 + ], + "type": "inline_equation", + "content": "55.74\\%" + }, + { + "bbox": [ + 313, + 287, + 553, + 323 + ], + "type": "text", + "content": " remains competitive for personalized face generation." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 324, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 396 + ], + "type": "text", + "content": "Overall, PersonaHOI bridges the gap between personalized face generation and interaction realism. Its training and tuning-free design enables seamless integration into different SD-based personalized face diffusion models, providing scalability and adaptability across diverse architectures and application scenarios." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 403, + 544, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 544, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 544, + 417 + ], + "type": "text", + "content": "5.3. Personalized Multi-Subject HOI Generation" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 422, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 601 + ], + "type": "text", + "content": "Generating images with multiple subjects engaged in human-object interactions (HOI) presents significant challenges in preserving distinct identities and accurately depicting interactions. As shown in Figure 6, our method enhances FastComposer [35] using its pre-trained SD v1.5 model, maintaining identity fidelity while generating coherent interactions. For instance, in the challenging \"cycling together\" (second line) scenario, PersonaHOI successfully preserves the unique identities of both subjects and generates two bicycles alongside a natural background. In contrast, FastComposer [35] produces two adjacent faces with minimal interaction details. These results highlight the potential of PersonaHOI in handling complex multi-subject HOI scenarios, ensuring both identity preservation and interaction realism." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 609, + 400, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 400, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 400, + 621 + ], + "type": "text", + "content": "5.4. More Results" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 714 + ], + "type": "text", + "content": "General Personalized Face Generation Performance. We evaluate PersonaHOI on the task of general personalized face generation [21, 35] across four categories of text prompts: Accessory, Style, Action, and Context. As shown in Table 2, the scores across all categories for the PersonaHOI-enhanced method are consistently comparable to or slightly better than the baseline FastComposer [35]." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23781" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 530, + 133 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 530, + 133 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 530, + 133 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 530, + 133 + ], + "type": "table", + "html": "
MethodAccessory (%)Style (%)Action (%)Context (%)Mean (%)
StableDiffusion v1.5 [25]NA / 26.70NA / 27.21NA / 23.66NA / 25.86NA / 25.86
FastComposer [35]54.65 / 24.2241.13 / 24.0155.35 / 21.3052.70 / 22.3150.95 / 22.96
+ Ours56.43 / 24.2546.07 / 23.9755.09 / 22.2153.77 / 22.5952.84 / 23.26
", + "image_path": "b046d87726dcdf498f5e0c912e1a3b88dc4672bcf75ea083efaa5ee130d3f4ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 194, + 294, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 194, + 294, + 254 + ], + "spans": [ + { + "bbox": [ + 54, + 194, + 294, + 254 + ], + "type": "text", + "content": "These results demonstrate that while our method is specifically designed for personalized HOI scenarios, it effectively preserves the generative capabilities of the base models and does not degrade their performance on general text prompts, showcasing its adaptability and robustness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 257, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 296, + 519 + ], + "type": "text", + "content": "Effectiveness of Individual Component. We conducted an ablation study on FastComposer [35] to evaluate the contributions of the Cross-Attention Constraint (CAC), Latent Merge (LM), and Residual Merge (RM) in PersonaHOI. As shown in Table ??, the full model, integrating all three components, achieves the best performance across all metrics, underscoring the complementary roles of our proposed modules. Removing LM or RM results in significant declines in both Identity Preservation and Interaction Alignment, highlighting the critical role of these merging strategies. The absence of CAC has a relatively smaller effect on Interaction Alignment, but it substantially impacts Identity Preservation. This is because without CAC, faces generated by PFD fail to align with the spatial layout provided by SD, disrupting the coherence between PFD and SD branches and leading to diminished alignment and identity fidelity in facial regions. In summary, the ablation results confirm the distinct and critical functionalities of each module. CAC ensures precise spatial alignment of faces between PFD and SD, while LM and RM are indispensable for consistently merging non-facial HOI-relevant features from SD with facial identity details from PFD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "Results on Additional Baselines. To further validate the efficacy of our approach, we conduct additional evaluations using InstantID [34] and PuLID [7]. Results in Table 3 and Figure 7 indicate PersonaHOI can benefit both PFD methods, notably improving HOI alignment and image-text consistency. InstantID shares design similarities with IP-Adapter, prioritizing portrait generation, which improves identity preservation but compromises interaction fidelity. The fusion mechanism of PersonaHOI significantly enhances HOI alignment and overall image-text consistency. PuLID, on the other hand, balances personalization against text-guided image synthesis capability during training. However, it can struggle to correctly place the object in some cases (second row in Figure 7). By incorporating PersonaHOI, we achieve consistent improvements across identity preservation, image-prompt consistency, and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 194, + 553, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 218 + ], + "type": "text", + "content": "interaction alignment, demonstrating the versatility and robustness of our method." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 225, + 550, + 290 + ], + "blocks": [ + { + "bbox": [ + 55, + 140, + 555, + 174 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 555, + 174 + ], + "type": "text", + "content": "Table 2. Comparison of Our Method with FastComposer [35] on General Personalized Face Generation. We compare across four categories of text prompts including Accessory, Style, Action, and Context, following [21, 35]. Results are formatted as \"Identity Preservation (\\%) / Prompt Consistency (\\%)\" and we bold the higher number for each pair of comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 225, + 550, + 290 + ], + "lines": [ + { + "bbox": [ + 318, + 225, + 550, + 290 + ], + "spans": [ + { + "bbox": [ + 318, + 225, + 550, + 290 + ], + "type": "table", + "html": "
MethodIdentity (%)Prompt (%)HOI (%)
InstantID76.2122.9345.73
+ Ours69.93 (-9.28)23.80 (+0.87)61.82 (+16.09)
PuLID55.9324.5072.10
+ Ours62.56 (+6.63)25.21 (+0.71)72.18 (+0.08)
", + "image_path": "deb446a2edefb06b1e589b58dab64f082fbd0d14ecfb5f717ec1b40ccede0dfc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 335, + 552, + 428 + ], + "blocks": [ + { + "bbox": [ + 313, + 298, + 554, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 298, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 554, + 319 + ], + "type": "text", + "content": "Table 3. Comparison of our method with InstantID and PuLID on HOI-Specific Personalized Face Generation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 335, + 552, + 428 + ], + "lines": [ + { + "bbox": [ + 319, + 335, + 552, + 428 + ], + "spans": [ + { + "bbox": [ + 319, + 335, + 552, + 428 + ], + "type": "table", + "html": "
Ref. ImageInstantID+OursRef. ImagePuLD+Ours
Caption: A man surfboarding with a surfboardCaption: A woman catching a frisbee
Caption: A woman riding a horseCaption: A woman snowboarding with a snowboard
", + "image_path": "c063ad6db310890186d59dafd9df446fdc172068c3fa15666d7cde1d58204723.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 441, + 554, + 462 + ], + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 462 + ], + "type": "text", + "content": "Figure 7. Visual comparison of InstantID [34] and PuLID [7] with our method." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 478, + 391, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 391, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 391, + 491 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "In this work, we propose a training&tuning-free framework that combines the strengths of Personalized Face Diffusion (PFD) models and StableDiffusion (SD) models to generate personalized images featuring realistic and complex human-object interactions (HOI). By introducing Cross-Attention Constraint, Latent Merge, and Residual Merge, our approach achieves seamless integration of identity-specific facial features from PFD with contextual interaction details from SD, ensuring both spatial alignment and contextual coherence. The flexibility of our framework allows adaptation to various PFD and SD architectures, demonstrating its robustness in handling diverse interaction scenarios. Without requiring additional fine-tuning or task-specific datasets, our method significantly improves identity retention and interaction realism compared to state-of-the-art approaches. This highlights the potential of leveraging pre-trained models to enable high-quality, personalized content generation across diverse applications." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23782" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 191 + ], + "type": "text", + "content": "[2] Li Chen, Mengyi Zhao, Yiheng Liu, Mingxu Ding, Yangyang Song, Shizun Wang, Xu Wang, Hao Yang, Jing Liu, Kang Du, et al. Photovise: Tuning-free image customization with text-to-image diffusion models, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 193, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 193, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 193, + 295, + 237 + ], + "type": "text", + "content": "[3] Zhuowei Chen, Shancheng Fang, Wei Liu, Qian He, Mengqi Huang, Yongdong Zhang, and Zhendong Mao. Dreamidentity: Improved editability for efficient face-identity preserved image generation. arXiv preprint arXiv:2307.00300, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 292 + ], + "type": "text", + "content": "[4] Dylan Campbell Frederic Z. Zhang and Stephen Gould. Spatially Conditioned Graphs for Detecting Human-Object Interactions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13319–13327, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 294, + 295, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 294, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 294, + 295, + 349 + ], + "type": "text", + "content": "[5] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel CohenOr. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 350, + 295, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 295, + 394 + ], + "type": "text", + "content": "[6] Chen Gao, Si Liu, Defa Zhu, Quan Liu, Jie Cao, Haoqian He, Ran He, and Shuicheng Yan. InteractGAN: Learning to Generate Human-Object Interaction. In Association for Computing Machinery, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 396, + 295, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 295, + 440 + ], + "type": "text", + "content": "[7] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. PuLID: Pure and Lightning ID Customization via Contrastive Alignment. In Advances in Neural Information Processing Systems, 2024. 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 441, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 295, + 463 + ], + "type": "text", + "content": "[8] Saurabh Gupta and Jitendra Malik. Visual Semantic Role Labeling. arXiv preprint arXiv:1505.04474, 2015. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 465, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 295, + 508 + ], + "type": "text", + "content": "[9] Junjie He, Yifeng Geng, and Liefeng Bo. UniPortrait: A Unified Framework for Identity-Preserving Single- and Multi-Human Image Personalization. arXiv preprint arXiv:2408.05939, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 510, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 296, + 564 + ], + "type": "text", + "content": "[10] Jiu Tian Hoe, Xudong Jiang, Chee Seng Chan, Yap-Peng Tan, and Weipeng Hu. InteractDiffusion: Interaction Control in Text-to-Image Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 567, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 609 + ], + "type": "text", + "content": "[11] Siteng Huang, Biao Gong, Yutong Feng, Xi Chen, Yuqian Fu, Yu Liu, and Donglin Wang. Learning disentangled identifiers for action-customized text-to-image generation. arXiv preprint arXiv:2311.15841, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 612, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 295, + 644 + ], + "type": "text", + "content": "[12] Junha Hyung, Jaeyo Shin, and Jaegul Choo. Magicapture: High-resolution multi-concept portrait customization. arXiv preprint arXiv:2309.06895, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[13] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing. arXiv preprint arXiv:2312.11392, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[14] Jian-Yu Jiang-Lin, Kang-Yang Huang, Ling Lo, Yi-Ning Huang, Terence Lin, Jhih-Ciang Wu, Hong-Han Shuai, and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Wen-Huang Cheng. ReCorD: Reasoning and Correcting Diffusion for HOI Generation. In Proceedings of the ACM International Conference on Multimedia (ACM MM), 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "type": "text", + "content": "[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[16] Xiaoming Li, Xinyu Hou, and Chen Change Loy. When stylegan meets stable diffusion: a w+ adapter for personalized image generation. arXiv preprint arXiv:2311.17461, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "type": "text", + "content": "[17] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-Set Grounded Text-to-Image Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "text", + "content": "[18] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 310, + 555, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 555, + 352 + ], + "type": "text", + "content": "[19] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep Learning Face Attributes in the Wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[20] A. Nagrani, J. S. Chung, and A. Zisserman. Voxceleb: a large-scale speaker identification dataset. In INTER-SPEECH, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "type": "text", + "content": "[21] Xu Peng, Junwei Zhu, Boyuan Jiang, Ying Tai, Donghao Luo, Jiangning Zhang, Wei Lin, Taisong Jin, Chengjie Wang, and Rongrong Ji. PortraitBooth: A Versatile Portrait Model for Fast Identity-preserved Personalization. arXiv preprint arXiv:2312.06354, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "type": "text", + "content": "[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 513, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 588 + ], + "type": "text", + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[24] Iasonas Kokkinos Riza Alp Güler, Natalia Neverova. Dense-Pose: Dense Human Pose Estimation In The Wild. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 4, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "type": "text", + "content": "[25] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[26] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23783" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "type": "text", + "content": "[27] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. arXiv preprint arXiv:2307.06949, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "type": "text", + "content": "[28] Florian Schroff, Dmitry Kalenichenko, and James Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "type": "text", + "content": "[29] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. arXiv preprint arXiv:2304.03411, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 293 + ], + "type": "text", + "content": "[30] Kaede Shiohara and Toshihiko Yamasaki. Face2diffusion for fast and editable face personalization. arXiv preprint arXiv:2403.05094, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 316 + ], + "type": "text", + "content": "[31] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. FreeU: Free Lunch in Diffusion U-Net. In CVPR, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "type": "text", + "content": "[32] Yu-Chuan Su, Kelvin CK Chan, Yandong Li, Yang Zhao, Han Zhang, Boqing Gong, Huisheng Wang, and Xuhui Jia. Identity encoder for personalized diffusion. arXiv preprint arXiv:2304.07429, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 295, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 404 + ], + "type": "text", + "content": "[33] Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. Face0: Instantaneously conditioning a text-to-image model on a face. In SIGGRAPH Asia 2023 Conference Papers, pages 1-10, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 405, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 295, + 448 + ], + "type": "text", + "content": "[34] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. InstantID: Zero-shot Identity-Preserving Generation in Seconds. arXiv preprint arXiv:2401.07519, 2024. 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "text", + "content": "[35] Guangxuan Xiao, Tianwei Yin, William T Freeman, Fredo Durand, and Song Han. FastComposer: Tuning-Free Multi-Subject Image Generation with Localized Attention. In International Journal of Computer Vision, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "type": "text", + "content": "[36] Yuxuan Yan, Chi Zhang, Rui Wang, Yichao Zhou, Gege Zhang, Pei Cheng, Gang Yu, and Bin Fu. Facestudio: Put your face everywhere in seconds. arXiv preprint arXiv:2312.02663, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "type": "text", + "content": "[37] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IpAdapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 593, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 295, + 646 + ], + "type": "text", + "content": "[38] Hangjie Yuan, Jianwen Jiang, Samuel Albanie, Tao Feng, Ziyuan Huang, Dong Ni, and Mingqian Tang. RLIP: Relational Language-Image Pre-training for Human-Object Interaction Detection. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "text", + "content": "[39] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Samuel Albanie, Yining Pan, Tao Feng, Jianwen Jiang, Dong Ni, Yingya Zhang, and Deli Zhao. RLIPv2: Fast Scaling of Relational Language-Image Pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 329 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "text", + "content": "[40] Frederic Z. Zhang, Dylan Campbell, and Stephen Gould. Efficient Two-Stage Detection of Human-Object Interactions with a Novel Unary-Pairwise Transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20104-20112, 2022. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "text", + "content": "[41] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2016. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 228 + ], + "type": "text", + "content": "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In IEEE International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 229, + 554, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 554, + 284 + ], + "type": "text", + "content": "[43] Xulu Zhang, Xiao-Yong Wei, Wengyu Zhang, Jinlin Wu, Zhaoxiang Zhang, Zhen Lei, and Qing Li. A Survey on Personalized Content Synthesis with Diffusion Models. arXiv preprint arXiv:2405.05538, 2024. https://arxiv.org/abs/2405.05538.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 285, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 554, + 329 + ], + "type": "text", + "content": "[44] Yuxin Zhang, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. MotionCrafter: One-Shot Motion Customization of Diffusion Models. arXiv preprint arXiv:2312.05288, 2023. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23784" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Personalized Preference Fine-tuning of Diffusion Models/005a06ec-5d3e-419e-a758-dd73f721c5f6_content_list.json b/2025/Personalized Preference Fine-tuning of Diffusion Models/005a06ec-5d3e-419e-a758-dd73f721c5f6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89ba4ed7758e2bac1ea0fb32b412e094f7875a8a --- /dev/null +++ b/2025/Personalized Preference Fine-tuning of Diffusion Models/005a06ec-5d3e-419e-a758-dd73f721c5f6_content_list.json @@ -0,0 +1,1447 @@ +[ + { + "type": "text", + "text": "Personalized Preference Fine-tuning of Diffusion Models", + "text_level": 1, + "bbox": [ + 212, + 130, + 784, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meihua Dang $^{1*}$ Anikait Singh $^{1*}$ Linqi Zhou $^{1,2}$ Stefano Ermon $^{1}$ Jiaming Song $^{2}$ $^{1}$ Stanford University, $^{2}$ Luma AI", + "bbox": [ + 130, + 179, + 867, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mhdang,anikait,ermon}@cs.stanford.edu,{alexzhou,jiaming}@lumalabs.ai", + "bbox": [ + 189, + 219, + 800, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "RLHF techniques like DPO can significantly improve the generation quality of text-to-image diffusion models. However, these methods optimize for a single reward that aligns model generation with population-level preferences, neglecting the nuances of individual users' beliefs or values. This lack of personalization limits the efficacy of these models. To bridge this gap, we introduce PPD, a multi-reward optimization objective that aligns diffusion models with personalized preferences. With PPD, a diffusion model learns the individual preferences of a population of users in a few-shot way, enabling generalization to unseen users. Specifically, our approach (1) leverages a vision-language model (VLM) to extract personal preference embeddings from a small set of pairwise preference examples, and then (2) incorporates the embeddings into diffusion models through cross attention. Conditioning on user embeddings, the text-to-image models are fine-tuned with the DPO objective, simultaneously optimizing for alignment with the preferences of multiple users. Empirical results demonstrate that our method effectively optimizes for multiple reward functions and can interpolate between them during inference. In real-world user scenarios, with as few as four preference examples from a new user, our approach achieves an average win rate of $76\\%$ over Stable Cascade, generating images that more accurately reflect specific user preferences.", + "bbox": [ + 89, + 300, + 485, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 705, + 220, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models [12, 40, 42, 43] have achieved great success in text-to-images generation tasks, as demonstrated by models such as DALL-E 2 [33], Imagen [35], Stable Diffusion [29, 34], and Stable Cascade [27]. These large-scale models, often trained through supervised fine-tuning on text-image datasets containing high-quality aesthetic images such as LAION-5B [38], demonstrate impressive image quality and a remarkable visual understanding of text prompts. Recent advances focus on fine-tuning diffusion", + "bbox": [ + 89, + 729, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models to better align with human preferences. Methods such as DDPO [2] and DPOK [9] employ reinforcement learning from human feedback (RLHF) techniques to optimize human-alignment reward functions. Alternatively, Diffusion-DPO [45] utilize direct preference optimization (DPO) [32] to directly optimize for human preferences using an offline ranking dataset. SD3 [7] shows that fine-tuning diffusion models with human feedback techniques can significantly enhance overall image quality.", + "bbox": [ + 511, + 268, + 906, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these promising results, Diffusion-DPO aligns a model with a single reward that represents population-level human preferences. These fine-tuning techniques are thus insufficient for achieving personalized preference alignment. As studied in domains such as language [10, 36, 44], individual users have unique preferences—for instance, some may favor images with brighter colors, while others might prefer images with centralized foregrounds. Finetuning a separate model for each person based on their preference data is impractical and does not scale well, limiting the feasibility of personalized alignment. Additionally, this may limit cross-user generalization.", + "bbox": [ + 511, + 409, + 908, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A complementary research direction introduces additional control mechanisms into diffusion models to enable customization and personalization. Approaches such as IP-Adapter [51] and CtrlNet [54] allow for controlled image generation by using an additional conditioning image that encodes personalization features. However, these methods are constrained by their reliance on single-image inputs and do not directly learn or adapt to human preferences.", + "bbox": [ + 511, + 595, + 908, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these limitations, we propose PPD, a general framework that incorporates personalization into a diffusion model preference alignment through multi-reward optimization. Our approach learns a diverse set of human preferences within a single model and generalizes effectively to new users. Personalization in our approach requires minimal additional data, as we label samples in the preference dataset with a unique identifier for each user. This annotation is inexpensive to obtain and commonly included in datasets such as Pick-a-Pic [16]. We leverage vision-language models (VLMs) such as LLaVA-OneVision [20] to extract user preference embeddings from", + "bbox": [ + 511, + 719, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions. \nWork done during MD's internship at Luma AI.", + "bbox": [ + 112, + 875, + 370, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8020", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "few-shot pairwise preference examples for each user. We show that these few-shot text embeddings from preference pairs serve as effective features to represent reward functions under the Bradley-Terry model [3] and thus are suitable for conditioning the diffusion model. Next, adopting Stable Cascade [27] as our base model, we condition the diffusion model on these user embeddings through additional cross-attention layers and fine-tune it with a variant of the Diffusion-DPO [45] objective for preference alignment.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our empirical evaluations demonstrate that PPD can effectively optimize multiple rewards including CLIP [13, 31], Aesthetic [37], and HPS [47], significantly outperforming baselines. It also interpolates smoothly between different rewards during inference. Additionally, in real-world user scenarios, PPD aligns more closely with individual user preferences on the dataset Pick-a-Pic [16]. Notably, with only a few-shot ranking examples for a new user, PPD generate images that align closely with that user's specific preference and outperforms prior approaches such as Stable Cascade, achieving an average win rate of $76\\%$ , as evaluated by GPT4o [26].", + "bbox": [ + 89, + 227, + 480, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 420, + 232, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Aligning Diffusion Models Aligning text-to-image diffusion models has been extensively studied. Some methods optimize with online RL using a Bradley Terry reward function [2, 9, 55]. Others fine-tune diffusion models to increase the reward of generated images by gradient backpropagation on the reward function [5, 30, 49]. Recently proposed approaches [45, 50] leverage the duality of the reward learning and policy optimization from DPO [32] to eliminate the need of on-policy samples and only utilize offline human preference data. Alternatively, DiffusionKTO [22] replaces DPO with KTO [8], eliminating the need for paired data. Additionally, some methods modify the training data distribution on visually appealing and textually cohered data [6, 18, 48], while others improve text accuracy by re-captioning pre-collected web images [11, 39].", + "bbox": [ + 89, + 445, + 482, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Controllable Generation for Diffusion Models To address the lack of fine-grained control in image synthesis, various methods add conditional controls to text-to-image diffusion models. CtrlNet [54] and T2I-Adapter [25] introduce additional UNet modules that encode spatial information, such as edges and human poses, derived from conditioning images. IP-Adapter [51] introduces additional cross-attention layers to incorporate high-level semantic information from reference images.", + "bbox": [ + 89, + 691, + 482, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLM Alignment The alignment of large language models (LLMs) using RLHF objectives has been extensively explored [1, 32, 56], demonstrating effective instruction-following and alignment to population values. However,", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a key challenge remains in representing the diversity of human preferences [36], as these models often capture an averaged population preference rather than accounting for individual differences, leading to biases and reduced alignment with underrepresented communities. Pluralistic Alignment [10, 44] addresses this issue by modeling a distribution of individuals, each with distinct values and perspectives, with methods like distributional and overton alignment enhancing calibration across diverse user groups. Several works have investigated LLM personalization through the reparameterization of reward models [23, 28, 52], enabling a more tailored user experience. However, these works leverage synthetic users and study a limited number of users, particularly in text and robotic control domains.", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 316, + 638, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce the preliminaries behind diffusion models, DPO and its adaption to diffusion models.", + "bbox": [ + 511, + 343, + 903, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models Diffusion models are generative models consisting two processes: a forward (diffusion) process and a backward (denoising) process [12, 40, 42, 43]. In the forward process, Gaussian noise $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ is progressively added to a data distribution $P_{\\mathrm{data}}(\\mathbf{x}_0)$ over $T$ steps, guided by noise scheduling functions $\\alpha_{t}$ and $\\sigma_t$ . At timestep $t\\in [0,T]$ , noisy images $\\mathbf{x}_t$ are generated via $\\mathbf{x}_t = \\alpha_t\\mathbf{x}_0 + \\sigma_t\\epsilon$ . In the backward process, samples are generated from Gaussian noise with a learnable model, which can also be conditioned on additional inputs, such as text in the case of text-to-image diffusion models. Typically, diffusion models - denoted as $\\epsilon_{\\theta}$ - are trained to predict the Gaussian noise $\\epsilon$ from noisy images $\\mathbf{x}_t$ at timestep $t$ with conditioning $\\mathbf{c}$ , using a denoising objective:", + "bbox": [ + 511, + 388, + 906, + 601 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {D M}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\omega (\\lambda_ {t}) \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 612, + 903, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\lambda_{t} = \\log \\left[\\alpha_{t}^{2} / \\sigma_{t}^{2}\\right]$ is a signal-to-noise ratio, $\\omega (\\lambda_t)$ is a pre-specified weighting function [14]. Once the model $\\epsilon_{\\theta}$ is trained, images can be generated from random noise using accelerated sampling methods such as DDIM [41].", + "bbox": [ + 511, + 642, + 903, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Direct Preference Optimization (DPO) DPO [32] is a fine-tuning paradigm that aligns autoregressive large language models (LLMs) with human preferences without relying on traditional reinforcement learning techniques. Unlike Reinforcement Learning from Human Feedback (RLHF), which involves training a reward model and then fine-tuning the LLM using reinforcement learning, DPO directly optimizes the model based on human preference data. Given a dataset containing ranking pairs $(\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-)$ where $\\mathbf{c}$ being the prompt and $\\mathbf{x}_0^+, \\mathbf{x}_0^-$ being the preferred and dispreferred responses labeled by human, the training objective optimizes a simple loss under which we can obtain an", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8021", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/34d385018c2e606d6fb799da3bd83b159e7de042ab8269021764468ded12aeef.jpg", + "image_caption": [ + "Stage 1: generate user profile with few-shot image pairs.", + "Stage 2: train personalized diffusion model.", + "Figure 1. The overall architecture of PPD. In Stage 1, user embedding are generated with few-shot preference examples using a VLM. In Stage 2, we fine-tune diffusion models on the preference datasets with the user embedding as conditioning added to cross-attention." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 911, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "optimal policy:", + "bbox": [ + 89, + 421, + 194, + 438 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\mathrm {D P O}} (\\theta) = - \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0} ^ {+}, \\mathbf {x} _ {0} ^ {-}} [ \\log \\sigma (\\beta \\Delta) ], \\text {w h e r e} \\\\ \\Delta = \\left(\\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0} ^ {+} \\mid \\mathbf {c}\\right)}{p _ {\\mathrm {r e f}} \\left(\\mathbf {x} _ {0} ^ {+} \\mid \\mathbf {c}\\right)} - \\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0} ^ {-} \\mid \\mathbf {c}\\right)}{p _ {\\mathrm {r e f}} \\left(\\mathbf {x} _ {0} ^ {-} \\mid \\mathbf {c}\\right)}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 446, + 480, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma$ is the sigmoid function and $\\beta$ is a hyperparameter determining how much $p_{\\theta}(\\mathbf{x}_0|\\mathbf{c})$ may deviate from a reference distribution $p_{\\mathrm{ref}}(\\mathbf{x}_0|\\mathbf{c})$ . Intuitively, the objective aims to increase the likelihoods of preferred responses while decreasing that of dispreferred ones.", + "bbox": [ + 89, + 513, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion DPO The DPO objective in Eq. (2) requires optimization of log-likelihoods $\\log p_{\\theta}(\\mathbf{x}_0^+|\\mathbf{c})$ , which is analytically intractable for diffusion models. Diffusion-DPO [45] leverages the fact that the (appropriately weighted) training objective of diffusion models as in Eq. (1) can be viewed as the evidence lower bound (ELBO) of their log-likelihoods. It derives an upper bound to DPO for diffusion models to align them with human preferences:", + "bbox": [ + 89, + 604, + 483, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {D i f f - D P O}} (\\theta) = - \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0} ^ {+}, \\mathbf {x} _ {0} ^ {-}} [ \\log \\sigma (- \\beta T \\omega (\\lambda_ {t}) \\Delta) ], \\text {w h e r e}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 736, + 477, + 757 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta = \\| \\epsilon^ {+} - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, t) \\| _ {2} ^ {2} - \\| \\epsilon^ {+} - \\epsilon_ {\\mathrm {r e f}} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\\\ - \\left(\\| \\epsilon^ {-} - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, t) \\| _ {2} ^ {2} - \\| \\epsilon^ {-} - \\epsilon_ {\\text {r e f}} (\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, t) \\| _ {2} ^ {2}\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 758, + 480, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{c}$ is the caption; $\\mathbf{x}_0^+$ and $\\mathbf{x}_0^-$ are the preferred and dispreferred images; $\\mathbf{x}_t^+ = \\alpha_t\\mathbf{x}_0^+ +\\sigma_t\\epsilon^+$ and $\\mathbf{x}_t^- = \\alpha_t\\mathbf{x}_0^- +\\sigma_t\\epsilon^-$ are noisy images obtained from forward process with Gaussian noise $\\epsilon^{+}$ and $\\epsilon^{-}$ at timestep $t$ . This objective computes 4 terms as in Eq. (1): the denoising loss of training and preference distributions $\\epsilon_{\\theta}$ and $\\epsilon_{\\mathrm{ref}}$ on preferred and dispreferred", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "noisy images $\\mathbf{x}_t^+$ and $\\mathbf{x}_t^-$ respectively. Intuitively, it aims to decrease the loss of preferred images while increasing that of dispreferred ones.", + "bbox": [ + 511, + 421, + 906, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 479, + 604, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite impressive image quality and alignment control achieved in prior works, a key limitation among works that use DPO for diffusion models is that it assumes a uniform preference among all users for generated images. In practice, this assumption often fails, as individuals may have diverse preferences related to style, color, lighting, and other visual attributes. While such preferences can be easily expressed by human through pairwise comparisons, it is challenging to summarize them accurately through text or single-image prompts, due to the inaccuracy of natural language and the limitations inherent in single-prompt formats.", + "bbox": [ + 511, + 506, + 906, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by this problem, we propose PPD, a method that aligns diffusion models with individual-level human preferences. We first formulate the problem of personalization and define a new DPO variant for this task (Sec. 4.1), and then show that individual-level preference learning can be achieved using few-shot comparison examples using a VLM (Sec. 4.2), finally we introduce a parameterization of denoiser that can takes advantage of the VLM embeddings (Sec. 4.3). The overall architecture is in Fig. 1.", + "bbox": [ + 511, + 672, + 908, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Problem Setup", + "text_level": 1, + "bbox": [ + 511, + 818, + 665, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a dataset containing $(\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-, \\mathbf{u})$ where $\\mathbf{c}$ is the text prompt, $\\mathbf{x}_0^+$ , $\\mathbf{x}_0^-$ are preferred and dispreferred generated images and $\\mathbf{u}$ represents the user labeling the preference. User features $\\mathbf{u}$ can be high-dimensional including", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8022", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "information like age or gender, but often $\\mathbf{u}$ consists only of scalar user IDs $\\mathbf{u} = [u]$ . We define $r(\\mathbf{c},\\mathbf{x}_0,\\mathbf{u})$ as the reward on image $\\mathbf{x}_0$ for user $\\mathbf{u}$ given prompt $\\mathbf{c}$ . We would like to fine-tune a text-to-image models $p_{\\theta_i}(\\mathbf{x}_0|\\mathbf{c})$ for individual user $\\mathbf{u}_i$ such that reward is maximized for this user while keeping close to a reference model $p_{\\mathrm{ref}}(\\mathbf{x}_0|\\mathbf{c})$ in terms of KL-divergence as regularization:", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {p _ {\\theta_ {i}}} \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0}} \\left[ r (\\mathbf {c}, \\mathbf {x} _ {0}, \\mathbf {u} _ {i}) \\right] \\tag {4} \\\\ - \\beta \\mathbb {D} _ {\\mathrm {K L}} \\left[ p _ {\\theta_ {i}} (\\mathbf {x} _ {0} | \\mathbf {c}) \\| p _ {\\text {r e f}} (\\mathbf {x} _ {0} | \\mathbf {c}) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 205, + 482, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ is a parameter controlling how much $p_{\\theta_i}(\\mathbf{x}_0|\\mathbf{c})$ deviates from $p_{\\mathrm{ref}}(\\mathbf{x}_0|\\mathbf{c})$ . Following [32], the optimal $p_{\\theta_i}$ for user $u_{i}$ takes the form", + "bbox": [ + 89, + 258, + 483, + 305 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta_ {i}} \\left(\\mathbf {x} _ {0} | \\mathbf {c}\\right) \\propto p _ {\\operatorname {r e f}} \\left(\\mathbf {x} _ {0} | \\mathbf {c}\\right) \\exp \\left(r \\left(\\mathbf {c}, \\mathbf {x} _ {0}, \\mathbf {u} _ {i}\\right) / \\beta\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 314, + 482, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Suppose model $p_{\\theta}$ has sufficient capacity, we can represent $p_{\\theta_i}(\\mathbf{x}_0|\\mathbf{c})$ as $p_{\\theta}(\\mathbf{x}_0|\\mathbf{c},\\mathbf{u}_i)$ , where $\\mathbf{u}_i$ serves as a conditioning. This allows us to jointly optimize a single model $p_{\\theta}(\\mathbf{x}_0|\\mathbf{c},\\mathbf{u})$ for all users. Following [45], we have the following objective to learn a personalized diffusion models:", + "bbox": [ + 89, + 339, + 483, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\mathrm {P P D}} (\\theta) = - \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0} ^ {+}, \\mathbf {x} _ {0} ^ {-}, \\mathbf {u}} [ \\log \\sigma (- \\beta T \\omega (\\lambda_ {t}) \\Delta) ], \\text {w h e r e} \\\\ \\Delta = \\| \\epsilon^ {+} - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, \\mathbf {u}, t) \\| _ {2} ^ {2} - \\| \\epsilon^ {+} - \\epsilon_ {\\mathrm {r e f}} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\\\ - \\left(\\| \\epsilon^ {-} - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, \\mathbf {u}, t\\right) \\| _ {2} ^ {2} - \\| \\epsilon^ {-} - \\epsilon_ {\\text {r e f}} \\left(\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, t\\right) \\| _ {2} ^ {2}\\right). \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 422, + 480, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This objective is similar to Eq. (3) but with the addition of $\\mathbf{u}$ , representing extra features conditioned on for each user. But what user features should be conditioned on?", + "bbox": [ + 89, + 492, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Generating User Features from a VLM with In-Context Examples", + "text_level": 1, + "bbox": [ + 89, + 546, + 483, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion on User Features in Preference Datasets One fundamental challenge in personalized image generation lies in the effective parameterization of the user representation, denoted as $\\mathbf{u}$ , for each individual. This parameterization is critical because it directly influences the model's capacity to generate outputs tailored to specific user preferences. Existing image preference datasets, such as Pick-a-Pic [16], typically lack comprehensive user information beyond basic identifiers, often focusing on generic preference data rather than capturing the nuanced individual preferences necessary for effective personalization.", + "bbox": [ + 89, + 583, + 482, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Even in preference datasets that include user information (e.g PRISM [15]), this information is frequently obtained through self-reported surveys, which can lead to incomplete or biased user profiles. Discrepancies often arise between stated preferences and actual behavior; users may not accurately articulate their likes and dislikes, or they may respond in a manner that they perceive as socially desirable rather than truthful. Such inconsistencies present significant challenges when attempting to model and predict user preferences accurately.", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8227e10fc14911015db064d4fdc1a615d4160ff02279b7591378f280cebe98fa.jpg", + "image_caption": [ + "Figure 2. Top-K accuracy of the User Classification. We fine-tune a user-classifier from the frozen embeddings from the VLM on few-shot preference examples for 300 users. This classifier significantly outperforms a random chance baseline." + ], + "image_footnote": [], + "bbox": [ + 537, + 89, + 883, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address these issues, we adopt a simplified approach by assuming access solely to a unique user ID for each individual in the dataset. This user ID serves as a key to differentiate users within the labeled preference dataset, represented as $\\mathcal{D}_{\\mathrm{pref}} = \\{(\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-, \\mathbf{s})^{(i)}\\}$ , where $\\mathbf{s}$ denotes a unique user identifier.", + "bbox": [ + 511, + 344, + 906, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditioning on a Unique User Identifier The primary challenge, then, is to leverage this user ID $i$ to construct a meaningful user embedding that can effectively guide the model's predictions. A naive approach might involve representing the embedding $\\mathbf{u}$ using a one-hot encoding of the user IDs, which can be formally defined by a function $f:\\mathbf{S}\\mapsto U$ , establishing a bijective mapping from the set of user IDs $\\mathbf{s}\\in \\{1,2,\\ldots ,|D|\\}$ to the space of one-hot encoded vectors $\\mathbf{u}_i\\in \\{0,1\\}^{|D|}$ , where $|D|$ represents the total number of unique users in the dataset. While this method uniquely encodes each user, it suffers from a critical limitation: the inability to generalize to new, unseen users, as the embedding space is inherently tied to the users present in the training data.", + "bbox": [ + 511, + 446, + 908, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance generalization and adaptability, we reconsider the parameterization of $u_{i}$ in the context of learning the distribution of reward functions, denoted as $r(\\mathbf{c},\\mathbf{x}_0,\\mathbf{u}_i)$ . Here, a more flexible and robust parameterization leverages few-shot examples that reflect an individual user's preferences. These examples can be extracted from the offline preference dataset for each user, providing concrete instances of preferred and dispreferred content. Conditioning on these examples allows for a more accurate estimation of the reward function $r(\\mathbf{c},\\mathbf{x}_0,\\mathbf{u}_i)$ for each user.", + "bbox": [ + 511, + 659, + 908, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we utilize the features of a pre-trained vision-language Model (VLM), specifically LLaVA-OneVision [21], an open-source multimodal model with multi-image processing capabilities. For each user, a set of $N = 4$ few-shot examples is selected, comprising a caption describing the context, a preferred image, and a", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8023", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dispreferred image. These examples are processed by the VLM, from which we extract an intermediate hidden state.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To validate that the user embedding is effective, we condition on the frozen embedding to classify which user provided the preference annotations. As seen in Figure 2, the classifier significantly outperforms chance with a top-16 accuracy of $90\\%$ , indicating that the embedding can disentangle different users in the preference dataset effectively. By incorporating this user-specific embedding into the conditioning of a text-to-image diffusion model, we enable user-aware reward modeling and generation.", + "bbox": [ + 88, + 119, + 482, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Personalization as Conditioning", + "text_level": 1, + "bbox": [ + 89, + 266, + 372, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to add the user embedding as conditioning to the diffusion model in image generation, we utilize additional cross-attention layers akin to IP-Adapter [51]. Here we learn to condition on the user information using the VLM embeddings and then integrate the image features into pretrained UNet by adapted modules with decoupled cross-attention. There are many other ways to add conditioning information to diffusion models; we chose this instantiation due to its training and parameter efficiency.", + "bbox": [ + 88, + 287, + 482, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize Stable Cascade [27] as our base model which comprises of three stages: Stages A-C. Stage A and B are image compressors, similar to the VAE in Stable Diffusion [34] but achieve a much higher compression rate; and Stage C is the generation module conditioned on text. In Stage C, the text features c from pretrained CLIP text encoder are plugged into the model by feeding into the cross-attention layers. Given the query features $\\mathbf{Z}$ and text features $\\mathbf{c}_t$ , the output is $\\mathbf{Z}' = \\text{Attention}(\\mathbf{Q}, \\mathbf{K}, \\mathbf{V}) = \\text{Softmax}(\\frac{\\mathbf{Q}\\mathbf{K}^T}{\\sqrt{d}})\\mathbf{V}$ , where $\\mathbf{Q} = \\mathbf{Z}\\mathbf{W}_q$ , $\\mathbf{K} = \\mathbf{c}_t\\mathbf{W}_k$ , $\\mathbf{V} = \\mathbf{c}_t\\mathbf{W}_v$ are query, key and values, and $\\mathbf{W}_q, \\mathbf{W}_k, \\mathbf{W}_v$ are trainable layers.", + "bbox": [ + 89, + 424, + 482, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following IP-Adapter [51], we utilize a decoupled crossattention mechanism to condition on the user embedding. First, we add a new cross-attention layer to each cross-attention layer in the Stage C model, using the original query $\\mathbf{Q}$ and introducing new keys $\\mathbf{K}'$ and values $\\mathbf{V}'$ for user features. Then, we simply add the output of user cross-attention to the output of text cross-attention. Hence the final formulation of cross-attention is", + "bbox": [ + 89, + 609, + 482, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} ^ {\\prime} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V} + \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\left(\\mathbf {K} ^ {\\prime}\\right) ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V} ^ {\\prime}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 750, + 482, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Q} = \\mathbf{Z}\\mathbf{W}_q$ , $\\mathbf{K} = \\mathbf{c}_t\\mathbf{W}_k$ , $\\mathbf{V} = \\mathbf{c}_t\\mathbf{W}_v$ , $\\mathbf{K}' = \\mathbf{u}_t\\mathbf{W}_k'$ , $\\mathbf{V}' = \\mathbf{u}_t\\mathbf{W}_v'$ and $\\mathbf{W}_k'$ , $\\mathbf{W}_v'$ are new introduced trainable parameters for $\\mathbf{u}$ . During training, we optimize only the added cross-attention layers while keeping the parameters of pretrained diffusion model frozen, using the training objective as in Eq. (6). We also randomly drop user features (zero out the user embedding) in the training stage as regularization.", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/68424b2137a30dde53f8c38f4dc2184411912ae8fdd4d13be66bf85f06d6d0e1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CLIPAestheticHPS
Stable Cascade31.975.3323.87
Diffusion-DPO32.485.4625.96
SFT32.265.5625.78
ours32.665.9227.51
Diffusion-DPO (CLIP)32.96--
Diffusion-DPO (Aesthetic)-6.42-
Diffusion-DPO (HPS)--28.61
", + "bbox": [ + 514, + 88, + 908, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Averaged scores on three reward function. We condition on each reward to generate images both in SFT and our methods. Diffusion-DPO (*) are upper-bounds.", + "bbox": [ + 511, + 242, + 906, + 285 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This allows us to also prompt the model unconditionally of the user, by passing a zero embedding for the user.", + "bbox": [ + 511, + 301, + 906, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 345, + 645, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models and Dataset In this section, we demonstrate the performance of PPD across a range of experiments designed to assess its capabilities in personalized preference alignment. We use the objective from Eq. (6) to fine-tune the open-source diffusion model, Stable Cascade [27] – specifically Stage C, which is responsible for text-conditioned generation. Following Wallace et al. [45], we fine-tune our model on the Pick-a-Pic v2 dataset [16], a preference dataset containing 58K text prompts and 0.8M image pairs labeled by 5K users.", + "bbox": [ + 511, + 371, + 906, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details During training of PPD, we only optimize the added cross-attention layers while keeping the pretrained diffusion model fixed, as described in Sec. 4.3. We use AdamW [24] with an effective batch size of 768 pairs and a learning rate of $1 \\times 10^{-5}$ for all experiments, training for one epoch. We find $\\beta \\in [0.1, 2]$ for hyperparameter tuning. Details are in Appendix B.", + "bbox": [ + 511, + 527, + 905, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "User Information Conditioning We begin by demonstrating the model's ability to condition on synthetic user information, optimizing three distinct reward functions, we further show that generalize by linearly interpolating the weights of different reward functions (Sec. 5.1). We then show that the model can also condition on real user preferences, generating personalized images based on few-shot examples from new users (Sec. 5.2).", + "bbox": [ + 511, + 641, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Synthetic User Alignment", + "text_level": 1, + "bbox": [ + 511, + 772, + 748, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic User We utilize three reward functions, each representing an aspect of image-to-text generation quality, to simulate three distinct synthetic users: (i) CLIP (Open-CLIP ViT-H/14) [13, 31] for prompt image alignment, (ii) Aesthetic Predictor [37] for image visual appeal, and (iii) HPSv2 (Human Preference Score v2) [47] for human preference estimate. We relabel pairs $(\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-, \\mathbf{u})$ based", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8024", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/20be1be546b5fead67b0bd5c4f13a60cb64e577ed7f117541aa39af3dbb5ba32.jpg", + "image_caption": [ + "Prompt Alignment (CLIP)" + ], + "image_footnote": [], + "bbox": [ + 94, + 103, + 480, + 190 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f420a367979d577308e125d5b7d491db21d5fea58fcc7c85aee2c073ed16c37.jpg", + "image_caption": [ + "Aesthetic Quality" + ], + "image_footnote": [], + "bbox": [ + 94, + 205, + 478, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2eea0ef63b8efad1dbc5e2d104f77df3138e1c6441dae699bdf45b3e25503108.jpg", + "image_caption": [ + "Human Preference Score (HPS)", + "Figure 3. Automatic win rate evaluation with reward functions." + ], + "image_footnote": [ + "We compare against Stable Cascade, Diffusion-DPO, and SFT." + ], + "bbox": [ + 94, + 309, + 478, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on the above reward functions, where $\\mathbf{u}$ is a one-hot vector indicating the reward function, taking one of 3 values: $\\mathbf{u} \\in \\{[1,0,0],[0,1,0],[0,0,1]\\}$ .", + "bbox": [ + 89, + 452, + 482, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline We compare our method with: (1) Stable Cascade [27] pretrained model, which is also the reference model during training; (2) Diffusion-DPO, fine-tuned on Pick-a-Pic label without user conditioning [45]; (3) SFT, supervised fine-tuning on relabeled data with user conditioning. We also compare with (4) Diffusion-DPO (CLIP/Aesthetic/HPS), each annotation indicating a separately fine-tuned model with corresponding reward, this serves as the upper-bound of our method. For a fair comparison, all baselines are fine-tuned using Stable Cascade with the same architecture as our model, optimizing only the added cross-attention layers.", + "bbox": [ + 89, + 510, + 483, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-to-Image Generation For each method, we generate images with 30 inference steps using captions from PartiPrompts [53], a benchmark containing 1.6K examples across various categories and challenge aspects. For SFT and our method, we generate three images per prompt, each conditioned on one reward function. For Diffusion-DPO (CLIP/Aesthetic/HPS), images for each annotation are generated from its corresponding model. We report averaged reward scores in Tab. 1 and win rate comparisons in Fig. 3, we can see that our method significantly outperforms Stable Cascade in different reward functions, achieving win rates of $60\\%$ , $90\\%$ , and $93\\%$ in terms of CLIP, Aesthetic, and HPS respectively. Our method effectively optimizes each", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "reward by conditioning on individual preference, outperforming supervised fine-tuning by a large margin. Unlike Diffusion-DPO, our method can jointly optimize multiple user rewards, achieving win rates comparable to Diffusion-DPO models fine-tuned separately on each user's data.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalization via Interpolation During training we fine-tune the model using one-hot encoding of each objective. Specifically, the user embedding vector $\\mathbf{u}$ being $[1,0,0],[0,1,0],[0,0,1]$ corresponding to optimization for CLIP, Aesthetic, and HPS. The user embedding can also be seen as weighting mechanism for each objective. However, during inference, we are also interested in whether we can trade-off among different objectives by interpolating their respective weights. In Fig. 4, we generate images conditioned on various objective weightings, namely interpolating the given user embedding, we can see that as the weight of an objective increases, the corresponding score also rises, and vice versa. Therefore, our method can generalize to different reward combinations, even only trained on single objective data. This can not be achieved by optimizing separate models and each with an objective.", + "bbox": [ + 511, + 179, + 906, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This shows that by linearly interpolating the user embedding, we can adjust the rewards of the generated images accordingly. This finding aligns with the real user experiments presented next, as real user embeddings exhibit a linear correlation in embedding space.", + "bbox": [ + 511, + 421, + 905, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Personalized Real User Alignment", + "text_level": 1, + "bbox": [ + 511, + 506, + 812, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now consider a realistic setting with preferences from the Pick-a-Pic dataset for each user. In this setting, we utilize the user features which are generated from a VLM on in-context preference examples as outlined in Section 4.2. We scale from the 3 synthetic \"users\" in Section 5.1 to a filtered subset of the 4800 unique users in Pick-a-Pic.", + "bbox": [ + 511, + 527, + 905, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment setting We repartition the Pick-a-Pic dataset into a set of 4 splits: (1) Train, (2) Validation, (3) Test with Seen Users + Held-out Captions and (4) Test with Unseen Users + Held-out Captions. This decomposition allows us to study independently the generalization of our methods to new prompts and unseen users. Additionally, for each user, we take sets of $N = 4$ few-shot preference examples for each user and use these groupings for the conditioning of the VLM on few-shot preference pairs. We thus have to filter out 824 users in Pick-a-Pic which have less than 4 examples per user, as no user embedding for this user can be formed. We generate the user embedding through the procedure outlined in Sec. 4.2.", + "bbox": [ + 511, + 631, + 905, + 827 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generation Results We perform qualitative analysis on the set of images generated for each user. We perform this evaluation by generating a user profile per user. We use a set of few-shot preference pairs per user which we condition a", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8025", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/70a6289357c9c72c33aff0e3d98e28ae1bd8abc357aec192681bac4b954b5c79.jpg", + "image_caption": [ + "Figure 4. PPD is able to interpolate among three distinct objectives during inference. (a) generated images conditioned on (b) various weights, with three axes representing CLIP, Aesthetic, and HPS; (c) reward scores for each image. The score for each objective increases as its respective weight increases, and decreasing otherwise. For each row from left to right, the CLIP score increases alongside its weight, leading to a decrease in the Aesthetic score as its weight decreases. From bottom to top, the HPS score increases with its weight." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7354d9f93bc8ec5d6af7da47b40884f69a8751e5c2b1532969da8b6594ce367.jpg", + "image_caption": [ + "Comparison of Win Rates for Real User Alignment" + ], + "image_footnote": [], + "bbox": [ + 94, + 469, + 228, + 556 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8c952716d2a9018bb07db669b11a97a52d0af5438f45dc4ac3a4e4ae6441965.jpg", + "image_caption": [ + "Figure 5. Automatic Win Rate Evaluation with GPT-4o We evaluate on seen users, held out users, and aggregate to see win-rates for generic users, comparing Stable Cascade and PPD." + ], + "image_footnote": [], + "bbox": [ + 230, + 470, + 354, + 555 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c361648958e60ee5242fa93059135f946fae62ee10d07ffd8dd87131580d02e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 470, + 480, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VLM on along with Chain-of-Thought prompting (COT) to construct a user-profile consistent with the preference pairs. The COT prompt asks the model to summarize each image in the preference pair and describe the differences between them, which is then aggregated to form a user profile. Sample user profiles and the full Chain-of-Thought prompt is found in Appendix B. As seen in Fig. 6, our method has closer alignment to the user profile with respect to Stable Cascade and Diffusion-DPO, being able to pick up details about the user such as the highlighted part. While Stable Cascade and Diffusion-DPO with augmented captions can capture certain aspects of the user description, they often overfit to the user profile, resulting in generated images that overlook the intended image caption.", + "bbox": [ + 89, + 625, + 483, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Personalized Preference Alignment Evaluation We evaluate whether the generated model aligns with specific user preference using a VLM as a judge, leveraging strong", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VLMs such as GPT4o-vision [26], similar to recipes in prior work [4, 19]. In particular, we condition the VLM on few-shot examples from the user as well as a pair of images for a given test time caption and ask the model to judge which of the two images would be preferred by the user. To ensure consistency of the response the model selects, we judge each pair of images twice using the VLM, flipping the ordering of the images, filtering out pairs where the VLM is inconsistent. With this paradigm and 4 shots per user, we find that the VLM is able to match the ground truth personalized human preference with $83\\%$ accuracy.", + "bbox": [ + 511, + 459, + 906, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We decompose our evaluation into two settings: (1) evaluating users that were seen during training but on captions that were held out and (2) evaluating held-out users and captions. We aggregate over these settings to also compute metrics for the general user in the dataset. For the 4341 users in Pick-a-Pic we select 192 held out and 277 in-distribution users. For each of these users, we choose $4 + 1$ examples, where 4 examples are used for the in-context conditioning for the user and 1 example is used as the caption for evaluation. We compute an automatic win-rate using pairs of models, generating one image for each caption and evaluating with the VLM-as-Judge pipeline outlined above.", + "bbox": [ + 511, + 642, + 908, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report win rate comparisons in Fig. 5. Here, PPD significantly outperforms Stable Cascade in all three evaluation settings, achieving a win rate of $83\\%$ on seen users, and $78\\%$ on unseen users, aggregated to $81\\%$ on all users.", + "bbox": [ + 511, + 840, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8026", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/716469cf2a56c86cc04800742a61409a2c1e7b74de810f57e55e1c27a6ff294f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
User ProfileStable CascadeDiffusion-DPODiffusion-DPOOurs
Append user profile to caption as augmentation
The preferred images share common themes of wood grain, classic or futuristic design elements, and a focus on aesthetics. They often feature unique front grille patterns, large windows, and a mix of traditional and modern design features. ...A needle-felted palm tree
This user prefers images that have a personal and intimate feel with a focus on the subject's appearance and expression. They also appreciate images with soft focus backgrounds and warm lighting that highlight texture and detail. ...An anime girl, masterpiece, good line art, trending in pixiv
The preferred images ... such as a softer color palette, more dynamic lighting, and a balanced composition that enhances the visual appeal of the scene. The characters are often positioned off-center, which adds to the dynamism of the composition. The backgrounds are rich in detail. ...Aelita from Code Lyoko in the forest, gathering flowers
The user is drawn to images that blend elements of reality with fantastical elements, creating a sense of wonder and curiosity. ... The user may be interested in fantasy art, animation, and storytelling through visual media.A cat on a propaganda poster
", + "bbox": [ + 89, + 88, + 903, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6. Qualitative Analysis of Images Generated by PPD and Baselines. Compared to Diffusion-DPO, PPD achieves closer alignment with the generated user profile, highlighted in green. The caption-augmented method captures user profile details; however, it often leads to unintended image alterations, causing the image to disregard the caption itself, as indicated in red.", + "bbox": [ + 89, + 700, + 906, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion, Conclusion, and Limitations", + "text_level": 1, + "bbox": [ + 89, + 768, + 450, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a framework for personalizing text-to-image diffusion models using user feature embeddings from a VLM. By conditioning on these embeddings via decoupled cross-attention and fine-tuning with a DPO objective, our method enables accurate reward modeling and high-quality personalized image generation. It generalizes to unseen users with only a few preference examples at test time.", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Several open questions and limitations remain. Our current approach that conditioning on few-shot examples is caption-agnostic, selecting random user preferences; future work could explore active preference selection to improve personalization. Extending this method to video generation and studying the trade-off between implicit vs. stated preferences are also promising directions.", + "bbox": [ + 511, + 770, + 906, + 878 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8027", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is done during MD's internship at Luma AI and supported in part by ARO (W911NF-21-1-0125), ONR (N00014-23-1-2159), and the CZ Biohub. AS gratefully acknowledges the support of the NSF Graduate Research Fellowship Program.", + "bbox": [ + 89, + 114, + 480, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 207, + 186, + 223 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. Training a helpful and harmless assistant with reinforcement learning from human feedback, 2022. 2", + "[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2", + "[3] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 2", + "[4] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Qinglan Huang, Canyu Chen, Qinghao Ye, Zhihong Zhu, Yuqing Zhang, Jiawei Zhou, Zhuokai Zhao, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. MJ-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation? arXiv e-prints, art. arXiv:2407.04842, 2024. 7", + "[5] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 2", + "[6] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2", + "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1", + "[8] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306, 2024. 2", + "[9] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion" + ], + "bbox": [ + 99, + 232, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models. In Advances in Neural Information Processing Systems, 2024. 1, 2", + "[10] Shangbin Feng, Taylor Sorensen, Yuhan Liu, Jillian Fisher, Chan Young Park, Yejin Choi, and Yulia Tsvetkov. Modular Pluralism: Pluralistic Alignment via Multi-LLM Collaboration. arXiv e-prints, art. arXiv:2406.15951, 2024. 1, 2", + "[11] Gabriel Goh, James Betker, Li Jing, Aditya Ramesh, Tim Brooks, Jianfeng Wang, Lindsey Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Prafulla Dhariwal, Casey Chu, Joy Jiao, Jong Wook Kim, Alex Nichol, Yang Song, Lijuan Wang, and Tao Xu. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 2", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in neural information processing systems, pages 6840-6851, 2020. 1, 2", + "[13] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 2,5", + "[14] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In Advances in Neural Information Processing Systems 34, 2021. 2", + "[15] Hannah Rose Kirk, Alexander Whitefield, Paul Röttger, Andrew Bean, Katerina Margatina, Juan Ciro, Rafael Mosquera, Max Bartolo, Adina Williams, He He, Bertie Vidgen, and Scott A. Hale. The PRISM Alignment Project: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models. arXiv e-prints, art. arXiv:2404.16019, 2024. 4", + "[16] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569, 2023. 1, 2, 4, 5, 3", + "[17] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners, 2023. 3", + "[18] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2", + "[19] Seongyun Lee, Seungone Kim, Sue Hyun Park, Geewook Kim, and Minjoon Seo. Prometheus-Vision: Vision-Language Model as a Judge for Fine-Grained Evaluation. arXiv e-prints, art. arXiv:2401.06591, 2024. 7", + "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 3", + "[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. arXiv e-prints, art. arXiv:2408.03326, 2024. 4" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8028", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. arXiv preprint arXiv:2404.04465, 2024. 2", + "[23] Xinyu Li, Zachary C Lipton, and Liu Leqi. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133, 2024. 2", + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[25] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2", + "[26] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 2, 7", + "[27] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6", + "[28] Sriyash Poddar, Yanming Wan, Hamish Ivison, Abhishek Gupta, and Natasha Jaques. Personalizing reinforcement learning from human feedback with variational preference learning. arXiv preprint arXiv:2408.10075, 2024. 2", + "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1", + "[30] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5", + "[32] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, 2024. 1, 2, 4", + "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5", + "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi." + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, pages 36479-36494, 2022. 1", + "[36] Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. Whose Opinions Do Language Models Reflect? arXiv e-prints, art. arXiv:2303.17548, 2023. 1, 2", + "[37] Christoph Schuhmann. CLIP MLP aesthetic score predictor. https://github.com/christophschuhmann/improved-aesthetic-predictor, 2022. Accessed: 2024-11-06. 2,5", + "[38] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278–25294, 2022. 1", + "[39] Eyal Segalis, Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. A picture is worth a thousand words: Principled recaptioning improves image generation. arXiv preprint arXiv:2310.16656, 2023. 2", + "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265, 2015. 1, 2", + "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2", + "[42] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 1, 2", + "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2", + "[44] Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Ryting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, and Yejin Choi. A Roadmap to Pluralistic Alignment. arXiv e-prints, art. arXiv:2402.05070, 2024. 1, 2", + "[45] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 1, 2, 3, 4, 5, 6", + "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 3", + "[47] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 2, 5" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8029", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 2", + "[49] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, 2024. 2", + "[50] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8941-8951, 2024. 2", + "[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 1, 2, 5", + "[52] Chao Yu, Hong Lu, Jiaxuan Gao, Qixin Tan, Xinting Yang, Yu Wang, Yi Wu, and Eugene Vinitsky. Few-shot in-context preference learning using large language models. arXiv preprint arXiv:2410.17233, 2024. 2", + "[53] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 6", + "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2", + "[55] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. arXiv preprint arXiv:2401.12244, 4, 2024. 2", + "[56] Rui Zheng, Shihan Dou, Songyang Gao, Yuan Hua, Wei Shen, Binghai Wang, Yan Liu, Senjie Jin, Qin Liu, Yuhao Zhou, Limao Xiong, Lu Chen, Zhiheng Xi, Nuo Xu, Wenbin Lai, Minghao Zhu, Cheng Chang, Zhangyue Yin, Rongxiang Weng, Wensen Cheng, Haoran Huang, Tianxiang Sun, Hang Yan, Tao Gui, Qi Zhang, Xipeng Qiu, and Xuanjing Huang. Secrets of rlhf in large language models part i: Ppo, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 724 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8030", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Personalized Preference Fine-tuning of Diffusion Models/005a06ec-5d3e-419e-a758-dd73f721c5f6_model.json b/2025/Personalized Preference Fine-tuning of Diffusion Models/005a06ec-5d3e-419e-a758-dd73f721c5f6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..70e534e3bbc0f60d2cb571c615270a04c0a5ecb0 --- /dev/null +++ b/2025/Personalized Preference Fine-tuning of Diffusion Models/005a06ec-5d3e-419e-a758-dd73f721c5f6_model.json @@ -0,0 +1,2103 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.131, + 0.785, + 0.154 + ], + "angle": 0, + "content": "Personalized Preference Fine-tuning of Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.18, + 0.868, + 0.218 + ], + "angle": 0, + "content": "Meihua Dang\\(^{1*}\\) Anikait Singh\\(^{1*}\\) Linqi Zhou\\(^{1,2}\\) Stefano Ermon\\(^{1}\\) Jiaming Song\\(^{2}\\) \n\\(^{1}\\)Stanford University, \\(^{2}\\)Luma AI" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.22, + 0.802, + 0.235 + ], + "angle": 0, + "content": "{mhdang,anikait,ermon}@cs.stanford.edu,{alexzhou,jiaming}@lumalabs.ai" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.681 + ], + "angle": 0, + "content": "RLHF techniques like DPO can significantly improve the generation quality of text-to-image diffusion models. However, these methods optimize for a single reward that aligns model generation with population-level preferences, neglecting the nuances of individual users' beliefs or values. This lack of personalization limits the efficacy of these models. To bridge this gap, we introduce PPD, a multi-reward optimization objective that aligns diffusion models with personalized preferences. With PPD, a diffusion model learns the individual preferences of a population of users in a few-shot way, enabling generalization to unseen users. Specifically, our approach (1) leverages a vision-language model (VLM) to extract personal preference embeddings from a small set of pairwise preference examples, and then (2) incorporates the embeddings into diffusion models through cross attention. Conditioning on user embeddings, the text-to-image models are fine-tuned with the DPO objective, simultaneously optimizing for alignment with the preferences of multiple users. Empirical results demonstrate that our method effectively optimizes for multiple reward functions and can interpolate between them during inference. In real-world user scenarios, with as few as four preference examples from a new user, our approach achieves an average win rate of \\(76\\%\\) over Stable Cascade, generating images that more accurately reflect specific user preferences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.706, + 0.222, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.485, + 0.868 + ], + "angle": 0, + "content": "Diffusion models [12, 40, 42, 43] have achieved great success in text-to-images generation tasks, as demonstrated by models such as DALL-E 2 [33], Imagen [35], Stable Diffusion [29, 34], and Stable Cascade [27]. These large-scale models, often trained through supervised fine-tuning on text-image datasets containing high-quality aesthetic images such as LAION-5B [38], demonstrate impressive image quality and a remarkable visual understanding of text prompts. Recent advances focus on fine-tuning diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.908, + 0.408 + ], + "angle": 0, + "content": "models to better align with human preferences. Methods such as DDPO [2] and DPOK [9] employ reinforcement learning from human feedback (RLHF) techniques to optimize human-alignment reward functions. Alternatively, Diffusion-DPO [45] utilize direct preference optimization (DPO) [32] to directly optimize for human preferences using an offline ranking dataset. SD3 [7] shows that fine-tuning diffusion models with human feedback techniques can significantly enhance overall image quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.909, + 0.593 + ], + "angle": 0, + "content": "Despite these promising results, Diffusion-DPO aligns a model with a single reward that represents population-level human preferences. These fine-tuning techniques are thus insufficient for achieving personalized preference alignment. As studied in domains such as language [10, 36, 44], individual users have unique preferences—for instance, some may favor images with brighter colors, while others might prefer images with centralized foregrounds. Finetuning a separate model for each person based on their preference data is impractical and does not scale well, limiting the feasibility of personalized alignment. Additionally, this may limit cross-user generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.716 + ], + "angle": 0, + "content": "A complementary research direction introduces additional control mechanisms into diffusion models to enable customization and personalization. Approaches such as IP-Adapter [51] and CtrlNet [54] allow for controlled image generation by using an additional conditioning image that encodes personalization features. However, these methods are constrained by their reliance on single-image inputs and do not directly learn or adapt to human preferences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.903 + ], + "angle": 0, + "content": "To address these limitations, we propose PPD, a general framework that incorporates personalization into a diffusion model preference alignment through multi-reward optimization. Our approach learns a diverse set of human preferences within a single model and generalizes effectively to new users. Personalization in our approach requires minimal additional data, as we label samples in the preference dataset with a unique identifier for each user. This annotation is inexpensive to obtain and commonly included in datasets such as Pick-a-Pic [16]. We leverage vision-language models (VLMs) such as LLaVA-OneVision [20] to extract user preference embeddings from" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.876, + 0.371, + 0.901 + ], + "angle": 0, + "content": "*Equal contributions. \nWork done during MD's internship at Luma AI." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8020" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "few-shot pairwise preference examples for each user. We show that these few-shot text embeddings from preference pairs serve as effective features to represent reward functions under the Bradley-Terry model [3] and thus are suitable for conditioning the diffusion model. Next, adopting Stable Cascade [27] as our base model, we condition the diffusion model on these user embeddings through additional cross-attention layers and fine-tune it with a variant of the Diffusion-DPO [45] objective for preference alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.482, + 0.41 + ], + "angle": 0, + "content": "Our empirical evaluations demonstrate that PPD can effectively optimize multiple rewards including CLIP [13, 31], Aesthetic [37], and HPS [47], significantly outperforming baselines. It also interpolates smoothly between different rewards during inference. Additionally, in real-world user scenarios, PPD aligns more closely with individual user preferences on the dataset Pick-a-Pic [16]. Notably, with only a few-shot ranking examples for a new user, PPD generate images that align closely with that user's specific preference and outperforms prior approaches such as Stable Cascade, achieving an average win rate of \\(76\\%\\), as evaluated by GPT4o [26]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.421, + 0.233, + 0.436 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Aligning Diffusion Models Aligning text-to-image diffusion models has been extensively studied. Some methods optimize with online RL using a Bradley Terry reward function [2, 9, 55]. Others fine-tune diffusion models to increase the reward of generated images by gradient backpropagation on the reward function [5, 30, 49]. Recently proposed approaches [45, 50] leverage the duality of the reward learning and policy optimization from DPO [32] to eliminate the need of on-policy samples and only utilize offline human preference data. Alternatively, DiffusionKTO [22] replaces DPO with KTO [8], eliminating the need for paired data. Additionally, some methods modify the training data distribution on visually appealing and textually cohered data [6, 18, 48], while others improve text accuracy by re-captioning pre-collected web images [11, 39]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.483, + 0.83 + ], + "angle": 0, + "content": "Controllable Generation for Diffusion Models To address the lack of fine-grained control in image synthesis, various methods add conditional controls to text-to-image diffusion models. CtrlNet [54] and T2I-Adapter [25] introduce additional UNet modules that encode spatial information, such as edges and human poses, derived from conditioning images. IP-Adapter [51] introduces additional cross-attention layers to incorporate high-level semantic information from reference images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "LLM Alignment The alignment of large language models (LLMs) using RLHF objectives has been extensively explored [1, 32, 56], demonstrating effective instruction-following and alignment to population values. However," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "a key challenge remains in representing the diversity of human preferences [36], as these models often capture an averaged population preference rather than accounting for individual differences, leading to biases and reduced alignment with underrepresented communities. Pluralistic Alignment [10, 44] addresses this issue by modeling a distribution of individuals, each with distinct values and perspectives, with methods like distributional and overton alignment enhancing calibration across diverse user groups. Several works have investigated LLM personalization through the reparameterization of reward models [23, 28, 52], enabling a more tailored user experience. However, these works leverage synthetic users and study a limited number of users, particularly in text and robotic control domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.318, + 0.64, + 0.336 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.344, + 0.905, + 0.374 + ], + "angle": 0, + "content": "We introduce the preliminaries behind diffusion models, DPO and its adaption to diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.907, + 0.602 + ], + "angle": 0, + "content": "Diffusion Models Diffusion models are generative models consisting two processes: a forward (diffusion) process and a backward (denoising) process [12, 40, 42, 43]. In the forward process, Gaussian noise \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) is progressively added to a data distribution \\(P_{\\mathrm{data}}(\\mathbf{x}_0)\\) over \\(T\\) steps, guided by noise scheduling functions \\(\\alpha_{t}\\) and \\(\\sigma_t\\). At timestep \\(t\\in [0,T]\\), noisy images \\(\\mathbf{x}_t\\) are generated via \\(\\mathbf{x}_t = \\alpha_t\\mathbf{x}_0 + \\sigma_t\\epsilon\\). In the backward process, samples are generated from Gaussian noise with a learnable model, which can also be conditioned on additional inputs, such as text in the case of text-to-image diffusion models. Typically, diffusion models - denoted as \\(\\epsilon_{\\theta}\\) - are trained to predict the Gaussian noise \\(\\epsilon\\) from noisy images \\(\\mathbf{x}_t\\) at timestep \\(t\\) with conditioning \\(\\mathbf{c}\\), using a denoising objective:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.613, + 0.905, + 0.632 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {D M}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\omega (\\lambda_ {t}) \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.905, + 0.706 + ], + "angle": 0, + "content": "where \\(\\lambda_{t} = \\log \\left[\\alpha_{t}^{2} / \\sigma_{t}^{2}\\right]\\) is a signal-to-noise ratio, \\(\\omega (\\lambda_t)\\) is a pre-specified weighting function [14]. Once the model \\(\\epsilon_{\\theta}\\) is trained, images can be generated from random noise using accelerated sampling methods such as DDIM [41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Direct Preference Optimization (DPO) DPO [32] is a fine-tuning paradigm that aligns autoregressive large language models (LLMs) with human preferences without relying on traditional reinforcement learning techniques. Unlike Reinforcement Learning from Human Feedback (RLHF), which involves training a reward model and then fine-tuning the LLM using reinforcement learning, DPO directly optimizes the model based on human preference data. Given a dataset containing ranking pairs \\((\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-)\\) where \\(\\mathbf{c}\\) being the prompt and \\(\\mathbf{x}_0^+, \\mathbf{x}_0^-\\) being the preferred and dispreferred responses labeled by human, the training objective optimizes a simple loss under which we can obtain an" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8021" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.912, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.339, + 0.506, + 0.355 + ], + "angle": 0, + "content": "Stage 1: generate user profile with few-shot image pairs." + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.339, + 0.885, + 0.354 + ], + "angle": 0, + "content": "Stage 2: train personalized diffusion model." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.367, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Figure 1. The overall architecture of PPD. In Stage 1, user embedding are generated with few-shot preference examples using a VLM. In Stage 2, we fine-tune diffusion models on the preference datasets with the user embedding as conditioning added to cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.422, + 0.196, + 0.439 + ], + "angle": 0, + "content": "optimal policy:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.448, + 0.482, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\mathrm {D P O}} (\\theta) = - \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0} ^ {+}, \\mathbf {x} _ {0} ^ {-}} [ \\log \\sigma (\\beta \\Delta) ], \\text {w h e r e} \\\\ \\Delta = \\left(\\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0} ^ {+} \\mid \\mathbf {c}\\right)}{p _ {\\mathrm {r e f}} \\left(\\mathbf {x} _ {0} ^ {+} \\mid \\mathbf {c}\\right)} - \\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0} ^ {-} \\mid \\mathbf {c}\\right)}{p _ {\\mathrm {r e f}} \\left(\\mathbf {x} _ {0} ^ {-} \\mid \\mathbf {c}\\right)}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.592 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the sigmoid function and \\(\\beta\\) is a hyperparameter determining how much \\(p_{\\theta}(\\mathbf{x}_0|\\mathbf{c})\\) may deviate from a reference distribution \\(p_{\\mathrm{ref}}(\\mathbf{x}_0|\\mathbf{c})\\). Intuitively, the objective aims to increase the likelihoods of preferred responses while decreasing that of dispreferred ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.727 + ], + "angle": 0, + "content": "Diffusion DPO The DPO objective in Eq. (2) requires optimization of log-likelihoods \\(\\log p_{\\theta}(\\mathbf{x}_0^+|\\mathbf{c})\\), which is analytically intractable for diffusion models. Diffusion-DPO [45] leverages the fact that the (appropriately weighted) training objective of diffusion models as in Eq. (1) can be viewed as the evidence lower bound (ELBO) of their log-likelihoods. It derives an upper bound to DPO for diffusion models to align them with human preferences:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.737, + 0.478, + 0.758 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {D i f f - D P O}} (\\theta) = - \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0} ^ {+}, \\mathbf {x} _ {0} ^ {-}} [ \\log \\sigma (- \\beta T \\omega (\\lambda_ {t}) \\Delta) ], \\text {w h e r e}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.759, + 0.482, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta = \\| \\epsilon^ {+} - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, t) \\| _ {2} ^ {2} - \\| \\epsilon^ {+} - \\epsilon_ {\\mathrm {r e f}} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\\\ - \\left(\\| \\epsilon^ {-} - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, t) \\| _ {2} ^ {2} - \\| \\epsilon^ {-} - \\epsilon_ {\\text {r e f}} (\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, t) \\| _ {2} ^ {2}\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}\\) is the caption; \\(\\mathbf{x}_0^+\\) and \\(\\mathbf{x}_0^-\\) are the preferred and dispreferred images; \\(\\mathbf{x}_t^+ = \\alpha_t\\mathbf{x}_0^+ +\\sigma_t\\epsilon^+\\) and \\(\\mathbf{x}_t^- = \\alpha_t\\mathbf{x}_0^- +\\sigma_t\\epsilon^-\\) are noisy images obtained from forward process with Gaussian noise \\(\\epsilon^{+}\\) and \\(\\epsilon^{-}\\) at timestep \\(t\\) . This objective computes 4 terms as in Eq. (1): the denoising loss of training and preference distributions \\(\\epsilon_{\\theta}\\) and \\(\\epsilon_{\\mathrm{ref}}\\) on preferred and dispreferred" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.907, + 0.467 + ], + "angle": 0, + "content": "noisy images \\(\\mathbf{x}_t^+\\) and \\(\\mathbf{x}_t^-\\) respectively. Intuitively, it aims to decrease the loss of preferred images while increasing that of dispreferred ones." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.481, + 0.605, + 0.496 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.673 + ], + "angle": 0, + "content": "Despite impressive image quality and alignment control achieved in prior works, a key limitation among works that use DPO for diffusion models is that it assumes a uniform preference among all users for generated images. In practice, this assumption often fails, as individuals may have diverse preferences related to style, color, lighting, and other visual attributes. While such preferences can be easily expressed by human through pairwise comparisons, it is challenging to summarize them accurately through text or single-image prompts, due to the inaccuracy of natural language and the limitations inherent in single-prompt formats." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Motivated by this problem, we propose PPD, a method that aligns diffusion models with individual-level human preferences. We first formulate the problem of personalization and define a new DPO variant for this task (Sec. 4.1), and then show that individual-level preference learning can be achieved using few-shot comparison examples using a VLM (Sec. 4.2), finally we introduce a parameterization of denoiser that can takes advantage of the VLM embeddings (Sec. 4.3). The overall architecture is in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.666, + 0.835 + ], + "angle": 0, + "content": "4.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.902 + ], + "angle": 0, + "content": "We consider a dataset containing \\((\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-, \\mathbf{u})\\) where \\(\\mathbf{c}\\) is the text prompt, \\(\\mathbf{x}_0^+\\), \\(\\mathbf{x}_0^-\\) are preferred and dispreferred generated images and \\(\\mathbf{u}\\) represents the user labeling the preference. User features \\(\\mathbf{u}\\) can be high-dimensional including" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8022" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "information like age or gender, but often \\(\\mathbf{u}\\) consists only of scalar user IDs \\(\\mathbf{u} = [u]\\). We define \\(r(\\mathbf{c},\\mathbf{x}_0,\\mathbf{u})\\) as the reward on image \\(\\mathbf{x}_0\\) for user \\(\\mathbf{u}\\) given prompt \\(\\mathbf{c}\\). We would like to fine-tune a text-to-image models \\(p_{\\theta_i}(\\mathbf{x}_0|\\mathbf{c})\\) for individual user \\(\\mathbf{u}_i\\) such that reward is maximized for this user while keeping close to a reference model \\(p_{\\mathrm{ref}}(\\mathbf{x}_0|\\mathbf{c})\\) in terms of KL-divergence as regularization:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.206, + 0.483, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {p _ {\\theta_ {i}}} \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0}} \\left[ r (\\mathbf {c}, \\mathbf {x} _ {0}, \\mathbf {u} _ {i}) \\right] \\tag {4} \\\\ - \\beta \\mathbb {D} _ {\\mathrm {K L}} \\left[ p _ {\\theta_ {i}} (\\mathbf {x} _ {0} | \\mathbf {c}) \\| p _ {\\text {r e f}} (\\mathbf {x} _ {0} | \\mathbf {c}) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.485, + 0.306 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a parameter controlling how much \\(p_{\\theta_i}(\\mathbf{x}_0|\\mathbf{c})\\) deviates from \\(p_{\\mathrm{ref}}(\\mathbf{x}_0|\\mathbf{c})\\). Following [32], the optimal \\(p_{\\theta_i}\\) for user \\(u_{i}\\) takes the form" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.315, + 0.483, + 0.332 + ], + "angle": 0, + "content": "\\[\np _ {\\theta_ {i}} \\left(\\mathbf {x} _ {0} | \\mathbf {c}\\right) \\propto p _ {\\operatorname {r e f}} \\left(\\mathbf {x} _ {0} | \\mathbf {c}\\right) \\exp \\left(r \\left(\\mathbf {c}, \\mathbf {x} _ {0}, \\mathbf {u} _ {i}\\right) / \\beta\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.416 + ], + "angle": 0, + "content": "Suppose model \\( p_{\\theta} \\) has sufficient capacity, we can represent \\( p_{\\theta_i}(\\mathbf{x}_0|\\mathbf{c}) \\) as \\( p_{\\theta}(\\mathbf{x}_0|\\mathbf{c},\\mathbf{u}_i) \\), where \\( \\mathbf{u}_i \\) serves as a conditioning. This allows us to jointly optimize a single model \\( p_{\\theta}(\\mathbf{x}_0|\\mathbf{c},\\mathbf{u}) \\) for all users. Following [45], we have the following objective to learn a personalized diffusion models:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.424, + 0.482, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\mathrm {P P D}} (\\theta) = - \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0} ^ {+}, \\mathbf {x} _ {0} ^ {-}, \\mathbf {u}} [ \\log \\sigma (- \\beta T \\omega (\\lambda_ {t}) \\Delta) ], \\text {w h e r e} \\\\ \\Delta = \\| \\epsilon^ {+} - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, \\mathbf {u}, t) \\| _ {2} ^ {2} - \\| \\epsilon^ {+} - \\epsilon_ {\\mathrm {r e f}} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\\\ - \\left(\\| \\epsilon^ {-} - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, \\mathbf {u}, t\\right) \\| _ {2} ^ {2} - \\| \\epsilon^ {-} - \\epsilon_ {\\text {r e f}} \\left(\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, t\\right) \\| _ {2} ^ {2}\\right). \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.539 + ], + "angle": 0, + "content": "This objective is similar to Eq. (3) but with the addition of \\(\\mathbf{u}\\), representing extra features conditioned on for each user. But what user features should be conditioned on?" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.579 + ], + "angle": 0, + "content": "4.2. Generating User Features from a VLM with In-Context Examples" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Discussion on User Features in Preference Datasets One fundamental challenge in personalized image generation lies in the effective parameterization of the user representation, denoted as \\(\\mathbf{u}\\), for each individual. This parameterization is critical because it directly influences the model's capacity to generate outputs tailored to specific user preferences. Existing image preference datasets, such as Pick-a-Pic [16], typically lack comprehensive user information beyond basic identifiers, often focusing on generic preference data rather than capturing the nuanced individual preferences necessary for effective personalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Even in preference datasets that include user information (e.g PRISM [15]), this information is frequently obtained through self-reported surveys, which can lead to incomplete or biased user profiles. Discrepancies often arise between stated preferences and actual behavior; users may not accurately articulate their likes and dislikes, or they may respond in a manner that they perceive as socially desirable rather than truthful. Such inconsistencies present significant challenges when attempting to model and predict user preferences accurately." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.09, + 0.885, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.268, + 0.908, + 0.326 + ], + "angle": 0, + "content": "Figure 2. Top-K accuracy of the User Classification. We fine-tune a user-classifier from the frozen embeddings from the VLM on few-shot preference examples for 300 users. This classifier significantly outperforms a random chance baseline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.437 + ], + "angle": 0, + "content": "To address these issues, we adopt a simplified approach by assuming access solely to a unique user ID for each individual in the dataset. This user ID serves as a key to differentiate users within the labeled preference dataset, represented as \\(\\mathcal{D}_{\\mathrm{pref}} = \\{(\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-, \\mathbf{s})^{(i)}\\}\\), where \\(\\mathbf{s}\\) denotes a unique user identifier." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.659 + ], + "angle": 0, + "content": "Conditioning on a Unique User Identifier The primary challenge, then, is to leverage this user ID \\(i\\) to construct a meaningful user embedding that can effectively guide the model's predictions. A naive approach might involve representing the embedding \\(\\mathbf{u}\\) using a one-hot encoding of the user IDs, which can be formally defined by a function \\(f:\\mathbf{S}\\mapsto U\\), establishing a bijective mapping from the set of user IDs \\(\\mathbf{s}\\in \\{1,2,\\ldots ,|D|\\}\\) to the space of one-hot encoded vectors \\(\\mathbf{u}_i\\in \\{0,1\\}^{|D|}\\), where \\(|D|\\) represents the total number of unique users in the dataset. While this method uniquely encodes each user, it suffers from a critical limitation: the inability to generalize to new, unseen users, as the embedding space is inherently tied to the users present in the training data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.81 + ], + "angle": 0, + "content": "To enhance generalization and adaptability, we reconsider the parameterization of \\( u_{i} \\) in the context of learning the distribution of reward functions, denoted as \\( r(\\mathbf{c},\\mathbf{x}_0,\\mathbf{u}_i) \\). Here, a more flexible and robust parameterization leverages few-shot examples that reflect an individual user's preferences. These examples can be extracted from the offline preference dataset for each user, providing concrete instances of preferred and dispreferred content. Conditioning on these examples allows for a more accurate estimation of the reward function \\( r(\\mathbf{c},\\mathbf{x}_0,\\mathbf{u}_i) \\) for each user." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this work, we utilize the features of a pre-trained vision-language Model (VLM), specifically LLaVA-OneVision [21], an open-source multimodal model with multi-image processing capabilities. For each user, a set of \\(N = 4\\) few-shot examples is selected, comprising a caption describing the context, a preferred image, and a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8023" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "dispreferred image. These examples are processed by the VLM, from which we extract an intermediate hidden state." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.121, + 0.483, + 0.258 + ], + "angle": 0, + "content": "To validate that the user embedding is effective, we condition on the frozen embedding to classify which user provided the preference annotations. As seen in Figure 2, the classifier significantly outperforms chance with a top-16 accuracy of \\(90\\%\\), indicating that the embedding can disentangle different users in the preference dataset effectively. By incorporating this user-specific embedding into the conditioning of a text-to-image diffusion model, we enable user-aware reward modeling and generation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.267, + 0.373, + 0.284 + ], + "angle": 0, + "content": "4.3. Personalization as Conditioning" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.289, + 0.483, + 0.424 + ], + "angle": 0, + "content": "In order to add the user embedding as conditioning to the diffusion model in image generation, we utilize additional cross-attention layers akin to IP-Adapter [51]. Here we learn to condition on the user information using the VLM embeddings and then integrate the image features into pretrained UNet by adapted modules with decoupled cross-attention. There are many other ways to add conditioning information to diffusion models; we chose this instantiation due to its training and parameter efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.61 + ], + "angle": 0, + "content": "We utilize Stable Cascade [27] as our base model which comprises of three stages: Stages A-C. Stage A and B are image compressors, similar to the VAE in Stable Diffusion [34] but achieve a much higher compression rate; and Stage C is the generation module conditioned on text. In Stage C, the text features c from pretrained CLIP text encoder are plugged into the model by feeding into the cross-attention layers. Given the query features \\(\\mathbf{Z}\\) and text features \\(\\mathbf{c}_t\\), the output is \\(\\mathbf{Z}' = \\text{Attention}(\\mathbf{Q}, \\mathbf{K}, \\mathbf{V}) = \\text{Softmax}(\\frac{\\mathbf{Q}\\mathbf{K}^T}{\\sqrt{d}})\\mathbf{V}\\), where \\(\\mathbf{Q} = \\mathbf{Z}\\mathbf{W}_q\\), \\(\\mathbf{K} = \\mathbf{c}_t\\mathbf{W}_k\\), \\(\\mathbf{V} = \\mathbf{c}_t\\mathbf{W}_v\\) are query, key and values, and \\(\\mathbf{W}_q, \\mathbf{W}_k, \\mathbf{W}_v\\) are trainable layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.731 + ], + "angle": 0, + "content": "Following IP-Adapter [51], we utilize a decoupled crossattention mechanism to condition on the user embedding. First, we add a new cross-attention layer to each cross-attention layer in the Stage C model, using the original query \\(\\mathbf{Q}\\) and introducing new keys \\(\\mathbf{K}'\\) and values \\(\\mathbf{V}'\\) for user features. Then, we simply add the output of user cross-attention to the output of text cross-attention. Hence the final formulation of cross-attention is" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.751, + 0.483, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} ^ {\\prime} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V} + \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\left(\\mathbf {K} ^ {\\prime}\\right) ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V} ^ {\\prime}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q} = \\mathbf{Z}\\mathbf{W}_q\\), \\(\\mathbf{K} = \\mathbf{c}_t\\mathbf{W}_k\\), \\(\\mathbf{V} = \\mathbf{c}_t\\mathbf{W}_v\\), \\(\\mathbf{K}' = \\mathbf{u}_t\\mathbf{W}_k'\\), \\(\\mathbf{V}' = \\mathbf{u}_t\\mathbf{W}_v'\\) and \\(\\mathbf{W}_k'\\), \\(\\mathbf{W}_v'\\) are new introduced trainable parameters for \\(\\mathbf{u}\\). During training, we optimize only the added cross-attention layers while keeping the parameters of pretrained diffusion model frozen, using the training objective as in Eq. (6). We also randomly drop user features (zero out the user embedding) in the training stage as regularization." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.91, + 0.233 + ], + "angle": 0, + "content": "
CLIPAestheticHPS
Stable Cascade31.975.3323.87
Diffusion-DPO32.485.4625.96
SFT32.265.5625.78
ours32.665.9227.51
Diffusion-DPO (CLIP)32.96--
Diffusion-DPO (Aesthetic)-6.42-
Diffusion-DPO (HPS)--28.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.243, + 0.908, + 0.286 + ], + "angle": 0, + "content": "Table 1. Averaged scores on three reward function. We condition on each reward to generate images both in SFT and our methods. Diffusion-DPO (*) are upper-bounds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.302, + 0.908, + 0.334 + ], + "angle": 0, + "content": "This allows us to also prompt the model unconditionally of the user, by passing a zero embedding for the user." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.346, + 0.646, + 0.364 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.524 + ], + "angle": 0, + "content": "Models and Dataset In this section, we demonstrate the performance of PPD across a range of experiments designed to assess its capabilities in personalized preference alignment. We use the objective from Eq. (6) to fine-tune the open-source diffusion model, Stable Cascade [27] – specifically Stage C, which is responsible for text-conditioned generation. Following Wallace et al. [45], we fine-tune our model on the Pick-a-Pic v2 dataset [16], a preference dataset containing 58K text prompts and 0.8M image pairs labeled by 5K users." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.906, + 0.636 + ], + "angle": 0, + "content": "Training Details During training of PPD, we only optimize the added cross-attention layers while keeping the pretrained diffusion model fixed, as described in Sec. 4.3. We use AdamW [24] with an effective batch size of 768 pairs and a learning rate of \\(1 \\times 10^{-5}\\) for all experiments, training for one epoch. We find \\(\\beta \\in [0.1, 2]\\) for hyperparameter tuning. Details are in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.906, + 0.764 + ], + "angle": 0, + "content": "User Information Conditioning We begin by demonstrating the model's ability to condition on synthetic user information, optimizing three distinct reward functions, we further show that generalize by linearly interpolating the weights of different reward functions (Sec. 5.1). We then show that the model can also condition on real user preferences, generating personalized images based on few-shot examples from new users (Sec. 5.2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.749, + 0.79 + ], + "angle": 0, + "content": "5.1. Synthetic User Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Synthetic User We utilize three reward functions, each representing an aspect of image-to-text generation quality, to simulate three distinct synthetic users: (i) CLIP (Open-CLIP ViT-H/14) [13, 31] for prompt image alignment, (ii) Aesthetic Predictor [37] for image visual appeal, and (iii) HPSv2 (Human Preference Score v2) [47] for human preference estimate. We relabel pairs \\((\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-, \\mathbf{u})\\) based" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8024" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.09, + 0.373, + 0.103 + ], + "angle": 0, + "content": "Prompt Alignment (CLIP)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.104, + 0.481, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.195, + 0.35, + 0.206 + ], + "angle": 0, + "content": "Aesthetic Quality" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.206, + 0.48, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.298, + 0.385, + 0.31 + ], + "angle": 0, + "content": "Human Preference Score (HPS)" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.31, + 0.48, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.41, + 0.482, + 0.423 + ], + "angle": 0, + "content": "Figure 3. Automatic win rate evaluation with reward functions." + }, + { + "type": "image_footnote", + "bbox": [ + 0.091, + 0.424, + 0.466, + 0.438 + ], + "angle": 0, + "content": "We compare against Stable Cascade, Diffusion-DPO, and SFT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.483, + 0.499 + ], + "angle": 0, + "content": "on the above reward functions, where \\(\\mathbf{u}\\) is a one-hot vector indicating the reward function, taking one of 3 values: \\(\\mathbf{u} \\in \\{[1,0,0],[0,1,0],[0,0,1]\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.692 + ], + "angle": 0, + "content": "Baseline We compare our method with: (1) Stable Cascade [27] pretrained model, which is also the reference model during training; (2) Diffusion-DPO, fine-tuned on Pick-a-Pic label without user conditioning [45]; (3) SFT, supervised fine-tuning on relabeled data with user conditioning. We also compare with (4) Diffusion-DPO (CLIP/Aesthetic/HPS), each annotation indicating a separately fine-tuned model with corresponding reward, this serves as the upper-bound of our method. For a fair comparison, all baselines are fine-tuned using Stable Cascade with the same architecture as our model, optimizing only the added cross-attention layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Text-to-Image Generation For each method, we generate images with 30 inference steps using captions from PartiPrompts [53], a benchmark containing 1.6K examples across various categories and challenge aspects. For SFT and our method, we generate three images per prompt, each conditioned on one reward function. For Diffusion-DPO (CLIP/Aesthetic/HPS), images for each annotation are generated from its corresponding model. We report averaged reward scores in Tab. 1 and win rate comparisons in Fig. 3, we can see that our method significantly outperforms Stable Cascade in different reward functions, achieving win rates of \\(60\\%\\), \\(90\\%\\), and \\(93\\%\\) in terms of CLIP, Aesthetic, and HPS respectively. Our method effectively optimizes each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "reward by conditioning on individual preference, outperforming supervised fine-tuning by a large margin. Unlike Diffusion-DPO, our method can jointly optimize multiple user rewards, achieving win rates comparable to Diffusion-DPO models fine-tuned separately on each user's data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.18, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Generalization via Interpolation During training we fine-tune the model using one-hot encoding of each objective. Specifically, the user embedding vector \\(\\mathbf{u}\\) being \\([1,0,0],[0,1,0],[0,0,1]\\) corresponding to optimization for CLIP, Aesthetic, and HPS. The user embedding can also be seen as weighting mechanism for each objective. However, during inference, we are also interested in whether we can trade-off among different objectives by interpolating their respective weights. In Fig. 4, we generate images conditioned on various objective weightings, namely interpolating the given user embedding, we can see that as the weight of an objective increases, the corresponding score also rises, and vice versa. Therefore, our method can generalize to different reward combinations, even only trained on single objective data. This can not be achieved by optimizing separate models and each with an objective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.906, + 0.497 + ], + "angle": 0, + "content": "This shows that by linearly interpolating the user embedding, we can adjust the rewards of the generated images accordingly. This finding aligns with the real user experiments presented next, as real user embeddings exhibit a linear correlation in embedding space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.507, + 0.813, + 0.523 + ], + "angle": 0, + "content": "5.2. Personalized Real User Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.906, + 0.62 + ], + "angle": 0, + "content": "We now consider a realistic setting with preferences from the Pick-a-Pic dataset for each user. In this setting, we utilize the user features which are generated from a VLM on in-context preference examples as outlined in Section 4.2. We scale from the 3 synthetic \"users\" in Section 5.1 to a filtered subset of the 4800 unique users in Pick-a-Pic." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.906, + 0.828 + ], + "angle": 0, + "content": "Experiment setting We repartition the Pick-a-Pic dataset into a set of 4 splits: (1) Train, (2) Validation, (3) Test with Seen Users + Held-out Captions and (4) Test with Unseen Users + Held-out Captions. This decomposition allows us to study independently the generalization of our methods to new prompts and unseen users. Additionally, for each user, we take sets of \\( N = 4 \\) few-shot preference examples for each user and use these groupings for the conditioning of the VLM on few-shot preference pairs. We thus have to filter out 824 users in Pick-a-Pic which have less than 4 examples per user, as no user embedding for this user can be formed. We generate the user embedding through the procedure outlined in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Generation Results We perform qualitative analysis on the set of images generated for each user. We perform this evaluation by generating a user profile per user. We use a set of few-shot preference pairs per user which we condition a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8025" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.905, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.377, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Figure 4. PPD is able to interpolate among three distinct objectives during inference. (a) generated images conditioned on (b) various weights, with three axes representing CLIP, Aesthetic, and HPS; (c) reward scores for each image. The score for each objective increases as its respective weight increases, and decreasing otherwise. For each row from left to right, the CLIP score increases alongside its weight, leading to a decrease in the Aesthetic score as its weight decreases. From bottom to top, the HPS score increases with its weight." + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.458, + 0.392, + 0.467 + ], + "angle": 0, + "content": "Comparison of Win Rates for Real User Alignment" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.47, + 0.229, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.471, + 0.356, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.471, + 0.482, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Figure 5. Automatic Win Rate Evaluation with GPT-4o We evaluate on seen users, held out users, and aggregate to see win-rates for generic users, comparing Stable Cascade and PPD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.837 + ], + "angle": 0, + "content": "VLM on along with Chain-of-Thought prompting (COT) to construct a user-profile consistent with the preference pairs. The COT prompt asks the model to summarize each image in the preference pair and describe the differences between them, which is then aggregated to form a user profile. Sample user profiles and the full Chain-of-Thought prompt is found in Appendix B. As seen in Fig. 6, our method has closer alignment to the user profile with respect to Stable Cascade and Diffusion-DPO, being able to pick up details about the user such as the highlighted part. While Stable Cascade and Diffusion-DPO with augmented captions can capture certain aspects of the user description, they often overfit to the user profile, resulting in generated images that overlook the intended image caption." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Personalized Preference Alignment Evaluation We evaluate whether the generated model aligns with specific user preference using a VLM as a judge, leveraging strong" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.626 + ], + "angle": 0, + "content": "VLMs such as GPT4o-vision [26], similar to recipes in prior work [4, 19]. In particular, we condition the VLM on few-shot examples from the user as well as a pair of images for a given test time caption and ask the model to judge which of the two images would be preferred by the user. To ensure consistency of the response the model selects, we judge each pair of images twice using the VLM, flipping the ordering of the images, filtering out pairs where the VLM is inconsistent. With this paradigm and 4 shots per user, we find that the VLM is able to match the ground truth personalized human preference with \\(83\\%\\) accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.825 + ], + "angle": 0, + "content": "We decompose our evaluation into two settings: (1) evaluating users that were seen during training but on captions that were held out and (2) evaluating held-out users and captions. We aggregate over these settings to also compute metrics for the general user in the dataset. For the 4341 users in Pick-a-Pic we select 192 held out and 277 in-distribution users. For each of these users, we choose \\(4 + 1\\) examples, where 4 examples are used for the in-context conditioning for the user and 1 example is used as the caption for evaluation. We compute an automatic win-rate using pairs of models, generating one image for each caption and evaluating with the VLM-as-Judge pipeline outlined above." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We report win rate comparisons in Fig. 5. Here, PPD significantly outperforms Stable Cascade in all three evaluation settings, achieving a win rate of \\(83\\%\\) on seen users, and \\(78\\%\\) on unseen users, aggregated to \\(81\\%\\) on all users." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8026" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.905, + 0.692 + ], + "angle": 0, + "content": "
User ProfileStable CascadeDiffusion-DPODiffusion-DPOOurs
Append user profile to caption as augmentation
The preferred images share common themes of wood grain, classic or futuristic design elements, and a focus on aesthetics. They often feature unique front grille patterns, large windows, and a mix of traditional and modern design features. ...A needle-felted palm tree
This user prefers images that have a personal and intimate feel with a focus on the subject's appearance and expression. They also appreciate images with soft focus backgrounds and warm lighting that highlight texture and detail. ...An anime girl, masterpiece, good line art, trending in pixiv
The preferred images ... such as a softer color palette, more dynamic lighting, and a balanced composition that enhances the visual appeal of the scene. The characters are often positioned off-center, which adds to the dynamism of the composition. The backgrounds are rich in detail. ...Aelita from Code Lyoko in the forest, gathering flowers
The user is drawn to images that blend elements of reality with fantastical elements, creating a sense of wonder and curiosity. ... The user may be interested in fantasy art, animation, and storytelling through visual media.A cat on a propaganda poster
" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.702, + 0.907, + 0.745 + ], + "angle": 0, + "content": "Figure 6. Qualitative Analysis of Images Generated by PPD and Baselines. Compared to Diffusion-DPO, PPD achieves closer alignment with the generated user profile, highlighted in green. The caption-augmented method captures user profile details; however, it often leads to unintended image alterations, causing the image to disregard the caption itself, as indicated in red." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.452, + 0.788 + ], + "angle": 0, + "content": "6. Discussion, Conclusion, and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We propose a framework for personalizing text-to-image diffusion models using user feature embeddings from a VLM. By conditioning on these embeddings via decoupled cross-attention and fine-tuning with a DPO objective, our method enables accurate reward modeling and high-quality personalized image generation. It generalizes to unseen users with only a few preference examples at test time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.771, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Several open questions and limitations remain. Our current approach that conditioning on few-shot examples is caption-agnostic, selecting random user preferences; future work could explore active preference selection to improve personalization. Extending this method to video generation and studying the trade-off between implicit vs. stated preferences are also promising directions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8027" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.482, + 0.192 + ], + "angle": 0, + "content": "This research is done during MD's internship at Luma AI and supported in part by ARO (W911NF-21-1-0125), ONR (N00014-23-1-2159), and the CZ Biohub. AS gratefully acknowledges the support of the NSF Graduate Research Fellowship Program." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.208, + 0.187, + 0.224 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[1] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. Training a helpful and harmless assistant with reinforcement learning from human feedback, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[3] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[4] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Qinglan Huang, Canyu Chen, Qinghao Ye, Zhihong Zhu, Yuqing Zhang, Jiawei Zhou, Zhuokai Zhao, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. MJ-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation? arXiv e-prints, art. arXiv:2407.04842, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[5] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[6] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.482, + 0.841 + ], + "angle": 0, + "content": "[8] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[9] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.233, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "models. In Advances in Neural Information Processing Systems, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[10] Shangbin Feng, Taylor Sorensen, Yuhan Liu, Jillian Fisher, Chan Young Park, Yejin Choi, and Yulia Tsvetkov. Modular Pluralism: Pluralistic Alignment via Multi-LLM Collaboration. arXiv e-prints, art. arXiv:2406.15951, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[11] Gabriel Goh, James Betker, Li Jing, Aditya Ramesh, Tim Brooks, Jianfeng Wang, Lindsey Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Prafulla Dhariwal, Casey Chu, Joy Jiao, Jong Wook Kim, Alex Nichol, Yang Song, Lijuan Wang, and Tao Xu. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.904, + 0.315 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in neural information processing systems, pages 6840-6851, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[13] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 2,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[14] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In Advances in Neural Information Processing Systems 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[15] Hannah Rose Kirk, Alexander Whitefield, Paul Röttger, Andrew Bean, Katerina Margatina, Juan Ciro, Rafael Mosquera, Max Bartolo, Adina Williams, He He, Bertie Vidgen, and Scott A. Hale. The PRISM Alignment Project: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models. arXiv e-prints, art. arXiv:2404.16019, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.553, + 0.904, + 0.607 + ], + "angle": 0, + "content": "[16] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569, 2023. 1, 2, 4, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[17] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[18] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[19] Seongyun Lee, Seungone Kim, Sue Hyun Park, Geewook Kim, and Minjoon Seo. Prometheus-Vision: Vision-Language Model as a Judge for Fine-Grained Evaluation. arXiv e-prints, art. arXiv:2401.06591, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. arXiv e-prints, art. arXiv:2408.03326, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8028" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[22] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. arXiv preprint arXiv:2404.04465, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[23] Xinyu Li, Zachary C Lipton, and Liu Leqi. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[25] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[26] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.386 + ], + "angle": 0, + "content": "[27] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.387, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[28] Sriyash Poddar, Yanming Wan, Hamish Ivison, Abhishek Gupta, and Natasha Jaques. Personalizing reinforcement learning from human feedback with variational preference learning. arXiv preprint arXiv:2408.10075, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.567 + ], + "angle": 0, + "content": "[30] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[32] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, pages 36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[36] Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. Whose Opinions Do Language Models Reflect? arXiv e-prints, art. arXiv:2303.17548, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[37] Christoph Schuhmann. CLIP MLP aesthetic score predictor. https://github.com/christophschuhmann/improved-aesthetic-predictor, 2022. Accessed: 2024-11-06. 2,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[38] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278–25294, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[39] Eyal Segalis, Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. A picture is worth a thousand words: Principled recaptioning improves image generation. arXiv preprint arXiv:2310.16656, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[42] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.549, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[44] Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Ryting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, and Yejin Choi. A Roadmap to Pluralistic Alignment. arXiv e-prints, art. arXiv:2402.05070, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[45] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[47] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8029" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[48] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[49] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[50] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8941-8951, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[52] Chao Yu, Hong Lu, Jiaxuan Gao, Qixin Tan, Xinting Yang, Yu Wang, Yi Wu, and Eugene Vinitsky. Few-shot in-context preference learning using large language models. arXiv preprint arXiv:2410.17233, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[53] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.53, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.587, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[55] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. arXiv preprint arXiv:2401.12244, 4, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[56] Rui Zheng, Shihan Dou, Songyang Gao, Yuan Hua, Wei Shen, Binghai Wang, Yan Liu, Senjie Jin, Qin Liu, Yuhao Zhou, Limao Xiong, Lu Chen, Zhiheng Xi, Nuo Xu, Wenbin Lai, Minghao Zhu, Cheng Chang, Zhangyue Yin, Rongxiang Weng, Wensen Cheng, Haoran Huang, Tianxiang Sun, Hang Yan, Tao Gui, Qi Zhang, Xipeng Qiu, and Xuanjing Huang. Secrets of rlhf in large language models part i: Ppo, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8030" + } + ] +] \ No newline at end of file diff --git a/2025/Personalized Preference Fine-tuning of Diffusion Models/005a06ec-5d3e-419e-a758-dd73f721c5f6_origin.pdf b/2025/Personalized Preference Fine-tuning of Diffusion Models/005a06ec-5d3e-419e-a758-dd73f721c5f6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..199764aea95ec5fa24d6eef39e8b0ba93f9ecb3d --- /dev/null +++ b/2025/Personalized Preference Fine-tuning of Diffusion Models/005a06ec-5d3e-419e-a758-dd73f721c5f6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5616b10ed3ef3bb60c8bb7d609e33ea175922e57df2c923ea6fc2028c813dc6f +size 2431251 diff --git a/2025/Personalized Preference Fine-tuning of Diffusion Models/full.md b/2025/Personalized Preference Fine-tuning of Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7f9130ee3678c0df91da028b6c49732270ddef1c --- /dev/null +++ b/2025/Personalized Preference Fine-tuning of Diffusion Models/full.md @@ -0,0 +1,290 @@ +# Personalized Preference Fine-tuning of Diffusion Models + +Meihua Dang $^{1*}$ Anikait Singh $^{1*}$ Linqi Zhou $^{1,2}$ Stefano Ermon $^{1}$ Jiaming Song $^{2}$ $^{1}$ Stanford University, $^{2}$ Luma AI + +{mhdang,anikait,ermon}@cs.stanford.edu,{alexzhou,jiaming}@lumalabs.ai + +# Abstract + +RLHF techniques like DPO can significantly improve the generation quality of text-to-image diffusion models. However, these methods optimize for a single reward that aligns model generation with population-level preferences, neglecting the nuances of individual users' beliefs or values. This lack of personalization limits the efficacy of these models. To bridge this gap, we introduce PPD, a multi-reward optimization objective that aligns diffusion models with personalized preferences. With PPD, a diffusion model learns the individual preferences of a population of users in a few-shot way, enabling generalization to unseen users. Specifically, our approach (1) leverages a vision-language model (VLM) to extract personal preference embeddings from a small set of pairwise preference examples, and then (2) incorporates the embeddings into diffusion models through cross attention. Conditioning on user embeddings, the text-to-image models are fine-tuned with the DPO objective, simultaneously optimizing for alignment with the preferences of multiple users. Empirical results demonstrate that our method effectively optimizes for multiple reward functions and can interpolate between them during inference. In real-world user scenarios, with as few as four preference examples from a new user, our approach achieves an average win rate of $76\%$ over Stable Cascade, generating images that more accurately reflect specific user preferences. + +# 1. Introduction + +Diffusion models [12, 40, 42, 43] have achieved great success in text-to-images generation tasks, as demonstrated by models such as DALL-E 2 [33], Imagen [35], Stable Diffusion [29, 34], and Stable Cascade [27]. These large-scale models, often trained through supervised fine-tuning on text-image datasets containing high-quality aesthetic images such as LAION-5B [38], demonstrate impressive image quality and a remarkable visual understanding of text prompts. Recent advances focus on fine-tuning diffusion + +models to better align with human preferences. Methods such as DDPO [2] and DPOK [9] employ reinforcement learning from human feedback (RLHF) techniques to optimize human-alignment reward functions. Alternatively, Diffusion-DPO [45] utilize direct preference optimization (DPO) [32] to directly optimize for human preferences using an offline ranking dataset. SD3 [7] shows that fine-tuning diffusion models with human feedback techniques can significantly enhance overall image quality. + +Despite these promising results, Diffusion-DPO aligns a model with a single reward that represents population-level human preferences. These fine-tuning techniques are thus insufficient for achieving personalized preference alignment. As studied in domains such as language [10, 36, 44], individual users have unique preferences—for instance, some may favor images with brighter colors, while others might prefer images with centralized foregrounds. Finetuning a separate model for each person based on their preference data is impractical and does not scale well, limiting the feasibility of personalized alignment. Additionally, this may limit cross-user generalization. + +A complementary research direction introduces additional control mechanisms into diffusion models to enable customization and personalization. Approaches such as IP-Adapter [51] and CtrlNet [54] allow for controlled image generation by using an additional conditioning image that encodes personalization features. However, these methods are constrained by their reliance on single-image inputs and do not directly learn or adapt to human preferences. + +To address these limitations, we propose PPD, a general framework that incorporates personalization into a diffusion model preference alignment through multi-reward optimization. Our approach learns a diverse set of human preferences within a single model and generalizes effectively to new users. Personalization in our approach requires minimal additional data, as we label samples in the preference dataset with a unique identifier for each user. This annotation is inexpensive to obtain and commonly included in datasets such as Pick-a-Pic [16]. We leverage vision-language models (VLMs) such as LLaVA-OneVision [20] to extract user preference embeddings from + +few-shot pairwise preference examples for each user. We show that these few-shot text embeddings from preference pairs serve as effective features to represent reward functions under the Bradley-Terry model [3] and thus are suitable for conditioning the diffusion model. Next, adopting Stable Cascade [27] as our base model, we condition the diffusion model on these user embeddings through additional cross-attention layers and fine-tune it with a variant of the Diffusion-DPO [45] objective for preference alignment. + +Our empirical evaluations demonstrate that PPD can effectively optimize multiple rewards including CLIP [13, 31], Aesthetic [37], and HPS [47], significantly outperforming baselines. It also interpolates smoothly between different rewards during inference. Additionally, in real-world user scenarios, PPD aligns more closely with individual user preferences on the dataset Pick-a-Pic [16]. Notably, with only a few-shot ranking examples for a new user, PPD generate images that align closely with that user's specific preference and outperforms prior approaches such as Stable Cascade, achieving an average win rate of $76\%$ , as evaluated by GPT4o [26]. + +# 2. Related Work + +Aligning Diffusion Models Aligning text-to-image diffusion models has been extensively studied. Some methods optimize with online RL using a Bradley Terry reward function [2, 9, 55]. Others fine-tune diffusion models to increase the reward of generated images by gradient backpropagation on the reward function [5, 30, 49]. Recently proposed approaches [45, 50] leverage the duality of the reward learning and policy optimization from DPO [32] to eliminate the need of on-policy samples and only utilize offline human preference data. Alternatively, DiffusionKTO [22] replaces DPO with KTO [8], eliminating the need for paired data. Additionally, some methods modify the training data distribution on visually appealing and textually cohered data [6, 18, 48], while others improve text accuracy by re-captioning pre-collected web images [11, 39]. + +Controllable Generation for Diffusion Models To address the lack of fine-grained control in image synthesis, various methods add conditional controls to text-to-image diffusion models. CtrlNet [54] and T2I-Adapter [25] introduce additional UNet modules that encode spatial information, such as edges and human poses, derived from conditioning images. IP-Adapter [51] introduces additional cross-attention layers to incorporate high-level semantic information from reference images. + +LLM Alignment The alignment of large language models (LLMs) using RLHF objectives has been extensively explored [1, 32, 56], demonstrating effective instruction-following and alignment to population values. However, + +a key challenge remains in representing the diversity of human preferences [36], as these models often capture an averaged population preference rather than accounting for individual differences, leading to biases and reduced alignment with underrepresented communities. Pluralistic Alignment [10, 44] addresses this issue by modeling a distribution of individuals, each with distinct values and perspectives, with methods like distributional and overton alignment enhancing calibration across diverse user groups. Several works have investigated LLM personalization through the reparameterization of reward models [23, 28, 52], enabling a more tailored user experience. However, these works leverage synthetic users and study a limited number of users, particularly in text and robotic control domains. + +# 3. Preliminary + +We introduce the preliminaries behind diffusion models, DPO and its adaption to diffusion models. + +Diffusion Models Diffusion models are generative models consisting two processes: a forward (diffusion) process and a backward (denoising) process [12, 40, 42, 43]. In the forward process, Gaussian noise $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ is progressively added to a data distribution $P_{\mathrm{data}}(\mathbf{x}_0)$ over $T$ steps, guided by noise scheduling functions $\alpha_{t}$ and $\sigma_t$ . At timestep $t\in [0,T]$ , noisy images $\mathbf{x}_t$ are generated via $\mathbf{x}_t = \alpha_t\mathbf{x}_0 + \sigma_t\epsilon$ . In the backward process, samples are generated from Gaussian noise with a learnable model, which can also be conditioned on additional inputs, such as text in the case of text-to-image diffusion models. Typically, diffusion models - denoted as $\epsilon_{\theta}$ - are trained to predict the Gaussian noise $\epsilon$ from noisy images $\mathbf{x}_t$ at timestep $t$ with conditioning $\mathbf{c}$ , using a denoising objective: + +$$ +L _ {\mathrm {D M}} = \mathbb {E} _ {\mathbf {x} _ {0}, \epsilon , t} \left[ \omega (\lambda_ {t}) \| \epsilon - \epsilon_ {\theta} (\mathbf {x} _ {t}, \mathbf {c}, t) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $\lambda_{t} = \log \left[\alpha_{t}^{2} / \sigma_{t}^{2}\right]$ is a signal-to-noise ratio, $\omega (\lambda_t)$ is a pre-specified weighting function [14]. Once the model $\epsilon_{\theta}$ is trained, images can be generated from random noise using accelerated sampling methods such as DDIM [41]. + +Direct Preference Optimization (DPO) DPO [32] is a fine-tuning paradigm that aligns autoregressive large language models (LLMs) with human preferences without relying on traditional reinforcement learning techniques. Unlike Reinforcement Learning from Human Feedback (RLHF), which involves training a reward model and then fine-tuning the LLM using reinforcement learning, DPO directly optimizes the model based on human preference data. Given a dataset containing ranking pairs $(\mathbf{c},\mathbf{x}_0^+, \mathbf{x}_0^-)$ where $\mathbf{c}$ being the prompt and $\mathbf{x}_0^+, \mathbf{x}_0^-$ being the preferred and dispreferred responses labeled by human, the training objective optimizes a simple loss under which we can obtain an + +![](images/34d385018c2e606d6fb799da3bd83b159e7de042ab8269021764468ded12aeef.jpg) +Stage 1: generate user profile with few-shot image pairs. +Stage 2: train personalized diffusion model. +Figure 1. The overall architecture of PPD. In Stage 1, user embedding are generated with few-shot preference examples using a VLM. In Stage 2, we fine-tune diffusion models on the preference datasets with the user embedding as conditioning added to cross-attention. + +optimal policy: + +$$ +\begin{array}{l} L _ {\mathrm {D P O}} (\theta) = - \mathbb {E} _ {\mathbf {c}, \mathbf {x} _ {0} ^ {+}, \mathbf {x} _ {0} ^ {-}} [ \log \sigma (\beta \Delta) ], \text {w h e r e} \\ \Delta = \left(\log \frac {p _ {\theta} \left(\mathbf {x} _ {0} ^ {+} \mid \mathbf {c}\right)}{p _ {\mathrm {r e f}} \left(\mathbf {x} _ {0} ^ {+} \mid \mathbf {c}\right)} - \log \frac {p _ {\theta} \left(\mathbf {x} _ {0} ^ {-} \mid \mathbf {c}\right)}{p _ {\mathrm {r e f}} \left(\mathbf {x} _ {0} ^ {-} \mid \mathbf {c}\right)}\right), \tag {2} \\ \end{array} +$$ + +where $\sigma$ is the sigmoid function and $\beta$ is a hyperparameter determining how much $p_{\theta}(\mathbf{x}_0|\mathbf{c})$ may deviate from a reference distribution $p_{\mathrm{ref}}(\mathbf{x}_0|\mathbf{c})$ . Intuitively, the objective aims to increase the likelihoods of preferred responses while decreasing that of dispreferred ones. + +Diffusion DPO The DPO objective in Eq. (2) requires optimization of log-likelihoods $\log p_{\theta}(\mathbf{x}_0^+|\mathbf{c})$ , which is analytically intractable for diffusion models. Diffusion-DPO [45] leverages the fact that the (appropriately weighted) training objective of diffusion models as in Eq. (1) can be viewed as the evidence lower bound (ELBO) of their log-likelihoods. It derives an upper bound to DPO for diffusion models to align them with human preferences: + +$$ +L _ {\text {D i f f - D P O}} (\theta) = - \mathbb {E} _ {\mathbf {c}, \mathbf {x} _ {0} ^ {+}, \mathbf {x} _ {0} ^ {-}} [ \log \sigma (- \beta T \omega (\lambda_ {t}) \Delta) ], \text {w h e r e} +$$ + +$$ +\begin{array}{l} \Delta = \| \epsilon^ {+} - \epsilon_ {\theta} (\mathbf {x} _ {t} ^ {+}, \mathbf {c}, t) \| _ {2} ^ {2} - \| \epsilon^ {+} - \epsilon_ {\mathrm {r e f}} (\mathbf {x} _ {t} ^ {+}, \mathbf {c}, t) \| _ {2} ^ {2} \\ - \left(\| \epsilon^ {-} - \epsilon_ {\theta} (\mathbf {x} _ {t} ^ {-}, \mathbf {c}, t) \| _ {2} ^ {2} - \| \epsilon^ {-} - \epsilon_ {\text {r e f}} (\mathbf {x} _ {t} ^ {-}, \mathbf {c}, t) \| _ {2} ^ {2}\right), \tag {3} \\ \end{array} +$$ + +where $\mathbf{c}$ is the caption; $\mathbf{x}_0^+$ and $\mathbf{x}_0^-$ are the preferred and dispreferred images; $\mathbf{x}_t^+ = \alpha_t\mathbf{x}_0^+ +\sigma_t\epsilon^+$ and $\mathbf{x}_t^- = \alpha_t\mathbf{x}_0^- +\sigma_t\epsilon^-$ are noisy images obtained from forward process with Gaussian noise $\epsilon^{+}$ and $\epsilon^{-}$ at timestep $t$ . This objective computes 4 terms as in Eq. (1): the denoising loss of training and preference distributions $\epsilon_{\theta}$ and $\epsilon_{\mathrm{ref}}$ on preferred and dispreferred + +noisy images $\mathbf{x}_t^+$ and $\mathbf{x}_t^-$ respectively. Intuitively, it aims to decrease the loss of preferred images while increasing that of dispreferred ones. + +# 4. Method + +Despite impressive image quality and alignment control achieved in prior works, a key limitation among works that use DPO for diffusion models is that it assumes a uniform preference among all users for generated images. In practice, this assumption often fails, as individuals may have diverse preferences related to style, color, lighting, and other visual attributes. While such preferences can be easily expressed by human through pairwise comparisons, it is challenging to summarize them accurately through text or single-image prompts, due to the inaccuracy of natural language and the limitations inherent in single-prompt formats. + +Motivated by this problem, we propose PPD, a method that aligns diffusion models with individual-level human preferences. We first formulate the problem of personalization and define a new DPO variant for this task (Sec. 4.1), and then show that individual-level preference learning can be achieved using few-shot comparison examples using a VLM (Sec. 4.2), finally we introduce a parameterization of denoiser that can takes advantage of the VLM embeddings (Sec. 4.3). The overall architecture is in Fig. 1. + +# 4.1. Problem Setup + +We consider a dataset containing $(\mathbf{c},\mathbf{x}_0^+, \mathbf{x}_0^-, \mathbf{u})$ where $\mathbf{c}$ is the text prompt, $\mathbf{x}_0^+$ , $\mathbf{x}_0^-$ are preferred and dispreferred generated images and $\mathbf{u}$ represents the user labeling the preference. User features $\mathbf{u}$ can be high-dimensional including + +information like age or gender, but often $\mathbf{u}$ consists only of scalar user IDs $\mathbf{u} = [u]$ . We define $r(\mathbf{c},\mathbf{x}_0,\mathbf{u})$ as the reward on image $\mathbf{x}_0$ for user $\mathbf{u}$ given prompt $\mathbf{c}$ . We would like to fine-tune a text-to-image models $p_{\theta_i}(\mathbf{x}_0|\mathbf{c})$ for individual user $\mathbf{u}_i$ such that reward is maximized for this user while keeping close to a reference model $p_{\mathrm{ref}}(\mathbf{x}_0|\mathbf{c})$ in terms of KL-divergence as regularization: + +$$ +\begin{array}{l} \max _ {p _ {\theta_ {i}}} \mathbb {E} _ {\mathbf {c}, \mathbf {x} _ {0}} \left[ r (\mathbf {c}, \mathbf {x} _ {0}, \mathbf {u} _ {i}) \right] \tag {4} \\ - \beta \mathbb {D} _ {\mathrm {K L}} \left[ p _ {\theta_ {i}} (\mathbf {x} _ {0} | \mathbf {c}) \| p _ {\text {r e f}} (\mathbf {x} _ {0} | \mathbf {c}) \right], \\ \end{array} +$$ + +where $\beta$ is a parameter controlling how much $p_{\theta_i}(\mathbf{x}_0|\mathbf{c})$ deviates from $p_{\mathrm{ref}}(\mathbf{x}_0|\mathbf{c})$ . Following [32], the optimal $p_{\theta_i}$ for user $u_{i}$ takes the form + +$$ +p _ {\theta_ {i}} \left(\mathbf {x} _ {0} | \mathbf {c}\right) \propto p _ {\operatorname {r e f}} \left(\mathbf {x} _ {0} | \mathbf {c}\right) \exp \left(r \left(\mathbf {c}, \mathbf {x} _ {0}, \mathbf {u} _ {i}\right) / \beta\right). \tag {5} +$$ + +Suppose model $p_{\theta}$ has sufficient capacity, we can represent $p_{\theta_i}(\mathbf{x}_0|\mathbf{c})$ as $p_{\theta}(\mathbf{x}_0|\mathbf{c},\mathbf{u}_i)$ , where $\mathbf{u}_i$ serves as a conditioning. This allows us to jointly optimize a single model $p_{\theta}(\mathbf{x}_0|\mathbf{c},\mathbf{u})$ for all users. Following [45], we have the following objective to learn a personalized diffusion models: + +$$ +\begin{array}{l} L _ {\mathrm {P P D}} (\theta) = - \mathbb {E} _ {\mathbf {c}, \mathbf {x} _ {0} ^ {+}, \mathbf {x} _ {0} ^ {-}, \mathbf {u}} [ \log \sigma (- \beta T \omega (\lambda_ {t}) \Delta) ], \text {w h e r e} \\ \Delta = \| \epsilon^ {+} - \epsilon_ {\theta} (\mathbf {x} _ {t} ^ {+}, \mathbf {c}, \mathbf {u}, t) \| _ {2} ^ {2} - \| \epsilon^ {+} - \epsilon_ {\mathrm {r e f}} (\mathbf {x} _ {t} ^ {+}, \mathbf {c}, t) \| _ {2} ^ {2} \\ - \left(\| \epsilon^ {-} - \epsilon_ {\theta} \left(\mathbf {x} _ {t} ^ {-}, \mathbf {c}, \mathbf {u}, t\right) \| _ {2} ^ {2} - \| \epsilon^ {-} - \epsilon_ {\text {r e f}} \left(\mathbf {x} _ {t} ^ {-}, \mathbf {c}, t\right) \| _ {2} ^ {2}\right). \tag {6} \\ \end{array} +$$ + +This objective is similar to Eq. (3) but with the addition of $\mathbf{u}$ , representing extra features conditioned on for each user. But what user features should be conditioned on? + +# 4.2. Generating User Features from a VLM with In-Context Examples + +Discussion on User Features in Preference Datasets One fundamental challenge in personalized image generation lies in the effective parameterization of the user representation, denoted as $\mathbf{u}$ , for each individual. This parameterization is critical because it directly influences the model's capacity to generate outputs tailored to specific user preferences. Existing image preference datasets, such as Pick-a-Pic [16], typically lack comprehensive user information beyond basic identifiers, often focusing on generic preference data rather than capturing the nuanced individual preferences necessary for effective personalization. + +Even in preference datasets that include user information (e.g PRISM [15]), this information is frequently obtained through self-reported surveys, which can lead to incomplete or biased user profiles. Discrepancies often arise between stated preferences and actual behavior; users may not accurately articulate their likes and dislikes, or they may respond in a manner that they perceive as socially desirable rather than truthful. Such inconsistencies present significant challenges when attempting to model and predict user preferences accurately. + +![](images/8227e10fc14911015db064d4fdc1a615d4160ff02279b7591378f280cebe98fa.jpg) +Figure 2. Top-K accuracy of the User Classification. We fine-tune a user-classifier from the frozen embeddings from the VLM on few-shot preference examples for 300 users. This classifier significantly outperforms a random chance baseline. + +To address these issues, we adopt a simplified approach by assuming access solely to a unique user ID for each individual in the dataset. This user ID serves as a key to differentiate users within the labeled preference dataset, represented as $\mathcal{D}_{\mathrm{pref}} = \{(\mathbf{c},\mathbf{x}_0^+, \mathbf{x}_0^-, \mathbf{s})^{(i)}\}$ , where $\mathbf{s}$ denotes a unique user identifier. + +Conditioning on a Unique User Identifier The primary challenge, then, is to leverage this user ID $i$ to construct a meaningful user embedding that can effectively guide the model's predictions. A naive approach might involve representing the embedding $\mathbf{u}$ using a one-hot encoding of the user IDs, which can be formally defined by a function $f:\mathbf{S}\mapsto U$ , establishing a bijective mapping from the set of user IDs $\mathbf{s}\in \{1,2,\ldots ,|D|\}$ to the space of one-hot encoded vectors $\mathbf{u}_i\in \{0,1\}^{|D|}$ , where $|D|$ represents the total number of unique users in the dataset. While this method uniquely encodes each user, it suffers from a critical limitation: the inability to generalize to new, unseen users, as the embedding space is inherently tied to the users present in the training data. + +To enhance generalization and adaptability, we reconsider the parameterization of $u_{i}$ in the context of learning the distribution of reward functions, denoted as $r(\mathbf{c},\mathbf{x}_0,\mathbf{u}_i)$ . Here, a more flexible and robust parameterization leverages few-shot examples that reflect an individual user's preferences. These examples can be extracted from the offline preference dataset for each user, providing concrete instances of preferred and dispreferred content. Conditioning on these examples allows for a more accurate estimation of the reward function $r(\mathbf{c},\mathbf{x}_0,\mathbf{u}_i)$ for each user. + +In this work, we utilize the features of a pre-trained vision-language Model (VLM), specifically LLaVA-OneVision [21], an open-source multimodal model with multi-image processing capabilities. For each user, a set of $N = 4$ few-shot examples is selected, comprising a caption describing the context, a preferred image, and a + +dispreferred image. These examples are processed by the VLM, from which we extract an intermediate hidden state. + +To validate that the user embedding is effective, we condition on the frozen embedding to classify which user provided the preference annotations. As seen in Figure 2, the classifier significantly outperforms chance with a top-16 accuracy of $90\%$ , indicating that the embedding can disentangle different users in the preference dataset effectively. By incorporating this user-specific embedding into the conditioning of a text-to-image diffusion model, we enable user-aware reward modeling and generation. + +# 4.3. Personalization as Conditioning + +In order to add the user embedding as conditioning to the diffusion model in image generation, we utilize additional cross-attention layers akin to IP-Adapter [51]. Here we learn to condition on the user information using the VLM embeddings and then integrate the image features into pretrained UNet by adapted modules with decoupled cross-attention. There are many other ways to add conditioning information to diffusion models; we chose this instantiation due to its training and parameter efficiency. + +We utilize Stable Cascade [27] as our base model which comprises of three stages: Stages A-C. Stage A and B are image compressors, similar to the VAE in Stable Diffusion [34] but achieve a much higher compression rate; and Stage C is the generation module conditioned on text. In Stage C, the text features c from pretrained CLIP text encoder are plugged into the model by feeding into the cross-attention layers. Given the query features $\mathbf{Z}$ and text features $\mathbf{c}_t$ , the output is $\mathbf{Z}' = \text{Attention}(\mathbf{Q}, \mathbf{K}, \mathbf{V}) = \text{Softmax}(\frac{\mathbf{Q}\mathbf{K}^T}{\sqrt{d}})\mathbf{V}$ , where $\mathbf{Q} = \mathbf{Z}\mathbf{W}_q$ , $\mathbf{K} = \mathbf{c}_t\mathbf{W}_k$ , $\mathbf{V} = \mathbf{c}_t\mathbf{W}_v$ are query, key and values, and $\mathbf{W}_q, \mathbf{W}_k, \mathbf{W}_v$ are trainable layers. + +Following IP-Adapter [51], we utilize a decoupled crossattention mechanism to condition on the user embedding. First, we add a new cross-attention layer to each cross-attention layer in the Stage C model, using the original query $\mathbf{Q}$ and introducing new keys $\mathbf{K}'$ and values $\mathbf{V}'$ for user features. Then, we simply add the output of user cross-attention to the output of text cross-attention. Hence the final formulation of cross-attention is + +$$ +\mathbf {Z} ^ {\prime} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q K} ^ {T}}{\sqrt {d}}\right) \mathbf {V} + \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \left(\mathbf {K} ^ {\prime}\right) ^ {T}}{\sqrt {d}}\right) \mathbf {V} ^ {\prime}, \tag {7} +$$ + +where $\mathbf{Q} = \mathbf{Z}\mathbf{W}_q$ , $\mathbf{K} = \mathbf{c}_t\mathbf{W}_k$ , $\mathbf{V} = \mathbf{c}_t\mathbf{W}_v$ , $\mathbf{K}' = \mathbf{u}_t\mathbf{W}_k'$ , $\mathbf{V}' = \mathbf{u}_t\mathbf{W}_v'$ and $\mathbf{W}_k'$ , $\mathbf{W}_v'$ are new introduced trainable parameters for $\mathbf{u}$ . During training, we optimize only the added cross-attention layers while keeping the parameters of pretrained diffusion model frozen, using the training objective as in Eq. (6). We also randomly drop user features (zero out the user embedding) in the training stage as regularization. + +
CLIPAestheticHPS
Stable Cascade31.975.3323.87
Diffusion-DPO32.485.4625.96
SFT32.265.5625.78
ours32.665.9227.51
Diffusion-DPO (CLIP)32.96--
Diffusion-DPO (Aesthetic)-6.42-
Diffusion-DPO (HPS)--28.61
+ +Table 1. Averaged scores on three reward function. We condition on each reward to generate images both in SFT and our methods. Diffusion-DPO (*) are upper-bounds. + +This allows us to also prompt the model unconditionally of the user, by passing a zero embedding for the user. + +# 5. Experiments + +Models and Dataset In this section, we demonstrate the performance of PPD across a range of experiments designed to assess its capabilities in personalized preference alignment. We use the objective from Eq. (6) to fine-tune the open-source diffusion model, Stable Cascade [27] – specifically Stage C, which is responsible for text-conditioned generation. Following Wallace et al. [45], we fine-tune our model on the Pick-a-Pic v2 dataset [16], a preference dataset containing 58K text prompts and 0.8M image pairs labeled by 5K users. + +Training Details During training of PPD, we only optimize the added cross-attention layers while keeping the pretrained diffusion model fixed, as described in Sec. 4.3. We use AdamW [24] with an effective batch size of 768 pairs and a learning rate of $1 \times 10^{-5}$ for all experiments, training for one epoch. We find $\beta \in [0.1, 2]$ for hyperparameter tuning. Details are in Appendix B. + +User Information Conditioning We begin by demonstrating the model's ability to condition on synthetic user information, optimizing three distinct reward functions, we further show that generalize by linearly interpolating the weights of different reward functions (Sec. 5.1). We then show that the model can also condition on real user preferences, generating personalized images based on few-shot examples from new users (Sec. 5.2). + +# 5.1. Synthetic User Alignment + +Synthetic User We utilize three reward functions, each representing an aspect of image-to-text generation quality, to simulate three distinct synthetic users: (i) CLIP (Open-CLIP ViT-H/14) [13, 31] for prompt image alignment, (ii) Aesthetic Predictor [37] for image visual appeal, and (iii) HPSv2 (Human Preference Score v2) [47] for human preference estimate. We relabel pairs $(\mathbf{c},\mathbf{x}_0^+, \mathbf{x}_0^-, \mathbf{u})$ based + +![](images/20be1be546b5fead67b0bd5c4f13a60cb64e577ed7f117541aa39af3dbb5ba32.jpg) +Prompt Alignment (CLIP) + +![](images/4f420a367979d577308e125d5b7d491db21d5fea58fcc7c85aee2c073ed16c37.jpg) +Aesthetic Quality + +Human Preference Score (HPS) +Figure 3. Automatic win rate evaluation with reward functions. +![](images/2eea0ef63b8efad1dbc5e2d104f77df3138e1c6441dae699bdf45b3e25503108.jpg) +We compare against Stable Cascade, Diffusion-DPO, and SFT. + +on the above reward functions, where $\mathbf{u}$ is a one-hot vector indicating the reward function, taking one of 3 values: $\mathbf{u} \in \{[1,0,0],[0,1,0],[0,0,1]\}$ . + +Baseline We compare our method with: (1) Stable Cascade [27] pretrained model, which is also the reference model during training; (2) Diffusion-DPO, fine-tuned on Pick-a-Pic label without user conditioning [45]; (3) SFT, supervised fine-tuning on relabeled data with user conditioning. We also compare with (4) Diffusion-DPO (CLIP/Aesthetic/HPS), each annotation indicating a separately fine-tuned model with corresponding reward, this serves as the upper-bound of our method. For a fair comparison, all baselines are fine-tuned using Stable Cascade with the same architecture as our model, optimizing only the added cross-attention layers. + +Text-to-Image Generation For each method, we generate images with 30 inference steps using captions from PartiPrompts [53], a benchmark containing 1.6K examples across various categories and challenge aspects. For SFT and our method, we generate three images per prompt, each conditioned on one reward function. For Diffusion-DPO (CLIP/Aesthetic/HPS), images for each annotation are generated from its corresponding model. We report averaged reward scores in Tab. 1 and win rate comparisons in Fig. 3, we can see that our method significantly outperforms Stable Cascade in different reward functions, achieving win rates of $60\%$ , $90\%$ , and $93\%$ in terms of CLIP, Aesthetic, and HPS respectively. Our method effectively optimizes each + +reward by conditioning on individual preference, outperforming supervised fine-tuning by a large margin. Unlike Diffusion-DPO, our method can jointly optimize multiple user rewards, achieving win rates comparable to Diffusion-DPO models fine-tuned separately on each user's data. + +Generalization via Interpolation During training we fine-tune the model using one-hot encoding of each objective. Specifically, the user embedding vector $\mathbf{u}$ being $[1,0,0],[0,1,0],[0,0,1]$ corresponding to optimization for CLIP, Aesthetic, and HPS. The user embedding can also be seen as weighting mechanism for each objective. However, during inference, we are also interested in whether we can trade-off among different objectives by interpolating their respective weights. In Fig. 4, we generate images conditioned on various objective weightings, namely interpolating the given user embedding, we can see that as the weight of an objective increases, the corresponding score also rises, and vice versa. Therefore, our method can generalize to different reward combinations, even only trained on single objective data. This can not be achieved by optimizing separate models and each with an objective. + +This shows that by linearly interpolating the user embedding, we can adjust the rewards of the generated images accordingly. This finding aligns with the real user experiments presented next, as real user embeddings exhibit a linear correlation in embedding space. + +# 5.2. Personalized Real User Alignment + +We now consider a realistic setting with preferences from the Pick-a-Pic dataset for each user. In this setting, we utilize the user features which are generated from a VLM on in-context preference examples as outlined in Section 4.2. We scale from the 3 synthetic "users" in Section 5.1 to a filtered subset of the 4800 unique users in Pick-a-Pic. + +Experiment setting We repartition the Pick-a-Pic dataset into a set of 4 splits: (1) Train, (2) Validation, (3) Test with Seen Users + Held-out Captions and (4) Test with Unseen Users + Held-out Captions. This decomposition allows us to study independently the generalization of our methods to new prompts and unseen users. Additionally, for each user, we take sets of $N = 4$ few-shot preference examples for each user and use these groupings for the conditioning of the VLM on few-shot preference pairs. We thus have to filter out 824 users in Pick-a-Pic which have less than 4 examples per user, as no user embedding for this user can be formed. We generate the user embedding through the procedure outlined in Sec. 4.2. + +Generation Results We perform qualitative analysis on the set of images generated for each user. We perform this evaluation by generating a user profile per user. We use a set of few-shot preference pairs per user which we condition a + +![](images/70a6289357c9c72c33aff0e3d98e28ae1bd8abc357aec192681bac4b954b5c79.jpg) +Figure 4. PPD is able to interpolate among three distinct objectives during inference. (a) generated images conditioned on (b) various weights, with three axes representing CLIP, Aesthetic, and HPS; (c) reward scores for each image. The score for each objective increases as its respective weight increases, and decreasing otherwise. For each row from left to right, the CLIP score increases alongside its weight, leading to a decrease in the Aesthetic score as its weight decreases. From bottom to top, the HPS score increases with its weight. + +![](images/e7354d9f93bc8ec5d6af7da47b40884f69a8751e5c2b1532969da8b6594ce367.jpg) +Comparison of Win Rates for Real User Alignment + +![](images/b8c952716d2a9018bb07db669b11a97a52d0af5438f45dc4ac3a4e4ae6441965.jpg) +Figure 5. Automatic Win Rate Evaluation with GPT-4o We evaluate on seen users, held out users, and aggregate to see win-rates for generic users, comparing Stable Cascade and PPD. + +![](images/5c361648958e60ee5242fa93059135f946fae62ee10d07ffd8dd87131580d02e.jpg) + +VLM on along with Chain-of-Thought prompting (COT) to construct a user-profile consistent with the preference pairs. The COT prompt asks the model to summarize each image in the preference pair and describe the differences between them, which is then aggregated to form a user profile. Sample user profiles and the full Chain-of-Thought prompt is found in Appendix B. As seen in Fig. 6, our method has closer alignment to the user profile with respect to Stable Cascade and Diffusion-DPO, being able to pick up details about the user such as the highlighted part. While Stable Cascade and Diffusion-DPO with augmented captions can capture certain aspects of the user description, they often overfit to the user profile, resulting in generated images that overlook the intended image caption. + +Personalized Preference Alignment Evaluation We evaluate whether the generated model aligns with specific user preference using a VLM as a judge, leveraging strong + +VLMs such as GPT4o-vision [26], similar to recipes in prior work [4, 19]. In particular, we condition the VLM on few-shot examples from the user as well as a pair of images for a given test time caption and ask the model to judge which of the two images would be preferred by the user. To ensure consistency of the response the model selects, we judge each pair of images twice using the VLM, flipping the ordering of the images, filtering out pairs where the VLM is inconsistent. With this paradigm and 4 shots per user, we find that the VLM is able to match the ground truth personalized human preference with $83\%$ accuracy. + +We decompose our evaluation into two settings: (1) evaluating users that were seen during training but on captions that were held out and (2) evaluating held-out users and captions. We aggregate over these settings to also compute metrics for the general user in the dataset. For the 4341 users in Pick-a-Pic we select 192 held out and 277 in-distribution users. For each of these users, we choose $4 + 1$ examples, where 4 examples are used for the in-context conditioning for the user and 1 example is used as the caption for evaluation. We compute an automatic win-rate using pairs of models, generating one image for each caption and evaluating with the VLM-as-Judge pipeline outlined above. + +We report win rate comparisons in Fig. 5. Here, PPD significantly outperforms Stable Cascade in all three evaluation settings, achieving a win rate of $83\%$ on seen users, and $78\%$ on unseen users, aggregated to $81\%$ on all users. + +
User ProfileStable CascadeDiffusion-DPODiffusion-DPOOurs
Append user profile to caption as augmentation
The preferred images share common themes of wood grain, classic or futuristic design elements, and a focus on aesthetics. They often feature unique front grille patterns, large windows, and a mix of traditional and modern design features. ...A needle-felted palm tree
This user prefers images that have a personal and intimate feel with a focus on the subject's appearance and expression. They also appreciate images with soft focus backgrounds and warm lighting that highlight texture and detail. ...An anime girl, masterpiece, good line art, trending in pixiv
The preferred images ... such as a softer color palette, more dynamic lighting, and a balanced composition that enhances the visual appeal of the scene. The characters are often positioned off-center, which adds to the dynamism of the composition. The backgrounds are rich in detail. ...Aelita from Code Lyoko in the forest, gathering flowers
The user is drawn to images that blend elements of reality with fantastical elements, creating a sense of wonder and curiosity. ... The user may be interested in fantasy art, animation, and storytelling through visual media.A cat on a propaganda poster
+ +Figure 6. Qualitative Analysis of Images Generated by PPD and Baselines. Compared to Diffusion-DPO, PPD achieves closer alignment with the generated user profile, highlighted in green. The caption-augmented method captures user profile details; however, it often leads to unintended image alterations, causing the image to disregard the caption itself, as indicated in red. + +# 6. Discussion, Conclusion, and Limitations + +We propose a framework for personalizing text-to-image diffusion models using user feature embeddings from a VLM. By conditioning on these embeddings via decoupled cross-attention and fine-tuning with a DPO objective, our method enables accurate reward modeling and high-quality personalized image generation. It generalizes to unseen users with only a few preference examples at test time. + +Several open questions and limitations remain. Our current approach that conditioning on few-shot examples is caption-agnostic, selecting random user preferences; future work could explore active preference selection to improve personalization. Extending this method to video generation and studying the trade-off between implicit vs. stated preferences are also promising directions. + +# Acknowledgment + +This research is done during MD's internship at Luma AI and supported in part by ARO (W911NF-21-1-0125), ONR (N00014-23-1-2159), and the CZ Biohub. AS gratefully acknowledges the support of the NSF Graduate Research Fellowship Program. + +# References + +[1] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. Training a helpful and harmless assistant with reinforcement learning from human feedback, 2022. 2 +[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2 +[3] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 2 +[4] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Qinglan Huang, Canyu Chen, Qinghao Ye, Zhihong Zhu, Yuqing Zhang, Jiawei Zhou, Zhuokai Zhao, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. MJ-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation? arXiv e-prints, art. arXiv:2407.04842, 2024. 7 +[5] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 2 +[6] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2 +[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1 +[8] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306, 2024. 2 +[9] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion + +models. In Advances in Neural Information Processing Systems, 2024. 1, 2 +[10] Shangbin Feng, Taylor Sorensen, Yuhan Liu, Jillian Fisher, Chan Young Park, Yejin Choi, and Yulia Tsvetkov. Modular Pluralism: Pluralistic Alignment via Multi-LLM Collaboration. arXiv e-prints, art. arXiv:2406.15951, 2024. 1, 2 +[11] Gabriel Goh, James Betker, Li Jing, Aditya Ramesh, Tim Brooks, Jianfeng Wang, Lindsey Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Prafulla Dhariwal, Casey Chu, Joy Jiao, Jong Wook Kim, Alex Nichol, Yang Song, Lijuan Wang, and Tao Xu. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 2 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in neural information processing systems, pages 6840-6851, 2020. 1, 2 +[13] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 2,5 +[14] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In Advances in Neural Information Processing Systems 34, 2021. 2 +[15] Hannah Rose Kirk, Alexander Whitefield, Paul Röttger, Andrew Bean, Katerina Margatina, Juan Ciro, Rafael Mosquera, Max Bartolo, Adina Williams, He He, Bertie Vidgen, and Scott A. Hale. The PRISM Alignment Project: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models. arXiv e-prints, art. arXiv:2404.16019, 2024. 4 +[16] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569, 2023. 1, 2, 4, 5, 3 +[17] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners, 2023. 3 +[18] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2 +[19] Seongyun Lee, Seungone Kim, Sue Hyun Park, Geewook Kim, and Minjoon Seo. Prometheus-Vision: Vision-Language Model as a Judge for Fine-Grained Evaluation. arXiv e-prints, art. arXiv:2401.06591, 2024. 7 +[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 3 +[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. arXiv e-prints, art. arXiv:2408.03326, 2024. 4 + +[22] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. arXiv preprint arXiv:2404.04465, 2024. 2 +[23] Xinyu Li, Zachary C Lipton, and Liu Leqi. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133, 2024. 2 +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[25] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2 +[26] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 2, 7 +[27] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6 +[28] Sriyash Poddar, Yanming Wan, Hamish Ivison, Abhishek Gupta, and Natasha Jaques. Personalizing reinforcement learning from human feedback with variational preference learning. arXiv preprint arXiv:2408.10075, 2024. 2 +[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1 +[30] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5 +[32] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, 2024. 1, 2, 4 +[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5 +[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. + +Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, pages 36479-36494, 2022. 1 +[36] Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. Whose Opinions Do Language Models Reflect? arXiv e-prints, art. arXiv:2303.17548, 2023. 1, 2 +[37] Christoph Schuhmann. CLIP MLP aesthetic score predictor. https://github.com/christophschuhmann/improved-aesthetic-predictor, 2022. Accessed: 2024-11-06. 2,5 +[38] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278–25294, 2022. 1 +[39] Eyal Segalis, Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. A picture is worth a thousand words: Principled recaptioning improves image generation. arXiv preprint arXiv:2310.16656, 2023. 2 +[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265, 2015. 1, 2 +[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2 +[42] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 1, 2 +[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2 +[44] Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Ryting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, and Yejin Choi. A Roadmap to Pluralistic Alignment. arXiv e-prints, art. arXiv:2402.05070, 2024. 1, 2 +[45] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 1, 2, 3, 4, 5, 6 +[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 3 +[47] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 2, 5 + +[48] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 2 +[49] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, 2024. 2 +[50] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8941-8951, 2024. 2 +[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 1, 2, 5 +[52] Chao Yu, Hong Lu, Jiaxuan Gao, Qixin Tan, Xinting Yang, Yu Wang, Yi Wu, and Eugene Vinitsky. Few-shot in-context preference learning using large language models. arXiv preprint arXiv:2410.17233, 2024. 2 +[53] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 6 +[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2 +[55] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. arXiv preprint arXiv:2401.12244, 4, 2024. 2 +[56] Rui Zheng, Shihan Dou, Songyang Gao, Yuan Hua, Wei Shen, Binghai Wang, Yan Liu, Senjie Jin, Qin Liu, Yuhao Zhou, Limao Xiong, Lu Chen, Zhiheng Xi, Nuo Xu, Wenbin Lai, Minghao Zhu, Cheng Chang, Zhangyue Yin, Rongxiang Weng, Wensen Cheng, Haoran Huang, Tianxiang Sun, Hang Yan, Tao Gui, Qi Zhang, Xipeng Qiu, and Xuanjing Huang. Secrets of rlhf in large language models part i: Ppo, 2023. 2 \ No newline at end of file diff --git a/2025/Personalized Preference Fine-tuning of Diffusion Models/images.zip b/2025/Personalized Preference Fine-tuning of Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f0cdba25c5090a9ede107ee8d9cd33c7fcb7bd70 --- /dev/null +++ b/2025/Personalized Preference Fine-tuning of Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ac1413825fb32fa0f1a77771b4b124b4c9218f44dbff12beb6a2e5b18a7f56d +size 674862 diff --git a/2025/Personalized Preference Fine-tuning of Diffusion Models/layout.json b/2025/Personalized Preference Fine-tuning of Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..037bf20806a8476824288234fa90c14191a78786 --- /dev/null +++ b/2025/Personalized Preference Fine-tuning of Diffusion Models/layout.json @@ -0,0 +1,8781 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 480, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 480, + 121 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 480, + 121 + ], + "type": "text", + "content": "Personalized Preference Fine-tuning of Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "spans": [ + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "text", + "content": "Meihua Dang" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "text", + "content": " Anikait Singh" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "text", + "content": " Linqi Zhou" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "text", + "content": " Stefano Ermon" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "text", + "content": " Jiaming Song" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "text", + "content": "Stanford University, " + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 142, + 531, + 172 + ], + "type": "text", + "content": "Luma AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 174, + 490, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 174, + 490, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 174, + 490, + 186 + ], + "type": "text", + "content": "{mhdang,anikait,ermon}@cs.stanford.edu,{alexzhou,jiaming}@lumalabs.ai" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 238, + 297, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 539 + ], + "type": "text", + "content": "RLHF techniques like DPO can significantly improve the generation quality of text-to-image diffusion models. However, these methods optimize for a single reward that aligns model generation with population-level preferences, neglecting the nuances of individual users' beliefs or values. This lack of personalization limits the efficacy of these models. To bridge this gap, we introduce PPD, a multi-reward optimization objective that aligns diffusion models with personalized preferences. With PPD, a diffusion model learns the individual preferences of a population of users in a few-shot way, enabling generalization to unseen users. Specifically, our approach (1) leverages a vision-language model (VLM) to extract personal preference embeddings from a small set of pairwise preference examples, and then (2) incorporates the embeddings into diffusion models through cross attention. Conditioning on user embeddings, the text-to-image models are fine-tuned with the DPO objective, simultaneously optimizing for alignment with the preferences of multiple users. Empirical results demonstrate that our method effectively optimizes for multiple reward functions and can interpolate between them during inference. In real-world user scenarios, with as few as four preference examples from a new user, our approach achieves an average win rate of " + }, + { + "bbox": [ + 55, + 238, + 297, + 539 + ], + "type": "inline_equation", + "content": "76\\%" + }, + { + "bbox": [ + 55, + 238, + 297, + 539 + ], + "type": "text", + "content": " over Stable Cascade, generating images that more accurately reflect specific user preferences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 559, + 135, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 135, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 135, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 578, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 687 + ], + "type": "text", + "content": "Diffusion models [12, 40, 42, 43] have achieved great success in text-to-images generation tasks, as demonstrated by models such as DALL-E 2 [33], Imagen [35], Stable Diffusion [29, 34], and Stable Cascade [27]. These large-scale models, often trained through supervised fine-tuning on text-image datasets containing high-quality aesthetic images such as LAION-5B [38], demonstrate impressive image quality and a remarkable visual understanding of text prompts. Recent advances focus on fine-tuning diffusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 323 + ], + "type": "text", + "content": "models to better align with human preferences. Methods such as DDPO [2] and DPOK [9] employ reinforcement learning from human feedback (RLHF) techniques to optimize human-alignment reward functions. Alternatively, Diffusion-DPO [45] utilize direct preference optimization (DPO) [32] to directly optimize for human preferences using an offline ranking dataset. SD3 [7] shows that fine-tuning diffusion models with human feedback techniques can significantly enhance overall image quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 324, + 556, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 556, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 556, + 469 + ], + "type": "text", + "content": "Despite these promising results, Diffusion-DPO aligns a model with a single reward that represents population-level human preferences. These fine-tuning techniques are thus insufficient for achieving personalized preference alignment. As studied in domains such as language [10, 36, 44], individual users have unique preferences—for instance, some may favor images with brighter colors, while others might prefer images with centralized foregrounds. Finetuning a separate model for each person based on their preference data is impractical and does not scale well, limiting the feasibility of personalized alignment. Additionally, this may limit cross-user generalization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 567 + ], + "type": "text", + "content": "A complementary research direction introduces additional control mechanisms into diffusion models to enable customization and personalization. Approaches such as IP-Adapter [51] and CtrlNet [54] allow for controlled image generation by using an additional conditioning image that encodes personalization features. However, these methods are constrained by their reliance on single-image inputs and do not directly learn or adapt to human preferences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": "To address these limitations, we propose PPD, a general framework that incorporates personalization into a diffusion model preference alignment through multi-reward optimization. Our approach learns a diverse set of human preferences within a single model and generalizes effectively to new users. Personalization in our approach requires minimal additional data, as we label samples in the preference dataset with a unique identifier for each user. This annotation is inexpensive to obtain and commonly included in datasets such as Pick-a-Pic [16]. We leverage vision-language models (VLMs) such as LLaVA-OneVision [20] to extract user preference embeddings from" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 227, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 227, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 227, + 713 + ], + "type": "text", + "content": "*Equal contributions. \nWork done during MD's internship at Luma AI." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8020" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "few-shot pairwise preference examples for each user. We show that these few-shot text embeddings from preference pairs serve as effective features to represent reward functions under the Bradley-Terry model [3] and thus are suitable for conditioning the diffusion model. Next, adopting Stable Cascade [27] as our base model, we condition the diffusion model on these user embeddings through additional cross-attention layers and fine-tune it with a variant of the Diffusion-DPO [45] objective for preference alignment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 324 + ], + "type": "text", + "content": "Our empirical evaluations demonstrate that PPD can effectively optimize multiple rewards including CLIP [13, 31], Aesthetic [37], and HPS [47], significantly outperforming baselines. It also interpolates smoothly between different rewards during inference. Additionally, in real-world user scenarios, PPD aligns more closely with individual user preferences on the dataset Pick-a-Pic [16]. Notably, with only a few-shot ranking examples for a new user, PPD generate images that align closely with that user's specific preference and outperforms prior approaches such as Stable Cascade, achieving an average win rate of " + }, + { + "bbox": [ + 55, + 180, + 294, + 324 + ], + "type": "inline_equation", + "content": "76\\%" + }, + { + "bbox": [ + 55, + 180, + 294, + 324 + ], + "type": "text", + "content": ", as evaluated by GPT4o [26]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 333, + 142, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 142, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 142, + 345 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 353, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 295, + 534 + ], + "type": "text", + "content": "Aligning Diffusion Models Aligning text-to-image diffusion models has been extensively studied. Some methods optimize with online RL using a Bradley Terry reward function [2, 9, 55]. Others fine-tune diffusion models to increase the reward of generated images by gradient backpropagation on the reward function [5, 30, 49]. Recently proposed approaches [45, 50] leverage the duality of the reward learning and policy optimization from DPO [32] to eliminate the need of on-policy samples and only utilize offline human preference data. Alternatively, DiffusionKTO [22] replaces DPO with KTO [8], eliminating the need for paired data. Additionally, some methods modify the training data distribution on visually appealing and textually cohered data [6, 18, 48], while others improve text accuracy by re-captioning pre-collected web images [11, 39]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 548, + 295, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 657 + ], + "type": "text", + "content": "Controllable Generation for Diffusion Models To address the lack of fine-grained control in image synthesis, various methods add conditional controls to text-to-image diffusion models. CtrlNet [54] and T2I-Adapter [25] introduce additional UNet modules that encode spatial information, such as edges and human poses, derived from conditioning images. IP-Adapter [51] introduces additional cross-attention layers to incorporate high-level semantic information from reference images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "LLM Alignment The alignment of large language models (LLMs) using RLHF objectives has been extensively explored [1, 32, 56], demonstrating effective instruction-following and alignment to population values. However," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "a key challenge remains in representing the diversity of human preferences [36], as these models often capture an averaged population preference rather than accounting for individual differences, leading to biases and reduced alignment with underrepresented communities. Pluralistic Alignment [10, 44] addresses this issue by modeling a distribution of individuals, each with distinct values and perspectives, with methods like distributional and overton alignment enhancing calibration across diverse user groups. Several works have investigated LLM personalization through the reparameterization of reward models [23, 28, 52], enabling a more tailored user experience. However, these works leverage synthetic users and study a limited number of users, particularly in text and robotic control domains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 251, + 391, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 391, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 391, + 266 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 272, + 553, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 553, + 296 + ], + "type": "text", + "content": "We introduce the preliminaries behind diffusion models, DPO and its adaption to diffusion models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": "Diffusion Models Diffusion models are generative models consisting two processes: a forward (diffusion) process and a backward (denoising) process [12, 40, 42, 43]. In the forward process, Gaussian noise " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": " is progressively added to a data distribution " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{data}}(\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": " steps, guided by noise scheduling functions " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": ". At timestep " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": ", noisy images " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": " are generated via " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\alpha_t\\mathbf{x}_0 + \\sigma_t\\epsilon" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": ". In the backward process, samples are generated from Gaussian noise with a learnable model, which can also be conditioned on additional inputs, such as text in the case of text-to-image diffusion models. Typically, diffusion models - denoted as " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": " - are trained to predict the Gaussian noise " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": " from noisy images " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": " with conditioning " + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 308, + 555, + 476 + ], + "type": "text", + "content": ", using a denoising objective:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 347, + 485, + 553, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 485, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 347, + 485, + 553, + 500 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {D M}} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\omega (\\lambda_ {t}) \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "4aae823a9d16aa72c822f4cc6d73fb56c29c36161b76e66b71267a92cedd0d5c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\lambda_{t} = \\log \\left[\\alpha_{t}^{2} / \\sigma_{t}^{2}\\right]" + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "text", + "content": " is a signal-to-noise ratio, " + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\omega (\\lambda_t)" + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "text", + "content": " is a pre-specified weighting function [14]. Once the model " + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 509, + 553, + 559 + ], + "type": "text", + "content": " is trained, images can be generated from random noise using accelerated sampling methods such as DDIM [41]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Direct Preference Optimization (DPO) DPO [32] is a fine-tuning paradigm that aligns autoregressive large language models (LLMs) with human preferences without relying on traditional reinforcement learning techniques. Unlike Reinforcement Learning from Human Feedback (RLHF), which involves training a reward model and then fine-tuning the LLM using reinforcement learning, DPO directly optimizes the model based on human preference data. Given a dataset containing ranking pairs " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-)" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " being the prompt and " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^+, \\mathbf{x}_0^-" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " being the preferred and dispreferred responses labeled by human, the training objective optimizes a simple loss under which we can obtain an" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8021" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 558, + 264 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 558, + 264 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 558, + 264 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 558, + 264 + ], + "type": "image", + "image_path": "34d385018c2e606d6fb799da3bd83b159e7de042ab8269021764468ded12aeef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 268, + 309, + 281 + ], + "lines": [ + { + "bbox": [ + 69, + 268, + 309, + 281 + ], + "spans": [ + { + "bbox": [ + 69, + 268, + 309, + 281 + ], + "type": "text", + "content": "Stage 1: generate user profile with few-shot image pairs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 357, + 268, + 541, + 280 + ], + "lines": [ + { + "bbox": [ + 357, + 268, + 541, + 280 + ], + "spans": [ + { + "bbox": [ + 357, + 268, + 541, + 280 + ], + "type": "text", + "content": "Stage 2: train personalized diffusion model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 290, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 54, + 290, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 54, + 290, + 555, + 314 + ], + "type": "text", + "content": "Figure 1. The overall architecture of PPD. In Stage 1, user embedding are generated with few-shot preference examples using a VLM. In Stage 2, we fine-tune diffusion models on the preference datasets with the user embedding as conditioning added to cross-attention." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 334, + 119, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 119, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 119, + 347 + ], + "type": "text", + "content": "optimal policy:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 354, + 294, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 354, + 294, + 400 + ], + "spans": [ + { + "bbox": [ + 70, + 354, + 294, + 400 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\mathrm {D P O}} (\\theta) = - \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0} ^ {+}, \\mathbf {x} _ {0} ^ {-}} [ \\log \\sigma (\\beta \\Delta) ], \\text {w h e r e} \\\\ \\Delta = \\left(\\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0} ^ {+} \\mid \\mathbf {c}\\right)}{p _ {\\mathrm {r e f}} \\left(\\mathbf {x} _ {0} ^ {+} \\mid \\mathbf {c}\\right)} - \\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0} ^ {-} \\mid \\mathbf {c}\\right)}{p _ {\\mathrm {r e f}} \\left(\\mathbf {x} _ {0} ^ {-} \\mid \\mathbf {c}\\right)}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "ee6f4ff3cda2298e57bfbac5e4fc0e64e01a36b3658d22a1916d319f3d1670e7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "text", + "content": " is the sigmoid function and " + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "text", + "content": " is a hyperparameter determining how much " + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0|\\mathbf{c})" + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "text", + "content": " may deviate from a reference distribution " + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{ref}}(\\mathbf{x}_0|\\mathbf{c})" + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "text", + "content": ". Intuitively, the objective aims to increase the likelihoods of preferred responses while decreasing that of dispreferred ones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 479, + 296, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 575 + ], + "type": "text", + "content": "Diffusion DPO The DPO objective in Eq. (2) requires optimization of log-likelihoods " + }, + { + "bbox": [ + 55, + 479, + 296, + 575 + ], + "type": "inline_equation", + "content": "\\log p_{\\theta}(\\mathbf{x}_0^+|\\mathbf{c})" + }, + { + "bbox": [ + 55, + 479, + 296, + 575 + ], + "type": "text", + "content": ", which is analytically intractable for diffusion models. Diffusion-DPO [45] leverages the fact that the (appropriately weighted) training objective of diffusion models as in Eq. (1) can be viewed as the evidence lower bound (ELBO) of their log-likelihoods. It derives an upper bound to DPO for diffusion models to align them with human preferences:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 583, + 292, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 583, + 292, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 583, + 292, + 600 + ], + "type": "interline_equation", + "content": "L _ {\\text {D i f f - D P O}} (\\theta) = - \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0} ^ {+}, \\mathbf {x} _ {0} ^ {-}} [ \\log \\sigma (- \\beta T \\omega (\\lambda_ {t}) \\Delta) ], \\text {w h e r e}", + "image_path": "efdf0bc5194c71c01d6ec30e6f646e5ad3894f68b0311c3a4bf54e5b13eacaf4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 601, + 294, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 601, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 63, + 601, + 294, + 632 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta = \\| \\epsilon^ {+} - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, t) \\| _ {2} ^ {2} - \\| \\epsilon^ {+} - \\epsilon_ {\\mathrm {r e f}} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\\\ - \\left(\\| \\epsilon^ {-} - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, t) \\| _ {2} ^ {2} - \\| \\epsilon^ {-} - \\epsilon_ {\\text {r e f}} (\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, t) \\| _ {2} ^ {2}\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "b080633b1b9c7c6a8eb812c8c331ca8af192f3977229fd13f483b98f114dbc4d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is the caption; " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^+" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^-" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " are the preferred and dispreferred images; " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^+ = \\alpha_t\\mathbf{x}_0^+ +\\sigma_t\\epsilon^+" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^- = \\alpha_t\\mathbf{x}_0^- +\\sigma_t\\epsilon^-" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " are noisy images obtained from forward process with Gaussian noise " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon^{+}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon^{-}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " . This objective computes 4 terms as in Eq. (1): the denoising loss of training and preference distributions " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " on preferred and dispreferred" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 334, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 369 + ], + "type": "text", + "content": "noisy images " + }, + { + "bbox": [ + 313, + 334, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^+" + }, + { + "bbox": [ + 313, + 334, + 555, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 334, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^-" + }, + { + "bbox": [ + 313, + 334, + 555, + 369 + ], + "type": "text", + "content": " respectively. Intuitively, it aims to decrease the loss of preferred images while increasing that of dispreferred ones." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 380, + 370, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 380, + 370, + 392 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 370, + 392 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 401, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 533 + ], + "type": "text", + "content": "Despite impressive image quality and alignment control achieved in prior works, a key limitation among works that use DPO for diffusion models is that it assumes a uniform preference among all users for generated images. In practice, this assumption often fails, as individuals may have diverse preferences related to style, color, lighting, and other visual attributes. While such preferences can be easily expressed by human through pairwise comparisons, it is challenging to summarize them accurately through text or single-image prompts, due to the inaccuracy of natural language and the limitations inherent in single-prompt formats." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 533, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 641 + ], + "type": "text", + "content": "Motivated by this problem, we propose PPD, a method that aligns diffusion models with individual-level human preferences. We first formulate the problem of personalization and define a new DPO variant for this task (Sec. 4.1), and then show that individual-level preference learning can be achieved using few-shot comparison examples using a VLM (Sec. 4.2), finally we introduce a parameterization of denoiser that can takes advantage of the VLM embeddings (Sec. 4.3). The overall architecture is in Fig. 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 648, + 407, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 407, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 407, + 661 + ], + "type": "text", + "content": "4.1. Problem Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "We consider a dataset containing " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "(\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-, \\mathbf{u})" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " is the text prompt, " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^+" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^-" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " are preferred and dispreferred generated images and " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " represents the user labeling the preference. User features " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " can be high-dimensional including" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8022" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "information like age or gender, but often " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " consists only of scalar user IDs " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = [u]" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "r(\\mathbf{c},\\mathbf{x}_0,\\mathbf{u})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " as the reward on image " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " for user " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " given prompt " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". We would like to fine-tune a text-to-image models " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "p_{\\theta_i}(\\mathbf{x}_0|\\mathbf{c})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " for individual user " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " such that reward is maximized for this user while keeping close to a reference model " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{ref}}(\\mathbf{x}_0|\\mathbf{c})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " in terms of KL-divergence as regularization:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 163, + 295, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 163, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 97, + 163, + 295, + 199 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {p _ {\\theta_ {i}}} \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0}} \\left[ r (\\mathbf {c}, \\mathbf {x} _ {0}, \\mathbf {u} _ {i}) \\right] \\tag {4} \\\\ - \\beta \\mathbb {D} _ {\\mathrm {K L}} \\left[ p _ {\\theta_ {i}} (\\mathbf {x} _ {0} | \\mathbf {c}) \\| p _ {\\text {r e f}} (\\mathbf {x} _ {0} | \\mathbf {c}) \\right], \\\\ \\end{array}", + "image_path": "66df14aa4c0473db7d690452f8f8ca74ba2a5e73369e445c00de9615a67690de.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "text", + "content": " is a parameter controlling how much " + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "inline_equation", + "content": "p_{\\theta_i}(\\mathbf{x}_0|\\mathbf{c})" + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "text", + "content": " deviates from " + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{ref}}(\\mathbf{x}_0|\\mathbf{c})" + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "text", + "content": ". Following [32], the optimal " + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "inline_equation", + "content": "p_{\\theta_i}" + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "text", + "content": " for user " + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 55, + 205, + 296, + 242 + ], + "type": "text", + "content": " takes the form" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 249, + 295, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 249, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 83, + 249, + 295, + 262 + ], + "type": "interline_equation", + "content": "p _ {\\theta_ {i}} \\left(\\mathbf {x} _ {0} | \\mathbf {c}\\right) \\propto p _ {\\operatorname {r e f}} \\left(\\mathbf {x} _ {0} | \\mathbf {c}\\right) \\exp \\left(r \\left(\\mathbf {c}, \\mathbf {x} _ {0}, \\mathbf {u} _ {i}\\right) / \\beta\\right). \\tag {5}", + "image_path": "25b2c370d52b6ec8e1b8a3e889ab859745c1f335a74ebd717ed3354ff7e84e85.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "text", + "content": "Suppose model " + }, + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "text", + "content": " has sufficient capacity, we can represent " + }, + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "inline_equation", + "content": "p_{\\theta_i}(\\mathbf{x}_0|\\mathbf{c})" + }, + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0|\\mathbf{c},\\mathbf{u}_i)" + }, + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "text", + "content": " serves as a conditioning. This allows us to jointly optimize a single model " + }, + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_0|\\mathbf{c},\\mathbf{u})" + }, + { + "bbox": [ + 55, + 269, + 296, + 329 + ], + "type": "text", + "content": " for all users. Following [45], we have the following objective to learn a personalized diffusion models:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 335, + 294, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 335, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 63, + 335, + 294, + 385 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\mathrm {P P D}} (\\theta) = - \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {0} ^ {+}, \\mathbf {x} _ {0} ^ {-}, \\mathbf {u}} [ \\log \\sigma (- \\beta T \\omega (\\lambda_ {t}) \\Delta) ], \\text {w h e r e} \\\\ \\Delta = \\| \\epsilon^ {+} - \\epsilon_ {\\theta} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, \\mathbf {u}, t) \\| _ {2} ^ {2} - \\| \\epsilon^ {+} - \\epsilon_ {\\mathrm {r e f}} (\\mathbf {x} _ {t} ^ {+}, \\mathbf {c}, t) \\| _ {2} ^ {2} \\\\ - \\left(\\| \\epsilon^ {-} - \\epsilon_ {\\theta} \\left(\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, \\mathbf {u}, t\\right) \\| _ {2} ^ {2} - \\| \\epsilon^ {-} - \\epsilon_ {\\text {r e f}} \\left(\\mathbf {x} _ {t} ^ {-}, \\mathbf {c}, t\\right) \\| _ {2} ^ {2}\\right). \\tag {6} \\\\ \\end{array}", + "image_path": "6803d50f3b249e00f0db8bd316e87a276d26ef97fb8601df2487349938df1312.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 390, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 426 + ], + "type": "text", + "content": "This objective is similar to Eq. (3) but with the addition of " + }, + { + "bbox": [ + 55, + 390, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 390, + 296, + 426 + ], + "type": "text", + "content": ", representing extra features conditioned on for each user. But what user features should be conditioned on?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 433, + 296, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 458 + ], + "type": "text", + "content": "4.2. Generating User Features from a VLM with In-Context Examples" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": "Discussion on User Features in Preference Datasets One fundamental challenge in personalized image generation lies in the effective parameterization of the user representation, denoted as " + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": ", for each individual. This parameterization is critical because it directly influences the model's capacity to generate outputs tailored to specific user preferences. Existing image preference datasets, such as Pick-a-Pic [16], typically lack comprehensive user information beyond basic identifiers, often focusing on generic preference data rather than capturing the nuanced individual preferences necessary for effective personalization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Even in preference datasets that include user information (e.g PRISM [15]), this information is frequently obtained through self-reported surveys, which can lead to incomplete or biased user profiles. Discrepancies often arise between stated preferences and actual behavior; users may not accurately articulate their likes and dislikes, or they may respond in a manner that they perceive as socially desirable rather than truthful. Such inconsistencies present significant challenges when attempting to model and predict user preferences accurately." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 329, + 71, + 541, + 203 + ], + "blocks": [ + { + "bbox": [ + 329, + 71, + 541, + 203 + ], + "lines": [ + { + "bbox": [ + 329, + 71, + 541, + 203 + ], + "spans": [ + { + "bbox": [ + 329, + 71, + 541, + 203 + ], + "type": "image", + "image_path": "8227e10fc14911015db064d4fdc1a615d4160ff02279b7591378f280cebe98fa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 212, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 258 + ], + "type": "text", + "content": "Figure 2. Top-K accuracy of the User Classification. We fine-tune a user-classifier from the frozen embeddings from the VLM on few-shot preference examples for 300 users. This classifier significantly outperforms a random chance baseline." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 273, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 346 + ], + "type": "text", + "content": "To address these issues, we adopt a simplified approach by assuming access solely to a unique user ID for each individual in the dataset. This user ID serves as a key to differentiate users within the labeled preference dataset, represented as " + }, + { + "bbox": [ + 313, + 273, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{pref}} = \\{(\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-, \\mathbf{s})^{(i)}\\}" + }, + { + "bbox": [ + 313, + 273, + 555, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 273, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 313, + 273, + 555, + 346 + ], + "type": "text", + "content": " denotes a unique user identifier." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "text", + "content": "Conditioning on a Unique User Identifier The primary challenge, then, is to leverage this user ID " + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "text", + "content": " to construct a meaningful user embedding that can effectively guide the model's predictions. A naive approach might involve representing the embedding " + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "text", + "content": " using a one-hot encoding of the user IDs, which can be formally defined by a function " + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "inline_equation", + "content": "f:\\mathbf{S}\\mapsto U" + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "text", + "content": ", establishing a bijective mapping from the set of user IDs " + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{s}\\in \\{1,2,\\ldots ,|D|\\}" + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "text", + "content": " to the space of one-hot encoded vectors " + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i\\in \\{0,1\\}^{|D|}" + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "inline_equation", + "content": "|D|" + }, + { + "bbox": [ + 313, + 354, + 556, + 521 + ], + "type": "text", + "content": " represents the total number of unique users in the dataset. While this method uniquely encodes each user, it suffers from a critical limitation: the inability to generalize to new, unseen users, as the embedding space is inherently tied to the users present in the training data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 522, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 641 + ], + "type": "text", + "content": "To enhance generalization and adaptability, we reconsider the parameterization of " + }, + { + "bbox": [ + 313, + 522, + 556, + 641 + ], + "type": "inline_equation", + "content": "u_{i}" + }, + { + "bbox": [ + 313, + 522, + 556, + 641 + ], + "type": "text", + "content": " in the context of learning the distribution of reward functions, denoted as " + }, + { + "bbox": [ + 313, + 522, + 556, + 641 + ], + "type": "inline_equation", + "content": "r(\\mathbf{c},\\mathbf{x}_0,\\mathbf{u}_i)" + }, + { + "bbox": [ + 313, + 522, + 556, + 641 + ], + "type": "text", + "content": ". Here, a more flexible and robust parameterization leverages few-shot examples that reflect an individual user's preferences. These examples can be extracted from the offline preference dataset for each user, providing concrete instances of preferred and dispreferred content. Conditioning on these examples allows for a more accurate estimation of the reward function " + }, + { + "bbox": [ + 313, + 522, + 556, + 641 + ], + "type": "inline_equation", + "content": "r(\\mathbf{c},\\mathbf{x}_0,\\mathbf{u}_i)" + }, + { + "bbox": [ + 313, + 522, + 556, + 641 + ], + "type": "text", + "content": " for each user." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "In this work, we utilize the features of a pre-trained vision-language Model (VLM), specifically LLaVA-OneVision [21], an open-source multimodal model with multi-image processing capabilities. For each user, a set of " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " few-shot examples is selected, comprising a caption describing the context, a preferred image, and a" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8023" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "dispreferred image. These examples are processed by the VLM, from which we extract an intermediate hidden state." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 95, + 295, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 95, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 54, + 95, + 295, + 204 + ], + "type": "text", + "content": "To validate that the user embedding is effective, we condition on the frozen embedding to classify which user provided the preference annotations. As seen in Figure 2, the classifier significantly outperforms chance with a top-16 accuracy of " + }, + { + "bbox": [ + 54, + 95, + 295, + 204 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 54, + 95, + 295, + 204 + ], + "type": "text", + "content": ", indicating that the embedding can disentangle different users in the preference dataset effectively. By incorporating this user-specific embedding into the conditioning of a text-to-image diffusion model, we enable user-aware reward modeling and generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 211, + 228, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 228, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 228, + 224 + ], + "type": "text", + "content": "4.3. Personalization as Conditioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 228, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 295, + 335 + ], + "type": "text", + "content": "In order to add the user embedding as conditioning to the diffusion model in image generation, we utilize additional cross-attention layers akin to IP-Adapter [51]. Here we learn to condition on the user information using the VLM embeddings and then integrate the image features into pretrained UNet by adapted modules with decoupled cross-attention. There are many other ways to add conditioning information to diffusion models; we chose this instantiation due to its training and parameter efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "text", + "content": "We utilize Stable Cascade [27] as our base model which comprises of three stages: Stages A-C. Stage A and B are image compressors, similar to the VAE in Stable Diffusion [34] but achieve a much higher compression rate; and Stage C is the generation module conditioned on text. In Stage C, the text features c from pretrained CLIP text encoder are plugged into the model by feeding into the cross-attention layers. Given the query features " + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "text", + "content": " and text features " + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t" + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "text", + "content": ", the output is " + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}' = \\text{Attention}(\\mathbf{Q}, \\mathbf{K}, \\mathbf{V}) = \\text{Softmax}(\\frac{\\mathbf{Q}\\mathbf{K}^T}{\\sqrt{d}})\\mathbf{V}" + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\mathbf{Z}\\mathbf{W}_q" + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = \\mathbf{c}_t\\mathbf{W}_k" + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\mathbf{c}_t\\mathbf{W}_v" + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "text", + "content": " are query, key and values, and " + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q, \\mathbf{W}_k, \\mathbf{W}_v" + }, + { + "bbox": [ + 55, + 336, + 295, + 483 + ], + "type": "text", + "content": " are trainable layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 578 + ], + "type": "text", + "content": "Following IP-Adapter [51], we utilize a decoupled crossattention mechanism to condition on the user embedding. First, we add a new cross-attention layer to each cross-attention layer in the Stage C model, using the original query " + }, + { + "bbox": [ + 55, + 483, + 295, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 483, + 295, + 578 + ], + "type": "text", + "content": " and introducing new keys " + }, + { + "bbox": [ + 55, + 483, + 295, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{K}'" + }, + { + "bbox": [ + 55, + 483, + 295, + 578 + ], + "type": "text", + "content": " and values " + }, + { + "bbox": [ + 55, + 483, + 295, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{V}'" + }, + { + "bbox": [ + 55, + 483, + 295, + 578 + ], + "type": "text", + "content": " for user features. Then, we simply add the output of user cross-attention to the output of text cross-attention. Hence the final formulation of cross-attention is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 594, + 295, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 594, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 65, + 594, + 295, + 624 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} ^ {\\prime} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V} + \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\left(\\mathbf {K} ^ {\\prime}\\right) ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V} ^ {\\prime}, \\tag {7}", + "image_path": "042b77f06d67298d8165e435f0f158b1dc9714553bd310e3bc1777ab3e1c91b4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\mathbf{Z}\\mathbf{W}_q" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = \\mathbf{c}_t\\mathbf{W}_k" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\mathbf{c}_t\\mathbf{W}_v" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{K}' = \\mathbf{u}_t\\mathbf{W}_k'" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}' = \\mathbf{u}_t\\mathbf{W}_v'" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k'" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v'" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": " are new introduced trainable parameters for " + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": ". During training, we optimize only the added cross-attention layers while keeping the parameters of pretrained diffusion model frozen, using the training objective as in Eq. (6). We also randomly drop user features (zero out the user embedding) in the training stage as regularization." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 556, + 184 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 556, + 184 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 556, + 184 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 556, + 184 + ], + "type": "table", + "html": "
CLIPAestheticHPS
Stable Cascade31.975.3323.87
Diffusion-DPO32.485.4625.96
SFT32.265.5625.78
ours32.665.9227.51
Diffusion-DPO (CLIP)32.96--
Diffusion-DPO (Aesthetic)-6.42-
Diffusion-DPO (HPS)--28.61
", + "image_path": "68424b2137a30dde53f8c38f4dc2184411912ae8fdd4d13be66bf85f06d6d0e1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 192, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 226 + ], + "type": "text", + "content": "Table 1. Averaged scores on three reward function. We condition on each reward to generate images both in SFT and our methods. Diffusion-DPO (*) are upper-bounds." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 239, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 264 + ], + "type": "text", + "content": "This allows us to also prompt the model unconditionally of the user, by passing a zero embedding for the user." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 274, + 395, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 395, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 395, + 288 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 294, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 415 + ], + "type": "text", + "content": "Models and Dataset In this section, we demonstrate the performance of PPD across a range of experiments designed to assess its capabilities in personalized preference alignment. We use the objective from Eq. (6) to fine-tune the open-source diffusion model, Stable Cascade [27] – specifically Stage C, which is responsible for text-conditioned generation. Following Wallace et al. [45], we fine-tune our model on the Pick-a-Pic v2 dataset [16], a preference dataset containing 58K text prompts and 0.8M image pairs labeled by 5K users." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 418, + 554, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 503 + ], + "type": "text", + "content": "Training Details During training of PPD, we only optimize the added cross-attention layers while keeping the pretrained diffusion model fixed, as described in Sec. 4.3. We use AdamW [24] with an effective batch size of 768 pairs and a learning rate of " + }, + { + "bbox": [ + 313, + 418, + 554, + 503 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 418, + 554, + 503 + ], + "type": "text", + "content": " for all experiments, training for one epoch. We find " + }, + { + "bbox": [ + 313, + 418, + 554, + 503 + ], + "type": "inline_equation", + "content": "\\beta \\in [0.1, 2]" + }, + { + "bbox": [ + 313, + 418, + 554, + 503 + ], + "type": "text", + "content": " for hyperparameter tuning. Details are in Appendix B." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 508, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 605 + ], + "type": "text", + "content": "User Information Conditioning We begin by demonstrating the model's ability to condition on synthetic user information, optimizing three distinct reward functions, we further show that generalize by linearly interpolating the weights of different reward functions (Sec. 5.1). We then show that the model can also condition on real user preferences, generating personalized images based on few-shot examples from new users (Sec. 5.2)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 612, + 458, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 458, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 458, + 625 + ], + "type": "text", + "content": "5.1. Synthetic User Alignment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "Synthetic User We utilize three reward functions, each representing an aspect of image-to-text generation quality, to simulate three distinct synthetic users: (i) CLIP (Open-CLIP ViT-H/14) [13, 31] for prompt image alignment, (ii) Aesthetic Predictor [37] for image visual appeal, and (iii) HPSv2 (Human Preference Score v2) [47] for human preference estimate. We relabel pairs " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{c},\\mathbf{x}_0^+, \\mathbf{x}_0^-, \\mathbf{u})" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " based" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8024" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 82, + 294, + 151 + ], + "blocks": [ + { + "bbox": [ + 135, + 71, + 228, + 81 + ], + "lines": [ + { + "bbox": [ + 135, + 71, + 228, + 81 + ], + "spans": [ + { + "bbox": [ + 135, + 71, + 228, + 81 + ], + "type": "text", + "content": "Prompt Alignment (CLIP)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 82, + 294, + 151 + ], + "lines": [ + { + "bbox": [ + 58, + 82, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 58, + 82, + 294, + 151 + ], + "type": "image", + "image_path": "20be1be546b5fead67b0bd5c4f13a60cb64e577ed7f117541aa39af3dbb5ba32.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 163, + 293, + 233 + ], + "blocks": [ + { + "bbox": [ + 149, + 154, + 214, + 163 + ], + "lines": [ + { + "bbox": [ + 149, + 154, + 214, + 163 + ], + "spans": [ + { + "bbox": [ + 149, + 154, + 214, + 163 + ], + "type": "text", + "content": "Aesthetic Quality" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 163, + 293, + 233 + ], + "lines": [ + { + "bbox": [ + 58, + 163, + 293, + 233 + ], + "spans": [ + { + "bbox": [ + 58, + 163, + 293, + 233 + ], + "type": "image", + "image_path": "4f420a367979d577308e125d5b7d491db21d5fea58fcc7c85aee2c073ed16c37.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 245, + 293, + 315 + ], + "blocks": [ + { + "bbox": [ + 120, + 236, + 235, + 245 + ], + "lines": [ + { + "bbox": [ + 120, + 236, + 235, + 245 + ], + "spans": [ + { + "bbox": [ + 120, + 236, + 235, + 245 + ], + "type": "text", + "content": "Human Preference Score (HPS)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 245, + 293, + 315 + ], + "lines": [ + { + "bbox": [ + 58, + 245, + 293, + 315 + ], + "spans": [ + { + "bbox": [ + 58, + 245, + 293, + 315 + ], + "type": "image", + "image_path": "2eea0ef63b8efad1dbc5e2d104f77df3138e1c6441dae699bdf45b3e25503108.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 324, + 294, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 324, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 294, + 335 + ], + "type": "text", + "content": "Figure 3. Automatic win rate evaluation with reward functions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 335, + 285, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 285, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 285, + 346 + ], + "type": "text", + "content": "We compare against Stable Cascade, Diffusion-DPO, and SFT." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 358, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 395 + ], + "type": "text", + "content": "on the above reward functions, where " + }, + { + "bbox": [ + 55, + 358, + 295, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 358, + 295, + 395 + ], + "type": "text", + "content": " is a one-hot vector indicating the reward function, taking one of 3 values: " + }, + { + "bbox": [ + 55, + 358, + 295, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{u} \\in \\{[1,0,0],[0,1,0],[0,0,1]\\}" + }, + { + "bbox": [ + 55, + 358, + 295, + 395 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 404, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 548 + ], + "type": "text", + "content": "Baseline We compare our method with: (1) Stable Cascade [27] pretrained model, which is also the reference model during training; (2) Diffusion-DPO, fine-tuned on Pick-a-Pic label without user conditioning [45]; (3) SFT, supervised fine-tuning on relabeled data with user conditioning. We also compare with (4) Diffusion-DPO (CLIP/Aesthetic/HPS), each annotation indicating a separately fine-tuned model with corresponding reward, this serves as the upper-bound of our method. For a fair comparison, all baselines are fine-tuned using Stable Cascade with the same architecture as our model, optimizing only the added cross-attention layers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Text-to-Image Generation For each method, we generate images with 30 inference steps using captions from PartiPrompts [53], a benchmark containing 1.6K examples across various categories and challenge aspects. For SFT and our method, we generate three images per prompt, each conditioned on one reward function. For Diffusion-DPO (CLIP/Aesthetic/HPS), images for each annotation are generated from its corresponding model. We report averaged reward scores in Tab. 1 and win rate comparisons in Fig. 3, we can see that our method significantly outperforms Stable Cascade in different reward functions, achieving win rates of " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " in terms of CLIP, Aesthetic, and HPS respectively. Our method effectively optimizes each" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "reward by conditioning on individual preference, outperforming supervised fine-tuning by a large margin. Unlike Diffusion-DPO, our method can jointly optimize multiple user rewards, achieving win rates comparable to Diffusion-DPO models fine-tuned separately on each user's data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 142, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 555, + 332 + ], + "type": "text", + "content": "Generalization via Interpolation During training we fine-tune the model using one-hot encoding of each objective. Specifically, the user embedding vector " + }, + { + "bbox": [ + 313, + 142, + 555, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 142, + 555, + 332 + ], + "type": "text", + "content": " being " + }, + { + "bbox": [ + 313, + 142, + 555, + 332 + ], + "type": "inline_equation", + "content": "[1,0,0],[0,1,0],[0,0,1]" + }, + { + "bbox": [ + 313, + 142, + 555, + 332 + ], + "type": "text", + "content": " corresponding to optimization for CLIP, Aesthetic, and HPS. The user embedding can also be seen as weighting mechanism for each objective. However, during inference, we are also interested in whether we can trade-off among different objectives by interpolating their respective weights. In Fig. 4, we generate images conditioned on various objective weightings, namely interpolating the given user embedding, we can see that as the weight of an objective increases, the corresponding score also rises, and vice versa. Therefore, our method can generalize to different reward combinations, even only trained on single objective data. This can not be achieved by optimizing separate models and each with an objective." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 334, + 554, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 554, + 393 + ], + "type": "text", + "content": "This shows that by linearly interpolating the user embedding, we can adjust the rewards of the generated images accordingly. This finding aligns with the real user experiments presented next, as real user embeddings exhibit a linear correlation in embedding space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 401, + 497, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 497, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 497, + 414 + ], + "type": "text", + "content": "5.2. Personalized Real User Alignment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": "We now consider a realistic setting with preferences from the Pick-a-Pic dataset for each user. In this setting, we utilize the user features which are generated from a VLM on in-context preference examples as outlined in Section 4.2. We scale from the 3 synthetic \"users\" in Section 5.1 to a filtered subset of the 4800 unique users in Pick-a-Pic." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 500, + 554, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 554, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 554, + 655 + ], + "type": "text", + "content": "Experiment setting We repartition the Pick-a-Pic dataset into a set of 4 splits: (1) Train, (2) Validation, (3) Test with Seen Users + Held-out Captions and (4) Test with Unseen Users + Held-out Captions. This decomposition allows us to study independently the generalization of our methods to new prompts and unseen users. Additionally, for each user, we take sets of " + }, + { + "bbox": [ + 313, + 500, + 554, + 655 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 313, + 500, + 554, + 655 + ], + "type": "text", + "content": " few-shot preference examples for each user and use these groupings for the conditioning of the VLM on few-shot preference pairs. We thus have to filter out 824 users in Pick-a-Pic which have less than 4 examples per user, as no user embedding for this user can be formed. We generate the user embedding through the procedure outlined in Sec. 4.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "Generation Results We perform qualitative analysis on the set of images generated for each user. We perform this evaluation by generating a user profile per user. We use a set of few-shot preference pairs per user which we condition a" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8025" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 289 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 289 + ], + "type": "image", + "image_path": "70a6289357c9c72c33aff0e3d98e28ae1bd8abc357aec192681bac4b954b5c79.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 298, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 298, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 555, + 344 + ], + "type": "text", + "content": "Figure 4. PPD is able to interpolate among three distinct objectives during inference. (a) generated images conditioned on (b) various weights, with three axes representing CLIP, Aesthetic, and HPS; (c) reward scores for each image. The score for each objective increases as its respective weight increases, and decreasing otherwise. For each row from left to right, the CLIP score increases alongside its weight, leading to a decrease in the Aesthetic score as its weight decreases. From bottom to top, the HPS score increases with its weight." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 372, + 140, + 441 + ], + "blocks": [ + { + "bbox": [ + 112, + 362, + 239, + 369 + ], + "lines": [ + { + "bbox": [ + 112, + 362, + 239, + 369 + ], + "spans": [ + { + "bbox": [ + 112, + 362, + 239, + 369 + ], + "type": "text", + "content": "Comparison of Win Rates for Real User Alignment" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 372, + 140, + 441 + ], + "lines": [ + { + "bbox": [ + 58, + 372, + 140, + 441 + ], + "spans": [ + { + "bbox": [ + 58, + 372, + 140, + 441 + ], + "type": "image", + "image_path": "e7354d9f93bc8ec5d6af7da47b40884f69a8751e5c2b1532969da8b6594ce367.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 141, + 373, + 217, + 440 + ], + "blocks": [ + { + "bbox": [ + 141, + 373, + 217, + 440 + ], + "lines": [ + { + "bbox": [ + 141, + 373, + 217, + 440 + ], + "spans": [ + { + "bbox": [ + 141, + 373, + 217, + 440 + ], + "type": "image", + "image_path": "b8c952716d2a9018bb07db669b11a97a52d0af5438f45dc4ac3a4e4ae6441965.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 449, + 295, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 483 + ], + "type": "text", + "content": "Figure 5. Automatic Win Rate Evaluation with GPT-4o We evaluate on seen users, held out users, and aggregate to see win-rates for generic users, comparing Stable Cascade and PPD." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 219, + 373, + 294, + 440 + ], + "blocks": [ + { + "bbox": [ + 219, + 373, + 294, + 440 + ], + "lines": [ + { + "bbox": [ + 219, + 373, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 219, + 373, + 294, + 440 + ], + "type": "image", + "image_path": "5c361648958e60ee5242fa93059135f946fae62ee10d07ffd8dd87131580d02e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 495, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 662 + ], + "type": "text", + "content": "VLM on along with Chain-of-Thought prompting (COT) to construct a user-profile consistent with the preference pairs. The COT prompt asks the model to summarize each image in the preference pair and describe the differences between them, which is then aggregated to form a user profile. Sample user profiles and the full Chain-of-Thought prompt is found in Appendix B. As seen in Fig. 6, our method has closer alignment to the user profile with respect to Stable Cascade and Diffusion-DPO, being able to pick up details about the user such as the highlighted part. While Stable Cascade and Diffusion-DPO with augmented captions can capture certain aspects of the user description, they often overfit to the user profile, resulting in generated images that overlook the intended image caption." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Personalized Preference Alignment Evaluation We evaluate whether the generated model aligns with specific user preference using a VLM as a judge, leveraging strong" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 364, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 495 + ], + "type": "text", + "content": "VLMs such as GPT4o-vision [26], similar to recipes in prior work [4, 19]. In particular, we condition the VLM on few-shot examples from the user as well as a pair of images for a given test time caption and ask the model to judge which of the two images would be preferred by the user. To ensure consistency of the response the model selects, we judge each pair of images twice using the VLM, flipping the ordering of the images, filtering out pairs where the VLM is inconsistent. With this paradigm and 4 shots per user, we find that the VLM is able to match the ground truth personalized human preference with " + }, + { + "bbox": [ + 313, + 364, + 555, + 495 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 495 + ], + "type": "text", + "content": " accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 509, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 653 + ], + "type": "text", + "content": "We decompose our evaluation into two settings: (1) evaluating users that were seen during training but on captions that were held out and (2) evaluating held-out users and captions. We aggregate over these settings to also compute metrics for the general user in the dataset. For the 4341 users in Pick-a-Pic we select 192 held out and 277 in-distribution users. For each of these users, we choose " + }, + { + "bbox": [ + 313, + 509, + 556, + 653 + ], + "type": "inline_equation", + "content": "4 + 1" + }, + { + "bbox": [ + 313, + 509, + 556, + 653 + ], + "type": "text", + "content": " examples, where 4 examples are used for the in-context conditioning for the user and 1 example is used as the caption for evaluation. We compute an automatic win-rate using pairs of models, generating one image for each caption and evaluating with the VLM-as-Judge pipeline outlined above." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": "We report win rate comparisons in Fig. 5. Here, PPD significantly outperforms Stable Cascade in all three evaluation settings, achieving a win rate of " + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": " on seen users, and " + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "inline_equation", + "content": "78\\%" + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": " on unseen users, aggregated to " + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "inline_equation", + "content": "81\\%" + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": " on all users." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8026" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 553, + 548 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 553, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 553, + 548 + ], + "type": "table", + "html": "
User ProfileStable CascadeDiffusion-DPODiffusion-DPOOurs
Append user profile to caption as augmentation
The preferred images share common themes of wood grain, classic or futuristic design elements, and a focus on aesthetics. They often feature unique front grille patterns, large windows, and a mix of traditional and modern design features. ...A needle-felted palm tree
This user prefers images that have a personal and intimate feel with a focus on the subject's appearance and expression. They also appreciate images with soft focus backgrounds and warm lighting that highlight texture and detail. ...An anime girl, masterpiece, good line art, trending in pixiv
The preferred images ... such as a softer color palette, more dynamic lighting, and a balanced composition that enhances the visual appeal of the scene. The characters are often positioned off-center, which adds to the dynamism of the composition. The backgrounds are rich in detail. ...Aelita from Code Lyoko in the forest, gathering flowers
The user is drawn to images that blend elements of reality with fantastical elements, creating a sense of wonder and curiosity. ... The user may be interested in fantasy art, animation, and storytelling through visual media.A cat on a propaganda poster
", + "image_path": "716469cf2a56c86cc04800742a61409a2c1e7b74de810f57e55e1c27a6ff294f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 555, + 555, + 590 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 555, + 590 + ], + "type": "text", + "content": "Figure 6. Qualitative Analysis of Images Generated by PPD and Baselines. Compared to Diffusion-DPO, PPD achieves closer alignment with the generated user profile, highlighted in green. The caption-augmented method captures user profile details; however, it often leads to unintended image alterations, causing the image to disregard the caption itself, as indicated in red." + } + ] + } + ], + "index": 1, + "type": "text" + }, + { + "bbox": [ + 55, + 609, + 276, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 276, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 276, + 624 + ], + "type": "text", + "content": "6. Discussion, Conclusion, and Limitations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "We propose a framework for personalizing text-to-image diffusion models using user feature embeddings from a VLM. By conditioning on these embeddings via decoupled cross-attention and fine-tuning with a DPO objective, our method enables accurate reward modeling and high-quality personalized image generation. It generalizes to unseen users with only a few preference examples at test time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 610, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 555, + 696 + ], + "type": "text", + "content": "Several open questions and limitations remain. Our current approach that conditioning on few-shot examples is caption-agnostic, selecting random user preferences; future work could explore active preference selection to improve personalization. Extending this method to video generation and studying the trade-off between implicit vs. stated preferences are also promising directions." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8027" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 152 + ], + "type": "text", + "content": "This research is done during MD's internship at Luma AI and supported in part by ARO (W911NF-21-1-0125), ONR (N00014-23-1-2159), and the CZ Biohub. AS gratefully acknowledges the support of the NSF Graduate Research Fellowship Program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 114, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 114, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 114, + 177 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 184, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 184, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 294, + 293 + ], + "type": "text", + "content": "[1] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. Training a helpful and harmless assistant with reinforcement learning from human feedback, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 296, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 339 + ], + "type": "text", + "content": "[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 342, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 373 + ], + "type": "text", + "content": "[3] Ralph Allan Bradley and Milton E Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 376, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 463 + ], + "type": "text", + "content": "[4] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Qinglan Huang, Canyu Chen, Qinghao Ye, Zhihong Zhu, Yuqing Zhang, Jiawei Zhou, Zhuokai Zhao, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. MJ-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation? arXiv e-prints, art. arXiv:2407.04842, 2024. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 465, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 294, + 498 + ], + "type": "text", + "content": "[5] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 500, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 294, + 554 + ], + "type": "text", + "content": "[6] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 556, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 294, + 621 + ], + "type": "text", + "content": "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 666 + ], + "type": "text", + "content": "[8] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 669, + 294, + 712 + ], + "type": "text", + "content": "[9] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "models. In Advances in Neural Information Processing Systems, 2024. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[10] Shangbin Feng, Taylor Sorensen, Yuhan Liu, Jillian Fisher, Chan Young Park, Yejin Choi, and Yulia Tsvetkov. Modular Pluralism: Pluralistic Alignment via Multi-LLM Collaboration. arXiv e-prints, art. arXiv:2406.15951, 2024. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 140, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 215 + ], + "type": "text", + "content": "[11] Gabriel Goh, James Betker, Li Jing, Aditya Ramesh, Tim Brooks, Jianfeng Wang, Lindsey Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Prafulla Dhariwal, Casey Chu, Joy Jiao, Jong Wook Kim, Alex Nichol, Yang Song, Lijuan Wang, and Tao Xu. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in neural information processing systems, pages 6840-6851, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 251, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 314 + ], + "type": "text", + "content": "[13] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 2,5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 316, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 348 + ], + "type": "text", + "content": "[14] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In Advances in Neural Information Processing Systems 34, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 350, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 436 + ], + "type": "text", + "content": "[15] Hannah Rose Kirk, Alexander Whitefield, Paul Röttger, Andrew Bean, Katerina Margatina, Juan Ciro, Rafael Mosquera, Max Bartolo, Adina Williams, He He, Bertie Vidgen, and Scott A. Hale. The PRISM Alignment Project: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models. arXiv e-prints, art. arXiv:2404.16019, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 553, + 480 + ], + "type": "text", + "content": "[16] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. arXiv preprint arXiv:2305.01569, 2023. 1, 2, 4, 5, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "text", + "content": "[17] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "text", + "content": "[18] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "text", + "content": "[19] Seongyun Lee, Seungone Kim, Sue Hyun Park, Geewook Kim, and Minjoon Seo. Prometheus-Vision: Vision-Language Model as a Judge for Fine-Grained Evaluation. arXiv e-prints, art. arXiv:2401.06591, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "text", + "content": "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[21] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy Visual Task Transfer. arXiv e-prints, art. arXiv:2408.03326, 2024. 4" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8028" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[22] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. arXiv preprint arXiv:2404.04465, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "text", + "content": "[23] Xinyu Li, Zachary C Lipton, and Liu Leqi. Personalized language modeling from personalized human feedback. arXiv preprint arXiv:2402.05133, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 173 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 228 + ], + "type": "text", + "content": "[25] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 249 + ], + "type": "text", + "content": "[26] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 251, + 294, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 305 + ], + "type": "text", + "content": "[27] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 306, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 306, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 294, + 350 + ], + "type": "text", + "content": "[28] Sriyash Poddar, Yanming Wan, Hamish Ivison, Abhishek Gupta, and Natasha Jaques. Personalizing reinforcement learning from human feedback with variational preference learning. arXiv preprint arXiv:2408.10075, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 404 + ], + "type": "text", + "content": "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "type": "text", + "content": "[30] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 514 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 569 + ], + "type": "text", + "content": "[32] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, 2024. 1, 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 669 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, pages 36479-36494, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[36] Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. Whose Opinions Do Language Models Reflect? arXiv e-prints, art. arXiv:2303.17548, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[37] Christoph Schuhmann. CLIP MLP aesthetic score predictor. https://github.com/christophschuhmann/improved-aesthetic-predictor, 2022. Accessed: 2024-11-06. 2,5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 274 + ], + "type": "text", + "content": "[38] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278–25294, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[39] Eyal Segalis, Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. A picture is worth a thousand words: Principled recaptioning improves image generation. arXiv preprint arXiv:2310.16656, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "type": "text", + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265, 2015. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 366, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 399 + ], + "type": "text", + "content": "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 433 + ], + "type": "text", + "content": "[42] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 488 + ], + "type": "text", + "content": "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "type": "text", + "content": "[44] Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Ryting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, and Yejin Choi. A Roadmap to Pluralistic Alignment. arXiv e-prints, art. arXiv:2402.05070, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 545, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 612 + ], + "type": "text", + "content": "[45] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "type": "text", + "content": "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[47] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 2, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8029" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 574 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[48] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[49] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 250 + ], + "type": "text", + "content": "[50] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8941-8951, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 295 + ], + "type": "text", + "content": "[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arxiv:2308.06721, 2023. 1, 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "text", + "content": "[52] Chao Yu, Hong Lu, Jiaxuan Gao, Qixin Tan, Xinting Yang, Yu Wang, Yi Wu, and Eugene Vinitsky. Few-shot in-context preference learning using large language models. arXiv preprint arXiv:2410.17233, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 418 + ], + "type": "text", + "content": "[53] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 419, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 419, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 419, + 294, + 463 + ], + "type": "text", + "content": "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 464, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 496 + ], + "type": "text", + "content": "[55] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. arXiv preprint arXiv:2401.12244, 4, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 498, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 574 + ], + "type": "text", + "content": "[56] Rui Zheng, Shihan Dou, Songyang Gao, Yuan Hua, Wei Shen, Binghai Wang, Yan Liu, Senjie Jin, Qin Liu, Yuhao Zhou, Limao Xiong, Lu Chen, Zhiheng Xi, Nuo Xu, Wenbin Lai, Minghao Zhu, Cheng Chang, Zhangyue Yin, Rongxiang Weng, Wensen Cheng, Haoran Huang, Tianxiang Sun, Hang Yan, Tao Gui, Qi Zhang, Xipeng Qiu, and Xuanjing Huang. Secrets of rlhf in large language models part i: Ppo, 2023. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8030" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/11693e39-cb59-4d3c-8f05-aea184580288_content_list.json b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/11693e39-cb59-4d3c-8f05-aea184580288_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ec4762dc75b3700dc9e31dfdc1d9283ec7145a59 --- /dev/null +++ b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/11693e39-cb59-4d3c-8f05-aea184580288_content_list.json @@ -0,0 +1,1532 @@ +[ + { + "type": "text", + "text": "Perturb-and-Revise: Flexible 3D Editing with Generative Trajectories", + "text_level": 1, + "bbox": [ + 145, + 130, + 851, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Susung Hong $^{1}$ Johanna Karras $^{1}$ Ricardo Martin-Brualla $^{2}$ Ira Kemelmacher-Shlizerman $^{1,2}$ $^{1}$ University of Washington $^{2}$ Google Research", + "bbox": [ + 125, + 179, + 869, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0456744913aec7f778273311d527dfbba574028498ee5fcd1bfbbcdb035a415a.jpg", + "image_caption": [ + "Figure 1. Perturb-and-Revise takes a source NeRF and an edit prompt as input and produces the edited result through: (1) versatile initialization via parameter perturbation, (2) multi-view consistent score distillation, and (3) refinement with the identity-preserving gradient." + ], + "image_footnote": [], + "bbox": [ + 147, + 250, + 861, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 494, + 326, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in text-based diffusion models have accelerated progress in 3D reconstruction and text-based 3D editing. Although existing 3D editing methods excel at modifying color, texture, and style, they struggle with extensive geometric or appearance changes, thus limiting their applications. To this end, we propose Perturb-and-Revise, which makes possible a variety of NeRF editing. First, we perturb the NeRF parameters with random initializations to create a versatile initialization. The level of perturbation is determined automatically through analysis of the local loss landscape. Then, we revise the edited NeRF via generative trajectories. Combined with the generative process, we impose identity-preserving gradients to refine the edited NeRF. Extensive experiments demonstrate that Perturb-and-Revise facilitates flexible, effective, and consistent editing of color, appearance, and geometry in 3D. For $360^{\\circ}$ results, please visit our project page: https://susunghong.github.io/Perturb-and-Revise.", + "bbox": [ + 88, + 527, + 485, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 829, + 220, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRFs) [34] have revolutionized the creation of high-quality 3D scenes, marking a significant advancement in 3D reconstruction technology. Beyond", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "their initial applications, NeRFs have enabled flexible generation of 3D content from models trained solely on image-text pairs [18, 40] through score distillation [25, 40, 58].", + "bbox": [ + 511, + 496, + 906, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, editing 3D content—an essential aspect of refinement and customization—remains a time-consuming and labor-intensive process across various industries, including animation, manufacturing, design, and gaming. This challenge is particularly pronounced with NeRFs, where color and density attributes are intricately encoded within their parameters. Consequently, there is an ongoing need for more intuitive and universally accessible tools for editing NeRFs, leveraging user-friendly interfaces such as text prompts for broader applicability and faster production.", + "bbox": [ + 509, + 545, + 908, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fortunately, recent innovations in text-conditioned diffusion models have enabled significant advances in 3D editing using natural language prompts. For instance, leveraging diffusion models [2, 43], recent works such as Instruct-NeRF2NeRF [8] and Posterior Distillation [25] propose methods that use text prompts to modify 3D scenes. While these methods excel in altering the color or style of an object, they struggle with edits involving significant geometric or appearance changes, limiting their applications.", + "bbox": [ + 509, + 699, + 908, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome these limitations, we propose a 3D object editing framework called Perturb-and-Revise (PnR), which leverages existing 3D generation methods for editing in a flexible and natural way (Figs. 1 and 2). Drawing on the", + "bbox": [ + 509, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16293", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a4e7f32f0d97019875e342d6fc409fff0c7540d4a344e57328fa3d4d59a5236.jpg", + "image_caption": [ + "Source 3D Scene" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 261, + 273 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c035d216508a2241fbcd98d489f27f96865c7f83e31224982997696332a481f9.jpg", + "image_caption": [ + "Instruct-NeRF2NeRF" + ], + "image_footnote": [], + "bbox": [ + 267, + 88, + 428, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ec4bffab462e5c59c8047e4cde5ccd3bb38998891a405abc95c8cf682238dee6.jpg", + "image_caption": [ + "Instruct-GS2GS" + ], + "image_footnote": [], + "bbox": [ + 428, + 88, + 584, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/67c69180d03e99875a45a5c6770ef55c48ff02528dd13b681672c91c4127b87d.jpg", + "image_caption": [ + "Posterior Distillation*" + ], + "image_footnote": [], + "bbox": [ + 586, + 88, + 743, + 272 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f8194064b3ac22c85c18725074a71d063ad53e82085a8302c85902a95e7e4d11.jpg", + "image_caption": [ + "Perturb-and-Revise (Ours)*" + ], + "image_footnote": [], + "bbox": [ + 745, + 89, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "notion that NeRF parameters optimized with score distillation can be considered as a particle [60] or data point [6], we propose a novel method that leverages perturbation at the parameter level to perform text-based edits requiring various changes. Specifically, we construct versatile NeRF initialization by interpolating the source NeRF and a random NeRF initialization. Intuitively, perturbation in the parameter space helps the particle escape local minima and facilitates its following of the natural trajectory of the generative ODE towards the distribution of the desired edit (Fig. 3), allowing for a wide range of challenging edits including changing the pose and introducing new objects.", + "bbox": [ + 88, + 335, + 482, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, to determine the amount of perturbation needed to escape the basin of attraction without costly searching algorithms, we propose an algorithm based on the loss landscape near the source parameters, by simulating a few score distillation steps ahead of the parameter perturbation and optimization. After the early step of our framework, to achieve more similarity to the original object and refine the quality of the result, we employ the Identity-Preserving Gradient (IPG) to increase the fidelity of the edited NeRF to the source NeRF. As a result, the output closely resembles the source while still maintaining fidelity to the intended edits.", + "bbox": [ + 88, + 517, + 482, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We extensively evaluate our method on 3D fashion objects, as well as general objects in Objaverse [4], on a variety of appearance- and geometry-based edits. Furthermore, we demonstrate that our method achieves state-of-the-art results on various 3D editing baselines.", + "bbox": [ + 88, + 696, + 482, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 784, + 232, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models [9, 49], which are closely associated with score-based models [21, 50, 51], have recently shown remarkable sample quality in image synthesis. Latent Diffusion Models (LDM) [43], which perform the diffusion forward and backward process in the latent space [7], have demonstrated their efficacy and genera", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bc107ecfc77875f8ae04e8bc2db533e300d8125079552eb0f28714c99971a0f6.jpg", + "image_caption": [ + "Figure 2. We qualitatively compare our method to Instruct-NeRF2NeRF [8], Instruct-GS2GS [54], and Posterior Distillation [25]. In this task, the goal is to edit the source \"bear\" into a \"polar bear.\" Compared to related works, our approach reconstructs a more realistic face and better overall geometry. An asterisk (*) denotes that we use an identical update rule and schedule.", + "Figure 3. Conceptual figure. The target distribution in the figures represents the conditional distribution of NeRF parameters relative to the edit prompt, and $\\mathcal{P}(\\Theta)$ denotes the distribution of randomly initialized NeRF parameters. First, parameter perturbation enables the parameters to escape from local minima and follow a natural generative path. Subsequently, during the refining process, the tug-of-war between two vectors, $\\lambda_{\\mathrm{d}}\\nabla_{\\theta}d(\\theta_{\\tau},\\theta_{\\mathrm{src}})$ (the red arrow) and $d\\theta_{\\tau}$ (the blue arrow), pushes the actual parameters into a region that is closer to either the source parameters or the high-density region specified by the edit prompt, following the purple arrow." + ], + "image_footnote": [], + "bbox": [ + 532, + 335, + 877, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion quality. This trend has been scaled up by training diffusion models on large-scale text-image datasets [45] for text-conditional generation. The prosperity of text-to-image diffusion models has led to their application in complex content creation, such as 3D scenes and videos [1, 10, 12, 14, 15, 23, 27, 33, 40, 46, 58, 60, 62]. Recently, many adapting methods of text-to-image diffusion models have been proposed to achieve either generation with additional conditions [2, 11, 13, 20, 35, 64, 66] or to maintain consistency needed for videos [14, 23, 41, 62] or 3D scenes [29, 47].", + "bbox": [ + 511, + 650, + 906, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Score distillation. Score distillation [40, 58, 60], a recent trend in synthesizing underlying parameterizations, mostly in 3D scene generation, uses only score-based models [43, 44] operating on 2D data. This approach has revolutionized text-to-3D generation by introducing an elegant method that does not require intense training on a", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16294", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D dataset. It enables the rich and flexible knowledge of 2D text-to-image models to be transferred to 3D generation [12, 27, 33, 58, 60]. However, due to bias in 2D models, a multi-view consistency problem has arisen, necessitating further analysis and generalization of the original method [12, 29, 47]. Another limitation of score distillation is its time-consuming optimization, with the controllability and editing capability remaining underexplored.", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-based NeRF editing. In recent years, neural radiance fields (NeRF) [34] have emerged as a groundbreaking approach for generating photorealistic novel views of a scene captured in photographs, and have been extended in many follow-up works [22, 32, 36, 39]. In response to these advancements, several techniques for modifying NeRF have been proposed. Traditional methodologies for editing NeRF include the alteration of materials and lighting [37, 52, 55], as well as the spatial manipulation of objects via bounding box frameworks [38, 63]. Additionally, there has been exploration in the stylization of NeRFs, including Edit-NeRF [30], Clip-NeRF [56], and NeRF-Art [57], as well as in distilling 2D features into radiance fields, such as in Distilled Feature Fields [24] and Neural Feature Fusion Fields [53], which enables nuanced, guided edits based on language or images. Crucially, Instruct-NeRF2NeRF [8] focuses on user-friendly, language-based editing commands, using an instruction-driven, 2D image-conditioned diffusion model [2] for more intuitive and context-aware 3D editing. However, even in a trend using 2D language-based models such as CLIP [42] or text-to-image diffusion models [2, 43, 44], these methodologies are unable to control large geometrical changes, such as changing the pose or introducing a massive object. Posterior Distillation [25], instead of naively matching the noise, proposes a way to use a loss that matches the stochastic latents [16, 61] of source and target images. Still, this method falls into issues such as slower convergence and large computational costs.", + "bbox": [ + 91, + 246, + 483, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 89, + 698, + 218, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models are a type of generative model that iteratively restores an image from Gaussian noise. The training objective of diffusion models involves training a neural network to estimate the score for a noised image $y$ given the noise level $\\sigma$ via denoising score matching [17]. Following the preconditioning convention [21], a diffusion model, parameterized by $\\phi$ , can be interpreted as a denoiser $D_{\\phi}(y;\\sigma)$ which minimizes the weighted L2 loss across different $\\sigma$ values in data samples.", + "bbox": [ + 89, + 729, + 482, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the context of text-based 3D scene generation, Score Distillation Sampling (SDS) [40] updates the 3D scene rep", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resentation given a text caption $c$ using the following rule:", + "bbox": [ + 511, + 90, + 897, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , c, x = g (\\theta , \\psi)) = \\\\ - \\mathbb {E} _ {\\sigma \\sim \\Sigma , n \\sim \\mathcal {N} (0, \\sigma^ {2} I)} \\left[ \\omega (\\sigma) \\left(D _ {\\phi} (x + n; \\sigma , c) - x\\right) \\frac {\\partial x}{\\partial \\theta} \\right]. \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 114, + 905, + 179 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this equation, $g(\\cdot)$ is the differentiable renderer, $\\psi$ is the random camera pose, $\\Sigma$ is a predefined distribution from which the noise level for the denoiser is sampled, and $\\omega(\\cdot)$ denotes the weighting function.", + "bbox": [ + 511, + 189, + 905, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SDS can be viewed as a form of particle-based variational inference [3, 5, 28, 31, 59, 60]. In this context, the update rule can be expressed as a generative ODE over an optimization step $\\tau$ , derived via Wasserstein gradient flow [3]:", + "bbox": [ + 511, + 250, + 906, + 311 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d \\theta_ {\\tau}}{d \\tau} = - \\mathbb {E} _ {\\sigma \\sim \\Sigma , n \\sim \\mathcal {N} (0, \\sigma^ {2} I)} \\left[ \\omega (\\sigma) \\left(D _ {\\phi} (x + n; \\sigma , c) - x\\right) \\frac {\\partial x}{\\partial \\theta_ {\\tau}} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 319, + 911, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For further details, please see the supplementary material.", + "bbox": [ + 511, + 364, + 893, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As background, multi-view diffusion models [47], denoted as $\\{D_{\\phi}^{\\mathrm{M}}(g(\\theta ,\\psi_i) + n;\\sigma ,c,\\psi_i)\\}_{i = 1}^{N}$ , consistently generate multi-view images simultaneously. Here, $N$ denotes the total number of multi-views generated simultaneously, and $\\psi_{i}$ denotes the $i$ -th viewpoint. Note that for $N = 1$ , these are ordinary diffusion models. For $N > 1$ , they are implemented by adapting the attention module to attend to different frames [47] and by training on 3D assets [4].", + "bbox": [ + 511, + 378, + 905, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 511, + 604, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In text-based editing scenarios, we assume that a user provides a text prompt, $c_{\\mathrm{edit}}$ , and a source 3D object represented by the parameterization $\\theta_{\\mathrm{src}}$ . The aim is to modify the 3D object into a new form $\\theta^{*} = \\mathcal{E}(\\theta_{\\mathrm{src}}, c_{\\mathrm{edit}})$ , ensuring that the transformation aligns with the text prompt and retains a resemblance to the original 3D object.", + "bbox": [ + 511, + 537, + 905, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At a high level, we leverage the generative ODE of 3D objects through novel parameter perturbation while preserving proximity to the source object. In Sec. 4.1, we introduce perturbation in parameter space that enables the parameters to escape local minima and converge to natural generative trajectories. In Sec. 4.2, we propose a novel algorithm to determine the degree of perturbation by analyzing the local loss landscape. Finally, in Sec. 4.3, we propose balancing fidelity between the text prompt and the source object to place the final output in the desired region.", + "bbox": [ + 511, + 628, + 905, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Parameter Perturbation for Flexible Object Editing", + "text_level": 1, + "bbox": [ + 511, + 787, + 905, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the generative ODE in Eq. 2, which minimizes the energy functional between distributions, a fully-optimized NeRF of a 3D object is considered to exhibit relatively low energy. This state corresponds to the particle residing in a local minimum, which remains stable despite alterations in", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16295", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the text prompts, as it is still capable of producing realistic renderings. As exemplified in the first row of Fig. 4, naively using $\\theta_{\\mathrm{src}}$ as initialization and making independent predictions on 2D noisy images with diffusion models are insufficient to handle the consistent changes in overall appearance and geometry that occur in a 3D object.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interpreting neural fields as a particle [60] or a data point [6], we propose parameter-level perturbation for an editing task, which facilitates changes in $\\theta_{\\mathrm{src}}$ by enabling the editing particle ODE to hijack the natural, coarse-to-fine generative process in Eq. 2 that the parameters are likely to follow. In other words, since adding more noise to the parameters is equivalent to undoing more of the optimization process, this approach enables the parameters to be flexible, allowing for significant changes.", + "bbox": [ + 89, + 181, + 482, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, let $\\mathcal{P}(\\Theta_0)$ denote the probability density function over $\\Theta_0$ , which is a set of random parameterizations. Further, assume that we have a fully optimized 3D scene, $\\theta_{\\mathrm{src}}$ . Then, we obtain the perturbed version of $\\theta_{\\mathrm{src}}$ by using linear interpolation:", + "bbox": [ + 89, + 316, + 482, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\text {p e r t u r b e d}} = \\operatorname {L e r p} \\left(\\theta_ {\\text {s r c}}, \\theta_ {0}, \\eta\\right) \\quad \\text {w h e r e} \\quad \\theta_ {0} \\sim \\mathcal {P} \\left(\\Theta_ {0}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 402, + 482, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta \\in [0,1]$ denotes the perturbation amount and $\\mathrm{Lerp}(\\theta_{\\mathrm{src}},\\theta_0,\\eta) = (1 - \\eta)\\cdot \\theta_{\\mathrm{src}} + \\eta \\cdot \\theta_0$ . Note that $\\eta = 1$ indicates $\\theta_{\\mathrm{perturbed}}$ is the random parameters. Theoretically, the interpolation makes the distribution of perturbed particles similar to the initializing distribution of NeRF, making it versatile. See the supplementary material for details.", + "bbox": [ + 89, + 430, + 482, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, we run the parameter ODE for textual editing, starting from $\\theta_{\\mathrm{perturbed}}$ , using a new prompt $c_{\\mathrm{edit}}$ . However, we observe that adopting ordinary single-view diffusion models (when $N = 1$ ) poses ambiguity when an edit introduces asymmetry or adds a new object. Therefore, we execute multi-view consistent updates as follows:", + "bbox": [ + 89, + 521, + 482, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d \\theta_ {\\tau}}{d \\tau} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {E} _ {\\sigma , n} \\left[ \\left(D _ {\\phi} ^ {\\mathrm {M}} \\left(z _ {i} + n; \\sigma , c _ {\\mathrm {e d i t}}, \\psi_ {i}\\right) - z _ {i}\\right) \\frac {\\partial z _ {i}}{\\partial \\theta_ {\\tau}} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 621, + 483, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we define $z_{i} \\coloneqq g(\\theta_{\\tau}, \\psi_{i})$ and omit the weighting factor for brevity.", + "bbox": [ + 89, + 688, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The conceptual illustration of parameter perturbation can be found in Fig. 3. This approach essentially reverts the optimization process by interpolating between the initial state and the fully optimized state. Consequently, the particles are relocated to a less-optimized state characterized by a more moderate energy landscape. This repositioning allows the particle ODE (Eq. 4) to intervene in the generative process with a new text prompt $c_{\\mathrm{edit}}$ , guiding the parameter trajectory. As shown in Fig. 4, increasing the noise level in the parameters inversely correlates with the degree of reversion of the optimization, allowing for more radical modifications in the model outputs.", + "bbox": [ + 88, + 719, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0f2e24a33f87c3b194df814caf850690c7bddaefba69dfbfb67f0e22290b21b6.jpg", + "image_caption": [ + "100 steps 200 steps 300 steps 500 steps 700 steps 900 steps 10000 steps", + "Figure 4. Effect of parameter perturbation. In this example, we aim to make a NeRF model of a standing person sit down using the word \"sitting.\" The scene converges quickly even with large perturbations $(\\eta = 0.6)$ , while complete regeneration yields blurry rendering results given the same number of optimization steps." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 901, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Determining $\\eta$ by Analyzing Loss Landscape", + "text_level": 1, + "bbox": [ + 511, + 391, + 890, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Sec. 4.1, we see that $\\eta$ controls parameter versatility by determining how similar the distribution of parameters is to the distribution of random NeRF initializations. The required degree of versatility depends on both the type of edits and the current source NeRF parameters. However, finding an optimal $\\eta$ through standard mechanisms, such as grid or random search, is computationally expensive. Therefore, we propose a novel algorithm that explores the basin of attraction surrounding local minima to determine $\\eta$ by analyzing the loss landscape.", + "bbox": [ + 511, + 414, + 905, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we leverage the loss function as a proxy to measure the depth and volume of the basin of attraction. Interestingly, for some examples, we observe that the loss function even increases during optimization steps, supporting our claim in Sec. 4.1 that the parameters are in a low-energy state. To address this, we simulate several optimization steps with $c_{\\mathrm{edit}}$ and calculate the total loss decrease prior to parameter perturbation. For algorithmic stability, we compute the difference between the averages of the first few steps and the last few steps. We then determine $\\eta(\\theta_{\\mathrm{src}}, c_{\\mathrm{edit}})$ based on the source parameters $\\theta_{\\mathrm{src}}$ and the text prompt $c_{\\mathrm{edit}}$ using an inverted exponential decay function. Subsequently, we apply parameter perturbation in Eq. 3 with $\\eta = \\eta(\\theta_{\\mathrm{src}}, c_{\\mathrm{edit}})$ . The complete algorithms for our parameter perturbation and $\\eta$ selection are provided in the supplementary material.", + "bbox": [ + 511, + 566, + 906, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Identity-Preserving Gradient (IPG)", + "text_level": 1, + "bbox": [ + 511, + 816, + 821, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While our parameter perturbation in Sec. 4.1 successfully enables flexible changes, there are still some estimation errors or biases arising from the diffusion model during this phase. On the other hand, imposing constraints that make", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16296", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/769013849dd445801c0939501c8f97c039dbbf494e20a2dba4de450a7afad531.jpg", + "image_caption": [ + "Source 3D Object", + "Score Distillation", + "Posterior Distillation", + "Instruct-NeRF2NeRF", + "Perturb-and-Revise (Ours)", + "Figure 5. Baseline comparisons with a wide range of edits. We compare our method with Score Distillation Sampling (SDS) [40], Posterior Distillation Sampling (PDS) [25], and Instruct-NeRF2NeRF [8]. For SDS and PDS, we use MVDream [47] as the backbone for fair comparison. SDS alters the appearance and texture of the source objects and is unable to handle edits that require extensive geometric changes (3rd, 4th, and 5th rows). PDS is not capable of making significant edits and cannot deviate far from local minima due to its high preservation term from the start [25]. While Instruct-NeRF2NeRF changes the texture of objects as desired, it cannot address geometric changes. In contrast, our method is capable of various types of edits, including those involving large geometric changes." + ], + "image_footnote": [], + "bbox": [ + 127, + 103, + 875, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "it similar to the source object (e.g., L2 distance) conflicts with the generative ODE, making the optimization challenging. To counteract the estimation error and to circumvent the conflict, we introduce the Identity-Preserving Gradient (IPG) term that is added at later steps to the editing gradient presented in Sec. 4.1.", + "bbox": [ + 89, + 652, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initially, we assume that the ideal parameters are closer to the source parameters and the high-density region of the target distribution. Then, as conceptually shown in Fig. 3, we instigate a tug-of-war between two vectors: one symbolizes the velocity consistently pointing towards the high-density region as dictated by the edit prompt, and the other symbolizes the velocity towards $\\theta_{\\mathrm{src}}$ . Specifically, the velocity in Eq. 4 propels the particle towards the region with a high likelihood given the text $c_{\\mathrm{edit}}$ , while the other gradient represents the pull towards the original NeRF parameteri", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "zation $\\theta_{\\mathrm{src}}$ . This effectively corrects the shift caused by the error in the editing process, ensuring it resembles the original while maintaining the intended edits. To this end, we extend the optimization process and compute IPGs during these additional steps.", + "bbox": [ + 511, + 652, + 906, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, we combine the result of Eq. 4 with the IPG to define a refinement step as follows:", + "bbox": [ + 511, + 729, + 906, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd \\theta_ {\\tau} ^ {\\text {r e f i n e}} = d \\theta_ {\\tau} + \\lambda_ {\\mathrm {d}} \\nabla_ {\\theta} d \\left(\\theta_ {\\tau}, \\theta_ {\\text {s r c}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 768, + 906, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(\\cdot, \\cdot)$ is a similarity metric. In practice, we observe that the combination of $L_{1}$ and perceptual loss [19], which is also the preferred choice in NeRF [34] training, is more robust and less susceptible to noise than using $L_{2}$ loss. Thus, given a random camera pose $\\psi$ , we use a combination of L1 distance and perceptual similarity between the rendered images $g(\\theta_{\\tau}, \\psi)$ and $g(\\theta_{\\mathrm{src}}, \\psi)$ , weighted by $\\lambda_{\\mathrm{L1}}$", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16297", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/70adcefa17fffdf7303e013f02071b2893134d7c9a9d34a4173f55fed37b23f6.jpg", + "image_caption": [ + "Source 3D Object", + "Score Distillation", + "Posterior Distillation", + "Instruct-NeRF2NeRF", + "Perturb-and-Revise (Ours)", + "Figure 6. Baseline comparisons of editing various general 3D objects from the Objaverse dataset [4]." + ], + "image_footnote": [], + "bbox": [ + 137, + 99, + 861, + 411 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cbd9dd176918061395080b343927fe8a912a908397b49ad8c0c978ca390c3e94.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricScore Distillation [40]Posterior Distillation [25]Instruct-NeRF2NeRF [8]Perturb-and-Revise (Ours)
CLIP-Dir-SimViT-B/32↑0.04800.02870.05830.0594
CLIP-Dir-SimViT-B/16↑0.04180.03040.05490.0534
CLIP-Dir-SimViT-L/14↑0.04150.02640.05390.0567
CLIP-Dir-Simaveraged↑0.04380.02850.05570.0565
LPIPSvgg↓0.12730.03370.10650.1060
LPIPSalex↓0.15330.02150.11120.1034
", + "bbox": [ + 125, + 446, + 867, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of different methods for fashion object editing. The best, second-best, and worst values are highlighted in bold, underlined, and gray, respectively. Values represent averages across all edit types and prompts. Our approach achieves a better trade-off between faithfulness to edit prompts and preservation of source 3D objects. While PDS exhibits lower LPIPS, it mostly generates edited objects that are nearly identical to the source, as evidenced by CLIP directional similarity.", + "bbox": [ + 89, + 555, + 906, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and $\\lambda_{\\mathfrak{p}}$ , respectively.", + "bbox": [ + 89, + 619, + 227, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Timestep Annealing", + "text_level": 1, + "bbox": [ + 89, + 660, + 281, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In text-based 3D generation, some recent works have revealed that using timestep annealing, which adjusts the noise level added to the 2D rendered images according to the global optimization step, boosts the quality of the results [47, 60]. In the editing task, while we use a smaller noise level at the start, we also identify the annealing method as crucial for maintaining the original quality of the 3D object while reducing blurry textures. To this end, we regard $\\Sigma$ as a function that depends on the editing step $\\tau$ , $\\Sigma(\\tau)$ , from which the noise level $\\sigma$ is sampled. Specifically, for $\\tau \\leq T$ , where $T$ is the final step of the schedule, $\\Sigma(\\tau)$ is gradually annealed from $\\mathcal{U}(\\sigma_{\\min}^{0}, \\sigma_{\\max}^{0})$ to $\\mathcal{U}(\\sigma_{\\min}^{\\mathrm{T}}, \\sigma_{\\max}^{\\mathrm{T}})$ . This approach initially facilitates low-frequency edits, followed by fine-grained ones.", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 618, + 643, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As baselines, we adopt two distillation-based methods: Score Distillation Sampling (SDS) [40] and Posterior Distillation Sampling (PDS) [25]. Bolded red in the prompts indicates the intended edits. We use the exact weighting factors for the losses as originally proposed in [25, 40]. For fair comparison, we use MVDream [47] as their backbones. Also, we use Instruct-NeRF2NeRF [8] as our baseline for the iterative dataset update strategy. We use several backbones for each metric and note them in the subscripts, while averaged refers to the average value across all backbones.", + "bbox": [ + 511, + 642, + 906, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.3D Object Editing", + "text_level": 1, + "bbox": [ + 511, + 801, + 689, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fashion object editing. In our fashion object editing experiment, we use MVDream [47] to generate a synthetic dataset of 3D fashion objects and perform edits using our framework. Specifically, we categorize the fashion object edits into color, pattern, shape, pose, and object edits, and", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16298", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/806296a49545bd21b6b7a6cea9fb55d7683afa6cf2d87510094b04186d496868.jpg", + "image_caption": [ + "Source 3D Object", + "Instruct-NeRF2NeRF", + "MVDream", + "Perturb-and-Revise (Ours)" + ], + "image_footnote": [], + "bbox": [ + 133, + 99, + 867, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8fbf0e6913b09ad0acfd7e882a93b931e6caf6dea7c7c0de99bcac2d1fc1707f.jpg", + "image_caption": [ + "Figure 7. Comparisons with Instruct-NeRF2NeRF (dataset update) [8] and MVDream (regeneration) [47] in editing Objaverse objects.", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 120, + 314, + 431, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e73f364f04c212157b4515bc368a3cdf1b1caa09313840583714def5ffa9f1f.jpg", + "image_caption": [ + "(b)", + "Figure 8. Ablation study on the selection of $\\eta$ . (a) and (b) show the CLIP direction similarity and the percentage of successful experiments (without errors) for different $\\eta$ values, respectively. When averaged across all types of edits, our adaptive method achieves near-maximum performance on these metrics compared to all fixed $\\eta$ values." + ], + "image_footnote": [], + "bbox": [ + 433, + 314, + 870, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "synthesize a total of 150 editing examples, with 30 for each type of edit. In this experiment, our framework takes as input a source object in NeRF with the desired edits to its characteristics, such as color, pattern, shape, pose, and added objects. Note that, unlike some previous work [25], we do not require a description of the original object.", + "bbox": [ + 88, + 523, + 482, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The qualitative results, as shown in Figs. 4 and 5, demonstrate the versatility introduced by parameter perturbation and our method's capability to perform various types of edits, respectively. In Fig. 5, we apply five types of edits: color, pattern, shape, pose, and object additions. We show that our framework supports a wide range of edits, including those requiring pose changes and introducing objects, corroborating its capability for significant geometric modifications, which have rarely been addressed in previous work [8, 25].", + "bbox": [ + 88, + 614, + 482, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "General object editing. We show a general object editing capability for our method, with objects from Objaverse [4]. We first optimize InstantNGP [36] with 200 rendered images and camera transforms extracted from an object. Then, we apply our editing method.", + "bbox": [ + 89, + 777, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We display the qualitative results of our method in Fig. 6. In addition to textural changes, our method successfully enables structural changes. In Fig. 7, we also demonstrate our", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3c73846066cb04da8987a186bef8f238a679d180917cee4626aeb9fb127e0819.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCLIP-Dir-Sim↑CLIP-Dir-Con↑LPIPS↓
w/o Refinement0.06240.75720.1147
w/ Refinement0.05650.76420.1047
", + "bbox": [ + 550, + 520, + 867, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study on the IPG steps. We present average values across all edit types, prompts, and evaluation backbones. Our approach yields a substantial reduction in LPIPS while balancing faithfulness to the source object and edit prompt.", + "bbox": [ + 511, + 585, + 906, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method's ability to perform creative edits anchored by the source 3D object. Here, to demonstrate flexibility, we use two prompts with different animal motifs. We can see that regenerating the object with MVDream produces results that appear completely different from the source 3D object. In contrast, our method successfully anchors the original shape while changing it to align with the edit prompt.", + "bbox": [ + 511, + 652, + 906, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison with Baselines", + "text_level": 1, + "bbox": [ + 511, + 770, + 759, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method against baselines to demonstrate our clear advantage in complex edits, as shown in Figs. 5 and 6. For the comparison, we display editing results using SDS with MVDream, Instruct-NeRF2NeRF, and PDS. Additionally, we compare our strategy with dataset update and regeneration strategies in Fig. 7.", + "bbox": [ + 511, + 792, + 905, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different from other methods, SDS exhibits blurry tex", + "bbox": [ + 531, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16299", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99b2f56980c3c891998109f9a90e0b7530343eb6f28c2e584710c27736cdde53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 88, + 192, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d3567d28f850c6fe854e7a44446b2948b1db22471d281dadb81e9a779f6bcd8.jpg", + "image_caption": [ + "\"... a cactus plant with flowers\"" + ], + "image_footnote": [], + "bbox": [ + 200, + 88, + 482, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65d0e7d37aacfd062ade932c86e92458aac049d20e4a15374913a4e451d1eb33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 208, + 192, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4bf341fc0340139126be6c2b353c56bc2df81ec3430ca94baae7b15366ed46b8.jpg", + "image_caption": [ + "\"… with windmills\"" + ], + "image_footnote": [], + "bbox": [ + 200, + 208, + 482, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f8d30c525a4522cd2c765500c7abf13a5dfee4aac830711660de2694c09918c.jpg", + "image_caption": [ + "NeRF2NeRF", + "Figure 9. Real scene editing results. An asterisk (*) indicates we use the exact same update rule and schedule." + ], + "image_footnote": [], + "bbox": [ + 94, + 316, + 191, + 415 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/925484d9a959e348142ed309dd5a217e4d99312fd116bfa52866e8912b11e3e1.jpg", + "image_caption": [ + "... with a snowman", + "Posterior", + "Perturb-and", + "lation\\*", + "Revise (Ours)*" + ], + "image_footnote": [], + "bbox": [ + 200, + 316, + 482, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tures and lacks the capability to perform complex edits. Although it can handle simple changes, it inadvertently changes the appearance and texture of the source objects and cannot perform edits that need extensive geometric changes, as seen in the 3rd, 4th, and 5th rows of Fig. 5. While Instruct-NeRF2NeRF successfully achieves simple color edits or symmetric edits, we observe that even when the editing process converges, it is not able to address geometric changes in the objects, such as a change in the pose. PDS, while maintaining similarity to the source object, is unable to make proper edits according to the prompt and cannot deviate far from local minima. As shown in Figs. 5, 6, and 7, our method is capable of various types of edits, including those involving large geometric changes, achieving state-of-the-art results for editing color, appearance, and geometry in a flexible and consistent manner.", + "bbox": [ + 88, + 489, + 482, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have to consider both the similarity to the source object and faithfulness to the edit prompt. To this end, we evaluate these aspects using CLIP directional similarity [8] across ViT-B/32, ViT-B/16, and ViT-L/14 [42] for measuring faithfulness to the edit prompts and LPIPS [65] across VGG [48] and AlexNet [26] for evaluating the similarity between original and edited objects. The results are shown in Table 1. We can see that our method excels at both metrics and achieves a better trade-off than other baselines. This, along with the quantitative evaluation, corroborates the effectiveness of our approach.", + "bbox": [ + 88, + 734, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Real Scene Editing", + "text_level": 1, + "bbox": [ + 511, + 90, + 694, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our parameter perturbation approach described in Sec. 4.1 can be readily extended to real scene editing scenarios [8, 25]. We present real scene editing results in Figs. 2 and 9, building upon Posterior Distillation [25]. Using the same number of iterations, we can better modify the scene geometry through parameter perturbation. We refer the readers to the supplementary material for details.", + "bbox": [ + 511, + 112, + 906, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 229, + 666, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameter perturbation. In Fig. 4, we display the rendered images based on the amount of parameter perturbation and optimization steps. This demonstrates that parameter perturbation facilitates easy alteration of the source object's structure while keeping crucial parts unchanged. Additionally, we present a quantitative study of $\\eta$ selection in Fig. 8, where we categorize edits into five categories: color, pattern, shape, pose, and object edits. We observe that pose changes and object additions generally require a larger degree of versatility and thus need a larger value of $\\eta$ to achieve maximum performance. Furthermore, we demonstrate that, when averaged across all types of edits, the adaptive $\\eta$ selection approach outperforms strategies using fixed $\\eta \\in \\{0.0, 0.2, 0.4, 0.6, 0.8\\}$ while maintaining high experimental success rates across various edit categories and significantly reducing computational costs compared to grid search. This indicates that our method dynamically selects an effective $\\eta$ value while remaining robust in terms of optimization. We present additional metrics in the supplementary material.", + "bbox": [ + 511, + 252, + 906, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Identity-preserving gradients. We demonstrate the effect of identity-preserving gradients. In Table 2, we present diverse metrics. Our results show that the refinement steps with IPG achieve a significant decrease in LPIPS while decreasing CLIP directional similarity, due to their trade-off relationship. In addition, we observe an overall increase in CLIP directional consistency [8].", + "bbox": [ + 511, + 563, + 908, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 684, + 633, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present Perturb-and-Revise (PnR), a framework for text-guided 3D object editing. By introducing adaptive parameter perturbation and identity-preserving gradients, our method enables extensive geometric and appearance changes that adhere to the text prompt while also maintaining fidelity to source objects. Our experiments corroborate that PnR achieves state-of-the-art results across diverse editing tasks without requiring model retraining or multiple input images. We discuss various limitations of our method in the supplementary material for future research.", + "bbox": [ + 511, + 710, + 906, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by the UW Reality Lab and Google.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16300", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. CVPR, 2023. 2", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. CVPR, 2023. 1, 2, 3", + "[3] Changyou Chen, Ruiyi Zhang, Wenlin Wang, Bai Li, and Liquun Chen. A unified particle-optimization framework for scalable bayesian sampling. arXiv preprint arXiv:1805.11659, 2018. 3", + "[4] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3, 6, 7", + "[5] Hanze Dong, Xi Wang, Yong Lin, and Tong Zhang. Particle-based variational inference with preconditioned functional gradient flow. arXiv preprint arXiv:2211.13954, 2022. 3", + "[6] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you should treat it like one. ICML, 2022. 2, 4", + "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2", + "[8] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 1, 2, 3, 5, 6, 7, 8", + "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2", + "[10] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[11] Susung Hong. Smoothed energy guidance: Guiding diffusion models with reduced energy curvature of attention. arXiv preprint arXiv:2408.00760, 2024. 2", + "[12] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debiasing scores and prompts of 2d diffusion for view-consistent text-to-3d generation. Advances in Neural Information Processing Systems, 36:11970-11987, 2023. 2, 3", + "[13] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 2", + "[14] Susung Hong, Junyoung Seo, Heeseong Shin, Sunghwan Hong, and Seungryong Kim. Direct2v: Large language models are frame-level directors for zero-shot text-to-video generation. arXiv preprint arXiv:2305.14330, 2023. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. arXiv preprint arXiv:2309.14494, 2023. 2", + "[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3", + "[17] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3", + "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 1", + "[19] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5", + "[20] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023. 2", + "[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 2, 3", + "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 3", + "[23] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2", + "[24] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 3", + "[25] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13352-13361, 2024. 1, 2, 3, 5, 6, 7, 8", + "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 8", + "[27] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2, 3", + "[28] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016. 3" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16301", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 2, 3", + "[30] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5773-5783, 2021. 3", + "[31] Chang Liua and Jun Zhub. Geometry in sampling methods: A review on manifold mcmc and particle-based variational inference methods. Advancements in Bayesian Methods and Implementations, 47:239, 2022. 3", + "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7210-7219, 2021. 3", + "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 2, 3", + "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 5", + "[35] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2", + "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 7", + "[37] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 3", + "[38] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 3", + "[39] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3", + "[40] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 5, 6", + "[41] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3, 8", + "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 3", + "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3", + "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[46] Junyoung Seo, Susung Hong, Wooseok Jang, Ines Hyeonsu Kim, Minseop Kwak, Doyup Lee, and Seungryong Kim. Retrieval-augmented score distillation for text-to-3d generation. arXiv preprint arXiv:2402.02972, 2024. 2", + "[47] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 5, 6, 7", + "[48] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 8", + "[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2", + "[50] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 32, 2019. 2", + "[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 2", + "[52] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 3", + "[53] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), pages 443-453. IEEE, 2022. 3", + "[54] Cyrus Vachha and Ayaan Haque. Instruct-gs2gs: Editing 3d gaussian splats with instructions, 2024. 2", + "[55] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields." + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16302", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 3", + "[56] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 3", + "[57] Can Wang, Ruixiang Jiang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 3", + "[58] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 1, 2, 3", + "[59] Ziyu Wang, Tongzheng Ren, Jun Zhu, and Bo Zhang. Function space particle optimization for bayesian neural networks. arXiv preprint arXiv:1902.09754, 2019. 3", + "[60] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 6", + "[61] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7378-7387, 2023. 3", + "[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2", + "[63] Hong-Xing Yu, Leonidas J Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. arXiv preprint arXiv:2107.07905, 2021. 3", + "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2", + "[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8", + "[66] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4606-4615, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 810 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16303", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/11693e39-cb59-4d3c-8f05-aea184580288_model.json b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/11693e39-cb59-4d3c-8f05-aea184580288_model.json new file mode 100644 index 0000000000000000000000000000000000000000..33eb3d25c206266e893df2b5511743dbd338fca1 --- /dev/null +++ b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/11693e39-cb59-4d3c-8f05-aea184580288_model.json @@ -0,0 +1,2521 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.131, + 0.852, + 0.154 + ], + "angle": 0, + "content": "Perturb-and-Revise: Flexible 3D Editing with Generative Trajectories" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.18, + 0.87, + 0.235 + ], + "angle": 0, + "content": "Susung Hong\\(^{1}\\) Johanna Karras\\(^{1}\\) Ricardo Martin-Brualla\\(^{2}\\) Ira Kemelmacher-Shlizerman\\(^{1,2}\\) \n\\(^{1}\\)University of Washington \\(^{2}\\)Google Research" + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.251, + 0.862, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.455, + 0.907, + 0.484 + ], + "angle": 0, + "content": "Figure 1. Perturb-and-Revise takes a source NeRF and an edit prompt as input and produces the edited result through: (1) versatile initialization via parameter perturbation, (2) multi-view consistent score distillation, and (3) refinement with the identity-preserving gradient." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.495, + 0.327, + 0.511 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.528, + 0.486, + 0.801 + ], + "angle": 0, + "content": "Recent advancements in text-based diffusion models have accelerated progress in 3D reconstruction and text-based 3D editing. Although existing 3D editing methods excel at modifying color, texture, and style, they struggle with extensive geometric or appearance changes, thus limiting their applications. To this end, we propose Perturb-and-Revise, which makes possible a variety of NeRF editing. First, we perturb the NeRF parameters with random initializations to create a versatile initialization. The level of perturbation is determined automatically through analysis of the local loss landscape. Then, we revise the edited NeRF via generative trajectories. Combined with the generative process, we impose identity-preserving gradients to refine the edited NeRF. Extensive experiments demonstrate that Perturb-and-Revise facilitates flexible, effective, and consistent editing of color, appearance, and geometry in 3D. For \\(360^{\\circ}\\) results, please visit our project page: https://susunghong.github.io/Perturb-and-Revise." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.222, + 0.845 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRFs) [34] have revolutionized the creation of high-quality 3D scenes, marking a significant advancement in 3D reconstruction technology. Beyond" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.907, + 0.543 + ], + "angle": 0, + "content": "their initial applications, NeRFs have enabled flexible generation of 3D content from models trained solely on image-text pairs [18, 40] through score distillation [25, 40, 58]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.546, + 0.909, + 0.699 + ], + "angle": 0, + "content": "However, editing 3D content—an essential aspect of refinement and customization—remains a time-consuming and labor-intensive process across various industries, including animation, manufacturing, design, and gaming. This challenge is particularly pronounced with NeRFs, where color and density attributes are intricately encoded within their parameters. Consequently, there is an ongoing need for more intuitive and universally accessible tools for editing NeRFs, leveraging user-friendly interfaces such as text prompts for broader applicability and faster production." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.7, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Fortunately, recent innovations in text-conditioned diffusion models have enabled significant advances in 3D editing using natural language prompts. For instance, leveraging diffusion models [2, 43], recent works such as Instruct-NeRF2NeRF [8] and Posterior Distillation [25] propose methods that use text prompts to modify 3D scenes. While these methods excel in altering the color or style of an object, they struggle with edits involving significant geometric or appearance changes, limiting their applications." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To overcome these limitations, we propose a 3D object editing framework called Perturb-and-Revise (PnR), which leverages existing 3D generation methods for editing in a flexible and natural way (Figs. 1 and 2). Drawing on the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16293" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.262, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.274, + 0.22, + 0.285 + ], + "angle": 0, + "content": "Source 3D Scene" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.089, + 0.429, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.275, + 0.406, + 0.285 + ], + "angle": 0, + "content": "Instruct-NeRF2NeRF" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.089, + 0.586, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.275, + 0.553, + 0.285 + ], + "angle": 0, + "content": "Instruct-GS2GS" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.089, + 0.745, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.275, + 0.723, + 0.285 + ], + "angle": 0, + "content": "Posterior Distillation*" + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.09, + 0.904, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.275, + 0.895, + 0.285 + ], + "angle": 0, + "content": "Perturb-and-Revise (Ours)*" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.287, + 0.907, + 0.328 + ], + "angle": 0, + "content": "Figure 2. We qualitatively compare our method to Instruct-NeRF2NeRF [8], Instruct-GS2GS [54], and Posterior Distillation [25]. In this task, the goal is to edit the source \"bear\" into a \"polar bear.\" Compared to related works, our approach reconstructs a more realistic face and better overall geometry. An asterisk (*) denotes that we use an identical update rule and schedule." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.336, + 0.483, + 0.517 + ], + "angle": 0, + "content": "notion that NeRF parameters optimized with score distillation can be considered as a particle [60] or data point [6], we propose a novel method that leverages perturbation at the parameter level to perform text-based edits requiring various changes. Specifically, we construct versatile NeRF initialization by interpolating the source NeRF and a random NeRF initialization. Intuitively, perturbation in the parameter space helps the particle escape local minima and facilitates its following of the natural trajectory of the generative ODE towards the distribution of the desired edit (Fig. 3), allowing for a wide range of challenging edits including changing the pose and introducing new objects." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.518, + 0.483, + 0.698 + ], + "angle": 0, + "content": "Additionally, to determine the amount of perturbation needed to escape the basin of attraction without costly searching algorithms, we propose an algorithm based on the loss landscape near the source parameters, by simulating a few score distillation steps ahead of the parameter perturbation and optimization. After the early step of our framework, to achieve more similarity to the original object and refine the quality of the result, we employ the Identity-Preserving Gradient (IPG) to increase the fidelity of the edited NeRF to the source NeRF. As a result, the output closely resembles the source while still maintaining fidelity to the intended edits." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.698, + 0.483, + 0.775 + ], + "angle": 0, + "content": "We extensively evaluate our method on 3D fashion objects, as well as general objects in Objaverse [4], on a variety of appearance- and geometry-based edits. Furthermore, we demonstrate that our method achieves state-of-the-art results on various 3D editing baselines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.233, + 0.801 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models [9, 49], which are closely associated with score-based models [21, 50, 51], have recently shown remarkable sample quality in image synthesis. Latent Diffusion Models (LDM) [43], which perform the diffusion forward and backward process in the latent space [7], have demonstrated their efficacy and genera" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.336, + 0.878, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.504, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Figure 3. Conceptual figure. The target distribution in the figures represents the conditional distribution of NeRF parameters relative to the edit prompt, and \\(\\mathcal{P}(\\Theta)\\) denotes the distribution of randomly initialized NeRF parameters. First, parameter perturbation enables the parameters to escape from local minima and follow a natural generative path. Subsequently, during the refining process, the tug-of-war between two vectors, \\(\\lambda_{\\mathrm{d}}\\nabla_{\\theta}d(\\theta_{\\tau},\\theta_{\\mathrm{src}})\\) (the red arrow) and \\(d\\theta_{\\tau}\\) (the blue arrow), pushes the actual parameters into a region that is closer to either the source parameters or the high-density region specified by the edit prompt, following the purple arrow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.907, + 0.803 + ], + "angle": 0, + "content": "tion quality. This trend has been scaled up by training diffusion models on large-scale text-image datasets [45] for text-conditional generation. The prosperity of text-to-image diffusion models has led to their application in complex content creation, such as 3D scenes and videos [1, 10, 12, 14, 15, 23, 27, 33, 40, 46, 58, 60, 62]. Recently, many adapting methods of text-to-image diffusion models have been proposed to achieve either generation with additional conditions [2, 11, 13, 20, 35, 64, 66] or to maintain consistency needed for videos [14, 23, 41, 62] or 3D scenes [29, 47]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Score distillation. Score distillation [40, 58, 60], a recent trend in synthesizing underlying parameterizations, mostly in 3D scene generation, uses only score-based models [43, 44] operating on 2D data. This approach has revolutionized text-to-3D generation by introducing an elegant method that does not require intense training on a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16294" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "3D dataset. It enables the rich and flexible knowledge of 2D text-to-image models to be transferred to 3D generation [12, 27, 33, 58, 60]. However, due to bias in 2D models, a multi-view consistency problem has arisen, necessitating further analysis and generalization of the original method [12, 29, 47]. Another limitation of score distillation is its time-consuming optimization, with the controllability and editing capability remaining underexplored." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.247, + 0.485, + 0.671 + ], + "angle": 0, + "content": "Text-based NeRF editing. In recent years, neural radiance fields (NeRF) [34] have emerged as a groundbreaking approach for generating photorealistic novel views of a scene captured in photographs, and have been extended in many follow-up works [22, 32, 36, 39]. In response to these advancements, several techniques for modifying NeRF have been proposed. Traditional methodologies for editing NeRF include the alteration of materials and lighting [37, 52, 55], as well as the spatial manipulation of objects via bounding box frameworks [38, 63]. Additionally, there has been exploration in the stylization of NeRFs, including Edit-NeRF [30], Clip-NeRF [56], and NeRF-Art [57], as well as in distilling 2D features into radiance fields, such as in Distilled Feature Fields [24] and Neural Feature Fusion Fields [53], which enables nuanced, guided edits based on language or images. Crucially, Instruct-NeRF2NeRF [8] focuses on user-friendly, language-based editing commands, using an instruction-driven, 2D image-conditioned diffusion model [2] for more intuitive and context-aware 3D editing. However, even in a trend using 2D language-based models such as CLIP [42] or text-to-image diffusion models [2, 43, 44], these methodologies are unable to control large geometrical changes, such as changing the pose or introducing a massive object. Posterior Distillation [25], instead of naively matching the noise, proposes a way to use a loss that matches the stochastic latents [16, 61] of source and target images. Still, this method falls into issues such as slower convergence and large computational costs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.699, + 0.22, + 0.716 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Diffusion models are a type of generative model that iteratively restores an image from Gaussian noise. The training objective of diffusion models involves training a neural network to estimate the score for a noised image \\(y\\) given the noise level \\(\\sigma\\) via denoising score matching [17]. Following the preconditioning convention [21], a diffusion model, parameterized by \\(\\phi\\), can be interpreted as a denoiser \\(D_{\\phi}(y;\\sigma)\\) which minimizes the weighted L2 loss across different \\(\\sigma\\) values in data samples." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In the context of text-based 3D scene generation, Score Distillation Sampling (SDS) [40] updates the 3D scene rep" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.898, + 0.108 + ], + "angle": 0, + "content": "resentation given a text caption \\(c\\) using the following rule:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.116, + 0.906, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , c, x = g (\\theta , \\psi)) = \\\\ - \\mathbb {E} _ {\\sigma \\sim \\Sigma , n \\sim \\mathcal {N} (0, \\sigma^ {2} I)} \\left[ \\omega (\\sigma) \\left(D _ {\\phi} (x + n; \\sigma , c) - x\\right) \\frac {\\partial x}{\\partial \\theta} \\right]. \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.906, + 0.251 + ], + "angle": 0, + "content": "In this equation, \\( g(\\cdot) \\) is the differentiable renderer, \\( \\psi \\) is the random camera pose, \\( \\Sigma \\) is a predefined distribution from which the noise level for the denoiser is sampled, and \\( \\omega(\\cdot) \\) denotes the weighting function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.907, + 0.313 + ], + "angle": 0, + "content": "SDS can be viewed as a form of particle-based variational inference [3, 5, 28, 31, 59, 60]. In this context, the update rule can be expressed as a generative ODE over an optimization step \\(\\tau\\), derived via Wasserstein gradient flow [3]:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.32, + 0.912, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\frac {d \\theta_ {\\tau}}{d \\tau} = - \\mathbb {E} _ {\\sigma \\sim \\Sigma , n \\sim \\mathcal {N} (0, \\sigma^ {2} I)} \\left[ \\omega (\\sigma) \\left(D _ {\\phi} (x + n; \\sigma , c) - x\\right) \\frac {\\partial x}{\\partial \\theta_ {\\tau}} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.365, + 0.895, + 0.379 + ], + "angle": 0, + "content": "For further details, please see the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.906, + 0.501 + ], + "angle": 0, + "content": "As background, multi-view diffusion models [47], denoted as \\(\\{D_{\\phi}^{\\mathrm{M}}(g(\\theta ,\\psi_i) + n;\\sigma ,c,\\psi_i)\\}_{i = 1}^{N}\\), consistently generate multi-view images simultaneously. Here, \\(N\\) denotes the total number of multi-views generated simultaneously, and \\(\\psi_{i}\\) denotes the \\(i\\)-th viewpoint. Note that for \\(N = 1\\), these are ordinary diffusion models. For \\(N > 1\\), they are implemented by adapting the attention module to attend to different frames [47] and by training on 3D assets [4]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.512, + 0.605, + 0.528 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.629 + ], + "angle": 0, + "content": "In text-based editing scenarios, we assume that a user provides a text prompt, \\( c_{\\mathrm{edit}} \\), and a source 3D object represented by the parameterization \\( \\theta_{\\mathrm{src}} \\). The aim is to modify the 3D object into a new form \\( \\theta^{*} = \\mathcal{E}(\\theta_{\\mathrm{src}}, c_{\\mathrm{edit}}) \\), ensuring that the transformation aligns with the text prompt and retains a resemblance to the original 3D object." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.781 + ], + "angle": 0, + "content": "At a high level, we leverage the generative ODE of 3D objects through novel parameter perturbation while preserving proximity to the source object. In Sec. 4.1, we introduce perturbation in parameter space that enables the parameters to escape local minima and converge to natural generative trajectories. In Sec. 4.2, we propose a novel algorithm to determine the degree of perturbation by analyzing the local loss landscape. Finally, in Sec. 4.3, we propose balancing fidelity between the text prompt and the source object to place the final output in the desired region." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.788, + 0.906, + 0.821 + ], + "angle": 0, + "content": "4.1. Parameter Perturbation for Flexible Object Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Given the generative ODE in Eq. 2, which minimizes the energy functional between distributions, a fully-optimized NeRF of a 3D object is considered to exhibit relatively low energy. This state corresponds to the particle residing in a local minimum, which remains stable despite alterations in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16295" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "the text prompts, as it is still capable of producing realistic renderings. As exemplified in the first row of Fig. 4, naively using \\(\\theta_{\\mathrm{src}}\\) as initialization and making independent predictions on 2D noisy images with diffusion models are insufficient to handle the consistent changes in overall appearance and geometry that occur in a 3D object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Interpreting neural fields as a particle [60] or a data point [6], we propose parameter-level perturbation for an editing task, which facilitates changes in \\(\\theta_{\\mathrm{src}}\\) by enabling the editing particle ODE to hijack the natural, coarse-to-fine generative process in Eq. 2 that the parameters are likely to follow. In other words, since adding more noise to the parameters is equivalent to undoing more of the optimization process, this approach enables the parameters to be flexible, allowing for significant changes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.483, + 0.395 + ], + "angle": 0, + "content": "Specifically, let \\(\\mathcal{P}(\\Theta_0)\\) denote the probability density function over \\(\\Theta_0\\), which is a set of random parameterizations. Further, assume that we have a fully optimized 3D scene, \\(\\theta_{\\mathrm{src}}\\). Then, we obtain the perturbed version of \\(\\theta_{\\mathrm{src}}\\) by using linear interpolation:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.404, + 0.483, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\text {p e r t u r b e d}} = \\operatorname {L e r p} \\left(\\theta_ {\\text {s r c}}, \\theta_ {0}, \\eta\\right) \\quad \\text {w h e r e} \\quad \\theta_ {0} \\sim \\mathcal {P} \\left(\\Theta_ {0}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.483, + 0.521 + ], + "angle": 0, + "content": "where \\(\\eta \\in [0,1]\\) denotes the perturbation amount and \\(\\mathrm{Lerp}(\\theta_{\\mathrm{src}},\\theta_0,\\eta) = (1 - \\eta)\\cdot \\theta_{\\mathrm{src}} + \\eta \\cdot \\theta_0\\). Note that \\(\\eta = 1\\) indicates \\(\\theta_{\\mathrm{perturbed}}\\) is the random parameters. Theoretically, the interpolation makes the distribution of perturbed particles similar to the initializing distribution of NeRF, making it versatile. See the supplementary material for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Subsequently, we run the parameter ODE for textual editing, starting from \\(\\theta_{\\mathrm{perturbed}}\\), using a new prompt \\(c_{\\mathrm{edit}}\\). However, we observe that adopting ordinary single-view diffusion models (when \\(N = 1\\)) poses ambiguity when an edit introduces asymmetry or adds a new object. Therefore, we execute multi-view consistent updates as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.622, + 0.485, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\frac {d \\theta_ {\\tau}}{d \\tau} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {E} _ {\\sigma , n} \\left[ \\left(D _ {\\phi} ^ {\\mathrm {M}} \\left(z _ {i} + n; \\sigma , c _ {\\mathrm {e d i t}}, \\psi_ {i}\\right) - z _ {i}\\right) \\frac {\\partial z _ {i}}{\\partial \\theta_ {\\tau}} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.719 + ], + "angle": 0, + "content": "where we define \\(z_{i} \\coloneqq g(\\theta_{\\tau}, \\psi_{i})\\) and omit the weighting factor for brevity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The conceptual illustration of parameter perturbation can be found in Fig. 3. This approach essentially reverts the optimization process by interpolating between the initial state and the fully optimized state. Consequently, the particles are relocated to a less-optimized state characterized by a more moderate energy landscape. This repositioning allows the particle ODE (Eq. 4) to intervene in the generative process with a new text prompt \\( c_{\\mathrm{edit}} \\), guiding the parameter trajectory. As shown in Fig. 4, increasing the noise level in the parameters inversely correlates with the degree of reversion of the optimization, allowing for more radical modifications in the model outputs." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.092, + 0.902, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.304, + 0.897, + 0.315 + ], + "angle": 0, + "content": "100 steps 200 steps 300 steps 500 steps 700 steps 900 steps 10000 steps" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.315, + 0.905, + 0.385 + ], + "angle": 0, + "content": "Figure 4. Effect of parameter perturbation. In this example, we aim to make a NeRF model of a standing person sit down using the word \"sitting.\" The scene converges quickly even with large perturbations \\((\\eta = 0.6)\\), while complete regeneration yields blurry rendering results given the same number of optimization steps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.392, + 0.892, + 0.409 + ], + "angle": 0, + "content": "4.2. Determining \\(\\eta\\) by Analyzing Loss Landscape" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.906, + 0.566 + ], + "angle": 0, + "content": "In Sec. 4.1, we see that \\(\\eta\\) controls parameter versatility by determining how similar the distribution of parameters is to the distribution of random NeRF initializations. The required degree of versatility depends on both the type of edits and the current source NeRF parameters. However, finding an optimal \\(\\eta\\) through standard mechanisms, such as grid or random search, is computationally expensive. Therefore, we propose a novel algorithm that explores the basin of attraction surrounding local minima to determine \\(\\eta\\) by analyzing the loss landscape." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Specifically, we leverage the loss function as a proxy to measure the depth and volume of the basin of attraction. Interestingly, for some examples, we observe that the loss function even increases during optimization steps, supporting our claim in Sec. 4.1 that the parameters are in a low-energy state. To address this, we simulate several optimization steps with \\( c_{\\mathrm{edit}} \\) and calculate the total loss decrease prior to parameter perturbation. For algorithmic stability, we compute the difference between the averages of the first few steps and the last few steps. We then determine \\( \\eta(\\theta_{\\mathrm{src}}, c_{\\mathrm{edit}}) \\) based on the source parameters \\( \\theta_{\\mathrm{src}} \\) and the text prompt \\( c_{\\mathrm{edit}} \\) using an inverted exponential decay function. Subsequently, we apply parameter perturbation in Eq. 3 with \\( \\eta = \\eta(\\theta_{\\mathrm{src}}, c_{\\mathrm{edit}}) \\). The complete algorithms for our parameter perturbation and \\( \\eta \\) selection are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.822, + 0.835 + ], + "angle": 0, + "content": "4.3. Identity-Preserving Gradient (IPG)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "While our parameter perturbation in Sec. 4.1 successfully enables flexible changes, there are still some estimation errors or biases arising from the diffusion model during this phase. On the other hand, imposing constraints that make" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16296" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.104, + 0.877, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.552, + 0.253, + 0.564 + ], + "angle": 0, + "content": "Source 3D Object" + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.552, + 0.434, + 0.562 + ], + "angle": 0, + "content": "Score Distillation" + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.552, + 0.582, + 0.562 + ], + "angle": 0, + "content": "Posterior Distillation" + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.552, + 0.721, + 0.562 + ], + "angle": 0, + "content": "Instruct-NeRF2NeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.552, + 0.875, + 0.563 + ], + "angle": 0, + "content": "Perturb-and-Revise (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.563, + 0.907, + 0.646 + ], + "angle": 0, + "content": "Figure 5. Baseline comparisons with a wide range of edits. We compare our method with Score Distillation Sampling (SDS) [40], Posterior Distillation Sampling (PDS) [25], and Instruct-NeRF2NeRF [8]. For SDS and PDS, we use MVDream [47] as the backbone for fair comparison. SDS alters the appearance and texture of the source objects and is unable to handle edits that require extensive geometric changes (3rd, 4th, and 5th rows). PDS is not capable of making significant edits and cannot deviate far from local minima due to its high preservation term from the start [25]. While Instruct-NeRF2NeRF changes the texture of objects as desired, it cannot address geometric changes. In contrast, our method is capable of various types of edits, including those involving large geometric changes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.745 + ], + "angle": 0, + "content": "it similar to the source object (e.g., L2 distance) conflicts with the generative ODE, making the optimization challenging. To counteract the estimation error and to circumvent the conflict, we introduce the Identity-Preserving Gradient (IPG) term that is added at later steps to the editing gradient presented in Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Initially, we assume that the ideal parameters are closer to the source parameters and the high-density region of the target distribution. Then, as conceptually shown in Fig. 3, we instigate a tug-of-war between two vectors: one symbolizes the velocity consistently pointing towards the high-density region as dictated by the edit prompt, and the other symbolizes the velocity towards \\(\\theta_{\\mathrm{src}}\\). Specifically, the velocity in Eq. 4 propels the particle towards the region with a high likelihood given the text \\(c_{\\mathrm{edit}}\\), while the other gradient represents the pull towards the original NeRF parameteri" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.907, + 0.729 + ], + "angle": 0, + "content": "zation \\(\\theta_{\\mathrm{src}}\\). This effectively corrects the shift caused by the error in the editing process, ensuring it resembles the original while maintaining the intended edits. To this end, we extend the optimization process and compute IPGs during these additional steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.907, + 0.76 + ], + "angle": 0, + "content": "Formally, we combine the result of Eq. 4 with the IPG to define a refinement step as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.769, + 0.907, + 0.787 + ], + "angle": 0, + "content": "\\[\nd \\theta_ {\\tau} ^ {\\text {r e f i n e}} = d \\theta_ {\\tau} + \\lambda_ {\\mathrm {d}} \\nabla_ {\\theta} d \\left(\\theta_ {\\tau}, \\theta_ {\\text {s r c}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(d(\\cdot, \\cdot)\\) is a similarity metric. In practice, we observe that the combination of \\(L_{1}\\) and perceptual loss [19], which is also the preferred choice in NeRF [34] training, is more robust and less susceptible to noise than using \\(L_{2}\\) loss. Thus, given a random camera pose \\(\\psi\\), we use a combination of L1 distance and perceptual similarity between the rendered images \\(g(\\theta_{\\tau}, \\psi)\\) and \\(g(\\theta_{\\mathrm{src}}, \\psi)\\), weighted by \\(\\lambda_{\\mathrm{L1}}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16297" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.101, + 0.862, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.415, + 0.255, + 0.427 + ], + "angle": 0, + "content": "Source 3D Object" + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.415, + 0.414, + 0.426 + ], + "angle": 0, + "content": "Score Distillation" + }, + { + "type": "image_caption", + "bbox": [ + 0.45, + 0.415, + 0.559, + 0.426 + ], + "angle": 0, + "content": "Posterior Distillation" + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.415, + 0.7, + 0.426 + ], + "angle": 0, + "content": "Instruct-NeRF2NeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.415, + 0.853, + 0.427 + ], + "angle": 0, + "content": "Perturb-and-Revise (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.427, + 0.796, + 0.44 + ], + "angle": 0, + "content": "Figure 6. Baseline comparisons of editing various general 3D objects from the Objaverse dataset [4]." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.448, + 0.868, + 0.554 + ], + "angle": 0, + "content": "
MetricScore Distillation [40]Posterior Distillation [25]Instruct-NeRF2NeRF [8]Perturb-and-Revise (Ours)
CLIP-Dir-SimViT-B/32↑0.04800.02870.05830.0594
CLIP-Dir-SimViT-B/16↑0.04180.03040.05490.0534
CLIP-Dir-SimViT-L/14↑0.04150.02640.05390.0567
CLIP-Dir-Simaveraged↑0.04380.02850.05570.0565
LPIPSvgg↓0.12730.03370.10650.1060
LPIPSalex↓0.15330.02150.11120.1034
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.556, + 0.907, + 0.613 + ], + "angle": 0, + "content": "Table 1. Comparison of different methods for fashion object editing. The best, second-best, and worst values are highlighted in bold, underlined, and gray, respectively. Values represent averages across all edit types and prompts. Our approach achieves a better trade-off between faithfulness to edit prompts and preservation of source 3D objects. While PDS exhibits lower LPIPS, it mostly generates edited objects that are nearly identical to the source, as evidenced by CLIP directional similarity." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.62, + 0.228, + 0.636 + ], + "angle": 0, + "content": "and \\(\\lambda_{\\mathfrak{p}}\\) , respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.661, + 0.282, + 0.679 + ], + "angle": 0, + "content": "4.4. Timestep Annealing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In text-based 3D generation, some recent works have revealed that using timestep annealing, which adjusts the noise level added to the 2D rendered images according to the global optimization step, boosts the quality of the results [47, 60]. In the editing task, while we use a smaller noise level at the start, we also identify the annealing method as crucial for maintaining the original quality of the 3D object while reducing blurry textures. To this end, we regard \\(\\Sigma\\) as a function that depends on the editing step \\(\\tau\\), \\(\\Sigma(\\tau)\\), from which the noise level \\(\\sigma\\) is sampled. Specifically, for \\(\\tau \\leq T\\), where \\(T\\) is the final step of the schedule, \\(\\Sigma(\\tau)\\) is gradually annealed from \\(\\mathcal{U}(\\sigma_{\\min}^{0}, \\sigma_{\\max}^{0})\\) to \\(\\mathcal{U}(\\sigma_{\\min}^{\\mathrm{T}}, \\sigma_{\\max}^{\\mathrm{T}})\\). This approach initially facilitates low-frequency edits, followed by fine-grained ones." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.619, + 0.645, + 0.635 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.795 + ], + "angle": 0, + "content": "As baselines, we adopt two distillation-based methods: Score Distillation Sampling (SDS) [40] and Posterior Distillation Sampling (PDS) [25]. Bolded red in the prompts indicates the intended edits. We use the exact weighting factors for the losses as originally proposed in [25, 40]. For fair comparison, we use MVDream [47] as their backbones. Also, we use Instruct-NeRF2NeRF [8] as our baseline for the iterative dataset update strategy. We use several backbones for each metric and note them in the subscripts, while averaged refers to the average value across all backbones." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.691, + 0.82 + ], + "angle": 0, + "content": "5.1.3D Object Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Fashion object editing. In our fashion object editing experiment, we use MVDream [47] to generate a synthetic dataset of 3D fashion objects and perform edits using our framework. Specifically, we categorize the fashion object edits into color, pattern, shape, pose, and object edits, and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16298" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.101, + 0.868, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.281, + 0.266, + 0.293 + ], + "angle": 0, + "content": "Source 3D Object" + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.281, + 0.471, + 0.291 + ], + "angle": 0, + "content": "Instruct-NeRF2NeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.281, + 0.622, + 0.291 + ], + "angle": 0, + "content": "MVDream" + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.281, + 0.851, + 0.292 + ], + "angle": 0, + "content": "Perturb-and-Revise (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.293, + 0.895, + 0.307 + ], + "angle": 0, + "content": "Figure 7. Comparisons with Instruct-NeRF2NeRF (dataset update) [8] and MVDream (regeneration) [47] in editing Objaverse objects." + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.315, + 0.432, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.464, + 0.285, + 0.475 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.315, + 0.871, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.464, + 0.662, + 0.475 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.475, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Figure 8. Ablation study on the selection of \\(\\eta\\). (a) and (b) show the CLIP direction similarity and the percentage of successful experiments (without errors) for different \\(\\eta\\) values, respectively. When averaged across all types of edits, our adaptive method achieves near-maximum performance on these metrics compared to all fixed \\(\\eta\\) values." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.483, + 0.614 + ], + "angle": 0, + "content": "synthesize a total of 150 editing examples, with 30 for each type of edit. In this experiment, our framework takes as input a source object in NeRF with the desired edits to its characteristics, such as color, pattern, shape, pose, and added objects. Note that, unlike some previous work [25], we do not require a description of the original object." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.616, + 0.483, + 0.766 + ], + "angle": 0, + "content": "The qualitative results, as shown in Figs. 4 and 5, demonstrate the versatility introduced by parameter perturbation and our method's capability to perform various types of edits, respectively. In Fig. 5, we apply five types of edits: color, pattern, shape, pose, and object additions. We show that our framework supports a wide range of edits, including those requiring pose changes and introducing objects, corroborating its capability for significant geometric modifications, which have rarely been addressed in previous work [8, 25]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.854 + ], + "angle": 0, + "content": "General object editing. We show a general object editing capability for our method, with objects from Objaverse [4]. We first optimize InstantNGP [36] with 200 rendered images and camera transforms extracted from an object. Then, we apply our editing method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We display the qualitative results of our method in Fig. 6. In addition to textural changes, our method successfully enables structural changes. In Fig. 7, we also demonstrate our" + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.521, + 0.868, + 0.584 + ], + "angle": 0, + "content": "
MethodCLIP-Dir-Sim↑CLIP-Dir-Con↑LPIPS↓
w/o Refinement0.06240.75720.1147
w/ Refinement0.05650.76420.1047
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.587, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Table 2. Ablation study on the IPG steps. We present average values across all edit types, prompts, and evaluation backbones. Our approach yields a substantial reduction in LPIPS while balancing faithfulness to the source object and edit prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.907, + 0.76 + ], + "angle": 0, + "content": "method's ability to perform creative edits anchored by the source 3D object. Here, to demonstrate flexibility, we use two prompts with different animal motifs. We can see that regenerating the object with MVDream produces results that appear completely different from the source 3D object. In contrast, our method successfully anchors the original shape while changing it to align with the edit prompt." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.771, + 0.76, + 0.787 + ], + "angle": 0, + "content": "5.2. Comparison with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.906, + 0.884 + ], + "angle": 0, + "content": "We compare our method against baselines to demonstrate our clear advantage in complex edits, as shown in Figs. 5 and 6. For the comparison, we display editing results using SDS with MVDream, Instruct-NeRF2NeRF, and PDS. Additionally, we compare our strategy with dataset update and regeneration strategies in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Different from other methods, SDS exhibits blurry tex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16299" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.194, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.089, + 0.483, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.198, + 0.422, + 0.207 + ], + "angle": 0, + "content": "\"... a cactus plant with flowers\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.209, + 0.193, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.209, + 0.483, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.308, + 0.391, + 0.315 + ], + "angle": 0, + "content": "\"… with windmills\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.317, + 0.192, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.318, + 0.483, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.416, + 0.395, + 0.424 + ], + "angle": 0, + "content": "... with a snowman" + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.425, + 0.367, + 0.433 + ], + "angle": 0, + "content": "Posterior" + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.425, + 0.465, + 0.434 + ], + "angle": 0, + "content": "Perturb-and" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.435, + 0.279, + 0.444 + ], + "angle": 0, + "content": "NeRF2NeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.435, + 0.376, + 0.444 + ], + "angle": 0, + "content": "lation\\*" + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.435, + 0.471, + 0.445 + ], + "angle": 0, + "content": "Revise (Ours)*" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.446, + 0.483, + 0.473 + ], + "angle": 0, + "content": "Figure 9. Real scene editing results. An asterisk (*) indicates we use the exact same update rule and schedule." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.483, + 0.732 + ], + "angle": 0, + "content": "tures and lacks the capability to perform complex edits. Although it can handle simple changes, it inadvertently changes the appearance and texture of the source objects and cannot perform edits that need extensive geometric changes, as seen in the 3rd, 4th, and 5th rows of Fig. 5. While Instruct-NeRF2NeRF successfully achieves simple color edits or symmetric edits, we observe that even when the editing process converges, it is not able to address geometric changes in the objects, such as a change in the pose. PDS, while maintaining similarity to the source object, is unable to make proper edits according to the prompt and cannot deviate far from local minima. As shown in Figs. 5, 6, and 7, our method is capable of various types of edits, including those involving large geometric changes, achieving state-of-the-art results for editing color, appearance, and geometry in a flexible and consistent manner." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We have to consider both the similarity to the source object and faithfulness to the edit prompt. To this end, we evaluate these aspects using CLIP directional similarity [8] across ViT-B/32, ViT-B/16, and ViT-L/14 [42] for measuring faithfulness to the edit prompts and LPIPS [65] across VGG [48] and AlexNet [26] for evaluating the similarity between original and edited objects. The results are shown in Table 1. We can see that our method excels at both metrics and achieves a better trade-off than other baselines. This, along with the quantitative evaluation, corroborates the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.696, + 0.108 + ], + "angle": 0, + "content": "5.3. Real Scene Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.22 + ], + "angle": 0, + "content": "Our parameter perturbation approach described in Sec. 4.1 can be readily extended to real scene editing scenarios [8, 25]. We present real scene editing results in Figs. 2 and 9, building upon Posterior Distillation [25]. Using the same number of iterations, we can better modify the scene geometry through parameter perturbation. We refer the readers to the supplementary material for details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.23, + 0.667, + 0.246 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.253, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Parameter perturbation. In Fig. 4, we display the rendered images based on the amount of parameter perturbation and optimization steps. This demonstrates that parameter perturbation facilitates easy alteration of the source object's structure while keeping crucial parts unchanged. Additionally, we present a quantitative study of \\(\\eta\\) selection in Fig. 8, where we categorize edits into five categories: color, pattern, shape, pose, and object edits. We observe that pose changes and object additions generally require a larger degree of versatility and thus need a larger value of \\(\\eta\\) to achieve maximum performance. Furthermore, we demonstrate that, when averaged across all types of edits, the adaptive \\(\\eta\\) selection approach outperforms strategies using fixed \\(\\eta \\in \\{0.0, 0.2, 0.4, 0.6, 0.8\\}\\) while maintaining high experimental success rates across various edit categories and significantly reducing computational costs compared to grid search. This indicates that our method dynamically selects an effective \\(\\eta\\) value while remaining robust in terms of optimization. We present additional metrics in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.909, + 0.672 + ], + "angle": 0, + "content": "Identity-preserving gradients. We demonstrate the effect of identity-preserving gradients. In Table 2, we present diverse metrics. Our results show that the refinement steps with IPG achieve a significant decrease in LPIPS while decreasing CLIP directional similarity, due to their trade-off relationship. In addition, we observe an overall increase in CLIP directional consistency [8]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.685, + 0.634, + 0.7 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.907, + 0.863 + ], + "angle": 0, + "content": "In this paper, we present Perturb-and-Revise (PnR), a framework for text-guided 3D object editing. By introducing adaptive parameter perturbation and identity-preserving gradients, our method enables extensive geometric and appearance changes that adhere to the text prompt while also maintaining fidelity to source objects. Our experiments corroborate that PnR achieves state-of-the-art results across diverse editing tasks without requiring model retraining or multiple input images. We discuss various limitations of our method in the supplementary material for future research." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by the UW Reality Lab and Google." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16300" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. CVPR, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.216, + 0.484, + 0.271 + ], + "angle": 0, + "content": "[3] Changyou Chen, Ruiyi Zhang, Wenlin Wang, Bai Li, and Liquun Chen. A unified particle-optimization framework for scalable bayesian sampling. arXiv preprint arXiv:1805.11659, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[4] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.4 + ], + "angle": 0, + "content": "[5] Hanze Dong, Xi Wang, Yong Lin, and Tong Zhang. Particle-based variational inference with preconditioned functional gradient flow. arXiv preprint arXiv:2211.13954, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[6] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you should treat it like one. ICML, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[8] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[10] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[11] Susung Hong. Smoothed energy guidance: Guiding diffusion models with reduced energy curvature of attention. arXiv preprint arXiv:2408.00760, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[12] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debiasing scores and prompts of 2d diffusion for view-consistent text-to-3d generation. Advances in Neural Information Processing Systems, 36:11970-11987, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[13] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Susung Hong, Junyoung Seo, Heeseong Shin, Sunghwan Hong, and Seungryong Kim. Direct2v: Large language models are frame-level directors for zero-shot text-to-video generation. arXiv preprint arXiv:2305.14330, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[15] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. arXiv preprint arXiv:2309.14494, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[17] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.906, + 0.313 + ], + "angle": 0, + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[19] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[20] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[23] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[24] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[25] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13352-13361, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[27] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[28] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16301" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[30] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5773-5783, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[31] Chang Liua and Jun Zhub. Geometry in sampling methods: A review on manifold mcmc and particle-based variational inference methods. Advancements in Bayesian Methods and Implementations, 47:239, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7210-7219, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.481 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.483, + 0.538 + ], + "angle": 0, + "content": "[35] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[37] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[38] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[39] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[40] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[41] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[46] Junyoung Seo, Susung Hong, Wooseok Jang, Ines Hyeonsu Kim, Minseop Kwak, Doyup Lee, and Seungryong Kim. Retrieval-augmented score distillation for text-to-3d generation. arXiv preprint arXiv:2402.02972, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[47] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[48] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[50] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[52] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[53] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), pages 443-453. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[54] Cyrus Vachha and Ayaan Haque. Instruct-gs2gs: Editing 3d gaussian splats with instructions, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[55] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16302" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[56] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[57] Can Wang, Ruixiang Jiang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[58] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[59] Ziyu Wang, Tongzheng Ren, Jun Zhu, and Bo Zhang. Function space particle optimization for bayesian neural networks. arXiv preprint arXiv:1902.09754, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[60] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[61] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7378-7387, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[63] Hong-Xing Yu, Leonidas J Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. arXiv preprint arXiv:2107.07905, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[66] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4606-4615, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16303" + } + ] +] \ No newline at end of file diff --git a/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/11693e39-cb59-4d3c-8f05-aea184580288_origin.pdf b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/11693e39-cb59-4d3c-8f05-aea184580288_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..67bb75895d1fa7899f399b8251953687b586a898 --- /dev/null +++ b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/11693e39-cb59-4d3c-8f05-aea184580288_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a4b105b92bd2116225f65ab004625bb68e4b8b250034e64863ab376b1ba7fc8 +size 7487605 diff --git a/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/full.md b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a68d0fa41c067bdaa55bdcdcf6fc449ac26a1983 --- /dev/null +++ b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/full.md @@ -0,0 +1,326 @@ +# Perturb-and-Revise: Flexible 3D Editing with Generative Trajectories + +Susung Hong $^{1}$ Johanna Karras $^{1}$ Ricardo Martin-Brualla $^{2}$ Ira Kemelmacher-Shlizerman $^{1,2}$ $^{1}$ University of Washington $^{2}$ Google Research + +![](images/0456744913aec7f778273311d527dfbba574028498ee5fcd1bfbbcdb035a415a.jpg) +Figure 1. Perturb-and-Revise takes a source NeRF and an edit prompt as input and produces the edited result through: (1) versatile initialization via parameter perturbation, (2) multi-view consistent score distillation, and (3) refinement with the identity-preserving gradient. + +# Abstract + +Recent advancements in text-based diffusion models have accelerated progress in 3D reconstruction and text-based 3D editing. Although existing 3D editing methods excel at modifying color, texture, and style, they struggle with extensive geometric or appearance changes, thus limiting their applications. To this end, we propose Perturb-and-Revise, which makes possible a variety of NeRF editing. First, we perturb the NeRF parameters with random initializations to create a versatile initialization. The level of perturbation is determined automatically through analysis of the local loss landscape. Then, we revise the edited NeRF via generative trajectories. Combined with the generative process, we impose identity-preserving gradients to refine the edited NeRF. Extensive experiments demonstrate that Perturb-and-Revise facilitates flexible, effective, and consistent editing of color, appearance, and geometry in 3D. For $360^{\circ}$ results, please visit our project page: https://susunghong.github.io/Perturb-and-Revise. + +# 1. Introduction + +Neural Radiance Fields (NeRFs) [34] have revolutionized the creation of high-quality 3D scenes, marking a significant advancement in 3D reconstruction technology. Beyond + +their initial applications, NeRFs have enabled flexible generation of 3D content from models trained solely on image-text pairs [18, 40] through score distillation [25, 40, 58]. + +However, editing 3D content—an essential aspect of refinement and customization—remains a time-consuming and labor-intensive process across various industries, including animation, manufacturing, design, and gaming. This challenge is particularly pronounced with NeRFs, where color and density attributes are intricately encoded within their parameters. Consequently, there is an ongoing need for more intuitive and universally accessible tools for editing NeRFs, leveraging user-friendly interfaces such as text prompts for broader applicability and faster production. + +Fortunately, recent innovations in text-conditioned diffusion models have enabled significant advances in 3D editing using natural language prompts. For instance, leveraging diffusion models [2, 43], recent works such as Instruct-NeRF2NeRF [8] and Posterior Distillation [25] propose methods that use text prompts to modify 3D scenes. While these methods excel in altering the color or style of an object, they struggle with edits involving significant geometric or appearance changes, limiting their applications. + +To overcome these limitations, we propose a 3D object editing framework called Perturb-and-Revise (PnR), which leverages existing 3D generation methods for editing in a flexible and natural way (Figs. 1 and 2). Drawing on the + +![](images/5a4e7f32f0d97019875e342d6fc409fff0c7540d4a344e57328fa3d4d59a5236.jpg) +Source 3D Scene + +![](images/c035d216508a2241fbcd98d489f27f96865c7f83e31224982997696332a481f9.jpg) +Instruct-NeRF2NeRF + +![](images/ec4bffab462e5c59c8047e4cde5ccd3bb38998891a405abc95c8cf682238dee6.jpg) +Instruct-GS2GS + +![](images/67c69180d03e99875a45a5c6770ef55c48ff02528dd13b681672c91c4127b87d.jpg) +Posterior Distillation* + +![](images/f8194064b3ac22c85c18725074a71d063ad53e82085a8302c85902a95e7e4d11.jpg) +Perturb-and-Revise (Ours)* + +notion that NeRF parameters optimized with score distillation can be considered as a particle [60] or data point [6], we propose a novel method that leverages perturbation at the parameter level to perform text-based edits requiring various changes. Specifically, we construct versatile NeRF initialization by interpolating the source NeRF and a random NeRF initialization. Intuitively, perturbation in the parameter space helps the particle escape local minima and facilitates its following of the natural trajectory of the generative ODE towards the distribution of the desired edit (Fig. 3), allowing for a wide range of challenging edits including changing the pose and introducing new objects. + +Additionally, to determine the amount of perturbation needed to escape the basin of attraction without costly searching algorithms, we propose an algorithm based on the loss landscape near the source parameters, by simulating a few score distillation steps ahead of the parameter perturbation and optimization. After the early step of our framework, to achieve more similarity to the original object and refine the quality of the result, we employ the Identity-Preserving Gradient (IPG) to increase the fidelity of the edited NeRF to the source NeRF. As a result, the output closely resembles the source while still maintaining fidelity to the intended edits. + +We extensively evaluate our method on 3D fashion objects, as well as general objects in Objaverse [4], on a variety of appearance- and geometry-based edits. Furthermore, we demonstrate that our method achieves state-of-the-art results on various 3D editing baselines. + +# 2. Related Work + +Diffusion models. Diffusion models [9, 49], which are closely associated with score-based models [21, 50, 51], have recently shown remarkable sample quality in image synthesis. Latent Diffusion Models (LDM) [43], which perform the diffusion forward and backward process in the latent space [7], have demonstrated their efficacy and genera + +![](images/bc107ecfc77875f8ae04e8bc2db533e300d8125079552eb0f28714c99971a0f6.jpg) +Figure 2. We qualitatively compare our method to Instruct-NeRF2NeRF [8], Instruct-GS2GS [54], and Posterior Distillation [25]. In this task, the goal is to edit the source "bear" into a "polar bear." Compared to related works, our approach reconstructs a more realistic face and better overall geometry. An asterisk (*) denotes that we use an identical update rule and schedule. +Figure 3. Conceptual figure. The target distribution in the figures represents the conditional distribution of NeRF parameters relative to the edit prompt, and $\mathcal{P}(\Theta)$ denotes the distribution of randomly initialized NeRF parameters. First, parameter perturbation enables the parameters to escape from local minima and follow a natural generative path. Subsequently, during the refining process, the tug-of-war between two vectors, $\lambda_{\mathrm{d}}\nabla_{\theta}d(\theta_{\tau},\theta_{\mathrm{src}})$ (the red arrow) and $d\theta_{\tau}$ (the blue arrow), pushes the actual parameters into a region that is closer to either the source parameters or the high-density region specified by the edit prompt, following the purple arrow. + +tion quality. This trend has been scaled up by training diffusion models on large-scale text-image datasets [45] for text-conditional generation. The prosperity of text-to-image diffusion models has led to their application in complex content creation, such as 3D scenes and videos [1, 10, 12, 14, 15, 23, 27, 33, 40, 46, 58, 60, 62]. Recently, many adapting methods of text-to-image diffusion models have been proposed to achieve either generation with additional conditions [2, 11, 13, 20, 35, 64, 66] or to maintain consistency needed for videos [14, 23, 41, 62] or 3D scenes [29, 47]. + +Score distillation. Score distillation [40, 58, 60], a recent trend in synthesizing underlying parameterizations, mostly in 3D scene generation, uses only score-based models [43, 44] operating on 2D data. This approach has revolutionized text-to-3D generation by introducing an elegant method that does not require intense training on a + +3D dataset. It enables the rich and flexible knowledge of 2D text-to-image models to be transferred to 3D generation [12, 27, 33, 58, 60]. However, due to bias in 2D models, a multi-view consistency problem has arisen, necessitating further analysis and generalization of the original method [12, 29, 47]. Another limitation of score distillation is its time-consuming optimization, with the controllability and editing capability remaining underexplored. + +Text-based NeRF editing. In recent years, neural radiance fields (NeRF) [34] have emerged as a groundbreaking approach for generating photorealistic novel views of a scene captured in photographs, and have been extended in many follow-up works [22, 32, 36, 39]. In response to these advancements, several techniques for modifying NeRF have been proposed. Traditional methodologies for editing NeRF include the alteration of materials and lighting [37, 52, 55], as well as the spatial manipulation of objects via bounding box frameworks [38, 63]. Additionally, there has been exploration in the stylization of NeRFs, including Edit-NeRF [30], Clip-NeRF [56], and NeRF-Art [57], as well as in distilling 2D features into radiance fields, such as in Distilled Feature Fields [24] and Neural Feature Fusion Fields [53], which enables nuanced, guided edits based on language or images. Crucially, Instruct-NeRF2NeRF [8] focuses on user-friendly, language-based editing commands, using an instruction-driven, 2D image-conditioned diffusion model [2] for more intuitive and context-aware 3D editing. However, even in a trend using 2D language-based models such as CLIP [42] or text-to-image diffusion models [2, 43, 44], these methodologies are unable to control large geometrical changes, such as changing the pose or introducing a massive object. Posterior Distillation [25], instead of naively matching the noise, proposes a way to use a loss that matches the stochastic latents [16, 61] of source and target images. Still, this method falls into issues such as slower convergence and large computational costs. + +# 3. Background + +Diffusion models are a type of generative model that iteratively restores an image from Gaussian noise. The training objective of diffusion models involves training a neural network to estimate the score for a noised image $y$ given the noise level $\sigma$ via denoising score matching [17]. Following the preconditioning convention [21], a diffusion model, parameterized by $\phi$ , can be interpreted as a denoiser $D_{\phi}(y;\sigma)$ which minimizes the weighted L2 loss across different $\sigma$ values in data samples. + +In the context of text-based 3D scene generation, Score Distillation Sampling (SDS) [40] updates the 3D scene rep + +resentation given a text caption $c$ using the following rule: + +$$ +\begin{array}{l} \nabla_ {\theta} \mathcal {L} _ {\mathrm {S D S}} (\phi , c, x = g (\theta , \psi)) = \\ - \mathbb {E} _ {\sigma \sim \Sigma , n \sim \mathcal {N} (0, \sigma^ {2} I)} \left[ \omega (\sigma) \left(D _ {\phi} (x + n; \sigma , c) - x\right) \frac {\partial x}{\partial \theta} \right]. \tag {1} \\ \end{array} +$$ + +In this equation, $g(\cdot)$ is the differentiable renderer, $\psi$ is the random camera pose, $\Sigma$ is a predefined distribution from which the noise level for the denoiser is sampled, and $\omega(\cdot)$ denotes the weighting function. + +SDS can be viewed as a form of particle-based variational inference [3, 5, 28, 31, 59, 60]. In this context, the update rule can be expressed as a generative ODE over an optimization step $\tau$ , derived via Wasserstein gradient flow [3]: + +$$ +\frac {d \theta_ {\tau}}{d \tau} = - \mathbb {E} _ {\sigma \sim \Sigma , n \sim \mathcal {N} (0, \sigma^ {2} I)} \left[ \omega (\sigma) \left(D _ {\phi} (x + n; \sigma , c) - x\right) \frac {\partial x}{\partial \theta_ {\tau}} \right]. \tag {2} +$$ + +For further details, please see the supplementary material. + +As background, multi-view diffusion models [47], denoted as $\{D_{\phi}^{\mathrm{M}}(g(\theta ,\psi_i) + n;\sigma ,c,\psi_i)\}_{i = 1}^{N}$ , consistently generate multi-view images simultaneously. Here, $N$ denotes the total number of multi-views generated simultaneously, and $\psi_{i}$ denotes the $i$ -th viewpoint. Note that for $N = 1$ , these are ordinary diffusion models. For $N > 1$ , they are implemented by adapting the attention module to attend to different frames [47] and by training on 3D assets [4]. + +# 4. Method + +In text-based editing scenarios, we assume that a user provides a text prompt, $c_{\mathrm{edit}}$ , and a source 3D object represented by the parameterization $\theta_{\mathrm{src}}$ . The aim is to modify the 3D object into a new form $\theta^{*} = \mathcal{E}(\theta_{\mathrm{src}}, c_{\mathrm{edit}})$ , ensuring that the transformation aligns with the text prompt and retains a resemblance to the original 3D object. + +At a high level, we leverage the generative ODE of 3D objects through novel parameter perturbation while preserving proximity to the source object. In Sec. 4.1, we introduce perturbation in parameter space that enables the parameters to escape local minima and converge to natural generative trajectories. In Sec. 4.2, we propose a novel algorithm to determine the degree of perturbation by analyzing the local loss landscape. Finally, in Sec. 4.3, we propose balancing fidelity between the text prompt and the source object to place the final output in the desired region. + +# 4.1. Parameter Perturbation for Flexible Object Editing + +Given the generative ODE in Eq. 2, which minimizes the energy functional between distributions, a fully-optimized NeRF of a 3D object is considered to exhibit relatively low energy. This state corresponds to the particle residing in a local minimum, which remains stable despite alterations in + +the text prompts, as it is still capable of producing realistic renderings. As exemplified in the first row of Fig. 4, naively using $\theta_{\mathrm{src}}$ as initialization and making independent predictions on 2D noisy images with diffusion models are insufficient to handle the consistent changes in overall appearance and geometry that occur in a 3D object. + +Interpreting neural fields as a particle [60] or a data point [6], we propose parameter-level perturbation for an editing task, which facilitates changes in $\theta_{\mathrm{src}}$ by enabling the editing particle ODE to hijack the natural, coarse-to-fine generative process in Eq. 2 that the parameters are likely to follow. In other words, since adding more noise to the parameters is equivalent to undoing more of the optimization process, this approach enables the parameters to be flexible, allowing for significant changes. + +Specifically, let $\mathcal{P}(\Theta_0)$ denote the probability density function over $\Theta_0$ , which is a set of random parameterizations. Further, assume that we have a fully optimized 3D scene, $\theta_{\mathrm{src}}$ . Then, we obtain the perturbed version of $\theta_{\mathrm{src}}$ by using linear interpolation: + +$$ +\theta_ {\text {p e r t u r b e d}} = \operatorname {L e r p} \left(\theta_ {\text {s r c}}, \theta_ {0}, \eta\right) \quad \text {w h e r e} \quad \theta_ {0} \sim \mathcal {P} \left(\Theta_ {0}\right). \tag {3} +$$ + +where $\eta \in [0,1]$ denotes the perturbation amount and $\mathrm{Lerp}(\theta_{\mathrm{src}},\theta_0,\eta) = (1 - \eta)\cdot \theta_{\mathrm{src}} + \eta \cdot \theta_0$ . Note that $\eta = 1$ indicates $\theta_{\mathrm{perturbed}}$ is the random parameters. Theoretically, the interpolation makes the distribution of perturbed particles similar to the initializing distribution of NeRF, making it versatile. See the supplementary material for details. + +Subsequently, we run the parameter ODE for textual editing, starting from $\theta_{\mathrm{perturbed}}$ , using a new prompt $c_{\mathrm{edit}}$ . However, we observe that adopting ordinary single-view diffusion models (when $N = 1$ ) poses ambiguity when an edit introduces asymmetry or adds a new object. Therefore, we execute multi-view consistent updates as follows: + +$$ +\frac {d \theta_ {\tau}}{d \tau} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \mathbb {E} _ {\sigma , n} \left[ \left(D _ {\phi} ^ {\mathrm {M}} \left(z _ {i} + n; \sigma , c _ {\mathrm {e d i t}}, \psi_ {i}\right) - z _ {i}\right) \frac {\partial z _ {i}}{\partial \theta_ {\tau}} \right], \tag {4} +$$ + +where we define $z_{i} \coloneqq g(\theta_{\tau}, \psi_{i})$ and omit the weighting factor for brevity. + +The conceptual illustration of parameter perturbation can be found in Fig. 3. This approach essentially reverts the optimization process by interpolating between the initial state and the fully optimized state. Consequently, the particles are relocated to a less-optimized state characterized by a more moderate energy landscape. This repositioning allows the particle ODE (Eq. 4) to intervene in the generative process with a new text prompt $c_{\mathrm{edit}}$ , guiding the parameter trajectory. As shown in Fig. 4, increasing the noise level in the parameters inversely correlates with the degree of reversion of the optimization, allowing for more radical modifications in the model outputs. + +![](images/0f2e24a33f87c3b194df814caf850690c7bddaefba69dfbfb67f0e22290b21b6.jpg) +100 steps 200 steps 300 steps 500 steps 700 steps 900 steps 10000 steps +Figure 4. Effect of parameter perturbation. In this example, we aim to make a NeRF model of a standing person sit down using the word "sitting." The scene converges quickly even with large perturbations $(\eta = 0.6)$ , while complete regeneration yields blurry rendering results given the same number of optimization steps. + +# 4.2. Determining $\eta$ by Analyzing Loss Landscape + +In Sec. 4.1, we see that $\eta$ controls parameter versatility by determining how similar the distribution of parameters is to the distribution of random NeRF initializations. The required degree of versatility depends on both the type of edits and the current source NeRF parameters. However, finding an optimal $\eta$ through standard mechanisms, such as grid or random search, is computationally expensive. Therefore, we propose a novel algorithm that explores the basin of attraction surrounding local minima to determine $\eta$ by analyzing the loss landscape. + +Specifically, we leverage the loss function as a proxy to measure the depth and volume of the basin of attraction. Interestingly, for some examples, we observe that the loss function even increases during optimization steps, supporting our claim in Sec. 4.1 that the parameters are in a low-energy state. To address this, we simulate several optimization steps with $c_{\mathrm{edit}}$ and calculate the total loss decrease prior to parameter perturbation. For algorithmic stability, we compute the difference between the averages of the first few steps and the last few steps. We then determine $\eta(\theta_{\mathrm{src}}, c_{\mathrm{edit}})$ based on the source parameters $\theta_{\mathrm{src}}$ and the text prompt $c_{\mathrm{edit}}$ using an inverted exponential decay function. Subsequently, we apply parameter perturbation in Eq. 3 with $\eta = \eta(\theta_{\mathrm{src}}, c_{\mathrm{edit}})$ . The complete algorithms for our parameter perturbation and $\eta$ selection are provided in the supplementary material. + +# 4.3. Identity-Preserving Gradient (IPG) + +While our parameter perturbation in Sec. 4.1 successfully enables flexible changes, there are still some estimation errors or biases arising from the diffusion model during this phase. On the other hand, imposing constraints that make + +![](images/769013849dd445801c0939501c8f97c039dbbf494e20a2dba4de450a7afad531.jpg) +Source 3D Object +Score Distillation +Posterior Distillation +Instruct-NeRF2NeRF +Perturb-and-Revise (Ours) +Figure 5. Baseline comparisons with a wide range of edits. We compare our method with Score Distillation Sampling (SDS) [40], Posterior Distillation Sampling (PDS) [25], and Instruct-NeRF2NeRF [8]. For SDS and PDS, we use MVDream [47] as the backbone for fair comparison. SDS alters the appearance and texture of the source objects and is unable to handle edits that require extensive geometric changes (3rd, 4th, and 5th rows). PDS is not capable of making significant edits and cannot deviate far from local minima due to its high preservation term from the start [25]. While Instruct-NeRF2NeRF changes the texture of objects as desired, it cannot address geometric changes. In contrast, our method is capable of various types of edits, including those involving large geometric changes. + +it similar to the source object (e.g., L2 distance) conflicts with the generative ODE, making the optimization challenging. To counteract the estimation error and to circumvent the conflict, we introduce the Identity-Preserving Gradient (IPG) term that is added at later steps to the editing gradient presented in Sec. 4.1. + +Initially, we assume that the ideal parameters are closer to the source parameters and the high-density region of the target distribution. Then, as conceptually shown in Fig. 3, we instigate a tug-of-war between two vectors: one symbolizes the velocity consistently pointing towards the high-density region as dictated by the edit prompt, and the other symbolizes the velocity towards $\theta_{\mathrm{src}}$ . Specifically, the velocity in Eq. 4 propels the particle towards the region with a high likelihood given the text $c_{\mathrm{edit}}$ , while the other gradient represents the pull towards the original NeRF parameteri + +zation $\theta_{\mathrm{src}}$ . This effectively corrects the shift caused by the error in the editing process, ensuring it resembles the original while maintaining the intended edits. To this end, we extend the optimization process and compute IPGs during these additional steps. + +Formally, we combine the result of Eq. 4 with the IPG to define a refinement step as follows: + +$$ +d \theta_ {\tau} ^ {\text {r e f i n e}} = d \theta_ {\tau} + \lambda_ {\mathrm {d}} \nabla_ {\theta} d \left(\theta_ {\tau}, \theta_ {\text {s r c}}\right), \tag {5} +$$ + +where $d(\cdot, \cdot)$ is a similarity metric. In practice, we observe that the combination of $L_{1}$ and perceptual loss [19], which is also the preferred choice in NeRF [34] training, is more robust and less susceptible to noise than using $L_{2}$ loss. Thus, given a random camera pose $\psi$ , we use a combination of L1 distance and perceptual similarity between the rendered images $g(\theta_{\tau}, \psi)$ and $g(\theta_{\mathrm{src}}, \psi)$ , weighted by $\lambda_{\mathrm{L1}}$ + +![](images/70adcefa17fffdf7303e013f02071b2893134d7c9a9d34a4173f55fed37b23f6.jpg) +Source 3D Object +Score Distillation +Posterior Distillation +Instruct-NeRF2NeRF +Perturb-and-Revise (Ours) +Figure 6. Baseline comparisons of editing various general 3D objects from the Objaverse dataset [4]. + +
MetricScore Distillation [40]Posterior Distillation [25]Instruct-NeRF2NeRF [8]Perturb-and-Revise (Ours)
CLIP-Dir-SimViT-B/32↑0.04800.02870.05830.0594
CLIP-Dir-SimViT-B/16↑0.04180.03040.05490.0534
CLIP-Dir-SimViT-L/14↑0.04150.02640.05390.0567
CLIP-Dir-Simaveraged↑0.04380.02850.05570.0565
LPIPSvgg↓0.12730.03370.10650.1060
LPIPSalex↓0.15330.02150.11120.1034
+ +Table 1. Comparison of different methods for fashion object editing. The best, second-best, and worst values are highlighted in bold, underlined, and gray, respectively. Values represent averages across all edit types and prompts. Our approach achieves a better trade-off between faithfulness to edit prompts and preservation of source 3D objects. While PDS exhibits lower LPIPS, it mostly generates edited objects that are nearly identical to the source, as evidenced by CLIP directional similarity. + +and $\lambda_{\mathfrak{p}}$ , respectively. + +# 4.4. Timestep Annealing + +In text-based 3D generation, some recent works have revealed that using timestep annealing, which adjusts the noise level added to the 2D rendered images according to the global optimization step, boosts the quality of the results [47, 60]. In the editing task, while we use a smaller noise level at the start, we also identify the annealing method as crucial for maintaining the original quality of the 3D object while reducing blurry textures. To this end, we regard $\Sigma$ as a function that depends on the editing step $\tau$ , $\Sigma(\tau)$ , from which the noise level $\sigma$ is sampled. Specifically, for $\tau \leq T$ , where $T$ is the final step of the schedule, $\Sigma(\tau)$ is gradually annealed from $\mathcal{U}(\sigma_{\min}^{0}, \sigma_{\max}^{0})$ to $\mathcal{U}(\sigma_{\min}^{\mathrm{T}}, \sigma_{\max}^{\mathrm{T}})$ . This approach initially facilitates low-frequency edits, followed by fine-grained ones. + +# 5. Experiments + +As baselines, we adopt two distillation-based methods: Score Distillation Sampling (SDS) [40] and Posterior Distillation Sampling (PDS) [25]. Bolded red in the prompts indicates the intended edits. We use the exact weighting factors for the losses as originally proposed in [25, 40]. For fair comparison, we use MVDream [47] as their backbones. Also, we use Instruct-NeRF2NeRF [8] as our baseline for the iterative dataset update strategy. We use several backbones for each metric and note them in the subscripts, while averaged refers to the average value across all backbones. + +# 5.1.3D Object Editing + +Fashion object editing. In our fashion object editing experiment, we use MVDream [47] to generate a synthetic dataset of 3D fashion objects and perform edits using our framework. Specifically, we categorize the fashion object edits into color, pattern, shape, pose, and object edits, and + +![](images/806296a49545bd21b6b7a6cea9fb55d7683afa6cf2d87510094b04186d496868.jpg) +Source 3D Object +Instruct-NeRF2NeRF +MVDream +Perturb-and-Revise (Ours) + +![](images/8fbf0e6913b09ad0acfd7e882a93b931e6caf6dea7c7c0de99bcac2d1fc1707f.jpg) +Figure 7. Comparisons with Instruct-NeRF2NeRF (dataset update) [8] and MVDream (regeneration) [47] in editing Objaverse objects. +(a) + +![](images/4e73f364f04c212157b4515bc368a3cdf1b1caa09313840583714def5ffa9f1f.jpg) +(b) +Figure 8. Ablation study on the selection of $\eta$ . (a) and (b) show the CLIP direction similarity and the percentage of successful experiments (without errors) for different $\eta$ values, respectively. When averaged across all types of edits, our adaptive method achieves near-maximum performance on these metrics compared to all fixed $\eta$ values. + +synthesize a total of 150 editing examples, with 30 for each type of edit. In this experiment, our framework takes as input a source object in NeRF with the desired edits to its characteristics, such as color, pattern, shape, pose, and added objects. Note that, unlike some previous work [25], we do not require a description of the original object. + +The qualitative results, as shown in Figs. 4 and 5, demonstrate the versatility introduced by parameter perturbation and our method's capability to perform various types of edits, respectively. In Fig. 5, we apply five types of edits: color, pattern, shape, pose, and object additions. We show that our framework supports a wide range of edits, including those requiring pose changes and introducing objects, corroborating its capability for significant geometric modifications, which have rarely been addressed in previous work [8, 25]. + +General object editing. We show a general object editing capability for our method, with objects from Objaverse [4]. We first optimize InstantNGP [36] with 200 rendered images and camera transforms extracted from an object. Then, we apply our editing method. + +We display the qualitative results of our method in Fig. 6. In addition to textural changes, our method successfully enables structural changes. In Fig. 7, we also demonstrate our + +
MethodCLIP-Dir-Sim↑CLIP-Dir-Con↑LPIPS↓
w/o Refinement0.06240.75720.1147
w/ Refinement0.05650.76420.1047
+ +Table 2. Ablation study on the IPG steps. We present average values across all edit types, prompts, and evaluation backbones. Our approach yields a substantial reduction in LPIPS while balancing faithfulness to the source object and edit prompt. + +method's ability to perform creative edits anchored by the source 3D object. Here, to demonstrate flexibility, we use two prompts with different animal motifs. We can see that regenerating the object with MVDream produces results that appear completely different from the source 3D object. In contrast, our method successfully anchors the original shape while changing it to align with the edit prompt. + +# 5.2. Comparison with Baselines + +We compare our method against baselines to demonstrate our clear advantage in complex edits, as shown in Figs. 5 and 6. For the comparison, we display editing results using SDS with MVDream, Instruct-NeRF2NeRF, and PDS. Additionally, we compare our strategy with dataset update and regeneration strategies in Fig. 7. + +Different from other methods, SDS exhibits blurry tex + +![](images/99b2f56980c3c891998109f9a90e0b7530343eb6f28c2e584710c27736cdde53.jpg) + +![](images/8d3567d28f850c6fe854e7a44446b2948b1db22471d281dadb81e9a779f6bcd8.jpg) +"... a cactus plant with flowers" + +![](images/65d0e7d37aacfd062ade932c86e92458aac049d20e4a15374913a4e451d1eb33.jpg) + +![](images/4bf341fc0340139126be6c2b353c56bc2df81ec3430ca94baae7b15366ed46b8.jpg) +"… with windmills" + +![](images/6f8d30c525a4522cd2c765500c7abf13a5dfee4aac830711660de2694c09918c.jpg) +NeRF2NeRF +Figure 9. Real scene editing results. An asterisk (*) indicates we use the exact same update rule and schedule. + +![](images/925484d9a959e348142ed309dd5a217e4d99312fd116bfa52866e8912b11e3e1.jpg) +... with a snowman +Posterior +Perturb-and +lation\* +Revise (Ours)* + +tures and lacks the capability to perform complex edits. Although it can handle simple changes, it inadvertently changes the appearance and texture of the source objects and cannot perform edits that need extensive geometric changes, as seen in the 3rd, 4th, and 5th rows of Fig. 5. While Instruct-NeRF2NeRF successfully achieves simple color edits or symmetric edits, we observe that even when the editing process converges, it is not able to address geometric changes in the objects, such as a change in the pose. PDS, while maintaining similarity to the source object, is unable to make proper edits according to the prompt and cannot deviate far from local minima. As shown in Figs. 5, 6, and 7, our method is capable of various types of edits, including those involving large geometric changes, achieving state-of-the-art results for editing color, appearance, and geometry in a flexible and consistent manner. + +We have to consider both the similarity to the source object and faithfulness to the edit prompt. To this end, we evaluate these aspects using CLIP directional similarity [8] across ViT-B/32, ViT-B/16, and ViT-L/14 [42] for measuring faithfulness to the edit prompts and LPIPS [65] across VGG [48] and AlexNet [26] for evaluating the similarity between original and edited objects. The results are shown in Table 1. We can see that our method excels at both metrics and achieves a better trade-off than other baselines. This, along with the quantitative evaluation, corroborates the effectiveness of our approach. + +# 5.3. Real Scene Editing + +Our parameter perturbation approach described in Sec. 4.1 can be readily extended to real scene editing scenarios [8, 25]. We present real scene editing results in Figs. 2 and 9, building upon Posterior Distillation [25]. Using the same number of iterations, we can better modify the scene geometry through parameter perturbation. We refer the readers to the supplementary material for details. + +# 5.4. Ablation Study + +Parameter perturbation. In Fig. 4, we display the rendered images based on the amount of parameter perturbation and optimization steps. This demonstrates that parameter perturbation facilitates easy alteration of the source object's structure while keeping crucial parts unchanged. Additionally, we present a quantitative study of $\eta$ selection in Fig. 8, where we categorize edits into five categories: color, pattern, shape, pose, and object edits. We observe that pose changes and object additions generally require a larger degree of versatility and thus need a larger value of $\eta$ to achieve maximum performance. Furthermore, we demonstrate that, when averaged across all types of edits, the adaptive $\eta$ selection approach outperforms strategies using fixed $\eta \in \{0.0, 0.2, 0.4, 0.6, 0.8\}$ while maintaining high experimental success rates across various edit categories and significantly reducing computational costs compared to grid search. This indicates that our method dynamically selects an effective $\eta$ value while remaining robust in terms of optimization. We present additional metrics in the supplementary material. + +Identity-preserving gradients. We demonstrate the effect of identity-preserving gradients. In Table 2, we present diverse metrics. Our results show that the refinement steps with IPG achieve a significant decrease in LPIPS while decreasing CLIP directional similarity, due to their trade-off relationship. In addition, we observe an overall increase in CLIP directional consistency [8]. + +# 6. Conclusion + +In this paper, we present Perturb-and-Revise (PnR), a framework for text-guided 3D object editing. By introducing adaptive parameter perturbation and identity-preserving gradients, our method enables extensive geometric and appearance changes that adhere to the text prompt while also maintaining fidelity to source objects. Our experiments corroborate that PnR achieves state-of-the-art results across diverse editing tasks without requiring model retraining or multiple input images. We discuss various limitations of our method in the supplementary material for future research. + +Acknowledgments. This work was supported by the UW Reality Lab and Google. + +# References + +[1] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. CVPR, 2023. 2 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. CVPR, 2023. 1, 2, 3 +[3] Changyou Chen, Ruiyi Zhang, Wenlin Wang, Bai Li, and Liquun Chen. A unified particle-optimization framework for scalable bayesian sampling. arXiv preprint arXiv:1805.11659, 2018. 3 +[4] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3, 6, 7 +[5] Hanze Dong, Xi Wang, Yong Lin, and Tong Zhang. Particle-based variational inference with preconditioned functional gradient flow. arXiv preprint arXiv:2211.13954, 2022. 3 +[6] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you should treat it like one. ICML, 2022. 2, 4 +[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2 +[8] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 1, 2, 3, 5, 6, 7, 8 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2 +[10] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[11] Susung Hong. Smoothed energy guidance: Guiding diffusion models with reduced energy curvature of attention. arXiv preprint arXiv:2408.00760, 2024. 2 +[12] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debiasing scores and prompts of 2d diffusion for view-consistent text-to-3d generation. Advances in Neural Information Processing Systems, 36:11970-11987, 2023. 2, 3 +[13] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 2 +[14] Susung Hong, Junyoung Seo, Heeseong Shin, Sunghwan Hong, and Seungryong Kim. Direct2v: Large language models are frame-level directors for zero-shot text-to-video generation. arXiv preprint arXiv:2305.14330, 2023. 2 + +[15] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. arXiv preprint arXiv:2309.14494, 2023. 2 +[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3 +[17] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3 +[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 1 +[19] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5 +[20] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023. 2 +[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 2, 3 +[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 3 +[23] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2 +[24] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 3 +[25] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13352-13361, 2024. 1, 2, 3, 5, 6, 7, 8 +[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 8 +[27] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2, 3 +[28] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016. 3 + +[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 2, 3 +[30] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5773-5783, 2021. 3 +[31] Chang Liua and Jun Zhub. Geometry in sampling methods: A review on manifold mcmc and particle-based variational inference methods. Advancements in Bayesian Methods and Implementations, 47:239, 2022. 3 +[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7210-7219, 2021. 3 +[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 2, 3 +[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 5 +[35] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2 +[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 7 +[37] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 3 +[38] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 3 +[39] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3 +[40] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 5, 6 +[41] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2 + +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3, 8 +[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 3 +[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3 +[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[46] Junyoung Seo, Susung Hong, Wooseok Jang, Ines Hyeonsu Kim, Minseop Kwak, Doyup Lee, and Seungryong Kim. Retrieval-augmented score distillation for text-to-3d generation. arXiv preprint arXiv:2402.02972, 2024. 2 +[47] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 5, 6, 7 +[48] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 8 +[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2 +[50] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 32, 2019. 2 +[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 2 +[52] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 3 +[53] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), pages 443-453. IEEE, 2022. 3 +[54] Cyrus Vachha and Ayaan Haque. Instruct-gs2gs: Editing 3d gaussian splats with instructions, 2024. 2 +[55] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. + +In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 3 +[56] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 3 +[57] Can Wang, Ruixiang Jiang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 3 +[58] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 1, 2, 3 +[59] Ziyu Wang, Tongzheng Ren, Jun Zhu, and Bo Zhang. Function space particle optimization for bayesian neural networks. arXiv preprint arXiv:1902.09754, 2019. 3 +[60] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 6 +[61] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7378-7387, 2023. 3 +[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2 +[63] Hong-Xing Yu, Leonidas J Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. arXiv preprint arXiv:2107.07905, 2021. 3 +[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8 +[66] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4606-4615, 2023. 2 \ No newline at end of file diff --git a/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/images.zip b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..77d42b1cfac6c04f3f6979276494e5ad25667f94 --- /dev/null +++ b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a07444772dcdff03e1abe4a9f16af58b44c470a205382c3d105a9c758dcc3dda +size 776146 diff --git a/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/layout.json b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..647204d3876f5af51b43b88e89cd17d5be420d92 --- /dev/null +++ b/2025/Perturb-and-Revise_ Flexible 3D Editing with Generative Trajectories/layout.json @@ -0,0 +1,9442 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "type": "text", + "content": "Perturb-and-Revise: Flexible 3D Editing with Generative Trajectories" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "text", + "content": "Susung Hong" + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "text", + "content": " Johanna Karras" + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "text", + "content": " Ricardo Martin-Brualla" + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "text", + "content": " Ira Kemelmacher-Shlizerman" + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "text", + "content": "University of Washington " + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 142, + 532, + 186 + ], + "type": "text", + "content": "Google Research" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 90, + 198, + 527, + 357 + ], + "blocks": [ + { + "bbox": [ + 90, + 198, + 527, + 357 + ], + "lines": [ + { + "bbox": [ + 90, + 198, + 527, + 357 + ], + "spans": [ + { + "bbox": [ + 90, + 198, + 527, + 357 + ], + "type": "image", + "image_path": "0456744913aec7f778273311d527dfbba574028498ee5fcd1bfbbcdb035a415a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 360, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 360, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 555, + 383 + ], + "type": "text", + "content": "Figure 1. Perturb-and-Revise takes a source NeRF and an edit prompt as input and produces the edited result through: (1) versatile initialization via parameter perturbation, (2) multi-view consistent score distillation, and (3) refinement with the identity-preserving gradient." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 392, + 200, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 392, + 200, + 404 + ], + "spans": [ + { + "bbox": [ + 151, + 392, + 200, + 404 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 418, + 297, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 418, + 297, + 634 + ], + "spans": [ + { + "bbox": [ + 54, + 418, + 297, + 634 + ], + "type": "text", + "content": "Recent advancements in text-based diffusion models have accelerated progress in 3D reconstruction and text-based 3D editing. Although existing 3D editing methods excel at modifying color, texture, and style, they struggle with extensive geometric or appearance changes, thus limiting their applications. To this end, we propose Perturb-and-Revise, which makes possible a variety of NeRF editing. First, we perturb the NeRF parameters with random initializations to create a versatile initialization. The level of perturbation is determined automatically through analysis of the local loss landscape. Then, we revise the edited NeRF via generative trajectories. Combined with the generative process, we impose identity-preserving gradients to refine the edited NeRF. Extensive experiments demonstrate that Perturb-and-Revise facilitates flexible, effective, and consistent editing of color, appearance, and geometry in 3D. For " + }, + { + "bbox": [ + 54, + 418, + 297, + 634 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 54, + 418, + 297, + 634 + ], + "type": "text", + "content": " results, please visit our project page: https://susunghong.github.io/Perturb-and-Revise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 657, + 135, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 135, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 135, + 669 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRFs) [34] have revolutionized the creation of high-quality 3D scenes, marking a significant advancement in 3D reconstruction technology. Beyond" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "type": "text", + "content": "their initial applications, NeRFs have enabled flexible generation of 3D content from models trained solely on image-text pairs [18, 40] through score distillation [25, 40, 58]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 432, + 556, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 432, + 556, + 553 + ], + "spans": [ + { + "bbox": [ + 312, + 432, + 556, + 553 + ], + "type": "text", + "content": "However, editing 3D content—an essential aspect of refinement and customization—remains a time-consuming and labor-intensive process across various industries, including animation, manufacturing, design, and gaming. This challenge is particularly pronounced with NeRFs, where color and density attributes are intricately encoded within their parameters. Consequently, there is an ongoing need for more intuitive and universally accessible tools for editing NeRFs, leveraging user-friendly interfaces such as text prompts for broader applicability and faster production." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 554, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 554, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 312, + 554, + 556, + 663 + ], + "type": "text", + "content": "Fortunately, recent innovations in text-conditioned diffusion models have enabled significant advances in 3D editing using natural language prompts. For instance, leveraging diffusion models [2, 43], recent works such as Instruct-NeRF2NeRF [8] and Posterior Distillation [25] propose methods that use text prompts to modify 3D scenes. While these methods excel in altering the color or style of an object, they struggle with edits involving significant geometric or appearance changes, limiting their applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 665, + 556, + 715 + ], + "type": "text", + "content": "To overcome these limitations, we propose a 3D object editing framework called Perturb-and-Revise (PnR), which leverages existing 3D generation methods for editing in a flexible and natural way (Figs. 1 and 2). Drawing on the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16293" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 160, + 217 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 160, + 217 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 160, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 160, + 217 + ], + "type": "image", + "image_path": "5a4e7f32f0d97019875e342d6fc409fff0c7540d4a344e57328fa3d4d59a5236.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 217, + 134, + 225 + ], + "lines": [ + { + "bbox": [ + 78, + 217, + 134, + 225 + ], + "spans": [ + { + "bbox": [ + 78, + 217, + 134, + 225 + ], + "type": "text", + "content": "Source 3D Scene" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 164, + 70, + 262, + 216 + ], + "blocks": [ + { + "bbox": [ + 164, + 70, + 262, + 216 + ], + "lines": [ + { + "bbox": [ + 164, + 70, + 262, + 216 + ], + "spans": [ + { + "bbox": [ + 164, + 70, + 262, + 216 + ], + "type": "image", + "image_path": "c035d216508a2241fbcd98d489f27f96865c7f83e31224982997696332a481f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 217, + 248, + 225 + ], + "lines": [ + { + "bbox": [ + 179, + 217, + 248, + 225 + ], + "spans": [ + { + "bbox": [ + 179, + 217, + 248, + 225 + ], + "type": "text", + "content": "Instruct-NeRF2NeRF" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 262, + 70, + 358, + 216 + ], + "blocks": [ + { + "bbox": [ + 262, + 70, + 358, + 216 + ], + "lines": [ + { + "bbox": [ + 262, + 70, + 358, + 216 + ], + "spans": [ + { + "bbox": [ + 262, + 70, + 358, + 216 + ], + "type": "image", + "image_path": "ec4bffab462e5c59c8047e4cde5ccd3bb38998891a405abc95c8cf682238dee6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 217, + 338, + 225 + ], + "lines": [ + { + "bbox": [ + 287, + 217, + 338, + 225 + ], + "spans": [ + { + "bbox": [ + 287, + 217, + 338, + 225 + ], + "type": "text", + "content": "Instruct-GS2GS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 359, + 70, + 455, + 216 + ], + "blocks": [ + { + "bbox": [ + 359, + 70, + 455, + 216 + ], + "lines": [ + { + "bbox": [ + 359, + 70, + 455, + 216 + ], + "spans": [ + { + "bbox": [ + 359, + 70, + 455, + 216 + ], + "type": "image", + "image_path": "67c69180d03e99875a45a5c6770ef55c48ff02528dd13b681672c91c4127b87d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 217, + 442, + 225 + ], + "lines": [ + { + "bbox": [ + 372, + 217, + 442, + 225 + ], + "spans": [ + { + "bbox": [ + 372, + 217, + 442, + 225 + ], + "type": "text", + "content": "Posterior Distillation*" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 456, + 71, + 553, + 216 + ], + "blocks": [ + { + "bbox": [ + 456, + 71, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 456, + 71, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 456, + 71, + 553, + 216 + ], + "type": "image", + "image_path": "f8194064b3ac22c85c18725074a71d063ad53e82085a8302c85902a95e7e4d11.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 217, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 460, + 217, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 460, + 217, + 547, + 225 + ], + "type": "text", + "content": "Perturb-and-Revise (Ours)*" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 266, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 295, + 409 + ], + "type": "text", + "content": "notion that NeRF parameters optimized with score distillation can be considered as a particle [60] or data point [6], we propose a novel method that leverages perturbation at the parameter level to perform text-based edits requiring various changes. Specifically, we construct versatile NeRF initialization by interpolating the source NeRF and a random NeRF initialization. Intuitively, perturbation in the parameter space helps the particle escape local minima and facilitates its following of the natural trajectory of the generative ODE towards the distribution of the desired edit (Fig. 3), allowing for a wide range of challenging edits including changing the pose and introducing new objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 410, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 410, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 410, + 295, + 552 + ], + "type": "text", + "content": "Additionally, to determine the amount of perturbation needed to escape the basin of attraction without costly searching algorithms, we propose an algorithm based on the loss landscape near the source parameters, by simulating a few score distillation steps ahead of the parameter perturbation and optimization. After the early step of our framework, to achieve more similarity to the original object and refine the quality of the result, we employ the Identity-Preserving Gradient (IPG) to increase the fidelity of the edited NeRF to the source NeRF. As a result, the output closely resembles the source while still maintaining fidelity to the intended edits." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 552, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 552, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 54, + 552, + 295, + 613 + ], + "type": "text", + "content": "We extensively evaluate our method on 3D fashion objects, as well as general objects in Objaverse [4], on a variety of appearance- and geometry-based edits. Furthermore, we demonstrate that our method achieves state-of-the-art results on various 3D editing baselines." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Diffusion models. Diffusion models [9, 49], which are closely associated with score-based models [21, 50, 51], have recently shown remarkable sample quality in image synthesis. Latent Diffusion Models (LDM) [43], which perform the diffusion forward and backward process in the latent space [7], have demonstrated their efficacy and genera" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 326, + 266, + 537, + 395 + ], + "blocks": [ + { + "bbox": [ + 55, + 227, + 555, + 259 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 555, + 259 + ], + "type": "text", + "content": "Figure 2. We qualitatively compare our method to Instruct-NeRF2NeRF [8], Instruct-GS2GS [54], and Posterior Distillation [25]. In this task, the goal is to edit the source \"bear\" into a \"polar bear.\" Compared to related works, our approach reconstructs a more realistic face and better overall geometry. An asterisk (*) denotes that we use an identical update rule and schedule." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 266, + 537, + 395 + ], + "lines": [ + { + "bbox": [ + 326, + 266, + 537, + 395 + ], + "spans": [ + { + "bbox": [ + 326, + 266, + 537, + 395 + ], + "type": "image", + "image_path": "bc107ecfc77875f8ae04e8bc2db533e300d8125079552eb0f28714c99971a0f6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 399, + 555, + 510 + ], + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 510 + ], + "type": "text", + "content": "Figure 3. Conceptual figure. The target distribution in the figures represents the conditional distribution of NeRF parameters relative to the edit prompt, and " + }, + { + "bbox": [ + 313, + 399, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\Theta)" + }, + { + "bbox": [ + 313, + 399, + 555, + 510 + ], + "type": "text", + "content": " denotes the distribution of randomly initialized NeRF parameters. First, parameter perturbation enables the parameters to escape from local minima and follow a natural generative path. Subsequently, during the refining process, the tug-of-war between two vectors, " + }, + { + "bbox": [ + 313, + 399, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{d}}\\nabla_{\\theta}d(\\theta_{\\tau},\\theta_{\\mathrm{src}})" + }, + { + "bbox": [ + 313, + 399, + 555, + 510 + ], + "type": "text", + "content": " (the red arrow) and " + }, + { + "bbox": [ + 313, + 399, + 555, + 510 + ], + "type": "inline_equation", + "content": "d\\theta_{\\tau}" + }, + { + "bbox": [ + 313, + 399, + 555, + 510 + ], + "type": "text", + "content": " (the blue arrow), pushes the actual parameters into a region that is closer to either the source parameters or the high-density region specified by the edit prompt, following the purple arrow." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 515, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 555, + 635 + ], + "type": "text", + "content": "tion quality. This trend has been scaled up by training diffusion models on large-scale text-image datasets [45] for text-conditional generation. The prosperity of text-to-image diffusion models has led to their application in complex content creation, such as 3D scenes and videos [1, 10, 12, 14, 15, 23, 27, 33, 40, 46, 58, 60, 62]. Recently, many adapting methods of text-to-image diffusion models have been proposed to achieve either generation with additional conditions [2, 11, 13, 20, 35, 64, 66] or to maintain consistency needed for videos [14, 23, 41, 62] or 3D scenes [29, 47]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Score distillation. Score distillation [40, 58, 60], a recent trend in synthesizing underlying parameterizations, mostly in 3D scene generation, uses only score-based models [43, 44] operating on 2D data. This approach has revolutionized text-to-3D generation by introducing an elegant method that does not require intense training on a" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16294" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "3D dataset. It enables the rich and flexible knowledge of 2D text-to-image models to be transferred to 3D generation [12, 27, 33, 58, 60]. However, due to bias in 2D models, a multi-view consistency problem has arisen, necessitating further analysis and generalization of the original method [12, 29, 47]. Another limitation of score distillation is its time-consuming optimization, with the controllability and editing capability remaining underexplored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 195, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 531 + ], + "type": "text", + "content": "Text-based NeRF editing. In recent years, neural radiance fields (NeRF) [34] have emerged as a groundbreaking approach for generating photorealistic novel views of a scene captured in photographs, and have been extended in many follow-up works [22, 32, 36, 39]. In response to these advancements, several techniques for modifying NeRF have been proposed. Traditional methodologies for editing NeRF include the alteration of materials and lighting [37, 52, 55], as well as the spatial manipulation of objects via bounding box frameworks [38, 63]. Additionally, there has been exploration in the stylization of NeRFs, including Edit-NeRF [30], Clip-NeRF [56], and NeRF-Art [57], as well as in distilling 2D features into radiance fields, such as in Distilled Feature Fields [24] and Neural Feature Fusion Fields [53], which enables nuanced, guided edits based on language or images. Crucially, Instruct-NeRF2NeRF [8] focuses on user-friendly, language-based editing commands, using an instruction-driven, 2D image-conditioned diffusion model [2] for more intuitive and context-aware 3D editing. However, even in a trend using 2D language-based models such as CLIP [42] or text-to-image diffusion models [2, 43, 44], these methodologies are unable to control large geometrical changes, such as changing the pose or introducing a massive object. Posterior Distillation [25], instead of naively matching the noise, proposes a way to use a loss that matches the stochastic latents [16, 61] of source and target images. Still, this method falls into issues such as slower convergence and large computational costs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 553, + 134, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 134, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 134, + 567 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "text", + "content": "Diffusion models are a type of generative model that iteratively restores an image from Gaussian noise. The training objective of diffusion models involves training a neural network to estimate the score for a noised image " + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "text", + "content": " given the noise level " + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "text", + "content": " via denoising score matching [17]. Following the preconditioning convention [21], a diffusion model, parameterized by " + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "text", + "content": ", can be interpreted as a denoiser " + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "D_{\\phi}(y;\\sigma)" + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "text", + "content": " which minimizes the weighted L2 loss across different " + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 578, + 295, + 685 + ], + "type": "text", + "content": " values in data samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "In the context of text-based 3D scene generation, Score Distillation Sampling (SDS) [40] updates the 3D scene rep" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 549, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 549, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 549, + 85 + ], + "type": "text", + "content": "resentation given a text caption " + }, + { + "bbox": [ + 313, + 72, + 549, + 85 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 72, + 549, + 85 + ], + "type": "text", + "content": " using the following rule:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 321, + 91, + 554, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 91, + 554, + 142 + ], + "spans": [ + { + "bbox": [ + 321, + 91, + 554, + 142 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , c, x = g (\\theta , \\psi)) = \\\\ - \\mathbb {E} _ {\\sigma \\sim \\Sigma , n \\sim \\mathcal {N} (0, \\sigma^ {2} I)} \\left[ \\omega (\\sigma) \\left(D _ {\\phi} (x + n; \\sigma , c) - x\\right) \\frac {\\partial x}{\\partial \\theta} \\right]. \\tag {1} \\\\ \\end{array}", + "image_path": "776276bd5efd0c2ef820e15590e1255c7441cfbd8c4a1253b9b37014b13ffc0a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 150, + 554, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 554, + 198 + ], + "type": "text", + "content": "In this equation, " + }, + { + "bbox": [ + 313, + 150, + 554, + 198 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 150, + 554, + 198 + ], + "type": "text", + "content": " is the differentiable renderer, " + }, + { + "bbox": [ + 313, + 150, + 554, + 198 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 150, + 554, + 198 + ], + "type": "text", + "content": " is the random camera pose, " + }, + { + "bbox": [ + 313, + 150, + 554, + 198 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 150, + 554, + 198 + ], + "type": "text", + "content": " is a predefined distribution from which the noise level for the denoiser is sampled, and " + }, + { + "bbox": [ + 313, + 150, + 554, + 198 + ], + "type": "inline_equation", + "content": "\\omega(\\cdot)" + }, + { + "bbox": [ + 313, + 150, + 554, + 198 + ], + "type": "text", + "content": " denotes the weighting function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 198, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 247 + ], + "type": "text", + "content": "SDS can be viewed as a form of particle-based variational inference [3, 5, 28, 31, 59, 60]. In this context, the update rule can be expressed as a generative ODE over an optimization step " + }, + { + "bbox": [ + 313, + 198, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 198, + 555, + 247 + ], + "type": "text", + "content": ", derived via Wasserstein gradient flow [3]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 253, + 558, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 253, + 558, + 287 + ], + "spans": [ + { + "bbox": [ + 315, + 253, + 558, + 287 + ], + "type": "interline_equation", + "content": "\\frac {d \\theta_ {\\tau}}{d \\tau} = - \\mathbb {E} _ {\\sigma \\sim \\Sigma , n \\sim \\mathcal {N} (0, \\sigma^ {2} I)} \\left[ \\omega (\\sigma) \\left(D _ {\\phi} (x + n; \\sigma , c) - x\\right) \\frac {\\partial x}{\\partial \\theta_ {\\tau}} \\right]. \\tag {2}", + "image_path": "1cfb5f42418d02a632f89530cf02ce928882308b126b2d555d8869b58fc9f142.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 289, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 547, + 300 + ], + "type": "text", + "content": "For further details, please see the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "text", + "content": "As background, multi-view diffusion models [47], denoted as " + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\{D_{\\phi}^{\\mathrm{M}}(g(\\theta ,\\psi_i) + n;\\sigma ,c,\\psi_i)\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "text", + "content": ", consistently generate multi-view images simultaneously. Here, " + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "text", + "content": " denotes the total number of multi-views generated simultaneously, and " + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\psi_{i}" + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "text", + "content": "-th viewpoint. Note that for " + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "text", + "content": ", these are ordinary diffusion models. For " + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "inline_equation", + "content": "N > 1" + }, + { + "bbox": [ + 313, + 300, + 554, + 396 + ], + "type": "text", + "content": ", they are implemented by adapting the attention module to attend to different frames [47] and by training on 3D assets [4]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 405, + 370, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 370, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 370, + 418 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "type": "text", + "content": "In text-based editing scenarios, we assume that a user provides a text prompt, " + }, + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{edit}}" + }, + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "type": "text", + "content": ", and a source 3D object represented by the parameterization " + }, + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "type": "text", + "content": ". The aim is to modify the 3D object into a new form " + }, + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\theta^{*} = \\mathcal{E}(\\theta_{\\mathrm{src}}, c_{\\mathrm{edit}})" + }, + { + "bbox": [ + 313, + 426, + 554, + 498 + ], + "type": "text", + "content": ", ensuring that the transformation aligns with the text prompt and retains a resemblance to the original 3D object." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 618 + ], + "type": "text", + "content": "At a high level, we leverage the generative ODE of 3D objects through novel parameter perturbation while preserving proximity to the source object. In Sec. 4.1, we introduce perturbation in parameter space that enables the parameters to escape local minima and converge to natural generative trajectories. In Sec. 4.2, we propose a novel algorithm to determine the degree of perturbation by analyzing the local loss landscape. Finally, in Sec. 4.3, we propose balancing fidelity between the text prompt and the source object to place the final output in the desired region." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 624, + 554, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 554, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 554, + 650 + ], + "type": "text", + "content": "4.1. Parameter Perturbation for Flexible Object Editing" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Given the generative ODE in Eq. 2, which minimizes the energy functional between distributions, a fully-optimized NeRF of a 3D object is considered to exhibit relatively low energy. This state corresponds to the particle residing in a local minimum, which remains stable despite alterations in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16295" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "the text prompts, as it is still capable of producing realistic renderings. As exemplified in the first row of Fig. 4, naively using " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " as initialization and making independent predictions on 2D noisy images with diffusion models are insufficient to handle the consistent changes in overall appearance and geometry that occur in a 3D object." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "content": "Interpreting neural fields as a particle [60] or a data point [6], we propose parameter-level perturbation for an editing task, which facilitates changes in " + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 144, + 295, + 251 + ], + "type": "text", + "content": " by enabling the editing particle ODE to hijack the natural, coarse-to-fine generative process in Eq. 2 that the parameters are likely to follow. In other words, since adding more noise to the parameters is equivalent to undoing more of the optimization process, this approach enables the parameters to be flexible, allowing for significant changes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "text", + "content": "Specifically, let " + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\Theta_0)" + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "text", + "content": " denote the probability density function over " + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\Theta_0" + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "text", + "content": ", which is a set of random parameterizations. Further, assume that we have a fully optimized 3D scene, " + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "text", + "content": ". Then, we obtain the perturbed version of " + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 251, + 295, + 312 + ], + "type": "text", + "content": " by using linear interpolation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 319, + 295, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 319, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 64, + 319, + 295, + 334 + ], + "type": "interline_equation", + "content": "\\theta_ {\\text {p e r t u r b e d}} = \\operatorname {L e r p} \\left(\\theta_ {\\text {s r c}}, \\theta_ {0}, \\eta\\right) \\quad \\text {w h e r e} \\quad \\theta_ {0} \\sim \\mathcal {P} \\left(\\Theta_ {0}\\right). \\tag {3}", + "image_path": "5ebae9eb78bc3e6b286cf7cd85ac59cad01d47a7040bfa0292b3c229ac48e7b4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 341, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 295, + 412 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 341, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\eta \\in [0,1]" + }, + { + "bbox": [ + 55, + 341, + 295, + 412 + ], + "type": "text", + "content": " denotes the perturbation amount and " + }, + { + "bbox": [ + 55, + 341, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\mathrm{Lerp}(\\theta_{\\mathrm{src}},\\theta_0,\\eta) = (1 - \\eta)\\cdot \\theta_{\\mathrm{src}} + \\eta \\cdot \\theta_0" + }, + { + "bbox": [ + 55, + 341, + 295, + 412 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 55, + 341, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\eta = 1" + }, + { + "bbox": [ + 55, + 341, + 295, + 412 + ], + "type": "text", + "content": " indicates " + }, + { + "bbox": [ + 55, + 341, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{perturbed}}" + }, + { + "bbox": [ + 55, + 341, + 295, + 412 + ], + "type": "text", + "content": " is the random parameters. Theoretically, the interpolation makes the distribution of perturbed particles similar to the initializing distribution of NeRF, making it versatile. See the supplementary material for details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "type": "text", + "content": "Subsequently, we run the parameter ODE for textual editing, starting from " + }, + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{perturbed}}" + }, + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "type": "text", + "content": ", using a new prompt " + }, + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{edit}}" + }, + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "type": "text", + "content": ". However, we observe that adopting ordinary single-view diffusion models (when " + }, + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 55, + 413, + 295, + 485 + ], + "type": "text", + "content": ") poses ambiguity when an edit introduces asymmetry or adds a new object. Therefore, we execute multi-view consistent updates as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 492, + 296, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 537 + ], + "type": "interline_equation", + "content": "\\frac {d \\theta_ {\\tau}}{d \\tau} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {E} _ {\\sigma , n} \\left[ \\left(D _ {\\phi} ^ {\\mathrm {M}} \\left(z _ {i} + n; \\sigma , c _ {\\mathrm {e d i t}}, \\psi_ {i}\\right) - z _ {i}\\right) \\frac {\\partial z _ {i}}{\\partial \\theta_ {\\tau}} \\right], \\tag {4}", + "image_path": "668314c8c707b8b6fdb490c94e0deb2c91de410f78f51d118fc08373b205c80a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 569 + ], + "type": "text", + "content": "where we define " + }, + { + "bbox": [ + 55, + 545, + 295, + 569 + ], + "type": "inline_equation", + "content": "z_{i} \\coloneqq g(\\theta_{\\tau}, \\psi_{i})" + }, + { + "bbox": [ + 55, + 545, + 295, + 569 + ], + "type": "text", + "content": " and omit the weighting factor for brevity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": "The conceptual illustration of parameter perturbation can be found in Fig. 3. This approach essentially reverts the optimization process by interpolating between the initial state and the fully optimized state. Consequently, the particles are relocated to a less-optimized state characterized by a more moderate energy landscape. This repositioning allows the particle ODE (Eq. 4) to intervene in the generative process with a new text prompt " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{edit}}" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": ", guiding the parameter trajectory. As shown in Fig. 4, increasing the noise level in the parameters inversely correlates with the degree of reversion of the optimization, allowing for more radical modifications in the model outputs." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 552, + 240 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 552, + 240 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 552, + 240 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 552, + 240 + ], + "type": "image", + "image_path": "0f2e24a33f87c3b194df814caf850690c7bddaefba69dfbfb67f0e22290b21b6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 240, + 548, + 249 + ], + "lines": [ + { + "bbox": [ + 358, + 240, + 548, + 249 + ], + "spans": [ + { + "bbox": [ + 358, + 240, + 548, + 249 + ], + "type": "text", + "content": "100 steps 200 steps 300 steps 500 steps 700 steps 900 steps 10000 steps" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 249, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 249, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 553, + 304 + ], + "type": "text", + "content": "Figure 4. Effect of parameter perturbation. In this example, we aim to make a NeRF model of a standing person sit down using the word \"sitting.\" The scene converges quickly even with large perturbations " + }, + { + "bbox": [ + 313, + 249, + 553, + 304 + ], + "type": "inline_equation", + "content": "(\\eta = 0.6)" + }, + { + "bbox": [ + 313, + 249, + 553, + 304 + ], + "type": "text", + "content": ", while complete regeneration yields blurry rendering results given the same number of optimization steps." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 310, + 545, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 545, + 323 + ], + "type": "text", + "content": "4.2. Determining " + }, + { + "bbox": [ + 313, + 310, + 545, + 323 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 310, + 545, + 323 + ], + "type": "text", + "content": " by Analyzing Loss Landscape" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 328, + 554, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 554, + 448 + ], + "type": "text", + "content": "In Sec. 4.1, we see that " + }, + { + "bbox": [ + 313, + 328, + 554, + 448 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 328, + 554, + 448 + ], + "type": "text", + "content": " controls parameter versatility by determining how similar the distribution of parameters is to the distribution of random NeRF initializations. The required degree of versatility depends on both the type of edits and the current source NeRF parameters. However, finding an optimal " + }, + { + "bbox": [ + 313, + 328, + 554, + 448 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 328, + 554, + 448 + ], + "type": "text", + "content": " through standard mechanisms, such as grid or random search, is computationally expensive. Therefore, we propose a novel algorithm that explores the basin of attraction surrounding local minima to determine " + }, + { + "bbox": [ + 313, + 328, + 554, + 448 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 328, + 554, + 448 + ], + "type": "text", + "content": " by analyzing the loss landscape." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "text", + "content": "Specifically, we leverage the loss function as a proxy to measure the depth and volume of the basin of attraction. Interestingly, for some examples, we observe that the loss function even increases during optimization steps, supporting our claim in Sec. 4.1 that the parameters are in a low-energy state. To address this, we simulate several optimization steps with " + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{edit}}" + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "text", + "content": " and calculate the total loss decrease prior to parameter perturbation. For algorithmic stability, we compute the difference between the averages of the first few steps and the last few steps. We then determine " + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\eta(\\theta_{\\mathrm{src}}, c_{\\mathrm{edit}})" + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "text", + "content": " based on the source parameters " + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "text", + "content": " and the text prompt " + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{edit}}" + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "text", + "content": " using an inverted exponential decay function. Subsequently, we apply parameter perturbation in Eq. 3 with " + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\eta = \\eta(\\theta_{\\mathrm{src}}, c_{\\mathrm{edit}})" + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "text", + "content": ". The complete algorithms for our parameter perturbation and " + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 449, + 555, + 640 + ], + "type": "text", + "content": " selection are provided in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 503, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 503, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 503, + 661 + ], + "type": "text", + "content": "4.3. Identity-Preserving Gradient (IPG)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "While our parameter perturbation in Sec. 4.1 successfully enables flexible changes, there are still some estimation errors or biases arising from the diffusion model during this phase. On the other hand, imposing constraints that make" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16296" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 82, + 536, + 434 + ], + "blocks": [ + { + "bbox": [ + 78, + 82, + 536, + 434 + ], + "lines": [ + { + "bbox": [ + 78, + 82, + 536, + 434 + ], + "spans": [ + { + "bbox": [ + 78, + 82, + 536, + 434 + ], + "type": "image", + "image_path": "769013849dd445801c0939501c8f97c039dbbf494e20a2dba4de450a7afad531.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 437, + 154, + 446 + ], + "lines": [ + { + "bbox": [ + 95, + 437, + 154, + 446 + ], + "spans": [ + { + "bbox": [ + 95, + 437, + 154, + 446 + ], + "type": "text", + "content": "Source 3D Object" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 208, + 437, + 265, + 445 + ], + "lines": [ + { + "bbox": [ + 208, + 437, + 265, + 445 + ], + "spans": [ + { + "bbox": [ + 208, + 437, + 265, + 445 + ], + "type": "text", + "content": "Score Distillation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 288, + 437, + 356, + 445 + ], + "lines": [ + { + "bbox": [ + 288, + 437, + 356, + 445 + ], + "spans": [ + { + "bbox": [ + 288, + 437, + 356, + 445 + ], + "type": "text", + "content": "Posterior Distillation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 437, + 441, + 445 + ], + "lines": [ + { + "bbox": [ + 371, + 437, + 441, + 445 + ], + "spans": [ + { + "bbox": [ + 371, + 437, + 441, + 445 + ], + "type": "text", + "content": "Instruct-NeRF2NeRF" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 450, + 437, + 535, + 445 + ], + "lines": [ + { + "bbox": [ + 450, + 437, + 535, + 445 + ], + "spans": [ + { + "bbox": [ + 450, + 437, + 535, + 445 + ], + "type": "text", + "content": "Perturb-and-Revise (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 445, + 555, + 511 + ], + "lines": [ + { + "bbox": [ + 55, + 445, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 555, + 511 + ], + "type": "text", + "content": "Figure 5. Baseline comparisons with a wide range of edits. We compare our method with Score Distillation Sampling (SDS) [40], Posterior Distillation Sampling (PDS) [25], and Instruct-NeRF2NeRF [8]. For SDS and PDS, we use MVDream [47] as the backbone for fair comparison. SDS alters the appearance and texture of the source objects and is unable to handle edits that require extensive geometric changes (3rd, 4th, and 5th rows). PDS is not capable of making significant edits and cannot deviate far from local minima due to its high preservation term from the start [25]. While Instruct-NeRF2NeRF changes the texture of objects as desired, it cannot address geometric changes. In contrast, our method is capable of various types of edits, including those involving large geometric changes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 517, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 590 + ], + "type": "text", + "content": "it similar to the source object (e.g., L2 distance) conflicts with the generative ODE, making the optimization challenging. To counteract the estimation error and to circumvent the conflict, we introduce the Identity-Preserving Gradient (IPG) term that is added at later steps to the editing gradient presented in Sec. 4.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Initially, we assume that the ideal parameters are closer to the source parameters and the high-density region of the target distribution. Then, as conceptually shown in Fig. 3, we instigate a tug-of-war between two vectors: one symbolizes the velocity consistently pointing towards the high-density region as dictated by the edit prompt, and the other symbolizes the velocity towards " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ". Specifically, the velocity in Eq. 4 propels the particle towards the region with a high likelihood given the text " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{edit}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", while the other gradient represents the pull towards the original NeRF parameteri" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 517, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 577 + ], + "type": "text", + "content": "zation " + }, + { + "bbox": [ + 313, + 517, + 555, + 577 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 517, + 555, + 577 + ], + "type": "text", + "content": ". This effectively corrects the shift caused by the error in the editing process, ensuring it resembles the original while maintaining the intended edits. To this end, we extend the optimization process and compute IPGs during these additional steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 578, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 601 + ], + "type": "text", + "content": "Formally, we combine the result of Eq. 4 with the IPG to define a refinement step as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 366, + 609, + 555, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 609, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 366, + 609, + 555, + 623 + ], + "type": "interline_equation", + "content": "d \\theta_ {\\tau} ^ {\\text {r e f i n e}} = d \\theta_ {\\tau} + \\lambda_ {\\mathrm {d}} \\nabla_ {\\theta} d \\left(\\theta_ {\\tau}, \\theta_ {\\text {s r c}}\\right), \\tag {5}", + "image_path": "dbc0358684e16db1ae87ba405deab37b62e9ec34cf48fd80479a08bcfc449162.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " is a similarity metric. In practice, we observe that the combination of " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " and perceptual loss [19], which is also the preferred choice in NeRF [34] training, is more robust and less susceptible to noise than using " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " loss. Thus, given a random camera pose " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ", we use a combination of L1 distance and perceptual similarity between the rendered images " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "g(\\theta_{\\tau}, \\psi)" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "g(\\theta_{\\mathrm{src}}, \\psi)" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ", weighted by " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{L1}}" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16297" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 79, + 527, + 326 + ], + "blocks": [ + { + "bbox": [ + 84, + 79, + 527, + 326 + ], + "lines": [ + { + "bbox": [ + 84, + 79, + 527, + 326 + ], + "spans": [ + { + "bbox": [ + 84, + 79, + 527, + 326 + ], + "type": "image", + "image_path": "70adcefa17fffdf7303e013f02071b2893134d7c9a9d34a4173f55fed37b23f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 328, + 156, + 338 + ], + "lines": [ + { + "bbox": [ + 97, + 328, + 156, + 338 + ], + "spans": [ + { + "bbox": [ + 97, + 328, + 156, + 338 + ], + "type": "text", + "content": "Source 3D Object" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 196, + 328, + 253, + 337 + ], + "lines": [ + { + "bbox": [ + 196, + 328, + 253, + 337 + ], + "spans": [ + { + "bbox": [ + 196, + 328, + 253, + 337 + ], + "type": "text", + "content": "Score Distillation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 275, + 328, + 342, + 337 + ], + "lines": [ + { + "bbox": [ + 275, + 328, + 342, + 337 + ], + "spans": [ + { + "bbox": [ + 275, + 328, + 342, + 337 + ], + "type": "text", + "content": "Posterior Distillation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 358, + 328, + 428, + 337 + ], + "lines": [ + { + "bbox": [ + 358, + 328, + 428, + 337 + ], + "spans": [ + { + "bbox": [ + 358, + 328, + 428, + 337 + ], + "type": "text", + "content": "Instruct-NeRF2NeRF" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 328, + 522, + 338 + ], + "lines": [ + { + "bbox": [ + 436, + 328, + 522, + 338 + ], + "spans": [ + { + "bbox": [ + 436, + 328, + 522, + 338 + ], + "type": "text", + "content": "Perturb-and-Revise (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 123, + 338, + 487, + 348 + ], + "lines": [ + { + "bbox": [ + 123, + 338, + 487, + 348 + ], + "spans": [ + { + "bbox": [ + 123, + 338, + 487, + 348 + ], + "type": "text", + "content": "Figure 6. Baseline comparisons of editing various general 3D objects from the Objaverse dataset [4]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 77, + 354, + 531, + 438 + ], + "blocks": [ + { + "bbox": [ + 77, + 354, + 531, + 438 + ], + "lines": [ + { + "bbox": [ + 77, + 354, + 531, + 438 + ], + "spans": [ + { + "bbox": [ + 77, + 354, + 531, + 438 + ], + "type": "table", + "html": "
MetricScore Distillation [40]Posterior Distillation [25]Instruct-NeRF2NeRF [8]Perturb-and-Revise (Ours)
CLIP-Dir-SimViT-B/32↑0.04800.02870.05830.0594
CLIP-Dir-SimViT-B/16↑0.04180.03040.05490.0534
CLIP-Dir-SimViT-L/14↑0.04150.02640.05390.0567
CLIP-Dir-Simaveraged↑0.04380.02850.05570.0565
LPIPSvgg↓0.12730.03370.10650.1060
LPIPSalex↓0.15330.02150.11120.1034
", + "image_path": "cbd9dd176918061395080b343927fe8a912a908397b49ad8c0c978ca390c3e94.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 440, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 55, + 440, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 555, + 485 + ], + "type": "text", + "content": "Table 1. Comparison of different methods for fashion object editing. The best, second-best, and worst values are highlighted in bold, underlined, and gray, respectively. Values represent averages across all edit types and prompts. Our approach achieves a better trade-off between faithfulness to edit prompts and preservation of source 3D objects. While PDS exhibits lower LPIPS, it mostly generates edited objects that are nearly identical to the source, as evidenced by CLIP directional similarity." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 491, + 139, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 139, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 139, + 503 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 55, + 491, + 139, + 503 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathfrak{p}}" + }, + { + "bbox": [ + 55, + 491, + 139, + 503 + ], + "type": "text", + "content": " , respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 523, + 172, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 172, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 172, + 537 + ], + "type": "text", + "content": "4.4. Timestep Annealing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "In text-based 3D generation, some recent works have revealed that using timestep annealing, which adjusts the noise level added to the 2D rendered images according to the global optimization step, boosts the quality of the results [47, 60]. In the editing task, while we use a smaller noise level at the start, we also identify the annealing method as crucial for maintaining the original quality of the 3D object while reducing blurry textures. To this end, we regard " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " as a function that depends on the editing step " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma(\\tau)" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ", from which the noise level " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " is sampled. Specifically, for " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tau \\leq T" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " is the final step of the schedule, " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma(\\tau)" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " is gradually annealed from " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\sigma_{\\min}^{0}, \\sigma_{\\max}^{0})" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\sigma_{\\min}^{\\mathrm{T}}, \\sigma_{\\max}^{\\mathrm{T}})" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ". This approach initially facilitates low-frequency edits, followed by fine-grained ones." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 490, + 394, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 490, + 394, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 490, + 394, + 502 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 509, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 629 + ], + "type": "text", + "content": "As baselines, we adopt two distillation-based methods: Score Distillation Sampling (SDS) [40] and Posterior Distillation Sampling (PDS) [25]. Bolded red in the prompts indicates the intended edits. We use the exact weighting factors for the losses as originally proposed in [25, 40]. For fair comparison, we use MVDream [47] as their backbones. Also, we use Instruct-NeRF2NeRF [8] as our baseline for the iterative dataset update strategy. We use several backbones for each metric and note them in the subscripts, while averaged refers to the average value across all backbones." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 635, + 422, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 422, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 422, + 649 + ], + "type": "text", + "content": "5.1.3D Object Editing" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "Fashion object editing. In our fashion object editing experiment, we use MVDream [47] to generate a synthetic dataset of 3D fashion objects and perform edits using our framework. Specifically, we categorize the fashion object edits into color, pattern, shape, pose, and object edits, and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16298" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 79, + 531, + 220 + ], + "blocks": [ + { + "bbox": [ + 82, + 79, + 531, + 220 + ], + "lines": [ + { + "bbox": [ + 82, + 79, + 531, + 220 + ], + "spans": [ + { + "bbox": [ + 82, + 79, + 531, + 220 + ], + "type": "image", + "image_path": "806296a49545bd21b6b7a6cea9fb55d7683afa6cf2d87510094b04186d496868.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 222, + 162, + 232 + ], + "lines": [ + { + "bbox": [ + 102, + 222, + 162, + 232 + ], + "spans": [ + { + "bbox": [ + 102, + 222, + 162, + 232 + ], + "type": "text", + "content": "Source 3D Object" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 214, + 222, + 288, + 230 + ], + "lines": [ + { + "bbox": [ + 214, + 222, + 288, + 230 + ], + "spans": [ + { + "bbox": [ + 214, + 222, + 288, + 230 + ], + "type": "text", + "content": "Instruct-NeRF2NeRF" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 222, + 380, + 230 + ], + "lines": [ + { + "bbox": [ + 343, + 222, + 380, + 230 + ], + "spans": [ + { + "bbox": [ + 343, + 222, + 380, + 230 + ], + "type": "text", + "content": "MVDream" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 222, + 520, + 231 + ], + "lines": [ + { + "bbox": [ + 431, + 222, + 520, + 231 + ], + "spans": [ + { + "bbox": [ + 431, + 222, + 520, + 231 + ], + "type": "text", + "content": "Perturb-and-Revise (Ours)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 74, + 249, + 264, + 365 + ], + "blocks": [ + { + "bbox": [ + 61, + 232, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 61, + 232, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 547, + 243 + ], + "type": "text", + "content": "Figure 7. Comparisons with Instruct-NeRF2NeRF (dataset update) [8] and MVDream (regeneration) [47] in editing Objaverse objects." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 249, + 264, + 365 + ], + "lines": [ + { + "bbox": [ + 74, + 249, + 264, + 365 + ], + "spans": [ + { + "bbox": [ + 74, + 249, + 264, + 365 + ], + "type": "image", + "image_path": "8fbf0e6913b09ad0acfd7e882a93b931e6caf6dea7c7c0de99bcac2d1fc1707f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 367, + 174, + 376 + ], + "lines": [ + { + "bbox": [ + 164, + 367, + 174, + 376 + ], + "spans": [ + { + "bbox": [ + 164, + 367, + 174, + 376 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 265, + 249, + 533, + 367 + ], + "blocks": [ + { + "bbox": [ + 265, + 249, + 533, + 367 + ], + "lines": [ + { + "bbox": [ + 265, + 249, + 533, + 367 + ], + "spans": [ + { + "bbox": [ + 265, + 249, + 533, + 367 + ], + "type": "image", + "image_path": "4e73f364f04c212157b4515bc368a3cdf1b1caa09313840583714def5ffa9f1f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 367, + 405, + 376 + ], + "lines": [ + { + "bbox": [ + 394, + 367, + 405, + 376 + ], + "spans": [ + { + "bbox": [ + 394, + 367, + 405, + 376 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 376, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 555, + 409 + ], + "type": "text", + "content": "Figure 8. Ablation study on the selection of " + }, + { + "bbox": [ + 55, + 376, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 376, + 555, + 409 + ], + "type": "text", + "content": ". (a) and (b) show the CLIP direction similarity and the percentage of successful experiments (without errors) for different " + }, + { + "bbox": [ + 55, + 376, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 376, + 555, + 409 + ], + "type": "text", + "content": " values, respectively. When averaged across all types of edits, our adaptive method achieves near-maximum performance on these metrics compared to all fixed " + }, + { + "bbox": [ + 55, + 376, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 376, + 555, + 409 + ], + "type": "text", + "content": " values." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 415, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 295, + 486 + ], + "type": "text", + "content": "synthesize a total of 150 editing examples, with 30 for each type of edit. In this experiment, our framework takes as input a source object in NeRF with the desired edits to its characteristics, such as color, pattern, shape, pose, and added objects. Note that, unlike some previous work [25], we do not require a description of the original object." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 487, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 295, + 606 + ], + "type": "text", + "content": "The qualitative results, as shown in Figs. 4 and 5, demonstrate the versatility introduced by parameter perturbation and our method's capability to perform various types of edits, respectively. In Fig. 5, we apply five types of edits: color, pattern, shape, pose, and object additions. We show that our framework supports a wide range of edits, including those requiring pose changes and introducing objects, corroborating its capability for significant geometric modifications, which have rarely been addressed in previous work [8, 25]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "content": "General object editing. We show a general object editing capability for our method, with objects from Objaverse [4]. We first optimize InstantNGP [36] with 200 rendered images and camera transforms extracted from an object. Then, we apply our editing method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "We display the qualitative results of our method in Fig. 6. In addition to textural changes, our method successfully enables structural changes. In Fig. 7, we also demonstrate our" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 337, + 412, + 531, + 462 + ], + "blocks": [ + { + "bbox": [ + 337, + 412, + 531, + 462 + ], + "lines": [ + { + "bbox": [ + 337, + 412, + 531, + 462 + ], + "spans": [ + { + "bbox": [ + 337, + 412, + 531, + 462 + ], + "type": "table", + "html": "
MethodCLIP-Dir-Sim↑CLIP-Dir-Con↑LPIPS↓
w/o Refinement0.06240.75720.1147
w/ Refinement0.05650.76420.1047
", + "image_path": "3c73846066cb04da8987a186bef8f238a679d180917cee4626aeb9fb127e0819.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 464, + 555, + 510 + ], + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 510 + ], + "type": "text", + "content": "Table 2. Ablation study on the IPG steps. We present average values across all edit types, prompts, and evaluation backbones. Our approach yields a substantial reduction in LPIPS while balancing faithfulness to the source object and edit prompt." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "type": "text", + "content": "method's ability to perform creative edits anchored by the source 3D object. Here, to demonstrate flexibility, we use two prompts with different animal motifs. We can see that regenerating the object with MVDream produces results that appear completely different from the source 3D object. In contrast, our method successfully anchors the original shape while changing it to align with the edit prompt." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 610, + 465, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 465, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 465, + 623 + ], + "type": "text", + "content": "5.2. Comparison with Baselines" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 628, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 700 + ], + "type": "text", + "content": "We compare our method against baselines to demonstrate our clear advantage in complex edits, as shown in Figs. 5 and 6. For the comparison, we display editing results using SDS with MVDream, Instruct-NeRF2NeRF, and PDS. Additionally, we compare our strategy with dataset update and regeneration strategies in Fig. 7." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "content": "Different from other methods, SDS exhibits blurry tex" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16299" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 118, + 156 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 118, + 156 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 118, + 156 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 118, + 156 + ], + "type": "image", + "image_path": "99b2f56980c3c891998109f9a90e0b7530343eb6f28c2e584710c27736cdde53.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 123, + 70, + 295, + 156 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 295, + 156 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 295, + 156 + ], + "type": "image", + "image_path": "8d3567d28f850c6fe854e7a44446b2948b1db22471d281dadb81e9a779f6bcd8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 156, + 258, + 163 + ], + "lines": [ + { + "bbox": [ + 161, + 156, + 258, + 163 + ], + "spans": [ + { + "bbox": [ + 161, + 156, + 258, + 163 + ], + "type": "text", + "content": "\"... a cactus plant with flowers\"" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 165, + 118, + 243 + ], + "blocks": [ + { + "bbox": [ + 58, + 165, + 118, + 243 + ], + "lines": [ + { + "bbox": [ + 58, + 165, + 118, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 165, + 118, + 243 + ], + "type": "image", + "image_path": "65d0e7d37aacfd062ade932c86e92458aac049d20e4a15374913a4e451d1eb33.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 123, + 165, + 295, + 243 + ], + "blocks": [ + { + "bbox": [ + 123, + 165, + 295, + 243 + ], + "lines": [ + { + "bbox": [ + 123, + 165, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 123, + 165, + 295, + 243 + ], + "type": "image", + "image_path": "4bf341fc0340139126be6c2b353c56bc2df81ec3430ca94baae7b15366ed46b8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 243, + 239, + 249 + ], + "lines": [ + { + "bbox": [ + 178, + 243, + 239, + 249 + ], + "spans": [ + { + "bbox": [ + 178, + 243, + 239, + 249 + ], + "type": "text", + "content": "\"… with windmills\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 251, + 117, + 329 + ], + "blocks": [ + { + "bbox": [ + 58, + 251, + 117, + 329 + ], + "lines": [ + { + "bbox": [ + 58, + 251, + 117, + 329 + ], + "spans": [ + { + "bbox": [ + 58, + 251, + 117, + 329 + ], + "type": "image", + "image_path": "6f8d30c525a4522cd2c765500c7abf13a5dfee4aac830711660de2694c09918c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 344, + 170, + 351 + ], + "lines": [ + { + "bbox": [ + 131, + 344, + 170, + 351 + ], + "spans": [ + { + "bbox": [ + 131, + 344, + 170, + 351 + ], + "type": "text", + "content": "NeRF2NeRF" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 353, + 295, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 353, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 295, + 374 + ], + "type": "text", + "content": "Figure 9. Real scene editing results. An asterisk (*) indicates we use the exact same update rule and schedule." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 123, + 251, + 295, + 329 + ], + "blocks": [ + { + "bbox": [ + 123, + 251, + 295, + 329 + ], + "lines": [ + { + "bbox": [ + 123, + 251, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 123, + 251, + 295, + 329 + ], + "type": "image", + "image_path": "925484d9a959e348142ed309dd5a217e4d99312fd116bfa52866e8912b11e3e1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 329, + 241, + 335 + ], + "lines": [ + { + "bbox": [ + 176, + 329, + 241, + 335 + ], + "spans": [ + { + "bbox": [ + 176, + 329, + 241, + 335 + ], + "type": "text", + "content": "... with a snowman" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 196, + 336, + 224, + 342 + ], + "lines": [ + { + "bbox": [ + 196, + 336, + 224, + 342 + ], + "spans": [ + { + "bbox": [ + 196, + 336, + 224, + 342 + ], + "type": "text", + "content": "Posterior" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 336, + 284, + 343 + ], + "lines": [ + { + "bbox": [ + 247, + 336, + 284, + 343 + ], + "spans": [ + { + "bbox": [ + 247, + 336, + 284, + 343 + ], + "type": "text", + "content": "Perturb-and" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 344, + 230, + 351 + ], + "lines": [ + { + "bbox": [ + 209, + 344, + 230, + 351 + ], + "spans": [ + { + "bbox": [ + 209, + 344, + 230, + 351 + ], + "type": "text", + "content": "lation\\*" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 243, + 344, + 288, + 352 + ], + "lines": [ + { + "bbox": [ + 243, + 344, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 243, + 344, + 288, + 352 + ], + "type": "text", + "content": "Revise (Ours)*" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 388, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 579 + ], + "type": "text", + "content": "tures and lacks the capability to perform complex edits. Although it can handle simple changes, it inadvertently changes the appearance and texture of the source objects and cannot perform edits that need extensive geometric changes, as seen in the 3rd, 4th, and 5th rows of Fig. 5. While Instruct-NeRF2NeRF successfully achieves simple color edits or symmetric edits, we observe that even when the editing process converges, it is not able to address geometric changes in the objects, such as a change in the pose. PDS, while maintaining similarity to the source object, is unable to make proper edits according to the prompt and cannot deviate far from local minima. As shown in Figs. 5, 6, and 7, our method is capable of various types of edits, including those involving large geometric changes, achieving state-of-the-art results for editing color, appearance, and geometry in a flexible and consistent manner." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "content": "We have to consider both the similarity to the source object and faithfulness to the edit prompt. To this end, we evaluate these aspects using CLIP directional similarity [8] across ViT-B/32, ViT-B/16, and ViT-L/14 [42] for measuring faithfulness to the edit prompts and LPIPS [65] across VGG [48] and AlexNet [26] for evaluating the similarity between original and edited objects. The results are shown in Table 1. We can see that our method excels at both metrics and achieves a better trade-off than other baselines. This, along with the quantitative evaluation, corroborates the effectiveness of our approach." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 425, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 425, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 425, + 85 + ], + "type": "text", + "content": "5.3. Real Scene Editing" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "type": "text", + "content": "Our parameter perturbation approach described in Sec. 4.1 can be readily extended to real scene editing scenarios [8, 25]. We present real scene editing results in Figs. 2 and 9, building upon Posterior Distillation [25]. Using the same number of iterations, we can better modify the scene geometry through parameter perturbation. We refer the readers to the supplementary material for details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 182, + 408, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 408, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 408, + 194 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "text", + "content": "Parameter perturbation. In Fig. 4, we display the rendered images based on the amount of parameter perturbation and optimization steps. This demonstrates that parameter perturbation facilitates easy alteration of the source object's structure while keeping crucial parts unchanged. Additionally, we present a quantitative study of " + }, + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "text", + "content": " selection in Fig. 8, where we categorize edits into five categories: color, pattern, shape, pose, and object edits. We observe that pose changes and object additions generally require a larger degree of versatility and thus need a larger value of " + }, + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "text", + "content": " to achieve maximum performance. Furthermore, we demonstrate that, when averaged across all types of edits, the adaptive " + }, + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "text", + "content": " selection approach outperforms strategies using fixed " + }, + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\eta \\in \\{0.0, 0.2, 0.4, 0.6, 0.8\\}" + }, + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "text", + "content": " while maintaining high experimental success rates across various edit categories and significantly reducing computational costs compared to grid search. This indicates that our method dynamically selects an effective " + }, + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 200, + 555, + 439 + ], + "type": "text", + "content": " value while remaining robust in terms of optimization. We present additional metrics in the supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 446, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 556, + 532 + ], + "type": "text", + "content": "Identity-preserving gradients. We demonstrate the effect of identity-preserving gradients. In Table 2, we present diverse metrics. Our results show that the refinement steps with IPG achieve a significant decrease in LPIPS while decreasing CLIP directional similarity, due to their trade-off relationship. In addition, we observe an overall increase in CLIP directional consistency [8]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 542, + 388, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 388, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 388, + 554 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 563, + 555, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 683 + ], + "type": "text", + "content": "In this paper, we present Perturb-and-Revise (PnR), a framework for text-guided 3D object editing. By introducing adaptive parameter perturbation and identity-preserving gradients, our method enables extensive geometric and appearance changes that adhere to the text prompt while also maintaining fidelity to source objects. Our experiments corroborate that PnR achieves state-of-the-art results across diverse editing tasks without requiring model retraining or multiple input images. We discuss various limitations of our method in the supplementary material for future research." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by the UW Reality Lab and Google." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16300" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. CVPR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 168 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. CVPR, 2023. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 296, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 296, + 214 + ], + "type": "text", + "content": "[3] Changyou Chen, Ruiyi Zhang, Wenlin Wang, Bai Li, and Liquun Chen. A unified particle-optimization framework for scalable bayesian sampling. arXiv preprint arXiv:1805.11659, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 296, + 281 + ], + "type": "text", + "content": "[4] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 316 + ], + "type": "text", + "content": "[5] Hanze Dong, Xi Wang, Yong Lin, and Tong Zhang. Particle-based variational inference with preconditioned functional gradient flow. arXiv preprint arXiv:2211.13954, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 318, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 295, + 360 + ], + "type": "text", + "content": "[6] Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you should treat it like one. ICML, 2022. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 295, + 396 + ], + "type": "text", + "content": "[7] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 441 + ], + "type": "text", + "content": "[8] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 295, + 474 + ], + "type": "text", + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 477, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 295, + 531 + ], + "type": "text", + "content": "[10] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "type": "text", + "content": "[11] Susung Hong. Smoothed energy guidance: Guiding diffusion models with reduced energy curvature of attention. arXiv preprint arXiv:2408.00760, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 568, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 295, + 611 + ], + "type": "text", + "content": "[12] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debiasing scores and prompts of 2d diffusion for view-consistent text-to-3d generation. Advances in Neural Information Processing Systems, 36:11970-11987, 2023. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "text", + "content": "[13] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462–7471, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[14] Susung Hong, Junyoung Seo, Heeseong Shin, Sunghwan Hong, and Seungryong Kim. Direct2v: Large language models are frame-level directors for zero-shot text-to-video generation. arXiv preprint arXiv:2305.14330, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[15] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. arXiv preprint arXiv:2309.14494, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "text", + "content": "[16] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12469-12478, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 173, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 554, + 205 + ], + "type": "text", + "content": "[17] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 206, + 554, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 554, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 554, + 247 + ], + "type": "text", + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 250, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 554, + 304 + ], + "type": "text", + "content": "[19] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 305, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 554, + 348 + ], + "type": "text", + "content": "[20] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. arXiv preprint arXiv:2304.06025, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 381 + ], + "type": "text", + "content": "[21] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "type": "text", + "content": "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 427, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 554, + 480 + ], + "type": "text", + "content": "[23] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 482, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 554, + 525 + ], + "type": "text", + "content": "[24] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 526, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 568 + ], + "type": "text", + "content": "[25] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13352-13361, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 570, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 554, + 602 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 604, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 554, + 669 + ], + "type": "text", + "content": "[27] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[28] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16301" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[29] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "text", + "content": "[30] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5773-5783, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "type": "text", + "content": "[31] Chang Liua and Jun Zhub. Geometry in sampling methods: A review on manifold mcmc and particle-based variational inference methods. Advancements in Bayesian Methods and Implementations, 47:239, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 295, + 282 + ], + "type": "text", + "content": "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7210-7219, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 295, + 326 + ], + "type": "text", + "content": "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 380 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 295, + 426 + ], + "type": "text", + "content": "[35] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "text", + "content": "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "type": "text", + "content": "[37] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "type": "text", + "content": "[38] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "type": "text", + "content": "[39] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 669 + ], + "type": "text", + "content": "[40] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[41] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 239 + ], + "type": "text", + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 306 + ], + "type": "text", + "content": "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[46] Junyoung Seo, Susung Hong, Wooseok Jang, Ines Hyeonsu Kim, Minseop Kwak, Doyup Lee, and Seungryong Kim. Retrieval-augmented score distillation for text-to-3d generation. arXiv preprint arXiv:2402.02972, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "type": "text", + "content": "[47] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 397, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 553, + 430 + ], + "type": "text", + "content": "[48] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 432, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 453 + ], + "type": "text", + "content": "[49] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 455, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 486 + ], + "type": "text", + "content": "[50] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 32, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 488, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 532 + ], + "type": "text", + "content": "[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 599 + ], + "type": "text", + "content": "[52] Pratul P Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7495-7504, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 600, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 655 + ], + "type": "text", + "content": "[53] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), pages 443-453. IEEE, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 657, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 553, + 678 + ], + "type": "text", + "content": "[54] Cyrus Vachha and Ayaan Haque. Instruct-gs2gs: Editing 3d gaussian splats with instructions, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[55] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16302" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 642 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[56] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[57] Can Wang, Ruixiang Jiang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 252 + ], + "type": "text", + "content": "[58] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12619-12629, 2023. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "text", + "content": "[59] Ziyu Wang, Tongzheng Ren, Jun Zhu, and Bo Zhang. Function space particle optimization for bayesian neural networks. arXiv preprint arXiv:1902.09754, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "text", + "content": "[60] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 387 + ], + "type": "text", + "content": "[61] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7378-7387, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "text", + "content": "[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 475 + ], + "type": "text", + "content": "[63] Hong-Xing Yu, Leonidas J Guibas, and Jiajun Wu. Unsupervised discovery of object radiance fields. arXiv preprint arXiv:2107.07905, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 476, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 294, + 521 + ], + "type": "text", + "content": "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 522, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 576 + ], + "type": "text", + "content": "[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 577, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 294, + 642 + ], + "type": "text", + "content": "[66] Luyang Zhu, Dawei Yang, Tyler Zhu, Fitsum Reda, William Chan, Chitwan Sahara, Mohammad Norouzi, and Ira Kemelmacher-Shlizerman. Tryondiffusion: A tale of two unets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4606-4615, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16303" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/0cae22af-0103-455d-a35f-1a510beabf2e_content_list.json b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/0cae22af-0103-455d-a35f-1a510beabf2e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b58a4784eaea7778e23fdfe61689074a90b0ff32 --- /dev/null +++ b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/0cae22af-0103-455d-a35f-1a510beabf2e_content_list.json @@ -0,0 +1,1934 @@ +[ + { + "type": "text", + "text": "PhD: A ChatGPT-Prompted Visual hallucination Evaluation Dataset", + "text_level": 1, + "bbox": [ + 151, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiazhen Liu $^{1*}$ , Yuhan Fu $^{1,2}$ , Ruobing Xie $^{2}$ , Runquan Xie $^{2}$ , Xingwu Sun $^{2}$ , Fengzong Lian $^{2}$ , Zhanhui Kang $^{2}$ , and Xirong Li $^{1\\dagger}$", + "bbox": [ + 137, + 179, + 857, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Renmin University of China \n $^{2}$ Machine Learning Platform Department, Tencent", + "bbox": [ + 299, + 233, + 697, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/jiazhen-code/PhD", + "bbox": [ + 336, + 272, + 650, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal Large Language Models (MLLMs) hallucinate, resulting in an emerging topic of visual hallucination evaluation (VHE). This paper contributes a ChatGPT-Prompted visual hallucination evaluation Dataset (PhD) for objective VHE at a large scale. The essence of VHE is to ask an MLLM questions about specific images to assess its susceptibility to hallucination. Depending on what to ask (objects, attributes, sentiment, etc.) and how the questions are asked, we structure PhD along two dimensions, i.e. task and mode. Five visual recognition tasks, ranging from low-level (object / attribute recognition) to middle-level (sentiment / position recognition and counting), are considered. Besides a normal visual QA mode, which we term PhD-base, PhD also asks questions with spcious context (PhD-sec) or with incorrect context (PhD-icc), or with AI-generated counter common sense images (PhD-ccs). We construct PhD by a ChatGPT-assisted semi-automated pipeline, encompassing four pivotal modules: task-specific hallucinatory item (hitem) selection, hitem-embedded question generation, spcious / incorrect context generation, and counter-common-sense (CCS) image generation. With over 14k daily images, 750 CCS images and 102k VQA triplets in total, PhD reveals considerable variability in MLLMs' performance across various modes and tasks, offering valuable insights into the nature of hallucination. As such, PhD stands as a potent tool not only for VHE but may also play a significant role in the refinement of MLLMs.", + "bbox": [ + 89, + 354, + 483, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 791, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Using a specific large language model (LLM) as its kernel, a multi-modal LLM (MLLM), exemplified by LLaVA [25], Qwen-VL [2] and MiniGPT-v2 [4] can now tackle a wide", + "bbox": [ + 89, + 816, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/51573b5c4bc12668482aa70070c9ac37cbe5d12a450d614eacfe6ba138b39e4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 321, + 589, + 371 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Are there five people in the image?", + "bbox": [ + 620, + 323, + 846, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yes, there are five people in the image.", + "bbox": [ + 611, + 344, + 831, + 357 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "They are standing near a bunch of bananas.", + "bbox": [ + 612, + 358, + 854, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2680ade224ae3647b4c20cb41abb6ee7eadff0ff185e54d85dd98db0b48659a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 874, + 349, + 900, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(a) Hallucination cause I: Visual ambiguity (MLLM: LLaVA-1.5) [25]", + "bbox": [ + 524, + 375, + 895, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/024fb122f8f287ca4e1de86ff2c9b38e902f56630418ed97f91a0bb71c9c373f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 396, + 583, + 430 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "After a long playtime with", + "bbox": [ + 584, + 397, + 683, + 406 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "its toy ball, the kitten, feeling", + "bbox": [ + 584, + 406, + 687, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thoroughly exhausted from", + "bbox": [ + 586, + 412, + 687, + 420 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thoroughly exhausted from all the fun, has finally found", + "bbox": [ + 586, + 420, + 687, + 429 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "its resting place. It has curled up on the table, seeking comfort and rest after its energetic adventures.", + "bbox": [ + 521, + 431, + 683, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/51556ece3eb4b096708dd3a43d21f7899e6822dabbbdc67f7ffe4356658ed86e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 398, + 710, + 416 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Is there a toy in the image?", + "bbox": [ + 725, + 401, + 882, + 414 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yes, the kitten is playing", + "bbox": [ + 705, + 424, + 849, + 438 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with a toy ball.", + "bbox": [ + 707, + 438, + 792, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/047bc6b5724bb26f26e69b59632039d8b495af91ce1499e71a4bb563e0261b2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 874, + 433, + 901, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(b) Hallucination cause II: Inconsistency in multi-modal input", + "bbox": [ + 545, + 458, + 872, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87bfd678fdbcb85444a4ba5ba0352e9f3533e3ea7e5e88b8dfb26f8286bbc7da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 479, + 591, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ad6cbdd5c082aceb923b355560b9b8013bdcab5928fb862d4394cf0058bb7b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 484, + 620, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Is the mouse in the image smaller than the cat?", + "bbox": [ + 640, + 482, + 789, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yes, the mouse is smaller than the cat.", + "bbox": [ + 643, + 513, + 856, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/68b0e82990025d2b9b673fa14738b3ab16ee1bd1f8d172f1dec2c0b9f6e4314e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 872, + 513, + 900, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(c) Hallucination cause III: Counter-common-sense content", + "bbox": [ + 553, + 534, + 866, + 546 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cbf98828d50186939dd14d9a1b6540911aad192ae7bca33e1ccbcc632766c070.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 556, + 895, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(d) Performance curves of the LLaVA series on two public VHE datasets (POPE [21]) and AMBER [34]) and the proposed PhD dataset.", + "bbox": [ + 514, + 733, + 901, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Illustrations of three major causes of an MLLM's visual hallucination and its evaluation. This paper contributes PhD, a binary VQA-based VHE benchmark, much larger and more challenging than its predecessors. In particular, it has four evaluation modes that explicitly measure an MLLM's performance w.r.t. the three causes, i.e. PhD-base for cause I, PhD-sec and PhD-icc for cause II and PhD-ccs for cause III.", + "bbox": [ + 511, + 768, + 906, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*J. Liu is currently pursuing his PhD at HKUST.", + "bbox": [ + 107, + 875, + 367, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author (xirong@ruc.edu.cn)", + "bbox": [ + 109, + 887, + 344, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19857", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "range of computer vision tasks in a unified visual-question-answering (VQA) manner. As LLMs are known to hallucinate [20, 37, 42], it is not surprising that MLLMs have visual hallucination, generating fabricated interpretation of the given visual content, see Fig. 1. Considering the rapidly growing use of MLLMs in varied scenarios, a comprehensive visual hallucination evaluation (VHE) is crucial. This paper develops a new dataset for VHE.", + "bbox": [ + 88, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VHE essentially involves posing a number of visual questions to an MLLM [11, 21]. A question of this kind shall include a hallucinatory item (hitem), in the form of a specific word or phrase, that induces the MLLM to generate a response discordant with the provided visual content. As the model typically has a strong visual recognition ability, how to identify an appropriate hitem and accordingly generate a proper question is nontrivial. Both the hitem and the question depend on the visual recognition task being considered. As shown in Tab. 1, we target at objective VHE in the context of low-level (object / attribute) to middle-level (sentiment / position / counting) visual recognition tasks. Such a target is chosen due to the following considerations. MLLMs generally work well for these tasks, so their erroneous responses can be largely attributed to their hallucinations instead of their incapability, e.g. asking a generic MLLM to read pathology images [14]. Meanwhile, a binary VQA based objective evaluation is more budget friendly and thus more suited for VHE at a large scale.", + "bbox": [ + 93, + 212, + 485, + 500 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/82205532c55138e48e80e1268da60463ee8dc39231e3ff79af6e666eb58028a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Vision tasksObjective evaluationSubjective evaluation
Low-/middle-level visual recognition+ POPE, EMNLP'23 [21]+ FAITHSCORE, EMNLP'24 [15]
+ AMBER, arXiv'23 [34]+ HaELM, arXiv'23 [35]
+ CIEM, ITIF'23 [13]+ M-HalDetect, AAAI'24 [12]
+ NOPE, ALVIR'24 [27]+ GAVIE, ICLR'24 [23]
+ ROME, EMNLP'23 [43]
+ PhD (this paper)
High-level visual reasoning+ MMMU, CVPR'24 [40]+ HallusionBench, CVPR'24 [11]
+ VLind-Bench, NAACL'25 [17]+ Bingo, arXiv'23 [6]
+ IllusionVQA, COLM'24 [31]
+ WHOOPS!, ICCV'23[3]
", + "bbox": [ + 91, + 513, + 483, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pioneered by POPE [21], good attempts exist in objective VHE [13, 27, 34, 43]. In these valuable datasets, hitem selection is largely untouched. POPE and ROME [43] select their hitems fully based on label co-occurrence in training data, AMBER [34] relies on manual annotation, whilst hitem selection is not considered in NOPE [27] and CIEM [13], see Tab. 2. Hence, there lacks an explicit link between hitem selection (and subsequent VQA triplets construction) and major causes of an MLLM's visual hallucination. As models rapidly evolve, the performance on these datasets quickly reaches saturation, see Fig. 1d.", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Analyzing an MLLM's typical dataflow of answering a visual question, we see three major causes of visual hallucination: I) visual ambiguity, II) inconsistency in multi-modal input and III) counter-common-sense (CCS) content, see Fig. 1. Firstly, the MLLM extracts tokenized visual features from a given image using a ViT based encoder. Recent studies [33, 41] show that the features tend to be high level, lacking sufficient details for fine-grained tasks such as counting. Secondly, the visual features, after vision-to-language adaptation, are mixed with the features of the associated textual prompt to form a multi-modal input to the LLM kernel. The LLM, pre-trained extensively on textual data, inevitably favors the textual part of the multi-modal input. Hence, when there is inconsistency between the visual and textual information, the former is more likely to be overruled. Lastly, at the decoding stage, the LLM might heavily rely on its internal (world) knowledge, ignoring the visual content especially when the content (showing a mouse much larger than a cat) contradicts the common sense. Our new dataset is developed with a close link to the three causes.", + "bbox": [ + 511, + 90, + 906, + 393 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/5aa706bb2c4620c69e58c0f787816ca90c73fe8ce9eceb219a6ec25eab031356.jpg", + "table_caption": [ + "Table 1. Taxonomy of VHE benchmarks. Our PhD benchmark performs an objective evaluation of MLLMs' hallucinations when they address visual recognition tasks ranging from low-level, i.e. object / attribute recognition to middle-level, i.e. sentiment / positional recognition and counting." + ], + "table_footnote": [], + "table_body": "
DatasetDaily imagesCCS imagesHItemsContextsVQA tripletsTasks
POPE500✘80✘3,000Obj.
NOPE*unknown✘unknown✘32,701Obj.
CIEM*4,929✘unknown✘72,941Obj. / Attr. / Pos.
AMBER1,004✘687✘14,216Obj. / Attr./Pos. / Count.
ROME✘1,563118✘1,563Attr./Pos.
PhD14,6487501,45233,688102,564Obj / Attr. / Pos. / Sent. / Count.
", + "bbox": [ + 513, + 404, + 906, + 497 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 2. PhD versus its predecessors. * indicates private dataset.", + "bbox": [ + 513, + 508, + 903, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The new dataset is constructed by adapting TDIUC, a popular multi-task VQA dataset [16], with a ChatGPT-assisted semi-automated pipeline, see Fig. 2. In particular, by prompting ChatGPT, we select diverse and visually challenging h items in an image-specific and task-specific manner, with minimal human involvement primarily spent on verifying ChatGPT-generated results. The selected h items are then automatically embedded into visual questions, specious context, and incorrect context, all generated by instructing ChatGPT. Moreover, we expand the daily image set with counter common sense (CCS) images, obtained by prompting AIGC tools with ChatGPT-generated CCS descriptions, e.g. \"trees growing underwater\" and \"a car with square-shaped wheels\". The dataset is dubbed as PhD (ChatGPT Prompted visual hallucination evaluation Dataset). Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base (questions about daily images w/o context), PhD-sec (PhD-base plus specious context), PhD-icc (PhD-base plus incorrect context), and PhD-ccs (questions about CCS images).", + "bbox": [ + 511, + 537, + 908, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, our major contributions are as follows:", + "bbox": [ + 535, + 856, + 867, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce PhD, a dataset with four evaluation modes across five visual recognition tasks, developed with a close", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19858", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "link to the three major causes of MLLM visual hallucination. Information on hallucinatory items (hItems) is provided per sample, enabling in-depth analytics to uncover the causes in more detail.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We offer a ChatGPT-assisted semi-automatic pipeline for dataset construction, with minimal human involvement, primarily focused on verifying the generated results. With 14,648 daily images, 750 CCS images and 102,564 VQA triplets in total, PhD is the largest of its kind.", + "- We conduct an extensive evaluation with 15 open-source MLLMs, 3 proprietary MLLMs, and 2 hallucination mitigation methods, showing the viability of PhD for VHE in varied manners including overall, mode-oriented, task-oriented, and model-wise zoom-in. The evaluation not only reveals inter-model performance divergence, but may also help developers of a specific MLLM to prioritize their efforts in refining the model." + ], + "bbox": [ + 88, + 151, + 480, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 361, + 232, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the increasing importance of VHE, new benchmarks are being actively developed. Depending on what vision tasks they focus on and how their evaluation is executed, we categorize existing achievements along two dimensions, i.e. task and evaluation, see Tab. 1. Concerning the task dimension, low- and middle-level visual recognition tasks, ranging from object / attribute recognition to sentiment / positional recognition and counting, assess an MLLM's basic visual skills. High-level visual reasoning is more domain-knowledge intensive, typically covering image-based math question solving, geography information understanding, meme interpretation, historical or folkloric contexts, etc. [3, 6, 11, 17]. As for the evaluation dimension, objective evaluation refers to objectively comparing the model's output with ground truth, mostly in the form of Yes/No answers [13, 21, 34, 43]. By contrast, subjective evaluation requires humans or LLMs to assess the model's output [3, 6, 11]. The proposed PhD, focusing on low/middle-level visual recognition and objective evaluation, belongs to the second quadrant of the taxonomy. In what follows, we briefly review peer benchmarks, i.e. POPE [21], ROME [43], NOPE [27], CIEM [13], and AMBER [34], in this quadrant and clarify our novelty accordingly.", + "bbox": [ + 91, + 387, + 482, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "POPE is probably the first dataset to evaluate object hallucination [21]. Given a specific MS-COCO image with object labels, POPE selects an adversarial object frequently co-occurring with the current labels. A binary question is then formed by filling out a predefined template with the selected object. Such co-occurrence based hitem selection is not image-specific by definition. Trivial hitems might thus be picked up, e.g. \"car\" selected for an indoor image labeled with \"person\", as the two objects often co-occur. For CCS image generation, ROME forms CCS descriptions by choosing attribute values having the lowest co-occurrence", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with a given object according to the Visual Genome dataset [43]. As low occurrence is not necessarily CCS, some of ROME images are indeed normal. NOPE [27] and CIEM [13] simply bypass hitem selection by asking an LLM to generate questions conditioned on the image captions (also from MS-COCO) and pre-specified answers.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To select h items in an image-specific manner, AMBER resorts to fully manual annotation [34]. Manual labeling is costly, while an annotator's personal vocabulary is relatively limited. All this makes it difficult to scale up w.r.t. the amount of test images and the number of distinct h items.", + "bbox": [ + 511, + 181, + 903, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In comparison, PhD, constructed by a ChatGPT-assisted semi-automatic pipeline (Sec. 3), is much larger (Tab. 2) and more challenging (Fig. 1d). With its unique mode-task structure, the new dataset enables a novel, structured and zoom-in understanding of inter-model difference.", + "bbox": [ + 511, + 258, + 903, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Roadmap to PhD", + "text_level": 1, + "bbox": [ + 513, + 348, + 723, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As MLLMs perform visual recognition in a VQA manner, a VQA sample for VHE naturally depends on the visual recognition task in consideration. We depart from TDIUC [16], a large-scale VQA dataset w.r.t. five tasks including object / attribute / sentiment / positional recognition and counting. Note that the images in TDIUC are sourced from MS-COCO [22], which plausibly have been seen by MLLMs in their development stage. As such, our adoption of TDIUC makes on-purpose data leakage: an MLLM's erroneous response w.r.t. a seen image can now be more safely attributed to its hallucination other than its incapability in visual recognition, say asking LLaVA to recognize glaucoma from color fundus photographs [36]. We construct PhD by adapting the TDIUC annotations with a ChatGPT1-assisted semi-automated pipeline, see Fig. 2.", + "bbox": [ + 511, + 375, + 903, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to compose a proper question that effectively makes an MLLM hallucinate about a given image, a hitem has to be first identified in an image-specific and task-specific manner. Then, the hitem has to be smoothly embedded in the form of a specific word or phrase into the question. We describe task-specific hitem selection in Sec. 3.1, followed by hitem-embedded question generation in Sec. 3.2. Furthermore, in order to simulate inconsistency in the multi-modal input, we prepend specious or incorrect context to the question, the generation of which is detailed in Sec. 3.3. Lastly, in order to explicitly create conflicts between the visual input and the internal (world) knowledge of the MLLM, we expand our image collection with autogenerated counter-common-sense (CCS) images (Sec. 3.4).", + "bbox": [ + 511, + 601, + 903, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task-specific Hitem Selection", + "text_level": 1, + "bbox": [ + 511, + 823, + 774, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Without loss of generality, we describe how a hitem is selected for color attribute recognition. Let us consider the", + "bbox": [ + 511, + 845, + 903, + 875 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1We use GPT-4o mini, released on 2024-05-13.", + "bbox": [ + 531, + 886, + 782, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19859", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/646c6dc4e7a35c48032ead0b4d1db978eb5c5caa8db07719269bf1c711925893.jpg", + "image_caption": [ + "(1) Task-specific hitem selection" + ], + "image_footnote": [], + "bbox": [ + 101, + 103, + 658, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a1fb24641db27d7e7c817515a0c3e9246c98bc3f0d50ba9cb3edf4792af3e5d.jpg", + "image_caption": [ + "(2) Hitem-embedded question generation", + "(a) Hitem-embedded question / context generation for MS-COCO daily images" + ], + "image_footnote": [], + "bbox": [ + 101, + 181, + 658, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4bf9d8d0909bb5ec79279fc3b476c78daea8d7ea2907d374a352ab0d79dd4389.jpg", + "image_caption": [ + "(b) QAs for AI-generated CCS images" + ], + "image_footnote": [], + "bbox": [ + 668, + 104, + 898, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a20f5410384ac7b3cacd15ece0e2ac60f70ab590b49b5a2ddf9a0be4e0eda7b3.jpg", + "image_caption": [ + "Task: Object recognition", + "Are there headphones in the image?" + ], + "image_footnote": [], + "bbox": [ + 99, + 282, + 222, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In recent years, this musician has often immersed himself in his own musical world, playing the guitar with headphones on as if cut off from the outside world. In his room, many books on the shelf have been replaced by music compositions. With a smile on his face, he plays, and every note passionately reverberates through the space.", + "bbox": [ + 220, + 285, + 375, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the photo, a man confidently plays his guitar with headphones on, fully immersed in his music. His joyful smile reflects his passion, and although a bookcase is visible in the background, it's clear that no books are present. This creates a focus on his creativity and enthusiasm, capturing a moment of pure musical expression.", + "bbox": [ + 220, + 332, + 375, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bbd20cbdfed5eade34256728b8a4b1d4e7689ed698daa4ce36e60cf75b07c54c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 382, + 282, + 656, + 383 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6a3c223e92b7b95afad980d10c040a3748a98106737e438852ad81adab4c6a36.jpg", + "image_caption": [ + "Are there books in the image?", + "Task: Sentiment recognition", + "Is the girl feeling lonely in the image?", + "Is the girl feeling happy in the image?" + ], + "image_footnote": [], + "bbox": [ + 99, + 387, + 220, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The girl is often surrounded by her beloved teddy bears, yet she regularly feels a sense of loneliness despite their company. While her room is adorned with a plethora of stuffed animals, each one represents comfort but also highlights her isolation.", + "bbox": [ + 220, + 388, + 375, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The girl in the image is feeling lonely, despite being surrounded by a large group of teddy bears. She lies on the bed near many stuffed animals, including dozens of brown teddy bears, and her contrast between her solitary emotions and the cheerful presence of her furry companions.", + "bbox": [ + 220, + 426, + 374, + 468 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/baa837148544a3bed93a35352ba8aa6b0d125a73da2247590bfc4c748b3f6e36.jpg", + "image_caption": [ + "Task: Positional", + "Is there a table behind", + "the looking dog in the image?", + "Is there a fence behind", + "the looking dog in the image?" + ], + "image_footnote": [], + "bbox": [ + 382, + 387, + 503, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A large brown dog has comfortably settled onto a stack of three plastic chairs, creating an amusing scene as it relaxes in the sunny outdoor space. Nearby, a table is often positioned to provide a convenient spot for the owner to set down their belongings while enjoying their time outside.", + "bbox": [ + 504, + 388, + 656, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a charming scene, a dog has made itself comfortable on a stack of plastic chairs, effortlessly showcasing its playful nature. Behind the relaxed dog captured in the photo, a table can be filled with a large box of cozy atmosphere. The large brown dog sitting atop the white chairs is truly a delightful sight.", + "bbox": [ + 504, + 426, + 656, + 468 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9cd613d5a37f9da3944ee086a421297a86ef13c5cd1187109a1bff7ef2e359b.jpg", + "image_caption": [ + "Is the ball being kicked a soccer ball?", + "Is the ball being kicked a basketball?" + ], + "image_footnote": [], + "bbox": [ + 669, + 304, + 779, + 367 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7df577b988cb09d7417eb469eacf4517b97d91a35d1b4eaca4d6886cf878a60b.jpg", + "image_caption": [ + "Is the mouse smaller than the cat?" + ], + "image_footnote": [], + "bbox": [ + 787, + 305, + 897, + 366 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aeece98e4fc0c4739e290283298a8866bb334f01e7c7b489bf23def3ae8aa6fa.jpg", + "image_caption": [ + "Is the highest number", + "of points on the dice six?", + "Is the highest number", + "points on the dice seven?" + ], + "image_footnote": [], + "bbox": [ + 669, + 390, + 779, + 441 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0c79a4a059a2df4dee4322880a54350267ee5874b09b19e5c68c90b2b52f62ab.jpg", + "image_caption": [ + "Is the cat smaller than the mouse?", + "Is the shape of the orange round?", + "Is the shape of the orange square?", + "(d) Showcase: CCS images and questions", + "Figure 2. Proposed semi-automatic pipeline for PhD construction. We use ChaptGPT (GPT-4o mini) to generate hitem-embedded questions / contexts for daily images, and Doubao and DALL-E3 for generating CCS images. Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base, i.e. questions about daily images w/o context, PhD-sec, i.e. PhD-base plus specious context, PhD-icc, i.e. PhD-base plus incorrect context, and PhD-ccs, i.e. questions about CCS images. By adapting TDIUC annotations, PhD supports binary VQA w.r.t. five visual recognition tasks including object / attribute / sentiment / positional recognition and counting. With 20 mode-task combinations in total, PhD enables a comprehensive VHE." + ], + "image_footnote": [], + "bbox": [ + 787, + 390, + 897, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "image in Fig. 2a, showing a black motorcycle followed by a red bus. As no red motorcycle is present while the red color is prominent near the black motorcycle, the word red will be a good choice of hitem to challenge an MLLM.", + "bbox": [ + 89, + 607, + 482, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vocabulary Construction per task. Started with a handful of manually specified colors such as red, green, and blue, we ask ChatGPT to expand the color vocabulary with instructions like \"Please expand the input vocabulary as much as possible by adding common items found in daily life. Avoid any duplication\", getting a set of 35 different colors.", + "bbox": [ + 89, + 669, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subject-Attribute Extraction. The image is associated with a TDIUC question-answer pair as \"what color is the motorcycle\" and \"black\". We use ChatGPT (with simple instructions) to extract with ease the subject (i.e. motorcycle) and its attribute (i.e. black) from the pair.", + "bbox": [ + 89, + 761, + 482, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Candidate Hitem Generation. We obtain candidate h items by excluding the ground-truth (GT) answer (and its synonyms if applicable) from the vocabulary.", + "bbox": [ + 89, + 838, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual-based Hitem Ranking. Intuitively, a hitem shall", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "be visually plausible in the given image. So for each candidate hitem, we compute its similarity to the image using a pre-trained CLIP [30]. In particular, the cosine similarity between the CLIP embeddings of hitem + subject (e.g. green motorcycle) and the image is adopted. Ranking the candidates by the CLIP similarity lets us to select the one visually closest to the image. It is worth noting that for emerging MLLMs equipped with stronger vision encoders [19], our pipeline is likely to produce even more effective h items by replacing CLIP with these advanced counterparts.", + "bbox": [ + 511, + 607, + 903, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Manual Inspection. While the above process is generally stable to produce satisfying results, manual inspection is performed to ensure the correctness of hitem selection. Note that due to errors in the original TDIUC annotations, occasionally the true label might be \"incorrectly\" selected. In such a case, we simply discard the VQA sample.", + "bbox": [ + 511, + 761, + 903, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With lightweight task-specific adaptation, the above hitem selection also works for other tasks. Overall, the joint use of ChatGPT and CLIP allows us to select 1,452 hitem", + "bbox": [ + 511, + 854, + 903, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19860", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that are more diverse and challenging than their counterparts in previous datasets, see Tab. 2.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Hitem-embedded Question Generation", + "text_level": 1, + "bbox": [ + 89, + 133, + 424, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a given subject (e.g. motorcycle) and a chosen hitem (e.g. red), generating a hitem-embedded question is trivial for ChatGPT. In particular, a No question is formed as \"Is the motorcycle in the image red?\" Meanwhile, a Yes question is simultaneously generated using the GT as \"Is the motorcycle in the image black?\". This ensures perfect Yes/No balance among the generated questions.", + "bbox": [ + 89, + 157, + 483, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Specious (Incorrect) Context Generation", + "text_level": 1, + "bbox": [ + 89, + 277, + 439, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When used as a document parser, an MLLM reading a specific image is often provided with the image's surrounding text. Inconsistency between the image and the text is not uncommon. A news article containing a general claim of \"red motorcycles frequently zip through the streets\" does not necessarily have each of its illustrated pictures match with the claim. To simulate such a scenario, for a given image we generate specious text as specious context and text contradicting the image as incorrect context, respectively. Next, we describe the generation of specious contexts, as their incorrect counterparts can be generated in a similar but more simplified manner.", + "bbox": [ + 89, + 300, + 483, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specious Text Generation. Using the previously generated hitem-embedded question and the original MS-COCO captions as input, we instruct ChatGPT to generate specious text for a given image. By \"specious\", we mean the text is specious or noisy, rather than directly contradicting the image content. As such, our instruction reads partially as \"Please generate the for the given question. It should be one sentence. The should answer the question, but it may not reflect the actual current status, thus making it specious.\"", + "bbox": [ + 89, + 483, + 483, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text Composition. ChatGPT is used to seamlessly merge the specious text with ground-truth captions, forming a longer context in which only a small portion (orange text in Fig. 2c) is mildly inconsistent with the image.", + "bbox": [ + 89, + 635, + 482, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Manual Inspection. We perform a spot check. If the context quality is low, we simply discard the entire sample.", + "bbox": [ + 89, + 696, + 483, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. CCS Image Generation", + "text_level": 1, + "bbox": [ + 89, + 739, + 307, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We generate CCS images by first generating CCS descriptions and then employing Text2Image tools to convert the descriptions to CCS images, see Fig. 2b.", + "bbox": [ + 89, + 763, + 482, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CCS Description Generation. A number of manually written task-specific samples, see Tab. 3, are used as in-context learning samples for ChatGPT to generate more descriptions. The descriptions have to be visually expressible, so bad cases like \"the more you eat, the thinner you get\" are filtered out manually. For each CCS text, its common-sense", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(CS) counterpart is simultaneously generated by ChatGPT, by providing the learning samples in pair.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text2Image. The generated CCS descriptions are used as prompts for AIGC tools (Doubao [9] and DALL-E3 [28]) to generate the corresponding CCS images. The quality of the generated images depends on various factors, making occasional failures inevitable. When this occurs, we attempt to refine the prompts or apply region-based inpainting. The sample will be discarded if the above attempts fail.", + "bbox": [ + 511, + 122, + 906, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question Generation. Per CCS description (e.g. A car with square wheels), we utilize ChatGPT to generate a Yes question (e.g. Does the car have square wheels?). Again, for balancing Yes/No questions, we generate a No question (e.g. Does the car have round wheels?) based on the CS description.", + "bbox": [ + 511, + 228, + 906, + 319 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8e4e50939d9f18731c969014c03e870a630d840b249170c1d51d389e2476e000.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CCS descriptionYes questionCS descriptionNo question
Task: Object recognition
Manually written:
Ice blocks in volcanic lavaAre there ice blocks in volcanic lava?Fire in volcanic lavaIs there fire in volcanic lava?
Grass in a tiger's mouthIs there grass in a tiger's mouth?Meat in the tiger's mouthIs there meat in the tiger's mouth?
ChatGPT generated:
Trees growing underwaterAre there trees growing underwater?Coral growing underwaterIs there coral growing underwater?
Books in a swimming poolAre there books in a swimming pool?Water in a swimming poolIs there water in a swimming pool?
Birds flying underwaterAre there birds flying underwater?Birds flying in the skyAre there birds flying in the sky?
Ice cream in a volcanoIs there ice cream in a volcano?Lava in a volcanoIs there lava in a volcano?
Computers in a forestAre there computers in a forest?Animals in a forestAre there animals in a forest?
Task: Attribute recognition
Manually written:
A car with square wheelsDoes the car have square wheels?A car with round wheelsDoes the car have round wheels?
Blue apples on the treeAre the apples on the tree blue?Red apples on the tree.Are the apples on the tree red?
ChatGPT generated:
A green skyIs the sky green?A blue skyIs the sky blue?
A bicycle with square wheelsDoes the bicycle have square wheels?A bicycle with round wheelsDoes the bicycle have square wheels?
A tree made of metalIs this tree made of metal?A wooden treeIs this tree a real wood tree?
A chocolate riverIs there chocolate flowing in the stream?A water riverIs there water flowing in the stream?
A house made of candyIs the house made of candy?A house made of bricksIs the house made of bricks?
", + "bbox": [ + 514, + 330, + 906, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Instances of descriptions used for generating CCS images and related questions. With manually written CCS / CS descriptions as instructions, ChatGPT is used to generate many more instances and subsequently convert them to Yes/No questions.", + "bbox": [ + 511, + 688, + 906, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Dataset Overview and PhD Index", + "text_level": 1, + "bbox": [ + 513, + 772, + 805, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An overview of the PhD dataset is given in Tab. 4. Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base, i.e. questions about daily images w/o context, PhD-sec, i.e. PhD-base plus spcious context, PhD-icc, i.e. PhD-base plus incorrect context, and PhD-ccs, i.e. questions about CCS images. With 20 mode-task com", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19861", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "binations in total, PhD supports a much more comprehensive VHE than its predecessors [21, 34].", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3eab5dea12d5ada065cf1dbf14c24553afe5da193988a4473da71ef41adea063.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TasksQuestions
ObjectAttributeSentimentPositionCountingYesNo
TDIUC samples used6,2714,3242,0952,8413,387--
Unique h items7451466548610--
VQA samples in PhD-base11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-sec11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-icc11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-ccs34473478220124750750
", + "bbox": [ + 91, + 133, + 483, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To measure the performance of an MLLM on PhD, we compute its recall w.r.t. the Yes and No questions, respectively. We term the harmonic mean of the two recalls PhD Index. A model simply saying Yes (or No) to all questions has a PhD Index of 0, while a random-guess score is 0.5.", + "bbox": [ + 89, + 250, + 483, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Evaluating MLLMs on PhD", + "text_level": 1, + "bbox": [ + 89, + 338, + 351, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Common Setup", + "text_level": 1, + "bbox": [ + 89, + 363, + 246, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Choices of MLLMs. For reproducible research, we focus on open-source MLLMs, compiling a list of 15 models that span varied sizes and architectures. see Tab. 5. We also evaluate two hallucination mitigation methods, VCD [18] and Woodpecker [39], currently supporting LLaVA-1.6-L and Qwen-VL. As additional references, we assess three proprietary MLLMs, i.e. GPT-4o [29], Claude 3.5 Sonnet [1], and Gemini 1.5 Pro [10], on a random subset of 2k samples (random-2k) subject to our budget. For the same reason we evaluate Woodpecker, which requires paid service from ChatGPT, on random-2k.", + "bbox": [ + 89, + 385, + 483, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test Protocol. For a fair comparison, per MLLM we use its designated prompt to wrap each test question. For instance, a question-specific prompt submitted to mPLUG-Owl2 will be in the form of \"USER: $< |image|>\\{question\\}$ ASSISTANT:\" We provide more prompts in the supplement. In addition, to help the models better handle PhD-sec and PhD-icc, we append to the test prompt an instruction as \"In case there is an inconsistency between the context and the image content, you should follow the image.\"", + "bbox": [ + 89, + 551, + 483, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Using PhD for Overall VHE", + "text_level": 1, + "bbox": [ + 89, + 712, + 344, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An overall VHE as shown in Tab. 5 is useful for providing a big picture of which MLLM hallucinates the most (or the least). The leading open-source MLLMs are LLaVA-OneVision, followed by Molmo and InternVL-1.5. Since their vision encoders and LLMs vary, the results are insufficient to conclude which component is the most effective to mitigate hallucinations. That said, comparisons among the same model series remains meaningful. Consider the LLaVA series for instance. While one would normally expect that a larger LLM yields a better MLLM, as LLaVA-1.6-L vs LLaVA-1.6, the difference between LLaVA-1.5-L", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and LLaVA-1.5 is marginal (0.270 vs 0.265). In order to analyze and consequently understand such an counterintuitive result, PhD enables a zoom-in analysis in mode-oriented (Sec. 4.3) and task-oriented (Sec. 4.4) styles, unavailable in the previous benchmarks.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/de797e374154d14af8b98f73f88317b4b828bdd4098ec16f803de465a50bec4b.jpg", + "table_caption": [ + "Table 4. Data statistics of the proposed PhD dataset." + ], + "table_footnote": [], + "table_body": "
ModelViTLLMPOPEAMBERPhD
Full-set evaluation:
LLaVA-OneVision [19]SoViT-400m/14Qwen2-72B0.840.900.698
Molmo [8]-L/14Qwen2-72B0.840.850.690
InternVL-1.5 [5]InternViT-6BInternLM2-20B0.860.890.561
Qwen-VL (VCD)-bigG/14Qwen-7B0.840.870.560
Cambrian-1 [32]HybridLlama-3-8B0.880.890.547
LLaVA-1.6-L (VCD)-L/14Vicuna-13B-1.50.820.810.511
LLaVA-1.6-XL [26]-L/14Nous-Hermes-2-Yi-34B0.860.840.492
Qwen-VL [2]-bigG/14Qwen-7B0.830.840.488
LLaVA-1.6-L [26]-L/14Vicuna-13B-1.50.830.800.423
MiniGPT-v2 [4]-G/14Llama-2-7B0.830.840.390
LLaVA-1.6 [26]-L/14Vicuna-7B-1.50.830.790.373
mPlug-Ow12 [38]-L/14Llama-2-7B0.780.770.320
InstructBLIP [7]-G/14Vicuna-7B-1.10.820.820.305
InstructBLIP-L [7]-G/14Vicuna-13B-1.10.800.790.278
LLaVA-1.5-L [25]-L/14Vicuna-13B-1.50.820.730.270
LLaVA-1.5 [25]-L/14Vicuna-7B-1.50.810.750.265
LLaVA-1.1 [24]-L/14Vicuna-7B-1.10.670.330.135
Random-2k evaluation:
GPT-4o [29]--0.880.870.812
Claude 3.5 Sonnet [1]--0.850.890.746
Gemini 1.5 Pro [10]--0.860.880.691
Qwen-VL (Woodpecker)-bigG/14Qwen-7B--0.531
LLaVA-1.6-L (Woodpecker)-L/14Vicuna-13B-1.5--0.409
", + "bbox": [ + 514, + 180, + 903, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Overall VHE. POPE (adversarial) and AMBER (discriminative) are used.", + "bbox": [ + 511, + 412, + 905, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One more advantage of PhD compared to its predecessors lies in its discrimination ability. The relatively small performance gap between GPT-4o and the top open-source models as measured by POPE and AMBER might lead to an overly optimistic interpretation that the open-source alternatives are catching up with the proprietary model. In fact, a substantial gap remains, as revealed by PhD. Fig. 3 further illustrate the qualitative results, where GPT-4o exhibits fewer hallucinations in its response.", + "bbox": [ + 511, + 458, + 906, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Using PhD for Mode-Oriented VHE", + "text_level": 1, + "bbox": [ + 511, + 604, + 826, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 4a illustrates mode-wise model performance. MLLMs working relatively well in the PhD-base mode tend to have stronger visual input. This is achieved either with stronger visual encoders, as the cases of LLaVA-OneVision, InternVL-1.5, and Cambrian-1 using SoViT-400m/14, InternViT-6B or hybrid vision structure, or supporting higher image resolutions, see Molmo and MiniGPT4-v2 that accept multiscale or larger input.", + "bbox": [ + 511, + 628, + 905, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast to the visual part, using a larger LLM alone does not necessarily lead to a better MLLM. As noted in Sec. 4.2, LLaVA-1.5-L with Vicuna-13B-1.5 has nearly the same PhD Index (0.270) as LLaVA-1.5 with Vicuna-7B-1.5 (0.265). We see from Fig. 4a that the larger LLM indeed improves the performance in PhD-sec $(0.082\\rightarrow 0.099)$ , PhD-icc $(0.011\\rightarrow 0.019)$ and PhD-ccs $(0.534\\rightarrow 0.542)$ , yet suffers loss in PhD-base $(0.443\\rightarrow 0.422)$ . Similar results can be more evidently observed in the case of InstructBLIP-L vs. InstructBLIP (PhD-base Index: $0.535\\rightarrow 0.324$ ). Our con", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19862", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/28682202df7b5fde01e490615c496a5370c413a8b54fb7d34ad2e55dd4bd55bc.jpg", + "image_caption": [ + "Figure 3. Qualitative results showing how an MLLM answers visual questions from PhD. The correctness of an answer is automatically determined by matching its first word, either Yes or No, with the ground truth (GT)." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 906, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "jecture is that although a larger LLM better understands user instructions, its successful use within an MLLM requires more targeted training for vision-language alignment.", + "bbox": [ + 88, + 369, + 482, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing the four modes, the performance of the open-source MLLMs on PhD-icc and PhD-sec is generally low. When provided with a multi-modal input, the models favor the textual part. Substituting a 13-B LLM for its 7-B counterpart helps tackling the inconsistency in the multi-modal input, see Fig. 5. However, a larger LLM might rely more on its internal knowledge for decoding, resulting in worse performance on PhD-base and PhD-ccs for which the image content shall carry more weights. In particular, PhD-ccs reveals a deeper and intrinsic challenge of VH: conflicts between the given image content and the model's internal knowledge. Solving the challenge demands a more comprehensive approach that goes beyond isolated improvements on the visual or language components.", + "bbox": [ + 88, + 416, + 482, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Among the open-source MLLMs, LLaVA-OneVision is the best, owing to its joint use of a stronger visual encoder (SoViT-400m/14), a more powerful LLM (Qwen2-72B), and better training strategies (much larger high-quality training data and multi-stage alignment). Nevertheless, LLaVA-OneVision remains inferior to GPT-4o, particularly in PhD-sec and PhD-icc, wherein inconsistency within the multi-modal input has to be properly addressed.", + "bbox": [ + 89, + 628, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Using PhD for Task-Oriented VHE", + "text_level": 1, + "bbox": [ + 89, + 757, + 395, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4b presents task-oriented VHE results. In general, to what extent an MLLM hallucinates is largely correlated to the required level (low / middle / high) of a specific task. The object recognition task has the overall highest PhD Index, followed by attribute recognition, positional recognition, counting, and sentiment recognition. Due to the complexity and subtlety of emotions, e.g. tears can be associated with both happiness and sadness, even GPT-4o performs", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "relatively worse in the sentiment task (first row of Fig. 3).", + "bbox": [ + 511, + 369, + 893, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Joint mode-task analytics per model is shown in Tab. 6. LLaVA-OneVision struggles with sentiment recognition and counting, especially when faced with textual or CCS distractions, underscoring the need for improvement in these areas. Similarly, Molmo also faces these challenges, but its counting performance under CCS distractions is notably better than LLaVA-OneVision's (0.737 vs 0.563). The above zoom-in analytics will be informative for MLLM developers to prioritize their efforts on model refinement.", + "bbox": [ + 511, + 386, + 905, + 523 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/950942eb7f2c43d25c9910bf0b713e71b491f73238e5e37a0a67f2e2aaf4866a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskPhD-basePhD-secPhD-iccPhD-ccs
LLAVA-OneVision
Object0.8720.8490.8240.727
Attribute0.8480.7440.6630.767
Sentiment0.6910.5810.5040.731
Positional0.7730.7300.6540.701
Counting0.7070.6520.5000.563
Molmo
Object0.8250.8800.8470.678
Attribute0.8420.7250.5560.791
Sentiment0.5470.6020.5680.746
Positional0.6970.6910.6540.742
Counting0.7270.5800.3500.737
", + "bbox": [ + 555, + 536, + 870, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Zoom-in analytics of specific models.", + "bbox": [ + 566, + 698, + 849, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Analysis of MLLM Answer Tendency", + "text_level": 1, + "bbox": [ + 511, + 741, + 836, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While the yes and no questions are perfectly balanced by design, we observe that the open-source MLLMs tend to answer yes, with the say-yes rate ranging from 0.462 (Molmo) to 0.811 (LLaVA-1.1) and an average value of 0.611. By contrast, the three proprietary MLLMs have a clearly lower say-yes rate. Similar observations are reported in [21, 23]. We go one step further by analyzing how the say-yes tendency is related to model performance. Ranking the models by their PhD Index and say-yes rate,", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19863", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5d732949e14a23a364068646dcb2c5ff8634ed802d965cb771589d5b17cdcdc.jpg", + "image_caption": [ + "(a) Mode-oriented VHE" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 495, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5e82218d4718b6c092c53ccbeb7bc9ef1daf9006999695476e3a37fe17c970b.jpg", + "image_caption": [ + "(b) Task-oriented VHE", + "Figure 4. PhD based VHE analytics. Models required paid services, shown in gray markers, are tested on random-2k.", + "Figure 6. MLLM say-yes rate vs. PhD Index, with Spearman correlation of -0.92. Proprietary models tested on random-2k." + ], + "image_footnote": [], + "bbox": [ + 96, + 371, + 495, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "respectively, we calculate the Spearman's correlation between the two ranks. A strong negative correlation exists, see Fig. 6. The result suggests that addressing VH requires balancing output tendencies, with a particular focus on enhancing an MLLM's ability to say no.", + "bbox": [ + 89, + 705, + 482, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Summary and Conclusions", + "text_level": 1, + "bbox": [ + 89, + 797, + 341, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced PhD, a large-scale benchmark developed with a close link to the three causes of visual hallucination, i.e. visual ambiguity, inconsistency in multi-modal input and CCS content. We propose a ChatGPT-assisted semi-automated pipeline to construct the new dataset with", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/acff34f0eb8e809b5a38ddef62c4204a3e20e42ef2c2dad88d16bb2c45770776.jpg", + "image_caption": [ + "Figure 5. Impact of LLM size (7B vs 13B) on LLaVA-1.5, LLaVA-1.6 and InstructBLIP. MLLMs using a 13B LLM tend to be better than their counterparts using a 7B LLM on PhD-sec and PhD-icc, yet worse on PhD-base and PhD-ccs." + ], + "image_footnote": [], + "bbox": [ + 555, + 90, + 867, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/307c17e5c327a1a5e1e9c5c60010a436ef845189864c64a1716ddf7a45eb0c54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 297, + 864, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "well affordable manual cost. The pipeline allows us to construct diverse and visually challenging h items in an image-specific and task-specific manner. Extensive experiments with 15 open-source MLLMs, 3 proprietary MLLMs, and 2 hallucination mitigation methods support our conclusions as follows. Larger visual encoders and higher input resolutions are helpful to reduce hallucination caused by visual ambiguity. The evaluation on PhD-sec and PhD-icc suggests the current models favor the textual part in the multimodal input. Resolving the conflicts between the CCS content and the model's internal knowledge demands a more comprehensive approach that is beyond isolated improvements on the visual or language components. Among the open-source MLLMs, LLaVA-OneVision is the best, followed by Molmo and InternVL-1.5. While existing benchmarks could lead to an overly optimistic expectation that the open-source models are catching up with GPT-4o, a substantial performance gap remains, as revealed by PhD.", + "bbox": [ + 511, + 564, + 906, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was supported by NSFC (62172420), Tencent Marketing Solution Rhino-Bird Focused Research Program and the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001).", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19864", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anthropic. Introducing Claude 3.5 Sonnet, 2024. https://www.anthropic.com/news/claude-3-5-sonnet.6", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 6", + "[3] Nitzan Bitton-Guetta, Yonatan Bitton, Jack Hessel, Ludwig Schmidt, Yuval Elovici, Gabriel Stanovsky, and Roy Schwartz. Breaking common sense: WHOOPS! a vision-and-language benchmark of synthetic and compositional images. In ICCV, 2023. 2, 3", + "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. MiniGPT-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 6", + "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. 6", + "[6] Chenhang Cui, Yiyang Zhou, Xinyu Yang, Shirley Wu, Linjun Zhang, James Zou, and Huaxiu Yao. Holistic analysis of hallucination in GPT-4v(ison): Bias and interference challenges. arXiv preprint arXiv:2311.03287, 2023. 2, 3", + "[7] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 6", + "[8] Matt Deitke, Christopher Clark, Sangho Lee, et al. Molmo and PixMo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 6", + "[9] Doubao Team. Doubao product, 2024. https://www.volcengine.com/product/doubao.5", + "[10] Google. Google Gemini 1.5 Pro, 2024. https://deepmind.google/technologies/gemini/pro/.6", + "[11] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. HallusionBench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In CVPR, 2024. 2, 3", + "[12] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. In AAAI, 2024. 2", + "[13] Hongyu Hu, Jiyuan Zhang, Minyi Zhao, and Zhenbang Sun. CIEM: Contrastive instruction evaluation method for better instruction tuning. In NeurIPS Workshop on ITIF, 2023. 2, 3", + "[14] Yutao Hu, Tianbin Li, Quanfeng Lu, Wenqi Shao, Junjun He, Yu Qiao, and Ping Luo. OmniMedVQA: A new large-scale" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "comprehensive evaluation benchmark for medical LVLM. In CVPR, 2024. 2", + "[15] Liqiang Jing, Ruosen Li, Yunmo Chen, Mengzhao Jia, and Xinya Du. FAITHSCORE: Evaluating hallucinations in large vision-language models. In Findings of the Association for Computational Linguistics: EMNLP, 2024. 2", + "[16] Kushal Kafle and Christopher Kanan. An analysis of visual question answering algorithms. In ICCV, 2017. 2, 3", + "[17] Kang-il Lee, Minbeom Kim, Seunghyun Yoon, Minsung Kim, Dongryeol Lee, Hyukhun Koh, and Kyomin Jung. VLind-Bench: Measuring language priors in large vision-language models. In NAACL Findings, 2025. 2, 3", + "[18] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In CVPR, 2024. 6", + "[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 4, 6", + "[20] Junyi Li, Jie Chen, Ruiyang Ren, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. The dawn after the dark: An empirical study on factuality hallucination in large language models. In ACL, 2024. 2", + "[21] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, 2023. 1, 2, 3, 6, 7", + "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 3", + "[23] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In ICLR, 2024. 2, 7", + "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 6", + "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 1, 6", + "[26] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. https://llava-v1.github.io/blog/2024-01-30-llava next/.6", + "[27] Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (NOPE) to measure object hallucination in vision-language models. In ALVR, 2024. 2, 3", + "[28] OpenAI. Dall-E 3 system, 2023. https://openai.com/index/dall-e-3/.5", + "[29] OpenAI. Hello GPT-4o, 2024. https://openai.com/index/hello-gpt-4o/.6", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + ], + "bbox": [ + 517, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19865", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. IllusionVQA: A challenging optical illusion dataset for vision language models. In COLM, 2024. 2", + "[32] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. In NeurIPS, 2024. 6", + "[33] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, 2024. 2", + "[34] Junyang Wang, Yuhang Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Ming Yan, Ji Zhang, and Jitao Sang. An LLM-free multi-dimensional benchmark for mllms hallucination evaluation. arXiv preprint arXiv:2311.07397, 2023. 1, 2, 3, 6", + "[35] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 2", + "[36] Qijie Wei, Kaiheng Qian, and Xirong Li. FunBench: Benchmarking fundus reading skills of MLLMs. arXiv preprint arXiv:2503.00901, 2025. 3", + "[37] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 2", + "[38] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren" + ], + "bbox": [ + 89, + 92, + 482, + 515 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhou. mPLUG-Owl2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 6", + "[39] Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. Woodpecker: Hallucination correction for multimodal large language models. Science China Information Sciences, 67(12):220105, 2024. 6", + "[40] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. MMMU: A massive multi-discipline multimodal understanding and reasoning benchmark for expert AGI. In CVPR, 2024. 2", + "[41] Le Zhang, Rabiul Awal, and Aishwarya Agrawal. Contrasting intra-modal and ranking cross-modal hard negatives to enhance visio-linguistic compositional understanding. In CVPR, 2024. 2", + "[42] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the AI ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2", + "[43] Kankan Zhou, Eason Lai, Wei Bin Au Yeong, Kyriakos Mouratidis, and Jing Jiang. ROME: Evaluating pre-trained vision-language models on reasoning beyond visual common sense. In Findings of the Association for Computational Linguistics: EMNLP, 2023. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 515 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19866", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/0cae22af-0103-455d-a35f-1a510beabf2e_model.json b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/0cae22af-0103-455d-a35f-1a510beabf2e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a68e96d6dc0c28f5875a1ad47d96fa5a2f9fdbda --- /dev/null +++ b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/0cae22af-0103-455d-a35f-1a510beabf2e_model.json @@ -0,0 +1,2673 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.131, + 0.847, + 0.154 + ], + "angle": 0, + "content": "PhD: A ChatGPT-Prompted Visual hallucination Evaluation Dataset" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.18, + 0.859, + 0.218 + ], + "angle": 0, + "content": "Jiazhen Liu\\(^{1*}\\), Yuhan Fu\\(^{1,2}\\), Ruobing Xie\\(^{2}\\), Runquan Xie\\(^{2}\\), Xingwu Sun\\(^{2}\\), Fengzong Lian\\(^{2}\\), Zhanhui Kang\\(^{2}\\), and Xirong Li\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.234, + 0.699, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)Renmin University of China \n\\(^{2}\\)Machine Learning Platform Department, Tencent" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.273, + 0.651, + 0.287 + ], + "angle": 0, + "content": "https://github.com/jiazhen-code/PhD" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.485, + 0.763 + ], + "angle": 0, + "content": "Multimodal Large Language Models (MLLMs) hallucinate, resulting in an emerging topic of visual hallucination evaluation (VHE). This paper contributes a ChatGPT-Prompted visual hallucination evaluation Dataset (PhD) for objective VHE at a large scale. The essence of VHE is to ask an MLLM questions about specific images to assess its susceptibility to hallucination. Depending on what to ask (objects, attributes, sentiment, etc.) and how the questions are asked, we structure PhD along two dimensions, i.e. task and mode. Five visual recognition tasks, ranging from low-level (object / attribute recognition) to middle-level (sentiment / position recognition and counting), are considered. Besides a normal visual QA mode, which we term PhD-base, PhD also asks questions with spcious context (PhD-sec) or with incorrect context (PhD-icc), or with AI-generated counter common sense images (PhD-ccs). We construct PhD by a ChatGPT-assisted semi-automated pipeline, encompassing four pivotal modules: task-specific hallucinatory item (hitem) selection, hitem-embedded question generation, spcious / incorrect context generation, and counter-common-sense (CCS) image generation. With over 14k daily images, 750 CCS images and 102k VQA triplets in total, PhD reveals considerable variability in MLLMs' performance across various modes and tasks, offering valuable insights into the nature of hallucination. As such, PhD stands as a potent tool not only for VHE but may also play a significant role in the refinement of MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.792, + 0.222, + 0.808 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.484, + 0.865 + ], + "angle": 0, + "content": "Using a specific large language model (LLM) as its kernel, a multi-modal LLM (MLLM), exemplified by LLaVA [25], Qwen-VL [2] and MiniGPT-v2 [4] can now tackle a wide" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.322, + 0.591, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.621, + 0.324, + 0.848, + 0.339 + ], + "angle": 0, + "content": "Are there five people in the image?" + }, + { + "type": "text", + "bbox": [ + 0.612, + 0.345, + 0.832, + 0.358 + ], + "angle": 0, + "content": "Yes, there are five people in the image." + }, + { + "type": "text", + "bbox": [ + 0.613, + 0.359, + 0.855, + 0.371 + ], + "angle": 0, + "content": "They are standing near a bunch of bananas." + }, + { + "type": "image", + "bbox": [ + 0.875, + 0.351, + 0.901, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.375, + 0.896, + 0.388 + ], + "angle": 0, + "content": "(a) Hallucination cause I: Visual ambiguity (MLLM: LLaVA-1.5) [25]" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.397, + 0.584, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.586, + 0.398, + 0.684, + 0.407 + ], + "angle": 0, + "content": "After a long playtime with" + }, + { + "type": "text", + "bbox": [ + 0.586, + 0.407, + 0.689, + 0.414 + ], + "angle": 0, + "content": "its toy ball, the kitten, feeling" + }, + { + "type": "text", + "bbox": [ + 0.587, + 0.414, + 0.689, + 0.421 + ], + "angle": 0, + "content": "thoroughly exhausted from" + }, + { + "type": "text", + "bbox": [ + 0.587, + 0.421, + 0.689, + 0.43 + ], + "angle": 0, + "content": "thoroughly exhausted from all the fun, has finally found" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.432, + 0.684, + 0.456 + ], + "angle": 0, + "content": "its resting place. It has curled up on the table, seeking comfort and rest after its energetic adventures." + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.4, + 0.712, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.727, + 0.402, + 0.883, + 0.415 + ], + "angle": 0, + "content": "Is there a toy in the image?" + }, + { + "type": "text", + "bbox": [ + 0.707, + 0.425, + 0.85, + 0.439 + ], + "angle": 0, + "content": "Yes, the kitten is playing" + }, + { + "type": "text", + "bbox": [ + 0.708, + 0.439, + 0.794, + 0.45 + ], + "angle": 0, + "content": "with a toy ball." + }, + { + "type": "image", + "bbox": [ + 0.875, + 0.434, + 0.902, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.459, + 0.874, + 0.472 + ], + "angle": 0, + "content": "(b) Hallucination cause II: Inconsistency in multi-modal input" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.48, + 0.592, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.486, + 0.621, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.483, + 0.79, + 0.507 + ], + "angle": 0, + "content": "Is the mouse in the image smaller than the cat?" + }, + { + "type": "text", + "bbox": [ + 0.645, + 0.515, + 0.857, + 0.528 + ], + "angle": 0, + "content": "Yes, the mouse is smaller than the cat." + }, + { + "type": "image", + "bbox": [ + 0.874, + 0.514, + 0.901, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.554, + 0.535, + 0.867, + 0.547 + ], + "angle": 0, + "content": "(c) Hallucination cause III: Counter-common-sense content" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.557, + 0.897, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.734, + 0.902, + 0.759 + ], + "angle": 0, + "content": "(d) Performance curves of the LLaVA series on two public VHE datasets (POPE [21]) and AMBER [34]) and the proposed PhD dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.907, + 0.867 + ], + "angle": 0, + "content": "Figure 1. Illustrations of three major causes of an MLLM's visual hallucination and its evaluation. This paper contributes PhD, a binary VQA-based VHE benchmark, much larger and more challenging than its predecessors. In particular, it has four evaluation modes that explicitly measure an MLLM's performance w.r.t. the three causes, i.e. PhD-base for cause I, PhD-sec and PhD-icc for cause II and PhD-ccs for cause III." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.368, + 0.888 + ], + "angle": 0, + "content": "*J. Liu is currently pursuing his PhD at HKUST." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.346, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author (xirong@ruc.edu.cn)" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.368, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19857" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "range of computer vision tasks in a unified visual-question-answering (VQA) manner. As LLMs are known to hallucinate [20, 37, 42], it is not surprising that MLLMs have visual hallucination, generating fabricated interpretation of the given visual content, see Fig. 1. Considering the rapidly growing use of MLLMs in varied scenarios, a comprehensive visual hallucination evaluation (VHE) is crucial. This paper develops a new dataset for VHE." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.213, + 0.486, + 0.501 + ], + "angle": 0, + "content": "VHE essentially involves posing a number of visual questions to an MLLM [11, 21]. A question of this kind shall include a hallucinatory item (hitem), in the form of a specific word or phrase, that induces the MLLM to generate a response discordant with the provided visual content. As the model typically has a strong visual recognition ability, how to identify an appropriate hitem and accordingly generate a proper question is nontrivial. Both the hitem and the question depend on the visual recognition task being considered. As shown in Tab. 1, we target at objective VHE in the context of low-level (object / attribute) to middle-level (sentiment / position / counting) visual recognition tasks. Such a target is chosen due to the following considerations. MLLMs generally work well for these tasks, so their erroneous responses can be largely attributed to their hallucinations instead of their incapability, e.g. asking a generic MLLM to read pathology images [14]. Meanwhile, a binary VQA based objective evaluation is more budget friendly and thus more suited for VHE at a large scale." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.514, + 0.485, + 0.638 + ], + "angle": 0, + "content": "
Vision tasksObjective evaluationSubjective evaluation
Low-/middle-level visual recognition+ POPE, EMNLP'23 [21]+ FAITHSCORE, EMNLP'24 [15]
+ AMBER, arXiv'23 [34]+ HaELM, arXiv'23 [35]
+ CIEM, ITIF'23 [13]+ M-HalDetect, AAAI'24 [12]
+ NOPE, ALVIR'24 [27]+ GAVIE, ICLR'24 [23]
+ ROME, EMNLP'23 [43]
+ PhD (this paper)
High-level visual reasoning+ MMMU, CVPR'24 [40]+ HallusionBench, CVPR'24 [11]
+ VLind-Bench, NAACL'25 [17]+ Bingo, arXiv'23 [6]
+ IllusionVQA, COLM'24 [31]
+ WHOOPS!, ICCV'23[3]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.647, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Table 1. Taxonomy of VHE benchmarks. Our PhD benchmark performs an objective evaluation of MLLMs' hallucinations when they address visual recognition tasks ranging from low-level, i.e. object / attribute recognition to middle-level, i.e. sentiment / positional recognition and counting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Pioneered by POPE [21], good attempts exist in objective VHE [13, 27, 34, 43]. In these valuable datasets, hitem selection is largely untouched. POPE and ROME [43] select their hitems fully based on label co-occurrence in training data, AMBER [34] relies on manual annotation, whilst hitem selection is not considered in NOPE [27] and CIEM [13], see Tab. 2. Hence, there lacks an explicit link between hitem selection (and subsequent VQA triplets construction) and major causes of an MLLM's visual hallucination. As models rapidly evolve, the performance on these datasets quickly reaches saturation, see Fig. 1d." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.394 + ], + "angle": 0, + "content": "Analyzing an MLLM's typical dataflow of answering a visual question, we see three major causes of visual hallucination: I) visual ambiguity, II) inconsistency in multi-modal input and III) counter-common-sense (CCS) content, see Fig. 1. Firstly, the MLLM extracts tokenized visual features from a given image using a ViT based encoder. Recent studies [33, 41] show that the features tend to be high level, lacking sufficient details for fine-grained tasks such as counting. Secondly, the visual features, after vision-to-language adaptation, are mixed with the features of the associated textual prompt to form a multi-modal input to the LLM kernel. The LLM, pre-trained extensively on textual data, inevitably favors the textual part of the multi-modal input. Hence, when there is inconsistency between the visual and textual information, the former is more likely to be overruled. Lastly, at the decoding stage, the LLM might heavily rely on its internal (world) knowledge, ignoring the visual content especially when the content (showing a mouse much larger than a cat) contradicts the common sense. Our new dataset is developed with a close link to the three causes." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.405, + 0.907, + 0.498 + ], + "angle": 0, + "content": "
DatasetDaily imagesCCS imagesHItemsContextsVQA tripletsTasks
POPE500✘80✘3,000Obj.
NOPE*unknown✘unknown✘32,701Obj.
CIEM*4,929✘unknown✘72,941Obj. / Attr. / Pos.
AMBER1,004✘687✘14,216Obj. / Attr./Pos. / Count.
ROME✘1,563118✘1,563Attr./Pos.
PhD14,6487501,45233,688102,564Obj / Attr. / Pos. / Sent. / Count.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.51, + 0.904, + 0.524 + ], + "angle": 0, + "content": "Table 2. PhD versus its predecessors. * indicates private dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.856 + ], + "angle": 0, + "content": "The new dataset is constructed by adapting TDIUC, a popular multi-task VQA dataset [16], with a ChatGPT-assisted semi-automated pipeline, see Fig. 2. In particular, by prompting ChatGPT, we select diverse and visually challenging h items in an image-specific and task-specific manner, with minimal human involvement primarily spent on verifying ChatGPT-generated results. The selected h items are then automatically embedded into visual questions, specious context, and incorrect context, all generated by instructing ChatGPT. Moreover, we expand the daily image set with counter common sense (CCS) images, obtained by prompting AIGC tools with ChatGPT-generated CCS descriptions, e.g. \"trees growing underwater\" and \"a car with square-shaped wheels\". The dataset is dubbed as PhD (ChatGPT Prompted visual hallucination evaluation Dataset). Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base (questions about daily images w/o context), PhD-sec (PhD-base plus specious context), PhD-icc (PhD-base plus incorrect context), and PhD-ccs (questions about CCS images)." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.857, + 0.868, + 0.871 + ], + "angle": 0, + "content": "To sum up, our major contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "- We introduce PhD, a dataset with four evaluation modes across five visual recognition tasks, developed with a close" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19858" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "link to the three major causes of MLLM visual hallucination. Information on hallucinatory items (hItems) is provided per sample, enabling in-depth analytics to uncover the causes in more detail." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.482, + 0.227 + ], + "angle": 0, + "content": "- We offer a ChatGPT-assisted semi-automatic pipeline for dataset construction, with minimal human involvement, primarily focused on verifying the generated results. With 14,648 daily images, 750 CCS images and 102,564 VQA triplets in total, PhD is the largest of its kind." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.228, + 0.482, + 0.348 + ], + "angle": 0, + "content": "- We conduct an extensive evaluation with 15 open-source MLLMs, 3 proprietary MLLMs, and 2 hallucination mitigation methods, showing the viability of PhD for VHE in varied manners including overall, mode-oriented, task-oriented, and model-wise zoom-in. The evaluation not only reveals inter-model performance divergence, but may also help developers of a specific MLLM to prioritize their efforts in refining the model." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.152, + 0.482, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.362, + 0.233, + 0.377 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.388, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Due to the increasing importance of VHE, new benchmarks are being actively developed. Depending on what vision tasks they focus on and how their evaluation is executed, we categorize existing achievements along two dimensions, i.e. task and evaluation, see Tab. 1. Concerning the task dimension, low- and middle-level visual recognition tasks, ranging from object / attribute recognition to sentiment / positional recognition and counting, assess an MLLM's basic visual skills. High-level visual reasoning is more domain-knowledge intensive, typically covering image-based math question solving, geography information understanding, meme interpretation, historical or folkloric contexts, etc. [3, 6, 11, 17]. As for the evaluation dimension, objective evaluation refers to objectively comparing the model's output with ground truth, mostly in the form of Yes/No answers [13, 21, 34, 43]. By contrast, subjective evaluation requires humans or LLMs to assess the model's output [3, 6, 11]. The proposed PhD, focusing on low/middle-level visual recognition and objective evaluation, belongs to the second quadrant of the taxonomy. In what follows, we briefly review peer benchmarks, i.e. POPE [21], ROME [43], NOPE [27], CIEM [13], and AMBER [34], in this quadrant and clarify our novelty accordingly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "POPE is probably the first dataset to evaluate object hallucination [21]. Given a specific MS-COCO image with object labels, POPE selects an adversarial object frequently co-occurring with the current labels. A binary question is then formed by filling out a predefined template with the selected object. Such co-occurrence based hitem selection is not image-specific by definition. Trivial hitems might thus be picked up, e.g. \"car\" selected for an indoor image labeled with \"person\", as the two objects often co-occur. For CCS image generation, ROME forms CCS descriptions by choosing attribute values having the lowest co-occurrence" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "with a given object according to the Visual Genome dataset [43]. As low occurrence is not necessarily CCS, some of ROME images are indeed normal. NOPE [27] and CIEM [13] simply bypass hitem selection by asking an LLM to generate questions conditioned on the image captions (also from MS-COCO) and pre-specified answers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.258 + ], + "angle": 0, + "content": "To select h items in an image-specific manner, AMBER resorts to fully manual annotation [34]. Manual labeling is costly, while an annotator's personal vocabulary is relatively limited. All this makes it difficult to scale up w.r.t. the amount of test images and the number of distinct h items." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.905, + 0.334 + ], + "angle": 0, + "content": "In comparison, PhD, constructed by a ChatGPT-assisted semi-automatic pipeline (Sec. 3), is much larger (Tab. 2) and more challenging (Fig. 1d). With its unique mode-task structure, the new dataset enables a novel, structured and zoom-in understanding of inter-model difference." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.349, + 0.724, + 0.366 + ], + "angle": 0, + "content": "3. Our Roadmap to PhD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.905, + 0.601 + ], + "angle": 0, + "content": "As MLLMs perform visual recognition in a VQA manner, a VQA sample for VHE naturally depends on the visual recognition task in consideration. We depart from TDIUC [16], a large-scale VQA dataset w.r.t. five tasks including object / attribute / sentiment / positional recognition and counting. Note that the images in TDIUC are sourced from MS-COCO [22], which plausibly have been seen by MLLMs in their development stage. As such, our adoption of TDIUC makes on-purpose data leakage: an MLLM's erroneous response w.r.t. a seen image can now be more safely attributed to its hallucination other than its incapability in visual recognition, say asking LLaVA to recognize glaucoma from color fundus photographs [36]. We construct PhD by adapting the TDIUC annotations with a ChatGPT1-assisted semi-automated pipeline, see Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.905, + 0.814 + ], + "angle": 0, + "content": "In order to compose a proper question that effectively makes an MLLM hallucinate about a given image, a hitem has to be first identified in an image-specific and task-specific manner. Then, the hitem has to be smoothly embedded in the form of a specific word or phrase into the question. We describe task-specific hitem selection in Sec. 3.1, followed by hitem-embedded question generation in Sec. 3.2. Furthermore, in order to simulate inconsistency in the multi-modal input, we prepend specious or incorrect context to the question, the generation of which is detailed in Sec. 3.3. Lastly, in order to explicitly create conflicts between the visual input and the internal (world) knowledge of the MLLM, we expand our image collection with autogenerated counter-common-sense (CCS) images (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.776, + 0.839 + ], + "angle": 0, + "content": "3.1. Task-specific Hitem Selection" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.846, + 0.905, + 0.876 + ], + "angle": 0, + "content": "Without loss of generality, we describe how a hitem is selected for color attribute recognition. Let us consider the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.783, + 0.899 + ], + "angle": 0, + "content": "1We use GPT-4o mini, released on 2024-05-13." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19859" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.093, + 0.24, + 0.102 + ], + "angle": 0, + "content": "(1) Task-specific hitem selection" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.104, + 0.66, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.169, + 0.278, + 0.179 + ], + "angle": 0, + "content": "(2) Hitem-embedded question generation" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.183, + 0.659, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.256, + 0.588, + 0.269 + ], + "angle": 0, + "content": "(a) Hitem-embedded question / context generation for MS-COCO daily images" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.105, + 0.9, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.256, + 0.886, + 0.269 + ], + "angle": 0, + "content": "(b) QAs for AI-generated CCS images" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.284, + 0.223, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.356, + 0.181, + 0.364 + ], + "angle": 0, + "content": "Task: Object recognition" + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.365, + 0.216, + 0.372 + ], + "angle": 0, + "content": "Are there headphones in the image?" + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.373, + 0.204, + 0.38 + ], + "angle": 0, + "content": "Are there books in the image?" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.286, + 0.376, + 0.332 + ], + "angle": 0, + "content": "In recent years, this musician has often immersed himself in his own musical world, playing the guitar with headphones on as if cut off from the outside world. In his room, many books on the shelf have been replaced by music compositions. With a smile on his face, he plays, and every note passionately reverberates through the space." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.333, + 0.376, + 0.38 + ], + "angle": 0, + "content": "In the photo, a man confidently plays his guitar with headphones on, fully immersed in his music. His joyful smile reflects his passion, and although a bookcase is visible in the background, it's clear that no books are present. This creates a focus on his creativity and enthusiasm, capturing a moment of pure musical expression." + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.284, + 0.658, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.388, + 0.222, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.446, + 0.192, + 0.453 + ], + "angle": 0, + "content": "Task: Sentiment recognition" + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.453, + 0.221, + 0.46 + ], + "angle": 0, + "content": "Is the girl feeling lonely in the image?" + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.46, + 0.219, + 0.468 + ], + "angle": 0, + "content": "Is the girl feeling happy in the image?" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.389, + 0.376, + 0.427 + ], + "angle": 0, + "content": "The girl is often surrounded by her beloved teddy bears, yet she regularly feels a sense of loneliness despite their company. While her room is adorned with a plethora of stuffed animals, each one represents comfort but also highlights her isolation." + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.428, + 0.375, + 0.469 + ], + "angle": 0, + "content": "The girl in the image is feeling lonely, despite being surrounded by a large group of teddy bears. She lies on the bed near many stuffed animals, including dozens of brown teddy bears, and her contrast between her solitary emotions and the cheerful presence of her furry companions." + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.388, + 0.504, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.433, + 0.435, + 0.439 + ], + "angle": 0, + "content": "Task: Positional" + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.44, + 0.478, + 0.446 + ], + "angle": 0, + "content": "Is there a table behind" + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.447, + 0.493, + 0.454 + ], + "angle": 0, + "content": "the looking dog in the image?" + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.455, + 0.489, + 0.462 + ], + "angle": 0, + "content": "Is there a fence behind" + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.462, + 0.489, + 0.469 + ], + "angle": 0, + "content": "the looking dog in the image?" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.389, + 0.658, + 0.427 + ], + "angle": 0, + "content": "A large brown dog has comfortably settled onto a stack of three plastic chairs, creating an amusing scene as it relaxes in the sunny outdoor space. Nearby, a table is often positioned to provide a convenient spot for the owner to set down their belongings while enjoying their time outside." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.428, + 0.658, + 0.469 + ], + "angle": 0, + "content": "In a charming scene, a dog has made itself comfortable on a stack of plastic chairs, effortlessly showcasing its playful nature. Behind the relaxed dog captured in the photo, a table can be filled with a large box of cozy atmosphere. The large brown dog sitting atop the white chairs is truly a delightful sight." + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.305, + 0.781, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.369, + 0.78, + 0.376 + ], + "angle": 0, + "content": "Is the ball being kicked a soccer ball?" + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.376, + 0.78, + 0.383 + ], + "angle": 0, + "content": "Is the ball being kicked a basketball?" + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.306, + 0.898, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.369, + 0.893, + 0.376 + ], + "angle": 0, + "content": "Is the mouse smaller than the cat?" + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.376, + 0.893, + 0.383 + ], + "angle": 0, + "content": "Is the cat smaller than the mouse?" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.391, + 0.781, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.443, + 0.758, + 0.449 + ], + "angle": 0, + "content": "Is the highest number" + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.45, + 0.763, + 0.456 + ], + "angle": 0, + "content": "of points on the dice six?" + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.456, + 0.758, + 0.462 + ], + "angle": 0, + "content": "Is the highest number" + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.462, + 0.766, + 0.469 + ], + "angle": 0, + "content": "points on the dice seven?" + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.391, + 0.898, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.455, + 0.891, + 0.462 + ], + "angle": 0, + "content": "Is the shape of the orange round?" + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.462, + 0.891, + 0.469 + ], + "angle": 0, + "content": "Is the shape of the orange square?" + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.472, + 0.894, + 0.485 + ], + "angle": 0, + "content": "(d) Showcase: CCS images and questions" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.497, + 0.907, + 0.581 + ], + "angle": 0, + "content": "Figure 2. Proposed semi-automatic pipeline for PhD construction. We use ChaptGPT (GPT-4o mini) to generate hitem-embedded questions / contexts for daily images, and Doubao and DALL-E3 for generating CCS images. Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base, i.e. questions about daily images w/o context, PhD-sec, i.e. PhD-base plus specious context, PhD-icc, i.e. PhD-base plus incorrect context, and PhD-ccs, i.e. questions about CCS images. By adapting TDIUC annotations, PhD supports binary VQA w.r.t. five visual recognition tasks including object / attribute / sentiment / positional recognition and counting. With 20 mode-task combinations in total, PhD enables a comprehensive VHE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.667 + ], + "angle": 0, + "content": "image in Fig. 2a, showing a black motorcycle followed by a red bus. As no red motorcycle is present while the red color is prominent near the black motorcycle, the word red will be a good choice of hitem to challenge an MLLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Vocabulary Construction per task. Started with a handful of manually specified colors such as red, green, and blue, we ask ChatGPT to expand the color vocabulary with instructions like \"Please expand the input vocabulary as much as possible by adding common items found in daily life. Avoid any duplication\", getting a set of 35 different colors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Subject-Attribute Extraction. The image is associated with a TDIUC question-answer pair as \"what color is the motorcycle\" and \"black\". We use ChatGPT (with simple instructions) to extract with ease the subject (i.e. motorcycle) and its attribute (i.e. black) from the pair." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.839, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Candidate Hitem Generation. We obtain candidate h items by excluding the ground-truth (GT) answer (and its synonyms if applicable) from the vocabulary." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Visual-based Hitem Ranking. Intuitively, a hitem shall" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.905, + 0.759 + ], + "angle": 0, + "content": "be visually plausible in the given image. So for each candidate hitem, we compute its similarity to the image using a pre-trained CLIP [30]. In particular, the cosine similarity between the CLIP embeddings of hitem + subject (e.g. green motorcycle) and the image is adopted. Ranking the candidates by the CLIP similarity lets us to select the one visually closest to the image. It is worth noting that for emerging MLLMs equipped with stronger vision encoders [19], our pipeline is likely to produce even more effective h items by replacing CLIP with these advanced counterparts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.905, + 0.853 + ], + "angle": 0, + "content": "Manual Inspection. While the above process is generally stable to produce satisfying results, manual inspection is performed to ensure the correctness of hitem selection. Note that due to errors in the original TDIUC annotations, occasionally the true label might be \"incorrectly\" selected. In such a case, we simply discard the VQA sample." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.9 + ], + "angle": 0, + "content": "With lightweight task-specific adaptation, the above hitem selection also works for other tasks. Overall, the joint use of ChatGPT and CLIP allows us to select 1,452 hitem" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19860" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "that are more diverse and challenging than their counterparts in previous datasets, see Tab. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.135, + 0.426, + 0.151 + ], + "angle": 0, + "content": "3.2. Hitem-embedded Question Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.159, + 0.485, + 0.266 + ], + "angle": 0, + "content": "For a given subject (e.g. motorcycle) and a chosen hitem (e.g. red), generating a hitem-embedded question is trivial for ChatGPT. In particular, a No question is formed as \"Is the motorcycle in the image red?\" Meanwhile, a Yes question is simultaneously generated using the GT as \"Is the motorcycle in the image black?\". This ensures perfect Yes/No balance among the generated questions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.278, + 0.44, + 0.294 + ], + "angle": 0, + "content": "3.3. Specious (Incorrect) Context Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.484, + 0.481 + ], + "angle": 0, + "content": "When used as a document parser, an MLLM reading a specific image is often provided with the image's surrounding text. Inconsistency between the image and the text is not uncommon. A news article containing a general claim of \"red motorcycles frequently zip through the streets\" does not necessarily have each of its illustrated pictures match with the claim. To simulate such a scenario, for a given image we generate specious text as specious context and text contradicting the image as incorrect context, respectively. Next, we describe the generation of specious contexts, as their incorrect counterparts can be generated in a similar but more simplified manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.635 + ], + "angle": 0, + "content": "Specious Text Generation. Using the previously generated hitem-embedded question and the original MS-COCO captions as input, we instruct ChatGPT to generate specious text for a given image. By \"specious\", we mean the text is specious or noisy, rather than directly contradicting the image content. As such, our instruction reads partially as \"Please generate the for the given question. It should be one sentence. The should answer the question, but it may not reflect the actual current status, thus making it specious.\"" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.483, + 0.697 + ], + "angle": 0, + "content": "Text Composition. ChatGPT is used to seamlessly merge the specious text with ground-truth captions, forming a longer context in which only a small portion (orange text in Fig. 2c) is mildly inconsistent with the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.729 + ], + "angle": 0, + "content": "Manual Inspection. We perform a spot check. If the context quality is low, we simply discard the entire sample." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.741, + 0.308, + 0.758 + ], + "angle": 0, + "content": "3.4. CCS Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.81 + ], + "angle": 0, + "content": "We generate CCS images by first generating CCS descriptions and then employing Text2Image tools to convert the descriptions to CCS images, see Fig. 2b." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "CCS Description Generation. A number of manually written task-specific samples, see Tab. 3, are used as in-context learning samples for ChatGPT to generate more descriptions. The descriptions have to be visually expressible, so bad cases like \"the more you eat, the thinner you get\" are filtered out manually. For each CCS text, its common-sense" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "(CS) counterpart is simultaneously generated by ChatGPT, by providing the learning samples in pair." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Text2Image. The generated CCS descriptions are used as prompts for AIGC tools (Doubao [9] and DALL-E3 [28]) to generate the corresponding CCS images. The quality of the generated images depends on various factors, making occasional failures inevitable. When this occurs, we attempt to refine the prompts or apply region-based inpainting. The sample will be discarded if the above attempts fail." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.229, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Question Generation. Per CCS description (e.g. A car with square wheels), we utilize ChatGPT to generate a Yes question (e.g. Does the car have square wheels?). Again, for balancing Yes/No questions, we generate a No question (e.g. Does the car have round wheels?) based on the CS description." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.332, + 0.907, + 0.681 + ], + "angle": 0, + "content": "
CCS descriptionYes questionCS descriptionNo question
Task: Object recognition
Manually written:
Ice blocks in volcanic lavaAre there ice blocks in volcanic lava?Fire in volcanic lavaIs there fire in volcanic lava?
Grass in a tiger's mouthIs there grass in a tiger's mouth?Meat in the tiger's mouthIs there meat in the tiger's mouth?
ChatGPT generated:
Trees growing underwaterAre there trees growing underwater?Coral growing underwaterIs there coral growing underwater?
Books in a swimming poolAre there books in a swimming pool?Water in a swimming poolIs there water in a swimming pool?
Birds flying underwaterAre there birds flying underwater?Birds flying in the skyAre there birds flying in the sky?
Ice cream in a volcanoIs there ice cream in a volcano?Lava in a volcanoIs there lava in a volcano?
Computers in a forestAre there computers in a forest?Animals in a forestAre there animals in a forest?
Task: Attribute recognition
Manually written:
A car with square wheelsDoes the car have square wheels?A car with round wheelsDoes the car have round wheels?
Blue apples on the treeAre the apples on the tree blue?Red apples on the tree.Are the apples on the tree red?
ChatGPT generated:
A green skyIs the sky green?A blue skyIs the sky blue?
A bicycle with square wheelsDoes the bicycle have square wheels?A bicycle with round wheelsDoes the bicycle have square wheels?
A tree made of metalIs this tree made of metal?A wooden treeIs this tree a real wood tree?
A chocolate riverIs there chocolate flowing in the stream?A water riverIs there water flowing in the stream?
A house made of candyIs the house made of candy?A house made of bricksIs the house made of bricks?
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.689, + 0.907, + 0.746 + ], + "angle": 0, + "content": "Table 3. Instances of descriptions used for generating CCS images and related questions. With manually written CCS / CS descriptions as instructions, ChatGPT is used to generate many more instances and subsequently convert them to Yes/No questions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.773, + 0.807, + 0.789 + ], + "angle": 0, + "content": "3.5. Dataset Overview and PhD Index" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "An overview of the PhD dataset is given in Tab. 4. Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base, i.e. questions about daily images w/o context, PhD-sec, i.e. PhD-base plus spcious context, PhD-icc, i.e. PhD-base plus incorrect context, and PhD-ccs, i.e. questions about CCS images. With 20 mode-task com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19861" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "binations in total, PhD supports a much more comprehensive VHE than its predecessors [21, 34]." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.134, + 0.485, + 0.208 + ], + "angle": 0, + "content": "
TasksQuestions
ObjectAttributeSentimentPositionCountingYesNo
TDIUC samples used6,2714,3242,0952,8413,387--
Unique h items7451466548610--
VQA samples in PhD-base11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-sec11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-icc11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-ccs34473478220124750750
" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.219, + 0.452, + 0.233 + ], + "angle": 0, + "content": "Table 4. Data statistics of the proposed PhD dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.25, + 0.484, + 0.327 + ], + "angle": 0, + "content": "To measure the performance of an MLLM on PhD, we compute its recall w.r.t. the Yes and No questions, respectively. We term the harmonic mean of the two recalls PhD Index. A model simply saying Yes (or No) to all questions has a PhD Index of 0, while a random-guess score is 0.5." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.339, + 0.352, + 0.357 + ], + "angle": 0, + "content": "4. Evaluating MLLMs on PhD" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.364, + 0.248, + 0.381 + ], + "angle": 0, + "content": "4.1. Common Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.484, + 0.55 + ], + "angle": 0, + "content": "Choices of MLLMs. For reproducible research, we focus on open-source MLLMs, compiling a list of 15 models that span varied sizes and architectures. see Tab. 5. We also evaluate two hallucination mitigation methods, VCD [18] and Woodpecker [39], currently supporting LLaVA-1.6-L and Qwen-VL. As additional references, we assess three proprietary MLLMs, i.e. GPT-4o [29], Claude 3.5 Sonnet [1], and Gemini 1.5 Pro [10], on a random subset of 2k samples (random-2k) subject to our budget. For the same reason we evaluate Woodpecker, which requires paid service from ChatGPT, on random-2k." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Test Protocol. For a fair comparison, per MLLM we use its designated prompt to wrap each test question. For instance, a question-specific prompt submitted to mPLUG-Owl2 will be in the form of \"USER: \\(< |image|>\\{question\\}\\) ASSISTANT:\" We provide more prompts in the supplement. In addition, to help the models better handle PhD-sec and PhD-icc, we append to the test prompt an instruction as \"In case there is an inconsistency between the context and the image content, you should follow the image.\"" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.345, + 0.729 + ], + "angle": 0, + "content": "4.2. Using PhD for Overall VHE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "An overall VHE as shown in Tab. 5 is useful for providing a big picture of which MLLM hallucinates the most (or the least). The leading open-source MLLMs are LLaVA-OneVision, followed by Molmo and InternVL-1.5. Since their vision encoders and LLMs vary, the results are insufficient to conclude which component is the most effective to mitigate hallucinations. That said, comparisons among the same model series remains meaningful. Consider the LLaVA series for instance. While one would normally expect that a larger LLM yields a better MLLM, as LLaVA-1.6-L vs LLaVA-1.6, the difference between LLaVA-1.5-L" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "and LLaVA-1.5 is marginal (0.270 vs 0.265). In order to analyze and consequently understand such an counterintuitive result, PhD enables a zoom-in analysis in mode-oriented (Sec. 4.3) and task-oriented (Sec. 4.4) styles, unavailable in the previous benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.181, + 0.905, + 0.402 + ], + "angle": 0, + "content": "
ModelViTLLMPOPEAMBERPhD
Full-set evaluation:
LLaVA-OneVision [19]SoViT-400m/14Qwen2-72B0.840.900.698
Molmo [8]-L/14Qwen2-72B0.840.850.690
InternVL-1.5 [5]InternViT-6BInternLM2-20B0.860.890.561
Qwen-VL (VCD)-bigG/14Qwen-7B0.840.870.560
Cambrian-1 [32]HybridLlama-3-8B0.880.890.547
LLaVA-1.6-L (VCD)-L/14Vicuna-13B-1.50.820.810.511
LLaVA-1.6-XL [26]-L/14Nous-Hermes-2-Yi-34B0.860.840.492
Qwen-VL [2]-bigG/14Qwen-7B0.830.840.488
LLaVA-1.6-L [26]-L/14Vicuna-13B-1.50.830.800.423
MiniGPT-v2 [4]-G/14Llama-2-7B0.830.840.390
LLaVA-1.6 [26]-L/14Vicuna-7B-1.50.830.790.373
mPlug-Ow12 [38]-L/14Llama-2-7B0.780.770.320
InstructBLIP [7]-G/14Vicuna-7B-1.10.820.820.305
InstructBLIP-L [7]-G/14Vicuna-13B-1.10.800.790.278
LLaVA-1.5-L [25]-L/14Vicuna-13B-1.50.820.730.270
LLaVA-1.5 [25]-L/14Vicuna-7B-1.50.810.750.265
LLaVA-1.1 [24]-L/14Vicuna-7B-1.10.670.330.135
Random-2k evaluation:
GPT-4o [29]--0.880.870.812
Claude 3.5 Sonnet [1]--0.850.890.746
Gemini 1.5 Pro [10]--0.860.880.691
Qwen-VL (Woodpecker)-bigG/14Qwen-7B--0.531
LLaVA-1.6-L (Woodpecker)-L/14Vicuna-13B-1.5--0.409
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.413, + 0.906, + 0.441 + ], + "angle": 0, + "content": "Table 5. Overall VHE. POPE (adversarial) and AMBER (discriminative) are used." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.596 + ], + "angle": 0, + "content": "One more advantage of PhD compared to its predecessors lies in its discrimination ability. The relatively small performance gap between GPT-4o and the top open-source models as measured by POPE and AMBER might lead to an overly optimistic interpretation that the open-source alternatives are catching up with the proprietary model. In fact, a substantial gap remains, as revealed by PhD. Fig. 3 further illustrate the qualitative results, where GPT-4o exhibits fewer hallucinations in its response." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.827, + 0.623 + ], + "angle": 0, + "content": "4.3. Using PhD for Mode-Oriented VHE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.75 + ], + "angle": 0, + "content": "Fig. 4a illustrates mode-wise model performance. MLLMs working relatively well in the PhD-base mode tend to have stronger visual input. This is achieved either with stronger visual encoders, as the cases of LLaVA-OneVision, InternVL-1.5, and Cambrian-1 using SoViT-400m/14, InternViT-6B or hybrid vision structure, or supporting higher image resolutions, see Molmo and MiniGPT4-v2 that accept multiscale or larger input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In contrast to the visual part, using a larger LLM alone does not necessarily lead to a better MLLM. As noted in Sec. 4.2, LLaVA-1.5-L with Vicuna-13B-1.5 has nearly the same PhD Index (0.270) as LLaVA-1.5 with Vicuna-7B-1.5 (0.265). We see from Fig. 4a that the larger LLM indeed improves the performance in PhD-sec \\((0.082\\rightarrow 0.099)\\), PhD-icc \\((0.011\\rightarrow 0.019)\\) and PhD-ccs \\((0.534\\rightarrow 0.542)\\), yet suffers loss in PhD-base \\((0.443\\rightarrow 0.422)\\). Similar results can be more evidently observed in the case of InstructBLIP-L vs. InstructBLIP (PhD-base Index: \\(0.535\\rightarrow 0.324\\)). Our con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19862" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.907, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.908, + 0.345 + ], + "angle": 0, + "content": "Figure 3. Qualitative results showing how an MLLM answers visual questions from PhD. The correctness of an answer is automatically determined by matching its first word, either Yes or No, with the ground truth (GT)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.483, + 0.417 + ], + "angle": 0, + "content": "jecture is that although a larger LLM better understands user instructions, its successful use within an MLLM requires more targeted training for vision-language alignment." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.417, + 0.483, + 0.627 + ], + "angle": 0, + "content": "Comparing the four modes, the performance of the open-source MLLMs on PhD-icc and PhD-sec is generally low. When provided with a multi-modal input, the models favor the textual part. Substituting a 13-B LLM for its 7-B counterpart helps tackling the inconsistency in the multi-modal input, see Fig. 5. However, a larger LLM might rely more on its internal knowledge for decoding, resulting in worse performance on PhD-base and PhD-ccs for which the image content shall carry more weights. In particular, PhD-ccs reveals a deeper and intrinsic challenge of VH: conflicts between the given image content and the model's internal knowledge. Solving the challenge demands a more comprehensive approach that goes beyond isolated improvements on the visual or language components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Among the open-source MLLMs, LLaVA-OneVision is the best, owing to its joint use of a stronger visual encoder (SoViT-400m/14), a more powerful LLM (Qwen2-72B), and better training strategies (much larger high-quality training data and multi-stage alignment). Nevertheless, LLaVA-OneVision remains inferior to GPT-4o, particularly in PhD-sec and PhD-icc, wherein inconsistency within the multi-modal input has to be properly addressed." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.396, + 0.773 + ], + "angle": 0, + "content": "4.4. Using PhD for Task-Oriented VHE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Fig. 4b presents task-oriented VHE results. In general, to what extent an MLLM hallucinates is largely correlated to the required level (low / middle / high) of a specific task. The object recognition task has the overall highest PhD Index, followed by attribute recognition, positional recognition, counting, and sentiment recognition. Due to the complexity and subtlety of emotions, e.g. tears can be associated with both happiness and sadness, even GPT-4o performs" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.371, + 0.895, + 0.387 + ], + "angle": 0, + "content": "relatively worse in the sentiment task (first row of Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.906, + 0.525 + ], + "angle": 0, + "content": "Joint mode-task analytics per model is shown in Tab. 6. LLaVA-OneVision struggles with sentiment recognition and counting, especially when faced with textual or CCS distractions, underscoring the need for improvement in these areas. Similarly, Molmo also faces these challenges, but its counting performance under CCS distractions is notably better than LLaVA-OneVision's (0.737 vs 0.563). The above zoom-in analytics will be informative for MLLM developers to prioritize their efforts on model refinement." + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.537, + 0.871, + 0.688 + ], + "angle": 0, + "content": "
TaskPhD-basePhD-secPhD-iccPhD-ccs
LLAVA-OneVision
Object0.8720.8490.8240.727
Attribute0.8480.7440.6630.767
Sentiment0.6910.5810.5040.731
Positional0.7730.7300.6540.701
Counting0.7070.6520.5000.563
Molmo
Object0.8250.8800.8470.678
Attribute0.8420.7250.5560.791
Sentiment0.5470.6020.5680.746
Positional0.6970.6910.6540.742
Counting0.7270.5800.3500.737
" + }, + { + "type": "table_caption", + "bbox": [ + 0.567, + 0.699, + 0.851, + 0.713 + ], + "angle": 0, + "content": "Table 6. Zoom-in analytics of specific models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.837, + 0.759 + ], + "angle": 0, + "content": "4.5. Analysis of MLLM Answer Tendency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "While the yes and no questions are perfectly balanced by design, we observe that the open-source MLLMs tend to answer yes, with the say-yes rate ranging from 0.462 (Molmo) to 0.811 (LLaVA-1.1) and an average value of 0.611. By contrast, the three proprietary MLLMs have a clearly lower say-yes rate. Similar observations are reported in [21, 23]. We go one step further by analyzing how the say-yes tendency is related to model performance. Ranking the models by their PhD Index and say-yes rate," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19863" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.496, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.343, + 0.361, + 0.354 + ], + "angle": 0, + "content": "(a) Mode-oriented VHE" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.372, + 0.496, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.624, + 0.359, + 0.635 + ], + "angle": 0, + "content": "(b) Task-oriented VHE" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.648, + 0.483, + 0.677 + ], + "angle": 0, + "content": "Figure 4. PhD based VHE analytics. Models required paid services, shown in gray markers, are tested on random-2k." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.483, + 0.782 + ], + "angle": 0, + "content": "respectively, we calculate the Spearman's correlation between the two ranks. A strong negative correlation exists, see Fig. 6. The result suggests that addressing VH requires balancing output tendencies, with a particular focus on enhancing an MLLM's ability to say no." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.799, + 0.342, + 0.816 + ], + "angle": 0, + "content": "5. Summary and Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We have introduced PhD, a large-scale benchmark developed with a close link to the three causes of visual hallucination, i.e. visual ambiguity, inconsistency in multi-modal input and CCS content. We propose a ChatGPT-assisted semi-automated pipeline to construct the new dataset with" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.091, + 0.868, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.22, + 0.907, + 0.275 + ], + "angle": 0, + "content": "Figure 5. Impact of LLM size (7B vs 13B) on LLaVA-1.5, LLaVA-1.6 and InstructBLIP. MLLMs using a 13B LLM tend to be better than their counterparts using a 7B LLM on PhD-sec and PhD-icc, yet worse on PhD-base and PhD-ccs." + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.298, + 0.865, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.503, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Figure 6. MLLM say-yes rate vs. PhD Index, with Spearman correlation of -0.92. Proprietary models tested on random-2k." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.838 + ], + "angle": 0, + "content": "well affordable manual cost. The pipeline allows us to construct diverse and visually challenging h items in an image-specific and task-specific manner. Extensive experiments with 15 open-source MLLMs, 3 proprietary MLLMs, and 2 hallucination mitigation methods support our conclusions as follows. Larger visual encoders and higher input resolutions are helpful to reduce hallucination caused by visual ambiguity. The evaluation on PhD-sec and PhD-icc suggests the current models favor the textual part in the multimodal input. Resolving the conflicts between the CCS content and the model's internal knowledge demands a more comprehensive approach that is beyond isolated improvements on the visual or language components. Among the open-source MLLMs, LLaVA-OneVision is the best, followed by Molmo and InternVL-1.5. While existing benchmarks could lead to an overly optimistic expectation that the open-source models are catching up with GPT-4o, a substantial performance gap remains, as revealed by PhD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements. This research was supported by NSFC (62172420), Tencent Marketing Solution Rhino-Bird Focused Research Program and the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19864" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Anthropic. Introducing Claude 3.5 Sonnet, 2024. https://www.anthropic.com/news/claude-3-5-sonnet.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.483, + 0.227 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Nitzan Bitton-Guetta, Yonatan Bitton, Jack Hessel, Ludwig Schmidt, Yuval Elovici, Gabriel Stanovsky, and Roy Schwartz. Breaking common sense: WHOOPS! a vision-and-language benchmark of synthetic and compositional images. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. MiniGPT-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[6] Chenhang Cui, Yiyang Zhou, Xinyu Yang, Shirley Wu, Linjun Zhang, James Zou, and Huaxiu Yao. Holistic analysis of hallucination in GPT-4v(ison): Bias and interference challenges. arXiv preprint arXiv:2311.03287, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.482, + 0.576 + ], + "angle": 0, + "content": "[7] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[8] Matt Deitke, Christopher Clark, Sangho Lee, et al. Molmo and PixMo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.48, + 0.662 + ], + "angle": 0, + "content": "[9] Doubao Team. Doubao product, 2024. https://www.volcengine.com/product/doubao.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[10] Google. Google Gemini 1.5 Pro, 2024. https://deepmind.google/technologies/gemini/pro/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[11] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. HallusionBench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[12] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[13] Hongyu Hu, Jiyuan Zhang, Minyi Zhao, and Zhenbang Sun. CIEM: Contrastive instruction evaluation method for better instruction tuning. In NeurIPS Workshop on ITIF, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Yutao Hu, Tianbin Li, Quanfeng Lu, Wenqi Shao, Junjun He, Yu Qiao, and Ping Luo. OmniMedVQA: A new large-scale" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "comprehensive evaluation benchmark for medical LVLM. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[15] Liqiang Jing, Ruosen Li, Yunmo Chen, Mengzhao Jia, and Xinya Du. FAITHSCORE: Evaluating hallucinations in large vision-language models. In Findings of the Association for Computational Linguistics: EMNLP, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[16] Kushal Kafle and Christopher Kanan. An analysis of visual question answering algorithms. In ICCV, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[17] Kang-il Lee, Minbeom Kim, Seunghyun Yoon, Minsung Kim, Dongryeol Lee, Hyukhun Koh, and Kyomin Jung. VLind-Bench: Measuring language priors in large vision-language models. In NAACL Findings, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[18] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[20] Junyi Li, Jie Chen, Ruiyang Ren, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. The dawn after the dark: An empirical study on factuality hallucination in large language models. In ACL, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[21] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.525, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[23] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In ICLR, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.904, + 0.608 + ], + "angle": 0, + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[26] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. https://llava-v1.github.io/blog/2024-01-30-llava next/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[27] Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (NOPE) to measure object hallucination in vision-language models. In ALVR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.902, + 0.803 + ], + "angle": 0, + "content": "[28] OpenAI. Dall-E 3 system, 2023. https://openai.com/index/dall-e-3/.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.904, + 0.831 + ], + "angle": 0, + "content": "[29] OpenAI. Hello GPT-4o, 2024. https://openai.com/index/hello-gpt-4o/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19865" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[31] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. IllusionVQA: A challenging optical illusion dataset for vision language models. In COLM, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[32] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. In NeurIPS, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.221, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[33] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.263, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[34] Junyang Wang, Yuhang Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Ming Yan, Ji Zhang, and Jitao Sang. An LLM-free multi-dimensional benchmark for mllms hallucination evaluation. arXiv preprint arXiv:2311.07397, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.333, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[35] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.403, + 0.483, + 0.444 + ], + "angle": 0, + "content": "[36] Qijie Wei, Kaiheng Qian, and Xirong Li. FunBench: Benchmarking fundus reading skills of MLLMs. arXiv preprint arXiv:2503.00901, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.446, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[37] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.489, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[38] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Zhou. mPLUG-Owl2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[39] Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. Woodpecker: Hallucination correction for multimodal large language models. Science China Information Sciences, 67(12):220105, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[40] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. MMMU: A massive multi-discipline multimodal understanding and reasoning benchmark for expert AGI. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[41] Le Zhang, Rabiul Awal, and Aishwarya Agrawal. Contrasting intra-modal and ranking cross-modal hard negatives to enhance visio-linguistic compositional understanding. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[42] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the AI ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[43] Kankan Zhou, Eason Lai, Wei Bin Au Yeong, Kyriakos Mouratidis, and Jing Jiang. ROME: Evaluating pre-trained vision-language models on reasoning beyond visual common sense. In Findings of the Association for Computational Linguistics: EMNLP, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19866" + } + ] +] \ No newline at end of file diff --git a/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/0cae22af-0103-455d-a35f-1a510beabf2e_origin.pdf b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/0cae22af-0103-455d-a35f-1a510beabf2e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..187491c14df4c85582900af146f6326c08650579 --- /dev/null +++ b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/0cae22af-0103-455d-a35f-1a510beabf2e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c39e98d20e20b8618a0d479e6bba84814a30246e1ee614b7ebd97c091ddb6dc3 +size 8137684 diff --git a/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/full.md b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f1f150efc627a2b2df502b545322b6d76273f628 --- /dev/null +++ b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/full.md @@ -0,0 +1,365 @@ +# PhD: A ChatGPT-Prompted Visual hallucination Evaluation Dataset + +Jiazhen Liu $^{1*}$ , Yuhan Fu $^{1,2}$ , Ruobing Xie $^{2}$ , Runquan Xie $^{2}$ , Xingwu Sun $^{2}$ , Fengzong Lian $^{2}$ , Zhanhui Kang $^{2}$ , and Xirong Li $^{1\dagger}$ + +$^{1}$ Renmin University of China + $^{2}$ Machine Learning Platform Department, Tencent + +https://github.com/jiazhen-code/PhD + +# Abstract + +Multimodal Large Language Models (MLLMs) hallucinate, resulting in an emerging topic of visual hallucination evaluation (VHE). This paper contributes a ChatGPT-Prompted visual hallucination evaluation Dataset (PhD) for objective VHE at a large scale. The essence of VHE is to ask an MLLM questions about specific images to assess its susceptibility to hallucination. Depending on what to ask (objects, attributes, sentiment, etc.) and how the questions are asked, we structure PhD along two dimensions, i.e. task and mode. Five visual recognition tasks, ranging from low-level (object / attribute recognition) to middle-level (sentiment / position recognition and counting), are considered. Besides a normal visual QA mode, which we term PhD-base, PhD also asks questions with spcious context (PhD-sec) or with incorrect context (PhD-icc), or with AI-generated counter common sense images (PhD-ccs). We construct PhD by a ChatGPT-assisted semi-automated pipeline, encompassing four pivotal modules: task-specific hallucinatory item (hitem) selection, hitem-embedded question generation, spcious / incorrect context generation, and counter-common-sense (CCS) image generation. With over 14k daily images, 750 CCS images and 102k VQA triplets in total, PhD reveals considerable variability in MLLMs' performance across various modes and tasks, offering valuable insights into the nature of hallucination. As such, PhD stands as a potent tool not only for VHE but may also play a significant role in the refinement of MLLMs. + +# 1. Introduction + +Using a specific large language model (LLM) as its kernel, a multi-modal LLM (MLLM), exemplified by LLaVA [25], Qwen-VL [2] and MiniGPT-v2 [4] can now tackle a wide + +![](images/51573b5c4bc12668482aa70070c9ac37cbe5d12a450d614eacfe6ba138b39e4e.jpg) + +Are there five people in the image? + +Yes, there are five people in the image. + +They are standing near a bunch of bananas. + +![](images/2680ade224ae3647b4c20cb41abb6ee7eadff0ff185e54d85dd98db0b48659a3.jpg) + +(a) Hallucination cause I: Visual ambiguity (MLLM: LLaVA-1.5) [25] + +![](images/024fb122f8f287ca4e1de86ff2c9b38e902f56630418ed97f91a0bb71c9c373f.jpg) + +After a long playtime with + +its toy ball, the kitten, feeling + +thoroughly exhausted from + +thoroughly exhausted from all the fun, has finally found + +its resting place. It has curled up on the table, seeking comfort and rest after its energetic adventures. + +![](images/51556ece3eb4b096708dd3a43d21f7899e6822dabbbdc67f7ffe4356658ed86e.jpg) + +Is there a toy in the image? + +Yes, the kitten is playing + +with a toy ball. + +![](images/047bc6b5724bb26f26e69b59632039d8b495af91ce1499e71a4bb563e0261b2c.jpg) + +(b) Hallucination cause II: Inconsistency in multi-modal input + +![](images/87bfd678fdbcb85444a4ba5ba0352e9f3533e3ea7e5e88b8dfb26f8286bbc7da.jpg) + +![](images/6ad6cbdd5c082aceb923b355560b9b8013bdcab5928fb862d4394cf0058bb7b2.jpg) + +Is the mouse in the image smaller than the cat? + +Yes, the mouse is smaller than the cat. + +![](images/68b0e82990025d2b9b673fa14738b3ab16ee1bd1f8d172f1dec2c0b9f6e4314e.jpg) + +(c) Hallucination cause III: Counter-common-sense content + +![](images/cbf98828d50186939dd14d9a1b6540911aad192ae7bca33e1ccbcc632766c070.jpg) + +(d) Performance curves of the LLaVA series on two public VHE datasets (POPE [21]) and AMBER [34]) and the proposed PhD dataset. + +Figure 1. Illustrations of three major causes of an MLLM's visual hallucination and its evaluation. This paper contributes PhD, a binary VQA-based VHE benchmark, much larger and more challenging than its predecessors. In particular, it has four evaluation modes that explicitly measure an MLLM's performance w.r.t. the three causes, i.e. PhD-base for cause I, PhD-sec and PhD-icc for cause II and PhD-ccs for cause III. + +range of computer vision tasks in a unified visual-question-answering (VQA) manner. As LLMs are known to hallucinate [20, 37, 42], it is not surprising that MLLMs have visual hallucination, generating fabricated interpretation of the given visual content, see Fig. 1. Considering the rapidly growing use of MLLMs in varied scenarios, a comprehensive visual hallucination evaluation (VHE) is crucial. This paper develops a new dataset for VHE. + +VHE essentially involves posing a number of visual questions to an MLLM [11, 21]. A question of this kind shall include a hallucinatory item (hitem), in the form of a specific word or phrase, that induces the MLLM to generate a response discordant with the provided visual content. As the model typically has a strong visual recognition ability, how to identify an appropriate hitem and accordingly generate a proper question is nontrivial. Both the hitem and the question depend on the visual recognition task being considered. As shown in Tab. 1, we target at objective VHE in the context of low-level (object / attribute) to middle-level (sentiment / position / counting) visual recognition tasks. Such a target is chosen due to the following considerations. MLLMs generally work well for these tasks, so their erroneous responses can be largely attributed to their hallucinations instead of their incapability, e.g. asking a generic MLLM to read pathology images [14]. Meanwhile, a binary VQA based objective evaluation is more budget friendly and thus more suited for VHE at a large scale. + +
Vision tasksObjective evaluationSubjective evaluation
Low-/middle-level visual recognition+ POPE, EMNLP'23 [21]+ FAITHSCORE, EMNLP'24 [15]
+ AMBER, arXiv'23 [34]+ HaELM, arXiv'23 [35]
+ CIEM, ITIF'23 [13]+ M-HalDetect, AAAI'24 [12]
+ NOPE, ALVIR'24 [27]+ GAVIE, ICLR'24 [23]
+ ROME, EMNLP'23 [43]
+ PhD (this paper)
High-level visual reasoning+ MMMU, CVPR'24 [40]+ HallusionBench, CVPR'24 [11]
+ VLind-Bench, NAACL'25 [17]+ Bingo, arXiv'23 [6]
+ IllusionVQA, COLM'24 [31]
+ WHOOPS!, ICCV'23[3]
+ +Pioneered by POPE [21], good attempts exist in objective VHE [13, 27, 34, 43]. In these valuable datasets, hitem selection is largely untouched. POPE and ROME [43] select their hitems fully based on label co-occurrence in training data, AMBER [34] relies on manual annotation, whilst hitem selection is not considered in NOPE [27] and CIEM [13], see Tab. 2. Hence, there lacks an explicit link between hitem selection (and subsequent VQA triplets construction) and major causes of an MLLM's visual hallucination. As models rapidly evolve, the performance on these datasets quickly reaches saturation, see Fig. 1d. + +Analyzing an MLLM's typical dataflow of answering a visual question, we see three major causes of visual hallucination: I) visual ambiguity, II) inconsistency in multi-modal input and III) counter-common-sense (CCS) content, see Fig. 1. Firstly, the MLLM extracts tokenized visual features from a given image using a ViT based encoder. Recent studies [33, 41] show that the features tend to be high level, lacking sufficient details for fine-grained tasks such as counting. Secondly, the visual features, after vision-to-language adaptation, are mixed with the features of the associated textual prompt to form a multi-modal input to the LLM kernel. The LLM, pre-trained extensively on textual data, inevitably favors the textual part of the multi-modal input. Hence, when there is inconsistency between the visual and textual information, the former is more likely to be overruled. Lastly, at the decoding stage, the LLM might heavily rely on its internal (world) knowledge, ignoring the visual content especially when the content (showing a mouse much larger than a cat) contradicts the common sense. Our new dataset is developed with a close link to the three causes. + +Table 1. Taxonomy of VHE benchmarks. Our PhD benchmark performs an objective evaluation of MLLMs' hallucinations when they address visual recognition tasks ranging from low-level, i.e. object / attribute recognition to middle-level, i.e. sentiment / positional recognition and counting. + +
DatasetDaily imagesCCS imagesHItemsContextsVQA tripletsTasks
POPE500✘80✘3,000Obj.
NOPE*unknown✘unknown✘32,701Obj.
CIEM*4,929✘unknown✘72,941Obj. / Attr. / Pos.
AMBER1,004✘687✘14,216Obj. / Attr./Pos. / Count.
ROME✘1,563118✘1,563Attr./Pos.
PhD14,6487501,45233,688102,564Obj / Attr. / Pos. / Sent. / Count.
+ +Table 2. PhD versus its predecessors. * indicates private dataset. + +The new dataset is constructed by adapting TDIUC, a popular multi-task VQA dataset [16], with a ChatGPT-assisted semi-automated pipeline, see Fig. 2. In particular, by prompting ChatGPT, we select diverse and visually challenging h items in an image-specific and task-specific manner, with minimal human involvement primarily spent on verifying ChatGPT-generated results. The selected h items are then automatically embedded into visual questions, specious context, and incorrect context, all generated by instructing ChatGPT. Moreover, we expand the daily image set with counter common sense (CCS) images, obtained by prompting AIGC tools with ChatGPT-generated CCS descriptions, e.g. "trees growing underwater" and "a car with square-shaped wheels". The dataset is dubbed as PhD (ChatGPT Prompted visual hallucination evaluation Dataset). Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base (questions about daily images w/o context), PhD-sec (PhD-base plus specious context), PhD-icc (PhD-base plus incorrect context), and PhD-ccs (questions about CCS images). + +To sum up, our major contributions are as follows: + +- We introduce PhD, a dataset with four evaluation modes across five visual recognition tasks, developed with a close + +link to the three major causes of MLLM visual hallucination. Information on hallucinatory items (hItems) is provided per sample, enabling in-depth analytics to uncover the causes in more detail. + +- We offer a ChatGPT-assisted semi-automatic pipeline for dataset construction, with minimal human involvement, primarily focused on verifying the generated results. With 14,648 daily images, 750 CCS images and 102,564 VQA triplets in total, PhD is the largest of its kind. +- We conduct an extensive evaluation with 15 open-source MLLMs, 3 proprietary MLLMs, and 2 hallucination mitigation methods, showing the viability of PhD for VHE in varied manners including overall, mode-oriented, task-oriented, and model-wise zoom-in. The evaluation not only reveals inter-model performance divergence, but may also help developers of a specific MLLM to prioritize their efforts in refining the model. + +# 2. Related Work + +Due to the increasing importance of VHE, new benchmarks are being actively developed. Depending on what vision tasks they focus on and how their evaluation is executed, we categorize existing achievements along two dimensions, i.e. task and evaluation, see Tab. 1. Concerning the task dimension, low- and middle-level visual recognition tasks, ranging from object / attribute recognition to sentiment / positional recognition and counting, assess an MLLM's basic visual skills. High-level visual reasoning is more domain-knowledge intensive, typically covering image-based math question solving, geography information understanding, meme interpretation, historical or folkloric contexts, etc. [3, 6, 11, 17]. As for the evaluation dimension, objective evaluation refers to objectively comparing the model's output with ground truth, mostly in the form of Yes/No answers [13, 21, 34, 43]. By contrast, subjective evaluation requires humans or LLMs to assess the model's output [3, 6, 11]. The proposed PhD, focusing on low/middle-level visual recognition and objective evaluation, belongs to the second quadrant of the taxonomy. In what follows, we briefly review peer benchmarks, i.e. POPE [21], ROME [43], NOPE [27], CIEM [13], and AMBER [34], in this quadrant and clarify our novelty accordingly. + +POPE is probably the first dataset to evaluate object hallucination [21]. Given a specific MS-COCO image with object labels, POPE selects an adversarial object frequently co-occurring with the current labels. A binary question is then formed by filling out a predefined template with the selected object. Such co-occurrence based hitem selection is not image-specific by definition. Trivial hitems might thus be picked up, e.g. "car" selected for an indoor image labeled with "person", as the two objects often co-occur. For CCS image generation, ROME forms CCS descriptions by choosing attribute values having the lowest co-occurrence + +with a given object according to the Visual Genome dataset [43]. As low occurrence is not necessarily CCS, some of ROME images are indeed normal. NOPE [27] and CIEM [13] simply bypass hitem selection by asking an LLM to generate questions conditioned on the image captions (also from MS-COCO) and pre-specified answers. + +To select h items in an image-specific manner, AMBER resorts to fully manual annotation [34]. Manual labeling is costly, while an annotator's personal vocabulary is relatively limited. All this makes it difficult to scale up w.r.t. the amount of test images and the number of distinct h items. + +In comparison, PhD, constructed by a ChatGPT-assisted semi-automatic pipeline (Sec. 3), is much larger (Tab. 2) and more challenging (Fig. 1d). With its unique mode-task structure, the new dataset enables a novel, structured and zoom-in understanding of inter-model difference. + +# 3. Our Roadmap to PhD + +As MLLMs perform visual recognition in a VQA manner, a VQA sample for VHE naturally depends on the visual recognition task in consideration. We depart from TDIUC [16], a large-scale VQA dataset w.r.t. five tasks including object / attribute / sentiment / positional recognition and counting. Note that the images in TDIUC are sourced from MS-COCO [22], which plausibly have been seen by MLLMs in their development stage. As such, our adoption of TDIUC makes on-purpose data leakage: an MLLM's erroneous response w.r.t. a seen image can now be more safely attributed to its hallucination other than its incapability in visual recognition, say asking LLaVA to recognize glaucoma from color fundus photographs [36]. We construct PhD by adapting the TDIUC annotations with a ChatGPT1-assisted semi-automated pipeline, see Fig. 2. + +In order to compose a proper question that effectively makes an MLLM hallucinate about a given image, a hitem has to be first identified in an image-specific and task-specific manner. Then, the hitem has to be smoothly embedded in the form of a specific word or phrase into the question. We describe task-specific hitem selection in Sec. 3.1, followed by hitem-embedded question generation in Sec. 3.2. Furthermore, in order to simulate inconsistency in the multi-modal input, we prepend specious or incorrect context to the question, the generation of which is detailed in Sec. 3.3. Lastly, in order to explicitly create conflicts between the visual input and the internal (world) knowledge of the MLLM, we expand our image collection with autogenerated counter-common-sense (CCS) images (Sec. 3.4). + +# 3.1. Task-specific Hitem Selection + +Without loss of generality, we describe how a hitem is selected for color attribute recognition. Let us consider the + +![](images/646c6dc4e7a35c48032ead0b4d1db978eb5c5caa8db07719269bf1c711925893.jpg) +(1) Task-specific hitem selection + +![](images/6a1fb24641db27d7e7c817515a0c3e9246c98bc3f0d50ba9cb3edf4792af3e5d.jpg) +(2) Hitem-embedded question generation +(a) Hitem-embedded question / context generation for MS-COCO daily images + +![](images/4bf9d8d0909bb5ec79279fc3b476c78daea8d7ea2907d374a352ab0d79dd4389.jpg) +(b) QAs for AI-generated CCS images + +![](images/a20f5410384ac7b3cacd15ece0e2ac60f70ab590b49b5a2ddf9a0be4e0eda7b3.jpg) +Task: Object recognition +Are there headphones in the image? + +In recent years, this musician has often immersed himself in his own musical world, playing the guitar with headphones on as if cut off from the outside world. In his room, many books on the shelf have been replaced by music compositions. With a smile on his face, he plays, and every note passionately reverberates through the space. + +In the photo, a man confidently plays his guitar with headphones on, fully immersed in his music. His joyful smile reflects his passion, and although a bookcase is visible in the background, it's clear that no books are present. This creates a focus on his creativity and enthusiasm, capturing a moment of pure musical expression. + +![](images/bbd20cbdfed5eade34256728b8a4b1d4e7689ed698daa4ce36e60cf75b07c54c.jpg) + +![](images/6a3c223e92b7b95afad980d10c040a3748a98106737e438852ad81adab4c6a36.jpg) +Are there books in the image? +Task: Sentiment recognition +Is the girl feeling lonely in the image? +Is the girl feeling happy in the image? + +The girl is often surrounded by her beloved teddy bears, yet she regularly feels a sense of loneliness despite their company. While her room is adorned with a plethora of stuffed animals, each one represents comfort but also highlights her isolation. + +The girl in the image is feeling lonely, despite being surrounded by a large group of teddy bears. She lies on the bed near many stuffed animals, including dozens of brown teddy bears, and her contrast between her solitary emotions and the cheerful presence of her furry companions. + +![](images/baa837148544a3bed93a35352ba8aa6b0d125a73da2247590bfc4c748b3f6e36.jpg) +Task: Positional +Is there a table behind +the looking dog in the image? +Is there a fence behind +the looking dog in the image? + +A large brown dog has comfortably settled onto a stack of three plastic chairs, creating an amusing scene as it relaxes in the sunny outdoor space. Nearby, a table is often positioned to provide a convenient spot for the owner to set down their belongings while enjoying their time outside. + +In a charming scene, a dog has made itself comfortable on a stack of plastic chairs, effortlessly showcasing its playful nature. Behind the relaxed dog captured in the photo, a table can be filled with a large box of cozy atmosphere. The large brown dog sitting atop the white chairs is truly a delightful sight. + +![](images/c9cd613d5a37f9da3944ee086a421297a86ef13c5cd1187109a1bff7ef2e359b.jpg) +Is the ball being kicked a soccer ball? +Is the ball being kicked a basketball? + +![](images/7df577b988cb09d7417eb469eacf4517b97d91a35d1b4eaca4d6886cf878a60b.jpg) +Is the mouse smaller than the cat? + +![](images/aeece98e4fc0c4739e290283298a8866bb334f01e7c7b489bf23def3ae8aa6fa.jpg) +Is the highest number +of points on the dice six? +Is the highest number +points on the dice seven? + +![](images/0c79a4a059a2df4dee4322880a54350267ee5874b09b19e5c68c90b2b52f62ab.jpg) +Is the cat smaller than the mouse? +Is the shape of the orange round? +Is the shape of the orange square? +(d) Showcase: CCS images and questions +Figure 2. Proposed semi-automatic pipeline for PhD construction. We use ChaptGPT (GPT-4o mini) to generate hitem-embedded questions / contexts for daily images, and Doubao and DALL-E3 for generating CCS images. Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base, i.e. questions about daily images w/o context, PhD-sec, i.e. PhD-base plus specious context, PhD-icc, i.e. PhD-base plus incorrect context, and PhD-ccs, i.e. questions about CCS images. By adapting TDIUC annotations, PhD supports binary VQA w.r.t. five visual recognition tasks including object / attribute / sentiment / positional recognition and counting. With 20 mode-task combinations in total, PhD enables a comprehensive VHE. + +image in Fig. 2a, showing a black motorcycle followed by a red bus. As no red motorcycle is present while the red color is prominent near the black motorcycle, the word red will be a good choice of hitem to challenge an MLLM. + +Vocabulary Construction per task. Started with a handful of manually specified colors such as red, green, and blue, we ask ChatGPT to expand the color vocabulary with instructions like "Please expand the input vocabulary as much as possible by adding common items found in daily life. Avoid any duplication", getting a set of 35 different colors. + +Subject-Attribute Extraction. The image is associated with a TDIUC question-answer pair as "what color is the motorcycle" and "black". We use ChatGPT (with simple instructions) to extract with ease the subject (i.e. motorcycle) and its attribute (i.e. black) from the pair. + +Candidate Hitem Generation. We obtain candidate h items by excluding the ground-truth (GT) answer (and its synonyms if applicable) from the vocabulary. + +Visual-based Hitem Ranking. Intuitively, a hitem shall + +be visually plausible in the given image. So for each candidate hitem, we compute its similarity to the image using a pre-trained CLIP [30]. In particular, the cosine similarity between the CLIP embeddings of hitem + subject (e.g. green motorcycle) and the image is adopted. Ranking the candidates by the CLIP similarity lets us to select the one visually closest to the image. It is worth noting that for emerging MLLMs equipped with stronger vision encoders [19], our pipeline is likely to produce even more effective h items by replacing CLIP with these advanced counterparts. + +Manual Inspection. While the above process is generally stable to produce satisfying results, manual inspection is performed to ensure the correctness of hitem selection. Note that due to errors in the original TDIUC annotations, occasionally the true label might be "incorrectly" selected. In such a case, we simply discard the VQA sample. + +With lightweight task-specific adaptation, the above hitem selection also works for other tasks. Overall, the joint use of ChatGPT and CLIP allows us to select 1,452 hitem + +that are more diverse and challenging than their counterparts in previous datasets, see Tab. 2. + +# 3.2. Hitem-embedded Question Generation + +For a given subject (e.g. motorcycle) and a chosen hitem (e.g. red), generating a hitem-embedded question is trivial for ChatGPT. In particular, a No question is formed as "Is the motorcycle in the image red?" Meanwhile, a Yes question is simultaneously generated using the GT as "Is the motorcycle in the image black?". This ensures perfect Yes/No balance among the generated questions. + +# 3.3. Specious (Incorrect) Context Generation + +When used as a document parser, an MLLM reading a specific image is often provided with the image's surrounding text. Inconsistency between the image and the text is not uncommon. A news article containing a general claim of "red motorcycles frequently zip through the streets" does not necessarily have each of its illustrated pictures match with the claim. To simulate such a scenario, for a given image we generate specious text as specious context and text contradicting the image as incorrect context, respectively. Next, we describe the generation of specious contexts, as their incorrect counterparts can be generated in a similar but more simplified manner. + +Specious Text Generation. Using the previously generated hitem-embedded question and the original MS-COCO captions as input, we instruct ChatGPT to generate specious text for a given image. By "specious", we mean the text is specious or noisy, rather than directly contradicting the image content. As such, our instruction reads partially as "Please generate the for the given question. It should be one sentence. The should answer the question, but it may not reflect the actual current status, thus making it specious." + +Text Composition. ChatGPT is used to seamlessly merge the specious text with ground-truth captions, forming a longer context in which only a small portion (orange text in Fig. 2c) is mildly inconsistent with the image. + +Manual Inspection. We perform a spot check. If the context quality is low, we simply discard the entire sample. + +# 3.4. CCS Image Generation + +We generate CCS images by first generating CCS descriptions and then employing Text2Image tools to convert the descriptions to CCS images, see Fig. 2b. + +CCS Description Generation. A number of manually written task-specific samples, see Tab. 3, are used as in-context learning samples for ChatGPT to generate more descriptions. The descriptions have to be visually expressible, so bad cases like "the more you eat, the thinner you get" are filtered out manually. For each CCS text, its common-sense + +(CS) counterpart is simultaneously generated by ChatGPT, by providing the learning samples in pair. + +Text2Image. The generated CCS descriptions are used as prompts for AIGC tools (Doubao [9] and DALL-E3 [28]) to generate the corresponding CCS images. The quality of the generated images depends on various factors, making occasional failures inevitable. When this occurs, we attempt to refine the prompts or apply region-based inpainting. The sample will be discarded if the above attempts fail. + +Question Generation. Per CCS description (e.g. A car with square wheels), we utilize ChatGPT to generate a Yes question (e.g. Does the car have square wheels?). Again, for balancing Yes/No questions, we generate a No question (e.g. Does the car have round wheels?) based on the CS description. + +
CCS descriptionYes questionCS descriptionNo question
Task: Object recognition
Manually written:
Ice blocks in volcanic lavaAre there ice blocks in volcanic lava?Fire in volcanic lavaIs there fire in volcanic lava?
Grass in a tiger's mouthIs there grass in a tiger's mouth?Meat in the tiger's mouthIs there meat in the tiger's mouth?
ChatGPT generated:
Trees growing underwaterAre there trees growing underwater?Coral growing underwaterIs there coral growing underwater?
Books in a swimming poolAre there books in a swimming pool?Water in a swimming poolIs there water in a swimming pool?
Birds flying underwaterAre there birds flying underwater?Birds flying in the skyAre there birds flying in the sky?
Ice cream in a volcanoIs there ice cream in a volcano?Lava in a volcanoIs there lava in a volcano?
Computers in a forestAre there computers in a forest?Animals in a forestAre there animals in a forest?
Task: Attribute recognition
Manually written:
A car with square wheelsDoes the car have square wheels?A car with round wheelsDoes the car have round wheels?
Blue apples on the treeAre the apples on the tree blue?Red apples on the tree.Are the apples on the tree red?
ChatGPT generated:
A green skyIs the sky green?A blue skyIs the sky blue?
A bicycle with square wheelsDoes the bicycle have square wheels?A bicycle with round wheelsDoes the bicycle have square wheels?
A tree made of metalIs this tree made of metal?A wooden treeIs this tree a real wood tree?
A chocolate riverIs there chocolate flowing in the stream?A water riverIs there water flowing in the stream?
A house made of candyIs the house made of candy?A house made of bricksIs the house made of bricks?
+ +Table 3. Instances of descriptions used for generating CCS images and related questions. With manually written CCS / CS descriptions as instructions, ChatGPT is used to generate many more instances and subsequently convert them to Yes/No questions. + +# 3.5. Dataset Overview and PhD Index + +An overview of the PhD dataset is given in Tab. 4. Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base, i.e. questions about daily images w/o context, PhD-sec, i.e. PhD-base plus spcious context, PhD-icc, i.e. PhD-base plus incorrect context, and PhD-ccs, i.e. questions about CCS images. With 20 mode-task com + +binations in total, PhD supports a much more comprehensive VHE than its predecessors [21, 34]. + +
TasksQuestions
ObjectAttributeSentimentPositionCountingYesNo
TDIUC samples used6,2714,3242,0952,8413,387--
Unique h items7451466548610--
VQA samples in PhD-base11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-sec11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-icc11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-ccs34473478220124750750
+ +To measure the performance of an MLLM on PhD, we compute its recall w.r.t. the Yes and No questions, respectively. We term the harmonic mean of the two recalls PhD Index. A model simply saying Yes (or No) to all questions has a PhD Index of 0, while a random-guess score is 0.5. + +# 4. Evaluating MLLMs on PhD + +# 4.1. Common Setup + +Choices of MLLMs. For reproducible research, we focus on open-source MLLMs, compiling a list of 15 models that span varied sizes and architectures. see Tab. 5. We also evaluate two hallucination mitigation methods, VCD [18] and Woodpecker [39], currently supporting LLaVA-1.6-L and Qwen-VL. As additional references, we assess three proprietary MLLMs, i.e. GPT-4o [29], Claude 3.5 Sonnet [1], and Gemini 1.5 Pro [10], on a random subset of 2k samples (random-2k) subject to our budget. For the same reason we evaluate Woodpecker, which requires paid service from ChatGPT, on random-2k. + +Test Protocol. For a fair comparison, per MLLM we use its designated prompt to wrap each test question. For instance, a question-specific prompt submitted to mPLUG-Owl2 will be in the form of "USER: $< |image|>\{question\}$ ASSISTANT:" We provide more prompts in the supplement. In addition, to help the models better handle PhD-sec and PhD-icc, we append to the test prompt an instruction as "In case there is an inconsistency between the context and the image content, you should follow the image." + +# 4.2. Using PhD for Overall VHE + +An overall VHE as shown in Tab. 5 is useful for providing a big picture of which MLLM hallucinates the most (or the least). The leading open-source MLLMs are LLaVA-OneVision, followed by Molmo and InternVL-1.5. Since their vision encoders and LLMs vary, the results are insufficient to conclude which component is the most effective to mitigate hallucinations. That said, comparisons among the same model series remains meaningful. Consider the LLaVA series for instance. While one would normally expect that a larger LLM yields a better MLLM, as LLaVA-1.6-L vs LLaVA-1.6, the difference between LLaVA-1.5-L + +and LLaVA-1.5 is marginal (0.270 vs 0.265). In order to analyze and consequently understand such an counterintuitive result, PhD enables a zoom-in analysis in mode-oriented (Sec. 4.3) and task-oriented (Sec. 4.4) styles, unavailable in the previous benchmarks. + +Table 4. Data statistics of the proposed PhD dataset. + +
ModelViTLLMPOPEAMBERPhD
Full-set evaluation:
LLaVA-OneVision [19]SoViT-400m/14Qwen2-72B0.840.900.698
Molmo [8]-L/14Qwen2-72B0.840.850.690
InternVL-1.5 [5]InternViT-6BInternLM2-20B0.860.890.561
Qwen-VL (VCD)-bigG/14Qwen-7B0.840.870.560
Cambrian-1 [32]HybridLlama-3-8B0.880.890.547
LLaVA-1.6-L (VCD)-L/14Vicuna-13B-1.50.820.810.511
LLaVA-1.6-XL [26]-L/14Nous-Hermes-2-Yi-34B0.860.840.492
Qwen-VL [2]-bigG/14Qwen-7B0.830.840.488
LLaVA-1.6-L [26]-L/14Vicuna-13B-1.50.830.800.423
MiniGPT-v2 [4]-G/14Llama-2-7B0.830.840.390
LLaVA-1.6 [26]-L/14Vicuna-7B-1.50.830.790.373
mPlug-Ow12 [38]-L/14Llama-2-7B0.780.770.320
InstructBLIP [7]-G/14Vicuna-7B-1.10.820.820.305
InstructBLIP-L [7]-G/14Vicuna-13B-1.10.800.790.278
LLaVA-1.5-L [25]-L/14Vicuna-13B-1.50.820.730.270
LLaVA-1.5 [25]-L/14Vicuna-7B-1.50.810.750.265
LLaVA-1.1 [24]-L/14Vicuna-7B-1.10.670.330.135
Random-2k evaluation:
GPT-4o [29]--0.880.870.812
Claude 3.5 Sonnet [1]--0.850.890.746
Gemini 1.5 Pro [10]--0.860.880.691
Qwen-VL (Woodpecker)-bigG/14Qwen-7B--0.531
LLaVA-1.6-L (Woodpecker)-L/14Vicuna-13B-1.5--0.409
+ +Table 5. Overall VHE. POPE (adversarial) and AMBER (discriminative) are used. + +One more advantage of PhD compared to its predecessors lies in its discrimination ability. The relatively small performance gap between GPT-4o and the top open-source models as measured by POPE and AMBER might lead to an overly optimistic interpretation that the open-source alternatives are catching up with the proprietary model. In fact, a substantial gap remains, as revealed by PhD. Fig. 3 further illustrate the qualitative results, where GPT-4o exhibits fewer hallucinations in its response. + +# 4.3. Using PhD for Mode-Oriented VHE + +Fig. 4a illustrates mode-wise model performance. MLLMs working relatively well in the PhD-base mode tend to have stronger visual input. This is achieved either with stronger visual encoders, as the cases of LLaVA-OneVision, InternVL-1.5, and Cambrian-1 using SoViT-400m/14, InternViT-6B or hybrid vision structure, or supporting higher image resolutions, see Molmo and MiniGPT4-v2 that accept multiscale or larger input. + +In contrast to the visual part, using a larger LLM alone does not necessarily lead to a better MLLM. As noted in Sec. 4.2, LLaVA-1.5-L with Vicuna-13B-1.5 has nearly the same PhD Index (0.270) as LLaVA-1.5 with Vicuna-7B-1.5 (0.265). We see from Fig. 4a that the larger LLM indeed improves the performance in PhD-sec $(0.082\rightarrow 0.099)$ , PhD-icc $(0.011\rightarrow 0.019)$ and PhD-ccs $(0.534\rightarrow 0.542)$ , yet suffers loss in PhD-base $(0.443\rightarrow 0.422)$ . Similar results can be more evidently observed in the case of InstructBLIP-L vs. InstructBLIP (PhD-base Index: $0.535\rightarrow 0.324$ ). Our con + +![](images/28682202df7b5fde01e490615c496a5370c413a8b54fb7d34ad2e55dd4bd55bc.jpg) +Figure 3. Qualitative results showing how an MLLM answers visual questions from PhD. The correctness of an answer is automatically determined by matching its first word, either Yes or No, with the ground truth (GT). + +jecture is that although a larger LLM better understands user instructions, its successful use within an MLLM requires more targeted training for vision-language alignment. + +Comparing the four modes, the performance of the open-source MLLMs on PhD-icc and PhD-sec is generally low. When provided with a multi-modal input, the models favor the textual part. Substituting a 13-B LLM for its 7-B counterpart helps tackling the inconsistency in the multi-modal input, see Fig. 5. However, a larger LLM might rely more on its internal knowledge for decoding, resulting in worse performance on PhD-base and PhD-ccs for which the image content shall carry more weights. In particular, PhD-ccs reveals a deeper and intrinsic challenge of VH: conflicts between the given image content and the model's internal knowledge. Solving the challenge demands a more comprehensive approach that goes beyond isolated improvements on the visual or language components. + +Among the open-source MLLMs, LLaVA-OneVision is the best, owing to its joint use of a stronger visual encoder (SoViT-400m/14), a more powerful LLM (Qwen2-72B), and better training strategies (much larger high-quality training data and multi-stage alignment). Nevertheless, LLaVA-OneVision remains inferior to GPT-4o, particularly in PhD-sec and PhD-icc, wherein inconsistency within the multi-modal input has to be properly addressed. + +# 4.4. Using PhD for Task-Oriented VHE + +Fig. 4b presents task-oriented VHE results. In general, to what extent an MLLM hallucinates is largely correlated to the required level (low / middle / high) of a specific task. The object recognition task has the overall highest PhD Index, followed by attribute recognition, positional recognition, counting, and sentiment recognition. Due to the complexity and subtlety of emotions, e.g. tears can be associated with both happiness and sadness, even GPT-4o performs + +relatively worse in the sentiment task (first row of Fig. 3). + +Joint mode-task analytics per model is shown in Tab. 6. LLaVA-OneVision struggles with sentiment recognition and counting, especially when faced with textual or CCS distractions, underscoring the need for improvement in these areas. Similarly, Molmo also faces these challenges, but its counting performance under CCS distractions is notably better than LLaVA-OneVision's (0.737 vs 0.563). The above zoom-in analytics will be informative for MLLM developers to prioritize their efforts on model refinement. + +
TaskPhD-basePhD-secPhD-iccPhD-ccs
LLAVA-OneVision
Object0.8720.8490.8240.727
Attribute0.8480.7440.6630.767
Sentiment0.6910.5810.5040.731
Positional0.7730.7300.6540.701
Counting0.7070.6520.5000.563
Molmo
Object0.8250.8800.8470.678
Attribute0.8420.7250.5560.791
Sentiment0.5470.6020.5680.746
Positional0.6970.6910.6540.742
Counting0.7270.5800.3500.737
+ +Table 6. Zoom-in analytics of specific models. + +# 4.5. Analysis of MLLM Answer Tendency + +While the yes and no questions are perfectly balanced by design, we observe that the open-source MLLMs tend to answer yes, with the say-yes rate ranging from 0.462 (Molmo) to 0.811 (LLaVA-1.1) and an average value of 0.611. By contrast, the three proprietary MLLMs have a clearly lower say-yes rate. Similar observations are reported in [21, 23]. We go one step further by analyzing how the say-yes tendency is related to model performance. Ranking the models by their PhD Index and say-yes rate, + +![](images/c5d732949e14a23a364068646dcb2c5ff8634ed802d965cb771589d5b17cdcdc.jpg) +(a) Mode-oriented VHE + +![](images/c5e82218d4718b6c092c53ccbeb7bc9ef1daf9006999695476e3a37fe17c970b.jpg) +(b) Task-oriented VHE +Figure 4. PhD based VHE analytics. Models required paid services, shown in gray markers, are tested on random-2k. +Figure 6. MLLM say-yes rate vs. PhD Index, with Spearman correlation of -0.92. Proprietary models tested on random-2k. + +respectively, we calculate the Spearman's correlation between the two ranks. A strong negative correlation exists, see Fig. 6. The result suggests that addressing VH requires balancing output tendencies, with a particular focus on enhancing an MLLM's ability to say no. + +# 5. Summary and Conclusions + +We have introduced PhD, a large-scale benchmark developed with a close link to the three causes of visual hallucination, i.e. visual ambiguity, inconsistency in multi-modal input and CCS content. We propose a ChatGPT-assisted semi-automated pipeline to construct the new dataset with + +![](images/acff34f0eb8e809b5a38ddef62c4204a3e20e42ef2c2dad88d16bb2c45770776.jpg) +Figure 5. Impact of LLM size (7B vs 13B) on LLaVA-1.5, LLaVA-1.6 and InstructBLIP. MLLMs using a 13B LLM tend to be better than their counterparts using a 7B LLM on PhD-sec and PhD-icc, yet worse on PhD-base and PhD-ccs. + +![](images/307c17e5c327a1a5e1e9c5c60010a436ef845189864c64a1716ddf7a45eb0c54.jpg) + +well affordable manual cost. The pipeline allows us to construct diverse and visually challenging h items in an image-specific and task-specific manner. Extensive experiments with 15 open-source MLLMs, 3 proprietary MLLMs, and 2 hallucination mitigation methods support our conclusions as follows. Larger visual encoders and higher input resolutions are helpful to reduce hallucination caused by visual ambiguity. The evaluation on PhD-sec and PhD-icc suggests the current models favor the textual part in the multimodal input. Resolving the conflicts between the CCS content and the model's internal knowledge demands a more comprehensive approach that is beyond isolated improvements on the visual or language components. Among the open-source MLLMs, LLaVA-OneVision is the best, followed by Molmo and InternVL-1.5. While existing benchmarks could lead to an overly optimistic expectation that the open-source models are catching up with GPT-4o, a substantial performance gap remains, as revealed by PhD. + +Acknowledgements. This research was supported by NSFC (62172420), Tencent Marketing Solution Rhino-Bird Focused Research Program and the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001). + +# References + +[1] Anthropic. Introducing Claude 3.5 Sonnet, 2024. https://www.anthropic.com/news/claude-3-5-sonnet.6 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 6 +[3] Nitzan Bitton-Guetta, Yonatan Bitton, Jack Hessel, Ludwig Schmidt, Yuval Elovici, Gabriel Stanovsky, and Roy Schwartz. Breaking common sense: WHOOPS! a vision-and-language benchmark of synthetic and compositional images. In ICCV, 2023. 2, 3 +[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. MiniGPT-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 6 +[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. 6 +[6] Chenhang Cui, Yiyang Zhou, Xinyu Yang, Shirley Wu, Linjun Zhang, James Zou, and Huaxiu Yao. Holistic analysis of hallucination in GPT-4v(ison): Bias and interference challenges. arXiv preprint arXiv:2311.03287, 2023. 2, 3 +[7] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 6 +[8] Matt Deitke, Christopher Clark, Sangho Lee, et al. Molmo and PixMo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 6 +[9] Doubao Team. Doubao product, 2024. https://www.volcengine.com/product/doubao.5 +[10] Google. Google Gemini 1.5 Pro, 2024. https://deepmind.google/technologies/gemini/pro/.6 +[11] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. HallusionBench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In CVPR, 2024. 2, 3 +[12] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. In AAAI, 2024. 2 +[13] Hongyu Hu, Jiyuan Zhang, Minyi Zhao, and Zhenbang Sun. CIEM: Contrastive instruction evaluation method for better instruction tuning. In NeurIPS Workshop on ITIF, 2023. 2, 3 +[14] Yutao Hu, Tianbin Li, Quanfeng Lu, Wenqi Shao, Junjun He, Yu Qiao, and Ping Luo. OmniMedVQA: A new large-scale + +comprehensive evaluation benchmark for medical LVLM. In CVPR, 2024. 2 +[15] Liqiang Jing, Ruosen Li, Yunmo Chen, Mengzhao Jia, and Xinya Du. FAITHSCORE: Evaluating hallucinations in large vision-language models. In Findings of the Association for Computational Linguistics: EMNLP, 2024. 2 +[16] Kushal Kafle and Christopher Kanan. An analysis of visual question answering algorithms. In ICCV, 2017. 2, 3 +[17] Kang-il Lee, Minbeom Kim, Seunghyun Yoon, Minsung Kim, Dongryeol Lee, Hyukhun Koh, and Kyomin Jung. VLind-Bench: Measuring language priors in large vision-language models. In NAACL Findings, 2025. 2, 3 +[18] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In CVPR, 2024. 6 +[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 4, 6 +[20] Junyi Li, Jie Chen, Ruiyang Ren, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. The dawn after the dark: An empirical study on factuality hallucination in large language models. In ACL, 2024. 2 +[21] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, 2023. 1, 2, 3, 6, 7 +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 3 +[23] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In ICLR, 2024. 2, 7 +[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 6 +[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 1, 6 +[26] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. https://llava-v1.github.io/blog/2024-01-30-llava next/.6 +[27] Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (NOPE) to measure object hallucination in vision-language models. In ALVR, 2024. 2, 3 +[28] OpenAI. Dall-E 3 system, 2023. https://openai.com/index/dall-e-3/.5 +[29] OpenAI. Hello GPT-4o, 2024. https://openai.com/index/hello-gpt-4o/.6 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4 + +[31] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. IllusionVQA: A challenging optical illusion dataset for vision language models. In COLM, 2024. 2 +[32] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. In NeurIPS, 2024. 6 +[33] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, 2024. 2 +[34] Junyang Wang, Yuhang Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Ming Yan, Ji Zhang, and Jitao Sang. An LLM-free multi-dimensional benchmark for mllms hallucination evaluation. arXiv preprint arXiv:2311.07397, 2023. 1, 2, 3, 6 +[35] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 2 +[36] Qijie Wei, Kaiheng Qian, and Xirong Li. FunBench: Benchmarking fundus reading skills of MLLMs. arXiv preprint arXiv:2503.00901, 2025. 3 +[37] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 2 +[38] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren + +Zhou. mPLUG-Owl2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 6 +[39] Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. Woodpecker: Hallucination correction for multimodal large language models. Science China Information Sciences, 67(12):220105, 2024. 6 +[40] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. MMMU: A massive multi-discipline multimodal understanding and reasoning benchmark for expert AGI. In CVPR, 2024. 2 +[41] Le Zhang, Rabiul Awal, and Aishwarya Agrawal. Contrasting intra-modal and ranking cross-modal hard negatives to enhance visio-linguistic compositional understanding. In CVPR, 2024. 2 +[42] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the AI ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2 +[43] Kankan Zhou, Eason Lai, Wei Bin Au Yeong, Kyriakos Mouratidis, and Jing Jiang. ROME: Evaluating pre-trained vision-language models on reasoning beyond visual common sense. In Findings of the Association for Computational Linguistics: EMNLP, 2023. 2, 3 \ No newline at end of file diff --git a/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/images.zip b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f7007d3c4af6f727cfb9386d3c4a00562f10a26a --- /dev/null +++ b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da47bb5f019060fdd27e4e5ab1f245afa6e740055bd278bff2b688418a62f1d5 +size 775822 diff --git a/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/layout.json b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d97ee140723c3137a3ac4ffc4ddff7ce8e8a4de8 --- /dev/null +++ b/2025/PhD_ A ChatGPT-Prompted Visual Hallucination Evaluation Dataset/layout.json @@ -0,0 +1,8619 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 518, + 121 + ], + "type": "text", + "content": "PhD: A ChatGPT-Prompted Visual hallucination Evaluation Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "spans": [ + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "text", + "content": "Jiazhen Liu" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "text", + "content": ", Yuhan Fu" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "text", + "content": ", Ruobing Xie" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "text", + "content": ", Runquan Xie" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "text", + "content": ", Xingwu Sun" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "text", + "content": ", Fengzong Lian" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "text", + "content": ", Zhanhui Kang" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "text", + "content": ", and Xirong Li" + }, + { + "bbox": [ + 84, + 142, + 525, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 183, + 185, + 427, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 185, + 427, + 213 + ], + "spans": [ + { + "bbox": [ + 183, + 185, + 427, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 183, + 185, + 427, + 213 + ], + "type": "text", + "content": "Renmin University of China \n" + }, + { + "bbox": [ + 183, + 185, + 427, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 183, + 185, + 427, + 213 + ], + "type": "text", + "content": "Machine Learning Platform Department, Tencent" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 216, + 398, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 216, + 398, + 227 + ], + "spans": [ + { + "bbox": [ + 206, + 216, + 398, + 227 + ], + "type": "text", + "content": "https://github.com/jiazhen-code/PhD" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 152, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 281, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 604 + ], + "type": "text", + "content": "Multimodal Large Language Models (MLLMs) hallucinate, resulting in an emerging topic of visual hallucination evaluation (VHE). This paper contributes a ChatGPT-Prompted visual hallucination evaluation Dataset (PhD) for objective VHE at a large scale. The essence of VHE is to ask an MLLM questions about specific images to assess its susceptibility to hallucination. Depending on what to ask (objects, attributes, sentiment, etc.) and how the questions are asked, we structure PhD along two dimensions, i.e. task and mode. Five visual recognition tasks, ranging from low-level (object / attribute recognition) to middle-level (sentiment / position recognition and counting), are considered. Besides a normal visual QA mode, which we term PhD-base, PhD also asks questions with spcious context (PhD-sec) or with incorrect context (PhD-icc), or with AI-generated counter common sense images (PhD-ccs). We construct PhD by a ChatGPT-assisted semi-automated pipeline, encompassing four pivotal modules: task-specific hallucinatory item (hitem) selection, hitem-embedded question generation, spcious / incorrect context generation, and counter-common-sense (CCS) image generation. With over 14k daily images, 750 CCS images and 102k VQA triplets in total, PhD reveals considerable variability in MLLMs' performance across various modes and tasks, offering valuable insights into the nature of hallucination. As such, PhD stands as a potent tool not only for VHE but may also play a significant role in the refinement of MLLMs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 685 + ], + "type": "text", + "content": "Using a specific large language model (LLM) as its kernel, a multi-modal LLM (MLLM), exemplified by LLaVA [25], Qwen-VL [2] and MiniGPT-v2 [4] can now tackle a wide" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 255, + 361, + 294 + ], + "blocks": [ + { + "bbox": [ + 318, + 255, + 361, + 294 + ], + "lines": [ + { + "bbox": [ + 318, + 255, + 361, + 294 + ], + "spans": [ + { + "bbox": [ + 318, + 255, + 361, + 294 + ], + "type": "image", + "image_path": "51573b5c4bc12668482aa70070c9ac37cbe5d12a450d614eacfe6ba138b39e4e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 380, + 256, + 518, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 256, + 518, + 268 + ], + "spans": [ + { + "bbox": [ + 380, + 256, + 518, + 268 + ], + "type": "text", + "content": "Are there five people in the image?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 374, + 273, + 509, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 273, + 509, + 283 + ], + "spans": [ + { + "bbox": [ + 374, + 273, + 509, + 283 + ], + "type": "text", + "content": "Yes, there are five people in the image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 375, + 284, + 523, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 284, + 523, + 293 + ], + "spans": [ + { + "bbox": [ + 375, + 284, + 523, + 293 + ], + "type": "text", + "content": "They are standing near a bunch of bananas." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 535, + 277, + 551, + 291 + ], + "blocks": [ + { + "bbox": [ + 535, + 277, + 551, + 291 + ], + "lines": [ + { + "bbox": [ + 535, + 277, + 551, + 291 + ], + "spans": [ + { + "bbox": [ + 535, + 277, + 551, + 291 + ], + "type": "image", + "image_path": "2680ade224ae3647b4c20cb41abb6ee7eadff0ff185e54d85dd98db0b48659a3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 297, + 548, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 297, + 548, + 307 + ], + "spans": [ + { + "bbox": [ + 321, + 297, + 548, + 307 + ], + "type": "text", + "content": "(a) Hallucination cause I: Visual ambiguity (MLLM: LLaVA-1.5) [25]" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 318, + 314, + 357, + 341 + ], + "blocks": [ + { + "bbox": [ + 318, + 314, + 357, + 341 + ], + "lines": [ + { + "bbox": [ + 318, + 314, + 357, + 341 + ], + "spans": [ + { + "bbox": [ + 318, + 314, + 357, + 341 + ], + "type": "image", + "image_path": "024fb122f8f287ca4e1de86ff2c9b38e902f56630418ed97f91a0bb71c9c373f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 358, + 315, + 418, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 315, + 418, + 322 + ], + "spans": [ + { + "bbox": [ + 358, + 315, + 418, + 322 + ], + "type": "text", + "content": "After a long playtime with" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 358, + 322, + 421, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 322, + 421, + 327 + ], + "spans": [ + { + "bbox": [ + 358, + 322, + 421, + 327 + ], + "type": "text", + "content": "its toy ball, the kitten, feeling" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 359, + 327, + 421, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 327, + 421, + 333 + ], + "spans": [ + { + "bbox": [ + 359, + 327, + 421, + 333 + ], + "type": "text", + "content": "thoroughly exhausted from" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 359, + 333, + 421, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 333, + 421, + 340 + ], + "spans": [ + { + "bbox": [ + 359, + 333, + 421, + 340 + ], + "type": "text", + "content": "thoroughly exhausted from all the fun, has finally found" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 319, + 342, + 418, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 342, + 418, + 361 + ], + "spans": [ + { + "bbox": [ + 319, + 342, + 418, + 361 + ], + "type": "text", + "content": "its resting place. It has curled up on the table, seeking comfort and rest after its energetic adventures." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 425, + 316, + 435, + 330 + ], + "blocks": [ + { + "bbox": [ + 425, + 316, + 435, + 330 + ], + "lines": [ + { + "bbox": [ + 425, + 316, + 435, + 330 + ], + "spans": [ + { + "bbox": [ + 425, + 316, + 435, + 330 + ], + "type": "image", + "image_path": "51556ece3eb4b096708dd3a43d21f7899e6822dabbbdc67f7ffe4356658ed86e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 444, + 318, + 540, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 444, + 318, + 540, + 328 + ], + "spans": [ + { + "bbox": [ + 444, + 318, + 540, + 328 + ], + "type": "text", + "content": "Is there a toy in the image?" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 432, + 336, + 520, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 336, + 520, + 347 + ], + "spans": [ + { + "bbox": [ + 432, + 336, + 520, + 347 + ], + "type": "text", + "content": "Yes, the kitten is playing" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 433, + 347, + 485, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 433, + 347, + 485, + 356 + ], + "spans": [ + { + "bbox": [ + 433, + 347, + 485, + 356 + ], + "type": "text", + "content": "with a toy ball." + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 535, + 343, + 552, + 357 + ], + "blocks": [ + { + "bbox": [ + 535, + 343, + 552, + 357 + ], + "lines": [ + { + "bbox": [ + 535, + 343, + 552, + 357 + ], + "spans": [ + { + "bbox": [ + 535, + 343, + 552, + 357 + ], + "type": "image", + "image_path": "047bc6b5724bb26f26e69b59632039d8b495af91ce1499e71a4bb563e0261b2c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 334, + 363, + 534, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 363, + 534, + 373 + ], + "spans": [ + { + "bbox": [ + 334, + 363, + 534, + 373 + ], + "type": "text", + "content": "(b) Hallucination cause II: Inconsistency in multi-modal input" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 318, + 380, + 362, + 420 + ], + "blocks": [ + { + "bbox": [ + 318, + 380, + 362, + 420 + ], + "lines": [ + { + "bbox": [ + 318, + 380, + 362, + 420 + ], + "spans": [ + { + "bbox": [ + 318, + 380, + 362, + 420 + ], + "type": "image", + "image_path": "87bfd678fdbcb85444a4ba5ba0352e9f3533e3ea7e5e88b8dfb26f8286bbc7da.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 369, + 384, + 380, + 400 + ], + "blocks": [ + { + "bbox": [ + 369, + 384, + 380, + 400 + ], + "lines": [ + { + "bbox": [ + 369, + 384, + 380, + 400 + ], + "spans": [ + { + "bbox": [ + 369, + 384, + 380, + 400 + ], + "type": "image", + "image_path": "6ad6cbdd5c082aceb923b355560b9b8013bdcab5928fb862d4394cf0058bb7b2.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 392, + 382, + 483, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 382, + 483, + 401 + ], + "spans": [ + { + "bbox": [ + 392, + 382, + 483, + 401 + ], + "type": "text", + "content": "Is the mouse in the image smaller than the cat?" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 394, + 407, + 524, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 407, + 524, + 418 + ], + "spans": [ + { + "bbox": [ + 394, + 407, + 524, + 418 + ], + "type": "text", + "content": "Yes, the mouse is smaller than the cat." + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 534, + 407, + 551, + 419 + ], + "blocks": [ + { + "bbox": [ + 534, + 407, + 551, + 419 + ], + "lines": [ + { + "bbox": [ + 534, + 407, + 551, + 419 + ], + "spans": [ + { + "bbox": [ + 534, + 407, + 551, + 419 + ], + "type": "image", + "image_path": "68b0e82990025d2b9b673fa14738b3ab16ee1bd1f8d172f1dec2c0b9f6e4314e.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 339, + 423, + 530, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 423, + 530, + 433 + ], + "spans": [ + { + "bbox": [ + 339, + 423, + 530, + 433 + ], + "type": "text", + "content": "(c) Hallucination cause III: Counter-common-sense content" + } + ] + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 318, + 441, + 548, + 578 + ], + "blocks": [ + { + "bbox": [ + 318, + 441, + 548, + 578 + ], + "lines": [ + { + "bbox": [ + 318, + 441, + 548, + 578 + ], + "spans": [ + { + "bbox": [ + 318, + 441, + 548, + 578 + ], + "type": "image", + "image_path": "cbf98828d50186939dd14d9a1b6540911aad192ae7bca33e1ccbcc632766c070.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 315, + 581, + 552, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 581, + 552, + 601 + ], + "spans": [ + { + "bbox": [ + 315, + 581, + 552, + 601 + ], + "type": "text", + "content": "(d) Performance curves of the LLaVA series on two public VHE datasets (POPE [21]) and AMBER [34]) and the proposed PhD dataset." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 609, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 555, + 686 + ], + "type": "text", + "content": "Figure 1. Illustrations of three major causes of an MLLM's visual hallucination and its evaluation. This paper contributes PhD, a binary VQA-based VHE benchmark, much larger and more challenging than its predecessors. In particular, it has four evaluation modes that explicitly measure an MLLM's performance w.r.t. the three causes, i.e. PhD-base for cause I, PhD-sec and PhD-icc for cause II and PhD-ccs for cause III." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 225, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 225, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 225, + 703 + ], + "type": "text", + "content": "*J. Liu is currently pursuing his PhD at HKUST." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 67, + 703, + 211, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 211, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 211, + 712 + ], + "type": "text", + "content": "† Corresponding author (xirong@ruc.edu.cn)" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19857" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": "range of computer vision tasks in a unified visual-question-answering (VQA) manner. As LLMs are known to hallucinate [20, 37, 42], it is not surprising that MLLMs have visual hallucination, generating fabricated interpretation of the given visual content, see Fig. 1. Considering the rapidly growing use of MLLMs in varied scenarios, a comprehensive visual hallucination evaluation (VHE) is crucial. This paper develops a new dataset for VHE." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 168, + 297, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 168, + 297, + 396 + ], + "spans": [ + { + "bbox": [ + 57, + 168, + 297, + 396 + ], + "type": "text", + "content": "VHE essentially involves posing a number of visual questions to an MLLM [11, 21]. A question of this kind shall include a hallucinatory item (hitem), in the form of a specific word or phrase, that induces the MLLM to generate a response discordant with the provided visual content. As the model typically has a strong visual recognition ability, how to identify an appropriate hitem and accordingly generate a proper question is nontrivial. Both the hitem and the question depend on the visual recognition task being considered. As shown in Tab. 1, we target at objective VHE in the context of low-level (object / attribute) to middle-level (sentiment / position / counting) visual recognition tasks. Such a target is chosen due to the following considerations. MLLMs generally work well for these tasks, so their erroneous responses can be largely attributed to their hallucinations instead of their incapability, e.g. asking a generic MLLM to read pathology images [14]. Meanwhile, a binary VQA based objective evaluation is more budget friendly and thus more suited for VHE at a large scale." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 407, + 296, + 505 + ], + "blocks": [ + { + "bbox": [ + 56, + 407, + 296, + 505 + ], + "lines": [ + { + "bbox": [ + 56, + 407, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 296, + 505 + ], + "type": "table", + "html": "
Vision tasksObjective evaluationSubjective evaluation
Low-/middle-level visual recognition+ POPE, EMNLP'23 [21]+ FAITHSCORE, EMNLP'24 [15]
+ AMBER, arXiv'23 [34]+ HaELM, arXiv'23 [35]
+ CIEM, ITIF'23 [13]+ M-HalDetect, AAAI'24 [12]
+ NOPE, ALVIR'24 [27]+ GAVIE, ICLR'24 [23]
+ ROME, EMNLP'23 [43]
+ PhD (this paper)
High-level visual reasoning+ MMMU, CVPR'24 [40]+ HallusionBench, CVPR'24 [11]
+ VLind-Bench, NAACL'25 [17]+ Bingo, arXiv'23 [6]
+ IllusionVQA, COLM'24 [31]
+ WHOOPS!, ICCV'23[3]
", + "image_path": "82205532c55138e48e80e1268da60463ee8dc39231e3ff79af6e666eb58028a2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Pioneered by POPE [21], good attempts exist in objective VHE [13, 27, 34, 43]. In these valuable datasets, hitem selection is largely untouched. POPE and ROME [43] select their hitems fully based on label co-occurrence in training data, AMBER [34] relies on manual annotation, whilst hitem selection is not considered in NOPE [27] and CIEM [13], see Tab. 2. Hence, there lacks an explicit link between hitem selection (and subsequent VQA triplets construction) and major causes of an MLLM's visual hallucination. As models rapidly evolve, the performance on these datasets quickly reaches saturation, see Fig. 1d." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": "Analyzing an MLLM's typical dataflow of answering a visual question, we see three major causes of visual hallucination: I) visual ambiguity, II) inconsistency in multi-modal input and III) counter-common-sense (CCS) content, see Fig. 1. Firstly, the MLLM extracts tokenized visual features from a given image using a ViT based encoder. Recent studies [33, 41] show that the features tend to be high level, lacking sufficient details for fine-grained tasks such as counting. Secondly, the visual features, after vision-to-language adaptation, are mixed with the features of the associated textual prompt to form a multi-modal input to the LLM kernel. The LLM, pre-trained extensively on textual data, inevitably favors the textual part of the multi-modal input. Hence, when there is inconsistency between the visual and textual information, the former is more likely to be overruled. Lastly, at the decoding stage, the LLM might heavily rely on its internal (world) knowledge, ignoring the visual content especially when the content (showing a mouse much larger than a cat) contradicts the common sense. Our new dataset is developed with a close link to the three causes." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 314, + 320, + 555, + 394 + ], + "blocks": [ + { + "bbox": [ + 55, + 512, + 295, + 567 + ], + "lines": [ + { + "bbox": [ + 55, + 512, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 295, + 567 + ], + "type": "text", + "content": "Table 1. Taxonomy of VHE benchmarks. Our PhD benchmark performs an objective evaluation of MLLMs' hallucinations when they address visual recognition tasks ranging from low-level, i.e. object / attribute recognition to middle-level, i.e. sentiment / positional recognition and counting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 320, + 555, + 394 + ], + "lines": [ + { + "bbox": [ + 314, + 320, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 314, + 320, + 555, + 394 + ], + "type": "table", + "html": "
DatasetDaily imagesCCS imagesHItemsContextsVQA tripletsTasks
POPE500✘80✘3,000Obj.
NOPE*unknown✘unknown✘32,701Obj.
CIEM*4,929✘unknown✘72,941Obj. / Attr. / Pos.
AMBER1,004✘687✘14,216Obj. / Attr./Pos. / Count.
ROME✘1,563118✘1,563Attr./Pos.
PhD14,6487501,45233,688102,564Obj / Attr. / Pos. / Sent. / Count.
", + "image_path": "5aa706bb2c4620c69e58c0f787816ca90c73fe8ce9eceb219a6ec25eab031356.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 403, + 553, + 415 + ], + "lines": [ + { + "bbox": [ + 314, + 403, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 314, + 403, + 553, + 415 + ], + "type": "text", + "content": "Table 2. PhD versus its predecessors. * indicates private dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 426, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 677 + ], + "type": "text", + "content": "The new dataset is constructed by adapting TDIUC, a popular multi-task VQA dataset [16], with a ChatGPT-assisted semi-automated pipeline, see Fig. 2. In particular, by prompting ChatGPT, we select diverse and visually challenging h items in an image-specific and task-specific manner, with minimal human involvement primarily spent on verifying ChatGPT-generated results. The selected h items are then automatically embedded into visual questions, specious context, and incorrect context, all generated by instructing ChatGPT. Moreover, we expand the daily image set with counter common sense (CCS) images, obtained by prompting AIGC tools with ChatGPT-generated CCS descriptions, e.g. \"trees growing underwater\" and \"a car with square-shaped wheels\". The dataset is dubbed as PhD (ChatGPT Prompted visual hallucination evaluation Dataset). Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base (questions about daily images w/o context), PhD-sec (PhD-base plus specious context), PhD-icc (PhD-base plus incorrect context), and PhD-ccs (questions about CCS images)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 678, + 531, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 678, + 531, + 689 + ], + "spans": [ + { + "bbox": [ + 328, + 678, + 531, + 689 + ], + "type": "text", + "content": "To sum up, our major contributions are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "- We introduce PhD, a dataset with four evaluation modes across five visual recognition tasks, developed with a close" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19858" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "link to the three major causes of MLLM visual hallucination. Information on hallucinatory items (hItems) is provided per sample, enabling in-depth analytics to uncover the causes in more detail." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 120, + 294, + 275 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 120, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 294, + 179 + ], + "type": "text", + "content": "- We offer a ChatGPT-assisted semi-automatic pipeline for dataset construction, with minimal human involvement, primarily focused on verifying the generated results. With 14,648 daily images, 750 CCS images and 102,564 VQA triplets in total, PhD is the largest of its kind." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 180, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 180, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 180, + 294, + 275 + ], + "type": "text", + "content": "- We conduct an extensive evaluation with 15 open-source MLLMs, 3 proprietary MLLMs, and 2 hallucination mitigation methods, showing the viability of PhD for VHE in varied manners including overall, mode-oriented, task-oriented, and model-wise zoom-in. The evaluation not only reveals inter-model performance divergence, but may also help developers of a specific MLLM to prioritize their efforts in refining the model." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 286, + 142, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 142, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 142, + 298 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 307, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 307, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 56, + 307, + 295, + 582 + ], + "type": "text", + "content": "Due to the increasing importance of VHE, new benchmarks are being actively developed. Depending on what vision tasks they focus on and how their evaluation is executed, we categorize existing achievements along two dimensions, i.e. task and evaluation, see Tab. 1. Concerning the task dimension, low- and middle-level visual recognition tasks, ranging from object / attribute recognition to sentiment / positional recognition and counting, assess an MLLM's basic visual skills. High-level visual reasoning is more domain-knowledge intensive, typically covering image-based math question solving, geography information understanding, meme interpretation, historical or folkloric contexts, etc. [3, 6, 11, 17]. As for the evaluation dimension, objective evaluation refers to objectively comparing the model's output with ground truth, mostly in the form of Yes/No answers [13, 21, 34, 43]. By contrast, subjective evaluation requires humans or LLMs to assess the model's output [3, 6, 11]. The proposed PhD, focusing on low/middle-level visual recognition and objective evaluation, belongs to the second quadrant of the taxonomy. In what follows, we briefly review peer benchmarks, i.e. POPE [21], ROME [43], NOPE [27], CIEM [13], and AMBER [34], in this quadrant and clarify our novelty accordingly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "POPE is probably the first dataset to evaluate object hallucination [21]. Given a specific MS-COCO image with object labels, POPE selects an adversarial object frequently co-occurring with the current labels. A binary question is then formed by filling out a predefined template with the selected object. Such co-occurrence based hitem selection is not image-specific by definition. Trivial hitems might thus be picked up, e.g. \"car\" selected for an indoor image labeled with \"person\", as the two objects often co-occur. For CCS image generation, ROME forms CCS descriptions by choosing attribute values having the lowest co-occurrence" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "with a given object according to the Visual Genome dataset [43]. As low occurrence is not necessarily CCS, some of ROME images are indeed normal. NOPE [27] and CIEM [13] simply bypass hitem selection by asking an LLM to generate questions conditioned on the image captions (also from MS-COCO) and pre-specified answers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 144, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 204 + ], + "type": "text", + "content": "To select h items in an image-specific manner, AMBER resorts to fully manual annotation [34]. Manual labeling is costly, while an annotator's personal vocabulary is relatively limited. All this makes it difficult to scale up w.r.t. the amount of test images and the number of distinct h items." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "content": "In comparison, PhD, constructed by a ChatGPT-assisted semi-automatic pipeline (Sec. 3), is much larger (Tab. 2) and more challenging (Fig. 1d). With its unique mode-task structure, the new dataset enables a novel, structured and zoom-in understanding of inter-model difference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 276, + 443, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 276, + 443, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 276, + 443, + 289 + ], + "type": "text", + "content": "3. Our Roadmap to PhD" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 297, + 553, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 553, + 475 + ], + "type": "text", + "content": "As MLLMs perform visual recognition in a VQA manner, a VQA sample for VHE naturally depends on the visual recognition task in consideration. We depart from TDIUC [16], a large-scale VQA dataset w.r.t. five tasks including object / attribute / sentiment / positional recognition and counting. Note that the images in TDIUC are sourced from MS-COCO [22], which plausibly have been seen by MLLMs in their development stage. As such, our adoption of TDIUC makes on-purpose data leakage: an MLLM's erroneous response w.r.t. a seen image can now be more safely attributed to its hallucination other than its incapability in visual recognition, say asking LLaVA to recognize glaucoma from color fundus photographs [36]. We construct PhD by adapting the TDIUC annotations with a ChatGPT1-assisted semi-automated pipeline, see Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 476, + 553, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 553, + 644 + ], + "type": "text", + "content": "In order to compose a proper question that effectively makes an MLLM hallucinate about a given image, a hitem has to be first identified in an image-specific and task-specific manner. Then, the hitem has to be smoothly embedded in the form of a specific word or phrase into the question. We describe task-specific hitem selection in Sec. 3.1, followed by hitem-embedded question generation in Sec. 3.2. Furthermore, in order to simulate inconsistency in the multi-modal input, we prepend specious or incorrect context to the question, the generation of which is detailed in Sec. 3.3. Lastly, in order to explicitly create conflicts between the visual input and the internal (world) knowledge of the MLLM, we expand our image collection with autogenerated counter-common-sense (CCS) images (Sec. 3.4)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 652, + 474, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 474, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 474, + 664 + ], + "type": "text", + "content": "3.1. Task-specific Hitem Selection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 670, + 553, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 553, + 693 + ], + "type": "text", + "content": "Without loss of generality, we describe how a hitem is selected for color attribute recognition. Let us consider the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 479, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 479, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 479, + 712 + ], + "type": "text", + "content": "1We use GPT-4o mini, released on 2024-05-13." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19859" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 82, + 403, + 129 + ], + "blocks": [ + { + "bbox": [ + 63, + 73, + 146, + 80 + ], + "lines": [ + { + "bbox": [ + 63, + 73, + 146, + 80 + ], + "spans": [ + { + "bbox": [ + 63, + 73, + 146, + 80 + ], + "type": "text", + "content": "(1) Task-specific hitem selection" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 82, + 403, + 129 + ], + "lines": [ + { + "bbox": [ + 62, + 82, + 403, + 129 + ], + "spans": [ + { + "bbox": [ + 62, + 82, + 403, + 129 + ], + "type": "image", + "image_path": "646c6dc4e7a35c48032ead0b4d1db978eb5c5caa8db07719269bf1c711925893.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 144, + 403, + 201 + ], + "blocks": [ + { + "bbox": [ + 63, + 133, + 170, + 141 + ], + "lines": [ + { + "bbox": [ + 63, + 133, + 170, + 141 + ], + "spans": [ + { + "bbox": [ + 63, + 133, + 170, + 141 + ], + "type": "text", + "content": "(2) Hitem-embedded question generation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 144, + 403, + 201 + ], + "lines": [ + { + "bbox": [ + 62, + 144, + 403, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 144, + 403, + 201 + ], + "type": "image", + "image_path": "6a1fb24641db27d7e7c817515a0c3e9246c98bc3f0d50ba9cb3edf4792af3e5d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 202, + 359, + 213 + ], + "lines": [ + { + "bbox": [ + 104, + 202, + 359, + 213 + ], + "spans": [ + { + "bbox": [ + 104, + 202, + 359, + 213 + ], + "type": "text", + "content": "(a) Hitem-embedded question / context generation for MS-COCO daily images" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 409, + 83, + 550, + 201 + ], + "blocks": [ + { + "bbox": [ + 409, + 83, + 550, + 201 + ], + "lines": [ + { + "bbox": [ + 409, + 83, + 550, + 201 + ], + "spans": [ + { + "bbox": [ + 409, + 83, + 550, + 201 + ], + "type": "image", + "image_path": "4bf9d8d0909bb5ec79279fc3b476c78daea8d7ea2907d374a352ab0d79dd4389.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 202, + 542, + 213 + ], + "lines": [ + { + "bbox": [ + 417, + 202, + 542, + 213 + ], + "spans": [ + { + "bbox": [ + 417, + 202, + 542, + 213 + ], + "type": "text", + "content": "(b) QAs for AI-generated CCS images" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 61, + 224, + 136, + 281 + ], + "blocks": [ + { + "bbox": [ + 61, + 224, + 136, + 281 + ], + "lines": [ + { + "bbox": [ + 61, + 224, + 136, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 136, + 281 + ], + "type": "image", + "image_path": "a20f5410384ac7b3cacd15ece0e2ac60f70ab590b49b5a2ddf9a0be4e0eda7b3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 281, + 110, + 288 + ], + "lines": [ + { + "bbox": [ + 63, + 281, + 110, + 288 + ], + "spans": [ + { + "bbox": [ + 63, + 281, + 110, + 288 + ], + "type": "text", + "content": "Task: Object recognition" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 289, + 132, + 294 + ], + "lines": [ + { + "bbox": [ + 65, + 289, + 132, + 294 + ], + "spans": [ + { + "bbox": [ + 65, + 289, + 132, + 294 + ], + "type": "text", + "content": "Are there headphones in the image?" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 135, + 226, + 230, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 226, + 230, + 262 + ], + "spans": [ + { + "bbox": [ + 135, + 226, + 230, + 262 + ], + "type": "text", + "content": "In recent years, this musician has often immersed himself in his own musical world, playing the guitar with headphones on as if cut off from the outside world. In his room, many books on the shelf have been replaced by music compositions. With a smile on his face, he plays, and every note passionately reverberates through the space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 135, + 263, + 230, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 263, + 230, + 300 + ], + "spans": [ + { + "bbox": [ + 135, + 263, + 230, + 300 + ], + "type": "text", + "content": "In the photo, a man confidently plays his guitar with headphones on, fully immersed in his music. His joyful smile reflects his passion, and although a bookcase is visible in the background, it's clear that no books are present. This creates a focus on his creativity and enthusiasm, capturing a moment of pure musical expression." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 234, + 224, + 402, + 304 + ], + "blocks": [ + { + "bbox": [ + 234, + 224, + 402, + 304 + ], + "lines": [ + { + "bbox": [ + 234, + 224, + 402, + 304 + ], + "spans": [ + { + "bbox": [ + 234, + 224, + 402, + 304 + ], + "type": "image", + "image_path": "bbd20cbdfed5eade34256728b8a4b1d4e7689ed698daa4ce36e60cf75b07c54c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 61, + 307, + 135, + 352 + ], + "blocks": [ + { + "bbox": [ + 65, + 295, + 124, + 300 + ], + "lines": [ + { + "bbox": [ + 65, + 295, + 124, + 300 + ], + "spans": [ + { + "bbox": [ + 65, + 295, + 124, + 300 + ], + "type": "text", + "content": "Are there books in the image?" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 307, + 135, + 352 + ], + "lines": [ + { + "bbox": [ + 61, + 307, + 135, + 352 + ], + "spans": [ + { + "bbox": [ + 61, + 307, + 135, + 352 + ], + "type": "image", + "image_path": "6a3c223e92b7b95afad980d10c040a3748a98106737e438852ad81adab4c6a36.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 353, + 117, + 358 + ], + "lines": [ + { + "bbox": [ + 63, + 353, + 117, + 358 + ], + "spans": [ + { + "bbox": [ + 63, + 353, + 117, + 358 + ], + "type": "text", + "content": "Task: Sentiment recognition" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 358, + 135, + 364 + ], + "lines": [ + { + "bbox": [ + 63, + 358, + 135, + 364 + ], + "spans": [ + { + "bbox": [ + 63, + 358, + 135, + 364 + ], + "type": "text", + "content": "Is the girl feeling lonely in the image?" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 364, + 134, + 370 + ], + "lines": [ + { + "bbox": [ + 63, + 364, + 134, + 370 + ], + "spans": [ + { + "bbox": [ + 63, + 364, + 134, + 370 + ], + "type": "text", + "content": "Is the girl feeling happy in the image?" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 308, + 230, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 308, + 230, + 338 + ], + "spans": [ + { + "bbox": [ + 135, + 308, + 230, + 338 + ], + "type": "text", + "content": "The girl is often surrounded by her beloved teddy bears, yet she regularly feels a sense of loneliness despite their company. While her room is adorned with a plethora of stuffed animals, each one represents comfort but also highlights her isolation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 135, + 338, + 229, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 338, + 229, + 371 + ], + "spans": [ + { + "bbox": [ + 135, + 338, + 229, + 371 + ], + "type": "text", + "content": "The girl in the image is feeling lonely, despite being surrounded by a large group of teddy bears. She lies on the bed near many stuffed animals, including dozens of brown teddy bears, and her contrast between her solitary emotions and the cheerful presence of her furry companions." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 234, + 307, + 308, + 342 + ], + "blocks": [ + { + "bbox": [ + 234, + 307, + 308, + 342 + ], + "lines": [ + { + "bbox": [ + 234, + 307, + 308, + 342 + ], + "spans": [ + { + "bbox": [ + 234, + 307, + 308, + 342 + ], + "type": "image", + "image_path": "baa837148544a3bed93a35352ba8aa6b0d125a73da2247590bfc4c748b3f6e36.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 342, + 266, + 347 + ], + "lines": [ + { + "bbox": [ + 235, + 342, + 266, + 347 + ], + "spans": [ + { + "bbox": [ + 235, + 342, + 266, + 347 + ], + "type": "text", + "content": "Task: Positional" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 348, + 292, + 353 + ], + "lines": [ + { + "bbox": [ + 249, + 348, + 292, + 353 + ], + "spans": [ + { + "bbox": [ + 249, + 348, + 292, + 353 + ], + "type": "text", + "content": "Is there a table behind" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 235, + 354, + 301, + 359 + ], + "lines": [ + { + "bbox": [ + 235, + 354, + 301, + 359 + ], + "spans": [ + { + "bbox": [ + 235, + 354, + 301, + 359 + ], + "type": "text", + "content": "the looking dog in the image?" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 243, + 360, + 299, + 365 + ], + "lines": [ + { + "bbox": [ + 243, + 360, + 299, + 365 + ], + "spans": [ + { + "bbox": [ + 243, + 360, + 299, + 365 + ], + "type": "text", + "content": "Is there a fence behind" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 243, + 365, + 299, + 371 + ], + "lines": [ + { + "bbox": [ + 243, + 365, + 299, + 371 + ], + "spans": [ + { + "bbox": [ + 243, + 365, + 299, + 371 + ], + "type": "text", + "content": "the looking dog in the image?" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 309, + 308, + 402, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 308, + 402, + 338 + ], + "spans": [ + { + "bbox": [ + 309, + 308, + 402, + 338 + ], + "type": "text", + "content": "A large brown dog has comfortably settled onto a stack of three plastic chairs, creating an amusing scene as it relaxes in the sunny outdoor space. Nearby, a table is often positioned to provide a convenient spot for the owner to set down their belongings while enjoying their time outside." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 309, + 338, + 402, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 338, + 402, + 371 + ], + "spans": [ + { + "bbox": [ + 309, + 338, + 402, + 371 + ], + "type": "text", + "content": "In a charming scene, a dog has made itself comfortable on a stack of plastic chairs, effortlessly showcasing its playful nature. Behind the relaxed dog captured in the photo, a table can be filled with a large box of cozy atmosphere. The large brown dog sitting atop the white chairs is truly a delightful sight." + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 410, + 241, + 477, + 291 + ], + "blocks": [ + { + "bbox": [ + 410, + 241, + 477, + 291 + ], + "lines": [ + { + "bbox": [ + 410, + 241, + 477, + 291 + ], + "spans": [ + { + "bbox": [ + 410, + 241, + 477, + 291 + ], + "type": "image", + "image_path": "c9cd613d5a37f9da3944ee086a421297a86ef13c5cd1187109a1bff7ef2e359b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 292, + 477, + 297 + ], + "lines": [ + { + "bbox": [ + 411, + 292, + 477, + 297 + ], + "spans": [ + { + "bbox": [ + 411, + 292, + 477, + 297 + ], + "type": "text", + "content": "Is the ball being kicked a soccer ball?" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 297, + 477, + 303 + ], + "lines": [ + { + "bbox": [ + 411, + 297, + 477, + 303 + ], + "spans": [ + { + "bbox": [ + 411, + 297, + 477, + 303 + ], + "type": "text", + "content": "Is the ball being kicked a basketball?" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 482, + 242, + 549, + 290 + ], + "blocks": [ + { + "bbox": [ + 482, + 242, + 549, + 290 + ], + "lines": [ + { + "bbox": [ + 482, + 242, + 549, + 290 + ], + "spans": [ + { + "bbox": [ + 482, + 242, + 549, + 290 + ], + "type": "image", + "image_path": "7df577b988cb09d7417eb469eacf4517b97d91a35d1b4eaca4d6886cf878a60b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 292, + 546, + 297 + ], + "lines": [ + { + "bbox": [ + 484, + 292, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 484, + 292, + 546, + 297 + ], + "type": "text", + "content": "Is the mouse smaller than the cat?" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 410, + 309, + 477, + 350 + ], + "blocks": [ + { + "bbox": [ + 410, + 309, + 477, + 350 + ], + "lines": [ + { + "bbox": [ + 410, + 309, + 477, + 350 + ], + "spans": [ + { + "bbox": [ + 410, + 309, + 477, + 350 + ], + "type": "image", + "image_path": "aeece98e4fc0c4739e290283298a8866bb334f01e7c7b489bf23def3ae8aa6fa.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 350, + 463, + 355 + ], + "lines": [ + { + "bbox": [ + 424, + 350, + 463, + 355 + ], + "spans": [ + { + "bbox": [ + 424, + 350, + 463, + 355 + ], + "type": "text", + "content": "Is the highest number" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 356, + 466, + 361 + ], + "lines": [ + { + "bbox": [ + 424, + 356, + 466, + 361 + ], + "spans": [ + { + "bbox": [ + 424, + 356, + 466, + 361 + ], + "type": "text", + "content": "of points on the dice six?" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 361, + 463, + 365 + ], + "lines": [ + { + "bbox": [ + 424, + 361, + 463, + 365 + ], + "spans": [ + { + "bbox": [ + 424, + 361, + 463, + 365 + ], + "type": "text", + "content": "Is the highest number" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 365, + 468, + 371 + ], + "lines": [ + { + "bbox": [ + 424, + 365, + 468, + 371 + ], + "spans": [ + { + "bbox": [ + 424, + 365, + 468, + 371 + ], + "type": "text", + "content": "points on the dice seven?" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 482, + 309, + 549, + 358 + ], + "blocks": [ + { + "bbox": [ + 484, + 297, + 546, + 303 + ], + "lines": [ + { + "bbox": [ + 484, + 297, + 546, + 303 + ], + "spans": [ + { + "bbox": [ + 484, + 297, + 546, + 303 + ], + "type": "text", + "content": "Is the cat smaller than the mouse?" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 482, + 309, + 549, + 358 + ], + "lines": [ + { + "bbox": [ + 482, + 309, + 549, + 358 + ], + "spans": [ + { + "bbox": [ + 482, + 309, + 549, + 358 + ], + "type": "image", + "image_path": "0c79a4a059a2df4dee4322880a54350267ee5874b09b19e5c68c90b2b52f62ab.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 360, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 484, + 360, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 484, + 360, + 545, + 365 + ], + "type": "text", + "content": "Is the shape of the orange round?" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 484, + 365, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 484, + 365, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 484, + 365, + 545, + 371 + ], + "type": "text", + "content": "Is the shape of the orange square?" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 373, + 547, + 384 + ], + "lines": [ + { + "bbox": [ + 413, + 373, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 413, + 373, + 547, + 384 + ], + "type": "text", + "content": "(d) Showcase: CCS images and questions" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 393, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 393, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 555, + 460 + ], + "type": "text", + "content": "Figure 2. Proposed semi-automatic pipeline for PhD construction. We use ChaptGPT (GPT-4o mini) to generate hitem-embedded questions / contexts for daily images, and Doubao and DALL-E3 for generating CCS images. Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base, i.e. questions about daily images w/o context, PhD-sec, i.e. PhD-base plus specious context, PhD-icc, i.e. PhD-base plus incorrect context, and PhD-ccs, i.e. questions about CCS images. By adapting TDIUC annotations, PhD supports binary VQA w.r.t. five visual recognition tasks including object / attribute / sentiment / positional recognition and counting. With 20 mode-task combinations in total, PhD enables a comprehensive VHE." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 481, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 528 + ], + "type": "text", + "content": "image in Fig. 2a, showing a black motorcycle followed by a red bus. As no red motorcycle is present while the red color is prominent near the black motorcycle, the word red will be a good choice of hitem to challenge an MLLM." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 55, + 530, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 295, + 601 + ], + "type": "text", + "content": "Vocabulary Construction per task. Started with a handful of manually specified colors such as red, green, and blue, we ask ChatGPT to expand the color vocabulary with instructions like \"Please expand the input vocabulary as much as possible by adding common items found in daily life. Avoid any duplication\", getting a set of 35 different colors." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "text", + "content": "Subject-Attribute Extraction. The image is associated with a TDIUC question-answer pair as \"what color is the motorcycle\" and \"black\". We use ChatGPT (with simple instructions) to extract with ease the subject (i.e. motorcycle) and its attribute (i.e. black) from the pair." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 55, + 664, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 295, + 700 + ], + "type": "text", + "content": "Candidate Hitem Generation. We obtain candidate h items by excluding the ground-truth (GT) answer (and its synonyms if applicable) from the vocabulary." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Visual-based Hitem Ranking. Intuitively, a hitem shall" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 481, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 553, + 601 + ], + "type": "text", + "content": "be visually plausible in the given image. So for each candidate hitem, we compute its similarity to the image using a pre-trained CLIP [30]. In particular, the cosine similarity between the CLIP embeddings of hitem + subject (e.g. green motorcycle) and the image is adopted. Ranking the candidates by the CLIP similarity lets us to select the one visually closest to the image. It is worth noting that for emerging MLLMs equipped with stronger vision encoders [19], our pipeline is likely to produce even more effective h items by replacing CLIP with these advanced counterparts." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 603, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 553, + 675 + ], + "type": "text", + "content": "Manual Inspection. While the above process is generally stable to produce satisfying results, manual inspection is performed to ensure the correctness of hitem selection. Note that due to errors in the original TDIUC annotations, occasionally the true label might be \"incorrectly\" selected. In such a case, we simply discard the VQA sample." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "type": "text", + "content": "With lightweight task-specific adaptation, the above hitem selection also works for other tasks. Overall, the joint use of ChatGPT and CLIP allows us to select 1,452 hitem" + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19860" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "that are more diverse and challenging than their counterparts in previous datasets, see Tab. 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 260, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 260, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 260, + 119 + ], + "type": "text", + "content": "3.2. Hitem-embedded Question Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 125, + 296, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 125, + 296, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 125, + 296, + 210 + ], + "type": "text", + "content": "For a given subject (e.g. motorcycle) and a chosen hitem (e.g. red), generating a hitem-embedded question is trivial for ChatGPT. In particular, a No question is formed as \"Is the motorcycle in the image red?\" Meanwhile, a Yes question is simultaneously generated using the GT as \"Is the motorcycle in the image black?\". This ensures perfect Yes/No balance among the generated questions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 220, + 269, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 269, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 269, + 232 + ], + "type": "text", + "content": "3.3. Specious (Incorrect) Context Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 238, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 380 + ], + "type": "text", + "content": "When used as a document parser, an MLLM reading a specific image is often provided with the image's surrounding text. Inconsistency between the image and the text is not uncommon. A news article containing a general claim of \"red motorcycles frequently zip through the streets\" does not necessarily have each of its illustrated pictures match with the claim. To simulate such a scenario, for a given image we generate specious text as specious context and text contradicting the image as incorrect context, respectively. Next, we describe the generation of specious contexts, as their incorrect counterparts can be generated in a similar but more simplified manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 383, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 502 + ], + "type": "text", + "content": "Specious Text Generation. Using the previously generated hitem-embedded question and the original MS-COCO captions as input, we instruct ChatGPT to generate specious text for a given image. By \"specious\", we mean the text is specious or noisy, rather than directly contradicting the image content. As such, our instruction reads partially as \"Please generate the for the given question. It should be one sentence. The should answer the question, but it may not reflect the actual current status, thus making it specious.\"" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 503, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 552 + ], + "type": "text", + "content": "Text Composition. ChatGPT is used to seamlessly merge the specious text with ground-truth captions, forming a longer context in which only a small portion (orange text in Fig. 2c) is mildly inconsistent with the image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 296, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 577 + ], + "type": "text", + "content": "Manual Inspection. We perform a spot check. If the context quality is low, we simply discard the entire sample." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 586, + 188, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 188, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 188, + 600 + ], + "type": "text", + "content": "3.4. CCS Image Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "content": "We generate CCS images by first generating CCS descriptions and then employing Text2Image tools to convert the descriptions to CCS images, see Fig. 2b." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "CCS Description Generation. A number of manually written task-specific samples, see Tab. 3, are used as in-context learning samples for ChatGPT to generate more descriptions. The descriptions have to be visually expressible, so bad cases like \"the more you eat, the thinner you get\" are filtered out manually. For each CCS text, its common-sense" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "(CS) counterpart is simultaneously generated by ChatGPT, by providing the learning samples in pair." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 97, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 555, + 180 + ], + "type": "text", + "content": "Text2Image. The generated CCS descriptions are used as prompts for AIGC tools (Doubao [9] and DALL-E3 [28]) to generate the corresponding CCS images. The quality of the generated images depends on various factors, making occasional failures inevitable. When this occurs, we attempt to refine the prompts or apply region-based inpainting. The sample will be discarded if the above attempts fail." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 181, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 253 + ], + "type": "text", + "content": "Question Generation. Per CCS description (e.g. A car with square wheels), we utilize ChatGPT to generate a Yes question (e.g. Does the car have square wheels?). Again, for balancing Yes/No questions, we generate a No question (e.g. Does the car have round wheels?) based on the CS description." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 315, + 262, + 555, + 539 + ], + "blocks": [ + { + "bbox": [ + 315, + 262, + 555, + 539 + ], + "lines": [ + { + "bbox": [ + 315, + 262, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 315, + 262, + 555, + 539 + ], + "type": "table", + "html": "
CCS descriptionYes questionCS descriptionNo question
Task: Object recognition
Manually written:
Ice blocks in volcanic lavaAre there ice blocks in volcanic lava?Fire in volcanic lavaIs there fire in volcanic lava?
Grass in a tiger's mouthIs there grass in a tiger's mouth?Meat in the tiger's mouthIs there meat in the tiger's mouth?
ChatGPT generated:
Trees growing underwaterAre there trees growing underwater?Coral growing underwaterIs there coral growing underwater?
Books in a swimming poolAre there books in a swimming pool?Water in a swimming poolIs there water in a swimming pool?
Birds flying underwaterAre there birds flying underwater?Birds flying in the skyAre there birds flying in the sky?
Ice cream in a volcanoIs there ice cream in a volcano?Lava in a volcanoIs there lava in a volcano?
Computers in a forestAre there computers in a forest?Animals in a forestAre there animals in a forest?
Task: Attribute recognition
Manually written:
A car with square wheelsDoes the car have square wheels?A car with round wheelsDoes the car have round wheels?
Blue apples on the treeAre the apples on the tree blue?Red apples on the tree.Are the apples on the tree red?
ChatGPT generated:
A green skyIs the sky green?A blue skyIs the sky blue?
A bicycle with square wheelsDoes the bicycle have square wheels?A bicycle with round wheelsDoes the bicycle have square wheels?
A tree made of metalIs this tree made of metal?A wooden treeIs this tree a real wood tree?
A chocolate riverIs there chocolate flowing in the stream?A water riverIs there water flowing in the stream?
A house made of candyIs the house made of candy?A house made of bricksIs the house made of bricks?
", + "image_path": "8e4e50939d9f18731c969014c03e870a630d840b249170c1d51d389e2476e000.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 555, + 590 + ], + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 590 + ], + "type": "text", + "content": "Table 3. Instances of descriptions used for generating CCS images and related questions. With manually written CCS / CS descriptions as instructions, ChatGPT is used to generate many more instances and subsequently convert them to Yes/No questions." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 612, + 493, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 612, + 493, + 624 + ], + "spans": [ + { + "bbox": [ + 314, + 612, + 493, + 624 + ], + "type": "text", + "content": "3.5. Dataset Overview and PhD Index" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "An overview of the PhD dataset is given in Tab. 4. Depending on what image (daily or CCS) is used and whether a specific context precedes a question, PhD supports four evaluation modes: PhD-base, i.e. questions about daily images w/o context, PhD-sec, i.e. PhD-base plus spcious context, PhD-icc, i.e. PhD-base plus incorrect context, and PhD-ccs, i.e. questions about CCS images. With 20 mode-task com" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19861" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "binations in total, PhD supports a much more comprehensive VHE than its predecessors [21, 34]." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 106, + 296, + 164 + ], + "blocks": [ + { + "bbox": [ + 56, + 106, + 296, + 164 + ], + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 164 + ], + "type": "table", + "html": "
TasksQuestions
ObjectAttributeSentimentPositionCountingYesNo
TDIUC samples used6,2714,3242,0952,8413,387--
Unique h items7451466548610--
VQA samples in PhD-base11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-sec11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-icc11,4727,9943,5504,9845,68816,84416,844
VQA samples in PhD-ccs34473478220124750750
", + "image_path": "3eab5dea12d5ada065cf1dbf14c24553afe5da193988a4473da71ef41adea063.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "content": "To measure the performance of an MLLM on PhD, we compute its recall w.r.t. the Yes and No questions, respectively. We term the harmonic mean of the two recalls PhD Index. A model simply saying Yes (or No) to all questions has a PhD Index of 0, while a random-guess score is 0.5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 268, + 215, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 215, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 215, + 282 + ], + "type": "text", + "content": "4. Evaluating MLLMs on PhD" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 288, + 151, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 151, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 151, + 301 + ], + "type": "text", + "content": "4.1. Common Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 305, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 435 + ], + "type": "text", + "content": "Choices of MLLMs. For reproducible research, we focus on open-source MLLMs, compiling a list of 15 models that span varied sizes and architectures. see Tab. 5. We also evaluate two hallucination mitigation methods, VCD [18] and Woodpecker [39], currently supporting LLaVA-1.6-L and Qwen-VL. As additional references, we assess three proprietary MLLMs, i.e. GPT-4o [29], Claude 3.5 Sonnet [1], and Gemini 1.5 Pro [10], on a random subset of 2k samples (random-2k) subject to our budget. For the same reason we evaluate Woodpecker, which requires paid service from ChatGPT, on random-2k." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 437, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 558 + ], + "type": "text", + "content": "Test Protocol. For a fair comparison, per MLLM we use its designated prompt to wrap each test question. For instance, a question-specific prompt submitted to mPLUG-Owl2 will be in the form of \"USER: " + }, + { + "bbox": [ + 55, + 437, + 296, + 558 + ], + "type": "inline_equation", + "content": "< |image|>\\{question\\}" + }, + { + "bbox": [ + 55, + 437, + 296, + 558 + ], + "type": "text", + "content": " ASSISTANT:\" We provide more prompts in the supplement. In addition, to help the models better handle PhD-sec and PhD-icc, we append to the test prompt an instruction as \"In case there is an inconsistency between the context and the image content, you should follow the image.\"" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 564, + 211, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 211, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 211, + 577 + ], + "type": "text", + "content": "4.2. Using PhD for Overall VHE" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "An overall VHE as shown in Tab. 5 is useful for providing a big picture of which MLLM hallucinates the most (or the least). The leading open-source MLLMs are LLaVA-OneVision, followed by Molmo and InternVL-1.5. Since their vision encoders and LLMs vary, the results are insufficient to conclude which component is the most effective to mitigate hallucinations. That said, comparisons among the same model series remains meaningful. Consider the LLaVA series for instance. While one would normally expect that a larger LLM yields a better MLLM, as LLaVA-1.6-L vs LLaVA-1.6, the difference between LLaVA-1.5-L" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "and LLaVA-1.5 is marginal (0.270 vs 0.265). In order to analyze and consequently understand such an counterintuitive result, PhD enables a zoom-in analysis in mode-oriented (Sec. 4.3) and task-oriented (Sec. 4.4) styles, unavailable in the previous benchmarks." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 143, + 553, + 318 + ], + "blocks": [ + { + "bbox": [ + 74, + 173, + 276, + 184 + ], + "lines": [ + { + "bbox": [ + 74, + 173, + 276, + 184 + ], + "spans": [ + { + "bbox": [ + 74, + 173, + 276, + 184 + ], + "type": "text", + "content": "Table 4. Data statistics of the proposed PhD dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 143, + 553, + 318 + ], + "lines": [ + { + "bbox": [ + 315, + 143, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 315, + 143, + 553, + 318 + ], + "type": "table", + "html": "
ModelViTLLMPOPEAMBERPhD
Full-set evaluation:
LLaVA-OneVision [19]SoViT-400m/14Qwen2-72B0.840.900.698
Molmo [8]-L/14Qwen2-72B0.840.850.690
InternVL-1.5 [5]InternViT-6BInternLM2-20B0.860.890.561
Qwen-VL (VCD)-bigG/14Qwen-7B0.840.870.560
Cambrian-1 [32]HybridLlama-3-8B0.880.890.547
LLaVA-1.6-L (VCD)-L/14Vicuna-13B-1.50.820.810.511
LLaVA-1.6-XL [26]-L/14Nous-Hermes-2-Yi-34B0.860.840.492
Qwen-VL [2]-bigG/14Qwen-7B0.830.840.488
LLaVA-1.6-L [26]-L/14Vicuna-13B-1.50.830.800.423
MiniGPT-v2 [4]-G/14Llama-2-7B0.830.840.390
LLaVA-1.6 [26]-L/14Vicuna-7B-1.50.830.790.373
mPlug-Ow12 [38]-L/14Llama-2-7B0.780.770.320
InstructBLIP [7]-G/14Vicuna-7B-1.10.820.820.305
InstructBLIP-L [7]-G/14Vicuna-13B-1.10.800.790.278
LLaVA-1.5-L [25]-L/14Vicuna-13B-1.50.820.730.270
LLaVA-1.5 [25]-L/14Vicuna-7B-1.50.810.750.265
LLaVA-1.1 [24]-L/14Vicuna-7B-1.10.670.330.135
Random-2k evaluation:
GPT-4o [29]--0.880.870.812
Claude 3.5 Sonnet [1]--0.850.890.746
Gemini 1.5 Pro [10]--0.860.880.691
Qwen-VL (Woodpecker)-bigG/14Qwen-7B--0.531
LLaVA-1.6-L (Woodpecker)-L/14Vicuna-13B-1.5--0.409
", + "image_path": "de797e374154d14af8b98f73f88317b4b828bdd4098ec16f803de465a50bec4b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 327, + 554, + 349 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 349 + ], + "type": "text", + "content": "Table 5. Overall VHE. POPE (adversarial) and AMBER (discriminative) are used." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 363, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 472 + ], + "type": "text", + "content": "One more advantage of PhD compared to its predecessors lies in its discrimination ability. The relatively small performance gap between GPT-4o and the top open-source models as measured by POPE and AMBER might lead to an overly optimistic interpretation that the open-source alternatives are catching up with the proprietary model. In fact, a substantial gap remains, as revealed by PhD. Fig. 3 further illustrate the qualitative results, where GPT-4o exhibits fewer hallucinations in its response." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 479, + 506, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 506, + 493 + ], + "type": "text", + "content": "4.3. Using PhD for Mode-Oriented VHE" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 498, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 594 + ], + "type": "text", + "content": "Fig. 4a illustrates mode-wise model performance. MLLMs working relatively well in the PhD-base mode tend to have stronger visual input. This is achieved either with stronger visual encoders, as the cases of LLaVA-OneVision, InternVL-1.5, and Cambrian-1 using SoViT-400m/14, InternViT-6B or hybrid vision structure, or supporting higher image resolutions, see Molmo and MiniGPT4-v2 that accept multiscale or larger input." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "In contrast to the visual part, using a larger LLM alone does not necessarily lead to a better MLLM. As noted in Sec. 4.2, LLaVA-1.5-L with Vicuna-13B-1.5 has nearly the same PhD Index (0.270) as LLaVA-1.5 with Vicuna-7B-1.5 (0.265). We see from Fig. 4a that the larger LLM indeed improves the performance in PhD-sec " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "(0.082\\rightarrow 0.099)" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", PhD-icc " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "(0.011\\rightarrow 0.019)" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " and PhD-ccs " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "(0.534\\rightarrow 0.542)" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", yet suffers loss in PhD-base " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "(0.443\\rightarrow 0.422)" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ". Similar results can be more evidently observed in the case of InstructBLIP-L vs. InstructBLIP (PhD-base Index: " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.535\\rightarrow 0.324" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "). Our con" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19862" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 555, + 242 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 555, + 242 + ], + "type": "image", + "image_path": "28682202df7b5fde01e490615c496a5370c413a8b54fb7d34ad2e55dd4bd55bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 273 + ], + "type": "text", + "content": "Figure 3. Qualitative results showing how an MLLM answers visual questions from PhD. The correctness of an answer is automatically determined by matching its first word, either Yes or No, with the ground truth (GT)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 293, + 295, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 295, + 330 + ], + "type": "text", + "content": "jecture is that although a larger LLM better understands user instructions, its successful use within an MLLM requires more targeted training for vision-language alignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 330, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 330, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 330, + 295, + 496 + ], + "type": "text", + "content": "Comparing the four modes, the performance of the open-source MLLMs on PhD-icc and PhD-sec is generally low. When provided with a multi-modal input, the models favor the textual part. Substituting a 13-B LLM for its 7-B counterpart helps tackling the inconsistency in the multi-modal input, see Fig. 5. However, a larger LLM might rely more on its internal knowledge for decoding, resulting in worse performance on PhD-base and PhD-ccs for which the image content shall carry more weights. In particular, PhD-ccs reveals a deeper and intrinsic challenge of VH: conflicts between the given image content and the model's internal knowledge. Solving the challenge demands a more comprehensive approach that goes beyond isolated improvements on the visual or language components." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 593 + ], + "type": "text", + "content": "Among the open-source MLLMs, LLaVA-OneVision is the best, owing to its joint use of a stronger visual encoder (SoViT-400m/14), a more powerful LLM (Qwen2-72B), and better training strategies (much larger high-quality training data and multi-stage alignment). Nevertheless, LLaVA-OneVision remains inferior to GPT-4o, particularly in PhD-sec and PhD-icc, wherein inconsistency within the multi-modal input has to be properly addressed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 242, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 242, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 242, + 612 + ], + "type": "text", + "content": "4.4. Using PhD for Task-Oriented VHE" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Fig. 4b presents task-oriented VHE results. In general, to what extent an MLLM hallucinates is largely correlated to the required level (low / middle / high) of a specific task. The object recognition task has the overall highest PhD Index, followed by attribute recognition, positional recognition, counting, and sentiment recognition. Due to the complexity and subtlety of emotions, e.g. tears can be associated with both happiness and sadness, even GPT-4o performs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 293, + 547, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 547, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 547, + 306 + ], + "type": "text", + "content": "relatively worse in the sentiment task (first row of Fig. 3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 306, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 415 + ], + "type": "text", + "content": "Joint mode-task analytics per model is shown in Tab. 6. LLaVA-OneVision struggles with sentiment recognition and counting, especially when faced with textual or CCS distractions, underscoring the need for improvement in these areas. Similarly, Molmo also faces these challenges, but its counting performance under CCS distractions is notably better than LLaVA-OneVision's (0.737 vs 0.563). The above zoom-in analytics will be informative for MLLM developers to prioritize their efforts on model refinement." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 340, + 425, + 533, + 544 + ], + "blocks": [ + { + "bbox": [ + 340, + 425, + 533, + 544 + ], + "lines": [ + { + "bbox": [ + 340, + 425, + 533, + 544 + ], + "spans": [ + { + "bbox": [ + 340, + 425, + 533, + 544 + ], + "type": "table", + "html": "
TaskPhD-basePhD-secPhD-iccPhD-ccs
LLAVA-OneVision
Object0.8720.8490.8240.727
Attribute0.8480.7440.6630.767
Sentiment0.6910.5810.5040.731
Positional0.7730.7300.6540.701
Counting0.7070.6520.5000.563
Molmo
Object0.8250.8800.8470.678
Attribute0.8420.7250.5560.791
Sentiment0.5470.6020.5680.746
Positional0.6970.6910.6540.742
Counting0.7270.5800.3500.737
", + "image_path": "950942eb7f2c43d25c9910bf0b713e71b491f73238e5e37a0a67f2e2aaf4866a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 347, + 553, + 520, + 564 + ], + "lines": [ + { + "bbox": [ + 347, + 553, + 520, + 564 + ], + "spans": [ + { + "bbox": [ + 347, + 553, + 520, + 564 + ], + "type": "text", + "content": "Table 6. Zoom-in analytics of specific models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 587, + 512, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 512, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 512, + 601 + ], + "type": "text", + "content": "4.5. Analysis of MLLM Answer Tendency" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "While the yes and no questions are perfectly balanced by design, we observe that the open-source MLLMs tend to answer yes, with the say-yes rate ranging from 0.462 (Molmo) to 0.811 (LLaVA-1.1) and an average value of 0.611. By contrast, the three proprietary MLLMs have a clearly lower say-yes rate. Similar observations are reported in [21, 23]. We go one step further by analyzing how the say-yes tendency is related to model performance. Ranking the models by their PhD Index and say-yes rate," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19863" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 303, + 267 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 303, + 267 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 303, + 267 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 303, + 267 + ], + "type": "image", + "image_path": "c5d732949e14a23a364068646dcb2c5ff8634ed802d965cb771589d5b17cdcdc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 271, + 220, + 280 + ], + "lines": [ + { + "bbox": [ + 143, + 271, + 220, + 280 + ], + "spans": [ + { + "bbox": [ + 143, + 271, + 220, + 280 + ], + "type": "text", + "content": "(a) Mode-oriented VHE" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 294, + 303, + 490 + ], + "blocks": [ + { + "bbox": [ + 59, + 294, + 303, + 490 + ], + "lines": [ + { + "bbox": [ + 59, + 294, + 303, + 490 + ], + "spans": [ + { + "bbox": [ + 59, + 294, + 303, + 490 + ], + "type": "image", + "image_path": "c5e82218d4718b6c092c53ccbeb7bc9ef1daf9006999695476e3a37fe17c970b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 494, + 219, + 502 + ], + "lines": [ + { + "bbox": [ + 144, + 494, + 219, + 502 + ], + "spans": [ + { + "bbox": [ + 144, + 494, + 219, + 502 + ], + "type": "text", + "content": "(b) Task-oriented VHE" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 513, + 295, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 513, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 295, + 536 + ], + "type": "text", + "content": "Figure 4. PhD based VHE analytics. Models required paid services, shown in gray markers, are tested on random-2k." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 398, + 555, + 421 + ], + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 421 + ], + "type": "text", + "content": "Figure 6. MLLM say-yes rate vs. PhD Index, with Spearman correlation of -0.92. Proprietary models tested on random-2k." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 559, + 295, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 619 + ], + "type": "text", + "content": "respectively, we calculate the Spearman's correlation between the two ranks. A strong negative correlation exists, see Fig. 6. The result suggests that addressing VH requires balancing output tendencies, with a particular focus on enhancing an MLLM's ability to say no." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 632, + 209, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 209, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 209, + 646 + ], + "type": "text", + "content": "5. Summary and Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "We have introduced PhD, a large-scale benchmark developed with a close link to the three causes of visual hallucination, i.e. visual ambiguity, inconsistency in multi-modal input and CCS content. We propose a ChatGPT-assisted semi-automated pipeline to construct the new dataset with" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 72, + 531, + 164 + ], + "blocks": [ + { + "bbox": [ + 340, + 72, + 531, + 164 + ], + "lines": [ + { + "bbox": [ + 340, + 72, + 531, + 164 + ], + "spans": [ + { + "bbox": [ + 340, + 72, + 531, + 164 + ], + "type": "image", + "image_path": "acff34f0eb8e809b5a38ddef62c4204a3e20e42ef2c2dad88d16bb2c45770776.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 174, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 217 + ], + "type": "text", + "content": "Figure 5. Impact of LLM size (7B vs 13B) on LLaVA-1.5, LLaVA-1.6 and InstructBLIP. MLLMs using a 13B LLM tend to be better than their counterparts using a 7B LLM on PhD-sec and PhD-icc, yet worse on PhD-base and PhD-ccs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 341, + 236, + 529, + 388 + ], + "blocks": [ + { + "bbox": [ + 341, + 236, + 529, + 388 + ], + "lines": [ + { + "bbox": [ + 341, + 236, + 529, + 388 + ], + "spans": [ + { + "bbox": [ + 341, + 236, + 529, + 388 + ], + "type": "image", + "image_path": "307c17e5c327a1a5e1e9c5c60010a436ef845189864c64a1716ddf7a45eb0c54.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 447, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 663 + ], + "type": "text", + "content": "well affordable manual cost. The pipeline allows us to construct diverse and visually challenging h items in an image-specific and task-specific manner. Extensive experiments with 15 open-source MLLMs, 3 proprietary MLLMs, and 2 hallucination mitigation methods support our conclusions as follows. Larger visual encoders and higher input resolutions are helpful to reduce hallucination caused by visual ambiguity. The evaluation on PhD-sec and PhD-icc suggests the current models favor the textual part in the multimodal input. Resolving the conflicts between the CCS content and the model's internal knowledge demands a more comprehensive approach that is beyond isolated improvements on the visual or language components. Among the open-source MLLMs, LLaVA-OneVision is the best, followed by Molmo and InternVL-1.5. While existing benchmarks could lead to an overly optimistic expectation that the open-source models are catching up with GPT-4o, a substantial performance gap remains, as revealed by PhD." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgements. This research was supported by NSFC (62172420), Tencent Marketing Solution Rhino-Bird Focused Research Program and the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19864" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Anthropic. Introducing Claude 3.5 Sonnet, 2024. https://www.anthropic.com/news/claude-3-5-sonnet.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "type": "text", + "content": "[3] Nitzan Bitton-Guetta, Yonatan Bitton, Jack Hessel, Ludwig Schmidt, Yuval Elovici, Gabriel Stanovsky, and Roy Schwartz. Breaking common sense: WHOOPS! a vision-and-language benchmark of synthetic and compositional images. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 300 + ], + "type": "text", + "content": "[4] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. MiniGPT-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "type": "text", + "content": "[5] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. InternVL: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 401 + ], + "type": "text", + "content": "[6] Chenhang Cui, Yiyang Zhou, Xinyu Yang, Shirley Wu, Linjun Zhang, James Zou, and Huaxiu Yao. Holistic analysis of hallucination in GPT-4v(ison): Bias and interference challenges. arXiv preprint arXiv:2311.03287, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 402, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 294, + 456 + ], + "type": "text", + "content": "[7] Wenliang Dai, Junnan Li, DONGXU LI, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 457, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 294, + 499 + ], + "type": "text", + "content": "[8] Matt Deitke, Christopher Clark, Sangho Lee, et al. Molmo and PixMo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 293, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 293, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 293, + 524 + ], + "type": "text", + "content": "[9] Doubao Team. Doubao product, 2024. https://www.volcengine.com/product/doubao.5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "type": "text", + "content": "[10] Google. Google Gemini 1.5 Pro, 2024. https://deepmind.google/technologies/gemini/pro/.6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 558, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 623 + ], + "type": "text", + "content": "[11] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. HallusionBench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "text", + "content": "[12] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. In AAAI, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "type": "text", + "content": "[13] Hongyu Hu, Jiyuan Zhang, Minyi Zhao, and Zhenbang Sun. CIEM: Contrastive instruction evaluation method for better instruction tuning. In NeurIPS Workshop on ITIF, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[14] Yutao Hu, Tianbin Li, Quanfeng Lu, Wenqi Shao, Junjun He, Yu Qiao, and Ping Luo. OmniMedVQA: A new large-scale" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "comprehensive evaluation benchmark for medical LVLM. In CVPR, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[15] Liqiang Jing, Ruosen Li, Yunmo Chen, Mengzhao Jia, and Xinya Du. FAITHSCORE: Evaluating hallucinations in large vision-language models. In Findings of the Association for Computational Linguistics: EMNLP, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 140, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 161 + ], + "type": "text", + "content": "[16] Kushal Kafle and Christopher Kanan. An analysis of visual question answering algorithms. In ICCV, 2017. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "text", + "content": "[17] Kang-il Lee, Minbeom Kim, Seunghyun Yoon, Minsung Kim, Dongryeol Lee, Hyukhun Koh, and Kyomin Jung. VLind-Bench: Measuring language priors in large vision-language models. In NAACL Findings, 2025. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "text", + "content": "[18] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In CVPR, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 251, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 294 + ], + "type": "text", + "content": "[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 294, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 337 + ], + "type": "text", + "content": "[20] Junyi Li, Jie Chen, Ruiyang Ren, Xiaoxue Cheng, Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. The dawn after the dark: An empirical study on factuality hallucination in large language models. In ACL, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 338, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 371 + ], + "type": "text", + "content": "[21] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 553, + 415 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 415, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 553, + 459 + ], + "type": "text", + "content": "[23] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In ICLR, 2024. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 460, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 481 + ], + "type": "text", + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "text", + "content": "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 1, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "text", + "content": "[26] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. https://llava-v1.github.io/blog/2024-01-30-llava next/.6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 613 + ], + "type": "text", + "content": "[27] Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (NOPE) to measure object hallucination in vision-language models. In ALVR, 2024. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 614, + 552, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 552, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 552, + 635 + ], + "type": "text", + "content": "[28] OpenAI. Dall-E 3 system, 2023. https://openai.com/index/dall-e-3/.5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 636, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 658 + ], + "type": "text", + "content": "[29] OpenAI. Hello GPT-4o, 2024. https://openai.com/index/hello-gpt-4o/.6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19865" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 408 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[31] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. IllusionVQA: A challenging optical illusion dataset for vision language models. In COLM, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 173 + ], + "type": "text", + "content": "[32] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. In NeurIPS, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 175, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 295, + 206 + ], + "type": "text", + "content": "[33] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 208, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 208, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 294, + 261 + ], + "type": "text", + "content": "[34] Junyang Wang, Yuhang Wang, Guohai Xu, Jing Zhang, Yukai Gu, Haitao Jia, Ming Yan, Ji Zhang, and Jitao Sang. An LLM-free multi-dimensional benchmark for mllms hallucination evaluation. arXiv preprint arXiv:2311.07397, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 263, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 263, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 57, + 263, + 295, + 317 + ], + "type": "text", + "content": "[35] Junyang Wang, Yiyang Zhou, Guohai Xu, Pengcheng Shi, Chenlin Zhao, Haiyang Xu, Qinghao Ye, Ming Yan, Ji Zhang, Jihua Zhu, et al. Evaluation and analysis of hallucination in large vision-language models. arXiv preprint arXiv:2308.15126, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 319, + 295, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 319, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 57, + 319, + 295, + 351 + ], + "type": "text", + "content": "[36] Qijie Wei, Kaiheng Qian, and Xirong Li. FunBench: Benchmarking fundus reading skills of MLLMs. arXiv preprint arXiv:2503.00901, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 353, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 353, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 57, + 353, + 294, + 385 + ], + "type": "text", + "content": "[37] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 387, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 387, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 57, + 387, + 294, + 408 + ], + "type": "text", + "content": "[38] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 408 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Zhou. mPLUG-Owl2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[39] Shukang Yin, Chaoyou Fu, Sirui Zhao, Tong Xu, Hao Wang, Dianbo Sui, Yunhang Shen, Ke Li, Xing Sun, and Enhong Chen. Woodpecker: Hallucination correction for multimodal large language models. Science China Information Sciences, 67(12):220105, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 163, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 250 + ], + "type": "text", + "content": "[40] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. MMMU: A massive multi-discipline multimodal understanding and reasoning benchmark for expert AGI. In CVPR, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "type": "text", + "content": "[41] Le Zhang, Rabiul Awal, and Aishwarya Agrawal. Contrasting intra-modal and ranking cross-modal hard negatives to enhance visio-linguistic compositional understanding. In CVPR, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "text", + "content": "[42] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the AI ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "text", + "content": "[43] Kankan Zhou, Eason Lai, Wei Bin Au Yeong, Kyriakos Mouratidis, and Jing Jiang. ROME: Evaluating pre-trained vision-language models on reasoning beyond visual common sense. In Findings of the Association for Computational Linguistics: EMNLP, 2023. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19866" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/497d038e-fa23-45f9-aee3-d3bfa187eaf5_content_list.json b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/497d038e-fa23-45f9-aee3-d3bfa187eaf5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..91b9b24c5446a6dd8b4bbc3dd7a683e44d21dc40 --- /dev/null +++ b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/497d038e-fa23-45f9-aee3-d3bfa187eaf5_content_list.json @@ -0,0 +1,1736 @@ +[ + { + "type": "image", + "img_path": "images/94f26f24f6f7452f7e01aa5535d32104cc00a0a2237b3a7d26cdaf862c4b8530.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 204, + 122, + 230, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Phoenix: A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction", + "text_level": 1, + "bbox": [ + 228, + 133, + 799, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenke Xia $^{1,2,*}$ , Ruoxuan Feng $^{1}$ , Dong Wang $^{2}$ , Di Hu $^{1,\\dagger}$", + "bbox": [ + 274, + 213, + 720, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Gaoling School of Artificial Intelligence, Renmin University of China, Beijing $^{2}$ Shanghai AI Laboratory", + "bbox": [ + 186, + 239, + 810, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 310, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building a generalizable self-correction system is crucial for robots to recover from failures. Despite advancements in Multimodal Large Language Models (MLLMs) that empower robots with semantic reflection ability for failure, translating semantic reflection into \"how to correct\" fine-grained robotic actions remains a significant challenge. To address this gap, we build the Phoenix framework, which leverages motion instruction as a bridge to connect high-level semantic reflection with low-level robotic action correction. In this motion-based self-reflection framework, we start with a dual-process motion adjustment mechanism with MLLMs to translate the semantic reflection into coarse-grained motion instruction adjustment. To leverage this motion instruction for guiding \"how to correct\" fine-grained robotic actions, a multi-task motion-conditioned diffusion policy is proposed to integrate visual observations for high-frequency robotic action correction. By combining these two models, we could shift the demand for generalization capability from the low-level manipulation policy to the MLLMs-driven motion adjustment model and facilitate precise, fine-grained robotic action correction. Utilizing this framework, we further develop a lifelong learning method to automatically improve the model's capability from interactions with dynamic environments. The experiments conducted in both the RoboMimic simulation and real-world scenarios prove the superior generalization and robustness of our framework across a variety of manipulation tasks. Our code is released at https://github.com/GeWuLab/Motion-based-Self-Reflection-Framework.", + "bbox": [ + 86, + 343, + 483, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 799, + 220, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"Failure is simply the opportunity to begin again, this time more intelligently.\" — Henry Ford", + "bbox": [ + 107, + 825, + 465, + 856 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc3cb80f6ac08831bf8e9a0baf087d94b24092c910fdf011164c138d348183ac.jpg", + "image_caption": [ + "Figure 1. Our motion-based self-reflection framework utilizes coarse-grained motion instruction as a bridge to convert the high-level semantic reflection into fine-grained robotic action correction, thereby facilitating generalizable and precise action correction with perceptual and inferential capabilities of MLLMs." + ], + "image_footnote": [], + "bbox": [ + 539, + 311, + 883, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans are naturally equipped with the ability to correct their behaviors by intentionally reflecting on actions that lead to failure. [12, 25]. By analyzing failure situations from high-level semantic reflection and low-level action correction perspective, humans can efficiently adapt to dynamic environments [8]. To emulate the correction capability and foster a continuous cycle of self-improvement in robots, researchers [26, 30, 35] have sought to develop self-reflection systems that enable robots to recover from and learn through their failure interactions.", + "bbox": [ + 509, + 732, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Among them, some existing self-correction systems [19,", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work is done during internship at Shanghai AI Laboratory", + "bbox": [ + 107, + 875, + 423, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6981", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "26, 34] leverage reinforcement learning to guide robots in correct low-level action execution through reward functions. However, the reliance on reinforcement learning limits the ability of these self-correction systems to generalize across long-horizon manipulation tasks, due to unstable training processes [10] and the need for task-specific prior knowledge [9, 35]. To construct a generalizable and stable self-correction system, recent works [17, 30, 31] borrow the inferential capability of Multi-modal Large Language Models (MLLMs) to propose closed-loop high-level semantic reflection framework for failure correction. Although these semantic self-reflection frameworks can decompose the failure correction process into semantic subgoals, they primarily rely on a predefined skill library to execute the detailed subgoals, which fails to utilize the generalization ability of MLLMs in fine-grained robotic action correction.", + "bbox": [ + 89, + 90, + 480, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To maximize the generalization potential of MLLMs for action correction, we propose motion instruction as a bridge to convert high-level semantic reflection to fine-grained robotic action correction. Motion instruction refers to coarse-grained robotic movement commands such as \"move arm backward\" and \"adjust gripper position\". Serving as an intermediate layer, motion instruction could provide general, low-frequency decision information for high-frequency robotic action execution, which makes it an excellent medium for embedding the knowledge of MLLMs into fine-grained action correction. As shown in Figure 1, we decompose the semantic reflection knowledge into coarse-grained motion instruction adjustment to indicate \"how to correct\" fine-grained action for low-level policy execution. This transition shifts the perceptual and decision-making requirements from low-level robotic policy to the MLLMs-driven motion adjustment model, thereby enabling generalizable, fine-grained robotic action correction.", + "bbox": [ + 91, + 335, + 482, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hence, in this work, we build the Phoenix framework, a motion-based self-reflection framework designed to convert the semantic reflection of MLLMs into fine-grained robotic action correction. Initially, we develop a dual-process motion adjustment mechanism to ensure efficient prediction through a motion prediction module, while addressing failure with a motion correction module. Concretely, we first utilize expert demonstration trajectories to train the motion prediction module for efficient motion instruction generation. Despite its efficiency in generating initial instructions, this module often struggles to handle failure scenarios. To recover from failures, we collect a comprehensive failure correction dataset and fine-tune the motion correction module, which thoroughly provides adjusted motion instructions through a chain-of-thought approach. By integrating these two modules, the dual-process motion adjustment mechanism guarantees both robustness and efficiency, facilitating the generation of accurate motion instructions. As the coarse-grained motion instructions only provide gen", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eral and low-frequency guidance for robotic manipulation, we further design a multi-task motion-conditioned diffusion policy that integrates visual observations to translate motion instruction into precise, high-frequency action corrections for manipulation tasks. Moreover, by leveraging these correction trajectories, we propose a lifelong learning method that iteratively enhances the model's capabilities through interaction, ensuring continuous improvements in performance and adaptability to dynamic environments.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate the efficacy of our framework, we conduct experiments across 9 contact-rich robotic manipulation tasks within the RoboMimic simulation [20]. The results demonstrate that our method could provide more precise action correction from failures through self-reflection and facilitate self-improvement through interactions with environments. Further, we conduct two novel manipulation tasks with color disruption and position distribution disruption, proving the generalization ability of our framework. The real-world experiments also demonstrate the applicability and robustness of our approach in practical scenarios.", + "bbox": [ + 511, + 227, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 409, + 653, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Robotic Self-correction Systems", + "text_level": 1, + "bbox": [ + 511, + 434, + 792, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-correction serves as a crucial mechanism enabling robots to recover from failures. To achieve self-correction on the low-level robotic action, Reinforcement Learning [1, 23] is proposed to guide robots in adjusting behaviors via reward signals. However, reinforcement learning strategies encounter difficulties in intricate robotic environments primarily due to learning inefficiency. Borrowing the common knowledge of MLLMs, Raman et al. [24] proposes to build a semantic self-reflection system for long-horizon task planning. To facilitate interaction with environments, some efforts generate robotic action through simulation APIs [17, 29, 30] and predefined action skill libraries [14, 16]. However, the reliance on predefined low-level skill libraries makes semantic self-reflection frameworks fail to directly provide fine-grained action correction. To provide low-level action feedback for robotic manipulation, recent works [6, 16, 32] suggest adjusting end-effector poses to refine actions. Nonetheless, these techniques are primarily restricted to simple manipulation tasks that employ motion planning and fail to generalize to contact-rich manipulation scenarios. In this work, we utilize motion instruction as an intermediate layer to guide robotic action correction, borrowing the perceptual and inferential capability of MLLMs for fine-grained robotic action correction.", + "bbox": [ + 511, + 458, + 903, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Robotic Manipulation Policy", + "text_level": 1, + "bbox": [ + 511, + 830, + 769, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The development of generalizable strategies for robotic manipulation remains a persistent challenge in robotics research. ACT [38] is proposed to predict action sequences", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6982", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/27ffee8de0eb2b660bfadeec206406a889c4db241c7cf888ebff4097a6c8a490.jpg", + "image_caption": [ + "Figure 2. The pipeline of our motion-based self-reflection framework. (a) demonstrates the dual-process motion refinement mechanism which leverages the motion prediction module for efficient motion instruction prediction and motion correction module for comprehensive failure correction. (b) illustrates the motion-conditioned diffusion policy which converts the low-frequency motion instruction guidance into high-frequency robotic action. The lifelong learning approach in (c) iteratively enhances the ability of the motion prediction module from the refined interaction trajectories." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 903, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to ensure temporal alignment of actions, while the Diffusion Policy [5] addresses multi-modal action distributions to enhance robust manipulation. Driven by the development of foundation models [13, 28], recent works [4, 7, 14, 22, 36, 37] leverage the world knowledge of MLLMs to facilitate task decomposition and planning for robotic manipulation. Concurrently, other works [3, 11, 21, 27] collect large-scale robotic manipulation demonstrations to train generalizable language-conditioned manipulation policy across different robots and tasks. Moreover, RT-H [2] employs detailed motion commands instead of semantic language inputs, fostering flexible and generalizable manipulation capabilities through the advanced perceptual abilities of MLLMs. However, the prohibitive costs associated with robotic data collection limit the scalability of existing imitation learning approaches. To address this, we propose the motion-based self-reflection framework to enable autonomous self-improvement through continuous interaction with environments without human intervention.", + "bbox": [ + 91, + 357, + 483, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Motion-based Self-Reflection Framework", + "text_level": 1, + "bbox": [ + 89, + 656, + 460, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Challenges in Robotic Self-Correction Model", + "text_level": 1, + "bbox": [ + 89, + 681, + 464, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building a generalizable and robust self-correction system is a key component in achieving failure correction for robots. Multi-modal Large Language Models (MLLMs) have already been applied to the construction of robotic self-reflection framework to recover from failures. However, existing systems mainly focus on semantic reflection, and their application to fine-grained action correction still faces the following two issues:", + "bbox": [ + 89, + 704, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- How to enable MLLMs to understand manipulation tasks and provide detailed correction information?", + "- How to convert the correction information provided by MLLMs into precise, high-frequency robotic actions?" + ], + "bbox": [ + 89, + 825, + 483, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these issues, we propose the Phoenix frame", + "bbox": [ + 109, + 886, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work, a motion-based self-reflection framework that integrates a dual-process motion adjustment mechanism and a multi-task motion-conditioned diffusion policy as illustrated in Figure 2. As detailed in Section 3.2, the dual-process motion adjustment mechanism is developed to ensure efficient and accurate motion instruction generation. Further, the motion-conditioned diffusion policy is proposed to translate the coarse-grained motion instructions into precise robotic actions, as explained in Section 3.3. Based on the refined manipulation trajectories, we propose a lifelong learning approach to facilitate robotic self-improvement, as outlined in Section 3.4.", + "bbox": [ + 511, + 357, + 906, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dual-process Motion Adjustment Mechanism", + "text_level": 1, + "bbox": [ + 511, + 545, + 898, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The dual-process motion adjustment mechanism is designed to ensure efficient motion prediction through a motion prediction module, while comprehensively addressing failure with a motion correction module. Given the observation $O$ and task description $T$ , we first train a Motion Prediction Module (MPM) with expert demonstration dataset $D_{e}$ to generate initial motion instruction $m_{i}$ . However, the MPM trained on expert demonstrations struggles to handle failure situations. Thus, we construct a comprehensive failure correction dataset $D_{c}$ to fine-tune the Motion Correction Module (MCM), enabling it to analyze the failure situation and adjust $m_{i}$ with a chain-of-thought approach. If $m_{i}$ is deemed correct, we adopt it as the decision motion instruction $m_{d}$ for further robotic action prediction. Otherwise, we employ the MCM to analyze the failure situation and generate adjusted motion instruction $m_{a}$ as decision motion instruction $m_{d}$ . Through the guidance of $m_{d}$ , our motion-based diffusion policy can generate high-frequency corrections to the robotic actions. As described in Algorithm 1, we establish the dual-process motion adjustment mechanism to guarantee the efficiency and accuracy of motion instruction generation for fine-grained robotic action prediction.", + "bbox": [ + 511, + 568, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6983", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Self-Reflection w/. Dual-Process Motion Adj." + ], + "code_body": "Require: Task description $T$ , Observation $O$ , Environment $E$ , Motion prediction module $MPM$ , Motion correction module $MCM$ , Motion-conditioned diffusion policy $\\pi$ , Exploration timestep $K$ ,", + "guess_lang": "verilog", + "bbox": [ + 91, + 109, + 483, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1: $O_{1}\\gets E.\\text{reset} ()$", + "bbox": [ + 102, + 171, + 233, + 185 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2: for $k = 1$ to $K$ do", + "bbox": [ + 102, + 186, + 243, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3: $m_{i}\\gets MPM(O_{k},T)$", + "bbox": [ + 102, + 200, + 290, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4: failure_flag, semantic_info $\\leftarrow$ MCM(Ok,m)", + "bbox": [ + 102, + 215, + 477, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5: if failure_flag is true then", + "bbox": [ + 102, + 232, + 334, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6: $m_{a}\\gets MCM(O_{k},\\text{semantic\\_info})$", + "bbox": [ + 102, + 246, + 401, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "7: $m_{d}\\gets m_{a}$", + "bbox": [ + 102, + 263, + 240, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "8: else", + "bbox": [ + 102, + 277, + 173, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "9: $m_{d}\\gets m_{i}$", + "bbox": [ + 102, + 294, + 238, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "10: end if", + "bbox": [ + 101, + 306, + 187, + 319 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "11: $a\\gets \\pi (O,m_d)$", + "12: $O_{k + 1}\\gets E.\\text{step}(a)$" + ], + "bbox": [ + 96, + 321, + 274, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "13: end for", + "bbox": [ + 96, + 353, + 173, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion Prediction Module (MPM). To fully harness the perceptual and decision-making capabilities of MLLMs for efficient motion instruction prediction, we develop a motion instruction dataset from the expert demonstrations dataset $D_{e}$ to fine-tune MLLMs for robotic manipulation tasks. To construct the expert dataset, we filter the robotic action to get dominant motion from expert demonstration with a threshold, generating a set of motion instructions that include arm direction and gripper control. In practice, we find that separating arm direction instruction and gripper control instruction would cause the misalignment between textual motion instruction and fine-grained robotic action. To address this issue, we combine the direction movement with gripper control, resulting in unified instruction formats for motion instruction such as \"move arm right with gripper closed\". In addition, we add the \"make slight adjustments to gripper position\" instruction to model the minor robotic actions below the threshold. Through the automatic construction method, we build 37 types of motion instructions as guidance for further robotic action prediction. By training on the expert dataset, the MPM acquires an understanding of robotic manipulation tasks and could efficiently generate an initial motion instruction $m_{i}$ .", + "bbox": [ + 89, + 398, + 483, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion Correction Module (MCM). During interactions with the environment, robots may execute incorrect actions, leading to a failure situation in the task. However, the MPM trained on success expert data often struggles to recover from these failure scenarios. Thus, we develop the motion correction module, to identify failure scenarios and correct behaviors from such situations. As shown in Figure 2(a), the MCM would evaluate the initial motion instruction $m_{i}$ and conduct a dual process based on the evaluation results to achieve efficient and accurate motion in-", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58b902bcb2c3bd382050d866be6c471215691932c42cb26724dfad1a2905896a.jpg", + "image_caption": [ + "(a) Online Human Interventions" + ], + "image_footnote": [], + "bbox": [ + 535, + 89, + 883, + 148 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b628c76d9c9b20eacf6ea52188f1e403b38ab35478911fa1cd95303fe633f455.jpg", + "image_caption": [ + "(b) Offline Human Annotations" + ], + "image_footnote": [], + "bbox": [ + 537, + 167, + 883, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f28590979d23cb25a51c3f2bc60b5abc50a1cf57d4b9215d6a3274676501cd39.jpg", + "image_caption": [ + "(c) Expert Demonstrations", + "Figure 3. Illustrations of our correction data: online human interventions, offline human annotations, and expert demonstrations." + ], + "image_footnote": [], + "bbox": [ + 537, + 247, + 883, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "struction adjustment. Once encountering the failure situation, the MCM would first analyze the type of failure and derive a semantic-level correction goal, such as \"insert the coffee pot into the coffee machine\". Based on this correction goal, the MCM further adjusts the motion instructions with its learned failure-correction knowledge, ultimately generating an accurate motion instruction through a hierarchical chain-of-thought approach.", + "bbox": [ + 511, + 361, + 905, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To equip MCM with the capabilities for failure detection and correction, we construct a comprehensive correction dataset as illustrated in Figure 3. This dataset includes three types of feedback data, encompassing both semantic and motion perspectives, and is categorized as follows:", + "bbox": [ + 511, + 483, + 905, + 559 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Online Human Intervention. We implement the human-in-the-loop method for trajectory collection. We first deploy the motion prediction model to interact with environments, then manually intervene to correct the motion instructions whenever the agent encounters failure situations. This method could collect accurate and high-quality motion correction data to ensure task completion. However, it requires frequent manual interactions with the environment, which leads to significant time consumption and makes it difficult to collect large-scale data.", + "- Offline Human Annotation. We utilize the motion prediction model to gather trajectory data, periodically sampling trajectories and annotating them with semantic reflections and motion correction details. While the accuracy of offline annotated data cannot be guaranteed due to its inability to interact with environments for verification, this method offers a significant volume of annotations.", + "- Expert Demonstration. We automate annotations on expert trajectories. Since these trajectories are successful, this data is used to provide accurate motion information to enhance the model's motion prediction capabilities." + ], + "bbox": [ + 513, + 563, + 906, + 881 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By fine-tuning MCM on this dataset, we enhance the", + "bbox": [ + 531, + 885, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6984", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MCM to thoroughly comprehend various types of failure situations and provide motion instruction corrections. Through the integration of MPM and MCM, the dual-process motion adjustment mechanism enables efficient motion instruction generation while ensuring comprehensive correction in failure situations.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Motion-conditioned Diffusion Policy", + "text_level": 1, + "bbox": [ + 89, + 191, + 405, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the motion instruction only provides general and low-frequency guidance for manipulation, we train a multi-task motion-conditioned diffusion policy $\\pi$ to convert the motion instructions into precise, high-frequency robotic actions. This policy takes observations $O$ and decision motion instructions $m_{d}$ to output robotic actions $a$ . To ensure the policy adheres to the motion instruction, we make adjustments as depicted in Figure 2(b):", + "bbox": [ + 89, + 215, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we observe that existing pre-trained language models often struggle to capture the discriminative features of various motion instructions. This limitation hampers their ability to follow various motion instructions. To address this issue, we introduce a learnable motion codebook designed to provide discriminative features for motion instructions. For a given decision motion instruction $m_d$ , the codebook would retrieve the corresponding motion feature to facilitate accurate robotic action prediction.", + "bbox": [ + 89, + 337, + 483, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further, we find that the direct concatenation of observation representation and motion instruction feature would cause the diffusion policy to prefer to rely on the vision information for action prediction, thereby hindering the effectiveness of the motion instruction guidance. To address this issue, we take the observation representation and motion instruction feature as separate conditions in different stages of the diffusion policy, allowing the model to better learn the guidance information from the motion instruction and thereby promote precise action correction.", + "bbox": [ + 89, + 474, + 483, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By integrating these two adjustments, we train the diffusion policy for action prediction with the following loss:", + "bbox": [ + 89, + 625, + 482, + 656 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\operatorname {M S E} \\left(\\mathcal {E} ^ {k}, \\pi (\\mathcal {O}, \\mathcal {M}, \\mathcal {A} ^ {0} + \\mathcal {E} ^ {k}, k)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 667, + 482, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{O}$ is the observation representation, $\\mathcal{M}$ is the motion instruction feature, $\\mathcal{A}^0$ is the ground truth robotic action, $\\mathcal{E}^k$ indicates the random noise at the denoising iteration $k$ . Through minimizing the loss function in Eq 1, the diffusion policy $\\pi$ could effectively predict precise, high-frequency robotic action guided by motion instruction.", + "bbox": [ + 89, + 685, + 482, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Action Correction for Lifelong Learning", + "text_level": 1, + "bbox": [ + 89, + 786, + 437, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The dual-process motion adjustment mechanism leverages the MPM to efficiently predict motion instructions and the MCM to adjust them with a comprehensive chain-of-thought approach. However, the reliance on the chain-of-thought poses challenges in adapting to real-time scenarios due to its time-consuming. Furthermore, the collection of", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "manipulation data and correction data is exceedingly labor-intensive. Thus, we propose a lifelong learning method that equips the MPM with both motion prediction and failure correction capabilities through learning from the refined interaction trajectories as illustrated in Figure 2(c), which enhances our model to adapt and react quickly to the environment without human intervention.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benefiting from the motion-conditioned diffusion policy which could adhere to the motion instruction to generate task-aware robotic action, we can enhance the robot's capabilities through only improving the MPM informed by the refined interaction trajectory. To address the issue of catastrophic forgetting, we mix the refined interaction trajectory with expert demonstration for co-fine-tuning, allowing the model to simultaneously learn failure correction and enhance the motion prediction capabilities. Through updates from refined interaction trajectories, our model can achieve self-improvement by learning from the knowledge of the motion correction module, achieving fast and accurate manipulation for contact-rich manipulation tasks.", + "bbox": [ + 511, + 196, + 905, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 406, + 645, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To comprehensively evaluate our framework, we propose experiments to answer the following questions:", + "bbox": [ + 511, + 431, + 903, + 460 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Does our motion-guided self-reflection model enhance the precision of action correction? Section 4.2", + "- Can our model achieve lifelong learning from interaction with environments? Section 4.3", + "- Does our framework can generalize across novel tasks? Section 4.4", + "- Can our framework ensure reliability and robustness in real-world scenarios? Section 4.5" + ], + "bbox": [ + 513, + 462, + 903, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 511, + 590, + 705, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we conduct experiments on 9 contact-rich manipulation tasks in RoboMimic [20], ranging from long-horizon tasks like ThreePieceAssembly to fine-grained manipulation tasks like Threading. To transform high-level semantic information into motion instructions, we filter expert demonstrations to obtain over 160,000 pairs of motion instructions and observations. The dataset includes 37 types of motion instructions, which are utilized to fine-tune the LLaVA-v1.5 model [15] as the motion prediction module. Furthermore, to develop the motion correction module that integrates semantic comprehension and motion instruction adjustment, we collect correction data comprising 3,644 online human intervention data, 7,365 offline human annotations, and 6,378 expert demonstrations. We filter out the correction dataset to balance the proportion of various failure situations to enhance the model's correction capabilities. Ultimately, to translate motion instructions into precise robotic actions, we train a multi-task motion-conditioned diffusion policy using a learnable motion instruction code", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6985", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ebab80a91437096570bd51489d2ec5bba181bd80abf35b37f5259ad8c11c25c8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCoffee_D0Coffee_D1Stack_D0Stack_D1StackThree_D0StackThree_D1Threading_D0ThreePieceAssembly_D0ThreePieceAssembly_D1Mean
OpenVLA [11]42%18%84%86%36%20%20%28%8%38.0%
Task-conditioned66%24%88%68%30%6%74%20%0%41.8%
Subgoal-conditioned76%26%88%74%24%6%78%20%2%43.8%
Motion-conditioned68%32%92%84%38%16%58%30%4%46.9%
Subgoal Self-reflection80%32%88%78%32%6%80%34%2%48.0%
Phoenix (Ours)94%48%96%86%50%20%68%52%6%57.8%
Human Intervention (Oracle)100%100%100%90%70%40%100%70%40%78.9%
", + "bbox": [ + 93, + 88, + 906, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison experiments results across 9 manipulation tasks in RoboMimic Simulation. The results demonstrate that our motion-based self-reflection method achieves better performance by facilitating precise correction of fine-grained robotic actions.", + "bbox": [ + 89, + 190, + 906, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "book, incorporating 500 demonstrations per task. During inference in simulation, our dual-process motion adjustment mechanism would provide motion instruction at $5Hz$ , and the diffusion policy would extend the motion instruction with visual observations to a $20Hz$ action sequence to control the robot. For each task, we conducted 50 trials and report the average success rate. More implementation details could refer to Supp.A.", + "bbox": [ + 88, + 232, + 480, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance of Motion Self-Reflection Model", + "text_level": 1, + "bbox": [ + 89, + 364, + 475, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Comparison Results", + "text_level": 1, + "bbox": [ + 89, + 388, + 285, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate our motion-based self-reflection framework, we compare our framework with other approaches. To ensure fairness, all our comparison methods are trained on the expert data from the simulation environment, with the decision model using LLaVA-v1.5 and the underlying policy employing a diffusion policy.", + "bbox": [ + 89, + 412, + 482, + 503 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- OpenVLA [11]. We fine-tune the OpenVLA model to provide baseline performance for multi-task experiments.", + "- Task-conditioned policy. We take the task description as the condition for diffusion policy without the reflection framework, as a variance of RT-1 [3] and Octo [27].", + "- Subgoal-conditioned policy. We fine-tune a LLaVAv1.5 to predict subgoals at $5\\mathrm{Hz}$ , which are utilized as the condition for diffusion policy without reflection framework. This method borrows the semantic comprehension capabilities of MLLMs, and is implemented as a variance of PaLM-E [7] with an individual diffusion policy.", + "- Motion-conditioned policy. We fine-tune a LLaVA-v1.5 as the motion prediction model to provide motion instructions at 5Hz, using these predictions to condition the diffusion policy without the reflection framework. This method employs the perceptual and inferential capacities of MLLMs, realized as a variation of RT-H [2] with an individual diffusion policy.", + "- Human Intervention. We manually correct the wrong motion instructions for the motion-conditioned policy. This method provides an upper bound on the performance of self-reflection methods. Due to labor costs, the results are presented as average success rates across 10 trials.", + "- Subgoal Self-reflection. We fine-tune a LLaVA-v1.5 as subgoal self-reflection model and apply it to the subgoal-condition policy. This method is designed to validate the" + ], + "bbox": [ + 89, + 507, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "effectiveness of the semantic self-reflection model.", + "bbox": [ + 526, + 233, + 862, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, we first compare three different condition methods. Borrowing the perception and inferential ability of MLLMs, the subgoal-conditioned and motion-conditioned policies are better than the task-conditioned policies. The results prove the potential applications of MLLMs in various complex robotic manipulation tasks.", + "bbox": [ + 511, + 251, + 905, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Focusing on specific tasks, we observe that the motion-conditioned policy excels in long-horizon tasks such as StackThree_D0 and ThreePieceAssembly_D0. However, this policy depends on consistent and accurate motion instruction predictions, which poses challenges in fine-grained manipulation tasks like Threading_D0.", + "bbox": [ + 511, + 342, + 903, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By providing correction subgoals, the subgoal self-reflection method consistently outperforms the subgoal-conditioned policy, particularly in long-horizon manipulation tasks such as \"StackThree_D0\", which demonstrates the efficacy of the self-reflection framework.", + "bbox": [ + 511, + 431, + 903, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The OpenVLA model demonstrates strong performance in certain long-horizon tasks, leveraging its end-to-end action token prediction capability. However, the lack of observation history and action chunking poses significant challenges in handling complex, fine-grained manipulation tasks like Threading_D0.", + "bbox": [ + 511, + 508, + 905, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, our Phoenix method achieves more substantial improvements than the subgoal self-reflection method, demonstrating the effectiveness of motion-conditioned method in long-horizon sequential tasks and fine-grained manipulation tasks. Benefiting from our motion-based correction method, agent could correct fine-grained action through motion instruction adjustment while the subgoal-conditioned self-reflection model fails to recover from most failure situations. Besides, the human intervention method achieves high success rates across multiple tasks, demonstrating that our motion-conditioned diffusion policy can effectively adhere to motion instructions for manipulation tasks. This result indicates that our method can perform well under the correct motion instructions, showcasing the significant potential of motion-conditioned self-reflection.", + "bbox": [ + 511, + 599, + 905, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Ablation Results", + "text_level": 1, + "bbox": [ + 511, + 845, + 684, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this work, we propose a motion prediction module to provide initial motion instruction, and a motion correction", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6986", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0ef03f3caa711c400e78dd9f77f37892bd120d4ed5a3ef8e9e158302329fb535.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCoffee_D0Coffee_D1Stack_D0Stack_D1StackThree_D0StackThree_D1Threading_D0ThreePieceAssembly_D0ThreePieceAssembly_D1Mean
Motion-conditioned68%32%92%84%38%16%58%30%4%46.9%
Expert-Correction Mixture74%36%94%86%38%22%64%30%2%49.6%
Expert-Correction Mixture with Self-Reflection76%30%92%90%46%26%64%34%4%51.3%
Phoenix (Ours)94%48%96%86%50%20%68%52%6%57.8%
", + "bbox": [ + 93, + 88, + 903, + 143 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. The ablation results of our dual-process motion adjustment mechanism. The results prove that our model, which separates motion prediction module and motion correction module, can provide more precise motion adjustment for action execution in manipulation tasks.", + "bbox": [ + 89, + 155, + 906, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c5e16dfc060a6d846025f0f7976d778a895c4c8af80581777bb3ef876c3239e.jpg", + "image_caption": [ + "Figure 4. The lifelong learning results. The results prove that our motion-based self-reflection method could iteratively improve performance through interactions with environments." + ], + "image_footnote": [], + "bbox": [ + 94, + 186, + 336, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5c7d13d24316b6a66738f0e5c7fd5271370f288a41756906c517321701d876f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 186, + 640, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e693f94a9c4eddc95f4716ffb5fd1cf83742b4809eed02076e406ab513d81989.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 186, + 900, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "module to provide fine-grained motion correction. Drawing upon prior research [18, 33], data mixture proportions could influence the efficacy of LLMs. In this section, we investigate whether integrating expert demonstrations with correction dataset, could also enhance the perception and decision-making capabilities of MLLMs for robotic manipulation with the following ablation methods:", + "bbox": [ + 88, + 364, + 482, + 470 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Expert-Correction Mixture. We mix the expert demonstration and correction data for co-training the motion prediction model.", + "- Expert-Correction Mixture with Self-Reflection. We mix the expert demonstration and correction data for constraining a unified model to provide initial motion instruction and adjust the instruction." + ], + "bbox": [ + 89, + 472, + 482, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Table 2, the results show that co-training with mixture data yields superior performance compared to models trained exclusively on expert demonstration data. This indicates that combining various types of feedback data can enhance the decision-making and perception capabilities of MLLMs. It also validates the viability of our approach to achieving self-improvement through interaction.", + "bbox": [ + 88, + 579, + 482, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, the mixture training model with self-reflection performs better than the one without self-reflection, which suggests that our designed motion-based self-reflection method can enhance the decision-making capabilities of robots and facilitate the correction of fine-grained actions.", + "bbox": [ + 88, + 686, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, we find that utilizing the mixture of data to train a unified model to serve as both the motion prediction module and motion correction module fails to provide accurate correction information compared to our separated motion correction module. This suggests that the mixture training strategy may not fully leverage the strengths of each dataset to achieve better correction effects under the significant data scale discrepancies (160,000 expert demonstrations vs. 16,000 feedback data). The results indicate", + "bbox": [ + 88, + 763, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that our dual-process motion adjustment mechanism can effectively leverage the expert demonstration and correction dataset, leading to accurate motion instruction adjustment.", + "bbox": [ + 511, + 364, + 903, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also provide ablation results of our designed motion codebook in Supp.B.", + "bbox": [ + 511, + 410, + 903, + 440 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. The Performance of Lifelong Learning", + "text_level": 1, + "bbox": [ + 511, + 450, + 846, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we explore whether our Phoenix framework can facilitate lifelong learning through interactions. Concretely, we deploy the motion self-reflection model to interact within the environments and utilize the successful trajectories to iteratively fine-tune our motion prediction model after 10, 30, and 50 rollouts. To avoid catastrophic forgetting, we combine 20 expert demonstrations to co-fine-tune the motion prediction module.", + "bbox": [ + 511, + 473, + 905, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the lifelong learning ability of our motion-based self-reflection model and subgoal-based self-reflection model. During testing, we record the average success rate over 50 trials. As shown in Figure 4, the subgoal-based lifelong learning fails to enhance model performance during the exploration phase due to its inability to provide fine-grained action correction. In contrast, our method corrects underlying action execution during interactions, allowing the robot to better learn from the refined trajectories, thereby achieving self-improvement.", + "bbox": [ + 511, + 595, + 905, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Generalization to Novel Tasks", + "text_level": 1, + "bbox": [ + 511, + 756, + 779, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate the generalization ability of our Phoenix framework in color disruption and position disruption novel tasks as shown in Figure 5. In the color disruption setting, we replace the red block with the blue block in the Stack_D0 task to verify whether our model could generalize to object manipulation tasks with different visual characteristics. In the position disruption setting, we change the fixed position of the coffee machine to a randomly placed posi", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6987", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff7531bd49dda9526510e06ff9a40c8e908dea632532c4a17bd62d929b7551d5.jpg", + "image_caption": [ + "(a) Color Disruption in Stack_D0" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 313, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e9b8d47a72d73e864fbab79799a0e51a61d4c2e4e1eff417e148c8dd8f66c9b.jpg", + "image_caption": [ + "(b) Position Disruption in Coffee_D0" + ], + "image_footnote": [], + "bbox": [ + 334, + 88, + 542, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18833c50dbe49e46ad8330bd52e6dea83265d86fb26e7537645d67b1bf3565ae.jpg", + "image_caption": [ + "(c) Performance on Novel Tasks" + ], + "image_footnote": [], + "bbox": [ + 558, + 89, + 903, + 185 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce0248ba09f15b6805f14b98a2d013596f17a668fba7d5b3266157b033ca99b7.jpg", + "image_caption": [ + "Figure 5. In the color disruption setting, we replace the red block with the blue block in the Stack_D0 task as shown in (a). In the position disruption setting, we change the position of the coffee machine from a fixed point $\\circ$ to a random position from the rectangle $\\square$ in the Coffee_D0 task as illustrated in (b). The results in (c) prove that our framework could generalize well to these novel task settings.", + "(a) Expert Demonstration" + ], + "image_footnote": [], + "bbox": [ + 96, + 263, + 537, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5ce8665994ad26c28bd8bec81a8ff2d3b3962a9d9339aed37c0aa3bb06e6dc76.jpg", + "image_caption": [ + "(b) In-Distribution", + "Figure 6. The real-world experiments with different variations." + ], + "image_footnote": [], + "bbox": [ + 96, + 325, + 187, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/024c4e6f4e7c4313979289d9106dc2e09a04924651f32266bf37c6fee355f761.jpg", + "image_caption": [ + "(c) Pose Disruption" + ], + "image_footnote": [], + "bbox": [ + 210, + 325, + 303, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf67ba9076b879daff6d955408fa819b7c6661dd39fff48b01a12572fd3222ca.jpg", + "image_caption": [ + "(d) Background Disruption" + ], + "image_footnote": [], + "bbox": [ + 328, + 325, + 419, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce12886b7cdfe6a0fcef35c7fd1eb1a26368a65c3e393dbf9588be3b08191c50.jpg", + "image_caption": [ + "(e) Texture Disruption" + ], + "image_footnote": [], + "bbox": [ + 444, + 325, + 535, + 378 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/af487ce48e3e96a076661ae88040af1cb18bcf7f3d8b2ca3aa480b589775fb44.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelIn-Dis.Pose Dis.Bg.Tex.
OpenVLA55%30%35%45%
Task60%25%25%45%
Motion60%35%30%50%
Ours75%55%45%65%
", + "bbox": [ + 586, + 265, + 901, + 338 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/213aa2ab43c3e53dddad730d8c7d8df044a8415001da7e972996d7bb71647ace.jpg", + "table_caption": [ + "Table 3. The real-world experiment results." + ], + "table_footnote": [], + "table_body": "
TaskMotion10 rollout30 rollout
In-Dis.60%65%75%
Pose Dis.35%45%50%
", + "bbox": [ + 586, + 354, + 883, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. The lifelong learning results", + "bbox": [ + 617, + 400, + 839, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion within a specific area in the Coffee_D0 task to verify whether our method could generalize to unseen scenarios.", + "bbox": [ + 89, + 431, + 480, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For these novel tasks, although the subgoal-conditioned policy could predict correct high-level semantic subgoal for manipulation, this method fails to predict precise robotic action to complete the tasks. Due to its limitation of providing high-level semantic correction information, the subgoal self-reflection method fails to effectively leverage the knowledge of MLLMs for action correction to manipulation tasks. In contrast, as shown in Figure 5(c), our motion-conditioned policy could generate fine-grained motion instruction to achieve generalizable manipulation benefiting from the perception and inferential capability of MLLMs. Besides, our method could achieve better performance on novel tasks by comprehensively refining the motion instruction with the motion-based self-reflection framework.", + "bbox": [ + 88, + 460, + 482, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Real-World Experiments", + "text_level": 1, + "bbox": [ + 89, + 681, + 318, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In real-world scenarios, we conduct the challenging \"drawer open\" articulated object manipulation task as shown in Fig 6(a), where the robot needs to align gripper with handle through precise rotations to open the drawer. We utilize the spacemouse device to collect 100 expert demonstrations with 14 motion instructions (e.g., \"move arm right\", \"rotate around x-axis\"). We train a motion-conditioned diffusion policy to convert instructions into robotic actions. During the inference, we introduce human-in-the-loop interventions to manually correct failure situations to collect 20 corresponding refined interaction trajectories to train our motion correction module. All models are only fine-tuned on the real-world data.", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate generalization, we design 4 settings as shown in Fig 6(b-e). In the pose disruption setting, we change the pose distribution of the drawer. For the background disruption setting, the background color was modified to green. In the texture disruption setting, the texture of the drawer was altered to evaluate performance under significant visual variations. The results in Tab 3 demonstrate the generalization ability of our method. We also evaluate lifelong learning, the results in Tab 4 show that our model achieves self-improvement in real world.", + "bbox": [ + 511, + 431, + 906, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also provide more real-world task experiments with a rule-based manipulation policy to prove the effectiveness of our motion-based self-reflection method in Supp.C.", + "bbox": [ + 511, + 582, + 905, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 640, + 633, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a motion-based self-reflection framework to convert the semantic reflection of MLLMs into fine-grained robotic action correction. Based on this framework, we further automatically improve the model's capability from interactions. We hope this motion-based self-reflection framework could bring insights for enhancing the generalization capabilities of agents in robotic manipulation tasks through the integration of MLLMs.", + "bbox": [ + 511, + 666, + 906, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 800, + 692, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The project was supported by National Natural Science Foundation of China (NO.62106272), Shanghai AI Laboratory, National Key R&D Program of China (2022ZD0160101) and CCF-Zhipu.AI Large Model Innovation Fund.", + "bbox": [ + 511, + 824, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6988", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andrew G Barto and Sridhar Mahadevan. Recent advances in hierarchical reinforcement learning. Discrete event dynamic systems, 13:341-379, 2003. 2", + "[2] Suneel Belkhale, Tianli Ding, Ted Xiao, Pierre Sermanet, Quon Vuong, Jonathan Tompson, Yevgen Chebotar, Debidatta Dwibedi, and Dorsa Sadigh. Rt-h: Action hierarchies using language. arXiv preprint arXiv:2403.01823, 2024. 3, 6", + "[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3, 6", + "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3", + "[5] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 3", + "[6] Yinpei Dai, Jayjun Lee, Nima Fazeli, and Joyce Chai. Racer: Rich language-guided failure recovery policies for imitation learning, 2024. 2", + "[7] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3, 6", + "[8] John H Flavell. Metacognition and cognitive monitoring: A new area of cognitive-developmental inquiry. American psychologist, 34(10):906, 1979. 1", + "[9] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for generalizable manipulation skills. In International Conference on Learning Representations, 2023. 2", + "[10] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado Van Hasselt, and David Silver. Distributed prioritized experience replay. arXiv preprint arXiv:1803.00933, 2018. 2", + "[11] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3, 6", + "[12] Daeyeol Lee, Hyojung Seo, and Min Whan Jung. Neural basis of reinforcement learning and decision making. Annual review of neuroscience, 35(1):287-308, 2012. 1", + "[13] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2, 3", + "[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 5", + "[16] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation. arXiv preprint arXiv:2405.17418, 2024. 2", + "[17] Kehui Liu, Zixin Tang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Coherent: Collaboration of heterogeneous multi-robot system with large language models. arXiv preprint arXiv:2409.15146, 2024. 2", + "[18] Qian Liu, Xiaosen Zheng, Niklas Muennighoff, Guangtao Zeng, Longxu Dou, Tianyu Pang, Jing Jiang, and Min Lin. Regmix: Data mixture as regression for language model pretraining. arXiv preprint arXiv:2407.01492, 2024. 7", + "[19] Jingxian Lu, Wenke Xia, Dong Wang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Koi: Accelerating online imitation learning via hybrid key-state guidance. arXiv preprint arXiv:2408.02912, 2024. 1", + "[20] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. In Conference on Robot Learning (CoRL), 2021. 2, 5", + "[21] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 3", + "[22] Xincheng Pang, Wenke Xia, Zhigang Wang, Bin Zhao, Di Hu, Dong Wang, and Xuelong Li. Depth helps: Improving pre-trained rgb-based policy with depth information injection. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7251-7256. IEEE, 2024. 3", + "[23] Shubham Pateria, Budhitama Subagdja, Ah-hwee Tan, and Chai Quek. Hierarchical reinforcement learning: A comprehensive survey. ACM Computing Surveys (CSUR), 54(5): 1-35, 2021. 2", + "[24] Shreyas Sundara Raman, Vanya Cohen, Eric Rosen, Ifrah Idrees, David Paulius, and Stefanie Telsex. Planning with large language models via corrective re-prompting. In NeurIPS 2022 Foundation Models for Decision Making Workshop, 2022. 2", + "[25] Roger C Schank. What we learn when we learn by doing. Technical report, Technical report, 1995. 1", + "[26] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 1, 2", + "[27] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6989", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 3, 6", + "[28] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[29] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 2", + "[30] Zihao Wang, Shaofei Cai, Guanzhou Chen, Anji Liu, Xiaojian Ma, and Yitao Liang. Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560, 2023. 1, 2", + "[31] Wenke Xia, Dong Wang, Xincheng Pang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Kinematic-aware prompting for generalizable articulated object manipulation with llms. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2073-2080, 2024. 2", + "[32] Chuyan Xiong, Chengyu Shen, Xiaqi Li, Kaichen Zhou, Jiaming Liu, Ruiping Wang, and Hao Dong. Aic mllm: Autonomous interactive correction mllm for robust robotic manipulation. arXiv preprint arXiv:2406.11548, 2024. 2", + "[33] Jiasheng Ye, Peiju Liu, Tianxiang Sun, Yunhua Zhou, Jun Zhan, and Xipeng Qiu. Data mixing laws: Optimizing data mixtures by predicting language modeling performance. arXiv preprint arXiv:2403.16952, 2024. 7", + "[34] Naoki Yokoyama, Alex Clegg, Joanne Truong, Eric Undersander, Tsung-Yen Yang, Sergio Arnaud, Sehoon Ha, Dhruv Batra, and Akshara Rai. Asc: Adaptive skill coordination for robotic mobile manipulation. IEEE Robotics and Automation Letters, 9(1):779-786, 2023. 2", + "[35] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 1, 2", + "[36] Jia Zeng, Qingwen Bu, Bangjun Wang, Wenke Xia, Li Chen, Hao Dong, Haoming Song, Dong Wang, Di Hu, Ping Luo, et al. Learning manipulation by predicting interaction. arXiv preprint arXiv:2406.00439, 2024. 3", + "[37] Junjie Zhang, Chenjia Bai, Haoran He, Wenke Xia, Zhigang Wang, Bin Zhao, Xiu Li, and Xuelong Li. Sam-e: leveraging visual foundation model with sequence imitation for embodied manipulation. arXiv preprint arXiv:2405.19586, 2024. 3", + "[38] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2" + ], + "bbox": [ + 91, + 90, + 483, + 838 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6990", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/497d038e-fa23-45f9-aee3-d3bfa187eaf5_model.json b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/497d038e-fa23-45f9-aee3-d3bfa187eaf5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..502a088e24030a84e87942339e65642d99c9732f --- /dev/null +++ b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/497d038e-fa23-45f9-aee3-d3bfa187eaf5_model.json @@ -0,0 +1,2431 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "image", + "bbox": [ + 0.205, + 0.123, + 0.232, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.23, + 0.135, + 0.8, + 0.178 + ], + "angle": 0, + "content": "Phoenix: A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.214, + 0.722, + 0.234 + ], + "angle": 0, + "content": "Wenke Xia\\(^{1,2,*}\\), Ruoxuan Feng\\(^{1}\\), Dong Wang\\(^{2}\\), Di Hu\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.24, + 0.812, + 0.276 + ], + "angle": 0, + "content": "\\(^{1}\\)Gaoling School of Artificial Intelligence, Renmin University of China, Beijing \\(^{2}\\) Shanghai AI Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.311, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.344, + 0.485, + 0.782 + ], + "angle": 0, + "content": "Building a generalizable self-correction system is crucial for robots to recover from failures. Despite advancements in Multimodal Large Language Models (MLLMs) that empower robots with semantic reflection ability for failure, translating semantic reflection into \"how to correct\" fine-grained robotic actions remains a significant challenge. To address this gap, we build the Phoenix framework, which leverages motion instruction as a bridge to connect high-level semantic reflection with low-level robotic action correction. In this motion-based self-reflection framework, we start with a dual-process motion adjustment mechanism with MLLMs to translate the semantic reflection into coarse-grained motion instruction adjustment. To leverage this motion instruction for guiding \"how to correct\" fine-grained robotic actions, a multi-task motion-conditioned diffusion policy is proposed to integrate visual observations for high-frequency robotic action correction. By combining these two models, we could shift the demand for generalization capability from the low-level manipulation policy to the MLLMs-driven motion adjustment model and facilitate precise, fine-grained robotic action correction. Utilizing this framework, we further develop a lifelong learning method to automatically improve the model's capability from interactions with dynamic environments. The experiments conducted in both the RoboMimic simulation and real-world scenarios prove the superior generalization and robustness of our framework across a variety of manipulation tasks. Our code is released at https://github.com/GeWuLab/Motion-based-Self-Reflection-Framework." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.8, + 0.222, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.826, + 0.466, + 0.857 + ], + "angle": 0, + "content": "\"Failure is simply the opportunity to begin again, this time more intelligently.\" — Henry Ford" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.312, + 0.885, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.73 + ], + "angle": 0, + "content": "Figure 1. Our motion-based self-reflection framework utilizes coarse-grained motion instruction as a bridge to convert the high-level semantic reflection into fine-grained robotic action correction, thereby facilitating generalizable and precise action correction with perceptual and inferential capabilities of MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.733, + 0.907, + 0.884 + ], + "angle": 0, + "content": "Humans are naturally equipped with the ability to correct their behaviors by intentionally reflecting on actions that lead to failure. [12, 25]. By analyzing failure situations from high-level semantic reflection and low-level action correction perspective, humans can efficiently adapt to dynamic environments [8]. To emulate the correction capability and foster a continuous cycle of self-improvement in robots, researchers [26, 30, 35] have sought to develop self-reflection systems that enable robots to recover from and learn through their failure interactions." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Among them, some existing self-correction systems [19," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.424, + 0.888 + ], + "angle": 0, + "content": "*Work is done during internship at Shanghai AI Laboratory" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.424, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6981" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.332 + ], + "angle": 0, + "content": "26, 34] leverage reinforcement learning to guide robots in correct low-level action execution through reward functions. However, the reliance on reinforcement learning limits the ability of these self-correction systems to generalize across long-horizon manipulation tasks, due to unstable training processes [10] and the need for task-specific prior knowledge [9, 35]. To construct a generalizable and stable self-correction system, recent works [17, 30, 31] borrow the inferential capability of Multi-modal Large Language Models (MLLMs) to propose closed-loop high-level semantic reflection framework for failure correction. Although these semantic self-reflection frameworks can decompose the failure correction process into semantic subgoals, they primarily rely on a predefined skill library to execute the detailed subgoals, which fails to utilize the generalization ability of MLLMs in fine-grained robotic action correction." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.337, + 0.483, + 0.609 + ], + "angle": 0, + "content": "To maximize the generalization potential of MLLMs for action correction, we propose motion instruction as a bridge to convert high-level semantic reflection to fine-grained robotic action correction. Motion instruction refers to coarse-grained robotic movement commands such as \"move arm backward\" and \"adjust gripper position\". Serving as an intermediate layer, motion instruction could provide general, low-frequency decision information for high-frequency robotic action execution, which makes it an excellent medium for embedding the knowledge of MLLMs into fine-grained action correction. As shown in Figure 1, we decompose the semantic reflection knowledge into coarse-grained motion instruction adjustment to indicate \"how to correct\" fine-grained action for low-level policy execution. This transition shifts the perceptual and decision-making requirements from low-level robotic policy to the MLLMs-driven motion adjustment model, thereby enabling generalizable, fine-grained robotic action correction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Hence, in this work, we build the Phoenix framework, a motion-based self-reflection framework designed to convert the semantic reflection of MLLMs into fine-grained robotic action correction. Initially, we develop a dual-process motion adjustment mechanism to ensure efficient prediction through a motion prediction module, while addressing failure with a motion correction module. Concretely, we first utilize expert demonstration trajectories to train the motion prediction module for efficient motion instruction generation. Despite its efficiency in generating initial instructions, this module often struggles to handle failure scenarios. To recover from failures, we collect a comprehensive failure correction dataset and fine-tune the motion correction module, which thoroughly provides adjusted motion instructions through a chain-of-thought approach. By integrating these two modules, the dual-process motion adjustment mechanism guarantees both robustness and efficiency, facilitating the generation of accurate motion instructions. As the coarse-grained motion instructions only provide gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "eral and low-frequency guidance for robotic manipulation, we further design a multi-task motion-conditioned diffusion policy that integrates visual observations to translate motion instruction into precise, high-frequency action corrections for manipulation tasks. Moreover, by leveraging these correction trajectories, we propose a lifelong learning method that iteratively enhances the model's capabilities through interaction, ensuring continuous improvements in performance and adaptability to dynamic environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.394 + ], + "angle": 0, + "content": "To validate the efficacy of our framework, we conduct experiments across 9 contact-rich robotic manipulation tasks within the RoboMimic simulation [20]. The results demonstrate that our method could provide more precise action correction from failures through self-reflection and facilitate self-improvement through interactions with environments. Further, we conduct two novel manipulation tasks with color disruption and position distribution disruption, proving the generalization ability of our framework. The real-world experiments also demonstrate the applicability and robustness of our approach in practical scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.655, + 0.424 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.435, + 0.794, + 0.451 + ], + "angle": 0, + "content": "2.1. Robotic Self-correction Systems" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.905, + 0.821 + ], + "angle": 0, + "content": "Self-correction serves as a crucial mechanism enabling robots to recover from failures. To achieve self-correction on the low-level robotic action, Reinforcement Learning [1, 23] is proposed to guide robots in adjusting behaviors via reward signals. However, reinforcement learning strategies encounter difficulties in intricate robotic environments primarily due to learning inefficiency. Borrowing the common knowledge of MLLMs, Raman et al. [24] proposes to build a semantic self-reflection system for long-horizon task planning. To facilitate interaction with environments, some efforts generate robotic action through simulation APIs [17, 29, 30] and predefined action skill libraries [14, 16]. However, the reliance on predefined low-level skill libraries makes semantic self-reflection frameworks fail to directly provide fine-grained action correction. To provide low-level action feedback for robotic manipulation, recent works [6, 16, 32] suggest adjusting end-effector poses to refine actions. Nonetheless, these techniques are primarily restricted to simple manipulation tasks that employ motion planning and fail to generalize to contact-rich manipulation scenarios. In this work, we utilize motion instruction as an intermediate layer to guide robotic action correction, borrowing the perceptual and inferential capability of MLLMs for fine-grained robotic action correction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.771, + 0.848 + ], + "angle": 0, + "content": "2.2. Robotic Manipulation Policy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The development of generalizable strategies for robotic manipulation remains a persistent challenge in robotics research. ACT [38] is proposed to predict action sequences" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6982" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.904, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.286, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Figure 2. The pipeline of our motion-based self-reflection framework. (a) demonstrates the dual-process motion refinement mechanism which leverages the motion prediction module for efficient motion instruction prediction and motion correction module for comprehensive failure correction. (b) illustrates the motion-conditioned diffusion policy which converts the low-frequency motion instruction guidance into high-frequency robotic action. The lifelong learning approach in (c) iteratively enhances the ability of the motion prediction module from the refined interaction trajectories." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.358, + 0.485, + 0.644 + ], + "angle": 0, + "content": "to ensure temporal alignment of actions, while the Diffusion Policy [5] addresses multi-modal action distributions to enhance robust manipulation. Driven by the development of foundation models [13, 28], recent works [4, 7, 14, 22, 36, 37] leverage the world knowledge of MLLMs to facilitate task decomposition and planning for robotic manipulation. Concurrently, other works [3, 11, 21, 27] collect large-scale robotic manipulation demonstrations to train generalizable language-conditioned manipulation policy across different robots and tasks. Moreover, RT-H [2] employs detailed motion commands instead of semantic language inputs, fostering flexible and generalizable manipulation capabilities through the advanced perceptual abilities of MLLMs. However, the prohibitive costs associated with robotic data collection limit the scalability of existing imitation learning approaches. To address this, we propose the motion-based self-reflection framework to enable autonomous self-improvement through continuous interaction with environments without human intervention." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.657, + 0.462, + 0.673 + ], + "angle": 0, + "content": "3. Motion-based Self-Reflection Framework" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.465, + 0.697 + ], + "angle": 0, + "content": "3.1. Challenges in Robotic Self-Correction Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Building a generalizable and robust self-correction system is a key component in achieving failure correction for robots. Multi-modal Large Language Models (MLLMs) have already been applied to the construction of robotic self-reflection framework to recover from failures. However, existing systems mainly focus on semantic reflection, and their application to fine-grained action correction still faces the following two issues:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.855 + ], + "angle": 0, + "content": "- How to enable MLLMs to understand manipulation tasks and provide detailed correction information?" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.885 + ], + "angle": 0, + "content": "- How to convert the correction information provided by MLLMs into precise, high-frequency robotic actions?" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.826, + 0.484, + 0.885 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address these issues, we propose the Phoenix frame" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.538 + ], + "angle": 0, + "content": "work, a motion-based self-reflection framework that integrates a dual-process motion adjustment mechanism and a multi-task motion-conditioned diffusion policy as illustrated in Figure 2. As detailed in Section 3.2, the dual-process motion adjustment mechanism is developed to ensure efficient and accurate motion instruction generation. Further, the motion-conditioned diffusion policy is proposed to translate the coarse-grained motion instructions into precise robotic actions, as explained in Section 3.3. Based on the refined manipulation trajectories, we propose a lifelong learning approach to facilitate robotic self-improvement, as outlined in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.546, + 0.899, + 0.562 + ], + "angle": 0, + "content": "3.2. Dual-process Motion Adjustment Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The dual-process motion adjustment mechanism is designed to ensure efficient motion prediction through a motion prediction module, while comprehensively addressing failure with a motion correction module. Given the observation \\( O \\) and task description \\( T \\), we first train a Motion Prediction Module (MPM) with expert demonstration dataset \\( D_{e} \\) to generate initial motion instruction \\( m_{i} \\). However, the MPM trained on expert demonstrations struggles to handle failure situations. Thus, we construct a comprehensive failure correction dataset \\( D_{c} \\) to fine-tune the Motion Correction Module (MCM), enabling it to analyze the failure situation and adjust \\( m_{i} \\) with a chain-of-thought approach. If \\( m_{i} \\) is deemed correct, we adopt it as the decision motion instruction \\( m_{d} \\) for further robotic action prediction. Otherwise, we employ the MCM to analyze the failure situation and generate adjusted motion instruction \\( m_{a} \\) as decision motion instruction \\( m_{d} \\). Through the guidance of \\( m_{d} \\), our motion-based diffusion policy can generate high-frequency corrections to the robotic actions. As described in Algorithm 1, we establish the dual-process motion adjustment mechanism to guarantee the efficiency and accuracy of motion instruction generation for fine-grained robotic action prediction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6983" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.092, + 0.09, + 0.482, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Self-Reflection w/. Dual-Process Motion Adj." + }, + { + "type": "code", + "bbox": [ + 0.092, + 0.111, + 0.484, + 0.171 + ], + "angle": 0, + "content": "Require: Task description \\(T\\), Observation \\(O\\), Environment \\(E\\), Motion prediction module \\(MPM\\), Motion correction module \\(MCM\\), Motion-conditioned diffusion policy \\(\\pi\\), Exploration timestep \\(K\\)," + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.172, + 0.235, + 0.186 + ], + "angle": 0, + "content": "1: \\(O_{1}\\gets E.\\text{reset} ()\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.188, + 0.245, + 0.2 + ], + "angle": 0, + "content": "2: for \\( k = 1 \\) to \\( K \\) do" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.202, + 0.292, + 0.216 + ], + "angle": 0, + "content": "3: \\(m_{i}\\gets MPM(O_{k},T)\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.217, + 0.478, + 0.232 + ], + "angle": 0, + "content": "4: failure_flag, semantic_info \\(\\leftarrow\\) MCM(Ok,m)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.233, + 0.336, + 0.246 + ], + "angle": 0, + "content": "5: if failure_flag is true then" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.247, + 0.402, + 0.262 + ], + "angle": 0, + "content": "6: \\(m_{a}\\gets MCM(O_{k},\\text{semantic\\_info})\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.265, + 0.241, + 0.276 + ], + "angle": 0, + "content": "7: \\(m_{d}\\gets m_{a}\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.278, + 0.174, + 0.29 + ], + "angle": 0, + "content": "8: else" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.295, + 0.239, + 0.306 + ], + "angle": 0, + "content": "9: \\(m_{d}\\gets m_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.308, + 0.188, + 0.32 + ], + "angle": 0, + "content": "10: end if" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.323, + 0.245, + 0.337 + ], + "angle": 0, + "content": "11: \\(a\\gets \\pi (O,m_d)\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.338, + 0.276, + 0.353 + ], + "angle": 0, + "content": "12: \\(O_{k + 1}\\gets E.\\text{step}(a)\\)" + }, + { + "type": "list", + "bbox": [ + 0.098, + 0.323, + 0.276, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.354, + 0.174, + 0.365 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.485, + 0.748 + ], + "angle": 0, + "content": "Motion Prediction Module (MPM). To fully harness the perceptual and decision-making capabilities of MLLMs for efficient motion instruction prediction, we develop a motion instruction dataset from the expert demonstrations dataset \\( D_{e} \\) to fine-tune MLLMs for robotic manipulation tasks. To construct the expert dataset, we filter the robotic action to get dominant motion from expert demonstration with a threshold, generating a set of motion instructions that include arm direction and gripper control. In practice, we find that separating arm direction instruction and gripper control instruction would cause the misalignment between textual motion instruction and fine-grained robotic action. To address this issue, we combine the direction movement with gripper control, resulting in unified instruction formats for motion instruction such as \"move arm right with gripper closed\". In addition, we add the \"make slight adjustments to gripper position\" instruction to model the minor robotic actions below the threshold. Through the automatic construction method, we build 37 types of motion instructions as guidance for further robotic action prediction. By training on the expert dataset, the MPM acquires an understanding of robotic manipulation tasks and could efficiently generate an initial motion instruction \\( m_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Motion Correction Module (MCM). During interactions with the environment, robots may execute incorrect actions, leading to a failure situation in the task. However, the MPM trained on success expert data often struggles to recover from these failure scenarios. Thus, we develop the motion correction module, to identify failure scenarios and correct behaviors from such situations. As shown in Figure 2(a), the MCM would evaluate the initial motion instruction \\( m_{i} \\) and conduct a dual process based on the evaluation results to achieve efficient and accurate motion in-" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.09, + 0.885, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.15, + 0.769, + 0.158 + ], + "angle": 0, + "content": "(a) Online Human Interventions" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.168, + 0.884, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.229, + 0.768, + 0.238 + ], + "angle": 0, + "content": "(b) Offline Human Annotations" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.248, + 0.884, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.308, + 0.758, + 0.317 + ], + "angle": 0, + "content": "(c) Expert Demonstrations" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.359 + ], + "angle": 0, + "content": "Figure 3. Illustrations of our correction data: online human interventions, offline human annotations, and expert demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.906, + 0.483 + ], + "angle": 0, + "content": "struction adjustment. Once encountering the failure situation, the MCM would first analyze the type of failure and derive a semantic-level correction goal, such as \"insert the coffee pot into the coffee machine\". Based on this correction goal, the MCM further adjusts the motion instructions with its learned failure-correction knowledge, ultimately generating an accurate motion instruction through a hierarchical chain-of-thought approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.906, + 0.56 + ], + "angle": 0, + "content": "To equip MCM with the capabilities for failure detection and correction, we construct a comprehensive correction dataset as illustrated in Figure 3. This dataset includes three types of feedback data, encompassing both semantic and motion perspectives, and is categorized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.564, + 0.907, + 0.715 + ], + "angle": 0, + "content": "- Online Human Intervention. We implement the human-in-the-loop method for trajectory collection. We first deploy the motion prediction model to interact with environments, then manually intervene to correct the motion instructions whenever the agent encounters failure situations. This method could collect accurate and high-quality motion correction data to ensure task completion. However, it requires frequent manual interactions with the environment, which leads to significant time consumption and makes it difficult to collect large-scale data." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.715, + 0.906, + 0.821 + ], + "angle": 0, + "content": "- Offline Human Annotation. We utilize the motion prediction model to gather trajectory data, periodically sampling trajectories and annotating them with semantic reflections and motion correction details. While the accuracy of offline annotated data cannot be guaranteed due to its inability to interact with environments for verification, this method offers a significant volume of annotations." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.821, + 0.905, + 0.882 + ], + "angle": 0, + "content": "- Expert Demonstration. We automate annotations on expert trajectories. Since these trajectories are successful, this data is used to provide accurate motion information to enhance the model's motion prediction capabilities." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.564, + 0.907, + 0.882 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "By fine-tuning MCM on this dataset, we enhance the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6984" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "MCM to thoroughly comprehend various types of failure situations and provide motion instruction corrections. Through the integration of MPM and MCM, the dual-process motion adjustment mechanism enables efficient motion instruction generation while ensuring comprehensive correction in failure situations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.193, + 0.406, + 0.21 + ], + "angle": 0, + "content": "3.3. Motion-conditioned Diffusion Policy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.484, + 0.337 + ], + "angle": 0, + "content": "As the motion instruction only provides general and low-frequency guidance for manipulation, we train a multi-task motion-conditioned diffusion policy \\(\\pi\\) to convert the motion instructions into precise, high-frequency robotic actions. This policy takes observations \\(O\\) and decision motion instructions \\(m_{d}\\) to output robotic actions \\(a\\). To ensure the policy adheres to the motion instruction, we make adjustments as depicted in Figure 2(b):" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.484, + 0.474 + ], + "angle": 0, + "content": "First, we observe that existing pre-trained language models often struggle to capture the discriminative features of various motion instructions. This limitation hampers their ability to follow various motion instructions. To address this issue, we introduce a learnable motion codebook designed to provide discriminative features for motion instructions. For a given decision motion instruction \\( m_d \\), the codebook would retrieve the corresponding motion feature to facilitate accurate robotic action prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.625 + ], + "angle": 0, + "content": "Further, we find that the direct concatenation of observation representation and motion instruction feature would cause the diffusion policy to prefer to rely on the vision information for action prediction, thereby hindering the effectiveness of the motion instruction guidance. To address this issue, we take the observation representation and motion instruction feature as separate conditions in different stages of the diffusion policy, allowing the model to better learn the guidance information from the motion instruction and thereby promote precise action correction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.657 + ], + "angle": 0, + "content": "By integrating these two adjustments, we train the diffusion policy for action prediction with the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.668, + 0.483, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\operatorname {M S E} \\left(\\mathcal {E} ^ {k}, \\pi (\\mathcal {O}, \\mathcal {M}, \\mathcal {A} ^ {0} + \\mathcal {E} ^ {k}, k)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.777 + ], + "angle": 0, + "content": "where \\(\\mathcal{O}\\) is the observation representation, \\(\\mathcal{M}\\) is the motion instruction feature, \\(\\mathcal{A}^0\\) is the ground truth robotic action, \\(\\mathcal{E}^k\\) indicates the random noise at the denoising iteration \\(k\\). Through minimizing the loss function in Eq 1, the diffusion policy \\(\\pi\\) could effectively predict precise, high-frequency robotic action guided by motion instruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.438, + 0.803 + ], + "angle": 0, + "content": "3.4. Action Correction for Lifelong Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The dual-process motion adjustment mechanism leverages the MPM to efficiently predict motion instructions and the MCM to adjust them with a comprehensive chain-of-thought approach. However, the reliance on the chain-of-thought poses challenges in adapting to real-time scenarios due to its time-consuming. Furthermore, the collection of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.197 + ], + "angle": 0, + "content": "manipulation data and correction data is exceedingly labor-intensive. Thus, we propose a lifelong learning method that equips the MPM with both motion prediction and failure correction capabilities through learning from the refined interaction trajectories as illustrated in Figure 2(c), which enhances our model to adapt and react quickly to the environment without human intervention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.906, + 0.395 + ], + "angle": 0, + "content": "Benefiting from the motion-conditioned diffusion policy which could adhere to the motion instruction to generate task-aware robotic action, we can enhance the robot's capabilities through only improving the MPM informed by the refined interaction trajectory. To address the issue of catastrophic forgetting, we mix the refined interaction trajectory with expert demonstration for co-fine-tuning, allowing the model to simultaneously learn failure correction and enhance the motion prediction capabilities. Through updates from refined interaction trajectories, our model can achieve self-improvement by learning from the knowledge of the motion correction module, achieving fast and accurate manipulation for contact-rich manipulation tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.407, + 0.646, + 0.423 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.432, + 0.905, + 0.462 + ], + "angle": 0, + "content": "To comprehensively evaluate our framework, we propose experiments to answer the following questions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.463, + 0.905, + 0.491 + ], + "angle": 0, + "content": "- Does our motion-guided self-reflection model enhance the precision of action correction? Section 4.2" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.492, + 0.905, + 0.52 + ], + "angle": 0, + "content": "- Can our model achieve lifelong learning from interaction with environments? Section 4.3" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.522, + 0.905, + 0.55 + ], + "angle": 0, + "content": "- Does our framework can generalize across novel tasks? Section 4.4" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.552, + 0.905, + 0.581 + ], + "angle": 0, + "content": "- Can our framework ensure reliability and robustness in real-world scenarios? Section 4.5" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.463, + 0.905, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.707, + 0.607 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we conduct experiments on 9 contact-rich manipulation tasks in RoboMimic [20], ranging from long-horizon tasks like ThreePieceAssembly to fine-grained manipulation tasks like Threading. To transform high-level semantic information into motion instructions, we filter expert demonstrations to obtain over 160,000 pairs of motion instructions and observations. The dataset includes 37 types of motion instructions, which are utilized to fine-tune the LLaVA-v1.5 model [15] as the motion prediction module. Furthermore, to develop the motion correction module that integrates semantic comprehension and motion instruction adjustment, we collect correction data comprising 3,644 online human intervention data, 7,365 offline human annotations, and 6,378 expert demonstrations. We filter out the correction dataset to balance the proportion of various failure situations to enhance the model's correction capabilities. Ultimately, to translate motion instructions into precise robotic actions, we train a multi-task motion-conditioned diffusion policy using a learnable motion instruction code" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6985" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.181 + ], + "angle": 0, + "content": "
MethodsCoffee_D0Coffee_D1Stack_D0Stack_D1StackThree_D0StackThree_D1Threading_D0ThreePieceAssembly_D0ThreePieceAssembly_D1Mean
OpenVLA [11]42%18%84%86%36%20%20%28%8%38.0%
Task-conditioned66%24%88%68%30%6%74%20%0%41.8%
Subgoal-conditioned76%26%88%74%24%6%78%20%2%43.8%
Motion-conditioned68%32%92%84%38%16%58%30%4%46.9%
Subgoal Self-reflection80%32%88%78%32%6%80%34%2%48.0%
Phoenix (Ours)94%48%96%86%50%20%68%52%6%57.8%
Human Intervention (Oracle)100%100%100%90%70%40%100%70%40%78.9%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.191, + 0.907, + 0.22 + ], + "angle": 0, + "content": "Table 1. Comparison experiments results across 9 manipulation tasks in RoboMimic Simulation. The results demonstrate that our motion-based self-reflection method achieves better performance by facilitating precise correction of fine-grained robotic actions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.233, + 0.482, + 0.355 + ], + "angle": 0, + "content": "book, incorporating 500 demonstrations per task. During inference in simulation, our dual-process motion adjustment mechanism would provide motion instruction at \\(5Hz\\), and the diffusion policy would extend the motion instruction with visual observations to a \\(20Hz\\) action sequence to control the robot. For each task, we conducted 50 trials and report the average success rate. More implementation details could refer to Supp.A." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.366, + 0.476, + 0.381 + ], + "angle": 0, + "content": "4.2. Performance of Motion Self-Reflection Model" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.39, + 0.286, + 0.404 + ], + "angle": 0, + "content": "4.2.1 Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.483, + 0.505 + ], + "angle": 0, + "content": "To evaluate our motion-based self-reflection framework, we compare our framework with other approaches. To ensure fairness, all our comparison methods are trained on the expert data from the simulation environment, with the decision model using LLaVA-v1.5 and the underlying policy employing a diffusion policy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.508, + 0.483, + 0.539 + ], + "angle": 0, + "content": "- OpenVLA [11]. We fine-tune the OpenVLA model to provide baseline performance for multi-task experiments." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.539, + 0.483, + 0.584 + ], + "angle": 0, + "content": "- Task-conditioned policy. We take the task description as the condition for diffusion policy without the reflection framework, as a variance of RT-1 [3] and Octo [27]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.584, + 0.483, + 0.675 + ], + "angle": 0, + "content": "- Subgoal-conditioned policy. We fine-tune a LLaVAv1.5 to predict subgoals at \\(5\\mathrm{Hz}\\), which are utilized as the condition for diffusion policy without reflection framework. This method borrows the semantic comprehension capabilities of MLLMs, and is implemented as a variance of PaLM-E [7] with an individual diffusion policy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.675, + 0.483, + 0.78 + ], + "angle": 0, + "content": "- Motion-conditioned policy. We fine-tune a LLaVA-v1.5 as the motion prediction model to provide motion instructions at 5Hz, using these predictions to condition the diffusion policy without the reflection framework. This method employs the perceptual and inferential capacities of MLLMs, realized as a variation of RT-H [2] with an individual diffusion policy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.483, + 0.855 + ], + "angle": 0, + "content": "- Human Intervention. We manually correct the wrong motion instructions for the motion-conditioned policy. This method provides an upper bound on the performance of self-reflection methods. Due to labor costs, the results are presented as average success rates across 10 trials." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Subgoal Self-reflection. We fine-tune a LLaVA-v1.5 as subgoal self-reflection model and apply it to the subgoal-condition policy. This method is designed to validate the" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.508, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.234, + 0.864, + 0.248 + ], + "angle": 0, + "content": "effectiveness of the semantic self-reflection model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.906, + 0.342 + ], + "angle": 0, + "content": "As shown in Table 1, we first compare three different condition methods. Borrowing the perception and inferential ability of MLLMs, the subgoal-conditioned and motion-conditioned policies are better than the task-conditioned policies. The results prove the potential applications of MLLMs in various complex robotic manipulation tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.905, + 0.433 + ], + "angle": 0, + "content": "Focusing on specific tasks, we observe that the motion-conditioned policy excels in long-horizon tasks such as StackThree_D0 and ThreePieceAssembly_D0. However, this policy depends on consistent and accurate motion instruction predictions, which poses challenges in fine-grained manipulation tasks like Threading_D0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.905, + 0.509 + ], + "angle": 0, + "content": "By providing correction subgoals, the subgoal self-reflection method consistently outperforms the subgoal-conditioned policy, particularly in long-horizon manipulation tasks such as \"StackThree_D0\", which demonstrates the efficacy of the self-reflection framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.906, + 0.6 + ], + "angle": 0, + "content": "The OpenVLA model demonstrates strong performance in certain long-horizon tasks, leveraging its end-to-end action token prediction capability. However, the lack of observation history and action chunking poses significant challenges in handling complex, fine-grained manipulation tasks like Threading_D0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.906, + 0.827 + ], + "angle": 0, + "content": "Notably, our Phoenix method achieves more substantial improvements than the subgoal self-reflection method, demonstrating the effectiveness of motion-conditioned method in long-horizon sequential tasks and fine-grained manipulation tasks. Benefiting from our motion-based correction method, agent could correct fine-grained action through motion instruction adjustment while the subgoal-conditioned self-reflection model fails to recover from most failure situations. Besides, the human intervention method achieves high success rates across multiple tasks, demonstrating that our motion-conditioned diffusion policy can effectively adhere to motion instructions for manipulation tasks. This result indicates that our method can perform well under the correct motion instructions, showcasing the significant potential of motion-conditioned self-reflection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.685, + 0.861 + ], + "angle": 0, + "content": "4.2.2 Ablation Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a motion prediction module to provide initial motion instruction, and a motion correction" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6986" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.145 + ], + "angle": 0, + "content": "
MethodsCoffee_D0Coffee_D1Stack_D0Stack_D1StackThree_D0StackThree_D1Threading_D0ThreePieceAssembly_D0ThreePieceAssembly_D1Mean
Motion-conditioned68%32%92%84%38%16%58%30%4%46.9%
Expert-Correction Mixture74%36%94%86%38%22%64%30%2%49.6%
Expert-Correction Mixture with Self-Reflection76%30%92%90%46%26%64%34%4%51.3%
Phoenix (Ours)94%48%96%86%50%20%68%52%6%57.8%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.156, + 0.907, + 0.184 + ], + "angle": 0, + "content": "Table 2. The ablation results of our dual-process motion adjustment mechanism. The results prove that our model, which separates motion prediction module and motion correction module, can provide more precise motion adjustment for action execution in manipulation tasks." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.187, + 0.337, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.187, + 0.642, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.187, + 0.901, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.323, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Figure 4. The lifelong learning results. The results prove that our motion-based self-reflection method could iteratively improve performance through interactions with environments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.483, + 0.472 + ], + "angle": 0, + "content": "module to provide fine-grained motion correction. Drawing upon prior research [18, 33], data mixture proportions could influence the efficacy of LLMs. In this section, we investigate whether integrating expert demonstrations with correction dataset, could also enhance the perception and decision-making capabilities of MLLMs for robotic manipulation with the following ablation methods:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.473, + 0.483, + 0.517 + ], + "angle": 0, + "content": "- Expert-Correction Mixture. We mix the expert demonstration and correction data for co-training the motion prediction model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.518, + 0.483, + 0.579 + ], + "angle": 0, + "content": "- Expert-Correction Mixture with Self-Reflection. We mix the expert demonstration and correction data for constraining a unified model to provide initial motion instruction and adjust the instruction." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.473, + 0.483, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.58, + 0.483, + 0.687 + ], + "angle": 0, + "content": "As illustrated in Table 2, the results show that co-training with mixture data yields superior performance compared to models trained exclusively on expert demonstration data. This indicates that combining various types of feedback data can enhance the decision-making and perception capabilities of MLLMs. It also validates the viability of our approach to achieving self-improvement through interaction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.688, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Besides, the mixture training model with self-reflection performs better than the one without self-reflection, which suggests that our designed motion-based self-reflection method can enhance the decision-making capabilities of robots and facilitate the correction of fine-grained actions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "However, we find that utilizing the mixture of data to train a unified model to serve as both the motion prediction module and motion correction module fails to provide accurate correction information compared to our separated motion correction module. This suggests that the mixture training strategy may not fully leverage the strengths of each dataset to achieve better correction effects under the significant data scale discrepancies (160,000 expert demonstrations vs. 16,000 feedback data). The results indicate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.905, + 0.411 + ], + "angle": 0, + "content": "that our dual-process motion adjustment mechanism can effectively leverage the expert demonstration and correction dataset, leading to accurate motion instruction adjustment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.905, + 0.441 + ], + "angle": 0, + "content": "We also provide ablation results of our designed motion codebook in Supp.B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.847, + 0.468 + ], + "angle": 0, + "content": "4.3. The Performance of Lifelong Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.906, + 0.595 + ], + "angle": 0, + "content": "In this section, we explore whether our Phoenix framework can facilitate lifelong learning through interactions. Concretely, we deploy the motion self-reflection model to interact within the environments and utilize the successful trajectories to iteratively fine-tune our motion prediction model after 10, 30, and 50 rollouts. To avoid catastrophic forgetting, we combine 20 expert demonstrations to co-fine-tune the motion prediction module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.747 + ], + "angle": 0, + "content": "We compare the lifelong learning ability of our motion-based self-reflection model and subgoal-based self-reflection model. During testing, we record the average success rate over 50 trials. As shown in Figure 4, the subgoal-based lifelong learning fails to enhance model performance during the exploration phase due to its inability to provide fine-grained action correction. In contrast, our method corrects underlying action execution during interactions, allowing the robot to better learn from the refined trajectories, thereby achieving self-improvement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.78, + 0.771 + ], + "angle": 0, + "content": "4.4. Generalization to Novel Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this section, we evaluate the generalization ability of our Phoenix framework in color disruption and position disruption novel tasks as shown in Figure 5. In the color disruption setting, we replace the red block with the blue block in the Stack_D0 task to verify whether our model could generalize to object manipulation tasks with different visual characteristics. In the position disruption setting, we change the fixed position of the coffee machine to a randomly placed posi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6987" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.315, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.189, + 0.303, + 0.203 + ], + "angle": 0, + "content": "(a) Color Disruption in Stack_D0" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.089, + 0.544, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.189, + 0.55, + 0.203 + ], + "angle": 0, + "content": "(b) Position Disruption in Coffee_D0" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.09, + 0.904, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.189, + 0.831, + 0.202 + ], + "angle": 0, + "content": "(c) Performance on Novel Tasks" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.219, + 0.907, + 0.261 + ], + "angle": 0, + "content": "Figure 5. In the color disruption setting, we replace the red block with the blue block in the Stack_D0 task as shown in (a). In the position disruption setting, we change the position of the coffee machine from a fixed point \\( \\circ \\) to a random position from the rectangle \\( \\square \\) in the Coffee_D0 task as illustrated in (b). The results in (c) prove that our framework could generalize well to these novel task settings." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.265, + 0.538, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.319, + 0.367, + 0.326 + ], + "angle": 0, + "content": "(a) Expert Demonstration" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.327, + 0.188, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.379, + 0.171, + 0.386 + ], + "angle": 0, + "content": "(b) In-Distribution" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.327, + 0.304, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.379, + 0.298, + 0.387 + ], + "angle": 0, + "content": "(c) Pose Disruption" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.327, + 0.42, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.379, + 0.43, + 0.387 + ], + "angle": 0, + "content": "(d) Background Disruption" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.327, + 0.536, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.379, + 0.537, + 0.387 + ], + "angle": 0, + "content": "(e) Texture Disruption" + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.403, + 0.503, + 0.417 + ], + "angle": 0, + "content": "Figure 6. The real-world experiments with different variations." + }, + { + "type": "table", + "bbox": [ + 0.587, + 0.266, + 0.903, + 0.339 + ], + "angle": 0, + "content": "
ModelIn-Dis.Pose Dis.Bg.Tex.
OpenVLA55%30%35%45%
Task60%25%25%45%
Motion60%35%30%50%
Ours75%55%45%65%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.601, + 0.341, + 0.857, + 0.353 + ], + "angle": 0, + "content": "Table 3. The real-world experiment results." + }, + { + "type": "table", + "bbox": [ + 0.587, + 0.355, + 0.884, + 0.4 + ], + "angle": 0, + "content": "
TaskMotion10 rollout30 rollout
In-Dis.60%65%75%
Pose Dis.35%45%50%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.618, + 0.401, + 0.84, + 0.414 + ], + "angle": 0, + "content": "Table 4. The lifelong learning results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.482, + 0.462 + ], + "angle": 0, + "content": "tion within a specific area in the Coffee_D0 task to verify whether our method could generalize to unseen scenarios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.673 + ], + "angle": 0, + "content": "For these novel tasks, although the subgoal-conditioned policy could predict correct high-level semantic subgoal for manipulation, this method fails to predict precise robotic action to complete the tasks. Due to its limitation of providing high-level semantic correction information, the subgoal self-reflection method fails to effectively leverage the knowledge of MLLMs for action correction to manipulation tasks. In contrast, as shown in Figure 5(c), our motion-conditioned policy could generate fine-grained motion instruction to achieve generalizable manipulation benefiting from the perception and inferential capability of MLLMs. Besides, our method could achieve better performance on novel tasks by comprehensively refining the motion instruction with the motion-based self-reflection framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.32, + 0.698 + ], + "angle": 0, + "content": "4.5. Real-World Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In real-world scenarios, we conduct the challenging \"drawer open\" articulated object manipulation task as shown in Fig 6(a), where the robot needs to align gripper with handle through precise rotations to open the drawer. We utilize the spacemouse device to collect 100 expert demonstrations with 14 motion instructions (e.g., \"move arm right\", \"rotate around x-axis\"). We train a motion-conditioned diffusion policy to convert instructions into robotic actions. During the inference, we introduce human-in-the-loop interventions to manually correct failure situations to collect 20 corresponding refined interaction trajectories to train our motion correction module. All models are only fine-tuned on the real-world data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.581 + ], + "angle": 0, + "content": "To validate generalization, we design 4 settings as shown in Fig 6(b-e). In the pose disruption setting, we change the pose distribution of the drawer. For the background disruption setting, the background color was modified to green. In the texture disruption setting, the texture of the drawer was altered to evaluate performance under significant visual variations. The results in Tab 3 demonstrate the generalization ability of our method. We also evaluate lifelong learning, the results in Tab 4 show that our model achieves self-improvement in real world." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.629 + ], + "angle": 0, + "content": "We also provide more real-world task experiments with a rule-based manipulation policy to prove the effectiveness of our motion-based self-reflection method in Supp.C." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.634, + 0.657 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.907, + 0.788 + ], + "angle": 0, + "content": "In this work, we propose a motion-based self-reflection framework to convert the semantic reflection of MLLMs into fine-grained robotic action correction. Based on this framework, we further automatically improve the model's capability from interactions. We hope this motion-based self-reflection framework could bring insights for enhancing the generalization capabilities of agents in robotic manipulation tasks through the integration of MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.694, + 0.817 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.9 + ], + "angle": 0, + "content": "The project was supported by National Natural Science Foundation of China (NO.62106272), Shanghai AI Laboratory, National Key R&D Program of China (2022ZD0160101) and CCF-Zhipu.AI Large Model Innovation Fund." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6988" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.483, + 0.156 + ], + "angle": 0, + "content": "[1] Andrew G Barto and Sridhar Mahadevan. Recent advances in hierarchical reinforcement learning. Discrete event dynamic systems, 13:341-379, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.226 + ], + "angle": 0, + "content": "[2] Suneel Belkhale, Tianli Ding, Ted Xiao, Pierre Sermanet, Quon Vuong, Jonathan Tompson, Yevgen Chebotar, Debidatta Dwibedi, and Dorsa Sadigh. Rt-h: Action hierarchies using language. arXiv preprint arXiv:2403.01823, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.482, + 0.299 + ], + "angle": 0, + "content": "[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[5] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[6] Yinpei Dai, Jayjun Lee, Nima Fazeli, and Joyce Chai. Racer: Rich language-guided failure recovery policies for imitation learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[7] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[8] John H Flavell. Metacognition and cognitive monitoring: A new area of cognitive-developmental inquiry. American psychologist, 34(10):906, 1979. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[9] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for generalizable manipulation skills. In International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[10] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado Van Hasselt, and David Silver. Distributed prioritized experience replay. arXiv preprint arXiv:1803.00933, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[11] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[12] Daeyeol Lee, Hyojung Seo, and Min Whan Jung. Neural basis of reinforcement learning and decision making. Annual review of neuroscience, 35(1):287-308, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[13] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[14] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.904, + 0.19 + ], + "angle": 0, + "content": "[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[16] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation. arXiv preprint arXiv:2405.17418, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[17] Kehui Liu, Zixin Tang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Coherent: Collaboration of heterogeneous multi-robot system with large language models. arXiv preprint arXiv:2409.15146, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[18] Qian Liu, Xiaosen Zheng, Niklas Muennighoff, Guangtao Zeng, Longxu Dou, Tianyu Pang, Jing Jiang, and Min Lin. Regmix: Data mixture as regression for language model pretraining. arXiv preprint arXiv:2407.01492, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[19] Jingxian Lu, Wenke Xia, Dong Wang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Koi: Accelerating online imitation learning via hybrid key-state guidance. arXiv preprint arXiv:2408.02912, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[20] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. In Conference on Robot Learning (CoRL), 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[21] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.905, + 0.67 + ], + "angle": 0, + "content": "[22] Xincheng Pang, Wenke Xia, Zhigang Wang, Bin Zhao, Di Hu, Dong Wang, and Xuelong Li. Depth helps: Improving pre-trained rgb-based policy with depth information injection. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7251-7256. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.674, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[23] Shubham Pateria, Budhitama Subagdja, Ah-hwee Tan, and Chai Quek. Hierarchical reinforcement learning: A comprehensive survey. ACM Computing Surveys (CSUR), 54(5): 1-35, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[24] Shreyas Sundara Raman, Vanya Cohen, Eric Rosen, Ifrah Idrees, David Paulius, and Stefanie Telsex. Planning with large language models via corrective re-prompting. In NeurIPS 2022 Foundation Models for Decision Making Workshop, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.904, + 0.828 + ], + "angle": 0, + "content": "[25] Roger C Schank. What we learn when we learn by doing. Technical report, Technical report, 1995. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.904, + 0.871 + ], + "angle": 0, + "content": "[26] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6989" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[28] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[29] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.484, + 0.331 + ], + "angle": 0, + "content": "[30] Zihao Wang, Shaofei Cai, Guanzhou Chen, Anji Liu, Xiaojian Ma, and Yitao Liang. Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[31] Wenke Xia, Dong Wang, Xincheng Pang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Kinematic-aware prompting for generalizable articulated object manipulation with llms. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2073-2080, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[32] Chuyan Xiong, Chengyu Shen, Xiaqi Li, Kaichen Zhou, Jiaming Liu, Ruiping Wang, and Hao Dong. Aic mllm: Autonomous interactive correction mllm for robust robotic manipulation. arXiv preprint arXiv:2406.11548, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.484, + 0.515 + ], + "angle": 0, + "content": "[33] Jiasheng Ye, Peiju Liu, Tianxiang Sun, Yunhua Zhou, Jun Zhan, and Xipeng Qiu. Data mixing laws: Optimizing data mixtures by predicting language modeling performance. arXiv preprint arXiv:2403.16952, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[34] Naoki Yokoyama, Alex Clegg, Joanne Truong, Eric Undersander, Tsung-Yen Yang, Sergio Arnaud, Sehoon Ha, Dhruv Batra, and Akshara Rai. Asc: Adaptive skill coordination for robotic mobile manipulation. IEEE Robotics and Automation Letters, 9(1):779-786, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.587, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[35] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[36] Jia Zeng, Qingwen Bu, Bangjun Wang, Wenke Xia, Li Chen, Hao Dong, Haoming Song, Dong Wang, Di Hu, Ping Luo, et al. Learning manipulation by predicting interaction. arXiv preprint arXiv:2406.00439, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.714, + 0.484, + 0.782 + ], + "angle": 0, + "content": "[37] Junjie Zhang, Chenjia Bai, Haoran He, Wenke Xia, Zhigang Wang, Bin Zhao, Xiu Li, and Xuelong Li. Sam-e: leveraging visual foundation model with sequence imitation for embodied manipulation. arXiv preprint arXiv:2405.19586, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.785, + 0.484, + 0.839 + ], + "angle": 0, + "content": "[38] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6990" + } + ] +] \ No newline at end of file diff --git a/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/497d038e-fa23-45f9-aee3-d3bfa187eaf5_origin.pdf b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/497d038e-fa23-45f9-aee3-d3bfa187eaf5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cca1dacb40115c9c5a89c5a1eda5285a06ed0fb2 --- /dev/null +++ b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/497d038e-fa23-45f9-aee3-d3bfa187eaf5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71bc01f15dfa6fd993d08f55e2dc646af4187a4195adebf7a4091ac794b5403d +size 850947 diff --git a/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/full.md b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..68c2dc11765b5a3dc4820da0d5f663f9c5f22ee8 --- /dev/null +++ b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/full.md @@ -0,0 +1,326 @@ +![](images/94f26f24f6f7452f7e01aa5535d32104cc00a0a2237b3a7d26cdaf862c4b8530.jpg) + +# Phoenix: A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction + +Wenke Xia $^{1,2,*}$ , Ruoxuan Feng $^{1}$ , Dong Wang $^{2}$ , Di Hu $^{1,\dagger}$ + +$^{1}$ Gaoling School of Artificial Intelligence, Renmin University of China, Beijing $^{2}$ Shanghai AI Laboratory + +# Abstract + +Building a generalizable self-correction system is crucial for robots to recover from failures. Despite advancements in Multimodal Large Language Models (MLLMs) that empower robots with semantic reflection ability for failure, translating semantic reflection into "how to correct" fine-grained robotic actions remains a significant challenge. To address this gap, we build the Phoenix framework, which leverages motion instruction as a bridge to connect high-level semantic reflection with low-level robotic action correction. In this motion-based self-reflection framework, we start with a dual-process motion adjustment mechanism with MLLMs to translate the semantic reflection into coarse-grained motion instruction adjustment. To leverage this motion instruction for guiding "how to correct" fine-grained robotic actions, a multi-task motion-conditioned diffusion policy is proposed to integrate visual observations for high-frequency robotic action correction. By combining these two models, we could shift the demand for generalization capability from the low-level manipulation policy to the MLLMs-driven motion adjustment model and facilitate precise, fine-grained robotic action correction. Utilizing this framework, we further develop a lifelong learning method to automatically improve the model's capability from interactions with dynamic environments. The experiments conducted in both the RoboMimic simulation and real-world scenarios prove the superior generalization and robustness of our framework across a variety of manipulation tasks. Our code is released at https://github.com/GeWuLab/Motion-based-Self-Reflection-Framework. + +# 1. Introduction + +"Failure is simply the opportunity to begin again, this time more intelligently." — Henry Ford + +![](images/fc3cb80f6ac08831bf8e9a0baf087d94b24092c910fdf011164c138d348183ac.jpg) +Figure 1. Our motion-based self-reflection framework utilizes coarse-grained motion instruction as a bridge to convert the high-level semantic reflection into fine-grained robotic action correction, thereby facilitating generalizable and precise action correction with perceptual and inferential capabilities of MLLMs. + +Humans are naturally equipped with the ability to correct their behaviors by intentionally reflecting on actions that lead to failure. [12, 25]. By analyzing failure situations from high-level semantic reflection and low-level action correction perspective, humans can efficiently adapt to dynamic environments [8]. To emulate the correction capability and foster a continuous cycle of self-improvement in robots, researchers [26, 30, 35] have sought to develop self-reflection systems that enable robots to recover from and learn through their failure interactions. + +Among them, some existing self-correction systems [19, + +26, 34] leverage reinforcement learning to guide robots in correct low-level action execution through reward functions. However, the reliance on reinforcement learning limits the ability of these self-correction systems to generalize across long-horizon manipulation tasks, due to unstable training processes [10] and the need for task-specific prior knowledge [9, 35]. To construct a generalizable and stable self-correction system, recent works [17, 30, 31] borrow the inferential capability of Multi-modal Large Language Models (MLLMs) to propose closed-loop high-level semantic reflection framework for failure correction. Although these semantic self-reflection frameworks can decompose the failure correction process into semantic subgoals, they primarily rely on a predefined skill library to execute the detailed subgoals, which fails to utilize the generalization ability of MLLMs in fine-grained robotic action correction. + +To maximize the generalization potential of MLLMs for action correction, we propose motion instruction as a bridge to convert high-level semantic reflection to fine-grained robotic action correction. Motion instruction refers to coarse-grained robotic movement commands such as "move arm backward" and "adjust gripper position". Serving as an intermediate layer, motion instruction could provide general, low-frequency decision information for high-frequency robotic action execution, which makes it an excellent medium for embedding the knowledge of MLLMs into fine-grained action correction. As shown in Figure 1, we decompose the semantic reflection knowledge into coarse-grained motion instruction adjustment to indicate "how to correct" fine-grained action for low-level policy execution. This transition shifts the perceptual and decision-making requirements from low-level robotic policy to the MLLMs-driven motion adjustment model, thereby enabling generalizable, fine-grained robotic action correction. + +Hence, in this work, we build the Phoenix framework, a motion-based self-reflection framework designed to convert the semantic reflection of MLLMs into fine-grained robotic action correction. Initially, we develop a dual-process motion adjustment mechanism to ensure efficient prediction through a motion prediction module, while addressing failure with a motion correction module. Concretely, we first utilize expert demonstration trajectories to train the motion prediction module for efficient motion instruction generation. Despite its efficiency in generating initial instructions, this module often struggles to handle failure scenarios. To recover from failures, we collect a comprehensive failure correction dataset and fine-tune the motion correction module, which thoroughly provides adjusted motion instructions through a chain-of-thought approach. By integrating these two modules, the dual-process motion adjustment mechanism guarantees both robustness and efficiency, facilitating the generation of accurate motion instructions. As the coarse-grained motion instructions only provide gen + +eral and low-frequency guidance for robotic manipulation, we further design a multi-task motion-conditioned diffusion policy that integrates visual observations to translate motion instruction into precise, high-frequency action corrections for manipulation tasks. Moreover, by leveraging these correction trajectories, we propose a lifelong learning method that iteratively enhances the model's capabilities through interaction, ensuring continuous improvements in performance and adaptability to dynamic environments. + +To validate the efficacy of our framework, we conduct experiments across 9 contact-rich robotic manipulation tasks within the RoboMimic simulation [20]. The results demonstrate that our method could provide more precise action correction from failures through self-reflection and facilitate self-improvement through interactions with environments. Further, we conduct two novel manipulation tasks with color disruption and position distribution disruption, proving the generalization ability of our framework. The real-world experiments also demonstrate the applicability and robustness of our approach in practical scenarios. + +# 2. Related Work + +# 2.1. Robotic Self-correction Systems + +Self-correction serves as a crucial mechanism enabling robots to recover from failures. To achieve self-correction on the low-level robotic action, Reinforcement Learning [1, 23] is proposed to guide robots in adjusting behaviors via reward signals. However, reinforcement learning strategies encounter difficulties in intricate robotic environments primarily due to learning inefficiency. Borrowing the common knowledge of MLLMs, Raman et al. [24] proposes to build a semantic self-reflection system for long-horizon task planning. To facilitate interaction with environments, some efforts generate robotic action through simulation APIs [17, 29, 30] and predefined action skill libraries [14, 16]. However, the reliance on predefined low-level skill libraries makes semantic self-reflection frameworks fail to directly provide fine-grained action correction. To provide low-level action feedback for robotic manipulation, recent works [6, 16, 32] suggest adjusting end-effector poses to refine actions. Nonetheless, these techniques are primarily restricted to simple manipulation tasks that employ motion planning and fail to generalize to contact-rich manipulation scenarios. In this work, we utilize motion instruction as an intermediate layer to guide robotic action correction, borrowing the perceptual and inferential capability of MLLMs for fine-grained robotic action correction. + +# 2.2. Robotic Manipulation Policy + +The development of generalizable strategies for robotic manipulation remains a persistent challenge in robotics research. ACT [38] is proposed to predict action sequences + +![](images/27ffee8de0eb2b660bfadeec206406a889c4db241c7cf888ebff4097a6c8a490.jpg) +Figure 2. The pipeline of our motion-based self-reflection framework. (a) demonstrates the dual-process motion refinement mechanism which leverages the motion prediction module for efficient motion instruction prediction and motion correction module for comprehensive failure correction. (b) illustrates the motion-conditioned diffusion policy which converts the low-frequency motion instruction guidance into high-frequency robotic action. The lifelong learning approach in (c) iteratively enhances the ability of the motion prediction module from the refined interaction trajectories. + +to ensure temporal alignment of actions, while the Diffusion Policy [5] addresses multi-modal action distributions to enhance robust manipulation. Driven by the development of foundation models [13, 28], recent works [4, 7, 14, 22, 36, 37] leverage the world knowledge of MLLMs to facilitate task decomposition and planning for robotic manipulation. Concurrently, other works [3, 11, 21, 27] collect large-scale robotic manipulation demonstrations to train generalizable language-conditioned manipulation policy across different robots and tasks. Moreover, RT-H [2] employs detailed motion commands instead of semantic language inputs, fostering flexible and generalizable manipulation capabilities through the advanced perceptual abilities of MLLMs. However, the prohibitive costs associated with robotic data collection limit the scalability of existing imitation learning approaches. To address this, we propose the motion-based self-reflection framework to enable autonomous self-improvement through continuous interaction with environments without human intervention. + +# 3. Motion-based Self-Reflection Framework + +# 3.1. Challenges in Robotic Self-Correction Model + +Building a generalizable and robust self-correction system is a key component in achieving failure correction for robots. Multi-modal Large Language Models (MLLMs) have already been applied to the construction of robotic self-reflection framework to recover from failures. However, existing systems mainly focus on semantic reflection, and their application to fine-grained action correction still faces the following two issues: + +- How to enable MLLMs to understand manipulation tasks and provide detailed correction information? +- How to convert the correction information provided by MLLMs into precise, high-frequency robotic actions? + +To address these issues, we propose the Phoenix frame + +work, a motion-based self-reflection framework that integrates a dual-process motion adjustment mechanism and a multi-task motion-conditioned diffusion policy as illustrated in Figure 2. As detailed in Section 3.2, the dual-process motion adjustment mechanism is developed to ensure efficient and accurate motion instruction generation. Further, the motion-conditioned diffusion policy is proposed to translate the coarse-grained motion instructions into precise robotic actions, as explained in Section 3.3. Based on the refined manipulation trajectories, we propose a lifelong learning approach to facilitate robotic self-improvement, as outlined in Section 3.4. + +# 3.2. Dual-process Motion Adjustment Mechanism + +The dual-process motion adjustment mechanism is designed to ensure efficient motion prediction through a motion prediction module, while comprehensively addressing failure with a motion correction module. Given the observation $O$ and task description $T$ , we first train a Motion Prediction Module (MPM) with expert demonstration dataset $D_{e}$ to generate initial motion instruction $m_{i}$ . However, the MPM trained on expert demonstrations struggles to handle failure situations. Thus, we construct a comprehensive failure correction dataset $D_{c}$ to fine-tune the Motion Correction Module (MCM), enabling it to analyze the failure situation and adjust $m_{i}$ with a chain-of-thought approach. If $m_{i}$ is deemed correct, we adopt it as the decision motion instruction $m_{d}$ for further robotic action prediction. Otherwise, we employ the MCM to analyze the failure situation and generate adjusted motion instruction $m_{a}$ as decision motion instruction $m_{d}$ . Through the guidance of $m_{d}$ , our motion-based diffusion policy can generate high-frequency corrections to the robotic actions. As described in Algorithm 1, we establish the dual-process motion adjustment mechanism to guarantee the efficiency and accuracy of motion instruction generation for fine-grained robotic action prediction. + +Algorithm 1 Self-Reflection w/. Dual-Process Motion Adj. +```verilog +Require: Task description $T$ , Observation $O$ , Environment $E$ , Motion prediction module $MPM$ , Motion correction module $MCM$ , Motion-conditioned diffusion policy $\pi$ , Exploration timestep $K$ , +``` + +1: $O_{1}\gets E.\text{reset} ()$ + +2: for $k = 1$ to $K$ do + +3: $m_{i}\gets MPM(O_{k},T)$ + +4: failure_flag, semantic_info $\leftarrow$ MCM(Ok,m) + +5: if failure_flag is true then + +6: $m_{a}\gets MCM(O_{k},\text{semantic\_info})$ + +7: $m_{d}\gets m_{a}$ + +8: else + +9: $m_{d}\gets m_{i}$ + +10: end if + +11: $a\gets \pi (O,m_d)$ +12: $O_{k + 1}\gets E.\text{step}(a)$ + +13: end for + +Motion Prediction Module (MPM). To fully harness the perceptual and decision-making capabilities of MLLMs for efficient motion instruction prediction, we develop a motion instruction dataset from the expert demonstrations dataset $D_{e}$ to fine-tune MLLMs for robotic manipulation tasks. To construct the expert dataset, we filter the robotic action to get dominant motion from expert demonstration with a threshold, generating a set of motion instructions that include arm direction and gripper control. In practice, we find that separating arm direction instruction and gripper control instruction would cause the misalignment between textual motion instruction and fine-grained robotic action. To address this issue, we combine the direction movement with gripper control, resulting in unified instruction formats for motion instruction such as "move arm right with gripper closed". In addition, we add the "make slight adjustments to gripper position" instruction to model the minor robotic actions below the threshold. Through the automatic construction method, we build 37 types of motion instructions as guidance for further robotic action prediction. By training on the expert dataset, the MPM acquires an understanding of robotic manipulation tasks and could efficiently generate an initial motion instruction $m_{i}$ . + +Motion Correction Module (MCM). During interactions with the environment, robots may execute incorrect actions, leading to a failure situation in the task. However, the MPM trained on success expert data often struggles to recover from these failure scenarios. Thus, we develop the motion correction module, to identify failure scenarios and correct behaviors from such situations. As shown in Figure 2(a), the MCM would evaluate the initial motion instruction $m_{i}$ and conduct a dual process based on the evaluation results to achieve efficient and accurate motion in- + +![](images/58b902bcb2c3bd382050d866be6c471215691932c42cb26724dfad1a2905896a.jpg) +(a) Online Human Interventions + +![](images/b628c76d9c9b20eacf6ea52188f1e403b38ab35478911fa1cd95303fe633f455.jpg) +(b) Offline Human Annotations + +![](images/f28590979d23cb25a51c3f2bc60b5abc50a1cf57d4b9215d6a3274676501cd39.jpg) +(c) Expert Demonstrations +Figure 3. Illustrations of our correction data: online human interventions, offline human annotations, and expert demonstrations. + +struction adjustment. Once encountering the failure situation, the MCM would first analyze the type of failure and derive a semantic-level correction goal, such as "insert the coffee pot into the coffee machine". Based on this correction goal, the MCM further adjusts the motion instructions with its learned failure-correction knowledge, ultimately generating an accurate motion instruction through a hierarchical chain-of-thought approach. + +To equip MCM with the capabilities for failure detection and correction, we construct a comprehensive correction dataset as illustrated in Figure 3. This dataset includes three types of feedback data, encompassing both semantic and motion perspectives, and is categorized as follows: + +- Online Human Intervention. We implement the human-in-the-loop method for trajectory collection. We first deploy the motion prediction model to interact with environments, then manually intervene to correct the motion instructions whenever the agent encounters failure situations. This method could collect accurate and high-quality motion correction data to ensure task completion. However, it requires frequent manual interactions with the environment, which leads to significant time consumption and makes it difficult to collect large-scale data. +- Offline Human Annotation. We utilize the motion prediction model to gather trajectory data, periodically sampling trajectories and annotating them with semantic reflections and motion correction details. While the accuracy of offline annotated data cannot be guaranteed due to its inability to interact with environments for verification, this method offers a significant volume of annotations. +- Expert Demonstration. We automate annotations on expert trajectories. Since these trajectories are successful, this data is used to provide accurate motion information to enhance the model's motion prediction capabilities. + +By fine-tuning MCM on this dataset, we enhance the + +MCM to thoroughly comprehend various types of failure situations and provide motion instruction corrections. Through the integration of MPM and MCM, the dual-process motion adjustment mechanism enables efficient motion instruction generation while ensuring comprehensive correction in failure situations. + +# 3.3. Motion-conditioned Diffusion Policy + +As the motion instruction only provides general and low-frequency guidance for manipulation, we train a multi-task motion-conditioned diffusion policy $\pi$ to convert the motion instructions into precise, high-frequency robotic actions. This policy takes observations $O$ and decision motion instructions $m_{d}$ to output robotic actions $a$ . To ensure the policy adheres to the motion instruction, we make adjustments as depicted in Figure 2(b): + +First, we observe that existing pre-trained language models often struggle to capture the discriminative features of various motion instructions. This limitation hampers their ability to follow various motion instructions. To address this issue, we introduce a learnable motion codebook designed to provide discriminative features for motion instructions. For a given decision motion instruction $m_d$ , the codebook would retrieve the corresponding motion feature to facilitate accurate robotic action prediction. + +Further, we find that the direct concatenation of observation representation and motion instruction feature would cause the diffusion policy to prefer to rely on the vision information for action prediction, thereby hindering the effectiveness of the motion instruction guidance. To address this issue, we take the observation representation and motion instruction feature as separate conditions in different stages of the diffusion policy, allowing the model to better learn the guidance information from the motion instruction and thereby promote precise action correction. + +By integrating these two adjustments, we train the diffusion policy for action prediction with the following loss: + +$$ +\mathcal {L} = \operatorname {M S E} \left(\mathcal {E} ^ {k}, \pi (\mathcal {O}, \mathcal {M}, \mathcal {A} ^ {0} + \mathcal {E} ^ {k}, k)\right), \tag {1} +$$ + +where $\mathcal{O}$ is the observation representation, $\mathcal{M}$ is the motion instruction feature, $\mathcal{A}^0$ is the ground truth robotic action, $\mathcal{E}^k$ indicates the random noise at the denoising iteration $k$ . Through minimizing the loss function in Eq 1, the diffusion policy $\pi$ could effectively predict precise, high-frequency robotic action guided by motion instruction. + +# 3.4. Action Correction for Lifelong Learning + +The dual-process motion adjustment mechanism leverages the MPM to efficiently predict motion instructions and the MCM to adjust them with a comprehensive chain-of-thought approach. However, the reliance on the chain-of-thought poses challenges in adapting to real-time scenarios due to its time-consuming. Furthermore, the collection of + +manipulation data and correction data is exceedingly labor-intensive. Thus, we propose a lifelong learning method that equips the MPM with both motion prediction and failure correction capabilities through learning from the refined interaction trajectories as illustrated in Figure 2(c), which enhances our model to adapt and react quickly to the environment without human intervention. + +Benefiting from the motion-conditioned diffusion policy which could adhere to the motion instruction to generate task-aware robotic action, we can enhance the robot's capabilities through only improving the MPM informed by the refined interaction trajectory. To address the issue of catastrophic forgetting, we mix the refined interaction trajectory with expert demonstration for co-fine-tuning, allowing the model to simultaneously learn failure correction and enhance the motion prediction capabilities. Through updates from refined interaction trajectories, our model can achieve self-improvement by learning from the knowledge of the motion correction module, achieving fast and accurate manipulation for contact-rich manipulation tasks. + +# 4. Experiments + +To comprehensively evaluate our framework, we propose experiments to answer the following questions: + +- Does our motion-guided self-reflection model enhance the precision of action correction? Section 4.2 +- Can our model achieve lifelong learning from interaction with environments? Section 4.3 +- Does our framework can generalize across novel tasks? Section 4.4 +- Can our framework ensure reliability and robustness in real-world scenarios? Section 4.5 + +# 4.1. Experiment Settings + +In this work, we conduct experiments on 9 contact-rich manipulation tasks in RoboMimic [20], ranging from long-horizon tasks like ThreePieceAssembly to fine-grained manipulation tasks like Threading. To transform high-level semantic information into motion instructions, we filter expert demonstrations to obtain over 160,000 pairs of motion instructions and observations. The dataset includes 37 types of motion instructions, which are utilized to fine-tune the LLaVA-v1.5 model [15] as the motion prediction module. Furthermore, to develop the motion correction module that integrates semantic comprehension and motion instruction adjustment, we collect correction data comprising 3,644 online human intervention data, 7,365 offline human annotations, and 6,378 expert demonstrations. We filter out the correction dataset to balance the proportion of various failure situations to enhance the model's correction capabilities. Ultimately, to translate motion instructions into precise robotic actions, we train a multi-task motion-conditioned diffusion policy using a learnable motion instruction code + +
MethodsCoffee_D0Coffee_D1Stack_D0Stack_D1StackThree_D0StackThree_D1Threading_D0ThreePieceAssembly_D0ThreePieceAssembly_D1Mean
OpenVLA [11]42%18%84%86%36%20%20%28%8%38.0%
Task-conditioned66%24%88%68%30%6%74%20%0%41.8%
Subgoal-conditioned76%26%88%74%24%6%78%20%2%43.8%
Motion-conditioned68%32%92%84%38%16%58%30%4%46.9%
Subgoal Self-reflection80%32%88%78%32%6%80%34%2%48.0%
Phoenix (Ours)94%48%96%86%50%20%68%52%6%57.8%
Human Intervention (Oracle)100%100%100%90%70%40%100%70%40%78.9%
+ +Table 1. Comparison experiments results across 9 manipulation tasks in RoboMimic Simulation. The results demonstrate that our motion-based self-reflection method achieves better performance by facilitating precise correction of fine-grained robotic actions. + +book, incorporating 500 demonstrations per task. During inference in simulation, our dual-process motion adjustment mechanism would provide motion instruction at $5Hz$ , and the diffusion policy would extend the motion instruction with visual observations to a $20Hz$ action sequence to control the robot. For each task, we conducted 50 trials and report the average success rate. More implementation details could refer to Supp.A. + +# 4.2. Performance of Motion Self-Reflection Model + +# 4.2.1 Comparison Results + +To evaluate our motion-based self-reflection framework, we compare our framework with other approaches. To ensure fairness, all our comparison methods are trained on the expert data from the simulation environment, with the decision model using LLaVA-v1.5 and the underlying policy employing a diffusion policy. + +- OpenVLA [11]. We fine-tune the OpenVLA model to provide baseline performance for multi-task experiments. +- Task-conditioned policy. We take the task description as the condition for diffusion policy without the reflection framework, as a variance of RT-1 [3] and Octo [27]. +- Subgoal-conditioned policy. We fine-tune a LLaVAv1.5 to predict subgoals at $5\mathrm{Hz}$ , which are utilized as the condition for diffusion policy without reflection framework. This method borrows the semantic comprehension capabilities of MLLMs, and is implemented as a variance of PaLM-E [7] with an individual diffusion policy. +- Motion-conditioned policy. We fine-tune a LLaVA-v1.5 as the motion prediction model to provide motion instructions at 5Hz, using these predictions to condition the diffusion policy without the reflection framework. This method employs the perceptual and inferential capacities of MLLMs, realized as a variation of RT-H [2] with an individual diffusion policy. +- Human Intervention. We manually correct the wrong motion instructions for the motion-conditioned policy. This method provides an upper bound on the performance of self-reflection methods. Due to labor costs, the results are presented as average success rates across 10 trials. +- Subgoal Self-reflection. We fine-tune a LLaVA-v1.5 as subgoal self-reflection model and apply it to the subgoal-condition policy. This method is designed to validate the + +effectiveness of the semantic self-reflection model. + +As shown in Table 1, we first compare three different condition methods. Borrowing the perception and inferential ability of MLLMs, the subgoal-conditioned and motion-conditioned policies are better than the task-conditioned policies. The results prove the potential applications of MLLMs in various complex robotic manipulation tasks. + +Focusing on specific tasks, we observe that the motion-conditioned policy excels in long-horizon tasks such as StackThree_D0 and ThreePieceAssembly_D0. However, this policy depends on consistent and accurate motion instruction predictions, which poses challenges in fine-grained manipulation tasks like Threading_D0. + +By providing correction subgoals, the subgoal self-reflection method consistently outperforms the subgoal-conditioned policy, particularly in long-horizon manipulation tasks such as "StackThree_D0", which demonstrates the efficacy of the self-reflection framework. + +The OpenVLA model demonstrates strong performance in certain long-horizon tasks, leveraging its end-to-end action token prediction capability. However, the lack of observation history and action chunking poses significant challenges in handling complex, fine-grained manipulation tasks like Threading_D0. + +Notably, our Phoenix method achieves more substantial improvements than the subgoal self-reflection method, demonstrating the effectiveness of motion-conditioned method in long-horizon sequential tasks and fine-grained manipulation tasks. Benefiting from our motion-based correction method, agent could correct fine-grained action through motion instruction adjustment while the subgoal-conditioned self-reflection model fails to recover from most failure situations. Besides, the human intervention method achieves high success rates across multiple tasks, demonstrating that our motion-conditioned diffusion policy can effectively adhere to motion instructions for manipulation tasks. This result indicates that our method can perform well under the correct motion instructions, showcasing the significant potential of motion-conditioned self-reflection. + +# 4.2.2 Ablation Results + +In this work, we propose a motion prediction module to provide initial motion instruction, and a motion correction + +
MethodsCoffee_D0Coffee_D1Stack_D0Stack_D1StackThree_D0StackThree_D1Threading_D0ThreePieceAssembly_D0ThreePieceAssembly_D1Mean
Motion-conditioned68%32%92%84%38%16%58%30%4%46.9%
Expert-Correction Mixture74%36%94%86%38%22%64%30%2%49.6%
Expert-Correction Mixture with Self-Reflection76%30%92%90%46%26%64%34%4%51.3%
Phoenix (Ours)94%48%96%86%50%20%68%52%6%57.8%
+ +Table 2. The ablation results of our dual-process motion adjustment mechanism. The results prove that our model, which separates motion prediction module and motion correction module, can provide more precise motion adjustment for action execution in manipulation tasks. + +![](images/6c5e16dfc060a6d846025f0f7976d778a895c4c8af80581777bb3ef876c3239e.jpg) +Figure 4. The lifelong learning results. The results prove that our motion-based self-reflection method could iteratively improve performance through interactions with environments. + +![](images/c5c7d13d24316b6a66738f0e5c7fd5271370f288a41756906c517321701d876f.jpg) + +![](images/e693f94a9c4eddc95f4716ffb5fd1cf83742b4809eed02076e406ab513d81989.jpg) + +module to provide fine-grained motion correction. Drawing upon prior research [18, 33], data mixture proportions could influence the efficacy of LLMs. In this section, we investigate whether integrating expert demonstrations with correction dataset, could also enhance the perception and decision-making capabilities of MLLMs for robotic manipulation with the following ablation methods: + +- Expert-Correction Mixture. We mix the expert demonstration and correction data for co-training the motion prediction model. +- Expert-Correction Mixture with Self-Reflection. We mix the expert demonstration and correction data for constraining a unified model to provide initial motion instruction and adjust the instruction. + +As illustrated in Table 2, the results show that co-training with mixture data yields superior performance compared to models trained exclusively on expert demonstration data. This indicates that combining various types of feedback data can enhance the decision-making and perception capabilities of MLLMs. It also validates the viability of our approach to achieving self-improvement through interaction. + +Besides, the mixture training model with self-reflection performs better than the one without self-reflection, which suggests that our designed motion-based self-reflection method can enhance the decision-making capabilities of robots and facilitate the correction of fine-grained actions. + +However, we find that utilizing the mixture of data to train a unified model to serve as both the motion prediction module and motion correction module fails to provide accurate correction information compared to our separated motion correction module. This suggests that the mixture training strategy may not fully leverage the strengths of each dataset to achieve better correction effects under the significant data scale discrepancies (160,000 expert demonstrations vs. 16,000 feedback data). The results indicate + +that our dual-process motion adjustment mechanism can effectively leverage the expert demonstration and correction dataset, leading to accurate motion instruction adjustment. + +We also provide ablation results of our designed motion codebook in Supp.B. + +# 4.3. The Performance of Lifelong Learning + +In this section, we explore whether our Phoenix framework can facilitate lifelong learning through interactions. Concretely, we deploy the motion self-reflection model to interact within the environments and utilize the successful trajectories to iteratively fine-tune our motion prediction model after 10, 30, and 50 rollouts. To avoid catastrophic forgetting, we combine 20 expert demonstrations to co-fine-tune the motion prediction module. + +We compare the lifelong learning ability of our motion-based self-reflection model and subgoal-based self-reflection model. During testing, we record the average success rate over 50 trials. As shown in Figure 4, the subgoal-based lifelong learning fails to enhance model performance during the exploration phase due to its inability to provide fine-grained action correction. In contrast, our method corrects underlying action execution during interactions, allowing the robot to better learn from the refined trajectories, thereby achieving self-improvement. + +# 4.4. Generalization to Novel Tasks + +In this section, we evaluate the generalization ability of our Phoenix framework in color disruption and position disruption novel tasks as shown in Figure 5. In the color disruption setting, we replace the red block with the blue block in the Stack_D0 task to verify whether our model could generalize to object manipulation tasks with different visual characteristics. In the position disruption setting, we change the fixed position of the coffee machine to a randomly placed posi + +![](images/ff7531bd49dda9526510e06ff9a40c8e908dea632532c4a17bd62d929b7551d5.jpg) +(a) Color Disruption in Stack_D0 + +![](images/3e9b8d47a72d73e864fbab79799a0e51a61d4c2e4e1eff417e148c8dd8f66c9b.jpg) +(b) Position Disruption in Coffee_D0 + +![](images/18833c50dbe49e46ad8330bd52e6dea83265d86fb26e7537645d67b1bf3565ae.jpg) +(c) Performance on Novel Tasks + +![](images/ce0248ba09f15b6805f14b98a2d013596f17a668fba7d5b3266157b033ca99b7.jpg) +Figure 5. In the color disruption setting, we replace the red block with the blue block in the Stack_D0 task as shown in (a). In the position disruption setting, we change the position of the coffee machine from a fixed point $\circ$ to a random position from the rectangle $\square$ in the Coffee_D0 task as illustrated in (b). The results in (c) prove that our framework could generalize well to these novel task settings. +(a) Expert Demonstration + +![](images/5ce8665994ad26c28bd8bec81a8ff2d3b3962a9d9339aed37c0aa3bb06e6dc76.jpg) +(b) In-Distribution +Figure 6. The real-world experiments with different variations. + +![](images/024c4e6f4e7c4313979289d9106dc2e09a04924651f32266bf37c6fee355f761.jpg) +(c) Pose Disruption + +![](images/bf67ba9076b879daff6d955408fa819b7c6661dd39fff48b01a12572fd3222ca.jpg) +(d) Background Disruption + +![](images/ce12886b7cdfe6a0fcef35c7fd1eb1a26368a65c3e393dbf9588be3b08191c50.jpg) +(e) Texture Disruption + +
ModelIn-Dis.Pose Dis.Bg.Tex.
OpenVLA55%30%35%45%
Task60%25%25%45%
Motion60%35%30%50%
Ours75%55%45%65%
+ +Table 3. The real-world experiment results. + +
TaskMotion10 rollout30 rollout
In-Dis.60%65%75%
Pose Dis.35%45%50%
+ +Table 4. The lifelong learning results + +tion within a specific area in the Coffee_D0 task to verify whether our method could generalize to unseen scenarios. + +For these novel tasks, although the subgoal-conditioned policy could predict correct high-level semantic subgoal for manipulation, this method fails to predict precise robotic action to complete the tasks. Due to its limitation of providing high-level semantic correction information, the subgoal self-reflection method fails to effectively leverage the knowledge of MLLMs for action correction to manipulation tasks. In contrast, as shown in Figure 5(c), our motion-conditioned policy could generate fine-grained motion instruction to achieve generalizable manipulation benefiting from the perception and inferential capability of MLLMs. Besides, our method could achieve better performance on novel tasks by comprehensively refining the motion instruction with the motion-based self-reflection framework. + +# 4.5. Real-World Experiments + +In real-world scenarios, we conduct the challenging "drawer open" articulated object manipulation task as shown in Fig 6(a), where the robot needs to align gripper with handle through precise rotations to open the drawer. We utilize the spacemouse device to collect 100 expert demonstrations with 14 motion instructions (e.g., "move arm right", "rotate around x-axis"). We train a motion-conditioned diffusion policy to convert instructions into robotic actions. During the inference, we introduce human-in-the-loop interventions to manually correct failure situations to collect 20 corresponding refined interaction trajectories to train our motion correction module. All models are only fine-tuned on the real-world data. + +To validate generalization, we design 4 settings as shown in Fig 6(b-e). In the pose disruption setting, we change the pose distribution of the drawer. For the background disruption setting, the background color was modified to green. In the texture disruption setting, the texture of the drawer was altered to evaluate performance under significant visual variations. The results in Tab 3 demonstrate the generalization ability of our method. We also evaluate lifelong learning, the results in Tab 4 show that our model achieves self-improvement in real world. + +We also provide more real-world task experiments with a rule-based manipulation policy to prove the effectiveness of our motion-based self-reflection method in Supp.C. + +# 5. Conclusion + +In this work, we propose a motion-based self-reflection framework to convert the semantic reflection of MLLMs into fine-grained robotic action correction. Based on this framework, we further automatically improve the model's capability from interactions. We hope this motion-based self-reflection framework could bring insights for enhancing the generalization capabilities of agents in robotic manipulation tasks through the integration of MLLMs. + +# 6. Acknowledgement + +The project was supported by National Natural Science Foundation of China (NO.62106272), Shanghai AI Laboratory, National Key R&D Program of China (2022ZD0160101) and CCF-Zhipu.AI Large Model Innovation Fund. + +# References + +[1] Andrew G Barto and Sridhar Mahadevan. Recent advances in hierarchical reinforcement learning. Discrete event dynamic systems, 13:341-379, 2003. 2 +[2] Suneel Belkhale, Tianli Ding, Ted Xiao, Pierre Sermanet, Quon Vuong, Jonathan Tompson, Yevgen Chebotar, Debidatta Dwibedi, and Dorsa Sadigh. Rt-h: Action hierarchies using language. arXiv preprint arXiv:2403.01823, 2024. 3, 6 +[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3, 6 +[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3 +[5] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 3 +[6] Yinpei Dai, Jayjun Lee, Nima Fazeli, and Joyce Chai. Racer: Rich language-guided failure recovery policies for imitation learning, 2024. 2 +[7] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3, 6 +[8] John H Flavell. Metacognition and cognitive monitoring: A new area of cognitive-developmental inquiry. American psychologist, 34(10):906, 1979. 1 +[9] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for generalizable manipulation skills. In International Conference on Learning Representations, 2023. 2 +[10] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado Van Hasselt, and David Silver. Distributed prioritized experience replay. arXiv preprint arXiv:1803.00933, 2018. 2 +[11] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3, 6 +[12] Daeyeol Lee, Hyojung Seo, and Min Whan Jung. Neural basis of reinforcement learning and decision making. Annual review of neuroscience, 35(1):287-308, 2012. 1 +[13] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3 + +[14] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2, 3 +[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 5 +[16] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation. arXiv preprint arXiv:2405.17418, 2024. 2 +[17] Kehui Liu, Zixin Tang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Coherent: Collaboration of heterogeneous multi-robot system with large language models. arXiv preprint arXiv:2409.15146, 2024. 2 +[18] Qian Liu, Xiaosen Zheng, Niklas Muennighoff, Guangtao Zeng, Longxu Dou, Tianyu Pang, Jing Jiang, and Min Lin. Regmix: Data mixture as regression for language model pretraining. arXiv preprint arXiv:2407.01492, 2024. 7 +[19] Jingxian Lu, Wenke Xia, Dong Wang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Koi: Accelerating online imitation learning via hybrid key-state guidance. arXiv preprint arXiv:2408.02912, 2024. 1 +[20] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. In Conference on Robot Learning (CoRL), 2021. 2, 5 +[21] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 3 +[22] Xincheng Pang, Wenke Xia, Zhigang Wang, Bin Zhao, Di Hu, Dong Wang, and Xuelong Li. Depth helps: Improving pre-trained rgb-based policy with depth information injection. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7251-7256. IEEE, 2024. 3 +[23] Shubham Pateria, Budhitama Subagdja, Ah-hwee Tan, and Chai Quek. Hierarchical reinforcement learning: A comprehensive survey. ACM Computing Surveys (CSUR), 54(5): 1-35, 2021. 2 +[24] Shreyas Sundara Raman, Vanya Cohen, Eric Rosen, Ifrah Idrees, David Paulius, and Stefanie Telsex. Planning with large language models via corrective re-prompting. In NeurIPS 2022 Foundation Models for Decision Making Workshop, 2022. 2 +[25] Roger C Schank. What we learn when we learn by doing. Technical report, Technical report, 1995. 1 +[26] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 1, 2 +[27] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey + +Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 3, 6 +[28] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[29] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 2 +[30] Zihao Wang, Shaofei Cai, Guanzhou Chen, Anji Liu, Xiaojian Ma, and Yitao Liang. Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560, 2023. 1, 2 +[31] Wenke Xia, Dong Wang, Xincheng Pang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Kinematic-aware prompting for generalizable articulated object manipulation with llms. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2073-2080, 2024. 2 +[32] Chuyan Xiong, Chengyu Shen, Xiaqi Li, Kaichen Zhou, Jiaming Liu, Ruiping Wang, and Hao Dong. Aic mllm: Autonomous interactive correction mllm for robust robotic manipulation. arXiv preprint arXiv:2406.11548, 2024. 2 +[33] Jiasheng Ye, Peiju Liu, Tianxiang Sun, Yunhua Zhou, Jun Zhan, and Xipeng Qiu. Data mixing laws: Optimizing data mixtures by predicting language modeling performance. arXiv preprint arXiv:2403.16952, 2024. 7 +[34] Naoki Yokoyama, Alex Clegg, Joanne Truong, Eric Undersander, Tsung-Yen Yang, Sergio Arnaud, Sehoon Ha, Dhruv Batra, and Akshara Rai. Asc: Adaptive skill coordination for robotic mobile manipulation. IEEE Robotics and Automation Letters, 9(1):779-786, 2023. 2 +[35] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 1, 2 +[36] Jia Zeng, Qingwen Bu, Bangjun Wang, Wenke Xia, Li Chen, Hao Dong, Haoming Song, Dong Wang, Di Hu, Ping Luo, et al. Learning manipulation by predicting interaction. arXiv preprint arXiv:2406.00439, 2024. 3 +[37] Junjie Zhang, Chenjia Bai, Haoran He, Wenke Xia, Zhigang Wang, Bin Zhao, Xiu Li, and Xuelong Li. Sam-e: leveraging visual foundation model with sequence imitation for embodied manipulation. arXiv preprint arXiv:2405.19586, 2024. 3 +[38] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2 \ No newline at end of file diff --git a/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/images.zip b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..00fe15a480a0b7c5f5e1840dd0e4636eb3da43fe --- /dev/null +++ b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cebb5531e083377e4774fd16a96ef0feb595326c69473f366d155858a6fadff +size 358414 diff --git a/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/layout.json b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3bcd71cd385901dcd9ae1c7226e5af8c6f3d1a42 --- /dev/null +++ b/2025/Phoenix_ A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction/layout.json @@ -0,0 +1,8380 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 125, + 97, + 141, + 119 + ], + "blocks": [ + { + "bbox": [ + 125, + 97, + 141, + 119 + ], + "lines": [ + { + "bbox": [ + 125, + 97, + 141, + 119 + ], + "spans": [ + { + "bbox": [ + 125, + 97, + 141, + 119 + ], + "type": "image", + "image_path": "94f26f24f6f7452f7e01aa5535d32104cc00a0a2237b3a7d26cdaf862c4b8530.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 106, + 489, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 106, + 489, + 140 + ], + "spans": [ + { + "bbox": [ + 140, + 106, + 489, + 140 + ], + "type": "text", + "content": "Phoenix: A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 168, + 169, + 441, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 169, + 441, + 185 + ], + "spans": [ + { + "bbox": [ + 168, + 169, + 441, + 185 + ], + "type": "text", + "content": "Wenke Xia" + }, + { + "bbox": [ + 168, + 169, + 441, + 185 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 168, + 169, + 441, + 185 + ], + "type": "text", + "content": ", Ruoxuan Feng" + }, + { + "bbox": [ + 168, + 169, + 441, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 169, + 441, + 185 + ], + "type": "text", + "content": ", Dong Wang" + }, + { + "bbox": [ + 168, + 169, + 441, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 168, + 169, + 441, + 185 + ], + "type": "text", + "content": ", Di Hu" + }, + { + "bbox": [ + 168, + 169, + 441, + 185 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 190, + 496, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 190, + 496, + 218 + ], + "spans": [ + { + "bbox": [ + 114, + 190, + 496, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 114, + 190, + 496, + 218 + ], + "type": "text", + "content": "Gaoling School of Artificial Intelligence, Renmin University of China, Beijing " + }, + { + "bbox": [ + 114, + 190, + 496, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 190, + 496, + 218 + ], + "type": "text", + "content": " Shanghai AI Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 246, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 246, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 152, + 246, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 272, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 296, + 619 + ], + "type": "text", + "content": "Building a generalizable self-correction system is crucial for robots to recover from failures. Despite advancements in Multimodal Large Language Models (MLLMs) that empower robots with semantic reflection ability for failure, translating semantic reflection into \"how to correct\" fine-grained robotic actions remains a significant challenge. To address this gap, we build the Phoenix framework, which leverages motion instruction as a bridge to connect high-level semantic reflection with low-level robotic action correction. In this motion-based self-reflection framework, we start with a dual-process motion adjustment mechanism with MLLMs to translate the semantic reflection into coarse-grained motion instruction adjustment. To leverage this motion instruction for guiding \"how to correct\" fine-grained robotic actions, a multi-task motion-conditioned diffusion policy is proposed to integrate visual observations for high-frequency robotic action correction. By combining these two models, we could shift the demand for generalization capability from the low-level manipulation policy to the MLLMs-driven motion adjustment model and facilitate precise, fine-grained robotic action correction. Utilizing this framework, we further develop a lifelong learning method to automatically improve the model's capability from interactions with dynamic environments. The experiments conducted in both the RoboMimic simulation and real-world scenarios prove the superior generalization and robustness of our framework across a variety of manipulation tasks. Our code is released at https://github.com/GeWuLab/Motion-based-Self-Reflection-Framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 654, + 285, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 654, + 285, + 678 + ], + "spans": [ + { + "bbox": [ + 66, + 654, + 285, + 678 + ], + "type": "text", + "content": "\"Failure is simply the opportunity to begin again, this time more intelligently.\" — Henry Ford" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 330, + 247, + 541, + 512 + ], + "blocks": [ + { + "bbox": [ + 330, + 247, + 541, + 512 + ], + "lines": [ + { + "bbox": [ + 330, + 247, + 541, + 512 + ], + "spans": [ + { + "bbox": [ + 330, + 247, + 541, + 512 + ], + "type": "image", + "image_path": "fc3cb80f6ac08831bf8e9a0baf087d94b24092c910fdf011164c138d348183ac.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 521, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 578 + ], + "type": "text", + "content": "Figure 1. Our motion-based self-reflection framework utilizes coarse-grained motion instruction as a bridge to convert the high-level semantic reflection into fine-grained robotic action correction, thereby facilitating generalizable and precise action correction with perceptual and inferential capabilities of MLLMs." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 580, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 580, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 312, + 580, + 555, + 700 + ], + "type": "text", + "content": "Humans are naturally equipped with the ability to correct their behaviors by intentionally reflecting on actions that lead to failure. [12, 25]. By analyzing failure situations from high-level semantic reflection and low-level action correction perspective, humans can efficiently adapt to dynamic environments [8]. To emulate the correction capability and foster a continuous cycle of self-improvement in robots, researchers [26, 30, 35] have sought to develop self-reflection systems that enable robots to recover from and learn through their failure interactions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "Among them, some existing self-correction systems [19," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 259, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 259, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 259, + 703 + ], + "type": "text", + "content": "*Work is done during internship at Shanghai AI Laboratory" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6981" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "content": "26, 34] leverage reinforcement learning to guide robots in correct low-level action execution through reward functions. However, the reliance on reinforcement learning limits the ability of these self-correction systems to generalize across long-horizon manipulation tasks, due to unstable training processes [10] and the need for task-specific prior knowledge [9, 35]. To construct a generalizable and stable self-correction system, recent works [17, 30, 31] borrow the inferential capability of Multi-modal Large Language Models (MLLMs) to propose closed-loop high-level semantic reflection framework for failure correction. Although these semantic self-reflection frameworks can decompose the failure correction process into semantic subgoals, they primarily rely on a predefined skill library to execute the detailed subgoals, which fails to utilize the generalization ability of MLLMs in fine-grained robotic action correction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 266, + 295, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 295, + 482 + ], + "type": "text", + "content": "To maximize the generalization potential of MLLMs for action correction, we propose motion instruction as a bridge to convert high-level semantic reflection to fine-grained robotic action correction. Motion instruction refers to coarse-grained robotic movement commands such as \"move arm backward\" and \"adjust gripper position\". Serving as an intermediate layer, motion instruction could provide general, low-frequency decision information for high-frequency robotic action execution, which makes it an excellent medium for embedding the knowledge of MLLMs into fine-grained action correction. As shown in Figure 1, we decompose the semantic reflection knowledge into coarse-grained motion instruction adjustment to indicate \"how to correct\" fine-grained action for low-level policy execution. This transition shifts the perceptual and decision-making requirements from low-level robotic policy to the MLLMs-driven motion adjustment model, thereby enabling generalizable, fine-grained robotic action correction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "Hence, in this work, we build the Phoenix framework, a motion-based self-reflection framework designed to convert the semantic reflection of MLLMs into fine-grained robotic action correction. Initially, we develop a dual-process motion adjustment mechanism to ensure efficient prediction through a motion prediction module, while addressing failure with a motion correction module. Concretely, we first utilize expert demonstration trajectories to train the motion prediction module for efficient motion instruction generation. Despite its efficiency in generating initial instructions, this module often struggles to handle failure scenarios. To recover from failures, we collect a comprehensive failure correction dataset and fine-tune the motion correction module, which thoroughly provides adjusted motion instructions through a chain-of-thought approach. By integrating these two modules, the dual-process motion adjustment mechanism guarantees both robustness and efficiency, facilitating the generation of accurate motion instructions. As the coarse-grained motion instructions only provide gen" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "eral and low-frequency guidance for robotic manipulation, we further design a multi-task motion-conditioned diffusion policy that integrates visual observations to translate motion instruction into precise, high-frequency action corrections for manipulation tasks. Moreover, by leveraging these correction trajectories, we propose a lifelong learning method that iteratively enhances the model's capabilities through interaction, ensuring continuous improvements in performance and adaptability to dynamic environments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 180, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 312 + ], + "type": "text", + "content": "To validate the efficacy of our framework, we conduct experiments across 9 contact-rich robotic manipulation tasks within the RoboMimic simulation [20]. The results demonstrate that our method could provide more precise action correction from failures through self-reflection and facilitate self-improvement through interactions with environments. Further, we conduct two novel manipulation tasks with color disruption and position distribution disruption, proving the generalization ability of our framework. The real-world experiments also demonstrate the applicability and robustness of our approach in practical scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 324, + 400, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 400, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 400, + 335 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 344, + 485, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 485, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 485, + 357 + ], + "type": "text", + "content": "2.1. Robotic Self-correction Systems" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 363, + 553, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 650 + ], + "type": "text", + "content": "Self-correction serves as a crucial mechanism enabling robots to recover from failures. To achieve self-correction on the low-level robotic action, Reinforcement Learning [1, 23] is proposed to guide robots in adjusting behaviors via reward signals. However, reinforcement learning strategies encounter difficulties in intricate robotic environments primarily due to learning inefficiency. Borrowing the common knowledge of MLLMs, Raman et al. [24] proposes to build a semantic self-reflection system for long-horizon task planning. To facilitate interaction with environments, some efforts generate robotic action through simulation APIs [17, 29, 30] and predefined action skill libraries [14, 16]. However, the reliance on predefined low-level skill libraries makes semantic self-reflection frameworks fail to directly provide fine-grained action correction. To provide low-level action feedback for robotic manipulation, recent works [6, 16, 32] suggest adjusting end-effector poses to refine actions. Nonetheless, these techniques are primarily restricted to simple manipulation tasks that employ motion planning and fail to generalize to contact-rich manipulation scenarios. In this work, we utilize motion instruction as an intermediate layer to guide robotic action correction, borrowing the perceptual and inferential capability of MLLMs for fine-grained robotic action correction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 658, + 471, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 471, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 471, + 671 + ], + "type": "text", + "content": "2.2. Robotic Manipulation Policy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "The development of generalizable strategies for robotic manipulation remains a persistent challenge in robotics research. ACT [38] is proposed to predict action sequences" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6982" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 553, + 218 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 553, + 218 + ], + "type": "image", + "image_path": "27ffee8de0eb2b660bfadeec206406a889c4db241c7cf888ebff4097a6c8a490.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 226, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 555, + 281 + ], + "type": "text", + "content": "Figure 2. The pipeline of our motion-based self-reflection framework. (a) demonstrates the dual-process motion refinement mechanism which leverages the motion prediction module for efficient motion instruction prediction and motion correction module for comprehensive failure correction. (b) illustrates the motion-conditioned diffusion policy which converts the low-frequency motion instruction guidance into high-frequency robotic action. The lifelong learning approach in (c) iteratively enhances the ability of the motion prediction module from the refined interaction trajectories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 283, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 510 + ], + "type": "text", + "content": "to ensure temporal alignment of actions, while the Diffusion Policy [5] addresses multi-modal action distributions to enhance robust manipulation. Driven by the development of foundation models [13, 28], recent works [4, 7, 14, 22, 36, 37] leverage the world knowledge of MLLMs to facilitate task decomposition and planning for robotic manipulation. Concurrently, other works [3, 11, 21, 27] collect large-scale robotic manipulation demonstrations to train generalizable language-conditioned manipulation policy across different robots and tasks. Moreover, RT-H [2] employs detailed motion commands instead of semantic language inputs, fostering flexible and generalizable manipulation capabilities through the advanced perceptual abilities of MLLMs. However, the prohibitive costs associated with robotic data collection limit the scalability of existing imitation learning approaches. To address this, we propose the motion-based self-reflection framework to enable autonomous self-improvement through continuous interaction with environments without human intervention." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 520, + 282, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 282, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 282, + 533 + ], + "type": "text", + "content": "3. Motion-based Self-Reflection Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 540, + 284, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 284, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 284, + 552 + ], + "type": "text", + "content": "3.1. Challenges in Robotic Self-Correction Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "content": "Building a generalizable and robust self-correction system is a key component in achieving failure correction for robots. Multi-modal Large Language Models (MLLMs) have already been applied to the construction of robotic self-reflection framework to recover from failures. However, existing systems mainly focus on semantic reflection, and their application to fine-grained action correction still faces the following two issues:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 700 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "type": "text", + "content": "- How to enable MLLMs to understand manipulation tasks and provide detailed correction information?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 700 + ], + "type": "text", + "content": "- How to convert the correction information provided by MLLMs into precise, high-frequency robotic actions?" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "type": "text", + "content": "To address these issues, we propose the Phoenix frame" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 283, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 426 + ], + "type": "text", + "content": "work, a motion-based self-reflection framework that integrates a dual-process motion adjustment mechanism and a multi-task motion-conditioned diffusion policy as illustrated in Figure 2. As detailed in Section 3.2, the dual-process motion adjustment mechanism is developed to ensure efficient and accurate motion instruction generation. Further, the motion-conditioned diffusion policy is proposed to translate the coarse-grained motion instructions into precise robotic actions, as explained in Section 3.3. Based on the refined manipulation trajectories, we propose a lifelong learning approach to facilitate robotic self-improvement, as outlined in Section 3.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 432, + 550, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 550, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 550, + 445 + ], + "type": "text", + "content": "3.2. Dual-process Motion Adjustment Mechanism" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": "The dual-process motion adjustment mechanism is designed to ensure efficient motion prediction through a motion prediction module, while comprehensively addressing failure with a motion correction module. Given the observation " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " and task description " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": ", we first train a Motion Prediction Module (MPM) with expert demonstration dataset " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_{e}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " to generate initial motion instruction " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": ". However, the MPM trained on expert demonstrations struggles to handle failure situations. Thus, we construct a comprehensive failure correction dataset " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " to fine-tune the Motion Correction Module (MCM), enabling it to analyze the failure situation and adjust " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " with a chain-of-thought approach. If " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " is deemed correct, we adopt it as the decision motion instruction " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{d}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " for further robotic action prediction. Otherwise, we employ the MCM to analyze the failure situation and generate adjusted motion instruction " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{a}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " as decision motion instruction " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{d}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": ". Through the guidance of " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{d}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": ", our motion-based diffusion policy can generate high-frequency corrections to the robotic actions. As described in Algorithm 1, we establish the dual-process motion adjustment mechanism to guarantee the efficiency and accuracy of motion instruction generation for fine-grained robotic action prediction." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6983" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 87, + 296, + 135 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 294, + 84 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 294, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 294, + 84 + ], + "type": "text", + "content": "Algorithm 1 Self-Reflection w/. Dual-Process Motion Adj." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "text", + "content": "Require: Task description " + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "text", + "content": ", Observation " + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "text", + "content": ", Environment " + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "text", + "content": ", Motion prediction module " + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "inline_equation", + "content": "MPM" + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "text", + "content": ", Motion correction module " + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "inline_equation", + "content": "MCM" + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "text", + "content": ", Motion-conditioned diffusion policy " + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "text", + "content": ", Exploration timestep " + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 87, + 296, + 135 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "code", + "guess_lang": "verilog" + }, + { + "bbox": [ + 63, + 136, + 143, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 136, + 143, + 147 + ], + "spans": [ + { + "bbox": [ + 63, + 136, + 143, + 147 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 63, + 136, + 143, + 147 + ], + "type": "inline_equation", + "content": "O_{1}\\gets E.\\text{reset} ()" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 148, + 149, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 148, + 149, + 158 + ], + "spans": [ + { + "bbox": [ + 63, + 148, + 149, + 158 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 63, + 148, + 149, + 158 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 63, + 148, + 149, + 158 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 63, + 148, + 149, + 158 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 63, + 148, + 149, + 158 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 159, + 178, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 159, + 178, + 171 + ], + "spans": [ + { + "bbox": [ + 63, + 159, + 178, + 171 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 63, + 159, + 178, + 171 + ], + "type": "inline_equation", + "content": "m_{i}\\gets MPM(O_{k},T)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 171, + 292, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 171, + 292, + 183 + ], + "spans": [ + { + "bbox": [ + 63, + 171, + 292, + 183 + ], + "type": "text", + "content": "4: failure_flag, semantic_info " + }, + { + "bbox": [ + 63, + 171, + 292, + 183 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 63, + 171, + 292, + 183 + ], + "type": "text", + "content": " MCM(Ok,m)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 184, + 205, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 184, + 205, + 194 + ], + "spans": [ + { + "bbox": [ + 63, + 184, + 205, + 194 + ], + "type": "text", + "content": "5: if failure_flag is true then" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 195, + 246, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 195, + 246, + 207 + ], + "spans": [ + { + "bbox": [ + 63, + 195, + 246, + 207 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 63, + 195, + 246, + 207 + ], + "type": "inline_equation", + "content": "m_{a}\\gets MCM(O_{k},\\text{semantic\\_info})" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 209, + 147, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 209, + 147, + 218 + ], + "spans": [ + { + "bbox": [ + 63, + 209, + 147, + 218 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 63, + 209, + 147, + 218 + ], + "type": "inline_equation", + "content": "m_{d}\\gets m_{a}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 220, + 106, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 220, + 106, + 229 + ], + "spans": [ + { + "bbox": [ + 63, + 220, + 106, + 229 + ], + "type": "text", + "content": "8: else" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 233, + 146, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 233, + 146, + 242 + ], + "spans": [ + { + "bbox": [ + 63, + 233, + 146, + 242 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 63, + 233, + 146, + 242 + ], + "type": "inline_equation", + "content": "m_{d}\\gets m_{i}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 243, + 115, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 115, + 253 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 115, + 253 + ], + "type": "text", + "content": "10: end if" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 255, + 168, + 279 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 59, + 255, + 149, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 255, + 149, + 266 + ], + "spans": [ + { + "bbox": [ + 59, + 255, + 149, + 266 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 59, + 255, + 149, + 266 + ], + "type": "inline_equation", + "content": "a\\gets \\pi (O,m_d)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 267, + 168, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 267, + 168, + 279 + ], + "spans": [ + { + "bbox": [ + 59, + 267, + 168, + 279 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 59, + 267, + 168, + 279 + ], + "type": "inline_equation", + "content": "O_{k + 1}\\gets E.\\text{step}(a)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 59, + 280, + 106, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 280, + 106, + 289 + ], + "spans": [ + { + "bbox": [ + 59, + 280, + 106, + 289 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 316, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 592 + ], + "type": "text", + "content": "Motion Prediction Module (MPM). To fully harness the perceptual and decision-making capabilities of MLLMs for efficient motion instruction prediction, we develop a motion instruction dataset from the expert demonstrations dataset " + }, + { + "bbox": [ + 55, + 316, + 296, + 592 + ], + "type": "inline_equation", + "content": "D_{e}" + }, + { + "bbox": [ + 55, + 316, + 296, + 592 + ], + "type": "text", + "content": " to fine-tune MLLMs for robotic manipulation tasks. To construct the expert dataset, we filter the robotic action to get dominant motion from expert demonstration with a threshold, generating a set of motion instructions that include arm direction and gripper control. In practice, we find that separating arm direction instruction and gripper control instruction would cause the misalignment between textual motion instruction and fine-grained robotic action. To address this issue, we combine the direction movement with gripper control, resulting in unified instruction formats for motion instruction such as \"move arm right with gripper closed\". In addition, we add the \"make slight adjustments to gripper position\" instruction to model the minor robotic actions below the threshold. Through the automatic construction method, we build 37 types of motion instructions as guidance for further robotic action prediction. By training on the expert dataset, the MPM acquires an understanding of robotic manipulation tasks and could efficiently generate an initial motion instruction " + }, + { + "bbox": [ + 55, + 316, + 296, + 592 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 55, + 316, + 296, + 592 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Motion Correction Module (MCM). During interactions with the environment, robots may execute incorrect actions, leading to a failure situation in the task. However, the MPM trained on success expert data often struggles to recover from these failure scenarios. Thus, we develop the motion correction module, to identify failure scenarios and correct behaviors from such situations. As shown in Figure 2(a), the MCM would evaluate the initial motion instruction " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " and conduct a dual process based on the evaluation results to achieve efficient and accurate motion in-" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 328, + 71, + 541, + 118 + ], + "blocks": [ + { + "bbox": [ + 328, + 71, + 541, + 118 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 541, + 118 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 541, + 118 + ], + "type": "image", + "image_path": "58b902bcb2c3bd382050d866be6c471215691932c42cb26724dfad1a2905896a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 118, + 470, + 125 + ], + "lines": [ + { + "bbox": [ + 395, + 118, + 470, + 125 + ], + "spans": [ + { + "bbox": [ + 395, + 118, + 470, + 125 + ], + "type": "text", + "content": "(a) Online Human Interventions" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 329, + 133, + 541, + 178 + ], + "blocks": [ + { + "bbox": [ + 329, + 133, + 541, + 178 + ], + "lines": [ + { + "bbox": [ + 329, + 133, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 329, + 133, + 541, + 178 + ], + "type": "image", + "image_path": "b628c76d9c9b20eacf6ea52188f1e403b38ab35478911fa1cd95303fe633f455.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 181, + 470, + 188 + ], + "lines": [ + { + "bbox": [ + 396, + 181, + 470, + 188 + ], + "spans": [ + { + "bbox": [ + 396, + 181, + 470, + 188 + ], + "type": "text", + "content": "(b) Offline Human Annotations" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 329, + 196, + 541, + 242 + ], + "blocks": [ + { + "bbox": [ + 329, + 196, + 541, + 242 + ], + "lines": [ + { + "bbox": [ + 329, + 196, + 541, + 242 + ], + "spans": [ + { + "bbox": [ + 329, + 196, + 541, + 242 + ], + "type": "image", + "image_path": "f28590979d23cb25a51c3f2bc60b5abc50a1cf57d4b9215d6a3274676501cd39.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 243, + 463, + 251 + ], + "lines": [ + { + "bbox": [ + 402, + 243, + 463, + 251 + ], + "spans": [ + { + "bbox": [ + 402, + 243, + 463, + 251 + ], + "type": "text", + "content": "(c) Expert Demonstrations" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 261, + 555, + 284 + ], + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 284 + ], + "type": "text", + "content": "Figure 3. Illustrations of our correction data: online human interventions, offline human annotations, and expert demonstrations." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 286, + 554, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 554, + 382 + ], + "type": "text", + "content": "struction adjustment. Once encountering the failure situation, the MCM would first analyze the type of failure and derive a semantic-level correction goal, such as \"insert the coffee pot into the coffee machine\". Based on this correction goal, the MCM further adjusts the motion instructions with its learned failure-correction knowledge, ultimately generating an accurate motion instruction through a hierarchical chain-of-thought approach." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 383, + 554, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 554, + 443 + ], + "type": "text", + "content": "To equip MCM with the capabilities for failure detection and correction, we construct a comprehensive correction dataset as illustrated in Figure 3. This dataset includes three types of feedback data, encompassing both semantic and motion perspectives, and is categorized as follows:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 314, + 446, + 555, + 698 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 314, + 446, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 446, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 314, + 446, + 555, + 566 + ], + "type": "text", + "content": "- Online Human Intervention. We implement the human-in-the-loop method for trajectory collection. We first deploy the motion prediction model to interact with environments, then manually intervene to correct the motion instructions whenever the agent encounters failure situations. This method could collect accurate and high-quality motion correction data to ensure task completion. However, it requires frequent manual interactions with the environment, which leads to significant time consumption and makes it difficult to collect large-scale data." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 566, + 554, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 566, + 554, + 650 + ], + "spans": [ + { + "bbox": [ + 314, + 566, + 554, + 650 + ], + "type": "text", + "content": "- Offline Human Annotation. We utilize the motion prediction model to gather trajectory data, periodically sampling trajectories and annotating them with semantic reflections and motion correction details. While the accuracy of offline annotated data cannot be guaranteed due to its inability to interact with environments for verification, this method offers a significant volume of annotations." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 650, + 553, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 650, + 553, + 698 + ], + "spans": [ + { + "bbox": [ + 314, + 650, + 553, + 698 + ], + "type": "text", + "content": "- Expert Demonstration. We automate annotations on expert trajectories. Since these trajectories are successful, this data is used to provide accurate motion information to enhance the model's motion prediction capabilities." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "content": "By fine-tuning MCM on this dataset, we enhance the" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6984" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "MCM to thoroughly comprehend various types of failure situations and provide motion instruction corrections. Through the integration of MPM and MCM, the dual-process motion adjustment mechanism enables efficient motion instruction generation while ensuring comprehensive correction in failure situations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 152, + 248, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 248, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 248, + 166 + ], + "type": "text", + "content": "3.3. Motion-conditioned Diffusion Policy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "text", + "content": "As the motion instruction only provides general and low-frequency guidance for manipulation, we train a multi-task motion-conditioned diffusion policy " + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "text", + "content": " to convert the motion instructions into precise, high-frequency robotic actions. This policy takes observations " + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "text", + "content": " and decision motion instructions " + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "inline_equation", + "content": "m_{d}" + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "text", + "content": " to output robotic actions " + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 171, + 296, + 266 + ], + "type": "text", + "content": ". To ensure the policy adheres to the motion instruction, we make adjustments as depicted in Figure 2(b):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 267, + 296, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 296, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 296, + 375 + ], + "type": "text", + "content": "First, we observe that existing pre-trained language models often struggle to capture the discriminative features of various motion instructions. This limitation hampers their ability to follow various motion instructions. To address this issue, we introduce a learnable motion codebook designed to provide discriminative features for motion instructions. For a given decision motion instruction " + }, + { + "bbox": [ + 55, + 267, + 296, + 375 + ], + "type": "inline_equation", + "content": "m_d" + }, + { + "bbox": [ + 55, + 267, + 296, + 375 + ], + "type": "text", + "content": ", the codebook would retrieve the corresponding motion feature to facilitate accurate robotic action prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 495 + ], + "type": "text", + "content": "Further, we find that the direct concatenation of observation representation and motion instruction feature would cause the diffusion policy to prefer to rely on the vision information for action prediction, thereby hindering the effectiveness of the motion instruction guidance. To address this issue, we take the observation representation and motion instruction feature as separate conditions in different stages of the diffusion policy, allowing the model to better learn the guidance information from the motion instruction and thereby promote precise action correction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 520 + ], + "type": "text", + "content": "By integrating these two adjustments, we train the diffusion policy for action prediction with the following loss:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 529, + 295, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 529, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 96, + 529, + 295, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\operatorname {M S E} \\left(\\mathcal {E} ^ {k}, \\pi (\\mathcal {O}, \\mathcal {M}, \\mathcal {A} ^ {0} + \\mathcal {E} ^ {k}, k)\\right), \\tag {1}", + "image_path": "85d05321176ec6d62c0e1c1ba6ace5eba6f259f19a5541f949deab0939328e1d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "content": " is the observation representation, " + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "content": " is the motion instruction feature, " + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^0" + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "content": " is the ground truth robotic action, " + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^k" + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "content": " indicates the random noise at the denoising iteration " + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "content": ". Through minimizing the loss function in Eq 1, the diffusion policy " + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "content": " could effectively predict precise, high-frequency robotic action guided by motion instruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 623, + 268, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 268, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 268, + 635 + ], + "type": "text", + "content": "3.4. Action Correction for Lifelong Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "The dual-process motion adjustment mechanism leverages the MPM to efficiently predict motion instructions and the MCM to adjust them with a comprehensive chain-of-thought approach. However, the reliance on the chain-of-thought poses challenges in adapting to real-time scenarios due to its time-consuming. Furthermore, the collection of" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "manipulation data and correction data is exceedingly labor-intensive. Thus, we propose a lifelong learning method that equips the MPM with both motion prediction and failure correction capabilities through learning from the refined interaction trajectories as illustrated in Figure 2(c), which enhances our model to adapt and react quickly to the environment without human intervention." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 554, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 554, + 312 + ], + "type": "text", + "content": "Benefiting from the motion-conditioned diffusion policy which could adhere to the motion instruction to generate task-aware robotic action, we can enhance the robot's capabilities through only improving the MPM informed by the refined interaction trajectory. To address the issue of catastrophic forgetting, we mix the refined interaction trajectory with expert demonstration for co-fine-tuning, allowing the model to simultaneously learn failure correction and enhance the motion prediction capabilities. Through updates from refined interaction trajectories, our model can achieve self-improvement by learning from the knowledge of the motion correction module, achieving fast and accurate manipulation for contact-rich manipulation tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 322, + 395, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 322, + 395, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 322, + 395, + 335 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 342, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 365 + ], + "type": "text", + "content": "To comprehensively evaluate our framework, we propose experiments to answer the following questions:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 366, + 553, + 460 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 366, + 553, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 366, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 314, + 366, + 553, + 388 + ], + "type": "text", + "content": "- Does our motion-guided self-reflection model enhance the precision of action correction? Section 4.2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 389, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 389, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 314, + 389, + 553, + 411 + ], + "type": "text", + "content": "- Can our model achieve lifelong learning from interaction with environments? Section 4.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 413, + 553, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 413, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 314, + 413, + 553, + 435 + ], + "type": "text", + "content": "- Does our framework can generalize across novel tasks? Section 4.4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 437, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 437, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 314, + 437, + 553, + 460 + ], + "type": "text", + "content": "- Can our framework ensure reliability and robustness in real-world scenarios? Section 4.5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 468, + 432, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 432, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 432, + 480 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "In this work, we conduct experiments on 9 contact-rich manipulation tasks in RoboMimic [20], ranging from long-horizon tasks like ThreePieceAssembly to fine-grained manipulation tasks like Threading. To transform high-level semantic information into motion instructions, we filter expert demonstrations to obtain over 160,000 pairs of motion instructions and observations. The dataset includes 37 types of motion instructions, which are utilized to fine-tune the LLaVA-v1.5 model [15] as the motion prediction module. Furthermore, to develop the motion correction module that integrates semantic comprehension and motion instruction adjustment, we collect correction data comprising 3,644 online human intervention data, 7,365 offline human annotations, and 6,378 expert demonstrations. We filter out the correction dataset to balance the proportion of various failure situations to enhance the model's correction capabilities. Ultimately, to translate motion instructions into precise robotic actions, we train a multi-task motion-conditioned diffusion policy using a learnable motion instruction code" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6985" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 555, + 143 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 143 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 143 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 143 + ], + "type": "table", + "html": "
MethodsCoffee_D0Coffee_D1Stack_D0Stack_D1StackThree_D0StackThree_D1Threading_D0ThreePieceAssembly_D0ThreePieceAssembly_D1Mean
OpenVLA [11]42%18%84%86%36%20%20%28%8%38.0%
Task-conditioned66%24%88%68%30%6%74%20%0%41.8%
Subgoal-conditioned76%26%88%74%24%6%78%20%2%43.8%
Motion-conditioned68%32%92%84%38%16%58%30%4%46.9%
Subgoal Self-reflection80%32%88%78%32%6%80%34%2%48.0%
Phoenix (Ours)94%48%96%86%50%20%68%52%6%57.8%
Human Intervention (Oracle)100%100%100%90%70%40%100%70%40%78.9%
", + "image_path": "ebab80a91437096570bd51489d2ec5bba181bd80abf35b37f5259ad8c11c25c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 151, + 555, + 174 + ], + "lines": [ + { + "bbox": [ + 55, + 151, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 555, + 174 + ], + "type": "text", + "content": "Table 1. Comparison experiments results across 9 manipulation tasks in RoboMimic Simulation. The results demonstrate that our motion-based self-reflection method achieves better performance by facilitating precise correction of fine-grained robotic actions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 184, + 294, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 184, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 184, + 294, + 281 + ], + "type": "text", + "content": "book, incorporating 500 demonstrations per task. During inference in simulation, our dual-process motion adjustment mechanism would provide motion instruction at " + }, + { + "bbox": [ + 54, + 184, + 294, + 281 + ], + "type": "inline_equation", + "content": "5Hz" + }, + { + "bbox": [ + 54, + 184, + 294, + 281 + ], + "type": "text", + "content": ", and the diffusion policy would extend the motion instruction with visual observations to a " + }, + { + "bbox": [ + 54, + 184, + 294, + 281 + ], + "type": "inline_equation", + "content": "20Hz" + }, + { + "bbox": [ + 54, + 184, + 294, + 281 + ], + "type": "text", + "content": " action sequence to control the robot. For each task, we conducted 50 trials and report the average success rate. More implementation details could refer to Supp.A." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 291, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 291, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 291, + 301 + ], + "type": "text", + "content": "4.2. Performance of Motion Self-Reflection Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 175, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 175, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 175, + 319 + ], + "type": "text", + "content": "4.2.1 Comparison Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 327, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 399 + ], + "type": "text", + "content": "To evaluate our motion-based self-reflection framework, we compare our framework with other approaches. To ensure fairness, all our comparison methods are trained on the expert data from the simulation environment, with the decision model using LLaVA-v1.5 and the underlying policy employing a diffusion policy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 402, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 402, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 426 + ], + "type": "text", + "content": "- OpenVLA [11]. We fine-tune the OpenVLA model to provide baseline performance for multi-task experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "text", + "content": "- Task-conditioned policy. We take the task description as the condition for diffusion policy without the reflection framework, as a variance of RT-1 [3] and Octo [27]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 462, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 534 + ], + "type": "text", + "content": "- Subgoal-conditioned policy. We fine-tune a LLaVAv1.5 to predict subgoals at " + }, + { + "bbox": [ + 55, + 462, + 295, + 534 + ], + "type": "inline_equation", + "content": "5\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 462, + 295, + 534 + ], + "type": "text", + "content": ", which are utilized as the condition for diffusion policy without reflection framework. This method borrows the semantic comprehension capabilities of MLLMs, and is implemented as a variance of PaLM-E [7] with an individual diffusion policy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "type": "text", + "content": "- Motion-conditioned policy. We fine-tune a LLaVA-v1.5 as the motion prediction model to provide motion instructions at 5Hz, using these predictions to condition the diffusion policy without the reflection framework. This method employs the perceptual and inferential capacities of MLLMs, realized as a variation of RT-H [2] with an individual diffusion policy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 677 + ], + "type": "text", + "content": "- Human Intervention. We manually correct the wrong motion instructions for the motion-conditioned policy. This method provides an upper bound on the performance of self-reflection methods. Due to labor costs, the results are presented as average success rates across 10 trials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "- Subgoal Self-reflection. We fine-tune a LLaVA-v1.5 as subgoal self-reflection model and apply it to the subgoal-condition policy. This method is designed to validate the" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 185, + 528, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 185, + 528, + 196 + ], + "spans": [ + { + "bbox": [ + 322, + 185, + 528, + 196 + ], + "type": "text", + "content": "effectiveness of the semantic self-reflection model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 199, + 554, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 554, + 270 + ], + "type": "text", + "content": "As shown in Table 1, we first compare three different condition methods. Borrowing the perception and inferential ability of MLLMs, the subgoal-conditioned and motion-conditioned policies are better than the task-conditioned policies. The results prove the potential applications of MLLMs in various complex robotic manipulation tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 271, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 553, + 342 + ], + "type": "text", + "content": "Focusing on specific tasks, we observe that the motion-conditioned policy excels in long-horizon tasks such as StackThree_D0 and ThreePieceAssembly_D0. However, this policy depends on consistent and accurate motion instruction predictions, which poses challenges in fine-grained manipulation tasks like Threading_D0." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 342, + 553, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 403 + ], + "type": "text", + "content": "By providing correction subgoals, the subgoal self-reflection method consistently outperforms the subgoal-conditioned policy, particularly in long-horizon manipulation tasks such as \"StackThree_D0\", which demonstrates the efficacy of the self-reflection framework." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "content": "The OpenVLA model demonstrates strong performance in certain long-horizon tasks, leveraging its end-to-end action token prediction capability. However, the lack of observation history and action chunking poses significant challenges in handling complex, fine-grained manipulation tasks like Threading_D0." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 475, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 654 + ], + "type": "text", + "content": "Notably, our Phoenix method achieves more substantial improvements than the subgoal self-reflection method, demonstrating the effectiveness of motion-conditioned method in long-horizon sequential tasks and fine-grained manipulation tasks. Benefiting from our motion-based correction method, agent could correct fine-grained action through motion instruction adjustment while the subgoal-conditioned self-reflection model fails to recover from most failure situations. Besides, the human intervention method achieves high success rates across multiple tasks, demonstrating that our motion-conditioned diffusion policy can effectively adhere to motion instructions for manipulation tasks. This result indicates that our method can perform well under the correct motion instructions, showcasing the significant potential of motion-conditioned self-reflection." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 670, + 419, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 419, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 419, + 681 + ], + "type": "text", + "content": "4.2.2 Ablation Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "In this work, we propose a motion prediction module to provide initial motion instruction, and a motion correction" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6986" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 553, + 114 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 114 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 114 + ], + "type": "table", + "html": "
MethodsCoffee_D0Coffee_D1Stack_D0Stack_D1StackThree_D0StackThree_D1Threading_D0ThreePieceAssembly_D0ThreePieceAssembly_D1Mean
Motion-conditioned68%32%92%84%38%16%58%30%4%46.9%
Expert-Correction Mixture74%36%94%86%38%22%64%30%2%49.6%
Expert-Correction Mixture with Self-Reflection76%30%92%90%46%26%64%34%4%51.3%
Phoenix (Ours)94%48%96%86%50%20%68%52%6%57.8%
", + "image_path": "0ef03f3caa711c400e78dd9f77f37892bd120d4ed5a3ef8e9e158302329fb535.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 123, + 555, + 145 + ], + "lines": [ + { + "bbox": [ + 55, + 123, + 555, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 555, + 145 + ], + "type": "text", + "content": "Table 2. The ablation results of our dual-process motion adjustment mechanism. The results prove that our model, which separates motion prediction module and motion correction module, can provide more precise motion adjustment for action execution in manipulation tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 148, + 206, + 245 + ], + "blocks": [ + { + "bbox": [ + 58, + 148, + 206, + 245 + ], + "lines": [ + { + "bbox": [ + 58, + 148, + 206, + 245 + ], + "spans": [ + { + "bbox": [ + 58, + 148, + 206, + 245 + ], + "type": "image", + "image_path": "6c5e16dfc060a6d846025f0f7976d778a895c4c8af80581777bb3ef876c3239e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 555, + 278 + ], + "type": "text", + "content": "Figure 4. The lifelong learning results. The results prove that our motion-based self-reflection method could iteratively improve performance through interactions with environments." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 213, + 148, + 392, + 245 + ], + "blocks": [ + { + "bbox": [ + 213, + 148, + 392, + 245 + ], + "lines": [ + { + "bbox": [ + 213, + 148, + 392, + 245 + ], + "spans": [ + { + "bbox": [ + 213, + 148, + 392, + 245 + ], + "type": "image", + "image_path": "c5c7d13d24316b6a66738f0e5c7fd5271370f288a41756906c517321701d876f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 402, + 148, + 551, + 245 + ], + "blocks": [ + { + "bbox": [ + 402, + 148, + 551, + 245 + ], + "lines": [ + { + "bbox": [ + 402, + 148, + 551, + 245 + ], + "spans": [ + { + "bbox": [ + 402, + 148, + 551, + 245 + ], + "type": "image", + "image_path": "e693f94a9c4eddc95f4716ffb5fd1cf83742b4809eed02076e406ab513d81989.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 289, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 295, + 373 + ], + "type": "text", + "content": "module to provide fine-grained motion correction. Drawing upon prior research [18, 33], data mixture proportions could influence the efficacy of LLMs. In this section, we investigate whether integrating expert demonstrations with correction dataset, could also enhance the perception and decision-making capabilities of MLLMs for robotic manipulation with the following ablation methods:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 374, + 295, + 458 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 374, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 409 + ], + "type": "text", + "content": "- Expert-Correction Mixture. We mix the expert demonstration and correction data for co-training the motion prediction model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 410, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 458 + ], + "type": "text", + "content": "- Expert-Correction Mixture with Self-Reflection. We mix the expert demonstration and correction data for constraining a unified model to provide initial motion instruction and adjust the instruction." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 459, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 459, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 459, + 295, + 544 + ], + "type": "text", + "content": "As illustrated in Table 2, the results show that co-training with mixture data yields superior performance compared to models trained exclusively on expert demonstration data. This indicates that combining various types of feedback data can enhance the decision-making and perception capabilities of MLLMs. It also validates the viability of our approach to achieving self-improvement through interaction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 544, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 544, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 544, + 295, + 605 + ], + "type": "text", + "content": "Besides, the mixture training model with self-reflection performs better than the one without self-reflection, which suggests that our designed motion-based self-reflection method can enhance the decision-making capabilities of robots and facilitate the correction of fine-grained actions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": "However, we find that utilizing the mixture of data to train a unified model to serve as both the motion prediction module and motion correction module fails to provide accurate correction information compared to our separated motion correction module. This suggests that the mixture training strategy may not fully leverage the strengths of each dataset to achieve better correction effects under the significant data scale discrepancies (160,000 expert demonstrations vs. 16,000 feedback data). The results indicate" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 289, + 553, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 325 + ], + "type": "text", + "content": "that our dual-process motion adjustment mechanism can effectively leverage the expert demonstration and correction dataset, leading to accurate motion instruction adjustment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 325, + 553, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 553, + 349 + ], + "type": "text", + "content": "We also provide ablation results of our designed motion codebook in Supp.B." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 357, + 518, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 518, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 518, + 370 + ], + "type": "text", + "content": "4.3. The Performance of Lifelong Learning" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 375, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 554, + 471 + ], + "type": "text", + "content": "In this section, we explore whether our Phoenix framework can facilitate lifelong learning through interactions. Concretely, we deploy the motion self-reflection model to interact within the environments and utilize the successful trajectories to iteratively fine-tune our motion prediction model after 10, 30, and 50 rollouts. To avoid catastrophic forgetting, we combine 20 expert demonstrations to co-fine-tune the motion prediction module." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 472, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 591 + ], + "type": "text", + "content": "We compare the lifelong learning ability of our motion-based self-reflection model and subgoal-based self-reflection model. During testing, we record the average success rate over 50 trials. As shown in Figure 4, the subgoal-based lifelong learning fails to enhance model performance during the exploration phase due to its inability to provide fine-grained action correction. In contrast, our method corrects underlying action execution during interactions, allowing the robot to better learn from the refined trajectories, thereby achieving self-improvement." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 599, + 477, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 477, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 477, + 610 + ], + "type": "text", + "content": "4.4. Generalization to Novel Tasks" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "In this section, we evaluate the generalization ability of our Phoenix framework in color disruption and position disruption novel tasks as shown in Figure 5. In the color disruption setting, we replace the red block with the blue block in the Stack_D0 task to verify whether our model could generalize to object manipulation tasks with different visual characteristics. In the position disruption setting, we change the fixed position of the coffee machine to a randomly placed posi" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6987" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 192, + 145 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 192, + 145 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 192, + 145 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 192, + 145 + ], + "type": "image", + "image_path": "ff7531bd49dda9526510e06ff9a40c8e908dea632532c4a17bd62d929b7551d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 149, + 185, + 160 + ], + "lines": [ + { + "bbox": [ + 62, + 149, + 185, + 160 + ], + "spans": [ + { + "bbox": [ + 62, + 149, + 185, + 160 + ], + "type": "text", + "content": "(a) Color Disruption in Stack_D0" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 70, + 332, + 144 + ], + "blocks": [ + { + "bbox": [ + 205, + 70, + 332, + 144 + ], + "lines": [ + { + "bbox": [ + 205, + 70, + 332, + 144 + ], + "spans": [ + { + "bbox": [ + 205, + 70, + 332, + 144 + ], + "type": "image", + "image_path": "3e9b8d47a72d73e864fbab79799a0e51a61d4c2e4e1eff417e148c8dd8f66c9b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 149, + 336, + 160 + ], + "lines": [ + { + "bbox": [ + 200, + 149, + 336, + 160 + ], + "spans": [ + { + "bbox": [ + 200, + 149, + 336, + 160 + ], + "type": "text", + "content": "(b) Position Disruption in Coffee_D0" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 342, + 71, + 553, + 147 + ], + "blocks": [ + { + "bbox": [ + 342, + 71, + 553, + 147 + ], + "lines": [ + { + "bbox": [ + 342, + 71, + 553, + 147 + ], + "spans": [ + { + "bbox": [ + 342, + 71, + 553, + 147 + ], + "type": "image", + "image_path": "18833c50dbe49e46ad8330bd52e6dea83265d86fb26e7537645d67b1bf3565ae.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 149, + 508, + 159 + ], + "lines": [ + { + "bbox": [ + 388, + 149, + 508, + 159 + ], + "spans": [ + { + "bbox": [ + 388, + 149, + 508, + 159 + ], + "type": "text", + "content": "(c) Performance on Novel Tasks" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 209, + 329, + 252 + ], + "blocks": [ + { + "bbox": [ + 55, + 173, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 555, + 206 + ], + "type": "text", + "content": "Figure 5. In the color disruption setting, we replace the red block with the blue block in the Stack_D0 task as shown in (a). In the position disruption setting, we change the position of the coffee machine from a fixed point " + }, + { + "bbox": [ + 55, + 173, + 555, + 206 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 55, + 173, + 555, + 206 + ], + "type": "text", + "content": " to a random position from the rectangle " + }, + { + "bbox": [ + 55, + 173, + 555, + 206 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 55, + 173, + 555, + 206 + ], + "type": "text", + "content": " in the Coffee_D0 task as illustrated in (b). The results in (c) prove that our framework could generalize well to these novel task settings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 209, + 329, + 252 + ], + "lines": [ + { + "bbox": [ + 59, + 209, + 329, + 252 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 329, + 252 + ], + "type": "image", + "image_path": "ce0248ba09f15b6805f14b98a2d013596f17a668fba7d5b3266157b033ca99b7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 252, + 224, + 258 + ], + "lines": [ + { + "bbox": [ + 163, + 252, + 224, + 258 + ], + "spans": [ + { + "bbox": [ + 163, + 252, + 224, + 258 + ], + "type": "text", + "content": "(a) Expert Demonstration" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 258, + 115, + 300 + ], + "blocks": [ + { + "bbox": [ + 59, + 258, + 115, + 300 + ], + "lines": [ + { + "bbox": [ + 59, + 258, + 115, + 300 + ], + "spans": [ + { + "bbox": [ + 59, + 258, + 115, + 300 + ], + "type": "image", + "image_path": "5ce8665994ad26c28bd8bec81a8ff2d3b3962a9d9339aed37c0aa3bb06e6dc76.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 300, + 104, + 305 + ], + "lines": [ + { + "bbox": [ + 60, + 300, + 104, + 305 + ], + "spans": [ + { + "bbox": [ + 60, + 300, + 104, + 305 + ], + "type": "text", + "content": "(b) In-Distribution" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 79, + 319, + 307, + 330 + ], + "lines": [ + { + "bbox": [ + 79, + 319, + 307, + 330 + ], + "spans": [ + { + "bbox": [ + 79, + 319, + 307, + 330 + ], + "type": "text", + "content": "Figure 6. The real-world experiments with different variations." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 129, + 258, + 186, + 300 + ], + "blocks": [ + { + "bbox": [ + 129, + 258, + 186, + 300 + ], + "lines": [ + { + "bbox": [ + 129, + 258, + 186, + 300 + ], + "spans": [ + { + "bbox": [ + 129, + 258, + 186, + 300 + ], + "type": "image", + "image_path": "024c4e6f4e7c4313979289d9106dc2e09a04924651f32266bf37c6fee355f761.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 300, + 182, + 306 + ], + "lines": [ + { + "bbox": [ + 136, + 300, + 182, + 306 + ], + "spans": [ + { + "bbox": [ + 136, + 300, + 182, + 306 + ], + "type": "text", + "content": "(c) Pose Disruption" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 201, + 258, + 257, + 300 + ], + "blocks": [ + { + "bbox": [ + 201, + 258, + 257, + 300 + ], + "lines": [ + { + "bbox": [ + 201, + 258, + 257, + 300 + ], + "spans": [ + { + "bbox": [ + 201, + 258, + 257, + 300 + ], + "type": "image", + "image_path": "bf67ba9076b879daff6d955408fa819b7c6661dd39fff48b01a12572fd3222ca.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 300, + 263, + 306 + ], + "lines": [ + { + "bbox": [ + 199, + 300, + 263, + 306 + ], + "spans": [ + { + "bbox": [ + 199, + 300, + 263, + 306 + ], + "type": "text", + "content": "(d) Background Disruption" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 272, + 258, + 328, + 300 + ], + "blocks": [ + { + "bbox": [ + 272, + 258, + 328, + 300 + ], + "lines": [ + { + "bbox": [ + 272, + 258, + 328, + 300 + ], + "spans": [ + { + "bbox": [ + 272, + 258, + 328, + 300 + ], + "type": "image", + "image_path": "ce12886b7cdfe6a0fcef35c7fd1eb1a26368a65c3e393dbf9588be3b08191c50.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 300, + 328, + 306 + ], + "lines": [ + { + "bbox": [ + 276, + 300, + 328, + 306 + ], + "spans": [ + { + "bbox": [ + 276, + 300, + 328, + 306 + ], + "type": "text", + "content": "(e) Texture Disruption" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 359, + 210, + 552, + 268 + ], + "blocks": [ + { + "bbox": [ + 359, + 210, + 552, + 268 + ], + "lines": [ + { + "bbox": [ + 359, + 210, + 552, + 268 + ], + "spans": [ + { + "bbox": [ + 359, + 210, + 552, + 268 + ], + "type": "table", + "html": "
ModelIn-Dis.Pose Dis.Bg.Tex.
OpenVLA55%30%35%45%
Task60%25%25%45%
Motion60%35%30%50%
Ours75%55%45%65%
", + "image_path": "af487ce48e3e96a076661ae88040af1cb18bcf7f3d8b2ca3aa480b589775fb44.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 359, + 281, + 541, + 316 + ], + "blocks": [ + { + "bbox": [ + 367, + 270, + 524, + 279 + ], + "lines": [ + { + "bbox": [ + 367, + 270, + 524, + 279 + ], + "spans": [ + { + "bbox": [ + 367, + 270, + 524, + 279 + ], + "type": "text", + "content": "Table 3. The real-world experiment results." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 359, + 281, + 541, + 316 + ], + "lines": [ + { + "bbox": [ + 359, + 281, + 541, + 316 + ], + "spans": [ + { + "bbox": [ + 359, + 281, + 541, + 316 + ], + "type": "table", + "html": "
TaskMotion10 rollout30 rollout
In-Dis.60%65%75%
Pose Dis.35%45%50%
", + "image_path": "213aa2ab43c3e53dddad730d8c7d8df044a8415001da7e972996d7bb71647ace.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 378, + 317, + 514, + 327 + ], + "lines": [ + { + "bbox": [ + 378, + 317, + 514, + 327 + ], + "spans": [ + { + "bbox": [ + 378, + 317, + 514, + 327 + ], + "type": "text", + "content": "Table 4. The lifelong learning results" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 342, + 294, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 294, + 365 + ], + "type": "text", + "content": "tion within a specific area in the Coffee_D0 task to verify whether our method could generalize to unseen scenarios." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 54, + 365, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 533 + ], + "type": "text", + "content": "For these novel tasks, although the subgoal-conditioned policy could predict correct high-level semantic subgoal for manipulation, this method fails to predict precise robotic action to complete the tasks. Due to its limitation of providing high-level semantic correction information, the subgoal self-reflection method fails to effectively leverage the knowledge of MLLMs for action correction to manipulation tasks. In contrast, as shown in Figure 5(c), our motion-conditioned policy could generate fine-grained motion instruction to achieve generalizable manipulation benefiting from the perception and inferential capability of MLLMs. Besides, our method could achieve better performance on novel tasks by comprehensively refining the motion instruction with the motion-based self-reflection framework." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 540, + 195, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 195, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 195, + 552 + ], + "type": "text", + "content": "4.5. Real-World Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "In real-world scenarios, we conduct the challenging \"drawer open\" articulated object manipulation task as shown in Fig 6(a), where the robot needs to align gripper with handle through precise rotations to open the drawer. We utilize the spacemouse device to collect 100 expert demonstrations with 14 motion instructions (e.g., \"move arm right\", \"rotate around x-axis\"). We train a motion-conditioned diffusion policy to convert instructions into robotic actions. During the inference, we introduce human-in-the-loop interventions to manually correct failure situations to collect 20 corresponding refined interaction trajectories to train our motion correction module. All models are only fine-tuned on the real-world data." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 342, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 460 + ], + "type": "text", + "content": "To validate generalization, we design 4 settings as shown in Fig 6(b-e). In the pose disruption setting, we change the pose distribution of the drawer. For the background disruption setting, the background color was modified to green. In the texture disruption setting, the texture of the drawer was altered to evaluate performance under significant visual variations. The results in Tab 3 demonstrate the generalization ability of our method. We also evaluate lifelong learning, the results in Tab 4 show that our model achieves self-improvement in real world." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 461, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 498 + ], + "type": "text", + "content": "We also provide more real-world task experiments with a rule-based manipulation policy to prove the effectiveness of our motion-based self-reflection method in Supp.C." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 528, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 555, + 624 + ], + "type": "text", + "content": "In this work, we propose a motion-based self-reflection framework to convert the semantic reflection of MLLMs into fine-grained robotic action correction. Based on this framework, we further automatically improve the model's capability from interactions. We hope this motion-based self-reflection framework could bring insights for enhancing the generalization capabilities of agents in robotic manipulation tasks through the integration of MLLMs." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 634, + 424, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 424, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 424, + 647 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "content": "The project was supported by National Natural Science Foundation of China (NO.62106272), Shanghai AI Laboratory, National Key R&D Program of China (2022ZD0160101) and CCF-Zhipu.AI Large Model Innovation Fund." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6988" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "text", + "content": "[1] Andrew G Barto and Sridhar Mahadevan. Recent advances in hierarchical reinforcement learning. Discrete event dynamic systems, 13:341-379, 2003. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 178 + ], + "type": "text", + "content": "[2] Suneel Belkhale, Tianli Ding, Ted Xiao, Pierre Sermanet, Quon Vuong, Jonathan Tompson, Yevgen Chebotar, Debidatta Dwibedi, and Dorsa Sadigh. Rt-h: Action hierarchies using language. arXiv preprint arXiv:2403.01823, 2024. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 294, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 294, + 236 + ], + "type": "text", + "content": "[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 293 + ], + "type": "text", + "content": "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 294, + 338 + ], + "type": "text", + "content": "[5] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. arXiv preprint arXiv:2303.04137, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 294, + 372 + ], + "type": "text", + "content": "[6] Yinpei Dai, Jayjun Lee, Nima Fazeli, and Joyce Chai. Racer: Rich language-guided failure recovery policies for imitation learning, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 375, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 294, + 429 + ], + "type": "text", + "content": "[7] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 431, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 294, + 464 + ], + "type": "text", + "content": "[8] John H Flavell. Metacognition and cognitive monitoring: A new area of cognitive-developmental inquiry. American psychologist, 34(10):906, 1979. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 466, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 294, + 531 + ], + "type": "text", + "content": "[9] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for generalizable manipulation skills. In International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 576 + ], + "type": "text", + "content": "[10] Dan Horgan, John Quan, David Budden, Gabriel Barth-Maron, Matteo Hessel, Hado Van Hasselt, and David Silver. Distributed prioritized experience replay. arXiv preprint arXiv:1803.00933, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 578, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 294, + 632 + ], + "type": "text", + "content": "[11] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "type": "text", + "content": "[12] Daeyeol Lee, Hyojung Seo, and Min Whan Jung. Neural basis of reinforcement learning and decision making. Annual review of neuroscience, 35(1):287-308, 2012. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[13] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[14] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 150 + ], + "type": "text", + "content": "[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 207 + ], + "type": "text", + "content": "[16] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation. arXiv preprint arXiv:2405.17418, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "type": "text", + "content": "[17] Kehui Liu, Zixin Tang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Coherent: Collaboration of heterogeneous multi-robot system with large language models. arXiv preprint arXiv:2409.15146, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "text", + "content": "[18] Qian Liu, Xiaosen Zheng, Niklas Muennighoff, Guangtao Zeng, Longxu Dou, Tianyu Pang, Jing Jiang, and Min Lin. Regmix: Data mixture as regression for language model pretraining. arXiv preprint arXiv:2407.01492, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "type": "text", + "content": "[19] Jingxian Lu, Wenke Xia, Dong Wang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Koi: Accelerating online imitation learning via hybrid key-state guidance. arXiv preprint arXiv:2408.02912, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "type": "text", + "content": "[20] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. In Conference on Robot Learning (CoRL), 2021. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "type": "text", + "content": "[21] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 466, + 553, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 553, + 530 + ], + "type": "text", + "content": "[22] Xincheng Pang, Wenke Xia, Zhigang Wang, Bin Zhao, Di Hu, Dong Wang, and Xuelong Li. Depth helps: Improving pre-trained rgb-based policy with depth information injection. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7251-7256. IEEE, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 533, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 553, + 576 + ], + "type": "text", + "content": "[23] Shubham Pateria, Budhitama Subagdja, Ah-hwee Tan, and Chai Quek. Hierarchical reinforcement learning: A comprehensive survey. ACM Computing Surveys (CSUR), 54(5): 1-35, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "type": "text", + "content": "[24] Shreyas Sundara Raman, Vanya Cohen, Eric Rosen, Ifrah Idrees, David Paulius, and Stefanie Telsex. Planning with large language models via corrective re-prompting. In NeurIPS 2022 Foundation Models for Decision Making Workshop, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 655 + ], + "type": "text", + "content": "[25] Roger C Schank. What we learn when we learn by doing. Technical report, Technical report, 1995. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "text", + "content": "[26] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[27] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6989" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 664 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[28] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "text", + "content": "[29] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 262 + ], + "type": "text", + "content": "[30] Zihao Wang, Shaofei Cai, Guanzhou Chen, Anji Liu, Xiaojian Ma, and Yitao Liang. Describe, explain, plan and select: Interactive planning with large language models enables open-world multi-task agents. arXiv preprint arXiv:2302.01560, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 296, + 319 + ], + "type": "text", + "content": "[31] Wenke Xia, Dong Wang, Xincheng Pang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Kinematic-aware prompting for generalizable articulated object manipulation with llms. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2073-2080, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 364 + ], + "type": "text", + "content": "[32] Chuyan Xiong, Chengyu Shen, Xiaqi Li, Kaichen Zhou, Jiaming Liu, Ruiping Wang, and Hao Dong. Aic mllm: Autonomous interactive correction mllm for robust robotic manipulation. arXiv preprint arXiv:2406.11548, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 407 + ], + "type": "text", + "content": "[33] Jiasheng Ye, Peiju Liu, Tianxiang Sun, Yunhua Zhou, Jun Zhan, and Xipeng Qiu. Data mixing laws: Optimizing data mixtures by predicting language modeling performance. arXiv preprint arXiv:2403.16952, 2024. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 296, + 463 + ], + "type": "text", + "content": "[34] Naoki Yokoyama, Alex Clegg, Joanne Truong, Eric Undersander, Tsung-Yen Yang, Sergio Arnaud, Sehoon Ha, Dhruv Batra, and Akshara Rai. Asc: Adaptive skill coordination for robotic mobile manipulation. IEEE Robotics and Automation Letters, 9(1):779-786, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 464, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 296, + 520 + ], + "type": "text", + "content": "[35] Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Metaworld: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning, pages 1094-1100. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 296, + 564 + ], + "type": "text", + "content": "[36] Jia Zeng, Qingwen Bu, Bangjun Wang, Wenke Xia, Li Chen, Hao Dong, Haoming Song, Dong Wang, Di Hu, Ping Luo, et al. Learning manipulation by predicting interaction. arXiv preprint arXiv:2406.00439, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 565, + 296, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 296, + 619 + ], + "type": "text", + "content": "[37] Junjie Zhang, Chenjia Bai, Haoran He, Wenke Xia, Zhigang Wang, Bin Zhao, Xiu Li, and Xuelong Li. Sam-e: leveraging visual foundation model with sequence imitation for embodied manipulation. arXiv preprint arXiv:2405.19586, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 621, + 296, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 621, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 621, + 296, + 664 + ], + "type": "text", + "content": "[38] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "6990" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/7ee131d2-1c01-43c3-8f06-0c7615cc10d7_content_list.json b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/7ee131d2-1c01-43c3-8f06-0c7615cc10d7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7520c40185b34d97c6a08bc5d43757418ea85986 --- /dev/null +++ b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/7ee131d2-1c01-43c3-8f06-0c7615cc10d7_content_list.json @@ -0,0 +1,1607 @@ +[ + { + "type": "text", + "text": "PhyS-EdiT: Physics-aware Semantic Image Editing with Text Description", + "text_level": 1, + "bbox": [ + 114, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziqi Cai $^{1,2}$ Shuchen Weng $^{3}$ Yifei Xia $^{1,2}$ Boxin Shi $^{1,2*}$", + "bbox": [ + 259, + 179, + 723, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University", + "bbox": [ + 156, + 205, + 815, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University", + "bbox": [ + 160, + 220, + 813, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Beijing Academy of Artificial Intelligence", + "bbox": [ + 357, + 234, + 614, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "czq@stu.pku.edu.cn, shuchenweng@pku.edu.cn, yfxia@pku.edu.cn, shiboxin@pku.edu.cn", + "bbox": [ + 132, + 255, + 841, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8d461302fdb0a4be85af24ea28eb1e89470b1cabb7ea3db8cc6b3404d6327ba9.jpg", + "image_caption": [ + "Figure 1. Illustration of PhyS-Edit's capabilities in disentangled control over lighting, material properties, and high-level semantics. The symbols $\\uparrow$ and $\\downarrow$ indicate an increase and a decrease in property value, respectively. Rows show variations in lighting environments, while columns show precise adjustments to individual material properties. The rightmost column presents a semantic edit with the prompt: \"The knight is holding the sword.\"" + ], + "image_footnote": [], + "bbox": [ + 81, + 281, + 890, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 527, + 313, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Achieving joint control over material properties, lighting, and high-level semantics in images is essential for applications in digital media, advertising, and interactive design. Existing methods often isolate these properties, lacking a cohesive approach to manipulating them simultaneously. We introduce PhyS-Edit, a novel diffusion-based model that enables precise control over four critical material properties: roughness, metallicity, albedo, and transparency while integrating lighting and semantic adjustments within a single framework. To facilitate this disentangled control, we present PR-TIPS, a large and diverse synthetic dataset designed to improve the disentanglement of material and lighting effects. PhyS-Edit incorporates a dual-network architecture and robust training strategies to balance low-level physical realism with high-level semantic coherence, supporting localized and continuous property adjustments. Extensive experiments demonstrate the superiority of PhyS-Edit in editing both synthetic and real-world images, achieving state-of-the-art performance on material, lighting, and semantic editing tasks.", + "bbox": [ + 75, + 558, + 472, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 527, + 630, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-based image editing models [3, 10, 28, 34, 51] have demonstrated significant success in various professional visual content creation applications, including business design, advertising, and entertainment. Given a text description to provide the editing instruction, these models can effectively manipulate high-level image semantics (e.g., category, pose, and layout), reducing the time and technical expertise required by users.", + "bbox": [ + 496, + 551, + 890, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Complementing the high-level \"what\" and \"where\" of the scene, the low-level physical properties (e.g., material, and lighting) are essential to realistically render \"how\" the visual appearance looks like. Recent research explores physical-aware image editing using pairwise rendered training data and incorporating physical principles. These studies have achieved separate control over material [38, 40] or lighting [4, 47, 48, 52].", + "bbox": [ + 496, + 672, + 892, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, despite these advances, two primary challenges remain: (i) Incomplete material-lighting disentanglement. The complex interplay between material and lighting often results in incomplete editing when only one aspect is considered [38, 52], limiting the control capability needed for precise adjustments and hindering the accurate rendering of low-level physical properties. (ii) Limited se", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 880, + 220, + 893 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7867", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mantic editing. Current physics-aware editing approaches focus primarily on material and lighting control, overlooking the alignment with high-level image semantics [38, 47]. It is still challenging to balance between physical realism and semantic consistency.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose PhysS-EdiT, a unified framework for Physics-aware Semantic image Editing with Text descriptions that leverages diffusion models for precise manipulation of material properties, lighting, and high-level semantics. As demonstrated in Fig. 1, our versatile approach allows for modifications ranging from varying lighting conditions (first column) to adjustments of metallic levels, roughness, transparency, albedo, and semantics (subsequent columns), producing consistent and realistic results. Specifically, for low-level physical properties, our PhysS-EdiT supports fine-grained disentangled control over lighting and material properties. These properties are represented as a joint feature maps, which guide a conditional network in injecting such low-level physical knowledge into the diffusion models. For high-level semantics, our PhysS-EdiT combines text-based instructions with accurate physical conditions using a fusion network, maintaining the full capability of semantic editing.", + "bbox": [ + 75, + 166, + 470, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For effective training and evaluating relevant editing models, we create and collect the PR-TIPS, a large-scale dataset of Physical Rendering-based Text and Image Pair Set. PR-TIPS contains an extensive collection of 3D-rendered scenes with systematically varied lighting and material properties, providing a rich source of paired data for disentangled control over physical properties. Our contributions can be summarized as follows:", + "bbox": [ + 75, + 438, + 468, + 556 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a unified framework for controllable editing of both disentangled low-level physical properties and flexible high-level image semantics.", + "- We develop a robust training strategy that minimizes the ambiguity between low-level and high-level editing goals, achieving versatility without compromising precision.", + "- We present PR-TIPS, a large-scale dataset tailored for disentangled control over physical properties, providing a rich source of paired data for training and evaluation." + ], + "bbox": [ + 76, + 559, + 468, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 715, + 212, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion Models for Image Generation", + "text_level": 1, + "bbox": [ + 76, + 741, + 413, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [18, 39] iteratively denoise a Gaussian distribution to generate images, learning to reverse the forward process where Gaussian noise is progressively added to an image. Recent advances have led to their dominance in image generation, outperforming other generative models [16, 22]. To further reduce computing cost and generate high-resolution images, Stable Diffusion [34] is designed. Its success inspires subsequent image generation works [9, 31], and image editing methods discussed later.", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Text-driven Semantic Image Editing", + "text_level": 1, + "bbox": [ + 500, + 90, + 815, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-driven image editing has gained significant research interest, as text descriptions offer an intuitive and flexible means to guide image manipulations. Early work [46, 49] focuses on controlling image category via text guidance. Subsequent research [26, 50] explores effective methods for injecting text features into image generation models. The emergence of pre-trained text-image models like CLIP [32] enables the open-vocabulary editing approaches with promising applications [6, 25]. The largescale image-text datasets [36, 37] further motivate the development of high-quality, open-source text-to-image models [35]. Leveraging the generative priors of these models, recent advancements have been made in super resolution [15, 41], image colorization [44, 45], and reflection removal [19, 54]. Despite substantial progress in high-level semantic editing, text-driven methods still struggle to comprehend and accurately manipulate low-level physical properties, such as lighting and material. This limitation often results in physically unrealistic outputs when precise control over such properties is needed.", + "bbox": [ + 496, + 112, + 890, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Physics-aware Image Property Editing", + "text_level": 1, + "bbox": [ + 498, + 428, + 831, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physics-aware image property editing, which includes material editing and lighting editing, confronts the highly ill-posed challenge of decomposing these elements from a single image. Early explorations [8, 12, 13, 29] utilize empirical formulas to approximate the separation of material and lighting properties. Recent advancements in diffusion models have demonstrated their potential in effectively separating material and lighting components. Notably, the recent works [1, 4, 27, 48] have demonstrated the powerful capabilities of diffusion models, producing impressive results in editing physical properties. However, these approaches typically focus on either lighting [47, 52] or material [38, 40], struggling to provide a comprehensive understanding of the full spectrum of physical properties inherent in images. The gap between high-level semantic editing and low-level physics-aware editing highlights the pressing need for an approach that effectively integrates these properties into a joint framework.", + "bbox": [ + 496, + 450, + 890, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 500, + 739, + 588, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although previous work [38] collects a dataset with basic material variations (e.g., roughness, metallicity, albedo, and transparency), it lacks pixel-aligned maps for fine-grained local editing and pairwise data for disentangled control of material and lighting. Similarly, DiLightNet [47] provides a large dataset with diverse lighting conditions, but it does not support material editing. To address these limitations, we introduce PR-TIPS, a dataset of Physical Rendering-based Text and Image Pair Set, designed to scale up the diversity", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7868", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and scope of scenes and objects. PR-TIPS provides extensive pairwise data, including pixel-aligned images and text-image pairs with varied instructions, supporting more fine-grained disentangled control over both material and lighting adjustments.", + "bbox": [ + 75, + 90, + 468, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Material diversity. We collect 3D assets from a subset of Objaverse [7], selecting 13K models that possess Principled BSDF material properties. To further adjust material properties continuously, we integrate three ShaderNodeMath nodes, applying random values to properties (e.g., Metallic, Roughness, and Transmission Weight). Additionally, we use a ShaderNodeMixRGB node to blend albedo color variations, supporting fine-grained color transitions. Rendering is performed in Blender's Cycles renderer [5], ensuring precise physical interaction with lighting.", + "bbox": [ + 75, + 167, + 467, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lighting diversity. To achieve lighting diversity, we utilize over 500 HDRI environment maps from Polyhaven1, applying random rotations to each to simulate diverse lighting conditions. Camera configurations are randomized for field of view, distance, and position around the scene origin, allowing a wide range of perspectives. This setup generates a final dataset, including 300K rendered images to provide substantial variation in materials, lighting, and viewpoints to support robust training.", + "bbox": [ + 75, + 319, + 467, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text annotation process. We employ a refined text annotation strategy using GPT-4o [11], a Multi-modal Large Language Model (MLLM) with Chain of Thought (CoT) prompting techniques [43]. The model generates detailed text descriptions for both the original and edited images in three structured steps: (i) Identify the scene context (e.g., day or night, indoor or outdoor) and the object type, (ii) Infer the object's material properties, leveraging knowledge of common materials and typical lighting effects, and (iii) Formulate an editing instruction based on the object type, material attributes, and lighting adjustments.", + "bbox": [ + 75, + 457, + 467, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset statistics. PR-TIPS consists of 300K pairs, with 298K for training and 2K for evaluation. To improve data diversity and promote model generalization, we apply various random augmentations to each sample on the fly during training. Specifically, we use random rotations within the range $r \\in [-30^{\\circ}, 30^{\\circ}]$ , scaling factors $s \\in [0.9, 1.1]$ and shear transformations $sh \\in [-10^{\\circ}, 10^{\\circ}]$ . The translations are applied in both horizontal and vertical directions, with the offsets $t_x$ and $t_y$ drawn from the range $[-0.1, 0.1] \\times$ image dimensions. Additionally, horizontal and vertical flips are applied with a probability of $p = 0.5$ . These augmentations are intended to improve the robustness and generalization of the model by simulating a wide range of transformations during training.", + "bbox": [ + 75, + 623, + 467, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Summary. PR-TIPS thus provides an extensive and versatile dataset for physics-aware image editing, supporting disentangled control over material and lighting properties.", + "bbox": [ + 75, + 835, + 467, + 881 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 500, + 89, + 633, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section begins with an overview of our framework in Sec. 4.1. Next, we discuss how our model extracts and injects physical properties in Sec. 4.2. Finally, we integrate physical knowledge to image semantic editing models in Sec. 4.3, supporting both low-level detail fidelity and high-level semantic consistency.", + "bbox": [ + 496, + 114, + 890, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 212, + 609, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, we propose PhyS-EdiT, a physics-based framework to improve text-based image editing with physical awareness.", + "bbox": [ + 496, + 234, + 890, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Physical embedding. The physical properties, i.e., lighting conditions and material conditions, are represented as condition maps $C_L$ and $C_M$ , respectively. These structured condition maps offer spatially aligned hints about the physical world, providing our model with robust priors for lighting and material properties.", + "bbox": [ + 496, + 281, + 890, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Forward process. In the forward process of our diffusion model [18, 39], the latent variable $z_{t}$ at each timestep $t$ is represented as a linear combination of the initial latent $z_{0}$ and Gaussian noise $\\epsilon$ :", + "bbox": [ + 496, + 371, + 890, + 431 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon . \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 438, + 890, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising networks. Our model incorporates a dual-denoising network design, composed of a low-level editing network $U_{\\mathrm{low}}$ and a high-level editing network $U_{\\mathrm{high}}$ . The low-level editing network $U_{\\mathrm{low}}$ , structured as a U-Net [3], integrates a new physical condition injection module to guide fine-grained adjustments based on material and lighting conditions. Concurrently, the high-level editing network $U_{\\mathrm{high}}$ performs broader semantic modifications. Finally, a fusion network $U_{\\mathrm{fusion}}$ combines the outputs of $U_{\\mathrm{low}}$ and $U_{\\mathrm{high}}$ , producing a refined result that balances physical condition editing with global semantic consistency. This dual-network approach enables the editing of precise physical conditions without compromising the ability of the model to make coherent semantic edits.", + "bbox": [ + 496, + 460, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Backward process. The backward process involves iterative denoising, where the dual networks $U_{\\mathrm{low}}$ and $U_{\\mathrm{high}}$ gradually refine the latent code. The denoising networks $U_{\\mathrm{low}}$ and $U_{\\mathrm{high}}$ first independently process the noisy latent code $\\hat{z}_t$ according to their respective roles. At each timestep $t$ , the fusion network $U_{\\mathrm{fusion}}$ then combines the outputs from $U_{\\mathrm{low}}$ and $U_{\\mathrm{high}}$ to yield the next-step latent code:", + "bbox": [ + 496, + 672, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\hat {z} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} U _ {\\text {f u s i o n}} (\\hat {z} _ {t}, t)\\right) + \\sigma_ {t} \\epsilon , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 782, + 890, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s$ denotes the accumulated noise factor, and $\\sigma_t$ is the standard deviation of the noise added at each step. Through $T$ iterations, the model refines the latent code to produce $\\hat{z}_0$ , achieving a balance between low-level physical editing and high-level semantic coherence.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://polyhaven.com", + "bbox": [ + 94, + 886, + 220, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7869", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7d7226d5f118a963093c860b476ab5b30f599026d0e1e824612ebfc7d00ecd6f.jpg", + "image_caption": [ + "Figure 2. Pipeline of PhyS-EdiT. Given an input image with the desired lighting condition $L$ and material condition $M$ , the model first uses the input image and $L$ to generate lighting condition maps and, concurrently, the input image and $M$ to generate material maps. These maps condition the low-level editing network $U_{\\mathrm{low}}$ . Simultaneously, a semantic text prompt modulates the high-level editing network $U_{\\mathrm{high}}$ . The outputs from $U_{\\mathrm{low}}$ and $U_{\\mathrm{high}}$ are then combined through the noise fusion module, resulting in the final edited image." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Material-lighting Joint Representation", + "text_level": 1, + "bbox": [ + 76, + 349, + 410, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While previous generative models [3, 34] excel at generating visually appealing images, they often fall short in adhering to real-world physical principles, such as accurate lighting dynamics and material interactions (e.g., realistic reflections). This limitation arises from their inability to interpret or integrate physical properties into the generative process. In this section, we introduce a novel approach that incorporates physical properties into image editing models, enabling physics-aware image editing.", + "bbox": [ + 75, + 372, + 468, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Physical property extraction. We categorize physical properties into two essential components: material and lighting. Previous approaches often focus on controlling either material or lighting independently; however, our goal is to achieve unified control over both. The primary challenge involves representing these properties in a format that diffusion models can interpret effectively. Traditional input formats, such as text prompts for material properties or environment maps for lighting conditions, often fail to maintain spatial alignment, thereby complicating the model's ability to understand and utilize the physical context effectively.", + "bbox": [ + 75, + 510, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we leverage pretrained models [48] to extract well-aligned physical property maps from the input image $I$ . We define the material condition map $C_M$ as $C_M = [A,R,M,T]$ , where $A$ , $R$ , $M$ , and $T$ represent albedo, roughness, metallicity, and transparency maps, respectively. For the lighting condition map $C_L$ , we utilize an off-the-shelf depth estimator [2] to reconstruct a coarse geometry of the scene following [47]. This geometry is then rendered under predefined proxy materials using environment maps. We define the lighting condition as follows:", + "bbox": [ + 75, + 676, + 470, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {L} = f _ {\\text {l i g h t i n g}} (I, E), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 844, + 468, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{lighting}}$ represents the function that determines the lighting maps based on the input image $I$ and environment", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "map $E$ . Finally we combine the lighting condition map $C_L$ with the material condition map $C_M$ to form the physical condition map $C_{\\mathrm{phys}}$ :", + "bbox": [ + 496, + 349, + 890, + 397 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {\\text {p h y s}} = C _ {M} \\oplus C _ {L}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 409, + 890, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes the channel-wise concatenation. By providing the model with material-lighting joint inputs $C_{\\mathrm{phys}}$ , we bridge the representation gap between different physical properties.", + "bbox": [ + 496, + 438, + 890, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Physical embedding injection. With available structured representations of the material and lighting conditions, the next step is to inject these conditions into the diffusion model. We achieve this by encoding the input image $I$ using an encoder $E(\\cdot)$ inspired by the deferred neural relighting architecture [14]. The resulting encoded feature map $F = E(I)$ captures richer information. To capture interactions between material and lighting, we perform a channelwise multiplication between the encoded features and the combined physical conditions:", + "bbox": [ + 496, + 500, + 892, + 651 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {\\text {e m b}} = F \\odot C _ {\\text {p h y s}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 667, + 890, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise multiplication. This operation allows the model to learn complex relationships between the physical properties and the visual appearance of the image. The resulting physical embedding $C_{\\mathrm{emb}}$ is then used as a conditioning input for ControlNet [51], which integrates this information into the initial feature spaces of the diffusion model.", + "bbox": [ + 496, + 691, + 890, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {t - 1} = \\operatorname {C o n t r o l N e t} \\left(U _ {\\text {l o w}}, \\hat {z} _ {t}, t, C _ {\\text {e m b}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 811, + 890, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This injection process effectively captures both local details and high-level contextual features, culminating in a compact, physics-aware embedding that meets the conditioning requirements of the diffusion model.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7870", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Bridging Physics to Semantic Image Editing", + "text_level": 1, + "bbox": [ + 76, + 90, + 450, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effective image editing models should seamlessly integrate both high-level semantic understanding and detailed physical property manipulations. Previous training strategies, such as fine-tuning attention layers [24] and using low-rank adaptations like LoRA [21], often struggle to preserve the original capabilities of the pre-trained model while adapting to new tasks. These methods can inadvertently weaken the model's ability to maintain generative priors, leading to suboptimal performance in scenarios that require fine-grained control over image properties, such as material textures and lighting effects. To address these challenges, we introduce the following strategies:", + "bbox": [ + 75, + 112, + 468, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preserving generative priors for semantic consistency. To maintain the semantic coherence and editing quality of the pre-trained diffusion model, we initially freeze the weights of the U-Net architecture, $\\theta_{\\mathrm{U - Net}}$ , for the first 30K iterations. This strategy preserves the foundational generative capabilities of the model during early training. Following this phase, we selectively unfreeze the U-Net's weights, keeping the decoder frozen, and continue training for an additional 15K iterations. This controlled fine-tuning improves the model's capacity to incorporate new conditioning inputs effectively while ensuring high-quality semantic consistency and robust generative performance.", + "bbox": [ + 75, + 296, + 468, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multitask supervision for learning diverse edits. We employ a multitask learning framework that leverages paired datasets for robust training. Our model is trained on an equal mix of our PR-TIPS synthetic dataset and the high-level editing dataset from [3], with a 1:1 ratio. To improve stability, we introduce a training augmentation where, at times, the edited image is identical to the input image, helping the model learn to preserve stability when no edits are required.", + "bbox": [ + 75, + 478, + 468, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sequential training. Our training initiates with material editing, focusing on mastering material representations under consistent lighting conditions. We quantify the material editing loss as:", + "bbox": [ + 75, + 614, + 468, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a t e r i a l}} = \\mathbb {E} _ {I, C _ {M}} \\left[ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, C _ {M}) | | _ {2} ^ {2} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 684, + 468, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C_M$ denotes material conditioning and $z_{t}$ is derived from the forward process described in Equation (1). After completing 15K training steps, we introduce variability in lighting conditions by integrating datasets with diverse lighting scenarios. The loss function then evolves to encompass both material and lighting factors:", + "bbox": [ + 75, + 712, + 468, + 803 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p h y s}} = \\mathbb {E} _ {I, C _ {\\text {p h y s}}} \\left[ | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, C _ {\\text {p h y s}}\\right) | | _ {2} ^ {2} \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 811, + 468, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following an additional 15K steps, we enable semantic control by unfreezing $U_{\\mathrm{fusion}}$ , incorporating semantic training data, and continuing the training over another 15K steps with all conditioning factors activated.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3014d375ff4f2557fcd3d140794d1c8d4cc8e2a6aabb04a484843331e9ca0fb9.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation of material editing methods. Throughout the paper, best performances are highlighted in boldface, and $\\uparrow (\\downarrow)$ means high (lower) is better. Evaluation considers only changes in metallicity and roughness properties." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
IP2P [3]20.850.690.3140.80
Subias et al. [40]†23.000.800.3244.17
Ours26.010.790.1523.38
", + "bbox": [ + 501, + 148, + 888, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This training strategy ensures the model first consolidates its skills in material editing before it is exposed to varying lighting conditions, and subsequently, semantic variations. This sequential training improves disentangle-ment and control, allowing for precise manipulation of material properties, lighting, and semantics.", + "bbox": [ + 496, + 220, + 890, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 318, + 633, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 343, + 718, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We initialize the parameters of PhyS-EdiT from IP2P [3] and perform image editing at a resolution of $512 \\times 512$ pixels. The training is conducted over 16 days on 4 NVIDIA RTX 4090 GPUs with a total batch size of 64. We use the Adam optimizer [23] with a learning rate of $5 \\times 10^{-5}$ . During inference, we utilize the DDIM sampling strategy [39] with 50 sampling steps. For real images lacking ground truth environment maps, we estimate lighting conditions using an off-the-shelf lighting estimator [30]. Additional training details are provided in the supplementary material.", + "bbox": [ + 496, + 366, + 890, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 526, + 684, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the performance of each method using five metrics: (i) FID[17], which measures the distributional distance between generated and real images, reflecting quality and diversity. (ii) SSIM[42], which quantifies the preservation of structural information between edited images and the ground truth. (iii) PSNR[20], which evaluates the visual quality of edited regions, with higher values indicating better fidelity. (iv) CLIP Score[33], which assesses semantic alignment between edited images and text descriptions in the CLIP embedding space. (v) LPIPS [53], which measures perceptual similarity between images, focusing on human visual perception.", + "bbox": [ + 496, + 549, + 890, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 739, + 720, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Material editing quality. To evaluate the quality of material editing, we use a test set of 1K samples. We compare our approach with existing feed-forward-based material editing methods [3, 40], using their publicly available implementations. The results are presented in Tab. 1. Our method achieves the best PSNR, LPIPS, and FID scores, as well as the second-best SSIM.", + "bbox": [ + 496, + 763, + 890, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lighting editing quality. We evaluate the performance of our lighting editing approach using an additional 1K", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7871", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b793aacd0437dc25d47ea1e88320b81831c9443ce9d0fe4ac4e3204a2624a550.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation of lighting editing methods." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
DiligNet [47]17.150.650.3951.08
Ours26.960.800.1321.59
", + "bbox": [ + 78, + 107, + 468, + 162 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7a941770e76da2dc815c194a4be84606bc23fc4e5ebe3503b958876706b50657.jpg", + "table_caption": [ + "Table 3. Quantitative evaluation of semantic editing methods." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓CLIP ↑FID ↓
IP2P [3]15.110.570.410.8145.14
SD 3 [10]14.640.500.470.8152.25
Ours15.370.580.390.8547.91
", + "bbox": [ + 78, + 186, + 468, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "test samples. We compare our results with the existing diffusion-based lighting editing method [47], which allows explicit lighting conditions as input. As shown in Tab. 2, our method achieves the highest scores across all metrics, demonstrating superior capability in lighting editing.", + "bbox": [ + 75, + 260, + 467, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic editing quality. We further evaluate the semantic editing capabilities of our model against state-of-the-art methods, with results shown in Tab. 3. Our method achieves the best PSNR, SSIM, LPIPS, and CLIP scores, and the second best FID.", + "bbox": [ + 75, + 335, + 467, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 421, + 287, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present a visual comparison of our method against baseline approaches for lighting, material, and semantic editing, as shown in Fig. 3. For lighting (red columns), our method effectively adjusts lighting to match the ground truth (GT) while preserving material properties, unlike DiLightNet [47], which is not material-aware and alters the material during lighting adjustments. For material (green columns), our approach produces high-quality images with material properties closely matching the GT while maintaining consistent lighting, whereas Subias et al. [40] fails to modify attributes like albedo and transparency, and IP2P [3] often misinterprets material commands. Finally, for semantic editing (blue columns), our method retains strong high-level editing capabilities, achieving results comparable to or better than IP2P [3] and Stable Diffusion 3 [10].", + "bbox": [ + 75, + 444, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5. User Study", + "text_level": 1, + "bbox": [ + 76, + 680, + 199, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further assess the perceptual quality and consistency of our results, we conduct a user study involving 20 participants. The study is divided into two parts: material editing and lighting editing. In each part, participants are presented with input conditions, original images, and candidate images synthesized by different methods, including our PhyS-EdiT. They are asked to select the most visually pleasing result that best matches the input conditions (preference) and to rate the consistency of the edited images with the original content (consistency).", + "bbox": [ + 75, + 703, + 467, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 4, our method receives the highest preference and consistency scores in both material and lighting editing tasks, indicating a strong user preference and better", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/89d3c8d55cf36fe824ad5810813c134980ea4d628ad9c99ed1e1d8a354bed3b9.jpg", + "table_caption": [ + "Table 4. User study results showing the preference and consistency percentages for each method in material and lighting editing tasks." + ], + "table_footnote": [], + "table_body": "
PropertyMetricMethodScore (%)
MaterialPreferenceOurs68
IP2P [3]18
Subias et al. [40]14
ConsistencyOurs61
IP2P [3]11
Subias et al. [40]28
LightingPreferenceOurs60
DiLightNet [47]40
ConsistencyOurs70
DiLightNet [47]30
", + "bbox": [ + 532, + 121, + 856, + 294 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c2c536d41c14680cf267159cf5198954633f30bcee3b7d9d2886d2cd99553cd0.jpg", + "table_caption": [ + "Table 5. Ablation study results grouped by low-level and high-level editing." + ], + "table_footnote": [], + "table_body": "
AblationPhysical EditingSemantic Editing
PSNR ↑LPIPS ↓CLIP ↑LPIPS ↓
Full Model26.390.360.850.39
w/o WF26.000.380.760.35
w/o JT25.800.390.830.40
", + "bbox": [ + 501, + 330, + 887, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "content preservation in our edited images.", + "bbox": [ + 500, + 415, + 777, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.6. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 436, + 651, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform an ablation study to evaluate the impact of different components of our method.", + "bbox": [ + 498, + 458, + 890, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Without weight freezing (WF). We train the entire network without freezing the pretrained weights. This leads to overfitting and loss of the model's generative priors, reducing the quality of the semantic edits.", + "bbox": [ + 498, + 489, + 890, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Without joint training (JT). We train the material and lighting editing components separately. The lack of joint training leads to inconsistent results when both material and lighting need to be edited simultaneously.", + "bbox": [ + 498, + 551, + 890, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative results of the ablation study are presented in Tab. 5, highlighting the performance of different model versions on PSNR, SSIM, LPIPS, and CLIP scores for low-level (physical) and high-level (semantic) editing tasks. The model version without weight freezing (w/o WF) achieves a slightly improved LPIPS score, as not freezing pretrained weights causes the model to forget its original semantic editing capabilities, producing nearly identical outputs regardless of the semantic editing instructions. Consequently, while the unedited image superficially resembles the edited ground truth, it fails to follow the intended instructions, leading to a notably lower CLIP score.", + "bbox": [ + 496, + 611, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.7. Discussion", + "text_level": 1, + "bbox": [ + 500, + 801, + 617, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Support of pixel-level editing. Our model encodes lighting and material properties as pixel-aligned condition maps, enabling precise, localized editing. For instance, in Fig. 5, we applied transparency adjustments to a plate, making it fully transparent (transparency $= 1$ ) except for the \"Phys", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7872", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/39284727e379a286303d1bf0cf171a4868663889a4bfb5ce09a9b429e85c6641.jpg", + "image_caption": [ + "Figure 3. Visual comparison of our method with baseline approaches. For lighting, we compare with DiLightNet [47]; for material, with InstructPix2Pix [3] and Subias et al. [40]; and for semantics, with InstructPix2Pix [3] and Stable Diffusion 3 [10]. Our method effectively disentangles these properties, allowing for the modification of one attribute (e.g., lighting) while preserving others (e.g., material and semantics) unchanged." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 302, + 738 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0567344f3312d766e2e19494c41789c16fe65cab54b486549424fcacf8ff8513.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 87, + 596, + 738 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c15fd6dff89b729c1bf10034d05eecdb2276c685b1b5e393d7e336f4ac4bb7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 88, + 890, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Edit\" text region, which remains opaque (transparency $= 0$ ).", + "bbox": [ + 76, + 803, + 470, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Support of continuous property control. Our model provides smooth, continuous control over lighting and material properties using a unified map format. By feeding in maps with gradually changing values, the model generates outputs that naturally reflect these subtle, progressive ad", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "justments. This also underscores the ability of the model to disentangle different physical conditions, allowing for independent and precise modifications. As shown in Fig. 4, row (a) demonstrates a lighting-only change, with the lighting condition visualized as a diffuse sphere. The remaining rows show combined changes in material properties along", + "bbox": [ + 496, + 803, + 893, + 897 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7873", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6fd319d7566ed0fb6f4cf54287b300f7a585045aca2fade484e816480e8e82f0.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 890, + 484 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c4497d011d7bf5cc37cbf650f49075448b0bc8159563bd41260240dcf0e07cb.jpg", + "image_caption": [ + "Figure 4. Demonstration of our model's capability for continuous editing across different attributes: (a) only lighting change, (b) metallicity increase, (c) roughness increase, (d) albedo increase, and (e) transparency increase. Our model produces consistent and seamless transitions." + ], + "image_footnote": [], + "bbox": [ + 155, + 521, + 250, + 592 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93330ce07b3fd9c9b3c20efec17c042358bb04618a7ecc7c628a2792ee8ee73f.jpg", + "image_caption": [ + "Figure 5. Visualization of our model's capability for pixel-level editing. By applying customized condition maps to a real image, we achieve localized editing in specific properties." + ], + "image_footnote": [], + "bbox": [ + 300, + 522, + 390, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with lighting: (b) increased metallicity, (c) increased roughness, (d) increased albedo, and (e) increased transparency.", + "bbox": [ + 76, + 648, + 468, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of multi-property control. Learning lighting and material control in conjunction improves the consistency and precision of the model. For example, by accounting for material properties, the model can adjust the lighting while keeping the material consistent. Without this material awareness, the model must rely on implicit assumptions about materials, which can lead to varied and inconsistent results across similar scenes.", + "bbox": [ + 76, + 683, + 468, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Dataset Scaling. Increasing dataset scale improves performance and generalization to unseen materiallighting combinations by exposing the model to diverse lighting and material interactions. A broader dataset enables realistic editing across materials like metal and glass, while a limited dataset risks overfitting.", + "bbox": [ + 76, + 810, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 523, + 617, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce PhyS-EdiT, a unified image editing model that enables simultaneous control over material properties, lighting, and high-level semantics. Alongside this, we present PR-TIPS, a large-scale synthetic dataset designed to improve the disentanglement of lighting and material properties. Our approach leverages the advantages of the diffusion model, i.e., disentanglement and controllability, to advance image editing grounded in physical principles. PhyS-EdiT effectively edits both synthetic and real images, supporting both localized and continuous edits.", + "bbox": [ + 496, + 547, + 890, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. While our model performs precise editing and understanding of single objects, it is limited in its ability to generalize to full scene-level editing. Additionally, our model requires two forward passes for processing, which makes it slower compared to single-pass approaches. Future work could enhance scene-level capabilities through richer datasets and optimized single-pass techniques.", + "bbox": [ + 496, + 700, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 815, + 658, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001 and Doubao Fund. PKU-affiliated authors thank openbayes.com for providing computing resources.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7874", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dina Bashkirova, Arijit Ray, Rupayan Mallick, Sarah Adel Bargal, Jianming Zhang, Ranjay Krishna, and Kate Saenko. Lasagna: Layered score distillation for disentangled object relighting. arXiv preprint arXiv:2312.00833, 2023. 2", + "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 4", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to follow image editing instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3, 4, 5, 6, 7", + "[4] Xi Chen, Sida Peng, Dongchen Yang, Yuan Liu, Bowen Pan, Chengfei Lv, and Xiaowei Zhou. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arXiv preprint arXiv:2404.11593, 2024. 1, 2", + "[5] Blender Online Community. *Blender - a 3D modelling and rendering package*. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 3", + "[6] Guillaume Couairon, Asya Grechka, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Flexit: Towards flexible semantic image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3", + "[8] Katja Doerschner, Huseyin Boyaci, and Laurence T Maloney. Estimating the glossiness transfer function induced by illumination change and testing its transitivity. Journal of Vision, 2010. 2", + "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, 2024. 2", + "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, 2024. 1, 6, 7", + "[11] OpenAI et al. GPT-4 technical report, 2024. 3", + "[12] Roland W Fleming. Visual perception of materials and their properties. Vision research, 2014. 2", + "[13] Roland W Fleming, Ron O Dror, and Edward H Adelson. Real-world illumination and the perception of surface reflectance properties. Journal of Vision, 2003. 2", + "[14] Duan Gao, Guojun Chen, Yue Dong, Pieter Peers, Kun Xu, and Xin Tong. Deferred neural lighting: free-viewpoint relighting from unstructured photographs. ACM Transactions on Graphics, 39(6):1-15, 2020. 4", + "[15] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "super-resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 2", + "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 2017. 5", + "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. 2, 3", + "[19] Yuchen Hong, Haofeng Zhong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. L-DiffER: Single image reflection removal with language-based diffusion model. European Conference on Computer Vision, 2024. 2", + "[20] Alain Hore and Djemel Ziou. Image quality metrics: PSNR vs. SSIM. In International Conference on Pattern Recognition, 2010. 5", + "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5", + "[22] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2", + "[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5", + "[24] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 5", + "[25] Gihyun Kwon and Jong Chul Ye. CLIPstyler: Image style transfer with a single text condition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[26] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. ManiGAN: Text-guided image manipulation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[27] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[28] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2I-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Association for the Advancement of Artificial Intelligence, 2024. 1", + "[29] Rohit Pandey, Sergio Orts-Escalano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul E Debevec, and Sean Ryan Fanello. Total Relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics, 2021. 2", + "[30] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7875", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5", + "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 5", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 4", + "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2", + "[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2", + "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 2", + "[38] Prafull Sharma, Varun Jampani, Yuanzhen Li, Xuhui Jia, Dmitry Lagun, Fredo Durand, Bill Freeman, and Mark Matthews. Alchemist: Parametric control of material properties with diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2", + "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3, 5", + "[40] J Daniel Subias and Manuel Lagunas. In-the-wild material appearance editing using perceptual attributes. In Computer Graphics Forum, 2023. 1, 2, 5, 6, 7", + "[41] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. SinSR: diffusion-based image superresolution in a single step. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5", + "[43] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 2022. 3", + "[44] Shuchen Weng, Jimeng Sun, Yu Li, Si Li, and Boxin Shi. $\\mathrm{CT^2}$ : Colorization transformer via color tokens. In European Conference on Computer Vision, 2022. 2", + "[45] Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, Boxin Shi, et al. L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In Advances in Neural Information Processing Systems, 2024. 2", + "[46] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 2", + "[47] Chong Zeng, Yue Dong, Pieter Peers, Youkang Kong, Hongzhi Wu, and Xin Tong. DilightNet: Fine-grained lighting control for diffusion-based image generation. In ACM SIGGRAPH Conference Papers, 2024. 1, 2, 4, 6, 7", + "[48] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. RGBleftrightarrowX: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH Conference Papers, 2024. 1, 2, 4", + "[49] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In International Conference on Computer Vision, 2017. 2", + "[50] Lisai Zhang, Qingcai Chen, Baotian Hu, and Shuoran Jiang. Text-guided neural image inpainting. In ACM International Conference on Multimedia, 2020. 2", + "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In International Conference on Computer Vision, 2023. 1, 4", + "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In International Conference on Learning Representations, 2025. 1, 2", + "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 5", + "[54] Haofeng Zhong, Yuchen Hong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. Language-guided image reflection separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + ], + "bbox": [ + 501, + 92, + 890, + 854 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7876", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/7ee131d2-1c01-43c3-8f06-0c7615cc10d7_model.json b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/7ee131d2-1c01-43c3-8f06-0c7615cc10d7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a4f3782a0eb8dfac36ff9357f72efbd307935618 --- /dev/null +++ b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/7ee131d2-1c01-43c3-8f06-0c7615cc10d7_model.json @@ -0,0 +1,2244 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "PhyS-EdiT: Physics-aware Semantic Image Editing with Text Description" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.18, + 0.725, + 0.2 + ], + "angle": 0, + "content": "Ziqi Cai\\(^{1,2}\\) Shuchen Weng\\(^{3}\\) Yifei Xia\\(^{1,2}\\) Boxin Shi\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.206, + 0.816, + 0.221 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.161, + 0.221, + 0.815, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.359, + 0.235, + 0.615, + 0.248 + ], + "angle": 0, + "content": "\\(^{3}\\)Beijing Academy of Artificial Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.256, + 0.842, + 0.27 + ], + "angle": 0, + "content": "czq@stu.pku.edu.cn, shuchenweng@pku.edu.cn, yfxia@pku.edu.cn, shiboxin@pku.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.282, + 0.891, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.459, + 0.893, + 0.515 + ], + "angle": 0, + "content": "Figure 1. Illustration of PhyS-Edit's capabilities in disentangled control over lighting, material properties, and high-level semantics. The symbols \\(\\uparrow\\) and \\(\\downarrow\\) indicate an increase and a decrease in property value, respectively. Rows show variations in lighting environments, while columns show precise adjustments to individual material properties. The rightmost column presents a semantic edit with the prompt: \"The knight is holding the sword.\"" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.528, + 0.314, + 0.543 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.473, + 0.862 + ], + "angle": 0, + "content": "Achieving joint control over material properties, lighting, and high-level semantics in images is essential for applications in digital media, advertising, and interactive design. Existing methods often isolate these properties, lacking a cohesive approach to manipulating them simultaneously. We introduce PhyS-Edit, a novel diffusion-based model that enables precise control over four critical material properties: roughness, metallicity, albedo, and transparency while integrating lighting and semantic adjustments within a single framework. To facilitate this disentangled control, we present PR-TIPS, a large and diverse synthetic dataset designed to improve the disentanglement of material and lighting effects. PhyS-Edit incorporates a dual-network architecture and robust training strategies to balance low-level physical realism with high-level semantic coherence, supporting localized and continuous property adjustments. Extensive experiments demonstrate the superiority of PhyS-Edit in editing both synthetic and real-world images, achieving state-of-the-art performance on material, lighting, and semantic editing tasks." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.528, + 0.631, + 0.543 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Text-based image editing models [3, 10, 28, 34, 51] have demonstrated significant success in various professional visual content creation applications, including business design, advertising, and entertainment. Given a text description to provide the editing instruction, these models can effectively manipulate high-level image semantics (e.g., category, pose, and layout), reducing the time and technical expertise required by users." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.893, + 0.794 + ], + "angle": 0, + "content": "Complementing the high-level \"what\" and \"where\" of the scene, the low-level physical properties (e.g., material, and lighting) are essential to realistically render \"how\" the visual appearance looks like. Recent research explores physical-aware image editing using pairwise rendered training data and incorporating physical principles. These studies have achieved separate control over material [38, 40] or lighting [4, 47, 48, 52]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.901 + ], + "angle": 0, + "content": "However, despite these advances, two primary challenges remain: (i) Incomplete material-lighting disentanglement. The complex interplay between material and lighting often results in incomplete editing when only one aspect is considered [38, 52], limiting the control capability needed for precise adjustments and hindering the accurate rendering of low-level physical properties. (ii) Limited se" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.881, + 0.222, + 0.894 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7867" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "mantic editing. Current physics-aware editing approaches focus primarily on material and lighting control, overlooking the alignment with high-level image semantics [38, 47]. It is still challenging to balance between physical realism and semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.167, + 0.471, + 0.438 + ], + "angle": 0, + "content": "In this paper, we propose PhysS-EdiT, a unified framework for Physics-aware Semantic image Editing with Text descriptions that leverages diffusion models for precise manipulation of material properties, lighting, and high-level semantics. As demonstrated in Fig. 1, our versatile approach allows for modifications ranging from varying lighting conditions (first column) to adjustments of metallic levels, roughness, transparency, albedo, and semantics (subsequent columns), producing consistent and realistic results. Specifically, for low-level physical properties, our PhysS-EdiT supports fine-grained disentangled control over lighting and material properties. These properties are represented as a joint feature maps, which guide a conditional network in injecting such low-level physical knowledge into the diffusion models. For high-level semantics, our PhysS-EdiT combines text-based instructions with accurate physical conditions using a fusion network, maintaining the full capability of semantic editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.558 + ], + "angle": 0, + "content": "For effective training and evaluating relevant editing models, we create and collect the PR-TIPS, a large-scale dataset of Physical Rendering-based Text and Image Pair Set. PR-TIPS contains an extensive collection of 3D-rendered scenes with systematically varied lighting and material properties, providing a rich source of paired data for disentangled control over physical properties. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.56, + 0.469, + 0.605 + ], + "angle": 0, + "content": "- We propose a unified framework for controllable editing of both disentangled low-level physical properties and flexible high-level image semantics." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.606, + 0.468, + 0.65 + ], + "angle": 0, + "content": "- We develop a robust training strategy that minimizes the ambiguity between low-level and high-level editing goals, achieving versatility without compromising precision." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.651, + 0.468, + 0.695 + ], + "angle": 0, + "content": "- We present PR-TIPS, a large-scale dataset tailored for disentangled control over physical properties, providing a rich source of paired data for training and evaluation." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.56, + 0.469, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.717, + 0.214, + 0.732 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.415, + 0.758 + ], + "angle": 0, + "content": "2.1. Diffusion Models for Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Diffusion models [18, 39] iteratively denoise a Gaussian distribution to generate images, learning to reverse the forward process where Gaussian noise is progressively added to an image. Recent advances have led to their dominance in image generation, outperforming other generative models [16, 22]. To further reduce computing cost and generate high-resolution images, Stable Diffusion [34] is designed. Its success inspires subsequent image generation works [9, 31], and image editing methods discussed later." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.816, + 0.108 + ], + "angle": 0, + "content": "2.2. Text-driven Semantic Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.113, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Text-driven image editing has gained significant research interest, as text descriptions offer an intuitive and flexible means to guide image manipulations. Early work [46, 49] focuses on controlling image category via text guidance. Subsequent research [26, 50] explores effective methods for injecting text features into image generation models. The emergence of pre-trained text-image models like CLIP [32] enables the open-vocabulary editing approaches with promising applications [6, 25]. The largescale image-text datasets [36, 37] further motivate the development of high-quality, open-source text-to-image models [35]. Leveraging the generative priors of these models, recent advancements have been made in super resolution [15, 41], image colorization [44, 45], and reflection removal [19, 54]. Despite substantial progress in high-level semantic editing, text-driven methods still struggle to comprehend and accurately manipulate low-level physical properties, such as lighting and material. This limitation often results in physically unrealistic outputs when precise control over such properties is needed." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.429, + 0.833, + 0.445 + ], + "angle": 0, + "content": "2.3. Physics-aware Image Property Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.892, + 0.724 + ], + "angle": 0, + "content": "Physics-aware image property editing, which includes material editing and lighting editing, confronts the highly ill-posed challenge of decomposing these elements from a single image. Early explorations [8, 12, 13, 29] utilize empirical formulas to approximate the separation of material and lighting properties. Recent advancements in diffusion models have demonstrated their potential in effectively separating material and lighting components. Notably, the recent works [1, 4, 27, 48] have demonstrated the powerful capabilities of diffusion models, producing impressive results in editing physical properties. However, these approaches typically focus on either lighting [47, 52] or material [38, 40], struggling to provide a comprehensive understanding of the full spectrum of physical properties inherent in images. The gap between high-level semantic editing and low-level physics-aware editing highlights the pressing need for an approach that effectively integrates these properties into a joint framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.589, + 0.755 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although previous work [38] collects a dataset with basic material variations (e.g., roughness, metallicity, albedo, and transparency), it lacks pixel-aligned maps for fine-grained local editing and pairwise data for disentangled control of material and lighting. Similarly, DiLightNet [47] provides a large dataset with diverse lighting conditions, but it does not support material editing. To address these limitations, we introduce PR-TIPS, a dataset of Physical Rendering-based Text and Image Pair Set, designed to scale up the diversity" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7868" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.167 + ], + "angle": 0, + "content": "and scope of scenes and objects. PR-TIPS provides extensive pairwise data, including pixel-aligned images and text-image pairs with varied instructions, supporting more fine-grained disentangled control over both material and lighting adjustments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.319 + ], + "angle": 0, + "content": "Material diversity. We collect 3D assets from a subset of Objaverse [7], selecting 13K models that possess Principled BSDF material properties. To further adjust material properties continuously, we integrate three ShaderNodeMath nodes, applying random values to properties (e.g., Metallic, Roughness, and Transmission Weight). Additionally, we use a ShaderNodeMixRGB node to blend albedo color variations, supporting fine-grained color transitions. Rendering is performed in Blender's Cycles renderer [5], ensuring precise physical interaction with lighting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.468, + 0.456 + ], + "angle": 0, + "content": "Lighting diversity. To achieve lighting diversity, we utilize over 500 HDRI environment maps from Polyhaven1, applying random rotations to each to simulate diverse lighting conditions. Camera configurations are randomized for field of view, distance, and position around the scene origin, allowing a wide range of perspectives. This setup generates a final dataset, including 300K rendered images to provide substantial variation in materials, lighting, and viewpoints to support robust training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.468, + 0.623 + ], + "angle": 0, + "content": "Text annotation process. We employ a refined text annotation strategy using GPT-4o [11], a Multi-modal Large Language Model (MLLM) with Chain of Thought (CoT) prompting techniques [43]. The model generates detailed text descriptions for both the original and edited images in three structured steps: (i) Identify the scene context (e.g., day or night, indoor or outdoor) and the object type, (ii) Infer the object's material properties, leveraging knowledge of common materials and typical lighting effects, and (iii) Formulate an editing instruction based on the object type, material attributes, and lighting adjustments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.468, + 0.836 + ], + "angle": 0, + "content": "Dataset statistics. PR-TIPS consists of 300K pairs, with 298K for training and 2K for evaluation. To improve data diversity and promote model generalization, we apply various random augmentations to each sample on the fly during training. Specifically, we use random rotations within the range \\( r \\in [-30^{\\circ}, 30^{\\circ}] \\), scaling factors \\( s \\in [0.9, 1.1] \\) and shear transformations \\( sh \\in [-10^{\\circ}, 10^{\\circ}] \\). The translations are applied in both horizontal and vertical directions, with the offsets \\( t_x \\) and \\( t_y \\) drawn from the range \\( [-0.1, 0.1] \\times \\) image dimensions. Additionally, horizontal and vertical flips are applied with a probability of \\( p = 0.5 \\). These augmentations are intended to improve the robustness and generalization of the model by simulating a wide range of transformations during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.837, + 0.468, + 0.882 + ], + "angle": 0, + "content": "Summary. PR-TIPS thus provides an extensive and versatile dataset for physics-aware image editing, supporting disentangled control over material and lighting properties." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.635, + 0.108 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.206 + ], + "angle": 0, + "content": "This section begins with an overview of our framework in Sec. 4.1. Next, we discuss how our model extracts and injects physical properties in Sec. 4.2. Finally, we integrate physical knowledge to image semantic editing models in Sec. 4.3, supporting both low-level detail fidelity and high-level semantic consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.213, + 0.61, + 0.227 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.281 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, we propose PhyS-EdiT, a physics-based framework to improve text-based image editing with physical awareness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Physical embedding. The physical properties, i.e., lighting conditions and material conditions, are represented as condition maps \\( C_L \\) and \\( C_M \\), respectively. These structured condition maps offer spatially aligned hints about the physical world, providing our model with robust priors for lighting and material properties." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Forward process. In the forward process of our diffusion model [18, 39], the latent variable \\( z_{t} \\) at each timestep \\( t \\) is represented as a linear combination of the initial latent \\( z_{0} \\) and Gaussian noise \\( \\epsilon \\):" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.439, + 0.892, + 0.455 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon . \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Denoising networks. Our model incorporates a dual-denoising network design, composed of a low-level editing network \\( U_{\\mathrm{low}} \\) and a high-level editing network \\( U_{\\mathrm{high}} \\). The low-level editing network \\( U_{\\mathrm{low}} \\), structured as a U-Net [3], integrates a new physical condition injection module to guide fine-grained adjustments based on material and lighting conditions. Concurrently, the high-level editing network \\( U_{\\mathrm{high}} \\) performs broader semantic modifications. Finally, a fusion network \\( U_{\\mathrm{fusion}} \\) combines the outputs of \\( U_{\\mathrm{low}} \\) and \\( U_{\\mathrm{high}} \\), producing a refined result that balances physical condition editing with global semantic consistency. This dual-network approach enables the editing of precise physical conditions without compromising the ability of the model to make coherent semantic edits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Backward process. The backward process involves iterative denoising, where the dual networks \\( U_{\\mathrm{low}} \\) and \\( U_{\\mathrm{high}} \\) gradually refine the latent code. The denoising networks \\( U_{\\mathrm{low}} \\) and \\( U_{\\mathrm{high}} \\) first independently process the noisy latent code \\( \\hat{z}_t \\) according to their respective roles. At each timestep \\( t \\), the fusion network \\( U_{\\mathrm{fusion}} \\) then combines the outputs from \\( U_{\\mathrm{low}} \\) and \\( U_{\\mathrm{high}} \\) to yield the next-step latent code:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.784, + 0.892, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\hat {z} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} U _ {\\text {f u s i o n}} (\\hat {z} _ {t}, t)\\right) + \\sigma_ {t} \\epsilon , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s\\) denotes the accumulated noise factor, and \\(\\sigma_t\\) is the standard deviation of the noise added at each step. Through \\(T\\) iterations, the model refines the latent code to produce \\(\\hat{z}_0\\), achieving a balance between low-level physical editing and high-level semantic coherence." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)https://polyhaven.com" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7869" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.892, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.895, + 0.345 + ], + "angle": 0, + "content": "Figure 2. Pipeline of PhyS-EdiT. Given an input image with the desired lighting condition \\( L \\) and material condition \\( M \\), the model first uses the input image and \\( L \\) to generate lighting condition maps and, concurrently, the input image and \\( M \\) to generate material maps. These maps condition the low-level editing network \\( U_{\\mathrm{low}} \\). Simultaneously, a semantic text prompt modulates the high-level editing network \\( U_{\\mathrm{high}} \\). The outputs from \\( U_{\\mathrm{low}} \\) and \\( U_{\\mathrm{high}} \\) are then combined through the noise fusion module, resulting in the final edited image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.35, + 0.411, + 0.367 + ], + "angle": 0, + "content": "4.2. Material-lighting Joint Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.47, + 0.51 + ], + "angle": 0, + "content": "While previous generative models [3, 34] excel at generating visually appealing images, they often fall short in adhering to real-world physical principles, such as accurate lighting dynamics and material interactions (e.g., realistic reflections). This limitation arises from their inability to interpret or integrate physical properties into the generative process. In this section, we introduce a novel approach that incorporates physical properties into image editing models, enabling physics-aware image editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.678 + ], + "angle": 0, + "content": "Physical property extraction. We categorize physical properties into two essential components: material and lighting. Previous approaches often focus on controlling either material or lighting independently; however, our goal is to achieve unified control over both. The primary challenge involves representing these properties in a format that diffusion models can interpret effectively. Traditional input formats, such as text prompts for material properties or environment maps for lighting conditions, often fail to maintain spatial alignment, thereby complicating the model's ability to understand and utilize the physical context effectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.471, + 0.83 + ], + "angle": 0, + "content": "To address this, we leverage pretrained models [48] to extract well-aligned physical property maps from the input image \\(I\\). We define the material condition map \\(C_M\\) as \\(C_M = [A,R,M,T]\\), where \\(A\\), \\(R\\), \\(M\\), and \\(T\\) represent albedo, roughness, metallicity, and transparency maps, respectively. For the lighting condition map \\(C_L\\), we utilize an off-the-shelf depth estimator [2] to reconstruct a coarse geometry of the scene following [47]. This geometry is then rendered under predefined proxy materials using environment maps. We define the lighting condition as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.845, + 0.47, + 0.863 + ], + "angle": 0, + "content": "\\[\nC _ {L} = f _ {\\text {l i g h t i n g}} (I, E), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{lighting}} \\) represents the function that determines the lighting maps based on the input image \\( I \\) and environment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.398 + ], + "angle": 0, + "content": "map \\(E\\). Finally we combine the lighting condition map \\(C_L\\) with the material condition map \\(C_M\\) to form the physical condition map \\(C_{\\mathrm{phys}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.41, + 0.891, + 0.427 + ], + "angle": 0, + "content": "\\[\nC _ {\\text {p h y s}} = C _ {M} \\oplus C _ {L}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.498 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes the channel-wise concatenation. By providing the model with material-lighting joint inputs \\(C_{\\mathrm{phys}}\\), we bridge the representation gap between different physical properties." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.893, + 0.652 + ], + "angle": 0, + "content": "Physical embedding injection. With available structured representations of the material and lighting conditions, the next step is to inject these conditions into the diffusion model. We achieve this by encoding the input image \\(I\\) using an encoder \\(E(\\cdot)\\) inspired by the deferred neural relighting architecture [14]. The resulting encoded feature map \\(F = E(I)\\) captures richer information. To capture interactions between material and lighting, we perform a channelwise multiplication between the encoded features and the combined physical conditions:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.668, + 0.891, + 0.686 + ], + "angle": 0, + "content": "\\[\nC _ {\\text {e m b}} = F \\odot C _ {\\text {p h y s}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.892, + 0.797 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise multiplication. This operation allows the model to learn complex relationships between the physical properties and the visual appearance of the image. The resulting physical embedding \\(C_{\\mathrm{emb}}\\) is then used as a conditioning input for ControlNet [51], which integrates this information into the initial feature spaces of the diffusion model." + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.812, + 0.891, + 0.828 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {t - 1} = \\operatorname {C o n t r o l N e t} \\left(U _ {\\text {l o w}}, \\hat {z} _ {t}, t, C _ {\\text {e m b}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This injection process effectively captures both local details and high-level contextual features, culminating in a compact, physics-aware embedding that meets the conditioning requirements of the diffusion model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7870" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.452, + 0.108 + ], + "angle": 0, + "content": "4.3. Bridging Physics to Semantic Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.47, + 0.295 + ], + "angle": 0, + "content": "Effective image editing models should seamlessly integrate both high-level semantic understanding and detailed physical property manipulations. Previous training strategies, such as fine-tuning attention layers [24] and using low-rank adaptations like LoRA [21], often struggle to preserve the original capabilities of the pre-trained model while adapting to new tasks. These methods can inadvertently weaken the model's ability to maintain generative priors, leading to suboptimal performance in scenarios that require fine-grained control over image properties, such as material textures and lighting effects. To address these challenges, we introduce the following strategies:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.47, + 0.477 + ], + "angle": 0, + "content": "Preserving generative priors for semantic consistency. To maintain the semantic coherence and editing quality of the pre-trained diffusion model, we initially freeze the weights of the U-Net architecture, \\(\\theta_{\\mathrm{U - Net}}\\) , for the first 30K iterations. This strategy preserves the foundational generative capabilities of the model during early training. Following this phase, we selectively unfreeze the U-Net's weights, keeping the decoder frozen, and continue training for an additional 15K iterations. This controlled fine-tuning improves the model's capacity to incorporate new conditioning inputs effectively while ensuring high-quality semantic consistency and robust generative performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.614 + ], + "angle": 0, + "content": "Multitask supervision for learning diverse edits. We employ a multitask learning framework that leverages paired datasets for robust training. Our model is trained on an equal mix of our PR-TIPS synthetic dataset and the high-level editing dataset from [3], with a 1:1 ratio. To improve stability, we introduce a training augmentation where, at times, the edited image is identical to the input image, helping the model learn to preserve stability when no edits are required." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.47, + 0.676 + ], + "angle": 0, + "content": "Sequential training. Our training initiates with material editing, focusing on mastering material representations under consistent lighting conditions. We quantify the material editing loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.685, + 0.469, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a t e r i a l}} = \\mathbb {E} _ {I, C _ {M}} \\left[ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, C _ {M}) | | _ {2} ^ {2} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.469, + 0.804 + ], + "angle": 0, + "content": "where \\( C_M \\) denotes material conditioning and \\( z_{t} \\) is derived from the forward process described in Equation (1). After completing 15K training steps, we introduce variability in lighting conditions by integrating datasets with diverse lighting scenarios. The loss function then evolves to encompass both material and lighting factors:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.813, + 0.469, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p h y s}} = \\mathbb {E} _ {I, C _ {\\text {p h y s}}} \\left[ | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, C _ {\\text {p h y s}}\\right) | | _ {2} ^ {2} \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Following an additional 15K steps, we enable semantic control by unfreezing \\( U_{\\mathrm{fusion}} \\), incorporating semantic training data, and continuing the training over another 15K steps with all conditioning factors activated." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation of material editing methods. Throughout the paper, best performances are highlighted in boldface, and \\(\\uparrow (\\downarrow)\\) means high (lower) is better. Evaluation considers only changes in metallicity and roughness properties." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.15, + 0.89, + 0.218 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
IP2P [3]20.850.690.3140.80
Subias et al. [40]†23.000.800.3244.17
Ours26.010.790.1523.38
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.222, + 0.892, + 0.314 + ], + "angle": 0, + "content": "This training strategy ensures the model first consolidates its skills in material editing before it is exposed to varying lighting conditions, and subsequently, semantic variations. This sequential training improves disentangle-ment and control, allowing for precise manipulation of material properties, lighting, and semantics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.319, + 0.634, + 0.336 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.344, + 0.719, + 0.36 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.519 + ], + "angle": 0, + "content": "We initialize the parameters of PhyS-EdiT from IP2P [3] and perform image editing at a resolution of \\(512 \\times 512\\) pixels. The training is conducted over 16 days on 4 NVIDIA RTX 4090 GPUs with a total batch size of 64. We use the Adam optimizer [23] with a learning rate of \\(5 \\times 10^{-5}\\). During inference, we utilize the DDIM sampling strategy [39] with 50 sampling steps. For real images lacking ground truth environment maps, we estimate lighting conditions using an off-the-shelf lighting estimator [30]. Additional training details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.527, + 0.685, + 0.542 + ], + "angle": 0, + "content": "5.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.732 + ], + "angle": 0, + "content": "We evaluate the performance of each method using five metrics: (i) FID[17], which measures the distributional distance between generated and real images, reflecting quality and diversity. (ii) SSIM[42], which quantifies the preservation of structural information between edited images and the ground truth. (iii) PSNR[20], which evaluates the visual quality of edited regions, with higher values indicating better fidelity. (iv) CLIP Score[33], which assesses semantic alignment between edited images and text descriptions in the CLIP embedding space. (v) LPIPS [53], which measures perceptual similarity between images, focusing on human visual perception." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.722, + 0.756 + ], + "angle": 0, + "content": "5.3.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Material editing quality. To evaluate the quality of material editing, we use a test set of 1K samples. We compare our approach with existing feed-forward-based material editing methods [3, 40], using their publicly available implementations. The results are presented in Tab. 1. Our method achieves the best PSNR, LPIPS, and FID scores, as well as the second-best SSIM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Lighting editing quality. We evaluate the performance of our lighting editing approach using an additional 1K" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7871" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.091, + 0.452, + 0.105 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation of lighting editing methods." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.108, + 0.47, + 0.163 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
DiligNet [47]17.150.650.3951.08
Ours26.960.800.1321.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.171, + 0.455, + 0.184 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation of semantic editing methods." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.187, + 0.47, + 0.254 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓CLIP ↑FID ↓
IP2P [3]15.110.570.410.8145.14
SD 3 [10]14.640.500.470.8152.25
Ours15.370.580.390.8547.91
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.468, + 0.336 + ], + "angle": 0, + "content": "test samples. We compare our results with the existing diffusion-based lighting editing method [47], which allows explicit lighting conditions as input. As shown in Tab. 2, our method achieves the highest scores across all metrics, demonstrating superior capability in lighting editing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.468, + 0.412 + ], + "angle": 0, + "content": "Semantic editing quality. We further evaluate the semantic editing capabilities of our model against state-of-the-art methods, with results shown in Tab. 3. Our method achieves the best PSNR, SSIM, LPIPS, and CLIP scores, and the second best FID." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.422, + 0.288, + 0.437 + ], + "angle": 0, + "content": "5.4. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.673 + ], + "angle": 0, + "content": "We present a visual comparison of our method against baseline approaches for lighting, material, and semantic editing, as shown in Fig. 3. For lighting (red columns), our method effectively adjusts lighting to match the ground truth (GT) while preserving material properties, unlike DiLightNet [47], which is not material-aware and alters the material during lighting adjustments. For material (green columns), our approach produces high-quality images with material properties closely matching the GT while maintaining consistent lighting, whereas Subias et al. [40] fails to modify attributes like albedo and transparency, and IP2P [3] often misinterprets material commands. Finally, for semantic editing (blue columns), our method retains strong high-level editing capabilities, achieving results comparable to or better than IP2P [3] and Stable Diffusion 3 [10]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.2, + 0.698 + ], + "angle": 0, + "content": "5.5. User Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.468, + 0.854 + ], + "angle": 0, + "content": "To further assess the perceptual quality and consistency of our results, we conduct a user study involving 20 participants. The study is divided into two parts: material editing and lighting editing. In each part, participants are presented with input conditions, original images, and candidate images synthesized by different methods, including our PhyS-EdiT. They are asked to select the most visually pleasing result that best matches the input conditions (preference) and to rate the consistency of the edited images with the original content (consistency)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 4, our method receives the highest preference and consistency scores in both material and lighting editing tasks, indicating a strong user preference and better" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Table 4. User study results showing the preference and consistency percentages for each method in material and lighting editing tasks." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.122, + 0.857, + 0.295 + ], + "angle": 0, + "content": "
PropertyMetricMethodScore (%)
MaterialPreferenceOurs68
IP2P [3]18
Subias et al. [40]14
ConsistencyOurs61
IP2P [3]11
Subias et al. [40]28
LightingPreferenceOurs60
DiLightNet [47]40
ConsistencyOurs70
DiLightNet [47]30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.301, + 0.892, + 0.329 + ], + "angle": 0, + "content": "Table 5. Ablation study results grouped by low-level and high-level editing." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.332, + 0.888, + 0.412 + ], + "angle": 0, + "content": "
AblationPhysical EditingSemantic Editing
PSNR ↑LPIPS ↓CLIP ↑LPIPS ↓
Full Model26.390.360.850.39
w/o WF26.000.380.760.35
w/o JT25.800.390.830.40
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.416, + 0.778, + 0.431 + ], + "angle": 0, + "content": "content preservation in our edited images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.437, + 0.653, + 0.453 + ], + "angle": 0, + "content": "5.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.891, + 0.489 + ], + "angle": 0, + "content": "We perform an ablation study to evaluate the impact of different components of our method." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.491, + 0.892, + 0.55 + ], + "angle": 0, + "content": "Without weight freezing (WF). We train the entire network without freezing the pretrained weights. This leads to overfitting and loss of the model's generative priors, reducing the quality of the semantic edits." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.552, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Without joint training (JT). We train the material and lighting editing components separately. The lack of joint training leads to inconsistent results when both material and lighting need to be edited simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.793 + ], + "angle": 0, + "content": "The quantitative results of the ablation study are presented in Tab. 5, highlighting the performance of different model versions on PSNR, SSIM, LPIPS, and CLIP scores for low-level (physical) and high-level (semantic) editing tasks. The model version without weight freezing (w/o WF) achieves a slightly improved LPIPS score, as not freezing pretrained weights causes the model to forget its original semantic editing capabilities, producing nearly identical outputs regardless of the semantic editing instructions. Consequently, while the unedited image superficially resembles the edited ground truth, it fails to follow the intended instructions, leading to a notably lower CLIP score." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.618, + 0.817 + ], + "angle": 0, + "content": "5.7. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Support of pixel-level editing. Our model encodes lighting and material properties as pixel-aligned condition maps, enabling precise, localized editing. For instance, in Fig. 5, we applied transparency adjustments to a plate, making it fully transparent (transparency \\(= 1\\)) except for the \"Phys" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7872" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.303, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.088, + 0.597, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.089, + 0.892, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.74, + 0.895, + 0.798 + ], + "angle": 0, + "content": "Figure 3. Visual comparison of our method with baseline approaches. For lighting, we compare with DiLightNet [47]; for material, with InstructPix2Pix [3] and Subias et al. [40]; and for semantics, with InstructPix2Pix [3] and Stable Diffusion 3 [10]. Our method effectively disentangles these properties, allowing for the modification of one attribute (e.g., lighting) while preserving others (e.g., material and semantics) unchanged." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.804, + 0.471, + 0.82 + ], + "angle": 0, + "content": "Edit\" text region, which remains opaque (transparency \\(= 0\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Support of continuous property control. Our model provides smooth, continuous control over lighting and material properties using a unified map format. By feeding in maps with gradually changing values, the model generates outputs that naturally reflect these subtle, progressive ad" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.804, + 0.895, + 0.898 + ], + "angle": 0, + "content": "justments. This also underscores the ability of the model to disentangle different physical conditions, allowing for independent and precise modifications. As shown in Fig. 4, row (a) demonstrates a lighting-only change, with the lighting condition visualized as a diffuse sphere. The remaining rows show combined changes in material properties along" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7873" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.127, + 0.099, + 0.138 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.091, + 0.891, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.489, + 0.893, + 0.517 + ], + "angle": 0, + "content": "Figure 4. Demonstration of our model's capability for continuous editing across different attributes: (a) only lighting change, (b) metallicity increase, (c) roughness increase, (d) albedo increase, and (e) transparency increase. Our model produces consistent and seamless transitions." + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.522, + 0.25, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.523, + 0.392, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.596, + 0.47, + 0.638 + ], + "angle": 0, + "content": "Figure 5. Visualization of our model's capability for pixel-level editing. By applying customized condition maps to a real image, we achieve localized editing in specific properties." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.649, + 0.469, + 0.68 + ], + "angle": 0, + "content": "with lighting: (b) increased metallicity, (c) increased roughness, (d) increased albedo, and (e) increased transparency." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.684, + 0.469, + 0.804 + ], + "angle": 0, + "content": "Impact of multi-property control. Learning lighting and material control in conjunction improves the consistency and precision of the model. For example, by accounting for material properties, the model can adjust the lighting while keeping the material consistent. Without this material awareness, the model must rely on implicit assumptions about materials, which can lead to varied and inconsistent results across similar scenes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Impact of Dataset Scaling. Increasing dataset scale improves performance and generalization to unseen materiallighting combinations by exposing the model to diverse lighting and material interactions. A broader dataset enables realistic editing across materials like metal and glass, while a limited dataset risks overfitting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.524, + 0.619, + 0.539 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.699 + ], + "angle": 0, + "content": "We introduce PhyS-EdiT, a unified image editing model that enables simultaneous control over material properties, lighting, and high-level semantics. Alongside this, we present PR-TIPS, a large-scale synthetic dataset designed to improve the disentanglement of lighting and material properties. Our approach leverages the advantages of the diffusion model, i.e., disentanglement and controllability, to advance image editing grounded in physical principles. PhyS-EdiT effectively edits both synthetic and real images, supporting both localized and continuous edits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Limitations. While our model performs precise editing and understanding of single objects, it is limited in its ability to generalize to full scene-level editing. Additionally, our model requires two forward passes for processing, which makes it slower compared to single-pass approaches. Future work could enhance scene-level capabilities through richer datasets and optimized single-pass techniques." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.66, + 0.832 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001 and Doubao Fund. PKU-affiliated authors thank openbayes.com for providing computing resources." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7874" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Dina Bashkirova, Arijit Ray, Rupayan Mallick, Sarah Adel Bargal, Jianming Zhang, Ranjay Krishna, and Kate Saenko. Lasagna: Layered score distillation for disentangled object relighting. arXiv preprint arXiv:2312.00833, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to follow image editing instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.339 + ], + "angle": 0, + "content": "[4] Xi Chen, Sida Peng, Dongchen Yang, Yuan Liu, Bowen Pan, Chengfei Lv, and Xiaowei Zhou. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arXiv preprint arXiv:2404.11593, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[5] Blender Online Community. *Blender - a 3D modelling and rendering package*. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[6] Guillaume Couairon, Asya Grechka, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Flexit: Towards flexible semantic image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[8] Katja Doerschner, Huseyin Boyaci, and Laurence T Maloney. Estimating the glossiness transfer function induced by illumination change and testing its transitivity. Journal of Vision, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.634 + ], + "angle": 0, + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.388, + 0.732 + ], + "angle": 0, + "content": "[11] OpenAI et al. GPT-4 technical report, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.471, + 0.76 + ], + "angle": 0, + "content": "[12] Roland W Fleming. Visual perception of materials and their properties. Vision research, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[13] Roland W Fleming, Ron O Dror, and Edward H Adelson. Real-world illumination and the perception of surface reflectance properties. Journal of Vision, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[14] Duan Gao, Guojun Chen, Yue Dong, Pieter Peers, Kun Xu, and Xin Tong. Deferred neural lighting: free-viewpoint relighting from unstructured photographs. ACM Transactions on Graphics, 39(6):1-15, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[15] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "super-resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[19] Yuchen Hong, Haofeng Zhong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. L-DiffER: Single image reflection removal with language-based diffusion model. European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[20] Alain Hore and Djemel Ziou. Image quality metrics: PSNR vs. SSIM. In International Conference on Pattern Recognition, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[22] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[24] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.618 + ], + "angle": 0, + "content": "[25] Gihyun Kwon and Jong Chul Ye. CLIPstyler: Image style transfer with a single text condition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[26] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. ManiGAN: Text-guided image manipulation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[27] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[28] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2I-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Association for the Advancement of Artificial Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Rohit Pandey, Sergio Orts-Escalano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul E Debevec, and Sean Ryan Fanello. Total Relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7875" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.471, + 0.217 + ], + "angle": 0, + "content": "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.383 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.471, + 0.441 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[38] Prafull Sharma, Varun Jampani, Yuanzhen Li, Xuhui Jia, Dmitry Lagun, Fredo Durand, Bill Freeman, and Mark Matthews. Alchemist: Parametric control of material properties with diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[40] J Daniel Subias and Manuel Lagunas. In-the-wild material appearance editing using perceptual attributes. In Computer Graphics Forum, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[41] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. SinSR: diffusion-based image superresolution in a single step. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[43] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[44] Shuchen Weng, Jimeng Sun, Yu Li, Si Li, and Boxin Shi. \\(\\mathrm{CT^2}\\): Colorization transformer via color tokens. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[45] Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, Boxin Shi, et al. L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In Advances in Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[46] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[47] Chong Zeng, Yue Dong, Pieter Peers, Youkang Kong, Hongzhi Wu, and Xin Tong. DilightNet: Fine-grained lighting control for diffusion-based image generation. In ACM SIGGRAPH Conference Papers, 2024. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[48] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. RGBleftrightarrowX: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH Conference Papers, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[49] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In International Conference on Computer Vision, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[50] Lisai Zhang, Qingcai Chen, Baotian Hu, and Shuoran Jiang. Text-guided neural image inpainting. In ACM International Conference on Multimedia, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In International Conference on Computer Vision, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In International Conference on Learning Representations, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.8, + 0.892, + 0.855 + ], + "angle": 0, + "content": "[54] Haofeng Zhong, Yuchen Hong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. Language-guided image reflection separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.855 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7876" + } + ] +] \ No newline at end of file diff --git a/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/7ee131d2-1c01-43c3-8f06-0c7615cc10d7_origin.pdf b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/7ee131d2-1c01-43c3-8f06-0c7615cc10d7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..11b157e78e0767b27a4d52839867ce49fa7fdee2 --- /dev/null +++ b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/7ee131d2-1c01-43c3-8f06-0c7615cc10d7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c4a604ff65f5504b05df4630eedf026185a6f2786d9aee6d67370b9a2659b7d +size 6469371 diff --git a/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/full.md b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/full.md new file mode 100644 index 0000000000000000000000000000000000000000..93744d3f060b1420141b8e8b716fc1c514b05ffc --- /dev/null +++ b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/full.md @@ -0,0 +1,321 @@ +# PhyS-EdiT: Physics-aware Semantic Image Editing with Text Description + +Ziqi Cai $^{1,2}$ Shuchen Weng $^{3}$ Yifei Xia $^{1,2}$ Boxin Shi $^{1,2*}$ + +$^{1}$ State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University + +$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University + +$^{3}$ Beijing Academy of Artificial Intelligence + +czq@stu.pku.edu.cn, shuchenweng@pku.edu.cn, yfxia@pku.edu.cn, shiboxin@pku.edu.cn + +![](images/8d461302fdb0a4be85af24ea28eb1e89470b1cabb7ea3db8cc6b3404d6327ba9.jpg) +Figure 1. Illustration of PhyS-Edit's capabilities in disentangled control over lighting, material properties, and high-level semantics. The symbols $\uparrow$ and $\downarrow$ indicate an increase and a decrease in property value, respectively. Rows show variations in lighting environments, while columns show precise adjustments to individual material properties. The rightmost column presents a semantic edit with the prompt: "The knight is holding the sword." + +# Abstract + +Achieving joint control over material properties, lighting, and high-level semantics in images is essential for applications in digital media, advertising, and interactive design. Existing methods often isolate these properties, lacking a cohesive approach to manipulating them simultaneously. We introduce PhyS-Edit, a novel diffusion-based model that enables precise control over four critical material properties: roughness, metallicity, albedo, and transparency while integrating lighting and semantic adjustments within a single framework. To facilitate this disentangled control, we present PR-TIPS, a large and diverse synthetic dataset designed to improve the disentanglement of material and lighting effects. PhyS-Edit incorporates a dual-network architecture and robust training strategies to balance low-level physical realism with high-level semantic coherence, supporting localized and continuous property adjustments. Extensive experiments demonstrate the superiority of PhyS-Edit in editing both synthetic and real-world images, achieving state-of-the-art performance on material, lighting, and semantic editing tasks. + +# 1. Introduction + +Text-based image editing models [3, 10, 28, 34, 51] have demonstrated significant success in various professional visual content creation applications, including business design, advertising, and entertainment. Given a text description to provide the editing instruction, these models can effectively manipulate high-level image semantics (e.g., category, pose, and layout), reducing the time and technical expertise required by users. + +Complementing the high-level "what" and "where" of the scene, the low-level physical properties (e.g., material, and lighting) are essential to realistically render "how" the visual appearance looks like. Recent research explores physical-aware image editing using pairwise rendered training data and incorporating physical principles. These studies have achieved separate control over material [38, 40] or lighting [4, 47, 48, 52]. + +However, despite these advances, two primary challenges remain: (i) Incomplete material-lighting disentanglement. The complex interplay between material and lighting often results in incomplete editing when only one aspect is considered [38, 52], limiting the control capability needed for precise adjustments and hindering the accurate rendering of low-level physical properties. (ii) Limited se + +mantic editing. Current physics-aware editing approaches focus primarily on material and lighting control, overlooking the alignment with high-level image semantics [38, 47]. It is still challenging to balance between physical realism and semantic consistency. + +In this paper, we propose PhysS-EdiT, a unified framework for Physics-aware Semantic image Editing with Text descriptions that leverages diffusion models for precise manipulation of material properties, lighting, and high-level semantics. As demonstrated in Fig. 1, our versatile approach allows for modifications ranging from varying lighting conditions (first column) to adjustments of metallic levels, roughness, transparency, albedo, and semantics (subsequent columns), producing consistent and realistic results. Specifically, for low-level physical properties, our PhysS-EdiT supports fine-grained disentangled control over lighting and material properties. These properties are represented as a joint feature maps, which guide a conditional network in injecting such low-level physical knowledge into the diffusion models. For high-level semantics, our PhysS-EdiT combines text-based instructions with accurate physical conditions using a fusion network, maintaining the full capability of semantic editing. + +For effective training and evaluating relevant editing models, we create and collect the PR-TIPS, a large-scale dataset of Physical Rendering-based Text and Image Pair Set. PR-TIPS contains an extensive collection of 3D-rendered scenes with systematically varied lighting and material properties, providing a rich source of paired data for disentangled control over physical properties. Our contributions can be summarized as follows: + +- We propose a unified framework for controllable editing of both disentangled low-level physical properties and flexible high-level image semantics. +- We develop a robust training strategy that minimizes the ambiguity between low-level and high-level editing goals, achieving versatility without compromising precision. +- We present PR-TIPS, a large-scale dataset tailored for disentangled control over physical properties, providing a rich source of paired data for training and evaluation. + +# 2. Related work + +# 2.1. Diffusion Models for Image Generation + +Diffusion models [18, 39] iteratively denoise a Gaussian distribution to generate images, learning to reverse the forward process where Gaussian noise is progressively added to an image. Recent advances have led to their dominance in image generation, outperforming other generative models [16, 22]. To further reduce computing cost and generate high-resolution images, Stable Diffusion [34] is designed. Its success inspires subsequent image generation works [9, 31], and image editing methods discussed later. + +# 2.2. Text-driven Semantic Image Editing + +Text-driven image editing has gained significant research interest, as text descriptions offer an intuitive and flexible means to guide image manipulations. Early work [46, 49] focuses on controlling image category via text guidance. Subsequent research [26, 50] explores effective methods for injecting text features into image generation models. The emergence of pre-trained text-image models like CLIP [32] enables the open-vocabulary editing approaches with promising applications [6, 25]. The largescale image-text datasets [36, 37] further motivate the development of high-quality, open-source text-to-image models [35]. Leveraging the generative priors of these models, recent advancements have been made in super resolution [15, 41], image colorization [44, 45], and reflection removal [19, 54]. Despite substantial progress in high-level semantic editing, text-driven methods still struggle to comprehend and accurately manipulate low-level physical properties, such as lighting and material. This limitation often results in physically unrealistic outputs when precise control over such properties is needed. + +# 2.3. Physics-aware Image Property Editing + +Physics-aware image property editing, which includes material editing and lighting editing, confronts the highly ill-posed challenge of decomposing these elements from a single image. Early explorations [8, 12, 13, 29] utilize empirical formulas to approximate the separation of material and lighting properties. Recent advancements in diffusion models have demonstrated their potential in effectively separating material and lighting components. Notably, the recent works [1, 4, 27, 48] have demonstrated the powerful capabilities of diffusion models, producing impressive results in editing physical properties. However, these approaches typically focus on either lighting [47, 52] or material [38, 40], struggling to provide a comprehensive understanding of the full spectrum of physical properties inherent in images. The gap between high-level semantic editing and low-level physics-aware editing highlights the pressing need for an approach that effectively integrates these properties into a joint framework. + +# 3. Dataset + +Although previous work [38] collects a dataset with basic material variations (e.g., roughness, metallicity, albedo, and transparency), it lacks pixel-aligned maps for fine-grained local editing and pairwise data for disentangled control of material and lighting. Similarly, DiLightNet [47] provides a large dataset with diverse lighting conditions, but it does not support material editing. To address these limitations, we introduce PR-TIPS, a dataset of Physical Rendering-based Text and Image Pair Set, designed to scale up the diversity + +and scope of scenes and objects. PR-TIPS provides extensive pairwise data, including pixel-aligned images and text-image pairs with varied instructions, supporting more fine-grained disentangled control over both material and lighting adjustments. + +Material diversity. We collect 3D assets from a subset of Objaverse [7], selecting 13K models that possess Principled BSDF material properties. To further adjust material properties continuously, we integrate three ShaderNodeMath nodes, applying random values to properties (e.g., Metallic, Roughness, and Transmission Weight). Additionally, we use a ShaderNodeMixRGB node to blend albedo color variations, supporting fine-grained color transitions. Rendering is performed in Blender's Cycles renderer [5], ensuring precise physical interaction with lighting. + +Lighting diversity. To achieve lighting diversity, we utilize over 500 HDRI environment maps from Polyhaven1, applying random rotations to each to simulate diverse lighting conditions. Camera configurations are randomized for field of view, distance, and position around the scene origin, allowing a wide range of perspectives. This setup generates a final dataset, including 300K rendered images to provide substantial variation in materials, lighting, and viewpoints to support robust training. + +Text annotation process. We employ a refined text annotation strategy using GPT-4o [11], a Multi-modal Large Language Model (MLLM) with Chain of Thought (CoT) prompting techniques [43]. The model generates detailed text descriptions for both the original and edited images in three structured steps: (i) Identify the scene context (e.g., day or night, indoor or outdoor) and the object type, (ii) Infer the object's material properties, leveraging knowledge of common materials and typical lighting effects, and (iii) Formulate an editing instruction based on the object type, material attributes, and lighting adjustments. + +Dataset statistics. PR-TIPS consists of 300K pairs, with 298K for training and 2K for evaluation. To improve data diversity and promote model generalization, we apply various random augmentations to each sample on the fly during training. Specifically, we use random rotations within the range $r \in [-30^{\circ}, 30^{\circ}]$ , scaling factors $s \in [0.9, 1.1]$ and shear transformations $sh \in [-10^{\circ}, 10^{\circ}]$ . The translations are applied in both horizontal and vertical directions, with the offsets $t_x$ and $t_y$ drawn from the range $[-0.1, 0.1] \times$ image dimensions. Additionally, horizontal and vertical flips are applied with a probability of $p = 0.5$ . These augmentations are intended to improve the robustness and generalization of the model by simulating a wide range of transformations during training. + +Summary. PR-TIPS thus provides an extensive and versatile dataset for physics-aware image editing, supporting disentangled control over material and lighting properties. + +# 4. Methodology + +This section begins with an overview of our framework in Sec. 4.1. Next, we discuss how our model extracts and injects physical properties in Sec. 4.2. Finally, we integrate physical knowledge to image semantic editing models in Sec. 4.3, supporting both low-level detail fidelity and high-level semantic consistency. + +# 4.1. Overview + +As illustrated in Fig. 2, we propose PhyS-EdiT, a physics-based framework to improve text-based image editing with physical awareness. + +Physical embedding. The physical properties, i.e., lighting conditions and material conditions, are represented as condition maps $C_L$ and $C_M$ , respectively. These structured condition maps offer spatially aligned hints about the physical world, providing our model with robust priors for lighting and material properties. + +Forward process. In the forward process of our diffusion model [18, 39], the latent variable $z_{t}$ at each timestep $t$ is represented as a linear combination of the initial latent $z_{0}$ and Gaussian noise $\epsilon$ : + +$$ +z _ {t} = \sqrt {\alpha_ {t}} z _ {0} + \sqrt {1 - \alpha_ {t}} \epsilon . \tag {1} +$$ + +Denoising networks. Our model incorporates a dual-denoising network design, composed of a low-level editing network $U_{\mathrm{low}}$ and a high-level editing network $U_{\mathrm{high}}$ . The low-level editing network $U_{\mathrm{low}}$ , structured as a U-Net [3], integrates a new physical condition injection module to guide fine-grained adjustments based on material and lighting conditions. Concurrently, the high-level editing network $U_{\mathrm{high}}$ performs broader semantic modifications. Finally, a fusion network $U_{\mathrm{fusion}}$ combines the outputs of $U_{\mathrm{low}}$ and $U_{\mathrm{high}}$ , producing a refined result that balances physical condition editing with global semantic consistency. This dual-network approach enables the editing of precise physical conditions without compromising the ability of the model to make coherent semantic edits. + +Backward process. The backward process involves iterative denoising, where the dual networks $U_{\mathrm{low}}$ and $U_{\mathrm{high}}$ gradually refine the latent code. The denoising networks $U_{\mathrm{low}}$ and $U_{\mathrm{high}}$ first independently process the noisy latent code $\hat{z}_t$ according to their respective roles. At each timestep $t$ , the fusion network $U_{\mathrm{fusion}}$ then combines the outputs from $U_{\mathrm{low}}$ and $U_{\mathrm{high}}$ to yield the next-step latent code: + +$$ +\hat {z} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\hat {z} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} U _ {\text {f u s i o n}} (\hat {z} _ {t}, t)\right) + \sigma_ {t} \epsilon , \tag {2} +$$ + +where $\bar{\alpha}_t = \prod_{s=1}^t \alpha_s$ denotes the accumulated noise factor, and $\sigma_t$ is the standard deviation of the noise added at each step. Through $T$ iterations, the model refines the latent code to produce $\hat{z}_0$ , achieving a balance between low-level physical editing and high-level semantic coherence. + +![](images/7d7226d5f118a963093c860b476ab5b30f599026d0e1e824612ebfc7d00ecd6f.jpg) +Figure 2. Pipeline of PhyS-EdiT. Given an input image with the desired lighting condition $L$ and material condition $M$ , the model first uses the input image and $L$ to generate lighting condition maps and, concurrently, the input image and $M$ to generate material maps. These maps condition the low-level editing network $U_{\mathrm{low}}$ . Simultaneously, a semantic text prompt modulates the high-level editing network $U_{\mathrm{high}}$ . The outputs from $U_{\mathrm{low}}$ and $U_{\mathrm{high}}$ are then combined through the noise fusion module, resulting in the final edited image. + +# 4.2. Material-lighting Joint Representation + +While previous generative models [3, 34] excel at generating visually appealing images, they often fall short in adhering to real-world physical principles, such as accurate lighting dynamics and material interactions (e.g., realistic reflections). This limitation arises from their inability to interpret or integrate physical properties into the generative process. In this section, we introduce a novel approach that incorporates physical properties into image editing models, enabling physics-aware image editing. + +Physical property extraction. We categorize physical properties into two essential components: material and lighting. Previous approaches often focus on controlling either material or lighting independently; however, our goal is to achieve unified control over both. The primary challenge involves representing these properties in a format that diffusion models can interpret effectively. Traditional input formats, such as text prompts for material properties or environment maps for lighting conditions, often fail to maintain spatial alignment, thereby complicating the model's ability to understand and utilize the physical context effectively. + +To address this, we leverage pretrained models [48] to extract well-aligned physical property maps from the input image $I$ . We define the material condition map $C_M$ as $C_M = [A,R,M,T]$ , where $A$ , $R$ , $M$ , and $T$ represent albedo, roughness, metallicity, and transparency maps, respectively. For the lighting condition map $C_L$ , we utilize an off-the-shelf depth estimator [2] to reconstruct a coarse geometry of the scene following [47]. This geometry is then rendered under predefined proxy materials using environment maps. We define the lighting condition as follows: + +$$ +C _ {L} = f _ {\text {l i g h t i n g}} (I, E), \tag {3} +$$ + +where $f_{\mathrm{lighting}}$ represents the function that determines the lighting maps based on the input image $I$ and environment + +map $E$ . Finally we combine the lighting condition map $C_L$ with the material condition map $C_M$ to form the physical condition map $C_{\mathrm{phys}}$ : + +$$ +C _ {\text {p h y s}} = C _ {M} \oplus C _ {L}, \tag {4} +$$ + +where $\oplus$ denotes the channel-wise concatenation. By providing the model with material-lighting joint inputs $C_{\mathrm{phys}}$ , we bridge the representation gap between different physical properties. + +Physical embedding injection. With available structured representations of the material and lighting conditions, the next step is to inject these conditions into the diffusion model. We achieve this by encoding the input image $I$ using an encoder $E(\cdot)$ inspired by the deferred neural relighting architecture [14]. The resulting encoded feature map $F = E(I)$ captures richer information. To capture interactions between material and lighting, we perform a channelwise multiplication between the encoded features and the combined physical conditions: + +$$ +C _ {\text {e m b}} = F \odot C _ {\text {p h y s}}, \tag {5} +$$ + +where $\odot$ denotes element-wise multiplication. This operation allows the model to learn complex relationships between the physical properties and the visual appearance of the image. The resulting physical embedding $C_{\mathrm{emb}}$ is then used as a conditioning input for ControlNet [51], which integrates this information into the initial feature spaces of the diffusion model. + +$$ +\hat {z} _ {t - 1} = \operatorname {C o n t r o l N e t} \left(U _ {\text {l o w}}, \hat {z} _ {t}, t, C _ {\text {e m b}}\right). \tag {6} +$$ + +This injection process effectively captures both local details and high-level contextual features, culminating in a compact, physics-aware embedding that meets the conditioning requirements of the diffusion model. + +# 4.3. Bridging Physics to Semantic Image Editing + +Effective image editing models should seamlessly integrate both high-level semantic understanding and detailed physical property manipulations. Previous training strategies, such as fine-tuning attention layers [24] and using low-rank adaptations like LoRA [21], often struggle to preserve the original capabilities of the pre-trained model while adapting to new tasks. These methods can inadvertently weaken the model's ability to maintain generative priors, leading to suboptimal performance in scenarios that require fine-grained control over image properties, such as material textures and lighting effects. To address these challenges, we introduce the following strategies: + +Preserving generative priors for semantic consistency. To maintain the semantic coherence and editing quality of the pre-trained diffusion model, we initially freeze the weights of the U-Net architecture, $\theta_{\mathrm{U - Net}}$ , for the first 30K iterations. This strategy preserves the foundational generative capabilities of the model during early training. Following this phase, we selectively unfreeze the U-Net's weights, keeping the decoder frozen, and continue training for an additional 15K iterations. This controlled fine-tuning improves the model's capacity to incorporate new conditioning inputs effectively while ensuring high-quality semantic consistency and robust generative performance. + +Multitask supervision for learning diverse edits. We employ a multitask learning framework that leverages paired datasets for robust training. Our model is trained on an equal mix of our PR-TIPS synthetic dataset and the high-level editing dataset from [3], with a 1:1 ratio. To improve stability, we introduce a training augmentation where, at times, the edited image is identical to the input image, helping the model learn to preserve stability when no edits are required. + +Sequential training. Our training initiates with material editing, focusing on mastering material representations under consistent lighting conditions. We quantify the material editing loss as: + +$$ +\mathcal {L} _ {\text {m a t e r i a l}} = \mathbb {E} _ {I, C _ {M}} \left[ | | \epsilon - \epsilon_ {\theta} (z _ {t}, t, C _ {M}) | | _ {2} ^ {2} \right], \tag {7} +$$ + +where $C_M$ denotes material conditioning and $z_{t}$ is derived from the forward process described in Equation (1). After completing 15K training steps, we introduce variability in lighting conditions by integrating datasets with diverse lighting scenarios. The loss function then evolves to encompass both material and lighting factors: + +$$ +\mathcal {L} _ {\text {p h y s}} = \mathbb {E} _ {I, C _ {\text {p h y s}}} \left[ | | \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, C _ {\text {p h y s}}\right) | | _ {2} ^ {2} \right]. \tag {8} +$$ + +Following an additional 15K steps, we enable semantic control by unfreezing $U_{\mathrm{fusion}}$ , incorporating semantic training data, and continuing the training over another 15K steps with all conditioning factors activated. + +Table 1. Quantitative evaluation of material editing methods. Throughout the paper, best performances are highlighted in boldface, and $\uparrow (\downarrow)$ means high (lower) is better. Evaluation considers only changes in metallicity and roughness properties. + +
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
IP2P [3]20.850.690.3140.80
Subias et al. [40]†23.000.800.3244.17
Ours26.010.790.1523.38
+ +This training strategy ensures the model first consolidates its skills in material editing before it is exposed to varying lighting conditions, and subsequently, semantic variations. This sequential training improves disentangle-ment and control, allowing for precise manipulation of material properties, lighting, and semantics. + +# 5. Experiments + +# 5.1. Implementation Details + +We initialize the parameters of PhyS-EdiT from IP2P [3] and perform image editing at a resolution of $512 \times 512$ pixels. The training is conducted over 16 days on 4 NVIDIA RTX 4090 GPUs with a total batch size of 64. We use the Adam optimizer [23] with a learning rate of $5 \times 10^{-5}$ . During inference, we utilize the DDIM sampling strategy [39] with 50 sampling steps. For real images lacking ground truth environment maps, we estimate lighting conditions using an off-the-shelf lighting estimator [30]. Additional training details are provided in the supplementary material. + +# 5.2. Evaluation Metrics + +We evaluate the performance of each method using five metrics: (i) FID[17], which measures the distributional distance between generated and real images, reflecting quality and diversity. (ii) SSIM[42], which quantifies the preservation of structural information between edited images and the ground truth. (iii) PSNR[20], which evaluates the visual quality of edited regions, with higher values indicating better fidelity. (iv) CLIP Score[33], which assesses semantic alignment between edited images and text descriptions in the CLIP embedding space. (v) LPIPS [53], which measures perceptual similarity between images, focusing on human visual perception. + +# 5.3.Quantitative Evaluation + +Material editing quality. To evaluate the quality of material editing, we use a test set of 1K samples. We compare our approach with existing feed-forward-based material editing methods [3, 40], using their publicly available implementations. The results are presented in Tab. 1. Our method achieves the best PSNR, LPIPS, and FID scores, as well as the second-best SSIM. + +Lighting editing quality. We evaluate the performance of our lighting editing approach using an additional 1K + +Table 2. Quantitative evaluation of lighting editing methods. + +
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
DiligNet [47]17.150.650.3951.08
Ours26.960.800.1321.59
+ +Table 3. Quantitative evaluation of semantic editing methods. + +
MethodPSNR ↑SSIM ↑LPIPS ↓CLIP ↑FID ↓
IP2P [3]15.110.570.410.8145.14
SD 3 [10]14.640.500.470.8152.25
Ours15.370.580.390.8547.91
+ +test samples. We compare our results with the existing diffusion-based lighting editing method [47], which allows explicit lighting conditions as input. As shown in Tab. 2, our method achieves the highest scores across all metrics, demonstrating superior capability in lighting editing. + +Semantic editing quality. We further evaluate the semantic editing capabilities of our model against state-of-the-art methods, with results shown in Tab. 3. Our method achieves the best PSNR, SSIM, LPIPS, and CLIP scores, and the second best FID. + +# 5.4. Qualitative Evaluation + +We present a visual comparison of our method against baseline approaches for lighting, material, and semantic editing, as shown in Fig. 3. For lighting (red columns), our method effectively adjusts lighting to match the ground truth (GT) while preserving material properties, unlike DiLightNet [47], which is not material-aware and alters the material during lighting adjustments. For material (green columns), our approach produces high-quality images with material properties closely matching the GT while maintaining consistent lighting, whereas Subias et al. [40] fails to modify attributes like albedo and transparency, and IP2P [3] often misinterprets material commands. Finally, for semantic editing (blue columns), our method retains strong high-level editing capabilities, achieving results comparable to or better than IP2P [3] and Stable Diffusion 3 [10]. + +# 5.5. User Study + +To further assess the perceptual quality and consistency of our results, we conduct a user study involving 20 participants. The study is divided into two parts: material editing and lighting editing. In each part, participants are presented with input conditions, original images, and candidate images synthesized by different methods, including our PhyS-EdiT. They are asked to select the most visually pleasing result that best matches the input conditions (preference) and to rate the consistency of the edited images with the original content (consistency). + +As shown in Tab. 4, our method receives the highest preference and consistency scores in both material and lighting editing tasks, indicating a strong user preference and better + +Table 4. User study results showing the preference and consistency percentages for each method in material and lighting editing tasks. + +
PropertyMetricMethodScore (%)
MaterialPreferenceOurs68
IP2P [3]18
Subias et al. [40]14
ConsistencyOurs61
IP2P [3]11
Subias et al. [40]28
LightingPreferenceOurs60
DiLightNet [47]40
ConsistencyOurs70
DiLightNet [47]30
+ +Table 5. Ablation study results grouped by low-level and high-level editing. + +
AblationPhysical EditingSemantic Editing
PSNR ↑LPIPS ↓CLIP ↑LPIPS ↓
Full Model26.390.360.850.39
w/o WF26.000.380.760.35
w/o JT25.800.390.830.40
+ +content preservation in our edited images. + +# 5.6. Ablation Study + +We perform an ablation study to evaluate the impact of different components of our method. + +Without weight freezing (WF). We train the entire network without freezing the pretrained weights. This leads to overfitting and loss of the model's generative priors, reducing the quality of the semantic edits. + +Without joint training (JT). We train the material and lighting editing components separately. The lack of joint training leads to inconsistent results when both material and lighting need to be edited simultaneously. + +The quantitative results of the ablation study are presented in Tab. 5, highlighting the performance of different model versions on PSNR, SSIM, LPIPS, and CLIP scores for low-level (physical) and high-level (semantic) editing tasks. The model version without weight freezing (w/o WF) achieves a slightly improved LPIPS score, as not freezing pretrained weights causes the model to forget its original semantic editing capabilities, producing nearly identical outputs regardless of the semantic editing instructions. Consequently, while the unedited image superficially resembles the edited ground truth, it fails to follow the intended instructions, leading to a notably lower CLIP score. + +# 5.7. Discussion + +Support of pixel-level editing. Our model encodes lighting and material properties as pixel-aligned condition maps, enabling precise, localized editing. For instance, in Fig. 5, we applied transparency adjustments to a plate, making it fully transparent (transparency $= 1$ ) except for the "Phys + +![](images/39284727e379a286303d1bf0cf171a4868663889a4bfb5ce09a9b429e85c6641.jpg) +Figure 3. Visual comparison of our method with baseline approaches. For lighting, we compare with DiLightNet [47]; for material, with InstructPix2Pix [3] and Subias et al. [40]; and for semantics, with InstructPix2Pix [3] and Stable Diffusion 3 [10]. Our method effectively disentangles these properties, allowing for the modification of one attribute (e.g., lighting) while preserving others (e.g., material and semantics) unchanged. + +![](images/0567344f3312d766e2e19494c41789c16fe65cab54b486549424fcacf8ff8513.jpg) + +![](images/7c15fd6dff89b729c1bf10034d05eecdb2276c685b1b5e393d7e336f4ac4bb7c.jpg) + +Edit" text region, which remains opaque (transparency $= 0$ ). + +Support of continuous property control. Our model provides smooth, continuous control over lighting and material properties using a unified map format. By feeding in maps with gradually changing values, the model generates outputs that naturally reflect these subtle, progressive ad + +justments. This also underscores the ability of the model to disentangle different physical conditions, allowing for independent and precise modifications. As shown in Fig. 4, row (a) demonstrates a lighting-only change, with the lighting condition visualized as a diffuse sphere. The remaining rows show combined changes in material properties along + +![](images/6fd319d7566ed0fb6f4cf54287b300f7a585045aca2fade484e816480e8e82f0.jpg) +(a) + +![](images/6c4497d011d7bf5cc37cbf650f49075448b0bc8159563bd41260240dcf0e07cb.jpg) +Figure 4. Demonstration of our model's capability for continuous editing across different attributes: (a) only lighting change, (b) metallicity increase, (c) roughness increase, (d) albedo increase, and (e) transparency increase. Our model produces consistent and seamless transitions. + +![](images/93330ce07b3fd9c9b3c20efec17c042358bb04618a7ecc7c628a2792ee8ee73f.jpg) +Figure 5. Visualization of our model's capability for pixel-level editing. By applying customized condition maps to a real image, we achieve localized editing in specific properties. + +with lighting: (b) increased metallicity, (c) increased roughness, (d) increased albedo, and (e) increased transparency. + +Impact of multi-property control. Learning lighting and material control in conjunction improves the consistency and precision of the model. For example, by accounting for material properties, the model can adjust the lighting while keeping the material consistent. Without this material awareness, the model must rely on implicit assumptions about materials, which can lead to varied and inconsistent results across similar scenes. + +Impact of Dataset Scaling. Increasing dataset scale improves performance and generalization to unseen materiallighting combinations by exposing the model to diverse lighting and material interactions. A broader dataset enables realistic editing across materials like metal and glass, while a limited dataset risks overfitting. + +# 6. Conclusion + +We introduce PhyS-EdiT, a unified image editing model that enables simultaneous control over material properties, lighting, and high-level semantics. Alongside this, we present PR-TIPS, a large-scale synthetic dataset designed to improve the disentanglement of lighting and material properties. Our approach leverages the advantages of the diffusion model, i.e., disentanglement and controllability, to advance image editing grounded in physical principles. PhyS-EdiT effectively edits both synthetic and real images, supporting both localized and continuous edits. + +Limitations. While our model performs precise editing and understanding of single objects, it is limited in its ability to generalize to full scene-level editing. Additionally, our model requires two forward passes for processing, which makes it slower compared to single-pass approaches. Future work could enhance scene-level capabilities through richer datasets and optimized single-pass techniques. + +# Acknowledgement + +This work is supported by National Natural Science Foundation of China under Grant No. 62136001 and Doubao Fund. PKU-affiliated authors thank openbayes.com for providing computing resources. + +# References + +[1] Dina Bashkirova, Arijit Ray, Rupayan Mallick, Sarah Adel Bargal, Jianming Zhang, Ranjay Krishna, and Kate Saenko. Lasagna: Layered score distillation for disentangled object relighting. arXiv preprint arXiv:2312.00833, 2023. 2 +[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 4 +[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to follow image editing instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3, 4, 5, 6, 7 +[4] Xi Chen, Sida Peng, Dongchen Yang, Yuan Liu, Bowen Pan, Chengfei Lv, and Xiaowei Zhou. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arXiv preprint arXiv:2404.11593, 2024. 1, 2 +[5] Blender Online Community. *Blender - a 3D modelling and rendering package*. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 3 +[6] Guillaume Couairon, Asya Grechka, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Flexit: Towards flexible semantic image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3 +[8] Katja Doerschner, Huseyin Boyaci, and Laurence T Maloney. Estimating the glossiness transfer function induced by illumination change and testing its transitivity. Journal of Vision, 2010. 2 +[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, 2024. 2 +[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, 2024. 1, 6, 7 +[11] OpenAI et al. GPT-4 technical report, 2024. 3 +[12] Roland W Fleming. Visual perception of materials and their properties. Vision research, 2014. 2 +[13] Roland W Fleming, Ron O Dror, and Edward H Adelson. Real-world illumination and the perception of surface reflectance properties. Journal of Vision, 2003. 2 +[14] Duan Gao, Guojun Chen, Yue Dong, Pieter Peers, Kun Xu, and Xin Tong. Deferred neural lighting: free-viewpoint relighting from unstructured photographs. ACM Transactions on Graphics, 39(6):1-15, 2020. 4 +[15] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous + +super-resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 2 +[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 2017. 5 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. 2, 3 +[19] Yuchen Hong, Haofeng Zhong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. L-DiffER: Single image reflection removal with language-based diffusion model. European Conference on Computer Vision, 2024. 2 +[20] Alain Hore and Djemel Ziou. Image quality metrics: PSNR vs. SSIM. In International Conference on Pattern Recognition, 2010. 5 +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5 +[22] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 +[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5 +[24] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 5 +[25] Gihyun Kwon and Jong Chul Ye. CLIPstyler: Image style transfer with a single text condition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[26] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. ManiGAN: Text-guided image manipulation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2 +[27] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[28] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2I-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Association for the Advancement of Artificial Intelligence, 2024. 1 +[29] Rohit Pandey, Sergio Orts-Escalano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul E Debevec, and Sean Ryan Fanello. Total Relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics, 2021. 2 +[30] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook + +Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5 +[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 5 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 4 +[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2 +[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2 +[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 2 +[38] Prafull Sharma, Varun Jampani, Yuanzhen Li, Xuhui Jia, Dmitry Lagun, Fredo Durand, Bill Freeman, and Mark Matthews. Alchemist: Parametric control of material properties with diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2 +[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3, 5 +[40] J Daniel Subias and Manuel Lagunas. In-the-wild material appearance editing using perceptual attributes. In Computer Graphics Forum, 2023. 1, 2, 5, 6, 7 +[41] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. SinSR: diffusion-based image superresolution in a single step. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 + +[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5 +[43] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 2022. 3 +[44] Shuchen Weng, Jimeng Sun, Yu Li, Si Li, and Boxin Shi. $\mathrm{CT^2}$ : Colorization transformer via color tokens. In European Conference on Computer Vision, 2022. 2 +[45] Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, Boxin Shi, et al. L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In Advances in Neural Information Processing Systems, 2024. 2 +[46] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 2 +[47] Chong Zeng, Yue Dong, Pieter Peers, Youkang Kong, Hongzhi Wu, and Xin Tong. DilightNet: Fine-grained lighting control for diffusion-based image generation. In ACM SIGGRAPH Conference Papers, 2024. 1, 2, 4, 6, 7 +[48] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. RGBleftrightarrowX: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH Conference Papers, 2024. 1, 2, 4 +[49] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In International Conference on Computer Vision, 2017. 2 +[50] Lisai Zhang, Qingcai Chen, Baotian Hu, and Shuoran Jiang. Text-guided neural image inpainting. In ACM International Conference on Multimedia, 2020. 2 +[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In International Conference on Computer Vision, 2023. 1, 4 +[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In International Conference on Learning Representations, 2025. 1, 2 +[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 5 +[54] Haofeng Zhong, Yuchen Hong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. Language-guided image reflection separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 \ No newline at end of file diff --git a/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/images.zip b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fad46b03cc34f7dbc20275960a66ab2a7f3f4506 --- /dev/null +++ b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2cab2af2a2d15b907317605867bf037c811184e33473ed9d9c002c1b9e297f2 +size 975359 diff --git a/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/layout.json b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..349dd632f77f1e446e2ff224e60958fbc9570126 --- /dev/null +++ b/2025/PhyS-EdiT_ Physics-aware Semantic Image Editing with Text Description/layout.json @@ -0,0 +1,8446 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "type": "text", + "content": "PhyS-EdiT: Physics-aware Semantic Image Editing with Text Description" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 142, + 443, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 142, + 443, + 158 + ], + "spans": [ + { + "bbox": [ + 159, + 142, + 443, + 158 + ], + "type": "text", + "content": "Ziqi Cai" + }, + { + "bbox": [ + 159, + 142, + 443, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 159, + 142, + 443, + 158 + ], + "type": "text", + "content": " Shuchen Weng" + }, + { + "bbox": [ + 159, + 142, + 443, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 159, + 142, + 443, + 158 + ], + "type": "text", + "content": " Yifei Xia" + }, + { + "bbox": [ + 159, + 142, + 443, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 159, + 142, + 443, + 158 + ], + "type": "text", + "content": " Boxin Shi" + }, + { + "bbox": [ + 159, + 142, + 443, + 158 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 163, + 499, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 163, + 499, + 175 + ], + "spans": [ + { + "bbox": [ + 96, + 163, + 499, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 163, + 499, + 175 + ], + "type": "text", + "content": "State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 98, + 175, + 498, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 175, + 498, + 186 + ], + "spans": [ + { + "bbox": [ + 98, + 175, + 498, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 98, + 175, + 498, + 186 + ], + "type": "text", + "content": "National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 219, + 186, + 376, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 186, + 376, + 196 + ], + "spans": [ + { + "bbox": [ + 219, + 186, + 376, + 196 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 219, + 186, + 376, + 196 + ], + "type": "text", + "content": "Beijing Academy of Artificial Intelligence" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 81, + 202, + 515, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 202, + 515, + 213 + ], + "spans": [ + { + "bbox": [ + 81, + 202, + 515, + 213 + ], + "type": "text", + "content": "czq@stu.pku.edu.cn, shuchenweng@pku.edu.cn, yfxia@pku.edu.cn, shiboxin@pku.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 223, + 545, + 361 + ], + "blocks": [ + { + "bbox": [ + 50, + 223, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 50, + 223, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 50, + 223, + 545, + 361 + ], + "type": "image", + "image_path": "8d461302fdb0a4be85af24ea28eb1e89470b1cabb7ea3db8cc6b3404d6327ba9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 363, + 546, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 363, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 546, + 407 + ], + "type": "text", + "content": "Figure 1. Illustration of PhyS-Edit's capabilities in disentangled control over lighting, material properties, and high-level semantics. The symbols " + }, + { + "bbox": [ + 46, + 363, + 546, + 407 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 363, + 546, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 363, + 546, + 407 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 363, + 546, + 407 + ], + "type": "text", + "content": " indicate an increase and a decrease in property value, respectively. Rows show variations in lighting environments, while columns show precise adjustments to individual material properties. The rightmost column presents a semantic edit with the prompt: \"The knight is holding the sword.\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 418, + 192, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 418, + 192, + 430 + ], + "spans": [ + { + "bbox": [ + 143, + 418, + 192, + 430 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 442, + 289, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 289, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 289, + 682 + ], + "type": "text", + "content": "Achieving joint control over material properties, lighting, and high-level semantics in images is essential for applications in digital media, advertising, and interactive design. Existing methods often isolate these properties, lacking a cohesive approach to manipulating them simultaneously. We introduce PhyS-Edit, a novel diffusion-based model that enables precise control over four critical material properties: roughness, metallicity, albedo, and transparency while integrating lighting and semantic adjustments within a single framework. To facilitate this disentangled control, we present PR-TIPS, a large and diverse synthetic dataset designed to improve the disentanglement of material and lighting effects. PhyS-Edit incorporates a dual-network architecture and robust training strategies to balance low-level physical realism with high-level semantic coherence, supporting localized and continuous property adjustments. Extensive experiments demonstrate the superiority of PhyS-Edit in editing both synthetic and real-world images, achieving state-of-the-art performance on material, lighting, and semantic editing tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 418, + 386, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 386, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 386, + 430 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": "Text-based image editing models [3, 10, 28, 34, 51] have demonstrated significant success in various professional visual content creation applications, including business design, advertising, and entertainment. Given a text description to provide the editing instruction, these models can effectively manipulate high-level image semantics (e.g., category, pose, and layout), reducing the time and technical expertise required by users." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 533, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 628 + ], + "type": "text", + "content": "Complementing the high-level \"what\" and \"where\" of the scene, the low-level physical properties (e.g., material, and lighting) are essential to realistically render \"how\" the visual appearance looks like. Recent research explores physical-aware image editing using pairwise rendered training data and incorporating physical principles. These studies have achieved separate control over material [38, 40] or lighting [4, 47, 48, 52]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "However, despite these advances, two primary challenges remain: (i) Incomplete material-lighting disentanglement. The complex interplay between material and lighting often results in incomplete editing when only one aspect is considered [38, 52], limiting the control capability needed for precise adjustments and hindering the accurate rendering of low-level physical properties. (ii) Limited se" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 697, + 135, + 708 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 697, + 135, + 708 + ], + "spans": [ + { + "bbox": [ + 58, + 697, + 135, + 708 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7867" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "mantic editing. Current physics-aware editing approaches focus primarily on material and lighting control, overlooking the alignment with high-level image semantics [38, 47]. It is still challenging to balance between physical realism and semantic consistency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 132, + 288, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 132, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 132, + 288, + 346 + ], + "type": "text", + "content": "In this paper, we propose PhysS-EdiT, a unified framework for Physics-aware Semantic image Editing with Text descriptions that leverages diffusion models for precise manipulation of material properties, lighting, and high-level semantics. As demonstrated in Fig. 1, our versatile approach allows for modifications ranging from varying lighting conditions (first column) to adjustments of metallic levels, roughness, transparency, albedo, and semantics (subsequent columns), producing consistent and realistic results. Specifically, for low-level physical properties, our PhysS-EdiT supports fine-grained disentangled control over lighting and material properties. These properties are represented as a joint feature maps, which guide a conditional network in injecting such low-level physical knowledge into the diffusion models. For high-level semantics, our PhysS-EdiT combines text-based instructions with accurate physical conditions using a fusion network, maintaining the full capability of semantic editing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 347, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 441 + ], + "type": "text", + "content": "For effective training and evaluating relevant editing models, we create and collect the PR-TIPS, a large-scale dataset of Physical Rendering-based Text and Image Pair Set. PR-TIPS contains an extensive collection of 3D-rendered scenes with systematically varied lighting and material properties, providing a rich source of paired data for disentangled control over physical properties. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 443, + 287, + 550 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 443, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 287, + 479 + ], + "type": "text", + "content": "- We propose a unified framework for controllable editing of both disentangled low-level physical properties and flexible high-level image semantics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 286, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 286, + 514 + ], + "type": "text", + "content": "- We develop a robust training strategy that minimizes the ambiguity between low-level and high-level editing goals, achieving versatility without compromising precision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 515, + 286, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 286, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 286, + 550 + ], + "type": "text", + "content": "- We present PR-TIPS, a large-scale dataset tailored for disentangled control over physical properties, providing a rich source of paired data for training and evaluation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 567, + 130, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 130, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 130, + 579 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 587, + 253, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 253, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 253, + 600 + ], + "type": "text", + "content": "2.1. Diffusion Models for Image Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Diffusion models [18, 39] iteratively denoise a Gaussian distribution to generate images, learning to reverse the forward process where Gaussian noise is progressively added to an image. Recent advances have led to their dominance in image generation, outperforming other generative models [16, 22]. To further reduce computing cost and generate high-resolution images, Stable Diffusion [34] is designed. Its success inspires subsequent image generation works [9, 31], and image editing methods discussed later." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 499, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 499, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 499, + 85 + ], + "type": "text", + "content": "2.2. Text-driven Semantic Image Editing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 89, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 89, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 89, + 545, + 329 + ], + "type": "text", + "content": "Text-driven image editing has gained significant research interest, as text descriptions offer an intuitive and flexible means to guide image manipulations. Early work [46, 49] focuses on controlling image category via text guidance. Subsequent research [26, 50] explores effective methods for injecting text features into image generation models. The emergence of pre-trained text-image models like CLIP [32] enables the open-vocabulary editing approaches with promising applications [6, 25]. The largescale image-text datasets [36, 37] further motivate the development of high-quality, open-source text-to-image models [35]. Leveraging the generative priors of these models, recent advancements have been made in super resolution [15, 41], image colorization [44, 45], and reflection removal [19, 54]. Despite substantial progress in high-level semantic editing, text-driven methods still struggle to comprehend and accurately manipulate low-level physical properties, such as lighting and material. This limitation often results in physically unrealistic outputs when precise control over such properties is needed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 339, + 509, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 339, + 509, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 339, + 509, + 352 + ], + "type": "text", + "content": "2.3. Physics-aware Image Property Editing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 357, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 573 + ], + "type": "text", + "content": "Physics-aware image property editing, which includes material editing and lighting editing, confronts the highly ill-posed challenge of decomposing these elements from a single image. Early explorations [8, 12, 13, 29] utilize empirical formulas to approximate the separation of material and lighting properties. Recent advancements in diffusion models have demonstrated their potential in effectively separating material and lighting components. Notably, the recent works [1, 4, 27, 48] have demonstrated the powerful capabilities of diffusion models, producing impressive results in editing physical properties. However, these approaches typically focus on either lighting [47, 52] or material [38, 40], struggling to provide a comprehensive understanding of the full spectrum of physical properties inherent in images. The gap between high-level semantic editing and low-level physics-aware editing highlights the pressing need for an approach that effectively integrates these properties into a joint framework." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 586, + 360, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 360, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 360, + 597 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Although previous work [38] collects a dataset with basic material variations (e.g., roughness, metallicity, albedo, and transparency), it lacks pixel-aligned maps for fine-grained local editing and pairwise data for disentangled control of material and lighting. Similarly, DiLightNet [47] provides a large dataset with diverse lighting conditions, but it does not support material editing. To address these limitations, we introduce PR-TIPS, a dataset of Physical Rendering-based Text and Image Pair Set, designed to scale up the diversity" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7868" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "content": "and scope of scenes and objects. PR-TIPS provides extensive pairwise data, including pixel-aligned images and text-image pairs with varied instructions, supporting more fine-grained disentangled control over both material and lighting adjustments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 252 + ], + "type": "text", + "content": "Material diversity. We collect 3D assets from a subset of Objaverse [7], selecting 13K models that possess Principled BSDF material properties. To further adjust material properties continuously, we integrate three ShaderNodeMath nodes, applying random values to properties (e.g., Metallic, Roughness, and Transmission Weight). Additionally, we use a ShaderNodeMixRGB node to blend albedo color variations, supporting fine-grained color transitions. Rendering is performed in Blender's Cycles renderer [5], ensuring precise physical interaction with lighting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 253, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 286, + 361 + ], + "type": "text", + "content": "Lighting diversity. To achieve lighting diversity, we utilize over 500 HDRI environment maps from Polyhaven1, applying random rotations to each to simulate diverse lighting conditions. Camera configurations are randomized for field of view, distance, and position around the scene origin, allowing a wide range of perspectives. This setup generates a final dataset, including 300K rendered images to provide substantial variation in materials, lighting, and viewpoints to support robust training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 362, + 286, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 286, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 286, + 493 + ], + "type": "text", + "content": "Text annotation process. We employ a refined text annotation strategy using GPT-4o [11], a Multi-modal Large Language Model (MLLM) with Chain of Thought (CoT) prompting techniques [43]. The model generates detailed text descriptions for both the original and edited images in three structured steps: (i) Identify the scene context (e.g., day or night, indoor or outdoor) and the object type, (ii) Infer the object's material properties, leveraging knowledge of common materials and typical lighting effects, and (iii) Formulate an editing instruction based on the object type, material attributes, and lighting adjustments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "text", + "content": "Dataset statistics. PR-TIPS consists of 300K pairs, with 298K for training and 2K for evaluation. To improve data diversity and promote model generalization, we apply various random augmentations to each sample on the fly during training. Specifically, we use random rotations within the range " + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "inline_equation", + "content": "r \\in [-30^{\\circ}, 30^{\\circ}]" + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "text", + "content": ", scaling factors " + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "inline_equation", + "content": "s \\in [0.9, 1.1]" + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "text", + "content": " and shear transformations " + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "inline_equation", + "content": "sh \\in [-10^{\\circ}, 10^{\\circ}]" + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "text", + "content": ". The translations are applied in both horizontal and vertical directions, with the offsets " + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "inline_equation", + "content": "t_x" + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "inline_equation", + "content": "t_y" + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "text", + "content": " drawn from the range " + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "inline_equation", + "content": "[-0.1, 0.1] \\times" + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "text", + "content": " image dimensions. Additionally, horizontal and vertical flips are applied with a probability of " + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "inline_equation", + "content": "p = 0.5" + }, + { + "bbox": [ + 46, + 494, + 286, + 662 + ], + "type": "text", + "content": ". These augmentations are intended to improve the robustness and generalization of the model by simulating a wide range of transformations during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 662, + 286, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 662, + 286, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 662, + 286, + 698 + ], + "type": "text", + "content": "Summary. PR-TIPS thus provides an extensive and versatile dataset for physics-aware image editing, supporting disentangled control over material and lighting properties." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 388, + 85 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "content": "This section begins with an overview of our framework in Sec. 4.1. Next, we discuss how our model extracts and injects physical properties in Sec. 4.2. Finally, we integrate physical knowledge to image semantic editing models in Sec. 4.3, supporting both low-level detail fidelity and high-level semantic consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 168, + 373, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 168, + 373, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 373, + 179 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 186, + 545, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 222 + ], + "type": "text", + "content": "As illustrated in Fig. 2, we propose PhyS-EdiT, a physics-based framework to improve text-based image editing with physical awareness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 223, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 294 + ], + "type": "text", + "content": "Physical embedding. The physical properties, i.e., lighting conditions and material conditions, are represented as condition maps " + }, + { + "bbox": [ + 304, + 223, + 545, + 294 + ], + "type": "inline_equation", + "content": "C_L" + }, + { + "bbox": [ + 304, + 223, + 545, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 223, + 545, + 294 + ], + "type": "inline_equation", + "content": "C_M" + }, + { + "bbox": [ + 304, + 223, + 545, + 294 + ], + "type": "text", + "content": ", respectively. These structured condition maps offer spatially aligned hints about the physical world, providing our model with robust priors for lighting and material properties." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "text", + "content": "Forward process. In the forward process of our diffusion model [18, 39], the latent variable " + }, + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "text", + "content": " at each timestep " + }, + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "text", + "content": " is represented as a linear combination of the initial latent " + }, + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "text", + "content": " and Gaussian noise " + }, + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 294, + 545, + 342 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 347, + 545, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 347, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 372, + 347, + 545, + 360 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon . \\tag {1}", + "image_path": "0ce97059946f6a5f34240bc6bf9cc1404b5aed4c443e7ce35a230059f67a33a9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "text", + "content": "Denoising networks. Our model incorporates a dual-denoising network design, composed of a low-level editing network " + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{low}}" + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "text", + "content": " and a high-level editing network " + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{high}}" + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "text", + "content": ". The low-level editing network " + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{low}}" + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "text", + "content": ", structured as a U-Net [3], integrates a new physical condition injection module to guide fine-grained adjustments based on material and lighting conditions. Concurrently, the high-level editing network " + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{high}}" + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "text", + "content": " performs broader semantic modifications. Finally, a fusion network " + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{fusion}}" + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "text", + "content": " combines the outputs of " + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{low}}" + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{high}}" + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "text", + "content": ", producing a refined result that balances physical condition editing with global semantic consistency. This dual-network approach enables the editing of precise physical conditions without compromising the ability of the model to make coherent semantic edits." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": "Backward process. The backward process involves iterative denoising, where the dual networks " + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{low}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{high}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": " gradually refine the latent code. The denoising networks " + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{low}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{high}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": " first independently process the noisy latent code " + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": " according to their respective roles. At each timestep " + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": ", the fusion network " + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{fusion}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": " then combines the outputs from " + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{low}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{high}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": " to yield the next-step latent code:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 620, + 545, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 620, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 316, + 620, + 545, + 648 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\hat {z} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} U _ {\\text {f u s i o n}} (\\hat {z} _ {t}, t)\\right) + \\sigma_ {t} \\epsilon , \\tag {2}", + "image_path": "d0c02ff17bba1449c36bb73a8ceb58408f37b52406ab0dff46a03702220b742d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s=1}^t \\alpha_s" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " denotes the accumulated noise factor, and " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " is the standard deviation of the noise added at each step. Through " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " iterations, the model refines the latent code to produce " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ", achieving a balance between low-level physical editing and high-level semantic coherence." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 135, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 702, + 135, + 713 + ], + "type": "text", + "content": "https://polyhaven.com" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7869" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 224 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 224 + ], + "type": "image", + "image_path": "7d7226d5f118a963093c860b476ab5b30f599026d0e1e824612ebfc7d00ecd6f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "text", + "content": "Figure 2. Pipeline of PhyS-EdiT. Given an input image with the desired lighting condition " + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "text", + "content": " and material condition " + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "text", + "content": ", the model first uses the input image and " + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "text", + "content": " to generate lighting condition maps and, concurrently, the input image and " + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "text", + "content": " to generate material maps. These maps condition the low-level editing network " + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{low}}" + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "text", + "content": ". Simultaneously, a semantic text prompt modulates the high-level editing network " + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{high}}" + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "text", + "content": ". The outputs from " + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{low}}" + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{high}}" + }, + { + "bbox": [ + 46, + 227, + 547, + 273 + ], + "type": "text", + "content": " are then combined through the noise fusion module, resulting in the final edited image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 277, + 251, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 251, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 251, + 290 + ], + "type": "text", + "content": "4.2. Material-lighting Joint Representation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 295, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 287, + 403 + ], + "type": "text", + "content": "While previous generative models [3, 34] excel at generating visually appealing images, they often fall short in adhering to real-world physical principles, such as accurate lighting dynamics and material interactions (e.g., realistic reflections). This limitation arises from their inability to interpret or integrate physical properties into the generative process. In this section, we introduce a novel approach that incorporates physical properties into image editing models, enabling physics-aware image editing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 404, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 536 + ], + "type": "text", + "content": "Physical property extraction. We categorize physical properties into two essential components: material and lighting. Previous approaches often focus on controlling either material or lighting independently; however, our goal is to achieve unified control over both. The primary challenge involves representing these properties in a format that diffusion models can interpret effectively. Traditional input formats, such as text prompts for material properties or environment maps for lighting conditions, often fail to maintain spatial alignment, thereby complicating the model's ability to understand and utilize the physical context effectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "text", + "content": "To address this, we leverage pretrained models [48] to extract well-aligned physical property maps from the input image " + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "text", + "content": ". We define the material condition map " + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "inline_equation", + "content": "C_M" + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "inline_equation", + "content": "C_M = [A,R,M,T]" + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "text", + "content": " represent albedo, roughness, metallicity, and transparency maps, respectively. For the lighting condition map " + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "inline_equation", + "content": "C_L" + }, + { + "bbox": [ + 46, + 536, + 288, + 657 + ], + "type": "text", + "content": ", we utilize an off-the-shelf depth estimator [2] to reconstruct a coarse geometry of the scene following [47]. This geometry is then rendered under predefined proxy materials using environment maps. We define the lighting condition as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 669, + 287, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 669, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 124, + 669, + 287, + 683 + ], + "type": "interline_equation", + "content": "C _ {L} = f _ {\\text {l i g h t i n g}} (I, E), \\tag {3}", + "image_path": "59404fd48358cd6b74bc37698d21fc0c7eda2add482963d104b00ff38aca6e17.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{lighting}}" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " represents the function that determines the lighting maps based on the input image " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " and environment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 277, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 315 + ], + "type": "text", + "content": "map " + }, + { + "bbox": [ + 304, + 277, + 545, + 315 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 277, + 545, + 315 + ], + "type": "text", + "content": ". Finally we combine the lighting condition map " + }, + { + "bbox": [ + 304, + 277, + 545, + 315 + ], + "type": "inline_equation", + "content": "C_L" + }, + { + "bbox": [ + 304, + 277, + 545, + 315 + ], + "type": "text", + "content": " with the material condition map " + }, + { + "bbox": [ + 304, + 277, + 545, + 315 + ], + "type": "inline_equation", + "content": "C_M" + }, + { + "bbox": [ + 304, + 277, + 545, + 315 + ], + "type": "text", + "content": " to form the physical condition map " + }, + { + "bbox": [ + 304, + 277, + 545, + 315 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{phys}}" + }, + { + "bbox": [ + 304, + 277, + 545, + 315 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 384, + 324, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 324, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 384, + 324, + 545, + 338 + ], + "type": "interline_equation", + "content": "C _ {\\text {p h y s}} = C _ {M} \\oplus C _ {L}, \\tag {4}", + "image_path": "f60f573c6cf3dd02b01b129209323740a81c7168d1596917782c88893d6797e1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 347, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 347, + 545, + 394 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 347, + 545, + 394 + ], + "type": "text", + "content": " denotes the channel-wise concatenation. By providing the model with material-lighting joint inputs " + }, + { + "bbox": [ + 304, + 347, + 545, + 394 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{phys}}" + }, + { + "bbox": [ + 304, + 347, + 545, + 394 + ], + "type": "text", + "content": ", we bridge the representation gap between different physical properties." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 396, + 546, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 546, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 546, + 516 + ], + "type": "text", + "content": "Physical embedding injection. With available structured representations of the material and lighting conditions, the next step is to inject these conditions into the diffusion model. We achieve this by encoding the input image " + }, + { + "bbox": [ + 304, + 396, + 546, + 516 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 396, + 546, + 516 + ], + "type": "text", + "content": " using an encoder " + }, + { + "bbox": [ + 304, + 396, + 546, + 516 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 304, + 396, + 546, + 516 + ], + "type": "text", + "content": " inspired by the deferred neural relighting architecture [14]. The resulting encoded feature map " + }, + { + "bbox": [ + 304, + 396, + 546, + 516 + ], + "type": "inline_equation", + "content": "F = E(I)" + }, + { + "bbox": [ + 304, + 396, + 546, + 516 + ], + "type": "text", + "content": " captures richer information. To capture interactions between material and lighting, we perform a channelwise multiplication between the encoded features and the combined physical conditions:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 386, + 529, + 545, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 529, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 386, + 529, + 545, + 543 + ], + "type": "interline_equation", + "content": "C _ {\\text {e m b}} = F \\odot C _ {\\text {p h y s}}, \\tag {5}", + "image_path": "1e4c034c4c5fc4df47339eaa223ca0ef04b4287ca0dcebd986a46029987f135c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 548, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 631 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 548, + 545, + 631 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 548, + 545, + 631 + ], + "type": "text", + "content": " denotes element-wise multiplication. This operation allows the model to learn complex relationships between the physical properties and the visual appearance of the image. The resulting physical embedding " + }, + { + "bbox": [ + 304, + 548, + 545, + 631 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{emb}}" + }, + { + "bbox": [ + 304, + 548, + 545, + 631 + ], + "type": "text", + "content": " is then used as a conditioning input for ControlNet [51], which integrates this information into the initial feature spaces of the diffusion model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 349, + 643, + 545, + 655 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 643, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 349, + 643, + 545, + 655 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {t - 1} = \\operatorname {C o n t r o l N e t} \\left(U _ {\\text {l o w}}, \\hat {z} _ {t}, t, C _ {\\text {e m b}}\\right). \\tag {6}", + "image_path": "02f8b3b927ea9bf0759c941d9f0ec2f1aa271eef1478c27f94785f7a211e272d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "This injection process effectively captures both local details and high-level contextual features, culminating in a compact, physics-aware embedding that meets the conditioning requirements of the diffusion model." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7870" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 276, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 276, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 276, + 85 + ], + "type": "text", + "content": "4.3. Bridging Physics to Semantic Image Editing" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 287, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 287, + 233 + ], + "type": "text", + "content": "Effective image editing models should seamlessly integrate both high-level semantic understanding and detailed physical property manipulations. Previous training strategies, such as fine-tuning attention layers [24] and using low-rank adaptations like LoRA [21], often struggle to preserve the original capabilities of the pre-trained model while adapting to new tasks. These methods can inadvertently weaken the model's ability to maintain generative priors, leading to suboptimal performance in scenarios that require fine-grained control over image properties, such as material textures and lighting effects. To address these challenges, we introduce the following strategies:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 235, + 287, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 287, + 377 + ], + "type": "text", + "content": "Preserving generative priors for semantic consistency. To maintain the semantic coherence and editing quality of the pre-trained diffusion model, we initially freeze the weights of the U-Net architecture, " + }, + { + "bbox": [ + 46, + 235, + 287, + 377 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{U - Net}}" + }, + { + "bbox": [ + 46, + 235, + 287, + 377 + ], + "type": "text", + "content": " , for the first 30K iterations. This strategy preserves the foundational generative capabilities of the model during early training. Following this phase, we selectively unfreeze the U-Net's weights, keeping the decoder frozen, and continue training for an additional 15K iterations. This controlled fine-tuning improves the model's capacity to incorporate new conditioning inputs effectively while ensuring high-quality semantic consistency and robust generative performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 379, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 486 + ], + "type": "text", + "content": "Multitask supervision for learning diverse edits. We employ a multitask learning framework that leverages paired datasets for robust training. Our model is trained on an equal mix of our PR-TIPS synthetic dataset and the high-level editing dataset from [3], with a 1:1 ratio. To improve stability, we introduce a training augmentation where, at times, the edited image is identical to the input image, helping the model learn to preserve stability when no edits are required." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 487, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 535 + ], + "type": "text", + "content": "Sequential training. Our training initiates with material editing, focusing on mastering material representations under consistent lighting conditions. We quantify the material editing loss as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 542, + 287, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 542, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 79, + 542, + 287, + 558 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a t e r i a l}} = \\mathbb {E} _ {I, C _ {M}} \\left[ | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, C _ {M}) | | _ {2} ^ {2} \\right], \\tag {7}", + "image_path": "18659955ead5b289bac237c4ab32d4e0d2671d7c86d7d00b3f9855068a9d15bd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 564, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 564, + 287, + 636 + ], + "type": "inline_equation", + "content": "C_M" + }, + { + "bbox": [ + 46, + 564, + 287, + 636 + ], + "type": "text", + "content": " denotes material conditioning and " + }, + { + "bbox": [ + 46, + 564, + 287, + 636 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 564, + 287, + 636 + ], + "type": "text", + "content": " is derived from the forward process described in Equation (1). After completing 15K training steps, we introduce variability in lighting conditions by integrating datasets with diverse lighting scenarios. The loss function then evolves to encompass both material and lighting factors:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 643, + 287, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 643, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 80, + 643, + 287, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p h y s}} = \\mathbb {E} _ {I, C _ {\\text {p h y s}}} \\left[ | | \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, C _ {\\text {p h y s}}\\right) | | _ {2} ^ {2} \\right]. \\tag {8}", + "image_path": "07fda00952a95c0bd6442e6be45994ce5ae12e44efcfec1506b3a1f0064dd243.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Following an additional 15K steps, we enable semantic control by unfreezing " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{fusion}}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ", incorporating semantic training data, and continuing the training over another 15K steps with all conditioning factors activated." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 118, + 544, + 172 + ], + "blocks": [ + { + "bbox": [ + 305, + 72, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 115 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation of material editing methods. Throughout the paper, best performances are highlighted in boldface, and " + }, + { + "bbox": [ + 305, + 72, + 545, + 115 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 305, + 72, + 545, + 115 + ], + "type": "text", + "content": " means high (lower) is better. Evaluation considers only changes in metallicity and roughness properties." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 118, + 544, + 172 + ], + "lines": [ + { + "bbox": [ + 307, + 118, + 544, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 544, + 172 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
IP2P [3]20.850.690.3140.80
Subias et al. [40]†23.000.800.3244.17
Ours26.010.790.1523.38
", + "image_path": "3014d375ff4f2557fcd3d140794d1c8d4cc8e2a6aabb04a484843331e9ca0fb9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 175, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 248 + ], + "type": "text", + "content": "This training strategy ensures the model first consolidates its skills in material editing before it is exposed to varying lighting conditions, and subsequently, semantic variations. This sequential training improves disentangle-ment and control, allowing for precise manipulation of material properties, lighting, and semantics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 252, + 388, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 252, + 388, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 252, + 388, + 266 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 272, + 440, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 272, + 440, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 440, + 285 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": "We initialize the parameters of PhyS-EdiT from IP2P [3] and perform image editing at a resolution of " + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": " pixels. The training is conducted over 16 days on 4 NVIDIA RTX 4090 GPUs with a total batch size of 64. We use the Adam optimizer [23] with a learning rate of " + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 290, + 545, + 411 + ], + "type": "text", + "content": ". During inference, we utilize the DDIM sampling strategy [39] with 50 sampling steps. For real images lacking ground truth environment maps, we estimate lighting conditions using an off-the-shelf lighting estimator [30]. Additional training details are provided in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 417, + 419, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 419, + 429 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 419, + 429 + ], + "type": "text", + "content": "5.2. Evaluation Metrics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 435, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 579 + ], + "type": "text", + "content": "We evaluate the performance of each method using five metrics: (i) FID[17], which measures the distributional distance between generated and real images, reflecting quality and diversity. (ii) SSIM[42], which quantifies the preservation of structural information between edited images and the ground truth. (iii) PSNR[20], which evaluates the visual quality of edited regions, with higher values indicating better fidelity. (iv) CLIP Score[33], which assesses semantic alignment between edited images and text descriptions in the CLIP embedding space. (v) LPIPS [53], which measures perceptual similarity between images, focusing on human visual perception." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 586, + 441, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 441, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 441, + 598 + ], + "type": "text", + "content": "5.3.Quantitative Evaluation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 688 + ], + "type": "text", + "content": "Material editing quality. To evaluate the quality of material editing, we use a test set of 1K samples. We compare our approach with existing feed-forward-based material editing methods [3, 40], using their publicly available implementations. The results are presented in Tab. 1. Our method achieves the best PSNR, LPIPS, and FID scores, as well as the second-best SSIM." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "Lighting editing quality. We evaluate the performance of our lighting editing approach using an additional 1K" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7871" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 85, + 287, + 129 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 276, + 83 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 276, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 276, + 83 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation of lighting editing methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 85, + 287, + 129 + ], + "lines": [ + { + "bbox": [ + 48, + 85, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 85, + 287, + 129 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓FID ↓
DiligNet [47]17.150.650.3951.08
Ours26.960.800.1321.59
", + "image_path": "b793aacd0437dc25d47ea1e88320b81831c9443ce9d0fe4ac4e3204a2624a550.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 148, + 287, + 201 + ], + "blocks": [ + { + "bbox": [ + 55, + 135, + 278, + 145 + ], + "lines": [ + { + "bbox": [ + 55, + 135, + 278, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 278, + 145 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation of semantic editing methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 148, + 287, + 201 + ], + "lines": [ + { + "bbox": [ + 48, + 148, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 148, + 287, + 201 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓CLIP ↑FID ↓
IP2P [3]15.110.570.410.8145.14
SD 3 [10]14.640.500.470.8152.25
Ours15.370.580.390.8547.91
", + "image_path": "7a941770e76da2dc815c194a4be84606bc23fc4e5ebe3503b958876706b50657.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 206, + 286, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 286, + 266 + ], + "type": "text", + "content": "test samples. We compare our results with the existing diffusion-based lighting editing method [47], which allows explicit lighting conditions as input. As shown in Tab. 2, our method achieves the highest scores across all metrics, demonstrating superior capability in lighting editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 286, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 286, + 326 + ], + "type": "text", + "content": "Semantic editing quality. We further evaluate the semantic editing capabilities of our model against state-of-the-art methods, with results shown in Tab. 3. Our method achieves the best PSNR, SSIM, LPIPS, and CLIP scores, and the second best FID." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 334, + 176, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 176, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 176, + 346 + ], + "type": "text", + "content": "5.4. Qualitative Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 352, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 533 + ], + "type": "text", + "content": "We present a visual comparison of our method against baseline approaches for lighting, material, and semantic editing, as shown in Fig. 3. For lighting (red columns), our method effectively adjusts lighting to match the ground truth (GT) while preserving material properties, unlike DiLightNet [47], which is not material-aware and alters the material during lighting adjustments. For material (green columns), our approach produces high-quality images with material properties closely matching the GT while maintaining consistent lighting, whereas Subias et al. [40] fails to modify attributes like albedo and transparency, and IP2P [3] often misinterprets material commands. Finally, for semantic editing (blue columns), our method retains strong high-level editing capabilities, achieving results comparable to or better than IP2P [3] and Stable Diffusion 3 [10]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 539, + 122, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 122, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 122, + 552 + ], + "type": "text", + "content": "5.5. User Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 557, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 286, + 676 + ], + "type": "text", + "content": "To further assess the perceptual quality and consistency of our results, we conduct a user study involving 20 participants. The study is divided into two parts: material editing and lighting editing. In each part, participants are presented with input conditions, original images, and candidate images synthesized by different methods, including our PhyS-EdiT. They are asked to select the most visually pleasing result that best matches the input conditions (preference) and to rate the consistency of the edited images with the original content (consistency)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "As shown in Tab. 4, our method receives the highest preference and consistency scores in both material and lighting editing tasks, indicating a strong user preference and better" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 326, + 96, + 524, + 233 + ], + "blocks": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "Table 4. User study results showing the preference and consistency percentages for each method in material and lighting editing tasks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 96, + 524, + 233 + ], + "lines": [ + { + "bbox": [ + 326, + 96, + 524, + 233 + ], + "spans": [ + { + "bbox": [ + 326, + 96, + 524, + 233 + ], + "type": "table", + "html": "
PropertyMetricMethodScore (%)
MaterialPreferenceOurs68
IP2P [3]18
Subias et al. [40]14
ConsistencyOurs61
IP2P [3]11
Subias et al. [40]28
LightingPreferenceOurs60
DiLightNet [47]40
ConsistencyOurs70
DiLightNet [47]30
", + "image_path": "89d3c8d55cf36fe824ad5810813c134980ea4d628ad9c99ed1e1d8a354bed3b9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 307, + 262, + 543, + 326 + ], + "blocks": [ + { + "bbox": [ + 305, + 238, + 545, + 260 + ], + "lines": [ + { + "bbox": [ + 305, + 238, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 238, + 545, + 260 + ], + "type": "text", + "content": "Table 5. Ablation study results grouped by low-level and high-level editing." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 262, + 543, + 326 + ], + "lines": [ + { + "bbox": [ + 307, + 262, + 543, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 543, + 326 + ], + "type": "table", + "html": "
AblationPhysical EditingSemantic Editing
PSNR ↑LPIPS ↓CLIP ↑LPIPS ↓
Full Model26.390.360.850.39
w/o WF26.000.380.760.35
w/o JT25.800.390.830.40
", + "image_path": "c2c536d41c14680cf267159cf5198954633f30bcee3b7d9d2886d2cd99553cd0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 329, + 476, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 329, + 476, + 341 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 476, + 341 + ], + "type": "text", + "content": "content preservation in our edited images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 346, + 399, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 399, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 399, + 358 + ], + "type": "text", + "content": "5.6. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 387 + ], + "type": "text", + "content": "We perform an ablation study to evaluate the impact of different components of our method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 388, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 545, + 435 + ], + "type": "text", + "content": "Without weight freezing (WF). We train the entire network without freezing the pretrained weights. This leads to overfitting and loss of the model's generative priors, reducing the quality of the semantic edits." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 437, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 484 + ], + "type": "text", + "content": "Without joint training (JT). We train the material and lighting editing components separately. The lack of joint training leads to inconsistent results when both material and lighting need to be edited simultaneously." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": "The quantitative results of the ablation study are presented in Tab. 5, highlighting the performance of different model versions on PSNR, SSIM, LPIPS, and CLIP scores for low-level (physical) and high-level (semantic) editing tasks. The model version without weight freezing (w/o WF) achieves a slightly improved LPIPS score, as not freezing pretrained weights causes the model to forget its original semantic editing capabilities, producing nearly identical outputs regardless of the semantic editing instructions. Consequently, while the unedited image superficially resembles the edited ground truth, it fails to follow the intended instructions, leading to a notably lower CLIP score." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 635, + 378, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 378, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 378, + 647 + ], + "type": "text", + "content": "5.7. Discussion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Support of pixel-level editing. Our model encodes lighting and material properties as pixel-aligned condition maps, enabling precise, localized editing. For instance, in Fig. 5, we applied transparency adjustments to a plate, making it fully transparent (transparency " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "= 1" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ") except for the \"Phys" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7872" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 185, + 585 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 185, + 585 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 185, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 185, + 585 + ], + "type": "image", + "image_path": "39284727e379a286303d1bf0cf171a4868663889a4bfb5ce09a9b429e85c6641.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 586, + 547, + 632 + ], + "lines": [ + { + "bbox": [ + 46, + 586, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 547, + 632 + ], + "type": "text", + "content": "Figure 3. Visual comparison of our method with baseline approaches. For lighting, we compare with DiLightNet [47]; for material, with InstructPix2Pix [3] and Subias et al. [40]; and for semantics, with InstructPix2Pix [3] and Stable Diffusion 3 [10]. Our method effectively disentangles these properties, allowing for the modification of one attribute (e.g., lighting) while preserving others (e.g., material and semantics) unchanged." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 185, + 69, + 365, + 585 + ], + "blocks": [ + { + "bbox": [ + 185, + 69, + 365, + 585 + ], + "lines": [ + { + "bbox": [ + 185, + 69, + 365, + 585 + ], + "spans": [ + { + "bbox": [ + 185, + 69, + 365, + 585 + ], + "type": "image", + "image_path": "0567344f3312d766e2e19494c41789c16fe65cab54b486549424fcacf8ff8513.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 365, + 70, + 545, + 585 + ], + "blocks": [ + { + "bbox": [ + 365, + 70, + 545, + 585 + ], + "lines": [ + { + "bbox": [ + 365, + 70, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 365, + 70, + 545, + 585 + ], + "type": "image", + "image_path": "7c15fd6dff89b729c1bf10034d05eecdb2276c685b1b5e393d7e336f4ac4bb7c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 636, + 288, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 636, + 288, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 636, + 288, + 649 + ], + "type": "text", + "content": "Edit\" text region, which remains opaque (transparency " + }, + { + "bbox": [ + 47, + 636, + 288, + 649 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 47, + 636, + 288, + 649 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "Support of continuous property control. Our model provides smooth, continuous control over lighting and material properties using a unified map format. By feeding in maps with gradually changing values, the model generates outputs that naturally reflect these subtle, progressive ad" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 636, + 547, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 636, + 547, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 636, + 547, + 711 + ], + "type": "text", + "content": "justments. This also underscores the ability of the model to disentangle different physical conditions, allowing for independent and precise modifications. As shown in Fig. 4, row (a) demonstrates a lighting-only change, with the lighting condition visualized as a diffuse sphere. The remaining rows show combined changes in material properties along" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7873" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 545, + 384 + ], + "blocks": [ + { + "bbox": [ + 49, + 100, + 60, + 109 + ], + "lines": [ + { + "bbox": [ + 49, + 100, + 60, + 109 + ], + "spans": [ + { + "bbox": [ + 49, + 100, + 60, + 109 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 72, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 545, + 384 + ], + "type": "image", + "image_path": "6fd319d7566ed0fb6f4cf54287b300f7a585045aca2fade484e816480e8e82f0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 95, + 413, + 153, + 469 + ], + "blocks": [ + { + "bbox": [ + 47, + 387, + 546, + 409 + ], + "lines": [ + { + "bbox": [ + 47, + 387, + 546, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 546, + 409 + ], + "type": "text", + "content": "Figure 4. Demonstration of our model's capability for continuous editing across different attributes: (a) only lighting change, (b) metallicity increase, (c) roughness increase, (d) albedo increase, and (e) transparency increase. Our model produces consistent and seamless transitions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 95, + 413, + 153, + 469 + ], + "lines": [ + { + "bbox": [ + 95, + 413, + 153, + 469 + ], + "spans": [ + { + "bbox": [ + 95, + 413, + 153, + 469 + ], + "type": "image", + "image_path": "6c4497d011d7bf5cc37cbf650f49075448b0bc8159563bd41260240dcf0e07cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 184, + 414, + 239, + 469 + ], + "blocks": [ + { + "bbox": [ + 184, + 414, + 239, + 469 + ], + "lines": [ + { + "bbox": [ + 184, + 414, + 239, + 469 + ], + "spans": [ + { + "bbox": [ + 184, + 414, + 239, + 469 + ], + "type": "image", + "image_path": "93330ce07b3fd9c9b3c20efec17c042358bb04618a7ecc7c628a2792ee8ee73f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 472, + 287, + 505 + ], + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 505 + ], + "type": "text", + "content": "Figure 5. Visualization of our model's capability for pixel-level editing. By applying customized condition maps to a real image, we achieve localized editing in specific properties." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 514, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 538 + ], + "type": "text", + "content": "with lighting: (b) increased metallicity, (c) increased roughness, (d) increased albedo, and (e) increased transparency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 541, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 541, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 541, + 287, + 636 + ], + "type": "text", + "content": "Impact of multi-property control. Learning lighting and material control in conjunction improves the consistency and precision of the model. For example, by accounting for material properties, the model can adjust the lighting while keeping the material consistent. Without this material awareness, the model must rely on implicit assumptions about materials, which can lead to varied and inconsistent results across similar scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 712 + ], + "type": "text", + "content": "Impact of Dataset Scaling. Increasing dataset scale improves performance and generalization to unseen materiallighting combinations by exposing the model to diverse lighting and material interactions. A broader dataset enables realistic editing across materials like metal and glass, while a limited dataset risks overfitting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 415, + 378, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 378, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 378, + 426 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 434, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 553 + ], + "type": "text", + "content": "We introduce PhyS-EdiT, a unified image editing model that enables simultaneous control over material properties, lighting, and high-level semantics. Alongside this, we present PR-TIPS, a large-scale synthetic dataset designed to improve the disentanglement of lighting and material properties. Our approach leverages the advantages of the diffusion model, i.e., disentanglement and controllability, to advance image editing grounded in physical principles. PhyS-EdiT effectively edits both synthetic and real images, supporting both localized and continuous edits." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": "Limitations. While our model performs precise editing and understanding of single objects, it is limited in its ability to generalize to full scene-level editing. Additionally, our model requires two forward passes for processing, which makes it slower compared to single-pass approaches. Future work could enhance scene-level capabilities through richer datasets and optimized single-pass techniques." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 646, + 403, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 403, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 403, + 658 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001 and Doubao Fund. PKU-affiliated authors thank openbayes.com for providing computing resources." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7874" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Dina Bashkirova, Arijit Ray, Rupayan Mallick, Sarah Adel Bargal, Jianming Zhang, Ranjay Krishna, and Kate Saenko. Lasagna: Layered score distillation for disentangled object relighting. arXiv preprint arXiv:2312.00833, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. ZoeDepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 179, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 288, + 224 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to follow image editing instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 268 + ], + "type": "text", + "content": "[4] Xi Chen, Sida Peng, Dongchen Yang, Yuan Liu, Bowen Pan, Chengfei Lv, and Xiaowei Zhou. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arXiv preprint arXiv:2404.11593, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 302 + ], + "type": "text", + "content": "[5] Blender Online Community. *Blender - a 3D modelling and rendering package*. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 346 + ], + "type": "text", + "content": "[6] Guillaume Couairon, Asya Grechka, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Flexit: Towards flexible semantic image translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 401 + ], + "type": "text", + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 446 + ], + "type": "text", + "content": "[8] Katja Doerschner, Huseyin Boyaci, and Laurence T Maloney. Estimating the glossiness transfer function induced by illumination change and testing its transitivity. Journal of Vision, 2010. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 502 + ], + "type": "text", + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 567 + ], + "type": "text", + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, 2024. 1, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 237, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 237, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 237, + 579 + ], + "type": "text", + "content": "[11] OpenAI et al. GPT-4 technical report, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 601 + ], + "type": "text", + "content": "[12] Roland W Fleming. Visual perception of materials and their properties. Vision research, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "type": "text", + "content": "[13] Roland W Fleming, Ron O Dror, and Edward H Adelson. Real-world illumination and the perception of surface reflectance properties. Journal of Vision, 2003. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "type": "text", + "content": "[14] Duan Gao, Guojun Chen, Yue Dong, Pieter Peers, Kun Xu, and Xin Tong. Deferred neural lighting: free-viewpoint relighting from unstructured photographs. ACM Transactions on Graphics, 39(6):1-15, 2020. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[15] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "super-resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 95, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 140 + ], + "type": "text", + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "text", + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 229 + ], + "type": "text", + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 230, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 275 + ], + "type": "text", + "content": "[19] Yuchen Hong, Haofeng Zhong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. L-DiffER: Single image reflection removal with language-based diffusion model. European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 308 + ], + "type": "text", + "content": "[20] Alain Hore and Djemel Ziou. Image quality metrics: PSNR vs. SSIM. In International Conference on Pattern Recognition, 2010. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "text", + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 365, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 388 + ], + "type": "text", + "content": "[22] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 388, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 410 + ], + "type": "text", + "content": "[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "type": "text", + "content": "[24] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 456, + 545, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 489 + ], + "type": "text", + "content": "[25] Gihyun Kwon and Jong Chul Ye. CLIPstyler: Image style transfer with a single text condition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "type": "text", + "content": "[26] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. ManiGAN: Text-guided image manipulation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 578 + ], + "type": "text", + "content": "[27] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 579, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 634 + ], + "type": "text", + "content": "[28] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2I-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Association for the Advancement of Artificial Intelligence, 2024. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[29] Rohit Pandey, Sergio Orts-Escalano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul E Debevec, and Sean Ryan Fanello. Total Relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[30] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7875" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "type": "text", + "content": "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 237 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 303 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 304, + 288, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 349 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 350, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 415 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "type": "text", + "content": "[36] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 471, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 536 + ], + "type": "text", + "content": "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "type": "text", + "content": "[38] Prafull Sharma, Varun Jampani, Yuanzhen Li, Xuhui Jia, Dmitry Lagun, Fredo Durand, Bill Freeman, and Mark Matthews. Alchemist: Parametric control of material properties with diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "text", + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 625, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 658 + ], + "type": "text", + "content": "[40] J Daniel Subias and Manuel Lagunas. In-the-wild material appearance editing using perceptual attributes. In Computer Graphics Forum, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[41] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. SinSR: diffusion-based image superresolution in a single step. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 677 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[42] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[43] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "type": "text", + "content": "[44] Shuchen Weng, Jimeng Sun, Yu Li, Si Li, and Boxin Shi. " + }, + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "type": "inline_equation", + "content": "\\mathrm{CT^2}" + }, + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "type": "text", + "content": ": Colorization transformer via color tokens. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "text", + "content": "[45] Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, Boxin Shi, et al. L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In Advances in Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "type": "text", + "content": "[46] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-grained text to image generation with attentional generative adversarial networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "text", + "content": "[47] Chong Zeng, Yue Dong, Pieter Peers, Youkang Kong, Hongzhi Wu, and Xin Tong. DilightNet: Fine-grained lighting control for diffusion-based image generation. In ACM SIGGRAPH Conference Papers, 2024. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 353, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 407 + ], + "type": "text", + "content": "[48] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. RGBleftrightarrowX: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH Conference Papers, 2024. 1, 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 409, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 409, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 409, + 545, + 463 + ], + "type": "text", + "content": "[49] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. StackGAN: Text to photo-realistic image synthesis with stacked generative adversarial networks. In International Conference on Computer Vision, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 464, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 497 + ], + "type": "text", + "content": "[50] Lisai Zhang, Qingcai Chen, Baotian Hu, and Shuoran Jiang. Text-guided neural image inpainting. In ACM International Conference on Multimedia, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 498, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 531 + ], + "type": "text", + "content": "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In International Conference on Computer Vision, 2023. 1, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "text", + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In International Conference on Learning Representations, 2025. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 587, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 632 + ], + "type": "text", + "content": "[53] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "type": "text", + "content": "[54] Haofeng Zhong, Yuchen Hong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. Language-guided image reflection separation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7876" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/d3a4c2ae-5aa4-4658-a693-f3535a28ce44_content_list.json b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/d3a4c2ae-5aa4-4658-a693-f3535a28ce44_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..934cd1735579d5bec6ea6450fe51ba8d6485fc22 --- /dev/null +++ b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/d3a4c2ae-5aa4-4658-a693-f3535a28ce44_content_list.json @@ -0,0 +1,1524 @@ +[ + { + "type": "text", + "text": "PhyT2V: LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation", + "text_level": 1, + "bbox": [ + 138, + 109, + 856, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiyao Xue, Xiangyu Yin, Boyuan Yang and Wei Gao \nUniversity of Pittsburgh", + "bbox": [ + 287, + 172, + 709, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qix63, eric.yin, by.yang, weigao}@pitt.edu", + "bbox": [ + 307, + 209, + 689, + 224 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8113c8c3ee5a7b963c2ebb52360142362a68b3e5e05009dfad7a0d2b66fac69b.jpg", + "image_caption": [ + "Figure 1. Left: videos generated by the current text-to-video generation model (CogVideoX-5B [50]) cannot adhere to the real-world physical rules (described in brackets following the user prompt). Right: our method PhyT2V, when applied to the same model, better reflects the real-world physical knowledge." + ], + "image_footnote": [], + "bbox": [ + 91, + 250, + 897, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 660, + 326, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-video (T2V) generation has been recently enabled by transformer-based diffusion models, but current T2V models lack capabilities in adhering to the real-world common knowledge and physical rules, due to their limited understanding of physical realism and deficiency in temporal modeling. Existing solutions are either data-driven or require extra model inputs, but cannot be generalizable to out-of-distribution domains. In this paper, we present PhyT2V, a new data-independent T2V technique that expands the current T2V model's capability of video generation to out-of-distribution domains, by enabling chain-of-thought and step-back reasoning in T2V prompting. Our experiments show that PhyT2V improves existing T2V models' adherence to real-world physical rules by $2.3x$ , and achieves $35\\%$ improvement compared to T2V prompt enhancers.", + "bbox": [ + 88, + 678, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 660, + 643, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-video (T2V) generation has recently marked a significant breakthrough of generative AI, with the advent of transformer-based diffusion models such as Sora [3], Pika [17] and CogVideoX [51] that can produce videos conditioned on textual prompts. These models demonstrate astonishing capabilities of generating complex and photorealistic scenes, and could even make it difficult for humans to distinguish between real-world and AI-generated videos, in the aspect of individual video frames' quality [1, 37].", + "bbox": [ + 509, + 676, + 906, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, as shown in Figure 1 - Left, current T2V models still have significant drawbacks in adhering to the real-world common knowledge and physical rules, such as quantity, material, fluid dynamics, gravity, motion, collision and causality, and such limitations fundamentally prevent current T2V models from being used for real-world", + "bbox": [ + 509, + 811, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18826", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "simulation [7, 19, 31]. Enforcement of real-world knowledge and physical rules in T2V generation, however, is challenging because it requires the models' understandings of not only individual objects but also how these objects move and interact with each other. Further, unlike generating static images, T2V generation requires frame-to-frame consistency in object appearance, shape, motion, lighting and other dynamics [11]. Current T2V models often lack such temporal modeling, especially over long sequences [20], and the generated videos often contain flickering, inconsistent motion and object deformations across frames [26].", + "bbox": [ + 89, + 90, + 480, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most of the existing solutions to these challenges are data-driven, by using large multimodal T2V datasets that cover different real-world domains to train the diffusion model [10, 12, 41, 49]. However, these solutions heavily rely on the volume, quality and diversity of datasets [42, 51]. Since real-world common knowledge and physical rules are not explicitly embedded in the T2V generation process, the quality of video generation would largely drop in out-of-distribution domains that are not covered by the training dataset, and the generalizability of T2V models is limited due to the vast diversity of real-world scenario domains. Alternatively, researchers also use the existing 3D engines (e.g., Blender [8], Unity3D [36] and Unreal [16]) or mathematical models of edge and depth maps [26-28] to inject real-world physical knowledge into the T2V model, but these approaches are limited to fixed physical categories and patterns such as predefined objects and movements [26, 49], similarly lacking generalizability.", + "bbox": [ + 91, + 257, + 482, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve generalizable enforcement of physics-grounded T2V generation, we propose a fundamentally different approach: instead of expanding the training dataset or further complicating the T2V model architecture, we aim to expand the current T2V model's capability of video generation from in-distribution to out-of-distribution domains, by embedding real-world knowledge and physical rules into the text prompts with sufficient and appropriate contexts. To avoid ambiguous and unexplainable prompt engineering [9, 32, 33], our basic idea is to enable chain-of-thought (CoT) and step-back reasoning in T2V generation prompting, to ensure that T2V models follow correct physical dynamics and inter-frame consistency by applying step-by-step guidance and iterative refinement.", + "bbox": [ + 89, + 527, + 482, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on this idea, this paper presents Physcial-grounded Text-to-Video (PhyT2V), a new T2V technique that harnesses the natural language reasoning capabilities of well-trained LLMs (e.g., ChatGPT-4o), to facilitate CoT and step-back reasoning as described above. As shown in Figure 2, such reasoning is iteratively conducted in PhyT2V, and each iteration autonomously refines both the T2V prompt and generated video in three steps. In Step 1, the LLM analyzes the T2V prompt to extract objects to be shown and physical rules to follow in the video via in-context learning. In Step 2, we first use a video captioning model to translate", + "bbox": [ + 89, + 738, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/36aad3ee25808e3108664a0a95864a636edf3867f4adeedb7f860368e96007ac.jpg", + "image_caption": [ + "Figure 2. One iteration of video and prompt self-refinement in PhyT2V. Such self-refinement will be iteratively conducted in multiple rounds until the quality of generated video is satisfactory." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the video's semantic contents into texts according to the list of objects obtained in Step 1, and then use the LLM to evaluate the mismatch between the video caption and current T2V prompt via CoT reasoning. In Step 3, the LLM refines the current T2V prompt, by incorporating the physical rules summarized in Step 1 and resolving the mismatch derived in Step 2, through step-back prompting. The refined T2V prompt is then used by the T2V model again for video generation, starting a new round of refinement. Such iterative refinement stops when the quality of generated video is satisfactory or the improvement of video quality converges.", + "bbox": [ + 511, + 291, + 906, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For physical-grounded video generation performance, we further evaluated PhyT2V by applying it onto multiple SOTA T2V models, by using ChatGPT4 o1-preview [18] for LLM reasoning and Tarsier [39] as the video captioning model. We used two major T2V prompt datasets that cover 7 different real-world domains, and compared PhyT2V with the most competitive baselines of prompt enhancers. Our main findings are as follows:", + "bbox": [ + 511, + 455, + 908, + 571 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- PhyT2V is highly effective. Without involving any model retraining efforts on any auxiliary model inputs, PhyT2V can improve the adherence of the existing T2V models' generated videos to real-world physical rules by up to $2.3\\mathrm{x}$ , by only refining the text prompts to the T2V model.", + "- PhyT2V is high generic. It can result in significant improvement of video quality in a large diversity of real-world domains, covering solid, liquid, mechanics, optics, thermal, etc. It is fully data independent, and its prompting templates can be applied to any existing T2V models with different architectures and input formats.", + "- Based on LLM-guided reasoning and self-refinement, PhyT2V is fully automated and involve the minimum amount of engineering and manual efforts." + ], + "bbox": [ + 511, + 575, + 906, + 781 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work and Motivation", + "text_level": 1, + "bbox": [ + 511, + 794, + 787, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. T2V Generation Models", + "text_level": 1, + "bbox": [ + 511, + 819, + 735, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early T2V techniques generate video frames from text-to-image model outputs with temporal extensions [35], but cannot maintain temporal consistency and coherence over time, often producing visually appealing but temporally dis", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18827", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "connected outputs. Diffusion Transformers (DiT) [30] improved such consistency with a transformer backbone capable of capturing more complex temporal dynamics and relationships across frames through attention mechanism and long-range dependency modeling [42, 51]. Based on the DiT architecture, recent T2V models, such as OpenSora [53] and VideoCrafter [4], demonstrated that T2V generation can be further improved by in-context learning when provided with sufficient contextual information [44].", + "bbox": [ + 89, + 90, + 483, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/af4d0ad4dec137b51df26c071e6ca57bdf56b2c75891f171a6ef9623aad6a3dc.jpg", + "image_caption": [ + "Figure 3. Examples of videos generated from in-distribution and out-of-distribution prompts, using the CogVideoX-5B model" + ], + "image_footnote": [], + "bbox": [ + 93, + 229, + 480, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, as shown in Figure 3, although these T2V models demonstrate strong capabilities in video generation when dealing with prompts aligned with the distributions found in the training data, they encounter significant challenges with out-of-distribution prompts that are not covered by training data1. In such cases, the outputs often contain physical illusions or artifacts, reflecting the model's limitations in generating realistic and coherent video contents under unfamiliar conditions. Such limitations can be addressed by enlarging the training datasets, improving T2V model architectures or developing new mechanisms for adaptation and error correction [43, 45], but these approaches are all prompt-specific and lack generalizability.", + "bbox": [ + 89, + 431, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f053aeeb42b356aeecd1aa01478efece326b8b30f3502fb76732588c06d29b49.jpg", + "image_caption": [ + "Figure 4. A video generated by enhancing the out-of-distribution prompt \"Whisking egg into milk for scramble\" in Figure 3" + ], + "image_footnote": [], + "bbox": [ + 93, + 640, + 482, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, as shown in Figure 4, recent research has demonstrated that the quality of video generation with an out-of-distribution prompt can be improved by refining the prompt with sufficient and appropriate details [11, 51]. These findings motivate our design of PhyT2V that embeds contexts of real-world knowledge and physical rules into", + "bbox": [ + 89, + 761, + 483, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "T2V prompts, to guide the T2V process for better physical accuracy and temporal alignment. The existing works, however, could still fail when tackling more intricate scenarios such as multi-object interactions, because the T2V model lacks an efficient feedback mechanism to learn how the generated video deviates from the real-world knowledge and physical rules. Researchers suggest to provide such feedback with extra input modalities to T2V models such as sampled video frames, depth map or scribble images [44, 52], but incur significant amounts of extra computing overhead and cannot be generalizable. Instead, in our design of PhyT2V, we aim to fully automate the feedback with only text prompts, and enable iterative feedback for the optimum video quality.", + "bbox": [ + 511, + 90, + 906, + 299 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d97ec0eb02e0075ebca8bde89d0348e0d0f9fe2cfc5837ed20b412d5b2b9978c.jpg", + "image_caption": [ + "Figure 5. Examples of CoT and step-back reasoning" + ], + "image_footnote": [], + "bbox": [ + 524, + 305, + 890, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Using LLM in T2V Generation", + "text_level": 1, + "bbox": [ + 511, + 542, + 790, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLMs with strong capabilities in natural language processing (NLP) have been a natural choice for prompt refinement in text-to-image and text-to-video generation, and existing work utilized LLMs to interpret text prompts and orchestrate the initial layout configurations [13, 14, 23-25, 46, 48, 54]. However, since current LLMs lack inherent understandings of the real-world physical laws, using LLMs with simple instructions usually result in videos that appear visually coherent but lack accurate physical realism, particularly when generating scenes with complex object interactions. Furthermore, these approaches frequently rely on static prompts or simple iterative refinements based on bounding box and segmentation map, which may capture basic visual attributes but fail to adapt to nuanced changes that require continuous physical modeling and adjustment.", + "bbox": [ + 511, + 560, + 906, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An effective approach to addressing these limitations and providing effective feedback for prompt refinement is to explicitly trigger in-context learning and reasoning in LLM. For example, as shown in Figure 5, CoT reasoning deconstructs complex prompts into stepwise logical tasks, and hence ensures a precise scheduling path to align generated content with the input prompt. However, CoT reasoning, in some cases, could make errors in some intermediate steps,", + "bbox": [ + 511, + 782, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1In Figure 3, the in-distribution prompts are picked from the ones listed in [50], and the out-of-distribution prompts are our crafted ones for similar scenarios as the in-distribution prompts.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18828", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 1: Identifying the Physical Rule and Main Object", + "text_level": 1, + "bbox": [ + 155, + 89, + 464, + 102 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/89f6e74f50e27e472e40e451539de5cb9d058b750440cf8fd26e92e8d4eaece4.jpg", + "image_caption": [ + "Figure 6. Our design of PhyT2V, illustrated by one round of video refinement consisting of three steps. Texts in brown are inputs from previous step. Texts in red are outputs to the next step; Texts in purple are prompts to trigger LLM reasoning" + ], + "image_footnote": [], + "bbox": [ + 109, + 103, + 503, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 3: Generating the Refined Prompt", + "text_level": 1, + "bbox": [ + 620, + 89, + 844, + 101 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task instruction", + "text_level": 1, + "bbox": [ + 593, + 102, + 687, + 112 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ed1642fa16d59bb50d36be50fcef2135f54234741e9fc8384f12f2ebf58833b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 113, + 883, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and step-back prompting can address this limitation by further deriving the step-back question at a higher level of abstraction and hence avoiding confusions and vagueness. In our design of PhyT2V, we will utilize such LLM reasoning to analyze the inconsistency of the generated video to real-world common knowledge and physical rules, and use the reasoning outcome as feedback for T2V prompt refinement.", + "bbox": [ + 88, + 470, + 482, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Chain-of-Thought (CoT) method is suitable for single data mode processing because it emphasizes linear decomposition and step-by-step reasoning, and it is especially effective for data processing flows that do not require cross-modal synchronization or interaction. However, multimodal data processing involves the fusion of data from different modalities and complex synchronization requirements, which cannot be completed through simple linear decomposition and requires frequent cross-modal information interaction and parallel processing, which exceeds the linear reasoning ability of the CoT method, resulting in limited performance in multimodal tasks. This is the reason why it is hard to directly apply CoT reasoning in T2V process itself as it requires multimodal alignment between the next and video modality. This also motivates us to adopt video captioning and use the video caption in the reasoning process, so that we can conduct CoT and step-back reasoning only in the text domain.", + "bbox": [ + 89, + 575, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 849, + 181, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present details of our PhyT2V design. In principal, PhyT2V's refinement of T2V generation is an", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "iterative process consisting of multiple rounds. In each round, as shown in Figure 6, the primary objective of our PhyT2V design is to guide a well-trained LLM (e.g., ChatGPT-4o) to generate a refined prompt that enables the pre-trained T2V model to generate videos that better match the given user prompt and real-world physical rules, and the refined prompt will be iteratively used as the new user prompt in the next round of refinement.", + "bbox": [ + 511, + 470, + 906, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each round of refinement is structured around decomposing the complex refinement problem into a series of simpler subproblems, more specifically, two parallel subproblems and one final subproblem. The two parallel subproblems are: Step 1) identifying the relevant physical rules that the generated video should follow based on the user prompt, and Step 2) identifying semantic mismatches between the user prompt and the generated video. Based on the knowledge about physical rules and semantic mismatches, the final subproblem (Step 3) generates the refined prompt to better adhere to the physical rules and resolve the mismatches.", + "bbox": [ + 511, + 589, + 908, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure proper identification in the parallel subproblems and prompt generation in the final subproblem, the core of PhyT2V design is two types of LLM reasoning processes within the prompt enhancement loop: the local CoT reasoning for individual subproblems and global step-back reasoning for the overall prompt refinement problem.", + "bbox": [ + 511, + 752, + 908, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local CoT reasoning is executed within the prompt for each subproblem, to prompt the LLM to generate a detailed reasoning chain in its latent embedding space [38]. Addressing the parallel subproblems facilitates LLM with a", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18829", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "more concentrated attention on prerequisites of prompt refinement, enabling a deeper comprehension of the physical laws that govern the video content as well as the identification of discrepancies between the generated video and the user prompt. The outcomes derived from these parallel subproblems reflect the language model's abstraction in stepback reasoning on the overarching prompt refinement.", + "bbox": [ + 89, + 90, + 480, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Global step-back reasoning: To integrate various subproblems into a coherent framework for prompt and video refinement, one intuitive approach involves employing CoT reasoning across these subproblems, allowing the LLM to engage in self-questioning. However, this method may lead to the risk of traversing incorrect reasoning pathways. Instead, we apply global step-back reasoning across subproblems, by using a self-augmented prompt to incorporate the LLM-generated responses to high-level questions about physical rules and semantic mismatches in earlier parallel problems, when generating the refined prompt in the final subproblem. In this way, we can improve the correctness of intermediate reasoning steps in CoT reasoning, and enable consistent improvement across steps in reasoning.", + "bbox": [ + 89, + 195, + 482, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both reasoning processes are facilitated through appropriate task instruction prompting tailored to different subproblems. In general, our prompting procedure follows the prompt modeling in [34], which divides task instructions into several components. More details about these components in our design of PhyT2V are elaborated as follows.", + "bbox": [ + 89, + 402, + 480, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared to the previous prompt enhancing methods, PhyT2V's key contribution is to analyze the semantic mismatch between currently generated video and the prompt, as well as refinements based on such mismatch. Previous methods can be formulated as $p' = f_{\\text{enhance}}(p, \\theta)$ , where $p$ and $p'$ are the original and enhanced prompts, $f_{\\text{enhance}}$ is the prompt enhancer, and $\\theta$ represents parameters or rules guiding the enhancement. In contrast, PhyT2V further involves the additional information about the T2V process, i.e., $p' = f_{\\text{enhance}}(p, f_{\\text{mismatch}}(C(V(p)), p), f_{\\text{phy}}(p), \\theta)$ , where $f_{\\text{phy}}(p)$ analyzes the physical rules to be followed given $p$ , $V(p)$ is the currently generated video given prompt $p$ , $C$ is the video captioning model and $f_{\\text{mismatch}}$ finds the semantic mismatch between $C(V(p))$ and $p$ . The key advantages are: (1) Semantic awareness: the refinement process explicitly incorporates the semantic mismatch to enable targeted T2V improvements; (2) Physical-world knowledge integration: physical rules derived from $p$ enable guided enhancement; (3) Guided reasoning: unlike prior methods that rely solely on templates or embeddings, PhyT2V dynamically adapts prompt refinement to the semantic mismatch.", + "bbox": [ + 89, + 489, + 482, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Prompting in Parallel Subproblems for Local CoT Reasoning", + "text_level": 1, + "bbox": [ + 89, + 808, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In both Step 1 and Step 2, the first part of prompt is a task instruction prompt $[I]$ to instruct the LLM to understand the task in the subproblem. $[I]$ is designed with multiple components, each of which corresponds to different functions.", + "bbox": [ + 89, + 840, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the first sentence, it provides general guidance to relate the current subproblem to the entire refinement problem, to better condition the subproblem answer. Afterwards, it will include detailed descriptions of the task: identifying the physical rule and main object in Step 1, and identifying the semantic mismatch between the user prompt and caption of the generated video (generated by the video captioning model) in Step 2. It will also contain the requirements about the expected information in LLM's output. For example, in Step 1, the LLM's output about the physical rule should be in a descriptive way without giving formulas.", + "bbox": [ + 511, + 90, + 903, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides, to ensure proper CoT reasoning, we follow the existing work [22, 40] and provide in-context examples $[E]$ about tasks. To facilitate LLM's in-context learning [5, 6], $[E]$ is given in the format of QA pairs. That is, instead of fine-tuning a separate LLM checkpoint for each new task, we prompt the LLM with a few input-output exemplars, to demonstrate the task and condition the task's input-output format to the LLM, to guide the LLM's reasoning process.", + "bbox": [ + 511, + 255, + 903, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the final part of the prompt, denoted as $[T]$ , is the information of the current instance of the task, usually with the current user prompt $(P_{i})$ being embedded. As a common practice of CoT reasoning, it also contains the hand-crafted trigger phrase $(t)$ , \"Let's think step by step\", to activate the local CoT reasoning in LLM.", + "bbox": [ + 511, + 373, + 903, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Prompting in the Final Subproblem for Global Step-Back Reasoning", + "text_level": 1, + "bbox": [ + 511, + 473, + 903, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the final subproblem, we enforce global step-back reasoning, by using the outputs of the two parallel subproblems above, i.e., knowledge about the physical rules and the prompt-video mismatch, as the high-level concepts and facts. Grounded on such high-level abstractions, we can make sure to improve the LLM's ability in following the correct reasoning path of generating the refined prompt.", + "bbox": [ + 511, + 506, + 903, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Being similar to the prompts used in the two parallel subproblems above, the prompt structure in the final subproblem also contains $[I]$ , $[E]$ and $[T]$ . Furthermore, to ensure the correct reasoning path, we also provide quantitative feedback to the LLM about the effectiveness of previous round's prompt refinement. Such effectiveness could be measured by the existing T2V evaluators, which judge the semantic alignment and quality of physical common sense of the currently generated video2. For example, the VideoCon-Physics evaluator [2] gives a score $([S])$ between 0 and 1. If $[S]$ is $< 0.5$ , it indicates that the refined prompt produced in the previous round is ineffective, hence guiding the LLM to take another alternative reasoning path.", + "bbox": [ + 511, + 611, + 903, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the prompt in the final subproblem is rich with reasoning and inherently very long-tailed, we removed the trigger prompt $[t]$ , to prevent incorporating the information in the final answer unrelated to the user's initial input prompt.", + "bbox": [ + 511, + 804, + 903, + 864 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2This video is generated using the prompt refined in the previous round, and is also used to generate the video caption as the input in Step 2.", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18830", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. The Stopping Condition", + "text_level": 1, + "bbox": [ + 89, + 90, + 310, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The process of iterative refinement normally continues until the quality of the generated video is satisfactory, measured by the T2V evaluator as described above. Furthermore, the current T2V models naturally have limitations in generating some complicated or subtle scenes. In these cases, it would be difficult, even for PhyT2V, to reach physical realism after multiple rounds of iterations, and PhyT2V's refinement would stop when the iterations converge, i.e., the improvement of video quality becomes little over rounds.", + "bbox": [ + 89, + 107, + 483, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 250, + 222, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models & Datasets: We applied PhyT2V on several DiT-based open-source T2V models, as listed below, and evaluated how PhyT2V improves the generated videos' adherence to real-world knowledge and physical rules. We use ChatGPT4 o1-preview [18] as the LLM for reasoning, and Tarsier [39] as the video captioning model. All generated videos last 6 seconds with 10 FPS and resolution of $720 \\times 480$ . Details of evaluation setup are in Appendix A.", + "bbox": [ + 89, + 268, + 483, + 388 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CogVideoX [51]: It generates 10-second videos from text prompts, with 16 FPS and $768 \\times 1360$ resolution. It offers two model variants with 2B and 5B parameters.", + "- OpenSora 1.2 [53]: As an alternative to OpenAI's Sora [3], it contains 1.1B parameters and produces videos with 16 seconds, 720p resolution and different aspect ratios.", + "- VideoCrafter [4]: With 1.8B parameters, it can generate both images and videos from text prompts, with $576 \\times 1024$ resolution and a focus on video dynamics." + ], + "bbox": [ + 91, + 390, + 482, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since we target enhancing the T2V models' capability of generating physics-grounded video contents, we use the following two prompt benchmarks that emphasize physical laws and adherence as the text prompts for T2V:", + "bbox": [ + 89, + 525, + 483, + 585 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- VideoPhy [2] is designed to assess whether the generated videos follow physical common sense for real-world activities. It consists 688 human-verified captions that describe interactions between various types of real-world objects, including solid and fluid.", + "- PhyGenBench [29], similarly, allows evaluating the correctness of following physical common sense in T2V generation. It comprises 160 carefully crafted prompts spanning four physical domains, namely mechanics, optics, thermal and material properties. Since the domain of material properties has been covered by VideoPhy, we use the first three domains listed above." + ], + "bbox": [ + 91, + 587, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metric: We use VideoCon-Physics evaluator provided with VideoPhy [2], to measure how the generated video adheres to physical common sense (PC) and achieves semantic adherence (SA). The PC metric evaluates whether the depicted actions and object's state follow the real-world physics laws. The SA metric measures if the actions, events, entities and their interactions specified in the prompt are present. Both metrics yield binary outputs: 1 indicates adherence and 0 indicates otherwise. On each T2V model and", + "bbox": [ + 89, + 767, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/94f6f5c704b6081d15b27cbb35f1aa877147ed07138efc9bc37801426b150ce3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Round1234
CogVideoX-5B [51]PC0.260.320.390.42
SA0.480.520.560.59
CogVideoX-2B [51]PC0.130.190.270.29
SA0.220.120.400.42
OpenSora [53]PC0.170.290.270.31
SA0.290.380.440.47
VideoCrafter [4]PC0.150.250.290.33
SA0.240.380.440.49
", + "bbox": [ + 535, + 79, + 880, + 241 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/27a8807dcd57c1a9e6164c24fa7300864779fdeea61405c6802a5f985a564069.jpg", + "table_caption": [ + "Table 1. The quality of videos generated by different T2V models using the VideoPhy prompt dataset, over multiple rounds of iterative refinement in PhyT2V" + ], + "table_footnote": [], + "table_body": "
Round1234
CogVideoX-5B [51]PC0.280.320.380.42
SA0.220.350.360.38
CogVideoX-2B [51]PC0.160.190.240.27
SA0.150.290.330.35
OpenSora [53]PC0.210.250.240.26
SA0.230.280.290.30
VideoCrafter [4]PC0.200.240.320.36
SA0.270.330.370.42
", + "bbox": [ + 535, + 286, + 880, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The quality of videos generated by different T2V models using the PhyGenBench prompt dataset, over multiple rounds of iterative refinement in PhyT2V", + "bbox": [ + 511, + 449, + 906, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset, the binary outputs from all prompts are averaged.", + "bbox": [ + 511, + 501, + 890, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we also evaluated PhyT2V using the widely used VBenchmark metrics and benchmarks [15], which allow comprehensive evaluations of the generated video in multiple aspects, including video quality, video-condition consistency, prompt following and human preference alignment.", + "bbox": [ + 511, + 517, + 903, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines: For fair comparison, we only use the existing T2V prompt enhancers as baselines, and other existing work with extra inputs to T2V models [7, 19, 26, 27, 31] are not applicable. We involve two prompt enhancers: 1) Directly using the existing LLM, particularly ChatGPT4, as the prompt enhancer [28, 47]; 2) Promptist [21], which uses reinforcement learning to automatically refine and enhance prompts in the model-preferred way.", + "bbox": [ + 511, + 592, + 905, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Improvement of the Generated Video Quality", + "text_level": 1, + "bbox": [ + 511, + 718, + 897, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1 and 2, when PhyT2V is applied to different T2V models, it can significantly improve the generated video's adherence to both the text prompt itself and the real-world physical rules, compared to the videos generated by vanilla T2V models (i.e., in Round 1 of PhyT2V's refinement). In particular, such improvement is the most significant on the CogVideoX-2B model, where PC improvement can be up to $2.2\\mathrm{x}$ and SA improvement can be up to $2.3\\mathrm{x}$ . On all the other models, PhyT2V can also reach noticeable improvement, ranging from $1.3\\mathrm{x}$ to $1.9\\mathrm{x}$ .", + "bbox": [ + 511, + 736, + 905, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Meanwhile, results in Table 1 and 2 showed that", + "bbox": [ + 531, + 885, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18831", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c60a4812cb07d2d79d7d2fd0e033a4b99cd033db6f25cba2cff25ed0ed1062f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CogVideoX-5BOpenSora
ChatGPT 4 [28]PC0.330.21
SA0.410.32
Promptist [21]PC0.250.19
SA0.390.33
", + "bbox": [ + 129, + 84, + 442, + 175 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d7b4952711ed70bf6a9a17eaad88391608f086ec8f01f0c6c569b7e781a73b63.jpg", + "table_caption": [ + "Table 3. The quality of videos generated by enhancing the prompts in the VideoPhy dataset using different prompt enhancers" + ], + "table_footnote": [], + "table_body": "
CogVideoX-5BOpenSora
ChatGPT 4 [28]PC0.270.20
SA0.230.23
Promptist [21]PC0.320.19
SA0.240.21
", + "bbox": [ + 127, + 210, + 442, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. The quality of videos generated by enhancing the prompts in the PhyGenBench dataset using different prompt enhancers", + "bbox": [ + 89, + 303, + 482, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PhyT2V's process of iterative refinement converge quickly and only takes few rounds. Most improvement of video quality happens in the first two rounds, and little improvement can be observed in the fourth round. Hence, in practice, we believe that 3-4 iterative rounds would be sufficient.", + "bbox": [ + 89, + 339, + 482, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, as shown in Table 3 and 4, PhyT2V also largely outperforms the existing prompt enhancers by at least $35\\%$ , when being applied to CogVideoX-5B and OpenSora models. In particular, ChatGPT 4, when being used as the prompt enhancer, delivers better performance than Promptist due to its stronger language processing capabilities, but still cannot ensure physics-grounded T2V, due to the lack of explicit reasoning on text-to-video alignment.", + "bbox": [ + 89, + 414, + 482, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our evaluation results on the VBenchmark metrics are shown in Figure 7, where numbers in Round 1 are the T2V model's original performance in current VBenchmark leaderboard, and iterative prompt refinements by PhyT2V in Round 2 & 3 noticeably improve the performance in many dimensions. In particular, large improvements are noted in most dimensions of Video-Condition Consistency, showing that PhyT2V improves T2V model's adherence to prompts and real-world physical rules underlying the prompts.", + "bbox": [ + 89, + 532, + 482, + 667 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37805667bedfd49b946dedca9679a1e972d7c9248a9f8d30d4ca8f25f556908a.jpg", + "image_caption": [ + "Figure 7. PhyT2V VBenchmark evaluation results with CogVideoX-5B (left) and OpenSora (right)" + ], + "image_footnote": [], + "bbox": [ + 94, + 670, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Different Domains of Physical Rules", + "text_level": 1, + "bbox": [ + 89, + 838, + 405, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conducted in-depth analysis on PhyT2V's performance on improving the generated video's quality in different domains of real-world physical rules, using the", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/35c8323bd586c74d247bb4e8ff65310a069fb50c8fc7dfa38dac39989d1436d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 79, + 901, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/623e59ce559975c0b379d113d4a4e4feea6eebf080f5960abc0157c0b52d4e02.jpg", + "image_caption": [ + "Figure 8. Examples of videos generated using different categories of prompts in the PhyGenBench dataset", + "Figure 9. Examples of videos generated using different categories of prompts in the VideoPhy dataset" + ], + "image_footnote": [], + "bbox": [ + 516, + 303, + 903, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CogVideoX-5B as the T2V model and ChatGPT 4 as the prompt enhancer. As shown in Table 5 and 6, PhyT2V achieves large improvements in most domains of physical rules. Especially in domains where physical interaction between objects are more subtle and difficult to be precisely captured, such as interaction with fluids and thermal-related scene changes, such improvements will be even higher.", + "bbox": [ + 511, + 551, + 905, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These improvements are also exemplified with sample videos and their related input prompts in Figure 9 and Figure 8. With LLM reasoning and iterative refinement, PhyT2V can largely enhance the T2V model's capability when encountering out-of-distribution prompts, by providing correct and sufficient contexts to ensure that the T2V model's video generation correctly captures the key objects and interaction between objects. For example, when the prompt of \"juice dropping from a bottle onto the counter\", PhyT2V correctly depicts the juice's slow diffusion on the counter. More examples can be found in Appendix B.", + "bbox": [ + 511, + 655, + 906, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 824, + 666, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct an ablation study to evaluate the necessity of both the physical rule reasoning (Step 1) and the mismatch reasoning (Step 2) within our PhyT2V workflow, by removing one of these steps from the refinement process to assess", + "bbox": [ + 511, + 840, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18832", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/60b97f80f1fc7a55442e997e78ba95b9ae7e76a85c458361b999053f8cbbe258.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CogVideoX-5BCogVideoX-2BOpenSoraVideoCrafter
Round1234123412341234
Solid-SolidPC0.210.280.340.320.090.130.140.220.120.270.290.300.190.220.270.28
SA0.240.480.490.470.180.250.360.330.160.340.370.350.240.400.450.47
Solid-FluidPC0.220.270.280.300.110.180.280.270.170.210.240.250.180.240.250.26
SA0.390.540.600.610.290.430.440.430.160.400.410.360.340.430.480.52
Fluid-FluidPC0.570.590.630.620.340.380.350.360.150.320.290.310.330.410.530.51
SA0.410.570.590.670.270.420.390.440.310.440.450.460.320.420.490.51
", + "bbox": [ + 102, + 88, + 890, + 233 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7f346fd04fa4c4570d37fbe685461e3a86ad5b8fa881e2150682e00d78a1f52e.jpg", + "table_caption": [ + "Table 5. The improvement of generated video quality in different categories of physical rules in the VideoPhy prompt dataset" + ], + "table_footnote": [], + "table_body": "
CogVideoX-5BCogVideoX-2BOpenSoraVideoCrafter
Round1234123412341234
MechanicsPC0.190.250.340.350.120.160.180.240.110.130.170.220.140.230.290.28
SA0.210.280.290.320.110.180.190.220.190.210.270.320.200.240.280.35
OpticsPC0.220.350.410.390.220.250.290.280.240.260.250.250.220.210.270.32
SA0.270.420.390.440.230.340.370.390.260.310.290.300.220.280.350.39
ThermalPC0.330.350.350.350.130.150.150.140.270.300.310.330.250.280.260.28
SA0.220.360.430.450.120.160.240.270.230.250.370.360.250.370.410.43
", + "bbox": [ + 104, + 253, + 888, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. The improvement of generated video quality in different categories of physical rules in the PhyGenBench prompt dataset", + "bbox": [ + 114, + 400, + 880, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "its impact on physical-grounded video generation.", + "bbox": [ + 89, + 422, + 421, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Physical rule reasoning (Step 1). As shown in Figure 10, the Step 1 of physical rule reasoning significantly enhances the T2V process by providing a more detailed and coherent description of the principal object's physical status, such as motion, states and deformation (red texts in Figure 10), all grounded in relevant physical laws. By anchoring the prompt in established physical rules, this step also helps avoid unnecessary texts (brown texts in Figure 10) and vague physical rule descriptions (purple texts in Figure 10), hence achieving a higher PC score.", + "bbox": [ + 89, + 436, + 483, + 585 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/184427b9654900e900410a525a756c99bc3a84987a626ce9ba2da482680bac76.jpg", + "image_caption": [ + "Figure 10. Ablation study on Step 1 of physical rule reasoning" + ], + "image_footnote": [], + "bbox": [ + 99, + 588, + 478, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mismatch reasoning (Step 2). The Step 2 of mismatch reasoning addresses details that may have been overlooked in the previous iteration of the generated video as shown in Figure 11. This step plays a critical role in the iterative refinement process by identifying and correcting discrepancies between expected and observed outputs. By enhancing the model's focus on the principal object, the mismatch reasoning step reduces the likelihood of losing attention to", + "bbox": [ + 89, + 781, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d346e6757ccf1d5eab11a5d1706259a20cc8e7a9c29c83a128cc0ec8588cd184.jpg", + "image_caption": [ + "Figure 11. Ablation study on Step 2 of mismatch reasoning" + ], + "image_footnote": [], + "bbox": [ + 513, + 419, + 903, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "important features (brown and purple texts in Figure 11), improving the fidelity and relevance of generated video content (red texts in Figure 11) towards a higher SA score.", + "bbox": [ + 511, + 609, + 903, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, our study shows that both reasoning steps are integral to PhyT2V, contributing to a more robust and semantically-aligned generation of refined prompts in Step 3. More detailed ablation studies are in Appendix C.", + "bbox": [ + 511, + 654, + 905, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 719, + 633, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present PhyT2V, a novel data-independent T2V generation framework designed to enhance the generalization capability of existing T2V models to out-of-distribution domains. By incorporating CoT reasoning and step-back prompting, PhyT2V systematically refines T2V prompts to ensure adherence to real-world physical principles without necessitating additional model retraining or reliance on additional conditions. Evaluation results indicate that PhyT2V achieves a $2.3\\mathrm{x}$ enhancement in physical realism compared to baseline T2V models and outperforms state-of-the-art T2V prompt enhancers by $35\\%$ .", + "bbox": [ + 511, + 739, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18833", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank the reviewers and the area chair for their insightful comments and feedback. This work was supported in part by National Science Foundation (NSF) under grant number IIS-2205360, CCF-2217003, CCF-2215042, and National Institutes of Health (NIH) under grant number R01HL170368.", + "bbox": [ + 89, + 109, + 483, + 198 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 227 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Luke Auburn. Ai video generation expert discusses the technology's rapid advances—and its current limitations, 2024. 1", + "[2] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, KaiWei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 5, 6", + "[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 1, 6", + "[4] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3, 6, 11", + "[5] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5", + "[6] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, et al. A survey on in-context learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1107-1128, 2024. 5", + "[7] Theodoros Dumas and Alexandros Sigalas. Blender, an open source design tool: Advances and integration in the architectural production pipeline. Aristoteleio University of Thessaloniki, 21:737-744, 2009. 2, 6", + "[8] Blender Foundation. Upbge: an open-source, 3d game engine forked from the old blender game engine, 2024. 2", + "[9] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models, 2023. 2", + "[10] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In European Conference on Computer Vision, pages 393–411. Springer, 2025. 2", + "[11] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 3", + "[12] Kai Huang, Haoming Wang, and Wei Gao. Freezeasguard: Mitigating illegal adaptation of diffusion models via selec" + ], + "bbox": [ + 93, + 236, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tive tensor freezing. arXiv preprint arXiv:2405.17472, 2024. 2", + "[13] Kai Huang, Hanyun Yin, Heng Huang, and Wei Gao. Towards green ai in fine-tuning large language models via adaptive backpropagation. ICLR, 2024. 3", + "[14] Kai Huang, Xiangyu Yin, Heng Huang, and Wei Gao. Modality plug-and-play: Runtime modality adaptation in LLM-driven autonomous mobile systems. In ACM MobiCom, 2025. 3", + "[15] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6", + "[16] Epic Games Inc. Unreal engine: The most powerful real-time 3d creation tool, 2024. 2", + "[17] Mellis Inc. Pika labs, 2023. 1", + "[18] OpenAI Inc. Introducing openai o1-preview, 2024. 2, 6", + "[19] Marcel Krüger, David Gilbert, Torsten W. Kuhlen, and Tim Gerrits. Game engines for immersive visualization: Using unreal engine beyond entertainment. PRESENCE: Virtual and Augmented Reality, 33:31-55, 2024. 2, 6", + "[20] Chengxuan Li, Di Huang, Zeyu Lu, Yang Xiao, Qingqi Pei, and Lei Bai. A survey on long video generation: Challenges, methods, and prospects, 2024. 2", + "[21] WeiJie Li, Jin Wang, and Xuejie Zhang. Promptist: Automated prompt optimization for text-to-image synthesis. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 295-306. Springer, 2024. 6, 7, 11", + "[22] Yingcong Li, Kartik Sreenivasan, Angeliki Giannou, Dimitris Papailiopoulos, and Samet Oymak. Dissecting chain-of-thought: Compositionality through in-context filtering and learning. Advances in Neural Information Processing Systems, 36, 2024. 5", + "[23] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 3", + "[24] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023.", + "[25] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 3", + "[26] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imaged-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 2, 6", + "[27] Jiaxi Lv, Yi Huang, Mingfu Yan, Jiancheng Huang, Jianzhuang Liu, Yifan Liu, Yafei Wen, Xiaoxin Chen, and Shifeng Chen. Gpt4: Scripting physical motions in text-to-video generation via blender-oriented gpt planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1430-1440, 2024. 6" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18834", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Jiaxi Lv, Yi Huang, Mingfu Yan, Jiancheng Huang, Jianzhuang Liu, Yifan Liu, Yafei Wen, Xiaoxin Chen, and Shifeng Chen. Gpt4motion: Scripting physical motions in text-to-video generation via blender-oriented gpt planning, 2024. 2, 6, 7", + "[29] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363, 2024. 6", + "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3", + "[31] Weichao Qiu and Alan Yuille. Unrealcv: Connecting computer vision to unreal engine. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 909-916. Springer, 2016. 2, 6", + "[32] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications, 2024. 2", + "[33] Sander Schulhoff, Michael Ilie, Nishant Balepur, Konstantine Kahadze, Amanda Liu, Chenglei Si, Yinheng Li, Aayush Gupta, HyoJung Han, Sevien Schulhoff, Pranav Sandeep Dulepet, Saurav Vidyadhara, Dayeon Ki, Sweta Agrawal, Chau Pham, Gerson Kroiz, Feileen Li, Hudson Tao, Ashay Srivastava, Hevander Da Costa, Saloni Gupta, Megan L. Rogers, Inna Goncearenco, Giuseppe Sarli, Igor Galynker, Denis Peskoff, Marine Carpuat, Jules White, Shyamal Anadkat, Alexander Hoyle, and Philip Resnik. The prompt report: A systematic survey of prompting techniques, 2024. 2", + "[34] Sander Schulhoff, Michael Ilie, Nishant Balepur, Konstantine Kahadze, Amanda Liu, Chenglei Si, Yinheng Li, Aayush Gupta, HyoJung Han, Sevien Schulhoff, et al. The prompt report: A systematic survey of prompting techniques. arXiv preprint arXiv:2406.06608, 2024. 5", + "[35] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[36] Unity Technologies. Unity real-time development platform, 2024. 2", + "[37] Stuart A. Thompson. A.i. can now create lifelike videos. can you tell what's real?, 2024. 1", + "[38] Boshi Wang, Sewon Min, Xiang Deng, Jiaming Shen, You Wu, Luke Zettlemoyer, and Huan Sun. Towards understanding chain-of-thought prompting: An empirical study of what matters. arXiv preprint arXiv:2212.10001, 2022. 4", + "[39] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 2, 6, 11", + "[40] Keheng Wang, Feiyu Duan, Sirui Wang, Peiguang Li, Yun-sen Xian, Chuantao Yin, Wenge Rong, and Zhang Xiong. Knowledge-driven cot: Exploring faithful reasoning in llms" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for knowledge-intensive question answering. arXiv preprint arXiv:2308.13259, 2023. 5", + "[41] Xiaofeng Wang, Zheng Zhu, Guan Huang, Boyuan Wang, Xinze Chen, and Jiwen Lu. Worlddreamer: Towards general world models for video generation via predicting masked tokens, 2024. 2", + "[42] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3", + "[43] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 3", + "[44] Zhendong Wang, Yifan Jiang, Yadong Lu, Pengcheng He, Weizhu Chen, Zhangyang Wang, Mingyuan Zhou, et al. Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems, 36:8542-8562, 2023. 3", + "[45] Zhao Wang, Aoxue Li, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. Customvideo: Customizing text-to-video generation with multiple subjects. arXiv preprint arXiv:2401.09962, 2024. 3", + "[46] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 3", + "[47] Deshun Yang, Luhui Hu, Yu Tian, Zihao Li, Chris Kelly, Bang Yang, Cindy Yang, and Yuexian Zou. Worldgpt: a sora-inspired video ai agent as rich world models from text and image inputs. arXiv preprint arXiv:2403.07944, 2024. 6, 11", + "[48] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In *Forty-first International Conference on Machine Learning*, 2024. 3", + "[49] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 2023. 2", + "[50] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer, 2024. 1, 3, 11", + "[51] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6", + "[52] Wentao Zhang, Junliang Guo, Tianyu He, Li Zhao, Linli Xu, and Jiang Bian. Video in-context learning. arXiv preprint arXiv:2407.07356, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18835", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 6, 11", + "[54] Hanxin Zhu, Tianyu He, Anni Tang, Junliang Guo, Zhibo Chen, and Jiang Bian. Compositional 3d-aware video generation with llm director. arXiv preprint arXiv:2409.00558, 2024. 3" + ], + "bbox": [ + 91, + 92, + 480, + 200 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18836", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/d3a4c2ae-5aa4-4658-a693-f3535a28ce44_model.json b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/d3a4c2ae-5aa4-4658-a693-f3535a28ce44_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8daac0b5bac27e47a34f9b659884840eef49106a --- /dev/null +++ b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/d3a4c2ae-5aa4-4658-a693-f3535a28ce44_model.json @@ -0,0 +1,2235 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.111, + 0.857, + 0.154 + ], + "angle": 0, + "content": "PhyT2V: LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.173, + 0.71, + 0.208 + ], + "angle": 0, + "content": "Qiyao Xue, Xiangyu Yin, Boyuan Yang and Wei Gao \nUniversity of Pittsburgh" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.21, + 0.69, + 0.226 + ], + "angle": 0, + "content": "{qix63, eric.yin, by.yang, weigao}@pitt.edu" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.251, + 0.898, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.604, + 0.908, + 0.648 + ], + "angle": 0, + "content": "Figure 1. Left: videos generated by the current text-to-video generation model (CogVideoX-5B [50]) cannot adhere to the real-world physical rules (described in brackets following the user prompt). Right: our method PhyT2V, when applied to the same model, better reflects the real-world physical knowledge." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.661, + 0.327, + 0.675 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.679, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Text-to-video (T2V) generation has been recently enabled by transformer-based diffusion models, but current T2V models lack capabilities in adhering to the real-world common knowledge and physical rules, due to their limited understanding of physical realism and deficiency in temporal modeling. Existing solutions are either data-driven or require extra model inputs, but cannot be generalizable to out-of-distribution domains. In this paper, we present PhyT2V, a new data-independent T2V technique that expands the current T2V model's capability of video generation to out-of-distribution domains, by enabling chain-of-thought and step-back reasoning in T2V prompting. Our experiments show that PhyT2V improves existing T2V models' adherence to real-world physical rules by \\(2.3x\\), and achieves \\(35\\%\\) improvement compared to T2V prompt enhancers." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.661, + 0.645, + 0.675 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.678, + 0.907, + 0.812 + ], + "angle": 0, + "content": "Text-to-video (T2V) generation has recently marked a significant breakthrough of generative AI, with the advent of transformer-based diffusion models such as Sora [3], Pika [17] and CogVideoX [51] that can produce videos conditioned on textual prompts. These models demonstrate astonishing capabilities of generating complex and photorealistic scenes, and could even make it difficult for humans to distinguish between real-world and AI-generated videos, in the aspect of individual video frames' quality [1, 37]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.812, + 0.909, + 0.901 + ], + "angle": 0, + "content": "On the other hand, as shown in Figure 1 - Left, current T2V models still have significant drawbacks in adhering to the real-world common knowledge and physical rules, such as quantity, material, fluid dynamics, gravity, motion, collision and causality, and such limitations fundamentally prevent current T2V models from being used for real-world" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18826" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.255 + ], + "angle": 0, + "content": "simulation [7, 19, 31]. Enforcement of real-world knowledge and physical rules in T2V generation, however, is challenging because it requires the models' understandings of not only individual objects but also how these objects move and interact with each other. Further, unlike generating static images, T2V generation requires frame-to-frame consistency in object appearance, shape, motion, lighting and other dynamics [11]. Current T2V models often lack such temporal modeling, especially over long sequences [20], and the generated videos often contain flickering, inconsistent motion and object deformations across frames [26]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.258, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Most of the existing solutions to these challenges are data-driven, by using large multimodal T2V datasets that cover different real-world domains to train the diffusion model [10, 12, 41, 49]. However, these solutions heavily rely on the volume, quality and diversity of datasets [42, 51]. Since real-world common knowledge and physical rules are not explicitly embedded in the T2V generation process, the quality of video generation would largely drop in out-of-distribution domains that are not covered by the training dataset, and the generalizability of T2V models is limited due to the vast diversity of real-world scenario domains. Alternatively, researchers also use the existing 3D engines (e.g., Blender [8], Unity3D [36] and Unreal [16]) or mathematical models of edge and depth maps [26-28] to inject real-world physical knowledge into the T2V model, but these approaches are limited to fixed physical categories and patterns such as predefined objects and movements [26, 49], similarly lacking generalizability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.483, + 0.736 + ], + "angle": 0, + "content": "To achieve generalizable enforcement of physics-grounded T2V generation, we propose a fundamentally different approach: instead of expanding the training dataset or further complicating the T2V model architecture, we aim to expand the current T2V model's capability of video generation from in-distribution to out-of-distribution domains, by embedding real-world knowledge and physical rules into the text prompts with sufficient and appropriate contexts. To avoid ambiguous and unexplainable prompt engineering [9, 32, 33], our basic idea is to enable chain-of-thought (CoT) and step-back reasoning in T2V generation prompting, to ensure that T2V models follow correct physical dynamics and inter-frame consistency by applying step-by-step guidance and iterative refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.739, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Based on this idea, this paper presents Physcial-grounded Text-to-Video (PhyT2V), a new T2V technique that harnesses the natural language reasoning capabilities of well-trained LLMs (e.g., ChatGPT-4o), to facilitate CoT and step-back reasoning as described above. As shown in Figure 2, such reasoning is iteratively conducted in PhyT2V, and each iteration autonomously refines both the T2V prompt and generated video in three steps. In Step 1, the LLM analyzes the T2V prompt to extract objects to be shown and physical rules to follow in the video via in-context learning. In Step 2, we first use a video captioning model to translate" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.244, + 0.907, + 0.286 + ], + "angle": 0, + "content": "Figure 2. One iteration of video and prompt self-refinement in PhyT2V. Such self-refinement will be iteratively conducted in multiple rounds until the quality of generated video is satisfactory." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.456 + ], + "angle": 0, + "content": "the video's semantic contents into texts according to the list of objects obtained in Step 1, and then use the LLM to evaluate the mismatch between the video caption and current T2V prompt via CoT reasoning. In Step 3, the LLM refines the current T2V prompt, by incorporating the physical rules summarized in Step 1 and resolving the mismatch derived in Step 2, through step-back prompting. The refined T2V prompt is then used by the T2V model again for video generation, starting a new round of refinement. Such iterative refinement stops when the quality of generated video is satisfactory or the improvement of video quality converges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.909, + 0.573 + ], + "angle": 0, + "content": "For physical-grounded video generation performance, we further evaluated PhyT2V by applying it onto multiple SOTA T2V models, by using ChatGPT4 o1-preview [18] for LLM reasoning and Tarsier [39] as the video captioning model. We used two major T2V prompt datasets that cover 7 different real-world domains, and compared PhyT2V with the most competitive baselines of prompt enhancers. Our main findings are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.576, + 0.907, + 0.65 + ], + "angle": 0, + "content": "- PhyT2V is highly effective. Without involving any model retraining efforts on any auxiliary model inputs, PhyT2V can improve the adherence of the existing T2V models' generated videos to real-world physical rules by up to \\(2.3\\mathrm{x}\\), by only refining the text prompts to the T2V model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.65, + 0.906, + 0.737 + ], + "angle": 0, + "content": "- PhyT2V is high generic. It can result in significant improvement of video quality in a large diversity of real-world domains, covering solid, liquid, mechanics, optics, thermal, etc. It is fully data independent, and its prompting templates can be applied to any existing T2V models with different architectures and input formats." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.739, + 0.905, + 0.782 + ], + "angle": 0, + "content": "- Based on LLM-guided reasoning and self-refinement, PhyT2V is fully automated and involve the minimum amount of engineering and manual efforts." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.576, + 0.907, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.795, + 0.789, + 0.81 + ], + "angle": 0, + "content": "2. Related Work and Motivation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.82, + 0.736, + 0.834 + ], + "angle": 0, + "content": "2.1. T2V Generation Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Early T2V techniques generate video frames from text-to-image model outputs with temporal extensions [35], but cannot maintain temporal consistency and coherence over time, often producing visually appealing but temporally dis" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18827" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.226 + ], + "angle": 0, + "content": "connected outputs. Diffusion Transformers (DiT) [30] improved such consistency with a transformer backbone capable of capturing more complex temporal dynamics and relationships across frames through attention mechanism and long-range dependency modeling [42, 51]. Based on the DiT architecture, recent T2V models, such as OpenSora [53] and VideoCrafter [4], demonstrated that T2V generation can be further improved by in-context learning when provided with sufficient contextual information [44]." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.231, + 0.482, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.424 + ], + "angle": 0, + "content": "Figure 3. Examples of videos generated from in-distribution and out-of-distribution prompts, using the CogVideoX-5B model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.626 + ], + "angle": 0, + "content": "However, as shown in Figure 3, although these T2V models demonstrate strong capabilities in video generation when dealing with prompts aligned with the distributions found in the training data, they encounter significant challenges with out-of-distribution prompts that are not covered by training data1. In such cases, the outputs often contain physical illusions or artifacts, reflecting the model's limitations in generating realistic and coherent video contents under unfamiliar conditions. Such limitations can be addressed by enlarging the training datasets, improving T2V model architectures or developing new mechanisms for adaptation and error correction [43, 45], but these approaches are all prompt-specific and lack generalizability." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.641, + 0.483, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.725, + 0.483, + 0.753 + ], + "angle": 0, + "content": "Figure 4. A video generated by enhancing the out-of-distribution prompt \"Whisking egg into milk for scramble\" in Figure 3" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.484, + 0.851 + ], + "angle": 0, + "content": "On the other hand, as shown in Figure 4, recent research has demonstrated that the quality of video generation with an out-of-distribution prompt can be improved by refining the prompt with sufficient and appropriate details [11, 51]. These findings motivate our design of PhyT2V that embeds contexts of real-world knowledge and physical rules into" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.3 + ], + "angle": 0, + "content": "T2V prompts, to guide the T2V process for better physical accuracy and temporal alignment. The existing works, however, could still fail when tackling more intricate scenarios such as multi-object interactions, because the T2V model lacks an efficient feedback mechanism to learn how the generated video deviates from the real-world knowledge and physical rules. Researchers suggest to provide such feedback with extra input modalities to T2V models such as sampled video frames, depth map or scribble images [44, 52], but incur significant amounts of extra computing overhead and cannot be generalizable. Instead, in our design of PhyT2V, we aim to fully automate the feedback with only text prompts, and enable iterative feedback for the optimum video quality." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.306, + 0.891, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.517, + 0.866, + 0.532 + ], + "angle": 0, + "content": "Figure 5. Examples of CoT and step-back reasoning" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.543, + 0.791, + 0.559 + ], + "angle": 0, + "content": "2.2. Using LLM in T2V Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.907, + 0.782 + ], + "angle": 0, + "content": "LLMs with strong capabilities in natural language processing (NLP) have been a natural choice for prompt refinement in text-to-image and text-to-video generation, and existing work utilized LLMs to interpret text prompts and orchestrate the initial layout configurations [13, 14, 23-25, 46, 48, 54]. However, since current LLMs lack inherent understandings of the real-world physical laws, using LLMs with simple instructions usually result in videos that appear visually coherent but lack accurate physical realism, particularly when generating scenes with complex object interactions. Furthermore, these approaches frequently rely on static prompts or simple iterative refinements based on bounding box and segmentation map, which may capture basic visual attributes but fail to adapt to nuanced changes that require continuous physical modeling and adjustment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.783, + 0.909, + 0.903 + ], + "angle": 0, + "content": "An effective approach to addressing these limitations and providing effective feedback for prompt refinement is to explicitly trigger in-context learning and reasoning in LLM. For example, as shown in Figure 5, CoT reasoning deconstructs complex prompts into stepwise logical tasks, and hence ensures a precise scheduling path to align generated content with the input prompt. However, CoT reasoning, in some cases, could make errors in some intermediate steps," + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1In Figure 3, the in-distribution prompts are picked from the ones listed in [50], and the out-of-distribution prompts are our crafted ones for similar scenarios as the in-distribution prompts." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18828" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.156, + 0.09, + 0.465, + 0.103 + ], + "angle": 0, + "content": "Step 1: Identifying the Physical Rule and Main Object" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.104, + 0.504, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.621, + 0.09, + 0.845, + 0.102 + ], + "angle": 0, + "content": "Step 3: Generating the Refined Prompt" + }, + { + "type": "title", + "bbox": [ + 0.594, + 0.103, + 0.688, + 0.113 + ], + "angle": 0, + "content": "Task instruction" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.114, + 0.884, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.435, + 0.907, + 0.464 + ], + "angle": 0, + "content": "Figure 6. Our design of PhyT2V, illustrated by one round of video refinement consisting of three steps. Texts in brown are inputs from previous step. Texts in red are outputs to the next step; Texts in purple are prompts to trigger LLM reasoning" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.472, + 0.483, + 0.575 + ], + "angle": 0, + "content": "and step-back prompting can address this limitation by further deriving the step-back question at a higher level of abstraction and hence avoiding confusions and vagueness. In our design of PhyT2V, we will utilize such LLM reasoning to analyze the inconsistency of the generated video to real-world common knowledge and physical rules, and use the reasoning outcome as feedback for T2V prompt refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.842 + ], + "angle": 0, + "content": "The Chain-of-Thought (CoT) method is suitable for single data mode processing because it emphasizes linear decomposition and step-by-step reasoning, and it is especially effective for data processing flows that do not require cross-modal synchronization or interaction. However, multimodal data processing involves the fusion of data from different modalities and complex synchronization requirements, which cannot be completed through simple linear decomposition and requires frequent cross-modal information interaction and parallel processing, which exceeds the linear reasoning ability of the CoT method, resulting in limited performance in multimodal tasks. This is the reason why it is hard to directly apply CoT reasoning in T2V process itself as it requires multimodal alignment between the next and video modality. This also motivates us to adopt video captioning and use the video caption in the reasoning process, so that we can conduct CoT and step-back reasoning only in the text domain." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.851, + 0.182, + 0.865 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we present details of our PhyT2V design. In principal, PhyT2V's refinement of T2V generation is an" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.59 + ], + "angle": 0, + "content": "iterative process consisting of multiple rounds. In each round, as shown in Figure 6, the primary objective of our PhyT2V design is to guide a well-trained LLM (e.g., ChatGPT-4o) to generate a refined prompt that enables the pre-trained T2V model to generate videos that better match the given user prompt and real-world physical rules, and the refined prompt will be iteratively used as the new user prompt in the next round of refinement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.909, + 0.752 + ], + "angle": 0, + "content": "Each round of refinement is structured around decomposing the complex refinement problem into a series of simpler subproblems, more specifically, two parallel subproblems and one final subproblem. The two parallel subproblems are: Step 1) identifying the relevant physical rules that the generated video should follow based on the user prompt, and Step 2) identifying semantic mismatches between the user prompt and the generated video. Based on the knowledge about physical rules and semantic mismatches, the final subproblem (Step 3) generates the refined prompt to better adhere to the physical rules and resolve the mismatches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.909, + 0.842 + ], + "angle": 0, + "content": "To ensure proper identification in the parallel subproblems and prompt generation in the final subproblem, the core of PhyT2V design is two types of LLM reasoning processes within the prompt enhancement loop: the local CoT reasoning for individual subproblems and global step-back reasoning for the overall prompt refinement problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Local CoT reasoning is executed within the prompt for each subproblem, to prompt the LLM to generate a detailed reasoning chain in its latent embedding space [38]. Addressing the parallel subproblems facilitates LLM with a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18829" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.195 + ], + "angle": 0, + "content": "more concentrated attention on prerequisites of prompt refinement, enabling a deeper comprehension of the physical laws that govern the video content as well as the identification of discrepancies between the generated video and the user prompt. The outcomes derived from these parallel subproblems reflect the language model's abstraction in stepback reasoning on the overarching prompt refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.196, + 0.483, + 0.403 + ], + "angle": 0, + "content": "Global step-back reasoning: To integrate various subproblems into a coherent framework for prompt and video refinement, one intuitive approach involves employing CoT reasoning across these subproblems, allowing the LLM to engage in self-questioning. However, this method may lead to the risk of traversing incorrect reasoning pathways. Instead, we apply global step-back reasoning across subproblems, by using a self-augmented prompt to incorporate the LLM-generated responses to high-level questions about physical rules and semantic mismatches in earlier parallel problems, when generating the refined prompt in the final subproblem. In this way, we can improve the correctness of intermediate reasoning steps in CoT reasoning, and enable consistent improvement across steps in reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.482, + 0.491 + ], + "angle": 0, + "content": "Both reasoning processes are facilitated through appropriate task instruction prompting tailored to different subproblems. In general, our prompting procedure follows the prompt modeling in [34], which divides task instructions into several components. More details about these components in our design of PhyT2V are elaborated as follows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.802 + ], + "angle": 0, + "content": "Compared to the previous prompt enhancing methods, PhyT2V's key contribution is to analyze the semantic mismatch between currently generated video and the prompt, as well as refinements based on such mismatch. Previous methods can be formulated as \\( p' = f_{\\text{enhance}}(p, \\theta) \\), where \\( p \\) and \\( p' \\) are the original and enhanced prompts, \\( f_{\\text{enhance}} \\) is the prompt enhancer, and \\( \\theta \\) represents parameters or rules guiding the enhancement. In contrast, PhyT2V further involves the additional information about the T2V process, i.e., \\( p' = f_{\\text{enhance}}(p, f_{\\text{mismatch}}(C(V(p)), p), f_{\\text{phy}}(p), \\theta) \\), where \\( f_{\\text{phy}}(p) \\) analyzes the physical rules to be followed given \\( p \\), \\( V(p) \\) is the currently generated video given prompt \\( p \\), \\( C \\) is the video captioning model and \\( f_{\\text{mismatch}} \\) finds the semantic mismatch between \\( C(V(p)) \\) and \\( p \\). The key advantages are: (1) Semantic awareness: the refinement process explicitly incorporates the semantic mismatch to enable targeted T2V improvements; (2) Physical-world knowledge integration: physical rules derived from \\( p \\) enable guided enhancement; (3) Guided reasoning: unlike prior methods that rely solely on templates or embeddings, PhyT2V dynamically adapts prompt refinement to the semantic mismatch." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.809, + 0.483, + 0.841 + ], + "angle": 0, + "content": "3.1. Prompting in Parallel Subproblems for Local CoT Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In both Step 1 and Step 2, the first part of prompt is a task instruction prompt \\([I]\\) to instruct the LLM to understand the task in the subproblem. \\([I]\\) is designed with multiple components, each of which corresponds to different functions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.255 + ], + "angle": 0, + "content": "In the first sentence, it provides general guidance to relate the current subproblem to the entire refinement problem, to better condition the subproblem answer. Afterwards, it will include detailed descriptions of the task: identifying the physical rule and main object in Step 1, and identifying the semantic mismatch between the user prompt and caption of the generated video (generated by the video captioning model) in Step 2. It will also contain the requirements about the expected information in LLM's output. For example, in Step 1, the LLM's output about the physical rule should be in a descriptive way without giving formulas." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.256, + 0.905, + 0.374 + ], + "angle": 0, + "content": "Besides, to ensure proper CoT reasoning, we follow the existing work [22, 40] and provide in-context examples \\([E]\\) about tasks. To facilitate LLM's in-context learning [5, 6], \\([E]\\) is given in the format of QA pairs. That is, instead of fine-tuning a separate LLM checkpoint for each new task, we prompt the LLM with a few input-output exemplars, to demonstrate the task and condition the task's input-output format to the LLM, to guide the LLM's reasoning process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.374, + 0.905, + 0.464 + ], + "angle": 0, + "content": "Then, the final part of the prompt, denoted as \\([T]\\), is the information of the current instance of the task, usually with the current user prompt \\((P_{i})\\) being embedded. As a common practice of CoT reasoning, it also contains the hand-crafted trigger phrase \\((t)\\), \"Let's think step by step\", to activate the local CoT reasoning in LLM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.474, + 0.905, + 0.505 + ], + "angle": 0, + "content": "3.2. Prompting in the Final Subproblem for Global Step-Back Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.905, + 0.611 + ], + "angle": 0, + "content": "In the final subproblem, we enforce global step-back reasoning, by using the outputs of the two parallel subproblems above, i.e., knowledge about the physical rules and the prompt-video mismatch, as the high-level concepts and facts. Grounded on such high-level abstractions, we can make sure to improve the LLM's ability in following the correct reasoning path of generating the refined prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.905, + 0.803 + ], + "angle": 0, + "content": "Being similar to the prompts used in the two parallel subproblems above, the prompt structure in the final subproblem also contains \\([I]\\), \\([E]\\) and \\([T]\\). Furthermore, to ensure the correct reasoning path, we also provide quantitative feedback to the LLM about the effectiveness of previous round's prompt refinement. Such effectiveness could be measured by the existing T2V evaluators, which judge the semantic alignment and quality of physical common sense of the currently generated video2. For example, the VideoCon-Physics evaluator [2] gives a score \\(([S])\\) between 0 and 1. If \\([S]\\) is \\(< 0.5\\), it indicates that the refined prompt produced in the previous round is ineffective, hence guiding the LLM to take another alternative reasoning path." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.805, + 0.905, + 0.865 + ], + "angle": 0, + "content": "Since the prompt in the final subproblem is rich with reasoning and inherently very long-tailed, we removed the trigger prompt \\([t]\\), to prevent incorporating the information in the final answer unrelated to the user's initial input prompt." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "2This video is generated using the prompt refined in the previous round, and is also used to generate the video caption as the input in Step 2." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18830" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.312, + 0.108 + ], + "angle": 0, + "content": "3.3. The Stopping Condition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.484, + 0.242 + ], + "angle": 0, + "content": "The process of iterative refinement normally continues until the quality of the generated video is satisfactory, measured by the T2V evaluator as described above. Furthermore, the current T2V models naturally have limitations in generating some complicated or subtle scenes. In these cases, it would be difficult, even for PhyT2V, to reach physical realism after multiple rounds of iterations, and PhyT2V's refinement would stop when the iterations converge, i.e., the improvement of video quality becomes little over rounds." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.25, + 0.223, + 0.267 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.27, + 0.484, + 0.389 + ], + "angle": 0, + "content": "Models & Datasets: We applied PhyT2V on several DiT-based open-source T2V models, as listed below, and evaluated how PhyT2V improves the generated videos' adherence to real-world knowledge and physical rules. We use ChatGPT4 o1-preview [18] as the LLM for reasoning, and Tarsier [39] as the video captioning model. All generated videos last 6 seconds with 10 FPS and resolution of \\(720 \\times 480\\). Details of evaluation setup are in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.391, + 0.483, + 0.435 + ], + "angle": 0, + "content": "- CogVideoX [51]: It generates 10-second videos from text prompts, with 16 FPS and \\(768 \\times 1360\\) resolution. It offers two model variants with 2B and 5B parameters." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.436, + 0.483, + 0.48 + ], + "angle": 0, + "content": "- OpenSora 1.2 [53]: As an alternative to OpenAI's Sora [3], it contains 1.1B parameters and produces videos with 16 seconds, 720p resolution and different aspect ratios." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.48, + 0.483, + 0.525 + ], + "angle": 0, + "content": "- VideoCrafter [4]: With 1.8B parameters, it can generate both images and videos from text prompts, with \\(576 \\times 1024\\) resolution and a focus on video dynamics." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.391, + 0.483, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.586 + ], + "angle": 0, + "content": "Since we target enhancing the T2V models' capability of generating physics-grounded video contents, we use the following two prompt benchmarks that emphasize physical laws and adherence as the text prompts for T2V:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.588, + 0.483, + 0.661 + ], + "angle": 0, + "content": "- VideoPhy [2] is designed to assess whether the generated videos follow physical common sense for real-world activities. It consists 688 human-verified captions that describe interactions between various types of real-world objects, including solid and fluid." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.662, + 0.483, + 0.765 + ], + "angle": 0, + "content": "- PhyGenBench [29], similarly, allows evaluating the correctness of following physical common sense in T2V generation. It comprises 160 carefully crafted prompts spanning four physical domains, namely mechanics, optics, thermal and material properties. Since the domain of material properties has been covered by VideoPhy, we use the first three domains listed above." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.588, + 0.483, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Evaluation metric: We use VideoCon-Physics evaluator provided with VideoPhy [2], to measure how the generated video adheres to physical common sense (PC) and achieves semantic adherence (SA). The PC metric evaluates whether the depicted actions and object's state follow the real-world physics laws. The SA metric measures if the actions, events, entities and their interactions specified in the prompt are present. Both metrics yield binary outputs: 1 indicates adherence and 0 indicates otherwise. On each T2V model and" + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.08, + 0.882, + 0.242 + ], + "angle": 0, + "content": "
Round1234
CogVideoX-5B [51]PC0.260.320.390.42
SA0.480.520.560.59
CogVideoX-2B [51]PC0.130.190.270.29
SA0.220.120.400.42
OpenSora [53]PC0.170.290.270.31
SA0.290.380.440.47
VideoCrafter [4]PC0.150.250.290.33
SA0.240.380.440.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.284 + ], + "angle": 0, + "content": "Table 1. The quality of videos generated by different T2V models using the VideoPhy prompt dataset, over multiple rounds of iterative refinement in PhyT2V" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.287, + 0.882, + 0.449 + ], + "angle": 0, + "content": "
Round1234
CogVideoX-5B [51]PC0.280.320.380.42
SA0.220.350.360.38
CogVideoX-2B [51]PC0.160.190.240.27
SA0.150.290.330.35
OpenSora [53]PC0.210.250.240.26
SA0.230.280.290.30
VideoCrafter [4]PC0.200.240.320.36
SA0.270.330.370.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.45, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Table 2. The quality of videos generated by different T2V models using the PhyGenBench prompt dataset, over multiple rounds of iterative refinement in PhyT2V" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.502, + 0.892, + 0.518 + ], + "angle": 0, + "content": "dataset, the binary outputs from all prompts are averaged." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.905, + 0.592 + ], + "angle": 0, + "content": "In addition, we also evaluated PhyT2V using the widely used VBenchmark metrics and benchmarks [15], which allow comprehensive evaluations of the generated video in multiple aspects, including video quality, video-condition consistency, prompt following and human preference alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.906, + 0.713 + ], + "angle": 0, + "content": "Baselines: For fair comparison, we only use the existing T2V prompt enhancers as baselines, and other existing work with extra inputs to T2V models [7, 19, 26, 27, 31] are not applicable. We involve two prompt enhancers: 1) Directly using the existing LLM, particularly ChatGPT4, as the prompt enhancer [28, 47]; 2) Promptist [21], which uses reinforcement learning to automatically refine and enhance prompts in the model-preferred way." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.719, + 0.898, + 0.736 + ], + "angle": 0, + "content": "4.1. Improvement of the Generated Video Quality" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.737, + 0.906, + 0.885 + ], + "angle": 0, + "content": "As shown in Table 1 and 2, when PhyT2V is applied to different T2V models, it can significantly improve the generated video's adherence to both the text prompt itself and the real-world physical rules, compared to the videos generated by vanilla T2V models (i.e., in Round 1 of PhyT2V's refinement). In particular, such improvement is the most significant on the CogVideoX-2B model, where PC improvement can be up to \\(2.2\\mathrm{x}\\) and SA improvement can be up to \\(2.3\\mathrm{x}\\). On all the other models, PhyT2V can also reach noticeable improvement, ranging from \\(1.3\\mathrm{x}\\) to \\(1.9\\mathrm{x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Meanwhile, results in Table 1 and 2 showed that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18831" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.13, + 0.085, + 0.443, + 0.176 + ], + "angle": 0, + "content": "
CogVideoX-5BOpenSora
ChatGPT 4 [28]PC0.330.21
SA0.410.32
Promptist [21]PC0.250.19
SA0.390.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.178, + 0.483, + 0.207 + ], + "angle": 0, + "content": "Table 3. The quality of videos generated by enhancing the prompts in the VideoPhy dataset using different prompt enhancers" + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.211, + 0.443, + 0.302 + ], + "angle": 0, + "content": "
CogVideoX-5BOpenSora
ChatGPT 4 [28]PC0.270.20
SA0.230.23
Promptist [21]PC0.320.19
SA0.240.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.304, + 0.483, + 0.334 + ], + "angle": 0, + "content": "Table 4. The quality of videos generated by enhancing the prompts in the PhyGenBench dataset using different prompt enhancers" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.483, + 0.415 + ], + "angle": 0, + "content": "PhyT2V's process of iterative refinement converge quickly and only takes few rounds. Most improvement of video quality happens in the first two rounds, and little improvement can be observed in the fourth round. Hence, in practice, we believe that 3-4 iterative rounds would be sufficient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Furthermore, as shown in Table 3 and 4, PhyT2V also largely outperforms the existing prompt enhancers by at least \\(35\\%\\), when being applied to CogVideoX-5B and OpenSora models. In particular, ChatGPT 4, when being used as the prompt enhancer, delivers better performance than Promptist due to its stronger language processing capabilities, but still cannot ensure physics-grounded T2V, due to the lack of explicit reasoning on text-to-video alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.483, + 0.668 + ], + "angle": 0, + "content": "Our evaluation results on the VBenchmark metrics are shown in Figure 7, where numbers in Round 1 are the T2V model's original performance in current VBenchmark leaderboard, and iterative prompt refinements by PhyT2V in Round 2 & 3 noticeably improve the performance in many dimensions. In particular, large improvements are noted in most dimensions of Video-Condition Consistency, showing that PhyT2V improves T2V model's adherence to prompts and real-world physical rules underlying the prompts." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.671, + 0.483, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.798, + 0.483, + 0.829 + ], + "angle": 0, + "content": "Figure 7. PhyT2V VBenchmark evaluation results with CogVideoX-5B (left) and OpenSora (right)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.839, + 0.406, + 0.855 + ], + "angle": 0, + "content": "4.2. Different Domains of Physical Rules" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We also conducted in-depth analysis on PhyT2V's performance on improving the generated video's quality in different domains of real-world physical rules, using the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.08, + 0.903, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.276, + 0.907, + 0.303 + ], + "angle": 0, + "content": "Figure 8. Examples of videos generated using different categories of prompts in the PhyGenBench dataset" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.511, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Figure 9. Examples of videos generated using different categories of prompts in the VideoPhy dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.656 + ], + "angle": 0, + "content": "CogVideoX-5B as the T2V model and ChatGPT 4 as the prompt enhancer. As shown in Table 5 and 6, PhyT2V achieves large improvements in most domains of physical rules. Especially in domains where physical interaction between objects are more subtle and difficult to be precisely captured, such as interaction with fluids and thermal-related scene changes, such improvements will be even higher." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.82 + ], + "angle": 0, + "content": "These improvements are also exemplified with sample videos and their related input prompts in Figure 9 and Figure 8. With LLM reasoning and iterative refinement, PhyT2V can largely enhance the T2V model's capability when encountering out-of-distribution prompts, by providing correct and sufficient contexts to ensure that the T2V model's video generation correctly captures the key objects and interaction between objects. For example, when the prompt of \"juice dropping from a bottle onto the counter\", PhyT2V correctly depicts the juice's slow diffusion on the counter. More examples can be found in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.825, + 0.667, + 0.84 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We conduct an ablation study to evaluate the necessity of both the physical rule reasoning (Step 1) and the mismatch reasoning (Step 2) within our PhyT2V workflow, by removing one of these steps from the refinement process to assess" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18832" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.891, + 0.234 + ], + "angle": 0, + "content": "
CogVideoX-5BCogVideoX-2BOpenSoraVideoCrafter
Round1234123412341234
Solid-SolidPC0.210.280.340.320.090.130.140.220.120.270.290.300.190.220.270.28
SA0.240.480.490.470.180.250.360.330.160.340.370.350.240.400.450.47
Solid-FluidPC0.220.270.280.300.110.180.280.270.170.210.240.250.180.240.250.26
SA0.390.540.600.610.290.430.440.430.160.400.410.360.340.430.480.52
Fluid-FluidPC0.570.590.630.620.340.380.350.360.150.320.290.310.330.410.530.51
SA0.410.570.590.670.270.420.390.440.310.440.450.460.320.420.490.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.236, + 0.868, + 0.251 + ], + "angle": 0, + "content": "Table 5. The improvement of generated video quality in different categories of physical rules in the VideoPhy prompt dataset" + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.254, + 0.89, + 0.4 + ], + "angle": 0, + "content": "
CogVideoX-5BCogVideoX-2BOpenSoraVideoCrafter
Round1234123412341234
MechanicsPC0.190.250.340.350.120.160.180.240.110.130.170.220.140.230.290.28
SA0.210.280.290.320.110.180.190.220.190.210.270.320.200.240.280.35
OpticsPC0.220.350.410.390.220.250.290.280.240.260.250.250.220.210.270.32
SA0.270.420.390.440.230.340.370.390.260.310.290.300.220.280.350.39
ThermalPC0.330.350.350.350.130.150.150.140.270.300.310.330.250.280.260.28
SA0.220.360.430.450.120.160.240.270.230.250.370.360.250.370.410.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.401, + 0.882, + 0.415 + ], + "angle": 0, + "content": "Table 6. The improvement of generated video quality in different categories of physical rules in the PhyGenBench prompt dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.423, + 0.438 + ], + "angle": 0, + "content": "its impact on physical-grounded video generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Physical rule reasoning (Step 1). As shown in Figure 10, the Step 1 of physical rule reasoning significantly enhances the T2V process by providing a more detailed and coherent description of the principal object's physical status, such as motion, states and deformation (red texts in Figure 10), all grounded in relevant physical laws. By anchoring the prompt in established physical rules, this step also helps avoid unnecessary texts (brown texts in Figure 10) and vague physical rule descriptions (purple texts in Figure 10), hence achieving a higher PC score." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.589, + 0.48, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.763, + 0.474, + 0.777 + ], + "angle": 0, + "content": "Figure 10. Ablation study on Step 1 of physical rule reasoning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.782, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Mismatch reasoning (Step 2). The Step 2 of mismatch reasoning addresses details that may have been overlooked in the previous iteration of the generated video as shown in Figure 11. This step plays a critical role in the iterative refinement process by identifying and correcting discrepancies between expected and observed outputs. By enhancing the model's focus on the principal object, the mismatch reasoning step reduces the likelihood of losing attention to" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.42, + 0.905, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.591, + 0.887, + 0.606 + ], + "angle": 0, + "content": "Figure 11. Ablation study on Step 2 of mismatch reasoning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.905, + 0.654 + ], + "angle": 0, + "content": "important features (brown and purple texts in Figure 11), improving the fidelity and relevance of generated video content (red texts in Figure 11) towards a higher SA score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.906, + 0.715 + ], + "angle": 0, + "content": "Overall, our study shows that both reasoning steps are integral to PhyT2V, contributing to a more robust and semantically-aligned generation of refined prompts in Step 3. More detailed ablation studies are in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.72, + 0.634, + 0.734 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.74, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present PhyT2V, a novel data-independent T2V generation framework designed to enhance the generalization capability of existing T2V models to out-of-distribution domains. By incorporating CoT reasoning and step-back prompting, PhyT2V systematically refines T2V prompts to ensure adherence to real-world physical principles without necessitating additional model retraining or reliance on additional conditions. Evaluation results indicate that PhyT2V achieves a \\(2.3\\mathrm{x}\\) enhancement in physical realism compared to baseline T2V models and outperforms state-of-the-art T2V prompt enhancers by \\(35\\%\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18833" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.111, + 0.484, + 0.199 + ], + "angle": 0, + "content": "We thank the reviewers and the area chair for their insightful comments and feedback. This work was supported in part by National Science Foundation (NSF) under grant number IIS-2205360, CCF-2217003, CCF-2215042, and National Institutes of Health (NIH) under grant number R01HL170368." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.212, + 0.188, + 0.228 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.237, + 0.484, + 0.277 + ], + "angle": 0, + "content": "[1] Luke Auburn. Ai video generation expert discusses the technology's rapid advances—and its current limitations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.278, + 0.484, + 0.348 + ], + "angle": 0, + "content": "[2] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, KaiWei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.348, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[4] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.484, + 0.54 + ], + "angle": 0, + "content": "[5] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.484, + 0.61 + ], + "angle": 0, + "content": "[6] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, et al. A survey on in-context learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1107-1128, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.61, + 0.485, + 0.664 + ], + "angle": 0, + "content": "[7] Theodoros Dumas and Alexandros Sigalas. Blender, an open source design tool: Advances and integration in the architectural production pipeline. Aristoteleio University of Thessaloniki, 21:737-744, 2009. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.484, + 0.694 + ], + "angle": 0, + "content": "[8] Blender Foundation. Upbge: an open-source, 3d game engine forked from the old blender game engine, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.694, + 0.484, + 0.749 + ], + "angle": 0, + "content": "[9] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.749, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[10] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In European Conference on Computer Vision, pages 393–411. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[11] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Kai Huang, Haoming Wang, and Wei Gao. Freezeasguard: Mitigating illegal adaptation of diffusion models via selec" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.237, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tive tensor freezing. arXiv preprint arXiv:2405.17472, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[13] Kai Huang, Hanyun Yin, Heng Huang, and Wei Gao. Towards green ai in fine-tuning large language models via adaptive backpropagation. ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[14] Kai Huang, Xiangyu Yin, Heng Huang, and Wei Gao. Modality plug-and-play: Runtime modality adaptation in LLM-driven autonomous mobile systems. In ACM MobiCom, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[15] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[16] Epic Games Inc. Unreal engine: The most powerful real-time 3d creation tool, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.726, + 0.348 + ], + "angle": 0, + "content": "[17] Mellis Inc. Pika labs, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.878, + 0.365 + ], + "angle": 0, + "content": "[18] OpenAI Inc. Introducing openai o1-preview, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.366, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[19] Marcel Krüger, David Gilbert, Torsten W. Kuhlen, and Tim Gerrits. Game engines for immersive visualization: Using unreal engine beyond entertainment. PRESENCE: Virtual and Augmented Reality, 33:31-55, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.422, + 0.904, + 0.464 + ], + "angle": 0, + "content": "[20] Chengxuan Li, Di Huang, Zeyu Lu, Yang Xiao, Qingqi Pei, and Lei Bai. A survey on long video generation: Challenges, methods, and prospects, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.466, + 0.905, + 0.534 + ], + "angle": 0, + "content": "[21] WeiJie Li, Jin Wang, and Xuejie Zhang. Promptist: Automated prompt optimization for text-to-image synthesis. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 295-306. Springer, 2024. 6, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.536, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[22] Yingcong Li, Kartik Sreenivasan, Angeliki Giannou, Dimitris Papailiopoulos, and Samet Oymak. Dissecting chain-of-thought: Compositionality through in-context filtering and learning. Advances in Neural Information Processing Systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.905, + 0.661 + ], + "angle": 0, + "content": "[23] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.662, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[24] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.705, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[25] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.818 + ], + "angle": 0, + "content": "[26] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imaged-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[27] Jiaxi Lv, Yi Huang, Mingfu Yan, Jiancheng Huang, Jianzhuang Liu, Yifan Liu, Yafei Wen, Xiaoxin Chen, and Shifeng Chen. Gpt4: Scripting physical motions in text-to-video generation via blender-oriented gpt planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1430-1440, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18834" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[28] Jiaxi Lv, Yi Huang, Mingfu Yan, Jiancheng Huang, Jianzhuang Liu, Yifan Liu, Yafei Wen, Xiaoxin Chen, and Shifeng Chen. Gpt4motion: Scripting physical motions in text-to-video generation via blender-oriented gpt planning, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.161, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[29] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.285 + ], + "angle": 0, + "content": "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[31] Weichao Qiu and Alan Yuille. Unrealcv: Connecting computer vision to unreal engine. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 909-916. Springer, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[32] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[33] Sander Schulhoff, Michael Ilie, Nishant Balepur, Konstantine Kahadze, Amanda Liu, Chenglei Si, Yinheng Li, Aayush Gupta, HyoJung Han, Sevien Schulhoff, Pranav Sandeep Dulepet, Saurav Vidyadhara, Dayeon Ki, Sweta Agrawal, Chau Pham, Gerson Kroiz, Feileen Li, Hudson Tao, Ashay Srivastava, Hevander Da Costa, Saloni Gupta, Megan L. Rogers, Inna Goncearenco, Giuseppe Sarli, Igor Galynker, Denis Peskoff, Marine Carpuat, Jules White, Shyamal Anadkat, Alexander Hoyle, and Philip Resnik. The prompt report: A systematic survey of prompting techniques, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.551, + 0.482, + 0.619 + ], + "angle": 0, + "content": "[34] Sander Schulhoff, Michael Ilie, Nishant Balepur, Konstantine Kahadze, Amanda Liu, Chenglei Si, Yinheng Li, Aayush Gupta, HyoJung Han, Sevien Schulhoff, et al. The prompt report: A systematic survey of prompting techniques. arXiv preprint arXiv:2406.06608, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.621, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[35] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.717 + ], + "angle": 0, + "content": "[36] Unity Technologies. Unity real-time development platform, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[37] Stuart A. Thompson. A.i. can now create lifelike videos. can you tell what's real?, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[38] Boshi Wang, Sewon Min, Xiang Deng, Jiaming Shen, You Wu, Luke Zettlemoyer, and Huan Sun. Towards understanding chain-of-thought prompting: An empirical study of what matters. arXiv preprint arXiv:2212.10001, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[39] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 2, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Keheng Wang, Feiyu Duan, Sirui Wang, Peiguang Li, Yun-sen Xian, Chuantao Yin, Wenge Rong, and Zhang Xiong. Knowledge-driven cot: Exploring faithful reasoning in llms" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "for knowledge-intensive question answering. arXiv preprint arXiv:2308.13259, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[41] Xiaofeng Wang, Zheng Zhu, Guan Huang, Boyuan Wang, Xinze Chen, and Jiwen Lu. Worlddreamer: Towards general world models for video generation via predicting masked tokens, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[42] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[43] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[44] Zhendong Wang, Yifan Jiang, Yadong Lu, Pengcheng He, Weizhu Chen, Zhangyang Wang, Mingyuan Zhou, et al. Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems, 36:8542-8562, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[45] Zhao Wang, Aoxue Li, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. Customvideo: Customizing text-to-video generation with multiple subjects. arXiv preprint arXiv:2401.09962, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[46] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[47] Deshun Yang, Luhui Hu, Yu Tian, Zihao Li, Chris Kelly, Bang Yang, Cindy Yang, and Yuexian Zou. Worldgpt: a sora-inspired video ai agent as rich world models from text and image inputs. arXiv preprint arXiv:2403.07944, 2024. 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[48] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In *Forty-first International Conference on Machine Learning*, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[49] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[50] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer, 2024. 1, 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[51] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Wentao Zhang, Junliang Guo, Tianyu He, Li Zhao, Linli Xu, and Jiang Bian. Video in-context learning. arXiv preprint arXiv:2407.07356, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18835" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[53] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.201 + ], + "angle": 0, + "content": "[54] Hanxin Zhu, Tianyu He, Anni Tang, Junliang Guo, Zhibo Chen, and Jiang Bian. Compositional 3d-aware video generation with llm director. arXiv preprint arXiv:2409.00558, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18836" + } + ] +] \ No newline at end of file diff --git a/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/d3a4c2ae-5aa4-4658-a693-f3535a28ce44_origin.pdf b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/d3a4c2ae-5aa4-4658-a693-f3535a28ce44_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..964d8c209114f874060a29968d2d307f19eec2d7 --- /dev/null +++ b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/d3a4c2ae-5aa4-4658-a693-f3535a28ce44_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf20d191d245360276d645142178bc8e6ef3de7f871cac50842771c16c0f09ac +size 2518532 diff --git a/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/full.md b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..12fe1a61fa27b706b48a6733fb52d7734e96d5e1 --- /dev/null +++ b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/full.md @@ -0,0 +1,289 @@ +# PhyT2V: LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation + +Qiyao Xue, Xiangyu Yin, Boyuan Yang and Wei Gao +University of Pittsburgh + +{qix63, eric.yin, by.yang, weigao}@pitt.edu + +![](images/8113c8c3ee5a7b963c2ebb52360142362a68b3e5e05009dfad7a0d2b66fac69b.jpg) +Figure 1. Left: videos generated by the current text-to-video generation model (CogVideoX-5B [50]) cannot adhere to the real-world physical rules (described in brackets following the user prompt). Right: our method PhyT2V, when applied to the same model, better reflects the real-world physical knowledge. + +# Abstract + +Text-to-video (T2V) generation has been recently enabled by transformer-based diffusion models, but current T2V models lack capabilities in adhering to the real-world common knowledge and physical rules, due to their limited understanding of physical realism and deficiency in temporal modeling. Existing solutions are either data-driven or require extra model inputs, but cannot be generalizable to out-of-distribution domains. In this paper, we present PhyT2V, a new data-independent T2V technique that expands the current T2V model's capability of video generation to out-of-distribution domains, by enabling chain-of-thought and step-back reasoning in T2V prompting. Our experiments show that PhyT2V improves existing T2V models' adherence to real-world physical rules by $2.3x$ , and achieves $35\%$ improvement compared to T2V prompt enhancers. + +# 1. Introduction + +Text-to-video (T2V) generation has recently marked a significant breakthrough of generative AI, with the advent of transformer-based diffusion models such as Sora [3], Pika [17] and CogVideoX [51] that can produce videos conditioned on textual prompts. These models demonstrate astonishing capabilities of generating complex and photorealistic scenes, and could even make it difficult for humans to distinguish between real-world and AI-generated videos, in the aspect of individual video frames' quality [1, 37]. + +On the other hand, as shown in Figure 1 - Left, current T2V models still have significant drawbacks in adhering to the real-world common knowledge and physical rules, such as quantity, material, fluid dynamics, gravity, motion, collision and causality, and such limitations fundamentally prevent current T2V models from being used for real-world + +simulation [7, 19, 31]. Enforcement of real-world knowledge and physical rules in T2V generation, however, is challenging because it requires the models' understandings of not only individual objects but also how these objects move and interact with each other. Further, unlike generating static images, T2V generation requires frame-to-frame consistency in object appearance, shape, motion, lighting and other dynamics [11]. Current T2V models often lack such temporal modeling, especially over long sequences [20], and the generated videos often contain flickering, inconsistent motion and object deformations across frames [26]. + +Most of the existing solutions to these challenges are data-driven, by using large multimodal T2V datasets that cover different real-world domains to train the diffusion model [10, 12, 41, 49]. However, these solutions heavily rely on the volume, quality and diversity of datasets [42, 51]. Since real-world common knowledge and physical rules are not explicitly embedded in the T2V generation process, the quality of video generation would largely drop in out-of-distribution domains that are not covered by the training dataset, and the generalizability of T2V models is limited due to the vast diversity of real-world scenario domains. Alternatively, researchers also use the existing 3D engines (e.g., Blender [8], Unity3D [36] and Unreal [16]) or mathematical models of edge and depth maps [26-28] to inject real-world physical knowledge into the T2V model, but these approaches are limited to fixed physical categories and patterns such as predefined objects and movements [26, 49], similarly lacking generalizability. + +To achieve generalizable enforcement of physics-grounded T2V generation, we propose a fundamentally different approach: instead of expanding the training dataset or further complicating the T2V model architecture, we aim to expand the current T2V model's capability of video generation from in-distribution to out-of-distribution domains, by embedding real-world knowledge and physical rules into the text prompts with sufficient and appropriate contexts. To avoid ambiguous and unexplainable prompt engineering [9, 32, 33], our basic idea is to enable chain-of-thought (CoT) and step-back reasoning in T2V generation prompting, to ensure that T2V models follow correct physical dynamics and inter-frame consistency by applying step-by-step guidance and iterative refinement. + +Based on this idea, this paper presents Physcial-grounded Text-to-Video (PhyT2V), a new T2V technique that harnesses the natural language reasoning capabilities of well-trained LLMs (e.g., ChatGPT-4o), to facilitate CoT and step-back reasoning as described above. As shown in Figure 2, such reasoning is iteratively conducted in PhyT2V, and each iteration autonomously refines both the T2V prompt and generated video in three steps. In Step 1, the LLM analyzes the T2V prompt to extract objects to be shown and physical rules to follow in the video via in-context learning. In Step 2, we first use a video captioning model to translate + +![](images/36aad3ee25808e3108664a0a95864a636edf3867f4adeedb7f860368e96007ac.jpg) +Figure 2. One iteration of video and prompt self-refinement in PhyT2V. Such self-refinement will be iteratively conducted in multiple rounds until the quality of generated video is satisfactory. + +the video's semantic contents into texts according to the list of objects obtained in Step 1, and then use the LLM to evaluate the mismatch between the video caption and current T2V prompt via CoT reasoning. In Step 3, the LLM refines the current T2V prompt, by incorporating the physical rules summarized in Step 1 and resolving the mismatch derived in Step 2, through step-back prompting. The refined T2V prompt is then used by the T2V model again for video generation, starting a new round of refinement. Such iterative refinement stops when the quality of generated video is satisfactory or the improvement of video quality converges. + +For physical-grounded video generation performance, we further evaluated PhyT2V by applying it onto multiple SOTA T2V models, by using ChatGPT4 o1-preview [18] for LLM reasoning and Tarsier [39] as the video captioning model. We used two major T2V prompt datasets that cover 7 different real-world domains, and compared PhyT2V with the most competitive baselines of prompt enhancers. Our main findings are as follows: + +- PhyT2V is highly effective. Without involving any model retraining efforts on any auxiliary model inputs, PhyT2V can improve the adherence of the existing T2V models' generated videos to real-world physical rules by up to $2.3\mathrm{x}$ , by only refining the text prompts to the T2V model. +- PhyT2V is high generic. It can result in significant improvement of video quality in a large diversity of real-world domains, covering solid, liquid, mechanics, optics, thermal, etc. It is fully data independent, and its prompting templates can be applied to any existing T2V models with different architectures and input formats. +- Based on LLM-guided reasoning and self-refinement, PhyT2V is fully automated and involve the minimum amount of engineering and manual efforts. + +# 2. Related Work and Motivation + +# 2.1. T2V Generation Models + +Early T2V techniques generate video frames from text-to-image model outputs with temporal extensions [35], but cannot maintain temporal consistency and coherence over time, often producing visually appealing but temporally dis + +connected outputs. Diffusion Transformers (DiT) [30] improved such consistency with a transformer backbone capable of capturing more complex temporal dynamics and relationships across frames through attention mechanism and long-range dependency modeling [42, 51]. Based on the DiT architecture, recent T2V models, such as OpenSora [53] and VideoCrafter [4], demonstrated that T2V generation can be further improved by in-context learning when provided with sufficient contextual information [44]. + +![](images/af4d0ad4dec137b51df26c071e6ca57bdf56b2c75891f171a6ef9623aad6a3dc.jpg) +Figure 3. Examples of videos generated from in-distribution and out-of-distribution prompts, using the CogVideoX-5B model + +However, as shown in Figure 3, although these T2V models demonstrate strong capabilities in video generation when dealing with prompts aligned with the distributions found in the training data, they encounter significant challenges with out-of-distribution prompts that are not covered by training data1. In such cases, the outputs often contain physical illusions or artifacts, reflecting the model's limitations in generating realistic and coherent video contents under unfamiliar conditions. Such limitations can be addressed by enlarging the training datasets, improving T2V model architectures or developing new mechanisms for adaptation and error correction [43, 45], but these approaches are all prompt-specific and lack generalizability. + +![](images/f053aeeb42b356aeecd1aa01478efece326b8b30f3502fb76732588c06d29b49.jpg) +Figure 4. A video generated by enhancing the out-of-distribution prompt "Whisking egg into milk for scramble" in Figure 3 + +On the other hand, as shown in Figure 4, recent research has demonstrated that the quality of video generation with an out-of-distribution prompt can be improved by refining the prompt with sufficient and appropriate details [11, 51]. These findings motivate our design of PhyT2V that embeds contexts of real-world knowledge and physical rules into + +T2V prompts, to guide the T2V process for better physical accuracy and temporal alignment. The existing works, however, could still fail when tackling more intricate scenarios such as multi-object interactions, because the T2V model lacks an efficient feedback mechanism to learn how the generated video deviates from the real-world knowledge and physical rules. Researchers suggest to provide such feedback with extra input modalities to T2V models such as sampled video frames, depth map or scribble images [44, 52], but incur significant amounts of extra computing overhead and cannot be generalizable. Instead, in our design of PhyT2V, we aim to fully automate the feedback with only text prompts, and enable iterative feedback for the optimum video quality. + +![](images/d97ec0eb02e0075ebca8bde89d0348e0d0f9fe2cfc5837ed20b412d5b2b9978c.jpg) +Figure 5. Examples of CoT and step-back reasoning + +# 2.2. Using LLM in T2V Generation + +LLMs with strong capabilities in natural language processing (NLP) have been a natural choice for prompt refinement in text-to-image and text-to-video generation, and existing work utilized LLMs to interpret text prompts and orchestrate the initial layout configurations [13, 14, 23-25, 46, 48, 54]. However, since current LLMs lack inherent understandings of the real-world physical laws, using LLMs with simple instructions usually result in videos that appear visually coherent but lack accurate physical realism, particularly when generating scenes with complex object interactions. Furthermore, these approaches frequently rely on static prompts or simple iterative refinements based on bounding box and segmentation map, which may capture basic visual attributes but fail to adapt to nuanced changes that require continuous physical modeling and adjustment. + +An effective approach to addressing these limitations and providing effective feedback for prompt refinement is to explicitly trigger in-context learning and reasoning in LLM. For example, as shown in Figure 5, CoT reasoning deconstructs complex prompts into stepwise logical tasks, and hence ensures a precise scheduling path to align generated content with the input prompt. However, CoT reasoning, in some cases, could make errors in some intermediate steps, + +# Step 1: Identifying the Physical Rule and Main Object + +![](images/89f6e74f50e27e472e40e451539de5cb9d058b750440cf8fd26e92e8d4eaece4.jpg) +Figure 6. Our design of PhyT2V, illustrated by one round of video refinement consisting of three steps. Texts in brown are inputs from previous step. Texts in red are outputs to the next step; Texts in purple are prompts to trigger LLM reasoning + +# Step 3: Generating the Refined Prompt + +# Task instruction + +![](images/ed1642fa16d59bb50d36be50fcef2135f54234741e9fc8384f12f2ebf58833b9.jpg) + +and step-back prompting can address this limitation by further deriving the step-back question at a higher level of abstraction and hence avoiding confusions and vagueness. In our design of PhyT2V, we will utilize such LLM reasoning to analyze the inconsistency of the generated video to real-world common knowledge and physical rules, and use the reasoning outcome as feedback for T2V prompt refinement. + +The Chain-of-Thought (CoT) method is suitable for single data mode processing because it emphasizes linear decomposition and step-by-step reasoning, and it is especially effective for data processing flows that do not require cross-modal synchronization or interaction. However, multimodal data processing involves the fusion of data from different modalities and complex synchronization requirements, which cannot be completed through simple linear decomposition and requires frequent cross-modal information interaction and parallel processing, which exceeds the linear reasoning ability of the CoT method, resulting in limited performance in multimodal tasks. This is the reason why it is hard to directly apply CoT reasoning in T2V process itself as it requires multimodal alignment between the next and video modality. This also motivates us to adopt video captioning and use the video caption in the reasoning process, so that we can conduct CoT and step-back reasoning only in the text domain. + +# 3. Method + +In this section, we present details of our PhyT2V design. In principal, PhyT2V's refinement of T2V generation is an + +iterative process consisting of multiple rounds. In each round, as shown in Figure 6, the primary objective of our PhyT2V design is to guide a well-trained LLM (e.g., ChatGPT-4o) to generate a refined prompt that enables the pre-trained T2V model to generate videos that better match the given user prompt and real-world physical rules, and the refined prompt will be iteratively used as the new user prompt in the next round of refinement. + +Each round of refinement is structured around decomposing the complex refinement problem into a series of simpler subproblems, more specifically, two parallel subproblems and one final subproblem. The two parallel subproblems are: Step 1) identifying the relevant physical rules that the generated video should follow based on the user prompt, and Step 2) identifying semantic mismatches between the user prompt and the generated video. Based on the knowledge about physical rules and semantic mismatches, the final subproblem (Step 3) generates the refined prompt to better adhere to the physical rules and resolve the mismatches. + +To ensure proper identification in the parallel subproblems and prompt generation in the final subproblem, the core of PhyT2V design is two types of LLM reasoning processes within the prompt enhancement loop: the local CoT reasoning for individual subproblems and global step-back reasoning for the overall prompt refinement problem. + +Local CoT reasoning is executed within the prompt for each subproblem, to prompt the LLM to generate a detailed reasoning chain in its latent embedding space [38]. Addressing the parallel subproblems facilitates LLM with a + +more concentrated attention on prerequisites of prompt refinement, enabling a deeper comprehension of the physical laws that govern the video content as well as the identification of discrepancies between the generated video and the user prompt. The outcomes derived from these parallel subproblems reflect the language model's abstraction in stepback reasoning on the overarching prompt refinement. + +Global step-back reasoning: To integrate various subproblems into a coherent framework for prompt and video refinement, one intuitive approach involves employing CoT reasoning across these subproblems, allowing the LLM to engage in self-questioning. However, this method may lead to the risk of traversing incorrect reasoning pathways. Instead, we apply global step-back reasoning across subproblems, by using a self-augmented prompt to incorporate the LLM-generated responses to high-level questions about physical rules and semantic mismatches in earlier parallel problems, when generating the refined prompt in the final subproblem. In this way, we can improve the correctness of intermediate reasoning steps in CoT reasoning, and enable consistent improvement across steps in reasoning. + +Both reasoning processes are facilitated through appropriate task instruction prompting tailored to different subproblems. In general, our prompting procedure follows the prompt modeling in [34], which divides task instructions into several components. More details about these components in our design of PhyT2V are elaborated as follows. + +Compared to the previous prompt enhancing methods, PhyT2V's key contribution is to analyze the semantic mismatch between currently generated video and the prompt, as well as refinements based on such mismatch. Previous methods can be formulated as $p' = f_{\text{enhance}}(p, \theta)$ , where $p$ and $p'$ are the original and enhanced prompts, $f_{\text{enhance}}$ is the prompt enhancer, and $\theta$ represents parameters or rules guiding the enhancement. In contrast, PhyT2V further involves the additional information about the T2V process, i.e., $p' = f_{\text{enhance}}(p, f_{\text{mismatch}}(C(V(p)), p), f_{\text{phy}}(p), \theta)$ , where $f_{\text{phy}}(p)$ analyzes the physical rules to be followed given $p$ , $V(p)$ is the currently generated video given prompt $p$ , $C$ is the video captioning model and $f_{\text{mismatch}}$ finds the semantic mismatch between $C(V(p))$ and $p$ . The key advantages are: (1) Semantic awareness: the refinement process explicitly incorporates the semantic mismatch to enable targeted T2V improvements; (2) Physical-world knowledge integration: physical rules derived from $p$ enable guided enhancement; (3) Guided reasoning: unlike prior methods that rely solely on templates or embeddings, PhyT2V dynamically adapts prompt refinement to the semantic mismatch. + +# 3.1. Prompting in Parallel Subproblems for Local CoT Reasoning + +In both Step 1 and Step 2, the first part of prompt is a task instruction prompt $[I]$ to instruct the LLM to understand the task in the subproblem. $[I]$ is designed with multiple components, each of which corresponds to different functions. + +In the first sentence, it provides general guidance to relate the current subproblem to the entire refinement problem, to better condition the subproblem answer. Afterwards, it will include detailed descriptions of the task: identifying the physical rule and main object in Step 1, and identifying the semantic mismatch between the user prompt and caption of the generated video (generated by the video captioning model) in Step 2. It will also contain the requirements about the expected information in LLM's output. For example, in Step 1, the LLM's output about the physical rule should be in a descriptive way without giving formulas. + +Besides, to ensure proper CoT reasoning, we follow the existing work [22, 40] and provide in-context examples $[E]$ about tasks. To facilitate LLM's in-context learning [5, 6], $[E]$ is given in the format of QA pairs. That is, instead of fine-tuning a separate LLM checkpoint for each new task, we prompt the LLM with a few input-output exemplars, to demonstrate the task and condition the task's input-output format to the LLM, to guide the LLM's reasoning process. + +Then, the final part of the prompt, denoted as $[T]$ , is the information of the current instance of the task, usually with the current user prompt $(P_{i})$ being embedded. As a common practice of CoT reasoning, it also contains the hand-crafted trigger phrase $(t)$ , "Let's think step by step", to activate the local CoT reasoning in LLM. + +# 3.2. Prompting in the Final Subproblem for Global Step-Back Reasoning + +In the final subproblem, we enforce global step-back reasoning, by using the outputs of the two parallel subproblems above, i.e., knowledge about the physical rules and the prompt-video mismatch, as the high-level concepts and facts. Grounded on such high-level abstractions, we can make sure to improve the LLM's ability in following the correct reasoning path of generating the refined prompt. + +Being similar to the prompts used in the two parallel subproblems above, the prompt structure in the final subproblem also contains $[I]$ , $[E]$ and $[T]$ . Furthermore, to ensure the correct reasoning path, we also provide quantitative feedback to the LLM about the effectiveness of previous round's prompt refinement. Such effectiveness could be measured by the existing T2V evaluators, which judge the semantic alignment and quality of physical common sense of the currently generated video2. For example, the VideoCon-Physics evaluator [2] gives a score $([S])$ between 0 and 1. If $[S]$ is $< 0.5$ , it indicates that the refined prompt produced in the previous round is ineffective, hence guiding the LLM to take another alternative reasoning path. + +Since the prompt in the final subproblem is rich with reasoning and inherently very long-tailed, we removed the trigger prompt $[t]$ , to prevent incorporating the information in the final answer unrelated to the user's initial input prompt. + +# 3.3. The Stopping Condition + +The process of iterative refinement normally continues until the quality of the generated video is satisfactory, measured by the T2V evaluator as described above. Furthermore, the current T2V models naturally have limitations in generating some complicated or subtle scenes. In these cases, it would be difficult, even for PhyT2V, to reach physical realism after multiple rounds of iterations, and PhyT2V's refinement would stop when the iterations converge, i.e., the improvement of video quality becomes little over rounds. + +# 4. Experiments + +Models & Datasets: We applied PhyT2V on several DiT-based open-source T2V models, as listed below, and evaluated how PhyT2V improves the generated videos' adherence to real-world knowledge and physical rules. We use ChatGPT4 o1-preview [18] as the LLM for reasoning, and Tarsier [39] as the video captioning model. All generated videos last 6 seconds with 10 FPS and resolution of $720 \times 480$ . Details of evaluation setup are in Appendix A. + +- CogVideoX [51]: It generates 10-second videos from text prompts, with 16 FPS and $768 \times 1360$ resolution. It offers two model variants with 2B and 5B parameters. +- OpenSora 1.2 [53]: As an alternative to OpenAI's Sora [3], it contains 1.1B parameters and produces videos with 16 seconds, 720p resolution and different aspect ratios. +- VideoCrafter [4]: With 1.8B parameters, it can generate both images and videos from text prompts, with $576 \times 1024$ resolution and a focus on video dynamics. + +Since we target enhancing the T2V models' capability of generating physics-grounded video contents, we use the following two prompt benchmarks that emphasize physical laws and adherence as the text prompts for T2V: + +- VideoPhy [2] is designed to assess whether the generated videos follow physical common sense for real-world activities. It consists 688 human-verified captions that describe interactions between various types of real-world objects, including solid and fluid. +- PhyGenBench [29], similarly, allows evaluating the correctness of following physical common sense in T2V generation. It comprises 160 carefully crafted prompts spanning four physical domains, namely mechanics, optics, thermal and material properties. Since the domain of material properties has been covered by VideoPhy, we use the first three domains listed above. + +Evaluation metric: We use VideoCon-Physics evaluator provided with VideoPhy [2], to measure how the generated video adheres to physical common sense (PC) and achieves semantic adherence (SA). The PC metric evaluates whether the depicted actions and object's state follow the real-world physics laws. The SA metric measures if the actions, events, entities and their interactions specified in the prompt are present. Both metrics yield binary outputs: 1 indicates adherence and 0 indicates otherwise. On each T2V model and + +
Round1234
CogVideoX-5B [51]PC0.260.320.390.42
SA0.480.520.560.59
CogVideoX-2B [51]PC0.130.190.270.29
SA0.220.120.400.42
OpenSora [53]PC0.170.290.270.31
SA0.290.380.440.47
VideoCrafter [4]PC0.150.250.290.33
SA0.240.380.440.49
+ +Table 1. The quality of videos generated by different T2V models using the VideoPhy prompt dataset, over multiple rounds of iterative refinement in PhyT2V + +
Round1234
CogVideoX-5B [51]PC0.280.320.380.42
SA0.220.350.360.38
CogVideoX-2B [51]PC0.160.190.240.27
SA0.150.290.330.35
OpenSora [53]PC0.210.250.240.26
SA0.230.280.290.30
VideoCrafter [4]PC0.200.240.320.36
SA0.270.330.370.42
+ +Table 2. The quality of videos generated by different T2V models using the PhyGenBench prompt dataset, over multiple rounds of iterative refinement in PhyT2V + +dataset, the binary outputs from all prompts are averaged. + +In addition, we also evaluated PhyT2V using the widely used VBenchmark metrics and benchmarks [15], which allow comprehensive evaluations of the generated video in multiple aspects, including video quality, video-condition consistency, prompt following and human preference alignment. + +Baselines: For fair comparison, we only use the existing T2V prompt enhancers as baselines, and other existing work with extra inputs to T2V models [7, 19, 26, 27, 31] are not applicable. We involve two prompt enhancers: 1) Directly using the existing LLM, particularly ChatGPT4, as the prompt enhancer [28, 47]; 2) Promptist [21], which uses reinforcement learning to automatically refine and enhance prompts in the model-preferred way. + +# 4.1. Improvement of the Generated Video Quality + +As shown in Table 1 and 2, when PhyT2V is applied to different T2V models, it can significantly improve the generated video's adherence to both the text prompt itself and the real-world physical rules, compared to the videos generated by vanilla T2V models (i.e., in Round 1 of PhyT2V's refinement). In particular, such improvement is the most significant on the CogVideoX-2B model, where PC improvement can be up to $2.2\mathrm{x}$ and SA improvement can be up to $2.3\mathrm{x}$ . On all the other models, PhyT2V can also reach noticeable improvement, ranging from $1.3\mathrm{x}$ to $1.9\mathrm{x}$ . + +Meanwhile, results in Table 1 and 2 showed that + +
CogVideoX-5BOpenSora
ChatGPT 4 [28]PC0.330.21
SA0.410.32
Promptist [21]PC0.250.19
SA0.390.33
+ +Table 3. The quality of videos generated by enhancing the prompts in the VideoPhy dataset using different prompt enhancers + +
CogVideoX-5BOpenSora
ChatGPT 4 [28]PC0.270.20
SA0.230.23
Promptist [21]PC0.320.19
SA0.240.21
+ +Table 4. The quality of videos generated by enhancing the prompts in the PhyGenBench dataset using different prompt enhancers + +PhyT2V's process of iterative refinement converge quickly and only takes few rounds. Most improvement of video quality happens in the first two rounds, and little improvement can be observed in the fourth round. Hence, in practice, we believe that 3-4 iterative rounds would be sufficient. + +Furthermore, as shown in Table 3 and 4, PhyT2V also largely outperforms the existing prompt enhancers by at least $35\%$ , when being applied to CogVideoX-5B and OpenSora models. In particular, ChatGPT 4, when being used as the prompt enhancer, delivers better performance than Promptist due to its stronger language processing capabilities, but still cannot ensure physics-grounded T2V, due to the lack of explicit reasoning on text-to-video alignment. + +Our evaluation results on the VBenchmark metrics are shown in Figure 7, where numbers in Round 1 are the T2V model's original performance in current VBenchmark leaderboard, and iterative prompt refinements by PhyT2V in Round 2 & 3 noticeably improve the performance in many dimensions. In particular, large improvements are noted in most dimensions of Video-Condition Consistency, showing that PhyT2V improves T2V model's adherence to prompts and real-world physical rules underlying the prompts. + +![](images/37805667bedfd49b946dedca9679a1e972d7c9248a9f8d30d4ca8f25f556908a.jpg) +Figure 7. PhyT2V VBenchmark evaluation results with CogVideoX-5B (left) and OpenSora (right) + +# 4.2. Different Domains of Physical Rules + +We also conducted in-depth analysis on PhyT2V's performance on improving the generated video's quality in different domains of real-world physical rules, using the + +![](images/35c8323bd586c74d247bb4e8ff65310a069fb50c8fc7dfa38dac39989d1436d2.jpg) + +![](images/623e59ce559975c0b379d113d4a4e4feea6eebf080f5960abc0157c0b52d4e02.jpg) +Figure 8. Examples of videos generated using different categories of prompts in the PhyGenBench dataset +Figure 9. Examples of videos generated using different categories of prompts in the VideoPhy dataset + +CogVideoX-5B as the T2V model and ChatGPT 4 as the prompt enhancer. As shown in Table 5 and 6, PhyT2V achieves large improvements in most domains of physical rules. Especially in domains where physical interaction between objects are more subtle and difficult to be precisely captured, such as interaction with fluids and thermal-related scene changes, such improvements will be even higher. + +These improvements are also exemplified with sample videos and their related input prompts in Figure 9 and Figure 8. With LLM reasoning and iterative refinement, PhyT2V can largely enhance the T2V model's capability when encountering out-of-distribution prompts, by providing correct and sufficient contexts to ensure that the T2V model's video generation correctly captures the key objects and interaction between objects. For example, when the prompt of "juice dropping from a bottle onto the counter", PhyT2V correctly depicts the juice's slow diffusion on the counter. More examples can be found in Appendix B. + +# 4.3. Ablation Study + +We conduct an ablation study to evaluate the necessity of both the physical rule reasoning (Step 1) and the mismatch reasoning (Step 2) within our PhyT2V workflow, by removing one of these steps from the refinement process to assess + +
CogVideoX-5BCogVideoX-2BOpenSoraVideoCrafter
Round1234123412341234
Solid-SolidPC0.210.280.340.320.090.130.140.220.120.270.290.300.190.220.270.28
SA0.240.480.490.470.180.250.360.330.160.340.370.350.240.400.450.47
Solid-FluidPC0.220.270.280.300.110.180.280.270.170.210.240.250.180.240.250.26
SA0.390.540.600.610.290.430.440.430.160.400.410.360.340.430.480.52
Fluid-FluidPC0.570.590.630.620.340.380.350.360.150.320.290.310.330.410.530.51
SA0.410.570.590.670.270.420.390.440.310.440.450.460.320.420.490.51
+ +Table 5. The improvement of generated video quality in different categories of physical rules in the VideoPhy prompt dataset + +
CogVideoX-5BCogVideoX-2BOpenSoraVideoCrafter
Round1234123412341234
MechanicsPC0.190.250.340.350.120.160.180.240.110.130.170.220.140.230.290.28
SA0.210.280.290.320.110.180.190.220.190.210.270.320.200.240.280.35
OpticsPC0.220.350.410.390.220.250.290.280.240.260.250.250.220.210.270.32
SA0.270.420.390.440.230.340.370.390.260.310.290.300.220.280.350.39
ThermalPC0.330.350.350.350.130.150.150.140.270.300.310.330.250.280.260.28
SA0.220.360.430.450.120.160.240.270.230.250.370.360.250.370.410.43
+ +Table 6. The improvement of generated video quality in different categories of physical rules in the PhyGenBench prompt dataset + +its impact on physical-grounded video generation. + +Physical rule reasoning (Step 1). As shown in Figure 10, the Step 1 of physical rule reasoning significantly enhances the T2V process by providing a more detailed and coherent description of the principal object's physical status, such as motion, states and deformation (red texts in Figure 10), all grounded in relevant physical laws. By anchoring the prompt in established physical rules, this step also helps avoid unnecessary texts (brown texts in Figure 10) and vague physical rule descriptions (purple texts in Figure 10), hence achieving a higher PC score. + +![](images/184427b9654900e900410a525a756c99bc3a84987a626ce9ba2da482680bac76.jpg) +Figure 10. Ablation study on Step 1 of physical rule reasoning + +Mismatch reasoning (Step 2). The Step 2 of mismatch reasoning addresses details that may have been overlooked in the previous iteration of the generated video as shown in Figure 11. This step plays a critical role in the iterative refinement process by identifying and correcting discrepancies between expected and observed outputs. By enhancing the model's focus on the principal object, the mismatch reasoning step reduces the likelihood of losing attention to + +![](images/d346e6757ccf1d5eab11a5d1706259a20cc8e7a9c29c83a128cc0ec8588cd184.jpg) +Figure 11. Ablation study on Step 2 of mismatch reasoning + +important features (brown and purple texts in Figure 11), improving the fidelity and relevance of generated video content (red texts in Figure 11) towards a higher SA score. + +Overall, our study shows that both reasoning steps are integral to PhyT2V, contributing to a more robust and semantically-aligned generation of refined prompts in Step 3. More detailed ablation studies are in Appendix C. + +# 5. Conclusion + +In this paper, we present PhyT2V, a novel data-independent T2V generation framework designed to enhance the generalization capability of existing T2V models to out-of-distribution domains. By incorporating CoT reasoning and step-back prompting, PhyT2V systematically refines T2V prompts to ensure adherence to real-world physical principles without necessitating additional model retraining or reliance on additional conditions. Evaluation results indicate that PhyT2V achieves a $2.3\mathrm{x}$ enhancement in physical realism compared to baseline T2V models and outperforms state-of-the-art T2V prompt enhancers by $35\%$ . + +# Acknowledgments + +We thank the reviewers and the area chair for their insightful comments and feedback. This work was supported in part by National Science Foundation (NSF) under grant number IIS-2205360, CCF-2217003, CCF-2215042, and National Institutes of Health (NIH) under grant number R01HL170368. + +# References + +[1] Luke Auburn. Ai video generation expert discusses the technology's rapid advances—and its current limitations, 2024. 1 +[2] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, KaiWei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 5, 6 +[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 1, 6 +[4] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3, 6, 11 +[5] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5 +[6] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, et al. A survey on in-context learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1107-1128, 2024. 5 +[7] Theodoros Dumas and Alexandros Sigalas. Blender, an open source design tool: Advances and integration in the architectural production pipeline. Aristoteleio University of Thessaloniki, 21:737-744, 2009. 2, 6 +[8] Blender Foundation. Upbge: an open-source, 3d game engine forked from the old blender game engine, 2024. 2 +[9] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models, 2023. 2 +[10] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In European Conference on Computer Vision, pages 393–411. Springer, 2025. 2 +[11] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 3 +[12] Kai Huang, Haoming Wang, and Wei Gao. Freezeasguard: Mitigating illegal adaptation of diffusion models via selec + +tive tensor freezing. arXiv preprint arXiv:2405.17472, 2024. 2 +[13] Kai Huang, Hanyun Yin, Heng Huang, and Wei Gao. Towards green ai in fine-tuning large language models via adaptive backpropagation. ICLR, 2024. 3 +[14] Kai Huang, Xiangyu Yin, Heng Huang, and Wei Gao. Modality plug-and-play: Runtime modality adaptation in LLM-driven autonomous mobile systems. In ACM MobiCom, 2025. 3 +[15] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6 +[16] Epic Games Inc. Unreal engine: The most powerful real-time 3d creation tool, 2024. 2 +[17] Mellis Inc. Pika labs, 2023. 1 +[18] OpenAI Inc. Introducing openai o1-preview, 2024. 2, 6 +[19] Marcel Krüger, David Gilbert, Torsten W. Kuhlen, and Tim Gerrits. Game engines for immersive visualization: Using unreal engine beyond entertainment. PRESENCE: Virtual and Augmented Reality, 33:31-55, 2024. 2, 6 +[20] Chengxuan Li, Di Huang, Zeyu Lu, Yang Xiao, Qingqi Pei, and Lei Bai. A survey on long video generation: Challenges, methods, and prospects, 2024. 2 +[21] WeiJie Li, Jin Wang, and Xuejie Zhang. Promptist: Automated prompt optimization for text-to-image synthesis. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 295-306. Springer, 2024. 6, 7, 11 +[22] Yingcong Li, Kartik Sreenivasan, Angeliki Giannou, Dimitris Papailiopoulos, and Samet Oymak. Dissecting chain-of-thought: Compositionality through in-context filtering and learning. Advances in Neural Information Processing Systems, 36, 2024. 5 +[23] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 3 +[24] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. +[25] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 3 +[26] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imaged-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 2, 6 +[27] Jiaxi Lv, Yi Huang, Mingfu Yan, Jiancheng Huang, Jianzhuang Liu, Yifan Liu, Yafei Wen, Xiaoxin Chen, and Shifeng Chen. Gpt4: Scripting physical motions in text-to-video generation via blender-oriented gpt planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1430-1440, 2024. 6 + +[28] Jiaxi Lv, Yi Huang, Mingfu Yan, Jiancheng Huang, Jianzhuang Liu, Yifan Liu, Yafei Wen, Xiaoxin Chen, and Shifeng Chen. Gpt4motion: Scripting physical motions in text-to-video generation via blender-oriented gpt planning, 2024. 2, 6, 7 +[29] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363, 2024. 6 +[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3 +[31] Weichao Qiu and Alan Yuille. Unrealcv: Connecting computer vision to unreal engine. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 909-916. Springer, 2016. 2, 6 +[32] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications, 2024. 2 +[33] Sander Schulhoff, Michael Ilie, Nishant Balepur, Konstantine Kahadze, Amanda Liu, Chenglei Si, Yinheng Li, Aayush Gupta, HyoJung Han, Sevien Schulhoff, Pranav Sandeep Dulepet, Saurav Vidyadhara, Dayeon Ki, Sweta Agrawal, Chau Pham, Gerson Kroiz, Feileen Li, Hudson Tao, Ashay Srivastava, Hevander Da Costa, Saloni Gupta, Megan L. Rogers, Inna Goncearenco, Giuseppe Sarli, Igor Galynker, Denis Peskoff, Marine Carpuat, Jules White, Shyamal Anadkat, Alexander Hoyle, and Philip Resnik. The prompt report: A systematic survey of prompting techniques, 2024. 2 +[34] Sander Schulhoff, Michael Ilie, Nishant Balepur, Konstantine Kahadze, Amanda Liu, Chenglei Si, Yinheng Li, Aayush Gupta, HyoJung Han, Sevien Schulhoff, et al. The prompt report: A systematic survey of prompting techniques. arXiv preprint arXiv:2406.06608, 2024. 5 +[35] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[36] Unity Technologies. Unity real-time development platform, 2024. 2 +[37] Stuart A. Thompson. A.i. can now create lifelike videos. can you tell what's real?, 2024. 1 +[38] Boshi Wang, Sewon Min, Xiang Deng, Jiaming Shen, You Wu, Luke Zettlemoyer, and Huan Sun. Towards understanding chain-of-thought prompting: An empirical study of what matters. arXiv preprint arXiv:2212.10001, 2022. 4 +[39] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 2, 6, 11 +[40] Keheng Wang, Feiyu Duan, Sirui Wang, Peiguang Li, Yun-sen Xian, Chuantao Yin, Wenge Rong, and Zhang Xiong. Knowledge-driven cot: Exploring faithful reasoning in llms + +for knowledge-intensive question answering. arXiv preprint arXiv:2308.13259, 2023. 5 +[41] Xiaofeng Wang, Zheng Zhu, Guan Huang, Boyuan Wang, Xinze Chen, and Jiwen Lu. Worlddreamer: Towards general world models for video generation via predicting masked tokens, 2024. 2 +[42] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3 +[43] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 3 +[44] Zhendong Wang, Yifan Jiang, Yadong Lu, Pengcheng He, Weizhu Chen, Zhangyang Wang, Mingyuan Zhou, et al. Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems, 36:8542-8562, 2023. 3 +[45] Zhao Wang, Aoxue Li, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. Customvideo: Customizing text-to-video generation with multiple subjects. arXiv preprint arXiv:2401.09962, 2024. 3 +[46] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 3 +[47] Deshun Yang, Luhui Hu, Yu Tian, Zihao Li, Chris Kelly, Bang Yang, Cindy Yang, and Yuexian Zou. Worldgpt: a sora-inspired video ai agent as rich world models from text and image inputs. arXiv preprint arXiv:2403.07944, 2024. 6, 11 +[48] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In *Forty-first International Conference on Machine Learning*, 2024. 3 +[49] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 2023. 2 +[50] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer, 2024. 1, 3, 11 +[51] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6 +[52] Wentao Zhang, Junliang Guo, Tianyu He, Li Zhao, Linli Xu, and Jiang Bian. Video in-context learning. arXiv preprint arXiv:2407.07356, 2024. 3 + +[53] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 6, 11 +[54] Hanxin Zhu, Tianyu He, Anni Tang, Junliang Guo, Zhibo Chen, and Jiang Bian. Compositional 3d-aware video generation with llm director. arXiv preprint arXiv:2409.00558, 2024. 3 \ No newline at end of file diff --git a/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/images.zip b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..061966dc2b35d2df0c8197694cfa856ecd20b417 --- /dev/null +++ b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36534c331e4222c431ae2995c8656d0aacbec753ff8468d5444d77d6910b5931 +size 1040445 diff --git a/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/layout.json b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a5037b1c46799bc8b21690eeaef20c036b5783e9 --- /dev/null +++ b/2025/PhyT2V_ LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation/layout.json @@ -0,0 +1,7670 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 87, + 524, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 87, + 524, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 87, + 524, + 121 + ], + "type": "text", + "content": "PhyT2V: LLM-Guided Iterative Self-Refinement for Physics-Grounded Text-to-Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 137, + 434, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 137, + 434, + 164 + ], + "spans": [ + { + "bbox": [ + 176, + 137, + 434, + 164 + ], + "type": "text", + "content": "Qiyao Xue, Xiangyu Yin, Boyuan Yang and Wei Gao \nUniversity of Pittsburgh" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 188, + 166, + 422, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 166, + 422, + 178 + ], + "spans": [ + { + "bbox": [ + 188, + 166, + 422, + 178 + ], + "type": "text", + "content": "{qix63, eric.yin, by.yang, weigao}@pitt.edu" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 198, + 549, + 473 + ], + "blocks": [ + { + "bbox": [ + 56, + 198, + 549, + 473 + ], + "lines": [ + { + "bbox": [ + 56, + 198, + 549, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 549, + 473 + ], + "type": "image", + "image_path": "8113c8c3ee5a7b963c2ebb52360142362a68b3e5e05009dfad7a0d2b66fac69b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 478, + 555, + 513 + ], + "lines": [ + { + "bbox": [ + 54, + 478, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 54, + 478, + 555, + 513 + ], + "type": "text", + "content": "Figure 1. Left: videos generated by the current text-to-video generation model (CogVideoX-5B [50]) cannot adhere to the real-world physical rules (described in brackets following the user prompt). Right: our method PhyT2V, when applied to the same model, better reflects the real-world physical knowledge." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 523, + 200, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 523, + 200, + 534 + ], + "spans": [ + { + "bbox": [ + 151, + 523, + 200, + 534 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 537, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 537, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 537, + 297, + 715 + ], + "type": "text", + "content": "Text-to-video (T2V) generation has been recently enabled by transformer-based diffusion models, but current T2V models lack capabilities in adhering to the real-world common knowledge and physical rules, due to their limited understanding of physical realism and deficiency in temporal modeling. Existing solutions are either data-driven or require extra model inputs, but cannot be generalizable to out-of-distribution domains. In this paper, we present PhyT2V, a new data-independent T2V technique that expands the current T2V model's capability of video generation to out-of-distribution domains, by enabling chain-of-thought and step-back reasoning in T2V prompting. Our experiments show that PhyT2V improves existing T2V models' adherence to real-world physical rules by " + }, + { + "bbox": [ + 54, + 537, + 297, + 715 + ], + "type": "inline_equation", + "content": "2.3x" + }, + { + "bbox": [ + 54, + 537, + 297, + 715 + ], + "type": "text", + "content": ", and achieves " + }, + { + "bbox": [ + 54, + 537, + 297, + 715 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 54, + 537, + 297, + 715 + ], + "type": "text", + "content": " improvement compared to T2V prompt enhancers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 523, + 394, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 523, + 394, + 534 + ], + "spans": [ + { + "bbox": [ + 314, + 523, + 394, + 534 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 536, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 536, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 312, + 536, + 555, + 643 + ], + "type": "text", + "content": "Text-to-video (T2V) generation has recently marked a significant breakthrough of generative AI, with the advent of transformer-based diffusion models such as Sora [3], Pika [17] and CogVideoX [51] that can produce videos conditioned on textual prompts. These models demonstrate astonishing capabilities of generating complex and photorealistic scenes, and could even make it difficult for humans to distinguish between real-world and AI-generated videos, in the aspect of individual video frames' quality [1, 37]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 643, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 643, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 643, + 556, + 713 + ], + "type": "text", + "content": "On the other hand, as shown in Figure 1 - Left, current T2V models still have significant drawbacks in adhering to the real-world common knowledge and physical rules, such as quantity, material, fluid dynamics, gravity, motion, collision and causality, and such limitations fundamentally prevent current T2V models from being used for real-world" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18826" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 201 + ], + "type": "text", + "content": "simulation [7, 19, 31]. Enforcement of real-world knowledge and physical rules in T2V generation, however, is challenging because it requires the models' understandings of not only individual objects but also how these objects move and interact with each other. Further, unlike generating static images, T2V generation requires frame-to-frame consistency in object appearance, shape, motion, lighting and other dynamics [11]. Current T2V models often lack such temporal modeling, especially over long sequences [20], and the generated videos often contain flickering, inconsistent motion and object deformations across frames [26]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 204, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 204, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 295, + 415 + ], + "type": "text", + "content": "Most of the existing solutions to these challenges are data-driven, by using large multimodal T2V datasets that cover different real-world domains to train the diffusion model [10, 12, 41, 49]. However, these solutions heavily rely on the volume, quality and diversity of datasets [42, 51]. Since real-world common knowledge and physical rules are not explicitly embedded in the T2V generation process, the quality of video generation would largely drop in out-of-distribution domains that are not covered by the training dataset, and the generalizability of T2V models is limited due to the vast diversity of real-world scenario domains. Alternatively, researchers also use the existing 3D engines (e.g., Blender [8], Unity3D [36] and Unreal [16]) or mathematical models of edge and depth maps [26-28] to inject real-world physical knowledge into the T2V model, but these approaches are limited to fixed physical categories and patterns such as predefined objects and movements [26, 49], similarly lacking generalizability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 418, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 582 + ], + "type": "text", + "content": "To achieve generalizable enforcement of physics-grounded T2V generation, we propose a fundamentally different approach: instead of expanding the training dataset or further complicating the T2V model architecture, we aim to expand the current T2V model's capability of video generation from in-distribution to out-of-distribution domains, by embedding real-world knowledge and physical rules into the text prompts with sufficient and appropriate contexts. To avoid ambiguous and unexplainable prompt engineering [9, 32, 33], our basic idea is to enable chain-of-thought (CoT) and step-back reasoning in T2V generation prompting, to ensure that T2V models follow correct physical dynamics and inter-frame consistency by applying step-by-step guidance and iterative refinement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 585, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 295, + 714 + ], + "type": "text", + "content": "Based on this idea, this paper presents Physcial-grounded Text-to-Video (PhyT2V), a new T2V technique that harnesses the natural language reasoning capabilities of well-trained LLMs (e.g., ChatGPT-4o), to facilitate CoT and step-back reasoning as described above. As shown in Figure 2, such reasoning is iteratively conducted in PhyT2V, and each iteration autonomously refines both the T2V prompt and generated video in three steps. In Step 1, the LLM analyzes the T2V prompt to extract objects to be shown and physical rules to follow in the video via in-context learning. In Step 2, we first use a video captioning model to translate" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 187 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 187 + ], + "type": "image", + "image_path": "36aad3ee25808e3108664a0a95864a636edf3867f4adeedb7f860368e96007ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 193, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 555, + 226 + ], + "type": "text", + "content": "Figure 2. One iteration of video and prompt self-refinement in PhyT2V. Such self-refinement will be iteratively conducted in multiple rounds until the quality of generated video is satisfactory." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 231, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 361 + ], + "type": "text", + "content": "the video's semantic contents into texts according to the list of objects obtained in Step 1, and then use the LLM to evaluate the mismatch between the video caption and current T2V prompt via CoT reasoning. In Step 3, the LLM refines the current T2V prompt, by incorporating the physical rules summarized in Step 1 and resolving the mismatch derived in Step 2, through step-back prompting. The refined T2V prompt is then used by the T2V model again for video generation, starting a new round of refinement. Such iterative refinement stops when the quality of generated video is satisfactory or the improvement of video quality converges." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 361, + 556, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 453 + ], + "type": "text", + "content": "For physical-grounded video generation performance, we further evaluated PhyT2V by applying it onto multiple SOTA T2V models, by using ChatGPT4 o1-preview [18] for LLM reasoning and Tarsier [39] as the video captioning model. We used two major T2V prompt datasets that cover 7 different real-world domains, and compared PhyT2V with the most competitive baselines of prompt enhancers. Our main findings are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 456, + 555, + 619 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 456, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 456, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 555, + 514 + ], + "type": "text", + "content": "- PhyT2V is highly effective. Without involving any model retraining efforts on any auxiliary model inputs, PhyT2V can improve the adherence of the existing T2V models' generated videos to real-world physical rules by up to " + }, + { + "bbox": [ + 314, + 456, + 555, + 514 + ], + "type": "inline_equation", + "content": "2.3\\mathrm{x}" + }, + { + "bbox": [ + 314, + 456, + 555, + 514 + ], + "type": "text", + "content": ", by only refining the text prompts to the T2V model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 514, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 583 + ], + "type": "text", + "content": "- PhyT2V is high generic. It can result in significant improvement of video quality in a large diversity of real-world domains, covering solid, liquid, mechanics, optics, thermal, etc. It is fully data independent, and its prompting templates can be applied to any existing T2V models with different architectures and input formats." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 585, + 553, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 553, + 619 + ], + "type": "text", + "content": "- Based on LLM-guided reasoning and self-refinement, PhyT2V is fully automated and involve the minimum amount of engineering and manual efforts." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 629, + 482, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 482, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 482, + 641 + ], + "type": "text", + "content": "2. Related Work and Motivation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 649, + 450, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 649, + 450, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 649, + 450, + 660 + ], + "type": "text", + "content": "2.1. T2V Generation Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "Early T2V techniques generate video frames from text-to-image model outputs with temporal extensions [35], but cannot maintain temporal consistency and coherence over time, often producing visually appealing but temporally dis" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18827" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 178 + ], + "type": "text", + "content": "connected outputs. Diffusion Transformers (DiT) [30] improved such consistency with a transformer backbone capable of capturing more complex temporal dynamics and relationships across frames through attention mechanism and long-range dependency modeling [42, 51]. Based on the DiT architecture, recent T2V models, such as OpenSora [53] and VideoCrafter [4], demonstrated that T2V generation can be further improved by in-context learning when provided with sufficient contextual information [44]." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 182, + 294, + 310 + ], + "blocks": [ + { + "bbox": [ + 57, + 182, + 294, + 310 + ], + "lines": [ + { + "bbox": [ + 57, + 182, + 294, + 310 + ], + "spans": [ + { + "bbox": [ + 57, + 182, + 294, + 310 + ], + "type": "image", + "image_path": "af4d0ad4dec137b51df26c071e6ca57bdf56b2c75891f171a6ef9623aad6a3dc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 313, + 295, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 335 + ], + "type": "text", + "content": "Figure 3. Examples of videos generated from in-distribution and out-of-distribution prompts, using the CogVideoX-5B model" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 342, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 495 + ], + "type": "text", + "content": "However, as shown in Figure 3, although these T2V models demonstrate strong capabilities in video generation when dealing with prompts aligned with the distributions found in the training data, they encounter significant challenges with out-of-distribution prompts that are not covered by training data1. In such cases, the outputs often contain physical illusions or artifacts, reflecting the model's limitations in generating realistic and coherent video contents under unfamiliar conditions. Such limitations can be addressed by enlarging the training datasets, improving T2V model architectures or developing new mechanisms for adaptation and error correction [43, 45], but these approaches are all prompt-specific and lack generalizability." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 507, + 295, + 571 + ], + "blocks": [ + { + "bbox": [ + 57, + 507, + 295, + 571 + ], + "lines": [ + { + "bbox": [ + 57, + 507, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 57, + 507, + 295, + 571 + ], + "type": "image", + "image_path": "f053aeeb42b356aeecd1aa01478efece326b8b30f3502fb76732588c06d29b49.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 574, + 295, + 596 + ], + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 596 + ], + "type": "text", + "content": "Figure 4. A video generated by enhancing the out-of-distribution prompt \"Whisking egg into milk for scramble\" in Figure 3" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 603, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 296, + 673 + ], + "type": "text", + "content": "On the other hand, as shown in Figure 4, recent research has demonstrated that the quality of video generation with an out-of-distribution prompt can be improved by refining the prompt with sufficient and appropriate details [11, 51]. These findings motivate our design of PhyT2V that embeds contexts of real-world knowledge and physical rules into" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 237 + ], + "type": "text", + "content": "T2V prompts, to guide the T2V process for better physical accuracy and temporal alignment. The existing works, however, could still fail when tackling more intricate scenarios such as multi-object interactions, because the T2V model lacks an efficient feedback mechanism to learn how the generated video deviates from the real-world knowledge and physical rules. Researchers suggest to provide such feedback with extra input modalities to T2V models such as sampled video frames, depth map or scribble images [44, 52], but incur significant amounts of extra computing overhead and cannot be generalizable. Instead, in our design of PhyT2V, we aim to fully automate the feedback with only text prompts, and enable iterative feedback for the optimum video quality." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 242, + 545, + 406 + ], + "blocks": [ + { + "bbox": [ + 321, + 242, + 545, + 406 + ], + "lines": [ + { + "bbox": [ + 321, + 242, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 321, + 242, + 545, + 406 + ], + "type": "image", + "image_path": "d97ec0eb02e0075ebca8bde89d0348e0d0f9fe2cfc5837ed20b412d5b2b9978c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 409, + 529, + 421 + ], + "lines": [ + { + "bbox": [ + 338, + 409, + 529, + 421 + ], + "spans": [ + { + "bbox": [ + 338, + 409, + 529, + 421 + ], + "type": "text", + "content": "Figure 5. Examples of CoT and step-back reasoning" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 430, + 484, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 484, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 484, + 442 + ], + "type": "text", + "content": "2.2. Using LLM in T2V Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 444, + 555, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 619 + ], + "type": "text", + "content": "LLMs with strong capabilities in natural language processing (NLP) have been a natural choice for prompt refinement in text-to-image and text-to-video generation, and existing work utilized LLMs to interpret text prompts and orchestrate the initial layout configurations [13, 14, 23-25, 46, 48, 54]. However, since current LLMs lack inherent understandings of the real-world physical laws, using LLMs with simple instructions usually result in videos that appear visually coherent but lack accurate physical realism, particularly when generating scenes with complex object interactions. Furthermore, these approaches frequently rely on static prompts or simple iterative refinements based on bounding box and segmentation map, which may capture basic visual attributes but fail to adapt to nuanced changes that require continuous physical modeling and adjustment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 620, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 556, + 715 + ], + "type": "text", + "content": "An effective approach to addressing these limitations and providing effective feedback for prompt refinement is to explicitly trigger in-context learning and reasoning in LLM. For example, as shown in Figure 5, CoT reasoning deconstructs complex prompts into stepwise logical tasks, and hence ensures a precise scheduling path to align generated content with the input prompt. However, CoT reasoning, in some cases, could make errors in some intermediate steps," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "1In Figure 3, the in-distribution prompts are picked from the ones listed in [50], and the out-of-distribution prompts are our crafted ones for similar scenarios as the in-distribution prompts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18828" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 95, + 71, + 284, + 81 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 71, + 284, + 81 + ], + "spans": [ + { + "bbox": [ + 95, + 71, + 284, + 81 + ], + "type": "text", + "content": "Step 1: Identifying the Physical Rule and Main Object" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 82, + 308, + 338 + ], + "blocks": [ + { + "bbox": [ + 67, + 82, + 308, + 338 + ], + "lines": [ + { + "bbox": [ + 67, + 82, + 308, + 338 + ], + "spans": [ + { + "bbox": [ + 67, + 82, + 308, + 338 + ], + "type": "image", + "image_path": "89f6e74f50e27e472e40e451539de5cb9d058b750440cf8fd26e92e8d4eaece4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 344, + 555, + 367 + ], + "lines": [ + { + "bbox": [ + 55, + 344, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 555, + 367 + ], + "type": "text", + "content": "Figure 6. Our design of PhyT2V, illustrated by one round of video refinement consisting of three steps. Texts in brown are inputs from previous step. Texts in red are outputs to the next step; Texts in purple are prompts to trigger LLM reasoning" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 380, + 71, + 517, + 80 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 71, + 517, + 80 + ], + "spans": [ + { + "bbox": [ + 380, + 71, + 517, + 80 + ], + "type": "text", + "content": "Step 3: Generating the Refined Prompt" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 363, + 81, + 421, + 89 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 81, + 421, + 89 + ], + "spans": [ + { + "bbox": [ + 363, + 81, + 421, + 89 + ], + "type": "text", + "content": "Task instruction" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 352, + 90, + 541, + 339 + ], + "blocks": [ + { + "bbox": [ + 352, + 90, + 541, + 339 + ], + "lines": [ + { + "bbox": [ + 352, + 90, + 541, + 339 + ], + "spans": [ + { + "bbox": [ + 352, + 90, + 541, + 339 + ], + "type": "image", + "image_path": "ed1642fa16d59bb50d36be50fcef2135f54234741e9fc8384f12f2ebf58833b9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 373, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 455 + ], + "type": "text", + "content": "and step-back prompting can address this limitation by further deriving the step-back question at a higher level of abstraction and hence avoiding confusions and vagueness. In our design of PhyT2V, we will utilize such LLM reasoning to analyze the inconsistency of the generated video to real-world common knowledge and physical rules, and use the reasoning outcome as feedback for T2V prompt refinement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 456, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 666 + ], + "type": "text", + "content": "The Chain-of-Thought (CoT) method is suitable for single data mode processing because it emphasizes linear decomposition and step-by-step reasoning, and it is especially effective for data processing flows that do not require cross-modal synchronization or interaction. However, multimodal data processing involves the fusion of data from different modalities and complex synchronization requirements, which cannot be completed through simple linear decomposition and requires frequent cross-modal information interaction and parallel processing, which exceeds the linear reasoning ability of the CoT method, resulting in limited performance in multimodal tasks. This is the reason why it is hard to directly apply CoT reasoning in T2V process itself as it requires multimodal alignment between the next and video modality. This also motivates us to adopt video captioning and use the video caption in the reasoning process, so that we can conduct CoT and step-back reasoning only in the text domain." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 673, + 111, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 111, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 111, + 685 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "In this section, we present details of our PhyT2V design. In principal, PhyT2V's refinement of T2V generation is an" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 373, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 467 + ], + "type": "text", + "content": "iterative process consisting of multiple rounds. In each round, as shown in Figure 6, the primary objective of our PhyT2V design is to guide a well-trained LLM (e.g., ChatGPT-4o) to generate a refined prompt that enables the pre-trained T2V model to generate videos that better match the given user prompt and real-world physical rules, and the refined prompt will be iteratively used as the new user prompt in the next round of refinement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 467, + 556, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 556, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 556, + 595 + ], + "type": "text", + "content": "Each round of refinement is structured around decomposing the complex refinement problem into a series of simpler subproblems, more specifically, two parallel subproblems and one final subproblem. The two parallel subproblems are: Step 1) identifying the relevant physical rules that the generated video should follow based on the user prompt, and Step 2) identifying semantic mismatches between the user prompt and the generated video. Based on the knowledge about physical rules and semantic mismatches, the final subproblem (Step 3) generates the refined prompt to better adhere to the physical rules and resolve the mismatches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 596, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 556, + 666 + ], + "type": "text", + "content": "To ensure proper identification in the parallel subproblems and prompt generation in the final subproblem, the core of PhyT2V design is two types of LLM reasoning processes within the prompt enhancement loop: the local CoT reasoning for individual subproblems and global step-back reasoning for the overall prompt refinement problem." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "Local CoT reasoning is executed within the prompt for each subproblem, to prompt the LLM to generate a detailed reasoning chain in its latent embedding space [38]. Addressing the parallel subproblems facilitates LLM with a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18829" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 154 + ], + "type": "text", + "content": "more concentrated attention on prerequisites of prompt refinement, enabling a deeper comprehension of the physical laws that govern the video content as well as the identification of discrepancies between the generated video and the user prompt. The outcomes derived from these parallel subproblems reflect the language model's abstraction in stepback reasoning on the overarching prompt refinement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 155, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 295, + 319 + ], + "type": "text", + "content": "Global step-back reasoning: To integrate various subproblems into a coherent framework for prompt and video refinement, one intuitive approach involves employing CoT reasoning across these subproblems, allowing the LLM to engage in self-questioning. However, this method may lead to the risk of traversing incorrect reasoning pathways. Instead, we apply global step-back reasoning across subproblems, by using a self-augmented prompt to incorporate the LLM-generated responses to high-level questions about physical rules and semantic mismatches in earlier parallel problems, when generating the refined prompt in the final subproblem. In this way, we can improve the correctness of intermediate reasoning steps in CoT reasoning, and enable consistent improvement across steps in reasoning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 294, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 294, + 388 + ], + "type": "text", + "content": "Both reasoning processes are facilitated through appropriate task instruction prompting tailored to different subproblems. In general, our prompting procedure follows the prompt modeling in [34], which divides task instructions into several components. More details about these components in our design of PhyT2V are elaborated as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": "Compared to the previous prompt enhancing methods, PhyT2V's key contribution is to analyze the semantic mismatch between currently generated video and the prompt, as well as refinements based on such mismatch. Previous methods can be formulated as " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "p' = f_{\\text{enhance}}(p, \\theta)" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": " are the original and enhanced prompts, " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "f_{\\text{enhance}}" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": " is the prompt enhancer, and " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": " represents parameters or rules guiding the enhancement. In contrast, PhyT2V further involves the additional information about the T2V process, i.e., " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "p' = f_{\\text{enhance}}(p, f_{\\text{mismatch}}(C(V(p)), p), f_{\\text{phy}}(p), \\theta)" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "f_{\\text{phy}}(p)" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": " analyzes the physical rules to be followed given " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "V(p)" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": " is the currently generated video given prompt " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": " is the video captioning model and " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "f_{\\text{mismatch}}" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": " finds the semantic mismatch between " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "C(V(p))" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": ". The key advantages are: (1) Semantic awareness: the refinement process explicitly incorporates the semantic mismatch to enable targeted T2V improvements; (2) Physical-world knowledge integration: physical rules derived from " + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 388, + 295, + 635 + ], + "type": "text", + "content": " enable guided enhancement; (3) Guided reasoning: unlike prior methods that rely solely on templates or embeddings, PhyT2V dynamically adapts prompt refinement to the semantic mismatch." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 640, + 295, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 295, + 666 + ], + "type": "text", + "content": "3.1. Prompting in Parallel Subproblems for Local CoT Reasoning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": "In both Step 1 and Step 2, the first part of prompt is a task instruction prompt " + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "inline_equation", + "content": "[I]" + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": " to instruct the LLM to understand the task in the subproblem. " + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "inline_equation", + "content": "[I]" + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": " is designed with multiple components, each of which corresponds to different functions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 201 + ], + "type": "text", + "content": "In the first sentence, it provides general guidance to relate the current subproblem to the entire refinement problem, to better condition the subproblem answer. Afterwards, it will include detailed descriptions of the task: identifying the physical rule and main object in Step 1, and identifying the semantic mismatch between the user prompt and caption of the generated video (generated by the video captioning model) in Step 2. It will also contain the requirements about the expected information in LLM's output. For example, in Step 1, the LLM's output about the physical rule should be in a descriptive way without giving formulas." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 202, + 553, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 553, + 296 + ], + "type": "text", + "content": "Besides, to ensure proper CoT reasoning, we follow the existing work [22, 40] and provide in-context examples " + }, + { + "bbox": [ + 313, + 202, + 553, + 296 + ], + "type": "inline_equation", + "content": "[E]" + }, + { + "bbox": [ + 313, + 202, + 553, + 296 + ], + "type": "text", + "content": " about tasks. To facilitate LLM's in-context learning [5, 6], " + }, + { + "bbox": [ + 313, + 202, + 553, + 296 + ], + "type": "inline_equation", + "content": "[E]" + }, + { + "bbox": [ + 313, + 202, + 553, + 296 + ], + "type": "text", + "content": " is given in the format of QA pairs. That is, instead of fine-tuning a separate LLM checkpoint for each new task, we prompt the LLM with a few input-output exemplars, to demonstrate the task and condition the task's input-output format to the LLM, to guide the LLM's reasoning process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 296, + 553, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 553, + 367 + ], + "type": "text", + "content": "Then, the final part of the prompt, denoted as " + }, + { + "bbox": [ + 313, + 296, + 553, + 367 + ], + "type": "inline_equation", + "content": "[T]" + }, + { + "bbox": [ + 313, + 296, + 553, + 367 + ], + "type": "text", + "content": ", is the information of the current instance of the task, usually with the current user prompt " + }, + { + "bbox": [ + 313, + 296, + 553, + 367 + ], + "type": "inline_equation", + "content": "(P_{i})" + }, + { + "bbox": [ + 313, + 296, + 553, + 367 + ], + "type": "text", + "content": " being embedded. As a common practice of CoT reasoning, it also contains the hand-crafted trigger phrase " + }, + { + "bbox": [ + 313, + 296, + 553, + 367 + ], + "type": "inline_equation", + "content": "(t)" + }, + { + "bbox": [ + 313, + 296, + 553, + 367 + ], + "type": "text", + "content": ", \"Let's think step by step\", to activate the local CoT reasoning in LLM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 375, + 553, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 553, + 399 + ], + "type": "text", + "content": "3.2. Prompting in the Final Subproblem for Global Step-Back Reasoning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 401, + 553, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 483 + ], + "type": "text", + "content": "In the final subproblem, we enforce global step-back reasoning, by using the outputs of the two parallel subproblems above, i.e., knowledge about the physical rules and the prompt-video mismatch, as the high-level concepts and facts. Grounded on such high-level abstractions, we can make sure to improve the LLM's ability in following the correct reasoning path of generating the refined prompt." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "text", + "content": "Being similar to the prompts used in the two parallel subproblems above, the prompt structure in the final subproblem also contains " + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "inline_equation", + "content": "[I]" + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "inline_equation", + "content": "[E]" + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "inline_equation", + "content": "[T]" + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "text", + "content": ". Furthermore, to ensure the correct reasoning path, we also provide quantitative feedback to the LLM about the effectiveness of previous round's prompt refinement. Such effectiveness could be measured by the existing T2V evaluators, which judge the semantic alignment and quality of physical common sense of the currently generated video2. For example, the VideoCon-Physics evaluator [2] gives a score " + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "inline_equation", + "content": "([S])" + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "text", + "content": " between 0 and 1. If " + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "inline_equation", + "content": "[S]" + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "inline_equation", + "content": "< 0.5" + }, + { + "bbox": [ + 313, + 484, + 553, + 635 + ], + "type": "text", + "content": ", it indicates that the refined prompt produced in the previous round is ineffective, hence guiding the LLM to take another alternative reasoning path." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 637, + 553, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 553, + 685 + ], + "type": "text", + "content": "Since the prompt in the final subproblem is rich with reasoning and inherently very long-tailed, we removed the trigger prompt " + }, + { + "bbox": [ + 313, + 637, + 553, + 685 + ], + "type": "inline_equation", + "content": "[t]" + }, + { + "bbox": [ + 313, + 637, + 553, + 685 + ], + "type": "text", + "content": ", to prevent incorporating the information in the final answer unrelated to the user's initial input prompt." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "2This video is generated using the prompt refined in the previous round, and is also used to generate the video caption as the input in Step 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18830" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 190, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 190, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 190, + 85 + ], + "type": "text", + "content": "3.3. The Stopping Condition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 296, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 296, + 191 + ], + "type": "text", + "content": "The process of iterative refinement normally continues until the quality of the generated video is satisfactory, measured by the T2V evaluator as described above. Furthermore, the current T2V models naturally have limitations in generating some complicated or subtle scenes. In these cases, it would be difficult, even for PhyT2V, to reach physical realism after multiple rounds of iterations, and PhyT2V's refinement would stop when the iterations converge, i.e., the improvement of video quality becomes little over rounds." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 136, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 136, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 136, + 211 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 213, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 308 + ], + "type": "text", + "content": "Models & Datasets: We applied PhyT2V on several DiT-based open-source T2V models, as listed below, and evaluated how PhyT2V improves the generated videos' adherence to real-world knowledge and physical rules. We use ChatGPT4 o1-preview [18] as the LLM for reasoning, and Tarsier [39] as the video captioning model. All generated videos last 6 seconds with 10 FPS and resolution of " + }, + { + "bbox": [ + 55, + 213, + 296, + 308 + ], + "type": "inline_equation", + "content": "720 \\times 480" + }, + { + "bbox": [ + 55, + 213, + 296, + 308 + ], + "type": "text", + "content": ". Details of evaluation setup are in Appendix A." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 309, + 295, + 415 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 309, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 344 + ], + "type": "text", + "content": "- CogVideoX [51]: It generates 10-second videos from text prompts, with 16 FPS and " + }, + { + "bbox": [ + 56, + 309, + 295, + 344 + ], + "type": "inline_equation", + "content": "768 \\times 1360" + }, + { + "bbox": [ + 56, + 309, + 295, + 344 + ], + "type": "text", + "content": " resolution. It offers two model variants with 2B and 5B parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 345, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 345, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 345, + 295, + 380 + ], + "type": "text", + "content": "- OpenSora 1.2 [53]: As an alternative to OpenAI's Sora [3], it contains 1.1B parameters and produces videos with 16 seconds, 720p resolution and different aspect ratios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 380, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 380, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 380, + 295, + 415 + ], + "type": "text", + "content": "- VideoCrafter [4]: With 1.8B parameters, it can generate both images and videos from text prompts, with " + }, + { + "bbox": [ + 56, + 380, + 295, + 415 + ], + "type": "inline_equation", + "content": "576 \\times 1024" + }, + { + "bbox": [ + 56, + 380, + 295, + 415 + ], + "type": "text", + "content": " resolution and a focus on video dynamics." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "content": "Since we target enhancing the T2V models' capability of generating physics-grounded video contents, we use the following two prompt benchmarks that emphasize physical laws and adherence as the text prompts for T2V:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 465, + 295, + 605 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 465, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 295, + 523 + ], + "type": "text", + "content": "- VideoPhy [2] is designed to assess whether the generated videos follow physical common sense for real-world activities. It consists 688 human-verified captions that describe interactions between various types of real-world objects, including solid and fluid." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 524, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 295, + 605 + ], + "type": "text", + "content": "- PhyGenBench [29], similarly, allows evaluating the correctness of following physical common sense in T2V generation. It comprises 160 carefully crafted prompts spanning four physical domains, namely mechanics, optics, thermal and material properties. Since the domain of material properties has been covered by VideoPhy, we use the first three domains listed above." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 608, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 296, + 713 + ], + "type": "text", + "content": "Evaluation metric: We use VideoCon-Physics evaluator provided with VideoPhy [2], to measure how the generated video adheres to physical common sense (PC) and achieves semantic adherence (SA). The PC metric evaluates whether the depicted actions and object's state follow the real-world physics laws. The SA metric measures if the actions, events, entities and their interactions specified in the prompt are present. Both metrics yield binary outputs: 1 indicates adherence and 0 indicates otherwise. On each T2V model and" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 328, + 63, + 539, + 191 + ], + "blocks": [ + { + "bbox": [ + 328, + 63, + 539, + 191 + ], + "lines": [ + { + "bbox": [ + 328, + 63, + 539, + 191 + ], + "spans": [ + { + "bbox": [ + 328, + 63, + 539, + 191 + ], + "type": "table", + "html": "
Round1234
CogVideoX-5B [51]PC0.260.320.390.42
SA0.480.520.560.59
CogVideoX-2B [51]PC0.130.190.270.29
SA0.220.120.400.42
OpenSora [53]PC0.170.290.270.31
SA0.290.380.440.47
VideoCrafter [4]PC0.150.250.290.33
SA0.240.380.440.49
", + "image_path": "94f6f5c704b6081d15b27cbb35f1aa877147ed07138efc9bc37801426b150ce3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 328, + 227, + 539, + 355 + ], + "blocks": [ + { + "bbox": [ + 313, + 192, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 224 + ], + "type": "text", + "content": "Table 1. The quality of videos generated by different T2V models using the VideoPhy prompt dataset, over multiple rounds of iterative refinement in PhyT2V" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 227, + 539, + 355 + ], + "lines": [ + { + "bbox": [ + 328, + 227, + 539, + 355 + ], + "spans": [ + { + "bbox": [ + 328, + 227, + 539, + 355 + ], + "type": "table", + "html": "
Round1234
CogVideoX-5B [51]PC0.280.320.380.42
SA0.220.350.360.38
CogVideoX-2B [51]PC0.160.190.240.27
SA0.150.290.330.35
OpenSora [53]PC0.210.250.240.26
SA0.230.280.290.30
VideoCrafter [4]PC0.200.240.320.36
SA0.270.330.370.42
", + "image_path": "27a8807dcd57c1a9e6164c24fa7300864779fdeea61405c6802a5f985a564069.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 356, + 555, + 389 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 389 + ], + "type": "text", + "content": "Table 2. The quality of videos generated by different T2V models using the PhyGenBench prompt dataset, over multiple rounds of iterative refinement in PhyT2V" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 397, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 545, + 410 + ], + "type": "text", + "content": "dataset, the binary outputs from all prompts are averaged." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 410, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 553, + 468 + ], + "type": "text", + "content": "In addition, we also evaluated PhyT2V using the widely used VBenchmark metrics and benchmarks [15], which allow comprehensive evaluations of the generated video in multiple aspects, including video quality, video-condition consistency, prompt following and human preference alignment." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 469, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 554, + 564 + ], + "type": "text", + "content": "Baselines: For fair comparison, we only use the existing T2V prompt enhancers as baselines, and other existing work with extra inputs to T2V models [7, 19, 26, 27, 31] are not applicable. We involve two prompt enhancers: 1) Directly using the existing LLM, particularly ChatGPT4, as the prompt enhancer [28, 47]; 2) Promptist [21], which uses reinforcement learning to automatically refine and enhance prompts in the model-preferred way." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 569, + 549, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 549, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 549, + 582 + ], + "type": "text", + "content": "4.1. Improvement of the Generated Video Quality" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 583, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 554, + 700 + ], + "type": "text", + "content": "As shown in Table 1 and 2, when PhyT2V is applied to different T2V models, it can significantly improve the generated video's adherence to both the text prompt itself and the real-world physical rules, compared to the videos generated by vanilla T2V models (i.e., in Round 1 of PhyT2V's refinement). In particular, such improvement is the most significant on the CogVideoX-2B model, where PC improvement can be up to " + }, + { + "bbox": [ + 313, + 583, + 554, + 700 + ], + "type": "inline_equation", + "content": "2.2\\mathrm{x}" + }, + { + "bbox": [ + 313, + 583, + 554, + 700 + ], + "type": "text", + "content": " and SA improvement can be up to " + }, + { + "bbox": [ + 313, + 583, + 554, + 700 + ], + "type": "inline_equation", + "content": "2.3\\mathrm{x}" + }, + { + "bbox": [ + 313, + 583, + 554, + 700 + ], + "type": "text", + "content": ". On all the other models, PhyT2V can also reach noticeable improvement, ranging from " + }, + { + "bbox": [ + 313, + 583, + 554, + 700 + ], + "type": "inline_equation", + "content": "1.3\\mathrm{x}" + }, + { + "bbox": [ + 313, + 583, + 554, + 700 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 583, + 554, + 700 + ], + "type": "inline_equation", + "content": "1.9\\mathrm{x}" + }, + { + "bbox": [ + 313, + 583, + 554, + 700 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "type": "text", + "content": "Meanwhile, results in Table 1 and 2 showed that" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18831" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 67, + 271, + 139 + ], + "blocks": [ + { + "bbox": [ + 79, + 67, + 271, + 139 + ], + "lines": [ + { + "bbox": [ + 79, + 67, + 271, + 139 + ], + "spans": [ + { + "bbox": [ + 79, + 67, + 271, + 139 + ], + "type": "table", + "html": "
CogVideoX-5BOpenSora
ChatGPT 4 [28]PC0.330.21
SA0.410.32
Promptist [21]PC0.250.19
SA0.390.33
", + "image_path": "c60a4812cb07d2d79d7d2fd0e033a4b99cd033db6f25cba2cff25ed0ed1062f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 78, + 167, + 271, + 239 + ], + "blocks": [ + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "type": "text", + "content": "Table 3. The quality of videos generated by enhancing the prompts in the VideoPhy dataset using different prompt enhancers" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 167, + 271, + 239 + ], + "lines": [ + { + "bbox": [ + 78, + 167, + 271, + 239 + ], + "spans": [ + { + "bbox": [ + 78, + 167, + 271, + 239 + ], + "type": "table", + "html": "
CogVideoX-5BOpenSora
ChatGPT 4 [28]PC0.270.20
SA0.230.23
Promptist [21]PC0.320.19
SA0.240.21
", + "image_path": "d7b4952711ed70bf6a9a17eaad88391608f086ec8f01f0c6c569b7e781a73b63.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 295, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 295, + 264 + ], + "type": "text", + "content": "Table 4. The quality of videos generated by enhancing the prompts in the PhyGenBench dataset using different prompt enhancers" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 269, + 295, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 295, + 328 + ], + "type": "text", + "content": "PhyT2V's process of iterative refinement converge quickly and only takes few rounds. Most improvement of video quality happens in the first two rounds, and little improvement can be observed in the fourth round. Hence, in practice, we believe that 3-4 iterative rounds would be sufficient." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 328, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 422 + ], + "type": "text", + "content": "Furthermore, as shown in Table 3 and 4, PhyT2V also largely outperforms the existing prompt enhancers by at least " + }, + { + "bbox": [ + 55, + 328, + 295, + 422 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 55, + 328, + 295, + 422 + ], + "type": "text", + "content": ", when being applied to CogVideoX-5B and OpenSora models. In particular, ChatGPT 4, when being used as the prompt enhancer, delivers better performance than Promptist due to its stronger language processing capabilities, but still cannot ensure physics-grounded T2V, due to the lack of explicit reasoning on text-to-video alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 422, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 529 + ], + "type": "text", + "content": "Our evaluation results on the VBenchmark metrics are shown in Figure 7, where numbers in Round 1 are the T2V model's original performance in current VBenchmark leaderboard, and iterative prompt refinements by PhyT2V in Round 2 & 3 noticeably improve the performance in many dimensions. In particular, large improvements are noted in most dimensions of Video-Condition Consistency, showing that PhyT2V improves T2V model's adherence to prompts and real-world physical rules underlying the prompts." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 531, + 295, + 628 + ], + "blocks": [ + { + "bbox": [ + 58, + 531, + 295, + 628 + ], + "lines": [ + { + "bbox": [ + 58, + 531, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 58, + 531, + 295, + 628 + ], + "type": "image", + "image_path": "37805667bedfd49b946dedca9679a1e972d7c9248a9f8d30d4ca8f25f556908a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 632, + 295, + 656 + ], + "lines": [ + { + "bbox": [ + 55, + 632, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 295, + 656 + ], + "type": "text", + "content": "Figure 7. PhyT2V VBenchmark evaluation results with CogVideoX-5B (left) and OpenSora (right)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 664, + 248, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 248, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 248, + 677 + ], + "type": "text", + "content": "4.2. Different Domains of Physical Rules" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "We also conducted in-depth analysis on PhyT2V's performance on improving the generated video's quality in different domains of real-world physical rules, using the" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 63, + 552, + 214 + ], + "blocks": [ + { + "bbox": [ + 316, + 63, + 552, + 214 + ], + "lines": [ + { + "bbox": [ + 316, + 63, + 552, + 214 + ], + "spans": [ + { + "bbox": [ + 316, + 63, + 552, + 214 + ], + "type": "image", + "image_path": "35c8323bd586c74d247bb4e8ff65310a069fb50c8fc7dfa38dac39989d1436d2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 240, + 553, + 399 + ], + "blocks": [ + { + "bbox": [ + 313, + 218, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 239 + ], + "type": "text", + "content": "Figure 8. Examples of videos generated using different categories of prompts in the PhyGenBench dataset" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 240, + 553, + 399 + ], + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 399 + ], + "type": "image", + "image_path": "623e59ce559975c0b379d113d4a4e4feea6eebf080f5960abc0157c0b52d4e02.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 404, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 313, + 404, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 555, + 427 + ], + "type": "text", + "content": "Figure 9. Examples of videos generated using different categories of prompts in the VideoPhy dataset" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 519 + ], + "type": "text", + "content": "CogVideoX-5B as the T2V model and ChatGPT 4 as the prompt enhancer. As shown in Table 5 and 6, PhyT2V achieves large improvements in most domains of physical rules. Especially in domains where physical interaction between objects are more subtle and difficult to be precisely captured, such as interaction with fluids and thermal-related scene changes, such improvements will be even higher." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 519, + 555, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 649 + ], + "type": "text", + "content": "These improvements are also exemplified with sample videos and their related input prompts in Figure 9 and Figure 8. With LLM reasoning and iterative refinement, PhyT2V can largely enhance the T2V model's capability when encountering out-of-distribution prompts, by providing correct and sufficient contexts to ensure that the T2V model's video generation correctly captures the key objects and interaction between objects. For example, when the prompt of \"juice dropping from a bottle onto the counter\", PhyT2V correctly depicts the juice's slow diffusion on the counter. More examples can be found in Appendix B." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 653, + 408, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 408, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 408, + 665 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": "We conduct an ablation study to evaluate the necessity of both the physical rule reasoning (Step 1) and the mismatch reasoning (Step 2) within our PhyT2V workflow, by removing one of these steps from the refinement process to assess" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18832" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 545, + 185 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 545, + 185 + ], + "type": "table", + "html": "
CogVideoX-5BCogVideoX-2BOpenSoraVideoCrafter
Round1234123412341234
Solid-SolidPC0.210.280.340.320.090.130.140.220.120.270.290.300.190.220.270.28
SA0.240.480.490.470.180.250.360.330.160.340.370.350.240.400.450.47
Solid-FluidPC0.220.270.280.300.110.180.280.270.170.210.240.250.180.240.250.26
SA0.390.540.600.610.290.430.440.430.160.400.410.360.340.430.480.52
Fluid-FluidPC0.570.590.630.620.340.380.350.360.150.320.290.310.330.410.530.51
SA0.410.570.590.670.270.420.390.440.310.440.450.460.320.420.490.51
", + "image_path": "60b97f80f1fc7a55442e997e78ba95b9ae7e76a85c458361b999053f8cbbe258.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 64, + 201, + 544, + 316 + ], + "blocks": [ + { + "bbox": [ + 79, + 186, + 531, + 198 + ], + "lines": [ + { + "bbox": [ + 79, + 186, + 531, + 198 + ], + "spans": [ + { + "bbox": [ + 79, + 186, + 531, + 198 + ], + "type": "text", + "content": "Table 5. The improvement of generated video quality in different categories of physical rules in the VideoPhy prompt dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 201, + 544, + 316 + ], + "lines": [ + { + "bbox": [ + 64, + 201, + 544, + 316 + ], + "spans": [ + { + "bbox": [ + 64, + 201, + 544, + 316 + ], + "type": "table", + "html": "
CogVideoX-5BCogVideoX-2BOpenSoraVideoCrafter
Round1234123412341234
MechanicsPC0.190.250.340.350.120.160.180.240.110.130.170.220.140.230.290.28
SA0.210.280.290.320.110.180.190.220.190.210.270.320.200.240.280.35
OpticsPC0.220.350.410.390.220.250.290.280.240.260.250.250.220.210.270.32
SA0.270.420.390.440.230.340.370.390.260.310.290.300.220.280.350.39
ThermalPC0.330.350.350.350.130.150.150.140.270.300.310.330.250.280.260.28
SA0.220.360.430.450.120.160.240.270.230.250.370.360.250.370.410.43
", + "image_path": "7f346fd04fa4c4570d37fbe685461e3a86ad5b8fa881e2150682e00d78a1f52e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 317, + 539, + 328 + ], + "lines": [ + { + "bbox": [ + 70, + 317, + 539, + 328 + ], + "spans": [ + { + "bbox": [ + 70, + 317, + 539, + 328 + ], + "type": "text", + "content": "Table 6. The improvement of generated video quality in different categories of physical rules in the PhyGenBench prompt dataset" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 335, + 258, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 258, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 258, + 346 + ], + "type": "text", + "content": "its impact on physical-grounded video generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 346, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 464 + ], + "type": "text", + "content": "Physical rule reasoning (Step 1). As shown in Figure 10, the Step 1 of physical rule reasoning significantly enhances the T2V process by providing a more detailed and coherent description of the principal object's physical status, such as motion, states and deformation (red texts in Figure 10), all grounded in relevant physical laws. By anchoring the prompt in established physical rules, this step also helps avoid unnecessary texts (brown texts in Figure 10) and vague physical rule descriptions (purple texts in Figure 10), hence achieving a higher PC score." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 61, + 466, + 293, + 601 + ], + "blocks": [ + { + "bbox": [ + 61, + 466, + 293, + 601 + ], + "lines": [ + { + "bbox": [ + 61, + 466, + 293, + 601 + ], + "spans": [ + { + "bbox": [ + 61, + 466, + 293, + 601 + ], + "type": "image", + "image_path": "184427b9654900e900410a525a756c99bc3a84987a626ce9ba2da482680bac76.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 604, + 290, + 615 + ], + "lines": [ + { + "bbox": [ + 61, + 604, + 290, + 615 + ], + "spans": [ + { + "bbox": [ + 61, + 604, + 290, + 615 + ], + "type": "text", + "content": "Figure 10. Ablation study on Step 1 of physical rule reasoning" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 619, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 296, + 713 + ], + "type": "text", + "content": "Mismatch reasoning (Step 2). The Step 2 of mismatch reasoning addresses details that may have been overlooked in the previous iteration of the generated video as shown in Figure 11. This step plays a critical role in the iterative refinement process by identifying and correcting discrepancies between expected and observed outputs. By enhancing the model's focus on the principal object, the mismatch reasoning step reduces the likelihood of losing attention to" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 332, + 553, + 464 + ], + "blocks": [ + { + "bbox": [ + 314, + 332, + 553, + 464 + ], + "lines": [ + { + "bbox": [ + 314, + 332, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 332, + 553, + 464 + ], + "type": "image", + "image_path": "d346e6757ccf1d5eab11a5d1706259a20cc8e7a9c29c83a128cc0ec8588cd184.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 468, + 542, + 479 + ], + "lines": [ + { + "bbox": [ + 326, + 468, + 542, + 479 + ], + "spans": [ + { + "bbox": [ + 326, + 468, + 542, + 479 + ], + "type": "text", + "content": "Figure 11. Ablation study on Step 2 of mismatch reasoning" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 553, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 553, + 517 + ], + "type": "text", + "content": "important features (brown and purple texts in Figure 11), improving the fidelity and relevance of generated video content (red texts in Figure 11) towards a higher SA score." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 518, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 554, + 566 + ], + "type": "text", + "content": "Overall, our study shows that both reasoning steps are integral to PhyT2V, contributing to a more robust and semantically-aligned generation of refined prompts in Step 3. More detailed ablation studies are in Appendix C." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 388, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 388, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 388, + 581 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 586, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we present PhyT2V, a novel data-independent T2V generation framework designed to enhance the generalization capability of existing T2V models to out-of-distribution domains. By incorporating CoT reasoning and step-back prompting, PhyT2V systematically refines T2V prompts to ensure adherence to real-world physical principles without necessitating additional model retraining or reliance on additional conditions. Evaluation results indicate that PhyT2V achieves a " + }, + { + "bbox": [ + 313, + 586, + 555, + 715 + ], + "type": "inline_equation", + "content": "2.3\\mathrm{x}" + }, + { + "bbox": [ + 313, + 586, + 555, + 715 + ], + "type": "text", + "content": " enhancement in physical realism compared to baseline T2V models and outperforms state-of-the-art T2V prompt enhancers by " + }, + { + "bbox": [ + 313, + 586, + 555, + 715 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 313, + 586, + 555, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18833" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 157 + ], + "type": "text", + "content": "We thank the reviewers and the area chair for their insightful comments and feedback. This work was supported in part by National Science Foundation (NSF) under grant number IIS-2205360, CCF-2217003, CCF-2215042, and National Institutes of Health (NIH) under grant number R01HL170368." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 187, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 187, + 296, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 187, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 61, + 187, + 296, + 219 + ], + "type": "text", + "content": "[1] Luke Auburn. Ai video generation expert discusses the technology's rapid advances—and its current limitations, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 220, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 296, + 275 + ], + "type": "text", + "content": "[2] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, KaiWei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 275, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 275, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 275, + 295, + 319 + ], + "type": "text", + "content": "[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 1, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 384 + ], + "type": "text", + "content": "[4] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3, 6, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 384, + 296, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 427 + ], + "type": "text", + "content": "[5] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 428, + 296, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 296, + 483 + ], + "type": "text", + "content": "[6] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, et al. A survey on in-context learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1107-1128, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 483, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 296, + 525 + ], + "type": "text", + "content": "[7] Theodoros Dumas and Alexandros Sigalas. Blender, an open source design tool: Advances and integration in the architectural production pipeline. Aristoteleio University of Thessaloniki, 21:737-744, 2009. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 526, + 296, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 296, + 549 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 296, + 549 + ], + "type": "text", + "content": "[8] Blender Foundation. Upbge: an open-source, 3d game engine forked from the old blender game engine, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 549, + 296, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 549, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 62, + 549, + 296, + 593 + ], + "type": "text", + "content": "[9] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 593, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 593, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 593, + 295, + 647 + ], + "type": "text", + "content": "[10] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In European Conference on Computer Vision, pages 393–411. Springer, 2025. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "text", + "content": "[11] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "type": "text", + "content": "[12] Kai Huang, Haoming Wang, and Wei Gao. Freezeasguard: Mitigating illegal adaptation of diffusion models via selec" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tive tensor freezing. arXiv preprint arXiv:2405.17472, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[13] Kai Huang, Hanyun Yin, Heng Huang, and Wei Gao. Towards green ai in fine-tuning large language models via adaptive backpropagation. ICLR, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "text", + "content": "[14] Kai Huang, Xiangyu Yin, Heng Huang, and Wei Gao. Modality plug-and-play: Runtime modality adaptation in LLM-driven autonomous mobile systems. In ACM MobiCom, 2025. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 240 + ], + "type": "text", + "content": "[15] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 262 + ], + "type": "text", + "content": "[16] Epic Games Inc. Unreal engine: The most powerful real-time 3d creation tool, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 444, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 444, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 444, + 275 + ], + "type": "text", + "content": "[17] Mellis Inc. Pika labs, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 277, + 537, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 537, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 537, + 289 + ], + "type": "text", + "content": "[18] OpenAI Inc. Introducing openai o1-preview, 2024. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 289, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 289, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 317, + 289, + 553, + 334 + ], + "type": "text", + "content": "[19] Marcel Krüger, David Gilbert, Torsten W. Kuhlen, and Tim Gerrits. Game engines for immersive visualization: Using unreal engine beyond entertainment. PRESENCE: Virtual and Augmented Reality, 33:31-55, 2024. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 334, + 553, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 334, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 316, + 334, + 553, + 367 + ], + "type": "text", + "content": "[20] Chengxuan Li, Di Huang, Zeyu Lu, Yang Xiao, Qingqi Pei, and Lei Bai. A survey on long video generation: Challenges, methods, and prospects, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 369, + 553, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 369, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 369, + 553, + 422 + ], + "type": "text", + "content": "[21] WeiJie Li, Jin Wang, and Xuejie Zhang. Promptist: Automated prompt optimization for text-to-image synthesis. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 295-306. Springer, 2024. 6, 7, 11" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 424, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 424, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 424, + 553, + 478 + ], + "type": "text", + "content": "[22] Yingcong Li, Kartik Sreenivasan, Angeliki Giannou, Dimitris Papailiopoulos, and Samet Oymak. Dissecting chain-of-thought: Compositionality through in-context filtering and learning. Advances in Neural Information Processing Systems, 36, 2024. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 479, + 553, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 523 + ], + "type": "text", + "content": "[23] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 524, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 524, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 524, + 553, + 557 + ], + "type": "text", + "content": "[24] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 558, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 558, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 558, + 553, + 601 + ], + "type": "text", + "content": "[25] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 602, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 647 + ], + "type": "text", + "content": "[26] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imaged-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 647, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 714 + ], + "type": "text", + "content": "[27] Jiaxi Lv, Yi Huang, Mingfu Yan, Jiancheng Huang, Jianzhuang Liu, Yifan Liu, Yafei Wen, Xiaoxin Chen, and Shifeng Chen. Gpt4: Scripting physical motions in text-to-video generation via blender-oriented gpt planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1430-1440, 2024. 6" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18834" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[28] Jiaxi Lv, Yi Huang, Mingfu Yan, Jiancheng Huang, Jianzhuang Liu, Yifan Liu, Yafei Wen, Xiaoxin Chen, and Shifeng Chen. Gpt4motion: Scripting physical motions in text-to-video generation via blender-oriented gpt planning, 2024. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 127, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 127, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 127, + 294, + 183 + ], + "type": "text", + "content": "[29] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 294, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 225 + ], + "type": "text", + "content": "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "text", + "content": "[31] Weichao Qiu and Alan Yuille. Unrealcv: Connecting computer vision to unreal engine. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 909-916. Springer, 2016. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "text", + "content": "[32] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 435 + ], + "type": "text", + "content": "[33] Sander Schulhoff, Michael Ilie, Nishant Balepur, Konstantine Kahadze, Amanda Liu, Chenglei Si, Yinheng Li, Aayush Gupta, HyoJung Han, Sevien Schulhoff, Pranav Sandeep Dulepet, Saurav Vidyadhara, Dayeon Ki, Sweta Agrawal, Chau Pham, Gerson Kroiz, Feileen Li, Hudson Tao, Ashay Srivastava, Hevander Da Costa, Saloni Gupta, Megan L. Rogers, Inna Goncearenco, Giuseppe Sarli, Igor Galynker, Denis Peskoff, Marine Carpuat, Jules White, Shyamal Anadkat, Alexander Hoyle, and Philip Resnik. The prompt report: A systematic survey of prompting techniques, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 436, + 294, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 436, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 56, + 436, + 294, + 490 + ], + "type": "text", + "content": "[34] Sander Schulhoff, Michael Ilie, Nishant Balepur, Konstantine Kahadze, Amanda Liu, Chenglei Si, Yinheng Li, Aayush Gupta, HyoJung Han, Sevien Schulhoff, et al. The prompt report: A systematic survey of prompting techniques. arXiv preprint arXiv:2406.06608, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 491, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 294, + 544 + ], + "type": "text", + "content": "[35] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 546, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 567 + ], + "type": "text", + "content": "[36] Unity Technologies. Unity real-time development platform, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 569, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 590 + ], + "type": "text", + "content": "[37] Stuart A. Thompson. A.i. can now create lifelike videos. can you tell what's real?, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[38] Boshi Wang, Sewon Min, Xiang Deng, Jiaming Shen, You Wu, Luke Zettlemoyer, and Huan Sun. Towards understanding chain-of-thought prompting: An empirical study of what matters. arXiv preprint arXiv:2212.10001, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "text", + "content": "[39] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 2, 6, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[40] Keheng Wang, Feiyu Duan, Sirui Wang, Peiguang Li, Yun-sen Xian, Chuantao Yin, Wenge Rong, and Zhang Xiong. Knowledge-driven cot: Exploring faithful reasoning in llms" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "for knowledge-intensive question answering. arXiv preprint arXiv:2308.13259, 2023. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[41] Xiaofeng Wang, Zheng Zhu, Guan Huang, Boyuan Wang, Xinze Chen, and Jiwen Lu. Worlddreamer: Towards general world models for video generation via predicting masked tokens, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[42] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "text", + "content": "[43] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 303 + ], + "type": "text", + "content": "[44] Zhendong Wang, Yifan Jiang, Yadong Lu, Pengcheng He, Weizhu Chen, Zhangyang Wang, Mingyuan Zhou, et al. Incontext learning unlocked for diffusion models. Advances in Neural Information Processing Systems, 36:8542-8562, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 304, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 349 + ], + "type": "text", + "content": "[45] Zhao Wang, Aoxue Li, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. Customvideo: Customizing text-to-video generation with multiple subjects. arXiv preprint arXiv:2401.09962, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[46] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "type": "text", + "content": "[47] Deshun Yang, Luhui Hu, Yu Tian, Zihao Li, Chris Kelly, Bang Yang, Cindy Yang, and Yuexian Zou. Worldgpt: a sora-inspired video ai agent as rich world models from text and image inputs. arXiv preprint arXiv:2403.07944, 2024. 6, 11" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[48] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In *Forty-first International Conference on Machine Learning*, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "text", + "content": "[49] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "text", + "content": "[50] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer, 2024. 1, 3, 11" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "type": "text", + "content": "[51] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[52] Wentao Zhang, Junliang Guo, Tianyu He, Li Zhao, Linli Xu, and Jiang Bian. Video in-context learning. arXiv preprint arXiv:2407.07356, 2024. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18835" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 159 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[53] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 6, 11" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 159 + ], + "type": "text", + "content": "[54] Hanxin Zhu, Tianyu He, Anni Tang, Junliang Guo, Zhibo Chen, and Jiang Bian. Compositional 3d-aware video generation with llm director. arXiv preprint arXiv:2409.00558, 2024. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18836" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/d8df7b02-5bad-4fed-8afa-940086807a3d_content_list.json b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/d8df7b02-5bad-4fed-8afa-940086807a3d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0c7afd8b26b4f0e7ff331a58dc5c487924582612 --- /dev/null +++ b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/d8df7b02-5bad-4fed-8afa-940086807a3d_content_list.json @@ -0,0 +1,2688 @@ +[ + { + "type": "text", + "text": "PhysAnimator: Physics-Guided Generative Cartoon Animation", + "text_level": 1, + "bbox": [ + 179, + 130, + 818, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyi Xie $^{1,2*}$", + "bbox": [ + 238, + 179, + 352, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiwei Zhao $^{1\\dagger}$ Ying Jiang $^{2}$", + "bbox": [ + 374, + 179, + 594, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenfanfu Jiang2", + "bbox": [ + 616, + 180, + 756, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Netflix, 2 UCLA", + "bbox": [ + 424, + 198, + 571, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tianyixie77@ucla.edu", + "bbox": [ + 169, + 219, + 352, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yiweiz@netflix.com", + "bbox": [ + 370, + 219, + 534, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yingjiang, cffjiang}@ucla.edu", + "bbox": [ + 555, + 219, + 820, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a9e715952303dadac7b0abec1e9ca13a20e23492b52a5df698ddcfbd7ab94904.jpg", + "image_caption": [ + "Rig Force", + "Figure 1. PhysAnimator is a novel framework that combines physics principles with video diffusion models to generate high-quality animations from static animeillustrations, allowing users to specify external forces or rigging points for custom effects." + ], + "image_footnote": [], + "bbox": [ + 94, + 271, + 254, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2571b5993f7dd2465a6b708ce88ca9acd9611039062dffc427db9514158bd22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 271, + 424, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92b152343db20f36e20ba59400c8c02cd1ebfb6738aa821026e6ed68f8e91633.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 349, + 424, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e74272d0eddf6ccf39a046da0ced18d8fbf6c955bc416571fd55d49ae9dda6ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 266, + 425, + 424, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76d2301538b8a695d2e2b0f899b878883d59fd49fab1073a7e79e4552106250d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 266, + 502, + 424, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/82a4db7032800870a7677346ec9d4aea10033d2165d93dab2845422fb5c3bd8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 271, + 581, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31132fdc53ef90a1da5c8dcaf39025bdc65dfa7c0956db9a5a5271433ddff276.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 349, + 581, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f4c636e8f00a9d6014fe742ebb998c531e2f5b04d1245dc770dcd97695ba261a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 425, + 581, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/085586728b176ffe03c4a51c8681f805e4665df120df0dfd6ca543fd1515f92f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 502, + 581, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae902493323f1f161c82965912dedd9fb0cb7d917cfe98d2b638c123cfa3b5f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 271, + 738, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd188b53557b541e076a377e4b34475758ac451ab94bd4aa71da060b4e1d6b71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 349, + 738, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94044e4096ddabd0f1e47a8dc19322762f298ec456a3c57fefc0ef000c5d530b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 425, + 738, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e932a9211f4d910a824f9921b9dc44b131292a9bda0c1d628a18f4c9c7ac222.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 502, + 738, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc3186f3e1485eeff8ca0fd49684b4b89673fd32756ba4ffa1b2ad46fdd8f5f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 271, + 897, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41be1be77d15251e163cfb20d4e8e3681458e3997d1fb683776e2e7aef9bb4a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 349, + 897, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a7b28c573c7e8feebb4e3b0d8c8b3c74caae4ed33545b4d8b30c4e9178eb00f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 425, + 897, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/228503e3ae85b052ba6e630cbac259a88d5fef2840a646235a7a48fe89fb9459.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 502, + 897, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 648, + 326, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating hand-drawn animation sequences is labor-intensive and demands professional expertise. We introduce PhysAnimator, a novel approach for generating physically plausible meanwhile anime-stylized animation from static animeillustrations.Our method seamlessly integrates physics-based simulations with data-driven generative models to produce dynamic and visually compelling animations.To capture the fluidity and exaggeration characteristic of anime,we perform image-space deformable body simulations on extracted mesh geometries.We enhance artistic control by introducing customizable energy strokes and incorporating rigging point support, enabling", + "bbox": [ + 89, + 681, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the creation of tailored animation effects such as wind interactions. Finally, we extract and warp sketches from the simulation sequence, generating a texture-agnostic representation, and employ a sketch-guided video diffusion model to synthesize high-quality animation frames. The resulting animations exhibit temporal consistency and visual plausibility, demonstrating the effectiveness of our method in creating dynamic anime-style animations. See our project page for more demos: https://xpandora.github.io/PhysAnimator/.", + "bbox": [ + 511, + 651, + 906, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 829, + 645, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dynamic visual effects are essential to the immersive quality of 2D animation. From the subtle sway of a character's hair to the fluid motion of garments in response to", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during the internship at Netflix.", + "bbox": [ + 112, + 875, + 352, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 116, + 888, + 245, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10793", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wind, realistic and pleasing dynamics create a captivating visual experience. Traditionally, these effects are achieved through meticulous, hand-drawn techniques, where animators painstakingly draw each frame to bring these dynamic elements to life. This process is labor-intensive and requires not only artistic skill but also a deep understanding of natural forces and environmental interactions.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate the challenges of manual animation, researchers have explored both traditional and data-driven approaches. Traditional animation tools [31, 58] provide interactive systems that assist artists in creating animated illustrations based on established principles of 2D animation. These methods often rely on user-provided stroke inputs to specify motion trajectories and utilize geometry constraints to produce deformation-based animation. However, such approaches typically assume simple inputs like lineart or drawings with separated layers, limiting their applicability to more complex, in-the-wild anime illustrations that feature intricate textures and details. In contrast, data-driven video generative models [2, 46, 71] offer a promising alternative by leveraging neural synthesis to generate dynamic effects directly from images, bypassing the need for manual sketching or layered inputs. Recent methods even enable interactive motion design, allowing users to specify object trajectories via drag-based inputs [49, 55, 67]. These approaches typically rely on predicting a sequence of optical flow fields to drive the desired motion and warp the frames accordingly. However, the quality of the generated results is often limited by inaccuracies in the predicted optical flow, which frequently exhibits artifacts due to a lack of geometric understanding and physical constraints. As a result, imprecise motion estimation can lead to noticeable distortions and unsatisfactory visual quality.", + "bbox": [ + 91, + 200, + 486, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recognizing the limitations of both traditional and purely data-driven methods, we introduce PhysAnimator, a novel framework that integrates physics-based animation with data-driven generative models to synthesize visually compelling dynamic animations driven by environmental forces such as wind or rigging from static animeillustrations. Our approach combines the physical consistency of simulation-based methods with the flexibility and expressiveness of generative models, overcoming the drawbacks of previous approaches. To achieve this, we model objects of interest in the anime illustration as deformable bodies, capturing the fluidity and exaggerated motion characteristic of anime. We solve the motion equations to compute a sequence of optical flow fields that represent consistent dynamics. Users can enhance these animations with customized energy strokes, defining the effects of external forces such as wind. To render the motion dynamics as high-quality frames, we first extract and warp the sketch using the optical flow fields to generate a texture-agnostic video sequence. We then apply a sketch-guided video diffu", + "bbox": [ + 89, + 598, + 486, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sion model to colorize the sketch sequence based on the reference illustration, ensuring stylistic coherence. Finally, to incorporate dynamic effects that cannot be easily captured by physical simulation, we employ a data-driven cartoon interpolation model, enriching the results with complementary, expressive dynamics. In summary, our contributions are as follows:", + "bbox": [ + 511, + 90, + 903, + 195 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel framework that combines physics-based simulations with data-driven generative models, specifically targeting animations driven by environmental forces and rigging controls, achieving both physical consistency and stylistic expressiveness.", + "- We develop an image-space deformable body simulation technique that models anime objects as deformable meshes, capturing fluid and exaggerated dynamics.", + "- We leverage a sketch-guided video diffusion model to render simulation dynamics as high-quality frames, combined with a cartoon interpolation model to introduce additional dynamic effects beyond physical laws." + ], + "bbox": [ + 511, + 196, + 903, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 393, + 655, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Video Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 419, + 730, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the success of image diffusion models [10, 53], recent work has introduced video diffusion models to streamline video synthesis from text prompts or images, significantly reducing labor and time costs compared to traditional commercial video editing and creation tools [23, 72]. These approaches either extend pre-trained image diffusion models by incorporating temporal mixing layers in various forms [3, 11, 13, 56] or train video diffusion models with temporal layers from scratch on large-scale text-video paired datasets [20, 79]. However, using text prompts or a single image as input provides limited control over fine-grained aspects of video generation, making it challenging to define complex structural attributes such as spatial layouts, poses, and shapes [17, 74]. To allow for more detailed control over object motion and camera movements, additional guidance in the form of motion trajectories [6, 28, 55, 74], pose [77], depth [21], and optical flow [17] has been integrated into video diffusion models to produce more controllable videos. These powerful video diffusion models have also been applied to various downstream tasks, such as video editing [16, 35], image animation [12, 66, 71], video understanding [38, 54, 61], video interpolation [27, 70] and 3D reconstruction and generation [8, 18, 36, 43, 60]. Nevertheless, these data-driven approaches usually produce artifacts due to a lack of geometric understanding and physical constraints.", + "bbox": [ + 511, + 443, + 906, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Physics-based Animation", + "text_level": 1, + "bbox": [ + 511, + 845, + 743, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Researchers have explored various methods for animation authoring. Xing et al. [68] extend local similarity tech", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10794", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "niques to global similarity, enabling automatic sketch completion based on previous frames. Building on this, Peng et al. [51] propose a keyframe-based sculpting system that autocompletes user edits using an intuitive brushing interface. Willett et al. [65] tackle the challenge of textured anime images by segmenting the illustration into layers for different objects and allowing users to provide scribbles as trajectory guides. Leveraging animation principles, some works [31, 45] use geometry-based deformations to create stylized animation effects such as \"squash\" and \"stretch\". Similarly, Xing et al. [69] employ physics-based simulations and introduce energy brushes to generate elemental dynamics [15], such as smoke and fire. Other approaches [1, 9, 29] incorporate physical laws like gravity, collision, and elasticity to animate deformable characters, enabling the efficient creation of complex scenes. For instance, Jones et al. [30] introduces an example-based plasticity model based on linear blend skinning for animating the failure of near-rigid, man-made materials. Coros et al. [9] exploit two rest-posed adaptation strategies using only internal energy to animate curve, shell, and solid-based characters. Physics-based modeling has also proven effective in adding secondary dynamics [64, 75] to enhance details in rigged animations. Despite their effectiveness, these 2D animation methods typically assume simple inputs, such as sketches, or require layered separation, which limits their applicability for animating complex, in-the-wild anime illustrations.", + "bbox": [ + 89, + 90, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Generative Image Dynamics", + "text_level": 1, + "bbox": [ + 89, + 513, + 344, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the advancement of generative models, there is growing interest in leveraging these powerful techniques to synthesize dynamic animations from static images, guided by motion features extracted from various user inputs. These inputs can be sparse, such as text prompts [7, 41, 71], trajectories [42, 55, 62, 67, 78], or camera movements [62, 73], or dense, like reference videos [28, 63, 80]. To achieve controllable dynamics, prior works often incorporate ControlNet [76] into image or video generative models during the decoding stage, utilizing motion features such as Canny edges, depth maps [17], 2D Gaussian maps [67], and optical flow maps [55]. In addition, several methods employ specialized motion fusion modules [41, 62, 78], spatial or temporal attention mechanisms [28, 63, 73], and feature injection strategies [42] to guide the generation of controllable dynamics. However, these approaches typically lack physics-based supervision, which can result in animations that violate physical laws or fail to align with user intentions [47]. Recently, PhysGen [44] integrated a rigid-body physics simulator to generate physics-consistent dynamics for foreground objects in a given image. While effective, PhysGen is limited to 2D rigid-body motions, making it unsuitable for the fluid, elastic effects commonly seen in anime, such as the waving of cloth or hair, which do not", + "bbox": [ + 89, + 537, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "conform to rigid-body dynamics. In this work, we address these limitations by leveraging deformable-body dynamics and a sketch-guided video diffusion model to create high-quality, physically consistent animation sequences that capture the fluidity and elasticity characteristic of anime.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 179, + 604, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a reference anime illustration $I_0$ , our goal is to generate a sequence of stylized video frames $\\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_T\\}$ with natural, user-guided motions. To achieve this, we introduce PhysAnimator, a novel framework that combines video diffusion models with physics-based animation. We start by segmenting the objects of interest in a given anime illustration and creating a 2D triangulated mesh representation. Using this geometry, we obtain dynamics through image-space deformable body simulation, generating a sequence of optical flow fields $\\{\\mathcal{F}_{0 \\to 1}, \\mathcal{F}_{0 \\to 2}, \\dots, \\mathcal{F}_{0 \\to T}\\}$ . To offer user control, we enable interactive inputs through customizable energy strokes for defining external forces and rigging points for specifying desired trajectories. To render dynamics as high-quality frames, we extract the sketch from the input illustration and warp it using the computed optical flow fields, yielding a sequence of dynamic sketch frames $\\{S_1, S_2, \\dots, S_T\\}$ . These sketches, together with the input illustration, are fed into a video diffusion model with a sketch-guided ControlNet [76], synthesizing a vivid animated sequence. Finally, an optional data-driven cartoon interpolation model [70] can be applied to enhance the anime style dynamics of the results by selecting keyframes from the animated sequence as input, yielding the final output frames $\\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_T\\}$ . An overview of our proposed framework is shown in Fig. 2.", + "bbox": [ + 511, + 204, + 906, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 590, + 653, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Video Diffusion Model Most video diffusion models build on the latent diffusion model (LDM) framework [53], which uses Variational Autoencoder [32] (VAE) to map input images into a latent space. In this space, data is transformed into Gaussian noise via a forward diffusion process, which the model learns to reverse through denoising. In the forward diffusion process, the latent code $z_0$ is perturbed as:", + "bbox": [ + 511, + 614, + 905, + 736 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 744, + 906, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_t)$ with $\\beta_t$ controlling the noise strength coefficient at step $t$ . The denoising model $\\epsilon_\\theta$ is trained to recover $z_t$ by minimizing the objective function:", + "bbox": [ + 511, + 773, + 905, + 833 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\epsilon} = \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}; c, t) \\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 843, + 903, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ denotes the learnable network parameters and $c$ represents conditioning input (e.g. text prompts or images).", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10795", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9ac3ae24b651918e23a47af2a20c77497bd842cbd723639be066e332273292e3.jpg", + "image_caption": [ + "Figure 2. Method Overview. We begin by segmenting the object and creating a triangulated deformable mesh. Physics-based simulations are then used to generate dynamic optical flow fields, with users given the option to guide the motion through customizable energy strokes (shown as orange arrows) and rigging points (shown as red dots). The extracted sketch is warped using the computed optical flow and refined with a sketch-guided video diffusion model, producing a smooth, stylized animation sequence. Optionally, a cartoon interpolation model can further be applied to enhance the animation with expressive dynamics." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After denoising, the VAE decoder reconstructs the latent code back into the image space. The latent video diffusion model (LVDM) extends the image LDM to videos by incorporating temporal modules to maintain temporal consistency.", + "bbox": [ + 88, + 425, + 483, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deformable Body Dynamics Mathematically, the dynamics of a continuum deformable body is described by a time-dependent continuous deformation map $\\boldsymbol{x} = \\phi(\\boldsymbol{X}, t)$ , which maps the undeformed material space $\\Omega^0$ to the deformed world space $\\Omega^t$ at time $t$ . The deformation gradient $\\boldsymbol{F} = \\frac{\\partial \\phi}{\\partial \\boldsymbol{X}}$ encodes the local deformation such as scaling, rotation, and shearing. In the context of a discretized setting, e.g. 2D meshes, the deformation map for each triangle can be expressed as", + "bbox": [ + 89, + 512, + 483, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {i} (\\boldsymbol {X}) = \\boldsymbol {F} _ {i} \\boldsymbol {X} + \\boldsymbol {b} _ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 660, + 482, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{b}_i\\in \\mathbb{R}^2$ accounts for translation of $i$ -th triangle, and the deformation gradient $\\pmb {F}_i\\in \\mathbb{R}^{2\\times 2}$ is assumed to be constant [40]. Undergoing deformation, the deformed body aims to recover its rest shape via resisting forces. Continuum mechanics model this behavior by first defining an energy density function $\\Psi (F(x))$ , which measures the strain energy per unit undeformed volume. The total potential energy for a deformable body is then obtained as", + "bbox": [ + 89, + 686, + 483, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE (\\boldsymbol {x}) = \\sum_ {i = 1} ^ {N} \\Psi \\left(\\boldsymbol {F} _ {i}\\right) V _ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 820, + 482, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V_{i}$ denotes volume of $i$ -th triangle. The internal resisting force is then defined as the negative gradient of the", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "potential energy with respect to the vertex position", + "bbox": [ + 511, + 426, + 849, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {\\text {i n t}} (\\boldsymbol {x}) = - \\frac {\\partial E (\\boldsymbol {x})}{\\partial \\boldsymbol {x}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 450, + 903, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The deformable body dynamics is governed by Newton's Second Law as", + "bbox": [ + 511, + 491, + 906, + 520 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d ^ {2} \\boldsymbol {x}}{d t ^ {2}} = \\boldsymbol {M} ^ {- 1} \\left(\\boldsymbol {f} _ {\\text {i n t}} (\\boldsymbol {x}) + \\boldsymbol {f} _ {\\text {e x t}} (\\boldsymbol {x})\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 529, + 903, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which can be solved numerically. Here $M$ is the mass matrix that represents the masses of all vertices and $\\pmb{f}_{\\mathrm{ext}}$ refers to external forces, such as gravity.", + "bbox": [ + 511, + 569, + 906, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Physics-based Animation", + "text_level": 1, + "bbox": [ + 511, + 625, + 743, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate dynamic animations from a single illustration, a straightforward approach is to use existing controllable video diffusion models [55, 67]. However, these purely data-driven methods often struggle due to a lack of physical understanding, leading to unrealistic results. To overcome this limitation, we incorporate physics-based animation to generate physically consistent and plausible motions.", + "bbox": [ + 511, + 647, + 906, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometry Registration The first step to animate objects of interest in a given illustration is to establish a geometric basis. While previous works [24, 59] have explored predicting 3D geometry from single-view images, these methods are typically constrained to reconstructing real-world objects and do not generalize to in-the-wild anime images, since characters and objects in anime often lack an inherent 3D representation due to their stylized and flat nature. As shown in prior works on cartoon", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10796", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "animation [31, 69], leveraging 2D animation techniques proves effective for generating anime-style dynamic effects. Inspired by this,", + "bbox": [ + 89, + 90, + 480, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we focus on extracting 2D meshes for the objects of interest. To achieve this, we first utilize the Segment Anything Model (SAM) [33, 52] to", + "bbox": [ + 89, + 137, + 233, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/140153bd5c609ced0978fbb8b624bc67b512235fdf3ee7991c891825297bd262.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 138, + 480, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "obtain segmentation masks for each target object, guided by user-specified query points. Along the contours of each segmentation mask, we uniformly sample boundary points, defining the outline of the intended mesh. We then employ conforming Delaunay triangulation [37] with these boundary constraints to generate well-structured triangular meshes, creating a simulation-ready format for subsequent animation.", + "bbox": [ + 89, + 258, + 483, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deformable Body Model Anime scenes often feature dynamic environmental effects such as external forces that interact with characters' clothing, hair, or other elements. Additionally, techniques like \"squash\" and \"stretch\" are commonly used to convey motion and energy, adding expressiveness to animated objects. Motivated by these stylized dynamics, we model anime objects as deformable bodies, allowing them to capture the fluidity and exaggerated motion characteristic of anime. A key property of deformable bodies is their ability to change shape in response to external forces and attempt to return to their original rest shape upon deformation. To represent this physical behavior, we employ the Fixed Corotated Constitutive model [57], which defines the energy density as:", + "bbox": [ + 89, + 393, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi (\\boldsymbol {F}) = \\mu \\| \\boldsymbol {F} - \\boldsymbol {R} \\| _ {F} ^ {2} + \\frac {\\lambda}{2} (\\det (\\boldsymbol {F}) - 1) ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 617, + 482, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu, \\lambda$ are the Lamé parameters, and $\\pmb{R}$ is the rotational part of $\\pmb{F}$ computed via polar decomposition. The first term models the stretching and compression resistance for the individual spatial directions and the second term describes resistance to volume change. The resulting internal forces $\\pmb{f}_{\\mathrm{int}}$ can then be derived via Eq. (5).", + "bbox": [ + 89, + 657, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Interactive Animation While $f_{\\mathrm{int}}$ governs the inherent physical behavior, $f_{\\mathrm{ext}}$ allows for user-defined interactions. Inspired by the concept of energy strokes [15, 69], we introduce customizable energy strokes that carry flow particles. These particles move along the user-specified strokes, propagating external forces to nearby vertices of the deformable mesh. This enables the creation of tailored animation effects, such as simulating wind interactions. Additionally, we incorporate rigging point support, allowing animators to", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "anchor specific regions or guide them along predefined trajectories, offering enhanced control and flexibility. Using the specified energy strokes and deformable modeling, we solve the motion equation (Eq. (6)) to evolve the dynamics. We employ the semi-implicit Euler method to compute the deformation map $\\phi_i(X,t)$ at each time step $t$ for every triangle. For each pixel $\\mathbf{X}_p$ in the reference image, if it lies within a triangle $\\mathcal{T}_i$ , we assign the displacement vector as $\\pmb{d}(\\pmb{p}) = \\phi_i(\\pmb{X}_p,t) - \\pmb{X}_p$ ; otherwise, we assign a zero vector. Collecting displacement vector $\\pmb{d}(\\pmb{p})$ of all pixels yields the optical flow $\\mathcal{F}_{0\\rightarrow t}$ , which defines the displacement fields for the reference illustration $I_0$ at time $t$ .", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Generative Rendering", + "text_level": 1, + "bbox": [ + 511, + 280, + 720, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe how to render the video sequence $\\{\\hat{I}_1,\\hat{I}_2,\\dots,\\hat{I}_T\\}$ given the reference image $I_0$ and the optical flow sequence $\\{\\mathcal{F}_{0\\to 1},\\mathcal{F}_{0\\to 2},\\ldots ,\\mathcal{F}_{0\\to T}\\}$ derived from simulation.", + "bbox": [ + 511, + 303, + 906, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sketch-Guided Rendering While the video sequence can be directly generated by warping the reference image $I_0$ using optical flow fields, the resulting frames often exhibit black hole artifacts caused by occlusions. To address this problem, we extract the sketch of $I_0$ , denoted as $S_0$ , and obtain the sketch at time $t$ as follows:", + "bbox": [ + 511, + 375, + 905, + 464 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {t} = \\mathcal {W} \\left(S _ {0}, \\mathcal {F} _ {0 \\rightarrow t}, w _ {0 \\rightarrow t}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 473, + 903, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{W}$ represents the forward-warping operator, and $w_{0\\rightarrow t}$ denotes the warping weights for each pixel in $I_0$ . In forward warping, multiple pixels may map to the same 2D location in the output frame [48], potentially leading to artifacts or distortions if the weights $w_{0\\rightarrow t}$ are not properly defined. Inspired by [42], we set the pixel weight as $w_{0\\rightarrow t}(\\pmb {p}) = \\| \\mathcal{F}_{0\\rightarrow t}(\\pmb {p})\\| _2$ , giving higher importance to pixels with larger motion, typically corresponding to foreground objects.", + "bbox": [ + 511, + 497, + 905, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we leverage a video diffusion model to generate the rendered frames $\\{I_1^r, I_2^r, \\dots, I_T^r\\}$ using the obtained sketch sequence $\\{S_1, S_2, \\dots, S_T\\}$ . The diffusion model uses the reference image $I_0$ as an input, and we employ a ControlNet [76] with the sketches as control signals, guiding the generation process to ensure that the results align with the sketch inputs. We observed that during the inference time, due to segmentation inaccuracies, the optical flow warping may introduce unintended distortions, or miss parts of the object contour, generating imperfect sketches. To address this, we apply Gaussian blur to the input sketches at both training and inference time, which smoothes out inconsistencies. The video diffusion model is then capable of refining the results, leveraging its generative capabilities to handle imperfections and produce coherent outputs.", + "bbox": [ + 511, + 633, + 906, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Complementary Dynamics Unlike motions in the real world, dynamic effects in animation do not strictly adhere to", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10797", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "physical laws and can not be fully captured by 2D animation methods. In the industrial animation pipeline, artists typically begin by creating a series of keyframes that define the primary motion trajectory, followed by drawing in-between frames to ensure smooth and fluid transitions. Inspired by this workflow, we leverage a data-driven module to enhance the physics-based animated results. Specifically, we select keyframes from the sketch-guided rendering results, forming the keyframe sequence $\\{I_0^r, I_n^r, I_{2n}^r, \\dots, I_{in}^r\\}$ , where $n$ denotes the gap between keyframes. We then employ a cartoon interpolation video diffusion model [70], which synthesizes intermediate frames $\\{\\hat{I}_{in+1}, \\hat{I}_{in+2}, \\dots, \\hat{I}_{i(n+1)-1}\\}$ between each adjacent keyframe pair $(I_{in}^r, I_{i(n+1)}^r)$ . For keyframes, we simply assign $\\hat{I}_{in} = I_{in}^r$ . This approach allows us to introduce expressive, data-driven complementary dynamics that go beyond what can be achieved through physics-based animation alone.", + "bbox": [ + 89, + 90, + 480, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 367, + 222, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct a comprehensive comparison of our method against existing video diffusion models and demonstrate that our approach generates high-quality and physically plausible animations.", + "bbox": [ + 89, + 393, + 482, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details We implement our deformable body simulator using Taichi [25]. During interactive animation, users can adjust the Lamé parameters $\\mu$ and $\\lambda$ to control the characteristics of the objects based on their specific needs. For generative rendering, we follow LVCD [26] to train a sketch-guided ControlNet for the stable video diffusion model [4], using blurred sketches as control signals to address potential imperfections in the predicted sketch sequence. When utilizing ToonCrafter [70] for generating in-between frames, similar to its original implementation, we also train an additional sketch-guided ControlNet, but set the control scale to 0.1 during inference. We also set $n = 15$ for the in-betweening step. This configuration helps the generated frames follow the rough motion indicated by the sketches while introducing extra stylized animation details.", + "bbox": [ + 89, + 460, + 482, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset We build our training dataset using the Sakuga-42M Dataset [50], which consists of 1.4 million animation video clips. To ensure high-quality training samples, we filter out clips with fewer than 24 frames and dynamics scores outside the range of 0.05 to 0.7. Next, we extract sketches from the selected clips using the method from [5], resulting in a final dataset of 380,000 pairs of sketches and corresponding video sequences.", + "bbox": [ + 89, + 710, + 482, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline We compare our method against two categories of video generation approaches. The first category includes state-of-the-art Image-to-Video (I2V) models, such as Cinemo [46] and DynamamiCrafter [71], which take an input im", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "age and use text prompts to guide the motion dynamics of the generated videos. For these models, we generate prompts describing the image content and expected motion using ChatGPT-4V. The second category, including Motion-I2V [55] and Drag Anything [67], usually trains an additional motion-control module for video generative models such that, given a single image input, a user-specified trajectory can be provided to control the movement of objects in the generated frames. To ensure a fair comparison, we extract trajectories from our animated results and use them as input for these methods.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 270, + 733, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the absence of established benchmarks for animestyle image-to-video generation, we follow Motion-I2V [55] and construct a test set comprising 20 anime images with stylized dynamic elements such as swaying hair, clothing, and plants. For each method, we generate 10 video samples per image, resulting in a total of 200 videos per method. To quantitatively evaluate the generated results, we use the Fréchet Inception Distance (FID) [22] to measure the similarity between the generated frames and the reference images. Additionally, we employ VideoScore [19] to assess multiple aspects of video quality, including visual quality, temporal consistency, dynamic degree, and factual consistency. The factual consistency evaluates the consistency of the video content with common sense and factual knowledge. We exclude the text alignment score as our method and the trajectory-controlled approach do not involve text input.", + "bbox": [ + 511, + 295, + 903, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative evaluation results, presented in Tab. 1, demonstrate that our method outperforms baseline approaches across most metrics, including visual quality, temporal consistency, and factual consistency, demonstrating the overall high quality of our generated videos. As shown in Fig. 3, Cinema tends to generate static videos that closely resemble the reference image, leading to better FID scores. While the dynamic degree score for DragAnything is notably higher than other methods, this is due to its tendency to misinterpret motion control as camera movement, causing shifts in the entire image space and resulting in an inflated dynamics score. Although our method's dynamic degree score is slightly lower than that of DynamiCrafter and Motion-I2V, these methods often exhibit implausible, distorted motions that create an illusion of increased dynamics. In contrast, our approach produces geometry-consistent motions and achieves high factual consistency, demonstrating the realistic motion dynamics enabled by our physics-based modeling. In addition, following [14, 34], we conducted an user study adopting a two-alternative forced choice (2AFC) protocol, where participants are asked to choose the preferred video based on temporal consistency, visual quality, motion plausibility, and overall feeling given", + "bbox": [ + 511, + 554, + 903, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10798", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/72996d9e6b9751afae14e0bb6aeb4463b985e19af012218324573ac3d0587af2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 80, + 210, + 128 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"A pirate flag flutters and waves vigorously in the wind.\"", + "bbox": [ + 218, + 88, + 496, + 119 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f6a3c80c2fc9498f6da9c46bc693d9bcb2ab8161a6d3403689f1b3c652b4d50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 79, + 607, + 128 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"The figure's hair flows gently with the breeze, strands swaying slightly.\"", + "bbox": [ + 611, + 88, + 875, + 122 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/404d853408afd38d611ea5f6b88aa006c8355ef4e9410a93b187c9d0c1710d1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 130, + 204, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc796bb3fda26fafe75e51079a48d33857fbc2b314c053260d53087b9faf919c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 130, + 302, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6fd15c76e952ce857490e897b6c7471314099af6fcf3a3220929bbfe951049f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 130, + 398, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82bb8f4e5333ca9eaf049c74d0c9edde9da2cd0d7c95121161c9d46f97344060.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 130, + 496, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24e88cc06546baeffba54954e68a91f4fa1173bf60f5393a63c08ea4cac07f8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 178, + 202, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04f3731173654c475ffb7fd9c30f2c1809341d9503951d44ef5af64a7bb35b0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 178, + 290, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/58c14f62647bcb91aa478a1bce85d8a15b11e36600ba8e5ae618c8e473bdfef1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 178, + 387, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddd92798a4ac91f99ac2850046ba3dfacf59eabd18b51c2226974f853b224327.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 178, + 496, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc33df2cfd66b29ecc6ccc76b6cb04203cbd81a62b3cec85e3c3d616d69ef6db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 229, + 204, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/30ac7c1cdb46509abfd92cc5cc9e7874aca21c1741866f285908461e8c9623da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 229, + 300, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5ec722bf208bce94f84944ac5248afaf92946df1149a95b829ceccdbc55dfb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 229, + 398, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05eb25d16d18b7c154fe6faf6ea1622b1ed77cfda47ea7e625b36ad869aeb1a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 229, + 496, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4bdc00d6c2ba77f25faa1eb3aae90c99a34fd37db77316ae68b160f8b86d2b4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 277, + 202, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15bc45417c47c863f62828292f6929272dbaccd8340dc48c75e2cbc2ac818e48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 279, + 300, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d870129f6963e724fc525220cb14d7525dde5a948e309d169f3248abeddd0f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 281, + 398, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9efaba4201d08f3f6d23866deb648175b56de9b41caff8efcbbdbcec1f6e45da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 280, + 496, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8b6606a892fd79f847cb649b378777cc8f78fdcdc9ebc911a1d6a242f9f5d9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 327, + 202, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3cd1cc252d0646a37f9d5f726694b71ed9f762d0fa82637ffc14fc67db9f1e04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 327, + 299, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34033b9d0993fcc5d3851566a841208ec90db1ac421575d44d6cb4cb9dcef814.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 327, + 397, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd5d0372f9e2e433499231924428215ff8d1079a59a1ef2d5ec3fab5049a4cdc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 327, + 496, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2779d2521a555952da799825be49094fb68ed374e8c3493021ece3d5b1f0a5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 130, + 599, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c945aa6c93950d127da908ea999fe2043baec3ea022b90228127cd398f03a29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 130, + 702, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c6471f274fc50fae62918089bf9301471bc8c5ce0ac9aaf3c3ee042f286c4de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 130, + 898, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68eb0c96e615cf2f829a969d738d1ce61702c94346109cd2eb24e6c4d14966c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 179, + 599, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31e80205d0e1c6f9453dab615a0ebedd0d5f054a7d854eb940505391c2420fca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 180, + 702, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a60bec4bdfcf020ae33ac6bc66cff6f159935fb56f7442c9efa6a6f92151b15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 180, + 898, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d71bcfe7097d17daf22ec939e35a84ef9cb4fa2026761227c83d560e575b562e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 228, + 599, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42f532c90fa66c9ae9d663fb821cdb2fdb2f07cce3a9f753c94df3f2c126fea0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 228, + 702, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3ee0012edc3b554001c934479da90ba48898af2cdeff2a7f9f34ff991d99ee5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 228, + 898, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e5f70c9c586a0c7f39093e324545c8d734b3f7079d5ee38c6a0da7472e863766.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 277, + 599, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a84de4bc1ad003696ce8d6ddde21e141cd24f1f6eee9ff95375c9a712039ecdb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 277, + 702, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/09ac87a5b6d11e3481aed0e7a6772231af811eb2884e0169dd48c4dd05d1c15e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 277, + 898, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2b2050c21d68f9ba72c5d18e4e6f9b7ef120224f12541b0ba74a3daa4348c63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 328, + 599, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ac3123f6415d2ec9ea771ea80ea05fd29678f41826e762891adde5ce141275b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 328, + 702, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50b29a0b6d550058c26ecee0a4cc7807b7c53a9b112d3d49fd997ca74964e8c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 328, + 898, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b6074ffcc6c1665a9ca40dba2f71b094d9c2bd940ece7ecd026f14af502750d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 377, + 215, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"A scarecrow stands in a field, swaying slightly as a gentle breeze rustles through its clothes.\"", + "bbox": [ + 218, + 387, + 495, + 419 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b27a1322749fa13ca2322da0c18f5cfd0330c808ce6275281baef84a28412f23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 377, + 607, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "\"A street sign and a large, friendly creature stand nearby, both swaying gently in the rain.\"", + "bbox": [ + 611, + 386, + 893, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab28d824abeac796127afd59b5c3139665748de17c0749402b75076f447be00e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 426, + 215, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1aa985894f2d540fcce6ed3b7a95c963278504213fb3b0c016d93e007894804.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 426, + 310, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/70ad98a5867db3df8a218e30cd96e8ed86f64fa91c47c32ee5d0509f83ca775e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 426, + 410, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7372c880b318355611d45da8c3256d28e8e1ff81e3135db2469b53454fc05bc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 426, + 506, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89fa153e5a2c5225cce4d8c73389f941b3c6f9d7bb1b89657366c13db8a90ff5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 426, + 607, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02aff4db14645901808384d9bbe39b1d672b1325e01fe7572a59a99c43140060.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 426, + 898, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5988bfca8895b7562a3d395bd776ddf39cdfcc007248eb704cad1d4caa9337c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 474, + 215, + 522 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3095b5322e89707f20e3e00a41f6047b16672805318b314beaa5cc941a605567.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 474, + 310, + 522 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd999b135e1d16eb44dd3bd848d842855dd4417a693cfae046669936289ab792.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 474, + 410, + 522 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20087750fe1f32460c5b0270fc5a2bc590a1e7da4af0ba609eed95670f74aa10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 474, + 506, + 522 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a992addcd2388e8719edf2da4f5a4df2965548abd1151f0a8c5b5af8761977d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 474, + 607, + 522 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c7b0b0c9ca6566e87b5bbbe32c83883bc36550cb84cf78432f90cbf7f05f1012.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 474, + 898, + 522 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b54ad140c594bb679f45ca47921d96c495f69a199c999b52823558c9001d46c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 523, + 215, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f267bc53ee02b69f23bed4570d02ad273a8dece970a555cd29d534aa03b8d106.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 523, + 310, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/809e411d6d5f68731831df46d1afbacfb6bb62fe4afca57273b98e89fa23cfaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 523, + 410, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66adc13db0aa5eafc456e877e9e6edc83b31eb403237185a5412c0c60b45a3df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 523, + 506, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57c8c81b1540a3209285830cfdaba4a2fd2200f22e1af1d4a77e233f061f9a62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 523, + 607, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/262454a7bc28d875471909c09aeac4dba219e971ba9d9ec0d151e0d8b3773b55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 523, + 898, + 571 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5843402af0ea4a17d5e638a9695913c6f0738b4d372961f9ed88d702d90ceec9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 574, + 215, + 622 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e17be56dc2ac6d4b18210dd2be4a755cd5ba0e4ae0eea794ef5795a5eb15a2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 574, + 310, + 622 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bed9f7bcb369e182c45cd8cfe6670b85cdadb5928c626c33c7e1d7df704b251.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 574, + 410, + 622 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/09277e53bf282a151fee93d4bf8af13c928d12fd84999866f72359970a6f065b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 574, + 506, + 622 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e753d89bdaeadaa816e33b8de8c04c75552a8cd593094ed40823f2aebc0f3ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 574, + 607, + 622 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84f9629ad708f9caf9ff666de7122d7c61dbd6354f5034ee61add850e0534ecc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 574, + 898, + 622 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84d61bf8e63f112c823eb89a2c4bb63a72717cc774429f12d0d9785fe8b303b2.jpg", + "image_caption": [ + "Figure 3. Qualitative Comparison. We compare our results against Cinemo [46], Drag Anything [67], DynamiCrafter [71] and Motion-I2V [55]. Text prompts for Cinemo and DynamiCrafter are generated using ChatGPT-4V, while trajectories for Drag Anything and Motion-I2V are extracted from our animated results." + ], + "image_footnote": [], + "bbox": [ + 116, + 625, + 215, + 671 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a61c30fb2e68d46c4d2f28ed133b0010146ecdabf110f16002b5236d3f0cce4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 625, + 310, + 671 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/58d3eb698fd4c6e43b15ce569c685f321c80855db8d6be23517971fecbca2ef2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 625, + 410, + 671 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17035d8104ef2956c454728813ff96a921669a0cf761a70eae26494a1b4dceb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 625, + 506, + 671 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/917e34fdf95927c94b74507b33529d1411abed38bc33985ec7bf411ca4fa18a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 625, + 607, + 671 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7662a5d73919ed2783883e5b2e9f79ab8c44d3e37a5bdfb0e7aa7353776443cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 625, + 898, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "two videos (one from our method and one from baselines). The user preference results, presented in Tab. 2, show that our proposed method consistently outperforms the baselines across all evaluation criteria.", + "bbox": [ + 89, + 741, + 482, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 89, + 815, + 312, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present a qualitative comparison with baseline methods in Fig. 3. Cinemo often produces static results with minimal motion dynamics. Drag Anything frequently misinterprets the motion trajectory as a camera movement, resulting in", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "unintended dynamic sequences. DynamiCrafter, while generating larger motions, struggles to maintain the geometry of input objects, leading to noticeable distortions. Motion-I2V has difficulty preserving the appearance of the input content, often yielding unsatisfactory results. In contrast, our physics-guided approach ensures both physically plausible motions and high-quality rendering, preserving the geometry and visual consistency of the input.", + "bbox": [ + 511, + 741, + 906, + 864 + ], + "page_idx": 6 + }, + { + "type": "aside_text", + "text": "Ours MotionI2V DC DA Cinemo Input Ours MotionI2V DC DA Cinemo Input", + "bbox": [ + 96, + 93, + 112, + 659 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10799", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6371ea29af5c5caf71b2c3e21ce2fd6d0138fd144f3d03a6f7b946bb9625632d.jpg", + "table_caption": [ + "Table 1. Quantitative Comparisons. We report FID to evaluate the similarity between the reference image and generated frames. VSVQ, VSTC, VSDD, and VSFC represent scores for visual quality, temporal consistency, dynamic degree, and factual consistency, respectively, as measured by VideoScore [19]." + ], + "table_footnote": [], + "table_body": "
MethodsFID↓VSVQ↑VSTC↑VSDD↑VSFC↑
Cinemoto [46]49.52.852.802.422.58
DragAnything [67]148.92.772.452.972.52
DynamiCrafter [71]94.92.782.682.532.51
Motion-I2V [55]121.82.702.502.662.39
Ours90.42.892.862.482.64
", + "bbox": [ + 98, + 162, + 473, + 252 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e8a4b7f5f93b3b4de528b79e3f3c1236cc0b6fb53b51c35ab2ea5351160bbf7f.jpg", + "table_caption": [ + "Table 2. User Study. We show the user preference for our method over the baseline methods in terms of visual quality (VQ), temporal consistency (TC), motion plausibility (MP), and overall feeling." + ], + "table_footnote": [], + "table_body": "
MethodsVQTCMPOverall
Cinemoto [46]86%83%82%81%
DragAnything [67]93%91%89%91%
DynamiCrafter [71]84%78%76%81%
Motion-I2V [55]95%94%97%96%
", + "bbox": [ + 112, + 314, + 457, + 390 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd091db1c3ae36cc9b4a5d8b67a436049ce8326a26cd689ee96ff8ce3d4d991f.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 94, + 405, + 220, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5ac7921e75f843573eed7141de0254603d0b0638db663b16a7ab4d5180fc883.jpg", + "image_caption": [ + "Warp" + ], + "image_footnote": [], + "bbox": [ + 222, + 405, + 349, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5bf0de7aaae797c51a605607d18d9754bc5ca467cb9a2f604f3056b1d7789fa9.jpg", + "image_caption": [ + "Warp + Hard Inpaint" + ], + "image_footnote": [], + "bbox": [ + 354, + 405, + 480, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09726d7a8718fa4ffa5f5190d08cdab74f9749af992288ba1daa8e2f0b7d1039.jpg", + "image_caption": [ + "Warp + Soft Inpaint", + "Figure 4. Sketch-Guided Rendering. Applying warping and inpainting introduces artifacts due to segmentation inaccuracy. Soft-inpainting [39] reduces these issues but can alter the content. Our sketch-guided rendering method produces high-quality results while preserving image details." + ], + "image_footnote": [], + "bbox": [ + 222, + 479, + 349, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85966a5bd34733c68c565dcf33d35234e86229de7e110abfa21ce73b4c792d6d.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 354, + 479, + 480, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Additional Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 664, + 359, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sketch-Guided Rendering After physics-based animation in our proposed framework, we obtain a sequence of optical flow $\\{\\mathcal{F}_{0\\to 1},\\mathcal{F}_{0\\to 2},\\dots ,\\mathcal{F}_{0\\to T}\\}$ . A straightforward approach would be to warp the input image $I_{0}$ using these optical flows and apply an inpainting algorithm to fill any occluded areas. However, this often results in suboptimal outputs. The main issue lies in segmentation inaccuracies, which can cause unintended regions to be warped or leave parts of the intended objects static, as shown in Fig. 4. While introducing a soft mask [39] into the inpainting process can help smooth the transition between the newly generated region and the original figure, it may also alter the content, making it diverge from the appearance of the reference image. Additionally, applying inpainting frame by", + "bbox": [ + 88, + 688, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/649990fdea61fcdb48bf9a9aedfc6332d7da33fe520d8e1f192348d31d392642.jpg", + "image_caption": [ + "Static Reference" + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 643, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84c44ba90a0a06041d57536943a5d89267f44585d43ea7b75caf193469a2aa55.jpg", + "image_caption": [ + "w/o Enhanced Dynamics", + "Figure 5. Complementary Dynamics Enhancement. While physics-based animation maintains geometric consistency, it may lack the fluidity and exaggeration commonly seen in anime. We employ a data-driven interpolation module to enhance the motion dynamics, creating more natural-looking animations that better resemble real anime." + ], + "image_footnote": [], + "bbox": [ + 651, + 90, + 776, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1ddccb858cb8dbf3c634553f15a094887f19fa1d4c14d986310ede26f6feb604.jpg", + "image_caption": [ + "w/ Enhanced Dynamics" + ], + "image_footnote": [], + "bbox": [ + 779, + 90, + 903, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "frame may also cause temporal consistency issues. To address this, our method employs a sketch-guided rendering module. We first extract a sparse geometric representation of the image and apply the animation dynamics directly to this sketch-based structure. This sparse representation is more robust to segmentation inaccuracies and helps preserve the intended motion. Our subsequent rendering module then synthesizes high-quality video sequences from the animated sketches, maintaining both temporal consistency and visual fidelity.", + "bbox": [ + 511, + 282, + 906, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Complementary Dynamics Enhancement While our physics-based animation ensures physically accurate and geometry-consistent motion, it is inherently limited to 2D representations and cannot fully capture 3D effects. Moreover, the exaggerated dynamics often seen in anime do not always conform to strict physical laws. To address these limitations, we incorporate a data-driven cartoon interpolation module [70] to introduce complementary dynamics. As illustrated in Fig. 5, this interpolation module enables fluid contour deformations of clothing during motion, rather than rigidly adhering to static geometry. It also dynamically generates new sketch lines, enhancing the expressiveness of the animation and bringing it closer to the aesthetic of traditional hand-drawn anime.", + "bbox": [ + 511, + 439, + 908, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented PhysAnimator, a novel framework for generating dynamic and stylized animations from static anime illustrations by integrating physics-based simulations with data-driven generative models. Our approach generates physically plausible, fluid and exaggerated animestyled motion through deformable body simulations, providing controllability to users through user-guided energy strokes. A sketch-guided video diffusion model ensures high-quality, temporally consistent frames, while a data-driven anime frame interpolation model adds expressive, non-physical dynamics. The experiments show that our proposed method outperforms existing video diffusion methods, offering a powerful tool for creating visually compelling and user-controllable anime-style animations.", + "bbox": [ + 509, + 689, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10800", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jernej Barbic, Marco da Silva, and Jovan Popovic. Deformable object animation using reduced optimal control. In ACM SIGGRAPH 2009 papers, pages 1-9. 2009. 3", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2", + "[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 6", + "[5] Caroline Chan, Fredo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. 2022. 6", + "[6] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv:2305.13840, 2023. 2", + "[7] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 3", + "[8] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv:2403.06738, 2024. 2", + "[9] Stelian Coros, Sebastian Martin, Bernhard Thomaszewski, Christian Schumacher, Robert Sumner, and Markus Gross. Deformable objects alive! ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 3", + "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[11] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 2" + ], + "bbox": [ + 93, + 246, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "We thank anonymous reviewers for their insightful comments. We acknowledge support from NSF 2153851. We would like to thank Rahul Garg, Hossein Taghavi, Roshni Cooper, Ritwik Kumar, Boris Chen, Amli Murphy, Oliver Banasiak and Taiki Sakurai for their valuable suggestions and support.", + "[12] Rinon Gal, Yael Vinker, Yuval Alaluf, Amit Bermano, Daniel Cohen-Or, Ariel Shamir, and Gal Chechik. Breathing life into sketches using text-to-video priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4325–4336, 2024. 2", + "[13] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 2", + "[14] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arxiv:2307.10373, 2023. 6", + "[15] Joseph Gilland. Elemental Magic, Volume 2: The Technique of Special Effects Animation. Routledge, 2012. 3, 5", + "[16] Yuchao Gu, Yipin Zhou, Bichen Wu, Licheng Yu, Jia-Wei Liu, Rui Zhao, Jay Zhangjie Wu, David Junhao Zhang, Mike Zheng Shou, and Kevin Tang. Videoswap: Customized video subject swapping with interactive semantic point correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7621-7630, 2024. 2", + "[17] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 2, 3", + "[18] Junlin Han, Filippos Kokkinos, and Philip Torr. Vfusion3d: Learning scalable 3d generative models from video diffusion models. In European Conference on Computer Vision, pages 333-350. Springer, 2025. 2", + "[19] Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, et al. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. arXiv:2406.15252, 2024. 6, 8", + "[20] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv:2211.13221, 2022. 2", + "[21] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. animate-a-story: Storytelling with retrieval-augmented video generation. arXiv:2307.06940, 2023. 2", + "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[23] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2", + "[24] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv:2311.04400, 2023. 4" + ], + "bbox": [ + 93, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10801", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG), 38(6): 201, 2019. 6", + "[26] Zhitong Huang, Mohan Zhang, and Jing Liao. Lycd: Reference-based linear video colorization with diffusion models. arXiv:2409.12960, 2024. 6", + "[27] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7341-7351, 2024. 2", + "[28] Hyeonho Jeong, Geon Yeong Park, and Jong Chul Ye. Vmc: Video motion customization using temporal attention adaptation for text-to-video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9212-9221, 2024. 2, 3", + "[29] Ben Jones, Jovan Popovic, James McCann, Wilmot Li, and Adam Bargteil. Dynamic sprites: artistic authoring of interactive animations. Computer Animation and Virtual Worlds, 26(2):97-108, 2015. 3", + "[30] Ben Jones, Nils Thuerey, Tamar Shinar, and Adam W Bargteil. Example-based plastic deformation of rigid bodies. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. 3", + "[31] Rubaiat Habib Kazi, Tovi Grossman, Nobuyuki Umetani, and George Fitzmaurice. Motion amplifiers: sketching dynamic illustrations using the principles of 2d animation. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, pages 4599-4609, 2016. 2, 3, 5", + "[32] Diederik P Kingma. Auto-encoding variational bayes. arXiv:1312.6114, 2013. 3", + "[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5", + "[34] Nicholas Kolkin, Jason Salavon, and Gregory Shakhnarovich. Style transfer by relaxed optimal transport and self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10051-10060, 2019. 6", + "[35] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv:2403.14468, 2024. 2", + "[36] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 2", + "[37] Der-Tsai Lee and Bruce J Schachter. Two algorithms for constructing a delaunay triangulation. International Journal of Computer & Information Sciences, 9(3):219-242, 1980. 5", + "[38] Minhyeok Lee, Suhwan Cho, Chajin Shin, Jungho Lee, Sunghun Yang, and Sangyoun Lee. Video diffusion models are strong video inpainter. arXiv:2408.11402, 2024. 2", + "[39] Eran Levin and Ohad Fried. Differential diffusion: Giving each pixel its strength. arXiv:2306.00950, 2023. 8" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Minchen Li, Chenfanfu Jiang, and Zhaofeng Luo. Physics-Based Simulation. 2024. 4", + "[41] Mingxiao Li, Bo Wan, Marie-Francine Moens, and Tinne Tuytelaars. animate your motion: Turning still images into dynamic videos. arXiv:2403.10179, 2024. 3", + "[42] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 3, 5", + "[43] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv:2408.16767, 2024. 2", + "[44] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imaged-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 3", + "[45] Jiaju Ma, Li-Yi Wei, and Rubaiat Habib Kazi. A layered authoring tool for stylized 3d animations. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1-14, 2022. 3", + "[46] Xin Ma, Yaohui Wang, Gengyu Jia, Xinyuan Chen, YuanFang Li, Cunjian Chen, and Yu Qiao. Cinemo: Consistent and controllable image animation with motion diffusion models. arXiv:2407.15642, 2024. 2, 6, 7, 8", + "[47] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv:2410.05363, 2024. 3", + "[48] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5437-5446, 2020. 5", + "[49] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adaptions in frozen image-to-video diffusion model. arXiv:2405.20222, 2024. 2", + "[50] Zhenglin Pan, Yu Zhu, and Yuxuan Mu. Sakuga-42m dataset: Scaling up cartoon research. arXiv:2405.07425, 2024.6", + "[51] Mengqi Peng, Li-yi Wei, Rubaiat Habib Kazi, and Vladimir G Kim. Autocomplete animated sculpting. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 760-777, 2020. 3", + "[52] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 5", + "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + ], + "bbox": [ + 514, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10802", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv:2406.01493, 2024. 2", + "[55] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 4, 6, 7, 8", + "[56] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv:2209.14792, 2022. 2", + "[57] Alexey Stomakhin, Russell Howes, Craig A Schroeder, and Joseph M Teran. Energetically consistent invertible elasticity. In Symposium on Computer Animation, 2012. 5", + "[58] Qingkun Su, Xue Bai, Hongbo Fu, Chiew-Lan Tai, and Jue Wang. Live sketch: Video-driven dynamic deformation of static drawings. In Proceedings of the 2018 chi conference on human factors in computing systems, pages 1-12, 2018. 2", + "[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 4", + "[60] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2", + "[61] Qian Wang, Abdelrahman Eldosokey, Mohit Mendiratta, Fangneng Zhan, Adam Kortylewski, Christian Theobalt, and Peter Wonka. Zero-shot video semantic segmentation based on pre-trained diffusion models. arXiv:2405.16947, 2024. 2", + "[62] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3", + "[63] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6537-6549, 2024. 3", + "[64] Nora S Willett, Wilmot Li, Jovan Popovic, Floraine Berthouzoz, and Adam Finkelstein. Secondary motion for performed 2d animation. In Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology, pages 97-108, 2017. 3", + "[65] Nora S Willett, Rubaiat Habib Kazi, Michael Chen, George Fitzmaurice, Adam Finkelstein, and Tovi Grossman. A mixed-initiative interface for animating static pictures. In Proceedings of the 31st annual ACM symposium on user interface software and technology, pages 649-661, 2018. 3" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. Aniclipart: Clipart animation with text-to-video priors. arXiv:2404.12347, 2024. 2", + "[67] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2025. 2, 3, 4, 6, 7, 8", + "[68] Jun Xing, Li-Yi Wei, Takaaki Shiratori, and Koji Yatani. Autocomplete hand-drawn animations. ACM Transactions on Graphics (TOG), 34(6):1-11, 2015. 2", + "[69] Jun Xing, Rubaiat Habib Kazi, Tovi Grossman, Li-Yi Wei, Jos Stam, and George Fitzmaurice. Energy-brushes: Interactive tools for illustrating stylized elemental dynamics. In Proceedings of the 29th Annual Symposium on User Interface Software and Technology, pages 755-766, 2016. 3, 5", + "[70] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. arXiv:2405.17933, 2024. 2, 3, 6, 8", + "[71] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 3, 6, 7, 8", + "[72] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 2023. 2", + "[73] Shiyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3", + "[74] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089, 2023. 2", + "[75] Jiayi Eris Zhang, Seungbae Bang, David IW Levin, and Alec Jacobson. Complementary dynamics. arXiv:2009.02462, 2020.3", + "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3, 5", + "[77] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv:2406.19680, 2024. 2", + "[78] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv:2407.21705, 2024. 3", + "[79] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. Trip: Temporal residual learning with image noise prior for image-to-video diffusion models. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10803", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 8671-8681, 2024. 2", + "[80] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv:2310.08465, 2023. 3" + ], + "bbox": [ + 91, + 90, + 483, + 176 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10804", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/d8df7b02-5bad-4fed-8afa-940086807a3d_model.json b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/d8df7b02-5bad-4fed-8afa-940086807a3d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..83a7eda8e69ea69bcd999daac9af28c6451a2807 --- /dev/null +++ b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/d8df7b02-5bad-4fed-8afa-940086807a3d_model.json @@ -0,0 +1,3480 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.82, + 0.152 + ], + "angle": 0, + "content": "PhysAnimator: Physics-Guided Generative Cartoon Animation" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.18, + 0.354, + 0.199 + ], + "angle": 0, + "content": "Tianyi Xie\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.18, + 0.596, + 0.199 + ], + "angle": 0, + "content": "Yiwei Zhao\\(^{1\\dagger}\\) Ying Jiang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.617, + 0.181, + 0.757, + 0.199 + ], + "angle": 0, + "content": "Chenfanfu Jiang2" + }, + { + "type": "text", + "bbox": [ + 0.426, + 0.199, + 0.573, + 0.216 + ], + "angle": 0, + "content": "1 Netflix, 2 UCLA" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.22, + 0.353, + 0.234 + ], + "angle": 0, + "content": "tianyixie77@ucla.edu" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.22, + 0.535, + 0.234 + ], + "angle": 0, + "content": "yiweiz@netflix.com" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.22, + 0.821, + 0.234 + ], + "angle": 0, + "content": "{yingjiang, cffjiang}@ucla.edu" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.272, + 0.256, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.582, + 0.252, + 0.596 + ], + "angle": 0, + "content": "Rig Force" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.272, + 0.425, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.35, + 0.425, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.426, + 0.425, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.503, + 0.425, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.272, + 0.583, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.35, + 0.583, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.426, + 0.583, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.503, + 0.582, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.272, + 0.739, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.35, + 0.739, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.426, + 0.739, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.503, + 0.739, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.272, + 0.898, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.35, + 0.898, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.426, + 0.898, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.503, + 0.898, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.609, + 0.907, + 0.639 + ], + "angle": 0, + "content": "Figure 1. PhysAnimator is a novel framework that combines physics principles with video diffusion models to generate high-quality animations from static animeillustrations, allowing users to specify external forces or rigging points for custom effects." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.65, + 0.327, + 0.667 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Creating hand-drawn animation sequences is labor-intensive and demands professional expertise. We introduce PhysAnimator, a novel approach for generating physically plausible meanwhile anime-stylized animation from static animeillustrations.Our method seamlessly integrates physics-based simulations with data-driven generative models to produce dynamic and visually compelling animations.To capture the fluidity and exaggeration characteristic of anime,we perform image-space deformable body simulations on extracted mesh geometries.We enhance artistic control by introducing customizable energy strokes and incorporating rigging point support, enabling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.907, + 0.803 + ], + "angle": 0, + "content": "the creation of tailored animation effects such as wind interactions. Finally, we extract and warp sketches from the simulation sequence, generating a texture-agnostic representation, and employ a sketch-guided video diffusion model to synthesize high-quality animation frames. The resulting animations exhibit temporal consistency and visual plausibility, demonstrating the effectiveness of our method in creating dynamic anime-style animations. See our project page for more demos: https://xpandora.github.io/PhysAnimator/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.646, + 0.846 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Dynamic visual effects are essential to the immersive quality of 2D animation. From the subtle sway of a character's hair to the fluid motion of garments in response to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.354, + 0.888 + ], + "angle": 0, + "content": "* Work done during the internship at Netflix." + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.889, + 0.246, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.354, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10793" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "wind, realistic and pleasing dynamics create a captivating visual experience. Traditionally, these effects are achieved through meticulous, hand-drawn techniques, where animators painstakingly draw each frame to bring these dynamic elements to life. This process is labor-intensive and requires not only artistic skill but also a deep understanding of natural forces and environmental interactions." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.202, + 0.487, + 0.595 + ], + "angle": 0, + "content": "To alleviate the challenges of manual animation, researchers have explored both traditional and data-driven approaches. Traditional animation tools [31, 58] provide interactive systems that assist artists in creating animated illustrations based on established principles of 2D animation. These methods often rely on user-provided stroke inputs to specify motion trajectories and utilize geometry constraints to produce deformation-based animation. However, such approaches typically assume simple inputs like lineart or drawings with separated layers, limiting their applicability to more complex, in-the-wild anime illustrations that feature intricate textures and details. In contrast, data-driven video generative models [2, 46, 71] offer a promising alternative by leveraging neural synthesis to generate dynamic effects directly from images, bypassing the need for manual sketching or layered inputs. Recent methods even enable interactive motion design, allowing users to specify object trajectories via drag-based inputs [49, 55, 67]. These approaches typically rely on predicting a sequence of optical flow fields to drive the desired motion and warp the frames accordingly. However, the quality of the generated results is often limited by inaccuracies in the predicted optical flow, which frequently exhibits artifacts due to a lack of geometric understanding and physical constraints. As a result, imprecise motion estimation can lead to noticeable distortions and unsatisfactory visual quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Recognizing the limitations of both traditional and purely data-driven methods, we introduce PhysAnimator, a novel framework that integrates physics-based animation with data-driven generative models to synthesize visually compelling dynamic animations driven by environmental forces such as wind or rigging from static animeillustrations. Our approach combines the physical consistency of simulation-based methods with the flexibility and expressiveness of generative models, overcoming the drawbacks of previous approaches. To achieve this, we model objects of interest in the anime illustration as deformable bodies, capturing the fluidity and exaggerated motion characteristic of anime. We solve the motion equations to compute a sequence of optical flow fields that represent consistent dynamics. Users can enhance these animations with customized energy strokes, defining the effects of external forces such as wind. To render the motion dynamics as high-quality frames, we first extract and warp the sketch using the optical flow fields to generate a texture-agnostic video sequence. We then apply a sketch-guided video diffu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.196 + ], + "angle": 0, + "content": "sion model to colorize the sketch sequence based on the reference illustration, ensuring stylistic coherence. Finally, to incorporate dynamic effects that cannot be easily captured by physical simulation, we employ a data-driven cartoon interpolation model, enriching the results with complementary, expressive dynamics. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.272 + ], + "angle": 0, + "content": "- We introduce a novel framework that combines physics-based simulations with data-driven generative models, specifically targeting animations driven by environmental forces and rigging controls, achieving both physical consistency and stylistic expressiveness." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.274, + 0.904, + 0.317 + ], + "angle": 0, + "content": "- We develop an image-space deformable body simulation technique that models anime objects as deformable meshes, capturing fluid and exaggerated dynamics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.904, + 0.378 + ], + "angle": 0, + "content": "- We leverage a sketch-guided video diffusion model to render simulation dynamics as high-quality frames, combined with a cartoon interpolation model to introduce additional dynamic effects beyond physical laws." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.394, + 0.656, + 0.41 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.42, + 0.731, + 0.435 + ], + "angle": 0, + "content": "2.1. Video Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.908, + 0.837 + ], + "angle": 0, + "content": "Building on the success of image diffusion models [10, 53], recent work has introduced video diffusion models to streamline video synthesis from text prompts or images, significantly reducing labor and time costs compared to traditional commercial video editing and creation tools [23, 72]. These approaches either extend pre-trained image diffusion models by incorporating temporal mixing layers in various forms [3, 11, 13, 56] or train video diffusion models with temporal layers from scratch on large-scale text-video paired datasets [20, 79]. However, using text prompts or a single image as input provides limited control over fine-grained aspects of video generation, making it challenging to define complex structural attributes such as spatial layouts, poses, and shapes [17, 74]. To allow for more detailed control over object motion and camera movements, additional guidance in the form of motion trajectories [6, 28, 55, 74], pose [77], depth [21], and optical flow [17] has been integrated into video diffusion models to produce more controllable videos. These powerful video diffusion models have also been applied to various downstream tasks, such as video editing [16, 35], image animation [12, 66, 71], video understanding [38, 54, 61], video interpolation [27, 70] and 3D reconstruction and generation [8, 18, 36, 43, 60]. Nevertheless, these data-driven approaches usually produce artifacts due to a lack of geometric understanding and physical constraints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.744, + 0.862 + ], + "angle": 0, + "content": "2.2. Physics-based Animation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Researchers have explored various methods for animation authoring. Xing et al. [68] extend local similarity tech" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "10794" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.501 + ], + "angle": 0, + "content": "niques to global similarity, enabling automatic sketch completion based on previous frames. Building on this, Peng et al. [51] propose a keyframe-based sculpting system that autocompletes user edits using an intuitive brushing interface. Willett et al. [65] tackle the challenge of textured anime images by segmenting the illustration into layers for different objects and allowing users to provide scribbles as trajectory guides. Leveraging animation principles, some works [31, 45] use geometry-based deformations to create stylized animation effects such as \"squash\" and \"stretch\". Similarly, Xing et al. [69] employ physics-based simulations and introduce energy brushes to generate elemental dynamics [15], such as smoke and fire. Other approaches [1, 9, 29] incorporate physical laws like gravity, collision, and elasticity to animate deformable characters, enabling the efficient creation of complex scenes. For instance, Jones et al. [30] introduces an example-based plasticity model based on linear blend skinning for animating the failure of near-rigid, man-made materials. Coros et al. [9] exploit two rest-posed adaptation strategies using only internal energy to animate curve, shell, and solid-based characters. Physics-based modeling has also proven effective in adding secondary dynamics [64, 75] to enhance details in rigged animations. Despite their effectiveness, these 2D animation methods typically assume simple inputs, such as sketches, or require layered separation, which limits their applicability for animating complex, in-the-wild anime illustrations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.514, + 0.345, + 0.53 + ], + "angle": 0, + "content": "2.3. Generative Image Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.485, + 0.903 + ], + "angle": 0, + "content": "With the advancement of generative models, there is growing interest in leveraging these powerful techniques to synthesize dynamic animations from static images, guided by motion features extracted from various user inputs. These inputs can be sparse, such as text prompts [7, 41, 71], trajectories [42, 55, 62, 67, 78], or camera movements [62, 73], or dense, like reference videos [28, 63, 80]. To achieve controllable dynamics, prior works often incorporate ControlNet [76] into image or video generative models during the decoding stage, utilizing motion features such as Canny edges, depth maps [17], 2D Gaussian maps [67], and optical flow maps [55]. In addition, several methods employ specialized motion fusion modules [41, 62, 78], spatial or temporal attention mechanisms [28, 63, 73], and feature injection strategies [42] to guide the generation of controllable dynamics. However, these approaches typically lack physics-based supervision, which can result in animations that violate physical laws or fail to align with user intentions [47]. Recently, PhysGen [44] integrated a rigid-body physics simulator to generate physics-consistent dynamics for foreground objects in a given image. While effective, PhysGen is limited to 2D rigid-body motions, making it unsuitable for the fluid, elastic effects commonly seen in anime, such as the waving of cloth or hair, which do not" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "conform to rigid-body dynamics. In this work, we address these limitations by leveraging deformable-body dynamics and a sketch-guided video diffusion model to create high-quality, physically consistent animation sequences that capture the fluidity and elasticity characteristic of anime." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.18, + 0.605, + 0.196 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.205, + 0.908, + 0.584 + ], + "angle": 0, + "content": "Given a reference anime illustration \\(I_0\\), our goal is to generate a sequence of stylized video frames \\(\\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_T\\}\\) with natural, user-guided motions. To achieve this, we introduce PhysAnimator, a novel framework that combines video diffusion models with physics-based animation. We start by segmenting the objects of interest in a given anime illustration and creating a 2D triangulated mesh representation. Using this geometry, we obtain dynamics through image-space deformable body simulation, generating a sequence of optical flow fields \\(\\{\\mathcal{F}_{0 \\to 1}, \\mathcal{F}_{0 \\to 2}, \\dots, \\mathcal{F}_{0 \\to T}\\}\\). To offer user control, we enable interactive inputs through customizable energy strokes for defining external forces and rigging points for specifying desired trajectories. To render dynamics as high-quality frames, we extract the sketch from the input illustration and warp it using the computed optical flow fields, yielding a sequence of dynamic sketch frames \\(\\{S_1, S_2, \\dots, S_T\\}\\). These sketches, together with the input illustration, are fed into a video diffusion model with a sketch-guided ControlNet [76], synthesizing a vivid animated sequence. Finally, an optional data-driven cartoon interpolation model [70] can be applied to enhance the anime style dynamics of the results by selecting keyframes from the animated sequence as input, yielding the final output frames \\(\\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_T\\}\\). An overview of our proposed framework is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.655, + 0.607 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.906, + 0.737 + ], + "angle": 0, + "content": "Latent Video Diffusion Model Most video diffusion models build on the latent diffusion model (LDM) framework [53], which uses Variational Autoencoder [32] (VAE) to map input images into a latent space. In this space, data is transformed into Gaussian noise via a forward diffusion process, which the model learns to reverse through denoising. In the forward diffusion process, the latent code \\( z_0 \\) is perturbed as:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.746, + 0.907, + 0.763 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.775, + 0.906, + 0.834 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_t)\\) with \\(\\beta_t\\) controlling the noise strength coefficient at step \\(t\\). The denoising model \\(\\epsilon_\\theta\\) is trained to recover \\(z_t\\) by minimizing the objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.844, + 0.905, + 0.861 + ], + "angle": 0, + "content": "\\[\nL _ {\\epsilon} = \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}; c, t) \\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "where \\(\\theta\\) denotes the learnable network parameters and \\(c\\) represents conditioning input (e.g. text prompts or images)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10795" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.912, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.33, + 0.908, + 0.402 + ], + "angle": 0, + "content": "Figure 2. Method Overview. We begin by segmenting the object and creating a triangulated deformable mesh. Physics-based simulations are then used to generate dynamic optical flow fields, with users given the option to guide the motion through customizable energy strokes (shown as orange arrows) and rigging points (shown as red dots). The extracted sketch is warped using the computed optical flow and refined with a sketch-guided video diffusion model, producing a smooth, stylized animation sequence. Optionally, a cartoon interpolation model can further be applied to enhance the animation with expressive dynamics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.426, + 0.484, + 0.503 + ], + "angle": 0, + "content": "After denoising, the VAE decoder reconstructs the latent code back into the image space. The latent video diffusion model (LVDM) extends the image LDM to videos by incorporating temporal modules to maintain temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.484, + 0.65 + ], + "angle": 0, + "content": "Deformable Body Dynamics Mathematically, the dynamics of a continuum deformable body is described by a time-dependent continuous deformation map \\( \\boldsymbol{x} = \\phi(\\boldsymbol{X}, t) \\), which maps the undeformed material space \\( \\Omega^0 \\) to the deformed world space \\( \\Omega^t \\) at time \\( t \\). The deformation gradient \\( \\boldsymbol{F} = \\frac{\\partial \\phi}{\\partial \\boldsymbol{X}} \\) encodes the local deformation such as scaling, rotation, and shearing. In the context of a discretized setting, e.g. 2D meshes, the deformation map for each triangle can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.661, + 0.483, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\phi_ {i} (\\boldsymbol {X}) = \\boldsymbol {F} _ {i} \\boldsymbol {X} + \\boldsymbol {b} _ {i}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.81 + ], + "angle": 0, + "content": "where \\(\\pmb{b}_i\\in \\mathbb{R}^2\\) accounts for translation of \\(i\\)-th triangle, and the deformation gradient \\(\\pmb {F}_i\\in \\mathbb{R}^{2\\times 2}\\) is assumed to be constant [40]. Undergoing deformation, the deformed body aims to recover its rest shape via resisting forces. Continuum mechanics model this behavior by first defining an energy density function \\(\\Psi (F(x))\\), which measures the strain energy per unit undeformed volume. The total potential energy for a deformable body is then obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.821, + 0.483, + 0.861 + ], + "angle": 0, + "content": "\\[\nE (\\boldsymbol {x}) = \\sum_ {i = 1} ^ {N} \\Psi \\left(\\boldsymbol {F} _ {i}\\right) V _ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(V_{i}\\) denotes volume of \\(i\\)-th triangle. The internal resisting force is then defined as the negative gradient of the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.427, + 0.851, + 0.442 + ], + "angle": 0, + "content": "potential energy with respect to the vertex position" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.452, + 0.905, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {\\text {i n t}} (\\boldsymbol {x}) = - \\frac {\\partial E (\\boldsymbol {x})}{\\partial \\boldsymbol {x}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.492, + 0.907, + 0.521 + ], + "angle": 0, + "content": "The deformable body dynamics is governed by Newton's Second Law as" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.53, + 0.905, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\frac {d ^ {2} \\boldsymbol {x}}{d t ^ {2}} = \\boldsymbol {M} ^ {- 1} \\left(\\boldsymbol {f} _ {\\text {i n t}} (\\boldsymbol {x}) + \\boldsymbol {f} _ {\\text {e x t}} (\\boldsymbol {x})\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.616 + ], + "angle": 0, + "content": "which can be solved numerically. Here \\(M\\) is the mass matrix that represents the masses of all vertices and \\(\\pmb{f}_{\\mathrm{ext}}\\) refers to external forces, such as gravity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.625, + 0.744, + 0.641 + ], + "angle": 0, + "content": "3.2. Physics-based Animation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.907, + 0.755 + ], + "angle": 0, + "content": "To generate dynamic animations from a single illustration, a straightforward approach is to use existing controllable video diffusion models [55, 67]. However, these purely data-driven methods often struggle due to a lack of physical understanding, leading to unrealistic results. To overcome this limitation, we incorporate physics-based animation to generate physically consistent and plausible motions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Geometry Registration The first step to animate objects of interest in a given illustration is to establish a geometric basis. While previous works [24, 59] have explored predicting 3D geometry from single-view images, these methods are typically constrained to reconstructing real-world objects and do not generalize to in-the-wild anime images, since characters and objects in anime often lack an inherent 3D representation due to their stylized and flat nature. As shown in prior works on cartoon" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10796" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "animation [31, 69], leveraging 2D animation techniques proves effective for generating anime-style dynamic effects. Inspired by this," + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.235, + 0.258 + ], + "angle": 0, + "content": "we focus on extracting 2D meshes for the objects of interest. To achieve this, we first utilize the Segment Anything Model (SAM) [33, 52] to" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.139, + 0.482, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.484, + 0.378 + ], + "angle": 0, + "content": "obtain segmentation masks for each target object, guided by user-specified query points. Along the contours of each segmentation mask, we uniformly sample boundary points, defining the outline of the intended mesh. We then employ conforming Delaunay triangulation [37] with these boundary constraints to generate well-structured triangular meshes, creating a simulation-ready format for subsequent animation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Deformable Body Model Anime scenes often feature dynamic environmental effects such as external forces that interact with characters' clothing, hair, or other elements. Additionally, techniques like \"squash\" and \"stretch\" are commonly used to convey motion and energy, adding expressiveness to animated objects. Motivated by these stylized dynamics, we model anime objects as deformable bodies, allowing them to capture the fluidity and exaggerated motion characteristic of anime. A key property of deformable bodies is their ability to change shape in response to external forces and attempt to return to their original rest shape upon deformation. To represent this physical behavior, we employ the Fixed Corotated Constitutive model [57], which defines the energy density as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.618, + 0.483, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\Psi (\\boldsymbol {F}) = \\mu \\| \\boldsymbol {F} - \\boldsymbol {R} \\| _ {F} ^ {2} + \\frac {\\lambda}{2} (\\det (\\boldsymbol {F}) - 1) ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.75 + ], + "angle": 0, + "content": "where \\(\\mu, \\lambda\\) are the Lamé parameters, and \\(\\pmb{R}\\) is the rotational part of \\(\\pmb{F}\\) computed via polar decomposition. The first term models the stretching and compression resistance for the individual spatial directions and the second term describes resistance to volume change. The resulting internal forces \\(\\pmb{f}_{\\mathrm{int}}\\) can then be derived via Eq. (5)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Interactive Animation While \\( f_{\\mathrm{int}} \\) governs the inherent physical behavior, \\( f_{\\mathrm{ext}} \\) allows for user-defined interactions. Inspired by the concept of energy strokes [15, 69], we introduce customizable energy strokes that carry flow particles. These particles move along the user-specified strokes, propagating external forces to nearby vertices of the deformable mesh. This enables the creation of tailored animation effects, such as simulating wind interactions. Additionally, we incorporate rigging point support, allowing animators to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "anchor specific regions or guide them along predefined trajectories, offering enhanced control and flexibility. Using the specified energy strokes and deformable modeling, we solve the motion equation (Eq. (6)) to evolve the dynamics. We employ the semi-implicit Euler method to compute the deformation map \\(\\phi_i(X,t)\\) at each time step \\(t\\) for every triangle. For each pixel \\(\\mathbf{X}_p\\) in the reference image, if it lies within a triangle \\(\\mathcal{T}_i\\), we assign the displacement vector as \\(\\pmb{d}(\\pmb{p}) = \\phi_i(\\pmb{X}_p,t) - \\pmb{X}_p\\); otherwise, we assign a zero vector. Collecting displacement vector \\(\\pmb{d}(\\pmb{p})\\) of all pixels yields the optical flow \\(\\mathcal{F}_{0\\rightarrow t}\\), which defines the displacement fields for the reference illustration \\(I_0\\) at time \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.281, + 0.722, + 0.297 + ], + "angle": 0, + "content": "3.3. Generative Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.907, + 0.364 + ], + "angle": 0, + "content": "In this section, we describe how to render the video sequence \\(\\{\\hat{I}_1,\\hat{I}_2,\\dots,\\hat{I}_T\\}\\) given the reference image \\(I_0\\) and the optical flow sequence \\(\\{\\mathcal{F}_{0\\to 1},\\mathcal{F}_{0\\to 2},\\ldots ,\\mathcal{F}_{0\\to T}\\}\\) derived from simulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.906, + 0.465 + ], + "angle": 0, + "content": "Sketch-Guided Rendering While the video sequence can be directly generated by warping the reference image \\( I_0 \\) using optical flow fields, the resulting frames often exhibit black hole artifacts caused by occlusions. To address this problem, we extract the sketch of \\( I_0 \\), denoted as \\( S_0 \\), and obtain the sketch at time \\( t \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.474, + 0.905, + 0.49 + ], + "angle": 0, + "content": "\\[\nS _ {t} = \\mathcal {W} \\left(S _ {0}, \\mathcal {F} _ {0 \\rightarrow t}, w _ {0 \\rightarrow t}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.906, + 0.633 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}\\) represents the forward-warping operator, and \\(w_{0\\rightarrow t}\\) denotes the warping weights for each pixel in \\(I_0\\). In forward warping, multiple pixels may map to the same 2D location in the output frame [48], potentially leading to artifacts or distortions if the weights \\(w_{0\\rightarrow t}\\) are not properly defined. Inspired by [42], we set the pixel weight as \\(w_{0\\rightarrow t}(\\pmb {p}) = \\| \\mathcal{F}_{0\\rightarrow t}(\\pmb {p})\\| _2\\), giving higher importance to pixels with larger motion, typically corresponding to foreground objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.86 + ], + "angle": 0, + "content": "Next, we leverage a video diffusion model to generate the rendered frames \\(\\{I_1^r, I_2^r, \\dots, I_T^r\\}\\) using the obtained sketch sequence \\(\\{S_1, S_2, \\dots, S_T\\}\\). The diffusion model uses the reference image \\(I_0\\) as an input, and we employ a ControlNet [76] with the sketches as control signals, guiding the generation process to ensure that the results align with the sketch inputs. We observed that during the inference time, due to segmentation inaccuracies, the optical flow warping may introduce unintended distortions, or miss parts of the object contour, generating imperfect sketches. To address this, we apply Gaussian blur to the input sketches at both training and inference time, which smoothes out inconsistencies. The video diffusion model is then capable of refining the results, leveraging its generative capabilities to handle imperfections and produce coherent outputs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Complementary Dynamics Unlike motions in the real world, dynamic effects in animation do not strictly adhere to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10797" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.354 + ], + "angle": 0, + "content": "physical laws and can not be fully captured by 2D animation methods. In the industrial animation pipeline, artists typically begin by creating a series of keyframes that define the primary motion trajectory, followed by drawing in-between frames to ensure smooth and fluid transitions. Inspired by this workflow, we leverage a data-driven module to enhance the physics-based animated results. Specifically, we select keyframes from the sketch-guided rendering results, forming the keyframe sequence \\(\\{I_0^r, I_n^r, I_{2n}^r, \\dots, I_{in}^r\\}\\), where \\(n\\) denotes the gap between keyframes. We then employ a cartoon interpolation video diffusion model [70], which synthesizes intermediate frames \\(\\{\\hat{I}_{in+1}, \\hat{I}_{in+2}, \\dots, \\hat{I}_{i(n+1)-1}\\}\\) between each adjacent keyframe pair \\((I_{in}^r, I_{i(n+1)}^r)\\). For keyframes, we simply assign \\(\\hat{I}_{in} = I_{in}^r\\). This approach allows us to introduce expressive, data-driven complementary dynamics that go beyond what can be achieved through physics-based animation alone." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.368, + 0.223, + 0.385 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.483, + 0.455 + ], + "angle": 0, + "content": "In this section, we conduct a comprehensive comparison of our method against existing video diffusion models and demonstrate that our approach generates high-quality and physically plausible animations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Implementation Details We implement our deformable body simulator using Taichi [25]. During interactive animation, users can adjust the Lamé parameters \\(\\mu\\) and \\(\\lambda\\) to control the characteristics of the objects based on their specific needs. For generative rendering, we follow LVCD [26] to train a sketch-guided ControlNet for the stable video diffusion model [4], using blurred sketches as control signals to address potential imperfections in the predicted sketch sequence. When utilizing ToonCrafter [70] for generating in-between frames, similar to its original implementation, we also train an additional sketch-guided ControlNet, but set the control scale to 0.1 during inference. We also set \\(n = 15\\) for the in-betweening step. This configuration helps the generated frames follow the rough motion indicated by the sketches while introducing extra stylized animation details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.483, + 0.833 + ], + "angle": 0, + "content": "Dataset We build our training dataset using the Sakuga-42M Dataset [50], which consists of 1.4 million animation video clips. To ensure high-quality training samples, we filter out clips with fewer than 24 frames and dynamics scores outside the range of 0.05 to 0.7. Next, we extract sketches from the selected clips using the method from [5], resulting in a final dataset of 380,000 pairs of sketches and corresponding video sequences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Baseline We compare our method against two categories of video generation approaches. The first category includes state-of-the-art Image-to-Video (I2V) models, such as Cinemo [46] and DynamamiCrafter [71], which take an input im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "age and use text prompts to guide the motion dynamics of the generated videos. For these models, we generate prompts describing the image content and expected motion using ChatGPT-4V. The second category, including Motion-I2V [55] and Drag Anything [67], usually trains an additional motion-control module for video generative models such that, given a single image input, a user-specified trajectory can be provided to control the movement of objects in the generated frames. To ensure a fair comparison, we extract trajectories from our animated results and use them as input for these methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.271, + 0.735, + 0.287 + ], + "angle": 0, + "content": "4.1.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.905, + 0.552 + ], + "angle": 0, + "content": "Given the absence of established benchmarks for animestyle image-to-video generation, we follow Motion-I2V [55] and construct a test set comprising 20 anime images with stylized dynamic elements such as swaying hair, clothing, and plants. For each method, we generate 10 video samples per image, resulting in a total of 200 videos per method. To quantitatively evaluate the generated results, we use the Fréchet Inception Distance (FID) [22] to measure the similarity between the generated frames and the reference images. Additionally, we employ VideoScore [19] to assess multiple aspects of video quality, including visual quality, temporal consistency, dynamic degree, and factual consistency. The factual consistency evaluates the consistency of the video content with common sense and factual knowledge. We exclude the text alignment score as our method and the trajectory-controlled approach do not involve text input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.905, + 0.903 + ], + "angle": 0, + "content": "The quantitative evaluation results, presented in Tab. 1, demonstrate that our method outperforms baseline approaches across most metrics, including visual quality, temporal consistency, and factual consistency, demonstrating the overall high quality of our generated videos. As shown in Fig. 3, Cinema tends to generate static videos that closely resemble the reference image, leading to better FID scores. While the dynamic degree score for DragAnything is notably higher than other methods, this is due to its tendency to misinterpret motion control as camera movement, causing shifts in the entire image space and resulting in an inflated dynamics score. Although our method's dynamic degree score is slightly lower than that of DynamiCrafter and Motion-I2V, these methods often exhibit implausible, distorted motions that create an illusion of increased dynamics. In contrast, our approach produces geometry-consistent motions and achieves high factual consistency, demonstrating the realistic motion dynamics enabled by our physics-based modeling. In addition, following [14, 34], we conducted an user study adopting a two-alternative forced choice (2AFC) protocol, where participants are asked to choose the preferred video based on temporal consistency, visual quality, motion plausibility, and overall feeling given" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10798" + } + ], + [ + { + "type": "aside_text", + "bbox": [ + 0.097, + 0.094, + 0.114, + 0.66 + ], + "angle": 270, + "content": "Ours MotionI2V DC DA Cinemo Input Ours MotionI2V DC DA Cinemo Input" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.081, + 0.211, + 0.129 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.089, + 0.498, + 0.121 + ], + "angle": 0, + "content": "\"A pirate flag flutters and waves vigorously in the wind.\"" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.08, + 0.609, + 0.129 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.612, + 0.089, + 0.876, + 0.123 + ], + "angle": 0, + "content": "\"The figure's hair flows gently with the breeze, strands swaying slightly.\"" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.131, + 0.205, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.131, + 0.303, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.131, + 0.4, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.131, + 0.497, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.179, + 0.204, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.179, + 0.291, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.179, + 0.388, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.179, + 0.498, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.23, + 0.205, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.23, + 0.302, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.231, + 0.4, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.23, + 0.497, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.279, + 0.204, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.28, + 0.301, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.282, + 0.399, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.281, + 0.497, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.328, + 0.204, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.328, + 0.3, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.328, + 0.398, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.328, + 0.497, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.131, + 0.6, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.131, + 0.704, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.131, + 0.899, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.18, + 0.6, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.181, + 0.704, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.181, + 0.899, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.229, + 0.6, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.229, + 0.704, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.229, + 0.899, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.279, + 0.6, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.279, + 0.704, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.279, + 0.899, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.329, + 0.6, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.329, + 0.704, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.329, + 0.899, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.378, + 0.216, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.388, + 0.496, + 0.42 + ], + "angle": 0, + "content": "\"A scarecrow stands in a field, swaying slightly as a gentle breeze rustles through its clothes.\"" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.378, + 0.609, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.612, + 0.387, + 0.895, + 0.421 + ], + "angle": 0, + "content": "\"A street sign and a large, friendly creature stand nearby, both swaying gently in the rain.\"" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.427, + 0.216, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.427, + 0.311, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.427, + 0.411, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.427, + 0.507, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.427, + 0.609, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.611, + 0.427, + 0.899, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.476, + 0.216, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.476, + 0.311, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.476, + 0.411, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.476, + 0.507, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.476, + 0.609, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.476, + 0.899, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.525, + 0.216, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.525, + 0.311, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.525, + 0.411, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.525, + 0.507, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.525, + 0.609, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.525, + 0.899, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.575, + 0.216, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.575, + 0.311, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.575, + 0.411, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.575, + 0.507, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.575, + 0.609, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.575, + 0.899, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.625, + 0.216, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.625, + 0.311, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.625, + 0.411, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.625, + 0.507, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.625, + 0.609, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.625, + 0.899, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.686, + 0.905, + 0.728 + ], + "angle": 0, + "content": "Figure 3. Qualitative Comparison. We compare our results against Cinemo [46], Drag Anything [67], DynamiCrafter [71] and Motion-I2V [55]. Text prompts for Cinemo and DynamiCrafter are generated using ChatGPT-4V, while trajectories for Drag Anything and Motion-I2V are extracted from our animated results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.483, + 0.803 + ], + "angle": 0, + "content": "two videos (one from our method and one from baselines). The user preference results, presented in Tab. 2, show that our proposed method consistently outperforms the baselines across all evaluation criteria." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.816, + 0.313, + 0.833 + ], + "angle": 0, + "content": "4.2. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We present a qualitative comparison with baseline methods in Fig. 3. Cinemo often produces static results with minimal motion dynamics. Drag Anything frequently misinterprets the motion trajectory as a camera movement, resulting in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.742, + 0.907, + 0.865 + ], + "angle": 0, + "content": "unintended dynamic sequences. DynamiCrafter, while generating larger motions, struggles to maintain the geometry of input objects, leading to noticeable distortions. Motion-I2V has difficulty preserving the appearance of the input content, often yielding unsatisfactory results. In contrast, our physics-guided approach ensures both physically plausible motions and high-quality rendering, preserving the geometry and visual consistency of the input." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10799" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.483, + 0.16 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparisons. We report FID to evaluate the similarity between the reference image and generated frames. VSVQ, VSTC, VSDD, and VSFC represent scores for visual quality, temporal consistency, dynamic degree, and factual consistency, respectively, as measured by VideoScore [19]." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.164, + 0.474, + 0.253 + ], + "angle": 0, + "content": "
MethodsFID↓VSVQ↑VSTC↑VSDD↑VSFC↑
Cinemoto [46]49.52.852.802.422.58
DragAnything [67]148.92.772.452.972.52
DynamiCrafter [71]94.92.782.682.532.51
Motion-I2V [55]121.82.702.502.662.39
Ours90.42.892.862.482.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.306 + ], + "angle": 0, + "content": "Table 2. User Study. We show the user preference for our method over the baseline methods in terms of visual quality (VQ), temporal consistency (TC), motion plausibility (MP), and overall feeling." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.315, + 0.458, + 0.391 + ], + "angle": 0, + "content": "
MethodsVQTCMPOverall
Cinemoto [46]86%83%82%81%
DragAnything [67]93%91%89%91%
DynamiCrafter [71]84%78%76%81%
Motion-I2V [55]95%94%97%96%
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.406, + 0.221, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.498, + 0.186, + 0.509 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.406, + 0.351, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.469, + 0.302, + 0.48 + ], + "angle": 0, + "content": "Warp" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.406, + 0.481, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.469, + 0.473, + 0.48 + ], + "angle": 0, + "content": "Warp + Hard Inpaint" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.481, + 0.351, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.543, + 0.341, + 0.554 + ], + "angle": 0, + "content": "Warp + Soft Inpaint" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.481, + 0.481, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.543, + 0.432, + 0.553 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.637 + ], + "angle": 0, + "content": "Figure 4. Sketch-Guided Rendering. Applying warping and inpainting introduces artifacts due to segmentation inaccuracy. Soft-inpainting [39] reduces these issues but can alter the content. Our sketch-guided rendering method produces high-quality results while preserving image details." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.665, + 0.36, + 0.681 + ], + "angle": 0, + "content": "4.3. Additional Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Sketch-Guided Rendering After physics-based animation in our proposed framework, we obtain a sequence of optical flow \\(\\{\\mathcal{F}_{0\\to 1},\\mathcal{F}_{0\\to 2},\\dots ,\\mathcal{F}_{0\\to T}\\}\\). A straightforward approach would be to warp the input image \\(I_{0}\\) using these optical flows and apply an inpainting algorithm to fill any occluded areas. However, this often results in suboptimal outputs. The main issue lies in segmentation inaccuracies, which can cause unintended regions to be warped or leave parts of the intended objects static, as shown in Fig. 4. While introducing a soft mask [39] into the inpainting process can help smooth the transition between the newly generated region and the original figure, it may also alter the content, making it diverge from the appearance of the reference image. Additionally, applying inpainting frame by" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.644, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.163, + 0.62, + 0.173 + ], + "angle": 0, + "content": "Static Reference" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.091, + 0.777, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.163, + 0.775, + 0.173 + ], + "angle": 0, + "content": "w/o Enhanced Dynamics" + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.091, + 0.905, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.163, + 0.9, + 0.173 + ], + "angle": 0, + "content": "w/ Enhanced Dynamics" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.185, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Figure 5. Complementary Dynamics Enhancement. While physics-based animation maintains geometric consistency, it may lack the fluidity and exaggeration commonly seen in anime. We employ a data-driven interpolation module to enhance the motion dynamics, creating more natural-looking animations that better resemble real anime." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.907, + 0.435 + ], + "angle": 0, + "content": "frame may also cause temporal consistency issues. To address this, our method employs a sketch-guided rendering module. We first extract a sparse geometric representation of the image and apply the animation dynamics directly to this sketch-based structure. This sparse representation is more robust to segmentation inaccuracies and helps preserve the intended motion. Our subsequent rendering module then synthesizes high-quality video sequences from the animated sketches, maintaining both temporal consistency and visual fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.909, + 0.651 + ], + "angle": 0, + "content": "Complementary Dynamics Enhancement While our physics-based animation ensures physically accurate and geometry-consistent motion, it is inherently limited to 2D representations and cannot fully capture 3D effects. Moreover, the exaggerated dynamics often seen in anime do not always conform to strict physical laws. To address these limitations, we incorporate a data-driven cartoon interpolation module [70] to introduce complementary dynamics. As illustrated in Fig. 5, this interpolation module enables fluid contour deformations of clothing during motion, rather than rigidly adhering to static geometry. It also dynamically generates new sketch lines, enhancing the expressiveness of the animation and bringing it closer to the aesthetic of traditional hand-drawn anime." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.69, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We presented PhysAnimator, a novel framework for generating dynamic and stylized animations from static anime illustrations by integrating physics-based simulations with data-driven generative models. Our approach generates physically plausible, fluid and exaggerated animestyled motion through deformable body simulations, providing controllability to users through user-guided energy strokes. A sketch-guided video diffusion model ensures high-quality, temporally consistent frames, while a data-driven anime frame interpolation model adds expressive, non-physical dynamics. The experiments show that our proposed method outperforms existing video diffusion methods, offering a powerful tool for creating visually compelling and user-controllable anime-style animations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10800" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.207 + ], + "angle": 0, + "content": "We thank anonymous reviewers for their insightful comments. We acknowledge support from NSF 2153851. We would like to thank Rahul Garg, Hossein Taghavi, Roshni Cooper, Ritwik Kumar, Boris Chen, Amli Murphy, Oliver Banasiak and Taiki Sakurai for their valuable suggestions and support." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.222, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.247, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[1] Jernej Barbic, Marco da Silva, and Jovan Popovic. Deformable object animation using reduced optimal control. In ACM SIGGRAPH 2009 papers, pages 1-9. 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.291, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[5] Caroline Chan, Fredo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[6] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv:2305.13840, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[7] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[8] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv:2403.06738, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[9] Stelian Coros, Sebastian Martin, Bernhard Thomaszewski, Christian Schumacher, Robert Sumner, and Markus Gross. Deformable objects alive! ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[11] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.247, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[12] Rinon Gal, Yael Vinker, Yuval Alaluf, Amit Bermano, Daniel Cohen-Or, Ariel Shamir, and Gal Chechik. Breathing life into sketches using text-to-video priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4325–4336, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[13] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[14] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arxiv:2307.10373, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[15] Joseph Gilland. Elemental Magic, Volume 2: The Technique of Special Effects Animation. Routledge, 2012. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[16] Yuchao Gu, Yipin Zhou, Bichen Wu, Licheng Yu, Jia-Wei Liu, Rui Zhao, Jay Zhangjie Wu, David Junhao Zhang, Mike Zheng Shou, and Kevin Tang. Videoswap: Customized video subject swapping with interactive semantic point correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7621-7630, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[17] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[18] Junlin Han, Filippos Kokkinos, and Philip Torr. Vfusion3d: Learning scalable 3d generative models from video diffusion models. In European Conference on Computer Vision, pages 333-350. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[19] Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, et al. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. arXiv:2406.15252, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[20] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv:2211.13221, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[21] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. animate-a-story: Storytelling with retrieval-augmented video generation. arXiv:2307.06940, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[24] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv:2311.04400, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10801" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[25] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG), 38(6): 201, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[26] Zhitong Huang, Mohan Zhang, and Jing Liao. Lycd: Reference-based linear video colorization with diffusion models. arXiv:2409.12960, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[27] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7341-7351, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[28] Hyeonho Jeong, Geon Yeong Park, and Jong Chul Ye. Vmc: Video motion customization using temporal attention adaptation for text-to-video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9212-9221, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[29] Ben Jones, Jovan Popovic, James McCann, Wilmot Li, and Adam Bargteil. Dynamic sprites: artistic authoring of interactive animations. Computer Animation and Virtual Worlds, 26(2):97-108, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[30] Ben Jones, Nils Thuerey, Tamar Shinar, and Adam W Bargteil. Example-based plastic deformation of rigid bodies. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[31] Rubaiat Habib Kazi, Tovi Grossman, Nobuyuki Umetani, and George Fitzmaurice. Motion amplifiers: sketching dynamic illustrations using the principles of 2d animation. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, pages 4599-4609, 2016. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[32] Diederik P Kingma. Auto-encoding variational bayes. arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[34] Nicholas Kolkin, Jason Salavon, and Gregory Shakhnarovich. Style transfer by relaxed optimal transport and self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10051-10060, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[35] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv:2403.14468, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[36] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[37] Der-Tsai Lee and Bruce J Schachter. Two algorithms for constructing a delaunay triangulation. International Journal of Computer & Information Sciences, 9(3):219-242, 1980. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[38] Minhyeok Lee, Suhwan Cho, Chajin Shin, Jungho Lee, Sunghun Yang, and Sangyoun Lee. Video diffusion models are strong video inpainter. arXiv:2408.11402, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Eran Levin and Ohad Fried. Differential diffusion: Giving each pixel its strength. arXiv:2306.00950, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[40] Minchen Li, Chenfanfu Jiang, and Zhaofeng Luo. Physics-Based Simulation. 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[41] Mingxiao Li, Bo Wan, Marie-Francine Moens, and Tinne Tuytelaars. animate your motion: Turning still images into dynamic videos. arXiv:2403.10179, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.167, + 0.907, + 0.222 + ], + "angle": 0, + "content": "[42] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.224, + 0.905, + 0.279 + ], + "angle": 0, + "content": "[43] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv:2408.16767, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.905, + 0.338 + ], + "angle": 0, + "content": "[44] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imaged-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.907, + 0.396 + ], + "angle": 0, + "content": "[45] Jiaju Ma, Li-Yi Wei, and Rubaiat Habib Kazi. A layered authoring tool for stylized 3d animations. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1-14, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[46] Xin Ma, Yaohui Wang, Gengyu Jia, Xinyuan Chen, YuanFang Li, Cunjian Chen, and Yu Qiao. Cinemo: Consistent and controllable image animation with motion diffusion models. arXiv:2407.15642, 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[47] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv:2410.05363, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.528, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[48] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5437-5446, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.654 + ], + "angle": 0, + "content": "[49] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adaptions in frozen image-to-video diffusion model. arXiv:2405.20222, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[50] Zhenglin Pan, Yu Zhu, and Yuxuan Mu. Sakuga-42m dataset: Scaling up cartoon research. arXiv:2405.07425, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[51] Mengqi Peng, Li-yi Wei, Rubaiat Habib Kazi, and Vladimir G Kim. Autocomplete animated sculpting. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 760-777, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[52] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10802" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[54] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv:2406.01493, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[55] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[56] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[57] Alexey Stomakhin, Russell Howes, Craig A Schroeder, and Joseph M Teran. Energetically consistent invertible elasticity. In Symposium on Computer Animation, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[58] Qingkun Su, Xue Bai, Hongbo Fu, Chiew-Lan Tai, and Jue Wang. Live sketch: Video-driven dynamic deformation of static drawings. In Proceedings of the 2018 chi conference on human factors in computing systems, pages 1-12, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[60] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[61] Qian Wang, Abdelrahman Eldosokey, Mohit Mendiratta, Fangneng Zhan, Adam Kortylewski, Christian Theobalt, and Peter Wonka. Zero-shot video semantic segmentation based on pre-trained diffusion models. arXiv:2405.16947, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[62] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[63] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6537-6549, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[64] Nora S Willett, Wilmot Li, Jovan Popovic, Floraine Berthouzoz, and Adam Finkelstein. Secondary motion for performed 2d animation. In Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology, pages 97-108, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[65] Nora S Willett, Rubaiat Habib Kazi, Michael Chen, George Fitzmaurice, Adam Finkelstein, and Tovi Grossman. A mixed-initiative interface for animating static pictures. In Proceedings of the 31st annual ACM symposium on user interface software and technology, pages 649-661, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[66] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. Aniclipart: Clipart animation with text-to-video priors. arXiv:2404.12347, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[67] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2025. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[68] Jun Xing, Li-Yi Wei, Takaaki Shiratori, and Koji Yatani. Autocomplete hand-drawn animations. ACM Transactions on Graphics (TOG), 34(6):1-11, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[69] Jun Xing, Rubaiat Habib Kazi, Tovi Grossman, Li-Yi Wei, Jos Stam, and George Fitzmaurice. Energy-brushes: Interactive tools for illustrating stylized elemental dynamics. In Proceedings of the 29th Annual Symposium on User Interface Software and Technology, pages 755-766, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[70] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. arXiv:2405.17933, 2024. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[71] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[72] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[73] Shiyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[74] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[75] Jiayi Eris Zhang, Seungbae Bang, David IW Levin, and Alec Jacobson. Complementary dynamics. arXiv:2009.02462, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[77] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv:2406.19680, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[78] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv:2407.21705, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[79] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. Trip: Temporal residual learning with image noise prior for image-to-video diffusion models. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10803" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 8671-8681, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[80] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv:2310.08465, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10804" + } + ] +] \ No newline at end of file diff --git a/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/d8df7b02-5bad-4fed-8afa-940086807a3d_origin.pdf b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/d8df7b02-5bad-4fed-8afa-940086807a3d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ba760c70a5744d4f54f551db42ca31af95dae9a8 --- /dev/null +++ b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/d8df7b02-5bad-4fed-8afa-940086807a3d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da5f866b3750e8938d7e634c37735ace4eb659d85b7185626a34d5d52dfbb443 +size 3311280 diff --git a/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/full.md b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0d8e15b3da1c294fb4eb11c6fa0fba666f93785f --- /dev/null +++ b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/full.md @@ -0,0 +1,500 @@ +# PhysAnimator: Physics-Guided Generative Cartoon Animation + +Tianyi Xie $^{1,2*}$ + +Yiwei Zhao $^{1\dagger}$ Ying Jiang $^{2}$ + +Chenfanfu Jiang2 + +1 Netflix, 2 UCLA + +tianyixie77@ucla.edu + +yiweiz@netflix.com + +{yingjiang, cffjiang}@ucla.edu + +![](images/a9e715952303dadac7b0abec1e9ca13a20e23492b52a5df698ddcfbd7ab94904.jpg) +Rig Force +Figure 1. PhysAnimator is a novel framework that combines physics principles with video diffusion models to generate high-quality animations from static animeillustrations, allowing users to specify external forces or rigging points for custom effects. + +![](images/a2571b5993f7dd2465a6b708ce88ca9acd9611039062dffc427db9514158bd22.jpg) + +![](images/92b152343db20f36e20ba59400c8c02cd1ebfb6738aa821026e6ed68f8e91633.jpg) + +![](images/e74272d0eddf6ccf39a046da0ced18d8fbf6c955bc416571fd55d49ae9dda6ec.jpg) + +![](images/76d2301538b8a695d2e2b0f899b878883d59fd49fab1073a7e79e4552106250d.jpg) + +![](images/82a4db7032800870a7677346ec9d4aea10033d2165d93dab2845422fb5c3bd8d.jpg) + +![](images/31132fdc53ef90a1da5c8dcaf39025bdc65dfa7c0956db9a5a5271433ddff276.jpg) + +![](images/f4c636e8f00a9d6014fe742ebb998c531e2f5b04d1245dc770dcd97695ba261a.jpg) + +![](images/085586728b176ffe03c4a51c8681f805e4665df120df0dfd6ca543fd1515f92f.jpg) + +![](images/ae902493323f1f161c82965912dedd9fb0cb7d917cfe98d2b638c123cfa3b5f4.jpg) + +![](images/dd188b53557b541e076a377e4b34475758ac451ab94bd4aa71da060b4e1d6b71.jpg) + +![](images/94044e4096ddabd0f1e47a8dc19322762f298ec456a3c57fefc0ef000c5d530b.jpg) + +![](images/2e932a9211f4d910a824f9921b9dc44b131292a9bda0c1d628a18f4c9c7ac222.jpg) + +![](images/fc3186f3e1485eeff8ca0fd49684b4b89673fd32756ba4ffa1b2ad46fdd8f5f9.jpg) + +![](images/41be1be77d15251e163cfb20d4e8e3681458e3997d1fb683776e2e7aef9bb4a3.jpg) + +![](images/1a7b28c573c7e8feebb4e3b0d8c8b3c74caae4ed33545b4d8b30c4e9178eb00f.jpg) + +![](images/228503e3ae85b052ba6e630cbac259a88d5fef2840a646235a7a48fe89fb9459.jpg) + +# Abstract + +Creating hand-drawn animation sequences is labor-intensive and demands professional expertise. We introduce PhysAnimator, a novel approach for generating physically plausible meanwhile anime-stylized animation from static animeillustrations.Our method seamlessly integrates physics-based simulations with data-driven generative models to produce dynamic and visually compelling animations.To capture the fluidity and exaggeration characteristic of anime,we perform image-space deformable body simulations on extracted mesh geometries.We enhance artistic control by introducing customizable energy strokes and incorporating rigging point support, enabling + +the creation of tailored animation effects such as wind interactions. Finally, we extract and warp sketches from the simulation sequence, generating a texture-agnostic representation, and employ a sketch-guided video diffusion model to synthesize high-quality animation frames. The resulting animations exhibit temporal consistency and visual plausibility, demonstrating the effectiveness of our method in creating dynamic anime-style animations. See our project page for more demos: https://xpandora.github.io/PhysAnimator/. + +# 1. Introduction + +Dynamic visual effects are essential to the immersive quality of 2D animation. From the subtle sway of a character's hair to the fluid motion of garments in response to + +wind, realistic and pleasing dynamics create a captivating visual experience. Traditionally, these effects are achieved through meticulous, hand-drawn techniques, where animators painstakingly draw each frame to bring these dynamic elements to life. This process is labor-intensive and requires not only artistic skill but also a deep understanding of natural forces and environmental interactions. + +To alleviate the challenges of manual animation, researchers have explored both traditional and data-driven approaches. Traditional animation tools [31, 58] provide interactive systems that assist artists in creating animated illustrations based on established principles of 2D animation. These methods often rely on user-provided stroke inputs to specify motion trajectories and utilize geometry constraints to produce deformation-based animation. However, such approaches typically assume simple inputs like lineart or drawings with separated layers, limiting their applicability to more complex, in-the-wild anime illustrations that feature intricate textures and details. In contrast, data-driven video generative models [2, 46, 71] offer a promising alternative by leveraging neural synthesis to generate dynamic effects directly from images, bypassing the need for manual sketching or layered inputs. Recent methods even enable interactive motion design, allowing users to specify object trajectories via drag-based inputs [49, 55, 67]. These approaches typically rely on predicting a sequence of optical flow fields to drive the desired motion and warp the frames accordingly. However, the quality of the generated results is often limited by inaccuracies in the predicted optical flow, which frequently exhibits artifacts due to a lack of geometric understanding and physical constraints. As a result, imprecise motion estimation can lead to noticeable distortions and unsatisfactory visual quality. + +Recognizing the limitations of both traditional and purely data-driven methods, we introduce PhysAnimator, a novel framework that integrates physics-based animation with data-driven generative models to synthesize visually compelling dynamic animations driven by environmental forces such as wind or rigging from static animeillustrations. Our approach combines the physical consistency of simulation-based methods with the flexibility and expressiveness of generative models, overcoming the drawbacks of previous approaches. To achieve this, we model objects of interest in the anime illustration as deformable bodies, capturing the fluidity and exaggerated motion characteristic of anime. We solve the motion equations to compute a sequence of optical flow fields that represent consistent dynamics. Users can enhance these animations with customized energy strokes, defining the effects of external forces such as wind. To render the motion dynamics as high-quality frames, we first extract and warp the sketch using the optical flow fields to generate a texture-agnostic video sequence. We then apply a sketch-guided video diffu + +sion model to colorize the sketch sequence based on the reference illustration, ensuring stylistic coherence. Finally, to incorporate dynamic effects that cannot be easily captured by physical simulation, we employ a data-driven cartoon interpolation model, enriching the results with complementary, expressive dynamics. In summary, our contributions are as follows: + +- We introduce a novel framework that combines physics-based simulations with data-driven generative models, specifically targeting animations driven by environmental forces and rigging controls, achieving both physical consistency and stylistic expressiveness. +- We develop an image-space deformable body simulation technique that models anime objects as deformable meshes, capturing fluid and exaggerated dynamics. +- We leverage a sketch-guided video diffusion model to render simulation dynamics as high-quality frames, combined with a cartoon interpolation model to introduce additional dynamic effects beyond physical laws. + +# 2. Related Work + +# 2.1. Video Diffusion Models + +Building on the success of image diffusion models [10, 53], recent work has introduced video diffusion models to streamline video synthesis from text prompts or images, significantly reducing labor and time costs compared to traditional commercial video editing and creation tools [23, 72]. These approaches either extend pre-trained image diffusion models by incorporating temporal mixing layers in various forms [3, 11, 13, 56] or train video diffusion models with temporal layers from scratch on large-scale text-video paired datasets [20, 79]. However, using text prompts or a single image as input provides limited control over fine-grained aspects of video generation, making it challenging to define complex structural attributes such as spatial layouts, poses, and shapes [17, 74]. To allow for more detailed control over object motion and camera movements, additional guidance in the form of motion trajectories [6, 28, 55, 74], pose [77], depth [21], and optical flow [17] has been integrated into video diffusion models to produce more controllable videos. These powerful video diffusion models have also been applied to various downstream tasks, such as video editing [16, 35], image animation [12, 66, 71], video understanding [38, 54, 61], video interpolation [27, 70] and 3D reconstruction and generation [8, 18, 36, 43, 60]. Nevertheless, these data-driven approaches usually produce artifacts due to a lack of geometric understanding and physical constraints. + +# 2.2. Physics-based Animation + +Researchers have explored various methods for animation authoring. Xing et al. [68] extend local similarity tech + +niques to global similarity, enabling automatic sketch completion based on previous frames. Building on this, Peng et al. [51] propose a keyframe-based sculpting system that autocompletes user edits using an intuitive brushing interface. Willett et al. [65] tackle the challenge of textured anime images by segmenting the illustration into layers for different objects and allowing users to provide scribbles as trajectory guides. Leveraging animation principles, some works [31, 45] use geometry-based deformations to create stylized animation effects such as "squash" and "stretch". Similarly, Xing et al. [69] employ physics-based simulations and introduce energy brushes to generate elemental dynamics [15], such as smoke and fire. Other approaches [1, 9, 29] incorporate physical laws like gravity, collision, and elasticity to animate deformable characters, enabling the efficient creation of complex scenes. For instance, Jones et al. [30] introduces an example-based plasticity model based on linear blend skinning for animating the failure of near-rigid, man-made materials. Coros et al. [9] exploit two rest-posed adaptation strategies using only internal energy to animate curve, shell, and solid-based characters. Physics-based modeling has also proven effective in adding secondary dynamics [64, 75] to enhance details in rigged animations. Despite their effectiveness, these 2D animation methods typically assume simple inputs, such as sketches, or require layered separation, which limits their applicability for animating complex, in-the-wild anime illustrations. + +# 2.3. Generative Image Dynamics + +With the advancement of generative models, there is growing interest in leveraging these powerful techniques to synthesize dynamic animations from static images, guided by motion features extracted from various user inputs. These inputs can be sparse, such as text prompts [7, 41, 71], trajectories [42, 55, 62, 67, 78], or camera movements [62, 73], or dense, like reference videos [28, 63, 80]. To achieve controllable dynamics, prior works often incorporate ControlNet [76] into image or video generative models during the decoding stage, utilizing motion features such as Canny edges, depth maps [17], 2D Gaussian maps [67], and optical flow maps [55]. In addition, several methods employ specialized motion fusion modules [41, 62, 78], spatial or temporal attention mechanisms [28, 63, 73], and feature injection strategies [42] to guide the generation of controllable dynamics. However, these approaches typically lack physics-based supervision, which can result in animations that violate physical laws or fail to align with user intentions [47]. Recently, PhysGen [44] integrated a rigid-body physics simulator to generate physics-consistent dynamics for foreground objects in a given image. While effective, PhysGen is limited to 2D rigid-body motions, making it unsuitable for the fluid, elastic effects commonly seen in anime, such as the waving of cloth or hair, which do not + +conform to rigid-body dynamics. In this work, we address these limitations by leveraging deformable-body dynamics and a sketch-guided video diffusion model to create high-quality, physically consistent animation sequences that capture the fluidity and elasticity characteristic of anime. + +# 3. Method + +Given a reference anime illustration $I_0$ , our goal is to generate a sequence of stylized video frames $\{\hat{I}_1, \hat{I}_2, \dots, \hat{I}_T\}$ with natural, user-guided motions. To achieve this, we introduce PhysAnimator, a novel framework that combines video diffusion models with physics-based animation. We start by segmenting the objects of interest in a given anime illustration and creating a 2D triangulated mesh representation. Using this geometry, we obtain dynamics through image-space deformable body simulation, generating a sequence of optical flow fields $\{\mathcal{F}_{0 \to 1}, \mathcal{F}_{0 \to 2}, \dots, \mathcal{F}_{0 \to T}\}$ . To offer user control, we enable interactive inputs through customizable energy strokes for defining external forces and rigging points for specifying desired trajectories. To render dynamics as high-quality frames, we extract the sketch from the input illustration and warp it using the computed optical flow fields, yielding a sequence of dynamic sketch frames $\{S_1, S_2, \dots, S_T\}$ . These sketches, together with the input illustration, are fed into a video diffusion model with a sketch-guided ControlNet [76], synthesizing a vivid animated sequence. Finally, an optional data-driven cartoon interpolation model [70] can be applied to enhance the anime style dynamics of the results by selecting keyframes from the animated sequence as input, yielding the final output frames $\{\hat{I}_1, \hat{I}_2, \dots, \hat{I}_T\}$ . An overview of our proposed framework is shown in Fig. 2. + +# 3.1. Preliminaries + +Latent Video Diffusion Model Most video diffusion models build on the latent diffusion model (LDM) framework [53], which uses Variational Autoencoder [32] (VAE) to map input images into a latent space. In this space, data is transformed into Gaussian noise via a forward diffusion process, which the model learns to reverse through denoising. In the forward diffusion process, the latent code $z_0$ is perturbed as: + +$$ +z _ {t} = \sqrt {\bar {\alpha} _ {t}} z _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \epsilon \sim \mathcal {N} (0, I), \tag {1} +$$ + +where $\bar{\alpha}_t = \prod_{i=1}^t (1 - \beta_t)$ with $\beta_t$ controlling the noise strength coefficient at step $t$ . The denoising model $\epsilon_\theta$ is trained to recover $z_t$ by minimizing the objective function: + +$$ +L _ {\epsilon} = \| \epsilon - \epsilon_ {\theta} (z _ {t}; c, t) \| _ {2} ^ {2}, \tag {2} +$$ + +where $\theta$ denotes the learnable network parameters and $c$ represents conditioning input (e.g. text prompts or images). + +![](images/9ac3ae24b651918e23a47af2a20c77497bd842cbd723639be066e332273292e3.jpg) +Figure 2. Method Overview. We begin by segmenting the object and creating a triangulated deformable mesh. Physics-based simulations are then used to generate dynamic optical flow fields, with users given the option to guide the motion through customizable energy strokes (shown as orange arrows) and rigging points (shown as red dots). The extracted sketch is warped using the computed optical flow and refined with a sketch-guided video diffusion model, producing a smooth, stylized animation sequence. Optionally, a cartoon interpolation model can further be applied to enhance the animation with expressive dynamics. + +After denoising, the VAE decoder reconstructs the latent code back into the image space. The latent video diffusion model (LVDM) extends the image LDM to videos by incorporating temporal modules to maintain temporal consistency. + +Deformable Body Dynamics Mathematically, the dynamics of a continuum deformable body is described by a time-dependent continuous deformation map $\boldsymbol{x} = \phi(\boldsymbol{X}, t)$ , which maps the undeformed material space $\Omega^0$ to the deformed world space $\Omega^t$ at time $t$ . The deformation gradient $\boldsymbol{F} = \frac{\partial \phi}{\partial \boldsymbol{X}}$ encodes the local deformation such as scaling, rotation, and shearing. In the context of a discretized setting, e.g. 2D meshes, the deformation map for each triangle can be expressed as + +$$ +\phi_ {i} (\boldsymbol {X}) = \boldsymbol {F} _ {i} \boldsymbol {X} + \boldsymbol {b} _ {i}, \tag {3} +$$ + +where $\pmb{b}_i\in \mathbb{R}^2$ accounts for translation of $i$ -th triangle, and the deformation gradient $\pmb {F}_i\in \mathbb{R}^{2\times 2}$ is assumed to be constant [40]. Undergoing deformation, the deformed body aims to recover its rest shape via resisting forces. Continuum mechanics model this behavior by first defining an energy density function $\Psi (F(x))$ , which measures the strain energy per unit undeformed volume. The total potential energy for a deformable body is then obtained as + +$$ +E (\boldsymbol {x}) = \sum_ {i = 1} ^ {N} \Psi \left(\boldsymbol {F} _ {i}\right) V _ {i}, \tag {4} +$$ + +where $V_{i}$ denotes volume of $i$ -th triangle. The internal resisting force is then defined as the negative gradient of the + +potential energy with respect to the vertex position + +$$ +\boldsymbol {f} _ {\text {i n t}} (\boldsymbol {x}) = - \frac {\partial E (\boldsymbol {x})}{\partial \boldsymbol {x}}. \tag {5} +$$ + +The deformable body dynamics is governed by Newton's Second Law as + +$$ +\frac {d ^ {2} \boldsymbol {x}}{d t ^ {2}} = \boldsymbol {M} ^ {- 1} \left(\boldsymbol {f} _ {\text {i n t}} (\boldsymbol {x}) + \boldsymbol {f} _ {\text {e x t}} (\boldsymbol {x})\right), \tag {6} +$$ + +which can be solved numerically. Here $M$ is the mass matrix that represents the masses of all vertices and $\pmb{f}_{\mathrm{ext}}$ refers to external forces, such as gravity. + +# 3.2. Physics-based Animation + +To generate dynamic animations from a single illustration, a straightforward approach is to use existing controllable video diffusion models [55, 67]. However, these purely data-driven methods often struggle due to a lack of physical understanding, leading to unrealistic results. To overcome this limitation, we incorporate physics-based animation to generate physically consistent and plausible motions. + +Geometry Registration The first step to animate objects of interest in a given illustration is to establish a geometric basis. While previous works [24, 59] have explored predicting 3D geometry from single-view images, these methods are typically constrained to reconstructing real-world objects and do not generalize to in-the-wild anime images, since characters and objects in anime often lack an inherent 3D representation due to their stylized and flat nature. As shown in prior works on cartoon + +animation [31, 69], leveraging 2D animation techniques proves effective for generating anime-style dynamic effects. Inspired by this, + +we focus on extracting 2D meshes for the objects of interest. To achieve this, we first utilize the Segment Anything Model (SAM) [33, 52] to + +![](images/140153bd5c609ced0978fbb8b624bc67b512235fdf3ee7991c891825297bd262.jpg) + +obtain segmentation masks for each target object, guided by user-specified query points. Along the contours of each segmentation mask, we uniformly sample boundary points, defining the outline of the intended mesh. We then employ conforming Delaunay triangulation [37] with these boundary constraints to generate well-structured triangular meshes, creating a simulation-ready format for subsequent animation. + +Deformable Body Model Anime scenes often feature dynamic environmental effects such as external forces that interact with characters' clothing, hair, or other elements. Additionally, techniques like "squash" and "stretch" are commonly used to convey motion and energy, adding expressiveness to animated objects. Motivated by these stylized dynamics, we model anime objects as deformable bodies, allowing them to capture the fluidity and exaggerated motion characteristic of anime. A key property of deformable bodies is their ability to change shape in response to external forces and attempt to return to their original rest shape upon deformation. To represent this physical behavior, we employ the Fixed Corotated Constitutive model [57], which defines the energy density as: + +$$ +\Psi (\boldsymbol {F}) = \mu \| \boldsymbol {F} - \boldsymbol {R} \| _ {F} ^ {2} + \frac {\lambda}{2} (\det (\boldsymbol {F}) - 1) ^ {2}, \tag {7} +$$ + +where $\mu, \lambda$ are the Lamé parameters, and $\pmb{R}$ is the rotational part of $\pmb{F}$ computed via polar decomposition. The first term models the stretching and compression resistance for the individual spatial directions and the second term describes resistance to volume change. The resulting internal forces $\pmb{f}_{\mathrm{int}}$ can then be derived via Eq. (5). + +Interactive Animation While $f_{\mathrm{int}}$ governs the inherent physical behavior, $f_{\mathrm{ext}}$ allows for user-defined interactions. Inspired by the concept of energy strokes [15, 69], we introduce customizable energy strokes that carry flow particles. These particles move along the user-specified strokes, propagating external forces to nearby vertices of the deformable mesh. This enables the creation of tailored animation effects, such as simulating wind interactions. Additionally, we incorporate rigging point support, allowing animators to + +anchor specific regions or guide them along predefined trajectories, offering enhanced control and flexibility. Using the specified energy strokes and deformable modeling, we solve the motion equation (Eq. (6)) to evolve the dynamics. We employ the semi-implicit Euler method to compute the deformation map $\phi_i(X,t)$ at each time step $t$ for every triangle. For each pixel $\mathbf{X}_p$ in the reference image, if it lies within a triangle $\mathcal{T}_i$ , we assign the displacement vector as $\pmb{d}(\pmb{p}) = \phi_i(\pmb{X}_p,t) - \pmb{X}_p$ ; otherwise, we assign a zero vector. Collecting displacement vector $\pmb{d}(\pmb{p})$ of all pixels yields the optical flow $\mathcal{F}_{0\rightarrow t}$ , which defines the displacement fields for the reference illustration $I_0$ at time $t$ . + +# 3.3. Generative Rendering + +In this section, we describe how to render the video sequence $\{\hat{I}_1,\hat{I}_2,\dots,\hat{I}_T\}$ given the reference image $I_0$ and the optical flow sequence $\{\mathcal{F}_{0\to 1},\mathcal{F}_{0\to 2},\ldots ,\mathcal{F}_{0\to T}\}$ derived from simulation. + +Sketch-Guided Rendering While the video sequence can be directly generated by warping the reference image $I_0$ using optical flow fields, the resulting frames often exhibit black hole artifacts caused by occlusions. To address this problem, we extract the sketch of $I_0$ , denoted as $S_0$ , and obtain the sketch at time $t$ as follows: + +$$ +S _ {t} = \mathcal {W} \left(S _ {0}, \mathcal {F} _ {0 \rightarrow t}, w _ {0 \rightarrow t}\right), \tag {8} +$$ + +where $\mathcal{W}$ represents the forward-warping operator, and $w_{0\rightarrow t}$ denotes the warping weights for each pixel in $I_0$ . In forward warping, multiple pixels may map to the same 2D location in the output frame [48], potentially leading to artifacts or distortions if the weights $w_{0\rightarrow t}$ are not properly defined. Inspired by [42], we set the pixel weight as $w_{0\rightarrow t}(\pmb {p}) = \| \mathcal{F}_{0\rightarrow t}(\pmb {p})\| _2$ , giving higher importance to pixels with larger motion, typically corresponding to foreground objects. + +Next, we leverage a video diffusion model to generate the rendered frames $\{I_1^r, I_2^r, \dots, I_T^r\}$ using the obtained sketch sequence $\{S_1, S_2, \dots, S_T\}$ . The diffusion model uses the reference image $I_0$ as an input, and we employ a ControlNet [76] with the sketches as control signals, guiding the generation process to ensure that the results align with the sketch inputs. We observed that during the inference time, due to segmentation inaccuracies, the optical flow warping may introduce unintended distortions, or miss parts of the object contour, generating imperfect sketches. To address this, we apply Gaussian blur to the input sketches at both training and inference time, which smoothes out inconsistencies. The video diffusion model is then capable of refining the results, leveraging its generative capabilities to handle imperfections and produce coherent outputs. + +Complementary Dynamics Unlike motions in the real world, dynamic effects in animation do not strictly adhere to + +physical laws and can not be fully captured by 2D animation methods. In the industrial animation pipeline, artists typically begin by creating a series of keyframes that define the primary motion trajectory, followed by drawing in-between frames to ensure smooth and fluid transitions. Inspired by this workflow, we leverage a data-driven module to enhance the physics-based animated results. Specifically, we select keyframes from the sketch-guided rendering results, forming the keyframe sequence $\{I_0^r, I_n^r, I_{2n}^r, \dots, I_{in}^r\}$ , where $n$ denotes the gap between keyframes. We then employ a cartoon interpolation video diffusion model [70], which synthesizes intermediate frames $\{\hat{I}_{in+1}, \hat{I}_{in+2}, \dots, \hat{I}_{i(n+1)-1}\}$ between each adjacent keyframe pair $(I_{in}^r, I_{i(n+1)}^r)$ . For keyframes, we simply assign $\hat{I}_{in} = I_{in}^r$ . This approach allows us to introduce expressive, data-driven complementary dynamics that go beyond what can be achieved through physics-based animation alone. + +# 4. Experiments + +In this section, we conduct a comprehensive comparison of our method against existing video diffusion models and demonstrate that our approach generates high-quality and physically plausible animations. + +Implementation Details We implement our deformable body simulator using Taichi [25]. During interactive animation, users can adjust the Lamé parameters $\mu$ and $\lambda$ to control the characteristics of the objects based on their specific needs. For generative rendering, we follow LVCD [26] to train a sketch-guided ControlNet for the stable video diffusion model [4], using blurred sketches as control signals to address potential imperfections in the predicted sketch sequence. When utilizing ToonCrafter [70] for generating in-between frames, similar to its original implementation, we also train an additional sketch-guided ControlNet, but set the control scale to 0.1 during inference. We also set $n = 15$ for the in-betweening step. This configuration helps the generated frames follow the rough motion indicated by the sketches while introducing extra stylized animation details. + +Dataset We build our training dataset using the Sakuga-42M Dataset [50], which consists of 1.4 million animation video clips. To ensure high-quality training samples, we filter out clips with fewer than 24 frames and dynamics scores outside the range of 0.05 to 0.7. Next, we extract sketches from the selected clips using the method from [5], resulting in a final dataset of 380,000 pairs of sketches and corresponding video sequences. + +Baseline We compare our method against two categories of video generation approaches. The first category includes state-of-the-art Image-to-Video (I2V) models, such as Cinemo [46] and DynamamiCrafter [71], which take an input im + +age and use text prompts to guide the motion dynamics of the generated videos. For these models, we generate prompts describing the image content and expected motion using ChatGPT-4V. The second category, including Motion-I2V [55] and Drag Anything [67], usually trains an additional motion-control module for video generative models such that, given a single image input, a user-specified trajectory can be provided to control the movement of objects in the generated frames. To ensure a fair comparison, we extract trajectories from our animated results and use them as input for these methods. + +# 4.1.Quantitative Evaluation + +Given the absence of established benchmarks for animestyle image-to-video generation, we follow Motion-I2V [55] and construct a test set comprising 20 anime images with stylized dynamic elements such as swaying hair, clothing, and plants. For each method, we generate 10 video samples per image, resulting in a total of 200 videos per method. To quantitatively evaluate the generated results, we use the Fréchet Inception Distance (FID) [22] to measure the similarity between the generated frames and the reference images. Additionally, we employ VideoScore [19] to assess multiple aspects of video quality, including visual quality, temporal consistency, dynamic degree, and factual consistency. The factual consistency evaluates the consistency of the video content with common sense and factual knowledge. We exclude the text alignment score as our method and the trajectory-controlled approach do not involve text input. + +The quantitative evaluation results, presented in Tab. 1, demonstrate that our method outperforms baseline approaches across most metrics, including visual quality, temporal consistency, and factual consistency, demonstrating the overall high quality of our generated videos. As shown in Fig. 3, Cinema tends to generate static videos that closely resemble the reference image, leading to better FID scores. While the dynamic degree score for DragAnything is notably higher than other methods, this is due to its tendency to misinterpret motion control as camera movement, causing shifts in the entire image space and resulting in an inflated dynamics score. Although our method's dynamic degree score is slightly lower than that of DynamiCrafter and Motion-I2V, these methods often exhibit implausible, distorted motions that create an illusion of increased dynamics. In contrast, our approach produces geometry-consistent motions and achieves high factual consistency, demonstrating the realistic motion dynamics enabled by our physics-based modeling. In addition, following [14, 34], we conducted an user study adopting a two-alternative forced choice (2AFC) protocol, where participants are asked to choose the preferred video based on temporal consistency, visual quality, motion plausibility, and overall feeling given + +![](images/72996d9e6b9751afae14e0bb6aeb4463b985e19af012218324573ac3d0587af2.jpg) + +"A pirate flag flutters and waves vigorously in the wind." + +![](images/5f6a3c80c2fc9498f6da9c46bc693d9bcb2ab8161a6d3403689f1b3c652b4d50.jpg) + +"The figure's hair flows gently with the breeze, strands swaying slightly." + +![](images/404d853408afd38d611ea5f6b88aa006c8355ef4e9410a93b187c9d0c1710d1e.jpg) + +![](images/bc796bb3fda26fafe75e51079a48d33857fbc2b314c053260d53087b9faf919c.jpg) + +![](images/e6fd15c76e952ce857490e897b6c7471314099af6fcf3a3220929bbfe951049f.jpg) + +![](images/82bb8f4e5333ca9eaf049c74d0c9edde9da2cd0d7c95121161c9d46f97344060.jpg) + +![](images/24e88cc06546baeffba54954e68a91f4fa1173bf60f5393a63c08ea4cac07f8c.jpg) + +![](images/04f3731173654c475ffb7fd9c30f2c1809341d9503951d44ef5af64a7bb35b0b.jpg) + +![](images/58c14f62647bcb91aa478a1bce85d8a15b11e36600ba8e5ae618c8e473bdfef1.jpg) + +![](images/ddd92798a4ac91f99ac2850046ba3dfacf59eabd18b51c2226974f853b224327.jpg) + +![](images/fc33df2cfd66b29ecc6ccc76b6cb04203cbd81a62b3cec85e3c3d616d69ef6db.jpg) + +![](images/30ac7c1cdb46509abfd92cc5cc9e7874aca21c1741866f285908461e8c9623da.jpg) + +![](images/b5ec722bf208bce94f84944ac5248afaf92946df1149a95b829ceccdbc55dfb9.jpg) + +![](images/05eb25d16d18b7c154fe6faf6ea1622b1ed77cfda47ea7e625b36ad869aeb1a4.jpg) + +![](images/4bdc00d6c2ba77f25faa1eb3aae90c99a34fd37db77316ae68b160f8b86d2b4d.jpg) + +![](images/15bc45417c47c863f62828292f6929272dbaccd8340dc48c75e2cbc2ac818e48.jpg) + +![](images/0d870129f6963e724fc525220cb14d7525dde5a948e309d169f3248abeddd0f0.jpg) + +![](images/9efaba4201d08f3f6d23866deb648175b56de9b41caff8efcbbdbcec1f6e45da.jpg) + +![](images/f8b6606a892fd79f847cb649b378777cc8f78fdcdc9ebc911a1d6a242f9f5d9d.jpg) + +![](images/3cd1cc252d0646a37f9d5f726694b71ed9f762d0fa82637ffc14fc67db9f1e04.jpg) + +![](images/34033b9d0993fcc5d3851566a841208ec90db1ac421575d44d6cb4cb9dcef814.jpg) + +![](images/fd5d0372f9e2e433499231924428215ff8d1079a59a1ef2d5ec3fab5049a4cdc.jpg) + +![](images/e2779d2521a555952da799825be49094fb68ed374e8c3493021ece3d5b1f0a5b.jpg) + +![](images/4c945aa6c93950d127da908ea999fe2043baec3ea022b90228127cd398f03a29.jpg) + +![](images/7c6471f274fc50fae62918089bf9301471bc8c5ce0ac9aaf3c3ee042f286c4de.jpg) + +![](images/68eb0c96e615cf2f829a969d738d1ce61702c94346109cd2eb24e6c4d14966c7.jpg) + +![](images/31e80205d0e1c6f9453dab615a0ebedd0d5f054a7d854eb940505391c2420fca.jpg) + +![](images/3a60bec4bdfcf020ae33ac6bc66cff6f159935fb56f7442c9efa6a6f92151b15.jpg) + +![](images/d71bcfe7097d17daf22ec939e35a84ef9cb4fa2026761227c83d560e575b562e.jpg) + +![](images/42f532c90fa66c9ae9d663fb821cdb2fdb2f07cce3a9f753c94df3f2c126fea0.jpg) + +![](images/d3ee0012edc3b554001c934479da90ba48898af2cdeff2a7f9f34ff991d99ee5.jpg) + +![](images/e5f70c9c586a0c7f39093e324545c8d734b3f7079d5ee38c6a0da7472e863766.jpg) + +![](images/a84de4bc1ad003696ce8d6ddde21e141cd24f1f6eee9ff95375c9a712039ecdb.jpg) + +![](images/09ac87a5b6d11e3481aed0e7a6772231af811eb2884e0169dd48c4dd05d1c15e.jpg) + +![](images/c2b2050c21d68f9ba72c5d18e4e6f9b7ef120224f12541b0ba74a3daa4348c63.jpg) + +![](images/6ac3123f6415d2ec9ea771ea80ea05fd29678f41826e762891adde5ce141275b.jpg) + +![](images/50b29a0b6d550058c26ecee0a4cc7807b7c53a9b112d3d49fd997ca74964e8c2.jpg) + +![](images/1b6074ffcc6c1665a9ca40dba2f71b094d9c2bd940ece7ecd026f14af502750d.jpg) + +"A scarecrow stands in a field, swaying slightly as a gentle breeze rustles through its clothes." + +![](images/b27a1322749fa13ca2322da0c18f5cfd0330c808ce6275281baef84a28412f23.jpg) + +"A street sign and a large, friendly creature stand nearby, both swaying gently in the rain." + +![](images/ab28d824abeac796127afd59b5c3139665748de17c0749402b75076f447be00e.jpg) + +![](images/d1aa985894f2d540fcce6ed3b7a95c963278504213fb3b0c016d93e007894804.jpg) + +![](images/70ad98a5867db3df8a218e30cd96e8ed86f64fa91c47c32ee5d0509f83ca775e.jpg) + +![](images/7372c880b318355611d45da8c3256d28e8e1ff81e3135db2469b53454fc05bc8.jpg) + +![](images/89fa153e5a2c5225cce4d8c73389f941b3c6f9d7bb1b89657366c13db8a90ff5.jpg) + +![](images/02aff4db14645901808384d9bbe39b1d672b1325e01fe7572a59a99c43140060.jpg) + +![](images/c5988bfca8895b7562a3d395bd776ddf39cdfcc007248eb704cad1d4caa9337c.jpg) + +![](images/3095b5322e89707f20e3e00a41f6047b16672805318b314beaa5cc941a605567.jpg) + +![](images/fd999b135e1d16eb44dd3bd848d842855dd4417a693cfae046669936289ab792.jpg) + +![](images/20087750fe1f32460c5b0270fc5a2bc590a1e7da4af0ba609eed95670f74aa10.jpg) + +![](images/7a992addcd2388e8719edf2da4f5a4df2965548abd1151f0a8c5b5af8761977d.jpg) + +![](images/c7b0b0c9ca6566e87b5bbbe32c83883bc36550cb84cf78432f90cbf7f05f1012.jpg) + +![](images/7b54ad140c594bb679f45ca47921d96c495f69a199c999b52823558c9001d46c.jpg) + +![](images/f267bc53ee02b69f23bed4570d02ad273a8dece970a555cd29d534aa03b8d106.jpg) + +![](images/809e411d6d5f68731831df46d1afbacfb6bb62fe4afca57273b98e89fa23cfaf.jpg) + +![](images/66adc13db0aa5eafc456e877e9e6edc83b31eb403237185a5412c0c60b45a3df.jpg) + +![](images/57c8c81b1540a3209285830cfdaba4a2fd2200f22e1af1d4a77e233f061f9a62.jpg) + +![](images/262454a7bc28d875471909c09aeac4dba219e971ba9d9ec0d151e0d8b3773b55.jpg) + +![](images/5843402af0ea4a17d5e638a9695913c6f0738b4d372961f9ed88d702d90ceec9.jpg) + +![](images/1e17be56dc2ac6d4b18210dd2be4a755cd5ba0e4ae0eea794ef5795a5eb15a2d.jpg) + +![](images/3bed9f7bcb369e182c45cd8cfe6670b85cdadb5928c626c33c7e1d7df704b251.jpg) + +![](images/09277e53bf282a151fee93d4bf8af13c928d12fd84999866f72359970a6f065b.jpg) + +![](images/9e753d89bdaeadaa816e33b8de8c04c75552a8cd593094ed40823f2aebc0f3ca.jpg) + +![](images/84f9629ad708f9caf9ff666de7122d7c61dbd6354f5034ee61add850e0534ecc.jpg) + +![](images/84d61bf8e63f112c823eb89a2c4bb63a72717cc774429f12d0d9785fe8b303b2.jpg) +Figure 3. Qualitative Comparison. We compare our results against Cinemo [46], Drag Anything [67], DynamiCrafter [71] and Motion-I2V [55]. Text prompts for Cinemo and DynamiCrafter are generated using ChatGPT-4V, while trajectories for Drag Anything and Motion-I2V are extracted from our animated results. + +![](images/a61c30fb2e68d46c4d2f28ed133b0010146ecdabf110f16002b5236d3f0cce4a.jpg) + +![](images/58d3eb698fd4c6e43b15ce569c685f321c80855db8d6be23517971fecbca2ef2.jpg) + +![](images/17035d8104ef2956c454728813ff96a921669a0cf761a70eae26494a1b4dceb1.jpg) + +![](images/917e34fdf95927c94b74507b33529d1411abed38bc33985ec7bf411ca4fa18a7.jpg) + +![](images/7662a5d73919ed2783883e5b2e9f79ab8c44d3e37a5bdfb0e7aa7353776443cd.jpg) + +two videos (one from our method and one from baselines). The user preference results, presented in Tab. 2, show that our proposed method consistently outperforms the baselines across all evaluation criteria. + +# 4.2. Qualitative Comparison + +We present a qualitative comparison with baseline methods in Fig. 3. Cinemo often produces static results with minimal motion dynamics. Drag Anything frequently misinterprets the motion trajectory as a camera movement, resulting in + +unintended dynamic sequences. DynamiCrafter, while generating larger motions, struggles to maintain the geometry of input objects, leading to noticeable distortions. Motion-I2V has difficulty preserving the appearance of the input content, often yielding unsatisfactory results. In contrast, our physics-guided approach ensures both physically plausible motions and high-quality rendering, preserving the geometry and visual consistency of the input. + +Table 1. Quantitative Comparisons. We report FID to evaluate the similarity between the reference image and generated frames. VSVQ, VSTC, VSDD, and VSFC represent scores for visual quality, temporal consistency, dynamic degree, and factual consistency, respectively, as measured by VideoScore [19]. + +
MethodsFID↓VSVQ↑VSTC↑VSDD↑VSFC↑
Cinemoto [46]49.52.852.802.422.58
DragAnything [67]148.92.772.452.972.52
DynamiCrafter [71]94.92.782.682.532.51
Motion-I2V [55]121.82.702.502.662.39
Ours90.42.892.862.482.64
+ +Table 2. User Study. We show the user preference for our method over the baseline methods in terms of visual quality (VQ), temporal consistency (TC), motion plausibility (MP), and overall feeling. + +
MethodsVQTCMPOverall
Cinemoto [46]86%83%82%81%
DragAnything [67]93%91%89%91%
DynamiCrafter [71]84%78%76%81%
Motion-I2V [55]95%94%97%96%
+ +![](images/cd091db1c3ae36cc9b4a5d8b67a436049ce8326a26cd689ee96ff8ce3d4d991f.jpg) +Reference + +![](images/b5ac7921e75f843573eed7141de0254603d0b0638db663b16a7ab4d5180fc883.jpg) +Warp + +![](images/5bf0de7aaae797c51a605607d18d9754bc5ca467cb9a2f604f3056b1d7789fa9.jpg) +Warp + Hard Inpaint + +![](images/09726d7a8718fa4ffa5f5190d08cdab74f9749af992288ba1daa8e2f0b7d1039.jpg) +Warp + Soft Inpaint +Figure 4. Sketch-Guided Rendering. Applying warping and inpainting introduces artifacts due to segmentation inaccuracy. Soft-inpainting [39] reduces these issues but can alter the content. Our sketch-guided rendering method produces high-quality results while preserving image details. + +![](images/85966a5bd34733c68c565dcf33d35234e86229de7e110abfa21ce73b4c792d6d.jpg) +Ours + +# 4.3. Additional Qualitative Results + +Sketch-Guided Rendering After physics-based animation in our proposed framework, we obtain a sequence of optical flow $\{\mathcal{F}_{0\to 1},\mathcal{F}_{0\to 2},\dots ,\mathcal{F}_{0\to T}\}$ . A straightforward approach would be to warp the input image $I_{0}$ using these optical flows and apply an inpainting algorithm to fill any occluded areas. However, this often results in suboptimal outputs. The main issue lies in segmentation inaccuracies, which can cause unintended regions to be warped or leave parts of the intended objects static, as shown in Fig. 4. While introducing a soft mask [39] into the inpainting process can help smooth the transition between the newly generated region and the original figure, it may also alter the content, making it diverge from the appearance of the reference image. Additionally, applying inpainting frame by + +![](images/649990fdea61fcdb48bf9a9aedfc6332d7da33fe520d8e1f192348d31d392642.jpg) +Static Reference + +![](images/84c44ba90a0a06041d57536943a5d89267f44585d43ea7b75caf193469a2aa55.jpg) +w/o Enhanced Dynamics +Figure 5. Complementary Dynamics Enhancement. While physics-based animation maintains geometric consistency, it may lack the fluidity and exaggeration commonly seen in anime. We employ a data-driven interpolation module to enhance the motion dynamics, creating more natural-looking animations that better resemble real anime. + +![](images/1ddccb858cb8dbf3c634553f15a094887f19fa1d4c14d986310ede26f6feb604.jpg) +w/ Enhanced Dynamics + +frame may also cause temporal consistency issues. To address this, our method employs a sketch-guided rendering module. We first extract a sparse geometric representation of the image and apply the animation dynamics directly to this sketch-based structure. This sparse representation is more robust to segmentation inaccuracies and helps preserve the intended motion. Our subsequent rendering module then synthesizes high-quality video sequences from the animated sketches, maintaining both temporal consistency and visual fidelity. + +Complementary Dynamics Enhancement While our physics-based animation ensures physically accurate and geometry-consistent motion, it is inherently limited to 2D representations and cannot fully capture 3D effects. Moreover, the exaggerated dynamics often seen in anime do not always conform to strict physical laws. To address these limitations, we incorporate a data-driven cartoon interpolation module [70] to introduce complementary dynamics. As illustrated in Fig. 5, this interpolation module enables fluid contour deformations of clothing during motion, rather than rigidly adhering to static geometry. It also dynamically generates new sketch lines, enhancing the expressiveness of the animation and bringing it closer to the aesthetic of traditional hand-drawn anime. + +# 5. Conclusion + +We presented PhysAnimator, a novel framework for generating dynamic and stylized animations from static anime illustrations by integrating physics-based simulations with data-driven generative models. Our approach generates physically plausible, fluid and exaggerated animestyled motion through deformable body simulations, providing controllability to users through user-guided energy strokes. A sketch-guided video diffusion model ensures high-quality, temporally consistent frames, while a data-driven anime frame interpolation model adds expressive, non-physical dynamics. The experiments show that our proposed method outperforms existing video diffusion methods, offering a powerful tool for creating visually compelling and user-controllable anime-style animations. + +# Acknowledgements + +# References + +[1] Jernej Barbic, Marco da Silva, and Jovan Popovic. Deformable object animation using reduced optimal control. In ACM SIGGRAPH 2009 papers, pages 1-9. 2009. 3 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2 +[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 6 +[5] Caroline Chan, Fredo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. 2022. 6 +[6] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv:2305.13840, 2023. 2 +[7] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 3 +[8] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv:2403.06738, 2024. 2 +[9] Stelian Coros, Sebastian Martin, Bernhard Thomaszewski, Christian Schumacher, Robert Sumner, and Markus Gross. Deformable objects alive! ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 3 +[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[11] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 2 + +We thank anonymous reviewers for their insightful comments. We acknowledge support from NSF 2153851. We would like to thank Rahul Garg, Hossein Taghavi, Roshni Cooper, Ritwik Kumar, Boris Chen, Amli Murphy, Oliver Banasiak and Taiki Sakurai for their valuable suggestions and support. +[12] Rinon Gal, Yael Vinker, Yuval Alaluf, Amit Bermano, Daniel Cohen-Or, Ariel Shamir, and Gal Chechik. Breathing life into sketches using text-to-video priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4325–4336, 2024. 2 +[13] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 2 +[14] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arxiv:2307.10373, 2023. 6 +[15] Joseph Gilland. Elemental Magic, Volume 2: The Technique of Special Effects Animation. Routledge, 2012. 3, 5 +[16] Yuchao Gu, Yipin Zhou, Bichen Wu, Licheng Yu, Jia-Wei Liu, Rui Zhao, Jay Zhangjie Wu, David Junhao Zhang, Mike Zheng Shou, and Kevin Tang. Videoswap: Customized video subject swapping with interactive semantic point correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7621-7630, 2024. 2 +[17] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 2, 3 +[18] Junlin Han, Filippos Kokkinos, and Philip Torr. Vfusion3d: Learning scalable 3d generative models from video diffusion models. In European Conference on Computer Vision, pages 333-350. Springer, 2025. 2 +[19] Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, et al. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. arXiv:2406.15252, 2024. 6, 8 +[20] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv:2211.13221, 2022. 2 +[21] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. animate-a-story: Storytelling with retrieval-augmented video generation. arXiv:2307.06940, 2023. 2 +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[23] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2 +[24] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv:2311.04400, 2023. 4 + +[25] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG), 38(6): 201, 2019. 6 +[26] Zhitong Huang, Mohan Zhang, and Jing Liao. Lycd: Reference-based linear video colorization with diffusion models. arXiv:2409.12960, 2024. 6 +[27] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7341-7351, 2024. 2 +[28] Hyeonho Jeong, Geon Yeong Park, and Jong Chul Ye. Vmc: Video motion customization using temporal attention adaptation for text-to-video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9212-9221, 2024. 2, 3 +[29] Ben Jones, Jovan Popovic, James McCann, Wilmot Li, and Adam Bargteil. Dynamic sprites: artistic authoring of interactive animations. Computer Animation and Virtual Worlds, 26(2):97-108, 2015. 3 +[30] Ben Jones, Nils Thuerey, Tamar Shinar, and Adam W Bargteil. Example-based plastic deformation of rigid bodies. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. 3 +[31] Rubaiat Habib Kazi, Tovi Grossman, Nobuyuki Umetani, and George Fitzmaurice. Motion amplifiers: sketching dynamic illustrations using the principles of 2d animation. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, pages 4599-4609, 2016. 2, 3, 5 +[32] Diederik P Kingma. Auto-encoding variational bayes. arXiv:1312.6114, 2013. 3 +[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5 +[34] Nicholas Kolkin, Jason Salavon, and Gregory Shakhnarovich. Style transfer by relaxed optimal transport and self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10051-10060, 2019. 6 +[35] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv:2403.14468, 2024. 2 +[36] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 2 +[37] Der-Tsai Lee and Bruce J Schachter. Two algorithms for constructing a delaunay triangulation. International Journal of Computer & Information Sciences, 9(3):219-242, 1980. 5 +[38] Minhyeok Lee, Suhwan Cho, Chajin Shin, Jungho Lee, Sunghun Yang, and Sangyoun Lee. Video diffusion models are strong video inpainter. arXiv:2408.11402, 2024. 2 +[39] Eran Levin and Ohad Fried. Differential diffusion: Giving each pixel its strength. arXiv:2306.00950, 2023. 8 + +[40] Minchen Li, Chenfanfu Jiang, and Zhaofeng Luo. Physics-Based Simulation. 2024. 4 +[41] Mingxiao Li, Bo Wan, Marie-Francine Moens, and Tinne Tuytelaars. animate your motion: Turning still images into dynamic videos. arXiv:2403.10179, 2024. 3 +[42] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 3, 5 +[43] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv:2408.16767, 2024. 2 +[44] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imaged-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 3 +[45] Jiaju Ma, Li-Yi Wei, and Rubaiat Habib Kazi. A layered authoring tool for stylized 3d animations. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1-14, 2022. 3 +[46] Xin Ma, Yaohui Wang, Gengyu Jia, Xinyuan Chen, YuanFang Li, Cunjian Chen, and Yu Qiao. Cinemo: Consistent and controllable image animation with motion diffusion models. arXiv:2407.15642, 2024. 2, 6, 7, 8 +[47] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv:2410.05363, 2024. 3 +[48] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5437-5446, 2020. 5 +[49] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adaptions in frozen image-to-video diffusion model. arXiv:2405.20222, 2024. 2 +[50] Zhenglin Pan, Yu Zhu, and Yuxuan Mu. Sakuga-42m dataset: Scaling up cartoon research. arXiv:2405.07425, 2024.6 +[51] Mengqi Peng, Li-yi Wei, Rubaiat Habib Kazi, and Vladimir G Kim. Autocomplete animated sculpting. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 760-777, 2020. 3 +[52] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 5 +[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3 + +[54] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv:2406.01493, 2024. 2 +[55] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 4, 6, 7, 8 +[56] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv:2209.14792, 2022. 2 +[57] Alexey Stomakhin, Russell Howes, Craig A Schroeder, and Joseph M Teran. Energetically consistent invertible elasticity. In Symposium on Computer Animation, 2012. 5 +[58] Qingkun Su, Xue Bai, Hongbo Fu, Chiew-Lan Tai, and Jue Wang. Live sketch: Video-driven dynamic deformation of static drawings. In Proceedings of the 2018 chi conference on human factors in computing systems, pages 1-12, 2018. 2 +[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 4 +[60] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2 +[61] Qian Wang, Abdelrahman Eldosokey, Mohit Mendiratta, Fangneng Zhan, Adam Kortylewski, Christian Theobalt, and Peter Wonka. Zero-shot video semantic segmentation based on pre-trained diffusion models. arXiv:2405.16947, 2024. 2 +[62] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[63] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6537-6549, 2024. 3 +[64] Nora S Willett, Wilmot Li, Jovan Popovic, Floraine Berthouzoz, and Adam Finkelstein. Secondary motion for performed 2d animation. In Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology, pages 97-108, 2017. 3 +[65] Nora S Willett, Rubaiat Habib Kazi, Michael Chen, George Fitzmaurice, Adam Finkelstein, and Tovi Grossman. A mixed-initiative interface for animating static pictures. In Proceedings of the 31st annual ACM symposium on user interface software and technology, pages 649-661, 2018. 3 + +[66] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. Aniclipart: Clipart animation with text-to-video priors. arXiv:2404.12347, 2024. 2 +[67] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2025. 2, 3, 4, 6, 7, 8 +[68] Jun Xing, Li-Yi Wei, Takaaki Shiratori, and Koji Yatani. Autocomplete hand-drawn animations. ACM Transactions on Graphics (TOG), 34(6):1-11, 2015. 2 +[69] Jun Xing, Rubaiat Habib Kazi, Tovi Grossman, Li-Yi Wei, Jos Stam, and George Fitzmaurice. Energy-brushes: Interactive tools for illustrating stylized elemental dynamics. In Proceedings of the 29th Annual Symposium on User Interface Software and Technology, pages 755-766, 2016. 3, 5 +[70] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. arXiv:2405.17933, 2024. 2, 3, 6, 8 +[71] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 3, 6, 7, 8 +[72] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 2023. 2 +[73] Shiyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3 +[74] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089, 2023. 2 +[75] Jiayi Eris Zhang, Seungbae Bang, David IW Levin, and Alec Jacobson. Complementary dynamics. arXiv:2009.02462, 2020.3 +[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3, 5 +[77] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv:2406.19680, 2024. 2 +[78] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv:2407.21705, 2024. 3 +[79] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. Trip: Temporal residual learning with image noise prior for image-to-video diffusion models. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 8671-8681, 2024. 2 +[80] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv:2310.08465, 2023. 3 \ No newline at end of file diff --git a/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/images.zip b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..23c48598aae8552202a3afa1bb4dc71b95c67dc4 --- /dev/null +++ b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11ef9348680aee8b919edf5e15e6b26cb69b959d1cc2257ef9d2f0609ede4461 +size 699041 diff --git a/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/layout.json b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..12426b80c14302b71bcde29037e7e1339aecad50 --- /dev/null +++ b/2025/PhysAnimator_ Physics-Guided Generative Cartoon Animation/layout.json @@ -0,0 +1,12861 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 501, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 501, + 120 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 501, + 120 + ], + "type": "text", + "content": "PhysAnimator: Physics-Guided Generative Cartoon Animation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 142, + 216, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 142, + 216, + 157 + ], + "spans": [ + { + "bbox": [ + 146, + 142, + 216, + 157 + ], + "type": "text", + "content": "Tianyi Xie" + }, + { + "bbox": [ + 146, + 142, + 216, + 157 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 229, + 142, + 364, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 142, + 364, + 157 + ], + "spans": [ + { + "bbox": [ + 229, + 142, + 364, + 157 + ], + "type": "text", + "content": "Yiwei Zhao" + }, + { + "bbox": [ + 229, + 142, + 364, + 157 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 229, + 142, + 364, + 157 + ], + "type": "text", + "content": " Ying Jiang" + }, + { + "bbox": [ + 229, + 142, + 364, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 377, + 143, + 463, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 143, + 463, + 157 + ], + "spans": [ + { + "bbox": [ + 377, + 143, + 463, + 157 + ], + "type": "text", + "content": "Chenfanfu Jiang2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 260, + 157, + 350, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 157, + 350, + 171 + ], + "spans": [ + { + "bbox": [ + 260, + 157, + 350, + 171 + ], + "type": "text", + "content": "1 Netflix, 2 UCLA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 174, + 216, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 174, + 216, + 185 + ], + "spans": [ + { + "bbox": [ + 104, + 174, + 216, + 185 + ], + "type": "text", + "content": "tianyixie77@ucla.edu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 227, + 174, + 327, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 174, + 327, + 185 + ], + "spans": [ + { + "bbox": [ + 227, + 174, + 327, + 185 + ], + "type": "text", + "content": "yiweiz@netflix.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 340, + 174, + 502, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 174, + 502, + 185 + ], + "spans": [ + { + "bbox": [ + 340, + 174, + 502, + 185 + ], + "type": "text", + "content": "{yingjiang, cffjiang}@ucla.edu" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 58, + 215, + 156, + 460 + ], + "blocks": [ + { + "bbox": [ + 58, + 215, + 156, + 460 + ], + "lines": [ + { + "bbox": [ + 58, + 215, + 156, + 460 + ], + "spans": [ + { + "bbox": [ + 58, + 215, + 156, + 460 + ], + "type": "image", + "image_path": "a9e715952303dadac7b0abec1e9ca13a20e23492b52a5df698ddcfbd7ab94904.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 460, + 154, + 472 + ], + "lines": [ + { + "bbox": [ + 65, + 460, + 154, + 472 + ], + "spans": [ + { + "bbox": [ + 65, + 460, + 154, + 472 + ], + "type": "text", + "content": "Rig Force" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 482, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 55, + 482, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 555, + 506 + ], + "type": "text", + "content": "Figure 1. PhysAnimator is a novel framework that combines physics principles with video diffusion models to generate high-quality animations from static animeillustrations, allowing users to specify external forces or rigging points for custom effects." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 162, + 215, + 260, + 277 + ], + "blocks": [ + { + "bbox": [ + 162, + 215, + 260, + 277 + ], + "lines": [ + { + "bbox": [ + 162, + 215, + 260, + 277 + ], + "spans": [ + { + "bbox": [ + 162, + 215, + 260, + 277 + ], + "type": "image", + "image_path": "a2571b5993f7dd2465a6b708ce88ca9acd9611039062dffc427db9514158bd22.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 164, + 277, + 260, + 337 + ], + "blocks": [ + { + "bbox": [ + 164, + 277, + 260, + 337 + ], + "lines": [ + { + "bbox": [ + 164, + 277, + 260, + 337 + ], + "spans": [ + { + "bbox": [ + 164, + 277, + 260, + 337 + ], + "type": "image", + "image_path": "92b152343db20f36e20ba59400c8c02cd1ebfb6738aa821026e6ed68f8e91633.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 163, + 337, + 260, + 397 + ], + "blocks": [ + { + "bbox": [ + 163, + 337, + 260, + 397 + ], + "lines": [ + { + "bbox": [ + 163, + 337, + 260, + 397 + ], + "spans": [ + { + "bbox": [ + 163, + 337, + 260, + 397 + ], + "type": "image", + "image_path": "e74272d0eddf6ccf39a046da0ced18d8fbf6c955bc416571fd55d49ae9dda6ec.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 163, + 398, + 260, + 460 + ], + "blocks": [ + { + "bbox": [ + 163, + 398, + 260, + 460 + ], + "lines": [ + { + "bbox": [ + 163, + 398, + 260, + 460 + ], + "spans": [ + { + "bbox": [ + 163, + 398, + 260, + 460 + ], + "type": "image", + "image_path": "76d2301538b8a695d2e2b0f899b878883d59fd49fab1073a7e79e4552106250d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 260, + 215, + 356, + 277 + ], + "blocks": [ + { + "bbox": [ + 260, + 215, + 356, + 277 + ], + "lines": [ + { + "bbox": [ + 260, + 215, + 356, + 277 + ], + "spans": [ + { + "bbox": [ + 260, + 215, + 356, + 277 + ], + "type": "image", + "image_path": "82a4db7032800870a7677346ec9d4aea10033d2165d93dab2845422fb5c3bd8d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 260, + 277, + 356, + 337 + ], + "blocks": [ + { + "bbox": [ + 260, + 277, + 356, + 337 + ], + "lines": [ + { + "bbox": [ + 260, + 277, + 356, + 337 + ], + "spans": [ + { + "bbox": [ + 260, + 277, + 356, + 337 + ], + "type": "image", + "image_path": "31132fdc53ef90a1da5c8dcaf39025bdc65dfa7c0956db9a5a5271433ddff276.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 260, + 337, + 356, + 398 + ], + "blocks": [ + { + "bbox": [ + 260, + 337, + 356, + 398 + ], + "lines": [ + { + "bbox": [ + 260, + 337, + 356, + 398 + ], + "spans": [ + { + "bbox": [ + 260, + 337, + 356, + 398 + ], + "type": "image", + "image_path": "f4c636e8f00a9d6014fe742ebb998c531e2f5b04d1245dc770dcd97695ba261a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 260, + 398, + 356, + 460 + ], + "blocks": [ + { + "bbox": [ + 260, + 398, + 356, + 460 + ], + "lines": [ + { + "bbox": [ + 260, + 398, + 356, + 460 + ], + "spans": [ + { + "bbox": [ + 260, + 398, + 356, + 460 + ], + "type": "image", + "image_path": "085586728b176ffe03c4a51c8681f805e4665df120df0dfd6ca543fd1515f92f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 356, + 215, + 452, + 277 + ], + "blocks": [ + { + "bbox": [ + 356, + 215, + 452, + 277 + ], + "lines": [ + { + "bbox": [ + 356, + 215, + 452, + 277 + ], + "spans": [ + { + "bbox": [ + 356, + 215, + 452, + 277 + ], + "type": "image", + "image_path": "ae902493323f1f161c82965912dedd9fb0cb7d917cfe98d2b638c123cfa3b5f4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 356, + 277, + 452, + 337 + ], + "blocks": [ + { + "bbox": [ + 356, + 277, + 452, + 337 + ], + "lines": [ + { + "bbox": [ + 356, + 277, + 452, + 337 + ], + "spans": [ + { + "bbox": [ + 356, + 277, + 452, + 337 + ], + "type": "image", + "image_path": "dd188b53557b541e076a377e4b34475758ac451ab94bd4aa71da060b4e1d6b71.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 356, + 337, + 452, + 398 + ], + "blocks": [ + { + "bbox": [ + 356, + 337, + 452, + 398 + ], + "lines": [ + { + "bbox": [ + 356, + 337, + 452, + 398 + ], + "spans": [ + { + "bbox": [ + 356, + 337, + 452, + 398 + ], + "type": "image", + "image_path": "94044e4096ddabd0f1e47a8dc19322762f298ec456a3c57fefc0ef000c5d530b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 356, + 398, + 452, + 460 + ], + "blocks": [ + { + "bbox": [ + 356, + 398, + 452, + 460 + ], + "lines": [ + { + "bbox": [ + 356, + 398, + 452, + 460 + ], + "spans": [ + { + "bbox": [ + 356, + 398, + 452, + 460 + ], + "type": "image", + "image_path": "2e932a9211f4d910a824f9921b9dc44b131292a9bda0c1d628a18f4c9c7ac222.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 452, + 215, + 549, + 277 + ], + "blocks": [ + { + "bbox": [ + 452, + 215, + 549, + 277 + ], + "lines": [ + { + "bbox": [ + 452, + 215, + 549, + 277 + ], + "spans": [ + { + "bbox": [ + 452, + 215, + 549, + 277 + ], + "type": "image", + "image_path": "fc3186f3e1485eeff8ca0fd49684b4b89673fd32756ba4ffa1b2ad46fdd8f5f9.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 452, + 277, + 549, + 337 + ], + "blocks": [ + { + "bbox": [ + 452, + 277, + 549, + 337 + ], + "lines": [ + { + "bbox": [ + 452, + 277, + 549, + 337 + ], + "spans": [ + { + "bbox": [ + 452, + 277, + 549, + 337 + ], + "type": "image", + "image_path": "41be1be77d15251e163cfb20d4e8e3681458e3997d1fb683776e2e7aef9bb4a3.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 452, + 337, + 549, + 398 + ], + "blocks": [ + { + "bbox": [ + 452, + 337, + 549, + 398 + ], + "lines": [ + { + "bbox": [ + 452, + 337, + 549, + 398 + ], + "spans": [ + { + "bbox": [ + 452, + 337, + 549, + 398 + ], + "type": "image", + "image_path": "1a7b28c573c7e8feebb4e3b0d8c8b3c74caae4ed33545b4d8b30c4e9178eb00f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 452, + 398, + 549, + 460 + ], + "blocks": [ + { + "bbox": [ + 452, + 398, + 549, + 460 + ], + "lines": [ + { + "bbox": [ + 452, + 398, + 549, + 460 + ], + "spans": [ + { + "bbox": [ + 452, + 398, + 549, + 460 + ], + "type": "image", + "image_path": "228503e3ae85b052ba6e630cbac259a88d5fef2840a646235a7a48fe89fb9459.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 152, + 514, + 200, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 514, + 200, + 528 + ], + "spans": [ + { + "bbox": [ + 152, + 514, + 200, + 528 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 540, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 684 + ], + "type": "text", + "content": "Creating hand-drawn animation sequences is labor-intensive and demands professional expertise. We introduce PhysAnimator, a novel approach for generating physically plausible meanwhile anime-stylized animation from static animeillustrations.Our method seamlessly integrates physics-based simulations with data-driven generative models to produce dynamic and visually compelling animations.To capture the fluidity and exaggeration characteristic of anime,we perform image-space deformable body simulations on extracted mesh geometries.We enhance artistic control by introducing customizable energy strokes and incorporating rigging point support, enabling" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 635 + ], + "type": "text", + "content": "the creation of tailored animation effects such as wind interactions. Finally, we extract and warp sketches from the simulation sequence, generating a texture-agnostic representation, and employ a sketch-guided video diffusion model to synthesize high-quality animation frames. The resulting animations exhibit temporal consistency and visual plausibility, demonstrating the effectiveness of our method in creating dynamic anime-style animations. See our project page for more demos: https://xpandora.github.io/PhysAnimator/." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "Dynamic visual effects are essential to the immersive quality of 2D animation. From the subtle sway of a character's hair to the fluid motion of garments in response to" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 216, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 216, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 216, + 703 + ], + "type": "text", + "content": "* Work done during the internship at Netflix." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 71, + 704, + 150, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 704, + 150, + 712 + ], + "spans": [ + { + "bbox": [ + 71, + 704, + 150, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10793" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "wind, realistic and pleasing dynamics create a captivating visual experience. Traditionally, these effects are achieved through meticulous, hand-drawn techniques, where animators painstakingly draw each frame to bring these dynamic elements to life. This process is labor-intensive and requires not only artistic skill but also a deep understanding of natural forces and environmental interactions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 298, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 298, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 298, + 471 + ], + "type": "text", + "content": "To alleviate the challenges of manual animation, researchers have explored both traditional and data-driven approaches. Traditional animation tools [31, 58] provide interactive systems that assist artists in creating animated illustrations based on established principles of 2D animation. These methods often rely on user-provided stroke inputs to specify motion trajectories and utilize geometry constraints to produce deformation-based animation. However, such approaches typically assume simple inputs like lineart or drawings with separated layers, limiting their applicability to more complex, in-the-wild anime illustrations that feature intricate textures and details. In contrast, data-driven video generative models [2, 46, 71] offer a promising alternative by leveraging neural synthesis to generate dynamic effects directly from images, bypassing the need for manual sketching or layered inputs. Recent methods even enable interactive motion design, allowing users to specify object trajectories via drag-based inputs [49, 55, 67]. These approaches typically rely on predicting a sequence of optical flow fields to drive the desired motion and warp the frames accordingly. However, the quality of the generated results is often limited by inaccuracies in the predicted optical flow, which frequently exhibits artifacts due to a lack of geometric understanding and physical constraints. As a result, imprecise motion estimation can lead to noticeable distortions and unsatisfactory visual quality." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 474, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 298, + 714 + ], + "type": "text", + "content": "Recognizing the limitations of both traditional and purely data-driven methods, we introduce PhysAnimator, a novel framework that integrates physics-based animation with data-driven generative models to synthesize visually compelling dynamic animations driven by environmental forces such as wind or rigging from static animeillustrations. Our approach combines the physical consistency of simulation-based methods with the flexibility and expressiveness of generative models, overcoming the drawbacks of previous approaches. To achieve this, we model objects of interest in the anime illustration as deformable bodies, capturing the fluidity and exaggerated motion characteristic of anime. We solve the motion equations to compute a sequence of optical flow fields that represent consistent dynamics. Users can enhance these animations with customized energy strokes, defining the effects of external forces such as wind. To render the motion dynamics as high-quality frames, we first extract and warp the sketch using the optical flow fields to generate a texture-agnostic video sequence. We then apply a sketch-guided video diffu" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "content": "sion model to colorize the sketch sequence based on the reference illustration, ensuring stylistic coherence. Finally, to incorporate dynamic effects that cannot be easily captured by physical simulation, we employ a data-driven cartoon interpolation model, enriching the results with complementary, expressive dynamics. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 156, + 553, + 299 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 156, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 553, + 215 + ], + "type": "text", + "content": "- We introduce a novel framework that combines physics-based simulations with data-driven generative models, specifically targeting animations driven by environmental forces and rigging controls, achieving both physical consistency and stylistic expressiveness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 217, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 251 + ], + "type": "text", + "content": "- We develop an image-space deformable body simulation technique that models anime objects as deformable meshes, capturing fluid and exaggerated dynamics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 251, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 299 + ], + "type": "text", + "content": "- We leverage a sketch-guided video diffusion model to render simulation dynamics as high-quality frames, combined with a cartoon interpolation model to introduce additional dynamic effects beyond physical laws." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 312, + 401, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 401, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 401, + 324 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 332, + 447, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 447, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 447, + 344 + ], + "type": "text", + "content": "2.1. Video Diffusion Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 351, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 662 + ], + "type": "text", + "content": "Building on the success of image diffusion models [10, 53], recent work has introduced video diffusion models to streamline video synthesis from text prompts or images, significantly reducing labor and time costs compared to traditional commercial video editing and creation tools [23, 72]. These approaches either extend pre-trained image diffusion models by incorporating temporal mixing layers in various forms [3, 11, 13, 56] or train video diffusion models with temporal layers from scratch on large-scale text-video paired datasets [20, 79]. However, using text prompts or a single image as input provides limited control over fine-grained aspects of video generation, making it challenging to define complex structural attributes such as spatial layouts, poses, and shapes [17, 74]. To allow for more detailed control over object motion and camera movements, additional guidance in the form of motion trajectories [6, 28, 55, 74], pose [77], depth [21], and optical flow [17] has been integrated into video diffusion models to produce more controllable videos. These powerful video diffusion models have also been applied to various downstream tasks, such as video editing [16, 35], image animation [12, 66, 71], video understanding [38, 54, 61], video interpolation [27, 70] and 3D reconstruction and generation [8, 18, 36, 43, 60]. Nevertheless, these data-driven approaches usually produce artifacts due to a lack of geometric understanding and physical constraints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 670, + 455, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 455, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 455, + 682 + ], + "type": "text", + "content": "2.2. Physics-based Animation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Researchers have explored various methods for animation authoring. Xing et al. [68] extend local similarity tech" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10794" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "content": "niques to global similarity, enabling automatic sketch completion based on previous frames. Building on this, Peng et al. [51] propose a keyframe-based sculpting system that autocompletes user edits using an intuitive brushing interface. Willett et al. [65] tackle the challenge of textured anime images by segmenting the illustration into layers for different objects and allowing users to provide scribbles as trajectory guides. Leveraging animation principles, some works [31, 45] use geometry-based deformations to create stylized animation effects such as \"squash\" and \"stretch\". Similarly, Xing et al. [69] employ physics-based simulations and introduce energy brushes to generate elemental dynamics [15], such as smoke and fire. Other approaches [1, 9, 29] incorporate physical laws like gravity, collision, and elasticity to animate deformable characters, enabling the efficient creation of complex scenes. For instance, Jones et al. [30] introduces an example-based plasticity model based on linear blend skinning for animating the failure of near-rigid, man-made materials. Coros et al. [9] exploit two rest-posed adaptation strategies using only internal energy to animate curve, shell, and solid-based characters. Physics-based modeling has also proven effective in adding secondary dynamics [64, 75] to enhance details in rigged animations. Despite their effectiveness, these 2D animation methods typically assume simple inputs, such as sketches, or require layered separation, which limits their applicability for animating complex, in-the-wild anime illustrations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 407, + 211, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 211, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 211, + 419 + ], + "type": "text", + "content": "2.3. Generative Image Dynamics" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 426, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 715 + ], + "type": "text", + "content": "With the advancement of generative models, there is growing interest in leveraging these powerful techniques to synthesize dynamic animations from static images, guided by motion features extracted from various user inputs. These inputs can be sparse, such as text prompts [7, 41, 71], trajectories [42, 55, 62, 67, 78], or camera movements [62, 73], or dense, like reference videos [28, 63, 80]. To achieve controllable dynamics, prior works often incorporate ControlNet [76] into image or video generative models during the decoding stage, utilizing motion features such as Canny edges, depth maps [17], 2D Gaussian maps [67], and optical flow maps [55]. In addition, several methods employ specialized motion fusion modules [41, 62, 78], spatial or temporal attention mechanisms [28, 63, 73], and feature injection strategies [42] to guide the generation of controllable dynamics. However, these approaches typically lack physics-based supervision, which can result in animations that violate physical laws or fail to align with user intentions [47]. Recently, PhysGen [44] integrated a rigid-body physics simulator to generate physics-consistent dynamics for foreground objects in a given image. While effective, PhysGen is limited to 2D rigid-body motions, making it unsuitable for the fluid, elastic effects commonly seen in anime, such as the waving of cloth or hair, which do not" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "conform to rigid-body dynamics. In this work, we address these limitations by leveraging deformable-body dynamics and a sketch-guided video diffusion model to create high-quality, physically consistent animation sequences that capture the fluidity and elasticity characteristic of anime." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 142, + 370, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 370, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 370, + 155 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "text", + "content": "Given a reference anime illustration " + }, + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "text", + "content": ", our goal is to generate a sequence of stylized video frames " + }, + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_T\\}" + }, + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "text", + "content": " with natural, user-guided motions. To achieve this, we introduce PhysAnimator, a novel framework that combines video diffusion models with physics-based animation. We start by segmenting the objects of interest in a given anime illustration and creating a 2D triangulated mesh representation. Using this geometry, we obtain dynamics through image-space deformable body simulation, generating a sequence of optical flow fields " + }, + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_{0 \\to 1}, \\mathcal{F}_{0 \\to 2}, \\dots, \\mathcal{F}_{0 \\to T}\\}" + }, + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "text", + "content": ". To offer user control, we enable interactive inputs through customizable energy strokes for defining external forces and rigging points for specifying desired trajectories. To render dynamics as high-quality frames, we extract the sketch from the input illustration and warp it using the computed optical flow fields, yielding a sequence of dynamic sketch frames " + }, + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\{S_1, S_2, \\dots, S_T\\}" + }, + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "text", + "content": ". These sketches, together with the input illustration, are fed into a video diffusion model with a sketch-guided ControlNet [76], synthesizing a vivid animated sequence. Finally, an optional data-driven cartoon interpolation model [70] can be applied to enhance the anime style dynamics of the results by selecting keyframes from the animated sequence as input, yielding the final output frames " + }, + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}_1, \\hat{I}_2, \\dots, \\hat{I}_T\\}" + }, + { + "bbox": [ + 313, + 162, + 555, + 462 + ], + "type": "text", + "content": ". An overview of our proposed framework is shown in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 468, + 400, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 400, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 400, + 480 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 487, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 583 + ], + "type": "text", + "content": "Latent Video Diffusion Model Most video diffusion models build on the latent diffusion model (LDM) framework [53], which uses Variational Autoencoder [32] (VAE) to map input images into a latent space. In this space, data is transformed into Gaussian noise via a forward diffusion process, which the model learns to reverse through denoising. In the forward diffusion process, the latent code " + }, + { + "bbox": [ + 313, + 487, + 554, + 583 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 487, + 554, + 583 + ], + "type": "text", + "content": " is perturbed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 354, + 590, + 555, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 590, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 354, + 590, + 555, + 604 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}", + "image_path": "56996176c6603f24b1ae92485ce82b6532b0b7a92de0f1b3885dc22d780edf1d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_t)" + }, + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "text", + "content": " controlling the noise strength coefficient at step " + }, + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "text", + "content": ". The denoising model " + }, + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta" + }, + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "text", + "content": " is trained to recover " + }, + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 313, + 613, + 554, + 660 + ], + "type": "text", + "content": " by minimizing the objective function:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 384, + 668, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 668, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 384, + 668, + 553, + 681 + ], + "type": "interline_equation", + "content": "L _ {\\epsilon} = \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}; c, t) \\| _ {2} ^ {2}, \\tag {2}", + "image_path": "aacefe65ee167d3c83287df15d8ac6503639a667e9de14d0437754b12eaa80c2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " denotes the learnable network parameters and " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " represents conditioning input (e.g. text prompts or images)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10795" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 254 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 254 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 254 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 254 + ], + "type": "image", + "image_path": "9ac3ae24b651918e23a47af2a20c77497bd842cbd723639be066e332273292e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 261, + 555, + 318 + ], + "lines": [ + { + "bbox": [ + 54, + 261, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 555, + 318 + ], + "type": "text", + "content": "Figure 2. Method Overview. We begin by segmenting the object and creating a triangulated deformable mesh. Physics-based simulations are then used to generate dynamic optical flow fields, with users given the option to guide the motion through customizable energy strokes (shown as orange arrows) and rigging points (shown as red dots). The extracted sketch is warped using the computed optical flow and refined with a sketch-guided video diffusion model, producing a smooth, stylized animation sequence. Optionally, a cartoon interpolation model can further be applied to enhance the animation with expressive dynamics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 337, + 296, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 337, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 54, + 337, + 296, + 398 + ], + "type": "text", + "content": "After denoising, the VAE decoder reconstructs the latent code back into the image space. The latent video diffusion model (LVDM) extends the image LDM to videos by incorporating temporal modules to maintain temporal consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "text", + "content": "Deformable Body Dynamics Mathematically, the dynamics of a continuum deformable body is described by a time-dependent continuous deformation map " + }, + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} = \\phi(\\boldsymbol{X}, t)" + }, + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "text", + "content": ", which maps the undeformed material space " + }, + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\Omega^0" + }, + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "text", + "content": " to the deformed world space " + }, + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\Omega^t" + }, + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "text", + "content": ". The deformation gradient " + }, + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\boldsymbol{F} = \\frac{\\partial \\phi}{\\partial \\boldsymbol{X}}" + }, + { + "bbox": [ + 55, + 406, + 296, + 514 + ], + "type": "text", + "content": " encodes the local deformation such as scaling, rotation, and shearing. In the context of a discretized setting, e.g. 2D meshes, the deformation map for each triangle can be expressed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 523, + 295, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 523, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 131, + 523, + 295, + 537 + ], + "type": "interline_equation", + "content": "\\phi_ {i} (\\boldsymbol {X}) = \\boldsymbol {F} _ {i} \\boldsymbol {X} + \\boldsymbol {b} _ {i}, \\tag {3}", + "image_path": "9e14473504977e14cdcfb5778c8cd513c9e95f1b86028918225459bc5281b9bd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 544, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{b}_i\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 544, + 296, + 641 + ], + "type": "text", + "content": " accounts for translation of " + }, + { + "bbox": [ + 55, + 544, + 296, + 641 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 544, + 296, + 641 + ], + "type": "text", + "content": "-th triangle, and the deformation gradient " + }, + { + "bbox": [ + 55, + 544, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\pmb {F}_i\\in \\mathbb{R}^{2\\times 2}" + }, + { + "bbox": [ + 55, + 544, + 296, + 641 + ], + "type": "text", + "content": " is assumed to be constant [40]. Undergoing deformation, the deformed body aims to recover its rest shape via resisting forces. Continuum mechanics model this behavior by first defining an energy density function " + }, + { + "bbox": [ + 55, + 544, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\Psi (F(x))" + }, + { + "bbox": [ + 55, + 544, + 296, + 641 + ], + "type": "text", + "content": ", which measures the strain energy per unit undeformed volume. The total potential energy for a deformable body is then obtained as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 650, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 650, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 129, + 650, + 295, + 681 + ], + "type": "interline_equation", + "content": "E (\\boldsymbol {x}) = \\sum_ {i = 1} ^ {N} \\Psi \\left(\\boldsymbol {F} _ {i}\\right) V _ {i}, \\tag {4}", + "image_path": "b26f4aeb399225c2a9c4079ee849c9aa0eb14945a4ab5b9e98197b83e0ae820b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " denotes volume of " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "-th triangle. The internal resisting force is then defined as the negative gradient of the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 338, + 520, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 520, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 520, + 350 + ], + "type": "text", + "content": "potential energy with respect to the vertex position" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 392, + 357, + 553, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 357, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 392, + 357, + 553, + 383 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {\\text {i n t}} (\\boldsymbol {x}) = - \\frac {\\partial E (\\boldsymbol {x})}{\\partial \\boldsymbol {x}}. \\tag {5}", + "image_path": "62f92e995ff22ab4b26c891ac0784b1422029899a7ed72f03c7d0122109ff46a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 389, + 555, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 412 + ], + "type": "text", + "content": "The deformable body dynamics is governed by Newton's Second Law as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 419, + 553, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 419, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 365, + 419, + 553, + 444 + ], + "type": "interline_equation", + "content": "\\frac {d ^ {2} \\boldsymbol {x}}{d t ^ {2}} = \\boldsymbol {M} ^ {- 1} \\left(\\boldsymbol {f} _ {\\text {i n t}} (\\boldsymbol {x}) + \\boldsymbol {f} _ {\\text {e x t}} (\\boldsymbol {x})\\right), \\tag {6}", + "image_path": "4322da25dbe299bbafab83e449dad59674d7ab1d62f53cad761884dae2677ba7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 451, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 487 + ], + "type": "text", + "content": "which can be solved numerically. Here " + }, + { + "bbox": [ + 313, + 451, + 555, + 487 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 451, + 555, + 487 + ], + "type": "text", + "content": " is the mass matrix that represents the masses of all vertices and " + }, + { + "bbox": [ + 313, + 451, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\mathrm{ext}}" + }, + { + "bbox": [ + 313, + 451, + 555, + 487 + ], + "type": "text", + "content": " refers to external forces, such as gravity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 495, + 455, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 455, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 455, + 507 + ], + "type": "text", + "content": "3.2. Physics-based Animation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 513, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 597 + ], + "type": "text", + "content": "To generate dynamic animations from a single illustration, a straightforward approach is to use existing controllable video diffusion models [55, 67]. However, these purely data-driven methods often struggle due to a lack of physical understanding, leading to unrealistic results. To overcome this limitation, we incorporate physics-based animation to generate physically consistent and plausible motions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Geometry Registration The first step to animate objects of interest in a given illustration is to establish a geometric basis. While previous works [24, 59] have explored predicting 3D geometry from single-view images, these methods are typically constrained to reconstructing real-world objects and do not generalize to in-the-wild anime images, since characters and objects in anime often lack an inherent 3D representation due to their stylized and flat nature. As shown in prior works on cartoon" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10796" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": "animation [31, 69], leveraging 2D animation techniques proves effective for generating anime-style dynamic effects. Inspired by this," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 143, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 143, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 143, + 204 + ], + "type": "text", + "content": "we focus on extracting 2D meshes for the objects of interest. To achieve this, we first utilize the Segment Anything Model (SAM) [33, 52] to" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 151, + 110, + 294, + 200 + ], + "blocks": [ + { + "bbox": [ + 151, + 110, + 294, + 200 + ], + "lines": [ + { + "bbox": [ + 151, + 110, + 294, + 200 + ], + "spans": [ + { + "bbox": [ + 151, + 110, + 294, + 200 + ], + "type": "image", + "image_path": "140153bd5c609ced0978fbb8b624bc67b512235fdf3ee7991c891825297bd262.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 205, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 299 + ], + "type": "text", + "content": "obtain segmentation masks for each target object, guided by user-specified query points. Along the contours of each segmentation mask, we uniformly sample boundary points, defining the outline of the intended mesh. We then employ conforming Delaunay triangulation [37] with these boundary constraints to generate well-structured triangular meshes, creating a simulation-ready format for subsequent animation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 479 + ], + "type": "text", + "content": "Deformable Body Model Anime scenes often feature dynamic environmental effects such as external forces that interact with characters' clothing, hair, or other elements. Additionally, techniques like \"squash\" and \"stretch\" are commonly used to convey motion and energy, adding expressiveness to animated objects. Motivated by these stylized dynamics, we model anime objects as deformable bodies, allowing them to capture the fluidity and exaggerated motion characteristic of anime. A key property of deformable bodies is their ability to change shape in response to external forces and attempt to return to their original rest shape upon deformation. To represent this physical behavior, we employ the Fixed Corotated Constitutive model [57], which defines the energy density as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 489, + 295, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 489, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 89, + 489, + 295, + 513 + ], + "type": "interline_equation", + "content": "\\Psi (\\boldsymbol {F}) = \\mu \\| \\boldsymbol {F} - \\boldsymbol {R} \\| _ {F} ^ {2} + \\frac {\\lambda}{2} (\\det (\\boldsymbol {F}) - 1) ^ {2}, \\tag {7}", + "image_path": "e8327e1894b6281db906e87f96737e892148508d841a7e88406acb129497d662.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mu, \\lambda" + }, + { + "bbox": [ + 55, + 521, + 296, + 594 + ], + "type": "text", + "content": " are the Lamé parameters, and " + }, + { + "bbox": [ + 55, + 521, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 55, + 521, + 296, + 594 + ], + "type": "text", + "content": " is the rotational part of " + }, + { + "bbox": [ + 55, + 521, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 55, + 521, + 296, + 594 + ], + "type": "text", + "content": " computed via polar decomposition. The first term models the stretching and compression resistance for the individual spatial directions and the second term describes resistance to volume change. The resulting internal forces " + }, + { + "bbox": [ + 55, + 521, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\mathrm{int}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 594 + ], + "type": "text", + "content": " can then be derived via Eq. (5)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Interactive Animation While " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{int}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " governs the inherent physical behavior, " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{ext}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " allows for user-defined interactions. Inspired by the concept of energy strokes [15, 69], we introduce customizable energy strokes that carry flow particles. These particles move along the user-specified strokes, propagating external forces to nearby vertices of the deformable mesh. This enables the creation of tailored animation effects, such as simulating wind interactions. Additionally, we incorporate rigging point support, allowing animators to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "anchor specific regions or guide them along predefined trajectories, offering enhanced control and flexibility. Using the specified energy strokes and deformable modeling, we solve the motion equation (Eq. (6)) to evolve the dynamics. We employ the semi-implicit Euler method to compute the deformation map " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\phi_i(X,t)" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " for every triangle. For each pixel " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_p" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " in the reference image, if it lies within a triangle " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": ", we assign the displacement vector as " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\pmb{d}(\\pmb{p}) = \\phi_i(\\pmb{X}_p,t) - \\pmb{X}_p" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "; otherwise, we assign a zero vector. Collecting displacement vector " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\pmb{d}(\\pmb{p})" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " of all pixels yields the optical flow " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{0\\rightarrow t}" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": ", which defines the displacement fields for the reference illustration " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 222, + 441, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 441, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 441, + 235 + ], + "type": "text", + "content": "3.3. Generative Rendering" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 240, + 555, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 288 + ], + "type": "text", + "content": "In this section, we describe how to render the video sequence " + }, + { + "bbox": [ + 313, + 240, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}_1,\\hat{I}_2,\\dots,\\hat{I}_T\\}" + }, + { + "bbox": [ + 313, + 240, + 555, + 288 + ], + "type": "text", + "content": " given the reference image " + }, + { + "bbox": [ + 313, + 240, + 555, + 288 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 313, + 240, + 555, + 288 + ], + "type": "text", + "content": " and the optical flow sequence " + }, + { + "bbox": [ + 313, + 240, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_{0\\to 1},\\mathcal{F}_{0\\to 2},\\ldots ,\\mathcal{F}_{0\\to T}\\}" + }, + { + "bbox": [ + 313, + 240, + 555, + 288 + ], + "type": "text", + "content": " derived from simulation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 297, + 554, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 554, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 554, + 368 + ], + "type": "text", + "content": "Sketch-Guided Rendering While the video sequence can be directly generated by warping the reference image " + }, + { + "bbox": [ + 313, + 297, + 554, + 368 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 313, + 297, + 554, + 368 + ], + "type": "text", + "content": " using optical flow fields, the resulting frames often exhibit black hole artifacts caused by occlusions. To address this problem, we extract the sketch of " + }, + { + "bbox": [ + 313, + 297, + 554, + 368 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 313, + 297, + 554, + 368 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 297, + 554, + 368 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 313, + 297, + 554, + 368 + ], + "type": "text", + "content": ", and obtain the sketch at time " + }, + { + "bbox": [ + 313, + 297, + 554, + 368 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 297, + 554, + 368 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 378, + 375, + 553, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 375, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 378, + 375, + 553, + 388 + ], + "type": "interline_equation", + "content": "S _ {t} = \\mathcal {W} \\left(S _ {0}, \\mathcal {F} _ {0 \\rightarrow t}, w _ {0 \\rightarrow t}\\right), \\tag {8}", + "image_path": "9cabe51c9aa3e74c6e5cc751a675095d2758206ab4e7031c2be535554f08bd89.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "content": " represents the forward-warping operator, and " + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "inline_equation", + "content": "w_{0\\rightarrow t}" + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "content": " denotes the warping weights for each pixel in " + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "content": ". In forward warping, multiple pixels may map to the same 2D location in the output frame [48], potentially leading to artifacts or distortions if the weights " + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "inline_equation", + "content": "w_{0\\rightarrow t}" + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "content": " are not properly defined. Inspired by [42], we set the pixel weight as " + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "inline_equation", + "content": "w_{0\\rightarrow t}(\\pmb {p}) = \\| \\mathcal{F}_{0\\rightarrow t}(\\pmb {p})\\| _2" + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "content": ", giving higher importance to pixels with larger motion, typically corresponding to foreground objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 502, + 555, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 681 + ], + "type": "text", + "content": "Next, we leverage a video diffusion model to generate the rendered frames " + }, + { + "bbox": [ + 313, + 502, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\{I_1^r, I_2^r, \\dots, I_T^r\\}" + }, + { + "bbox": [ + 313, + 502, + 555, + 681 + ], + "type": "text", + "content": " using the obtained sketch sequence " + }, + { + "bbox": [ + 313, + 502, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\{S_1, S_2, \\dots, S_T\\}" + }, + { + "bbox": [ + 313, + 502, + 555, + 681 + ], + "type": "text", + "content": ". The diffusion model uses the reference image " + }, + { + "bbox": [ + 313, + 502, + 555, + 681 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 313, + 502, + 555, + 681 + ], + "type": "text", + "content": " as an input, and we employ a ControlNet [76] with the sketches as control signals, guiding the generation process to ensure that the results align with the sketch inputs. We observed that during the inference time, due to segmentation inaccuracies, the optical flow warping may introduce unintended distortions, or miss parts of the object contour, generating imperfect sketches. To address this, we apply Gaussian blur to the input sketches at both training and inference time, which smoothes out inconsistencies. The video diffusion model is then capable of refining the results, leveraging its generative capabilities to handle imperfections and produce coherent outputs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Complementary Dynamics Unlike motions in the real world, dynamic effects in animation do not strictly adhere to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10797" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "text", + "content": "physical laws and can not be fully captured by 2D animation methods. In the industrial animation pipeline, artists typically begin by creating a series of keyframes that define the primary motion trajectory, followed by drawing in-between frames to ensure smooth and fluid transitions. Inspired by this workflow, we leverage a data-driven module to enhance the physics-based animated results. Specifically, we select keyframes from the sketch-guided rendering results, forming the keyframe sequence " + }, + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "inline_equation", + "content": "\\{I_0^r, I_n^r, I_{2n}^r, \\dots, I_{in}^r\\}" + }, + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "text", + "content": " denotes the gap between keyframes. We then employ a cartoon interpolation video diffusion model [70], which synthesizes intermediate frames " + }, + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}_{in+1}, \\hat{I}_{in+2}, \\dots, \\hat{I}_{i(n+1)-1}\\}" + }, + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "text", + "content": " between each adjacent keyframe pair " + }, + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "inline_equation", + "content": "(I_{in}^r, I_{i(n+1)}^r)" + }, + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "text", + "content": ". For keyframes, we simply assign " + }, + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "inline_equation", + "content": "\\hat{I}_{in} = I_{in}^r" + }, + { + "bbox": [ + 55, + 72, + 294, + 280 + ], + "type": "text", + "content": ". This approach allows us to introduce expressive, data-driven complementary dynamics that go beyond what can be achieved through physics-based animation alone." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 136, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 136, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 136, + 304 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 360 + ], + "type": "text", + "content": "In this section, we conduct a comprehensive comparison of our method against existing video diffusion models and demonstrate that our approach generates high-quality and physically plausible animations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 556 + ], + "type": "text", + "content": "Implementation Details We implement our deformable body simulator using Taichi [25]. During interactive animation, users can adjust the Lamé parameters " + }, + { + "bbox": [ + 55, + 365, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 365, + 295, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 365, + 295, + 556 + ], + "type": "text", + "content": " to control the characteristics of the objects based on their specific needs. For generative rendering, we follow LVCD [26] to train a sketch-guided ControlNet for the stable video diffusion model [4], using blurred sketches as control signals to address potential imperfections in the predicted sketch sequence. When utilizing ToonCrafter [70] for generating in-between frames, similar to its original implementation, we also train an additional sketch-guided ControlNet, but set the control scale to 0.1 during inference. We also set " + }, + { + "bbox": [ + 55, + 365, + 295, + 556 + ], + "type": "inline_equation", + "content": "n = 15" + }, + { + "bbox": [ + 55, + 365, + 295, + 556 + ], + "type": "text", + "content": " for the in-betweening step. This configuration helps the generated frames follow the rough motion indicated by the sketches while introducing extra stylized animation details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 563, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 659 + ], + "type": "text", + "content": "Dataset We build our training dataset using the Sakuga-42M Dataset [50], which consists of 1.4 million animation video clips. To ensure high-quality training samples, we filter out clips with fewer than 24 frames and dynamics scores outside the range of 0.05 to 0.7. Next, we extract sketches from the selected clips using the method from [5], resulting in a final dataset of 380,000 pairs of sketches and corresponding video sequences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Baseline We compare our method against two categories of video generation approaches. The first category includes state-of-the-art Image-to-Video (I2V) models, such as Cinemo [46] and DynamamiCrafter [71], which take an input im" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "age and use text prompts to guide the motion dynamics of the generated videos. For these models, we generate prompts describing the image content and expected motion using ChatGPT-4V. The second category, including Motion-I2V [55] and Drag Anything [67], usually trains an additional motion-control module for video generative models such that, given a single image input, a user-specified trajectory can be provided to control the movement of objects in the generated frames. To ensure a fair comparison, we extract trajectories from our animated results and use them as input for these methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 214, + 449, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 449, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 449, + 227 + ], + "type": "text", + "content": "4.1.Quantitative Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 234, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 553, + 437 + ], + "type": "text", + "content": "Given the absence of established benchmarks for animestyle image-to-video generation, we follow Motion-I2V [55] and construct a test set comprising 20 anime images with stylized dynamic elements such as swaying hair, clothing, and plants. For each method, we generate 10 video samples per image, resulting in a total of 200 videos per method. To quantitatively evaluate the generated results, we use the Fréchet Inception Distance (FID) [22] to measure the similarity between the generated frames and the reference images. Additionally, we employ VideoScore [19] to assess multiple aspects of video quality, including visual quality, temporal consistency, dynamic degree, and factual consistency. The factual consistency evaluates the consistency of the video content with common sense and factual knowledge. We exclude the text alignment score as our method and the trajectory-controlled approach do not involve text input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 439, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 553, + 715 + ], + "type": "text", + "content": "The quantitative evaluation results, presented in Tab. 1, demonstrate that our method outperforms baseline approaches across most metrics, including visual quality, temporal consistency, and factual consistency, demonstrating the overall high quality of our generated videos. As shown in Fig. 3, Cinema tends to generate static videos that closely resemble the reference image, leading to better FID scores. While the dynamic degree score for DragAnything is notably higher than other methods, this is due to its tendency to misinterpret motion control as camera movement, causing shifts in the entire image space and resulting in an inflated dynamics score. Although our method's dynamic degree score is slightly lower than that of DynamiCrafter and Motion-I2V, these methods often exhibit implausible, distorted motions that create an illusion of increased dynamics. In contrast, our approach produces geometry-consistent motions and achieves high factual consistency, demonstrating the realistic motion dynamics enabled by our physics-based modeling. In addition, following [14, 34], we conducted an user study adopting a two-alternative forced choice (2AFC) protocol, where participants are asked to choose the preferred video based on temporal consistency, visual quality, motion plausibility, and overall feeling given" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10798" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 64, + 129, + 102 + ], + "blocks": [ + { + "bbox": [ + 86, + 64, + 129, + 102 + ], + "lines": [ + { + "bbox": [ + 86, + 64, + 129, + 102 + ], + "spans": [ + { + "bbox": [ + 86, + 64, + 129, + 102 + ], + "type": "image", + "image_path": "72996d9e6b9751afae14e0bb6aeb4463b985e19af012218324573ac3d0587af2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 134, + 70, + 304, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 70, + 304, + 95 + ], + "spans": [ + { + "bbox": [ + 134, + 70, + 304, + 95 + ], + "type": "text", + "content": "\"A pirate flag flutters and waves vigorously in the wind.\"" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 63, + 372, + 102 + ], + "blocks": [ + { + "bbox": [ + 310, + 63, + 372, + 102 + ], + "lines": [ + { + "bbox": [ + 310, + 63, + 372, + 102 + ], + "spans": [ + { + "bbox": [ + 310, + 63, + 372, + 102 + ], + "type": "image", + "image_path": "5f6a3c80c2fc9498f6da9c46bc693d9bcb2ab8161a6d3403689f1b3c652b4d50.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 374, + 70, + 536, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 70, + 536, + 97 + ], + "spans": [ + { + "bbox": [ + 374, + 70, + 536, + 97 + ], + "type": "text", + "content": "\"The figure's hair flows gently with the breeze, strands swaying slightly.\"" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 86, + 103, + 125, + 140 + ], + "blocks": [ + { + "bbox": [ + 86, + 103, + 125, + 140 + ], + "lines": [ + { + "bbox": [ + 86, + 103, + 125, + 140 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 125, + 140 + ], + "type": "image", + "image_path": "404d853408afd38d611ea5f6b88aa006c8355ef4e9410a93b187c9d0c1710d1e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 147, + 103, + 185, + 140 + ], + "blocks": [ + { + "bbox": [ + 147, + 103, + 185, + 140 + ], + "lines": [ + { + "bbox": [ + 147, + 103, + 185, + 140 + ], + "spans": [ + { + "bbox": [ + 147, + 103, + 185, + 140 + ], + "type": "image", + "image_path": "bc796bb3fda26fafe75e51079a48d33857fbc2b314c053260d53087b9faf919c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 206, + 103, + 244, + 140 + ], + "blocks": [ + { + "bbox": [ + 206, + 103, + 244, + 140 + ], + "lines": [ + { + "bbox": [ + 206, + 103, + 244, + 140 + ], + "spans": [ + { + "bbox": [ + 206, + 103, + 244, + 140 + ], + "type": "image", + "image_path": "e6fd15c76e952ce857490e897b6c7471314099af6fcf3a3220929bbfe951049f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 266, + 103, + 304, + 140 + ], + "blocks": [ + { + "bbox": [ + 266, + 103, + 304, + 140 + ], + "lines": [ + { + "bbox": [ + 266, + 103, + 304, + 140 + ], + "spans": [ + { + "bbox": [ + 266, + 103, + 304, + 140 + ], + "type": "image", + "image_path": "82bb8f4e5333ca9eaf049c74d0c9edde9da2cd0d7c95121161c9d46f97344060.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 86, + 141, + 124, + 179 + ], + "blocks": [ + { + "bbox": [ + 86, + 141, + 124, + 179 + ], + "lines": [ + { + "bbox": [ + 86, + 141, + 124, + 179 + ], + "spans": [ + { + "bbox": [ + 86, + 141, + 124, + 179 + ], + "type": "image", + "image_path": "24e88cc06546baeffba54954e68a91f4fa1173bf60f5393a63c08ea4cac07f8c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 134, + 141, + 178, + 179 + ], + "blocks": [ + { + "bbox": [ + 134, + 141, + 178, + 179 + ], + "lines": [ + { + "bbox": [ + 134, + 141, + 178, + 179 + ], + "spans": [ + { + "bbox": [ + 134, + 141, + 178, + 179 + ], + "type": "image", + "image_path": "04f3731173654c475ffb7fd9c30f2c1809341d9503951d44ef5af64a7bb35b0b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 192, + 141, + 237, + 179 + ], + "blocks": [ + { + "bbox": [ + 192, + 141, + 237, + 179 + ], + "lines": [ + { + "bbox": [ + 192, + 141, + 237, + 179 + ], + "spans": [ + { + "bbox": [ + 192, + 141, + 237, + 179 + ], + "type": "image", + "image_path": "58c14f62647bcb91aa478a1bce85d8a15b11e36600ba8e5ae618c8e473bdfef1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 266, + 141, + 304, + 179 + ], + "blocks": [ + { + "bbox": [ + 266, + 141, + 304, + 179 + ], + "lines": [ + { + "bbox": [ + 266, + 141, + 304, + 179 + ], + "spans": [ + { + "bbox": [ + 266, + 141, + 304, + 179 + ], + "type": "image", + "image_path": "ddd92798a4ac91f99ac2850046ba3dfacf59eabd18b51c2226974f853b224327.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 86, + 182, + 125, + 219 + ], + "blocks": [ + { + "bbox": [ + 86, + 182, + 125, + 219 + ], + "lines": [ + { + "bbox": [ + 86, + 182, + 125, + 219 + ], + "spans": [ + { + "bbox": [ + 86, + 182, + 125, + 219 + ], + "type": "image", + "image_path": "fc33df2cfd66b29ecc6ccc76b6cb04203cbd81a62b3cec85e3c3d616d69ef6db.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 145, + 182, + 184, + 219 + ], + "blocks": [ + { + "bbox": [ + 145, + 182, + 184, + 219 + ], + "lines": [ + { + "bbox": [ + 145, + 182, + 184, + 219 + ], + "spans": [ + { + "bbox": [ + 145, + 182, + 184, + 219 + ], + "type": "image", + "image_path": "30ac7c1cdb46509abfd92cc5cc9e7874aca21c1741866f285908461e8c9623da.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 204, + 182, + 244, + 219 + ], + "blocks": [ + { + "bbox": [ + 204, + 182, + 244, + 219 + ], + "lines": [ + { + "bbox": [ + 204, + 182, + 244, + 219 + ], + "spans": [ + { + "bbox": [ + 204, + 182, + 244, + 219 + ], + "type": "image", + "image_path": "b5ec722bf208bce94f84944ac5248afaf92946df1149a95b829ceccdbc55dfb9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 266, + 182, + 304, + 219 + ], + "blocks": [ + { + "bbox": [ + 266, + 182, + 304, + 219 + ], + "lines": [ + { + "bbox": [ + 266, + 182, + 304, + 219 + ], + "spans": [ + { + "bbox": [ + 266, + 182, + 304, + 219 + ], + "type": "image", + "image_path": "05eb25d16d18b7c154fe6faf6ea1622b1ed77cfda47ea7e625b36ad869aeb1a4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 86, + 220, + 124, + 258 + ], + "blocks": [ + { + "bbox": [ + 86, + 220, + 124, + 258 + ], + "lines": [ + { + "bbox": [ + 86, + 220, + 124, + 258 + ], + "spans": [ + { + "bbox": [ + 86, + 220, + 124, + 258 + ], + "type": "image", + "image_path": "4bdc00d6c2ba77f25faa1eb3aae90c99a34fd37db77316ae68b160f8b86d2b4d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 147, + 221, + 184, + 258 + ], + "blocks": [ + { + "bbox": [ + 147, + 221, + 184, + 258 + ], + "lines": [ + { + "bbox": [ + 147, + 221, + 184, + 258 + ], + "spans": [ + { + "bbox": [ + 147, + 221, + 184, + 258 + ], + "type": "image", + "image_path": "15bc45417c47c863f62828292f6929272dbaccd8340dc48c75e2cbc2ac818e48.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 205, + 223, + 244, + 258 + ], + "blocks": [ + { + "bbox": [ + 205, + 223, + 244, + 258 + ], + "lines": [ + { + "bbox": [ + 205, + 223, + 244, + 258 + ], + "spans": [ + { + "bbox": [ + 205, + 223, + 244, + 258 + ], + "type": "image", + "image_path": "0d870129f6963e724fc525220cb14d7525dde5a948e309d169f3248abeddd0f0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 266, + 222, + 304, + 258 + ], + "blocks": [ + { + "bbox": [ + 266, + 222, + 304, + 258 + ], + "lines": [ + { + "bbox": [ + 266, + 222, + 304, + 258 + ], + "spans": [ + { + "bbox": [ + 266, + 222, + 304, + 258 + ], + "type": "image", + "image_path": "9efaba4201d08f3f6d23866deb648175b56de9b41caff8efcbbdbcec1f6e45da.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 86, + 259, + 124, + 297 + ], + "blocks": [ + { + "bbox": [ + 86, + 259, + 124, + 297 + ], + "lines": [ + { + "bbox": [ + 86, + 259, + 124, + 297 + ], + "spans": [ + { + "bbox": [ + 86, + 259, + 124, + 297 + ], + "type": "image", + "image_path": "f8b6606a892fd79f847cb649b378777cc8f78fdcdc9ebc911a1d6a242f9f5d9d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 143, + 259, + 183, + 297 + ], + "blocks": [ + { + "bbox": [ + 143, + 259, + 183, + 297 + ], + "lines": [ + { + "bbox": [ + 143, + 259, + 183, + 297 + ], + "spans": [ + { + "bbox": [ + 143, + 259, + 183, + 297 + ], + "type": "image", + "image_path": "3cd1cc252d0646a37f9d5f726694b71ed9f762d0fa82637ffc14fc67db9f1e04.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 202, + 259, + 243, + 297 + ], + "blocks": [ + { + "bbox": [ + 202, + 259, + 243, + 297 + ], + "lines": [ + { + "bbox": [ + 202, + 259, + 243, + 297 + ], + "spans": [ + { + "bbox": [ + 202, + 259, + 243, + 297 + ], + "type": "image", + "image_path": "34033b9d0993fcc5d3851566a841208ec90db1ac421575d44d6cb4cb9dcef814.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 260, + 259, + 304, + 297 + ], + "blocks": [ + { + "bbox": [ + 260, + 259, + 304, + 297 + ], + "lines": [ + { + "bbox": [ + 260, + 259, + 304, + 297 + ], + "spans": [ + { + "bbox": [ + 260, + 259, + 304, + 297 + ], + "type": "image", + "image_path": "fd5d0372f9e2e433499231924428215ff8d1079a59a1ef2d5ec3fab5049a4cdc.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 310, + 103, + 367, + 140 + ], + "blocks": [ + { + "bbox": [ + 310, + 103, + 367, + 140 + ], + "lines": [ + { + "bbox": [ + 310, + 103, + 367, + 140 + ], + "spans": [ + { + "bbox": [ + 310, + 103, + 367, + 140 + ], + "type": "image", + "image_path": "e2779d2521a555952da799825be49094fb68ed374e8c3493021ece3d5b1f0a5b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 368, + 103, + 430, + 140 + ], + "blocks": [ + { + "bbox": [ + 368, + 103, + 430, + 140 + ], + "lines": [ + { + "bbox": [ + 368, + 103, + 430, + 140 + ], + "spans": [ + { + "bbox": [ + 368, + 103, + 430, + 140 + ], + "type": "image", + "image_path": "4c945aa6c93950d127da908ea999fe2043baec3ea022b90228127cd398f03a29.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 432, + 103, + 550, + 140 + ], + "blocks": [ + { + "bbox": [ + 432, + 103, + 550, + 140 + ], + "lines": [ + { + "bbox": [ + 432, + 103, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 432, + 103, + 550, + 140 + ], + "type": "image", + "image_path": "7c6471f274fc50fae62918089bf9301471bc8c5ce0ac9aaf3c3ee042f286c4de.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 310, + 142, + 367, + 179 + ], + "blocks": [ + { + "bbox": [ + 310, + 142, + 367, + 179 + ], + "lines": [ + { + "bbox": [ + 310, + 142, + 367, + 179 + ], + "spans": [ + { + "bbox": [ + 310, + 142, + 367, + 179 + ], + "type": "image", + "image_path": "68eb0c96e615cf2f829a969d738d1ce61702c94346109cd2eb24e6c4d14966c7.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 368, + 143, + 430, + 179 + ], + "blocks": [ + { + "bbox": [ + 368, + 143, + 430, + 179 + ], + "lines": [ + { + "bbox": [ + 368, + 143, + 430, + 179 + ], + "spans": [ + { + "bbox": [ + 368, + 143, + 430, + 179 + ], + "type": "image", + "image_path": "31e80205d0e1c6f9453dab615a0ebedd0d5f054a7d854eb940505391c2420fca.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 432, + 143, + 550, + 179 + ], + "blocks": [ + { + "bbox": [ + 432, + 143, + 550, + 179 + ], + "lines": [ + { + "bbox": [ + 432, + 143, + 550, + 179 + ], + "spans": [ + { + "bbox": [ + 432, + 143, + 550, + 179 + ], + "type": "image", + "image_path": "3a60bec4bdfcf020ae33ac6bc66cff6f159935fb56f7442c9efa6a6f92151b15.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 310, + 181, + 367, + 220 + ], + "blocks": [ + { + "bbox": [ + 310, + 181, + 367, + 220 + ], + "lines": [ + { + "bbox": [ + 310, + 181, + 367, + 220 + ], + "spans": [ + { + "bbox": [ + 310, + 181, + 367, + 220 + ], + "type": "image", + "image_path": "d71bcfe7097d17daf22ec939e35a84ef9cb4fa2026761227c83d560e575b562e.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 368, + 181, + 430, + 220 + ], + "blocks": [ + { + "bbox": [ + 368, + 181, + 430, + 220 + ], + "lines": [ + { + "bbox": [ + 368, + 181, + 430, + 220 + ], + "spans": [ + { + "bbox": [ + 368, + 181, + 430, + 220 + ], + "type": "image", + "image_path": "42f532c90fa66c9ae9d663fb821cdb2fdb2f07cce3a9f753c94df3f2c126fea0.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 432, + 181, + 550, + 220 + ], + "blocks": [ + { + "bbox": [ + 432, + 181, + 550, + 220 + ], + "lines": [ + { + "bbox": [ + 432, + 181, + 550, + 220 + ], + "spans": [ + { + "bbox": [ + 432, + 181, + 550, + 220 + ], + "type": "image", + "image_path": "d3ee0012edc3b554001c934479da90ba48898af2cdeff2a7f9f34ff991d99ee5.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 310, + 220, + 367, + 258 + ], + "blocks": [ + { + "bbox": [ + 310, + 220, + 367, + 258 + ], + "lines": [ + { + "bbox": [ + 310, + 220, + 367, + 258 + ], + "spans": [ + { + "bbox": [ + 310, + 220, + 367, + 258 + ], + "type": "image", + "image_path": "e5f70c9c586a0c7f39093e324545c8d734b3f7079d5ee38c6a0da7472e863766.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 368, + 220, + 430, + 258 + ], + "blocks": [ + { + "bbox": [ + 368, + 220, + 430, + 258 + ], + "lines": [ + { + "bbox": [ + 368, + 220, + 430, + 258 + ], + "spans": [ + { + "bbox": [ + 368, + 220, + 430, + 258 + ], + "type": "image", + "image_path": "a84de4bc1ad003696ce8d6ddde21e141cd24f1f6eee9ff95375c9a712039ecdb.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 432, + 220, + 550, + 258 + ], + "blocks": [ + { + "bbox": [ + 432, + 220, + 550, + 258 + ], + "lines": [ + { + "bbox": [ + 432, + 220, + 550, + 258 + ], + "spans": [ + { + "bbox": [ + 432, + 220, + 550, + 258 + ], + "type": "image", + "image_path": "09ac87a5b6d11e3481aed0e7a6772231af811eb2884e0169dd48c4dd05d1c15e.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 310, + 260, + 367, + 297 + ], + "blocks": [ + { + "bbox": [ + 310, + 260, + 367, + 297 + ], + "lines": [ + { + "bbox": [ + 310, + 260, + 367, + 297 + ], + "spans": [ + { + "bbox": [ + 310, + 260, + 367, + 297 + ], + "type": "image", + "image_path": "c2b2050c21d68f9ba72c5d18e4e6f9b7ef120224f12541b0ba74a3daa4348c63.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 368, + 260, + 430, + 297 + ], + "blocks": [ + { + "bbox": [ + 368, + 260, + 430, + 297 + ], + "lines": [ + { + "bbox": [ + 368, + 260, + 430, + 297 + ], + "spans": [ + { + "bbox": [ + 368, + 260, + 430, + 297 + ], + "type": "image", + "image_path": "6ac3123f6415d2ec9ea771ea80ea05fd29678f41826e762891adde5ce141275b.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 432, + 260, + 550, + 297 + ], + "blocks": [ + { + "bbox": [ + 432, + 260, + 550, + 297 + ], + "lines": [ + { + "bbox": [ + 432, + 260, + 550, + 297 + ], + "spans": [ + { + "bbox": [ + 432, + 260, + 550, + 297 + ], + "type": "image", + "image_path": "50b29a0b6d550058c26ecee0a4cc7807b7c53a9b112d3d49fd997ca74964e8c2.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 71, + 299, + 132, + 336 + ], + "blocks": [ + { + "bbox": [ + 71, + 299, + 132, + 336 + ], + "lines": [ + { + "bbox": [ + 71, + 299, + 132, + 336 + ], + "spans": [ + { + "bbox": [ + 71, + 299, + 132, + 336 + ], + "type": "image", + "image_path": "1b6074ffcc6c1665a9ca40dba2f71b094d9c2bd940ece7ecd026f14af502750d.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "bbox": [ + 134, + 307, + 303, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 307, + 303, + 332 + ], + "spans": [ + { + "bbox": [ + 134, + 307, + 303, + 332 + ], + "type": "text", + "content": "\"A scarecrow stands in a field, swaying slightly as a gentle breeze rustles through its clothes.\"" + } + ] + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 310, + 299, + 372, + 337 + ], + "blocks": [ + { + "bbox": [ + 310, + 299, + 372, + 337 + ], + "lines": [ + { + "bbox": [ + 310, + 299, + 372, + 337 + ], + "spans": [ + { + "bbox": [ + 310, + 299, + 372, + 337 + ], + "type": "image", + "image_path": "b27a1322749fa13ca2322da0c18f5cfd0330c808ce6275281baef84a28412f23.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 374, + 306, + 547, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 306, + 547, + 333 + ], + "spans": [ + { + "bbox": [ + 374, + 306, + 547, + 333 + ], + "type": "text", + "content": "\"A street sign and a large, friendly creature stand nearby, both swaying gently in the rain.\"" + } + ] + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 71, + 338, + 132, + 374 + ], + "blocks": [ + { + "bbox": [ + 71, + 338, + 132, + 374 + ], + "lines": [ + { + "bbox": [ + 71, + 338, + 132, + 374 + ], + "spans": [ + { + "bbox": [ + 71, + 338, + 132, + 374 + ], + "type": "image", + "image_path": "ab28d824abeac796127afd59b5c3139665748de17c0749402b75076f447be00e.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 134, + 338, + 190, + 374 + ], + "blocks": [ + { + "bbox": [ + 134, + 338, + 190, + 374 + ], + "lines": [ + { + "bbox": [ + 134, + 338, + 190, + 374 + ], + "spans": [ + { + "bbox": [ + 134, + 338, + 190, + 374 + ], + "type": "image", + "image_path": "d1aa985894f2d540fcce6ed3b7a95c963278504213fb3b0c016d93e007894804.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 191, + 338, + 251, + 374 + ], + "blocks": [ + { + "bbox": [ + 191, + 338, + 251, + 374 + ], + "lines": [ + { + "bbox": [ + 191, + 338, + 251, + 374 + ], + "spans": [ + { + "bbox": [ + 191, + 338, + 251, + 374 + ], + "type": "image", + "image_path": "70ad98a5867db3df8a218e30cd96e8ed86f64fa91c47c32ee5d0509f83ca775e.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 252, + 338, + 310, + 374 + ], + "blocks": [ + { + "bbox": [ + 252, + 338, + 310, + 374 + ], + "lines": [ + { + "bbox": [ + 252, + 338, + 310, + 374 + ], + "spans": [ + { + "bbox": [ + 252, + 338, + 310, + 374 + ], + "type": "image", + "image_path": "7372c880b318355611d45da8c3256d28e8e1ff81e3135db2469b53454fc05bc8.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 310, + 338, + 372, + 374 + ], + "blocks": [ + { + "bbox": [ + 310, + 338, + 372, + 374 + ], + "lines": [ + { + "bbox": [ + 310, + 338, + 372, + 374 + ], + "spans": [ + { + "bbox": [ + 310, + 338, + 372, + 374 + ], + "type": "image", + "image_path": "89fa153e5a2c5225cce4d8c73389f941b3c6f9d7bb1b89657366c13db8a90ff5.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 373, + 338, + 550, + 374 + ], + "blocks": [ + { + "bbox": [ + 373, + 338, + 550, + 374 + ], + "lines": [ + { + "bbox": [ + 373, + 338, + 550, + 374 + ], + "spans": [ + { + "bbox": [ + 373, + 338, + 550, + 374 + ], + "type": "image", + "image_path": "02aff4db14645901808384d9bbe39b1d672b1325e01fe7572a59a99c43140060.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 71, + 376, + 132, + 414 + ], + "blocks": [ + { + "bbox": [ + 71, + 376, + 132, + 414 + ], + "lines": [ + { + "bbox": [ + 71, + 376, + 132, + 414 + ], + "spans": [ + { + "bbox": [ + 71, + 376, + 132, + 414 + ], + "type": "image", + "image_path": "c5988bfca8895b7562a3d395bd776ddf39cdfcc007248eb704cad1d4caa9337c.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 134, + 376, + 190, + 414 + ], + "blocks": [ + { + "bbox": [ + 134, + 376, + 190, + 414 + ], + "lines": [ + { + "bbox": [ + 134, + 376, + 190, + 414 + ], + "spans": [ + { + "bbox": [ + 134, + 376, + 190, + 414 + ], + "type": "image", + "image_path": "3095b5322e89707f20e3e00a41f6047b16672805318b314beaa5cc941a605567.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 191, + 376, + 251, + 414 + ], + "blocks": [ + { + "bbox": [ + 191, + 376, + 251, + 414 + ], + "lines": [ + { + "bbox": [ + 191, + 376, + 251, + 414 + ], + "spans": [ + { + "bbox": [ + 191, + 376, + 251, + 414 + ], + "type": "image", + "image_path": "fd999b135e1d16eb44dd3bd848d842855dd4417a693cfae046669936289ab792.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 252, + 376, + 310, + 414 + ], + "blocks": [ + { + "bbox": [ + 252, + 376, + 310, + 414 + ], + "lines": [ + { + "bbox": [ + 252, + 376, + 310, + 414 + ], + "spans": [ + { + "bbox": [ + 252, + 376, + 310, + 414 + ], + "type": "image", + "image_path": "20087750fe1f32460c5b0270fc5a2bc590a1e7da4af0ba609eed95670f74aa10.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 310, + 376, + 372, + 414 + ], + "blocks": [ + { + "bbox": [ + 310, + 376, + 372, + 414 + ], + "lines": [ + { + "bbox": [ + 310, + 376, + 372, + 414 + ], + "spans": [ + { + "bbox": [ + 310, + 376, + 372, + 414 + ], + "type": "image", + "image_path": "7a992addcd2388e8719edf2da4f5a4df2965548abd1151f0a8c5b5af8761977d.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 372, + 376, + 550, + 414 + ], + "blocks": [ + { + "bbox": [ + 372, + 376, + 550, + 414 + ], + "lines": [ + { + "bbox": [ + 372, + 376, + 550, + 414 + ], + "spans": [ + { + "bbox": [ + 372, + 376, + 550, + 414 + ], + "type": "image", + "image_path": "c7b0b0c9ca6566e87b5bbbe32c83883bc36550cb84cf78432f90cbf7f05f1012.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 71, + 415, + 132, + 453 + ], + "blocks": [ + { + "bbox": [ + 71, + 415, + 132, + 453 + ], + "lines": [ + { + "bbox": [ + 71, + 415, + 132, + 453 + ], + "spans": [ + { + "bbox": [ + 71, + 415, + 132, + 453 + ], + "type": "image", + "image_path": "7b54ad140c594bb679f45ca47921d96c495f69a199c999b52823558c9001d46c.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 134, + 415, + 190, + 453 + ], + "blocks": [ + { + "bbox": [ + 134, + 415, + 190, + 453 + ], + "lines": [ + { + "bbox": [ + 134, + 415, + 190, + 453 + ], + "spans": [ + { + "bbox": [ + 134, + 415, + 190, + 453 + ], + "type": "image", + "image_path": "f267bc53ee02b69f23bed4570d02ad273a8dece970a555cd29d534aa03b8d106.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 191, + 415, + 251, + 453 + ], + "blocks": [ + { + "bbox": [ + 191, + 415, + 251, + 453 + ], + "lines": [ + { + "bbox": [ + 191, + 415, + 251, + 453 + ], + "spans": [ + { + "bbox": [ + 191, + 415, + 251, + 453 + ], + "type": "image", + "image_path": "809e411d6d5f68731831df46d1afbacfb6bb62fe4afca57273b98e89fa23cfaf.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 252, + 415, + 310, + 453 + ], + "blocks": [ + { + "bbox": [ + 252, + 415, + 310, + 453 + ], + "lines": [ + { + "bbox": [ + 252, + 415, + 310, + 453 + ], + "spans": [ + { + "bbox": [ + 252, + 415, + 310, + 453 + ], + "type": "image", + "image_path": "66adc13db0aa5eafc456e877e9e6edc83b31eb403237185a5412c0c60b45a3df.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 310, + 415, + 372, + 453 + ], + "blocks": [ + { + "bbox": [ + 310, + 415, + 372, + 453 + ], + "lines": [ + { + "bbox": [ + 310, + 415, + 372, + 453 + ], + "spans": [ + { + "bbox": [ + 310, + 415, + 372, + 453 + ], + "type": "image", + "image_path": "57c8c81b1540a3209285830cfdaba4a2fd2200f22e1af1d4a77e233f061f9a62.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 372, + 415, + 550, + 453 + ], + "blocks": [ + { + "bbox": [ + 372, + 415, + 550, + 453 + ], + "lines": [ + { + "bbox": [ + 372, + 415, + 550, + 453 + ], + "spans": [ + { + "bbox": [ + 372, + 415, + 550, + 453 + ], + "type": "image", + "image_path": "262454a7bc28d875471909c09aeac4dba219e971ba9d9ec0d151e0d8b3773b55.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 71, + 455, + 132, + 493 + ], + "blocks": [ + { + "bbox": [ + 71, + 455, + 132, + 493 + ], + "lines": [ + { + "bbox": [ + 71, + 455, + 132, + 493 + ], + "spans": [ + { + "bbox": [ + 71, + 455, + 132, + 493 + ], + "type": "image", + "image_path": "5843402af0ea4a17d5e638a9695913c6f0738b4d372961f9ed88d702d90ceec9.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 134, + 455, + 190, + 493 + ], + "blocks": [ + { + "bbox": [ + 134, + 455, + 190, + 493 + ], + "lines": [ + { + "bbox": [ + 134, + 455, + 190, + 493 + ], + "spans": [ + { + "bbox": [ + 134, + 455, + 190, + 493 + ], + "type": "image", + "image_path": "1e17be56dc2ac6d4b18210dd2be4a755cd5ba0e4ae0eea794ef5795a5eb15a2d.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 191, + 455, + 251, + 493 + ], + "blocks": [ + { + "bbox": [ + 191, + 455, + 251, + 493 + ], + "lines": [ + { + "bbox": [ + 191, + 455, + 251, + 493 + ], + "spans": [ + { + "bbox": [ + 191, + 455, + 251, + 493 + ], + "type": "image", + "image_path": "3bed9f7bcb369e182c45cd8cfe6670b85cdadb5928c626c33c7e1d7df704b251.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 252, + 455, + 310, + 493 + ], + "blocks": [ + { + "bbox": [ + 252, + 455, + 310, + 493 + ], + "lines": [ + { + "bbox": [ + 252, + 455, + 310, + 493 + ], + "spans": [ + { + "bbox": [ + 252, + 455, + 310, + 493 + ], + "type": "image", + "image_path": "09277e53bf282a151fee93d4bf8af13c928d12fd84999866f72359970a6f065b.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 310, + 455, + 372, + 493 + ], + "blocks": [ + { + "bbox": [ + 310, + 455, + 372, + 493 + ], + "lines": [ + { + "bbox": [ + 310, + 455, + 372, + 493 + ], + "spans": [ + { + "bbox": [ + 310, + 455, + 372, + 493 + ], + "type": "image", + "image_path": "9e753d89bdaeadaa816e33b8de8c04c75552a8cd593094ed40823f2aebc0f3ca.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 372, + 455, + 550, + 493 + ], + "blocks": [ + { + "bbox": [ + 372, + 455, + 550, + 493 + ], + "lines": [ + { + "bbox": [ + 372, + 455, + 550, + 493 + ], + "spans": [ + { + "bbox": [ + 372, + 455, + 550, + 493 + ], + "type": "image", + "image_path": "84f9629ad708f9caf9ff666de7122d7c61dbd6354f5034ee61add850e0534ecc.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 71, + 495, + 132, + 532 + ], + "blocks": [ + { + "bbox": [ + 71, + 495, + 132, + 532 + ], + "lines": [ + { + "bbox": [ + 71, + 495, + 132, + 532 + ], + "spans": [ + { + "bbox": [ + 71, + 495, + 132, + 532 + ], + "type": "image", + "image_path": "84d61bf8e63f112c823eb89a2c4bb63a72717cc774429f12d0d9785fe8b303b2.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 543, + 553, + 576 + ], + "lines": [ + { + "bbox": [ + 55, + 543, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 553, + 576 + ], + "type": "text", + "content": "Figure 3. Qualitative Comparison. We compare our results against Cinemo [46], Drag Anything [67], DynamiCrafter [71] and Motion-I2V [55]. Text prompts for Cinemo and DynamiCrafter are generated using ChatGPT-4V, while trajectories for Drag Anything and Motion-I2V are extracted from our animated results." + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_caption" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 134, + 495, + 190, + 532 + ], + "blocks": [ + { + "bbox": [ + 134, + 495, + 190, + 532 + ], + "lines": [ + { + "bbox": [ + 134, + 495, + 190, + 532 + ], + "spans": [ + { + "bbox": [ + 134, + 495, + 190, + 532 + ], + "type": "image", + "image_path": "a61c30fb2e68d46c4d2f28ed133b0010146ecdabf110f16002b5236d3f0cce4a.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 191, + 495, + 251, + 532 + ], + "blocks": [ + { + "bbox": [ + 191, + 495, + 251, + 532 + ], + "lines": [ + { + "bbox": [ + 191, + 495, + 251, + 532 + ], + "spans": [ + { + "bbox": [ + 191, + 495, + 251, + 532 + ], + "type": "image", + "image_path": "58d3eb698fd4c6e43b15ce569c685f321c80855db8d6be23517971fecbca2ef2.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + } + ], + "index": 70 + }, + { + "type": "image", + "bbox": [ + 252, + 495, + 310, + 532 + ], + "blocks": [ + { + "bbox": [ + 252, + 495, + 310, + 532 + ], + "lines": [ + { + "bbox": [ + 252, + 495, + 310, + 532 + ], + "spans": [ + { + "bbox": [ + 252, + 495, + 310, + 532 + ], + "type": "image", + "image_path": "17035d8104ef2956c454728813ff96a921669a0cf761a70eae26494a1b4dceb1.jpg" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_body" + } + ], + "index": 71 + }, + { + "type": "image", + "bbox": [ + 310, + 495, + 372, + 532 + ], + "blocks": [ + { + "bbox": [ + 310, + 495, + 372, + 532 + ], + "lines": [ + { + "bbox": [ + 310, + 495, + 372, + 532 + ], + "spans": [ + { + "bbox": [ + 310, + 495, + 372, + 532 + ], + "type": "image", + "image_path": "917e34fdf95927c94b74507b33529d1411abed38bc33985ec7bf411ca4fa18a7.jpg" + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_body" + } + ], + "index": 72 + }, + { + "type": "image", + "bbox": [ + 372, + 495, + 550, + 532 + ], + "blocks": [ + { + "bbox": [ + 372, + 495, + 550, + 532 + ], + "lines": [ + { + "bbox": [ + 372, + 495, + 550, + 532 + ], + "spans": [ + { + "bbox": [ + 372, + 495, + 550, + 532 + ], + "type": "image", + "image_path": "7662a5d73919ed2783883e5b2e9f79ab8c44d3e37a5bdfb0e7aa7353776443cd.jpg" + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "image_body" + } + ], + "index": 73 + }, + { + "bbox": [ + 55, + 587, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 295, + 635 + ], + "type": "text", + "content": "two videos (one from our method and one from baselines). The user preference results, presented in Tab. 2, show that our proposed method consistently outperforms the baselines across all evaluation criteria." + } + ] + } + ], + "index": 75 + }, + { + "bbox": [ + 55, + 646, + 191, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 191, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 191, + 659 + ], + "type": "text", + "content": "4.2. Qualitative Comparison" + } + ] + } + ], + "index": 76 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "We present a qualitative comparison with baseline methods in Fig. 3. Cinemo often produces static results with minimal motion dynamics. Drag Anything frequently misinterprets the motion trajectory as a camera movement, resulting in" + } + ] + } + ], + "index": 77 + }, + { + "bbox": [ + 313, + 587, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 555, + 685 + ], + "type": "text", + "content": "unintended dynamic sequences. DynamiCrafter, while generating larger motions, struggles to maintain the geometry of input objects, leading to noticeable distortions. Motion-I2V has difficulty preserving the appearance of the input content, often yielding unsatisfactory results. In contrast, our physics-guided approach ensures both physically plausible motions and high-quality rendering, preserving the geometry and visual consistency of the input." + } + ] + } + ], + "index": 78 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 59, + 74, + 69, + 522 + ], + "type": "aside_text", + "angle": 270, + "lines": [ + { + "bbox": [ + 59, + 74, + 69, + 522 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 69, + 522 + ], + "type": "text", + "content": "Ours MotionI2V DC DA Cinemo Input Ours MotionI2V DC DA Cinemo Input" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10799" + } + ] + } + ], + "index": 79 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 129, + 290, + 200 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 126 + ], + "type": "text", + "content": "Table 1. Quantitative Comparisons. We report FID to evaluate the similarity between the reference image and generated frames. VSVQ, VSTC, VSDD, and VSFC represent scores for visual quality, temporal consistency, dynamic degree, and factual consistency, respectively, as measured by VideoScore [19]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 129, + 290, + 200 + ], + "lines": [ + { + "bbox": [ + 60, + 129, + 290, + 200 + ], + "spans": [ + { + "bbox": [ + 60, + 129, + 290, + 200 + ], + "type": "table", + "html": "
MethodsFID↓VSVQ↑VSTC↑VSDD↑VSFC↑
Cinemoto [46]49.52.852.802.422.58
DragAnything [67]148.92.772.452.972.52
DynamiCrafter [71]94.92.782.682.532.51
Motion-I2V [55]121.82.702.502.662.39
Ours90.42.892.862.482.64
", + "image_path": "6371ea29af5c5caf71b2c3e21ce2fd6d0138fd144f3d03a6f7b946bb9625632d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 249, + 280, + 309 + ], + "blocks": [ + { + "bbox": [ + 55, + 207, + 295, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 242 + ], + "type": "text", + "content": "Table 2. User Study. We show the user preference for our method over the baseline methods in terms of visual quality (VQ), temporal consistency (TC), motion plausibility (MP), and overall feeling." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 249, + 280, + 309 + ], + "lines": [ + { + "bbox": [ + 69, + 249, + 280, + 309 + ], + "spans": [ + { + "bbox": [ + 69, + 249, + 280, + 309 + ], + "type": "table", + "html": "
MethodsVQTCMPOverall
Cinemoto [46]86%83%82%81%
DragAnything [67]93%91%89%91%
DynamiCrafter [71]84%78%76%81%
Motion-I2V [55]95%94%97%96%
", + "image_path": "e8a4b7f5f93b3b4de528b79e3f3c1236cc0b6fb53b51c35ab2ea5351160bbf7f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 321, + 135, + 393 + ], + "blocks": [ + { + "bbox": [ + 58, + 321, + 135, + 393 + ], + "lines": [ + { + "bbox": [ + 58, + 321, + 135, + 393 + ], + "spans": [ + { + "bbox": [ + 58, + 321, + 135, + 393 + ], + "type": "image", + "image_path": "cd091db1c3ae36cc9b4a5d8b67a436049ce8326a26cd689ee96ff8ce3d4d991f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 394, + 113, + 403 + ], + "lines": [ + { + "bbox": [ + 79, + 394, + 113, + 403 + ], + "spans": [ + { + "bbox": [ + 79, + 394, + 113, + 403 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 136, + 321, + 214, + 371 + ], + "blocks": [ + { + "bbox": [ + 136, + 321, + 214, + 371 + ], + "lines": [ + { + "bbox": [ + 136, + 321, + 214, + 371 + ], + "spans": [ + { + "bbox": [ + 136, + 321, + 214, + 371 + ], + "type": "image", + "image_path": "b5ac7921e75f843573eed7141de0254603d0b0638db663b16a7ab4d5180fc883.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 371, + 184, + 380 + ], + "lines": [ + { + "bbox": [ + 165, + 371, + 184, + 380 + ], + "spans": [ + { + "bbox": [ + 165, + 371, + 184, + 380 + ], + "type": "text", + "content": "Warp" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 217, + 321, + 294, + 371 + ], + "blocks": [ + { + "bbox": [ + 217, + 321, + 294, + 371 + ], + "lines": [ + { + "bbox": [ + 217, + 321, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 217, + 321, + 294, + 371 + ], + "type": "image", + "image_path": "5bf0de7aaae797c51a605607d18d9754bc5ca467cb9a2f604f3056b1d7789fa9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 371, + 289, + 380 + ], + "lines": [ + { + "bbox": [ + 222, + 371, + 289, + 380 + ], + "spans": [ + { + "bbox": [ + 222, + 371, + 289, + 380 + ], + "type": "text", + "content": "Warp + Hard Inpaint" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 136, + 380, + 214, + 430 + ], + "blocks": [ + { + "bbox": [ + 136, + 380, + 214, + 430 + ], + "lines": [ + { + "bbox": [ + 136, + 380, + 214, + 430 + ], + "spans": [ + { + "bbox": [ + 136, + 380, + 214, + 430 + ], + "type": "image", + "image_path": "09726d7a8718fa4ffa5f5190d08cdab74f9749af992288ba1daa8e2f0b7d1039.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 430, + 208, + 438 + ], + "lines": [ + { + "bbox": [ + 144, + 430, + 208, + 438 + ], + "spans": [ + { + "bbox": [ + 144, + 430, + 208, + 438 + ], + "type": "text", + "content": "Warp + Soft Inpaint" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 448, + 295, + 504 + ], + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 504 + ], + "type": "text", + "content": "Figure 4. Sketch-Guided Rendering. Applying warping and inpainting introduces artifacts due to segmentation inaccuracy. Soft-inpainting [39] reduces these issues but can alter the content. Our sketch-guided rendering method produces high-quality results while preserving image details." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 217, + 380, + 294, + 430 + ], + "blocks": [ + { + "bbox": [ + 217, + 380, + 294, + 430 + ], + "lines": [ + { + "bbox": [ + 217, + 380, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 217, + 380, + 294, + 430 + ], + "type": "image", + "image_path": "85966a5bd34733c68c565dcf33d35234e86229de7e110abfa21ce73b4c792d6d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 430, + 264, + 437 + ], + "lines": [ + { + "bbox": [ + 247, + 430, + 264, + 437 + ], + "spans": [ + { + "bbox": [ + 247, + 430, + 264, + 437 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 526, + 220, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 220, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 220, + 539 + ], + "type": "text", + "content": "4.3. Additional Qualitative Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 545, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 715 + ], + "type": "text", + "content": "Sketch-Guided Rendering After physics-based animation in our proposed framework, we obtain a sequence of optical flow " + }, + { + "bbox": [ + 54, + 545, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_{0\\to 1},\\mathcal{F}_{0\\to 2},\\dots ,\\mathcal{F}_{0\\to T}\\}" + }, + { + "bbox": [ + 54, + 545, + 295, + 715 + ], + "type": "text", + "content": ". A straightforward approach would be to warp the input image " + }, + { + "bbox": [ + 54, + 545, + 295, + 715 + ], + "type": "inline_equation", + "content": "I_{0}" + }, + { + "bbox": [ + 54, + 545, + 295, + 715 + ], + "type": "text", + "content": " using these optical flows and apply an inpainting algorithm to fill any occluded areas. However, this often results in suboptimal outputs. The main issue lies in segmentation inaccuracies, which can cause unintended regions to be warped or leave parts of the intended objects static, as shown in Fig. 4. While introducing a soft mask [39] into the inpainting process can help smooth the transition between the newly generated region and the original figure, it may also alter the content, making it diverge from the appearance of the reference image. Additionally, applying inpainting frame by" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 394, + 127 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 394, + 127 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 394, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 394, + 127 + ], + "type": "image", + "image_path": "649990fdea61fcdb48bf9a9aedfc6332d7da33fe520d8e1f192348d31d392642.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 129, + 379, + 137 + ], + "lines": [ + { + "bbox": [ + 329, + 129, + 379, + 137 + ], + "spans": [ + { + "bbox": [ + 329, + 129, + 379, + 137 + ], + "type": "text", + "content": "Static Reference" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 399, + 72, + 475, + 128 + ], + "blocks": [ + { + "bbox": [ + 399, + 72, + 475, + 128 + ], + "lines": [ + { + "bbox": [ + 399, + 72, + 475, + 128 + ], + "spans": [ + { + "bbox": [ + 399, + 72, + 475, + 128 + ], + "type": "image", + "image_path": "84c44ba90a0a06041d57536943a5d89267f44585d43ea7b75caf193469a2aa55.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 129, + 474, + 137 + ], + "lines": [ + { + "bbox": [ + 400, + 129, + 474, + 137 + ], + "spans": [ + { + "bbox": [ + 400, + 129, + 474, + 137 + ], + "type": "text", + "content": "w/o Enhanced Dynamics" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 146, + 555, + 212 + ], + "lines": [ + { + "bbox": [ + 313, + 146, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 555, + 212 + ], + "type": "text", + "content": "Figure 5. Complementary Dynamics Enhancement. While physics-based animation maintains geometric consistency, it may lack the fluidity and exaggeration commonly seen in anime. We employ a data-driven interpolation module to enhance the motion dynamics, creating more natural-looking animations that better resemble real anime." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 477, + 72, + 553, + 128 + ], + "blocks": [ + { + "bbox": [ + 477, + 72, + 553, + 128 + ], + "lines": [ + { + "bbox": [ + 477, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 477, + 72, + 553, + 128 + ], + "type": "image", + "image_path": "1ddccb858cb8dbf3c634553f15a094887f19fa1d4c14d986310ede26f6feb604.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 129, + 550, + 137 + ], + "lines": [ + { + "bbox": [ + 480, + 129, + 550, + 137 + ], + "spans": [ + { + "bbox": [ + 480, + 129, + 550, + 137 + ], + "type": "text", + "content": "w/ Enhanced Dynamics" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 224, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 344 + ], + "type": "text", + "content": "frame may also cause temporal consistency issues. To address this, our method employs a sketch-guided rendering module. We first extract a sparse geometric representation of the image and apply the animation dynamics directly to this sketch-based structure. This sparse representation is more robust to segmentation inaccuracies and helps preserve the intended motion. Our subsequent rendering module then synthesizes high-quality video sequences from the animated sketches, maintaining both temporal consistency and visual fidelity." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 556, + 515 + ], + "type": "text", + "content": "Complementary Dynamics Enhancement While our physics-based animation ensures physically accurate and geometry-consistent motion, it is inherently limited to 2D representations and cannot fully capture 3D effects. Moreover, the exaggerated dynamics often seen in anime do not always conform to strict physical laws. To address these limitations, we incorporate a data-driven cartoon interpolation module [70] to introduce complementary dynamics. As illustrated in Fig. 5, this interpolation module enables fluid contour deformations of clothing during motion, rather than rigidly adhering to static geometry. It also dynamically generates new sketch lines, enhancing the expressiveness of the animation and bringing it closer to the aesthetic of traditional hand-drawn anime." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 546, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 546, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 546, + 555, + 714 + ], + "type": "text", + "content": "We presented PhysAnimator, a novel framework for generating dynamic and stylized animations from static anime illustrations by integrating physics-based simulations with data-driven generative models. Our approach generates physically plausible, fluid and exaggerated animestyled motion through deformable body simulations, providing controllability to users through user-guided energy strokes. A sketch-guided video diffusion model ensures high-quality, temporally consistent frames, while a data-driven anime frame interpolation model adds expressive, non-physical dynamics. The experiments show that our proposed method outperforms existing video diffusion methods, offering a powerful tool for creating visually compelling and user-controllable anime-style animations." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10800" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 195, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 195, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 195, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 295, + 228 + ], + "type": "text", + "content": "[1] Jernej Barbic, Marco da Silva, and Jovan Popovic. Deformable object animation using reduced optimal control. In ACM SIGGRAPH 2009 papers, pages 1-9. 2009. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 230, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 230, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 230, + 296, + 283 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv:2311.15127, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 286, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 295, + 350 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "type": "text", + "content": "[4] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 430 + ], + "type": "text", + "content": "[5] Caroline Chan, Fredo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 433, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 294, + 475 + ], + "type": "text", + "content": "[6] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv:2305.13840, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 478, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 294, + 542 + ], + "type": "text", + "content": "[7] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 294, + 577 + ], + "type": "text", + "content": "[8] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv:2403.06738, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 579, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 294, + 622 + ], + "type": "text", + "content": "[9] Stelian Coros, Sebastian Martin, Bernhard Thomaszewski, Christian Schumacher, Robert Sumner, and Markus Gross. Deformable objects alive! ACM Transactions on Graphics (TOG), 31(4):1-9, 2012. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "text", + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[11] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "We thank anonymous reviewers for their insightful comments. We acknowledge support from NSF 2153851. We would like to thank Rahul Garg, Hossein Taghavi, Roshni Cooper, Ritwik Kumar, Boris Chen, Amli Murphy, Oliver Banasiak and Taiki Sakurai for their valuable suggestions and support." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[12] Rinon Gal, Yael Vinker, Yuval Alaluf, Amit Bermano, Daniel Cohen-Or, Ariel Shamir, and Gal Chechik. Breathing life into sketches using text-to-video priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4325–4336, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "type": "text", + "content": "[13] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22930-22941, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "text", + "content": "[14] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arxiv:2307.10373, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "type": "text", + "content": "[15] Joseph Gilland. Elemental Magic, Volume 2: The Technique of Special Effects Animation. Routledge, 2012. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 330 + ], + "type": "text", + "content": "[16] Yuchao Gu, Yipin Zhou, Bichen Wu, Licheng Yu, Jia-Wei Liu, Rui Zhao, Jay Zhangjie Wu, David Junhao Zhang, Mike Zheng Shou, and Kevin Tang. Videoswap: Customized video subject swapping with interactive semantic point correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7621-7630, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "text", + "content": "[17] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "type": "text", + "content": "[18] Junlin Han, Filippos Kokkinos, and Philip Torr. Vfusion3d: Learning scalable 3d generative models from video diffusion models. In European Conference on Computer Vision, pages 333-350. Springer, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "text", + "content": "[19] Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, et al. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. arXiv:2406.15252, 2024. 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[20] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv:2211.13221, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 513, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 566 + ], + "type": "text", + "content": "[21] Yingqing He, Menghan Xia, Haoxin Chen, Xiaodong Cun, Yuan Gong, Jinbo Xing, Yong Zhang, Xintao Wang, Chao Weng, Ying Shan, et al. animate-a-story: Storytelling with retrieval-augmented video generation. arXiv:2307.06940, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 568, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 621 + ], + "type": "text", + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[23] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[24] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv:2311.04400, 2023. 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10801" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[25] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG), 38(6): 201, 2019. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 160 + ], + "type": "text", + "content": "[26] Zhitong Huang, Mohan Zhang, and Jing Liao. Lycd: Reference-based linear video colorization with diffusion models. arXiv:2409.12960, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 215 + ], + "type": "text", + "content": "[27] Siddhant Jain, Daniel Watson, Eric Tabellion, Ben Poole, Janne Kontkanen, et al. Video interpolation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7341-7351, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "text", + "content": "[28] Hyeonho Jeong, Geon Yeong Park, and Jong Chul Ye. Vmc: Video motion customization using temporal attention adaptation for text-to-video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9212-9221, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 271, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 294, + 315 + ], + "type": "text", + "content": "[29] Ben Jones, Jovan Popovic, James McCann, Wilmot Li, and Adam Bargteil. Dynamic sprites: artistic authoring of interactive animations. Computer Animation and Virtual Worlds, 26(2):97-108, 2015. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "type": "text", + "content": "[30] Ben Jones, Nils Thuerey, Tamar Shinar, and Adam W Bargteil. Example-based plastic deformation of rigid bodies. ACM Transactions on Graphics (TOG), 35(4):1-11, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "text", + "content": "[31] Rubaiat Habib Kazi, Tovi Grossman, Nobuyuki Umetani, and George Fitzmaurice. Motion amplifiers: sketching dynamic illustrations using the principles of 2d animation. In Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, pages 4599-4609, 2016. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 425 + ], + "type": "text", + "content": "[32] Diederik P Kingma. Auto-encoding variational bayes. arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "type": "text", + "content": "[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "text", + "content": "[34] Nicholas Kolkin, Jason Salavon, and Gregory Shakhnarovich. Style transfer by relaxed optimal transport and self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10051-10060, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "type": "text", + "content": "[35] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv:2403.14468, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "text", + "content": "[36] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "text", + "content": "[37] Der-Tsai Lee and Bruce J Schachter. Two algorithms for constructing a delaunay triangulation. International Journal of Computer & Information Sciences, 9(3):219-242, 1980. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "text", + "content": "[38] Minhyeok Lee, Suhwan Cho, Chajin Shin, Jungho Lee, Sunghun Yang, and Sangyoun Lee. Video diffusion models are strong video inpainter. arXiv:2408.11402, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] Eran Levin and Ohad Fried. Differential diffusion: Giving each pixel its strength. arXiv:2306.00950, 2023. 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "type": "text", + "content": "[40] Minchen Li, Chenfanfu Jiang, and Zhaofeng Luo. Physics-Based Simulation. 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[41] Mingxiao Li, Bo Wan, Marie-Francine Moens, and Tinne Tuytelaars. animate your motion: Turning still images into dynamic videos. arXiv:2403.10179, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 132, + 555, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 132, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 132, + 555, + 175 + ], + "type": "text", + "content": "[42] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 177, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 177, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 553, + 220 + ], + "type": "text", + "content": "[43] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv:2408.16767, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 223, + 553, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 267 + ], + "type": "text", + "content": "[44] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imaged-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 269, + 555, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 555, + 313 + ], + "type": "text", + "content": "[45] Jiaju Ma, Li-Yi Wei, and Rubaiat Habib Kazi. A layered authoring tool for stylized 3d animations. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1-14, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "text", + "content": "[46] Xin Ma, Yaohui Wang, Gengyu Jia, Xinyuan Chen, YuanFang Li, Cunjian Chen, and Yu Qiao. Cinemo: Consistent and controllable image animation with motion diffusion models. arXiv:2407.15642, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "text", + "content": "[47] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv:2410.05363, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 418, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 553, + 460 + ], + "type": "text", + "content": "[48] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5437-5446, 2020. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 464, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 517 + ], + "type": "text", + "content": "[49] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adaptions in frozen image-to-video diffusion model. arXiv:2405.20222, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 520, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 552 + ], + "type": "text", + "content": "[50] Zhenglin Pan, Yu Zhu, and Yuxuan Mu. Sakuga-42m dataset: Scaling up cartoon research. arXiv:2405.07425, 2024.6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "text", + "content": "[51] Mengqi Peng, Li-yi Wei, Rubaiat Habib Kazi, and Vladimir G Kim. Autocomplete animated sculpting. In Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology, pages 760-777, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 601, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 655 + ], + "type": "text", + "content": "[52] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv:2408.00714, 2024. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "text", + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10802" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[54] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv:2406.01493, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "type": "text", + "content": "[55] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[56] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 261 + ], + "type": "text", + "content": "[57] Alexey Stomakhin, Russell Howes, Craig A Schroeder, and Joseph M Teran. Energetically consistent invertible elasticity. In Symposium on Computer Animation, 2012. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 314 + ], + "type": "text", + "content": "[58] Qingkun Su, Xue Bai, Hongbo Fu, Chiew-Lan Tai, and Jue Wang. Live sketch: Video-driven dynamic deformation of static drawings. In Proceedings of the 2018 chi conference on human factors in computing systems, pages 1-12, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "type": "text", + "content": "[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 437 + ], + "type": "text", + "content": "[60] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "type": "text", + "content": "[61] Qian Wang, Abdelrahman Eldosokey, Mohit Mendiratta, Fangneng Zhan, Adam Kortylewski, Christian Theobalt, and Peter Wonka. Zero-shot video semantic segmentation based on pre-trained diffusion models. arXiv:2405.16947, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "text", + "content": "[62] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 602 + ], + "type": "text", + "content": "[63] Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6537-6549, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "text", + "content": "[64] Nora S Willett, Wilmot Li, Jovan Popovic, Floraine Berthouzoz, and Adam Finkelstein. Secondary motion for performed 2d animation. In Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology, pages 97-108, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[65] Nora S Willett, Rubaiat Habib Kazi, Michael Chen, George Fitzmaurice, Adam Finkelstein, and Tovi Grossman. A mixed-initiative interface for animating static pictures. In Proceedings of the 31st annual ACM symposium on user interface software and technology, pages 649-661, 2018. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[66] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. Aniclipart: Clipart animation with text-to-video priors. arXiv:2404.12347, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 172 + ], + "type": "text", + "content": "[67] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2025. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 174, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 207 + ], + "type": "text", + "content": "[68] Jun Xing, Li-Yi Wei, Takaaki Shiratori, and Koji Yatani. Autocomplete hand-drawn animations. ACM Transactions on Graphics (TOG), 34(6):1-11, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[69] Jun Xing, Rubaiat Habib Kazi, Tovi Grossman, Li-Yi Wei, Jos Stam, and George Fitzmaurice. Energy-brushes: Interactive tools for illustrating stylized elemental dynamics. In Proceedings of the 29th Annual Symposium on User Interface Software and Technology, pages 755-766, 2016. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[70] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. arXiv:2405.17933, 2024. 2, 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 309, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 374 + ], + "type": "text", + "content": "[71] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In European Conference on Computer Vision, pages 399-417. Springer, 2025. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[72] Zhen Xing, Qijun Feng, Haoran Chen, Qi Dai, Han Hu, Hang Xu, Zuxuan Wu, and Yu-Gang Jiang. A survey on video diffusion models. ACM Computing Surveys, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "text", + "content": "[73] Shiyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "text", + "content": "[74] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv:2308.08089, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 511, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 543 + ], + "type": "text", + "content": "[75] Jiayi Eris Zhang, Seungbae Bang, David IW Levin, and Alec Jacobson. Complementary dynamics. arXiv:2009.02462, 2020.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "text", + "content": "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "text", + "content": "[77] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance. arXiv:2406.19680, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "text", + "content": "[78] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv:2407.21705, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[79] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. Trip: Temporal residual learning with image noise prior for image-to-video diffusion models. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10803" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 140 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 8671-8681, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 140 + ], + "type": "text", + "content": "[80] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv:2310.08465, 2023. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10804" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/1bf0def3-a278-47bf-9ffd-13a93c05e545_content_list.json b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/1bf0def3-a278-47bf-9ffd-13a93c05e545_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..877a8d658300bb383b06c329b6027247357da03c --- /dev/null +++ b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/1bf0def3-a278-47bf-9ffd-13a93c05e545_content_list.json @@ -0,0 +1,1407 @@ +[ + { + "type": "text", + "text": "PhysGen3D: Crafting a Miniature Interactive World from a Single Image", + "text_level": 1, + "bbox": [ + 114, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boyuan Chen $^{1}$ , Hanxiao Jiang $^{2,3}$ , Shaowei Liu $^{2}$ , Saurabh Gupta $^{2}$ , Yunzhu Li $^{3}$ , Hao Zhao $^{1}$ , Shenlong Wang $^{2}$", + "bbox": [ + 250, + 179, + 715, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University, $^{2}$ University of Illinois Urbana-Champaign, $^{3}$ Columbia University", + "bbox": [ + 222, + 217, + 743, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49f193c9257b3209f2c5f4245cae84a3c27bd0eaedebae202fd6ad7fe17b923a.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 80, + 277, + 215, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a342b395edbf963f5c4f60ab92bbb3d57b4f2344e9375768511d415d71dde92c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 383, + 215, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e9c23e75751deed27b69f4b1d0178d42f89e4779a6c24d56f9424f9d3243c3a.jpg", + "image_caption": [ + "Figure 1. PhysGen3D generates realistic, physically plausible motion from a single image and a text prompt by reasoning about geometry, semantics, and material properties. (a) An apple rolls under the influence of its initial velocity, friction, and shape, producing a natural progression over time. (b) Three animal figures interact dynamically, colliding after being propelled upwards and forwards. (c) A toy potato bounces back with soft-body dynamics in response to an initial downward force, capturing material-specific behaviors. PhysGen3D lets users quickly explore physics-driven object interactions and behaviors in a compact virtual scene generated from a single input image." + ], + "image_footnote": [], + "bbox": [ + 106, + 491, + 189, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4400e282a98a46655b270e898c9a7a4cd37d9152d740d431b7c0ae06a7bbac86.jpg", + "image_caption": [ + "Generation (left $\\rightarrow$ right: time steps)" + ], + "image_footnote": [], + "bbox": [ + 220, + 277, + 890, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd7fbdbb261151180d950def447ba3044456996c965db97a12b40b3a01795464.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 383, + 890, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef2c847f1dd27ef933fad6291247779d72fe434ea5e5eac39611947b87e2228b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 491, + 890, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 693, + 313, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Envisioning physically plausible outcomes from a single image requires a deep understanding of the world's dynamics. To address this, we introduce PhysGen3D, a novel framework that transforms a single image into an amodal, camera-centric, interactive 3D scene. By combining advanced image-based geometric and semantic understanding with physics-based simulation, PhysGen3D creates an interactive 3D world from a static image, enabling us to \"imagine\" and simulate future scenarios based on user input. At its core, PhysGen3D estimates 3D shapes, poses, physical and lighting properties of objects, thereby", + "bbox": [ + 75, + 736, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "capturing essential physical attributes that drive realistic object interactions. This framework allows users to specify precise initial conditions, such as object speed or material properties, for enhanced control over generated video outcomes. We evaluate PhysGen3D's performance against closed-source state-of-the-art (SOTA) image-to-video models, including Pika, Kling, and Gen-3, showing PhysGen3D's capacity to generate videos with realistic physics while offering greater flexibility and fine-grained control. Our results show that PhysGen3D achieves a unique balance of photorealism, physical plausibility, and user-driven interactivity, opening new possibilities for generating dynamic, physics-grounded video from an image. Project page: https://by-luckk.github.io/PhysGen3D.", + "bbox": [ + 496, + 694, + 895, + 905 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6178", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Photographs capture snapshots of our physical world, preserving specific moments in time but leaving out the alternative outcomes that could have unfolded. For instance, looking at a photo, we might wonder, \"What if I poke the apples to make them roll across the ground?\" or \"What if I squeeze the three stuffed animals closer together?\" or \"What if I drop my cute potato toys onto the floor?\" Humans intuitively understand how these scenarios would play out because we have an innate sense of the physical world beyond what we see in a single image. We develop a computational model that can answer such \"what-if\" questions by generating video outcomes from a single static image.", + "bbox": [ + 76, + 114, + 468, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A promising approach toward this goal is data-driven image-to-video (I2V) generation [17, 18, 29, 61, 68, 74, 84]. I2V leverages diffusion-based generative models trained on vast datasets of internet images and videos, enabling the production of photorealistic videos with remarkable detail. However, I2V still has limitations in precise control and lacks physical grounding. As a result, users cannot interact freely and accurately to achieve specific physical effects, nor can I2V guarantee physical realism.", + "bbox": [ + 76, + 296, + 470, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, recent research has focused on modeling the physical world from visual inputs to create digital twins, allowing for precise interactions [42, 81, 92, 96]. These approaches can generate virtual scenes with convincing physical interactions, but they typically require complete 3D scans from multi-view images or depth sensors, making them data-intensive. While some methods [43, 48, 66] enable interaction with a single image, they are often limited by physical constraints (e.g., rigid bodies or springs), specific object types (e.g., waterfalls), or a 2D scope. This gap highlights the need for a generic, controllable, physically grounded, and photorealistic approach to generate video from a single image while maintaining physical realism.", + "bbox": [ + 76, + 431, + 470, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce PhysGen3D, a novel framework that transforms a single image into an amodal, camera-centric, interactive 3D scene, enabling realistic simulation and rendering. Our approach combines the strengths of image-based geometric and semantic understanding [40, 58, 67, 71, 83] with physics-based simulation [31-33]. At its core, PhysGen3D is a digital twinning method that estimates an object's 3D shape, pose, physical and lighting properties, infers background geometry and appearance, deduces physical characteristics, and performs dimensional analysis—all from a single input image. This task is conventionally challenging due to its inherently ill-posed nature. To tackle this, we leverage various pretrained vision models, integrating their outputs to create an image-centric digital twin.", + "bbox": [ + 76, + 628, + 468, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For physical simulation, we employ material point methods [30, 37], a robust point-voxel-based framework that models counterfactual physical behaviors of objects in the image. Through precise inference of physical properties,", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "simulations in the PhysGen3D environment achieve a high degree of realism and stability. We further enhance realism by applying physics-based rendering, seamlessly integrating dynamic effects back into the original image. PhysGen3D produces results that are not only visually realistic in terms of dynamics and lighting but also highly controllable, allowing users to specify initial conditions like speed and material properties. Due to the use of large pretrained models, our pipeline operates effectively without task-specific training.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments, based on a carefully designed and rigorous user study, demonstrate that, compared to closed-source state-of-the-art video AIGC models such as Pika, Kling, and Gen-3, our framework provides significantly more flexible control over object motions, generates videos that better align with user intentions, and achieves superior physical realism—all while maintaining comparable rendering quality.", + "bbox": [ + 496, + 229, + 893, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 358, + 635, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-view 3D reconstruction Great progress has been made in object-centric single view 3D reconstructions [45, 46, 62, 83]. However, applying these techniques to single-view 3D scene reconstruction becomes more challenging. Most existing works only focus on one part of 3D scene understanding. Geometric approaches reconstruct the 3D scene holistically but neglect individual object understanding, or they focus solely on foreground objects without considering the interaction of objects with the complex environment. Additionally, works have been done for scene relighting [49, 93, 95] and segmentation [40, 47, 58, 85, 97], but none of these provide a full 3D understanding within the image. In our work, we also perform perception reasoning for object materials, backgrounds, rendering properties, and physical stability.", + "bbox": [ + 496, + 387, + 893, + 614 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Controllable video generation. Video generation has made significant progress in recent years [5, 9, 10, 15, 21, 23, 25, 28, 41, 54, 63, 69, 73, 82, 90, 91]. The state-of-the-art framework [1, 3] can generate photorealistic and coherent videos from text instructions using diffusion models [27, 55, 59, 64, 65]. Controllable video generation is achieved through leveraging pre-trained video generation models with conditioning information, including but not limited to depth maps [13, 86], linear translation [39], layouts [44, 70, 87], and multiple combinations [50]. However, most existing generation methods implicitly generate image space dynamics, which might lead to unrealistic, hallucinated motions. In contrast, our method explicitly controls the motion and interaction with simulation, allowing us to create more sophisticated effects without the need for extensive training data. Our approach is training-free and generalizable to all objects in the world.", + "bbox": [ + 496, + 643, + 895, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6179", + "bbox": [ + 480, + 946, + 514, + 958 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fd7253eba88cec6ff23009b105abf5f7acd3dfa849c9a4e32bc4faa5e1ba38c8.jpg", + "image_caption": [ + "Figure 2. Method Overview. PhysGen3D's framework consists of three modules: a) 3D world creation, which infers geometry, semantics, rendering and physical parameters from the input image; b) dynamics simulation using Taichi-Elements for particle-based physics; and c) physics-based rendering with a two-pass shadow mapping technique." + ], + "image_footnote": [], + "bbox": [ + 83, + 77, + 890, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable image animation Image-based animation aims to animate objects that appear in images. Numerous works [17, 18, 29, 61, 68, 74, 84] have focused on this task. To improve quality, recent research has adopted data-driven solutions, training temporal neural networks to directly predict subsequent video frames [6, 14, 19, 24, 29, 72], or incorporating physical heuristics [17, 36]. Recently, there has been increased focus on interactive [7, 8, 43] and controllable [4, 18] image-to-video synthesis. Additional priors such as motion fields [19, 29, 51, 52, 66, 84], optical flow [11, 22, 94], 3D geometry information [57, 76], and user annotations [43] have also been introduced. The closest work to ours is Liu et al. [48], which focuses on image understanding and uses an existing 2D rigid simulator to generate 2D animations. However, their approach is limited to 2D dynamics, does not account for real-world 3D physics, and is restricted to rigid bodies. In contrast, our method provides more realistic and flexible control over 3D motions and object materials and extends beyond rigid-body physics.", + "bbox": [ + 75, + 429, + 473, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 737, + 187, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to reconstruct an interactive, camera-centric miniature world from a single input image. We aim to control object materials, dynamics, and motions to simulate diverse, photorealistic, and physically plausible videos from user input. The key challenge lies in partial observations and the ill-posed nature of physical reasoning without observed dynamics. To address this, we propose a holistic reconstruction method leveraging large pretrained visual models to jointly infer geometry, dynamic materials, lighting, and", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PBR materials from a single image (Sec. 3.1). The reconstructed scene is input into a material-point-method (MPM) simulator to generate realistic physics phenomena (Sec. 3.2). Finally, we render dynamic object behaviors based on the simulation and reintegrate them into the scene, producing realistic videos with accurate appearance and motion (Sec. 3.3). Fig. 2 depicts our framework.", + "bbox": [ + 496, + 429, + 895, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Interactive 3D World from a Single Image", + "text_level": 1, + "bbox": [ + 498, + 544, + 859, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A full amodal reconstruction of the 3D world depicted in the image is critical for the next step in simulation. An ideal reconstruction should capture a comprehensive understanding of objects' relationships, geometry, appearance, material, and physical properties. However, obtaining this understanding from a single image is highly ill-posed. Our key idea is to leverage priors from pretrained vision foundation models to help infer these properties, as shown in Fig. 2.", + "bbox": [ + 496, + 568, + 893, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Segmentation. We leverage vision foundation models to recognize object categories and segment object instances. Specifically, we use GPT-4o [89] to identify the foreground object categories and use Grounded-SAM [40, 47, 58] to further detect and segment each individual instance $\\{o^i\\in \\mathbb{R}^{W\\times H\\times 3}\\}_{i}^{N}$ , where $o^i$ is the image of i-th object.", + "bbox": [ + 496, + 689, + 893, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mesh Generation. Unlike previous work [48], which used only 2D rigid-body physics, we extend to general-purpose 3D simulation. This requires a complete 3D representation of foreground objects. We adopt InstantMesh [83], which uses Zero123++ [46, 62] to synthesize multi-view images from the segmented object image $o^i$ and uses these images to reconstruct the 3D mesh $\\mathcal{O}$ of the objects. For multiple object occlusions, we adopt the iterative inpainting", + "bbox": [ + 496, + 780, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6180", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "strategy to extract the 3D mesh of the objects (see supp for more details).", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background Handling. The background serves important roles in both dynamic simulation and rendering. In simulation, background geometry acts as a support and collider, where accuracy ensures realistic object-scene interactions. In rendering, it serves as a static backdrop while the foreground objects move and helps simulate realistic global illumination effects like cast shadows. For simulation purposes, we use Dust3r [71] to estimate iamage depth. The output depth map $z \\in \\mathbb{R}^{W \\times H}$ is unprojected to the 3D world as a 3D point cloud $\\mathcal{P}$ , and Bilateral Normal Integration [12] is applied to generate a smooth surface $S$ serving as the collider. For rendering, to fill background regions occluded by moving objects, we use the LaMA inpainting model[67]. This model generates a complete background after masking out all objects and their shadows.", + "bbox": [ + 75, + 122, + 470, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Pose and Scale Estimation. The generative 3D reconstruction step provides a complete mesh with a normalized scale in an object-centric coordinate system, but it does not infer the object's location and scale in the camera coordinate. To ensure coherence with the input image, it is necessary to accurately place the 3D meshes $\\mathcal{O}$ into the 3D scene $\\mathcal{P}$ with the correct scale and 6DoF pose. This registration task is challenging, as the generated object mesh may not perfectly match the real-world object in the image.", + "bbox": [ + 75, + 349, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this challenge, we designed a multi-stage coarse-to-fine alignment strategy. In the coarse stage, we perform 2D-3D feature point matching. Firstly, we render multiple images from viewpoints uniformly distributed over a unit sphere surrounding the object. For each rendered image, we match its feature points with the original object image $o^i$ using SuperGlue [60], and the viewpoint with the most matches is selected. Matched pairs $(p_i^N \\in \\mathbb{R}^2, p_i^N \\in \\mathbb{R}^2)$ are then projected back into 3D points $(P_i^N \\in \\mathbb{R}^3, P_i^N \\in \\mathbb{R}^3)$ in object and camera coordinate, respectively. Perspective-n-Point (PnP) algorithm [20] is applied between $P_i^N$ and $p_i^N$ in image $o^i$ to estimate the 6DoF pose with scale ambiguity. Then we adjust the scale and translation factor simultaneously to minimize the L2 loss $\\sum_{i}^{N} = ||P_i^N - P_i'^N||_2$ , without altering its projection.", + "bbox": [ + 75, + 486, + 470, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the fine alignment stage, we render the mask and depth on the image plane using the current estimation through a differentiable renderer and jointly minimize two losses: $\\mathcal{L} = \\mathcal{L}_{\\mathrm{dice}} + \\mathcal{L}_{\\mathrm{depth}}$ . Here, $\\mathcal{L}_{\\mathrm{dice}} = 1 - \\frac{2\\times|M_A\\cap M_B|}{|M_A| + |M_B|}$ measures the discrepancy between the rendered mask $M_{A}$ and the observed mask $M_B$ from Grounded-SAM. The depth consistency loss is $\\mathcal{L}_{\\mathrm{depth}} = \\frac{\\mathrm{mse}(M_B*z_A - M_B*z_B)}{|M_B|}$ , where $z_{A}$ and $z_{B}$ are the rendered and Dust3r-predicted depths. This joint optimization ensures consistency between the estimated mesh in the camera coordinate and the point cloud from Dust3r, while maximizing the alignment between the 3D pose and observed object mask, ensuring accurate simu", + "bbox": [ + 76, + 715, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lation and rendering alignment.", + "bbox": [ + 498, + 90, + 709, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Appearance Optimization. The texture of the generated 3D meshes may differ from the input image. To enhance rendering quality, we use the inverse rendering pipeline in Mitsuba3 [35] to estimate material properties. Lighting parameters are estimated with DiffusionLight [56], while object PBR materials (albedo, roughness, and metallic) are optimized via differentiable rendering. To handle unknown back views and reduce complexity, we assume uniform roughness and metallic values per object and apply tone mapping for albedo optimization: $y(x) = ax^3 + bx^2 + cx + d$ , where $y(0) = 0$ , $y(1) = 1$ . With reconstructed lighting and ground surface, the optimized materials improve asset appearance, capturing realistic object-surface interactions during rendering.", + "bbox": [ + 496, + 107, + 893, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Physics Reasoning. Accurately simulating real-world dynamics requires estimating physical parameters. We focus on two aspects: 1) Following [79], we use GPT-4o to query each object's elasticity and density, and the friction coefficient for the surface $S$ . 2) We ensure reconstructed 3D assets match real-world proportions, as depth inaccuracies from Dust3r can cause unrealistic behaviors. To address this, we estimate a scale factor $k$ by comparing asset size with typical real-world sizes from GPT-4o and use $k$ for dimensionless scaling of gravity and velocity-related parameters.", + "bbox": [ + 496, + 319, + 893, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dynamics Simulation", + "text_level": 1, + "bbox": [ + 500, + 481, + 700, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the 3D assets with reasoned physical properties and the scale factor, we use the physics engine Taichi Elements [31-33], based on the Material Point Method (MPM) [37], as our simulator to support a variety of different materials, including but not limited to rigid, soft, and granular.", + "bbox": [ + 496, + 505, + 893, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Particle representation. The simulator is based on a particle-based representation. To convert our 3D assets into a simulatable particle-based representation, we apply floating points removing, internal filling and voxel downsampling. We use downsampling to handle uneven particle distributions, where downsample rate is adjusted according to the grid size of the simulator. For the convenience of rendering, we prioritize surface points.", + "bbox": [ + 496, + 580, + 893, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Physical parameters. To enhance stability, we apply the scale factor $k$ to physical parameters in simulator instead of scaling the assets to real size. For example, the gravitational acceleration is multiplied by $k$ , making to $k * 9.8$ (nondimensionalized here). With this tuning, the motion of falling or collapse remains realistic for all scales.", + "bbox": [ + 496, + 703, + 893, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "External disturbance. For each object, we set a different initial velocity based on the user input to make the object move as specified by the user.", + "bbox": [ + 496, + 794, + 893, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Other Visual Effects. Besides real world physics simulation, our pipeline allows special effects like collapsing and melting. To simulate different materials (rigid, soft, or granular), we can easily change the material type to modify", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6181", + "bbox": [ + 482, + 946, + 513, + 958 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0ee7dc5270c6c9c2dba35d2ac98e011544b60151acdb5dead4e9656f78f5d80d.jpg", + "image_caption": [ + "Figure 3. Video generation results. Left: Input initial frame. Right: Generated future frames. We apply an initial velocity to each movable object and use the physically grounded parameters outlined in 3.2 to generate physically plausible results." + ], + "image_footnote": [], + "bbox": [ + 158, + 71, + 807, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the physical properties of an object. This approach provides more flexibility for the user to edit the scene.", + "bbox": [ + 75, + 383, + 470, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Physics-Based Rendering", + "text_level": 1, + "bbox": [ + 76, + 422, + 308, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After dynamics simulation, we obtain object point trajectories and apply motion interpolation to compute vertex motions, deforming the mesh accordingly. With optimized PBR materials, we use Mitsuba3 for Physically-Based Rendering under environment lighting. Following prior insertion rendering work [16, 38, 75, 88], we avoid converting the entire static background into the rendering pipeline. Instead, we build a 3D shadow catcher surface from the background depth. During rendering, no texture is applied to the background; two-pass shadow mapping extracts shadows and global illumination effects. The foreground objects and shadows are then composited onto the inpainted background to produce the final video with realistic lighting.", + "bbox": [ + 75, + 446, + 473, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 656, + 282, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setups", + "text_level": 1, + "bbox": [ + 76, + 681, + 163, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The test images come from a diverse set of our own photography, online collections, and generative models. Our pipeline is primarily designed for object-centric scenes with one or a few objects. Due to our limitations (Sec. 5), we excluded those with excessive objects, heavy occlusion between objects, or highly uneven surfaces.", + "bbox": [ + 75, + 704, + 470, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Post-processing. We used VEnhancer [26] as an optional post-processing module, which takes the produced video and a text prompt to perform enhancement. As shown in Table 2, while it restores some details, it also introduces extra hallucinations.", + "bbox": [ + 75, + 795, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. Our method is one of the first of its kind, as existing model-simulate approaches require multi-view im", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ages [78, 80, 92] or specific settings [43, 48, 66]. Therefore, we evaluate ours against Image-to-Video (I2V) models: two open-source motion controller models DragAnything[77], MOFA-Video[53] and three state-of-the-art (SOTA) commercial models Kling 1.0 [3], Gen-3 [2], and Pika 1.5 [1]. We manually set correct motion trajectories and select applied regions for DragAnything, MOFA-Video and Kling to provide privileged motion guidance. We give text descriptions to Pika 1.5 and Gen-3, as they lack direct motion control capabilities. Additionally, Pika 1.5 offers \"Pikafect\" effects, such as \"Melt it\" and \"Deflate it.\"", + "bbox": [ + 496, + 382, + 893, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 500, + 559, + 591, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our system generates a miniature interactive world from a single image, enabling the simulation of various phenomena. Fig. 3 presents videos generated from different types of images. These images encompass single or multiple objects and various physical materials (i.e., rigid or soft). We also show applications like dynamic changes, object editing, and dense 3D tracking, illustrating the adaptability and creative potential of our approach for generating customized, interactive video content.", + "bbox": [ + 496, + 583, + 893, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparisons. We compare the results in two dimensions: motion control alone and physical materials. Fig. 4 shows that our system produces more physically plausible and controllable videos compared to SOTA I2V models. Despite prompt tuning, learning-based models often hallucinate, failing to adhere to physical laws or user intent. For example, in the toy dog case, we manually adjust the material and accurately simulate a collapse, whereas other models fall short. Similarly, for the book case, our results are the most physically realistic.", + "bbox": [ + 496, + 719, + 893, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamics. Fig. 5 demonstrates the varying dynamics generated from the same input image, highlighting the high", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6182", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7d701def37dd7cf51e8cb9c57d4b5ee8c67cb19161258ad41486b2731a9aef22.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison. We compare videos generated from our framework with three state-of-the-art I2V models: Gen-3 [2], Pika [1] and Kling [3]. We carefully designed the prompt to describe the motion outcome, and uses motion brush to control Kling. Our framework employs initial velocity control. Results show that our method can follow text instructions while maintaining plausible physics." + ], + "image_footnote": [], + "bbox": [ + 156, + 70, + 815, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "controllability of our method over physical parameters and motion trajectories. In the three rows on the left, we set different elasticities for the two objects, while keeping their initial positions and velocities the same. In the three rows on the right, we alter velocity directions of the objects and", + "bbox": [ + 75, + 816, + 472, + 893 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "keep physical parameters the same.", + "bbox": [ + 500, + 816, + 735, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Editing. Our method enables modifications to videos by removing, adding, or replacing objects, as illustrated in Fig. 6. The generated 3D assets can be easily manipulated, allowing for diverse video edits.", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6183", + "bbox": [ + 482, + 946, + 514, + 959 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f40b3e95e99456ead552db3794caf0cac6a6c2f80a03035004970785dc473f0.jpg", + "image_caption": [ + "Figure 5. Dynamics Effects. We can generate various dynamics from the same input image. The left three columns share the same initial positions and velocities, but are in different materials. The right three columns have the same material, but defer in velocity directions. This showcases the potential of our method for generating diverse physical scenarios." + ], + "image_footnote": [], + "bbox": [ + 107, + 78, + 857, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd89ae7d75def4d460580a26442d603f49deb985da903ce1225730fa86ed802e.jpg", + "image_caption": [ + "Figure 6. Video edition. As illustrated, our method supports video editing. In the top row, we remove the chair and allow the toy to fall from a static position. In the bottom two rows, we exchange one object between two scenes while keeping the other unchanged. This demonstrates the great flexibility of our video generation approach." + ], + "image_footnote": [], + "bbox": [ + 161, + 415, + 810, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tracking. Our framework uses an explicit 3D representation and works with a particle-based physics simulator. This allows our method to easily create videos with detailed 3D tracking results. Fig. 7 showcases two examples, demonstrating the accuracy and reliability of the tracking in different scenarios.", + "bbox": [ + 75, + 753, + 472, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation. Each step we designed in perception, simulation and rendering is intended to mimic the real world. Fig. 8 shows the results of ablation study involving position", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "optimization and inverse texture. Without pose optimization using differentiable rendering, the two objects are roughly at the initial place shown in input image, but cannot fully replicate the scene. Without inverse texture, the generated object mesh does not match the input in terms of color tone and brightness. We also conducted ablations on point sampling. However, without sampling, a large number of points become crowded in several MPM simulation grids, causing the simulator to crash and fail to produce final outputs.", + "bbox": [ + 496, + 753, + 893, + 892 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6184", + "bbox": [ + 482, + 946, + 514, + 959 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ccd443d8f46e08f3d34a1c25be6065d405314fa428a6c77905917dbc08af94d.jpg", + "image_caption": [ + "Figure 7. Dense 3D Tracking. Our method can naturally generate videos with dense 3D tracking results. Here we show tracking results for collapse and bounce cases." + ], + "image_footnote": [], + "bbox": [ + 117, + 68, + 225, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d6fe1b6722bb10feb967a35d7ede758d22f7ad745f8ed82126f3daeaf8f91a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 68, + 325, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26c42bb64e6966cb03323f952e349a13dc34e0bffb5252a28abcfb91e30d45fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 69, + 421, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/87f9c188d3cb717ded077e3d3c21d92b24ddb579536ac105935fa56acff25b97.jpg", + "image_caption": [ + "Figure 8. Ablation study. Without pose optimization, the two objects cannot fully replicate the scene. Without inverse texture, the object mesh does not match the input color." + ], + "image_footnote": [], + "bbox": [ + 78, + 214, + 467, + 301 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5d7c57c950b377a7b5474a97043cedfe6d5c72396bf614c4ca5149a83f553d3a.jpg", + "table_caption": [ + "Table 1. Human Evaluation Results. The three criteria are: Physical Realism, Photorealism and Semantic Consistency" + ], + "table_footnote": [], + "table_body": "
MethodsPhysRealPhotoRealAlign
Kling 1.0 [3]2.8113.5662.467
Runway Gen-3 [2]2.2833.5821.886
Pika 1.5 [1]2.4123.3142.016
Ours3.7073.4113.866
", + "bbox": [ + 135, + 393, + 403, + 460 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bc5664ff1dbde64ec177069c29ca34b7272416b0e2447ee115f1bd081ce1b0e7.jpg", + "table_caption": [ + "Table 2. VBenchmark Scores and GPT-4o Evaluation Results. Motion and Imaging refers to Motion Smoothness and Imaging Quality scores in VBenchmark. The three criteria on the right are the same as Table 1, with results given by GPT-4o." + ], + "table_footnote": [], + "table_body": "
MethodsMotion↑Imaging↑PhysReal↑PhotoReal↑Align↑
Kling 1.00.9960.6710.5630.8740.596
Runway Gen-30.9910.7230.1410.8960.144
Pika 1.50.9940.6710.5440.8630.563
MOFA-Video0.9940.6340.3840.7640.304
DragAnything0.9850.4280.6450.7560.380
Ours0.9950.6660.7520.8670.796
Ours+VEnhancer0.9940.6770.7660.8800.745
", + "bbox": [ + 81, + 527, + 465, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.Quantitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 635, + 308, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess the quality of the generated videos, we performed human evaluation, GPT-4o evaluation and provided two VBenchmark[34] evaluation scores.", + "bbox": [ + 76, + 657, + 468, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benchmarks. We designed three criteria for human and GPT-4o evaluation. (1) Physical Realism (PhysReal) measures how realistically the video follows the physical rules and whether the video represents real physical properties like elasticity and friction. (2) Photorealism (Photoreal) assesses the overall visual quality of the video, including the visual artifacts, discontinuities, and how accurately the video replicates details of light, shadow, texture, and materials. (3) Semantic Consistency (Align) evaluates how well the content of the generated video aligns with the intended meaning of the text prompt. We also chose two Quality Scores in VBenchmark: Motion Smoothness and Imaging Quality.", + "bbox": [ + 75, + 704, + 468, + 885 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Details in evaluation. Following a similar methodology", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/515f589cee55e3624b71f3fa169f8d8ff40886f9004a775088e0ac648b8552ec.jpg", + "image_caption": [ + "Figure 9. Limitations. The first set of two images show rendering errors: extreme lighting and heavy shading cause incorrect painting and artifacts from the pillow penetrating the ground. The second set reveals simulation limitations: the MPM method struggles with fine details, such as the thin kettle handle, leading to failures." + ], + "image_footnote": [], + "bbox": [ + 504, + 69, + 888, + 154 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to [48], we designed a questionnaire with 27 videos covering various scenes, motion conditions, and effects. Each video includes an input image, a motion prompt, and outputs from three SOTA commercial baselines and ours, shown in random order. 31 participants rated three criteria on a five-point scale from strongly disagree (1) to strongly agree (5). We also evaluated all five baselines and our diffusion-enhanced version using GPT-4o and VBenchmark. GPT-4o assessed videos on the same criteria based on 10 evenly sampled frames, with the input image and prompt provided. We used GPT version gpt-4o-2024-08-06, with detailed prompts in the supp.", + "bbox": [ + 496, + 246, + 892, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Result analysis. The results in Table 1 demonstrate our ability to generate physically accurate and controllable videos. In physical realism (PhysReal) and semantic consistency (Align), our method achieves the highest scores and outperforms all the commercial models by a large margin. Table 2 presents the results of GPT-4o and VBench evaluation. GPT score is aligned with human evaluation, and we also beat open-sourced in VBench. Among the baseline models, Kling 1.0 performs the best overall, likely due to its use of the \"motion brush\", which specifies motion trajectories.", + "bbox": [ + 496, + 412, + 893, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 500, + 577, + 622, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our single-image-based interactive miniature world is designed for object-centric scenes with simple spatial geometry and inter-object relationships. Reconstructing complete scenes for more complex scenarios remains an open challenge. Fig. 9 illustrates several failure cases, such as rendering errors under challenging shading, perception failures, and simulation limitations. More detailed analysis on limitations can be found in the supp.", + "bbox": [ + 496, + 602, + 893, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 737, + 625, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present PhysGen3D, a framework that transforms a static image into an interactive 3D scene for simulating and rendering future motions based on user input. PhysGen3D integrates modules for 3D world reconstruction, model-based dynamic simulation, and physics-based rendering to generate realistic, controllable videos. By extending the 2D image-to-video paradigm to 3D, PhysGen3D enables more realistic motion and diverse material behaviors. We hope this work inspires future research.", + "bbox": [ + 496, + 762, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6185", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 78, + 90, + 236, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project is supported by the Intel AI SRS gift, Amazon-Illinois AICE grant, Meta Research Grant, IBM IIDAI Grant, and NSF Awards #2331878, #2340254, #2312102, #2414227, and #2404385. We greatly appreciate the NCSA for providing computing resources.", + "bbox": [ + 76, + 114, + 470, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 199, + 173, + 215 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pika, 2024. 2, 5, 6, 8", + "[2] Runway, 2024. 5, 6, 8", + "[3] Klingai,2024.2,5,6,8", + "[4] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. ACM TOG, 2020. 3", + "[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2", + "[6] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In CVPR, 2021. 3", + "[7] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In ICCV, 2021. 3", + "[8] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In CVPR, 2021. 3", + "[9] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2", + "[10] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2", + "[11] Richard Strong Bowen, Richard Tucker, Ramin Zabih, and Noah Snavely. Dimensions of motion: Monocular prediction through flow subspaces. In 3DV, 2022. 3", + "[12] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In European Conference on Computer Vision, pages 552-567. Springer, 2022. 4", + "[13] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2", + "[14] Xi Chen, Zhiheng Liu, Mengting Chen, Yutong Feng, Yu Liu, Yujun Shen, and Hengshuang Zhao. Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928, 2023. 3", + "[15] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion" + ], + "bbox": [ + 78, + 224, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 2", + "[16] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7230-7240, 2021. 5", + "[17] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. ACM TOG, 2005. 2, 3", + "[18] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM TOG, 2015. 2, 3", + "[19] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. arXiv preprint arXiv:1910.07192, 2019. 3", + "[20] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 4", + "[21] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 2", + "[22] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In ICLR, 2024. 3", + "[23] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2", + "[24] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning, 2023. 3", + "[25] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2", + "[26] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation. arXiv preprint arXiv:2407.07667, 2024. 5", + "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 2", + "[28] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[29] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In CVPR, 2021. 2, 3" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6186", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Yuanming Hu, Yu Fang, Ziheng Ge, Ziyin Qu, Yixin Zhu, Andre Pradhana, and Chenfanfu Jiang. A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG, 2018. 2", + "[31] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG), 38(6):201, 2019. 2, 4", + "[32] Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Frédo Durand. Difftaichi: Differentiable programming for physical simulation. *ICLR*, 2020.", + "[33] Yuanming Hu, Jiafeng Liu, Xuanda Yang, Mingkuan Xu, Ye Kuang, Weiwei Xu, Qiang Dai, William T. Freeman, and Frédo Durand. Quantaichi: A compiler for quantized simulations. ACM Transactions on Graphics (TOG), 40(4), 2021. 2, 4", + "[34] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 8", + "[35] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, and Delio Vicini. Dr.jit: A just-in-time compiler for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH), 41(4), 2022. 4", + "[36] Wei-Cih Jhou and Wen-Huang Cheng. Animating still landscape photographs through cloud motion creation. IEEE Transactions on Multimedia, 2015. 3", + "[37] Chenfanfu Jiang, Craig Schroeder, Joseph Teran, Alexey Stomakhin, and Andrew Selle. The material point method for simulating continuum materials. In Acm siggraph 2016 courses, pages 1-52. 2016. 2, 4", + "[38] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on graphics (TOG), 30(6):1-12, 2011. 5", + "[39] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2", + "[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 3", + "[41] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2", + "[42] Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In ICLR, 2023. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. arXiv preprint arXiv:2309.07906, 2023. 2, 3, 5", + "[44] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 2", + "[45] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10072–10083, 2024. 2", + "[46] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2, 3", + "[47] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3", + "[48] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 2, 3, 5, 8", + "[49] Xinyu Liu, Jinlong Li, Jin Ma, Huiming Sun, Zhigang Xu, Tianyun Zhang, and Hongkai Yu. Deep transfer learning for intelligent vehicle perception: A survey. Green Energy and Intelligent Transportation, page 100125, 2023. 2", + "[50] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 2", + "[51] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In CVPR, 2022. 3", + "[52] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit warping for animation with image sets. 2022. 3", + "[53] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. In European Conference on Computer Vision, pages 111-128. Springer, 2024. 5", + "[54] OpenAI. Creating video from text. https://openai.com/sora, 2024.2", + "[55] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 2", + "[56] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 98-108, 2024. 4", + "[57] Haonan Qiu, Zhaoxi Chen, Yuming Jiang, Hang Zhou, Xiangyu Fan, Lei Yang, Wayne Wu, and Ziwei Liu. Relitalk: Relightable talking portrait generation from a single video, 2023. 3", + "[58] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6187", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 2, 3", + "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[60] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 4", + "[61] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In PACMCGIT, 2000. 2, 3", + "[62] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 2, 3", + "[63] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[64] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. 2015. 2", + "[65] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. 2019. 2", + "[66] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia, 2022. 2, 3, 5", + "[67] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 2, 4", + "[68] Martin Szummer and Rosalind W. Picard. Temporal texture modeling. In ICIP, 1996. 2, 3", + "[69] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2", + "[70] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. arXiv preprint arXiv:2402.01566, 2024. 2", + "[71] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 4", + "[72] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3", + "[73] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2", + "[74] Li-Yi Wei and Marc Levoy. Fast texture synthesis using tree-structured vector quantization. In PACMCGIT, 2000. 2, 3", + "[75] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via llm-agent collaboration. In The First Workshop on Populating Empty Cities—Virtual Humans for Robotics and Autonomous Driving at CVPR 2024, 2nd Round. 5", + "[76] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In CVPR, 2019. 3", + "[77] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 5", + "[78] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In CVPR, 2024. 5", + "[79] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time interactive realistic and browser-compatible environment from a single video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4588, 2024. 4", + "[80] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In CVPR, 2024. 5", + "[81] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 2", + "[82] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccrafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2", + "[83] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 3", + "[84] Tianfan Xue, Jiajun Wu, Katherine L Bouman, and William T Freeman. Visual dynamics: Stochastic future generation via layered cross convolutional networks. T-PAMI, 2018. 2, 3", + "[85] Chenggang Yan, Biyao Shao, Hao Zhao, Ruixin Ning, Yongdong Zhang, and Feng Xu. 3d room layout estimation from a single rgb image. IEEE Transactions on Multimedia, 22(11): 3014-3024, 2020. 2", + "[86] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. Magicprop: Diffusion-based video edit" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6188", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing via motion-aware appearance propagation. arXiv preprint arXiv:2309.00908, 2023. 2", + "[87] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. Entropy, 25 (10):1469, 2023. 2", + "[88] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 5", + "[89] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9(1):1, 2023. 3", + "[90] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2", + "[91] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qing, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2", + "[92] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2025. 2, 5", + "[93] Xianling Zhang, Nathan Tseng, Ameerah Syed, Rohan Bhasin, and Nikita Jaipuria. Simbar: Single image-based scene relighting for effective data augmentation for automated driving vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3718-3728, 2022. 2", + "[94] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In CVPR, 2022. 3", + "[95] Yupeng Zheng, Chengliang Zhong, Pengfei Li, Huan-ang Gao, Yuhang Zheng, Bu Jin, Ling Wang, Hao Zhao, Guyue Zhou, Qichao Zhang, et al. Steps: Joint self-supervised nighttime image enhancement and depth estimation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4916–4923. IEEE, 2023. 2", + "[96] Chengliang Zhong, Yuhang Zheng, Yupeng Zheng, Hao Zhao, Li Yi, Xiaodong Mu, Ling Wang, Pengfei Li, Guyue Zhou, Chao Yang, et al. 3d implicit transporter for temporally consistent keypoint discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3869-3880, 2023. 2", + "[97] Leisheng Zhong, Yu Zhang, Hao Zhao, An Chang, Wenhao Xiang, Shunli Zhang, and Li Zhang. Seeing through the occluders: Robust monocular 6-dof object pose tracking via model-guided video object segmentation. IEEE Robotics and Automation Letters, 5(4):5159-5166, 2020. 2" + ], + "bbox": [ + 78, + 92, + 470, + 853 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6189", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/1bf0def3-a278-47bf-9ffd-13a93c05e545_model.json b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/1bf0def3-a278-47bf-9ffd-13a93c05e545_model.json new file mode 100644 index 0000000000000000000000000000000000000000..41e3f5816fa086ca45fcb34292fe52f02845f1bb --- /dev/null +++ b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/1bf0def3-a278-47bf-9ffd-13a93c05e545_model.json @@ -0,0 +1,2490 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "PhysGen3D: Crafting a Miniature Interactive World from a Single Image" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.18, + 0.717, + 0.217 + ], + "angle": 0, + "content": "Boyuan Chen\\(^{1}\\), Hanxiao Jiang\\(^{2,3}\\), Shaowei Liu\\(^{2}\\), Saurabh Gupta\\(^{2}\\), Yunzhu Li\\(^{3}\\), Hao Zhao\\(^{1}\\), Shenlong Wang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.218, + 0.744, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University, \\(^{2}\\)University of Illinois Urbana-Champaign, \\(^{3}\\)Columbia University" + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.264, + 0.165, + 0.277 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.279, + 0.216, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.385, + 0.216, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.492, + 0.191, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.264, + 0.657, + 0.277 + ], + "angle": 0, + "content": "Generation (left \\(\\rightarrow\\) right: time steps)" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.278, + 0.891, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.385, + 0.891, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.492, + 0.891, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.607, + 0.894, + 0.679 + ], + "angle": 0, + "content": "Figure 1. PhysGen3D generates realistic, physically plausible motion from a single image and a text prompt by reasoning about geometry, semantics, and material properties. (a) An apple rolls under the influence of its initial velocity, friction, and shape, producing a natural progression over time. (b) Three animal figures interact dynamically, colliding after being propelled upwards and forwards. (c) A toy potato bounces back with soft-body dynamics in response to an initial downward force, capturing material-specific behaviors. PhysGen3D lets users quickly explore physics-driven object interactions and behaviors in a compact virtual scene generated from a single input image." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.694, + 0.314, + 0.71 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Envisioning physically plausible outcomes from a single image requires a deep understanding of the world's dynamics. To address this, we introduce PhysGen3D, a novel framework that transforms a single image into an amodal, camera-centric, interactive 3D scene. By combining advanced image-based geometric and semantic understanding with physics-based simulation, PhysGen3D creates an interactive 3D world from a static image, enabling us to \"imagine\" and simulate future scenarios based on user input. At its core, PhysGen3D estimates 3D shapes, poses, physical and lighting properties of objects, thereby" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.896, + 0.906 + ], + "angle": 0, + "content": "capturing essential physical attributes that drive realistic object interactions. This framework allows users to specify precise initial conditions, such as object speed or material properties, for enhanced control over generated video outcomes. We evaluate PhysGen3D's performance against closed-source state-of-the-art (SOTA) image-to-video models, including Pika, Kling, and Gen-3, showing PhysGen3D's capacity to generate videos with realistic physics while offering greater flexibility and fine-grained control. Our results show that PhysGen3D achieves a unique balance of photorealism, physical plausibility, and user-driven interactivity, opening new possibilities for generating dynamic, physics-grounded video from an image. Project page: https://by-luckk.github.io/PhysGen3D." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "6178" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.297 + ], + "angle": 0, + "content": "Photographs capture snapshots of our physical world, preserving specific moments in time but leaving out the alternative outcomes that could have unfolded. For instance, looking at a photo, we might wonder, \"What if I poke the apples to make them roll across the ground?\" or \"What if I squeeze the three stuffed animals closer together?\" or \"What if I drop my cute potato toys onto the floor?\" Humans intuitively understand how these scenarios would play out because we have an innate sense of the physical world beyond what we see in a single image. We develop a computational model that can answer such \"what-if\" questions by generating video outcomes from a single static image." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.297, + 0.471, + 0.432 + ], + "angle": 0, + "content": "A promising approach toward this goal is data-driven image-to-video (I2V) generation [17, 18, 29, 61, 68, 74, 84]. I2V leverages diffusion-based generative models trained on vast datasets of internet images and videos, enabling the production of photorealistic videos with remarkable detail. However, I2V still has limitations in precise control and lacks physical grounding. As a result, users cannot interact freely and accurately to achieve specific physical effects, nor can I2V guarantee physical realism." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.433, + 0.471, + 0.629 + ], + "angle": 0, + "content": "On the other hand, recent research has focused on modeling the physical world from visual inputs to create digital twins, allowing for precise interactions [42, 81, 92, 96]. These approaches can generate virtual scenes with convincing physical interactions, but they typically require complete 3D scans from multi-view images or depth sensors, making them data-intensive. While some methods [43, 48, 66] enable interaction with a single image, they are often limited by physical constraints (e.g., rigid bodies or springs), specific object types (e.g., waterfalls), or a 2D scope. This gap highlights the need for a generic, controllable, physically grounded, and photorealistic approach to generate video from a single image while maintaining physical realism." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.63, + 0.47, + 0.84 + ], + "angle": 0, + "content": "In this work, we introduce PhysGen3D, a novel framework that transforms a single image into an amodal, camera-centric, interactive 3D scene, enabling realistic simulation and rendering. Our approach combines the strengths of image-based geometric and semantic understanding [40, 58, 67, 71, 83] with physics-based simulation [31-33]. At its core, PhysGen3D is a digital twinning method that estimates an object's 3D shape, pose, physical and lighting properties, infers background geometry and appearance, deduces physical characteristics, and performs dimensional analysis—all from a single input image. This task is conventionally challenging due to its inherently ill-posed nature. To tackle this, we leverage various pretrained vision models, integrating their outputs to create an image-centric digital twin." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "For physical simulation, we employ material point methods [30, 37], a robust point-voxel-based framework that models counterfactual physical behaviors of objects in the image. Through precise inference of physical properties," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "simulations in the PhysGen3D environment achieve a high degree of realism and stability. We further enhance realism by applying physics-based rendering, seamlessly integrating dynamic effects back into the original image. PhysGen3D produces results that are not only visually realistic in terms of dynamics and lighting but also highly controllable, allowing users to specify initial conditions like speed and material properties. Due to the use of large pretrained models, our pipeline operates effectively without task-specific training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.895, + 0.337 + ], + "angle": 0, + "content": "Our experiments, based on a carefully designed and rigorous user study, demonstrate that, compared to closed-source state-of-the-art video AIGC models such as Pika, Kling, and Gen-3, our framework provides significantly more flexible control over object motions, generates videos that better align with user intentions, and achieves superior physical realism—all while maintaining comparable rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.359, + 0.637, + 0.375 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.616 + ], + "angle": 0, + "content": "Single-view 3D reconstruction Great progress has been made in object-centric single view 3D reconstructions [45, 46, 62, 83]. However, applying these techniques to single-view 3D scene reconstruction becomes more challenging. Most existing works only focus on one part of 3D scene understanding. Geometric approaches reconstruct the 3D scene holistically but neglect individual object understanding, or they focus solely on foreground objects without considering the interaction of objects with the complex environment. Additionally, works have been done for scene relighting [49, 93, 95] and segmentation [40, 47, 58, 85, 97], but none of these provide a full 3D understanding within the image. In our work, we also perform perception reasoning for object materials, backgrounds, rendering properties, and physical stability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Controllable video generation. Video generation has made significant progress in recent years [5, 9, 10, 15, 21, 23, 25, 28, 41, 54, 63, 69, 73, 82, 90, 91]. The state-of-the-art framework [1, 3] can generate photorealistic and coherent videos from text instructions using diffusion models [27, 55, 59, 64, 65]. Controllable video generation is achieved through leveraging pre-trained video generation models with conditioning information, including but not limited to depth maps [13, 86], linear translation [39], layouts [44, 70, 87], and multiple combinations [50]. However, most existing generation methods implicitly generate image space dynamics, which might lead to unrealistic, hallucinated motions. In contrast, our method explicitly controls the motion and interaction with simulation, allowing us to create more sophisticated effects without the need for extensive training data. Our approach is training-free and generalizable to all objects in the world." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "6179" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.078, + 0.891, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.373, + 0.897, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Method Overview. PhysGen3D's framework consists of three modules: a) 3D world creation, which infers geometry, semantics, rendering and physical parameters from the input image; b) dynamics simulation using Taichi-Elements for particle-based physics; and c) physics-based rendering with a two-pass shadow mapping technique." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.474, + 0.718 + ], + "angle": 0, + "content": "Controllable image animation Image-based animation aims to animate objects that appear in images. Numerous works [17, 18, 29, 61, 68, 74, 84] have focused on this task. To improve quality, recent research has adopted data-driven solutions, training temporal neural networks to directly predict subsequent video frames [6, 14, 19, 24, 29, 72], or incorporating physical heuristics [17, 36]. Recently, there has been increased focus on interactive [7, 8, 43] and controllable [4, 18] image-to-video synthesis. Additional priors such as motion fields [19, 29, 51, 52, 66, 84], optical flow [11, 22, 94], 3D geometry information [57, 76], and user annotations [43] have also been introduced. The closest work to ours is Liu et al. [48], which focuses on image understanding and uses an existing 2D rigid simulator to generate 2D animations. However, their approach is limited to 2D dynamics, does not account for real-world 3D physics, and is restricted to rigid bodies. In contrast, our method provides more realistic and flexible control over 3D motions and object materials and extends beyond rigid-body physics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.738, + 0.188, + 0.755 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our goal is to reconstruct an interactive, camera-centric miniature world from a single input image. We aim to control object materials, dynamics, and motions to simulate diverse, photorealistic, and physically plausible videos from user input. The key challenge lies in partial observations and the ill-posed nature of physical reasoning without observed dynamics. To address this, we propose a holistic reconstruction method leveraging large pretrained visual models to jointly infer geometry, dynamic materials, lighting, and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.896, + 0.537 + ], + "angle": 0, + "content": "PBR materials from a single image (Sec. 3.1). The reconstructed scene is input into a material-point-method (MPM) simulator to generate realistic physics phenomena (Sec. 3.2). Finally, we render dynamic object behaviors based on the simulation and reintegrate them into the scene, producing realistic videos with accurate appearance and motion (Sec. 3.3). Fig. 2 depicts our framework." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.545, + 0.86, + 0.562 + ], + "angle": 0, + "content": "3.1. Interactive 3D World from a Single Image" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.689 + ], + "angle": 0, + "content": "A full amodal reconstruction of the 3D world depicted in the image is critical for the next step in simulation. An ideal reconstruction should capture a comprehensive understanding of objects' relationships, geometry, appearance, material, and physical properties. However, obtaining this understanding from a single image is highly ill-posed. Our key idea is to leverage priors from pretrained vision foundation models to help infer these properties, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Segmentation. We leverage vision foundation models to recognize object categories and segment object instances. Specifically, we use GPT-4o [89] to identify the foreground object categories and use Grounded-SAM [40, 47, 58] to further detect and segment each individual instance \\(\\{o^i\\in \\mathbb{R}^{W\\times H\\times 3}\\}_{i}^{N}\\), where \\(o^i\\) is the image of i-th object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Mesh Generation. Unlike previous work [48], which used only 2D rigid-body physics, we extend to general-purpose 3D simulation. This requires a complete 3D representation of foreground objects. We adopt InstantMesh [83], which uses Zero123++ [46, 62] to synthesize multi-view images from the segmented object image \\( o^i \\) and uses these images to reconstruct the 3D mesh \\( \\mathcal{O} \\) of the objects. For multiple object occlusions, we adopt the iterative inpainting" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "6180" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "strategy to extract the 3D mesh of the objects (see supp for more details)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.471, + 0.349 + ], + "angle": 0, + "content": "Background Handling. The background serves important roles in both dynamic simulation and rendering. In simulation, background geometry acts as a support and collider, where accuracy ensures realistic object-scene interactions. In rendering, it serves as a static backdrop while the foreground objects move and helps simulate realistic global illumination effects like cast shadows. For simulation purposes, we use Dust3r [71] to estimate iamage depth. The output depth map \\( z \\in \\mathbb{R}^{W \\times H} \\) is unprojected to the 3D world as a 3D point cloud \\( \\mathcal{P} \\), and Bilateral Normal Integration [12] is applied to generate a smooth surface \\( S \\) serving as the collider. For rendering, to fill background regions occluded by moving objects, we use the LaMA inpainting model[67]. This model generates a complete background after masking out all objects and their shadows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.486 + ], + "angle": 0, + "content": "Object Pose and Scale Estimation. The generative 3D reconstruction step provides a complete mesh with a normalized scale in an object-centric coordinate system, but it does not infer the object's location and scale in the camera coordinate. To ensure coherence with the input image, it is necessary to accurately place the 3D meshes \\(\\mathcal{O}\\) into the 3D scene \\(\\mathcal{P}\\) with the correct scale and 6DoF pose. This registration task is challenging, as the generated object mesh may not perfectly match the real-world object in the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.471, + 0.715 + ], + "angle": 0, + "content": "To address this challenge, we designed a multi-stage coarse-to-fine alignment strategy. In the coarse stage, we perform 2D-3D feature point matching. Firstly, we render multiple images from viewpoints uniformly distributed over a unit sphere surrounding the object. For each rendered image, we match its feature points with the original object image \\(o^i\\) using SuperGlue [60], and the viewpoint with the most matches is selected. Matched pairs \\((p_i^N \\in \\mathbb{R}^2, p_i^N \\in \\mathbb{R}^2)\\) are then projected back into 3D points \\((P_i^N \\in \\mathbb{R}^3, P_i^N \\in \\mathbb{R}^3)\\) in object and camera coordinate, respectively. Perspective-n-Point (PnP) algorithm [20] is applied between \\(P_i^N\\) and \\(p_i^N\\) in image \\(o^i\\) to estimate the 6DoF pose with scale ambiguity. Then we adjust the scale and translation factor simultaneously to minimize the L2 loss \\(\\sum_{i}^{N} = ||P_i^N - P_i'^N||_2\\), without altering its projection." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.716, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In the fine alignment stage, we render the mask and depth on the image plane using the current estimation through a differentiable renderer and jointly minimize two losses: \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{dice}} + \\mathcal{L}_{\\mathrm{depth}}\\). Here, \\(\\mathcal{L}_{\\mathrm{dice}} = 1 - \\frac{2\\times|M_A\\cap M_B|}{|M_A| + |M_B|}\\) measures the discrepancy between the rendered mask \\(M_{A}\\) and the observed mask \\(M_B\\) from Grounded-SAM. The depth consistency loss is \\(\\mathcal{L}_{\\mathrm{depth}} = \\frac{\\mathrm{mse}(M_B*z_A - M_B*z_B)}{|M_B|}\\), where \\(z_{A}\\) and \\(z_{B}\\) are the rendered and Dust3r-predicted depths. This joint optimization ensures consistency between the estimated mesh in the camera coordinate and the point cloud from Dust3r, while maximizing the alignment between the 3D pose and observed object mask, ensuring accurate simu" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.71, + 0.107 + ], + "angle": 0, + "content": "lation and rendering alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.895, + 0.319 + ], + "angle": 0, + "content": "Appearance Optimization. The texture of the generated 3D meshes may differ from the input image. To enhance rendering quality, we use the inverse rendering pipeline in Mitsuba3 [35] to estimate material properties. Lighting parameters are estimated with DiffusionLight [56], while object PBR materials (albedo, roughness, and metallic) are optimized via differentiable rendering. To handle unknown back views and reduce complexity, we assume uniform roughness and metallic values per object and apply tone mapping for albedo optimization: \\( y(x) = ax^3 + bx^2 + cx + d \\), where \\( y(0) = 0 \\), \\( y(1) = 1 \\). With reconstructed lighting and ground surface, the optimized materials improve asset appearance, capturing realistic object-surface interactions during rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.895, + 0.471 + ], + "angle": 0, + "content": "Physics Reasoning. Accurately simulating real-world dynamics requires estimating physical parameters. We focus on two aspects: 1) Following [79], we use GPT-4o to query each object's elasticity and density, and the friction coefficient for the surface \\( S \\). 2) We ensure reconstructed 3D assets match real-world proportions, as depth inaccuracies from Dust3r can cause unrealistic behaviors. To address this, we estimate a scale factor \\( k \\) by comparing asset size with typical real-world sizes from GPT-4o and use \\( k \\) for dimensionless scaling of gravity and velocity-related parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.702, + 0.497 + ], + "angle": 0, + "content": "3.2. Dynamics Simulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.894, + 0.581 + ], + "angle": 0, + "content": "Given the 3D assets with reasoned physical properties and the scale factor, we use the physics engine Taichi Elements [31-33], based on the Material Point Method (MPM) [37], as our simulator to support a variety of different materials, including but not limited to rigid, soft, and granular." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.895, + 0.703 + ], + "angle": 0, + "content": "Particle representation. The simulator is based on a particle-based representation. To convert our 3D assets into a simulatable particle-based representation, we apply floating points removing, internal filling and voxel downsampling. We use downsampling to handle uneven particle distributions, where downsample rate is adjusted according to the grid size of the simulator. For the convenience of rendering, we prioritize surface points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.894, + 0.794 + ], + "angle": 0, + "content": "Physical parameters. To enhance stability, we apply the scale factor \\( k \\) to physical parameters in simulator instead of scaling the assets to real size. For example, the gravitational acceleration is multiplied by \\( k \\), making to \\( k * 9.8 \\) (nondimensionalized here). With this tuning, the motion of falling or collapse remains realistic for all scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.84 + ], + "angle": 0, + "content": "External disturbance. For each object, we set a different initial velocity based on the user input to make the object move as specified by the user." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Other Visual Effects. Besides real world physics simulation, our pipeline allows special effects like collapsing and melting. To simulate different materials (rigid, soft, or granular), we can easily change the material type to modify" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.514, + 0.959 + ], + "angle": 0, + "content": "6181" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.072, + 0.808, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.345, + 0.895, + 0.377 + ], + "angle": 0, + "content": "Figure 3. Video generation results. Left: Input initial frame. Right: Generated future frames. We apply an initial velocity to each movable object and use the physically grounded parameters outlined in 3.2 to generate physically plausible results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.471, + 0.414 + ], + "angle": 0, + "content": "the physical properties of an object. This approach provides more flexibility for the user to edit the scene." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.423, + 0.31, + 0.44 + ], + "angle": 0, + "content": "3.3. Physics-Based Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.474, + 0.645 + ], + "angle": 0, + "content": "After dynamics simulation, we obtain object point trajectories and apply motion interpolation to compute vertex motions, deforming the mesh accordingly. With optimized PBR materials, we use Mitsuba3 for Physically-Based Rendering under environment lighting. Following prior insertion rendering work [16, 38, 75, 88], we avoid converting the entire static background into the rendering pipeline. Instead, we build a 3D shadow catcher surface from the background depth. During rendering, no texture is applied to the background; two-pass shadow mapping extracts shadows and global illumination effects. The foreground objects and shadows are then composited onto the inpainted background to produce the final video with realistic lighting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.657, + 0.283, + 0.673 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.165, + 0.697 + ], + "angle": 0, + "content": "4.1. Setups" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.795 + ], + "angle": 0, + "content": "The test images come from a diverse set of our own photography, online collections, and generative models. Our pipeline is primarily designed for object-centric scenes with one or a few objects. Due to our limitations (Sec. 5), we excluded those with excessive objects, heavy occlusion between objects, or highly uneven surfaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Post-processing. We used VEnhancer [26] as an optional post-processing module, which takes the produced video and a text prompt to perform enhancement. As shown in Table 2, while it restores some details, it also introduces extra hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Baselines. Our method is one of the first of its kind, as existing model-simulate approaches require multi-view im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.895, + 0.55 + ], + "angle": 0, + "content": "ages [78, 80, 92] or specific settings [43, 48, 66]. Therefore, we evaluate ours against Image-to-Video (I2V) models: two open-source motion controller models DragAnything[77], MOFA-Video[53] and three state-of-the-art (SOTA) commercial models Kling 1.0 [3], Gen-3 [2], and Pika 1.5 [1]. We manually set correct motion trajectories and select applied regions for DragAnything, MOFA-Video and Kling to provide privileged motion guidance. We give text descriptions to Pika 1.5 and Gen-3, as they lack direct motion control capabilities. Additionally, Pika 1.5 offers \"Pikafect\" effects, such as \"Melt it\" and \"Deflate it.\"" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.593, + 0.575 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.718 + ], + "angle": 0, + "content": "Our system generates a miniature interactive world from a single image, enabling the simulation of various phenomena. Fig. 3 presents videos generated from different types of images. These images encompass single or multiple objects and various physical materials (i.e., rigid or soft). We also show applications like dynamic changes, object editing, and dense 3D tracking, illustrating the adaptability and creative potential of our approach for generating customized, interactive video content." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Comparisons. We compare the results in two dimensions: motion control alone and physical materials. Fig. 4 shows that our system produces more physically plausible and controllable videos compared to SOTA I2V models. Despite prompt tuning, learning-based models often hallucinate, failing to adhere to physical laws or user intent. For example, in the toy dog case, we manually adjust the material and accurately simulate a collapse, whereas other models fall short. Similarly, for the book case, our results are the most physically realistic." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Dynamics. Fig. 5 demonstrates the varying dynamics generated from the same input image, highlighting the high" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "6182" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.158, + 0.071, + 0.816, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.766, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison. We compare videos generated from our framework with three state-of-the-art I2V models: Gen-3 [2], Pika [1] and Kling [3]. We carefully designed the prompt to describe the motion outcome, and uses motion brush to control Kling. Our framework employs initial velocity control. Results show that our method can follow text instructions while maintaining plausible physics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.818, + 0.473, + 0.895 + ], + "angle": 0, + "content": "controllability of our method over physical parameters and motion trajectories. In the three rows on the left, we set different elasticities for the two objects, while keeping their initial positions and velocities the same. In the three rows on the right, we alter velocity directions of the objects and" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.818, + 0.736, + 0.835 + ], + "angle": 0, + "content": "keep physical parameters the same." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Editing. Our method enables modifications to videos by removing, adding, or replacing objects, as illustrated in Fig. 6. The generated 3D assets can be easily manipulated, allowing for diverse video edits." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.96 + ], + "angle": 0, + "content": "6183" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.079, + 0.859, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.364, + 0.895, + 0.406 + ], + "angle": 0, + "content": "Figure 5. Dynamics Effects. We can generate various dynamics from the same input image. The left three columns share the same initial positions and velocities, but are in different materials. The right three columns have the same material, but defer in velocity directions. This showcases the potential of our method for generating diverse physical scenarios." + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.416, + 0.812, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.697, + 0.895, + 0.74 + ], + "angle": 0, + "content": "Figure 6. Video edition. As illustrated, our method supports video editing. In the top row, we remove the chair and allow the toy to fall from a static position. In the bottom two rows, we exchange one object between two scenes while keeping the other unchanged. This demonstrates the great flexibility of our video generation approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.473, + 0.844 + ], + "angle": 0, + "content": "Tracking. Our framework uses an explicit 3D representation and works with a particle-based physics simulator. This allows our method to easily create videos with detailed 3D tracking results. Fig. 7 showcases two examples, demonstrating the accuracy and reliability of the tracking in different scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Ablation. Each step we designed in perception, simulation and rendering is intended to mimic the real world. Fig. 8 shows the results of ablation study involving position" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.755, + 0.895, + 0.893 + ], + "angle": 0, + "content": "optimization and inverse texture. Without pose optimization using differentiable rendering, the two objects are roughly at the initial place shown in input image, but cannot fully replicate the scene. Without inverse texture, the generated object mesh does not match the input in terms of color tone and brightness. We also conducted ablations on point sampling. However, without sampling, a large number of points become crowded in several MPM simulation grids, causing the simulator to crash and fail to produce final outputs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.96 + ], + "angle": 0, + "content": "6184" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.069, + 0.227, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.069, + 0.326, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.07, + 0.423, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.163, + 0.47, + 0.204 + ], + "angle": 0, + "content": "Figure 7. Dense 3D Tracking. Our method can naturally generate videos with dense 3D tracking results. Here we show tracking results for collapse and bounce cases." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.215, + 0.468, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.311, + 0.47, + 0.354 + ], + "angle": 0, + "content": "Figure 8. Ablation study. Without pose optimization, the two objects cannot fully replicate the scene. Without inverse texture, the object mesh does not match the input color." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.362, + 0.47, + 0.39 + ], + "angle": 0, + "content": "Table 1. Human Evaluation Results. The three criteria are: Physical Realism, Photorealism and Semantic Consistency" + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.395, + 0.405, + 0.461 + ], + "angle": 0, + "content": "
MethodsPhysRealPhotoRealAlign
Kling 1.0 [3]2.8113.5662.467
Runway Gen-3 [2]2.2833.5821.886
Pika 1.5 [1]2.4123.3142.016
Ours3.7073.4113.866
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.468, + 0.469, + 0.524 + ], + "angle": 0, + "content": "Table 2. VBenchmark Scores and GPT-4o Evaluation Results. Motion and Imaging refers to Motion Smoothness and Imaging Quality scores in VBenchmark. The three criteria on the right are the same as Table 1, with results given by GPT-4o." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.529, + 0.466, + 0.622 + ], + "angle": 0, + "content": "
MethodsMotion↑Imaging↑PhysReal↑PhotoReal↑Align↑
Kling 1.00.9960.6710.5630.8740.596
Runway Gen-30.9910.7230.1410.8960.144
Pika 1.50.9940.6710.5440.8630.563
MOFA-Video0.9940.6340.3840.7640.304
DragAnything0.9850.4280.6450.7560.380
Ours0.9950.6660.7520.8670.796
Ours+VEnhancer0.9940.6770.7660.8800.745
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.31, + 0.652 + ], + "angle": 0, + "content": "4.3.Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.659, + 0.469, + 0.703 + ], + "angle": 0, + "content": "To assess the quality of the generated videos, we performed human evaluation, GPT-4o evaluation and provided two VBenchmark[34] evaluation scores." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.886 + ], + "angle": 0, + "content": "Benchmarks. We designed three criteria for human and GPT-4o evaluation. (1) Physical Realism (PhysReal) measures how realistically the video follows the physical rules and whether the video represents real physical properties like elasticity and friction. (2) Photorealism (Photoreal) assesses the overall visual quality of the video, including the visual artifacts, discontinuities, and how accurately the video replicates details of light, shadow, texture, and materials. (3) Semantic Consistency (Align) evaluates how well the content of the generated video aligns with the intended meaning of the text prompt. We also chose two Quality Scores in VBenchmark: Motion Smoothness and Imaging Quality." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Details in evaluation. Following a similar methodology" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.07, + 0.89, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.164, + 0.892, + 0.234 + ], + "angle": 0, + "content": "Figure 9. Limitations. The first set of two images show rendering errors: extreme lighting and heavy shading cause incorrect painting and artifacts from the pillow penetrating the ground. The second set reveals simulation limitations: the MPM method struggles with fine details, such as the thin kettle handle, leading to failures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.893, + 0.414 + ], + "angle": 0, + "content": "to [48], we designed a questionnaire with 27 videos covering various scenes, motion conditions, and effects. Each video includes an input image, a motion prompt, and outputs from three SOTA commercial baselines and ours, shown in random order. 31 participants rated three criteria on a five-point scale from strongly disagree (1) to strongly agree (5). We also evaluated all five baselines and our diffusion-enhanced version using GPT-4o and VBenchmark. GPT-4o assessed videos on the same criteria based on 10 evenly sampled frames, with the input image and prompt provided. We used GPT version gpt-4o-2024-08-06, with detailed prompts in the supp." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.894, + 0.565 + ], + "angle": 0, + "content": "Result analysis. The results in Table 1 demonstrate our ability to generate physically accurate and controllable videos. In physical realism (PhysReal) and semantic consistency (Align), our method achieves the highest scores and outperforms all the commercial models by a large margin. Table 2 presents the results of GPT-4o and VBench evaluation. GPT score is aligned with human evaluation, and we also beat open-sourced in VBench. Among the baseline models, Kling 1.0 performs the best overall, likely due to its use of the \"motion brush\", which specifies motion trajectories." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.578, + 0.623, + 0.594 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.895, + 0.726 + ], + "angle": 0, + "content": "Our single-image-based interactive miniature world is designed for object-centric scenes with simple spatial geometry and inter-object relationships. Reconstructing complete scenes for more complex scenarios remains an open challenge. Fig. 9 illustrates several failure cases, such as rendering errors under challenging shading, perception failures, and simulation limitations. More detailed analysis on limitations can be found in the supp." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.738, + 0.627, + 0.754 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.894, + 0.9 + ], + "angle": 0, + "content": "We present PhysGen3D, a framework that transforms a static image into an interactive 3D scene for simulating and rendering future motions based on user input. PhysGen3D integrates modules for 3D world reconstruction, model-based dynamic simulation, and physics-based rendering to generate realistic, controllable videos. By extending the 2D image-to-video paradigm to 3D, PhysGen3D enables more realistic motion and diverse material behaviors. We hope this work inspires future research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "6185" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.091, + 0.237, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.471, + 0.173 + ], + "angle": 0, + "content": "This project is supported by the Intel AI SRS gift, Amazon-Illinois AICE grant, Meta Research Grant, IBM IIDAI Grant, and NSF Awards #2331878, #2340254, #2312102, #2414227, and #2404385. We greatly appreciate the NCSA for providing computing resources." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.2, + 0.174, + 0.216 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.226, + 0.238, + 0.239 + ], + "angle": 0, + "content": "[1] Pika, 2024. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.245, + 0.255 + ], + "angle": 0, + "content": "[2] Runway, 2024. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.26, + 0.271 + ], + "angle": 0, + "content": "[3] Klingai,2024.2,5,6,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[4] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. ACM TOG, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.471, + 0.383 + ], + "angle": 0, + "content": "[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.469, + 0.428 + ], + "angle": 0, + "content": "[6] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.469, + 0.471 + ], + "angle": 0, + "content": "[7] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[8] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.469, + 0.585 + ], + "angle": 0, + "content": "[9] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.471, + 0.642 + ], + "angle": 0, + "content": "[10] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[11] Richard Strong Bowen, Richard Tucker, Ramin Zabih, and Noah Snavely. Dimensions of motion: Monocular prediction through flow subspaces. In 3DV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.471, + 0.741 + ], + "angle": 0, + "content": "[12] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In European Conference on Computer Vision, pages 552-567. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[13] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[14] Xi Chen, Zhiheng Liu, Mengting Chen, Yutong Feng, Yu Liu, Yujun Shen, and Hengshuang Zhao. Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.226, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[16] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7230-7240, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.894, + 0.26 + ], + "angle": 0, + "content": "[17] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. ACM TOG, 2005. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.894, + 0.304 + ], + "angle": 0, + "content": "[18] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM TOG, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[19] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. arXiv preprint arXiv:1910.07192, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[20] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.894, + 0.474 + ], + "angle": 0, + "content": "[21] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.894, + 0.517 + ], + "angle": 0, + "content": "[22] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.894, + 0.587 + ], + "angle": 0, + "content": "[23] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[24] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.894, + 0.701 + ], + "angle": 0, + "content": "[25] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.894, + 0.758 + ], + "angle": 0, + "content": "[26] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation. arXiv preprint arXiv:2407.07667, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.894, + 0.786 + ], + "angle": 0, + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[28] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[29] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In CVPR, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "6186" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[30] Yuanming Hu, Yu Fang, Ziheng Ge, Ziyin Qu, Yixin Zhu, Andre Pradhana, and Chenfanfu Jiang. A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[31] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG), 38(6):201, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[32] Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Frédo Durand. Difftaichi: Differentiable programming for physical simulation. *ICLR*, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.471, + 0.331 + ], + "angle": 0, + "content": "[33] Yuanming Hu, Jiafeng Liu, Xuanda Yang, Mingkuan Xu, Ye Kuang, Weiwei Xu, Qiang Dai, William T. Freeman, and Frédo Durand. Quantaichi: A compiler for quantized simulations. ACM Transactions on Graphics (TOG), 40(4), 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[34] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.471, + 0.474 + ], + "angle": 0, + "content": "[35] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, and Delio Vicini. Dr.jit: A just-in-time compiler for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH), 41(4), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.471, + 0.517 + ], + "angle": 0, + "content": "[36] Wei-Cih Jhou and Wen-Huang Cheng. Animating still landscape photographs through cloud motion creation. IEEE Transactions on Multimedia, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.471, + 0.574 + ], + "angle": 0, + "content": "[37] Chenfanfu Jiang, Craig Schroeder, Joseph Teran, Alexey Stomakhin, and Andrew Selle. The material point method for simulating continuum materials. In Acm siggraph 2016 courses, pages 1-52. 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.471, + 0.618 + ], + "angle": 0, + "content": "[38] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on graphics (TOG), 30(6):1-12, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[39] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[41] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[42] Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In ICLR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. arXiv preprint arXiv:2309.07906, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[44] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[45] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10072–10083, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.894, + 0.304 + ], + "angle": 0, + "content": "[46] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[47] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[48] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[49] Xinyu Liu, Jinlong Li, Jin Ma, Huiming Sun, Zhigang Xu, Tianyun Zhang, and Hongkai Yu. Deep transfer learning for intelligent vehicle perception: A survey. Green Energy and Intelligent Transportation, page 100125, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[50] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[51] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[52] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit warping for animation with image sets. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[53] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. In European Conference on Computer Vision, pages 111-128. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[54] OpenAI. Creating video from text. https://openai.com/sora, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[55] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[56] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 98-108, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[57] Haonan Qiu, Zhaoxi Chen, Yuming Jiang, Hang Zhou, Xiangyu Fan, Lei Yang, Wayne Wu, and Ziwei Liu. Relitalk: Relightable talking portrait generation from a single video, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[58] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "6187" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.163 + ], + "angle": 0, + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[60] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[61] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In PACMCGIT, 2000. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[62] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[63] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.471, + 0.433 + ], + "angle": 0, + "content": "[64] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.471, + 0.462 + ], + "angle": 0, + "content": "[65] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.471, + 0.505 + ], + "angle": 0, + "content": "[66] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.507, + 0.471, + 0.575 + ], + "angle": 0, + "content": "[67] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.471, + 0.604 + ], + "angle": 0, + "content": "[68] Martin Szummer and Rosalind W. Picard. Temporal texture modeling. In ICIP, 1996. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[69] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[70] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. arXiv preprint arXiv:2402.01566, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[71] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.87 + ], + "angle": 0, + "content": "[72] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[73] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[74] Li-Yi Wei and Marc Levoy. Fast texture synthesis using tree-structured vector quantization. In PACMCGIT, 2000. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[75] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via llm-agent collaboration. In The First Workshop on Populating Empty Cities—Virtual Humans for Robotics and Autonomous Driving at CVPR 2024, 2nd Round. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.894, + 0.305 + ], + "angle": 0, + "content": "[76] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.894, + 0.376 + ], + "angle": 0, + "content": "[77] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.893, + 0.431 + ], + "angle": 0, + "content": "[78] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.894, + 0.504 + ], + "angle": 0, + "content": "[79] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time interactive realistic and browser-compatible environment from a single video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4588, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[80] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.894, + 0.631 + ], + "angle": 0, + "content": "[81] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.894, + 0.701 + ], + "angle": 0, + "content": "[82] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccrafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.894, + 0.771 + ], + "angle": 0, + "content": "[83] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.893, + 0.815 + ], + "angle": 0, + "content": "[84] Tianfan Xue, Jiajun Wu, Katherine L Bouman, and William T Freeman. Visual dynamics: Stochastic future generation via layered cross convolutional networks. T-PAMI, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.894, + 0.87 + ], + "angle": 0, + "content": "[85] Chenggang Yan, Biyao Shao, Hao Zhao, Ruixin Ning, Yongdong Zhang, and Feng Xu. 3d room layout estimation from a single rgb image. IEEE Transactions on Multimedia, 22(11): 3014-3024, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[86] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. Magicprop: Diffusion-based video edit" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "6188" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "ing via motion-aware appearance propagation. arXiv preprint arXiv:2309.00908, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.163 + ], + "angle": 0, + "content": "[87] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. Entropy, 25 (10):1469, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[88] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.472, + 0.29 + ], + "angle": 0, + "content": "[89] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9(1):1, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.472, + 0.36 + ], + "angle": 0, + "content": "[90] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.472, + 0.431 + ], + "angle": 0, + "content": "[91] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qing, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[92] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.472, + 0.585 + ], + "angle": 0, + "content": "[93] Xianling Zhang, Nathan Tseng, Ameerah Syed, Rohan Bhasin, and Nikita Jaipuria. Simbar: Single image-based scene relighting for effective data augmentation for automated driving vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3718-3728, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.587, + 0.472, + 0.614 + ], + "angle": 0, + "content": "[94] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.472, + 0.699 + ], + "angle": 0, + "content": "[95] Yupeng Zheng, Chengliang Zhong, Pengfei Li, Huan-ang Gao, Yuhang Zheng, Bu Jin, Ling Wang, Hao Zhao, Guyue Zhou, Qichao Zhang, et al. Steps: Joint self-supervised nighttime image enhancement and depth estimation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4916–4923. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.472, + 0.783 + ], + "angle": 0, + "content": "[96] Chengliang Zhong, Yuhang Zheng, Yupeng Zheng, Hao Zhao, Li Yi, Xiaodong Mu, Ling Wang, Pengfei Li, Guyue Zhou, Chao Yang, et al. 3d implicit transporter for temporally consistent keypoint discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3869-3880, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.785, + 0.472, + 0.854 + ], + "angle": 0, + "content": "[97] Leisheng Zhong, Yu Zhang, Hao Zhao, An Chang, Wenhao Xiang, Shunli Zhang, and Li Zhang. Seeing through the occluders: Robust monocular 6-dof object pose tracking via model-guided video object segmentation. IEEE Robotics and Automation Letters, 5(4):5159-5166, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "6189" + } + ] +] \ No newline at end of file diff --git a/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/1bf0def3-a278-47bf-9ffd-13a93c05e545_origin.pdf b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/1bf0def3-a278-47bf-9ffd-13a93c05e545_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..acea29d026227ca0d702a89fd7ba31fd5240d478 --- /dev/null +++ b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/1bf0def3-a278-47bf-9ffd-13a93c05e545_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47edb282122a9ed5db083455adb04a778596f422f07ac7fcc58ee9f3225c40b6 +size 3634758 diff --git a/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/full.md b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a2548d33c5f00c140eb8192137fcb01f83a35949 --- /dev/null +++ b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/full.md @@ -0,0 +1,300 @@ +# PhysGen3D: Crafting a Miniature Interactive World from a Single Image + +Boyuan Chen $^{1}$ , Hanxiao Jiang $^{2,3}$ , Shaowei Liu $^{2}$ , Saurabh Gupta $^{2}$ , Yunzhu Li $^{3}$ , Hao Zhao $^{1}$ , Shenlong Wang $^{2}$ + +$^{1}$ Tsinghua University, $^{2}$ University of Illinois Urbana-Champaign, $^{3}$ Columbia University + +![](images/49f193c9257b3209f2c5f4245cae84a3c27bd0eaedebae202fd6ad7fe17b923a.jpg) +Input + +![](images/a342b395edbf963f5c4f60ab92bbb3d57b4f2344e9375768511d415d71dde92c.jpg) + +![](images/8e9c23e75751deed27b69f4b1d0178d42f89e4779a6c24d56f9424f9d3243c3a.jpg) +Figure 1. PhysGen3D generates realistic, physically plausible motion from a single image and a text prompt by reasoning about geometry, semantics, and material properties. (a) An apple rolls under the influence of its initial velocity, friction, and shape, producing a natural progression over time. (b) Three animal figures interact dynamically, colliding after being propelled upwards and forwards. (c) A toy potato bounces back with soft-body dynamics in response to an initial downward force, capturing material-specific behaviors. PhysGen3D lets users quickly explore physics-driven object interactions and behaviors in a compact virtual scene generated from a single input image. + +![](images/4400e282a98a46655b270e898c9a7a4cd37d9152d740d431b7c0ae06a7bbac86.jpg) +Generation (left $\rightarrow$ right: time steps) + +![](images/bd7fbdbb261151180d950def447ba3044456996c965db97a12b40b3a01795464.jpg) + +![](images/ef2c847f1dd27ef933fad6291247779d72fe434ea5e5eac39611947b87e2228b.jpg) + +# Abstract + +Envisioning physically plausible outcomes from a single image requires a deep understanding of the world's dynamics. To address this, we introduce PhysGen3D, a novel framework that transforms a single image into an amodal, camera-centric, interactive 3D scene. By combining advanced image-based geometric and semantic understanding with physics-based simulation, PhysGen3D creates an interactive 3D world from a static image, enabling us to "imagine" and simulate future scenarios based on user input. At its core, PhysGen3D estimates 3D shapes, poses, physical and lighting properties of objects, thereby + +capturing essential physical attributes that drive realistic object interactions. This framework allows users to specify precise initial conditions, such as object speed or material properties, for enhanced control over generated video outcomes. We evaluate PhysGen3D's performance against closed-source state-of-the-art (SOTA) image-to-video models, including Pika, Kling, and Gen-3, showing PhysGen3D's capacity to generate videos with realistic physics while offering greater flexibility and fine-grained control. Our results show that PhysGen3D achieves a unique balance of photorealism, physical plausibility, and user-driven interactivity, opening new possibilities for generating dynamic, physics-grounded video from an image. Project page: https://by-luckk.github.io/PhysGen3D. + +# 1. Introduction + +Photographs capture snapshots of our physical world, preserving specific moments in time but leaving out the alternative outcomes that could have unfolded. For instance, looking at a photo, we might wonder, "What if I poke the apples to make them roll across the ground?" or "What if I squeeze the three stuffed animals closer together?" or "What if I drop my cute potato toys onto the floor?" Humans intuitively understand how these scenarios would play out because we have an innate sense of the physical world beyond what we see in a single image. We develop a computational model that can answer such "what-if" questions by generating video outcomes from a single static image. + +A promising approach toward this goal is data-driven image-to-video (I2V) generation [17, 18, 29, 61, 68, 74, 84]. I2V leverages diffusion-based generative models trained on vast datasets of internet images and videos, enabling the production of photorealistic videos with remarkable detail. However, I2V still has limitations in precise control and lacks physical grounding. As a result, users cannot interact freely and accurately to achieve specific physical effects, nor can I2V guarantee physical realism. + +On the other hand, recent research has focused on modeling the physical world from visual inputs to create digital twins, allowing for precise interactions [42, 81, 92, 96]. These approaches can generate virtual scenes with convincing physical interactions, but they typically require complete 3D scans from multi-view images or depth sensors, making them data-intensive. While some methods [43, 48, 66] enable interaction with a single image, they are often limited by physical constraints (e.g., rigid bodies or springs), specific object types (e.g., waterfalls), or a 2D scope. This gap highlights the need for a generic, controllable, physically grounded, and photorealistic approach to generate video from a single image while maintaining physical realism. + +In this work, we introduce PhysGen3D, a novel framework that transforms a single image into an amodal, camera-centric, interactive 3D scene, enabling realistic simulation and rendering. Our approach combines the strengths of image-based geometric and semantic understanding [40, 58, 67, 71, 83] with physics-based simulation [31-33]. At its core, PhysGen3D is a digital twinning method that estimates an object's 3D shape, pose, physical and lighting properties, infers background geometry and appearance, deduces physical characteristics, and performs dimensional analysis—all from a single input image. This task is conventionally challenging due to its inherently ill-posed nature. To tackle this, we leverage various pretrained vision models, integrating their outputs to create an image-centric digital twin. + +For physical simulation, we employ material point methods [30, 37], a robust point-voxel-based framework that models counterfactual physical behaviors of objects in the image. Through precise inference of physical properties, + +simulations in the PhysGen3D environment achieve a high degree of realism and stability. We further enhance realism by applying physics-based rendering, seamlessly integrating dynamic effects back into the original image. PhysGen3D produces results that are not only visually realistic in terms of dynamics and lighting but also highly controllable, allowing users to specify initial conditions like speed and material properties. Due to the use of large pretrained models, our pipeline operates effectively without task-specific training. + +Our experiments, based on a carefully designed and rigorous user study, demonstrate that, compared to closed-source state-of-the-art video AIGC models such as Pika, Kling, and Gen-3, our framework provides significantly more flexible control over object motions, generates videos that better align with user intentions, and achieves superior physical realism—all while maintaining comparable rendering quality. + +# 2. Related work + +Single-view 3D reconstruction Great progress has been made in object-centric single view 3D reconstructions [45, 46, 62, 83]. However, applying these techniques to single-view 3D scene reconstruction becomes more challenging. Most existing works only focus on one part of 3D scene understanding. Geometric approaches reconstruct the 3D scene holistically but neglect individual object understanding, or they focus solely on foreground objects without considering the interaction of objects with the complex environment. Additionally, works have been done for scene relighting [49, 93, 95] and segmentation [40, 47, 58, 85, 97], but none of these provide a full 3D understanding within the image. In our work, we also perform perception reasoning for object materials, backgrounds, rendering properties, and physical stability. + +Controllable video generation. Video generation has made significant progress in recent years [5, 9, 10, 15, 21, 23, 25, 28, 41, 54, 63, 69, 73, 82, 90, 91]. The state-of-the-art framework [1, 3] can generate photorealistic and coherent videos from text instructions using diffusion models [27, 55, 59, 64, 65]. Controllable video generation is achieved through leveraging pre-trained video generation models with conditioning information, including but not limited to depth maps [13, 86], linear translation [39], layouts [44, 70, 87], and multiple combinations [50]. However, most existing generation methods implicitly generate image space dynamics, which might lead to unrealistic, hallucinated motions. In contrast, our method explicitly controls the motion and interaction with simulation, allowing us to create more sophisticated effects without the need for extensive training data. Our approach is training-free and generalizable to all objects in the world. + +![](images/fd7253eba88cec6ff23009b105abf5f7acd3dfa849c9a4e32bc4faa5e1ba38c8.jpg) +Figure 2. Method Overview. PhysGen3D's framework consists of three modules: a) 3D world creation, which infers geometry, semantics, rendering and physical parameters from the input image; b) dynamics simulation using Taichi-Elements for particle-based physics; and c) physics-based rendering with a two-pass shadow mapping technique. + +Controllable image animation Image-based animation aims to animate objects that appear in images. Numerous works [17, 18, 29, 61, 68, 74, 84] have focused on this task. To improve quality, recent research has adopted data-driven solutions, training temporal neural networks to directly predict subsequent video frames [6, 14, 19, 24, 29, 72], or incorporating physical heuristics [17, 36]. Recently, there has been increased focus on interactive [7, 8, 43] and controllable [4, 18] image-to-video synthesis. Additional priors such as motion fields [19, 29, 51, 52, 66, 84], optical flow [11, 22, 94], 3D geometry information [57, 76], and user annotations [43] have also been introduced. The closest work to ours is Liu et al. [48], which focuses on image understanding and uses an existing 2D rigid simulator to generate 2D animations. However, their approach is limited to 2D dynamics, does not account for real-world 3D physics, and is restricted to rigid bodies. In contrast, our method provides more realistic and flexible control over 3D motions and object materials and extends beyond rigid-body physics. + +# 3. Approach + +Our goal is to reconstruct an interactive, camera-centric miniature world from a single input image. We aim to control object materials, dynamics, and motions to simulate diverse, photorealistic, and physically plausible videos from user input. The key challenge lies in partial observations and the ill-posed nature of physical reasoning without observed dynamics. To address this, we propose a holistic reconstruction method leveraging large pretrained visual models to jointly infer geometry, dynamic materials, lighting, and + +PBR materials from a single image (Sec. 3.1). The reconstructed scene is input into a material-point-method (MPM) simulator to generate realistic physics phenomena (Sec. 3.2). Finally, we render dynamic object behaviors based on the simulation and reintegrate them into the scene, producing realistic videos with accurate appearance and motion (Sec. 3.3). Fig. 2 depicts our framework. + +# 3.1. Interactive 3D World from a Single Image + +A full amodal reconstruction of the 3D world depicted in the image is critical for the next step in simulation. An ideal reconstruction should capture a comprehensive understanding of objects' relationships, geometry, appearance, material, and physical properties. However, obtaining this understanding from a single image is highly ill-posed. Our key idea is to leverage priors from pretrained vision foundation models to help infer these properties, as shown in Fig. 2. + +Segmentation. We leverage vision foundation models to recognize object categories and segment object instances. Specifically, we use GPT-4o [89] to identify the foreground object categories and use Grounded-SAM [40, 47, 58] to further detect and segment each individual instance $\{o^i\in \mathbb{R}^{W\times H\times 3}\}_{i}^{N}$ , where $o^i$ is the image of i-th object. + +Mesh Generation. Unlike previous work [48], which used only 2D rigid-body physics, we extend to general-purpose 3D simulation. This requires a complete 3D representation of foreground objects. We adopt InstantMesh [83], which uses Zero123++ [46, 62] to synthesize multi-view images from the segmented object image $o^i$ and uses these images to reconstruct the 3D mesh $\mathcal{O}$ of the objects. For multiple object occlusions, we adopt the iterative inpainting + +strategy to extract the 3D mesh of the objects (see supp for more details). + +Background Handling. The background serves important roles in both dynamic simulation and rendering. In simulation, background geometry acts as a support and collider, where accuracy ensures realistic object-scene interactions. In rendering, it serves as a static backdrop while the foreground objects move and helps simulate realistic global illumination effects like cast shadows. For simulation purposes, we use Dust3r [71] to estimate iamage depth. The output depth map $z \in \mathbb{R}^{W \times H}$ is unprojected to the 3D world as a 3D point cloud $\mathcal{P}$ , and Bilateral Normal Integration [12] is applied to generate a smooth surface $S$ serving as the collider. For rendering, to fill background regions occluded by moving objects, we use the LaMA inpainting model[67]. This model generates a complete background after masking out all objects and their shadows. + +Object Pose and Scale Estimation. The generative 3D reconstruction step provides a complete mesh with a normalized scale in an object-centric coordinate system, but it does not infer the object's location and scale in the camera coordinate. To ensure coherence with the input image, it is necessary to accurately place the 3D meshes $\mathcal{O}$ into the 3D scene $\mathcal{P}$ with the correct scale and 6DoF pose. This registration task is challenging, as the generated object mesh may not perfectly match the real-world object in the image. + +To address this challenge, we designed a multi-stage coarse-to-fine alignment strategy. In the coarse stage, we perform 2D-3D feature point matching. Firstly, we render multiple images from viewpoints uniformly distributed over a unit sphere surrounding the object. For each rendered image, we match its feature points with the original object image $o^i$ using SuperGlue [60], and the viewpoint with the most matches is selected. Matched pairs $(p_i^N \in \mathbb{R}^2, p_i^N \in \mathbb{R}^2)$ are then projected back into 3D points $(P_i^N \in \mathbb{R}^3, P_i^N \in \mathbb{R}^3)$ in object and camera coordinate, respectively. Perspective-n-Point (PnP) algorithm [20] is applied between $P_i^N$ and $p_i^N$ in image $o^i$ to estimate the 6DoF pose with scale ambiguity. Then we adjust the scale and translation factor simultaneously to minimize the L2 loss $\sum_{i}^{N} = ||P_i^N - P_i'^N||_2$ , without altering its projection. + +In the fine alignment stage, we render the mask and depth on the image plane using the current estimation through a differentiable renderer and jointly minimize two losses: $\mathcal{L} = \mathcal{L}_{\mathrm{dice}} + \mathcal{L}_{\mathrm{depth}}$ . Here, $\mathcal{L}_{\mathrm{dice}} = 1 - \frac{2\times|M_A\cap M_B|}{|M_A| + |M_B|}$ measures the discrepancy between the rendered mask $M_{A}$ and the observed mask $M_B$ from Grounded-SAM. The depth consistency loss is $\mathcal{L}_{\mathrm{depth}} = \frac{\mathrm{mse}(M_B*z_A - M_B*z_B)}{|M_B|}$ , where $z_{A}$ and $z_{B}$ are the rendered and Dust3r-predicted depths. This joint optimization ensures consistency between the estimated mesh in the camera coordinate and the point cloud from Dust3r, while maximizing the alignment between the 3D pose and observed object mask, ensuring accurate simu + +lation and rendering alignment. + +Appearance Optimization. The texture of the generated 3D meshes may differ from the input image. To enhance rendering quality, we use the inverse rendering pipeline in Mitsuba3 [35] to estimate material properties. Lighting parameters are estimated with DiffusionLight [56], while object PBR materials (albedo, roughness, and metallic) are optimized via differentiable rendering. To handle unknown back views and reduce complexity, we assume uniform roughness and metallic values per object and apply tone mapping for albedo optimization: $y(x) = ax^3 + bx^2 + cx + d$ , where $y(0) = 0$ , $y(1) = 1$ . With reconstructed lighting and ground surface, the optimized materials improve asset appearance, capturing realistic object-surface interactions during rendering. + +Physics Reasoning. Accurately simulating real-world dynamics requires estimating physical parameters. We focus on two aspects: 1) Following [79], we use GPT-4o to query each object's elasticity and density, and the friction coefficient for the surface $S$ . 2) We ensure reconstructed 3D assets match real-world proportions, as depth inaccuracies from Dust3r can cause unrealistic behaviors. To address this, we estimate a scale factor $k$ by comparing asset size with typical real-world sizes from GPT-4o and use $k$ for dimensionless scaling of gravity and velocity-related parameters. + +# 3.2. Dynamics Simulation + +Given the 3D assets with reasoned physical properties and the scale factor, we use the physics engine Taichi Elements [31-33], based on the Material Point Method (MPM) [37], as our simulator to support a variety of different materials, including but not limited to rigid, soft, and granular. + +Particle representation. The simulator is based on a particle-based representation. To convert our 3D assets into a simulatable particle-based representation, we apply floating points removing, internal filling and voxel downsampling. We use downsampling to handle uneven particle distributions, where downsample rate is adjusted according to the grid size of the simulator. For the convenience of rendering, we prioritize surface points. + +Physical parameters. To enhance stability, we apply the scale factor $k$ to physical parameters in simulator instead of scaling the assets to real size. For example, the gravitational acceleration is multiplied by $k$ , making to $k * 9.8$ (nondimensionalized here). With this tuning, the motion of falling or collapse remains realistic for all scales. + +External disturbance. For each object, we set a different initial velocity based on the user input to make the object move as specified by the user. + +Other Visual Effects. Besides real world physics simulation, our pipeline allows special effects like collapsing and melting. To simulate different materials (rigid, soft, or granular), we can easily change the material type to modify + +![](images/0ee7dc5270c6c9c2dba35d2ac98e011544b60151acdb5dead4e9656f78f5d80d.jpg) +Figure 3. Video generation results. Left: Input initial frame. Right: Generated future frames. We apply an initial velocity to each movable object and use the physically grounded parameters outlined in 3.2 to generate physically plausible results. + +the physical properties of an object. This approach provides more flexibility for the user to edit the scene. + +# 3.3. Physics-Based Rendering + +After dynamics simulation, we obtain object point trajectories and apply motion interpolation to compute vertex motions, deforming the mesh accordingly. With optimized PBR materials, we use Mitsuba3 for Physically-Based Rendering under environment lighting. Following prior insertion rendering work [16, 38, 75, 88], we avoid converting the entire static background into the rendering pipeline. Instead, we build a 3D shadow catcher surface from the background depth. During rendering, no texture is applied to the background; two-pass shadow mapping extracts shadows and global illumination effects. The foreground objects and shadows are then composited onto the inpainted background to produce the final video with realistic lighting. + +# 4. Experimental Results + +# 4.1. Setups + +The test images come from a diverse set of our own photography, online collections, and generative models. Our pipeline is primarily designed for object-centric scenes with one or a few objects. Due to our limitations (Sec. 5), we excluded those with excessive objects, heavy occlusion between objects, or highly uneven surfaces. + +Post-processing. We used VEnhancer [26] as an optional post-processing module, which takes the produced video and a text prompt to perform enhancement. As shown in Table 2, while it restores some details, it also introduces extra hallucinations. + +Baselines. Our method is one of the first of its kind, as existing model-simulate approaches require multi-view im + +ages [78, 80, 92] or specific settings [43, 48, 66]. Therefore, we evaluate ours against Image-to-Video (I2V) models: two open-source motion controller models DragAnything[77], MOFA-Video[53] and three state-of-the-art (SOTA) commercial models Kling 1.0 [3], Gen-3 [2], and Pika 1.5 [1]. We manually set correct motion trajectories and select applied regions for DragAnything, MOFA-Video and Kling to provide privileged motion guidance. We give text descriptions to Pika 1.5 and Gen-3, as they lack direct motion control capabilities. Additionally, Pika 1.5 offers "Pikafect" effects, such as "Melt it" and "Deflate it." + +# 4.2. Results + +Our system generates a miniature interactive world from a single image, enabling the simulation of various phenomena. Fig. 3 presents videos generated from different types of images. These images encompass single or multiple objects and various physical materials (i.e., rigid or soft). We also show applications like dynamic changes, object editing, and dense 3D tracking, illustrating the adaptability and creative potential of our approach for generating customized, interactive video content. + +Comparisons. We compare the results in two dimensions: motion control alone and physical materials. Fig. 4 shows that our system produces more physically plausible and controllable videos compared to SOTA I2V models. Despite prompt tuning, learning-based models often hallucinate, failing to adhere to physical laws or user intent. For example, in the toy dog case, we manually adjust the material and accurately simulate a collapse, whereas other models fall short. Similarly, for the book case, our results are the most physically realistic. + +Dynamics. Fig. 5 demonstrates the varying dynamics generated from the same input image, highlighting the high + +![](images/7d701def37dd7cf51e8cb9c57d4b5ee8c67cb19161258ad41486b2731a9aef22.jpg) +Figure 4. Qualitative comparison. We compare videos generated from our framework with three state-of-the-art I2V models: Gen-3 [2], Pika [1] and Kling [3]. We carefully designed the prompt to describe the motion outcome, and uses motion brush to control Kling. Our framework employs initial velocity control. Results show that our method can follow text instructions while maintaining plausible physics. + +controllability of our method over physical parameters and motion trajectories. In the three rows on the left, we set different elasticities for the two objects, while keeping their initial positions and velocities the same. In the three rows on the right, we alter velocity directions of the objects and + +keep physical parameters the same. + +Editing. Our method enables modifications to videos by removing, adding, or replacing objects, as illustrated in Fig. 6. The generated 3D assets can be easily manipulated, allowing for diverse video edits. + +![](images/7f40b3e95e99456ead552db3794caf0cac6a6c2f80a03035004970785dc473f0.jpg) +Figure 5. Dynamics Effects. We can generate various dynamics from the same input image. The left three columns share the same initial positions and velocities, but are in different materials. The right three columns have the same material, but defer in velocity directions. This showcases the potential of our method for generating diverse physical scenarios. + +![](images/bd89ae7d75def4d460580a26442d603f49deb985da903ce1225730fa86ed802e.jpg) +Figure 6. Video edition. As illustrated, our method supports video editing. In the top row, we remove the chair and allow the toy to fall from a static position. In the bottom two rows, we exchange one object between two scenes while keeping the other unchanged. This demonstrates the great flexibility of our video generation approach. + +Tracking. Our framework uses an explicit 3D representation and works with a particle-based physics simulator. This allows our method to easily create videos with detailed 3D tracking results. Fig. 7 showcases two examples, demonstrating the accuracy and reliability of the tracking in different scenarios. + +Ablation. Each step we designed in perception, simulation and rendering is intended to mimic the real world. Fig. 8 shows the results of ablation study involving position + +optimization and inverse texture. Without pose optimization using differentiable rendering, the two objects are roughly at the initial place shown in input image, but cannot fully replicate the scene. Without inverse texture, the generated object mesh does not match the input in terms of color tone and brightness. We also conducted ablations on point sampling. However, without sampling, a large number of points become crowded in several MPM simulation grids, causing the simulator to crash and fail to produce final outputs. + +![](images/5ccd443d8f46e08f3d34a1c25be6065d405314fa428a6c77905917dbc08af94d.jpg) +Figure 7. Dense 3D Tracking. Our method can naturally generate videos with dense 3D tracking results. Here we show tracking results for collapse and bounce cases. + +![](images/3d6fe1b6722bb10feb967a35d7ede758d22f7ad745f8ed82126f3daeaf8f91a1.jpg) + +![](images/26c42bb64e6966cb03323f952e349a13dc34e0bffb5252a28abcfb91e30d45fc.jpg) + +![](images/87f9c188d3cb717ded077e3d3c21d92b24ddb579536ac105935fa56acff25b97.jpg) +Figure 8. Ablation study. Without pose optimization, the two objects cannot fully replicate the scene. Without inverse texture, the object mesh does not match the input color. + +Table 1. Human Evaluation Results. The three criteria are: Physical Realism, Photorealism and Semantic Consistency + +
MethodsPhysRealPhotoRealAlign
Kling 1.0 [3]2.8113.5662.467
Runway Gen-3 [2]2.2833.5821.886
Pika 1.5 [1]2.4123.3142.016
Ours3.7073.4113.866
+ +Table 2. VBenchmark Scores and GPT-4o Evaluation Results. Motion and Imaging refers to Motion Smoothness and Imaging Quality scores in VBenchmark. The three criteria on the right are the same as Table 1, with results given by GPT-4o. + +
MethodsMotion↑Imaging↑PhysReal↑PhotoReal↑Align↑
Kling 1.00.9960.6710.5630.8740.596
Runway Gen-30.9910.7230.1410.8960.144
Pika 1.50.9940.6710.5440.8630.563
MOFA-Video0.9940.6340.3840.7640.304
DragAnything0.9850.4280.6450.7560.380
Ours0.9950.6660.7520.8670.796
Ours+VEnhancer0.9940.6770.7660.8800.745
+ +# 4.3.Quantitative Comparison + +To assess the quality of the generated videos, we performed human evaluation, GPT-4o evaluation and provided two VBenchmark[34] evaluation scores. + +Benchmarks. We designed three criteria for human and GPT-4o evaluation. (1) Physical Realism (PhysReal) measures how realistically the video follows the physical rules and whether the video represents real physical properties like elasticity and friction. (2) Photorealism (Photoreal) assesses the overall visual quality of the video, including the visual artifacts, discontinuities, and how accurately the video replicates details of light, shadow, texture, and materials. (3) Semantic Consistency (Align) evaluates how well the content of the generated video aligns with the intended meaning of the text prompt. We also chose two Quality Scores in VBenchmark: Motion Smoothness and Imaging Quality. + +Details in evaluation. Following a similar methodology + +![](images/515f589cee55e3624b71f3fa169f8d8ff40886f9004a775088e0ac648b8552ec.jpg) +Figure 9. Limitations. The first set of two images show rendering errors: extreme lighting and heavy shading cause incorrect painting and artifacts from the pillow penetrating the ground. The second set reveals simulation limitations: the MPM method struggles with fine details, such as the thin kettle handle, leading to failures. + +to [48], we designed a questionnaire with 27 videos covering various scenes, motion conditions, and effects. Each video includes an input image, a motion prompt, and outputs from three SOTA commercial baselines and ours, shown in random order. 31 participants rated three criteria on a five-point scale from strongly disagree (1) to strongly agree (5). We also evaluated all five baselines and our diffusion-enhanced version using GPT-4o and VBenchmark. GPT-4o assessed videos on the same criteria based on 10 evenly sampled frames, with the input image and prompt provided. We used GPT version gpt-4o-2024-08-06, with detailed prompts in the supp. + +Result analysis. The results in Table 1 demonstrate our ability to generate physically accurate and controllable videos. In physical realism (PhysReal) and semantic consistency (Align), our method achieves the highest scores and outperforms all the commercial models by a large margin. Table 2 presents the results of GPT-4o and VBench evaluation. GPT score is aligned with human evaluation, and we also beat open-sourced in VBench. Among the baseline models, Kling 1.0 performs the best overall, likely due to its use of the "motion brush", which specifies motion trajectories. + +# 5. Limitations + +Our single-image-based interactive miniature world is designed for object-centric scenes with simple spatial geometry and inter-object relationships. Reconstructing complete scenes for more complex scenarios remains an open challenge. Fig. 9 illustrates several failure cases, such as rendering errors under challenging shading, perception failures, and simulation limitations. More detailed analysis on limitations can be found in the supp. + +# 6. Conclusions + +We present PhysGen3D, a framework that transforms a static image into an interactive 3D scene for simulating and rendering future motions based on user input. PhysGen3D integrates modules for 3D world reconstruction, model-based dynamic simulation, and physics-based rendering to generate realistic, controllable videos. By extending the 2D image-to-video paradigm to 3D, PhysGen3D enables more realistic motion and diverse material behaviors. We hope this work inspires future research. + +# Acknowledgement + +This project is supported by the Intel AI SRS gift, Amazon-Illinois AICE grant, Meta Research Grant, IBM IIDAI Grant, and NSF Awards #2331878, #2340254, #2312102, #2414227, and #2404385. We greatly appreciate the NCSA for providing computing resources. + +# References + +[1] Pika, 2024. 2, 5, 6, 8 +[2] Runway, 2024. 5, 6, 8 +[3] Klingai,2024.2,5,6,8 +[4] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. ACM TOG, 2020. 3 +[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2 +[6] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In CVPR, 2021. 3 +[7] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In ICCV, 2021. 3 +[8] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In CVPR, 2021. 3 +[9] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2 +[10] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2 +[11] Richard Strong Bowen, Richard Tucker, Ramin Zabih, and Noah Snavely. Dimensions of motion: Monocular prediction through flow subspaces. In 3DV, 2022. 3 +[12] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In European Conference on Computer Vision, pages 552-567. Springer, 2022. 4 +[13] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2 +[14] Xi Chen, Zhiheng Liu, Mengting Chen, Yutong Feng, Yu Liu, Yujun Shen, and Hengshuang Zhao. Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928, 2023. 3 +[15] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion + +model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 2 +[16] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7230-7240, 2021. 5 +[17] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. ACM TOG, 2005. 2, 3 +[18] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM TOG, 2015. 2, 3 +[19] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. arXiv preprint arXiv:1910.07192, 2019. 3 +[20] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 4 +[21] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 2 +[22] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In ICLR, 2024. 3 +[23] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2 +[24] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning, 2023. 3 +[25] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2 +[26] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation. arXiv preprint arXiv:2407.07667, 2024. 5 +[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 2 +[28] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[29] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In CVPR, 2021. 2, 3 + +[30] Yuanming Hu, Yu Fang, Ziheng Ge, Ziyin Qu, Yixin Zhu, Andre Pradhana, and Chenfanfu Jiang. A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG, 2018. 2 +[31] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG), 38(6):201, 2019. 2, 4 +[32] Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Frédo Durand. Difftaichi: Differentiable programming for physical simulation. *ICLR*, 2020. +[33] Yuanming Hu, Jiafeng Liu, Xuanda Yang, Mingkuan Xu, Ye Kuang, Weiwei Xu, Qiang Dai, William T. Freeman, and Frédo Durand. Quantaichi: A compiler for quantized simulations. ACM Transactions on Graphics (TOG), 40(4), 2021. 2, 4 +[34] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 8 +[35] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, and Delio Vicini. Dr.jit: A just-in-time compiler for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH), 41(4), 2022. 4 +[36] Wei-Cih Jhou and Wen-Huang Cheng. Animating still landscape photographs through cloud motion creation. IEEE Transactions on Multimedia, 2015. 3 +[37] Chenfanfu Jiang, Craig Schroeder, Joseph Teran, Alexey Stomakhin, and Andrew Selle. The material point method for simulating continuum materials. In Acm siggraph 2016 courses, pages 1-52. 2016. 2, 4 +[38] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on graphics (TOG), 30(6):1-12, 2011. 5 +[39] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2 +[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 3 +[41] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2 +[42] Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In ICLR, 2023. 2 + +[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. arXiv preprint arXiv:2309.07906, 2023. 2, 3, 5 +[44] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 2 +[45] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10072–10083, 2024. 2 +[46] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2, 3 +[47] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3 +[48] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 2, 3, 5, 8 +[49] Xinyu Liu, Jinlong Li, Jin Ma, Huiming Sun, Zhigang Xu, Tianyun Zhang, and Hongkai Yu. Deep transfer learning for intelligent vehicle perception: A survey. Green Energy and Intelligent Transportation, page 100125, 2023. 2 +[50] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 2 +[51] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In CVPR, 2022. 3 +[52] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit warping for animation with image sets. 2022. 3 +[53] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. In European Conference on Computer Vision, pages 111-128. Springer, 2024. 5 +[54] OpenAI. Creating video from text. https://openai.com/sora, 2024.2 +[55] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 2 +[56] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 98-108, 2024. 4 +[57] Haonan Qiu, Zhaoxi Chen, Yuming Jiang, Hang Zhou, Xiangyu Fan, Lei Yang, Wayne Wu, and Ziwei Liu. Relitalk: Relightable talking portrait generation from a single video, 2023. 3 +[58] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang + +Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 2, 3 +[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[60] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 4 +[61] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In PACMCGIT, 2000. 2, 3 +[62] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 2, 3 +[63] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[64] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. 2015. 2 +[65] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. 2019. 2 +[66] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia, 2022. 2, 3, 5 +[67] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 2, 4 +[68] Martin Szummer and Rosalind W. Picard. Temporal texture modeling. In ICIP, 1996. 2, 3 +[69] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2 +[70] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. arXiv preprint arXiv:2402.01566, 2024. 2 +[71] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 4 +[72] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3 +[73] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo + +Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2 +[74] Li-Yi Wei and Marc Levoy. Fast texture synthesis using tree-structured vector quantization. In PACMCGIT, 2000. 2, 3 +[75] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via llm-agent collaboration. In The First Workshop on Populating Empty Cities—Virtual Humans for Robotics and Autonomous Driving at CVPR 2024, 2nd Round. 5 +[76] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In CVPR, 2019. 3 +[77] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 5 +[78] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In CVPR, 2024. 5 +[79] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time interactive realistic and browser-compatible environment from a single video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4588, 2024. 4 +[80] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In CVPR, 2024. 5 +[81] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 2 +[82] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccrafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2 +[83] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 3 +[84] Tianfan Xue, Jiajun Wu, Katherine L Bouman, and William T Freeman. Visual dynamics: Stochastic future generation via layered cross convolutional networks. T-PAMI, 2018. 2, 3 +[85] Chenggang Yan, Biyao Shao, Hao Zhao, Ruixin Ning, Yongdong Zhang, and Feng Xu. 3d room layout estimation from a single rgb image. IEEE Transactions on Multimedia, 22(11): 3014-3024, 2020. 2 +[86] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. Magicprop: Diffusion-based video edit + +ing via motion-aware appearance propagation. arXiv preprint arXiv:2309.00908, 2023. 2 +[87] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. Entropy, 25 (10):1469, 2023. 2 +[88] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 5 +[89] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9(1):1, 2023. 3 +[90] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2 +[91] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qing, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2 +[92] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2025. 2, 5 +[93] Xianling Zhang, Nathan Tseng, Ameerah Syed, Rohan Bhasin, and Nikita Jaipuria. Simbar: Single image-based scene relighting for effective data augmentation for automated driving vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3718-3728, 2022. 2 +[94] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In CVPR, 2022. 3 +[95] Yupeng Zheng, Chengliang Zhong, Pengfei Li, Huan-ang Gao, Yuhang Zheng, Bu Jin, Ling Wang, Hao Zhao, Guyue Zhou, Qichao Zhang, et al. Steps: Joint self-supervised nighttime image enhancement and depth estimation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4916–4923. IEEE, 2023. 2 +[96] Chengliang Zhong, Yuhang Zheng, Yupeng Zheng, Hao Zhao, Li Yi, Xiaodong Mu, Ling Wang, Pengfei Li, Guyue Zhou, Chao Yang, et al. 3d implicit transporter for temporally consistent keypoint discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3869-3880, 2023. 2 +[97] Leisheng Zhong, Yu Zhang, Hao Zhao, An Chang, Wenhao Xiang, Shunli Zhang, and Li Zhang. Seeing through the occluders: Robust monocular 6-dof object pose tracking via model-guided video object segmentation. IEEE Robotics and Automation Letters, 5(4):5159-5166, 2020. 2 \ No newline at end of file diff --git a/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/images.zip b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aaeb78c129e970adadc458d4fea7d0b3a2f2a712 --- /dev/null +++ b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3045fac4b9046ff863ae14694d8c8bb15c08fbb33956f315fc558670d0e7fe1b +size 889519 diff --git a/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/layout.json b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..48fdd5287d2199d745f6fd1f80a068e298223657 --- /dev/null +++ b/2025/PhysGen3D_ Crafting a Miniature Interactive World from a Single Image/layout.json @@ -0,0 +1,8455 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 523, + 121 + ], + "type": "text", + "content": "PhysGen3D: Crafting a Miniature Interactive World from a Single Image" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "spans": [ + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "text", + "content": "Boyuan Chen" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "text", + "content": ", Hanxiao Jiang" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "text", + "content": ", Shaowei Liu" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "text", + "content": ", Saurabh Gupta" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "text", + "content": ", Yunzhu Li" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "text", + "content": ", Hao Zhao" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "text", + "content": ", Shenlong Wang" + }, + { + "bbox": [ + 153, + 142, + 438, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 172, + 455, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 172, + 455, + 185 + ], + "spans": [ + { + "bbox": [ + 136, + 172, + 455, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 172, + 455, + 185 + ], + "type": "text", + "content": "Tsinghua University, " + }, + { + "bbox": [ + 136, + 172, + 455, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 172, + 455, + 185 + ], + "type": "text", + "content": "University of Illinois Urbana-Champaign, " + }, + { + "bbox": [ + 136, + 172, + 455, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 172, + 455, + 185 + ], + "type": "text", + "content": "Columbia University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 220, + 132, + 304 + ], + "blocks": [ + { + "bbox": [ + 78, + 209, + 100, + 219 + ], + "lines": [ + { + "bbox": [ + 78, + 209, + 100, + 219 + ], + "spans": [ + { + "bbox": [ + 78, + 209, + 100, + 219 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 220, + 132, + 304 + ], + "lines": [ + { + "bbox": [ + 49, + 220, + 132, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 132, + 304 + ], + "type": "image", + "image_path": "49f193c9257b3209f2c5f4245cae84a3c27bd0eaedebae202fd6ad7fe17b923a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 304, + 132, + 388 + ], + "blocks": [ + { + "bbox": [ + 49, + 304, + 132, + 388 + ], + "lines": [ + { + "bbox": [ + 49, + 304, + 132, + 388 + ], + "spans": [ + { + "bbox": [ + 49, + 304, + 132, + 388 + ], + "type": "image", + "image_path": "a342b395edbf963f5c4f60ab92bbb3d57b4f2344e9375768511d415d71dde92c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 65, + 389, + 116, + 464 + ], + "blocks": [ + { + "bbox": [ + 65, + 389, + 116, + 464 + ], + "lines": [ + { + "bbox": [ + 65, + 389, + 116, + 464 + ], + "spans": [ + { + "bbox": [ + 65, + 389, + 116, + 464 + ], + "type": "image", + "image_path": "8e9c23e75751deed27b69f4b1d0178d42f89e4779a6c24d56f9424f9d3243c3a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 480, + 547, + 537 + ], + "lines": [ + { + "bbox": [ + 45, + 480, + 547, + 537 + ], + "spans": [ + { + "bbox": [ + 45, + 480, + 547, + 537 + ], + "type": "text", + "content": "Figure 1. PhysGen3D generates realistic, physically plausible motion from a single image and a text prompt by reasoning about geometry, semantics, and material properties. (a) An apple rolls under the influence of its initial velocity, friction, and shape, producing a natural progression over time. (b) Three animal figures interact dynamically, colliding after being propelled upwards and forwards. (c) A toy potato bounces back with soft-body dynamics in response to an initial downward force, capturing material-specific behaviors. PhysGen3D lets users quickly explore physics-driven object interactions and behaviors in a compact virtual scene generated from a single input image." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 135, + 220, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 267, + 209, + 402, + 219 + ], + "lines": [ + { + "bbox": [ + 267, + 209, + 402, + 219 + ], + "spans": [ + { + "bbox": [ + 267, + 209, + 402, + 219 + ], + "type": "text", + "content": "Generation (left " + }, + { + "bbox": [ + 267, + 209, + 402, + 219 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 267, + 209, + 402, + 219 + ], + "type": "text", + "content": " right: time steps)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 220, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 135, + 220, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 135, + 220, + 545, + 304 + ], + "type": "image", + "image_path": "4400e282a98a46655b270e898c9a7a4cd37d9152d740d431b7c0ae06a7bbac86.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 135, + 304, + 545, + 388 + ], + "blocks": [ + { + "bbox": [ + 135, + 304, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 135, + 304, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 135, + 304, + 545, + 388 + ], + "type": "image", + "image_path": "bd7fbdbb261151180d950def447ba3044456996c965db97a12b40b3a01795464.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 135, + 389, + 545, + 472 + ], + "blocks": [ + { + "bbox": [ + 135, + 389, + 545, + 472 + ], + "lines": [ + { + "bbox": [ + 135, + 389, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 135, + 389, + 545, + 472 + ], + "type": "image", + "image_path": "ef2c847f1dd27ef933fad6291247779d72fe434ea5e5eac39611947b87e2228b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 549, + 192, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 549, + 192, + 562 + ], + "spans": [ + { + "bbox": [ + 143, + 549, + 192, + 562 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 583, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 289, + 715 + ], + "type": "text", + "content": "Envisioning physically plausible outcomes from a single image requires a deep understanding of the world's dynamics. To address this, we introduce PhysGen3D, a novel framework that transforms a single image into an amodal, camera-centric, interactive 3D scene. By combining advanced image-based geometric and semantic understanding with physics-based simulation, PhysGen3D creates an interactive 3D world from a static image, enabling us to \"imagine\" and simulate future scenarios based on user input. At its core, PhysGen3D estimates 3D shapes, poses, physical and lighting properties of objects, thereby" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 550, + 548, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 548, + 717 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 548, + 717 + ], + "type": "text", + "content": "capturing essential physical attributes that drive realistic object interactions. This framework allows users to specify precise initial conditions, such as object speed or material properties, for enhanced control over generated video outcomes. We evaluate PhysGen3D's performance against closed-source state-of-the-art (SOTA) image-to-video models, including Pika, Kling, and Gen-3, showing PhysGen3D's capacity to generate videos with realistic physics while offering greater flexibility and fine-grained control. Our results show that PhysGen3D achieves a unique balance of photorealism, physical plausibility, and user-driven interactivity, opening new possibilities for generating dynamic, physics-grounded video from an image. Project page: https://by-luckk.github.io/PhysGen3D." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "6178" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 235 + ], + "type": "text", + "content": "Photographs capture snapshots of our physical world, preserving specific moments in time but leaving out the alternative outcomes that could have unfolded. For instance, looking at a photo, we might wonder, \"What if I poke the apples to make them roll across the ground?\" or \"What if I squeeze the three stuffed animals closer together?\" or \"What if I drop my cute potato toys onto the floor?\" Humans intuitively understand how these scenarios would play out because we have an innate sense of the physical world beyond what we see in a single image. We develop a computational model that can answer such \"what-if\" questions by generating video outcomes from a single static image." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 235, + 288, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 235, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 288, + 342 + ], + "type": "text", + "content": "A promising approach toward this goal is data-driven image-to-video (I2V) generation [17, 18, 29, 61, 68, 74, 84]. I2V leverages diffusion-based generative models trained on vast datasets of internet images and videos, enabling the production of photorealistic videos with remarkable detail. However, I2V still has limitations in precise control and lacks physical grounding. As a result, users cannot interact freely and accurately to achieve specific physical effects, nor can I2V guarantee physical realism." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 342, + 288, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 288, + 498 + ], + "type": "text", + "content": "On the other hand, recent research has focused on modeling the physical world from visual inputs to create digital twins, allowing for precise interactions [42, 81, 92, 96]. These approaches can generate virtual scenes with convincing physical interactions, but they typically require complete 3D scans from multi-view images or depth sensors, making them data-intensive. While some methods [43, 48, 66] enable interaction with a single image, they are often limited by physical constraints (e.g., rigid bodies or springs), specific object types (e.g., waterfalls), or a 2D scope. This gap highlights the need for a generic, controllable, physically grounded, and photorealistic approach to generate video from a single image while maintaining physical realism." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 287, + 665 + ], + "type": "text", + "content": "In this work, we introduce PhysGen3D, a novel framework that transforms a single image into an amodal, camera-centric, interactive 3D scene, enabling realistic simulation and rendering. Our approach combines the strengths of image-based geometric and semantic understanding [40, 58, 67, 71, 83] with physics-based simulation [31-33]. At its core, PhysGen3D is a digital twinning method that estimates an object's 3D shape, pose, physical and lighting properties, infers background geometry and appearance, deduces physical characteristics, and performs dimensional analysis—all from a single input image. This task is conventionally challenging due to its inherently ill-posed nature. To tackle this, we leverage various pretrained vision models, integrating their outputs to create an image-centric digital twin." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "For physical simulation, we employ material point methods [30, 37], a robust point-voxel-based framework that models counterfactual physical behaviors of objects in the image. Through precise inference of physical properties," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "simulations in the PhysGen3D environment achieve a high degree of realism and stability. We further enhance realism by applying physics-based rendering, seamlessly integrating dynamic effects back into the original image. PhysGen3D produces results that are not only visually realistic in terms of dynamics and lighting but also highly controllable, allowing users to specify initial conditions like speed and material properties. Due to the use of large pretrained models, our pipeline operates effectively without task-specific training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 182, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 266 + ], + "type": "text", + "content": "Our experiments, based on a carefully designed and rigorous user study, demonstrate that, compared to closed-source state-of-the-art video AIGC models such as Pika, Kling, and Gen-3, our framework provides significantly more flexible control over object motions, generates videos that better align with user intentions, and achieves superior physical realism—all while maintaining comparable rendering quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 284, + 389, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 284, + 389, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 284, + 389, + 297 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 307, + 547, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 487 + ], + "type": "text", + "content": "Single-view 3D reconstruction Great progress has been made in object-centric single view 3D reconstructions [45, 46, 62, 83]. However, applying these techniques to single-view 3D scene reconstruction becomes more challenging. Most existing works only focus on one part of 3D scene understanding. Geometric approaches reconstruct the 3D scene holistically but neglect individual object understanding, or they focus solely on foreground objects without considering the interaction of objects with the complex environment. Additionally, works have been done for scene relighting [49, 93, 95] and segmentation [40, 47, 58, 85, 97], but none of these provide a full 3D understanding within the image. In our work, we also perform perception reasoning for object materials, backgrounds, rendering properties, and physical stability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 713 + ], + "type": "text", + "content": "Controllable video generation. Video generation has made significant progress in recent years [5, 9, 10, 15, 21, 23, 25, 28, 41, 54, 63, 69, 73, 82, 90, 91]. The state-of-the-art framework [1, 3] can generate photorealistic and coherent videos from text instructions using diffusion models [27, 55, 59, 64, 65]. Controllable video generation is achieved through leveraging pre-trained video generation models with conditioning information, including but not limited to depth maps [13, 86], linear translation [39], layouts [44, 70, 87], and multiple combinations [50]. However, most existing generation methods implicitly generate image space dynamics, which might lead to unrealistic, hallucinated motions. In contrast, our method explicitly controls the motion and interaction with simulation, allowing us to create more sophisticated effects without the need for extensive training data. Our approach is training-free and generalizable to all objects in the world." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 315, + 759 + ], + "type": "text", + "content": "6179" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 61, + 545, + 281 + ], + "blocks": [ + { + "bbox": [ + 51, + 61, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 51, + 61, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 51, + 61, + 545, + 281 + ], + "type": "image", + "image_path": "fd7253eba88cec6ff23009b105abf5f7acd3dfa849c9a4e32bc4faa5e1ba38c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 295, + 548, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 548, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 548, + 329 + ], + "type": "text", + "content": "Figure 2. Method Overview. PhysGen3D's framework consists of three modules: a) 3D world creation, which infers geometry, semantics, rendering and physical parameters from the input image; b) dynamics simulation using Taichi-Elements for particle-based physics; and c) physics-based rendering with a two-pass shadow mapping technique." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 340, + 290, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 290, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 290, + 568 + ], + "type": "text", + "content": "Controllable image animation Image-based animation aims to animate objects that appear in images. Numerous works [17, 18, 29, 61, 68, 74, 84] have focused on this task. To improve quality, recent research has adopted data-driven solutions, training temporal neural networks to directly predict subsequent video frames [6, 14, 19, 24, 29, 72], or incorporating physical heuristics [17, 36]. Recently, there has been increased focus on interactive [7, 8, 43] and controllable [4, 18] image-to-video synthesis. Additional priors such as motion fields [19, 29, 51, 52, 66, 84], optical flow [11, 22, 94], 3D geometry information [57, 76], and user annotations [43] have also been introduced. The closest work to ours is Liu et al. [48], which focuses on image understanding and uses an existing 2D rigid simulator to generate 2D animations. However, their approach is limited to 2D dynamics, does not account for real-world 3D physics, and is restricted to rigid bodies. In contrast, our method provides more realistic and flexible control over 3D motions and object materials and extends beyond rigid-body physics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 584, + 115, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 115, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 115, + 597 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Our goal is to reconstruct an interactive, camera-centric miniature world from a single input image. We aim to control object materials, dynamics, and motions to simulate diverse, photorealistic, and physically plausible videos from user input. The key challenge lies in partial observations and the ill-posed nature of physical reasoning without observed dynamics. To address this, we propose a holistic reconstruction method leveraging large pretrained visual models to jointly infer geometry, dynamic materials, lighting, and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 340, + 548, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 548, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 548, + 425 + ], + "type": "text", + "content": "PBR materials from a single image (Sec. 3.1). The reconstructed scene is input into a material-point-method (MPM) simulator to generate realistic physics phenomena (Sec. 3.2). Finally, we render dynamic object behaviors based on the simulation and reintegrate them into the scene, producing realistic videos with accurate appearance and motion (Sec. 3.3). Fig. 2 depicts our framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 431, + 526, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 431, + 526, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 526, + 445 + ], + "type": "text", + "content": "3.1. Interactive 3D World from a Single Image" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 450, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 545 + ], + "type": "text", + "content": "A full amodal reconstruction of the 3D world depicted in the image is critical for the next step in simulation. An ideal reconstruction should capture a comprehensive understanding of objects' relationships, geometry, appearance, material, and physical properties. However, obtaining this understanding from a single image is highly ill-posed. Our key idea is to leverage priors from pretrained vision foundation models to help infer these properties, as shown in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 546, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 617 + ], + "type": "text", + "content": "Segmentation. We leverage vision foundation models to recognize object categories and segment object instances. Specifically, we use GPT-4o [89] to identify the foreground object categories and use Grounded-SAM [40, 47, 58] to further detect and segment each individual instance " + }, + { + "bbox": [ + 304, + 546, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\{o^i\\in \\mathbb{R}^{W\\times H\\times 3}\\}_{i}^{N}" + }, + { + "bbox": [ + 304, + 546, + 547, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 546, + 547, + 617 + ], + "type": "inline_equation", + "content": "o^i" + }, + { + "bbox": [ + 304, + 546, + 547, + 617 + ], + "type": "text", + "content": " is the image of i-th object." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "text", + "content": "Mesh Generation. Unlike previous work [48], which used only 2D rigid-body physics, we extend to general-purpose 3D simulation. This requires a complete 3D representation of foreground objects. We adopt InstantMesh [83], which uses Zero123++ [46, 62] to synthesize multi-view images from the segmented object image " + }, + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "inline_equation", + "content": "o^i" + }, + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "text", + "content": " and uses these images to reconstruct the 3D mesh " + }, + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "text", + "content": " of the objects. For multiple object occlusions, we adopt the iterative inpainting" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "6180" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "strategy to extract the 3D mesh of the objects (see supp for more details)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 288, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 288, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 288, + 276 + ], + "type": "text", + "content": "Background Handling. The background serves important roles in both dynamic simulation and rendering. In simulation, background geometry acts as a support and collider, where accuracy ensures realistic object-scene interactions. In rendering, it serves as a static backdrop while the foreground objects move and helps simulate realistic global illumination effects like cast shadows. For simulation purposes, we use Dust3r [71] to estimate iamage depth. The output depth map " + }, + { + "bbox": [ + 46, + 97, + 288, + 276 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{W \\times H}" + }, + { + "bbox": [ + 46, + 97, + 288, + 276 + ], + "type": "text", + "content": " is unprojected to the 3D world as a 3D point cloud " + }, + { + "bbox": [ + 46, + 97, + 288, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 97, + 288, + 276 + ], + "type": "text", + "content": ", and Bilateral Normal Integration [12] is applied to generate a smooth surface " + }, + { + "bbox": [ + 46, + 97, + 288, + 276 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 97, + 288, + 276 + ], + "type": "text", + "content": " serving as the collider. For rendering, to fill background regions occluded by moving objects, we use the LaMA inpainting model[67]. This model generates a complete background after masking out all objects and their shadows." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 277, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 384 + ], + "type": "text", + "content": "Object Pose and Scale Estimation. The generative 3D reconstruction step provides a complete mesh with a normalized scale in an object-centric coordinate system, but it does not infer the object's location and scale in the camera coordinate. To ensure coherence with the input image, it is necessary to accurately place the 3D meshes " + }, + { + "bbox": [ + 46, + 277, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 46, + 277, + 287, + 384 + ], + "type": "text", + "content": " into the 3D scene " + }, + { + "bbox": [ + 46, + 277, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 277, + 287, + 384 + ], + "type": "text", + "content": " with the correct scale and 6DoF pose. This registration task is challenging, as the generated object mesh may not perfectly match the real-world object in the image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "text", + "content": "To address this challenge, we designed a multi-stage coarse-to-fine alignment strategy. In the coarse stage, we perform 2D-3D feature point matching. Firstly, we render multiple images from viewpoints uniformly distributed over a unit sphere surrounding the object. For each rendered image, we match its feature points with the original object image " + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "inline_equation", + "content": "o^i" + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "text", + "content": " using SuperGlue [60], and the viewpoint with the most matches is selected. Matched pairs " + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "inline_equation", + "content": "(p_i^N \\in \\mathbb{R}^2, p_i^N \\in \\mathbb{R}^2)" + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "text", + "content": " are then projected back into 3D points " + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "inline_equation", + "content": "(P_i^N \\in \\mathbb{R}^3, P_i^N \\in \\mathbb{R}^3)" + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "text", + "content": " in object and camera coordinate, respectively. Perspective-n-Point (PnP) algorithm [20] is applied between " + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "inline_equation", + "content": "P_i^N" + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "inline_equation", + "content": "p_i^N" + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "inline_equation", + "content": "o^i" + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "text", + "content": " to estimate the 6DoF pose with scale ambiguity. Then we adjust the scale and translation factor simultaneously to minimize the L2 loss " + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "inline_equation", + "content": "\\sum_{i}^{N} = ||P_i^N - P_i'^N||_2" + }, + { + "bbox": [ + 46, + 385, + 288, + 566 + ], + "type": "text", + "content": ", without altering its projection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "text", + "content": "In the fine alignment stage, we render the mask and depth on the image plane using the current estimation through a differentiable renderer and jointly minimize two losses: " + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{dice}} + \\mathcal{L}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dice}} = 1 - \\frac{2\\times|M_A\\cap M_B|}{|M_A| + |M_B|}" + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "text", + "content": " measures the discrepancy between the rendered mask " + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "inline_equation", + "content": "M_{A}" + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "text", + "content": " and the observed mask " + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "inline_equation", + "content": "M_B" + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "text", + "content": " from Grounded-SAM. The depth consistency loss is " + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{depth}} = \\frac{\\mathrm{mse}(M_B*z_A - M_B*z_B)}{|M_B|}" + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "inline_equation", + "content": "z_{A}" + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "inline_equation", + "content": "z_{B}" + }, + { + "bbox": [ + 47, + 567, + 288, + 714 + ], + "type": "text", + "content": " are the rendered and Dust3r-predicted depths. This joint optimization ensures consistency between the estimated mesh in the camera coordinate and the point cloud from Dust3r, while maximizing the alignment between the 3D pose and observed object mask, ensuring accurate simu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 434, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 434, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 434, + 84 + ], + "type": "text", + "content": "lation and rendering alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 85, + 547, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 547, + 252 + ], + "type": "text", + "content": "Appearance Optimization. The texture of the generated 3D meshes may differ from the input image. To enhance rendering quality, we use the inverse rendering pipeline in Mitsuba3 [35] to estimate material properties. Lighting parameters are estimated with DiffusionLight [56], while object PBR materials (albedo, roughness, and metallic) are optimized via differentiable rendering. To handle unknown back views and reduce complexity, we assume uniform roughness and metallic values per object and apply tone mapping for albedo optimization: " + }, + { + "bbox": [ + 304, + 85, + 547, + 252 + ], + "type": "inline_equation", + "content": "y(x) = ax^3 + bx^2 + cx + d" + }, + { + "bbox": [ + 304, + 85, + 547, + 252 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 85, + 547, + 252 + ], + "type": "inline_equation", + "content": "y(0) = 0" + }, + { + "bbox": [ + 304, + 85, + 547, + 252 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 85, + 547, + 252 + ], + "type": "inline_equation", + "content": "y(1) = 1" + }, + { + "bbox": [ + 304, + 85, + 547, + 252 + ], + "type": "text", + "content": ". With reconstructed lighting and ground surface, the optimized materials improve asset appearance, capturing realistic object-surface interactions during rendering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 253, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 547, + 373 + ], + "type": "text", + "content": "Physics Reasoning. Accurately simulating real-world dynamics requires estimating physical parameters. We focus on two aspects: 1) Following [79], we use GPT-4o to query each object's elasticity and density, and the friction coefficient for the surface " + }, + { + "bbox": [ + 304, + 253, + 547, + 373 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 253, + 547, + 373 + ], + "type": "text", + "content": ". 2) We ensure reconstructed 3D assets match real-world proportions, as depth inaccuracies from Dust3r can cause unrealistic behaviors. To address this, we estimate a scale factor " + }, + { + "bbox": [ + 304, + 253, + 547, + 373 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 253, + 547, + 373 + ], + "type": "text", + "content": " by comparing asset size with typical real-world sizes from GPT-4o and use " + }, + { + "bbox": [ + 304, + 253, + 547, + 373 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 253, + 547, + 373 + ], + "type": "text", + "content": " for dimensionless scaling of gravity and velocity-related parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 381, + 429, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 429, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 429, + 393 + ], + "type": "text", + "content": "3.2. Dynamics Simulation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 400, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 547, + 460 + ], + "type": "text", + "content": "Given the 3D assets with reasoned physical properties and the scale factor, we use the physics engine Taichi Elements [31-33], based on the Material Point Method (MPM) [37], as our simulator to support a variety of different materials, including but not limited to rigid, soft, and granular." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 460, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 556 + ], + "type": "text", + "content": "Particle representation. The simulator is based on a particle-based representation. To convert our 3D assets into a simulatable particle-based representation, we apply floating points removing, internal filling and voxel downsampling. We use downsampling to handle uneven particle distributions, where downsample rate is adjusted according to the grid size of the simulator. For the convenience of rendering, we prioritize surface points." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "type": "text", + "content": "Physical parameters. To enhance stability, we apply the scale factor " + }, + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "type": "text", + "content": " to physical parameters in simulator instead of scaling the assets to real size. For example, the gravitational acceleration is multiplied by " + }, + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "type": "text", + "content": ", making to " + }, + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "type": "inline_equation", + "content": "k * 9.8" + }, + { + "bbox": [ + 304, + 557, + 547, + 628 + ], + "type": "text", + "content": " (nondimensionalized here). With this tuning, the motion of falling or collapse remains realistic for all scales." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 665 + ], + "type": "text", + "content": "External disturbance. For each object, we set a different initial velocity based on the user input to make the object move as specified by the user." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Other Visual Effects. Besides real world physics simulation, our pipeline allows special effects like collapsing and melting. To simulate different materials (rigid, soft, or granular), we can easily change the material type to modify" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 314, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 314, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 314, + 759 + ], + "type": "text", + "content": "6181" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 57, + 494, + 268 + ], + "blocks": [ + { + "bbox": [ + 97, + 57, + 494, + 268 + ], + "lines": [ + { + "bbox": [ + 97, + 57, + 494, + 268 + ], + "spans": [ + { + "bbox": [ + 97, + 57, + 494, + 268 + ], + "type": "image", + "image_path": "0ee7dc5270c6c9c2dba35d2ac98e011544b60151acdb5dead4e9656f78f5d80d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 273, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 273, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 547, + 298 + ], + "type": "text", + "content": "Figure 3. Video generation results. Left: Input initial frame. Right: Generated future frames. We apply an initial velocity to each movable object and use the physically grounded parameters outlined in 3.2 to generate physically plausible results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 288, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 327 + ], + "type": "text", + "content": "the physical properties of an object. This approach provides more flexibility for the user to edit the scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 335, + 189, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 189, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 189, + 348 + ], + "type": "text", + "content": "3.3. Physics-Based Rendering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 290, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 290, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 290, + 510 + ], + "type": "text", + "content": "After dynamics simulation, we obtain object point trajectories and apply motion interpolation to compute vertex motions, deforming the mesh accordingly. With optimized PBR materials, we use Mitsuba3 for Physically-Based Rendering under environment lighting. Following prior insertion rendering work [16, 38, 75, 88], we avoid converting the entire static background into the rendering pipeline. Instead, we build a 3D shadow catcher surface from the background depth. During rendering, no texture is applied to the background; two-pass shadow mapping extracts shadows and global illumination effects. The foreground objects and shadows are then composited onto the inpainted background to produce the final video with realistic lighting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 520, + 173, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 173, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 173, + 533 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 540, + 100, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 100, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 100, + 552 + ], + "type": "text", + "content": "4.1. Setups" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 629 + ], + "type": "text", + "content": "The test images come from a diverse set of our own photography, online collections, and generative models. Our pipeline is primarily designed for object-centric scenes with one or a few objects. Due to our limitations (Sec. 5), we excluded those with excessive objects, heavy occlusion between objects, or highly uneven surfaces." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 689 + ], + "type": "text", + "content": "Post-processing. We used VEnhancer [26] as an optional post-processing module, which takes the produced video and a text prompt to perform enhancement. As shown in Table 2, while it restores some details, it also introduces extra hallucinations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Baselines. Our method is one of the first of its kind, as existing model-simulate approaches require multi-view im" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 303, + 547, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 547, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 547, + 435 + ], + "type": "text", + "content": "ages [78, 80, 92] or specific settings [43, 48, 66]. Therefore, we evaluate ours against Image-to-Video (I2V) models: two open-source motion controller models DragAnything[77], MOFA-Video[53] and three state-of-the-art (SOTA) commercial models Kling 1.0 [3], Gen-3 [2], and Pika 1.5 [1]. We manually set correct motion trajectories and select applied regions for DragAnything, MOFA-Video and Kling to provide privileged motion guidance. We give text descriptions to Pika 1.5 and Gen-3, as they lack direct motion control capabilities. Additionally, Pika 1.5 offers \"Pikafect\" effects, such as \"Melt it\" and \"Deflate it.\"" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 443, + 362, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 362, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 362, + 455 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 568 + ], + "type": "text", + "content": "Our system generates a miniature interactive world from a single image, enabling the simulation of various phenomena. Fig. 3 presents videos generated from different types of images. These images encompass single or multiple objects and various physical materials (i.e., rigid or soft). We also show applications like dynamic changes, object editing, and dense 3D tracking, illustrating the adaptability and creative potential of our approach for generating customized, interactive video content." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 689 + ], + "type": "text", + "content": "Comparisons. We compare the results in two dimensions: motion control alone and physical materials. Fig. 4 shows that our system produces more physically plausible and controllable videos compared to SOTA I2V models. Despite prompt tuning, learning-based models often hallucinate, failing to adhere to physical laws or user intent. For example, in the toy dog case, we manually adjust the material and accurately simulate a collapse, whereas other models fall short. Similarly, for the book case, our results are the most physically realistic." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Dynamics. Fig. 5 demonstrates the varying dynamics generated from the same input image, highlighting the high" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "6182" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 56, + 499, + 597 + ], + "blocks": [ + { + "bbox": [ + 96, + 56, + 499, + 597 + ], + "lines": [ + { + "bbox": [ + 96, + 56, + 499, + 597 + ], + "spans": [ + { + "bbox": [ + 96, + 56, + 499, + 597 + ], + "type": "image", + "image_path": "7d701def37dd7cf51e8cb9c57d4b5ee8c67cb19161258ad41486b2731a9aef22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 606, + 547, + 641 + ], + "lines": [ + { + "bbox": [ + 46, + 606, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 547, + 641 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison. We compare videos generated from our framework with three state-of-the-art I2V models: Gen-3 [2], Pika [1] and Kling [3]. We carefully designed the prompt to describe the motion outcome, and uses motion brush to control Kling. Our framework employs initial velocity control. Results show that our method can follow text instructions while maintaining plausible physics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 647, + 289, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 289, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 289, + 708 + ], + "type": "text", + "content": "controllability of our method over physical parameters and motion trajectories. In the three rows on the left, we set different elasticities for the two objects, while keeping their initial positions and velocities the same. In the three rows on the right, we alter velocity directions of the objects and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 647, + 450, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 450, + 661 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 450, + 661 + ], + "type": "text", + "content": "keep physical parameters the same." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "Editing. Our method enables modifications to videos by removing, adding, or replacing objects, as illustrated in Fig. 6. The generated 3D assets can be easily manipulated, allowing for diverse video edits." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 760 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 760 + ], + "type": "text", + "content": "6183" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 62, + 525, + 277 + ], + "blocks": [ + { + "bbox": [ + 66, + 62, + 525, + 277 + ], + "lines": [ + { + "bbox": [ + 66, + 62, + 525, + 277 + ], + "spans": [ + { + "bbox": [ + 66, + 62, + 525, + 277 + ], + "type": "image", + "image_path": "7f40b3e95e99456ead552db3794caf0cac6a6c2f80a03035004970785dc473f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 288, + 547, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 547, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 547, + 321 + ], + "type": "text", + "content": "Figure 5. Dynamics Effects. We can generate various dynamics from the same input image. The left three columns share the same initial positions and velocities, but are in different materials. The right three columns have the same material, but defer in velocity directions. This showcases the potential of our method for generating diverse physical scenarios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 99, + 329, + 496, + 541 + ], + "blocks": [ + { + "bbox": [ + 99, + 329, + 496, + 541 + ], + "lines": [ + { + "bbox": [ + 99, + 329, + 496, + 541 + ], + "spans": [ + { + "bbox": [ + 99, + 329, + 496, + 541 + ], + "type": "image", + "image_path": "bd89ae7d75def4d460580a26442d603f49deb985da903ce1225730fa86ed802e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 552, + 547, + 586 + ], + "lines": [ + { + "bbox": [ + 46, + 552, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 547, + 586 + ], + "type": "text", + "content": "Figure 6. Video edition. As illustrated, our method supports video editing. In the top row, we remove the chair and allow the toy to fall from a static position. In the bottom two rows, we exchange one object between two scenes while keeping the other unchanged. This demonstrates the great flexibility of our video generation approach." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 597, + 289, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 289, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 289, + 668 + ], + "type": "text", + "content": "Tracking. Our framework uses an explicit 3D representation and works with a particle-based physics simulator. This allows our method to easily create videos with detailed 3D tracking results. Fig. 7 showcases two examples, demonstrating the accuracy and reliability of the tracking in different scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "Ablation. Each step we designed in perception, simulation and rendering is intended to mimic the real world. Fig. 8 shows the results of ablation study involving position" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 597, + 547, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 547, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 547, + 707 + ], + "type": "text", + "content": "optimization and inverse texture. Without pose optimization using differentiable rendering, the two objects are roughly at the initial place shown in input image, but cannot fully replicate the scene. Without inverse texture, the generated object mesh does not match the input in terms of color tone and brightness. We also conducted ablations on point sampling. However, without sampling, a large number of points become crowded in several MPM simulation grids, causing the simulator to crash and fail to produce final outputs." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 760 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 760 + ], + "type": "text", + "content": "6184" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 54, + 138, + 124 + ], + "blocks": [ + { + "bbox": [ + 72, + 54, + 138, + 124 + ], + "lines": [ + { + "bbox": [ + 72, + 54, + 138, + 124 + ], + "spans": [ + { + "bbox": [ + 72, + 54, + 138, + 124 + ], + "type": "image", + "image_path": "5ccd443d8f46e08f3d34a1c25be6065d405314fa428a6c77905917dbc08af94d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 129, + 287, + 161 + ], + "lines": [ + { + "bbox": [ + 47, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 287, + 161 + ], + "type": "text", + "content": "Figure 7. Dense 3D Tracking. Our method can naturally generate videos with dense 3D tracking results. Here we show tracking results for collapse and bounce cases." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 54, + 199, + 124 + ], + "blocks": [ + { + "bbox": [ + 139, + 54, + 199, + 124 + ], + "lines": [ + { + "bbox": [ + 139, + 54, + 199, + 124 + ], + "spans": [ + { + "bbox": [ + 139, + 54, + 199, + 124 + ], + "type": "image", + "image_path": "3d6fe1b6722bb10feb967a35d7ede758d22f7ad745f8ed82126f3daeaf8f91a1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 199, + 55, + 258, + 124 + ], + "blocks": [ + { + "bbox": [ + 199, + 55, + 258, + 124 + ], + "lines": [ + { + "bbox": [ + 199, + 55, + 258, + 124 + ], + "spans": [ + { + "bbox": [ + 199, + 55, + 258, + 124 + ], + "type": "image", + "image_path": "26c42bb64e6966cb03323f952e349a13dc34e0bffb5252a28abcfb91e30d45fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 170, + 286, + 239 + ], + "blocks": [ + { + "bbox": [ + 48, + 170, + 286, + 239 + ], + "lines": [ + { + "bbox": [ + 48, + 170, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 170, + 286, + 239 + ], + "type": "image", + "image_path": "87f9c188d3cb717ded077e3d3c21d92b24ddb579536ac105935fa56acff25b97.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 246, + 287, + 280 + ], + "lines": [ + { + "bbox": [ + 47, + 246, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 287, + 280 + ], + "type": "text", + "content": "Figure 8. Ablation study. Without pose optimization, the two objects cannot fully replicate the scene. Without inverse texture, the object mesh does not match the input color." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 83, + 312, + 247, + 365 + ], + "blocks": [ + { + "bbox": [ + 47, + 286, + 287, + 308 + ], + "lines": [ + { + "bbox": [ + 47, + 286, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 287, + 308 + ], + "type": "text", + "content": "Table 1. Human Evaluation Results. The three criteria are: Physical Realism, Photorealism and Semantic Consistency" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 312, + 247, + 365 + ], + "lines": [ + { + "bbox": [ + 83, + 312, + 247, + 365 + ], + "spans": [ + { + "bbox": [ + 83, + 312, + 247, + 365 + ], + "type": "table", + "html": "
MethodsPhysRealPhotoRealAlign
Kling 1.0 [3]2.8113.5662.467
Runway Gen-3 [2]2.2833.5821.886
Pika 1.5 [1]2.4123.3142.016
Ours3.7073.4113.866
", + "image_path": "5d7c57c950b377a7b5474a97043cedfe6d5c72396bf614c4ca5149a83f553d3a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 50, + 418, + 285, + 492 + ], + "blocks": [ + { + "bbox": [ + 47, + 370, + 287, + 415 + ], + "lines": [ + { + "bbox": [ + 47, + 370, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 287, + 415 + ], + "type": "text", + "content": "Table 2. VBenchmark Scores and GPT-4o Evaluation Results. Motion and Imaging refers to Motion Smoothness and Imaging Quality scores in VBenchmark. The three criteria on the right are the same as Table 1, with results given by GPT-4o." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 418, + 285, + 492 + ], + "lines": [ + { + "bbox": [ + 50, + 418, + 285, + 492 + ], + "spans": [ + { + "bbox": [ + 50, + 418, + 285, + 492 + ], + "type": "table", + "html": "
MethodsMotion↑Imaging↑PhysReal↑PhotoReal↑Align↑
Kling 1.00.9960.6710.5630.8740.596
Runway Gen-30.9910.7230.1410.8960.144
Pika 1.50.9940.6710.5440.8630.563
MOFA-Video0.9940.6340.3840.7640.304
DragAnything0.9850.4280.6450.7560.380
Ours0.9950.6660.7520.8670.796
Ours+VEnhancer0.9940.6770.7660.8800.745
", + "image_path": "bc5664ff1dbde64ec177069c29ca34b7272416b0e2447ee115f1bd081ce1b0e7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 503, + 189, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 189, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 189, + 516 + ], + "type": "text", + "content": "4.3.Quantitative Comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 521, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 556 + ], + "type": "text", + "content": "To assess the quality of the generated videos, we performed human evaluation, GPT-4o evaluation and provided two VBenchmark[34] evaluation scores." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 558, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 701 + ], + "type": "text", + "content": "Benchmarks. We designed three criteria for human and GPT-4o evaluation. (1) Physical Realism (PhysReal) measures how realistically the video follows the physical rules and whether the video represents real physical properties like elasticity and friction. (2) Photorealism (Photoreal) assesses the overall visual quality of the video, including the visual artifacts, discontinuities, and how accurately the video replicates details of light, shadow, texture, and materials. (3) Semantic Consistency (Align) evaluates how well the content of the generated video aligns with the intended meaning of the text prompt. We also chose two Quality Scores in VBenchmark: Motion Smoothness and Imaging Quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Details in evaluation. Following a similar methodology" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 55, + 544, + 122 + ], + "blocks": [ + { + "bbox": [ + 309, + 55, + 544, + 122 + ], + "lines": [ + { + "bbox": [ + 309, + 55, + 544, + 122 + ], + "spans": [ + { + "bbox": [ + 309, + 55, + 544, + 122 + ], + "type": "image", + "image_path": "515f589cee55e3624b71f3fa169f8d8ff40886f9004a775088e0ac648b8552ec.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 129, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 185 + ], + "type": "text", + "content": "Figure 9. Limitations. The first set of two images show rendering errors: extreme lighting and heavy shading cause incorrect painting and artifacts from the pillow penetrating the ground. The second set reveals simulation limitations: the MPM method struggles with fine details, such as the thin kettle handle, leading to failures." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 195, + 546, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 546, + 327 + ], + "type": "text", + "content": "to [48], we designed a questionnaire with 27 videos covering various scenes, motion conditions, and effects. Each video includes an input image, a motion prompt, and outputs from three SOTA commercial baselines and ours, shown in random order. 31 participants rated three criteria on a five-point scale from strongly disagree (1) to strongly agree (5). We also evaluated all five baselines and our diffusion-enhanced version using GPT-4o and VBenchmark. GPT-4o assessed videos on the same criteria based on 10 evenly sampled frames, with the input image and prompt provided. We used GPT version gpt-4o-2024-08-06, with detailed prompts in the supp." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 327, + 547, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 447 + ], + "type": "text", + "content": "Result analysis. The results in Table 1 demonstrate our ability to generate physically accurate and controllable videos. In physical realism (PhysReal) and semantic consistency (Align), our method achieves the highest scores and outperforms all the commercial models by a large margin. Table 2 presents the results of GPT-4o and VBench evaluation. GPT score is aligned with human evaluation, and we also beat open-sourced in VBench. Among the baseline models, Kling 1.0 performs the best overall, likely due to its use of the \"motion brush\", which specifies motion trajectories." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 457, + 381, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 457, + 381, + 470 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 381, + 470 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 477, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 547, + 574 + ], + "type": "text", + "content": "Our single-image-based interactive miniature world is designed for object-centric scenes with simple spatial geometry and inter-object relationships. Reconstructing complete scenes for more complex scenarios remains an open challenge. Fig. 9 illustrates several failure cases, such as rendering errors under challenging shading, perception failures, and simulation limitations. More detailed analysis on limitations can be found in the supp." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 584, + 383, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 584, + 383, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 584, + 383, + 597 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 604, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 547, + 712 + ], + "type": "text", + "content": "We present PhysGen3D, a framework that transforms a static image into an interactive 3D scene for simulating and rendering future motions based on user input. PhysGen3D integrates modules for 3D world reconstruction, model-based dynamic simulation, and physics-based rendering to generate realistic, controllable videos. By extending the 2D image-to-video paradigm to 3D, PhysGen3D enables more realistic motion and diverse material behaviors. We hope this work inspires future research." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "6185" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 145, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 145, + 85 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 145, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 288, + 137 + ], + "type": "text", + "content": "This project is supported by the Intel AI SRS gift, Amazon-Illinois AICE grant, Meta Research Grant, IBM IIDAI Grant, and NSF Awards #2331878, #2340254, #2312102, #2414227, and #2404385. We greatly appreciate the NCSA for providing computing resources." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 158, + 106, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 158, + 106, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 158, + 106, + 171 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 178, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 178, + 145, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 178, + 145, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 178, + 145, + 189 + ], + "type": "text", + "content": "[1] Pika, 2024. 2, 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 149, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 149, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 149, + 201 + ], + "type": "text", + "content": "[2] Runway, 2024. 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 203, + 159, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 159, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 159, + 214 + ], + "type": "text", + "content": "[3] Klingai,2024.2,5,6,8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 216, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 288, + 248 + ], + "type": "text", + "content": "[4] Kfir Aberman, Yijia Weng, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Unpaired motion style transfer from video to animation. ACM TOG, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 250, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 303 + ], + "type": "text", + "content": "[5] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A space-time diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 306, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 338 + ], + "type": "text", + "content": "[6] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In CVPR, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 340, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 373 + ], + "type": "text", + "content": "[7] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In ICCV, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 375, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 407 + ], + "type": "text", + "content": "[8] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Bjorn Ommer. Understanding object dynamics for interactive image-to-video synthesis. In CVPR, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 409, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 463 + ], + "type": "text", + "content": "[9] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 465, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 288, + 508 + ], + "type": "text", + "content": "[10] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "type": "text", + "content": "[11] Richard Strong Bowen, Richard Tucker, Ramin Zabih, and Noah Snavely. Dimensions of motion: Monocular prediction through flow subspaces. In 3DV, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 544, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 586 + ], + "type": "text", + "content": "[12] Xu Cao, Hiroaki Santo, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Bilateral normal integration. In European Conference on Computer Vision, pages 552-567. Springer, 2022. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 633 + ], + "type": "text", + "content": "[13] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "text", + "content": "[14] Xi Chen, Zhiheng Liu, Mengting Chen, Yutong Feng, Yu Liu, Yujun Shen, and Hengshuang Zhao. Livephoto: Real image animation with text-guided motion control. arXiv preprint arXiv:2312.02928, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[15] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "model for generative transition and prediction. arXiv preprint arXiv:2310.20700, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 96, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 162 + ], + "type": "text", + "content": "[16] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7230-7240, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 163, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 547, + 205 + ], + "type": "text", + "content": "[17] Yung-Yu Chuang, Dan B Goldman, Ke Colin Zheng, Brian Curless, David H Salesin, and Richard Szeliski. Animating pictures with stochastic motion textures. ACM TOG, 2005. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 208, + 547, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 547, + 240 + ], + "type": "text", + "content": "[18] Abe Davis, Justin G Chen, and Frédo Durand. Image-space modal bases for plausible manipulation of objects in video. ACM TOG, 2015. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "type": "text", + "content": "[19] Yuki Endo, Yoshihiro Kanamori, and Shigeru Kuriyama. Animating landscape: self-supervised learning of decoupled motion and appearance for single-image video synthesis. arXiv preprint arXiv:1910.07192, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 330 + ], + "type": "text", + "content": "[20] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 332, + 547, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 547, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 547, + 375 + ], + "type": "text", + "content": "[21] Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. Preserve your own correlation: A noise prior for video diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 376, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 547, + 409 + ], + "type": "text", + "content": "[22] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In ICLR, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 411, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 547, + 464 + ], + "type": "text", + "content": "[23] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 467, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 547, + 510 + ], + "type": "text", + "content": "[24] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 512, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 547, + 555 + ], + "type": "text", + "content": "[25] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 556, + 547, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 547, + 600 + ], + "type": "text", + "content": "[26] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. Venhancer: Generative space-time enhancement for video generation. arXiv preprint arXiv:2407.07667, 2024. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 601, + 547, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 547, + 622 + ], + "type": "text", + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 624, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 547, + 678 + ], + "type": "text", + "content": "[28] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 680, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 547, + 713 + ], + "type": "text", + "content": "[29] Aleksander Holynski, Brian L Curless, Steven M Seitz, and Richard Szeliski. Animating pictures with eulerian motion fields. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "6186" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[30] Yuanming Hu, Yu Fang, Ziheng Ge, Ziyin Qu, Yixin Zhu, Andre Pradhana, and Chenfanfu Jiang. A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM TOG, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "text", + "content": "[31] Yuanming Hu, Tzu-Mao Li, Luke Anderson, Jonathan Ragan-Kelley, and Frédo Durand. Taichi: a language for high-performance computation on spatially sparse data structures. ACM Transactions on Graphics (TOG), 38(6):201, 2019. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 206 + ], + "type": "text", + "content": "[32] Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Frédo Durand. Difftaichi: Differentiable programming for physical simulation. *ICLR*, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 288, + 262 + ], + "type": "text", + "content": "[33] Yuanming Hu, Jiafeng Liu, Xuanda Yang, Mingkuan Xu, Ye Kuang, Weiwei Xu, Qiang Dai, William T. Freeman, and Frédo Durand. Quantaichi: A compiler for quantized simulations. ACM Transactions on Graphics (TOG), 40(4), 2021. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 264, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 330 + ], + "type": "text", + "content": "[34] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 288, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 375 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 375 + ], + "type": "text", + "content": "[35] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, and Delio Vicini. Dr.jit: A just-in-time compiler for differentiable rendering. Transactions on Graphics (Proceedings of SIGGRAPH), 41(4), 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 288, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 288, + 409 + ], + "type": "text", + "content": "[36] Wei-Cih Jhou and Wen-Huang Cheng. Animating still landscape photographs through cloud motion creation. IEEE Transactions on Multimedia, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 454 + ], + "type": "text", + "content": "[37] Chenfanfu Jiang, Craig Schroeder, Joseph Teran, Alexey Stomakhin, and Andrew Selle. The material point method for simulating continuum materials. In Acm siggraph 2016 courses, pages 1-52. 2016. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 288, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 489 + ], + "type": "text", + "content": "[38] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on graphics (TOG), 30(6):1-12, 2011. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 491, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 555 + ], + "type": "text", + "content": "[39] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "text", + "content": "[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "text", + "content": "[41] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[42] Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. In ICLR, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. arXiv preprint arXiv:2309.07906, 2023. 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[44] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 141, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 547, + 206 + ], + "type": "text", + "content": "[45] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10072–10083, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 208, + 547, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 547, + 240 + ], + "type": "text", + "content": "[46] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 243, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 296 + ], + "type": "text", + "content": "[47] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "type": "text", + "content": "[48] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 2, 3, 5, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "text", + "content": "[49] Xinyu Liu, Jinlong Li, Jin Ma, Huiming Sun, Zhigang Xu, Tianyun Zhang, and Hongkai Yu. Deep transfer learning for intelligent vehicle perception: A survey. Green Energy and Intelligent Transportation, page 100125, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 420 + ], + "type": "text", + "content": "[50] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 422, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 443 + ], + "type": "text", + "content": "[51] Aniruddha Mahapatra and Kuldeep Kulkarni. Controllable animation of fluid elements in still images. In CVPR, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 445, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 466 + ], + "type": "text", + "content": "[52] Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Implicit warping for animation with image sets. 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 521 + ], + "type": "text", + "content": "[53] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. In European Conference on Computer Vision, pages 111-128. Springer, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 523, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 544 + ], + "type": "text", + "content": "[54] OpenAI. Creating video from text. https://openai.com/sora, 2024.2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 545, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 567 + ], + "type": "text", + "content": "[55] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 569, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 545, + 634 + ], + "type": "text", + "content": "[56] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 98-108, 2024. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 678 + ], + "type": "text", + "content": "[57] Haonan Qiu, Zhaoxi Chen, Yuming Jiang, Hang Zhou, Xiangyu Fan, Lei Yang, Wayne Wu, and Ziwei Liu. Relitalk: Relightable talking portrait generation from a single video, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[58] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "6187" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "type": "text", + "content": "[60] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 207 + ], + "type": "text", + "content": "[61] Arno Schödl, Richard Szeliski, David H Salesin, and Irfan Essa. Video textures. In PACMCGIT, 2000. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "text", + "content": "[62] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "text", + "content": "[63] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "type": "text", + "content": "[64] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 344, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 288, + 365 + ], + "type": "text", + "content": "[65] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 367, + 288, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 367, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 367, + 288, + 399 + ], + "type": "text", + "content": "[66] Ryusuke Sugimoto, Mingming He, Jing Liao, and Pedro V Sander. Water simulation and rendering from a still photograph. In SIGGRAPH Asia, 2022. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 401, + 288, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 401, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 401, + 288, + 455 + ], + "type": "text", + "content": "[67] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 288, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 478 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 478 + ], + "type": "text", + "content": "[68] Martin Szummer and Rosalind W. Picard. Temporal texture modeling. In ICIP, 1996. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 288, + 534 + ], + "type": "text", + "content": "[69] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 578 + ], + "type": "text", + "content": "[70] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. arXiv preprint arXiv:2402.01566, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 634 + ], + "type": "text", + "content": "[71] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "text", + "content": "[72] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. arXiv preprint arXiv:2306.02018, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[73] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "text", + "content": "Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "type": "text", + "content": "[74] Li-Yi Wei and Marc Levoy. Fast texture synthesis using tree-structured vector quantization. In PACMCGIT, 2000. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 206 + ], + "type": "text", + "content": "[75] Yuxi Wei, Zi Wang, Yifan Lu, Chenxin Xu, Changxing Liu, Hao Zhao, Siheng Chen, and Yanfeng Wang. Editable scene simulation for autonomous driving via llm-agent collaboration. In The First Workshop on Populating Empty Cities—Virtual Humans for Robotics and Autonomous Driving at CVPR 2024, 2nd Round. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 209, + 547, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 547, + 241 + ], + "type": "text", + "content": "[76] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In CVPR, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 243, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 547, + 297 + ], + "type": "text", + "content": "[77] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 299, + 546, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 546, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 546, + 341 + ], + "type": "text", + "content": "[78] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video. In CVPR, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 343, + 547, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 547, + 399 + ], + "type": "text", + "content": "[79] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time interactive realistic and browser-compatible environment from a single video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4588, 2024. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 547, + 442 + ], + "type": "text", + "content": "[80] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In CVPR, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 445, + 547, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 547, + 499 + ], + "type": "text", + "content": "[81] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 501, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 547, + 555 + ], + "type": "text", + "content": "[82] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccrafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 556, + 547, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 547, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 547, + 610 + ], + "type": "text", + "content": "[83] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 613, + 546, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 645 + ], + "type": "text", + "content": "[84] Tianfan Xue, Jiajun Wu, Katherine L Bouman, and William T Freeman. Visual dynamics: Stochastic future generation via layered cross convolutional networks. T-PAMI, 2018. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 689 + ], + "type": "text", + "content": "[85] Chenggang Yan, Biyao Shao, Hao Zhao, Ruixin Ning, Yongdong Zhang, and Feng Xu. 3d room layout estimation from a single rgb image. IEEE Transactions on Multimedia, 22(11): 3014-3024, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "text", + "content": "[86] Hanshu Yan, Jun Hao Liew, Long Mai, Shanchuan Lin, and Jiashi Feng. Magicprop: Diffusion-based video edit" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "6188" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 676 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "ing via motion-aware appearance propagation. arXiv preprint arXiv:2309.00908, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 129 + ], + "type": "text", + "content": "[87] Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation. Entropy, 25 (10):1469, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 185 + ], + "type": "text", + "content": "[88] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1389-1399, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "text", + "content": "[89] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9(1):1, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "text", + "content": "[90] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 341 + ], + "type": "text", + "content": "[91] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qing, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 397 + ], + "type": "text", + "content": "[92] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2025. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 463 + ], + "type": "text", + "content": "[93] Xianling Zhang, Nathan Tseng, Ameerah Syed, Rohan Bhasin, and Nikita Jaipuria. Simbar: Single image-based scene relighting for effective data augmentation for automated driving vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3718-3728, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 464, + 288, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 288, + 486 + ], + "type": "text", + "content": "[94] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 487, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 288, + 553 + ], + "type": "text", + "content": "[95] Yupeng Zheng, Chengliang Zhong, Pengfei Li, Huan-ang Gao, Yuhang Zheng, Bu Jin, Ling Wang, Hao Zhao, Guyue Zhou, Qichao Zhang, et al. Steps: Joint self-supervised nighttime image enhancement and depth estimation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4916–4923. IEEE, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 620 + ], + "type": "text", + "content": "[96] Chengliang Zhong, Yuhang Zheng, Yupeng Zheng, Hao Zhao, Li Yi, Xiaodong Mu, Ling Wang, Pengfei Li, Guyue Zhou, Chao Yang, et al. 3d implicit transporter for temporally consistent keypoint discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3869-3880, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 621, + 288, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 621, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 48, + 621, + 288, + 676 + ], + "type": "text", + "content": "[97] Leisheng Zhong, Yu Zhang, Hao Zhao, An Chang, Wenhao Xiang, Shunli Zhang, and Li Zhang. Seeing through the occluders: Robust monocular 6-dof object pose tracking via model-guided video object segmentation. IEEE Robotics and Automation Letters, 5(4):5159-5166, 2020. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "6189" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/14176105-f8ea-4fe1-b4ab-02889a03b21d_content_list.json b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/14176105-f8ea-4fe1-b4ab-02889a03b21d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3dd80863e9a4fddb6c804077c7c069127e7a9c42 --- /dev/null +++ b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/14176105-f8ea-4fe1-b4ab-02889a03b21d_content_list.json @@ -0,0 +1,1681 @@ +[ + { + "type": "text", + "text": "PhysVLM: Enabling Visual Language Models to Understand Robotic Physical Reachability", + "text_level": 1, + "bbox": [ + 106, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weijie Zhou $^{1,2}$ , Manli Tao $^{2}$ , Chaoyang Zhao $^{2,3,\\dagger}$ , Haiyun Guo $^{2}$ , Honghui Dong $^{1}$ , Ming Tang $^{2}$ , Jinqiao Wang $^{1,2,3,4,\\dagger}$", + "bbox": [ + 250, + 202, + 746, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Traffic and Transportation, Beijing Jiaotong University", + "bbox": [ + 235, + 238, + 761, + 257 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{2}$ Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences", + "3 ObjectEye Inc. Guangdong Provincial Key Laboratory of Intellectual Property & Big Data, Guangdong Polytechnic Normal University" + ], + "bbox": [ + 117, + 256, + 877, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "†Corresponding authors: chaoyang.zhao@nlpr.ia.ac.cn, jqwang@nlpr.ia.ac.cn", + "bbox": [ + 168, + 311, + 823, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 361, + 326, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding the environment and a robot's physical reachability is crucial for task execution. While state-of-the-art vision-language models (VLMs) excel in environmental perception, they often generate inaccurate or impractical responses in embodied visual reasoning tasks due to a lack of understanding of robotic physical reachability. To address this issue, we propose a unified representation of physical reachability across diverse robots, i.e., Space-Physical Reachability Map (S-P Map), and PhysVLM, a vision-language model that integrates this reachability information into visual reasoning. Specifically, the S-P Map abstracts a robot's physical reachability into a generalized spatial representation, independent of specific robot configurations, allowing the model to focus on reachability features rather than robot-specific parameters. Subsequently, PhysVLM extends traditional VLM architectures by incorporating an additional feature encoder to process the S-P Map, enabling the model to reason about physical reachability without compromising its general vision-language capabilities. To train and evaluate PhysVLM, we constructed a large-scale multi-robot dataset, Phys100K, and a challenging benchmark, EQA-phys, which includes tasks for six different robots in both simulated and real-world environments. Experimental results demonstrate that PhysVLM outperforms existing models, achieving a $14\\%$ improvement over GPT-4o on EQA-phys and surpassing advanced embodied VLMs such as RoboMamba and SpatialVLM on the RoboVQA-val and OpenEQA benchmarks. Additionally, the S-P Map shows strong compatibility with various VLMs, and its integration into GPT-4o-mini yields a $7.1\\%$ performance improvement.", + "bbox": [ + 88, + 414, + 485, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 361, + 643, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/535fe186e22b80d0d2dc3d275570fe4336156db5e4ba7393c2b943b0f6824f8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 419, + 707, + 534 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a4a1a6975f734f86cd0b7b70f79cb8747df78880f35a5e4c15c3cc05f56e77d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 421, + 901, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Embodied QA Can robot pick up the bowl now?", + "bbox": [ + 527, + 537, + 803, + 550 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GPT40:", + "bbox": [ + 527, + 554, + 573, + 566 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e6c350836bfcc35a52933f2f35e9a7b8584f301ee426085ade0cb6bbebd3b5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 555, + 759, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yes, the bowl is placed on a flat surface and is accessible.", + "bbox": [ + 527, + 566, + 844, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "PhysVLM:", + "bbox": [ + 527, + 579, + 589, + 590 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1291bb3947edae24b71075d5f018e1ff11332a56f0d724dd2f50e66eb5ca376c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 579, + 759, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "No, the bowl is not within the robot arm's reachable space.", + "bbox": [ + 527, + 592, + 849, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Task Planning", + "bbox": [ + 527, + 612, + 609, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Task goal is to put the can into the bowl, what are the next steps?", + "bbox": [ + 614, + 611, + 888, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GPT40:", + "bbox": [ + 527, + 638, + 571, + 650 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7b9bd18dc43b0146055c501fbf80616557b2dc560cbb32b5b8c81fb1bdc1b60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 640, + 759, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Execution failure", + "bbox": [ + 763, + 641, + 844, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Picking up the can and place it inside the bowl.", + "bbox": [ + 527, + 652, + 785, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "PhysVLM:", + "bbox": [ + 527, + 662, + 589, + 675 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eea56502ab52faa3fdf13e7ae0d6a176450fb57dd6a95f4b3c3661caeb8a29fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 664, + 759, + 675 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1-Move close to the can. 2-Pick up the can. 3-Place the can into the bowl. 4-Down.", + "bbox": [ + 527, + 676, + 877, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Existing VLM models, such as GPT-4o, may generate inaccurate or impractical responses due to poor comprehension of robotic physical reachability. The proposed PhysVLM integrates vision-language capabilities with an understanding of robotic physical reachability.", + "bbox": [ + 511, + 713, + 906, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate perception of physical reachability is essential for robots to perform tasks effectively. Similar to how humans adjust their actions based on bodily conditions and environmental factors, robots must account for their physical reachability within an environment to ensure efficient and reliable task execution. For instance, in grasping tasks,", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6940", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a robot that fails to assess its reachability may attempt to grasp an object from an unreachable position, leading to task failure or equipment damage [32, 35]. Thus, enhancing a robot's understanding of physical reachability is crucial for successful task planning and execution in complex environments [31, 41].", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) have shown remarkable progress in environmental understanding [6, 17, 19, 21, 37], and many studies have applied these models to embodied AI to assist robots in perceiving environments and planning tasks [11, 18, 25, 36, 43]. However, while VLMs excel in general environmental perception, they often struggle with tasks that require an understanding of robotic physical reachability (see Figure 1). We identify two key challenges that must be addressed for VLMs to be effective in robotic tasks: (1) how to develop a unified and efficient representation of physical reachability. Robots vary significantly in size, joint types, and other characteristics, making it difficult for VLMs to directly learn these differences; (2) how to enable VLMs to improve their understanding of physical reachability without compromising general vision-language capabilities. Existing VLMs typically combine pre-trained unimodal encoders for vision and language tasks. However, introducing a new modality like physical reachability requires careful architectural and training adjustments to ensure the model can reason about reachability while maintaining its general capabilities.", + "bbox": [ + 91, + 184, + 483, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose the Space-Physical Reachability Map (S-P Map), a unified representation that abstracts the physical reachability of diverse robots into a generalized spatial form. The S-P Map is generated by combining robot parameters with egocentric depth images, but crucially, the model learns to focus on the abstracted reachability features (i.e., the gray regions in the S-P Map) rather than the specific robot configurations. This abstraction allows the model to generalize across different robots, as it only needs to reason about which areas are reachable, independent of the robot's specific characteristics. We introduce PhysVLM, a vision-language model that extends traditional VLM architectures by incorporating an additional feature encoder to process the S-P Map. This design enables PhysVLM to integrate physical reachability information into its reasoning process without compromising its general vision-language capabilities. To train and evaluate PhysVLM, we constructed a large-scale multi-robot dataset, Phys100K, and a challenging benchmark, EQAphys, which includes tasks for six different robots in both simulated and real-world environments.", + "bbox": [ + 89, + 503, + 483, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 109, + 823, + 401, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a unified and robot-agnostic formulation, the Space-Physical Reachability Map (S-P Map), which abstracts robotic physical reachability in a way that is independent of specific robot configurations, promoting the", + "bbox": [ + 91, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learning of generalized features.", + "bbox": [ + 527, + 90, + 743, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce PhysVLM, a vision-language model that integrates physical reachability with general vision-language capabilities via an additional feature encoder, improving task execution reliability.", + "- We release the EQA-phys benchmark, which includes six robots and 1.3K question-answer pairs, designed to test the model's understanding of physical reachability in simulated and real-world environments.", + "- Our model achieves a $14\\%$ improvement over GPT-4o on the EQA-phys benchmark. In embodied visual reasoning tasks on the RoboVQA-val and OpenEQA benchmarks, it outperforms advanced embodied VLMs such as RoboMamba and SpatialVLM. Furthermore, the S-P Map demonstrates strong compatibility with various VLMs, and integrating it into GPT-4o-mini results in a $7.1\\%$ performance improvement." + ], + "bbox": [ + 513, + 106, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 363, + 653, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. VLMs in Robotics", + "text_level": 1, + "bbox": [ + 513, + 388, + 689, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Embodied Question Answering (EQA) tasks require agents to interact with environments to answer questions [10, 13, 38]. RoboVQA offers a large, diverse dataset for robotic visual question answering. 3D-VLA [44] integrates 3D perception with a generative world model for embodied reasoning, while SpatialVLM [5] enhances VLMs' spatial understanding using extensive 3D data.", + "bbox": [ + 511, + 411, + 903, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Robot task planning involves sequencing subtasks to achieve goals [3, 30, 40]. Code as Policies (CaP) [20] employs OpenAI's Codex (code-davinci-002) to generate planning code. SayCan [2] combines the Pathways Language Model (PaLM) [7] with robotic affordances to create feasible action plans based on the robot's capabilities. However, these methods often assume all objects are within the robot's operational area, ignoring physical reachability and potentially leading to suboptimal or infeasible plans.", + "bbox": [ + 511, + 518, + 905, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Understanding Physical Reachability", + "text_level": 1, + "bbox": [ + 511, + 665, + 833, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies use voxel grids with open-vocabulary detection models to assign task-specific attributes, enabling environmental constraint understanding. ReKep [15] generates keypoint proposals and constraints using voxel grids and VLMs, while VoxPoser [14] synthesizes robot trajectories by integrating OWL-ViT [26] and VLMs with voxel-based environment representations. However, these approaches focus on environment modeling without explicitly addressing the robot's physical reachability.", + "bbox": [ + 511, + 688, + 905, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Explicitly representing reachable workspaces remains challenging. Reachability maps [41] model spatial capabilities, and occupancy grids [16] account for obstacles to ensure safe navigation. Additionally, methods like online model predictive control with offline workspace analysis", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6941", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[31] and Reachability Expression-based Motion Planning (REMP) [12] address workspace constraints. Despite these advancements, integrating physical reachability into visual reasoning for complex embodied tasks is still limited, primarily due to the lack of large-scale datasets that include robotic physical parameters in VLM pretraining.", + "bbox": [ + 89, + 90, + 480, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 196, + 181, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PhysVLM is a large-scale vision-language model designed for visual reasoning that accounts for physical constraints in embodied tasks. As illustrated in Figure 2, PhysVLM integrates instruction text, visual input (RGB image), and an S-P Map abstracts the robotic physical reachability into a unified spatial representation. By combining these inputs, PhysVLM generates responses consistent with both the visual context and the robot's physical reachability, without being tied to specific robot configurations. The S-P Map is constructed using a unified physical reachability encoding method, which abstracts the physical parameters of various robots alongside their egocentric depth maps into a generalized form. This abstraction allows the model to generalize across different robots, addressing the challenge of learning and reasoning about physical reachability in a robot-agnostic manner.", + "bbox": [ + 89, + 220, + 483, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the core components of PhysVLM. Section 3.1 describes the S-P Map encoding method, Section 3.2 describes the model architecture, and Section 3.3 discusses the training procedures.", + "bbox": [ + 89, + 464, + 483, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. S-P Map Encoding", + "text_level": 1, + "bbox": [ + 89, + 534, + 272, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, we model the physical reachability of various robots using a unified approach that abstracts robot-specific parameters into a generalized spatial representation. This abstraction allows the model to focus on the spatial regions that are physically reachable, independent of the specific robot configuration.", + "bbox": [ + 89, + 556, + 483, + 647 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S} - \\mathrm {P} \\quad \\operatorname {M a p} = F \\left(\\mathcal {P} _ {\\text {r a w}}, \\left\\{\\theta_ {i} ^ {\\min }, \\theta_ {i} ^ {\\max } \\right\\}, \\mathrm {D H}, \\mathbf {E}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 674, + 482, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{P}_{\\mathrm{raw}}$ denotes the raw point cloud data from the robot's RGB-D camera. $\\{\\theta_i^{\\mathrm{min}},\\theta_i^{\\mathrm{max}}\\}$ represents the range of motion for each joint $i$ . DH refers to the Denavit-Hartenberg parameters that describe the geometric structure of each joint, and $\\mathbf{E}$ is the extrinsic calibration matrix that transforms coordinates from the camera to the robot's coordinate system. The function $F$ maps these inputs to produce the S-P Map, which abstracts the robot's physical reachability into a spatial form that is independent of the specific robot configuration.", + "bbox": [ + 89, + 704, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a robot arm with $n$ degrees of freedom, where each joint $i$ has DH parameters $\\{\\theta_i, d_i, a_i, \\alpha_i\\}$ ( $\\theta_i$ is the joint angle, $d_i$ is the offset along the z-axis, $a_i$ is the link", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "length, and $\\alpha_{i}$ is the twist angle). The homogeneous transformation matrix for each joint is defined as:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {i} = G \\left(\\theta_ {i}, d _ {i}, a _ {i}, \\alpha_ {i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 136, + 903, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $G$ is the standard Denavit-Hartenberg transformation function. By multiplying the transformation matrices of all joints, we obtain the transformation matrix from the base frame to the end-effector frame:", + "bbox": [ + 511, + 157, + 906, + 218 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} = \\mathbf {T} _ {1} \\mathbf {T} _ {2} \\dots \\mathbf {T} _ {n}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 233, + 903, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate joint configurations, we sample the joint angles $\\theta_{i}$ from their respective motion ranges $\\theta_{i} \\in [\\theta_{i}^{\\min}, \\theta_{i}^{\\max}]$ , resulting in configurations $\\{\\theta_{1}, \\theta_{2}, \\ldots, \\theta_{n}\\}$ . By substituting these joint configurations into the forward kinematics equations, we compute the corresponding end-effector positions:", + "bbox": [ + 511, + 255, + 905, + 330 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} _ {\\text {v o x e l}} = \\left\\{\\mathbf {p} \\mid \\mathbf {p} = \\mathbf {T} \\left(\\theta_ {1}, \\theta_ {2}, \\dots , \\theta_ {n}\\right) \\cdot \\mathbf {p} _ {0} \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 351, + 903, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_0$ is the origin point in the end-effector frame. We precompute these joint configurations offline, discretize the workspace into a voxel grid $\\mathcal{W}_{\\mathrm{voxel}}$ , and store it for efficient computation in subsequent steps.", + "bbox": [ + 511, + 393, + 903, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, as shown in Figure 2, the robot's raw point cloud $\\mathcal{P}_{\\mathrm{raw}}$ is captured from its egocentric RGB-D camera in the camera coordinate system and transformed into the robot's coordinate system using the extrinsic calibration matrix $\\mathbf{E}$ , resulting in the transformed point cloud $\\mathcal{P}$ :", + "bbox": [ + 511, + 454, + 905, + 529 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} = \\mathbf {E} \\cdot \\mathcal {P} _ {\\text {r a w}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 660, + 544, + 903, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure physical feasibility, we perform a voxel grid lookup to determine whether each point in $\\mathcal{P}$ lies within the precomputed reachable workspace $\\mathcal{W}_{\\mathrm{voxel}}$ :", + "bbox": [ + 511, + 566, + 905, + 612 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {\\text {v a l i d}} = \\left\\{\\mathbf {p} \\in \\mathcal {P} \\mid \\mathbf {p} \\in \\mathcal {W} _ {\\text {v o x e l}} \\right\\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 621, + 903, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This step filters the point cloud to include only points within the robot's reachable workspace, abstracting the robot's physical reachability into a generalized spatial form.", + "bbox": [ + 511, + 659, + 905, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we transform the valid point cloud $\\mathcal{P}_{\\mathrm{valid}}$ back into the camera coordinate system and use the camera's intrinsic parameters to project these points onto the image plane. We then mark the regions that comply with physical reachability on the original depth map. For areas that are not reachable, we apply a gray mask and outline their boundaries. The resulting S-P Map clearly highlights regions that are beyond the robot's physical reach, providing a unified and abstracted representation of reachability that is independent of the specific robot configuration. This allows the model to focus on the spatial constraints of the task, without needing to account for the detailed physical parameters of each robot.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6942", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/88d43d2171b258406312fcfa33b9d14fb4805f02ceb4d972a68f9f1cff100ba1.jpg", + "image_caption": [ + "Figure 2. Overview of PhysVLM. Starting with robot parameters and an egocentric depth map, an S-P Map is generated via unified physical reachability encoding. Using the S-P Map, image, and instruction text, PhysVLM generates textual output considering the robot's physical reachability." + ], + "image_footnote": [], + "bbox": [ + 96, + 108, + 416, + 344 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/32cc7a7325b209c2bfe43dd9c2f2cbef4029ba37c6ccdc2828b60eaa5e7b3c9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 108, + 903, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Model Architecture", + "text_level": 1, + "bbox": [ + 89, + 428, + 277, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To seamlessly integrate robotic physical reachability into PhysVLM while preserving its visual reasoning capabilities, we design a dual-branch architecture: one branch dedicated to vision processing and the other to physical reachability (see Figure 2). These branches operate independently, extracting features from their respective inputs, which are then fused and passed to a unified decoder for final reasoning and response generation.", + "bbox": [ + 89, + 453, + 483, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The vision branch leverages a pre-trained Vision Transformer (ViT) [9], specifically the SigLip-400M model [42], to extract high-level visual features from egocentric images. To reduce computational overhead, a Max Pooling layer is applied, followed by a two-layer Multi-Layer Perceptron (MLP) that transforms the visual features into token representations suitable for multimodal fusion.", + "bbox": [ + 89, + 577, + 483, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The physical reachability branch processes the S-P Map, which abstracts the robot's physical reachability into a generalized spatial form. This branch also utilizes the SigLip-400M model for feature extraction, followed by Max Pooling and a feature fusion layer. The fusion layer combines the visual and reachability features, and a two-layer MLP further refines these fused features into reachability-specific tokens.", + "bbox": [ + 89, + 686, + 483, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the language decoding, we employ the Qwen-2.5-Instruct-3B model [33, 39] as PhysVLM's large language model (LLM) decoder, using the Qwen-2.5 tokenizer to process natural language instructions. The decoder integrates multimodal tokens from the vision branch, S-P Maps, and language inputs, generating coherent and contextually rele", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vant textual responses that account for both visual and physical reachability information.", + "bbox": [ + 511, + 429, + 906, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 511, + 469, + 617, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Data Construction. The training data for PhysVLM consists of our Phys100K dataset and general VQA datasets, such as LLaVA-Pretrain, ShareGPT4V, and RoboVQA. Phys100K focuses on question-answering related to physical reachability, aggregating data from RoboVQA (20K samples), ScanNet [8] (10K samples), OpenX-Embodiment [27] (60K samples), and an additional 10K samples from PyBullet.", + "bbox": [ + 511, + 491, + 906, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depth maps are essential inputs for generating the S-P Map. For datasets lacking depth maps, we generate them using DepthAnything-v2. Additionally, we employ Grounding DINO [23] and SAM2 [34] to obtain 2D bounding boxes and segmentation results for objects in the images. In PyBullet, we simulate work scenarios using four robotic arms (UR5, FR5, CR5, and FRANKA) to collect RGB images, depth maps, and segmentation results.", + "bbox": [ + 511, + 613, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the PyBullet data in Phys100K, precise robot configurations can be obtained from the simulator. Therefore, we directly generate the S-P Map using the method described in Section 3.2, and labels indicating whether objects are reachable are obtained through simulated motion. The advantage of the S-P Map is that it abstracts physical reachability into a region-based representation, decoupling the learning process from specific robot configurations. This allows us to generate pseudo-labels for datasets without precise robot parameters. We approximate reachability using the segmentation results, marking regions and the objects within", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6943", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2fc2e9b0c0203553b91d189e788739d096524d5d6068b12942995c3621fb84ac.jpg", + "image_caption": [ + "Figure 3. Details of the Phys100K Dataset and EQA-Phys Benchmark." + ], + "image_footnote": [], + "bbox": [ + 93, + 108, + 906, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "them as \"reachable\" or \"unreachable\" based on depth values. Next, we generate question-answer pairs for two main categories:", + "bbox": [ + 89, + 468, + 482, + 513 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Embodied QA. GPT-4 generates question-answer pairs for ScanNet and RoboVQA, covering categories include Function Reasoning, World Knowledge, Object Recognition, Object Localization, Attribute Recognition, Spatial Reasoning, Object State Recognition, and Hallucination (See Figure 3). Detailed prompts and examples are provided in the appendix to guide the generation process.", + "- Tasks Involving Physical Reachability. We use the \"reachable\" label with five fixed task templates to generate question-answer pairs, such as \"USER:\\n\\n Is the [Object] in the robot's reachable space? ASSISTANT: Yes, it is.\" Here, [Object] represents the relevant object category, while and serve as placeholders for image patch tokens and S-P Map patch tokens, respectively. Figure 3 provides examples of question-answer pairs for each object category." + ], + "bbox": [ + 89, + 517, + 482, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Pipeline. We adopt a two-stage training process to fully leverage the S-P Map and ensure PhysVLM generalizes across different robots. In the first stage, we align multimodal features using the LLaVA-Pretrain and OpenX-Embodiment datasets from Phys100K. This stage only trains the projection layers, allowing the model to build", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a foundational understanding of visual inputs and physical reachability, independent of specific robot configurations.", + "bbox": [ + 511, + 468, + 906, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the second stage, we unfreeze all parameters and train the entire model using data from Phys100K, ShareGPT4V, and RoboVQA. This stage enhances PhysVLM's ability to handle complex visual reasoning tasks with physical reachability constraints, ensuring the model can generalize across diverse environments and robots.", + "bbox": [ + 511, + 502, + 906, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. PhysVLM is trained for 48 hours using eight A800 GPUs. The training process consists of two stages, each lasting one epoch. The batch size and learning rate are set at 128 and 1e-3 in the first stage, and 64 and 1e-5 in the second stage. The final model is PhysVLM-3B.", + "bbox": [ + 511, + 628, + 906, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. EQA-phys Benchmark", + "text_level": 1, + "bbox": [ + 511, + 739, + 723, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Figure 3, we introduce an embodied QA task focused on physical reachability, termed EQA-phys, which emphasizes QA tasks constrained by physical limitations. This benchmark includes a simulator dataset with 200 samples and 1,000 questions from the PyBullet validation set, as well as a zero-shot evaluation set based on real-world data from UR3 and XArm6 robots in two scenarios. The evaluation set contains 60 samples and 300 questions, all manually annotated by domain experts.", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6944", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f593b02147f76713dcfdf7ea942cc4efcdb3662a1f3172f6219f4cf60d7c1167.jpg", + "table_caption": [ + "Table 1. Results of EQA-phys. Comparison of PhysVLM-3B (ours) with API-based VLMs and embodied VLMs." + ], + "table_footnote": [], + "table_body": "
REAL-WORLDSIMULATORALL
UR3XARM6UR5FR5CR5FRANKA
API-BASED VLMSGPT-4O-MINI54.356.049.455.454.647.152.8
CLAUDE-3.556.260.554.058.155.754.356.4
GPT-4O56.761.555.758.357.552.657.0
GPT-4O-MINI + S-P MAP60.0↑5.760.5↑4.557.0↑7.659.1↑3.759.2↑4.653.3↑6.259.8↑7.0
CLAUDE-3.5 + S-P MAP65.3↑9.167.3↑6.854.9↑0.958.3↑0.258.2↑2.558.1↑3.860.3↑3.4
GPT-4O + S-P MAP66.6↑9.968.1↑6.655.8↑0.160.7↑1.459.4↑1.957.6↑5.061.3↑4.1
EMBODIED VLMSSPATIALVLM56.355.154.659.152.047.554.1
SPATIALBOT51.150.250.048.153.354.451.1
PHYSVLM-3B64.163.071.475.774.078.171.0
", + "bbox": [ + 102, + 127, + 893, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 329, + 222, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 89, + 354, + 290, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tasks. We compare the performance of PhysVLM with other methods across three categories of tasks:", + "bbox": [ + 89, + 376, + 482, + 407 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- EQA-phys. This benchmark tests the model's ability to integrate visual reasoning with robotic physical reachability. The real-robot component is used to assess PhysVLM's zero-shot generalization, highlighting its ability to handle unseen robots and environments.", + "- Embodied QA. We evaluate the model's general visual reasoning ability in embodied tasks using the OpenEQA [25] and RoboVQA-val [29] benchmarks.", + "- Robot Task Planning. For real-world tasks such as \"Pick A into B\", we assess the model's ability to understand robotic physical reachability and generate reasonable task plans. As this study does not focus on robot control strategies, we use the natural language planning approach from [29]." + ], + "bbox": [ + 89, + 410, + 483, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our model to several baselines, including API-accessible VLMs such as Claude 3.5 [28], GPT-4o-mini [1], and GPT-4o [1], as well as embodied VLMs like SpatialVLM [5], SpatialBot [4], 3D-VLA [44], and RoboMamba [22]. SpatialVLM and SpatialBot both use the 3B version, which has a similar parameter count to our model. Since the executable versions of 3D-VLA and RoboMamba are unavailable, we compare their reported results on RoboVQA-val.", + "bbox": [ + 89, + 638, + 482, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For tasks involving physical reachability, we use LLM scoring, following the approach used in existing studies [24, 25]. Assigning 5 points for completely correct responses and 1 point for incorrect responses, we calculate the average score and express it as a percentage. For Embodied QA, we follow the benchmark settings of the corresponding datasets [25, 29]. For task planning, each", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "task type is executed 10 times, and the average success rate serves as the evaluation metric.", + "bbox": [ + 511, + 330, + 905, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on EQA-phys", + "text_level": 1, + "bbox": [ + 511, + 390, + 714, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the results on EQA-phys. Neither API-based nor embodied VLMs can handle the robot's parameter constraints, resulting in suboptimal outputs with scores around $55\\%$ . In contrast, our model successfully completes visual reasoning tasks involving physical reachability, obtaining an average score of $71\\%$ . As discussed, these tasks require the model to perform visual reasoning based on an understanding of the robotic physical reachability. The model can only effectively accomplish these tasks by truly understanding robotic physical reachability.", + "bbox": [ + 511, + 419, + 906, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results in Table 1 demonstrate that prompting API-based VLMs, such as GPT-40, with the S-P Map (detailed in 4) significantly enhances their performance. This improvement stems from the S-P Map's ability to abstract physical reachability into a robot-agnostic representation, enabling VLMs to reason about physical constraints that would otherwise be beyond their capabilities. By decoupling reachability from specific robot parameters, the S-P Map facilitates generalization across diverse environments, allowing models to better comprehend physical reachability, even in previously unseen scenarios.", + "bbox": [ + 511, + 575, + 908, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 also demonstrates that PhysVLM-3B achieved scores of over $63\\%$ in zero-shot evaluations for UR3 and XArm6 robots, despite operating in new environments with different robot parameters. This performance is attributed to two key factors: (1) the S-P Map abstracts various robot parameters into a unified, transferable representation of physical reachability, and (2) the model's independent visual and constraint encoding branches allow it to learn generalizable visual features from diverse image-text data, enabling effective reasoning in novel environments.", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6945", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6ee41cde6cc2048172e628d13e387d3c0a1e9d2a3cecdb9ea3b2971b02c012d7.jpg", + "table_caption": [ + "Table 2. Embodied QA results on the RoboVQA-val set, comparison of PhysVLM (ours) with existing methods. An asterisk (*) indicates models not pre-trained on the RoboVQA dataset." + ], + "table_footnote": [], + "table_body": "
BLEU1BLEU2BLEU3BLEU4
SPATIALVLM*5.13.01.91.2
SPATIALBOT*12.49.38.07.2
3D-VLA48.338.531.726.8
ROBOMAMBA54.944.239.536.3
PHYSVLM-3B65.362.450.943.5
", + "bbox": [ + 91, + 155, + 488, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/63103c3fe4040698bcab4fc17acca0a367830fa12d6ad0787c12fd053a770c0e.jpg", + "table_caption": [ + "Table 3. Embodied QA results on the OpenEQA benchmark, comparison of PhysVLM (ours) with existing methods. An asterisk (*) indicates that only the first 200 samples were tested due to API limitations." + ], + "table_footnote": [], + "table_body": "
EM-EQA (SCANNET)EM-EQA (HM3D)ALL
SPATIALVLM42.944.343.8
SPATIALBOT45.351.049.1
GPT4V57.451.355.3
GPT-4O*68.265.266.7
PHYSVLM-3B60.751.257.4
", + "bbox": [ + 112, + 354, + 460, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Results on Embodied QA", + "text_level": 1, + "bbox": [ + 89, + 497, + 323, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate our model's effectiveness in handling general embodied visual reasoning tasks. Additionally, we show that incorporating an understanding of physical constraints does not diminish its general visual reasoning capabilities. We compare our model with state-of-the-art embodied VLMs (see Tables 2 and 3). Our model achieves the best performance on the RoboVQA-val benchmark, surpassing other models by $7.2\\%$ in BLEU-4. On the OpenEQA benchmark, our model outperforms existing embodied VLMs and GPT-4V, ranking second, behind only GPT-4o.", + "bbox": [ + 88, + 518, + 485, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Results on Robot Task Planning", + "text_level": 1, + "bbox": [ + 89, + 696, + 375, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 presents the performance of PhysVLM and baseline models on real-world task planning scenarios. When all objects are within the robot's physical reach, PhysVLM performs similarly to other models, as directly grabbing or placing the objects succeeds. However, when some objects are outside the physical reach, the model must suggest that the robot move closer before grabbing or placing them. In these cases, our model performs exceptionally well, whereas the task success rates of other models decline significantly. This is attributed to our model's understanding of robotic physical reachability and its ability to incorporate this understanding in task planning.", + "bbox": [ + 88, + 719, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ddf4325ab7f84dbf086c41468d85adf41ecb9d3ac0da06bf098ed35734feceb8.jpg", + "table_caption": [ + "Table 4. Task planning results. Comparison of PhysVLM (ours) with other VLMs." + ], + "table_footnote": [], + "table_body": "
ALL OBJECTS IN RANGEPART OBJECTS IN RANGE
GPT-4O-MINI70.523.2
CLAUDE-3.573.632.1
GPT-4O75.935.8
SPATIALVLM64.421.5
SPATIALBOT65.625.3
PHYSVLM-3B69.248.4
", + "bbox": [ + 540, + 140, + 877, + 268 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/34f2c2e11c1922adda8e1044d30ea755d567d0f868bc922912cf03a52869c843.jpg", + "table_caption": [ + "Table 5. Ablation study on the S-P Map. We compare the constraint encoder's performance when using the S-P Map, replacing it with a Depth Map, or providing no input." + ], + "table_footnote": [], + "table_body": "
IDS-P MAPDEPTH MAPEQA-PHYS REALEQA-PHYS SIM
1✓63.574.8
2✓58.162.4
354.258.8
", + "bbox": [ + 540, + 352, + 877, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 468, + 668, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to evaluate the contribution of each component in PhysVLM. We report the average LLM score on tasks involving physical reachability.", + "bbox": [ + 511, + 491, + 906, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effectiveness of S-P Map. To demonstrate the S-P Map's contribution, we compare results obtained with and without its input. As shown in Table 5, Experiments 1 and 3 demonstrate that omitting the S-P Map leads to a significant performance decrease for both zero-shot real-world robots and simulators. Specifically, the overall average score drops by $16\\%$ in simulation results and $9.3\\%$ in real-world robot evaluations. Without the S-P Map input, the model struggles to handle the robotic physical reachability.", + "bbox": [ + 511, + 559, + 908, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, Experiments 1 and 2 demonstrate that replacing the S-P Map with a Depth Map significantly degrades the model's performance on zero-shot tasks. Since the Depth Map does not accurately represent robotic physical reachability, the model cannot rely solely on depth information to understand it.", + "bbox": [ + 511, + 696, + 908, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of an additional feature encoder. To demonstrate the effectiveness of the model architecture, we compare the performance of a feature encoder that shares weights with the visual feature encoder for the S-P Map. In these experiments, we evaluate the average scores on both the EQA-phys and OpenEQA benchmarks. The results in", + "bbox": [ + 511, + 809, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6946", + "bbox": [ + 482, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55e73a6e3b0b10b01833172bb38edd36b9835b507e965ff9dd727e592c519a18.jpg", + "image_caption": [ + "Figure 4. Visual comparison of PhysVLM (ours), GPT-4o, and SpatialBot." + ], + "image_footnote": [], + "bbox": [ + 117, + 107, + 883, + 333 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6bac1a881dbcbe0a03e69626bf144fa0171cfa6f268acf06f6845af7db9459da.jpg", + "table_caption": [ + "Table 6. Ablation study on the effectiveness of an additional feature encoder. Share indicates shared network and weights with the visual feature encoder." + ], + "table_footnote": [], + "table_body": "
EQA-PHYSOPENEQA
INDEPENDENT71.057.4
SHARE68.256.5
", + "bbox": [ + 129, + 453, + 442, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6 show that sharing the feature encoder not only decreases performance on EQA-phys but also impairs general visual reasoning capabilities. This is because S-P Map features differ from images, and the training data contains significantly more image-text pairs than S-P Map data.", + "bbox": [ + 89, + 541, + 483, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of training data. To evaluate the effectiveness of Phys100K, we conduct experiments by selectively removing data from various sources in Phys100K. As shown in Table 7, removing data from PyBullet or other embodied datasets leads to a reduction in overall performance, highlighting the critical role of each data component in the model's performance.", + "bbox": [ + 89, + 635, + 483, + 742 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d73143653ef2f520880acbe459efbb08d5606c6541cca936738abdcb1974ad20.jpg", + "table_caption": [ + "Table 7. Ablation study on the effectiveness of training data." + ], + "table_footnote": [], + "table_body": "
PART OF PHYS100KEQA-PHYS REALEQA-PHYS SIM
ALL63.574.8
W/O PYBULET62.165.4
W/O OTHER DATASETS58.671.5
", + "bbox": [ + 112, + 792, + 459, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Qualitative Results", + "text_level": 1, + "bbox": [ + 513, + 388, + 694, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 compares our method with SpatialBot and GPT-4o. SpatialBot uses depth maps and images, while GPT-4o uses standard images. Both struggle with tasks requiring physical reachability, causing visual reasoning errors. In contrast, our approach delivers accurate results. Additionally, incorporating the S-P Map into GPT-4o improves its handling of physical reachability and response accuracy.", + "bbox": [ + 511, + 414, + 906, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 540, + 632, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce PhysVLM, a VLM that incorporates physical reachability into visual reasoning for robotic tasks. The S-P Map provides a unified representation of robotic reachability, facilitating the learning of generalizable features. PhysVLM extends traditional VLMs by adding a physical reachability encoder, enabling the simultaneous processing of visual, reachability, and textual information. Additionally, we present EQA-phys, a benchmark for evaluating embodied QA tasks involving physical reachability. Our experiments show that PhysVLM outperforms existing models, achieving a $14\\%$ higher score than GPT-4o on EQA-phys. A limitation is its reduced zero-shot performance on real robots compared to simulations, likely due to the domain gap. Future work will focus on expanding datasets, enhancing real-world performance, and improving the understanding of physical accessibility in vision-language-action models. PhysVLM's reachability awareness supports safer and more reliable robotic decision-making in industrial and assistive settings, while its unified representation ensures cross-platform adaptability for real-world deployment, bridging crucial gaps between environmental perception and actionable robotic intelligence.", + "bbox": [ + 511, + 568, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6947", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by National Key R&D Program of China under Grant No.2022ZD0160601, 2022YFB4300400, and 2018B030322016, National Natural Science Foundation of China under Grants 62176254, 62276260, and U1701266.", + "bbox": [ + 89, + 114, + 483, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6", + "[2] Michael Ahn, Brohan, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2", + "[3] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2", + "[4] Wenxiao Cai, Yaroslav Ponomarenko, Jianhao Yuan, Xiaqi Li, Wankou Yang, Hao Dong, and Bo Zhao. Spatialbot: Precise spatial understanding with vision language models. arXiv preprint arXiv:2406.13642, 2024. 6", + "[5] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2, 6", + "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2", + "[7] Aakanksha Chowdhery, Narang, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1-113, 2023. 2", + "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 4", + "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4", + "[10] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2", + "[11] Kiana Ehsani, Tanmay Gupta, Rose Hendrix, Jordi Salvador, Luca Weihs, Kuo-Hao Zeng, Kunal Pratap Singh, Yejin Kim, Winson Han, Alvaro Herrasti, et al. Spoc: Imitating shortest" + ], + "bbox": [ + 94, + 227, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "paths in simulation enables effective navigation and manipulation in the real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16238-16250, 2024. 2", + "[12] Xiaofeng Gao, Luyao Yuan, Tianmin Shu, Hongjing Lu, and Song-Chun Zhu. Show me what you can do: Capability calibration on reachable workspace for human-robot collaboration. IEEE Robotics and Automation Letters, 7(2):2644-2651, 2022. 3", + "[13] Yining Hong, Zishuo Zheng, Peihao Chen, Yian Wang, Junyan Li, and Chuang Gan. Multiply: A multisensory object-centric embodied large language model in 3d world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26406-26416, 2024. 2", + "[14] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2", + "[15] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 2", + "[16] Lorenzo Jamone, Martim Brandao, Lorenzo Natale, Kenji Hashimoto, Giulio Sandini, and Atsuo Takanishi. Autonomous online generation of a motor representation of the workspace for intelligent whole-body reaching. Robotics and Autonomous Systems, 62(4):556-567, 2014. 2", + "[17] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2", + "[18] Xiaohan Lei, Min Wang, Wengang Zhou, Li Li, and Houqiang Li. Instance-aware exploration-verification-exploitation for instance imagegoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16329–16339, 2024. 2", + "[19] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2", + "[20] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2", + "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[22] Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Lily Lee, Kaichen Zhou, Pengju An, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. Robomamba: Multimodal state space model for efficient robot reasoning and manipulation. arXiv preprint arXiv:2406.04339, 2024. 6", + "[23] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding din: Marrying din with" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6948", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4", + "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6", + "[25] Arjun Majumdar, Ajay, et al. Openeca: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2, 6", + "[26] Matthias Minderer, Alexey Gritsenko, Austin Stone, et al. Simple open-vocabulary object detection with vision transformers, 2022. 2", + "[27] Abby O'Neill, Rehman, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 4", + "[28] Alec Radford, Kim, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[29] Pierre Sermanet, Ding, et al. Robovqa: Multimodal long-horizon reasoning for robotics. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 645-652. IEEE, 2024. 6", + "[30] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2", + "[31] Chen Song and Darwin Lau. Workspace-based model predictive control for cable-driven robots. IEEE Transactions on Robotics, 38(4):2577-2596, 2022. 2, 3", + "[32] Giovanni Sutanto, Austin Wang, Yixin Lin, Mustafa Mukadam, Gaurav Sukhatme, Akshara Rai, and Franziska Meier. Encoding physical constraints in differentiable newton-euler algorithm. In Proceedings of the 2nd Conference on Learning for Dynamics and Control, pages 804-813. PMLR, 2020. 2", + "[33] Qwen Team. qwen2.5, 2024. 4", + "[34] Dominique Thomas, Rodney Rothstein, Nathan Rosenberg, and Yolande Surdin-Kerjan. Sam2 encodes the second methionine s-adenosyl transferase in saccharomyces cerevisiae: physiology and regulation of both enzymes. Molecular and Cellular Biology, 8(12):5132-5139, 1988. 4", + "[35] Yunsheng Tian, Karl DD Willis, Bassel Al Omari, Jieliang Luo, Pingchuan Ma, Yichen Li, Farhad Javid, Edward Gu, Joshua Jacob, Shinjiro Sueda, et al. Asap: Automated sequence planning for complex robotic assembly with physical feasibility. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4380-4386. IEEE, 2024. 2", + "[36] An Dinh Vuong, Minh Nhat Vu, Baoru Huang, Nghia Nguyen, Hieu Le, Thieu Vo, and Anh Nguyen. Language-driven grasp detection. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 17902-17912, 2024. 2", + "[37] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2", + "[38] Yu Wu, Lu Jiang, and Yi Yang. Revisiting embodiedqa: A simple baseline and beyond. IEEE Transactions on Image Processing, 29:3984-3992, 2020. 2", + "[39] An Yang, Baosong Yang, Binyuan Hui, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 4", + "[40] Yijun Yang, Tianyi Zhou, Kanxue Li, Dapeng Tao, Lusong Li, Li Shen, Xiaodong He, Jing Jiang, and Yuhui Shi. Embodied multi-modal agent trained by an llm from a parallel textworld. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26275-26285, 2024. 2", + "[41] Franziska Zacharias, Christoph Borst, and Gerd Hirzinger. Capturing robot workspace structure: representing robot capabilities. In 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3229-3236. IEEE, 2007. 2", + "[42] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023. 4", + "[43] Ganlong Zhao, Guanbin Li, Weikai Chen, and Yizhou Yu. Over-nav: Elevating iterative vision-and-language navigation with open-vocabulary detection and structured representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16296–16306, 2024. 2", + "[44] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 2, 6" + ], + "bbox": [ + 516, + 92, + 903, + 599 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6949", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/14176105-f8ea-4fe1-b4ab-02889a03b21d_model.json b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/14176105-f8ea-4fe1-b4ab-02889a03b21d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..55d1350160fb0eacb2f6cf9b5195b9e69585a4b5 --- /dev/null +++ b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/14176105-f8ea-4fe1-b4ab-02889a03b21d_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.895, + 0.178 + ], + "angle": 0, + "content": "PhysVLM: Enabling Visual Language Models to Understand Robotic Physical Reachability" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.203, + 0.747, + 0.24 + ], + "angle": 0, + "content": "Weijie Zhou\\(^{1,2}\\), Manli Tao\\(^{2}\\), Chaoyang Zhao\\(^{2,3,\\dagger}\\), Haiyun Guo\\(^{2}\\), Honghui Dong\\(^{1}\\), Ming Tang\\(^{2}\\), Jinqiao Wang\\(^{1,2,3,4,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.239, + 0.762, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Traffic and Transportation, Beijing Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.257, + 0.867, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\) Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.275, + 0.879, + 0.311 + ], + "angle": 0, + "content": "3 ObjectEye Inc. Guangdong Provincial Key Laboratory of Intellectual Property & Big Data, Guangdong Polytechnic Normal University" + }, + { + "type": "list", + "bbox": [ + 0.119, + 0.257, + 0.879, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.313, + 0.825, + 0.327 + ], + "angle": 0, + "content": "†Corresponding authors: chaoyang.zhao@nlpr.ia.ac.cn, jqwang@nlpr.ia.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.362, + 0.327, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.415, + 0.486, + 0.884 + ], + "angle": 0, + "content": "Understanding the environment and a robot's physical reachability is crucial for task execution. While state-of-the-art vision-language models (VLMs) excel in environmental perception, they often generate inaccurate or impractical responses in embodied visual reasoning tasks due to a lack of understanding of robotic physical reachability. To address this issue, we propose a unified representation of physical reachability across diverse robots, i.e., Space-Physical Reachability Map (S-P Map), and PhysVLM, a vision-language model that integrates this reachability information into visual reasoning. Specifically, the S-P Map abstracts a robot's physical reachability into a generalized spatial representation, independent of specific robot configurations, allowing the model to focus on reachability features rather than robot-specific parameters. Subsequently, PhysVLM extends traditional VLM architectures by incorporating an additional feature encoder to process the S-P Map, enabling the model to reason about physical reachability without compromising its general vision-language capabilities. To train and evaluate PhysVLM, we constructed a large-scale multi-robot dataset, Phys100K, and a challenging benchmark, EQA-phys, which includes tasks for six different robots in both simulated and real-world environments. Experimental results demonstrate that PhysVLM outperforms existing models, achieving a \\(14\\%\\) improvement over GPT-4o on EQA-phys and surpassing advanced embodied VLMs such as RoboMamba and SpatialVLM on the RoboVQA-val and OpenEQA benchmarks. Additionally, the S-P Map shows strong compatibility with various VLMs, and its integration into GPT-4o-mini yields a \\(7.1\\%\\) performance improvement." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.362, + 0.645, + 0.378 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.42, + 0.709, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.422, + 0.902, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.538, + 0.804, + 0.551 + ], + "angle": 0, + "content": "Embodied QA Can robot pick up the bowl now?" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.555, + 0.574, + 0.567 + ], + "angle": 0, + "content": "GPT40:" + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.556, + 0.761, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.568, + 0.845, + 0.579 + ], + "angle": 0, + "content": "Yes, the bowl is placed on a flat surface and is accessible." + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.58, + 0.591, + 0.592 + ], + "angle": 0, + "content": "PhysVLM:" + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.58, + 0.761, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.593, + 0.851, + 0.605 + ], + "angle": 0, + "content": "No, the bowl is not within the robot arm's reachable space." + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.613, + 0.61, + 0.626 + ], + "angle": 0, + "content": "Task Planning" + }, + { + "type": "text", + "bbox": [ + 0.616, + 0.612, + 0.89, + 0.636 + ], + "angle": 0, + "content": "Task goal is to put the can into the bowl, what are the next steps?" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.64, + 0.573, + 0.651 + ], + "angle": 0, + "content": "GPT40:" + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.641, + 0.761, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.764, + 0.642, + 0.845, + 0.652 + ], + "angle": 0, + "content": "Execution failure" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.653, + 0.787, + 0.664 + ], + "angle": 0, + "content": "Picking up the can and place it inside the bowl." + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.664, + 0.591, + 0.676 + ], + "angle": 0, + "content": "PhysVLM:" + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.665, + 0.761, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.677, + 0.879, + 0.7 + ], + "angle": 0, + "content": "1-Move close to the can. 2-Pick up the can. 3-Place the can into the bowl. 4-Down." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.907, + 0.784 + ], + "angle": 0, + "content": "Figure 1. Existing VLM models, such as GPT-4o, may generate inaccurate or impractical responses due to poor comprehension of robotic physical reachability. The proposed PhysVLM integrates vision-language capabilities with an understanding of robotic physical reachability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Accurate perception of physical reachability is essential for robots to perform tasks effectively. Similar to how humans adjust their actions based on bodily conditions and environmental factors, robots must account for their physical reachability within an environment to ensure efficient and reliable task execution. For instance, in grasping tasks," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6940" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "a robot that fails to assess its reachability may attempt to grasp an object from an unreachable position, leading to task failure or equipment damage [32, 35]. Thus, enhancing a robot's understanding of physical reachability is crucial for successful task planning and execution in complex environments [31, 41]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.185, + 0.485, + 0.502 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) have shown remarkable progress in environmental understanding [6, 17, 19, 21, 37], and many studies have applied these models to embodied AI to assist robots in perceiving environments and planning tasks [11, 18, 25, 36, 43]. However, while VLMs excel in general environmental perception, they often struggle with tasks that require an understanding of robotic physical reachability (see Figure 1). We identify two key challenges that must be addressed for VLMs to be effective in robotic tasks: (1) how to develop a unified and efficient representation of physical reachability. Robots vary significantly in size, joint types, and other characteristics, making it difficult for VLMs to directly learn these differences; (2) how to enable VLMs to improve their understanding of physical reachability without compromising general vision-language capabilities. Existing VLMs typically combine pre-trained unimodal encoders for vision and language tasks. However, introducing a new modality like physical reachability requires careful architectural and training adjustments to ensure the model can reason about reachability while maintaining its general capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.82 + ], + "angle": 0, + "content": "To address these challenges, we propose the Space-Physical Reachability Map (S-P Map), a unified representation that abstracts the physical reachability of diverse robots into a generalized spatial form. The S-P Map is generated by combining robot parameters with egocentric depth images, but crucially, the model learns to focus on the abstracted reachability features (i.e., the gray regions in the S-P Map) rather than the specific robot configurations. This abstraction allows the model to generalize across different robots, as it only needs to reason about which areas are reachable, independent of the robot's specific characteristics. We introduce PhysVLM, a vision-language model that extends traditional VLM architectures by incorporating an additional feature encoder to process the S-P Map. This design enables PhysVLM to integrate physical reachability information into its reasoning process without compromising its general vision-language capabilities. To train and evaluate PhysVLM, we constructed a large-scale multi-robot dataset, Phys100K, and a challenging benchmark, EQAphys, which includes tasks for six different robots in both simulated and real-world environments." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.824, + 0.403, + 0.837 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose a unified and robot-agnostic formulation, the Space-Physical Reachability Map (S-P Map), which abstracts robotic physical reachability in a way that is independent of specific robot configurations, promoting the" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.744, + 0.106 + ], + "angle": 0, + "content": "learning of generalized features." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.167 + ], + "angle": 0, + "content": "- We introduce PhysVLM, a vision-language model that integrates physical reachability with general vision-language capabilities via an additional feature encoder, improving task execution reliability." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- We release the EQA-phys benchmark, which includes six robots and 1.3K question-answer pairs, designed to test the model's understanding of physical reachability in simulated and real-world environments." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.228, + 0.905, + 0.348 + ], + "angle": 0, + "content": "- Our model achieves a \\(14\\%\\) improvement over GPT-4o on the EQA-phys benchmark. In embodied visual reasoning tasks on the RoboVQA-val and OpenEQA benchmarks, it outperforms advanced embodied VLMs such as RoboMamba and SpatialVLM. Furthermore, the S-P Map demonstrates strong compatibility with various VLMs, and integrating it into GPT-4o-mini results in a \\(7.1\\%\\) performance improvement." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.364, + 0.655, + 0.379 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.389, + 0.691, + 0.404 + ], + "angle": 0, + "content": "2.1. VLMs in Robotics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.905, + 0.517 + ], + "angle": 0, + "content": "Embodied Question Answering (EQA) tasks require agents to interact with environments to answer questions [10, 13, 38]. RoboVQA offers a large, diverse dataset for robotic visual question answering. 3D-VLA [44] integrates 3D perception with a generative world model for embodied reasoning, while SpatialVLM [5] enhances VLMs' spatial understanding using extensive 3D data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Robot task planning involves sequencing subtasks to achieve goals [3, 30, 40]. Code as Policies (CaP) [20] employs OpenAI's Codex (code-davinci-002) to generate planning code. SayCan [2] combines the Pathways Language Model (PaLM) [7] with robotic affordances to create feasible action plans based on the robot's capabilities. However, these methods often assume all objects are within the robot's operational area, ignoring physical reachability and potentially leading to suboptimal or infeasible plans." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.834, + 0.683 + ], + "angle": 0, + "content": "2.2. Understanding Physical Reachability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Recent studies use voxel grids with open-vocabulary detection models to assign task-specific attributes, enabling environmental constraint understanding. ReKep [15] generates keypoint proposals and constraints using voxel grids and VLMs, while VoxPoser [14] synthesizes robot trajectories by integrating OWL-ViT [26] and VLMs with voxel-based environment representations. However, these approaches focus on environment modeling without explicitly addressing the robot's physical reachability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Explicitly representing reachable workspaces remains challenging. Reachability maps [41] model spatial capabilities, and occupancy grids [16] account for obstacles to ensure safe navigation. Additionally, methods like online model predictive control with offline workspace analysis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6941" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.184 + ], + "angle": 0, + "content": "[31] and Reachability Expression-based Motion Planning (REMP) [12] address workspace constraints. Despite these advancements, integrating physical reachability into visual reasoning for complex embodied tasks is still limited, primarily due to the lack of large-scale datasets that include robotic physical parameters in VLM pretraining." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.197, + 0.182, + 0.212 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.484, + 0.463 + ], + "angle": 0, + "content": "PhysVLM is a large-scale vision-language model designed for visual reasoning that accounts for physical constraints in embodied tasks. As illustrated in Figure 2, PhysVLM integrates instruction text, visual input (RGB image), and an S-P Map abstracts the robotic physical reachability into a unified spatial representation. By combining these inputs, PhysVLM generates responses consistent with both the visual context and the robot's physical reachability, without being tied to specific robot configurations. The S-P Map is constructed using a unified physical reachability encoding method, which abstracts the physical parameters of various robots alongside their egocentric depth maps into a generalized form. This abstraction allows the model to generalize across different robots, addressing the challenge of learning and reasoning about physical reachability in a robot-agnostic manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.526 + ], + "angle": 0, + "content": "In this section, we present the core components of PhysVLM. Section 3.1 describes the S-P Map encoding method, Section 3.2 describes the model architecture, and Section 3.3 discusses the training procedures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.273, + 0.551 + ], + "angle": 0, + "content": "3.1. S-P Map Encoding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.648 + ], + "angle": 0, + "content": "As illustrated in Figure 2, we model the physical reachability of various robots using a unified approach that abstracts robot-specific parameters into a generalized spatial representation. This abstraction allows the model to focus on the spatial regions that are physically reachable, independent of the specific robot configuration." + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.675, + 0.483, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S} - \\mathrm {P} \\quad \\operatorname {M a p} = F \\left(\\mathcal {P} _ {\\text {r a w}}, \\left\\{\\theta_ {i} ^ {\\min }, \\theta_ {i} ^ {\\max } \\right\\}, \\mathrm {D H}, \\mathbf {E}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}_{\\mathrm{raw}}\\) denotes the raw point cloud data from the robot's RGB-D camera. \\(\\{\\theta_i^{\\mathrm{min}},\\theta_i^{\\mathrm{max}}\\}\\) represents the range of motion for each joint \\(i\\). DH refers to the Denavit-Hartenberg parameters that describe the geometric structure of each joint, and \\(\\mathbf{E}\\) is the extrinsic calibration matrix that transforms coordinates from the camera to the robot's coordinate system. The function \\(F\\) maps these inputs to produce the S-P Map, which abstracts the robot's physical reachability into a spatial form that is independent of the specific robot configuration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Consider a robot arm with \\( n \\) degrees of freedom, where each joint \\( i \\) has DH parameters \\( \\{\\theta_i, d_i, a_i, \\alpha_i\\} \\) (\\( \\theta_i \\) is the joint angle, \\( d_i \\) is the offset along the z-axis, \\( a_i \\) is the link" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "length, and \\(\\alpha_{i}\\) is the twist angle). The homogeneous transformation matrix for each joint is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.137, + 0.905, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {i} = G \\left(\\theta_ {i}, d _ {i}, a _ {i}, \\alpha_ {i}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.159, + 0.907, + 0.219 + ], + "angle": 0, + "content": "where \\( G \\) is the standard Denavit-Hartenberg transformation function. By multiplying the transformation matrices of all joints, we obtain the transformation matrix from the base frame to the end-effector frame:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.234, + 0.905, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} = \\mathbf {T} _ {1} \\mathbf {T} _ {2} \\dots \\mathbf {T} _ {n}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.256, + 0.906, + 0.332 + ], + "angle": 0, + "content": "To generate joint configurations, we sample the joint angles \\(\\theta_{i}\\) from their respective motion ranges \\(\\theta_{i} \\in [\\theta_{i}^{\\min}, \\theta_{i}^{\\max}]\\), resulting in configurations \\(\\{\\theta_{1}, \\theta_{2}, \\ldots, \\theta_{n}\\}\\). By substituting these joint configurations into the forward kinematics equations, we compute the corresponding end-effector positions:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.352, + 0.905, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} _ {\\text {v o x e l}} = \\left\\{\\mathbf {p} \\mid \\mathbf {p} = \\mathbf {T} \\left(\\theta_ {1}, \\theta_ {2}, \\dots , \\theta_ {n}\\right) \\cdot \\mathbf {p} _ {0} \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.395, + 0.905, + 0.454 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_0\\) is the origin point in the end-effector frame. We precompute these joint configurations offline, discretize the workspace into a voxel grid \\(\\mathcal{W}_{\\mathrm{voxel}}\\), and store it for efficient computation in subsequent steps." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.455, + 0.906, + 0.53 + ], + "angle": 0, + "content": "Next, as shown in Figure 2, the robot's raw point cloud \\(\\mathcal{P}_{\\mathrm{raw}}\\) is captured from its egocentric RGB-D camera in the camera coordinate system and transformed into the robot's coordinate system using the extrinsic calibration matrix \\(\\mathbf{E}\\), resulting in the transformed point cloud \\(\\mathcal{P}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.661, + 0.545, + 0.905, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} = \\mathbf {E} \\cdot \\mathcal {P} _ {\\text {r a w}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.613 + ], + "angle": 0, + "content": "To ensure physical feasibility, we perform a voxel grid lookup to determine whether each point in \\(\\mathcal{P}\\) lies within the precomputed reachable workspace \\(\\mathcal{W}_{\\mathrm{voxel}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.622, + 0.905, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {\\text {v a l i d}} = \\left\\{\\mathbf {p} \\in \\mathcal {P} \\mid \\mathbf {p} \\in \\mathcal {W} _ {\\text {v o x e l}} \\right\\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.705 + ], + "angle": 0, + "content": "This step filters the point cloud to include only points within the robot's reachable workspace, abstracting the robot's physical reachability into a generalized spatial form." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Finally, we transform the valid point cloud \\(\\mathcal{P}_{\\mathrm{valid}}\\) back into the camera coordinate system and use the camera's intrinsic parameters to project these points onto the image plane. We then mark the regions that comply with physical reachability on the original depth map. For areas that are not reachable, we apply a gray mask and outline their boundaries. The resulting S-P Map clearly highlights regions that are beyond the robot's physical reach, providing a unified and abstracted representation of reachability that is independent of the specific robot configuration. This allows the model to focus on the spatial constraints of the task, without needing to account for the detailed physical parameters of each robot." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6942" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.109, + 0.418, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.109, + 0.904, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.355, + 0.908, + 0.399 + ], + "angle": 0, + "content": "Figure 2. Overview of PhysVLM. Starting with robot parameters and an egocentric depth map, an S-P Map is generated via unified physical reachability encoding. Using the S-P Map, image, and instruction text, PhysVLM generates textual output considering the robot's physical reachability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.429, + 0.279, + 0.443 + ], + "angle": 0, + "content": "3.2. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.484, + 0.576 + ], + "angle": 0, + "content": "To seamlessly integrate robotic physical reachability into PhysVLM while preserving its visual reasoning capabilities, we design a dual-branch architecture: one branch dedicated to vision processing and the other to physical reachability (see Figure 2). These branches operate independently, extracting features from their respective inputs, which are then fused and passed to a unified decoder for final reasoning and response generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.484, + 0.683 + ], + "angle": 0, + "content": "The vision branch leverages a pre-trained Vision Transformer (ViT) [9], specifically the SigLip-400M model [42], to extract high-level visual features from egocentric images. To reduce computational overhead, a Max Pooling layer is applied, followed by a two-layer Multi-Layer Perceptron (MLP) that transforms the visual features into token representations suitable for multimodal fusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.806 + ], + "angle": 0, + "content": "The physical reachability branch processes the S-P Map, which abstracts the robot's physical reachability into a generalized spatial form. This branch also utilizes the SigLip-400M model for feature extraction, followed by Max Pooling and a feature fusion layer. The fusion layer combines the visual and reachability features, and a two-layer MLP further refines these fused features into reachability-specific tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For the language decoding, we employ the Qwen-2.5-Instruct-3B model [33, 39] as PhysVLM's large language model (LLM) decoder, using the Qwen-2.5 tokenizer to process natural language instructions. The decoder integrates multimodal tokens from the vision branch, S-P Maps, and language inputs, generating coherent and contextually rele" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.43, + 0.907, + 0.46 + ], + "angle": 0, + "content": "vant textual responses that account for both visual and physical reachability information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.618, + 0.485 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.613 + ], + "angle": 0, + "content": "Training Data Construction. The training data for PhysVLM consists of our Phys100K dataset and general VQA datasets, such as LLaVA-Pretrain, ShareGPT4V, and RoboVQA. Phys100K focuses on question-answering related to physical reachability, aggregating data from RoboVQA (20K samples), ScanNet [8] (10K samples), OpenX-Embodiment [27] (60K samples), and an additional 10K samples from PyBullet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Depth maps are essential inputs for generating the S-P Map. For datasets lacking depth maps, we generate them using DepthAnything-v2. Additionally, we employ Grounding DINO [23] and SAM2 [34] to obtain 2D bounding boxes and segmentation results for objects in the images. In PyBullet, we simulate work scenarios using four robotic arms (UR5, FR5, CR5, and FRANKA) to collect RGB images, depth maps, and segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "For the PyBullet data in Phys100K, precise robot configurations can be obtained from the simulator. Therefore, we directly generate the S-P Map using the method described in Section 3.2, and labels indicating whether objects are reachable are obtained through simulated motion. The advantage of the S-P Map is that it abstracts physical reachability into a region-based representation, decoupling the learning process from specific robot configurations. This allows us to generate pseudo-labels for datasets without precise robot parameters. We approximate reachability using the segmentation results, marking regions and the objects within" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6943" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.109, + 0.907, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.423, + 0.708, + 0.437 + ], + "angle": 0, + "content": "Figure 3. Details of the Phys100K Dataset and EQA-Phys Benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.515 + ], + "angle": 0, + "content": "them as \"reachable\" or \"unreachable\" based on depth values. Next, we generate question-answer pairs for two main categories:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.518, + 0.483, + 0.623 + ], + "angle": 0, + "content": "- Embodied QA. GPT-4 generates question-answer pairs for ScanNet and RoboVQA, covering categories include Function Reasoning, World Knowledge, Object Recognition, Object Localization, Attribute Recognition, Spatial Reasoning, Object State Recognition, and Hallucination (See Figure 3). Detailed prompts and examples are provided in the appendix to guide the generation process." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.624, + 0.483, + 0.79 + ], + "angle": 0, + "content": "- Tasks Involving Physical Reachability. We use the \"reachable\" label with five fixed task templates to generate question-answer pairs, such as \"USER:\\n\\n Is the [Object] in the robot's reachable space? ASSISTANT: Yes, it is.\" Here, [Object] represents the relevant object category, while and serve as placeholders for image patch tokens and S-P Map patch tokens, respectively. Figure 3 provides examples of question-answer pairs for each object category." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.518, + 0.483, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Training Pipeline. We adopt a two-stage training process to fully leverage the S-P Map and ensure PhysVLM generalizes across different robots. In the first stage, we align multimodal features using the LLaVA-Pretrain and OpenX-Embodiment datasets from Phys100K. This stage only trains the projection layers, allowing the model to build" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.469, + 0.907, + 0.5 + ], + "angle": 0, + "content": "a foundational understanding of visual inputs and physical reachability, independent of specific robot configurations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.593 + ], + "angle": 0, + "content": "In the second stage, we unfreeze all parameters and train the entire model using data from Phys100K, ShareGPT4V, and RoboVQA. This stage enhances PhysVLM's ability to handle complex visual reasoning tasks with physical reachability constraints, ensuring the model can generalize across diverse environments and robots." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.907, + 0.721 + ], + "angle": 0, + "content": "Implementation Details. PhysVLM is trained for 48 hours using eight A800 GPUs. The training process consists of two stages, each lasting one epoch. The batch size and learning rate are set at 128 and 1e-3 in the first stage, and 64 and 1e-5 in the second stage. The final model is PhysVLM-3B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.725, + 0.757 + ], + "angle": 0, + "content": "3.4. EQA-phys Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Figure 3, we introduce an embodied QA task focused on physical reachability, termed EQA-phys, which emphasizes QA tasks constrained by physical limitations. This benchmark includes a simulator dataset with 200 samples and 1,000 questions from the PyBullet validation set, as well as a zero-shot evaluation set based on real-world data from UR3 and XArm6 robots in two scenarios. The evaluation set contains 60 samples and 300 questions, all manually annotated by domain experts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6944" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.163, + 0.089, + 0.835, + 0.104 + ], + "angle": 0, + "content": "Table 1. Results of EQA-phys. Comparison of PhysVLM-3B (ours) with API-based VLMs and embodied VLMs." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.128, + 0.894, + 0.304 + ], + "angle": 0, + "content": "
REAL-WORLDSIMULATORALL
UR3XARM6UR5FR5CR5FRANKA
API-BASED VLMSGPT-4O-MINI54.356.049.455.454.647.152.8
CLAUDE-3.556.260.554.058.155.754.356.4
GPT-4O56.761.555.758.357.552.657.0
GPT-4O-MINI + S-P MAP60.0↑5.760.5↑4.557.0↑7.659.1↑3.759.2↑4.653.3↑6.259.8↑7.0
CLAUDE-3.5 + S-P MAP65.3↑9.167.3↑6.854.9↑0.958.3↑0.258.2↑2.558.1↑3.860.3↑3.4
GPT-4O + S-P MAP66.6↑9.968.1↑6.655.8↑0.160.7↑1.459.4↑1.957.6↑5.061.3↑4.1
EMBODIED VLMSSPATIALVLM56.355.154.659.152.047.554.1
SPATIALBOT51.150.250.048.153.354.451.1
PHYSVLM-3B64.163.071.475.774.078.171.0
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.33, + 0.223, + 0.347 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.355, + 0.292, + 0.372 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.483, + 0.408 + ], + "angle": 0, + "content": "Tasks. We compare the performance of PhysVLM with other methods across three categories of tasks:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.411, + 0.483, + 0.485 + ], + "angle": 0, + "content": "- EQA-phys. This benchmark tests the model's ability to integrate visual reasoning with robotic physical reachability. The real-robot component is used to assess PhysVLM's zero-shot generalization, highlighting its ability to handle unseen robots and environments." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.486, + 0.483, + 0.53 + ], + "angle": 0, + "content": "- Embodied QA. We evaluate the model's general visual reasoning ability in embodied tasks using the OpenEQA [25] and RoboVQA-val [29] benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.531, + 0.484, + 0.622 + ], + "angle": 0, + "content": "- Robot Task Planning. For real-world tasks such as \"Pick A into B\", we assess the model's ability to understand robotic physical reachability and generate reasonable task plans. As this study does not focus on robot control strategies, we use the natural language planning approach from [29]." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.411, + 0.484, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.777 + ], + "angle": 0, + "content": "Baselines. We compare our model to several baselines, including API-accessible VLMs such as Claude 3.5 [28], GPT-4o-mini [1], and GPT-4o [1], as well as embodied VLMs like SpatialVLM [5], SpatialBot [4], 3D-VLA [44], and RoboMamba [22]. SpatialVLM and SpatialBot both use the 3B version, which has a similar parameter count to our model. Since the executable versions of 3D-VLA and RoboMamba are unavailable, we compare their reported results on RoboVQA-val." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. For tasks involving physical reachability, we use LLM scoring, following the approach used in existing studies [24, 25]. Assigning 5 points for completely correct responses and 1 point for incorrect responses, we calculate the average score and express it as a percentage. For Embodied QA, we follow the benchmark settings of the corresponding datasets [25, 29]. For task planning, each" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.332, + 0.906, + 0.362 + ], + "angle": 0, + "content": "task type is executed 10 times, and the average success rate serves as the evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.391, + 0.715, + 0.408 + ], + "angle": 0, + "content": "4.2. Results on EQA-phys" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.571 + ], + "angle": 0, + "content": "Table 1 shows the results on EQA-phys. Neither API-based nor embodied VLMs can handle the robot's parameter constraints, resulting in suboptimal outputs with scores around \\(55\\%\\). In contrast, our model successfully completes visual reasoning tasks involving physical reachability, obtaining an average score of \\(71\\%\\). As discussed, these tasks require the model to perform visual reasoning based on an understanding of the robotic physical reachability. The model can only effectively accomplish these tasks by truly understanding robotic physical reachability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.909, + 0.744 + ], + "angle": 0, + "content": "The results in Table 1 demonstrate that prompting API-based VLMs, such as GPT-40, with the S-P Map (detailed in 4) significantly enhances their performance. This improvement stems from the S-P Map's ability to abstract physical reachability into a robot-agnostic representation, enabling VLMs to reason about physical constraints that would otherwise be beyond their capabilities. By decoupling reachability from specific robot parameters, the S-P Map facilitates generalization across diverse environments, allowing models to better comprehend physical reachability, even in previously unseen scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Table 1 also demonstrates that PhysVLM-3B achieved scores of over \\(63\\%\\) in zero-shot evaluations for UR3 and XArm6 robots, despite operating in new environments with different robot parameters. This performance is attributed to two key factors: (1) the S-P Map abstracts various robot parameters into a unified, transferable representation of physical reachability, and (2) the model's independent visual and constraint encoding branches allow it to learn generalizable visual features from diverse image-text data, enabling effective reasoning in novel environments." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6945" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 2. Embodied QA results on the RoboVQA-val set, comparison of PhysVLM (ours) with existing methods. An asterisk (*) indicates models not pre-trained on the RoboVQA dataset." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.156, + 0.49, + 0.259 + ], + "angle": 0, + "content": "
BLEU1BLEU2BLEU3BLEU4
SPATIALVLM*5.13.01.91.2
SPATIALBOT*12.49.38.07.2
3D-VLA48.338.531.726.8
ROBOMAMBA54.944.239.536.3
PHYSVLM-3B65.362.450.943.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.275, + 0.485, + 0.332 + ], + "angle": 0, + "content": "Table 3. Embodied QA results on the OpenEQA benchmark, comparison of PhysVLM (ours) with existing methods. An asterisk (*) indicates that only the first 200 samples were tested due to API limitations." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.355, + 0.462, + 0.471 + ], + "angle": 0, + "content": "
EM-EQA (SCANNET)EM-EQA (HM3D)ALL
SPATIALVLM42.944.343.8
SPATIALBOT45.351.049.1
GPT4V57.451.355.3
GPT-4O*68.265.266.7
PHYSVLM-3B60.751.257.4
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.498, + 0.325, + 0.515 + ], + "angle": 0, + "content": "4.3. Results on Embodied QA" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.52, + 0.486, + 0.687 + ], + "angle": 0, + "content": "We demonstrate our model's effectiveness in handling general embodied visual reasoning tasks. Additionally, we show that incorporating an understanding of physical constraints does not diminish its general visual reasoning capabilities. We compare our model with state-of-the-art embodied VLMs (see Tables 2 and 3). Our model achieves the best performance on the RoboVQA-val benchmark, surpassing other models by \\(7.2\\%\\) in BLEU-4. On the OpenEQA benchmark, our model outperforms existing embodied VLMs and GPT-4V, ranking second, behind only GPT-4o." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.698, + 0.376, + 0.715 + ], + "angle": 0, + "content": "4.4. Results on Robot Task Planning" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Table 4 presents the performance of PhysVLM and baseline models on real-world task planning scenarios. When all objects are within the robot's physical reach, PhysVLM performs similarly to other models, as directly grabbing or placing the objects succeeds. However, when some objects are outside the physical reach, the model must suggest that the robot move closer before grabbing or placing them. In these cases, our model performs exceptionally well, whereas the task success rates of other models decline significantly. This is attributed to our model's understanding of robotic physical reachability and its ability to incorporate this understanding in task planning." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 4. Task planning results. Comparison of PhysVLM (ours) with other VLMs." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.141, + 0.878, + 0.27 + ], + "angle": 0, + "content": "
ALL OBJECTS IN RANGEPART OBJECTS IN RANGE
GPT-4O-MINI70.523.2
CLAUDE-3.573.632.1
GPT-4O75.935.8
SPATIALVLM64.421.5
SPATIALBOT65.625.3
PHYSVLM-3B69.248.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.287, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Table 5. Ablation study on the S-P Map. We compare the constraint encoder's performance when using the S-P Map, replacing it with a Depth Map, or providing no input." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.353, + 0.878, + 0.44 + ], + "angle": 0, + "content": "
IDS-P MAPDEPTH MAPEQA-PHYS REALEQA-PHYS SIM
1✓63.574.8
2✓58.162.4
354.258.8
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.669, + 0.485 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.539 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to evaluate the contribution of each component in PhysVLM. We report the average LLM score on tasks involving physical reachability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.909, + 0.697 + ], + "angle": 0, + "content": "The effectiveness of S-P Map. To demonstrate the S-P Map's contribution, we compare results obtained with and without its input. As shown in Table 5, Experiments 1 and 3 demonstrate that omitting the S-P Map leads to a significant performance decrease for both zero-shot real-world robots and simulators. Specifically, the overall average score drops by \\(16\\%\\) in simulation results and \\(9.3\\%\\) in real-world robot evaluations. Without the S-P Map input, the model struggles to handle the robotic physical reachability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.909, + 0.788 + ], + "angle": 0, + "content": "Additionally, Experiments 1 and 2 demonstrate that replacing the S-P Map with a Depth Map significantly degrades the model's performance on zero-shot tasks. Since the Depth Map does not accurately represent robotic physical reachability, the model cannot rely solely on depth information to understand it." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of an additional feature encoder. To demonstrate the effectiveness of the model architecture, we compare the performance of a feature encoder that shares weights with the visual feature encoder for the S-P Map. In these experiments, we evaluate the average scores on both the EQA-phys and OpenEQA benchmarks. The results in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "6946" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.108, + 0.885, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.345, + 0.72, + 0.358 + ], + "angle": 0, + "content": "Figure 4. Visual comparison of PhysVLM (ours), GPT-4o, and SpatialBot." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.389, + 0.483, + 0.43 + ], + "angle": 0, + "content": "Table 6. Ablation study on the effectiveness of an additional feature encoder. Share indicates shared network and weights with the visual feature encoder." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.454, + 0.443, + 0.514 + ], + "angle": 0, + "content": "
EQA-PHYSOPENEQA
INDEPENDENT71.057.4
SHARE68.256.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.618 + ], + "angle": 0, + "content": "Table 6 show that sharing the feature encoder not only decreases performance on EQA-phys but also impairs general visual reasoning capabilities. This is because S-P Map features differ from images, and the training data contains significantly more image-text pairs than S-P Map data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.484, + 0.743 + ], + "angle": 0, + "content": "Effectiveness of training data. To evaluate the effectiveness of Phys100K, we conduct experiments by selectively removing data from various sources in Phys100K. As shown in Table 7, removing data from PyBullet or other embodied datasets leads to a reduction in overall performance, highlighting the critical role of each data component in the model's performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.755, + 0.466, + 0.77 + ], + "angle": 0, + "content": "Table 7. Ablation study on the effectiveness of training data." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.793, + 0.46, + 0.88 + ], + "angle": 0, + "content": "
PART OF PHYS100KEQA-PHYS REALEQA-PHYS SIM
ALL63.574.8
W/O PYBULET62.165.4
W/O OTHER DATASETS58.671.5
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.39, + 0.696, + 0.405 + ], + "angle": 0, + "content": "4.6. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Figure 4 compares our method with SpatialBot and GPT-4o. SpatialBot uses depth maps and images, while GPT-4o uses standard images. Both struggle with tasks requiring physical reachability, causing visual reasoning errors. In contrast, our approach delivers accurate results. Additionally, incorporating the S-P Map into GPT-4o improves its handling of physical reachability and response accuracy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.541, + 0.633, + 0.557 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce PhysVLM, a VLM that incorporates physical reachability into visual reasoning for robotic tasks. The S-P Map provides a unified representation of robotic reachability, facilitating the learning of generalizable features. PhysVLM extends traditional VLMs by adding a physical reachability encoder, enabling the simultaneous processing of visual, reachability, and textual information. Additionally, we present EQA-phys, a benchmark for evaluating embodied QA tasks involving physical reachability. Our experiments show that PhysVLM outperforms existing models, achieving a \\(14\\%\\) higher score than GPT-4o on EQA-phys. A limitation is its reduced zero-shot performance on real robots compared to simulations, likely due to the domain gap. Future work will focus on expanding datasets, enhancing real-world performance, and improving the understanding of physical accessibility in vision-language-action models. PhysVLM's reachability awareness supports safer and more reliable robotic decision-making in industrial and assistive settings, while its unified representation ensures cross-platform adaptability for real-world deployment, bridging crucial gaps between environmental perception and actionable robotic intelligence." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6947" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.191 + ], + "angle": 0, + "content": "This work was supported by National Key R&D Program of China under Grant No.2022ZD0160601, 2022YFB4300400, and 2018B030322016, National Natural Science Foundation of China under Grants 62176254, 62276260, and U1701266." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.482, + 0.296 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.299, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[2] Michael Ahn, Brohan, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.342, + 0.483, + 0.41 + ], + "angle": 0, + "content": "[3] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[4] Wenxiao Cai, Yaroslav Ponomarenko, Jianhao Yuan, Xiaqi Li, Wankou Yang, Hao Dong, and Bo Zhao. Spatialbot: Precise spatial understanding with vision language models. arXiv preprint arXiv:2406.13642, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.484, + 0.549 + ], + "angle": 0, + "content": "[5] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.634 + ], + "angle": 0, + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[7] Aakanksha Chowdhery, Narang, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1-113, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.79, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[10] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Kiana Ehsani, Tanmay Gupta, Rose Hendrix, Jordi Salvador, Luca Weihs, Kuo-Hao Zeng, Kunal Pratap Singh, Yejin Kim, Winson Han, Alvaro Herrasti, et al. Spoc: Imitating shortest" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.228, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "paths in simulation enables effective navigation and manipulation in the real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16238-16250, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[12] Xiaofeng Gao, Luyao Yuan, Tianmin Shu, Hongjing Lu, and Song-Chun Zhu. Show me what you can do: Capability calibration on reachable workspace for human-robot collaboration. IEEE Robotics and Automation Letters, 7(2):2644-2651, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[13] Yining Hong, Zishuo Zheng, Peihao Chen, Yian Wang, Junyan Li, and Chuang Gan. Multiply: A multisensory object-centric embodied large language model in 3d world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26406-26416, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[14] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[15] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[16] Lorenzo Jamone, Martim Brandao, Lorenzo Natale, Kenji Hashimoto, Giulio Sandini, and Atsuo Takanishi. Autonomous online generation of a motor representation of the workspace for intelligent whole-body reaching. Robotics and Autonomous Systems, 62(4):556-567, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[17] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.616 + ], + "angle": 0, + "content": "[18] Xiaohan Lei, Min Wang, Wengang Zhou, Li Li, and Houqiang Li. Instance-aware exploration-verification-exploitation for instance imagegoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16329–16339, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[19] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.906, + 0.743 + ], + "angle": 0, + "content": "[20] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[22] Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Lily Lee, Kaichen Zhou, Pengju An, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. Robomamba: Multimodal state space model for efficient robot reasoning and manipulation. arXiv preprint arXiv:2406.04339, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding din: Marrying din with" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6948" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.483, + 0.245 + ], + "angle": 0, + "content": "[25] Arjun Majumdar, Ajay, et al. Openeca: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[26] Matthias Minderer, Alexey Gritsenko, Austin Stone, et al. Simple open-vocabulary object detection with vision transformers, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.355 + ], + "angle": 0, + "content": "[27] Abby O'Neill, Rehman, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[28] Alec Radford, Kim, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[29] Pierre Sermanet, Ding, et al. Robovqa: Multimodal long-horizon reasoning for robotics. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 645-652. IEEE, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.551 + ], + "angle": 0, + "content": "[30] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[31] Chen Song and Darwin Lau. Workspace-based model predictive control for cable-driven robots. IEEE Transactions on Robotics, 38(4):2577-2596, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[32] Giovanni Sutanto, Austin Wang, Yixin Lin, Mustafa Mukadam, Gaurav Sukhatme, Akshara Rai, and Franziska Meier. Encoding physical constraints in differentiable newton-euler algorithm. In Proceedings of the 2nd Conference on Learning for Dynamics and Control, pages 804-813. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.311, + 0.692 + ], + "angle": 0, + "content": "[33] Qwen Team. qwen2.5, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[34] Dominique Thomas, Rodney Rothstein, Nathan Rosenberg, and Yolande Surdin-Kerjan. Sam2 encodes the second methionine s-adenosyl transferase in saccharomyces cerevisiae: physiology and regulation of both enzymes. Molecular and Cellular Biology, 8(12):5132-5139, 1988. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[35] Yunsheng Tian, Karl DD Willis, Bassel Al Omari, Jieliang Luo, Pingchuan Ma, Yichen Li, Farhad Javid, Edward Gu, Joshua Jacob, Shinjiro Sueda, et al. Asap: Automated sequence planning for complex robotic assembly with physical feasibility. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4380-4386. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[36] An Dinh Vuong, Minh Nhat Vu, Baoru Huang, Nghia Nguyen, Hieu Le, Thieu Vo, and Anh Nguyen. Language-driven grasp detection. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 17902-17912, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[37] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[38] Yu Wu, Lu Jiang, and Yi Yang. Revisiting embodiedqa: A simple baseline and beyond. IEEE Transactions on Image Processing, 29:3984-3992, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[39] An Yang, Baosong Yang, Binyuan Hui, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[40] Yijun Yang, Tianyi Zhou, Kanxue Li, Dapeng Tao, Lusong Li, Li Shen, Xiaodong He, Jing Jiang, and Yuhui Shi. Embodied multi-modal agent trained by an llm from a parallel textworld. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26275-26285, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[41] Franziska Zacharias, Christoph Borst, and Gerd Hirzinger. Capturing robot workspace structure: representing robot capabilities. In 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3229-3236. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[42] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[43] Ganlong Zhao, Guanbin Li, Weikai Chen, and Yizhou Yu. Over-nav: Elevating iterative vision-and-language navigation with open-vocabulary detection and structured representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16296–16306, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[44] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6949" + } + ] +] \ No newline at end of file diff --git a/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/14176105-f8ea-4fe1-b4ab-02889a03b21d_origin.pdf b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/14176105-f8ea-4fe1-b4ab-02889a03b21d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3e73e84f439491812eff1af6feda023bf39c1c22 --- /dev/null +++ b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/14176105-f8ea-4fe1-b4ab-02889a03b21d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56490b49324d317c9ea68d601be0fc939b93a9e02bb3fc472363e498e594204a +size 2333598 diff --git a/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/full.md b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4eb32049fcf52317a01aff7c38f1c32c39b5bd82 --- /dev/null +++ b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/full.md @@ -0,0 +1,329 @@ +# PhysVLM: Enabling Visual Language Models to Understand Robotic Physical Reachability + +Weijie Zhou $^{1,2}$ , Manli Tao $^{2}$ , Chaoyang Zhao $^{2,3,\dagger}$ , Haiyun Guo $^{2}$ , Honghui Dong $^{1}$ , Ming Tang $^{2}$ , Jinqiao Wang $^{1,2,3,4,\dagger}$ + +$^{1}$ School of Traffic and Transportation, Beijing Jiaotong University + +$^{2}$ Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences +3 ObjectEye Inc. Guangdong Provincial Key Laboratory of Intellectual Property & Big Data, Guangdong Polytechnic Normal University + +†Corresponding authors: chaoyang.zhao@nlpr.ia.ac.cn, jqwang@nlpr.ia.ac.cn + +# Abstract + +Understanding the environment and a robot's physical reachability is crucial for task execution. While state-of-the-art vision-language models (VLMs) excel in environmental perception, they often generate inaccurate or impractical responses in embodied visual reasoning tasks due to a lack of understanding of robotic physical reachability. To address this issue, we propose a unified representation of physical reachability across diverse robots, i.e., Space-Physical Reachability Map (S-P Map), and PhysVLM, a vision-language model that integrates this reachability information into visual reasoning. Specifically, the S-P Map abstracts a robot's physical reachability into a generalized spatial representation, independent of specific robot configurations, allowing the model to focus on reachability features rather than robot-specific parameters. Subsequently, PhysVLM extends traditional VLM architectures by incorporating an additional feature encoder to process the S-P Map, enabling the model to reason about physical reachability without compromising its general vision-language capabilities. To train and evaluate PhysVLM, we constructed a large-scale multi-robot dataset, Phys100K, and a challenging benchmark, EQA-phys, which includes tasks for six different robots in both simulated and real-world environments. Experimental results demonstrate that PhysVLM outperforms existing models, achieving a $14\%$ improvement over GPT-4o on EQA-phys and surpassing advanced embodied VLMs such as RoboMamba and SpatialVLM on the RoboVQA-val and OpenEQA benchmarks. Additionally, the S-P Map shows strong compatibility with various VLMs, and its integration into GPT-4o-mini yields a $7.1\%$ performance improvement. + +# 1. Introduction + +![](images/535fe186e22b80d0d2dc3d275570fe4336156db5e4ba7393c2b943b0f6824f8d.jpg) + +![](images/0a4a1a6975f734f86cd0b7b70f79cb8747df78880f35a5e4c15c3cc05f56e77d.jpg) + +Embodied QA Can robot pick up the bowl now? + +GPT40: + +![](images/7e6c350836bfcc35a52933f2f35e9a7b8584f301ee426085ade0cb6bbebd3b5b.jpg) + +Yes, the bowl is placed on a flat surface and is accessible. + +PhysVLM: + +![](images/1291bb3947edae24b71075d5f018e1ff11332a56f0d724dd2f50e66eb5ca376c.jpg) + +No, the bowl is not within the robot arm's reachable space. + +Task Planning + +Task goal is to put the can into the bowl, what are the next steps? + +GPT40: + +![](images/e7b9bd18dc43b0146055c501fbf80616557b2dc560cbb32b5b8c81fb1bdc1b60.jpg) + +Execution failure + +Picking up the can and place it inside the bowl. + +PhysVLM: + +![](images/eea56502ab52faa3fdf13e7ae0d6a176450fb57dd6a95f4b3c3661caeb8a29fb.jpg) + +1-Move close to the can. 2-Pick up the can. 3-Place the can into the bowl. 4-Down. + +Figure 1. Existing VLM models, such as GPT-4o, may generate inaccurate or impractical responses due to poor comprehension of robotic physical reachability. The proposed PhysVLM integrates vision-language capabilities with an understanding of robotic physical reachability. + +Accurate perception of physical reachability is essential for robots to perform tasks effectively. Similar to how humans adjust their actions based on bodily conditions and environmental factors, robots must account for their physical reachability within an environment to ensure efficient and reliable task execution. For instance, in grasping tasks, + +a robot that fails to assess its reachability may attempt to grasp an object from an unreachable position, leading to task failure or equipment damage [32, 35]. Thus, enhancing a robot's understanding of physical reachability is crucial for successful task planning and execution in complex environments [31, 41]. + +Vision-Language Models (VLMs) have shown remarkable progress in environmental understanding [6, 17, 19, 21, 37], and many studies have applied these models to embodied AI to assist robots in perceiving environments and planning tasks [11, 18, 25, 36, 43]. However, while VLMs excel in general environmental perception, they often struggle with tasks that require an understanding of robotic physical reachability (see Figure 1). We identify two key challenges that must be addressed for VLMs to be effective in robotic tasks: (1) how to develop a unified and efficient representation of physical reachability. Robots vary significantly in size, joint types, and other characteristics, making it difficult for VLMs to directly learn these differences; (2) how to enable VLMs to improve their understanding of physical reachability without compromising general vision-language capabilities. Existing VLMs typically combine pre-trained unimodal encoders for vision and language tasks. However, introducing a new modality like physical reachability requires careful architectural and training adjustments to ensure the model can reason about reachability while maintaining its general capabilities. + +To address these challenges, we propose the Space-Physical Reachability Map (S-P Map), a unified representation that abstracts the physical reachability of diverse robots into a generalized spatial form. The S-P Map is generated by combining robot parameters with egocentric depth images, but crucially, the model learns to focus on the abstracted reachability features (i.e., the gray regions in the S-P Map) rather than the specific robot configurations. This abstraction allows the model to generalize across different robots, as it only needs to reason about which areas are reachable, independent of the robot's specific characteristics. We introduce PhysVLM, a vision-language model that extends traditional VLM architectures by incorporating an additional feature encoder to process the S-P Map. This design enables PhysVLM to integrate physical reachability information into its reasoning process without compromising its general vision-language capabilities. To train and evaluate PhysVLM, we constructed a large-scale multi-robot dataset, Phys100K, and a challenging benchmark, EQAphys, which includes tasks for six different robots in both simulated and real-world environments. + +We summarize our contributions as follows: + +- We propose a unified and robot-agnostic formulation, the Space-Physical Reachability Map (S-P Map), which abstracts robotic physical reachability in a way that is independent of specific robot configurations, promoting the + +learning of generalized features. + +- We introduce PhysVLM, a vision-language model that integrates physical reachability with general vision-language capabilities via an additional feature encoder, improving task execution reliability. +- We release the EQA-phys benchmark, which includes six robots and 1.3K question-answer pairs, designed to test the model's understanding of physical reachability in simulated and real-world environments. +- Our model achieves a $14\%$ improvement over GPT-4o on the EQA-phys benchmark. In embodied visual reasoning tasks on the RoboVQA-val and OpenEQA benchmarks, it outperforms advanced embodied VLMs such as RoboMamba and SpatialVLM. Furthermore, the S-P Map demonstrates strong compatibility with various VLMs, and integrating it into GPT-4o-mini results in a $7.1\%$ performance improvement. + +# 2. Related Work + +# 2.1. VLMs in Robotics + +Embodied Question Answering (EQA) tasks require agents to interact with environments to answer questions [10, 13, 38]. RoboVQA offers a large, diverse dataset for robotic visual question answering. 3D-VLA [44] integrates 3D perception with a generative world model for embodied reasoning, while SpatialVLM [5] enhances VLMs' spatial understanding using extensive 3D data. + +Robot task planning involves sequencing subtasks to achieve goals [3, 30, 40]. Code as Policies (CaP) [20] employs OpenAI's Codex (code-davinci-002) to generate planning code. SayCan [2] combines the Pathways Language Model (PaLM) [7] with robotic affordances to create feasible action plans based on the robot's capabilities. However, these methods often assume all objects are within the robot's operational area, ignoring physical reachability and potentially leading to suboptimal or infeasible plans. + +# 2.2. Understanding Physical Reachability + +Recent studies use voxel grids with open-vocabulary detection models to assign task-specific attributes, enabling environmental constraint understanding. ReKep [15] generates keypoint proposals and constraints using voxel grids and VLMs, while VoxPoser [14] synthesizes robot trajectories by integrating OWL-ViT [26] and VLMs with voxel-based environment representations. However, these approaches focus on environment modeling without explicitly addressing the robot's physical reachability. + +Explicitly representing reachable workspaces remains challenging. Reachability maps [41] model spatial capabilities, and occupancy grids [16] account for obstacles to ensure safe navigation. Additionally, methods like online model predictive control with offline workspace analysis + +[31] and Reachability Expression-based Motion Planning (REMP) [12] address workspace constraints. Despite these advancements, integrating physical reachability into visual reasoning for complex embodied tasks is still limited, primarily due to the lack of large-scale datasets that include robotic physical parameters in VLM pretraining. + +# 3. Method + +PhysVLM is a large-scale vision-language model designed for visual reasoning that accounts for physical constraints in embodied tasks. As illustrated in Figure 2, PhysVLM integrates instruction text, visual input (RGB image), and an S-P Map abstracts the robotic physical reachability into a unified spatial representation. By combining these inputs, PhysVLM generates responses consistent with both the visual context and the robot's physical reachability, without being tied to specific robot configurations. The S-P Map is constructed using a unified physical reachability encoding method, which abstracts the physical parameters of various robots alongside their egocentric depth maps into a generalized form. This abstraction allows the model to generalize across different robots, addressing the challenge of learning and reasoning about physical reachability in a robot-agnostic manner. + +In this section, we present the core components of PhysVLM. Section 3.1 describes the S-P Map encoding method, Section 3.2 describes the model architecture, and Section 3.3 discusses the training procedures. + +# 3.1. S-P Map Encoding + +As illustrated in Figure 2, we model the physical reachability of various robots using a unified approach that abstracts robot-specific parameters into a generalized spatial representation. This abstraction allows the model to focus on the spatial regions that are physically reachable, independent of the specific robot configuration. + +$$ +\mathrm {S} - \mathrm {P} \quad \operatorname {M a p} = F \left(\mathcal {P} _ {\text {r a w}}, \left\{\theta_ {i} ^ {\min }, \theta_ {i} ^ {\max } \right\}, \mathrm {D H}, \mathbf {E}\right), \tag {1} +$$ + +where $\mathcal{P}_{\mathrm{raw}}$ denotes the raw point cloud data from the robot's RGB-D camera. $\{\theta_i^{\mathrm{min}},\theta_i^{\mathrm{max}}\}$ represents the range of motion for each joint $i$ . DH refers to the Denavit-Hartenberg parameters that describe the geometric structure of each joint, and $\mathbf{E}$ is the extrinsic calibration matrix that transforms coordinates from the camera to the robot's coordinate system. The function $F$ maps these inputs to produce the S-P Map, which abstracts the robot's physical reachability into a spatial form that is independent of the specific robot configuration. + +Consider a robot arm with $n$ degrees of freedom, where each joint $i$ has DH parameters $\{\theta_i, d_i, a_i, \alpha_i\}$ ( $\theta_i$ is the joint angle, $d_i$ is the offset along the z-axis, $a_i$ is the link + +length, and $\alpha_{i}$ is the twist angle). The homogeneous transformation matrix for each joint is defined as: + +$$ +\mathbf {T} _ {i} = G \left(\theta_ {i}, d _ {i}, a _ {i}, \alpha_ {i}\right), \tag {2} +$$ + +where $G$ is the standard Denavit-Hartenberg transformation function. By multiplying the transformation matrices of all joints, we obtain the transformation matrix from the base frame to the end-effector frame: + +$$ +\mathbf {T} = \mathbf {T} _ {1} \mathbf {T} _ {2} \dots \mathbf {T} _ {n}. \tag {3} +$$ + +To generate joint configurations, we sample the joint angles $\theta_{i}$ from their respective motion ranges $\theta_{i} \in [\theta_{i}^{\min}, \theta_{i}^{\max}]$ , resulting in configurations $\{\theta_{1}, \theta_{2}, \ldots, \theta_{n}\}$ . By substituting these joint configurations into the forward kinematics equations, we compute the corresponding end-effector positions: + +$$ +\mathcal {W} _ {\text {v o x e l}} = \left\{\mathbf {p} \mid \mathbf {p} = \mathbf {T} \left(\theta_ {1}, \theta_ {2}, \dots , \theta_ {n}\right) \cdot \mathbf {p} _ {0} \right\}, \tag {4} +$$ + +where $\mathbf{p}_0$ is the origin point in the end-effector frame. We precompute these joint configurations offline, discretize the workspace into a voxel grid $\mathcal{W}_{\mathrm{voxel}}$ , and store it for efficient computation in subsequent steps. + +Next, as shown in Figure 2, the robot's raw point cloud $\mathcal{P}_{\mathrm{raw}}$ is captured from its egocentric RGB-D camera in the camera coordinate system and transformed into the robot's coordinate system using the extrinsic calibration matrix $\mathbf{E}$ , resulting in the transformed point cloud $\mathcal{P}$ : + +$$ +\mathcal {P} = \mathbf {E} \cdot \mathcal {P} _ {\text {r a w}}. \tag {5} +$$ + +To ensure physical feasibility, we perform a voxel grid lookup to determine whether each point in $\mathcal{P}$ lies within the precomputed reachable workspace $\mathcal{W}_{\mathrm{voxel}}$ : + +$$ +\mathcal {P} _ {\text {v a l i d}} = \left\{\mathbf {p} \in \mathcal {P} \mid \mathbf {p} \in \mathcal {W} _ {\text {v o x e l}} \right\}. \tag {6} +$$ + +This step filters the point cloud to include only points within the robot's reachable workspace, abstracting the robot's physical reachability into a generalized spatial form. + +Finally, we transform the valid point cloud $\mathcal{P}_{\mathrm{valid}}$ back into the camera coordinate system and use the camera's intrinsic parameters to project these points onto the image plane. We then mark the regions that comply with physical reachability on the original depth map. For areas that are not reachable, we apply a gray mask and outline their boundaries. The resulting S-P Map clearly highlights regions that are beyond the robot's physical reach, providing a unified and abstracted representation of reachability that is independent of the specific robot configuration. This allows the model to focus on the spatial constraints of the task, without needing to account for the detailed physical parameters of each robot. + +![](images/88d43d2171b258406312fcfa33b9d14fb4805f02ceb4d972a68f9f1cff100ba1.jpg) +Figure 2. Overview of PhysVLM. Starting with robot parameters and an egocentric depth map, an S-P Map is generated via unified physical reachability encoding. Using the S-P Map, image, and instruction text, PhysVLM generates textual output considering the robot's physical reachability. + +![](images/32cc7a7325b209c2bfe43dd9c2f2cbef4029ba37c6ccdc2828b60eaa5e7b3c9d.jpg) + +# 3.2. Model Architecture + +To seamlessly integrate robotic physical reachability into PhysVLM while preserving its visual reasoning capabilities, we design a dual-branch architecture: one branch dedicated to vision processing and the other to physical reachability (see Figure 2). These branches operate independently, extracting features from their respective inputs, which are then fused and passed to a unified decoder for final reasoning and response generation. + +The vision branch leverages a pre-trained Vision Transformer (ViT) [9], specifically the SigLip-400M model [42], to extract high-level visual features from egocentric images. To reduce computational overhead, a Max Pooling layer is applied, followed by a two-layer Multi-Layer Perceptron (MLP) that transforms the visual features into token representations suitable for multimodal fusion. + +The physical reachability branch processes the S-P Map, which abstracts the robot's physical reachability into a generalized spatial form. This branch also utilizes the SigLip-400M model for feature extraction, followed by Max Pooling and a feature fusion layer. The fusion layer combines the visual and reachability features, and a two-layer MLP further refines these fused features into reachability-specific tokens. + +For the language decoding, we employ the Qwen-2.5-Instruct-3B model [33, 39] as PhysVLM's large language model (LLM) decoder, using the Qwen-2.5 tokenizer to process natural language instructions. The decoder integrates multimodal tokens from the vision branch, S-P Maps, and language inputs, generating coherent and contextually rele + +vant textual responses that account for both visual and physical reachability information. + +# 3.3. Training + +Training Data Construction. The training data for PhysVLM consists of our Phys100K dataset and general VQA datasets, such as LLaVA-Pretrain, ShareGPT4V, and RoboVQA. Phys100K focuses on question-answering related to physical reachability, aggregating data from RoboVQA (20K samples), ScanNet [8] (10K samples), OpenX-Embodiment [27] (60K samples), and an additional 10K samples from PyBullet. + +Depth maps are essential inputs for generating the S-P Map. For datasets lacking depth maps, we generate them using DepthAnything-v2. Additionally, we employ Grounding DINO [23] and SAM2 [34] to obtain 2D bounding boxes and segmentation results for objects in the images. In PyBullet, we simulate work scenarios using four robotic arms (UR5, FR5, CR5, and FRANKA) to collect RGB images, depth maps, and segmentation results. + +For the PyBullet data in Phys100K, precise robot configurations can be obtained from the simulator. Therefore, we directly generate the S-P Map using the method described in Section 3.2, and labels indicating whether objects are reachable are obtained through simulated motion. The advantage of the S-P Map is that it abstracts physical reachability into a region-based representation, decoupling the learning process from specific robot configurations. This allows us to generate pseudo-labels for datasets without precise robot parameters. We approximate reachability using the segmentation results, marking regions and the objects within + +![](images/2fc2e9b0c0203553b91d189e788739d096524d5d6068b12942995c3621fb84ac.jpg) +Figure 3. Details of the Phys100K Dataset and EQA-Phys Benchmark. + +them as "reachable" or "unreachable" based on depth values. Next, we generate question-answer pairs for two main categories: + +- Embodied QA. GPT-4 generates question-answer pairs for ScanNet and RoboVQA, covering categories include Function Reasoning, World Knowledge, Object Recognition, Object Localization, Attribute Recognition, Spatial Reasoning, Object State Recognition, and Hallucination (See Figure 3). Detailed prompts and examples are provided in the appendix to guide the generation process. +- Tasks Involving Physical Reachability. We use the "reachable" label with five fixed task templates to generate question-answer pairs, such as "USER:\n\n Is the [Object] in the robot's reachable space? ASSISTANT: Yes, it is." Here, [Object] represents the relevant object category, while and serve as placeholders for image patch tokens and S-P Map patch tokens, respectively. Figure 3 provides examples of question-answer pairs for each object category. + +Training Pipeline. We adopt a two-stage training process to fully leverage the S-P Map and ensure PhysVLM generalizes across different robots. In the first stage, we align multimodal features using the LLaVA-Pretrain and OpenX-Embodiment datasets from Phys100K. This stage only trains the projection layers, allowing the model to build + +a foundational understanding of visual inputs and physical reachability, independent of specific robot configurations. + +In the second stage, we unfreeze all parameters and train the entire model using data from Phys100K, ShareGPT4V, and RoboVQA. This stage enhances PhysVLM's ability to handle complex visual reasoning tasks with physical reachability constraints, ensuring the model can generalize across diverse environments and robots. + +Implementation Details. PhysVLM is trained for 48 hours using eight A800 GPUs. The training process consists of two stages, each lasting one epoch. The batch size and learning rate are set at 128 and 1e-3 in the first stage, and 64 and 1e-5 in the second stage. The final model is PhysVLM-3B. + +# 3.4. EQA-phys Benchmark + +As illustrated in Figure 3, we introduce an embodied QA task focused on physical reachability, termed EQA-phys, which emphasizes QA tasks constrained by physical limitations. This benchmark includes a simulator dataset with 200 samples and 1,000 questions from the PyBullet validation set, as well as a zero-shot evaluation set based on real-world data from UR3 and XArm6 robots in two scenarios. The evaluation set contains 60 samples and 300 questions, all manually annotated by domain experts. + +Table 1. Results of EQA-phys. Comparison of PhysVLM-3B (ours) with API-based VLMs and embodied VLMs. + +
REAL-WORLDSIMULATORALL
UR3XARM6UR5FR5CR5FRANKA
API-BASED VLMSGPT-4O-MINI54.356.049.455.454.647.152.8
CLAUDE-3.556.260.554.058.155.754.356.4
GPT-4O56.761.555.758.357.552.657.0
GPT-4O-MINI + S-P MAP60.0↑5.760.5↑4.557.0↑7.659.1↑3.759.2↑4.653.3↑6.259.8↑7.0
CLAUDE-3.5 + S-P MAP65.3↑9.167.3↑6.854.9↑0.958.3↑0.258.2↑2.558.1↑3.860.3↑3.4
GPT-4O + S-P MAP66.6↑9.968.1↑6.655.8↑0.160.7↑1.459.4↑1.957.6↑5.061.3↑4.1
EMBODIED VLMSSPATIALVLM56.355.154.659.152.047.554.1
SPATIALBOT51.150.250.048.153.354.451.1
PHYSVLM-3B64.163.071.475.774.078.171.0
+ +# 4. Experiments + +# 4.1. Experimental Setting + +Tasks. We compare the performance of PhysVLM with other methods across three categories of tasks: + +- EQA-phys. This benchmark tests the model's ability to integrate visual reasoning with robotic physical reachability. The real-robot component is used to assess PhysVLM's zero-shot generalization, highlighting its ability to handle unseen robots and environments. +- Embodied QA. We evaluate the model's general visual reasoning ability in embodied tasks using the OpenEQA [25] and RoboVQA-val [29] benchmarks. +- Robot Task Planning. For real-world tasks such as "Pick A into B", we assess the model's ability to understand robotic physical reachability and generate reasonable task plans. As this study does not focus on robot control strategies, we use the natural language planning approach from [29]. + +Baselines. We compare our model to several baselines, including API-accessible VLMs such as Claude 3.5 [28], GPT-4o-mini [1], and GPT-4o [1], as well as embodied VLMs like SpatialVLM [5], SpatialBot [4], 3D-VLA [44], and RoboMamba [22]. SpatialVLM and SpatialBot both use the 3B version, which has a similar parameter count to our model. Since the executable versions of 3D-VLA and RoboMamba are unavailable, we compare their reported results on RoboVQA-val. + +Evaluation Metrics. For tasks involving physical reachability, we use LLM scoring, following the approach used in existing studies [24, 25]. Assigning 5 points for completely correct responses and 1 point for incorrect responses, we calculate the average score and express it as a percentage. For Embodied QA, we follow the benchmark settings of the corresponding datasets [25, 29]. For task planning, each + +task type is executed 10 times, and the average success rate serves as the evaluation metric. + +# 4.2. Results on EQA-phys + +Table 1 shows the results on EQA-phys. Neither API-based nor embodied VLMs can handle the robot's parameter constraints, resulting in suboptimal outputs with scores around $55\%$ . In contrast, our model successfully completes visual reasoning tasks involving physical reachability, obtaining an average score of $71\%$ . As discussed, these tasks require the model to perform visual reasoning based on an understanding of the robotic physical reachability. The model can only effectively accomplish these tasks by truly understanding robotic physical reachability. + +The results in Table 1 demonstrate that prompting API-based VLMs, such as GPT-40, with the S-P Map (detailed in 4) significantly enhances their performance. This improvement stems from the S-P Map's ability to abstract physical reachability into a robot-agnostic representation, enabling VLMs to reason about physical constraints that would otherwise be beyond their capabilities. By decoupling reachability from specific robot parameters, the S-P Map facilitates generalization across diverse environments, allowing models to better comprehend physical reachability, even in previously unseen scenarios. + +Table 1 also demonstrates that PhysVLM-3B achieved scores of over $63\%$ in zero-shot evaluations for UR3 and XArm6 robots, despite operating in new environments with different robot parameters. This performance is attributed to two key factors: (1) the S-P Map abstracts various robot parameters into a unified, transferable representation of physical reachability, and (2) the model's independent visual and constraint encoding branches allow it to learn generalizable visual features from diverse image-text data, enabling effective reasoning in novel environments. + +Table 2. Embodied QA results on the RoboVQA-val set, comparison of PhysVLM (ours) with existing methods. An asterisk (*) indicates models not pre-trained on the RoboVQA dataset. + +
BLEU1BLEU2BLEU3BLEU4
SPATIALVLM*5.13.01.91.2
SPATIALBOT*12.49.38.07.2
3D-VLA48.338.531.726.8
ROBOMAMBA54.944.239.536.3
PHYSVLM-3B65.362.450.943.5
+ +Table 3. Embodied QA results on the OpenEQA benchmark, comparison of PhysVLM (ours) with existing methods. An asterisk (*) indicates that only the first 200 samples were tested due to API limitations. + +
EM-EQA (SCANNET)EM-EQA (HM3D)ALL
SPATIALVLM42.944.343.8
SPATIALBOT45.351.049.1
GPT4V57.451.355.3
GPT-4O*68.265.266.7
PHYSVLM-3B60.751.257.4
+ +# 4.3. Results on Embodied QA + +We demonstrate our model's effectiveness in handling general embodied visual reasoning tasks. Additionally, we show that incorporating an understanding of physical constraints does not diminish its general visual reasoning capabilities. We compare our model with state-of-the-art embodied VLMs (see Tables 2 and 3). Our model achieves the best performance on the RoboVQA-val benchmark, surpassing other models by $7.2\%$ in BLEU-4. On the OpenEQA benchmark, our model outperforms existing embodied VLMs and GPT-4V, ranking second, behind only GPT-4o. + +# 4.4. Results on Robot Task Planning + +Table 4 presents the performance of PhysVLM and baseline models on real-world task planning scenarios. When all objects are within the robot's physical reach, PhysVLM performs similarly to other models, as directly grabbing or placing the objects succeeds. However, when some objects are outside the physical reach, the model must suggest that the robot move closer before grabbing or placing them. In these cases, our model performs exceptionally well, whereas the task success rates of other models decline significantly. This is attributed to our model's understanding of robotic physical reachability and its ability to incorporate this understanding in task planning. + +Table 4. Task planning results. Comparison of PhysVLM (ours) with other VLMs. + +
ALL OBJECTS IN RANGEPART OBJECTS IN RANGE
GPT-4O-MINI70.523.2
CLAUDE-3.573.632.1
GPT-4O75.935.8
SPATIALVLM64.421.5
SPATIALBOT65.625.3
PHYSVLM-3B69.248.4
+ +Table 5. Ablation study on the S-P Map. We compare the constraint encoder's performance when using the S-P Map, replacing it with a Depth Map, or providing no input. + +
IDS-P MAPDEPTH MAPEQA-PHYS REALEQA-PHYS SIM
1✓63.574.8
2✓58.162.4
354.258.8
+ +# 4.5. Ablation Study + +In this section, we conduct ablation studies to evaluate the contribution of each component in PhysVLM. We report the average LLM score on tasks involving physical reachability. + +The effectiveness of S-P Map. To demonstrate the S-P Map's contribution, we compare results obtained with and without its input. As shown in Table 5, Experiments 1 and 3 demonstrate that omitting the S-P Map leads to a significant performance decrease for both zero-shot real-world robots and simulators. Specifically, the overall average score drops by $16\%$ in simulation results and $9.3\%$ in real-world robot evaluations. Without the S-P Map input, the model struggles to handle the robotic physical reachability. + +Additionally, Experiments 1 and 2 demonstrate that replacing the S-P Map with a Depth Map significantly degrades the model's performance on zero-shot tasks. Since the Depth Map does not accurately represent robotic physical reachability, the model cannot rely solely on depth information to understand it. + +Effectiveness of an additional feature encoder. To demonstrate the effectiveness of the model architecture, we compare the performance of a feature encoder that shares weights with the visual feature encoder for the S-P Map. In these experiments, we evaluate the average scores on both the EQA-phys and OpenEQA benchmarks. The results in + +![](images/55e73a6e3b0b10b01833172bb38edd36b9835b507e965ff9dd727e592c519a18.jpg) +Figure 4. Visual comparison of PhysVLM (ours), GPT-4o, and SpatialBot. + +Table 6. Ablation study on the effectiveness of an additional feature encoder. Share indicates shared network and weights with the visual feature encoder. + +
EQA-PHYSOPENEQA
INDEPENDENT71.057.4
SHARE68.256.5
+ +Table 6 show that sharing the feature encoder not only decreases performance on EQA-phys but also impairs general visual reasoning capabilities. This is because S-P Map features differ from images, and the training data contains significantly more image-text pairs than S-P Map data. + +Effectiveness of training data. To evaluate the effectiveness of Phys100K, we conduct experiments by selectively removing data from various sources in Phys100K. As shown in Table 7, removing data from PyBullet or other embodied datasets leads to a reduction in overall performance, highlighting the critical role of each data component in the model's performance. + +Table 7. Ablation study on the effectiveness of training data. + +
PART OF PHYS100KEQA-PHYS REALEQA-PHYS SIM
ALL63.574.8
W/O PYBULET62.165.4
W/O OTHER DATASETS58.671.5
+ +# 4.6. Qualitative Results + +Figure 4 compares our method with SpatialBot and GPT-4o. SpatialBot uses depth maps and images, while GPT-4o uses standard images. Both struggle with tasks requiring physical reachability, causing visual reasoning errors. In contrast, our approach delivers accurate results. Additionally, incorporating the S-P Map into GPT-4o improves its handling of physical reachability and response accuracy. + +# 5. Conclusion + +We introduce PhysVLM, a VLM that incorporates physical reachability into visual reasoning for robotic tasks. The S-P Map provides a unified representation of robotic reachability, facilitating the learning of generalizable features. PhysVLM extends traditional VLMs by adding a physical reachability encoder, enabling the simultaneous processing of visual, reachability, and textual information. Additionally, we present EQA-phys, a benchmark for evaluating embodied QA tasks involving physical reachability. Our experiments show that PhysVLM outperforms existing models, achieving a $14\%$ higher score than GPT-4o on EQA-phys. A limitation is its reduced zero-shot performance on real robots compared to simulations, likely due to the domain gap. Future work will focus on expanding datasets, enhancing real-world performance, and improving the understanding of physical accessibility in vision-language-action models. PhysVLM's reachability awareness supports safer and more reliable robotic decision-making in industrial and assistive settings, while its unified representation ensures cross-platform adaptability for real-world deployment, bridging crucial gaps between environmental perception and actionable robotic intelligence. + +# 6. Acknowledgments + +This work was supported by National Key R&D Program of China under Grant No.2022ZD0160601, 2022YFB4300400, and 2018B030322016, National Natural Science Foundation of China under Grants 62176254, 62276260, and U1701266. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6 +[2] Michael Ahn, Brohan, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2 +[3] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2 +[4] Wenxiao Cai, Yaroslav Ponomarenko, Jianhao Yuan, Xiaqi Li, Wankou Yang, Hao Dong, and Bo Zhao. Spatialbot: Precise spatial understanding with vision language models. arXiv preprint arXiv:2406.13642, 2024. 6 +[5] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2, 6 +[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2 +[7] Aakanksha Chowdhery, Narang, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1-113, 2023. 2 +[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 4 +[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4 +[10] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2 +[11] Kiana Ehsani, Tanmay Gupta, Rose Hendrix, Jordi Salvador, Luca Weihs, Kuo-Hao Zeng, Kunal Pratap Singh, Yejin Kim, Winson Han, Alvaro Herrasti, et al. Spoc: Imitating shortest + +paths in simulation enables effective navigation and manipulation in the real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16238-16250, 2024. 2 +[12] Xiaofeng Gao, Luyao Yuan, Tianmin Shu, Hongjing Lu, and Song-Chun Zhu. Show me what you can do: Capability calibration on reachable workspace for human-robot collaboration. IEEE Robotics and Automation Letters, 7(2):2644-2651, 2022. 3 +[13] Yining Hong, Zishuo Zheng, Peihao Chen, Yian Wang, Junyan Li, and Chuang Gan. Multiply: A multisensory object-centric embodied large language model in 3d world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26406-26416, 2024. 2 +[14] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2 +[15] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 2 +[16] Lorenzo Jamone, Martim Brandao, Lorenzo Natale, Kenji Hashimoto, Giulio Sandini, and Atsuo Takanishi. Autonomous online generation of a motor representation of the workspace for intelligent whole-body reaching. Robotics and Autonomous Systems, 62(4):556-567, 2014. 2 +[17] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2 +[18] Xiaohan Lei, Min Wang, Wengang Zhou, Li Li, and Houqiang Li. Instance-aware exploration-verification-exploitation for instance imagegoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16329–16339, 2024. 2 +[19] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2 +[20] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2 +[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[22] Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Lily Lee, Kaichen Zhou, Pengju An, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. Robomamba: Multimodal state space model for efficient robot reasoning and manipulation. arXiv preprint arXiv:2406.04339, 2024. 6 +[23] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding din: Marrying din with + +grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4 +[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6 +[25] Arjun Majumdar, Ajay, et al. Openeca: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2, 6 +[26] Matthias Minderer, Alexey Gritsenko, Austin Stone, et al. Simple open-vocabulary object detection with vision transformers, 2022. 2 +[27] Abby O'Neill, Rehman, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 4 +[28] Alec Radford, Kim, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[29] Pierre Sermanet, Ding, et al. Robovqa: Multimodal long-horizon reasoning for robotics. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 645-652. IEEE, 2024. 6 +[30] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2 +[31] Chen Song and Darwin Lau. Workspace-based model predictive control for cable-driven robots. IEEE Transactions on Robotics, 38(4):2577-2596, 2022. 2, 3 +[32] Giovanni Sutanto, Austin Wang, Yixin Lin, Mustafa Mukadam, Gaurav Sukhatme, Akshara Rai, and Franziska Meier. Encoding physical constraints in differentiable newton-euler algorithm. In Proceedings of the 2nd Conference on Learning for Dynamics and Control, pages 804-813. PMLR, 2020. 2 +[33] Qwen Team. qwen2.5, 2024. 4 +[34] Dominique Thomas, Rodney Rothstein, Nathan Rosenberg, and Yolande Surdin-Kerjan. Sam2 encodes the second methionine s-adenosyl transferase in saccharomyces cerevisiae: physiology and regulation of both enzymes. Molecular and Cellular Biology, 8(12):5132-5139, 1988. 4 +[35] Yunsheng Tian, Karl DD Willis, Bassel Al Omari, Jieliang Luo, Pingchuan Ma, Yichen Li, Farhad Javid, Edward Gu, Joshua Jacob, Shinjiro Sueda, et al. Asap: Automated sequence planning for complex robotic assembly with physical feasibility. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4380-4386. IEEE, 2024. 2 +[36] An Dinh Vuong, Minh Nhat Vu, Baoru Huang, Nghia Nguyen, Hieu Le, Thieu Vo, and Anh Nguyen. Language-driven grasp detection. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 17902-17912, 2024. 2 +[37] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2 +[38] Yu Wu, Lu Jiang, and Yi Yang. Revisiting embodiedqa: A simple baseline and beyond. IEEE Transactions on Image Processing, 29:3984-3992, 2020. 2 +[39] An Yang, Baosong Yang, Binyuan Hui, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 4 +[40] Yijun Yang, Tianyi Zhou, Kanxue Li, Dapeng Tao, Lusong Li, Li Shen, Xiaodong He, Jing Jiang, and Yuhui Shi. Embodied multi-modal agent trained by an llm from a parallel textworld. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26275-26285, 2024. 2 +[41] Franziska Zacharias, Christoph Borst, and Gerd Hirzinger. Capturing robot workspace structure: representing robot capabilities. In 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3229-3236. IEEE, 2007. 2 +[42] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023. 4 +[43] Ganlong Zhao, Guanbin Li, Weikai Chen, and Yizhou Yu. Over-nav: Elevating iterative vision-and-language navigation with open-vocabulary detection and structured representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16296–16306, 2024. 2 +[44] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 2, 6 \ No newline at end of file diff --git a/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/images.zip b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..987351b106cfc225bd8013f0386a34fe5434ac11 --- /dev/null +++ b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2793d86fc2a0630e4fe6a2ca4866cc1d266763bfab1a6e8d36216c072cc73caf +size 590297 diff --git a/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/layout.json b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4abcaac8765559149275ca4334046b82ed935df7 --- /dev/null +++ b/2025/PhysVLM_ Enabling Visual Language Models to Understand Robotic Physical Reachability/layout.json @@ -0,0 +1,7796 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "type": "text", + "content": "PhysVLM: Enabling Visual Language Models to Understand Robotic Physical Reachability" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "spans": [ + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": "Weijie Zhou" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": ", Manli Tao" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": ", Chaoyang Zhao" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{2,3,\\dagger}" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": ", Haiyun Guo" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": ", Honghui Dong" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": ", Ming Tang" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": ", Jinqiao Wang" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,3,4,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 189, + 466, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 189, + 466, + 204 + ], + "spans": [ + { + "bbox": [ + 144, + 189, + 466, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 189, + 466, + 204 + ], + "type": "text", + "content": " School of Traffic and Transportation, Beijing Jiaotong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 203, + 537, + 246 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 79, + 203, + 530, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 203, + 530, + 217 + ], + "spans": [ + { + "bbox": [ + 79, + 203, + 530, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 203, + 530, + 217 + ], + "type": "text", + "content": " Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 217, + 537, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 217, + 537, + 246 + ], + "spans": [ + { + "bbox": [ + 72, + 217, + 537, + 246 + ], + "type": "text", + "content": "3 ObjectEye Inc. Guangdong Provincial Key Laboratory of Intellectual Property & Big Data, Guangdong Polytechnic Normal University" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 103, + 247, + 504, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 247, + 504, + 258 + ], + "spans": [ + { + "bbox": [ + 103, + 247, + 504, + 258 + ], + "type": "text", + "content": "†Corresponding authors: chaoyang.zhao@nlpr.ia.ac.cn, jqwang@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 328, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 328, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 297, + 700 + ], + "type": "text", + "content": "Understanding the environment and a robot's physical reachability is crucial for task execution. While state-of-the-art vision-language models (VLMs) excel in environmental perception, they often generate inaccurate or impractical responses in embodied visual reasoning tasks due to a lack of understanding of robotic physical reachability. To address this issue, we propose a unified representation of physical reachability across diverse robots, i.e., Space-Physical Reachability Map (S-P Map), and PhysVLM, a vision-language model that integrates this reachability information into visual reasoning. Specifically, the S-P Map abstracts a robot's physical reachability into a generalized spatial representation, independent of specific robot configurations, allowing the model to focus on reachability features rather than robot-specific parameters. Subsequently, PhysVLM extends traditional VLM architectures by incorporating an additional feature encoder to process the S-P Map, enabling the model to reason about physical reachability without compromising its general vision-language capabilities. To train and evaluate PhysVLM, we constructed a large-scale multi-robot dataset, Phys100K, and a challenging benchmark, EQA-phys, which includes tasks for six different robots in both simulated and real-world environments. Experimental results demonstrate that PhysVLM outperforms existing models, achieving a " + }, + { + "bbox": [ + 54, + 328, + 297, + 700 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 54, + 328, + 297, + 700 + ], + "type": "text", + "content": " improvement over GPT-4o on EQA-phys and surpassing advanced embodied VLMs such as RoboMamba and SpatialVLM on the RoboVQA-val and OpenEQA benchmarks. Additionally, the S-P Map shows strong compatibility with various VLMs, and its integration into GPT-4o-mini yields a " + }, + { + "bbox": [ + 54, + 328, + 297, + 700 + ], + "type": "inline_equation", + "content": "7.1\\%" + }, + { + "bbox": [ + 54, + 328, + 297, + 700 + ], + "type": "text", + "content": " performance improvement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 286, + 394, + 299 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 286, + 394, + 299 + ], + "spans": [ + { + "bbox": [ + 315, + 286, + 394, + 299 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 332, + 433, + 423 + ], + "blocks": [ + { + "bbox": [ + 318, + 332, + 433, + 423 + ], + "lines": [ + { + "bbox": [ + 318, + 332, + 433, + 423 + ], + "spans": [ + { + "bbox": [ + 318, + 332, + 433, + 423 + ], + "type": "image", + "image_path": "535fe186e22b80d0d2dc3d275570fe4336156db5e4ba7393c2b943b0f6824f8d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 441, + 334, + 552, + 423 + ], + "blocks": [ + { + "bbox": [ + 441, + 334, + 552, + 423 + ], + "lines": [ + { + "bbox": [ + 441, + 334, + 552, + 423 + ], + "spans": [ + { + "bbox": [ + 441, + 334, + 552, + 423 + ], + "type": "image", + "image_path": "0a4a1a6975f734f86cd0b7b70f79cb8747df78880f35a5e4c15c3cc05f56e77d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 426, + 492, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 426, + 492, + 436 + ], + "spans": [ + { + "bbox": [ + 323, + 426, + 492, + 436 + ], + "type": "text", + "content": "Embodied QA Can robot pick up the bowl now?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 439, + 351, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 439, + 351, + 449 + ], + "spans": [ + { + "bbox": [ + 323, + 439, + 351, + 449 + ], + "type": "text", + "content": "GPT40:" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 455, + 440, + 465, + 449 + ], + "blocks": [ + { + "bbox": [ + 455, + 440, + 465, + 449 + ], + "lines": [ + { + "bbox": [ + 455, + 440, + 465, + 449 + ], + "spans": [ + { + "bbox": [ + 455, + 440, + 465, + 449 + ], + "type": "image", + "image_path": "7e6c350836bfcc35a52933f2f35e9a7b8584f301ee426085ade0cb6bbebd3b5b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 323, + 449, + 517, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 449, + 517, + 458 + ], + "spans": [ + { + "bbox": [ + 323, + 449, + 517, + 458 + ], + "type": "text", + "content": "Yes, the bowl is placed on a flat surface and is accessible." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 323, + 459, + 361, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 459, + 361, + 468 + ], + "spans": [ + { + "bbox": [ + 323, + 459, + 361, + 468 + ], + "type": "text", + "content": "PhysVLM:" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 455, + 459, + 465, + 468 + ], + "blocks": [ + { + "bbox": [ + 455, + 459, + 465, + 468 + ], + "lines": [ + { + "bbox": [ + 455, + 459, + 465, + 468 + ], + "spans": [ + { + "bbox": [ + 455, + 459, + 465, + 468 + ], + "type": "image", + "image_path": "1291bb3947edae24b71075d5f018e1ff11332a56f0d724dd2f50e66eb5ca376c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 323, + 469, + 520, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 469, + 520, + 479 + ], + "spans": [ + { + "bbox": [ + 323, + 469, + 520, + 479 + ], + "type": "text", + "content": "No, the bowl is not within the robot arm's reachable space." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 323, + 485, + 373, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 485, + 373, + 495 + ], + "spans": [ + { + "bbox": [ + 323, + 485, + 373, + 495 + ], + "type": "text", + "content": "Task Planning" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 376, + 484, + 544, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 484, + 544, + 503 + ], + "spans": [ + { + "bbox": [ + 376, + 484, + 544, + 503 + ], + "type": "text", + "content": "Task goal is to put the can into the bowl, what are the next steps?" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 323, + 506, + 350, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 506, + 350, + 515 + ], + "spans": [ + { + "bbox": [ + 323, + 506, + 350, + 515 + ], + "type": "text", + "content": "GPT40:" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 455, + 507, + 465, + 516 + ], + "blocks": [ + { + "bbox": [ + 455, + 507, + 465, + 516 + ], + "lines": [ + { + "bbox": [ + 455, + 507, + 465, + 516 + ], + "spans": [ + { + "bbox": [ + 455, + 507, + 465, + 516 + ], + "type": "image", + "image_path": "e7b9bd18dc43b0146055c501fbf80616557b2dc560cbb32b5b8c81fb1bdc1b60.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 467, + 508, + 517, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 467, + 508, + 517, + 516 + ], + "spans": [ + { + "bbox": [ + 467, + 508, + 517, + 516 + ], + "type": "text", + "content": "Execution failure" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 323, + 517, + 481, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 517, + 481, + 525 + ], + "spans": [ + { + "bbox": [ + 323, + 517, + 481, + 525 + ], + "type": "text", + "content": "Picking up the can and place it inside the bowl." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 323, + 525, + 361, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 525, + 361, + 535 + ], + "spans": [ + { + "bbox": [ + 323, + 525, + 361, + 535 + ], + "type": "text", + "content": "PhysVLM:" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 455, + 526, + 465, + 535 + ], + "blocks": [ + { + "bbox": [ + 455, + 526, + 465, + 535 + ], + "lines": [ + { + "bbox": [ + 455, + 526, + 465, + 535 + ], + "spans": [ + { + "bbox": [ + 455, + 526, + 465, + 535 + ], + "type": "image", + "image_path": "eea56502ab52faa3fdf13e7ae0d6a176450fb57dd6a95f4b3c3661caeb8a29fb.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 323, + 536, + 537, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 536, + 537, + 554 + ], + "spans": [ + { + "bbox": [ + 323, + 536, + 537, + 554 + ], + "type": "text", + "content": "1-Move close to the can. 2-Pick up the can. 3-Place the can into the bowl. 4-Down." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 565, + 555, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 555, + 620 + ], + "type": "text", + "content": "Figure 1. Existing VLM models, such as GPT-4o, may generate inaccurate or impractical responses due to poor comprehension of robotic physical reachability. The proposed PhysVLM integrates vision-language capabilities with an understanding of robotic physical reachability." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Accurate perception of physical reachability is essential for robots to perform tasks effectively. Similar to how humans adjust their actions based on bodily conditions and environmental factors, robots must account for their physical reachability within an environment to ensure efficient and reliable task execution. For instance, in grasping tasks," + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6940" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "a robot that fails to assess its reachability may attempt to grasp an object from an unreachable position, leading to task failure or equipment damage [32, 35]. Thus, enhancing a robot's understanding of physical reachability is crucial for successful task planning and execution in complex environments [31, 41]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 146, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 146, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 146, + 296, + 397 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) have shown remarkable progress in environmental understanding [6, 17, 19, 21, 37], and many studies have applied these models to embodied AI to assist robots in perceiving environments and planning tasks [11, 18, 25, 36, 43]. However, while VLMs excel in general environmental perception, they often struggle with tasks that require an understanding of robotic physical reachability (see Figure 1). We identify two key challenges that must be addressed for VLMs to be effective in robotic tasks: (1) how to develop a unified and efficient representation of physical reachability. Robots vary significantly in size, joint types, and other characteristics, making it difficult for VLMs to directly learn these differences; (2) how to enable VLMs to improve their understanding of physical reachability without compromising general vision-language capabilities. Existing VLMs typically combine pre-trained unimodal encoders for vision and language tasks. However, introducing a new modality like physical reachability requires careful architectural and training adjustments to ensure the model can reason about reachability while maintaining its general capabilities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 399, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 649 + ], + "type": "text", + "content": "To address these challenges, we propose the Space-Physical Reachability Map (S-P Map), a unified representation that abstracts the physical reachability of diverse robots into a generalized spatial form. The S-P Map is generated by combining robot parameters with egocentric depth images, but crucially, the model learns to focus on the abstracted reachability features (i.e., the gray regions in the S-P Map) rather than the specific robot configurations. This abstraction allows the model to generalize across different robots, as it only needs to reason about which areas are reachable, independent of the robot's specific characteristics. We introduce PhysVLM, a vision-language model that extends traditional VLM architectures by incorporating an additional feature encoder to process the S-P Map. This design enables PhysVLM to integrate physical reachability information into its reasoning process without compromising its general vision-language capabilities. To train and evaluate PhysVLM, we constructed a large-scale multi-robot dataset, Phys100K, and a challenging benchmark, EQAphys, which includes tasks for six different robots in both simulated and real-world environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 652, + 246, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 652, + 246, + 662 + ], + "spans": [ + { + "bbox": [ + 67, + 652, + 246, + 662 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 665, + 295, + 714 + ], + "type": "text", + "content": "- We propose a unified and robot-agnostic formulation, the Space-Physical Reachability Map (S-P Map), which abstracts robotic physical reachability in a way that is independent of specific robot configurations, promoting the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 323, + 72, + 455, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 455, + 83 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 455, + 83 + ], + "type": "text", + "content": "learning of generalized features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 84, + 553, + 275 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 84, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 553, + 132 + ], + "type": "text", + "content": "- We introduce PhysVLM, a vision-language model that integrates physical reachability with general vision-language capabilities via an additional feature encoder, improving task execution reliability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "type": "text", + "content": "- We release the EQA-phys benchmark, which includes six robots and 1.3K question-answer pairs, designed to test the model's understanding of physical reachability in simulated and real-world environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 180, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 180, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 314, + 180, + 553, + 275 + ], + "type": "text", + "content": "- Our model achieves a " + }, + { + "bbox": [ + 314, + 180, + 553, + 275 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 314, + 180, + 553, + 275 + ], + "type": "text", + "content": " improvement over GPT-4o on the EQA-phys benchmark. In embodied visual reasoning tasks on the RoboVQA-val and OpenEQA benchmarks, it outperforms advanced embodied VLMs such as RoboMamba and SpatialVLM. Furthermore, the S-P Map demonstrates strong compatibility with various VLMs, and integrating it into GPT-4o-mini results in a " + }, + { + "bbox": [ + 314, + 180, + 553, + 275 + ], + "type": "inline_equation", + "content": "7.1\\%" + }, + { + "bbox": [ + 314, + 180, + 553, + 275 + ], + "type": "text", + "content": " performance improvement." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 288, + 400, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 288, + 400, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 288, + 400, + 300 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 308, + 422, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 308, + 422, + 319 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 422, + 319 + ], + "type": "text", + "content": "2.1. VLMs in Robotics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 326, + 553, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 553, + 409 + ], + "type": "text", + "content": "Embodied Question Answering (EQA) tasks require agents to interact with environments to answer questions [10, 13, 38]. RoboVQA offers a large, diverse dataset for robotic visual question answering. 3D-VLA [44] integrates 3D perception with a generative world model for embodied reasoning, while SpatialVLM [5] enhances VLMs' spatial understanding using extensive 3D data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 411, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 518 + ], + "type": "text", + "content": "Robot task planning involves sequencing subtasks to achieve goals [3, 30, 40]. Code as Policies (CaP) [20] employs OpenAI's Codex (code-davinci-002) to generate planning code. SayCan [2] combines the Pathways Language Model (PaLM) [7] with robotic affordances to create feasible action plans based on the robot's capabilities. However, these methods often assume all objects are within the robot's operational area, ignoring physical reachability and potentially leading to suboptimal or infeasible plans." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 527, + 510, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 510, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 510, + 540 + ], + "type": "text", + "content": "2.2. Understanding Physical Reachability" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "text", + "content": "Recent studies use voxel grids with open-vocabulary detection models to assign task-specific attributes, enabling environmental constraint understanding. ReKep [15] generates keypoint proposals and constraints using voxel grids and VLMs, while VoxPoser [14] synthesizes robot trajectories by integrating OWL-ViT [26] and VLMs with voxel-based environment representations. However, these approaches focus on environment modeling without explicitly addressing the robot's physical reachability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Explicitly representing reachable workspaces remains challenging. Reachability maps [41] model spatial capabilities, and occupancy grids [16] account for obstacles to ensure safe navigation. Additionally, methods like online model predictive control with offline workspace analysis" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6941" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "content": "[31] and Reachability Expression-based Motion Planning (REMP) [12] address workspace constraints. Despite these advancements, integrating physical reachability into visual reasoning for complex embodied tasks is still limited, primarily due to the lack of large-scale datasets that include robotic physical parameters in VLM pretraining." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 111, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 111, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 111, + 167 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 296, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 366 + ], + "type": "text", + "content": "PhysVLM is a large-scale vision-language model designed for visual reasoning that accounts for physical constraints in embodied tasks. As illustrated in Figure 2, PhysVLM integrates instruction text, visual input (RGB image), and an S-P Map abstracts the robotic physical reachability into a unified spatial representation. By combining these inputs, PhysVLM generates responses consistent with both the visual context and the robot's physical reachability, without being tied to specific robot configurations. The S-P Map is constructed using a unified physical reachability encoding method, which abstracts the physical parameters of various robots alongside their egocentric depth maps into a generalized form. This abstraction allows the model to generalize across different robots, addressing the challenge of learning and reasoning about physical reachability in a robot-agnostic manner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 368, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 416 + ], + "type": "text", + "content": "In this section, we present the core components of PhysVLM. Section 3.1 describes the S-P Map encoding method, Section 3.2 describes the model architecture, and Section 3.3 discusses the training procedures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 167, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 167, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 167, + 436 + ], + "type": "text", + "content": "3.1. S-P Map Encoding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 513 + ], + "type": "text", + "content": "As illustrated in Figure 2, we model the physical reachability of various robots using a unified approach that abstracts robot-specific parameters into a generalized spatial representation. This abstraction allows the model to focus on the spatial regions that are physically reachable, independent of the specific robot configuration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 534, + 295, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 534, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 87, + 534, + 295, + 548 + ], + "type": "interline_equation", + "content": "\\mathrm {S} - \\mathrm {P} \\quad \\operatorname {M a p} = F \\left(\\mathcal {P} _ {\\text {r a w}}, \\left\\{\\theta_ {i} ^ {\\min }, \\theta_ {i} ^ {\\max } \\right\\}, \\mathrm {D H}, \\mathbf {E}\\right), \\tag {1}", + "image_path": "a042bce2c32303fe48372bed9062bf7a8b33268631b86691fdceba7c6856a9bc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{raw}}" + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "text", + "content": " denotes the raw point cloud data from the robot's RGB-D camera. " + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\{\\theta_i^{\\mathrm{min}},\\theta_i^{\\mathrm{max}}\\}" + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "text", + "content": " represents the range of motion for each joint " + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "text", + "content": ". DH refers to the Denavit-Hartenberg parameters that describe the geometric structure of each joint, and " + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "text", + "content": " is the extrinsic calibration matrix that transforms coordinates from the camera to the robot's coordinate system. The function " + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 558, + 295, + 677 + ], + "type": "text", + "content": " maps these inputs to produce the S-P Map, which abstracts the robot's physical reachability into a spatial form that is independent of the specific robot configuration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Consider a robot arm with " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " degrees of freedom, where each joint " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " has DH parameters " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\theta_i, d_i, a_i, \\alpha_i\\}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta_i" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is the joint angle, " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is the offset along the z-axis, " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is the link" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "length, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " is the twist angle). The homogeneous transformation matrix for each joint is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 387, + 108, + 553, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 108, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 387, + 108, + 553, + 121 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {i} = G \\left(\\theta_ {i}, d _ {i}, a _ {i}, \\alpha_ {i}\\right), \\tag {2}", + "image_path": "0b95ba6e9a9bc5a5378e7dd404dcfdfc35467b5936235c57b053f2c9e419291e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 125, + 555, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 555, + 173 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 125, + 555, + 173 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 125, + 555, + 173 + ], + "type": "text", + "content": " is the standard Denavit-Hartenberg transformation function. By multiplying the transformation matrices of all joints, we obtain the transformation matrix from the base frame to the end-effector frame:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 394, + 185, + 553, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 185, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 394, + 185, + 553, + 198 + ], + "type": "interline_equation", + "content": "\\mathbf {T} = \\mathbf {T} _ {1} \\mathbf {T} _ {2} \\dots \\mathbf {T} _ {n}. \\tag {3}", + "image_path": "357039c16e7d7b8e946c73abe53ac6d8cf0719b15eb29d0dca81b3d16580eb4f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 202, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 554, + 262 + ], + "type": "text", + "content": "To generate joint configurations, we sample the joint angles " + }, + { + "bbox": [ + 313, + 202, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 202, + 554, + 262 + ], + "type": "text", + "content": " from their respective motion ranges " + }, + { + "bbox": [ + 313, + 202, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\theta_{i} \\in [\\theta_{i}^{\\min}, \\theta_{i}^{\\max}]" + }, + { + "bbox": [ + 313, + 202, + 554, + 262 + ], + "type": "text", + "content": ", resulting in configurations " + }, + { + "bbox": [ + 313, + 202, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\{\\theta_{1}, \\theta_{2}, \\ldots, \\theta_{n}\\}" + }, + { + "bbox": [ + 313, + 202, + 554, + 262 + ], + "type": "text", + "content": ". By substituting these joint configurations into the forward kinematics equations, we compute the corresponding end-effector positions:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 341, + 278, + 553, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 278, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 341, + 278, + 553, + 306 + ], + "type": "interline_equation", + "content": "\\mathcal {W} _ {\\text {v o x e l}} = \\left\\{\\mathbf {p} \\mid \\mathbf {p} = \\mathbf {T} \\left(\\theta_ {1}, \\theta_ {2}, \\dots , \\theta_ {n}\\right) \\cdot \\mathbf {p} _ {0} \\right\\}, \\tag {4}", + "image_path": "ac2565b630fd0729dee9a6fced29dfe9d3eaa11af9be42d6c5ef028852d3d204.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 312, + 553, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 359 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 312, + 553, + 359 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_0" + }, + { + "bbox": [ + 313, + 312, + 553, + 359 + ], + "type": "text", + "content": " is the origin point in the end-effector frame. We precompute these joint configurations offline, discretize the workspace into a voxel grid " + }, + { + "bbox": [ + 313, + 312, + 553, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{\\mathrm{voxel}}" + }, + { + "bbox": [ + 313, + 312, + 553, + 359 + ], + "type": "text", + "content": ", and store it for efficient computation in subsequent steps." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "text", + "content": "Next, as shown in Figure 2, the robot's raw point cloud " + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{raw}}" + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "text", + "content": " is captured from its egocentric RGB-D camera in the camera coordinate system and transformed into the robot's coordinate system using the extrinsic calibration matrix " + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "text", + "content": ", resulting in the transformed point cloud " + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 404, + 431, + 553, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 431, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 404, + 431, + 553, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {P} = \\mathbf {E} \\cdot \\mathcal {P} _ {\\text {r a w}}. \\tag {5}", + "image_path": "265d939882a52f54813ea1a99042f9c2c20acf6b99f824862ae88e36e941f512.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 449, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 485 + ], + "type": "text", + "content": "To ensure physical feasibility, we perform a voxel grid lookup to determine whether each point in " + }, + { + "bbox": [ + 313, + 449, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 449, + 554, + 485 + ], + "type": "text", + "content": " lies within the precomputed reachable workspace " + }, + { + "bbox": [ + 313, + 449, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{\\mathrm{voxel}}" + }, + { + "bbox": [ + 313, + 449, + 554, + 485 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 365, + 492, + 553, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 492, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 365, + 492, + 553, + 519 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {\\text {v a l i d}} = \\left\\{\\mathbf {p} \\in \\mathcal {P} \\mid \\mathbf {p} \\in \\mathcal {W} _ {\\text {v o x e l}} \\right\\}. \\tag {6}", + "image_path": "1a567154217658f790eae6b556fdd1946be741be66a69f077ec01422ce4dd0fb.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 522, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 558 + ], + "type": "text", + "content": "This step filters the point cloud to include only points within the robot's reachable workspace, abstracting the robot's physical reachability into a generalized spatial form." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Finally, we transform the valid point cloud " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{valid}}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " back into the camera coordinate system and use the camera's intrinsic parameters to project these points onto the image plane. We then mark the regions that comply with physical reachability on the original depth map. For areas that are not reachable, we apply a gray mask and outline their boundaries. The resulting S-P Map clearly highlights regions that are beyond the robot's physical reach, providing a unified and abstracted representation of reachability that is independent of the specific robot configuration. This allows the model to focus on the spatial constraints of the task, without needing to account for the detailed physical parameters of each robot." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6942" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 86, + 255, + 273 + ], + "blocks": [ + { + "bbox": [ + 59, + 86, + 255, + 273 + ], + "lines": [ + { + "bbox": [ + 59, + 86, + 255, + 273 + ], + "spans": [ + { + "bbox": [ + 59, + 86, + 255, + 273 + ], + "type": "image", + "image_path": "88d43d2171b258406312fcfa33b9d14fb4805f02ceb4d972a68f9f1cff100ba1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 281, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 54, + 281, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 555, + 316 + ], + "type": "text", + "content": "Figure 2. Overview of PhysVLM. Starting with robot parameters and an egocentric depth map, an S-P Map is generated via unified physical reachability encoding. Using the S-P Map, image, and instruction text, PhysVLM generates textual output considering the robot's physical reachability." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 258, + 86, + 553, + 274 + ], + "blocks": [ + { + "bbox": [ + 258, + 86, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 258, + 86, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 258, + 86, + 553, + 274 + ], + "type": "image", + "image_path": "32cc7a7325b209c2bfe43dd9c2f2cbef4029ba37c6ccdc2828b60eaa5e7b3c9d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 339, + 170, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 170, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 170, + 350 + ], + "type": "text", + "content": "3.2. Model Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 359, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 456 + ], + "type": "text", + "content": "To seamlessly integrate robotic physical reachability into PhysVLM while preserving its visual reasoning capabilities, we design a dual-branch architecture: one branch dedicated to vision processing and the other to physical reachability (see Figure 2). These branches operate independently, extracting features from their respective inputs, which are then fused and passed to a unified decoder for final reasoning and response generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 457, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 296, + 540 + ], + "type": "text", + "content": "The vision branch leverages a pre-trained Vision Transformer (ViT) [9], specifically the SigLip-400M model [42], to extract high-level visual features from egocentric images. To reduce computational overhead, a Max Pooling layer is applied, followed by a two-layer Multi-Layer Perceptron (MLP) that transforms the visual features into token representations suitable for multimodal fusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 638 + ], + "type": "text", + "content": "The physical reachability branch processes the S-P Map, which abstracts the robot's physical reachability into a generalized spatial form. This branch also utilizes the SigLip-400M model for feature extraction, followed by Max Pooling and a feature fusion layer. The fusion layer combines the visual and reachability features, and a two-layer MLP further refines these fused features into reachability-specific tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "For the language decoding, we employ the Qwen-2.5-Instruct-3B model [33, 39] as PhysVLM's large language model (LLM) decoder, using the Qwen-2.5 tokenizer to process natural language instructions. The decoder integrates multimodal tokens from the vision branch, S-P Maps, and language inputs, generating coherent and contextually rele" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 340, + 555, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 364 + ], + "type": "text", + "content": "vant textual responses that account for both visual and physical reachability information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 372, + 378, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 378, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 378, + 384 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 389, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 485 + ], + "type": "text", + "content": "Training Data Construction. The training data for PhysVLM consists of our Phys100K dataset and general VQA datasets, such as LLaVA-Pretrain, ShareGPT4V, and RoboVQA. Phys100K focuses on question-answering related to physical reachability, aggregating data from RoboVQA (20K samples), ScanNet [8] (10K samples), OpenX-Embodiment [27] (60K samples), and an additional 10K samples from PyBullet." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": "Depth maps are essential inputs for generating the S-P Map. For datasets lacking depth maps, we generate them using DepthAnything-v2. Additionally, we employ Grounding DINO [23] and SAM2 [34] to obtain 2D bounding boxes and segmentation results for objects in the images. In PyBullet, we simulate work scenarios using four robotic arms (UR5, FR5, CR5, and FRANKA) to collect RGB images, depth maps, and segmentation results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "For the PyBullet data in Phys100K, precise robot configurations can be obtained from the simulator. Therefore, we directly generate the S-P Map using the method described in Section 3.2, and labels indicating whether objects are reachable are obtained through simulated motion. The advantage of the S-P Map is that it abstracts physical reachability into a region-based representation, decoupling the learning process from specific robot configurations. This allows us to generate pseudo-labels for datasets without precise robot parameters. We approximate reachability using the segmentation results, marking regions and the objects within" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6943" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 86, + 555, + 326 + ], + "blocks": [ + { + "bbox": [ + 57, + 86, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 57, + 86, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 57, + 86, + 555, + 326 + ], + "type": "image", + "image_path": "2fc2e9b0c0203553b91d189e788739d096524d5d6068b12942995c3621fb84ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 335, + 433, + 346 + ], + "lines": [ + { + "bbox": [ + 176, + 335, + 433, + 346 + ], + "spans": [ + { + "bbox": [ + 176, + 335, + 433, + 346 + ], + "type": "text", + "content": "Figure 3. Details of the Phys100K Dataset and EQA-Phys Benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 371, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 407 + ], + "type": "text", + "content": "them as \"reachable\" or \"unreachable\" based on depth values. Next, we generate question-answer pairs for two main categories:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 410, + 295, + 625 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 410, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 493 + ], + "type": "text", + "content": "- Embodied QA. GPT-4 generates question-answer pairs for ScanNet and RoboVQA, covering categories include Function Reasoning, World Knowledge, Object Recognition, Object Localization, Attribute Recognition, Spatial Reasoning, Object State Recognition, and Hallucination (See Figure 3). Detailed prompts and examples are provided in the appendix to guide the generation process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 625 + ], + "type": "text", + "content": "- Tasks Involving Physical Reachability. We use the \"reachable\" label with five fixed task templates to generate question-answer pairs, such as \"USER:\\n\\n Is the [Object] in the robot's reachable space? ASSISTANT: Yes, it is.\" Here, [Object] represents the relevant object category, while and serve as placeholders for image patch tokens and S-P Map patch tokens, respectively. Figure 3 provides examples of question-answer pairs for each object category." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Training Pipeline. We adopt a two-stage training process to fully leverage the S-P Map and ensure PhysVLM generalizes across different robots. In the first stage, we align multimodal features using the LLaVA-Pretrain and OpenX-Embodiment datasets from Phys100K. This stage only trains the projection layers, allowing the model to build" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 371, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 396 + ], + "type": "text", + "content": "a foundational understanding of visual inputs and physical reachability, independent of specific robot configurations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 398, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 469 + ], + "type": "text", + "content": "In the second stage, we unfreeze all parameters and train the entire model using data from Phys100K, ShareGPT4V, and RoboVQA. This stage enhances PhysVLM's ability to handle complex visual reasoning tasks with physical reachability constraints, ensuring the model can generalize across diverse environments and robots." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 555, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 571 + ], + "type": "text", + "content": "Implementation Details. PhysVLM is trained for 48 hours using eight A800 GPUs. The training process consists of two stages, each lasting one epoch. The batch size and learning rate are set at 128 and 1e-3 in the first stage, and 64 and 1e-5 in the second stage. The final model is PhysVLM-3B." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 586, + 443, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 443, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 443, + 599 + ], + "type": "text", + "content": "3.4. EQA-phys Benchmark" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "As illustrated in Figure 3, we introduce an embodied QA task focused on physical reachability, termed EQA-phys, which emphasizes QA tasks constrained by physical limitations. This benchmark includes a simulator dataset with 200 samples and 1,000 questions from the PyBullet validation set, as well as a zero-shot evaluation set based on real-world data from UR3 and XArm6 robots in two scenarios. The evaluation set contains 60 samples and 300 questions, all manually annotated by domain experts." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6944" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 101, + 547, + 240 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 511, + 82 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 511, + 82 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 511, + 82 + ], + "type": "text", + "content": "Table 1. Results of EQA-phys. Comparison of PhysVLM-3B (ours) with API-based VLMs and embodied VLMs." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 101, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 63, + 101, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 63, + 101, + 547, + 240 + ], + "type": "table", + "html": "
REAL-WORLDSIMULATORALL
UR3XARM6UR5FR5CR5FRANKA
API-BASED VLMSGPT-4O-MINI54.356.049.455.454.647.152.8
CLAUDE-3.556.260.554.058.155.754.356.4
GPT-4O56.761.555.758.357.552.657.0
GPT-4O-MINI + S-P MAP60.0↑5.760.5↑4.557.0↑7.659.1↑3.759.2↑4.653.3↑6.259.8↑7.0
CLAUDE-3.5 + S-P MAP65.3↑9.167.3↑6.854.9↑0.958.3↑0.258.2↑2.558.1↑3.860.3↑3.4
GPT-4O + S-P MAP66.6↑9.968.1↑6.655.8↑0.160.7↑1.459.4↑1.957.6↑5.061.3↑4.1
EMBODIED VLMSSPATIALVLM56.355.154.659.152.047.554.1
SPATIALBOT51.150.250.048.153.354.451.1
PHYSVLM-3B64.163.071.475.774.078.171.0
", + "image_path": "f593b02147f76713dcfdf7ea942cc4efcdb3662a1f3172f6219f4cf60d7c1167.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 261, + 136, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 136, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 136, + 274 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 281, + 178, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 178, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 178, + 294 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 298, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 295, + 323 + ], + "type": "text", + "content": "Tasks. We compare the performance of PhysVLM with other methods across three categories of tasks:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 325, + 296, + 492 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 325, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 295, + 384 + ], + "type": "text", + "content": "- EQA-phys. This benchmark tests the model's ability to integrate visual reasoning with robotic physical reachability. The real-robot component is used to assess PhysVLM's zero-shot generalization, highlighting its ability to handle unseen robots and environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 384, + 295, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 419 + ], + "type": "text", + "content": "- Embodied QA. We evaluate the model's general visual reasoning ability in embodied tasks using the OpenEQA [25] and RoboVQA-val [29] benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 296, + 492 + ], + "type": "text", + "content": "- Robot Task Planning. For real-world tasks such as \"Pick A into B\", we assess the model's ability to understand robotic physical reachability and generate reasonable task plans. As this study does not focus on robot control strategies, we use the natural language planning approach from [29]." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 506, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 615 + ], + "type": "text", + "content": "Baselines. We compare our model to several baselines, including API-accessible VLMs such as Claude 3.5 [28], GPT-4o-mini [1], and GPT-4o [1], as well as embodied VLMs like SpatialVLM [5], SpatialBot [4], 3D-VLA [44], and RoboMamba [22]. SpatialVLM and SpatialBot both use the 3B version, which has a similar parameter count to our model. Since the executable versions of 3D-VLA and RoboMamba are unavailable, we compare their reported results on RoboVQA-val." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. For tasks involving physical reachability, we use LLM scoring, following the approach used in existing studies [24, 25]. Assigning 5 points for completely correct responses and 1 point for incorrect responses, we calculate the average score and express it as a percentage. For Embodied QA, we follow the benchmark settings of the corresponding datasets [25, 29]. For task planning, each" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 262, + 554, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 286 + ], + "type": "text", + "content": "task type is executed 10 times, and the average success rate serves as the evaluation metric." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 309, + 437, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 437, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 437, + 323 + ], + "type": "text", + "content": "4.2. Results on EQA-phys" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 332, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 452 + ], + "type": "text", + "content": "Table 1 shows the results on EQA-phys. Neither API-based nor embodied VLMs can handle the robot's parameter constraints, resulting in suboptimal outputs with scores around " + }, + { + "bbox": [ + 313, + 332, + 555, + 452 + ], + "type": "inline_equation", + "content": "55\\%" + }, + { + "bbox": [ + 313, + 332, + 555, + 452 + ], + "type": "text", + "content": ". In contrast, our model successfully completes visual reasoning tasks involving physical reachability, obtaining an average score of " + }, + { + "bbox": [ + 313, + 332, + 555, + 452 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 313, + 332, + 555, + 452 + ], + "type": "text", + "content": ". As discussed, these tasks require the model to perform visual reasoning based on an understanding of the robotic physical reachability. The model can only effectively accomplish these tasks by truly understanding robotic physical reachability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 456, + 556, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 556, + 589 + ], + "type": "text", + "content": "The results in Table 1 demonstrate that prompting API-based VLMs, such as GPT-40, with the S-P Map (detailed in 4) significantly enhances their performance. This improvement stems from the S-P Map's ability to abstract physical reachability into a robot-agnostic representation, enabling VLMs to reason about physical constraints that would otherwise be beyond their capabilities. By decoupling reachability from specific robot parameters, the S-P Map facilitates generalization across diverse environments, allowing models to better comprehend physical reachability, even in previously unseen scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "Table 1 also demonstrates that PhysVLM-3B achieved scores of over " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "63\\%" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": " in zero-shot evaluations for UR3 and XArm6 robots, despite operating in new environments with different robot parameters. This performance is attributed to two key factors: (1) the S-P Map abstracts various robot parameters into a unified, transferable representation of physical reachability, and (2) the model's independent visual and constraint encoding branches allow it to learn generalizable visual features from diverse image-text data, enabling effective reasoning in novel environments." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6945" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 123, + 299, + 205 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 2. Embodied QA results on the RoboVQA-val set, comparison of PhysVLM (ours) with existing methods. An asterisk (*) indicates models not pre-trained on the RoboVQA dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 123, + 299, + 205 + ], + "lines": [ + { + "bbox": [ + 56, + 123, + 299, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 123, + 299, + 205 + ], + "type": "table", + "html": "
BLEU1BLEU2BLEU3BLEU4
SPATIALVLM*5.13.01.91.2
SPATIALBOT*12.49.38.07.2
3D-VLA48.338.531.726.8
ROBOMAMBA54.944.239.536.3
PHYSVLM-3B65.362.450.943.5
", + "image_path": "6ee41cde6cc2048172e628d13e387d3c0a1e9d2a3cecdb9ea3b2971b02c012d7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 281, + 282, + 373 + ], + "blocks": [ + { + "bbox": [ + 55, + 217, + 296, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 262 + ], + "type": "text", + "content": "Table 3. Embodied QA results on the OpenEQA benchmark, comparison of PhysVLM (ours) with existing methods. An asterisk (*) indicates that only the first 200 samples were tested due to API limitations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 281, + 282, + 373 + ], + "lines": [ + { + "bbox": [ + 69, + 281, + 282, + 373 + ], + "spans": [ + { + "bbox": [ + 69, + 281, + 282, + 373 + ], + "type": "table", + "html": "
EM-EQA (SCANNET)EM-EQA (HM3D)ALL
SPATIALVLM42.944.343.8
SPATIALBOT45.351.049.1
GPT4V57.451.355.3
GPT-4O*68.265.266.7
PHYSVLM-3B60.751.257.4
", + "image_path": "63103c3fe4040698bcab4fc17acca0a367830fa12d6ad0787c12fd053a770c0e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 394, + 198, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 198, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 198, + 407 + ], + "type": "text", + "content": "4.3. Results on Embodied QA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 411, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 297, + 544 + ], + "type": "text", + "content": "We demonstrate our model's effectiveness in handling general embodied visual reasoning tasks. Additionally, we show that incorporating an understanding of physical constraints does not diminish its general visual reasoning capabilities. We compare our model with state-of-the-art embodied VLMs (see Tables 2 and 3). Our model achieves the best performance on the RoboVQA-val benchmark, surpassing other models by " + }, + { + "bbox": [ + 54, + 411, + 297, + 544 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 54, + 411, + 297, + 544 + ], + "type": "text", + "content": " in BLEU-4. On the OpenEQA benchmark, our model outperforms existing embodied VLMs and GPT-4V, ranking second, behind only GPT-4o." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 230, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 230, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 230, + 566 + ], + "type": "text", + "content": "4.4. Results on Robot Task Planning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 570, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 297, + 715 + ], + "type": "text", + "content": "Table 4 presents the performance of PhysVLM and baseline models on real-world task planning scenarios. When all objects are within the robot's physical reach, PhysVLM performs similarly to other models, as directly grabbing or placing the objects succeeds. However, when some objects are outside the physical reach, the model must suggest that the robot move closer before grabbing or placing them. In these cases, our model performs exceptionally well, whereas the task success rates of other models decline significantly. This is attributed to our model's understanding of robotic physical reachability and its ability to incorporate this understanding in task planning." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 331, + 111, + 537, + 213 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 4. Task planning results. Comparison of PhysVLM (ours) with other VLMs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 111, + 537, + 213 + ], + "lines": [ + { + "bbox": [ + 331, + 111, + 537, + 213 + ], + "spans": [ + { + "bbox": [ + 331, + 111, + 537, + 213 + ], + "type": "table", + "html": "
ALL OBJECTS IN RANGEPART OBJECTS IN RANGE
GPT-4O-MINI70.523.2
CLAUDE-3.573.632.1
GPT-4O75.935.8
SPATIALVLM64.421.5
SPATIALBOT65.625.3
PHYSVLM-3B69.248.4
", + "image_path": "ddf4325ab7f84dbf086c41468d85adf41ecb9d3ac0da06bf098ed35734feceb8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 331, + 279, + 537, + 348 + ], + "blocks": [ + { + "bbox": [ + 313, + 227, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 227, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 555, + 262 + ], + "type": "text", + "content": "Table 5. Ablation study on the S-P Map. We compare the constraint encoder's performance when using the S-P Map, replacing it with a Depth Map, or providing no input." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 279, + 537, + 348 + ], + "lines": [ + { + "bbox": [ + 331, + 279, + 537, + 348 + ], + "spans": [ + { + "bbox": [ + 331, + 279, + 537, + 348 + ], + "type": "table", + "html": "
IDS-P MAPDEPTH MAPEQA-PHYS REALEQA-PHYS SIM
1✓63.574.8
2✓58.162.4
354.258.8
", + "image_path": "34f2c2e11c1922adda8e1044d30ea755d567d0f868bc922912cf03a52869c843.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 371, + 409, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 409, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 409, + 384 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 389, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 426 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to evaluate the contribution of each component in PhysVLM. We report the average LLM score on tasks involving physical reachability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "type": "text", + "content": "The effectiveness of S-P Map. To demonstrate the S-P Map's contribution, we compare results obtained with and without its input. As shown in Table 5, Experiments 1 and 3 demonstrate that omitting the S-P Map leads to a significant performance decrease for both zero-shot real-world robots and simulators. Specifically, the overall average score drops by " + }, + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "type": "text", + "content": " in simulation results and " + }, + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "type": "inline_equation", + "content": "9.3\\%" + }, + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "type": "text", + "content": " in real-world robot evaluations. Without the S-P Map input, the model struggles to handle the robotic physical reachability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 552, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 556, + 624 + ], + "type": "text", + "content": "Additionally, Experiments 1 and 2 demonstrate that replacing the S-P Map with a Depth Map significantly degrades the model's performance on zero-shot tasks. Since the Depth Map does not accurately represent robotic physical reachability, the model cannot rely solely on depth information to understand it." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "content": "Effectiveness of an additional feature encoder. To demonstrate the effectiveness of the model architecture, we compare the performance of a feature encoder that shares weights with the visual feature encoder for the S-P Map. In these experiments, we evaluate the average scores on both the EQA-phys and OpenEQA benchmarks. The results in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "text", + "content": "6946" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 85, + 541, + 264 + ], + "blocks": [ + { + "bbox": [ + 72, + 85, + 541, + 264 + ], + "lines": [ + { + "bbox": [ + 72, + 85, + 541, + 264 + ], + "spans": [ + { + "bbox": [ + 72, + 85, + 541, + 264 + ], + "type": "image", + "image_path": "55e73a6e3b0b10b01833172bb38edd36b9835b507e965ff9dd727e592c519a18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 273, + 440, + 283 + ], + "lines": [ + { + "bbox": [ + 169, + 273, + 440, + 283 + ], + "spans": [ + { + "bbox": [ + 169, + 273, + 440, + 283 + ], + "type": "text", + "content": "Figure 4. Visual comparison of PhysVLM (ours), GPT-4o, and SpatialBot." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 79, + 359, + 271, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 308, + 295, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 295, + 340 + ], + "type": "text", + "content": "Table 6. Ablation study on the effectiveness of an additional feature encoder. Share indicates shared network and weights with the visual feature encoder." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 359, + 271, + 407 + ], + "lines": [ + { + "bbox": [ + 79, + 359, + 271, + 407 + ], + "spans": [ + { + "bbox": [ + 79, + 359, + 271, + 407 + ], + "type": "table", + "html": "
EQA-PHYSOPENEQA
INDEPENDENT71.057.4
SHARE68.256.5
", + "image_path": "6bac1a881dbcbe0a03e69626bf144fa0171cfa6f268acf06f6845af7db9459da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 489 + ], + "type": "text", + "content": "Table 6 show that sharing the feature encoder not only decreases performance on EQA-phys but also impairs general visual reasoning capabilities. This is because S-P Map features differ from images, and the training data contains significantly more image-text pairs than S-P Map data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 296, + 588 + ], + "type": "text", + "content": "Effectiveness of training data. To evaluate the effectiveness of Phys100K, we conduct experiments by selectively removing data from various sources in Phys100K. As shown in Table 7, removing data from PyBullet or other embodied datasets leads to a reduction in overall performance, highlighting the critical role of each data component in the model's performance." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 69, + 628, + 281, + 696 + ], + "blocks": [ + { + "bbox": [ + 65, + 597, + 285, + 609 + ], + "lines": [ + { + "bbox": [ + 65, + 597, + 285, + 609 + ], + "spans": [ + { + "bbox": [ + 65, + 597, + 285, + 609 + ], + "type": "text", + "content": "Table 7. Ablation study on the effectiveness of training data." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 628, + 281, + 696 + ], + "lines": [ + { + "bbox": [ + 69, + 628, + 281, + 696 + ], + "spans": [ + { + "bbox": [ + 69, + 628, + 281, + 696 + ], + "type": "table", + "html": "
PART OF PHYS100KEQA-PHYS REALEQA-PHYS SIM
ALL63.574.8
W/O PYBULET62.165.4
W/O OTHER DATASETS58.671.5
", + "image_path": "d73143653ef2f520880acbe459efbb08d5606c6541cca936738abdcb1974ad20.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 308, + 425, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 308, + 425, + 320 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 425, + 320 + ], + "type": "text", + "content": "4.6. Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 328, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 411 + ], + "type": "text", + "content": "Figure 4 compares our method with SpatialBot and GPT-4o. SpatialBot uses depth maps and images, while GPT-4o uses standard images. Both struggle with tasks requiring physical reachability, causing visual reasoning errors. In contrast, our approach delivers accurate results. Additionally, incorporating the S-P Map into GPT-4o improves its handling of physical reachability and response accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 428, + 387, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 428, + 387, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 428, + 387, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "type": "text", + "content": "We introduce PhysVLM, a VLM that incorporates physical reachability into visual reasoning for robotic tasks. The S-P Map provides a unified representation of robotic reachability, facilitating the learning of generalizable features. PhysVLM extends traditional VLMs by adding a physical reachability encoder, enabling the simultaneous processing of visual, reachability, and textual information. Additionally, we present EQA-phys, a benchmark for evaluating embodied QA tasks involving physical reachability. Our experiments show that PhysVLM outperforms existing models, achieving a " + }, + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "type": "text", + "content": " higher score than GPT-4o on EQA-phys. A limitation is its reduced zero-shot performance on real robots compared to simulations, likely due to the domain gap. Future work will focus on expanding datasets, enhancing real-world performance, and improving the understanding of physical accessibility in vision-language-action models. PhysVLM's reachability awareness supports safer and more reliable robotic decision-making in industrial and assistive settings, while its unified representation ensures cross-platform adaptability for real-world deployment, bridging crucial gaps between environmental perception and actionable robotic intelligence." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6947" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "type": "text", + "content": "This work was supported by National Key R&D Program of China under Grant No.2022ZD0160601, 2022YFB4300400, and 2018B030322016, National Natural Science Foundation of China under Grants 62176254, 62276260, and U1701266." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 180, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 180, + 294, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 294, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 294, + 234 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 236, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 236, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 296, + 269 + ], + "type": "text", + "content": "[2] Michael Ahn, Brohan, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 270, + 295, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 295, + 324 + ], + "type": "text", + "content": "[3] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 326, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 295, + 369 + ], + "type": "text", + "content": "[4] Wenxiao Cai, Yaroslav Ponomarenko, Jianhao Yuan, Xiaqi Li, Wankou Yang, Hao Dong, and Bo Zhao. Spatialbot: Precise spatial understanding with vision language models. arXiv preprint arXiv:2406.13642, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 370, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 296, + 434 + ], + "type": "text", + "content": "[5] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 502 + ], + "type": "text", + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 535 + ], + "type": "text", + "content": "[7] Aakanksha Chowdhery, Narang, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1-113, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 591 + ], + "type": "text", + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 592, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 295, + 624 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 60, + 625, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 625, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 60, + 625, + 295, + 678 + ], + "type": "text", + "content": "[10] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 58, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 680, + 295, + 713 + ], + "type": "text", + "content": "[11] Kiana Ehsani, Tanmay Gupta, Rose Hendrix, Jordi Salvador, Luca Weihs, Kuo-Hao Zeng, Kunal Pratap Singh, Yejin Kim, Winson Han, Alvaro Herrasti, et al. Spoc: Imitating shortest" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "text", + "content": "paths in simulation enables effective navigation and manipulation in the real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16238-16250, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 119, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 554, + 172 + ], + "type": "text", + "content": "[12] Xiaofeng Gao, Luyao Yuan, Tianmin Shu, Hongjing Lu, and Song-Chun Zhu. Show me what you can do: Capability calibration on reachable workspace for human-robot collaboration. IEEE Robotics and Automation Letters, 7(2):2644-2651, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "text", + "content": "[13] Yining Hong, Zishuo Zheng, Peihao Chen, Yian Wang, Junyan Li, and Chuang Gan. Multiply: A multisensory object-centric embodied large language model in 3d world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26406-26416, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[14] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[15] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "type": "text", + "content": "[16] Lorenzo Jamone, Martim Brandao, Lorenzo Natale, Kenji Hashimoto, Giulio Sandini, and Atsuo Takanishi. Autonomous online generation of a motor representation of the workspace for intelligent whole-body reaching. Robotics and Autonomous Systems, 62(4):556-567, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "type": "text", + "content": "[17] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 487 + ], + "type": "text", + "content": "[18] Xiaohan Lei, Min Wang, Wengang Zhou, Li Li, and Houqiang Li. Instance-aware exploration-verification-exploitation for instance imagegoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16329–16339, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "text", + "content": "[19] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 554, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 554, + 588 + ], + "type": "text", + "content": "[20] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "text", + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[22] Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Lily Lee, Kaichen Zhou, Pengju An, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. Robomamba: Multimodal state space model for efficient robot reasoning and manipulation. arXiv preprint arXiv:2406.04339, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[23] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding din: Marrying din with" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6948" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 295, + 194 + ], + "type": "text", + "content": "[25] Arjun Majumdar, Ajay, et al. Openeca: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 226 + ], + "type": "text", + "content": "[26] Matthias Minderer, Alexey Gritsenko, Austin Stone, et al. Simple open-vocabulary object detection with vision transformers, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "text", + "content": "[27] Abby O'Neill, Rehman, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "text", + "content": "[28] Alec Radford, Kim, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 369 + ], + "type": "text", + "content": "[29] Pierre Sermanet, Ding, et al. Robovqa: Multimodal long-horizon reasoning for robotics. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 645-652. IEEE, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 295, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 436 + ], + "type": "text", + "content": "[30] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "text", + "content": "[31] Chen Song and Darwin Lau. Workspace-based model predictive control for cable-driven robots. IEEE Transactions on Robotics, 38(4):2577-2596, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "type": "text", + "content": "[32] Giovanni Sutanto, Austin Wang, Yixin Lin, Mustafa Mukadam, Gaurav Sukhatme, Akshara Rai, and Franziska Meier. Encoding physical constraints in differentiable newton-euler algorithm. In Proceedings of the 2nd Conference on Learning for Dynamics and Control, pages 804-813. PMLR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 190, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 190, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 190, + 548 + ], + "type": "text", + "content": "[33] Qwen Team. qwen2.5, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 602 + ], + "type": "text", + "content": "[34] Dominique Thomas, Rodney Rothstein, Nathan Rosenberg, and Yolande Surdin-Kerjan. Sam2 encodes the second methionine s-adenosyl transferase in saccharomyces cerevisiae: physiology and regulation of both enzymes. Molecular and Cellular Biology, 8(12):5132-5139, 1988. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 678 + ], + "type": "text", + "content": "[35] Yunsheng Tian, Karl DD Willis, Bassel Al Omari, Jieliang Luo, Pingchuan Ma, Yichen Li, Farhad Javid, Edward Gu, Joshua Jacob, Shinjiro Sueda, et al. Asap: Automated sequence planning for complex robotic assembly with physical feasibility. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4380-4386. IEEE, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "type": "text", + "content": "[36] An Dinh Vuong, Minh Nhat Vu, Baoru Huang, Nghia Nguyen, Hieu Le, Thieu Vo, and Anh Nguyen. Language-driven grasp detection. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 475 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 17902-17912, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[37] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "type": "text", + "content": "[38] Yu Wu, Lu Jiang, and Yi Yang. Revisiting embodiedqa: A simple baseline and beyond. IEEE Transactions on Image Processing, 29:3984-3992, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 207 + ], + "type": "text", + "content": "[39] An Yang, Baosong Yang, Binyuan Hui, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 273 + ], + "type": "text", + "content": "[40] Yijun Yang, Tianyi Zhou, Kanxue Li, Dapeng Tao, Lusong Li, Li Shen, Xiaodong He, Jing Jiang, and Yuhui Shi. Embodied multi-modal agent trained by an llm from a parallel textworld. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26275-26285, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 275, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 328 + ], + "type": "text", + "content": "[41] Franziska Zacharias, Christoph Borst, and Gerd Hirzinger. Capturing robot workspace structure: representing robot capabilities. In 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3229-3236. IEEE, 2007. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 363 + ], + "type": "text", + "content": "[42] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 430 + ], + "type": "text", + "content": "[43] Ganlong Zhao, Guanbin Li, Weikai Chen, and Yizhou Yu. Over-nav: Elevating iterative vision-and-language navigation with open-vocabulary detection and structured representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16296–16306, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "type": "text", + "content": "[44] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 2, 6" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6949" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/88abcf25-51f5-40e1-9600-4a2e38b6cc9a_content_list.json b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/88abcf25-51f5-40e1-9600-4a2e38b6cc9a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4a25c8d63fc8ee5e537076bc15d1457c74f24ca6 --- /dev/null +++ b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/88abcf25-51f5-40e1-9600-4a2e38b6cc9a_content_list.json @@ -0,0 +1,1645 @@ +[ + { + "type": "text", + "text": "Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment", + "text_level": 1, + "bbox": [ + 99, + 130, + 897, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hiromu Taketsugu $^{1}$ Takeru Oba $^{1}$ Takahiro Maeda $^{1}$ Shohei Nobuhara $^{2}$ Norimichi Ukita $^{1}$", + "bbox": [ + 258, + 179, + 736, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Toyota Technological Institute, Japan $^{2}$ Kyoto Institute of Technology, Japan", + "bbox": [ + 176, + 215, + 820, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sd23426, sd21502, sd21601}@toyota-ti.ac.jp, nob@kit.ac.jp, ukita@toyota-ti.ac.jp", + "bbox": [ + 138, + 236, + 852, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 279, + 326, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans can predict future human trajectories even from momentary observations by using human pose-related cues. However, previous Human Trajectory Prediction (HTP) methods leverage the pose cues implicitly, resulting in implausible predictions. To address this, we propose Locomotion Embodiment, a framework that explicitly evaluates the physical plausibility of the predicted trajectory by locomotion generation under the laws of physics. While the plausibility of locomotion is learned with an indifferentiable physics simulator, it is replaced by our differentiable Locomotion Value function to train an HTP network in a data-driven manner. In particular, our proposed Embodied Locomotion loss is beneficial for efficiently training a stochastic HTP network using multiple heads. Furthermore, the Locomotion Value filter is proposed to filter out implausible trajectories at inference. Experiments demonstrate that our method enhances even the state-of-the-art HTP methods across diverse datasets and problem settings. Our code is available at: https://github.com/ImIntheMiddle/EmLoco.", + "bbox": [ + 89, + 311, + 485, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 618, + 220, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human Trajectory Prediction (HTP), which predicts future human trajectories, is actively studied [2, 18, 29, 37, 49, 63, 64, 70] due to its importance for applications such as autonomous driving [3, 8, 41, 60, 61, 70], robotics [34, 37, 50, 73], and security systems [10, 66, 72]. These HTP methods assume the availability of sufficiently long past observations and rely solely on a past trajectory. Due to this limitation, these methods cannot handle safety-critical situations in a momentary observation setting, where only a few frames of the past trajectory are available, such as when people suddenly emerge from behind obstacles.", + "bbox": [ + 88, + 643, + 485, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Regarding this issue, prior work [26, 35, 42, 47, 50] validates that human pose-related cues, such as body orientations and limb motions are beneficial for inferring future trajectories, especially in a momentary observation setting. These methods learn the consistency between human past poses and future trajectories in a data-driven manner alone.", + "bbox": [ + 88, + 810, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b0fb1b8b9167e5f301ade0bd7c15d736ad54e579c65d1900ed1b5bebc01d9b53.jpg", + "image_caption": [ + "Figure 1. Overview of our method. Unlike existing methods, which often predict physically implausible trajectories, our framework uses locomotion generation in a physics simulator to incorporate the laws of physics to HTP by training the plausibility score as the consistency between the observed 3D pose and future possible trajectories, which are indicated by $\\boldsymbol{j}_0$ and $\\tau_f^{1,2,3}$ , respectively, in the figure. Additionally, at inference, our method can evaluate predicted trajectories, $\\hat{\\tau}_f^{1,2,3}$ , to filter out implausible ones." + ], + "image_footnote": [], + "bbox": [ + 516, + 286, + 905, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The data-driven approach can prove its ability with a large-scale dataset that covers most possible combinations of poses and trajectories. However, such datasets are unavailable in the literature because human locomotion is diverse. In particular, it is difficult to collect a large number of accurate pose annotations, which are collected by time-consuming delicate motion capture systems [30]. Instead, we propose to fully leverage the beneficial cues of human poses, by assessing the physical consistency between observed poses and future trajectories for physical plausibility-aware HTP.", + "bbox": [ + 509, + 656, + 908, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For this physical plausibility-aware HTP, we present Locomotion Embodiment, a framework that incorporates the laws of physics into an HTP network. This is achieved by physically simulating the locomotion conditioned by observed poses and future trajectories in a physics simulator [31]; refer to Fig. 1 to see the effect of our framework. As a typical", + "bbox": [ + 511, + 810, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12324", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "example, one of the human poses included in a future possible trajectory $\\tau_f^1$ (indicated by $j_{t}$ in Fig. 1) is a little unstably tilted, leading to a lower plausibility score of $\\tau_f^1 = 0.3$ . By training the HTP network in a data-driven manner with following the laws of physics, we benefit from the advantages of both the data-driven and physics-based approaches.", + "bbox": [ + 88, + 90, + 480, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By evaluating and back-propagating the plausibility of locomotion generated in a physics simulator, we can train the physical plausibility-aware HTP network. However, locomotion generation using physics simulators requires various motion parameters, most of which are unavailable or noisy (e.g., joint angular velocities estimated from joint positions provided by a 3D pose estimator) in image-based HTP scenarios. In addition, rigorous physics simulators are indifferent to the physical plausibility, this paper proposes the Locomotion Value (LocoVal) function implemented with a differentiable network to serve as a surrogate for a physics simulator for the plausibility evaluation (Sec. 4.1). Given the plausibility evaluation provided by the LocoVal function, our EmLoco loss penalizes implausible trajectories (Sec. 4.2).", + "bbox": [ + 91, + 185, + 483, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This EmLoco loss is used in conjunction with the MSE loss for supervised trajectory prediction learning. While the MSE loss behaves as a data term, our EmLoco loss can be a regularization term that incorporates prior knowledge about physics. This is crucial for maintaining the diversity of the predicted trajectories. Although human trajectories are inherently stochastic due to many factors, such as dynamic environments, training with the MSE loss alone forces all predicted trajectories to be aligned with a single ground truth. That is, minimizing the MSE essentially decreases the diversity of the predicted trajectories.", + "bbox": [ + 89, + 412, + 483, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, for stochastically predicting multiple future trajectories, our EmLoco loss allows us to supervise all predicted trajectories to improve their physical plausibility while preserving diversity. In addition, our LocoVal function can also serve as the LocoVal filter that evaluates and filters the candidate trajectories at inference (Sec. 4.3).", + "bbox": [ + 89, + 579, + 483, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework consists of three key components:", + "bbox": [ + 109, + 671, + 439, + 686 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. LocoVal function as a surrogate for an indifferentiable physics simulator in HTP network training.", + "2. EmLoco loss for back-propagating the plausibility of predicted trajectories. This loss is also useful for deterministic predictors but more beneficial for stochastic predictors because of its applicability to all possible predicted trajectories while preserving their stochastic diversity.", + "3. LocoVal filter for filtering out implausible trajectories at inference. This filter can be applied to any pre-trained HTP network in a plug-and-play manner." + ], + "bbox": [ + 89, + 686, + 483, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work is the first attempt to use simulated humanoid control in a physics simulator for HTP. Since the physics simulator is used only during training, an additional cost (i.e., the cost of the LocoVal function) is negligible at inference.", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Human Trajectory Prediction (HTP)", + "text_level": 1, + "bbox": [ + 511, + 114, + 830, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "HTP methods can be divided into deterministic [2, 23, 37, 47, 63] and stochastic [18, 29, 49, 50, 64] methods. Deterministic HTP methods [2, 20, 23, 37, 46, 63] predict the most likely trajectory. For example, Social-LSTM uses the Long Short Term Memory [2] to represent the consistency between observed and future trajectories. On the other hand, stochastic HTP methods [12, 49, 50, 64] produce multiple trajectories or probability distributions. To model the inherent stochasticity in HTP, these methods employ probabilistic frameworks such as Generative Adversarial Networks [18, 48] and Diffusion models [15, 33], as well as Normalizing Flows [29, 51] to draw multiple trajectories at inference. Our method is model-agnostic and applicable to both deterministic and stochastic methods.", + "bbox": [ + 511, + 137, + 906, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. HTP with Pose Information", + "text_level": 1, + "bbox": [ + 511, + 359, + 761, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior studies demonstrate that human poses benefit HTP [26, 42, 47, 50, 65]. Yagi et al. [65] and Liang et al. [26] incorporate human poses estimated by CNNs into HTP. Transformer-based methods such as Social-Transmotion [47] and Human Scene Transformer (HST) [50] further improve HTP conditioned by human poses. In particular, HST [50] reports the effectiveness of human pose conditioning for momentary observation scenarios.", + "bbox": [ + 511, + 381, + 906, + 502 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, since these methods include poses simply as an additional input to the HTP network, predicted future trajectories can be physically inconsistent with past poses. Inspired by this observation, our Locomotion Embodiment framework represents the physical plausibility of predicted trajectories as pose-trajectory consistent locomotion.", + "bbox": [ + 511, + 503, + 905, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that we aim to improve HTP by fully leveraging human poses, distinct from Human Motion Prediction which prioritizes predicting future poses [1, 4, 17, 28, 32, 69].", + "bbox": [ + 511, + 595, + 905, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Locomotion Generation", + "text_level": 1, + "bbox": [ + 511, + 651, + 733, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Locomotion generation is a task to generate walking motions for humanoid robots [16, 43, 44] and virtual characters [22, 38, 71]. Locomotion can be controlled on various conditions such as destinations [7, 19, 56], trajectories [21, 45], and poses [27, 61]. Generative models such as Diffusion models [5, 21, 53, 58, 67] can generate such conditioned locomotion, but they are prone to generate physically implausible locomotion (e.g., skating motions). This is because the laws of physics are not explicitly formulated in generative models.", + "bbox": [ + 511, + 672, + 908, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, physics-based locomotion generation [27, 40, 45, 60-62] follows the laws of physics in simulators [6, 31]. However, a generated locomotion would fail to follow a given trajectory or fall due to kinematic limitations if a humanoid is controlled to walk along an implausible", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12325", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trajectory. Therefore, a physics-based locomotion generator can evaluate whether the conditioned trajectory is plausible. Our Locomotion Embodiment framework utilizes this ability to evaluate the plausibility of future trajectories.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The aforementioned physics-based locomotion generator, e.g., PACER [27, 40, 45, 61], can be trained with reinforcement learning in a physics simulator. A straightforward way to use this locomotion generator for training an HTP network is to directly incorporate the locomotion generator into the HTP network. However, this direct integration has two problems: (1) As mentioned in Sec. 1, the physics simulator is indifferent, preventing training the HTP network by back-propagation. (2) The locomotion generator using the physics simulator requires physical humanoid states, e.g., joint angular velocities, which are unavailable or noisy in image-based HTP scenarios. To address these two problems, our framework uses the physics simulator and the locomotion generator only before training the HTP network and replaces them with the LocoVal function in training the HTP network.", + "bbox": [ + 89, + 154, + 485, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 415, + 228, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HTP. An HTP network, $\\mathcal{H}$ , takes the past $T_{\\mathrm{p}}$ frames as input, including the past trajectory and 3D poses of a target person, and predicts the trajectory of the target person in the future $T_{\\mathrm{f}}$ frames. The past trajectory $\\pmb{\\tau}_{\\mathrm{p}} \\in \\mathbb{R}^{T_{\\mathrm{p}} \\times 2}$ and the future target trajectory $\\pmb{\\tau}_{\\mathrm{f}} \\in \\mathbb{R}^{T_{\\mathrm{f}} \\times 2}$ are on the ground $xy$ plane. The 3D pose involves $J$ joint positions $\\pmb{j}_t \\in \\mathbb{R}^{J \\times 3}$ at the past time $t$ , where $(-T_{\\mathrm{p}} + 1) \\leq t \\leq 0$ .", + "bbox": [ + 89, + 441, + 483, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The predicted future trajectory $\\hat{\\tau}_{\\mathrm{f}}$ is expected to be as close as possible to its ground-truth trajectory $\\tau_{\\mathrm{f}}$ .", + "bbox": [ + 89, + 550, + 483, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Locomotion Generation. In our framework, humanoid locomotion is generated to follow any given future trajectory $\\pmb{\\tau}_{\\mathrm{s}} \\in \\mathbb{R}^{T_{\\mathrm{f}} \\times 2}$ in a physics simulator [31]. As such a locomotion generator $\\mathcal{G}$ , our method employs PACER [45], which is implemented with the policy network [36] trained via reinforcement learning [52]. $\\mathcal{G}$ takes $\\pmb{\\tau}_{\\mathrm{s}}$ and the initial humanoid state $h_0$ to control the humanoid joints. $h_0$ involves the positions, rotations, velocities, and angular velocities of all joints. In PACER [45], $\\mathcal{G}$ is trained by maximizing the discounted cumulative reward $\\Omega$ consisting of a trajectory following reward, a motion style reward [40], and an energy penalty [11]. Refer to PACER [45] for details.", + "bbox": [ + 89, + 611, + 483, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In PACER, the humanoid movement is constrained by the limits of joint angles and torques, preventing it from producing kinematically impossible locomotion. Due to this limitation, when conditioned on a physically implausible trajectory, the humanoid fails to follow it, resulting in a lower reward. Therefore, we define the discounted cumulative reward $\\Omega$ [52] as the plausibility of $\\tau_{\\mathrm{s}}$ .", + "bbox": [ + 89, + 795, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b132474f5f9792dfc6aa265d95795383915f0cf2e15abb00dd1b1eef279e64e7.jpg", + "image_caption": [ + "Figure 2. The overview of the proposed framework. $①$ and $⑤$ mean the weights are trained and fixed, respectively." + ], + "image_footnote": [], + "bbox": [ + 521, + 93, + 901, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Physical Plausibility-aware HTP", + "text_level": 1, + "bbox": [ + 511, + 404, + 808, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. Figure 2 shows the overview of our framework consisting of two training stages and an inference phase. In (a) the first training stage, given the physics-based locomotion generator $\\mathcal{G}$ trained as mentioned in Sec. 3, a current humanoid state is fed into it to obtain the reward as the physical plausibility. With this physical plausibility as the ground truth, the LocoVal function $\\nu$ is trained. In (b) the second stage, observed pose-related cues in training data are fed into both $\\nu$ and $\\mathcal{H}$ . $\\nu$ also takes the future trajectories predicted by $\\mathcal{H}$ to compute the EmLoco losses. The EmLoco loss supports both the MSE and minMSE [18, 50, 63] losses used for deterministic and stochastic predictions, respectively, to train $\\mathcal{H}$ . At (c) inference, candidate trajectories predicted by $\\mathcal{H}$ are fed into $\\nu$ to filter out implausible trajectories.", + "bbox": [ + 511, + 429, + 908, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. LocoVal Function", + "text_level": 1, + "bbox": [ + 511, + 651, + 687, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The detail of training the LocoVal function $\\mathcal{V}$ is illustrated in Fig. 3. As shown in the top half of Fig. 3, the discounted cumulative reward $\\Omega$ [52] is used as the plausibility of Embodied Locomotion that is generated through $\\mathcal{G}$ using a physics simulator. $\\mathcal{V}$ is trained to estimate the reward $\\hat{\\Omega}$ close to $\\Omega$ from only available cues in image-based HTP scenarios because $\\mathcal{V}$ serves as a surrogate for $\\mathcal{G}$ at inference.", + "bbox": [ + 511, + 674, + 908, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As such available cues, our method uses the future trajectory $\\tau_{\\mathrm{s}}$ , the initial pose $j_0$ , and the initial velocity of the root joint $\\boldsymbol{v}_{\\mathrm{root},0} \\in \\mathbb{R}^2$ , where $j_0$ and $\\boldsymbol{v}_{\\mathrm{root},0}$ are in a subset $h_0'$ extracted from $h_0$ . In training $\\mathcal{V}$ , these cues are available from the physics simulator. At inference, $\\tau_{\\mathrm{s}}$ is replaced by $\\tau_{\\mathrm{f}}$ , $j_0$ can be obtained by a 3D pose estimator [14, 50], and $\\boldsymbol{v}_{\\mathrm{root},0}$ can be obtained from $\\tau_p$ . $\\mathcal{V}$ is implemented as a differentiable MLP, taking these cues as input.", + "bbox": [ + 511, + 780, + 910, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12326", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2f7a4a352c71f865c61e391bc5a80910b1147b2718c4c3e5b1c5717ef2fb9290.jpg", + "image_caption": [ + "Figure 3. The overview of training our LocoVal function, $\\nu$ ." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 483, + 205 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c55a620292924e8da6a607c2aa66c53643b22e615d323f6ea73ea44b4ed6add2.jpg", + "image_caption": [ + "Figure 4. The overview of training the HTP network, $\\mathcal{H}$" + ], + "image_footnote": [], + "bbox": [ + 94, + 238, + 480, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Process. Unlike PACER [45], which trains $\\mathcal{G}$ to obtain high rewards even for implausible trajectories, e.g., dangerously sharp turns, such trajectories should result in low rewards in our framework. This is because humans do not typically walk along such implausible trajectories in daily life. For this reason, we train $\\mathcal{G}$ with only plausible pairs of $\\tau_{s}$ and $h_0$ to give high rewards only to plausible pose-trajectory pairs. On the other hand, since $\\mathcal{V}$ must discriminate between the plausible and implausible pose-trajectory pairs, it is trained with plausible pairs and randomly paired diverse implausible pose-trajectory pairs.", + "bbox": [ + 88, + 424, + 482, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{V}$ is trained by minimizing the following loss $\\mathcal{L}_{\\mathcal{V}}$ so that it can estimate the plausibility of the Embodied Locomotion $\\varOmega=\\mathcal{G}(\\pmb{\\tau}_{\\mathrm{s}},\\pmb{h}_{0})$ obtained by rolling out $\\mathcal{G}$ :", + "bbox": [ + 89, + 590, + 483, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {V}} = \\operatorname {M S E} \\left(\\mathcal {V} \\left(\\boldsymbol {\\tau} _ {\\mathrm {s}}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right), \\mathcal {G} \\left(\\boldsymbol {\\tau} _ {\\mathrm {s}}, \\boldsymbol {h} _ {0}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 645, + 482, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where MSE denotes the Mean Squared Error function.", + "bbox": [ + 89, + 671, + 449, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. HTP Network Training with EmLoco Loss", + "text_level": 1, + "bbox": [ + 89, + 694, + 452, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 4 illustrates the detail of training the HTP network $\\mathcal{H}$ . Our HTP network takes the past information including $\\tau_{\\mathrm{p}}$ and 3D poses $j_{\\mathrm{p}} = \\{j_{-T_{\\mathrm{p}} + 1},\\dots ,j_0\\}$ as input to predict the future trajectory $\\hat{\\tau}_{\\mathrm{f}}$ of a target person. Following traditional HTP training methods [2, 47], $\\hat{\\tau}_{\\mathrm{f}}$ is compared with its ground truth $\\tau_{\\mathrm{f}}$ to back-propagate the MSE loss $\\mathcal{L}_{\\mathrm{T}} = \\mathrm{MSE}(\\hat{\\tau}_{\\mathrm{f}},\\tau_{\\mathrm{f}})$ .", + "bbox": [ + 89, + 715, + 483, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "EmLoco loss. The EmLoco loss $\\mathcal{L}_{\\mathrm{E}}$ evaluates predicted trajectories by using $\\mathcal{V}$ , which has prior knowledge about plausibility. Note again that a physics simulator is not required in this second stage or the inference phase because $\\mathcal{V}$ serves as a surrogate for the plausibility evaluation process.", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the prediction of $\\mathcal{H}$ , the predicted trajectory $\\hat{\\tau}_{\\mathrm{f}}$ and the subset of $h_0$ (i.e., $h_0^{\\prime}$ ) observed in the current frame are fed into $\\nu$ to compute $\\mathcal{L}_{\\mathrm{E}}$ as follows:", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E}} = - \\mathcal {V} \\left(\\hat {\\tau} _ {\\mathrm {f}}, h _ {0} ^ {\\prime}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 143, + 906, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{E}}$ encourages $\\mathcal{H}$ to predict $\\hat{\\tau}_{\\mathrm{f}}$ that increases the plausibility.", + "bbox": [ + 513, + 167, + 906, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the total training objective, $\\mathcal{L}_{\\mathrm{E}}$ is combined with the MSE loss $\\mathcal{L}_{\\mathrm{T}}$ with a hyper-parameter $\\alpha$ as follows:", + "bbox": [ + 511, + 183, + 905, + 214 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {T}} + \\alpha \\mathcal {L} _ {\\mathrm {E}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 222, + 906, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-head HTP network. Our method with the EmLoco loss can be extended to a stochastic HTP network using multiple heads. As also mentioned in Sec. 1, the MSE loss (Eq. (1)) is inappropriate for multiple stochastic predictions because it forces all of them to converge to a unique ground truth. To mitigate this problem, following prior work [18, 50, 63], the multiple heads in our method are trained by replacing $\\mathcal{L}_{\\mathrm{T}}$ in Eq.(3) with the following objective function:", + "bbox": [ + 511, + 255, + 906, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T}} = \\min _ {k} \\left(\\operatorname {M S E} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {\\tau} _ {\\mathrm {f}}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 388, + 906, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\tau}_{\\mathrm{f}}^{k}$ denotes the prediction of $k$ -th head. In Eq. (4), the minMSE loss supervises only one prediction closest to its ground truth $\\tau_{\\mathrm{f}}$ among multiple predictions.", + "bbox": [ + 511, + 417, + 903, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike the MSE loss, this minMSE loss allows us to maintain diverse predictions. However, the minMSE loss is inefficient. What is even worse is that several heads may be under-fitted if other heads are trained in early training iterations so that these heads are always selected as those minimizing the MSE in Eq. (4) in later iterations.", + "bbox": [ + 511, + 463, + 905, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the other hand, our EmLoco loss does not require the unique ground truth, i.e., $\\tau_{\\mathrm{f}}$ in Eq. (4). As a regularization term, the EmLoco loss trains all the heads to predict plausible trajectories. This simultaneous optimization allows us to avoid the under-fitting problem mentioned above. In our method, this simultaneous optimization is implemented by averaging the EmLoco losses using $\\mathcal{V}$ for all predicted trajectories, and this average is used as $\\mathcal{L}_{\\mathrm{E}}$ in Eq.(3).", + "bbox": [ + 511, + 554, + 906, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. LocoVal Filter at Inference", + "text_level": 1, + "bbox": [ + 511, + 681, + 756, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The trained LocoVal function can also be applied to trajectory filtering, called the LocoVal filter, at inference using multi-head HTP networks. Similar to the usage of $\\mathcal{V}$ for training multiple heads simultaneously at training, it computes the plausibilities of all predicted trajectories at inference. Each of the predicted trajectories with a plausibility lower than a threshold $\\lambda$ is filtered out as follows:", + "bbox": [ + 511, + 704, + 906, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} \\left\\{\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k} \\mid \\mathcal {V} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right) \\geq \\lambda \\right\\} & \\text {i f} \\exists k, \\mathcal {V} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right) \\geq \\lambda , \\\\ \\operatorname {a r g m a x} _ {\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}} \\mathcal {V} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 815, + 906, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the LocoVal filter can be used with any stochastic HTP networks in a plug-and-play manner.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12327", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 225, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec. 5.2.1, we benchmark our HTP network trained with the EmLoco loss under the standard HTP setting. Section 5.2.2 presents results for multi-head HTP networks, and Sec. 5.2.3 examines the performance with momentary observations. Then, we report ablation results in Sec. 5.2.4. In Sec. 5.3.1 and Sec. 5.3.2, we evaluate the effectiveness of the LocoVal filter in various scenarios.", + "bbox": [ + 89, + 117, + 485, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 244, + 299, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. As trajectory datasets, the Joint Track Auto (JTA) dataset [9], the JackRabbit Dataset and Benchmark (JRDB) [34], and the ETH/UCY dataset [24, 39] are used. For the JRDB dataset, a subset provided by the SocialTrans [47] was used. The ETH/UCY dataset is used only in Sec. 5.3.2 to evaluate the performance of the LocoVal filter on pre-trained models.", + "bbox": [ + 89, + 270, + 483, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While JTA also provides 3D human poses synchronized with their trajectories, no 3D pose is annotated in JRDB. Instead, in JRDB, our experiments use 3D poses estimated from images and 2D pose annotations by using [14, 50]. While the estimated 3D poses are provided by HST [50], misestimated poses are filtered out in our experiments; see the Supp. for the details. All the 3D poses mentioned above are converted to the SMPL-style poses [25] for a fair comparison with the baseline.", + "bbox": [ + 89, + 378, + 485, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the 3D poses described above are used in the second stage to train the HTP network $\\mathcal{H}$ , more accurate 3D poses are expected to train $\\mathcal{G}$ and $\\mathcal{V}$ to learn the physical plausibility of 3D poses. Such accurate 3D poses are provided by the Archive of Motion Capture as Surface Shapes (AMASS) [30]. While accurate 3D poses are available and can be used to compute the trajectories in motion capture datasets such as AMASS, motion capture data do not scale compared with trajectory datasets, as described in Sec. 1.", + "bbox": [ + 89, + 518, + 485, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To avoid this problem, 3D poses in AMASS are combined with trajectories in the trajectory datasets, i.e., JTA and JRDB, as follows. Plausible and implausible pose-trajectory pairs are required to train $\\mathcal{G}$ and $\\mathcal{V}$ , as described in Sec. 4.1. The plausible pairs are prepared by spatially aligning the 3D pose and the trajectories, which are randomly sampled from AMASS and the trajectory datasets, respectively. This spatial alignment consists of the orientation and velocity alignments. Given the heading vector of each pedestrian in these two datasets, the heading orientations are aligned between the motion capture pose from AMASS and the trajectory data. For the velocity alignment, on the other hand, the trajectory data is expanded or contracted to make the vector norms equal between the motion capture pose and the trajectory data. Refer to the Supp. for detailed training conditions on the aforementioned datasets.", + "bbox": [ + 89, + 657, + 485, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HTP networks. Since multiple people are observed in each frame of the aforementioned datasets, all baseline HTP networks of our method take not only the past trajectory of a target person but also those of other people as input. This allows us to predict the target trajectory affected by the dynamic environment. While our method is introduced to take only the target person as input, all people in the scene are fed into $\\mathcal{H}$ following the previous methods. In the evaluations with the JTA and JRDB datasets, Social-Transmotion (Social-Trans), is used as our baseline without any architectural changes. We extend this baseline of the trajectory predictor $\\mathcal{H}$ by plugging the EmLoco loss into the training loss. Following the implementation of Social-Trans, past trajectories, 2D and 3D human poses, and 2D and 3D bounding boxes are fed into $\\mathcal{H}$ in the JTA dataset [9], while the past trajectories, 2D bounding boxes, and 3D human poses are used in the JRDB dataset. Note that the 3D human poses are estimated for the JRDB dataset [34] as mentioned earlier. For the stochastic version of our method, we modify Social-Trans [47] by parallelizing the prediction heads with different initial weights. For comparison, all HTP networks predict the future $T_{\\mathrm{f}} = 12$ frames at 2.5 fps on the JTA and JRDB datasets. The number of past frames differs depending on the experiment; e.g., 9 frames in Sec. 5.2.1 and Sec. 5.2.2, and 2 frames as momentary observations [55] in Sec. 5.2.3.", + "bbox": [ + 511, + 90, + 906, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In experiments on the ETH/UCY dataset [24, 39], where a bird's-eye view hinders the use of 3D poses, only past trajectories are fed into $\\mathcal{H}$ . Here, we employ EqMotion [63], a state-of-the-art stochastic HTP method that only uses past trajectories, as a pre-trained HTP network. Each past trajectory consists of $T_{\\mathrm{p}} = 8$ frames, while each future trajectory remains at $T_{\\mathrm{f}} = 12$ frames at 2.5 fps.", + "bbox": [ + 511, + 469, + 908, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. Two major metrics, the Average Displacement Error (ADE) and the Final Displacement Error (FDE), both measured in meters, evaluate the HTP results. Consistent with the other HTP literature [29, 49, 50, 63], in evaluating the stochastic HTP networks, we report minADE and minFDE among multiple trajectories. Also, following MATRIX [64], we report $\\chi^2$ distances for the four physics primitives of each predicted trajectory, namely velocities, accelerations, angular velocities, and angular accelerations, representing the plausibility of the predicted trajectories.", + "bbox": [ + 511, + 595, + 910, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. HTP Training with the EmLoco Loss", + "text_level": 1, + "bbox": [ + 511, + 758, + 834, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we evaluate our HTP network trained with the proposed EmLoco loss. Note that the LocoVal filter is not applied in these experiments. $\\alpha$ in Eq.(3) is set to 100.", + "bbox": [ + 511, + 781, + 908, + 827 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2.1. Comparison with Existing HTP Methods", + "text_level": 1, + "bbox": [ + 511, + 835, + 844, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the standard protocol [2, 12, 18, 47, 63] using 9 past frames, where past trajectories are sufficiently long, we compare the accuracy of deterministic predictions by our", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12328", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9eef455644a15eeb783a45b80bd6cf5a9d23e8064458295ad8fa203943e71403.jpg", + "table_caption": [ + "Table 1. HTP results under the standard setting. Red and blue indicate the best and second-best. 'Social-Trans' and 'Ours' utilize pose input. For methods other than 'Social-Trans' and 'Ours', we report the results from the Social-Trans literature [47]." + ], + "table_footnote": [], + "table_body": "
MethodJTA [9]JRDB [34]
ADEFDEADEFDE
Social-GAN-det [18]1.663.760.500.99
Transformer [13]1.563.540.561.10
Vanilla-LSTM [2]1.443.250.420.83
Occupancy-LSTM [2]1.413.150.430.85
Directional-LSTM [23]1.373.060.450.87
Dir-social-LSTM [23]1.232.590.480.95
Social-LSTM [2]1.212.540.470.95
Autobots [12]1.202.700.390.80
Trajectory++ [49]1.182.530.400.78
EqMotion [63]1.132.390.400.77
Social-Trans [47]1.112.260.400.76
Ours0.971.910.370.72
", + "bbox": [ + 91, + 151, + 480, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HTP network with other HTP networks [2, 12, 18, 23, 47, 49, 59, 63]. Table 1 shows that our method outperforms all other methods including the state-of-the-art baseline, Social-Trans [47], on both the JTA [9] and the JRDB [34] datasets. This demonstrates that the EmLoco loss effectively provides the prior knowledge of the trajectory-pose consistency during training. Notably, the performance improvements on the JRDB dataset [34] with imperfect 3D poses suggest the applicability of our EmLoco loss to real-world applications.", + "bbox": [ + 88, + 388, + 486, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.2. Multi-head HTP Networks", + "text_level": 1, + "bbox": [ + 89, + 532, + 326, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the EmLoco loss on a multi-head HTP network, 5 and 20 prediction heads are implemented on Social-Trans [47]. As described in Sec. 4.2, our proposed method is expected to improve the plausibility of predicted trajectories closer to human locomotion. From this perspective, we evaluate the plausibility of the predicted trajectories by the $\\chi^2$ distance employed in MATRIX [64].", + "bbox": [ + 89, + 551, + 485, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 presents the comparisons with the baseline. Our method achieves the best values across almost all metrics on both the JTA [9] and JRDB [34] datasets. Particularly, the improvements in ADE and FDE reveal that all heads are trained properly because these metrics are evaluated with the prediction of all heads. Also, improvements in $\\chi^2$ distances, the plausibility regarding physics primitives, demonstrate that $\\mathcal{H}$ trained with our EmLoco loss predicts trajectories similar to actual human locomotion than the baseline [47].", + "bbox": [ + 89, + 657, + 483, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5 shows qualitative results comparing stochastic HTP with 5 heads for both Social-Trans [47] and our method. While the Social-Trans trained solely with minMSE can predict diverse trajectories, many of them are implausible, including sharp turns. In contrast, the predictions of our HTP network maintain diversity across all heads while remaining plausible, resulting in trajectories closer to the ground truth.", + "bbox": [ + 89, + 795, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e545968f64e3b6e38a64bf326a3604d1be02541b6ead9ec3a8aa856f56dad392.jpg", + "image_caption": [ + "Figure 5. Comparison of the stochastic HTP between Social-Trans [47] and ours with 5 heads on the JTA dataset [9]. Human poses are shown at a doubled scale for visualization purposes." + ], + "image_footnote": [], + "bbox": [ + 513, + 95, + 906, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b42ac27e3786c76a1711e01516b66deb639db7991942c5be018ce7758c758df7.jpg", + "image_caption": [ + "Figure 6. Comparison of the prediction results between the baseline Social-Trans [47] and our method with momentary observations on the JTA dataset [9]. The observed pose are displayed at a doubled scale for visualization purposes." + ], + "image_footnote": [], + "bbox": [ + 511, + 315, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.3. HTP with Momentary Observations", + "text_level": 1, + "bbox": [ + 511, + 571, + 812, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As noted in Sec. 1, existing HTP methods struggle with momentary observations. On the contrary, because the proposed EmLoco loss encourages physical plausibility-aware HTP, $\\mathcal{H}$ trained with the EmLoco loss is expected to be robust against this setting. From this point of view, we experiment with 2 past frames following the prior work [55].", + "bbox": [ + 511, + 594, + 908, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are shown in Table 3. Consistent with Tables 1 and 2, the EmLoco loss allows our method to outperform the baseline [47] across almost all categories. In addition, Fig. 6 presents qualitative results in two cases: when a pedestrian is (a) going straight and (b) turning. In both cases, (i) Social-Trans [47] with trajectory-only input results in significant prediction errors, and (ii) Social-Trans with pose input still predicts implausible trajectories, especially during sharp turns. In contrast, (iii) our method provides plausible predictions that closely follow the ground-truth trajectory and align with the observed pose. This demonstrates that the proposed method can effectively leverage human poses and achieve plausibility-aware HTP, thanks to the physics-based LocoVal function training.", + "bbox": [ + 511, + 688, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12329", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dc7e31e98663d28c3d2cab5f599d91877e90dd787405857adcda4db3a9100214.jpg", + "table_caption": [ + "Table 2. Comparison of our HTP network using the EmLoco loss with baseline [47] in the stochastic HTP under 9 frames of observations. Red indicates the best result for each category. The results are presented as left/right values, where the left denotes evaluations with 5 heads and the right denotes 20 heads. 'Vel,' 'Acc,' 'Ang. Vel,' and 'Ang. Acc.' represent the $\\chi^2$ distances between the distributions of the predicted and ground-truth trajectories for velocity, acceleration, angular velocity, and angular acceleration, respectively." + ], + "table_footnote": [], + "table_body": "
MethodADEFDEminADEminFDE\\( \\chi^2 \\) distances
Vel.Acc.Ang. Vel.Ang. Acc.
JTASocial-Trans1.86/2.143.51/4.260.88/0.711.04/0.540.134/0.1690.009/0.0090.011/0.0070.026/0.008
Ours1.68/1.803.34/3.560.79/0.661.02/0.540.100/0.0870.002/0.0030.009/0.0070.016/0.011
JRDBSocial-Trans0.61/0.711.19/1.490.25/0.180.44/0.260.025/0.0770.034/0.0130.021/0.0190.029/0.137
Ours0.56/0.701.16/1.480.23/0.170.44/0.250.024/0.0670.011/0.0040.021/0.0200.019/0.036
", + "bbox": [ + 94, + 151, + 911, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5a934815e4301e01a05f32d6337ea0d9b6f371a6947d6781ef7d6dc74d8f0489.jpg", + "table_caption": [ + "Table 3. HTP results with momentary observations where only 2 frames of past trajectories are observed. Red and blue indicate the best and second-best for each category. 'Traj.' indicates the input is only past trajectories. 'Det.' denotes the deterministic prediction.", + "(a) HTP with momentary observations on the JTA dataset." + ], + "table_footnote": [], + "table_body": "
MethodADEFDEminADEminFDE
Det.Social-Trans (Traj.)1.573.26--
Social-Trans1.362.80--
Ours1.332.72--
5headsSocial-Trans (Traj.)2.254.620.911.51
Social-Trans2.234.550.921.53
Ours1.943.960.871.30
20headsSocial-Trans (Traj.)2.585.280.710.76
Social-Trans2.465.070.690.73
Ours2.124.470.650.68
", + "bbox": [ + 98, + 349, + 473, + 484 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c3c881a14397c9c89b2c4242fff6e2e2b92d05b50252675ddfda2231555748b1.jpg", + "table_caption": [ + "(b) HTP with momentary observations on the JRDB dataset." + ], + "table_footnote": [], + "table_body": "
MethodADEFDEminADEminFDE
Det.Social-Trans (Traj.)0.500.94--
Social-Trans0.500.94--
Ours0.440.85--
5headsSocial-Trans (Traj.)0.571.190.250.47
Social-Trans0.581.200.250.47
Ours0.551.120.240.48
20headsSocial-Trans (Traj.)0.681.460.180.30
Social-Trans0.681.460.170.28
Ours0.681.450.180.28
", + "bbox": [ + 98, + 510, + 473, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 686, + 253, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Supervision by the EmLoco loss. To provide deeper insights into the impact of the EmLoco loss $\\mathcal{L}_{\\mathrm{E}}$ on HTP network training, we train $\\mathcal{H}$ only with $\\mathcal{L}_{\\mathrm{E}}$ i.e., omitting the loss $\\mathcal{L}_{\\mathrm{T}}$ with the ground-truth trajectories. The results are presented in Table 4(B). Importantly, even the proposed method without $\\mathcal{L}_{\\mathrm{T}}$ can improve performance from a randomly initialized network (Table 4(A)). This demonstrates that the supervision provided by $\\mathcal{L}_{\\mathrm{E}}$ contributes to the improvements in HTP performance, while it is designed to be used in conjunction with $\\mathcal{L}_{\\mathrm{T}}$ . We emphasize that the combination of $\\mathcal{L}_{\\mathrm{T}}$ and $\\mathcal{L}_{\\mathrm{E}}$ (Table 4(G)) improves HTP performance compared to the baseline [47] (Table 4(C)).", + "bbox": [ + 88, + 719, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/793576e0bc0a71b2bb0b1cacf163ae21f69c0fc5a834902a4d4f07116403c310.jpg", + "table_caption": [ + "Table 4. Ablation studies on JTA [9] and JRDB [34] datasets. Red and blue indicate the best and second-best results for each category. 'initialized' represents the initialized HTP network for reference." + ], + "table_footnote": [], + "table_body": "
MethodJTAJRDB
ADEFDEADEFDE
(A) Initialized5.078.921.311.87
(B) w/o GT loss LT3.528.200.921.55
(C) w/o EmLoco loss LE [47]1.112.260.400.76
(D) w/o Implausible trajectories1.012.050.380.74
(E) w/o Pose input j01.072.240.390.75
(F) w/o Root velocity vroot,00.992.030.380.74
(G) Ours0.971.910.370.72
", + "bbox": [ + 517, + 327, + 906, + 454 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/92e66b48e8242bff4b25ce51d1d71a150f877540b85edef83039878c6867b17b.jpg", + "table_caption": [ + "Table 5. Results of the LocoVal filter for stochastic HTP on the JTA [9] dataset with 9 and 2-frame observations. The results are presented as left/right values, where the left denotes evaluations with 5 heads and the right denotes 20 heads. Red indicates the best result. 'Rejected' refers to the performance of rejected trajectories." + ], + "table_footnote": [], + "table_body": "
Method9 frames2 frames
ADEFDEADEFDE
Social-Trans(w/o LE)w/o Filtering1.86/2.143.51/4.262.23/2.464.55/5.07
w/ Filtering1.81/1.973.49/3.932.06/2.214.27/4.56
Rejected2.37/3.413.75/6.653.83/4.327.16/8.87
Ours(w/ LE)w/o Filtering1.68/1.803.34/3.561.94/2.123.96/4.47
w/ Filtering1.65/1.763.31/3.521.90/2.083.88/4.41
Rejected2.32/2.403.86/4.193.22/3.256.18/6.41
", + "bbox": [ + 517, + 575, + 903, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inputs into the LocoVal function. We ablate the inputs of $\\mathcal{V}$ as follows: (1) implausible pose-trajectory pairs used in training $\\mathcal{V}$ (Table 4(D)), (2) the human pose $j_0$ at both training and inference (Table 4(E)), (3) and the root velocity $\\pmb{v}_{\\mathrm{root},0}$ at both training and inference (Table 4(F)). These results indicate that not only plausible but also implausible pose-trajectory pairs are required for training $\\mathcal{V}$ to serve as a surrogate for the locomotion generator. Also, both $j_0$ and $\\pmb{v}_{\\mathrm{root},0}$ contribute to performance improvements. Notably, the HTP performance decreases when $j_0$ is not used, demonstrating that the pose is important to estimate the plausibility of predicted trajectories.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12330", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b6b483f9fa0b93e97d7bb68591819433605ae8c45202d8c40b19e75a3175bb62.jpg", + "table_caption": [ + "Table 6. Results of zero-shot filtering by the LocoVal filter on the predictions of a pre-trained trajectory predictor [63] on the ETH/UCY dataset [24, 39]. 'Mean' represents the average performance across the 5 subsets." + ], + "table_footnote": [], + "table_body": "
MethodETHHOTELUNIVZARA1ZARA2Mean
ADEFDEADEFDEADEFDEADEFDEADEFDEADEFDE
Pretrained EqMotion [63]2.184.630.641.311.302.810.821.840.651.471.122.41
Ours (w/ LocoVal filter)1.412.880.611.260.932.040.801.800.641.450.881.89
Rejected trajectories8.8919.722.695.534.339.181.703.672.214.723.968.56
", + "bbox": [ + 104, + 127, + 888, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. LocoVal Filter at Inference", + "text_level": 1, + "bbox": [ + 89, + 232, + 334, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.1. LocoVal Filter on the JTA Dataset", + "text_level": 1, + "bbox": [ + 89, + 255, + 375, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since our LocoVal filter can be applied in a plug-and-play manner to any HTP network, we apply it to Social-Trans [47] and our method (i.e., Social-Trans trained with $\\mathcal{L}_{\\mathrm{E}}$ ).", + "bbox": [ + 89, + 276, + 482, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5 presents the filtering results on the JTA dataset [9] with $\\lambda$ set to 0.7 in Eq.(5). We observe that the LocoVal filter enhances performance across all metrics in both Social-Trans and our method. Notably, although $\\mathcal{L}_{\\mathrm{E}}$ using $\\nu$ is used in training $\\mathcal{H}$ in our method, the LocoVal filter using $\\nu$ can further enhance the performance of our method through filtering at inference. Table 5 also shows the evaluation scores of the rejected trajectories. These scores exhibit substantially poorer performance than the filtered results, indicating that the LocoVal filter rejects implausible, inaccurate trajectories.", + "bbox": [ + 88, + 321, + 485, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.2. Zero-shot Filtering on the ETH/UCY Dataset", + "text_level": 1, + "bbox": [ + 89, + 484, + 452, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following the aforementioned experiments on the JTA dataset [9], we further evaluate the LocoVal filter on the ETH/UCY dataset [24, 39] using another baseline, EqMotion [63]. Note that this dataset is not used for training $\\nu$ , and due to the bird's-eye-view of the ETH/UCY dataset, we experimented with $\\nu$ that does not use human poses as input.", + "bbox": [ + 89, + 506, + 483, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6 presents the results of applying the LocoVal filter under this zero-shot setting, with $\\lambda$ set to 0.8. Although the dataset and human poses are not used for training $\\mathcal{V}$ , we observe improvements in both ADE and FDE with the application of the LocoVal filter across all 5 subsets. While the importance of human poses can be observed in the comparison between Table 4(E) and (G), Table 6 suggests that it is even possible to identify implausible trajectories based only on the current root velocity, even without the human poses.", + "bbox": [ + 89, + 598, + 485, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 89, + 753, + 204, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Estimated Plausibility. We examine the plausibility estimated by $\\nu$ and the actual reward in the locomotion generator. To test this, 200 episodes of locomotion are generated from randomly sampled trajectories and initial humanoid states. The correlation coefficient between the estimated plausibilities and the actual rewards is 0.85. This demonstrates that $\\nu$ can serve as a surrogate for the locomotion generator, contributing to performance improvements.", + "bbox": [ + 89, + 779, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cost-benefit Analysis. The additional training time compared to the baseline [47] is 10 hours in the first stage on a single NVIDIA RTX A6000 GPU. The additional cost in the second stage is nearly negligible because only the computation of $\\mathcal{V}$ is added, which is a shallow MLP. Importantly, $\\mathcal{V}$ can be used with any HTP network, which means only the HTP network needs training in the second stage. Considering these benefits, despite the potential for further cost reductions, our method is highly useful.", + "bbox": [ + 511, + 232, + 908, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Infeasible Trajectory in Dataset. Some trajectories in the synthetic JTA dataset [9] are infeasible rather than implausible for a humanoid and result in low rewards. However, since our target is plausible locomotion, like most pedestrians' locomotion in daily scenes, it is reasonable not to reproduce such infeasible trajectories. During HTP training, these infeasible trajectories are deprioritized since they have low plausibilities and the HTP network prioritizes more plausible trajectories instead. This weighting mechanism can be one reason for the performance improvement of the proposed method, as shown in Tables 1, 2, and 3.", + "bbox": [ + 511, + 386, + 908, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. The smaller performance improvement on the JRDB dataset [34] compared to the JTA dataset [9] (especially in Table 3) is considered to be derived from the assumption of accurate 3D human poses, which is one limitation of the proposed method. In this perspective, future research may extend the EmLoco loss to account for the uncertainty of human poses [54, 57, 68].", + "bbox": [ + 511, + 571, + 908, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Concluding Remarks", + "text_level": 1, + "bbox": [ + 511, + 690, + 718, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presented Locomotion Embodiment, a framework that evaluates the plausibility of predicted trajectories as human locomotion. Our experiments demonstrated that the proposed EmLoco loss improves HTP performance even with the state-of-the-art HTP method [47], without any inference speed deceleration. Furthermore, we established the use of the LocoVal function for trajectory filtering.", + "bbox": [ + 511, + 715, + 906, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importantly, our EmLoco loss improved HTP performance with limited observations even in the real-world dataset [34], demonstrating its effectiveness in real-world scenarios. We hope our findings inspire future physics-aware HTP methods for safe real-world applications.", + "bbox": [ + 511, + 821, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12331", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vida Adeli, Mahsa Ehsanpour, Ian D. Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In ICCV, 2021. 2", + "[2] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social LSTM: human trajectory prediction in crowded spaces. In CVPR, 2016. 1, 2, 4, 5, 6", + "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1", + "[4] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In ECCV, 2020. 2", + "[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2", + "[6] Erwin Coumans. Bullet physics simulation. In SIGGRAPH, 2015. 2", + "[7] Markos Diomataris, Nikos Athanasiou, Omid Taheri, Xi Wang, Otmar Hilliges, and Michael J. Black. WANDR: Intention-guided human motion generation. In CVPR, 2024. 2", + "[8] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R. Qi, Yin Zhou, Zoey Yang, Aurelien Chouard, Pei Sun, Jiquan Ngiam, Vijay Vasudevan, Alexander McCauley, Jonathon Shlens, and Dragomir Anguelov. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In ICCV, 2021. 1", + "[9] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In ECCV, 2018. 5, 6, 7, 8", + "[10] Carlos Flores, Pierre Merdrignac, Raoul de Charette, Francisco M. Navas Matos, Vicente Milanés, and Fawzi Nashashibi. A cooperative car-following/emergency braking system with prediction-based pedestrian avoidance capabilities. TITS, 20(5), 2019. 1", + "[11] Zipeng Fu, Xuxin Cheng, and Deepak Pathak. Deep whole-body control: Learning a unified policy for manipulation and locomotion. In CoRL, 2022. 3", + "[12] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In ICLR, 2022. 2, 5, 6", + "[13] Francesco Giuliani, Irtiza Hasan, Marco Cristani, and Fabio Galasso. Transformer networks for trajectory forecasting. In ICPR, 2021. 6", + "[14] Ivan Grishchenko, Valentin Bazarevsky, Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Zanfir, Richard Yee, Karthik Raveendran, Matsvei Zhdanovich, Matthias Grundmann, and Cristian Sminchisescu. Blazeposition ghum holistic: Real-time" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d human landmarks and pose estimation. arXiv:2206.11678, 2022. 3, 5", + "[15] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In CVPR, 2022. 2", + "[16] Xinyang Gu, Yen-Jen Wang, Xiang Zhu, Chengming Shi, Yanjiang Guo, Yichen Liu, and Jianyu Chen. Advancing humanoid locomotion: Mastering challenging terrains with denoising world model learning. In RSS, 2024. 2", + "[17] Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Back to mlp: A simple baseline for human motion prediction. In WACV, 2023. 2", + "[18] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social GAN: socially acceptable trajectories with generative adversarial networks. In CVPR, 2018. 1, 2, 3, 4, 5, 6", + "[19] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2", + "[20] Dirk Helbing and Péter Molnár. Social force model for pedestrian dynamics. Phys. Rev. E, 51, 1995. 2", + "[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 2", + "[22] Minkwan Kim and Yoonsang Lee. Learning human-like locomotion based on biological actuation and rewards. In SIGGRAPH, 2023. 2", + "[23] Parth Kothari, Sven Kreiss, and Alexandre Alahi. Human trajectory forecasting in crowds: A deep learning perspective. TITS, 23(7), 2021. 2, 6", + "[24] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. In Computer graphics forum, 2007. 5, 8", + "[25] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yikun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. arXiv:2304.00553, 2023. 5", + "[26] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G. Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In CVPR, 2019. 1, 2", + "[27] Zhengyi Luo, Jinkun Cao, Alexander W. Winkler, Kris Kitani, and Weipeng Xu. Perpetual humanoid control for real-time simulated avatars. In ICCV, 2023. 2, 3", + "[28] Takahiro Maeda and Norimichi Ukita. Motionaug: Augmentation with physical correction for human motion prediction. In CVPR, 2022. 2", + "[29] Takahiro Maeda and Norimichi Ukita. Fast inference and update of probabilistic density estimation on trajectory prediction. In ICCV, 2023. 1, 2, 5", + "[30] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: archive of motion capture as surface shapes. In ICCV, 2019. 1, 5", + "[31] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12332", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac gym: High performance GPU based physics simulation for robot learning. In NeurIPS, 2021. 1, 2, 3", + "[32] Karttikeya Mangalam, Ehsan Adeli, Kuan-Hui Lee, Adrien Gaidon, and Juan Carlos Niebles. Disentangling human dynamics for pedestrian locomotion forecasting with noisy supervision. In WACV, 2020. 2", + "[33] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In CVPR, 2023. 2", + "[34] Roberto Martin-Martin, Mihir Patel, Hamid Rezatofighi, Abhijeet Shenoi, JunYoung Gwak, Eric Frankel, Amir Sadeghian, and Silvio Savarese. JRDB: A dataset and benchmark of egocentric robot visual perception of humans in built environments. TPAMI, 45(6), 2023. 1, 5, 6, 7, 8", + "[35] Raul Quintero Minguez, Ignacio Parra Alonso, David Fernández Llorca, and Miguel Ángel Sotelo. Pedestrian path, pose, and intention prediction through gaussian process dynamical models and pedestrian activity recognition. TITS, 20(5), 2019. 1", + "[36] Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In ICML, 2016. 3", + "[37] Mai Nishimura, Shohei Nobuhara, and Ko Nishino. Viewbirdformer: Learning to recover ground-plane crowd trajectories and ego-motion from a single ego-centric view. RA-L, 8(1), 2023. 1, 2", + "[38] Sang II Park, Hyun Joon Shin, and Sung Yong Shin. Online locomotion generation based on motion blending. In SIGGRAPH, 2002. 2", + "[39] Stefano Pellegrini, Andreas Ess, and Luc Van Gool. Improving data association by joint modeling of pedestrian trajectories and groupings. In ECCV, 2010. 5, 8", + "[40] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ToG, 40(4), 2021. 2, 3", + "[41] Tran Phong, Haoran Wu, Cunjun Yu, Panpan Cai, Sifa Zheng, and David Hsu. What truly matters in trajectory prediction for autonomous driving? In NeurIPS, 2023. 1", + "[42] Raul Quintero, Jorge Almeida, David Fernández Llorca, and MA Sotelo. Pedestrian path prediction using body language traits. In IV, 2014. 1, 2", + "[43] Ilija Radosavovic, Tete Xiao, Bike Zhang, Trevor Darrell, Jitendra Malik, and Koushil Sreenath. Real-world humanoid locomotion with reinforcement learning. Sci. Robotics, 9(89), 2024. 2", + "[44] Ilija Radosavovic, Bike Zhang, Baifeng Shi, Jathushan Rajasegaran, Sarthak Kamat, Trevor Darrell, Koushil Sreenath, and Jitendra Malik. Humanoid locomotion as next token prediction. arXiv/2402.19469, 2024. 2", + "[45] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In CVPR, 2023. 2, 3, 4" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Andrey Rudenko, Luigi Palmieri, Michael Herman, Kris M. Kitani, Dariu M. Gavrila, and Kai Oliver Arras. Human motion trajectory prediction: a survey. Int. J. Robotics Res., 39(8), 2020. 2", + "[47] Saeed Saadatnejad, Yang Gao, Kaouther Messaoud, and Alexandre Alahi. Social-transmotion: Promptable human trajectory prediction. In ICLR, 2024. 1, 2, 4, 5, 6, 7, 8", + "[48] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. Sophie: An attentive gan for predicting paths compliant to social and physical constraints. In CVPR, 2019. 2", + "[49] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In ECCV, 2020. 1, 2, 5, 6", + "[50] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. RA-L, 8(11), 2023. 1, 2, 3, 4, 5", + "[51] Christoph Schöller and Alois Knoll. Flomo: Tractable motion prediction with normalizing flows. In IROS, 2021. 2", + "[52] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv:1707.06347, 2017. 3", + "[53] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 2", + "[54] Megh Shukla, Roshan Roy, Pankaj Singh, Shuaib Ahmed, and Alexandre Alahi. VL4pose: Active learning through out-of-distribution detection for pose estimation. In BMVC, 2022, 8", + "[55] Jianhua Sun, Yuxuan Li, Liang Chai, Haoshu Fang, Yong-Lu Li, and Cewu Lu. Human trajectory prediction with momentary observation. In CVPR, 2022. 5, 6", + "[56] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 2", + "[57] Hiromu Taketsugu and Norimichi Ukita. Active transfer learning for efficient video-specific human pose estimation. In WACV, 2024. 8", + "[58] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 2", + "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6", + "[60] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In ICCV, 2023. 1, 2", + "[61] Jingbo Wang, Zhengyi Luo, Yixuan Li, Ye Yuan, and Bo Dai. Pacer+: On-demand pedestrian animation controller in driving scenarios. In CVPR, 2024. 1, 2, 3", + "[62] Jungdam Won, Deepak Gopinath, and Jessica K. Hodgins. Physics-based character controllers using conditional vaes. ToG, 41(4), 2022. 2" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12333", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8", + "[64] Zhuo Xu, Rui Zhou, Yida Yin, Huidong Gao, Masayoshi Tomizuka, and Jiachen Li. MATRIX: multi-agent trajectory generation with diverse contexts. In ICRA, 2024. 1, 2, 5, 6", + "[65] Takuma Yagi, Karttikeya Mangalam, Ryo Yonetani, and Yoichi Sato. Future person localization in first-person videos. In CVPR, 2018. 2", + "[66] Shuai Yi, Hongsheng Li, and Xiaogang Wang. Understanding pedestrian behaviors from stationary crowd groups. In CVPR, 2015. 1", + "[67] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, 2023. 2", + "[68] Yufei Zhang, Hanjing Wang, Jeffrey O. Kephart, and Qiang Ji. Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In ICCV, 2023. 8", + "[69] Yufei Zhang, Jeffrey O. Kephart, and Qiang Ji. Incorporating physics principles for precise human motion prediction. In WACV, 2024. 2", + "[70] Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Benjamin Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yuning Chai, Cordelia Schmid, Congcong Li, and Dragomir Anguelov. TNT: target-driven trajectory prediction. In CoRL, 2020. 1", + "[71] Jingbo Zhao, Zhetao Wang, Yiqin Peng, and Yaojun Wang. Real-time generation of leg animation for walking-in-place techniques. In SIGGRAPH, 2022. 2", + "[72] Bolei Zhou, Xiaou Tang, and Xiaogang Wang. Learning collective crowd behaviors with dynamic pedestrian-agents. IJCV, 111(1):50-68, 2015. 1", + "[73] Brian D Ziebart, Nathan Ratliff, Garratt Gallagher, Christoph Mertz, Kevin Peterson, J Andrew Bagnell, Martial Hebert, Anind K Dey, and Siddhartha Srinivasa. Planning-based prediction for pedestrians. In IROS, 2009. 1" + ], + "bbox": [ + 91, + 90, + 485, + 603 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12334", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/88abcf25-51f5-40e1-9600-4a2e38b6cc9a_model.json b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/88abcf25-51f5-40e1-9600-4a2e38b6cc9a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cf5a80cb03a6ac16a678cae1dd6561f73a3eb41d --- /dev/null +++ b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/88abcf25-51f5-40e1-9600-4a2e38b6cc9a_model.json @@ -0,0 +1,2499 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.898, + 0.152 + ], + "angle": 0, + "content": "Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.18, + 0.738, + 0.215 + ], + "angle": 0, + "content": "Hiromu Taketsugu\\(^{1}\\) Takeru Oba\\(^{1}\\) Takahiro Maeda\\(^{1}\\) Shohei Nobuhara\\(^{2}\\) Norimichi Ukita\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.217, + 0.821, + 0.236 + ], + "angle": 0, + "content": "\\(^{1}\\)Toyota Technological Institute, Japan \\(^{2}\\)Kyoto Institute of Technology, Japan" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.237, + 0.853, + 0.253 + ], + "angle": 0, + "content": "{sd23426, sd21502, sd21601}@toyota-ti.ac.jp, nob@kit.ac.jp, ukita@toyota-ti.ac.jp" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.28, + 0.327, + 0.296 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.486, + 0.599 + ], + "angle": 0, + "content": "Humans can predict future human trajectories even from momentary observations by using human pose-related cues. However, previous Human Trajectory Prediction (HTP) methods leverage the pose cues implicitly, resulting in implausible predictions. To address this, we propose Locomotion Embodiment, a framework that explicitly evaluates the physical plausibility of the predicted trajectory by locomotion generation under the laws of physics. While the plausibility of locomotion is learned with an indifferentiable physics simulator, it is replaced by our differentiable Locomotion Value function to train an HTP network in a data-driven manner. In particular, our proposed Embodied Locomotion loss is beneficial for efficiently training a stochastic HTP network using multiple heads. Furthermore, the Locomotion Value filter is proposed to filter out implausible trajectories at inference. Experiments demonstrate that our method enhances even the state-of-the-art HTP methods across diverse datasets and problem settings. Our code is available at: https://github.com/ImIntheMiddle/EmLoco." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.222, + 0.635 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.486, + 0.81 + ], + "angle": 0, + "content": "Human Trajectory Prediction (HTP), which predicts future human trajectories, is actively studied [2, 18, 29, 37, 49, 63, 64, 70] due to its importance for applications such as autonomous driving [3, 8, 41, 60, 61, 70], robotics [34, 37, 50, 73], and security systems [10, 66, 72]. These HTP methods assume the availability of sufficiently long past observations and rely solely on a past trajectory. Due to this limitation, these methods cannot handle safety-critical situations in a momentary observation setting, where only a few frames of the past trajectory are available, such as when people suddenly emerge from behind obstacles." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Regarding this issue, prior work [26, 35, 42, 47, 50] validates that human pose-related cues, such as body orientations and limb motions are beneficial for inferring future trajectories, especially in a momentary observation setting. These methods learn the consistency between human past poses and future trajectories in a data-driven manner alone." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.287, + 0.906, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.525, + 0.907, + 0.638 + ], + "angle": 0, + "content": "Figure 1. Overview of our method. Unlike existing methods, which often predict physically implausible trajectories, our framework uses locomotion generation in a physics simulator to incorporate the laws of physics to HTP by training the plausibility score as the consistency between the observed 3D pose and future possible trajectories, which are indicated by \\( \\boldsymbol{j}_0 \\) and \\( \\tau_f^{1,2,3} \\), respectively, in the figure. Additionally, at inference, our method can evaluate predicted trajectories, \\( \\hat{\\tau}_f^{1,2,3} \\), to filter out implausible ones." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.657, + 0.909, + 0.809 + ], + "angle": 0, + "content": "The data-driven approach can prove its ability with a large-scale dataset that covers most possible combinations of poses and trajectories. However, such datasets are unavailable in the literature because human locomotion is diverse. In particular, it is difficult to collect a large number of accurate pose annotations, which are collected by time-consuming delicate motion capture systems [30]. Instead, we propose to fully leverage the beneficial cues of human poses, by assessing the physical consistency between observed poses and future trajectories for physical plausibility-aware HTP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.911, + 0.902 + ], + "angle": 0, + "content": "For this physical plausibility-aware HTP, we present Locomotion Embodiment, a framework that incorporates the laws of physics into an HTP network. This is achieved by physically simulating the locomotion conditioned by observed poses and future trajectories in a physics simulator [31]; refer to Fig. 1 to see the effect of our framework. As a typical" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12324" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.184 + ], + "angle": 0, + "content": "example, one of the human poses included in a future possible trajectory \\(\\tau_f^1\\) (indicated by \\(j_{t}\\) in Fig. 1) is a little unstably tilted, leading to a lower plausibility score of \\(\\tau_f^1 = 0.3\\). By training the HTP network in a data-driven manner with following the laws of physics, we benefit from the advantages of both the data-driven and physics-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.485, + 0.412 + ], + "angle": 0, + "content": "By evaluating and back-propagating the plausibility of locomotion generated in a physics simulator, we can train the physical plausibility-aware HTP network. However, locomotion generation using physics simulators requires various motion parameters, most of which are unavailable or noisy (e.g., joint angular velocities estimated from joint positions provided by a 3D pose estimator) in image-based HTP scenarios. In addition, rigorous physics simulators are indifferent to the physical plausibility, this paper proposes the Locomotion Value (LocoVal) function implemented with a differentiable network to serve as a surrogate for a physics simulator for the plausibility evaluation (Sec. 4.1). Given the plausibility evaluation provided by the LocoVal function, our EmLoco loss penalizes implausible trajectories (Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.579 + ], + "angle": 0, + "content": "This EmLoco loss is used in conjunction with the MSE loss for supervised trajectory prediction learning. While the MSE loss behaves as a data term, our EmLoco loss can be a regularization term that incorporates prior knowledge about physics. This is crucial for maintaining the diversity of the predicted trajectories. Although human trajectories are inherently stochastic due to many factors, such as dynamic environments, training with the MSE loss alone forces all predicted trajectories to be aligned with a single ground truth. That is, minimizing the MSE essentially decreases the diversity of the predicted trajectories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.671 + ], + "angle": 0, + "content": "On the other hand, for stochastically predicting multiple future trajectories, our EmLoco loss allows us to supervise all predicted trajectories to improve their physical plausibility while preserving diversity. In addition, our LocoVal function can also serve as the LocoVal filter that evaluates and filters the candidate trajectories at inference (Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.672, + 0.441, + 0.687 + ], + "angle": 0, + "content": "Our framework consists of three key components:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.688, + 0.483, + 0.718 + ], + "angle": 0, + "content": "1. LocoVal function as a surrogate for an indifferentiable physics simulator in HTP network training." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.719, + 0.484, + 0.794 + ], + "angle": 0, + "content": "2. EmLoco loss for back-propagating the plausibility of predicted trajectories. This loss is also useful for deterministic predictors but more beneficial for stochastic predictors because of its applicability to all possible predicted trajectories while preserving their stochastic diversity." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.795, + 0.483, + 0.839 + ], + "angle": 0, + "content": "3. LocoVal filter for filtering out implausible trajectories at inference. This filter can be applied to any pre-trained HTP network in a plug-and-play manner." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.688, + 0.484, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "This work is the first attempt to use simulated humanoid control in a physics simulator for HTP. Since the physics simulator is used only during training, an additional cost (i.e., the cost of the LocoVal function) is negligible at inference." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.115, + 0.831, + 0.132 + ], + "angle": 0, + "content": "2.1. Human Trajectory Prediction (HTP)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.908, + 0.349 + ], + "angle": 0, + "content": "HTP methods can be divided into deterministic [2, 23, 37, 47, 63] and stochastic [18, 29, 49, 50, 64] methods. Deterministic HTP methods [2, 20, 23, 37, 46, 63] predict the most likely trajectory. For example, Social-LSTM uses the Long Short Term Memory [2] to represent the consistency between observed and future trajectories. On the other hand, stochastic HTP methods [12, 49, 50, 64] produce multiple trajectories or probability distributions. To model the inherent stochasticity in HTP, these methods employ probabilistic frameworks such as Generative Adversarial Networks [18, 48] and Diffusion models [15, 33], as well as Normalizing Flows [29, 51] to draw multiple trajectories at inference. Our method is model-agnostic and applicable to both deterministic and stochastic methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.762, + 0.374 + ], + "angle": 0, + "content": "2.2. HTP with Pose Information" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.907, + 0.503 + ], + "angle": 0, + "content": "Prior studies demonstrate that human poses benefit HTP [26, 42, 47, 50, 65]. Yagi et al. [65] and Liang et al. [26] incorporate human poses estimated by CNNs into HTP. Transformer-based methods such as Social-Transmotion [47] and Human Scene Transformer (HST) [50] further improve HTP conditioned by human poses. In particular, HST [50] reports the effectiveness of human pose conditioning for momentary observation scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.906, + 0.595 + ], + "angle": 0, + "content": "However, since these methods include poses simply as an additional input to the HTP network, predicted future trajectories can be physically inconsistent with past poses. Inspired by this observation, our Locomotion Embodiment framework represents the physical plausibility of predicted trajectories as pose-trajectory consistent locomotion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.641 + ], + "angle": 0, + "content": "Note that we aim to improve HTP by fully leveraging human poses, distinct from Human Motion Prediction which prioritizes predicting future poses [1, 4, 17, 28, 32, 69]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.734, + 0.667 + ], + "angle": 0, + "content": "2.3. Locomotion Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.824 + ], + "angle": 0, + "content": "Locomotion generation is a task to generate walking motions for humanoid robots [16, 43, 44] and virtual characters [22, 38, 71]. Locomotion can be controlled on various conditions such as destinations [7, 19, 56], trajectories [21, 45], and poses [27, 61]. Generative models such as Diffusion models [5, 21, 53, 58, 67] can generate such conditioned locomotion, but they are prone to generate physically implausible locomotion (e.g., skating motions). This is because the laws of physics are not explicitly formulated in generative models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "On the other hand, physics-based locomotion generation [27, 40, 45, 60-62] follows the laws of physics in simulators [6, 31]. However, a generated locomotion would fail to follow a given trajectory or fall due to kinematic limitations if a humanoid is controlled to walk along an implausible" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12325" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "trajectory. Therefore, a physics-based locomotion generator can evaluate whether the conditioned trajectory is plausible. Our Locomotion Embodiment framework utilizes this ability to evaluate the plausibility of future trajectories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.486, + 0.395 + ], + "angle": 0, + "content": "The aforementioned physics-based locomotion generator, e.g., PACER [27, 40, 45, 61], can be trained with reinforcement learning in a physics simulator. A straightforward way to use this locomotion generator for training an HTP network is to directly incorporate the locomotion generator into the HTP network. However, this direct integration has two problems: (1) As mentioned in Sec. 1, the physics simulator is indifferent, preventing training the HTP network by back-propagation. (2) The locomotion generator using the physics simulator requires physical humanoid states, e.g., joint angular velocities, which are unavailable or noisy in image-based HTP scenarios. To address these two problems, our framework uses the physics simulator and the locomotion generator only before training the HTP network and replaces them with the LocoVal function in training the HTP network." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.416, + 0.23, + 0.431 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.549 + ], + "angle": 0, + "content": "HTP. An HTP network, \\(\\mathcal{H}\\), takes the past \\(T_{\\mathrm{p}}\\) frames as input, including the past trajectory and 3D poses of a target person, and predicts the trajectory of the target person in the future \\(T_{\\mathrm{f}}\\) frames. The past trajectory \\(\\pmb{\\tau}_{\\mathrm{p}} \\in \\mathbb{R}^{T_{\\mathrm{p}} \\times 2}\\) and the future target trajectory \\(\\pmb{\\tau}_{\\mathrm{f}} \\in \\mathbb{R}^{T_{\\mathrm{f}} \\times 2}\\) are on the ground \\(xy\\) plane. The 3D pose involves \\(J\\) joint positions \\(\\pmb{j}_t \\in \\mathbb{R}^{J \\times 3}\\) at the past time \\(t\\), where \\((-T_{\\mathrm{p}} + 1) \\leq t \\leq 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.583 + ], + "angle": 0, + "content": "The predicted future trajectory \\(\\hat{\\tau}_{\\mathrm{f}}\\) is expected to be as close as possible to its ground-truth trajectory \\(\\tau_{\\mathrm{f}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.484, + 0.793 + ], + "angle": 0, + "content": "Locomotion Generation. In our framework, humanoid locomotion is generated to follow any given future trajectory \\(\\pmb{\\tau}_{\\mathrm{s}} \\in \\mathbb{R}^{T_{\\mathrm{f}} \\times 2}\\) in a physics simulator [31]. As such a locomotion generator \\(\\mathcal{G}\\), our method employs PACER [45], which is implemented with the policy network [36] trained via reinforcement learning [52]. \\(\\mathcal{G}\\) takes \\(\\pmb{\\tau}_{\\mathrm{s}}\\) and the initial humanoid state \\(h_0\\) to control the humanoid joints. \\(h_0\\) involves the positions, rotations, velocities, and angular velocities of all joints. In PACER [45], \\(\\mathcal{G}\\) is trained by maximizing the discounted cumulative reward \\(\\Omega\\) consisting of a trajectory following reward, a motion style reward [40], and an energy penalty [11]. Refer to PACER [45] for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In PACER, the humanoid movement is constrained by the limits of joint angles and torques, preventing it from producing kinematically impossible locomotion. Due to this limitation, when conditioned on a physically implausible trajectory, the humanoid fails to follow it, resulting in a lower reward. Therefore, we define the discounted cumulative reward \\(\\Omega\\) [52] as the plausibility of \\(\\tau_{\\mathrm{s}}\\)." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.094, + 0.903, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.351, + 0.907, + 0.38 + ], + "angle": 0, + "content": "Figure 2. The overview of the proposed framework. \\(①\\) and \\(⑤\\) mean the weights are trained and fixed, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.405, + 0.81, + 0.423 + ], + "angle": 0, + "content": "4. Physical Plausibility-aware HTP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.909, + 0.644 + ], + "angle": 0, + "content": "Overview. Figure 2 shows the overview of our framework consisting of two training stages and an inference phase. In (a) the first training stage, given the physics-based locomotion generator \\(\\mathcal{G}\\) trained as mentioned in Sec. 3, a current humanoid state is fed into it to obtain the reward as the physical plausibility. With this physical plausibility as the ground truth, the LocoVal function \\(\\nu\\) is trained. In (b) the second stage, observed pose-related cues in training data are fed into both \\(\\nu\\) and \\(\\mathcal{H}\\). \\(\\nu\\) also takes the future trajectories predicted by \\(\\mathcal{H}\\) to compute the EmLoco losses. The EmLoco loss supports both the MSE and minMSE [18, 50, 63] losses used for deterministic and stochastic predictions, respectively, to train \\(\\mathcal{H}\\). At (c) inference, candidate trajectories predicted by \\(\\mathcal{H}\\) are fed into \\(\\nu\\) to filter out implausible trajectories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.688, + 0.667 + ], + "angle": 0, + "content": "4.1. LocoVal Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.78 + ], + "angle": 0, + "content": "The detail of training the LocoVal function \\(\\mathcal{V}\\) is illustrated in Fig. 3. As shown in the top half of Fig. 3, the discounted cumulative reward \\(\\Omega\\) [52] is used as the plausibility of Embodied Locomotion that is generated through \\(\\mathcal{G}\\) using a physics simulator. \\(\\mathcal{V}\\) is trained to estimate the reward \\(\\hat{\\Omega}\\) close to \\(\\Omega\\) from only available cues in image-based HTP scenarios because \\(\\mathcal{V}\\) serves as a surrogate for \\(\\mathcal{G}\\) at inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.911, + 0.902 + ], + "angle": 0, + "content": "As such available cues, our method uses the future trajectory \\(\\tau_{\\mathrm{s}}\\), the initial pose \\(j_0\\), and the initial velocity of the root joint \\(\\boldsymbol{v}_{\\mathrm{root},0} \\in \\mathbb{R}^2\\), where \\(j_0\\) and \\(\\boldsymbol{v}_{\\mathrm{root},0}\\) are in a subset \\(h_0'\\) extracted from \\(h_0\\). In training \\(\\mathcal{V}\\), these cues are available from the physics simulator. At inference, \\(\\tau_{\\mathrm{s}}\\) is replaced by \\(\\tau_{\\mathrm{f}}\\), \\(j_0\\) can be obtained by a 3D pose estimator [14, 50], and \\(\\boldsymbol{v}_{\\mathrm{root},0}\\) can be obtained from \\(\\tau_p\\). \\(\\mathcal{V}\\) is implemented as a differentiable MLP, taking these cues as input." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12326" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.485, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.211, + 0.466, + 0.226 + ], + "angle": 0, + "content": "Figure 3. The overview of training our LocoVal function, \\(\\nu\\)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.239, + 0.481, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.387, + 0.454, + 0.401 + ], + "angle": 0, + "content": "Figure 4. The overview of training the HTP network, \\(\\mathcal{H}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.425, + 0.483, + 0.591 + ], + "angle": 0, + "content": "Training Process. Unlike PACER [45], which trains \\(\\mathcal{G}\\) to obtain high rewards even for implausible trajectories, e.g., dangerously sharp turns, such trajectories should result in low rewards in our framework. This is because humans do not typically walk along such implausible trajectories in daily life. For this reason, we train \\(\\mathcal{G}\\) with only plausible pairs of \\(\\tau_{s}\\) and \\(h_0\\) to give high rewards only to plausible pose-trajectory pairs. On the other hand, since \\(\\mathcal{V}\\) must discriminate between the plausible and implausible pose-trajectory pairs, it is trained with plausible pairs and randomly paired diverse implausible pose-trajectory pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.484, + 0.639 + ], + "angle": 0, + "content": "\\(\\mathcal{V}\\) is trained by minimizing the following loss \\(\\mathcal{L}_{\\mathcal{V}}\\) so that it can estimate the plausibility of the Embodied Locomotion \\(\\varOmega=\\mathcal{G}(\\pmb{\\tau}_{\\mathrm{s}},\\pmb{h}_{0})\\) obtained by rolling out \\(\\mathcal{G}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.646, + 0.483, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {V}} = \\operatorname {M S E} \\left(\\mathcal {V} \\left(\\boldsymbol {\\tau} _ {\\mathrm {s}}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right), \\mathcal {G} \\left(\\boldsymbol {\\tau} _ {\\mathrm {s}}, \\boldsymbol {h} _ {0}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.45, + 0.687 + ], + "angle": 0, + "content": "where MSE denotes the Mean Squared Error function." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.695, + 0.454, + 0.711 + ], + "angle": 0, + "content": "4.2. HTP Network Training with EmLoco Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.484, + 0.809 + ], + "angle": 0, + "content": "Figure 4 illustrates the detail of training the HTP network \\(\\mathcal{H}\\). Our HTP network takes the past information including \\(\\tau_{\\mathrm{p}}\\) and 3D poses \\(j_{\\mathrm{p}} = \\{j_{-T_{\\mathrm{p}} + 1},\\dots ,j_0\\}\\) as input to predict the future trajectory \\(\\hat{\\tau}_{\\mathrm{f}}\\) of a target person. Following traditional HTP training methods [2, 47], \\(\\hat{\\tau}_{\\mathrm{f}}\\) is compared with its ground truth \\(\\tau_{\\mathrm{f}}\\) to back-propagate the MSE loss \\(\\mathcal{L}_{\\mathrm{T}} = \\mathrm{MSE}(\\hat{\\tau}_{\\mathrm{f}},\\tau_{\\mathrm{f}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "EmLoco loss. The EmLoco loss \\(\\mathcal{L}_{\\mathrm{E}}\\) evaluates predicted trajectories by using \\(\\mathcal{V}\\), which has prior knowledge about plausibility. Note again that a physics simulator is not required in this second stage or the inference phase because \\(\\mathcal{V}\\) serves as a surrogate for the plausibility evaluation process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "Given the prediction of \\(\\mathcal{H}\\), the predicted trajectory \\(\\hat{\\tau}_{\\mathrm{f}}\\) and the subset of \\(h_0\\) (i.e., \\(h_0^{\\prime}\\)) observed in the current frame are fed into \\(\\nu\\) to compute \\(\\mathcal{L}_{\\mathrm{E}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.145, + 0.907, + 0.162 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E}} = - \\mathcal {V} \\left(\\hat {\\tau} _ {\\mathrm {f}}, h _ {0} ^ {\\prime}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.169, + 0.907, + 0.184 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{E}}\\) encourages \\(\\mathcal{H}\\) to predict \\(\\hat{\\tau}_{\\mathrm{f}}\\) that increases the plausibility." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.184, + 0.906, + 0.215 + ], + "angle": 0, + "content": "As the total training objective, \\(\\mathcal{L}_{\\mathrm{E}}\\) is combined with the MSE loss \\(\\mathcal{L}_{\\mathrm{T}}\\) with a hyper-parameter \\(\\alpha\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.223, + 0.907, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {T}} + \\alpha \\mathcal {L} _ {\\mathrm {E}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.256, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Multi-head HTP network. Our method with the EmLoco loss can be extended to a stochastic HTP network using multiple heads. As also mentioned in Sec. 1, the MSE loss (Eq. (1)) is inappropriate for multiple stochastic predictions because it forces all of them to converge to a unique ground truth. To mitigate this problem, following prior work [18, 50, 63], the multiple heads in our method are trained by replacing \\(\\mathcal{L}_{\\mathrm{T}}\\) in Eq.(3) with the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.39, + 0.907, + 0.414 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T}} = \\min _ {k} \\left(\\operatorname {M S E} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {\\tau} _ {\\mathrm {f}}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.905, + 0.463 + ], + "angle": 0, + "content": "where \\(\\hat{\\tau}_{\\mathrm{f}}^{k}\\) denotes the prediction of \\(k\\)-th head. In Eq. (4), the minMSE loss supervises only one prediction closest to its ground truth \\(\\tau_{\\mathrm{f}}\\) among multiple predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.906, + 0.554 + ], + "angle": 0, + "content": "Unlike the MSE loss, this minMSE loss allows us to maintain diverse predictions. However, the minMSE loss is inefficient. What is even worse is that several heads may be under-fitted if other heads are trained in early training iterations so that these heads are always selected as those minimizing the MSE in Eq. (4) in later iterations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.676 + ], + "angle": 0, + "content": "On the other hand, our EmLoco loss does not require the unique ground truth, i.e., \\(\\tau_{\\mathrm{f}}\\) in Eq. (4). As a regularization term, the EmLoco loss trains all the heads to predict plausible trajectories. This simultaneous optimization allows us to avoid the under-fitting problem mentioned above. In our method, this simultaneous optimization is implemented by averaging the EmLoco losses using \\(\\mathcal{V}\\) for all predicted trajectories, and this average is used as \\(\\mathcal{L}_{\\mathrm{E}}\\) in Eq.(3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.757, + 0.698 + ], + "angle": 0, + "content": "4.3. LocoVal Filter at Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.81 + ], + "angle": 0, + "content": "The trained LocoVal function can also be applied to trajectory filtering, called the LocoVal filter, at inference using multi-head HTP networks. Similar to the usage of \\(\\mathcal{V}\\) for training multiple heads simultaneously at training, it computes the plausibilities of all predicted trajectories at inference. Each of the predicted trajectories with a plausibility lower than a threshold \\(\\lambda\\) is filtered out as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.816, + 0.907, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} \\left\\{\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k} \\mid \\mathcal {V} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right) \\geq \\lambda \\right\\} & \\text {i f} \\exists k, \\mathcal {V} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right) \\geq \\lambda , \\\\ \\operatorname {a r g m a x} _ {\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}} \\mathcal {V} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Note that the LocoVal filter can be used with any stochastic HTP networks in a plug-and-play manner." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12327" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.226, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.118, + 0.486, + 0.225 + ], + "angle": 0, + "content": "In Sec. 5.2.1, we benchmark our HTP network trained with the EmLoco loss under the standard HTP setting. Section 5.2.2 presents results for multi-head HTP networks, and Sec. 5.2.3 examines the performance with momentary observations. Then, we report ablation results in Sec. 5.2.4. In Sec. 5.3.1 and Sec. 5.3.2, we evaluate the effectiveness of the LocoVal filter in various scenarios." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.245, + 0.3, + 0.263 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.271, + 0.485, + 0.378 + ], + "angle": 0, + "content": "Datasets. As trajectory datasets, the Joint Track Auto (JTA) dataset [9], the JackRabbit Dataset and Benchmark (JRDB) [34], and the ETH/UCY dataset [24, 39] are used. For the JRDB dataset, a subset provided by the SocialTrans [47] was used. The ETH/UCY dataset is used only in Sec. 5.3.2 to evaluate the performance of the LocoVal filter on pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.486, + 0.517 + ], + "angle": 0, + "content": "While JTA also provides 3D human poses synchronized with their trajectories, no 3D pose is annotated in JRDB. Instead, in JRDB, our experiments use 3D poses estimated from images and 2D pose annotations by using [14, 50]. While the estimated 3D poses are provided by HST [50], misestimated poses are filtered out in our experiments; see the Supp. for the details. All the 3D poses mentioned above are converted to the SMPL-style poses [25] for a fair comparison with the baseline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.486, + 0.656 + ], + "angle": 0, + "content": "While the 3D poses described above are used in the second stage to train the HTP network \\(\\mathcal{H}\\), more accurate 3D poses are expected to train \\(\\mathcal{G}\\) and \\(\\mathcal{V}\\) to learn the physical plausibility of 3D poses. Such accurate 3D poses are provided by the Archive of Motion Capture as Surface Shapes (AMASS) [30]. While accurate 3D poses are available and can be used to compute the trajectories in motion capture datasets such as AMASS, motion capture data do not scale compared with trajectory datasets, as described in Sec. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.486, + 0.901 + ], + "angle": 0, + "content": "To avoid this problem, 3D poses in AMASS are combined with trajectories in the trajectory datasets, i.e., JTA and JRDB, as follows. Plausible and implausible pose-trajectory pairs are required to train \\(\\mathcal{G}\\) and \\(\\mathcal{V}\\), as described in Sec. 4.1. The plausible pairs are prepared by spatially aligning the 3D pose and the trajectories, which are randomly sampled from AMASS and the trajectory datasets, respectively. This spatial alignment consists of the orientation and velocity alignments. Given the heading vector of each pedestrian in these two datasets, the heading orientations are aligned between the motion capture pose from AMASS and the trajectory data. For the velocity alignment, on the other hand, the trajectory data is expanded or contracted to make the vector norms equal between the motion capture pose and the trajectory data. Refer to the Supp. for detailed training conditions on the aforementioned datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.469 + ], + "angle": 0, + "content": "HTP networks. Since multiple people are observed in each frame of the aforementioned datasets, all baseline HTP networks of our method take not only the past trajectory of a target person but also those of other people as input. This allows us to predict the target trajectory affected by the dynamic environment. While our method is introduced to take only the target person as input, all people in the scene are fed into \\(\\mathcal{H}\\) following the previous methods. In the evaluations with the JTA and JRDB datasets, Social-Transmotion (Social-Trans), is used as our baseline without any architectural changes. We extend this baseline of the trajectory predictor \\(\\mathcal{H}\\) by plugging the EmLoco loss into the training loss. Following the implementation of Social-Trans, past trajectories, 2D and 3D human poses, and 2D and 3D bounding boxes are fed into \\(\\mathcal{H}\\) in the JTA dataset [9], while the past trajectories, 2D bounding boxes, and 3D human poses are used in the JRDB dataset. Note that the 3D human poses are estimated for the JRDB dataset [34] as mentioned earlier. For the stochastic version of our method, we modify Social-Trans [47] by parallelizing the prediction heads with different initial weights. For comparison, all HTP networks predict the future \\(T_{\\mathrm{f}} = 12\\) frames at 2.5 fps on the JTA and JRDB datasets. The number of past frames differs depending on the experiment; e.g., 9 frames in Sec. 5.2.1 and Sec. 5.2.2, and 2 frames as momentary observations [55] in Sec. 5.2.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.909, + 0.577 + ], + "angle": 0, + "content": "In experiments on the ETH/UCY dataset [24, 39], where a bird's-eye view hinders the use of 3D poses, only past trajectories are fed into \\(\\mathcal{H}\\). Here, we employ EqMotion [63], a state-of-the-art stochastic HTP method that only uses past trajectories, as a pre-trained HTP network. Each past trajectory consists of \\(T_{\\mathrm{p}} = 8\\) frames, while each future trajectory remains at \\(T_{\\mathrm{f}} = 12\\) frames at 2.5 fps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.911, + 0.75 + ], + "angle": 0, + "content": "Evaluation. Two major metrics, the Average Displacement Error (ADE) and the Final Displacement Error (FDE), both measured in meters, evaluate the HTP results. Consistent with the other HTP literature [29, 49, 50, 63], in evaluating the stochastic HTP networks, we report minADE and minFDE among multiple trajectories. Also, following MATRIX [64], we report \\(\\chi^2\\) distances for the four physics primitives of each predicted trajectory, namely velocities, accelerations, angular velocities, and angular accelerations, representing the plausibility of the predicted trajectories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.759, + 0.835, + 0.775 + ], + "angle": 0, + "content": "5.2. HTP Training with the EmLoco Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.782, + 0.91, + 0.828 + ], + "angle": 0, + "content": "Here, we evaluate our HTP network trained with the proposed EmLoco loss. Note that the LocoVal filter is not applied in these experiments. \\(\\alpha\\) in Eq.(3) is set to 100." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.836, + 0.845, + 0.852 + ], + "angle": 0, + "content": "5.2.1. Comparison with Existing HTP Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Following the standard protocol [2, 12, 18, 47, 63] using 9 past frames, where past trajectories are sufficiently long, we compare the accuracy of deterministic predictions by our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12328" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 1. HTP results under the standard setting. Red and blue indicate the best and second-best. 'Social-Trans' and 'Ours' utilize pose input. For methods other than 'Social-Trans' and 'Ours', we report the results from the Social-Trans literature [47]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.152, + 0.482, + 0.372 + ], + "angle": 0, + "content": "
MethodJTA [9]JRDB [34]
ADEFDEADEFDE
Social-GAN-det [18]1.663.760.500.99
Transformer [13]1.563.540.561.10
Vanilla-LSTM [2]1.443.250.420.83
Occupancy-LSTM [2]1.413.150.430.85
Directional-LSTM [23]1.373.060.450.87
Dir-social-LSTM [23]1.232.590.480.95
Social-LSTM [2]1.212.540.470.95
Autobots [12]1.202.700.390.80
Trajectory++ [49]1.182.530.400.78
EqMotion [63]1.132.390.400.77
Social-Trans [47]1.112.260.400.76
Ours0.971.910.370.72
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.389, + 0.488, + 0.529 + ], + "angle": 0, + "content": "HTP network with other HTP networks [2, 12, 18, 23, 47, 49, 59, 63]. Table 1 shows that our method outperforms all other methods including the state-of-the-art baseline, Social-Trans [47], on both the JTA [9] and the JRDB [34] datasets. This demonstrates that the EmLoco loss effectively provides the prior knowledge of the trajectory-pose consistency during training. Notably, the performance improvements on the JRDB dataset [34] with imperfect 3D poses suggest the applicability of our EmLoco loss to real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.534, + 0.327, + 0.549 + ], + "angle": 0, + "content": "5.2.2. Multi-head HTP Networks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.486, + 0.659 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the EmLoco loss on a multi-head HTP network, 5 and 20 prediction heads are implemented on Social-Trans [47]. As described in Sec. 4.2, our proposed method is expected to improve the plausibility of predicted trajectories closer to human locomotion. From this perspective, we evaluate the plausibility of the predicted trajectories by the \\(\\chi^2\\) distance employed in MATRIX [64]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.485, + 0.795 + ], + "angle": 0, + "content": "Table 2 presents the comparisons with the baseline. Our method achieves the best values across almost all metrics on both the JTA [9] and JRDB [34] datasets. Particularly, the improvements in ADE and FDE reveal that all heads are trained properly because these metrics are evaluated with the prediction of all heads. Also, improvements in \\(\\chi^2\\) distances, the plausibility regarding physics primitives, demonstrate that \\(\\mathcal{H}\\) trained with our EmLoco loss predicts trajectories similar to actual human locomotion than the baseline [47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Figure 5 shows qualitative results comparing stochastic HTP with 5 heads for both Social-Trans [47] and our method. While the Social-Trans trained solely with minMSE can predict diverse trajectories, many of them are implausible, including sharp turns. In contrast, the predictions of our HTP network maintain diversity across all heads while remaining plausible, resulting in trajectories closer to the ground truth." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.097, + 0.907, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.244, + 0.909, + 0.288 + ], + "angle": 0, + "content": "Figure 5. Comparison of the stochastic HTP between Social-Trans [47] and ours with 5 heads on the JTA dataset [9]. Human poses are shown at a doubled scale for visualization purposes." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.316, + 0.907, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.481, + 0.909, + 0.538 + ], + "angle": 0, + "content": "Figure 6. Comparison of the prediction results between the baseline Social-Trans [47] and our method with momentary observations on the JTA dataset [9]. The observed pose are displayed at a doubled scale for visualization purposes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.813, + 0.587 + ], + "angle": 0, + "content": "5.2.3. HTP with Momentary Observations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.909, + 0.687 + ], + "angle": 0, + "content": "As noted in Sec. 1, existing HTP methods struggle with momentary observations. On the contrary, because the proposed EmLoco loss encourages physical plausibility-aware HTP, \\(\\mathcal{H}\\) trained with the EmLoco loss is expected to be robust against this setting. From this point of view, we experiment with 2 past frames following the prior work [55]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.911, + 0.903 + ], + "angle": 0, + "content": "The results are shown in Table 3. Consistent with Tables 1 and 2, the EmLoco loss allows our method to outperform the baseline [47] across almost all categories. In addition, Fig. 6 presents qualitative results in two cases: when a pedestrian is (a) going straight and (b) turning. In both cases, (i) Social-Trans [47] with trajectory-only input results in significant prediction errors, and (ii) Social-Trans with pose input still predicts implausible trajectories, especially during sharp turns. In contrast, (iii) our method provides plausible predictions that closely follow the ground-truth trajectory and align with the observed pose. This demonstrates that the proposed method can effectively leverage human poses and achieve plausibility-aware HTP, thanks to the physics-based LocoVal function training." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12329" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.146 + ], + "angle": 0, + "content": "Table 2. Comparison of our HTP network using the EmLoco loss with baseline [47] in the stochastic HTP under 9 frames of observations. Red indicates the best result for each category. The results are presented as left/right values, where the left denotes evaluations with 5 heads and the right denotes 20 heads. 'Vel,' 'Acc,' 'Ang. Vel,' and 'Ang. Acc.' represent the \\( \\chi^2 \\) distances between the distributions of the predicted and ground-truth trajectories for velocity, acceleration, angular velocity, and angular acceleration, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.152, + 0.912, + 0.248 + ], + "angle": 0, + "content": "
MethodADEFDEminADEminFDE\\( \\chi^2 \\) distances
Vel.Acc.Ang. Vel.Ang. Acc.
JTASocial-Trans1.86/2.143.51/4.260.88/0.711.04/0.540.134/0.1690.009/0.0090.011/0.0070.026/0.008
Ours1.68/1.803.34/3.560.79/0.661.02/0.540.100/0.0870.002/0.0030.009/0.0070.016/0.011
JRDBSocial-Trans0.61/0.711.19/1.490.25/0.180.44/0.260.025/0.0770.034/0.0130.021/0.0190.029/0.137
Ours0.56/0.701.16/1.480.23/0.170.44/0.250.024/0.0670.011/0.0040.021/0.0200.019/0.036
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.272, + 0.485, + 0.328 + ], + "angle": 0, + "content": "Table 3. HTP results with momentary observations where only 2 frames of past trajectories are observed. Red and blue indicate the best and second-best for each category. 'Traj.' indicates the input is only past trajectories. 'Det.' denotes the deterministic prediction." + }, + { + "type": "table_caption", + "bbox": [ + 0.132, + 0.339, + 0.439, + 0.35 + ], + "angle": 0, + "content": "(a) HTP with momentary observations on the JTA dataset." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.35, + 0.474, + 0.485 + ], + "angle": 0, + "content": "
MethodADEFDEminADEminFDE
Det.Social-Trans (Traj.)1.573.26--
Social-Trans1.362.80--
Ours1.332.72--
5headsSocial-Trans (Traj.)2.254.620.911.51
Social-Trans2.234.550.921.53
Ours1.943.960.871.30
20headsSocial-Trans (Traj.)2.585.280.710.76
Social-Trans2.465.070.690.73
Ours2.124.470.650.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.126, + 0.494, + 0.443, + 0.506 + ], + "angle": 0, + "content": "(b) HTP with momentary observations on the JRDB dataset." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.511, + 0.474, + 0.646 + ], + "angle": 0, + "content": "
MethodADEFDEminADEminFDE
Det.Social-Trans (Traj.)0.500.94--
Social-Trans0.500.94--
Ours0.440.85--
5headsSocial-Trans (Traj.)0.571.190.250.47
Social-Trans0.581.200.250.47
Ours0.551.120.240.48
20headsSocial-Trans (Traj.)0.681.460.180.30
Social-Trans0.681.460.170.28
Ours0.681.450.180.28
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.688, + 0.254, + 0.702 + ], + "angle": 0, + "content": "5.2.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Supervision by the EmLoco loss. To provide deeper insights into the impact of the EmLoco loss \\(\\mathcal{L}_{\\mathrm{E}}\\) on HTP network training, we train \\(\\mathcal{H}\\) only with \\(\\mathcal{L}_{\\mathrm{E}}\\) i.e., omitting the loss \\(\\mathcal{L}_{\\mathrm{T}}\\) with the ground-truth trajectories. The results are presented in Table 4(B). Importantly, even the proposed method without \\(\\mathcal{L}_{\\mathrm{T}}\\) can improve performance from a randomly initialized network (Table 4(A)). This demonstrates that the supervision provided by \\(\\mathcal{L}_{\\mathrm{E}}\\) contributes to the improvements in HTP performance, while it is designed to be used in conjunction with \\(\\mathcal{L}_{\\mathrm{T}}\\). We emphasize that the combination of \\(\\mathcal{L}_{\\mathrm{T}}\\) and \\(\\mathcal{L}_{\\mathrm{E}}\\) (Table 4(G)) improves HTP performance compared to the baseline [47] (Table 4(C))." + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.272, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Table 4. Ablation studies on JTA [9] and JRDB [34] datasets. Red and blue indicate the best and second-best results for each category. 'initialized' represents the initialized HTP network for reference." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.328, + 0.907, + 0.455 + ], + "angle": 0, + "content": "
MethodJTAJRDB
ADEFDEADEFDE
(A) Initialized5.078.921.311.87
(B) w/o GT loss LT3.528.200.921.55
(C) w/o EmLoco loss LE [47]1.112.260.400.76
(D) w/o Implausible trajectories1.012.050.380.74
(E) w/o Pose input j01.072.240.390.75
(F) w/o Root velocity vroot,00.992.030.380.74
(G) Ours0.971.910.370.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.495, + 0.907, + 0.566 + ], + "angle": 0, + "content": "Table 5. Results of the LocoVal filter for stochastic HTP on the JTA [9] dataset with 9 and 2-frame observations. The results are presented as left/right values, where the left denotes evaluations with 5 heads and the right denotes 20 heads. Red indicates the best result. 'Rejected' refers to the performance of rejected trajectories." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.671 + ], + "angle": 0, + "content": "
Method9 frames2 frames
ADEFDEADEFDE
Social-Trans(w/o LE)w/o Filtering1.86/2.143.51/4.262.23/2.464.55/5.07
w/ Filtering1.81/1.973.49/3.932.06/2.214.27/4.56
Rejected2.37/3.413.75/6.653.83/4.327.16/8.87
Ours(w/ LE)w/o Filtering1.68/1.803.34/3.561.94/2.123.96/4.47
w/ Filtering1.65/1.763.31/3.521.90/2.083.88/4.41
Rejected2.32/2.403.86/4.193.22/3.256.18/6.41
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Inputs into the LocoVal function. We ablate the inputs of \\(\\mathcal{V}\\) as follows: (1) implausible pose-trajectory pairs used in training \\(\\mathcal{V}\\) (Table 4(D)), (2) the human pose \\(j_0\\) at both training and inference (Table 4(E)), (3) and the root velocity \\(\\pmb{v}_{\\mathrm{root},0}\\) at both training and inference (Table 4(F)). These results indicate that not only plausible but also implausible pose-trajectory pairs are required for training \\(\\mathcal{V}\\) to serve as a surrogate for the locomotion generator. Also, both \\(j_0\\) and \\(\\pmb{v}_{\\mathrm{root},0}\\) contribute to performance improvements. Notably, the HTP performance decreases when \\(j_0\\) is not used, demonstrating that the pose is important to estimate the plausibility of predicted trajectories." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12330" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 6. Results of zero-shot filtering by the LocoVal filter on the predictions of a pre-trained trajectory predictor [63] on the ETH/UCY dataset [24, 39]. 'Mean' represents the average performance across the 5 subsets." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.128, + 0.89, + 0.21 + ], + "angle": 0, + "content": "
MethodETHHOTELUNIVZARA1ZARA2Mean
ADEFDEADEFDEADEFDEADEFDEADEFDEADEFDE
Pretrained EqMotion [63]2.184.630.641.311.302.810.821.840.651.471.122.41
Ours (w/ LocoVal filter)1.412.880.611.260.932.040.801.800.641.450.881.89
Rejected trajectories8.8919.722.695.534.339.181.703.672.214.723.968.56
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.233, + 0.336, + 0.248 + ], + "angle": 0, + "content": "5.3. LocoVal Filter at Inference" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.256, + 0.376, + 0.271 + ], + "angle": 0, + "content": "5.3.1. LocoVal Filter on the JTA Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.277, + 0.483, + 0.322 + ], + "angle": 0, + "content": "Since our LocoVal filter can be applied in a plug-and-play manner to any HTP network, we apply it to Social-Trans [47] and our method (i.e., Social-Trans trained with \\(\\mathcal{L}_{\\mathrm{E}}\\))." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.486, + 0.475 + ], + "angle": 0, + "content": "Table 5 presents the filtering results on the JTA dataset [9] with \\(\\lambda\\) set to 0.7 in Eq.(5). We observe that the LocoVal filter enhances performance across all metrics in both Social-Trans and our method. Notably, although \\(\\mathcal{L}_{\\mathrm{E}}\\) using \\(\\nu\\) is used in training \\(\\mathcal{H}\\) in our method, the LocoVal filter using \\(\\nu\\) can further enhance the performance of our method through filtering at inference. Table 5 also shows the evaluation scores of the rejected trajectories. These scores exhibit substantially poorer performance than the filtered results, indicating that the LocoVal filter rejects implausible, inaccurate trajectories." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.486, + 0.453, + 0.501 + ], + "angle": 0, + "content": "5.3.2. Zero-shot Filtering on the ETH/UCY Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.485, + 0.598 + ], + "angle": 0, + "content": "Following the aforementioned experiments on the JTA dataset [9], we further evaluate the LocoVal filter on the ETH/UCY dataset [24, 39] using another baseline, EqMotion [63]. Note that this dataset is not used for training \\(\\nu\\), and due to the bird's-eye-view of the ETH/UCY dataset, we experimented with \\(\\nu\\) that does not use human poses as input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.486, + 0.735 + ], + "angle": 0, + "content": "Table 6 presents the results of applying the LocoVal filter under this zero-shot setting, with \\(\\lambda\\) set to 0.8. Although the dataset and human poses are not used for training \\(\\mathcal{V}\\), we observe improvements in both ADE and FDE with the application of the LocoVal filter across all 5 subsets. While the importance of human poses can be observed in the comparison between Table 4(E) and (G), Table 6 suggests that it is even possible to identify implausible trajectories based only on the current root velocity, even without the human poses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.205, + 0.769 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Analysis of Estimated Plausibility. We examine the plausibility estimated by \\(\\nu\\) and the actual reward in the locomotion generator. To test this, 200 episodes of locomotion are generated from randomly sampled trajectories and initial humanoid states. The correlation coefficient between the estimated plausibilities and the actual rewards is 0.85. This demonstrates that \\(\\nu\\) can serve as a surrogate for the locomotion generator, contributing to performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.909, + 0.37 + ], + "angle": 0, + "content": "Cost-benefit Analysis. The additional training time compared to the baseline [47] is 10 hours in the first stage on a single NVIDIA RTX A6000 GPU. The additional cost in the second stage is nearly negligible because only the computation of \\(\\mathcal{V}\\) is added, which is a shallow MLP. Importantly, \\(\\mathcal{V}\\) can be used with any HTP network, which means only the HTP network needs training in the second stage. Considering these benefits, despite the potential for further cost reductions, our method is highly useful." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.909, + 0.553 + ], + "angle": 0, + "content": "Infeasible Trajectory in Dataset. Some trajectories in the synthetic JTA dataset [9] are infeasible rather than implausible for a humanoid and result in low rewards. However, since our target is plausible locomotion, like most pedestrians' locomotion in daily scenes, it is reasonable not to reproduce such infeasible trajectories. During HTP training, these infeasible trajectories are deprioritized since they have low plausibilities and the HTP network prioritizes more plausible trajectories instead. This weighting mechanism can be one reason for the performance improvement of the proposed method, as shown in Tables 1, 2, and 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.909, + 0.678 + ], + "angle": 0, + "content": "Limitations and Future Work. The smaller performance improvement on the JRDB dataset [34] compared to the JTA dataset [9] (especially in Table 3) is considered to be derived from the assumption of accurate 3D human poses, which is one limitation of the proposed method. In this perspective, future research may extend the EmLoco loss to account for the uncertainty of human poses [54, 57, 68]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.691, + 0.719, + 0.709 + ], + "angle": 0, + "content": "7. Concluding Remarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.908, + 0.822 + ], + "angle": 0, + "content": "This paper presented Locomotion Embodiment, a framework that evaluates the plausibility of predicted trajectories as human locomotion. Our experiments demonstrated that the proposed EmLoco loss improves HTP performance even with the state-of-the-art HTP method [47], without any inference speed deceleration. Furthermore, we established the use of the LocoVal function for trajectory filtering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.822, + 0.909, + 0.899 + ], + "angle": 0, + "content": "Importantly, our EmLoco loss improved HTP performance with limited observations even in the real-world dataset [34], demonstrating its effectiveness in real-world scenarios. We hope our findings inspire future physics-aware HTP methods for safe real-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12331" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Vida Adeli, Mahsa Ehsanpour, Ian D. Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social LSTM: human trajectory prediction in crowded spaces. In CVPR, 2016. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[4] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[6] Erwin Coumans. Bullet physics simulation. In SIGGRAPH, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[7] Markos Diomataris, Nikos Athanasiou, Omid Taheri, Xi Wang, Otmar Hilliges, and Michael J. Black. WANDR: Intention-guided human motion generation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.484, + 0.557 + ], + "angle": 0, + "content": "[8] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R. Qi, Yin Zhou, Zoey Yang, Aurelien Chouard, Pei Sun, Jiquan Ngiam, Vijay Vasudevan, Alexander McCauley, Jonathon Shlens, and Dragomir Anguelov. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[9] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In ECCV, 2018. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[10] Carlos Flores, Pierre Merdrignac, Raoul de Charette, Francisco M. Navas Matos, Vicente Milanés, and Fawzi Nashashibi. A cooperative car-following/emergency braking system with prediction-based pedestrian avoidance capabilities. TITS, 20(5), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[11] Zipeng Fu, Xuxin Cheng, and Deepak Pathak. Deep whole-body control: Learning a unified policy for manipulation and locomotion. In CoRL, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[12] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In ICLR, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[13] Francesco Giuliani, Irtiza Hasan, Marco Cristani, and Fabio Galasso. Transformer networks for trajectory forecasting. In ICPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[14] Ivan Grishchenko, Valentin Bazarevsky, Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Zanfir, Richard Yee, Karthik Raveendran, Matsvei Zhdanovich, Matthias Grundmann, and Cristian Sminchisescu. Blazeposition ghum holistic: Real-time" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.907, + 0.119 + ], + "angle": 0, + "content": "3d human landmarks and pose estimation. arXiv:2206.11678, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.907, + 0.175 + ], + "angle": 0, + "content": "[15] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[16] Xinyang Gu, Yen-Jen Wang, Xiang Zhu, Chengming Shi, Yanjiang Guo, Yichen Liu, and Jianyu Chen. Advancing humanoid locomotion: Mastering challenging terrains with denoising world model learning. In RSS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.907, + 0.288 + ], + "angle": 0, + "content": "[17] Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Back to mlp: A simple baseline for human motion prediction. In WACV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.907, + 0.345 + ], + "angle": 0, + "content": "[18] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social GAN: socially acceptable trajectories with generative adversarial networks. In CVPR, 2018. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.907, + 0.388 + ], + "angle": 0, + "content": "[19] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[20] Dirk Helbing and Péter Molnár. Social force model for pedestrian dynamics. Phys. Rev. E, 51, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.46 + ], + "angle": 0, + "content": "[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[22] Minkwan Kim and Yoonsang Lee. Learning human-like locomotion based on biological actuation and rewards. In SIGGRAPH, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.907, + 0.545 + ], + "angle": 0, + "content": "[23] Parth Kothari, Sven Kreiss, and Alexandre Alahi. Human trajectory forecasting in crowds: A deep learning perspective. TITS, 23(7), 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.907, + 0.587 + ], + "angle": 0, + "content": "[24] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. In Computer graphics forum, 2007. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[25] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yikun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. arXiv:2304.00553, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.907, + 0.7 + ], + "angle": 0, + "content": "[26] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G. Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.907, + 0.743 + ], + "angle": 0, + "content": "[27] Zhengyi Luo, Jinkun Cao, Alexander W. Winkler, Kris Kitani, and Weipeng Xu. Perpetual humanoid control for real-time simulated avatars. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.907, + 0.786 + ], + "angle": 0, + "content": "[28] Takahiro Maeda and Norimichi Ukita. Motionaug: Augmentation with physical correction for human motion prediction. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.907, + 0.828 + ], + "angle": 0, + "content": "[29] Takahiro Maeda and Norimichi Ukita. Fast inference and update of probabilistic density estimation on trajectory prediction. In ICCV, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[30] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: archive of motion capture as surface shapes. In ICCV, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[31] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12332" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.134 + ], + "angle": 0, + "content": "Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac gym: High performance GPU based physics simulation for robot learning. In NeurIPS, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.192 + ], + "angle": 0, + "content": "[32] Karttikeya Mangalam, Ehsan Adeli, Kuan-Hui Lee, Adrien Gaidon, and Juan Carlos Niebles. Disentangling human dynamics for pedestrian locomotion forecasting with noisy supervision. In WACV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.483, + 0.235 + ], + "angle": 0, + "content": "[33] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[34] Roberto Martin-Martin, Mihir Patel, Hamid Rezatofighi, Abhijeet Shenoi, JunYoung Gwak, Eric Frankel, Amir Sadeghian, and Silvio Savarese. JRDB: A dataset and benchmark of egocentric robot visual perception of humans in built environments. TPAMI, 45(6), 2023. 1, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.31, + 0.483, + 0.379 + ], + "angle": 0, + "content": "[35] Raul Quintero Minguez, Ignacio Parra Alonso, David Fernández Llorca, and Miguel Ángel Sotelo. Pedestrian path, pose, and intention prediction through gaussian process dynamical models and pedestrian activity recognition. TITS, 20(5), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.381, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[36] Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In ICML, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.439, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[37] Mai Nishimura, Shohei Nobuhara, and Ko Nishino. Viewbirdformer: Learning to recover ground-plane crowd trajectories and ego-motion from a single ego-centric view. RA-L, 8(1), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[38] Sang II Park, Hyun Joon Shin, and Sung Yong Shin. Online locomotion generation based on motion blending. In SIGGRAPH, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.483, + 0.582 + ], + "angle": 0, + "content": "[39] Stefano Pellegrini, Andreas Ess, and Luc Van Gool. Improving data association by joint modeling of pedestrian trajectories and groupings. In ECCV, 2010. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.584, + 0.483, + 0.638 + ], + "angle": 0, + "content": "[40] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ToG, 40(4), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.642, + 0.483, + 0.684 + ], + "angle": 0, + "content": "[41] Tran Phong, Haoran Wu, Cunjun Yu, Panpan Cai, Sifa Zheng, and David Hsu. What truly matters in trajectory prediction for autonomous driving? In NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.483, + 0.727 + ], + "angle": 0, + "content": "[42] Raul Quintero, Jorge Almeida, David Fernández Llorca, and MA Sotelo. Pedestrian path prediction using body language traits. In IV, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.73, + 0.483, + 0.784 + ], + "angle": 0, + "content": "[43] Ilija Radosavovic, Tete Xiao, Bike Zhang, Trevor Darrell, Jitendra Malik, and Koushil Sreenath. Real-world humanoid locomotion with reinforcement learning. Sci. Robotics, 9(89), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[44] Ilija Radosavovic, Bike Zhang, Baifeng Shi, Jathushan Rajasegaran, Sarthak Kamat, Trevor Darrell, Koushil Sreenath, and Jitendra Malik. Humanoid locomotion as next token prediction. arXiv/2402.19469, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[45] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In CVPR, 2023. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[46] Andrey Rudenko, Luigi Palmieri, Michael Herman, Kris M. Kitani, Dariu M. Gavrila, and Kai Oliver Arras. Human motion trajectory prediction: a survey. Int. J. Robotics Res., 39(8), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[47] Saeed Saadatnejad, Yang Gao, Kaouther Messaoud, and Alexandre Alahi. Social-transmotion: Promptable human trajectory prediction. In ICLR, 2024. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[48] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. Sophie: An attentive gan for predicting paths compliant to social and physical constraints. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[49] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In ECCV, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.309, + 0.905, + 0.363 + ], + "angle": 0, + "content": "[50] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. RA-L, 8(11), 2023. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.366, + 0.905, + 0.394 + ], + "angle": 0, + "content": "[51] Christoph Schöller and Alois Knoll. Flomo: Tractable motion prediction with normalizing flows. In IROS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.396, + 0.905, + 0.436 + ], + "angle": 0, + "content": "[52] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv:1707.06347, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.439, + 0.905, + 0.479 + ], + "angle": 0, + "content": "[53] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.482, + 0.905, + 0.536 + ], + "angle": 0, + "content": "[54] Megh Shukla, Roshan Roy, Pankaj Singh, Shuaib Ahmed, and Alexandre Alahi. VL4pose: Active learning through out-of-distribution detection for pose estimation. In BMVC, 2022, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.581 + ], + "angle": 0, + "content": "[55] Jianhua Sun, Yuxuan Li, Liang Chai, Haoshu Fang, Yong-Lu Li, and Cewu Lu. Human trajectory prediction with momentary observation. In CVPR, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.584, + 0.905, + 0.625 + ], + "angle": 0, + "content": "[56] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.628, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[57] Hiromu Taketsugu and Norimichi Ukita. Active transfer learning for efficient video-specific human pose estimation. In WACV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.712 + ], + "angle": 0, + "content": "[58] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[60] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[61] Jingbo Wang, Zhengyi Luo, Yixuan Li, Ye Yuan, and Bo Dai. Pacer+: On-demand pedestrian animation controller in driving scenarios. In CVPR, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[62] Jungdam Won, Deepak Gopinath, and Jessica K. Hodgins. Physics-based character controllers using conditional vaes. ToG, 41(4), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12333" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[63] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[64] Zhuo Xu, Rui Zhou, Yida Yin, Huidong Gao, Masayoshi Tomizuka, and Jiachen Li. MATRIX: multi-agent trajectory generation with diverse contexts. In ICRA, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[65] Takuma Yagi, Karttikeya Mangalam, Ryo Yonetani, and Yoichi Sato. Future person localization in first-person videos. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.274 + ], + "angle": 0, + "content": "[66] Shuai Yi, Hongsheng Li, and Xiaogang Wang. Understanding pedestrian behaviors from stationary crowd groups. In CVPR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.483, + 0.318 + ], + "angle": 0, + "content": "[67] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.361 + ], + "angle": 0, + "content": "[68] Yufei Zhang, Hanjing Wang, Jeffrey O. Kephart, and Qiang Ji. Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[69] Yufei Zhang, Jeffrey O. Kephart, and Qiang Ji. Incorporating physics principles for precise human motion prediction. In WACV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[70] Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Benjamin Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yuning Chai, Cordelia Schmid, Congcong Li, and Dragomir Anguelov. TNT: target-driven trajectory prediction. In CoRL, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[71] Jingbo Zhao, Zhetao Wang, Yiqin Peng, and Yaojun Wang. Real-time generation of leg animation for walking-in-place techniques. In SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[72] Bolei Zhou, Xiaou Tang, and Xiaogang Wang. Learning collective crowd behaviors with dynamic pedestrian-agents. IJCV, 111(1):50-68, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.604 + ], + "angle": 0, + "content": "[73] Brian D Ziebart, Nathan Ratliff, Garratt Gallagher, Christoph Mertz, Kevin Peterson, J Andrew Bagnell, Martial Hebert, Anind K Dey, and Siddhartha Srinivasa. Planning-based prediction for pedestrians. In IROS, 2009. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12334" + } + ] +] \ No newline at end of file diff --git a/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/88abcf25-51f5-40e1-9600-4a2e38b6cc9a_origin.pdf b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/88abcf25-51f5-40e1-9600-4a2e38b6cc9a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7ce5833915d4602907d21eb71627a65c0b7ed03b --- /dev/null +++ b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/88abcf25-51f5-40e1-9600-4a2e38b6cc9a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d63867e27b542f215c3598e7f3d72fa70311d782f41af03c2d8bef4c02d92e3f +size 2581820 diff --git a/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/full.md b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..28b585aff6672de009d8a909584004b9b6be76c7 --- /dev/null +++ b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/full.md @@ -0,0 +1,342 @@ +# Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment + +Hiromu Taketsugu $^{1}$ Takeru Oba $^{1}$ Takahiro Maeda $^{1}$ Shohei Nobuhara $^{2}$ Norimichi Ukita $^{1}$ + +$^{1}$ Toyota Technological Institute, Japan $^{2}$ Kyoto Institute of Technology, Japan + +{sd23426, sd21502, sd21601}@toyota-ti.ac.jp, nob@kit.ac.jp, ukita@toyota-ti.ac.jp + +# Abstract + +Humans can predict future human trajectories even from momentary observations by using human pose-related cues. However, previous Human Trajectory Prediction (HTP) methods leverage the pose cues implicitly, resulting in implausible predictions. To address this, we propose Locomotion Embodiment, a framework that explicitly evaluates the physical plausibility of the predicted trajectory by locomotion generation under the laws of physics. While the plausibility of locomotion is learned with an indifferentiable physics simulator, it is replaced by our differentiable Locomotion Value function to train an HTP network in a data-driven manner. In particular, our proposed Embodied Locomotion loss is beneficial for efficiently training a stochastic HTP network using multiple heads. Furthermore, the Locomotion Value filter is proposed to filter out implausible trajectories at inference. Experiments demonstrate that our method enhances even the state-of-the-art HTP methods across diverse datasets and problem settings. Our code is available at: https://github.com/ImIntheMiddle/EmLoco. + +# 1. Introduction + +Human Trajectory Prediction (HTP), which predicts future human trajectories, is actively studied [2, 18, 29, 37, 49, 63, 64, 70] due to its importance for applications such as autonomous driving [3, 8, 41, 60, 61, 70], robotics [34, 37, 50, 73], and security systems [10, 66, 72]. These HTP methods assume the availability of sufficiently long past observations and rely solely on a past trajectory. Due to this limitation, these methods cannot handle safety-critical situations in a momentary observation setting, where only a few frames of the past trajectory are available, such as when people suddenly emerge from behind obstacles. + +Regarding this issue, prior work [26, 35, 42, 47, 50] validates that human pose-related cues, such as body orientations and limb motions are beneficial for inferring future trajectories, especially in a momentary observation setting. These methods learn the consistency between human past poses and future trajectories in a data-driven manner alone. + +![](images/b0fb1b8b9167e5f301ade0bd7c15d736ad54e579c65d1900ed1b5bebc01d9b53.jpg) +Figure 1. Overview of our method. Unlike existing methods, which often predict physically implausible trajectories, our framework uses locomotion generation in a physics simulator to incorporate the laws of physics to HTP by training the plausibility score as the consistency between the observed 3D pose and future possible trajectories, which are indicated by $\boldsymbol{j}_0$ and $\tau_f^{1,2,3}$ , respectively, in the figure. Additionally, at inference, our method can evaluate predicted trajectories, $\hat{\tau}_f^{1,2,3}$ , to filter out implausible ones. + +The data-driven approach can prove its ability with a large-scale dataset that covers most possible combinations of poses and trajectories. However, such datasets are unavailable in the literature because human locomotion is diverse. In particular, it is difficult to collect a large number of accurate pose annotations, which are collected by time-consuming delicate motion capture systems [30]. Instead, we propose to fully leverage the beneficial cues of human poses, by assessing the physical consistency between observed poses and future trajectories for physical plausibility-aware HTP. + +For this physical plausibility-aware HTP, we present Locomotion Embodiment, a framework that incorporates the laws of physics into an HTP network. This is achieved by physically simulating the locomotion conditioned by observed poses and future trajectories in a physics simulator [31]; refer to Fig. 1 to see the effect of our framework. As a typical + +example, one of the human poses included in a future possible trajectory $\tau_f^1$ (indicated by $j_{t}$ in Fig. 1) is a little unstably tilted, leading to a lower plausibility score of $\tau_f^1 = 0.3$ . By training the HTP network in a data-driven manner with following the laws of physics, we benefit from the advantages of both the data-driven and physics-based approaches. + +By evaluating and back-propagating the plausibility of locomotion generated in a physics simulator, we can train the physical plausibility-aware HTP network. However, locomotion generation using physics simulators requires various motion parameters, most of which are unavailable or noisy (e.g., joint angular velocities estimated from joint positions provided by a 3D pose estimator) in image-based HTP scenarios. In addition, rigorous physics simulators are indifferent to the physical plausibility, this paper proposes the Locomotion Value (LocoVal) function implemented with a differentiable network to serve as a surrogate for a physics simulator for the plausibility evaluation (Sec. 4.1). Given the plausibility evaluation provided by the LocoVal function, our EmLoco loss penalizes implausible trajectories (Sec. 4.2). + +This EmLoco loss is used in conjunction with the MSE loss for supervised trajectory prediction learning. While the MSE loss behaves as a data term, our EmLoco loss can be a regularization term that incorporates prior knowledge about physics. This is crucial for maintaining the diversity of the predicted trajectories. Although human trajectories are inherently stochastic due to many factors, such as dynamic environments, training with the MSE loss alone forces all predicted trajectories to be aligned with a single ground truth. That is, minimizing the MSE essentially decreases the diversity of the predicted trajectories. + +On the other hand, for stochastically predicting multiple future trajectories, our EmLoco loss allows us to supervise all predicted trajectories to improve their physical plausibility while preserving diversity. In addition, our LocoVal function can also serve as the LocoVal filter that evaluates and filters the candidate trajectories at inference (Sec. 4.3). + +Our framework consists of three key components: + +1. LocoVal function as a surrogate for an indifferentiable physics simulator in HTP network training. +2. EmLoco loss for back-propagating the plausibility of predicted trajectories. This loss is also useful for deterministic predictors but more beneficial for stochastic predictors because of its applicability to all possible predicted trajectories while preserving their stochastic diversity. +3. LocoVal filter for filtering out implausible trajectories at inference. This filter can be applied to any pre-trained HTP network in a plug-and-play manner. + +This work is the first attempt to use simulated humanoid control in a physics simulator for HTP. Since the physics simulator is used only during training, an additional cost (i.e., the cost of the LocoVal function) is negligible at inference. + +# 2. Related Work + +# 2.1. Human Trajectory Prediction (HTP) + +HTP methods can be divided into deterministic [2, 23, 37, 47, 63] and stochastic [18, 29, 49, 50, 64] methods. Deterministic HTP methods [2, 20, 23, 37, 46, 63] predict the most likely trajectory. For example, Social-LSTM uses the Long Short Term Memory [2] to represent the consistency between observed and future trajectories. On the other hand, stochastic HTP methods [12, 49, 50, 64] produce multiple trajectories or probability distributions. To model the inherent stochasticity in HTP, these methods employ probabilistic frameworks such as Generative Adversarial Networks [18, 48] and Diffusion models [15, 33], as well as Normalizing Flows [29, 51] to draw multiple trajectories at inference. Our method is model-agnostic and applicable to both deterministic and stochastic methods. + +# 2.2. HTP with Pose Information + +Prior studies demonstrate that human poses benefit HTP [26, 42, 47, 50, 65]. Yagi et al. [65] and Liang et al. [26] incorporate human poses estimated by CNNs into HTP. Transformer-based methods such as Social-Transmotion [47] and Human Scene Transformer (HST) [50] further improve HTP conditioned by human poses. In particular, HST [50] reports the effectiveness of human pose conditioning for momentary observation scenarios. + +However, since these methods include poses simply as an additional input to the HTP network, predicted future trajectories can be physically inconsistent with past poses. Inspired by this observation, our Locomotion Embodiment framework represents the physical plausibility of predicted trajectories as pose-trajectory consistent locomotion. + +Note that we aim to improve HTP by fully leveraging human poses, distinct from Human Motion Prediction which prioritizes predicting future poses [1, 4, 17, 28, 32, 69]. + +# 2.3. Locomotion Generation + +Locomotion generation is a task to generate walking motions for humanoid robots [16, 43, 44] and virtual characters [22, 38, 71]. Locomotion can be controlled on various conditions such as destinations [7, 19, 56], trajectories [21, 45], and poses [27, 61]. Generative models such as Diffusion models [5, 21, 53, 58, 67] can generate such conditioned locomotion, but they are prone to generate physically implausible locomotion (e.g., skating motions). This is because the laws of physics are not explicitly formulated in generative models. + +On the other hand, physics-based locomotion generation [27, 40, 45, 60-62] follows the laws of physics in simulators [6, 31]. However, a generated locomotion would fail to follow a given trajectory or fall due to kinematic limitations if a humanoid is controlled to walk along an implausible + +trajectory. Therefore, a physics-based locomotion generator can evaluate whether the conditioned trajectory is plausible. Our Locomotion Embodiment framework utilizes this ability to evaluate the plausibility of future trajectories. + +The aforementioned physics-based locomotion generator, e.g., PACER [27, 40, 45, 61], can be trained with reinforcement learning in a physics simulator. A straightforward way to use this locomotion generator for training an HTP network is to directly incorporate the locomotion generator into the HTP network. However, this direct integration has two problems: (1) As mentioned in Sec. 1, the physics simulator is indifferent, preventing training the HTP network by back-propagation. (2) The locomotion generator using the physics simulator requires physical humanoid states, e.g., joint angular velocities, which are unavailable or noisy in image-based HTP scenarios. To address these two problems, our framework uses the physics simulator and the locomotion generator only before training the HTP network and replaces them with the LocoVal function in training the HTP network. + +# 3. Preliminaries + +HTP. An HTP network, $\mathcal{H}$ , takes the past $T_{\mathrm{p}}$ frames as input, including the past trajectory and 3D poses of a target person, and predicts the trajectory of the target person in the future $T_{\mathrm{f}}$ frames. The past trajectory $\pmb{\tau}_{\mathrm{p}} \in \mathbb{R}^{T_{\mathrm{p}} \times 2}$ and the future target trajectory $\pmb{\tau}_{\mathrm{f}} \in \mathbb{R}^{T_{\mathrm{f}} \times 2}$ are on the ground $xy$ plane. The 3D pose involves $J$ joint positions $\pmb{j}_t \in \mathbb{R}^{J \times 3}$ at the past time $t$ , where $(-T_{\mathrm{p}} + 1) \leq t \leq 0$ . + +The predicted future trajectory $\hat{\tau}_{\mathrm{f}}$ is expected to be as close as possible to its ground-truth trajectory $\tau_{\mathrm{f}}$ . + +Locomotion Generation. In our framework, humanoid locomotion is generated to follow any given future trajectory $\pmb{\tau}_{\mathrm{s}} \in \mathbb{R}^{T_{\mathrm{f}} \times 2}$ in a physics simulator [31]. As such a locomotion generator $\mathcal{G}$ , our method employs PACER [45], which is implemented with the policy network [36] trained via reinforcement learning [52]. $\mathcal{G}$ takes $\pmb{\tau}_{\mathrm{s}}$ and the initial humanoid state $h_0$ to control the humanoid joints. $h_0$ involves the positions, rotations, velocities, and angular velocities of all joints. In PACER [45], $\mathcal{G}$ is trained by maximizing the discounted cumulative reward $\Omega$ consisting of a trajectory following reward, a motion style reward [40], and an energy penalty [11]. Refer to PACER [45] for details. + +In PACER, the humanoid movement is constrained by the limits of joint angles and torques, preventing it from producing kinematically impossible locomotion. Due to this limitation, when conditioned on a physically implausible trajectory, the humanoid fails to follow it, resulting in a lower reward. Therefore, we define the discounted cumulative reward $\Omega$ [52] as the plausibility of $\tau_{\mathrm{s}}$ . + +![](images/b132474f5f9792dfc6aa265d95795383915f0cf2e15abb00dd1b1eef279e64e7.jpg) +Figure 2. The overview of the proposed framework. $①$ and $⑤$ mean the weights are trained and fixed, respectively. + +# 4. Physical Plausibility-aware HTP + +Overview. Figure 2 shows the overview of our framework consisting of two training stages and an inference phase. In (a) the first training stage, given the physics-based locomotion generator $\mathcal{G}$ trained as mentioned in Sec. 3, a current humanoid state is fed into it to obtain the reward as the physical plausibility. With this physical plausibility as the ground truth, the LocoVal function $\nu$ is trained. In (b) the second stage, observed pose-related cues in training data are fed into both $\nu$ and $\mathcal{H}$ . $\nu$ also takes the future trajectories predicted by $\mathcal{H}$ to compute the EmLoco losses. The EmLoco loss supports both the MSE and minMSE [18, 50, 63] losses used for deterministic and stochastic predictions, respectively, to train $\mathcal{H}$ . At (c) inference, candidate trajectories predicted by $\mathcal{H}$ are fed into $\nu$ to filter out implausible trajectories. + +# 4.1. LocoVal Function + +The detail of training the LocoVal function $\mathcal{V}$ is illustrated in Fig. 3. As shown in the top half of Fig. 3, the discounted cumulative reward $\Omega$ [52] is used as the plausibility of Embodied Locomotion that is generated through $\mathcal{G}$ using a physics simulator. $\mathcal{V}$ is trained to estimate the reward $\hat{\Omega}$ close to $\Omega$ from only available cues in image-based HTP scenarios because $\mathcal{V}$ serves as a surrogate for $\mathcal{G}$ at inference. + +As such available cues, our method uses the future trajectory $\tau_{\mathrm{s}}$ , the initial pose $j_0$ , and the initial velocity of the root joint $\boldsymbol{v}_{\mathrm{root},0} \in \mathbb{R}^2$ , where $j_0$ and $\boldsymbol{v}_{\mathrm{root},0}$ are in a subset $h_0'$ extracted from $h_0$ . In training $\mathcal{V}$ , these cues are available from the physics simulator. At inference, $\tau_{\mathrm{s}}$ is replaced by $\tau_{\mathrm{f}}$ , $j_0$ can be obtained by a 3D pose estimator [14, 50], and $\boldsymbol{v}_{\mathrm{root},0}$ can be obtained from $\tau_p$ . $\mathcal{V}$ is implemented as a differentiable MLP, taking these cues as input. + +![](images/2f7a4a352c71f865c61e391bc5a80910b1147b2718c4c3e5b1c5717ef2fb9290.jpg) +Figure 3. The overview of training our LocoVal function, $\nu$ . + +![](images/c55a620292924e8da6a607c2aa66c53643b22e615d323f6ea73ea44b4ed6add2.jpg) +Figure 4. The overview of training the HTP network, $\mathcal{H}$ + +Training Process. Unlike PACER [45], which trains $\mathcal{G}$ to obtain high rewards even for implausible trajectories, e.g., dangerously sharp turns, such trajectories should result in low rewards in our framework. This is because humans do not typically walk along such implausible trajectories in daily life. For this reason, we train $\mathcal{G}$ with only plausible pairs of $\tau_{s}$ and $h_0$ to give high rewards only to plausible pose-trajectory pairs. On the other hand, since $\mathcal{V}$ must discriminate between the plausible and implausible pose-trajectory pairs, it is trained with plausible pairs and randomly paired diverse implausible pose-trajectory pairs. + +$\mathcal{V}$ is trained by minimizing the following loss $\mathcal{L}_{\mathcal{V}}$ so that it can estimate the plausibility of the Embodied Locomotion $\varOmega=\mathcal{G}(\pmb{\tau}_{\mathrm{s}},\pmb{h}_{0})$ obtained by rolling out $\mathcal{G}$ : + +$$ +\mathcal {L} _ {\mathcal {V}} = \operatorname {M S E} \left(\mathcal {V} \left(\boldsymbol {\tau} _ {\mathrm {s}}, \boldsymbol {h} _ {0} ^ {\prime}\right), \mathcal {G} \left(\boldsymbol {\tau} _ {\mathrm {s}}, \boldsymbol {h} _ {0}\right)\right), \tag {1} +$$ + +where MSE denotes the Mean Squared Error function. + +# 4.2. HTP Network Training with EmLoco Loss + +Figure 4 illustrates the detail of training the HTP network $\mathcal{H}$ . Our HTP network takes the past information including $\tau_{\mathrm{p}}$ and 3D poses $j_{\mathrm{p}} = \{j_{-T_{\mathrm{p}} + 1},\dots ,j_0\}$ as input to predict the future trajectory $\hat{\tau}_{\mathrm{f}}$ of a target person. Following traditional HTP training methods [2, 47], $\hat{\tau}_{\mathrm{f}}$ is compared with its ground truth $\tau_{\mathrm{f}}$ to back-propagate the MSE loss $\mathcal{L}_{\mathrm{T}} = \mathrm{MSE}(\hat{\tau}_{\mathrm{f}},\tau_{\mathrm{f}})$ . + +EmLoco loss. The EmLoco loss $\mathcal{L}_{\mathrm{E}}$ evaluates predicted trajectories by using $\mathcal{V}$ , which has prior knowledge about plausibility. Note again that a physics simulator is not required in this second stage or the inference phase because $\mathcal{V}$ serves as a surrogate for the plausibility evaluation process. + +Given the prediction of $\mathcal{H}$ , the predicted trajectory $\hat{\tau}_{\mathrm{f}}$ and the subset of $h_0$ (i.e., $h_0^{\prime}$ ) observed in the current frame are fed into $\nu$ to compute $\mathcal{L}_{\mathrm{E}}$ as follows: + +$$ +\mathcal {L} _ {\mathrm {E}} = - \mathcal {V} \left(\hat {\tau} _ {\mathrm {f}}, h _ {0} ^ {\prime}\right). \tag {2} +$$ + +$\mathcal{L}_{\mathrm{E}}$ encourages $\mathcal{H}$ to predict $\hat{\tau}_{\mathrm{f}}$ that increases the plausibility. + +As the total training objective, $\mathcal{L}_{\mathrm{E}}$ is combined with the MSE loss $\mathcal{L}_{\mathrm{T}}$ with a hyper-parameter $\alpha$ as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\mathrm {T}} + \alpha \mathcal {L} _ {\mathrm {E}}, \tag {3} +$$ + +Multi-head HTP network. Our method with the EmLoco loss can be extended to a stochastic HTP network using multiple heads. As also mentioned in Sec. 1, the MSE loss (Eq. (1)) is inappropriate for multiple stochastic predictions because it forces all of them to converge to a unique ground truth. To mitigate this problem, following prior work [18, 50, 63], the multiple heads in our method are trained by replacing $\mathcal{L}_{\mathrm{T}}$ in Eq.(3) with the following objective function: + +$$ +\mathcal {L} _ {\mathrm {T}} = \min _ {k} \left(\operatorname {M S E} \left(\hat {\boldsymbol {\tau}} _ {\mathrm {f}} ^ {k}, \boldsymbol {\tau} _ {\mathrm {f}}\right)\right), \tag {4} +$$ + +where $\hat{\tau}_{\mathrm{f}}^{k}$ denotes the prediction of $k$ -th head. In Eq. (4), the minMSE loss supervises only one prediction closest to its ground truth $\tau_{\mathrm{f}}$ among multiple predictions. + +Unlike the MSE loss, this minMSE loss allows us to maintain diverse predictions. However, the minMSE loss is inefficient. What is even worse is that several heads may be under-fitted if other heads are trained in early training iterations so that these heads are always selected as those minimizing the MSE in Eq. (4) in later iterations. + +On the other hand, our EmLoco loss does not require the unique ground truth, i.e., $\tau_{\mathrm{f}}$ in Eq. (4). As a regularization term, the EmLoco loss trains all the heads to predict plausible trajectories. This simultaneous optimization allows us to avoid the under-fitting problem mentioned above. In our method, this simultaneous optimization is implemented by averaging the EmLoco losses using $\mathcal{V}$ for all predicted trajectories, and this average is used as $\mathcal{L}_{\mathrm{E}}$ in Eq.(3). + +# 4.3. LocoVal Filter at Inference + +The trained LocoVal function can also be applied to trajectory filtering, called the LocoVal filter, at inference using multi-head HTP networks. Similar to the usage of $\mathcal{V}$ for training multiple heads simultaneously at training, it computes the plausibilities of all predicted trajectories at inference. Each of the predicted trajectories with a plausibility lower than a threshold $\lambda$ is filtered out as follows: + +$$ +\left\{ \begin{array}{l l} \left\{\hat {\boldsymbol {\tau}} _ {\mathrm {f}} ^ {k} \mid \mathcal {V} \left(\hat {\boldsymbol {\tau}} _ {\mathrm {f}} ^ {k}, \boldsymbol {h} _ {0} ^ {\prime}\right) \geq \lambda \right\} & \text {i f} \exists k, \mathcal {V} \left(\hat {\boldsymbol {\tau}} _ {\mathrm {f}} ^ {k}, \boldsymbol {h} _ {0} ^ {\prime}\right) \geq \lambda , \\ \operatorname {a r g m a x} _ {\hat {\boldsymbol {\tau}} _ {\mathrm {f}} ^ {k}} \mathcal {V} \left(\hat {\boldsymbol {\tau}} _ {\mathrm {f}} ^ {k}, \boldsymbol {h} _ {0} ^ {\prime}\right) & \text {o t h e r w i s e .} \end{array} \right. \tag {5} +$$ + +Note that the LocoVal filter can be used with any stochastic HTP networks in a plug-and-play manner. + +# 5. Experiments + +In Sec. 5.2.1, we benchmark our HTP network trained with the EmLoco loss under the standard HTP setting. Section 5.2.2 presents results for multi-head HTP networks, and Sec. 5.2.3 examines the performance with momentary observations. Then, we report ablation results in Sec. 5.2.4. In Sec. 5.3.1 and Sec. 5.3.2, we evaluate the effectiveness of the LocoVal filter in various scenarios. + +# 5.1. Experimental Settings + +Datasets. As trajectory datasets, the Joint Track Auto (JTA) dataset [9], the JackRabbit Dataset and Benchmark (JRDB) [34], and the ETH/UCY dataset [24, 39] are used. For the JRDB dataset, a subset provided by the SocialTrans [47] was used. The ETH/UCY dataset is used only in Sec. 5.3.2 to evaluate the performance of the LocoVal filter on pre-trained models. + +While JTA also provides 3D human poses synchronized with their trajectories, no 3D pose is annotated in JRDB. Instead, in JRDB, our experiments use 3D poses estimated from images and 2D pose annotations by using [14, 50]. While the estimated 3D poses are provided by HST [50], misestimated poses are filtered out in our experiments; see the Supp. for the details. All the 3D poses mentioned above are converted to the SMPL-style poses [25] for a fair comparison with the baseline. + +While the 3D poses described above are used in the second stage to train the HTP network $\mathcal{H}$ , more accurate 3D poses are expected to train $\mathcal{G}$ and $\mathcal{V}$ to learn the physical plausibility of 3D poses. Such accurate 3D poses are provided by the Archive of Motion Capture as Surface Shapes (AMASS) [30]. While accurate 3D poses are available and can be used to compute the trajectories in motion capture datasets such as AMASS, motion capture data do not scale compared with trajectory datasets, as described in Sec. 1. + +To avoid this problem, 3D poses in AMASS are combined with trajectories in the trajectory datasets, i.e., JTA and JRDB, as follows. Plausible and implausible pose-trajectory pairs are required to train $\mathcal{G}$ and $\mathcal{V}$ , as described in Sec. 4.1. The plausible pairs are prepared by spatially aligning the 3D pose and the trajectories, which are randomly sampled from AMASS and the trajectory datasets, respectively. This spatial alignment consists of the orientation and velocity alignments. Given the heading vector of each pedestrian in these two datasets, the heading orientations are aligned between the motion capture pose from AMASS and the trajectory data. For the velocity alignment, on the other hand, the trajectory data is expanded or contracted to make the vector norms equal between the motion capture pose and the trajectory data. Refer to the Supp. for detailed training conditions on the aforementioned datasets. + +HTP networks. Since multiple people are observed in each frame of the aforementioned datasets, all baseline HTP networks of our method take not only the past trajectory of a target person but also those of other people as input. This allows us to predict the target trajectory affected by the dynamic environment. While our method is introduced to take only the target person as input, all people in the scene are fed into $\mathcal{H}$ following the previous methods. In the evaluations with the JTA and JRDB datasets, Social-Transmotion (Social-Trans), is used as our baseline without any architectural changes. We extend this baseline of the trajectory predictor $\mathcal{H}$ by plugging the EmLoco loss into the training loss. Following the implementation of Social-Trans, past trajectories, 2D and 3D human poses, and 2D and 3D bounding boxes are fed into $\mathcal{H}$ in the JTA dataset [9], while the past trajectories, 2D bounding boxes, and 3D human poses are used in the JRDB dataset. Note that the 3D human poses are estimated for the JRDB dataset [34] as mentioned earlier. For the stochastic version of our method, we modify Social-Trans [47] by parallelizing the prediction heads with different initial weights. For comparison, all HTP networks predict the future $T_{\mathrm{f}} = 12$ frames at 2.5 fps on the JTA and JRDB datasets. The number of past frames differs depending on the experiment; e.g., 9 frames in Sec. 5.2.1 and Sec. 5.2.2, and 2 frames as momentary observations [55] in Sec. 5.2.3. + +In experiments on the ETH/UCY dataset [24, 39], where a bird's-eye view hinders the use of 3D poses, only past trajectories are fed into $\mathcal{H}$ . Here, we employ EqMotion [63], a state-of-the-art stochastic HTP method that only uses past trajectories, as a pre-trained HTP network. Each past trajectory consists of $T_{\mathrm{p}} = 8$ frames, while each future trajectory remains at $T_{\mathrm{f}} = 12$ frames at 2.5 fps. + +Evaluation. Two major metrics, the Average Displacement Error (ADE) and the Final Displacement Error (FDE), both measured in meters, evaluate the HTP results. Consistent with the other HTP literature [29, 49, 50, 63], in evaluating the stochastic HTP networks, we report minADE and minFDE among multiple trajectories. Also, following MATRIX [64], we report $\chi^2$ distances for the four physics primitives of each predicted trajectory, namely velocities, accelerations, angular velocities, and angular accelerations, representing the plausibility of the predicted trajectories. + +# 5.2. HTP Training with the EmLoco Loss + +Here, we evaluate our HTP network trained with the proposed EmLoco loss. Note that the LocoVal filter is not applied in these experiments. $\alpha$ in Eq.(3) is set to 100. + +# 5.2.1. Comparison with Existing HTP Methods + +Following the standard protocol [2, 12, 18, 47, 63] using 9 past frames, where past trajectories are sufficiently long, we compare the accuracy of deterministic predictions by our + +Table 1. HTP results under the standard setting. Red and blue indicate the best and second-best. 'Social-Trans' and 'Ours' utilize pose input. For methods other than 'Social-Trans' and 'Ours', we report the results from the Social-Trans literature [47]. + +
MethodJTA [9]JRDB [34]
ADEFDEADEFDE
Social-GAN-det [18]1.663.760.500.99
Transformer [13]1.563.540.561.10
Vanilla-LSTM [2]1.443.250.420.83
Occupancy-LSTM [2]1.413.150.430.85
Directional-LSTM [23]1.373.060.450.87
Dir-social-LSTM [23]1.232.590.480.95
Social-LSTM [2]1.212.540.470.95
Autobots [12]1.202.700.390.80
Trajectory++ [49]1.182.530.400.78
EqMotion [63]1.132.390.400.77
Social-Trans [47]1.112.260.400.76
Ours0.971.910.370.72
+ +HTP network with other HTP networks [2, 12, 18, 23, 47, 49, 59, 63]. Table 1 shows that our method outperforms all other methods including the state-of-the-art baseline, Social-Trans [47], on both the JTA [9] and the JRDB [34] datasets. This demonstrates that the EmLoco loss effectively provides the prior knowledge of the trajectory-pose consistency during training. Notably, the performance improvements on the JRDB dataset [34] with imperfect 3D poses suggest the applicability of our EmLoco loss to real-world applications. + +# 5.2.2. Multi-head HTP Networks + +To demonstrate the effectiveness of the EmLoco loss on a multi-head HTP network, 5 and 20 prediction heads are implemented on Social-Trans [47]. As described in Sec. 4.2, our proposed method is expected to improve the plausibility of predicted trajectories closer to human locomotion. From this perspective, we evaluate the plausibility of the predicted trajectories by the $\chi^2$ distance employed in MATRIX [64]. + +Table 2 presents the comparisons with the baseline. Our method achieves the best values across almost all metrics on both the JTA [9] and JRDB [34] datasets. Particularly, the improvements in ADE and FDE reveal that all heads are trained properly because these metrics are evaluated with the prediction of all heads. Also, improvements in $\chi^2$ distances, the plausibility regarding physics primitives, demonstrate that $\mathcal{H}$ trained with our EmLoco loss predicts trajectories similar to actual human locomotion than the baseline [47]. + +Figure 5 shows qualitative results comparing stochastic HTP with 5 heads for both Social-Trans [47] and our method. While the Social-Trans trained solely with minMSE can predict diverse trajectories, many of them are implausible, including sharp turns. In contrast, the predictions of our HTP network maintain diversity across all heads while remaining plausible, resulting in trajectories closer to the ground truth. + +![](images/e545968f64e3b6e38a64bf326a3604d1be02541b6ead9ec3a8aa856f56dad392.jpg) +Figure 5. Comparison of the stochastic HTP between Social-Trans [47] and ours with 5 heads on the JTA dataset [9]. Human poses are shown at a doubled scale for visualization purposes. + +![](images/b42ac27e3786c76a1711e01516b66deb639db7991942c5be018ce7758c758df7.jpg) +Figure 6. Comparison of the prediction results between the baseline Social-Trans [47] and our method with momentary observations on the JTA dataset [9]. The observed pose are displayed at a doubled scale for visualization purposes. + +# 5.2.3. HTP with Momentary Observations + +As noted in Sec. 1, existing HTP methods struggle with momentary observations. On the contrary, because the proposed EmLoco loss encourages physical plausibility-aware HTP, $\mathcal{H}$ trained with the EmLoco loss is expected to be robust against this setting. From this point of view, we experiment with 2 past frames following the prior work [55]. + +The results are shown in Table 3. Consistent with Tables 1 and 2, the EmLoco loss allows our method to outperform the baseline [47] across almost all categories. In addition, Fig. 6 presents qualitative results in two cases: when a pedestrian is (a) going straight and (b) turning. In both cases, (i) Social-Trans [47] with trajectory-only input results in significant prediction errors, and (ii) Social-Trans with pose input still predicts implausible trajectories, especially during sharp turns. In contrast, (iii) our method provides plausible predictions that closely follow the ground-truth trajectory and align with the observed pose. This demonstrates that the proposed method can effectively leverage human poses and achieve plausibility-aware HTP, thanks to the physics-based LocoVal function training. + +Table 2. Comparison of our HTP network using the EmLoco loss with baseline [47] in the stochastic HTP under 9 frames of observations. Red indicates the best result for each category. The results are presented as left/right values, where the left denotes evaluations with 5 heads and the right denotes 20 heads. 'Vel,' 'Acc,' 'Ang. Vel,' and 'Ang. Acc.' represent the $\chi^2$ distances between the distributions of the predicted and ground-truth trajectories for velocity, acceleration, angular velocity, and angular acceleration, respectively. + +
MethodADEFDEminADEminFDE\( \chi^2 \) distances
Vel.Acc.Ang. Vel.Ang. Acc.
JTASocial-Trans1.86/2.143.51/4.260.88/0.711.04/0.540.134/0.1690.009/0.0090.011/0.0070.026/0.008
Ours1.68/1.803.34/3.560.79/0.661.02/0.540.100/0.0870.002/0.0030.009/0.0070.016/0.011
JRDBSocial-Trans0.61/0.711.19/1.490.25/0.180.44/0.260.025/0.0770.034/0.0130.021/0.0190.029/0.137
Ours0.56/0.701.16/1.480.23/0.170.44/0.250.024/0.0670.011/0.0040.021/0.0200.019/0.036
+ +Table 3. HTP results with momentary observations where only 2 frames of past trajectories are observed. Red and blue indicate the best and second-best for each category. 'Traj.' indicates the input is only past trajectories. 'Det.' denotes the deterministic prediction. +(a) HTP with momentary observations on the JTA dataset. + +
MethodADEFDEminADEminFDE
Det.Social-Trans (Traj.)1.573.26--
Social-Trans1.362.80--
Ours1.332.72--
5headsSocial-Trans (Traj.)2.254.620.911.51
Social-Trans2.234.550.921.53
Ours1.943.960.871.30
20headsSocial-Trans (Traj.)2.585.280.710.76
Social-Trans2.465.070.690.73
Ours2.124.470.650.68
+ +(b) HTP with momentary observations on the JRDB dataset. + +
MethodADEFDEminADEminFDE
Det.Social-Trans (Traj.)0.500.94--
Social-Trans0.500.94--
Ours0.440.85--
5headsSocial-Trans (Traj.)0.571.190.250.47
Social-Trans0.581.200.250.47
Ours0.551.120.240.48
20headsSocial-Trans (Traj.)0.681.460.180.30
Social-Trans0.681.460.170.28
Ours0.681.450.180.28
+ +# 5.2.4. Ablation Studies + +Supervision by the EmLoco loss. To provide deeper insights into the impact of the EmLoco loss $\mathcal{L}_{\mathrm{E}}$ on HTP network training, we train $\mathcal{H}$ only with $\mathcal{L}_{\mathrm{E}}$ i.e., omitting the loss $\mathcal{L}_{\mathrm{T}}$ with the ground-truth trajectories. The results are presented in Table 4(B). Importantly, even the proposed method without $\mathcal{L}_{\mathrm{T}}$ can improve performance from a randomly initialized network (Table 4(A)). This demonstrates that the supervision provided by $\mathcal{L}_{\mathrm{E}}$ contributes to the improvements in HTP performance, while it is designed to be used in conjunction with $\mathcal{L}_{\mathrm{T}}$ . We emphasize that the combination of $\mathcal{L}_{\mathrm{T}}$ and $\mathcal{L}_{\mathrm{E}}$ (Table 4(G)) improves HTP performance compared to the baseline [47] (Table 4(C)). + +Table 4. Ablation studies on JTA [9] and JRDB [34] datasets. Red and blue indicate the best and second-best results for each category. 'initialized' represents the initialized HTP network for reference. + +
MethodJTAJRDB
ADEFDEADEFDE
(A) Initialized5.078.921.311.87
(B) w/o GT loss LT3.528.200.921.55
(C) w/o EmLoco loss LE [47]1.112.260.400.76
(D) w/o Implausible trajectories1.012.050.380.74
(E) w/o Pose input j01.072.240.390.75
(F) w/o Root velocity vroot,00.992.030.380.74
(G) Ours0.971.910.370.72
+ +Table 5. Results of the LocoVal filter for stochastic HTP on the JTA [9] dataset with 9 and 2-frame observations. The results are presented as left/right values, where the left denotes evaluations with 5 heads and the right denotes 20 heads. Red indicates the best result. 'Rejected' refers to the performance of rejected trajectories. + +
Method9 frames2 frames
ADEFDEADEFDE
Social-Trans(w/o LE)w/o Filtering1.86/2.143.51/4.262.23/2.464.55/5.07
w/ Filtering1.81/1.973.49/3.932.06/2.214.27/4.56
Rejected2.37/3.413.75/6.653.83/4.327.16/8.87
Ours(w/ LE)w/o Filtering1.68/1.803.34/3.561.94/2.123.96/4.47
w/ Filtering1.65/1.763.31/3.521.90/2.083.88/4.41
Rejected2.32/2.403.86/4.193.22/3.256.18/6.41
+ +Inputs into the LocoVal function. We ablate the inputs of $\mathcal{V}$ as follows: (1) implausible pose-trajectory pairs used in training $\mathcal{V}$ (Table 4(D)), (2) the human pose $j_0$ at both training and inference (Table 4(E)), (3) and the root velocity $\pmb{v}_{\mathrm{root},0}$ at both training and inference (Table 4(F)). These results indicate that not only plausible but also implausible pose-trajectory pairs are required for training $\mathcal{V}$ to serve as a surrogate for the locomotion generator. Also, both $j_0$ and $\pmb{v}_{\mathrm{root},0}$ contribute to performance improvements. Notably, the HTP performance decreases when $j_0$ is not used, demonstrating that the pose is important to estimate the plausibility of predicted trajectories. + +Table 6. Results of zero-shot filtering by the LocoVal filter on the predictions of a pre-trained trajectory predictor [63] on the ETH/UCY dataset [24, 39]. 'Mean' represents the average performance across the 5 subsets. + +
MethodETHHOTELUNIVZARA1ZARA2Mean
ADEFDEADEFDEADEFDEADEFDEADEFDEADEFDE
Pretrained EqMotion [63]2.184.630.641.311.302.810.821.840.651.471.122.41
Ours (w/ LocoVal filter)1.412.880.611.260.932.040.801.800.641.450.881.89
Rejected trajectories8.8919.722.695.534.339.181.703.672.214.723.968.56
+ +# 5.3. LocoVal Filter at Inference + +# 5.3.1. LocoVal Filter on the JTA Dataset + +Since our LocoVal filter can be applied in a plug-and-play manner to any HTP network, we apply it to Social-Trans [47] and our method (i.e., Social-Trans trained with $\mathcal{L}_{\mathrm{E}}$ ). + +Table 5 presents the filtering results on the JTA dataset [9] with $\lambda$ set to 0.7 in Eq.(5). We observe that the LocoVal filter enhances performance across all metrics in both Social-Trans and our method. Notably, although $\mathcal{L}_{\mathrm{E}}$ using $\nu$ is used in training $\mathcal{H}$ in our method, the LocoVal filter using $\nu$ can further enhance the performance of our method through filtering at inference. Table 5 also shows the evaluation scores of the rejected trajectories. These scores exhibit substantially poorer performance than the filtered results, indicating that the LocoVal filter rejects implausible, inaccurate trajectories. + +# 5.3.2. Zero-shot Filtering on the ETH/UCY Dataset + +Following the aforementioned experiments on the JTA dataset [9], we further evaluate the LocoVal filter on the ETH/UCY dataset [24, 39] using another baseline, EqMotion [63]. Note that this dataset is not used for training $\nu$ , and due to the bird's-eye-view of the ETH/UCY dataset, we experimented with $\nu$ that does not use human poses as input. + +Table 6 presents the results of applying the LocoVal filter under this zero-shot setting, with $\lambda$ set to 0.8. Although the dataset and human poses are not used for training $\mathcal{V}$ , we observe improvements in both ADE and FDE with the application of the LocoVal filter across all 5 subsets. While the importance of human poses can be observed in the comparison between Table 4(E) and (G), Table 6 suggests that it is even possible to identify implausible trajectories based only on the current root velocity, even without the human poses. + +# 6. Discussion + +Analysis of Estimated Plausibility. We examine the plausibility estimated by $\nu$ and the actual reward in the locomotion generator. To test this, 200 episodes of locomotion are generated from randomly sampled trajectories and initial humanoid states. The correlation coefficient between the estimated plausibilities and the actual rewards is 0.85. This demonstrates that $\nu$ can serve as a surrogate for the locomotion generator, contributing to performance improvements. + +Cost-benefit Analysis. The additional training time compared to the baseline [47] is 10 hours in the first stage on a single NVIDIA RTX A6000 GPU. The additional cost in the second stage is nearly negligible because only the computation of $\mathcal{V}$ is added, which is a shallow MLP. Importantly, $\mathcal{V}$ can be used with any HTP network, which means only the HTP network needs training in the second stage. Considering these benefits, despite the potential for further cost reductions, our method is highly useful. + +Infeasible Trajectory in Dataset. Some trajectories in the synthetic JTA dataset [9] are infeasible rather than implausible for a humanoid and result in low rewards. However, since our target is plausible locomotion, like most pedestrians' locomotion in daily scenes, it is reasonable not to reproduce such infeasible trajectories. During HTP training, these infeasible trajectories are deprioritized since they have low plausibilities and the HTP network prioritizes more plausible trajectories instead. This weighting mechanism can be one reason for the performance improvement of the proposed method, as shown in Tables 1, 2, and 3. + +Limitations and Future Work. The smaller performance improvement on the JRDB dataset [34] compared to the JTA dataset [9] (especially in Table 3) is considered to be derived from the assumption of accurate 3D human poses, which is one limitation of the proposed method. In this perspective, future research may extend the EmLoco loss to account for the uncertainty of human poses [54, 57, 68]. + +# 7. Concluding Remarks + +This paper presented Locomotion Embodiment, a framework that evaluates the plausibility of predicted trajectories as human locomotion. Our experiments demonstrated that the proposed EmLoco loss improves HTP performance even with the state-of-the-art HTP method [47], without any inference speed deceleration. Furthermore, we established the use of the LocoVal function for trajectory filtering. + +Importantly, our EmLoco loss improved HTP performance with limited observations even in the real-world dataset [34], demonstrating its effectiveness in real-world scenarios. We hope our findings inspire future physics-aware HTP methods for safe real-world applications. + +# References + +[1] Vida Adeli, Mahsa Ehsanpour, Ian D. Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In ICCV, 2021. 2 +[2] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social LSTM: human trajectory prediction in crowded spaces. In CVPR, 2016. 1, 2, 4, 5, 6 +[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1 +[4] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In ECCV, 2020. 2 +[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2 +[6] Erwin Coumans. Bullet physics simulation. In SIGGRAPH, 2015. 2 +[7] Markos Diomataris, Nikos Athanasiou, Omid Taheri, Xi Wang, Otmar Hilliges, and Michael J. Black. WANDR: Intention-guided human motion generation. In CVPR, 2024. 2 +[8] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R. Qi, Yin Zhou, Zoey Yang, Aurelien Chouard, Pei Sun, Jiquan Ngiam, Vijay Vasudevan, Alexander McCauley, Jonathon Shlens, and Dragomir Anguelov. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In ICCV, 2021. 1 +[9] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In ECCV, 2018. 5, 6, 7, 8 +[10] Carlos Flores, Pierre Merdrignac, Raoul de Charette, Francisco M. Navas Matos, Vicente Milanés, and Fawzi Nashashibi. A cooperative car-following/emergency braking system with prediction-based pedestrian avoidance capabilities. TITS, 20(5), 2019. 1 +[11] Zipeng Fu, Xuxin Cheng, and Deepak Pathak. Deep whole-body control: Learning a unified policy for manipulation and locomotion. In CoRL, 2022. 3 +[12] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In ICLR, 2022. 2, 5, 6 +[13] Francesco Giuliani, Irtiza Hasan, Marco Cristani, and Fabio Galasso. Transformer networks for trajectory forecasting. In ICPR, 2021. 6 +[14] Ivan Grishchenko, Valentin Bazarevsky, Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Zanfir, Richard Yee, Karthik Raveendran, Matsvei Zhdanovich, Matthias Grundmann, and Cristian Sminchisescu. Blazeposition ghum holistic: Real-time + +3d human landmarks and pose estimation. arXiv:2206.11678, 2022. 3, 5 +[15] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In CVPR, 2022. 2 +[16] Xinyang Gu, Yen-Jen Wang, Xiang Zhu, Chengming Shi, Yanjiang Guo, Yichen Liu, and Jianyu Chen. Advancing humanoid locomotion: Mastering challenging terrains with denoising world model learning. In RSS, 2024. 2 +[17] Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Back to mlp: A simple baseline for human motion prediction. In WACV, 2023. 2 +[18] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social GAN: socially acceptable trajectories with generative adversarial networks. In CVPR, 2018. 1, 2, 3, 4, 5, 6 +[19] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2 +[20] Dirk Helbing and Péter Molnár. Social force model for pedestrian dynamics. Phys. Rev. E, 51, 1995. 2 +[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 2 +[22] Minkwan Kim and Yoonsang Lee. Learning human-like locomotion based on biological actuation and rewards. In SIGGRAPH, 2023. 2 +[23] Parth Kothari, Sven Kreiss, and Alexandre Alahi. Human trajectory forecasting in crowds: A deep learning perspective. TITS, 23(7), 2021. 2, 6 +[24] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. In Computer graphics forum, 2007. 5, 8 +[25] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yikun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. arXiv:2304.00553, 2023. 5 +[26] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G. Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In CVPR, 2019. 1, 2 +[27] Zhengyi Luo, Jinkun Cao, Alexander W. Winkler, Kris Kitani, and Weipeng Xu. Perpetual humanoid control for real-time simulated avatars. In ICCV, 2023. 2, 3 +[28] Takahiro Maeda and Norimichi Ukita. Motionaug: Augmentation with physical correction for human motion prediction. In CVPR, 2022. 2 +[29] Takahiro Maeda and Norimichi Ukita. Fast inference and update of probabilistic density estimation on trajectory prediction. In ICCV, 2023. 1, 2, 5 +[30] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: archive of motion capture as surface shapes. In ICCV, 2019. 1, 5 +[31] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, + +Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac gym: High performance GPU based physics simulation for robot learning. In NeurIPS, 2021. 1, 2, 3 +[32] Karttikeya Mangalam, Ehsan Adeli, Kuan-Hui Lee, Adrien Gaidon, and Juan Carlos Niebles. Disentangling human dynamics for pedestrian locomotion forecasting with noisy supervision. In WACV, 2020. 2 +[33] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In CVPR, 2023. 2 +[34] Roberto Martin-Martin, Mihir Patel, Hamid Rezatofighi, Abhijeet Shenoi, JunYoung Gwak, Eric Frankel, Amir Sadeghian, and Silvio Savarese. JRDB: A dataset and benchmark of egocentric robot visual perception of humans in built environments. TPAMI, 45(6), 2023. 1, 5, 6, 7, 8 +[35] Raul Quintero Minguez, Ignacio Parra Alonso, David Fernández Llorca, and Miguel Ángel Sotelo. Pedestrian path, pose, and intention prediction through gaussian process dynamical models and pedestrian activity recognition. TITS, 20(5), 2019. 1 +[36] Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In ICML, 2016. 3 +[37] Mai Nishimura, Shohei Nobuhara, and Ko Nishino. Viewbirdformer: Learning to recover ground-plane crowd trajectories and ego-motion from a single ego-centric view. RA-L, 8(1), 2023. 1, 2 +[38] Sang II Park, Hyun Joon Shin, and Sung Yong Shin. Online locomotion generation based on motion blending. In SIGGRAPH, 2002. 2 +[39] Stefano Pellegrini, Andreas Ess, and Luc Van Gool. Improving data association by joint modeling of pedestrian trajectories and groupings. In ECCV, 2010. 5, 8 +[40] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ToG, 40(4), 2021. 2, 3 +[41] Tran Phong, Haoran Wu, Cunjun Yu, Panpan Cai, Sifa Zheng, and David Hsu. What truly matters in trajectory prediction for autonomous driving? In NeurIPS, 2023. 1 +[42] Raul Quintero, Jorge Almeida, David Fernández Llorca, and MA Sotelo. Pedestrian path prediction using body language traits. In IV, 2014. 1, 2 +[43] Ilija Radosavovic, Tete Xiao, Bike Zhang, Trevor Darrell, Jitendra Malik, and Koushil Sreenath. Real-world humanoid locomotion with reinforcement learning. Sci. Robotics, 9(89), 2024. 2 +[44] Ilija Radosavovic, Bike Zhang, Baifeng Shi, Jathushan Rajasegaran, Sarthak Kamat, Trevor Darrell, Koushil Sreenath, and Jitendra Malik. Humanoid locomotion as next token prediction. arXiv/2402.19469, 2024. 2 +[45] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In CVPR, 2023. 2, 3, 4 + +[46] Andrey Rudenko, Luigi Palmieri, Michael Herman, Kris M. Kitani, Dariu M. Gavrila, and Kai Oliver Arras. Human motion trajectory prediction: a survey. Int. J. Robotics Res., 39(8), 2020. 2 +[47] Saeed Saadatnejad, Yang Gao, Kaouther Messaoud, and Alexandre Alahi. Social-transmotion: Promptable human trajectory prediction. In ICLR, 2024. 1, 2, 4, 5, 6, 7, 8 +[48] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. Sophie: An attentive gan for predicting paths compliant to social and physical constraints. In CVPR, 2019. 2 +[49] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In ECCV, 2020. 1, 2, 5, 6 +[50] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. RA-L, 8(11), 2023. 1, 2, 3, 4, 5 +[51] Christoph Schöller and Alois Knoll. Flomo: Tractable motion prediction with normalizing flows. In IROS, 2021. 2 +[52] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv:1707.06347, 2017. 3 +[53] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 2 +[54] Megh Shukla, Roshan Roy, Pankaj Singh, Shuaib Ahmed, and Alexandre Alahi. VL4pose: Active learning through out-of-distribution detection for pose estimation. In BMVC, 2022, 8 +[55] Jianhua Sun, Yuxuan Li, Liang Chai, Haoshu Fang, Yong-Lu Li, and Cewu Lu. Human trajectory prediction with momentary observation. In CVPR, 2022. 5, 6 +[56] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 2 +[57] Hiromu Taketsugu and Norimichi Ukita. Active transfer learning for efficient video-specific human pose estimation. In WACV, 2024. 8 +[58] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 2 +[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6 +[60] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In ICCV, 2023. 1, 2 +[61] Jingbo Wang, Zhengyi Luo, Yixuan Li, Ye Yuan, and Bo Dai. Pacer+: On-demand pedestrian animation controller in driving scenarios. In CVPR, 2024. 1, 2, 3 +[62] Jungdam Won, Deepak Gopinath, and Jessica K. Hodgins. Physics-based character controllers using conditional vaes. ToG, 41(4), 2022. 2 + +[63] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8 +[64] Zhuo Xu, Rui Zhou, Yida Yin, Huidong Gao, Masayoshi Tomizuka, and Jiachen Li. MATRIX: multi-agent trajectory generation with diverse contexts. In ICRA, 2024. 1, 2, 5, 6 +[65] Takuma Yagi, Karttikeya Mangalam, Ryo Yonetani, and Yoichi Sato. Future person localization in first-person videos. In CVPR, 2018. 2 +[66] Shuai Yi, Hongsheng Li, and Xiaogang Wang. Understanding pedestrian behaviors from stationary crowd groups. In CVPR, 2015. 1 +[67] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, 2023. 2 +[68] Yufei Zhang, Hanjing Wang, Jeffrey O. Kephart, and Qiang Ji. Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In ICCV, 2023. 8 +[69] Yufei Zhang, Jeffrey O. Kephart, and Qiang Ji. Incorporating physics principles for precise human motion prediction. In WACV, 2024. 2 +[70] Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Benjamin Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yuning Chai, Cordelia Schmid, Congcong Li, and Dragomir Anguelov. TNT: target-driven trajectory prediction. In CoRL, 2020. 1 +[71] Jingbo Zhao, Zhetao Wang, Yiqin Peng, and Yaojun Wang. Real-time generation of leg animation for walking-in-place techniques. In SIGGRAPH, 2022. 2 +[72] Bolei Zhou, Xiaou Tang, and Xiaogang Wang. Learning collective crowd behaviors with dynamic pedestrian-agents. IJCV, 111(1):50-68, 2015. 1 +[73] Brian D Ziebart, Nathan Ratliff, Garratt Gallagher, Christoph Mertz, Kevin Peterson, J Andrew Bagnell, Martial Hebert, Anind K Dey, and Siddhartha Srinivasa. Planning-based prediction for pedestrians. In IROS, 2009. 1 \ No newline at end of file diff --git a/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/images.zip b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3c1a3c74383ea9bfc2648762c4cf8ab63887c20d --- /dev/null +++ b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07c233666f016e0124ea3ec0fee3c073d222684026f737a911077b7f0f19b4f8 +size 591082 diff --git a/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/layout.json b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5855957c9fcabb6b62f99692e3fe3712bda8a7f7 --- /dev/null +++ b/2025/Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment/layout.json @@ -0,0 +1,10756 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 549, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 549, + 120 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 549, + 120 + ], + "type": "text", + "content": "Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "spans": [ + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "type": "text", + "content": "Hiromu Taketsugu" + }, + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "type": "text", + "content": " Takeru Oba" + }, + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "type": "text", + "content": " Takahiro Maeda" + }, + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "type": "text", + "content": " Shohei Nobuhara" + }, + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "type": "text", + "content": " Norimichi Ukita" + }, + { + "bbox": [ + 158, + 142, + 451, + 170 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "spans": [ + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "text", + "content": "Toyota Technological Institute, Japan " + }, + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "text", + "content": "Kyoto Institute of Technology, Japan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 187, + 522, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 187, + 522, + 200 + ], + "spans": [ + { + "bbox": [ + 85, + 187, + 522, + 200 + ], + "type": "text", + "content": "{sd23426, sd21502, sd21601}@toyota-ti.ac.jp, nob@kit.ac.jp, ukita@toyota-ti.ac.jp" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 221, + 200, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 221, + 200, + 234 + ], + "spans": [ + { + "bbox": [ + 151, + 221, + 200, + 234 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 247, + 297, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 297, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 297, + 474 + ], + "type": "text", + "content": "Humans can predict future human trajectories even from momentary observations by using human pose-related cues. However, previous Human Trajectory Prediction (HTP) methods leverage the pose cues implicitly, resulting in implausible predictions. To address this, we propose Locomotion Embodiment, a framework that explicitly evaluates the physical plausibility of the predicted trajectory by locomotion generation under the laws of physics. While the plausibility of locomotion is learned with an indifferentiable physics simulator, it is replaced by our differentiable Locomotion Value function to train an HTP network in a data-driven manner. In particular, our proposed Embodied Locomotion loss is beneficial for efficiently training a stochastic HTP network using multiple heads. Furthermore, the Locomotion Value filter is proposed to filter out implausible trajectories at inference. Experiments demonstrate that our method enhances even the state-of-the-art HTP methods across diverse datasets and problem settings. Our code is available at: https://github.com/ImIntheMiddle/EmLoco." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 490, + 135, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 135, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 135, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 510, + 297, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 297, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 297, + 641 + ], + "type": "text", + "content": "Human Trajectory Prediction (HTP), which predicts future human trajectories, is actively studied [2, 18, 29, 37, 49, 63, 64, 70] due to its importance for applications such as autonomous driving [3, 8, 41, 60, 61, 70], robotics [34, 37, 50, 73], and security systems [10, 66, 72]. These HTP methods assume the availability of sufficiently long past observations and rely solely on a past trajectory. Due to this limitation, these methods cannot handle safety-critical situations in a momentary observation setting, where only a few frames of the past trajectory are available, such as when people suddenly emerge from behind obstacles." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 642, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 297, + 714 + ], + "type": "text", + "content": "Regarding this issue, prior work [26, 35, 42, 47, 50] validates that human pose-related cues, such as body orientations and limb motions are beneficial for inferring future trajectories, especially in a momentary observation setting. These methods learn the consistency between human past poses and future trajectories in a data-driven manner alone." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 227, + 554, + 412 + ], + "blocks": [ + { + "bbox": [ + 316, + 227, + 554, + 412 + ], + "lines": [ + { + "bbox": [ + 316, + 227, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 554, + 412 + ], + "type": "image", + "image_path": "b0fb1b8b9167e5f301ade0bd7c15d736ad54e579c65d1900ed1b5bebc01d9b53.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 415, + 555, + 505 + ], + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 505 + ], + "type": "text", + "content": "Figure 1. Overview of our method. Unlike existing methods, which often predict physically implausible trajectories, our framework uses locomotion generation in a physics simulator to incorporate the laws of physics to HTP by training the plausibility score as the consistency between the observed 3D pose and future possible trajectories, which are indicated by " + }, + { + "bbox": [ + 313, + 415, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\boldsymbol{j}_0" + }, + { + "bbox": [ + 313, + 415, + 555, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\tau_f^{1,2,3}" + }, + { + "bbox": [ + 313, + 415, + 555, + 505 + ], + "type": "text", + "content": ", respectively, in the figure. Additionally, at inference, our method can evaluate predicted trajectories, " + }, + { + "bbox": [ + 313, + 415, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_f^{1,2,3}" + }, + { + "bbox": [ + 313, + 415, + 555, + 505 + ], + "type": "text", + "content": ", to filter out implausible ones." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 520, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 520, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 312, + 520, + 556, + 640 + ], + "type": "text", + "content": "The data-driven approach can prove its ability with a large-scale dataset that covers most possible combinations of poses and trajectories. However, such datasets are unavailable in the literature because human locomotion is diverse. In particular, it is difficult to collect a large number of accurate pose annotations, which are collected by time-consuming delicate motion capture systems [30]. Instead, we propose to fully leverage the beneficial cues of human poses, by assessing the physical consistency between observed poses and future trajectories for physical plausibility-aware HTP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "content": "For this physical plausibility-aware HTP, we present Locomotion Embodiment, a framework that incorporates the laws of physics into an HTP network. This is achieved by physically simulating the locomotion conditioned by observed poses and future trajectories in a physics simulator [31]; refer to Fig. 1 to see the effect of our framework. As a typical" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12324" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 145 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 145 + ], + "type": "text", + "content": "example, one of the human poses included in a future possible trajectory " + }, + { + "bbox": [ + 54, + 72, + 294, + 145 + ], + "type": "inline_equation", + "content": "\\tau_f^1" + }, + { + "bbox": [ + 54, + 72, + 294, + 145 + ], + "type": "text", + "content": " (indicated by " + }, + { + "bbox": [ + 54, + 72, + 294, + 145 + ], + "type": "inline_equation", + "content": "j_{t}" + }, + { + "bbox": [ + 54, + 72, + 294, + 145 + ], + "type": "text", + "content": " in Fig. 1) is a little unstably tilted, leading to a lower plausibility score of " + }, + { + "bbox": [ + 54, + 72, + 294, + 145 + ], + "type": "inline_equation", + "content": "\\tau_f^1 = 0.3" + }, + { + "bbox": [ + 54, + 72, + 294, + 145 + ], + "type": "text", + "content": ". By training the HTP network in a data-driven manner with following the laws of physics, we benefit from the advantages of both the data-driven and physics-based approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 147, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 296, + 326 + ], + "type": "text", + "content": "By evaluating and back-propagating the plausibility of locomotion generated in a physics simulator, we can train the physical plausibility-aware HTP network. However, locomotion generation using physics simulators requires various motion parameters, most of which are unavailable or noisy (e.g., joint angular velocities estimated from joint positions provided by a 3D pose estimator) in image-based HTP scenarios. In addition, rigorous physics simulators are indifferent to the physical plausibility, this paper proposes the Locomotion Value (LocoVal) function implemented with a differentiable network to serve as a surrogate for a physics simulator for the plausibility evaluation (Sec. 4.1). Given the plausibility evaluation provided by the LocoVal function, our EmLoco loss penalizes implausible trajectories (Sec. 4.2)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 458 + ], + "type": "text", + "content": "This EmLoco loss is used in conjunction with the MSE loss for supervised trajectory prediction learning. While the MSE loss behaves as a data term, our EmLoco loss can be a regularization term that incorporates prior knowledge about physics. This is crucial for maintaining the diversity of the predicted trajectories. Although human trajectories are inherently stochastic due to many factors, such as dynamic environments, training with the MSE loss alone forces all predicted trajectories to be aligned with a single ground truth. That is, minimizing the MSE essentially decreases the diversity of the predicted trajectories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 459, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 531 + ], + "type": "text", + "content": "On the other hand, for stochastically predicting multiple future trajectories, our EmLoco loss allows us to supervise all predicted trajectories to improve their physical plausibility while preserving diversity. In addition, our LocoVal function can also serve as the LocoVal filter that evaluates and filters the candidate trajectories at inference (Sec. 4.3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 532, + 269, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 532, + 269, + 544 + ], + "spans": [ + { + "bbox": [ + 67, + 532, + 269, + 544 + ], + "type": "text", + "content": "Our framework consists of three key components:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 544, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 568 + ], + "type": "text", + "content": "1. LocoVal function as a surrogate for an indifferentiable physics simulator in HTP network training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 569, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 628 + ], + "type": "text", + "content": "2. EmLoco loss for back-propagating the plausibility of predicted trajectories. This loss is also useful for deterministic predictors but more beneficial for stochastic predictors because of its applicability to all possible predicted trajectories while preserving their stochastic diversity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 664 + ], + "type": "text", + "content": "3. LocoVal filter for filtering out implausible trajectories at inference. This filter can be applied to any pre-trained HTP network in a plug-and-play manner." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "This work is the first attempt to use simulated humanoid control in a physics simulator for HTP. Since the physics simulator is used only during training, an additional cost (i.e., the cost of the LocoVal function) is negligible at inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 91, + 508, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 508, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 508, + 104 + ], + "type": "text", + "content": "2.1. Human Trajectory Prediction (HTP)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 109, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 276 + ], + "type": "text", + "content": "HTP methods can be divided into deterministic [2, 23, 37, 47, 63] and stochastic [18, 29, 49, 50, 64] methods. Deterministic HTP methods [2, 20, 23, 37, 46, 63] predict the most likely trajectory. For example, Social-LSTM uses the Long Short Term Memory [2] to represent the consistency between observed and future trajectories. On the other hand, stochastic HTP methods [12, 49, 50, 64] produce multiple trajectories or probability distributions. To model the inherent stochasticity in HTP, these methods employ probabilistic frameworks such as Generative Adversarial Networks [18, 48] and Diffusion models [15, 33], as well as Normalizing Flows [29, 51] to draw multiple trajectories at inference. Our method is model-agnostic and applicable to both deterministic and stochastic methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 466, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 466, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 466, + 296 + ], + "type": "text", + "content": "2.2. HTP with Pose Information" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 302, + 555, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 398 + ], + "type": "text", + "content": "Prior studies demonstrate that human poses benefit HTP [26, 42, 47, 50, 65]. Yagi et al. [65] and Liang et al. [26] incorporate human poses estimated by CNNs into HTP. Transformer-based methods such as Social-Transmotion [47] and Human Scene Transformer (HST) [50] further improve HTP conditioned by human poses. In particular, HST [50] reports the effectiveness of human pose conditioning for momentary observation scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 471 + ], + "type": "text", + "content": "However, since these methods include poses simply as an additional input to the HTP network, predicted future trajectories can be physically inconsistent with past poses. Inspired by this observation, our Locomotion Embodiment framework represents the physical plausibility of predicted trajectories as pose-trajectory consistent locomotion." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 472, + 554, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 507 + ], + "type": "text", + "content": "Note that we aim to improve HTP by fully leveraging human poses, distinct from Human Motion Prediction which prioritizes predicting future poses [1, 4, 17, 28, 32, 69]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 516, + 449, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 449, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 449, + 528 + ], + "type": "text", + "content": "2.3. Locomotion Generation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 533, + 556, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 652 + ], + "type": "text", + "content": "Locomotion generation is a task to generate walking motions for humanoid robots [16, 43, 44] and virtual characters [22, 38, 71]. Locomotion can be controlled on various conditions such as destinations [7, 19, 56], trajectories [21, 45], and poses [27, 61]. Generative models such as Diffusion models [5, 21, 53, 58, 67] can generate such conditioned locomotion, but they are prone to generate physically implausible locomotion (e.g., skating motions). This is because the laws of physics are not explicitly formulated in generative models." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "On the other hand, physics-based locomotion generation [27, 40, 45, 60-62] follows the laws of physics in simulators [6, 31]. However, a generated locomotion would fail to follow a given trajectory or fall due to kinematic limitations if a humanoid is controlled to walk along an implausible" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12325" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "trajectory. Therefore, a physics-based locomotion generator can evaluate whether the conditioned trajectory is plausible. Our Locomotion Embodiment framework utilizes this ability to evaluate the plausibility of future trajectories." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 297, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 297, + 312 + ], + "type": "text", + "content": "The aforementioned physics-based locomotion generator, e.g., PACER [27, 40, 45, 61], can be trained with reinforcement learning in a physics simulator. A straightforward way to use this locomotion generator for training an HTP network is to directly incorporate the locomotion generator into the HTP network. However, this direct integration has two problems: (1) As mentioned in Sec. 1, the physics simulator is indifferent, preventing training the HTP network by back-propagation. (2) The locomotion generator using the physics simulator requires physical humanoid states, e.g., joint angular velocities, which are unavailable or noisy in image-based HTP scenarios. To address these two problems, our framework uses the physics simulator and the locomotion generator only before training the HTP network and replaces them with the LocoVal function in training the HTP network." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 329, + 140, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 140, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 140, + 341 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": "HTP. An HTP network, " + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": ", takes the past " + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{p}}" + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": " frames as input, including the past trajectory and 3D poses of a target person, and predicts the trajectory of the target person in the future " + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": " frames. The past trajectory " + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}_{\\mathrm{p}} \\in \\mathbb{R}^{T_{\\mathrm{p}} \\times 2}" + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": " and the future target trajectory " + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}_{\\mathrm{f}} \\in \\mathbb{R}^{T_{\\mathrm{f}} \\times 2}" + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": " are on the ground " + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": " plane. The 3D pose involves " + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": " joint positions " + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{j}_t \\in \\mathbb{R}^{J \\times 3}" + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": " at the past time " + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "inline_equation", + "content": "(-T_{\\mathrm{p}} + 1) \\leq t \\leq 0" + }, + { + "bbox": [ + 55, + 350, + 296, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 436, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 461 + ], + "type": "text", + "content": "The predicted future trajectory " + }, + { + "bbox": [ + 55, + 436, + 296, + 461 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 436, + 296, + 461 + ], + "type": "text", + "content": " is expected to be as close as possible to its ground-truth trajectory " + }, + { + "bbox": [ + 55, + 436, + 296, + 461 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 436, + 296, + 461 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "content": "Locomotion Generation. In our framework, humanoid locomotion is generated to follow any given future trajectory " + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}_{\\mathrm{s}} \\in \\mathbb{R}^{T_{\\mathrm{f}} \\times 2}" + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "content": " in a physics simulator [31]. As such a locomotion generator " + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "content": ", our method employs PACER [45], which is implemented with the policy network [36] trained via reinforcement learning [52]. " + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{\\tau}_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "content": " and the initial humanoid state " + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "content": " to control the humanoid joints. " + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "content": " involves the positions, rotations, velocities, and angular velocities of all joints. In PACER [45], " + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "content": " is trained by maximizing the discounted cumulative reward " + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 484, + 296, + 628 + ], + "type": "text", + "content": " consisting of a trajectory following reward, a motion style reward [40], and an energy penalty [11]. Refer to PACER [45] for details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "content": "In PACER, the humanoid movement is constrained by the limits of joint angles and torques, preventing it from producing kinematically impossible locomotion. Due to this limitation, when conditioned on a physically implausible trajectory, the humanoid fails to follow it, resulting in a lower reward. Therefore, we define the discounted cumulative reward " + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "content": " [52] as the plausibility of " + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 74, + 552, + 273 + ], + "blocks": [ + { + "bbox": [ + 319, + 74, + 552, + 273 + ], + "lines": [ + { + "bbox": [ + 319, + 74, + 552, + 273 + ], + "spans": [ + { + "bbox": [ + 319, + 74, + 552, + 273 + ], + "type": "image", + "image_path": "b132474f5f9792dfc6aa265d95795383915f0cf2e15abb00dd1b1eef279e64e7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 277, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 300 + ], + "type": "text", + "content": "Figure 2. The overview of the proposed framework. " + }, + { + "bbox": [ + 313, + 277, + 555, + 300 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 313, + 277, + 555, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 277, + 555, + 300 + ], + "type": "inline_equation", + "content": "⑤" + }, + { + "bbox": [ + 313, + 277, + 555, + 300 + ], + "type": "text", + "content": " mean the weights are trained and fixed, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 320, + 495, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 495, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 495, + 335 + ], + "type": "text", + "content": "4. Physical Plausibility-aware HTP" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "text", + "content": "Overview. Figure 2 shows the overview of our framework consisting of two training stages and an inference phase. In (a) the first training stage, given the physics-based locomotion generator " + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "text", + "content": " trained as mentioned in Sec. 3, a current humanoid state is fed into it to obtain the reward as the physical plausibility. With this physical plausibility as the ground truth, the LocoVal function " + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "text", + "content": " is trained. In (b) the second stage, observed pose-related cues in training data are fed into both " + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "text", + "content": " also takes the future trajectories predicted by " + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "text", + "content": " to compute the EmLoco losses. The EmLoco loss supports both the MSE and minMSE [18, 50, 63] losses used for deterministic and stochastic predictions, respectively, to train " + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "text", + "content": ". At (c) inference, candidate trajectories predicted by " + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "text", + "content": " are fed into " + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 340, + 556, + 510 + ], + "type": "text", + "content": " to filter out implausible trajectories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 516, + 421, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 421, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 421, + 528 + ], + "type": "text", + "content": "4.1. LocoVal Function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "text", + "content": "The detail of training the LocoVal function " + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "text", + "content": " is illustrated in Fig. 3. As shown in the top half of Fig. 3, the discounted cumulative reward " + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "text", + "content": " [52] is used as the plausibility of Embodied Locomotion that is generated through " + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "text", + "content": " using a physics simulator. " + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "text", + "content": " is trained to estimate the reward " + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\Omega}" + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "text", + "content": " close to " + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "text", + "content": " from only available cues in image-based HTP scenarios because " + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "text", + "content": " serves as a surrogate for " + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 534, + 556, + 617 + ], + "type": "text", + "content": " at inference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": "As such available cues, our method uses the future trajectory " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": ", the initial pose " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "j_0" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": ", and the initial velocity of the root joint " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_{\\mathrm{root},0} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "j_0" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_{\\mathrm{root},0}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " are in a subset " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "h_0'" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " extracted from " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": ". In training " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": ", these cues are available from the physics simulator. At inference, " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " is replaced by " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{f}}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "j_0" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " can be obtained by a 3D pose estimator [14, 50], and " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_{\\mathrm{root},0}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " can be obtained from " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\tau_p" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 618, + 557, + 714 + ], + "type": "text", + "content": " is implemented as a differentiable MLP, taking these cues as input." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12326" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 296, + 163 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 163 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 296, + 163 + ], + "type": "image", + "image_path": "2f7a4a352c71f865c61e391bc5a80910b1147b2718c4c3e5b1c5717ef2fb9290.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 167, + 285, + 178 + ], + "lines": [ + { + "bbox": [ + 66, + 167, + 285, + 178 + ], + "spans": [ + { + "bbox": [ + 66, + 167, + 285, + 178 + ], + "type": "text", + "content": "Figure 3. The overview of training our LocoVal function, " + }, + { + "bbox": [ + 66, + 167, + 285, + 178 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 66, + 167, + 285, + 178 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 189, + 294, + 299 + ], + "blocks": [ + { + "bbox": [ + 58, + 189, + 294, + 299 + ], + "lines": [ + { + "bbox": [ + 58, + 189, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 58, + 189, + 294, + 299 + ], + "type": "image", + "image_path": "c55a620292924e8da6a607c2aa66c53643b22e615d323f6ea73ea44b4ed6add2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 306, + 277, + 317 + ], + "lines": [ + { + "bbox": [ + 72, + 306, + 277, + 317 + ], + "spans": [ + { + "bbox": [ + 72, + 306, + 277, + 317 + ], + "type": "text", + "content": "Figure 4. The overview of training the HTP network, " + }, + { + "bbox": [ + 72, + 306, + 277, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "text", + "content": "Training Process. Unlike PACER [45], which trains " + }, + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "text", + "content": " to obtain high rewards even for implausible trajectories, e.g., dangerously sharp turns, such trajectories should result in low rewards in our framework. This is because humans do not typically walk along such implausible trajectories in daily life. For this reason, we train " + }, + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "text", + "content": " with only plausible pairs of " + }, + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "inline_equation", + "content": "\\tau_{s}" + }, + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "text", + "content": " to give high rewards only to plausible pose-trajectory pairs. On the other hand, since " + }, + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 54, + 336, + 295, + 468 + ], + "type": "text", + "content": " must discriminate between the plausible and implausible pose-trajectory pairs, it is trained with plausible pairs and randomly paired diverse implausible pose-trajectory pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "content": " is trained by minimizing the following loss " + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "content": " so that it can estimate the plausibility of the Embodied Locomotion " + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\varOmega=\\mathcal{G}(\\pmb{\\tau}_{\\mathrm{s}},\\pmb{h}_{0})" + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "content": " obtained by rolling out " + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 511, + 295, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 511, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 104, + 511, + 295, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {V}} = \\operatorname {M S E} \\left(\\mathcal {V} \\left(\\boldsymbol {\\tau} _ {\\mathrm {s}}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right), \\mathcal {G} \\left(\\boldsymbol {\\tau} _ {\\mathrm {s}}, \\boldsymbol {h} _ {0}\\right)\\right), \\tag {1}", + "image_path": "bb454d83dd4f2e7f56712a8bfcd33378db2fad08cf83fa9167241e10581f2294.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 532, + 275, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 275, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 275, + 544 + ], + "type": "text", + "content": "where MSE denotes the Mean Squared Error function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 550, + 277, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 277, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 277, + 563 + ], + "type": "text", + "content": "4.2. HTP Network Training with EmLoco Loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "text", + "content": "Figure 4 illustrates the detail of training the HTP network " + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "text", + "content": ". Our HTP network takes the past information including " + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{p}}" + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "text", + "content": " and 3D poses " + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "inline_equation", + "content": "j_{\\mathrm{p}} = \\{j_{-T_{\\mathrm{p}} + 1},\\dots ,j_0\\}" + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "text", + "content": " as input to predict the future trajectory " + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "text", + "content": " of a target person. Following traditional HTP training methods [2, 47], " + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "text", + "content": " is compared with its ground truth " + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "text", + "content": " to back-propagate the MSE loss " + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}} = \\mathrm{MSE}(\\hat{\\tau}_{\\mathrm{f}},\\tau_{\\mathrm{f}})" + }, + { + "bbox": [ + 55, + 567, + 296, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "EmLoco loss. The EmLoco loss " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " evaluates predicted trajectories by using " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": ", which has prior knowledge about plausibility. Note again that a physics simulator is not required in this second stage or the inference phase because " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " serves as a surrogate for the plausibility evaluation process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "Given the prediction of " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": ", the predicted trajectory " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_{\\mathrm{f}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " and the subset of " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "h_0^{\\prime}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": ") observed in the current frame are fed into " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " to compute " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 395, + 114, + 555, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 114, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 395, + 114, + 555, + 128 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E}} = - \\mathcal {V} \\left(\\hat {\\tau} _ {\\mathrm {f}}, h _ {0} ^ {\\prime}\\right). \\tag {2}", + "image_path": "ca364eb62991ec897608a05ed5072ed4e3677e7e8a3bd6f72c3ce13fd4475f1e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 133, + 555, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 555, + 145 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 555, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 314, + 133, + 555, + 145 + ], + "type": "text", + "content": " encourages " + }, + { + "bbox": [ + 314, + 133, + 555, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 314, + 133, + 555, + 145 + ], + "type": "text", + "content": " to predict " + }, + { + "bbox": [ + 314, + 133, + 555, + 145 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_{\\mathrm{f}}" + }, + { + "bbox": [ + 314, + 133, + 555, + 145 + ], + "type": "text", + "content": " that increases the plausibility." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 145, + 554, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 554, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 554, + 170 + ], + "type": "text", + "content": "As the total training objective, " + }, + { + "bbox": [ + 313, + 145, + 554, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 313, + 145, + 554, + 170 + ], + "type": "text", + "content": " is combined with the MSE loss " + }, + { + "bbox": [ + 313, + 145, + 554, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 145, + 554, + 170 + ], + "type": "text", + "content": " with a hyper-parameter " + }, + { + "bbox": [ + 313, + 145, + 554, + 170 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 145, + 554, + 170 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 393, + 176, + 555, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 176, + 555, + 189 + ], + "spans": [ + { + "bbox": [ + 393, + 176, + 555, + 189 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {T}} + \\alpha \\mathcal {L} _ {\\mathrm {E}}, \\tag {3}", + "image_path": "ffa6ef519c5626692e050c7d17fb8977b5d9acc834f8641f52bd771c756e4047.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 202, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 555, + 299 + ], + "type": "text", + "content": "Multi-head HTP network. Our method with the EmLoco loss can be extended to a stochastic HTP network using multiple heads. As also mentioned in Sec. 1, the MSE loss (Eq. (1)) is inappropriate for multiple stochastic predictions because it forces all of them to converge to a unique ground truth. To mitigate this problem, following prior work [18, 50, 63], the multiple heads in our method are trained by replacing " + }, + { + "bbox": [ + 313, + 202, + 555, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 202, + 555, + 299 + ], + "type": "text", + "content": " in Eq.(3) with the following objective function:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 380, + 308, + 555, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 308, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 380, + 308, + 555, + 327 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T}} = \\min _ {k} \\left(\\operatorname {M S E} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {\\tau} _ {\\mathrm {f}}\\right)\\right), \\tag {4}", + "image_path": "c99800db37f7c3cc9f7edd861318c670b9b59c1164dd0a940c05fde0d3ca4d92.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 331, + 553, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 331, + 553, + 366 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_{\\mathrm{f}}^{k}" + }, + { + "bbox": [ + 313, + 331, + 553, + 366 + ], + "type": "text", + "content": " denotes the prediction of " + }, + { + "bbox": [ + 313, + 331, + 553, + 366 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 331, + 553, + 366 + ], + "type": "text", + "content": "-th head. In Eq. (4), the minMSE loss supervises only one prediction closest to its ground truth " + }, + { + "bbox": [ + 313, + 331, + 553, + 366 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{f}}" + }, + { + "bbox": [ + 313, + 331, + 553, + 366 + ], + "type": "text", + "content": " among multiple predictions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 367, + 554, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 554, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 554, + 438 + ], + "type": "text", + "content": "Unlike the MSE loss, this minMSE loss allows us to maintain diverse predictions. However, the minMSE loss is inefficient. What is even worse is that several heads may be under-fitted if other heads are trained in early training iterations so that these heads are always selected as those minimizing the MSE in Eq. (4) in later iterations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 439, + 555, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 535 + ], + "type": "text", + "content": "On the other hand, our EmLoco loss does not require the unique ground truth, i.e., " + }, + { + "bbox": [ + 313, + 439, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{f}}" + }, + { + "bbox": [ + 313, + 439, + 555, + 535 + ], + "type": "text", + "content": " in Eq. (4). As a regularization term, the EmLoco loss trains all the heads to predict plausible trajectories. This simultaneous optimization allows us to avoid the under-fitting problem mentioned above. In our method, this simultaneous optimization is implemented by averaging the EmLoco losses using " + }, + { + "bbox": [ + 313, + 439, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 439, + 555, + 535 + ], + "type": "text", + "content": " for all predicted trajectories, and this average is used as " + }, + { + "bbox": [ + 313, + 439, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 313, + 439, + 555, + 535 + ], + "type": "text", + "content": " in Eq.(3)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 540, + 463, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 463, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 463, + 552 + ], + "type": "text", + "content": "4.3. LocoVal Filter at Inference" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 558, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 641 + ], + "type": "text", + "content": "The trained LocoVal function can also be applied to trajectory filtering, called the LocoVal filter, at inference using multi-head HTP networks. Similar to the usage of " + }, + { + "bbox": [ + 313, + 558, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 558, + 555, + 641 + ], + "type": "text", + "content": " for training multiple heads simultaneously at training, it computes the plausibilities of all predicted trajectories at inference. Each of the predicted trajectories with a plausibility lower than a threshold " + }, + { + "bbox": [ + 313, + 558, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 558, + 555, + 641 + ], + "type": "text", + "content": " is filtered out as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 328, + 646, + 555, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 646, + 555, + 684 + ], + "spans": [ + { + "bbox": [ + 328, + 646, + 555, + 684 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} \\left\\{\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k} \\mid \\mathcal {V} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right) \\geq \\lambda \\right\\} & \\text {i f} \\exists k, \\mathcal {V} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right) \\geq \\lambda , \\\\ \\operatorname {a r g m a x} _ {\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}} \\mathcal {V} \\left(\\hat {\\boldsymbol {\\tau}} _ {\\mathrm {f}} ^ {k}, \\boldsymbol {h} _ {0} ^ {\\prime}\\right) & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {5}", + "image_path": "a103ad1e549583c44ec4bdf995fe569f68e7ef01282e0a1e35181703607559e9.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Note that the LocoVal filter can be used with any stochastic HTP networks in a plug-and-play manner." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12327" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 93, + 297, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 93, + 297, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 93, + 297, + 178 + ], + "type": "text", + "content": "In Sec. 5.2.1, we benchmark our HTP network trained with the EmLoco loss under the standard HTP setting. Section 5.2.2 presents results for multi-head HTP networks, and Sec. 5.2.3 examines the performance with momentary observations. Then, we report ablation results in Sec. 5.2.4. In Sec. 5.3.1 and Sec. 5.3.2, we evaluate the effectiveness of the LocoVal filter in various scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 183, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 183, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 183, + 208 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 214, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 296, + 299 + ], + "type": "text", + "content": "Datasets. As trajectory datasets, the Joint Track Auto (JTA) dataset [9], the JackRabbit Dataset and Benchmark (JRDB) [34], and the ETH/UCY dataset [24, 39] are used. For the JRDB dataset, a subset provided by the SocialTrans [47] was used. The ETH/UCY dataset is used only in Sec. 5.3.2 to evaluate the performance of the LocoVal filter on pre-trained models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 297, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 297, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 297, + 409 + ], + "type": "text", + "content": "While JTA also provides 3D human poses synchronized with their trajectories, no 3D pose is annotated in JRDB. Instead, in JRDB, our experiments use 3D poses estimated from images and 2D pose annotations by using [14, 50]. While the estimated 3D poses are provided by HST [50], misestimated poses are filtered out in our experiments; see the Supp. for the details. All the 3D poses mentioned above are converted to the SMPL-style poses [25] for a fair comparison with the baseline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 411, + 297, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 297, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 297, + 519 + ], + "type": "text", + "content": "While the 3D poses described above are used in the second stage to train the HTP network " + }, + { + "bbox": [ + 55, + 411, + 297, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 411, + 297, + 519 + ], + "type": "text", + "content": ", more accurate 3D poses are expected to train " + }, + { + "bbox": [ + 55, + 411, + 297, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 411, + 297, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 411, + 297, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 411, + 297, + 519 + ], + "type": "text", + "content": " to learn the physical plausibility of 3D poses. Such accurate 3D poses are provided by the Archive of Motion Capture as Surface Shapes (AMASS) [30]. While accurate 3D poses are available and can be used to compute the trajectories in motion capture datasets such as AMASS, motion capture data do not scale compared with trajectory datasets, as described in Sec. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "content": "To avoid this problem, 3D poses in AMASS are combined with trajectories in the trajectory datasets, i.e., JTA and JRDB, as follows. Plausible and implausible pose-trajectory pairs are required to train " + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "content": ", as described in Sec. 4.1. The plausible pairs are prepared by spatially aligning the 3D pose and the trajectories, which are randomly sampled from AMASS and the trajectory datasets, respectively. This spatial alignment consists of the orientation and velocity alignments. Given the heading vector of each pedestrian in these two datasets, the heading orientations are aligned between the motion capture pose from AMASS and the trajectory data. For the velocity alignment, on the other hand, the trajectory data is expanded or contracted to make the vector norms equal between the motion capture pose and the trajectory data. Refer to the Supp. for detailed training conditions on the aforementioned datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 371 + ], + "type": "text", + "content": "HTP networks. Since multiple people are observed in each frame of the aforementioned datasets, all baseline HTP networks of our method take not only the past trajectory of a target person but also those of other people as input. This allows us to predict the target trajectory affected by the dynamic environment. While our method is introduced to take only the target person as input, all people in the scene are fed into " + }, + { + "bbox": [ + 313, + 72, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 72, + 555, + 371 + ], + "type": "text", + "content": " following the previous methods. In the evaluations with the JTA and JRDB datasets, Social-Transmotion (Social-Trans), is used as our baseline without any architectural changes. We extend this baseline of the trajectory predictor " + }, + { + "bbox": [ + 313, + 72, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 72, + 555, + 371 + ], + "type": "text", + "content": " by plugging the EmLoco loss into the training loss. Following the implementation of Social-Trans, past trajectories, 2D and 3D human poses, and 2D and 3D bounding boxes are fed into " + }, + { + "bbox": [ + 313, + 72, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 72, + 555, + 371 + ], + "type": "text", + "content": " in the JTA dataset [9], while the past trajectories, 2D bounding boxes, and 3D human poses are used in the JRDB dataset. Note that the 3D human poses are estimated for the JRDB dataset [34] as mentioned earlier. For the stochastic version of our method, we modify Social-Trans [47] by parallelizing the prediction heads with different initial weights. For comparison, all HTP networks predict the future " + }, + { + "bbox": [ + 313, + 72, + 555, + 371 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{f}} = 12" + }, + { + "bbox": [ + 313, + 72, + 555, + 371 + ], + "type": "text", + "content": " frames at 2.5 fps on the JTA and JRDB datasets. The number of past frames differs depending on the experiment; e.g., 9 frames in Sec. 5.2.1 and Sec. 5.2.2, and 2 frames as momentary observations [55] in Sec. 5.2.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 372, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 556, + 456 + ], + "type": "text", + "content": "In experiments on the ETH/UCY dataset [24, 39], where a bird's-eye view hinders the use of 3D poses, only past trajectories are fed into " + }, + { + "bbox": [ + 313, + 372, + 556, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 372, + 556, + 456 + ], + "type": "text", + "content": ". Here, we employ EqMotion [63], a state-of-the-art stochastic HTP method that only uses past trajectories, as a pre-trained HTP network. Each past trajectory consists of " + }, + { + "bbox": [ + 313, + 372, + 556, + 456 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{p}} = 8" + }, + { + "bbox": [ + 313, + 372, + 556, + 456 + ], + "type": "text", + "content": " frames, while each future trajectory remains at " + }, + { + "bbox": [ + 313, + 372, + 556, + 456 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{f}} = 12" + }, + { + "bbox": [ + 313, + 372, + 556, + 456 + ], + "type": "text", + "content": " frames at 2.5 fps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 557, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 557, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 557, + 594 + ], + "type": "text", + "content": "Evaluation. Two major metrics, the Average Displacement Error (ADE) and the Final Displacement Error (FDE), both measured in meters, evaluate the HTP results. Consistent with the other HTP literature [29, 49, 50, 63], in evaluating the stochastic HTP networks, we report minADE and minFDE among multiple trajectories. Also, following MATRIX [64], we report " + }, + { + "bbox": [ + 313, + 472, + 557, + 594 + ], + "type": "inline_equation", + "content": "\\chi^2" + }, + { + "bbox": [ + 313, + 472, + 557, + 594 + ], + "type": "text", + "content": " distances for the four physics primitives of each predicted trajectory, namely velocities, accelerations, angular velocities, and angular accelerations, representing the plausibility of the predicted trajectories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 601, + 511, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 511, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 511, + 613 + ], + "type": "text", + "content": "5.2. HTP Training with the EmLoco Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 619, + 556, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 556, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 556, + 655 + ], + "type": "text", + "content": "Here, we evaluate our HTP network trained with the proposed EmLoco loss. Note that the LocoVal filter is not applied in these experiments. " + }, + { + "bbox": [ + 313, + 619, + 556, + 655 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 619, + 556, + 655 + ], + "type": "text", + "content": " in Eq.(3) is set to 100." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 662, + 517, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 517, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 517, + 674 + ], + "type": "text", + "content": "5.2.1. Comparison with Existing HTP Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Following the standard protocol [2, 12, 18, 47, 63] using 9 past frames, where past trajectories are sufficiently long, we compare the accuracy of deterministic predictions by our" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12328" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 120, + 294, + 294 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 1. HTP results under the standard setting. Red and blue indicate the best and second-best. 'Social-Trans' and 'Ours' utilize pose input. For methods other than 'Social-Trans' and 'Ours', we report the results from the Social-Trans literature [47]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 120, + 294, + 294 + ], + "lines": [ + { + "bbox": [ + 56, + 120, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 120, + 294, + 294 + ], + "type": "table", + "html": "
MethodJTA [9]JRDB [34]
ADEFDEADEFDE
Social-GAN-det [18]1.663.760.500.99
Transformer [13]1.563.540.561.10
Vanilla-LSTM [2]1.443.250.420.83
Occupancy-LSTM [2]1.413.150.430.85
Directional-LSTM [23]1.373.060.450.87
Dir-social-LSTM [23]1.232.590.480.95
Social-LSTM [2]1.212.540.470.95
Autobots [12]1.202.700.390.80
Trajectory++ [49]1.182.530.400.78
EqMotion [63]1.132.390.400.77
Social-Trans [47]1.112.260.400.76
Ours0.971.910.370.72
", + "image_path": "9eef455644a15eeb783a45b80bd6cf5a9d23e8064458295ad8fa203943e71403.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 308, + 298, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 298, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 298, + 418 + ], + "type": "text", + "content": "HTP network with other HTP networks [2, 12, 18, 23, 47, 49, 59, 63]. Table 1 shows that our method outperforms all other methods including the state-of-the-art baseline, Social-Trans [47], on both the JTA [9] and the JRDB [34] datasets. This demonstrates that the EmLoco loss effectively provides the prior knowledge of the trajectory-pose consistency during training. Notably, the performance improvements on the JRDB dataset [34] with imperfect 3D poses suggest the applicability of our EmLoco loss to real-world applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 200, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 200, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 200, + 434 + ], + "type": "text", + "content": "5.2.2. Multi-head HTP Networks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 297, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 297, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 297, + 521 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the EmLoco loss on a multi-head HTP network, 5 and 20 prediction heads are implemented on Social-Trans [47]. As described in Sec. 4.2, our proposed method is expected to improve the plausibility of predicted trajectories closer to human locomotion. From this perspective, we evaluate the plausibility of the predicted trajectories by the " + }, + { + "bbox": [ + 55, + 437, + 297, + 521 + ], + "type": "inline_equation", + "content": "\\chi^2" + }, + { + "bbox": [ + 55, + 437, + 297, + 521 + ], + "type": "text", + "content": " distance employed in MATRIX [64]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": "Table 2 presents the comparisons with the baseline. Our method achieves the best values across almost all metrics on both the JTA [9] and JRDB [34] datasets. Particularly, the improvements in ADE and FDE reveal that all heads are trained properly because these metrics are evaluated with the prediction of all heads. Also, improvements in " + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\chi^2" + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": " distances, the plausibility regarding physics primitives, demonstrate that " + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": " trained with our EmLoco loss predicts trajectories similar to actual human locomotion than the baseline [47]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "content": "Figure 5 shows qualitative results comparing stochastic HTP with 5 heads for both Social-Trans [47] and our method. While the Social-Trans trained solely with minMSE can predict diverse trajectories, many of them are implausible, including sharp turns. In contrast, the predictions of our HTP network maintain diversity across all heads while remaining plausible, resulting in trajectories closer to the ground truth." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 314, + 76, + 555, + 187 + ], + "blocks": [ + { + "bbox": [ + 314, + 76, + 555, + 187 + ], + "lines": [ + { + "bbox": [ + 314, + 76, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 314, + 76, + 555, + 187 + ], + "type": "image", + "image_path": "e545968f64e3b6e38a64bf326a3604d1be02541b6ead9ec3a8aa856f56dad392.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 193, + 556, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 556, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 556, + 228 + ], + "type": "text", + "content": "Figure 5. Comparison of the stochastic HTP between Social-Trans [47] and ours with 5 heads on the JTA dataset [9]. Human poses are shown at a doubled scale for visualization purposes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 250, + 555, + 373 + ], + "blocks": [ + { + "bbox": [ + 313, + 250, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 313, + 250, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 555, + 373 + ], + "type": "image", + "image_path": "b42ac27e3786c76a1711e01516b66deb639db7991942c5be018ce7758c758df7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 380, + 556, + 426 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 426 + ], + "type": "text", + "content": "Figure 6. Comparison of the prediction results between the baseline Social-Trans [47] and our method with momentary observations on the JTA dataset [9]. The observed pose are displayed at a doubled scale for visualization purposes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 453, + 497, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 497, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 497, + 464 + ], + "type": "text", + "content": "5.2.3. HTP with Momentary Observations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 471, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 544 + ], + "type": "text", + "content": "As noted in Sec. 1, existing HTP methods struggle with momentary observations. On the contrary, because the proposed EmLoco loss encourages physical plausibility-aware HTP, " + }, + { + "bbox": [ + 313, + 471, + 556, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 471, + 556, + 544 + ], + "type": "text", + "content": " trained with the EmLoco loss is expected to be robust against this setting. From this point of view, we experiment with 2 past frames following the prior work [55]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 557, + 715 + ], + "type": "text", + "content": "The results are shown in Table 3. Consistent with Tables 1 and 2, the EmLoco loss allows our method to outperform the baseline [47] across almost all categories. In addition, Fig. 6 presents qualitative results in two cases: when a pedestrian is (a) going straight and (b) turning. In both cases, (i) Social-Trans [47] with trajectory-only input results in significant prediction errors, and (ii) Social-Trans with pose input still predicts implausible trajectories, especially during sharp turns. In contrast, (iii) our method provides plausible predictions that closely follow the ground-truth trajectory and align with the observed pose. This demonstrates that the proposed method can effectively leverage human poses and achieve plausibility-aware HTP, thanks to the physics-based LocoVal function training." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12329" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 120, + 558, + 196 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 2. Comparison of our HTP network using the EmLoco loss with baseline [47] in the stochastic HTP under 9 frames of observations. Red indicates the best result for each category. The results are presented as left/right values, where the left denotes evaluations with 5 heads and the right denotes 20 heads. 'Vel,' 'Acc,' 'Ang. Vel,' and 'Ang. Acc.' represent the " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\chi^2" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " distances between the distributions of the predicted and ground-truth trajectories for velocity, acceleration, angular velocity, and angular acceleration, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 120, + 558, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 120, + 558, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 120, + 558, + 196 + ], + "type": "table", + "html": "
MethodADEFDEminADEminFDE\\( \\chi^2 \\) distances
Vel.Acc.Ang. Vel.Ang. Acc.
JTASocial-Trans1.86/2.143.51/4.260.88/0.711.04/0.540.134/0.1690.009/0.0090.011/0.0070.026/0.008
Ours1.68/1.803.34/3.560.79/0.661.02/0.540.100/0.0870.002/0.0030.009/0.0070.016/0.011
JRDBSocial-Trans0.61/0.711.19/1.490.25/0.180.44/0.260.025/0.0770.034/0.0130.021/0.0190.029/0.137
Ours0.56/0.701.16/1.480.23/0.170.44/0.250.024/0.0670.011/0.0040.021/0.0200.019/0.036
", + "image_path": "dc7e31e98663d28c3d2cab5f599d91877e90dd787405857adcda4db3a9100214.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 60, + 277, + 290, + 384 + ], + "blocks": [ + { + "bbox": [ + 55, + 215, + 296, + 259 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 259 + ], + "type": "text", + "content": "Table 3. HTP results with momentary observations where only 2 frames of past trajectories are observed. Red and blue indicate the best and second-best for each category. 'Traj.' indicates the input is only past trajectories. 'Det.' denotes the deterministic prediction." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 268, + 268, + 277 + ], + "lines": [ + { + "bbox": [ + 80, + 268, + 268, + 277 + ], + "spans": [ + { + "bbox": [ + 80, + 268, + 268, + 277 + ], + "type": "text", + "content": "(a) HTP with momentary observations on the JTA dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 277, + 290, + 384 + ], + "lines": [ + { + "bbox": [ + 60, + 277, + 290, + 384 + ], + "spans": [ + { + "bbox": [ + 60, + 277, + 290, + 384 + ], + "type": "table", + "html": "
MethodADEFDEminADEminFDE
Det.Social-Trans (Traj.)1.573.26--
Social-Trans1.362.80--
Ours1.332.72--
5headsSocial-Trans (Traj.)2.254.620.911.51
Social-Trans2.234.550.921.53
Ours1.943.960.871.30
20headsSocial-Trans (Traj.)2.585.280.710.76
Social-Trans2.465.070.690.73
Ours2.124.470.650.68
", + "image_path": "5a934815e4301e01a05f32d6337ea0d9b6f371a6947d6781ef7d6dc74d8f0489.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 60, + 404, + 290, + 511 + ], + "blocks": [ + { + "bbox": [ + 77, + 391, + 271, + 400 + ], + "lines": [ + { + "bbox": [ + 77, + 391, + 271, + 400 + ], + "spans": [ + { + "bbox": [ + 77, + 391, + 271, + 400 + ], + "type": "text", + "content": "(b) HTP with momentary observations on the JRDB dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 404, + 290, + 511 + ], + "lines": [ + { + "bbox": [ + 60, + 404, + 290, + 511 + ], + "spans": [ + { + "bbox": [ + 60, + 404, + 290, + 511 + ], + "type": "table", + "html": "
MethodADEFDEminADEminFDE
Det.Social-Trans (Traj.)0.500.94--
Social-Trans0.500.94--
Ours0.440.85--
5headsSocial-Trans (Traj.)0.571.190.250.47
Social-Trans0.581.200.250.47
Ours0.551.120.240.48
20headsSocial-Trans (Traj.)0.681.460.180.30
Social-Trans0.681.460.170.28
Ours0.681.450.180.28
", + "image_path": "c3c881a14397c9c89b2c4242fff6e2e2b92d05b50252675ddfda2231555748b1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 155, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 155, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 155, + 555 + ], + "type": "text", + "content": "5.2.4. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "content": "Supervision by the EmLoco loss. To provide deeper insights into the impact of the EmLoco loss " + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "content": " on HTP network training, we train " + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "content": " only with " + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "content": " i.e., omitting the loss " + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "content": " with the ground-truth trajectories. The results are presented in Table 4(B). Importantly, even the proposed method without " + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "content": " can improve performance from a randomly initialized network (Table 4(A)). This demonstrates that the supervision provided by " + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "content": " contributes to the improvements in HTP performance, while it is designed to be used in conjunction with " + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "content": ". We emphasize that the combination of " + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 54, + 570, + 297, + 713 + ], + "type": "text", + "content": " (Table 4(G)) improves HTP performance compared to the baseline [47] (Table 4(C))." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 259, + 555, + 360 + ], + "blocks": [ + { + "bbox": [ + 313, + 215, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 248 + ], + "type": "text", + "content": "Table 4. Ablation studies on JTA [9] and JRDB [34] datasets. Red and blue indicate the best and second-best results for each category. 'initialized' represents the initialized HTP network for reference." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 259, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 317, + 259, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 259, + 555, + 360 + ], + "type": "table", + "html": "
MethodJTAJRDB
ADEFDEADEFDE
(A) Initialized5.078.921.311.87
(B) w/o GT loss LT3.528.200.921.55
(C) w/o EmLoco loss LE [47]1.112.260.400.76
(D) w/o Implausible trajectories1.012.050.380.74
(E) w/o Pose input j01.072.240.390.75
(F) w/o Root velocity vroot,00.992.030.380.74
(G) Ours0.971.910.370.72
", + "image_path": "793576e0bc0a71b2bb0b1cacf163ae21f69c0fc5a834902a4d4f07116403c310.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 456, + 553, + 531 + ], + "blocks": [ + { + "bbox": [ + 313, + 392, + 555, + 448 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 448 + ], + "type": "text", + "content": "Table 5. Results of the LocoVal filter for stochastic HTP on the JTA [9] dataset with 9 and 2-frame observations. The results are presented as left/right values, where the left denotes evaluations with 5 heads and the right denotes 20 heads. Red indicates the best result. 'Rejected' refers to the performance of rejected trajectories." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 456, + 553, + 531 + ], + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 531 + ], + "type": "table", + "html": "
Method9 frames2 frames
ADEFDEADEFDE
Social-Trans(w/o LE)w/o Filtering1.86/2.143.51/4.262.23/2.464.55/5.07
w/ Filtering1.81/1.973.49/3.932.06/2.214.27/4.56
Rejected2.37/3.413.75/6.653.83/4.327.16/8.87
Ours(w/ LE)w/o Filtering1.68/1.803.34/3.561.94/2.123.96/4.47
w/ Filtering1.65/1.763.31/3.521.90/2.083.88/4.41
Rejected2.32/2.403.86/4.193.22/3.256.18/6.41
", + "image_path": "92e66b48e8242bff4b25ce51d1d71a150f877540b85edef83039878c6867b17b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Inputs into the LocoVal function. We ablate the inputs of " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " as follows: (1) implausible pose-trajectory pairs used in training " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " (Table 4(D)), (2) the human pose " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "j_0" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " at both training and inference (Table 4(E)), (3) and the root velocity " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{\\mathrm{root},0}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " at both training and inference (Table 4(F)). These results indicate that not only plausible but also implausible pose-trajectory pairs are required for training " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " to serve as a surrogate for the locomotion generator. Also, both " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "j_0" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{\\mathrm{root},0}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " contribute to performance improvements. Notably, the HTP performance decreases when " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "j_0" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " is not used, demonstrating that the pose is important to estimate the plausibility of predicted trajectories." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12330" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 101, + 544, + 166 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 6. Results of zero-shot filtering by the LocoVal filter on the predictions of a pre-trained trajectory predictor [63] on the ETH/UCY dataset [24, 39]. 'Mean' represents the average performance across the 5 subsets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 101, + 544, + 166 + ], + "lines": [ + { + "bbox": [ + 64, + 101, + 544, + 166 + ], + "spans": [ + { + "bbox": [ + 64, + 101, + 544, + 166 + ], + "type": "table", + "html": "
MethodETHHOTELUNIVZARA1ZARA2Mean
ADEFDEADEFDEADEFDEADEFDEADEFDEADEFDE
Pretrained EqMotion [63]2.184.630.641.311.302.810.821.840.651.471.122.41
Ours (w/ LocoVal filter)1.412.880.611.260.932.040.801.800.641.450.881.89
Rejected trajectories8.8919.722.695.534.339.181.703.672.214.723.968.56
", + "image_path": "b6b483f9fa0b93e97d7bb68591819433605ae8c45202d8c40b19e75a3175bb62.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 184, + 205, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 184, + 205, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 205, + 196 + ], + "type": "text", + "content": "5.3. LocoVal Filter at Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 202, + 230, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 230, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 230, + 214 + ], + "type": "text", + "content": "5.3.1. LocoVal Filter on the JTA Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 219, + 295, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 295, + 255 + ], + "type": "text", + "content": "Since our LocoVal filter can be applied in a plug-and-play manner to any HTP network, we apply it to Social-Trans [47] and our method (i.e., Social-Trans trained with " + }, + { + "bbox": [ + 55, + 219, + 295, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 55, + 219, + 295, + 255 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "text", + "content": "Table 5 presents the filtering results on the JTA dataset [9] with " + }, + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "text", + "content": " set to 0.7 in Eq.(5). We observe that the LocoVal filter enhances performance across all metrics in both Social-Trans and our method. Notably, although " + }, + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "text", + "content": " is used in training " + }, + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "text", + "content": " in our method, the LocoVal filter using " + }, + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 54, + 255, + 297, + 376 + ], + "type": "text", + "content": " can further enhance the performance of our method through filtering at inference. Table 5 also shows the evaluation scores of the rejected trajectories. These scores exhibit substantially poorer performance than the filtered results, indicating that the LocoVal filter rejects implausible, inaccurate trajectories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 384, + 277, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 277, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 277, + 396 + ], + "type": "text", + "content": "5.3.2. Zero-shot Filtering on the ETH/UCY Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 401, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 473 + ], + "type": "text", + "content": "Following the aforementioned experiments on the JTA dataset [9], we further evaluate the LocoVal filter on the ETH/UCY dataset [24, 39] using another baseline, EqMotion [63]. Note that this dataset is not used for training " + }, + { + "bbox": [ + 55, + 401, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 401, + 296, + 473 + ], + "type": "text", + "content": ", and due to the bird's-eye-view of the ETH/UCY dataset, we experimented with " + }, + { + "bbox": [ + 55, + 401, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 401, + 296, + 473 + ], + "type": "text", + "content": " that does not use human poses as input." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 474, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 297, + 582 + ], + "type": "text", + "content": "Table 6 presents the results of applying the LocoVal filter under this zero-shot setting, with " + }, + { + "bbox": [ + 55, + 474, + 297, + 582 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 474, + 297, + 582 + ], + "type": "text", + "content": " set to 0.8. Although the dataset and human poses are not used for training " + }, + { + "bbox": [ + 55, + 474, + 297, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 474, + 297, + 582 + ], + "type": "text", + "content": ", we observe improvements in both ADE and FDE with the application of the LocoVal filter across all 5 subsets. While the importance of human poses can be observed in the comparison between Table 4(E) and (G), Table 6 suggests that it is even possible to identify implausible trajectories based only on the current root velocity, even without the human poses." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 597, + 125, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 125, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 125, + 609 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "content": "Analysis of Estimated Plausibility. We examine the plausibility estimated by " + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "content": " and the actual reward in the locomotion generator. To test this, 200 episodes of locomotion are generated from randomly sampled trajectories and initial humanoid states. The correlation coefficient between the estimated plausibilities and the actual rewards is 0.85. This demonstrates that " + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "content": " can serve as a surrogate for the locomotion generator, contributing to performance improvements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 184, + 556, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 556, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 556, + 293 + ], + "type": "text", + "content": "Cost-benefit Analysis. The additional training time compared to the baseline [47] is 10 hours in the first stage on a single NVIDIA RTX A6000 GPU. The additional cost in the second stage is nearly negligible because only the computation of " + }, + { + "bbox": [ + 313, + 184, + 556, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 184, + 556, + 293 + ], + "type": "text", + "content": " is added, which is a shallow MLP. Importantly, " + }, + { + "bbox": [ + 313, + 184, + 556, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 184, + 556, + 293 + ], + "type": "text", + "content": " can be used with any HTP network, which means only the HTP network needs training in the second stage. Considering these benefits, despite the potential for further cost reductions, our method is highly useful." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "text", + "content": "Infeasible Trajectory in Dataset. Some trajectories in the synthetic JTA dataset [9] are infeasible rather than implausible for a humanoid and result in low rewards. However, since our target is plausible locomotion, like most pedestrians' locomotion in daily scenes, it is reasonable not to reproduce such infeasible trajectories. During HTP training, these infeasible trajectories are deprioritized since they have low plausibilities and the HTP network prioritizes more plausible trajectories instead. This weighting mechanism can be one reason for the performance improvement of the proposed method, as shown in Tables 1, 2, and 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 453, + 556, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 556, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 556, + 536 + ], + "type": "text", + "content": "Limitations and Future Work. The smaller performance improvement on the JRDB dataset [34] compared to the JTA dataset [9] (especially in Table 3) is considered to be derived from the assumption of accurate 3D human poses, which is one limitation of the proposed method. In this perspective, future research may extend the EmLoco loss to account for the uncertainty of human poses [54, 57, 68]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 547, + 440, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 440, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 440, + 561 + ], + "type": "text", + "content": "7. Concluding Remarks" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 567, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 651 + ], + "type": "text", + "content": "This paper presented Locomotion Embodiment, a framework that evaluates the plausibility of predicted trajectories as human locomotion. Our experiments demonstrated that the proposed EmLoco loss improves HTP performance even with the state-of-the-art HTP method [47], without any inference speed deceleration. Furthermore, we established the use of the LocoVal function for trajectory filtering." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 651, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 556, + 712 + ], + "type": "text", + "content": "Importantly, our EmLoco loss improved HTP performance with limited observations even in the real-world dataset [34], demonstrating its effectiveness in real-world scenarios. We hope our findings inspire future physics-aware HTP methods for safe real-world applications." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12331" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Vida Adeli, Mahsa Ehsanpour, Ian D. Reid, Juan Carlos Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In ICCV, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "text", + "content": "[2] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social LSTM: human trajectory prediction in crowded spaces. In CVPR, 2016. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 296, + 225 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 258 + ], + "type": "text", + "content": "[4] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 294, + 294 + ], + "type": "text", + "content": "[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 297, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 296, + 316 + ], + "type": "text", + "content": "[6] Erwin Coumans. Bullet physics simulation. In SIGGRAPH, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 319, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 361 + ], + "type": "text", + "content": "[7] Markos Diomataris, Nikos Athanasiou, Omid Taheri, Xi Wang, Otmar Hilliges, and Michael J. Black. WANDR: Intention-guided human motion generation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 365, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 296, + 441 + ], + "type": "text", + "content": "[8] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R. Qi, Yin Zhou, Zoey Yang, Aurelien Chouard, Pei Sun, Jiquan Ngiam, Vijay Vasudevan, Alexander McCauley, Jonathon Shlens, and Dragomir Anguelov. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In ICCV, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 294, + 485 + ], + "type": "text", + "content": "[9] Matteo Fabbri, Fabio Lanzi, Simone Calderara, Andrea Palazzi, Roberto Vezzani, and Rita Cucchiara. Learning to detect and track visible and occluded body joints in a virtual world. In ECCV, 2018. 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 488, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 296, + 542 + ], + "type": "text", + "content": "[10] Carlos Flores, Pierre Merdrignac, Raoul de Charette, Francisco M. Navas Matos, Vicente Milanés, and Fawzi Nashashibi. A cooperative car-following/emergency braking system with prediction-based pedestrian avoidance capabilities. TITS, 20(5), 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 544, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 296, + 576 + ], + "type": "text", + "content": "[11] Zipeng Fu, Xuxin Cheng, and Deepak Pathak. Deep whole-body control: Learning a unified policy for manipulation and locomotion. In CoRL, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 296, + 632 + ], + "type": "text", + "content": "[12] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. In ICLR, 2022. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "type": "text", + "content": "[13] Francesco Giuliani, Irtiza Hasan, Marco Cristani, and Fabio Galasso. Transformer networks for trajectory forecasting. In ICPR, 2021. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[14] Ivan Grishchenko, Valentin Bazarevsky, Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Zanfir, Richard Yee, Karthik Raveendran, Matsvei Zhdanovich, Matthias Grundmann, and Cristian Sminchisescu. Blazeposition ghum holistic: Real-time" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 94 + ], + "type": "text", + "content": "3d human landmarks and pose estimation. arXiv:2206.11678, 2022. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 555, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 555, + 138 + ], + "type": "text", + "content": "[15] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In CVPR, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 554, + 184 + ], + "type": "text", + "content": "[16] Xinyang Gu, Yen-Jen Wang, Xiang Zhu, Chengming Shi, Yanjiang Guo, Yichen Liu, and Jianyu Chen. Advancing humanoid locomotion: Mastering challenging terrains with denoising world model learning. In RSS, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 228 + ], + "type": "text", + "content": "[17] Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Back to mlp: A simple baseline for human motion prediction. In WACV, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 231, + 555, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 555, + 273 + ], + "type": "text", + "content": "[18] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social GAN: socially acceptable trajectories with generative adversarial networks. In CVPR, 2018. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 276, + 555, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 555, + 307 + ], + "type": "text", + "content": "[19] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In ICCV, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 555, + 330 + ], + "type": "text", + "content": "[20] Dirk Helbing and Péter Molnár. Social force model for pedestrian dynamics. Phys. Rev. E, 51, 1995. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 332, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 364 + ], + "type": "text", + "content": "[21] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 366, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 554, + 397 + ], + "type": "text", + "content": "[22] Minkwan Kim and Yoonsang Lee. Learning human-like locomotion based on biological actuation and rewards. In SIGGRAPH, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 399, + 555, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 555, + 431 + ], + "type": "text", + "content": "[23] Parth Kothari, Sven Kreiss, and Alexandre Alahi. Human trajectory forecasting in crowds: A deep learning perspective. TITS, 23(7), 2021. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 434, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 555, + 464 + ], + "type": "text", + "content": "[24] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. In Computer graphics forum, 2007. 5, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 468, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 554, + 510 + ], + "type": "text", + "content": "[25] Yong-Lu Li, Xiaogian Wu, Xinpeng Liu, Yiming Dou, Yikun Ji, Junyi Zhang, Yixing Li, Jingru Tan, Xudong Lu, and Cewu Lu. From isolated islands to pangea: Unifying semantic space for human action understanding. arXiv:2304.00553, 2023. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 512, + 555, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 555, + 554 + ], + "type": "text", + "content": "[26] Junwei Liang, Lu Jiang, Juan Carlos Niebles, Alexander G. Hauptmann, and Li Fei-Fei. Peeking into the future: Predicting future person activities and locations in videos. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 556, + 555, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 555, + 588 + ], + "type": "text", + "content": "[27] Zhengyi Luo, Jinkun Cao, Alexander W. Winkler, Kris Kitani, and Weipeng Xu. Perpetual humanoid control for real-time simulated avatars. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 590, + 555, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 555, + 622 + ], + "type": "text", + "content": "[28] Takahiro Maeda and Norimichi Ukita. Motionaug: Augmentation with physical correction for human motion prediction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 624, + 555, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 555, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 555, + 655 + ], + "type": "text", + "content": "[29] Takahiro Maeda and Norimichi Ukita. Fast inference and update of probabilistic density estimation on trajectory prediction. In ICCV, 2023. 1, 2, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 658, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 689 + ], + "type": "text", + "content": "[30] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: archive of motion capture as surface shapes. In ICCV, 2019. 1, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[31] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller," + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12332" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 106 + ], + "type": "text", + "content": "Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac gym: High performance GPU based physics simulation for robot learning. In NeurIPS, 2021. 1, 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "type": "text", + "content": "[32] Karttikeya Mangalam, Ehsan Adeli, Kuan-Hui Lee, Adrien Gaidon, and Juan Carlos Niebles. Disentangling human dynamics for pedestrian locomotion forecasting with noisy supervision. In WACV, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 295, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 295, + 186 + ], + "type": "text", + "content": "[33] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 188, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 296, + 243 + ], + "type": "text", + "content": "[34] Roberto Martin-Martin, Mihir Patel, Hamid Rezatofighi, Abhijeet Shenoi, JunYoung Gwak, Eric Frankel, Amir Sadeghian, and Silvio Savarese. JRDB: A dataset and benchmark of egocentric robot visual perception of humans in built environments. TPAMI, 45(6), 2023. 1, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 245, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 295, + 300 + ], + "type": "text", + "content": "[35] Raul Quintero Minguez, Ignacio Parra Alonso, David Fernández Llorca, and Miguel Ángel Sotelo. Pedestrian path, pose, and intention prediction through gaussian process dynamical models and pedestrian activity recognition. TITS, 20(5), 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 301, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 295, + 345 + ], + "type": "text", + "content": "[36] Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In ICML, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 347, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 347, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 56, + 347, + 295, + 390 + ], + "type": "text", + "content": "[37] Mai Nishimura, Shohei Nobuhara, and Ko Nishino. Viewbirdformer: Learning to recover ground-plane crowd trajectories and ego-motion from a single ego-centric view. RA-L, 8(1), 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 425 + ], + "type": "text", + "content": "[38] Sang II Park, Hyun Joon Shin, and Sung Yong Shin. Online locomotion generation based on motion blending. In SIGGRAPH, 2002. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 428, + 295, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 295, + 460 + ], + "type": "text", + "content": "[39] Stefano Pellegrini, Andreas Ess, and Luc Van Gool. Improving data association by joint modeling of pedestrian trajectories and groupings. In ECCV, 2010. 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 462, + 295, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 295, + 505 + ], + "type": "text", + "content": "[40] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ToG, 40(4), 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 508, + 295, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 508, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 508, + 295, + 541 + ], + "type": "text", + "content": "[41] Tran Phong, Haoran Wu, Cunjun Yu, Panpan Cai, Sifa Zheng, and David Hsu. What truly matters in trajectory prediction for autonomous driving? In NeurIPS, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 543, + 295, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 295, + 575 + ], + "type": "text", + "content": "[42] Raul Quintero, Jorge Almeida, David Fernández Llorca, and MA Sotelo. Pedestrian path prediction using body language traits. In IV, 2014. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 578, + 295, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 295, + 620 + ], + "type": "text", + "content": "[43] Ilija Radosavovic, Tete Xiao, Bike Zhang, Trevor Darrell, Jitendra Malik, and Koushil Sreenath. Real-world humanoid locomotion with reinforcement learning. Sci. Robotics, 9(89), 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "text", + "content": "[44] Ilija Radosavovic, Bike Zhang, Baifeng Shi, Jathushan Rajasegaran, Sarthak Kamat, Trevor Darrell, Koushil Sreenath, and Jitendra Malik. Humanoid locomotion as next token prediction. arXiv/2402.19469, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[45] Davis Rempe, Zhengyi Luo, Xue Bin Peng, Ye Yuan, Kris Kitani, Karsten Kreis, Sanja Fidler, and Or Litany. Trace and pace: Controllable pedestrian animation via guided trajectory diffusion. In CVPR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "type": "text", + "content": "[46] Andrey Rudenko, Luigi Palmieri, Michael Herman, Kris M. Kitani, Dariu M. Gavrila, and Kai Oliver Arras. Human motion trajectory prediction: a survey. Int. J. Robotics Res., 39(8), 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 152 + ], + "type": "text", + "content": "[47] Saeed Saadatnejad, Yang Gao, Kaouther Messaoud, and Alexandre Alahi. Social-transmotion: Promptable human trajectory prediction. In ICLR, 2024. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "type": "text", + "content": "[48] Amir Sadeghian, Vineet Kosaraju, Ali Sadeghian, Noriaki Hirose, Hamid Rezatofighi, and Silvio Savarese. Sophie: An attentive gan for predicting paths compliant to social and physical constraints. In CVPR, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[49] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In ECCV, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 244, + 553, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 244, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 316, + 244, + 553, + 287 + ], + "type": "text", + "content": "[50] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. RA-L, 8(11), 2023. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 289, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 289, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 553, + 312 + ], + "type": "text", + "content": "[51] Christoph Schöller and Alois Knoll. Flomo: Tractable motion prediction with normalizing flows. In IROS, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 313, + 553, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 313, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 316, + 313, + 553, + 345 + ], + "type": "text", + "content": "[52] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv:1707.06347, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 347, + 553, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 347, + 553, + 379 + ], + "spans": [ + { + "bbox": [ + 316, + 347, + 553, + 379 + ], + "type": "text", + "content": "[53] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 381, + 553, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 381, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 316, + 381, + 553, + 424 + ], + "type": "text", + "content": "[54] Megh Shukla, Roshan Roy, Pankaj Singh, Shuaib Ahmed, and Alexandre Alahi. VL4pose: Active learning through out-of-distribution detection for pose estimation. In BMVC, 2022, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 427, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 460 + ], + "type": "text", + "content": "[55] Jianhua Sun, Yuxuan Li, Liang Chai, Haoshu Fang, Yong-Lu Li, and Cewu Lu. Human trajectory prediction with momentary observation. In CVPR, 2022. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "type": "text", + "content": "[56] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: generating 4d whole-body motion for hand-object grasping. In CVPR, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 497, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 553, + 529 + ], + "type": "text", + "content": "[57] Hiromu Taketsugu and Norimichi Ukita. Active transfer learning for efficient video-specific human pose estimation. In WACV, 2024. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 531, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 563 + ], + "type": "text", + "content": "[58] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 566, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 598 + ], + "type": "text", + "content": "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 601, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 643 + ], + "type": "text", + "content": "[60] Jingbo Wang, Ye Yuan, Zhengyi Luo, Kevin Xie, Dahua Lin, Umar Iqbal, Sanja Fidler, and Sameh Khamis. Learning human dynamics in autonomous driving scenarios. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "type": "text", + "content": "[61] Jingbo Wang, Zhengyi Luo, Yixuan Li, Ye Yuan, and Bo Dai. Pacer+: On-demand pedestrian animation controller in driving scenarios. In CVPR, 2024. 1, 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[62] Jungdam Won, Deepak Gopinath, and Jessica K. Hodgins. Physics-based character controllers using conditional vaes. ToG, 41(4), 2022. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12333" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 478 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[63] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "text", + "content": "[64] Zhuo Xu, Rui Zhou, Yida Yin, Huidong Gao, Masayoshi Tomizuka, and Jiachen Li. MATRIX: multi-agent trajectory generation with diverse contexts. In ICRA, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 184 + ], + "type": "text", + "content": "[65] Takuma Yagi, Karttikeya Mangalam, Ryo Yonetani, and Yoichi Sato. Future person localization in first-person videos. In CVPR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 217 + ], + "type": "text", + "content": "[66] Shuai Yi, Hongsheng Li, and Xiaogang Wang. Understanding pedestrian behaviors from stationary crowd groups. In CVPR, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 219, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 295, + 251 + ], + "type": "text", + "content": "[67] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 285 + ], + "type": "text", + "content": "[68] Yufei Zhang, Hanjing Wang, Jeffrey O. Kephart, and Qiang Ji. Body knowledge and uncertainty modeling for monocular 3d human body reconstruction. In ICCV, 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 287, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 295, + 319 + ], + "type": "text", + "content": "[69] Yufei Zhang, Jeffrey O. Kephart, and Qiang Ji. Incorporating physics principles for precise human motion prediction. In WACV, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "type": "text", + "content": "[70] Hang Zhao, Jiyang Gao, Tian Lan, Chen Sun, Benjamin Sapp, Balakrishnan Varadarajan, Yue Shen, Yi Shen, Yuning Chai, Cordelia Schmid, Congcong Li, and Dragomir Anguelov. TNT: target-driven trajectory prediction. In CoRL, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 366, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 295, + 399 + ], + "type": "text", + "content": "[71] Jingbo Zhao, Zhetao Wang, Yiqin Peng, and Yaojun Wang. Real-time generation of leg animation for walking-in-place techniques. In SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 400, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 295, + 431 + ], + "type": "text", + "content": "[72] Bolei Zhou, Xiaou Tang, and Xiaogang Wang. Learning collective crowd behaviors with dynamic pedestrian-agents. IJCV, 111(1):50-68, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 434, + 295, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 478 + ], + "type": "text", + "content": "[73] Brian D Ziebart, Nathan Ratliff, Garratt Gallagher, Christoph Mertz, Kevin Peterson, J Andrew Bagnell, Martial Hebert, Anind K Dey, and Siddhartha Srinivasa. Planning-based prediction for pedestrians. In IROS, 2009. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12334" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/d8bec7dd-2170-41a3-a07a-e8a3813c711e_content_list.json b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/d8bec7dd-2170-41a3-a07a-e8a3813c711e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cb8f561aaa09ec4e74fd89bc974e1c2d3ac49f27 --- /dev/null +++ b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/d8bec7dd-2170-41a3-a07a-e8a3813c711e_content_list.json @@ -0,0 +1,1317 @@ +[ + { + "type": "text", + "text": "PhysicsGen: Can Generative Models Learn from Images to Predict Complex Physical Relations?", + "text_level": 1, + "bbox": [ + 232, + 130, + 763, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Martin Spitznagel1, Jan Vaillant2 and Janis Keuper1,3 \n1IMLA, Offenburg University, 2Herrenknecht AG, 3Mannheim University \n{first.last}@hs-offenburg.de", + "bbox": [ + 205, + 202, + 789, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The image-to-image translation abilities of generative learning models have recently made significant progress in the estimation of complex (steered) mappings between image distributions. While appearance based tasks like image inpainting or style transfer have been studied at length, we propose to investigate the potential of generative models in the context of physical simulations. Providing a dataset of 300k image-pairs and baseline evaluations for three different physical simulation tasks, we propose a benchmark to investigate the following research questions: i) are generative models able to learn complex physical relations from input-output image pairs? ii) what speedups can be achieved by replacing differential equation based simulations? While baseline evaluations of different current models show the potential for high speedups (ii), these results also show strong limitations toward the physical correctness (i). This underlines the need for new methods to enforce physical correctness. Data, baseline models and evaluation code: http://www.physics-gen.org.", + "bbox": [ + 88, + 325, + 483, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 646, + 220, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The numerical simulation of physical processes and relations is a crucial tool in a wide range of scientific and engineering tasks. Traditionally depending on the solution of various types of differential equations, research of physical simulation methods is increasingly shifting towards data-driven and learnable approaches [9].", + "bbox": [ + 89, + 674, + 482, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we focus on a narrow but practically important sub-field: the inference of physical relations from images. Extending Tenenbaum's seminal idea of a intuitive physics engine [2], where machine learning models have been trained to predict coarse physical properties from images, we investigate the capabilities of modern generative models to learn detailed, complex and physically correct mappings from image pairs. Given such abilities, generative models like GANs [15] and Diffusion models [19] could be", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3433c8f4aee8bc1cbab065d16d4233edf15af085231732bc8fe61661585766f.jpg", + "image_caption": [ + "Figure 1. Overview of the physical problems, baseline generative models and their evaluation. A: We introduce three complex physical simulation tasks with 100k input-output image pairs each, providing ground truth simulations based on differential equations with varying complexity. B: We evaluate all tasks on independently trained image translation models; only results for the sound propagation task are visualized in this figure. C: While the evaluation of the baseline models shows a general ability of generative image models to learn physical relations from images, we observe significant performance drops for tasks that require a higher order term in the differential equations of their simulation." + ], + "image_footnote": [], + "bbox": [ + 519, + 291, + 897, + 724 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11125", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "utilized to introduce hard to formalize but observable relations into physical models. Due to the fast inference times, generative models could also be used to accelerate simulations [34].", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Introducing the PhysicsGen Benchmark. A first step towards such data-driven models is a thorough analysis of the capabilities of current generative approaches. Hence, we introduce a novel benchmark that allows to systematically investigate the inference of a diverse set of physical problems from images (as visualized in Fig. 1):", + "bbox": [ + 89, + 152, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce PhysicsGen, a novel Benchmark for the evaluation of generative image models learning to infer physical simulations tasks from images.", + "- PhysicsGen provides a Dataset with a total of 300k image-pairs, representing three diverse physical simulation problems: i) iterative wave propagation (Sec. 3), ii) closed form lens distortion (Sec. 4) and iii) time series prediction of motion dynamics (Sec. 5).", + "- We provide a baseline evaluation that measures speedups and physical accuracy compared to full physical simulations across multiple generative image-to-image models. The evaluated models include various generative image-to-image models: Generative Adversarial Networks (GANs) [16], specifically Pix2Pix [20]; an U-Net [31]; Convolutional Autoencoders (convAE) and Variational Autoencoders (VAEs) [25]. We also assess diffusion models, including Denoising Diffusion Probabilistic Models (DDPMs) [18], Stable Diffusion [30], and Denoising Diffusion Bridge Models (DDBMs) [38]. (see Appendix A for details of the model architectures and training - the model architectures and training procedures are identical for all three tasks).", + "- Our Analysis of the baseline results shows potentially high speed-ups at good accuracies for simple 1st order simulation tasks. However, our experiments also show that current generative models have fundamental problems to learn higher order physical relations." + ], + "bbox": [ + 89, + 244, + 482, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1. Related Work.", + "text_level": 1, + "bbox": [ + 91, + 664, + 238, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The integration of learnable methods into physical modeling is a wide field with a rich body of literature which is very hard to capture in the limited space of this paper. In the following, we try to refer to the most important works in direct relation to our benchmark and some more general key concepts. We refer the reader to [34] and [9] for a comprehensive overview of the field.", + "bbox": [ + 89, + 688, + 480, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative models for physical simulations. The integration of physical principles within generative models is an active and burgeoning area of research. Recent models such as PUGAN [11] and FEM-GAN [1] have demonstrated the potential of combining GANs with physical modeling to enhance performance in environments governed by complex physical laws. Advances in fluid dynamics and", + "bbox": [ + 89, + 795, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "structural system identification using physics-guided GANs have shown significant improvements in efficiency and precision by incorporating physics-based loss functions and simulations [24, 36].", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Predicting physical relations. Also, significant progress has been made in understanding physics through machine learning. Initial research demonstrated that models could comprehend the dynamics of block towers, progressing beyond simple memorization to genuinely understanding physical interactions [27]. Practical applications have been explored, including fall detection using body part tracking through machine learning [28]. Additionally, advancements in generating physically plausible human animations highlight the potential of incorporating physical principles into machine learning frameworks [37].", + "bbox": [ + 511, + 152, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Grey-Box Models. Physics-guided approaches in AI have laid a solid foundation for incorporating physical laws into model training. Prior research has demonstrated the effectiveness of embedding physical constraints within AI models to enhance their performance and reliability [10, 22]. In generative models, diffusion models have shown promise but often face challenges when generating complex geometrical structures due to issues like multi-modality and noise sensitivity. The development of Geometric Bayesian Flow Networks addresses these challenges by enhancing the robustness and efficiency of molecule geometry modeling [32].", + "bbox": [ + 511, + 316, + 903, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physical Simulations for Reinforcement Learning. Deep reinforcement learning settings benefit from a substantial availability of physics-based environments and datasets, which facilitate the training of models to understand and predict physical phenomena [13]. Resources such as OpenAI Gym, Robosuite, and PyBullet provide robust platforms for developing and testing reinforcement learning models in simulated physical settings [6, 17, 39]. However, in the domain of generative AI, there is a notable scarcity of physics-informed datasets and environments, limiting the ability to train models on complex physical systems.", + "bbox": [ + 511, + 500, + 903, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Benchmark Tasks Overview", + "text_level": 1, + "bbox": [ + 511, + 691, + 772, + 707 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Selecting an appropriate set of tasks for the PhysicsGen benchmark is crucial to comprehensively evaluate the capabilities of generative image models in inferring complex physical relations. Given the vast diversity of possible physical simulation tasks, it is impractical to encompass the entire spectrum. Therefore, we strategically selected three distinct tasks based on the following criteria:", + "bbox": [ + 511, + 717, + 903, + 823 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Diversity of Physical Problems: To cover different types of physical phenomena, ensuring that models are tested across various domains.", + "- Variety of Numerical Solution Strategies: Incorporating both iterative and closed-form (non-iterative) solvers" + ], + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11126", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to assess models' adaptability to different computational approaches.", + "bbox": [ + 102, + 90, + 480, + 119 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Simplicity of Evaluation Metrics: Facilitating straightforward and effective model performance evaluation.", + "- Accessibility to the ML Community: Choosing intuitively understandable tasks without requiring specialized physics knowledge, thereby making the benchmark widely applicable to the machine learning community." + ], + "bbox": [ + 91, + 121, + 480, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on these objectives, we introduce the following three simulation tasks within the PhysicsGen benchmark:", + "bbox": [ + 89, + 215, + 480, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Urban Sound Propagation: This task models the behavior of sound waves in urban environments, accounting for phenomena such as diffraction and reflections (Sec. 3). By providing 100,000 input-output image pairs based on differential equation-based simulations, it challenges generative models to accurately capture intricate sound distribution patterns influenced by urban structures [26, 33].", + "bbox": [ + 89, + 244, + 480, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lens Distortion: Focusing on optical aberrations, this task utilizes the Brown-Conrady distortion model to simulate lens-induced geometric distortions in images (Sec. 4). With 100,000 image pairs depicting varying degrees of distortion, models are tasked with learning precise, closed-form physical relations that alter image geometry based on lens parameters.", + "bbox": [ + 89, + 351, + 480, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamics of Rolling and Bouncing Movements: This task involves simulating the motion of a rolling or bouncing ball on an inclined surface, incorporating both linear and rotational dynamics (Sec. 5). Providing 100,000 image pairs that capture the ball's position and rotation over time, it assesses models' ability to predict time-series dynamics and handle higher-order physical relations inherent in motion equations.", + "bbox": [ + 89, + 457, + 480, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Collectively, these tasks offer a diverse framework for assessing the strengths and limitations of generative models in learning and replicating complex physical simulations from image data. The subsequent sections delve into each task in detail, outlining dataset generation, underlying physics, evaluation metrics, and baseline model performances.", + "bbox": [ + 89, + 579, + 480, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Urban Sound Propagation", + "text_level": 1, + "bbox": [ + 89, + 684, + 334, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The simulation of wave propagation in complex scenes, including diffraction and reflection at multiple objects is a representative task for a wide range of physical problems which are typically solved via (expensive) iterative solving of higher order differential equations. Due to its intuitive setting and accessible data, we investigate wave propagation problems by a concrete example where sound is propagated in urban environments.", + "bbox": [ + 89, + 709, + 480, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset. We sampled $25k$ geo-locations across 10 cities, leveraging open-source geo-data from the Overpass API1 and processed it with GeoPandas [21] to represent urban en", + "bbox": [ + 89, + 830, + 480, + 876 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "virements in a $500\\mathrm{m}^2$ area with buildings and open spaces encoded in black and white pixels, respectively. Our selection criteria ensured diverse urban scenarios by mandating a minimum of 10 buildings within 200 meters and a 50-meter clearance from buildings to the sample point. We utilized the NoiseModelling v4 framework [4], adhering to the CNOSSOS-EU standards [23], to simulate sound propagation maps, placing sound receivers on a grid with a step-size of 5 meters and at building boundaries, excluding indoor placements (see Appendix B.3 for details). The resulting simulation output was then interpolated into $512\\mathrm{x}512$ or $256\\mathrm{x}256$ pixel maps, representing decibel levels on a $[0,255]$ scale. Four simulation tasks (Baseline, Diffraction, Reflection, Combined) were defined to model sound propagation under varying conditions, accumulating in a dataset of 100,000 samples. The Baseline task models a constant $95\\mathrm{dB}$ noise source at $500\\mathrm{Hz}$ , excluding diffraction and reflection. The Diffraction task adds horizontal sound wave diffraction, the Reflection task simulates sound reflections with a standard absorption coefficient, and the Combined task introduces variable sound levels (60-115 dB) and environmental factors, incorporating both reflection and diffraction to mimic realistic urban sound behavior. Refer to Fig. 2 and Appendix B.6 for a detailed description and visualization of the dataset generation pipeline.", + "bbox": [ + 511, + 90, + 903, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Physics of Sound Propagation. Mathematically, the propagation of sound over time is described via partial differential wave equations. Due to space constraints and the practical nature of our problem setting, we will neglect the derivation from continuous wave equations and directly focus on the discrete and iterative implementations of sound propagation which have been applied for our ground-truth simulations. Following [35], for a discrete set of receivers $R$ , the amplitude $L_{R_k}^j$ of receiver $R_k$ at frequency $j$ is computed via iterative differences:", + "bbox": [ + 511, + 468, + 903, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {R _ {k}} ^ {j} = L _ {W} ^ {j} - A _ {\\text {d i v} R _ {k}} - A _ {\\text {a t m} R _ {k}} ^ {j} - A _ {\\text {d i f} R _ {k}} ^ {j} \\quad (j, k \\in \\mathbb {N}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 625, + 903, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L_W^j$ represents the source level, $A_{div_{R_k}}$ captures the geometrical spreading, $A_{atm_{R_k}}^j$ denotes the atmospheric absorption, and $A_{dif_{R_k}}^j$ models diffraction. Note that the ground effect $A_{grd_{R_k}}^j$ is neglected in our study.", + "bbox": [ + 511, + 654, + 903, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, the model accounts for reflections by adjusting the power level $L_W^j$ based on the absorption coefficient $\\alpha_{vert}$ of the surfaces involved. This adjustment is performed using the equation", + "bbox": [ + 511, + 732, + 903, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {W} ^ {(n _ {r e f})} = L _ {W} ^ {(n _ {r e f} - 1)} + n _ {r e f} \\times 1 0 \\log_ {1 0} (1 - \\alpha_ {v e r t}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 797, + 903, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n_{ref}$ indicates the number of reflections considered. Specular reflections are modeled using the image receiver method, which provides a computationally efficient way to account for the angle of incidence being equal to the angle of reflection [35]. The sound level at each receiver", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "http://overpass-api.de/api/map", + "bbox": [ + 107, + 886, + 354, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11127", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b4a0020648eab743670178776b22464d496d013addaa94f228fd2952c73b1b06.jpg", + "image_caption": [ + "Figure 2. The sampling pipeline for the sound propagation dataset utilizes the NoiseModelling framework [4] to generate sound propagation maps based on specific urban layouts. The generators are then trained to replicate these sound propagation patterns for given locations and source parameters. Predictions are evaluated by specifically analyzing errors in relation to the line of sight (see Appendix B.1 for details)." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 602, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reflects the cumulative effect of direct, diffracted, and reflected sound paths. As the number of reflections increases, the complexity of calculating the power level $L_W^j$ also increases, growing in $O(N^2)$ , where $N$ denotes the number of reflective surfaces considered in the computation.", + "bbox": [ + 88, + 227, + 480, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation. Our baseline evaluation focuses on the ability of standard image-to-image generative models to accurately predict sound propagations. Prediction quality is determined by the pixel-wise difference between actual and predicted sound distributions. We use Mean Absolute Error (MAE) to quantify the average magnitude of prediction errors. Additionally, we introduce Weighted Mean Absolute Percentage Error (wMAPE) to specifically penalize errors in predictions that inaccurately show high sound amplitudes in regions expected to have low amplitudes, such as areas behind buildings. To evaluate model performance in capturing sound reflections or diffractions, metrics were specifically calculated for areas both within and outside the direct line of sight (LoS and NLoS) from the sound source. This methodology allows us to assess model predictions for sound propagations through direct paths as well as reflections and diffractions.", + "bbox": [ + 93, + 304, + 483, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analysis. The results in Tab. 1 show that generative models significantly outperform traditional sound propagation simulations in processing speed, achieving up to a $20k$ factor improvement in runtime (see Tab. 9 in the Appendix for further analysis). Regarding the prediction quality, we observe that models encountering tasks with reflections or diffractions show an increase in MAE for NLoS regions across", + "bbox": [ + 89, + 561, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/42bfc572d48512a9a46443aad13f8ef05dc6fd99c2d2bddce6041606c71f34ed.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation across all tasks for all architectures with a batch size of 16 during inference. The complete results containing the Combined task are given in Appendix B.7. Best overall results in bold, second best underlined." + ], + "table_footnote": [], + "table_body": "
ModelTaskMAE ↓wMAPE ↓Runtime/ ↓ Sample (ms)
LoSNLoSLoSNLoS
SimulationBase0.000.000.000.00204700
convAEBase3.672.7420.2467.130.128
VAE [25]Base3.922.8421.3375.580.124
UNet [31]Base2.291.7312.9137.570.138
Pix2Pix [20]Base1.731.199.366.750.138
DDPM [18]Base2.423.2615.5751.083986.353
SD(w,CA) [30]Base3.763.3417.4235.182961.027
SDBase2.121.0813.2332.462970.86
DDBM [38]Base1.612.1717.5065.243732.21
SimulationDif.0.000.000.000.00206000
convAEDif.3.598.0413.7732.090.128
VAEDif.3.928.2214.4632.570.124
UNetDif.0.943.274.2222.360.138
Pix2PixDif.0.913.363.5118.060.138
DDPMDif.1.593.278.2520.303986.353
SD(w,CA)Dif.2.467.7210.1431.232961.027
SDDif.1.335.078.1524.452970.86
DDBMDif.1.353.3511.2223.563732.21
SimulationRefl.0.000.000.000.00251000
convAERefl.3.836.5620.6793.540.128
VAERefl.4.156.3221.5792.470.124
UNetRefl.2.295.7212.7580.460.138
Pix2PixRefl.2.144.7911.3030.670.138
DDPMRefl.2.747.9317.8580.383986.353
SD(w,CA)Refl.3.816.8219.7881.612961.027
SDRefl.2.535.2615.0455.272970.86
DDBMRefl.1.936.3818.3479.133732.21
", + "bbox": [ + 516, + 292, + 903, + 645 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/92c5beb2928dbe44ce383c4da7fbc92cb7d8fbe07a70808739e69d48f2cebc24.jpg", + "image_caption": [ + "Figure 3. Qualitative results comparing the ground-truth simulation with the prediction for a single sample within the reflection task. Additional results can be found in Appendix B.7" + ], + "image_footnote": [], + "bbox": [ + 94, + 694, + 478, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "all tested architectures. For the Pix2Pix model this rise in NLoS error is moderate, while this effect is more pronounced in the UNet and Diffusion models. Models without skip connections, such as VAE and ConvAE, tend to blur most of the image and struggle to capture any reflection patterns accurately. This trend highlights how reflections and diffractions introduce complex higher-order dependencies (see Eq. (12) and Appendix B.4) that complicate the image-to-image sound propagation tasks. Although the models successfully replicate sound reflection patterns visually (see Fig. 3), there exists a significant gap between its visual outputs and the actual acoustic accuracy. This discrepancy suggests that while visually detailed, the models do not fully capture the complexities of sound physics.", + "bbox": [ + 511, + 689, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11128", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Lens Distortion", + "text_level": 1, + "bbox": [ + 89, + 89, + 246, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our second benchmark is designed to investigate a class of physical problems for which solutions can be computed directly via closed equations, rather than iterative solvers. We find a representative and intuitive task in the modeling of optical lens distortions [7, 14], a common problem in computer vision. Generative models are trained to simulate the appearance of faces distorted by lens effects, using the Brown-Conrady [12] model. These distortions are computed using OpenCV [5], providing ground truth for training and evaluation of generative models.", + "bbox": [ + 88, + 114, + 482, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We leverage a subset of the CelebA dataset [29], comprising 50,000 facial images, which were cropped to center the faces within a $256 \\times 256$ pixel frame. The resulting images where split into two distinct groups to isolate the effects of tangential distortion: one focusing on horizontal distortion and the other on vertical distortion. This allows us to specifically measure the impact of each type of distortion on the accuracy of the generative models (see Fig. 4 and Appendix C for more details).", + "bbox": [ + 88, + 266, + 482, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Physics of Lens Distortions. The Brown-Conrady model is a widely adopted method to describe and correct lens distortions. It provides a mathematical framework to represent both radial and tangential distortions through a series of coefficients. For our study, we concentrate exclusively on the tangential distortion components, utilizing relevant coefficients to model the distortion characteristics of specific lenses. The mathematical representation of tangential distortion is expressed through the following equation, where $x_{dist}$ and $y_{dist}$ denote the distorted coordinates:", + "bbox": [ + 89, + 402, + 483, + 553 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {d i s t} = x + \\left[ 2 p _ {1} x y + p _ {2} \\left(r ^ {2} + 2 x ^ {2}\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 563, + 482, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {d i s t} = y + \\left[ p _ {1} \\left(r ^ {2} + 2 y ^ {2}\\right) + 2 p _ {2} x y \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 599, + 421, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r^2 = x^2 + y^2$ represents the squared radial distance from the image center, and $p_1$ and $p_2$ are the coefficients that model the tangential distortion effects. By strategically adjusting $p_1$ and $p_2$ and isolating their effects via controlled nullification, we can identify how each parameter individually contributes to the image distortion:", + "bbox": [ + 89, + 643, + 483, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {d i s t} = \\left\\{ \\begin{array}{l l} x + 2 p _ {1} x y & \\text {i f} p _ {2} = 0, \\\\ x + p _ {2} \\left(r ^ {2} + 2 x ^ {2}\\right) & \\text {i f} p _ {1} = 0, \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 746, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {d i s t} = \\left\\{ \\begin{array}{l l} y + p _ {1} (r ^ {2} + 2 y ^ {2}) & \\text {i f} p _ {2} = 0, \\\\ y + 2 p _ {2} x y & \\text {i f} p _ {1} = 0. \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 790, + 419, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These conditional equations illustrate the isolated impact of each coefficient, revealing that the correction complexity for $y_{dist}$ escalates with the term $p_1(r^2 + 2y^2)$ when $p_1$ is non-zero and $p_2$ is zero. Conversely, the complexity of $x_{dist}$", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "correction intensifies with $p_2(r^2 + 2x^2)$ under the condition that $p_2$ is non-zero and $p_1$ is zero.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. Given the challenges of accurately capturing lens distortion through traditional pixel-based metrics, our evaluation strategy is specifically designed to focus on accurately reproducing designated landmarks in the images. We detect facial landmarks using a 2D facial landmark detection method based on the Facial Alignment Network (FAN) [8], applied to the faces in the CelebA dataset (see Appendix C for details). We use these landmarks to assess the geometric distortions introduced by lens settings, represented by the coefficients $p_1$ and $p_2$ . We then compute the Euclidean distance between the actual position of a landmark in the true distorted image and its predicted position in the image generated by the models as a quality metric. This measurement is performed separately for the horizontal (X) and vertical (Y) coordinates, resulting in X Error and Y Error respectively. This approach not only allows us to gauge the accuracy of the models in simulating the specific effects of lens distortion on facial features but also helps in understanding how different distortion coefficients impact the positional accuracy of landmarks along each axis.", + "bbox": [ + 511, + 121, + 905, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis. Our experiments with the CelebA dataset under varying conditions of $p_1$ and $p_2$ showcase different patterns in error rates that align with the theoretical complexity of the distortion model. As shown in Tab. 2 and Fig. 5, the generative models exhibit differing accuracies in landmark prediction under the influence of lens distortion parameters $p_1$ and $p_2$ . Notably, the increased complexity in $x_{dist}$ and $y_{dist}$ equations under separate conditions ( $p_1$ nonzero vs. $p_2$ nonzero) slightly correlates with heightened error rates in respective dimensions. The models exhibit higher Y Error when only $p_1$ is active, indicative of the complex correction challenge introduced by the higher order dependency on $y$ . Similarly, a slight increase in X error and a slight decrease in Y error are observed when $p_2$ is exclusively nonzero.", + "bbox": [ + 511, + 422, + 905, + 635 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cc25ea096a192917527fde18b87997a2ecf22ccdbb5c567c7e5ddc1765986527.jpg", + "image_caption": [ + "Figure 4. Sampling and evaluation pipeline for the lens distortion dataset. The Brown-Conrady distortion model generates the true distorted images based on parameters $p_1$ and $p_2$ (we depict a chess pattern for visualization). The conditioned generators are then trained to replicate these distortions for given images and parameters. The models are evaluated by comparing predicted against true facial landmarks, using a 2D facial landmark detection based on the Facial Alignment Network (FAN) [8]." + ], + "image_footnote": [], + "bbox": [ + 516, + 638, + 903, + 781 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11129", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/83fe1329682073524b4b8ea7e8f635861464b06e3fc097388d51e2dab1135714.jpg", + "image_caption": [ + "Figure 5. Qualitative visualization of lens distortion predictions using different generative models on a CelebA dataset sample. $p_2$ distortion: Original image at top-left; subsequent panels show U-Net, Pix2Pix, Diffusion models. Red dots mark actual landmark positions and blue dots for predictions." + ], + "image_footnote": [], + "bbox": [ + 114, + 89, + 460, + 281 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/077749b4ebfe3668979b4322e3609bd87492f1e9ed12bf98ca52b12a88153324.jpg", + "table_caption": [ + "Table 2. Quantitative X-Y error and shift comparison across different models for varying activation of parameters $p_1$ and $p_2$ with a batch size of 16 during inference. Best overall results in bold, second best underlined." + ], + "table_footnote": [], + "table_body": "
ModelComb ↓X ↓ ErrY ↓ Err|X-Y| ShiftRuntime/ Sample (ms)
Sim.0.000.000.000.00153.205
convAE11.936.758.131.380.110
VAE11.536.557.831.280.122
UNet2.821.282.150.870.118
Pix2Pix2.000.991.430.440.122
DDPM1.930.941.390.453970.603
SD(w,CA)3.091.592.210.622991.678
SD2.791.412.010.602997.576
Sim.p1 = 0, p2 ≠ 0
convAE0.000.000.000.00153.205
10.568.354.772.210.110
VAE10.408.264.623.640.122
UNet2.361.331.600.270.117
Pix2Pix1.771.021.140.130.123
DDPM2.131.391.230.163970.603
SD(w,CA)2.851.601.940.342991.678
SD2.441.381.640.262997.576
", + "bbox": [ + 114, + 445, + 460, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Dynamics of bouncing movements", + "text_level": 1, + "bbox": [ + 89, + 707, + 401, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The third physical problem under investigation is the movement of a rolling or bouncing ball. The aim here is to evaluate the ability of generative models to map kinematic movements of physics - a task which should generalize well to the investigation of various time series problems. Specifically, we define the task to predict the position and rotation of a ball along an inclined surface for a defined time interval after the input image. Only the bouncing case is discussed below; however, all results for the rolling ball can be found in Appendix D.3.", + "bbox": [ + 89, + 734, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We generated our data set with Pymunk [3], a", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8dbd5c8c665714692e47243196b5096b0fd8654c58e8739a751a2c945931c44b.jpg", + "image_caption": [ + "Figure 6. Physical equations describing the bouncing ball movement divided in 3 parts: free fall (green), impact & bounce of the ball on the ground (red dot) and oblique throw with initial velocity from ball bounce (orange)." + ], + "image_footnote": [], + "bbox": [ + 553, + 90, + 869, + 186 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {g} = m * a \\Leftrightarrow m * \\ddot {y} = - m * g \\Leftrightarrow \\ddot {y} = - g \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 204, + 906, + 222 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} m * \\ddot {y} + d * \\dot {y} - c * (r - y) = - m * g \\tag {6} \\\\ m * \\ddot {x} = m * g * \\sin (\\beta) \\text {a l o n g x a x i s} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 237, + 906, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Python physics simulation toolbox. For the bouncing case we use 50k pairs of training images (256x256px) and perform the evaluation of the generative networks with 1600 images. Three parameters on the input images are variable: the ground inclination, the start height of the ball movement and the ball position. Additionally, the time interval between the input and target images is also variable. All other physical properties of the problem (ball's mass or friction coefficients) are kept constant. A multi-coloured structure with randomly distributed lines is superimposed on the simulation images in order to obtain greater diversity in the training data and is also used to analyze consistency, e.g. how well the generative networks can map fine multicoloured patterns (see Fig. 7). The total time required to create a sample using the available physics simulation CPU code (with a customized background) is 3.8 seconds, while the inference times for the generative models remain consistent with those reported in Tab. 1.", + "bbox": [ + 511, + 386, + 906, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Physics of bouncing balls. To model the physics behind the movement of a bouncing ball, the motion is divided into different sections which are represented with different colors in Fig. 6. First, the ball is released from a defined height without any velocity and falls down according to the Newton's law of the free fall (green part) (see equation 5). The impact of the ball on the ground and its rebound can be described in simplified terms by a spring-damper model. The aim of this second part of the movement (red dot) is to calculate the rebound speed and angle. This is done with equation 6, which contains the constant of the elements of the simplified model (see details in Appendix D.1) and the angle corresponds to the ground inclination for the first impact. Finally, the orange part of the movement is assumed to be an oblique throw with the initial velocity $v_{0}$ calculated with equation 6. The equations of motion describing", + "bbox": [ + 511, + 657, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11130", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/294f33bcc5ed3449a6de618b8c53515dc21024d86629cf3012bb64ce2fd83e2c.jpg", + "image_caption": [ + "Figure 7. The sampling and evaluation pipeline for the rolling and bouncing ball movements simulates ball trajectories. A physics-based model is used to predict the ball's movement. Evaluations focus on the accuracy of predictions related to key movement metrics, including bounce height and horizontal displacement (see Appendix D for details)." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 604, + 188 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "this oblique throw along the X and Y axes are also derived from the Newton's law (see equation 7). When the ball hits the ground for the second time, the principle \"angle of incidence = angle of reflection\" is used to determine the angle of rebound $\\theta$ and the next part of the motion is again defined by equations 6 and 7.", + "bbox": [ + 88, + 214, + 480, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation. The generated images are examined according to 4 precise error criteria: the ball position split in X and Y directions, where the error is defined as the number of pixels between the theoretical and the predicted ball center. The second error measure is the ball rotation describing its angle in degrees during the movement due to the inclined ground surface. The next quality indicator of the algorithms is the out-of-roundness of the ball also measured in pixels by taking all identified points of the ball contour and calculating the difference between the theoretical ball radius of 15px and the predicted one. The standard deviation of all radius errors is taken as the error measure. Finally, the correct representation of the ground inclination is checked: the angular error between simulation and prediction is also measured in degrees. In addition, there are three \"general\" error criteria that give preciser indications of the correct representation of the physics which are explained with the complete results in the Appendix D.", + "bbox": [ + 93, + 305, + 482, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results of the seven examined generative algorithms are summarized in table 3 (only for the four most important error criteria). The last column of the table contains the average number of non-evaluable predicted images over the four error criteria. A detailed analysis of the advantages and disadvantages of the individual methods with all the evaluation parameters can be found in the Appendix D, as well as the evaluations obtained for individual runs for the rolling and bouncing case.", + "bbox": [ + 89, + 577, + 482, + 713 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/72ce5141c374744b552daf989718985e67c873d583655e5d1434bfb10df33fc8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Position XPosition YRotationRoundnessError
convAE4.24 ± 3.96.08 ± 5.912.2 ± 8.61.06 ± 0.099%
VAE4.69 ± 6.16.25 ± 6.931.0 ± 400.90 ± 0.195%
UNet5.53 ± 7.510.8 ± 1215.2 ± 230.74 ± 0.228%
Pix2Pix6.28 ± 8.011.7 ± 1317.2 ± 210.56 ± 0.111%
DDPM7.91 ± 9.015.5 ± 1432.9 ± 340.61 ± 0.25.7%
SD(w.CA)40.0 ± 4924.8 ± 2361.1 ± 520.53 ± 0.27.3%
SD8.55 ± 1216.2 ± 1434.2 ± 380.47 ± 0.12%
", + "bbox": [ + 91, + 731, + 483, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Prediction results for the bouncing ball experiment in mean±std and mean percentage of non-evaluable predictions for the four criteria (best overall results in bold). Average runtime of the simulation: 3.8s, generative models: same as inTab. 1", + "bbox": [ + 89, + 840, + 483, + 897 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, the GAN (Pix2Pix model) provides the most stable representation of the bouncing ball problem. It achieves the best results across all evaluation criteria, with the most reliable physics predictions and fewest failed samples. The UNet also represents the ball position and angle on the generated samples very well, but produces more error images that cannot be evaluated with mostly either no ball or only a shredded shape. The greatest weakness of this algorithm lies in the clean representation of the ball roundness. The auto-encoder approaches (VAE and convAE) produce very blurred images in which the ball is often distorted and hardly recognizable. Therefore, the mean error values are not very representative, as over $90\\%$ of the predictions cannot be analyzed.", + "bbox": [ + 511, + 214, + 903, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The denoising diffusion approach (DDPM) produces slightly poorer results in absolute terms. Especially the prediction of the ball rotation is worse, with an error approximately twice as high compared to the two other approaches. Finally, to test another type of diffusion, two variations of stable diffusion were compared. The one with cross-attention is generating more error pictures and show high error values. In contrast, the normal stable diffusion performs very well: the mapping of the position is almost as accurate as with the Pix2Pix approach and the representation of the ball roundness is the best. In addition, the network generates very few error images, so that it could almost be placed on the same level as the GAN but comes as second due to a poorer angle mapping.", + "bbox": [ + 511, + 426, + 903, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As can be seen in the qualitative result images below (figure 8), GAN and diffusion (especially the SD) networks depict small details much more accurately. The background structure of the images with the different colors is completely blurred by the UNet but not by the GAN and the diffusion. The DDPM predictions are blurred because the images have been scaled up from $64\\mathrm{px}$ to $256\\mathrm{px}$ . If the output of the diffusion model is set to $256\\mathrm{px}$ , the target image is sharp again but several or no balls and many more artefacts appear on the net predictions (see figure 25 in the Appendix). To summarize, the Pix2Pix and Stable Diffusion approaches also are visually performing the best.", + "bbox": [ + 511, + 638, + 903, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, slight correlations can be observed between the simulation's parameters and the generated errors. For example the ball position in the X direction is predicted worse for all nets when the time interval between input and prediction images becomes larger. More details on the correla", + "bbox": [ + 511, + 818, + 903, + 893 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11131", + "bbox": [ + 480, + 943, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/528cde0b04af22a6d9af7c0f885c082acc046ffc275f10d212ea074d10668656.jpg", + "image_caption": [ + "Figure 8. Qualitative results for predictions of the bouncing ball problem. Initial State and True label image at top-left; subsequent panels show results for VAE, UNet, Pix2Pix and Diffusion models." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 478, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions of each method can be found in the Appendix D.2.1. Analysis. When comparing the result in table 3 (and tables 11, 13 in the Appendix) with the physical motion equations, it shows that the errors of the generative networks increase with higher order terms. The second-order terms, such as the movement along the X-axis and the ball angle represent the largest errors for the rolling scenario. In the bouncing case, a second-order derivative is also present to describe the ball movement in the Y-direction due to the ball's impact on the ground. Therefore the calculation rule along the Y-direction consists of zero, first and second derivative terms, making it more complex than the equation in the X-direction. This is reflected in the error values of all the investigated approaches: the Y-error clearly exceeds that in the X-direction, as indicated by the complexity of the motion equations, although the X-position and angular errors remain roughly the same between the rolling and bouncing cases. In general, it can be concluded that the errors of the analyzed approaches increase with the order and complexity of the physical formulation.", + "bbox": [ + 91, + 314, + 483, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 89, + 631, + 204, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With PhysicsGen, we present a comprehensive benchmark for the estimation of complex physical relations by image-to-image generative models. The three different datasets contained in PhysicsGen cover a wide range of diverse simulation scenarios and hence allow a systematic evaluation of learning based image generation models.", + "bbox": [ + 89, + 657, + 482, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Key insights from the baseline-evaluation. Despite the diversity of the three given physical problems and the fundamentally different network types used the base-line evaluation experiments, we can report several consistent findings:", + "bbox": [ + 89, + 750, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- Due to the unavailability of optimized GPU code for the used simulations it is impossible to provide a theoretically sound runtime comparison. Since a highly parallel implementation of the simulation algorithms is highly nontrivial (if possible at all), our analysis focuses on the runtime analysis of the generative models. However, replac", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing physical simulations with GPU powered generative models allows high computational speed-ups in practice, especially when replacing highly iterative solvers like in the urban sound propagation with Pix2Pix or UNets (where we report speed-up factors of up to 20k).", + "bbox": [ + 524, + 90, + 903, + 167 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In terms of physical correctness and consistency, our evaluation shows good results for simple 0th and 1st order problems, but also indicates that generative models mostly fail to predict more complex physical relations (which are typically formulated by higher order terms). This finding needs further investigation whether this is an effect of the principal theoretical limitations of current models or just a lack of sufficient training data.", + "- We also observe distinct error pattern for different model types which represents how the different model types are dealing with complexity and uncertainty. Please refer to Appendix D.2.1 for examples in the following exemplary discussion for the bouncing ball problem (similar effects can be observed in Appendix B.7 for the sound propagation tasks as well): while GANs often produce physically possible outputs, they still can fail to generate the correct output for given inputs. This results in consistent scenes with a complete ball at the wrong position. UNets, VAE and convAE on the other hand, often fail to produce consistent scenes and tend to blur uncertain predictions, resulting in deformed balls and destructed backgrounds. Lastly, diffusion models often produce multiple candidate solutions (multiple balls in one scene) and tend to fail to estimate more complex relations like ball rotation. This phenomenon also needs further investigation." + ], + "bbox": [ + 511, + 167, + 908, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. The provided run-time analysis is based on the comparison of practically available implementations and is not fair from a theoretical perspective as no efforts were made to optimize the simulation codes to GPUs. Also, even though we carefully selected the three proposed physical benchmark problems such that they are likely to transfer to a wide range of other simulation tasks, our evaluation provides only a small subset of possible applications of generative models in physical modeling. Hence, results from our benchmark might not transfer to all possible scenarios.", + "bbox": [ + 511, + 583, + 906, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader Impact Statement. The simulation of physical systems forms the backbone for a wide range of scientific and engineering tasks. Hence, it is important to have benchmarks for the emerging field of neural enhanced simulations to compare and validate these novel approaches. This work provides such a benchmark for the specific, but widely applicable case, where physical simulations are used to map between 2D input and output structures (visualized as images). Our analysis shows that while current generative models are promising significant speedups, there are still many open questions regarding their physical correctness.", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11132", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Funding Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 89, + 326, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors acknowledge the financial support by the German Federal Ministry of Education and Research (BMBF) in the program \"Forschung an Fachhochschulen in Kooperation mit Unternehmen (FH-Kooperativ)\" within the joint project \"KI-Bohrer\" under grant 13FH525KX1.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] A. Argilaga. Fem-gan: A physics-supervised deep learning generative model for elastic porous materials. Materials, 2023. 2", + "[2] Peter W Battaglia, Jessica B Hamrick, and Joshua B Tenenbaum. Simulation as an engine of physical scene understanding. Proceedings of the National Academy of Sciences, 110 (45):18327-18332, 2013. 1", + "[3] Victor Blomqvist. Pymunk pythonic 2d physics library, 2007 - 2024. 6, 20", + "[4] Erwan Bocher, Gwenael Guillaume, Judicael Picaut, Gwendall Petit, and Nicolas Fortin. Noise modelling: An open source gis based tool to produce environmental noise maps. ISPRS International Journal of Geo-Information, 8(3):130, 2019. 3, 4, 13, 15", + "[5] G. Bradski. The OpenCV Library. Dr. Dobb's Journal of Software Tools, 2000. 5, 17", + "[6] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym, 2016. 2", + "[7] Duane Brown. Close-range camera calibration. 1971. 5", + "[8] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d and 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, 2017. 5", + "[9] Giuseppe Carleo, Ignacio Cirac, Kyle Cranmer, Laurent Daudet, Maria Schuld, Naftali Tishby, Leslie Vogt-Maranto, and Lenka Zdeborova. Machine learning and the physical sciences. Reviews of Modern Physics, 91(4):045002, 2019. 1, 2", + "[10] Yilun Chen, Qi Wu, Yujie Sui, Yizhou Wang, and Song-Chun Zhu. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 2", + "[11] Runmin Cong, Wenyu Yang, Chongyi Zhang, Weiand Li, Chunle Guo, and Sam Huang, Qingmingand Kwong. Pugan: Physical model-guided underwater image enhancement using gan with dual-discriminators. IEEE Transactions on Image Processing, 2023. 2", + "[12] A. E. Conrady. Decentred Lens-Systems. Monthly Notices of the Royal Astronomical Society, 79(5):384-390, 1919. 5, 17", + "[13] Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep reinforcement learning for continuous control, 2016. 2", + "[14] Wolfgang Faig. Calibration of close-range photogrammetric systems: Mathematical formulation. Photogrammetric Engineering and Remote Sensing, 41, 1975. 5" + ], + "bbox": [ + 93, + 227, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. 1", + "[16] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. 2", + "[17] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti, Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator, 2022. 2", + "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. CoRR, abs/2006.11239, 2020. 2, 4, 11, 16", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1", + "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. CoRR, abs/1611.07004, 2016. 2, 4, 11, 16", + "[21] Kelsey Jordahl, Joris Van den Bossche, Martin Fleischmann, Jacob Wasserman, James McBride, Jeffrey Gerard, Jeff Stratner, Matthew Perry, Adrian Garcia Badaracco, Carson Farmer, Geir Arne Hjelle, Alan D. Snow, Micah Cochran, Sean Gillies, Lucas Culbertson, Matt Bartos, Nick Eubank, maxalbert, Aleksey Bilogur, Sergio Rey, Christopher Ren, Dani Arribas-Bel, Leah Wasser, Levi John Wolf, Martin Journois, Joshua Wilson, Adam Greenhall, Chris Holdgraf, Filipe, and François Leblanc. geopandas/geopandas: v0.8.1, 2020.3", + "[22] Anuj Karpatne, Imme Ebert-Uphoff, Sai Ravela, Hassan A. Babaie, and Vipin Kumar. Physics-guided neural networks (pgnn): An application in lake temperature modeling. arXiv preprint arXiv:2212.02500, 2022. 2", + "[23] S Kephalopoulos, M Paviotti, and F Anfosso-Ledée. Common noise assessment methods in europe (cnossos-eu). Technical Report EUR 25379 EN, Publications Office of the European Union, Luxembourg (Luxembourg), 2012. JRC72550. 3", + "[24] Byungsoo Kim, V. C. Azevedo, N. Thürey, Theodore Kim, M. Gross, and B. Solenthaler. Deep fluids: A generative network for parameterized fluid simulations. Computer Graphics Forum, 38, 2018. 2", + "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 2, 4, 11, 16", + "[26] Daniel Ladwig, Martin Spitznagel, Jan Vaillant, Klaus Dorer, and Janis Keuper. Ai-guided noise reduction for urban geothermal drilling. 2024. 3", + "[27] Adam Lerer, Sam Gross, and Rob Fergus. Learning physical intuition of block towers by example, 2016. 2", + "[28] Wenbin Li, Seyedmajid Azimi, Ales Leonardis, and Mario Fritz. To fall or not to fall: A visual approach to physical stability prediction, 2016. 2" + ], + "bbox": [ + 514, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11133", + "bbox": [ + 480, + 943, + 517, + 954 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2, 4, 11, 16", + "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015. 2, 4, 11, 16", + "[32] Yuxuan Song, Jingjing Gong, Yanru Qu, Hao Zhou, Mingyue Zheng, Jingjing Liu, and Wei-Ying Ma. Unified generative modeling of 3d molecules via bayesian flow networks, 2024. 2", + "[33] Martin Spitznagel and Janis Keuper. Urban sound propagation: a benchmark for 1-step generative modeling of complex physical systems, 2024. 3", + "[34] Nils Thuerey, Philipp Holl, Maximilian Mueller, Patrick Schnell, Felix Trost, and Kiwon Um. Physics-based Deep Learning. WWW, 2021. 2", + "[35] Michael Vorländer. Auralization: Fundamentals of acoustics, modelling, simulation, algorithms and acoustic virtual reality. Springer Science & Business Media, 2007. 3, 14", + "[36] Yang Yu and Yongming Liu. Physics-guided generative adversarial network for probabilistic structural system identification. Expert Systems with Applications, 239:122339, 2024. 2", + "[37] Yufeng Zhang, Wei Li, Yangang Gao, and Shihong Li. Phys-diff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2106.03704, 2021. 2", + "[38] Linqi Zhou, Aaron Lou, Samar Khanna, and Stefano Ermon. Denoising diffusion bridge models, 2023. 2, 4, 11, 16", + "[39] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martin-Martin, Abhishek Joshi, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning, 2022. 2" + ], + "bbox": [ + 91, + 92, + 485, + 603 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11134", + "bbox": [ + 480, + 943, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/d8bec7dd-2170-41a3-a07a-e8a3813c711e_model.json b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/d8bec7dd-2170-41a3-a07a-e8a3813c711e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bea8e92ab09ab4221b0503afae36b9a74885e1aa --- /dev/null +++ b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/d8bec7dd-2170-41a3-a07a-e8a3813c711e_model.json @@ -0,0 +1,1859 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.131, + 0.764, + 0.177 + ], + "angle": 0, + "content": "PhysicsGen: Can Generative Models Learn from Images to Predict Complex Physical Relations?" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.203, + 0.79, + 0.257 + ], + "angle": 0, + "content": "Martin Spitznagel1, Jan Vaillant2 and Janis Keuper1,3 \n1IMLA, Offenburg University, 2Herrenknecht AG, 3Mannheim University \n{first.last}@hs-offenburg.de" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.485, + 0.613 + ], + "angle": 0, + "content": "The image-to-image translation abilities of generative learning models have recently made significant progress in the estimation of complex (steered) mappings between image distributions. While appearance based tasks like image inpainting or style transfer have been studied at length, we propose to investigate the potential of generative models in the context of physical simulations. Providing a dataset of 300k image-pairs and baseline evaluations for three different physical simulation tasks, we propose a benchmark to investigate the following research questions: i) are generative models able to learn complex physical relations from input-output image pairs? ii) what speedups can be achieved by replacing differential equation based simulations? While baseline evaluations of different current models show the potential for high speedups (ii), these results also show strong limitations toward the physical correctness (i). This underlines the need for new methods to enforce physical correctness. Data, baseline models and evaluation code: http://www.physics-gen.org." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.647, + 0.222, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.765 + ], + "angle": 0, + "content": "The numerical simulation of physical processes and relations is a crucial tool in a wide range of scientific and engineering tasks. Traditionally depending on the solution of various types of differential equations, research of physical simulation methods is increasingly shifting towards data-driven and learnable approaches [9]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this work, we focus on a narrow but practically important sub-field: the inference of physical relations from images. Extending Tenenbaum's seminal idea of a intuitive physics engine [2], where machine learning models have been trained to predict coarse physical properties from images, we investigate the capabilities of modern generative models to learn detailed, complex and physically correct mappings from image pairs. Given such abilities, generative models like GANs [15] and Diffusion models [19] could be" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.292, + 0.898, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.74, + 0.908, + 0.893 + ], + "angle": 0, + "content": "Figure 1. Overview of the physical problems, baseline generative models and their evaluation. A: We introduce three complex physical simulation tasks with 100k input-output image pairs each, providing ground truth simulations based on differential equations with varying complexity. B: We evaluate all tasks on independently trained image translation models; only results for the sound propagation task are visualized in this figure. C: While the evaluation of the baseline models shows a general ability of generative image models to learn physical relations from images, we observe significant performance drops for tasks that require a higher order term in the differential equations of their simulation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11125" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "utilized to introduce hard to formalize but observable relations into physical models. Due to the fast inference times, generative models could also be used to accelerate simulations [34]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.482, + 0.243 + ], + "angle": 0, + "content": "Introducing the PhysicsGen Benchmark. A first step towards such data-driven models is a thorough analysis of the capabilities of current generative approaches. Hence, we introduce a novel benchmark that allows to systematically investigate the inference of a diverse set of physical problems from images (as visualized in Fig. 1):" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.245, + 0.483, + 0.29 + ], + "angle": 0, + "content": "- We introduce PhysicsGen, a novel Benchmark for the evaluation of generative image models learning to infer physical simulations tasks from images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.29, + 0.483, + 0.366 + ], + "angle": 0, + "content": "- PhysicsGen provides a Dataset with a total of 300k image-pairs, representing three diverse physical simulation problems: i) iterative wave propagation (Sec. 3), ii) closed form lens distortion (Sec. 4) and iii) time series prediction of motion dynamics (Sec. 5)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.366, + 0.483, + 0.576 + ], + "angle": 0, + "content": "- We provide a baseline evaluation that measures speedups and physical accuracy compared to full physical simulations across multiple generative image-to-image models. The evaluated models include various generative image-to-image models: Generative Adversarial Networks (GANs) [16], specifically Pix2Pix [20]; an U-Net [31]; Convolutional Autoencoders (convAE) and Variational Autoencoders (VAEs) [25]. We also assess diffusion models, including Denoising Diffusion Probabilistic Models (DDPMs) [18], Stable Diffusion [30], and Denoising Diffusion Bridge Models (DDBMs) [38]. (see Appendix A for details of the model architectures and training - the model architectures and training procedures are identical for all three tasks)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.577, + 0.483, + 0.653 + ], + "angle": 0, + "content": "- Our Analysis of the baseline results shows potentially high speed-ups at good accuracies for simple 1st order simulation tasks. However, our experiments also show that current generative models have fundamental problems to learn higher order physical relations." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.245, + 0.483, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.665, + 0.24, + 0.68 + ], + "angle": 0, + "content": "1.1. Related Work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.482, + 0.795 + ], + "angle": 0, + "content": "The integration of learnable methods into physical modeling is a wide field with a rich body of literature which is very hard to capture in the limited space of this paper. In the following, we try to refer to the most important works in direct relation to our benchmark and some more general key concepts. We refer the reader to [34] and [9] for a comprehensive overview of the field." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Generative models for physical simulations. The integration of physical principles within generative models is an active and burgeoning area of research. Recent models such as PUGAN [11] and FEM-GAN [1] have demonstrated the potential of combining GANs with physical modeling to enhance performance in environments governed by complex physical laws. Advances in fluid dynamics and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "structural system identification using physics-guided GANs have shown significant improvements in efficiency and precision by incorporating physics-based loss functions and simulations [24, 36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Predicting physical relations. Also, significant progress has been made in understanding physics through machine learning. Initial research demonstrated that models could comprehend the dynamics of block towers, progressing beyond simple memorization to genuinely understanding physical interactions [27]. Practical applications have been explored, including fall detection using body part tracking through machine learning [28]. Additionally, advancements in generating physically plausible human animations highlight the potential of incorporating physical principles into machine learning frameworks [37]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.499 + ], + "angle": 0, + "content": "Grey-Box Models. Physics-guided approaches in AI have laid a solid foundation for incorporating physical laws into model training. Prior research has demonstrated the effectiveness of embedding physical constraints within AI models to enhance their performance and reliability [10, 22]. In generative models, diffusion models have shown promise but often face challenges when generating complex geometrical structures due to issues like multi-modality and noise sensitivity. The development of Geometric Bayesian Flow Networks addresses these challenges by enhancing the robustness and efficiency of molecule geometry modeling [32]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.905, + 0.665 + ], + "angle": 0, + "content": "Physical Simulations for Reinforcement Learning. Deep reinforcement learning settings benefit from a substantial availability of physics-based environments and datasets, which facilitate the training of models to understand and predict physical phenomena [13]. Resources such as OpenAI Gym, Robosuite, and PyBullet provide robust platforms for developing and testing reinforcement learning models in simulated physical settings [6, 17, 39]. However, in the domain of generative AI, there is a notable scarcity of physics-informed datasets and environments, limiting the ability to train models on complex physical systems." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.774, + 0.708 + ], + "angle": 0, + "content": "2. Benchmark Tasks Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.905, + 0.824 + ], + "angle": 0, + "content": "Selecting an appropriate set of tasks for the PhysicsGen benchmark is crucial to comprehensively evaluate the capabilities of generative image models in inferring complex physical relations. Given the vast diversity of possible physical simulation tasks, it is impractical to encompass the entire spectrum. Therefore, we strategically selected three distinct tasks based on the following criteria:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.869 + ], + "angle": 0, + "content": "- Diversity of Physical Problems: To cover different types of physical phenomena, ensuring that models are tested across various domains." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Variety of Numerical Solution Strategies: Incorporating both iterative and closed-form (non-iterative) solvers" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "11126" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "to assess models' adaptability to different computational approaches." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.122, + 0.482, + 0.151 + ], + "angle": 0, + "content": "- Simplicity of Evaluation Metrics: Facilitating straightforward and effective model performance evaluation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.153, + 0.482, + 0.213 + ], + "angle": 0, + "content": "- Accessibility to the ML Community: Choosing intuitively understandable tasks without requiring specialized physics knowledge, thereby making the benchmark widely applicable to the machine learning community." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.122, + 0.482, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.216, + 0.482, + 0.245 + ], + "angle": 0, + "content": "Based on these objectives, we introduce the following three simulation tasks within the PhysicsGen benchmark:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.482, + 0.351 + ], + "angle": 0, + "content": "Urban Sound Propagation: This task models the behavior of sound waves in urban environments, accounting for phenomena such as diffraction and reflections (Sec. 3). By providing 100,000 input-output image pairs based on differential equation-based simulations, it challenges generative models to accurately capture intricate sound distribution patterns influenced by urban structures [26, 33]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.482, + 0.456 + ], + "angle": 0, + "content": "Lens Distortion: Focusing on optical aberrations, this task utilizes the Brown-Conrady distortion model to simulate lens-induced geometric distortions in images (Sec. 4). With 100,000 image pairs depicting varying degrees of distortion, models are tasked with learning precise, closed-form physical relations that alter image geometry based on lens parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.482, + 0.578 + ], + "angle": 0, + "content": "Dynamics of Rolling and Bouncing Movements: This task involves simulating the motion of a rolling or bouncing ball on an inclined surface, incorporating both linear and rotational dynamics (Sec. 5). Providing 100,000 image pairs that capture the ball's position and rotation over time, it assesses models' ability to predict time-series dynamics and handle higher-order physical relations inherent in motion equations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.482, + 0.67 + ], + "angle": 0, + "content": "Collectively, these tasks offer a diverse framework for assessing the strengths and limitations of generative models in learning and replicating complex physical simulations from image data. The subsequent sections delve into each task in detail, outlining dataset generation, underlying physics, evaluation metrics, and baseline model performances." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.336, + 0.703 + ], + "angle": 0, + "content": "3. Urban Sound Propagation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.71, + 0.482, + 0.83 + ], + "angle": 0, + "content": "The simulation of wave propagation in complex scenes, including diffraction and reflection at multiple objects is a representative task for a wide range of physical problems which are typically solved via (expensive) iterative solving of higher order differential equations. Due to its intuitive setting and accessible data, we investigate wave propagation problems by a concrete example where sound is propagated in urban environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.831, + 0.482, + 0.877 + ], + "angle": 0, + "content": "Dataset. We sampled \\(25k\\) geo-locations across 10 cities, leveraging open-source geo-data from the Overpass API1 and processed it with GeoPandas [21] to represent urban en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.469 + ], + "angle": 0, + "content": "virements in a \\(500\\mathrm{m}^2\\) area with buildings and open spaces encoded in black and white pixels, respectively. Our selection criteria ensured diverse urban scenarios by mandating a minimum of 10 buildings within 200 meters and a 50-meter clearance from buildings to the sample point. We utilized the NoiseModelling v4 framework [4], adhering to the CNOSSOS-EU standards [23], to simulate sound propagation maps, placing sound receivers on a grid with a step-size of 5 meters and at building boundaries, excluding indoor placements (see Appendix B.3 for details). The resulting simulation output was then interpolated into \\(512\\mathrm{x}512\\) or \\(256\\mathrm{x}256\\) pixel maps, representing decibel levels on a \\([0,255]\\) scale. Four simulation tasks (Baseline, Diffraction, Reflection, Combined) were defined to model sound propagation under varying conditions, accumulating in a dataset of 100,000 samples. The Baseline task models a constant \\(95\\mathrm{dB}\\) noise source at \\(500\\mathrm{Hz}\\), excluding diffraction and reflection. The Diffraction task adds horizontal sound wave diffraction, the Reflection task simulates sound reflections with a standard absorption coefficient, and the Combined task introduces variable sound levels (60-115 dB) and environmental factors, incorporating both reflection and diffraction to mimic realistic urban sound behavior. Refer to Fig. 2 and Appendix B.6 for a detailed description and visualization of the dataset generation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.469, + 0.905, + 0.62 + ], + "angle": 0, + "content": "Physics of Sound Propagation. Mathematically, the propagation of sound over time is described via partial differential wave equations. Due to space constraints and the practical nature of our problem setting, we will neglect the derivation from continuous wave equations and directly focus on the discrete and iterative implementations of sound propagation which have been applied for our ground-truth simulations. Following [35], for a discrete set of receivers \\( R \\), the amplitude \\( L_{R_k}^j \\) of receiver \\( R_k \\) at frequency \\( j \\) is computed via iterative differences:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.626, + 0.905, + 0.648 + ], + "angle": 0, + "content": "\\[\nL _ {R _ {k}} ^ {j} = L _ {W} ^ {j} - A _ {\\text {d i v} R _ {k}} - A _ {\\text {a t m} R _ {k}} ^ {j} - A _ {\\text {d i f} R _ {k}} ^ {j} \\quad (j, k \\in \\mathbb {N}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.655, + 0.905, + 0.732 + ], + "angle": 0, + "content": "where \\( L_W^j \\) represents the source level, \\( A_{div_{R_k}} \\) captures the geometrical spreading, \\( A_{atm_{R_k}}^j \\) denotes the atmospheric absorption, and \\( A_{dif_{R_k}}^j \\) models diffraction. Note that the ground effect \\( A_{grd_{R_k}}^j \\) is neglected in our study." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.733, + 0.905, + 0.793 + ], + "angle": 0, + "content": "Additionally, the model accounts for reflections by adjusting the power level \\( L_W^j \\) based on the absorption coefficient \\( \\alpha_{vert} \\) of the surfaces involved. This adjustment is performed using the equation" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.798, + 0.905, + 0.82 + ], + "angle": 0, + "content": "\\[\nL _ {W} ^ {(n _ {r e f})} = L _ {W} ^ {(n _ {r e f} - 1)} + n _ {r e f} \\times 1 0 \\log_ {1 0} (1 - \\alpha_ {v e r t}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\( n_{ref} \\) indicates the number of reflections considered. Specular reflections are modeled using the image receiver method, which provides a computationally efficient way to account for the angle of incidence being equal to the angle of reflection [35]. The sound level at each receiver" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.355, + 0.901 + ], + "angle": 0, + "content": "http://overpass-api.de/api/map" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11127" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.603, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.065, + 0.907, + 0.19 + ], + "angle": 0, + "content": "Figure 2. The sampling pipeline for the sound propagation dataset utilizes the NoiseModelling framework [4] to generate sound propagation maps based on specific urban layouts. The generators are then trained to replicate these sound propagation patterns for given locations and source parameters. Predictions are evaluated by specifically analyzing errors in relation to the line of sight (see Appendix B.1 for details)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.228, + 0.482, + 0.304 + ], + "angle": 0, + "content": "reflects the cumulative effect of direct, diffracted, and reflected sound paths. As the number of reflections increases, the complexity of calculating the power level \\( L_W^j \\) also increases, growing in \\( O(N^2) \\), where \\( N \\) denotes the number of reflective surfaces considered in the computation." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.305, + 0.484, + 0.56 + ], + "angle": 0, + "content": "Evaluation. Our baseline evaluation focuses on the ability of standard image-to-image generative models to accurately predict sound propagations. Prediction quality is determined by the pixel-wise difference between actual and predicted sound distributions. We use Mean Absolute Error (MAE) to quantify the average magnitude of prediction errors. Additionally, we introduce Weighted Mean Absolute Percentage Error (wMAPE) to specifically penalize errors in predictions that inaccurately show high sound amplitudes in regions expected to have low amplitudes, such as areas behind buildings. To evaluate model performance in capturing sound reflections or diffractions, metrics were specifically calculated for areas both within and outside the direct line of sight (LoS and NLoS) from the sound source. This methodology allows us to assess model predictions for sound propagations through direct paths as well as reflections and diffractions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Analysis. The results in Tab. 1 show that generative models significantly outperform traditional sound propagation simulations in processing speed, achieving up to a \\(20k\\) factor improvement in runtime (see Tab. 9 in the Appendix for further analysis). Regarding the prediction quality, we observe that models encountering tasks with reflections or diffractions show an increase in MAE for NLoS regions across" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.227, + 0.907, + 0.283 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation across all tasks for all architectures with a batch size of 16 during inference. The complete results containing the Combined task are given in Appendix B.7. Best overall results in bold, second best underlined." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.646 + ], + "angle": 0, + "content": "
ModelTaskMAE ↓wMAPE ↓Runtime/ ↓ Sample (ms)
LoSNLoSLoSNLoS
SimulationBase0.000.000.000.00204700
convAEBase3.672.7420.2467.130.128
VAE [25]Base3.922.8421.3375.580.124
UNet [31]Base2.291.7312.9137.570.138
Pix2Pix [20]Base1.731.199.366.750.138
DDPM [18]Base2.423.2615.5751.083986.353
SD(w,CA) [30]Base3.763.3417.4235.182961.027
SDBase2.121.0813.2332.462970.86
DDBM [38]Base1.612.1717.5065.243732.21
SimulationDif.0.000.000.000.00206000
convAEDif.3.598.0413.7732.090.128
VAEDif.3.928.2214.4632.570.124
UNetDif.0.943.274.2222.360.138
Pix2PixDif.0.913.363.5118.060.138
DDPMDif.1.593.278.2520.303986.353
SD(w,CA)Dif.2.467.7210.1431.232961.027
SDDif.1.335.078.1524.452970.86
DDBMDif.1.353.3511.2223.563732.21
SimulationRefl.0.000.000.000.00251000
convAERefl.3.836.5620.6793.540.128
VAERefl.4.156.3221.5792.470.124
UNetRefl.2.295.7212.7580.460.138
Pix2PixRefl.2.144.7911.3030.670.138
DDPMRefl.2.747.9317.8580.383986.353
SD(w,CA)Refl.3.816.8219.7881.612961.027
SDRefl.2.535.2615.0455.272970.86
DDBMRefl.1.936.3818.3479.133732.21
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.695, + 0.48, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.855, + 0.483, + 0.898 + ], + "angle": 0, + "content": "Figure 3. Qualitative results comparing the ground-truth simulation with the prediction for a single sample within the reflection task. Additional results can be found in Appendix B.7" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.903 + ], + "angle": 0, + "content": "all tested architectures. For the Pix2Pix model this rise in NLoS error is moderate, while this effect is more pronounced in the UNet and Diffusion models. Models without skip connections, such as VAE and ConvAE, tend to blur most of the image and struggle to capture any reflection patterns accurately. This trend highlights how reflections and diffractions introduce complex higher-order dependencies (see Eq. (12) and Appendix B.4) that complicate the image-to-image sound propagation tasks. Although the models successfully replicate sound reflection patterns visually (see Fig. 3), there exists a significant gap between its visual outputs and the actual acoustic accuracy. This discrepancy suggests that while visually detailed, the models do not fully capture the complexities of sound physics." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "11128" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.248, + 0.106 + ], + "angle": 0, + "content": "4. Lens Distortion" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.115, + 0.483, + 0.266 + ], + "angle": 0, + "content": "Our second benchmark is designed to investigate a class of physical problems for which solutions can be computed directly via closed equations, rather than iterative solvers. We find a representative and intuitive task in the modeling of optical lens distortions [7, 14], a common problem in computer vision. Generative models are trained to simulate the appearance of faces distorted by lens effects, using the Brown-Conrady [12] model. These distortions are computed using OpenCV [5], providing ground truth for training and evaluation of generative models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.267, + 0.483, + 0.401 + ], + "angle": 0, + "content": "Dataset. We leverage a subset of the CelebA dataset [29], comprising 50,000 facial images, which were cropped to center the faces within a \\(256 \\times 256\\) pixel frame. The resulting images where split into two distinct groups to isolate the effects of tangential distortion: one focusing on horizontal distortion and the other on vertical distortion. This allows us to specifically measure the impact of each type of distortion on the accuracy of the generative models (see Fig. 4 and Appendix C for more details)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.554 + ], + "angle": 0, + "content": "Physics of Lens Distortions. The Brown-Conrady model is a widely adopted method to describe and correct lens distortions. It provides a mathematical framework to represent both radial and tangential distortions through a series of coefficients. For our study, we concentrate exclusively on the tangential distortion components, utilizing relevant coefficients to model the distortion characteristics of specific lenses. The mathematical representation of tangential distortion is expressed through the following equation, where \\( x_{dist} \\) and \\( y_{dist} \\) denote the distorted coordinates:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.564, + 0.483, + 0.605 + ], + "angle": 0, + "content": "\\[\nx _ {d i s t} = x + \\left[ 2 p _ {1} x y + p _ {2} \\left(r ^ {2} + 2 x ^ {2}\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.601, + 0.422, + 0.634 + ], + "angle": 0, + "content": "\\[\ny _ {d i s t} = y + \\left[ p _ {1} \\left(r ^ {2} + 2 y ^ {2}\\right) + 2 p _ {2} x y \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.737 + ], + "angle": 0, + "content": "where \\( r^2 = x^2 + y^2 \\) represents the squared radial distance from the image center, and \\( p_1 \\) and \\( p_2 \\) are the coefficients that model the tangential distortion effects. By strategically adjusting \\( p_1 \\) and \\( p_2 \\) and isolating their effects via controlled nullification, we can identify how each parameter individually contributes to the image distortion:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.747, + 0.483, + 0.794 + ], + "angle": 0, + "content": "\\[\nx _ {d i s t} = \\left\\{ \\begin{array}{l l} x + 2 p _ {1} x y & \\text {i f} p _ {2} = 0, \\\\ x + p _ {2} \\left(r ^ {2} + 2 x ^ {2}\\right) & \\text {i f} p _ {1} = 0, \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.791, + 0.42, + 0.83 + ], + "angle": 0, + "content": "\\[\ny _ {d i s t} = \\left\\{ \\begin{array}{l l} y + p _ {1} (r ^ {2} + 2 y ^ {2}) & \\text {i f} p _ {2} = 0, \\\\ y + 2 p _ {2} x y & \\text {i f} p _ {1} = 0. \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "These conditional equations illustrate the isolated impact of each coefficient, revealing that the correction complexity for \\( y_{dist} \\) escalates with the term \\( p_1(r^2 + 2y^2) \\) when \\( p_1 \\) is non-zero and \\( p_2 \\) is zero. Conversely, the complexity of \\( x_{dist} \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.121 + ], + "angle": 0, + "content": "correction intensifies with \\( p_2(r^2 + 2x^2) \\) under the condition that \\( p_2 \\) is non-zero and \\( p_1 \\) is zero." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.906, + 0.424 + ], + "angle": 0, + "content": "Evaluation. Given the challenges of accurately capturing lens distortion through traditional pixel-based metrics, our evaluation strategy is specifically designed to focus on accurately reproducing designated landmarks in the images. We detect facial landmarks using a 2D facial landmark detection method based on the Facial Alignment Network (FAN) [8], applied to the faces in the CelebA dataset (see Appendix C for details). We use these landmarks to assess the geometric distortions introduced by lens settings, represented by the coefficients \\( p_1 \\) and \\( p_2 \\). We then compute the Euclidean distance between the actual position of a landmark in the true distorted image and its predicted position in the image generated by the models as a quality metric. This measurement is performed separately for the horizontal (X) and vertical (Y) coordinates, resulting in X Error and Y Error respectively. This approach not only allows us to gauge the accuracy of the models in simulating the specific effects of lens distortion on facial features but also helps in understanding how different distortion coefficients impact the positional accuracy of landmarks along each axis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.636 + ], + "angle": 0, + "content": "Analysis. Our experiments with the CelebA dataset under varying conditions of \\( p_1 \\) and \\( p_2 \\) showcase different patterns in error rates that align with the theoretical complexity of the distortion model. As shown in Tab. 2 and Fig. 5, the generative models exhibit differing accuracies in landmark prediction under the influence of lens distortion parameters \\( p_1 \\) and \\( p_2 \\). Notably, the increased complexity in \\( x_{dist} \\) and \\( y_{dist} \\) equations under separate conditions (\\( p_1 \\) nonzero vs. \\( p_2 \\) nonzero) slightly correlates with heightened error rates in respective dimensions. The models exhibit higher Y Error when only \\( p_1 \\) is active, indicative of the complex correction challenge introduced by the higher order dependency on \\( y \\). Similarly, a slight increase in X error and a slight decrease in Y error are observed when \\( p_2 \\) is exclusively nonzero." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.64, + 0.905, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.792, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Figure 4. Sampling and evaluation pipeline for the lens distortion dataset. The Brown-Conrady distortion model generates the true distorted images based on parameters \\( p_1 \\) and \\( p_2 \\) (we depict a chess pattern for visualization). The conditioned generators are then trained to replicate these distortions for given images and parameters. The models are evaluated by comparing predicted against true facial landmarks, using a 2D facial landmark detection based on the Facial Alignment Network (FAN) [8]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "11129" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.09, + 0.462, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.292, + 0.485, + 0.362 + ], + "angle": 0, + "content": "Figure 5. Qualitative visualization of lens distortion predictions using different generative models on a CelebA dataset sample. \\( p_2 \\) distortion: Original image at top-left; subsequent panels show U-Net, Pix2Pix, Diffusion models. Red dots mark actual landmark positions and blue dots for predictions." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.437 + ], + "angle": 0, + "content": "Table 2. Quantitative X-Y error and shift comparison across different models for varying activation of parameters \\( p_1 \\) and \\( p_2 \\) with a batch size of 16 during inference. Best overall results in bold, second best underlined." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.446, + 0.462, + 0.702 + ], + "angle": 0, + "content": "
ModelComb ↓X ↓ ErrY ↓ Err|X-Y| ShiftRuntime/ Sample (ms)
Sim.0.000.000.000.00153.205
convAE11.936.758.131.380.110
VAE11.536.557.831.280.122
UNet2.821.282.150.870.118
Pix2Pix2.000.991.430.440.122
DDPM1.930.941.390.453970.603
SD(w,CA)3.091.592.210.622991.678
SD2.791.412.010.602997.576
Sim.p1 = 0, p2 ≠ 0
convAE0.000.000.000.00153.205
10.568.354.772.210.110
VAE10.408.264.623.640.122
UNet2.361.331.600.270.117
Pix2Pix1.771.021.140.130.123
DDPM2.131.391.230.163970.603
SD(w,CA)2.851.601.940.342991.678
SD2.441.381.640.262997.576
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.402, + 0.725 + ], + "angle": 0, + "content": "5. Dynamics of bouncing movements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.885 + ], + "angle": 0, + "content": "The third physical problem under investigation is the movement of a rolling or bouncing ball. The aim here is to evaluate the ability of generative models to map kinematic movements of physics - a task which should generalize well to the investigation of various time series problems. Specifically, we define the task to predict the position and rotation of a ball along an inclined surface for a defined time interval after the input image. Only the bouncing case is discussed below; however, all results for the rolling ball can be found in Appendix D.3." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Dataset. We generated our data set with Pymunk [3], a" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.091, + 0.87, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.205, + 0.907, + 0.223 + ], + "angle": 0, + "content": "\\[\nF _ {g} = m * a \\Leftrightarrow m * \\ddot {y} = - m * g \\Leftrightarrow \\ddot {y} = - g \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.238, + 0.907, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} m * \\ddot {y} + d * \\dot {y} - c * (r - y) = - m * g \\tag {6} \\\\ m * \\ddot {x} = m * g * \\sin (\\beta) \\text {a l o n g x a x i s} \\\\ \\end{array}\n\\]" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.379 + ], + "angle": 0, + "content": "Figure 6. Physical equations describing the bouncing ball movement divided in 3 parts: free fall (green), impact & bounce of the ball on the ground (red dot) and oblique throw with initial velocity from ball bounce (orange)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.658 + ], + "angle": 0, + "content": "Python physics simulation toolbox. For the bouncing case we use 50k pairs of training images (256x256px) and perform the evaluation of the generative networks with 1600 images. Three parameters on the input images are variable: the ground inclination, the start height of the ball movement and the ball position. Additionally, the time interval between the input and target images is also variable. All other physical properties of the problem (ball's mass or friction coefficients) are kept constant. A multi-coloured structure with randomly distributed lines is superimposed on the simulation images in order to obtain greater diversity in the training data and is also used to analyze consistency, e.g. how well the generative networks can map fine multicoloured patterns (see Fig. 7). The total time required to create a sample using the available physics simulation CPU code (with a customized background) is 3.8 seconds, while the inference times for the generative models remain consistent with those reported in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Physics of bouncing balls. To model the physics behind the movement of a bouncing ball, the motion is divided into different sections which are represented with different colors in Fig. 6. First, the ball is released from a defined height without any velocity and falls down according to the Newton's law of the free fall (green part) (see equation 5). The impact of the ball on the ground and its rebound can be described in simplified terms by a spring-damper model. The aim of this second part of the movement (red dot) is to calculate the rebound speed and angle. This is done with equation 6, which contains the constant of the elements of the simplified model (see details in Appendix D.1) and the angle corresponds to the ground inclination for the first impact. Finally, the orange part of the movement is assumed to be an oblique throw with the initial velocity \\( v_{0} \\) calculated with equation 6. The equations of motion describing" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "11130" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.605, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.065, + 0.907, + 0.177 + ], + "angle": 0, + "content": "Figure 7. The sampling and evaluation pipeline for the rolling and bouncing ball movements simulates ball trajectories. A physics-based model is used to predict the ball's movement. Evaluations focus on the accuracy of predictions related to key movement metrics, including bounce height and horizontal displacement (see Appendix D for details)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.215, + 0.482, + 0.305 + ], + "angle": 0, + "content": "this oblique throw along the X and Y axes are also derived from the Newton's law (see equation 7). When the ball hits the ground for the second time, the principle \"angle of incidence = angle of reflection\" is used to determine the angle of rebound \\(\\theta\\) and the next part of the motion is again defined by equations 6 and 7." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.306, + 0.483, + 0.576 + ], + "angle": 0, + "content": "Evaluation. The generated images are examined according to 4 precise error criteria: the ball position split in X and Y directions, where the error is defined as the number of pixels between the theoretical and the predicted ball center. The second error measure is the ball rotation describing its angle in degrees during the movement due to the inclined ground surface. The next quality indicator of the algorithms is the out-of-roundness of the ball also measured in pixels by taking all identified points of the ball contour and calculating the difference between the theoretical ball radius of 15px and the predicted one. The standard deviation of all radius errors is taken as the error measure. Finally, the correct representation of the ground inclination is checked: the angular error between simulation and prediction is also measured in degrees. In addition, there are three \"general\" error criteria that give preciser indications of the correct representation of the physics which are explained with the complete results in the Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.483, + 0.714 + ], + "angle": 0, + "content": "The results of the seven examined generative algorithms are summarized in table 3 (only for the four most important error criteria). The last column of the table contains the average number of non-evaluable predicted images over the four error criteria. A detailed analysis of the advantages and disadvantages of the individual methods with all the evaluation parameters can be found in the Appendix D, as well as the evaluations obtained for individual runs for the rolling and bouncing case." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.732, + 0.485, + 0.832 + ], + "angle": 0, + "content": "
Position XPosition YRotationRoundnessError
convAE4.24 ± 3.96.08 ± 5.912.2 ± 8.61.06 ± 0.099%
VAE4.69 ± 6.16.25 ± 6.931.0 ± 400.90 ± 0.195%
UNet5.53 ± 7.510.8 ± 1215.2 ± 230.74 ± 0.228%
Pix2Pix6.28 ± 8.011.7 ± 1317.2 ± 210.56 ± 0.111%
DDPM7.91 ± 9.015.5 ± 1432.9 ± 340.61 ± 0.25.7%
SD(w.CA)40.0 ± 4924.8 ± 2361.1 ± 520.53 ± 0.27.3%
SD8.55 ± 1216.2 ± 1434.2 ± 380.47 ± 0.12%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.841, + 0.485, + 0.898 + ], + "angle": 0, + "content": "Table 3. Prediction results for the bouncing ball experiment in mean±std and mean percentage of non-evaluable predictions for the four criteria (best overall results in bold). Average runtime of the simulation: 3.8s, generative models: same as inTab. 1" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.905, + 0.426 + ], + "angle": 0, + "content": "Overall, the GAN (Pix2Pix model) provides the most stable representation of the bouncing ball problem. It achieves the best results across all evaluation criteria, with the most reliable physics predictions and fewest failed samples. The UNet also represents the ball position and angle on the generated samples very well, but produces more error images that cannot be evaluated with mostly either no ball or only a shredded shape. The greatest weakness of this algorithm lies in the clean representation of the ball roundness. The auto-encoder approaches (VAE and convAE) produce very blurred images in which the ball is often distorted and hardly recognizable. Therefore, the mean error values are not very representative, as over \\(90\\%\\) of the predictions cannot be analyzed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.905, + 0.638 + ], + "angle": 0, + "content": "The denoising diffusion approach (DDPM) produces slightly poorer results in absolute terms. Especially the prediction of the ball rotation is worse, with an error approximately twice as high compared to the two other approaches. Finally, to test another type of diffusion, two variations of stable diffusion were compared. The one with cross-attention is generating more error pictures and show high error values. In contrast, the normal stable diffusion performs very well: the mapping of the position is almost as accurate as with the Pix2Pix approach and the representation of the ball roundness is the best. In addition, the network generates very few error images, so that it could almost be placed on the same level as the GAN but comes as second due to a poorer angle mapping." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.905, + 0.818 + ], + "angle": 0, + "content": "As can be seen in the qualitative result images below (figure 8), GAN and diffusion (especially the SD) networks depict small details much more accurately. The background structure of the images with the different colors is completely blurred by the UNet but not by the GAN and the diffusion. The DDPM predictions are blurred because the images have been scaled up from \\(64\\mathrm{px}\\) to \\(256\\mathrm{px}\\). If the output of the diffusion model is set to \\(256\\mathrm{px}\\), the target image is sharp again but several or no balls and many more artefacts appear on the net predictions (see figure 25 in the Appendix). To summarize, the Pix2Pix and Stable Diffusion approaches also are visually performing the best." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.819, + 0.905, + 0.895 + ], + "angle": 0, + "content": "Finally, slight correlations can be observed between the simulation's parameters and the generated errors. For example the ball position in the X direction is predicted worse for all nets when the time interval between input and prediction images becomes larger. More details on the correla" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11131" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.48, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.307 + ], + "angle": 0, + "content": "Figure 8. Qualitative results for predictions of the bouncing ball problem. Initial State and True label image at top-left; subsequent panels show results for VAE, UNet, Pix2Pix and Diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.315, + 0.484, + 0.617 + ], + "angle": 0, + "content": "tions of each method can be found in the Appendix D.2.1. Analysis. When comparing the result in table 3 (and tables 11, 13 in the Appendix) with the physical motion equations, it shows that the errors of the generative networks increase with higher order terms. The second-order terms, such as the movement along the X-axis and the ball angle represent the largest errors for the rolling scenario. In the bouncing case, a second-order derivative is also present to describe the ball movement in the Y-direction due to the ball's impact on the ground. Therefore the calculation rule along the Y-direction consists of zero, first and second derivative terms, making it more complex than the equation in the X-direction. This is reflected in the error values of all the investigated approaches: the Y-error clearly exceeds that in the X-direction, as indicated by the complexity of the motion equations, although the X-position and angular errors remain roughly the same between the rolling and bouncing cases. In general, it can be concluded that the errors of the analyzed approaches increase with the order and complexity of the physical formulation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.632, + 0.205, + 0.648 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.748 + ], + "angle": 0, + "content": "With PhysicsGen, we present a comprehensive benchmark for the estimation of complex physical relations by image-to-image generative models. The three different datasets contained in PhysicsGen cover a wide range of diverse simulation scenarios and hence allow a systematic evaluation of learning based image generation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Key insights from the baseline-evaluation. Despite the diversity of the three given physical problems and the fundamentally different network types used the base-line evaluation experiments, we can report several consistent findings:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Due to the unavailability of optimized GPU code for the used simulations it is impossible to provide a theoretically sound runtime comparison. Since a highly parallel implementation of the simulation algorithms is highly nontrivial (if possible at all), our analysis focuses on the runtime analysis of the generative models. However, replac" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "ing physical simulations with GPU powered generative models allows high computational speed-ups in practice, especially when replacing highly iterative solvers like in the urban sound propagation with Pix2Pix or UNets (where we report speed-up factors of up to 20k)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.907, + 0.288 + ], + "angle": 0, + "content": "- In terms of physical correctness and consistency, our evaluation shows good results for simple 0th and 1st order problems, but also indicates that generative models mostly fail to predict more complex physical relations (which are typically formulated by higher order terms). This finding needs further investigation whether this is an effect of the principal theoretical limitations of current models or just a lack of sufficient training data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.288, + 0.909, + 0.545 + ], + "angle": 0, + "content": "- We also observe distinct error pattern for different model types which represents how the different model types are dealing with complexity and uncertainty. Please refer to Appendix D.2.1 for examples in the following exemplary discussion for the bouncing ball problem (similar effects can be observed in Appendix B.7 for the sound propagation tasks as well): while GANs often produce physically possible outputs, they still can fail to generate the correct output for given inputs. This results in consistent scenes with a complete ball at the wrong position. UNets, VAE and convAE on the other hand, often fail to produce consistent scenes and tend to blur uncertain predictions, resulting in deformed balls and destructed backgrounds. Lastly, diffusion models often produce multiple candidate solutions (multiple balls in one scene) and tend to fail to estimate more complex relations like ball rotation. This phenomenon also needs further investigation." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.168, + 0.909, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Limitations. The provided run-time analysis is based on the comparison of practically available implementations and is not fair from a theoretical perspective as no efforts were made to optimize the simulation codes to GPUs. Also, even though we carefully selected the three proposed physical benchmark problems such that they are likely to transfer to a wide range of other simulation tasks, our evaluation provides only a small subset of possible applications of generative models in physical modeling. Hence, results from our benchmark might not transfer to all possible scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Broader Impact Statement. The simulation of physical systems forms the backbone for a wide range of scientific and engineering tasks. Hence, it is important to have benchmarks for the emerging field of neural enhanced simulations to compare and validate these novel approaches. This work provides such a benchmark for the specific, but widely applicable case, where physical simulations are used to map between 2D input and output structures (visualized as images). Our analysis shows that while current generative models are promising significant speedups, there are still many open questions regarding their physical correctness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "11132" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.327, + 0.108 + ], + "angle": 0, + "content": "Funding Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.192 + ], + "angle": 0, + "content": "The authors acknowledge the financial support by the German Federal Ministry of Education and Research (BMBF) in the program \"Forschung an Fachhochschulen in Kooperation mit Unternehmen (FH-Kooperativ)\" within the joint project \"KI-Bohrer\" under grant 13FH525KX1." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[1] A. Argilaga. Fem-gan: A physics-supervised deep learning generative model for elastic porous materials. Materials, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.271, + 0.484, + 0.326 + ], + "angle": 0, + "content": "[2] Peter W Battaglia, Jessica B Hamrick, and Joshua B Tenenbaum. Simulation as an engine of physical scene understanding. Proceedings of the National Academy of Sciences, 110 (45):18327-18332, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.327, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[3] Victor Blomqvist. Pymunk pythonic 2d physics library, 2007 - 2024. 6, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.355, + 0.483, + 0.423 + ], + "angle": 0, + "content": "[4] Erwan Bocher, Gwenael Guillaume, Judicael Picaut, Gwendall Petit, and Nicolas Fortin. Noise modelling: An open source gis based tool to produce environmental noise maps. ISPRS International Journal of Geo-Information, 8(3):130, 2019. 3, 4, 13, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.485, + 0.452 + ], + "angle": 0, + "content": "[5] G. Bradski. The OpenCV Library. Dr. Dobb's Journal of Software Tools, 2000. 5, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[6] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.458, + 0.509 + ], + "angle": 0, + "content": "[7] Duane Brown. Close-range camera calibration. 1971. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.484, + 0.578 + ], + "angle": 0, + "content": "[8] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d and 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.484, + 0.648 + ], + "angle": 0, + "content": "[9] Giuseppe Carleo, Ignacio Cirac, Kyle Cranmer, Laurent Daudet, Maria Schuld, Naftali Tishby, Leslie Vogt-Maranto, and Lenka Zdeborova. Machine learning and the physical sciences. Reviews of Modern Physics, 91(4):045002, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[10] Yilun Chen, Qi Wu, Yujie Sui, Yizhou Wang, and Song-Chun Zhu. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[11] Runmin Cong, Wenyu Yang, Chongyi Zhang, Weiand Li, Chunle Guo, and Sam Huang, Qingmingand Kwong. Pugan: Physical model-guided underwater image enhancement using gan with dual-discriminators. IEEE Transactions on Image Processing, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[12] A. E. Conrady. Decentred Lens-Systems. Monthly Notices of the Royal Astronomical Society, 79(5):384-390, 1919. 5, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[13] Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep reinforcement learning for continuous control, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Wolfgang Faig. Calibration of close-range photogrammetric systems: Mathematical formulation. Photogrammetric Engineering and Remote Sensing, 41, 1975. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[16] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.327 + ], + "angle": 0, + "content": "[17] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti, Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. CoRR, abs/2006.11239, 2020. 2, 4, 11, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. CoRR, abs/1611.07004, 2016. 2, 4, 11, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.906, + 0.578 + ], + "angle": 0, + "content": "[21] Kelsey Jordahl, Joris Van den Bossche, Martin Fleischmann, Jacob Wasserman, James McBride, Jeffrey Gerard, Jeff Stratner, Matthew Perry, Adrian Garcia Badaracco, Carson Farmer, Geir Arne Hjelle, Alan D. Snow, Micah Cochran, Sean Gillies, Lucas Culbertson, Matt Bartos, Nick Eubank, maxalbert, Aleksey Bilogur, Sergio Rey, Christopher Ren, Dani Arribas-Bel, Leah Wasser, Levi John Wolf, Martin Journois, Joshua Wilson, Adam Greenhall, Chris Holdgraf, Filipe, and François Leblanc. geopandas/geopandas: v0.8.1, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[22] Anuj Karpatne, Imme Ebert-Uphoff, Sai Ravela, Hassan A. Babaie, and Vipin Kumar. Physics-guided neural networks (pgnn): An application in lake temperature modeling. arXiv preprint arXiv:2212.02500, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[23] S Kephalopoulos, M Paviotti, and F Anfosso-Ledée. Common noise assessment methods in europe (cnossos-eu). Technical Report EUR 25379 EN, Publications Office of the European Union, Luxembourg (Luxembourg), 2012. JRC72550. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.705, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[24] Byungsoo Kim, V. C. Azevedo, N. Thürey, Theodore Kim, M. Gross, and B. Solenthaler. Deep fluids: A generative network for parameterized fluid simulations. Computer Graphics Forum, 38, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.761, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 2, 4, 11, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[26] Daniel Ladwig, Martin Spitznagel, Jan Vaillant, Klaus Dorer, and Janis Keuper. Ai-guided noise reduction for urban geothermal drilling. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.832, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[27] Adam Lerer, Sam Gross, and Rob Fergus. Learning physical intuition of block towers by example, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Wenbin Li, Seyedmajid Azimi, Ales Leonardis, and Mario Fritz. To fall or not to fall: A visual approach to physical stability prediction, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.519, + 0.955 + ], + "angle": 0, + "content": "11133" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[29] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2, 4, 11, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015. 2, 4, 11, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[32] Yuxuan Song, Jingjing Gong, Yanru Qu, Hao Zhou, Mingyue Zheng, Jingjing Liu, and Wei-Ying Ma. Unified generative modeling of 3d molecules via bayesian flow networks, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[33] Martin Spitznagel and Janis Keuper. Urban sound propagation: a benchmark for 1-step generative modeling of complex physical systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[34] Nils Thuerey, Philipp Holl, Maximilian Mueller, Patrick Schnell, Felix Trost, and Kiwon Um. Physics-based Deep Learning. WWW, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[35] Michael Vorländer. Auralization: Fundamentals of acoustics, modelling, simulation, algorithms and acoustic virtual reality. Springer Science & Business Media, 2007. 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[36] Yang Yu and Yongming Liu. Physics-guided generative adversarial network for probabilistic structural system identification. Expert Systems with Applications, 239:122339, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[37] Yufeng Zhang, Wei Li, Yangang Gao, and Shihong Li. Phys-diff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2106.03704, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[38] Linqi Zhou, Aaron Lou, Samar Khanna, and Stefano Ermon. Denoising diffusion bridge models, 2023. 2, 4, 11, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.604 + ], + "angle": 0, + "content": "[39] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martin-Martin, Abhishek Joshi, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "11134" + } + ] +] \ No newline at end of file diff --git a/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/d8bec7dd-2170-41a3-a07a-e8a3813c711e_origin.pdf b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/d8bec7dd-2170-41a3-a07a-e8a3813c711e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f20ae4f0f90edee7bb1420db2040e9913ee0871b --- /dev/null +++ b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/d8bec7dd-2170-41a3-a07a-e8a3813c711e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13eaf301e3cbd41503d17bb057876fb1b1db952da4c70f713c4066a0e281ebc3 +size 8973547 diff --git a/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/full.md b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad001c2f7425179c30adc5ba5ecfd9d22d621c43 --- /dev/null +++ b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/full.md @@ -0,0 +1,261 @@ +# PhysicsGen: Can Generative Models Learn from Images to Predict Complex Physical Relations? + +Martin Spitznagel1, Jan Vaillant2 and Janis Keuper1,3 +1IMLA, Offenburg University, 2Herrenknecht AG, 3Mannheim University +{first.last}@hs-offenburg.de + +# Abstract + +The image-to-image translation abilities of generative learning models have recently made significant progress in the estimation of complex (steered) mappings between image distributions. While appearance based tasks like image inpainting or style transfer have been studied at length, we propose to investigate the potential of generative models in the context of physical simulations. Providing a dataset of 300k image-pairs and baseline evaluations for three different physical simulation tasks, we propose a benchmark to investigate the following research questions: i) are generative models able to learn complex physical relations from input-output image pairs? ii) what speedups can be achieved by replacing differential equation based simulations? While baseline evaluations of different current models show the potential for high speedups (ii), these results also show strong limitations toward the physical correctness (i). This underlines the need for new methods to enforce physical correctness. Data, baseline models and evaluation code: http://www.physics-gen.org. + +# 1. Introduction + +The numerical simulation of physical processes and relations is a crucial tool in a wide range of scientific and engineering tasks. Traditionally depending on the solution of various types of differential equations, research of physical simulation methods is increasingly shifting towards data-driven and learnable approaches [9]. + +In this work, we focus on a narrow but practically important sub-field: the inference of physical relations from images. Extending Tenenbaum's seminal idea of a intuitive physics engine [2], where machine learning models have been trained to predict coarse physical properties from images, we investigate the capabilities of modern generative models to learn detailed, complex and physically correct mappings from image pairs. Given such abilities, generative models like GANs [15] and Diffusion models [19] could be + +![](images/b3433c8f4aee8bc1cbab065d16d4233edf15af085231732bc8fe61661585766f.jpg) +Figure 1. Overview of the physical problems, baseline generative models and their evaluation. A: We introduce three complex physical simulation tasks with 100k input-output image pairs each, providing ground truth simulations based on differential equations with varying complexity. B: We evaluate all tasks on independently trained image translation models; only results for the sound propagation task are visualized in this figure. C: While the evaluation of the baseline models shows a general ability of generative image models to learn physical relations from images, we observe significant performance drops for tasks that require a higher order term in the differential equations of their simulation. + +utilized to introduce hard to formalize but observable relations into physical models. Due to the fast inference times, generative models could also be used to accelerate simulations [34]. + +Introducing the PhysicsGen Benchmark. A first step towards such data-driven models is a thorough analysis of the capabilities of current generative approaches. Hence, we introduce a novel benchmark that allows to systematically investigate the inference of a diverse set of physical problems from images (as visualized in Fig. 1): + +- We introduce PhysicsGen, a novel Benchmark for the evaluation of generative image models learning to infer physical simulations tasks from images. +- PhysicsGen provides a Dataset with a total of 300k image-pairs, representing three diverse physical simulation problems: i) iterative wave propagation (Sec. 3), ii) closed form lens distortion (Sec. 4) and iii) time series prediction of motion dynamics (Sec. 5). +- We provide a baseline evaluation that measures speedups and physical accuracy compared to full physical simulations across multiple generative image-to-image models. The evaluated models include various generative image-to-image models: Generative Adversarial Networks (GANs) [16], specifically Pix2Pix [20]; an U-Net [31]; Convolutional Autoencoders (convAE) and Variational Autoencoders (VAEs) [25]. We also assess diffusion models, including Denoising Diffusion Probabilistic Models (DDPMs) [18], Stable Diffusion [30], and Denoising Diffusion Bridge Models (DDBMs) [38]. (see Appendix A for details of the model architectures and training - the model architectures and training procedures are identical for all three tasks). +- Our Analysis of the baseline results shows potentially high speed-ups at good accuracies for simple 1st order simulation tasks. However, our experiments also show that current generative models have fundamental problems to learn higher order physical relations. + +# 1.1. Related Work. + +The integration of learnable methods into physical modeling is a wide field with a rich body of literature which is very hard to capture in the limited space of this paper. In the following, we try to refer to the most important works in direct relation to our benchmark and some more general key concepts. We refer the reader to [34] and [9] for a comprehensive overview of the field. + +Generative models for physical simulations. The integration of physical principles within generative models is an active and burgeoning area of research. Recent models such as PUGAN [11] and FEM-GAN [1] have demonstrated the potential of combining GANs with physical modeling to enhance performance in environments governed by complex physical laws. Advances in fluid dynamics and + +structural system identification using physics-guided GANs have shown significant improvements in efficiency and precision by incorporating physics-based loss functions and simulations [24, 36]. + +Predicting physical relations. Also, significant progress has been made in understanding physics through machine learning. Initial research demonstrated that models could comprehend the dynamics of block towers, progressing beyond simple memorization to genuinely understanding physical interactions [27]. Practical applications have been explored, including fall detection using body part tracking through machine learning [28]. Additionally, advancements in generating physically plausible human animations highlight the potential of incorporating physical principles into machine learning frameworks [37]. + +Grey-Box Models. Physics-guided approaches in AI have laid a solid foundation for incorporating physical laws into model training. Prior research has demonstrated the effectiveness of embedding physical constraints within AI models to enhance their performance and reliability [10, 22]. In generative models, diffusion models have shown promise but often face challenges when generating complex geometrical structures due to issues like multi-modality and noise sensitivity. The development of Geometric Bayesian Flow Networks addresses these challenges by enhancing the robustness and efficiency of molecule geometry modeling [32]. + +Physical Simulations for Reinforcement Learning. Deep reinforcement learning settings benefit from a substantial availability of physics-based environments and datasets, which facilitate the training of models to understand and predict physical phenomena [13]. Resources such as OpenAI Gym, Robosuite, and PyBullet provide robust platforms for developing and testing reinforcement learning models in simulated physical settings [6, 17, 39]. However, in the domain of generative AI, there is a notable scarcity of physics-informed datasets and environments, limiting the ability to train models on complex physical systems. + +# 2. Benchmark Tasks Overview + +Selecting an appropriate set of tasks for the PhysicsGen benchmark is crucial to comprehensively evaluate the capabilities of generative image models in inferring complex physical relations. Given the vast diversity of possible physical simulation tasks, it is impractical to encompass the entire spectrum. Therefore, we strategically selected three distinct tasks based on the following criteria: + +- Diversity of Physical Problems: To cover different types of physical phenomena, ensuring that models are tested across various domains. +- Variety of Numerical Solution Strategies: Incorporating both iterative and closed-form (non-iterative) solvers + +to assess models' adaptability to different computational approaches. + +- Simplicity of Evaluation Metrics: Facilitating straightforward and effective model performance evaluation. +- Accessibility to the ML Community: Choosing intuitively understandable tasks without requiring specialized physics knowledge, thereby making the benchmark widely applicable to the machine learning community. + +Based on these objectives, we introduce the following three simulation tasks within the PhysicsGen benchmark: + +Urban Sound Propagation: This task models the behavior of sound waves in urban environments, accounting for phenomena such as diffraction and reflections (Sec. 3). By providing 100,000 input-output image pairs based on differential equation-based simulations, it challenges generative models to accurately capture intricate sound distribution patterns influenced by urban structures [26, 33]. + +Lens Distortion: Focusing on optical aberrations, this task utilizes the Brown-Conrady distortion model to simulate lens-induced geometric distortions in images (Sec. 4). With 100,000 image pairs depicting varying degrees of distortion, models are tasked with learning precise, closed-form physical relations that alter image geometry based on lens parameters. + +Dynamics of Rolling and Bouncing Movements: This task involves simulating the motion of a rolling or bouncing ball on an inclined surface, incorporating both linear and rotational dynamics (Sec. 5). Providing 100,000 image pairs that capture the ball's position and rotation over time, it assesses models' ability to predict time-series dynamics and handle higher-order physical relations inherent in motion equations. + +Collectively, these tasks offer a diverse framework for assessing the strengths and limitations of generative models in learning and replicating complex physical simulations from image data. The subsequent sections delve into each task in detail, outlining dataset generation, underlying physics, evaluation metrics, and baseline model performances. + +# 3. Urban Sound Propagation + +The simulation of wave propagation in complex scenes, including diffraction and reflection at multiple objects is a representative task for a wide range of physical problems which are typically solved via (expensive) iterative solving of higher order differential equations. Due to its intuitive setting and accessible data, we investigate wave propagation problems by a concrete example where sound is propagated in urban environments. + +Dataset. We sampled $25k$ geo-locations across 10 cities, leveraging open-source geo-data from the Overpass API1 and processed it with GeoPandas [21] to represent urban en + +virements in a $500\mathrm{m}^2$ area with buildings and open spaces encoded in black and white pixels, respectively. Our selection criteria ensured diverse urban scenarios by mandating a minimum of 10 buildings within 200 meters and a 50-meter clearance from buildings to the sample point. We utilized the NoiseModelling v4 framework [4], adhering to the CNOSSOS-EU standards [23], to simulate sound propagation maps, placing sound receivers on a grid with a step-size of 5 meters and at building boundaries, excluding indoor placements (see Appendix B.3 for details). The resulting simulation output was then interpolated into $512\mathrm{x}512$ or $256\mathrm{x}256$ pixel maps, representing decibel levels on a $[0,255]$ scale. Four simulation tasks (Baseline, Diffraction, Reflection, Combined) were defined to model sound propagation under varying conditions, accumulating in a dataset of 100,000 samples. The Baseline task models a constant $95\mathrm{dB}$ noise source at $500\mathrm{Hz}$ , excluding diffraction and reflection. The Diffraction task adds horizontal sound wave diffraction, the Reflection task simulates sound reflections with a standard absorption coefficient, and the Combined task introduces variable sound levels (60-115 dB) and environmental factors, incorporating both reflection and diffraction to mimic realistic urban sound behavior. Refer to Fig. 2 and Appendix B.6 for a detailed description and visualization of the dataset generation pipeline. + +Physics of Sound Propagation. Mathematically, the propagation of sound over time is described via partial differential wave equations. Due to space constraints and the practical nature of our problem setting, we will neglect the derivation from continuous wave equations and directly focus on the discrete and iterative implementations of sound propagation which have been applied for our ground-truth simulations. Following [35], for a discrete set of receivers $R$ , the amplitude $L_{R_k}^j$ of receiver $R_k$ at frequency $j$ is computed via iterative differences: + +$$ +L _ {R _ {k}} ^ {j} = L _ {W} ^ {j} - A _ {\text {d i v} R _ {k}} - A _ {\text {a t m} R _ {k}} ^ {j} - A _ {\text {d i f} R _ {k}} ^ {j} \quad (j, k \in \mathbb {N}), \tag {1} +$$ + +where $L_W^j$ represents the source level, $A_{div_{R_k}}$ captures the geometrical spreading, $A_{atm_{R_k}}^j$ denotes the atmospheric absorption, and $A_{dif_{R_k}}^j$ models diffraction. Note that the ground effect $A_{grd_{R_k}}^j$ is neglected in our study. + +Additionally, the model accounts for reflections by adjusting the power level $L_W^j$ based on the absorption coefficient $\alpha_{vert}$ of the surfaces involved. This adjustment is performed using the equation + +$$ +L _ {W} ^ {(n _ {r e f})} = L _ {W} ^ {(n _ {r e f} - 1)} + n _ {r e f} \times 1 0 \log_ {1 0} (1 - \alpha_ {v e r t}) \tag {2} +$$ + +where $n_{ref}$ indicates the number of reflections considered. Specular reflections are modeled using the image receiver method, which provides a computationally efficient way to account for the angle of incidence being equal to the angle of reflection [35]. The sound level at each receiver + +![](images/b4a0020648eab743670178776b22464d496d013addaa94f228fd2952c73b1b06.jpg) +Figure 2. The sampling pipeline for the sound propagation dataset utilizes the NoiseModelling framework [4] to generate sound propagation maps based on specific urban layouts. The generators are then trained to replicate these sound propagation patterns for given locations and source parameters. Predictions are evaluated by specifically analyzing errors in relation to the line of sight (see Appendix B.1 for details). + +reflects the cumulative effect of direct, diffracted, and reflected sound paths. As the number of reflections increases, the complexity of calculating the power level $L_W^j$ also increases, growing in $O(N^2)$ , where $N$ denotes the number of reflective surfaces considered in the computation. + +Evaluation. Our baseline evaluation focuses on the ability of standard image-to-image generative models to accurately predict sound propagations. Prediction quality is determined by the pixel-wise difference between actual and predicted sound distributions. We use Mean Absolute Error (MAE) to quantify the average magnitude of prediction errors. Additionally, we introduce Weighted Mean Absolute Percentage Error (wMAPE) to specifically penalize errors in predictions that inaccurately show high sound amplitudes in regions expected to have low amplitudes, such as areas behind buildings. To evaluate model performance in capturing sound reflections or diffractions, metrics were specifically calculated for areas both within and outside the direct line of sight (LoS and NLoS) from the sound source. This methodology allows us to assess model predictions for sound propagations through direct paths as well as reflections and diffractions. + +Analysis. The results in Tab. 1 show that generative models significantly outperform traditional sound propagation simulations in processing speed, achieving up to a $20k$ factor improvement in runtime (see Tab. 9 in the Appendix for further analysis). Regarding the prediction quality, we observe that models encountering tasks with reflections or diffractions show an increase in MAE for NLoS regions across + +Table 1. Quantitative evaluation across all tasks for all architectures with a batch size of 16 during inference. The complete results containing the Combined task are given in Appendix B.7. Best overall results in bold, second best underlined. + +
ModelTaskMAE ↓wMAPE ↓Runtime/ ↓ Sample (ms)
LoSNLoSLoSNLoS
SimulationBase0.000.000.000.00204700
convAEBase3.672.7420.2467.130.128
VAE [25]Base3.922.8421.3375.580.124
UNet [31]Base2.291.7312.9137.570.138
Pix2Pix [20]Base1.731.199.366.750.138
DDPM [18]Base2.423.2615.5751.083986.353
SD(w,CA) [30]Base3.763.3417.4235.182961.027
SDBase2.121.0813.2332.462970.86
DDBM [38]Base1.612.1717.5065.243732.21
SimulationDif.0.000.000.000.00206000
convAEDif.3.598.0413.7732.090.128
VAEDif.3.928.2214.4632.570.124
UNetDif.0.943.274.2222.360.138
Pix2PixDif.0.913.363.5118.060.138
DDPMDif.1.593.278.2520.303986.353
SD(w,CA)Dif.2.467.7210.1431.232961.027
SDDif.1.335.078.1524.452970.86
DDBMDif.1.353.3511.2223.563732.21
SimulationRefl.0.000.000.000.00251000
convAERefl.3.836.5620.6793.540.128
VAERefl.4.156.3221.5792.470.124
UNetRefl.2.295.7212.7580.460.138
Pix2PixRefl.2.144.7911.3030.670.138
DDPMRefl.2.747.9317.8580.383986.353
SD(w,CA)Refl.3.816.8219.7881.612961.027
SDRefl.2.535.2615.0455.272970.86
DDBMRefl.1.936.3818.3479.133732.21
+ +![](images/92c5beb2928dbe44ce383c4da7fbc92cb7d8fbe07a70808739e69d48f2cebc24.jpg) +Figure 3. Qualitative results comparing the ground-truth simulation with the prediction for a single sample within the reflection task. Additional results can be found in Appendix B.7 + +all tested architectures. For the Pix2Pix model this rise in NLoS error is moderate, while this effect is more pronounced in the UNet and Diffusion models. Models without skip connections, such as VAE and ConvAE, tend to blur most of the image and struggle to capture any reflection patterns accurately. This trend highlights how reflections and diffractions introduce complex higher-order dependencies (see Eq. (12) and Appendix B.4) that complicate the image-to-image sound propagation tasks. Although the models successfully replicate sound reflection patterns visually (see Fig. 3), there exists a significant gap between its visual outputs and the actual acoustic accuracy. This discrepancy suggests that while visually detailed, the models do not fully capture the complexities of sound physics. + +# 4. Lens Distortion + +Our second benchmark is designed to investigate a class of physical problems for which solutions can be computed directly via closed equations, rather than iterative solvers. We find a representative and intuitive task in the modeling of optical lens distortions [7, 14], a common problem in computer vision. Generative models are trained to simulate the appearance of faces distorted by lens effects, using the Brown-Conrady [12] model. These distortions are computed using OpenCV [5], providing ground truth for training and evaluation of generative models. + +Dataset. We leverage a subset of the CelebA dataset [29], comprising 50,000 facial images, which were cropped to center the faces within a $256 \times 256$ pixel frame. The resulting images where split into two distinct groups to isolate the effects of tangential distortion: one focusing on horizontal distortion and the other on vertical distortion. This allows us to specifically measure the impact of each type of distortion on the accuracy of the generative models (see Fig. 4 and Appendix C for more details). + +Physics of Lens Distortions. The Brown-Conrady model is a widely adopted method to describe and correct lens distortions. It provides a mathematical framework to represent both radial and tangential distortions through a series of coefficients. For our study, we concentrate exclusively on the tangential distortion components, utilizing relevant coefficients to model the distortion characteristics of specific lenses. The mathematical representation of tangential distortion is expressed through the following equation, where $x_{dist}$ and $y_{dist}$ denote the distorted coordinates: + +$$ +x _ {d i s t} = x + \left[ 2 p _ {1} x y + p _ {2} \left(r ^ {2} + 2 x ^ {2}\right) \right], \tag {3} +$$ + +$$ +y _ {d i s t} = y + \left[ p _ {1} \left(r ^ {2} + 2 y ^ {2}\right) + 2 p _ {2} x y \right], +$$ + +where $r^2 = x^2 + y^2$ represents the squared radial distance from the image center, and $p_1$ and $p_2$ are the coefficients that model the tangential distortion effects. By strategically adjusting $p_1$ and $p_2$ and isolating their effects via controlled nullification, we can identify how each parameter individually contributes to the image distortion: + +$$ +x _ {d i s t} = \left\{ \begin{array}{l l} x + 2 p _ {1} x y & \text {i f} p _ {2} = 0, \\ x + p _ {2} \left(r ^ {2} + 2 x ^ {2}\right) & \text {i f} p _ {1} = 0, \end{array} \right. \tag {4} +$$ + +$$ +y _ {d i s t} = \left\{ \begin{array}{l l} y + p _ {1} (r ^ {2} + 2 y ^ {2}) & \text {i f} p _ {2} = 0, \\ y + 2 p _ {2} x y & \text {i f} p _ {1} = 0. \end{array} \right. +$$ + +These conditional equations illustrate the isolated impact of each coefficient, revealing that the correction complexity for $y_{dist}$ escalates with the term $p_1(r^2 + 2y^2)$ when $p_1$ is non-zero and $p_2$ is zero. Conversely, the complexity of $x_{dist}$ + +correction intensifies with $p_2(r^2 + 2x^2)$ under the condition that $p_2$ is non-zero and $p_1$ is zero. + +Evaluation. Given the challenges of accurately capturing lens distortion through traditional pixel-based metrics, our evaluation strategy is specifically designed to focus on accurately reproducing designated landmarks in the images. We detect facial landmarks using a 2D facial landmark detection method based on the Facial Alignment Network (FAN) [8], applied to the faces in the CelebA dataset (see Appendix C for details). We use these landmarks to assess the geometric distortions introduced by lens settings, represented by the coefficients $p_1$ and $p_2$ . We then compute the Euclidean distance between the actual position of a landmark in the true distorted image and its predicted position in the image generated by the models as a quality metric. This measurement is performed separately for the horizontal (X) and vertical (Y) coordinates, resulting in X Error and Y Error respectively. This approach not only allows us to gauge the accuracy of the models in simulating the specific effects of lens distortion on facial features but also helps in understanding how different distortion coefficients impact the positional accuracy of landmarks along each axis. + +Analysis. Our experiments with the CelebA dataset under varying conditions of $p_1$ and $p_2$ showcase different patterns in error rates that align with the theoretical complexity of the distortion model. As shown in Tab. 2 and Fig. 5, the generative models exhibit differing accuracies in landmark prediction under the influence of lens distortion parameters $p_1$ and $p_2$ . Notably, the increased complexity in $x_{dist}$ and $y_{dist}$ equations under separate conditions ( $p_1$ nonzero vs. $p_2$ nonzero) slightly correlates with heightened error rates in respective dimensions. The models exhibit higher Y Error when only $p_1$ is active, indicative of the complex correction challenge introduced by the higher order dependency on $y$ . Similarly, a slight increase in X error and a slight decrease in Y error are observed when $p_2$ is exclusively nonzero. + +![](images/cc25ea096a192917527fde18b87997a2ecf22ccdbb5c567c7e5ddc1765986527.jpg) +Figure 4. Sampling and evaluation pipeline for the lens distortion dataset. The Brown-Conrady distortion model generates the true distorted images based on parameters $p_1$ and $p_2$ (we depict a chess pattern for visualization). The conditioned generators are then trained to replicate these distortions for given images and parameters. The models are evaluated by comparing predicted against true facial landmarks, using a 2D facial landmark detection based on the Facial Alignment Network (FAN) [8]. + +![](images/83fe1329682073524b4b8ea7e8f635861464b06e3fc097388d51e2dab1135714.jpg) +Figure 5. Qualitative visualization of lens distortion predictions using different generative models on a CelebA dataset sample. $p_2$ distortion: Original image at top-left; subsequent panels show U-Net, Pix2Pix, Diffusion models. Red dots mark actual landmark positions and blue dots for predictions. + +Table 2. Quantitative X-Y error and shift comparison across different models for varying activation of parameters $p_1$ and $p_2$ with a batch size of 16 during inference. Best overall results in bold, second best underlined. + +
ModelComb ↓X ↓ ErrY ↓ Err|X-Y| ShiftRuntime/ Sample (ms)
Sim.0.000.000.000.00153.205
convAE11.936.758.131.380.110
VAE11.536.557.831.280.122
UNet2.821.282.150.870.118
Pix2Pix2.000.991.430.440.122
DDPM1.930.941.390.453970.603
SD(w,CA)3.091.592.210.622991.678
SD2.791.412.010.602997.576
Sim.p1 = 0, p2 ≠ 0
convAE0.000.000.000.00153.205
10.568.354.772.210.110
VAE10.408.264.623.640.122
UNet2.361.331.600.270.117
Pix2Pix1.771.021.140.130.123
DDPM2.131.391.230.163970.603
SD(w,CA)2.851.601.940.342991.678
SD2.441.381.640.262997.576
+ +# 5. Dynamics of bouncing movements + +The third physical problem under investigation is the movement of a rolling or bouncing ball. The aim here is to evaluate the ability of generative models to map kinematic movements of physics - a task which should generalize well to the investigation of various time series problems. Specifically, we define the task to predict the position and rotation of a ball along an inclined surface for a defined time interval after the input image. Only the bouncing case is discussed below; however, all results for the rolling ball can be found in Appendix D.3. + +Dataset. We generated our data set with Pymunk [3], a + +![](images/8dbd5c8c665714692e47243196b5096b0fd8654c58e8739a751a2c945931c44b.jpg) +Figure 6. Physical equations describing the bouncing ball movement divided in 3 parts: free fall (green), impact & bounce of the ball on the ground (red dot) and oblique throw with initial velocity from ball bounce (orange). + +$$ +F _ {g} = m * a \Leftrightarrow m * \ddot {y} = - m * g \Leftrightarrow \ddot {y} = - g \tag {5} +$$ + +$$ +\begin{array}{l} m * \ddot {y} + d * \dot {y} - c * (r - y) = - m * g \tag {6} \\ m * \ddot {x} = m * g * \sin (\beta) \text {a l o n g x a x i s} \\ \end{array} +$$ + +Python physics simulation toolbox. For the bouncing case we use 50k pairs of training images (256x256px) and perform the evaluation of the generative networks with 1600 images. Three parameters on the input images are variable: the ground inclination, the start height of the ball movement and the ball position. Additionally, the time interval between the input and target images is also variable. All other physical properties of the problem (ball's mass or friction coefficients) are kept constant. A multi-coloured structure with randomly distributed lines is superimposed on the simulation images in order to obtain greater diversity in the training data and is also used to analyze consistency, e.g. how well the generative networks can map fine multicoloured patterns (see Fig. 7). The total time required to create a sample using the available physics simulation CPU code (with a customized background) is 3.8 seconds, while the inference times for the generative models remain consistent with those reported in Tab. 1. + +Physics of bouncing balls. To model the physics behind the movement of a bouncing ball, the motion is divided into different sections which are represented with different colors in Fig. 6. First, the ball is released from a defined height without any velocity and falls down according to the Newton's law of the free fall (green part) (see equation 5). The impact of the ball on the ground and its rebound can be described in simplified terms by a spring-damper model. The aim of this second part of the movement (red dot) is to calculate the rebound speed and angle. This is done with equation 6, which contains the constant of the elements of the simplified model (see details in Appendix D.1) and the angle corresponds to the ground inclination for the first impact. Finally, the orange part of the movement is assumed to be an oblique throw with the initial velocity $v_{0}$ calculated with equation 6. The equations of motion describing + +![](images/294f33bcc5ed3449a6de618b8c53515dc21024d86629cf3012bb64ce2fd83e2c.jpg) +Figure 7. The sampling and evaluation pipeline for the rolling and bouncing ball movements simulates ball trajectories. A physics-based model is used to predict the ball's movement. Evaluations focus on the accuracy of predictions related to key movement metrics, including bounce height and horizontal displacement (see Appendix D for details). + +this oblique throw along the X and Y axes are also derived from the Newton's law (see equation 7). When the ball hits the ground for the second time, the principle "angle of incidence = angle of reflection" is used to determine the angle of rebound $\theta$ and the next part of the motion is again defined by equations 6 and 7. + +Evaluation. The generated images are examined according to 4 precise error criteria: the ball position split in X and Y directions, where the error is defined as the number of pixels between the theoretical and the predicted ball center. The second error measure is the ball rotation describing its angle in degrees during the movement due to the inclined ground surface. The next quality indicator of the algorithms is the out-of-roundness of the ball also measured in pixels by taking all identified points of the ball contour and calculating the difference between the theoretical ball radius of 15px and the predicted one. The standard deviation of all radius errors is taken as the error measure. Finally, the correct representation of the ground inclination is checked: the angular error between simulation and prediction is also measured in degrees. In addition, there are three "general" error criteria that give preciser indications of the correct representation of the physics which are explained with the complete results in the Appendix D. + +The results of the seven examined generative algorithms are summarized in table 3 (only for the four most important error criteria). The last column of the table contains the average number of non-evaluable predicted images over the four error criteria. A detailed analysis of the advantages and disadvantages of the individual methods with all the evaluation parameters can be found in the Appendix D, as well as the evaluations obtained for individual runs for the rolling and bouncing case. + +
Position XPosition YRotationRoundnessError
convAE4.24 ± 3.96.08 ± 5.912.2 ± 8.61.06 ± 0.099%
VAE4.69 ± 6.16.25 ± 6.931.0 ± 400.90 ± 0.195%
UNet5.53 ± 7.510.8 ± 1215.2 ± 230.74 ± 0.228%
Pix2Pix6.28 ± 8.011.7 ± 1317.2 ± 210.56 ± 0.111%
DDPM7.91 ± 9.015.5 ± 1432.9 ± 340.61 ± 0.25.7%
SD(w.CA)40.0 ± 4924.8 ± 2361.1 ± 520.53 ± 0.27.3%
SD8.55 ± 1216.2 ± 1434.2 ± 380.47 ± 0.12%
+ +Table 3. Prediction results for the bouncing ball experiment in mean±std and mean percentage of non-evaluable predictions for the four criteria (best overall results in bold). Average runtime of the simulation: 3.8s, generative models: same as inTab. 1 + +Overall, the GAN (Pix2Pix model) provides the most stable representation of the bouncing ball problem. It achieves the best results across all evaluation criteria, with the most reliable physics predictions and fewest failed samples. The UNet also represents the ball position and angle on the generated samples very well, but produces more error images that cannot be evaluated with mostly either no ball or only a shredded shape. The greatest weakness of this algorithm lies in the clean representation of the ball roundness. The auto-encoder approaches (VAE and convAE) produce very blurred images in which the ball is often distorted and hardly recognizable. Therefore, the mean error values are not very representative, as over $90\%$ of the predictions cannot be analyzed. + +The denoising diffusion approach (DDPM) produces slightly poorer results in absolute terms. Especially the prediction of the ball rotation is worse, with an error approximately twice as high compared to the two other approaches. Finally, to test another type of diffusion, two variations of stable diffusion were compared. The one with cross-attention is generating more error pictures and show high error values. In contrast, the normal stable diffusion performs very well: the mapping of the position is almost as accurate as with the Pix2Pix approach and the representation of the ball roundness is the best. In addition, the network generates very few error images, so that it could almost be placed on the same level as the GAN but comes as second due to a poorer angle mapping. + +As can be seen in the qualitative result images below (figure 8), GAN and diffusion (especially the SD) networks depict small details much more accurately. The background structure of the images with the different colors is completely blurred by the UNet but not by the GAN and the diffusion. The DDPM predictions are blurred because the images have been scaled up from $64\mathrm{px}$ to $256\mathrm{px}$ . If the output of the diffusion model is set to $256\mathrm{px}$ , the target image is sharp again but several or no balls and many more artefacts appear on the net predictions (see figure 25 in the Appendix). To summarize, the Pix2Pix and Stable Diffusion approaches also are visually performing the best. + +Finally, slight correlations can be observed between the simulation's parameters and the generated errors. For example the ball position in the X direction is predicted worse for all nets when the time interval between input and prediction images becomes larger. More details on the correla + +![](images/528cde0b04af22a6d9af7c0f885c082acc046ffc275f10d212ea074d10668656.jpg) +Figure 8. Qualitative results for predictions of the bouncing ball problem. Initial State and True label image at top-left; subsequent panels show results for VAE, UNet, Pix2Pix and Diffusion models. + +tions of each method can be found in the Appendix D.2.1. Analysis. When comparing the result in table 3 (and tables 11, 13 in the Appendix) with the physical motion equations, it shows that the errors of the generative networks increase with higher order terms. The second-order terms, such as the movement along the X-axis and the ball angle represent the largest errors for the rolling scenario. In the bouncing case, a second-order derivative is also present to describe the ball movement in the Y-direction due to the ball's impact on the ground. Therefore the calculation rule along the Y-direction consists of zero, first and second derivative terms, making it more complex than the equation in the X-direction. This is reflected in the error values of all the investigated approaches: the Y-error clearly exceeds that in the X-direction, as indicated by the complexity of the motion equations, although the X-position and angular errors remain roughly the same between the rolling and bouncing cases. In general, it can be concluded that the errors of the analyzed approaches increase with the order and complexity of the physical formulation. + +# 6. Discussion + +With PhysicsGen, we present a comprehensive benchmark for the estimation of complex physical relations by image-to-image generative models. The three different datasets contained in PhysicsGen cover a wide range of diverse simulation scenarios and hence allow a systematic evaluation of learning based image generation models. + +Key insights from the baseline-evaluation. Despite the diversity of the three given physical problems and the fundamentally different network types used the base-line evaluation experiments, we can report several consistent findings: + +- Due to the unavailability of optimized GPU code for the used simulations it is impossible to provide a theoretically sound runtime comparison. Since a highly parallel implementation of the simulation algorithms is highly nontrivial (if possible at all), our analysis focuses on the runtime analysis of the generative models. However, replac + +ing physical simulations with GPU powered generative models allows high computational speed-ups in practice, especially when replacing highly iterative solvers like in the urban sound propagation with Pix2Pix or UNets (where we report speed-up factors of up to 20k). + +- In terms of physical correctness and consistency, our evaluation shows good results for simple 0th and 1st order problems, but also indicates that generative models mostly fail to predict more complex physical relations (which are typically formulated by higher order terms). This finding needs further investigation whether this is an effect of the principal theoretical limitations of current models or just a lack of sufficient training data. +- We also observe distinct error pattern for different model types which represents how the different model types are dealing with complexity and uncertainty. Please refer to Appendix D.2.1 for examples in the following exemplary discussion for the bouncing ball problem (similar effects can be observed in Appendix B.7 for the sound propagation tasks as well): while GANs often produce physically possible outputs, they still can fail to generate the correct output for given inputs. This results in consistent scenes with a complete ball at the wrong position. UNets, VAE and convAE on the other hand, often fail to produce consistent scenes and tend to blur uncertain predictions, resulting in deformed balls and destructed backgrounds. Lastly, diffusion models often produce multiple candidate solutions (multiple balls in one scene) and tend to fail to estimate more complex relations like ball rotation. This phenomenon also needs further investigation. + +Limitations. The provided run-time analysis is based on the comparison of practically available implementations and is not fair from a theoretical perspective as no efforts were made to optimize the simulation codes to GPUs. Also, even though we carefully selected the three proposed physical benchmark problems such that they are likely to transfer to a wide range of other simulation tasks, our evaluation provides only a small subset of possible applications of generative models in physical modeling. Hence, results from our benchmark might not transfer to all possible scenarios. + +Broader Impact Statement. The simulation of physical systems forms the backbone for a wide range of scientific and engineering tasks. Hence, it is important to have benchmarks for the emerging field of neural enhanced simulations to compare and validate these novel approaches. This work provides such a benchmark for the specific, but widely applicable case, where physical simulations are used to map between 2D input and output structures (visualized as images). Our analysis shows that while current generative models are promising significant speedups, there are still many open questions regarding their physical correctness. + +# Funding Acknowledgement + +The authors acknowledge the financial support by the German Federal Ministry of Education and Research (BMBF) in the program "Forschung an Fachhochschulen in Kooperation mit Unternehmen (FH-Kooperativ)" within the joint project "KI-Bohrer" under grant 13FH525KX1. + +# References + +[1] A. Argilaga. Fem-gan: A physics-supervised deep learning generative model for elastic porous materials. Materials, 2023. 2 +[2] Peter W Battaglia, Jessica B Hamrick, and Joshua B Tenenbaum. Simulation as an engine of physical scene understanding. Proceedings of the National Academy of Sciences, 110 (45):18327-18332, 2013. 1 +[3] Victor Blomqvist. Pymunk pythonic 2d physics library, 2007 - 2024. 6, 20 +[4] Erwan Bocher, Gwenael Guillaume, Judicael Picaut, Gwendall Petit, and Nicolas Fortin. Noise modelling: An open source gis based tool to produce environmental noise maps. ISPRS International Journal of Geo-Information, 8(3):130, 2019. 3, 4, 13, 15 +[5] G. Bradski. The OpenCV Library. Dr. Dobb's Journal of Software Tools, 2000. 5, 17 +[6] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym, 2016. 2 +[7] Duane Brown. Close-range camera calibration. 1971. 5 +[8] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d and 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, 2017. 5 +[9] Giuseppe Carleo, Ignacio Cirac, Kyle Cranmer, Laurent Daudet, Maria Schuld, Naftali Tishby, Leslie Vogt-Maranto, and Lenka Zdeborova. Machine learning and the physical sciences. Reviews of Modern Physics, 91(4):045002, 2019. 1, 2 +[10] Yilun Chen, Qi Wu, Yujie Sui, Yizhou Wang, and Song-Chun Zhu. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 2 +[11] Runmin Cong, Wenyu Yang, Chongyi Zhang, Weiand Li, Chunle Guo, and Sam Huang, Qingmingand Kwong. Pugan: Physical model-guided underwater image enhancement using gan with dual-discriminators. IEEE Transactions on Image Processing, 2023. 2 +[12] A. E. Conrady. Decentred Lens-Systems. Monthly Notices of the Royal Astronomical Society, 79(5):384-390, 1919. 5, 17 +[13] Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep reinforcement learning for continuous control, 2016. 2 +[14] Wolfgang Faig. Calibration of close-range photogrammetric systems: Mathematical formulation. Photogrammetric Engineering and Remote Sensing, 41, 1975. 5 + +[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. 1 +[16] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. 2 +[17] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti, Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator, 2022. 2 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. CoRR, abs/2006.11239, 2020. 2, 4, 11, 16 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1 +[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. CoRR, abs/1611.07004, 2016. 2, 4, 11, 16 +[21] Kelsey Jordahl, Joris Van den Bossche, Martin Fleischmann, Jacob Wasserman, James McBride, Jeffrey Gerard, Jeff Stratner, Matthew Perry, Adrian Garcia Badaracco, Carson Farmer, Geir Arne Hjelle, Alan D. Snow, Micah Cochran, Sean Gillies, Lucas Culbertson, Matt Bartos, Nick Eubank, maxalbert, Aleksey Bilogur, Sergio Rey, Christopher Ren, Dani Arribas-Bel, Leah Wasser, Levi John Wolf, Martin Journois, Joshua Wilson, Adam Greenhall, Chris Holdgraf, Filipe, and François Leblanc. geopandas/geopandas: v0.8.1, 2020.3 +[22] Anuj Karpatne, Imme Ebert-Uphoff, Sai Ravela, Hassan A. Babaie, and Vipin Kumar. Physics-guided neural networks (pgnn): An application in lake temperature modeling. arXiv preprint arXiv:2212.02500, 2022. 2 +[23] S Kephalopoulos, M Paviotti, and F Anfosso-Ledée. Common noise assessment methods in europe (cnossos-eu). Technical Report EUR 25379 EN, Publications Office of the European Union, Luxembourg (Luxembourg), 2012. JRC72550. 3 +[24] Byungsoo Kim, V. C. Azevedo, N. Thürey, Theodore Kim, M. Gross, and B. Solenthaler. Deep fluids: A generative network for parameterized fluid simulations. Computer Graphics Forum, 38, 2018. 2 +[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 2, 4, 11, 16 +[26] Daniel Ladwig, Martin Spitznagel, Jan Vaillant, Klaus Dorer, and Janis Keuper. Ai-guided noise reduction for urban geothermal drilling. 2024. 3 +[27] Adam Lerer, Sam Gross, and Rob Fergus. Learning physical intuition of block towers by example, 2016. 2 +[28] Wenbin Li, Seyedmajid Azimi, Ales Leonardis, and Mario Fritz. To fall or not to fall: A visual approach to physical stability prediction, 2016. 2 + +[29] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2, 4, 11, 16 +[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015. 2, 4, 11, 16 +[32] Yuxuan Song, Jingjing Gong, Yanru Qu, Hao Zhou, Mingyue Zheng, Jingjing Liu, and Wei-Ying Ma. Unified generative modeling of 3d molecules via bayesian flow networks, 2024. 2 +[33] Martin Spitznagel and Janis Keuper. Urban sound propagation: a benchmark for 1-step generative modeling of complex physical systems, 2024. 3 +[34] Nils Thuerey, Philipp Holl, Maximilian Mueller, Patrick Schnell, Felix Trost, and Kiwon Um. Physics-based Deep Learning. WWW, 2021. 2 +[35] Michael Vorländer. Auralization: Fundamentals of acoustics, modelling, simulation, algorithms and acoustic virtual reality. Springer Science & Business Media, 2007. 3, 14 +[36] Yang Yu and Yongming Liu. Physics-guided generative adversarial network for probabilistic structural system identification. Expert Systems with Applications, 239:122339, 2024. 2 +[37] Yufeng Zhang, Wei Li, Yangang Gao, and Shihong Li. Phys-diff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2106.03704, 2021. 2 +[38] Linqi Zhou, Aaron Lou, Samar Khanna, and Stefano Ermon. Denoising diffusion bridge models, 2023. 2, 4, 11, 16 +[39] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martin-Martin, Abhishek Joshi, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning, 2022. 2 \ No newline at end of file diff --git a/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/images.zip b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f24a96cd01a3fac476f71b387da7beedd697d96d --- /dev/null +++ b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:242b9b6d85aa2f7b6f594ccf61c157c2a5681dd2ad986183dae1a8545b9ee59c +size 620935 diff --git a/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/layout.json b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..baf69f4275ac10845a4657b97d31326bbfd61769 --- /dev/null +++ b/2025/PhysicsGen_ Can Generative Models Learn from Images to Predict Complex Physical Relations_/layout.json @@ -0,0 +1,6887 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 103, + 467, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 103, + 467, + 140 + ], + "spans": [ + { + "bbox": [ + 142, + 103, + 467, + 140 + ], + "type": "text", + "content": "PhysicsGen: Can Generative Models Learn from Images to Predict Complex Physical Relations?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 160, + 483, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 160, + 483, + 203 + ], + "spans": [ + { + "bbox": [ + 126, + 160, + 483, + 203 + ], + "type": "text", + "content": "Martin Spitznagel1, Jan Vaillant2 and Janis Keuper1,3 \n1IMLA, Offenburg University, 2Herrenknecht AG, 3Mannheim University \n{first.last}@hs-offenburg.de" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 258, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 296, + 485 + ], + "type": "text", + "content": "The image-to-image translation abilities of generative learning models have recently made significant progress in the estimation of complex (steered) mappings between image distributions. While appearance based tasks like image inpainting or style transfer have been studied at length, we propose to investigate the potential of generative models in the context of physical simulations. Providing a dataset of 300k image-pairs and baseline evaluations for three different physical simulation tasks, we propose a benchmark to investigate the following research questions: i) are generative models able to learn complex physical relations from input-output image pairs? ii) what speedups can be achieved by replacing differential equation based simulations? While baseline evaluations of different current models show the potential for high speedups (ii), these results also show strong limitations toward the physical correctness (i). This underlines the need for new methods to enforce physical correctness. Data, baseline models and evaluation code: http://www.physics-gen.org." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 512, + 135, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 135, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 135, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 605 + ], + "type": "text", + "content": "The numerical simulation of physical processes and relations is a crucial tool in a wide range of scientific and engineering tasks. Traditionally depending on the solution of various types of differential equations, research of physical simulation methods is increasingly shifting towards data-driven and learnable approaches [9]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "In this work, we focus on a narrow but practically important sub-field: the inference of physical relations from images. Extending Tenenbaum's seminal idea of a intuitive physics engine [2], where machine learning models have been trained to predict coarse physical properties from images, we investigate the capabilities of modern generative models to learn detailed, complex and physically correct mappings from image pairs. Given such abilities, generative models like GANs [15] and Diffusion models [19] could be" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 231, + 549, + 574 + ], + "blocks": [ + { + "bbox": [ + 318, + 231, + 549, + 574 + ], + "lines": [ + { + "bbox": [ + 318, + 231, + 549, + 574 + ], + "spans": [ + { + "bbox": [ + 318, + 231, + 549, + 574 + ], + "type": "image", + "image_path": "b3433c8f4aee8bc1cbab065d16d4233edf15af085231732bc8fe61661585766f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 586, + 555, + 707 + ], + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 707 + ], + "type": "text", + "content": "Figure 1. Overview of the physical problems, baseline generative models and their evaluation. A: We introduce three complex physical simulation tasks with 100k input-output image pairs each, providing ground truth simulations based on differential equations with varying complexity. B: We evaluate all tasks on independently trained image translation models; only results for the sound propagation task are visualized in this figure. C: While the evaluation of the baseline models shows a general ability of generative image models to learn physical relations from images, we observe significant performance drops for tasks that require a higher order term in the differential equations of their simulation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "11125" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "utilized to introduce hard to formalize but observable relations into physical models. Due to the fast inference times, generative models could also be used to accelerate simulations [34]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 192 + ], + "type": "text", + "content": "Introducing the PhysicsGen Benchmark. A first step towards such data-driven models is a thorough analysis of the capabilities of current generative approaches. Hence, we introduce a novel benchmark that allows to systematically investigate the inference of a diverse set of physical problems from images (as visualized in Fig. 1):" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 295, + 517 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 194, + 295, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 229 + ], + "type": "text", + "content": "- We introduce PhysicsGen, a novel Benchmark for the evaluation of generative image models learning to infer physical simulations tasks from images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 229, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 295, + 289 + ], + "type": "text", + "content": "- PhysicsGen provides a Dataset with a total of 300k image-pairs, representing three diverse physical simulation problems: i) iterative wave propagation (Sec. 3), ii) closed form lens distortion (Sec. 4) and iii) time series prediction of motion dynamics (Sec. 5)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 289, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 456 + ], + "type": "text", + "content": "- We provide a baseline evaluation that measures speedups and physical accuracy compared to full physical simulations across multiple generative image-to-image models. The evaluated models include various generative image-to-image models: Generative Adversarial Networks (GANs) [16], specifically Pix2Pix [20]; an U-Net [31]; Convolutional Autoencoders (convAE) and Variational Autoencoders (VAEs) [25]. We also assess diffusion models, including Denoising Diffusion Probabilistic Models (DDPMs) [18], Stable Diffusion [30], and Denoising Diffusion Bridge Models (DDBMs) [38]. (see Appendix A for details of the model architectures and training - the model architectures and training procedures are identical for all three tasks)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 517 + ], + "type": "text", + "content": "- Our Analysis of the baseline results shows potentially high speed-ups at good accuracies for simple 1st order simulation tasks. However, our experiments also show that current generative models have fundamental problems to learn higher order physical relations." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 56, + 526, + 146, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 146, + 538 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 146, + 538 + ], + "type": "text", + "content": "1.1. Related Work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 294, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 294, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 294, + 629 + ], + "type": "text", + "content": "The integration of learnable methods into physical modeling is a wide field with a rich body of literature which is very hard to capture in the limited space of this paper. In the following, we try to refer to the most important works in direct relation to our benchmark and some more general key concepts. We refer the reader to [34] and [9] for a comprehensive overview of the field." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": "Generative models for physical simulations. The integration of physical principles within generative models is an active and burgeoning area of research. Recent models such as PUGAN [11] and FEM-GAN [1] have demonstrated the potential of combining GANs with physical modeling to enhance performance in environments governed by complex physical laws. Advances in fluid dynamics and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "structural system identification using physics-guided GANs have shown significant improvements in efficiency and precision by incorporating physics-based loss functions and simulations [24, 36]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 121, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 251 + ], + "type": "text", + "content": "Predicting physical relations. Also, significant progress has been made in understanding physics through machine learning. Initial research demonstrated that models could comprehend the dynamics of block towers, progressing beyond simple memorization to genuinely understanding physical interactions [27]. Practical applications have been explored, including fall detection using body part tracking through machine learning [28]. Additionally, advancements in generating physically plausible human animations highlight the potential of incorporating physical principles into machine learning frameworks [37]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 251, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 395 + ], + "type": "text", + "content": "Grey-Box Models. Physics-guided approaches in AI have laid a solid foundation for incorporating physical laws into model training. Prior research has demonstrated the effectiveness of embedding physical constraints within AI models to enhance their performance and reliability [10, 22]. In generative models, diffusion models have shown promise but often face challenges when generating complex geometrical structures due to issues like multi-modality and noise sensitivity. The development of Geometric Bayesian Flow Networks addresses these challenges by enhancing the robustness and efficiency of molecule geometry modeling [32]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 396, + 553, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 526 + ], + "type": "text", + "content": "Physical Simulations for Reinforcement Learning. Deep reinforcement learning settings benefit from a substantial availability of physics-based environments and datasets, which facilitate the training of models to understand and predict physical phenomena [13]. Resources such as OpenAI Gym, Robosuite, and PyBullet provide robust platforms for developing and testing reinforcement learning models in simulated physical settings [6, 17, 39]. However, in the domain of generative AI, there is a notable scarcity of physics-informed datasets and environments, limiting the ability to train models on complex physical systems." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 548, + 473, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 473, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 473, + 560 + ], + "type": "text", + "content": "2. Benchmark Tasks Overview" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 568, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 553, + 652 + ], + "type": "text", + "content": "Selecting an appropriate set of tasks for the PhysicsGen benchmark is crucial to comprehensively evaluate the capabilities of generative image models in inferring complex physical relations. Given the vast diversity of possible physical simulation tasks, it is impractical to encompass the entire spectrum. Therefore, we strategically selected three distinct tasks based on the following criteria:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 654, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 688 + ], + "type": "text", + "content": "- Diversity of Physical Problems: To cover different types of physical phenomena, ensuring that models are tested across various domains." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "- Variety of Numerical Solution Strategies: Incorporating both iterative and closed-form (non-iterative) solvers" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "11126" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 294, + 95 + ], + "type": "text", + "content": "to assess models' adaptability to different computational approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 294, + 168 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 96, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 294, + 119 + ], + "type": "text", + "content": "- Simplicity of Evaluation Metrics: Facilitating straightforward and effective model performance evaluation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 121, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 294, + 168 + ], + "type": "text", + "content": "- Accessibility to the ML Community: Choosing intuitively understandable tasks without requiring specialized physics knowledge, thereby making the benchmark widely applicable to the machine learning community." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 171, + 294, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 294, + 194 + ], + "type": "text", + "content": "Based on these objectives, we introduce the following three simulation tasks within the PhysicsGen benchmark:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 194, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 294, + 277 + ], + "type": "text", + "content": "Urban Sound Propagation: This task models the behavior of sound waves in urban environments, accounting for phenomena such as diffraction and reflections (Sec. 3). By providing 100,000 input-output image pairs based on differential equation-based simulations, it challenges generative models to accurately capture intricate sound distribution patterns influenced by urban structures [26, 33]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 278, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 294, + 361 + ], + "type": "text", + "content": "Lens Distortion: Focusing on optical aberrations, this task utilizes the Brown-Conrady distortion model to simulate lens-induced geometric distortions in images (Sec. 4). With 100,000 image pairs depicting varying degrees of distortion, models are tasked with learning precise, closed-form physical relations that alter image geometry based on lens parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 362, + 294, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 294, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 294, + 457 + ], + "type": "text", + "content": "Dynamics of Rolling and Bouncing Movements: This task involves simulating the motion of a rolling or bouncing ball on an inclined surface, incorporating both linear and rotational dynamics (Sec. 5). Providing 100,000 image pairs that capture the ball's position and rotation over time, it assesses models' ability to predict time-series dynamics and handle higher-order physical relations inherent in motion equations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 459, + 294, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 294, + 530 + ], + "type": "text", + "content": "Collectively, these tasks offer a diverse framework for assessing the strengths and limitations of generative models in learning and replicating complex physical simulations from image data. The subsequent sections delve into each task in detail, outlining dataset generation, underlying physics, evaluation metrics, and baseline model performances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 542, + 205, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 205, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 205, + 556 + ], + "type": "text", + "content": "3. Urban Sound Propagation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 562, + 294, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 294, + 657 + ], + "type": "text", + "content": "The simulation of wave propagation in complex scenes, including diffraction and reflection at multiple objects is a representative task for a wide range of physical problems which are typically solved via (expensive) iterative solving of higher order differential equations. Due to its intuitive setting and accessible data, we investigate wave propagation problems by a concrete example where sound is propagated in urban environments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 658, + 294, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 294, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 294, + 694 + ], + "type": "text", + "content": "Dataset. We sampled " + }, + { + "bbox": [ + 55, + 658, + 294, + 694 + ], + "type": "inline_equation", + "content": "25k" + }, + { + "bbox": [ + 55, + 658, + 294, + 694 + ], + "type": "text", + "content": " geo-locations across 10 cities, leveraging open-source geo-data from the Overpass API1 and processed it with GeoPandas [21] to represent urban en" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "text", + "content": "virements in a " + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "inline_equation", + "content": "500\\mathrm{m}^2" + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "text", + "content": " area with buildings and open spaces encoded in black and white pixels, respectively. Our selection criteria ensured diverse urban scenarios by mandating a minimum of 10 buildings within 200 meters and a 50-meter clearance from buildings to the sample point. We utilized the NoiseModelling v4 framework [4], adhering to the CNOSSOS-EU standards [23], to simulate sound propagation maps, placing sound receivers on a grid with a step-size of 5 meters and at building boundaries, excluding indoor placements (see Appendix B.3 for details). The resulting simulation output was then interpolated into " + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "inline_equation", + "content": "512\\mathrm{x}512" + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "inline_equation", + "content": "256\\mathrm{x}256" + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "text", + "content": " pixel maps, representing decibel levels on a " + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "inline_equation", + "content": "[0,255]" + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "text", + "content": " scale. Four simulation tasks (Baseline, Diffraction, Reflection, Combined) were defined to model sound propagation under varying conditions, accumulating in a dataset of 100,000 samples. The Baseline task models a constant " + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "inline_equation", + "content": "95\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "text", + "content": " noise source at " + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "inline_equation", + "content": "500\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 72, + 553, + 371 + ], + "type": "text", + "content": ", excluding diffraction and reflection. The Diffraction task adds horizontal sound wave diffraction, the Reflection task simulates sound reflections with a standard absorption coefficient, and the Combined task introduces variable sound levels (60-115 dB) and environmental factors, incorporating both reflection and diffraction to mimic realistic urban sound behavior. Refer to Fig. 2 and Appendix B.6 for a detailed description and visualization of the dataset generation pipeline." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 371, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 553, + 491 + ], + "type": "text", + "content": "Physics of Sound Propagation. Mathematically, the propagation of sound over time is described via partial differential wave equations. Due to space constraints and the practical nature of our problem setting, we will neglect the derivation from continuous wave equations and directly focus on the discrete and iterative implementations of sound propagation which have been applied for our ground-truth simulations. Following [35], for a discrete set of receivers " + }, + { + "bbox": [ + 313, + 371, + 553, + 491 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 371, + 553, + 491 + ], + "type": "text", + "content": ", the amplitude " + }, + { + "bbox": [ + 313, + 371, + 553, + 491 + ], + "type": "inline_equation", + "content": "L_{R_k}^j" + }, + { + "bbox": [ + 313, + 371, + 553, + 491 + ], + "type": "text", + "content": " of receiver " + }, + { + "bbox": [ + 313, + 371, + 553, + 491 + ], + "type": "inline_equation", + "content": "R_k" + }, + { + "bbox": [ + 313, + 371, + 553, + 491 + ], + "type": "text", + "content": " at frequency " + }, + { + "bbox": [ + 313, + 371, + 553, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 371, + 553, + 491 + ], + "type": "text", + "content": " is computed via iterative differences:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 495, + 553, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 495, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 319, + 495, + 553, + 513 + ], + "type": "interline_equation", + "content": "L _ {R _ {k}} ^ {j} = L _ {W} ^ {j} - A _ {\\text {d i v} R _ {k}} - A _ {\\text {a t m} R _ {k}} ^ {j} - A _ {\\text {d i f} R _ {k}} ^ {j} \\quad (j, k \\in \\mathbb {N}), \\tag {1}", + "image_path": "adf1c9706c2e969b8bcf366a7620d48e3c6f96ee0ed38b11a4eab141da5d3d77.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "inline_equation", + "content": "L_W^j" + }, + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "text", + "content": " represents the source level, " + }, + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "inline_equation", + "content": "A_{div_{R_k}}" + }, + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "text", + "content": " captures the geometrical spreading, " + }, + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "inline_equation", + "content": "A_{atm_{R_k}}^j" + }, + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "text", + "content": " denotes the atmospheric absorption, and " + }, + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "inline_equation", + "content": "A_{dif_{R_k}}^j" + }, + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "text", + "content": " models diffraction. Note that the ground effect " + }, + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "inline_equation", + "content": "A_{grd_{R_k}}^j" + }, + { + "bbox": [ + 313, + 518, + 553, + 579 + ], + "type": "text", + "content": " is neglected in our study." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "type": "text", + "content": "Additionally, the model accounts for reflections by adjusting the power level " + }, + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "type": "inline_equation", + "content": "L_W^j" + }, + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "type": "text", + "content": " based on the absorption coefficient " + }, + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "type": "inline_equation", + "content": "\\alpha_{vert}" + }, + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "type": "text", + "content": " of the surfaces involved. This adjustment is performed using the equation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 323, + 632, + 553, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 632, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 323, + 632, + 553, + 649 + ], + "type": "interline_equation", + "content": "L _ {W} ^ {(n _ {r e f})} = L _ {W} ^ {(n _ {r e f} - 1)} + n _ {r e f} \\times 1 0 \\log_ {1 0} (1 - \\alpha_ {v e r t}) \\tag {2}", + "image_path": "d1464541b2a0858256fcf6e080c3180cad35fa14a9f962cb02d3be628267c087.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "n_{ref}" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " indicates the number of reflections considered. Specular reflections are modeled using the image receiver method, which provides a computationally efficient way to account for the angle of incidence being equal to the angle of reflection [35]. The sound level at each receiver" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "type": "text", + "content": "http://overpass-api.de/api/map" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "11127" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 369, + 160 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 369, + 160 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 369, + 160 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 369, + 160 + ], + "type": "image", + "image_path": "b4a0020648eab743670178776b22464d496d013addaa94f228fd2952c73b1b06.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 51, + 555, + 150 + ], + "lines": [ + { + "bbox": [ + 376, + 51, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 376, + 51, + 555, + 150 + ], + "type": "text", + "content": "Figure 2. The sampling pipeline for the sound propagation dataset utilizes the NoiseModelling framework [4] to generate sound propagation maps based on specific urban layouts. The generators are then trained to replicate these sound propagation patterns for given locations and source parameters. Predictions are evaluated by specifically analyzing errors in relation to the line of sight (see Appendix B.1 for details)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 180, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 180, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 180, + 294, + 240 + ], + "type": "text", + "content": "reflects the cumulative effect of direct, diffracted, and reflected sound paths. As the number of reflections increases, the complexity of calculating the power level " + }, + { + "bbox": [ + 54, + 180, + 294, + 240 + ], + "type": "inline_equation", + "content": "L_W^j" + }, + { + "bbox": [ + 54, + 180, + 294, + 240 + ], + "type": "text", + "content": " also increases, growing in " + }, + { + "bbox": [ + 54, + 180, + 294, + 240 + ], + "type": "inline_equation", + "content": "O(N^2)" + }, + { + "bbox": [ + 54, + 180, + 294, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 180, + 294, + 240 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 180, + 294, + 240 + ], + "type": "text", + "content": " denotes the number of reflective surfaces considered in the computation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 241, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 241, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 57, + 241, + 296, + 443 + ], + "type": "text", + "content": "Evaluation. Our baseline evaluation focuses on the ability of standard image-to-image generative models to accurately predict sound propagations. Prediction quality is determined by the pixel-wise difference between actual and predicted sound distributions. We use Mean Absolute Error (MAE) to quantify the average magnitude of prediction errors. Additionally, we introduce Weighted Mean Absolute Percentage Error (wMAPE) to specifically penalize errors in predictions that inaccurately show high sound amplitudes in regions expected to have low amplitudes, such as areas behind buildings. To evaluate model performance in capturing sound reflections or diffractions, metrics were specifically calculated for areas both within and outside the direct line of sight (LoS and NLoS) from the sound source. This methodology allows us to assess model predictions for sound propagations through direct paths as well as reflections and diffractions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": "Analysis. The results in Tab. 1 show that generative models significantly outperform traditional sound propagation simulations in processing speed, achieving up to a " + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "inline_equation", + "content": "20k" + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": " factor improvement in runtime (see Tab. 9 in the Appendix for further analysis). Regarding the prediction quality, we observe that models encountering tasks with reflections or diffractions show an increase in MAE for NLoS regions across" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 232, + 553, + 511 + ], + "blocks": [ + { + "bbox": [ + 313, + 179, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 313, + 179, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 555, + 224 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation across all tasks for all architectures with a batch size of 16 during inference. The complete results containing the Combined task are given in Appendix B.7. Best overall results in bold, second best underlined." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 232, + 553, + 511 + ], + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 511 + ], + "type": "table", + "html": "
ModelTaskMAE ↓wMAPE ↓Runtime/ ↓ Sample (ms)
LoSNLoSLoSNLoS
SimulationBase0.000.000.000.00204700
convAEBase3.672.7420.2467.130.128
VAE [25]Base3.922.8421.3375.580.124
UNet [31]Base2.291.7312.9137.570.138
Pix2Pix [20]Base1.731.199.366.750.138
DDPM [18]Base2.423.2615.5751.083986.353
SD(w,CA) [30]Base3.763.3417.4235.182961.027
SDBase2.121.0813.2332.462970.86
DDBM [38]Base1.612.1717.5065.243732.21
SimulationDif.0.000.000.000.00206000
convAEDif.3.598.0413.7732.090.128
VAEDif.3.928.2214.4632.570.124
UNetDif.0.943.274.2222.360.138
Pix2PixDif.0.913.363.5118.060.138
DDPMDif.1.593.278.2520.303986.353
SD(w,CA)Dif.2.467.7210.1431.232961.027
SDDif.1.335.078.1524.452970.86
DDBMDif.1.353.3511.2223.563732.21
SimulationRefl.0.000.000.000.00251000
convAERefl.3.836.5620.6793.540.128
VAERefl.4.156.3221.5792.470.124
UNetRefl.2.295.7212.7580.460.138
Pix2PixRefl.2.144.7911.3030.670.138
DDPMRefl.2.747.9317.8580.383986.353
SD(w,CA)Refl.3.816.8219.7881.612961.027
SDRefl.2.535.2615.0455.272970.86
DDBMRefl.1.936.3818.3479.133732.21
", + "image_path": "42bfc572d48512a9a46443aad13f8ef05dc6fd99c2d2bddce6041606c71f34ed.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 550, + 293, + 668 + ], + "blocks": [ + { + "bbox": [ + 58, + 550, + 293, + 668 + ], + "lines": [ + { + "bbox": [ + 58, + 550, + 293, + 668 + ], + "spans": [ + { + "bbox": [ + 58, + 550, + 293, + 668 + ], + "type": "image", + "image_path": "92c5beb2928dbe44ce383c4da7fbc92cb7d8fbe07a70808739e69d48f2cebc24.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 711 + ], + "type": "text", + "content": "Figure 3. Qualitative results comparing the ground-truth simulation with the prediction for a single sample within the reflection task. Additional results can be found in Appendix B.7" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 546, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 715 + ], + "type": "text", + "content": "all tested architectures. For the Pix2Pix model this rise in NLoS error is moderate, while this effect is more pronounced in the UNet and Diffusion models. Models without skip connections, such as VAE and ConvAE, tend to blur most of the image and struggle to capture any reflection patterns accurately. This trend highlights how reflections and diffractions introduce complex higher-order dependencies (see Eq. (12) and Appendix B.4) that complicate the image-to-image sound propagation tasks. Although the models successfully replicate sound reflection patterns visually (see Fig. 3), there exists a significant gap between its visual outputs and the actual acoustic accuracy. This discrepancy suggests that while visually detailed, the models do not fully capture the complexities of sound physics." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "11128" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 151, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 151, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 151, + 83 + ], + "type": "text", + "content": "4. Lens Distortion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 91, + 295, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 91, + 295, + 210 + ], + "spans": [ + { + "bbox": [ + 54, + 91, + 295, + 210 + ], + "type": "text", + "content": "Our second benchmark is designed to investigate a class of physical problems for which solutions can be computed directly via closed equations, rather than iterative solvers. We find a representative and intuitive task in the modeling of optical lens distortions [7, 14], a common problem in computer vision. Generative models are trained to simulate the appearance of faces distorted by lens effects, using the Brown-Conrady [12] model. These distortions are computed using OpenCV [5], providing ground truth for training and evaluation of generative models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 211, + 295, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 211, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 54, + 211, + 295, + 317 + ], + "type": "text", + "content": "Dataset. We leverage a subset of the CelebA dataset [29], comprising 50,000 facial images, which were cropped to center the faces within a " + }, + { + "bbox": [ + 54, + 211, + 295, + 317 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 211, + 295, + 317 + ], + "type": "text", + "content": " pixel frame. The resulting images where split into two distinct groups to isolate the effects of tangential distortion: one focusing on horizontal distortion and the other on vertical distortion. This allows us to specifically measure the impact of each type of distortion on the accuracy of the generative models (see Fig. 4 and Appendix C for more details)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 438 + ], + "type": "text", + "content": "Physics of Lens Distortions. The Brown-Conrady model is a widely adopted method to describe and correct lens distortions. It provides a mathematical framework to represent both radial and tangential distortions through a series of coefficients. For our study, we concentrate exclusively on the tangential distortion components, utilizing relevant coefficients to model the distortion characteristics of specific lenses. The mathematical representation of tangential distortion is expressed through the following equation, where " + }, + { + "bbox": [ + 55, + 319, + 296, + 438 + ], + "type": "inline_equation", + "content": "x_{dist}" + }, + { + "bbox": [ + 55, + 319, + 296, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 296, + 438 + ], + "type": "inline_equation", + "content": "y_{dist}" + }, + { + "bbox": [ + 55, + 319, + 296, + 438 + ], + "type": "text", + "content": " denote the distorted coordinates:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 446, + 295, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 446, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 91, + 446, + 295, + 479 + ], + "type": "interline_equation", + "content": "x _ {d i s t} = x + \\left[ 2 p _ {1} x y + p _ {2} \\left(r ^ {2} + 2 x ^ {2}\\right) \\right], \\tag {3}", + "image_path": "f69d3151491549bf46d9b7ae90b96a2422c9191cf30edad5b4ee5211843436cb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 475, + 258, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 475, + 258, + 502 + ], + "spans": [ + { + "bbox": [ + 94, + 475, + 258, + 502 + ], + "type": "interline_equation", + "content": "y _ {d i s t} = y + \\left[ p _ {1} \\left(r ^ {2} + 2 y ^ {2}\\right) + 2 p _ {2} x y \\right],", + "image_path": "04c3d585c46e1ee9e96bce8b3ca403f5a64d2f61a0907bef59c5abb5577c0a46.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "inline_equation", + "content": "r^2 = x^2 + y^2" + }, + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "text", + "content": " represents the squared radial distance from the image center, and " + }, + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "text", + "content": " are the coefficients that model the tangential distortion effects. By strategically adjusting " + }, + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 55, + 510, + 296, + 583 + ], + "type": "text", + "content": " and isolating their effects via controlled nullification, we can identify how each parameter individually contributes to the image distortion:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 591, + 295, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 591, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 91, + 591, + 295, + 628 + ], + "type": "interline_equation", + "content": "x _ {d i s t} = \\left\\{ \\begin{array}{l l} x + 2 p _ {1} x y & \\text {i f} p _ {2} = 0, \\\\ x + p _ {2} \\left(r ^ {2} + 2 x ^ {2}\\right) & \\text {i f} p _ {1} = 0, \\end{array} \\right. \\tag {4}", + "image_path": "f2377df3472b965e37dd787d26c0f0a57b087062ef32986df6c7cac7662ca788.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 626, + 257, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 626, + 257, + 657 + ], + "spans": [ + { + "bbox": [ + 94, + 626, + 257, + 657 + ], + "type": "interline_equation", + "content": "y _ {d i s t} = \\left\\{ \\begin{array}{l l} y + p _ {1} (r ^ {2} + 2 y ^ {2}) & \\text {i f} p _ {2} = 0, \\\\ y + 2 p _ {2} x y & \\text {i f} p _ {1} = 0. \\end{array} \\right.", + "image_path": "6d56853ab48d9506354370152298357cda3f09929f243dc9b4fb75ebe595e520.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "These conditional equations illustrate the isolated impact of each coefficient, revealing that the correction complexity for " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "y_{dist}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " escalates with the term " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_1(r^2 + 2y^2)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is non-zero and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is zero. Conversely, the complexity of " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{dist}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "correction intensifies with " + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "inline_equation", + "content": "p_2(r^2 + 2x^2)" + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": " under the condition that " + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": " is non-zero and " + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": " is zero." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 96, + 554, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 554, + 335 + ], + "type": "text", + "content": "Evaluation. Given the challenges of accurately capturing lens distortion through traditional pixel-based metrics, our evaluation strategy is specifically designed to focus on accurately reproducing designated landmarks in the images. We detect facial landmarks using a 2D facial landmark detection method based on the Facial Alignment Network (FAN) [8], applied to the faces in the CelebA dataset (see Appendix C for details). We use these landmarks to assess the geometric distortions introduced by lens settings, represented by the coefficients " + }, + { + "bbox": [ + 313, + 96, + 554, + 335 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 96, + 554, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 96, + 554, + 335 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 313, + 96, + 554, + 335 + ], + "type": "text", + "content": ". We then compute the Euclidean distance between the actual position of a landmark in the true distorted image and its predicted position in the image generated by the models as a quality metric. This measurement is performed separately for the horizontal (X) and vertical (Y) coordinates, resulting in X Error and Y Error respectively. This approach not only allows us to gauge the accuracy of the models in simulating the specific effects of lens distortion on facial features but also helps in understanding how different distortion coefficients impact the positional accuracy of landmarks along each axis." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": "Analysis. Our experiments with the CelebA dataset under varying conditions of " + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": " showcase different patterns in error rates that align with the theoretical complexity of the distortion model. As shown in Tab. 2 and Fig. 5, the generative models exhibit differing accuracies in landmark prediction under the influence of lens distortion parameters " + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": ". Notably, the increased complexity in " + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "inline_equation", + "content": "x_{dist}" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "inline_equation", + "content": "y_{dist}" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": " equations under separate conditions (" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": " nonzero vs. " + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": " nonzero) slightly correlates with heightened error rates in respective dimensions. The models exhibit higher Y Error when only " + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": " is active, indicative of the complex correction challenge introduced by the higher order dependency on " + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": ". Similarly, a slight increase in X error and a slight decrease in Y error are observed when " + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 313, + 335, + 554, + 503 + ], + "type": "text", + "content": " is exclusively nonzero." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 506, + 553, + 619 + ], + "blocks": [ + { + "bbox": [ + 316, + 506, + 553, + 619 + ], + "lines": [ + { + "bbox": [ + 316, + 506, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 316, + 506, + 553, + 619 + ], + "type": "image", + "image_path": "cc25ea096a192917527fde18b87997a2ecf22ccdbb5c567c7e5ddc1765986527.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 627, + 555, + 715 + ], + "lines": [ + { + "bbox": [ + 313, + 627, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 555, + 715 + ], + "type": "text", + "content": "Figure 4. Sampling and evaluation pipeline for the lens distortion dataset. The Brown-Conrady distortion model generates the true distorted images based on parameters " + }, + { + "bbox": [ + 313, + 627, + 555, + 715 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 627, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 627, + 555, + 715 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 313, + 627, + 555, + 715 + ], + "type": "text", + "content": " (we depict a chess pattern for visualization). The conditioned generators are then trained to replicate these distortions for given images and parameters. The models are evaluated by comparing predicted against true facial landmarks, using a 2D facial landmark detection based on the Facial Alignment Network (FAN) [8]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "11129" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 282, + 223 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 282, + 223 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 282, + 223 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 282, + 223 + ], + "type": "image", + "image_path": "83fe1329682073524b4b8ea7e8f635861464b06e3fc097388d51e2dab1135714.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 231, + 296, + 286 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 286 + ], + "type": "text", + "content": "Figure 5. Qualitative visualization of lens distortion predictions using different generative models on a CelebA dataset sample. " + }, + { + "bbox": [ + 55, + 231, + 296, + 286 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 55, + 231, + 296, + 286 + ], + "type": "text", + "content": " distortion: Original image at top-left; subsequent panels show U-Net, Pix2Pix, Diffusion models. Red dots mark actual landmark positions and blue dots for predictions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 353, + 282, + 555 + ], + "blocks": [ + { + "bbox": [ + 55, + 300, + 295, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 346 + ], + "type": "text", + "content": "Table 2. Quantitative X-Y error and shift comparison across different models for varying activation of parameters " + }, + { + "bbox": [ + 55, + 300, + 295, + 346 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 300, + 295, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 300, + 295, + 346 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 55, + 300, + 295, + 346 + ], + "type": "text", + "content": " with a batch size of 16 during inference. Best overall results in bold, second best underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 353, + 282, + 555 + ], + "lines": [ + { + "bbox": [ + 70, + 353, + 282, + 555 + ], + "spans": [ + { + "bbox": [ + 70, + 353, + 282, + 555 + ], + "type": "table", + "html": "
ModelComb ↓X ↓ ErrY ↓ Err|X-Y| ShiftRuntime/ Sample (ms)
Sim.0.000.000.000.00153.205
convAE11.936.758.131.380.110
VAE11.536.557.831.280.122
UNet2.821.282.150.870.118
Pix2Pix2.000.991.430.440.122
DDPM1.930.941.390.453970.603
SD(w,CA)3.091.592.210.622991.678
SD2.791.412.010.602997.576
Sim.p1 = 0, p2 ≠ 0
convAE0.000.000.000.00153.205
10.568.354.772.210.110
VAE10.408.264.623.640.122
UNet2.361.331.600.270.117
Pix2Pix1.771.021.140.130.123
DDPM2.131.391.230.163970.603
SD(w,CA)2.851.601.940.342991.678
SD2.441.381.640.262997.576
", + "image_path": "077749b4ebfe3668979b4322e3609bd87492f1e9ed12bf98ca52b12a88153324.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 560, + 246, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 246, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 246, + 574 + ], + "type": "text", + "content": "5. Dynamics of bouncing movements" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 700 + ], + "type": "text", + "content": "The third physical problem under investigation is the movement of a rolling or bouncing ball. The aim here is to evaluate the ability of generative models to map kinematic movements of physics - a task which should generalize well to the investigation of various time series problems. Specifically, we define the task to predict the position and rotation of a ball along an inclined surface for a defined time interval after the input image. Only the bouncing case is discussed below; however, all results for the rolling ball can be found in Appendix D.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Dataset. We generated our data set with Pymunk [3], a" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 339, + 72, + 532, + 148 + ], + "blocks": [ + { + "bbox": [ + 339, + 72, + 532, + 148 + ], + "lines": [ + { + "bbox": [ + 339, + 72, + 532, + 148 + ], + "spans": [ + { + "bbox": [ + 339, + 72, + 532, + 148 + ], + "type": "image", + "image_path": "8dbd5c8c665714692e47243196b5096b0fd8654c58e8739a751a2c945931c44b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 255, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 300 + ], + "type": "text", + "content": "Figure 6. Physical equations describing the bouncing ball movement divided in 3 parts: free fall (green), impact & bounce of the ball on the ground (red dot) and oblique throw with initial velocity from ball bounce (orange)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 343, + 162, + 555, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 162, + 555, + 176 + ], + "spans": [ + { + "bbox": [ + 343, + 162, + 555, + 176 + ], + "type": "interline_equation", + "content": "F _ {g} = m * a \\Leftrightarrow m * \\ddot {y} = - m * g \\Leftrightarrow \\ddot {y} = - g \\tag {5}", + "image_path": "089831f97823e9f217615fa73f9ab862ba4f19c3bdda7a9ef3b094bc76bde7ff.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 354, + 188, + 555, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 188, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 354, + 188, + 555, + 228 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} m * \\ddot {y} + d * \\dot {y} - c * (r - y) = - m * g \\tag {6} \\\\ m * \\ddot {x} = m * g * \\sin (\\beta) \\text {a l o n g x a x i s} \\\\ \\end{array}", + "image_path": "826ce4b5336f14647255c81a123fd1a9fd812615cf0ab2d15c63a4618c10f0a0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 306, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 521 + ], + "type": "text", + "content": "Python physics simulation toolbox. For the bouncing case we use 50k pairs of training images (256x256px) and perform the evaluation of the generative networks with 1600 images. Three parameters on the input images are variable: the ground inclination, the start height of the ball movement and the ball position. Additionally, the time interval between the input and target images is also variable. All other physical properties of the problem (ball's mass or friction coefficients) are kept constant. A multi-coloured structure with randomly distributed lines is superimposed on the simulation images in order to obtain greater diversity in the training data and is also used to analyze consistency, e.g. how well the generative networks can map fine multicoloured patterns (see Fig. 7). The total time required to create a sample using the available physics simulation CPU code (with a customized background) is 3.8 seconds, while the inference times for the generative models remain consistent with those reported in Tab. 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "content": "Physics of bouncing balls. To model the physics behind the movement of a bouncing ball, the motion is divided into different sections which are represented with different colors in Fig. 6. First, the ball is released from a defined height without any velocity and falls down according to the Newton's law of the free fall (green part) (see equation 5). The impact of the ball on the ground and its rebound can be described in simplified terms by a spring-damper model. The aim of this second part of the movement (red dot) is to calculate the rebound speed and angle. This is done with equation 6, which contains the constant of the elements of the simplified model (see details in Appendix D.1) and the angle corresponds to the ground inclination for the first impact. Finally, the orange part of the movement is assumed to be an oblique throw with the initial velocity " + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "inline_equation", + "content": "v_{0}" + }, + { + "bbox": [ + 313, + 521, + 556, + 714 + ], + "type": "text", + "content": " calculated with equation 6. The equations of motion describing" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "11130" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 370, + 149 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 370, + 149 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 370, + 149 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 370, + 149 + ], + "type": "image", + "image_path": "294f33bcc5ed3449a6de618b8c53515dc21024d86629cf3012bb64ce2fd83e2c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 51, + 555, + 140 + ], + "lines": [ + { + "bbox": [ + 376, + 51, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 376, + 51, + 555, + 140 + ], + "type": "text", + "content": "Figure 7. The sampling and evaluation pipeline for the rolling and bouncing ball movements simulates ball trajectories. A physics-based model is used to predict the ball's movement. Evaluations focus on the accuracy of predictions related to key movement metrics, including bounce height and horizontal displacement (see Appendix D for details)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 170, + 294, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 170, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 54, + 170, + 294, + 241 + ], + "type": "text", + "content": "this oblique throw along the X and Y axes are also derived from the Newton's law (see equation 7). When the ball hits the ground for the second time, the principle \"angle of incidence = angle of reflection\" is used to determine the angle of rebound " + }, + { + "bbox": [ + 54, + 170, + 294, + 241 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 170, + 294, + 241 + ], + "type": "text", + "content": " and the next part of the motion is again defined by equations 6 and 7." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 242, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 242, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 57, + 242, + 295, + 456 + ], + "type": "text", + "content": "Evaluation. The generated images are examined according to 4 precise error criteria: the ball position split in X and Y directions, where the error is defined as the number of pixels between the theoretical and the predicted ball center. The second error measure is the ball rotation describing its angle in degrees during the movement due to the inclined ground surface. The next quality indicator of the algorithms is the out-of-roundness of the ball also measured in pixels by taking all identified points of the ball contour and calculating the difference between the theoretical ball radius of 15px and the predicted one. The standard deviation of all radius errors is taken as the error measure. Finally, the correct representation of the ground inclination is checked: the angular error between simulation and prediction is also measured in degrees. In addition, there are three \"general\" error criteria that give preciser indications of the correct representation of the physics which are explained with the complete results in the Appendix D." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 457, + 295, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 295, + 565 + ], + "type": "text", + "content": "The results of the seven examined generative algorithms are summarized in table 3 (only for the four most important error criteria). The last column of the table contains the average number of non-evaluable predicted images over the four error criteria. A detailed analysis of the advantages and disadvantages of the individual methods with all the evaluation parameters can be found in the Appendix D, as well as the evaluations obtained for individual runs for the rolling and bouncing case." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 579, + 296, + 658 + ], + "blocks": [ + { + "bbox": [ + 56, + 579, + 296, + 658 + ], + "lines": [ + { + "bbox": [ + 56, + 579, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 296, + 658 + ], + "type": "table", + "html": "
Position XPosition YRotationRoundnessError
convAE4.24 ± 3.96.08 ± 5.912.2 ± 8.61.06 ± 0.099%
VAE4.69 ± 6.16.25 ± 6.931.0 ± 400.90 ± 0.195%
UNet5.53 ± 7.510.8 ± 1215.2 ± 230.74 ± 0.228%
Pix2Pix6.28 ± 8.011.7 ± 1317.2 ± 210.56 ± 0.111%
DDPM7.91 ± 9.015.5 ± 1432.9 ± 340.61 ± 0.25.7%
SD(w.CA)40.0 ± 4924.8 ± 2361.1 ± 520.53 ± 0.27.3%
SD8.55 ± 1216.2 ± 1434.2 ± 380.47 ± 0.12%
", + "image_path": "72ce5141c374744b552daf989718985e67c873d583655e5d1434bfb10df33fc8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 296, + 711 + ], + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 711 + ], + "type": "text", + "content": "Table 3. Prediction results for the bouncing ball experiment in mean±std and mean percentage of non-evaluable predictions for the four criteria (best overall results in bold). Average runtime of the simulation: 3.8s, generative models: same as inTab. 1" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 170, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 337 + ], + "type": "text", + "content": "Overall, the GAN (Pix2Pix model) provides the most stable representation of the bouncing ball problem. It achieves the best results across all evaluation criteria, with the most reliable physics predictions and fewest failed samples. The UNet also represents the ball position and angle on the generated samples very well, but produces more error images that cannot be evaluated with mostly either no ball or only a shredded shape. The greatest weakness of this algorithm lies in the clean representation of the ball roundness. The auto-encoder approaches (VAE and convAE) produce very blurred images in which the ball is often distorted and hardly recognizable. Therefore, the mean error values are not very representative, as over " + }, + { + "bbox": [ + 313, + 170, + 553, + 337 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 170, + 553, + 337 + ], + "type": "text", + "content": " of the predictions cannot be analyzed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 338, + 553, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 553, + 505 + ], + "type": "text", + "content": "The denoising diffusion approach (DDPM) produces slightly poorer results in absolute terms. Especially the prediction of the ball rotation is worse, with an error approximately twice as high compared to the two other approaches. Finally, to test another type of diffusion, two variations of stable diffusion were compared. The one with cross-attention is generating more error pictures and show high error values. In contrast, the normal stable diffusion performs very well: the mapping of the position is almost as accurate as with the Pix2Pix approach and the representation of the ball roundness is the best. In addition, the network generates very few error images, so that it could almost be placed on the same level as the GAN but comes as second due to a poorer angle mapping." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 506, + 553, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 553, + 647 + ], + "type": "text", + "content": "As can be seen in the qualitative result images below (figure 8), GAN and diffusion (especially the SD) networks depict small details much more accurately. The background structure of the images with the different colors is completely blurred by the UNet but not by the GAN and the diffusion. The DDPM predictions are blurred because the images have been scaled up from " + }, + { + "bbox": [ + 313, + 506, + 553, + 647 + ], + "type": "inline_equation", + "content": "64\\mathrm{px}" + }, + { + "bbox": [ + 313, + 506, + 553, + 647 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 506, + 553, + 647 + ], + "type": "inline_equation", + "content": "256\\mathrm{px}" + }, + { + "bbox": [ + 313, + 506, + 553, + 647 + ], + "type": "text", + "content": ". If the output of the diffusion model is set to " + }, + { + "bbox": [ + 313, + 506, + 553, + 647 + ], + "type": "inline_equation", + "content": "256\\mathrm{px}" + }, + { + "bbox": [ + 313, + 506, + 553, + 647 + ], + "type": "text", + "content": ", the target image is sharp again but several or no balls and many more artefacts appear on the net predictions (see figure 25 in the Appendix). To summarize, the Pix2Pix and Stable Diffusion approaches also are visually performing the best." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 648, + 553, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 553, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 553, + 708 + ], + "type": "text", + "content": "Finally, slight correlations can be observed between the simulation's parameters and the generated errors. For example the ball position in the X direction is predicted worse for all nets when the time interval between input and prediction images becomes larger. More details on the correla" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 757 + ], + "type": "text", + "content": "11131" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 293, + 198 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 293, + 198 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 293, + 198 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 293, + 198 + ], + "type": "image", + "image_path": "528cde0b04af22a6d9af7c0f885c082acc046ffc275f10d212ea074d10668656.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 208, + 295, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 243 + ], + "type": "text", + "content": "Figure 8. Qualitative results for predictions of the bouncing ball problem. Initial State and True label image at top-left; subsequent panels show results for VAE, UNet, Pix2Pix and Diffusion models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 249, + 296, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 249, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 249, + 296, + 488 + ], + "type": "text", + "content": "tions of each method can be found in the Appendix D.2.1. Analysis. When comparing the result in table 3 (and tables 11, 13 in the Appendix) with the physical motion equations, it shows that the errors of the generative networks increase with higher order terms. The second-order terms, such as the movement along the X-axis and the ball angle represent the largest errors for the rolling scenario. In the bouncing case, a second-order derivative is also present to describe the ball movement in the Y-direction due to the ball's impact on the ground. Therefore the calculation rule along the Y-direction consists of zero, first and second derivative terms, making it more complex than the equation in the X-direction. This is reflected in the error values of all the investigated approaches: the Y-error clearly exceeds that in the X-direction, as indicated by the complexity of the motion equations, although the X-position and angular errors remain roughly the same between the rolling and bouncing cases. In general, it can be concluded that the errors of the analyzed approaches increase with the order and complexity of the physical formulation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 500, + 125, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 125, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 125, + 513 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "type": "text", + "content": "With PhysicsGen, we present a comprehensive benchmark for the estimation of complex physical relations by image-to-image generative models. The three different datasets contained in PhysicsGen cover a wide range of diverse simulation scenarios and hence allow a systematic evaluation of learning based image generation models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": "Key insights from the baseline-evaluation. Despite the diversity of the three given physical problems and the fundamentally different network types used the base-line evaluation experiments, we can report several consistent findings:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "- Due to the unavailability of optimized GPU code for the used simulations it is impossible to provide a theoretically sound runtime comparison. Since a highly parallel implementation of the simulation algorithms is highly nontrivial (if possible at all), our analysis focuses on the runtime analysis of the generative models. However, replac" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 321, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 553, + 133 + ], + "type": "text", + "content": "ing physical simulations with GPU powered generative models allows high computational speed-ups in practice, especially when replacing highly iterative solvers like in the urban sound propagation with Pix2Pix or UNets (where we report speed-up factors of up to 20k)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 133, + 556, + 431 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 133, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 555, + 228 + ], + "type": "text", + "content": "- In terms of physical correctness and consistency, our evaluation shows good results for simple 0th and 1st order problems, but also indicates that generative models mostly fail to predict more complex physical relations (which are typically formulated by higher order terms). This finding needs further investigation whether this is an effect of the principal theoretical limitations of current models or just a lack of sufficient training data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 556, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 556, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 556, + 431 + ], + "type": "text", + "content": "- We also observe distinct error pattern for different model types which represents how the different model types are dealing with complexity and uncertainty. Please refer to Appendix D.2.1 for examples in the following exemplary discussion for the bouncing ball problem (similar effects can be observed in Appendix B.7 for the sound propagation tasks as well): while GANs often produce physically possible outputs, they still can fail to generate the correct output for given inputs. This results in consistent scenes with a complete ball at the wrong position. UNets, VAE and convAE on the other hand, often fail to produce consistent scenes and tend to blur uncertain predictions, resulting in deformed balls and destructed backgrounds. Lastly, diffusion models often produce multiple candidate solutions (multiple balls in one scene) and tend to fail to estimate more complex relations like ball rotation. This phenomenon also needs further investigation." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 462, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 582 + ], + "type": "text", + "content": "Limitations. The provided run-time analysis is based on the comparison of practically available implementations and is not fair from a theoretical perspective as no efforts were made to optimize the simulation codes to GPUs. Also, even though we carefully selected the three proposed physical benchmark problems such that they are likely to transfer to a wide range of other simulation tasks, our evaluation provides only a small subset of possible applications of generative models in physical modeling. Hence, results from our benchmark might not transfer to all possible scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Broader Impact Statement. The simulation of physical systems forms the backbone for a wide range of scientific and engineering tasks. Hence, it is important to have benchmarks for the emerging field of neural enhanced simulations to compare and validate these novel approaches. This work provides such a benchmark for the specific, but widely applicable case, where physical simulations are used to map between 2D input and output structures (visualized as images). Our analysis shows that while current generative models are promising significant speedups, there are still many open questions regarding their physical correctness." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "11132" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 200, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 200, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 200, + 85 + ], + "type": "text", + "content": "Funding Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "The authors acknowledge the financial support by the German Federal Ministry of Education and Research (BMBF) in the program \"Forschung an Fachhochschulen in Kooperation mit Unternehmen (FH-Kooperativ)\" within the joint project \"KI-Bohrer\" under grant 13FH525KX1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 180, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 212 + ], + "type": "text", + "content": "[1] A. Argilaga. Fem-gan: A physics-supervised deep learning generative model for elastic porous materials. Materials, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 214, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 214, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 214, + 296, + 258 + ], + "type": "text", + "content": "[2] Peter W Battaglia, Jessica B Hamrick, and Joshua B Tenenbaum. Simulation as an engine of physical scene understanding. Proceedings of the National Academy of Sciences, 110 (45):18327-18332, 2013. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 258, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 258, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 295, + 280 + ], + "type": "text", + "content": "[3] Victor Blomqvist. Pymunk pythonic 2d physics library, 2007 - 2024. 6, 20" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 281, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 295, + 335 + ], + "type": "text", + "content": "[4] Erwan Bocher, Gwenael Guillaume, Judicael Picaut, Gwendall Petit, and Nicolas Fortin. Noise modelling: An open source gis based tool to produce environmental noise maps. ISPRS International Journal of Geo-Information, 8(3):130, 2019. 3, 4, 13, 15" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 296, + 357 + ], + "type": "text", + "content": "[5] G. Bradski. The OpenCV Library. Dr. Dobb's Journal of Software Tools, 2000. 5, 17" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 358, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 392 + ], + "type": "text", + "content": "[6] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 392, + 280, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 280, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 280, + 403 + ], + "type": "text", + "content": "[7] Duane Brown. Close-range camera calibration. 1971. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 403, + 296, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 457 + ], + "type": "text", + "content": "[8] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d and 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 459, + 296, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 296, + 513 + ], + "type": "text", + "content": "[9] Giuseppe Carleo, Ignacio Cirac, Kyle Cranmer, Laurent Daudet, Maria Schuld, Naftali Tishby, Leslie Vogt-Maranto, and Lenka Zdeborova. Machine learning and the physical sciences. Reviews of Modern Physics, 91(4):045002, 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 557 + ], + "type": "text", + "content": "[10] Yilun Chen, Qi Wu, Yujie Sui, Yizhou Wang, and Song-Chun Zhu. Physically grounded vision-language models for robotic manipulation. arXiv preprint arXiv:2309.02561, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "type": "text", + "content": "[11] Runmin Cong, Wenyu Yang, Chongyi Zhang, Weiand Li, Chunle Guo, and Sam Huang, Qingmingand Kwong. Pugan: Physical model-guided underwater image enhancement using gan with dual-discriminators. IEEE Transactions on Image Processing, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 646 + ], + "type": "text", + "content": "[12] A. E. Conrady. Decentred Lens-Systems. Monthly Notices of the Royal Astronomical Society, 79(5):384-390, 1919. 5, 17" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "text", + "content": "[13] Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep reinforcement learning for continuous control, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[14] Wolfgang Faig. Calibration of close-range photogrammetric systems: Mathematical formulation. Photogrammetric Engineering and Remote Sensing, 41, 1975. 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 139 + ], + "type": "text", + "content": "[16] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 140, + 554, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 258 + ], + "type": "text", + "content": "[17] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti, Liu, Henning Meyer, Yishu Miao, Derek Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 260, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 553, + 293 + ], + "type": "text", + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. CoRR, abs/2006.11239, 2020. 2, 4, 11, 16" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 293, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 316 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "text", + "content": "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks. CoRR, abs/1611.07004, 2016. 2, 4, 11, 16" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 349, + 554, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 554, + 457 + ], + "type": "text", + "content": "[21] Kelsey Jordahl, Joris Van den Bossche, Martin Fleischmann, Jacob Wasserman, James McBride, Jeffrey Gerard, Jeff Stratner, Matthew Perry, Adrian Garcia Badaracco, Carson Farmer, Geir Arne Hjelle, Alan D. Snow, Micah Cochran, Sean Gillies, Lucas Culbertson, Matt Bartos, Nick Eubank, maxalbert, Aleksey Bilogur, Sergio Rey, Christopher Ren, Dani Arribas-Bel, Leah Wasser, Levi John Wolf, Martin Journois, Joshua Wilson, Adam Greenhall, Chris Holdgraf, Filipe, and François Leblanc. geopandas/geopandas: v0.8.1, 2020.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 459, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 553, + 502 + ], + "type": "text", + "content": "[22] Anuj Karpatne, Imme Ebert-Uphoff, Sai Ravela, Hassan A. Babaie, and Vipin Kumar. Physics-guided neural networks (pgnn): An application in lake temperature modeling. arXiv preprint arXiv:2212.02500, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 503, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 554, + 557 + ], + "type": "text", + "content": "[23] S Kephalopoulos, M Paviotti, and F Anfosso-Ledée. Common noise assessment methods in europe (cnossos-eu). Technical Report EUR 25379 EN, Publications Office of the European Union, Luxembourg (Luxembourg), 2012. JRC72550. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 558, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 558, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 558, + 553, + 602 + ], + "type": "text", + "content": "[24] Byungsoo Kim, V. C. Azevedo, N. Thürey, Theodore Kim, M. Gross, and B. Solenthaler. Deep fluids: A generative network for parameterized fluid simulations. Computer Graphics Forum, 38, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 602, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 602, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 315, + 602, + 553, + 624 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 2, 4, 11, 16" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 315, + 625, + 553, + 658 + ], + "type": "text", + "content": "[26] Daniel Ladwig, Martin Spitznagel, Jan Vaillant, Klaus Dorer, and Janis Keuper. Ai-guided noise reduction for urban geothermal drilling. 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 658, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 658, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 315, + 658, + 553, + 680 + ], + "type": "text", + "content": "[27] Adam Lerer, Sam Gross, and Rob Fergus. Learning physical intuition of block towers by example, 2016. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 681, + 553, + 713 + ], + "type": "text", + "content": "[28] Wenbin Li, Seyedmajid Azimi, Ales Leonardis, and Mario Fritz. To fall or not to fall: A visual approach to physical stability prediction, 2016. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 317, + 756 + ], + "type": "text", + "content": "11133" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 478 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "text", + "content": "[29] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2, 4, 11, 16" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. CoRR, abs/1505.04597, 2015. 2, 4, 11, 16" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[32] Yuxuan Song, Jingjing Gong, Yanru Qu, Hao Zhou, Mingyue Zheng, Jingjing Liu, and Wei-Ying Ma. Unified generative modeling of 3d molecules via bayesian flow networks, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "text", + "content": "[33] Martin Spitznagel and Janis Keuper. Urban sound propagation: a benchmark for 1-step generative modeling of complex physical systems, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[34] Nils Thuerey, Philipp Holl, Maximilian Mueller, Patrick Schnell, Felix Trost, and Kiwon Um. Physics-based Deep Learning. WWW, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "text", + "content": "[35] Michael Vorländer. Auralization: Fundamentals of acoustics, modelling, simulation, algorithms and acoustic virtual reality. Springer Science & Business Media, 2007. 3, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "text", + "content": "[36] Yang Yu and Yongming Liu. Physics-guided generative adversarial network for probabilistic structural system identification. Expert Systems with Applications, 239:122339, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "type": "text", + "content": "[37] Yufeng Zhang, Wei Li, Yangang Gao, and Shihong Li. Phys-diff: Physics-guided human motion diffusion model. arXiv preprint arXiv:2106.03704, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "type": "text", + "content": "[38] Linqi Zhou, Aaron Lou, Samar Khanna, and Stefano Ermon. Denoising diffusion bridge models, 2023. 2, 4, 11, 16" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 434, + 294, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 478 + ], + "type": "text", + "content": "[39] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martin-Martin, Abhishek Joshi, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning, 2022. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 747, + 318, + 757 + ], + "type": "text", + "content": "11134" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/18874676-27a7-4b87-b788-f15decae2835_content_list.json b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/18874676-27a7-4b87-b788-f15decae2835_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c854d7f25c415ab0328b2ad78e7503e08a4523b4 --- /dev/null +++ b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/18874676-27a7-4b87-b788-f15decae2835_content_list.json @@ -0,0 +1,1307 @@ +[ + { + "type": "text", + "text": "PillarHist: A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram", + "text_level": 1, + "bbox": [ + 210, + 130, + 785, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sifan Zhou $^{1\\dagger*}$ Zhihang Yuan $^{2\\dagger}$ Dawei Yang $^{2\\dagger}$ Xing Hu $^{2}$ Jian Qian $^{3}$ Ziyu Zhao $^{1}$", + "bbox": [ + 127, + 200, + 867, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Automation, Southeast University $^{2}$ Houmo AI $^{3}$ Fudan University", + "bbox": [ + 320, + 238, + 676, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sifanjay, hahnyuan, zhaoziyu.950207}@gmail.com,{dawei.yang, xing.hu}@houmo.ai", + "bbox": [ + 148, + 277, + 836, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Real-time and high-performance 3D object detection plays a critical role in autonomous driving and robotics. Recent pillar-based 3D object detectors have gained significant attention due to their compact representation and low computational overhead, making them suitable for onboard deployment and quantization. However, existing pillar-based detectors still suffer from information loss along height dimension and large numerical distribution difference during pillar feature encoding (PFE), which severely limits their performance and quantization potential. To address above issue, we first unveil the importance of different input information during PFE and identify the height dimension as a key factor in enhancing 3D detection performance. Motivated by this observation, we propose a height-aware pillar feature encoder, called PillarHist. Specifically, PillarHist statistics the discrete distribution of points at different heights within one pillar with the information entropy guidance. This simple yet effective design greatly preserves the information along the height dimension while significantly reducing the computation overhead of the PFE. Meanwhile, PillarHist also constrains the arithmetic distribution of PFE input to a stable range, making it quantization-friendly. Notably, PillarHist operates exclusively within the PFE stage to enhance performance, enabling seamless integration into existing pillar-based methods without introducing complex operations. Extensive experiments show the effectiveness of PillarHist in terms of both efficiency and performance.", + "bbox": [ + 89, + 357, + 485, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 795, + 220, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LiDAR-based 3D object detection plays a pivotal role in various applications [5, 7, 9, 13, 29, 37, 50], particularly in autonomous driving and robotics. It is important to detect", + "bbox": [ + 89, + 820, + 483, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the objects in the surrounding environment fastly and accurately, which places a high demand for both performance and latency. Existing mainstream methods can be categorized into two types based on the point cloud representation: (1) point-based [8, 28, 31, 32, 42] and (2) grid-based models [2, 6, 14, 15, 21, 41, 41, 51]. Point-based models generally inherit the success in point cloud representation learning using deep neural networks[26, 44] and detect objects directly from raw points. However, these methods often face challenges in real-time onboard deployment as they often require point query/retrieval in 3D space (e.g., PointNet++ [27]). Grid-based methods rasterize the points into regular grid (voxel/pillar) and employ 2D or 3D convolution for feature learning. Compared to voxel-based methods, pillar-based detectors exhibit a more compact grid representation and can seamlessly integrate advanced 2D detection techniques without much modification. Therefore, those methods offer a favorable trade-off between performance and speed, making them highly suitable for practical deployment in autonomous systems. Furthermore, it is equally crucial to deploy them on resource-limited edge devices by improve their efficiency (e.g., reduce memory and computation cost).", + "bbox": [ + 511, + 328, + 906, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Pillar Feature Encoding (PFE) module plays a huge role in pillar-based networks [15, 17, 30, 45, 49]. Its main purpose is to transform the input point clouds, which are initially unordered (x, y, z, intensity) information, into pillars with a regular spatial arrangement. Each pillar has a fixed length of feature dimensions. In this paper, we observe that existing pillar-based methods suffer from two key problem in PFE module. 1 Information loss significantly hampers their performance. 2 Encoding scheme leads to a large variation of activations, making quantization challenging. As shown in Figure 1, the problem 1 can be attributed to two factors: (1) Lack of detailed height-aware feature encoding. The height information is not well handled, resulting in an incomplete representation of the 3D scene. It is", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done as an intern at Houmo AI. †Equal Contribution. ‡Corresponding author.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27336", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5686df97058c02679acbf77ed3a8da0e002a5b124a2e9e06889d5431749e19c.jpg", + "image_caption": [ + "(1) Lack of height-aware feature encoding" + ], + "image_footnote": [], + "bbox": [ + 181, + 109, + 813, + 204 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0a58559b08e8eb8e5efcd382426f84b8acce917b642f673fab80657b303da536.jpg", + "image_caption": [ + "(2) Information loss with max-pooling operator", + "Figure 1. The drawbacks of pillar feature encoding are as follows: (1) Compared with the voxel feature encoding, the encoding process of pillars lacks height awareness, as it does not explicitly model the height dimension of the input point cloud. (2) PFE employs element-wise max-pooling operation, which inevitably results in information loss. As illustrated in the right side of (2), the max-pooling operation tends to preserve the feature information of points with higher height values within a pillar while disregarding points with lower height values, resulting in the loss of geometric information within the pillar." + ], + "image_footnote": [], + "bbox": [ + 178, + 222, + 815, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "difficult to distinguish the height differences in various situations, like a pedestrian and a person standing on a car. (2) Max-pooling operator causes the loss of height information. The adoption of max-pooling operations within each pillar to derive pillar features leads to a loss of detailed information, which limits the ability of the encoded features to distinguish different objects.", + "bbox": [ + 89, + 426, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The problem 2 can be attributed the various arithmetic range of the PFE input. For instance, in nuScenes [1] dataset, the input of PFE includes $\\hat{p}_i = \\left\\{[x_i,y_i,z_i,r_i,x_i^m,y_i^m,z_i^m,x_i^c,y_i^c,z_i^c]\\in \\mathbb{R}^{N_v\\times 10}\\right\\}$ , where $[x_i,y_i,z_i]\\in [-51.2,51.2]$ is the point coordinates, $r_i\\in [0,255]$ is the intensity, $[x_i^m,y_i^m,z_i^m]\\in [-0.2,0.2]$ (pillar size $= 0.2\\mathrm{m}$ ) denotes the offset of $[x_i,y_i,z_i]$ and the mean value of $[x_i,y_i,z_i]$ within one pillar, and $[x_i^c,y_i^c,z_i^c]\\in [-4.0,4.0]$ is the spatial offset of $[x_i,y_i,z_i]$ and the pillar center. Details refer to Sec 3.1. The different part of input represents the global- and local-contextual information, but exhibit significant difference in numerical distribution. Therefore, the arithmetic disparity makes it more susceptible to the effects of quantization (e.g., clipping error).", + "bbox": [ + 89, + 535, + 482, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we propose a novel quantization-aware pillar feature encoder called PillarHist. It leverages a histogram technique to encode the height-aware information of each pillar into multiple bins, and the number of points falling within each bin is counted. The resulting histogram representation provides both semantic and geometric information from the height dimension. PillarHist incorporates the LiDAR reflection intensity of each point to obtain a weighted histogram of intensity, which adds distinguished height-aware information for object de", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tection. These histograms, along with the pillar center coordinates, are concatenated and passed through a linear layer to obtain a feature vector representing the pillar. Specially, PillarHist constrains the arithmetic distribution of PFE input to a stable range, making it friendly for further quantization. Besides, by operating linear projection at the pillar level instead of individual points, PillarHist significantly reduces the computation overhead, resulting in improved efficiency. It can be seamlessly integrated into existing pillar-based methods as a plugin module. Extensive experiments demonstrate the effectiveness of our proposed PillarHist in terms of both efficiency and performance. The contributions of our method can be summarised as follows:", + "bbox": [ + 511, + 426, + 906, + 623 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conduct a comprehensive analysis of existing Pillar Feature Encoding (PFE). Our findings indicate that they suffer from information loss in the height dimension, and exits large variation of activation for model quantization.", + "- We propose a novel PFE module called PillarHist, which can mitigate the height information loss and reduce the computation cost of PFE. It is noteworthy that we make it a plugin module to conveniently insert into existing pillar-based methods while seamlessly boosting performance.", + "- We explore the quantization-friendly properties of PillarHist. PillarHist could effectively eliminate the severe performance degradation after quantization caused by the raw PFE, bringing performance gain effortlessly.", + "- We conducted extensive experiments to evaluate the effectiveness of PillarHist. Our results show that it achieved an average improvement of approximately 1 point in different dataset without extra latency." + ], + "bbox": [ + 511, + 647, + 903, + 905 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27337", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 233, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Grid Feature Encoding. Grid feature encoding is crucial for grid-based 3D object detectors, and significantly impacts model accuracy[12, 22]. Various methods have been proposed to extract features from two grid representations: Voxel and Pillar. VoxelNet [51] is a pioneering work that proposes a novel voxel feature encoding (VFE) layer, influencing many subsequent studies [31, 36, 38, 39, 43, 48, 52, 53]. Pillars serve as special voxels without vertical division; PointPillars [15] utilizes PointNet [26] for features extraction and employs max-pooling to obtain pillar features, leading to its widespread adoption in later works [10, 16, 17, 20, 30, 45, 49]. TANet [20] introduces a Triple Attention (TA) module to capture spatial correlations, but its reliance on point-wise operation and max-pooling incurs high computational costs and limits pillar representations. PiFeNet [16] proposes a lighter Pillar Aware Attention (PAA) module using channel-wise and point-wise attention, yet still employs pooling, affecting inference speed. Recently, PillarNet [30] introduced a dynamic PFE module with less information loss, while Fast-Pillars [49] presented a MAPE module for finer-grained geometric preservation. However, these methods do not adequately account for height-aware information and primarily use element-wise max-pooling or similar pooling operations, leading to further information loss. This paper focuses on enhancing detailed height-aware information and mitigate the information loss caused by pooling operations.", + "bbox": [ + 93, + 114, + 483, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quantization for 3D Object Detection. With the wide application of 3D object detection in autonomous driving and robotics, some quantization methods [46, 47] are designed to improve inference speed for onboard applications. In LiDAR-based 3D detection, [33] find that directly using INT8 quantization for 2D CNN significantly degrades performance on PointPillars [15], with an even greater reduction when using an entropy calibrator. Recently, LiDAR-PTQ [50] unveils the cause of performance collapse and propose an effective post-training quantization (PTQ) method. However, advanced quantization techniques[18, 19, 23, 40] are still needed to optimize quantization parameter within LiDAR-PTQ. Our PillarHist aim to make the network with PillarHist are quantization-friendly.", + "bbox": [ + 89, + 529, + 483, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 747, + 225, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 772, + 220, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LiDAR-based 3D Object Detection. Here, we present the basic task definitions of LiDAR-based 3D object detection. Given a point set with $N$ points in the 3D space, which is defined as $\\mathbf{P} = \\{\\mathbf{p}_i = [x_i,y_i,z_i,r_i]^T\\in \\mathbb{R}^{N\\times 4}\\}$ , where $x_{i},y_{i},z_{i}$ denote the coordinate values of each point along the axes X, Y, Z, respectively, and $r_i$ is the LiDAR reflection intensity. Given a set of objects in the 3D scene $\\mathbf{B} = \\{\\mathbf{b}_j =$", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$[x_j, y_j, z_j, h_j, w_j, l_j, r_j, c_j]^T \\in \\mathbb{R}^{M \\times 8}$ , where $M$ is the total number of objects, $b_i$ is the $i$ -th object in the scene, $x_j, y_j, z_j$ is the object's center, $h_j, w_j, l_j$ is the object's size, $r_j$ is the object's heading angle and $c_j$ is the object's class. 3D object detection aims to predict the 3D bounding boxes $\\mathbf{B}$ from the point cloud $\\mathbf{P}$ in driving scenarios accurately.", + "bbox": [ + 511, + 89, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Cloud Pillarization. We aim to transform the point cloud $\\mathbf{P}$ into a regular 3D pillar representation. Assuming that $\\mathbf{P}$ is in the 3D space with the range of $L, W, H$ along the x, y, z axes, respectively, The point cloud $\\mathbf{P}$ is uniformly partitioned into a specific pillar grid of size $(\\delta_x, \\delta_y, \\mathrm{H})$ , we can compute the no-empty pillar index using:", + "bbox": [ + 511, + 191, + 906, + 282 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(x _ {i n t}, y _ {i n t}, z _ {i n t}\\right) = \\left(\\lfloor \\frac {x _ {i}}{\\delta_ {x}} \\rceil , \\lfloor \\frac {y _ {i}}{\\delta_ {y}} \\rceil , \\lfloor \\frac {H}{2} \\rceil\\right), i = 0 \\dots N - 1 \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 289, + 906, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lfloor \\cdot \\rfloor$ is the floor operation, $x_{int},y_{int},z_{int}$ is the 3D pillar index of $p_i$ after pillarization. Notably, point cloud $\\mathbf{P}$ is only pillarized along the XY plane without the height dimension. Let $V = \\{v_{i} = [x_{i},y_{i},z_{i},r_{i}]\\in \\mathbb{R}^{N_{v}\\times 4}\\} ,i\\in$ $\\{1,\\dots,N_v\\}$ , be a non-empty pillar contains $N_{v}$ points, $N_{v}$ is number of points in pillar $V$", + "bbox": [ + 511, + 328, + 906, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pillar Feature Encoding. Pillar Feature Encoding (PFE) plays a crucial role in pillar-based 3D detection methods[15, 30, 49], directly impacting the representation learning of BEV feature and final detection accuracy. Here, taking PointPillars[15] as an example, we illustrate the details of PFE. The points in each pillar are initially concatenated along the channel dimension to form $\\hat{p}_i = \\{[x_i,y_i,z_i,r_i,x_i^m,y_i^m,z_i^m,x_i^c,y_i^c,z_i^c]\\in \\mathbb{R}^{N_v\\times 10}\\}$ , where $[x_i,y_i,z_i]$ is the original point coordinates in the ego frame, $[x_i^m,y_i^m,z_i^m]$ denotes the offset of $p_i$ from the mean of the xyz coordinates of all points within the current pillar, and $[x_i^c,y_i^c,z_i^c]$ is the offset of $p_i$ from the current pillar center. Notably, for each pillar, a sampling or padding strategy is employed to ensure that all pillars contain the same number of points. The concatenated point-wise features $\\hat{p}_i$ within a non-empty pillar $V$ are mapped to the high-dimensional feature space using MLP layers. This process can be mathematically formulated as:", + "bbox": [ + 511, + 433, + 908, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} ^ {c} = m \\left(\\hat {p} _ {i}; w _ {m}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 714, + 903, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the function $m(\\cdot)$ denotes MLP, $w_{m}$ denotes the learnable weights of $m(\\cdot)$ , and $p_i^c \\in \\mathbb{R}^{N_v \\times D}$ is the pointwise feature. 3 An element-wise max-pooling operation is applied along the dimension of point features, which is formulated as:", + "bbox": [ + 511, + 738, + 905, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {m} = \\max \\left(p _ {i} ^ {c}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 811, + 903, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\max(\\cdot)$ means the max-pooling operator, and $f^m \\in \\mathbb{R}^D$ is the resulting feature vector to represent non-empty pillar $V$ . A non-empty pillar $V \\in \\mathbb{R}^{N_v \\times 4}$ consisting of $N_v$ points is encoded into a feature vector $f^m \\in \\mathbb{R}^D$ .", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27338", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/3478325fcc75040ddcd29a5d5934b5aca1cfa1743e0c7ff2f5abb81ae5ed28b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(xi, xi^m, xi^c)(yi, yi^m, yi^c)(zi, zi^m, zi^c)3D_{R40}3D_{R11}
mAPEasyMod.HardmAPEasyMod.Hard
✓57.5067.9853.3051.2458.2067.9453.8252.85
✓59.4069.8555.4552.8959.9669.4156.1754.31
✓79.6386.6877.6274.5979.1586.6776.4974.28
PointPillars79.9887.0877.9074.9779.4386.2576.7675.02
", + "bbox": [ + 169, + 88, + 821, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Ablation results of different input for PFE in PointPillars on the KITTI val set in ${\\mathrm{{AP}}}_{3D}$ (R40) and ${\\mathrm{{AP}}}_{3D}$ (R11) for car.", + "bbox": [ + 120, + 193, + 874, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. What Kind of Information Do We Need in PFE?", + "text_level": 1, + "bbox": [ + 89, + 224, + 480, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(1) Height Information is More Important. As aforementioned, the pioneering pillar-based method PointPillars[15] concatenates two extra relation information along the channel dimension of the raw point cloud, i.e. $\\left([x_i^m, y_i^m, z_i^m]\\right)$ and $\\left([x_i^c, y_i^c, z_i^c]\\right)$ . Here, we decouple the input information along the $x$ , $y$ , and $z$ axes, respectively. aiming to explore their impact on the final 3D detection performance.", + "bbox": [ + 89, + 248, + 482, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Table 1, when we only input the information along the $x$ axis $(x_{i}, x_{i}^{m}, x_{i}^{c})$ or the information along the $y$ axis $(y_{i}, y_{i}^{m}, y_{i}^{c})$ , there is a significant degradation in the final 3D detection performance. The former results in a large decrease of $22.48 \\mathrm{~mAP}$ of $3D_{R40}$ and $21.23 \\mathrm{~mAP}$ of $3D_{R11}$ , while the latter leads to a large decrease of $20.58 \\mathrm{~mAP}$ of $3D_{R40}$ and $19.47 \\mathrm{~mAP}$ of $3D_{R11}$ . However, when we only input the information along the $z$ axis $(z_{i}, z_{i}^{m}, z_{i}^{c})$ , it also achieves comparable performance with the baseline. This indicates that in the pillar feature encoding process, the information along the $z$ axis, a.k.a the height dimension, plays a key role in the final 3D detection performance.", + "bbox": [ + 89, + 354, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2183e81d51230276621f5f7cda7cdcfcccd9335f3ef2b6f51e7c1da76449a5ce.jpg", + "image_caption": [ + "Figure 2. Visualization of the weights on (a) vehicle and (b) cyclist using Max-pooling in PFE. The size of points means scores. The point will be paid more attention if it has a higher score. Compared with original object's points, max-pooling operator in PFE inevitably result in the loss of fine-grained information the object and local geometries. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 94, + 542, + 480, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(2) Pillar Information Loss by Max-pooling. As mentioned in Section 1, current pillar-based methods [15, 30, 45, 49] heavily rely on max pooling to aggregate point features within each pillar. However, this max-pooling operation (Equation 3) inevitably leads to the loss of detailed information, particularly the local geometric patterns that play an important role in object recognition. To illustrate this point, we have visualized the effect of the max-pooling operation in Figure 2. In the figure, we represent the points that are more likely to be selected by max pooling as larger and red spheres. It is evident from the visualization that a", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "significant number of points are not selected after the max-pooling operation, while only a small fraction of points have a high probability of being chosen. This comparison clearly shows that the max-pooling operator results in a substantial loss of object geometry information compared to the original points of the object.", + "bbox": [ + 511, + 226, + 903, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Height-aware Discrete Histogram for PFE", + "text_level": 1, + "bbox": [ + 511, + 323, + 872, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, the challenge lies in finding a pillar encoding method that addresses these problems. Our goal is to develop a pillar feature encoding approach that possesses the following two key characteristics:", + "bbox": [ + 511, + 339, + 903, + 398 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "It should eliminate the requirement for sampling operations while still maintaining the ability to capture the height-aware of points within each pillar. This means that the encoding method should effectively handle the variation in point density across different height levels within a pillar.", + "It should keep the semantic and geometric information within the pillar, without resorting to max pooling or similar pooling operations. This is important to ensure that the encoded representation retains the relevant details and characteristics of the original pillar." + ], + "bbox": [ + 511, + 400, + 903, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve these objectives, we propose leveraging the histogram distribution. We can effectively capture the distribution of points within a pillar at different height levels. This allows us to maintain the height-awareness of the points without the need for sampling operations. The implementation of the histogram operation is straightforward. Initially, we divide the height range of the pillar into multiple bins, where each bin corresponds to a specific range of heights. For instance, if the pillar's height spans from 0 to 10 meters, we can partition it into several bins such as $0 - 1\\mathrm{m}$ , $1 - 2\\mathrm{m}$ , $2 - 3\\mathrm{m}$ , and so forth. Next, we proceed to count the number of points that fall within each bin for a given pillar. This counting process yields a count value for each bin, indicating the number of points residing within that particular height range. For instance, if a larger number of points are found in the lower height range bins, it signifies a higher density of points near the bottom of the pillar. Conversely, if there is a greater concentration of points in the higher height range bins, it suggests a higher density of points towards the top of the pillar. This approach provides the model with both the semantic and geometric information derived from the height dimension, thereby enhancing its understanding and representation of the point cloud data.", + "bbox": [ + 511, + 553, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27339", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3ed8bbfcf0f259111549c49cb22e3b931b993dd07c316d63c3b53e0764381a6f.jpg", + "image_caption": [ + "Figure 3. The overall architecture of our proposed PillarHist. The PFE (a) in PointPillars first extracts point-wise feature, and then the point-wise features are max-pooled along the point dimension to obtain the pillar-wise representation. This process has three key drawbacks: (1) it lacks detailed height-aware information, (2) the MLP operates on point-wise level, incurring computational burden, (3) the max-pooling operation leads to information loss. In contrast, our proposed PillarHist (b) directly encodes the height-aware information based on histogram, and then perform the MLP operation on the pillar level rather than point level. Finally, the max-pooling operation is also eliminated in proposed PillarHist." + ], + "image_footnote": [], + "bbox": [ + 186, + 99, + 807, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By using the histogram technique, we propose a height-aware pillar feature encoder named PillarHist. The algorithm is summarized in Algorithm 1: (1) We compute the histogram of each pillar, denoted as $H^p \\in \\mathbb{R}^B$ , where $B$ is the number of bins. (2) In addition, we incorporates the LiDAR reflection intensity of each point, $r_i$ , to obtain the weighted histogram of intensity, denoted as $H^i \\in \\mathbb{R}^B$ . The weighted histogram represents the mean LiDAR reflection intensity for each bin, providing valuable height-aware information for object detection. (3) To provide the global position of each pillar, we utilize the x and y coordinates of the pillar center, $x_{center}$ and $y_{center}$ . Since the coordinates of the pillar are constant, this step does not require any additional computational effort. (4) We concatenate the computed histograms and the pillar center coordinates, resulting in $[H^p | H^i | x_{center} | y_{center}] \\in \\mathbb{R}^{2B + 2}$ . (5) To represent the features of the pillars, we utilize a linear layer $f_\\theta$ that projects the combined information into a feature vector. $\\theta \\in \\mathbb{R}^{D \\times (2B + 2)}$ and $D$ is the number of output channels.", + "bbox": [ + 91, + 561, + 483, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the pillar center $x_{center}$ and $y_{center}$ provide the relative spatial relationships between different points within the point cloud. This spatial location information", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the pillars reflects the relative position of the targets in the scene, which can help the model better understand environmental context. For instance, if there are two pillars with the same point cloud distribution located at [0,10] and [10,0] in space, without the pillar coordinate to indicate the spatial relationship, these two pillars would be indistinguishable. This would pose significant challenges for the detector's object perception capabilities. Therefore, combining pillar center $x_{center}$ and $y_{center}$ with other features such as height and intensity can further enhance the model's perception capability. This is necessary to compensate for the missing spatial position information inherent in the raw point cloud data. By incorporating the pillar center coordinates, the model can better capture the geometric structure and spatial relationships between different parts of the scene. This is a crucial aspect for enabling accurate object detection in point cloud data.", + "bbox": [ + 511, + 561, + 908, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig 3, our PillarHist approach explicitly captures the histogram of points in different height intervals. This allows us to keep the semantic and geometric information within the pillar. PillarHist operates only during the PFE stage without introducing complex operations", + "bbox": [ + 511, + 824, + 910, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27340", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1: PillarHist Algorithm", + "text_level": 1, + "bbox": [ + 101, + 94, + 334, + 109 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input: Raw point data in a pillar $V\\in \\mathbb{R}^{N_v\\times 4}$ consisting of $N_{v}$ points, number of height bins $B$ ,number of output channels $D$ ,height of the pillar $H$ , linear layer $f_{\\theta}$", + "bbox": [ + 109, + 112, + 459, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Output: Pillar's feature vector $Y \\in \\mathbb{R}^{D}$ to represent the non-empty pillar $V$ .", + "bbox": [ + 109, + 172, + 459, + 204 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Divide the height $H$ of the pillar into $B$ bins, each bin having height $h$ , where $h = \\frac{H}{B}$", + "2: Initialize $H^p \\in \\mathbb{R}^B$ to zero, where $H^p$ represents the histogram of points in each bin", + "3: Initialize $H^i\\in \\mathbb{R}^B$ to zero, where $H^{i}$ represents the weighted histogram of intensity in each bin", + "4: for $p_i = \\{p_i | i = 1, 2, \\dots, N_v\\}$ do" + ], + "bbox": [ + 119, + 205, + 495, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5: Compute the corresponding bin index $b_{i}$ based on the $z$ coordinate $z_{i}$ of point $p_{i}$ : $b_{i} = \\lfloor \\frac{z_{i}}{h} \\rfloor$", + "bbox": [ + 119, + 309, + 455, + 354 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "6: $H^{p}[b_{i}]\\gets H^{p}[b_{i}] + 1$", + "7: $H^{i}[b_{i}]\\gets H^{i}[b_{i}] + r_{i},\\quad \\triangleright r_{i}$ is the LiDAR" + ], + "bbox": [ + 143, + 354, + 457, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "8: reflection intensity of $p_i$", + "bbox": [ + 120, + 385, + 321, + 398 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "9: $H^{i}[b_{i}]\\coloneqq H^{i}[b_{i}] / H^{p}[b_{i}]$ where $H^{i}[b_{i}]! = 0$", + "10: Compute the pillar center coordinates $x_{center}, y_{center}$", + "11: Concatenate: $[H^p |H^i |x_{center}|y_{center}]\\in \\mathbb{R}^{2B + 2}$", + "12: Linear projection: $Y = f_{\\theta}([H^{p}|H^{i}|x_{center}|y_{center}])$" + ], + "bbox": [ + 116, + 398, + 491, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "or interaction with other modules, allowing for seamless inserted into existing pillar-based methods serves as a plugin module with performance improvements.", + "bbox": [ + 89, + 496, + 482, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Efficiency of PillarHist: Here, we further analyze the efficiency of PillarHist. (1). The linear projection in PillarHist operates on the pillar level, not on individual points. In contrast, the previous approach [15, 20, 30, 45] performed the linear projection for each point within a pillar. By employing the linear projection at the pillar level, we significantly decrease the number of floating-point operations (FLOPs) required by the PFE, resulting in improved efficiency (Table 5). (2). PillarHist also avoids various memory-intensive operations, such as copying the original point cloud data, calculating distances $(x^{m},y^{m},z^{m},x^{c},y^{c},z^{c})$ , and performing max pooling. This further enhances the efficiency of PillarHist in real-world deployment.", + "bbox": [ + 89, + 542, + 482, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 768, + 223, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and Training Details. We evaluate our PillarHist on KITTI [11], nuScenes [1] and Waymo dataset [34]. We report the official metrics. We adopt OpenPCDet [35] as the training framework and faithfully reproduced the different baseline methods based on the provided official setting on kitti and nuScenes dataset in the OpenPCDet [35]. we directly follow the setting of the four pillar-based 3D de", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tectors. (i.e., PointPillars [15], CenterPoint-Pillar [45], PillarNet [30], FastPillars [49]). In nuScenes [1] dataset, we uniformly sample 64 frames point cloud data from the validation set as the calibration set. Calibration is a standard step in quantization, where a small subset of dataset is used to determine activation quantization parameters (scale factors and zero points). This is a common practice in model quantization [24]. The calibration set proportions is $1.06\\%$ (64/6019). We set the first and the last layer of the network to keep unquantized. Specifically, the quantization methods is widely used round-to-nearest (RTN) method, a common setting in prior works, such as PD-Quant [19] and LiDAPTQ [50]. We test all experiments on a single Nvidia Tesla V100 GPU. Please refer to Appendix for more quantization details and experiment analysis.", + "bbox": [ + 511, + 90, + 906, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Overall Results", + "text_level": 1, + "bbox": [ + 511, + 330, + 668, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on nuScenes test set. The quantitative results on nuScenes test set are shown in Table 2. It can be seen that by integrating our proposed PillarHist into the current representative pillar-based detector PoinPillar [15], CenterPoint-Pillar [45], PillarNet [30], the proposed PH-PillarHist achieves significant performance gains on nuScenes test set across different pillar-based methods, respectively. Meanwhile, PH-PillarHist brings 6ms, 9ms, 5ms and 4ms reduction on inference latency, respectively. Notably, our method focuses on the pillar feature encoding module, rather than designing a state-of-the-art pillar-based detector. The proposed PillarHist is a plugin module to conveniently insert into existing mainstream pillar-based methods, seamlessly bringing performance gains and latency reduction. This presents the effectiveness and efficiency of PillarHist in not only boosting the detection accuracy, but also improving the computational efficiency, making it practical for real-world application.", + "bbox": [ + 511, + 354, + 906, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on nuScenes val set. Here, we evaluate our PillarHist on different pillar-based baselines in nuScenes val set. We selected three representative pillar-based detectors: PointPillars [15], CP-Pillar [45], and PillarNet [30]. In detail, we replace the raw PFE module in each baseline model with our PillarHist, resulting in PH-PointPillars, PH-CP-Pillar, and PH-PillarNet, respectively. As shown in Table 3, the results demonstrate the effectiveness of proposed PillarHist, as we observe performance improvements of 1.7 NDS, 1.4 NDS, and 1.6 NDS on the three different baselines, respectively. These results highlight the effectiveness and broad applicability of our PillarHist in enhancing the represent power of pillar-based 3D object detectors, leading to consistent performance gains across different baseline architectures. Notably, On the nuScenes dataset, the nuScenes detection score (NDS) is the primary evaluation metric.", + "bbox": [ + 511, + 627, + 908, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Quantization Effectiveness of PillarHist. Furthermore, we also conduct quantization experiments to evaluate", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27341", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/58f809cc00b7a560bca7a47a9fee6ede767c155ad64da3b188e9d2c70fb629fd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNDSmAPLatencyCarTruckBusTrailerC.V.Ped.MotBycT.C.Bar
Focals Conv[3]70.063.8138 ms86.756.367.759.523.887.564.536.381.474.1
VoxelNeXt[4]70.064.566 ms84.653.064.755.828.785.873.245.779.074.6
FG-PFE[25]71.865.786 ms85.255.561.661.829.286.572.548.682.573.8
FastPillars [49]71.866.856 ms87.358.066.062.334.587.470.347.981.572.8
PointPillars [15]45.330.531 ms68.423.028.223.44.159.727.41.130.838.9
PH-PointPillars47.032.425 ms69.424.230.024.89.660.928.15.231.540.1
CenterPoint [45]65.558.096 ms84.651.060.253.217.583.453.728.776.770.9
PH-CenterPoint66.659.287 ms85.252.561.154.219.083.954.830.179.271.5
\\( PillarNet-18[30]^\\dagger \\)70.865.078 ms87.456.760.961.830.487.267.440.382.176.0
\\( PH-PillarNet-18^\\dagger \\)71.565.273 ms87.657.060.862.030.687.368.140.381.976.2
\\( PillarNet-34[30]^\\dagger \\)71.466.096 ms87.657.563.663.127.987.370.142.383.377.2
\\( PH-PillarNet-34^\\dagger \\)72.066.392 ms87.557.664.063.328.487.170.543.083.677.8
", + "bbox": [ + 91, + 88, + 903, + 297 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a1db4b57357cbcecec55e40a3daa4033d45f17052572559931caa7f05543806.jpg", + "table_caption": [ + "Table 2. Performance of 3D object detection methods on nuScenes test set. ${}^{ \\dagger }$ means the method that uses double-flip testing. All models listed take LIDAR data as input without image fusion or any model ensemble." + ], + "table_footnote": [], + "table_body": "
ModelsBits (W/A)NDSmAP
PointPillars [15]32/3252.137.4
8/844.8 (-7.3)29.8 (-7.6)
PH-PointPillars32/3253.8 (+1.7)38.8 (+1.4)
8/852.2 (-1.6)37.0 (-1.8)
CP-Pillar[45]32/3256.444.9
8/848.6 (-7.8)36.2 (-8.7)
PH-CP-Pillar32/3257.8 (+1.4)45.8 (+0.9)
8/856.3 (-1.5)44.3 (-1.5)
PillarNet[30]32/3261.452.6
8/854.6 (-6.8)43.1 (-9.5)
PH-PillarNet32/3263.0 (+1.6)52.8 (+0.2)
8/861.6 (-1.4)51.4 (-1.3)
", + "bbox": [ + 101, + 337, + 468, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the quantization-friendly ability. Specifically, we convert the weights and activations of the models from floating-point 32-bit to integer 8-bit and applied a simple posttraining quantization setting. As shown in Table 3, we find that the PointPillars, CP-Pillar, and PillarNet models suffered significant performance degradation after weight and activation quantization, with drops of 7.3 NDS and $7.6\\mathrm{mAP}$ for PointPillars, 7.8 NDS and $8.7\\mathrm{mAP}$ for CP-Pillar, and 6.8 NDS and $9.5\\mathrm{mAP}$ for PillarNet. In contrast, the PH-PointPillars, PH-CP-Pillar, and PH-PillarNet models that integrate PillarHist only perform performance drops of 1.6 NDS, 1.5 NDS, and 1.4 NDS, respectively. Notably, the quantized models with PillarHist integration achieved comparable performance to original floating-point counterparts without PillarHist (52.2 v.s. 52.1 NDS on PointPillars, 56.3 v.s. 56.4 NDS on CenterPoint-Pillars and 61.6 v.s. 61.4 NDS on PillarNet). Specifically, PillarHist could effectively eliminate the severe performance degradation after quantization caused by other PFE design, bringing quantization performance gain seamlessly. These results demonstrate the quantization-friendly nature of PillarHist, which maintains", + "bbox": [ + 88, + 583, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the model's representation power and performance even under quantization compression. This property is particularly valuable and practicable for deploying high-performance 3D object detectors on resource-constrained edge devices, where accuracy and latency are critical factors.", + "bbox": [ + 511, + 340, + 906, + 417 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/748000c08b08bb39faa7fcf49fa71b53683c43b58a7dc6f9d1bda0d464b922dc.jpg", + "table_caption": [ + "Table 3. Generalization and quantization performance comparison on nuScene val set." + ], + "table_footnote": [], + "table_body": "
MethodsMean (L2) mAPHVehicle (L2) mAPmAPHPed. (L2) mAPmAPHCyc. (L2) mAPmAPH
CP-Pillar55.962.562.061.851.556.154.3
PH-CP-Pillar57.363.262.863.854.057.055.3
↑+1.4+0.7+0.8+2.0+2.5+0.9+1.0
", + "bbox": [ + 514, + 425, + 903, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization of PillarHist on Waymo val set. Here, we also evaluate our PillarHist on Waymo dataset. We select CenterPoint-Pillars as our baseline and only replaced their pillar feature encoding module with our PillarHist for comparison, as show in table our method can bring $1.2\\mathrm{mAP}$ L2 and $1.5\\mathrm{mAPH}$ L2 performance improvements. Notably, it can achieve $2.0\\mathrm{mAP}$ L2 and $2.5\\mathrm{mAPH}$ L2 performance gains in the pedestrian category. This result demonstrates the effectiveness of our method.", + "bbox": [ + 511, + 529, + 908, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 665, + 666, + 681 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ae92e16005866d65758a738f7fee3ed8878fec5385526a8ffae53f4cf271e381.jpg", + "table_caption": [ + "Table 4. Apply PillarHist to CeterPoint-Pillars on Waymo val set trained with $20\\%$ training data." + ], + "table_footnote": [], + "table_body": "
PFE moduleCar 3D R40Ped 3D R40PFE GFLOPs
mAPEasyMod.HardmAPEasyMod.Hard
PFE[15]79.9887.0877.9074.9749.4154.7149.0144.520.152
PAA[16]81.0488.6279.2475.2748.5654.2647.9343.493.391
TA[20]80.7187.3379.1175.6949.4554.0049.3145.057.061
DynPFE*[30]79.9687.6377.7074.5549.3654.2749.0144.810.013
DynPFE[45]80.4687.8678.1675.3847.7752.2847.5343.520.112
FG-PFE[25]----50.7556.8349.8845.561.136
MAPE[49]81.1989.3978.5775.6348.1852.6848.0243.830.165
PillarHist81.4288.8079.1376.3051.0757.4550.4245.360.065
", + "bbox": [ + 517, + 689, + 903, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Performance of different PFE on the KITTI val set.", + "bbox": [ + 529, + 804, + 887, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of Different PFE modules. We compare the computation efficiency and performance of existing PFE modules. Specifically, we evaluate the following PFE modules: PFE in PointPillars [15], DynPFE in CenterPoint-Pillar [45], TA module in TANet [20], PAA module in", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27342", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PiFeNet [16], DynPFE* in PillarNet [30], and MAPE in FastPillars [49]. To ensure a fair comparison, we select PointPillars as the baseline method and conduct experiments by replacing the PFE in PointPillars with other PFE module (DynPFE, TA, PAA, DynpFE*, MAPE and our PillarHist). This setting eliminates the performance difference caused by different network design. As shown in Table 5, our PillarHist achieves the highest 3D AP compared to other PFE methods while maintaining superior computational efficiency. Notably, although DynPFE* used in PillarNet [30] has lower GFLOPs, it suffers from poor detection performance. Notably, our PillarHist method performs linear projection at the pillar-wise level rather than point-wise level in other PFE. This design can efficiently reduce the computational cost while maintaining the necessary structural information within pillar without max-pooling operations.", + "bbox": [ + 89, + 90, + 483, + 333 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/288099c089d9c8260e669a7d07a018374db91de289f89c70fde93d08f59e737f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
heightintensitycoordsCar 3D40Ped 3D40
mAPEasyMod.HardmAPEasyMod.Hard
✓80.1887.4778.1875.2248.0653.7547.5142.92
✓✓80.8188.2978.6275.5349.6255.4249.1044.36
✓✓✓81.4188.8079.1376.3051.0757.4550.4245.36
", + "bbox": [ + 91, + 345, + 480, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation results of different input for PillarHist on the KITTI val set for car and pedestrian.", + "bbox": [ + 89, + 422, + 482, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Effects of Different Input in PillarHist. Here, we conduct ablation study to analyze the impact of different input. As shown in Table 6, when we incorporate the intensity information, it leads to performance improvements for both the vehicle and pedestrian categories. Particularly for the pedestrian, approximately 1.7, 1.6, and 1.6 points for the easy, moderate, and hard levels, respectively. This notable improvement indicates that for small objects like pedestrians, the intensity values are crucial in distinguishing them from the background and enhancing their distinctive features. Furthermore, by adding the pillar coordinates information, there is a performance improvement of 0.6 and 1.5 points for the vehicle and pedestrian. We believe this improvement is because explicitly incorporating the global coordinate of the pillars helps the model better identify the object positions with the enhanced encoding information that includes height awareness. Notably, after adding the pillar coordinates, there is an improvement of 2.03 points in easy AP for pedestrians. This can be attributed that the pillars encode not only varying height distribution but also including the difference between empty and non-empty pillars.", + "bbox": [ + 89, + 462, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Map Visualization of PillarHist. Here, as shown in Figure 4, we apply our PillarHist method to PointPillars [15] (a) and compare the features from the max-pooling operation in their PFE. We find that the max pooling operation tends to focus more on the points at higher positions within the same pillar in BEV (with larger coordinates), resulting in a biased feature representation. In contrast, our PillarHist can focus more evenly on each point within", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e1b6e0bce78164ffc54c7e0885a4ea68909a150b5dcf8091594c6a283ed8b27.jpg", + "image_caption": [ + "Figure 4. (a) Visualization of the learned attention scores in PillarHist on different scenes. The color of points represents their scores. The point will be paid more attention if it has a higher score. (b) Visualization of the learned BEV feature in PointPillars and PH-PointPillars. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 516, + 82, + 903, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the pillar, encoding the height information into an unbiased histogram distribution, thereby obtaining richer geometric-semantic feature representation. This difference is particularly obvious at the car door location, the points within the pillar is typically uniformly distributed along height; however, max pooling emphasizes the higher contours, whereas PillarHist provides a better representation of the complete geometry of the car door Besides, as shown in Figure 4 (b), we also visualize the BEV features maps. Our method can obtain a more informative BEV feature representation compared to the PFE in PointPillars [15]. This also demonstrates the underlying reason for the better performance of our method.", + "bbox": [ + 511, + 375, + 908, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 585, + 633, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we addressed the issue of information loss and large numerical distribution difference in existing pillar-based 3D object detectors during the PFE stage. We identify the height dimension as a crucial factor for enhancing 3D detection performance and propose a height-aware pillar feature encoder called PillarHist. By leveraging histogram, PillarHist preserve the information along the height dimension while effectively mitigating the information loss introduced by max-pooling operations. Importantly, PillarHist operated exclusively within the PFE stage, allowing seamless insert into existing pillar-based methods serves as a plug-in module. Meanwhile, PillarHist also constrains the arithmetic distribution of PFE input to a stable range, making it quantization-friendly. Extensive experiments demonstrated the effectiveness of PillarHist in terms of both efficiency and performance. Our proposed method achieved an average improvement of approximately $1.0\\%$ NDS across various baselines, showcasing its ability to enhance 3D object detection accuracy and computation efficiency.", + "bbox": [ + 511, + 609, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27343", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank anonymous reviewers for their kind help of this work. This work was supported by National Natural Science Foundation of China (No.62271143), Frontier Technologies RD Program of Jiangsu (No. BF2024060) and the Big Data Computing Center of Southeast University.", + "bbox": [ + 89, + 113, + 483, + 185 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 209, + 187, + 224 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Holger Caesar, Varun Bankiti, Alex Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. Nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 2, 6", + "[2] Qi Cai, Yingwei Pan, Ting Yao, and Tao Mei. 3d cascade rconn: High quality object detection in point clouds. IEEE Transactions on Image Processing, 31:5706-5719, 2022. 1", + "[3] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 7", + "[4] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 7", + "[5] Zehui Chen, Zheng Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, Feng Wu, and Feng Zhao. Graph-detr4d: Spatio-temporal graph modeling for multiview 3d object detection. IEEE Transactions on Image Processing, 33:4488-4500, 2024. 1", + "[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3D object detection. In AAAI, 2021. 1", + "[7] Xiaoying Ding, Weisi Lin, Zhenzhong Chen, and Xinfeng Zhang. Point cloud saliency detection by local and global feature fusion. IEEE Transactions on Image Processing, 28 (11):5379-5393, 2019. 1", + "[8] Lue Fan, Feng Wang, Naiyan Wang, and ZHAO-XIANG ZHANG. Fully sparse 3d object detection. Advances in Neural Information Processing Systems, 35:351-363, 2022. 1", + "[9] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3d-siamrpn: An end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 1", + "[10] Jiahui Fu, Guanghui Ren, Yunpeng Chen, and Si Liu. Improved pillar with fine-grained feature for 3d object detection. arXiv preprint arXiv:2110.06049, 2021. 3", + "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 6", + "[12] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 3" + ], + "bbox": [ + 93, + 234, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Shan Jiayao, Sifan Zhou, Yubo Cui, and Zheng Fang. Realtime 3d single object tracking with transformer. IEEE Transactions on Multimedia, 2022. 1", + "[14] Hongwu Kuang, Bei Wang, Jianping An, Ming Zhang, and Zehan Zhang. Voxel-fpn: Multi-scale voxel feature aggregation for 3d object detection from lidar point clouds. Sensors, 20(3):704, 2020. 1", + "[15] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 1, 3, 4, 6, 7, 8", + "[16] Duy Tho Le, Hengcan Shi, Hamid Rezatofighi, and Jianfei Cai. Accurate and real-time 3d pedestrian detection using an efficient attentive pillar network. IEEE Robotics and Automation Letters, 8(2):1159-1166, 2022. 3, 7, 8", + "[17] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3", + "[18] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv preprint arXiv:2102.05426, 2021. 3", + "[19] Jiawei Liu, Lin Niu, Zhihang Yuan, Dawei Yang, Xinggang Wang, and Wenyu Liu. Pd-quant: Post-training quantization based on prediction difference metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24427–24437, 2023. 3, 6", + "[20] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In AAAI, pages 11677-11684, 2020. 3, 6, 7", + "[21] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In ICCV, pages 3164-3173, 2021. 1", + "[22] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, pages 1-55, 2023. 3", + "[23] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 3", + "[24] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yely-sei Bondarenko, Mart Van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 6", + "[25] Konyul Park, Yechel Kim, Junho Koh, Byungwoo Park, and Jun Won Choi. Fine-grained pillar feature encoding via spatio-temporal virtual grid for 3d object detection. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4259-4265, 2024. 7", + "[26] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, pages 652-660, 2017. 1, 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27344", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1", + "[28] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1", + "[29] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. Ptt: Point-track-transformer module for 3d single object tracking in point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 1", + "[30] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: Real-time and high-performance pillar-based 3d object detection. In ECCV, 2022. 1, 3, 4, 6, 7, 8", + "[31] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE T-PAMI, 2020. 1, 3", + "[32] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3D object detection in a point cloud. In CVPR, pages 1711-1719, 2020. 1", + "[33] Lukas Stäcker, Juncong Fei, Philipp Heidenreich, Frank Bonarens, Jason Rambach, Didier Stricker, and Christoph Stiller. Deployment of deep neural networks for object detection on edge ai devices with runtime optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshop, pages 1015-1022, 2021. 3", + "[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 6", + "[35] OpenPCDet Development Team. OpenPCDet: An opensource toolbox for 3D object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.6", + "[36] Yonglin Tian, Lichao Huang, Hui Yu, Xiangbin Wu, Xuesong Li, Kunfeng Wang, Zilei Wang, and Fei-Yue Wang. Context-aware dynamic feature extraction for 3d object detection in point clouds. IEEE Transactions on Intelligent Transportation Systems, 23(8):10773-10785, 2021. 3", + "[37] Zhi Tian, Xiangxiang Chu, Xiaoming Wang, Xiaolin Wei, and Chunhua Shen. Fully convolutional one-stage 3d object detection on lidar range images. NeurIPS, 2022. 1", + "[38] Tai Wang, Xinge Zhu, and Dahua Lin. Reconfigurable voxels: A new representation for lidar-based point clouds. arXiv preprint arXiv:2004.02724, 2020. 3", + "[39] Yue Wang and Justin M Solomon. Object dgCNN: 3d object detection using dynamic graphs. NeurIPS, 2021. 3", + "[40] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization. arXiv preprint arXiv:2203.05740, 2022. 3", + "[41] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3D single stage object detector. In CVPR, pages 11040-11048, 2020. 1", + "[43] Maosheng Ye, Shuangjie Xu, and Tongyi Cao. Hvnet: Hybrid voxel network for lidar based 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1631-1640, 2020. 3", + "[44] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Centerpoint++ submission to the waymo real-time 3d detection challenge. 2021. Accessed: 2021-12-05. 1", + "[45] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 1, 3, 4, 6, 7", + "[46] Jiangyong Yu, Changyong Shu, Dawei Yang, Sifan Zhou, Zichen Yu, Xing Hu, and Yan Chen. Q-petr: Quant-aware position embedding transformation for multi-view 3d object detection. arXiv e-prints, pages arXiv-2502, 2025. 3", + "[47] Yifan Zhang, Zhen Dong, Huanrui Yang, Ming Lu, Cheng-Ching Tseng, Yandong Guo, Kurt Keutzer, Li Du, and Shanghang Zhang. Qd-bev: Quantization-aware view-guided distillation for multi-view 3d object detection. 2023. 3", + "[48] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In AAAI, 2021. 3", + "[49] Sifan Zhou, Zhi Tian, Xiangxiang Chu, Xinyu Zhang, Bo Zhang, Xiaobo Lu, Chengjian Feng, Zequn Jie, Patrick Yin Chiang, and Lin Ma. Fastpillars: A deployment-friendly pillar-based 3d detector. arXiv preprint arXiv:2302.02367, 2023. 1, 3, 4, 6, 7, 8", + "[50] Sifan Zhou, Liang Li, Xinyu Zhang, Bo Zhang, Shipeng Bai, Miao Sun, Ziyu Zhao, Xiaobo Lu, and Xiangxiang Chu. Lidar-ptq: Post-training quantization for point cloud 3d object detection. International Conference on Learning Representations (ICLR), 2024. 1, 3, 6", + "[51] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3D object detection. In CVPR, pages 4490-4499, 2018. 1, 3", + "[52] Yin Zhou, Pei Sun, Yu Zhang, Dragomir Anguelov, Jiyang Gao, Tom Ouyang, James Guo, Jiquan Ngiam, and Vijay Vasudevan. End-to-end multi-view fusion for 3d object detection in lidar point clouds. In Conference on Robot Learning, pages 923-932. PMLR, 2020. 3", + "[53] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv:1908.09492, 2019. 3" + ], + "bbox": [ + 516, + 90, + 903, + 742 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27345", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/18874676-27a7-4b87-b788-f15decae2835_model.json b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/18874676-27a7-4b87-b788-f15decae2835_model.json new file mode 100644 index 0000000000000000000000000000000000000000..670eca2c277608f09acc4a697d68f907d3f73a68 --- /dev/null +++ b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/18874676-27a7-4b87-b788-f15decae2835_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.212, + 0.131, + 0.787, + 0.175 + ], + "angle": 0, + "content": "PillarHist: A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.202, + 0.868, + 0.223 + ], + "angle": 0, + "content": "Sifan Zhou\\(^{1\\dagger*}\\) Zhihang Yuan\\(^{2\\dagger}\\) Dawei Yang\\(^{2\\dagger}\\) Xing Hu\\(^{2}\\) Jian Qian\\(^{3}\\) Ziyu Zhao\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.239, + 0.677, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Automation, Southeast University \\(^{2}\\)Houmo AI \\(^{3}\\)Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.278, + 0.838, + 0.292 + ], + "angle": 0, + "content": "{sifanjay, hahnyuan, zhaoziyu.950207}@gmail.com,{dawei.yang, xing.hu}@houmo.ai" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.486, + 0.783 + ], + "angle": 0, + "content": "Real-time and high-performance 3D object detection plays a critical role in autonomous driving and robotics. Recent pillar-based 3D object detectors have gained significant attention due to their compact representation and low computational overhead, making them suitable for onboard deployment and quantization. However, existing pillar-based detectors still suffer from information loss along height dimension and large numerical distribution difference during pillar feature encoding (PFE), which severely limits their performance and quantization potential. To address above issue, we first unveil the importance of different input information during PFE and identify the height dimension as a key factor in enhancing 3D detection performance. Motivated by this observation, we propose a height-aware pillar feature encoder, called PillarHist. Specifically, PillarHist statistics the discrete distribution of points at different heights within one pillar with the information entropy guidance. This simple yet effective design greatly preserves the information along the height dimension while significantly reducing the computation overhead of the PFE. Meanwhile, PillarHist also constrains the arithmetic distribution of PFE input to a stable range, making it quantization-friendly. Notably, PillarHist operates exclusively within the PFE stage to enhance performance, enabling seamless integration into existing pillar-based methods without introducing complex operations. Extensive experiments show the effectiveness of PillarHist in terms of both efficiency and performance." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.796, + 0.222, + 0.811 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.821, + 0.484, + 0.867 + ], + "angle": 0, + "content": "LiDAR-based 3D object detection plays a pivotal role in various applications [5, 7, 9, 13, 29, 37, 50], particularly in autonomous driving and robotics. It is important to detect" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.908, + 0.678 + ], + "angle": 0, + "content": "the objects in the surrounding environment fastly and accurately, which places a high demand for both performance and latency. Existing mainstream methods can be categorized into two types based on the point cloud representation: (1) point-based [8, 28, 31, 32, 42] and (2) grid-based models [2, 6, 14, 15, 21, 41, 41, 51]. Point-based models generally inherit the success in point cloud representation learning using deep neural networks[26, 44] and detect objects directly from raw points. However, these methods often face challenges in real-time onboard deployment as they often require point query/retrieval in 3D space (e.g., PointNet++ [27]). Grid-based methods rasterize the points into regular grid (voxel/pillar) and employ 2D or 3D convolution for feature learning. Compared to voxel-based methods, pillar-based detectors exhibit a more compact grid representation and can seamlessly integrate advanced 2D detection techniques without much modification. Therefore, those methods offer a favorable trade-off between performance and speed, making them highly suitable for practical deployment in autonomous systems. Furthermore, it is equally crucial to deploy them on resource-limited edge devices by improve their efficiency (e.g., reduce memory and computation cost)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The Pillar Feature Encoding (PFE) module plays a huge role in pillar-based networks [15, 17, 30, 45, 49]. Its main purpose is to transform the input point clouds, which are initially unordered (x, y, z, intensity) information, into pillars with a regular spatial arrangement. Each pillar has a fixed length of feature dimensions. In this paper, we observe that existing pillar-based methods suffer from two key problem in PFE module. 1 Information loss significantly hampers their performance. 2 Encoding scheme leads to a large variation of activations, making quantization challenging. As shown in Figure 1, the problem 1 can be attributed to two factors: (1) Lack of detailed height-aware feature encoding. The height information is not well handled, resulting in an incomplete representation of the 3D scene. It is" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Work done as an intern at Houmo AI. †Equal Contribution. ‡Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27336" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.094, + 0.62, + 0.108 + ], + "angle": 0, + "content": "(1) Lack of height-aware feature encoding" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.11, + 0.815, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.209, + 0.637, + 0.221 + ], + "angle": 0, + "content": "(2) Information loss with max-pooling operator" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.223, + 0.816, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.907, + 0.415 + ], + "angle": 0, + "content": "Figure 1. The drawbacks of pillar feature encoding are as follows: (1) Compared with the voxel feature encoding, the encoding process of pillars lacks height awareness, as it does not explicitly model the height dimension of the input point cloud. (2) PFE employs element-wise max-pooling operation, which inevitably results in information loss. As illustrated in the right side of (2), the max-pooling operation tends to preserve the feature information of points with higher height values within a pillar while disregarding points with lower height values, resulting in the loss of geometric information within the pillar." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.483, + 0.534 + ], + "angle": 0, + "content": "difficult to distinguish the height differences in various situations, like a pedestrian and a person standing on a car. (2) Max-pooling operator causes the loss of height information. The adoption of max-pooling operations within each pillar to derive pillar features leads to a loss of detailed information, which limits the ability of the encoded features to distinguish different objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.483, + 0.748 + ], + "angle": 0, + "content": "The problem 2 can be attributed the various arithmetic range of the PFE input. For instance, in nuScenes [1] dataset, the input of PFE includes \\(\\hat{p}_i = \\left\\{[x_i,y_i,z_i,r_i,x_i^m,y_i^m,z_i^m,x_i^c,y_i^c,z_i^c]\\in \\mathbb{R}^{N_v\\times 10}\\right\\}\\), where \\([x_i,y_i,z_i]\\in [-51.2,51.2]\\) is the point coordinates, \\(r_i\\in [0,255]\\) is the intensity, \\([x_i^m,y_i^m,z_i^m]\\in [-0.2,0.2]\\) (pillar size \\(= 0.2\\mathrm{m}\\)) denotes the offset of \\([x_i,y_i,z_i]\\) and the mean value of \\([x_i,y_i,z_i]\\) within one pillar, and \\([x_i^c,y_i^c,z_i^c]\\in [-4.0,4.0]\\) is the spatial offset of \\([x_i,y_i,z_i]\\) and the pillar center. Details refer to Sec 3.1. The different part of input represents the global- and local-contextual information, but exhibit significant difference in numerical distribution. Therefore, the arithmetic disparity makes it more susceptible to the effects of quantization (e.g., clipping error)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address the above issues, we propose a novel quantization-aware pillar feature encoder called PillarHist. It leverages a histogram technique to encode the height-aware information of each pillar into multiple bins, and the number of points falling within each bin is counted. The resulting histogram representation provides both semantic and geometric information from the height dimension. PillarHist incorporates the LiDAR reflection intensity of each point to obtain a weighted histogram of intensity, which adds distinguished height-aware information for object de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.624 + ], + "angle": 0, + "content": "tection. These histograms, along with the pillar center coordinates, are concatenated and passed through a linear layer to obtain a feature vector representing the pillar. Specially, PillarHist constrains the arithmetic distribution of PFE input to a stable range, making it friendly for further quantization. Besides, by operating linear projection at the pillar level instead of individual points, PillarHist significantly reduces the computation overhead, resulting in improved efficiency. It can be seamlessly integrated into existing pillar-based methods as a plugin module. Extensive experiments demonstrate the effectiveness of our proposed PillarHist in terms of both efficiency and performance. The contributions of our method can be summarised as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.648, + 0.905, + 0.709 + ], + "angle": 0, + "content": "- We conduct a comprehensive analysis of existing Pillar Feature Encoding (PFE). Our findings indicate that they suffer from information loss in the height dimension, and exits large variation of activation for model quantization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.71, + 0.905, + 0.785 + ], + "angle": 0, + "content": "- We propose a novel PFE module called PillarHist, which can mitigate the height information loss and reduce the computation cost of PFE. It is noteworthy that we make it a plugin module to conveniently insert into existing pillar-based methods while seamlessly boosting performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.785, + 0.905, + 0.845 + ], + "angle": 0, + "content": "- We explore the quantization-friendly properties of PillarHist. PillarHist could effectively eliminate the severe performance degradation after quantization caused by the raw PFE, bringing performance gain effortlessly." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.846, + 0.905, + 0.906 + ], + "angle": 0, + "content": "- We conducted extensive experiments to evaluate the effectiveness of PillarHist. Our results show that it achieved an average improvement of approximately 1 point in different dataset without extra latency." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.648, + 0.905, + 0.906 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27337" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.234, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.116, + 0.485, + 0.525 + ], + "angle": 0, + "content": "Grid Feature Encoding. Grid feature encoding is crucial for grid-based 3D object detectors, and significantly impacts model accuracy[12, 22]. Various methods have been proposed to extract features from two grid representations: Voxel and Pillar. VoxelNet [51] is a pioneering work that proposes a novel voxel feature encoding (VFE) layer, influencing many subsequent studies [31, 36, 38, 39, 43, 48, 52, 53]. Pillars serve as special voxels without vertical division; PointPillars [15] utilizes PointNet [26] for features extraction and employs max-pooling to obtain pillar features, leading to its widespread adoption in later works [10, 16, 17, 20, 30, 45, 49]. TANet [20] introduces a Triple Attention (TA) module to capture spatial correlations, but its reliance on point-wise operation and max-pooling incurs high computational costs and limits pillar representations. PiFeNet [16] proposes a lighter Pillar Aware Attention (PAA) module using channel-wise and point-wise attention, yet still employs pooling, affecting inference speed. Recently, PillarNet [30] introduced a dynamic PFE module with less information loss, while Fast-Pillars [49] presented a MAPE module for finer-grained geometric preservation. However, these methods do not adequately account for height-aware information and primarily use element-wise max-pooling or similar pooling operations, leading to further information loss. This paper focuses on enhancing detailed height-aware information and mitigate the information loss caused by pooling operations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.742 + ], + "angle": 0, + "content": "Quantization for 3D Object Detection. With the wide application of 3D object detection in autonomous driving and robotics, some quantization methods [46, 47] are designed to improve inference speed for onboard applications. In LiDAR-based 3D detection, [33] find that directly using INT8 quantization for 2D CNN significantly degrades performance on PointPillars [15], with an even greater reduction when using an entropy calibrator. Recently, LiDAR-PTQ [50] unveils the cause of performance collapse and propose an effective post-training quantization (PTQ) method. However, advanced quantization techniques[18, 19, 23, 40] are still needed to optimize quantization parameter within LiDAR-PTQ. Our PillarHist aim to make the network with PillarHist are quantization-friendly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.226, + 0.765 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.221, + 0.789 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "LiDAR-based 3D Object Detection. Here, we present the basic task definitions of LiDAR-based 3D object detection. Given a point set with \\(N\\) points in the 3D space, which is defined as \\(\\mathbf{P} = \\{\\mathbf{p}_i = [x_i,y_i,z_i,r_i]^T\\in \\mathbb{R}^{N\\times 4}\\}\\), where \\(x_{i},y_{i},z_{i}\\) denote the coordinate values of each point along the axes X, Y, Z, respectively, and \\(r_i\\) is the LiDAR reflection intensity. Given a set of objects in the 3D scene \\(\\mathbf{B} = \\{\\mathbf{b}_j =\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.183 + ], + "angle": 0, + "content": "\\([x_j, y_j, z_j, h_j, w_j, l_j, r_j, c_j]^T \\in \\mathbb{R}^{M \\times 8}\\), where \\(M\\) is the total number of objects, \\(b_i\\) is the \\(i\\)-th object in the scene, \\(x_j, y_j, z_j\\) is the object's center, \\(h_j, w_j, l_j\\) is the object's size, \\(r_j\\) is the object's heading angle and \\(c_j\\) is the object's class. 3D object detection aims to predict the 3D bounding boxes \\(\\mathbf{B}\\) from the point cloud \\(\\mathbf{P}\\) in driving scenarios accurately." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.907, + 0.284 + ], + "angle": 0, + "content": "Point Cloud Pillarization. We aim to transform the point cloud \\(\\mathbf{P}\\) into a regular 3D pillar representation. Assuming that \\(\\mathbf{P}\\) is in the 3D space with the range of \\(L, W, H\\) along the x, y, z axes, respectively, The point cloud \\(\\mathbf{P}\\) is uniformly partitioned into a specific pillar grid of size \\((\\delta_x, \\delta_y, \\mathrm{H})\\), we can compute the no-empty pillar index using:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.29, + 0.907, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\left(x _ {i n t}, y _ {i n t}, z _ {i n t}\\right) = \\left(\\lfloor \\frac {x _ {i}}{\\delta_ {x}} \\rceil , \\lfloor \\frac {y _ {i}}{\\delta_ {y}} \\rceil , \\lfloor \\frac {H}{2} \\rceil\\right), i = 0 \\dots N - 1 \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.329, + 0.907, + 0.421 + ], + "angle": 0, + "content": "where \\(\\lfloor \\cdot \\rfloor\\) is the floor operation, \\(x_{int},y_{int},z_{int}\\) is the 3D pillar index of \\(p_i\\) after pillarization. Notably, point cloud \\(\\mathbf{P}\\) is only pillarized along the XY plane without the height dimension. Let \\(V = \\{v_{i} = [x_{i},y_{i},z_{i},r_{i}]\\in \\mathbb{R}^{N_{v}\\times 4}\\} ,i\\in\\) \\(\\{1,\\dots,N_v\\}\\) , be a non-empty pillar contains \\(N_{v}\\) points, \\(N_{v}\\) is number of points in pillar \\(V\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.434, + 0.909, + 0.706 + ], + "angle": 0, + "content": "Pillar Feature Encoding. Pillar Feature Encoding (PFE) plays a crucial role in pillar-based 3D detection methods[15, 30, 49], directly impacting the representation learning of BEV feature and final detection accuracy. Here, taking PointPillars[15] as an example, we illustrate the details of PFE. The points in each pillar are initially concatenated along the channel dimension to form \\(\\hat{p}_i = \\{[x_i,y_i,z_i,r_i,x_i^m,y_i^m,z_i^m,x_i^c,y_i^c,z_i^c]\\in \\mathbb{R}^{N_v\\times 10}\\}\\), where \\([x_i,y_i,z_i]\\) is the original point coordinates in the ego frame, \\([x_i^m,y_i^m,z_i^m]\\) denotes the offset of \\(p_i\\) from the mean of the xyz coordinates of all points within the current pillar, and \\([x_i^c,y_i^c,z_i^c]\\) is the offset of \\(p_i\\) from the current pillar center. Notably, for each pillar, a sampling or padding strategy is employed to ensure that all pillars contain the same number of points. The concatenated point-wise features \\(\\hat{p}_i\\) within a non-empty pillar \\(V\\) are mapped to the high-dimensional feature space using MLP layers. This process can be mathematically formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.715, + 0.905, + 0.731 + ], + "angle": 0, + "content": "\\[\np _ {i} ^ {c} = m \\left(\\hat {p} _ {i}; w _ {m}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.739, + 0.906, + 0.812 + ], + "angle": 0, + "content": "where the function \\(m(\\cdot)\\) denotes MLP, \\(w_{m}\\) denotes the learnable weights of \\(m(\\cdot)\\), and \\(p_i^c \\in \\mathbb{R}^{N_v \\times D}\\) is the pointwise feature. 3 An element-wise max-pooling operation is applied along the dimension of point features, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.812, + 0.905, + 0.829 + ], + "angle": 0, + "content": "\\[\nf ^ {m} = \\max \\left(p _ {i} ^ {c}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( \\max(\\cdot) \\) means the max-pooling operator, and \\( f^m \\in \\mathbb{R}^D \\) is the resulting feature vector to represent non-empty pillar \\( V \\). A non-empty pillar \\( V \\in \\mathbb{R}^{N_v \\times 4} \\) consisting of \\( N_v \\) points is encoded into a feature vector \\( f^m \\in \\mathbb{R}^D \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27338" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.171, + 0.089, + 0.822, + 0.189 + ], + "angle": 0, + "content": "
(xi, xi^m, xi^c)(yi, yi^m, yi^c)(zi, zi^m, zi^c)3D_{R40}3D_{R11}
mAPEasyMod.HardmAPEasyMod.Hard
✓57.5067.9853.3051.2458.2067.9453.8252.85
✓59.4069.8555.4552.8959.9669.4156.1754.31
✓79.6386.6877.6274.5979.1586.6776.4974.28
PointPillars79.9887.0877.9074.9779.4386.2576.7675.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.194, + 0.875, + 0.208 + ], + "angle": 0, + "content": "Table 1. Ablation results of different input for PFE in PointPillars on the KITTI val set in \\( {\\mathrm{{AP}}}_{3D} \\) (R40) and \\( {\\mathrm{{AP}}}_{3D} \\) (R11) for car." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.226, + 0.482, + 0.243 + ], + "angle": 0, + "content": "3.2. What Kind of Information Do We Need in PFE?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.483, + 0.355 + ], + "angle": 0, + "content": "(1) Height Information is More Important. As aforementioned, the pioneering pillar-based method PointPillars[15] concatenates two extra relation information along the channel dimension of the raw point cloud, i.e. \\(\\left([x_i^m, y_i^m, z_i^m]\\right)\\) and \\(\\left([x_i^c, y_i^c, z_i^c]\\right)\\). Here, we decouple the input information along the \\(x\\), \\(y\\), and \\(z\\) axes, respectively. aiming to explore their impact on the final 3D detection performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.484, + 0.538 + ], + "angle": 0, + "content": "As shown in Table 1, when we only input the information along the \\(x\\) axis \\((x_{i}, x_{i}^{m}, x_{i}^{c})\\) or the information along the \\(y\\) axis \\((y_{i}, y_{i}^{m}, y_{i}^{c})\\), there is a significant degradation in the final 3D detection performance. The former results in a large decrease of \\(22.48 \\mathrm{~mAP}\\) of \\(3D_{R40}\\) and \\(21.23 \\mathrm{~mAP}\\) of \\(3D_{R11}\\), while the latter leads to a large decrease of \\(20.58 \\mathrm{~mAP}\\) of \\(3D_{R40}\\) and \\(19.47 \\mathrm{~mAP}\\) of \\(3D_{R11}\\). However, when we only input the information along the \\(z\\) axis \\((z_{i}, z_{i}^{m}, z_{i}^{c})\\), it also achieves comparable performance with the baseline. This indicates that in the pillar feature encoding process, the information along the \\(z\\) axis, a.k.a the height dimension, plays a key role in the final 3D detection performance." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.544, + 0.482, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.725 + ], + "angle": 0, + "content": "Figure 2. Visualization of the weights on (a) vehicle and (b) cyclist using Max-pooling in PFE. The size of points means scores. The point will be paid more attention if it has a higher score. Compared with original object's points, max-pooling operator in PFE inevitably result in the loss of fine-grained information the object and local geometries. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "(2) Pillar Information Loss by Max-pooling. As mentioned in Section 1, current pillar-based methods [15, 30, 45, 49] heavily rely on max pooling to aggregate point features within each pillar. However, this max-pooling operation (Equation 3) inevitably leads to the loss of detailed information, particularly the local geometric patterns that play an important role in object recognition. To illustrate this point, we have visualized the effect of the max-pooling operation in Figure 2. In the figure, we represent the points that are more likely to be selected by max pooling as larger and red spheres. It is evident from the visualization that a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.227, + 0.905, + 0.318 + ], + "angle": 0, + "content": "significant number of points are not selected after the max-pooling operation, while only a small fraction of points have a high probability of being chosen. This comparison clearly shows that the max-pooling operator results in a substantial loss of object geometry information compared to the original points of the object." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.324, + 0.874, + 0.339 + ], + "angle": 0, + "content": "3.3. Height-aware Discrete Histogram for PFE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.905, + 0.4 + ], + "angle": 0, + "content": "Hence, the challenge lies in finding a pillar encoding method that addresses these problems. Our goal is to develop a pillar feature encoding approach that possesses the following two key characteristics:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.905, + 0.477 + ], + "angle": 0, + "content": "It should eliminate the requirement for sampling operations while still maintaining the ability to capture the height-aware of points within each pillar. This means that the encoding method should effectively handle the variation in point density across different height levels within a pillar." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.905, + 0.553 + ], + "angle": 0, + "content": "It should keep the semantic and geometric information within the pillar, without resorting to max pooling or similar pooling operations. This is important to ensure that the encoded representation retains the relevant details and characteristics of the original pillar." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.401, + 0.905, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To achieve these objectives, we propose leveraging the histogram distribution. We can effectively capture the distribution of points within a pillar at different height levels. This allows us to maintain the height-awareness of the points without the need for sampling operations. The implementation of the histogram operation is straightforward. Initially, we divide the height range of the pillar into multiple bins, where each bin corresponds to a specific range of heights. For instance, if the pillar's height spans from 0 to 10 meters, we can partition it into several bins such as \\(0 - 1\\mathrm{m}\\), \\(1 - 2\\mathrm{m}\\), \\(2 - 3\\mathrm{m}\\), and so forth. Next, we proceed to count the number of points that fall within each bin for a given pillar. This counting process yields a count value for each bin, indicating the number of points residing within that particular height range. For instance, if a larger number of points are found in the lower height range bins, it signifies a higher density of points near the bottom of the pillar. Conversely, if there is a greater concentration of points in the higher height range bins, it suggests a higher density of points towards the top of the pillar. This approach provides the model with both the semantic and geometric information derived from the height dimension, thereby enhancing its understanding and representation of the point cloud data." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27339" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.187, + 0.101, + 0.808, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.469, + 0.908, + 0.555 + ], + "angle": 0, + "content": "Figure 3. The overall architecture of our proposed PillarHist. The PFE (a) in PointPillars first extracts point-wise feature, and then the point-wise features are max-pooled along the point dimension to obtain the pillar-wise representation. This process has three key drawbacks: (1) it lacks detailed height-aware information, (2) the MLP operates on point-wise level, incurring computational burden, (3) the max-pooling operation leads to information loss. In contrast, our proposed PillarHist (b) directly encodes the height-aware information based on histogram, and then perform the MLP operation on the pillar level rather than point level. Finally, the max-pooling operation is also eliminated in proposed PillarHist." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.562, + 0.485, + 0.851 + ], + "angle": 0, + "content": "By using the histogram technique, we propose a height-aware pillar feature encoder named PillarHist. The algorithm is summarized in Algorithm 1: (1) We compute the histogram of each pillar, denoted as \\(H^p \\in \\mathbb{R}^B\\), where \\(B\\) is the number of bins. (2) In addition, we incorporates the LiDAR reflection intensity of each point, \\(r_i\\), to obtain the weighted histogram of intensity, denoted as \\(H^i \\in \\mathbb{R}^B\\). The weighted histogram represents the mean LiDAR reflection intensity for each bin, providing valuable height-aware information for object detection. (3) To provide the global position of each pillar, we utilize the x and y coordinates of the pillar center, \\(x_{center}\\) and \\(y_{center}\\). Since the coordinates of the pillar are constant, this step does not require any additional computational effort. (4) We concatenate the computed histograms and the pillar center coordinates, resulting in \\([H^p | H^i | x_{center} | y_{center}] \\in \\mathbb{R}^{2B + 2}\\). (5) To represent the features of the pillars, we utilize a linear layer \\(f_\\theta\\) that projects the combined information into a feature vector. \\(\\theta \\in \\mathbb{R}^{D \\times (2B + 2)}\\) and \\(D\\) is the number of output channels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Specifically, the pillar center \\( x_{center} \\) and \\( y_{center} \\) provide the relative spatial relationships between different points within the point cloud. This spatial location information" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.909, + 0.82 + ], + "angle": 0, + "content": "of the pillars reflects the relative position of the targets in the scene, which can help the model better understand environmental context. For instance, if there are two pillars with the same point cloud distribution located at [0,10] and [10,0] in space, without the pillar coordinate to indicate the spatial relationship, these two pillars would be indistinguishable. This would pose significant challenges for the detector's object perception capabilities. Therefore, combining pillar center \\( x_{center} \\) and \\( y_{center} \\) with other features such as height and intensity can further enhance the model's perception capability. This is necessary to compensate for the missing spatial position information inherent in the raw point cloud data. By incorporating the pillar center coordinates, the model can better capture the geometric structure and spatial relationships between different parts of the scene. This is a crucial aspect for enabling accurate object detection in point cloud data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.911, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig 3, our PillarHist approach explicitly captures the histogram of points in different height intervals. This allows us to keep the semantic and geometric information within the pillar. PillarHist operates only during the PFE stage without introducing complex operations" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27340" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.102, + 0.095, + 0.335, + 0.111 + ], + "angle": 0, + "content": "Algorithm 1: PillarHist Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.113, + 0.46, + 0.174 + ], + "angle": 0, + "content": "Input: Raw point data in a pillar \\(V\\in \\mathbb{R}^{N_v\\times 4}\\) consisting of \\(N_{v}\\) points, number of height bins \\(B\\) ,number of output channels \\(D\\) ,height of the pillar \\(H\\) , linear layer \\(f_{\\theta}\\)" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.174, + 0.46, + 0.205 + ], + "angle": 0, + "content": "Output: Pillar's feature vector \\( Y \\in \\mathbb{R}^{D} \\) to represent the non-empty pillar \\( V \\)." + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.206, + 0.496, + 0.235 + ], + "angle": 0, + "content": "1: Divide the height \\( H \\) of the pillar into \\( B \\) bins, each bin having height \\( h \\), where \\( h = \\frac{H}{B} \\)" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.235, + 0.488, + 0.265 + ], + "angle": 0, + "content": "2: Initialize \\(H^p \\in \\mathbb{R}^B\\) to zero, where \\(H^p\\) represents the histogram of points in each bin" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.265, + 0.483, + 0.295 + ], + "angle": 0, + "content": "3: Initialize \\(H^i\\in \\mathbb{R}^B\\) to zero, where \\(H^{i}\\) represents the weighted histogram of intensity in each bin" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.296, + 0.368, + 0.31 + ], + "angle": 0, + "content": "4: for \\( p_i = \\{p_i | i = 1, 2, \\dots, N_v\\} \\) do" + }, + { + "type": "list", + "bbox": [ + 0.12, + 0.206, + 0.496, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.31, + 0.457, + 0.355 + ], + "angle": 0, + "content": "5: Compute the corresponding bin index \\( b_{i} \\) based on the \\( z \\) coordinate \\( z_{i} \\) of point \\( p_{i} \\): \\( b_{i} = \\lfloor \\frac{z_{i}}{h} \\rfloor \\)" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.356, + 0.307, + 0.369 + ], + "angle": 0, + "content": "6: \\(H^{p}[b_{i}]\\gets H^{p}[b_{i}] + 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.369, + 0.458, + 0.385 + ], + "angle": 0, + "content": "7: \\(H^{i}[b_{i}]\\gets H^{i}[b_{i}] + r_{i},\\quad \\triangleright r_{i}\\) is the LiDAR" + }, + { + "type": "list", + "bbox": [ + 0.144, + 0.356, + 0.458, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.386, + 0.322, + 0.399 + ], + "angle": 0, + "content": "8: reflection intensity of \\( p_i \\)" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.399, + 0.429, + 0.414 + ], + "angle": 0, + "content": "9: \\( H^{i}[b_{i}]\\coloneqq H^{i}[b_{i}] / H^{p}[b_{i}] \\) where \\( H^{i}[b_{i}]! = 0 \\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.416, + 0.493, + 0.429 + ], + "angle": 0, + "content": "10: Compute the pillar center coordinates \\(x_{center}, y_{center}\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.429, + 0.459, + 0.444 + ], + "angle": 0, + "content": "11: Concatenate: \\([H^p |H^i |x_{center}|y_{center}]\\in \\mathbb{R}^{2B + 2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.444, + 0.487, + 0.46 + ], + "angle": 0, + "content": "12: Linear projection: \\( Y = f_{\\theta}([H^{p}|H^{i}|x_{center}|y_{center}]) \\)" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.399, + 0.493, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.483, + 0.543 + ], + "angle": 0, + "content": "or interaction with other modules, allowing for seamless inserted into existing pillar-based methods serves as a plugin module with performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.483, + 0.756 + ], + "angle": 0, + "content": "The Efficiency of PillarHist: Here, we further analyze the efficiency of PillarHist. (1). The linear projection in PillarHist operates on the pillar level, not on individual points. In contrast, the previous approach [15, 20, 30, 45] performed the linear projection for each point within a pillar. By employing the linear projection at the pillar level, we significantly decrease the number of floating-point operations (FLOPs) required by the PFE, resulting in improved efficiency (Table 5). (2). PillarHist also avoids various memory-intensive operations, such as copying the original point cloud data, calculating distances \\((x^{m},y^{m},z^{m},x^{c},y^{c},z^{c})\\), and performing max pooling. This further enhances the efficiency of PillarHist in real-world deployment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.224, + 0.786 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Dataset and Training Details. We evaluate our PillarHist on KITTI [11], nuScenes [1] and Waymo dataset [34]. We report the official metrics. We adopt OpenPCDet [35] as the training framework and faithfully reproduced the different baseline methods based on the provided official setting on kitti and nuScenes dataset in the OpenPCDet [35]. we directly follow the setting of the four pillar-based 3D de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.319 + ], + "angle": 0, + "content": "tectors. (i.e., PointPillars [15], CenterPoint-Pillar [45], PillarNet [30], FastPillars [49]). In nuScenes [1] dataset, we uniformly sample 64 frames point cloud data from the validation set as the calibration set. Calibration is a standard step in quantization, where a small subset of dataset is used to determine activation quantization parameters (scale factors and zero points). This is a common practice in model quantization [24]. The calibration set proportions is \\(1.06\\%\\) (64/6019). We set the first and the last layer of the network to keep unquantized. Specifically, the quantization methods is widely used round-to-nearest (RTN) method, a common setting in prior works, such as PD-Quant [19] and LiDAPTQ [50]. We test all experiments on a single Nvidia Tesla V100 GPU. Please refer to Appendix for more quantization details and experiment analysis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.332, + 0.669, + 0.346 + ], + "angle": 0, + "content": "4.1. Overall Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Evaluation on nuScenes test set. The quantitative results on nuScenes test set are shown in Table 2. It can be seen that by integrating our proposed PillarHist into the current representative pillar-based detector PoinPillar [15], CenterPoint-Pillar [45], PillarNet [30], the proposed PH-PillarHist achieves significant performance gains on nuScenes test set across different pillar-based methods, respectively. Meanwhile, PH-PillarHist brings 6ms, 9ms, 5ms and 4ms reduction on inference latency, respectively. Notably, our method focuses on the pillar feature encoding module, rather than designing a state-of-the-art pillar-based detector. The proposed PillarHist is a plugin module to conveniently insert into existing mainstream pillar-based methods, seamlessly bringing performance gains and latency reduction. This presents the effectiveness and efficiency of PillarHist in not only boosting the detection accuracy, but also improving the computational efficiency, making it practical for real-world application." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.869 + ], + "angle": 0, + "content": "Evaluation on nuScenes val set. Here, we evaluate our PillarHist on different pillar-based baselines in nuScenes val set. We selected three representative pillar-based detectors: PointPillars [15], CP-Pillar [45], and PillarNet [30]. In detail, we replace the raw PFE module in each baseline model with our PillarHist, resulting in PH-PointPillars, PH-CP-Pillar, and PH-PillarNet, respectively. As shown in Table 3, the results demonstrate the effectiveness of proposed PillarHist, as we observe performance improvements of 1.7 NDS, 1.4 NDS, and 1.6 NDS on the three different baselines, respectively. These results highlight the effectiveness and broad applicability of our PillarHist in enhancing the represent power of pillar-based 3D object detectors, leading to consistent performance gains across different baseline architectures. Notably, On the nuScenes dataset, the nuScenes detection score (NDS) is the primary evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The Quantization Effectiveness of PillarHist. Furthermore, we also conduct quantization experiments to evaluate" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27341" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.299 + ], + "angle": 0, + "content": "
MethodNDSmAPLatencyCarTruckBusTrailerC.V.Ped.MotBycT.C.Bar
Focals Conv[3]70.063.8138 ms86.756.367.759.523.887.564.536.381.474.1
VoxelNeXt[4]70.064.566 ms84.653.064.755.828.785.873.245.779.074.6
FG-PFE[25]71.865.786 ms85.255.561.661.829.286.572.548.682.573.8
FastPillars [49]71.866.856 ms87.358.066.062.334.587.470.347.981.572.8
PointPillars [15]45.330.531 ms68.423.028.223.44.159.727.41.130.838.9
PH-PointPillars47.032.425 ms69.424.230.024.89.660.928.15.231.540.1
CenterPoint [45]65.558.096 ms84.651.060.253.217.583.453.728.776.770.9
PH-CenterPoint66.659.287 ms85.252.561.154.219.083.954.830.179.271.5
\\( PillarNet-18[30]^\\dagger \\)70.865.078 ms87.456.760.961.830.487.267.440.382.176.0
\\( PH-PillarNet-18^\\dagger \\)71.565.273 ms87.657.060.862.030.687.368.140.381.976.2
\\( PillarNet-34[30]^\\dagger \\)71.466.096 ms87.657.563.663.127.987.370.142.383.377.2
\\( PH-PillarNet-34^\\dagger \\)72.066.392 ms87.557.664.063.328.487.170.543.083.677.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.303, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Table 2. Performance of 3D object detection methods on nuScenes test set. \\( {}^{ \\dagger } \\) means the method that uses double-flip testing. All models listed take LIDAR data as input without image fusion or any model ensemble." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.338, + 0.47, + 0.547 + ], + "angle": 0, + "content": "
ModelsBits (W/A)NDSmAP
PointPillars [15]32/3252.137.4
8/844.8 (-7.3)29.8 (-7.6)
PH-PointPillars32/3253.8 (+1.7)38.8 (+1.4)
8/852.2 (-1.6)37.0 (-1.8)
CP-Pillar[45]32/3256.444.9
8/848.6 (-7.8)36.2 (-8.7)
PH-CP-Pillar32/3257.8 (+1.4)45.8 (+0.9)
8/856.3 (-1.5)44.3 (-1.5)
PillarNet[30]32/3261.452.6
8/854.6 (-6.8)43.1 (-9.5)
PH-PillarNet32/3263.0 (+1.6)52.8 (+0.2)
8/861.6 (-1.4)51.4 (-1.3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.579 + ], + "angle": 0, + "content": "Table 3. Generalization and quantization performance comparison on nuScene val set." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.486, + 0.902 + ], + "angle": 0, + "content": "the quantization-friendly ability. Specifically, we convert the weights and activations of the models from floating-point 32-bit to integer 8-bit and applied a simple posttraining quantization setting. As shown in Table 3, we find that the PointPillars, CP-Pillar, and PillarNet models suffered significant performance degradation after weight and activation quantization, with drops of 7.3 NDS and \\(7.6\\mathrm{mAP}\\) for PointPillars, 7.8 NDS and \\(8.7\\mathrm{mAP}\\) for CP-Pillar, and 6.8 NDS and \\(9.5\\mathrm{mAP}\\) for PillarNet. In contrast, the PH-PointPillars, PH-CP-Pillar, and PH-PillarNet models that integrate PillarHist only perform performance drops of 1.6 NDS, 1.5 NDS, and 1.4 NDS, respectively. Notably, the quantized models with PillarHist integration achieved comparable performance to original floating-point counterparts without PillarHist (52.2 v.s. 52.1 NDS on PointPillars, 56.3 v.s. 56.4 NDS on CenterPoint-Pillars and 61.6 v.s. 61.4 NDS on PillarNet). Specifically, PillarHist could effectively eliminate the severe performance degradation after quantization caused by other PFE design, bringing quantization performance gain seamlessly. These results demonstrate the quantization-friendly nature of PillarHist, which maintains" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.341, + 0.907, + 0.418 + ], + "angle": 0, + "content": "the model's representation power and performance even under quantization compression. This property is particularly valuable and practicable for deploying high-performance 3D object detectors on resource-constrained edge devices, where accuracy and latency are critical factors." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.426, + 0.905, + 0.495 + ], + "angle": 0, + "content": "
MethodsMean (L2) mAPHVehicle (L2) mAPmAPHPed. (L2) mAPmAPHCyc. (L2) mAPmAPH
CP-Pillar55.962.562.061.851.556.154.3
PH-CP-Pillar57.363.262.863.854.057.055.3
↑+1.4+0.7+0.8+2.0+2.5+0.9+1.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.495, + 0.907, + 0.524 + ], + "angle": 0, + "content": "Table 4. Apply PillarHist to CeterPoint-Pillars on Waymo val set trained with \\(20\\%\\) training data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.909, + 0.666 + ], + "angle": 0, + "content": "Generalization of PillarHist on Waymo val set. Here, we also evaluate our PillarHist on Waymo dataset. We select CenterPoint-Pillars as our baseline and only replaced their pillar feature encoding module with our PillarHist for comparison, as show in table our method can bring \\(1.2\\mathrm{mAP}\\) L2 and \\(1.5\\mathrm{mAPH}\\) L2 performance improvements. Notably, it can achieve \\(2.0\\mathrm{mAP}\\) L2 and \\(2.5\\mathrm{mAPH}\\) L2 performance gains in the pedestrian category. This result demonstrates the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.666, + 0.667, + 0.682 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.801 + ], + "angle": 0, + "content": "
PFE moduleCar 3D R40Ped 3D R40PFE GFLOPs
mAPEasyMod.HardmAPEasyMod.Hard
PFE[15]79.9887.0877.9074.9749.4154.7149.0144.520.152
PAA[16]81.0488.6279.2475.2748.5654.2647.9343.493.391
TA[20]80.7187.3379.1175.6949.4554.0049.3145.057.061
DynPFE*[30]79.9687.6377.7074.5549.3654.2749.0144.810.013
DynPFE[45]80.4687.8678.1675.3847.7752.2847.5343.520.112
FG-PFE[25]----50.7556.8349.8845.561.136
MAPE[49]81.1989.3978.5775.6348.1852.6848.0243.830.165
PillarHist81.4288.8079.1376.3051.0757.4550.4245.360.065
" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.805, + 0.888, + 0.819 + ], + "angle": 0, + "content": "Table 5. Performance of different PFE on the KITTI val set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparison of Different PFE modules. We compare the computation efficiency and performance of existing PFE modules. Specifically, we evaluate the following PFE modules: PFE in PointPillars [15], DynPFE in CenterPoint-Pillar [45], TA module in TANet [20], PAA module in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27342" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.334 + ], + "angle": 0, + "content": "PiFeNet [16], DynPFE* in PillarNet [30], and MAPE in FastPillars [49]. To ensure a fair comparison, we select PointPillars as the baseline method and conduct experiments by replacing the PFE in PointPillars with other PFE module (DynPFE, TA, PAA, DynpFE*, MAPE and our PillarHist). This setting eliminates the performance difference caused by different network design. As shown in Table 5, our PillarHist achieves the highest 3D AP compared to other PFE methods while maintaining superior computational efficiency. Notably, although DynPFE* used in PillarNet [30] has lower GFLOPs, it suffers from poor detection performance. Notably, our PillarHist method performs linear projection at the pillar-wise level rather than point-wise level in other PFE. This design can efficiently reduce the computational cost while maintaining the necessary structural information within pillar without max-pooling operations." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.414 + ], + "angle": 0, + "content": "
heightintensitycoordsCar 3D40Ped 3D40
mAPEasyMod.HardmAPEasyMod.Hard
✓80.1887.4778.1875.2248.0653.7547.5142.92
✓✓80.8188.2978.6275.5349.6255.4249.1044.36
✓✓✓81.4188.8079.1376.3051.0757.4550.4245.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.423, + 0.483, + 0.452 + ], + "angle": 0, + "content": "Table 6. Ablation results of different input for PillarHist on the KITTI val set for car and pedestrian." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.78 + ], + "angle": 0, + "content": "The Effects of Different Input in PillarHist. Here, we conduct ablation study to analyze the impact of different input. As shown in Table 6, when we incorporate the intensity information, it leads to performance improvements for both the vehicle and pedestrian categories. Particularly for the pedestrian, approximately 1.7, 1.6, and 1.6 points for the easy, moderate, and hard levels, respectively. This notable improvement indicates that for small objects like pedestrians, the intensity values are crucial in distinguishing them from the background and enhancing their distinctive features. Furthermore, by adding the pillar coordinates information, there is a performance improvement of 0.6 and 1.5 points for the vehicle and pedestrian. We believe this improvement is because explicitly incorporating the global coordinate of the pillars helps the model better identify the object positions with the enhanced encoding information that includes height awareness. Notably, after adding the pillar coordinates, there is an improvement of 2.03 points in easy AP for pedestrians. This can be attributed that the pillars encode not only varying height distribution but also including the difference between empty and non-empty pillars." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Feature Map Visualization of PillarHist. Here, as shown in Figure 4, we apply our PillarHist method to PointPillars [15] (a) and compare the features from the max-pooling operation in their PFE. We find that the max pooling operation tends to focus more on the points at higher positions within the same pillar in BEV (with larger coordinates), resulting in a biased feature representation. In contrast, our PillarHist can focus more evenly on each point within" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.083, + 0.905, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.297, + 0.908, + 0.368 + ], + "angle": 0, + "content": "Figure 4. (a) Visualization of the learned attention scores in PillarHist on different scenes. The color of points represents their scores. The point will be paid more attention if it has a higher score. (b) Visualization of the learned BEV feature in PointPillars and PH-PointPillars. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.909, + 0.571 + ], + "angle": 0, + "content": "the pillar, encoding the height information into an unbiased histogram distribution, thereby obtaining richer geometric-semantic feature representation. This difference is particularly obvious at the car door location, the points within the pillar is typically uniformly distributed along height; however, max pooling emphasizes the higher contours, whereas PillarHist provides a better representation of the complete geometry of the car door Besides, as shown in Figure 4 (b), we also visualize the BEV features maps. Our method can obtain a more informative BEV feature representation compared to the PFE in PointPillars [15]. This also demonstrates the underlying reason for the better performance of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.586, + 0.634, + 0.601 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.909, + 0.899 + ], + "angle": 0, + "content": "In this paper, we addressed the issue of information loss and large numerical distribution difference in existing pillar-based 3D object detectors during the PFE stage. We identify the height dimension as a crucial factor for enhancing 3D detection performance and propose a height-aware pillar feature encoder called PillarHist. By leveraging histogram, PillarHist preserve the information along the height dimension while effectively mitigating the information loss introduced by max-pooling operations. Importantly, PillarHist operated exclusively within the PFE stage, allowing seamless insert into existing pillar-based methods serves as a plug-in module. Meanwhile, PillarHist also constrains the arithmetic distribution of PFE input to a stable range, making it quantization-friendly. Extensive experiments demonstrated the effectiveness of PillarHist in terms of both efficiency and performance. Our proposed method achieved an average improvement of approximately \\(1.0\\%\\) NDS across various baselines, showcasing its ability to enhance 3D object detection accuracy and computation efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27343" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.186 + ], + "angle": 0, + "content": "We thank anonymous reviewers for their kind help of this work. This work was supported by National Natural Science Foundation of China (No.62271143), Frontier Technologies RD Program of Jiangsu (No. BF2024060) and the Big Data Computing Center of Southeast University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.21, + 0.188, + 0.226 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.235, + 0.484, + 0.304 + ], + "angle": 0, + "content": "[1] Holger Caesar, Varun Bankiti, Alex Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. Nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.305, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[2] Qi Cai, Yingwei Pan, Ting Yao, and Tao Mei. 3d cascade rconn: High quality object detection in point clouds. IEEE Transactions on Image Processing, 31:5706-5719, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[3] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[4] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.471, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[5] Zehui Chen, Zheng Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, Feng Wu, and Feng Zhao. Graph-detr4d: Spatio-temporal graph modeling for multiview 3d object detection. IEEE Transactions on Image Processing, 33:4488-4500, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.482, + 0.595 + ], + "angle": 0, + "content": "[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3D object detection. In AAAI, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.482, + 0.651 + ], + "angle": 0, + "content": "[7] Xiaoying Ding, Weisi Lin, Zhenzhong Chen, and Xinfeng Zhang. Point cloud saliency detection by local and global feature fusion. IEEE Transactions on Image Processing, 28 (11):5379-5393, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.652, + 0.482, + 0.693 + ], + "angle": 0, + "content": "[8] Lue Fan, Feng Wang, Naiyan Wang, and ZHAO-XIANG ZHANG. Fully sparse 3d object detection. Advances in Neural Information Processing Systems, 35:351-363, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.694, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[9] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3d-siamrpn: An end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.749, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Jiahui Fu, Guanghui Ren, Yunpeng Chen, and Si Liu. Improved pillar with fine-grained feature for 3d object detection. arXiv preprint arXiv:2110.06049, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.235, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[13] Shan Jiayao, Sifan Zhou, Yubo Cui, and Zheng Fang. Realtime 3d single object tracking with transformer. IEEE Transactions on Multimedia, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[14] Hongwu Kuang, Bei Wang, Jianping An, Ming Zhang, and Zehan Zhang. Voxel-fpn: Multi-scale voxel feature aggregation for 3d object detection from lidar point clouds. Sensors, 20(3):704, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[15] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 1, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[16] Duy Tho Le, Hengcan Shi, Hamid Rezatofighi, and Jianfei Cai. Accurate and real-time 3d pedestrian detection using an efficient attentive pillar network. IEEE Robotics and Automation Letters, 8(2):1159-1166, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[17] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[18] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv preprint arXiv:2102.05426, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.489 + ], + "angle": 0, + "content": "[19] Jiawei Liu, Lin Niu, Zhihang Yuan, Dawei Yang, Xinggang Wang, and Wenyu Liu. Pd-quant: Post-training quantization based on prediction difference metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24427–24437, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[20] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In AAAI, pages 11677-11684, 2020. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[21] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In ICCV, pages 3164-3173, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[22] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, pages 1-55, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[23] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[24] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yely-sei Bondarenko, Mart Van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[25] Konyul Park, Yechel Kim, Junho Koh, Byungwoo Park, and Jun Won Choi. Fine-grained pillar feature encoding via spatio-temporal virtual grid for 3d object detection. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4259-4265, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, pages 652-660, 2017. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27344" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[28] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[29] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. Ptt: Point-track-transformer module for 3d single object tracking in point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[30] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: Real-time and high-performance pillar-based 3d object detection. In ECCV, 2022. 1, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[31] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE T-PAMI, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[32] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3D object detection in a point cloud. In CVPR, pages 1711-1719, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[33] Lukas Stäcker, Juncong Fei, Philipp Heidenreich, Frank Bonarens, Jason Rambach, Didier Stricker, and Christoph Stiller. Deployment of deep neural networks for object detection on edge ai devices with runtime optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshop, pages 1015-1022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[35] OpenPCDet Development Team. OpenPCDet: An opensource toolbox for 3D object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[36] Yonglin Tian, Lichao Huang, Hui Yu, Xiangbin Wu, Xuesong Li, Kunfeng Wang, Zilei Wang, and Fei-Yue Wang. Context-aware dynamic feature extraction for 3d object detection in point clouds. IEEE Transactions on Intelligent Transportation Systems, 23(8):10773-10785, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[37] Zhi Tian, Xiangxiang Chu, Xiaoming Wang, Xiaolin Wei, and Chunhua Shen. Fully convolutional one-stage 3d object detection on lidar range images. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[38] Tai Wang, Xinge Zhu, and Dahua Lin. Reconfigurable voxels: A new representation for lidar-based point clouds. arXiv preprint arXiv:2004.02724, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[39] Yue Wang and Justin M Solomon. Object dgCNN: 3d object detection using dynamic graphs. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[40] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization. arXiv preprint arXiv:2203.05740, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[41] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[42] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3D single stage object detector. In CVPR, pages 11040-11048, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[43] Maosheng Ye, Shuangjie Xu, and Tongyi Cao. Hvnet: Hybrid voxel network for lidar based 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1631-1640, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[44] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Centerpoint++ submission to the waymo real-time 3d detection challenge. 2021. Accessed: 2021-12-05. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[45] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[46] Jiangyong Yu, Changyong Shu, Dawei Yang, Sifan Zhou, Zichen Yu, Xing Hu, and Yan Chen. Q-petr: Quant-aware position embedding transformation for multi-view 3d object detection. arXiv e-prints, pages arXiv-2502, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[47] Yifan Zhang, Zhen Dong, Huanrui Yang, Ming Lu, Cheng-Ching Tseng, Yandong Guo, Kurt Keutzer, Li Du, and Shanghang Zhang. Qd-bev: Quantization-aware view-guided distillation for multi-view 3d object detection. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[48] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[49] Sifan Zhou, Zhi Tian, Xiangxiang Chu, Xinyu Zhang, Bo Zhang, Xiaobo Lu, Chengjian Feng, Zequn Jie, Patrick Yin Chiang, and Lin Ma. Fastpillars: A deployment-friendly pillar-based 3d detector. arXiv preprint arXiv:2302.02367, 2023. 1, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[50] Sifan Zhou, Liang Li, Xinyu Zhang, Bo Zhang, Shipeng Bai, Miao Sun, Ziyu Zhao, Xiaobo Lu, and Xiangxiang Chu. Lidar-ptq: Post-training quantization for point cloud 3d object detection. International Conference on Learning Representations (ICLR), 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[51] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3D object detection. In CVPR, pages 4490-4499, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.7 + ], + "angle": 0, + "content": "[52] Yin Zhou, Pei Sun, Yu Zhang, Dragomir Anguelov, Jiyang Gao, Tom Ouyang, James Guo, Jiquan Ngiam, and Vijay Vasudevan. End-to-end multi-view fusion for 3d object detection in lidar point clouds. In Conference on Robot Learning, pages 923-932. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[53] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv:1908.09492, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27345" + } + ] +] \ No newline at end of file diff --git a/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/18874676-27a7-4b87-b788-f15decae2835_origin.pdf b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/18874676-27a7-4b87-b788-f15decae2835_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e9b86d2950305f375c1a69ab50686f24cce99c9f --- /dev/null +++ b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/18874676-27a7-4b87-b788-f15decae2835_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c72b41651fe0a4d01ff68884422a337fc8ab1e28d61c6da898a062149f0891d5 +size 709648 diff --git a/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/full.md b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/full.md new file mode 100644 index 0000000000000000000000000000000000000000..39d8679352ae048d7b760ce7f377af9023d0ea1b --- /dev/null +++ b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/full.md @@ -0,0 +1,261 @@ +# PillarHist: A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram + +Sifan Zhou $^{1\dagger*}$ Zhihang Yuan $^{2\dagger}$ Dawei Yang $^{2\dagger}$ Xing Hu $^{2}$ Jian Qian $^{3}$ Ziyu Zhao $^{1}$ + +$^{1}$ School of Automation, Southeast University $^{2}$ Houmo AI $^{3}$ Fudan University + +{sifanjay, hahnyuan, zhaoziyu.950207}@gmail.com,{dawei.yang, xing.hu}@houmo.ai + +# Abstract + +Real-time and high-performance 3D object detection plays a critical role in autonomous driving and robotics. Recent pillar-based 3D object detectors have gained significant attention due to their compact representation and low computational overhead, making them suitable for onboard deployment and quantization. However, existing pillar-based detectors still suffer from information loss along height dimension and large numerical distribution difference during pillar feature encoding (PFE), which severely limits their performance and quantization potential. To address above issue, we first unveil the importance of different input information during PFE and identify the height dimension as a key factor in enhancing 3D detection performance. Motivated by this observation, we propose a height-aware pillar feature encoder, called PillarHist. Specifically, PillarHist statistics the discrete distribution of points at different heights within one pillar with the information entropy guidance. This simple yet effective design greatly preserves the information along the height dimension while significantly reducing the computation overhead of the PFE. Meanwhile, PillarHist also constrains the arithmetic distribution of PFE input to a stable range, making it quantization-friendly. Notably, PillarHist operates exclusively within the PFE stage to enhance performance, enabling seamless integration into existing pillar-based methods without introducing complex operations. Extensive experiments show the effectiveness of PillarHist in terms of both efficiency and performance. + +# 1. Introduction + +LiDAR-based 3D object detection plays a pivotal role in various applications [5, 7, 9, 13, 29, 37, 50], particularly in autonomous driving and robotics. It is important to detect + +the objects in the surrounding environment fastly and accurately, which places a high demand for both performance and latency. Existing mainstream methods can be categorized into two types based on the point cloud representation: (1) point-based [8, 28, 31, 32, 42] and (2) grid-based models [2, 6, 14, 15, 21, 41, 41, 51]. Point-based models generally inherit the success in point cloud representation learning using deep neural networks[26, 44] and detect objects directly from raw points. However, these methods often face challenges in real-time onboard deployment as they often require point query/retrieval in 3D space (e.g., PointNet++ [27]). Grid-based methods rasterize the points into regular grid (voxel/pillar) and employ 2D or 3D convolution for feature learning. Compared to voxel-based methods, pillar-based detectors exhibit a more compact grid representation and can seamlessly integrate advanced 2D detection techniques without much modification. Therefore, those methods offer a favorable trade-off between performance and speed, making them highly suitable for practical deployment in autonomous systems. Furthermore, it is equally crucial to deploy them on resource-limited edge devices by improve their efficiency (e.g., reduce memory and computation cost). + +The Pillar Feature Encoding (PFE) module plays a huge role in pillar-based networks [15, 17, 30, 45, 49]. Its main purpose is to transform the input point clouds, which are initially unordered (x, y, z, intensity) information, into pillars with a regular spatial arrangement. Each pillar has a fixed length of feature dimensions. In this paper, we observe that existing pillar-based methods suffer from two key problem in PFE module. 1 Information loss significantly hampers their performance. 2 Encoding scheme leads to a large variation of activations, making quantization challenging. As shown in Figure 1, the problem 1 can be attributed to two factors: (1) Lack of detailed height-aware feature encoding. The height information is not well handled, resulting in an incomplete representation of the 3D scene. It is + +![](images/b5686df97058c02679acbf77ed3a8da0e002a5b124a2e9e06889d5431749e19c.jpg) +(1) Lack of height-aware feature encoding + +![](images/0a58559b08e8eb8e5efcd382426f84b8acce917b642f673fab80657b303da536.jpg) +(2) Information loss with max-pooling operator +Figure 1. The drawbacks of pillar feature encoding are as follows: (1) Compared with the voxel feature encoding, the encoding process of pillars lacks height awareness, as it does not explicitly model the height dimension of the input point cloud. (2) PFE employs element-wise max-pooling operation, which inevitably results in information loss. As illustrated in the right side of (2), the max-pooling operation tends to preserve the feature information of points with higher height values within a pillar while disregarding points with lower height values, resulting in the loss of geometric information within the pillar. + +difficult to distinguish the height differences in various situations, like a pedestrian and a person standing on a car. (2) Max-pooling operator causes the loss of height information. The adoption of max-pooling operations within each pillar to derive pillar features leads to a loss of detailed information, which limits the ability of the encoded features to distinguish different objects. + +The problem 2 can be attributed the various arithmetic range of the PFE input. For instance, in nuScenes [1] dataset, the input of PFE includes $\hat{p}_i = \left\{[x_i,y_i,z_i,r_i,x_i^m,y_i^m,z_i^m,x_i^c,y_i^c,z_i^c]\in \mathbb{R}^{N_v\times 10}\right\}$ , where $[x_i,y_i,z_i]\in [-51.2,51.2]$ is the point coordinates, $r_i\in [0,255]$ is the intensity, $[x_i^m,y_i^m,z_i^m]\in [-0.2,0.2]$ (pillar size $= 0.2\mathrm{m}$ ) denotes the offset of $[x_i,y_i,z_i]$ and the mean value of $[x_i,y_i,z_i]$ within one pillar, and $[x_i^c,y_i^c,z_i^c]\in [-4.0,4.0]$ is the spatial offset of $[x_i,y_i,z_i]$ and the pillar center. Details refer to Sec 3.1. The different part of input represents the global- and local-contextual information, but exhibit significant difference in numerical distribution. Therefore, the arithmetic disparity makes it more susceptible to the effects of quantization (e.g., clipping error). + +To address the above issues, we propose a novel quantization-aware pillar feature encoder called PillarHist. It leverages a histogram technique to encode the height-aware information of each pillar into multiple bins, and the number of points falling within each bin is counted. The resulting histogram representation provides both semantic and geometric information from the height dimension. PillarHist incorporates the LiDAR reflection intensity of each point to obtain a weighted histogram of intensity, which adds distinguished height-aware information for object de + +tection. These histograms, along with the pillar center coordinates, are concatenated and passed through a linear layer to obtain a feature vector representing the pillar. Specially, PillarHist constrains the arithmetic distribution of PFE input to a stable range, making it friendly for further quantization. Besides, by operating linear projection at the pillar level instead of individual points, PillarHist significantly reduces the computation overhead, resulting in improved efficiency. It can be seamlessly integrated into existing pillar-based methods as a plugin module. Extensive experiments demonstrate the effectiveness of our proposed PillarHist in terms of both efficiency and performance. The contributions of our method can be summarised as follows: + +- We conduct a comprehensive analysis of existing Pillar Feature Encoding (PFE). Our findings indicate that they suffer from information loss in the height dimension, and exits large variation of activation for model quantization. +- We propose a novel PFE module called PillarHist, which can mitigate the height information loss and reduce the computation cost of PFE. It is noteworthy that we make it a plugin module to conveniently insert into existing pillar-based methods while seamlessly boosting performance. +- We explore the quantization-friendly properties of PillarHist. PillarHist could effectively eliminate the severe performance degradation after quantization caused by the raw PFE, bringing performance gain effortlessly. +- We conducted extensive experiments to evaluate the effectiveness of PillarHist. Our results show that it achieved an average improvement of approximately 1 point in different dataset without extra latency. + +# 2. Related Work + +Grid Feature Encoding. Grid feature encoding is crucial for grid-based 3D object detectors, and significantly impacts model accuracy[12, 22]. Various methods have been proposed to extract features from two grid representations: Voxel and Pillar. VoxelNet [51] is a pioneering work that proposes a novel voxel feature encoding (VFE) layer, influencing many subsequent studies [31, 36, 38, 39, 43, 48, 52, 53]. Pillars serve as special voxels without vertical division; PointPillars [15] utilizes PointNet [26] for features extraction and employs max-pooling to obtain pillar features, leading to its widespread adoption in later works [10, 16, 17, 20, 30, 45, 49]. TANet [20] introduces a Triple Attention (TA) module to capture spatial correlations, but its reliance on point-wise operation and max-pooling incurs high computational costs and limits pillar representations. PiFeNet [16] proposes a lighter Pillar Aware Attention (PAA) module using channel-wise and point-wise attention, yet still employs pooling, affecting inference speed. Recently, PillarNet [30] introduced a dynamic PFE module with less information loss, while Fast-Pillars [49] presented a MAPE module for finer-grained geometric preservation. However, these methods do not adequately account for height-aware information and primarily use element-wise max-pooling or similar pooling operations, leading to further information loss. This paper focuses on enhancing detailed height-aware information and mitigate the information loss caused by pooling operations. + +Quantization for 3D Object Detection. With the wide application of 3D object detection in autonomous driving and robotics, some quantization methods [46, 47] are designed to improve inference speed for onboard applications. In LiDAR-based 3D detection, [33] find that directly using INT8 quantization for 2D CNN significantly degrades performance on PointPillars [15], with an even greater reduction when using an entropy calibrator. Recently, LiDAR-PTQ [50] unveils the cause of performance collapse and propose an effective post-training quantization (PTQ) method. However, advanced quantization techniques[18, 19, 23, 40] are still needed to optimize quantization parameter within LiDAR-PTQ. Our PillarHist aim to make the network with PillarHist are quantization-friendly. + +# 3. Methodology + +# 3.1. Preliminary + +LiDAR-based 3D Object Detection. Here, we present the basic task definitions of LiDAR-based 3D object detection. Given a point set with $N$ points in the 3D space, which is defined as $\mathbf{P} = \{\mathbf{p}_i = [x_i,y_i,z_i,r_i]^T\in \mathbb{R}^{N\times 4}\}$ , where $x_{i},y_{i},z_{i}$ denote the coordinate values of each point along the axes X, Y, Z, respectively, and $r_i$ is the LiDAR reflection intensity. Given a set of objects in the 3D scene $\mathbf{B} = \{\mathbf{b}_j =$ + +$[x_j, y_j, z_j, h_j, w_j, l_j, r_j, c_j]^T \in \mathbb{R}^{M \times 8}$ , where $M$ is the total number of objects, $b_i$ is the $i$ -th object in the scene, $x_j, y_j, z_j$ is the object's center, $h_j, w_j, l_j$ is the object's size, $r_j$ is the object's heading angle and $c_j$ is the object's class. 3D object detection aims to predict the 3D bounding boxes $\mathbf{B}$ from the point cloud $\mathbf{P}$ in driving scenarios accurately. + +Point Cloud Pillarization. We aim to transform the point cloud $\mathbf{P}$ into a regular 3D pillar representation. Assuming that $\mathbf{P}$ is in the 3D space with the range of $L, W, H$ along the x, y, z axes, respectively, The point cloud $\mathbf{P}$ is uniformly partitioned into a specific pillar grid of size $(\delta_x, \delta_y, \mathrm{H})$ , we can compute the no-empty pillar index using: + +$$ +\left(x _ {i n t}, y _ {i n t}, z _ {i n t}\right) = \left(\lfloor \frac {x _ {i}}{\delta_ {x}} \rceil , \lfloor \frac {y _ {i}}{\delta_ {y}} \rceil , \lfloor \frac {H}{2} \rceil\right), i = 0 \dots N - 1 \tag {1} +$$ + +where $\lfloor \cdot \rfloor$ is the floor operation, $x_{int},y_{int},z_{int}$ is the 3D pillar index of $p_i$ after pillarization. Notably, point cloud $\mathbf{P}$ is only pillarized along the XY plane without the height dimension. Let $V = \{v_{i} = [x_{i},y_{i},z_{i},r_{i}]\in \mathbb{R}^{N_{v}\times 4}\} ,i\in$ $\{1,\dots,N_v\}$ , be a non-empty pillar contains $N_{v}$ points, $N_{v}$ is number of points in pillar $V$ + +Pillar Feature Encoding. Pillar Feature Encoding (PFE) plays a crucial role in pillar-based 3D detection methods[15, 30, 49], directly impacting the representation learning of BEV feature and final detection accuracy. Here, taking PointPillars[15] as an example, we illustrate the details of PFE. The points in each pillar are initially concatenated along the channel dimension to form $\hat{p}_i = \{[x_i,y_i,z_i,r_i,x_i^m,y_i^m,z_i^m,x_i^c,y_i^c,z_i^c]\in \mathbb{R}^{N_v\times 10}\}$ , where $[x_i,y_i,z_i]$ is the original point coordinates in the ego frame, $[x_i^m,y_i^m,z_i^m]$ denotes the offset of $p_i$ from the mean of the xyz coordinates of all points within the current pillar, and $[x_i^c,y_i^c,z_i^c]$ is the offset of $p_i$ from the current pillar center. Notably, for each pillar, a sampling or padding strategy is employed to ensure that all pillars contain the same number of points. The concatenated point-wise features $\hat{p}_i$ within a non-empty pillar $V$ are mapped to the high-dimensional feature space using MLP layers. This process can be mathematically formulated as: + +$$ +p _ {i} ^ {c} = m \left(\hat {p} _ {i}; w _ {m}\right) \tag {2} +$$ + +where the function $m(\cdot)$ denotes MLP, $w_{m}$ denotes the learnable weights of $m(\cdot)$ , and $p_i^c \in \mathbb{R}^{N_v \times D}$ is the pointwise feature. 3 An element-wise max-pooling operation is applied along the dimension of point features, which is formulated as: + +$$ +f ^ {m} = \max \left(p _ {i} ^ {c}\right) \tag {3} +$$ + +where $\max(\cdot)$ means the max-pooling operator, and $f^m \in \mathbb{R}^D$ is the resulting feature vector to represent non-empty pillar $V$ . A non-empty pillar $V \in \mathbb{R}^{N_v \times 4}$ consisting of $N_v$ points is encoded into a feature vector $f^m \in \mathbb{R}^D$ . + +
(xi, xi^m, xi^c)(yi, yi^m, yi^c)(zi, zi^m, zi^c)3D_{R40}3D_{R11}
mAPEasyMod.HardmAPEasyMod.Hard
✓57.5067.9853.3051.2458.2067.9453.8252.85
✓59.4069.8555.4552.8959.9669.4156.1754.31
✓79.6386.6877.6274.5979.1586.6776.4974.28
PointPillars79.9887.0877.9074.9779.4386.2576.7675.02
+ +Table 1. Ablation results of different input for PFE in PointPillars on the KITTI val set in ${\mathrm{{AP}}}_{3D}$ (R40) and ${\mathrm{{AP}}}_{3D}$ (R11) for car. + +# 3.2. What Kind of Information Do We Need in PFE? + +(1) Height Information is More Important. As aforementioned, the pioneering pillar-based method PointPillars[15] concatenates two extra relation information along the channel dimension of the raw point cloud, i.e. $\left([x_i^m, y_i^m, z_i^m]\right)$ and $\left([x_i^c, y_i^c, z_i^c]\right)$ . Here, we decouple the input information along the $x$ , $y$ , and $z$ axes, respectively. aiming to explore their impact on the final 3D detection performance. + +As shown in Table 1, when we only input the information along the $x$ axis $(x_{i}, x_{i}^{m}, x_{i}^{c})$ or the information along the $y$ axis $(y_{i}, y_{i}^{m}, y_{i}^{c})$ , there is a significant degradation in the final 3D detection performance. The former results in a large decrease of $22.48 \mathrm{~mAP}$ of $3D_{R40}$ and $21.23 \mathrm{~mAP}$ of $3D_{R11}$ , while the latter leads to a large decrease of $20.58 \mathrm{~mAP}$ of $3D_{R40}$ and $19.47 \mathrm{~mAP}$ of $3D_{R11}$ . However, when we only input the information along the $z$ axis $(z_{i}, z_{i}^{m}, z_{i}^{c})$ , it also achieves comparable performance with the baseline. This indicates that in the pillar feature encoding process, the information along the $z$ axis, a.k.a the height dimension, plays a key role in the final 3D detection performance. + +![](images/2183e81d51230276621f5f7cda7cdcfcccd9335f3ef2b6f51e7c1da76449a5ce.jpg) +Figure 2. Visualization of the weights on (a) vehicle and (b) cyclist using Max-pooling in PFE. The size of points means scores. The point will be paid more attention if it has a higher score. Compared with original object's points, max-pooling operator in PFE inevitably result in the loss of fine-grained information the object and local geometries. Best viewed in color. + +(2) Pillar Information Loss by Max-pooling. As mentioned in Section 1, current pillar-based methods [15, 30, 45, 49] heavily rely on max pooling to aggregate point features within each pillar. However, this max-pooling operation (Equation 3) inevitably leads to the loss of detailed information, particularly the local geometric patterns that play an important role in object recognition. To illustrate this point, we have visualized the effect of the max-pooling operation in Figure 2. In the figure, we represent the points that are more likely to be selected by max pooling as larger and red spheres. It is evident from the visualization that a + +significant number of points are not selected after the max-pooling operation, while only a small fraction of points have a high probability of being chosen. This comparison clearly shows that the max-pooling operator results in a substantial loss of object geometry information compared to the original points of the object. + +# 3.3. Height-aware Discrete Histogram for PFE + +Hence, the challenge lies in finding a pillar encoding method that addresses these problems. Our goal is to develop a pillar feature encoding approach that possesses the following two key characteristics: + +It should eliminate the requirement for sampling operations while still maintaining the ability to capture the height-aware of points within each pillar. This means that the encoding method should effectively handle the variation in point density across different height levels within a pillar. +It should keep the semantic and geometric information within the pillar, without resorting to max pooling or similar pooling operations. This is important to ensure that the encoded representation retains the relevant details and characteristics of the original pillar. + +To achieve these objectives, we propose leveraging the histogram distribution. We can effectively capture the distribution of points within a pillar at different height levels. This allows us to maintain the height-awareness of the points without the need for sampling operations. The implementation of the histogram operation is straightforward. Initially, we divide the height range of the pillar into multiple bins, where each bin corresponds to a specific range of heights. For instance, if the pillar's height spans from 0 to 10 meters, we can partition it into several bins such as $0 - 1\mathrm{m}$ , $1 - 2\mathrm{m}$ , $2 - 3\mathrm{m}$ , and so forth. Next, we proceed to count the number of points that fall within each bin for a given pillar. This counting process yields a count value for each bin, indicating the number of points residing within that particular height range. For instance, if a larger number of points are found in the lower height range bins, it signifies a higher density of points near the bottom of the pillar. Conversely, if there is a greater concentration of points in the higher height range bins, it suggests a higher density of points towards the top of the pillar. This approach provides the model with both the semantic and geometric information derived from the height dimension, thereby enhancing its understanding and representation of the point cloud data. + +![](images/3ed8bbfcf0f259111549c49cb22e3b931b993dd07c316d63c3b53e0764381a6f.jpg) +Figure 3. The overall architecture of our proposed PillarHist. The PFE (a) in PointPillars first extracts point-wise feature, and then the point-wise features are max-pooled along the point dimension to obtain the pillar-wise representation. This process has three key drawbacks: (1) it lacks detailed height-aware information, (2) the MLP operates on point-wise level, incurring computational burden, (3) the max-pooling operation leads to information loss. In contrast, our proposed PillarHist (b) directly encodes the height-aware information based on histogram, and then perform the MLP operation on the pillar level rather than point level. Finally, the max-pooling operation is also eliminated in proposed PillarHist. + +By using the histogram technique, we propose a height-aware pillar feature encoder named PillarHist. The algorithm is summarized in Algorithm 1: (1) We compute the histogram of each pillar, denoted as $H^p \in \mathbb{R}^B$ , where $B$ is the number of bins. (2) In addition, we incorporates the LiDAR reflection intensity of each point, $r_i$ , to obtain the weighted histogram of intensity, denoted as $H^i \in \mathbb{R}^B$ . The weighted histogram represents the mean LiDAR reflection intensity for each bin, providing valuable height-aware information for object detection. (3) To provide the global position of each pillar, we utilize the x and y coordinates of the pillar center, $x_{center}$ and $y_{center}$ . Since the coordinates of the pillar are constant, this step does not require any additional computational effort. (4) We concatenate the computed histograms and the pillar center coordinates, resulting in $[H^p | H^i | x_{center} | y_{center}] \in \mathbb{R}^{2B + 2}$ . (5) To represent the features of the pillars, we utilize a linear layer $f_\theta$ that projects the combined information into a feature vector. $\theta \in \mathbb{R}^{D \times (2B + 2)}$ and $D$ is the number of output channels. + +Specifically, the pillar center $x_{center}$ and $y_{center}$ provide the relative spatial relationships between different points within the point cloud. This spatial location information + +of the pillars reflects the relative position of the targets in the scene, which can help the model better understand environmental context. For instance, if there are two pillars with the same point cloud distribution located at [0,10] and [10,0] in space, without the pillar coordinate to indicate the spatial relationship, these two pillars would be indistinguishable. This would pose significant challenges for the detector's object perception capabilities. Therefore, combining pillar center $x_{center}$ and $y_{center}$ with other features such as height and intensity can further enhance the model's perception capability. This is necessary to compensate for the missing spatial position information inherent in the raw point cloud data. By incorporating the pillar center coordinates, the model can better capture the geometric structure and spatial relationships between different parts of the scene. This is a crucial aspect for enabling accurate object detection in point cloud data. + +As shown in Fig 3, our PillarHist approach explicitly captures the histogram of points in different height intervals. This allows us to keep the semantic and geometric information within the pillar. PillarHist operates only during the PFE stage without introducing complex operations + +# Algorithm 1: PillarHist Algorithm + +Input: Raw point data in a pillar $V\in \mathbb{R}^{N_v\times 4}$ consisting of $N_{v}$ points, number of height bins $B$ ,number of output channels $D$ ,height of the pillar $H$ , linear layer $f_{\theta}$ + +Output: Pillar's feature vector $Y \in \mathbb{R}^{D}$ to represent the non-empty pillar $V$ . + +1: Divide the height $H$ of the pillar into $B$ bins, each bin having height $h$ , where $h = \frac{H}{B}$ +2: Initialize $H^p \in \mathbb{R}^B$ to zero, where $H^p$ represents the histogram of points in each bin +3: Initialize $H^i\in \mathbb{R}^B$ to zero, where $H^{i}$ represents the weighted histogram of intensity in each bin +4: for $p_i = \{p_i | i = 1, 2, \dots, N_v\}$ do + +5: Compute the corresponding bin index $b_{i}$ based on the $z$ coordinate $z_{i}$ of point $p_{i}$ : $b_{i} = \lfloor \frac{z_{i}}{h} \rfloor$ + +6: $H^{p}[b_{i}]\gets H^{p}[b_{i}] + 1$ +7: $H^{i}[b_{i}]\gets H^{i}[b_{i}] + r_{i},\quad \triangleright r_{i}$ is the LiDAR + +8: reflection intensity of $p_i$ + +9: $H^{i}[b_{i}]\coloneqq H^{i}[b_{i}] / H^{p}[b_{i}]$ where $H^{i}[b_{i}]! = 0$ +10: Compute the pillar center coordinates $x_{center}, y_{center}$ +11: Concatenate: $[H^p |H^i |x_{center}|y_{center}]\in \mathbb{R}^{2B + 2}$ +12: Linear projection: $Y = f_{\theta}([H^{p}|H^{i}|x_{center}|y_{center}])$ + +or interaction with other modules, allowing for seamless inserted into existing pillar-based methods serves as a plugin module with performance improvements. + +The Efficiency of PillarHist: Here, we further analyze the efficiency of PillarHist. (1). The linear projection in PillarHist operates on the pillar level, not on individual points. In contrast, the previous approach [15, 20, 30, 45] performed the linear projection for each point within a pillar. By employing the linear projection at the pillar level, we significantly decrease the number of floating-point operations (FLOPs) required by the PFE, resulting in improved efficiency (Table 5). (2). PillarHist also avoids various memory-intensive operations, such as copying the original point cloud data, calculating distances $(x^{m},y^{m},z^{m},x^{c},y^{c},z^{c})$ , and performing max pooling. This further enhances the efficiency of PillarHist in real-world deployment. + +# 4. Experiments + +Dataset and Training Details. We evaluate our PillarHist on KITTI [11], nuScenes [1] and Waymo dataset [34]. We report the official metrics. We adopt OpenPCDet [35] as the training framework and faithfully reproduced the different baseline methods based on the provided official setting on kitti and nuScenes dataset in the OpenPCDet [35]. we directly follow the setting of the four pillar-based 3D de + +tectors. (i.e., PointPillars [15], CenterPoint-Pillar [45], PillarNet [30], FastPillars [49]). In nuScenes [1] dataset, we uniformly sample 64 frames point cloud data from the validation set as the calibration set. Calibration is a standard step in quantization, where a small subset of dataset is used to determine activation quantization parameters (scale factors and zero points). This is a common practice in model quantization [24]. The calibration set proportions is $1.06\%$ (64/6019). We set the first and the last layer of the network to keep unquantized. Specifically, the quantization methods is widely used round-to-nearest (RTN) method, a common setting in prior works, such as PD-Quant [19] and LiDAPTQ [50]. We test all experiments on a single Nvidia Tesla V100 GPU. Please refer to Appendix for more quantization details and experiment analysis. + +# 4.1. Overall Results + +Evaluation on nuScenes test set. The quantitative results on nuScenes test set are shown in Table 2. It can be seen that by integrating our proposed PillarHist into the current representative pillar-based detector PoinPillar [15], CenterPoint-Pillar [45], PillarNet [30], the proposed PH-PillarHist achieves significant performance gains on nuScenes test set across different pillar-based methods, respectively. Meanwhile, PH-PillarHist brings 6ms, 9ms, 5ms and 4ms reduction on inference latency, respectively. Notably, our method focuses on the pillar feature encoding module, rather than designing a state-of-the-art pillar-based detector. The proposed PillarHist is a plugin module to conveniently insert into existing mainstream pillar-based methods, seamlessly bringing performance gains and latency reduction. This presents the effectiveness and efficiency of PillarHist in not only boosting the detection accuracy, but also improving the computational efficiency, making it practical for real-world application. + +Evaluation on nuScenes val set. Here, we evaluate our PillarHist on different pillar-based baselines in nuScenes val set. We selected three representative pillar-based detectors: PointPillars [15], CP-Pillar [45], and PillarNet [30]. In detail, we replace the raw PFE module in each baseline model with our PillarHist, resulting in PH-PointPillars, PH-CP-Pillar, and PH-PillarNet, respectively. As shown in Table 3, the results demonstrate the effectiveness of proposed PillarHist, as we observe performance improvements of 1.7 NDS, 1.4 NDS, and 1.6 NDS on the three different baselines, respectively. These results highlight the effectiveness and broad applicability of our PillarHist in enhancing the represent power of pillar-based 3D object detectors, leading to consistent performance gains across different baseline architectures. Notably, On the nuScenes dataset, the nuScenes detection score (NDS) is the primary evaluation metric. + +The Quantization Effectiveness of PillarHist. Furthermore, we also conduct quantization experiments to evaluate + +
MethodNDSmAPLatencyCarTruckBusTrailerC.V.Ped.MotBycT.C.Bar
Focals Conv[3]70.063.8138 ms86.756.367.759.523.887.564.536.381.474.1
VoxelNeXt[4]70.064.566 ms84.653.064.755.828.785.873.245.779.074.6
FG-PFE[25]71.865.786 ms85.255.561.661.829.286.572.548.682.573.8
FastPillars [49]71.866.856 ms87.358.066.062.334.587.470.347.981.572.8
PointPillars [15]45.330.531 ms68.423.028.223.44.159.727.41.130.838.9
PH-PointPillars47.032.425 ms69.424.230.024.89.660.928.15.231.540.1
CenterPoint [45]65.558.096 ms84.651.060.253.217.583.453.728.776.770.9
PH-CenterPoint66.659.287 ms85.252.561.154.219.083.954.830.179.271.5
\( PillarNet-18[30]^\dagger \)70.865.078 ms87.456.760.961.830.487.267.440.382.176.0
\( PH-PillarNet-18^\dagger \)71.565.273 ms87.657.060.862.030.687.368.140.381.976.2
\( PillarNet-34[30]^\dagger \)71.466.096 ms87.657.563.663.127.987.370.142.383.377.2
\( PH-PillarNet-34^\dagger \)72.066.392 ms87.557.664.063.328.487.170.543.083.677.8
+ +Table 2. Performance of 3D object detection methods on nuScenes test set. ${}^{ \dagger }$ means the method that uses double-flip testing. All models listed take LIDAR data as input without image fusion or any model ensemble. + +
ModelsBits (W/A)NDSmAP
PointPillars [15]32/3252.137.4
8/844.8 (-7.3)29.8 (-7.6)
PH-PointPillars32/3253.8 (+1.7)38.8 (+1.4)
8/852.2 (-1.6)37.0 (-1.8)
CP-Pillar[45]32/3256.444.9
8/848.6 (-7.8)36.2 (-8.7)
PH-CP-Pillar32/3257.8 (+1.4)45.8 (+0.9)
8/856.3 (-1.5)44.3 (-1.5)
PillarNet[30]32/3261.452.6
8/854.6 (-6.8)43.1 (-9.5)
PH-PillarNet32/3263.0 (+1.6)52.8 (+0.2)
8/861.6 (-1.4)51.4 (-1.3)
+ +the quantization-friendly ability. Specifically, we convert the weights and activations of the models from floating-point 32-bit to integer 8-bit and applied a simple posttraining quantization setting. As shown in Table 3, we find that the PointPillars, CP-Pillar, and PillarNet models suffered significant performance degradation after weight and activation quantization, with drops of 7.3 NDS and $7.6\mathrm{mAP}$ for PointPillars, 7.8 NDS and $8.7\mathrm{mAP}$ for CP-Pillar, and 6.8 NDS and $9.5\mathrm{mAP}$ for PillarNet. In contrast, the PH-PointPillars, PH-CP-Pillar, and PH-PillarNet models that integrate PillarHist only perform performance drops of 1.6 NDS, 1.5 NDS, and 1.4 NDS, respectively. Notably, the quantized models with PillarHist integration achieved comparable performance to original floating-point counterparts without PillarHist (52.2 v.s. 52.1 NDS on PointPillars, 56.3 v.s. 56.4 NDS on CenterPoint-Pillars and 61.6 v.s. 61.4 NDS on PillarNet). Specifically, PillarHist could effectively eliminate the severe performance degradation after quantization caused by other PFE design, bringing quantization performance gain seamlessly. These results demonstrate the quantization-friendly nature of PillarHist, which maintains + +the model's representation power and performance even under quantization compression. This property is particularly valuable and practicable for deploying high-performance 3D object detectors on resource-constrained edge devices, where accuracy and latency are critical factors. + +Table 3. Generalization and quantization performance comparison on nuScene val set. + +
MethodsMean (L2) mAPHVehicle (L2) mAPmAPHPed. (L2) mAPmAPHCyc. (L2) mAPmAPH
CP-Pillar55.962.562.061.851.556.154.3
PH-CP-Pillar57.363.262.863.854.057.055.3
↑+1.4+0.7+0.8+2.0+2.5+0.9+1.0
+ +Generalization of PillarHist on Waymo val set. Here, we also evaluate our PillarHist on Waymo dataset. We select CenterPoint-Pillars as our baseline and only replaced their pillar feature encoding module with our PillarHist for comparison, as show in table our method can bring $1.2\mathrm{mAP}$ L2 and $1.5\mathrm{mAPH}$ L2 performance improvements. Notably, it can achieve $2.0\mathrm{mAP}$ L2 and $2.5\mathrm{mAPH}$ L2 performance gains in the pedestrian category. This result demonstrates the effectiveness of our method. + +# 4.2. Ablation Study + +Table 4. Apply PillarHist to CeterPoint-Pillars on Waymo val set trained with $20\%$ training data. + +
PFE moduleCar 3D R40Ped 3D R40PFE GFLOPs
mAPEasyMod.HardmAPEasyMod.Hard
PFE[15]79.9887.0877.9074.9749.4154.7149.0144.520.152
PAA[16]81.0488.6279.2475.2748.5654.2647.9343.493.391
TA[20]80.7187.3379.1175.6949.4554.0049.3145.057.061
DynPFE*[30]79.9687.6377.7074.5549.3654.2749.0144.810.013
DynPFE[45]80.4687.8678.1675.3847.7752.2847.5343.520.112
FG-PFE[25]----50.7556.8349.8845.561.136
MAPE[49]81.1989.3978.5775.6348.1852.6848.0243.830.165
PillarHist81.4288.8079.1376.3051.0757.4550.4245.360.065
+ +Table 5. Performance of different PFE on the KITTI val set. + +Comparison of Different PFE modules. We compare the computation efficiency and performance of existing PFE modules. Specifically, we evaluate the following PFE modules: PFE in PointPillars [15], DynPFE in CenterPoint-Pillar [45], TA module in TANet [20], PAA module in + +PiFeNet [16], DynPFE* in PillarNet [30], and MAPE in FastPillars [49]. To ensure a fair comparison, we select PointPillars as the baseline method and conduct experiments by replacing the PFE in PointPillars with other PFE module (DynPFE, TA, PAA, DynpFE*, MAPE and our PillarHist). This setting eliminates the performance difference caused by different network design. As shown in Table 5, our PillarHist achieves the highest 3D AP compared to other PFE methods while maintaining superior computational efficiency. Notably, although DynPFE* used in PillarNet [30] has lower GFLOPs, it suffers from poor detection performance. Notably, our PillarHist method performs linear projection at the pillar-wise level rather than point-wise level in other PFE. This design can efficiently reduce the computational cost while maintaining the necessary structural information within pillar without max-pooling operations. + +
heightintensitycoordsCar 3D40Ped 3D40
mAPEasyMod.HardmAPEasyMod.Hard
✓80.1887.4778.1875.2248.0653.7547.5142.92
✓✓80.8188.2978.6275.5349.6255.4249.1044.36
✓✓✓81.4188.8079.1376.3051.0757.4550.4245.36
+ +Table 6. Ablation results of different input for PillarHist on the KITTI val set for car and pedestrian. + +The Effects of Different Input in PillarHist. Here, we conduct ablation study to analyze the impact of different input. As shown in Table 6, when we incorporate the intensity information, it leads to performance improvements for both the vehicle and pedestrian categories. Particularly for the pedestrian, approximately 1.7, 1.6, and 1.6 points for the easy, moderate, and hard levels, respectively. This notable improvement indicates that for small objects like pedestrians, the intensity values are crucial in distinguishing them from the background and enhancing their distinctive features. Furthermore, by adding the pillar coordinates information, there is a performance improvement of 0.6 and 1.5 points for the vehicle and pedestrian. We believe this improvement is because explicitly incorporating the global coordinate of the pillars helps the model better identify the object positions with the enhanced encoding information that includes height awareness. Notably, after adding the pillar coordinates, there is an improvement of 2.03 points in easy AP for pedestrians. This can be attributed that the pillars encode not only varying height distribution but also including the difference between empty and non-empty pillars. + +Feature Map Visualization of PillarHist. Here, as shown in Figure 4, we apply our PillarHist method to PointPillars [15] (a) and compare the features from the max-pooling operation in their PFE. We find that the max pooling operation tends to focus more on the points at higher positions within the same pillar in BEV (with larger coordinates), resulting in a biased feature representation. In contrast, our PillarHist can focus more evenly on each point within + +![](images/4e1b6e0bce78164ffc54c7e0885a4ea68909a150b5dcf8091594c6a283ed8b27.jpg) +Figure 4. (a) Visualization of the learned attention scores in PillarHist on different scenes. The color of points represents their scores. The point will be paid more attention if it has a higher score. (b) Visualization of the learned BEV feature in PointPillars and PH-PointPillars. Best viewed in color. + +the pillar, encoding the height information into an unbiased histogram distribution, thereby obtaining richer geometric-semantic feature representation. This difference is particularly obvious at the car door location, the points within the pillar is typically uniformly distributed along height; however, max pooling emphasizes the higher contours, whereas PillarHist provides a better representation of the complete geometry of the car door Besides, as shown in Figure 4 (b), we also visualize the BEV features maps. Our method can obtain a more informative BEV feature representation compared to the PFE in PointPillars [15]. This also demonstrates the underlying reason for the better performance of our method. + +# 5. Conclusion + +In this paper, we addressed the issue of information loss and large numerical distribution difference in existing pillar-based 3D object detectors during the PFE stage. We identify the height dimension as a crucial factor for enhancing 3D detection performance and propose a height-aware pillar feature encoder called PillarHist. By leveraging histogram, PillarHist preserve the information along the height dimension while effectively mitigating the information loss introduced by max-pooling operations. Importantly, PillarHist operated exclusively within the PFE stage, allowing seamless insert into existing pillar-based methods serves as a plug-in module. Meanwhile, PillarHist also constrains the arithmetic distribution of PFE input to a stable range, making it quantization-friendly. Extensive experiments demonstrated the effectiveness of PillarHist in terms of both efficiency and performance. Our proposed method achieved an average improvement of approximately $1.0\%$ NDS across various baselines, showcasing its ability to enhance 3D object detection accuracy and computation efficiency. + +# Acknowledgments + +We thank anonymous reviewers for their kind help of this work. This work was supported by National Natural Science Foundation of China (No.62271143), Frontier Technologies RD Program of Jiangsu (No. BF2024060) and the Big Data Computing Center of Southeast University. + +# References + +[1] Holger Caesar, Varun Bankiti, Alex Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. Nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 2, 6 +[2] Qi Cai, Yingwei Pan, Ting Yao, and Tao Mei. 3d cascade rconn: High quality object detection in point clouds. IEEE Transactions on Image Processing, 31:5706-5719, 2022. 1 +[3] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 7 +[4] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 7 +[5] Zehui Chen, Zheng Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, Feng Wu, and Feng Zhao. Graph-detr4d: Spatio-temporal graph modeling for multiview 3d object detection. IEEE Transactions on Image Processing, 33:4488-4500, 2024. 1 +[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3D object detection. In AAAI, 2021. 1 +[7] Xiaoying Ding, Weisi Lin, Zhenzhong Chen, and Xinfeng Zhang. Point cloud saliency detection by local and global feature fusion. IEEE Transactions on Image Processing, 28 (11):5379-5393, 2019. 1 +[8] Lue Fan, Feng Wang, Naiyan Wang, and ZHAO-XIANG ZHANG. Fully sparse 3d object detection. Advances in Neural Information Processing Systems, 35:351-363, 2022. 1 +[9] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3d-siamrpn: An end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 1 +[10] Jiahui Fu, Guanghui Ren, Yunpeng Chen, and Si Liu. Improved pillar with fine-grained feature for 3d object detection. arXiv preprint arXiv:2110.06049, 2021. 3 +[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 6 +[12] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 3 + +[13] Shan Jiayao, Sifan Zhou, Yubo Cui, and Zheng Fang. Realtime 3d single object tracking with transformer. IEEE Transactions on Multimedia, 2022. 1 +[14] Hongwu Kuang, Bei Wang, Jianping An, Ming Zhang, and Zehan Zhang. Voxel-fpn: Multi-scale voxel feature aggregation for 3d object detection from lidar point clouds. Sensors, 20(3):704, 2020. 1 +[15] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 1, 3, 4, 6, 7, 8 +[16] Duy Tho Le, Hengcan Shi, Hamid Rezatofighi, and Jianfei Cai. Accurate and real-time 3d pedestrian detection using an efficient attentive pillar network. IEEE Robotics and Automation Letters, 8(2):1159-1166, 2022. 3, 7, 8 +[17] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3 +[18] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv preprint arXiv:2102.05426, 2021. 3 +[19] Jiawei Liu, Lin Niu, Zhihang Yuan, Dawei Yang, Xinggang Wang, and Wenyu Liu. Pd-quant: Post-training quantization based on prediction difference metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24427–24437, 2023. 3, 6 +[20] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In AAAI, pages 11677-11684, 2020. 3, 6, 7 +[21] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In ICCV, pages 3164-3173, 2021. 1 +[22] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, pages 1-55, 2023. 3 +[23] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 3 +[24] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yely-sei Bondarenko, Mart Van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 6 +[25] Konyul Park, Yechel Kim, Junho Koh, Byungwoo Park, and Jun Won Choi. Fine-grained pillar feature encoding via spatio-temporal virtual grid for 3d object detection. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4259-4265, 2024. 7 +[26] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, pages 652-660, 2017. 1, 3 + +[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1 +[28] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1 +[29] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. Ptt: Point-track-transformer module for 3d single object tracking in point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 1 +[30] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: Real-time and high-performance pillar-based 3d object detection. In ECCV, 2022. 1, 3, 4, 6, 7, 8 +[31] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE T-PAMI, 2020. 1, 3 +[32] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3D object detection in a point cloud. In CVPR, pages 1711-1719, 2020. 1 +[33] Lukas Stäcker, Juncong Fei, Philipp Heidenreich, Frank Bonarens, Jason Rambach, Didier Stricker, and Christoph Stiller. Deployment of deep neural networks for object detection on edge ai devices with runtime optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshop, pages 1015-1022, 2021. 3 +[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 6 +[35] OpenPCDet Development Team. OpenPCDet: An opensource toolbox for 3D object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.6 +[36] Yonglin Tian, Lichao Huang, Hui Yu, Xiangbin Wu, Xuesong Li, Kunfeng Wang, Zilei Wang, and Fei-Yue Wang. Context-aware dynamic feature extraction for 3d object detection in point clouds. IEEE Transactions on Intelligent Transportation Systems, 23(8):10773-10785, 2021. 3 +[37] Zhi Tian, Xiangxiang Chu, Xiaoming Wang, Xiaolin Wei, and Chunhua Shen. Fully convolutional one-stage 3d object detection on lidar range images. NeurIPS, 2022. 1 +[38] Tai Wang, Xinge Zhu, and Dahua Lin. Reconfigurable voxels: A new representation for lidar-based point clouds. arXiv preprint arXiv:2004.02724, 2020. 3 +[39] Yue Wang and Justin M Solomon. Object dgCNN: 3d object detection using dynamic graphs. NeurIPS, 2021. 3 +[40] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization. arXiv preprint arXiv:2203.05740, 2022. 3 +[41] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1 + +[42] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3D single stage object detector. In CVPR, pages 11040-11048, 2020. 1 +[43] Maosheng Ye, Shuangjie Xu, and Tongyi Cao. Hvnet: Hybrid voxel network for lidar based 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1631-1640, 2020. 3 +[44] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Centerpoint++ submission to the waymo real-time 3d detection challenge. 2021. Accessed: 2021-12-05. 1 +[45] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 1, 3, 4, 6, 7 +[46] Jiangyong Yu, Changyong Shu, Dawei Yang, Sifan Zhou, Zichen Yu, Xing Hu, and Yan Chen. Q-petr: Quant-aware position embedding transformation for multi-view 3d object detection. arXiv e-prints, pages arXiv-2502, 2025. 3 +[47] Yifan Zhang, Zhen Dong, Huanrui Yang, Ming Lu, Cheng-Ching Tseng, Yandong Guo, Kurt Keutzer, Li Du, and Shanghang Zhang. Qd-bev: Quantization-aware view-guided distillation for multi-view 3d object detection. 2023. 3 +[48] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In AAAI, 2021. 3 +[49] Sifan Zhou, Zhi Tian, Xiangxiang Chu, Xinyu Zhang, Bo Zhang, Xiaobo Lu, Chengjian Feng, Zequn Jie, Patrick Yin Chiang, and Lin Ma. Fastpillars: A deployment-friendly pillar-based 3d detector. arXiv preprint arXiv:2302.02367, 2023. 1, 3, 4, 6, 7, 8 +[50] Sifan Zhou, Liang Li, Xinyu Zhang, Bo Zhang, Shipeng Bai, Miao Sun, Ziyu Zhao, Xiaobo Lu, and Xiangxiang Chu. Lidar-ptq: Post-training quantization for point cloud 3d object detection. International Conference on Learning Representations (ICLR), 2024. 1, 3, 6 +[51] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3D object detection. In CVPR, pages 4490-4499, 2018. 1, 3 +[52] Yin Zhou, Pei Sun, Yu Zhang, Dragomir Anguelov, Jiyang Gao, Tom Ouyang, James Guo, Jiquan Ngiam, and Vijay Vasudevan. End-to-end multi-view fusion for 3d object detection in lidar point clouds. In Conference on Robot Learning, pages 923-932. PMLR, 2020. 3 +[53] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv:1908.09492, 2019. 3 \ No newline at end of file diff --git a/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/images.zip b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..52d2b19965c7ef78184ff533711c5ddd32e71e29 --- /dev/null +++ b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b72e73e274ae04fb487bf524b110fe38017c53571f9b9498c38007d54e3599cf +size 552342 diff --git a/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/layout.json b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e285385e692ce69c00b59b9786f1de8538edb2e9 --- /dev/null +++ b/2025/PillarHist_ A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram/layout.json @@ -0,0 +1,8895 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 481, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 481, + 138 + ], + "type": "text", + "content": "PillarHist: A Quantization-aware Pillar Feature Encoder based on Height-aware Histogram" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "spans": [ + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "content": "Sifan Zhou" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{1\\dagger*}" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "content": " Zhihang Yuan" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "content": " Dawei Yang" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "content": " Xing Hu" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "content": " Jian Qian" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "content": " Ziyu Zhao" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 189, + 414, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 189, + 414, + 217 + ], + "spans": [ + { + "bbox": [ + 196, + 189, + 414, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 196, + 189, + 414, + 217 + ], + "type": "text", + "content": "School of Automation, Southeast University " + }, + { + "bbox": [ + 196, + 189, + 414, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 196, + 189, + 414, + 217 + ], + "type": "text", + "content": "Houmo AI " + }, + { + "bbox": [ + 196, + 189, + 414, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 196, + 189, + 414, + 217 + ], + "type": "text", + "content": "Fudan University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 220, + 512, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 220, + 512, + 231 + ], + "spans": [ + { + "bbox": [ + 91, + 220, + 512, + 231 + ], + "type": "text", + "content": "{sifanjay, hahnyuan, zhaoziyu.950207}@gmail.com,{dawei.yang, xing.hu}@houmo.ai" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 283, + 297, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 297, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 297, + 620 + ], + "type": "text", + "content": "Real-time and high-performance 3D object detection plays a critical role in autonomous driving and robotics. Recent pillar-based 3D object detectors have gained significant attention due to their compact representation and low computational overhead, making them suitable for onboard deployment and quantization. However, existing pillar-based detectors still suffer from information loss along height dimension and large numerical distribution difference during pillar feature encoding (PFE), which severely limits their performance and quantization potential. To address above issue, we first unveil the importance of different input information during PFE and identify the height dimension as a key factor in enhancing 3D detection performance. Motivated by this observation, we propose a height-aware pillar feature encoder, called PillarHist. Specifically, PillarHist statistics the discrete distribution of points at different heights within one pillar with the information entropy guidance. This simple yet effective design greatly preserves the information along the height dimension while significantly reducing the computation overhead of the PFE. Meanwhile, PillarHist also constrains the arithmetic distribution of PFE input to a stable range, making it quantization-friendly. Notably, PillarHist operates exclusively within the PFE stage to enhance performance, enabling seamless integration into existing pillar-based methods without introducing complex operations. Extensive experiments show the effectiveness of PillarHist in terms of both efficiency and performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 630, + 135, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 135, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 135, + 642 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "text", + "content": "LiDAR-based 3D object detection plays a pivotal role in various applications [5, 7, 9, 13, 29, 37, 50], particularly in autonomous driving and robotics. It is important to detect" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 260, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 536 + ], + "type": "text", + "content": "the objects in the surrounding environment fastly and accurately, which places a high demand for both performance and latency. Existing mainstream methods can be categorized into two types based on the point cloud representation: (1) point-based [8, 28, 31, 32, 42] and (2) grid-based models [2, 6, 14, 15, 21, 41, 41, 51]. Point-based models generally inherit the success in point cloud representation learning using deep neural networks[26, 44] and detect objects directly from raw points. However, these methods often face challenges in real-time onboard deployment as they often require point query/retrieval in 3D space (e.g., PointNet++ [27]). Grid-based methods rasterize the points into regular grid (voxel/pillar) and employ 2D or 3D convolution for feature learning. Compared to voxel-based methods, pillar-based detectors exhibit a more compact grid representation and can seamlessly integrate advanced 2D detection techniques without much modification. Therefore, those methods offer a favorable trade-off between performance and speed, making them highly suitable for practical deployment in autonomous systems. Furthermore, it is equally crucial to deploy them on resource-limited edge devices by improve their efficiency (e.g., reduce memory and computation cost)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "The Pillar Feature Encoding (PFE) module plays a huge role in pillar-based networks [15, 17, 30, 45, 49]. Its main purpose is to transform the input point clouds, which are initially unordered (x, y, z, intensity) information, into pillars with a regular spatial arrangement. Each pillar has a fixed length of feature dimensions. In this paper, we observe that existing pillar-based methods suffer from two key problem in PFE module. 1 Information loss significantly hampers their performance. 2 Encoding scheme leads to a large variation of activations, making quantization challenging. As shown in Figure 1, the problem 1 can be attributed to two factors: (1) Lack of detailed height-aware feature encoding. The height information is not well handled, resulting in an incomplete representation of the 3D scene. It is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*Work done as an intern at Houmo AI. †Equal Contribution. ‡Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27336" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 87, + 498, + 162 + ], + "blocks": [ + { + "bbox": [ + 230, + 74, + 379, + 85 + ], + "lines": [ + { + "bbox": [ + 230, + 74, + 379, + 85 + ], + "spans": [ + { + "bbox": [ + 230, + 74, + 379, + 85 + ], + "type": "text", + "content": "(1) Lack of height-aware feature encoding" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 111, + 87, + 498, + 162 + ], + "lines": [ + { + "bbox": [ + 111, + 87, + 498, + 162 + ], + "spans": [ + { + "bbox": [ + 111, + 87, + 498, + 162 + ], + "type": "image", + "image_path": "b5686df97058c02679acbf77ed3a8da0e002a5b124a2e9e06889d5431749e19c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 109, + 176, + 499, + 260 + ], + "blocks": [ + { + "bbox": [ + 222, + 165, + 389, + 175 + ], + "lines": [ + { + "bbox": [ + 222, + 165, + 389, + 175 + ], + "spans": [ + { + "bbox": [ + 222, + 165, + 389, + 175 + ], + "type": "text", + "content": "(2) Information loss with max-pooling operator" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 109, + 176, + 499, + 260 + ], + "lines": [ + { + "bbox": [ + 109, + 176, + 499, + 260 + ], + "spans": [ + { + "bbox": [ + 109, + 176, + 499, + 260 + ], + "type": "image", + "image_path": "0a58559b08e8eb8e5efcd382426f84b8acce917b642f673fab80657b303da536.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 273, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 555, + 328 + ], + "type": "text", + "content": "Figure 1. The drawbacks of pillar feature encoding are as follows: (1) Compared with the voxel feature encoding, the encoding process of pillars lacks height awareness, as it does not explicitly model the height dimension of the input point cloud. (2) PFE employs element-wise max-pooling operation, which inevitably results in information loss. As illustrated in the right side of (2), the max-pooling operation tends to preserve the feature information of points with higher height values within a pillar while disregarding points with lower height values, resulting in the loss of geometric information within the pillar." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 338, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 422 + ], + "type": "text", + "content": "difficult to distinguish the height differences in various situations, like a pedestrian and a person standing on a car. (2) Max-pooling operator causes the loss of height information. The adoption of max-pooling operations within each pillar to derive pillar features leads to a loss of detailed information, which limits the ability of the encoded features to distinguish different objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "text", + "content": "The problem 2 can be attributed the various arithmetic range of the PFE input. For instance, in nuScenes [1] dataset, the input of PFE includes " + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\hat{p}_i = \\left\\{[x_i,y_i,z_i,r_i,x_i^m,y_i^m,z_i^m,x_i^c,y_i^c,z_i^c]\\in \\mathbb{R}^{N_v\\times 10}\\right\\}" + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "inline_equation", + "content": "[x_i,y_i,z_i]\\in [-51.2,51.2]" + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "text", + "content": " is the point coordinates, " + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "inline_equation", + "content": "r_i\\in [0,255]" + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "text", + "content": " is the intensity, " + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "inline_equation", + "content": "[x_i^m,y_i^m,z_i^m]\\in [-0.2,0.2]" + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "text", + "content": " (pillar size " + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "inline_equation", + "content": "= 0.2\\mathrm{m}" + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "text", + "content": ") denotes the offset of " + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "inline_equation", + "content": "[x_i,y_i,z_i]" + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "text", + "content": " and the mean value of " + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "inline_equation", + "content": "[x_i,y_i,z_i]" + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "text", + "content": " within one pillar, and " + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "inline_equation", + "content": "[x_i^c,y_i^c,z_i^c]\\in [-4.0,4.0]" + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "text", + "content": " is the spatial offset of " + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "inline_equation", + "content": "[x_i,y_i,z_i]" + }, + { + "bbox": [ + 55, + 424, + 295, + 592 + ], + "type": "text", + "content": " and the pillar center. Details refer to Sec 3.1. The different part of input represents the global- and local-contextual information, but exhibit significant difference in numerical distribution. Therefore, the arithmetic disparity makes it more susceptible to the effects of quantization (e.g., clipping error)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "To address the above issues, we propose a novel quantization-aware pillar feature encoder called PillarHist. It leverages a histogram technique to encode the height-aware information of each pillar into multiple bins, and the number of points falling within each bin is counted. The resulting histogram representation provides both semantic and geometric information from the height dimension. PillarHist incorporates the LiDAR reflection intensity of each point to obtain a weighted histogram of intensity, which adds distinguished height-aware information for object de" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 338, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 494 + ], + "type": "text", + "content": "tection. These histograms, along with the pillar center coordinates, are concatenated and passed through a linear layer to obtain a feature vector representing the pillar. Specially, PillarHist constrains the arithmetic distribution of PFE input to a stable range, making it friendly for further quantization. Besides, by operating linear projection at the pillar level instead of individual points, PillarHist significantly reduces the computation overhead, resulting in improved efficiency. It can be seamlessly integrated into existing pillar-based methods as a plugin module. Extensive experiments demonstrate the effectiveness of our proposed PillarHist in terms of both efficiency and performance. The contributions of our method can be summarised as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 513, + 553, + 717 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 513, + 553, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 553, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 553, + 561 + ], + "type": "text", + "content": "- We conduct a comprehensive analysis of existing Pillar Feature Encoding (PFE). Our findings indicate that they suffer from information loss in the height dimension, and exits large variation of activation for model quantization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 562, + 553, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 553, + 621 + ], + "type": "text", + "content": "- We propose a novel PFE module called PillarHist, which can mitigate the height information loss and reduce the computation cost of PFE. It is noteworthy that we make it a plugin module to conveniently insert into existing pillar-based methods while seamlessly boosting performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 621, + 553, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 553, + 669 + ], + "type": "text", + "content": "- We explore the quantization-friendly properties of PillarHist. PillarHist could effectively eliminate the severe performance degradation after quantization caused by the raw PFE, bringing performance gain effortlessly." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 670, + 553, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 553, + 717 + ], + "type": "text", + "content": "- We conducted extensive experiments to evaluate the effectiveness of PillarHist. Our results show that it achieved an average improvement of approximately 1 point in different dataset without extra latency." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27337" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 91, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 91, + 296, + 415 + ], + "type": "text", + "content": "Grid Feature Encoding. Grid feature encoding is crucial for grid-based 3D object detectors, and significantly impacts model accuracy[12, 22]. Various methods have been proposed to extract features from two grid representations: Voxel and Pillar. VoxelNet [51] is a pioneering work that proposes a novel voxel feature encoding (VFE) layer, influencing many subsequent studies [31, 36, 38, 39, 43, 48, 52, 53]. Pillars serve as special voxels without vertical division; PointPillars [15] utilizes PointNet [26] for features extraction and employs max-pooling to obtain pillar features, leading to its widespread adoption in later works [10, 16, 17, 20, 30, 45, 49]. TANet [20] introduces a Triple Attention (TA) module to capture spatial correlations, but its reliance on point-wise operation and max-pooling incurs high computational costs and limits pillar representations. PiFeNet [16] proposes a lighter Pillar Aware Attention (PAA) module using channel-wise and point-wise attention, yet still employs pooling, affecting inference speed. Recently, PillarNet [30] introduced a dynamic PFE module with less information loss, while Fast-Pillars [49] presented a MAPE module for finer-grained geometric preservation. However, these methods do not adequately account for height-aware information and primarily use element-wise max-pooling or similar pooling operations, leading to further information loss. This paper focuses on enhancing detailed height-aware information and mitigate the information loss caused by pooling operations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 419, + 296, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 587 + ], + "type": "text", + "content": "Quantization for 3D Object Detection. With the wide application of 3D object detection in autonomous driving and robotics, some quantization methods [46, 47] are designed to improve inference speed for onboard applications. In LiDAR-based 3D detection, [33] find that directly using INT8 quantization for 2D CNN significantly degrades performance on PointPillars [15], with an even greater reduction when using an entropy calibrator. Recently, LiDAR-PTQ [50] unveils the cause of performance collapse and propose an effective post-training quantization (PTQ) method. However, advanced quantization techniques[18, 19, 23, 40] are still needed to optimize quantization parameter within LiDAR-PTQ. Our PillarHist aim to make the network with PillarHist are quantization-friendly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 592, + 138, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 138, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 138, + 605 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 612, + 135, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 135, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 135, + 624 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "LiDAR-based 3D Object Detection. Here, we present the basic task definitions of LiDAR-based 3D object detection. Given a point set with " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " points in the 3D space, which is defined as " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\mathbf{p}_i = [x_i,y_i,z_i,r_i]^T\\in \\mathbb{R}^{N\\times 4}\\}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_{i},y_{i},z_{i}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " denote the coordinate values of each point along the axes X, Y, Z, respectively, and " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " is the LiDAR reflection intensity. Given a set of objects in the 3D scene " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{B} = \\{\\mathbf{b}_j =" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "inline_equation", + "content": "[x_j, y_j, z_j, h_j, w_j, l_j, r_j, c_j]^T \\in \\mathbb{R}^{M \\times 8}" + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": " is the total number of objects, " + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": "-th object in the scene, " + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "inline_equation", + "content": "x_j, y_j, z_j" + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": " is the object's center, " + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "inline_equation", + "content": "h_j, w_j, l_j" + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": " is the object's size, " + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "inline_equation", + "content": "r_j" + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": " is the object's heading angle and " + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": " is the object's class. 3D object detection aims to predict the 3D bounding boxes " + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": " from the point cloud " + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": " in driving scenarios accurately." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": "Point Cloud Pillarization. We aim to transform the point cloud " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": " into a regular 3D pillar representation. Assuming that " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": " is in the 3D space with the range of " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "L, W, H" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": " along the x, y, z axes, respectively, The point cloud " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": " is uniformly partitioned into a specific pillar grid of size " + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "inline_equation", + "content": "(\\delta_x, \\delta_y, \\mathrm{H})" + }, + { + "bbox": [ + 313, + 152, + 555, + 224 + ], + "type": "text", + "content": ", we can compute the no-empty pillar index using:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 229, + 555, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 229, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 319, + 229, + 555, + 255 + ], + "type": "interline_equation", + "content": "\\left(x _ {i n t}, y _ {i n t}, z _ {i n t}\\right) = \\left(\\lfloor \\frac {x _ {i}}{\\delta_ {x}} \\rceil , \\lfloor \\frac {y _ {i}}{\\delta_ {y}} \\rceil , \\lfloor \\frac {H}{2} \\rceil\\right), i = 0 \\dots N - 1 \\tag {1}", + "image_path": "58394f2fa7180a5af434e8514e932d19419b209f8eeb1a4a6f52bc9b03362d5a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\lfloor \\cdot \\rfloor" + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "text", + "content": " is the floor operation, " + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "inline_equation", + "content": "x_{int},y_{int},z_{int}" + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "text", + "content": " is the 3D pillar index of " + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "text", + "content": " after pillarization. Notably, point cloud " + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "text", + "content": " is only pillarized along the XY plane without the height dimension. Let " + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "inline_equation", + "content": "V = \\{v_{i} = [x_{i},y_{i},z_{i},r_{i}]\\in \\mathbb{R}^{N_{v}\\times 4}\\} ,i\\in" + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\{1,\\dots,N_v\\}" + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "text", + "content": " , be a non-empty pillar contains " + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "text", + "content": " points, " + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "text", + "content": " is number of points in pillar " + }, + { + "bbox": [ + 313, + 260, + 555, + 333 + ], + "type": "inline_equation", + "content": "V" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "text", + "content": "Pillar Feature Encoding. Pillar Feature Encoding (PFE) plays a crucial role in pillar-based 3D detection methods[15, 30, 49], directly impacting the representation learning of BEV feature and final detection accuracy. Here, taking PointPillars[15] as an example, we illustrate the details of PFE. The points in each pillar are initially concatenated along the channel dimension to form " + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\hat{p}_i = \\{[x_i,y_i,z_i,r_i,x_i^m,y_i^m,z_i^m,x_i^c,y_i^c,z_i^c]\\in \\mathbb{R}^{N_v\\times 10}\\}" + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "inline_equation", + "content": "[x_i,y_i,z_i]" + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "text", + "content": " is the original point coordinates in the ego frame, " + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "inline_equation", + "content": "[x_i^m,y_i^m,z_i^m]" + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "text", + "content": " denotes the offset of " + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "text", + "content": " from the mean of the xyz coordinates of all points within the current pillar, and " + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "inline_equation", + "content": "[x_i^c,y_i^c,z_i^c]" + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "text", + "content": " is the offset of " + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "text", + "content": " from the current pillar center. Notably, for each pillar, a sampling or padding strategy is employed to ensure that all pillars contain the same number of points. The concatenated point-wise features " + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "text", + "content": " within a non-empty pillar " + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 343, + 556, + 559 + ], + "type": "text", + "content": " are mapped to the high-dimensional feature space using MLP layers. This process can be mathematically formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 399, + 566, + 553, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 566, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 399, + 566, + 553, + 578 + ], + "type": "interline_equation", + "content": "p _ {i} ^ {c} = m \\left(\\hat {p} _ {i}; w _ {m}\\right) \\tag {2}", + "image_path": "4c1549916a60308ee8b07a0cca5022575c4228cb9a5d299aa217be1d062b1c78.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 585, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 554, + 643 + ], + "type": "text", + "content": "where the function " + }, + { + "bbox": [ + 313, + 585, + 554, + 643 + ], + "type": "inline_equation", + "content": "m(\\cdot)" + }, + { + "bbox": [ + 313, + 585, + 554, + 643 + ], + "type": "text", + "content": " denotes MLP, " + }, + { + "bbox": [ + 313, + 585, + 554, + 643 + ], + "type": "inline_equation", + "content": "w_{m}" + }, + { + "bbox": [ + 313, + 585, + 554, + 643 + ], + "type": "text", + "content": " denotes the learnable weights of " + }, + { + "bbox": [ + 313, + 585, + 554, + 643 + ], + "type": "inline_equation", + "content": "m(\\cdot)" + }, + { + "bbox": [ + 313, + 585, + 554, + 643 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 585, + 554, + 643 + ], + "type": "inline_equation", + "content": "p_i^c \\in \\mathbb{R}^{N_v \\times D}" + }, + { + "bbox": [ + 313, + 585, + 554, + 643 + ], + "type": "text", + "content": " is the pointwise feature. 3 An element-wise max-pooling operation is applied along the dimension of point features, which is formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 400, + 643, + 553, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 643, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 400, + 643, + 553, + 656 + ], + "type": "interline_equation", + "content": "f ^ {m} = \\max \\left(p _ {i} ^ {c}\\right) \\tag {3}", + "image_path": "fab105b244f6c89fb36298e0006f21ba175f37e848441b132fc9a1c5979f4540.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\max(\\cdot)" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " means the max-pooling operator, and " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "f^m \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " is the resulting feature vector to represent non-empty pillar " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ". A non-empty pillar " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{N_v \\times 4}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " points is encoded into a feature vector " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "f^m \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27338" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 104, + 70, + 503, + 149 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 503, + 149 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 503, + 149 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 503, + 149 + ], + "type": "table", + "html": "
(xi, xi^m, xi^c)(yi, yi^m, yi^c)(zi, zi^m, zi^c)3D_{R40}3D_{R11}
mAPEasyMod.HardmAPEasyMod.Hard
✓57.5067.9853.3051.2458.2067.9453.8252.85
✓59.4069.8555.4552.8959.9669.4156.1754.31
✓79.6386.6877.6274.5979.1586.6776.4974.28
PointPillars79.9887.0877.9074.9779.4386.2576.7675.02
", + "image_path": "3478325fcc75040ddcd29a5d5934b5aca1cfa1743e0c7ff2f5abb81ae5ed28b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 153, + 535, + 164 + ], + "lines": [ + { + "bbox": [ + 74, + 153, + 535, + 164 + ], + "spans": [ + { + "bbox": [ + 74, + 153, + 535, + 164 + ], + "type": "text", + "content": "Table 1. Ablation results of different input for PFE in PointPillars on the KITTI val set in " + }, + { + "bbox": [ + 74, + 153, + 535, + 164 + ], + "type": "inline_equation", + "content": "{\\mathrm{{AP}}}_{3D}" + }, + { + "bbox": [ + 74, + 153, + 535, + 164 + ], + "type": "text", + "content": " (R40) and " + }, + { + "bbox": [ + 74, + 153, + 535, + 164 + ], + "type": "inline_equation", + "content": "{\\mathrm{{AP}}}_{3D}" + }, + { + "bbox": [ + 74, + 153, + 535, + 164 + ], + "type": "text", + "content": " (R11) for car." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 178, + 294, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 294, + 192 + ], + "type": "text", + "content": "3.2. What Kind of Information Do We Need in PFE?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "text", + "content": "(1) Height Information is More Important. As aforementioned, the pioneering pillar-based method PointPillars[15] concatenates two extra relation information along the channel dimension of the raw point cloud, i.e. " + }, + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\left([x_i^m, y_i^m, z_i^m]\\right)" + }, + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "inline_equation", + "content": "\\left([x_i^c, y_i^c, z_i^c]\\right)" + }, + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "text", + "content": ". Here, we decouple the input information along the " + }, + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 197, + 295, + 281 + ], + "type": "text", + "content": " axes, respectively. aiming to explore their impact on the final 3D detection performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": "As shown in Table 1, when we only input the information along the " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": " axis " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "(x_{i}, x_{i}^{m}, x_{i}^{c})" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": " or the information along the " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": " axis " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "(y_{i}, y_{i}^{m}, y_{i}^{c})" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": ", there is a significant degradation in the final 3D detection performance. The former results in a large decrease of " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "22.48 \\mathrm{~mAP}" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "3D_{R40}" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "21.23 \\mathrm{~mAP}" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "3D_{R11}" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": ", while the latter leads to a large decrease of " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "20.58 \\mathrm{~mAP}" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "3D_{R40}" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "19.47 \\mathrm{~mAP}" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "3D_{R11}" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": ". However, when we only input the information along the " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": " axis " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "(z_{i}, z_{i}^{m}, z_{i}^{c})" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": ", it also achieves comparable performance with the baseline. This indicates that in the pillar feature encoding process, the information along the " + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 281, + 296, + 426 + ], + "type": "text", + "content": " axis, a.k.a the height dimension, plays a key role in the final 3D detection performance." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 430, + 294, + 498 + ], + "blocks": [ + { + "bbox": [ + 58, + 430, + 294, + 498 + ], + "lines": [ + { + "bbox": [ + 58, + 430, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 58, + 430, + 294, + 498 + ], + "type": "image", + "image_path": "2183e81d51230276621f5f7cda7cdcfcccd9335f3ef2b6f51e7c1da76449a5ce.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 507, + 295, + 574 + ], + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 574 + ], + "type": "text", + "content": "Figure 2. Visualization of the weights on (a) vehicle and (b) cyclist using Max-pooling in PFE. The size of points means scores. The point will be paid more attention if it has a higher score. Compared with original object's points, max-pooling operator in PFE inevitably result in the loss of fine-grained information the object and local geometries. Best viewed in color." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "(2) Pillar Information Loss by Max-pooling. As mentioned in Section 1, current pillar-based methods [15, 30, 45, 49] heavily rely on max pooling to aggregate point features within each pillar. However, this max-pooling operation (Equation 3) inevitably leads to the loss of detailed information, particularly the local geometric patterns that play an important role in object recognition. To illustrate this point, we have visualized the effect of the max-pooling operation in Figure 2. In the figure, we represent the points that are more likely to be selected by max pooling as larger and red spheres. It is evident from the visualization that a" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 553, + 251 + ], + "type": "text", + "content": "significant number of points are not selected after the max-pooling operation, while only a small fraction of points have a high probability of being chosen. This comparison clearly shows that the max-pooling operator results in a substantial loss of object geometry information compared to the original points of the object." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 256, + 534, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 534, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 534, + 268 + ], + "type": "text", + "content": "3.3. Height-aware Discrete Histogram for PFE" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 269, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 316 + ], + "type": "text", + "content": "Hence, the challenge lies in finding a pillar encoding method that addresses these problems. Our goal is to develop a pillar feature encoding approach that possesses the following two key characteristics:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 317, + 553, + 437 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 317, + 553, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 553, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 553, + 377 + ], + "type": "text", + "content": "It should eliminate the requirement for sampling operations while still maintaining the ability to capture the height-aware of points within each pillar. This means that the encoding method should effectively handle the variation in point density across different height levels within a pillar." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 378, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 553, + 437 + ], + "type": "text", + "content": "It should keep the semantic and geometric information within the pillar, without resorting to max pooling or similar pooling operations. This is important to ensure that the encoded representation retains the relevant details and characteristics of the original pillar." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": "To achieve these objectives, we propose leveraging the histogram distribution. We can effectively capture the distribution of points within a pillar at different height levels. This allows us to maintain the height-awareness of the points without the need for sampling operations. The implementation of the histogram operation is straightforward. Initially, we divide the height range of the pillar into multiple bins, where each bin corresponds to a specific range of heights. For instance, if the pillar's height spans from 0 to 10 meters, we can partition it into several bins such as " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "0 - 1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "1 - 2\\mathrm{m}" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "2 - 3\\mathrm{m}" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": ", and so forth. Next, we proceed to count the number of points that fall within each bin for a given pillar. This counting process yields a count value for each bin, indicating the number of points residing within that particular height range. For instance, if a larger number of points are found in the lower height range bins, it signifies a higher density of points near the bottom of the pillar. Conversely, if there is a greater concentration of points in the higher height range bins, it suggests a higher density of points towards the top of the pillar. This approach provides the model with both the semantic and geometric information derived from the height dimension, thereby enhancing its understanding and representation of the point cloud data." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27339" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 79, + 494, + 363 + ], + "blocks": [ + { + "bbox": [ + 114, + 79, + 494, + 363 + ], + "lines": [ + { + "bbox": [ + 114, + 79, + 494, + 363 + ], + "spans": [ + { + "bbox": [ + 114, + 79, + 494, + 363 + ], + "type": "image", + "image_path": "3ed8bbfcf0f259111549c49cb22e3b931b993dd07c316d63c3b53e0764381a6f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 371, + 555, + 439 + ], + "lines": [ + { + "bbox": [ + 54, + 371, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 54, + 371, + 555, + 439 + ], + "type": "text", + "content": "Figure 3. The overall architecture of our proposed PillarHist. The PFE (a) in PointPillars first extracts point-wise feature, and then the point-wise features are max-pooled along the point dimension to obtain the pillar-wise representation. This process has three key drawbacks: (1) it lacks detailed height-aware information, (2) the MLP operates on point-wise level, incurring computational burden, (3) the max-pooling operation leads to information loss. In contrast, our proposed PillarHist (b) directly encodes the height-aware information based on histogram, and then perform the MLP operation on the pillar level rather than point level. Finally, the max-pooling operation is also eliminated in proposed PillarHist." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "content": "By using the histogram technique, we propose a height-aware pillar feature encoder named PillarHist. The algorithm is summarized in Algorithm 1: (1) We compute the histogram of each pillar, denoted as " + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "inline_equation", + "content": "H^p \\in \\mathbb{R}^B" + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "content": " is the number of bins. (2) In addition, we incorporates the LiDAR reflection intensity of each point, " + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "content": ", to obtain the weighted histogram of intensity, denoted as " + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "inline_equation", + "content": "H^i \\in \\mathbb{R}^B" + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "content": ". The weighted histogram represents the mean LiDAR reflection intensity for each bin, providing valuable height-aware information for object detection. (3) To provide the global position of each pillar, we utilize the x and y coordinates of the pillar center, " + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "inline_equation", + "content": "x_{center}" + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "inline_equation", + "content": "y_{center}" + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "content": ". Since the coordinates of the pillar are constant, this step does not require any additional computational effort. (4) We concatenate the computed histograms and the pillar center coordinates, resulting in " + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "inline_equation", + "content": "[H^p | H^i | x_{center} | y_{center}] \\in \\mathbb{R}^{2B + 2}" + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "content": ". (5) To represent the features of the pillars, we utilize a linear layer " + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "inline_equation", + "content": "f_\\theta" + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "content": " that projects the combined information into a feature vector. " + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{D \\times (2B + 2)}" + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 56, + 445, + 296, + 673 + ], + "type": "text", + "content": " is the number of output channels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Specifically, the pillar center " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_{center}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "y_{center}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " provide the relative spatial relationships between different points within the point cloud. This spatial location information" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 445, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 649 + ], + "type": "text", + "content": "of the pillars reflects the relative position of the targets in the scene, which can help the model better understand environmental context. For instance, if there are two pillars with the same point cloud distribution located at [0,10] and [10,0] in space, without the pillar coordinate to indicate the spatial relationship, these two pillars would be indistinguishable. This would pose significant challenges for the detector's object perception capabilities. Therefore, combining pillar center " + }, + { + "bbox": [ + 313, + 445, + 556, + 649 + ], + "type": "inline_equation", + "content": "x_{center}" + }, + { + "bbox": [ + 313, + 445, + 556, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 445, + 556, + 649 + ], + "type": "inline_equation", + "content": "y_{center}" + }, + { + "bbox": [ + 313, + 445, + 556, + 649 + ], + "type": "text", + "content": " with other features such as height and intensity can further enhance the model's perception capability. This is necessary to compensate for the missing spatial position information inherent in the raw point cloud data. By incorporating the pillar center coordinates, the model can better capture the geometric structure and spatial relationships between different parts of the scene. This is a crucial aspect for enabling accurate object detection in point cloud data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 653, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 557, + 715 + ], + "type": "text", + "content": "As shown in Fig 3, our PillarHist approach explicitly captures the histogram of points in different height intervals. This allows us to keep the semantic and geometric information within the pillar. PillarHist operates only during the PFE stage without introducing complex operations" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27340" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 75, + 205, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 75, + 205, + 87 + ], + "spans": [ + { + "bbox": [ + 62, + 75, + 205, + 87 + ], + "type": "text", + "content": "Algorithm 1: PillarHist Algorithm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "spans": [ + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "text", + "content": "Input: Raw point data in a pillar " + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "inline_equation", + "content": "V\\in \\mathbb{R}^{N_v\\times 4}" + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "text", + "content": " points, number of height bins " + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "text", + "content": " ,number of output channels " + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "text", + "content": " ,height of the pillar " + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "text", + "content": " , linear layer " + }, + { + "bbox": [ + 67, + 89, + 281, + 137 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 137, + 281, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 137, + 281, + 162 + ], + "spans": [ + { + "bbox": [ + 67, + 137, + 281, + 162 + ], + "type": "text", + "content": "Output: Pillar's feature vector " + }, + { + "bbox": [ + 67, + 137, + 281, + 162 + ], + "type": "inline_equation", + "content": "Y \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 67, + 137, + 281, + 162 + ], + "type": "text", + "content": " to represent the non-empty pillar " + }, + { + "bbox": [ + 67, + 137, + 281, + 162 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 67, + 137, + 281, + 162 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 163, + 303, + 245 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 73, + 163, + 303, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 163, + 303, + 186 + ], + "spans": [ + { + "bbox": [ + 73, + 163, + 303, + 186 + ], + "type": "text", + "content": "1: Divide the height " + }, + { + "bbox": [ + 73, + 163, + 303, + 186 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 73, + 163, + 303, + 186 + ], + "type": "text", + "content": " of the pillar into " + }, + { + "bbox": [ + 73, + 163, + 303, + 186 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 73, + 163, + 303, + 186 + ], + "type": "text", + "content": " bins, each bin having height " + }, + { + "bbox": [ + 73, + 163, + 303, + 186 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 73, + 163, + 303, + 186 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 73, + 163, + 303, + 186 + ], + "type": "inline_equation", + "content": "h = \\frac{H}{B}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 186, + 298, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 186, + 298, + 209 + ], + "spans": [ + { + "bbox": [ + 73, + 186, + 298, + 209 + ], + "type": "text", + "content": "2: Initialize " + }, + { + "bbox": [ + 73, + 186, + 298, + 209 + ], + "type": "inline_equation", + "content": "H^p \\in \\mathbb{R}^B" + }, + { + "bbox": [ + 73, + 186, + 298, + 209 + ], + "type": "text", + "content": " to zero, where " + }, + { + "bbox": [ + 73, + 186, + 298, + 209 + ], + "type": "inline_equation", + "content": "H^p" + }, + { + "bbox": [ + 73, + 186, + 298, + 209 + ], + "type": "text", + "content": " represents the histogram of points in each bin" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 209, + 295, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 209, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 73, + 209, + 295, + 233 + ], + "type": "text", + "content": "3: Initialize " + }, + { + "bbox": [ + 73, + 209, + 295, + 233 + ], + "type": "inline_equation", + "content": "H^i\\in \\mathbb{R}^B" + }, + { + "bbox": [ + 73, + 209, + 295, + 233 + ], + "type": "text", + "content": " to zero, where " + }, + { + "bbox": [ + 73, + 209, + 295, + 233 + ], + "type": "inline_equation", + "content": "H^{i}" + }, + { + "bbox": [ + 73, + 209, + 295, + 233 + ], + "type": "text", + "content": " represents the weighted histogram of intensity in each bin" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 234, + 225, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 234, + 225, + 245 + ], + "spans": [ + { + "bbox": [ + 73, + 234, + 225, + 245 + ], + "type": "text", + "content": "4: for " + }, + { + "bbox": [ + 73, + 234, + 225, + 245 + ], + "type": "inline_equation", + "content": "p_i = \\{p_i | i = 1, 2, \\dots, N_v\\}" + }, + { + "bbox": [ + 73, + 234, + 225, + 245 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "spans": [ + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "type": "text", + "content": "5: Compute the corresponding bin index " + }, + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "type": "text", + "content": " based on the " + }, + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "type": "text", + "content": " coordinate " + }, + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "type": "text", + "content": " of point " + }, + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 73, + 245, + 279, + 281 + ], + "type": "inline_equation", + "content": "b_{i} = \\lfloor \\frac{z_{i}}{h} \\rfloor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 281, + 280, + 304 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 88, + 281, + 187, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 281, + 187, + 292 + ], + "spans": [ + { + "bbox": [ + 88, + 281, + 187, + 292 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 88, + 281, + 187, + 292 + ], + "type": "inline_equation", + "content": "H^{p}[b_{i}]\\gets H^{p}[b_{i}] + 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 88, + 292, + 280, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 292, + 280, + 304 + ], + "spans": [ + { + "bbox": [ + 88, + 292, + 280, + 304 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 88, + 292, + 280, + 304 + ], + "type": "inline_equation", + "content": "H^{i}[b_{i}]\\gets H^{i}[b_{i}] + r_{i},\\quad \\triangleright r_{i}" + }, + { + "bbox": [ + 88, + 292, + 280, + 304 + ], + "type": "text", + "content": " is the LiDAR" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 74, + 305, + 197, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 305, + 197, + 316 + ], + "spans": [ + { + "bbox": [ + 74, + 305, + 197, + 316 + ], + "type": "text", + "content": "8: reflection intensity of " + }, + { + "bbox": [ + 74, + 305, + 197, + 316 + ], + "type": "inline_equation", + "content": "p_i" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 71, + 316, + 301, + 364 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 73, + 316, + 262, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 316, + 262, + 327 + ], + "spans": [ + { + "bbox": [ + 73, + 316, + 262, + 327 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 73, + 316, + 262, + 327 + ], + "type": "inline_equation", + "content": "H^{i}[b_{i}]\\coloneqq H^{i}[b_{i}] / H^{p}[b_{i}]" + }, + { + "bbox": [ + 73, + 316, + 262, + 327 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 73, + 316, + 262, + 327 + ], + "type": "inline_equation", + "content": "H^{i}[b_{i}]! = 0" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 71, + 329, + 301, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 329, + 301, + 339 + ], + "spans": [ + { + "bbox": [ + 71, + 329, + 301, + 339 + ], + "type": "text", + "content": "10: Compute the pillar center coordinates " + }, + { + "bbox": [ + 71, + 329, + 301, + 339 + ], + "type": "inline_equation", + "content": "x_{center}, y_{center}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 71, + 339, + 280, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 339, + 280, + 351 + ], + "spans": [ + { + "bbox": [ + 71, + 339, + 280, + 351 + ], + "type": "text", + "content": "11: Concatenate: " + }, + { + "bbox": [ + 71, + 339, + 280, + 351 + ], + "type": "inline_equation", + "content": "[H^p |H^i |x_{center}|y_{center}]\\in \\mathbb{R}^{2B + 2}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 71, + 351, + 298, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 351, + 298, + 364 + ], + "spans": [ + { + "bbox": [ + 71, + 351, + 298, + 364 + ], + "type": "text", + "content": "12: Linear projection: " + }, + { + "bbox": [ + 71, + 351, + 298, + 364 + ], + "type": "inline_equation", + "content": "Y = f_{\\theta}([H^{p}|H^{i}|x_{center}|y_{center}])" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 393, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 295, + 430 + ], + "type": "text", + "content": "or interaction with other modules, allowing for seamless inserted into existing pillar-based methods serves as a plugin module with performance improvements." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 430, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 598 + ], + "type": "text", + "content": "The Efficiency of PillarHist: Here, we further analyze the efficiency of PillarHist. (1). The linear projection in PillarHist operates on the pillar level, not on individual points. In contrast, the previous approach [15, 20, 30, 45] performed the linear projection for each point within a pillar. By employing the linear projection at the pillar level, we significantly decrease the number of floating-point operations (FLOPs) required by the PFE, resulting in improved efficiency (Table 5). (2). PillarHist also avoids various memory-intensive operations, such as copying the original point cloud data, calculating distances " + }, + { + "bbox": [ + 55, + 430, + 295, + 598 + ], + "type": "inline_equation", + "content": "(x^{m},y^{m},z^{m},x^{c},y^{c},z^{c})" + }, + { + "bbox": [ + 55, + 430, + 295, + 598 + ], + "type": "text", + "content": ", and performing max pooling. This further enhances the efficiency of PillarHist in real-world deployment." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 609, + 137, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 137, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 137, + 622 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Dataset and Training Details. We evaluate our PillarHist on KITTI [11], nuScenes [1] and Waymo dataset [34]. We report the official metrics. We adopt OpenPCDet [35] as the training framework and faithfully reproduced the different baseline methods based on the provided official setting on kitti and nuScenes dataset in the OpenPCDet [35]. we directly follow the setting of the four pillar-based 3D de" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": "tectors. (i.e., PointPillars [15], CenterPoint-Pillar [45], PillarNet [30], FastPillars [49]). In nuScenes [1] dataset, we uniformly sample 64 frames point cloud data from the validation set as the calibration set. Calibration is a standard step in quantization, where a small subset of dataset is used to determine activation quantization parameters (scale factors and zero points). This is a common practice in model quantization [24]. The calibration set proportions is " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "1.06\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": " (64/6019). We set the first and the last layer of the network to keep unquantized. Specifically, the quantization methods is widely used round-to-nearest (RTN) method, a common setting in prior works, such as PD-Quant [19] and LiDAPTQ [50]. We test all experiments on a single Nvidia Tesla V100 GPU. Please refer to Appendix for more quantization details and experiment analysis." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 262, + 409, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 409, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 409, + 274 + ], + "type": "text", + "content": "4.1. Overall Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 281, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 496 + ], + "type": "text", + "content": "Evaluation on nuScenes test set. The quantitative results on nuScenes test set are shown in Table 2. It can be seen that by integrating our proposed PillarHist into the current representative pillar-based detector PoinPillar [15], CenterPoint-Pillar [45], PillarNet [30], the proposed PH-PillarHist achieves significant performance gains on nuScenes test set across different pillar-based methods, respectively. Meanwhile, PH-PillarHist brings 6ms, 9ms, 5ms and 4ms reduction on inference latency, respectively. Notably, our method focuses on the pillar feature encoding module, rather than designing a state-of-the-art pillar-based detector. The proposed PillarHist is a plugin module to conveniently insert into existing mainstream pillar-based methods, seamlessly bringing performance gains and latency reduction. This presents the effectiveness and efficiency of PillarHist in not only boosting the detection accuracy, but also improving the computational efficiency, making it practical for real-world application." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 497, + 556, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 688 + ], + "type": "text", + "content": "Evaluation on nuScenes val set. Here, we evaluate our PillarHist on different pillar-based baselines in nuScenes val set. We selected three representative pillar-based detectors: PointPillars [15], CP-Pillar [45], and PillarNet [30]. In detail, we replace the raw PFE module in each baseline model with our PillarHist, resulting in PH-PointPillars, PH-CP-Pillar, and PH-PillarNet, respectively. As shown in Table 3, the results demonstrate the effectiveness of proposed PillarHist, as we observe performance improvements of 1.7 NDS, 1.4 NDS, and 1.6 NDS on the three different baselines, respectively. These results highlight the effectiveness and broad applicability of our PillarHist in enhancing the represent power of pillar-based 3D object detectors, leading to consistent performance gains across different baseline architectures. Notably, On the nuScenes dataset, the nuScenes detection score (NDS) is the primary evaluation metric." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "The Quantization Effectiveness of PillarHist. Furthermore, we also conduct quantization experiments to evaluate" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27341" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 236 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 236 + ], + "type": "table", + "html": "
MethodNDSmAPLatencyCarTruckBusTrailerC.V.Ped.MotBycT.C.Bar
Focals Conv[3]70.063.8138 ms86.756.367.759.523.887.564.536.381.474.1
VoxelNeXt[4]70.064.566 ms84.653.064.755.828.785.873.245.779.074.6
FG-PFE[25]71.865.786 ms85.255.561.661.829.286.572.548.682.573.8
FastPillars [49]71.866.856 ms87.358.066.062.334.587.470.347.981.572.8
PointPillars [15]45.330.531 ms68.423.028.223.44.159.727.41.130.838.9
PH-PointPillars47.032.425 ms69.424.230.024.89.660.928.15.231.540.1
CenterPoint [45]65.558.096 ms84.651.060.253.217.583.453.728.776.770.9
PH-CenterPoint66.659.287 ms85.252.561.154.219.083.954.830.179.271.5
\\( PillarNet-18[30]^\\dagger \\)70.865.078 ms87.456.760.961.830.487.267.440.382.176.0
\\( PH-PillarNet-18^\\dagger \\)71.565.273 ms87.657.060.862.030.687.368.140.381.976.2
\\( PillarNet-34[30]^\\dagger \\)71.466.096 ms87.657.563.663.127.987.370.142.383.377.2
\\( PH-PillarNet-34^\\dagger \\)72.066.392 ms87.557.664.063.328.487.170.543.083.677.8
", + "image_path": "58f809cc00b7a560bca7a47a9fee6ede767c155ad64da3b188e9d2c70fb629fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 267, + 287, + 433 + ], + "blocks": [ + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "type": "text", + "content": "Table 2. Performance of 3D object detection methods on nuScenes test set. " + }, + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "type": "text", + "content": " means the method that uses double-flip testing. All models listed take LIDAR data as input without image fusion or any model ensemble." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 267, + 287, + 433 + ], + "lines": [ + { + "bbox": [ + 62, + 267, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 62, + 267, + 287, + 433 + ], + "type": "table", + "html": "
ModelsBits (W/A)NDSmAP
PointPillars [15]32/3252.137.4
8/844.8 (-7.3)29.8 (-7.6)
PH-PointPillars32/3253.8 (+1.7)38.8 (+1.4)
8/852.2 (-1.6)37.0 (-1.8)
CP-Pillar[45]32/3256.444.9
8/848.6 (-7.8)36.2 (-8.7)
PH-CP-Pillar32/3257.8 (+1.4)45.8 (+0.9)
8/856.3 (-1.5)44.3 (-1.5)
PillarNet[30]32/3261.452.6
8/854.6 (-6.8)43.1 (-9.5)
PH-PillarNet32/3263.0 (+1.6)52.8 (+0.2)
8/861.6 (-1.4)51.4 (-1.3)
", + "image_path": "9a1db4b57357cbcecec55e40a3daa4033d45f17052572559931caa7f05543806.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "content": "the quantization-friendly ability. Specifically, we convert the weights and activations of the models from floating-point 32-bit to integer 8-bit and applied a simple posttraining quantization setting. As shown in Table 3, we find that the PointPillars, CP-Pillar, and PillarNet models suffered significant performance degradation after weight and activation quantization, with drops of 7.3 NDS and " + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "inline_equation", + "content": "7.6\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "content": " for PointPillars, 7.8 NDS and " + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "inline_equation", + "content": "8.7\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "content": " for CP-Pillar, and 6.8 NDS and " + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "inline_equation", + "content": "9.5\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "content": " for PillarNet. In contrast, the PH-PointPillars, PH-CP-Pillar, and PH-PillarNet models that integrate PillarHist only perform performance drops of 1.6 NDS, 1.5 NDS, and 1.4 NDS, respectively. Notably, the quantized models with PillarHist integration achieved comparable performance to original floating-point counterparts without PillarHist (52.2 v.s. 52.1 NDS on PointPillars, 56.3 v.s. 56.4 NDS on CenterPoint-Pillars and 61.6 v.s. 61.4 NDS on PillarNet). Specifically, PillarHist could effectively eliminate the severe performance degradation after quantization caused by other PFE design, bringing quantization performance gain seamlessly. These results demonstrate the quantization-friendly nature of PillarHist, which maintains" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 331 + ], + "type": "text", + "content": "the model's representation power and performance even under quantization compression. This property is particularly valuable and practicable for deploying high-performance 3D object detectors on resource-constrained edge devices, where accuracy and latency are critical factors." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 337, + 553, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 435, + 295, + 458 + ], + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 458 + ], + "type": "text", + "content": "Table 3. Generalization and quantization performance comparison on nuScene val set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 337, + 553, + 392 + ], + "lines": [ + { + "bbox": [ + 315, + 337, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 337, + 553, + 392 + ], + "type": "table", + "html": "
MethodsMean (L2) mAPHVehicle (L2) mAPmAPHPed. (L2) mAPmAPHCyc. (L2) mAPmAPH
CP-Pillar55.962.562.061.851.556.154.3
PH-CP-Pillar57.363.262.863.854.057.055.3
↑+1.4+0.7+0.8+2.0+2.5+0.9+1.0
", + "image_path": "748000c08b08bb39faa7fcf49fa71b53683c43b58a7dc6f9d1bda0d464b922dc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 419, + 556, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 556, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 556, + 527 + ], + "type": "text", + "content": "Generalization of PillarHist on Waymo val set. Here, we also evaluate our PillarHist on Waymo dataset. We select CenterPoint-Pillars as our baseline and only replaced their pillar feature encoding module with our PillarHist for comparison, as show in table our method can bring " + }, + { + "bbox": [ + 313, + 419, + 556, + 527 + ], + "type": "inline_equation", + "content": "1.2\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 419, + 556, + 527 + ], + "type": "text", + "content": " L2 and " + }, + { + "bbox": [ + 313, + 419, + 556, + 527 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{mAPH}" + }, + { + "bbox": [ + 313, + 419, + 556, + 527 + ], + "type": "text", + "content": " L2 performance improvements. Notably, it can achieve " + }, + { + "bbox": [ + 313, + 419, + 556, + 527 + ], + "type": "inline_equation", + "content": "2.0\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 419, + 556, + 527 + ], + "type": "text", + "content": " L2 and " + }, + { + "bbox": [ + 313, + 419, + 556, + 527 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{mAPH}" + }, + { + "bbox": [ + 313, + 419, + 556, + 527 + ], + "type": "text", + "content": " L2 performance gains in the pedestrian category. This result demonstrates the effectiveness of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 527, + 408, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 527, + 408, + 540 + ], + "spans": [ + { + "bbox": [ + 314, + 527, + 408, + 540 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 546, + 553, + 634 + ], + "blocks": [ + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "type": "text", + "content": "Table 4. Apply PillarHist to CeterPoint-Pillars on Waymo val set trained with " + }, + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "type": "text", + "content": " training data." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 546, + 553, + 634 + ], + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 634 + ], + "type": "table", + "html": "
PFE moduleCar 3D R40Ped 3D R40PFE GFLOPs
mAPEasyMod.HardmAPEasyMod.Hard
PFE[15]79.9887.0877.9074.9749.4154.7149.0144.520.152
PAA[16]81.0488.6279.2475.2748.5654.2647.9343.493.391
TA[20]80.7187.3379.1175.6949.4554.0049.3145.057.061
DynPFE*[30]79.9687.6377.7074.5549.3654.2749.0144.810.013
DynPFE[45]80.4687.8678.1675.3847.7752.2847.5343.520.112
FG-PFE[25]----50.7556.8349.8845.561.136
MAPE[49]81.1989.3978.5775.6348.1852.6848.0243.830.165
PillarHist81.4288.8079.1376.3051.0757.4550.4245.360.065
", + "image_path": "ae92e16005866d65758a738f7fee3ed8878fec5385526a8ffae53f4cf271e381.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 637, + 543, + 648 + ], + "lines": [ + { + "bbox": [ + 324, + 637, + 543, + 648 + ], + "spans": [ + { + "bbox": [ + 324, + 637, + 543, + 648 + ], + "type": "text", + "content": "Table 5. Performance of different PFE on the KITTI val set." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Comparison of Different PFE modules. We compare the computation efficiency and performance of existing PFE modules. Specifically, we evaluate the following PFE modules: PFE in PointPillars [15], DynPFE in CenterPoint-Pillar [45], TA module in TANet [20], PAA module in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27342" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": "PiFeNet [16], DynPFE* in PillarNet [30], and MAPE in FastPillars [49]. To ensure a fair comparison, we select PointPillars as the baseline method and conduct experiments by replacing the PFE in PointPillars with other PFE module (DynPFE, TA, PAA, DynpFE*, MAPE and our PillarHist). This setting eliminates the performance difference caused by different network design. As shown in Table 5, our PillarHist achieves the highest 3D AP compared to other PFE methods while maintaining superior computational efficiency. Notably, although DynPFE* used in PillarNet [30] has lower GFLOPs, it suffers from poor detection performance. Notably, our PillarHist method performs linear projection at the pillar-wise level rather than point-wise level in other PFE. This design can efficiently reduce the computational cost while maintaining the necessary structural information within pillar without max-pooling operations." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 274, + 294, + 327 + ], + "blocks": [ + { + "bbox": [ + 56, + 274, + 294, + 327 + ], + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 327 + ], + "type": "table", + "html": "
heightintensitycoordsCar 3D40Ped 3D40
mAPEasyMod.HardmAPEasyMod.Hard
✓80.1887.4778.1875.2248.0653.7547.5142.92
✓✓80.8188.2978.6275.5349.6255.4249.1044.36
✓✓✓81.4188.8079.1376.3051.0757.4550.4245.36
", + "image_path": "288099c089d9c8260e669a7d07a018374db91de289f89c70fde93d08f59e737f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 335, + 295, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 357 + ], + "type": "text", + "content": "Table 6. Ablation results of different input for PillarHist on the KITTI val set for car and pedestrian." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 366, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 617 + ], + "type": "text", + "content": "The Effects of Different Input in PillarHist. Here, we conduct ablation study to analyze the impact of different input. As shown in Table 6, when we incorporate the intensity information, it leads to performance improvements for both the vehicle and pedestrian categories. Particularly for the pedestrian, approximately 1.7, 1.6, and 1.6 points for the easy, moderate, and hard levels, respectively. This notable improvement indicates that for small objects like pedestrians, the intensity values are crucial in distinguishing them from the background and enhancing their distinctive features. Furthermore, by adding the pillar coordinates information, there is a performance improvement of 0.6 and 1.5 points for the vehicle and pedestrian. We believe this improvement is because explicitly incorporating the global coordinate of the pillars helps the model better identify the object positions with the enhanced encoding information that includes height awareness. Notably, after adding the pillar coordinates, there is an improvement of 2.03 points in easy AP for pedestrians. This can be attributed that the pillars encode not only varying height distribution but also including the difference between empty and non-empty pillars." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Feature Map Visualization of PillarHist. Here, as shown in Figure 4, we apply our PillarHist method to PointPillars [15] (a) and compare the features from the max-pooling operation in their PFE. We find that the max pooling operation tends to focus more on the points at higher positions within the same pillar in BEV (with larger coordinates), resulting in a biased feature representation. In contrast, our PillarHist can focus more evenly on each point within" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 65, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 316, + 65, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 316, + 65, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 316, + 65, + 553, + 224 + ], + "type": "image", + "image_path": "4e1b6e0bce78164ffc54c7e0885a4ea68909a150b5dcf8091594c6a283ed8b27.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 235, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 291 + ], + "type": "text", + "content": "Figure 4. (a) Visualization of the learned attention scores in PillarHist on different scenes. The color of points represents their scores. The point will be paid more attention if it has a higher score. (b) Visualization of the learned BEV feature in PointPillars and PH-PointPillars. Best viewed in color." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 297, + 556, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 556, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 556, + 452 + ], + "type": "text", + "content": "the pillar, encoding the height information into an unbiased histogram distribution, thereby obtaining richer geometric-semantic feature representation. This difference is particularly obvious at the car door location, the points within the pillar is typically uniformly distributed along height; however, max pooling emphasizes the higher contours, whereas PillarHist provides a better representation of the complete geometry of the car door Besides, as shown in Figure 4 (b), we also visualize the BEV features maps. Our method can obtain a more informative BEV feature representation compared to the PFE in PointPillars [15]. This also demonstrates the underlying reason for the better performance of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 464, + 388, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 464, + 388, + 475 + ], + "spans": [ + { + "bbox": [ + 314, + 464, + 388, + 475 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 483, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 712 + ], + "type": "text", + "content": "In this paper, we addressed the issue of information loss and large numerical distribution difference in existing pillar-based 3D object detectors during the PFE stage. We identify the height dimension as a crucial factor for enhancing 3D detection performance and propose a height-aware pillar feature encoder called PillarHist. By leveraging histogram, PillarHist preserve the information along the height dimension while effectively mitigating the information loss introduced by max-pooling operations. Importantly, PillarHist operated exclusively within the PFE stage, allowing seamless insert into existing pillar-based methods serves as a plug-in module. Meanwhile, PillarHist also constrains the arithmetic distribution of PFE input to a stable range, making it quantization-friendly. Extensive experiments demonstrated the effectiveness of PillarHist in terms of both efficiency and performance. Our proposed method achieved an average improvement of approximately " + }, + { + "bbox": [ + 313, + 483, + 556, + 712 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 313, + 483, + 556, + 712 + ], + "type": "text", + "content": " NDS across various baselines, showcasing its ability to enhance 3D object detection accuracy and computation efficiency." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27343" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "type": "text", + "content": "We thank anonymous reviewers for their kind help of this work. This work was supported by National Natural Science Foundation of China (No.62271143), Frontier Technologies RD Program of Jiangsu (No. BF2024060) and the Big Data Computing Center of Southeast University." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 186, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "type": "text", + "content": "[1] Holger Caesar, Varun Bankiti, Alex Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. Nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 241, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 241, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 296, + 274 + ], + "type": "text", + "content": "[2] Qi Cai, Yingwei Pan, Ting Yao, and Tao Mei. 3d cascade rconn: High quality object detection in point clouds. IEEE Transactions on Image Processing, 31:5706-5719, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "text", + "content": "[3] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 318, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 295, + 370 + ], + "type": "text", + "content": "[4] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 294, + 426 + ], + "type": "text", + "content": "[5] Zehui Chen, Zheng Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, Feng Wu, and Feng Zhao. Graph-detr4d: Spatio-temporal graph modeling for multiview 3d object detection. IEEE Transactions on Image Processing, 33:4488-4500, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "type": "text", + "content": "[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3D object detection. In AAAI, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 472, + 294, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 294, + 515 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 294, + 515 + ], + "type": "text", + "content": "[7] Xiaoying Ding, Weisi Lin, Zhenzhong Chen, and Xinfeng Zhang. Point cloud saliency detection by local and global feature fusion. IEEE Transactions on Image Processing, 28 (11):5379-5393, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 516, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 516, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 294, + 548 + ], + "type": "text", + "content": "[8] Lue Fan, Feng Wang, Naiyan Wang, and ZHAO-XIANG ZHANG. Fully sparse 3d object detection. Advances in Neural Information Processing Systems, 35:351-363, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 549, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 549, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 62, + 549, + 294, + 592 + ], + "type": "text", + "content": "[9] Zheng Fang, Sifan Zhou, Yubo Cui, and Sebastian Scherer. 3d-siamrpn: An end-to-end learning method for real-time 3d single object tracking using raw point cloud. IEEE Sensors Journal, 21(4):4995-5011, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 593, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 593, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 593, + 295, + 624 + ], + "type": "text", + "content": "[10] Jiahui Fu, Guanghui Ren, Yunpeng Chen, and Si Liu. Improved pillar with fine-grained feature for 3d object detection. arXiv preprint arXiv:2110.06049, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "text", + "content": "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[13] Shan Jiayao, Sifan Zhou, Yubo Cui, and Zheng Fang. Realtime 3d single object tracking with transformer. IEEE Transactions on Multimedia, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[14] Hongwu Kuang, Bei Wang, Jianping An, Ming Zhang, and Zehan Zhang. Voxel-fpn: Multi-scale voxel feature aggregation for 3d object detection from lidar point clouds. Sensors, 20(3):704, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[15] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "text", + "content": "[16] Duy Tho Le, Hengcan Shi, Hamid Rezatofighi, and Jianfei Cai. Accurate and real-time 3d pedestrian detection using an efficient attentive pillar network. IEEE Robotics and Automation Letters, 8(2):1159-1166, 2022. 3, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[17] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "type": "text", + "content": "[18] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv preprint arXiv:2102.05426, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 555, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 387 + ], + "type": "text", + "content": "[19] Jiawei Liu, Lin Niu, Zhihang Yuan, Dawei Yang, Xinggang Wang, and Wenyu Liu. Pd-quant: Post-training quantization based on prediction difference metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24427–24437, 2023. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[20] Zhe Liu, Xin Zhao, Tengteng Huang, Ruolan Hu, Yu Zhou, and Xiang Bai. Tanet: Robust 3d object detection from point clouds with triple attention. In AAAI, pages 11677-11684, 2020. 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 434, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 476 + ], + "type": "text", + "content": "[21] Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. Voxel transformer for 3d object detection. In ICCV, pages 3164-3173, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[22] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A comprehensive survey. International Journal of Computer Vision, pages 1-55, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 523, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 577 + ], + "type": "text", + "content": "[23] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "text", + "content": "[24] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yely-sei Bondarenko, Mart Van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "text", + "content": "[25] Konyul Park, Yechel Kim, Junho Koh, Byungwoo Park, and Jun Won Choi. Fine-grained pillar feature encoding via spatio-temporal virtual grid for 3d object detection. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 4259-4265, 2024. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, pages 652-660, 2017. 1, 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27344" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[27] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "text", + "content": "[28] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "text", + "content": "[29] Jiayao Shan, Sifan Zhou, Zheng Fang, and Yubo Cui. Ptt: Point-track-transformer module for 3d single object tracking in point clouds. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1310-1316. IEEE, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "type": "text", + "content": "[30] Guangsheng Shi, Ruifeng Li, and Chao Ma. Pillarnet: Real-time and high-performance pillar-based 3d object detection. In ECCV, 2022. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "text", + "content": "[31] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE T-PAMI, 2020. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[32] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3D object detection in a point cloud. In CVPR, pages 1711-1719, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "type": "text", + "content": "[33] Lukas Stäcker, Juncong Fei, Philipp Heidenreich, Frank Bonarens, Jason Rambach, Didier Stricker, and Christoph Stiller. Deployment of deep neural networks for object detection on edge ai devices with runtime optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshop, pages 1015-1022, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "text", + "content": "[34] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 497 + ], + "type": "text", + "content": "[35] OpenPCDet Development Team. OpenPCDet: An opensource toolbox for 3D object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 499, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 294, + 554 + ], + "type": "text", + "content": "[36] Yonglin Tian, Lichao Huang, Hui Yu, Xiangbin Wu, Xuesong Li, Kunfeng Wang, Zilei Wang, and Fei-Yue Wang. Context-aware dynamic feature extraction for 3d object detection in point clouds. IEEE Transactions on Intelligent Transportation Systems, 23(8):10773-10785, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 588 + ], + "type": "text", + "content": "[37] Zhi Tian, Xiangxiang Chu, Xiaoming Wang, Xiaolin Wei, and Chunhua Shen. Fully convolutional one-stage 3d object detection on lidar range images. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "text", + "content": "[38] Tai Wang, Xinge Zhu, and Dahua Lin. Reconfigurable voxels: A new representation for lidar-based point clouds. arXiv preprint arXiv:2004.02724, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 645 + ], + "type": "text", + "content": "[39] Yue Wang and Justin M Solomon. Object dgCNN: 3d object detection using dynamic graphs. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[40] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization. arXiv preprint arXiv:2203.05740, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "text", + "content": "[41] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 588 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "text", + "content": "[42] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3D single stage object detector. In CVPR, pages 11040-11048, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[43] Maosheng Ye, Shuangjie Xu, and Tongyi Cao. Hvnet: Hybrid voxel network for lidar based 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1631-1640, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "type": "text", + "content": "[44] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Centerpoint++ submission to the waymo real-time 3d detection challenge. 2021. Accessed: 2021-12-05. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "type": "text", + "content": "[45] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[46] Jiangyong Yu, Changyong Shu, Dawei Yang, Sifan Zhou, Zichen Yu, Xing Hu, and Yan Chen. Q-petr: Quant-aware position embedding transformation for multi-view 3d object detection. arXiv e-prints, pages arXiv-2502, 2025. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "type": "text", + "content": "[47] Yifan Zhang, Zhen Dong, Huanrui Yang, Ming Lu, Cheng-Ching Tseng, Yandong Guo, Kurt Keutzer, Li Du, and Shanghang Zhang. Qd-bev: Quantization-aware view-guided distillation for multi-view 3d object detection. 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 319, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 353 + ], + "type": "text", + "content": "[48] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In AAAI, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 408 + ], + "type": "text", + "content": "[49] Sifan Zhou, Zhi Tian, Xiangxiang Chu, Xinyu Zhang, Bo Zhang, Xiaobo Lu, Chengjian Feng, Zequn Jie, Patrick Yin Chiang, and Lin Ma. Fastpillars: A deployment-friendly pillar-based 3d detector. arXiv preprint arXiv:2302.02367, 2023. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "text", + "content": "[50] Sifan Zhou, Liang Li, Xinyu Zhang, Bo Zhang, Shipeng Bai, Miao Sun, Ziyu Zhao, Xiaobo Lu, and Xiangxiang Chu. Lidar-ptq: Post-training quantization for point cloud 3d object detection. International Conference on Learning Representations (ICLR), 2024. 1, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "type": "text", + "content": "[51] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3D object detection. In CVPR, pages 4490-4499, 2018. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 499, + 553, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 554 + ], + "type": "text", + "content": "[52] Yin Zhou, Pei Sun, Yu Zhang, Dragomir Anguelov, Jiyang Gao, Tom Ouyang, James Guo, Jiquan Ngiam, and Vijay Vasudevan. End-to-end multi-view fusion for 3d object detection in lidar point clouds. In Conference on Robot Learning, pages 923-932. PMLR, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 555, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 588 + ], + "type": "text", + "content": "[53] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv:1908.09492, 2019. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27345" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/6042cf5b-516c-4fe8-bfad-ecfc854fc855_content_list.json b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/6042cf5b-516c-4fe8-bfad-ecfc854fc855_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d790d9fa31b5db1d12f871142c5f3f604abd4c09 --- /dev/null +++ b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/6042cf5b-516c-4fe8-bfad-ecfc854fc855_content_list.json @@ -0,0 +1,1541 @@ +[ + { + "type": "text", + "text": "Pioneering 4-Bit FP Quantization for Diffusion Models: Mixup-Sign Quantization and Timestep-Aware Fine-Tuning", + "text_level": 1, + "bbox": [ + 156, + 128, + 843, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maosen Zhao $^{1*}$ , Pengtao Chen $^{1*}$ , Chong Yu $^{2}$ , Yan Wen $^{1}$ , Xudong Tan $^{1}$ , Tao Chen $^{1\\dagger}$", + "bbox": [ + 151, + 202, + 836, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Information Science and Technology, Fudan University", + "bbox": [ + 233, + 220, + 761, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Academy for Engineering and Technology, Fudan University", + "bbox": [ + 251, + 238, + 745, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "20307130202@fudan.edu.cn, eetchen@fudan.edu.cn", + "bbox": [ + 290, + 258, + 699, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model quantization reduces the bit-width of weights and activations, improving memory efficiency and inference speed in diffusion models. However, achieving 4-bit quantization remains challenging. Existing methods, primarily based on integer quantization and post-training quantization fine-tuning, struggle with inconsistent performance. Inspired by the success of floating-point (FP) quantization in large language models, we explore low-bit FP quantization for diffusion models and identify key challenges: the failure of signed FP quantization to handle asymmetric activation distributions, the insufficient consideration of temporal complexity in the denoising process during fine-tuning, and the misalignment between fine-tuning loss and quantization error. To address these challenges, we propose the mixup-sign floating-point quantization (MSFP) framework, first introducing unsigned FP quantization in model quantization, along with timestep-aware LoRA (TALoRA) and denoising-factor loss alignment (DFA), which ensure precise and stable fine-tuning. Extensive experiments show that we are the first to achieve superior performance in 4-bit FP quantization for diffusion models, outperforming existing PTQ fine-tuning methods in 4-bit INT quantization.", + "bbox": [ + 88, + 340, + 485, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 700, + 220, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the impressive performance of diffusion models (DMs) in image generation [30, 41], their computational and memory demands, particularly for high-resolution outputs, pose significant challenges for deployment on resource-constrained edge devices, highlighting the need for model compression to address these limitations. Model quantization, a key technique within model compression, reduces the bit-width of model weights and activations, typically stored in 32-bit format, to lower precision. This reduction lowers memory usage and accelerates inference", + "bbox": [ + 89, + 726, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "speed. By decreasing the bit-width, quantization improves both temporal and memory efficiency in mainstream models, maintaining robust performance, particularly in resource-constrained environments [13, 18, 21].", + "bbox": [ + 511, + 310, + 906, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The current quantization methods for diffusion models can be broadly categorized into two primary approaches. The first, post-training quantization (PTQ), optimizes the quantization parameters after the model has been trained, typically by minimizing the quantization error [19, 32]. While PTQ is effective for 4-bit quantization of weights, it is limited by its reliance on 8-bit quantization for activations, as further reduction in activation bit-width leads to significant performance degradation. In contrast, quantization-aware training (QAT) integrates quantization into the training process, enabling the model to learn with 4-bit precision from scratch [20]. Although QAT can achieve high-performance 4-bit models, it incurs substantial computational overhead, rendering it less practical for many real-world applications [6, 15].", + "bbox": [ + 511, + 371, + 908, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To achieve fully quantized 4-bit diffusion models with minimal overhead, fine-tuning has emerged as a promising solution. This approach leverages pre-trained models and adjusts a small subset of parameters to narrow the performance gap between the quantized model and its full-precision counterpart. While some studies have explored fine-tuning for 4-bit quantization in diffusion models [8, 39], these methods often fail to achieve consistent performance under standard configurations (e.g., quantizing all layers but not some). Consequently, developing a universally effective and scalable fine-tuning method for 4-bit quantization in diffusion models remains an open challenge.", + "bbox": [ + 511, + 598, + 908, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To be noticed, existing quantization methods for diffusion models primarily rely on integer (INT) quantization, which has long been the dominant approach. However, recent developments have demonstrated the considerable potential of floating-point (FP) quantization. Compared to INT quantization, FP quantization offers greater flexibility in modeling complex weight and activation distributions [27], leading to improved performance in visual", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author. *Equal contribution.", + "bbox": [ + 114, + 886, + 362, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18134", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tasks at 8-bit precision [17, 36], and remarkable results in large language models under 4-bit precision [23, 40]. Moreover, FP quantization provides significant advantages in inference acceleration, with NVIDIA H100 achieving a $1.45 \\times$ speedup with FP8 quantization, outperforming INT8 [29, 42]. Despite these advantages, the application of low-bit FP quantization in diffusion models remains largely unexplored, presenting a promising avenue for future research.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, to address the challenges in achieving 4-bit diffusion models, we propose a baseline method using a search-based signed FP quantization framework [2, 23] combined with single-LoRA fine-tuning, aiming to realize 4-bit FP quantized diffusion models. Through this exploration, we uncover several findings that are significant for both diffusion model quantization and FP quantization: (1) There exists many layers which exhibit asymmetric distributions, due to the nonlinear activation function $SiLU$ . The application of traditional signed FP quantization with a symmetric distribution leads to significant precision loss in the sub-zero area, causing substantial performance degradation after quantization. (2) Fine-tuning is conducted based on the denoising process, which is regarded as a complex task involving the restoration from outlines to details [37]. However, current methods typically apply a single LoRA to fine-tune severely degraded models across all timesteps, which leads to suboptimal learning at certain timesteps. (3) Predicted noise plays a varying role at different timesteps during denoising, which is the key to diffusion models. Arising from the neglect of this variation across timesteps, we identify a mismatch between the impact of quantization and the loss function in current methods, which undermines the effectiveness of fine-tuning.", + "bbox": [ + 91, + 214, + 483, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Facing the challenges in achieving 4-bit FP quantization, we propose constructive strategies: (1) To handle different distribution effectively, we propose a mixup-sign floating-point quantization (MSFP) framework, where unsigned FP quantization with an added zero point leads to a more compatible distribution of discrete points with the anomalous distributions in activations and signed FP quantization continues to exhibit strong representation capacity on other distributions. (2) Realizing that the current single-LoRA fine-tuning approach is in lack of flexibility across timesteps, we introduce timestep-aware LoRA (TALoRA), incorporating multiple LoRAs and a timestep-aware router to dynamically select the appropriate LoRA for each timestep in the denoising process. (3) To further improve the effectiveness of fine-tuning, we introduce a denoising-factor loss alignment (DFA), ensuring the loss function, and the guidance of the fine-tuning, are consistent with the actual quantization deterioration across timesteps.", + "bbox": [ + 89, + 579, + 482, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 106, + 853, + 408, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(i) We are the first to identify that signed FP quantization struggles with asymmetric activations, which", + "bbox": [ + 107, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "arise from the nonlinear behavior of activation functions. To address this, we introduce the MSFP framework, which is also the first effective application of unsigned FP quantization in quantization, offering a novel approach for achieving low-bit quantization.", + "bbox": [ + 552, + 90, + 906, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(ii) For the fine-tuning of low-bit diffusion models, which is based on the denoising process, we precisely define it as a multi-task process across timesteps and introduce an efficient TALoRA module. Furthermore, we improve the alignment of loss function with quantization error via DFA strategy, enabling stable and reliable fine-tuning to achieve low-bit diffusion models.", + "bbox": [ + 526, + 167, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(iii) We focus on identifying and eliminating three major barriers to effective low-bit FP quantization in diffusion models. Extensive experiments on DDIM and LDM demonstrate that our work achieves SOTA results for 4-bit quantization. The proposed MSFP, TALoRA and DFA have greatly advanced the progress of low-bit quantization in diffusion models.", + "bbox": [ + 522, + 272, + 903, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 396, + 653, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion Model Quantization", + "text_level": 1, + "bbox": [ + 513, + 422, + 779, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are two main approaches in the quantization of diffusion models: QAT [6, 14] and PTQ [28]. QAT is particularly effective for low-bit quantization while it retransmits the model from scratch, consuming extensive computational resources and time [20]. In contrast, PTQ offers greater time efficiency, making it more practical for large models in real-world applications. Recent advancements in PTQ have concentrated on optimizing calibration datasets to enhance reconstruction accuracy [9, 19, 25, 32] and mitigating quantization errors stemming from the temporal and structural properties of diffusion models [3, 19, 33, 35]. To further achieve fully 4-bit quantization, finetuning techniques has been introduced in PTQ-based quantization. EfficientDM [8] develops a LoRA-based finetuning framework, while QuEST [39] focuses on optimizing quantization-unfriendly activations. Both works stagnate in addressing data distribution during fine-tuning, failing to account for the specific challenges involved in finetuning within the context of diffusion model quantization.", + "bbox": [ + 511, + 446, + 906, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, recent advancements in FP quantization have made significant strides in model quantization [17, 23], highlighting its potential for diffusion models. While existing low-bit quantization methods perform well for linear models, applying FP quantization to convolutional diffusion models remains more challenging. To date, only one study has explored 8-bit activation quantization in diffusion models using a basic search-based approach [2], with no research addressing lower-bit quantization. This underscores both the challenges and the untapped potential of achieving 4-bit FP quantization for diffusion models.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18135", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4e3effa50b5d8326230a2591b257ff69ba8f6f732d73b7721d59900932d8a318.jpg", + "image_caption": [ + "Figure 1. The activation distributions in NALs and AALs, results on the CelebA dataset. (a) The paradigm of NALs with symmetric activations. (b) The typical paradigm of AALs with asymmetric activations, where unsigned FP quantization is more suitable. (c) The infrequent paradigm of AALs with relatively symmetric activations, where either signed or unsigned FP quantization could be applicable." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 364, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c83abc04fa0f7347d12b7a595d24dbb92053e026580e0418646e8eedbc838016.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 90, + 633, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/87fcb6d0f687d38b802457a5ff7c2538c91eaaa900b6a761ca3f35cb4afa78e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 90, + 898, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Parameter-Efficient Fine-Tuning", + "text_level": 1, + "bbox": [ + 89, + 330, + 377, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameter-efficient fine-tuning (PEFT) has emerged as an effective alternative to full model fine-tuning, focusing on adjusting only a subset of parameters while keeping the majority frozen, thereby reducing storage overhead. Low-rank adapters (LoRA) [12], originally developed for large language models, have become one of the most widely used PEFT methods. Leveraging LoRA's strong transferability, QLoRA [5] can be effectively applied to fine-tune low-bit diffusion models. However, previous LoRA-based fine-tuning is suboptimal in practice. In this paper, we make a further exploration on this and incorporate timestep-level adaptation inspired by MoELoRA [7], enhancing performance of low-bit quantized DMs.", + "bbox": [ + 88, + 353, + 483, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Challenges & Exploration", + "text_level": 1, + "bbox": [ + 89, + 563, + 333, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 588, + 220, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models. The diffusion model is a new generation framework that completes learning by adding noise and completes generation in a denoising manner. As for the forward process, the noise is injected into ground-truth images $\\boldsymbol{x}_0$ at random timesteps. This enables the DMs to learn the distributions of noise through noisy images $\\boldsymbol{x}_t$ , which can be obtained as follows [11]:", + "bbox": [ + 89, + 609, + 483, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\varepsilon . \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 729, + 480, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $\\varepsilon$ represents a standard Gaussian noise and $\\overline{\\alpha}_t$ is the accumulated noise intensity, calculated by:", + "bbox": [ + 89, + 753, + 482, + 782 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {t} \\alpha_ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 794, + 482, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{s}$ governs the noise intensity under each timestep. Once the DMs are well-trained, the Gaussian noise image will be inputted to the DMs, and undergo the iterative denoising process. Specifically, the noise can be predicted", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by DMs and used to obtain the image $\\boldsymbol{x}_t$ under timestep $t$ , which could range from $T$ to 0, with the objective of obtaining the image $\\boldsymbol{x}_{t-1}$ in a superior quality:", + "bbox": [ + 511, + 332, + 906, + 378 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\overline {{\\alpha}} _ {t}}} \\cdot \\boldsymbol {\\varepsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right) + \\sigma_ {t} \\boldsymbol {\\delta}, (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 397, + 906, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\varepsilon_{\\theta}(\\boldsymbol{x}_t,t)$ is the predicted noise at timestep $t$ and $\\delta$ is a newly added noise with the factor $\\sigma_t$ to ensure diverse results. Here we define a denoising factor $\\gamma_t$ , which is formulated as:", + "bbox": [ + 511, + 436, + 906, + 497 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {t} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\cdot \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 508, + 903, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to Equation 3, $\\gamma_{t}$ indicates the impact of the prediction noise under timestep $t$ . The greater the factor $\\gamma_{t}$ , the stronger the predicted noise effect in the denoising.", + "bbox": [ + 511, + 544, + 905, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model Quantization. The fundamental paradigm of model quantization is the process of transforming a continuous data distribution into a finite set of discrete points. Consequently, the quality of the quantized model is inextricably related to the distribution of discrete points. According to the discrete point type, quantization is defined as two categories, widely-used INT quantization and emerging FP quantization. Analogous to the process of INT quantization, a floating-point vector $x$ can be quantized as follows:", + "bbox": [ + 511, + 589, + 906, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\boldsymbol {x}} = \\operatorname {C l i p} \\left(\\left\\lfloor \\frac {\\boldsymbol {x}}{s} \\right\\rceil + z, l, u\\right) \\cdot s. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 734, + 903, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $\\lfloor\\rfloor$ is the rounding operation. $l$ and $u$ are the minimum and maximum quantization thresholds while scaling factor $s$ and zero-point $z$ together constitute the quantization parameters. As shown in Equation 5, INT quantization results in an evenly spaced distribution of discrete points. While INT quantization is straightforward to implement, it may be too simplistic for continuous distributions that have significant variations in density, where evenly spaced intervals in INT quantization are not effective. On the contrary,", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18136", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a26f2dab317b9b9951cc7615de13e4369501d93b6d456f95088259c55a95c58.jpg", + "image_caption": [ + "Figure 2. Effect of bit-width reduction on Figure 3. Two loss, and performance degradation representation capacity in AALs (blue) and NALs (orange) under signed FP models across steps. Compared with metric, quantization, evaluated on CelebA dataset. the original loss shows an inverse trend, while the aligned loss remains consistent." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 356, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/53d938ec9cd6120a695018d1bc6c77e257b06fb20289876e696a0b67c248c4a1.jpg", + "image_caption": [ + "Figure 4. The MSE of activations before and after quantization across all AALs under four different strategies, normalized against the baseline of signed FP quantization without zero point (purple)." + ], + "image_footnote": [], + "bbox": [ + 359, + 89, + 625, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6c53f0d4bffe1523270302989a390884558814a6d1ecb309c6a260e6e9bed735.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 90, + 903, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the discrete points under FP quantization are not uniformly spaced in distribution, as shown below [22]:", + "bbox": [ + 89, + 325, + 482, + 357 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf = (- 1) ^ {s} 2 ^ {p - b} \\left(1 + \\frac {d _ {1}}{2} + \\frac {d _ {2}}{2 ^ {2}} + \\dots \\frac {d _ {m}}{2 ^ {m}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 377, + 482, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s$ is the sign bit, $d_{i}$ is the $m$ -bit mantissa, $p$ is the $e$ -bit exponent and $b$ is the bias that serves as both scaling factor and threshold in INT quantization. Previous studies in FP quantization rely on signed FP quantization, where $s$ is set to 1. For an $n$ -bit FP quantization, the bit-width is distributed across the mantissa, exponent, and sign bit, with the condition $m + e + s = n$ . The different combinations of $m$ and $e$ allow FP quantization to represent data in multiple formats with a fixed number of bits, denoted as $E_{i}M_{j}$ , where $i$ -bit exponent and $j$ -bit mantissa are specified. A larger $j$ results in higher precision within each interval, while a larger $i$ expands the range of covered intervals.", + "bbox": [ + 89, + 417, + 482, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There appears to be an inherent suitability of FP quantization for DMs, as the majority of weights and activations follow a normal distribution symmetric around value $= 0$ . This aligns well with the unevenly spaced distribution, where discrete points are dense in the small-value region and sparse in the large-value region, in FP quantization [42]. Additionally, the flexible formatting mechanism allows FP quantization to better accommodate complex distribution scenarios.", + "bbox": [ + 89, + 598, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Barriers to Effective Low-Bit FP Quantization in Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 742, + 482, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to implement high-performing 4-bit diffusion models, we implemented FP quantization that is more compatible with the data distribution of diffusion models, with the search-based strategy in previous work [23]. Building on this, we deployed a LoRA-based fine-tuning strategy to address the performance degradation caused by 4-bit quantization. Despite this, we find that the performance of the 4-bit FP diffusion model remains satisfactory and", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we identify two main issues that we are facing: (1) For FP quantization, although FP quantization works well at 8 bits, it experiences a sharp performance degradation at 4 bits. How to improve the ability of FP quantization to represent data under low-bit quantization? (2) For standard post-training LoRA-based fine-tuning, exhibits instability and suboptimal results when applied to low-bit quantized diffusion models. How can we make LoRA more efficient and accurate in learning the loss information at different denoising timesteps? In the following parts, we will explore the underlying causes of these two issues and the feasible strategies to address them.", + "bbox": [ + 511, + 325, + 906, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 1: Previous signed FP quantization fails to achieve effective low-bit quantization in Activation-Anomalous Layers.", + "text_level": 1, + "bbox": [ + 511, + 510, + 905, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In diffusion models, we observed that the nonlinear activation layer $SiLU$ , defined as $SiLU(\\boldsymbol{x}) = \\frac{\\boldsymbol{x}}{1 + e^{-\\boldsymbol{x}}}$ , is commonly situated between layers.", + "bbox": [ + 511, + 559, + 905, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$SiLU$ causes the abnormal activations for the subsequent layer. As depicted in Panel (b) of Figure 1, all values below 0 are compressed into the range of $[-0.278, 0)$ . In this paper, we refer to layers with such asymmetric activations as Anomalous-Activation-Distribution Layers (AALs) and the other layers as Normal-Activation-Distribution Layers (NALs). Mainstream signed FP quantization typically sets the maximum threshold of the quantizer high to accurately represent normal positive activations. However, this approach results in a significant precision loss when dealing with values below 0. Figure 2 illustrates the representation capacity of FP signed quantization in both NALs and AALs under different bit widths. When the bit width drops below 6 bits, AALs suffer more severe performance degradation compared to NALs, which ultimately leads to the failure in low-bit FP quantization. This phenomenon suggests that mitigating the performance decline in AALs is a critical step towards improving low-bit FP quantization in diffusion models.", + "bbox": [ + 511, + 613, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18137", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 2: The single-LoRA-based strategy is overly simplistic for fine-tuning quantized diffusion models across different timesteps.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous work has focused on adapting LoRA for the quantization of diffusion models but has not fully explored LoRA's performance in the context of denoising. Considering that the denoising process of diffusion models starts with recovering outlines and progresses to restoring details, we question whether the single-LoRA fine-tuning strategy can handle this complexity. In Table 1, we compare the baseline model, which uses a single-LoRA strategy for fine-tuning, with two alternative strategies. The second strategy assigns a separate LoRA for the first and last 50 timesteps, resulting in a significant improvement over the baseline. In contrast, the third strategy also introduces a dual-LoRA strategy, it randomly selects one for each timestep, resulting in much worse performance. These results suggest that applying multiple LoRAs, allocated in a structured manner across timesteps, enhances model performance, while disordered selection of LoRAs could lead to suboptimal results. This motivates us to approach the fine-tuning of quantized diffusion models as a multi-task process and assign multiple LoRAs across different timesteps with a rational approach in allocation.", + "bbox": [ + 91, + 137, + 482, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Observation 3: The MSE of the predicted noise of full-precision and quantized models does not reflect the actual impact of quantization at different timesteps.", + "bbox": [ + 89, + 453, + 482, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In fine-tuning, a commonly used loss function calculates the MSE between the noise predictions of the full-precision and quantized diffusion models, using denoised images from the full-precision model at the previous timestep as inputs:", + "bbox": [ + 89, + 500, + 482, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\boldsymbol {\\varepsilon} _ {\\theta}} ^ {t} = \\left\\| \\boldsymbol {\\varepsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\widehat {\\boldsymbol {\\varepsilon}} _ {\\theta} (\\boldsymbol {x} _ {t}, t) \\right\\| ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 588, + 482, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By observing the variation in this loss during single-LoRA fine-tuning (see Figure 3), we identify an unexpected trend: the loss increases progressively faster as denoising advances. This contradicts the principle of denoising, where image quality should improve with each step, and the impact of predicted noise should diminish over time. By the final step, the impact of quantization on model performance should be negligible, as the predicted noise no longer affects the input. However, the loss is at its maximum, indicating that quantization error is most significant at this stage, contradicting the expectation that its influence should diminish over time.", + "bbox": [ + 89, + 613, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To highlight the discrepancy between the loss and actual quantization errors, we define the performance gap at each step as the difference in denoising quality between the quantized and full-precision models, as the ultimate goal of denoising is to yield high-quality images. We input the previous output image $\\boldsymbol{x}_t$ from the full-precision model and calculate the performance gap between the de", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d673ed7a4ad95207386a65a232d0f92875a9330231a7be884fbb61924e6895d4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBits (W/A)FID ↓
FP32/326.49
Single-LoRA4/419.41
Dual-LoRA (Split Steps in Half)4/417.07
Dual-LoRA (Random Allocation)4/441.96
", + "bbox": [ + 532, + 88, + 887, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. The impact of the number of LoRAs and their allocation across timesteps on the performance of the fine-tuning. The results is evaluated by 4-bit quantization on CelebA dataset.", + "bbox": [ + 511, + 191, + 903, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "noised image $\\boldsymbol{x}_{t-1}$ from the full-precision model and the denoised image $\\widehat{\\boldsymbol{x}}_{t-1}$ from the quantized model, measured by $MSE(\\boldsymbol{x}_{t-1}, \\widehat{\\boldsymbol{x}}_{t-1})$ . As shown in Figure 3, this misalignment between the loss and the actual performance gap leads to deviations in LoRA's learning. This necessitates aligning our loss function with the denoising process during fine-tuning.", + "bbox": [ + 511, + 260, + 903, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 380, + 648, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we explore the issues identified in Section 3.2 and propose corresponding solutions. As illustrated in Figure 5, we introduce a mixup-sign FP quantization framework to address the diverse activation distributions in the first stage. During fine-tuning, the timestep-aware routing mechanism and denoising-factor loss alignment work in tandem to enable high-quality learning, ultimately enabling the realization of optimized 4-bit FP diffusion models.", + "bbox": [ + 511, + 406, + 905, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Mixup-Sign Floating Point Quantization", + "text_level": 1, + "bbox": [ + 511, + 537, + 859, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the challenges of low-bit FP quantization failure in AALs, we leverage FP quantization by allocating more discrete points to areas with high data concentration. Motivated by the half-normal distribution of activations in AALs, we introduce unsigned floating-point quantization. However, as shown in Equation 6, when using unsigned FP quantization with $s$ set to 0, we round all data in the sub-zero range to zero, losing important negative information. To address this, we introduce a zero point in the range of [-0.278, 0) to recover most sub-zero activations. The updated quantization formula becomes:", + "bbox": [ + 511, + 559, + 905, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {u n s i g n} = (- 1) ^ {s} 2 ^ {p - b} \\left(1 + \\frac {d _ {1}}{2} + \\frac {d _ {2}}{2 ^ {2}} + \\dots \\frac {d _ {m}}{2 ^ {m}}\\right) + z, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 750, + 903, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s$ is set to 0 and $z$ is the newly added zero point. By freeing the 1-bit sign bit, which is ineffective in signed FP quantization, and using it as additional exponent / mantissa bit width, we fully utilize the representation capacity. As illustrated in Figure 4, unsigned FP quantization with a zero point significantly improves representation in over $95\\%$ of AALs, compared to traditional signed FP quantization.", + "bbox": [ + 511, + 795, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18138", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/42465d15d6c8f7fb8080a1fedefa5e8e6339dcd958c96419d84c15eba4ea4e57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 88, + 883, + 301 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/732cdcf2680b76c4b427472166d45ba4ed70817c700a3c08e2fe4f54cfeffcf4.jpg", + "image_caption": [ + "Figure 5. The pipeline of our proposed method. UNets are applied to the Mixup-Sign Floating-Point Quantization (MSFP), where distinct floating-point quantization schemes are employed for Anomalous-Activation-Distribution Layers (AALs) and Normal-Activation-Distribution Layers (NALs). During the fine-tuning stage, multiple LoRA modules are introduced, and a timestep-aware routing mechanism is used for dynamic LoRA allocation across different timesteps. Additionally, a denoising-factor alignment technique is employed to align the loss function with quantization-induced performance degradation." + ], + "image_footnote": [], + "bbox": [ + 122, + 305, + 344, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5a2e2cf7780dc2a47983d69b83aaf68c13032b4fb54287b32209c352dacbaea0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 304, + 651, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/94473f604c4f4dd4b84b70a27c3788b0bbcbfba33e6c6a81635e404c58974655.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 304, + 859, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, there are rare cases where performance slightly goes worse due to the diversity of anomalous distributions. Panel (c) of Figure 1 shows that in such cases, the activation distribution may resemble a normal distribution, where signed FP quantization might perform better. Figure 4 further indicates that introducing a zero point into signed FP quantization is unnecessary, offering minimal improvement in a few cases.", + "bbox": [ + 88, + 492, + 482, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the strong performance of unsigned FP quantization with a zero point and the diversity of AALs, we propose a mixup-sign FP quantization framework. During the search-based initialization phase, we use signed FP quantization for NALs and introduce both unsigned FP quantization with a zero point and signed FP quantization for AALs. This approach addresses AAL challenges in low-bit quantization while minimizing computational overhead by adding the zero point only to unsigned FP quantization.", + "bbox": [ + 88, + 613, + 482, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Timestep-Aware Router for LoRA Allocation", + "text_level": 1, + "bbox": [ + 89, + 757, + 472, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Section 3.2, we observe that a single LoRA cannot fully capture all the information across timesteps due to the diverse generative characteristics at different timesteps. We also find that a reasonable allocation of different LoRAs for timesteps will be beneficial to fine-tuning. In this section, we introduce a timestep-aware LoRA allocation method that dynamically assigns optimal LoRA to each timestep, maximizing fine-tuning effectiveness.", + "bbox": [ + 88, + 780, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method relies on a learnable router (illustrated in Figure 5), a module shared across all timesteps. It takes the timestep as input and outputs selection probabilities for each LoRA across UNet layers. For each timestep, the LoRA with the highest probability corresponding to the router's output is inserted into the quantized model for fin-tuning or inference. In a router network, the main components are a time embedding layer and an MLP layer. The time embedding layer, derived from a pre-trained diffusion model, converts the scalar into a $d$ -dimensional embedding. The MLP layer then maps the embedding to a LoRA allocation distribution, where $y$ is the number of quantized layers and $h$ is the size of the LoRA Hub. Using an STE method [1], this distribution is converted into 0/1 probabilities to allocate suitable LoRAs to the diffusion model's quantized layers.", + "bbox": [ + 511, + 492, + 906, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Denoising Factor Aligned Loss", + "text_level": 1, + "bbox": [ + 511, + 742, + 784, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address the mismatch between the actual performance gap and the loss used during quantization, we review the denoising principle depicted in Equation 3 and pinpoint the cause of the mismatch: the time-step-dependent constraint on the predicted noise is not sufficiently accounted for during the denoising process. Therefore, we implement a modification to the loss function based on predicted noise:", + "bbox": [ + 511, + 763, + 905, + 869 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL ^ {t} = \\gamma_ {t} \\cdot L _ {\\varepsilon_ {\\theta}} ^ {t}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 883, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18139", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By introducing $\\gamma_{t}$ , which accurately reflects the utilization of the predicted noise at each time step, we achieve a preliminary alignment between the loss and the actual quantization error, as shown in Figure 3. This facilitates more accurate fine-tuning, leading to better performance recovery in the low-bit diffusion model.", + "bbox": [ + 89, + 90, + 482, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 196, + 215, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 222, + 279, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Models and Metrics. To verify the effectiveness of the proposed method, we evaluate it with two widely adopted diffusion paradigms: DDIM [34] and LDM [30]. For DDIM experiments, we evaluate on CIFAR-10 [16] and CelebA [26]. For LDM, we test unconditional generation on LSUN-Bedroom [41] and LSUN-Church [41] and conditional generation on ImageNet [4]. The performance of the diffusion models is evaluated with Inception Score (IS) [31], Fréchet Inception Distance (FID) [10] and Sliding FID (sFID) [31]. All metrics are evaluated based on 50k samples generated by the DDIM solver [34].", + "bbox": [ + 89, + 244, + 483, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantization Detail. We employ standard layer-wise quantization for both weights and activations. Except for the input and output layers, which are typically set to 8-bit, all other convolution and linear layers are quantized to the target bit-width. Furthermore, we generate 256 samples for the calibration set based on Q-Diffusion [19] for bias initialization and use the method from [2] to obtain the optimal quantization parameters.", + "bbox": [ + 89, + 412, + 483, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline. We compare two main types of quantization methods: PTQ methods (Q-Diffusion [19] and EDADM [38]) and fine-tuning methods (EfficientDM [8] and QuEST [39]). Since these fine-tuning methods involve special settings like non-full-layer quantization, we standardize the settings of EfficientDM in a consistent manner and procure other standardized results from DilateQuant [24]. Comparison with special settings is provided in Appendix.", + "bbox": [ + 89, + 534, + 483, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Quantization Performance", + "text_level": 1, + "bbox": [ + 89, + 666, + 331, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unconditional Generation. Table 2 presents the results of unconditional image generation across multiple datasets. With 6-bit quantization, our method achieves nearly identical performance to full precision. Our 4-bit quantized models achieve SOTA results in all tasks, significantly outperforming previous baseline methods. Notably, on CIFAR-10, our 4-bit quantized model results in an FID score that is only 1.84 worse than full precision, an almost negligible degradation, while previous methods struggle with 4-bit quantization. Compared to the fine-tuning method [8], our 4-bit quantized models improve FID by 32.38, 24.15, and 9.63 on three datasets, respectively, and still maintain remarkable IS. Additionally, we provide the performance of 4-bit and 6-bit quantized models on CelebA in Appendix.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b49d4f1e67b3fa6778bc2b1c403f4ec0574f08edddb692b324b227fcc7ebb321.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskMethodPrec. (W/A)FID ↓IS ↑
CIFAR-10 32x32FP32/324.269.03
Q-Diffusion6/69.198.76
EDA-DM6/626.689.35
EfficientDM6/625.038.08
Ours (h=2)6/64.269.04
DDIM steps = 100Ours (h=4)6/64.239.06
Q-Diffusion4/4N/AN/A
EDA-DM4/4120.244.42
EfficientDM4/438.407.32
Ours (h=2)4/46.028.79
Ours (h=4)4/46.108.90
LSUN (Bedroom) 256x256FP32/323.022.29
Q-Diffusion6/610.102.11
EDA-DM6/610.562.12
QuEST6/610.102.20
EfficientDM6/612.952.57
Ours (h=2)6/68.422.49
Ours (h=4)6/68.402.49
LDM-4 steps = 100 eta = 1.0Q-Diffusion4/4N/AN/A
EDA-DM4/4N/AN/A
QuEST4/4N/AN/A
EfficientDM4/436.362.69
Ours (h=2)4/412.212.47
Ours (h=4)4/412.342.48
LSUN (Church) 256x256FP32/324.062.70
Q-Diffusion6/610.902.47
EDA-DM6/610.762.43
QuEST6/66.832.65
EfficientDM6/67.452.80
Ours (h=2)6/66.242.73
Ours (h=4)6/66.382.73
LDM-8 steps = 100 eta = 0.0Q-Diffusion4/4N/AN/A
EDA-DM4/4N/AN/A
QuEST4/413.032.63
EfficientDM4/418.402.97
Ours (h=2)4/48.812.70
Ours (h=4)4/48.772.71
", + "bbox": [ + 535, + 90, + 893, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantization performance of unconditional generation. 'Prec. (W/A)' denotes the quantization bit-width. 'N/A' denotes failed image generation. $h$ denotes the size of LoRA Hub.", + "bbox": [ + 511, + 722, + 903, + 766 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conditional Generation. Table 3 presents the results of our conditional generation experiments on ImageNet. We observe that FID is not always a reliable metric in this context, as an unexpected trend emerges: as the bit-width of the quantized model decreases, the FID score improves, which contradicts the expected trend. Therefore, our discussion of the ImageNet results focuses on other evaluation metrics. As shown by the IS and sFID, our method achieves", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18140", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b44320fb47421170619870e8e825a81b7066ba6a69d1b107979d2956c7c66a28.jpg", + "image_caption": [ + "Figure 6. A visual comparison of generation results using our method across different quantization bit-widths, with the LSUN-Church dataset as an example." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 483, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performance comparable to that of the full-precision model with 6-bit quantization. Even with 4-bit quantization, we achieve SOTA results in terms of sFID, improving by 7.00 over the previous best method, EfficientDM [8]. Furthermore, our method significantly outperforms two other fin-tuning methods in the IS metric. Visual evaluations further confirm that the generated images maintain high quality, exhibiting clear and coherent content, as shown in Figure 6. More visualization results will be presented in Appendix.", + "bbox": [ + 89, + 347, + 483, + 484 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d3c1c0b4956f47ea59a398b248da49522363d6cc522746e3da290c1e90ed3b52.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPrec. (W/A)sFID ↓FID ↓IS ↑
FP32/327.6711.69364.72
EDA-DM6/68.0211.52360.77
QuEST6/69.368.45310.12
EfficientDM6/66.889.54351.79
Ours (h=2)6/67.4310.10349.91
Ours (h=4)6/66.6510.10351.79
EDA-DM4/436.6620.02204.93
QuEST4/429.2738.4369.58
EfficientDM4/414.4212.73139.45
Ours (h=2)4/47.426.50190.74
Ours (h=4)4/48.237.43177.40
", + "bbox": [ + 106, + 494, + 470, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 772, + 243, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ablation experiments are conducted on the 4-bit quantization using the CelebA dataset, which is challenging for low-bit quantization, further demonstrating the effectiveness of our approach. The baseline uses signed FP quantization combined with single LoRA fine-tuning. As shown in Table 4, all three proposed modules lead to significant performance improvements, with their combination", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2a40426ae05b4a2168b648b643243d91d50736d785496ada4c0e7c8bca69f292.jpg", + "table_caption": [ + "Table 3. Quantization performance of conditional generation for fully-quantized LDM-4 models on ImageNet $256 \\times 256$ with 20 steps. 'Prec. (W/A)' denotes the quantization bit-width. $h$ denotes the size of LoRA Hub." + ], + "table_footnote": [], + "table_body": "
MSFPMethod TALoRADFAPrec. (W/A)FID ↓
XXX4/416.02
✓XX4/49.60
X✓X4/410.66
✓X✓4/48.39
✓✓X4/48.79
✓✓✓4/47.69
", + "bbox": [ + 535, + 88, + 887, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on different modules we proposed. Testing on CelebA dataset with $h = 2$ LoRA Hub size.", + "bbox": [ + 511, + 236, + 905, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "yielding a synergistic effect. The baseline FID score is 9.53 higher than that of the full-precision model (6.49). By applying our technique, we reduce the FID by 8.18 compared to the baseline. More interesting is that we visualize the LoRA allocation distribution learned by the router as shown in Figure 7. We find that the distribution of the router-learned allocation over timesteps is consistent with the finding that the diffusion model focuses on contour generation early and on detail generation later [37].", + "bbox": [ + 511, + 281, + 906, + 417 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80925daab5f8ace3bc40f76b239165ee63f0e19ef2fd0dbcf7d5fd7f43f265d2.jpg", + "image_caption": [ + "Figure 7. Distribution of LoRA allocations over timesteps obtained after router training on different datasets, when $h = 2$ ." + ], + "image_footnote": [], + "bbox": [ + 540, + 439, + 880, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 674, + 633, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we focus on exploring low-bit FP quantization for diffusion models. For model initialization, we innovatively introduce unsigned FP quantization with zero point to address AALs. For the fine-tuning based on the denoising process, we formulate it as a multi-task procedure. We introduce multiple LoRAs along with a router for their allocation at different timesteps, and further align the loss function, originally based on estimated noise, with the actual quantization error. We introduce unsigned FP quantization and achieve 4-bit FP quantized diffusion models. Our FP PTQ-based fine-tuning method sets a new precedent for 4-bit diffusion models, offering insight into the deployment of low-bit diffusion models in the future.", + "bbox": [ + 511, + 704, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18141", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by Shanghai Science and Technology Commission Explorer Program Project (24TS1401300), National Key Research and Development Program of China (No.2022ZD0160101). The computations in this research were performed using the CFFF platform of Fudan University.", + "bbox": [ + 89, + 113, + 483, + 213 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 224, + 187, + 242 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 6", + "[2] Cheng Chen, Christina Giannoula, and Andreas Moshovos. Low-bitwidth floating point quantization for efficient high-quality diffusion models. arXiv preprint arXiv:2408.06995, 2024. 2, 7", + "[3] Huanpeng Chu, Wei Wu, Chengjie Zang, and Kun Yuan. Qncd: Quantization noise correction for diffusion models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10995-11003, 2024. 2", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7", + "[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: efficient finetuning of quantized llms (2023). arXiv preprint arXiv:2305.14314, 52:3982-3992, 2023. 3", + "[6] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. In International Conference on Learning Representations, 2019. 1, 2", + "[7] Wenfeng Feng, Chuzhan Hao, Yuewei Zhang, Yu Han, and Hao Wang. Mixture-of-loras: An efficient multitask tuning method for large language models. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 11371-11380, 2024. 3", + "[8] Yefei He, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Efficientdm: Efficient quantization-aware fine-tuning of low-bit diffusion models. arXiv preprint arXiv:2310.03270, 2023. 1, 2, 7, 8", + "[9] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + ], + "bbox": [ + 94, + 250, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[13] Wei Huang, Xingyu Zheng, Xudong Ma, Haotong Qin, Chengtao Lv, Hong Chen, Jie Luo, Xiaojuan Qi, Xianglong Liu, and Michele Magno. An empirical study of llama3 quantization: From llms to mllms. Visual Intelligence, 2(1): 36, 2024. 1", + "[14] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2704-2713, 2018. 2", + "[15] Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342, 2018. 1", + "[16] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7", + "[17] Andrey Kuzmin, Mart Van Baalen, Yuwei Ren, Markus Nagel, Jorn Peters, and Tijmen Blankevoort. Fp8 quantization: The power of the exponent. Advances in Neural Information Processing Systems, 35:14651-14662, 2022. 2", + "[18] Min Li, Zihao Huang, Lin Chen, Junxing Ren, Miao Jiang, Fengfa Li, Jitao Fu, and Chenghua Gao. Contemporary advances in neural network quantization: A survey. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-10. IEEE, 2024. 1", + "[19] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1, 2, 7", + "[20] Yanjing Li, Sheng Xu, Xianbin Cao, Xiao Sun, and Baochang Zhang. Q-dm: An efficient low-bit quantized diffusion model. Advances in Neural Information Processing Systems, 36, 2024. 1, 2", + "[21] Tailin Liang, John Glossner, Lei Wang, Shaobo Shi, and Xiaotong Zhang. Pruning and quantization for deep neural network acceleration: A survey. Neurocomputing, 461:370-403, 2021. 1", + "[22] Fangxin Liu, Wenbo Zhao, Zhezhi He, Yanzhi Wang, Zongwu Wang, Changzhi Dai, Xiaoyao Liang, and Li Jiang. Improving neural network efficiency via post-training quantization with adaptive floating-point. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5281-5290, 2021. 4", + "[23] Shih-yang Liu, Zechun Liu, Xijie Huang, Pingcheng Dong, and Kwang-Ting Cheng. LIm-fp4: 4-bit floating-point quantized transformers. arXiv preprint arXiv:2310.16836, 2023. 2, 4", + "[24] Xuewen Liu, Zhikai Li, and Qingyi Gu. Dilatequant: Accurate and efficient diffusion quantization via weight dilation. arXiv preprint arXiv:2409.14307, 2024. 7", + "[25] Xuewen Liu, Zhikai Li, Junrui Xiao, and Qingyi Gu. Enhanced distribution alignment for post-training quantization" + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18142", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of diffusion models. arXiv preprint arXiv:2401.04585, 2024. 2", + "[26] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 7", + "[27] Paulius Micikevicius, Dusan Stosic, Neil Burgess, Marius Cornea, Pradeep Dubey, Richard Grisenthwaite, Sangwon Ha, Alexander Heinecke, Patrick Judd, John Kamalu, et al. Fp8 formats for deep learning. arXiv preprint arXiv:2209.05433, 2022. 1", + "[28] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 2", + "[29] NVIDIA. Blackwell platform sets new llm inference records in mlperf inference v4.1, 2024. Available at: https://developer.nvidia.com/blog/nvidia-blackwell-platform-sets-new-llm-inference-records-in-mlperf-inference-v4-1, Accessed: 2024-11-14. 2", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 7", + "[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 7", + "[32] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 1, 2", + "[33] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 7", + "[35] Haojun Sun, Chen Tang, Zhi Wang, Yuan Meng, Xinzhu Ma, Wenwu Zhu, et al. Tmpq-dm: Joint timestep reduction and quantization precision selection for efficient diffusion models. arXiv preprint arXiv:2404.09532, 2024. 2", + "[36] Mart van Baalen, Andrey Kuzmin, Suparna S Nair, Yuwei Ren, Eric Mahurin, Chirag Patel, Sundar Subramanian, Sanghyuk Lee, Markus Nagel, Joseph Soriaga, et al. Fp8 versus int8 for efficient deep learning inference. arXiv preprint arXiv:2303.17951, 2023. 2", + "[37] Binxu Wang and John J Vastola. Diffusion models generate images like painters: an analytical theory of outline first, details later. arXiv preprint arXiv:2303.02490, 2023. 2, 8", + "[38] Changyuan Wang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, and Jiwen Lu. Towards accurate data-free quantization for diffusion models. arXiv preprint arXiv:2305.18723, 2(5), 2023. 7" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, Junchi Yan, and Yan Yan. Quest: Low-bit diffusion model quantization via efficient selective finetuning. arXiv preprint arXiv:2402.03666, 2024. 1, 2, 7", + "[40] Jie Wang, Huanxi Liu, Dawei Feng, Jie Ding, and Bo Ding. Fp4-quantization: Lossless 4bit quantization for large language models. In 2024 IEEE International Conference on Joint Cloud Computing (JCC), pages 61–67. IEEE, 2024. 2", + "[41] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 1, 7", + "[42] Yijia Zhang, Lingran Zhao, Shijie Cao, Sicheng Zhang, Wenqiang Wang, Ting Cao, Fan Yang, Mao Yang, Shang-hang Zhang, and Ningyi Xu. Integer or floating point? new outlooks for low-bit quantization on large language models. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2, 4" + ], + "bbox": [ + 516, + 92, + 905, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18143", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/6042cf5b-516c-4fe8-bfad-ecfc854fc855_model.json b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/6042cf5b-516c-4fe8-bfad-ecfc854fc855_model.json new file mode 100644 index 0000000000000000000000000000000000000000..586005fa87ed9b37a0442987e41a1b3d77e474da --- /dev/null +++ b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/6042cf5b-516c-4fe8-bfad-ecfc854fc855_model.json @@ -0,0 +1,1980 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.13, + 0.844, + 0.177 + ], + "angle": 0, + "content": "Pioneering 4-Bit FP Quantization for Diffusion Models: Mixup-Sign Quantization and Timestep-Aware Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.203, + 0.838, + 0.221 + ], + "angle": 0, + "content": "Maosen Zhao\\(^{1*}\\), Pengtao Chen\\(^{1*}\\), Chong Yu\\(^{2}\\), Yan Wen\\(^{1}\\), Xudong Tan\\(^{1}\\), Tao Chen\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.222, + 0.763, + 0.239 + ], + "angle": 0, + "content": "1 School of Information Science and Technology, Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.239, + 0.746, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\) Academy for Engineering and Technology, Fudan University" + }, + { + "type": "text", + "bbox": [ + 0.292, + 0.26, + 0.7, + 0.274 + ], + "angle": 0, + "content": "20307130202@fudan.edu.cn, eetchen@fudan.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.486, + 0.675 + ], + "angle": 0, + "content": "Model quantization reduces the bit-width of weights and activations, improving memory efficiency and inference speed in diffusion models. However, achieving 4-bit quantization remains challenging. Existing methods, primarily based on integer quantization and post-training quantization fine-tuning, struggle with inconsistent performance. Inspired by the success of floating-point (FP) quantization in large language models, we explore low-bit FP quantization for diffusion models and identify key challenges: the failure of signed FP quantization to handle asymmetric activation distributions, the insufficient consideration of temporal complexity in the denoising process during fine-tuning, and the misalignment between fine-tuning loss and quantization error. To address these challenges, we propose the mixup-sign floating-point quantization (MSFP) framework, first introducing unsigned FP quantization in model quantization, along with timestep-aware LoRA (TALoRA) and denoising-factor loss alignment (DFA), which ensure precise and stable fine-tuning. Extensive experiments show that we are the first to achieve superior performance in 4-bit FP quantization for diffusion models, outperforming existing PTQ fine-tuning methods in 4-bit INT quantization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.702, + 0.222, + 0.717 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Despite the impressive performance of diffusion models (DMs) in image generation [30, 41], their computational and memory demands, particularly for high-resolution outputs, pose significant challenges for deployment on resource-constrained edge devices, highlighting the need for model compression to address these limitations. Model quantization, a key technique within model compression, reduces the bit-width of model weights and activations, typically stored in 32-bit format, to lower precision. This reduction lowers memory usage and accelerates inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.371 + ], + "angle": 0, + "content": "speed. By decreasing the bit-width, quantization improves both temporal and memory efficiency in mainstream models, maintaining robust performance, particularly in resource-constrained environments [13, 18, 21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.909, + 0.599 + ], + "angle": 0, + "content": "The current quantization methods for diffusion models can be broadly categorized into two primary approaches. The first, post-training quantization (PTQ), optimizes the quantization parameters after the model has been trained, typically by minimizing the quantization error [19, 32]. While PTQ is effective for 4-bit quantization of weights, it is limited by its reliance on 8-bit quantization for activations, as further reduction in activation bit-width leads to significant performance degradation. In contrast, quantization-aware training (QAT) integrates quantization into the training process, enabling the model to learn with 4-bit precision from scratch [20]. Although QAT can achieve high-performance 4-bit models, it incurs substantial computational overhead, rendering it less practical for many real-world applications [6, 15]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.78 + ], + "angle": 0, + "content": "To achieve fully quantized 4-bit diffusion models with minimal overhead, fine-tuning has emerged as a promising solution. This approach leverages pre-trained models and adjusts a small subset of parameters to narrow the performance gap between the quantized model and its full-precision counterpart. While some studies have explored fine-tuning for 4-bit quantization in diffusion models [8, 39], these methods often fail to achieve consistent performance under standard configurations (e.g., quantizing all layers but not some). Consequently, developing a universally effective and scalable fine-tuning method for 4-bit quantization in diffusion models remains an open challenge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "To be noticed, existing quantization methods for diffusion models primarily rely on integer (INT) quantization, which has long been the dominant approach. However, recent developments have demonstrated the considerable potential of floating-point (FP) quantization. Compared to INT quantization, FP quantization offers greater flexibility in modeling complex weight and activation distributions [27], leading to improved performance in visual" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.363, + 0.901 + ], + "angle": 0, + "content": "†Corresponding author. *Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18134" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "tasks at 8-bit precision [17, 36], and remarkable results in large language models under 4-bit precision [23, 40]. Moreover, FP quantization provides significant advantages in inference acceleration, with NVIDIA H100 achieving a \\(1.45 \\times\\) speedup with FP8 quantization, outperforming INT8 [29, 42]. Despite these advantages, the application of low-bit FP quantization in diffusion models remains largely unexplored, presenting a promising avenue for future research." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.215, + 0.485, + 0.576 + ], + "angle": 0, + "content": "In summary, to address the challenges in achieving 4-bit diffusion models, we propose a baseline method using a search-based signed FP quantization framework [2, 23] combined with single-LoRA fine-tuning, aiming to realize 4-bit FP quantized diffusion models. Through this exploration, we uncover several findings that are significant for both diffusion model quantization and FP quantization: (1) There exists many layers which exhibit asymmetric distributions, due to the nonlinear activation function \\( SiLU \\). The application of traditional signed FP quantization with a symmetric distribution leads to significant precision loss in the sub-zero area, causing substantial performance degradation after quantization. (2) Fine-tuning is conducted based on the denoising process, which is regarded as a complex task involving the restoration from outlines to details [37]. However, current methods typically apply a single LoRA to fine-tune severely degraded models across all timesteps, which leads to suboptimal learning at certain timesteps. (3) Predicted noise plays a varying role at different timesteps during denoising, which is the key to diffusion models. Arising from the neglect of this variation across timesteps, we identify a mismatch between the impact of quantization and the loss function in current methods, which undermines the effectiveness of fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Facing the challenges in achieving 4-bit FP quantization, we propose constructive strategies: (1) To handle different distribution effectively, we propose a mixup-sign floating-point quantization (MSFP) framework, where unsigned FP quantization with an added zero point leads to a more compatible distribution of discrete points with the anomalous distributions in activations and signed FP quantization continues to exhibit strong representation capacity on other distributions. (2) Realizing that the current single-LoRA fine-tuning approach is in lack of flexibility across timesteps, we introduce timestep-aware LoRA (TALoRA), incorporating multiple LoRAs and a timestep-aware router to dynamically select the appropriate LoRA for each timestep in the denoising process. (3) To further improve the effectiveness of fine-tuning, we introduce a denoising-factor loss alignment (DFA), ensuring the loss function, and the guidance of the fine-tuning, are consistent with the actual quantization deterioration across timesteps." + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.854, + 0.409, + 0.868 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "(i) We are the first to identify that signed FP quantization struggles with asymmetric activations, which" + }, + { + "type": "text", + "bbox": [ + 0.553, + 0.092, + 0.907, + 0.167 + ], + "angle": 0, + "content": "arise from the nonlinear behavior of activation functions. To address this, we introduce the MSFP framework, which is also the first effective application of unsigned FP quantization in quantization, offering a novel approach for achieving low-bit quantization." + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.168, + 0.905, + 0.272 + ], + "angle": 0, + "content": "(ii) For the fine-tuning of low-bit diffusion models, which is based on the denoising process, we precisely define it as a multi-task process across timesteps and introduce an efficient TALoRA module. Furthermore, we improve the alignment of loss function with quantization error via DFA strategy, enabling stable and reliable fine-tuning to achieve low-bit diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.273, + 0.905, + 0.378 + ], + "angle": 0, + "content": "(iii) We focus on identifying and eliminating three major barriers to effective low-bit FP quantization in diffusion models. Extensive experiments on DDIM and LDM demonstrate that our work achieves SOTA results for 4-bit quantization. The proposed MSFP, TALoRA and DFA have greatly advanced the progress of low-bit quantization in diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.397, + 0.655, + 0.412 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.423, + 0.78, + 0.438 + ], + "angle": 0, + "content": "2.1. Diffusion Model Quantization" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.447, + 0.907, + 0.733 + ], + "angle": 0, + "content": "There are two main approaches in the quantization of diffusion models: QAT [6, 14] and PTQ [28]. QAT is particularly effective for low-bit quantization while it retransmits the model from scratch, consuming extensive computational resources and time [20]. In contrast, PTQ offers greater time efficiency, making it more practical for large models in real-world applications. Recent advancements in PTQ have concentrated on optimizing calibration datasets to enhance reconstruction accuracy [9, 19, 25, 32] and mitigating quantization errors stemming from the temporal and structural properties of diffusion models [3, 19, 33, 35]. To further achieve fully 4-bit quantization, finetuning techniques has been introduced in PTQ-based quantization. EfficientDM [8] develops a LoRA-based finetuning framework, while QuEST [39] focuses on optimizing quantization-unfriendly activations. Both works stagnate in addressing data distribution during fine-tuning, failing to account for the specific challenges involved in finetuning within the context of diffusion model quantization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Meanwhile, recent advancements in FP quantization have made significant strides in model quantization [17, 23], highlighting its potential for diffusion models. While existing low-bit quantization methods perform well for linear models, applying FP quantization to convolutional diffusion models remains more challenging. To date, only one study has explored 8-bit activation quantization in diffusion models using a basic search-based approach [2], with no research addressing lower-bit quantization. This underscores both the challenges and the untapped potential of achieving 4-bit FP quantization for diffusion models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18135" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.365, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.092, + 0.634, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.092, + 0.9, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.264, + 0.908, + 0.309 + ], + "angle": 0, + "content": "Figure 1. The activation distributions in NALs and AALs, results on the CelebA dataset. (a) The paradigm of NALs with symmetric activations. (b) The typical paradigm of AALs with asymmetric activations, where unsigned FP quantization is more suitable. (c) The infrequent paradigm of AALs with relatively symmetric activations, where either signed or unsigned FP quantization could be applicable." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.332, + 0.379, + 0.349 + ], + "angle": 0, + "content": "2.2. Parameter-Efficient Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.484, + 0.55 + ], + "angle": 0, + "content": "Parameter-efficient fine-tuning (PEFT) has emerged as an effective alternative to full model fine-tuning, focusing on adjusting only a subset of parameters while keeping the majority frozen, thereby reducing storage overhead. Low-rank adapters (LoRA) [12], originally developed for large language models, have become one of the most widely used PEFT methods. Leveraging LoRA's strong transferability, QLoRA [5] can be effectively applied to fine-tune low-bit diffusion models. However, previous LoRA-based fine-tuning is suboptimal in practice. In this paper, we make a further exploration on this and incorporate timestep-level adaptation inspired by MoELoRA [7], enhancing performance of low-bit quantized DMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.564, + 0.334, + 0.581 + ], + "angle": 0, + "content": "3. Challenges & Exploration" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.221, + 0.605 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Diffusion Models. The diffusion model is a new generation framework that completes learning by adding noise and completes generation in a denoising manner. As for the forward process, the noise is injected into ground-truth images \\( \\boldsymbol{x}_0 \\) at random timesteps. This enables the DMs to learn the distributions of noise through noisy images \\( \\boldsymbol{x}_t \\), which can be obtained as follows [11]:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.73, + 0.482, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\varepsilon . \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.754, + 0.483, + 0.784 + ], + "angle": 0, + "content": "Here \\(\\varepsilon\\) represents a standard Gaussian noise and \\(\\overline{\\alpha}_t\\) is the accumulated noise intensity, calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.242, + 0.795, + 0.483, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {t} \\alpha_ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha_{s}\\) governs the noise intensity under each timestep. Once the DMs are well-trained, the Gaussian noise image will be inputted to the DMs, and undergo the iterative denoising process. Specifically, the noise can be predicted" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.907, + 0.38 + ], + "angle": 0, + "content": "by DMs and used to obtain the image \\( \\boldsymbol{x}_t \\) under timestep \\( t \\), which could range from \\( T \\) to 0, with the objective of obtaining the image \\( \\boldsymbol{x}_{t-1} \\) in a superior quality:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.398, + 0.907, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\overline {{\\alpha}} _ {t}}} \\cdot \\boldsymbol {\\varepsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right) + \\sigma_ {t} \\boldsymbol {\\delta}, (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.498 + ], + "angle": 0, + "content": "where \\(\\varepsilon_{\\theta}(\\boldsymbol{x}_t,t)\\) is the predicted noise at timestep \\(t\\) and \\(\\delta\\) is a newly added noise with the factor \\(\\sigma_t\\) to ensure diverse results. Here we define a denoising factor \\(\\gamma_t\\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.51, + 0.905, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {t} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\cdot \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.906, + 0.589 + ], + "angle": 0, + "content": "According to Equation 3, \\(\\gamma_{t}\\) indicates the impact of the prediction noise under timestep \\(t\\). The greater the factor \\(\\gamma_{t}\\), the stronger the predicted noise effect in the denoising." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.727 + ], + "angle": 0, + "content": "Model Quantization. The fundamental paradigm of model quantization is the process of transforming a continuous data distribution into a finite set of discrete points. Consequently, the quality of the quantized model is inextricably related to the distribution of discrete points. According to the discrete point type, quantization is defined as two categories, widely-used INT quantization and emerging FP quantization. Analogous to the process of INT quantization, a floating-point vector \\( x \\) can be quantized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.735, + 0.905, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\boldsymbol {x}} = \\operatorname {C l i p} \\left(\\left\\lfloor \\frac {\\boldsymbol {x}}{s} \\right\\rceil + z, l, u\\right) \\cdot s. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Here \\(\\lfloor\\rfloor\\) is the rounding operation. \\(l\\) and \\(u\\) are the minimum and maximum quantization thresholds while scaling factor \\(s\\) and zero-point \\(z\\) together constitute the quantization parameters. As shown in Equation 5, INT quantization results in an evenly spaced distribution of discrete points. While INT quantization is straightforward to implement, it may be too simplistic for continuous distributions that have significant variations in density, where evenly spaced intervals in INT quantization are not effective. On the contrary," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18136" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.357, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.09, + 0.627, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.091, + 0.904, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.231, + 0.63, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Effect of bit-width reduction on Figure 3. Two loss, and performance degradation representation capacity in AALs (blue) and NALs (orange) under signed FP models across steps. Compared with metric, quantization, evaluated on CelebA dataset. the original loss shows an inverse trend, while the aligned loss remains consistent." + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.231, + 0.906, + 0.301 + ], + "angle": 0, + "content": "Figure 4. The MSE of activations before and after quantization across all AALs under four different strategies, normalized against the baseline of signed FP quantization without zero point (purple)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.483, + 0.358 + ], + "angle": 0, + "content": "the discrete points under FP quantization are not uniformly spaced in distribution, as shown below [22]:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.378, + 0.483, + 0.412 + ], + "angle": 0, + "content": "\\[\nf = (- 1) ^ {s} 2 ^ {p - b} \\left(1 + \\frac {d _ {1}}{2} + \\frac {d _ {2}}{2 ^ {2}} + \\dots \\frac {d _ {m}}{2 ^ {m}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.483, + 0.599 + ], + "angle": 0, + "content": "where \\( s \\) is the sign bit, \\( d_{i} \\) is the \\( m \\)-bit mantissa, \\( p \\) is the \\( e \\)-bit exponent and \\( b \\) is the bias that serves as both scaling factor and threshold in INT quantization. Previous studies in FP quantization rely on signed FP quantization, where \\( s \\) is set to 1. For an \\( n \\)-bit FP quantization, the bit-width is distributed across the mantissa, exponent, and sign bit, with the condition \\( m + e + s = n \\). The different combinations of \\( m \\) and \\( e \\) allow FP quantization to represent data in multiple formats with a fixed number of bits, denoted as \\( E_{i}M_{j} \\), where \\( i \\)-bit exponent and \\( j \\)-bit mantissa are specified. A larger \\( j \\) results in higher precision within each interval, while a larger \\( i \\) expands the range of covered intervals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.734 + ], + "angle": 0, + "content": "There appears to be an inherent suitability of FP quantization for DMs, as the majority of weights and activations follow a normal distribution symmetric around value \\( = 0 \\). This aligns well with the unevenly spaced distribution, where discrete points are dense in the small-value region and sparse in the large-value region, in FP quantization [42]. Additionally, the flexible formatting mechanism allows FP quantization to better accommodate complex distribution scenarios." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.773 + ], + "angle": 0, + "content": "3.2. Barriers to Effective Low-Bit FP Quantization in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In order to implement high-performing 4-bit diffusion models, we implemented FP quantization that is more compatible with the data distribution of diffusion models, with the search-based strategy in previous work [23]. Building on this, we deployed a LoRA-based fine-tuning strategy to address the performance degradation caused by 4-bit quantization. Despite this, we find that the performance of the 4-bit FP diffusion model remains satisfactory and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.509 + ], + "angle": 0, + "content": "we identify two main issues that we are facing: (1) For FP quantization, although FP quantization works well at 8 bits, it experiences a sharp performance degradation at 4 bits. How to improve the ability of FP quantization to represent data under low-bit quantization? (2) For standard post-training LoRA-based fine-tuning, exhibits instability and suboptimal results when applied to low-bit quantized diffusion models. How can we make LoRA more efficient and accurate in learning the loss information at different denoising timesteps? In the following parts, we will explore the underlying causes of these two issues and the feasible strategies to address them." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.511, + 0.906, + 0.557 + ], + "angle": 0, + "content": "Observation 1: Previous signed FP quantization fails to achieve effective low-bit quantization in Activation-Anomalous Layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.906, + 0.612 + ], + "angle": 0, + "content": "In diffusion models, we observed that the nonlinear activation layer \\( SiLU \\), defined as \\( SiLU(\\boldsymbol{x}) = \\frac{\\boldsymbol{x}}{1 + e^{-\\boldsymbol{x}}} \\), is commonly situated between layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.9 + ], + "angle": 0, + "content": "\\(SiLU\\) causes the abnormal activations for the subsequent layer. As depicted in Panel (b) of Figure 1, all values below 0 are compressed into the range of \\([-0.278, 0)\\). In this paper, we refer to layers with such asymmetric activations as Anomalous-Activation-Distribution Layers (AALs) and the other layers as Normal-Activation-Distribution Layers (NALs). Mainstream signed FP quantization typically sets the maximum threshold of the quantizer high to accurately represent normal positive activations. However, this approach results in a significant precision loss when dealing with values below 0. Figure 2 illustrates the representation capacity of FP signed quantization in both NALs and AALs under different bit widths. When the bit width drops below 6 bits, AALs suffer more severe performance degradation compared to NALs, which ultimately leads to the failure in low-bit FP quantization. This phenomenon suggests that mitigating the performance decline in AALs is a critical step towards improving low-bit FP quantization in diffusion models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18137" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "Observation 2: The single-LoRA-based strategy is overly simplistic for fine-tuning quantized diffusion models across different timesteps." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.138, + 0.483, + 0.452 + ], + "angle": 0, + "content": "Previous work has focused on adapting LoRA for the quantization of diffusion models but has not fully explored LoRA's performance in the context of denoising. Considering that the denoising process of diffusion models starts with recovering outlines and progresses to restoring details, we question whether the single-LoRA fine-tuning strategy can handle this complexity. In Table 1, we compare the baseline model, which uses a single-LoRA strategy for fine-tuning, with two alternative strategies. The second strategy assigns a separate LoRA for the first and last 50 timesteps, resulting in a significant improvement over the baseline. In contrast, the third strategy also introduces a dual-LoRA strategy, it randomly selects one for each timestep, resulting in much worse performance. These results suggest that applying multiple LoRAs, allocated in a structured manner across timesteps, enhances model performance, while disordered selection of LoRAs could lead to suboptimal results. This motivates us to approach the fine-tuning of quantized diffusion models as a multi-task process and assign multiple LoRAs across different timesteps with a rational approach in allocation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.483, + 0.5 + ], + "angle": 0, + "content": "Observation 3: The MSE of the predicted noise of full-precision and quantized models does not reflect the actual impact of quantization at different timesteps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.576 + ], + "angle": 0, + "content": "In fine-tuning, a commonly used loss function calculates the MSE between the noise predictions of the full-precision and quantized diffusion models, using denoised images from the full-precision model at the previous timestep as inputs:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.589, + 0.483, + 0.609 + ], + "angle": 0, + "content": "\\[\nL _ {\\boldsymbol {\\varepsilon} _ {\\theta}} ^ {t} = \\left\\| \\boldsymbol {\\varepsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\widehat {\\boldsymbol {\\varepsilon}} _ {\\theta} (\\boldsymbol {x} _ {t}, t) \\right\\| ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.794 + ], + "angle": 0, + "content": "By observing the variation in this loss during single-LoRA fine-tuning (see Figure 3), we identify an unexpected trend: the loss increases progressively faster as denoising advances. This contradicts the principle of denoising, where image quality should improve with each step, and the impact of predicted noise should diminish over time. By the final step, the impact of quantization on model performance should be negligible, as the predicted noise no longer affects the input. However, the loss is at its maximum, indicating that quantization error is most significant at this stage, contradicting the expectation that its influence should diminish over time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To highlight the discrepancy between the loss and actual quantization errors, we define the performance gap at each step as the difference in denoising quality between the quantized and full-precision models, as the ultimate goal of denoising is to yield high-quality images. We input the previous output image \\( \\boldsymbol{x}_t \\) from the full-precision model and calculate the performance gap between the de" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.089, + 0.888, + 0.182 + ], + "angle": 0, + "content": "
MethodBits (W/A)FID ↓
FP32/326.49
Single-LoRA4/419.41
Dual-LoRA (Split Steps in Half)4/417.07
Dual-LoRA (Random Allocation)4/441.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "Table 1. The impact of the number of LoRAs and their allocation across timesteps on the performance of the fine-tuning. The results is evaluated by 4-bit quantization on CelebA dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.905, + 0.367 + ], + "angle": 0, + "content": "noised image \\( \\boldsymbol{x}_{t-1} \\) from the full-precision model and the denoised image \\( \\widehat{\\boldsymbol{x}}_{t-1} \\) from the quantized model, measured by \\( MSE(\\boldsymbol{x}_{t-1}, \\widehat{\\boldsymbol{x}}_{t-1}) \\). As shown in Figure 3, this misalignment between the loss and the actual performance gap leads to deviations in LoRA's learning. This necessitates aligning our loss function with the denoising process during fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.649, + 0.399 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.906, + 0.527 + ], + "angle": 0, + "content": "In this section, we explore the issues identified in Section 3.2 and propose corresponding solutions. As illustrated in Figure 5, we introduce a mixup-sign FP quantization framework to address the diverse activation distributions in the first stage. During fine-tuning, the timestep-aware routing mechanism and denoising-factor loss alignment work in tandem to enable high-quality learning, ultimately enabling the realization of optimized 4-bit FP diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.538, + 0.86, + 0.554 + ], + "angle": 0, + "content": "4.1. Mixup-Sign Floating Point Quantization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.906, + 0.726 + ], + "angle": 0, + "content": "To address the challenges of low-bit FP quantization failure in AALs, we leverage FP quantization by allocating more discrete points to areas with high data concentration. Motivated by the half-normal distribution of activations in AALs, we introduce unsigned floating-point quantization. However, as shown in Equation 6, when using unsigned FP quantization with \\( s \\) set to 0, we round all data in the sub-zero range to zero, losing important negative information. To address this, we introduce a zero point in the range of [-0.278, 0) to recover most sub-zero activations. The updated quantization formula becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.75, + 0.905, + 0.795 + ], + "angle": 0, + "content": "\\[\nf _ {u n s i g n} = (- 1) ^ {s} 2 ^ {p - b} \\left(1 + \\frac {d _ {1}}{2} + \\frac {d _ {2}}{2 ^ {2}} + \\dots \\frac {d _ {m}}{2 ^ {m}}\\right) + z, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( s \\) is set to 0 and \\( z \\) is the newly added zero point. By freeing the 1-bit sign bit, which is ineffective in signed FP quantization, and using it as additional exponent / mantissa bit width, we fully utilize the representation capacity. As illustrated in Figure 4, unsigned FP quantization with a zero point significantly improves representation in over \\( 95\\% \\) of AALs, compared to traditional signed FP quantization." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18138" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.885, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.306, + 0.345, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.305, + 0.652, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.305, + 0.861, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.397, + 0.908, + 0.468 + ], + "angle": 0, + "content": "Figure 5. The pipeline of our proposed method. UNets are applied to the Mixup-Sign Floating-Point Quantization (MSFP), where distinct floating-point quantization schemes are employed for Anomalous-Activation-Distribution Layers (AALs) and Normal-Activation-Distribution Layers (NALs). During the fine-tuning stage, multiple LoRA modules are introduced, and a timestep-aware routing mechanism is used for dynamic LoRA allocation across different timesteps. Additionally, a denoising-factor alignment technique is employed to align the loss function with quantization-induced performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.493, + 0.483, + 0.612 + ], + "angle": 0, + "content": "However, there are rare cases where performance slightly goes worse due to the diversity of anomalous distributions. Panel (c) of Figure 1 shows that in such cases, the activation distribution may resemble a normal distribution, where signed FP quantization might perform better. Figure 4 further indicates that introducing a zero point into signed FP quantization is unnecessary, offering minimal improvement in a few cases." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Given the strong performance of unsigned FP quantization with a zero point and the diversity of AALs, we propose a mixup-sign FP quantization framework. During the search-based initialization phase, we use signed FP quantization for NALs and introduce both unsigned FP quantization with a zero point and signed FP quantization for AALs. This approach addresses AAL challenges in low-bit quantization while minimizing computational overhead by adding the zero point only to unsigned FP quantization." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.473, + 0.774 + ], + "angle": 0, + "content": "4.2. Timestep-Aware Router for LoRA Allocation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In Section 3.2, we observe that a single LoRA cannot fully capture all the information across timesteps due to the diverse generative characteristics at different timesteps. We also find that a reasonable allocation of different LoRAs for timesteps will be beneficial to fine-tuning. In this section, we introduce a timestep-aware LoRA allocation method that dynamically assigns optimal LoRA to each timestep, maximizing fine-tuning effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Our method relies on a learnable router (illustrated in Figure 5), a module shared across all timesteps. It takes the timestep as input and outputs selection probabilities for each LoRA across UNet layers. For each timestep, the LoRA with the highest probability corresponding to the router's output is inserted into the quantized model for fin-tuning or inference. In a router network, the main components are a time embedding layer and an MLP layer. The time embedding layer, derived from a pre-trained diffusion model, converts the scalar into a \\(d\\)-dimensional embedding. The MLP layer then maps the embedding to a LoRA allocation distribution, where \\(y\\) is the number of quantized layers and \\(h\\) is the size of the LoRA Hub. Using an STE method [1], this distribution is converted into 0/1 probabilities to allocate suitable LoRAs to the diffusion model's quantized layers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.785, + 0.76 + ], + "angle": 0, + "content": "4.3. Denoising Factor Aligned Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.871 + ], + "angle": 0, + "content": "To address the mismatch between the actual performance gap and the loss used during quantization, we review the denoising principle depicted in Equation 3 and pinpoint the cause of the mismatch: the time-step-dependent constraint on the predicted noise is not sufficiently accounted for during the denoising process. Therefore, we implement a modification to the loss function based on predicted noise:" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.885, + 0.906, + 0.904 + ], + "angle": 0, + "content": "\\[\nL ^ {t} = \\gamma_ {t} \\cdot L _ {\\varepsilon_ {\\theta}} ^ {t}. \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18139" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.182 + ], + "angle": 0, + "content": "By introducing \\(\\gamma_{t}\\), which accurately reflects the utilization of the predicted noise at each time step, we achieve a preliminary alignment between the loss and the actual quantization error, as shown in Figure 3. This facilitates more accurate fine-tuning, leading to better performance recovery in the low-bit diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.198, + 0.216, + 0.215 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.223, + 0.281, + 0.24 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.484, + 0.412 + ], + "angle": 0, + "content": "Models and Metrics. To verify the effectiveness of the proposed method, we evaluate it with two widely adopted diffusion paradigms: DDIM [34] and LDM [30]. For DDIM experiments, we evaluate on CIFAR-10 [16] and CelebA [26]. For LDM, we test unconditional generation on LSUN-Bedroom [41] and LSUN-Church [41] and conditional generation on ImageNet [4]. The performance of the diffusion models is evaluated with Inception Score (IS) [31], Fréchet Inception Distance (FID) [10] and Sliding FID (sFID) [31]. All metrics are evaluated based on 50k samples generated by the DDIM solver [34]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Quantization Detail. We employ standard layer-wise quantization for both weights and activations. Except for the input and output layers, which are typically set to 8-bit, all other convolution and linear layers are quantized to the target bit-width. Furthermore, we generate 256 samples for the calibration set based on Q-Diffusion [19] for bias initialization and use the method from [2] to obtain the optimal quantization parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Baseline. We compare two main types of quantization methods: PTQ methods (Q-Diffusion [19] and EDADM [38]) and fine-tuning methods (EfficientDM [8] and QuEST [39]). Since these fine-tuning methods involve special settings like non-full-layer quantization, we standardize the settings of EfficientDM in a consistent manner and procure other standardized results from DilateQuant [24]. Comparison with special settings is provided in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.332, + 0.683 + ], + "angle": 0, + "content": "5.2. Quantization Performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Unconditional Generation. Table 2 presents the results of unconditional image generation across multiple datasets. With 6-bit quantization, our method achieves nearly identical performance to full precision. Our 4-bit quantized models achieve SOTA results in all tasks, significantly outperforming previous baseline methods. Notably, on CIFAR-10, our 4-bit quantized model results in an FID score that is only 1.84 worse than full precision, an almost negligible degradation, while previous methods struggle with 4-bit quantization. Compared to the fine-tuning method [8], our 4-bit quantized models improve FID by 32.38, 24.15, and 9.63 on three datasets, respectively, and still maintain remarkable IS. Additionally, we provide the performance of 4-bit and 6-bit quantized models on CelebA in Appendix." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.091, + 0.895, + 0.711 + ], + "angle": 0, + "content": "
TaskMethodPrec. (W/A)FID ↓IS ↑
CIFAR-10 32x32FP32/324.269.03
Q-Diffusion6/69.198.76
EDA-DM6/626.689.35
EfficientDM6/625.038.08
Ours (h=2)6/64.269.04
DDIM steps = 100Ours (h=4)6/64.239.06
Q-Diffusion4/4N/AN/A
EDA-DM4/4120.244.42
EfficientDM4/438.407.32
Ours (h=2)4/46.028.79
Ours (h=4)4/46.108.90
LSUN (Bedroom) 256x256FP32/323.022.29
Q-Diffusion6/610.102.11
EDA-DM6/610.562.12
QuEST6/610.102.20
EfficientDM6/612.952.57
Ours (h=2)6/68.422.49
Ours (h=4)6/68.402.49
LDM-4 steps = 100 eta = 1.0Q-Diffusion4/4N/AN/A
EDA-DM4/4N/AN/A
QuEST4/4N/AN/A
EfficientDM4/436.362.69
Ours (h=2)4/412.212.47
Ours (h=4)4/412.342.48
LSUN (Church) 256x256FP32/324.062.70
Q-Diffusion6/610.902.47
EDA-DM6/610.762.43
QuEST6/66.832.65
EfficientDM6/67.452.80
Ours (h=2)6/66.242.73
Ours (h=4)6/66.382.73
LDM-8 steps = 100 eta = 0.0Q-Diffusion4/4N/AN/A
EDA-DM4/4N/AN/A
QuEST4/413.032.63
EfficientDM4/418.402.97
Ours (h=2)4/48.812.70
Ours (h=4)4/48.772.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.723, + 0.905, + 0.767 + ], + "angle": 0, + "content": "Table 2. Quantization performance of unconditional generation. 'Prec. (W/A)' denotes the quantization bit-width. 'N/A' denotes failed image generation. \\( h \\) denotes the size of LoRA Hub." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Conditional Generation. Table 3 presents the results of our conditional generation experiments on ImageNet. We observe that FID is not always a reliable metric in this context, as an unexpected trend emerges: as the bit-width of the quantized model decreases, the FID score improves, which contradicts the expected trend. Therefore, our discussion of the ImageNet results focuses on other evaluation metrics. As shown by the IS and sFID, our method achieves" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18140" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.485, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.293, + 0.484, + 0.337 + ], + "angle": 0, + "content": "Figure 6. A visual comparison of generation results using our method across different quantization bit-widths, with the LSUN-Church dataset as an example." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.484, + 0.485 + ], + "angle": 0, + "content": "performance comparable to that of the full-precision model with 6-bit quantization. Even with 4-bit quantization, we achieve SOTA results in terms of sFID, improving by 7.00 over the previous best method, EfficientDM [8]. Furthermore, our method significantly outperforms two other fin-tuning methods in the IS metric. Visual evaluations further confirm that the generated images maintain high quality, exhibiting clear and coherent content, as shown in Figure 6. More visualization results will be presented in Appendix." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.496, + 0.472, + 0.691 + ], + "angle": 0, + "content": "
MethodPrec. (W/A)sFID ↓FID ↓IS ↑
FP32/327.6711.69364.72
EDA-DM6/68.0211.52360.77
QuEST6/69.368.45310.12
EfficientDM6/66.889.54351.79
Ours (h=2)6/67.4310.10349.91
Ours (h=4)6/66.6510.10351.79
EDA-DM4/436.6620.02204.93
QuEST4/429.2738.4369.58
EfficientDM4/414.4212.73139.45
Ours (h=2)4/47.426.50190.74
Ours (h=4)4/48.237.43177.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.699, + 0.484, + 0.755 + ], + "angle": 0, + "content": "Table 3. Quantization performance of conditional generation for fully-quantized LDM-4 models on ImageNet \\(256 \\times 256\\) with 20 steps. 'Prec. (W/A)' denotes the quantization bit-width. \\(h\\) denotes the size of LoRA Hub." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.245, + 0.79 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The ablation experiments are conducted on the 4-bit quantization using the CelebA dataset, which is challenging for low-bit quantization, further demonstrating the effectiveness of our approach. The baseline uses signed FP quantization combined with single LoRA fine-tuning. As shown in Table 4, all three proposed modules lead to significant performance improvements, with their combination" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.089, + 0.888, + 0.227 + ], + "angle": 0, + "content": "
MSFPMethod TALoRADFAPrec. (W/A)FID ↓
XXX4/416.02
✓XX4/49.60
X✓X4/410.66
✓X✓4/48.39
✓✓X4/48.79
✓✓✓4/47.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.237, + 0.906, + 0.265 + ], + "angle": 0, + "content": "Table 4. Ablation study on different modules we proposed. Testing on CelebA dataset with \\( h = 2 \\) LoRA Hub size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.907, + 0.419 + ], + "angle": 0, + "content": "yielding a synergistic effect. The baseline FID score is 9.53 higher than that of the full-precision model (6.49). By applying our technique, we reduce the FID by 8.18 compared to the baseline. More interesting is that we visualize the LoRA allocation distribution learned by the router as shown in Figure 7. We find that the distribution of the router-learned allocation over timesteps is consistent with the finding that the diffusion model focuses on contour generation early and on detail generation later [37]." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.44, + 0.882, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.616, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Figure 7. Distribution of LoRA allocations over timesteps obtained after router training on different datasets, when \\( h = 2 \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.675, + 0.634, + 0.69 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this paper, we focus on exploring low-bit FP quantization for diffusion models. For model initialization, we innovatively introduce unsigned FP quantization with zero point to address AALs. For the fine-tuning based on the denoising process, we formulate it as a multi-task procedure. We introduce multiple LoRAs along with a router for their allocation at different timesteps, and further align the loss function, originally based on estimated noise, with the actual quantization error. We introduce unsigned FP quantization and achieve 4-bit FP quantized diffusion models. Our FP PTQ-based fine-tuning method sets a new precedent for 4-bit diffusion models, offering insight into the deployment of low-bit diffusion models in the future." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18141" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.214 + ], + "angle": 0, + "content": "This work is supported by Shanghai Science and Technology Commission Explorer Program Project (24TS1401300), National Key Research and Development Program of China (No.2022ZD0160101). The computations in this research were performed using the CFFF platform of Fudan University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.226, + 0.188, + 0.243 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.251, + 0.483, + 0.306 + ], + "angle": 0, + "content": "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.308, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[2] Cheng Chen, Christina Giannoula, and Andreas Moshovos. Low-bitwidth floating point quantization for efficient high-quality diffusion models. arXiv preprint arXiv:2408.06995, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.365, + 0.483, + 0.42 + ], + "angle": 0, + "content": "[3] Huanpeng Chu, Wei Wu, Chengjie Zang, and Kun Yuan. Qncd: Quantization noise correction for diffusion models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10995-11003, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.422, + 0.484, + 0.478 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.479, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: efficient finetuning of quantized llms (2023). arXiv preprint arXiv:2305.14314, 52:3982-3992, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.535, + 0.484, + 0.59 + ], + "angle": 0, + "content": "[6] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. In International Conference on Learning Representations, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.485, + 0.675 + ], + "angle": 0, + "content": "[7] Wenfeng Feng, Chuzhan Hao, Yuewei Zhang, Yu Han, and Hao Wang. Mixture-of-loras: An efficient multitask tuning method for large language models. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 11371-11380, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[8] Yefei He, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Efficientdm: Efficient quantization-aware fine-tuning of low-bit diffusion models. arXiv preprint arXiv:2310.03270, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[9] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.789, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.251, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[13] Wei Huang, Xingyu Zheng, Xudong Ma, Haotong Qin, Chengtao Lv, Hong Chen, Jie Luo, Xiaojuan Qi, Xianglong Liu, and Michele Magno. An empirical study of llama3 quantization: From llms to mllms. Visual Intelligence, 2(1): 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.304 + ], + "angle": 0, + "content": "[14] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2704-2713, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[15] Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[17] Andrey Kuzmin, Mart Van Baalen, Yuwei Ren, Markus Nagel, Jorn Peters, and Tijmen Blankevoort. Fp8 quantization: The power of the exponent. Advances in Neural Information Processing Systems, 35:14651-14662, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[18] Min Li, Zihao Huang, Lin Chen, Junxing Ren, Miao Jiang, Fengfa Li, Jitao Fu, and Chenghua Gao. Contemporary advances in neural network quantization: A survey. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-10. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[19] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[20] Yanjing Li, Sheng Xu, Xianbin Cao, Xiao Sun, and Baochang Zhang. Q-dm: An efficient low-bit quantized diffusion model. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[21] Tailin Liang, John Glossner, Lei Wang, Shaobo Shi, and Xiaotong Zhang. Pruning and quantization for deep neural network acceleration: A survey. Neurocomputing, 461:370-403, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.906, + 0.772 + ], + "angle": 0, + "content": "[22] Fangxin Liu, Wenbo Zhao, Zhezhi He, Yanzhi Wang, Zongwu Wang, Changzhi Dai, Xiaoyao Liang, and Li Jiang. Improving neural network efficiency via post-training quantization with adaptive floating-point. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5281-5290, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[23] Shih-yang Liu, Zechun Liu, Xijie Huang, Pingcheng Dong, and Kwang-Ting Cheng. LIm-fp4: 4-bit floating-point quantized transformers. arXiv preprint arXiv:2310.16836, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Xuewen Liu, Zhikai Li, and Qingyi Gu. Dilatequant: Accurate and efficient diffusion quantization via weight dilation. arXiv preprint arXiv:2409.14307, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Xuewen Liu, Zhikai Li, Junrui Xiao, and Qingyi Gu. Enhanced distribution alignment for post-training quantization" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18142" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "of diffusion models. arXiv preprint arXiv:2401.04585, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[26] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[27] Paulius Micikevicius, Dusan Stosic, Neil Burgess, Marius Cornea, Pradeep Dubey, Richard Grisenthwaite, Sangwon Ha, Alexander Heinecke, Patrick Judd, John Kamalu, et al. Fp8 formats for deep learning. arXiv preprint arXiv:2209.05433, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[28] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[29] NVIDIA. Blackwell platform sets new llm inference records in mlperf inference v4.1, 2024. Available at: https://developer.nvidia.com/blog/nvidia-blackwell-platform-sets-new-llm-inference-records-in-mlperf-inference-v4-1, Accessed: 2024-11-14. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[32] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[33] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[35] Haojun Sun, Chen Tang, Zhi Wang, Yuan Meng, Xinzhu Ma, Wenwu Zhu, et al. Tmpq-dm: Joint timestep reduction and quantization precision selection for efficient diffusion models. arXiv preprint arXiv:2404.09532, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[36] Mart van Baalen, Andrey Kuzmin, Suparna S Nair, Yuwei Ren, Eric Mahurin, Chirag Patel, Sundar Subramanian, Sanghyuk Lee, Markus Nagel, Joseph Soriaga, et al. Fp8 versus int8 for efficient deep learning inference. arXiv preprint arXiv:2303.17951, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[37] Binxu Wang and John J Vastola. Diffusion models generate images like painters: an analytical theory of outline first, details later. arXiv preprint arXiv:2303.02490, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Changyuan Wang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, and Jiwen Lu. Towards accurate data-free quantization for diffusion models. arXiv preprint arXiv:2305.18723, 2(5), 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[39] Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, Junchi Yan, and Yan Yan. Quest: Low-bit diffusion model quantization via efficient selective finetuning. arXiv preprint arXiv:2402.03666, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[40] Jie Wang, Huanxi Liu, Dawei Feng, Jie Ding, and Bo Ding. Fp4-quantization: Lossless 4bit quantization for large language models. In 2024 IEEE International Conference on Joint Cloud Computing (JCC), pages 61–67. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[41] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[42] Yijia Zhang, Lingran Zhao, Shijie Cao, Sicheng Zhang, Wenqiang Wang, Ting Cao, Fan Yang, Mao Yang, Shang-hang Zhang, and Ningyi Xu. Integer or floating point? new outlooks for low-bit quantization on large language models. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18143" + } + ] +] \ No newline at end of file diff --git a/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/6042cf5b-516c-4fe8-bfad-ecfc854fc855_origin.pdf b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/6042cf5b-516c-4fe8-bfad-ecfc854fc855_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fe3d247247707e73150f07d48c8f3cac046d7331 --- /dev/null +++ b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/6042cf5b-516c-4fe8-bfad-ecfc854fc855_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbe47b1ac81180fc4768211671baa8d18d94c90e0a7e46b5827e89450df60520 +size 6513195 diff --git a/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/full.md b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..efcca5e3fe7e204a2cb10febaf0252c90dbcd435 --- /dev/null +++ b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/full.md @@ -0,0 +1,294 @@ +# Pioneering 4-Bit FP Quantization for Diffusion Models: Mixup-Sign Quantization and Timestep-Aware Fine-Tuning + +Maosen Zhao $^{1*}$ , Pengtao Chen $^{1*}$ , Chong Yu $^{2}$ , Yan Wen $^{1}$ , Xudong Tan $^{1}$ , Tao Chen $^{1\dagger}$ + +1 School of Information Science and Technology, Fudan University + +$^{2}$ Academy for Engineering and Technology, Fudan University + +20307130202@fudan.edu.cn, eetchen@fudan.edu.cn + +# Abstract + +Model quantization reduces the bit-width of weights and activations, improving memory efficiency and inference speed in diffusion models. However, achieving 4-bit quantization remains challenging. Existing methods, primarily based on integer quantization and post-training quantization fine-tuning, struggle with inconsistent performance. Inspired by the success of floating-point (FP) quantization in large language models, we explore low-bit FP quantization for diffusion models and identify key challenges: the failure of signed FP quantization to handle asymmetric activation distributions, the insufficient consideration of temporal complexity in the denoising process during fine-tuning, and the misalignment between fine-tuning loss and quantization error. To address these challenges, we propose the mixup-sign floating-point quantization (MSFP) framework, first introducing unsigned FP quantization in model quantization, along with timestep-aware LoRA (TALoRA) and denoising-factor loss alignment (DFA), which ensure precise and stable fine-tuning. Extensive experiments show that we are the first to achieve superior performance in 4-bit FP quantization for diffusion models, outperforming existing PTQ fine-tuning methods in 4-bit INT quantization. + +# 1. Introduction + +Despite the impressive performance of diffusion models (DMs) in image generation [30, 41], their computational and memory demands, particularly for high-resolution outputs, pose significant challenges for deployment on resource-constrained edge devices, highlighting the need for model compression to address these limitations. Model quantization, a key technique within model compression, reduces the bit-width of model weights and activations, typically stored in 32-bit format, to lower precision. This reduction lowers memory usage and accelerates inference + +speed. By decreasing the bit-width, quantization improves both temporal and memory efficiency in mainstream models, maintaining robust performance, particularly in resource-constrained environments [13, 18, 21]. + +The current quantization methods for diffusion models can be broadly categorized into two primary approaches. The first, post-training quantization (PTQ), optimizes the quantization parameters after the model has been trained, typically by minimizing the quantization error [19, 32]. While PTQ is effective for 4-bit quantization of weights, it is limited by its reliance on 8-bit quantization for activations, as further reduction in activation bit-width leads to significant performance degradation. In contrast, quantization-aware training (QAT) integrates quantization into the training process, enabling the model to learn with 4-bit precision from scratch [20]. Although QAT can achieve high-performance 4-bit models, it incurs substantial computational overhead, rendering it less practical for many real-world applications [6, 15]. + +To achieve fully quantized 4-bit diffusion models with minimal overhead, fine-tuning has emerged as a promising solution. This approach leverages pre-trained models and adjusts a small subset of parameters to narrow the performance gap between the quantized model and its full-precision counterpart. While some studies have explored fine-tuning for 4-bit quantization in diffusion models [8, 39], these methods often fail to achieve consistent performance under standard configurations (e.g., quantizing all layers but not some). Consequently, developing a universally effective and scalable fine-tuning method for 4-bit quantization in diffusion models remains an open challenge. + +To be noticed, existing quantization methods for diffusion models primarily rely on integer (INT) quantization, which has long been the dominant approach. However, recent developments have demonstrated the considerable potential of floating-point (FP) quantization. Compared to INT quantization, FP quantization offers greater flexibility in modeling complex weight and activation distributions [27], leading to improved performance in visual + +tasks at 8-bit precision [17, 36], and remarkable results in large language models under 4-bit precision [23, 40]. Moreover, FP quantization provides significant advantages in inference acceleration, with NVIDIA H100 achieving a $1.45 \times$ speedup with FP8 quantization, outperforming INT8 [29, 42]. Despite these advantages, the application of low-bit FP quantization in diffusion models remains largely unexplored, presenting a promising avenue for future research. + +In summary, to address the challenges in achieving 4-bit diffusion models, we propose a baseline method using a search-based signed FP quantization framework [2, 23] combined with single-LoRA fine-tuning, aiming to realize 4-bit FP quantized diffusion models. Through this exploration, we uncover several findings that are significant for both diffusion model quantization and FP quantization: (1) There exists many layers which exhibit asymmetric distributions, due to the nonlinear activation function $SiLU$ . The application of traditional signed FP quantization with a symmetric distribution leads to significant precision loss in the sub-zero area, causing substantial performance degradation after quantization. (2) Fine-tuning is conducted based on the denoising process, which is regarded as a complex task involving the restoration from outlines to details [37]. However, current methods typically apply a single LoRA to fine-tune severely degraded models across all timesteps, which leads to suboptimal learning at certain timesteps. (3) Predicted noise plays a varying role at different timesteps during denoising, which is the key to diffusion models. Arising from the neglect of this variation across timesteps, we identify a mismatch between the impact of quantization and the loss function in current methods, which undermines the effectiveness of fine-tuning. + +Facing the challenges in achieving 4-bit FP quantization, we propose constructive strategies: (1) To handle different distribution effectively, we propose a mixup-sign floating-point quantization (MSFP) framework, where unsigned FP quantization with an added zero point leads to a more compatible distribution of discrete points with the anomalous distributions in activations and signed FP quantization continues to exhibit strong representation capacity on other distributions. (2) Realizing that the current single-LoRA fine-tuning approach is in lack of flexibility across timesteps, we introduce timestep-aware LoRA (TALoRA), incorporating multiple LoRAs and a timestep-aware router to dynamically select the appropriate LoRA for each timestep in the denoising process. (3) To further improve the effectiveness of fine-tuning, we introduce a denoising-factor loss alignment (DFA), ensuring the loss function, and the guidance of the fine-tuning, are consistent with the actual quantization deterioration across timesteps. + +In summary, our contributions are as follows: + +(i) We are the first to identify that signed FP quantization struggles with asymmetric activations, which + +arise from the nonlinear behavior of activation functions. To address this, we introduce the MSFP framework, which is also the first effective application of unsigned FP quantization in quantization, offering a novel approach for achieving low-bit quantization. + +(ii) For the fine-tuning of low-bit diffusion models, which is based on the denoising process, we precisely define it as a multi-task process across timesteps and introduce an efficient TALoRA module. Furthermore, we improve the alignment of loss function with quantization error via DFA strategy, enabling stable and reliable fine-tuning to achieve low-bit diffusion models. + +(iii) We focus on identifying and eliminating three major barriers to effective low-bit FP quantization in diffusion models. Extensive experiments on DDIM and LDM demonstrate that our work achieves SOTA results for 4-bit quantization. The proposed MSFP, TALoRA and DFA have greatly advanced the progress of low-bit quantization in diffusion models. + +# 2. Related Work + +# 2.1. Diffusion Model Quantization + +There are two main approaches in the quantization of diffusion models: QAT [6, 14] and PTQ [28]. QAT is particularly effective for low-bit quantization while it retransmits the model from scratch, consuming extensive computational resources and time [20]. In contrast, PTQ offers greater time efficiency, making it more practical for large models in real-world applications. Recent advancements in PTQ have concentrated on optimizing calibration datasets to enhance reconstruction accuracy [9, 19, 25, 32] and mitigating quantization errors stemming from the temporal and structural properties of diffusion models [3, 19, 33, 35]. To further achieve fully 4-bit quantization, finetuning techniques has been introduced in PTQ-based quantization. EfficientDM [8] develops a LoRA-based finetuning framework, while QuEST [39] focuses on optimizing quantization-unfriendly activations. Both works stagnate in addressing data distribution during fine-tuning, failing to account for the specific challenges involved in finetuning within the context of diffusion model quantization. + +Meanwhile, recent advancements in FP quantization have made significant strides in model quantization [17, 23], highlighting its potential for diffusion models. While existing low-bit quantization methods perform well for linear models, applying FP quantization to convolutional diffusion models remains more challenging. To date, only one study has explored 8-bit activation quantization in diffusion models using a basic search-based approach [2], with no research addressing lower-bit quantization. This underscores both the challenges and the untapped potential of achieving 4-bit FP quantization for diffusion models. + +![](images/4e3effa50b5d8326230a2591b257ff69ba8f6f732d73b7721d59900932d8a318.jpg) +Figure 1. The activation distributions in NALs and AALs, results on the CelebA dataset. (a) The paradigm of NALs with symmetric activations. (b) The typical paradigm of AALs with asymmetric activations, where unsigned FP quantization is more suitable. (c) The infrequent paradigm of AALs with relatively symmetric activations, where either signed or unsigned FP quantization could be applicable. + +![](images/c83abc04fa0f7347d12b7a595d24dbb92053e026580e0418646e8eedbc838016.jpg) + +![](images/87fcb6d0f687d38b802457a5ff7c2538c91eaaa900b6a761ca3f35cb4afa78e5.jpg) + +# 2.2. Parameter-Efficient Fine-Tuning + +Parameter-efficient fine-tuning (PEFT) has emerged as an effective alternative to full model fine-tuning, focusing on adjusting only a subset of parameters while keeping the majority frozen, thereby reducing storage overhead. Low-rank adapters (LoRA) [12], originally developed for large language models, have become one of the most widely used PEFT methods. Leveraging LoRA's strong transferability, QLoRA [5] can be effectively applied to fine-tune low-bit diffusion models. However, previous LoRA-based fine-tuning is suboptimal in practice. In this paper, we make a further exploration on this and incorporate timestep-level adaptation inspired by MoELoRA [7], enhancing performance of low-bit quantized DMs. + +# 3. Challenges & Exploration + +# 3.1. Preliminary + +Diffusion Models. The diffusion model is a new generation framework that completes learning by adding noise and completes generation in a denoising manner. As for the forward process, the noise is injected into ground-truth images $\boldsymbol{x}_0$ at random timesteps. This enables the DMs to learn the distributions of noise through noisy images $\boldsymbol{x}_t$ , which can be obtained as follows [11]: + +$$ +\boldsymbol {x} _ {t} = \sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \varepsilon . \tag {1} +$$ + +Here $\varepsilon$ represents a standard Gaussian noise and $\overline{\alpha}_t$ is the accumulated noise intensity, calculated by: + +$$ +\bar {\alpha} _ {t} = \prod_ {i = 1} ^ {t} \alpha_ {i}, \tag {2} +$$ + +where $\alpha_{s}$ governs the noise intensity under each timestep. Once the DMs are well-trained, the Gaussian noise image will be inputted to the DMs, and undergo the iterative denoising process. Specifically, the noise can be predicted + +by DMs and used to obtain the image $\boldsymbol{x}_t$ under timestep $t$ , which could range from $T$ to 0, with the objective of obtaining the image $\boldsymbol{x}_{t-1}$ in a superior quality: + +$$ +\boldsymbol {x} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\boldsymbol {x} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \overline {{\alpha}} _ {t}}} \cdot \boldsymbol {\varepsilon} _ {\theta} \left(\boldsymbol {x} _ {t}, t\right)\right) + \sigma_ {t} \boldsymbol {\delta}, (3) +$$ + +where $\varepsilon_{\theta}(\boldsymbol{x}_t,t)$ is the predicted noise at timestep $t$ and $\delta$ is a newly added noise with the factor $\sigma_t$ to ensure diverse results. Here we define a denoising factor $\gamma_t$ , which is formulated as: + +$$ +\gamma_ {t} = \frac {1}{\sqrt {\alpha_ {t}}} \cdot \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}}. \tag {4} +$$ + +According to Equation 3, $\gamma_{t}$ indicates the impact of the prediction noise under timestep $t$ . The greater the factor $\gamma_{t}$ , the stronger the predicted noise effect in the denoising. + +Model Quantization. The fundamental paradigm of model quantization is the process of transforming a continuous data distribution into a finite set of discrete points. Consequently, the quality of the quantized model is inextricably related to the distribution of discrete points. According to the discrete point type, quantization is defined as two categories, widely-used INT quantization and emerging FP quantization. Analogous to the process of INT quantization, a floating-point vector $x$ can be quantized as follows: + +$$ +\widehat {\boldsymbol {x}} = \operatorname {C l i p} \left(\left\lfloor \frac {\boldsymbol {x}}{s} \right\rceil + z, l, u\right) \cdot s. \tag {5} +$$ + +Here $\lfloor\rfloor$ is the rounding operation. $l$ and $u$ are the minimum and maximum quantization thresholds while scaling factor $s$ and zero-point $z$ together constitute the quantization parameters. As shown in Equation 5, INT quantization results in an evenly spaced distribution of discrete points. While INT quantization is straightforward to implement, it may be too simplistic for continuous distributions that have significant variations in density, where evenly spaced intervals in INT quantization are not effective. On the contrary, + +![](images/6a26f2dab317b9b9951cc7615de13e4369501d93b6d456f95088259c55a95c58.jpg) +Figure 2. Effect of bit-width reduction on Figure 3. Two loss, and performance degradation representation capacity in AALs (blue) and NALs (orange) under signed FP models across steps. Compared with metric, quantization, evaluated on CelebA dataset. the original loss shows an inverse trend, while the aligned loss remains consistent. + +![](images/53d938ec9cd6120a695018d1bc6c77e257b06fb20289876e696a0b67c248c4a1.jpg) +Figure 4. The MSE of activations before and after quantization across all AALs under four different strategies, normalized against the baseline of signed FP quantization without zero point (purple). + +![](images/6c53f0d4bffe1523270302989a390884558814a6d1ecb309c6a260e6e9bed735.jpg) + +the discrete points under FP quantization are not uniformly spaced in distribution, as shown below [22]: + +$$ +f = (- 1) ^ {s} 2 ^ {p - b} \left(1 + \frac {d _ {1}}{2} + \frac {d _ {2}}{2 ^ {2}} + \dots \frac {d _ {m}}{2 ^ {m}}\right), \tag {6} +$$ + +where $s$ is the sign bit, $d_{i}$ is the $m$ -bit mantissa, $p$ is the $e$ -bit exponent and $b$ is the bias that serves as both scaling factor and threshold in INT quantization. Previous studies in FP quantization rely on signed FP quantization, where $s$ is set to 1. For an $n$ -bit FP quantization, the bit-width is distributed across the mantissa, exponent, and sign bit, with the condition $m + e + s = n$ . The different combinations of $m$ and $e$ allow FP quantization to represent data in multiple formats with a fixed number of bits, denoted as $E_{i}M_{j}$ , where $i$ -bit exponent and $j$ -bit mantissa are specified. A larger $j$ results in higher precision within each interval, while a larger $i$ expands the range of covered intervals. + +There appears to be an inherent suitability of FP quantization for DMs, as the majority of weights and activations follow a normal distribution symmetric around value $= 0$ . This aligns well with the unevenly spaced distribution, where discrete points are dense in the small-value region and sparse in the large-value region, in FP quantization [42]. Additionally, the flexible formatting mechanism allows FP quantization to better accommodate complex distribution scenarios. + +# 3.2. Barriers to Effective Low-Bit FP Quantization in Diffusion Models + +In order to implement high-performing 4-bit diffusion models, we implemented FP quantization that is more compatible with the data distribution of diffusion models, with the search-based strategy in previous work [23]. Building on this, we deployed a LoRA-based fine-tuning strategy to address the performance degradation caused by 4-bit quantization. Despite this, we find that the performance of the 4-bit FP diffusion model remains satisfactory and + +we identify two main issues that we are facing: (1) For FP quantization, although FP quantization works well at 8 bits, it experiences a sharp performance degradation at 4 bits. How to improve the ability of FP quantization to represent data under low-bit quantization? (2) For standard post-training LoRA-based fine-tuning, exhibits instability and suboptimal results when applied to low-bit quantized diffusion models. How can we make LoRA more efficient and accurate in learning the loss information at different denoising timesteps? In the following parts, we will explore the underlying causes of these two issues and the feasible strategies to address them. + +# Observation 1: Previous signed FP quantization fails to achieve effective low-bit quantization in Activation-Anomalous Layers. + +In diffusion models, we observed that the nonlinear activation layer $SiLU$ , defined as $SiLU(\boldsymbol{x}) = \frac{\boldsymbol{x}}{1 + e^{-\boldsymbol{x}}}$ , is commonly situated between layers. + +$SiLU$ causes the abnormal activations for the subsequent layer. As depicted in Panel (b) of Figure 1, all values below 0 are compressed into the range of $[-0.278, 0)$ . In this paper, we refer to layers with such asymmetric activations as Anomalous-Activation-Distribution Layers (AALs) and the other layers as Normal-Activation-Distribution Layers (NALs). Mainstream signed FP quantization typically sets the maximum threshold of the quantizer high to accurately represent normal positive activations. However, this approach results in a significant precision loss when dealing with values below 0. Figure 2 illustrates the representation capacity of FP signed quantization in both NALs and AALs under different bit widths. When the bit width drops below 6 bits, AALs suffer more severe performance degradation compared to NALs, which ultimately leads to the failure in low-bit FP quantization. This phenomenon suggests that mitigating the performance decline in AALs is a critical step towards improving low-bit FP quantization in diffusion models. + +Observation 2: The single-LoRA-based strategy is overly simplistic for fine-tuning quantized diffusion models across different timesteps. + +Previous work has focused on adapting LoRA for the quantization of diffusion models but has not fully explored LoRA's performance in the context of denoising. Considering that the denoising process of diffusion models starts with recovering outlines and progresses to restoring details, we question whether the single-LoRA fine-tuning strategy can handle this complexity. In Table 1, we compare the baseline model, which uses a single-LoRA strategy for fine-tuning, with two alternative strategies. The second strategy assigns a separate LoRA for the first and last 50 timesteps, resulting in a significant improvement over the baseline. In contrast, the third strategy also introduces a dual-LoRA strategy, it randomly selects one for each timestep, resulting in much worse performance. These results suggest that applying multiple LoRAs, allocated in a structured manner across timesteps, enhances model performance, while disordered selection of LoRAs could lead to suboptimal results. This motivates us to approach the fine-tuning of quantized diffusion models as a multi-task process and assign multiple LoRAs across different timesteps with a rational approach in allocation. + +Observation 3: The MSE of the predicted noise of full-precision and quantized models does not reflect the actual impact of quantization at different timesteps. + +In fine-tuning, a commonly used loss function calculates the MSE between the noise predictions of the full-precision and quantized diffusion models, using denoised images from the full-precision model at the previous timestep as inputs: + +$$ +L _ {\boldsymbol {\varepsilon} _ {\theta}} ^ {t} = \left\| \boldsymbol {\varepsilon} _ {\theta} (\boldsymbol {x} _ {t}, t) - \widehat {\boldsymbol {\varepsilon}} _ {\theta} (\boldsymbol {x} _ {t}, t) \right\| ^ {2}. \tag {7} +$$ + +By observing the variation in this loss during single-LoRA fine-tuning (see Figure 3), we identify an unexpected trend: the loss increases progressively faster as denoising advances. This contradicts the principle of denoising, where image quality should improve with each step, and the impact of predicted noise should diminish over time. By the final step, the impact of quantization on model performance should be negligible, as the predicted noise no longer affects the input. However, the loss is at its maximum, indicating that quantization error is most significant at this stage, contradicting the expectation that its influence should diminish over time. + +To highlight the discrepancy between the loss and actual quantization errors, we define the performance gap at each step as the difference in denoising quality between the quantized and full-precision models, as the ultimate goal of denoising is to yield high-quality images. We input the previous output image $\boldsymbol{x}_t$ from the full-precision model and calculate the performance gap between the de + +
MethodBits (W/A)FID ↓
FP32/326.49
Single-LoRA4/419.41
Dual-LoRA (Split Steps in Half)4/417.07
Dual-LoRA (Random Allocation)4/441.96
+ +Table 1. The impact of the number of LoRAs and their allocation across timesteps on the performance of the fine-tuning. The results is evaluated by 4-bit quantization on CelebA dataset. + +noised image $\boldsymbol{x}_{t-1}$ from the full-precision model and the denoised image $\widehat{\boldsymbol{x}}_{t-1}$ from the quantized model, measured by $MSE(\boldsymbol{x}_{t-1}, \widehat{\boldsymbol{x}}_{t-1})$ . As shown in Figure 3, this misalignment between the loss and the actual performance gap leads to deviations in LoRA's learning. This necessitates aligning our loss function with the denoising process during fine-tuning. + +# 4. Methodology + +In this section, we explore the issues identified in Section 3.2 and propose corresponding solutions. As illustrated in Figure 5, we introduce a mixup-sign FP quantization framework to address the diverse activation distributions in the first stage. During fine-tuning, the timestep-aware routing mechanism and denoising-factor loss alignment work in tandem to enable high-quality learning, ultimately enabling the realization of optimized 4-bit FP diffusion models. + +# 4.1. Mixup-Sign Floating Point Quantization + +To address the challenges of low-bit FP quantization failure in AALs, we leverage FP quantization by allocating more discrete points to areas with high data concentration. Motivated by the half-normal distribution of activations in AALs, we introduce unsigned floating-point quantization. However, as shown in Equation 6, when using unsigned FP quantization with $s$ set to 0, we round all data in the sub-zero range to zero, losing important negative information. To address this, we introduce a zero point in the range of [-0.278, 0) to recover most sub-zero activations. The updated quantization formula becomes: + +$$ +f _ {u n s i g n} = (- 1) ^ {s} 2 ^ {p - b} \left(1 + \frac {d _ {1}}{2} + \frac {d _ {2}}{2 ^ {2}} + \dots \frac {d _ {m}}{2 ^ {m}}\right) + z, \tag {8} +$$ + +where $s$ is set to 0 and $z$ is the newly added zero point. By freeing the 1-bit sign bit, which is ineffective in signed FP quantization, and using it as additional exponent / mantissa bit width, we fully utilize the representation capacity. As illustrated in Figure 4, unsigned FP quantization with a zero point significantly improves representation in over $95\%$ of AALs, compared to traditional signed FP quantization. + +![](images/42465d15d6c8f7fb8080a1fedefa5e8e6339dcd958c96419d84c15eba4ea4e57.jpg) + +![](images/732cdcf2680b76c4b427472166d45ba4ed70817c700a3c08e2fe4f54cfeffcf4.jpg) +Figure 5. The pipeline of our proposed method. UNets are applied to the Mixup-Sign Floating-Point Quantization (MSFP), where distinct floating-point quantization schemes are employed for Anomalous-Activation-Distribution Layers (AALs) and Normal-Activation-Distribution Layers (NALs). During the fine-tuning stage, multiple LoRA modules are introduced, and a timestep-aware routing mechanism is used for dynamic LoRA allocation across different timesteps. Additionally, a denoising-factor alignment technique is employed to align the loss function with quantization-induced performance degradation. + +![](images/5a2e2cf7780dc2a47983d69b83aaf68c13032b4fb54287b32209c352dacbaea0.jpg) + +![](images/94473f604c4f4dd4b84b70a27c3788b0bbcbfba33e6c6a81635e404c58974655.jpg) + +However, there are rare cases where performance slightly goes worse due to the diversity of anomalous distributions. Panel (c) of Figure 1 shows that in such cases, the activation distribution may resemble a normal distribution, where signed FP quantization might perform better. Figure 4 further indicates that introducing a zero point into signed FP quantization is unnecessary, offering minimal improvement in a few cases. + +Given the strong performance of unsigned FP quantization with a zero point and the diversity of AALs, we propose a mixup-sign FP quantization framework. During the search-based initialization phase, we use signed FP quantization for NALs and introduce both unsigned FP quantization with a zero point and signed FP quantization for AALs. This approach addresses AAL challenges in low-bit quantization while minimizing computational overhead by adding the zero point only to unsigned FP quantization. + +# 4.2. Timestep-Aware Router for LoRA Allocation + +In Section 3.2, we observe that a single LoRA cannot fully capture all the information across timesteps due to the diverse generative characteristics at different timesteps. We also find that a reasonable allocation of different LoRAs for timesteps will be beneficial to fine-tuning. In this section, we introduce a timestep-aware LoRA allocation method that dynamically assigns optimal LoRA to each timestep, maximizing fine-tuning effectiveness. + +Our method relies on a learnable router (illustrated in Figure 5), a module shared across all timesteps. It takes the timestep as input and outputs selection probabilities for each LoRA across UNet layers. For each timestep, the LoRA with the highest probability corresponding to the router's output is inserted into the quantized model for fin-tuning or inference. In a router network, the main components are a time embedding layer and an MLP layer. The time embedding layer, derived from a pre-trained diffusion model, converts the scalar into a $d$ -dimensional embedding. The MLP layer then maps the embedding to a LoRA allocation distribution, where $y$ is the number of quantized layers and $h$ is the size of the LoRA Hub. Using an STE method [1], this distribution is converted into 0/1 probabilities to allocate suitable LoRAs to the diffusion model's quantized layers. + +# 4.3. Denoising Factor Aligned Loss + +To address the mismatch between the actual performance gap and the loss used during quantization, we review the denoising principle depicted in Equation 3 and pinpoint the cause of the mismatch: the time-step-dependent constraint on the predicted noise is not sufficiently accounted for during the denoising process. Therefore, we implement a modification to the loss function based on predicted noise: + +$$ +L ^ {t} = \gamma_ {t} \cdot L _ {\varepsilon_ {\theta}} ^ {t}. \tag {9} +$$ + +By introducing $\gamma_{t}$ , which accurately reflects the utilization of the predicted noise at each time step, we achieve a preliminary alignment between the loss and the actual quantization error, as shown in Figure 3. This facilitates more accurate fine-tuning, leading to better performance recovery in the low-bit diffusion model. + +# 5. Experiment + +# 5.1. Experimental Setup + +Models and Metrics. To verify the effectiveness of the proposed method, we evaluate it with two widely adopted diffusion paradigms: DDIM [34] and LDM [30]. For DDIM experiments, we evaluate on CIFAR-10 [16] and CelebA [26]. For LDM, we test unconditional generation on LSUN-Bedroom [41] and LSUN-Church [41] and conditional generation on ImageNet [4]. The performance of the diffusion models is evaluated with Inception Score (IS) [31], Fréchet Inception Distance (FID) [10] and Sliding FID (sFID) [31]. All metrics are evaluated based on 50k samples generated by the DDIM solver [34]. + +Quantization Detail. We employ standard layer-wise quantization for both weights and activations. Except for the input and output layers, which are typically set to 8-bit, all other convolution and linear layers are quantized to the target bit-width. Furthermore, we generate 256 samples for the calibration set based on Q-Diffusion [19] for bias initialization and use the method from [2] to obtain the optimal quantization parameters. + +Baseline. We compare two main types of quantization methods: PTQ methods (Q-Diffusion [19] and EDADM [38]) and fine-tuning methods (EfficientDM [8] and QuEST [39]). Since these fine-tuning methods involve special settings like non-full-layer quantization, we standardize the settings of EfficientDM in a consistent manner and procure other standardized results from DilateQuant [24]. Comparison with special settings is provided in Appendix. + +# 5.2. Quantization Performance + +Unconditional Generation. Table 2 presents the results of unconditional image generation across multiple datasets. With 6-bit quantization, our method achieves nearly identical performance to full precision. Our 4-bit quantized models achieve SOTA results in all tasks, significantly outperforming previous baseline methods. Notably, on CIFAR-10, our 4-bit quantized model results in an FID score that is only 1.84 worse than full precision, an almost negligible degradation, while previous methods struggle with 4-bit quantization. Compared to the fine-tuning method [8], our 4-bit quantized models improve FID by 32.38, 24.15, and 9.63 on three datasets, respectively, and still maintain remarkable IS. Additionally, we provide the performance of 4-bit and 6-bit quantized models on CelebA in Appendix. + +
TaskMethodPrec. (W/A)FID ↓IS ↑
CIFAR-10 32x32FP32/324.269.03
Q-Diffusion6/69.198.76
EDA-DM6/626.689.35
EfficientDM6/625.038.08
Ours (h=2)6/64.269.04
DDIM steps = 100Ours (h=4)6/64.239.06
Q-Diffusion4/4N/AN/A
EDA-DM4/4120.244.42
EfficientDM4/438.407.32
Ours (h=2)4/46.028.79
Ours (h=4)4/46.108.90
LSUN (Bedroom) 256x256FP32/323.022.29
Q-Diffusion6/610.102.11
EDA-DM6/610.562.12
QuEST6/610.102.20
EfficientDM6/612.952.57
Ours (h=2)6/68.422.49
Ours (h=4)6/68.402.49
LDM-4 steps = 100 eta = 1.0Q-Diffusion4/4N/AN/A
EDA-DM4/4N/AN/A
QuEST4/4N/AN/A
EfficientDM4/436.362.69
Ours (h=2)4/412.212.47
Ours (h=4)4/412.342.48
LSUN (Church) 256x256FP32/324.062.70
Q-Diffusion6/610.902.47
EDA-DM6/610.762.43
QuEST6/66.832.65
EfficientDM6/67.452.80
Ours (h=2)6/66.242.73
Ours (h=4)6/66.382.73
LDM-8 steps = 100 eta = 0.0Q-Diffusion4/4N/AN/A
EDA-DM4/4N/AN/A
QuEST4/413.032.63
EfficientDM4/418.402.97
Ours (h=2)4/48.812.70
Ours (h=4)4/48.772.71
+ +Table 2. Quantization performance of unconditional generation. 'Prec. (W/A)' denotes the quantization bit-width. 'N/A' denotes failed image generation. $h$ denotes the size of LoRA Hub. + +Conditional Generation. Table 3 presents the results of our conditional generation experiments on ImageNet. We observe that FID is not always a reliable metric in this context, as an unexpected trend emerges: as the bit-width of the quantized model decreases, the FID score improves, which contradicts the expected trend. Therefore, our discussion of the ImageNet results focuses on other evaluation metrics. As shown by the IS and sFID, our method achieves + +![](images/b44320fb47421170619870e8e825a81b7066ba6a69d1b107979d2956c7c66a28.jpg) +Figure 6. A visual comparison of generation results using our method across different quantization bit-widths, with the LSUN-Church dataset as an example. + +performance comparable to that of the full-precision model with 6-bit quantization. Even with 4-bit quantization, we achieve SOTA results in terms of sFID, improving by 7.00 over the previous best method, EfficientDM [8]. Furthermore, our method significantly outperforms two other fin-tuning methods in the IS metric. Visual evaluations further confirm that the generated images maintain high quality, exhibiting clear and coherent content, as shown in Figure 6. More visualization results will be presented in Appendix. + +
MethodPrec. (W/A)sFID ↓FID ↓IS ↑
FP32/327.6711.69364.72
EDA-DM6/68.0211.52360.77
QuEST6/69.368.45310.12
EfficientDM6/66.889.54351.79
Ours (h=2)6/67.4310.10349.91
Ours (h=4)6/66.6510.10351.79
EDA-DM4/436.6620.02204.93
QuEST4/429.2738.4369.58
EfficientDM4/414.4212.73139.45
Ours (h=2)4/47.426.50190.74
Ours (h=4)4/48.237.43177.40
+ +# 5.3. Ablation Study + +The ablation experiments are conducted on the 4-bit quantization using the CelebA dataset, which is challenging for low-bit quantization, further demonstrating the effectiveness of our approach. The baseline uses signed FP quantization combined with single LoRA fine-tuning. As shown in Table 4, all three proposed modules lead to significant performance improvements, with their combination + +Table 3. Quantization performance of conditional generation for fully-quantized LDM-4 models on ImageNet $256 \times 256$ with 20 steps. 'Prec. (W/A)' denotes the quantization bit-width. $h$ denotes the size of LoRA Hub. + +
MSFPMethod TALoRADFAPrec. (W/A)FID ↓
XXX4/416.02
✓XX4/49.60
X✓X4/410.66
✓X✓4/48.39
✓✓X4/48.79
✓✓✓4/47.69
+ +Table 4. Ablation study on different modules we proposed. Testing on CelebA dataset with $h = 2$ LoRA Hub size. + +yielding a synergistic effect. The baseline FID score is 9.53 higher than that of the full-precision model (6.49). By applying our technique, we reduce the FID by 8.18 compared to the baseline. More interesting is that we visualize the LoRA allocation distribution learned by the router as shown in Figure 7. We find that the distribution of the router-learned allocation over timesteps is consistent with the finding that the diffusion model focuses on contour generation early and on detail generation later [37]. + +![](images/80925daab5f8ace3bc40f76b239165ee63f0e19ef2fd0dbcf7d5fd7f43f265d2.jpg) +Figure 7. Distribution of LoRA allocations over timesteps obtained after router training on different datasets, when $h = 2$ . + +# 6. Conclusion + +In this paper, we focus on exploring low-bit FP quantization for diffusion models. For model initialization, we innovatively introduce unsigned FP quantization with zero point to address AALs. For the fine-tuning based on the denoising process, we formulate it as a multi-task procedure. We introduce multiple LoRAs along with a router for their allocation at different timesteps, and further align the loss function, originally based on estimated noise, with the actual quantization error. We introduce unsigned FP quantization and achieve 4-bit FP quantized diffusion models. Our FP PTQ-based fine-tuning method sets a new precedent for 4-bit diffusion models, offering insight into the deployment of low-bit diffusion models in the future. + +# 7. Acknowledgments + +This work is supported by Shanghai Science and Technology Commission Explorer Program Project (24TS1401300), National Key Research and Development Program of China (No.2022ZD0160101). The computations in this research were performed using the CFFF platform of Fudan University. + +# References + +[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 6 +[2] Cheng Chen, Christina Giannoula, and Andreas Moshovos. Low-bitwidth floating point quantization for efficient high-quality diffusion models. arXiv preprint arXiv:2408.06995, 2024. 2, 7 +[3] Huanpeng Chu, Wei Wu, Chengjie Zang, and Kun Yuan. Qncd: Quantization noise correction for diffusion models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10995-11003, 2024. 2 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7 +[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: efficient finetuning of quantized llms (2023). arXiv preprint arXiv:2305.14314, 52:3982-3992, 2023. 3 +[6] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. In International Conference on Learning Representations, 2019. 1, 2 +[7] Wenfeng Feng, Chuzhan Hao, Yuewei Zhang, Yu Han, and Hao Wang. Mixture-of-loras: An efficient multitask tuning method for large language models. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 11371-11380, 2024. 3 +[8] Yefei He, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Efficientdm: Efficient quantization-aware fine-tuning of low-bit diffusion models. arXiv preprint arXiv:2310.03270, 2023. 1, 2, 7, 8 +[9] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 + +[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[13] Wei Huang, Xingyu Zheng, Xudong Ma, Haotong Qin, Chengtao Lv, Hong Chen, Jie Luo, Xiaojuan Qi, Xianglong Liu, and Michele Magno. An empirical study of llama3 quantization: From llms to mllms. Visual Intelligence, 2(1): 36, 2024. 1 +[14] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2704-2713, 2018. 2 +[15] Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342, 2018. 1 +[16] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7 +[17] Andrey Kuzmin, Mart Van Baalen, Yuwei Ren, Markus Nagel, Jorn Peters, and Tijmen Blankevoort. Fp8 quantization: The power of the exponent. Advances in Neural Information Processing Systems, 35:14651-14662, 2022. 2 +[18] Min Li, Zihao Huang, Lin Chen, Junxing Ren, Miao Jiang, Fengfa Li, Jitao Fu, and Chenghua Gao. Contemporary advances in neural network quantization: A survey. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-10. IEEE, 2024. 1 +[19] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1, 2, 7 +[20] Yanjing Li, Sheng Xu, Xianbin Cao, Xiao Sun, and Baochang Zhang. Q-dm: An efficient low-bit quantized diffusion model. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 +[21] Tailin Liang, John Glossner, Lei Wang, Shaobo Shi, and Xiaotong Zhang. Pruning and quantization for deep neural network acceleration: A survey. Neurocomputing, 461:370-403, 2021. 1 +[22] Fangxin Liu, Wenbo Zhao, Zhezhi He, Yanzhi Wang, Zongwu Wang, Changzhi Dai, Xiaoyao Liang, and Li Jiang. Improving neural network efficiency via post-training quantization with adaptive floating-point. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5281-5290, 2021. 4 +[23] Shih-yang Liu, Zechun Liu, Xijie Huang, Pingcheng Dong, and Kwang-Ting Cheng. LIm-fp4: 4-bit floating-point quantized transformers. arXiv preprint arXiv:2310.16836, 2023. 2, 4 +[24] Xuewen Liu, Zhikai Li, and Qingyi Gu. Dilatequant: Accurate and efficient diffusion quantization via weight dilation. arXiv preprint arXiv:2409.14307, 2024. 7 +[25] Xuewen Liu, Zhikai Li, Junrui Xiao, and Qingyi Gu. Enhanced distribution alignment for post-training quantization + +of diffusion models. arXiv preprint arXiv:2401.04585, 2024. 2 +[26] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 7 +[27] Paulius Micikevicius, Dusan Stosic, Neil Burgess, Marius Cornea, Pradeep Dubey, Richard Grisenthwaite, Sangwon Ha, Alexander Heinecke, Patrick Judd, John Kamalu, et al. Fp8 formats for deep learning. arXiv preprint arXiv:2209.05433, 2022. 1 +[28] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 2 +[29] NVIDIA. Blackwell platform sets new llm inference records in mlperf inference v4.1, 2024. Available at: https://developer.nvidia.com/blog/nvidia-blackwell-platform-sets-new-llm-inference-records-in-mlperf-inference-v4-1, Accessed: 2024-11-14. 2 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 7 +[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 7 +[32] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 1, 2 +[33] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 7 +[35] Haojun Sun, Chen Tang, Zhi Wang, Yuan Meng, Xinzhu Ma, Wenwu Zhu, et al. Tmpq-dm: Joint timestep reduction and quantization precision selection for efficient diffusion models. arXiv preprint arXiv:2404.09532, 2024. 2 +[36] Mart van Baalen, Andrey Kuzmin, Suparna S Nair, Yuwei Ren, Eric Mahurin, Chirag Patel, Sundar Subramanian, Sanghyuk Lee, Markus Nagel, Joseph Soriaga, et al. Fp8 versus int8 for efficient deep learning inference. arXiv preprint arXiv:2303.17951, 2023. 2 +[37] Binxu Wang and John J Vastola. Diffusion models generate images like painters: an analytical theory of outline first, details later. arXiv preprint arXiv:2303.02490, 2023. 2, 8 +[38] Changyuan Wang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, and Jiwen Lu. Towards accurate data-free quantization for diffusion models. arXiv preprint arXiv:2305.18723, 2(5), 2023. 7 + +[39] Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, Junchi Yan, and Yan Yan. Quest: Low-bit diffusion model quantization via efficient selective finetuning. arXiv preprint arXiv:2402.03666, 2024. 1, 2, 7 +[40] Jie Wang, Huanxi Liu, Dawei Feng, Jie Ding, and Bo Ding. Fp4-quantization: Lossless 4bit quantization for large language models. In 2024 IEEE International Conference on Joint Cloud Computing (JCC), pages 61–67. IEEE, 2024. 2 +[41] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 1, 7 +[42] Yijia Zhang, Lingran Zhao, Shijie Cao, Sicheng Zhang, Wenqiang Wang, Ting Cao, Fan Yang, Mao Yang, Shang-hang Zhang, and Ningyi Xu. Integer or floating point? new outlooks for low-bit quantization on large language models. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2, 4 \ No newline at end of file diff --git a/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/images.zip b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0462c2ea83b4a52fff3e817ae5e84c75498847a2 --- /dev/null +++ b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc3b3fa173f04d0552943da5a6d39905180723894af1ad449706d1473b552896 +size 559948 diff --git a/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/layout.json b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8118898842f6356b9ace1257b07dc28c47556f64 --- /dev/null +++ b/2025/Pioneering 4-Bit FP Quantization for Diffusion Models_ Mixup-Sign Quantization and Timestep-Aware Fine-Tuning/layout.json @@ -0,0 +1,7489 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 102, + 516, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 102, + 516, + 140 + ], + "spans": [ + { + "bbox": [ + 96, + 102, + 516, + 140 + ], + "type": "text", + "content": "Pioneering 4-Bit FP Quantization for Diffusion Models: Mixup-Sign Quantization and Timestep-Aware Fine-Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "spans": [ + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "text", + "content": "Maosen Zhao" + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "text", + "content": ", Pengtao Chen" + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "text", + "content": ", Chong Yu" + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "text", + "content": ", Yan Wen" + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "text", + "content": ", Xudong Tan" + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "text", + "content": ", Tao Chen" + }, + { + "bbox": [ + 93, + 160, + 512, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 175, + 466, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 175, + 466, + 189 + ], + "spans": [ + { + "bbox": [ + 143, + 175, + 466, + 189 + ], + "type": "text", + "content": "1 School of Information Science and Technology, Fudan University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 154, + 189, + 456, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 189, + 456, + 203 + ], + "spans": [ + { + "bbox": [ + 154, + 189, + 456, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 189, + 456, + 203 + ], + "type": "text", + "content": " Academy for Engineering and Technology, Fudan University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 178, + 205, + 428, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 205, + 428, + 217 + ], + "spans": [ + { + "bbox": [ + 178, + 205, + 428, + 217 + ], + "type": "text", + "content": "20307130202@fudan.edu.cn, eetchen@fudan.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 270, + 297, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 534 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 534 + ], + "type": "text", + "content": "Model quantization reduces the bit-width of weights and activations, improving memory efficiency and inference speed in diffusion models. However, achieving 4-bit quantization remains challenging. Existing methods, primarily based on integer quantization and post-training quantization fine-tuning, struggle with inconsistent performance. Inspired by the success of floating-point (FP) quantization in large language models, we explore low-bit FP quantization for diffusion models and identify key challenges: the failure of signed FP quantization to handle asymmetric activation distributions, the insufficient consideration of temporal complexity in the denoising process during fine-tuning, and the misalignment between fine-tuning loss and quantization error. To address these challenges, we propose the mixup-sign floating-point quantization (MSFP) framework, first introducing unsigned FP quantization in model quantization, along with timestep-aware LoRA (TALoRA) and denoising-factor loss alignment (DFA), which ensure precise and stable fine-tuning. Extensive experiments show that we are the first to achieve superior performance in 4-bit FP quantization for diffusion models, outperforming existing PTQ fine-tuning methods in 4-bit INT quantization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 575, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 296, + 696 + ], + "type": "text", + "content": "Despite the impressive performance of diffusion models (DMs) in image generation [30, 41], their computational and memory demands, particularly for high-resolution outputs, pose significant challenges for deployment on resource-constrained edge devices, highlighting the need for model compression to address these limitations. Model quantization, a key technique within model compression, reduces the bit-width of model weights and activations, typically stored in 32-bit format, to lower precision. This reduction lowers memory usage and accelerates inference" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 246, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 293 + ], + "type": "text", + "content": "speed. By decreasing the bit-width, quantization improves both temporal and memory efficiency in mainstream models, maintaining robust performance, particularly in resource-constrained environments [13, 18, 21]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 294, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 556, + 474 + ], + "type": "text", + "content": "The current quantization methods for diffusion models can be broadly categorized into two primary approaches. The first, post-training quantization (PTQ), optimizes the quantization parameters after the model has been trained, typically by minimizing the quantization error [19, 32]. While PTQ is effective for 4-bit quantization of weights, it is limited by its reliance on 8-bit quantization for activations, as further reduction in activation bit-width leads to significant performance degradation. In contrast, quantization-aware training (QAT) integrates quantization into the training process, enabling the model to learn with 4-bit precision from scratch [20]. Although QAT can achieve high-performance 4-bit models, it incurs substantial computational overhead, rendering it less practical for many real-world applications [6, 15]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 474, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 617 + ], + "type": "text", + "content": "To achieve fully quantized 4-bit diffusion models with minimal overhead, fine-tuning has emerged as a promising solution. This approach leverages pre-trained models and adjusts a small subset of parameters to narrow the performance gap between the quantized model and its full-precision counterpart. While some studies have explored fine-tuning for 4-bit quantization in diffusion models [8, 39], these methods often fail to achieve consistent performance under standard configurations (e.g., quantizing all layers but not some). Consequently, developing a universally effective and scalable fine-tuning method for 4-bit quantization in diffusion models remains an open challenge." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "To be noticed, existing quantization methods for diffusion models primarily rely on integer (INT) quantization, which has long been the dominant approach. However, recent developments have demonstrated the considerable potential of floating-point (FP) quantization. Compared to INT quantization, FP quantization offers greater flexibility in modeling complex weight and activation distributions [27], leading to improved performance in visual" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "type": "text", + "content": "†Corresponding author. *Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18134" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "tasks at 8-bit precision [17, 36], and remarkable results in large language models under 4-bit precision [23, 40]. Moreover, FP quantization provides significant advantages in inference acceleration, with NVIDIA H100 achieving a " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "1.45 \\times" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " speedup with FP8 quantization, outperforming INT8 [29, 42]. Despite these advantages, the application of low-bit FP quantization in diffusion models remains largely unexplored, presenting a promising avenue for future research." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 170, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 296, + 456 + ], + "type": "text", + "content": "In summary, to address the challenges in achieving 4-bit diffusion models, we propose a baseline method using a search-based signed FP quantization framework [2, 23] combined with single-LoRA fine-tuning, aiming to realize 4-bit FP quantized diffusion models. Through this exploration, we uncover several findings that are significant for both diffusion model quantization and FP quantization: (1) There exists many layers which exhibit asymmetric distributions, due to the nonlinear activation function " + }, + { + "bbox": [ + 56, + 170, + 296, + 456 + ], + "type": "inline_equation", + "content": "SiLU" + }, + { + "bbox": [ + 56, + 170, + 296, + 456 + ], + "type": "text", + "content": ". The application of traditional signed FP quantization with a symmetric distribution leads to significant precision loss in the sub-zero area, causing substantial performance degradation after quantization. (2) Fine-tuning is conducted based on the denoising process, which is regarded as a complex task involving the restoration from outlines to details [37]. However, current methods typically apply a single LoRA to fine-tune severely degraded models across all timesteps, which leads to suboptimal learning at certain timesteps. (3) Predicted noise plays a varying role at different timesteps during denoising, which is the key to diffusion models. Arising from the neglect of this variation across timesteps, we identify a mismatch between the impact of quantization and the loss function in current methods, which undermines the effectiveness of fine-tuning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 459, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 673 + ], + "type": "text", + "content": "Facing the challenges in achieving 4-bit FP quantization, we propose constructive strategies: (1) To handle different distribution effectively, we propose a mixup-sign floating-point quantization (MSFP) framework, where unsigned FP quantization with an added zero point leads to a more compatible distribution of discrete points with the anomalous distributions in activations and signed FP quantization continues to exhibit strong representation capacity on other distributions. (2) Realizing that the current single-LoRA fine-tuning approach is in lack of flexibility across timesteps, we introduce timestep-aware LoRA (TALoRA), incorporating multiple LoRAs and a timestep-aware router to dynamically select the appropriate LoRA for each timestep in the denoising process. (3) To further improve the effectiveness of fine-tuning, we introduce a denoising-factor loss alignment (DFA), ensuring the loss function, and the guidance of the fine-tuning, are consistent with the actual quantization deterioration across timesteps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 676, + 250, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 676, + 250, + 687 + ], + "spans": [ + { + "bbox": [ + 65, + 676, + 250, + 687 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 689, + 295, + 714 + ], + "type": "text", + "content": "(i) We are the first to identify that signed FP quantization struggles with asymmetric activations, which" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 338, + 72, + 555, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 72, + 555, + 132 + ], + "spans": [ + { + "bbox": [ + 338, + 72, + 555, + 132 + ], + "type": "text", + "content": "arise from the nonlinear behavior of activation functions. To address this, we introduce the MSFP framework, which is also the first effective application of unsigned FP quantization in quantization, offering a novel approach for achieving low-bit quantization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 322, + 133, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 133, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 322, + 133, + 553, + 215 + ], + "type": "text", + "content": "(ii) For the fine-tuning of low-bit diffusion models, which is based on the denoising process, we precisely define it as a multi-task process across timesteps and introduce an efficient TALoRA module. Furthermore, we improve the alignment of loss function with quantization error via DFA strategy, enabling stable and reliable fine-tuning to achieve low-bit diffusion models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 320, + 216, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 216, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 320, + 216, + 553, + 299 + ], + "type": "text", + "content": "(iii) We focus on identifying and eliminating three major barriers to effective low-bit FP quantization in diffusion models. Extensive experiments on DDIM and LDM demonstrate that our work achieves SOTA results for 4-bit quantization. The proposed MSFP, TALoRA and DFA have greatly advanced the progress of low-bit quantization in diffusion models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 314, + 400, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 314, + 400, + 326 + ], + "spans": [ + { + "bbox": [ + 314, + 314, + 400, + 326 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 335, + 477, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 335, + 477, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 335, + 477, + 346 + ], + "type": "text", + "content": "2.1. Diffusion Model Quantization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 354, + 555, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 580 + ], + "type": "text", + "content": "There are two main approaches in the quantization of diffusion models: QAT [6, 14] and PTQ [28]. QAT is particularly effective for low-bit quantization while it retransmits the model from scratch, consuming extensive computational resources and time [20]. In contrast, PTQ offers greater time efficiency, making it more practical for large models in real-world applications. Recent advancements in PTQ have concentrated on optimizing calibration datasets to enhance reconstruction accuracy [9, 19, 25, 32] and mitigating quantization errors stemming from the temporal and structural properties of diffusion models [3, 19, 33, 35]. To further achieve fully 4-bit quantization, finetuning techniques has been introduced in PTQ-based quantization. EfficientDM [8] develops a LoRA-based finetuning framework, while QuEST [39] focuses on optimizing quantization-unfriendly activations. Both works stagnate in addressing data distribution during fine-tuning, failing to account for the specific challenges involved in finetuning within the context of diffusion model quantization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Meanwhile, recent advancements in FP quantization have made significant strides in model quantization [17, 23], highlighting its potential for diffusion models. While existing low-bit quantization methods perform well for linear models, applying FP quantization to convolutional diffusion models remains more challenging. To date, only one study has explored 8-bit activation quantization in diffusion models using a basic search-based approach [2], with no research addressing lower-bit quantization. This underscores both the challenges and the untapped potential of achieving 4-bit FP quantization for diffusion models." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18135" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 223, + 205 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 223, + 205 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 223, + 205 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 223, + 205 + ], + "type": "image", + "image_path": "4e3effa50b5d8326230a2591b257ff69ba8f6f732d73b7721d59900932d8a318.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 555, + 244 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 555, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 555, + 244 + ], + "type": "text", + "content": "Figure 1. The activation distributions in NALs and AALs, results on the CelebA dataset. (a) The paradigm of NALs with symmetric activations. (b) The typical paradigm of AALs with asymmetric activations, where unsigned FP quantization is more suitable. (c) The infrequent paradigm of AALs with relatively symmetric activations, where either signed or unsigned FP quantization could be applicable." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 72, + 388, + 205 + ], + "blocks": [ + { + "bbox": [ + 224, + 72, + 388, + 205 + ], + "lines": [ + { + "bbox": [ + 224, + 72, + 388, + 205 + ], + "spans": [ + { + "bbox": [ + 224, + 72, + 388, + 205 + ], + "type": "image", + "image_path": "c83abc04fa0f7347d12b7a595d24dbb92053e026580e0418646e8eedbc838016.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 388, + 72, + 550, + 205 + ], + "blocks": [ + { + "bbox": [ + 388, + 72, + 550, + 205 + ], + "lines": [ + { + "bbox": [ + 388, + 72, + 550, + 205 + ], + "spans": [ + { + "bbox": [ + 388, + 72, + 550, + 205 + ], + "type": "image", + "image_path": "87fcb6d0f687d38b802457a5ff7c2538c91eaaa900b6a761ca3f35cb4afa78e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 231, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 231, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 231, + 276 + ], + "type": "text", + "content": "2.2. Parameter-Efficient Fine-Tuning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 280, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 296, + 435 + ], + "type": "text", + "content": "Parameter-efficient fine-tuning (PEFT) has emerged as an effective alternative to full model fine-tuning, focusing on adjusting only a subset of parameters while keeping the majority frozen, thereby reducing storage overhead. Low-rank adapters (LoRA) [12], originally developed for large language models, have become one of the most widely used PEFT methods. Leveraging LoRA's strong transferability, QLoRA [5] can be effectively applied to fine-tune low-bit diffusion models. However, previous LoRA-based fine-tuning is suboptimal in practice. In this paper, we make a further exploration on this and incorporate timestep-level adaptation inspired by MoELoRA [7], enhancing performance of low-bit quantized DMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 446, + 204, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 204, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 204, + 460 + ], + "type": "text", + "content": "3. Challenges & Exploration" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 466, + 135, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 135, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 135, + 479 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "content": "Diffusion Models. The diffusion model is a new generation framework that completes learning by adding noise and completes generation in a denoising manner. As for the forward process, the noise is injected into ground-truth images " + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "content": " at random timesteps. This enables the DMs to learn the distributions of noise through noisy images " + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "content": ", which can be obtained as follows [11]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 119, + 578, + 294, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 578, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 119, + 578, + 294, + 591 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\varepsilon . \\tag {1}", + "image_path": "0797612b5d0475d6e59cfceb6c704bba4fce18dcb4550bca84f2179fb88f1713.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 597, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 620 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 597, + 295, + 620 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 597, + 295, + 620 + ], + "type": "text", + "content": " represents a standard Gaussian noise and " + }, + { + "bbox": [ + 55, + 597, + 295, + 620 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}_t" + }, + { + "bbox": [ + 55, + 597, + 295, + 620 + ], + "type": "text", + "content": " is the accumulated noise intensity, calculated by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 148, + 629, + 295, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 629, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 148, + 629, + 295, + 661 + ], + "type": "interline_equation", + "content": "\\bar {\\alpha} _ {t} = \\prod_ {i = 1} ^ {t} \\alpha_ {i}, \\tag {2}", + "image_path": "9ddef0e15338300035a6b6d5c76bc6d7175e145a70168e83751e209ff872cc85.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{s}" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " governs the noise intensity under each timestep. Once the DMs are well-trained, the Gaussian noise image will be inputted to the DMs, and undergo the iterative denoising process. Specifically, the noise can be predicted" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 263, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 300 + ], + "type": "text", + "content": "by DMs and used to obtain the image " + }, + { + "bbox": [ + 313, + 263, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 313, + 263, + 555, + 300 + ], + "type": "text", + "content": " under timestep " + }, + { + "bbox": [ + 313, + 263, + 555, + 300 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 263, + 555, + 300 + ], + "type": "text", + "content": ", which could range from " + }, + { + "bbox": [ + 313, + 263, + 555, + 300 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 263, + 555, + 300 + ], + "type": "text", + "content": " to 0, with the objective of obtaining the image " + }, + { + "bbox": [ + 313, + 263, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{t-1}" + }, + { + "bbox": [ + 313, + 263, + 555, + 300 + ], + "type": "text", + "content": " in a superior quality:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 315, + 555, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 315, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 323, + 315, + 555, + 342 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\boldsymbol {x} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\overline {{\\alpha}} _ {t}}} \\cdot \\boldsymbol {\\varepsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right) + \\sigma_ {t} \\boldsymbol {\\delta}, (3)", + "image_path": "665fd1cc774671074c78accc4e6dafe5611a0f96d04bed6f99680df383774a95.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\theta}(\\boldsymbol{x}_t,t)" + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "text", + "content": " is the predicted noise at timestep " + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "text", + "content": " is a newly added noise with the factor " + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "text", + "content": " to ensure diverse results. Here we define a denoising factor " + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "inline_equation", + "content": "\\gamma_t" + }, + { + "bbox": [ + 313, + 346, + 555, + 394 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 387, + 403, + 553, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 403, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 387, + 403, + 553, + 430 + ], + "type": "interline_equation", + "content": "\\gamma_ {t} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\cdot \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}. \\tag {4}", + "image_path": "6a0b76ccd6adbb9ed17151d4c8f17d9065dbb6d4790b1a6386945ccf1f2c0f4e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 431, + 554, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 554, + 466 + ], + "type": "text", + "content": "According to Equation 3, " + }, + { + "bbox": [ + 313, + 431, + 554, + 466 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 313, + 431, + 554, + 466 + ], + "type": "text", + "content": " indicates the impact of the prediction noise under timestep " + }, + { + "bbox": [ + 313, + 431, + 554, + 466 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 431, + 554, + 466 + ], + "type": "text", + "content": ". The greater the factor " + }, + { + "bbox": [ + 313, + 431, + 554, + 466 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 313, + 431, + 554, + 466 + ], + "type": "text", + "content": ", the stronger the predicted noise effect in the denoising." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 467, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 575 + ], + "type": "text", + "content": "Model Quantization. The fundamental paradigm of model quantization is the process of transforming a continuous data distribution into a finite set of discrete points. Consequently, the quality of the quantized model is inextricably related to the distribution of discrete points. According to the discrete point type, quantization is defined as two categories, widely-used INT quantization and emerging FP quantization. Analogous to the process of INT quantization, a floating-point vector " + }, + { + "bbox": [ + 313, + 467, + 555, + 575 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 467, + 555, + 575 + ], + "type": "text", + "content": " can be quantized as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 371, + 582, + 553, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 582, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 371, + 582, + 553, + 604 + ], + "type": "interline_equation", + "content": "\\widehat {\\boldsymbol {x}} = \\operatorname {C l i p} \\left(\\left\\lfloor \\frac {\\boldsymbol {x}}{s} \\right\\rceil + z, l, u\\right) \\cdot s. \\tag {5}", + "image_path": "432b679179fe2dc301eccfd0afd97ac67e66c19247cfc420e91b0a6e0b324c0a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lfloor\\rfloor" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " is the rounding operation. " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " are the minimum and maximum quantization thresholds while scaling factor " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " and zero-point " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " together constitute the quantization parameters. As shown in Equation 5, INT quantization results in an evenly spaced distribution of discrete points. While INT quantization is straightforward to implement, it may be too simplistic for continuous distributions that have significant variations in density, where evenly spaced intervals in INT quantization are not effective. On the contrary," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18136" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 218, + 171 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 218, + 171 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 218, + 171 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 218, + 171 + ], + "type": "image", + "image_path": "6a26f2dab317b9b9951cc7615de13e4369501d93b6d456f95088259c55a95c58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 182, + 385, + 237 + ], + "lines": [ + { + "bbox": [ + 59, + 182, + 385, + 237 + ], + "spans": [ + { + "bbox": [ + 59, + 182, + 385, + 237 + ], + "type": "text", + "content": "Figure 2. Effect of bit-width reduction on Figure 3. Two loss, and performance degradation representation capacity in AALs (blue) and NALs (orange) under signed FP models across steps. Compared with metric, quantization, evaluated on CelebA dataset. the original loss shows an inverse trend, while the aligned loss remains consistent." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 71, + 383, + 171 + ], + "blocks": [ + { + "bbox": [ + 220, + 71, + 383, + 171 + ], + "lines": [ + { + "bbox": [ + 220, + 71, + 383, + 171 + ], + "spans": [ + { + "bbox": [ + 220, + 71, + 383, + 171 + ], + "type": "image", + "image_path": "53d938ec9cd6120a695018d1bc6c77e257b06fb20289876e696a0b67c248c4a1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 182, + 554, + 238 + ], + "lines": [ + { + "bbox": [ + 387, + 182, + 554, + 238 + ], + "spans": [ + { + "bbox": [ + 387, + 182, + 554, + 238 + ], + "type": "text", + "content": "Figure 4. The MSE of activations before and after quantization across all AALs under four different strategies, normalized against the baseline of signed FP quantization without zero point (purple)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 389, + 72, + 553, + 171 + ], + "blocks": [ + { + "bbox": [ + 389, + 72, + 553, + 171 + ], + "lines": [ + { + "bbox": [ + 389, + 72, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 389, + 72, + 553, + 171 + ], + "type": "image", + "image_path": "6c53f0d4bffe1523270302989a390884558814a6d1ecb309c6a260e6e9bed735.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 258, + 295, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 283 + ], + "type": "text", + "content": "the discrete points under FP quantization are not uniformly spaced in distribution, as shown below [22]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 299, + 295, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 299, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 83, + 299, + 295, + 326 + ], + "type": "interline_equation", + "content": "f = (- 1) ^ {s} 2 ^ {p - b} \\left(1 + \\frac {d _ {1}}{2} + \\frac {d _ {2}}{2 ^ {2}} + \\dots \\frac {d _ {m}}{2 ^ {m}}\\right), \\tag {6}", + "image_path": "96de71949d835b19cf663619700dbfd74edea7a53758e9a31b1b448bb3fca8f5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": " is the sign bit, " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": "-bit mantissa, " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": "-bit exponent and " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": " is the bias that serves as both scaling factor and threshold in INT quantization. Previous studies in FP quantization rely on signed FP quantization, where " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": " is set to 1. For an " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": "-bit FP quantization, the bit-width is distributed across the mantissa, exponent, and sign bit, with the condition " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "m + e + s = n" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": ". The different combinations of " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": " allow FP quantization to represent data in multiple formats with a fixed number of bits, denoted as " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "E_{i}M_{j}" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": "-bit exponent and " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": "-bit mantissa are specified. A larger " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": " results in higher precision within each interval, while a larger " + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 331, + 295, + 474 + ], + "type": "text", + "content": " expands the range of covered intervals." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 474, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 581 + ], + "type": "text", + "content": "There appears to be an inherent suitability of FP quantization for DMs, as the majority of weights and activations follow a normal distribution symmetric around value " + }, + { + "bbox": [ + 55, + 474, + 295, + 581 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 55, + 474, + 295, + 581 + ], + "type": "text", + "content": ". This aligns well with the unevenly spaced distribution, where discrete points are dense in the small-value region and sparse in the large-value region, in FP quantization [42]. Additionally, the flexible formatting mechanism allows FP quantization to better accommodate complex distribution scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "type": "text", + "content": "3.2. Barriers to Effective Low-Bit FP Quantization in Diffusion Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "In order to implement high-performing 4-bit diffusion models, we implemented FP quantization that is more compatible with the data distribution of diffusion models, with the search-based strategy in previous work [23]. Building on this, we deployed a LoRA-based fine-tuning strategy to address the performance degradation caused by 4-bit quantization. Despite this, we find that the performance of the 4-bit FP diffusion model remains satisfactory and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "type": "text", + "content": "we identify two main issues that we are facing: (1) For FP quantization, although FP quantization works well at 8 bits, it experiences a sharp performance degradation at 4 bits. How to improve the ability of FP quantization to represent data under low-bit quantization? (2) For standard post-training LoRA-based fine-tuning, exhibits instability and suboptimal results when applied to low-bit quantized diffusion models. How can we make LoRA more efficient and accurate in learning the loss information at different denoising timesteps? In the following parts, we will explore the underlying causes of these two issues and the feasible strategies to address them." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 404, + 554, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 554, + 441 + ], + "type": "text", + "content": "Observation 1: Previous signed FP quantization fails to achieve effective low-bit quantization in Activation-Anomalous Layers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 443, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 554, + 484 + ], + "type": "text", + "content": "In diffusion models, we observed that the nonlinear activation layer " + }, + { + "bbox": [ + 313, + 443, + 554, + 484 + ], + "type": "inline_equation", + "content": "SiLU" + }, + { + "bbox": [ + 313, + 443, + 554, + 484 + ], + "type": "text", + "content": ", defined as " + }, + { + "bbox": [ + 313, + 443, + 554, + 484 + ], + "type": "inline_equation", + "content": "SiLU(\\boldsymbol{x}) = \\frac{\\boldsymbol{x}}{1 + e^{-\\boldsymbol{x}}}" + }, + { + "bbox": [ + 313, + 443, + 554, + 484 + ], + "type": "text", + "content": ", is commonly situated between layers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 486, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 712 + ], + "type": "inline_equation", + "content": "SiLU" + }, + { + "bbox": [ + 313, + 486, + 555, + 712 + ], + "type": "text", + "content": " causes the abnormal activations for the subsequent layer. As depicted in Panel (b) of Figure 1, all values below 0 are compressed into the range of " + }, + { + "bbox": [ + 313, + 486, + 555, + 712 + ], + "type": "inline_equation", + "content": "[-0.278, 0)" + }, + { + "bbox": [ + 313, + 486, + 555, + 712 + ], + "type": "text", + "content": ". In this paper, we refer to layers with such asymmetric activations as Anomalous-Activation-Distribution Layers (AALs) and the other layers as Normal-Activation-Distribution Layers (NALs). Mainstream signed FP quantization typically sets the maximum threshold of the quantizer high to accurately represent normal positive activations. However, this approach results in a significant precision loss when dealing with values below 0. Figure 2 illustrates the representation capacity of FP signed quantization in both NALs and AALs under different bit widths. When the bit width drops below 6 bits, AALs suffer more severe performance degradation compared to NALs, which ultimately leads to the failure in low-bit FP quantization. This phenomenon suggests that mitigating the performance decline in AALs is a critical step towards improving low-bit FP quantization in diffusion models." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18137" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "Observation 2: The single-LoRA-based strategy is overly simplistic for fine-tuning quantized diffusion models across different timesteps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 109, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 109, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 109, + 295, + 357 + ], + "type": "text", + "content": "Previous work has focused on adapting LoRA for the quantization of diffusion models but has not fully explored LoRA's performance in the context of denoising. Considering that the denoising process of diffusion models starts with recovering outlines and progresses to restoring details, we question whether the single-LoRA fine-tuning strategy can handle this complexity. In Table 1, we compare the baseline model, which uses a single-LoRA strategy for fine-tuning, with two alternative strategies. The second strategy assigns a separate LoRA for the first and last 50 timesteps, resulting in a significant improvement over the baseline. In contrast, the third strategy also introduces a dual-LoRA strategy, it randomly selects one for each timestep, resulting in much worse performance. These results suggest that applying multiple LoRAs, allocated in a structured manner across timesteps, enhances model performance, while disordered selection of LoRAs could lead to suboptimal results. This motivates us to approach the fine-tuning of quantized diffusion models as a multi-task process and assign multiple LoRAs across different timesteps with a rational approach in allocation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 359, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 295, + 396 + ], + "type": "text", + "content": "Observation 3: The MSE of the predicted noise of full-precision and quantized models does not reflect the actual impact of quantization at different timesteps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": "In fine-tuning, a commonly used loss function calculates the MSE between the noise predictions of the full-precision and quantized diffusion models, using denoised images from the full-precision model at the previous timestep as inputs:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 466, + 295, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 466, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 107, + 466, + 295, + 482 + ], + "type": "interline_equation", + "content": "L _ {\\boldsymbol {\\varepsilon} _ {\\theta}} ^ {t} = \\left\\| \\boldsymbol {\\varepsilon} _ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\widehat {\\boldsymbol {\\varepsilon}} _ {\\theta} (\\boldsymbol {x} _ {t}, t) \\right\\| ^ {2}. \\tag {7}", + "image_path": "d3c8540f2c029fe9f85db28763891a50eb4a494aa20cbcbc38edfbd1aaf853f0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 628 + ], + "type": "text", + "content": "By observing the variation in this loss during single-LoRA fine-tuning (see Figure 3), we identify an unexpected trend: the loss increases progressively faster as denoising advances. This contradicts the principle of denoising, where image quality should improve with each step, and the impact of predicted noise should diminish over time. By the final step, the impact of quantization on model performance should be negligible, as the predicted noise no longer affects the input. However, the loss is at its maximum, indicating that quantization error is most significant at this stage, contradicting the expectation that its influence should diminish over time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "To highlight the discrepancy between the loss and actual quantization errors, we define the performance gap at each step as the difference in denoising quality between the quantized and full-precision models, as the ultimate goal of denoising is to yield high-quality images. We input the previous output image " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": " from the full-precision model and calculate the performance gap between the de" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 70, + 543, + 144 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 543, + 144 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 543, + 144 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 543, + 144 + ], + "type": "table", + "html": "
MethodBits (W/A)FID ↓
FP32/326.49
Single-LoRA4/419.41
Dual-LoRA (Split Steps in Half)4/417.07
Dual-LoRA (Random Allocation)4/441.96
", + "image_path": "d673ed7a4ad95207386a65a232d0f92875a9330231a7be884fbb61924e6895d4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 152, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 185 + ], + "type": "text", + "content": "Table 1. The impact of the number of LoRAs and their allocation across timesteps on the performance of the fine-tuning. The results is evaluated by 4-bit quantization on CelebA dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 206, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 553, + 290 + ], + "type": "text", + "content": "noised image " + }, + { + "bbox": [ + 313, + 206, + 553, + 290 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_{t-1}" + }, + { + "bbox": [ + 313, + 206, + 553, + 290 + ], + "type": "text", + "content": " from the full-precision model and the denoised image " + }, + { + "bbox": [ + 313, + 206, + 553, + 290 + ], + "type": "inline_equation", + "content": "\\widehat{\\boldsymbol{x}}_{t-1}" + }, + { + "bbox": [ + 313, + 206, + 553, + 290 + ], + "type": "text", + "content": " from the quantized model, measured by " + }, + { + "bbox": [ + 313, + 206, + 553, + 290 + ], + "type": "inline_equation", + "content": "MSE(\\boldsymbol{x}_{t-1}, \\widehat{\\boldsymbol{x}}_{t-1})" + }, + { + "bbox": [ + 313, + 206, + 553, + 290 + ], + "type": "text", + "content": ". As shown in Figure 3, this misalignment between the loss and the actual performance gap leads to deviations in LoRA's learning. This necessitates aligning our loss function with the denoising process during fine-tuning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 301, + 397, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 397, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 397, + 316 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 322, + 554, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 554, + 417 + ], + "type": "text", + "content": "In this section, we explore the issues identified in Section 3.2 and propose corresponding solutions. As illustrated in Figure 5, we introduce a mixup-sign FP quantization framework to address the diverse activation distributions in the first stage. During fine-tuning, the timestep-aware routing mechanism and denoising-factor loss alignment work in tandem to enable high-quality learning, ultimately enabling the realization of optimized 4-bit FP diffusion models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 526, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 526, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 526, + 438 + ], + "type": "text", + "content": "4.1. Mixup-Sign Floating Point Quantization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 443, + 554, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 554, + 574 + ], + "type": "text", + "content": "To address the challenges of low-bit FP quantization failure in AALs, we leverage FP quantization by allocating more discrete points to areas with high data concentration. Motivated by the half-normal distribution of activations in AALs, we introduce unsigned floating-point quantization. However, as shown in Equation 6, when using unsigned FP quantization with " + }, + { + "bbox": [ + 313, + 443, + 554, + 574 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 443, + 554, + 574 + ], + "type": "text", + "content": " set to 0, we round all data in the sub-zero range to zero, losing important negative information. To address this, we introduce a zero point in the range of [-0.278, 0) to recover most sub-zero activations. The updated quantization formula becomes:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 322, + 594, + 553, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 594, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 322, + 594, + 553, + 629 + ], + "type": "interline_equation", + "content": "f _ {u n s i g n} = (- 1) ^ {s} 2 ^ {p - b} \\left(1 + \\frac {d _ {1}}{2} + \\frac {d _ {2}}{2 ^ {2}} + \\dots \\frac {d _ {m}}{2 ^ {m}}\\right) + z, \\tag {8}", + "image_path": "56403bf6ea26b15d99f80c23c0e8d67f8143da3b049e506d1b9d7b284ac693db.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " is set to 0 and " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " is the newly added zero point. By freeing the 1-bit sign bit, which is ineffective in signed FP quantization, and using it as additional exponent / mantissa bit width, we fully utilize the representation capacity. As illustrated in Figure 4, unsigned FP quantization with a zero point significantly improves representation in over " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " of AALs, compared to traditional signed FP quantization." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18138" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 541, + 239 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 239 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 239 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 239 + ], + "type": "image", + "image_path": "42465d15d6c8f7fb8080a1fedefa5e8e6339dcd958c96419d84c15eba4ea4e57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 75, + 242, + 211, + 302 + ], + "blocks": [ + { + "bbox": [ + 75, + 242, + 211, + 302 + ], + "lines": [ + { + "bbox": [ + 75, + 242, + 211, + 302 + ], + "spans": [ + { + "bbox": [ + 75, + 242, + 211, + 302 + ], + "type": "image", + "image_path": "732cdcf2680b76c4b427472166d45ba4ed70817c700a3c08e2fe4f54cfeffcf4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 314, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 54, + 314, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 314, + 555, + 370 + ], + "type": "text", + "content": "Figure 5. The pipeline of our proposed method. UNets are applied to the Mixup-Sign Floating-Point Quantization (MSFP), where distinct floating-point quantization schemes are employed for Anomalous-Activation-Distribution Layers (AALs) and Normal-Activation-Distribution Layers (NALs). During the fine-tuning stage, multiple LoRA modules are introduced, and a timestep-aware routing mechanism is used for dynamic LoRA allocation across different timesteps. Additionally, a denoising-factor alignment technique is employed to align the loss function with quantization-induced performance degradation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 219, + 241, + 399, + 304 + ], + "blocks": [ + { + "bbox": [ + 219, + 241, + 399, + 304 + ], + "lines": [ + { + "bbox": [ + 219, + 241, + 399, + 304 + ], + "spans": [ + { + "bbox": [ + 219, + 241, + 399, + 304 + ], + "type": "image", + "image_path": "5a2e2cf7780dc2a47983d69b83aaf68c13032b4fb54287b32209c352dacbaea0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 404, + 241, + 526, + 304 + ], + "blocks": [ + { + "bbox": [ + 404, + 241, + 526, + 304 + ], + "lines": [ + { + "bbox": [ + 404, + 241, + 526, + 304 + ], + "spans": [ + { + "bbox": [ + 404, + 241, + 526, + 304 + ], + "type": "image", + "image_path": "94473f604c4f4dd4b84b70a27c3788b0bbcbfba33e6c6a81635e404c58974655.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 390, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 390, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 295, + 484 + ], + "type": "text", + "content": "However, there are rare cases where performance slightly goes worse due to the diversity of anomalous distributions. Panel (c) of Figure 1 shows that in such cases, the activation distribution may resemble a normal distribution, where signed FP quantization might perform better. Figure 4 further indicates that introducing a zero point into signed FP quantization is unnecessary, offering minimal improvement in a few cases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 486, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 295, + 594 + ], + "type": "text", + "content": "Given the strong performance of unsigned FP quantization with a zero point and the diversity of AALs, we propose a mixup-sign FP quantization framework. During the search-based initialization phase, we use signed FP quantization for NALs and introduce both unsigned FP quantization with a zero point and signed FP quantization for AALs. This approach addresses AAL challenges in low-bit quantization while minimizing computational overhead by adding the zero point only to unsigned FP quantization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 600, + 289, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 289, + 613 + ], + "type": "text", + "content": "4.2. Timestep-Aware Router for LoRA Allocation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "text", + "content": "In Section 3.2, we observe that a single LoRA cannot fully capture all the information across timesteps due to the diverse generative characteristics at different timesteps. We also find that a reasonable allocation of different LoRAs for timesteps will be beneficial to fine-tuning. In this section, we introduce a timestep-aware LoRA allocation method that dynamically assigns optimal LoRA to each timestep, maximizing fine-tuning effectiveness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "content": "Our method relies on a learnable router (illustrated in Figure 5), a module shared across all timesteps. It takes the timestep as input and outputs selection probabilities for each LoRA across UNet layers. For each timestep, the LoRA with the highest probability corresponding to the router's output is inserted into the quantized model for fin-tuning or inference. In a router network, the main components are a time embedding layer and an MLP layer. The time embedding layer, derived from a pre-trained diffusion model, converts the scalar into a " + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "content": "-dimensional embedding. The MLP layer then maps the embedding to a LoRA allocation distribution, where " + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "content": " is the number of quantized layers and " + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "content": " is the size of the LoRA Hub. Using an STE method [1], this distribution is converted into 0/1 probabilities to allocate suitable LoRAs to the diffusion model's quantized layers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 588, + 480, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 480, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 480, + 601 + ], + "type": "text", + "content": "4.3. Denoising Factor Aligned Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": "To address the mismatch between the actual performance gap and the loss used during quantization, we review the denoising principle depicted in Equation 3 and pinpoint the cause of the mismatch: the time-step-dependent constraint on the predicted noise is not sufficiently accounted for during the denoising process. Therefore, we implement a modification to the loss function based on predicted noise:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 403, + 700, + 554, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 700, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 403, + 700, + 554, + 715 + ], + "type": "interline_equation", + "content": "L ^ {t} = \\gamma_ {t} \\cdot L _ {\\varepsilon_ {\\theta}} ^ {t}. \\tag {9}", + "image_path": "1a37fe29711348245b2da46f86cf6c2ea65014bbb51f8f2674b9426d6bd329da.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18139" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": "By introducing " + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": ", which accurately reflects the utilization of the predicted noise at each time step, we achieve a preliminary alignment between the loss and the actual quantization error, as shown in Figure 3. This facilitates more accurate fine-tuning, leading to better performance recovery in the low-bit diffusion model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 132, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 132, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 132, + 170 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 176, + 171, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 176, + 171, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 171, + 190 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 194, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 326 + ], + "type": "text", + "content": "Models and Metrics. To verify the effectiveness of the proposed method, we evaluate it with two widely adopted diffusion paradigms: DDIM [34] and LDM [30]. For DDIM experiments, we evaluate on CIFAR-10 [16] and CelebA [26]. For LDM, we test unconditional generation on LSUN-Bedroom [41] and LSUN-Church [41] and conditional generation on ImageNet [4]. The performance of the diffusion models is evaluated with Inception Score (IS) [31], Fréchet Inception Distance (FID) [10] and Sliding FID (sFID) [31]. All metrics are evaluated based on 50k samples generated by the DDIM solver [34]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 422 + ], + "type": "text", + "content": "Quantization Detail. We employ standard layer-wise quantization for both weights and activations. Except for the input and output layers, which are typically set to 8-bit, all other convolution and linear layers are quantized to the target bit-width. Furthermore, we generate 256 samples for the calibration set based on Q-Diffusion [19] for bias initialization and use the method from [2] to obtain the optimal quantization parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 519 + ], + "type": "text", + "content": "Baseline. We compare two main types of quantization methods: PTQ methods (Q-Diffusion [19] and EDADM [38]) and fine-tuning methods (EfficientDM [8] and QuEST [39]). Since these fine-tuning methods involve special settings like non-full-layer quantization, we standardize the settings of EfficientDM in a consistent manner and procure other standardized results from DilateQuant [24]. Comparison with special settings is provided in Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 203, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 203, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 203, + 540 + ], + "type": "text", + "content": "5.2. Quantization Performance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Unconditional Generation. Table 2 presents the results of unconditional image generation across multiple datasets. With 6-bit quantization, our method achieves nearly identical performance to full precision. Our 4-bit quantized models achieve SOTA results in all tasks, significantly outperforming previous baseline methods. Notably, on CIFAR-10, our 4-bit quantized model results in an FID score that is only 1.84 worse than full precision, an almost negligible degradation, while previous methods struggle with 4-bit quantization. Compared to the fine-tuning method [8], our 4-bit quantized models improve FID by 32.38, 24.15, and 9.63 on three datasets, respectively, and still maintain remarkable IS. Additionally, we provide the performance of 4-bit and 6-bit quantized models on CelebA in Appendix." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 328, + 72, + 547, + 563 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 547, + 563 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 547, + 563 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 547, + 563 + ], + "type": "table", + "html": "
TaskMethodPrec. (W/A)FID ↓IS ↑
CIFAR-10 32x32FP32/324.269.03
Q-Diffusion6/69.198.76
EDA-DM6/626.689.35
EfficientDM6/625.038.08
Ours (h=2)6/64.269.04
DDIM steps = 100Ours (h=4)6/64.239.06
Q-Diffusion4/4N/AN/A
EDA-DM4/4120.244.42
EfficientDM4/438.407.32
Ours (h=2)4/46.028.79
Ours (h=4)4/46.108.90
LSUN (Bedroom) 256x256FP32/323.022.29
Q-Diffusion6/610.102.11
EDA-DM6/610.562.12
QuEST6/610.102.20
EfficientDM6/612.952.57
Ours (h=2)6/68.422.49
Ours (h=4)6/68.402.49
LDM-4 steps = 100 eta = 1.0Q-Diffusion4/4N/AN/A
EDA-DM4/4N/AN/A
QuEST4/4N/AN/A
EfficientDM4/436.362.69
Ours (h=2)4/412.212.47
Ours (h=4)4/412.342.48
LSUN (Church) 256x256FP32/324.062.70
Q-Diffusion6/610.902.47
EDA-DM6/610.762.43
QuEST6/66.832.65
EfficientDM6/67.452.80
Ours (h=2)6/66.242.73
Ours (h=4)6/66.382.73
LDM-8 steps = 100 eta = 0.0Q-Diffusion4/4N/AN/A
EDA-DM4/4N/AN/A
QuEST4/413.032.63
EfficientDM4/418.402.97
Ours (h=2)4/48.812.70
Ours (h=4)4/48.772.71
", + "image_path": "b49d4f1e67b3fa6778bc2b1c403f4ec0574f08edddb692b324b227fcc7ebb321.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 572, + 553, + 607 + ], + "lines": [ + { + "bbox": [ + 313, + 572, + 553, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 553, + 607 + ], + "type": "text", + "content": "Table 2. Quantization performance of unconditional generation. 'Prec. (W/A)' denotes the quantization bit-width. 'N/A' denotes failed image generation. " + }, + { + "bbox": [ + 313, + 572, + 553, + 607 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 572, + 553, + 607 + ], + "type": "text", + "content": " denotes the size of LoRA Hub." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Conditional Generation. Table 3 presents the results of our conditional generation experiments on ImageNet. We observe that FID is not always a reliable metric in this context, as an unexpected trend emerges: as the bit-width of the quantized model decreases, the FID score improves, which contradicts the expected trend. Therefore, our discussion of the ImageNet results focuses on other evaluation metrics. As shown by the IS and sFID, our method achieves" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18140" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 296, + 224 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 296, + 224 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 296, + 224 + ], + "type": "image", + "image_path": "b44320fb47421170619870e8e825a81b7066ba6a69d1b107979d2956c7c66a28.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 266 + ], + "type": "text", + "content": "Figure 6. A visual comparison of generation results using our method across different quantization bit-widths, with the LSUN-Church dataset as an example." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 384 + ], + "type": "text", + "content": "performance comparable to that of the full-precision model with 6-bit quantization. Even with 4-bit quantization, we achieve SOTA results in terms of sFID, improving by 7.00 over the previous best method, EfficientDM [8]. Furthermore, our method significantly outperforms two other fin-tuning methods in the IS metric. Visual evaluations further confirm that the generated images maintain high quality, exhibiting clear and coherent content, as shown in Figure 6. More visualization results will be presented in Appendix." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 65, + 392, + 288, + 547 + ], + "blocks": [ + { + "bbox": [ + 65, + 392, + 288, + 547 + ], + "lines": [ + { + "bbox": [ + 65, + 392, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 65, + 392, + 288, + 547 + ], + "type": "table", + "html": "
MethodPrec. (W/A)sFID ↓FID ↓IS ↑
FP32/327.6711.69364.72
EDA-DM6/68.0211.52360.77
QuEST6/69.368.45310.12
EfficientDM6/66.889.54351.79
Ours (h=2)6/67.4310.10349.91
Ours (h=4)6/66.6510.10351.79
EDA-DM4/436.6620.02204.93
QuEST4/429.2738.4369.58
EfficientDM4/414.4212.73139.45
Ours (h=2)4/47.426.50190.74
Ours (h=4)4/48.237.43177.40
", + "image_path": "d3c1c0b4956f47ea59a398b248da49522363d6cc522746e3da290c1e90ed3b52.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 612, + 149, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 149, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 149, + 625 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "The ablation experiments are conducted on the 4-bit quantization using the CelebA dataset, which is challenging for low-bit quantization, further demonstrating the effectiveness of our approach. The baseline uses signed FP quantization combined with single LoRA fine-tuning. As shown in Table 4, all three proposed modules lead to significant performance improvements, with their combination" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 328, + 70, + 543, + 179 + ], + "blocks": [ + { + "bbox": [ + 55, + 553, + 296, + 597 + ], + "lines": [ + { + "bbox": [ + 55, + 553, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 296, + 597 + ], + "type": "text", + "content": "Table 3. Quantization performance of conditional generation for fully-quantized LDM-4 models on ImageNet " + }, + { + "bbox": [ + 55, + 553, + 296, + 597 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 553, + 296, + 597 + ], + "type": "text", + "content": " with 20 steps. 'Prec. (W/A)' denotes the quantization bit-width. " + }, + { + "bbox": [ + 55, + 553, + 296, + 597 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 553, + 296, + 597 + ], + "type": "text", + "content": " denotes the size of LoRA Hub." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 70, + 543, + 179 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 543, + 179 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 543, + 179 + ], + "type": "table", + "html": "
MSFPMethod TALoRADFAPrec. (W/A)FID ↓
XXX4/416.02
✓XX4/49.60
X✓X4/410.66
✓X✓4/48.39
✓✓X4/48.79
✓✓✓4/47.69
", + "image_path": "2a40426ae05b4a2168b648b643243d91d50736d785496ada4c0e7c8bca69f292.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 187, + 554, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 187, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 554, + 209 + ], + "type": "text", + "content": "Table 4. Ablation study on different modules we proposed. Testing on CelebA dataset with " + }, + { + "bbox": [ + 313, + 187, + 554, + 209 + ], + "type": "inline_equation", + "content": "h = 2" + }, + { + "bbox": [ + 313, + 187, + 554, + 209 + ], + "type": "text", + "content": " LoRA Hub size." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 223, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 331 + ], + "type": "text", + "content": "yielding a synergistic effect. The baseline FID score is 9.53 higher than that of the full-precision model (6.49). By applying our technique, we reduce the FID by 8.18 compared to the baseline. More interesting is that we visualize the LoRA allocation distribution learned by the router as shown in Figure 7. We find that the distribution of the router-learned allocation over timesteps is consistent with the finding that the diffusion model focuses on contour generation early and on detail generation later [37]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 348, + 539, + 479 + ], + "blocks": [ + { + "bbox": [ + 331, + 348, + 539, + 479 + ], + "lines": [ + { + "bbox": [ + 331, + 348, + 539, + 479 + ], + "spans": [ + { + "bbox": [ + 331, + 348, + 539, + 479 + ], + "type": "image", + "image_path": "80925daab5f8ace3bc40f76b239165ee63f0e19ef2fd0dbcf7d5fd7f43f265d2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 487, + 554, + 510 + ], + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 510 + ], + "type": "text", + "content": "Figure 7. Distribution of LoRA allocations over timesteps obtained after router training on different datasets, when " + }, + { + "bbox": [ + 313, + 487, + 554, + 510 + ], + "type": "inline_equation", + "content": "h = 2" + }, + { + "bbox": [ + 313, + 487, + 554, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 534, + 388, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 534, + 388, + 546 + ], + "spans": [ + { + "bbox": [ + 314, + 534, + 388, + 546 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "type": "text", + "content": "In this paper, we focus on exploring low-bit FP quantization for diffusion models. For model initialization, we innovatively introduce unsigned FP quantization with zero point to address AALs. For the fine-tuning based on the denoising process, we formulate it as a multi-task procedure. We introduce multiple LoRAs along with a router for their allocation at different timesteps, and further align the loss function, originally based on estimated noise, with the actual quantization error. We introduce unsigned FP quantization and achieve 4-bit FP quantized diffusion models. Our FP PTQ-based fine-tuning method sets a new precedent for 4-bit diffusion models, offering insight into the deployment of low-bit diffusion models in the future." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18141" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 169 + ], + "type": "text", + "content": "This work is supported by Shanghai Science and Technology Commission Explorer Program Project (24TS1401300), National Key Research and Development Program of China (No.2022ZD0160101). The computations in this research were performed using the CFFF platform of Fudan University." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 198, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 198, + 295, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 198, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 295, + 242 + ], + "type": "text", + "content": "[1] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 243, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 296, + 287 + ], + "type": "text", + "content": "[2] Cheng Chen, Christina Giannoula, and Andreas Moshovos. Low-bitwidth floating point quantization for efficient high-quality diffusion models. arXiv preprint arXiv:2408.06995, 2024. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 289, + 295, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 289, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 61, + 289, + 295, + 332 + ], + "type": "text", + "content": "[3] Huanpeng Chu, Wei Wu, Chengjie Zang, and Kun Yuan. Qncd: Quantization noise correction for diffusion models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10995-11003, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 334, + 296, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 334, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 61, + 334, + 296, + 378 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 379, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 379, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 379, + 296, + 421 + ], + "type": "text", + "content": "[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: efficient finetuning of quantized llms (2023). arXiv preprint arXiv:2305.14314, 52:3982-3992, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 423, + 296, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 296, + 467 + ], + "type": "text", + "content": "[6] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. In International Conference on Learning Representations, 2019. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 468, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 534 + ], + "type": "text", + "content": "[7] Wenfeng Feng, Chuzhan Hao, Yuewei Zhang, Yu Han, and Hao Wang. Mixture-of-loras: An efficient multitask tuning method for large language models. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 11371-11380, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 535, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 296, + 578 + ], + "type": "text", + "content": "[8] Yefei He, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Efficientdm: Efficient quantization-aware fine-tuning of low-bit diffusion models. arXiv preprint arXiv:2310.03270, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 580, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 624 + ], + "type": "text", + "content": "[9] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 60, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 60, + 624, + 295, + 678 + ], + "type": "text", + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 58, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 680, + 295, + 714 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 554, + 172 + ], + "type": "text", + "content": "[13] Wei Huang, Xingyu Zheng, Xudong Ma, Haotong Qin, Chengtao Lv, Hong Chen, Jie Luo, Xiaojuan Qi, Xianglong Liu, and Michele Magno. An empirical study of llama3 quantization: From llms to mllms. Visual Intelligence, 2(1): 36, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 554, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 240 + ], + "type": "text", + "content": "[14] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2704-2713, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[15] Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 297 + ], + "type": "text", + "content": "[16] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "text", + "content": "[17] Andrey Kuzmin, Mart Van Baalen, Yuwei Ren, Markus Nagel, Jorn Peters, and Tijmen Blankevoort. Fp8 quantization: The power of the exponent. Advances in Neural Information Processing Systems, 35:14651-14662, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "text", + "content": "[18] Min Li, Zihao Huang, Lin Chen, Junxing Ren, Miao Jiang, Fengfa Li, Jitao Fu, and Chenghua Gao. Contemporary advances in neural network quantization: A survey. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-10. IEEE, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "text", + "content": "[19] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1, 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "text", + "content": "[20] Yanjing Li, Sheng Xu, Xianbin Cao, Xiao Sun, and Baochang Zhang. Q-dm: An efficient low-bit quantized diffusion model. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[21] Tailin Liang, John Glossner, Lei Wang, Shaobo Shi, and Xiaotong Zhang. Pruning and quantization for deep neural network acceleration: A survey. Neurocomputing, 461:370-403, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 545, + 554, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 554, + 611 + ], + "type": "text", + "content": "[22] Fangxin Liu, Wenbo Zhao, Zhezhi He, Yanzhi Wang, Zongwu Wang, Changzhi Dai, Xiaoyao Liang, and Li Jiang. Improving neural network efficiency via post-training quantization with adaptive floating-point. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5281-5290, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "text", + "content": "[23] Shih-yang Liu, Zechun Liu, Xijie Huang, Pingcheng Dong, and Kwang-Ting Cheng. LIm-fp4: 4-bit floating-point quantized transformers. arXiv preprint arXiv:2310.16836, 2023. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "text", + "content": "[24] Xuewen Liu, Zhikai Li, and Qingyi Gu. Dilatequant: Accurate and efficient diffusion quantization via weight dilation. arXiv preprint arXiv:2409.14307, 2024. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[25] Xuewen Liu, Zhikai Li, Junrui Xiao, and Qingyi Gu. Enhanced distribution alignment for post-training quantization" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18142" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "of diffusion models. arXiv preprint arXiv:2401.04585, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 297, + 139 + ], + "type": "text", + "content": "[26] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "type": "text", + "content": "[27] Paulius Micikevicius, Dusan Stosic, Neil Burgess, Marius Cornea, Pradeep Dubey, Richard Grisenthwaite, Sangwon Ha, Alexander Heinecke, Patrick Judd, John Kamalu, et al. Fp8 formats for deep learning. arXiv preprint arXiv:2209.05433, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "type": "text", + "content": "[28] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 308 + ], + "type": "text", + "content": "[29] NVIDIA. Blackwell platform sets new llm inference records in mlperf inference v4.1, 2024. Available at: https://developer.nvidia.com/blog/nvidia-blackwell-platform-sets-new-llm-inference-records-in-mlperf-inference-v4-1, Accessed: 2024-11-14. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 365 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "text", + "content": "[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 454 + ], + "type": "text", + "content": "[32] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 498 + ], + "type": "text", + "content": "[33] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 533 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "text", + "content": "[35] Haojun Sun, Chen Tang, Zhi Wang, Yuan Meng, Xinzhu Ma, Wenwu Zhu, et al. Tmpq-dm: Joint timestep reduction and quantization precision selection for efficient diffusion models. arXiv preprint arXiv:2404.09532, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 633 + ], + "type": "text", + "content": "[36] Mart van Baalen, Andrey Kuzmin, Suparna S Nair, Yuwei Ren, Eric Mahurin, Chirag Patel, Sundar Subramanian, Sanghyuk Lee, Markus Nagel, Joseph Soriaga, et al. Fp8 versus int8 for efficient deep learning inference. arXiv preprint arXiv:2303.17951, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 669 + ], + "type": "text", + "content": "[37] Binxu Wang and John J Vastola. Diffusion models generate images like painters: an analytical theory of outline first, details later. arXiv preprint arXiv:2303.02490, 2023. 2, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[38] Changyuan Wang, Ziwei Wang, Xiuwei Xu, Yansong Tang, Jie Zhou, and Jiwen Lu. Towards accurate data-free quantization for diffusion models. arXiv preprint arXiv:2305.18723, 2(5), 2023. 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 274 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[39] Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, Junchi Yan, and Yan Yan. Quest: Low-bit diffusion model quantization via efficient selective finetuning. arXiv preprint arXiv:2402.03666, 2024. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "text", + "content": "[40] Jie Wang, Huanxi Liu, Dawei Feng, Jie Ding, and Bo Ding. Fp4-quantization: Lossless 4bit quantization for large language models. In 2024 IEEE International Conference on Joint Cloud Computing (JCC), pages 61–67. IEEE, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "type": "text", + "content": "[41] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 1, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 554, + 274 + ], + "type": "text", + "content": "[42] Yijia Zhang, Lingran Zhao, Shijie Cao, Sicheng Zhang, Wenqiang Wang, Ting Cao, Fan Yang, Mao Yang, Shang-hang Zhang, and Ningyi Xu. Integer or floating point? new outlooks for low-bit quantization on large language models. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2, 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18143" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/2679a0a3-d3fd-49c2-aeba-9db847b48300_content_list.json b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/2679a0a3-d3fd-49c2-aeba-9db847b48300_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e00e8498bfa9346e91c20a3717c8ec4f8ce7eb96 --- /dev/null +++ b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/2679a0a3-d3fd-49c2-aeba-9db847b48300_content_list.json @@ -0,0 +1,1556 @@ +[ + { + "type": "text", + "text": "Pippo: High-Resolution Multi-View Humans from a Single Image", + "text_level": 1, + "bbox": [ + 166, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yash Kant $^{1,2,3}$ , Ethan Weber $^{1,4}$ , Jin Kyu Kim $^{1}$ , Rawal Khirodkar $^{1}$ , Su Zhaoen $^{1}$ , Julieta Martinez $^{1}$ , Igor Gilitschenski $^{*,2,3}$ , Shunsuke Saito $^{*,1}$ , Timur Bagautdinov $^{*,1}$ $^{1}$ Meta Reality Labs, $^{2}$ University of Toronto, $^{3}$ Vector Institute, $^{4}$ UC Berkeley https://yashkant.github.io/pippo/ \n*Joint Advising", + "bbox": [ + 116, + 179, + 880, + 277 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c5cebda302024c92c2938cc7a88a490127e4df0e93ba112539057ae62f0db3a8.jpg", + "image_caption": [ + "Figure 1. Pippo produces high-resolution, multi-view, studio-quality images from a single image. First row shows full-body generation of an unseen subject captured from an iPhone picture. Second and third rows show face-only and partial full-body input images from unseen studio subject(s) to novel views respectively. All visuals are at 1K resolution." + ], + "image_footnote": [], + "bbox": [ + 91, + 286, + 903, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 794, + 326, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Pippo, a generative model capable of producing 1K resolution dense turnaround videos of a person from a single casually clicked photo. Pippo is a multi-view diffusion transformer and does not require any additional", + "bbox": [ + 88, + 840, + 480, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "inputs — e.g., a fitted parametric model or camera parameters of the input image. We pre-train Pippo on 3B human images without captions, and conduct multi-view mid-training and post-training on studio captured humans. During mid-training, to quickly absorb the studio dataset, we denoise several (upto 48) views at low-resolution, and encode tar", + "bbox": [ + 511, + 796, + 906, + 887 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16418", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "get cameras coarsely using a shallow MLP. During posttraining, we denoise fewer views at high-resolution and use pixel-aligned controls (e.g., Spatial anchor and Plucker rays) to enable 3D consistent generations. At inference, we propose an attention biasing technique that allows Pippo to simultaneously generate greater than $5 \\times$ as many views as seen during training. Finally, we also introduce an improved metric to evaluate 3D consistency of multi-view generations, and show that Pippo outperforms existing works on multi-view human generation from a single image.", + "bbox": [ + 89, + 90, + 480, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 277, + 222, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Creating photorealistic human representations with the ability to control viewpoints has numerous applications in entertainment, healthcare, fashion and social media. Building such representations, first and foremost, requires high-quality multi-view studio data [2, 29], which is costly to acquire. This significantly limits the scalability in terms of the number identities for high-quality studio data [29, 43, 84]. In contrast, large-scale, unstructured, and diverse human images and videos are available online. However, the raw data of such in-the-wild images does not offer ground-truth 3D or multi-view representations of humans.", + "bbox": [ + 89, + 306, + 482, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present a novel approach leveraging the best of two worlds: generalizability from in-the-wild unstructured images, and fidelity and view-controllability from studio capture data. Specifically, our model Pippo is a diffusion transformer, which can generate several 1K resolution multi-view consistent images jointly during inference. Pippo takes as input a single image of an individual, camera poses of the target viewpoints to be generated. Since the scale and placement of the subject is ambiguous from a single image, Pippo uses a Spatial Anchor which roughly specifies the location and orientation of the subject in 3D space. Our model does not rely on additional conditioning such as body priors or camera parameters of the input images, to scale our training pipeline to in-the-wild data and support unconstrained inputs at test time.", + "bbox": [ + 89, + 474, + 482, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We employ a multi-stage training recipe to train Pippo. First, we pre-train the model for latent-to-image generation task, similar to [54], on a large human-centric dataset of in-the-wild images. Next, in Mid-training stage, we jointly generate multiple consistent images of the subject, conditioned on target viewpoints and a single input image using high-quality studio dataset. Finally, in Post-training stage, the model is provided with a minimal placement signal - spatial anchor - encoding rough head orientation - which further improves 3D consistency. Our architecture is also carefully tailored to the conditional multi-view generation - we propose several simple but effective modifications to the basic DiT architecture, including self-attention-based", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "conditioning, lightweight spatial controls and camera conditioning with Plucker coordinates.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For evaluating such a multi-view diffusion model, 3D consistency is a critical metric for understanding the geometric correctness of the generation. For generation task in an ill-posed setup (e.g., a single image input), there may be multiple possibilities that are equally plausible. Existing multi-view generation methods typically report reconstruction metrics (i.e., PSNR, SSIM and LPIPS) [19] or FID [41], which either (1) penalize any new content that is actually 3D consistent or (2) are unable to measure the 3D consistency of the generation. To address this problem, we design a metric measuring 3D consistency: we compute 2D keypoint matches with an off-the-shelf method [59], triangulate them, and then reproject back into the other views to measure the reprojection error in pixels. Some generation methods measure reprojection errors [17] from SfM [60] or epipolar distance [45, 89], but our metric uses camera pose as input and is more precise than measuring distance to a epipolar line. We show our metric helps to quantify our results, finding that our method is favorable compared with existing approaches and other baseline methods.", + "bbox": [ + 511, + 122, + 903, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we make the following contributions:", + "bbox": [ + 513, + 426, + 862, + 441 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- a generative model capable of generating high-resolution and multi-view consistent humans from a single image and its effective training strategy.", + "- a diffusion transformer architecture designed to enhance multi-view generation and viewpoint control.", + "- a novel 3D consistency metric tailored to accurately measure the level of 3D consistency in generative tasks." + ], + "bbox": [ + 514, + 443, + 903, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 566, + 653, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We review multi-view human datasets and generative models for human synthesis, categorizing methods by their data requirements and prior constraints—such as parametric human models or explicit 3D structures. In this work, we minimize reliance on complex priors by training the model on large amounts of human-centric data and impose minimal camera and spatial controls.", + "bbox": [ + 511, + 593, + 906, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-view Human Datasets. While many large 2D human datasets exist [12, 39, 62], 3d captures of people (i.e., captures using dozens of views at a time) are still relatively rare, since they are expensive to collect and are thus mostly obtained in specialized research labs. Nevertheless, this kind of data has recently become more common, going from hundreds to thousands of publicly-available captures. For face-only captures, a few datasets [43, 47, 87, 88, 90] provide up to 600 subjects captured from over 60 views each, while for full-body captures, 4D-Dress [77] provides 32 subjects with 2 outfits each, and detailed annotations for their garments, while DNA-rendering [8] provides 500 subjects from 60 views. MVHumanNet [84] stands apart by", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16419", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/274551e9bf4dd5a03572aa716c816b4fd81239c27280c6a69e5ac8d6d24eb537.jpg", + "image_caption": [ + "Figure 2. Pipeline overview. This is an illustration of how we train our model. (Left) we use data from a studio capture and train our multi-view diffusion model (right). We condition on a full reference photo and a cropped face, as well as the target view cameras and 2D projected spatial anchor indicating head position and orientation. Our diffusion model also takes in noisy target views and a timestep in order to predict the denoised views (top). In practice, we apply a segmentation mask around the person." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 883, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "providing 4,500 people captured from 48 views in and over 9,000 sets of clothing, outpacing other datasets by an order of magnitude in terms of people diversity. We use internal multi-view data of $\\sim 1000$ identities with dense ( $\\sim 160$ ) camera setup, but we expect our model to produce reasonable results with publicly-available captures.", + "bbox": [ + 88, + 375, + 482, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generating Novel Poses and Expressions. Although different from our task, this related work builds 3D animatable models of people and faces. These methods use neural radiance fields [70, 71, 81], or 3d Gaussians [34, 51, 80, 94], using canonical \"T\"-poses and a corresponding forward and backward mapping to model local appearance changes. Many of these methods, however, rely on accurate human shape and pose estimation, and only achieve high photorealism for personalized models from studio captures. Relatedly, a family of methods focus on generating animatable faces from a single image by disentangling viewpoints and expressions from large collections of 2d portraits [14, 15, 32, 48, 86, 92], allowing photorealistic real-time reenactment, while being limited to faces and small viewpoint variations. Unlike these methods, we complete missing parts of a person given either one or a few partial views. We do not repose the person or animate the faces but can instead recover entirely missing views, which is orthogonal to these methods. Moreover, our approach addresses both full-body and facial reconstruction, rather than specializing in either domain.", + "bbox": [ + 91, + 473, + 482, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generations with Explicit 3D Structures. Similar to early methods that extracted 3d representations from 2d models via score distillation [7, 11, 40, 50, 53, 74, 78], 3d representations of human appearance have been extracted from GANs by co-training a neural renderer from learned TriPlane [5] or HexPlane [1, 37] latent spaces. Others train generative models that operate directly in 3D space", + "bbox": [ + 88, + 794, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with diffusion or GAN objectives [18, 44, 72, 76, 82, 83] or regression models to directly predict 3D representations [56, 57, 63]. While these methods produce 3d consistent faces by design, their quality is limited by the relatively small amounts of 3d training data and/or the limitations of their respective 3d representations. In contrast, we focus on generating 3d-consistent 2d images and thus avoid the downsides of explicit 3d modelling.", + "bbox": [ + 511, + 375, + 906, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generations with 2D Models. Another option is to model viewpoint changes in 2d, without an underlying 3d representation [41], sometimes modelling either photometric [30] or epipolar [28] constraints explicitly. Diff-Portrait3D [20] is a portrait-generation method that falls roughly within this category, by fine-tuning a 2d diffusion model for 3d-aware face generation using ControlNet-style [93] conditioning for viewpoint control, as well as cross-view attention [21] and initialization from a 3d-consistent GAN [5]. Other methods are more general and not specific to humans, focusing on single-view to 3D generation [19, 41, 58, 64, 65, 73, 85]. In addition to these single-view methods, video models have been fine-tuned for camera control [23, 79].", + "bbox": [ + 511, + 498, + 908, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 729, + 604, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We train our models following a three-stage strategy:", + "bbox": [ + 511, + 757, + 864, + 773 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Image-only Pre-training (P1). We pretrain on a large-scale human-centric dataset with image conditioning.", + "- Multiview Mid-training (M2). We train models at a low-resolution of $128 \\times 128$ to denoise 48 target views with coarse camera control (no pixel-aligned spatial control).", + "- Multiview Post-training (P3). We train at a high-resolution of $1024 \\times 1024$ to denoise $1 - 3$ target views with spatial control injected via ControlMLP layers." + ], + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16420", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote the training stage and resolution of any given model as $\\{\\mathrm{stage}\\} @\\{\\mathrm{resolution}\\}$ . For instance, M2@128 represents a mid-trained model at 128 resolution.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Base Model", + "text_level": 1, + "bbox": [ + 89, + 146, + 217, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture. We adopt a DiT-like [49] architecture with scale, shift, and gate modulation for timestep conditioning inspired by Stable Diffusion 3 [16] and Flux [35]. We simplify the architecture by employing MLP and attention in parallel [91], and removing second LayerNorm after attention layers. We use VAE for training in latent space with 8x spatial compression, and patchify latent images via a linear layer and patch size of 2. We use fixed sinusoidal positional encoding during training. We provide more details in ??, and the exact design of our DiT block in Fig. 3.", + "bbox": [ + 89, + 167, + 482, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-only Pre-training. During pre-training, the model learns to denoise an image conditioned on its corresponding image embedding from DINOv2 [9, 46], this is similar in principle to the image decoder of DALL-E 2 [54]. We project both embeddings to the model dimension using with a linear layer to create a joint conditioning. Importantly, our pretraining setup does not require any annotations or captions for the images, and is well-aligned with our downstream objective of generating consistent multiview images given a single reference image as input.", + "bbox": [ + 89, + 319, + 482, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, given an image $\\mathbf{y} \\in \\mathbb{R}^{H \\times W \\times C}$ , and the joint conditioning as $\\mathbf{e}^{\\mathrm{img}} \\in \\mathbb{R}^{N \\times D}$ . We pre-train our diffusion model $\\epsilon_{\\theta}$ with the following objective:", + "bbox": [ + 89, + 469, + 482, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D M}} = \\left| \\left| \\boldsymbol {\\epsilon} ^ {t} - \\epsilon_ {\\theta} \\left(\\mathbf {y} ^ {t}, \\mathbf {e} ^ {\\mathrm {i m g}}, t\\right) \\right| \\right| ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 526, + 482, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $t \\in [0, T]$ is diffusion timestep, $\\epsilon^t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ is the noise added at the given timestep. We use the DDPM [26, 69] formulation to define discrete timesteps and set $T = 1000$ . We first pretrain our model at $256 \\times 256$ and then at $512 \\times 512$ on a large corpus of human-centric images. Exact training details can be found in ??", + "bbox": [ + 89, + 555, + 483, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Multiview Model", + "text_level": 1, + "bbox": [ + 89, + 656, + 259, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to generate many high resolution and unseen novel viewpoints of a human subject (akin to a studio capture) given a single input image.", + "bbox": [ + 89, + 678, + 482, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input Reference. We denote the input image as $\\mathbf{x}^{\\mathrm{ref}}$ and corresponding face crop as $\\mathbf{x}^{\\mathrm{face}}$ . The face crop is obtained using FaceNet [61] and resized to the same size as input image, such that: $\\mathbf{x}^{\\mathrm{face}}, \\mathbf{x}^{\\mathrm{ref}} \\in \\mathbb{R}^{H \\times W \\times C}$ .", + "bbox": [ + 89, + 726, + 482, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Target Cameras. We denote the target viewpoints to be synthesized using distinct cameras (intrinsic and extrinsic), represented as $\\mathbf{c}_{1:N}$ . Each camera is used to generate its Plücker coordinates $\\mathbf{P}_i\\in \\mathbb{R}^{H\\times W\\times 6}$ .", + "bbox": [ + 89, + 791, + 482, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Target Spatial Anchor. In addition to target cameras, we provide an oriented 3D point denoted as $\\mathbf{a}_i = [\\mathbf{R}_i|\\mathbf{t}_i]$ , which roughly defines the center of the subject's head, as", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a9c12939a1366ca61e3051df02041f892cb9bd5c4751d1a1f106a75a05c4bc3c.jpg", + "image_caption": [ + "Figure 3. DiT and ControlMLP Block. Our DiT block (left) loosely follows [55], with a AdaIn-based timestep modulation. We apply attention and MLP blocks in parallel [10], and jointly apply self-attention to the noisy generated and identity conditioning tokens. ControlMLP block (right) is used to provide lightweight spatially-aligned conditioning - Plucker and Spatial Anchor." + ], + "image_footnote": [], + "bbox": [ + 535, + 95, + 874, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "well as their gaze direction. We show an example of our Spatial Anchor in Fig. 2 in the studio on the left. This anchor is color-coded and projected into a 2D image, which is used as conditioning for our target view generations. During inference, the Spatial Anchor can be placed at any given point that lies within the field-of-view of the target cameras.", + "bbox": [ + 511, + 426, + 905, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multiview Diffusion Model. Given the above inputs, we train a multiview diffusion model $\\epsilon_{\\theta}$ to jointly denoise all the target views $\\mathbf{y}_{1:N} \\in \\mathbb{R}^{N \\times H \\times W \\times C}$ with the objective:", + "bbox": [ + 511, + 520, + 905, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D M}} = \\left\\| \\boldsymbol {\\epsilon} ^ {t} - \\epsilon_ {\\theta} \\left(\\mathbf {y} _ {1: N} ^ {t}, \\mathbf {c} _ {1: N}, \\mathbf {x} ^ {\\text {r e f}}, \\mathbf {x} ^ {\\text {f a c e}}, t\\right) \\right\\| ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 577, + 903, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{y}_{1:N}^{t}$ are noisy target images. We condition the base model on the provided reference image and its face crop by concatenating their patchified latent tokens with the noisy input latent tokens to the model. This is shown in Fig. 2", + "bbox": [ + 511, + 606, + 905, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mid-training. In mid-training stage, we want to train a strong multiview model that can denoise several images together, and absorb the dataset quickly at a lower resolution. During this phase, we do not use any pixel-aligned spatial control such as Plücker or Spatial Anchor. We use an MLP to encode the flattened 16-dimensional target camera intrinsics and extrinsics into a single token. We fuse this camera token into each noisy latent token (for the corresponding view) as positional encoding, which makes our multiview model 3D-aware of the target viewpoints. We mid-train our model at $128 \\times 128$ resolution to jointly denoise 24 views.", + "bbox": [ + 511, + 670, + 906, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Post-training. In the post-training stage, our objective is to create a high-resolution model which is 3D-consistent starting with a low-resolution and a 3D-aware (but not consistent) model. For this, we design a lightweight Control-", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16421", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Net [93]-inspired module, which takes as input the pixel-aligned Plücker and Spatial Anchor controls, and the denoising timestep to create a separate modulation signal for the multiview model. We name this module ControlMLP, as it uses a single MLP to generate scale-and-shift modulated control for each multiview-DiT block as shown in Fig. 3. Each layer of ControlMLP is zero-initialized at the start. We find that the Post-training phase is crucial in reducing flicker and 3D inconsistencies in generations. We post-train models at $512 \\times 512$ and $1024 \\times 1024$ resolutions to jointly denoise 10 and 2 views respectively. Increasing the number of views further leads to GPU out-of-memory issues.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Encoding Plücker and Spatial Anchor. We notice that the relative differences between neighboring pixels in Plücker coordinates are tiny. To amplify these differences better, we use a SIREN [67] layer to first process the 6D grid into a 32D feature grid. Then, we downsample it by $8\\mathrm{x}$ to match the size of latent tokens and feed it as input to the ControlMLP. In addition, use the Spatial Anchor to fix the position and orientation of the subject's head in 3D. We only use the Spatial Anchor for generations, and not for the input reference view. We encode the Spatial Anchor image into the latent space of our model via VAE, and concatenate it with Plücker input and pass it through an MLP to create the modulation signal at each layer.", + "bbox": [ + 89, + 275, + 482, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Understanding and Improving Spatial Control", + "text_level": 1, + "bbox": [ + 89, + 479, + 482, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section presents our design choices and examines alternative approaches for injecting pixel-aligned spatial controls during Post-training stage. We demonstrate the effectiveness of spatial control through a focused overfitting experiment with quantitative evaluations in Tab. 1.", + "bbox": [ + 89, + 502, + 482, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene Overfitting Task. We use 160 frames from a fixed 3D scene of a given subject and timestamp, split into 100 training and 60 validation views. We overfit our mid-trained model to the training views while testing various spatial control methods, training only the control modules while keeping other weights frozen. After overfitting for 10K iterations, we evaluate the model on validation views for novel view synthesis. Strong generalization to validation viewpoints indicates effective spatial control and appropriate camera viewpoint sensitivity. Through this task, we evaluate different spatial control injection methods in Tab. 1, starting with simple to advanced modulation designs.", + "bbox": [ + 89, + 580, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- No overfitting (Row 1). The Mid-trained model without scene-specific overfitting achieves comparable PSNR of 19.2 and 19.7 on train and validation views respectively. We treat this setup as baseline to improve over.", + "- Encoding Camera with MLP (Row 2). We encode camera using an MLP similar to prior works [41, 66] and our Mid-training stage (Sec. 3.2). After overfitting, the model achieves slightly better PSNR on training views as expected, however the validation PSNR drops by 1.28" + ], + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/623166a5a6e79e45848ec54594c42e21a9fb62a418aaaef0ddcdd98907d6712c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Method (M2@ 128)PSNRval ↑PSNRtrain ↑
1.Mid-trained (No Overfitting)19.2319.70
2.+ Camera (w/ MLP) [41, 66]17.95-1.2819.92+0.22
3.+ Plücker (w/ MLP) [6, 23, 31, 68]18.89-0.3420.74+1.04
4.+ ControlMLP19.45+0.2229.36+9.66
5.+ SIREN20.13+0.9030.19+10.49
6.+ Spatial Anchor (Ours)22.60+3.3730.49+10.79
", + "bbox": [ + 540, + 89, + 880, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Evaluating and Designing Spatial Controls (Sec. 3.3). We overfit our multi-view model for 10K iterations on 100 views of a single scene from our Head-only dataset, and evaluate on 60 novel views of this scene by only varying the spatial controls (camera and Spatial Anchor). We use this setup to test the modulation strength of different spatial controls. Subscript values green/red show deviations from Row 1.", + "bbox": [ + 511, + 199, + 906, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "points to 17.95. This suggests that an MLP does not provide enough modulation for camera control.", + "bbox": [ + 526, + 311, + 903, + 340 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Plücker as Positional Encoding (Row 3). In this setup, we use downsampled and patchified Plücker coordinates processed through MLP to create positional encoding, which is added to the noisy latent tokens. This setup is inspired from prior works [3, 6, 23, 31, 68], and it further improves the validation PSNR compared to MLP at 18.89, but lags behind the non-overfitted baseline.", + "- Plücker with ControlMLP and SIREN (Row 4, 5). Here, we use our ControlMLP module to inject spatial control at each multiview-DiT block output. Moreover, encoding Plücker coordinate with a SIREN [67] amplifies the relative differences between neighboring pixels (Sec. 3.2). This setup achieves PSNR of 20.13 with an improvement of 0.9 over baseline.", + "- Adding Spatial Anchor (Row 6). Finally using the Spatial Anchor gives validation PSNR of 22.6 (gain of 3.3 points over baseline) and enables strong spatial control. Thus, we adopt this configuration for Post-training stage." + ], + "bbox": [ + 513, + 340, + 903, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Enhanced 3D Consistency Metric", + "text_level": 1, + "bbox": [ + 513, + 622, + 805, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traditionally, the 3D consistency of multiview generation models is evaluated using 2D image metrics such as PSNR, LPIPS, and SSIM against a fixed set of ground truth images. However, this approach unfairly penalizes models that generate plausible and 3D-consistent novel content deviating from the fixed ground truth images. Some works try to address this by measuring SfM [17, 60] or epipolar error [45, 89], but these methods solve for pose or are not robust since they measure against the entire epipolar line. To address these limitations, we use our GT camera poses as input and compute Reprojection Error (RE) given our known camera poses and predicted correspondences. Our computation of RE involves the following steps:", + "bbox": [ + 511, + 643, + 906, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Landmarks and Correspondence estimation. We use SuperPoint [13] to detect landmarks in the generated images and employ SuperGlue [59] to establish pairwise correspondences between landmarks across images.", + "bbox": [ + 513, + 840, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16422", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. Triangulation. Given the correspondences and camera parameters, we apply Triangulation based on Direct Linear Transformation (DLT) [22] to obtain the corresponding 3D points for each landmark.", + "3. Reprojection and Error calculation. We reproject these 3D points onto each image and compute the RE as the L2 distance between the original landmark and the reprojected 3D point normalized by image resolution, and average error across all images." + ], + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This approach evaluates multiview generation models by focusing on their ability to produce 3D-consistent results rather than adhering to a fixed ground truth. The Reprojection Error provides a valuable basis for comparison across different methods. Furthermore, by computing RE on a set of real-world images that are independent of our generated images, we can establish a baseline that quantifies the error due to the noisy predictions from SuperGlue and SuperPoint rather than the quality of our generated images.", + "bbox": [ + 89, + 229, + 482, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Naming Convention (RE@SG). Please note that the use of SuperPoint [13] and SuperGlue [59] estimation modules is a particular instantiation of our metric, and these could be replaced in the future with stronger counterparts such as MAST3R [36, 75] or domain-specific keypoint detectors such as Sapiens [33]. Thus we use the naming convention of RE@SG to denote the Reprojection Error (RE) under SuperGlue (SG) estimation, which could be modified accordingly in the future for different estimators.", + "bbox": [ + 89, + 367, + 482, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 523, + 222, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present details of the datasets used in all training and validation stages, followed by discussion of evaluation metrics - with particular emphasis on the 3D consistency metric and conclude with core experimental results and ablations.", + "bbox": [ + 89, + 551, + 482, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.Data", + "text_level": 1, + "bbox": [ + 89, + 628, + 163, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Humans-3B Dataset. We utilize a large proprietary dataset of approximately 3 billion human-centric in-the-wild images for pretraining. We provide more details on data filtering and curation in ??", + "bbox": [ + 89, + 652, + 482, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Head and Full-body Studio Datasets. We rely on high-quality proprietary studio captures as our primary source of data for learning 3D consistency. Our model comes in two variants: head-only and full-body, each trained (for Mid-training and Post-training stages) on corresponding datasets. For our full-body model, we utilize a dataset of 861 subjects (811 train, 50 test), nearly 1000 frames per subject. For our head-only model, we use a dataset of 1450 subjects (1400 train, 50 test), nearly 40000 frames per subject. The studio setup is similar to [43], with two capture domes for capturing full-body and head-only high-resolution 4K images with 230 and 160 cameras.", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "iPhone Dataset. To evaluate real-world performance, we collect casual images of 50 test subjects in an indoor office environment using an iPhone 13 Pro. We preprocess these images with Sapiens-2B [33] for background segmentation before model inference. This dataset serves exclusively for evaluating our model's performance on in-the-wild inputs.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Setup and Metrics", + "text_level": 1, + "bbox": [ + 511, + 194, + 779, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Consistency. Following prior works, we report standard metrics like PSNR, SSIM ( $\\times 100$ ), and LPIPS ( $\\times 100$ ) metrics. However, these metrics unfairly penalize plausible novel views generated under incomplete inputs. We therefore introduce the Reprojection Error metric, described in Sec. 3.4, which validates 3D consistency without directly relying on ground truth. Our evaluation generates 4 randomly selected views from the test split.", + "bbox": [ + 511, + 215, + 903, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Identity Preservation. We use two metrics that measure identity preservation across generated views by computing cosine distance between features extracted via FaceNet [4, 61] for face similarity; and CLIP [52] vision encoder for full-body similarity.", + "bbox": [ + 511, + 340, + 903, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretrained Model Evaluation. We measure the effectiveness of our pre-training strategy by reporting the FID [24]. We use a smaller annotated 30M subset of Humans-3B for training image and text-conditioned P1@128 models, a 30M subset of unfiltered dataset for training No Filtering P1@128, and a 1000 sample test set from iPhone dataset.", + "bbox": [ + 511, + 421, + 903, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results", + "text_level": 1, + "bbox": [ + 511, + 523, + 606, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretraining and Data Filtering. Tab. 2 presents our pretrained model results (Row 1) and ablations on Human-centric data filtering and Image-conditioned pretraining (Rows 2-5). Human-centric filtering and image-based conditioning are both critical to achieve high-quality generation. The qualitative results are shown in ??", + "bbox": [ + 511, + 547, + 903, + 638 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/04edfcfda9dd5d56935de8d60fe8fd6ae881072f0bda819771f15c84a5583f58.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Method (Stage @ Resolution)FID ↓
1.Pippo (P1@512)51.164
2.Human-centric Filtering (P1@128)75.639
3.No Filtering (P1@128)86.838
4.Image-conditioned (P1@128)75.639
5.Text-conditioned (P1@128)109.720
", + "bbox": [ + 578, + 652, + 841, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Pretraining and Data Filtering. We report results of the full pretrained model P1@512, and compare several variants of P1@128 models. We report FID on iPhone dataset (1k samples).", + "bbox": [ + 511, + 758, + 906, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "High-resolution Multi-view Generation. In Tab. 3, we evaluate 3D reconstruction and identity preservation for unseen subjects from the studio datasets. We show that increasing the output resolution of generation in our approach does not hurt 3D consistency or similarity. We put corresponding visuals in ??, Rows 2 and 3.", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16423", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c9adac0176b9e1c4121c417c4796215c96fc0a30364aa36d0d1f106b3bb157ac.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Results on Unseen Studio and iPhone data (Sec. 4.3.) We report results on Post-trained Pippo models at three different resolutions. We report 3D metrics PSNR, SSIM, and LPIPS; as well as our proposed Reprojection Error (RE@SG) under SuperGlue estimation which does not require ground truth (Sec. 3.4). We report 2D Face and Body similarities. Red subscript show deviation against ground truth value." + ], + "table_body": "
# Split & Resolution3D ConsistencySimilarity
RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
Head-only
1. Studio (128)3.3+0.821.067.614.862.0-13.361.5-16.2
2. Studio (1K)3.4+0.520.372.026.273.5-2.579.4-0.1
3. iPhone Face (1K)3.0---67.6-
Full-body
4. Studio (128)3.6+0.122.884.010.041.7-8.567.1-9.4
5. Studio (1K)1.5+0.122.491.711.164.7-6.074.1-2.2
6. iPhone (1K)1.7---58.068.1
", + "bbox": [ + 98, + 88, + 477, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generations from Casual iPhone Photos. We present in Tab. 3 (Rows 3,6) Reprojection Error and similarity scores for casually taken images from the iPhone dataset with 1K resolution model. In this scenario, the standard reconstruction error metrics cannot be assessed due to missing ground truth. We find that the reprojection error on iPhone captures remains comparable to the studio dataset - demonstrating 3D consistency. This illustrates the generalizability of Pippo beyond the multi-view training data domain, where our pretraining with large-scale in-the-wild human data is critical. We put corresponding visuals in ??, Row 1 (Appendix).", + "bbox": [ + 89, + 333, + 482, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons with External Benchmarks. In Fig. 4, we compare Pippo with individual state-of-the-art baselines in Full-body and Head-only generation. SiTH [25] reconstructs textured human-mesh using ControlNet paired with a SDF representation. Compared to SiTH, Pippo facilitates high-resolution and accurate multiview synthesis. DiffPortrait3D [20] inverts a 3D-GAN based on a given input image. Compared to it, our model supports greater viewpoint variability and ensures closer adherence to the input image.", + "bbox": [ + 89, + 517, + 482, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparisons and baselines. Existing SoTA Human methods [20, 25] use explicit SMPL priors, and thus are difficult to compare with directly. Qualitatively, we found that they cannot handle novel views or preserve details (Fig. 4), and hence do not quantitatively compare against them. In Pippo, we focus on creating a strong multiview human generator, and we benchmark four state-of-the-art multi-view diffusion models on the iPhone full-body dataset in Tab. 4. We find that Pippo preserves identity (i.e., face and body similarity) and 3D consistency (RE) better while also operating at a higher resolution compared to baselines.", + "bbox": [ + 89, + 656, + 482, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmarking Pippo on public datasets. In Tab. 5, we benchmark Pippo on the public Codec Avatar datasets [43] to aid future comparisons. Specifically, we use Ava-256 containing 256 head-only captures and Goliath containing 4", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f7cfe4dc9030a24476f896b1d7e6352896f9fd0fd1091ac86ed64200fafa42cc.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Quantitative comparison with SoTA multi-view models. We quantitatively compare Pippo against state-of-the-art multi-view diffusion models. We find that Pippo preserves identity (i.e., face and body similarity) and 3D consistency (RE) better while operating at a higher resolution compared to baselines." + ], + "table_body": "
#Method & ResolutionRE@SG ↓Face ↑Body ↑
1.MV-Adapter [27] (768)4.743.064.1
2.Era3D [38] (512)4.138.164.4
3.Wonder3D [42] (256)5.334.758.8
4.Pippo (P3@1K)3.058.068.1
", + "bbox": [ + 568, + 88, + 851, + 160 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "full-body. During evaluation, we only use subsets of these datasets that were not used for training. We find that Pippo's performance on these datasets is inline with its performance on internal studio datasets.", + "bbox": [ + 511, + 267, + 903, + 327 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/138833e6f8a20e686357191950441214fb3945269252631b82f4182a58b4c394.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Benchmarking Pippo on public Ava-256 and Goliath datasets. We find that Pippo's performance on these datasets is inline with its performance on internal studio datasets. This will aid future comparisons against Pippo." + ], + "table_body": "
#Dataset (P3@1K)RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
1.Ava-256 (Head-only)3.820.168.726.672.976.8
2.Goliath (Full-body)1.220.489.715.587.577.7
", + "bbox": [ + 514, + 338, + 903, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablations", + "text_level": 1, + "bbox": [ + 511, + 482, + 624, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We examine design choices at each training stage (Sec. 3) and present our ablation results in Tab. 6. Experiments are conducted at $128 \\times 128$ resolution on the Head-only dataset.", + "bbox": [ + 511, + 503, + 903, + 550 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/08467f4f8305391de885ca1743be65be613156a1e9606bf39b6b5f8b9dc13fdc.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Ablation on design choices and training stages. We evaluate several multi-view models at $128 \\times 128$ and $512 \\times 512$ resolution at different training stages on Head-only dataset. We ablate the choice of doing Mid-training and Pre-training, along with different spatial controls and reference conditioning methods..The training stages are labeled as M2 (Mid-training) and P3 (Post-training), followed by @ specified resolution." + ], + "table_body": "
#Method (Stage @ Res.)3D ConsistencySimilarity
RE@SG↓PSNR↑SSIM↑LPIPS↓Face ↑Body ↑
1.Pippo (P3@512)2.7+0.721.771.721.574.1-2.077.4-2.4
2.Pippo Face-only (P3@512)2.4+0.420.570.325.374.3-1.775.8-4.0
3.No Mid-train (P3@512)5.6+3.614.559.544.220.4-55.663.4-16.4
4.Pippo (P3@128)3.3+0.821.067.614.862.0-13.361.5-16.2
5.No Anchor (P3@128)11.5+9.017.154.021.163.1-12.268.3-9.3
6.No Plücker (P3@128)4.2+1.720.264.516.163.5-11.866.1-11.5
7.Pippo (M2@128)3.4+0.919.161.417.260.0-15.362.6-15.0
8.No Pre-train (M2@128)5.9+3.413.139.344.919.5-55.849.0-28.6
9.Cross-Attn (M2@128)3.5+1.018.059.222.166.2-9.170.7-7.0
10.Non-frontal (M2@128)5.8+3.315.252.430.149.2-26.160.7-16.9
", + "bbox": [ + 516, + 561, + 903, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Significance of Pre-training and Mid-training. Pretraining our model on the Humans-3B dataset enables robust generalization to novel identities, as demonstrated in", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16424", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a1556fb832c03b8b9171a80936dc21cd43442942792f4423be8d34086dead0f.jpg", + "image_caption": [ + "Figure 4. Visual comparison with state-of-the-art methods. We visually compare Pippo with state-of-the-art baselines DiffPortrait3D [20] (head-only) and SiTH [25] (full-body) generation." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 906, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 6, Row 8. Without pre-training, model generalization deteriorates, resulting in unclear facial features. Skipping mid-training at lower resolution impairs consistent multiview generation, as shown in Row 2.", + "bbox": [ + 88, + 347, + 480, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of Frontal Input Reference. Our ablation in Tab. 6, Row 10 demonstrates that completely randomizing the view point of an input reference image leads to overfitting to training identities. Non-frontal views, especially rear views, have very limited information about the identity which forces the network to pick up spurious correlations.", + "bbox": [ + 89, + 409, + 482, + 500 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8f1698f351e4d427ed5447ad16096ad58ca475bbdf358843b809bce5e61a597b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# Pretrain Data (M2@128)RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
1. 30M (1% HQ)3.821.381.813.330.666.2
2. 900M (30% Random)4.021.682.012.130.667.2
3. 2.1B (70% Random)3.721.682.112.137.667.5
4. Humans-3B (100%)3.721.982.312.741.767.2
", + "bbox": [ + 91, + 515, + 480, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablating the size of pre-training dataset on Humans-3B. We conduct Full-body mid-training with pretrained checkpoints trained on $1\\%$ , $30\\%$ , $70\\%$ , and $100\\%$ subsets of Humans-3B dataset. We found large-scale data is crucial for generalization to novel identities (i.e., face similarity).", + "bbox": [ + 89, + 585, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of Self-attention. Replacing self-attention with cross-attention for reference image encoding, using the same routing as image-conditioned pretraining from Sec. 3.1, leads to degraded performance as shown in Tab. 6, Row 9. We observe that this setup causes the model to ignore input conditioning, generating images that only vaguely resemble training subjects.", + "bbox": [ + 89, + 672, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Role of large scale Humans-3B pre-training dataset. We utilized intermediate checkpoints from pre-training stage (P1) trained on $30\\%$ and $70\\%$ of the data, and a separate checkpoint trained on $1\\%$ high-quality subset of Humans-3B. Starting from these checkpoints, we mid-trained Pippo to denoise 4 views at $128 \\times 128$ resolution on Full-body dataset for two days. We report their respective results in Tab. 7. We found large-scale data is crucial for general", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ization to novel identities – indicated by high gains in face similarity metric.", + "bbox": [ + 511, + 348, + 906, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 392, + 633, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Pippo, a diffusion transformer model that generates a dense set of high-resolution multi-view consistent images of a person from a single image. Our experiments show that our multi-stage training strategy, which combines largescale in-the-wild data with high-quality multi-view studio captures, enables generalizable high-resolution multi-view synthesis. Analysis of the diffusion transformer architecture reveals that self-attention with the reference image, Plücker coordinates with SIREN, and Spatial Anchor are all essential for high-fidelity multi-view human generation. Pippo achieves, for the first time, consistent multi-view human image generation at 1K resolution. We also show that our proposed 3D consistency metric enables evaluation of 3D consistency without paired ground-truth data. One of the limitations of our approach is the limited number of simultaneously generated views, originating from large context length and memory constraints, which can be tackled with parallelization techniques and autoregressive generation. Extending our approach to multi-view consistent video generation will be addressed in future work.", + "bbox": [ + 511, + 417, + 906, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 513, + 734, + 609, + 750 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panohed: Geometry-aware 3d fullhead synthesis in 360deg. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20950-20959, 2023. 3", + "[2] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabián Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM Transactions on Graphics (TOG), 40(4):1-17, 2021. 2" + ], + "bbox": [ + 521, + 760, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16425", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[3] Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, et al. Vd3d: Taming large video diffusion transformers for 3d camera control. arXiv preprint arXiv:2407.12781, 2024.5", + "[4] Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2018), pages 67-74, 2018. 6", + "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 3", + "[6] Eric Ming Chen, Sidhanth Holalkere, Ruyu Yan, Kai Zhang, and Abe Davis. Ray conditioning: Trading photoconsistency for photo-realism in multi-view image generation. arXiv preprint arXiv:2304.13681, 2023. 5", + "[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 3", + "[8] Wei Cheng, Ruixiang Chen, Wanqi Yin, Siming Fan, Keyu Chen, Honglin He, Huiwen Luo, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, Daxuan Ren, Lei Yang, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, Bo Dai, and Kwan-Yee Lin. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. arXiv preprint, arXiv:2307.10173, 2023. 2", + "[9] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 4", + "[10] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme Ruiz, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd Van Steenkiste, Gamaleldin Fathy Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmiŋn Bastings, Mark Collier, Alexey A. Gritsenko, Vighnesh Birodkar, Cristina Nader Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lucic, Xiaohua Zhai, Daniel Keysers, Jeremiah J. Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters. In Proceedings of the 40th International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 4", + "[11] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 3", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2", + "[13] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection" + ], + "bbox": [ + 93, + 92, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 5, 6", + "[14] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In ACM MM, 2022. 3", + "[15] Nikita Drobyshev, Antoni Bigata Casademunt, Konstantinos Vougioukas, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Emoportraits: Emotion-enhanced multimodal one-shot head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8498-8507, 2024. 3", + "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 4", + "[17] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2, 5", + "[18] Matheus Gadelha, Rui Wang, and Subhransu Maji. Multiresolution tree networks for 3d point cloud processing. In ECCV, 2018. 3", + "[19] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 3", + "[20] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10456-10465, 2024. 3, 7, 8", + "[21] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 3", + "[22] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 6", + "[23] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3, 5", + "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. 2017. 6", + "[25] I Ho, Jie Song, Otmar Hilliges, et al. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 538-549, 2024. 7, 8" + ], + "bbox": [ + 516, + 93, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16426", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 4", + "[27] Zehuan Huang, Yuanchen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 7", + "[28] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9784-9794, 2024. 3", + "[29] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2", + "[30] Yash Kant, Aliaksandr Siarohin, Michael Vasilkovsky, Riza Alp Guler, Jian Ren, Sergey Tulyakov, and Igor Gilitschenski. invs: Repurposing diffusion inpainters for novel view synthesis. In SIGGRAPH Asia 2023 Conference Papers, 2023. 3", + "[31] Yash Kant, Aliaksandr Siarohin, Ziyi Wu, Michael Vasilkovsky, Guocheng Qian, Jian Ren, Riza Alp Guler, Bernard Ghanem, Sergey Tulyakov, and Igor Gilitschenski. Spad: Spatially aware multi-view diffusers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10026-10038, 2024. 5", + "[32] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 3", + "[33] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 6", + "[34] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 3", + "[35] Black Forest Labs. Flux model, 2024. Accessed: 2024-11-10. 4", + "[36] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 6", + "[37] Heyuan Li, Ce Chen, Tianhao Shi, Yuda Qiu, Sizhe An, Guanying Chen, and Xiaoguang Han. Spherehead: Stable 3d full-head synthesis with spherical tri-plane representation. In European Conference on Computer Vision, pages 324-341. Springer, 2025. 3", + "[38] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, Wenping Wang, Qifeng Liu, and Yike Guo. Era3d: High-resolution multiview diffusion using efficient row-wise attention, 2024. 7", + "[39] Chieh Hubert Lin, Chia-Che Chang, Yu-Sheng Chen, Da-Cheng Juan, Wei Wei, and Hwann-Tzong Chen. Coco-gan: Generation by parts via conditional coordinating. In ICCV, 2019. 2" + ], + "bbox": [ + 89, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3", + "[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 2, 3, 5", + "[42] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion, 2023. 7", + "[43] Julieta Martinez, Emily Kim, Javier Romero, Timur Bagautdinov, Shunsuke Saito, Shoou-I Yu, Stuart Anderson, Michael Zollhöfer, Te-Li Wang, Shaojie Bai, Chenghui Li, Shih-En Wei, Rohan Joshi, Wyatt Borsos, Tomas Simon, Jason Saragih, Paul Theodosis, Alexander Greene, Anjani Josyula, Silvio Mano Maeta, Andrew I. Jewett, Simon Venshtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mohamed Ezzeldin A. Elshaer, Tingfang Du, Longhua Wu, Shen-Chi Chen, Kai Kang, Michael Wu, Youssef Emad, Steven Longay, Ashley Brewer, Hitesh Shah, James Booth, Taylor Koska, Kayla Haidle, Matt Andromalos, Joanna Hsu, Thomas Dauer, Peter Selednik, Tim Godisart, Scott Ardisson, Matthew Cipperly, Ben Humberston, Lon Farr, Bob Hansen, Peihong Guo, Dave Braun, Steven Krenn, He Wen, Lucas Evans, Natalia Fadeeva, Matthew Stewart, Gabriel Schwartz, Divam Gupta, Gyeongsik Moon, Kaiwen Guo, Yuan Dong, Yichen Xu, Takaaki Shiratori, Fabian Prada, Bernardo R. Pires, Bo Peng, Julia Buffalini, Autumn Trimble, Kevyn McPhail, Melissa Schoeller, and Yaser Sheikh. Codec Avatar Studio: Paired Human Captures for Complete, Driveable, and Generalizable Avatars. NeurIPS Track on Datasets and Benchmarks, 2024. 2, 6, 7", + "[44] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 3", + "[45] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024. 2, 5", + "[46] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4", + "[47] Dongwei Pan, Long Zhuo, Jingtan Piao, Huiwen Luo, Wei Cheng, Yuxin Wang, Siming Fan, Shengqi Liu, Lei Yang, Bo Dai, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, and Kwan-Yee Lin. RenderMe-360: Large digital asset library and benchmark towards high-fidelity head avatars. In Thirty-seventh Conference on Neural Information" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16427", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Processing Systems Datasets and Benchmarks Track, 2023. 2", + "[48] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3", + "[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4", + "[50] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3", + "[51] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 3", + "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 6", + "[53] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 3", + "[54] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4", + "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4", + "[56] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In CVPR, 2019. 3", + "[57] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 3", + "[58] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 3", + "[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 5, 6", + "[60] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 5" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 4, 6", + "[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. 2022. 2", + "[63] Akash Sengupta, Thiemo Alldieck, Nikos Kolotouros, Enric Corona, Andrei Zanfir, and Cristian Sminchisescu. Diffhuman: Probabilistic photorealistic 3d reconstruction of humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1439-1449, 2024. 3", + "[64] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 3", + "[65] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 3", + "[66] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 5", + "[67] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. 2020. 5", + "[68] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. 2021. 5", + "[69] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. 2015. 4", + "[70] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In European Conference on Computer Vision, pages 107-124. Springer, 2022. 3", + "[71] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14795-14805, 2023. 3", + "[72] Qingyang Tan, Lin Gao, Yu-Kun Lai, and Shihong Xia. Variational autoencoders for deforming 3d mesh models. In CVPR, 2018. 3", + "[73] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. 2023. 3", + "[74] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 2023. 3", + "[75] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 6" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16428", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 3", + "[77] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[78] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. 2023. 3", + "[79] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3", + "[80] Jing Wen, Xiaoming Zhao, Zhongzheng Ren, Alexander G Schwing, and Shenlong Wang. Gomavatar: Efficient animatable human modeling from monocular video using gaussians-on-mesh. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2059-2069, 2024. 3", + "[81] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 3", + "[82] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. 2016. 3", + "[83] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Song-Chun Zhu, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In CVPR, 2018. 3", + "[84] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, et al. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19801-19811, 2024. 2", + "[85] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 3", + "[86] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024.3", + "[87] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. FaceScape: A largescale high quality 3D face dataset and detailed riggable 3D" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "face prediction. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[88] T Yenamandra, A Tewari, F Bernard, HP Seidel, M Elgharib, D Cremers, and C Theobalt. i3DMM: Deep implicit 3D morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[89] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2, 5", + "[90] Zhixuan Yu, Jae Shin Yoon, In Kyu Lee, Prashanth Venkatesh, Jaesik Park, Jihun Yu, and Hyun Soo Park. HUMBI: A large multiview dataset of human body expressions. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[91] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 4", + "[92] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22096-22105, 2023. 3", + "[93] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3, 5", + "[94] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + ], + "bbox": [ + 516, + 92, + 906, + 556 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "16429", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/2679a0a3-d3fd-49c2-aeba-9db847b48300_model.json b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/2679a0a3-d3fd-49c2-aeba-9db847b48300_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4b1ca3f7999fb1cbf7cec2a2d9447820c409c1b1 --- /dev/null +++ b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/2679a0a3-d3fd-49c2-aeba-9db847b48300_model.json @@ -0,0 +1,2677 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.83, + 0.154 + ], + "angle": 0, + "content": "Pippo: High-Resolution Multi-View Humans from a Single Image" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.18, + 0.882, + 0.278 + ], + "angle": 0, + "content": "Yash Kant\\(^{1,2,3}\\), Ethan Weber\\(^{1,4}\\), Jin Kyu Kim\\(^{1}\\), Rawal Khirodkar\\(^{1}\\), Su Zhaoen\\(^{1}\\), Julieta Martinez\\(^{1}\\), Igor Gilitschenski\\(^{*,2,3}\\), Shunsuke Saito\\(^{*,1}\\), Timur Bagautdinov\\(^{*,1}\\) \n\\(^{1}\\)Meta Reality Labs, \\(^{2}\\)University of Toronto, \\(^{3}\\)Vector Institute, \\(^{4}\\)UC Berkeley https://yashkant.github.io/pippo/ \n*Joint Advising" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.287, + 0.905, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.728, + 0.908, + 0.77 + ], + "angle": 0, + "content": "Figure 1. Pippo produces high-resolution, multi-view, studio-quality images from a single image. First row shows full-body generation of an unseen subject captured from an iPhone picture. Second and third rows show face-only and partial full-body input images from unseen studio subject(s) to novel views respectively. All visuals are at 1K resolution." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.795, + 0.327, + 0.811 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.841, + 0.482, + 0.903 + ], + "angle": 0, + "content": "We present Pippo, a generative model capable of producing 1K resolution dense turnaround videos of a person from a single casually clicked photo. Pippo is a multi-view diffusion transformer and does not require any additional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.797, + 0.907, + 0.888 + ], + "angle": 0, + "content": "inputs — e.g., a fitted parametric model or camera parameters of the input image. We pre-train Pippo on 3B human images without captions, and conduct multi-view mid-training and post-training on studio captured humans. During mid-training, to quickly absorb the studio dataset, we denoise several (upto 48) views at low-resolution, and encode tar" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16418" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.244 + ], + "angle": 0, + "content": "get cameras coarsely using a shallow MLP. During posttraining, we denoise fewer views at high-resolution and use pixel-aligned controls (e.g., Spatial anchor and Plucker rays) to enable 3D consistent generations. At inference, we propose an attention biasing technique that allows Pippo to simultaneously generate greater than \\(5 \\times\\) as many views as seen during training. Finally, we also introduce an improved metric to evaluate 3D consistency of multi-view generations, and show that Pippo outperforms existing works on multi-view human generation from a single image." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.279, + 0.223, + 0.294 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.483, + 0.473 + ], + "angle": 0, + "content": "Creating photorealistic human representations with the ability to control viewpoints has numerous applications in entertainment, healthcare, fashion and social media. Building such representations, first and foremost, requires high-quality multi-view studio data [2, 29], which is costly to acquire. This significantly limits the scalability in terms of the number identities for high-quality studio data [29, 43, 84]. In contrast, large-scale, unstructured, and diverse human images and videos are available online. However, the raw data of such in-the-wild images does not offer ground-truth 3D or multi-view representations of humans." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.703 + ], + "angle": 0, + "content": "In this work, we present a novel approach leveraging the best of two worlds: generalizability from in-the-wild unstructured images, and fidelity and view-controllability from studio capture data. Specifically, our model Pippo is a diffusion transformer, which can generate several 1K resolution multi-view consistent images jointly during inference. Pippo takes as input a single image of an individual, camera poses of the target viewpoints to be generated. Since the scale and placement of the subject is ambiguous from a single image, Pippo uses a Spatial Anchor which roughly specifies the location and orientation of the subject in 3D space. Our model does not rely on additional conditioning such as body priors or camera parameters of the input images, to scale our training pipeline to in-the-wild data and support unconstrained inputs at test time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We employ a multi-stage training recipe to train Pippo. First, we pre-train the model for latent-to-image generation task, similar to [54], on a large human-centric dataset of in-the-wild images. Next, in Mid-training stage, we jointly generate multiple consistent images of the subject, conditioned on target viewpoints and a single input image using high-quality studio dataset. Finally, in Post-training stage, the model is provided with a minimal placement signal - spatial anchor - encoding rough head orientation - which further improves 3D consistency. Our architecture is also carefully tailored to the conditional multi-view generation - we propose several simple but effective modifications to the basic DiT architecture, including self-attention-based" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "conditioning, lightweight spatial controls and camera conditioning with Plucker coordinates." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.905, + 0.426 + ], + "angle": 0, + "content": "For evaluating such a multi-view diffusion model, 3D consistency is a critical metric for understanding the geometric correctness of the generation. For generation task in an ill-posed setup (e.g., a single image input), there may be multiple possibilities that are equally plausible. Existing multi-view generation methods typically report reconstruction metrics (i.e., PSNR, SSIM and LPIPS) [19] or FID [41], which either (1) penalize any new content that is actually 3D consistent or (2) are unable to measure the 3D consistency of the generation. To address this problem, we design a metric measuring 3D consistency: we compute 2D keypoint matches with an off-the-shelf method [59], triangulate them, and then reproject back into the other views to measure the reprojection error in pixels. Some generation methods measure reprojection errors [17] from SfM [60] or epipolar distance [45, 89], but our metric uses camera pose as input and is more precise than measuring distance to a epipolar line. We show our metric helps to quantify our results, finding that our method is favorable compared with existing approaches and other baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.427, + 0.864, + 0.442 + ], + "angle": 0, + "content": "To summarize, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.444, + 0.905, + 0.489 + ], + "angle": 0, + "content": "- a generative model capable of generating high-resolution and multi-view consistent humans from a single image and its effective training strategy." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.489, + 0.905, + 0.518 + ], + "angle": 0, + "content": "- a diffusion transformer architecture designed to enhance multi-view generation and viewpoint control." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.519, + 0.905, + 0.549 + ], + "angle": 0, + "content": "- a novel 3D consistency metric tailored to accurately measure the level of 3D consistency in generative tasks." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.444, + 0.905, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.567, + 0.655, + 0.582 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.7 + ], + "angle": 0, + "content": "We review multi-view human datasets and generative models for human synthesis, categorizing methods by their data requirements and prior constraints—such as parametric human models or explicit 3D structures. In this work, we minimize reliance on complex priors by training the model on large amounts of human-centric data and impose minimal camera and spatial controls." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Multi-view Human Datasets. While many large 2D human datasets exist [12, 39, 62], 3d captures of people (i.e., captures using dozens of views at a time) are still relatively rare, since they are expensive to collect and are thus mostly obtained in specialized research labs. Nevertheless, this kind of data has recently become more common, going from hundreds to thousands of publicly-available captures. For face-only captures, a few datasets [43, 47, 87, 88, 90] provide up to 600 subjects captured from over 60 views each, while for full-body captures, 4D-Dress [77] provides 32 subjects with 2 outfits each, and detailed annotations for their garments, while DNA-rendering [8] provides 500 subjects from 60 views. MVHumanNet [84] stands apart by" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16419" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.885, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.293, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Pipeline overview. This is an illustration of how we train our model. (Left) we use data from a studio capture and train our multi-view diffusion model (right). We condition on a full reference photo and a cropped face, as well as the target view cameras and 2D projected spatial anchor indicating head position and orientation. Our diffusion model also takes in noisy target views and a timestep in order to predict the denoised views (top). In practice, we apply a segmentation mask around the person." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.376, + 0.483, + 0.468 + ], + "angle": 0, + "content": "providing 4,500 people captured from 48 views in and over 9,000 sets of clothing, outpacing other datasets by an order of magnitude in terms of people diversity. We use internal multi-view data of \\(\\sim 1000\\) identities with dense (\\(\\sim 160\\)) camera setup, but we expect our model to produce reasonable results with publicly-available captures." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.474, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Generating Novel Poses and Expressions. Although different from our task, this related work builds 3D animatable models of people and faces. These methods use neural radiance fields [70, 71, 81], or 3d Gaussians [34, 51, 80, 94], using canonical \"T\"-poses and a corresponding forward and backward mapping to model local appearance changes. Many of these methods, however, rely on accurate human shape and pose estimation, and only achieve high photorealism for personalized models from studio captures. Relatedly, a family of methods focus on generating animatable faces from a single image by disentangling viewpoints and expressions from large collections of 2d portraits [14, 15, 32, 48, 86, 92], allowing photorealistic real-time reenactment, while being limited to faces and small viewpoint variations. Unlike these methods, we complete missing parts of a person given either one or a few partial views. We do not repose the person or animate the faces but can instead recover entirely missing views, which is orthogonal to these methods. Moreover, our approach addresses both full-body and facial reconstruction, rather than specializing in either domain." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.904 + ], + "angle": 0, + "content": "Generations with Explicit 3D Structures. Similar to early methods that extracted 3d representations from 2d models via score distillation [7, 11, 40, 50, 53, 74, 78], 3d representations of human appearance have been extracted from GANs by co-training a neural renderer from learned TriPlane [5] or HexPlane [1, 37] latent spaces. Others train generative models that operate directly in 3D space" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.497 + ], + "angle": 0, + "content": "with diffusion or GAN objectives [18, 44, 72, 76, 82, 83] or regression models to directly predict 3D representations [56, 57, 63]. While these methods produce 3d consistent faces by design, their quality is limited by the relatively small amounts of 3d training data and/or the limitations of their respective 3d representations. In contrast, we focus on generating 3d-consistent 2d images and thus avoid the downsides of explicit 3d modelling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.909, + 0.712 + ], + "angle": 0, + "content": "Generations with 2D Models. Another option is to model viewpoint changes in 2d, without an underlying 3d representation [41], sometimes modelling either photometric [30] or epipolar [28] constraints explicitly. Diff-Portrait3D [20] is a portrait-generation method that falls roughly within this category, by fine-tuning a 2d diffusion model for 3d-aware face generation using ControlNet-style [93] conditioning for viewpoint control, as well as cross-view attention [21] and initialization from a 3d-consistent GAN [5]. Other methods are more general and not specific to humans, focusing on single-view to 3D generation [19, 41, 58, 64, 65, 73, 85]. In addition to these single-view methods, video models have been fine-tuned for camera control [23, 79]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.731, + 0.605, + 0.745 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.758, + 0.865, + 0.774 + ], + "angle": 0, + "content": "We train our models following a three-stage strategy:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.78, + 0.905, + 0.81 + ], + "angle": 0, + "content": "- Image-only Pre-training (P1). We pretrain on a large-scale human-centric dataset with image conditioning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.856 + ], + "angle": 0, + "content": "- Multiview Mid-training (M2). We train models at a low-resolution of \\(128 \\times 128\\) to denoise 48 target views with coarse camera control (no pixel-aligned spatial control)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Multiview Post-training (P3). We train at a high-resolution of \\(1024 \\times 1024\\) to denoise \\(1 - 3\\) target views with spatial control injected via ControlMLP layers." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16420" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "We denote the training stage and resolution of any given model as \\(\\{\\mathrm{stage}\\} @\\{\\mathrm{resolution}\\}\\). For instance, M2@128 represents a mid-trained model at 128 resolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.147, + 0.218, + 0.16 + ], + "angle": 0, + "content": "3.1. Base Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.319 + ], + "angle": 0, + "content": "Architecture. We adopt a DiT-like [49] architecture with scale, shift, and gate modulation for timestep conditioning inspired by Stable Diffusion 3 [16] and Flux [35]. We simplify the architecture by employing MLP and attention in parallel [91], and removing second LayerNorm after attention layers. We use VAE for training in latent space with 8x spatial compression, and patchify latent images via a linear layer and patch size of 2. We use fixed sinusoidal positional encoding during training. We provide more details in ??, and the exact design of our DiT block in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.483, + 0.47 + ], + "angle": 0, + "content": "Image-only Pre-training. During pre-training, the model learns to denoise an image conditioned on its corresponding image embedding from DINOv2 [9, 46], this is similar in principle to the image decoder of DALL-E 2 [54]. We project both embeddings to the model dimension using with a linear layer to create a joint conditioning. Importantly, our pretraining setup does not require any annotations or captions for the images, and is well-aligned with our downstream objective of generating consistent multiview images given a single reference image as input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.517 + ], + "angle": 0, + "content": "Formally, given an image \\(\\mathbf{y} \\in \\mathbb{R}^{H \\times W \\times C}\\), and the joint conditioning as \\(\\mathbf{e}^{\\mathrm{img}} \\in \\mathbb{R}^{N \\times D}\\). We pre-train our diffusion model \\(\\epsilon_{\\theta}\\) with the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.527, + 0.483, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D M}} = \\left| \\left| \\boldsymbol {\\epsilon} ^ {t} - \\epsilon_ {\\theta} \\left(\\mathbf {y} ^ {t}, \\mathbf {e} ^ {\\mathrm {i m g}}, t\\right) \\right| \\right| ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.484, + 0.648 + ], + "angle": 0, + "content": "Where \\(t \\in [0, T]\\) is diffusion timestep, \\(\\epsilon^t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) is the noise added at the given timestep. We use the DDPM [26, 69] formulation to define discrete timesteps and set \\(T = 1000\\). We first pretrain our model at \\(256 \\times 256\\) and then at \\(512 \\times 512\\) on a large corpus of human-centric images. Exact training details can be found in ??" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.657, + 0.26, + 0.672 + ], + "angle": 0, + "content": "3.2. Multiview Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.679, + 0.483, + 0.725 + ], + "angle": 0, + "content": "Our goal is to generate many high resolution and unseen novel viewpoints of a human subject (akin to a studio capture) given a single input image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.483, + 0.789 + ], + "angle": 0, + "content": "Input Reference. We denote the input image as \\(\\mathbf{x}^{\\mathrm{ref}}\\) and corresponding face crop as \\(\\mathbf{x}^{\\mathrm{face}}\\). The face crop is obtained using FaceNet [61] and resized to the same size as input image, such that: \\(\\mathbf{x}^{\\mathrm{face}}, \\mathbf{x}^{\\mathrm{ref}} \\in \\mathbb{R}^{H \\times W \\times C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.792, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Target Cameras. We denote the target viewpoints to be synthesized using distinct cameras (intrinsic and extrinsic), represented as \\(\\mathbf{c}_{1:N}\\). Each camera is used to generate its Plücker coordinates \\(\\mathbf{P}_i\\in \\mathbb{R}^{H\\times W\\times 6}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Target Spatial Anchor. In addition to target cameras, we provide an oriented 3D point denoted as \\(\\mathbf{a}_i = [\\mathbf{R}_i|\\mathbf{t}_i]\\), which roughly defines the center of the subject's head, as" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.097, + 0.875, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.335, + 0.906, + 0.42 + ], + "angle": 0, + "content": "Figure 3. DiT and ControlMLP Block. Our DiT block (left) loosely follows [55], with a AdaIn-based timestep modulation. We apply attention and MLP blocks in parallel [10], and jointly apply self-attention to the noisy generated and identity conditioning tokens. ControlMLP block (right) is used to provide lightweight spatially-aligned conditioning - Plucker and Spatial Anchor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.906, + 0.518 + ], + "angle": 0, + "content": "well as their gaze direction. We show an example of our Spatial Anchor in Fig. 2 in the studio on the left. This anchor is color-coded and projected into a 2D image, which is used as conditioning for our target view generations. During inference, the Spatial Anchor can be placed at any given point that lies within the field-of-view of the target cameras." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.521, + 0.906, + 0.567 + ], + "angle": 0, + "content": "Multiview Diffusion Model. Given the above inputs, we train a multiview diffusion model \\(\\epsilon_{\\theta}\\) to jointly denoise all the target views \\(\\mathbf{y}_{1:N} \\in \\mathbb{R}^{N \\times H \\times W \\times C}\\) with the objective:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.578, + 0.905, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D M}} = \\left\\| \\boldsymbol {\\epsilon} ^ {t} - \\epsilon_ {\\theta} \\left(\\mathbf {y} _ {1: N} ^ {t}, \\mathbf {c} _ {1: N}, \\mathbf {x} ^ {\\text {r e f}}, \\mathbf {x} ^ {\\text {f a c e}}, t\\right) \\right\\| ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.669 + ], + "angle": 0, + "content": "where \\(\\mathbf{y}_{1:N}^{t}\\) are noisy target images. We condition the base model on the provided reference image and its face crop by concatenating their patchified latent tokens with the noisy input latent tokens to the model. This is shown in Fig. 2" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Mid-training. In mid-training stage, we want to train a strong multiview model that can denoise several images together, and absorb the dataset quickly at a lower resolution. During this phase, we do not use any pixel-aligned spatial control such as Plücker or Spatial Anchor. We use an MLP to encode the flattened 16-dimensional target camera intrinsics and extrinsics into a single token. We fuse this camera token into each noisy latent token (for the corresponding view) as positional encoding, which makes our multiview model 3D-aware of the target viewpoints. We mid-train our model at \\(128 \\times 128\\) resolution to jointly denoise 24 views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Post-training. In the post-training stage, our objective is to create a high-resolution model which is 3D-consistent starting with a low-resolution and a 3D-aware (but not consistent) model. For this, we design a lightweight Control-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16421" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "Net [93]-inspired module, which takes as input the pixel-aligned Plücker and Spatial Anchor controls, and the denoising timestep to create a separate modulation signal for the multiview model. We name this module ControlMLP, as it uses a single MLP to generate scale-and-shift modulated control for each multiview-DiT block as shown in Fig. 3. Each layer of ControlMLP is zero-initialized at the start. We find that the Post-training phase is crucial in reducing flicker and 3D inconsistencies in generations. We post-train models at \\(512 \\times 512\\) and \\(1024 \\times 1024\\) resolutions to jointly denoise 10 and 2 views respectively. Increasing the number of views further leads to GPU out-of-memory issues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.483, + 0.474 + ], + "angle": 0, + "content": "Encoding Plücker and Spatial Anchor. We notice that the relative differences between neighboring pixels in Plücker coordinates are tiny. To amplify these differences better, we use a SIREN [67] layer to first process the 6D grid into a 32D feature grid. Then, we downsample it by \\(8\\mathrm{x}\\) to match the size of latent tokens and feed it as input to the ControlMLP. In addition, use the Spatial Anchor to fix the position and orientation of the subject's head in 3D. We only use the Spatial Anchor for generations, and not for the input reference view. We encode the Spatial Anchor image into the latent space of our model via VAE, and concatenate it with Plücker input and pass it through an MLP to create the modulation signal at each layer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.481, + 0.483, + 0.498 + ], + "angle": 0, + "content": "3.3. Understanding and Improving Spatial Control" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.483, + 0.579 + ], + "angle": 0, + "content": "This section presents our design choices and examines alternative approaches for injecting pixel-aligned spatial controls during Post-training stage. We demonstrate the effectiveness of spatial control through a focused overfitting experiment with quantitative evaluations in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Scene Overfitting Task. We use 160 frames from a fixed 3D scene of a given subject and timestamp, split into 100 training and 60 validation views. We overfit our mid-trained model to the training views while testing various spatial control methods, training only the control modules while keeping other weights frozen. After overfitting for 10K iterations, we evaluate the model on validation views for novel view synthesis. Strong generalization to validation viewpoints indicates effective spatial control and appropriate camera viewpoint sensitivity. Through this task, we evaluate different spatial control injection methods in Tab. 1, starting with simple to advanced modulation designs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.826 + ], + "angle": 0, + "content": "- No overfitting (Row 1). The Mid-trained model without scene-specific overfitting achieves comparable PSNR of 19.2 and 19.7 on train and validation views respectively. We treat this setup as baseline to improve over." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Encoding Camera with MLP (Row 2). We encode camera using an MLP similar to prior works [41, 66] and our Mid-training stage (Sec. 3.2). After overfitting, the model achieves slightly better PSNR on training views as expected, however the validation PSNR drops by 1.28" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.09, + 0.882, + 0.19 + ], + "angle": 0, + "content": "
#Method (M2@ 128)PSNRval ↑PSNRtrain ↑
1.Mid-trained (No Overfitting)19.2319.70
2.+ Camera (w/ MLP) [41, 66]17.95-1.2819.92+0.22
3.+ Plücker (w/ MLP) [6, 23, 31, 68]18.89-0.3420.74+1.04
4.+ ControlMLP19.45+0.2229.36+9.66
5.+ SIREN20.13+0.9030.19+10.49
6.+ Spatial Anchor (Ours)22.60+3.3730.49+10.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.2, + 0.907, + 0.298 + ], + "angle": 0, + "content": "Table 1. Evaluating and Designing Spatial Controls (Sec. 3.3). We overfit our multi-view model for 10K iterations on 100 views of a single scene from our Head-only dataset, and evaluate on 60 novel views of this scene by only varying the spatial controls (camera and Spatial Anchor). We use this setup to test the modulation strength of different spatial controls. Subscript values green/red show deviations from Row 1." + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.312, + 0.905, + 0.341 + ], + "angle": 0, + "content": "points to 17.95. This suggests that an MLP does not provide enough modulation for camera control." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.342, + 0.905, + 0.446 + ], + "angle": 0, + "content": "- Plücker as Positional Encoding (Row 3). In this setup, we use downsampled and patchified Plücker coordinates processed through MLP to create positional encoding, which is added to the noisy latent tokens. This setup is inspired from prior works [3, 6, 23, 31, 68], and it further improves the validation PSNR compared to MLP at 18.89, but lags behind the non-overfitted baseline." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.447, + 0.905, + 0.552 + ], + "angle": 0, + "content": "- Plücker with ControlMLP and SIREN (Row 4, 5). Here, we use our ControlMLP module to inject spatial control at each multiview-DiT block output. Moreover, encoding Plücker coordinate with a SIREN [67] amplifies the relative differences between neighboring pixels (Sec. 3.2). This setup achieves PSNR of 20.13 with an improvement of 0.9 over baseline." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.553, + 0.905, + 0.614 + ], + "angle": 0, + "content": "- Adding Spatial Anchor (Row 6). Finally using the Spatial Anchor gives validation PSNR of 22.6 (gain of 3.3 points over baseline) and enables strong spatial control. Thus, we adopt this configuration for Post-training stage." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.342, + 0.905, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.623, + 0.807, + 0.639 + ], + "angle": 0, + "content": "3.4. Enhanced 3D Consistency Metric" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.841 + ], + "angle": 0, + "content": "Traditionally, the 3D consistency of multiview generation models is evaluated using 2D image metrics such as PSNR, LPIPS, and SSIM against a fixed set of ground truth images. However, this approach unfairly penalizes models that generate plausible and 3D-consistent novel content deviating from the fixed ground truth images. Some works try to address this by measuring SfM [17, 60] or epipolar error [45, 89], but these methods solve for pose or are not robust since they measure against the entire epipolar line. To address these limitations, we use our GT camera poses as input and compute Reprojection Error (RE) given our known camera poses and predicted correspondences. Our computation of RE involves the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.841, + 0.905, + 0.902 + ], + "angle": 0, + "content": "1. Landmarks and Correspondence estimation. We use SuperPoint [13] to detect landmarks in the generated images and employ SuperGlue [59] to establish pairwise correspondences between landmarks across images." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16422" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "2. Triangulation. Given the correspondences and camera parameters, we apply Triangulation based on Direct Linear Transformation (DLT) [22] to obtain the corresponding 3D points for each landmark." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.152, + 0.482, + 0.227 + ], + "angle": 0, + "content": "3. Reprojection and Error calculation. We reproject these 3D points onto each image and compute the RE as the L2 distance between the original landmark and the reprojected 3D point normalized by image resolution, and average error across all images." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.483, + 0.366 + ], + "angle": 0, + "content": "This approach evaluates multiview generation models by focusing on their ability to produce 3D-consistent results rather than adhering to a fixed ground truth. The Reprojection Error provides a valuable basis for comparison across different methods. Furthermore, by computing RE on a set of real-world images that are independent of our generated images, we can establish a baseline that quantifies the error due to the noisy predictions from SuperGlue and SuperPoint rather than the quality of our generated images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.504 + ], + "angle": 0, + "content": "Naming Convention (RE@SG). Please note that the use of SuperPoint [13] and SuperGlue [59] estimation modules is a particular instantiation of our metric, and these could be replaced in the future with stronger counterparts such as MAST3R [36, 75] or domain-specific keypoint detectors such as Sapiens [33]. Thus we use the naming convention of RE@SG to denote the Reprojection Error (RE) under SuperGlue (SG) estimation, which could be modified accordingly in the future for different estimators." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.223, + 0.541 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.613 + ], + "angle": 0, + "content": "We present details of the datasets used in all training and validation stages, followed by discussion of evaluation metrics - with particular emphasis on the 3D consistency metric and conclude with core experimental results and ablations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.629, + 0.165, + 0.644 + ], + "angle": 0, + "content": "4.1.Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.714 + ], + "angle": 0, + "content": "Humans-3B Dataset. We utilize a large proprietary dataset of approximately 3 billion human-centric in-the-wild images for pretraining. We provide more details on data filtering and curation in ??" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Head and Full-body Studio Datasets. We rely on high-quality proprietary studio captures as our primary source of data for learning 3D consistency. Our model comes in two variants: head-only and full-body, each trained (for Mid-training and Post-training stages) on corresponding datasets. For our full-body model, we utilize a dataset of 861 subjects (811 train, 50 test), nearly 1000 frames per subject. For our head-only model, we use a dataset of 1450 subjects (1400 train, 50 test), nearly 40000 frames per subject. The studio setup is similar to [43], with two capture domes for capturing full-body and head-only high-resolution 4K images with 230 and 160 cameras." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "iPhone Dataset. To evaluate real-world performance, we collect casual images of 50 test subjects in an indoor office environment using an iPhone 13 Pro. We preprocess these images with Sapiens-2B [33] for background segmentation before model inference. This dataset serves exclusively for evaluating our model's performance on in-the-wild inputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.78, + 0.21 + ], + "angle": 0, + "content": "4.2. Evaluation Setup and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.217, + 0.905, + 0.338 + ], + "angle": 0, + "content": "3D Consistency. Following prior works, we report standard metrics like PSNR, SSIM (\\( \\times 100 \\)), and LPIPS (\\( \\times 100 \\)) metrics. However, these metrics unfairly penalize plausible novel views generated under incomplete inputs. We therefore introduce the Reprojection Error metric, described in Sec. 3.4, which validates 3D consistency without directly relying on ground truth. Our evaluation generates 4 randomly selected views from the test split." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.418 + ], + "angle": 0, + "content": "Identity Preservation. We use two metrics that measure identity preservation across generated views by computing cosine distance between features extracted via FaceNet [4, 61] for face similarity; and CLIP [52] vision encoder for full-body similarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.905, + 0.513 + ], + "angle": 0, + "content": "Pretrained Model Evaluation. We measure the effectiveness of our pre-training strategy by reporting the FID [24]. We use a smaller annotated 30M subset of Humans-3B for training image and text-conditioned P1@128 models, a 30M subset of unfiltered dataset for training No Filtering P1@128, and a 1000 sample test set from iPhone dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.525, + 0.607, + 0.54 + ], + "angle": 0, + "content": "4.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.905, + 0.639 + ], + "angle": 0, + "content": "Pretraining and Data Filtering. Tab. 2 presents our pretrained model results (Row 1) and ablations on Human-centric data filtering and Image-conditioned pretraining (Rows 2-5). Human-centric filtering and image-based conditioning are both critical to achieve high-quality generation. The qualitative results are shown in ??" + }, + { + "type": "table", + "bbox": [ + 0.58, + 0.653, + 0.842, + 0.749 + ], + "angle": 0, + "content": "
#Method (Stage @ Resolution)FID ↓
1.Pippo (P1@512)51.164
2.Human-centric Filtering (P1@128)75.639
3.No Filtering (P1@128)86.838
4.Image-conditioned (P1@128)75.639
5.Text-conditioned (P1@128)109.720
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.759, + 0.907, + 0.801 + ], + "angle": 0, + "content": "Table 2. Pretraining and Data Filtering. We report results of the full pretrained model P1@512, and compare several variants of P1@128 models. We report FID on iPhone dataset (1k samples)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "High-resolution Multi-view Generation. In Tab. 3, we evaluate 3D reconstruction and identity preservation for unseen subjects from the studio datasets. We show that increasing the output resolution of generation in our approach does not hurt 3D consistency or similarity. We put corresponding visuals in ??, Rows 2 and 3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16423" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.478, + 0.206 + ], + "angle": 0, + "content": "
# Split & Resolution3D ConsistencySimilarity
RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
Head-only
1. Studio (128)3.3+0.821.067.614.862.0-13.361.5-16.2
2. Studio (1K)3.4+0.520.372.026.273.5-2.579.4-0.1
3. iPhone Face (1K)3.0---67.6-
Full-body
4. Studio (128)3.6+0.122.884.010.041.7-8.567.1-9.4
5. Studio (1K)1.5+0.122.491.711.164.7-6.074.1-2.2
6. iPhone (1K)1.7---58.068.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.217, + 0.482, + 0.314 + ], + "angle": 0, + "content": "Table 3. Results on Unseen Studio and iPhone data (Sec. 4.3.) We report results on Post-trained Pippo models at three different resolutions. We report 3D metrics PSNR, SSIM, and LPIPS; as well as our proposed Reprojection Error (RE@SG) under SuperGlue estimation which does not require ground truth (Sec. 3.4). We report 2D Face and Body similarities. Red subscript show deviation against ground truth value." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.483, + 0.515 + ], + "angle": 0, + "content": "Generations from Casual iPhone Photos. We present in Tab. 3 (Rows 3,6) Reprojection Error and similarity scores for casually taken images from the iPhone dataset with 1K resolution model. In this scenario, the standard reconstruction error metrics cannot be assessed due to missing ground truth. We find that the reprojection error on iPhone captures remains comparable to the studio dataset - demonstrating 3D consistency. This illustrates the generalizability of Pippo beyond the multi-view training data domain, where our pretraining with large-scale in-the-wild human data is critical. We put corresponding visuals in ??, Row 1 (Appendix)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Comparisons with External Benchmarks. In Fig. 4, we compare Pippo with individual state-of-the-art baselines in Full-body and Head-only generation. SiTH [25] reconstructs textured human-mesh using ControlNet paired with a SDF representation. Compared to SiTH, Pippo facilitates high-resolution and accurate multiview synthesis. DiffPortrait3D [20] inverts a 3D-GAN based on a given input image. Compared to it, our model supports greater viewpoint variability and ensures closer adherence to the input image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Quantitative comparisons and baselines. Existing SoTA Human methods [20, 25] use explicit SMPL priors, and thus are difficult to compare with directly. Qualitatively, we found that they cannot handle novel views or preserve details (Fig. 4), and hence do not quantitatively compare against them. In Pippo, we focus on creating a strong multiview human generator, and we benchmark four state-of-the-art multi-view diffusion models on the iPhone full-body dataset in Tab. 4. We find that Pippo preserves identity (i.e., face and body similarity) and 3D consistency (RE) better while also operating at a higher resolution compared to baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Benchmarking Pippo on public datasets. In Tab. 5, we benchmark Pippo on the public Codec Avatar datasets [43] to aid future comparisons. Specifically, we use Ava-256 containing 256 head-only captures and Goliath containing 4" + }, + { + "type": "table", + "bbox": [ + 0.569, + 0.089, + 0.852, + 0.161 + ], + "angle": 0, + "content": "
#Method & ResolutionRE@SG ↓Face ↑Body ↑
1.MV-Adapter [27] (768)4.743.064.1
2.Era3D [38] (512)4.138.164.4
3.Wonder3D [42] (256)5.334.758.8
4.Pippo (P3@1K)3.058.068.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.512, + 0.172, + 0.905, + 0.241 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison with SoTA multi-view models. We quantitatively compare Pippo against state-of-the-art multi-view diffusion models. We find that Pippo preserves identity (i.e., face and body similarity) and 3D consistency (RE) better while operating at a higher resolution compared to baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.905, + 0.328 + ], + "angle": 0, + "content": "full-body. During evaluation, we only use subsets of these datasets that were not used for training. We find that Pippo's performance on these datasets is inline with its performance on internal studio datasets." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.339, + 0.905, + 0.388 + ], + "angle": 0, + "content": "
#Dataset (P3@1K)RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
1.Ava-256 (Head-only)3.820.168.726.672.976.8
2.Goliath (Full-body)1.220.489.715.587.577.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.512, + 0.399, + 0.905, + 0.456 + ], + "angle": 0, + "content": "Table 5. Benchmarking Pippo on public Ava-256 and Goliath datasets. We find that Pippo's performance on these datasets is inline with its performance on internal studio datasets. This will aid future comparisons against Pippo." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.625, + 0.497 + ], + "angle": 0, + "content": "4.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.905, + 0.551 + ], + "angle": 0, + "content": "We examine design choices at each training stage (Sec. 3) and present our ablation results in Tab. 6. Experiments are conducted at \\(128 \\times 128\\) resolution on the Head-only dataset." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.563, + 0.905, + 0.732 + ], + "angle": 0, + "content": "
#Method (Stage @ Res.)3D ConsistencySimilarity
RE@SG↓PSNR↑SSIM↑LPIPS↓Face ↑Body ↑
1.Pippo (P3@512)2.7+0.721.771.721.574.1-2.077.4-2.4
2.Pippo Face-only (P3@512)2.4+0.420.570.325.374.3-1.775.8-4.0
3.No Mid-train (P3@512)5.6+3.614.559.544.220.4-55.663.4-16.4
4.Pippo (P3@128)3.3+0.821.067.614.862.0-13.361.5-16.2
5.No Anchor (P3@128)11.5+9.017.154.021.163.1-12.268.3-9.3
6.No Plücker (P3@128)4.2+1.720.264.516.163.5-11.866.1-11.5
7.Pippo (M2@128)3.4+0.919.161.417.260.0-15.362.6-15.0
8.No Pre-train (M2@128)5.9+3.413.139.344.919.5-55.849.0-28.6
9.Cross-Attn (M2@128)3.5+1.018.059.222.166.2-9.170.7-7.0
10.Non-frontal (M2@128)5.8+3.315.252.430.149.2-26.160.7-16.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.512, + 0.742, + 0.905, + 0.84 + ], + "angle": 0, + "content": "Table 6. Ablation on design choices and training stages. We evaluate several multi-view models at \\(128 \\times 128\\) and \\(512 \\times 512\\) resolution at different training stages on Head-only dataset. We ablate the choice of doing Mid-training and Pre-training, along with different spatial controls and reference conditioning methods..The training stages are labeled as M2 (Mid-training) and P3 (Post-training), followed by @ specified resolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Significance of Pre-training and Mid-training. Pretraining our model on the Humans-3B dataset enables robust generalization to novel identities, as demonstrated in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16424" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.086, + 0.907, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.294, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Visual comparison with state-of-the-art methods. We visually compare Pippo with state-of-the-art baselines DiffPortrait3D [20] (head-only) and SiTH [25] (full-body) generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.348, + 0.482, + 0.408 + ], + "angle": 0, + "content": "Tab. 6, Row 8. Without pre-training, model generalization deteriorates, resulting in unclear facial features. Skipping mid-training at lower resolution impairs consistent multiview generation, as shown in Row 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.501 + ], + "angle": 0, + "content": "Importance of Frontal Input Reference. Our ablation in Tab. 6, Row 10 demonstrates that completely randomizing the view point of an input reference image leads to overfitting to training identities. Non-frontal views, especially rear views, have very limited information about the identity which forces the network to pick up spurious correlations." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.516, + 0.482, + 0.585 + ], + "angle": 0, + "content": "
# Pretrain Data (M2@128)RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
1. 30M (1% HQ)3.821.381.813.330.666.2
2. 900M (30% Random)4.021.682.012.130.667.2
3. 2.1B (70% Random)3.721.682.112.137.667.5
4. Humans-3B (100%)3.721.982.312.741.767.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.587, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Table 7. Ablating the size of pre-training dataset on Humans-3B. We conduct Full-body mid-training with pretrained checkpoints trained on \\(1\\%\\), \\(30\\%\\), \\(70\\%\\), and \\(100\\%\\) subsets of Humans-3B dataset. We found large-scale data is crucial for generalization to novel identities (i.e., face similarity)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Importance of Self-attention. Replacing self-attention with cross-attention for reference image encoding, using the same routing as image-conditioned pretraining from Sec. 3.1, leads to degraded performance as shown in Tab. 6, Row 9. We observe that this setup causes the model to ignore input conditioning, generating images that only vaguely resemble training subjects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Role of large scale Humans-3B pre-training dataset. We utilized intermediate checkpoints from pre-training stage (P1) trained on \\(30\\%\\) and \\(70\\%\\) of the data, and a separate checkpoint trained on \\(1\\%\\) high-quality subset of Humans-3B. Starting from these checkpoints, we mid-trained Pippo to denoise 4 views at \\(128 \\times 128\\) resolution on Full-body dataset for two days. We report their respective results in Tab. 7. We found large-scale data is crucial for general" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.38 + ], + "angle": 0, + "content": "ization to novel identities – indicated by high gains in face similarity metric." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.393, + 0.634, + 0.41 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.722 + ], + "angle": 0, + "content": "We present Pippo, a diffusion transformer model that generates a dense set of high-resolution multi-view consistent images of a person from a single image. Our experiments show that our multi-stage training strategy, which combines largescale in-the-wild data with high-quality multi-view studio captures, enables generalizable high-resolution multi-view synthesis. Analysis of the diffusion transformer architecture reveals that self-attention with the reference image, Plücker coordinates with SIREN, and Spatial Anchor are all essential for high-fidelity multi-view human generation. Pippo achieves, for the first time, consistent multi-view human image generation at 1K resolution. We also show that our proposed 3D consistency metric enables evaluation of 3D consistency without paired ground-truth data. One of the limitations of our approach is the limited number of simultaneously generated views, originating from large context length and memory constraints, which can be tackled with parallelization techniques and autoregressive generation. Extending our approach to multi-view consistent video generation will be addressed in future work." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.736, + 0.611, + 0.751 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.761, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panohed: Geometry-aware 3d fullhead synthesis in 360deg. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20950-20959, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.831, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[2] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabián Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM Transactions on Graphics (TOG), 40(4):1-17, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.761, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16425" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[3] Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, et al. Vd3d: Taming large video diffusion transformers for 3d camera control. arXiv preprint arXiv:2407.12781, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[4] Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2018), pages 67-74, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.247, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[6] Eric Ming Chen, Sidhanth Holalkere, Ruyu Yan, Kai Zhang, and Abe Davis. Ray conditioning: Trading photoconsistency for photo-realism in multi-view image generation. arXiv preprint arXiv:2304.13681, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[8] Wei Cheng, Ruixiang Chen, Wanqi Yin, Siming Fan, Keyu Chen, Honglin He, Huiwen Luo, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, Daxuan Ren, Lei Yang, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, Bo Dai, and Kwan-Yee Lin. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. arXiv preprint, arXiv:2307.10173, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.526, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[9] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.554, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[10] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme Ruiz, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd Van Steenkiste, Gamaleldin Fathy Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmiŋn Bastings, Mark Collier, Alexey A. Gritsenko, Vighnesh Birodkar, Cristina Nader Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lucic, Xiaohua Zhai, Daniel Keysers, Jeremiah J. Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters. In Proceedings of the 40th International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[11] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[13] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.133 + ], + "angle": 0, + "content": "and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.135, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[14] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In ACM MM, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[15] Nikita Drobyshev, Antoni Bigata Casademunt, Konstantinos Vougioukas, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Emoportraits: Emotion-enhanced multimodal one-shot head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8498-8507, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.274, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.904, + 0.411 + ], + "angle": 0, + "content": "[17] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[18] Matheus Gadelha, Rui Wang, and Subhransu Maji. Multiresolution tree networks for 3d point cloud processing. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[19] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.525, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[20] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10456-10465, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[21] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[22] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[23] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[25] I Ho, Jie Song, Otmar Hilliges, et al. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 538-549, 2024. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16426" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[27] Zehuan Huang, Yuanchen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.178, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[28] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9784-9794, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.261, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[29] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.331, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[30] Yash Kant, Aliaksandr Siarohin, Michael Vasilkovsky, Riza Alp Guler, Jian Ren, Sergey Tulyakov, and Igor Gilitschenski. invs: Repurposing diffusion inpainters for novel view synthesis. In SIGGRAPH Asia 2023 Conference Papers, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.401, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[31] Yash Kant, Aliaksandr Siarohin, Ziyi Wu, Michael Vasilkovsky, Guocheng Qian, Jian Ren, Riza Alp Guler, Bernard Ghanem, Sergey Tulyakov, and Igor Gilitschenski. Spad: Spatially aware multi-view diffusers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10026-10038, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.484, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[32] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.526, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[33] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.596, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[34] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.651, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[35] Black Forest Labs. Flux model, 2024. Accessed: 2024-11-10. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.679, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[36] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.707, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[37] Heyuan Li, Ce Chen, Tianhao Shi, Yuda Qiu, Sizhe An, Guanying Chen, and Xiaoguang Han. Spherehead: Stable 3d full-head synthesis with spherical tri-plane representation. In European Conference on Computer Vision, pages 324-341. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.777, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[38] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, Wenping Wang, Qifeng Liu, and Yike Guo. Era3d: High-resolution multiview diffusion using efficient row-wise attention, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[39] Chieh Hubert Lin, Chia-Che Chang, Yu-Sheng Chen, Da-Cheng Juan, Wei Wei, and Hwann-Tzong Chen. Coco-gan: Generation by parts via conditional coordinating. In ICCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.196, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[42] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.268, + 0.905, + 0.57 + ], + "angle": 0, + "content": "[43] Julieta Martinez, Emily Kim, Javier Romero, Timur Bagautdinov, Shunsuke Saito, Shoou-I Yu, Stuart Anderson, Michael Zollhöfer, Te-Li Wang, Shaojie Bai, Chenghui Li, Shih-En Wei, Rohan Joshi, Wyatt Borsos, Tomas Simon, Jason Saragih, Paul Theodosis, Alexander Greene, Anjani Josyula, Silvio Mano Maeta, Andrew I. Jewett, Simon Venshtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mohamed Ezzeldin A. Elshaer, Tingfang Du, Longhua Wu, Shen-Chi Chen, Kai Kang, Michael Wu, Youssef Emad, Steven Longay, Ashley Brewer, Hitesh Shah, James Booth, Taylor Koska, Kayla Haidle, Matt Andromalos, Joanna Hsu, Thomas Dauer, Peter Selednik, Tim Godisart, Scott Ardisson, Matthew Cipperly, Ben Humberston, Lon Farr, Bob Hansen, Peihong Guo, Dave Braun, Steven Krenn, He Wen, Lucas Evans, Natalia Fadeeva, Matthew Stewart, Gabriel Schwartz, Divam Gupta, Gyeongsik Moon, Kaiwen Guo, Yuan Dong, Yichen Xu, Takaaki Shiratori, Fabian Prada, Bernardo R. Pires, Bo Peng, Julia Buffalini, Autumn Trimble, Kevyn McPhail, Melissa Schoeller, and Yaser Sheikh. Codec Avatar Studio: Paired Human Captures for Complete, Driveable, and Generalizable Avatars. NeurIPS Track on Datasets and Benchmarks, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.905, + 0.656 + ], + "angle": 0, + "content": "[44] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[45] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[46] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[47] Dongwei Pan, Long Zhuo, Jingtan Piao, Huiwen Luo, Wei Cheng, Yuxin Wang, Siming Fan, Shengqi Liu, Lei Yang, Bo Dai, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, and Kwan-Yee Lin. RenderMe-360: Large digital asset library and benchmark towards high-fidelity head avatars. In Thirty-seventh Conference on Neural Information" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16427" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Processing Systems Datasets and Benchmarks Track, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[48] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.235 + ], + "angle": 0, + "content": "[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.483, + 0.279 + ], + "angle": 0, + "content": "[50] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.282, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[51] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.354, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[53] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[54] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.556, + 0.483, + 0.597 + ], + "angle": 0, + "content": "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[56] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.483, + 0.727 + ], + "angle": 0, + "content": "[57] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[58] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[60] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[61] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.906, + 0.22 + ], + "angle": 0, + "content": "[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.906, + 0.303 + ], + "angle": 0, + "content": "[63] Akash Sengupta, Thiemo Alldieck, Nikos Kolotouros, Enric Corona, Andrei Zanfir, and Cristian Sminchisescu. Diffhuman: Probabilistic photorealistic 3d reconstruction of humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1439-1449, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.906, + 0.376 + ], + "angle": 0, + "content": "[64] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.379, + 0.906, + 0.42 + ], + "angle": 0, + "content": "[65] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.422, + 0.906, + 0.464 + ], + "angle": 0, + "content": "[66] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.466, + 0.906, + 0.508 + ], + "angle": 0, + "content": "[67] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.51, + 0.906, + 0.564 + ], + "angle": 0, + "content": "[68] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.906, + 0.61 + ], + "angle": 0, + "content": "[69] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.612, + 0.906, + 0.668 + ], + "angle": 0, + "content": "[70] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In European Conference on Computer Vision, pages 107-124. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.669, + 0.906, + 0.725 + ], + "angle": 0, + "content": "[71] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14795-14805, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.727, + 0.906, + 0.768 + ], + "angle": 0, + "content": "[72] Qingyang Tan, Lin Gao, Yu-Kun Lai, and Shihong Xia. Variational autoencoders for deforming 3d mesh models. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.771, + 0.906, + 0.813 + ], + "angle": 0, + "content": "[73] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[74] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[75] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16428" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[76] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[77] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[78] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[79] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[80] Jing Wen, Xiaoming Zhao, Zhongzheng Ren, Alexander G Schwing, and Shenlong Wang. Gomavatar: Efficient animatable human modeling from monocular video using gaussians-on-mesh. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2059-2069, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[81] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[82] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[83] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Song-Chun Zhu, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[84] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, et al. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19801-19811, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[85] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[86] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[87] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. FaceScape: A largescale high quality 3D face dataset and detailed riggable 3D" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "face prediction. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[88] T Yenamandra, A Tewari, F Bernard, HP Seidel, M Elgharib, D Cremers, and C Theobalt. i3DMM: Deep implicit 3D morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[89] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[90] Zhixuan Yu, Jae Shin Yoon, In Kyu Lee, Prashanth Venkatesh, Jaesik Park, Jihun Yu, and Hyun Soo Park. HUMBI: A large multiview dataset of human body expressions. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.907, + 0.375 + ], + "angle": 0, + "content": "[91] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[92] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22096-22105, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[93] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[94] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16429" + } + ] +] \ No newline at end of file diff --git a/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/2679a0a3-d3fd-49c2-aeba-9db847b48300_origin.pdf b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/2679a0a3-d3fd-49c2-aeba-9db847b48300_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8718a3f376b2aab0c2431e2af55b69ecbe7b84ba --- /dev/null +++ b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/2679a0a3-d3fd-49c2-aeba-9db847b48300_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5615684accef3d041325e540ddcaaad01433d31eb3c5f43ecb883df51b3b613f +size 1203930 diff --git a/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/full.md b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6627bf748f69c698267175436f24109a2ce591d0 --- /dev/null +++ b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/full.md @@ -0,0 +1,330 @@ +# Pippo: High-Resolution Multi-View Humans from a Single Image + +Yash Kant $^{1,2,3}$ , Ethan Weber $^{1,4}$ , Jin Kyu Kim $^{1}$ , Rawal Khirodkar $^{1}$ , Su Zhaoen $^{1}$ , Julieta Martinez $^{1}$ , Igor Gilitschenski $^{*,2,3}$ , Shunsuke Saito $^{*,1}$ , Timur Bagautdinov $^{*,1}$ $^{1}$ Meta Reality Labs, $^{2}$ University of Toronto, $^{3}$ Vector Institute, $^{4}$ UC Berkeley https://yashkant.github.io/pippo/ +*Joint Advising + +![](images/c5cebda302024c92c2938cc7a88a490127e4df0e93ba112539057ae62f0db3a8.jpg) +Figure 1. Pippo produces high-resolution, multi-view, studio-quality images from a single image. First row shows full-body generation of an unseen subject captured from an iPhone picture. Second and third rows show face-only and partial full-body input images from unseen studio subject(s) to novel views respectively. All visuals are at 1K resolution. + +# Abstract + +We present Pippo, a generative model capable of producing 1K resolution dense turnaround videos of a person from a single casually clicked photo. Pippo is a multi-view diffusion transformer and does not require any additional + +inputs — e.g., a fitted parametric model or camera parameters of the input image. We pre-train Pippo on 3B human images without captions, and conduct multi-view mid-training and post-training on studio captured humans. During mid-training, to quickly absorb the studio dataset, we denoise several (upto 48) views at low-resolution, and encode tar + +get cameras coarsely using a shallow MLP. During posttraining, we denoise fewer views at high-resolution and use pixel-aligned controls (e.g., Spatial anchor and Plucker rays) to enable 3D consistent generations. At inference, we propose an attention biasing technique that allows Pippo to simultaneously generate greater than $5 \times$ as many views as seen during training. Finally, we also introduce an improved metric to evaluate 3D consistency of multi-view generations, and show that Pippo outperforms existing works on multi-view human generation from a single image. + +# 1. Introduction + +Creating photorealistic human representations with the ability to control viewpoints has numerous applications in entertainment, healthcare, fashion and social media. Building such representations, first and foremost, requires high-quality multi-view studio data [2, 29], which is costly to acquire. This significantly limits the scalability in terms of the number identities for high-quality studio data [29, 43, 84]. In contrast, large-scale, unstructured, and diverse human images and videos are available online. However, the raw data of such in-the-wild images does not offer ground-truth 3D or multi-view representations of humans. + +In this work, we present a novel approach leveraging the best of two worlds: generalizability from in-the-wild unstructured images, and fidelity and view-controllability from studio capture data. Specifically, our model Pippo is a diffusion transformer, which can generate several 1K resolution multi-view consistent images jointly during inference. Pippo takes as input a single image of an individual, camera poses of the target viewpoints to be generated. Since the scale and placement of the subject is ambiguous from a single image, Pippo uses a Spatial Anchor which roughly specifies the location and orientation of the subject in 3D space. Our model does not rely on additional conditioning such as body priors or camera parameters of the input images, to scale our training pipeline to in-the-wild data and support unconstrained inputs at test time. + +We employ a multi-stage training recipe to train Pippo. First, we pre-train the model for latent-to-image generation task, similar to [54], on a large human-centric dataset of in-the-wild images. Next, in Mid-training stage, we jointly generate multiple consistent images of the subject, conditioned on target viewpoints and a single input image using high-quality studio dataset. Finally, in Post-training stage, the model is provided with a minimal placement signal - spatial anchor - encoding rough head orientation - which further improves 3D consistency. Our architecture is also carefully tailored to the conditional multi-view generation - we propose several simple but effective modifications to the basic DiT architecture, including self-attention-based + +conditioning, lightweight spatial controls and camera conditioning with Plucker coordinates. + +For evaluating such a multi-view diffusion model, 3D consistency is a critical metric for understanding the geometric correctness of the generation. For generation task in an ill-posed setup (e.g., a single image input), there may be multiple possibilities that are equally plausible. Existing multi-view generation methods typically report reconstruction metrics (i.e., PSNR, SSIM and LPIPS) [19] or FID [41], which either (1) penalize any new content that is actually 3D consistent or (2) are unable to measure the 3D consistency of the generation. To address this problem, we design a metric measuring 3D consistency: we compute 2D keypoint matches with an off-the-shelf method [59], triangulate them, and then reproject back into the other views to measure the reprojection error in pixels. Some generation methods measure reprojection errors [17] from SfM [60] or epipolar distance [45, 89], but our metric uses camera pose as input and is more precise than measuring distance to a epipolar line. We show our metric helps to quantify our results, finding that our method is favorable compared with existing approaches and other baseline methods. + +To summarize, we make the following contributions: + +- a generative model capable of generating high-resolution and multi-view consistent humans from a single image and its effective training strategy. +- a diffusion transformer architecture designed to enhance multi-view generation and viewpoint control. +- a novel 3D consistency metric tailored to accurately measure the level of 3D consistency in generative tasks. + +# 2. Related Work + +We review multi-view human datasets and generative models for human synthesis, categorizing methods by their data requirements and prior constraints—such as parametric human models or explicit 3D structures. In this work, we minimize reliance on complex priors by training the model on large amounts of human-centric data and impose minimal camera and spatial controls. + +Multi-view Human Datasets. While many large 2D human datasets exist [12, 39, 62], 3d captures of people (i.e., captures using dozens of views at a time) are still relatively rare, since they are expensive to collect and are thus mostly obtained in specialized research labs. Nevertheless, this kind of data has recently become more common, going from hundreds to thousands of publicly-available captures. For face-only captures, a few datasets [43, 47, 87, 88, 90] provide up to 600 subjects captured from over 60 views each, while for full-body captures, 4D-Dress [77] provides 32 subjects with 2 outfits each, and detailed annotations for their garments, while DNA-rendering [8] provides 500 subjects from 60 views. MVHumanNet [84] stands apart by + +![](images/274551e9bf4dd5a03572aa716c816b4fd81239c27280c6a69e5ac8d6d24eb537.jpg) +Figure 2. Pipeline overview. This is an illustration of how we train our model. (Left) we use data from a studio capture and train our multi-view diffusion model (right). We condition on a full reference photo and a cropped face, as well as the target view cameras and 2D projected spatial anchor indicating head position and orientation. Our diffusion model also takes in noisy target views and a timestep in order to predict the denoised views (top). In practice, we apply a segmentation mask around the person. + +providing 4,500 people captured from 48 views in and over 9,000 sets of clothing, outpacing other datasets by an order of magnitude in terms of people diversity. We use internal multi-view data of $\sim 1000$ identities with dense ( $\sim 160$ ) camera setup, but we expect our model to produce reasonable results with publicly-available captures. + +Generating Novel Poses and Expressions. Although different from our task, this related work builds 3D animatable models of people and faces. These methods use neural radiance fields [70, 71, 81], or 3d Gaussians [34, 51, 80, 94], using canonical "T"-poses and a corresponding forward and backward mapping to model local appearance changes. Many of these methods, however, rely on accurate human shape and pose estimation, and only achieve high photorealism for personalized models from studio captures. Relatedly, a family of methods focus on generating animatable faces from a single image by disentangling viewpoints and expressions from large collections of 2d portraits [14, 15, 32, 48, 86, 92], allowing photorealistic real-time reenactment, while being limited to faces and small viewpoint variations. Unlike these methods, we complete missing parts of a person given either one or a few partial views. We do not repose the person or animate the faces but can instead recover entirely missing views, which is orthogonal to these methods. Moreover, our approach addresses both full-body and facial reconstruction, rather than specializing in either domain. + +Generations with Explicit 3D Structures. Similar to early methods that extracted 3d representations from 2d models via score distillation [7, 11, 40, 50, 53, 74, 78], 3d representations of human appearance have been extracted from GANs by co-training a neural renderer from learned TriPlane [5] or HexPlane [1, 37] latent spaces. Others train generative models that operate directly in 3D space + +with diffusion or GAN objectives [18, 44, 72, 76, 82, 83] or regression models to directly predict 3D representations [56, 57, 63]. While these methods produce 3d consistent faces by design, their quality is limited by the relatively small amounts of 3d training data and/or the limitations of their respective 3d representations. In contrast, we focus on generating 3d-consistent 2d images and thus avoid the downsides of explicit 3d modelling. + +Generations with 2D Models. Another option is to model viewpoint changes in 2d, without an underlying 3d representation [41], sometimes modelling either photometric [30] or epipolar [28] constraints explicitly. Diff-Portrait3D [20] is a portrait-generation method that falls roughly within this category, by fine-tuning a 2d diffusion model for 3d-aware face generation using ControlNet-style [93] conditioning for viewpoint control, as well as cross-view attention [21] and initialization from a 3d-consistent GAN [5]. Other methods are more general and not specific to humans, focusing on single-view to 3D generation [19, 41, 58, 64, 65, 73, 85]. In addition to these single-view methods, video models have been fine-tuned for camera control [23, 79]. + +# 3. Method + +We train our models following a three-stage strategy: + +- Image-only Pre-training (P1). We pretrain on a large-scale human-centric dataset with image conditioning. +- Multiview Mid-training (M2). We train models at a low-resolution of $128 \times 128$ to denoise 48 target views with coarse camera control (no pixel-aligned spatial control). +- Multiview Post-training (P3). We train at a high-resolution of $1024 \times 1024$ to denoise $1 - 3$ target views with spatial control injected via ControlMLP layers. + +We denote the training stage and resolution of any given model as $\{\mathrm{stage}\} @\{\mathrm{resolution}\}$ . For instance, M2@128 represents a mid-trained model at 128 resolution. + +# 3.1. Base Model + +Architecture. We adopt a DiT-like [49] architecture with scale, shift, and gate modulation for timestep conditioning inspired by Stable Diffusion 3 [16] and Flux [35]. We simplify the architecture by employing MLP and attention in parallel [91], and removing second LayerNorm after attention layers. We use VAE for training in latent space with 8x spatial compression, and patchify latent images via a linear layer and patch size of 2. We use fixed sinusoidal positional encoding during training. We provide more details in ??, and the exact design of our DiT block in Fig. 3. + +Image-only Pre-training. During pre-training, the model learns to denoise an image conditioned on its corresponding image embedding from DINOv2 [9, 46], this is similar in principle to the image decoder of DALL-E 2 [54]. We project both embeddings to the model dimension using with a linear layer to create a joint conditioning. Importantly, our pretraining setup does not require any annotations or captions for the images, and is well-aligned with our downstream objective of generating consistent multiview images given a single reference image as input. + +Formally, given an image $\mathbf{y} \in \mathbb{R}^{H \times W \times C}$ , and the joint conditioning as $\mathbf{e}^{\mathrm{img}} \in \mathbb{R}^{N \times D}$ . We pre-train our diffusion model $\epsilon_{\theta}$ with the following objective: + +$$ +\mathcal {L} _ {\mathrm {D M}} = \left| \left| \boldsymbol {\epsilon} ^ {t} - \epsilon_ {\theta} \left(\mathbf {y} ^ {t}, \mathbf {e} ^ {\mathrm {i m g}}, t\right) \right| \right| ^ {2} \tag {1} +$$ + +Where $t \in [0, T]$ is diffusion timestep, $\epsilon^t \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ is the noise added at the given timestep. We use the DDPM [26, 69] formulation to define discrete timesteps and set $T = 1000$ . We first pretrain our model at $256 \times 256$ and then at $512 \times 512$ on a large corpus of human-centric images. Exact training details can be found in ?? + +# 3.2. Multiview Model + +Our goal is to generate many high resolution and unseen novel viewpoints of a human subject (akin to a studio capture) given a single input image. + +Input Reference. We denote the input image as $\mathbf{x}^{\mathrm{ref}}$ and corresponding face crop as $\mathbf{x}^{\mathrm{face}}$ . The face crop is obtained using FaceNet [61] and resized to the same size as input image, such that: $\mathbf{x}^{\mathrm{face}}, \mathbf{x}^{\mathrm{ref}} \in \mathbb{R}^{H \times W \times C}$ . + +Target Cameras. We denote the target viewpoints to be synthesized using distinct cameras (intrinsic and extrinsic), represented as $\mathbf{c}_{1:N}$ . Each camera is used to generate its Plücker coordinates $\mathbf{P}_i\in \mathbb{R}^{H\times W\times 6}$ . + +Target Spatial Anchor. In addition to target cameras, we provide an oriented 3D point denoted as $\mathbf{a}_i = [\mathbf{R}_i|\mathbf{t}_i]$ , which roughly defines the center of the subject's head, as + +![](images/a9c12939a1366ca61e3051df02041f892cb9bd5c4751d1a1f106a75a05c4bc3c.jpg) +Figure 3. DiT and ControlMLP Block. Our DiT block (left) loosely follows [55], with a AdaIn-based timestep modulation. We apply attention and MLP blocks in parallel [10], and jointly apply self-attention to the noisy generated and identity conditioning tokens. ControlMLP block (right) is used to provide lightweight spatially-aligned conditioning - Plucker and Spatial Anchor. + +well as their gaze direction. We show an example of our Spatial Anchor in Fig. 2 in the studio on the left. This anchor is color-coded and projected into a 2D image, which is used as conditioning for our target view generations. During inference, the Spatial Anchor can be placed at any given point that lies within the field-of-view of the target cameras. + +Multiview Diffusion Model. Given the above inputs, we train a multiview diffusion model $\epsilon_{\theta}$ to jointly denoise all the target views $\mathbf{y}_{1:N} \in \mathbb{R}^{N \times H \times W \times C}$ with the objective: + +$$ +\mathcal {L} _ {\mathrm {D M}} = \left\| \boldsymbol {\epsilon} ^ {t} - \epsilon_ {\theta} \left(\mathbf {y} _ {1: N} ^ {t}, \mathbf {c} _ {1: N}, \mathbf {x} ^ {\text {r e f}}, \mathbf {x} ^ {\text {f a c e}}, t\right) \right\| ^ {2} \tag {2} +$$ + +where $\mathbf{y}_{1:N}^{t}$ are noisy target images. We condition the base model on the provided reference image and its face crop by concatenating their patchified latent tokens with the noisy input latent tokens to the model. This is shown in Fig. 2 + +Mid-training. In mid-training stage, we want to train a strong multiview model that can denoise several images together, and absorb the dataset quickly at a lower resolution. During this phase, we do not use any pixel-aligned spatial control such as Plücker or Spatial Anchor. We use an MLP to encode the flattened 16-dimensional target camera intrinsics and extrinsics into a single token. We fuse this camera token into each noisy latent token (for the corresponding view) as positional encoding, which makes our multiview model 3D-aware of the target viewpoints. We mid-train our model at $128 \times 128$ resolution to jointly denoise 24 views. + +Post-training. In the post-training stage, our objective is to create a high-resolution model which is 3D-consistent starting with a low-resolution and a 3D-aware (but not consistent) model. For this, we design a lightweight Control- + +Net [93]-inspired module, which takes as input the pixel-aligned Plücker and Spatial Anchor controls, and the denoising timestep to create a separate modulation signal for the multiview model. We name this module ControlMLP, as it uses a single MLP to generate scale-and-shift modulated control for each multiview-DiT block as shown in Fig. 3. Each layer of ControlMLP is zero-initialized at the start. We find that the Post-training phase is crucial in reducing flicker and 3D inconsistencies in generations. We post-train models at $512 \times 512$ and $1024 \times 1024$ resolutions to jointly denoise 10 and 2 views respectively. Increasing the number of views further leads to GPU out-of-memory issues. + +Encoding Plücker and Spatial Anchor. We notice that the relative differences between neighboring pixels in Plücker coordinates are tiny. To amplify these differences better, we use a SIREN [67] layer to first process the 6D grid into a 32D feature grid. Then, we downsample it by $8\mathrm{x}$ to match the size of latent tokens and feed it as input to the ControlMLP. In addition, use the Spatial Anchor to fix the position and orientation of the subject's head in 3D. We only use the Spatial Anchor for generations, and not for the input reference view. We encode the Spatial Anchor image into the latent space of our model via VAE, and concatenate it with Plücker input and pass it through an MLP to create the modulation signal at each layer. + +# 3.3. Understanding and Improving Spatial Control + +This section presents our design choices and examines alternative approaches for injecting pixel-aligned spatial controls during Post-training stage. We demonstrate the effectiveness of spatial control through a focused overfitting experiment with quantitative evaluations in Tab. 1. + +Scene Overfitting Task. We use 160 frames from a fixed 3D scene of a given subject and timestamp, split into 100 training and 60 validation views. We overfit our mid-trained model to the training views while testing various spatial control methods, training only the control modules while keeping other weights frozen. After overfitting for 10K iterations, we evaluate the model on validation views for novel view synthesis. Strong generalization to validation viewpoints indicates effective spatial control and appropriate camera viewpoint sensitivity. Through this task, we evaluate different spatial control injection methods in Tab. 1, starting with simple to advanced modulation designs. + +- No overfitting (Row 1). The Mid-trained model without scene-specific overfitting achieves comparable PSNR of 19.2 and 19.7 on train and validation views respectively. We treat this setup as baseline to improve over. +- Encoding Camera with MLP (Row 2). We encode camera using an MLP similar to prior works [41, 66] and our Mid-training stage (Sec. 3.2). After overfitting, the model achieves slightly better PSNR on training views as expected, however the validation PSNR drops by 1.28 + +
#Method (M2@ 128)PSNRval ↑PSNRtrain ↑
1.Mid-trained (No Overfitting)19.2319.70
2.+ Camera (w/ MLP) [41, 66]17.95-1.2819.92+0.22
3.+ Plücker (w/ MLP) [6, 23, 31, 68]18.89-0.3420.74+1.04
4.+ ControlMLP19.45+0.2229.36+9.66
5.+ SIREN20.13+0.9030.19+10.49
6.+ Spatial Anchor (Ours)22.60+3.3730.49+10.79
+ +Table 1. Evaluating and Designing Spatial Controls (Sec. 3.3). We overfit our multi-view model for 10K iterations on 100 views of a single scene from our Head-only dataset, and evaluate on 60 novel views of this scene by only varying the spatial controls (camera and Spatial Anchor). We use this setup to test the modulation strength of different spatial controls. Subscript values green/red show deviations from Row 1. + +points to 17.95. This suggests that an MLP does not provide enough modulation for camera control. + +- Plücker as Positional Encoding (Row 3). In this setup, we use downsampled and patchified Plücker coordinates processed through MLP to create positional encoding, which is added to the noisy latent tokens. This setup is inspired from prior works [3, 6, 23, 31, 68], and it further improves the validation PSNR compared to MLP at 18.89, but lags behind the non-overfitted baseline. +- Plücker with ControlMLP and SIREN (Row 4, 5). Here, we use our ControlMLP module to inject spatial control at each multiview-DiT block output. Moreover, encoding Plücker coordinate with a SIREN [67] amplifies the relative differences between neighboring pixels (Sec. 3.2). This setup achieves PSNR of 20.13 with an improvement of 0.9 over baseline. +- Adding Spatial Anchor (Row 6). Finally using the Spatial Anchor gives validation PSNR of 22.6 (gain of 3.3 points over baseline) and enables strong spatial control. Thus, we adopt this configuration for Post-training stage. + +# 3.4. Enhanced 3D Consistency Metric + +Traditionally, the 3D consistency of multiview generation models is evaluated using 2D image metrics such as PSNR, LPIPS, and SSIM against a fixed set of ground truth images. However, this approach unfairly penalizes models that generate plausible and 3D-consistent novel content deviating from the fixed ground truth images. Some works try to address this by measuring SfM [17, 60] or epipolar error [45, 89], but these methods solve for pose or are not robust since they measure against the entire epipolar line. To address these limitations, we use our GT camera poses as input and compute Reprojection Error (RE) given our known camera poses and predicted correspondences. Our computation of RE involves the following steps: + +1. Landmarks and Correspondence estimation. We use SuperPoint [13] to detect landmarks in the generated images and employ SuperGlue [59] to establish pairwise correspondences between landmarks across images. + +2. Triangulation. Given the correspondences and camera parameters, we apply Triangulation based on Direct Linear Transformation (DLT) [22] to obtain the corresponding 3D points for each landmark. +3. Reprojection and Error calculation. We reproject these 3D points onto each image and compute the RE as the L2 distance between the original landmark and the reprojected 3D point normalized by image resolution, and average error across all images. + +This approach evaluates multiview generation models by focusing on their ability to produce 3D-consistent results rather than adhering to a fixed ground truth. The Reprojection Error provides a valuable basis for comparison across different methods. Furthermore, by computing RE on a set of real-world images that are independent of our generated images, we can establish a baseline that quantifies the error due to the noisy predictions from SuperGlue and SuperPoint rather than the quality of our generated images. + +Naming Convention (RE@SG). Please note that the use of SuperPoint [13] and SuperGlue [59] estimation modules is a particular instantiation of our metric, and these could be replaced in the future with stronger counterparts such as MAST3R [36, 75] or domain-specific keypoint detectors such as Sapiens [33]. Thus we use the naming convention of RE@SG to denote the Reprojection Error (RE) under SuperGlue (SG) estimation, which could be modified accordingly in the future for different estimators. + +# 4. Experiments + +We present details of the datasets used in all training and validation stages, followed by discussion of evaluation metrics - with particular emphasis on the 3D consistency metric and conclude with core experimental results and ablations. + +# 4.1.Data + +Humans-3B Dataset. We utilize a large proprietary dataset of approximately 3 billion human-centric in-the-wild images for pretraining. We provide more details on data filtering and curation in ?? + +Head and Full-body Studio Datasets. We rely on high-quality proprietary studio captures as our primary source of data for learning 3D consistency. Our model comes in two variants: head-only and full-body, each trained (for Mid-training and Post-training stages) on corresponding datasets. For our full-body model, we utilize a dataset of 861 subjects (811 train, 50 test), nearly 1000 frames per subject. For our head-only model, we use a dataset of 1450 subjects (1400 train, 50 test), nearly 40000 frames per subject. The studio setup is similar to [43], with two capture domes for capturing full-body and head-only high-resolution 4K images with 230 and 160 cameras. + +iPhone Dataset. To evaluate real-world performance, we collect casual images of 50 test subjects in an indoor office environment using an iPhone 13 Pro. We preprocess these images with Sapiens-2B [33] for background segmentation before model inference. This dataset serves exclusively for evaluating our model's performance on in-the-wild inputs. + +# 4.2. Evaluation Setup and Metrics + +3D Consistency. Following prior works, we report standard metrics like PSNR, SSIM ( $\times 100$ ), and LPIPS ( $\times 100$ ) metrics. However, these metrics unfairly penalize plausible novel views generated under incomplete inputs. We therefore introduce the Reprojection Error metric, described in Sec. 3.4, which validates 3D consistency without directly relying on ground truth. Our evaluation generates 4 randomly selected views from the test split. + +Identity Preservation. We use two metrics that measure identity preservation across generated views by computing cosine distance between features extracted via FaceNet [4, 61] for face similarity; and CLIP [52] vision encoder for full-body similarity. + +Pretrained Model Evaluation. We measure the effectiveness of our pre-training strategy by reporting the FID [24]. We use a smaller annotated 30M subset of Humans-3B for training image and text-conditioned P1@128 models, a 30M subset of unfiltered dataset for training No Filtering P1@128, and a 1000 sample test set from iPhone dataset. + +# 4.3. Results + +Pretraining and Data Filtering. Tab. 2 presents our pretrained model results (Row 1) and ablations on Human-centric data filtering and Image-conditioned pretraining (Rows 2-5). Human-centric filtering and image-based conditioning are both critical to achieve high-quality generation. The qualitative results are shown in ?? + +
#Method (Stage @ Resolution)FID ↓
1.Pippo (P1@512)51.164
2.Human-centric Filtering (P1@128)75.639
3.No Filtering (P1@128)86.838
4.Image-conditioned (P1@128)75.639
5.Text-conditioned (P1@128)109.720
+ +Table 2. Pretraining and Data Filtering. We report results of the full pretrained model P1@512, and compare several variants of P1@128 models. We report FID on iPhone dataset (1k samples). + +High-resolution Multi-view Generation. In Tab. 3, we evaluate 3D reconstruction and identity preservation for unseen subjects from the studio datasets. We show that increasing the output resolution of generation in our approach does not hurt 3D consistency or similarity. We put corresponding visuals in ??, Rows 2 and 3. + +
# Split & Resolution3D ConsistencySimilarity
RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
Head-only
1. Studio (128)3.3+0.821.067.614.862.0-13.361.5-16.2
2. Studio (1K)3.4+0.520.372.026.273.5-2.579.4-0.1
3. iPhone Face (1K)3.0---67.6-
Full-body
4. Studio (128)3.6+0.122.884.010.041.7-8.567.1-9.4
5. Studio (1K)1.5+0.122.491.711.164.7-6.074.1-2.2
6. iPhone (1K)1.7---58.068.1
+ +Table 3. Results on Unseen Studio and iPhone data (Sec. 4.3.) We report results on Post-trained Pippo models at three different resolutions. We report 3D metrics PSNR, SSIM, and LPIPS; as well as our proposed Reprojection Error (RE@SG) under SuperGlue estimation which does not require ground truth (Sec. 3.4). We report 2D Face and Body similarities. Red subscript show deviation against ground truth value. + +Generations from Casual iPhone Photos. We present in Tab. 3 (Rows 3,6) Reprojection Error and similarity scores for casually taken images from the iPhone dataset with 1K resolution model. In this scenario, the standard reconstruction error metrics cannot be assessed due to missing ground truth. We find that the reprojection error on iPhone captures remains comparable to the studio dataset - demonstrating 3D consistency. This illustrates the generalizability of Pippo beyond the multi-view training data domain, where our pretraining with large-scale in-the-wild human data is critical. We put corresponding visuals in ??, Row 1 (Appendix). + +Comparisons with External Benchmarks. In Fig. 4, we compare Pippo with individual state-of-the-art baselines in Full-body and Head-only generation. SiTH [25] reconstructs textured human-mesh using ControlNet paired with a SDF representation. Compared to SiTH, Pippo facilitates high-resolution and accurate multiview synthesis. DiffPortrait3D [20] inverts a 3D-GAN based on a given input image. Compared to it, our model supports greater viewpoint variability and ensures closer adherence to the input image. + +Quantitative comparisons and baselines. Existing SoTA Human methods [20, 25] use explicit SMPL priors, and thus are difficult to compare with directly. Qualitatively, we found that they cannot handle novel views or preserve details (Fig. 4), and hence do not quantitatively compare against them. In Pippo, we focus on creating a strong multiview human generator, and we benchmark four state-of-the-art multi-view diffusion models on the iPhone full-body dataset in Tab. 4. We find that Pippo preserves identity (i.e., face and body similarity) and 3D consistency (RE) better while also operating at a higher resolution compared to baselines. + +Benchmarking Pippo on public datasets. In Tab. 5, we benchmark Pippo on the public Codec Avatar datasets [43] to aid future comparisons. Specifically, we use Ava-256 containing 256 head-only captures and Goliath containing 4 + +
#Method & ResolutionRE@SG ↓Face ↑Body ↑
1.MV-Adapter [27] (768)4.743.064.1
2.Era3D [38] (512)4.138.164.4
3.Wonder3D [42] (256)5.334.758.8
4.Pippo (P3@1K)3.058.068.1
+ +Table 4. Quantitative comparison with SoTA multi-view models. We quantitatively compare Pippo against state-of-the-art multi-view diffusion models. We find that Pippo preserves identity (i.e., face and body similarity) and 3D consistency (RE) better while operating at a higher resolution compared to baselines. + +full-body. During evaluation, we only use subsets of these datasets that were not used for training. We find that Pippo's performance on these datasets is inline with its performance on internal studio datasets. + +
#Dataset (P3@1K)RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
1.Ava-256 (Head-only)3.820.168.726.672.976.8
2.Goliath (Full-body)1.220.489.715.587.577.7
+ +Table 5. Benchmarking Pippo on public Ava-256 and Goliath datasets. We find that Pippo's performance on these datasets is inline with its performance on internal studio datasets. This will aid future comparisons against Pippo. + +# 4.4. Ablations + +We examine design choices at each training stage (Sec. 3) and present our ablation results in Tab. 6. Experiments are conducted at $128 \times 128$ resolution on the Head-only dataset. + +
#Method (Stage @ Res.)3D ConsistencySimilarity
RE@SG↓PSNR↑SSIM↑LPIPS↓Face ↑Body ↑
1.Pippo (P3@512)2.7+0.721.771.721.574.1-2.077.4-2.4
2.Pippo Face-only (P3@512)2.4+0.420.570.325.374.3-1.775.8-4.0
3.No Mid-train (P3@512)5.6+3.614.559.544.220.4-55.663.4-16.4
4.Pippo (P3@128)3.3+0.821.067.614.862.0-13.361.5-16.2
5.No Anchor (P3@128)11.5+9.017.154.021.163.1-12.268.3-9.3
6.No Plücker (P3@128)4.2+1.720.264.516.163.5-11.866.1-11.5
7.Pippo (M2@128)3.4+0.919.161.417.260.0-15.362.6-15.0
8.No Pre-train (M2@128)5.9+3.413.139.344.919.5-55.849.0-28.6
9.Cross-Attn (M2@128)3.5+1.018.059.222.166.2-9.170.7-7.0
10.Non-frontal (M2@128)5.8+3.315.252.430.149.2-26.160.7-16.9
+ +Table 6. Ablation on design choices and training stages. We evaluate several multi-view models at $128 \times 128$ and $512 \times 512$ resolution at different training stages on Head-only dataset. We ablate the choice of doing Mid-training and Pre-training, along with different spatial controls and reference conditioning methods..The training stages are labeled as M2 (Mid-training) and P3 (Post-training), followed by @ specified resolution. + +Significance of Pre-training and Mid-training. Pretraining our model on the Humans-3B dataset enables robust generalization to novel identities, as demonstrated in + +![](images/0a1556fb832c03b8b9171a80936dc21cd43442942792f4423be8d34086dead0f.jpg) +Figure 4. Visual comparison with state-of-the-art methods. We visually compare Pippo with state-of-the-art baselines DiffPortrait3D [20] (head-only) and SiTH [25] (full-body) generation. + +Tab. 6, Row 8. Without pre-training, model generalization deteriorates, resulting in unclear facial features. Skipping mid-training at lower resolution impairs consistent multiview generation, as shown in Row 2. + +Importance of Frontal Input Reference. Our ablation in Tab. 6, Row 10 demonstrates that completely randomizing the view point of an input reference image leads to overfitting to training identities. Non-frontal views, especially rear views, have very limited information about the identity which forces the network to pick up spurious correlations. + +
# Pretrain Data (M2@128)RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
1. 30M (1% HQ)3.821.381.813.330.666.2
2. 900M (30% Random)4.021.682.012.130.667.2
3. 2.1B (70% Random)3.721.682.112.137.667.5
4. Humans-3B (100%)3.721.982.312.741.767.2
+ +Table 7. Ablating the size of pre-training dataset on Humans-3B. We conduct Full-body mid-training with pretrained checkpoints trained on $1\%$ , $30\%$ , $70\%$ , and $100\%$ subsets of Humans-3B dataset. We found large-scale data is crucial for generalization to novel identities (i.e., face similarity). + +Importance of Self-attention. Replacing self-attention with cross-attention for reference image encoding, using the same routing as image-conditioned pretraining from Sec. 3.1, leads to degraded performance as shown in Tab. 6, Row 9. We observe that this setup causes the model to ignore input conditioning, generating images that only vaguely resemble training subjects. + +Role of large scale Humans-3B pre-training dataset. We utilized intermediate checkpoints from pre-training stage (P1) trained on $30\%$ and $70\%$ of the data, and a separate checkpoint trained on $1\%$ high-quality subset of Humans-3B. Starting from these checkpoints, we mid-trained Pippo to denoise 4 views at $128 \times 128$ resolution on Full-body dataset for two days. We report their respective results in Tab. 7. We found large-scale data is crucial for general + +ization to novel identities – indicated by high gains in face similarity metric. + +# 5. Conclusion + +We present Pippo, a diffusion transformer model that generates a dense set of high-resolution multi-view consistent images of a person from a single image. Our experiments show that our multi-stage training strategy, which combines largescale in-the-wild data with high-quality multi-view studio captures, enables generalizable high-resolution multi-view synthesis. Analysis of the diffusion transformer architecture reveals that self-attention with the reference image, Plücker coordinates with SIREN, and Spatial Anchor are all essential for high-fidelity multi-view human generation. Pippo achieves, for the first time, consistent multi-view human image generation at 1K resolution. We also show that our proposed 3D consistency metric enables evaluation of 3D consistency without paired ground-truth data. One of the limitations of our approach is the limited number of simultaneously generated views, originating from large context length and memory constraints, which can be tackled with parallelization techniques and autoregressive generation. Extending our approach to multi-view consistent video generation will be addressed in future work. + +# References + +[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panohed: Geometry-aware 3d fullhead synthesis in 360deg. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20950-20959, 2023. 3 +[2] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabián Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM Transactions on Graphics (TOG), 40(4):1-17, 2021. 2 + +[3] Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, et al. Vd3d: Taming large video diffusion transformers for 3d camera control. arXiv preprint arXiv:2407.12781, 2024.5 +[4] Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2018), pages 67-74, 2018. 6 +[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 3 +[6] Eric Ming Chen, Sidhanth Holalkere, Ruyu Yan, Kai Zhang, and Abe Davis. Ray conditioning: Trading photoconsistency for photo-realism in multi-view image generation. arXiv preprint arXiv:2304.13681, 2023. 5 +[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 3 +[8] Wei Cheng, Ruixiang Chen, Wanqi Yin, Siming Fan, Keyu Chen, Honglin He, Huiwen Luo, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, Daxuan Ren, Lei Yang, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, Bo Dai, and Kwan-Yee Lin. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. arXiv preprint, arXiv:2307.10173, 2023. 2 +[9] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 4 +[10] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme Ruiz, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd Van Steenkiste, Gamaleldin Fathy Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmiŋn Bastings, Mark Collier, Alexey A. Gritsenko, Vighnesh Birodkar, Cristina Nader Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lucic, Xiaohua Zhai, Daniel Keysers, Jeremiah J. Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters. In Proceedings of the 40th International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 4 +[11] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 3 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2 +[13] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection + +and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 5, 6 +[14] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In ACM MM, 2022. 3 +[15] Nikita Drobyshev, Antoni Bigata Casademunt, Konstantinos Vougioukas, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Emoportraits: Emotion-enhanced multimodal one-shot head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8498-8507, 2024. 3 +[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 4 +[17] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2, 5 +[18] Matheus Gadelha, Rui Wang, and Subhransu Maji. Multiresolution tree networks for 3d point cloud processing. In ECCV, 2018. 3 +[19] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 3 +[20] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10456-10465, 2024. 3, 7, 8 +[21] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 3 +[22] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 6 +[23] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3, 5 +[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. 2017. 6 +[25] I Ho, Jie Song, Otmar Hilliges, et al. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 538-549, 2024. 7, 8 + +[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 4 +[27] Zehuan Huang, Yuanchen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 7 +[28] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9784-9794, 2024. 3 +[29] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2 +[30] Yash Kant, Aliaksandr Siarohin, Michael Vasilkovsky, Riza Alp Guler, Jian Ren, Sergey Tulyakov, and Igor Gilitschenski. invs: Repurposing diffusion inpainters for novel view synthesis. In SIGGRAPH Asia 2023 Conference Papers, 2023. 3 +[31] Yash Kant, Aliaksandr Siarohin, Ziyi Wu, Michael Vasilkovsky, Guocheng Qian, Jian Ren, Riza Alp Guler, Bernard Ghanem, Sergey Tulyakov, and Igor Gilitschenski. Spad: Spatially aware multi-view diffusers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10026-10038, 2024. 5 +[32] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 3 +[33] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 6 +[34] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 3 +[35] Black Forest Labs. Flux model, 2024. Accessed: 2024-11-10. 4 +[36] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 6 +[37] Heyuan Li, Ce Chen, Tianhao Shi, Yuda Qiu, Sizhe An, Guanying Chen, and Xiaoguang Han. Spherehead: Stable 3d full-head synthesis with spherical tri-plane representation. In European Conference on Computer Vision, pages 324-341. Springer, 2025. 3 +[38] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, Wenping Wang, Qifeng Liu, and Yike Guo. Era3d: High-resolution multiview diffusion using efficient row-wise attention, 2024. 7 +[39] Chieh Hubert Lin, Chia-Che Chang, Yu-Sheng Chen, Da-Cheng Juan, Wei Wei, and Hwann-Tzong Chen. Coco-gan: Generation by parts via conditional coordinating. In ICCV, 2019. 2 + +[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3 +[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 2, 3, 5 +[42] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion, 2023. 7 +[43] Julieta Martinez, Emily Kim, Javier Romero, Timur Bagautdinov, Shunsuke Saito, Shoou-I Yu, Stuart Anderson, Michael Zollhöfer, Te-Li Wang, Shaojie Bai, Chenghui Li, Shih-En Wei, Rohan Joshi, Wyatt Borsos, Tomas Simon, Jason Saragih, Paul Theodosis, Alexander Greene, Anjani Josyula, Silvio Mano Maeta, Andrew I. Jewett, Simon Venshtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mohamed Ezzeldin A. Elshaer, Tingfang Du, Longhua Wu, Shen-Chi Chen, Kai Kang, Michael Wu, Youssef Emad, Steven Longay, Ashley Brewer, Hitesh Shah, James Booth, Taylor Koska, Kayla Haidle, Matt Andromalos, Joanna Hsu, Thomas Dauer, Peter Selednik, Tim Godisart, Scott Ardisson, Matthew Cipperly, Ben Humberston, Lon Farr, Bob Hansen, Peihong Guo, Dave Braun, Steven Krenn, He Wen, Lucas Evans, Natalia Fadeeva, Matthew Stewart, Gabriel Schwartz, Divam Gupta, Gyeongsik Moon, Kaiwen Guo, Yuan Dong, Yichen Xu, Takaaki Shiratori, Fabian Prada, Bernardo R. Pires, Bo Peng, Julia Buffalini, Autumn Trimble, Kevyn McPhail, Melissa Schoeller, and Yaser Sheikh. Codec Avatar Studio: Paired Human Captures for Complete, Driveable, and Generalizable Avatars. NeurIPS Track on Datasets and Benchmarks, 2024. 2, 6, 7 +[44] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 3 +[45] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024. 2, 5 +[46] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4 +[47] Dongwei Pan, Long Zhuo, Jingtan Piao, Huiwen Luo, Wei Cheng, Yuxin Wang, Siming Fan, Shengqi Liu, Lei Yang, Bo Dai, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, and Kwan-Yee Lin. RenderMe-360: Large digital asset library and benchmark towards high-fidelity head avatars. In Thirty-seventh Conference on Neural Information + +Processing Systems Datasets and Benchmarks Track, 2023. 2 +[48] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3 +[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4 +[50] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3 +[51] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 3 +[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 6 +[53] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 3 +[54] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4 +[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4 +[56] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In CVPR, 2019. 3 +[57] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 3 +[58] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 3 +[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 5, 6 +[60] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 5 + +[61] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 4, 6 +[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. 2022. 2 +[63] Akash Sengupta, Thiemo Alldieck, Nikos Kolotouros, Enric Corona, Andrei Zanfir, and Cristian Sminchisescu. Diffhuman: Probabilistic photorealistic 3d reconstruction of humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1439-1449, 2024. 3 +[64] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 3 +[65] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 3 +[66] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 5 +[67] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. 2020. 5 +[68] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. 2021. 5 +[69] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. 2015. 4 +[70] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In European Conference on Computer Vision, pages 107-124. Springer, 2022. 3 +[71] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14795-14805, 2023. 3 +[72] Qingyang Tan, Lin Gao, Yu-Kun Lai, and Shihong Xia. Variational autoencoders for deforming 3d mesh models. In CVPR, 2018. 3 +[73] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. 2023. 3 +[74] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 2023. 3 +[75] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 6 + +[76] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 3 +[77] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[78] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. 2023. 3 +[79] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[80] Jing Wen, Xiaoming Zhao, Zhongzheng Ren, Alexander G Schwing, and Shenlong Wang. Gomavatar: Efficient animatable human modeling from monocular video using gaussians-on-mesh. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2059-2069, 2024. 3 +[81] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 3 +[82] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. 2016. 3 +[83] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Song-Chun Zhu, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In CVPR, 2018. 3 +[84] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, et al. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19801-19811, 2024. 2 +[85] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 3 +[86] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024.3 +[87] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. FaceScape: A largescale high quality 3D face dataset and detailed riggable 3D + +face prediction. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2 +[88] T Yenamandra, A Tewari, F Bernard, HP Seidel, M Elgharib, D Cremers, and C Theobalt. i3DMM: Deep implicit 3D morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[89] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2, 5 +[90] Zhixuan Yu, Jae Shin Yoon, In Kyu Lee, Prashanth Venkatesh, Jaesik Park, Jihun Yu, and Hyun Soo Park. HUMBI: A large multiview dataset of human body expressions. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2 +[91] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 4 +[92] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22096-22105, 2023. 3 +[93] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3, 5 +[94] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3 \ No newline at end of file diff --git a/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/images.zip b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cd490a556472c485e45d98a13f6e6e6642e474e4 --- /dev/null +++ b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18c33065d38032c31774a63a87856db698d9234425d8d2b4c74217315d673ba7 +size 516576 diff --git a/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/layout.json b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a96330097ff39c54307553f16321f8c0046a37d6 --- /dev/null +++ b/2025/Pippo_ High-Resolution Multi-View Humans from a Single Image/layout.json @@ -0,0 +1,9036 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "type": "text", + "content": "Pippo: High-Resolution Multi-View Humans from a Single Image" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "spans": [ + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": "Yash Kant" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": ", Ethan Weber" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": ", Jin Kyu Kim" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": ", Rawal Khirodkar" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": ", Su Zhaoen" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": ", Julieta Martinez" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": ", Igor Gilitschenski" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{*,2,3}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": ", Shunsuke Saito" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{*,1}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": ", Timur Bagautdinov" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{*,1}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": "Meta Reality Labs, " + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": "University of Toronto, " + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": "Vector Institute, " + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 71, + 142, + 539, + 220 + ], + "type": "text", + "content": "UC Berkeley https://yashkant.github.io/pippo/ \n*Joint Advising" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 227, + 553, + 567 + ], + "blocks": [ + { + "bbox": [ + 56, + 227, + 553, + 567 + ], + "lines": [ + { + "bbox": [ + 56, + 227, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 553, + 567 + ], + "type": "image", + "image_path": "c5cebda302024c92c2938cc7a88a490127e4df0e93ba112539057ae62f0db3a8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 576, + 555, + 609 + ], + "lines": [ + { + "bbox": [ + 55, + 576, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 555, + 609 + ], + "type": "text", + "content": "Figure 1. Pippo produces high-resolution, multi-view, studio-quality images from a single image. First row shows full-body generation of an unseen subject captured from an iPhone picture. Second and third rows show face-only and partial full-body input images from unseen studio subject(s) to novel views respectively. All visuals are at 1K resolution." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 629, + 200, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 629, + 200, + 642 + ], + "spans": [ + { + "bbox": [ + 152, + 629, + 200, + 642 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 666, + 294, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 666, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 666, + 294, + 715 + ], + "type": "text", + "content": "We present Pippo, a generative model capable of producing 1K resolution dense turnaround videos of a person from a single casually clicked photo. Pippo is a multi-view diffusion transformer and does not require any additional" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 631, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 555, + 703 + ], + "type": "text", + "content": "inputs — e.g., a fitted parametric model or camera parameters of the input image. We pre-train Pippo on 3B human images without captions, and conduct multi-view mid-training and post-training on studio captured humans. During mid-training, to quickly absorb the studio dataset, we denoise several (upto 48) views at low-resolution, and encode tar" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16418" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": "get cameras coarsely using a shallow MLP. During posttraining, we denoise fewer views at high-resolution and use pixel-aligned controls (e.g., Spatial anchor and Plucker rays) to enable 3D consistent generations. At inference, we propose an attention biasing technique that allows Pippo to simultaneously generate greater than " + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": " as many views as seen during training. Finally, we also introduce an improved metric to evaluate 3D consistency of multi-view generations, and show that Pippo outperforms existing works on multi-view human generation from a single image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 220, + 136, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 136, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 136, + 232 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 295, + 374 + ], + "type": "text", + "content": "Creating photorealistic human representations with the ability to control viewpoints has numerous applications in entertainment, healthcare, fashion and social media. Building such representations, first and foremost, requires high-quality multi-view studio data [2, 29], which is costly to acquire. This significantly limits the scalability in terms of the number identities for high-quality studio data [29, 43, 84]. In contrast, large-scale, unstructured, and diverse human images and videos are available online. However, the raw data of such in-the-wild images does not offer ground-truth 3D or multi-view representations of humans." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 556 + ], + "type": "text", + "content": "In this work, we present a novel approach leveraging the best of two worlds: generalizability from in-the-wild unstructured images, and fidelity and view-controllability from studio capture data. Specifically, our model Pippo is a diffusion transformer, which can generate several 1K resolution multi-view consistent images jointly during inference. Pippo takes as input a single image of an individual, camera poses of the target viewpoints to be generated. Since the scale and placement of the subject is ambiguous from a single image, Pippo uses a Spatial Anchor which roughly specifies the location and orientation of the subject in 3D space. Our model does not rely on additional conditioning such as body priors or camera parameters of the input images, to scale our training pipeline to in-the-wild data and support unconstrained inputs at test time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "We employ a multi-stage training recipe to train Pippo. First, we pre-train the model for latent-to-image generation task, similar to [54], on a large human-centric dataset of in-the-wild images. Next, in Mid-training stage, we jointly generate multiple consistent images of the subject, conditioned on target viewpoints and a single input image using high-quality studio dataset. Finally, in Post-training stage, the model is provided with a minimal placement signal - spatial anchor - encoding rough head orientation - which further improves 3D consistency. Our architecture is also carefully tailored to the conditional multi-view generation - we propose several simple but effective modifications to the basic DiT architecture, including self-attention-based" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "conditioning, lightweight spatial controls and camera conditioning with Plucker coordinates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 97, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 337 + ], + "type": "text", + "content": "For evaluating such a multi-view diffusion model, 3D consistency is a critical metric for understanding the geometric correctness of the generation. For generation task in an ill-posed setup (e.g., a single image input), there may be multiple possibilities that are equally plausible. Existing multi-view generation methods typically report reconstruction metrics (i.e., PSNR, SSIM and LPIPS) [19] or FID [41], which either (1) penalize any new content that is actually 3D consistent or (2) are unable to measure the 3D consistency of the generation. To address this problem, we design a metric measuring 3D consistency: we compute 2D keypoint matches with an off-the-shelf method [59], triangulate them, and then reproject back into the other views to measure the reprojection error in pixels. Some generation methods measure reprojection errors [17] from SfM [60] or epipolar distance [45, 89], but our metric uses camera pose as input and is more precise than measuring distance to a epipolar line. We show our metric helps to quantify our results, finding that our method is favorable compared with existing approaches and other baseline methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 338, + 528, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 528, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 528, + 350 + ], + "type": "text", + "content": "To summarize, we make the following contributions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 351, + 553, + 434 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 351, + 553, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 351, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 315, + 351, + 553, + 387 + ], + "type": "text", + "content": "- a generative model capable of generating high-resolution and multi-view consistent humans from a single image and its effective training strategy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 387, + 553, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 387, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 315, + 387, + 553, + 410 + ], + "type": "text", + "content": "- a diffusion transformer architecture designed to enhance multi-view generation and viewpoint control." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 411, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 411, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 315, + 411, + 553, + 434 + ], + "type": "text", + "content": "- a novel 3D consistency metric tailored to accurately measure the level of 3D consistency in generative tasks." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 449, + 400, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 449, + 400, + 460 + ], + "spans": [ + { + "bbox": [ + 314, + 449, + 400, + 460 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 470, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 554 + ], + "type": "text", + "content": "We review multi-view human datasets and generative models for human synthesis, categorizing methods by their data requirements and prior constraints—such as parametric human models or explicit 3D structures. In this work, we minimize reliance on complex priors by training the model on large amounts of human-centric data and impose minimal camera and spatial controls." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Multi-view Human Datasets. While many large 2D human datasets exist [12, 39, 62], 3d captures of people (i.e., captures using dozens of views at a time) are still relatively rare, since they are expensive to collect and are thus mostly obtained in specialized research labs. Nevertheless, this kind of data has recently become more common, going from hundreds to thousands of publicly-available captures. For face-only captures, a few datasets [43, 47, 87, 88, 90] provide up to 600 subjects captured from over 60 views each, while for full-body captures, 4D-Dress [77] provides 32 subjects with 2 outfits each, and detailed annotations for their garments, while DNA-rendering [8] provides 500 subjects from 60 views. MVHumanNet [84] stands apart by" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16419" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 541, + 223 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 541, + 223 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 541, + 223 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 541, + 223 + ], + "type": "image", + "image_path": "274551e9bf4dd5a03572aa716c816b4fd81239c27280c6a69e5ac8d6d24eb537.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "type": "text", + "content": "Figure 2. Pipeline overview. This is an illustration of how we train our model. (Left) we use data from a studio capture and train our multi-view diffusion model (right). We condition on a full reference photo and a cropped face, as well as the target view cameras and 2D projected spatial anchor indicating head position and orientation. Our diffusion model also takes in noisy target views and a timestep in order to predict the denoised views (top). In practice, we apply a segmentation mask around the person." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 297, + 295, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 295, + 370 + ], + "type": "text", + "content": "providing 4,500 people captured from 48 views in and over 9,000 sets of clothing, outpacing other datasets by an order of magnitude in terms of people diversity. We use internal multi-view data of " + }, + { + "bbox": [ + 54, + 297, + 295, + 370 + ], + "type": "inline_equation", + "content": "\\sim 1000" + }, + { + "bbox": [ + 54, + 297, + 295, + 370 + ], + "type": "text", + "content": " identities with dense (" + }, + { + "bbox": [ + 54, + 297, + 295, + 370 + ], + "type": "inline_equation", + "content": "\\sim 160" + }, + { + "bbox": [ + 54, + 297, + 295, + 370 + ], + "type": "text", + "content": ") camera setup, but we expect our model to produce reasonable results with publicly-available captures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 375, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 295, + 626 + ], + "type": "text", + "content": "Generating Novel Poses and Expressions. Although different from our task, this related work builds 3D animatable models of people and faces. These methods use neural radiance fields [70, 71, 81], or 3d Gaussians [34, 51, 80, 94], using canonical \"T\"-poses and a corresponding forward and backward mapping to model local appearance changes. Many of these methods, however, rely on accurate human shape and pose estimation, and only achieve high photorealism for personalized models from studio captures. Relatedly, a family of methods focus on generating animatable faces from a single image by disentangling viewpoints and expressions from large collections of 2d portraits [14, 15, 32, 48, 86, 92], allowing photorealistic real-time reenactment, while being limited to faces and small viewpoint variations. Unlike these methods, we complete missing parts of a person given either one or a few partial views. We do not repose the person or animate the faces but can instead recover entirely missing views, which is orthogonal to these methods. Moreover, our approach addresses both full-body and facial reconstruction, rather than specializing in either domain." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "text", + "content": "Generations with Explicit 3D Structures. Similar to early methods that extracted 3d representations from 2d models via score distillation [7, 11, 40, 50, 53, 74, 78], 3d representations of human appearance have been extracted from GANs by co-training a neural renderer from learned TriPlane [5] or HexPlane [1, 37] latent spaces. Others train generative models that operate directly in 3D space" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 297, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 393 + ], + "type": "text", + "content": "with diffusion or GAN objectives [18, 44, 72, 76, 82, 83] or regression models to directly predict 3D representations [56, 57, 63]. While these methods produce 3d consistent faces by design, their quality is limited by the relatively small amounts of 3d training data and/or the limitations of their respective 3d representations. In contrast, we focus on generating 3d-consistent 2d images and thus avoid the downsides of explicit 3d modelling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 395, + 556, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 556, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 556, + 563 + ], + "type": "text", + "content": "Generations with 2D Models. Another option is to model viewpoint changes in 2d, without an underlying 3d representation [41], sometimes modelling either photometric [30] or epipolar [28] constraints explicitly. Diff-Portrait3D [20] is a portrait-generation method that falls roughly within this category, by fine-tuning a 2d diffusion model for 3d-aware face generation using ControlNet-style [93] conditioning for viewpoint control, as well as cross-view attention [21] and initialization from a 3d-consistent GAN [5]. Other methods are more general and not specific to humans, focusing on single-view to 3D generation [19, 41, 58, 64, 65, 73, 85]. In addition to these single-view methods, video models have been fine-tuned for camera control [23, 79]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 578, + 370, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 578, + 370, + 590 + ], + "spans": [ + { + "bbox": [ + 314, + 578, + 370, + 590 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 600, + 529, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 529, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 529, + 613 + ], + "type": "text", + "content": "We train our models following a three-stage strategy:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 617, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 641 + ], + "type": "text", + "content": "- Image-only Pre-training (P1). We pretrain on a large-scale human-centric dataset with image conditioning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 677 + ], + "type": "text", + "content": "- Multiview Mid-training (M2). We train models at a low-resolution of " + }, + { + "bbox": [ + 313, + 642, + 553, + 677 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 642, + 553, + 677 + ], + "type": "text", + "content": " to denoise 48 target views with coarse camera control (no pixel-aligned spatial control)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "- Multiview Post-training (P3). We train at a high-resolution of " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " to denoise " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "1 - 3" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " target views with spatial control injected via ControlMLP layers." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16420" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "We denote the training stage and resolution of any given model as " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{stage}\\} @\\{\\mathrm{resolution}\\}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": ". For instance, M2@128 represents a mid-trained model at 128 resolution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 116, + 133, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 116, + 133, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 116, + 133, + 126 + ], + "type": "text", + "content": "3.1. Base Model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": "Architecture. We adopt a DiT-like [49] architecture with scale, shift, and gate modulation for timestep conditioning inspired by Stable Diffusion 3 [16] and Flux [35]. We simplify the architecture by employing MLP and attention in parallel [91], and removing second LayerNorm after attention layers. We use VAE for training in latent space with 8x spatial compression, and patchify latent images via a linear layer and patch size of 2. We use fixed sinusoidal positional encoding during training. We provide more details in ??, and the exact design of our DiT block in Fig. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 253, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 372 + ], + "type": "text", + "content": "Image-only Pre-training. During pre-training, the model learns to denoise an image conditioned on its corresponding image embedding from DINOv2 [9, 46], this is similar in principle to the image decoder of DALL-E 2 [54]. We project both embeddings to the model dimension using with a linear layer to create a joint conditioning. Importantly, our pretraining setup does not require any annotations or captions for the images, and is well-aligned with our downstream objective of generating consistent multiview images given a single reference image as input." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "text", + "content": "Formally, given an image " + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "text", + "content": ", and the joint conditioning as " + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{\\mathrm{img}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "text", + "content": ". We pre-train our diffusion model " + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 372, + 295, + 409 + ], + "type": "text", + "content": " with the following objective:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 417, + 295, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 417, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 111, + 417, + 295, + 431 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D M}} = \\left| \\left| \\boldsymbol {\\epsilon} ^ {t} - \\epsilon_ {\\theta} \\left(\\mathbf {y} ^ {t}, \\mathbf {e} ^ {\\mathrm {i m g}}, t\\right) \\right| \\right| ^ {2} \\tag {1}", + "image_path": "70b1957c196ab1c532e8ef47e1a47d084339a7a6124f152665a128c6a0fb9b4c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "t \\in [0, T]" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": " is diffusion timestep, " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\epsilon^t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": " is the noise added at the given timestep. We use the DDPM [26, 69] formulation to define discrete timesteps and set " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": ". We first pretrain our model at " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": " and then at " + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 440, + 296, + 513 + ], + "type": "text", + "content": " on a large corpus of human-centric images. Exact training details can be found in ??" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 520, + 159, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 159, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 159, + 532 + ], + "type": "text", + "content": "3.2. Multiview Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "type": "text", + "content": "Our goal is to generate many high resolution and unseen novel viewpoints of a human subject (akin to a studio capture) given a single input image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 575, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 295, + 624 + ], + "type": "text", + "content": "Input Reference. We denote the input image as " + }, + { + "bbox": [ + 55, + 575, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 575, + 295, + 624 + ], + "type": "text", + "content": " and corresponding face crop as " + }, + { + "bbox": [ + 55, + 575, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{face}}" + }, + { + "bbox": [ + 55, + 575, + 295, + 624 + ], + "type": "text", + "content": ". The face crop is obtained using FaceNet [61] and resized to the same size as input image, such that: " + }, + { + "bbox": [ + 55, + 575, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{face}}, \\mathbf{x}^{\\mathrm{ref}} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 575, + 295, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 627, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 295, + 675 + ], + "type": "text", + "content": "Target Cameras. We denote the target viewpoints to be synthesized using distinct cameras (intrinsic and extrinsic), represented as " + }, + { + "bbox": [ + 55, + 627, + 295, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{1:N}" + }, + { + "bbox": [ + 55, + 627, + 295, + 675 + ], + "type": "text", + "content": ". Each camera is used to generate its Plücker coordinates " + }, + { + "bbox": [ + 55, + 627, + 295, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i\\in \\mathbb{R}^{H\\times W\\times 6}" + }, + { + "bbox": [ + 55, + 627, + 295, + 675 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Target Spatial Anchor. In addition to target cameras, we provide an oriented 3D point denoted as " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i = [\\mathbf{R}_i|\\mathbf{t}_i]" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", which roughly defines the center of the subject's head, as" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 76, + 535, + 262 + ], + "blocks": [ + { + "bbox": [ + 328, + 76, + 535, + 262 + ], + "lines": [ + { + "bbox": [ + 328, + 76, + 535, + 262 + ], + "spans": [ + { + "bbox": [ + 328, + 76, + 535, + 262 + ], + "type": "image", + "image_path": "a9c12939a1366ca61e3051df02041f892cb9bd5c4751d1a1f106a75a05c4bc3c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 265, + 554, + 332 + ], + "lines": [ + { + "bbox": [ + 313, + 265, + 554, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 554, + 332 + ], + "type": "text", + "content": "Figure 3. DiT and ControlMLP Block. Our DiT block (left) loosely follows [55], with a AdaIn-based timestep modulation. We apply attention and MLP blocks in parallel [10], and jointly apply self-attention to the noisy generated and identity conditioning tokens. ControlMLP block (right) is used to provide lightweight spatially-aligned conditioning - Plucker and Spatial Anchor." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 338, + 554, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 410 + ], + "type": "text", + "content": "well as their gaze direction. We show an example of our Spatial Anchor in Fig. 2 in the studio on the left. This anchor is color-coded and projected into a 2D image, which is used as conditioning for our target view generations. During inference, the Spatial Anchor can be placed at any given point that lies within the field-of-view of the target cameras." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 412, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 449 + ], + "type": "text", + "content": "Multiview Diffusion Model. Given the above inputs, we train a multiview diffusion model " + }, + { + "bbox": [ + 313, + 412, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 412, + 554, + 449 + ], + "type": "text", + "content": " to jointly denoise all the target views " + }, + { + "bbox": [ + 313, + 412, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{1:N} \\in \\mathbb{R}^{N \\times H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 412, + 554, + 449 + ], + "type": "text", + "content": " with the objective:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 457, + 553, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 457, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 343, + 457, + 553, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D M}} = \\left\\| \\boldsymbol {\\epsilon} ^ {t} - \\epsilon_ {\\theta} \\left(\\mathbf {y} _ {1: N} ^ {t}, \\mathbf {c} _ {1: N}, \\mathbf {x} ^ {\\text {r e f}}, \\mathbf {x} ^ {\\text {f a c e}}, t\\right) \\right\\| ^ {2} \\tag {2}", + "image_path": "b99c49313a76a8a4df616f766605ea48581634c2bfe847aab4337ea8d168b97d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 480, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 480, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{1:N}^{t}" + }, + { + "bbox": [ + 313, + 480, + 554, + 529 + ], + "type": "text", + "content": " are noisy target images. We condition the base model on the provided reference image and its face crop by concatenating their patchified latent tokens with the noisy input latent tokens to the model. This is shown in Fig. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 531, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 663 + ], + "type": "text", + "content": "Mid-training. In mid-training stage, we want to train a strong multiview model that can denoise several images together, and absorb the dataset quickly at a lower resolution. During this phase, we do not use any pixel-aligned spatial control such as Plücker or Spatial Anchor. We use an MLP to encode the flattened 16-dimensional target camera intrinsics and extrinsics into a single token. We fuse this camera token into each noisy latent token (for the corresponding view) as positional encoding, which makes our multiview model 3D-aware of the target viewpoints. We mid-train our model at " + }, + { + "bbox": [ + 313, + 531, + 555, + 663 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 531, + 555, + 663 + ], + "type": "text", + "content": " resolution to jointly denoise 24 views." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Post-training. In the post-training stage, our objective is to create a high-resolution model which is 3D-consistent starting with a low-resolution and a 3D-aware (but not consistent) model. For this, we design a lightweight Control-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16421" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "Net [93]-inspired module, which takes as input the pixel-aligned Plücker and Spatial Anchor controls, and the denoising timestep to create a separate modulation signal for the multiview model. We name this module ControlMLP, as it uses a single MLP to generate scale-and-shift modulated control for each multiview-DiT block as shown in Fig. 3. Each layer of ControlMLP is zero-initialized at the start. We find that the Post-training phase is crucial in reducing flicker and 3D inconsistencies in generations. We post-train models at " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": " resolutions to jointly denoise 10 and 2 views respectively. Increasing the number of views further leads to GPU out-of-memory issues." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 218, + 295, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 375 + ], + "type": "text", + "content": "Encoding Plücker and Spatial Anchor. We notice that the relative differences between neighboring pixels in Plücker coordinates are tiny. To amplify these differences better, we use a SIREN [67] layer to first process the 6D grid into a 32D feature grid. Then, we downsample it by " + }, + { + "bbox": [ + 55, + 218, + 295, + 375 + ], + "type": "inline_equation", + "content": "8\\mathrm{x}" + }, + { + "bbox": [ + 55, + 218, + 295, + 375 + ], + "type": "text", + "content": " to match the size of latent tokens and feed it as input to the ControlMLP. In addition, use the Spatial Anchor to fix the position and orientation of the subject's head in 3D. We only use the Spatial Anchor for generations, and not for the input reference view. We encode the Spatial Anchor image into the latent space of our model via VAE, and concatenate it with Plücker input and pass it through an MLP to create the modulation signal at each layer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 380, + 295, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 394 + ], + "type": "text", + "content": "3.3. Understanding and Improving Spatial Control" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 398, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 458 + ], + "type": "text", + "content": "This section presents our design choices and examines alternative approaches for injecting pixel-aligned spatial controls during Post-training stage. We demonstrate the effectiveness of spatial control through a focused overfitting experiment with quantitative evaluations in Tab. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 605 + ], + "type": "text", + "content": "Scene Overfitting Task. We use 160 frames from a fixed 3D scene of a given subject and timestamp, split into 100 training and 60 validation views. We overfit our mid-trained model to the training views while testing various spatial control methods, training only the control modules while keeping other weights frozen. After overfitting for 10K iterations, we evaluate the model on validation views for novel view synthesis. Strong generalization to validation viewpoints indicates effective spatial control and appropriate camera viewpoint sensitivity. Through this task, we evaluate different spatial control injection methods in Tab. 1, starting with simple to advanced modulation designs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 654 + ], + "type": "text", + "content": "- No overfitting (Row 1). The Mid-trained model without scene-specific overfitting achieves comparable PSNR of 19.2 and 19.7 on train and validation views respectively. We treat this setup as baseline to improve over." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "- Encoding Camera with MLP (Row 2). We encode camera using an MLP similar to prior works [41, 66] and our Mid-training stage (Sec. 3.2). After overfitting, the model achieves slightly better PSNR on training views as expected, however the validation PSNR drops by 1.28" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 331, + 71, + 539, + 150 + ], + "blocks": [ + { + "bbox": [ + 331, + 71, + 539, + 150 + ], + "lines": [ + { + "bbox": [ + 331, + 71, + 539, + 150 + ], + "spans": [ + { + "bbox": [ + 331, + 71, + 539, + 150 + ], + "type": "table", + "html": "
#Method (M2@ 128)PSNRval ↑PSNRtrain ↑
1.Mid-trained (No Overfitting)19.2319.70
2.+ Camera (w/ MLP) [41, 66]17.95-1.2819.92+0.22
3.+ Plücker (w/ MLP) [6, 23, 31, 68]18.89-0.3420.74+1.04
4.+ ControlMLP19.45+0.2229.36+9.66
5.+ SIREN20.13+0.9030.19+10.49
6.+ Spatial Anchor (Ours)22.60+3.3730.49+10.79
", + "image_path": "623166a5a6e79e45848ec54594c42e21a9fb62a418aaaef0ddcdd98907d6712c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 158, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 236 + ], + "type": "text", + "content": "Table 1. Evaluating and Designing Spatial Controls (Sec. 3.3). We overfit our multi-view model for 10K iterations on 100 views of a single scene from our Head-only dataset, and evaluate on 60 novel views of this scene by only varying the spatial controls (camera and Spatial Anchor). We use this setup to test the modulation strength of different spatial controls. Subscript values green/red show deviations from Row 1." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 322, + 247, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 247, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 322, + 247, + 553, + 270 + ], + "type": "text", + "content": "points to 17.95. This suggests that an MLP does not provide enough modulation for camera control." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 270, + 553, + 486 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 270, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 270, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 314, + 270, + 553, + 353 + ], + "type": "text", + "content": "- Plücker as Positional Encoding (Row 3). In this setup, we use downsampled and patchified Plücker coordinates processed through MLP to create positional encoding, which is added to the noisy latent tokens. This setup is inspired from prior works [3, 6, 23, 31, 68], and it further improves the validation PSNR compared to MLP at 18.89, but lags behind the non-overfitted baseline." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 354, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 354, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 314, + 354, + 553, + 437 + ], + "type": "text", + "content": "- Plücker with ControlMLP and SIREN (Row 4, 5). Here, we use our ControlMLP module to inject spatial control at each multiview-DiT block output. Moreover, encoding Plücker coordinate with a SIREN [67] amplifies the relative differences between neighboring pixels (Sec. 3.2). This setup achieves PSNR of 20.13 with an improvement of 0.9 over baseline." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 437, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 437, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 314, + 437, + 553, + 486 + ], + "type": "text", + "content": "- Adding Spatial Anchor (Row 6). Finally using the Spatial Anchor gives validation PSNR of 22.6 (gain of 3.3 points over baseline) and enables strong spatial control. Thus, we adopt this configuration for Post-training stage." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 493, + 493, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 493, + 493, + 506 + ], + "spans": [ + { + "bbox": [ + 314, + 493, + 493, + 506 + ], + "type": "text", + "content": "3.4. Enhanced 3D Consistency Metric" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 666 + ], + "type": "text", + "content": "Traditionally, the 3D consistency of multiview generation models is evaluated using 2D image metrics such as PSNR, LPIPS, and SSIM against a fixed set of ground truth images. However, this approach unfairly penalizes models that generate plausible and 3D-consistent novel content deviating from the fixed ground truth images. Some works try to address this by measuring SfM [17, 60] or epipolar error [45, 89], but these methods solve for pose or are not robust since they measure against the entire epipolar line. To address these limitations, we use our GT camera poses as input and compute Reprojection Error (RE) given our known camera poses and predicted correspondences. Our computation of RE involves the following steps:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 666, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 666, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 666, + 553, + 714 + ], + "type": "text", + "content": "1. Landmarks and Correspondence estimation. We use SuperPoint [13] to detect landmarks in the generated images and employ SuperGlue [59] to establish pairwise correspondences between landmarks across images." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16422" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "2. Triangulation. Given the correspondences and camera parameters, we apply Triangulation based on Direct Linear Transformation (DLT) [22] to obtain the corresponding 3D points for each landmark." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 294, + 179 + ], + "type": "text", + "content": "3. Reprojection and Error calculation. We reproject these 3D points onto each image and compute the RE as the L2 distance between the original landmark and the reprojected 3D point normalized by image resolution, and average error across all images." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 182, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 289 + ], + "type": "text", + "content": "This approach evaluates multiview generation models by focusing on their ability to produce 3D-consistent results rather than adhering to a fixed ground truth. The Reprojection Error provides a valuable basis for comparison across different methods. Furthermore, by computing RE on a set of real-world images that are independent of our generated images, we can establish a baseline that quantifies the error due to the noisy predictions from SuperGlue and SuperPoint rather than the quality of our generated images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 291, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 399 + ], + "type": "text", + "content": "Naming Convention (RE@SG). Please note that the use of SuperPoint [13] and SuperGlue [59] estimation modules is a particular instantiation of our metric, and these could be replaced in the future with stronger counterparts such as MAST3R [36, 75] or domain-specific keypoint detectors such as Sapiens [33]. Thus we use the naming convention of RE@SG to denote the Reprojection Error (RE) under SuperGlue (SG) estimation, which could be modified accordingly in the future for different estimators." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 136, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 136, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 136, + 428 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 485 + ], + "type": "text", + "content": "We present details of the datasets used in all training and validation stages, followed by discussion of evaluation metrics - with particular emphasis on the 3D consistency metric and conclude with core experimental results and ablations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 100, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 100, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 100, + 510 + ], + "type": "text", + "content": "4.1.Data" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 565 + ], + "type": "text", + "content": "Humans-3B Dataset. We utilize a large proprietary dataset of approximately 3 billion human-centric in-the-wild images for pretraining. We provide more details on data filtering and curation in ??" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Head and Full-body Studio Datasets. We rely on high-quality proprietary studio captures as our primary source of data for learning 3D consistency. Our model comes in two variants: head-only and full-body, each trained (for Mid-training and Post-training stages) on corresponding datasets. For our full-body model, we utilize a dataset of 861 subjects (811 train, 50 test), nearly 1000 frames per subject. For our head-only model, we use a dataset of 1450 subjects (1400 train, 50 test), nearly 40000 frames per subject. The studio setup is similar to [43], with two capture domes for capturing full-body and head-only high-resolution 4K images with 230 and 160 cameras." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "iPhone Dataset. To evaluate real-world performance, we collect casual images of 50 test subjects in an indoor office environment using an iPhone 13 Pro. We preprocess these images with Sapiens-2B [33] for background segmentation before model inference. This dataset serves exclusively for evaluating our model's performance on in-the-wild inputs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 154, + 477, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 477, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 477, + 166 + ], + "type": "text", + "content": "4.2. Evaluation Setup and Metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 171, + 553, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 553, + 267 + ], + "type": "text", + "content": "3D Consistency. Following prior works, we report standard metrics like PSNR, SSIM (" + }, + { + "bbox": [ + 313, + 171, + 553, + 267 + ], + "type": "inline_equation", + "content": "\\times 100" + }, + { + "bbox": [ + 313, + 171, + 553, + 267 + ], + "type": "text", + "content": "), and LPIPS (" + }, + { + "bbox": [ + 313, + 171, + 553, + 267 + ], + "type": "inline_equation", + "content": "\\times 100" + }, + { + "bbox": [ + 313, + 171, + 553, + 267 + ], + "type": "text", + "content": ") metrics. However, these metrics unfairly penalize plausible novel views generated under incomplete inputs. We therefore introduce the Reprojection Error metric, described in Sec. 3.4, which validates 3D consistency without directly relying on ground truth. Our evaluation generates 4 randomly selected views from the test split." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 270, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 331 + ], + "type": "text", + "content": "Identity Preservation. We use two metrics that measure identity preservation across generated views by computing cosine distance between features extracted via FaceNet [4, 61] for face similarity; and CLIP [52] vision encoder for full-body similarity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 334, + 553, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 553, + 406 + ], + "type": "text", + "content": "Pretrained Model Evaluation. We measure the effectiveness of our pre-training strategy by reporting the FID [24]. We use a smaller annotated 30M subset of Humans-3B for training image and text-conditioned P1@128 models, a 30M subset of unfiltered dataset for training No Filtering P1@128, and a 1000 sample test set from iPhone dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 415, + 371, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 371, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 371, + 427 + ], + "type": "text", + "content": "4.3. Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 434, + 553, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 553, + 506 + ], + "type": "text", + "content": "Pretraining and Data Filtering. Tab. 2 presents our pretrained model results (Row 1) and ablations on Human-centric data filtering and Image-conditioned pretraining (Rows 2-5). Human-centric filtering and image-based conditioning are both critical to achieve high-quality generation. The qualitative results are shown in ??" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 354, + 517, + 515, + 593 + ], + "blocks": [ + { + "bbox": [ + 354, + 517, + 515, + 593 + ], + "lines": [ + { + "bbox": [ + 354, + 517, + 515, + 593 + ], + "spans": [ + { + "bbox": [ + 354, + 517, + 515, + 593 + ], + "type": "table", + "html": "
#Method (Stage @ Resolution)FID ↓
1.Pippo (P1@512)51.164
2.Human-centric Filtering (P1@128)75.639
3.No Filtering (P1@128)86.838
4.Image-conditioned (P1@128)75.639
5.Text-conditioned (P1@128)109.720
", + "image_path": "04edfcfda9dd5d56935de8d60fe8fd6ae881072f0bda819771f15c84a5583f58.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 601, + 555, + 634 + ], + "lines": [ + { + "bbox": [ + 313, + 601, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 555, + 634 + ], + "type": "text", + "content": "Table 2. Pretraining and Data Filtering. We report results of the full pretrained model P1@512, and compare several variants of P1@128 models. We report FID on iPhone dataset (1k samples)." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "High-resolution Multi-view Generation. In Tab. 3, we evaluate 3D reconstruction and identity preservation for unseen subjects from the studio datasets. We show that increasing the output resolution of generation in our approach does not hurt 3D consistency or similarity. We put corresponding visuals in ??, Rows 2 and 3." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16423" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 292, + 163 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 292, + 163 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 292, + 163 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 292, + 163 + ], + "type": "table", + "html": "
# Split & Resolution3D ConsistencySimilarity
RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
Head-only
1. Studio (128)3.3+0.821.067.614.862.0-13.361.5-16.2
2. Studio (1K)3.4+0.520.372.026.273.5-2.579.4-0.1
3. iPhone Face (1K)3.0---67.6-
Full-body
4. Studio (128)3.6+0.122.884.010.041.7-8.567.1-9.4
5. Studio (1K)1.5+0.122.491.711.164.7-6.074.1-2.2
6. iPhone (1K)1.7---58.068.1
", + "image_path": "c9adac0176b9e1c4121c417c4796215c96fc0a30364aa36d0d1f106b3bb157ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 171, + 294, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 294, + 248 + ], + "type": "text", + "content": "Table 3. Results on Unseen Studio and iPhone data (Sec. 4.3.) We report results on Post-trained Pippo models at three different resolutions. We report 3D metrics PSNR, SSIM, and LPIPS; as well as our proposed Reprojection Error (RE@SG) under SuperGlue estimation which does not require ground truth (Sec. 3.4). We report 2D Face and Body similarities. Red subscript show deviation against ground truth value." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 264, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 295, + 407 + ], + "type": "text", + "content": "Generations from Casual iPhone Photos. We present in Tab. 3 (Rows 3,6) Reprojection Error and similarity scores for casually taken images from the iPhone dataset with 1K resolution model. In this scenario, the standard reconstruction error metrics cannot be assessed due to missing ground truth. We find that the reprojection error on iPhone captures remains comparable to the studio dataset - demonstrating 3D consistency. This illustrates the generalizability of Pippo beyond the multi-view training data domain, where our pretraining with large-scale in-the-wild human data is critical. We put corresponding visuals in ??, Row 1 (Appendix)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 410, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 518 + ], + "type": "text", + "content": "Comparisons with External Benchmarks. In Fig. 4, we compare Pippo with individual state-of-the-art baselines in Full-body and Head-only generation. SiTH [25] reconstructs textured human-mesh using ControlNet paired with a SDF representation. Compared to SiTH, Pippo facilitates high-resolution and accurate multiview synthesis. DiffPortrait3D [20] inverts a 3D-GAN based on a given input image. Compared to it, our model supports greater viewpoint variability and ensures closer adherence to the input image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 520, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 662 + ], + "type": "text", + "content": "Quantitative comparisons and baselines. Existing SoTA Human methods [20, 25] use explicit SMPL priors, and thus are difficult to compare with directly. Qualitatively, we found that they cannot handle novel views or preserve details (Fig. 4), and hence do not quantitatively compare against them. In Pippo, we focus on creating a strong multiview human generator, and we benchmark four state-of-the-art multi-view diffusion models on the iPhone full-body dataset in Tab. 4. We find that Pippo preserves identity (i.e., face and body similarity) and 3D consistency (RE) better while also operating at a higher resolution compared to baselines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Benchmarking Pippo on public datasets. In Tab. 5, we benchmark Pippo on the public Codec Avatar datasets [43] to aid future comparisons. Specifically, we use Ava-256 containing 256 head-only captures and Goliath containing 4" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 348, + 70, + 521, + 127 + ], + "blocks": [ + { + "bbox": [ + 348, + 70, + 521, + 127 + ], + "lines": [ + { + "bbox": [ + 348, + 70, + 521, + 127 + ], + "spans": [ + { + "bbox": [ + 348, + 70, + 521, + 127 + ], + "type": "table", + "html": "
#Method & ResolutionRE@SG ↓Face ↑Body ↑
1.MV-Adapter [27] (768)4.743.064.1
2.Era3D [38] (512)4.138.164.4
3.Wonder3D [42] (256)5.334.758.8
4.Pippo (P3@1K)3.058.068.1
", + "image_path": "f7cfe4dc9030a24476f896b1d7e6352896f9fd0fd1091ac86ed64200fafa42cc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 136, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 313, + 136, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 553, + 190 + ], + "type": "text", + "content": "Table 4. Quantitative comparison with SoTA multi-view models. We quantitatively compare Pippo against state-of-the-art multi-view diffusion models. We find that Pippo preserves identity (i.e., face and body similarity) and 3D consistency (RE) better while operating at a higher resolution compared to baselines." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 212, + 553, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 553, + 259 + ], + "type": "text", + "content": "full-body. During evaluation, we only use subsets of these datasets that were not used for training. We find that Pippo's performance on these datasets is inline with its performance on internal studio datasets." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 268, + 553, + 307 + ], + "blocks": [ + { + "bbox": [ + 315, + 268, + 553, + 307 + ], + "lines": [ + { + "bbox": [ + 315, + 268, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 315, + 268, + 553, + 307 + ], + "type": "table", + "html": "
#Dataset (P3@1K)RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
1.Ava-256 (Head-only)3.820.168.726.672.976.8
2.Goliath (Full-body)1.220.489.715.587.577.7
", + "image_path": "138833e6f8a20e686357191950441214fb3945269252631b82f4182a58b4c394.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 316, + 553, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 361 + ], + "type": "text", + "content": "Table 5. Benchmarking Pippo on public Ava-256 and Goliath datasets. We find that Pippo's performance on these datasets is inline with its performance on internal studio datasets. This will aid future comparisons against Pippo." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 382, + 382, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 382, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 382, + 393 + ], + "type": "text", + "content": "4.4. Ablations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 399, + 553, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 553, + 436 + ], + "type": "text", + "content": "We examine design choices at each training stage (Sec. 3) and present our ablation results in Tab. 6. Experiments are conducted at " + }, + { + "bbox": [ + 313, + 399, + 553, + 436 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 399, + 553, + 436 + ], + "type": "text", + "content": " resolution on the Head-only dataset." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 316, + 445, + 553, + 579 + ], + "blocks": [ + { + "bbox": [ + 316, + 445, + 553, + 579 + ], + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 579 + ], + "type": "table", + "html": "
#Method (Stage @ Res.)3D ConsistencySimilarity
RE@SG↓PSNR↑SSIM↑LPIPS↓Face ↑Body ↑
1.Pippo (P3@512)2.7+0.721.771.721.574.1-2.077.4-2.4
2.Pippo Face-only (P3@512)2.4+0.420.570.325.374.3-1.775.8-4.0
3.No Mid-train (P3@512)5.6+3.614.559.544.220.4-55.663.4-16.4
4.Pippo (P3@128)3.3+0.821.067.614.862.0-13.361.5-16.2
5.No Anchor (P3@128)11.5+9.017.154.021.163.1-12.268.3-9.3
6.No Plücker (P3@128)4.2+1.720.264.516.163.5-11.866.1-11.5
7.Pippo (M2@128)3.4+0.919.161.417.260.0-15.362.6-15.0
8.No Pre-train (M2@128)5.9+3.413.139.344.919.5-55.849.0-28.6
9.Cross-Attn (M2@128)3.5+1.018.059.222.166.2-9.170.7-7.0
10.Non-frontal (M2@128)5.8+3.315.252.430.149.2-26.160.7-16.9
", + "image_path": "08467f4f8305391de885ca1743be65be613156a1e9606bf39b6b5f8b9dc13fdc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 587, + 553, + 665 + ], + "lines": [ + { + "bbox": [ + 313, + 587, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 553, + 665 + ], + "type": "text", + "content": "Table 6. Ablation on design choices and training stages. We evaluate several multi-view models at " + }, + { + "bbox": [ + 313, + 587, + 553, + 665 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 587, + 553, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 587, + 553, + 665 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 587, + 553, + 665 + ], + "type": "text", + "content": " resolution at different training stages on Head-only dataset. We ablate the choice of doing Mid-training and Pre-training, along with different spatial controls and reference conditioning methods..The training stages are labeled as M2 (Mid-training) and P3 (Post-training), followed by @ specified resolution." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Significance of Pre-training and Mid-training. Pretraining our model on the Humans-3B dataset enables robust generalization to novel identities, as demonstrated in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16424" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 555, + 227 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 555, + 227 + ], + "type": "image", + "image_path": "0a1556fb832c03b8b9171a80936dc21cd43442942792f4423be8d34086dead0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 257 + ], + "type": "text", + "content": "Figure 4. Visual comparison with state-of-the-art methods. We visually compare Pippo with state-of-the-art baselines DiffPortrait3D [20] (head-only) and SiTH [25] (full-body) generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 275, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 275, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 294, + 323 + ], + "type": "text", + "content": "Tab. 6, Row 8. Without pre-training, model generalization deteriorates, resulting in unclear facial features. Skipping mid-training at lower resolution impairs consistent multiview generation, as shown in Row 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 396 + ], + "type": "text", + "content": "Importance of Frontal Input Reference. Our ablation in Tab. 6, Row 10 demonstrates that completely randomizing the view point of an input reference image leads to overfitting to training identities. Non-frontal views, especially rear views, have very limited information about the identity which forces the network to pick up spurious correlations." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 408, + 294, + 463 + ], + "blocks": [ + { + "bbox": [ + 56, + 408, + 294, + 463 + ], + "lines": [ + { + "bbox": [ + 56, + 408, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 294, + 463 + ], + "type": "table", + "html": "
# Pretrain Data (M2@128)RE@SG ↓PSNR ↑SSIM ↑LPIPS ↓Face ↑Body ↑
1. 30M (1% HQ)3.821.381.813.330.666.2
2. 900M (30% Random)4.021.682.012.130.667.2
3. 2.1B (70% Random)3.721.682.112.137.667.5
4. Humans-3B (100%)3.721.982.312.741.767.2
", + "image_path": "8f1698f351e4d427ed5447ad16096ad58ca475bbdf358843b809bce5e61a597b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 464, + 295, + 521 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 521 + ], + "type": "text", + "content": "Table 7. Ablating the size of pre-training dataset on Humans-3B. We conduct Full-body mid-training with pretrained checkpoints trained on " + }, + { + "bbox": [ + 55, + 464, + 295, + 521 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 464, + 295, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 464, + 295, + 521 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 464, + 295, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 464, + 295, + 521 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 464, + 295, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 464, + 295, + 521 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 464, + 295, + 521 + ], + "type": "text", + "content": " subsets of Humans-3B dataset. We found large-scale data is crucial for generalization to novel identities (i.e., face similarity)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "type": "text", + "content": "Importance of Self-attention. Replacing self-attention with cross-attention for reference image encoding, using the same routing as image-conditioned pretraining from Sec. 3.1, leads to degraded performance as shown in Tab. 6, Row 9. We observe that this setup causes the model to ignore input conditioning, generating images that only vaguely resemble training subjects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Role of large scale Humans-3B pre-training dataset. We utilized intermediate checkpoints from pre-training stage (P1) trained on " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " of the data, and a separate checkpoint trained on " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " high-quality subset of Humans-3B. Starting from these checkpoints, we mid-trained Pippo to denoise 4 views at " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " resolution on Full-body dataset for two days. We report their respective results in Tab. 7. We found large-scale data is crucial for general" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 300 + ], + "type": "text", + "content": "ization to novel identities – indicated by high gains in face similarity metric." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 311, + 388, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 388, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 388, + 324 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 331, + 555, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 571 + ], + "type": "text", + "content": "We present Pippo, a diffusion transformer model that generates a dense set of high-resolution multi-view consistent images of a person from a single image. Our experiments show that our multi-stage training strategy, which combines largescale in-the-wild data with high-quality multi-view studio captures, enables generalizable high-resolution multi-view synthesis. Analysis of the diffusion transformer architecture reveals that self-attention with the reference image, Plücker coordinates with SIREN, and Spatial Anchor are all essential for high-fidelity multi-view human generation. Pippo achieves, for the first time, consistent multi-view human image generation at 1K resolution. We also show that our proposed 3D consistency metric enables evaluation of 3D consistency without paired ground-truth data. One of the limitations of our approach is the limited number of simultaneously generated views, originating from large context length and memory constraints, which can be tackled with parallelization techniques and autoregressive generation. Extending our approach to multi-view consistent video generation will be addressed in future work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 582, + 373, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 582, + 373, + 594 + ], + "spans": [ + { + "bbox": [ + 314, + 582, + 373, + 594 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 602, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 319, + 602, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 602, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 319, + 602, + 555, + 657 + ], + "type": "text", + "content": "[1] Sizhe An, Hongyi Xu, Yichun Shi, Guoxian Song, Umit Y. Ogras, and Linjie Luo. Panohed: Geometry-aware 3d fullhead synthesis in 360deg. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20950-20959, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 319, + 658, + 554, + 712 + ], + "type": "text", + "content": "[2] Timur Bagautdinov, Chenglei Wu, Tomas Simon, Fabián Prada, Takaaki Shiratori, Shih-En Wei, Weipeng Xu, Yaser Sheikh, and Jason Saragih. Driving-signal aware full-body avatars. ACM Transactions on Graphics (TOG), 40(4):1-17, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16425" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 137 + ], + "type": "text", + "content": "[3] Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, et al. Vd3d: Taming large video diffusion transformers for 3d camera control. arXiv preprint arXiv:2407.12781, 2024.5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 294, + 194 + ], + "type": "text", + "content": "[4] Qiong Cao, Li Shen, Weidi Xie, Omkar M. Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2018), pages 67-74, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 195, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 294, + 248 + ], + "type": "text", + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 251, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 294, + 293 + ], + "type": "text", + "content": "[6] Eric Ming Chen, Sidhanth Holalkere, Ruyu Yan, Kai Zhang, and Abe Davis. Ray conditioning: Trading photoconsistency for photo-realism in multi-view image generation. arXiv preprint arXiv:2304.13681, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 295, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 294, + 326 + ], + "type": "text", + "content": "[7] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 327, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 413 + ], + "type": "text", + "content": "[8] Wei Cheng, Ruixiang Chen, Wanqi Yin, Siming Fan, Keyu Chen, Honglin He, Huiwen Luo, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, Daxuan Ren, Lei Yang, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, Bo Dai, and Kwan-Yee Lin. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. arXiv preprint, arXiv:2307.10173, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 416, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 416, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 416, + 294, + 437 + ], + "type": "text", + "content": "[9] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 438, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 438, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 438, + 294, + 612 + ], + "type": "text", + "content": "[10] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme Ruiz, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd Van Steenkiste, Gamaleldin Fathy Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmiŋn Bastings, Mark Collier, Alexey A. Gritsenko, Vighnesh Birodkar, Cristina Nader Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lucic, Xiaohua Zhai, Daniel Keysers, Jeremiah J. Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters. In Proceedings of the 40th International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "type": "text", + "content": "[11] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. In CVPR, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "text", + "content": "[13] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "type": "text", + "content": "and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 106, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 553, + 148 + ], + "type": "text", + "content": "[14] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. In ACM MM, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 151, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 215 + ], + "type": "text", + "content": "[15] Nikita Drobyshev, Antoni Bigata Casademunt, Konstantinos Vougioukas, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Emoportraits: Emotion-enhanced multimodal one-shot head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8498-8507, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 217, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 281 + ], + "type": "text", + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 283, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 325 + ], + "type": "text", + "content": "[17] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "type": "text", + "content": "[18] Matheus Gadelha, Rui Wang, and Subhransu Maji. Multiresolution tree networks for 3d point cloud processing. In ECCV, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "text", + "content": "[19] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 415, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 553, + 480 + ], + "type": "text", + "content": "[20] Yuming Gu, Hongyi Xu, You Xie, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10456-10465, 2024. 3, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "text", + "content": "[21] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 537, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 568 + ], + "type": "text", + "content": "[22] R. I. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "text", + "content": "[23] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "text", + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[25] I Ho, Jie Song, Otmar Hilliges, et al. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 538-549, 2024. 7, 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16426" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 94 + ], + "type": "text", + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 2020. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 95, + 295, + 139 + ], + "type": "text", + "content": "[27] Zehuan Huang, Yuanchen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 140, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 140, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 140, + 295, + 205 + ], + "type": "text", + "content": "[28] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9784-9794, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 206, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 206, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 57, + 206, + 294, + 259 + ], + "type": "text", + "content": "[29] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 262, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 262, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 57, + 262, + 295, + 315 + ], + "type": "text", + "content": "[30] Yash Kant, Aliaksandr Siarohin, Michael Vasilkovsky, Riza Alp Guler, Jian Ren, Sergey Tulyakov, and Igor Gilitschenski. invs: Repurposing diffusion inpainters for novel view synthesis. In SIGGRAPH Asia 2023 Conference Papers, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 317, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 317, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 57, + 317, + 294, + 380 + ], + "type": "text", + "content": "[31] Yash Kant, Aliaksandr Siarohin, Ziyi Wu, Michael Vasilkovsky, Guocheng Qian, Jian Ren, Riza Alp Guler, Bernard Ghanem, Sergey Tulyakov, and Igor Gilitschenski. Spad: Spatially aware multi-view diffusers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10026-10038, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 383, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 383, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 57, + 383, + 294, + 413 + ], + "type": "text", + "content": "[32] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 416, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 416, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 57, + 416, + 294, + 469 + ], + "type": "text", + "content": "[33] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 472, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 472, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 57, + 472, + 294, + 514 + ], + "type": "text", + "content": "[34] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Transactions on Graphics (TOG), 42(4):1-14, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 515, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 515, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 515, + 294, + 535 + ], + "type": "text", + "content": "[35] Black Forest Labs. Flux model, 2024. Accessed: 2024-11-10. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 537, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 537, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 57, + 537, + 294, + 558 + ], + "type": "text", + "content": "[36] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 294, + 613 + ], + "type": "text", + "content": "[37] Heyuan Li, Ce Chen, Tianhao Shi, Yuda Qiu, Sizhe An, Guanying Chen, and Xiaoguang Han. Spherehead: Stable 3d full-head synthesis with spherical tri-plane representation. In European Conference on Computer Vision, pages 324-341. Springer, 2025. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 615, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 615, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 615, + 294, + 667 + ], + "type": "text", + "content": "[38] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, Wenping Wang, Qifeng Liu, and Yike Guo. Era3d: High-resolution multiview diffusion using efficient row-wise attention, 2024. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[39] Chieh Hubert Lin, Chia-Che Chang, Yu-Sheng Chen, Da-Cheng Juan, Wei Wei, and Hwann-Tzong Chen. Coco-gan: Generation by parts via conditional coordinating. In ICCV, 2019. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "type": "text", + "content": "[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 155, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 155, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 155, + 553, + 208 + ], + "type": "text", + "content": "[42] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 212, + 553, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 212, + 553, + 451 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 553, + 451 + ], + "type": "text", + "content": "[43] Julieta Martinez, Emily Kim, Javier Romero, Timur Bagautdinov, Shunsuke Saito, Shoou-I Yu, Stuart Anderson, Michael Zollhöfer, Te-Li Wang, Shaojie Bai, Chenghui Li, Shih-En Wei, Rohan Joshi, Wyatt Borsos, Tomas Simon, Jason Saragih, Paul Theodosis, Alexander Greene, Anjani Josyula, Silvio Mano Maeta, Andrew I. Jewett, Simon Venshtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mohamed Ezzeldin A. Elshaer, Tingfang Du, Longhua Wu, Shen-Chi Chen, Kai Kang, Michael Wu, Youssef Emad, Steven Longay, Ashley Brewer, Hitesh Shah, James Booth, Taylor Koska, Kayla Haidle, Matt Andromalos, Joanna Hsu, Thomas Dauer, Peter Selednik, Tim Godisart, Scott Ardisson, Matthew Cipperly, Ben Humberston, Lon Farr, Bob Hansen, Peihong Guo, Dave Braun, Steven Krenn, He Wen, Lucas Evans, Natalia Fadeeva, Matthew Stewart, Gabriel Schwartz, Divam Gupta, Gyeongsik Moon, Kaiwen Guo, Yuan Dong, Yichen Xu, Takaaki Shiratori, Fabian Prada, Bernardo R. Pires, Bo Peng, Julia Buffalini, Autumn Trimble, Kevyn McPhail, Melissa Schoeller, and Yaser Sheikh. Codec Avatar Studio: Paired Human Captures for Complete, Driveable, and Generalizable Avatars. NeurIPS Track on Datasets and Benchmarks, 2024. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 455, + 553, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 553, + 519 + ], + "type": "text", + "content": "[44] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 523, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 586 + ], + "type": "text", + "content": "[45] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "type": "text", + "content": "[46] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "text", + "content": "[47] Dongwei Pan, Long Zhuo, Jingtan Piao, Huiwen Luo, Wei Cheng, Yuxin Wang, Siming Fan, Shengqi Liu, Lei Yang, Bo Dai, Ziwei Liu, Chen Change Loy, Chen Qian, Wayne Wu, Dahua Lin, and Kwan-Yee Lin. RenderMe-360: Large digital asset library and benchmark towards high-fidelity head avatars. In Thirty-seventh Conference on Neural Information" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16427" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "text", + "content": "Processing Systems Datasets and Benchmarks Track, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[48] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "type": "text", + "content": "[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 189, + 295, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 295, + 220 + ], + "type": "text", + "content": "[50] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 223, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 295, + 278 + ], + "type": "text", + "content": "[51] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs- avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 280, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 295, + 334 + ], + "type": "text", + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 337, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 295, + 392 + ], + "type": "text", + "content": "[53] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. arXiv preprint arXiv:2303.13508, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 295, + 437 + ], + "type": "text", + "content": "[54] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 440, + 295, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 440, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 440, + 295, + 472 + ], + "type": "text", + "content": "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 475, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 295, + 518 + ], + "type": "text", + "content": "[56] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In CVPR, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 520, + 295, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 295, + 575 + ], + "type": "text", + "content": "[57] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 577, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 295, + 632 + ], + "type": "text", + "content": "[58] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "text", + "content": "[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[60] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "text", + "content": "[61] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 119, + 554, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 554, + 174 + ], + "type": "text", + "content": "[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 239 + ], + "type": "text", + "content": "[63] Akash Sengupta, Thiemo Alldieck, Nikos Kolotouros, Enric Corona, Andrei Zanfir, and Cristian Sminchisescu. Diffhuman: Probabilistic photorealistic 3d reconstruction of humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1439-1449, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 554, + 297 + ], + "type": "text", + "content": "[64] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 300, + 554, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 554, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 554, + 332 + ], + "type": "text", + "content": "[65] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 334, + 554, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 334, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 316, + 334, + 554, + 367 + ], + "type": "text", + "content": "[66] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 369, + 554, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 369, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 369, + 554, + 402 + ], + "type": "text", + "content": "[67] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 403, + 554, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 403, + 554, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 403, + 554, + 446 + ], + "type": "text", + "content": "[68] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 554, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 554, + 483 + ], + "type": "text", + "content": "[69] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. 2015. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 484, + 554, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 484, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 484, + 554, + 529 + ], + "type": "text", + "content": "[70] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Danbo: Disentangled articulated neural body representations via graph neural networks. In European Conference on Computer Vision, pages 107-124. Springer, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 529, + 554, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 529, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 529, + 554, + 574 + ], + "type": "text", + "content": "[71] Shih-Yang Su, Timur Bagautdinov, and Helge Rhodin. Npc: Neural point characters from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14795-14805, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 575, + 554, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 575, + 554, + 608 + ], + "spans": [ + { + "bbox": [ + 316, + 575, + 554, + 608 + ], + "type": "text", + "content": "[72] Qingyang Tan, Lin Gao, Yu-Kun Lai, and Shihong Xia. Variational autoencoders for deforming 3d mesh models. In CVPR, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 610, + 554, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 610, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 610, + 554, + 643 + ], + "type": "text", + "content": "[73] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 646, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 554, + 678 + ], + "type": "text", + "content": "[74] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[75] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16428" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[76] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 194 + ], + "type": "text", + "content": "[77] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4d-dress: A 4d dataset of real-world human clothing with semantic annotations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[78] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "text", + "content": "[79] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 363 + ], + "type": "text", + "content": "[80] Jing Wen, Xiaoming Zhao, Zhongzheng Ren, Alexander G Schwing, and Shenlong Wang. Gomavatar: Efficient animatable human modeling from monocular video using gaussians-on-mesh. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2059-2069, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[81] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[82] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "text", + "content": "[83] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Song-Chun Zhu, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In CVPR, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 576 + ], + "type": "text", + "content": "[84] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, et al. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19801-19811, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "text", + "content": "[85] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "text", + "content": "[86] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[87] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. FaceScape: A largescale high quality 3D face dataset and detailed riggable 3D" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 441 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "face prediction. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[88] T Yenamandra, A Tewari, F Bernard, HP Seidel, M Elgharib, D Cremers, and C Theobalt. i3DMM: Deep implicit 3D morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "text", + "content": "[89] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[90] Zhixuan Yu, Jae Shin Yoon, In Kyu Lee, Prashanth Venkatesh, Jaesik Park, Jihun Yu, and Hyun Soo Park. HUMBI: A large multiview dataset of human body expressions. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 555, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 297 + ], + "type": "text", + "content": "[91] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 363 + ], + "type": "text", + "content": "[92] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22096-22105, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 396 + ], + "type": "text", + "content": "[93] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "type": "text", + "content": "[94] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16429" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/640cd516-44b5-47ce-aa6a-b65d88b8da0a_content_list.json b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/640cd516-44b5-47ce-aa6a-b65d88b8da0a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4a8301aa18e091a3ba4d2058601ca656f18b0abd --- /dev/null +++ b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/640cd516-44b5-47ce-aa6a-b65d88b8da0a_content_list.json @@ -0,0 +1,1817 @@ +[ + { + "type": "text", + "text": "Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision", + "text_level": 1, + "bbox": [ + 135, + 128, + 833, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinnyeong Kim Seung-Hwan Baek POSTECH", + "bbox": [ + 330, + 180, + 637, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 250, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Integrating RGB and NIR stereo imaging provides complementary spectral information, potentially enhancing robotic 3D vision in challenging lighting conditions. However, existing datasets and imaging systems lack pixel-level alignment between RGB and NIR images, posing challenges for downstream vision tasks. In this paper, we introduce a robotic vision system equipped with pixel-aligned RGB-NIR stereo cameras and a LiDAR sensor mounted on a mobile robot. The system simultaneously captures pixel-aligned pairs of RGB stereo images, NIR stereo images, and temporally synchronized LiDAR points. Utilizing the mobility of the robot, we present a dataset containing continuous video frames under diverse lighting conditions. We then introduce two methods that utilize the pixel-aligned RGB-NIR images: an RGB-NIR image fusion method and a feature fusion method. The first approach enables existing RGB-pretrained vision models to directly utilize RGB-NIR information without fine-tuning. The second approach fine-tunes existing vision models to more effectively utilize RGB-NIR information. Experimental results demonstrate the effectiveness of using pixel-aligned RGB-NIR images across diverse lighting conditions.", + "bbox": [ + 73, + 282, + 473, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 646, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "RGB imaging captures visible light in wavelengths from $400\\mathrm{nm}$ to $700\\mathrm{nm}$ and serves as the primary data format for visual computing. Near-infrared (NIR) imaging with active illumination captures light in the range of $750\\mathrm{nm}$ to $1000\\mathrm{nm}$ , offering effective information in conditions where where RGB imaging is inadequate, such as low-light environments during night or poorly lit indoor conditions. NIR light is imperceptible to human vision, allowing for using active NIR illumination without disturbing people, which facilitates robust 3D imaging [5, 32, 38, 54] and eye tracking [3]. Additionally, NIR light offers supplementary scene information due to different spectral reflectance between RGB and NIR wavelengths [18, 27, 33].", + "bbox": [ + 75, + 672, + 468, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Combining these benefits, RGB-NIR imaging emerges as a promising modality for challenging environments re", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quiring robust and efficient 2D and 3D vision capabilities. RGB-NIR imaging has been studied for applications including object detection [50, 61] and 3D reconstruction [5, 21, 22, 43]. However, existing RGB-NIR imaging systems and datasets face significant challenges compared to standard RGB vision systems. A major issue is the use of separate RGB and NIR cameras at different poses, leading to pixel misalignment between the NIR and RGB images. Previous methods address such misalignments using image registration, pose estimation, and other processing techniques, which can accumulate errors and limit effectiveness [5, 48, 63].", + "bbox": [ + 496, + 252, + 893, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we develop a robotic vision system capable of capturing pixel-aligned RGB-NIR stereo images and synchronized LiDAR point clouds. Mounted on a mobile platform, the system enables exploration in both indoor and outdoor environments, facilitating data collection under diverse lighting conditions. Using this system, we collect and share a dataset of pixel-aligned RGB-NIR stereo images with synchronized LiDAR point clouds as ground-truth depth for 82 real-world scenes including 90,000 frames. Additionally, we provide a synthetic RGB-NIR dataset with ground-truth dense depth labels for 2,238 synthetic scenes. To demonstrate the effectiveness of pixel-aligned RGB-NIR imaging and datasets, we propose an RGB-NIR image fusion method and an RGB-NIR feature fusion method.", + "bbox": [ + 496, + 436, + 893, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The first method fuses the pixel-aligned RGB-NIR images into a three-channel image that can be input into pretrained RGB vision models without finetuning. The second method involves developing a stereo depth estimation technique by finetuning pretrained stereo networks [7, 36, 57] to utilize pixel-aligned RGB-NIR features. Our experiments show improvements in various downstream tasks over using only RGB or NIR data, as well as over methods relying on pixel-misaligned RGB-NIR datasets, particularly in challenging scenarios.", + "bbox": [ + 496, + 651, + 895, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows.", + "bbox": [ + 517, + 806, + 820, + 821 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a robotic vision system with pixel-aligned RGB-NIR stereo cameras and LiDAR, capturing synchronized RGB-NIR stereo images and LiDAR point clouds under diverse lighting conditions.", + "- We present a large-scale dataset of pixel-aligned RGB-" + ], + "bbox": [ + 500, + 824, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11482", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NIR stereo images and LiDAR point clouds, collected in various indoor and outdoor environments.", + "bbox": [ + 89, + 90, + 467, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an RGB-NIR image fusion method and an RGB-NIR stereo depth estimation network using RGB-NIR feature fusion.", + "- We evaluate the proposed methods on diverse downstream tasks and illumination conditions, on synthetic and real-world datasets, outperforming baselines." + ], + "bbox": [ + 76, + 121, + 467, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 228, + 217, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RGB-NIR Imaging Conventional RGB-NIR imaging systems employ separate RGB and NIR cameras positioned at different viewpoints, such as the Kinect sensor, Intel RealSense D415, and various multi-view RGB-NIR systems [8, 9, 11, 12, 29, 48, 56, 58, 63]. However, the captured RGB and NIR images from these systems exhibit pixel misalignment, necessitating depth-dependent registration between the RGB and NIR images. This introduces a significant challenge for effectively utilizing RGB and NIR information, creating a chicken-and-egg problem between depth estimation using spectrally fused images and spectral pose alignment using estimated depth.", + "bbox": [ + 76, + 253, + 467, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the pixel misalignment, sequential capture with interchangeable RGB and NIR filters has been used [4, 25]; however, this approach is limited to capturing static scenes due to the time delay between captures. Designing custom color filter arrays for image sensors to record four channels (RGB-NIR) has also been explored [40]. However, using a single sensor with a fixed global exposure for all four channels makes it unsuitable for challenging environments where the dynamic ranges for RGB and NIR channels differ significantly, such as in dark indoor or outdoor conditions with active NIR illumination. In the domain of RGB-thermal imaging, beam splitters have been used to combine an RGB camera and a thermal camera, enabling pixel-aligned RGB-thermal imaging [23, 28, 60].", + "bbox": [ + 76, + 436, + 467, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We capture pixel-aligned RGB-NIR stereo images alongside LiDAR point clouds using two prism-based RGB-NIR dual-sensor cameras. The system is mounted on a mobile robot, allowing scalable data acquisition of pixel-aligned RGB-NIR stereo images in both indoor and outdoor conditions.", + "bbox": [ + 76, + 648, + 467, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RGB-NIR Fusion To utilize information from both RGB and NIR spectra, fusing RGB-NIR images has been extensively studied with applications in image enhancement for long-distance visibility and fog penetration [1, 10, 20, 62], as well as in information visualization leveraging different RGB-NIR reflectance properties [18, 19, 34, 47]. Conventional methods apply simple arithmetic operations such as addition and subtraction in RGB, HSL, HSV, and YUV color spaces to create images that contain both RGB and", + "bbox": [ + 76, + 763, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NIR information [15, 19, 20, 37, 42, 52]. Edge-aware filtering techniques have been employed to refine noisy and low-contrast RGB images using guidance from clean and high-contrast NIR images, showing promising performance in various imaging scenarios [10, 24, 31, 34]. More recently, learning-based approaches have attempted to enhance RGB images and their 3D vision applications using NIR guidance by fusing RGB-NIR features [25, 26, 30, 49, 55] or RGB-thermal features [14, 23, 59].", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging our pixel-aligned RGB-NIR images, we develop a RGB-NIR image fusion method that effectively encodes information from both RGB and NIR images into a single three-channel image with learned spatially-varying weights. The resulting fused image can be directly used as input to neural networks pretrained on RGB images.", + "bbox": [ + 496, + 227, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RGB-NIR Depth Estimation Cross-spectral correspondence matching between RGB and NIR images using an RGB-NIR color transform network has demonstrated promising depth estimation results when the RGB-NIR reflectances are similar [63]. Liang et al. showed that GAN based architecture can enhance RGB-NIR color transform on cross-spectral stereo matching [35]. Active stereo systems using structured NIR illumination, such as the Kinect and Intel RealSense D415, employ multiple RGB and NIR cameras at different viewpoints. These systems often suffer from the aforementioned chicken-and-egg problem between depth estimation using spectrally fused images and spectral fusion using the estimated depth. Brucker et al. [5] fuse red, blue, and clean channels of stereo RCCB cameras with gating-based time-of-flight cameras positioned differently but still face a similar chicken-and-egg problem. In the domain of RGB-thermal imaging, Guo et al. [17] demonstrate effective depth estimation using pixel-aligned RGB-thermal stereo cameras with beam splitters.", + "bbox": [ + 496, + 334, + 890, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We develop a depth estimation method leveraging our pixel-aligned RGB-NIR images for robust depth estimation, thereby bypassing the chicken-and-egg problem between depth estimation and spectral fusion.", + "bbox": [ + 496, + 622, + 890, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Pixel-aligned RGB-NIR Stereo Imaging", + "text_level": 1, + "bbox": [ + 500, + 694, + 854, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Robotic Imaging System Figure 1(a) illustrates our robotic imaging system, which integrates a stereo pair of pixel-aligned RGB-NIR cameras, NIR illumination, and a LiDAR mounted on a mobile robot (AgileX Ranger Mini 2.0). Each RGB-NIR camera (JAI FS-1600D-10GE), as shown in Figure 1(c), employs a dichroic prism to separate RGB and NIR light, independently captured by RGB and NIR CMOS sensors. To enhance NIR visibility under varying lighting conditions, we include active NIR illumination (Advanced Illumination AL295-150850IC) without disturbing human vision. The spectral sensitivities of the camera and the illumination are shown in Figure 1(b). We", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11483", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6bdb5b8e2650c642bc857dac65b417eff7ccc9592c68034991f7fa79f9dfc93b.jpg", + "image_caption": [ + "(a) Imaging system" + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 336, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/42756beefa69eac440a5b29aba76cb7cca9c884b61accf35d4b800f38456f62b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 88, + 593, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/66f05de8d39b4afbd616c20250f24d4293ff51d299ff7ea5bab71ebc5e5efdeb.jpg", + "image_caption": [ + "(c) Pixel-aligned RGB-NIR camera", + "Figure 1. RGB-NIR stereo imaging system. (a) Stereo camera setup integrated with LiDAR and active NIR illumination. (b) Spectral sensitivity profiles of the camera sensors and the irradiance profile of the NIR illumination. (c) Diagram of the pixel-aligned RGB-NIR camera featuring a dichroism prism, where RGB light penetrates the prism and NIR light is reflected, achieving spectral separation." + ], + "image_footnote": [], + "bbox": [ + 553, + 88, + 880, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "installed a LiDAR (OUSTER OS1) to acquire ground-truth sparse depth maps, along with angular velocity and linear acceleration data from its inertial measurement unit. We use RJ-45 interfaces for the data transmission of the cameras and LiDAR. We set up an AC powerbank to supply power to the devices that require an AC adapter. These devices are connected to a laptop, which triggers the capture pipeline and manages real-time data storage. We manage separate threads on the laptop for the left camera, right camera, and LiDAR, acquiring synchronized data.", + "bbox": [ + 75, + 362, + 472, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Formation We model the image formation of the pixel-aligned RGB-NIR camera, which captures RGB and NIR images on separate CMOS sensors. This configuration results in the same exposure time and gain across the RGB channels, with separate exposure time and gain for the NIR channel: $t_R = t_G = t_B \\neq t_{\\mathrm{NIR}}$ and $g_R = g_G = g_B \\neq g_{\\mathrm{NIR}}$ . The image captured by camera $c \\in \\{\\text{left}, \\text{right}\\}$ for channel $i \\in \\{\\text{R}, \\text{G}, \\text{B}, \\text{NIR}\\}$ is given by", + "bbox": [ + 75, + 534, + 470, + 656 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} ^ {c} (p ^ {c}) = \\eta_ {1} + g _ {i} \\left(\\eta_ {2} + t _ {i} \\left(R _ {i} ^ {c} (p ^ {c}) \\left(E _ {i} ^ {c} (p ^ {c}) + L _ {i} ^ {c} (p ^ {c})\\right)\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 667, + 468, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p^c$ is a pixel of camera $c$ , $R_i^c(p^c)$ is the reflectance at pixel $p^c$ for channel $i$ , and $\\eta_1, \\eta_2$ are Gaussian noise in the post-gamma and pre-gamma stages, respectively. The intensities measured in each channel are influenced by environmental illumination, denoted as $E_i^c(p^c)$ for channel $i$ . The active illumination $L_i^c(p^c)$ is present only in the NIR channel, satisfying $L_i^c(p^c) = 0$ for $i \\in \\{\\mathrm{R}, \\mathrm{G}, \\mathrm{B}\\}$ , since the active illumination is confined to the NIR spectrum (see Figure 1(c)).", + "bbox": [ + 75, + 698, + 468, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Calibration One advantage of using pixel-aligned cameras is that calibration between RGB and NIR images is unnecessary. Thus, we only calibrate the stereo cameras", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "using checkerboard calibration [2]. The cameras contain precision time protocol, which enables maintaining temporal difference between stereo cameras less than 1 microsecond. We also obtain the LiDAR extrinsics by using corresponding point pairs between left-camera images and LiDAR point clouds [44].", + "bbox": [ + 496, + 362, + 890, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Real-world Dataset Using the system, we acquire a dataset consisting of 82 videos totaling 90,000 frames under various lighting conditions for training and 7000 frames for testing. For each frame, we provide pixel-aligned RGB-NIR stereo images, a sparse LiDAR point cloud, sensor timestamps, and exposure values for the RGB and NIR CMOS sensors. Each video was recorded at frame rates of $5 - 10\\mathrm{Hz}$ , with durations ranging from five to over thirty minutes. Figure 2(a) shows samples from our real dataset and augmented synthetic dataset. We categorize our dataset with lighting conditions: well-lit, dark and complex, based on indoor or outdoor environment. Figure 2(b) shows the statistics of our dataset. We use auto exposure on both RGB and NIR sensors independently, with fixed gain values, and recorded the exposure time for each image. The distribution of exposure times in our dataset varies depending on lighting conditions. For the indoor dataset, many frames have short exposure times for NIR images, as active illumination is sufficient to illuminate nearby indoor scenes. In the daytime dataset, strong sunlight causes the exposure times to shift towards very short durations under direct sunlight and longer durations in shaded areas. For the night dataset, although many frames exhibit long exposure times for both RGB and NIR, the average exposure time for NIR frames is shorter than that for RGB frames.", + "bbox": [ + 496, + 472, + 892, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic Dataset We augmented synthetic RGB stereo-image dataset [39] to obtain RGB-NIR stereo dataset with", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11484", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1c75b476738be0abdc435c478e4617fc2723b0bec0962dbce8ce9d32ea430fac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 89, + 225, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aad553ce203d37f77d9ed9982f1335477b01f22c386982d95e7f035658ee7d67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 90, + 354, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e13cb34444c805ad299dd21657c619920c9f4a68aaaf5150eb2900b78471de57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 90, + 482, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/40d0db20c310a85a23607e2284c369416d9c36b55fd0a07b53e4ecf7bd9b8145.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 90, + 609, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d804875de43fcbbb5aa4e4fdd80528794fa90db7b37fc1d673acc3405d42dd6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 90, + 738, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6ecfc06e90d4fcf20464b0391d988c0b3487f2e122e3ab680b8457158a8be9e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 90, + 892, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea1a5f990256691827bf3d0605b0c601c612ede73b6f180281da74c5fdb72f5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 262, + 228, + 371 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ada75e867842a84b31feea414fce65219add8101645002a856dfa61aaddb256d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 263, + 390, + 371 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c5d6541d77b8d09028d8ef85881c709f21d91b56c1560a4051863fefdf603e17.jpg", + "image_caption": [ + "Figure 2. RGB-NIR stereo dataset. (a) Our dataset comprises stereo RGB image pairs, stereo NIR image pairs, and sparse depth point clouds, all captured from continuous video sequences. (b) A quantitative analysis of our dataset is presented, illustrating the distribution of frames under varying lighting conditions, a histogram showing the number of frames per video sequence, and the frame exposure time distributions for RGB and NIR sensors under three distinct lighting scenarios." + ], + "image_footnote": [], + "bbox": [ + 392, + 263, + 712, + 383 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2467d5d1632b6bad731ac6be6c9441cf88192a0621712efa78bd7d07d1ccdf1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 263, + 888, + 369 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3daeb539e1be2c69fbeaafbb0ee3c68d4fe9261d1405af47907076e20baa4cb1.jpg", + "table_caption": [ + "Table 1. RGB-NIR image dataset comparison. Our dataset is the only dataset with pixel-aligned RGB-NIR stereo image pairs." + ], + "table_footnote": [], + "table_body": "
DatasetsRGB-NIR imagesGT depthEnvironments
Pixel alignedMulti viewRGB stereoNIR stereo
[6]OOXXOOutdoor, day
[50]XOXXXOutdoor, day, night
[53]XOOXOOutdoor, day
[51]OXXXXOutdoor, day, night
[63]XOXXXOutdoor, sunny, dark
[8]XOXXXOutdoor, day, low-light
[48]XOOOOOutdoor, day, night, rain
[41]OOOXOOutdoor, offload, day
Ours (syn.)OOOOOOutdoor, indoor, day, night
Ours (real)OOOOOOutdoor, indoor, day, night
", + "bbox": [ + 78, + 488, + 465, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ground-truth dense depth. Figure 2(a) shows a sample of synthetic scenes. We compute the RGB reflectance map $R_{\\mathrm{RGB}}$ and the NIR reflectance map and $R_{\\mathrm{NIR}}$ using the RGB-to-NIR color synthesis method [16]. We then simulate diverse environmental lighting and obtain the RGB-NIR stereo images using Equation (1). Our augmented dataset comprising 2,238 short sequences (10 frames each) and four long sequences with total 2,200 frames. Our augmentation method is detailed in the supplemental material.", + "bbox": [ + 75, + 691, + 468, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset Comparison Table 1 compares our dataset with existing RGB-NIR stereo datasets in terms of alignment, environment diversity, and the availability of ground truth", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "depth. Unlike datasets requiring color transformation or material segmentation [8, 63], we capture stereo images directly in each spectral band, enhancing quality and reducing computational demands. While others need pose transformations due to differing camera positions [29, 48], our system uses pixel-aligned RGB-NIR camera. By overcoming these challenges, our dataset offers high-quality RGB-NIR stereo images suitable for various applications, including autonomous vehicles and robotics.", + "bbox": [ + 496, + 455, + 892, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Pixel-aligned RGB-NIR Fusion", + "text_level": 1, + "bbox": [ + 500, + 607, + 785, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce two methods fusing pixel-aligned RGB-NIR images to improve performance of vision models.", + "bbox": [ + 498, + 633, + 890, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Image Fusion with Learned Weights", + "text_level": 1, + "bbox": [ + 500, + 675, + 815, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we fuse pixel-aligned RGB-NIR images into a single three-channel image that can be fed to vision models pretrained for RGB images without finetuning. Figure 3 shows the overview of our image-fusion method.", + "bbox": [ + 498, + 699, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baseline As a baseline, we use the HSV channel blending method [15] to enhance performance on vision application without losing photometric consistency of both RGB and NIR. Given RGB and NIR images for the camera $c$ , the baseline method converts the RGB image into HSV representation:", + "bbox": [ + 496, + 780, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ I _ {H} ^ {c}, I _ {S} ^ {c}, I _ {V} ^ {c} \\right] ^ {\\intercal} = M \\left[ I _ {R} ^ {c}, I _ {G} ^ {c}, I _ {B} ^ {c} \\right] ^ {\\intercal}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 883, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11485", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7ae64f2252e6b081b2ae0728505fec4063e4eb267cc9bc6726e6a18b73229188.jpg", + "image_caption": [ + "Figure 3. RGB-NIR image fusion. We fuse the RGB image $I_{\\mathrm{RGB}}$ and the image $I_{\\mathrm{NIR}}$ as a weighted sum in the brightness domain $V$ after converting RGB image into HSV channel. The spatially-varying weights $\\alpha, \\beta$ are learned to effectively fuse RGB and NIR images. The fused image can be used as inputs to vision models such as object detection, stereo depth estimation, and structure from motion." + ], + "image_footnote": [], + "bbox": [ + 80, + 87, + 890, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M$ is the RGB-to-HSV conversion matrix. $I_H^c, I_S^c, I_V^c$ are the hue, saturation, and brightness images. We then combine the brightness channel $I_V^c$ with the NIR image $I_{\\mathrm{NIR}}^c$ to include details from both while the hue and saturation channels remain unchanged. The fused HSV image is converted back to the RGB domain obtaining the fused image $I_{\\mathrm{fusion}}^c$ :", + "bbox": [ + 76, + 385, + 468, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\text {f u s i o n}} ^ {c} = M ^ {- 1} \\left[ I _ {H} ^ {c}, I _ {S} ^ {c}, \\alpha I _ {V} ^ {c} + \\beta I _ {\\mathrm {N I R}} ^ {c} \\right] ^ {\\intercal} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 503, + 468, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are fusion weights.", + "bbox": [ + 76, + 532, + 305, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The baseline method has a limitation because the weights $\\alpha$ and $\\beta$ are predefined. This makes the fusion process non-adaptable to various environments and lighting conditions which make image details in NIR and RGB images vary significantly.", + "bbox": [ + 75, + 547, + 468, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learned Spatially-varying Weights Instead of using predefined constants $\\alpha$ and $\\beta$ , we propose to learn spatially-varying $\\alpha$ and $\\beta$ using an attention-based MLP. We first encode each image into a 256-channel feature map as follows:", + "bbox": [ + 76, + 641, + 468, + 702 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {e n c}} \\left(I _ {\\mathrm {R G B}} ^ {c}\\right) = F _ {\\mathrm {R G B}} ^ {c} \\quad \\text {a n d} \\quad f _ {\\mathrm {e n c}} \\left(I _ {\\mathrm {N I R}} ^ {c}\\right) = F _ {\\mathrm {N I R}} ^ {c} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 715, + 468, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{enc}}$ is a ResNet-based feature extractor. For the NIR feature encoding, we copy the NIR channel three times as an input to the pretrained three-channel feature extractor. We then obtain fused feature map $F_{\\mathrm{fusion}}^{c}$ by combining $F_{\\mathrm{RGB}}^{c}$ and $F_{\\mathrm{NIR}}^{c}$ using attentional fusion:", + "bbox": [ + 76, + 739, + 467, + 814 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {f u s i o n}} ^ {c} = f _ {\\text {f u s i o n}} \\left(F _ {\\mathrm {R G B}} ^ {c}, F _ {\\mathrm {N I R}} ^ {c}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 827, + 468, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the feature fusion module $f_{\\mathrm{fusion}}$ computes a sum of RGB and NIR feature maps with feature attention [13]. Finally, spatially varying weights $\\alpha$ and $\\beta$ are derived by the", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "decoder $f_{\\mathrm{dec}}$ consisting of several convolution layers:", + "bbox": [ + 500, + 386, + 849, + 401 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {d e c}} \\left(F _ {\\text {f u s i o n}} ^ {c}\\right) = \\{\\alpha , \\beta \\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 414, + 890, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the spatially-varying scene-dependent values of $\\alpha, \\beta$ are obtained, we estimate the fused image using Equation (3). Then, we filter $I_{\\mathrm{fusion}}^c$ using guided filtering with the reference NIR image $I_{\\mathrm{NIR}}^c$ [31].", + "bbox": [ + 498, + 441, + 892, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our RGB-NIR image fusion method runs at 100 fps on a laptop with RTX4060, operating at a resolution of $720 \\times 480$ . We evaluate the proposed method on object detection, depth estimation, and structure from motion described in Section 5. We only train the image-fusion models using photometric loss and stereo depth reconstruction loss, without training the downstream vision models. Refer to the Supplemental Document for details.", + "bbox": [ + 498, + 503, + 890, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Feature Fusion for Stereo Depth Estimation", + "text_level": 1, + "bbox": [ + 500, + 633, + 870, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the image fusion method proposed in Section 4.1 allows the use of existing vision models without finetuning, it reduces the image dimensions from four channels (RGB and NIR) to three output channels, which may result in information loss.", + "bbox": [ + 498, + 657, + 890, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we develop a method that fuses RGB and NIR features and finetunes a downstream model. We use the RAFT-Stereo network [36] as an example. Figure 4 shows a diagram of our proposed stereo depth estimation network. We extract features for the RGB stereo images $I_{\\mathrm{RGB}}^c$ and NIR stereo images $I_{\\mathrm{NIR}}^c$ , where $c \\in \\{\\text{left}, \\text{right}\\}$ , using the same ResNet-based feature extractor: $F_{\\mathrm{RGB}}^c = f_{\\mathrm{enc}}(I_{\\mathrm{RGB}}^c)$ , $F_{\\mathrm{NIR}}^c = f_{\\mathrm{enc}}(I_{\\mathrm{NIR}}^c)$ . We then apply the attention-based fusion method from Section 4.1: $F_{\\mathrm{fusion}}^c = f_{\\mathrm{fusion}}(F_{\\mathrm{RGB}}^c, F_{\\mathrm{NIR}}^c)$ [13].", + "bbox": [ + 498, + 734, + 890, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compute correlation volumes $V_{s\\in \\{\\mathrm{fusion,NIR}\\}}$ for not", + "bbox": [ + 517, + 886, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11486", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ec6670c65c016f4e6653b5ed86bbb06f87b5e9b84ac37f2b144d0d69f3a9f3c.jpg", + "image_caption": [ + "Figure 4. RGB-NIR stereo depth estimation model. We modified RAFT-Stereo [36] with attentional feature fusion and alternative correlation search for RGB-NIR depth estimation. We extract features from RGB and NIR images, fuse them, and build cost volumes. We estimate disparity by repeatedly feeding the cost volume of fused features and cost volume of NIR features to the GRU unit whose hidden state is initialized with $F_{\\mathrm{fusion}}^{\\mathrm{left}}$ ." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 467, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "only the fused features and but also NIR features:", + "bbox": [ + 76, + 375, + 403, + 388 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nV _ {s} (x, y, k) = F _ {s} ^ {\\text {l e f t}} (x, y) \\cdot F _ {s} ^ {\\text {r i g h t}} (x + k, y), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 396, + 468, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $(x,y)$ denotes the pixel location, $k$ is the disparity, and $\\cdot$ represents the inner product. We found that using the two correlation volumes for the fused features and the NIR features effectively utilize cross-spectral information, resulting in better depth accuracy as shown in Table 5.", + "bbox": [ + 75, + 421, + 468, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We estimate a series of disparity maps $\\{d_1,\\dots ,d_N\\}$ using the GRU structure of the RAFT-Stereo network [36]. By feeding $F_{\\mathrm{fusion}}^{\\mathrm{left}}$ as context feature in the RAFT-Stereo, we alternate between the fused and NIR correlation volumes, $V_{\\mathrm{fusion}}$ and $V_{\\mathrm{NIR}}$ as input to the GRU at each iteration, inspired by the cross-spectral time-of-flight imaging method [5].", + "bbox": [ + 75, + 497, + 468, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method leverages spectral information from both RGB and NIR features, placing more emphasis on NIR images that are robust to environmental lighting. For more details on the loss functions and optimization, we refer to the Supplemental Document.", + "bbox": [ + 75, + 603, + 468, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 76, + 690, + 161, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the evaluation of our proposed pixel-aligned RGB-NIR image fusion and feature fusion methods, along with comprehensive ablation studies.", + "bbox": [ + 75, + 715, + 468, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Pixel-aligned RGB-NIR Image Fusion", + "text_level": 1, + "bbox": [ + 76, + 768, + 403, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate our pixel-aligned RGB-NIR image fusion method in Section 4.1, we conducted assessments across three downstream tasks: depth estimation, object detection, and structure from motion.", + "bbox": [ + 75, + 792, + 468, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Depth Estimation Figure 5 shows that using our fused image as an input to the pre-trained stereo depth esti", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/37ee7886c42f4a0059420d5de8f04a63b34060974c15bdb21dd882bac974e748.jpg", + "table_caption": [ + "Table 2. Comparison of image fusion methods for pretrained RAFT-stereo [36] and YOLO [45]. Our RGB-NIR image fusion method outperforms other image-fusion methods for (a) stereo depth estimation and (b) object detection." + ], + "table_footnote": [], + "table_body": "
Methods(a) Depth MAE [m] ↓(b) Detection mAP ↑
RGB5.3110.756
NIR4.2400.703
YCrCb [20]5.2090.571
Bayesian[62]7.4090.745
DarkVision[25]6.1960.762
Adaptive[1]5.5010.773
VGG-NIR[30]5.3840.726
HSV(our baseline) [15]4.1610.744
Ours4.0340.809
", + "bbox": [ + 501, + 148, + 887, + 308 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a9b4ef6013c07ee7e5eb60e1c1cadc29b172ccdee8ddd27b7109e784afb563f2.jpg", + "table_caption": [ + "Table 3. Depth accuracy under diverse environments. Our image and feature fusion methods (last two rows) enable accurate depth estimation, outperforming single-modality RGB or NIR inputs (first two rows) as well as HSV-based RGB-NIR image fusion [15] (third row). RAFT-Stereo [36] is employed across all methods as the depth estimation network." + ], + "table_footnote": [], + "table_body": "
Metric: depth MAE [m]↓Day outdoorNight outdoorCloudy outdoorWell-lit indoorDark indoorComplex indoor
RGB5.7107.4913.1013.2923.2014.698
NIR4.5765.5633.1393.0622.3334.271
HSV fusion [15]4.4685.4173.1213.0912.2934.293
Our image fusion4.3565.1562.8162.9902.4314.181
Our feature fusion3.6514.7712.8382.6952.0593.920
", + "bbox": [ + 501, + 402, + 888, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mation network [36] yields improved results over single-modality predictions using only RGB or NIR images. Table 2(a) shows quantitative analysis compared to those single-modality predictions as well as using other RGB-NIR image fusion methods. Our proposed RGB-NIR image fusion outperforms all the baselines. Refer to the Supplemental Document for corresponding qualitative results.", + "bbox": [ + 496, + 497, + 890, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Detection We use theYOLO object detector [45] to test our RGB-NIR image fusion method. Figure 6 and Table 2(b) show that our image fusion method consistently outperforms single-modality predictions of either RGB or NIR, and other RGB-NIR image fusion methods.", + "bbox": [ + 496, + 621, + 890, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Structure from Motion Figure 7 shows that our RGB-NIR fused images enables robust reconstruction for a structure-from-motion method [46] on challenging lighting conditions, where RGB and NIR images provide different scene visibility. This suggested the potential utility of our fused RGB-NIR images for downstream applications, including scene reconstruction and view synthesis.", + "bbox": [ + 496, + 715, + 890, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Pixel-aligned RGB-NIR Feature Fusion", + "text_level": 1, + "bbox": [ + 500, + 830, + 836, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 8 shows that our feature-based RGB-NIR stereo method (Section 4.2) outperforms not only the single-modality predictions using either RGB or NIR channels,", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11487", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1596889a46cf11b1c298ea5073aab3fc4b1611cc76cca769648d3d476e0180f6.jpg", + "image_caption": [ + "(a) RGB" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 227, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/207efbf37db92f0dae928758c1b40bb9d7b3a0f1bf3a13db3117f33af22610da.jpg", + "image_caption": [ + "(b) NIR" + ], + "image_footnote": [], + "bbox": [ + 227, + 88, + 357, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5945d8199d38be3a1e03aa7a5cd2a27991ea19fbe3e1033c211f8344daf177b8.jpg", + "image_caption": [ + "(c) Image fusion (ours)" + ], + "image_footnote": [], + "bbox": [ + 359, + 88, + 491, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4746cd236517eb8325cfa5de324ad9425de6313f8868a8ba073de0f359bb4e0e.jpg", + "image_caption": [ + "(d) Detailed view" + ], + "image_footnote": [], + "bbox": [ + 495, + 88, + 581, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a63c0b252faaccb5a6ab3ee5be3a9125f22f26129c2e203a6d51ff8d23104f9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 90, + 669, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/335deca18d0f39719bc2901f00ae43e71b21493cbb5d405c0f012c8893b5bf09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 90, + 756, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ab9838840bae7f36b5d0fb02d57c31f0f3318ce37b353a64bcfb0fa5521bdab.jpg", + "image_caption": [ + "(e) LiDAR (GT)" + ], + "image_footnote": [], + "bbox": [ + 759, + 90, + 888, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/567798df9bedff439ac758c92708ad680ec81a890f233062f232f14fecd02fa8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 304, + 277, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1b7b686e136cda264100de42d66ece8ec4be6c538c3a3481c8abe309ce2c61a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 402, + 277, + 502 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94e61948c75e6bf9f733ddc54b30941afcd36e4930c8edbdac57bc870286041f.jpg", + "image_caption": [ + "Figure 6. Image fusion for pretrained YOLO [45]. (a) Using the RGB images struggle with object detection in scenes with low light conditions or high saturation due to the under- and overexposed regions by the limited dynamic range of the cameras. (b) NIR images suffer from challenges due to variations in material reflectance and the absence of detailed chromatic information for robust object detection. (c) Our RGB-NIR image fusion integrates the strengths of both RGB and NIR modalities, enabling robust object detection without retraining the model." + ], + "image_footnote": [], + "bbox": [ + 81, + 502, + 277, + 601 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/297bc4a947b14e8479543e9375d69e18145fbc0e1c51673bbd5135facb417d36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 304, + 465, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02d1747eda0a77b883de036d375e71c118f307b2ef11a570bb040d39fd012b2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 402, + 465, + 502 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0da687a51e5e07f20400274034179c32346b26eff14744371fcd98ee6efdf1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 502, + 465, + 601 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4fd577da27649a67d696efab9a6a2627b274eccca6ee52320a25c6a1a226469f.jpg", + "table_caption": [ + "Table 4. Dependency on depth-estimation networks. Our image and feature fusion methods achieve accurate depth estimation across three tested depth networks [7, 36, 57]." + ], + "table_footnote": [], + "table_body": "
Depth MAE [m]↓RGBNIRHSV fusion [15]Our image fusionOur feature fusion
RAFT-Stereo [36]5.2724.3024.2194.0443.619
MoCha-Stereo [7]8.3176.0986.4195.3914.611
IGEV++ [57]6.1855.4025.2565.4872.446
", + "bbox": [ + 78, + 795, + 467, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "but also our RGB-NIR image-fusion method (Section 4.1). While using single-modality RGB/NIR images easily struggle with saturated areas and shadows, our RGB-NIR fused", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96a4cc3407715ed1820a7c87934e56d825478394a01499bb1ec9cf27260f7b5c.jpg", + "image_caption": [ + "Figure 5. Image fusion for pretrained RAFT-stereo [36]. (a)&(b) Using a single modality either RGB or NIR images often results in sub-optimal disparity estimation in challenging lighting conditions. (c) Our RGB-NIR fused image enables robust disparity estimation without finetuning the pretrained model. (d) Close-up views. (e) Ground-truth LiDAR sparse disparity.", + "(a) COLMAP reconstruction with RGB images" + ], + "image_footnote": [], + "bbox": [ + 498, + 301, + 660, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b0e609e973cf8614d1b846f90a09b40a105f06abf0d737e5d9d74a455d67027.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 303, + 890, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b67fcf32ce7dee1be94e4fcc2d4df20f782cb21f5c080aa88d5a69c98a26cfc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 417, + 656, + 522 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fca06e5ea3df449f442045b5f63dfdc846bdfe5f553d53e5b2452534e7abbb60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 421, + 890, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b3d0e975dec39dd254cd9876696d486fe95136ae1168dcf3c34400bfd59098b.jpg", + "image_caption": [ + "(b) COLMAP reconstruction with NIR images", + "(c) COLMAP reconstruction with our RGB-NIR fusion images", + "Figure 7. Image fusion for structure from motion [46]. Compared to using (a) RGB images and (b) NIR images only, our RGB-NIR fused images (c) enable accurate reconstruction of both regions indicated by red and green rectangles. The green region is located closer to the imaging system, where active illumination enhances the visibility of distinct features in the NIR domain. The red region is situated farther from the imaging system, where features are well illuminated in the RGB domain due to ambient light. (c) Our RGB-NIR image fusion integrates these complementary features, enabling robust 3D reconstruction." + ], + "image_footnote": [], + "bbox": [ + 501, + 536, + 666, + 643 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d85a5137f1bc4ffc9a5d4ffd27e23836e94a8bf0ddaaf1a6db9586f3a56d521c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 537, + 890, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "images provide better depth estimation without finetuning. Our feature-based stereo depth estimation method further improves reconstruction quality by effectively exploiting the strengths of RGB and NIR channels via finetuning.", + "bbox": [ + 496, + 823, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate depth estimation accuracy across diverse", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11488", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/454b089a9cd36a48ac2b739d57ea6df35527891a5e97d6370a47f6c146fe0498.jpg", + "image_caption": [ + "Figure 8. RGB-NIR feature-based depth estimation. (a)&(b) Using single-modality RGB/NIR images for stereo depth estimation [36] often fails to obtain high-quality depth. (c) Our RGB-NIR image-fusion method (Section 4.1) mitigates the issue without finetuning the RAFT stereo, although it does not fully compensate for the strong saturation seen in (a) and (b). (d) Our RGB-NIR feature-based method (Section 4.2) further improves accuracy using an attentional method, at the cost of additional training." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 385 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/04bd80954aaaa5932596e7544534476e2c61c6c60409235d7db07d0015820b8d.jpg", + "table_caption": [ + "Table 5. Ablation study on alternating correlation volumes. Using fused and NIR correlation volumes in an alternating manner provides the highest depth-estimation accuracy." + ], + "table_footnote": [], + "table_body": "
Methods on using the fused correlation volumeDepth error [m]↓
Fused correlation volume only7.440
Alternating RGB-NIR correlation volumes8.571
Alternating Fused-RGB-NIR correlation volumes7.426
Alternating Fused-NIR correlation volumes6.747
", + "bbox": [ + 78, + 508, + 465, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "real-world environments using our dataset. Table 3 shows that our image fusion and feature fusion methods outperform single-modality RGB or NIR inputs, as well as HSV-based RGB-NIR image fusion [15].", + "bbox": [ + 75, + 606, + 468, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also verify the robustness of our fusion methods with two other depth-estimation networks [7, 57] alongside RAFT-Stereo [36]. Table 4 shows that our proposed fusion methods achieve accurate estimation compared to single-modality inputs and HSV-based RGB-NIR fusion [15] across all three network architectures.", + "bbox": [ + 75, + 667, + 468, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alternating Correlation Volumes In Section 4.2, we investigate the effect of using different types of correlation volumes for stereo depth estimation [36]. Specifically, we experiment with several strategies: using only the fusion correlation volume, alternating between RGB and NIR correlation volumes, alternating among fusion, RGB, and NIR correlation volumes, and alternating between fusion and NIR correlation volumes. Table 5 shows that alternating", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fusion-NIR correlation volumes yields the highest accuracy compared to the other methods.", + "bbox": [ + 500, + 465, + 888, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 513, + 617, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have presented a pixel-aligned RGB-NIR stereo imaging system, collect datasets, and develop RGB-NIR image fusion for pretrained models without any fin-tuning and feature fusion methods with additional training for improved performance. We demonstrate the effectiveness of pixel-aligned RGB-NIR images and our RGB-NIR fusion methods for three downstream tasks: depth estimation, object detection, and structure from motion. We hope our work shows the potential of pixel-aligned RGB-NIR imaging.", + "bbox": [ + 496, + 541, + 890, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work It will be an interesting future work to bring the success of RGB-image generative models for pixel-aligned RGB-NIR images. Another interesting direction is to improve the accuracy of pixel-aligned RGB-NIR 3D imaging with LiDAR measurements.", + "bbox": [ + 496, + 720, + 890, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments Seung-Hwan Baek was supported by Korea NRF (RS-2023-00211658, RS-2024-00438532), Samsung Electronics, and Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT1801-52.", + "bbox": [ + 496, + 825, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11489", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohamed Awad, Ahmed Eliethy, and Hussein A Aly. Adaptive near-infrared and visible fusion for fast image enhancement. IEEE Transactions on Computational Imaging, 6:408-418, 2019. 2, 6", + "[2] Jean-Yves Bouguet. Camera calibration toolbox for matlab. http://www.vision.caltech.edu/bouguetj/calib_doc/, 2004.3", + "[3] David P Broadbent, Giorgia D'Innocenzo, Toby J Ellmers, Justin Parsler, Andre J Szameitat, and Daniel T Bishop. Cognitive load, working memory capacity and driving performance: A preliminary fNIRS and eye tracking study. Transportation research part F: traffic psychology and behaviour, 92:121-132, 2023. 1", + "[4] Matthew Brown and Sabine Susstrunk. Multi-spectral SIFT for scene category recognition. In CVPR 2011, pages 177-184. IEEE, 2011. 2", + "[5] Samuel Brucker, Stefanie Walz, Mario Bijelic, and Felix Heide. Cross-spectral gated-RGB stereo depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21654-21665, 2024. 1, 2, 6", + "[6] Nived Chebrolu, Philipp Lottes, Alexander Schaefer, Wera Winterhalter, Wolfram Burgard, and Cyril Stachniss. Agricultural robot dataset for plant classification, localization and mapping on sugar beet fields. The International Journal of Robotics Research, 36(10):1045-1052, 2017. 4", + "[7] Ziyang Chen, Wei Long, He Yao, Yongjun Zhang, Bingshu Wang, Yongbin Qin, and Jia Wu. Mocha-stereo: Motif channel attention network for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27768-27777, 2024. 1, 7, 8", + "[8] Gyeongmin Choe, Seong-Heum Kim, Sunghoon Im, Joon-Young Lee, Srinivasa G Narasimhan, and In So Kweon. RANUS: RGB and NIR urban scene dataset for deep scene parsing. IEEE Robotics and Automation Letters, 3(3):1808-1815, 2018. 2, 4", + "[9] Yukyung Choi, Namil Kim, Soonmin Hwang, Kibaek Park, Jae Shin Yoon, Kyounghwan An, and In So Kweon. Kaist multi-spectral day/night data set for autonomous and assisted driving. IEEE Transactions on Intelligent Transportation Systems, 19(3):934-948, Mar. 2018. 2", + "[10] David Connah, Mark Samuel Drew, and Graham David Finlayson. Spectral edge image fusion: Theory and applications. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision - ECCV 2014, page 65-80, Cham, 2014. Springer International Publishing. 2", + "[11] Camille Couprie, Clément Farabet, Laurent Najman, and Yann Lecun. Indoor semantic segmentation using depth information. In First International Conference on Learning Representations (ICLR 2013), pages 1-8, 2013. 2", + "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2", + "[13] Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. In 2021" + ], + "bbox": [ + 78, + 114, + 470, + 887 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Winter Conference on Applications of Computer Vision (WACV), page 3559-3568, Waikoloa, HI, USA, Jan. 2021. IEEE. 5", + "[14] Sri Aditya Deevi, Connor Lee, Lu Gan, Sushruth Nagesh, Gaurav Pandey, and Soon-Jo Chung. Rgb-x object detection via scene-specific fusion modules. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7366-7375, 2024. 2", + "[15] Clément Fredembach and Sabine Susstrunk. Colouring the near-infrared. In Color and imaging conference, volume 16, pages 176-182. Society of Imaging Science and Technology, 2008. 2, 4, 6, 7, 8", + "[16] Tobias Gruber, Frank Julca-Aguilar, Mario Bijelic, and Felix Heide. Gated2depth: Real-time dense lidar from gated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1506-1516, 2019. 4", + "[17] Yubin Guo, Xinlei Qi, Jin Xie, Cheng-Zhong Xu, and Hui Kong. Unsupervised cross-spectrum depth estimation by visible-light and thermal cameras. IEEE Transactions on Intelligent Transportation Systems, 24(10):10937-10947, Oct. 2023. 2", + "[18] Mark F Hansen, Gary A Atkinson, Lyndon N Smith, and Melvyn L Smith. 3d face reconstructions from photometric stereo using near infrared and visible light. Computer Vision and Image Understanding, 114(8):942-951, 2010. 1, 2", + "[19] María Herrera-Arellano, Hayde Peregrina-Barreto, and Iván Terol-Villalobos. Visible-nir image fusion based on top-hat transform. IEEE Transactions on Image Processing, 30:4962-4972, 2021. 2", + "[20] María A Herrera-Arellano, Hayde Peregrina-Barreto, and Iván Terol-Villalobos. Color outdoor image enhancement by v-nir fusion and weighted luminance. In 2019 IEEE International Autumn Meeting on Power, Electronics and Computing (ROPEC), pages 1-6. IEEE, 2019. 2, 6", + "[21] Yuchen Hong, Youwei Lyu, Si Li, Gang Cao, and Boxin Shi. Reflection removal with nir and rgb image feature fusion. IEEE Transactions on Multimedia, 2022. 1", + "[22] Xuanlun Huang, Chenyang Wu, Xiaolan Xu, Baishun Wang, Sui Zhang, Chihchiang Shen, Chiennan Yu, Jiaxing Wang, Nan Chi, Shaohua Yu, et al. Polarization structured light 3d depth image sensor for scenes with reflective surfaces. Nature Communications, 14(1):6855, 2023. 1", + "[23] Soonmin Hwang, Jaesik Park, Namil Kim, Yukyung Choi, and In So Kweon. Multispectral pedestrian detection: Benchmark dataset and baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1037-1045, 2015. 2", + "[24] Dong-Won Jang and Rae-Hong Park. Colour image dehazing using near-infrared fusion. IET Image Processing, 11(8):587-594, 2017. 2", + "[25] Shuangping Jin, Bingbing Yu, Minhao Jing, Yi Zhou, Jiajun Liang, and Renhe Ji. Darkvisionnet: Low-light imaging via rgb-nir fusion with deep inconsistency prior. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1104-1112, 2022. 2, 6", + "[26] Cheolkon Jung, Kailong Zhou, and Jiawei Feng. Fusionnet: Multispectral fusion of rgb and nir images using two stage convolutional neural networks. IEEE Access, 8:23912-23919, 2020. 2" + ], + "bbox": [ + 503, + 92, + 890, + 891 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11490", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Min H Kim, Todd Alan Harvey, David S Kittle, Holly Rushmeier, Julie Dorsey, Richard O Prum, and David J Brady. 3d imaging spectroscopy for measuring hyperspectral patterns on solid objects. ACM Transactions on Graphics (TOG), 31(4):1-11, 2012. 1", + "[28] Namil Kim, Yukyung Choi, Soonmin Hwang, and In So Kweon. Multispectral transfer network: Unsupervised depth estimation for all-day vision. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 2", + "[29] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. Vivid++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, July 2022. 2, 4", + "[30] Hui Li, Xiao-Jun Wu, and Josef Kittler. Infrared and visible image fusion using a deep learning framework. In 2018 24th international conference on pattern recognition (ICPR), pages 2705-2710. IEEE, 2018. 2, 6", + "[31] Shutao Li, Xudong Kang, and Jianwen Hu. Image fusion with guided filtering. IEEE Transactions on Image Processing, 22(7):2864-2875, July 2013. 2, 5", + "[32] Stan Z Li, RuFeng Chu, ShengCai Liao, and Lun Zhang. Illumination invariant face recognition using near-infrared images. IEEE Transactions on pattern analysis and machine intelligence, 29(4):627-639, 2007. 1", + "[33] Xuan Li, Fei Liu, Pingli Han, Shichao Zhang, and Xiaopeng Shao. Near-infrared monocular 3d computational polarization imaging of surfaces exhibiting nonuniform reflectance. Optics Express, 29(10):15616-15630, 2021. 1", + "[34] Zhuo Li, Hai-Miao Hu, Wei Zhang, Shiliang Pu, and Bo Li. Spectrum characteristics preserved visible and near-infrared image fusion algorithm. IEEE Transactions on Multimedia, 23:306-319, 2021. 2", + "[35] Mingyang Liang, Xiaoyang Guo, Hongsheng Li, Xiaogang Wang, and You Song. Unsupervised cross-spectral stereo matching by learning to synthesize. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8706-8713, 2019. 2", + "[36] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 1, 5, 6, 7, 8", + "[37] Gang Liu and Guohong Huang. Color fusion based on em algorithm for ir and visible image. In 2010 The 2nd International Conference on Computer and Automation Engineering (ICCAE), volume 2, page 253-258, Feb. 2010. 2", + "[38] Jingjing Liu, Shaoting Zhang, Shu Wang, and Dimitris N Metaxas. Multispectral deep neural networks for pedestrian detection. In 27th British Machine Vision Conference, BMVC 2016, 2016. 1", + "[39] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 3", + "[40] Yusuke Monno, Hayato Teranaka, Kazunori Yoshizaki, Masayuki Tanaka, and Masatoshi Okutomi. Single-sensor rgb-nir imaging: High-quality system design and prototype implementation. IEEE Sensors Journal, 19(2):497-507," + ], + "bbox": [ + 78, + 90, + 470, + 892 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2018. 2", + "[41] Peter Mortimer, Raphael Hagmanns, Miguel Granero, Thorsten Luettel, Janko Petereit, and Hans-Joachim Wuensch. The goose dataset for perception in unstructured environments. 2024. 4", + "[42] Chulhee Park and Moon Gi Kang. Color restoration of rgbn multispectral filter array sensor images based on spectral decomposition. Sensors, 16(5):719, 2016. 2", + "[43] Matteo Poggi, Pierluigi Zama Ramirez, Fabio Tosi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Cross-spectral neural radiance fields. In 2022 International Conference on 3D Vision (3DV), pages 606-616. IEEE, 2022. 1", + "[44] Long Quan and Zhongdan Lan. Linear n-point camera pose determination. IEEE Transactions on pattern analysis and machine intelligence, 21(8):774-780, 1999. 3", + "[45] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 6, 7", + "[46] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6, 7", + "[47] Takashi Shibata, Masayuki Tanaka, and Masatoshi Okutomi. Versatile visible and near-infrared image fusion based on high visibility area selection. Journal of Electronic Imaging, 25(1):013016-013016, 2016. 2", + "[48] Ukcheol Shin, Jinsun Park, and In So Kweon. Deep depth estimation from thermal image. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 1043-1053, Vancouver, BC, Canada, June 2023. IEEE. 1, 2, 4", + "[49] Haonan Su, Cheolkon Jung, and Long Yu. Multi-spectral fusion and denoising of color and near-infrared images using multi-scale wavelet analysis. Sensors, 21(11):3610, 2021. 2", + "[50] Karasawa Takumi, Kohei Watanabe, Qishen Ha, Antonio Tejero-De-Pablos, Yoshitaka Ushiku, and Tatsuya Harada. Multispectral object detection for autonomous vehicles. In Proceedings of the on Thematic Workshops of ACM Multimedia 2017, pages 35–43, 2017. 1, 4", + "[51] Alexander Toet, Maarten A Hogervorst, and Alan R Pinkus. The triclobs dynamic multi-band image data set for the development and evaluation of image fusion methods. PloS one, 11(12):e0165016, 2016. 4", + "[52] Alexander Toet and Jan Walraven. New false color mapping for image fusion. Optical engineering, 35(3):650-658, 1996. 2", + "[53] Abhinav Valada, Gabriel L Oliveira, Thomas Brox, and Wolfram Burgard. Deep multispectral semantic scene understanding of forested environments using multimodal fusion. In 2016 International Symposium on Experimental Robotics, pages 465-477. Springer, 2017. 4", + "[54] Stefanie Walz, Mario Bijelic, Andrea Ramazzina, Amanpreet Walia, Fahim Mannan, and Felix Heide. Gated stereo: Joint depth estimation from gated and wide-baseline active stereo cues. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 13252-13262, Vancouver, BC, Canada, June 2023. IEEE. 1", + "[55] Linbo Wang, Tao Wang, Deyun Yang, Xianyong Fang, and" + ], + "bbox": [ + 503, + 92, + 890, + 890 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11491", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shaohua Wan. Near-infrared fusion for deep lightness enhancement. International Journal of Machine Learning and Cybernetics, 14(5):1621-1633, May 2023. 2", + "[56] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 2", + "[57] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: iterative multi-range geometry encoding volumes for stereo matching. arXiv preprint arXiv:2409.00638, 2024. 1, 7, 8", + "[58] Karmesh Yadav, Ram Ramrakhya, Santhosh Kumar Ramakrishnan, Theo Gervet, John Turner, Aaron Gokaslan, Noah Maestre, Angel Xuan Chang, Dhruv Batra, Manolis Savva, et al. Habitat-matterport 3d semantics dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4927-4936, 2023. 2", + "[59] Xue Zhang, Xiaohan Zhang, Jiangtao Wang, Jiacheng Ying, Zehua Sheng, Heng Yu, Chunguang Li, and Hui-Liang Shen. Tfdet: Target-aware fusion for rgb-t pedestrian detection. IEEE Transactions on Neural Networks and Learning Systems, page 1-15, 2024. 2", + "[60] Yigong Zhang, Yicheng Gao, Shuo Gu, Yubin Guo, Minghao Liu, Zezhou Sun, Zhixing Hou, Hang Yang, Ying Wang, Jian Yang, Jean Ponce, and Hui Kong. Build your own hybrid thermal/eo camera for autonomous vehicle. In 2019 International Conference on Robotics and Automation (ICRA), page 6555-6560, May 2019. 2", + "[61] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 1", + "[62] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, and Jiangshe Zhang. Bayesian fusion for infrared and visible images. Signal Processing, 177:107734, 2020. 2, 6", + "[63] Tiancheng Zhi, Bernardo R Pires, Martial Hebert, and Srinivasa G Narasimhan. Deep material-aware cross-spectral stereo matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1916-1925, 2018. 1, 2, 4" + ], + "bbox": [ + 78, + 90, + 470, + 662 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11492", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/640cd516-44b5-47ce-aa6a-b65d88b8da0a_model.json b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/640cd516-44b5-47ce-aa6a-b65d88b8da0a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d6dfd625f5333113b55fe5ea1d1b2feeb11262b8 --- /dev/null +++ b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/640cd516-44b5-47ce-aa6a-b65d88b8da0a_model.json @@ -0,0 +1,2631 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.13, + 0.834, + 0.152 + ], + "angle": 0, + "content": "Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.181, + 0.638, + 0.216 + ], + "angle": 0, + "content": "Jinnyeong Kim Seung-Hwan Baek POSTECH" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.251, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.284, + 0.474, + 0.617 + ], + "angle": 0, + "content": "Integrating RGB and NIR stereo imaging provides complementary spectral information, potentially enhancing robotic 3D vision in challenging lighting conditions. However, existing datasets and imaging systems lack pixel-level alignment between RGB and NIR images, posing challenges for downstream vision tasks. In this paper, we introduce a robotic vision system equipped with pixel-aligned RGB-NIR stereo cameras and a LiDAR sensor mounted on a mobile robot. The system simultaneously captures pixel-aligned pairs of RGB stereo images, NIR stereo images, and temporally synchronized LiDAR points. Utilizing the mobility of the robot, we present a dataset containing continuous video frames under diverse lighting conditions. We then introduce two methods that utilize the pixel-aligned RGB-NIR images: an RGB-NIR image fusion method and a feature fusion method. The first approach enables existing RGB-pretrained vision models to directly utilize RGB-NIR information without fine-tuning. The second approach fine-tunes existing vision models to more effectively utilize RGB-NIR information. Experimental results demonstrate the effectiveness of using pixel-aligned RGB-NIR images across diverse lighting conditions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.647, + 0.21, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.869 + ], + "angle": 0, + "content": "RGB imaging captures visible light in wavelengths from \\(400\\mathrm{nm}\\) to \\(700\\mathrm{nm}\\) and serves as the primary data format for visual computing. Near-infrared (NIR) imaging with active illumination captures light in the range of \\(750\\mathrm{nm}\\) to \\(1000\\mathrm{nm}\\), offering effective information in conditions where where RGB imaging is inadequate, such as low-light environments during night or poorly lit indoor conditions. NIR light is imperceptible to human vision, allowing for using active NIR illumination without disturbing people, which facilitates robust 3D imaging [5, 32, 38, 54] and eye tracking [3]. Additionally, NIR light offers supplementary scene information due to different spectral reflectance between RGB and NIR wavelengths [18, 27, 33]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Combining these benefits, RGB-NIR imaging emerges as a promising modality for challenging environments re" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.253, + 0.895, + 0.435 + ], + "angle": 0, + "content": "quiring robust and efficient 2D and 3D vision capabilities. RGB-NIR imaging has been studied for applications including object detection [50, 61] and 3D reconstruction [5, 21, 22, 43]. However, existing RGB-NIR imaging systems and datasets face significant challenges compared to standard RGB vision systems. A major issue is the use of separate RGB and NIR cameras at different poses, leading to pixel misalignment between the NIR and RGB images. Previous methods address such misalignments using image registration, pose estimation, and other processing techniques, which can accumulate errors and limit effectiveness [5, 48, 63]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.895, + 0.649 + ], + "angle": 0, + "content": "In this work, we develop a robotic vision system capable of capturing pixel-aligned RGB-NIR stereo images and synchronized LiDAR point clouds. Mounted on a mobile platform, the system enables exploration in both indoor and outdoor environments, facilitating data collection under diverse lighting conditions. Using this system, we collect and share a dataset of pixel-aligned RGB-NIR stereo images with synchronized LiDAR point clouds as ground-truth depth for 82 real-world scenes including 90,000 frames. Additionally, we provide a synthetic RGB-NIR dataset with ground-truth dense depth labels for 2,238 synthetic scenes. To demonstrate the effectiveness of pixel-aligned RGB-NIR imaging and datasets, we propose an RGB-NIR image fusion method and an RGB-NIR feature fusion method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.896, + 0.804 + ], + "angle": 0, + "content": "The first method fuses the pixel-aligned RGB-NIR images into a three-channel image that can be input into pretrained RGB vision models without finetuning. The second method involves developing a stereo depth estimation technique by finetuning pretrained stereo networks [7, 36, 57] to utilize pixel-aligned RGB-NIR features. Our experiments show improvements in various downstream tasks over using only RGB or NIR data, as well as over methods relying on pixel-misaligned RGB-NIR datasets, particularly in challenging scenarios." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.807, + 0.821, + 0.822 + ], + "angle": 0, + "content": "In summary, our contributions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.825, + 0.892, + 0.884 + ], + "angle": 0, + "content": "- We develop a robotic vision system with pixel-aligned RGB-NIR stereo cameras and LiDAR, capturing synchronized RGB-NIR stereo images and LiDAR point clouds under diverse lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- We present a large-scale dataset of pixel-aligned RGB-" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11482" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "NIR stereo images and LiDAR point clouds, collected in various indoor and outdoor environments." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.122, + 0.468, + 0.165 + ], + "angle": 0, + "content": "- We propose an RGB-NIR image fusion method and an RGB-NIR stereo depth estimation network using RGB-NIR feature fusion." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.167, + 0.468, + 0.213 + ], + "angle": 0, + "content": "- We evaluate the proposed methods on diverse downstream tasks and illumination conditions, on synthetic and real-world datasets, outperforming baselines." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.122, + 0.468, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.229, + 0.218, + 0.244 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.255, + 0.468, + 0.436 + ], + "angle": 0, + "content": "RGB-NIR Imaging Conventional RGB-NIR imaging systems employ separate RGB and NIR cameras positioned at different viewpoints, such as the Kinect sensor, Intel RealSense D415, and various multi-view RGB-NIR systems [8, 9, 11, 12, 29, 48, 56, 58, 63]. However, the captured RGB and NIR images from these systems exhibit pixel misalignment, necessitating depth-dependent registration between the RGB and NIR images. This introduces a significant challenge for effectively utilizing RGB and NIR information, creating a chicken-and-egg problem between depth estimation using spectrally fused images and spectral pose alignment using estimated depth." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.438, + 0.468, + 0.649 + ], + "angle": 0, + "content": "To address the pixel misalignment, sequential capture with interchangeable RGB and NIR filters has been used [4, 25]; however, this approach is limited to capturing static scenes due to the time delay between captures. Designing custom color filter arrays for image sensors to record four channels (RGB-NIR) has also been explored [40]. However, using a single sensor with a fixed global exposure for all four channels makes it unsuitable for challenging environments where the dynamic ranges for RGB and NIR channels differ significantly, such as in dark indoor or outdoor conditions with active NIR illumination. In the domain of RGB-thermal imaging, beam splitters have been used to combine an RGB camera and a thermal camera, enabling pixel-aligned RGB-thermal imaging [23, 28, 60]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.65, + 0.468, + 0.739 + ], + "angle": 0, + "content": "We capture pixel-aligned RGB-NIR stereo images alongside LiDAR point clouds using two prism-based RGB-NIR dual-sensor cameras. The system is mounted on a mobile robot, allowing scalable data acquisition of pixel-aligned RGB-NIR stereo images in both indoor and outdoor conditions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "RGB-NIR Fusion To utilize information from both RGB and NIR spectra, fusing RGB-NIR images has been extensively studied with applications in image enhancement for long-distance visibility and fog penetration [1, 10, 20, 62], as well as in information visualization leveraging different RGB-NIR reflectance properties [18, 19, 34, 47]. Conventional methods apply simple arithmetic operations such as addition and subtraction in RGB, HSL, HSV, and YUV color spaces to create images that contain both RGB and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "NIR information [15, 19, 20, 37, 42, 52]. Edge-aware filtering techniques have been employed to refine noisy and low-contrast RGB images using guidance from clean and high-contrast NIR images, showing promising performance in various imaging scenarios [10, 24, 31, 34]. More recently, learning-based approaches have attempted to enhance RGB images and their 3D vision applications using NIR guidance by fusing RGB-NIR features [25, 26, 30, 49, 55] or RGB-thermal features [14, 23, 59]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.318 + ], + "angle": 0, + "content": "Leveraging our pixel-aligned RGB-NIR images, we develop a RGB-NIR image fusion method that effectively encodes information from both RGB and NIR images into a single three-channel image with learned spatially-varying weights. The resulting fused image can be directly used as input to neural networks pretrained on RGB images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.622 + ], + "angle": 0, + "content": "RGB-NIR Depth Estimation Cross-spectral correspondence matching between RGB and NIR images using an RGB-NIR color transform network has demonstrated promising depth estimation results when the RGB-NIR reflectances are similar [63]. Liang et al. showed that GAN based architecture can enhance RGB-NIR color transform on cross-spectral stereo matching [35]. Active stereo systems using structured NIR illumination, such as the Kinect and Intel RealSense D415, employ multiple RGB and NIR cameras at different viewpoints. These systems often suffer from the aforementioned chicken-and-egg problem between depth estimation using spectrally fused images and spectral fusion using the estimated depth. Brucker et al. [5] fuse red, blue, and clean channels of stereo RCCB cameras with gating-based time-of-flight cameras positioned differently but still face a similar chicken-and-egg problem. In the domain of RGB-thermal imaging, Guo et al. [17] demonstrate effective depth estimation using pixel-aligned RGB-thermal stereo cameras with beam splitters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.682 + ], + "angle": 0, + "content": "We develop a depth estimation method leveraging our pixel-aligned RGB-NIR images for robust depth estimation, thereby bypassing the chicken-and-egg problem between depth estimation and spectral fusion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.855, + 0.712 + ], + "angle": 0, + "content": "3. Pixel-aligned RGB-NIR Stereo Imaging" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Robotic Imaging System Figure 1(a) illustrates our robotic imaging system, which integrates a stereo pair of pixel-aligned RGB-NIR cameras, NIR illumination, and a LiDAR mounted on a mobile robot (AgileX Ranger Mini 2.0). Each RGB-NIR camera (JAI FS-1600D-10GE), as shown in Figure 1(c), employs a dichroic prism to separate RGB and NIR light, independently captured by RGB and NIR CMOS sensors. To enhance NIR visibility under varying lighting conditions, we include active NIR illumination (Advanced Illumination AL295-150850IC) without disturbing human vision. The spectral sensitivities of the camera and the illumination are shown in Figure 1(b). We" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11483" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.089, + 0.338, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.293, + 0.395, + 0.305 + ], + "angle": 0, + "content": "(a) Imaging system" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.089, + 0.594, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.089, + 0.882, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.293, + 0.833, + 0.305 + ], + "angle": 0, + "content": "(c) Pixel-aligned RGB-NIR camera" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.308, + 0.894, + 0.352 + ], + "angle": 0, + "content": "Figure 1. RGB-NIR stereo imaging system. (a) Stereo camera setup integrated with LiDAR and active NIR illumination. (b) Spectral sensitivity profiles of the camera sensors and the irradiance profile of the NIR illumination. (c) Diagram of the pixel-aligned RGB-NIR camera featuring a dichroism prism, where RGB light penetrates the prism and NIR light is reflected, achieving spectral separation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.473, + 0.514 + ], + "angle": 0, + "content": "installed a LiDAR (OUSTER OS1) to acquire ground-truth sparse depth maps, along with angular velocity and linear acceleration data from its inertial measurement unit. We use RJ-45 interfaces for the data transmission of the cameras and LiDAR. We set up an AC powerbank to supply power to the devices that require an AC adapter. These devices are connected to a laptop, which triggers the capture pipeline and manages real-time data storage. We manage separate threads on the laptop for the left camera, right camera, and LiDAR, acquiring synchronized data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.472, + 0.657 + ], + "angle": 0, + "content": "Image Formation We model the image formation of the pixel-aligned RGB-NIR camera, which captures RGB and NIR images on separate CMOS sensors. This configuration results in the same exposure time and gain across the RGB channels, with separate exposure time and gain for the NIR channel: \\( t_R = t_G = t_B \\neq t_{\\mathrm{NIR}} \\) and \\( g_R = g_G = g_B \\neq g_{\\mathrm{NIR}} \\). The image captured by camera \\( c \\in \\{\\text{left}, \\text{right}\\} \\) for channel \\( i \\in \\{\\text{R}, \\text{G}, \\text{B}, \\text{NIR}\\} \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.669, + 0.47, + 0.687 + ], + "angle": 0, + "content": "\\[\nI _ {i} ^ {c} (p ^ {c}) = \\eta_ {1} + g _ {i} \\left(\\eta_ {2} + t _ {i} \\left(R _ {i} ^ {c} (p ^ {c}) \\left(E _ {i} ^ {c} (p ^ {c}) + L _ {i} ^ {c} (p ^ {c})\\right)\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.835 + ], + "angle": 0, + "content": "where \\(p^c\\) is a pixel of camera \\(c\\), \\(R_i^c(p^c)\\) is the reflectance at pixel \\(p^c\\) for channel \\(i\\), and \\(\\eta_1, \\eta_2\\) are Gaussian noise in the post-gamma and pre-gamma stages, respectively. The intensities measured in each channel are influenced by environmental illumination, denoted as \\(E_i^c(p^c)\\) for channel \\(i\\). The active illumination \\(L_i^c(p^c)\\) is present only in the NIR channel, satisfying \\(L_i^c(p^c) = 0\\) for \\(i \\in \\{\\mathrm{R}, \\mathrm{G}, \\mathrm{B}\\}\\), since the active illumination is confined to the NIR spectrum (see Figure 1(c))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Calibration One advantage of using pixel-aligned cameras is that calibration between RGB and NIR images is unnecessary. Thus, we only calibrate the stereo cameras" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.892, + 0.454 + ], + "angle": 0, + "content": "using checkerboard calibration [2]. The cameras contain precision time protocol, which enables maintaining temporal difference between stereo cameras less than 1 microsecond. We also obtain the LiDAR extrinsics by using corresponding point pairs between left-camera images and LiDAR point clouds [44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.893, + 0.851 + ], + "angle": 0, + "content": "Real-world Dataset Using the system, we acquire a dataset consisting of 82 videos totaling 90,000 frames under various lighting conditions for training and 7000 frames for testing. For each frame, we provide pixel-aligned RGB-NIR stereo images, a sparse LiDAR point cloud, sensor timestamps, and exposure values for the RGB and NIR CMOS sensors. Each video was recorded at frame rates of \\(5 - 10\\mathrm{Hz}\\), with durations ranging from five to over thirty minutes. Figure 2(a) shows samples from our real dataset and augmented synthetic dataset. We categorize our dataset with lighting conditions: well-lit, dark and complex, based on indoor or outdoor environment. Figure 2(b) shows the statistics of our dataset. We use auto exposure on both RGB and NIR sensors independently, with fixed gain values, and recorded the exposure time for each image. The distribution of exposure times in our dataset varies depending on lighting conditions. For the indoor dataset, many frames have short exposure times for NIR images, as active illumination is sufficient to illuminate nearby indoor scenes. In the daytime dataset, strong sunlight causes the exposure times to shift towards very short durations under direct sunlight and longer durations in shaded areas. For the night dataset, although many frames exhibit long exposure times for both RGB and NIR, the average exposure time for NIR frames is shorter than that for RGB frames." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Synthetic Dataset We augmented synthetic RGB stereo-image dataset [39] to obtain RGB-NIR stereo dataset with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11484" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.227, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.092, + 0.355, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.092, + 0.483, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.092, + 0.611, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.092, + 0.739, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.092, + 0.893, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.263, + 0.229, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.265, + 0.391, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.265, + 0.714, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.265, + 0.89, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.389, + 0.894, + 0.445 + ], + "angle": 0, + "content": "Figure 2. RGB-NIR stereo dataset. (a) Our dataset comprises stereo RGB image pairs, stereo NIR image pairs, and sparse depth point clouds, all captured from continuous video sequences. (b) A quantitative analysis of our dataset is presented, illustrating the distribution of frames under varying lighting conditions, a histogram showing the number of frames per video sequence, and the frame exposure time distributions for RGB and NIR sensors under three distinct lighting scenarios." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.455, + 0.47, + 0.484 + ], + "angle": 0, + "content": "Table 1. RGB-NIR image dataset comparison. Our dataset is the only dataset with pixel-aligned RGB-NIR stereo image pairs." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.489, + 0.466, + 0.677 + ], + "angle": 0, + "content": "
DatasetsRGB-NIR imagesGT depthEnvironments
Pixel alignedMulti viewRGB stereoNIR stereo
[6]OOXXOOutdoor, day
[50]XOXXXOutdoor, day, night
[53]XOOXOOutdoor, day
[51]OXXXXOutdoor, day, night
[63]XOXXXOutdoor, sunny, dark
[8]XOXXXOutdoor, day, low-light
[48]XOOOOOutdoor, day, night, rain
[41]OOOXOOutdoor, offload, day
Ours (syn.)OOOOOOutdoor, indoor, day, night
Ours (real)OOOOOOutdoor, indoor, day, night
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.47, + 0.83 + ], + "angle": 0, + "content": "ground-truth dense depth. Figure 2(a) shows a sample of synthetic scenes. We compute the RGB reflectance map \\( R_{\\mathrm{RGB}} \\) and the NIR reflectance map and \\( R_{\\mathrm{NIR}} \\) using the RGB-to-NIR color synthesis method [16]. We then simulate diverse environmental lighting and obtain the RGB-NIR stereo images using Equation (1). Our augmented dataset comprising 2,238 short sequences (10 frames each) and four long sequences with total 2,200 frames. Our augmentation method is detailed in the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Dataset Comparison Table 1 compares our dataset with existing RGB-NIR stereo datasets in terms of alignment, environment diversity, and the availability of ground truth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.893, + 0.593 + ], + "angle": 0, + "content": "depth. Unlike datasets requiring color transformation or material segmentation [8, 63], we capture stereo images directly in each spectral band, enhancing quality and reducing computational demands. While others need pose transformations due to differing camera positions [29, 48], our system uses pixel-aligned RGB-NIR camera. By overcoming these challenges, our dataset offers high-quality RGB-NIR stereo images suitable for various applications, including autonomous vehicles and robotics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.608, + 0.787, + 0.626 + ], + "angle": 0, + "content": "4. Pixel-aligned RGB-NIR Fusion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.634, + 0.892, + 0.665 + ], + "angle": 0, + "content": "We introduce two methods fusing pixel-aligned RGB-NIR images to improve performance of vision models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.676, + 0.816, + 0.693 + ], + "angle": 0, + "content": "4.1. Image Fusion with Learned Weights" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.7, + 0.892, + 0.76 + ], + "angle": 0, + "content": "First, we fuse pixel-aligned RGB-NIR images into a single three-channel image that can be fed to vision models pretrained for RGB images without finetuning. Figure 3 shows the overview of our image-fusion method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Baseline As a baseline, we use the HSV channel blending method [15] to enhance performance on vision application without losing photometric consistency of both RGB and NIR. Given RGB and NIR images for the camera \\(c\\), the baseline method converts the RGB image into HSV representation:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.885, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\left[ I _ {H} ^ {c}, I _ {S} ^ {c}, I _ {V} ^ {c} \\right] ^ {\\intercal} = M \\left[ I _ {R} ^ {c}, I _ {G} ^ {c}, I _ {B} ^ {c} \\right] ^ {\\intercal}, \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11485" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.892, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.893, + 0.375 + ], + "angle": 0, + "content": "Figure 3. RGB-NIR image fusion. We fuse the RGB image \\( I_{\\mathrm{RGB}} \\) and the image \\( I_{\\mathrm{NIR}} \\) as a weighted sum in the brightness domain \\( V \\) after converting RGB image into HSV channel. The spatially-varying weights \\( \\alpha, \\beta \\) are learned to effectively fuse RGB and NIR images. The fused image can be used as inputs to vision models such as object detection, stereo depth estimation, and structure from motion." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.386, + 0.469, + 0.493 + ], + "angle": 0, + "content": "where \\(M\\) is the RGB-to-HSV conversion matrix. \\(I_H^c, I_S^c, I_V^c\\) are the hue, saturation, and brightness images. We then combine the brightness channel \\(I_V^c\\) with the NIR image \\(I_{\\mathrm{NIR}}^c\\) to include details from both while the hue and saturation channels remain unchanged. The fused HSV image is converted back to the RGB domain obtaining the fused image \\(I_{\\mathrm{fusion}}^c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.504, + 0.469, + 0.521 + ], + "angle": 0, + "content": "\\[\nI _ {\\text {f u s i o n}} ^ {c} = M ^ {- 1} \\left[ I _ {H} ^ {c}, I _ {S} ^ {c}, \\alpha I _ {V} ^ {c} + \\beta I _ {\\mathrm {N I R}} ^ {c} \\right] ^ {\\intercal} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.533, + 0.306, + 0.547 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are fusion weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.469, + 0.623 + ], + "angle": 0, + "content": "The baseline method has a limitation because the weights \\(\\alpha\\) and \\(\\beta\\) are predefined. This makes the fusion process non-adaptable to various environments and lighting conditions which make image details in NIR and RGB images vary significantly." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.642, + 0.469, + 0.703 + ], + "angle": 0, + "content": "Learned Spatially-varying Weights Instead of using predefined constants \\(\\alpha\\) and \\(\\beta\\), we propose to learn spatially-varying \\(\\alpha\\) and \\(\\beta\\) using an attention-based MLP. We first encode each image into a 256-channel feature map as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.717, + 0.469, + 0.733 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {e n c}} \\left(I _ {\\mathrm {R G B}} ^ {c}\\right) = F _ {\\mathrm {R G B}} ^ {c} \\quad \\text {a n d} \\quad f _ {\\mathrm {e n c}} \\left(I _ {\\mathrm {N I R}} ^ {c}\\right) = F _ {\\mathrm {N I R}} ^ {c} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.74, + 0.468, + 0.815 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{enc}} \\) is a ResNet-based feature extractor. For the NIR feature encoding, we copy the NIR channel three times as an input to the pretrained three-channel feature extractor. We then obtain fused feature map \\( F_{\\mathrm{fusion}}^{c} \\) by combining \\( F_{\\mathrm{RGB}}^{c} \\) and \\( F_{\\mathrm{NIR}}^{c} \\) using attentional fusion:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.828, + 0.469, + 0.844 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {f u s i o n}} ^ {c} = f _ {\\text {f u s i o n}} \\left(F _ {\\mathrm {R G B}} ^ {c}, F _ {\\mathrm {N I R}} ^ {c}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where the feature fusion module \\( f_{\\mathrm{fusion}} \\) computes a sum of RGB and NIR feature maps with feature attention [13]. Finally, spatially varying weights \\( \\alpha \\) and \\( \\beta \\) are derived by the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.387, + 0.851, + 0.402 + ], + "angle": 0, + "content": "decoder \\( f_{\\mathrm{dec}} \\) consisting of several convolution layers:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.415, + 0.891, + 0.431 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {d e c}} \\left(F _ {\\text {f u s i o n}} ^ {c}\\right) = \\{\\alpha , \\beta \\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.443, + 0.893, + 0.504 + ], + "angle": 0, + "content": "Once the spatially-varying scene-dependent values of \\(\\alpha, \\beta\\) are obtained, we estimate the fused image using Equation (3). Then, we filter \\(I_{\\mathrm{fusion}}^c\\) using guided filtering with the reference NIR image \\(I_{\\mathrm{NIR}}^c\\) [31]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.505, + 0.892, + 0.624 + ], + "angle": 0, + "content": "Our RGB-NIR image fusion method runs at 100 fps on a laptop with RTX4060, operating at a resolution of \\(720 \\times 480\\). We evaluate the proposed method on object detection, depth estimation, and structure from motion described in Section 5. We only train the image-fusion models using photometric loss and stereo depth reconstruction loss, without training the downstream vision models. Refer to the Supplemental Document for details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.871, + 0.651 + ], + "angle": 0, + "content": "4.2. Feature Fusion for Stereo Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.659, + 0.891, + 0.733 + ], + "angle": 0, + "content": "While the image fusion method proposed in Section 4.1 allows the use of existing vision models without finetuning, it reduces the image dimensions from four channels (RGB and NIR) to three output channels, which may result in information loss." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.892, + 0.886 + ], + "angle": 0, + "content": "Here, we develop a method that fuses RGB and NIR features and finetunes a downstream model. We use the RAFT-Stereo network [36] as an example. Figure 4 shows a diagram of our proposed stereo depth estimation network. We extract features for the RGB stereo images \\( I_{\\mathrm{RGB}}^c \\) and NIR stereo images \\( I_{\\mathrm{NIR}}^c \\), where \\( c \\in \\{\\text{left}, \\text{right}\\} \\), using the same ResNet-based feature extractor: \\( F_{\\mathrm{RGB}}^c = f_{\\mathrm{enc}}(I_{\\mathrm{RGB}}^c) \\), \\( F_{\\mathrm{NIR}}^c = f_{\\mathrm{enc}}(I_{\\mathrm{NIR}}^c) \\). We then apply the attention-based fusion method from Section 4.1: \\( F_{\\mathrm{fusion}}^c = f_{\\mathrm{fusion}}(F_{\\mathrm{RGB}}^c, F_{\\mathrm{NIR}}^c) \\) [13]." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.887, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We compute correlation volumes \\(V_{s\\in \\{\\mathrm{fusion,NIR}\\}}\\) for not" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11486" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.47, + 0.366 + ], + "angle": 0, + "content": "Figure 4. RGB-NIR stereo depth estimation model. We modified RAFT-Stereo [36] with attentional feature fusion and alternative correlation search for RGB-NIR depth estimation. We extract features from RGB and NIR images, fuse them, and build cost volumes. We estimate disparity by repeatedly feeding the cost volume of fused features and cost volume of NIR features to the GRU unit whose hidden state is initialized with \\( F_{\\mathrm{fusion}}^{\\mathrm{left}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.375, + 0.405, + 0.39 + ], + "angle": 0, + "content": "only the fused features and but also NIR features:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.397, + 0.469, + 0.415 + ], + "angle": 0, + "content": "\\[\nV _ {s} (x, y, k) = F _ {s} ^ {\\text {l e f t}} (x, y) \\cdot F _ {s} ^ {\\text {r i g h t}} (x + k, y), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.469, + 0.497 + ], + "angle": 0, + "content": "where \\((x,y)\\) denotes the pixel location, \\(k\\) is the disparity, and \\(\\cdot\\) represents the inner product. We found that using the two correlation volumes for the fused features and the NIR features effectively utilize cross-spectral information, resulting in better depth accuracy as shown in Table 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.469, + 0.603 + ], + "angle": 0, + "content": "We estimate a series of disparity maps \\(\\{d_1,\\dots ,d_N\\}\\) using the GRU structure of the RAFT-Stereo network [36]. By feeding \\(F_{\\mathrm{fusion}}^{\\mathrm{left}}\\) as context feature in the RAFT-Stereo, we alternate between the fused and NIR correlation volumes, \\(V_{\\mathrm{fusion}}\\) and \\(V_{\\mathrm{NIR}}\\) as input to the GRU at each iteration, inspired by the cross-spectral time-of-flight imaging method [5]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.469, + 0.679 + ], + "angle": 0, + "content": "Our method leverages spectral information from both RGB and NIR features, placing more emphasis on NIR images that are robust to environmental lighting. For more details on the loss functions and optimization, we refer to the Supplemental Document." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.691, + 0.163, + 0.706 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.469, + 0.762 + ], + "angle": 0, + "content": "We present the evaluation of our proposed pixel-aligned RGB-NIR image fusion and feature fusion methods, along with comprehensive ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.405, + 0.786 + ], + "angle": 0, + "content": "5.1. Pixel-aligned RGB-NIR Image Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.469, + 0.853 + ], + "angle": 0, + "content": "To evaluate our pixel-aligned RGB-NIR image fusion method in Section 4.1, we conducted assessments across three downstream tasks: depth estimation, object detection, and structure from motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Depth Estimation Figure 5 shows that using our fused image as an input to the pre-trained stereo depth esti" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.145 + ], + "angle": 0, + "content": "Table 2. Comparison of image fusion methods for pretrained RAFT-stereo [36] and YOLO [45]. Our RGB-NIR image fusion method outperforms other image-fusion methods for (a) stereo depth estimation and (b) object detection." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.15, + 0.888, + 0.309 + ], + "angle": 0, + "content": "
Methods(a) Depth MAE [m] ↓(b) Detection mAP ↑
RGB5.3110.756
NIR4.2400.703
YCrCb [20]5.2090.571
Bayesian[62]7.4090.745
DarkVision[25]6.1960.762
Adaptive[1]5.5010.773
VGG-NIR[30]5.3840.726
HSV(our baseline) [15]4.1610.744
Ours4.0340.809
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.317, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Table 3. Depth accuracy under diverse environments. Our image and feature fusion methods (last two rows) enable accurate depth estimation, outperforming single-modality RGB or NIR inputs (first two rows) as well as HSV-based RGB-NIR image fusion [15] (third row). RAFT-Stereo [36] is employed across all methods as the depth estimation network." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.404, + 0.89, + 0.48 + ], + "angle": 0, + "content": "
Metric: depth MAE [m]↓Day outdoorNight outdoorCloudy outdoorWell-lit indoorDark indoorComplex indoor
RGB5.7107.4913.1013.2923.2014.698
NIR4.5765.5633.1393.0622.3334.271
HSV fusion [15]4.4685.4173.1213.0912.2934.293
Our image fusion4.3565.1562.8162.9902.4314.181
Our feature fusion3.6514.7712.8382.6952.0593.920
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.604 + ], + "angle": 0, + "content": "mation network [36] yields improved results over single-modality predictions using only RGB or NIR images. Table 2(a) shows quantitative analysis compared to those single-modality predictions as well as using other RGB-NIR image fusion methods. Our proposed RGB-NIR image fusion outperforms all the baselines. Refer to the Supplemental Document for corresponding qualitative results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.698 + ], + "angle": 0, + "content": "Object Detection We use theYOLO object detector [45] to test our RGB-NIR image fusion method. Figure 6 and Table 2(b) show that our image fusion method consistently outperforms single-modality predictions of either RGB or NIR, and other RGB-NIR image fusion methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Structure from Motion Figure 7 shows that our RGB-NIR fused images enables robust reconstruction for a structure-from-motion method [46] on challenging lighting conditions, where RGB and NIR images provide different scene visibility. This suggested the potential utility of our fused RGB-NIR images for downstream applications, including scene reconstruction and view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.838, + 0.848 + ], + "angle": 0, + "content": "5.2. Pixel-aligned RGB-NIR Feature Fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Figure 8 shows that our feature-based RGB-NIR stereo method (Section 4.2) outperforms not only the single-modality predictions using either RGB or NIR channels," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11487" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.228, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.226, + 0.188, + 0.237 + ], + "angle": 0, + "content": "(a) RGB" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.089, + 0.359, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.226, + 0.315, + 0.237 + ], + "angle": 0, + "content": "(b) NIR" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.089, + 0.493, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.226, + 0.49, + 0.238 + ], + "angle": 0, + "content": "(c) Image fusion (ours)" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.089, + 0.582, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.091, + 0.671, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.091, + 0.758, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.091, + 0.89, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.226, + 0.675, + 0.237 + ], + "angle": 0, + "content": "(d) Detailed view" + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.226, + 0.862, + 0.238 + ], + "angle": 0, + "content": "(e) LiDAR (GT)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.895, + 0.293 + ], + "angle": 0, + "content": "Figure 5. Image fusion for pretrained RAFT-stereo [36]. (a)&(b) Using a single modality either RGB or NIR images often results in sub-optimal disparity estimation in challenging lighting conditions. (c) Our RGB-NIR fused image enables robust disparity estimation without finetuning the pretrained model. (d) Close-up views. (e) Ground-truth LiDAR sparse disparity." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.305, + 0.278, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.404, + 0.278, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.503, + 0.278, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.305, + 0.466, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.404, + 0.466, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.503, + 0.466, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.613, + 0.473, + 0.738 + ], + "angle": 0, + "content": "Figure 6. Image fusion for pretrained YOLO [45]. (a) Using the RGB images struggle with object detection in scenes with low light conditions or high saturation due to the under- and overexposed regions by the limited dynamic range of the cameras. (b) NIR images suffer from challenges due to variations in material reflectance and the absence of detailed chromatic information for robust object detection. (c) Our RGB-NIR image fusion integrates the strengths of both RGB and NIR modalities, enabling robust object detection without retraining the model." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.751, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Table 4. Dependency on depth-estimation networks. Our image and feature fusion methods achieve accurate depth estimation across three tested depth networks [7, 36, 57]." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.796, + 0.468, + 0.837 + ], + "angle": 0, + "content": "
Depth MAE [m]↓RGBNIRHSV fusion [15]Our image fusionOur feature fusion
RAFT-Stereo [36]5.2724.3024.2194.0443.619
MoCha-Stereo [7]8.3176.0986.4195.3914.611
IGEV++ [57]6.1855.4025.2565.4872.446
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "but also our RGB-NIR image-fusion method (Section 4.1). While using single-modality RGB/NIR images easily struggle with saturated areas and shadows, our RGB-NIR fused" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.303, + 0.661, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.304, + 0.892, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.409, + 0.819, + 0.421 + ], + "angle": 0, + "content": "(a) COLMAP reconstruction with RGB images" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.419, + 0.657, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.422, + 0.892, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.526, + 0.818, + 0.537 + ], + "angle": 0, + "content": "(b) COLMAP reconstruction with NIR images" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.537, + 0.667, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.538, + 0.891, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.644, + 0.87, + 0.656 + ], + "angle": 0, + "content": "(c) COLMAP reconstruction with our RGB-NIR fusion images" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Figure 7. Image fusion for structure from motion [46]. Compared to using (a) RGB images and (b) NIR images only, our RGB-NIR fused images (c) enable accurate reconstruction of both regions indicated by red and green rectangles. The green region is located closer to the imaging system, where active illumination enhances the visibility of distinct features in the NIR domain. The red region is situated farther from the imaging system, where features are well illuminated in the RGB domain due to ambient light. (c) Our RGB-NIR image fusion integrates these complementary features, enabling robust 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.824, + 0.892, + 0.884 + ], + "angle": 0, + "content": "images provide better depth estimation without finetuning. Our feature-based stereo depth estimation method further improves reconstruction quality by effectively exploiting the strengths of RGB and NIR channels via finetuning." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We evaluate depth estimation accuracy across diverse" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11488" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.089, + 0.895, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.397, + 0.893, + 0.456 + ], + "angle": 0, + "content": "Figure 8. RGB-NIR feature-based depth estimation. (a)&(b) Using single-modality RGB/NIR images for stereo depth estimation [36] often fails to obtain high-quality depth. (c) Our RGB-NIR image-fusion method (Section 4.1) mitigates the issue without finetuning the RAFT stereo, although it does not fully compensate for the strong saturation seen in (a) and (b). (d) Our RGB-NIR feature-based method (Section 4.2) further improves accuracy using an attentional method, at the cost of additional training." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.464, + 0.47, + 0.506 + ], + "angle": 0, + "content": "Table 5. Ablation study on alternating correlation volumes. Using fused and NIR correlation volumes in an alternating manner provides the highest depth-estimation accuracy." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.51, + 0.466, + 0.594 + ], + "angle": 0, + "content": "
Methods on using the fused correlation volumeDepth error [m]↓
Fused correlation volume only7.440
Alternating RGB-NIR correlation volumes8.571
Alternating Fused-RGB-NIR correlation volumes7.426
Alternating Fused-NIR correlation volumes6.747
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.469, + 0.669 + ], + "angle": 0, + "content": "real-world environments using our dataset. Table 3 shows that our image fusion and feature fusion methods outperform single-modality RGB or NIR inputs, as well as HSV-based RGB-NIR image fusion [15]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.469, + 0.759 + ], + "angle": 0, + "content": "We also verify the robustness of our fusion methods with two other depth-estimation networks [7, 57] alongside RAFT-Stereo [36]. Table 4 shows that our proposed fusion methods achieve accurate estimation compared to single-modality inputs and HSV-based RGB-NIR fusion [15] across all three network architectures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Alternating Correlation Volumes In Section 4.2, we investigate the effect of using different types of correlation volumes for stereo depth estimation [36]. Specifically, we experiment with several strategies: using only the fusion correlation volume, alternating between RGB and NIR correlation volumes, alternating among fusion, RGB, and NIR correlation volumes, and alternating between fusion and NIR correlation volumes. Table 5 shows that alternating" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.466, + 0.89, + 0.496 + ], + "angle": 0, + "content": "fusion-NIR correlation volumes yields the highest accuracy compared to the other methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.618, + 0.53 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.694 + ], + "angle": 0, + "content": "In this paper, we have presented a pixel-aligned RGB-NIR stereo imaging system, collect datasets, and develop RGB-NIR image fusion for pretrained models without any fin-tuning and feature fusion methods with additional training for improved performance. We demonstrate the effectiveness of pixel-aligned RGB-NIR images and our RGB-NIR fusion methods for three downstream tasks: depth estimation, object detection, and structure from motion. We hope our work shows the potential of pixel-aligned RGB-NIR imaging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.892, + 0.798 + ], + "angle": 0, + "content": "Limitations and Future Work It will be an interesting future work to bring the success of RGB-image generative models for pixel-aligned RGB-NIR images. Another interesting direction is to improve the accuracy of pixel-aligned RGB-NIR 3D imaging with LiDAR measurements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments Seung-Hwan Baek was supported by Korea NRF (RS-2023-00211658, RS-2024-00438532), Samsung Electronics, and Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT1801-52." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11489" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Mohamed Awad, Ahmed Eliethy, and Hussein A Aly. Adaptive near-infrared and visible fusion for fast image enhancement. IEEE Transactions on Computational Imaging, 6:408-418, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[2] Jean-Yves Bouguet. Camera calibration toolbox for matlab. http://www.vision.caltech.edu/bouguetj/calib_doc/, 2004.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.198, + 0.47, + 0.277 + ], + "angle": 0, + "content": "[3] David P Broadbent, Giorgia D'Innocenzo, Toby J Ellmers, Justin Parsler, Andre J Szameitat, and Daniel T Bishop. Cognitive load, working memory capacity and driving performance: A preliminary fNIRS and eye tracking study. Transportation research part F: traffic psychology and behaviour, 92:121-132, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[4] Matthew Brown and Sabine Susstrunk. Multi-spectral SIFT for scene category recognition. In CVPR 2011, pages 177-184. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.319, + 0.47, + 0.386 + ], + "angle": 0, + "content": "[5] Samuel Brucker, Stefanie Walz, Mario Bijelic, and Felix Heide. Cross-spectral gated-RGB stereo depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21654-21665, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[6] Nived Chebrolu, Philipp Lottes, Alexander Schaefer, Wera Winterhalter, Wolfram Burgard, and Cyril Stachniss. Agricultural robot dataset for plant classification, localization and mapping on sugar beet fields. The International Journal of Robotics Research, 36(10):1045-1052, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.455, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[7] Ziyang Chen, Wei Long, He Yao, Yongjun Zhang, Bingshu Wang, Yongbin Qin, and Jia Wu. Mocha-stereo: Motif channel attention network for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27768-27777, 2024. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.589 + ], + "angle": 0, + "content": "[8] Gyeongmin Choe, Seong-Heum Kim, Sunghoon Im, Joon-Young Lee, Srinivasa G Narasimhan, and In So Kweon. RANUS: RGB and NIR urban scene dataset for deep scene parsing. IEEE Robotics and Automation Letters, 3(3):1808-1815, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[9] Yukyung Choi, Namil Kim, Soonmin Hwang, Kibaek Park, Jae Shin Yoon, Kyounghwan An, and In So Kweon. Kaist multi-spectral day/night data set for autonomous and assisted driving. IEEE Transactions on Intelligent Transportation Systems, 19(3):934-948, Mar. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.738 + ], + "angle": 0, + "content": "[10] David Connah, Mark Samuel Drew, and Graham David Finlayson. Spectral edge image fusion: Theory and applications. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision - ECCV 2014, page 65-80, Cham, 2014. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.74, + 0.469, + 0.794 + ], + "angle": 0, + "content": "[11] Camille Couprie, Clément Farabet, Laurent Najman, and Yann Lecun. Indoor semantic segmentation using depth information. In First International Conference on Learning Representations (ICLR 2013), pages 1-8, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.795, + 0.469, + 0.862 + ], + "angle": 0, + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.863, + 0.469, + 0.888 + ], + "angle": 0, + "content": "[13] Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. In 2021" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "IEEE Winter Conference on Applications of Computer Vision (WACV), page 3559-3568, Waikoloa, HI, USA, Jan. 2021. IEEE. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.134, + 0.892, + 0.201 + ], + "angle": 0, + "content": "[14] Sri Aditya Deevi, Connor Lee, Lu Gan, Sushruth Nagesh, Gaurav Pandey, and Soon-Jo Chung. Rgb-x object detection via scene-specific fusion modules. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7366-7375, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.202, + 0.892, + 0.255 + ], + "angle": 0, + "content": "[15] Clément Fredembach and Sabine Susstrunk. Colouring the near-infrared. In Color and imaging conference, volume 16, pages 176-182. Society of Imaging Science and Technology, 2008. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.256, + 0.892, + 0.31 + ], + "angle": 0, + "content": "[16] Tobias Gruber, Frank Julca-Aguilar, Mario Bijelic, and Felix Heide. Gated2depth: Real-time dense lidar from gated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1506-1516, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.31, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[17] Yubin Guo, Xinlei Qi, Jin Xie, Cheng-Zhong Xu, and Hui Kong. Unsupervised cross-spectrum depth estimation by visible-light and thermal cameras. IEEE Transactions on Intelligent Transportation Systems, 24(10):10937-10947, Oct. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[18] Mark F Hansen, Gary A Atkinson, Lyndon N Smith, and Melvyn L Smith. 3d face reconstructions from photometric stereo using near infrared and visible light. Computer Vision and Image Understanding, 114(8):942-951, 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[19] María Herrera-Arellano, Hayde Peregrina-Barreto, and Iván Terol-Villalobos. Visible-nir image fusion based on top-hat transform. IEEE Transactions on Image Processing, 30:4962-4972, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.486, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[20] María A Herrera-Arellano, Hayde Peregrina-Barreto, and Iván Terol-Villalobos. Color outdoor image enhancement by v-nir fusion and weighted luminance. In 2019 IEEE International Autumn Meeting on Power, Electronics and Computing (ROPEC), pages 1-6. IEEE, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.555, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[21] Yuchen Hong, Youwei Lyu, Si Li, Gang Cao, and Boxin Shi. Reflection removal with nir and rgb image feature fusion. IEEE Transactions on Multimedia, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[22] Xuanlun Huang, Chenyang Wu, Xiaolan Xu, Baishun Wang, Sui Zhang, Chihchiang Shen, Chiennan Yu, Jiaxing Wang, Nan Chi, Shaohua Yu, et al. Polarization structured light 3d depth image sensor for scenes with reflective surfaces. Nature Communications, 14(1):6855, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[23] Soonmin Hwang, Jaesik Park, Namil Kim, Yukyung Choi, and In So Kweon. Multispectral pedestrian detection: Benchmark dataset and baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1037-1045, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[24] Dong-Won Jang and Rae-Hong Park. Colour image dehazing using near-infrared fusion. IET Image Processing, 11(8):587-594, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.771, + 0.892, + 0.838 + ], + "angle": 0, + "content": "[25] Shuangping Jin, Bingbing Yu, Minhao Jing, Yi Zhou, Jiajun Liang, and Renhe Ji. Darkvisionnet: Low-light imaging via rgb-nir fusion with deep inconsistency prior. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1104-1112, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.839, + 0.892, + 0.892 + ], + "angle": 0, + "content": "[26] Cheolkon Jung, Kailong Zhou, and Jiawei Feng. Fusionnet: Multispectral fusion of rgb and nir images using two stage convolutional neural networks. IEEE Access, 8:23912-23919, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11490" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Min H Kim, Todd Alan Harvey, David S Kittle, Holly Rushmeier, Julie Dorsey, Richard O Prum, and David J Brady. 3d imaging spectroscopy for measuring hyperspectral patterns on solid objects. ACM Transactions on Graphics (TOG), 31(4):1-11, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.161, + 0.471, + 0.216 + ], + "angle": 0, + "content": "[28] Namil Kim, Yukyung Choi, Soonmin Hwang, and In So Kweon. Multispectral transfer network: Unsupervised depth estimation for all-day vision. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.216, + 0.471, + 0.27 + ], + "angle": 0, + "content": "[29] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. Vivid++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, July 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.27, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[30] Hui Li, Xiao-Jun Wu, and Josef Kittler. Infrared and visible image fusion using a deep learning framework. In 2018 24th international conference on pattern recognition (ICPR), pages 2705-2710. IEEE, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.471, + 0.364 + ], + "angle": 0, + "content": "[31] Shutao Li, Xudong Kang, and Jianwen Hu. Image fusion with guided filtering. IEEE Transactions on Image Processing, 22(7):2864-2875, July 2013. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.471, + 0.418 + ], + "angle": 0, + "content": "[32] Stan Z Li, RuFeng Chu, ShengCai Liao, and Lun Zhang. Illumination invariant face recognition using near-infrared images. IEEE Transactions on pattern analysis and machine intelligence, 29(4):627-639, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.471, + 0.472 + ], + "angle": 0, + "content": "[33] Xuan Li, Fei Liu, Pingli Han, Shichao Zhang, and Xiaopeng Shao. Near-infrared monocular 3d computational polarization imaging of surfaces exhibiting nonuniform reflectance. Optics Express, 29(10):15616-15630, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.472, + 0.471, + 0.525 + ], + "angle": 0, + "content": "[34] Zhuo Li, Hai-Miao Hu, Wei Zhang, Shiliang Pu, and Bo Li. Spectrum characteristics preserved visible and near-infrared image fusion algorithm. IEEE Transactions on Multimedia, 23:306-319, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[35] Mingyang Liang, Xiaoyang Guo, Hongsheng Li, Xiaogang Wang, and You Song. Unsupervised cross-spectral stereo matching by learning to synthesize. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8706-8713, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[36] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 1, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.471, + 0.702 + ], + "angle": 0, + "content": "[37] Gang Liu and Guohong Huang. Color fusion based on em algorithm for ir and visible image. In 2010 The 2nd International Conference on Computer and Automation Engineering (ICCAE), volume 2, page 253-258, Feb. 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.471, + 0.757 + ], + "angle": 0, + "content": "[38] Jingjing Liu, Shaoting Zhang, Shu Wang, and Dimitris N Metaxas. Multispectral deep neural networks for pedestrian detection. In 27th British Machine Vision Conference, BMVC 2016, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.471, + 0.838 + ], + "angle": 0, + "content": "[39] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.838, + 0.471, + 0.893 + ], + "angle": 0, + "content": "[40] Yusuke Monno, Hayato Teranaka, Kazunori Yoshizaki, Masayuki Tanaka, and Masatoshi Okutomi. Single-sensor rgb-nir imaging: High-quality system design and prototype implementation. IEEE Sensors Journal, 19(2):497-507," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.584, + 0.105 + ], + "angle": 0, + "content": "2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.106, + 0.892, + 0.159 + ], + "angle": 0, + "content": "[41] Peter Mortimer, Raphael Hagmanns, Miguel Granero, Thorsten Luettel, Janko Petereit, and Hans-Joachim Wuensch. The goose dataset for perception in unstructured environments. 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.16, + 0.892, + 0.2 + ], + "angle": 0, + "content": "[42] Chulhee Park and Moon Gi Kang. Color restoration of rgbn multispectral filter array sensor images based on spectral decomposition. Sensors, 16(5):719, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.201, + 0.892, + 0.255 + ], + "angle": 0, + "content": "[43] Matteo Poggi, Pierluigi Zama Ramirez, Fabio Tosi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Cross-spectral neural radiance fields. In 2022 International Conference on 3D Vision (3DV), pages 606-616. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.255, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[44] Long Quan and Zhongdan Lan. Linear n-point camera pose determination. IEEE Transactions on pattern analysis and machine intelligence, 21(8):774-780, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.295, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[45] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[46] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.403, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[47] Takashi Shibata, Masayuki Tanaka, and Masatoshi Okutomi. Versatile visible and near-infrared image fusion based on high visibility area selection. Journal of Electronic Imaging, 25(1):013016-013016, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.457, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[48] Ukcheol Shin, Jinsun Park, and In So Kweon. Deep depth estimation from thermal image. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 1043-1053, Vancouver, BC, Canada, June 2023. IEEE. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[49] Haonan Su, Cheolkon Jung, and Long Yu. Multi-spectral fusion and denoising of color and near-infrared images using multi-scale wavelet analysis. Sensors, 21(11):3610, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.566, + 0.892, + 0.633 + ], + "angle": 0, + "content": "[50] Karasawa Takumi, Kohei Watanabe, Qishen Ha, Antonio Tejero-De-Pablos, Yoshitaka Ushiku, and Tatsuya Harada. Multispectral object detection for autonomous vehicles. In Proceedings of the on Thematic Workshops of ACM Multimedia 2017, pages 35–43, 2017. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[51] Alexander Toet, Maarten A Hogervorst, and Alan R Pinkus. The triclobs dynamic multi-band image data set for the development and evaluation of image fusion methods. PloS one, 11(12):e0165016, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[52] Alexander Toet and Jan Walraven. New false color mapping for image fusion. Optical engineering, 35(3):650-658, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.728, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[53] Abhinav Valada, Gabriel L Oliveira, Thomas Brox, and Wolfram Burgard. Deep multispectral semantic scene understanding of forested environments using multimodal fusion. In 2016 International Symposium on Experimental Robotics, pages 465-477. Springer, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.796, + 0.892, + 0.876 + ], + "angle": 0, + "content": "[54] Stefanie Walz, Mario Bijelic, Andrea Ramazzina, Amanpreet Walia, Fahim Mannan, and Felix Heide. Gated stereo: Joint depth estimation from gated and wide-baseline active stereo cues. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 13252-13262, Vancouver, BC, Canada, June 2023. IEEE. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.876, + 0.892, + 0.891 + ], + "angle": 0, + "content": "[55] Linbo Wang, Tao Wang, Deyun Yang, Xianyong Fang, and" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11491" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Shaohua Wan. Near-infrared fusion for deep lightness enhancement. International Journal of Machine Learning and Cybernetics, 14(5):1621-1633, May 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.201 + ], + "angle": 0, + "content": "[56] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.202, + 0.471, + 0.257 + ], + "angle": 0, + "content": "[57] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: iterative multi-range geometry encoding volumes for stereo matching. arXiv preprint arXiv:2409.00638, 2024. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.257, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[58] Karmesh Yadav, Ram Ramrakhya, Santhosh Kumar Ramakrishnan, Theo Gervet, John Turner, Aaron Gokaslan, Noah Maestre, Angel Xuan Chang, Dhruv Batra, Manolis Savva, et al. Habitat-matterport 3d semantics dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4927-4936, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.338, + 0.471, + 0.405 + ], + "angle": 0, + "content": "[59] Xue Zhang, Xiaohan Zhang, Jiangtao Wang, Jiacheng Ying, Zehua Sheng, Heng Yu, Chunguang Li, and Hui-Liang Shen. Tfdet: Target-aware fusion for rgb-t pedestrian detection. IEEE Transactions on Neural Networks and Learning Systems, page 1-15, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.471, + 0.488 + ], + "angle": 0, + "content": "[60] Yigong Zhang, Yicheng Gao, Shuo Gu, Yubin Guo, Minghao Liu, Zezhou Sun, Zhixing Hou, Hang Yang, Ying Wang, Jian Yang, Jean Ponce, and Hui Kong. Build your own hybrid thermal/eo camera for autonomous vehicle. In 2019 International Conference on Robotics and Automation (ICRA), page 6555-6560, May 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.488, + 0.471, + 0.555 + ], + "angle": 0, + "content": "[61] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.471, + 0.596 + ], + "angle": 0, + "content": "[62] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, and Jiangshe Zhang. Bayesian fusion for infrared and visible images. Signal Processing, 177:107734, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.596, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[63] Tiancheng Zhi, Bernardo R Pires, Martial Hebert, and Srinivasa G Narasimhan. Deep material-aware cross-spectral stereo matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1916-1925, 2018. 1, 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11492" + } + ] +] \ No newline at end of file diff --git a/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/640cd516-44b5-47ce-aa6a-b65d88b8da0a_origin.pdf b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/640cd516-44b5-47ce-aa6a-b65d88b8da0a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0d10de39b214555ab8a20de5970f65b0de1e4238 --- /dev/null +++ b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/640cd516-44b5-47ce-aa6a-b65d88b8da0a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0202221d639faaba080afcdd5d7eae30a276040f2ac452402ed91470c373adaa +size 3765704 diff --git a/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/full.md b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3354681482c211dc870b67f5833b7c8166addf7f --- /dev/null +++ b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/full.md @@ -0,0 +1,366 @@ +# Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision + +Jinnyeong Kim Seung-Hwan Baek POSTECH + +# Abstract + +Integrating RGB and NIR stereo imaging provides complementary spectral information, potentially enhancing robotic 3D vision in challenging lighting conditions. However, existing datasets and imaging systems lack pixel-level alignment between RGB and NIR images, posing challenges for downstream vision tasks. In this paper, we introduce a robotic vision system equipped with pixel-aligned RGB-NIR stereo cameras and a LiDAR sensor mounted on a mobile robot. The system simultaneously captures pixel-aligned pairs of RGB stereo images, NIR stereo images, and temporally synchronized LiDAR points. Utilizing the mobility of the robot, we present a dataset containing continuous video frames under diverse lighting conditions. We then introduce two methods that utilize the pixel-aligned RGB-NIR images: an RGB-NIR image fusion method and a feature fusion method. The first approach enables existing RGB-pretrained vision models to directly utilize RGB-NIR information without fine-tuning. The second approach fine-tunes existing vision models to more effectively utilize RGB-NIR information. Experimental results demonstrate the effectiveness of using pixel-aligned RGB-NIR images across diverse lighting conditions. + +# 1. Introduction + +RGB imaging captures visible light in wavelengths from $400\mathrm{nm}$ to $700\mathrm{nm}$ and serves as the primary data format for visual computing. Near-infrared (NIR) imaging with active illumination captures light in the range of $750\mathrm{nm}$ to $1000\mathrm{nm}$ , offering effective information in conditions where where RGB imaging is inadequate, such as low-light environments during night or poorly lit indoor conditions. NIR light is imperceptible to human vision, allowing for using active NIR illumination without disturbing people, which facilitates robust 3D imaging [5, 32, 38, 54] and eye tracking [3]. Additionally, NIR light offers supplementary scene information due to different spectral reflectance between RGB and NIR wavelengths [18, 27, 33]. + +Combining these benefits, RGB-NIR imaging emerges as a promising modality for challenging environments re + +quiring robust and efficient 2D and 3D vision capabilities. RGB-NIR imaging has been studied for applications including object detection [50, 61] and 3D reconstruction [5, 21, 22, 43]. However, existing RGB-NIR imaging systems and datasets face significant challenges compared to standard RGB vision systems. A major issue is the use of separate RGB and NIR cameras at different poses, leading to pixel misalignment between the NIR and RGB images. Previous methods address such misalignments using image registration, pose estimation, and other processing techniques, which can accumulate errors and limit effectiveness [5, 48, 63]. + +In this work, we develop a robotic vision system capable of capturing pixel-aligned RGB-NIR stereo images and synchronized LiDAR point clouds. Mounted on a mobile platform, the system enables exploration in both indoor and outdoor environments, facilitating data collection under diverse lighting conditions. Using this system, we collect and share a dataset of pixel-aligned RGB-NIR stereo images with synchronized LiDAR point clouds as ground-truth depth for 82 real-world scenes including 90,000 frames. Additionally, we provide a synthetic RGB-NIR dataset with ground-truth dense depth labels for 2,238 synthetic scenes. To demonstrate the effectiveness of pixel-aligned RGB-NIR imaging and datasets, we propose an RGB-NIR image fusion method and an RGB-NIR feature fusion method. + +The first method fuses the pixel-aligned RGB-NIR images into a three-channel image that can be input into pretrained RGB vision models without finetuning. The second method involves developing a stereo depth estimation technique by finetuning pretrained stereo networks [7, 36, 57] to utilize pixel-aligned RGB-NIR features. Our experiments show improvements in various downstream tasks over using only RGB or NIR data, as well as over methods relying on pixel-misaligned RGB-NIR datasets, particularly in challenging scenarios. + +In summary, our contributions are as follows. + +- We develop a robotic vision system with pixel-aligned RGB-NIR stereo cameras and LiDAR, capturing synchronized RGB-NIR stereo images and LiDAR point clouds under diverse lighting conditions. +- We present a large-scale dataset of pixel-aligned RGB- + +NIR stereo images and LiDAR point clouds, collected in various indoor and outdoor environments. + +- We propose an RGB-NIR image fusion method and an RGB-NIR stereo depth estimation network using RGB-NIR feature fusion. +- We evaluate the proposed methods on diverse downstream tasks and illumination conditions, on synthetic and real-world datasets, outperforming baselines. + +# 2. Related Work + +RGB-NIR Imaging Conventional RGB-NIR imaging systems employ separate RGB and NIR cameras positioned at different viewpoints, such as the Kinect sensor, Intel RealSense D415, and various multi-view RGB-NIR systems [8, 9, 11, 12, 29, 48, 56, 58, 63]. However, the captured RGB and NIR images from these systems exhibit pixel misalignment, necessitating depth-dependent registration between the RGB and NIR images. This introduces a significant challenge for effectively utilizing RGB and NIR information, creating a chicken-and-egg problem between depth estimation using spectrally fused images and spectral pose alignment using estimated depth. + +To address the pixel misalignment, sequential capture with interchangeable RGB and NIR filters has been used [4, 25]; however, this approach is limited to capturing static scenes due to the time delay between captures. Designing custom color filter arrays for image sensors to record four channels (RGB-NIR) has also been explored [40]. However, using a single sensor with a fixed global exposure for all four channels makes it unsuitable for challenging environments where the dynamic ranges for RGB and NIR channels differ significantly, such as in dark indoor or outdoor conditions with active NIR illumination. In the domain of RGB-thermal imaging, beam splitters have been used to combine an RGB camera and a thermal camera, enabling pixel-aligned RGB-thermal imaging [23, 28, 60]. + +We capture pixel-aligned RGB-NIR stereo images alongside LiDAR point clouds using two prism-based RGB-NIR dual-sensor cameras. The system is mounted on a mobile robot, allowing scalable data acquisition of pixel-aligned RGB-NIR stereo images in both indoor and outdoor conditions. + +RGB-NIR Fusion To utilize information from both RGB and NIR spectra, fusing RGB-NIR images has been extensively studied with applications in image enhancement for long-distance visibility and fog penetration [1, 10, 20, 62], as well as in information visualization leveraging different RGB-NIR reflectance properties [18, 19, 34, 47]. Conventional methods apply simple arithmetic operations such as addition and subtraction in RGB, HSL, HSV, and YUV color spaces to create images that contain both RGB and + +NIR information [15, 19, 20, 37, 42, 52]. Edge-aware filtering techniques have been employed to refine noisy and low-contrast RGB images using guidance from clean and high-contrast NIR images, showing promising performance in various imaging scenarios [10, 24, 31, 34]. More recently, learning-based approaches have attempted to enhance RGB images and their 3D vision applications using NIR guidance by fusing RGB-NIR features [25, 26, 30, 49, 55] or RGB-thermal features [14, 23, 59]. + +Leveraging our pixel-aligned RGB-NIR images, we develop a RGB-NIR image fusion method that effectively encodes information from both RGB and NIR images into a single three-channel image with learned spatially-varying weights. The resulting fused image can be directly used as input to neural networks pretrained on RGB images. + +RGB-NIR Depth Estimation Cross-spectral correspondence matching between RGB and NIR images using an RGB-NIR color transform network has demonstrated promising depth estimation results when the RGB-NIR reflectances are similar [63]. Liang et al. showed that GAN based architecture can enhance RGB-NIR color transform on cross-spectral stereo matching [35]. Active stereo systems using structured NIR illumination, such as the Kinect and Intel RealSense D415, employ multiple RGB and NIR cameras at different viewpoints. These systems often suffer from the aforementioned chicken-and-egg problem between depth estimation using spectrally fused images and spectral fusion using the estimated depth. Brucker et al. [5] fuse red, blue, and clean channels of stereo RCCB cameras with gating-based time-of-flight cameras positioned differently but still face a similar chicken-and-egg problem. In the domain of RGB-thermal imaging, Guo et al. [17] demonstrate effective depth estimation using pixel-aligned RGB-thermal stereo cameras with beam splitters. + +We develop a depth estimation method leveraging our pixel-aligned RGB-NIR images for robust depth estimation, thereby bypassing the chicken-and-egg problem between depth estimation and spectral fusion. + +# 3. Pixel-aligned RGB-NIR Stereo Imaging + +Robotic Imaging System Figure 1(a) illustrates our robotic imaging system, which integrates a stereo pair of pixel-aligned RGB-NIR cameras, NIR illumination, and a LiDAR mounted on a mobile robot (AgileX Ranger Mini 2.0). Each RGB-NIR camera (JAI FS-1600D-10GE), as shown in Figure 1(c), employs a dichroic prism to separate RGB and NIR light, independently captured by RGB and NIR CMOS sensors. To enhance NIR visibility under varying lighting conditions, we include active NIR illumination (Advanced Illumination AL295-150850IC) without disturbing human vision. The spectral sensitivities of the camera and the illumination are shown in Figure 1(b). We + +![](images/6bdb5b8e2650c642bc857dac65b417eff7ccc9592c68034991f7fa79f9dfc93b.jpg) +(a) Imaging system + +![](images/42756beefa69eac440a5b29aba76cb7cca9c884b61accf35d4b800f38456f62b.jpg) + +![](images/66f05de8d39b4afbd616c20250f24d4293ff51d299ff7ea5bab71ebc5e5efdeb.jpg) +(c) Pixel-aligned RGB-NIR camera +Figure 1. RGB-NIR stereo imaging system. (a) Stereo camera setup integrated with LiDAR and active NIR illumination. (b) Spectral sensitivity profiles of the camera sensors and the irradiance profile of the NIR illumination. (c) Diagram of the pixel-aligned RGB-NIR camera featuring a dichroism prism, where RGB light penetrates the prism and NIR light is reflected, achieving spectral separation. + +installed a LiDAR (OUSTER OS1) to acquire ground-truth sparse depth maps, along with angular velocity and linear acceleration data from its inertial measurement unit. We use RJ-45 interfaces for the data transmission of the cameras and LiDAR. We set up an AC powerbank to supply power to the devices that require an AC adapter. These devices are connected to a laptop, which triggers the capture pipeline and manages real-time data storage. We manage separate threads on the laptop for the left camera, right camera, and LiDAR, acquiring synchronized data. + +Image Formation We model the image formation of the pixel-aligned RGB-NIR camera, which captures RGB and NIR images on separate CMOS sensors. This configuration results in the same exposure time and gain across the RGB channels, with separate exposure time and gain for the NIR channel: $t_R = t_G = t_B \neq t_{\mathrm{NIR}}$ and $g_R = g_G = g_B \neq g_{\mathrm{NIR}}$ . The image captured by camera $c \in \{\text{left}, \text{right}\}$ for channel $i \in \{\text{R}, \text{G}, \text{B}, \text{NIR}\}$ is given by + +$$ +I _ {i} ^ {c} (p ^ {c}) = \eta_ {1} + g _ {i} \left(\eta_ {2} + t _ {i} \left(R _ {i} ^ {c} (p ^ {c}) \left(E _ {i} ^ {c} (p ^ {c}) + L _ {i} ^ {c} (p ^ {c})\right)\right)\right), \tag {1} +$$ + +where $p^c$ is a pixel of camera $c$ , $R_i^c(p^c)$ is the reflectance at pixel $p^c$ for channel $i$ , and $\eta_1, \eta_2$ are Gaussian noise in the post-gamma and pre-gamma stages, respectively. The intensities measured in each channel are influenced by environmental illumination, denoted as $E_i^c(p^c)$ for channel $i$ . The active illumination $L_i^c(p^c)$ is present only in the NIR channel, satisfying $L_i^c(p^c) = 0$ for $i \in \{\mathrm{R}, \mathrm{G}, \mathrm{B}\}$ , since the active illumination is confined to the NIR spectrum (see Figure 1(c)). + +Calibration One advantage of using pixel-aligned cameras is that calibration between RGB and NIR images is unnecessary. Thus, we only calibrate the stereo cameras + +using checkerboard calibration [2]. The cameras contain precision time protocol, which enables maintaining temporal difference between stereo cameras less than 1 microsecond. We also obtain the LiDAR extrinsics by using corresponding point pairs between left-camera images and LiDAR point clouds [44]. + +Real-world Dataset Using the system, we acquire a dataset consisting of 82 videos totaling 90,000 frames under various lighting conditions for training and 7000 frames for testing. For each frame, we provide pixel-aligned RGB-NIR stereo images, a sparse LiDAR point cloud, sensor timestamps, and exposure values for the RGB and NIR CMOS sensors. Each video was recorded at frame rates of $5 - 10\mathrm{Hz}$ , with durations ranging from five to over thirty minutes. Figure 2(a) shows samples from our real dataset and augmented synthetic dataset. We categorize our dataset with lighting conditions: well-lit, dark and complex, based on indoor or outdoor environment. Figure 2(b) shows the statistics of our dataset. We use auto exposure on both RGB and NIR sensors independently, with fixed gain values, and recorded the exposure time for each image. The distribution of exposure times in our dataset varies depending on lighting conditions. For the indoor dataset, many frames have short exposure times for NIR images, as active illumination is sufficient to illuminate nearby indoor scenes. In the daytime dataset, strong sunlight causes the exposure times to shift towards very short durations under direct sunlight and longer durations in shaded areas. For the night dataset, although many frames exhibit long exposure times for both RGB and NIR, the average exposure time for NIR frames is shorter than that for RGB frames. + +Synthetic Dataset We augmented synthetic RGB stereo-image dataset [39] to obtain RGB-NIR stereo dataset with + +![](images/1c75b476738be0abdc435c478e4617fc2723b0bec0962dbce8ce9d32ea430fac.jpg) + +![](images/aad553ce203d37f77d9ed9982f1335477b01f22c386982d95e7f035658ee7d67.jpg) + +![](images/e13cb34444c805ad299dd21657c619920c9f4a68aaaf5150eb2900b78471de57.jpg) + +![](images/40d0db20c310a85a23607e2284c369416d9c36b55fd0a07b53e4ecf7bd9b8145.jpg) + +![](images/d804875de43fcbbb5aa4e4fdd80528794fa90db7b37fc1d673acc3405d42dd6e.jpg) + +![](images/6ecfc06e90d4fcf20464b0391d988c0b3487f2e122e3ab680b8457158a8be9e9.jpg) + +![](images/ea1a5f990256691827bf3d0605b0c601c612ede73b6f180281da74c5fdb72f5e.jpg) + +![](images/ada75e867842a84b31feea414fce65219add8101645002a856dfa61aaddb256d.jpg) + +![](images/c5d6541d77b8d09028d8ef85881c709f21d91b56c1560a4051863fefdf603e17.jpg) +Figure 2. RGB-NIR stereo dataset. (a) Our dataset comprises stereo RGB image pairs, stereo NIR image pairs, and sparse depth point clouds, all captured from continuous video sequences. (b) A quantitative analysis of our dataset is presented, illustrating the distribution of frames under varying lighting conditions, a histogram showing the number of frames per video sequence, and the frame exposure time distributions for RGB and NIR sensors under three distinct lighting scenarios. + +![](images/2467d5d1632b6bad731ac6be6c9441cf88192a0621712efa78bd7d07d1ccdf1b.jpg) + +Table 1. RGB-NIR image dataset comparison. Our dataset is the only dataset with pixel-aligned RGB-NIR stereo image pairs. + +
DatasetsRGB-NIR imagesGT depthEnvironments
Pixel alignedMulti viewRGB stereoNIR stereo
[6]OOXXOOutdoor, day
[50]XOXXXOutdoor, day, night
[53]XOOXOOutdoor, day
[51]OXXXXOutdoor, day, night
[63]XOXXXOutdoor, sunny, dark
[8]XOXXXOutdoor, day, low-light
[48]XOOOOOutdoor, day, night, rain
[41]OOOXOOutdoor, offload, day
Ours (syn.)OOOOOOutdoor, indoor, day, night
Ours (real)OOOOOOutdoor, indoor, day, night
+ +ground-truth dense depth. Figure 2(a) shows a sample of synthetic scenes. We compute the RGB reflectance map $R_{\mathrm{RGB}}$ and the NIR reflectance map and $R_{\mathrm{NIR}}$ using the RGB-to-NIR color synthesis method [16]. We then simulate diverse environmental lighting and obtain the RGB-NIR stereo images using Equation (1). Our augmented dataset comprising 2,238 short sequences (10 frames each) and four long sequences with total 2,200 frames. Our augmentation method is detailed in the supplemental material. + +Dataset Comparison Table 1 compares our dataset with existing RGB-NIR stereo datasets in terms of alignment, environment diversity, and the availability of ground truth + +depth. Unlike datasets requiring color transformation or material segmentation [8, 63], we capture stereo images directly in each spectral band, enhancing quality and reducing computational demands. While others need pose transformations due to differing camera positions [29, 48], our system uses pixel-aligned RGB-NIR camera. By overcoming these challenges, our dataset offers high-quality RGB-NIR stereo images suitable for various applications, including autonomous vehicles and robotics. + +# 4. Pixel-aligned RGB-NIR Fusion + +We introduce two methods fusing pixel-aligned RGB-NIR images to improve performance of vision models. + +# 4.1. Image Fusion with Learned Weights + +First, we fuse pixel-aligned RGB-NIR images into a single three-channel image that can be fed to vision models pretrained for RGB images without finetuning. Figure 3 shows the overview of our image-fusion method. + +Baseline As a baseline, we use the HSV channel blending method [15] to enhance performance on vision application without losing photometric consistency of both RGB and NIR. Given RGB and NIR images for the camera $c$ , the baseline method converts the RGB image into HSV representation: + +$$ +\left[ I _ {H} ^ {c}, I _ {S} ^ {c}, I _ {V} ^ {c} \right] ^ {\intercal} = M \left[ I _ {R} ^ {c}, I _ {G} ^ {c}, I _ {B} ^ {c} \right] ^ {\intercal}, \tag {2} +$$ + +![](images/7ae64f2252e6b081b2ae0728505fec4063e4eb267cc9bc6726e6a18b73229188.jpg) +Figure 3. RGB-NIR image fusion. We fuse the RGB image $I_{\mathrm{RGB}}$ and the image $I_{\mathrm{NIR}}$ as a weighted sum in the brightness domain $V$ after converting RGB image into HSV channel. The spatially-varying weights $\alpha, \beta$ are learned to effectively fuse RGB and NIR images. The fused image can be used as inputs to vision models such as object detection, stereo depth estimation, and structure from motion. + +where $M$ is the RGB-to-HSV conversion matrix. $I_H^c, I_S^c, I_V^c$ are the hue, saturation, and brightness images. We then combine the brightness channel $I_V^c$ with the NIR image $I_{\mathrm{NIR}}^c$ to include details from both while the hue and saturation channels remain unchanged. The fused HSV image is converted back to the RGB domain obtaining the fused image $I_{\mathrm{fusion}}^c$ : + +$$ +I _ {\text {f u s i o n}} ^ {c} = M ^ {- 1} \left[ I _ {H} ^ {c}, I _ {S} ^ {c}, \alpha I _ {V} ^ {c} + \beta I _ {\mathrm {N I R}} ^ {c} \right] ^ {\intercal} \tag {3} +$$ + +where $\alpha$ and $\beta$ are fusion weights. + +The baseline method has a limitation because the weights $\alpha$ and $\beta$ are predefined. This makes the fusion process non-adaptable to various environments and lighting conditions which make image details in NIR and RGB images vary significantly. + +Learned Spatially-varying Weights Instead of using predefined constants $\alpha$ and $\beta$ , we propose to learn spatially-varying $\alpha$ and $\beta$ using an attention-based MLP. We first encode each image into a 256-channel feature map as follows: + +$$ +f _ {\mathrm {e n c}} \left(I _ {\mathrm {R G B}} ^ {c}\right) = F _ {\mathrm {R G B}} ^ {c} \quad \text {a n d} \quad f _ {\mathrm {e n c}} \left(I _ {\mathrm {N I R}} ^ {c}\right) = F _ {\mathrm {N I R}} ^ {c} \tag {4} +$$ + +where $f_{\mathrm{enc}}$ is a ResNet-based feature extractor. For the NIR feature encoding, we copy the NIR channel three times as an input to the pretrained three-channel feature extractor. We then obtain fused feature map $F_{\mathrm{fusion}}^{c}$ by combining $F_{\mathrm{RGB}}^{c}$ and $F_{\mathrm{NIR}}^{c}$ using attentional fusion: + +$$ +F _ {\text {f u s i o n}} ^ {c} = f _ {\text {f u s i o n}} \left(F _ {\mathrm {R G B}} ^ {c}, F _ {\mathrm {N I R}} ^ {c}\right), \tag {5} +$$ + +where the feature fusion module $f_{\mathrm{fusion}}$ computes a sum of RGB and NIR feature maps with feature attention [13]. Finally, spatially varying weights $\alpha$ and $\beta$ are derived by the + +decoder $f_{\mathrm{dec}}$ consisting of several convolution layers: + +$$ +f _ {\mathrm {d e c}} \left(F _ {\text {f u s i o n}} ^ {c}\right) = \{\alpha , \beta \}. \tag {6} +$$ + +Once the spatially-varying scene-dependent values of $\alpha, \beta$ are obtained, we estimate the fused image using Equation (3). Then, we filter $I_{\mathrm{fusion}}^c$ using guided filtering with the reference NIR image $I_{\mathrm{NIR}}^c$ [31]. + +Our RGB-NIR image fusion method runs at 100 fps on a laptop with RTX4060, operating at a resolution of $720 \times 480$ . We evaluate the proposed method on object detection, depth estimation, and structure from motion described in Section 5. We only train the image-fusion models using photometric loss and stereo depth reconstruction loss, without training the downstream vision models. Refer to the Supplemental Document for details. + +# 4.2. Feature Fusion for Stereo Depth Estimation + +While the image fusion method proposed in Section 4.1 allows the use of existing vision models without finetuning, it reduces the image dimensions from four channels (RGB and NIR) to three output channels, which may result in information loss. + +Here, we develop a method that fuses RGB and NIR features and finetunes a downstream model. We use the RAFT-Stereo network [36] as an example. Figure 4 shows a diagram of our proposed stereo depth estimation network. We extract features for the RGB stereo images $I_{\mathrm{RGB}}^c$ and NIR stereo images $I_{\mathrm{NIR}}^c$ , where $c \in \{\text{left}, \text{right}\}$ , using the same ResNet-based feature extractor: $F_{\mathrm{RGB}}^c = f_{\mathrm{enc}}(I_{\mathrm{RGB}}^c)$ , $F_{\mathrm{NIR}}^c = f_{\mathrm{enc}}(I_{\mathrm{NIR}}^c)$ . We then apply the attention-based fusion method from Section 4.1: $F_{\mathrm{fusion}}^c = f_{\mathrm{fusion}}(F_{\mathrm{RGB}}^c, F_{\mathrm{NIR}}^c)$ [13]. + +We compute correlation volumes $V_{s\in \{\mathrm{fusion,NIR}\}}$ for not + +![](images/8ec6670c65c016f4e6653b5ed86bbb06f87b5e9b84ac37f2b144d0d69f3a9f3c.jpg) +Figure 4. RGB-NIR stereo depth estimation model. We modified RAFT-Stereo [36] with attentional feature fusion and alternative correlation search for RGB-NIR depth estimation. We extract features from RGB and NIR images, fuse them, and build cost volumes. We estimate disparity by repeatedly feeding the cost volume of fused features and cost volume of NIR features to the GRU unit whose hidden state is initialized with $F_{\mathrm{fusion}}^{\mathrm{left}}$ . + +only the fused features and but also NIR features: + +$$ +V _ {s} (x, y, k) = F _ {s} ^ {\text {l e f t}} (x, y) \cdot F _ {s} ^ {\text {r i g h t}} (x + k, y), \tag {7} +$$ + +where $(x,y)$ denotes the pixel location, $k$ is the disparity, and $\cdot$ represents the inner product. We found that using the two correlation volumes for the fused features and the NIR features effectively utilize cross-spectral information, resulting in better depth accuracy as shown in Table 5. + +We estimate a series of disparity maps $\{d_1,\dots ,d_N\}$ using the GRU structure of the RAFT-Stereo network [36]. By feeding $F_{\mathrm{fusion}}^{\mathrm{left}}$ as context feature in the RAFT-Stereo, we alternate between the fused and NIR correlation volumes, $V_{\mathrm{fusion}}$ and $V_{\mathrm{NIR}}$ as input to the GRU at each iteration, inspired by the cross-spectral time-of-flight imaging method [5]. + +Our method leverages spectral information from both RGB and NIR features, placing more emphasis on NIR images that are robust to environmental lighting. For more details on the loss functions and optimization, we refer to the Supplemental Document. + +# 5. Results + +We present the evaluation of our proposed pixel-aligned RGB-NIR image fusion and feature fusion methods, along with comprehensive ablation studies. + +# 5.1. Pixel-aligned RGB-NIR Image Fusion + +To evaluate our pixel-aligned RGB-NIR image fusion method in Section 4.1, we conducted assessments across three downstream tasks: depth estimation, object detection, and structure from motion. + +Depth Estimation Figure 5 shows that using our fused image as an input to the pre-trained stereo depth esti + +Table 2. Comparison of image fusion methods for pretrained RAFT-stereo [36] and YOLO [45]. Our RGB-NIR image fusion method outperforms other image-fusion methods for (a) stereo depth estimation and (b) object detection. + +
Methods(a) Depth MAE [m] ↓(b) Detection mAP ↑
RGB5.3110.756
NIR4.2400.703
YCrCb [20]5.2090.571
Bayesian[62]7.4090.745
DarkVision[25]6.1960.762
Adaptive[1]5.5010.773
VGG-NIR[30]5.3840.726
HSV(our baseline) [15]4.1610.744
Ours4.0340.809
+ +Table 3. Depth accuracy under diverse environments. Our image and feature fusion methods (last two rows) enable accurate depth estimation, outperforming single-modality RGB or NIR inputs (first two rows) as well as HSV-based RGB-NIR image fusion [15] (third row). RAFT-Stereo [36] is employed across all methods as the depth estimation network. + +
Metric: depth MAE [m]↓Day outdoorNight outdoorCloudy outdoorWell-lit indoorDark indoorComplex indoor
RGB5.7107.4913.1013.2923.2014.698
NIR4.5765.5633.1393.0622.3334.271
HSV fusion [15]4.4685.4173.1213.0912.2934.293
Our image fusion4.3565.1562.8162.9902.4314.181
Our feature fusion3.6514.7712.8382.6952.0593.920
+ +mation network [36] yields improved results over single-modality predictions using only RGB or NIR images. Table 2(a) shows quantitative analysis compared to those single-modality predictions as well as using other RGB-NIR image fusion methods. Our proposed RGB-NIR image fusion outperforms all the baselines. Refer to the Supplemental Document for corresponding qualitative results. + +Object Detection We use theYOLO object detector [45] to test our RGB-NIR image fusion method. Figure 6 and Table 2(b) show that our image fusion method consistently outperforms single-modality predictions of either RGB or NIR, and other RGB-NIR image fusion methods. + +Structure from Motion Figure 7 shows that our RGB-NIR fused images enables robust reconstruction for a structure-from-motion method [46] on challenging lighting conditions, where RGB and NIR images provide different scene visibility. This suggested the potential utility of our fused RGB-NIR images for downstream applications, including scene reconstruction and view synthesis. + +# 5.2. Pixel-aligned RGB-NIR Feature Fusion + +Figure 8 shows that our feature-based RGB-NIR stereo method (Section 4.2) outperforms not only the single-modality predictions using either RGB or NIR channels, + +![](images/1596889a46cf11b1c298ea5073aab3fc4b1611cc76cca769648d3d476e0180f6.jpg) +(a) RGB + +![](images/207efbf37db92f0dae928758c1b40bb9d7b3a0f1bf3a13db3117f33af22610da.jpg) +(b) NIR + +![](images/5945d8199d38be3a1e03aa7a5cd2a27991ea19fbe3e1033c211f8344daf177b8.jpg) +(c) Image fusion (ours) + +![](images/4746cd236517eb8325cfa5de324ad9425de6313f8868a8ba073de0f359bb4e0e.jpg) +(d) Detailed view + +![](images/a63c0b252faaccb5a6ab3ee5be3a9125f22f26129c2e203a6d51ff8d23104f9b.jpg) + +![](images/335deca18d0f39719bc2901f00ae43e71b21493cbb5d405c0f012c8893b5bf09.jpg) + +![](images/7ab9838840bae7f36b5d0fb02d57c31f0f3318ce37b353a64bcfb0fa5521bdab.jpg) +(e) LiDAR (GT) + +![](images/567798df9bedff439ac758c92708ad680ec81a890f233062f232f14fecd02fa8.jpg) + +![](images/f1b7b686e136cda264100de42d66ece8ec4be6c538c3a3481c8abe309ce2c61a.jpg) + +![](images/94e61948c75e6bf9f733ddc54b30941afcd36e4930c8edbdac57bc870286041f.jpg) +Figure 6. Image fusion for pretrained YOLO [45]. (a) Using the RGB images struggle with object detection in scenes with low light conditions or high saturation due to the under- and overexposed regions by the limited dynamic range of the cameras. (b) NIR images suffer from challenges due to variations in material reflectance and the absence of detailed chromatic information for robust object detection. (c) Our RGB-NIR image fusion integrates the strengths of both RGB and NIR modalities, enabling robust object detection without retraining the model. + +![](images/297bc4a947b14e8479543e9375d69e18145fbc0e1c51673bbd5135facb417d36.jpg) + +![](images/02d1747eda0a77b883de036d375e71c118f307b2ef11a570bb040d39fd012b2f.jpg) + +![](images/b0da687a51e5e07f20400274034179c32346b26eff14744371fcd98ee6efdf1c.jpg) + +Table 4. Dependency on depth-estimation networks. Our image and feature fusion methods achieve accurate depth estimation across three tested depth networks [7, 36, 57]. + +
Depth MAE [m]↓RGBNIRHSV fusion [15]Our image fusionOur feature fusion
RAFT-Stereo [36]5.2724.3024.2194.0443.619
MoCha-Stereo [7]8.3176.0986.4195.3914.611
IGEV++ [57]6.1855.4025.2565.4872.446
+ +but also our RGB-NIR image-fusion method (Section 4.1). While using single-modality RGB/NIR images easily struggle with saturated areas and shadows, our RGB-NIR fused + +![](images/96a4cc3407715ed1820a7c87934e56d825478394a01499bb1ec9cf27260f7b5c.jpg) +Figure 5. Image fusion for pretrained RAFT-stereo [36]. (a)&(b) Using a single modality either RGB or NIR images often results in sub-optimal disparity estimation in challenging lighting conditions. (c) Our RGB-NIR fused image enables robust disparity estimation without finetuning the pretrained model. (d) Close-up views. (e) Ground-truth LiDAR sparse disparity. +(a) COLMAP reconstruction with RGB images + +![](images/3b0e609e973cf8614d1b846f90a09b40a105f06abf0d737e5d9d74a455d67027.jpg) + +![](images/9b67fcf32ce7dee1be94e4fcc2d4df20f782cb21f5c080aa88d5a69c98a26cfc.jpg) + +![](images/fca06e5ea3df449f442045b5f63dfdc846bdfe5f553d53e5b2452534e7abbb60.jpg) + +![](images/9b3d0e975dec39dd254cd9876696d486fe95136ae1168dcf3c34400bfd59098b.jpg) +(b) COLMAP reconstruction with NIR images +(c) COLMAP reconstruction with our RGB-NIR fusion images +Figure 7. Image fusion for structure from motion [46]. Compared to using (a) RGB images and (b) NIR images only, our RGB-NIR fused images (c) enable accurate reconstruction of both regions indicated by red and green rectangles. The green region is located closer to the imaging system, where active illumination enhances the visibility of distinct features in the NIR domain. The red region is situated farther from the imaging system, where features are well illuminated in the RGB domain due to ambient light. (c) Our RGB-NIR image fusion integrates these complementary features, enabling robust 3D reconstruction. + +![](images/d85a5137f1bc4ffc9a5d4ffd27e23836e94a8bf0ddaaf1a6db9586f3a56d521c.jpg) + +images provide better depth estimation without finetuning. Our feature-based stereo depth estimation method further improves reconstruction quality by effectively exploiting the strengths of RGB and NIR channels via finetuning. + +We evaluate depth estimation accuracy across diverse + +![](images/454b089a9cd36a48ac2b739d57ea6df35527891a5e97d6370a47f6c146fe0498.jpg) +Figure 8. RGB-NIR feature-based depth estimation. (a)&(b) Using single-modality RGB/NIR images for stereo depth estimation [36] often fails to obtain high-quality depth. (c) Our RGB-NIR image-fusion method (Section 4.1) mitigates the issue without finetuning the RAFT stereo, although it does not fully compensate for the strong saturation seen in (a) and (b). (d) Our RGB-NIR feature-based method (Section 4.2) further improves accuracy using an attentional method, at the cost of additional training. + +Table 5. Ablation study on alternating correlation volumes. Using fused and NIR correlation volumes in an alternating manner provides the highest depth-estimation accuracy. + +
Methods on using the fused correlation volumeDepth error [m]↓
Fused correlation volume only7.440
Alternating RGB-NIR correlation volumes8.571
Alternating Fused-RGB-NIR correlation volumes7.426
Alternating Fused-NIR correlation volumes6.747
+ +real-world environments using our dataset. Table 3 shows that our image fusion and feature fusion methods outperform single-modality RGB or NIR inputs, as well as HSV-based RGB-NIR image fusion [15]. + +We also verify the robustness of our fusion methods with two other depth-estimation networks [7, 57] alongside RAFT-Stereo [36]. Table 4 shows that our proposed fusion methods achieve accurate estimation compared to single-modality inputs and HSV-based RGB-NIR fusion [15] across all three network architectures. + +Alternating Correlation Volumes In Section 4.2, we investigate the effect of using different types of correlation volumes for stereo depth estimation [36]. Specifically, we experiment with several strategies: using only the fusion correlation volume, alternating between RGB and NIR correlation volumes, alternating among fusion, RGB, and NIR correlation volumes, and alternating between fusion and NIR correlation volumes. Table 5 shows that alternating + +fusion-NIR correlation volumes yields the highest accuracy compared to the other methods. + +# 6. Conclusion + +In this paper, we have presented a pixel-aligned RGB-NIR stereo imaging system, collect datasets, and develop RGB-NIR image fusion for pretrained models without any fin-tuning and feature fusion methods with additional training for improved performance. We demonstrate the effectiveness of pixel-aligned RGB-NIR images and our RGB-NIR fusion methods for three downstream tasks: depth estimation, object detection, and structure from motion. We hope our work shows the potential of pixel-aligned RGB-NIR imaging. + +Limitations and Future Work It will be an interesting future work to bring the success of RGB-image generative models for pixel-aligned RGB-NIR images. Another interesting direction is to improve the accuracy of pixel-aligned RGB-NIR 3D imaging with LiDAR measurements. + +Acknowledgments Seung-Hwan Baek was supported by Korea NRF (RS-2023-00211658, RS-2024-00438532), Samsung Electronics, and Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT1801-52. + +# References + +[1] Mohamed Awad, Ahmed Eliethy, and Hussein A Aly. Adaptive near-infrared and visible fusion for fast image enhancement. IEEE Transactions on Computational Imaging, 6:408-418, 2019. 2, 6 +[2] Jean-Yves Bouguet. Camera calibration toolbox for matlab. http://www.vision.caltech.edu/bouguetj/calib_doc/, 2004.3 +[3] David P Broadbent, Giorgia D'Innocenzo, Toby J Ellmers, Justin Parsler, Andre J Szameitat, and Daniel T Bishop. Cognitive load, working memory capacity and driving performance: A preliminary fNIRS and eye tracking study. Transportation research part F: traffic psychology and behaviour, 92:121-132, 2023. 1 +[4] Matthew Brown and Sabine Susstrunk. Multi-spectral SIFT for scene category recognition. In CVPR 2011, pages 177-184. IEEE, 2011. 2 +[5] Samuel Brucker, Stefanie Walz, Mario Bijelic, and Felix Heide. Cross-spectral gated-RGB stereo depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21654-21665, 2024. 1, 2, 6 +[6] Nived Chebrolu, Philipp Lottes, Alexander Schaefer, Wera Winterhalter, Wolfram Burgard, and Cyril Stachniss. Agricultural robot dataset for plant classification, localization and mapping on sugar beet fields. The International Journal of Robotics Research, 36(10):1045-1052, 2017. 4 +[7] Ziyang Chen, Wei Long, He Yao, Yongjun Zhang, Bingshu Wang, Yongbin Qin, and Jia Wu. Mocha-stereo: Motif channel attention network for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27768-27777, 2024. 1, 7, 8 +[8] Gyeongmin Choe, Seong-Heum Kim, Sunghoon Im, Joon-Young Lee, Srinivasa G Narasimhan, and In So Kweon. RANUS: RGB and NIR urban scene dataset for deep scene parsing. IEEE Robotics and Automation Letters, 3(3):1808-1815, 2018. 2, 4 +[9] Yukyung Choi, Namil Kim, Soonmin Hwang, Kibaek Park, Jae Shin Yoon, Kyounghwan An, and In So Kweon. Kaist multi-spectral day/night data set for autonomous and assisted driving. IEEE Transactions on Intelligent Transportation Systems, 19(3):934-948, Mar. 2018. 2 +[10] David Connah, Mark Samuel Drew, and Graham David Finlayson. Spectral edge image fusion: Theory and applications. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision - ECCV 2014, page 65-80, Cham, 2014. Springer International Publishing. 2 +[11] Camille Couprie, Clément Farabet, Laurent Najman, and Yann Lecun. Indoor semantic segmentation using depth information. In First International Conference on Learning Representations (ICLR 2013), pages 1-8, 2013. 2 +[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2 +[13] Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. In 2021 + +IEEE Winter Conference on Applications of Computer Vision (WACV), page 3559-3568, Waikoloa, HI, USA, Jan. 2021. IEEE. 5 +[14] Sri Aditya Deevi, Connor Lee, Lu Gan, Sushruth Nagesh, Gaurav Pandey, and Soon-Jo Chung. Rgb-x object detection via scene-specific fusion modules. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7366-7375, 2024. 2 +[15] Clément Fredembach and Sabine Susstrunk. Colouring the near-infrared. In Color and imaging conference, volume 16, pages 176-182. Society of Imaging Science and Technology, 2008. 2, 4, 6, 7, 8 +[16] Tobias Gruber, Frank Julca-Aguilar, Mario Bijelic, and Felix Heide. Gated2depth: Real-time dense lidar from gated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1506-1516, 2019. 4 +[17] Yubin Guo, Xinlei Qi, Jin Xie, Cheng-Zhong Xu, and Hui Kong. Unsupervised cross-spectrum depth estimation by visible-light and thermal cameras. IEEE Transactions on Intelligent Transportation Systems, 24(10):10937-10947, Oct. 2023. 2 +[18] Mark F Hansen, Gary A Atkinson, Lyndon N Smith, and Melvyn L Smith. 3d face reconstructions from photometric stereo using near infrared and visible light. Computer Vision and Image Understanding, 114(8):942-951, 2010. 1, 2 +[19] María Herrera-Arellano, Hayde Peregrina-Barreto, and Iván Terol-Villalobos. Visible-nir image fusion based on top-hat transform. IEEE Transactions on Image Processing, 30:4962-4972, 2021. 2 +[20] María A Herrera-Arellano, Hayde Peregrina-Barreto, and Iván Terol-Villalobos. Color outdoor image enhancement by v-nir fusion and weighted luminance. In 2019 IEEE International Autumn Meeting on Power, Electronics and Computing (ROPEC), pages 1-6. IEEE, 2019. 2, 6 +[21] Yuchen Hong, Youwei Lyu, Si Li, Gang Cao, and Boxin Shi. Reflection removal with nir and rgb image feature fusion. IEEE Transactions on Multimedia, 2022. 1 +[22] Xuanlun Huang, Chenyang Wu, Xiaolan Xu, Baishun Wang, Sui Zhang, Chihchiang Shen, Chiennan Yu, Jiaxing Wang, Nan Chi, Shaohua Yu, et al. Polarization structured light 3d depth image sensor for scenes with reflective surfaces. Nature Communications, 14(1):6855, 2023. 1 +[23] Soonmin Hwang, Jaesik Park, Namil Kim, Yukyung Choi, and In So Kweon. Multispectral pedestrian detection: Benchmark dataset and baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1037-1045, 2015. 2 +[24] Dong-Won Jang and Rae-Hong Park. Colour image dehazing using near-infrared fusion. IET Image Processing, 11(8):587-594, 2017. 2 +[25] Shuangping Jin, Bingbing Yu, Minhao Jing, Yi Zhou, Jiajun Liang, and Renhe Ji. Darkvisionnet: Low-light imaging via rgb-nir fusion with deep inconsistency prior. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1104-1112, 2022. 2, 6 +[26] Cheolkon Jung, Kailong Zhou, and Jiawei Feng. Fusionnet: Multispectral fusion of rgb and nir images using two stage convolutional neural networks. IEEE Access, 8:23912-23919, 2020. 2 + +[27] Min H Kim, Todd Alan Harvey, David S Kittle, Holly Rushmeier, Julie Dorsey, Richard O Prum, and David J Brady. 3d imaging spectroscopy for measuring hyperspectral patterns on solid objects. ACM Transactions on Graphics (TOG), 31(4):1-11, 2012. 1 +[28] Namil Kim, Yukyung Choi, Soonmin Hwang, and In So Kweon. Multispectral transfer network: Unsupervised depth estimation for all-day vision. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 2 +[29] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. Vivid++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, July 2022. 2, 4 +[30] Hui Li, Xiao-Jun Wu, and Josef Kittler. Infrared and visible image fusion using a deep learning framework. In 2018 24th international conference on pattern recognition (ICPR), pages 2705-2710. IEEE, 2018. 2, 6 +[31] Shutao Li, Xudong Kang, and Jianwen Hu. Image fusion with guided filtering. IEEE Transactions on Image Processing, 22(7):2864-2875, July 2013. 2, 5 +[32] Stan Z Li, RuFeng Chu, ShengCai Liao, and Lun Zhang. Illumination invariant face recognition using near-infrared images. IEEE Transactions on pattern analysis and machine intelligence, 29(4):627-639, 2007. 1 +[33] Xuan Li, Fei Liu, Pingli Han, Shichao Zhang, and Xiaopeng Shao. Near-infrared monocular 3d computational polarization imaging of surfaces exhibiting nonuniform reflectance. Optics Express, 29(10):15616-15630, 2021. 1 +[34] Zhuo Li, Hai-Miao Hu, Wei Zhang, Shiliang Pu, and Bo Li. Spectrum characteristics preserved visible and near-infrared image fusion algorithm. IEEE Transactions on Multimedia, 23:306-319, 2021. 2 +[35] Mingyang Liang, Xiaoyang Guo, Hongsheng Li, Xiaogang Wang, and You Song. Unsupervised cross-spectral stereo matching by learning to synthesize. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8706-8713, 2019. 2 +[36] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 1, 5, 6, 7, 8 +[37] Gang Liu and Guohong Huang. Color fusion based on em algorithm for ir and visible image. In 2010 The 2nd International Conference on Computer and Automation Engineering (ICCAE), volume 2, page 253-258, Feb. 2010. 2 +[38] Jingjing Liu, Shaoting Zhang, Shu Wang, and Dimitris N Metaxas. Multispectral deep neural networks for pedestrian detection. In 27th British Machine Vision Conference, BMVC 2016, 2016. 1 +[39] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 3 +[40] Yusuke Monno, Hayato Teranaka, Kazunori Yoshizaki, Masayuki Tanaka, and Masatoshi Okutomi. Single-sensor rgb-nir imaging: High-quality system design and prototype implementation. IEEE Sensors Journal, 19(2):497-507, + +2018. 2 +[41] Peter Mortimer, Raphael Hagmanns, Miguel Granero, Thorsten Luettel, Janko Petereit, and Hans-Joachim Wuensch. The goose dataset for perception in unstructured environments. 2024. 4 +[42] Chulhee Park and Moon Gi Kang. Color restoration of rgbn multispectral filter array sensor images based on spectral decomposition. Sensors, 16(5):719, 2016. 2 +[43] Matteo Poggi, Pierluigi Zama Ramirez, Fabio Tosi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Cross-spectral neural radiance fields. In 2022 International Conference on 3D Vision (3DV), pages 606-616. IEEE, 2022. 1 +[44] Long Quan and Zhongdan Lan. Linear n-point camera pose determination. IEEE Transactions on pattern analysis and machine intelligence, 21(8):774-780, 1999. 3 +[45] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 6, 7 +[46] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6, 7 +[47] Takashi Shibata, Masayuki Tanaka, and Masatoshi Okutomi. Versatile visible and near-infrared image fusion based on high visibility area selection. Journal of Electronic Imaging, 25(1):013016-013016, 2016. 2 +[48] Ukcheol Shin, Jinsun Park, and In So Kweon. Deep depth estimation from thermal image. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 1043-1053, Vancouver, BC, Canada, June 2023. IEEE. 1, 2, 4 +[49] Haonan Su, Cheolkon Jung, and Long Yu. Multi-spectral fusion and denoising of color and near-infrared images using multi-scale wavelet analysis. Sensors, 21(11):3610, 2021. 2 +[50] Karasawa Takumi, Kohei Watanabe, Qishen Ha, Antonio Tejero-De-Pablos, Yoshitaka Ushiku, and Tatsuya Harada. Multispectral object detection for autonomous vehicles. In Proceedings of the on Thematic Workshops of ACM Multimedia 2017, pages 35–43, 2017. 1, 4 +[51] Alexander Toet, Maarten A Hogervorst, and Alan R Pinkus. The triclobs dynamic multi-band image data set for the development and evaluation of image fusion methods. PloS one, 11(12):e0165016, 2016. 4 +[52] Alexander Toet and Jan Walraven. New false color mapping for image fusion. Optical engineering, 35(3):650-658, 1996. 2 +[53] Abhinav Valada, Gabriel L Oliveira, Thomas Brox, and Wolfram Burgard. Deep multispectral semantic scene understanding of forested environments using multimodal fusion. In 2016 International Symposium on Experimental Robotics, pages 465-477. Springer, 2017. 4 +[54] Stefanie Walz, Mario Bijelic, Andrea Ramazzina, Amanpreet Walia, Fahim Mannan, and Felix Heide. Gated stereo: Joint depth estimation from gated and wide-baseline active stereo cues. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 13252-13262, Vancouver, BC, Canada, June 2023. IEEE. 1 +[55] Linbo Wang, Tao Wang, Deyun Yang, Xianyong Fang, and + +Shaohua Wan. Near-infrared fusion for deep lightness enhancement. International Journal of Machine Learning and Cybernetics, 14(5):1621-1633, May 2023. 2 +[56] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 2 +[57] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: iterative multi-range geometry encoding volumes for stereo matching. arXiv preprint arXiv:2409.00638, 2024. 1, 7, 8 +[58] Karmesh Yadav, Ram Ramrakhya, Santhosh Kumar Ramakrishnan, Theo Gervet, John Turner, Aaron Gokaslan, Noah Maestre, Angel Xuan Chang, Dhruv Batra, Manolis Savva, et al. Habitat-matterport 3d semantics dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4927-4936, 2023. 2 +[59] Xue Zhang, Xiaohan Zhang, Jiangtao Wang, Jiacheng Ying, Zehua Sheng, Heng Yu, Chunguang Li, and Hui-Liang Shen. Tfdet: Target-aware fusion for rgb-t pedestrian detection. IEEE Transactions on Neural Networks and Learning Systems, page 1-15, 2024. 2 +[60] Yigong Zhang, Yicheng Gao, Shuo Gu, Yubin Guo, Minghao Liu, Zezhou Sun, Zhixing Hou, Hang Yang, Ying Wang, Jian Yang, Jean Ponce, and Hui Kong. Build your own hybrid thermal/eo camera for autonomous vehicle. In 2019 International Conference on Robotics and Automation (ICRA), page 6555-6560, May 2019. 2 +[61] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 1 +[62] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, and Jiangshe Zhang. Bayesian fusion for infrared and visible images. Signal Processing, 177:107734, 2020. 2, 6 +[63] Tiancheng Zhi, Bernardo R Pires, Martial Hebert, and Srinivasa G Narasimhan. Deep material-aware cross-spectral stereo matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1916-1925, 2018. 1, 2, 4 \ No newline at end of file diff --git a/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/images.zip b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f94f59bff2de91ec0440c7b39713b7c2045a59ba --- /dev/null +++ b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fee81f6cd3d8ac2c24a623e38f289bedf2d9abdd3ef4a2987c13e98a0294d1e5 +size 901525 diff --git a/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/layout.json b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..18065a70ef40154beb715e0d8206e72999c3459c --- /dev/null +++ b/2025/Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision/layout.json @@ -0,0 +1,9647 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 102, + 510, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 102, + 510, + 120 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 510, + 120 + ], + "type": "text", + "content": "Pixel-aligned RGB-NIR Stereo Imaging and Dataset for Robot Vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 202, + 143, + 390, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 143, + 390, + 171 + ], + "spans": [ + { + "bbox": [ + 202, + 143, + 390, + 171 + ], + "type": "text", + "content": "Jinnyeong Kim Seung-Hwan Baek POSTECH" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 198, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 198, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 198, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 224, + 290, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 224, + 290, + 488 + ], + "spans": [ + { + "bbox": [ + 45, + 224, + 290, + 488 + ], + "type": "text", + "content": "Integrating RGB and NIR stereo imaging provides complementary spectral information, potentially enhancing robotic 3D vision in challenging lighting conditions. However, existing datasets and imaging systems lack pixel-level alignment between RGB and NIR images, posing challenges for downstream vision tasks. In this paper, we introduce a robotic vision system equipped with pixel-aligned RGB-NIR stereo cameras and a LiDAR sensor mounted on a mobile robot. The system simultaneously captures pixel-aligned pairs of RGB stereo images, NIR stereo images, and temporally synchronized LiDAR points. Utilizing the mobility of the robot, we present a dataset containing continuous video frames under diverse lighting conditions. We then introduce two methods that utilize the pixel-aligned RGB-NIR images: an RGB-NIR image fusion method and a feature fusion method. The first approach enables existing RGB-pretrained vision models to directly utilize RGB-NIR information without fine-tuning. The second approach fine-tunes existing vision models to more effectively utilize RGB-NIR information. Experimental results demonstrate the effectiveness of using pixel-aligned RGB-NIR images across diverse lighting conditions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "text", + "content": "RGB imaging captures visible light in wavelengths from " + }, + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "inline_equation", + "content": "400\\mathrm{nm}" + }, + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "inline_equation", + "content": "700\\mathrm{nm}" + }, + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "text", + "content": " and serves as the primary data format for visual computing. Near-infrared (NIR) imaging with active illumination captures light in the range of " + }, + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "inline_equation", + "content": "750\\mathrm{nm}" + }, + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "inline_equation", + "content": "1000\\mathrm{nm}" + }, + { + "bbox": [ + 46, + 533, + 287, + 688 + ], + "type": "text", + "content": ", offering effective information in conditions where where RGB imaging is inadequate, such as low-light environments during night or poorly lit indoor conditions. NIR light is imperceptible to human vision, allowing for using active NIR illumination without disturbing people, which facilitates robust 3D imaging [5, 32, 38, 54] and eye tracking [3]. Additionally, NIR light offers supplementary scene information due to different spectral reflectance between RGB and NIR wavelengths [18, 27, 33]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Combining these benefits, RGB-NIR imaging emerges as a promising modality for challenging environments re" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 200, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 547, + 344 + ], + "type": "text", + "content": "quiring robust and efficient 2D and 3D vision capabilities. RGB-NIR imaging has been studied for applications including object detection [50, 61] and 3D reconstruction [5, 21, 22, 43]. However, existing RGB-NIR imaging systems and datasets face significant challenges compared to standard RGB vision systems. A major issue is the use of separate RGB and NIR cameras at different poses, leading to pixel misalignment between the NIR and RGB images. Previous methods address such misalignments using image registration, pose estimation, and other processing techniques, which can accumulate errors and limit effectiveness [5, 48, 63]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 346, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 547, + 514 + ], + "type": "text", + "content": "In this work, we develop a robotic vision system capable of capturing pixel-aligned RGB-NIR stereo images and synchronized LiDAR point clouds. Mounted on a mobile platform, the system enables exploration in both indoor and outdoor environments, facilitating data collection under diverse lighting conditions. Using this system, we collect and share a dataset of pixel-aligned RGB-NIR stereo images with synchronized LiDAR point clouds as ground-truth depth for 82 real-world scenes including 90,000 frames. Additionally, we provide a synthetic RGB-NIR dataset with ground-truth dense depth labels for 2,238 synthetic scenes. To demonstrate the effectiveness of pixel-aligned RGB-NIR imaging and datasets, we propose an RGB-NIR image fusion method and an RGB-NIR feature fusion method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 516, + 548, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 548, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 548, + 636 + ], + "type": "text", + "content": "The first method fuses the pixel-aligned RGB-NIR images into a three-channel image that can be input into pretrained RGB vision models without finetuning. The second method involves developing a stereo depth estimation technique by finetuning pretrained stereo networks [7, 36, 57] to utilize pixel-aligned RGB-NIR features. Our experiments show improvements in various downstream tasks over using only RGB or NIR data, as well as over methods relying on pixel-misaligned RGB-NIR datasets, particularly in challenging scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 639, + 502, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 639, + 502, + 651 + ], + "spans": [ + { + "bbox": [ + 317, + 639, + 502, + 651 + ], + "type": "text", + "content": "In summary, our contributions are as follows." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 306, + 653, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 545, + 700 + ], + "type": "text", + "content": "- We develop a robotic vision system with pixel-aligned RGB-NIR stereo cameras and LiDAR, capturing synchronized RGB-NIR stereo images and LiDAR point clouds under diverse lighting conditions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "- We present a large-scale dataset of pixel-aligned RGB-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11482" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 95 + ], + "type": "text", + "content": "NIR stereo images and LiDAR point clouds, collected in various indoor and outdoor environments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 96, + 286, + 168 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 47, + 96, + 286, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 96, + 286, + 130 + ], + "spans": [ + { + "bbox": [ + 47, + 96, + 286, + 130 + ], + "type": "text", + "content": "- We propose an RGB-NIR image fusion method and an RGB-NIR stereo depth estimation network using RGB-NIR feature fusion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 132, + 286, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 132, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 132, + 286, + 168 + ], + "type": "text", + "content": "- We evaluate the proposed methods on diverse downstream tasks and illumination conditions, on synthetic and real-world datasets, outperforming baselines." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 181, + 133, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 181, + 133, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 181, + 133, + 193 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 201, + 286, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 286, + 345 + ], + "type": "text", + "content": "RGB-NIR Imaging Conventional RGB-NIR imaging systems employ separate RGB and NIR cameras positioned at different viewpoints, such as the Kinect sensor, Intel RealSense D415, and various multi-view RGB-NIR systems [8, 9, 11, 12, 29, 48, 56, 58, 63]. However, the captured RGB and NIR images from these systems exhibit pixel misalignment, necessitating depth-dependent registration between the RGB and NIR images. This introduces a significant challenge for effectively utilizing RGB and NIR information, creating a chicken-and-egg problem between depth estimation using spectrally fused images and spectral pose alignment using estimated depth." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 346, + 286, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 286, + 514 + ], + "type": "text", + "content": "To address the pixel misalignment, sequential capture with interchangeable RGB and NIR filters has been used [4, 25]; however, this approach is limited to capturing static scenes due to the time delay between captures. Designing custom color filter arrays for image sensors to record four channels (RGB-NIR) has also been explored [40]. However, using a single sensor with a fixed global exposure for all four channels makes it unsuitable for challenging environments where the dynamic ranges for RGB and NIR channels differ significantly, such as in dark indoor or outdoor conditions with active NIR illumination. In the domain of RGB-thermal imaging, beam splitters have been used to combine an RGB camera and a thermal camera, enabling pixel-aligned RGB-thermal imaging [23, 28, 60]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 514, + 286, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 286, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 286, + 585 + ], + "type": "text", + "content": "We capture pixel-aligned RGB-NIR stereo images alongside LiDAR point clouds using two prism-based RGB-NIR dual-sensor cameras. The system is mounted on a mobile robot, allowing scalable data acquisition of pixel-aligned RGB-NIR stereo images in both indoor and outdoor conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": "RGB-NIR Fusion To utilize information from both RGB and NIR spectra, fusing RGB-NIR images has been extensively studied with applications in image enhancement for long-distance visibility and fog penetration [1, 10, 20, 62], as well as in information visualization leveraging different RGB-NIR reflectance properties [18, 19, 34, 47]. Conventional methods apply simple arithmetic operations such as addition and subtraction in RGB, HSL, HSV, and YUV color spaces to create images that contain both RGB and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "NIR information [15, 19, 20, 37, 42, 52]. Edge-aware filtering techniques have been employed to refine noisy and low-contrast RGB images using guidance from clean and high-contrast NIR images, showing promising performance in various imaging scenarios [10, 24, 31, 34]. More recently, learning-based approaches have attempted to enhance RGB images and their 3D vision applications using NIR guidance by fusing RGB-NIR features [25, 26, 30, 49, 55] or RGB-thermal features [14, 23, 59]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 180, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 251 + ], + "type": "text", + "content": "Leveraging our pixel-aligned RGB-NIR images, we develop a RGB-NIR image fusion method that effectively encodes information from both RGB and NIR images into a single three-channel image with learned spatially-varying weights. The resulting fused image can be directly used as input to neural networks pretrained on RGB images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 265, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 492 + ], + "type": "text", + "content": "RGB-NIR Depth Estimation Cross-spectral correspondence matching between RGB and NIR images using an RGB-NIR color transform network has demonstrated promising depth estimation results when the RGB-NIR reflectances are similar [63]. Liang et al. showed that GAN based architecture can enhance RGB-NIR color transform on cross-spectral stereo matching [35]. Active stereo systems using structured NIR illumination, such as the Kinect and Intel RealSense D415, employ multiple RGB and NIR cameras at different viewpoints. These systems often suffer from the aforementioned chicken-and-egg problem between depth estimation using spectrally fused images and spectral fusion using the estimated depth. Brucker et al. [5] fuse red, blue, and clean channels of stereo RCCB cameras with gating-based time-of-flight cameras positioned differently but still face a similar chicken-and-egg problem. In the domain of RGB-thermal imaging, Guo et al. [17] demonstrate effective depth estimation using pixel-aligned RGB-thermal stereo cameras with beam splitters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 493, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 540 + ], + "type": "text", + "content": "We develop a depth estimation method leveraging our pixel-aligned RGB-NIR images for robust depth estimation, thereby bypassing the chicken-and-egg problem between depth estimation and spectral fusion." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 550, + 523, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 523, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 523, + 563 + ], + "type": "text", + "content": "3. Pixel-aligned RGB-NIR Stereo Imaging" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Robotic Imaging System Figure 1(a) illustrates our robotic imaging system, which integrates a stereo pair of pixel-aligned RGB-NIR cameras, NIR illumination, and a LiDAR mounted on a mobile robot (AgileX Ranger Mini 2.0). Each RGB-NIR camera (JAI FS-1600D-10GE), as shown in Figure 1(c), employs a dichroic prism to separate RGB and NIR light, independently captured by RGB and NIR CMOS sensors. To enhance NIR visibility under varying lighting conditions, we include active NIR illumination (Advanced Illumination AL295-150850IC) without disturbing human vision. The spectral sensitivities of the camera and the illumination are shown in Figure 1(b). We" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11483" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 206, + 232 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 206, + 232 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 206, + 232 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 206, + 232 + ], + "type": "image", + "image_path": "6bdb5b8e2650c642bc857dac65b417eff7ccc9592c68034991f7fa79f9dfc93b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 232, + 241, + 241 + ], + "lines": [ + { + "bbox": [ + 175, + 232, + 241, + 241 + ], + "spans": [ + { + "bbox": [ + 175, + 232, + 241, + 241 + ], + "type": "text", + "content": "(a) Imaging system" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 70, + 363, + 232 + ], + "blocks": [ + { + "bbox": [ + 211, + 70, + 363, + 232 + ], + "lines": [ + { + "bbox": [ + 211, + 70, + 363, + 232 + ], + "spans": [ + { + "bbox": [ + 211, + 70, + 363, + 232 + ], + "type": "image", + "image_path": "42756beefa69eac440a5b29aba76cb7cca9c884b61accf35d4b800f38456f62b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 339, + 70, + 539, + 232 + ], + "blocks": [ + { + "bbox": [ + 339, + 70, + 539, + 232 + ], + "lines": [ + { + "bbox": [ + 339, + 70, + 539, + 232 + ], + "spans": [ + { + "bbox": [ + 339, + 70, + 539, + 232 + ], + "type": "image", + "image_path": "66f05de8d39b4afbd616c20250f24d4293ff51d299ff7ea5bab71ebc5e5efdeb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 232, + 509, + 241 + ], + "lines": [ + { + "bbox": [ + 392, + 232, + 509, + 241 + ], + "spans": [ + { + "bbox": [ + 392, + 232, + 509, + 241 + ], + "type": "text", + "content": "(c) Pixel-aligned RGB-NIR camera" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 243, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 278 + ], + "type": "text", + "content": "Figure 1. RGB-NIR stereo imaging system. (a) Stereo camera setup integrated with LiDAR and active NIR illumination. (b) Spectral sensitivity profiles of the camera sensors and the irradiance profile of the NIR illumination. (c) Diagram of the pixel-aligned RGB-NIR camera featuring a dichroism prism, where RGB light penetrates the prism and NIR light is reflected, achieving spectral separation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 287, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 289, + 407 + ], + "type": "text", + "content": "installed a LiDAR (OUSTER OS1) to acquire ground-truth sparse depth maps, along with angular velocity and linear acceleration data from its inertial measurement unit. We use RJ-45 interfaces for the data transmission of the cameras and LiDAR. We set up an AC powerbank to supply power to the devices that require an AC adapter. These devices are connected to a laptop, which triggers the capture pipeline and manages real-time data storage. We manage separate threads on the laptop for the left camera, right camera, and LiDAR, acquiring synchronized data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 423, + 288, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 288, + 520 + ], + "type": "text", + "content": "Image Formation We model the image formation of the pixel-aligned RGB-NIR camera, which captures RGB and NIR images on separate CMOS sensors. This configuration results in the same exposure time and gain across the RGB channels, with separate exposure time and gain for the NIR channel: " + }, + { + "bbox": [ + 46, + 423, + 288, + 520 + ], + "type": "inline_equation", + "content": "t_R = t_G = t_B \\neq t_{\\mathrm{NIR}}" + }, + { + "bbox": [ + 46, + 423, + 288, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 423, + 288, + 520 + ], + "type": "inline_equation", + "content": "g_R = g_G = g_B \\neq g_{\\mathrm{NIR}}" + }, + { + "bbox": [ + 46, + 423, + 288, + 520 + ], + "type": "text", + "content": ". The image captured by camera " + }, + { + "bbox": [ + 46, + 423, + 288, + 520 + ], + "type": "inline_equation", + "content": "c \\in \\{\\text{left}, \\text{right}\\}" + }, + { + "bbox": [ + 46, + 423, + 288, + 520 + ], + "type": "text", + "content": " for channel " + }, + { + "bbox": [ + 46, + 423, + 288, + 520 + ], + "type": "inline_equation", + "content": "i \\in \\{\\text{R}, \\text{G}, \\text{B}, \\text{NIR}\\}" + }, + { + "bbox": [ + 46, + 423, + 288, + 520 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 529, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 287, + 544 + ], + "type": "interline_equation", + "content": "I _ {i} ^ {c} (p ^ {c}) = \\eta_ {1} + g _ {i} \\left(\\eta_ {2} + t _ {i} \\left(R _ {i} ^ {c} (p ^ {c}) \\left(E _ {i} ^ {c} (p ^ {c}) + L _ {i} ^ {c} (p ^ {c})\\right)\\right)\\right), \\tag {1}", + "image_path": "5a09f7ca9597f847d6eba03f3a3926dfa44191a4d084c58fd9e7dea047642432.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "inline_equation", + "content": "p^c" + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": " is a pixel of camera " + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "inline_equation", + "content": "R_i^c(p^c)" + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": " is the reflectance at pixel " + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "inline_equation", + "content": "p^c" + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": " for channel " + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\eta_1, \\eta_2" + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": " are Gaussian noise in the post-gamma and pre-gamma stages, respectively. The intensities measured in each channel are influenced by environmental illumination, denoted as " + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "inline_equation", + "content": "E_i^c(p^c)" + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": " for channel " + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": ". The active illumination " + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "inline_equation", + "content": "L_i^c(p^c)" + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": " is present only in the NIR channel, satisfying " + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "inline_equation", + "content": "L_i^c(p^c) = 0" + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "inline_equation", + "content": "i \\in \\{\\mathrm{R}, \\mathrm{G}, \\mathrm{B}\\}" + }, + { + "bbox": [ + 46, + 553, + 287, + 661 + ], + "type": "text", + "content": ", since the active illumination is confined to the NIR spectrum (see Figure 1(c))." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Calibration One advantage of using pixel-aligned cameras is that calibration between RGB and NIR images is unnecessary. Thus, we only calibrate the stereo cameras" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 287, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 545, + 359 + ], + "type": "text", + "content": "using checkerboard calibration [2]. The cameras contain precision time protocol, which enables maintaining temporal difference between stereo cameras less than 1 microsecond. We also obtain the LiDAR extrinsics by using corresponding point pairs between left-camera images and LiDAR point clouds [44]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 374, + 546, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 546, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 546, + 673 + ], + "type": "text", + "content": "Real-world Dataset Using the system, we acquire a dataset consisting of 82 videos totaling 90,000 frames under various lighting conditions for training and 7000 frames for testing. For each frame, we provide pixel-aligned RGB-NIR stereo images, a sparse LiDAR point cloud, sensor timestamps, and exposure values for the RGB and NIR CMOS sensors. Each video was recorded at frame rates of " + }, + { + "bbox": [ + 304, + 374, + 546, + 673 + ], + "type": "inline_equation", + "content": "5 - 10\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 374, + 546, + 673 + ], + "type": "text", + "content": ", with durations ranging from five to over thirty minutes. Figure 2(a) shows samples from our real dataset and augmented synthetic dataset. We categorize our dataset with lighting conditions: well-lit, dark and complex, based on indoor or outdoor environment. Figure 2(b) shows the statistics of our dataset. We use auto exposure on both RGB and NIR sensors independently, with fixed gain values, and recorded the exposure time for each image. The distribution of exposure times in our dataset varies depending on lighting conditions. For the indoor dataset, many frames have short exposure times for NIR images, as active illumination is sufficient to illuminate nearby indoor scenes. In the daytime dataset, strong sunlight causes the exposure times to shift towards very short durations under direct sunlight and longer durations in shaded areas. For the night dataset, although many frames exhibit long exposure times for both RGB and NIR, the average exposure time for NIR frames is shorter than that for RGB frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Synthetic Dataset We augmented synthetic RGB stereo-image dataset [39] to obtain RGB-NIR stereo dataset with" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11484" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 138, + 198 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 138, + 198 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 138, + 198 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 138, + 198 + ], + "type": "image", + "image_path": "1c75b476738be0abdc435c478e4617fc2723b0bec0962dbce8ce9d32ea430fac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 72, + 217, + 198 + ], + "blocks": [ + { + "bbox": [ + 139, + 72, + 217, + 198 + ], + "lines": [ + { + "bbox": [ + 139, + 72, + 217, + 198 + ], + "spans": [ + { + "bbox": [ + 139, + 72, + 217, + 198 + ], + "type": "image", + "image_path": "aad553ce203d37f77d9ed9982f1335477b01f22c386982d95e7f035658ee7d67.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 217, + 72, + 295, + 198 + ], + "blocks": [ + { + "bbox": [ + 217, + 72, + 295, + 198 + ], + "lines": [ + { + "bbox": [ + 217, + 72, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 217, + 72, + 295, + 198 + ], + "type": "image", + "image_path": "e13cb34444c805ad299dd21657c619920c9f4a68aaaf5150eb2900b78471de57.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 295, + 72, + 373, + 198 + ], + "blocks": [ + { + "bbox": [ + 295, + 72, + 373, + 198 + ], + "lines": [ + { + "bbox": [ + 295, + 72, + 373, + 198 + ], + "spans": [ + { + "bbox": [ + 295, + 72, + 373, + 198 + ], + "type": "image", + "image_path": "40d0db20c310a85a23607e2284c369416d9c36b55fd0a07b53e4ecf7bd9b8145.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 374, + 72, + 452, + 198 + ], + "blocks": [ + { + "bbox": [ + 374, + 72, + 452, + 198 + ], + "lines": [ + { + "bbox": [ + 374, + 72, + 452, + 198 + ], + "spans": [ + { + "bbox": [ + 374, + 72, + 452, + 198 + ], + "type": "image", + "image_path": "d804875de43fcbbb5aa4e4fdd80528794fa90db7b37fc1d673acc3405d42dd6e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 452, + 72, + 546, + 198 + ], + "blocks": [ + { + "bbox": [ + 452, + 72, + 546, + 198 + ], + "lines": [ + { + "bbox": [ + 452, + 72, + 546, + 198 + ], + "spans": [ + { + "bbox": [ + 452, + 72, + 546, + 198 + ], + "type": "image", + "image_path": "6ecfc06e90d4fcf20464b0391d988c0b3487f2e122e3ab680b8457158a8be9e9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 51, + 208, + 140, + 294 + ], + "blocks": [ + { + "bbox": [ + 51, + 208, + 140, + 294 + ], + "lines": [ + { + "bbox": [ + 51, + 208, + 140, + 294 + ], + "spans": [ + { + "bbox": [ + 51, + 208, + 140, + 294 + ], + "type": "image", + "image_path": "ea1a5f990256691827bf3d0605b0c601c612ede73b6f180281da74c5fdb72f5e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 141, + 209, + 239, + 294 + ], + "blocks": [ + { + "bbox": [ + 141, + 209, + 239, + 294 + ], + "lines": [ + { + "bbox": [ + 141, + 209, + 239, + 294 + ], + "spans": [ + { + "bbox": [ + 141, + 209, + 239, + 294 + ], + "type": "image", + "image_path": "ada75e867842a84b31feea414fce65219add8101645002a856dfa61aaddb256d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 240, + 209, + 436, + 304 + ], + "blocks": [ + { + "bbox": [ + 240, + 209, + 436, + 304 + ], + "lines": [ + { + "bbox": [ + 240, + 209, + 436, + 304 + ], + "spans": [ + { + "bbox": [ + 240, + 209, + 436, + 304 + ], + "type": "image", + "image_path": "c5d6541d77b8d09028d8ef85881c709f21d91b56c1560a4051863fefdf603e17.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 308, + 547, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 547, + 352 + ], + "type": "text", + "content": "Figure 2. RGB-NIR stereo dataset. (a) Our dataset comprises stereo RGB image pairs, stereo NIR image pairs, and sparse depth point clouds, all captured from continuous video sequences. (b) A quantitative analysis of our dataset is presented, illustrating the distribution of frames under varying lighting conditions, a histogram showing the number of frames per video sequence, and the frame exposure time distributions for RGB and NIR sensors under three distinct lighting scenarios." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 442, + 209, + 544, + 293 + ], + "blocks": [ + { + "bbox": [ + 442, + 209, + 544, + 293 + ], + "lines": [ + { + "bbox": [ + 442, + 209, + 544, + 293 + ], + "spans": [ + { + "bbox": [ + 442, + 209, + 544, + 293 + ], + "type": "image", + "image_path": "2467d5d1632b6bad731ac6be6c9441cf88192a0621712efa78bd7d07d1ccdf1b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 48, + 387, + 285, + 536 + ], + "blocks": [ + { + "bbox": [ + 47, + 360, + 287, + 383 + ], + "lines": [ + { + "bbox": [ + 47, + 360, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 287, + 383 + ], + "type": "text", + "content": "Table 1. RGB-NIR image dataset comparison. Our dataset is the only dataset with pixel-aligned RGB-NIR stereo image pairs." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 387, + 285, + 536 + ], + "lines": [ + { + "bbox": [ + 48, + 387, + 285, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 285, + 536 + ], + "type": "table", + "html": "
DatasetsRGB-NIR imagesGT depthEnvironments
Pixel alignedMulti viewRGB stereoNIR stereo
[6]OOXXOOutdoor, day
[50]XOXXXOutdoor, day, night
[53]XOOXOOutdoor, day
[51]OXXXXOutdoor, day, night
[63]XOXXXOutdoor, sunny, dark
[8]XOXXXOutdoor, day, low-light
[48]XOOOOOutdoor, day, night, rain
[41]OOOXOOutdoor, offload, day
Ours (syn.)OOOOOOutdoor, indoor, day, night
Ours (real)OOOOOOutdoor, indoor, day, night
", + "image_path": "3daeb539e1be2c69fbeaafbb0ee3c68d4fe9261d1405af47907076e20baa4cb1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 548, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 657 + ], + "type": "text", + "content": "ground-truth dense depth. Figure 2(a) shows a sample of synthetic scenes. We compute the RGB reflectance map " + }, + { + "bbox": [ + 46, + 548, + 287, + 657 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 46, + 548, + 287, + 657 + ], + "type": "text", + "content": " and the NIR reflectance map and " + }, + { + "bbox": [ + 46, + 548, + 287, + 657 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{NIR}}" + }, + { + "bbox": [ + 46, + 548, + 287, + 657 + ], + "type": "text", + "content": " using the RGB-to-NIR color synthesis method [16]. We then simulate diverse environmental lighting and obtain the RGB-NIR stereo images using Equation (1). Our augmented dataset comprising 2,238 short sequences (10 frames each) and four long sequences with total 2,200 frames. Our augmentation method is detailed in the supplemental material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Dataset Comparison Table 1 compares our dataset with existing RGB-NIR stereo datasets in terms of alignment, environment diversity, and the availability of ground truth" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 361, + 546, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 469 + ], + "type": "text", + "content": "depth. Unlike datasets requiring color transformation or material segmentation [8, 63], we capture stereo images directly in each spectral band, enhancing quality and reducing computational demands. While others need pose transformations due to differing camera positions [29, 48], our system uses pixel-aligned RGB-NIR camera. By overcoming these challenges, our dataset offers high-quality RGB-NIR stereo images suitable for various applications, including autonomous vehicles and robotics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 481, + 481, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 481, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 481, + 495 + ], + "type": "text", + "content": "4. Pixel-aligned RGB-NIR Fusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "text", + "content": "We introduce two methods fusing pixel-aligned RGB-NIR images to improve performance of vision models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 535, + 499, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 535, + 499, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 535, + 499, + 548 + ], + "type": "text", + "content": "4.1. Image Fusion with Learned Weights" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 554, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 554, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 554, + 545, + 601 + ], + "type": "text", + "content": "First, we fuse pixel-aligned RGB-NIR images into a single three-channel image that can be fed to vision models pretrained for RGB images without finetuning. Figure 3 shows the overview of our image-fusion method." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 618, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 689 + ], + "type": "text", + "content": "Baseline As a baseline, we use the HSV channel blending method [15] to enhance performance on vision application without losing photometric consistency of both RGB and NIR. Given RGB and NIR images for the camera " + }, + { + "bbox": [ + 304, + 618, + 545, + 689 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 618, + 545, + 689 + ], + "type": "text", + "content": ", the baseline method converts the RGB image into HSV representation:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 358, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 358, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\left[ I _ {H} ^ {c}, I _ {S} ^ {c}, I _ {V} ^ {c} \\right] ^ {\\intercal} = M \\left[ I _ {R} ^ {c}, I _ {G} ^ {c}, I _ {B} ^ {c} \\right] ^ {\\intercal}, \\tag {2}", + "image_path": "96a4c9c3e77fbb9b7ee40e29b7eb12bf8c6d8d268b33aac119496013e3de4703.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11485" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 545, + 255 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 545, + 255 + ], + "type": "image", + "image_path": "7ae64f2252e6b081b2ae0728505fec4063e4eb267cc9bc6726e6a18b73229188.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 546, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 546, + 297 + ], + "type": "text", + "content": "Figure 3. RGB-NIR image fusion. We fuse the RGB image " + }, + { + "bbox": [ + 46, + 262, + 546, + 297 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 46, + 262, + 546, + 297 + ], + "type": "text", + "content": " and the image " + }, + { + "bbox": [ + 46, + 262, + 546, + 297 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{NIR}}" + }, + { + "bbox": [ + 46, + 262, + 546, + 297 + ], + "type": "text", + "content": " as a weighted sum in the brightness domain " + }, + { + "bbox": [ + 46, + 262, + 546, + 297 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 262, + 546, + 297 + ], + "type": "text", + "content": " after converting RGB image into HSV channel. The spatially-varying weights " + }, + { + "bbox": [ + 46, + 262, + 546, + 297 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 46, + 262, + 546, + 297 + ], + "type": "text", + "content": " are learned to effectively fuse RGB and NIR images. The fused image can be used as inputs to vision models such as object detection, stereo depth estimation, and structure from motion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "text", + "content": " is the RGB-to-HSV conversion matrix. " + }, + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "I_H^c, I_S^c, I_V^c" + }, + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "text", + "content": " are the hue, saturation, and brightness images. We then combine the brightness channel " + }, + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "I_V^c" + }, + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "text", + "content": " with the NIR image " + }, + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{NIR}}^c" + }, + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "text", + "content": " to include details from both while the hue and saturation channels remain unchanged. The fused HSV image is converted back to the RGB domain obtaining the fused image " + }, + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{fusion}}^c" + }, + { + "bbox": [ + 47, + 305, + 287, + 390 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 399, + 287, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 399, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 89, + 399, + 287, + 412 + ], + "type": "interline_equation", + "content": "I _ {\\text {f u s i o n}} ^ {c} = M ^ {- 1} \\left[ I _ {H} ^ {c}, I _ {S} ^ {c}, \\alpha I _ {V} ^ {c} + \\beta I _ {\\mathrm {N I R}} ^ {c} \\right] ^ {\\intercal} \\tag {3}", + "image_path": "80d764e4d057086959f73af45aeca33590d5eb8bb41ba7779c7def28ee6e418f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 422, + 187, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 187, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 187, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 422, + 187, + 433 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 422, + 187, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 422, + 187, + 433 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 422, + 187, + 433 + ], + "type": "text", + "content": " are fusion weights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 434, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 493 + ], + "type": "text", + "content": "The baseline method has a limitation because the weights " + }, + { + "bbox": [ + 46, + 434, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 434, + 287, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 434, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 434, + 287, + 493 + ], + "type": "text", + "content": " are predefined. This makes the fusion process non-adaptable to various environments and lighting conditions which make image details in NIR and RGB images vary significantly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 508, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 287, + 556 + ], + "type": "text", + "content": "Learned Spatially-varying Weights Instead of using predefined constants " + }, + { + "bbox": [ + 47, + 508, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 508, + 287, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 508, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 508, + 287, + 556 + ], + "type": "text", + "content": ", we propose to learn spatially-varying " + }, + { + "bbox": [ + 47, + 508, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 508, + 287, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 508, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 508, + 287, + 556 + ], + "type": "text", + "content": " using an attention-based MLP. We first encode each image into a 256-channel feature map as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 567, + 287, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 567, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 74, + 567, + 287, + 580 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {e n c}} \\left(I _ {\\mathrm {R G B}} ^ {c}\\right) = F _ {\\mathrm {R G B}} ^ {c} \\quad \\text {a n d} \\quad f _ {\\mathrm {e n c}} \\left(I _ {\\mathrm {N I R}} ^ {c}\\right) = F _ {\\mathrm {N I R}} ^ {c} \\tag {4}", + "image_path": "8b89342c3412087af6dcd3c17e27aecc8e40c39ed673800688b000ddf44e11c1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 586, + 286, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 286, + 645 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 586, + 286, + 645 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}" + }, + { + "bbox": [ + 47, + 586, + 286, + 645 + ], + "type": "text", + "content": " is a ResNet-based feature extractor. For the NIR feature encoding, we copy the NIR channel three times as an input to the pretrained three-channel feature extractor. We then obtain fused feature map " + }, + { + "bbox": [ + 47, + 586, + 286, + 645 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fusion}}^{c}" + }, + { + "bbox": [ + 47, + 586, + 286, + 645 + ], + "type": "text", + "content": " by combining " + }, + { + "bbox": [ + 47, + 586, + 286, + 645 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{RGB}}^{c}" + }, + { + "bbox": [ + 47, + 586, + 286, + 645 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 586, + 286, + 645 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{NIR}}^{c}" + }, + { + "bbox": [ + 47, + 586, + 286, + 645 + ], + "type": "text", + "content": " using attentional fusion:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 655, + 287, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 655, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 107, + 655, + 287, + 668 + ], + "type": "interline_equation", + "content": "F _ {\\text {f u s i o n}} ^ {c} = f _ {\\text {f u s i o n}} \\left(F _ {\\mathrm {R G B}} ^ {c}, F _ {\\mathrm {N I R}} ^ {c}\\right), \\tag {5}", + "image_path": "8950c18547ef9fb9d8970c1f4c612d3d4510f576a39a294921e642a277a4be26.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "where the feature fusion module " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{fusion}}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " computes a sum of RGB and NIR feature maps with feature attention [13]. Finally, spatially varying weights " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " are derived by the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 306, + 520, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 306, + 520, + 318 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 520, + 318 + ], + "type": "text", + "content": "decoder " + }, + { + "bbox": [ + 306, + 306, + 520, + 318 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{dec}}" + }, + { + "bbox": [ + 306, + 306, + 520, + 318 + ], + "type": "text", + "content": " consisting of several convolution layers:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 378, + 328, + 545, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 328, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 378, + 328, + 545, + 341 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {d e c}} \\left(F _ {\\text {f u s i o n}} ^ {c}\\right) = \\{\\alpha , \\beta \\}. \\tag {6}", + "image_path": "2d227b2aeed3a319ec8665ebd1cbe80e0bb6596caa367d4547ecc50b67509bb6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 350, + 546, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 350, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 350, + 546, + 399 + ], + "type": "text", + "content": "Once the spatially-varying scene-dependent values of " + }, + { + "bbox": [ + 305, + 350, + 546, + 399 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 305, + 350, + 546, + 399 + ], + "type": "text", + "content": " are obtained, we estimate the fused image using Equation (3). Then, we filter " + }, + { + "bbox": [ + 305, + 350, + 546, + 399 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{fusion}}^c" + }, + { + "bbox": [ + 305, + 350, + 546, + 399 + ], + "type": "text", + "content": " using guided filtering with the reference NIR image " + }, + { + "bbox": [ + 305, + 350, + 546, + 399 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{NIR}}^c" + }, + { + "bbox": [ + 305, + 350, + 546, + 399 + ], + "type": "text", + "content": " [31]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 399, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 494 + ], + "type": "text", + "content": "Our RGB-NIR image fusion method runs at 100 fps on a laptop with RTX4060, operating at a resolution of " + }, + { + "bbox": [ + 305, + 399, + 545, + 494 + ], + "type": "inline_equation", + "content": "720 \\times 480" + }, + { + "bbox": [ + 305, + 399, + 545, + 494 + ], + "type": "text", + "content": ". We evaluate the proposed method on object detection, depth estimation, and structure from motion described in Section 5. We only train the image-fusion models using photometric loss and stereo depth reconstruction loss, without training the downstream vision models. Refer to the Supplemental Document for details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 502, + 533, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 533, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 533, + 515 + ], + "type": "text", + "content": "4.2. Feature Fusion for Stereo Depth Estimation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 521, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 580 + ], + "type": "text", + "content": "While the image fusion method proposed in Section 4.1 allows the use of existing vision models without finetuning, it reduces the image dimensions from four channels (RGB and NIR) to three output channels, which may result in information loss." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "text", + "content": "Here, we develop a method that fuses RGB and NIR features and finetunes a downstream model. We use the RAFT-Stereo network [36] as an example. Figure 4 shows a diagram of our proposed stereo depth estimation network. We extract features for the RGB stereo images " + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{RGB}}^c" + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "text", + "content": " and NIR stereo images " + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{NIR}}^c" + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "inline_equation", + "content": "c \\in \\{\\text{left}, \\text{right}\\}" + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "text", + "content": ", using the same ResNet-based feature extractor: " + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{RGB}}^c = f_{\\mathrm{enc}}(I_{\\mathrm{RGB}}^c)" + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{NIR}}^c = f_{\\mathrm{enc}}(I_{\\mathrm{NIR}}^c)" + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "text", + "content": ". We then apply the attention-based fusion method from Section 4.1: " + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fusion}}^c = f_{\\mathrm{fusion}}(F_{\\mathrm{RGB}}^c, F_{\\mathrm{NIR}}^c)" + }, + { + "bbox": [ + 305, + 582, + 545, + 701 + ], + "type": "text", + "content": " [13]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 702, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 545, + 715 + ], + "type": "text", + "content": "We compute correlation volumes " + }, + { + "bbox": [ + 317, + 702, + 545, + 715 + ], + "type": "inline_equation", + "content": "V_{s\\in \\{\\mathrm{fusion,NIR}\\}}" + }, + { + "bbox": [ + 317, + 702, + 545, + 715 + ], + "type": "text", + "content": " for not" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11486" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 286, + 204 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 204 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 204 + ], + "type": "image", + "image_path": "8ec6670c65c016f4e6653b5ed86bbb06f87b5e9b84ac37f2b144d0d69f3a9f3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 287, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 289 + ], + "type": "text", + "content": "Figure 4. RGB-NIR stereo depth estimation model. We modified RAFT-Stereo [36] with attentional feature fusion and alternative correlation search for RGB-NIR depth estimation. We extract features from RGB and NIR images, fuse them, and build cost volumes. We estimate disparity by repeatedly feeding the cost volume of fused features and cost volume of NIR features to the GRU unit whose hidden state is initialized with " + }, + { + "bbox": [ + 46, + 213, + 287, + 289 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fusion}}^{\\mathrm{left}}" + }, + { + "bbox": [ + 46, + 213, + 287, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 297, + 247, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 247, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 247, + 308 + ], + "type": "text", + "content": "only the fused features and but also NIR features:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 314, + 287, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 314, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 80, + 314, + 287, + 328 + ], + "type": "interline_equation", + "content": "V _ {s} (x, y, k) = F _ {s} ^ {\\text {l e f t}} (x, y) \\cdot F _ {s} ^ {\\text {r i g h t}} (x + k, y), \\tag {7}", + "image_path": "3a60ae9bf24896390621b7679ac3a74cf13a59a64c2d93ef18e207414b4abd31.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "type": "text", + "content": " denotes the pixel location, " + }, + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "type": "text", + "content": " is the disparity, and " + }, + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 334, + 287, + 393 + ], + "type": "text", + "content": " represents the inner product. We found that using the two correlation volumes for the fused features and the NIR features effectively utilize cross-spectral information, resulting in better depth accuracy as shown in Table 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "text", + "content": "We estimate a series of disparity maps " + }, + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\{d_1,\\dots ,d_N\\}" + }, + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "text", + "content": " using the GRU structure of the RAFT-Stereo network [36]. By feeding " + }, + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fusion}}^{\\mathrm{left}}" + }, + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "text", + "content": " as context feature in the RAFT-Stereo, we alternate between the fused and NIR correlation volumes, " + }, + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{fusion}}" + }, + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{NIR}}" + }, + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "text", + "content": " as input to the GRU at each iteration, inspired by the cross-spectral time-of-flight imaging method [5]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 478, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 287, + 537 + ], + "type": "text", + "content": "Our method leverages spectral information from both RGB and NIR features, placing more emphasis on NIR images that are robust to environmental lighting. For more details on the loss functions and optimization, we refer to the Supplemental Document." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 547, + 99, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 99, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 99, + 559 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 567, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 603 + ], + "type": "text", + "content": "We present the evaluation of our proposed pixel-aligned RGB-NIR image fusion and feature fusion methods, along with comprehensive ablation studies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 609, + 247, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 247, + 622 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 247, + 622 + ], + "type": "text", + "content": "5.1. Pixel-aligned RGB-NIR Image Fusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 628, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 675 + ], + "type": "text", + "content": "To evaluate our pixel-aligned RGB-NIR image fusion method in Section 4.1, we conducted assessments across three downstream tasks: depth estimation, object detection, and structure from motion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Depth Estimation Figure 5 shows that using our fused image as an input to the pre-trained stereo depth esti" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 118, + 543, + 244 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 114 + ], + "type": "text", + "content": "Table 2. Comparison of image fusion methods for pretrained RAFT-stereo [36] and YOLO [45]. Our RGB-NIR image fusion method outperforms other image-fusion methods for (a) stereo depth estimation and (b) object detection." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 118, + 543, + 244 + ], + "lines": [ + { + "bbox": [ + 307, + 118, + 543, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 543, + 244 + ], + "type": "table", + "html": "
Methods(a) Depth MAE [m] ↓(b) Detection mAP ↑
RGB5.3110.756
NIR4.2400.703
YCrCb [20]5.2090.571
Bayesian[62]7.4090.745
DarkVision[25]6.1960.762
Adaptive[1]5.5010.773
VGG-NIR[30]5.3840.726
HSV(our baseline) [15]4.1610.744
Ours4.0340.809
", + "image_path": "37ee7886c42f4a0059420d5de8f04a63b34060974c15bdb21dd882bac974e748.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 307, + 319, + 544, + 380 + ], + "blocks": [ + { + "bbox": [ + 305, + 251, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 316 + ], + "type": "text", + "content": "Table 3. Depth accuracy under diverse environments. Our image and feature fusion methods (last two rows) enable accurate depth estimation, outperforming single-modality RGB or NIR inputs (first two rows) as well as HSV-based RGB-NIR image fusion [15] (third row). RAFT-Stereo [36] is employed across all methods as the depth estimation network." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 319, + 544, + 380 + ], + "lines": [ + { + "bbox": [ + 307, + 319, + 544, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 544, + 380 + ], + "type": "table", + "html": "
Metric: depth MAE [m]↓Day outdoorNight outdoorCloudy outdoorWell-lit indoorDark indoorComplex indoor
RGB5.7107.4913.1013.2923.2014.698
NIR4.5765.5633.1393.0622.3334.271
HSV fusion [15]4.4685.4173.1213.0912.2934.293
Our image fusion4.3565.1562.8162.9902.4314.181
Our feature fusion3.6514.7712.8382.6952.0593.920
", + "image_path": "a9b4ef6013c07ee7e5eb60e1c1cadc29b172ccdee8ddd27b7109e784afb563f2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 394, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 478 + ], + "type": "text", + "content": "mation network [36] yields improved results over single-modality predictions using only RGB or NIR images. Table 2(a) shows quantitative analysis compared to those single-modality predictions as well as using other RGB-NIR image fusion methods. Our proposed RGB-NIR image fusion outperforms all the baselines. Refer to the Supplemental Document for corresponding qualitative results." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 492, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 552 + ], + "type": "text", + "content": "Object Detection We use theYOLO object detector [45] to test our RGB-NIR image fusion method. Figure 6 and Table 2(b) show that our image fusion method consistently outperforms single-modality predictions of either RGB or NIR, and other RGB-NIR image fusion methods." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 651 + ], + "type": "text", + "content": "Structure from Motion Figure 7 shows that our RGB-NIR fused images enables robust reconstruction for a structure-from-motion method [46] on challenging lighting conditions, where RGB and NIR images provide different scene visibility. This suggested the potential utility of our fused RGB-NIR images for downstream applications, including scene reconstruction and view synthesis." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 658, + 512, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 512, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 512, + 671 + ], + "type": "text", + "content": "5.2. Pixel-aligned RGB-NIR Feature Fusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Figure 8 shows that our feature-based RGB-NIR stereo method (Section 4.2) outperforms not only the single-modality predictions using either RGB or NIR channels," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11487" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 139, + 178 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 139, + 178 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 139, + 178 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 139, + 178 + ], + "type": "image", + "image_path": "1596889a46cf11b1c298ea5073aab3fc4b1611cc76cca769648d3d476e0180f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 178, + 115, + 187 + ], + "lines": [ + { + "bbox": [ + 86, + 178, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 86, + 178, + 115, + 187 + ], + "type": "text", + "content": "(a) RGB" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 70, + 219, + 178 + ], + "blocks": [ + { + "bbox": [ + 139, + 70, + 219, + 178 + ], + "lines": [ + { + "bbox": [ + 139, + 70, + 219, + 178 + ], + "spans": [ + { + "bbox": [ + 139, + 70, + 219, + 178 + ], + "type": "image", + "image_path": "207efbf37db92f0dae928758c1b40bb9d7b3a0f1bf3a13db3117f33af22610da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 178, + 192, + 187 + ], + "lines": [ + { + "bbox": [ + 166, + 178, + 192, + 187 + ], + "spans": [ + { + "bbox": [ + 166, + 178, + 192, + 187 + ], + "type": "text", + "content": "(b) NIR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 70, + 301, + 178 + ], + "blocks": [ + { + "bbox": [ + 220, + 70, + 301, + 178 + ], + "lines": [ + { + "bbox": [ + 220, + 70, + 301, + 178 + ], + "spans": [ + { + "bbox": [ + 220, + 70, + 301, + 178 + ], + "type": "image", + "image_path": "5945d8199d38be3a1e03aa7a5cd2a27991ea19fbe3e1033c211f8344daf177b8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 178, + 299, + 188 + ], + "lines": [ + { + "bbox": [ + 221, + 178, + 299, + 188 + ], + "spans": [ + { + "bbox": [ + 221, + 178, + 299, + 188 + ], + "type": "text", + "content": "(c) Image fusion (ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 303, + 70, + 356, + 178 + ], + "blocks": [ + { + "bbox": [ + 303, + 70, + 356, + 178 + ], + "lines": [ + { + "bbox": [ + 303, + 70, + 356, + 178 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 356, + 178 + ], + "type": "image", + "image_path": "4746cd236517eb8325cfa5de324ad9425de6313f8868a8ba073de0f359bb4e0e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 178, + 413, + 187 + ], + "lines": [ + { + "bbox": [ + 354, + 178, + 413, + 187 + ], + "spans": [ + { + "bbox": [ + 354, + 178, + 413, + 187 + ], + "type": "text", + "content": "(d) Detailed view" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 356, + 72, + 410, + 178 + ], + "blocks": [ + { + "bbox": [ + 356, + 72, + 410, + 178 + ], + "lines": [ + { + "bbox": [ + 356, + 72, + 410, + 178 + ], + "spans": [ + { + "bbox": [ + 356, + 72, + 410, + 178 + ], + "type": "image", + "image_path": "a63c0b252faaccb5a6ab3ee5be3a9125f22f26129c2e203a6d51ff8d23104f9b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 411, + 72, + 463, + 178 + ], + "blocks": [ + { + "bbox": [ + 411, + 72, + 463, + 178 + ], + "lines": [ + { + "bbox": [ + 411, + 72, + 463, + 178 + ], + "spans": [ + { + "bbox": [ + 411, + 72, + 463, + 178 + ], + "type": "image", + "image_path": "335deca18d0f39719bc2901f00ae43e71b21493cbb5d405c0f012c8893b5bf09.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 465, + 72, + 544, + 178 + ], + "blocks": [ + { + "bbox": [ + 465, + 72, + 544, + 178 + ], + "lines": [ + { + "bbox": [ + 465, + 72, + 544, + 178 + ], + "spans": [ + { + "bbox": [ + 465, + 72, + 544, + 178 + ], + "type": "image", + "image_path": "7ab9838840bae7f36b5d0fb02d57c31f0f3318ce37b353a64bcfb0fa5521bdab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 178, + 527, + 188 + ], + "lines": [ + { + "bbox": [ + 474, + 178, + 527, + 188 + ], + "spans": [ + { + "bbox": [ + 474, + 178, + 527, + 188 + ], + "type": "text", + "content": "(e) LiDAR (GT)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 241, + 170, + 319 + ], + "blocks": [ + { + "bbox": [ + 50, + 241, + 170, + 319 + ], + "lines": [ + { + "bbox": [ + 50, + 241, + 170, + 319 + ], + "spans": [ + { + "bbox": [ + 50, + 241, + 170, + 319 + ], + "type": "image", + "image_path": "567798df9bedff439ac758c92708ad680ec81a890f233062f232f14fecd02fa8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 50, + 319, + 170, + 398 + ], + "blocks": [ + { + "bbox": [ + 50, + 319, + 170, + 398 + ], + "lines": [ + { + "bbox": [ + 50, + 319, + 170, + 398 + ], + "spans": [ + { + "bbox": [ + 50, + 319, + 170, + 398 + ], + "type": "image", + "image_path": "f1b7b686e136cda264100de42d66ece8ec4be6c538c3a3481c8abe309ce2c61a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 50, + 398, + 170, + 476 + ], + "blocks": [ + { + "bbox": [ + 50, + 398, + 170, + 476 + ], + "lines": [ + { + "bbox": [ + 50, + 398, + 170, + 476 + ], + "spans": [ + { + "bbox": [ + 50, + 398, + 170, + 476 + ], + "type": "image", + "image_path": "94e61948c75e6bf9f733ddc54b30941afcd36e4930c8edbdac57bc870286041f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 485, + 289, + 584 + ], + "lines": [ + { + "bbox": [ + 46, + 485, + 289, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 289, + 584 + ], + "type": "text", + "content": "Figure 6. Image fusion for pretrained YOLO [45]. (a) Using the RGB images struggle with object detection in scenes with low light conditions or high saturation due to the under- and overexposed regions by the limited dynamic range of the cameras. (b) NIR images suffer from challenges due to variations in material reflectance and the absence of detailed chromatic information for robust object detection. (c) Our RGB-NIR image fusion integrates the strengths of both RGB and NIR modalities, enabling robust object detection without retraining the model." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 173, + 241, + 285, + 319 + ], + "blocks": [ + { + "bbox": [ + 173, + 241, + 285, + 319 + ], + "lines": [ + { + "bbox": [ + 173, + 241, + 285, + 319 + ], + "spans": [ + { + "bbox": [ + 173, + 241, + 285, + 319 + ], + "type": "image", + "image_path": "297bc4a947b14e8479543e9375d69e18145fbc0e1c51673bbd5135facb417d36.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 173, + 319, + 285, + 398 + ], + "blocks": [ + { + "bbox": [ + 173, + 319, + 285, + 398 + ], + "lines": [ + { + "bbox": [ + 173, + 319, + 285, + 398 + ], + "spans": [ + { + "bbox": [ + 173, + 319, + 285, + 398 + ], + "type": "image", + "image_path": "02d1747eda0a77b883de036d375e71c118f307b2ef11a570bb040d39fd012b2f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 173, + 398, + 285, + 476 + ], + "blocks": [ + { + "bbox": [ + 173, + 398, + 285, + 476 + ], + "lines": [ + { + "bbox": [ + 173, + 398, + 285, + 476 + ], + "spans": [ + { + "bbox": [ + 173, + 398, + 285, + 476 + ], + "type": "image", + "image_path": "b0da687a51e5e07f20400274034179c32346b26eff14744371fcd98ee6efdf1c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 48, + 630, + 286, + 662 + ], + "blocks": [ + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 628 + ], + "type": "text", + "content": "Table 4. Dependency on depth-estimation networks. Our image and feature fusion methods achieve accurate depth estimation across three tested depth networks [7, 36, 57]." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 630, + 286, + 662 + ], + "lines": [ + { + "bbox": [ + 48, + 630, + 286, + 662 + ], + "spans": [ + { + "bbox": [ + 48, + 630, + 286, + 662 + ], + "type": "table", + "html": "
Depth MAE [m]↓RGBNIRHSV fusion [15]Our image fusionOur feature fusion
RAFT-Stereo [36]5.2724.3024.2194.0443.619
MoCha-Stereo [7]8.3176.0986.4195.3914.611
IGEV++ [57]6.1855.4025.2565.4872.446
", + "image_path": "4fd577da27649a67d696efab9a6a2627b274eccca6ee52320a25c6a1a226469f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "but also our RGB-NIR image-fusion method (Section 4.1). While using single-modality RGB/NIR images easily struggle with saturated areas and shadows, our RGB-NIR fused" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 305, + 239, + 404, + 323 + ], + "blocks": [ + { + "bbox": [ + 46, + 197, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 547, + 232 + ], + "type": "text", + "content": "Figure 5. Image fusion for pretrained RAFT-stereo [36]. (a)&(b) Using a single modality either RGB or NIR images often results in sub-optimal disparity estimation in challenging lighting conditions. (c) Our RGB-NIR fused image enables robust disparity estimation without finetuning the pretrained model. (d) Close-up views. (e) Ground-truth LiDAR sparse disparity." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 239, + 404, + 323 + ], + "lines": [ + { + "bbox": [ + 305, + 239, + 404, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 404, + 323 + ], + "type": "image", + "image_path": "96a4cc3407715ed1820a7c87934e56d825478394a01499bb1ec9cf27260f7b5c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 323, + 501, + 333 + ], + "lines": [ + { + "bbox": [ + 349, + 323, + 501, + 333 + ], + "spans": [ + { + "bbox": [ + 349, + 323, + 501, + 333 + ], + "type": "text", + "content": "(a) COLMAP reconstruction with RGB images" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 419, + 240, + 545, + 323 + ], + "blocks": [ + { + "bbox": [ + 419, + 240, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 419, + 240, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 419, + 240, + 545, + 323 + ], + "type": "image", + "image_path": "3b0e609e973cf8614d1b846f90a09b40a105f06abf0d737e5d9d74a455d67027.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 306, + 331, + 402, + 414 + ], + "blocks": [ + { + "bbox": [ + 306, + 331, + 402, + 414 + ], + "lines": [ + { + "bbox": [ + 306, + 331, + 402, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 402, + 414 + ], + "type": "image", + "image_path": "9b67fcf32ce7dee1be94e4fcc2d4df20f782cb21f5c080aa88d5a69c98a26cfc.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 419, + 334, + 545, + 412 + ], + "blocks": [ + { + "bbox": [ + 419, + 334, + 545, + 412 + ], + "lines": [ + { + "bbox": [ + 419, + 334, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 419, + 334, + 545, + 412 + ], + "type": "image", + "image_path": "fca06e5ea3df449f442045b5f63dfdc846bdfe5f553d53e5b2452534e7abbb60.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 307, + 425, + 408, + 510 + ], + "blocks": [ + { + "bbox": [ + 353, + 416, + 500, + 425 + ], + "lines": [ + { + "bbox": [ + 353, + 416, + 500, + 425 + ], + "spans": [ + { + "bbox": [ + 353, + 416, + 500, + 425 + ], + "type": "text", + "content": "(b) COLMAP reconstruction with NIR images" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 425, + 408, + 510 + ], + "lines": [ + { + "bbox": [ + 307, + 425, + 408, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 425, + 408, + 510 + ], + "type": "image", + "image_path": "9b3d0e975dec39dd254cd9876696d486fe95136ae1168dcf3c34400bfd59098b.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 510, + 532, + 519 + ], + "lines": [ + { + "bbox": [ + 331, + 510, + 532, + 519 + ], + "spans": [ + { + "bbox": [ + 331, + 510, + 532, + 519 + ], + "type": "text", + "content": "(c) COLMAP reconstruction with our RGB-NIR fusion images" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 529, + 545, + 639 + ], + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 639 + ], + "type": "text", + "content": "Figure 7. Image fusion for structure from motion [46]. Compared to using (a) RGB images and (b) NIR images only, our RGB-NIR fused images (c) enable accurate reconstruction of both regions indicated by red and green rectangles. The green region is located closer to the imaging system, where active illumination enhances the visibility of distinct features in the NIR domain. The red region is situated farther from the imaging system, where features are well illuminated in the RGB domain due to ambient light. (c) Our RGB-NIR image fusion integrates these complementary features, enabling robust 3D reconstruction." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 419, + 426, + 545, + 508 + ], + "blocks": [ + { + "bbox": [ + 419, + 426, + 545, + 508 + ], + "lines": [ + { + "bbox": [ + 419, + 426, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 419, + 426, + 545, + 508 + ], + "type": "image", + "image_path": "d85a5137f1bc4ffc9a5d4ffd27e23836e94a8bf0ddaaf1a6db9586f3a56d521c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 652, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 652, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 652, + 545, + 700 + ], + "type": "text", + "content": "images provide better depth estimation without finetuning. Our feature-based stereo depth estimation method further improves reconstruction quality by effectively exploiting the strengths of RGB and NIR channels via finetuning." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "We evaluate depth estimation accuracy across diverse" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11488" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 305 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 305 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 305 + ], + "type": "image", + "image_path": "454b089a9cd36a48ac2b739d57ea6df35527891a5e97d6370a47f6c146fe0498.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 314, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 314, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 546, + 361 + ], + "type": "text", + "content": "Figure 8. RGB-NIR feature-based depth estimation. (a)&(b) Using single-modality RGB/NIR images for stereo depth estimation [36] often fails to obtain high-quality depth. (c) Our RGB-NIR image-fusion method (Section 4.1) mitigates the issue without finetuning the RAFT stereo, although it does not fully compensate for the strong saturation seen in (a) and (b). (d) Our RGB-NIR feature-based method (Section 4.2) further improves accuracy using an attentional method, at the cost of additional training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 403, + 285, + 470 + ], + "blocks": [ + { + "bbox": [ + 47, + 367, + 287, + 400 + ], + "lines": [ + { + "bbox": [ + 47, + 367, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 287, + 400 + ], + "type": "text", + "content": "Table 5. Ablation study on alternating correlation volumes. Using fused and NIR correlation volumes in an alternating manner provides the highest depth-estimation accuracy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 403, + 285, + 470 + ], + "lines": [ + { + "bbox": [ + 48, + 403, + 285, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 285, + 470 + ], + "type": "table", + "html": "
Methods on using the fused correlation volumeDepth error [m]↓
Fused correlation volume only7.440
Alternating RGB-NIR correlation volumes8.571
Alternating Fused-RGB-NIR correlation volumes7.426
Alternating Fused-NIR correlation volumes6.747
", + "image_path": "04bd80954aaaa5932596e7544534476e2c61c6c60409235d7db07d0015820b8d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 529 + ], + "type": "text", + "content": "real-world environments using our dataset. Table 3 shows that our image fusion and feature fusion methods outperform single-modality RGB or NIR inputs, as well as HSV-based RGB-NIR image fusion [15]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 601 + ], + "type": "text", + "content": "We also verify the robustness of our fusion methods with two other depth-estimation networks [7, 57] alongside RAFT-Stereo [36]. Table 4 shows that our proposed fusion methods achieve accurate estimation compared to single-modality inputs and HSV-based RGB-NIR fusion [15] across all three network architectures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Alternating Correlation Volumes In Section 4.2, we investigate the effect of using different types of correlation volumes for stereo depth estimation [36]. Specifically, we experiment with several strategies: using only the fusion correlation volume, alternating between RGB and NIR correlation volumes, alternating among fusion, RGB, and NIR correlation volumes, and alternating between fusion and NIR correlation volumes. Table 5 shows that alternating" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 369, + 544, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 544, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 544, + 392 + ], + "type": "text", + "content": "fusion-NIR correlation volumes yields the highest accuracy compared to the other methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 407, + 378, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 378, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 378, + 419 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 429, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 549 + ], + "type": "text", + "content": "In this paper, we have presented a pixel-aligned RGB-NIR stereo imaging system, collect datasets, and develop RGB-NIR image fusion for pretrained models without any fin-tuning and feature fusion methods with additional training for improved performance. We demonstrate the effectiveness of pixel-aligned RGB-NIR images and our RGB-NIR fusion methods for three downstream tasks: depth estimation, object detection, and structure from motion. We hope our work shows the potential of pixel-aligned RGB-NIR imaging." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 571, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 545, + 632 + ], + "type": "text", + "content": "Limitations and Future Work It will be an interesting future work to bring the success of RGB-image generative models for pixel-aligned RGB-NIR images. Another interesting direction is to improve the accuracy of pixel-aligned RGB-NIR 3D imaging with LiDAR measurements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 712 + ], + "type": "text", + "content": "Acknowledgments Seung-Hwan Baek was supported by Korea NRF (RS-2023-00211658, RS-2024-00438532), Samsung Electronics, and Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT1801-52." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11489" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 703 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Mohamed Awad, Ahmed Eliethy, and Hussein A Aly. Adaptive near-infrared and visible fusion for fast image enhancement. IEEE Transactions on Computational Imaging, 6:408-418, 2019. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 156 + ], + "type": "text", + "content": "[2] Jean-Yves Bouguet. Camera calibration toolbox for matlab. http://www.vision.caltech.edu/bouguetj/calib_doc/, 2004.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 156, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 156, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 156, + 287, + 219 + ], + "type": "text", + "content": "[3] David P Broadbent, Giorgia D'Innocenzo, Toby J Ellmers, Justin Parsler, Andre J Szameitat, and Daniel T Bishop. Cognitive load, working memory capacity and driving performance: A preliminary fNIRS and eye tracking study. Transportation research part F: traffic psychology and behaviour, 92:121-132, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 287, + 251 + ], + "type": "text", + "content": "[4] Matthew Brown and Sabine Susstrunk. Multi-spectral SIFT for scene category recognition. In CVPR 2011, pages 177-184. IEEE, 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 252, + 287, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 252, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 252, + 287, + 305 + ], + "type": "text", + "content": "[5] Samuel Brucker, Stefanie Walz, Mario Bijelic, and Felix Heide. Cross-spectral gated-RGB stereo depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21654-21665, 2024. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 306, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 358 + ], + "type": "text", + "content": "[6] Nived Chebrolu, Philipp Lottes, Alexander Schaefer, Wera Winterhalter, Wolfram Burgard, and Cyril Stachniss. Agricultural robot dataset for plant classification, localization and mapping on sugar beet fields. The International Journal of Robotics Research, 36(10):1045-1052, 2017. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 360, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 288, + 413 + ], + "type": "text", + "content": "[7] Ziyang Chen, Wei Long, He Yao, Yongjun Zhang, Bingshu Wang, Yongbin Qin, and Jia Wu. Mocha-stereo: Motif channel attention network for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27768-27777, 2024. 1, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 466 + ], + "type": "text", + "content": "[8] Gyeongmin Choe, Seong-Heum Kim, Sunghoon Im, Joon-Young Lee, Srinivasa G Narasimhan, and In So Kweon. RANUS: RGB and NIR urban scene dataset for deep scene parsing. IEEE Robotics and Automation Letters, 3(3):1808-1815, 2018. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 468, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 521 + ], + "type": "text", + "content": "[9] Yukyung Choi, Namil Kim, Soonmin Hwang, Kibaek Park, Jae Shin Yoon, Kyounghwan An, and In So Kweon. Kaist multi-spectral day/night data set for autonomous and assisted driving. IEEE Transactions on Intelligent Transportation Systems, 19(3):934-948, Mar. 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 521, + 287, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 584 + ], + "type": "text", + "content": "[10] David Connah, Mark Samuel Drew, and Graham David Finlayson. Spectral edge image fusion: Theory and applications. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision - ECCV 2014, page 65-80, Cham, 2014. Springer International Publishing. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 586, + 287, + 628 + ], + "type": "text", + "content": "[11] Camille Couprie, Clément Farabet, Laurent Najman, and Yann Lecun. Indoor semantic segmentation using depth information. In First International Conference on Learning Representations (ICLR 2013), pages 1-8, 2013. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 629, + 287, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 629, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 48, + 629, + 287, + 682 + ], + "type": "text", + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 683, + 287, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 683, + 287, + 703 + ], + "spans": [ + { + "bbox": [ + 48, + 683, + 287, + 703 + ], + "type": "text", + "content": "[13] Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional feature fusion. In 2021" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 706 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "IEEE Winter Conference on Applications of Computer Vision (WACV), page 3559-3568, Waikoloa, HI, USA, Jan. 2021. IEEE. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 106, + 545, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 159 + ], + "type": "text", + "content": "[14] Sri Aditya Deevi, Connor Lee, Lu Gan, Sushruth Nagesh, Gaurav Pandey, and Soon-Jo Chung. Rgb-x object detection via scene-specific fusion modules. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7366-7375, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "type": "text", + "content": "[15] Clément Fredembach and Sabine Susstrunk. Colouring the near-infrared. In Color and imaging conference, volume 16, pages 176-182. Society of Imaging Science and Technology, 2008. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 202, + 545, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 202, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 308, + 202, + 545, + 245 + ], + "type": "text", + "content": "[16] Tobias Gruber, Frank Julca-Aguilar, Mario Bijelic, and Felix Heide. Gated2depth: Real-time dense lidar from gated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1506-1516, 2019. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 245, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 245, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 545, + 298 + ], + "type": "text", + "content": "[17] Yubin Guo, Xinlei Qi, Jin Xie, Cheng-Zhong Xu, and Hui Kong. Unsupervised cross-spectrum depth estimation by visible-light and thermal cameras. IEEE Transactions on Intelligent Transportation Systems, 24(10):10937-10947, Oct. 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 342 + ], + "type": "text", + "content": "[18] Mark F Hansen, Gary A Atkinson, Lyndon N Smith, and Melvyn L Smith. 3d face reconstructions from photometric stereo using near infrared and visible light. Computer Vision and Image Understanding, 114(8):942-951, 2010. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 342, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 384 + ], + "type": "text", + "content": "[19] María Herrera-Arellano, Hayde Peregrina-Barreto, and Iván Terol-Villalobos. Visible-nir image fusion based on top-hat transform. IEEE Transactions on Image Processing, 30:4962-4972, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 384, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 438 + ], + "type": "text", + "content": "[20] María A Herrera-Arellano, Hayde Peregrina-Barreto, and Iván Terol-Villalobos. Color outdoor image enhancement by v-nir fusion and weighted luminance. In 2019 IEEE International Autumn Meeting on Power, Electronics and Computing (ROPEC), pages 1-6. IEEE, 2019. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 439, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 439, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 439, + 545, + 470 + ], + "type": "text", + "content": "[21] Yuchen Hong, Youwei Lyu, Si Li, Gang Cao, and Boxin Shi. Reflection removal with nir and rgb image feature fusion. IEEE Transactions on Multimedia, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 523 + ], + "type": "text", + "content": "[22] Xuanlun Huang, Chenyang Wu, Xiaolan Xu, Baishun Wang, Sui Zhang, Chihchiang Shen, Chiennan Yu, Jiaxing Wang, Nan Chi, Shaohua Yu, et al. Polarization structured light 3d depth image sensor for scenes with reflective surfaces. Nature Communications, 14(1):6855, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 524, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 545, + 578 + ], + "type": "text", + "content": "[23] Soonmin Hwang, Jaesik Park, Namil Kim, Yukyung Choi, and In So Kweon. Multispectral pedestrian detection: Benchmark dataset and baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1037-1045, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 578, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 609 + ], + "type": "text", + "content": "[24] Dong-Won Jang and Rae-Hong Park. Colour image dehazing using near-infrared fusion. IET Image Processing, 11(8):587-594, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 610, + 545, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 610, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 308, + 610, + 545, + 663 + ], + "type": "text", + "content": "[25] Shuangping Jin, Bingbing Yu, Minhao Jing, Yi Zhou, Jiajun Liang, and Renhe Ji. Darkvisionnet: Low-light imaging via rgb-nir fusion with deep inconsistency prior. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1104-1112, 2022. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 664, + 545, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 664, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 308, + 664, + 545, + 706 + ], + "type": "text", + "content": "[26] Cheolkon Jung, Kailong Zhou, and Jiawei Feng. Fusionnet: Multispectral fusion of rgb and nir images using two stage convolutional neural networks. IEEE Access, 8:23912-23919, 2020. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11490" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 707 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Min H Kim, Todd Alan Harvey, David S Kittle, Holly Rushmeier, Julie Dorsey, Richard O Prum, and David J Brady. 3d imaging spectroscopy for measuring hyperspectral patterns on solid objects. ACM Transactions on Graphics (TOG), 31(4):1-11, 2012. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 127, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 127, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 127, + 288, + 171 + ], + "type": "text", + "content": "[28] Namil Kim, Yukyung Choi, Soonmin Hwang, and In So Kweon. Multispectral transfer network: Unsupervised depth estimation for all-day vision. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 171, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 171, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 171, + 288, + 213 + ], + "type": "text", + "content": "[29] Alex Junho Lee, Younggun Cho, Young-sik Shin, Ayoung Kim, and Hyun Myung. Vivid++: Vision for visibility dataset. IEEE Robotics and Automation Letters, 7(3):6282-6289, July 2022. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 213, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 213, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 288, + 256 + ], + "type": "text", + "content": "[30] Hui Li, Xiao-Jun Wu, and Josef Kittler. Infrared and visible image fusion using a deep learning framework. In 2018 24th international conference on pattern recognition (ICPR), pages 2705-2710. IEEE, 2018. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 288, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 288, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 288, + 288 + ], + "type": "text", + "content": "[31] Shutao Li, Xudong Kang, and Jianwen Hu. Image fusion with guided filtering. IEEE Transactions on Image Processing, 22(7):2864-2875, July 2013. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 288, + 331 + ], + "type": "text", + "content": "[32] Stan Z Li, RuFeng Chu, ShengCai Liao, and Lun Zhang. Illumination invariant face recognition using near-infrared images. IEEE Transactions on pattern analysis and machine intelligence, 29(4):627-639, 2007. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 373 + ], + "type": "text", + "content": "[33] Xuan Li, Fei Liu, Pingli Han, Shichao Zhang, and Xiaopeng Shao. Near-infrared monocular 3d computational polarization imaging of surfaces exhibiting nonuniform reflectance. Optics Express, 29(10):15616-15630, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 373, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 288, + 415 + ], + "type": "text", + "content": "[34] Zhuo Li, Hai-Miao Hu, Wei Zhang, Shiliang Pu, and Bo Li. Spectrum characteristics preserved visible and near-infrared image fusion algorithm. IEEE Transactions on Multimedia, 23:306-319, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 288, + 470 + ], + "type": "text", + "content": "[35] Mingyang Liang, Xiaoyang Guo, Hongsheng Li, Xiaogang Wang, and You Song. Unsupervised cross-spectral stereo matching by learning to synthesize. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8706-8713, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 470, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 288, + 513 + ], + "type": "text", + "content": "[36] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227. IEEE, 2021. 1, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 555 + ], + "type": "text", + "content": "[37] Gang Liu and Guohong Huang. Color fusion based on em algorithm for ir and visible image. In 2010 The 2nd International Conference on Computer and Automation Engineering (ICCAE), volume 2, page 253-258, Feb. 2010. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "type": "text", + "content": "[38] Jingjing Liu, Shaoting Zhang, Shu Wang, and Dimitris N Metaxas. Multispectral deep neural networks for pedestrian detection. In 27th British Machine Vision Conference, BMVC 2016, 2016. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 599, + 288, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 288, + 663 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 288, + 663 + ], + "type": "text", + "content": "[39] Nikolaus Mayer, Eddy Ig, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 663, + 288, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 663, + 288, + 707 + ], + "spans": [ + { + "bbox": [ + 48, + 663, + 288, + 707 + ], + "type": "text", + "content": "[40] Yusuke Monno, Hayato Teranaka, Kazunori Yoshizaki, Masayuki Tanaka, and Masatoshi Okutomi. Single-sensor rgb-nir imaging: High-quality system design and prototype implementation. IEEE Sensors Journal, 19(2):497-507," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 705 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 357, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 357, + 83 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 357, + 83 + ], + "type": "text", + "content": "2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 83, + 545, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 83, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 545, + 125 + ], + "type": "text", + "content": "[41] Peter Mortimer, Raphael Hagmanns, Miguel Granero, Thorsten Luettel, Janko Petereit, and Hans-Joachim Wuensch. The goose dataset for perception in unstructured environments. 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 126, + 545, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 126, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 308, + 126, + 545, + 158 + ], + "type": "text", + "content": "[42] Chulhee Park and Moon Gi Kang. Color restoration of rgbn multispectral filter array sensor images based on spectral decomposition. Sensors, 16(5):719, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "type": "text", + "content": "[43] Matteo Poggi, Pierluigi Zama Ramirez, Fabio Tosi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Cross-spectral neural radiance fields. In 2022 International Conference on 3D Vision (3DV), pages 606-616. IEEE, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 201, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 233 + ], + "type": "text", + "content": "[44] Long Quan and Zhongdan Lan. Linear n-point camera pose determination. IEEE Transactions on pattern analysis and machine intelligence, 21(8):774-780, 1999. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 233, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 233, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 233, + 545, + 276 + ], + "type": "text", + "content": "[45] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, 2016. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[46] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 319, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 361 + ], + "type": "text", + "content": "[47] Takashi Shibata, Masayuki Tanaka, and Masatoshi Okutomi. Versatile visible and near-infrared image fusion based on high visibility area selection. Journal of Electronic Imaging, 25(1):013016-013016, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[48] Ukcheol Shin, Jinsun Park, and In So Kweon. Deep depth estimation from thermal image. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 1043-1053, Vancouver, BC, Canada, June 2023. IEEE. 1, 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 415, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 448 + ], + "type": "text", + "content": "[49] Haonan Su, Cheolkon Jung, and Long Yu. Multi-spectral fusion and denoising of color and near-infrared images using multi-scale wavelet analysis. Sensors, 21(11):3610, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 448, + 545, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 448, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 308, + 448, + 545, + 501 + ], + "type": "text", + "content": "[50] Karasawa Takumi, Kohei Watanabe, Qishen Ha, Antonio Tejero-De-Pablos, Yoshitaka Ushiku, and Tatsuya Harada. Multispectral object detection for autonomous vehicles. In Proceedings of the on Thematic Workshops of ACM Multimedia 2017, pages 35–43, 2017. 1, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[51] Alexander Toet, Maarten A Hogervorst, and Alan R Pinkus. The triclobs dynamic multi-band image data set for the development and evaluation of image fusion methods. PloS one, 11(12):e0165016, 2016. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 544, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 575 + ], + "type": "text", + "content": "[52] Alexander Toet and Jan Walraven. New false color mapping for image fusion. Optical engineering, 35(3):650-658, 1996. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 576, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 576, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 308, + 576, + 545, + 630 + ], + "type": "text", + "content": "[53] Abhinav Valada, Gabriel L Oliveira, Thomas Brox, and Wolfram Burgard. Deep multispectral semantic scene understanding of forested environments using multimodal fusion. In 2016 International Symposium on Experimental Robotics, pages 465-477. Springer, 2017. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 630, + 545, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 630, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 308, + 630, + 545, + 693 + ], + "type": "text", + "content": "[54] Stefanie Walz, Mario Bijelic, Andrea Ramazzina, Amanpreet Walia, Fahim Mannan, and Felix Heide. Gated stereo: Joint depth estimation from gated and wide-baseline active stereo cues. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 13252-13262, Vancouver, BC, Canada, June 2023. IEEE. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 693, + 545, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 693, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 308, + 693, + 545, + 705 + ], + "type": "text", + "content": "[55] Linbo Wang, Tao Wang, Deyun Yang, Xianyong Fang, and" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11491" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 525 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Shaohua Wan. Near-infrared fusion for deep lightness enhancement. International Journal of Machine Learning and Cybernetics, 14(5):1621-1633, May 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "type": "text", + "content": "[56] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9068-9079, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 159, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 159, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 288, + 203 + ], + "type": "text", + "content": "[57] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: iterative multi-range geometry encoding volumes for stereo matching. arXiv preprint arXiv:2409.00638, 2024. 1, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 203, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 203, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 203, + 288, + 267 + ], + "type": "text", + "content": "[58] Karmesh Yadav, Ram Ramrakhya, Santhosh Kumar Ramakrishnan, Theo Gervet, John Turner, Aaron Gokaslan, Noah Maestre, Angel Xuan Chang, Dhruv Batra, Manolis Savva, et al. Habitat-matterport 3d semantics dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4927-4936, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 267, + 288, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 288, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 288, + 320 + ], + "type": "text", + "content": "[59] Xue Zhang, Xiaohan Zhang, Jiangtao Wang, Jiacheng Ying, Zehua Sheng, Heng Yu, Chunguang Li, and Hui-Liang Shen. Tfdet: Target-aware fusion for rgb-t pedestrian detection. IEEE Transactions on Neural Networks and Learning Systems, page 1-15, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 288, + 386 + ], + "type": "text", + "content": "[60] Yigong Zhang, Yicheng Gao, Shuo Gu, Yubin Guo, Minghao Liu, Zezhou Sun, Zhixing Hou, Hang Yang, Ying Wang, Jian Yang, Jean Ponce, and Hui Kong. Build your own hybrid thermal/eo camera for autonomous vehicle. In 2019 International Conference on Robotics and Automation (ICRA), page 6555-6560, May 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 386, + 288, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 288, + 439 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 288, + 439 + ], + "type": "text", + "content": "[61] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 440, + 288, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 288, + 472 + ], + "type": "text", + "content": "[62] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, and Jiangshe Zhang. Bayesian fusion for infrared and visible images. Signal Processing, 177:107734, 2020. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 472, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 288, + 525 + ], + "type": "text", + "content": "[63] Tiancheng Zhi, Bernardo R Pires, Martial Hebert, and Srinivasa G Narasimhan. Deep material-aware cross-spectral stereo matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1916-1925, 2018. 1, 2, 4" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11492" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/6dc2236b-e40d-40ee-98e0-34b3414c88ae_content_list.json b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/6dc2236b-e40d-40ee-98e0-34b3414c88ae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d39ea48ff0de19b31073d58a78628d10c38a9fb5 --- /dev/null +++ b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/6dc2236b-e40d-40ee-98e0-34b3414c88ae_content_list.json @@ -0,0 +1,1412 @@ +[ + { + "type": "text", + "text": "Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach", + "text_level": 1, + "bbox": [ + 200, + 130, + 799, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lingchen Sun $^{1,2}$ , Rongyuan Wu $^{1,2}$ , Zhiyuan Ma $^{1}$ , Shuaizheng Liu $^{1,2}$ , Qiaosi Yi $^{1,2}$ , Lei Zhang $^{1,2,*}$ $^{1}$ The Hong Kong Polytechnic University \n $^{2}$ OPPO Research Institute", + "bbox": [ + 120, + 202, + 880, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ling-chen.sun, rong-yuan.wu, zm2354.ma} @connect.polyu.hk", + "bbox": [ + 305, + 241, + 684, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shuaizhengliu21,qiaosiyijoyies} $@$ gmail.com cslzhang@comp.polyu.edu.hk", + "bbox": [ + 263, + 258, + 733, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 309, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion prior-based methods have shown impressive results in real-world image super-resolution (SR). However, most existing methods entangle pixel-level and semantic-level SR objectives in the training process, struggling to balance pixel-wise fidelity and perceptual quality. Meanwhile, users have varying preferences on SR results, thus it is demanded to develop an adjustable SR model that can be tailored to different fidelity-perception preferences during inference without re-training. We present Pixel-level and Semantic-level Adjustable SR (PiSA-SR), which learns two LoRA modules upon the pre-trained stable-diffusion (SD) model to achieve improved and adjustable SR results. We first formulate the SD-based SR problem as learning the residual between the low-quality input and the high-quality output, then show that the learning objective can be decoupled into two distinct LoRA weight spaces: one is characterized by the $\\ell_2$ -loss for pixel-level regression, and another is characterized by the LPIPS and classifier score distillation losses to extract semantic information from pre-trained classification and SD models. In its default setting, PiSA-SR can be performed in a single diffusion step, achieving leading real-world SR results in both quality and efficiency. By introducing two adjustable guidance scales on the two LoRA modules to control the strengths of pixel-wise fidelity and semantic-level details during inference, PiSA-SR can offer flexible SR results according to user preference without re-training. The source code of our method can be found at https://github.com/csslc/PiSA-SR.", + "bbox": [ + 88, + 340, + 485, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 795, + 220, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single image super-resolution (SR) [53] aims to reconstruct a high-quality (HQ) image from its low-quality (LQ) counterpart suffering from various degradations, e.g., noise, blur,", + "bbox": [ + 89, + 816, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3f1dbd479a492cd7e51b28cdbaef7cbf576e830eb36022417feeb227c97368cb.jpg", + "image_caption": [ + "Figure 1. Visual illustration of our pixel- and semantic-level adjustable method for real-world SR. By increasing the pixel-level guidance scale $\\lambda_{pix}$ , the image degradations such as noise and compression artifacts can be gradually removed; however, a too-strong $\\lambda_{pix}$ will make the SR image over-smoothed. By increasing the semantic-level guidance scale $\\lambda_{sem}$ , the SR images will have more semantic details; nonetheless, a too-high $\\lambda_{sem}$ will generate visual artifacts. Please zoom in for a better view." + ], + "image_footnote": [], + "bbox": [ + 531, + 309, + 893, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "downsampling, etc. SR is a challenging ill-posed problem due to the inherent ambiguity in reconstructing details [3, 29], and various deep neural networks (DNNs) [5, 8, 10, 22, 27, 68] have been developed to address it. The pixel-level regression losses, such as $\\ell_1$ and $\\ell_2$ losses, are essential to keep the pixel-level fidelity of SR outputs, yet they tend to produce over-smoothed details [3, 29]. The SSIM [52] and perceptual losses can alleviate this problem. In specific, SSIM measures the local structural similarity between the SR and ground-truth (GT) images, and the perceptual loss [17] extracts semantic features with a pre-trained classification model [39] to improve the per", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. This work is supported by the PolyU-OPPO Joint Innovative Research Center.", + "bbox": [ + 89, + 875, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2333", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ceptual quality of SR results. The adversarial loss associated with the generative adversarial network (GAN) [24, 28, 29, 48, 50] provides a more effective solution than SSIM/perceptual losses to align SR images with natural image distribution, thereby leading to perceptually more realistic SR results. However, it simultaneously introduces many undesirable visual artifacts [29] because the adversarial training is unstable and the limited capacity of GAN models in characterizing the space of natural images.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently developed large-scale stable diffusion (SD) models [37], which are pre-trained on text-to-image (T2I) tasks, have demonstrated impressive semantic understanding and have been successfully applied to many downstream tasks [35, 66], including SR [30, 42, 46, 56, 57, 59]. While SD-based SR approaches have exhibited perceptually more realistic SR results than GAN-based methods [7], they commonly entangle the pixel-level fidelity and semantic-level enhancement objectives in the diffusion process [46, 56, 57], which can be contradictory in optimization [3, 43]. As a result, they struggle to balance the pixel-wise fidelity and semantic-level perception in the final SR output. While some methods [30, 42, 59] perform pixel-level restoration before semantic-level enhancement, the performance of the latter stage depends heavily on the accuracy of the earlier stage.", + "bbox": [ + 89, + 229, + 482, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides the trade-off between pixel-level fidelity and semantic-level details, users in practice often have different preferences on the SR results: some prioritize content fidelity over detail generation, while others favor semantically rich details over pixel-wise fidelity. This diversity in user preferences underscores the importance and demands for a more flexible SR approach that can accommodate individual tastes during inference. Although some previous methods [11, 34] have explored interactive SR methods to control the restoration strength, they are primarily limited to degradation-level adjustments (such as noise and blur), resulting in either smoother or sharper outputs. Several multi-step SD-based SR methods [30, 57, 59, 61] employ guidance-based strategies [14] at each sampling step to achieve varying levels of semantic richness, but these methods often struggle with precise control and efficiency.", + "bbox": [ + 89, + 474, + 482, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present Pixel-level and Semantic-level Adjustable Super-Resolution, namely PiSA-SR, which separates pixel-level and semantic-level enhancement into two distinct Low-Rank Adapter (LoRA) [16] weight spaces with the pre-trained SD model, providing an effective SR model for various user preferences. We begin by formulating the SD-based SR problem to learn the residual between LQ and HQ latent features [12, 64, 69]. With this formulation, we can not only accelerate the convergence of model training but also introduce scale factors on the model outputs, enabling flexible adjustment on the SR results during the inference stage without re-training. Then, we fine-tune the", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pre-trained SD model by introducing two separate LoRA modules dedicated to pixel-level regression and semantic-level enhancement. In specific, the $\\ell_2$ loss is employed for the LoRA module of pixel-level regression, while the LPIPS [67] and classifier score distillation (CSD) [62] losses are employed for another LoRA module, leveraging the semantic priors encoded in the pre-trained VGG classification model [39] and SD image generation model [37]. A decoupled training approach is proposed to train pixel- and semantic-level LoRAs, effectively improving semantic information while maintaining pixel-level fidelity.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments demonstrate that PiSA-SR not only achieves superior SR performance to existing SD-based models but also offers an effective way for users to adjust the SR styles based on their tastes. An example is shown in Fig. 1. The horizontal and vertical axes indicate the semantic-level and pixel-level guidance scales, respectively. Increasing the pixel-level scale removes noise and compression artifacts, but a too-strong scale over-smoothes image details. Conversely, increasing the semantic-level scale enriches the image details. Nonetheless, a too-high semantic scale over-enhances visual artifacts. PiSA-SR provides the flexibility to adjust both pixel-level and semantic-level guidance scales based on user preferences.", + "bbox": [ + 511, + 258, + 906, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 469, + 653, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image super-resolution. Early deep learning-based SR methods [5, 10, 22, 27] aim to improve image fidelity metrics such as PSNR and SSIM [52]. By using bicubic down-sampling to generate LQ-HQ image training pairs, various approaches have been developed to enhance the SR performance, such as dense [44], residual [69], recursive connections [22], non-local strategies [47], and attention mechanisms [8, 27, 68]. Though these advancements have significantly improved the SR performance, the real-world LQ images with complex degradations are much more challenging to enhance. While researchers have proposed to collect real-world LQ-HQ paired datasets [4, 55] to train SR models, simulating realistic LQ-HQ pairs is a more economic way [50, 65]. BSRGAN [65] randomly shuffles some basic degradation operators to synthesize LQ-HQ pairs, while RealESRGAN [50] implements high-order degradation modeling. Compared with pixel-level losses, the perceptual loss [17] and GAN loss [24, 48] are more effective in improving image perceptual quality. However, the adversarial training also leads to unnatural visual artifacts in the SR results. Many of the later works aim to reduce GAN-generated artifacts from the perspectives of frequency division [25], network optimization [70], the weighting of different losses [43, 49], and image local statistics [29].", + "bbox": [ + 511, + 492, + 903, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion model based super-resolution. Recent studies have explored the use of diffusion models (DMs) [15, 37] for SR. Early approaches adjust pre-trained DMs'", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2334", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "reverse transitions [20] using gradient descent, offering training-free solutions by assuming predefined image degradations. ResShift [63] trains a DM from scratch on paired LQ-HQ data. Recently, SD models pre-trained on T2I tasks have been widely used for SR tasks due to their strong image priors [2, 6, 30, 32, 36, 46, 51, 56, 57, 59]. StableSR [46] introduces a trainable encoder and takes LQ image as a condition in SD. DiffBIR [30] first applies a restoration module to mitigate degradations, then utilizes SD to enhance details. PASD [59] employs an encoder for degradation removal and introduces a pixel-aware cross-attention module to incorporate low-level and high-level image features into the SD process. SeeSR [57] enhances semantic robustness through degradation-aware tag-style prompts to guide diffusion. Considering that the multi-step diffusion process will increase the computational cost and the risk of synthesizing unfaithful contents, researchers started to develop fewer-step DM-based SR models. SinSR [51] applies consistency-preserving distillation to shorten the diffusion process of ResShift. OSEDiff [56] employs the LQ image as direct input, eliminating the random noise sampling. It adopts the VSD loss [54, 60] to distill the generative capacity from the multi-step diffusion process, offering a one-step DM-based SR solution. However, all the above-mentioned methods face a contradiction between pixel-level and semantic-level enhancement and most of them lack the flexibility to adjust the SR style for diverse user preferences.", + "bbox": [ + 89, + 90, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 513, + 225, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section first formulates the SD-based SR as a residual learning model, then introduces the dual-LoRA approach to disentangle the learning objectives of pixel-level regression and semantic-level enhancement. Finally, pixel and semantic guidances are proposed for flexible SR results generation. In our following development, we denote by $x_{L}$ and $x_{H}$ the LQ and HQ images, respectively, and by $z_{L}$ and $z_{H}$ their corresponding latent codes. Let $E$ and $D$ be the encoder and decoder of a well-trained variational auto-encoder (VAE), we can nearly have $z_{L} = E(x_{L})$ , $z_{H} = E(x_{H})$ , $x_{L} = D(z_{L})$ and $x_{H} = D(z_{H})$ .", + "bbox": [ + 89, + 536, + 483, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Formulation", + "text_level": 1, + "bbox": [ + 89, + 714, + 276, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DMs such SD [37] apply a $T$ -step forward process to progressively transform an initial latent code $z_0$ into Gaussian noise $z_T$ over $T$ steps. In each timestep $t$ , the noisy latent code $z_t$ is obtained by directly adding noise $\\epsilon \\sim \\mathcal{N}(0, I)$ to $z_0$ by $z_t = \\sqrt{\\bar{\\alpha}_t} \\cdot z_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\cdot \\epsilon$ , where $\\bar{\\alpha}_t$ is the cumulative parameter controlling the noise level at timestep $t$ . During the reverse process, multi-step SD-based SR models [2, 30, 36, 46, 57, 59] usually perform a $T$ -step denoising process to gradually transform Gaussian noise $z_T$ into the HQ latent code $z_H$ , conditioned on the LQ image $x_L$ with a ControlNet [66] (see Fig. 2 (a)). At stage $t$ , the latent", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/21ba05524be1c5cd4a637aa37b3630e901b28adaf25b29f1639e41a8831cbd49.jpg", + "image_caption": [ + "(a) Multi-step DM-based SR methods" + ], + "image_footnote": [], + "bbox": [ + 521, + 87, + 901, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d12052bcddc86787f481f76a8a7c1394d95c7c81e5e521133daba8c5c641e221.jpg", + "image_caption": [ + "(b) OSEDiff" + ], + "image_footnote": [], + "bbox": [ + 521, + 198, + 898, + 242 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5ebbbb204c307be2d71598fccbd173d6060bec51e5349b7a8a0029c77a195c19.jpg", + "image_caption": [ + "(c) PiSA-SR", + "Figure 2. Comparison of the pipeline of different DM-based SR methods. (a) Multi-step methods [2, 30, 36, 46, 57, 59] perform $T$ denoising steps starting from Gaussian noise $z_{T}$ , conditioned on the LQ image $x_{L}$ . (b) OSEDiff [56] starts from LQ latent representation $z_{L}$ with only one-step diffusion. (c) Our proposed PiSA-SR formulates the SD-based SR as learning the residual between the LQ latent $z_{L}$ and HQ latent $z_{H}$ ." + ], + "image_footnote": [], + "bbox": [ + 521, + 257, + 898, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\hat{z}_0$ can be computed from the noise estimated by SD UNet $\\epsilon_{\\theta}(z_t,x_L,t)$ and the current latent $z_{t}$ as follows:", + "bbox": [ + 511, + 433, + 906, + 465 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {0} = f (z _ {t}, \\epsilon_ {\\theta}) = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (z _ {t} , x _ {L} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 474, + 905, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then the latent $z_{t-1}$ at timestep $t - 1$ can be sampled from $p(z_{t-1} | z_t, \\hat{z}_0) = N(\\mu_t(z_t, \\hat{z}_0), \\sigma_t^2)$ , where $\\mu_t$ and $\\sigma_t^2$ denote the mean and variance of $z_{t-1}$ , respectively.", + "bbox": [ + 511, + 517, + 905, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these multi-step DM-based SR methods are computationally expensive and often yield unstable results due to the random noise sampled in the diffusion process [42, 57]. To address this issue, OSEDiff [56] starts from the LQ latent code $z_{L}$ and completes the SR process with only one diffusion step (see Fig. 2 (b)), improving both the efficiency and stability. The transformation between $z_{L}$ and $z_{H}$ can be described as follows:", + "bbox": [ + 511, + 563, + 905, + 681 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {H} = f \\left(z _ {L}, \\epsilon_ {\\theta}\\right) = \\frac {z _ {L} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(z _ {L}\\right)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 691, + 905, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that we omit the timestep $t$ in the above formula since $t$ can be viewed as a constant in one-step diffusion.", + "bbox": [ + 511, + 734, + 905, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning the residual between the LQ and GT features has been successfully applied in deep learning-based restoration methods [64, 69]. However, DMs typically perform multi-step iterations, making applying residual learning difficult. The recently developed OSEDiff [56] employs a single diffusion step to achieve end-to-end SR training, making it feasible to adopt the residual learning strategy. Here, we formulate the SR problem as learning the residual between $z_{L}$ and $z_{H}$ , as illustrated in Fig. 2 (c). We use", + "bbox": [ + 511, + 765, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2335", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "subtraction in the global residual connection because DMs are trained to denoise in the inverse diffusion process. Such a residual learning formulation helps the model focus on learning the desired high-frequency information from the HQ latent features, avoiding extracting less relevant information from the LQ latent. It can also accelerate the convergence of the model training process [12]. Furthermore, we introduce a scaling factor $\\lambda$ to adjust the residual $\\epsilon_{\\theta}$ added to the LQ latent, which can be expressed as:", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {H} = z _ {L} - \\lambda \\epsilon_ {\\theta} (z _ {L}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 237, + 482, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The $\\lambda$ is fixed as 1 during training. Users can adjust the output in the inference stage by using a smaller $\\lambda$ (to preserve more original contents) or a higher $\\lambda$ (to enhance details more aggressively) based on their preferences.", + "bbox": [ + 89, + 261, + 483, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dual-LoRA Training", + "text_level": 1, + "bbox": [ + 89, + 329, + 290, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous SD-based SR methods mostly entangle the pixel-level and semantic-level enhancement in the training process [46, 56], making the balance of content fidelity and perpetual quality difficult. Some works sequentially perform pixel-level and semantic-level enhancement using two separate networks [30]. However, errors in the earlier stage may propagate to the next stage, limiting overall performance. Furthermore, using two separate networks increases the computational burden and memory requirements, making the approach less efficient. It has been successfully applied in customized T2I tasks using distinct LoRA modules for various generation styles [18, 38, 41]. Inspired by this, we propose a decoupled training approach by utilizing two LoRA modules under the pre-trained SD model, targeting pixel-level and semantic-level enhancement for the SR task respectively. Our approach, namely Dual-LoRA, introduces only a small number of additional parameters during training. These LoRA parameters can also be merged into the pre-trained SD model during inference.", + "bbox": [ + 89, + 352, + 483, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 (a) illustrates the training process of our approach. We freeze the parameters of the pre-trained VAE and introduce two trainable LoRA modules to the UNet of SD. Since the LQ image is corrupted by degradations such as noise, blur, and downsampling, we optimize the pixel-level LoRA first to reduce the effect of corruptions, followed by the optimization of semantic-level LoRA. We train the pixel-level LoRA, denoted by $\\Delta \\theta_{pix}$ , by the pixel-level loss function. Together with the pre-trained SD parameters, the full parameter set can be represented as $\\theta_{pix} = \\{\\theta_{sd}, \\Delta \\theta_{pix}\\}$ . The HQ latent can then be estimated by $z_H^{pix} = z_L - \\epsilon_{\\theta_{pix}}(z_L)$ and decoded by the VAE decoder with $x_H^{pix} = D(z_H^{pix})$ .", + "bbox": [ + 89, + 640, + 483, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For semantic-level enhancement, we train another LoRA, denoted by $\\Delta \\theta_{sem}$ , by the semantic-level loss function. To disentangle the pixel-level and semantic-level objectives, we fix the already trained pixel-level LoRA $\\Delta \\theta_{pix}$ and combine it with the semantic-level LoRA $\\Delta \\theta_{sem}$ to be", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trained, forming a PiSA-LoRA group (see Fig. 3 (a)). Note that only the semantic-level LoRA module within the PiSA-LoRA group is updated in this stage. This ensures that the optimization process focuses on the semantic details without much interference from the pixel-level degradations. The full parameter set during this PiSA-LoRA training stage can be expressed as $\\theta_{PiSA} = \\{\\theta_{sd},\\Delta \\theta_{pix},\\Delta \\theta_{sem}\\}$ . The HQ latent after this stage is given by $z_H^{sem} = z_L - \\epsilon_{\\theta_{PiSA}}(z_L)$ , which is then decoded by the VAE decoder with $x_H^{sem} = D(z_H^{sem})$ .", + "bbox": [ + 511, + 90, + 906, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pixel-level and Semantic-level Losses", + "text_level": 1, + "bbox": [ + 511, + 253, + 831, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train the pixel-level LoRA using the $\\ell_2$ loss. As shown by $D(\\epsilon_{\\theta_{pix}}(z_L))$ in Fig. 4, the $\\ell_2$ loss can effectively remove degradation and enhance edges. However, it is insufficient to generate semantic-level details [24, 29, 48], resulting in smooth SR outputs. The LPIPS loss [67] can be used to activate semantic details by aligning the high-level features with a pre-trained classification VGG network [39], which is however trained on a limited set of image categories. The GAN loss [24] captures semantic information through adversarial training, encouraging the generator to produce more realistic images. However, GAN loss can be unstable in training and generate much artifacts [29].", + "bbox": [ + 511, + 277, + 906, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently developed SD models [37] have shown superior performance in generating open-category images with enhanced semantic details. As a conditional generation model, SD can be used to form up an implicit classifier [14] for the posterior distribution modeling, and its gradient w.r.t. the synthesized image $x_{t}$ is as follows:", + "bbox": [ + 511, + 460, + 906, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log p _ {\\theta} (c \\mid x _ {t}) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t} \\mid c) - \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ = - \\left(\\epsilon_ {\\theta} \\left(x _ {t}, c, t\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right) / \\sigma_ {t}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 563, + 903, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ is the text prompt, $t$ is the current timestep, and $\\sigma_{t} = \\sqrt{1 - \\bar{\\alpha}_{t}}$ . The SD model, parameterized by $\\theta$ , model score functions of $\\nabla_{x_t}\\log p_\\theta (x_t\\mid c)$ and $\\nabla_{x_t}\\log p_\\theta (x_t)$ using noise predictions [40]. By taking the expectation of the gradient in Eq. (4) over all possible $t$ , we have the CSD [14] loss, which is originally used in 3D generation tasks to optimize the posterior probability of rendered images aligning their semantic content with text prompts.", + "bbox": [ + 511, + 613, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the functionality of CSD loss in DMs for generation, we investigate the usage of CSD loss for semantic-level enhancement in the SR task. Following [56, 60], and for the clarity of the following context, we formulate the gradient of CSD in the latent space instead of the noise domain. Specifically, the gradient of CSD loss can be formulated in Eq. (5) below (the detailed derivation can be found in the supplementary materials):", + "bbox": [ + 511, + 734, + 906, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\ell_ {C S D} ^ {\\lambda_ {c f g}} = \\mathbb {E} _ {t, \\epsilon , z _ {t}, c} \\left[ w _ {t} \\left(f (z _ {t}, \\epsilon_ {r e a l}) - f (z _ {t}, \\epsilon_ {r e a l} ^ {\\lambda_ {c f g}})\\right) \\frac {\\partial z _ {H} ^ {s e m}}{\\partial \\theta_ {P i S A}} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 861, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2336", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/92a9c6f205b7f5bdf5817974d47b3fdc250ff0c02ef5bd1438db8d12d2d67032.jpg", + "image_caption": [ + "(a) Training process of the PiSA-SR" + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 888, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ceec4e76b76f126dfa4803000686b836e72b86bc51c0bfb975de412a3e0692e8.jpg", + "image_caption": [ + "(b) Inference process of PiSA-SR" + ], + "image_footnote": [], + "bbox": [ + 112, + 311, + 888, + 452 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/607e5f283805224f3efbfca2cf926a683d5a9f2476049342273a5ba62b4ff2b3.jpg", + "image_caption": [ + "Figure 3. The (a) training and (b) inference procedures of PiSA-SR. During the training process, two LoRA modules are respectively optimized for pixel-level and semantic-level enhancement. During the inference stage, users can use the default setting to reconstruct the HQ image in one-step diffusion or adjust $\\lambda_{pix}$ and $\\lambda_{sem}$ to control the strengths of pixel-level and semantic-level enhancement.", + "Figure 4. The model outputs with pixel-wise and semantic-level losses for a given LQ image." + ], + "image_footnote": [], + "bbox": [ + 94, + 531, + 480, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the gradient expectation is calculated considering all diffusion timesteps $t = 1,\\dots,T$ , the noise is sampled from $\\epsilon \\sim \\mathcal{N}(0,I),z_{t}$ is obtained by $z_{t} = \\sqrt{\\bar{\\alpha}_{t}}\\cdot z_{H}^{sem} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\cdot \\epsilon$ $c$ is the text prompt extracted from $z_H^{sem}$ $f(\\cdot)$ is the function as Eq. (1), $w_{t}$ is the timestep-dependent scalar weight, and $\\epsilon_{real}^{\\lambda_{cfg}}$ denotes the pre-trained SD output with the CFG term $\\epsilon_{real}^{cls}$ same as Eq. (4) [14, 62]:", + "bbox": [ + 89, + 750, + 482, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {r e a l} ^ {\\lambda_ {c f g}} = \\epsilon_ {r e a l} (z _ {t}, t) + \\underbrace {\\lambda_ {c f g} (\\epsilon_ {r e a l} (z _ {t} , t , c) - \\epsilon_ {r e a l} (z _ {t} , t))}. \\qquad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 863, + 482, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the VSD loss [54, 60], which also aims to align the distributions of enhanced and natural images within the latent space, has been verified effective for SR by OSEDiff [56]. The gradient of VSD loss can be written as:", + "bbox": [ + 511, + 535, + 906, + 597 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\ell_ {V S D} ^ {\\lambda_ {c f g}} = \\mathbb {E} _ {t, \\epsilon , z _ {t}, c} \\left[ w _ {t} \\left(f (z _ {t}, \\epsilon_ {f a k e}) - f (z _ {t}, \\epsilon_ {r e a l} ^ {\\lambda_ {c f g}})\\right) \\frac {\\partial z _ {H} ^ {s e m}}{\\partial \\theta_ {P i S A}} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 614, + 905, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\text{fake}}$ denotes the output of fine-tuned SD aligned with the distribution of synthetic images. The VSD loss with $\\lambda_{cfg}$ can be divided into two components: the VSD loss with $\\lambda_{cfg} = 0.0$ , i.e., $\\ell_{VSD}^{\\lambda_{cfg} = 0.0}$ , and the CSD loss. Coupling with the LPIPS loss, we visualize the semantic-level LoRA optimization results of the two components of VSD in Fig. 4. We see that the CSD loss with the normal guidance scale of $\\lambda_{cfg} = 7.5$ contributes more significantly to semantic enhancement. In contrast, the VSD loss with $\\lambda_{cfg} = 0.0$ weakens the semantic details. More importantly, the component $\\ell_{VSD}^{\\lambda_{cfg} = 0.0}$ requires bi-level optimization, leading to intensive memory usage and unstable training [31]. In comparison, the CSD loss is free of bi-level optimization, reducing significantly memory usage and improving training stability. Therefore, we integrate LPIPS and CSD losses for the semantic-level LoRA optimization.", + "bbox": [ + 511, + 654, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2337", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. The Inference Process of PiSA-SR", + "text_level": 1, + "bbox": [ + 89, + 90, + 387, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3(b) shows the inference process of PiSA-SR. In the default inference setting (represented by the green dash-dot line), only the PiSA-LoRA, which merges the pixel-level and semantic-level LoRA modules, is used to process the input together with the pre-trained SD model, achieving state-of-the-art SR performance in just one-step diffusion (see our experiments in Section 4.3). To achieve flexible SR with diverse user preferences, we extend the strategy in Eq. (3) by introducing a pair of pixel and semantic guidance factors, denoted by $\\lambda_{pix}$ and $\\lambda_{sem}$ , to control the SR results:", + "bbox": [ + 89, + 109, + 483, + 261 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\theta} \\left(z _ {L}\\right) = \\lambda_ {p i x} \\epsilon_ {\\theta_ {p i x}} \\left(z _ {L}\\right) + \\lambda_ {s e m} \\left(\\epsilon_ {\\theta_ {P s S A}} \\left(z _ {L}\\right) - \\epsilon_ {\\theta_ {p i x}} \\left(z _ {L}\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 272, + 480, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Eq. (8), $\\epsilon_{\\theta_{pix}}(z_L)$ is the output with only pixel-level LoRA, while $\\epsilon_{\\theta_{PiSA}}(z_L)$ is the output with both pixel and semantic level enhancement. Inspired by CFG [14], the difference between them, i.e., $\\epsilon_{\\theta_{PiSA}}(z_L) - \\epsilon_{\\theta_{pix}}(z_L)$ , can well separate pixel-level degradations from semantic details (see the red box in Fig. 4). Therefore, as illustrated in Fig. 3(b) (represented by the black dashed line), by adjusting $\\lambda_{pix}$ and $\\lambda_{sem}$ , we can control the contributions of pixel-level LoRA and semantic-level LoRA, and consequently deliver the SR results with different styles, as shown in Fig. 1.", + "bbox": [ + 89, + 303, + 483, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 467, + 223, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 488, + 297, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training settings. We train PiSA-SR upon SD 2.1-base for the $\\times 4$ SR task. The two trainable LoRA modules are applied to the weights of all convolution and MLP layers. Both modules are initialized using a Gaussian distribution with a rank of 4. Following SeeSR [57] and OSEDiff [56], we employ LSDIR [26] and the first 10K images from FFHQ [19] as the training data. We generate paired training data using the degradation pipeline from RealESRGAN [50]. The batch size is 16 and the training patch size is $512 \\times 512$ . The Adam optimizer [23] is used with a learning rate 5e-5. The pixel-level and semantic-level LoRA modules undergo 4K and 8.5K training iterations.", + "bbox": [ + 89, + 507, + 483, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. We compare PiSA-SR with leading multi-step DM-based methods StableSR [46], ResShift [63], DiffBIR [30], PASD [59], and SeeSR [57], one-step DM-based methods SinSR [51] and OSEDiff [56], and GAN-based SR methods RealESRGAN [50], BSRGAN [65] and LDL [29]. All comparative results are obtained using officially released codes or models. Due to space constraints, comparisons with GAN-based methods are presented in the supplementary materials.", + "bbox": [ + 89, + 688, + 483, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test datasets and evaluation metrics. Following previous arts [46, 56, 57], we evaluate the competing methods using synthetic and real-world test data. The synthetic dataset comprises 3000 images cropped from DIV2K [1] with size $512 \\times 512$ , degraded using RealESRGAN", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e2fd0c5b79a32616af4a4425e0a9358cee20993e6120b6e63485966c249f793b.jpg", + "table_caption": [ + "Table 1. Results of PiSA-SR with different pixel-semantic guidance scales on the RealSR test dataset." + ], + "table_footnote": [], + "table_body": "
λpixλsemPSNR↑LIPIS↓CLIPIQA↑MUSIQ↑
0.01.025.960.34260.412946.45
0.21.026.480.30420.486854.05
0.51.026.750.26460.570563.82
0.81.026.180.26120.629268.95
1.01.025.500.26720.670270.15
1.21.024.760.27230.674670.33
1.51.023.740.27690.630569.23
1.00.026.920.30180.322749.62
1.00.226.950.27840.359153.64
1.00.526.770.24760.432258.76
1.00.826.200.24650.580666.33
1.01.025.500.26720.670270.15
1.01.224.590.30000.701571.60
1.01.523.080.35410.683571.76
", + "bbox": [ + 517, + 119, + 903, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "degradation pipeline [50]. The real-world data are center-cropped from RealSR [4] and DrealSR [55] datasets with size $128 \\times 128$ for LQ images and $512 \\times 512$ for HQ images. PSNR and SSIM [52], computed on the Y channel in the YCbCr space, are used to measure the fidelity of SR results; LPIPS [67] and DISTS [9], computed in the RGB space, are used to assess the perceptual quality of SR results; FID [13] evaluates the distance of distributions between GT and restored images; NIQE [33], CLIPIQA [45], MUSIQ [21] and MANIQA [58] evaluate the image quality without reference images. Additionally, a user study is conducted to validate the effectiveness of the proposed adjustable SR in the supplementary materials.", + "bbox": [ + 509, + 290, + 906, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experiments on Adjustable SR", + "text_level": 1, + "bbox": [ + 511, + 501, + 785, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments by fixing one guidance scale ( $\\lambda_{pix}$ or $\\lambda_{sem}$ ) at 1 and varying the other to observe how the reconstructed images change. PSNR, LPIPS, CLIPIQA, and MUSIQ are used to evaluate performance. PSNR measures pixel-level fidelity; LPIPS evaluates the image perceptual quality with GT as reference; CLIPIQA and MUSIQ measure the image quality without using a reference. The results are shown in Table 1. First, we observe that increasing the pixel-level scale $\\lambda_{pix}$ leads to a continuous improvement in the no-reference metrics CLIPIQA and MUSIQ. This is because increasing $\\lambda_{pix}$ can remove image degradation and enhance edge, which is favorable to these two metrics. However, the reference-based metrics PSNR and LPIPS exhibit a rise-and-fall pattern. PSNR peaks at $\\lambda_{pix} = 0.5$ , indicating the best pixel-level fidelity. LPIPS reaches its best value at $\\lambda_{pix} = 0.8$ (0.2612), suggesting that at this point the restored image is perceptually most similar to the GT with rich details. Further increasing $\\lambda_{pix}$ reduces both PSNR and LPIPS scores. Second, increasing the semantic-level scale $\\lambda_{sem}$ also results in a continuous improvement in CLIPIQA and MUSIQ, but with a higher upper bound than pixel-level adjustments. This is because increasing $\\lambda_{sem}$ can synthesize more semantic-level details (see Fig. 1). However, PSNR decreases as $\\lambda_{sem}$ increases, while LPIPS first improves and peaks at $\\lambda_{sem} = 0.8$ (0.2465), and then", + "bbox": [ + 509, + 523, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2338", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d0b1567fb7c1d01c4c998105c7c5446209d6208495b59b3165c67be8b17c2d97.jpg", + "table_caption": [ + "Table 2. Quantitative comparison among the state-of-the-art DM-based SR methods on synthetic and real-world test datasets. 'S' denotes the number of diffusion steps. The best and the second-best results are highlighted in red and blue, respectively." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodsPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓CLIPQA↑MUSIQ↑MANIQA↑
DIV2KResShift-S1524.690.61750.33740.221536.016.820.608960.920.5450
StableSR-S20023.310.57280.31290.213824.674.760.668265.630.6188
DiffBIR-S5023.670.56530.35410.212930.934.710.665265.660.6204
PASD-S2023.140.54890.36070.221929.324.400.671168.830.6484
SeeSR-S5023.710.60450.32070.196725.834.820.685768.490.6239
SinSR-S124.430.60120.32620.206635.456.020.649962.800.5395
OSEDiff-S123.720.61080.29410.197626.324.710.668367.970.6148
PiSA-SR-S123.870.60580.28230.193425.074.550.692769.680.6400
RealSRResShift-S1526.310.74110.34890.2498142.817.270.545058.100.5305
StableSR-S20024.690.70520.30910.2167127.205.760.619565.420.6211
DiffBIR-S5024.880.66730.35670.2290124.565.630.641264.660.6231
PASD-S2025.220.68090.33920.2259123.085.180.650268.740.6461
SeeSR-S5025.330.72730.29850.2213125.665.380.659469.370.6439
SinSR-S126.300.73540.32120.2346137.056.310.620460.410.5389
OSEDiff-S125.150.73410.29210.2128123.505.650.669369.090.6339
PiSA-SR-S125.500.74170.26720.2044124.095.500.670270.150.6560
DrealSRResShift-S1528.450.76320.40730.2700175.928.280.525949.860.4573
StableSR-S20028.040.74600.33540.2287147.036.510.617158.500.5602
DiffBIR-S5026.840.66600.44460.2706167.386.020.629260.680.5902
PASD-S2027.480.70510.38540.2535157.365.570.671464.550.6130
SeeSR-S5028.260.76980.31970.2306149.866.520.667264.840.6026
SinSR-S128.410.74950.37410.2488177.057.020.636755.340.4898
OSEDiff-S127.920.78350.29680.2165135.296.490.696364.650.5899
PiSA-SR-S128.310.78040.29600.2169130.616.200.697066.110.6156
", + "bbox": [ + 94, + 121, + 903, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "declines. This is because excessive semantic details may induce changes in image content and reduce pixel-level fidelity. Meanwhile, the over-enhanced images may exhibit many details different from the GT, making the reference-based LPIPS index deteriorate.", + "bbox": [ + 88, + 415, + 480, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 1 illustrates the visual comparisons of different pixel-semantic scales. More comparisons can be found in the supplementary materials. From bottom to top, as the pixel-level scale $\\lambda_{pix}$ increases, the model progressively removes degradation from the input LQ image, resulting in clearer and sharper visual output. However, a too-high $\\lambda_{pix}$ can lead to loss of details, and diminish the effectiveness of semantic improvements, as seen in the top-right corner of Fig. 1. From left to right, increasing the semantic-level scale $\\lambda_{sem}$ enhances the semantic richness of the image, making the details of the wrinkles, beard, and hair of Albert Einstein more pronounced. However, an excessive semantic level such as $\\lambda_{sem} = 1.5$ introduces unnatural artifacts, making the SR result less realistic. With the flexibility of PiSASR, users can customize their preferences, either preserving more fidelity or emphasizing richer semantic enhancements, depending on their specific use case.", + "bbox": [ + 91, + 491, + 482, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparisons with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 89, + 758, + 401, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Eq. (8), letting $\\lambda_{pix} = 1$ and $\\lambda_{sem} = 1$ , we have the default version of PiSA-SR, which can be performed in just one-step diffusion. Here, we compare this default version of PiSA-SR with the competing methods.", + "bbox": [ + 89, + 777, + 482, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparisons. Table 2 compares the performance of our default PiSA-SR model with other DM-based Real-SR methods. The following findings can be made. ResShift [63] and its distilled version SinSR [51] show lim", + "bbox": [ + 89, + 840, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ited performance in no-reference metrics, implying poor perceptual quality. PASD [59] and SeeSR [57] leverage the pre-trained SD as their base model, and extract additional high-level information [71] to improve the perceptual quality of images. Their no-reference metric scores, such as MUSIQ and CLIPIQA, are highly improved. However, the multi-step inference makes them less efficient. In addition, their relatively low LPIPS and DISTS scores suggest their less faithful reconstruction to the GT. OSEDiff [56] distills the generation capacity of multi-step DM into just one-step diffusion by the VSD loss [54, 60]. While effective and efficient, its no-reference metrics are limited. Our proposed PiSA-SR also exhibits high efficiency by requiring only one inference step, while showing impressive pixel-level fidelity and semantic-level perceptual quality. It outperforms other methods not only in reference-based metrics such as LPIPS and DISTS but also in no-reference metrics such as CLIP-IQA, MUSIQ, and MANIQA.", + "bbox": [ + 511, + 415, + 906, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparisons. We present visual comparisons in Fig. 5. ResShift and SinSR fail to generate richer textures and details because of their limited generation capacity. StableSR and DiffBIR can generate finer details (e.g., more realistic penguin feathers in the second example), due to the leveraged strong SD prior. PASD and SeeSR incorporate high-level information guidance, resulting in more comprehensive details. However, the inherent randomness in PASD and SeeSR may lead to over-enhanced (e.g., unnatural penguin feathers generated by SeeSR) or over-smoothed details (e.g., the blurred texture of the wooden stool generated by SeeSR). OSEDiff can generate more consistent results in both examples. However, the restored details can be limited. In contrast, the proposed PiSA-SR delivers high-quality, re", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2339", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83e03c8fc3d1a51dc0c8c80547632ab54021d8500c334b73108bd595dc10a4a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 88, + 318, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c94bb7aeb1ca51df5be79f5d4df3861f8e4419d41172956c38f4a290d045125.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 88, + 504, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8821bc87bfc13e4e8b9d850820813b08cca674b5887ae7eecf6034a4ec839543.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 88, + 844, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ab7a306b01b1a7ba098f4159268641d8d75cf459971510514f8f9d395220843.jpg", + "image_caption": [ + "Figure 5. Visual comparisons of different DM-based SR methods. Please zoom in for a better view." + ], + "image_footnote": [], + "bbox": [ + 130, + 246, + 316, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6537849122bd587a018edc04a77c6cc65c4c3c0f4587192d3892eef8d5b2b73c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 246, + 506, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15f6fa77e3e01fcdfce6f529044557eda47a85d4d439d6140062f7156a52b17a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 246, + 844, + 401 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/efd6af792f58936aa5d84c8e55e81607cae63649f6139f61a1ccec13cb0dd4bd.jpg", + "table_caption": [ + "Table 3. The inference time and the number of parameters of DM-based SR methods." + ], + "table_footnote": [], + "table_body": "
StableSRResShiftDiffBIRPASDSeeSRSinSROSEDiffPiSA-SR-def.PiSA-SR-adj.
Inference Steps200155020501112
Inference time(s)/Image10.030.762.722.804.300.130.120.090.13
#Params(B)1.560.181.682.312.510.181.771.301.30
", + "bbox": [ + 94, + 439, + 906, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "alistic SR results. With the dual-LoRA design, the restored structures align well with the input LQ image, providing more accurate fidelity information and generating more natural and richer details. More visualization comparisons can be found in supplementary materials.", + "bbox": [ + 89, + 505, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Complexity comparisons. We compare the number of parameters and inference time of competing DM-based SR models in Table 3, where PiSA-SR-def. and PiSA-SR-adj. represent its default setting and adjustable setting, respectively. Inference time is measured on the $\\times 4$ SR task with $128\\times 128$ LQ images using a single NVIDIA A100 80G GPU. ResShift has fewer parameters because it is trained from scratch without using the pre-trained SD model. SinSR inherits the parameters from its parental ResShift model. Among the SD-based SR methods, PiSA-SR-def. has the fewest parameters and achieves the fastest inference time. Unlike OSEDiff, PiSA-SR-def. does not require a semantic extractor (e.g., RAM [71]) during inference. Instead, it distills the semantic information into the semantic-level LoRA with CSD. PiSA-SR-adj., which requires two diffusion steps to compute outputs from pixellevel and semantic-level LoRA modules, incurs a slightly longer inference time over the default setting. However, this small increase in time offers users the advantage of generating SR outputs tailored to their specific needs.", + "bbox": [ + 89, + 580, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. We performed a series of ablation stud", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ies for a better understanding of the dual-LoRA training approach, and the roles of the two LoRA modules. The details can be found in the supplementary materials.", + "bbox": [ + 511, + 505, + 906, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 511, + 561, + 763, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented PiSA-SR, a novel SR framework that disentangles pixel-level and semantic-level objectives by learning two LoRA modules upon the pre-trained SD model. By decoupling the learning process into two distinct spaces, PiSA-SR effectively balanced pixel-wise regression and perceptual quality. The commonly used $\\ell_2$ loss was used to optimize pixel-level LoRA, while the LPIPS and CSD losses were employed to optimize the semantic-level LoRA. PiSA-SR demonstrated impressive performance in both effectiveness and efficiency. Furthermore, it offered high flexibility during inference through adjustable guidance scales, allowing users to customize the SR results without model re-training.", + "bbox": [ + 509, + 583, + 906, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While PiSA-SR provided a flexible solution for adjusting pixel-level fidelity and semantic-level perception, it incurs a slight inference time increment over its default version. In addition, we use a single pixel-level LoRA to handle various degradations in the LQ, which may be insufficient for heavy degradations. In the future, we will explore refining separate LoRA spaces for different image degradations and the possibility of image-adaptive scale factors.", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2340", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. NTIRE 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 6", + "[2] Yang Ai, Xiaogiang Zhou, Huaibo Huang, Xiaotian Han, Zhengyu Chen, Quanzeng You, and Hongxia Yang. Dreamclear: High-capacity real-world image restoration with privacy-safe dataset curation. arXiv preprint arXiv:2410.18666, 2024.3", + "[3] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 1, 2", + "[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 6", + "[5] Dong Chao, Loy Chen Change, He Kaiming, and Tang Xiaou. Learning a deep convolutional network for image super-resolution. In ECCV, pages 184-199, 2014. 1, 2", + "[6] Bin Chen, Gehui Li, Rongyuan Wu, Xindong Zhang, Jie Chen, Jian Zhang, and Lei Zhang. Adversarial diffusion compression for real-world image super-resolution. arXiv preprint arXiv:2411.13383, 2024. 3", + "[7] Du Chen, Tianhe Wu, Kede Ma, and Lei Zhang. Toward generalized image quality assessment: Relaxing the perfect reference quality assumption, 2025. 2", + "[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 1, 2", + "[9] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 6", + "[10] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 1, 2", + "[11] Jingwen He, Chao Dong, Yihao Liu, and Yu Qiao. Interactive multi-dimension modulation for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9363-9379, 2021. 2", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 4", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 4, 5, 6" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2", + "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 1, 2", + "[18] Maxwell Jones, Sheng-Yu Wang, Nupur Kumari, David Bau, and Jun-Yan Zhu. Customizing text-to-image models with a single image pair. arXiv preprint arXiv:2405.01536, 2024. 4", + "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6", + "[20] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 3", + "[21] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 6", + "[22] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 1, 2", + "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[24] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2, 4", + "[25] Wenbo Li, Kun Zhou, Lu Qi, Liying Lu, and Jiangbo Lu. Best-Buddy GANs for highly detailed image superresolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1412-1420, 2022. 2", + "[26] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Lsdir: A large scale dataset for image restoration. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1775-1787, 2023. 6", + "[27] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2341", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, 2022. 2", + "[29] Jie Liang, Hui Zeng, and Lei Zhang. Details or Artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 1, 2, 4, 6", + "[30] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 2, 3, 4, 6", + "[31] Zhiyuan Ma, Yuxiang Wei, Yabin Zhang, Xiangyu Zhu, Zhen Lei, and Lei Zhang. Scaledreamer: Scalable text-to-3d synthesis with asynchronous score distillation. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 5", + "[32] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Conditional diffusion distillation. arXiv preprint arXiv:2310.01407, 2023. 3", + "[33] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6", + "[34] Chong Mou, Yanze Wu, Xintao Wang, Chao Dong, Jian Zhang, and Ying Shan. Metric learning based interactive modulation for real-world super-resolution. In European Conference on Computer Vision, pages 723-740. Springer, 2022. 2", + "[35] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2", + "[36] Yunpeng Qu, Kun Yuan, Kai Zhao, Qizhi Xie, Jinhua Hao, Ming Sun, and Chao Zhou. Xpsr: Cross-modal priors for diffusion-based image super-resolution. arXiv preprint arXiv:2403.05049, 2024. 3", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4", + "[38] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In European Conference on Computer Vision, pages 422-438. Springer, 2025. 4", + "[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 2, 4", + "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Nick Stracke, Stefan Andreas Baumann, Joshua M Susskind, Miguel Angel Bautista, and Björn Ommer. Ctrloralter: Conditional loradapter for efficient 0-shot control & altering of t2i models. arXiv preprint arXiv:2405.07913, 2024. 4", + "[42] Lingchen Sun, Rongyuan Wu, Zhengqiang Zhang, Hongwei Yong, and Lei Zhang. Improving the stability of diffusion models for content consistent super-resolution. arXiv e-prints, pages arXiv-2401, 2023. 2, 3", + "[43] Lingchen Sun, Jie Liang, Shuaizheng Liu, Hongwei Yong, and Lei Zhang. Perception-distortion balanced superresolution: A multi-objective optimization perspective. IEEE Transactions on Image Processing, 33:4444-4458, 2024. 2", + "[44] Tong Tong, Gen Li, Xiejie Liu, and Qinquan Gao. Image super-resolution using dense skip connections. In Proceedings of the IEEE international conference on computer vision, pages 4799-4807, 2017. 2", + "[45] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 6", + "[46] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 4, 6", + "[47] Risheng Wang, Tao Lei, Wenzheng Zhou, Qi Wang, Hongying Meng, and Asoke K Nandi. Lightweight non-local network for image super-resolution. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1625-1629. IEEE, 2021. 2", + "[48] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. ESRGAN: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 4", + "[49] Xintao Wang, Ke Yu, Chao Dong, Xiaou Tang, and Chen Change Loy. Deep network interpolation for continuous imagery effect transition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1701, 2019. 2", + "[50] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 2, 6", + "[51] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 3, 6, 7", + "[52] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 1, 2, 6", + "[53] Zhihao Wang, Jian Chen, and Steven C. H. Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions" + ], + "bbox": [ + 514, + 90, + 905, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2342", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, 2021. 1", + "[54] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3, 5, 7", + "[55] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 2, 6", + "[56] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 4, 5, 6, 7", + "[57] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 2, 3, 6, 7", + "[58] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6", + "[59] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 2, 3, 6, 7", + "[60] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 3, 4, 5, 7", + "[61] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. arXiv preprint arXiv:2401.13627, 2024. 2", + "[62] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and XIAOJUAN QI. Text-to-3d with classifier score distillation. In The Twelfth International Conference on Learning Representations, 2024. 2, 5", + "[63] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. arXiv preprint arXiv:2307.12348, 2023. 3, 6, 7", + "[64] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 2, 3", + "[65] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision, pages 4791-4800, 2021. 2, 6", + "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3", + "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 4, 6", + "[68] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. arXiv preprint arXiv:2203.06697, 2022. 1, 2", + "[69] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 2, 3", + "[70] Yuehan Zhang, Bo Ji, Jia Hao, and Angela Yao. Perception-distortion balanced ADMM optimization for single-image super-resolution. In European Conference on Computer Vision, pages 108-125. Springer, 2022. 2", + "[71] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 7, 8" + ], + "bbox": [ + 516, + 92, + 903, + 472 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2343", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/6dc2236b-e40d-40ee-98e0-34b3414c88ae_model.json b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/6dc2236b-e40d-40ee-98e0-34b3414c88ae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6c63e27d712de9f98c6272729256510e580e6681 --- /dev/null +++ b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/6dc2236b-e40d-40ee-98e0-34b3414c88ae_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.201, + 0.131, + 0.8, + 0.176 + ], + "angle": 0, + "content": "Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.203, + 0.882, + 0.24 + ], + "angle": 0, + "content": "Lingchen Sun\\(^{1,2}\\), Rongyuan Wu\\(^{1,2}\\), Zhiyuan Ma\\(^{1}\\), Shuaizheng Liu\\(^{1,2}\\), Qiaosi Yi\\(^{1,2}\\), Lei Zhang\\(^{1,2,*}\\) \n\\(^{1}\\)The Hong Kong Polytechnic University \n\\(^{2}\\)OPPO Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.242, + 0.686, + 0.258 + ], + "angle": 0, + "content": "{ling-chen.sun, rong-yuan.wu, zm2354.ma} @connect.polyu.hk" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.26, + 0.734, + 0.275 + ], + "angle": 0, + "content": "{shuaizhengliu21,qiaosiyijoyies} \\(@\\) gmail.com cslzhang@comp.polyu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.31, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.486, + 0.765 + ], + "angle": 0, + "content": "Diffusion prior-based methods have shown impressive results in real-world image super-resolution (SR). However, most existing methods entangle pixel-level and semantic-level SR objectives in the training process, struggling to balance pixel-wise fidelity and perceptual quality. Meanwhile, users have varying preferences on SR results, thus it is demanded to develop an adjustable SR model that can be tailored to different fidelity-perception preferences during inference without re-training. We present Pixel-level and Semantic-level Adjustable SR (PiSA-SR), which learns two LoRA modules upon the pre-trained stable-diffusion (SD) model to achieve improved and adjustable SR results. We first formulate the SD-based SR problem as learning the residual between the low-quality input and the high-quality output, then show that the learning objective can be decoupled into two distinct LoRA weight spaces: one is characterized by the \\(\\ell_2\\)-loss for pixel-level regression, and another is characterized by the LPIPS and classifier score distillation losses to extract semantic information from pre-trained classification and SD models. In its default setting, PiSA-SR can be performed in a single diffusion step, achieving leading real-world SR results in both quality and efficiency. By introducing two adjustable guidance scales on the two LoRA modules to control the strengths of pixel-wise fidelity and semantic-level details during inference, PiSA-SR can offer flexible SR results according to user preference without re-training. The source code of our method can be found at https://github.com/csslc/PiSA-SR." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.796, + 0.222, + 0.811 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Single image super-resolution (SR) [53] aims to reconstruct a high-quality (HQ) image from its low-quality (LQ) counterpart suffering from various degradations, e.g., noise, blur," + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.31, + 0.895, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.702 + ], + "angle": 0, + "content": "Figure 1. Visual illustration of our pixel- and semantic-level adjustable method for real-world SR. By increasing the pixel-level guidance scale \\(\\lambda_{pix}\\), the image degradations such as noise and compression artifacts can be gradually removed; however, a too-strong \\(\\lambda_{pix}\\) will make the SR image over-smoothed. By increasing the semantic-level guidance scale \\(\\lambda_{sem}\\), the SR images will have more semantic details; nonetheless, a too-high \\(\\lambda_{sem}\\) will generate visual artifacts. Please zoom in for a better view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "downsampling, etc. SR is a challenging ill-posed problem due to the inherent ambiguity in reconstructing details [3, 29], and various deep neural networks (DNNs) [5, 8, 10, 22, 27, 68] have been developed to address it. The pixel-level regression losses, such as \\(\\ell_1\\) and \\(\\ell_2\\) losses, are essential to keep the pixel-level fidelity of SR outputs, yet they tend to produce over-smoothed details [3, 29]. The SSIM [52] and perceptual losses can alleviate this problem. In specific, SSIM measures the local structural similarity between the SR and ground-truth (GT) images, and the perceptual loss [17] extracts semantic features with a pre-trained classification model [39] to improve the per" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author. This work is supported by the PolyU-OPPO Joint Innovative Research Center." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2333" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "ceptual quality of SR results. The adversarial loss associated with the generative adversarial network (GAN) [24, 28, 29, 48, 50] provides a more effective solution than SSIM/perceptual losses to align SR images with natural image distribution, thereby leading to perceptually more realistic SR results. However, it simultaneously introduces many undesirable visual artifacts [29] because the adversarial training is unstable and the limited capacity of GAN models in characterizing the space of natural images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.483, + 0.473 + ], + "angle": 0, + "content": "Recently developed large-scale stable diffusion (SD) models [37], which are pre-trained on text-to-image (T2I) tasks, have demonstrated impressive semantic understanding and have been successfully applied to many downstream tasks [35, 66], including SR [30, 42, 46, 56, 57, 59]. While SD-based SR approaches have exhibited perceptually more realistic SR results than GAN-based methods [7], they commonly entangle the pixel-level fidelity and semantic-level enhancement objectives in the diffusion process [46, 56, 57], which can be contradictory in optimization [3, 43]. As a result, they struggle to balance the pixel-wise fidelity and semantic-level perception in the final SR output. While some methods [30, 42, 59] perform pixel-level restoration before semantic-level enhancement, the performance of the latter stage depends heavily on the accuracy of the earlier stage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Besides the trade-off between pixel-level fidelity and semantic-level details, users in practice often have different preferences on the SR results: some prioritize content fidelity over detail generation, while others favor semantically rich details over pixel-wise fidelity. This diversity in user preferences underscores the importance and demands for a more flexible SR approach that can accommodate individual tastes during inference. Although some previous methods [11, 34] have explored interactive SR methods to control the restoration strength, they are primarily limited to degradation-level adjustments (such as noise and blur), resulting in either smoother or sharper outputs. Several multi-step SD-based SR methods [30, 57, 59, 61] employ guidance-based strategies [14] at each sampling step to achieve varying levels of semantic richness, but these methods often struggle with precise control and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present Pixel-level and Semantic-level Adjustable Super-Resolution, namely PiSA-SR, which separates pixel-level and semantic-level enhancement into two distinct Low-Rank Adapter (LoRA) [16] weight spaces with the pre-trained SD model, providing an effective SR model for various user preferences. We begin by formulating the SD-based SR problem to learn the residual between LQ and HQ latent features [12, 64, 69]. With this formulation, we can not only accelerate the convergence of model training but also introduce scale factors on the model outputs, enabling flexible adjustment on the SR results during the inference stage without re-training. Then, we fine-tune the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "pre-trained SD model by introducing two separate LoRA modules dedicated to pixel-level regression and semantic-level enhancement. In specific, the \\(\\ell_2\\) loss is employed for the LoRA module of pixel-level regression, while the LPIPS [67] and classifier score distillation (CSD) [62] losses are employed for another LoRA module, leveraging the semantic priors encoded in the pre-trained VGG classification model [39] and SD image generation model [37]. A decoupled training approach is proposed to train pixel- and semantic-level LoRAs, effectively improving semantic information while maintaining pixel-level fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Our experiments demonstrate that PiSA-SR not only achieves superior SR performance to existing SD-based models but also offers an effective way for users to adjust the SR styles based on their tastes. An example is shown in Fig. 1. The horizontal and vertical axes indicate the semantic-level and pixel-level guidance scales, respectively. Increasing the pixel-level scale removes noise and compression artifacts, but a too-strong scale over-smoothes image details. Conversely, increasing the semantic-level scale enriches the image details. Nonetheless, a too-high semantic scale over-enhances visual artifacts. PiSA-SR provides the flexibility to adjust both pixel-level and semantic-level guidance scales based on user preferences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.655, + 0.485 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.905, + 0.854 + ], + "angle": 0, + "content": "Image super-resolution. Early deep learning-based SR methods [5, 10, 22, 27] aim to improve image fidelity metrics such as PSNR and SSIM [52]. By using bicubic down-sampling to generate LQ-HQ image training pairs, various approaches have been developed to enhance the SR performance, such as dense [44], residual [69], recursive connections [22], non-local strategies [47], and attention mechanisms [8, 27, 68]. Though these advancements have significantly improved the SR performance, the real-world LQ images with complex degradations are much more challenging to enhance. While researchers have proposed to collect real-world LQ-HQ paired datasets [4, 55] to train SR models, simulating realistic LQ-HQ pairs is a more economic way [50, 65]. BSRGAN [65] randomly shuffles some basic degradation operators to synthesize LQ-HQ pairs, while RealESRGAN [50] implements high-order degradation modeling. Compared with pixel-level losses, the perceptual loss [17] and GAN loss [24, 48] are more effective in improving image perceptual quality. However, the adversarial training also leads to unnatural visual artifacts in the SR results. Many of the later works aim to reduce GAN-generated artifacts from the perspectives of frequency division [25], network optimization [70], the weighting of different losses [43, 49], and image local statistics [29]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Diffusion model based super-resolution. Recent studies have explored the use of diffusion models (DMs) [15, 37] for SR. Early approaches adjust pre-trained DMs'" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2334" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.501 + ], + "angle": 0, + "content": "reverse transitions [20] using gradient descent, offering training-free solutions by assuming predefined image degradations. ResShift [63] trains a DM from scratch on paired LQ-HQ data. Recently, SD models pre-trained on T2I tasks have been widely used for SR tasks due to their strong image priors [2, 6, 30, 32, 36, 46, 51, 56, 57, 59]. StableSR [46] introduces a trainable encoder and takes LQ image as a condition in SD. DiffBIR [30] first applies a restoration module to mitigate degradations, then utilizes SD to enhance details. PASD [59] employs an encoder for degradation removal and introduces a pixel-aware cross-attention module to incorporate low-level and high-level image features into the SD process. SeeSR [57] enhances semantic robustness through degradation-aware tag-style prompts to guide diffusion. Considering that the multi-step diffusion process will increase the computational cost and the risk of synthesizing unfaithful contents, researchers started to develop fewer-step DM-based SR models. SinSR [51] applies consistency-preserving distillation to shorten the diffusion process of ResShift. OSEDiff [56] employs the LQ image as direct input, eliminating the random noise sampling. It adopts the VSD loss [54, 60] to distill the generative capacity from the multi-step diffusion process, offering a one-step DM-based SR solution. However, all the above-mentioned methods face a contradiction between pixel-level and semantic-level enhancement and most of them lack the flexibility to adjust the SR style for diverse user preferences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.515, + 0.226, + 0.532 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.705 + ], + "angle": 0, + "content": "This section first formulates the SD-based SR as a residual learning model, then introduces the dual-LoRA approach to disentangle the learning objectives of pixel-level regression and semantic-level enhancement. Finally, pixel and semantic guidances are proposed for flexible SR results generation. In our following development, we denote by \\( x_{L} \\) and \\( x_{H} \\) the LQ and HQ images, respectively, and by \\( z_{L} \\) and \\( z_{H} \\) their corresponding latent codes. Let \\( E \\) and \\( D \\) be the encoder and decoder of a well-trained variational auto-encoder (VAE), we can nearly have \\( z_{L} = E(x_{L}) \\), \\( z_{H} = E(x_{H}) \\), \\( x_{L} = D(z_{L}) \\) and \\( x_{H} = D(z_{H}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.715, + 0.277, + 0.729 + ], + "angle": 0, + "content": "3.1. Model Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "DMs such SD [37] apply a \\(T\\)-step forward process to progressively transform an initial latent code \\(z_0\\) into Gaussian noise \\(z_T\\) over \\(T\\) steps. In each timestep \\(t\\), the noisy latent code \\(z_t\\) is obtained by directly adding noise \\(\\epsilon \\sim \\mathcal{N}(0, I)\\) to \\(z_0\\) by \\(z_t = \\sqrt{\\bar{\\alpha}_t} \\cdot z_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\cdot \\epsilon\\), where \\(\\bar{\\alpha}_t\\) is the cumulative parameter controlling the noise level at timestep \\(t\\). During the reverse process, multi-step SD-based SR models [2, 30, 36, 46, 57, 59] usually perform a \\(T\\)-step denoising process to gradually transform Gaussian noise \\(z_T\\) into the HQ latent code \\(z_H\\), conditioned on the LQ image \\(x_L\\) with a ControlNet [66] (see Fig. 2 (a)). At stage \\(t\\), the latent" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.088, + 0.902, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.183, + 0.797, + 0.194 + ], + "angle": 0, + "content": "(a) Multi-step DM-based SR methods" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.199, + 0.9, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.241, + 0.736, + 0.251 + ], + "angle": 0, + "content": "(b) OSEDiff" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.258, + 0.9, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.313, + 0.736, + 0.323 + ], + "angle": 0, + "content": "(c) PiSA-SR" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.324, + 0.907, + 0.421 + ], + "angle": 0, + "content": "Figure 2. Comparison of the pipeline of different DM-based SR methods. (a) Multi-step methods [2, 30, 36, 46, 57, 59] perform \\( T \\) denoising steps starting from Gaussian noise \\( z_{T} \\), conditioned on the LQ image \\( x_{L} \\). (b) OSEDiff [56] starts from LQ latent representation \\( z_{L} \\) with only one-step diffusion. (c) Our proposed PiSA-SR formulates the SD-based SR as learning the residual between the LQ latent \\( z_{L} \\) and HQ latent \\( z_{H} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.434, + 0.907, + 0.466 + ], + "angle": 0, + "content": "\\(\\hat{z}_0\\) can be computed from the noise estimated by SD UNet \\(\\epsilon_{\\theta}(z_t,x_L,t)\\) and the current latent \\(z_{t}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.475, + 0.906, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {0} = f (z _ {t}, \\epsilon_ {\\theta}) = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (z _ {t} , x _ {L} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.906, + 0.564 + ], + "angle": 0, + "content": "Then the latent \\( z_{t-1} \\) at timestep \\( t - 1 \\) can be sampled from \\( p(z_{t-1} | z_t, \\hat{z}_0) = N(\\mu_t(z_t, \\hat{z}_0), \\sigma_t^2) \\), where \\( \\mu_t \\) and \\( \\sigma_t^2 \\) denote the mean and variance of \\( z_{t-1} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.906, + 0.683 + ], + "angle": 0, + "content": "However, these multi-step DM-based SR methods are computationally expensive and often yield unstable results due to the random noise sampled in the diffusion process [42, 57]. To address this issue, OSEDiff [56] starts from the LQ latent code \\( z_{L} \\) and completes the SR process with only one diffusion step (see Fig. 2 (b)), improving both the efficiency and stability. The transformation between \\( z_{L} \\) and \\( z_{H} \\) can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.693, + 0.906, + 0.726 + ], + "angle": 0, + "content": "\\[\nz _ {H} = f \\left(z _ {L}, \\epsilon_ {\\theta}\\right) = \\frac {z _ {L} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(z _ {L}\\right)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.765 + ], + "angle": 0, + "content": "Note that we omit the timestep \\( t \\) in the above formula since \\( t \\) can be viewed as a constant in one-step diffusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Learning the residual between the LQ and GT features has been successfully applied in deep learning-based restoration methods [64, 69]. However, DMs typically perform multi-step iterations, making applying residual learning difficult. The recently developed OSEDiff [56] employs a single diffusion step to achieve end-to-end SR training, making it feasible to adopt the residual learning strategy. Here, we formulate the SR problem as learning the residual between \\( z_{L} \\) and \\( z_{H} \\), as illustrated in Fig. 2 (c). We use" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2335" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "subtraction in the global residual connection because DMs are trained to denoise in the inverse diffusion process. Such a residual learning formulation helps the model focus on learning the desired high-frequency information from the HQ latent features, avoiding extracting less relevant information from the LQ latent. It can also accelerate the convergence of the model training process [12]. Furthermore, we introduce a scaling factor \\(\\lambda\\) to adjust the residual \\(\\epsilon_{\\theta}\\) added to the LQ latent, which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.238, + 0.483, + 0.253 + ], + "angle": 0, + "content": "\\[\nz _ {H} = z _ {L} - \\lambda \\epsilon_ {\\theta} (z _ {L}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.484, + 0.323 + ], + "angle": 0, + "content": "The \\(\\lambda\\) is fixed as 1 during training. Users can adjust the output in the inference stage by using a smaller \\(\\lambda\\) (to preserve more original contents) or a higher \\(\\lambda\\) (to enhance details more aggressively) based on their preferences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.33, + 0.291, + 0.347 + ], + "angle": 0, + "content": "3.2. Dual-LoRA Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.485, + 0.64 + ], + "angle": 0, + "content": "Previous SD-based SR methods mostly entangle the pixel-level and semantic-level enhancement in the training process [46, 56], making the balance of content fidelity and perpetual quality difficult. Some works sequentially perform pixel-level and semantic-level enhancement using two separate networks [30]. However, errors in the earlier stage may propagate to the next stage, limiting overall performance. Furthermore, using two separate networks increases the computational burden and memory requirements, making the approach less efficient. It has been successfully applied in customized T2I tasks using distinct LoRA modules for various generation styles [18, 38, 41]. Inspired by this, we propose a decoupled training approach by utilizing two LoRA modules under the pre-trained SD model, targeting pixel-level and semantic-level enhancement for the SR task respectively. Our approach, namely Dual-LoRA, introduces only a small number of additional parameters during training. These LoRA parameters can also be merged into the pre-trained SD model during inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.484, + 0.827 + ], + "angle": 0, + "content": "Fig. 3 (a) illustrates the training process of our approach. We freeze the parameters of the pre-trained VAE and introduce two trainable LoRA modules to the UNet of SD. Since the LQ image is corrupted by degradations such as noise, blur, and downsampling, we optimize the pixel-level LoRA first to reduce the effect of corruptions, followed by the optimization of semantic-level LoRA. We train the pixel-level LoRA, denoted by \\(\\Delta \\theta_{pix}\\), by the pixel-level loss function. Together with the pre-trained SD parameters, the full parameter set can be represented as \\(\\theta_{pix} = \\{\\theta_{sd}, \\Delta \\theta_{pix}\\}\\). The HQ latent can then be estimated by \\(z_H^{pix} = z_L - \\epsilon_{\\theta_{pix}}(z_L)\\) and decoded by the VAE decoder with \\(x_H^{pix} = D(z_H^{pix})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.827, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For semantic-level enhancement, we train another LoRA, denoted by \\(\\Delta \\theta_{sem}\\), by the semantic-level loss function. To disentangle the pixel-level and semantic-level objectives, we fix the already trained pixel-level LoRA \\(\\Delta \\theta_{pix}\\) and combine it with the semantic-level LoRA \\(\\Delta \\theta_{sem}\\) to be" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.245 + ], + "angle": 0, + "content": "trained, forming a PiSA-LoRA group (see Fig. 3 (a)). Note that only the semantic-level LoRA module within the PiSA-LoRA group is updated in this stage. This ensures that the optimization process focuses on the semantic details without much interference from the pixel-level degradations. The full parameter set during this PiSA-LoRA training stage can be expressed as \\(\\theta_{PiSA} = \\{\\theta_{sd},\\Delta \\theta_{pix},\\Delta \\theta_{sem}\\}\\). The HQ latent after this stage is given by \\(z_H^{sem} = z_L - \\epsilon_{\\theta_{PiSA}}(z_L)\\), which is then decoded by the VAE decoder with \\(x_H^{sem} = D(z_H^{sem})\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.254, + 0.833, + 0.269 + ], + "angle": 0, + "content": "3.3. Pixel-level and Semantic-level Losses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.907, + 0.46 + ], + "angle": 0, + "content": "We train the pixel-level LoRA using the \\(\\ell_2\\) loss. As shown by \\(D(\\epsilon_{\\theta_{pix}}(z_L))\\) in Fig. 4, the \\(\\ell_2\\) loss can effectively remove degradation and enhance edges. However, it is insufficient to generate semantic-level details [24, 29, 48], resulting in smooth SR outputs. The LPIPS loss [67] can be used to activate semantic details by aligning the high-level features with a pre-trained classification VGG network [39], which is however trained on a limited set of image categories. The GAN loss [24] captures semantic information through adversarial training, encouraging the generator to produce more realistic images. However, GAN loss can be unstable in training and generate much artifacts [29]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.552 + ], + "angle": 0, + "content": "Recently developed SD models [37] have shown superior performance in generating open-category images with enhanced semantic details. As a conditional generation model, SD can be used to form up an implicit classifier [14] for the posterior distribution modeling, and its gradient w.r.t. the synthesized image \\( x_{t} \\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.564, + 0.905, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log p _ {\\theta} (c \\mid x _ {t}) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t} \\mid c) - \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ = - \\left(\\epsilon_ {\\theta} \\left(x _ {t}, c, t\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right) / \\sigma_ {t}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.734 + ], + "angle": 0, + "content": "where \\( c \\) is the text prompt, \\( t \\) is the current timestep, and \\( \\sigma_{t} = \\sqrt{1 - \\bar{\\alpha}_{t}} \\). The SD model, parameterized by \\( \\theta \\), model score functions of \\( \\nabla_{x_t}\\log p_\\theta (x_t\\mid c) \\) and \\( \\nabla_{x_t}\\log p_\\theta (x_t) \\) using noise predictions [40]. By taking the expectation of the gradient in Eq. (4) over all possible \\( t \\), we have the CSD [14] loss, which is originally used in 3D generation tasks to optimize the posterior probability of rendered images aligning their semantic content with text prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Motivated by the functionality of CSD loss in DMs for generation, we investigate the usage of CSD loss for semantic-level enhancement in the SR task. Following [56, 60], and for the clarity of the following context, we formulate the gradient of CSD in the latent space instead of the noise domain. Specifically, the gradient of CSD loss can be formulated in Eq. (5) below (the detailed derivation can be found in the supplementary materials):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.862, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\nabla \\ell_ {C S D} ^ {\\lambda_ {c f g}} = \\mathbb {E} _ {t, \\epsilon , z _ {t}, c} \\left[ w _ {t} \\left(f (z _ {t}, \\epsilon_ {r e a l}) - f (z _ {t}, \\epsilon_ {r e a l} ^ {\\lambda_ {c f g}})\\right) \\frac {\\partial z _ {H} ^ {s e m}}{\\partial \\theta_ {P i S A}} \\right], \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2336" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.889, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.294, + 0.613, + 0.309 + ], + "angle": 0, + "content": "(a) Training process of the PiSA-SR" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.313, + 0.889, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.453, + 0.603, + 0.466 + ], + "angle": 0, + "content": "(b) Inference process of PiSA-SR" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.467, + 0.908, + 0.511 + ], + "angle": 0, + "content": "Figure 3. The (a) training and (b) inference procedures of PiSA-SR. During the training process, two LoRA modules are respectively optimized for pixel-level and semantic-level enhancement. During the inference stage, users can use the default setting to reconstruct the HQ image in one-step diffusion or adjust \\(\\lambda_{pix}\\) and \\(\\lambda_{sem}\\) to control the strengths of pixel-level and semantic-level enhancement." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.532, + 0.481, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.731 + ], + "angle": 0, + "content": "Figure 4. The model outputs with pixel-wise and semantic-level losses for a given LQ image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.86 + ], + "angle": 0, + "content": "where the gradient expectation is calculated considering all diffusion timesteps \\(t = 1,\\dots,T\\) , the noise is sampled from \\(\\epsilon \\sim \\mathcal{N}(0,I),z_{t}\\) is obtained by \\(z_{t} = \\sqrt{\\bar{\\alpha}_{t}}\\cdot z_{H}^{sem} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\cdot \\epsilon\\) \\(c\\) is the text prompt extracted from \\(z_H^{sem}\\) \\(f(\\cdot)\\) is the function as Eq. (1), \\(w_{t}\\) is the timestep-dependent scalar weight, and \\(\\epsilon_{real}^{\\lambda_{cfg}}\\) denotes the pre-trained SD output with the CFG term \\(\\epsilon_{real}^{cls}\\) same as Eq. (4) [14, 62]:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.864, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {r e a l} ^ {\\lambda_ {c f g}} = \\epsilon_ {r e a l} (z _ {t}, t) + \\underbrace {\\lambda_ {c f g} (\\epsilon_ {r e a l} (z _ {t} , t , c) - \\epsilon_ {r e a l} (z _ {t} , t))}. \\qquad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Note that the VSD loss [54, 60], which also aims to align the distributions of enhanced and natural images within the latent space, has been verified effective for SR by OSEDiff [56]. The gradient of VSD loss can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.616, + 0.906, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\nabla \\ell_ {V S D} ^ {\\lambda_ {c f g}} = \\mathbb {E} _ {t, \\epsilon , z _ {t}, c} \\left[ w _ {t} \\left(f (z _ {t}, \\epsilon_ {f a k e}) - f (z _ {t}, \\epsilon_ {r e a l} ^ {\\lambda_ {c f g}})\\right) \\frac {\\partial z _ {H} ^ {s e m}}{\\partial \\theta_ {P i S A}} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\text{fake}}\\) denotes the output of fine-tuned SD aligned with the distribution of synthetic images. The VSD loss with \\(\\lambda_{cfg}\\) can be divided into two components: the VSD loss with \\(\\lambda_{cfg} = 0.0\\), i.e., \\(\\ell_{VSD}^{\\lambda_{cfg} = 0.0}\\), and the CSD loss. Coupling with the LPIPS loss, we visualize the semantic-level LoRA optimization results of the two components of VSD in Fig. 4. We see that the CSD loss with the normal guidance scale of \\(\\lambda_{cfg} = 7.5\\) contributes more significantly to semantic enhancement. In contrast, the VSD loss with \\(\\lambda_{cfg} = 0.0\\) weakens the semantic details. More importantly, the component \\(\\ell_{VSD}^{\\lambda_{cfg} = 0.0}\\) requires bi-level optimization, leading to intensive memory usage and unstable training [31]. In comparison, the CSD loss is free of bi-level optimization, reducing significantly memory usage and improving training stability. Therefore, we integrate LPIPS and CSD losses for the semantic-level LoRA optimization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2337" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.388, + 0.105 + ], + "angle": 0, + "content": "3.4. The Inference Process of PiSA-SR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.484, + 0.262 + ], + "angle": 0, + "content": "Fig. 3(b) shows the inference process of PiSA-SR. In the default inference setting (represented by the green dash-dot line), only the PiSA-LoRA, which merges the pixel-level and semantic-level LoRA modules, is used to process the input together with the pre-trained SD model, achieving state-of-the-art SR performance in just one-step diffusion (see our experiments in Section 4.3). To achieve flexible SR with diverse user preferences, we extend the strategy in Eq. (3) by introducing a pair of pixel and semantic guidance factors, denoted by \\(\\lambda_{pix}\\) and \\(\\lambda_{sem}\\), to control the SR results:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.273, + 0.482, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\theta} \\left(z _ {L}\\right) = \\lambda_ {p i x} \\epsilon_ {\\theta_ {p i x}} \\left(z _ {L}\\right) + \\lambda_ {s e m} \\left(\\epsilon_ {\\theta_ {P s S A}} \\left(z _ {L}\\right) - \\epsilon_ {\\theta_ {p i x}} \\left(z _ {L}\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.456 + ], + "angle": 0, + "content": "In Eq. (8), \\(\\epsilon_{\\theta_{pix}}(z_L)\\) is the output with only pixel-level LoRA, while \\(\\epsilon_{\\theta_{PiSA}}(z_L)\\) is the output with both pixel and semantic level enhancement. Inspired by CFG [14], the difference between them, i.e., \\(\\epsilon_{\\theta_{PiSA}}(z_L) - \\epsilon_{\\theta_{pix}}(z_L)\\), can well separate pixel-level degradations from semantic details (see the red box in Fig. 4). Therefore, as illustrated in Fig. 3(b) (represented by the black dashed line), by adjusting \\(\\lambda_{pix}\\) and \\(\\lambda_{sem}\\), we can control the contributions of pixel-level LoRA and semantic-level LoRA, and consequently deliver the SR results with different styles, as shown in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.224, + 0.485 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.489, + 0.298, + 0.505 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.689 + ], + "angle": 0, + "content": "Training settings. We train PiSA-SR upon SD 2.1-base for the \\(\\times 4\\) SR task. The two trainable LoRA modules are applied to the weights of all convolution and MLP layers. Both modules are initialized using a Gaussian distribution with a rank of 4. Following SeeSR [57] and OSEDiff [56], we employ LSDIR [26] and the first 10K images from FFHQ [19] as the training data. We generate paired training data using the degradation pipeline from RealESRGAN [50]. The batch size is 16 and the training patch size is \\(512 \\times 512\\). The Adam optimizer [23] is used with a learning rate 5e-5. The pixel-level and semantic-level LoRA modules undergo 4K and 8.5K training iterations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Compared methods. We compare PiSA-SR with leading multi-step DM-based methods StableSR [46], ResShift [63], DiffBIR [30], PASD [59], and SeeSR [57], one-step DM-based methods SinSR [51] and OSEDiff [56], and GAN-based SR methods RealESRGAN [50], BSRGAN [65] and LDL [29]. All comparative results are obtained using officially released codes or models. Due to space constraints, comparisons with GAN-based methods are presented in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Test datasets and evaluation metrics. Following previous arts [46, 56, 57], we evaluate the competing methods using synthetic and real-world test data. The synthetic dataset comprises 3000 images cropped from DIV2K [1] with size \\(512 \\times 512\\), degraded using RealESRGAN" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.116 + ], + "angle": 0, + "content": "Table 1. Results of PiSA-SR with different pixel-semantic guidance scales on the RealSR test dataset." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.12, + 0.905, + 0.273 + ], + "angle": 0, + "content": "
λpixλsemPSNR↑LIPIS↓CLIPIQA↑MUSIQ↑
0.01.025.960.34260.412946.45
0.21.026.480.30420.486854.05
0.51.026.750.26460.570563.82
0.81.026.180.26120.629268.95
1.01.025.500.26720.670270.15
1.21.024.760.27230.674670.33
1.51.023.740.27690.630569.23
1.00.026.920.30180.322749.62
1.00.226.950.27840.359153.64
1.00.526.770.24760.432258.76
1.00.826.200.24650.580666.33
1.01.025.500.26720.670270.15
1.01.224.590.30000.701571.60
1.01.523.080.35410.683571.76
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.291, + 0.907, + 0.488 + ], + "angle": 0, + "content": "degradation pipeline [50]. The real-world data are center-cropped from RealSR [4] and DrealSR [55] datasets with size \\(128 \\times 128\\) for LQ images and \\(512 \\times 512\\) for HQ images. PSNR and SSIM [52], computed on the Y channel in the YCbCr space, are used to measure the fidelity of SR results; LPIPS [67] and DISTS [9], computed in the RGB space, are used to assess the perceptual quality of SR results; FID [13] evaluates the distance of distributions between GT and restored images; NIQE [33], CLIPIQA [45], MUSIQ [21] and MANIQA [58] evaluate the image quality without reference images. Additionally, a user study is conducted to validate the effectiveness of the proposed adjustable SR in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.787, + 0.518 + ], + "angle": 0, + "content": "4.2. Experiments on Adjustable SR" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.524, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We conduct experiments by fixing one guidance scale (\\(\\lambda_{pix}\\) or \\(\\lambda_{sem}\\)) at 1 and varying the other to observe how the reconstructed images change. PSNR, LPIPS, CLIPIQA, and MUSIQ are used to evaluate performance. PSNR measures pixel-level fidelity; LPIPS evaluates the image perceptual quality with GT as reference; CLIPIQA and MUSIQ measure the image quality without using a reference. The results are shown in Table 1. First, we observe that increasing the pixel-level scale \\(\\lambda_{pix}\\) leads to a continuous improvement in the no-reference metrics CLIPIQA and MUSIQ. This is because increasing \\(\\lambda_{pix}\\) can remove image degradation and enhance edge, which is favorable to these two metrics. However, the reference-based metrics PSNR and LPIPS exhibit a rise-and-fall pattern. PSNR peaks at \\(\\lambda_{pix} = 0.5\\), indicating the best pixel-level fidelity. LPIPS reaches its best value at \\(\\lambda_{pix} = 0.8\\) (0.2612), suggesting that at this point the restored image is perceptually most similar to the GT with rich details. Further increasing \\(\\lambda_{pix}\\) reduces both PSNR and LPIPS scores. Second, increasing the semantic-level scale \\(\\lambda_{sem}\\) also results in a continuous improvement in CLIPIQA and MUSIQ, but with a higher upper bound than pixel-level adjustments. This is because increasing \\(\\lambda_{sem}\\) can synthesize more semantic-level details (see Fig. 1). However, PSNR decreases as \\(\\lambda_{sem}\\) increases, while LPIPS first improves and peaks at \\(\\lambda_{sem} = 0.8\\) (0.2465), and then" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2338" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison among the state-of-the-art DM-based SR methods on synthetic and real-world test datasets. 'S' denotes the number of diffusion steps. The best and the second-best results are highlighted in red and blue, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.122, + 0.904, + 0.403 + ], + "angle": 0, + "content": "
DatasetsMethodsPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓CLIPQA↑MUSIQ↑MANIQA↑
DIV2KResShift-S1524.690.61750.33740.221536.016.820.608960.920.5450
StableSR-S20023.310.57280.31290.213824.674.760.668265.630.6188
DiffBIR-S5023.670.56530.35410.212930.934.710.665265.660.6204
PASD-S2023.140.54890.36070.221929.324.400.671168.830.6484
SeeSR-S5023.710.60450.32070.196725.834.820.685768.490.6239
SinSR-S124.430.60120.32620.206635.456.020.649962.800.5395
OSEDiff-S123.720.61080.29410.197626.324.710.668367.970.6148
PiSA-SR-S123.870.60580.28230.193425.074.550.692769.680.6400
RealSRResShift-S1526.310.74110.34890.2498142.817.270.545058.100.5305
StableSR-S20024.690.70520.30910.2167127.205.760.619565.420.6211
DiffBIR-S5024.880.66730.35670.2290124.565.630.641264.660.6231
PASD-S2025.220.68090.33920.2259123.085.180.650268.740.6461
SeeSR-S5025.330.72730.29850.2213125.665.380.659469.370.6439
SinSR-S126.300.73540.32120.2346137.056.310.620460.410.5389
OSEDiff-S125.150.73410.29210.2128123.505.650.669369.090.6339
PiSA-SR-S125.500.74170.26720.2044124.095.500.670270.150.6560
DrealSRResShift-S1528.450.76320.40730.2700175.928.280.525949.860.4573
StableSR-S20028.040.74600.33540.2287147.036.510.617158.500.5602
DiffBIR-S5026.840.66600.44460.2706167.386.020.629260.680.5902
PASD-S2027.480.70510.38540.2535157.365.570.671464.550.6130
SeeSR-S5028.260.76980.31970.2306149.866.520.667264.840.6026
SinSR-S128.410.74950.37410.2488177.057.020.636755.340.4898
OSEDiff-S127.920.78350.29680.2165135.296.490.696364.650.5899
PiSA-SR-S128.310.78040.29600.2169130.616.200.697066.110.6156
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.482, + 0.492 + ], + "angle": 0, + "content": "declines. This is because excessive semantic details may induce changes in image content and reduce pixel-level fidelity. Meanwhile, the over-enhanced images may exhibit many details different from the GT, making the reference-based LPIPS index deteriorate." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.492, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Fig. 1 illustrates the visual comparisons of different pixel-semantic scales. More comparisons can be found in the supplementary materials. From bottom to top, as the pixel-level scale \\(\\lambda_{pix}\\) increases, the model progressively removes degradation from the input LQ image, resulting in clearer and sharper visual output. However, a too-high \\(\\lambda_{pix}\\) can lead to loss of details, and diminish the effectiveness of semantic improvements, as seen in the top-right corner of Fig. 1. From left to right, increasing the semantic-level scale \\(\\lambda_{sem}\\) enhances the semantic richness of the image, making the details of the wrinkles, beard, and hair of Albert Einstein more pronounced. However, an excessive semantic level such as \\(\\lambda_{sem} = 1.5\\) introduces unnatural artifacts, making the SR result less realistic. With the flexibility of PiSASR, users can customize their preferences, either preserving more fidelity or emphasizing richer semantic enhancements, depending on their specific use case." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.76, + 0.402, + 0.776 + ], + "angle": 0, + "content": "4.3. Comparisons with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.841 + ], + "angle": 0, + "content": "As shown in Eq. (8), letting \\(\\lambda_{pix} = 1\\) and \\(\\lambda_{sem} = 1\\), we have the default version of PiSA-SR, which can be performed in just one-step diffusion. Here, we compare this default version of PiSA-SR with the competing methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Quantitative comparisons. Table 2 compares the performance of our default PiSA-SR model with other DM-based Real-SR methods. The following findings can be made. ResShift [63] and its distilled version SinSR [51] show lim" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.688 + ], + "angle": 0, + "content": "ited performance in no-reference metrics, implying poor perceptual quality. PASD [59] and SeeSR [57] leverage the pre-trained SD as their base model, and extract additional high-level information [71] to improve the perceptual quality of images. Their no-reference metric scores, such as MUSIQ and CLIPIQA, are highly improved. However, the multi-step inference makes them less efficient. In addition, their relatively low LPIPS and DISTS scores suggest their less faithful reconstruction to the GT. OSEDiff [56] distills the generation capacity of multi-step DM into just one-step diffusion by the VSD loss [54, 60]. While effective and efficient, its no-reference metrics are limited. Our proposed PiSA-SR also exhibits high efficiency by requiring only one inference step, while showing impressive pixel-level fidelity and semantic-level perceptual quality. It outperforms other methods not only in reference-based metrics such as LPIPS and DISTS but also in no-reference metrics such as CLIP-IQA, MUSIQ, and MANIQA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative comparisons. We present visual comparisons in Fig. 5. ResShift and SinSR fail to generate richer textures and details because of their limited generation capacity. StableSR and DiffBIR can generate finer details (e.g., more realistic penguin feathers in the second example), due to the leveraged strong SD prior. PASD and SeeSR incorporate high-level information guidance, resulting in more comprehensive details. However, the inherent randomness in PASD and SeeSR may lead to over-enhanced (e.g., unnatural penguin feathers generated by SeeSR) or over-smoothed details (e.g., the blurred texture of the wooden stool generated by SeeSR). OSEDiff can generate more consistent results in both examples. However, the restored details can be limited. In contrast, the proposed PiSA-SR delivers high-quality, re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2339" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.089, + 0.319, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.089, + 0.506, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.089, + 0.846, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.247, + 0.318, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.32, + 0.247, + 0.507, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.247, + 0.846, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.408, + 0.793, + 0.421 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons of different DM-based SR methods. Please zoom in for a better view." + }, + { + "type": "table_caption", + "bbox": [ + 0.245, + 0.424, + 0.751, + 0.437 + ], + "angle": 0, + "content": "Table 3. The inference time and the number of parameters of DM-based SR methods." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.44, + 0.907, + 0.491 + ], + "angle": 0, + "content": "
StableSRResShiftDiffBIRPASDSeeSRSinSROSEDiffPiSA-SR-def.PiSA-SR-adj.
Inference Steps200155020501112
Inference time(s)/Image10.030.762.722.804.300.130.120.090.13
#Params(B)1.560.181.682.312.510.181.771.301.30
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.581 + ], + "angle": 0, + "content": "alistic SR results. With the dual-LoRA design, the restored structures align well with the input LQ image, providing more accurate fidelity information and generating more natural and richer details. More visualization comparisons can be found in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Complexity comparisons. We compare the number of parameters and inference time of competing DM-based SR models in Table 3, where PiSA-SR-def. and PiSA-SR-adj. represent its default setting and adjustable setting, respectively. Inference time is measured on the \\(\\times 4\\) SR task with \\(128\\times 128\\) LQ images using a single NVIDIA A100 80G GPU. ResShift has fewer parameters because it is trained from scratch without using the pre-trained SD model. SinSR inherits the parameters from its parental ResShift model. Among the SD-based SR methods, PiSA-SR-def. has the fewest parameters and achieves the fastest inference time. Unlike OSEDiff, PiSA-SR-def. does not require a semantic extractor (e.g., RAM [71]) during inference. Instead, it distills the semantic information into the semantic-level LoRA with CSD. PiSA-SR-adj., which requires two diffusion steps to compute outputs from pixellevel and semantic-level LoRA modules, incurs a slightly longer inference time over the default setting. However, this small increase in time offers users the advantage of generating SR outputs tailored to their specific needs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation study. We performed a series of ablation stud" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.551 + ], + "angle": 0, + "content": "ies for a better understanding of the dual-LoRA training approach, and the roles of the two LoRA modules. The details can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.563, + 0.764, + 0.578 + ], + "angle": 0, + "content": "5. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.584, + 0.907, + 0.78 + ], + "angle": 0, + "content": "In this work, we presented PiSA-SR, a novel SR framework that disentangles pixel-level and semantic-level objectives by learning two LoRA modules upon the pre-trained SD model. By decoupling the learning process into two distinct spaces, PiSA-SR effectively balanced pixel-wise regression and perceptual quality. The commonly used \\(\\ell_2\\) loss was used to optimize pixel-level LoRA, while the LPIPS and CSD losses were employed to optimize the semantic-level LoRA. PiSA-SR demonstrated impressive performance in both effectiveness and efficiency. Furthermore, it offered high flexibility during inference through adjustable guidance scales, allowing users to customize the SR results without model re-training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "While PiSA-SR provided a flexible solution for adjusting pixel-level fidelity and semantic-level perception, it incurs a slight inference time increment over its default version. In addition, we use a single pixel-level LoRA to handle various degradations in the LQ, which may be insufficient for heavy degradations. In the future, we will explore refining separate LoRA spaces for different image degradations and the possibility of image-adaptive scale factors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2340" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. NTIRE 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Yang Ai, Xiaogiang Zhou, Huaibo Huang, Xiaotian Han, Zhengyu Chen, Quanzeng You, and Hongxia Yang. Dreamclear: High-capacity real-world image restoration with privacy-safe dataset curation. arXiv preprint arXiv:2410.18666, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[3] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.4 + ], + "angle": 0, + "content": "[5] Dong Chao, Loy Chen Change, He Kaiming, and Tang Xiaou. Learning a deep convolutional network for image super-resolution. In ECCV, pages 184-199, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.483, + 0.457 + ], + "angle": 0, + "content": "[6] Bin Chen, Gehui Li, Rongyuan Wu, Xindong Zhang, Jie Chen, Jian Zhang, and Lei Zhang. Adversarial diffusion compression for real-world image super-resolution. arXiv preprint arXiv:2411.13383, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[7] Du Chen, Tianhe Wu, Kede Ma, and Lei Zhang. Toward generalized image quality assessment: Relaxing the perfect reference quality assumption, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.503, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[9] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.631, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[10] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.483, + 0.742 + ], + "angle": 0, + "content": "[11] Jingwen He, Chao Dong, Yihao Liu, and Yu Qiao. Interactive multi-dimension modulation for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9363-9379, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.196, + 0.905, + 0.25 + ], + "angle": 0, + "content": "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.254, + 0.905, + 0.295 + ], + "angle": 0, + "content": "[18] Maxwell Jones, Sheng-Yu Wang, Nupur Kumari, David Bau, and Jun-Yan Zhu. Customizing text-to-image models with a single image pair. arXiv preprint arXiv:2405.01536, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.298, + 0.905, + 0.353 + ], + "angle": 0, + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.356, + 0.905, + 0.41 + ], + "angle": 0, + "content": "[20] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[21] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.473, + 0.905, + 0.527 + ], + "angle": 0, + "content": "[22] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[24] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.674, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[25] Wenbo Li, Kun Zhou, Lu Qi, Liying Lu, and Jiangbo Lu. Best-Buddy GANs for highly detailed image superresolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1412-1420, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[26] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Lsdir: A large scale dataset for image restoration. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1775-1787, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2341" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[28] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[29] Jie Liang, Hui Zeng, and Lei Zhang. Details or Artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.223, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[30] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[31] Zhiyuan Ma, Yuxiang Wei, Yabin Zhang, Xiangyu Zhu, Zhen Lei, and Lei Zhang. Scaledreamer: Scalable text-to-3d synthesis with asynchronous score distillation. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.367, + 0.482, + 0.421 + ], + "angle": 0, + "content": "[32] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Conditional diffusion distillation. arXiv preprint arXiv:2310.01407, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.425, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[33] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.536 + ], + "angle": 0, + "content": "[34] Chong Mou, Yanze Wu, Xintao Wang, Chao Dong, Jian Zhang, and Ying Shan. Metric learning based interactive modulation for real-world super-resolution. In European Conference on Computer Vision, pages 723-740. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[35] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.482, + 0.654 + ], + "angle": 0, + "content": "[36] Yunpeng Qu, Kun Yuan, Kai Zhao, Qizhi Xie, Jinhua Hao, Ming Sun, and Chao Zhou. Xpsr: Cross-modal priors for diffusion-based image super-resolution. arXiv preprint arXiv:2403.05049, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[38] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In European Conference on Computer Vision, pages 422-438. Springer, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[41] Nick Stracke, Stefan Andreas Baumann, Joshua M Susskind, Miguel Angel Bautista, and Björn Ommer. Ctrloralter: Conditional loradapter for efficient 0-shot control & altering of t2i models. arXiv preprint arXiv:2405.07913, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[42] Lingchen Sun, Rongyuan Wu, Zhengqiang Zhang, Hongwei Yong, and Lei Zhang. Improving the stability of diffusion models for content consistent super-resolution. arXiv e-prints, pages arXiv-2401, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.205, + 0.906, + 0.272 + ], + "angle": 0, + "content": "[43] Lingchen Sun, Jie Liang, Shuaizheng Liu, Hongwei Yong, and Lei Zhang. Perception-distortion balanced superresolution: A multi-objective optimization perspective. IEEE Transactions on Image Processing, 33:4444-4458, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.274, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[44] Tong Tong, Gen Li, Xiejie Liu, and Qinquan Gao. Image super-resolution using dense skip connections. In Proceedings of the IEEE international conference on computer vision, pages 4799-4807, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.33, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[45] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.386, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[46] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.442, + 0.906, + 0.522 + ], + "angle": 0, + "content": "[47] Risheng Wang, Tao Lei, Wenzheng Zhou, Qi Wang, Hongying Meng, and Asoke K Nandi. Lightweight non-local network for image super-resolution. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1625-1629. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.525, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[48] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. ESRGAN: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.595, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[49] Xintao Wang, Ke Yu, Chao Dong, Xiaou Tang, and Chen Change Loy. Deep network interpolation for continuous imagery effect transition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1701, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.665, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[50] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.735, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[51] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.818, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[52] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.873, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[53] Zhihao Wang, Jian Chen, and Steven C. H. Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2342" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[54] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[55] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[56] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[57] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[58] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[59] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[60] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[61] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. arXiv preprint arXiv:2401.13627, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[62] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and XIAOJUAN QI. Text-to-3d with classifier score distillation. In The Twelfth International Conference on Learning Representations, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[63] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. arXiv preprint arXiv:2307.12348, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[64] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[65] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision, pages 4791-4800, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[68] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. arXiv preprint arXiv:2203.06697, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[69] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[70] Yuehan Zhang, Bo Ji, Jia Hao, and Angela Yao. Perception-distortion balanced ADMM optimization for single-image super-resolution. In European Conference on Computer Vision, pages 108-125. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[71] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2343" + } + ] +] \ No newline at end of file diff --git a/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/6dc2236b-e40d-40ee-98e0-34b3414c88ae_origin.pdf b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/6dc2236b-e40d-40ee-98e0-34b3414c88ae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..105644aea68050d37a65c7f864d498b97a0ba755 --- /dev/null +++ b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/6dc2236b-e40d-40ee-98e0-34b3414c88ae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7f8a5b8fb172b182d39b30617c01c826ead1c2719ade37c605d94fd3b12b614 +size 8955911 diff --git a/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/full.md b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/full.md new file mode 100644 index 0000000000000000000000000000000000000000..893360163788ac6a9e0cbd6566c479c908414cf9 --- /dev/null +++ b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/full.md @@ -0,0 +1,297 @@ +# Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach + +Lingchen Sun $^{1,2}$ , Rongyuan Wu $^{1,2}$ , Zhiyuan Ma $^{1}$ , Shuaizheng Liu $^{1,2}$ , Qiaosi Yi $^{1,2}$ , Lei Zhang $^{1,2,*}$ $^{1}$ The Hong Kong Polytechnic University + $^{2}$ OPPO Research Institute + +{ling-chen.sun, rong-yuan.wu, zm2354.ma} @connect.polyu.hk + +{shuaizhengliu21,qiaosiyijoyies} $@$ gmail.com cslzhang@comp.polyu.edu.hk + +# Abstract + +Diffusion prior-based methods have shown impressive results in real-world image super-resolution (SR). However, most existing methods entangle pixel-level and semantic-level SR objectives in the training process, struggling to balance pixel-wise fidelity and perceptual quality. Meanwhile, users have varying preferences on SR results, thus it is demanded to develop an adjustable SR model that can be tailored to different fidelity-perception preferences during inference without re-training. We present Pixel-level and Semantic-level Adjustable SR (PiSA-SR), which learns two LoRA modules upon the pre-trained stable-diffusion (SD) model to achieve improved and adjustable SR results. We first formulate the SD-based SR problem as learning the residual between the low-quality input and the high-quality output, then show that the learning objective can be decoupled into two distinct LoRA weight spaces: one is characterized by the $\ell_2$ -loss for pixel-level regression, and another is characterized by the LPIPS and classifier score distillation losses to extract semantic information from pre-trained classification and SD models. In its default setting, PiSA-SR can be performed in a single diffusion step, achieving leading real-world SR results in both quality and efficiency. By introducing two adjustable guidance scales on the two LoRA modules to control the strengths of pixel-wise fidelity and semantic-level details during inference, PiSA-SR can offer flexible SR results according to user preference without re-training. The source code of our method can be found at https://github.com/csslc/PiSA-SR. + +# 1. Introduction + +Single image super-resolution (SR) [53] aims to reconstruct a high-quality (HQ) image from its low-quality (LQ) counterpart suffering from various degradations, e.g., noise, blur, + +![](images/3f1dbd479a492cd7e51b28cdbaef7cbf576e830eb36022417feeb227c97368cb.jpg) +Figure 1. Visual illustration of our pixel- and semantic-level adjustable method for real-world SR. By increasing the pixel-level guidance scale $\lambda_{pix}$ , the image degradations such as noise and compression artifacts can be gradually removed; however, a too-strong $\lambda_{pix}$ will make the SR image over-smoothed. By increasing the semantic-level guidance scale $\lambda_{sem}$ , the SR images will have more semantic details; nonetheless, a too-high $\lambda_{sem}$ will generate visual artifacts. Please zoom in for a better view. + +downsampling, etc. SR is a challenging ill-posed problem due to the inherent ambiguity in reconstructing details [3, 29], and various deep neural networks (DNNs) [5, 8, 10, 22, 27, 68] have been developed to address it. The pixel-level regression losses, such as $\ell_1$ and $\ell_2$ losses, are essential to keep the pixel-level fidelity of SR outputs, yet they tend to produce over-smoothed details [3, 29]. The SSIM [52] and perceptual losses can alleviate this problem. In specific, SSIM measures the local structural similarity between the SR and ground-truth (GT) images, and the perceptual loss [17] extracts semantic features with a pre-trained classification model [39] to improve the per + +ceptual quality of SR results. The adversarial loss associated with the generative adversarial network (GAN) [24, 28, 29, 48, 50] provides a more effective solution than SSIM/perceptual losses to align SR images with natural image distribution, thereby leading to perceptually more realistic SR results. However, it simultaneously introduces many undesirable visual artifacts [29] because the adversarial training is unstable and the limited capacity of GAN models in characterizing the space of natural images. + +Recently developed large-scale stable diffusion (SD) models [37], which are pre-trained on text-to-image (T2I) tasks, have demonstrated impressive semantic understanding and have been successfully applied to many downstream tasks [35, 66], including SR [30, 42, 46, 56, 57, 59]. While SD-based SR approaches have exhibited perceptually more realistic SR results than GAN-based methods [7], they commonly entangle the pixel-level fidelity and semantic-level enhancement objectives in the diffusion process [46, 56, 57], which can be contradictory in optimization [3, 43]. As a result, they struggle to balance the pixel-wise fidelity and semantic-level perception in the final SR output. While some methods [30, 42, 59] perform pixel-level restoration before semantic-level enhancement, the performance of the latter stage depends heavily on the accuracy of the earlier stage. + +Besides the trade-off between pixel-level fidelity and semantic-level details, users in practice often have different preferences on the SR results: some prioritize content fidelity over detail generation, while others favor semantically rich details over pixel-wise fidelity. This diversity in user preferences underscores the importance and demands for a more flexible SR approach that can accommodate individual tastes during inference. Although some previous methods [11, 34] have explored interactive SR methods to control the restoration strength, they are primarily limited to degradation-level adjustments (such as noise and blur), resulting in either smoother or sharper outputs. Several multi-step SD-based SR methods [30, 57, 59, 61] employ guidance-based strategies [14] at each sampling step to achieve varying levels of semantic richness, but these methods often struggle with precise control and efficiency. + +In this paper, we present Pixel-level and Semantic-level Adjustable Super-Resolution, namely PiSA-SR, which separates pixel-level and semantic-level enhancement into two distinct Low-Rank Adapter (LoRA) [16] weight spaces with the pre-trained SD model, providing an effective SR model for various user preferences. We begin by formulating the SD-based SR problem to learn the residual between LQ and HQ latent features [12, 64, 69]. With this formulation, we can not only accelerate the convergence of model training but also introduce scale factors on the model outputs, enabling flexible adjustment on the SR results during the inference stage without re-training. Then, we fine-tune the + +pre-trained SD model by introducing two separate LoRA modules dedicated to pixel-level regression and semantic-level enhancement. In specific, the $\ell_2$ loss is employed for the LoRA module of pixel-level regression, while the LPIPS [67] and classifier score distillation (CSD) [62] losses are employed for another LoRA module, leveraging the semantic priors encoded in the pre-trained VGG classification model [39] and SD image generation model [37]. A decoupled training approach is proposed to train pixel- and semantic-level LoRAs, effectively improving semantic information while maintaining pixel-level fidelity. + +Our experiments demonstrate that PiSA-SR not only achieves superior SR performance to existing SD-based models but also offers an effective way for users to adjust the SR styles based on their tastes. An example is shown in Fig. 1. The horizontal and vertical axes indicate the semantic-level and pixel-level guidance scales, respectively. Increasing the pixel-level scale removes noise and compression artifacts, but a too-strong scale over-smoothes image details. Conversely, increasing the semantic-level scale enriches the image details. Nonetheless, a too-high semantic scale over-enhances visual artifacts. PiSA-SR provides the flexibility to adjust both pixel-level and semantic-level guidance scales based on user preferences. + +# 2. Related Work + +Image super-resolution. Early deep learning-based SR methods [5, 10, 22, 27] aim to improve image fidelity metrics such as PSNR and SSIM [52]. By using bicubic down-sampling to generate LQ-HQ image training pairs, various approaches have been developed to enhance the SR performance, such as dense [44], residual [69], recursive connections [22], non-local strategies [47], and attention mechanisms [8, 27, 68]. Though these advancements have significantly improved the SR performance, the real-world LQ images with complex degradations are much more challenging to enhance. While researchers have proposed to collect real-world LQ-HQ paired datasets [4, 55] to train SR models, simulating realistic LQ-HQ pairs is a more economic way [50, 65]. BSRGAN [65] randomly shuffles some basic degradation operators to synthesize LQ-HQ pairs, while RealESRGAN [50] implements high-order degradation modeling. Compared with pixel-level losses, the perceptual loss [17] and GAN loss [24, 48] are more effective in improving image perceptual quality. However, the adversarial training also leads to unnatural visual artifacts in the SR results. Many of the later works aim to reduce GAN-generated artifacts from the perspectives of frequency division [25], network optimization [70], the weighting of different losses [43, 49], and image local statistics [29]. + +Diffusion model based super-resolution. Recent studies have explored the use of diffusion models (DMs) [15, 37] for SR. Early approaches adjust pre-trained DMs' + +reverse transitions [20] using gradient descent, offering training-free solutions by assuming predefined image degradations. ResShift [63] trains a DM from scratch on paired LQ-HQ data. Recently, SD models pre-trained on T2I tasks have been widely used for SR tasks due to their strong image priors [2, 6, 30, 32, 36, 46, 51, 56, 57, 59]. StableSR [46] introduces a trainable encoder and takes LQ image as a condition in SD. DiffBIR [30] first applies a restoration module to mitigate degradations, then utilizes SD to enhance details. PASD [59] employs an encoder for degradation removal and introduces a pixel-aware cross-attention module to incorporate low-level and high-level image features into the SD process. SeeSR [57] enhances semantic robustness through degradation-aware tag-style prompts to guide diffusion. Considering that the multi-step diffusion process will increase the computational cost and the risk of synthesizing unfaithful contents, researchers started to develop fewer-step DM-based SR models. SinSR [51] applies consistency-preserving distillation to shorten the diffusion process of ResShift. OSEDiff [56] employs the LQ image as direct input, eliminating the random noise sampling. It adopts the VSD loss [54, 60] to distill the generative capacity from the multi-step diffusion process, offering a one-step DM-based SR solution. However, all the above-mentioned methods face a contradiction between pixel-level and semantic-level enhancement and most of them lack the flexibility to adjust the SR style for diverse user preferences. + +# 3. Methodology + +This section first formulates the SD-based SR as a residual learning model, then introduces the dual-LoRA approach to disentangle the learning objectives of pixel-level regression and semantic-level enhancement. Finally, pixel and semantic guidances are proposed for flexible SR results generation. In our following development, we denote by $x_{L}$ and $x_{H}$ the LQ and HQ images, respectively, and by $z_{L}$ and $z_{H}$ their corresponding latent codes. Let $E$ and $D$ be the encoder and decoder of a well-trained variational auto-encoder (VAE), we can nearly have $z_{L} = E(x_{L})$ , $z_{H} = E(x_{H})$ , $x_{L} = D(z_{L})$ and $x_{H} = D(z_{H})$ . + +# 3.1. Model Formulation + +DMs such SD [37] apply a $T$ -step forward process to progressively transform an initial latent code $z_0$ into Gaussian noise $z_T$ over $T$ steps. In each timestep $t$ , the noisy latent code $z_t$ is obtained by directly adding noise $\epsilon \sim \mathcal{N}(0, I)$ to $z_0$ by $z_t = \sqrt{\bar{\alpha}_t} \cdot z_0 + \sqrt{1 - \bar{\alpha}_t} \cdot \epsilon$ , where $\bar{\alpha}_t$ is the cumulative parameter controlling the noise level at timestep $t$ . During the reverse process, multi-step SD-based SR models [2, 30, 36, 46, 57, 59] usually perform a $T$ -step denoising process to gradually transform Gaussian noise $z_T$ into the HQ latent code $z_H$ , conditioned on the LQ image $x_L$ with a ControlNet [66] (see Fig. 2 (a)). At stage $t$ , the latent + +![](images/21ba05524be1c5cd4a637aa37b3630e901b28adaf25b29f1639e41a8831cbd49.jpg) +(a) Multi-step DM-based SR methods + +![](images/d12052bcddc86787f481f76a8a7c1394d95c7c81e5e521133daba8c5c641e221.jpg) +(b) OSEDiff + +![](images/5ebbbb204c307be2d71598fccbd173d6060bec51e5349b7a8a0029c77a195c19.jpg) +(c) PiSA-SR +Figure 2. Comparison of the pipeline of different DM-based SR methods. (a) Multi-step methods [2, 30, 36, 46, 57, 59] perform $T$ denoising steps starting from Gaussian noise $z_{T}$ , conditioned on the LQ image $x_{L}$ . (b) OSEDiff [56] starts from LQ latent representation $z_{L}$ with only one-step diffusion. (c) Our proposed PiSA-SR formulates the SD-based SR as learning the residual between the LQ latent $z_{L}$ and HQ latent $z_{H}$ . + +$\hat{z}_0$ can be computed from the noise estimated by SD UNet $\epsilon_{\theta}(z_t,x_L,t)$ and the current latent $z_{t}$ as follows: + +$$ +\hat {z} _ {0} = f (z _ {t}, \epsilon_ {\theta}) = \frac {z _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} (z _ {t} , x _ {L} , t)}{\sqrt {\bar {\alpha} _ {t}}}. \tag {1} +$$ + +Then the latent $z_{t-1}$ at timestep $t - 1$ can be sampled from $p(z_{t-1} | z_t, \hat{z}_0) = N(\mu_t(z_t, \hat{z}_0), \sigma_t^2)$ , where $\mu_t$ and $\sigma_t^2$ denote the mean and variance of $z_{t-1}$ , respectively. + +However, these multi-step DM-based SR methods are computationally expensive and often yield unstable results due to the random noise sampled in the diffusion process [42, 57]. To address this issue, OSEDiff [56] starts from the LQ latent code $z_{L}$ and completes the SR process with only one diffusion step (see Fig. 2 (b)), improving both the efficiency and stability. The transformation between $z_{L}$ and $z_{H}$ can be described as follows: + +$$ +z _ {H} = f \left(z _ {L}, \epsilon_ {\theta}\right) = \frac {z _ {L} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} \left(z _ {L}\right)}{\sqrt {\bar {\alpha} _ {t}}}. \tag {2} +$$ + +Note that we omit the timestep $t$ in the above formula since $t$ can be viewed as a constant in one-step diffusion. + +Learning the residual between the LQ and GT features has been successfully applied in deep learning-based restoration methods [64, 69]. However, DMs typically perform multi-step iterations, making applying residual learning difficult. The recently developed OSEDiff [56] employs a single diffusion step to achieve end-to-end SR training, making it feasible to adopt the residual learning strategy. Here, we formulate the SR problem as learning the residual between $z_{L}$ and $z_{H}$ , as illustrated in Fig. 2 (c). We use + +subtraction in the global residual connection because DMs are trained to denoise in the inverse diffusion process. Such a residual learning formulation helps the model focus on learning the desired high-frequency information from the HQ latent features, avoiding extracting less relevant information from the LQ latent. It can also accelerate the convergence of the model training process [12]. Furthermore, we introduce a scaling factor $\lambda$ to adjust the residual $\epsilon_{\theta}$ added to the LQ latent, which can be expressed as: + +$$ +z _ {H} = z _ {L} - \lambda \epsilon_ {\theta} (z _ {L}). \tag {3} +$$ + +The $\lambda$ is fixed as 1 during training. Users can adjust the output in the inference stage by using a smaller $\lambda$ (to preserve more original contents) or a higher $\lambda$ (to enhance details more aggressively) based on their preferences. + +# 3.2. Dual-LoRA Training + +Previous SD-based SR methods mostly entangle the pixel-level and semantic-level enhancement in the training process [46, 56], making the balance of content fidelity and perpetual quality difficult. Some works sequentially perform pixel-level and semantic-level enhancement using two separate networks [30]. However, errors in the earlier stage may propagate to the next stage, limiting overall performance. Furthermore, using two separate networks increases the computational burden and memory requirements, making the approach less efficient. It has been successfully applied in customized T2I tasks using distinct LoRA modules for various generation styles [18, 38, 41]. Inspired by this, we propose a decoupled training approach by utilizing two LoRA modules under the pre-trained SD model, targeting pixel-level and semantic-level enhancement for the SR task respectively. Our approach, namely Dual-LoRA, introduces only a small number of additional parameters during training. These LoRA parameters can also be merged into the pre-trained SD model during inference. + +Fig. 3 (a) illustrates the training process of our approach. We freeze the parameters of the pre-trained VAE and introduce two trainable LoRA modules to the UNet of SD. Since the LQ image is corrupted by degradations such as noise, blur, and downsampling, we optimize the pixel-level LoRA first to reduce the effect of corruptions, followed by the optimization of semantic-level LoRA. We train the pixel-level LoRA, denoted by $\Delta \theta_{pix}$ , by the pixel-level loss function. Together with the pre-trained SD parameters, the full parameter set can be represented as $\theta_{pix} = \{\theta_{sd}, \Delta \theta_{pix}\}$ . The HQ latent can then be estimated by $z_H^{pix} = z_L - \epsilon_{\theta_{pix}}(z_L)$ and decoded by the VAE decoder with $x_H^{pix} = D(z_H^{pix})$ . + +For semantic-level enhancement, we train another LoRA, denoted by $\Delta \theta_{sem}$ , by the semantic-level loss function. To disentangle the pixel-level and semantic-level objectives, we fix the already trained pixel-level LoRA $\Delta \theta_{pix}$ and combine it with the semantic-level LoRA $\Delta \theta_{sem}$ to be + +trained, forming a PiSA-LoRA group (see Fig. 3 (a)). Note that only the semantic-level LoRA module within the PiSA-LoRA group is updated in this stage. This ensures that the optimization process focuses on the semantic details without much interference from the pixel-level degradations. The full parameter set during this PiSA-LoRA training stage can be expressed as $\theta_{PiSA} = \{\theta_{sd},\Delta \theta_{pix},\Delta \theta_{sem}\}$ . The HQ latent after this stage is given by $z_H^{sem} = z_L - \epsilon_{\theta_{PiSA}}(z_L)$ , which is then decoded by the VAE decoder with $x_H^{sem} = D(z_H^{sem})$ . + +# 3.3. Pixel-level and Semantic-level Losses + +We train the pixel-level LoRA using the $\ell_2$ loss. As shown by $D(\epsilon_{\theta_{pix}}(z_L))$ in Fig. 4, the $\ell_2$ loss can effectively remove degradation and enhance edges. However, it is insufficient to generate semantic-level details [24, 29, 48], resulting in smooth SR outputs. The LPIPS loss [67] can be used to activate semantic details by aligning the high-level features with a pre-trained classification VGG network [39], which is however trained on a limited set of image categories. The GAN loss [24] captures semantic information through adversarial training, encouraging the generator to produce more realistic images. However, GAN loss can be unstable in training and generate much artifacts [29]. + +Recently developed SD models [37] have shown superior performance in generating open-category images with enhanced semantic details. As a conditional generation model, SD can be used to form up an implicit classifier [14] for the posterior distribution modeling, and its gradient w.r.t. the synthesized image $x_{t}$ is as follows: + +$$ +\begin{array}{l} \nabla_ {x _ {t}} \log p _ {\theta} (c \mid x _ {t}) = \nabla_ {x _ {t}} \log p _ {\theta} (x _ {t} \mid c) - \nabla_ {x _ {t}} \log p _ {\theta} (x _ {t}) \\ = - \left(\epsilon_ {\theta} \left(x _ {t}, c, t\right) - \epsilon_ {\theta} \left(x _ {t}, t\right)\right) / \sigma_ {t}, \tag {4} \\ \end{array} +$$ + +where $c$ is the text prompt, $t$ is the current timestep, and $\sigma_{t} = \sqrt{1 - \bar{\alpha}_{t}}$ . The SD model, parameterized by $\theta$ , model score functions of $\nabla_{x_t}\log p_\theta (x_t\mid c)$ and $\nabla_{x_t}\log p_\theta (x_t)$ using noise predictions [40]. By taking the expectation of the gradient in Eq. (4) over all possible $t$ , we have the CSD [14] loss, which is originally used in 3D generation tasks to optimize the posterior probability of rendered images aligning their semantic content with text prompts. + +Motivated by the functionality of CSD loss in DMs for generation, we investigate the usage of CSD loss for semantic-level enhancement in the SR task. Following [56, 60], and for the clarity of the following context, we formulate the gradient of CSD in the latent space instead of the noise domain. Specifically, the gradient of CSD loss can be formulated in Eq. (5) below (the detailed derivation can be found in the supplementary materials): + +$$ +\nabla \ell_ {C S D} ^ {\lambda_ {c f g}} = \mathbb {E} _ {t, \epsilon , z _ {t}, c} \left[ w _ {t} \left(f (z _ {t}, \epsilon_ {r e a l}) - f (z _ {t}, \epsilon_ {r e a l} ^ {\lambda_ {c f g}})\right) \frac {\partial z _ {H} ^ {s e m}}{\partial \theta_ {P i S A}} \right], \tag {5} +$$ + +![](images/92a9c6f205b7f5bdf5817974d47b3fdc250ff0c02ef5bd1438db8d12d2d67032.jpg) +(a) Training process of the PiSA-SR + +![](images/ceec4e76b76f126dfa4803000686b836e72b86bc51c0bfb975de412a3e0692e8.jpg) +(b) Inference process of PiSA-SR + +![](images/607e5f283805224f3efbfca2cf926a683d5a9f2476049342273a5ba62b4ff2b3.jpg) +Figure 3. The (a) training and (b) inference procedures of PiSA-SR. During the training process, two LoRA modules are respectively optimized for pixel-level and semantic-level enhancement. During the inference stage, users can use the default setting to reconstruct the HQ image in one-step diffusion or adjust $\lambda_{pix}$ and $\lambda_{sem}$ to control the strengths of pixel-level and semantic-level enhancement. +Figure 4. The model outputs with pixel-wise and semantic-level losses for a given LQ image. + +where the gradient expectation is calculated considering all diffusion timesteps $t = 1,\dots,T$ , the noise is sampled from $\epsilon \sim \mathcal{N}(0,I),z_{t}$ is obtained by $z_{t} = \sqrt{\bar{\alpha}_{t}}\cdot z_{H}^{sem} + \sqrt{1 - \bar{\alpha}_{t}}\cdot \epsilon$ $c$ is the text prompt extracted from $z_H^{sem}$ $f(\cdot)$ is the function as Eq. (1), $w_{t}$ is the timestep-dependent scalar weight, and $\epsilon_{real}^{\lambda_{cfg}}$ denotes the pre-trained SD output with the CFG term $\epsilon_{real}^{cls}$ same as Eq. (4) [14, 62]: + +$$ +\epsilon_ {r e a l} ^ {\lambda_ {c f g}} = \epsilon_ {r e a l} (z _ {t}, t) + \underbrace {\lambda_ {c f g} (\epsilon_ {r e a l} (z _ {t} , t , c) - \epsilon_ {r e a l} (z _ {t} , t))}. \qquad (6) +$$ + +Note that the VSD loss [54, 60], which also aims to align the distributions of enhanced and natural images within the latent space, has been verified effective for SR by OSEDiff [56]. The gradient of VSD loss can be written as: + +$$ +\nabla \ell_ {V S D} ^ {\lambda_ {c f g}} = \mathbb {E} _ {t, \epsilon , z _ {t}, c} \left[ w _ {t} \left(f (z _ {t}, \epsilon_ {f a k e}) - f (z _ {t}, \epsilon_ {r e a l} ^ {\lambda_ {c f g}})\right) \frac {\partial z _ {H} ^ {s e m}}{\partial \theta_ {P i S A}} \right], \tag {7} +$$ + +where $\epsilon_{\text{fake}}$ denotes the output of fine-tuned SD aligned with the distribution of synthetic images. The VSD loss with $\lambda_{cfg}$ can be divided into two components: the VSD loss with $\lambda_{cfg} = 0.0$ , i.e., $\ell_{VSD}^{\lambda_{cfg} = 0.0}$ , and the CSD loss. Coupling with the LPIPS loss, we visualize the semantic-level LoRA optimization results of the two components of VSD in Fig. 4. We see that the CSD loss with the normal guidance scale of $\lambda_{cfg} = 7.5$ contributes more significantly to semantic enhancement. In contrast, the VSD loss with $\lambda_{cfg} = 0.0$ weakens the semantic details. More importantly, the component $\ell_{VSD}^{\lambda_{cfg} = 0.0}$ requires bi-level optimization, leading to intensive memory usage and unstable training [31]. In comparison, the CSD loss is free of bi-level optimization, reducing significantly memory usage and improving training stability. Therefore, we integrate LPIPS and CSD losses for the semantic-level LoRA optimization. + +# 3.4. The Inference Process of PiSA-SR + +Fig. 3(b) shows the inference process of PiSA-SR. In the default inference setting (represented by the green dash-dot line), only the PiSA-LoRA, which merges the pixel-level and semantic-level LoRA modules, is used to process the input together with the pre-trained SD model, achieving state-of-the-art SR performance in just one-step diffusion (see our experiments in Section 4.3). To achieve flexible SR with diverse user preferences, we extend the strategy in Eq. (3) by introducing a pair of pixel and semantic guidance factors, denoted by $\lambda_{pix}$ and $\lambda_{sem}$ , to control the SR results: + +$$ +\epsilon_ {\theta} \left(z _ {L}\right) = \lambda_ {p i x} \epsilon_ {\theta_ {p i x}} \left(z _ {L}\right) + \lambda_ {s e m} \left(\epsilon_ {\theta_ {P s S A}} \left(z _ {L}\right) - \epsilon_ {\theta_ {p i x}} \left(z _ {L}\right)\right). \tag {8} +$$ + +In Eq. (8), $\epsilon_{\theta_{pix}}(z_L)$ is the output with only pixel-level LoRA, while $\epsilon_{\theta_{PiSA}}(z_L)$ is the output with both pixel and semantic level enhancement. Inspired by CFG [14], the difference between them, i.e., $\epsilon_{\theta_{PiSA}}(z_L) - \epsilon_{\theta_{pix}}(z_L)$ , can well separate pixel-level degradations from semantic details (see the red box in Fig. 4). Therefore, as illustrated in Fig. 3(b) (represented by the black dashed line), by adjusting $\lambda_{pix}$ and $\lambda_{sem}$ , we can control the contributions of pixel-level LoRA and semantic-level LoRA, and consequently deliver the SR results with different styles, as shown in Fig. 1. + +# 4. Experiments + +# 4.1. Experimental Settings + +Training settings. We train PiSA-SR upon SD 2.1-base for the $\times 4$ SR task. The two trainable LoRA modules are applied to the weights of all convolution and MLP layers. Both modules are initialized using a Gaussian distribution with a rank of 4. Following SeeSR [57] and OSEDiff [56], we employ LSDIR [26] and the first 10K images from FFHQ [19] as the training data. We generate paired training data using the degradation pipeline from RealESRGAN [50]. The batch size is 16 and the training patch size is $512 \times 512$ . The Adam optimizer [23] is used with a learning rate 5e-5. The pixel-level and semantic-level LoRA modules undergo 4K and 8.5K training iterations. + +Compared methods. We compare PiSA-SR with leading multi-step DM-based methods StableSR [46], ResShift [63], DiffBIR [30], PASD [59], and SeeSR [57], one-step DM-based methods SinSR [51] and OSEDiff [56], and GAN-based SR methods RealESRGAN [50], BSRGAN [65] and LDL [29]. All comparative results are obtained using officially released codes or models. Due to space constraints, comparisons with GAN-based methods are presented in the supplementary materials. + +Test datasets and evaluation metrics. Following previous arts [46, 56, 57], we evaluate the competing methods using synthetic and real-world test data. The synthetic dataset comprises 3000 images cropped from DIV2K [1] with size $512 \times 512$ , degraded using RealESRGAN + +Table 1. Results of PiSA-SR with different pixel-semantic guidance scales on the RealSR test dataset. + +
λpixλsemPSNR↑LIPIS↓CLIPIQA↑MUSIQ↑
0.01.025.960.34260.412946.45
0.21.026.480.30420.486854.05
0.51.026.750.26460.570563.82
0.81.026.180.26120.629268.95
1.01.025.500.26720.670270.15
1.21.024.760.27230.674670.33
1.51.023.740.27690.630569.23
1.00.026.920.30180.322749.62
1.00.226.950.27840.359153.64
1.00.526.770.24760.432258.76
1.00.826.200.24650.580666.33
1.01.025.500.26720.670270.15
1.01.224.590.30000.701571.60
1.01.523.080.35410.683571.76
+ +degradation pipeline [50]. The real-world data are center-cropped from RealSR [4] and DrealSR [55] datasets with size $128 \times 128$ for LQ images and $512 \times 512$ for HQ images. PSNR and SSIM [52], computed on the Y channel in the YCbCr space, are used to measure the fidelity of SR results; LPIPS [67] and DISTS [9], computed in the RGB space, are used to assess the perceptual quality of SR results; FID [13] evaluates the distance of distributions between GT and restored images; NIQE [33], CLIPIQA [45], MUSIQ [21] and MANIQA [58] evaluate the image quality without reference images. Additionally, a user study is conducted to validate the effectiveness of the proposed adjustable SR in the supplementary materials. + +# 4.2. Experiments on Adjustable SR + +We conduct experiments by fixing one guidance scale ( $\lambda_{pix}$ or $\lambda_{sem}$ ) at 1 and varying the other to observe how the reconstructed images change. PSNR, LPIPS, CLIPIQA, and MUSIQ are used to evaluate performance. PSNR measures pixel-level fidelity; LPIPS evaluates the image perceptual quality with GT as reference; CLIPIQA and MUSIQ measure the image quality without using a reference. The results are shown in Table 1. First, we observe that increasing the pixel-level scale $\lambda_{pix}$ leads to a continuous improvement in the no-reference metrics CLIPIQA and MUSIQ. This is because increasing $\lambda_{pix}$ can remove image degradation and enhance edge, which is favorable to these two metrics. However, the reference-based metrics PSNR and LPIPS exhibit a rise-and-fall pattern. PSNR peaks at $\lambda_{pix} = 0.5$ , indicating the best pixel-level fidelity. LPIPS reaches its best value at $\lambda_{pix} = 0.8$ (0.2612), suggesting that at this point the restored image is perceptually most similar to the GT with rich details. Further increasing $\lambda_{pix}$ reduces both PSNR and LPIPS scores. Second, increasing the semantic-level scale $\lambda_{sem}$ also results in a continuous improvement in CLIPIQA and MUSIQ, but with a higher upper bound than pixel-level adjustments. This is because increasing $\lambda_{sem}$ can synthesize more semantic-level details (see Fig. 1). However, PSNR decreases as $\lambda_{sem}$ increases, while LPIPS first improves and peaks at $\lambda_{sem} = 0.8$ (0.2465), and then + +Table 2. Quantitative comparison among the state-of-the-art DM-based SR methods on synthetic and real-world test datasets. 'S' denotes the number of diffusion steps. The best and the second-best results are highlighted in red and blue, respectively. + +
DatasetsMethodsPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓CLIPQA↑MUSIQ↑MANIQA↑
DIV2KResShift-S1524.690.61750.33740.221536.016.820.608960.920.5450
StableSR-S20023.310.57280.31290.213824.674.760.668265.630.6188
DiffBIR-S5023.670.56530.35410.212930.934.710.665265.660.6204
PASD-S2023.140.54890.36070.221929.324.400.671168.830.6484
SeeSR-S5023.710.60450.32070.196725.834.820.685768.490.6239
SinSR-S124.430.60120.32620.206635.456.020.649962.800.5395
OSEDiff-S123.720.61080.29410.197626.324.710.668367.970.6148
PiSA-SR-S123.870.60580.28230.193425.074.550.692769.680.6400
RealSRResShift-S1526.310.74110.34890.2498142.817.270.545058.100.5305
StableSR-S20024.690.70520.30910.2167127.205.760.619565.420.6211
DiffBIR-S5024.880.66730.35670.2290124.565.630.641264.660.6231
PASD-S2025.220.68090.33920.2259123.085.180.650268.740.6461
SeeSR-S5025.330.72730.29850.2213125.665.380.659469.370.6439
SinSR-S126.300.73540.32120.2346137.056.310.620460.410.5389
OSEDiff-S125.150.73410.29210.2128123.505.650.669369.090.6339
PiSA-SR-S125.500.74170.26720.2044124.095.500.670270.150.6560
DrealSRResShift-S1528.450.76320.40730.2700175.928.280.525949.860.4573
StableSR-S20028.040.74600.33540.2287147.036.510.617158.500.5602
DiffBIR-S5026.840.66600.44460.2706167.386.020.629260.680.5902
PASD-S2027.480.70510.38540.2535157.365.570.671464.550.6130
SeeSR-S5028.260.76980.31970.2306149.866.520.667264.840.6026
SinSR-S128.410.74950.37410.2488177.057.020.636755.340.4898
OSEDiff-S127.920.78350.29680.2165135.296.490.696364.650.5899
PiSA-SR-S128.310.78040.29600.2169130.616.200.697066.110.6156
+ +declines. This is because excessive semantic details may induce changes in image content and reduce pixel-level fidelity. Meanwhile, the over-enhanced images may exhibit many details different from the GT, making the reference-based LPIPS index deteriorate. + +Fig. 1 illustrates the visual comparisons of different pixel-semantic scales. More comparisons can be found in the supplementary materials. From bottom to top, as the pixel-level scale $\lambda_{pix}$ increases, the model progressively removes degradation from the input LQ image, resulting in clearer and sharper visual output. However, a too-high $\lambda_{pix}$ can lead to loss of details, and diminish the effectiveness of semantic improvements, as seen in the top-right corner of Fig. 1. From left to right, increasing the semantic-level scale $\lambda_{sem}$ enhances the semantic richness of the image, making the details of the wrinkles, beard, and hair of Albert Einstein more pronounced. However, an excessive semantic level such as $\lambda_{sem} = 1.5$ introduces unnatural artifacts, making the SR result less realistic. With the flexibility of PiSASR, users can customize their preferences, either preserving more fidelity or emphasizing richer semantic enhancements, depending on their specific use case. + +# 4.3. Comparisons with State-of-the-Arts + +As shown in Eq. (8), letting $\lambda_{pix} = 1$ and $\lambda_{sem} = 1$ , we have the default version of PiSA-SR, which can be performed in just one-step diffusion. Here, we compare this default version of PiSA-SR with the competing methods. + +Quantitative comparisons. Table 2 compares the performance of our default PiSA-SR model with other DM-based Real-SR methods. The following findings can be made. ResShift [63] and its distilled version SinSR [51] show lim + +ited performance in no-reference metrics, implying poor perceptual quality. PASD [59] and SeeSR [57] leverage the pre-trained SD as their base model, and extract additional high-level information [71] to improve the perceptual quality of images. Their no-reference metric scores, such as MUSIQ and CLIPIQA, are highly improved. However, the multi-step inference makes them less efficient. In addition, their relatively low LPIPS and DISTS scores suggest their less faithful reconstruction to the GT. OSEDiff [56] distills the generation capacity of multi-step DM into just one-step diffusion by the VSD loss [54, 60]. While effective and efficient, its no-reference metrics are limited. Our proposed PiSA-SR also exhibits high efficiency by requiring only one inference step, while showing impressive pixel-level fidelity and semantic-level perceptual quality. It outperforms other methods not only in reference-based metrics such as LPIPS and DISTS but also in no-reference metrics such as CLIP-IQA, MUSIQ, and MANIQA. + +Qualitative comparisons. We present visual comparisons in Fig. 5. ResShift and SinSR fail to generate richer textures and details because of their limited generation capacity. StableSR and DiffBIR can generate finer details (e.g., more realistic penguin feathers in the second example), due to the leveraged strong SD prior. PASD and SeeSR incorporate high-level information guidance, resulting in more comprehensive details. However, the inherent randomness in PASD and SeeSR may lead to over-enhanced (e.g., unnatural penguin feathers generated by SeeSR) or over-smoothed details (e.g., the blurred texture of the wooden stool generated by SeeSR). OSEDiff can generate more consistent results in both examples. However, the restored details can be limited. In contrast, the proposed PiSA-SR delivers high-quality, re + +![](images/83e03c8fc3d1a51dc0c8c80547632ab54021d8500c334b73108bd595dc10a4a6.jpg) + +![](images/8c94bb7aeb1ca51df5be79f5d4df3861f8e4419d41172956c38f4a290d045125.jpg) + +![](images/8821bc87bfc13e4e8b9d850820813b08cca674b5887ae7eecf6034a4ec839543.jpg) + +![](images/8ab7a306b01b1a7ba098f4159268641d8d75cf459971510514f8f9d395220843.jpg) +Figure 5. Visual comparisons of different DM-based SR methods. Please zoom in for a better view. + +![](images/6537849122bd587a018edc04a77c6cc65c4c3c0f4587192d3892eef8d5b2b73c.jpg) + +![](images/15f6fa77e3e01fcdfce6f529044557eda47a85d4d439d6140062f7156a52b17a.jpg) + +Table 3. The inference time and the number of parameters of DM-based SR methods. + +
StableSRResShiftDiffBIRPASDSeeSRSinSROSEDiffPiSA-SR-def.PiSA-SR-adj.
Inference Steps200155020501112
Inference time(s)/Image10.030.762.722.804.300.130.120.090.13
#Params(B)1.560.181.682.312.510.181.771.301.30
+ +alistic SR results. With the dual-LoRA design, the restored structures align well with the input LQ image, providing more accurate fidelity information and generating more natural and richer details. More visualization comparisons can be found in supplementary materials. + +Complexity comparisons. We compare the number of parameters and inference time of competing DM-based SR models in Table 3, where PiSA-SR-def. and PiSA-SR-adj. represent its default setting and adjustable setting, respectively. Inference time is measured on the $\times 4$ SR task with $128\times 128$ LQ images using a single NVIDIA A100 80G GPU. ResShift has fewer parameters because it is trained from scratch without using the pre-trained SD model. SinSR inherits the parameters from its parental ResShift model. Among the SD-based SR methods, PiSA-SR-def. has the fewest parameters and achieves the fastest inference time. Unlike OSEDiff, PiSA-SR-def. does not require a semantic extractor (e.g., RAM [71]) during inference. Instead, it distills the semantic information into the semantic-level LoRA with CSD. PiSA-SR-adj., which requires two diffusion steps to compute outputs from pixellevel and semantic-level LoRA modules, incurs a slightly longer inference time over the default setting. However, this small increase in time offers users the advantage of generating SR outputs tailored to their specific needs. + +Ablation study. We performed a series of ablation stud + +ies for a better understanding of the dual-LoRA training approach, and the roles of the two LoRA modules. The details can be found in the supplementary materials. + +# 5. Conclusion and Limitation + +In this work, we presented PiSA-SR, a novel SR framework that disentangles pixel-level and semantic-level objectives by learning two LoRA modules upon the pre-trained SD model. By decoupling the learning process into two distinct spaces, PiSA-SR effectively balanced pixel-wise regression and perceptual quality. The commonly used $\ell_2$ loss was used to optimize pixel-level LoRA, while the LPIPS and CSD losses were employed to optimize the semantic-level LoRA. PiSA-SR demonstrated impressive performance in both effectiveness and efficiency. Furthermore, it offered high flexibility during inference through adjustable guidance scales, allowing users to customize the SR results without model re-training. + +While PiSA-SR provided a flexible solution for adjusting pixel-level fidelity and semantic-level perception, it incurs a slight inference time increment over its default version. In addition, we use a single pixel-level LoRA to handle various degradations in the LQ, which may be insufficient for heavy degradations. In the future, we will explore refining separate LoRA spaces for different image degradations and the possibility of image-adaptive scale factors. + +# References + +[1] Eirikur Agustsson and Radu Timofte. NTIRE 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 6 +[2] Yang Ai, Xiaogiang Zhou, Huaibo Huang, Xiaotian Han, Zhengyu Chen, Quanzeng You, and Hongxia Yang. Dreamclear: High-capacity real-world image restoration with privacy-safe dataset curation. arXiv preprint arXiv:2410.18666, 2024.3 +[3] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 1, 2 +[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 6 +[5] Dong Chao, Loy Chen Change, He Kaiming, and Tang Xiaou. Learning a deep convolutional network for image super-resolution. In ECCV, pages 184-199, 2014. 1, 2 +[6] Bin Chen, Gehui Li, Rongyuan Wu, Xindong Zhang, Jie Chen, Jian Zhang, and Lei Zhang. Adversarial diffusion compression for real-world image super-resolution. arXiv preprint arXiv:2411.13383, 2024. 3 +[7] Du Chen, Tianhe Wu, Kede Ma, and Lei Zhang. Toward generalized image quality assessment: Relaxing the perfect reference quality assumption, 2025. 2 +[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 1, 2 +[9] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 6 +[10] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 1, 2 +[11] Jingwen He, Chao Dong, Yihao Liu, and Yu Qiao. Interactive multi-dimension modulation for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9363-9379, 2021. 2 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 4 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 4, 5, 6 + +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2 +[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 1, 2 +[18] Maxwell Jones, Sheng-Yu Wang, Nupur Kumari, David Bau, and Jun-Yan Zhu. Customizing text-to-image models with a single image pair. arXiv preprint arXiv:2405.01536, 2024. 4 +[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6 +[20] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 3 +[21] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 6 +[22] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 1, 2 +[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[24] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2, 4 +[25] Wenbo Li, Kun Zhou, Lu Qi, Liying Lu, and Jiangbo Lu. Best-Buddy GANs for highly detailed image superresolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1412-1420, 2022. 2 +[26] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Lsdir: A large scale dataset for image restoration. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1775-1787, 2023. 6 +[27] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 1, 2 + +[28] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, 2022. 2 +[29] Jie Liang, Hui Zeng, and Lei Zhang. Details or Artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 1, 2, 4, 6 +[30] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 2, 3, 4, 6 +[31] Zhiyuan Ma, Yuxiang Wei, Yabin Zhang, Xiangyu Zhu, Zhen Lei, and Lei Zhang. Scaledreamer: Scalable text-to-3d synthesis with asynchronous score distillation. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 5 +[32] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Conditional diffusion distillation. arXiv preprint arXiv:2310.01407, 2023. 3 +[33] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6 +[34] Chong Mou, Yanze Wu, Xintao Wang, Chao Dong, Jian Zhang, and Ying Shan. Metric learning based interactive modulation for real-world super-resolution. In European Conference on Computer Vision, pages 723-740. Springer, 2022. 2 +[35] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2 +[36] Yunpeng Qu, Kun Yuan, Kai Zhao, Qizhi Xie, Jinhua Hao, Ming Sun, and Chao Zhou. Xpsr: Cross-modal priors for diffusion-based image super-resolution. arXiv preprint arXiv:2403.05049, 2024. 3 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4 +[38] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In European Conference on Computer Vision, pages 422-438. Springer, 2025. 4 +[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 2, 4 +[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4 + +[41] Nick Stracke, Stefan Andreas Baumann, Joshua M Susskind, Miguel Angel Bautista, and Björn Ommer. Ctrloralter: Conditional loradapter for efficient 0-shot control & altering of t2i models. arXiv preprint arXiv:2405.07913, 2024. 4 +[42] Lingchen Sun, Rongyuan Wu, Zhengqiang Zhang, Hongwei Yong, and Lei Zhang. Improving the stability of diffusion models for content consistent super-resolution. arXiv e-prints, pages arXiv-2401, 2023. 2, 3 +[43] Lingchen Sun, Jie Liang, Shuaizheng Liu, Hongwei Yong, and Lei Zhang. Perception-distortion balanced superresolution: A multi-objective optimization perspective. IEEE Transactions on Image Processing, 33:4444-4458, 2024. 2 +[44] Tong Tong, Gen Li, Xiejie Liu, and Qinquan Gao. Image super-resolution using dense skip connections. In Proceedings of the IEEE international conference on computer vision, pages 4799-4807, 2017. 2 +[45] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 6 +[46] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 4, 6 +[47] Risheng Wang, Tao Lei, Wenzheng Zhou, Qi Wang, Hongying Meng, and Asoke K Nandi. Lightweight non-local network for image super-resolution. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1625-1629. IEEE, 2021. 2 +[48] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. ESRGAN: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 4 +[49] Xintao Wang, Ke Yu, Chao Dong, Xiaou Tang, and Chen Change Loy. Deep network interpolation for continuous imagery effect transition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1701, 2019. 2 +[50] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 2, 6 +[51] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 3, 6, 7 +[52] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 1, 2, 6 +[53] Zhihao Wang, Jian Chen, and Steven C. H. Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions + +on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, 2021. 1 +[54] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3, 5, 7 +[55] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 2, 6 +[56] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 4, 5, 6, 7 +[57] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 2, 3, 6, 7 +[58] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6 +[59] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 2, 3, 6, 7 +[60] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 3, 4, 5, 7 +[61] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. arXiv preprint arXiv:2401.13627, 2024. 2 +[62] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and XIAOJUAN QI. Text-to-3d with classifier score distillation. In The Twelfth International Conference on Learning Representations, 2024. 2, 5 +[63] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. arXiv preprint arXiv:2307.12348, 2023. 3, 6, 7 +[64] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 2, 3 +[65] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF + +International Conference on Computer Vision, pages 4791-4800, 2021. 2, 6 +[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3 +[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 4, 6 +[68] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. arXiv preprint arXiv:2203.06697, 2022. 1, 2 +[69] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 2, 3 +[70] Yuehan Zhang, Bo Ji, Jia Hao, and Angela Yao. Perception-distortion balanced ADMM optimization for single-image super-resolution. In European Conference on Computer Vision, pages 108-125. Springer, 2022. 2 +[71] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 7, 8 \ No newline at end of file diff --git a/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/images.zip b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..082a20be261af88b9cf2523970aff7acbf3fab75 --- /dev/null +++ b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3381ba06478052406b073d62f0763980ec219463af63d4d1c5ab02a9e6e34ae +size 640573 diff --git a/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/layout.json b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..03a5e2355dbe3b3c3c49e25f0533de61fee5ca9c --- /dev/null +++ b/2025/Pixel-level and Semantic-level Adjustable Super-resolution_ A Dual-LoRA Approach/layout.json @@ -0,0 +1,9563 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 489, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 489, + 139 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 489, + 139 + ], + "type": "text", + "content": "Pixel-level and Semantic-level Adjustable Super-resolution: A Dual-LoRA Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "spans": [ + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "text", + "content": "Lingchen Sun" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "text", + "content": ", Rongyuan Wu" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "text", + "content": ", Zhiyuan Ma" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "text", + "content": ", Shuaizheng Liu" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "text", + "content": ", Qiaosi Yi" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "text", + "content": ", Lei Zhang" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "text", + "content": "The Hong Kong Polytechnic University \n" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 160, + 539, + 190 + ], + "type": "text", + "content": "OPPO Research Institute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 191, + 419, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 191, + 419, + 204 + ], + "spans": [ + { + "bbox": [ + 187, + 191, + 419, + 204 + ], + "type": "text", + "content": "{ling-chen.sun, rong-yuan.wu, zm2354.ma} @connect.polyu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 161, + 205, + 449, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 205, + 449, + 217 + ], + "spans": [ + { + "bbox": [ + 161, + 205, + 449, + 217 + ], + "type": "text", + "content": "{shuaizhengliu21,qiaosiyijoyies} " + }, + { + "bbox": [ + 161, + 205, + 449, + 217 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 161, + 205, + 449, + 217 + ], + "type": "text", + "content": " gmail.com cslzhang@comp.polyu.edu.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 270, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 605 + ], + "type": "text", + "content": "Diffusion prior-based methods have shown impressive results in real-world image super-resolution (SR). However, most existing methods entangle pixel-level and semantic-level SR objectives in the training process, struggling to balance pixel-wise fidelity and perceptual quality. Meanwhile, users have varying preferences on SR results, thus it is demanded to develop an adjustable SR model that can be tailored to different fidelity-perception preferences during inference without re-training. We present Pixel-level and Semantic-level Adjustable SR (PiSA-SR), which learns two LoRA modules upon the pre-trained stable-diffusion (SD) model to achieve improved and adjustable SR results. We first formulate the SD-based SR problem as learning the residual between the low-quality input and the high-quality output, then show that the learning objective can be decoupled into two distinct LoRA weight spaces: one is characterized by the " + }, + { + "bbox": [ + 54, + 270, + 297, + 605 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 54, + 270, + 297, + 605 + ], + "type": "text", + "content": "-loss for pixel-level regression, and another is characterized by the LPIPS and classifier score distillation losses to extract semantic information from pre-trained classification and SD models. In its default setting, PiSA-SR can be performed in a single diffusion step, achieving leading real-world SR results in both quality and efficiency. By introducing two adjustable guidance scales on the two LoRA modules to control the strengths of pixel-wise fidelity and semantic-level details during inference, PiSA-SR can offer flexible SR results according to user preference without re-training. The source code of our method can be found at https://github.com/csslc/PiSA-SR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 630, + 135, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 135, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 135, + 642 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "type": "text", + "content": "Single image super-resolution (SR) [53] aims to reconstruct a high-quality (HQ) image from its low-quality (LQ) counterpart suffering from various degradations, e.g., noise, blur," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 325, + 245, + 547, + 467 + ], + "blocks": [ + { + "bbox": [ + 325, + 245, + 547, + 467 + ], + "lines": [ + { + "bbox": [ + 325, + 245, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 325, + 245, + 547, + 467 + ], + "type": "image", + "image_path": "3f1dbd479a492cd7e51b28cdbaef7cbf576e830eb36022417feeb227c97368cb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 468, + 555, + 555 + ], + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 555 + ], + "type": "text", + "content": "Figure 1. Visual illustration of our pixel- and semantic-level adjustable method for real-world SR. By increasing the pixel-level guidance scale " + }, + { + "bbox": [ + 313, + 468, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\lambda_{pix}" + }, + { + "bbox": [ + 313, + 468, + 555, + 555 + ], + "type": "text", + "content": ", the image degradations such as noise and compression artifacts can be gradually removed; however, a too-strong " + }, + { + "bbox": [ + 313, + 468, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\lambda_{pix}" + }, + { + "bbox": [ + 313, + 468, + 555, + 555 + ], + "type": "text", + "content": " will make the SR image over-smoothed. By increasing the semantic-level guidance scale " + }, + { + "bbox": [ + 313, + 468, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\lambda_{sem}" + }, + { + "bbox": [ + 313, + 468, + 555, + 555 + ], + "type": "text", + "content": ", the SR images will have more semantic details; nonetheless, a too-high " + }, + { + "bbox": [ + 313, + 468, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\lambda_{sem}" + }, + { + "bbox": [ + 313, + 468, + 555, + 555 + ], + "type": "text", + "content": " will generate visual artifacts. Please zoom in for a better view." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "downsampling, etc. SR is a challenging ill-posed problem due to the inherent ambiguity in reconstructing details [3, 29], and various deep neural networks (DNNs) [5, 8, 10, 22, 27, 68] have been developed to address it. The pixel-level regression losses, such as " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " losses, are essential to keep the pixel-level fidelity of SR outputs, yet they tend to produce over-smoothed details [3, 29]. The SSIM [52] and perceptual losses can alleviate this problem. In specific, SSIM measures the local structural similarity between the SR and ground-truth (GT) images, and the perceptual loss [17] extracts semantic features with a pre-trained classification model [39] to improve the per" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "text", + "content": "*Corresponding author. This work is supported by the PolyU-OPPO Joint Innovative Research Center." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2333" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "ceptual quality of SR results. The adversarial loss associated with the generative adversarial network (GAN) [24, 28, 29, 48, 50] provides a more effective solution than SSIM/perceptual losses to align SR images with natural image distribution, thereby leading to perceptually more realistic SR results. However, it simultaneously introduces many undesirable visual artifacts [29] because the adversarial training is unstable and the limited capacity of GAN models in characterizing the space of natural images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 374 + ], + "type": "text", + "content": "Recently developed large-scale stable diffusion (SD) models [37], which are pre-trained on text-to-image (T2I) tasks, have demonstrated impressive semantic understanding and have been successfully applied to many downstream tasks [35, 66], including SR [30, 42, 46, 56, 57, 59]. While SD-based SR approaches have exhibited perceptually more realistic SR results than GAN-based methods [7], they commonly entangle the pixel-level fidelity and semantic-level enhancement objectives in the diffusion process [46, 56, 57], which can be contradictory in optimization [3, 43]. As a result, they struggle to balance the pixel-wise fidelity and semantic-level perception in the final SR output. While some methods [30, 42, 59] perform pixel-level restoration before semantic-level enhancement, the performance of the latter stage depends heavily on the accuracy of the earlier stage." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 376, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 568 + ], + "type": "text", + "content": "Besides the trade-off between pixel-level fidelity and semantic-level details, users in practice often have different preferences on the SR results: some prioritize content fidelity over detail generation, while others favor semantically rich details over pixel-wise fidelity. This diversity in user preferences underscores the importance and demands for a more flexible SR approach that can accommodate individual tastes during inference. Although some previous methods [11, 34] have explored interactive SR methods to control the restoration strength, they are primarily limited to degradation-level adjustments (such as noise and blur), resulting in either smoother or sharper outputs. Several multi-step SD-based SR methods [30, 57, 59, 61] employ guidance-based strategies [14] at each sampling step to achieve varying levels of semantic richness, but these methods often struggle with precise control and efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we present Pixel-level and Semantic-level Adjustable Super-Resolution, namely PiSA-SR, which separates pixel-level and semantic-level enhancement into two distinct Low-Rank Adapter (LoRA) [16] weight spaces with the pre-trained SD model, providing an effective SR model for various user preferences. We begin by formulating the SD-based SR problem to learn the residual between LQ and HQ latent features [12, 64, 69]. With this formulation, we can not only accelerate the convergence of model training but also introduce scale factors on the model outputs, enabling flexible adjustment on the SR results during the inference stage without re-training. Then, we fine-tune the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "pre-trained SD model by introducing two separate LoRA modules dedicated to pixel-level regression and semantic-level enhancement. In specific, the " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " loss is employed for the LoRA module of pixel-level regression, while the LPIPS [67] and classifier score distillation (CSD) [62] losses are employed for another LoRA module, leveraging the semantic priors encoded in the pre-trained VGG classification model [39] and SD image generation model [37]. A decoupled training approach is proposed to train pixel- and semantic-level LoRAs, effectively improving semantic information while maintaining pixel-level fidelity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 205, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 361 + ], + "type": "text", + "content": "Our experiments demonstrate that PiSA-SR not only achieves superior SR performance to existing SD-based models but also offers an effective way for users to adjust the SR styles based on their tastes. An example is shown in Fig. 1. The horizontal and vertical axes indicate the semantic-level and pixel-level guidance scales, respectively. Increasing the pixel-level scale removes noise and compression artifacts, but a too-strong scale over-smoothes image details. Conversely, increasing the semantic-level scale enriches the image details. Nonetheless, a too-high semantic scale over-enhances visual artifacts. PiSA-SR provides the flexibility to adjust both pixel-level and semantic-level guidance scales based on user preferences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 372, + 400, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 400, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 400, + 384 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 390, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 553, + 676 + ], + "type": "text", + "content": "Image super-resolution. Early deep learning-based SR methods [5, 10, 22, 27] aim to improve image fidelity metrics such as PSNR and SSIM [52]. By using bicubic down-sampling to generate LQ-HQ image training pairs, various approaches have been developed to enhance the SR performance, such as dense [44], residual [69], recursive connections [22], non-local strategies [47], and attention mechanisms [8, 27, 68]. Though these advancements have significantly improved the SR performance, the real-world LQ images with complex degradations are much more challenging to enhance. While researchers have proposed to collect real-world LQ-HQ paired datasets [4, 55] to train SR models, simulating realistic LQ-HQ pairs is a more economic way [50, 65]. BSRGAN [65] randomly shuffles some basic degradation operators to synthesize LQ-HQ pairs, while RealESRGAN [50] implements high-order degradation modeling. Compared with pixel-level losses, the perceptual loss [17] and GAN loss [24, 48] are more effective in improving image perceptual quality. However, the adversarial training also leads to unnatural visual artifacts in the SR results. Many of the later works aim to reduce GAN-generated artifacts from the perspectives of frequency division [25], network optimization [70], the weighting of different losses [43, 49], and image local statistics [29]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Diffusion model based super-resolution. Recent studies have explored the use of diffusion models (DMs) [15, 37] for SR. Early approaches adjust pre-trained DMs'" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2334" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "content": "reverse transitions [20] using gradient descent, offering training-free solutions by assuming predefined image degradations. ResShift [63] trains a DM from scratch on paired LQ-HQ data. Recently, SD models pre-trained on T2I tasks have been widely used for SR tasks due to their strong image priors [2, 6, 30, 32, 36, 46, 51, 56, 57, 59]. StableSR [46] introduces a trainable encoder and takes LQ image as a condition in SD. DiffBIR [30] first applies a restoration module to mitigate degradations, then utilizes SD to enhance details. PASD [59] employs an encoder for degradation removal and introduces a pixel-aware cross-attention module to incorporate low-level and high-level image features into the SD process. SeeSR [57] enhances semantic robustness through degradation-aware tag-style prompts to guide diffusion. Considering that the multi-step diffusion process will increase the computational cost and the risk of synthesizing unfaithful contents, researchers started to develop fewer-step DM-based SR models. SinSR [51] applies consistency-preserving distillation to shorten the diffusion process of ResShift. OSEDiff [56] employs the LQ image as direct input, eliminating the random noise sampling. It adopts the VSD loss [54, 60] to distill the generative capacity from the multi-step diffusion process, offering a one-step DM-based SR solution. However, all the above-mentioned methods face a contradiction between pixel-level and semantic-level enhancement and most of them lack the flexibility to adjust the SR style for diverse user preferences." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 407, + 138, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 138, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 138, + 421 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "content": "This section first formulates the SD-based SR as a residual learning model, then introduces the dual-LoRA approach to disentangle the learning objectives of pixel-level regression and semantic-level enhancement. Finally, pixel and semantic guidances are proposed for flexible SR results generation. In our following development, we denote by " + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "inline_equation", + "content": "x_{L}" + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "inline_equation", + "content": "x_{H}" + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "content": " the LQ and HQ images, respectively, and by " + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "inline_equation", + "content": "z_{L}" + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "inline_equation", + "content": "z_{H}" + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "content": " their corresponding latent codes. Let " + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "content": " be the encoder and decoder of a well-trained variational auto-encoder (VAE), we can nearly have " + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "inline_equation", + "content": "z_{L} = E(x_{L})" + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "inline_equation", + "content": "z_{H} = E(x_{H})" + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "inline_equation", + "content": "x_{L} = D(z_{L})" + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "inline_equation", + "content": "x_{H} = D(z_{H})" + }, + { + "bbox": [ + 55, + 425, + 296, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 566, + 169, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 169, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 169, + 577 + ], + "type": "text", + "content": "3.1. Model Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "DMs such SD [37] apply a " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "-step forward process to progressively transform an initial latent code " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " into Gaussian noise " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " steps. In each timestep " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", the noisy latent code " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " is obtained by directly adding noise " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_t = \\sqrt{\\bar{\\alpha}_t} \\cdot z_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\cdot \\epsilon" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " is the cumulative parameter controlling the noise level at timestep " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". During the reverse process, multi-step SD-based SR models [2, 30, 36, 46, 57, 59] usually perform a " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "-step denoising process to gradually transform Gaussian noise " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_T" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " into the HQ latent code " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_H" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", conditioned on the LQ image " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_L" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " with a ControlNet [66] (see Fig. 2 (a)). At stage " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", the latent" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 319, + 69, + 552, + 144 + ], + "blocks": [ + { + "bbox": [ + 319, + 69, + 552, + 144 + ], + "lines": [ + { + "bbox": [ + 319, + 69, + 552, + 144 + ], + "spans": [ + { + "bbox": [ + 319, + 69, + 552, + 144 + ], + "type": "image", + "image_path": "21ba05524be1c5cd4a637aa37b3630e901b28adaf25b29f1639e41a8831cbd49.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 144, + 487, + 153 + ], + "lines": [ + { + "bbox": [ + 381, + 144, + 487, + 153 + ], + "spans": [ + { + "bbox": [ + 381, + 144, + 487, + 153 + ], + "type": "text", + "content": "(a) Multi-step DM-based SR methods" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 157, + 550, + 192 + ], + "blocks": [ + { + "bbox": [ + 319, + 157, + 550, + 192 + ], + "lines": [ + { + "bbox": [ + 319, + 157, + 550, + 192 + ], + "spans": [ + { + "bbox": [ + 319, + 157, + 550, + 192 + ], + "type": "image", + "image_path": "d12052bcddc86787f481f76a8a7c1394d95c7c81e5e521133daba8c5c641e221.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 190, + 450, + 198 + ], + "lines": [ + { + "bbox": [ + 414, + 190, + 450, + 198 + ], + "spans": [ + { + "bbox": [ + 414, + 190, + 450, + 198 + ], + "type": "text", + "content": "(b) OSEDiff" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 204, + 550, + 246 + ], + "blocks": [ + { + "bbox": [ + 319, + 204, + 550, + 246 + ], + "lines": [ + { + "bbox": [ + 319, + 204, + 550, + 246 + ], + "spans": [ + { + "bbox": [ + 319, + 204, + 550, + 246 + ], + "type": "image", + "image_path": "5ebbbb204c307be2d71598fccbd173d6060bec51e5349b7a8a0029c77a195c19.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 247, + 450, + 255 + ], + "lines": [ + { + "bbox": [ + 414, + 247, + 450, + 255 + ], + "spans": [ + { + "bbox": [ + 414, + 247, + 450, + 255 + ], + "type": "text", + "content": "(c) PiSA-SR" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "text", + "content": "Figure 2. Comparison of the pipeline of different DM-based SR methods. (a) Multi-step methods [2, 30, 36, 46, 57, 59] perform " + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "text", + "content": " denoising steps starting from Gaussian noise " + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "text", + "content": ", conditioned on the LQ image " + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "inline_equation", + "content": "x_{L}" + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "text", + "content": ". (b) OSEDiff [56] starts from LQ latent representation " + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "inline_equation", + "content": "z_{L}" + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "text", + "content": " with only one-step diffusion. (c) Our proposed PiSA-SR formulates the SD-based SR as learning the residual between the LQ latent " + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "inline_equation", + "content": "z_{L}" + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "text", + "content": " and HQ latent " + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "inline_equation", + "content": "z_{H}" + }, + { + "bbox": [ + 313, + 256, + 555, + 333 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 343, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 313, + 343, + 555, + 369 + ], + "type": "text", + "content": " can be computed from the noise estimated by SD UNet " + }, + { + "bbox": [ + 313, + 343, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,x_L,t)" + }, + { + "bbox": [ + 313, + 343, + 555, + 369 + ], + "type": "text", + "content": " and the current latent " + }, + { + "bbox": [ + 313, + 343, + 555, + 369 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 343, + 555, + 369 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 342, + 376, + 554, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 376, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 342, + 376, + 554, + 403 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {0} = f (z _ {t}, \\epsilon_ {\\theta}) = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (z _ {t} , x _ {L} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {1}", + "image_path": "ed4657e4b6b0715730edb75e4d42d44934f70623a7b37b430d09cd467ca4e581.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "content": "Then the latent " + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "content": " can be sampled from " + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "inline_equation", + "content": "p(z_{t-1} | z_t, \\hat{z}_0) = N(\\mu_t(z_t, \\hat{z}_0), \\sigma_t^2)" + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "inline_equation", + "content": "\\mu_t" + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "content": " denote the mean and variance of " + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 446, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 554, + 540 + ], + "type": "text", + "content": "However, these multi-step DM-based SR methods are computationally expensive and often yield unstable results due to the random noise sampled in the diffusion process [42, 57]. To address this issue, OSEDiff [56] starts from the LQ latent code " + }, + { + "bbox": [ + 313, + 446, + 554, + 540 + ], + "type": "inline_equation", + "content": "z_{L}" + }, + { + "bbox": [ + 313, + 446, + 554, + 540 + ], + "type": "text", + "content": " and completes the SR process with only one diffusion step (see Fig. 2 (b)), improving both the efficiency and stability. The transformation between " + }, + { + "bbox": [ + 313, + 446, + 554, + 540 + ], + "type": "inline_equation", + "content": "z_{L}" + }, + { + "bbox": [ + 313, + 446, + 554, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 446, + 554, + 540 + ], + "type": "inline_equation", + "content": "z_{H}" + }, + { + "bbox": [ + 313, + 446, + 554, + 540 + ], + "type": "text", + "content": " can be described as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 548, + 554, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 548, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 349, + 548, + 554, + 574 + ], + "type": "interline_equation", + "content": "z _ {H} = f \\left(z _ {L}, \\epsilon_ {\\theta}\\right) = \\frac {z _ {L} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(z _ {L}\\right)}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {2}", + "image_path": "7b5c1d2a390f77da356971b685ed39e5b94282c35feb13d3314fd819b784839f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": "Note that we omit the timestep " + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": " in the above formula since " + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": " can be viewed as a constant in one-step diffusion." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": "Learning the residual between the LQ and GT features has been successfully applied in deep learning-based restoration methods [64, 69]. However, DMs typically perform multi-step iterations, making applying residual learning difficult. The recently developed OSEDiff [56] employs a single diffusion step to achieve end-to-end SR training, making it feasible to adopt the residual learning strategy. Here, we formulate the SR problem as learning the residual between " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "z_{L}" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "z_{H}" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": ", as illustrated in Fig. 2 (c). We use" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2335" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "subtraction in the global residual connection because DMs are trained to denoise in the inverse diffusion process. Such a residual learning formulation helps the model focus on learning the desired high-frequency information from the HQ latent features, avoiding extracting less relevant information from the LQ latent. It can also accelerate the convergence of the model training process [12]. Furthermore, we introduce a scaling factor " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " to adjust the residual " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " added to the LQ latent, which can be expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 188, + 295, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 188, + 295, + 200 + ], + "spans": [ + { + "bbox": [ + 132, + 188, + 295, + 200 + ], + "type": "interline_equation", + "content": "z _ {H} = z _ {L} - \\lambda \\epsilon_ {\\theta} (z _ {L}). \\tag {3}", + "image_path": "599b64118e30bbb7222ee71b12b2a80f840b2b668fd05ad646b4dfa45eb9ce08.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 207, + 296, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 255 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 55, + 207, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 207, + 296, + 255 + ], + "type": "text", + "content": " is fixed as 1 during training. Users can adjust the output in the inference stage by using a smaller " + }, + { + "bbox": [ + 55, + 207, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 207, + 296, + 255 + ], + "type": "text", + "content": " (to preserve more original contents) or a higher " + }, + { + "bbox": [ + 55, + 207, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 207, + 296, + 255 + ], + "type": "text", + "content": " (to enhance details more aggressively) based on their preferences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 261, + 178, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 178, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 178, + 274 + ], + "type": "text", + "content": "3.2. Dual-LoRA Training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 279, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 506 + ], + "type": "text", + "content": "Previous SD-based SR methods mostly entangle the pixel-level and semantic-level enhancement in the training process [46, 56], making the balance of content fidelity and perpetual quality difficult. Some works sequentially perform pixel-level and semantic-level enhancement using two separate networks [30]. However, errors in the earlier stage may propagate to the next stage, limiting overall performance. Furthermore, using two separate networks increases the computational burden and memory requirements, making the approach less efficient. It has been successfully applied in customized T2I tasks using distinct LoRA modules for various generation styles [18, 38, 41]. Inspired by this, we propose a decoupled training approach by utilizing two LoRA modules under the pre-trained SD model, targeting pixel-level and semantic-level enhancement for the SR task respectively. Our approach, namely Dual-LoRA, introduces only a small number of additional parameters during training. These LoRA parameters can also be merged into the pre-trained SD model during inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 507, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 296, + 654 + ], + "type": "text", + "content": "Fig. 3 (a) illustrates the training process of our approach. We freeze the parameters of the pre-trained VAE and introduce two trainable LoRA modules to the UNet of SD. Since the LQ image is corrupted by degradations such as noise, blur, and downsampling, we optimize the pixel-level LoRA first to reduce the effect of corruptions, followed by the optimization of semantic-level LoRA. We train the pixel-level LoRA, denoted by " + }, + { + "bbox": [ + 55, + 507, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{pix}" + }, + { + "bbox": [ + 55, + 507, + 296, + 654 + ], + "type": "text", + "content": ", by the pixel-level loss function. Together with the pre-trained SD parameters, the full parameter set can be represented as " + }, + { + "bbox": [ + 55, + 507, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\theta_{pix} = \\{\\theta_{sd}, \\Delta \\theta_{pix}\\}" + }, + { + "bbox": [ + 55, + 507, + 296, + 654 + ], + "type": "text", + "content": ". The HQ latent can then be estimated by " + }, + { + "bbox": [ + 55, + 507, + 296, + 654 + ], + "type": "inline_equation", + "content": "z_H^{pix} = z_L - \\epsilon_{\\theta_{pix}}(z_L)" + }, + { + "bbox": [ + 55, + 507, + 296, + 654 + ], + "type": "text", + "content": " and decoded by the VAE decoder with " + }, + { + "bbox": [ + 55, + 507, + 296, + 654 + ], + "type": "inline_equation", + "content": "x_H^{pix} = D(z_H^{pix})" + }, + { + "bbox": [ + 55, + 507, + 296, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "For semantic-level enhancement, we train another LoRA, denoted by " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{sem}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", by the semantic-level loss function. To disentangle the pixel-level and semantic-level objectives, we fix the already trained pixel-level LoRA " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{pix}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " and combine it with the semantic-level LoRA " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{sem}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " to be" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": "trained, forming a PiSA-LoRA group (see Fig. 3 (a)). Note that only the semantic-level LoRA module within the PiSA-LoRA group is updated in this stage. This ensures that the optimization process focuses on the semantic details without much interference from the pixel-level degradations. The full parameter set during this PiSA-LoRA training stage can be expressed as " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\theta_{PiSA} = \\{\\theta_{sd},\\Delta \\theta_{pix},\\Delta \\theta_{sem}\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": ". The HQ latent after this stage is given by " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "z_H^{sem} = z_L - \\epsilon_{\\theta_{PiSA}}(z_L)" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": ", which is then decoded by the VAE decoder with " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "x_H^{sem} = D(z_H^{sem})" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 201, + 509, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 509, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 509, + 213 + ], + "type": "text", + "content": "3.3. Pixel-level and Semantic-level Losses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 220, + 555, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 364 + ], + "type": "text", + "content": "We train the pixel-level LoRA using the " + }, + { + "bbox": [ + 313, + 220, + 555, + 364 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 220, + 555, + 364 + ], + "type": "text", + "content": " loss. As shown by " + }, + { + "bbox": [ + 313, + 220, + 555, + 364 + ], + "type": "inline_equation", + "content": "D(\\epsilon_{\\theta_{pix}}(z_L))" + }, + { + "bbox": [ + 313, + 220, + 555, + 364 + ], + "type": "text", + "content": " in Fig. 4, the " + }, + { + "bbox": [ + 313, + 220, + 555, + 364 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 220, + 555, + 364 + ], + "type": "text", + "content": " loss can effectively remove degradation and enhance edges. However, it is insufficient to generate semantic-level details [24, 29, 48], resulting in smooth SR outputs. The LPIPS loss [67] can be used to activate semantic details by aligning the high-level features with a pre-trained classification VGG network [39], which is however trained on a limited set of image categories. The GAN loss [24] captures semantic information through adversarial training, encouraging the generator to produce more realistic images. However, GAN loss can be unstable in training and generate much artifacts [29]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": "Recently developed SD models [37] have shown superior performance in generating open-category images with enhanced semantic details. As a conditional generation model, SD can be used to form up an implicit classifier [14] for the posterior distribution modeling, and its gradient w.r.t. the synthesized image " + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 446, + 553, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 446, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 318, + 446, + 553, + 485 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {x _ {t}} \\log p _ {\\theta} (c \\mid x _ {t}) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t} \\mid c) - \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ = - \\left(\\epsilon_ {\\theta} \\left(x _ {t}, c, t\\right) - \\epsilon_ {\\theta} \\left(x _ {t}, t\\right)\\right) / \\sigma_ {t}, \\tag {4} \\\\ \\end{array}", + "image_path": "623dc53c7b121f44b42fd48fbb22ffe3d376ec99acd937c24f867be83ebf4cee.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": " is the text prompt, " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": " is the current timestep, and " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\sigma_{t} = \\sqrt{1 - \\bar{\\alpha}_{t}}" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": ". The SD model, parameterized by " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": ", model score functions of " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}\\log p_\\theta (x_t\\mid c)" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}\\log p_\\theta (x_t)" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": " using noise predictions [40]. By taking the expectation of the gradient in Eq. (4) over all possible " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": ", we have the CSD [14] loss, which is originally used in 3D generation tasks to optimize the posterior probability of rendered images aligning their semantic content with text prompts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 677 + ], + "type": "text", + "content": "Motivated by the functionality of CSD loss in DMs for generation, we investigate the usage of CSD loss for semantic-level enhancement in the SR task. Following [56, 60], and for the clarity of the following context, we formulate the gradient of CSD in the latent space instead of the noise domain. Specifically, the gradient of CSD loss can be formulated in Eq. (5) below (the detailed derivation can be found in the supplementary materials):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 322, + 682, + 553, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 682, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 322, + 682, + 553, + 713 + ], + "type": "interline_equation", + "content": "\\nabla \\ell_ {C S D} ^ {\\lambda_ {c f g}} = \\mathbb {E} _ {t, \\epsilon , z _ {t}, c} \\left[ w _ {t} \\left(f (z _ {t}, \\epsilon_ {r e a l}) - f (z _ {t}, \\epsilon_ {r e a l} ^ {\\lambda_ {c f g}})\\right) \\frac {\\partial z _ {H} ^ {s e m}}{\\partial \\theta_ {P i S A}} \\right], \\tag {5}", + "image_path": "8f74e1fde638529ba0e34b949396f86467872776d1494914720bccf4397f32da.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2336" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 544, + 228 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 544, + 228 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 544, + 228 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 544, + 228 + ], + "type": "image", + "image_path": "92a9c6f205b7f5bdf5817974d47b3fdc250ff0c02ef5bd1438db8d12d2d67032.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 232, + 375, + 244 + ], + "lines": [ + { + "bbox": [ + 234, + 232, + 375, + 244 + ], + "spans": [ + { + "bbox": [ + 234, + 232, + 375, + 244 + ], + "type": "text", + "content": "(a) Training process of the PiSA-SR" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 247, + 544, + 358 + ], + "blocks": [ + { + "bbox": [ + 69, + 247, + 544, + 358 + ], + "lines": [ + { + "bbox": [ + 69, + 247, + 544, + 358 + ], + "spans": [ + { + "bbox": [ + 69, + 247, + 544, + 358 + ], + "type": "image", + "image_path": "ceec4e76b76f126dfa4803000686b836e72b86bc51c0bfb975de412a3e0692e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 358, + 369, + 369 + ], + "lines": [ + { + "bbox": [ + 237, + 358, + 369, + 369 + ], + "spans": [ + { + "bbox": [ + 237, + 358, + 369, + 369 + ], + "type": "text", + "content": "(b) Inference process of PiSA-SR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 421, + 294, + 555 + ], + "blocks": [ + { + "bbox": [ + 54, + 369, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 54, + 369, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 555, + 404 + ], + "type": "text", + "content": "Figure 3. The (a) training and (b) inference procedures of PiSA-SR. During the training process, two LoRA modules are respectively optimized for pixel-level and semantic-level enhancement. During the inference stage, users can use the default setting to reconstruct the HQ image in one-step diffusion or adjust " + }, + { + "bbox": [ + 54, + 369, + 555, + 404 + ], + "type": "inline_equation", + "content": "\\lambda_{pix}" + }, + { + "bbox": [ + 54, + 369, + 555, + 404 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 369, + 555, + 404 + ], + "type": "inline_equation", + "content": "\\lambda_{sem}" + }, + { + "bbox": [ + 54, + 369, + 555, + 404 + ], + "type": "text", + "content": " to control the strengths of pixel-level and semantic-level enhancement." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 421, + 294, + 555 + ], + "lines": [ + { + "bbox": [ + 58, + 421, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 58, + 421, + 294, + 555 + ], + "type": "image", + "image_path": "607e5f283805224f3efbfca2cf926a683d5a9f2476049342273a5ba62b4ff2b3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "type": "text", + "content": "Figure 4. The model outputs with pixel-wise and semantic-level losses for a given LQ image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "text", + "content": "where the gradient expectation is calculated considering all diffusion timesteps " + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "inline_equation", + "content": "t = 1,\\dots,T" + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "text", + "content": " , the noise is sampled from " + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I),z_{t}" + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "text", + "content": " is obtained by " + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "inline_equation", + "content": "z_{t} = \\sqrt{\\bar{\\alpha}_{t}}\\cdot z_{H}^{sem} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\cdot \\epsilon" + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "text", + "content": " is the text prompt extracted from " + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "inline_equation", + "content": "z_H^{sem}" + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "text", + "content": " is the function as Eq. (1), " + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "text", + "content": " is the timestep-dependent scalar weight, and " + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "inline_equation", + "content": "\\epsilon_{real}^{\\lambda_{cfg}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "text", + "content": " denotes the pre-trained SD output with the CFG term " + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "inline_equation", + "content": "\\epsilon_{real}^{cls}" + }, + { + "bbox": [ + 55, + 594, + 295, + 681 + ], + "type": "text", + "content": " same as Eq. (4) [14, 62]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 684, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 684, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 68, + 684, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\epsilon_ {r e a l} ^ {\\lambda_ {c f g}} = \\epsilon_ {r e a l} (z _ {t}, t) + \\underbrace {\\lambda_ {c f g} (\\epsilon_ {r e a l} (z _ {t} , t , c) - \\epsilon_ {r e a l} (z _ {t} , t))}. \\qquad (6)", + "image_path": "38db472e4f10c8bc8e42301d134c11d18d890f5e68b6cfed47b237a57dbaa38d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 424, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 473 + ], + "type": "text", + "content": "Note that the VSD loss [54, 60], which also aims to align the distributions of enhanced and natural images within the latent space, has been verified effective for SR by OSEDiff [56]. The gradient of VSD loss can be written as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 322, + 487, + 554, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 487, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 322, + 487, + 554, + 518 + ], + "type": "interline_equation", + "content": "\\nabla \\ell_ {V S D} ^ {\\lambda_ {c f g}} = \\mathbb {E} _ {t, \\epsilon , z _ {t}, c} \\left[ w _ {t} \\left(f (z _ {t}, \\epsilon_ {f a k e}) - f (z _ {t}, \\epsilon_ {r e a l} ^ {\\lambda_ {c f g}})\\right) \\frac {\\partial z _ {H} ^ {s e m}}{\\partial \\theta_ {P i S A}} \\right], \\tag {7}", + "image_path": "b39038be197e6e289cbbcf28ee15870b2474a413b4d840ef079ec6bd245f11dc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{fake}}" + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "text", + "content": " denotes the output of fine-tuned SD aligned with the distribution of synthetic images. The VSD loss with " + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{cfg}" + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "text", + "content": " can be divided into two components: the VSD loss with " + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{cfg} = 0.0" + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{VSD}^{\\lambda_{cfg} = 0.0}" + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "text", + "content": ", and the CSD loss. Coupling with the LPIPS loss, we visualize the semantic-level LoRA optimization results of the two components of VSD in Fig. 4. We see that the CSD loss with the normal guidance scale of " + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{cfg} = 7.5" + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "text", + "content": " contributes more significantly to semantic enhancement. In contrast, the VSD loss with " + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{cfg} = 0.0" + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "text", + "content": " weakens the semantic details. More importantly, the component " + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{VSD}^{\\lambda_{cfg} = 0.0}" + }, + { + "bbox": [ + 313, + 518, + 555, + 714 + ], + "type": "text", + "content": " requires bi-level optimization, leading to intensive memory usage and unstable training [31]. In comparison, the CSD loss is free of bi-level optimization, reducing significantly memory usage and improving training stability. Therefore, we integrate LPIPS and CSD losses for the semantic-level LoRA optimization." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2337" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 237, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 237, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 237, + 83 + ], + "type": "text", + "content": "3.4. The Inference Process of PiSA-SR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 207 + ], + "type": "text", + "content": "Fig. 3(b) shows the inference process of PiSA-SR. In the default inference setting (represented by the green dash-dot line), only the PiSA-LoRA, which merges the pixel-level and semantic-level LoRA modules, is used to process the input together with the pre-trained SD model, achieving state-of-the-art SR performance in just one-step diffusion (see our experiments in Section 4.3). To achieve flexible SR with diverse user preferences, we extend the strategy in Eq. (3) by introducing a pair of pixel and semantic guidance factors, denoted by " + }, + { + "bbox": [ + 55, + 87, + 296, + 207 + ], + "type": "inline_equation", + "content": "\\lambda_{pix}" + }, + { + "bbox": [ + 55, + 87, + 296, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 87, + 296, + 207 + ], + "type": "inline_equation", + "content": "\\lambda_{sem}" + }, + { + "bbox": [ + 55, + 87, + 296, + 207 + ], + "type": "text", + "content": ", to control the SR results:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 216, + 294, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 216, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 58, + 216, + 294, + 240 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\theta} \\left(z _ {L}\\right) = \\lambda_ {p i x} \\epsilon_ {\\theta_ {p i x}} \\left(z _ {L}\\right) + \\lambda_ {s e m} \\left(\\epsilon_ {\\theta_ {P s S A}} \\left(z _ {L}\\right) - \\epsilon_ {\\theta_ {p i x}} \\left(z _ {L}\\right)\\right). \\tag {8}", + "image_path": "4eaa0b07b6857088fdfa55f500018e8970011c6ea782a106251369d9271b2cae.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "text", + "content": "In Eq. (8), " + }, + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_{pix}}(z_L)" + }, + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "text", + "content": " is the output with only pixel-level LoRA, while " + }, + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_{PiSA}}(z_L)" + }, + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "text", + "content": " is the output with both pixel and semantic level enhancement. Inspired by CFG [14], the difference between them, i.e., " + }, + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_{PiSA}}(z_L) - \\epsilon_{\\theta_{pix}}(z_L)" + }, + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "text", + "content": ", can well separate pixel-level degradations from semantic details (see the red box in Fig. 4). Therefore, as illustrated in Fig. 3(b) (represented by the black dashed line), by adjusting " + }, + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\lambda_{pix}" + }, + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\lambda_{sem}" + }, + { + "bbox": [ + 55, + 240, + 296, + 361 + ], + "type": "text", + "content": ", we can control the contributions of pixel-level LoRA and semantic-level LoRA, and consequently deliver the SR results with different styles, as shown in Fig. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 370, + 137, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 137, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 137, + 384 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 387, + 182, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 182, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 182, + 399 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "text", + "content": "Training settings. We train PiSA-SR upon SD 2.1-base for the " + }, + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "text", + "content": " SR task. The two trainable LoRA modules are applied to the weights of all convolution and MLP layers. Both modules are initialized using a Gaussian distribution with a rank of 4. Following SeeSR [57] and OSEDiff [56], we employ LSDIR [26] and the first 10K images from FFHQ [19] as the training data. We generate paired training data using the degradation pipeline from RealESRGAN [50]. The batch size is 16 and the training patch size is " + }, + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 402, + 296, + 545 + ], + "type": "text", + "content": ". The Adam optimizer [23] is used with a learning rate 5e-5. The pixel-level and semantic-level LoRA modules undergo 4K and 8.5K training iterations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "text", + "content": "Compared methods. We compare PiSA-SR with leading multi-step DM-based methods StableSR [46], ResShift [63], DiffBIR [30], PASD [59], and SeeSR [57], one-step DM-based methods SinSR [51] and OSEDiff [56], and GAN-based SR methods RealESRGAN [50], BSRGAN [65] and LDL [29]. All comparative results are obtained using officially released codes or models. Due to space constraints, comparisons with GAN-based methods are presented in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Test datasets and evaluation metrics. Following previous arts [46, 56, 57], we evaluate the competing methods using synthetic and real-world test data. The synthetic dataset comprises 3000 images cropped from DIV2K [1] with size " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", degraded using RealESRGAN" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 95, + 553, + 216 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 91 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 91 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 91 + ], + "type": "text", + "content": "Table 1. Results of PiSA-SR with different pixel-semantic guidance scales on the RealSR test dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 95, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 317, + 95, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 95, + 553, + 216 + ], + "type": "table", + "html": "
λpixλsemPSNR↑LIPIS↓CLIPIQA↑MUSIQ↑
0.01.025.960.34260.412946.45
0.21.026.480.30420.486854.05
0.51.026.750.26460.570563.82
0.81.026.180.26120.629268.95
1.01.025.500.26720.670270.15
1.21.024.760.27230.674670.33
1.51.023.740.27690.630569.23
1.00.026.920.30180.322749.62
1.00.226.950.27840.359153.64
1.00.526.770.24760.432258.76
1.00.826.200.24650.580666.33
1.01.025.500.26720.670270.15
1.01.224.590.30000.701571.60
1.01.523.080.35410.683571.76
", + "image_path": "e2fd0c5b79a32616af4a4425e0a9358cee20993e6120b6e63485966c249f793b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 230, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 230, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 312, + 230, + 555, + 386 + ], + "type": "text", + "content": "degradation pipeline [50]. The real-world data are center-cropped from RealSR [4] and DrealSR [55] datasets with size " + }, + { + "bbox": [ + 312, + 230, + 555, + 386 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 312, + 230, + 555, + 386 + ], + "type": "text", + "content": " for LQ images and " + }, + { + "bbox": [ + 312, + 230, + 555, + 386 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 312, + 230, + 555, + 386 + ], + "type": "text", + "content": " for HQ images. PSNR and SSIM [52], computed on the Y channel in the YCbCr space, are used to measure the fidelity of SR results; LPIPS [67] and DISTS [9], computed in the RGB space, are used to assess the perceptual quality of SR results; FID [13] evaluates the distance of distributions between GT and restored images; NIQE [33], CLIPIQA [45], MUSIQ [21] and MANIQA [58] evaluate the image quality without reference images. Additionally, a user study is conducted to validate the effectiveness of the proposed adjustable SR in the supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 397, + 481, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 481, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 481, + 410 + ], + "type": "text", + "content": "4.2. Experiments on Adjustable SR" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": "We conduct experiments by fixing one guidance scale (" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{pix}" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{sem}" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": ") at 1 and varying the other to observe how the reconstructed images change. PSNR, LPIPS, CLIPIQA, and MUSIQ are used to evaluate performance. PSNR measures pixel-level fidelity; LPIPS evaluates the image perceptual quality with GT as reference; CLIPIQA and MUSIQ measure the image quality without using a reference. The results are shown in Table 1. First, we observe that increasing the pixel-level scale " + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{pix}" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": " leads to a continuous improvement in the no-reference metrics CLIPIQA and MUSIQ. This is because increasing " + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{pix}" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": " can remove image degradation and enhance edge, which is favorable to these two metrics. However, the reference-based metrics PSNR and LPIPS exhibit a rise-and-fall pattern. PSNR peaks at " + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{pix} = 0.5" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": ", indicating the best pixel-level fidelity. LPIPS reaches its best value at " + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{pix} = 0.8" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": " (0.2612), suggesting that at this point the restored image is perceptually most similar to the GT with rich details. Further increasing " + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{pix}" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": " reduces both PSNR and LPIPS scores. Second, increasing the semantic-level scale " + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{sem}" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": " also results in a continuous improvement in CLIPIQA and MUSIQ, but with a higher upper bound than pixel-level adjustments. This is because increasing " + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{sem}" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": " can synthesize more semantic-level details (see Fig. 1). However, PSNR decreases as " + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{sem}" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": " increases, while LPIPS first improves and peaks at " + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{sem} = 0.8" + }, + { + "bbox": [ + 312, + 415, + 555, + 715 + ], + "type": "text", + "content": " (0.2465), and then" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2338" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 96, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Quantitative comparison among the state-of-the-art DM-based SR methods on synthetic and real-world test datasets. 'S' denotes the number of diffusion steps. The best and the second-best results are highlighted in red and blue, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 96, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 58, + 96, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 58, + 96, + 553, + 319 + ], + "type": "table", + "html": "
DatasetsMethodsPSNR↑SSIM↑LPIPS↓DISTS↓FID↓NIQE↓CLIPQA↑MUSIQ↑MANIQA↑
DIV2KResShift-S1524.690.61750.33740.221536.016.820.608960.920.5450
StableSR-S20023.310.57280.31290.213824.674.760.668265.630.6188
DiffBIR-S5023.670.56530.35410.212930.934.710.665265.660.6204
PASD-S2023.140.54890.36070.221929.324.400.671168.830.6484
SeeSR-S5023.710.60450.32070.196725.834.820.685768.490.6239
SinSR-S124.430.60120.32620.206635.456.020.649962.800.5395
OSEDiff-S123.720.61080.29410.197626.324.710.668367.970.6148
PiSA-SR-S123.870.60580.28230.193425.074.550.692769.680.6400
RealSRResShift-S1526.310.74110.34890.2498142.817.270.545058.100.5305
StableSR-S20024.690.70520.30910.2167127.205.760.619565.420.6211
DiffBIR-S5024.880.66730.35670.2290124.565.630.641264.660.6231
PASD-S2025.220.68090.33920.2259123.085.180.650268.740.6461
SeeSR-S5025.330.72730.29850.2213125.665.380.659469.370.6439
SinSR-S126.300.73540.32120.2346137.056.310.620460.410.5389
OSEDiff-S125.150.73410.29210.2128123.505.650.669369.090.6339
PiSA-SR-S125.500.74170.26720.2044124.095.500.670270.150.6560
DrealSRResShift-S1528.450.76320.40730.2700175.928.280.525949.860.4573
StableSR-S20028.040.74600.33540.2287147.036.510.617158.500.5602
DiffBIR-S5026.840.66600.44460.2706167.386.020.629260.680.5902
PASD-S2027.480.70510.38540.2535157.365.570.671464.550.6130
SeeSR-S5028.260.76980.31970.2306149.866.520.667264.840.6026
SinSR-S128.410.74950.37410.2488177.057.020.636755.340.4898
OSEDiff-S127.920.78350.29680.2165135.296.490.696364.650.5899
PiSA-SR-S128.310.78040.29600.2169130.616.200.697066.110.6156
", + "image_path": "d0b1567fb7c1d01c4c998105c7c5446209d6208495b59b3165c67be8b17c2d97.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 329, + 294, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 294, + 389 + ], + "type": "text", + "content": "declines. This is because excessive semantic details may induce changes in image content and reduce pixel-level fidelity. Meanwhile, the over-enhanced images may exhibit many details different from the GT, making the reference-based LPIPS index deteriorate." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 389, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 295, + 594 + ], + "type": "text", + "content": "Fig. 1 illustrates the visual comparisons of different pixel-semantic scales. More comparisons can be found in the supplementary materials. From bottom to top, as the pixel-level scale " + }, + { + "bbox": [ + 56, + 389, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\lambda_{pix}" + }, + { + "bbox": [ + 56, + 389, + 295, + 594 + ], + "type": "text", + "content": " increases, the model progressively removes degradation from the input LQ image, resulting in clearer and sharper visual output. However, a too-high " + }, + { + "bbox": [ + 56, + 389, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\lambda_{pix}" + }, + { + "bbox": [ + 56, + 389, + 295, + 594 + ], + "type": "text", + "content": " can lead to loss of details, and diminish the effectiveness of semantic improvements, as seen in the top-right corner of Fig. 1. From left to right, increasing the semantic-level scale " + }, + { + "bbox": [ + 56, + 389, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\lambda_{sem}" + }, + { + "bbox": [ + 56, + 389, + 295, + 594 + ], + "type": "text", + "content": " enhances the semantic richness of the image, making the details of the wrinkles, beard, and hair of Albert Einstein more pronounced. However, an excessive semantic level such as " + }, + { + "bbox": [ + 56, + 389, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\lambda_{sem} = 1.5" + }, + { + "bbox": [ + 56, + 389, + 295, + 594 + ], + "type": "text", + "content": " introduces unnatural artifacts, making the SR result less realistic. With the flexibility of PiSASR, users can customize their preferences, either preserving more fidelity or emphasizing richer semantic enhancements, depending on their specific use case." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 601, + 246, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 246, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 246, + 614 + ], + "type": "text", + "content": "4.3. Comparisons with State-of-the-Arts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 616, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 666 + ], + "type": "text", + "content": "As shown in Eq. (8), letting " + }, + { + "bbox": [ + 55, + 616, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{pix} = 1" + }, + { + "bbox": [ + 55, + 616, + 295, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 616, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\lambda_{sem} = 1" + }, + { + "bbox": [ + 55, + 616, + 295, + 666 + ], + "type": "text", + "content": ", we have the default version of PiSA-SR, which can be performed in just one-step diffusion. Here, we compare this default version of PiSA-SR with the competing methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": "Quantitative comparisons. Table 2 compares the performance of our default PiSA-SR model with other DM-based Real-SR methods. The following findings can be made. ResShift [63] and its distilled version SinSR [51] show lim" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 329, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 544 + ], + "type": "text", + "content": "ited performance in no-reference metrics, implying poor perceptual quality. PASD [59] and SeeSR [57] leverage the pre-trained SD as their base model, and extract additional high-level information [71] to improve the perceptual quality of images. Their no-reference metric scores, such as MUSIQ and CLIPIQA, are highly improved. However, the multi-step inference makes them less efficient. In addition, their relatively low LPIPS and DISTS scores suggest their less faithful reconstruction to the GT. OSEDiff [56] distills the generation capacity of multi-step DM into just one-step diffusion by the VSD loss [54, 60]. While effective and efficient, its no-reference metrics are limited. Our proposed PiSA-SR also exhibits high efficiency by requiring only one inference step, while showing impressive pixel-level fidelity and semantic-level perceptual quality. It outperforms other methods not only in reference-based metrics such as LPIPS and DISTS but also in no-reference metrics such as CLIP-IQA, MUSIQ, and MANIQA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Qualitative comparisons. We present visual comparisons in Fig. 5. ResShift and SinSR fail to generate richer textures and details because of their limited generation capacity. StableSR and DiffBIR can generate finer details (e.g., more realistic penguin feathers in the second example), due to the leveraged strong SD prior. PASD and SeeSR incorporate high-level information guidance, resulting in more comprehensive details. However, the inherent randomness in PASD and SeeSR may lead to over-enhanced (e.g., unnatural penguin feathers generated by SeeSR) or over-smoothed details (e.g., the blurred texture of the wooden stool generated by SeeSR). OSEDiff can generate more consistent results in both examples. However, the restored details can be limited. In contrast, the proposed PiSA-SR delivers high-quality, re" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2339" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 195, + 194 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 195, + 194 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 195, + 194 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 195, + 194 + ], + "type": "image", + "image_path": "83e03c8fc3d1a51dc0c8c80547632ab54021d8500c334b73108bd595dc10a4a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 195, + 70, + 309, + 194 + ], + "blocks": [ + { + "bbox": [ + 195, + 70, + 309, + 194 + ], + "lines": [ + { + "bbox": [ + 195, + 70, + 309, + 194 + ], + "spans": [ + { + "bbox": [ + 195, + 70, + 309, + 194 + ], + "type": "image", + "image_path": "8c94bb7aeb1ca51df5be79f5d4df3861f8e4419d41172956c38f4a290d045125.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 517, + 194 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 517, + 194 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 517, + 194 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 517, + 194 + ], + "type": "image", + "image_path": "8821bc87bfc13e4e8b9d850820813b08cca674b5887ae7eecf6034a4ec839543.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 80, + 195, + 194, + 319 + ], + "blocks": [ + { + "bbox": [ + 80, + 195, + 194, + 319 + ], + "lines": [ + { + "bbox": [ + 80, + 195, + 194, + 319 + ], + "spans": [ + { + "bbox": [ + 80, + 195, + 194, + 319 + ], + "type": "image", + "image_path": "8ab7a306b01b1a7ba098f4159268641d8d75cf459971510514f8f9d395220843.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 323, + 485, + 333 + ], + "lines": [ + { + "bbox": [ + 124, + 323, + 485, + 333 + ], + "spans": [ + { + "bbox": [ + 124, + 323, + 485, + 333 + ], + "type": "text", + "content": "Figure 5. Visual comparisons of different DM-based SR methods. Please zoom in for a better view." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 195, + 195, + 310, + 318 + ], + "blocks": [ + { + "bbox": [ + 195, + 195, + 310, + 318 + ], + "lines": [ + { + "bbox": [ + 195, + 195, + 310, + 318 + ], + "spans": [ + { + "bbox": [ + 195, + 195, + 310, + 318 + ], + "type": "image", + "image_path": "6537849122bd587a018edc04a77c6cc65c4c3c0f4587192d3892eef8d5b2b73c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 195, + 517, + 318 + ], + "blocks": [ + { + "bbox": [ + 310, + 195, + 517, + 318 + ], + "lines": [ + { + "bbox": [ + 310, + 195, + 517, + 318 + ], + "spans": [ + { + "bbox": [ + 310, + 195, + 517, + 318 + ], + "type": "image", + "image_path": "15f6fa77e3e01fcdfce6f529044557eda47a85d4d439d6140062f7156a52b17a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 348, + 555, + 388 + ], + "blocks": [ + { + "bbox": [ + 149, + 335, + 459, + 346 + ], + "lines": [ + { + "bbox": [ + 149, + 335, + 459, + 346 + ], + "spans": [ + { + "bbox": [ + 149, + 335, + 459, + 346 + ], + "type": "text", + "content": "Table 3. The inference time and the number of parameters of DM-based SR methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 348, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 58, + 348, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 58, + 348, + 555, + 388 + ], + "type": "table", + "html": "
StableSRResShiftDiffBIRPASDSeeSRSinSROSEDiffPiSA-SR-def.PiSA-SR-adj.
Inference Steps200155020501112
Inference time(s)/Image10.030.762.722.804.300.130.120.090.13
#Params(B)1.560.181.682.312.510.181.771.301.30
", + "image_path": "efd6af792f58936aa5d84c8e55e81607cae63649f6139f61a1ccec13cb0dd4bd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 400, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 460 + ], + "type": "text", + "content": "alistic SR results. With the dual-LoRA design, the restored structures align well with the input LQ image, providing more accurate fidelity information and generating more natural and richer details. More visualization comparisons can be found in supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 460, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 700 + ], + "type": "text", + "content": "Complexity comparisons. We compare the number of parameters and inference time of competing DM-based SR models in Table 3, where PiSA-SR-def. and PiSA-SR-adj. represent its default setting and adjustable setting, respectively. Inference time is measured on the " + }, + { + "bbox": [ + 55, + 460, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 460, + 295, + 700 + ], + "type": "text", + "content": " SR task with " + }, + { + "bbox": [ + 55, + 460, + 295, + 700 + ], + "type": "inline_equation", + "content": "128\\times 128" + }, + { + "bbox": [ + 55, + 460, + 295, + 700 + ], + "type": "text", + "content": " LQ images using a single NVIDIA A100 80G GPU. ResShift has fewer parameters because it is trained from scratch without using the pre-trained SD model. SinSR inherits the parameters from its parental ResShift model. Among the SD-based SR methods, PiSA-SR-def. has the fewest parameters and achieves the fastest inference time. Unlike OSEDiff, PiSA-SR-def. does not require a semantic extractor (e.g., RAM [71]) during inference. Instead, it distills the semantic information into the semantic-level LoRA with CSD. PiSA-SR-adj., which requires two diffusion steps to compute outputs from pixellevel and semantic-level LoRA modules, incurs a slightly longer inference time over the default setting. However, this small increase in time offers users the advantage of generating SR outputs tailored to their specific needs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Ablation study. We performed a series of ablation stud" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 400, + 555, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 436 + ], + "type": "text", + "content": "ies for a better understanding of the dual-LoRA training approach, and the roles of the two LoRA modules. The details can be found in the supplementary materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 445, + 467, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 467, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 467, + 457 + ], + "type": "text", + "content": "5. Conclusion and Limitation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 462, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 462, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 312, + 462, + 555, + 617 + ], + "type": "text", + "content": "In this work, we presented PiSA-SR, a novel SR framework that disentangles pixel-level and semantic-level objectives by learning two LoRA modules upon the pre-trained SD model. By decoupling the learning process into two distinct spaces, PiSA-SR effectively balanced pixel-wise regression and perceptual quality. The commonly used " + }, + { + "bbox": [ + 312, + 462, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 312, + 462, + 555, + 617 + ], + "type": "text", + "content": " loss was used to optimize pixel-level LoRA, while the LPIPS and CSD losses were employed to optimize the semantic-level LoRA. PiSA-SR demonstrated impressive performance in both effectiveness and efficiency. Furthermore, it offered high flexibility during inference through adjustable guidance scales, allowing users to customize the SR results without model re-training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "While PiSA-SR provided a flexible solution for adjusting pixel-level fidelity and semantic-level perception, it incurs a slight inference time increment over its default version. In addition, we use a single pixel-level LoRA to handle various degradations in the LQ, which may be insufficient for heavy degradations. In the future, we will explore refining separate LoRA spaces for different image degradations and the possibility of image-adaptive scale factors." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2340" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. NTIRE 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "text", + "content": "[2] Yang Ai, Xiaogiang Zhou, Huaibo Huang, Xiaotian Han, Zhengyu Chen, Quanzeng You, and Hongxia Yang. Dreamclear: High-capacity real-world image restoration with privacy-safe dataset curation. arXiv preprint arXiv:2410.18666, 2024.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 225 + ], + "type": "text", + "content": "[3] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6228-6237, 2018. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 296, + 281 + ], + "type": "text", + "content": "[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 316 + ], + "type": "text", + "content": "[5] Dong Chao, Loy Chen Change, He Kaiming, and Tang Xiaou. Learning a deep convolutional network for image super-resolution. In ECCV, pages 184-199, 2014. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 318, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 295, + 361 + ], + "type": "text", + "content": "[6] Bin Chen, Gehui Li, Rongyuan Wu, Xindong Zhang, Jie Chen, Jian Zhang, and Lei Zhang. Adversarial diffusion compression for real-world image super-resolution. arXiv preprint arXiv:2411.13383, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 364, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 295, + 396 + ], + "type": "text", + "content": "[7] Du Chen, Tianhe Wu, Kede Ma, and Lei Zhang. Toward generalized image quality assessment: Relaxing the perfect reference quality assumption, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 398, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 398, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 398, + 295, + 452 + ], + "type": "text", + "content": "[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 454, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 295, + 497 + ], + "type": "text", + "content": "[9] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 499, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 499, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 57, + 499, + 295, + 543 + ], + "type": "text", + "content": "[10] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 544, + 295, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 587 + ], + "type": "text", + "content": "[11] Jingwen He, Chao Dong, Yihao Liu, and Yu Qiao. Interactive multi-dimension modulation for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9363-9379, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 4, 5, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 155, + 553, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 155, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 317, + 155, + 553, + 198 + ], + "type": "text", + "content": "[17] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "type": "text", + "content": "[18] Maxwell Jones, Sheng-Yu Wang, Nupur Kumari, David Bau, and Jun-Yan Zhu. Customizing text-to-image models with a single image pair. arXiv preprint arXiv:2405.01536, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 236, + 553, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 236, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 317, + 236, + 553, + 279 + ], + "type": "text", + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 281, + 553, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 281, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 281, + 553, + 324 + ], + "type": "text", + "content": "[20] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "text", + "content": "[21] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5148-5157, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 374, + 553, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 374, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 553, + 417 + ], + "type": "text", + "content": "[22] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 420, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 452 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 455, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 553, + 529 + ], + "type": "text", + "content": "[24] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 533, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 553, + 578 + ], + "type": "text", + "content": "[25] Wenbo Li, Kun Zhou, Lu Qi, Liying Lu, and Jiangbo Lu. Best-Buddy GANs for highly detailed image superresolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1412-1420, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 580, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 655 + ], + "type": "text", + "content": "[26] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Lsdir: A large scale dataset for image restoration. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1775-1787, 2023. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[27] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2341" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[28] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 173 + ], + "type": "text", + "content": "[29] Jie Liang, Hui Zeng, and Lei Zhang. Details or Artifacts: A locally discriminative learning approach to realistic image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5657-5666, 2022. 1, 2, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 294, + 230 + ], + "type": "text", + "content": "[30] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Ben Fei, Bo Dai, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 233, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 294, + 286 + ], + "type": "text", + "content": "[31] Zhiyuan Ma, Yuxiang Wei, Yabin Zhang, Xiangyu Zhu, Zhen Lei, and Lei Zhang. Scaledreamer: Scalable text-to-3d synthesis with asynchronous score distillation. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 290, + 294, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 290, + 294, + 333 + ], + "spans": [ + { + "bbox": [ + 56, + 290, + 294, + 333 + ], + "type": "text", + "content": "[32] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Conditional diffusion distillation. arXiv preprint arXiv:2310.01407, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 336, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 336, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 336, + 294, + 369 + ], + "type": "text", + "content": "[33] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 424 + ], + "type": "text", + "content": "[34] Chong Mou, Yanze Wu, Xintao Wang, Chao Dong, Jian Zhang, and Ying Shan. Metric learning based interactive modulation for real-world super-resolution. In European Conference on Computer Vision, pages 723-740. Springer, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "type": "text", + "content": "[35] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 474, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 294, + 517 + ], + "type": "text", + "content": "[36] Yunpeng Qu, Kun Yuan, Kai Zhao, Qizhi Xie, Jinhua Hao, Ming Sun, and Chao Zhou. Xpsr: Cross-modal priors for diffusion-based image super-resolution. arXiv preprint arXiv:2403.05049, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 577, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 294, + 632 + ], + "type": "text", + "content": "[38] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In European Conference on Computer Vision, pages 422-438. Springer, 2025. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "type": "text", + "content": "[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "type": "text", + "content": "[41] Nick Stracke, Stefan Andreas Baumann, Joshua M Susskind, Miguel Angel Bautista, and Björn Ommer. Ctrloralter: Conditional loradapter for efficient 0-shot control & altering of t2i models. arXiv preprint arXiv:2405.07913, 2024. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 554, + 161 + ], + "type": "text", + "content": "[42] Lingchen Sun, Rongyuan Wu, Zhengqiang Zhang, Hongwei Yong, and Lei Zhang. Improving the stability of diffusion models for content consistent super-resolution. arXiv e-prints, pages arXiv-2401, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 162, + 554, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 162, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 315, + 162, + 554, + 215 + ], + "type": "text", + "content": "[43] Lingchen Sun, Jie Liang, Shuaizheng Liu, Hongwei Yong, and Lei Zhang. Perception-distortion balanced superresolution: A multi-objective optimization perspective. IEEE Transactions on Image Processing, 33:4444-4458, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 217, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 217, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 315, + 217, + 554, + 259 + ], + "type": "text", + "content": "[44] Tong Tong, Gen Li, Xiejie Liu, and Qinquan Gao. Image super-resolution using dense skip connections. In Proceedings of the IEEE international conference on computer vision, pages 4799-4807, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 261, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 261, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 315, + 261, + 554, + 304 + ], + "type": "text", + "content": "[45] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 305, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 305, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 315, + 305, + 554, + 348 + ], + "type": "text", + "content": "[46] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 350, + 554, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 350, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 315, + 350, + 554, + 413 + ], + "type": "text", + "content": "[47] Risheng Wang, Tao Lei, Wenzheng Zhou, Qi Wang, Hongying Meng, and Asoke K Nandi. Lightweight non-local network for image super-resolution. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1625-1629. IEEE, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 415, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 415, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 315, + 415, + 554, + 470 + ], + "type": "text", + "content": "[48] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. ESRGAN: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 471, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 471, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 315, + 471, + 554, + 525 + ], + "type": "text", + "content": "[49] Xintao Wang, Ke Yu, Chao Dong, Xiaou Tang, and Chen Change Loy. Deep network interpolation for continuous imagery effect transition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1701, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 526, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 526, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 315, + 526, + 554, + 580 + ], + "type": "text", + "content": "[50] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 582, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 582, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 315, + 582, + 554, + 646 + ], + "type": "text", + "content": "[51] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 647, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 554, + 689 + ], + "type": "text", + "content": "[52] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 1, 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 691, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 691, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 691, + 554, + 712 + ], + "type": "text", + "content": "[53] Zhihao Wang, Jian Chen, and Steven C. H. Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2342" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[54] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 207 + ], + "type": "text", + "content": "[55] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII 16, pages 101-117. Springer, 2020. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 252 + ], + "type": "text", + "content": "[56] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. arXiv preprint arXiv:2406.08177, 2024. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "type": "text", + "content": "[57] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456-25467, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 376 + ], + "type": "text", + "content": "[58] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "type": "text", + "content": "[59] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image superresolution and personalized stylization. arXiv preprint arXiv:2308.14469, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "type": "text", + "content": "[60] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 3, 4, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 544 + ], + "type": "text", + "content": "[61] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. arXiv preprint arXiv:2401.13627, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "type": "text", + "content": "[62] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and XIAOJUAN QI. Text-to-3d with classifier score distillation. In The Twelfth International Conference on Learning Representations, 2024. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 634 + ], + "type": "text", + "content": "[63] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Resshift: Efficient diffusion model for image super-resolution by residual shifting. arXiv preprint arXiv:2307.12348, 2023. 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 679 + ], + "type": "text", + "content": "[64] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[65] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 374 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision, pages 4791-4800, 2021. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "type": "text", + "content": "[68] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. arXiv preprint arXiv:2203.06697, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[69] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[70] Yuehan Zhang, Bo Ji, Jia Hao, and Angela Yao. Perception-distortion balanced ADMM optimization for single-image super-resolution. In European Conference on Computer Vision, pages 108-125. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "text", + "content": "[71] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 7, 8" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2343" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/3af07a07-d377-440f-9ac8-47f29938e643_content_list.json b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/3af07a07-d377-440f-9ac8-47f29938e643_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ebc0e70f516fea7ec21055d02068ad7b2a6d5317 --- /dev/null +++ b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/3af07a07-d377-440f-9ac8-47f29938e643_content_list.json @@ -0,0 +1,2024 @@ +[ + { + "type": "text", + "text": "PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes", + "text_level": 1, + "bbox": [ + 140, + 130, + 854, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bin Tan $^{1}$ Rui Yu $^{2}$ Yujun Shen $^{1}$ Nan Xue $^{\\dagger,1}$", + "bbox": [ + 300, + 179, + 694, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Ant Group $^{2}$ University of Louisville", + "bbox": [ + 341, + 214, + 656, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b827c1d16c0df3e53f13f20d1d328eaa178a00e3a8e2a42aad9c6a6c185dbaae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 273, + 184, + 505 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3dddd3f4530dbc79fd5a16b7b9fbb9a27311b9783876e02963049084f9d3489.jpg", + "image_caption": [ + "Input Images" + ], + "image_footnote": [], + "bbox": [ + 91, + 506, + 184, + 570 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0451cdae4cbd7334c7243a32e757d3d445c4ee3ec651a55faca10f7db8636092.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 273, + 375, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79aa4b4f4e46379333f2b65e99ba502bed86d79f234d3637c59a33aef65efb4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 396, + 364, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4916d6cbdba2112540472625f4b4183a3e4ef568bbb2e0857eb68ba39c726d95.jpg", + "image_caption": [ + "PlanarRecon", + "Figure 1. We introduce PlanarSplatting, a fast and accurate optimization-based planar reconstruction method for indoor scenes. We show our planar reconstruction results on the ScanNetV2 [2] dataset achieved in 3 minutes. Compared to prior art PlanarRecon [37] and AirPlanes [35], our PlanarSplatting reconstructs more complete and detailed 3D planes." + ], + "image_footnote": [], + "bbox": [ + 212, + 481, + 364, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1781cecd8eb6918046d57c32821de614ad47b67aa2e9518f3034e52851b0e9a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 277, + 549, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e95653c421b1bbb220c8ce549fef0c28c7d86d8bbcd139a8c1464ff140bc80be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 396, + 542, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/468b392657c3a7cb0b2a5dae055a6b06df7e94753aca59826605cccd7f5ff261.jpg", + "image_caption": [ + "AirPlanes" + ], + "image_footnote": [], + "bbox": [ + 390, + 481, + 540, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/352fddffc5b6279c78776f2f43f6aadfac74c4d1c15e1a3e5f01be4955fcabc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 275, + 732, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74bee36ad248c719a6aa81152b3bff713ad696a2e5053bdbb878d5ec3626f41a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 396, + 720, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2541fda366f5214e314e57546836a2d3c5aded25a626ff05d9f762adb5b64aba.jpg", + "image_caption": [ + "Ours (3min)" + ], + "image_footnote": [], + "bbox": [ + 570, + 481, + 720, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a4123ad3d6a2e9ea0e457cba4140e1753e6f8b69f01afdb72c448e61339d145.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 275, + 906, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9765116973bfe08b642c5744ecb8719655ee8b6df897d50875c60d2e9a7a2384.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 396, + 900, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2f33164e35e29dcb882b5948ad8c26a7f9dcfb2317e8e72ef6760342e42cd16.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 750, + 481, + 900, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 669, + 326, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents PlanarSplatting, an ultra-fast and accurate surface reconstruction approach for multi-view indoor images. We take the 3D planes as the main objective due to their compactness and structural expressiveness in indoor scenes, and develop an explicit optimization framework that learns to fit the expected surface of indoor scenes by splatting the 3D planes into 2.5D depth and normal maps. As our PlanarSplatting operates directly on the 3D plane primitives, it eliminates the dependencies on 2D/3D plane detection and plane matching/tracking for planar surface reconstruction. Furthermore, with the essential", + "bbox": [ + 88, + 702, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "merits of plane-based representation coupled with CUDA-based implementation of planar splatting functions, Planar-Splatting reconstructs an indoor scene in 3 minutes while having significantly better geometric accuracy. Thanks to our ultra-fast reconstruction speed, the largest quantitative evaluation on the ScanNet and ScanNet++ datasets over hundreds of scenes clearly demonstrated the advantages of our method. We believe that our accurate and ultrafast planar surface reconstruction method will be applied in the structured data curation for surface reconstruction in the future. The code of our CUDA implementation will be publicly available.", + "bbox": [ + 511, + 670, + 908, + 852 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author.", + "bbox": [ + 114, + 886, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1190", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb3d1f06b998f755890b6cac6dfe7e9370a3513152a72585733a531311d3cd2d.jpg", + "image_caption": [ + "Figure 2. Novel View Synthesis: Seamless integration of our PlanarSplatting with recent Gaussian Splatting methods (e.g., 3DGS [13] and 2DGS [11]) achieves indoor-scene rendering with improved performance and significant optimization time reduction simultaneously." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 263, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/57d8d611aa27ba0110bf11462e32e60d5fa3707ccf20e762ed0f2d739edc2ef4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 88, + 418, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4351adb86ce454a84d5604a730584fc91e2048110c3ce9379b9f12b3aabafe57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 88, + 573, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/45fb1fec7ac4f6e2bce35f9247eb24272ca72a120fdcbd4962f410603be3952f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 88, + 730, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a211d1496809b85aed57ff390bad18cb58749e4691c062529158b66765332b05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 88, + 885, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 335, + 222, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We humans are long-term immersed in structured indoor scenes, ranging from bedrooms to offices. This fact has ignited a wealth of studies for reconstructing the indoor environment with various 3D structures such as lines [9, 20, 39], planes [12, 16, 18, 19, 33, 37, 42] and blocks [7, 23]. Among them, the 3D plane is the most common representation because of its simplicity and completeness in describing physical surfaces, thus motivating us to study the problem of 3D reconstruction for indoor scenes using 3D planes, i.e., planar 3D reconstruction.", + "bbox": [ + 89, + 364, + 483, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Planar 3D reconstruction has been extensively studied for years as a model fitting problem, in which a 3D scene geometry (e.g., point clouds, or meshes) was assumed to be known and the main goal is fitting the scene in a set of 3D planes [25, 30, 38]. Recently, the paradigm has been gradually simplified in image-based solutions in single-view and multi-view 3D reconstruction, eliminating the acquisition of known 3D scene geometry. To make the problem trackable, existing methods were extensively based on the image-level characterization of 2D/3D planes. That is to say, those methods have to detect 3D planes for each input image, match or track planes in across viewpoints, and finally reconstruct and merge 3D planes as the compact 3D representation of indoor scenes. Recent PlanarRecon [37] attempted to address these problems end-to-end by learning a 3D volume from monocular videos. Then, 3D planes can be detected, tracked, and fused in a consistent 3D space.", + "bbox": [ + 89, + 518, + 483, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-based planar 3D reconstruction mostly followed keypoint-based 3D reconstruction pipelines and treated planes as a special kind of visual features. However, because of the essential difference between the local point features and regional plane masks in image space, we argue, the existing approaches did not fully leverage the advantages of planar representations. Some evidences could be observed from the results of PlanarRecon [37], which are", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "usually very coarse and lose many details of the scene.", + "bbox": [ + 511, + 338, + 874, + 352 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we are going to address the issues remained in image-based planar 3D reconstruction, aiming at obtaining a complete, structural, and compact indoor scene reconstruction from multi-view images. Our main idea is approximating the indoor scenes with a collection of solid 3D planar primitives from multi-view input images, directly optimizing them to have consistent 3D planes, but eliminating any suboptimal precomputing of 2D plane primitives (e.g., plane masks). We introduce PlanarSplatting that explicitly optimizes rectangular plane primitives in 3D space by differentiably splatting them into 2.5D depth and normal maps. Thanks to our well-designed plane splatting function, PlanarSplatting effectively leverage monocular geometry cues from modern foundational models [3, 10, 41] for accurate plane optimization. As shown in Fig. 1, the final high-quality planar surface can be reconstructed by simply merging adjacent 3D plane primitives.", + "bbox": [ + 509, + 354, + 906, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our PlanarSplatting is directly designed on the 3D plane primitives and efficiently implemented with CUDA, which accordingly induces a seamlessly integration with recent Gaussian Splatting (GS) methods for high-quality indoor novel view synthesis (NVS). As shown in Fig. 2, benefiting from our fast and accurate scene reconstruction (within 3 minutes), GS-based methods can be well-initialized and optimized without densification, leading to better rendering results and significantly less training time. It demonstrates the strong potential to promote the unity of surface reconstruction and novel view synthesis for indoor scenes.", + "bbox": [ + 509, + 612, + 908, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the experiments, our PlanarSplatting shows its powerful ability for accurate indoor planar surface reconstruction for hundreds of scenes from two real-world indoor datasets, i.e. ScanNetV2 [2] and ScanNet++[40]. Furthermore, we show that with the combination of our PlanarSplatting and GS-based methods (e.g., 3DGS [13] and 2DGS [11]), we can effectively improve the rendering quality with less training time and fewer points.", + "bbox": [ + 511, + 780, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1191", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Indoor Planar Reconstruction. Reconstructing indoor scenes with 3D plane primitives has been studied for a long time [4-6, 29, 36]. Traditionally, it is usually achieved by deducing and fitting the plane geometry directly from 3D data (e.g., point clouds [24, 25, 30, 31] and line clouds [15]), or from 2D single-view images with strict scene assumptions (e.g., the Manhattan World constrain) [16, 22, 26] which seriously limited their application. In recent years, some learnable-based methods are proposed to formulate this problem as single-view 3D plane segmentation [17, 18, 28, 32, 42, 43] and cross-view plane instance matching [1, 12, 33]. Although impressive results have been achieved, these works are hard to extend to multiple views. PlanarRecon [37] was the first end-to-end work proposed to deal with holistic indoor plane recovery by learning a plane-related 3D volume from posed RGB videos. Then, 3D planes can be extracted, tracked, and fused from the learned 3D volume incrementally. Most recently, AirPlanes [35] developed a two-step method that first reconstructed a dense scene mesh and then learned consistent 3D plane embeddings from 2D plane embeddings for plane extraction from the dense mesh. Note that the aforementioned learning-based methods require 2D/3D plane annotations as their supervision, leading to performance bottlenecks due to the difficulty of obtaining a large scale of plane annotations. In contrast, we proposed PlanarSplatting to reconstruct accurate and complete indoor planar surfaces by directly optimizing a set of solid 3D plane primitives from posed multi-view images without any extra plane detection or matching operations. Benefiting from our differentiable planar primitive rendering, PlanarSplatting directly leverages monocular depth/normal cues from modern foundation models for optimization without plane annotations. Concurrent work, AlphaTablets [8], explores planar representations, which involves learning planar tablets equipped with alpha maps for shape modeling and pseudo-mesh construction for rendering. Our PlanarSplatting distinguishes itself from AlphaTablets by learning plane radii, providing a more explicit parameterization of 3D planes. As a result, we achieve a much more efficient rasterization scheme at the level of minutes while AlphaTablets takes about 2 hours per scene.", + "bbox": [ + 89, + 114, + 485, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Primitive-based Scene Representation. Optimizing explicit primitives such as points [11, 13, 34, 44], volumes [21], and superquadric [23] to represent the 3D scene has been studied for a long time. The core of these methods is to design a differentiable rendering process to optimize the attributes of primitives by gradient descent. Among them, a typical paradigm is to render images from primitives with splatting techniques which is realized with radial basis functions defined on the primitive (e.g., the Gaussian function). Inspired by these works, we proposed", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PlanarSplatting to optimize solid 3D plane primitives in a differentiable rendering manner with a novel shape-aware plane splating function to better fit the scene geometry. Benefiting from our efficient CUDA implementation, PlanarSplatting can reconstruct the accurate indoor planar scene within 3 minutes and empower recent Gaussian-based works (e.g., 3DGS [13]) to further improve the rendering quality of novel views as shown in Fig. 2.", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed PlanarSplatting", + "text_level": 1, + "bbox": [ + 511, + 224, + 794, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 3, given a set of posed multi-view images, our PlanarSplatting can reconstruct the indoor planar scene from them by optimizing a set of learnable 3D planar primitives (Sec. 3.1). With the proposed Differentiable Planar Primitive Rendering (Sec. 3.2), we can explicitly learn these planar primitives from a coarse initialization to accurately recover the scene geometry. These optimized 3D planar primitives are then merged to achieve the 3D plane instances for the final planar reconstruction.", + "bbox": [ + 511, + 250, + 908, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Learnable Planar Scene Representation", + "text_level": 1, + "bbox": [ + 511, + 393, + 854, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since our key idea is directly optimizing explicit 3D primitives for planar scene reconstruction, we first present the representation of our learnable 3D planar primitives and then introduce how we initialize the scene with these planar primitives.", + "bbox": [ + 511, + 416, + 906, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learnable Planar Primitive. As shown in Fig. 4, we formulate a planar primitive $\\pi$ as a 3D rectangle which is equipped with several learnable parameters including the plane center $\\mathbf{p}_{\\pi} \\in \\mathbb{R}^3$ , the plane rotation $\\mathbf{q}_{\\pi} \\in \\mathbb{R}^4$ (in quaternion representation) and the plane radii $\\mathbf{r}_{\\pi}$ . Specifically, to improve the optimization flexibility of plane shape, we use the design of double direction plane radii (Double Radii) for $\\mathbf{r}_{\\pi}$ as:", + "bbox": [ + 511, + 491, + 908, + 613 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {\\pi} = \\left\\{r _ {\\pi} ^ {x +}, r _ {\\pi} ^ {x -}, r _ {\\pi} ^ {y +}, r _ {\\pi} ^ {y -} \\right\\} \\in \\mathbb {R} _ {+} ^ {4}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 621, + 906, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $r_{\\pi}^{x+}, r_{\\pi}^{x-}, r_{\\pi}^{y+}, r_{\\pi}^{y-}$ are the radii defined on the positive/negative direction of the X-axis/Y-axis of the rectangle as shown in Fig. 4. For the sake of description, we further define the positive direction of the X-axis and Y-axis of the 3D planar primitive as two orthogonal unit vectors $\\mathbf{v}_{\\pi}^{x}, \\mathbf{v}_{\\pi}^{y} \\in \\mathbb{R}^{3}$ which can be calculated as:", + "bbox": [ + 511, + 647, + 908, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\pi} ^ {x} = \\mathbf {R} (\\mathbf {q} _ {\\pi}) [ 1, 0, 0 ] ^ {\\top}, \\quad \\mathbf {v} _ {\\pi} ^ {y} = \\mathbf {R} (\\mathbf {q} _ {\\pi}) [ 0, 1, 0 ] ^ {\\top}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 744, + 906, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{R}(\\mathbf{q}_{\\pi}) \\in \\mathbb{R}^{3 \\times 3}$ means the rotation matrix of the quaternion $\\mathbf{q}_{\\pi}$ . Similarly, the normal of the planar primitive $\\mathbf{n}_{\\pi} \\in \\mathbb{R}^3$ can be calculated as:", + "bbox": [ + 511, + 773, + 906, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} _ {\\pi} = \\mathbf {R} (\\mathbf {q} _ {\\pi}) [ 0, 0, 1 ] ^ {\\top}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 825, + 906, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With these learnable parameters, the 3D planar primitive can be moved to align with the potential scene surface and deformed to fit the surface shape during optimization.", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1192", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/447b99c3bbe6bc3176621af2c5497560f491541fce2e96b126a2b29bb12331e1.jpg", + "image_caption": [ + "Figure 3. Illustration of our proposed PlanarSplatting. Given a set of posed multi-view images of indoor scenes, our method renders depth and normal maps from 3D plane primitives. Then, with the supervision of monocular cues, these 3D plane primitives are gradually optimized to recover the scene geometry and finally merged to get the planar reconstruction result." + ], + "image_footnote": [], + "bbox": [ + 125, + 106, + 874, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/35b7220572873baffb5a9a2f549ec7e342306f7efcbe2faba124ceba94a1acd5.jpg", + "image_caption": [ + "Figure 4. Representation of 3D plane primitive with learnable shape parameters including plane center, radii, and rotation." + ], + "image_footnote": [], + "bbox": [ + 200, + 446, + 370, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Initialization. We use monocular depth from recent foundation models [10] to fast initialize our 3D planar primitives at the beginning of optimization. Specifically, we use depths from Metric3Dv2 [10] to get a very coarse scene geometry. Then we randomly sample 2,000 points on the coarse mesh to achieve the plane centers of our 3D planar primitives. The initial radii of each primitive $\\pi$ is set to $0.5Dist(\\pi)$ . Here, $Dist(\\pi)$ means the distance closest to $\\pi$ to its neighbors. We use the normal direction on the coarse mesh to initialize the plane rotation. At the bottom of Fig. 3, we show an example of our initial planar primitives. With such a coarse initialization, our PlanarSplatting can finally reconstruct the accurate and complete scene surface.", + "bbox": [ + 89, + 580, + 483, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Differentiable Planar Primitive Rendering", + "text_level": 1, + "bbox": [ + 89, + 786, + 452, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimize the learnable planar primitives $\\Pi = \\{\\pi_i\\}_{i=1}^K$ , we introduce the differentiable planar primitive rendering with a carefully designed plane splating function which enables the planar primitives to accurately fit the scene geometry with the supervision only from 2D multi-view images.", + "bbox": [ + 89, + 806, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b5dfbdc26ede9d699243bd82c88bb907ec43a8b2147a12fa0fb486aa0b484ef8.jpg", + "image_caption": [ + "Figure 5. Illustration of the proposed plane splatting function. Naive Gaussian Splatting can not effectively approximate the boundary of our rectangular plane primitive (shown in black dashed border). In contrast, our proposed plane splatting function can approximate the boundary of the rectangle as the number of iterations increases, allowing our 3D planar primitives to better fit the surface of the scene." + ], + "image_footnote": [], + "bbox": [ + 532, + 445, + 879, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ray-to-Plane Intersection. To project 3D planar primitives to the 2D image space, we first calculate the intersections between planar primitives and the rays emitted from image pixels. Specifically, given a ray $\\mathbf{r} = \\{\\mathbf{o},\\mathbf{d}\\}$ starting from the camera center $\\mathbf{o}\\in \\mathbb{R}^3$ with direction $\\mathbf{d}\\in \\mathbb{R}^3$ , its intersection $\\mathbf{x}_{\\pi}^{\\mathbf{r}}\\in \\mathbb{R}^3$ to one planar primitive $\\pi$ is:", + "bbox": [ + 511, + 662, + 906, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} = \\mathbf {o} + \\frac {\\left(\\mathbf {p} _ {\\pi} - \\mathbf {o}\\right) \\cdot \\mathbf {n} _ {\\pi}}{\\mathbf {d} \\cdot \\mathbf {n} _ {\\pi}} \\mathbf {d}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 766, + 906, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_{\\pi}$ and $\\mathbf{n}_{\\pi}$ are the center and the normal of the planar primitive $\\pi$ .", + "bbox": [ + 511, + 809, + 906, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Plane Splatting Function. After achieving the ray-to-plane intersection $\\mathbf{x}_{\\pi}^{\\mathbf{r}}$ , we calculate its splatting weight with our plane splatting function which will be used for rendering.", + "bbox": [ + 511, + 839, + 905, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A vanilla selection of the splatting function is the", + "bbox": [ + 531, + 885, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1193", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "anisotropic Gaussian function:", + "bbox": [ + 89, + 90, + 295, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}, \\pi\\right) = \\exp \\left(- \\frac {1}{2} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right) ^ {\\top} \\Sigma^ {- 1} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 116, + 482, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_{\\pi}$ is the center of the planar primitive $\\pi$ . The covariance matrix $\\Sigma$ can be calculated like [11, 13]. However, as shown in Fig. 5 and Fig. 6, the Gaussian-based splatting function will make ambiguous boundaries of our rectangular planar primitive leading to the degeneration of reconstruction quality.", + "bbox": [ + 89, + 159, + 483, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, we propose to calculate the splating weight with a novel rectangle-based plane splating function. To a given intersection $\\mathbf{x}_{\\pi}^{\\mathbf{r}}$ between ray $\\mathbf{r}$ and planar primitive $\\pi$ , we first calculate its projection distance $\\mathcal{P}_X, \\mathcal{P}_Y \\in \\mathbb{R}$ to the X-axis and Y-axis of the planar primitive as:", + "bbox": [ + 89, + 250, + 483, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {X} = \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right) \\cdot \\mathbf {v} _ {\\pi} ^ {x}, \\quad \\mathcal {P} _ {Y} = \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right) \\cdot \\mathbf {v} _ {\\pi} ^ {y}. \\quad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 335, + 482, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we calculate the splating weight along the X-axis of the plane $\\pi$ as:", + "bbox": [ + 89, + 364, + 483, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {X} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}\\right) = \\left\\{ \\begin{array}{l l} 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {x +} - | \\mathcal {P} _ {X} |\\right)\\right), & \\text {i f} \\mathcal {P} _ {X} > 0, \\\\ 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {x -} - | \\mathcal {P} _ {X} |\\right)\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 405, + 482, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ is the Sigmoid function and $\\lambda$ is the hyperparameter to control the splattering weight. Similarly, we continue to calculate the splattering weight along the Y-axis of the plane $\\pi$ as:", + "bbox": [ + 89, + 455, + 483, + 516 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {Y} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}\\right) = \\left\\{ \\begin{array}{l l} 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {y +} - | \\mathcal {P} _ {Y} |\\right)\\right), & \\text {i f} \\mathcal {P} _ {Y} > 0, \\\\ 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {y -} - | \\mathcal {P} _ {Y} |\\right)\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 527, + 482, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r_{\\pi}^{x + },r_{\\pi}^{x - },r_{\\pi}^{y + },r_{\\pi}^{y - }$ are the radii parameters of the plane $\\pi$ . At last, the final splatting weight can be calculated as:", + "bbox": [ + 89, + 578, + 483, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}\\right) = \\left\\{ \\begin{array}{l l} w _ {X}, & \\text {i f} w _ {X} < w _ {Y}, \\\\ w _ {Y}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 619, + 482, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Fig. 5, we show that with the increment of hyperparameter $\\lambda$ , our plane splating function gradually approximates the shape of the rectangular plane primitive. In practice, we increase the value of $\\lambda$ with an exponential function during optimization up to the maximum value of 300 as:", + "bbox": [ + 89, + 664, + 483, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda = \\min \\left(2 0 e ^ {\\left(- (1 - 0. 0 0 1 * i t e)\\right)}, 3 0 0\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 750, + 482, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where ite means the iteration number during optimization.", + "bbox": [ + 89, + 779, + 478, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Blending Composition. For all ray-to-plane intersections, we filter them with splating weight lower than 0.0001 and then sort the remaining intersections according to their depth from near to far. Then, $M$ nearest intersections of each ray are selected for rendering ( $M = 30$ in this paper). Denote the selected intersections of a ray $\\mathbf{r}$ as $P^{\\mathbf{r}} = \\{\\mathbf{x}_{\\pi_{\\tau(j)}}^{\\mathbf{r}}\\}_{j=1}^{M}$ . Here, $\\tau(j)$ indicates the index of plane", + "bbox": [ + 89, + 795, + 483, + 904 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d609f4ee2501591c5c13063024999d195c9498e0f1d5caa7aa8de2cd9d516373.jpg", + "image_caption": [ + "(a) w/ GS Splitting" + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 629, + 150 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d900e5a7255e421cb3ed95003d65a2540f5bbdfc7cb0a5620067d8653d14481f.jpg", + "image_caption": [ + "(b) w/ Plane Splatting", + "Figure 6. Reconstruction comparison with different splatting functions. 'w/' means 'with'." + ], + "image_footnote": [], + "bbox": [ + 651, + 89, + 764, + 150 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0fc6fd89f60d5c112dce3c6820bbc905110efc178afc07d7fe501693f4f81889.jpg", + "image_caption": [ + "(c) GT Mesh" + ], + "image_footnote": [], + "bbox": [ + 785, + 88, + 898, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "among all planar primitives. At last, we render the depth and normal map of a certain image $\\mathbf{I}$ as:", + "bbox": [ + 511, + 215, + 903, + 246 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) = \\sum_ {j = 1} ^ {M} T _ {j} w \\left(\\mathbf {x} _ {\\pi_ {\\tau (j)}} ^ {\\mathbf {r}}\\right) t _ {j}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 257, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {N} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) = \\sum_ {j = 1} ^ {M} T _ {j} w \\left(\\mathbf {x} _ {\\pi_ {\\tau (j)}} ^ {\\mathbf {r}}\\right) \\mathbf {n} _ {\\pi_ {\\tau (j)}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 316, + 903, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 366, + 558, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {j} = \\prod_ {i = 1} ^ {j - 1} (1 - w \\left(\\mathbf {x} _ {\\pi_ {\\tau (i)}} ^ {\\mathbf {r}}\\right)). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 378, + 903, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $t_j$ is the depth of the intersection and $\\mathbf{n}_{\\pi_{\\tau(j)}}$ is the normal of planar primitive $\\pi_{\\tau(j)}$ . To supervise the rendered depth and normal map, we use the pretrained model of Metric3Dv2 [10] to predict the depth map $\\mathbf{D}_{\\mathrm{pre}}$ and normal map $\\mathbf{N}_{\\mathrm{pre}}$ to serve as pseudo labels. Besides, we also test the normal map predicted by Omnidata [3]. Finally, the render loss can be calculated as:", + "bbox": [ + 511, + 426, + 905, + 531 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e n d e r}} ^ {\\Pi} = \\alpha_ {1} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\| 1 - \\mathbf {N} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) ^ {\\top} \\mathbf {N} _ {\\text {p r e}} (\\mathbf {r})) \\| _ {1} + \\\\ \\alpha_ {1} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\| \\mathbf {N} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) - \\mathbf {N} _ {\\text {p r e}} (\\mathbf {r})) \\| _ {1} + \\tag {14} \\\\ \\alpha_ {2} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\| (\\mathbf {D} _ {\\mathrm {r e n d e r}} ^ {\\Pi} (\\mathbf {r})) - \\mathbf {D} _ {\\mathrm {p r e}} (\\mathbf {r}) \\| _ {1}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 541, + 903, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{1} = 2.0$ , $\\alpha_{2} = 2.0$ , $\\mathbf{r}$ is the ray/pixel emitted from image $\\mathbf{I}$ .", + "bbox": [ + 511, + 655, + 903, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Optimization", + "text_level": 1, + "bbox": [ + 513, + 696, + 653, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function. To each scene, we optimize our Planar-Splatting with the Adam optimizer [14] for 5,000 iterations on the ScanNetV2 [2] dataset and 10,000 iterations on the ScanNet++ [40] dataset using the loss as shown in Eq. (14).", + "bbox": [ + 511, + 719, + 905, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Plane Splitting. During optimization, we introduce a splitting operation on planes according to the gradients of their radii to better fit the scene geometry. If the average radii gradients on the X-axis $(r^{x+}$ and $r^{x-})$ of the plane are greater than 0.2, we split the plane along the Y-axis. Similarly, we split the planes along the X-axis, if their radii gradients on the Y-axis $(r^{y+}$ and $r^{y-})$ are larger than 0.2. We conduct the splitting operation every 1,000 iterations.", + "bbox": [ + 511, + 780, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1194", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/56602deac7da37e8139cc4358e94a4b3bd8bcd627f136fc40c5965cd9927ef8e.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of planar reconstruction results on the ScanNetV2 [2] dataset. 'P. Ann.' means using 2D/3D plane annotations in training stages." + ], + "table_footnote": [], + "table_body": "
MethodP. Ann.GeometrySegmentationPlanar
Chamfer ↓F-score ↑VOI ↓RI ↑SC ↑Fidelity ↓Acc ↓Chamfer ↓
PlanarRecon [37]✓9.8943.473.2010.9190.40518.8616.2117.53
AirPlanes [35]✓5.3064.922.2680.9570.5688.767.988.37
2DGS [11] + RANSACX14.1531.334.0300.9240.25740.0214.7727.40
Metric3Dv2 [10] + RANSACX5.9365.452.7440.9270.4889.2911.3610.33
SR [27] + RANSACX5.4065.452.5070.9460.5159.4210.139.78
Ours@OmnidataX4.8368.852.5020.9480.5326.6411.769.20
Ours@Metric3Dv2X4.2373.762.3100.9580.5605.7210.658.19
", + "bbox": [ + 99, + 121, + 897, + 242 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bac4030c307e0b6bba8c4fb58eb80f13a79d2f5c827b46b2e4f139a2831cb6e2.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of planar reconstruction results on the ScanNet++ [40] dataset. 'P. Ann.' means using 2D/3D plane annotations in training stages." + ], + "table_footnote": [], + "table_body": "
MethodP. Ann.GeometrySegmentationPlanar
Chamfer ↓F-score ↑VOI ↓RI ↑SC ↑Fidelity ↓Acc ↓Chamfer ↓
PlanarRecon [37]✓17.8531.103.5420.9190.36734.4419.3526.90
AirPlanes [35]✓13.7532.582.8590.9410.47028.1612.5820.37
2DGS [11] + RANSACX20.3926.464.4560.9270.24155.9016.8836.39
Metric3Dv2 [10] + RANSACX11.1441.543.0210.9300.44127.5010.8119.16
SR [27] + RANSACX13.1535.933.0130.9380.44230.2511.6220.94
Ours@OmnidataX9.3347.042.7720.9460.52317.2412.2614.75
Ours@Metric3Dv2X9.0150.682.7820.9440.52117.1712.3614.77
", + "bbox": [ + 98, + 287, + 897, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Plane Merge. After optimization, we greedily merge the adjacent 3D plane primitives with normal angle error lower than $25^{\\circ}$ and offset distance error lower than $0.1m$ . Here, offset means the projection distance from the scene center to the plane surface.", + "bbox": [ + 89, + 422, + 482, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CUDA Implementation. For fast optimization of solid 3D planar primitives, we implement the forward and backward process of our Differentiable Planar Primitive Rendering with CUDA, which enables our PlanarSplatting to reconstruct one scene within 3 minutes. We will release the code of our CUDA implementation for solid 3D planar primitives optimization after publication.", + "bbox": [ + 89, + 500, + 483, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 623, + 223, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset, Metrics and Baselines", + "text_level": 1, + "bbox": [ + 89, + 648, + 362, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our PlanarSplatting on two large indoor datasets including ScanNetV2 [2] and ScanNet++ [40] which provide posed images. On the ScanNetV2 dataset, we use the test split according to AirPlanes [35] which includes 100 scenes with ground truth 3D plane annotations provided by [18]. On the ScanNet++ dataset, we randomly select 30 scenes for evaluation and extract 3D plane annotations from the ground truth meshes like [18].", + "bbox": [ + 89, + 672, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. According to PlanarRecon [37], we evaluate the geometry quality of all reconstructed 3D planes with Chamfer Distance and F-score. Following AirPlanes [35], we also evaluate the reconstruction quality of Top-20 largest planes from the ground truth and use the metrics including Planar Fidelity, Planar Accuracy, and Planar Chamfer. To evaluate 3D plane segmentation, we use", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the metrics including Variation of Information (VOI), Rand Index (RI), and Segmentation Covering (SC) like [18, 37].", + "bbox": [ + 511, + 422, + 906, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Since our method is purely built upon geometry cues, we mainly compare our PlanarSplatting with those geometry-based methods including SR+RANSAC [27], Metric3Dv2+RANSAC [10] and 2DGS+RANSAC [11]. These baselines build dense scene meshes from multi-view images at first and then extract 3D planes with the RANSAC algorithm from the reconstructed meshes. We use the RANSAC implementation provided by AirPlanes [35] for all these geometry-based baselines. Besides, we also compare our PlanarSplatting with some plane annotation based methods that use 2D/3D plane labels/priors in their training stage including PlanarRecon [37] and AirPlanes [35], and report their results for reference.", + "bbox": [ + 511, + 455, + 908, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with Baselines", + "text_level": 1, + "bbox": [ + 511, + 665, + 766, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. We first evaluate our PlanarSplatting on the ScanNetV2 [2] dataset. Specifically, we evaluate two versions of our PlanarSplatting, utilizing normal maps from different pretrained models for supervision: Metric3Dv2 [10] and Omnidata [3], respectively. As shown in Tab. 1, our method achieves the best geometry performance in the metric of Chamfer and F-score compared to all baselines. To the results of Segmentation and Planar, our method outperforms geometry-based baselines in most metrics and is comparable to the state-of-the-art AirPlanes which uses plane embeddings trained on the same ScanNetV2 dataset. In Tab. 2, we show the results on the ScanNet++ [40] dataset. Benefiting from our novel design for differentiable plane rendering and", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1195", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0ee8655e0fcd8dbd7e212abbbe46f584c8dfa0127e044df27e2adac3fc79a426.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 92, + 300, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/087c91d570fb06d8e10fc2d319e95eb4f2b9b5723a70934bcb8e3946112ac1bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 98, + 488, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/892d7ded0fdf2f3d18ec68b91a7330b9f365816a231e449bf600e68677b3ec5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 98, + 674, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da5de82a644b7150ff18daa56e0511f75ae9d0726e8bac4c267911ceb7178d4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 101, + 857, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4097cac7631077f509a2caa59c63170033411d36b9859551fbc664496aa43be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 196, + 308, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f3e4491a28c4e4b0d5190aa4816425a298cccad82b1ecca8548b3405017158d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 195, + 491, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7c4c1acfe3e66c27dd96655c5d682bcc44b16aa1979fb331ced419ae71dc75d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 195, + 673, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6578c113348bbfda366dd36c9d4e9f5604b5318858afedc87e3ae402696d8048.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 196, + 856, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7035b4089af551553830c3fe32e725dfcf28f9fb46405490618e6670b64afb61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 132, + 325, + 308, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2212230165801d77165153c0cfc6d97b1165a0932d68e2e939bd106736becfab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 325, + 488, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86a40c7a3b4d65f485daa94a0525111123ff34d3335f27934c44cad584a03cdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 321, + 669, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55ef8f1c13195e3561c5e8ce0ad07ab908815af9e058669b6aa66fb4543f5fb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 325, + 856, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/867d41b776217bfa3a818d9b59086e370c1f89e8d0735f82cb395f24c69333c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 420, + 308, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/062632401fc894e0bdd10083f67c0615f7aa9ed1c7853df05392c083d8d6706c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 420, + 491, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b69d97ab7aa638401571b550c41a3caedf16d20446824f28652ab87f89f2e0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 420, + 673, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f64063875bec07baae356f33420c6b621ee6436333f04a8ee543ed2c61fe8682.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 420, + 856, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1916fe74eeb04f44110a79719014c03925d0a996dda6753e1605faf9b226e3af.jpg", + "image_caption": [ + "(a) PlanarRecon [37]" + ], + "image_footnote": [], + "bbox": [ + 140, + 532, + 303, + 608 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a453f1e1cc145e922161c0554e42a454937035287caf033fb21fe3f83da8edef.jpg", + "image_caption": [ + "(b) AirPlanes [35]" + ], + "image_footnote": [], + "bbox": [ + 320, + 531, + 482, + 614 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1249e5a5b2cf1181a6b7449c1648c21409bc4849354ccf63f956d508358845aa.jpg", + "image_caption": [ + "(c) Ours" + ], + "image_footnote": [], + "bbox": [ + 504, + 532, + 666, + 607 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94530101150dcfc2422cf01a4ed422e2492d771d4de9df59440eff52e2840fe9.jpg", + "image_caption": [ + "(d) Ground Truth", + "Figure 7. Qualitative comparison on the ScanNetV2 [2] (rows 1-4) and ScanNet++ [40] (last row) datasets." + ], + "image_footnote": [], + "bbox": [ + 681, + 532, + 846, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "optimization via monocular cues provided by modern foundation models, our PlanarSplatting still achieves the best Geometry performance among all baselines. Furthermore, our method outperforms both geometry-based and plane annotation based methods in most metrics of Segmentation and Planar, demonstrating the robustness and superiority of our proposed PlanarSplatting.", + "bbox": [ + 89, + 659, + 482, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. In Fig. 7, we show the comparisons of PlanarRecon [37], AirPlanes [35] and our PlanarSplatting with normal supervision of Omnidata [3]. PlanarRecon reconstructs coarse 3D planes from input multi-view images while our PlanarSplatting accurately reconstructs scene geometry via directly optimizing 3D plane primitives in the whole 3D space. With the help of a plane embedding model, AirPlanes achieves more semantic plane segmentation from", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "co-plane regions. However, as shown in the zoom-in results in Fig. 7, AirPlanes loses many geometry details in the scene. In contrast, our PlanarSplatting successfully reconstructs detailed structures such as the chairs and the legs of the bed, resulting in both accuracy and completeness in results. These comparisons effectively indicate the superiority of our method.", + "bbox": [ + 511, + 659, + 903, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 772, + 678, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We verify the design of our PlanarSplatting on 10 scenes randomly selected from the ScanNetV2 [2] dataset. We optimize 3D plane primitives with 5,000 iterations with the normal supervision of Omnidata [3] in default.", + "bbox": [ + 511, + 794, + 903, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Plane Initialization. We first assess the sensitivity of the plane initialization strategy of our PlanarSplatting. We ablate the used Metric3Dv2 [10] initialization to the sphere", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1196", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2ecd1d1c9145758ca6f8b209284d2e4b656175ab07081826be09bf1aea0dc025.jpg", + "table_caption": [ + "Table 3. Ablation studies of the proposed PlanarSplatting on the ScanNetV2 dataset. 'w/' means 'with' and 'w/o' means 'without'." + ], + "table_footnote": [], + "table_body": "
Chamfer ↓F-score ↑VOI ↓
w/ Sphere Init.8.3957.533.073
w/ Sphere Init. (30K)4.5869.912.522
w/o Double Radii4.7370.252.491
w/o Plane Splitting4.6970.302.477
Ours (Full)4.6671.302.473
", + "bbox": [ + 96, + 121, + 478, + 204 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/70fcc33d63c13fd50df316f206f38d6de7286f30ecb0d0e2def89ead1330624a.jpg", + "table_caption": [ + "Table 4. Quantitative comparison of novel view synthesis on the ScanNetV2 [2] dataset. 'Plane' means time for our plane optimization. 'GS' means time for 2D/3D Gaussian optimization. #P' means the average number of Gaussian points." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓Avg. Time (min)#P
PlaneGSTotal
3DGS [13]24.4170.7810.321-12.212.21.27M
Ours+3DGS [13]25.4710.8160.2962.53.15.60.37M
2DGS [11]24.7660.7960.323-14.214.20.76M
Ours+2DGS [11]25.3800.8130.2962.54.87.30.37M
", + "bbox": [ + 94, + 279, + 478, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "initialization, in which 3D plane primitives are initialized by setting their radii as 0.05 and are placed on the bounding sphere of the scene with their normals pointing to the center of the scene. As reported in the top two rows of Tab. 3, initialization with Metric3Dv2 [10] mainly improves our method to be faster converged, while sphere initialization could also obtain promising results with more iterations.", + "bbox": [ + 89, + 380, + 482, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Plane Radii. We then evaluate the design of double-direction plane radii (Double Radii) by replacing it with the vanilla single-direction plane radii. It means that we learn one radius at each axis of the plane primitive. As shown in Tab. 3, optimizing without Double Radii leads to a decrease in the performance of geometry and segmentation, demonstrating the effectiveness of our Double Radii.", + "bbox": [ + 89, + 487, + 482, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Plane Splitting. As shown in the last two rows of Tab. 3, applying the Plane Splitting in optimization can further improve the performance in both geometry and segmentation.", + "bbox": [ + 89, + 593, + 483, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. PlanarSplatting for Novel View Synthesis", + "text_level": 1, + "bbox": [ + 89, + 651, + 442, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benefiting from our accurate and fast planar reconstruction, we show that our PlanarSplatting can seamlessly integrate with recent Gaussian Splatting methods for more efficient and better quality novel view synthesis in indoor scenes. We select two typical methods (3DGS [13] and 2DGS [11]) for evaluation and compare them with two variants including 'Ours+3DGS' and 'Ours+2DGS'. Specifically, for our two variants, we directly sample points from our reconstructed 3D plane primitives to use as the initialization of these Gaussian Splatting methods. Different from the original 3DGS and 2DGS, we fix the position of 3D points and exclude the densification operation when optimizing our two variants. For the baselines of 3DGS and 2DGS, we initialize them with SfM and run with densification to achieve the best performance. We conduct the experiments", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4229a7f08aa8f9b36891862c2f620d7c1356dce12a0b05c09d435859f0dacc50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 88, + 707, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a0ac663c9fc8963e1f882135e47e455848d5fcb04147030b99df4253518b5639.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 88, + 870, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/861144c33b8304525270ef59a448c080c172c483ae21b137e1b44ceb0425e4fb.jpg", + "image_caption": [ + "(a) 3DGS [13]", + "(c) Ours+3DGS [13]", + "Figure 8. Comparison of novel view synthesis on the ScanNetV2 dataset [2]. With initialization from our planar reconstruction, we significantly improve the rendering result of 3DGS [13]." + ], + "image_footnote": [], + "bbox": [ + 545, + 196, + 707, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5762a9ab220771dc100ac680dc36ab7e552770b398ec31489a577cc2eb4ecb7e.jpg", + "image_caption": [ + "(b) 2DGS [11]", + "(d) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 709, + 196, + 870, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "using the same scenes as those used in the ablation study. As shown in Tab. 4, our two variants 'Ours+3DGS' and 'Ours+2DGS' significantly outperform the original 3DGS and 2DGS in all metrics with even less optimization time and Gaussian points, demonstrating the superiority and potential of our PlanarSplatting. In Fig. 8, we further show that 'Ours+3DGS' effectively improves the rendering quality on the scene from the ScanNetV2 dataset.", + "bbox": [ + 511, + 363, + 906, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 511, + 492, + 777, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our PlanarSplatting can reconstruct the accurate indoor planar surface, it is not suitable for complex shapes such as curved surfaces. We leave this challenging problem in our future work for more flexible geometric modeling.", + "bbox": [ + 511, + 515, + 905, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 588, + 633, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present PlanarSplatting, a novel approach for multi-view 3D reconstruction of indoor scenes. By formulating the problem through differentiable rendering with plane splatting, we demonstrate the powerful capabilities of 3D planar representation for both accurate geometry reconstruction and compact structural scene modeling. Our efficient CUDA implementation enables fast 3D surface reconstruction, allowing comprehensive evaluation across over 100 scenes within hours using a single GPU. Furthermore, the seamless integration of PlanarSplatting with Gaussian Splatting significantly enhances both the quality and efficiency of indoor novel view synthesis, highlighting the broader potential of our approach for indoor scene understanding.", + "bbox": [ + 511, + 613, + 906, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 833, + 660, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by Ant Group Postdoctoral Programme. The authors would love to thank the anonymous reviewers and area chairs for their constructive comments.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1197", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samir Agarwala, Linyi Jin, Chris Rockwell, and David F. Fouhey. Planeformers: From sparse view planes to 3d reconstruction. In Eur. Conf. Comput. Vis., pages 192-209, 2022. 3", + "[2] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2432-2443, 2017. 1, 2, 5, 6, 7, 8", + "[3] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Int. Conf. Comput. Vis., pages 10766-10776, 2021. 2, 5, 6, 7", + "[4] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Manhattan-world stereo. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1422-1429, 2009. 3", + "[5] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Reconstructing building interiors from images. In Int. Conf. Comput. Vis., pages 80-87, 2009.", + "[6] David Gallup, Jan-Michael Frahm, and Marc Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1418-1425, 2010. 3", + "[7] Abhinav Gupta, Alexei A. Efros, and Martial Hebert. Blocks world revisited: Image understanding using qualitative geometry and mechanics. In *Eur. Conf. Comput. Vis.*, pages 482-496, 2010. 2", + "[8] Yuze He, Wang Zhao, Shaohui Liu, Yubin Hu, Yushi Bai, Yu-Hui Wen, and Yongjin Liu. Alphatables: A generic plane representation for 3d planar reconstruction from monocular videos. In Adv. Neural Inform. Process. Syst., 2024. 3", + "[9] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Comput. Vis. Image Underst., 157:167-178, 2017. 2", + "[10] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):10579-10596, 2024. 2, 4, 5, 6, 7, 8", + "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH Conference Papers, page 32, 2024. 2, 3, 5, 6, 8", + "[12] Linyi Jin, Shengyi Qian, Andrew Owens, and David F. Fouhey. Planar surface reconstruction from sparse views. In Int. Conf. Comput. Vis., pages 12971-12980, 2021. 2, 3", + "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 2, 3, 5, 8", + "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Int. Conf. Learn. Represent., 2015. 5" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Pierre-Alain Langlois, Alexandre Boulch, and Renaud Marlet. Surface reconstruction from 3d line segments. In Int. Conf. 3D Vis., pages 553-563, 2019. 3", + "[16] David C. Lee, Martial Hebert, and Takeo Kanade. Geometric reasoning for single image structure recovery. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2136-2143, 2009. 2, 3", + "[17] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single RGB image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2579-2588, 2018. 3", + "[18] Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz. Planercnn: 3d plane detection and reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4450-4459, 2019. 2, 3, 6", + "[19] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8655-8665. IEEE, 2022. 2", + "[20] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21445-21455. IEEE, 2023. 2", + "[21] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhöfer, Yaser Sheikh, and Jason M. Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 40(4):59:1-59:13, 2021. 3", + "[22] Branislav Micusik, Horst Wildenauer, and Markus Vincze. Towards detection of orthogonal planes in monocular images of indoor environments. In Int. Conf. on Robot. and Auto., pages 999-1004, 2008. 3", + "[23] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable blocks world: Qualitative 3d decomposition by rendering primitives. In Adv. Neural Inform. Process. Syst., 2023. 2, 3", + "[24] Aron Monszpart, Nicolas Mellado, Gabriel J. Brostow, and Niloy J. Mitra. Rapter: rebuilding man-made scenes with regular arrangements of planes. ACM Trans. Graph., 34(4): 103:1-103:12, 2015. 3", + "[25] Jann Poppinga, Narunas Vaskevicius, Andreas Birk, and Kaustubh Pathak. Fast plane detection and polygonalization in noisy 3d range images. In Int. Conf. Intell. Robots and Syst., pages 3378-3383, 2008. 2, 3", + "[26] Yiming Qian, Srikumar Ramalingam, and James H. Elder. LS3D: single-view gestalt 3d surface reconstruction from Manhattan line segments. In *Asian Conf. Comput. Vis.*, pages 399-416, 2018. 3", + "[27] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplerecon: 3d reconstruction without 3d convolutions. In Eur. Conf. Comput. Vis., pages 1-19, 2022. 6", + "[28] Jingjia Shi, Shuaifeng Zhi, and Kai Xu. Planerectr: Unified query learning for 3d plane recovery from a single view. In Int. Conf. Comput. Vis., pages 9343-9352, 2023. 3", + "[29] Sudipta N. Sinha, Drew Steedly, and Richard Szeliski. Piecewise planar stereo for image-based rendering. In Int. Conf. Comput. Vis., 2009. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1198", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Christiane Sommer, Yumin Sun, Leonidas J. Guibas, Daniel Cremers, and Tolga Birdal. From planes to corners: Multipurpose primitive detection in unorganized 3d point clouds. IEEE Robotics Autom. Lett., 5(2):1764-1771, 2020. 2, 3", + "[31] Bo Sun and Philippos Mordohai. Oriented point sampling for plane detection in unorganized point clouds. In Int. Conf. on Robot. and Auto., pages 2917-2923, 2019. 3", + "[32] Bin Tan, Nan Xue, Song Bai, Tianfu Wu, and Gui-Song Xia. Planetr: Structure-guided transformers for 3d plane recovery. In Int. Conf. Comput. Vis., pages 4166-4175, 2021. 3", + "[33] Bin Tan, Nan Xue, Tianfu Wu, and Gui-Song Xia. NOPE-SAC: neural one-plane RANSAC for sparse-view planar 3d reconstruction. IEEE Trans. Pattern Anal. Mach. Intell., 45 (12):15233-15248, 2023. 2, 3", + "[34] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Trans. Graph., 38 (6):230:1-230:14, 2019. 3", + "[35] Jamie Watson, Filippo Aleotti, Mohamed Sayed, Zawar Qureshi, Oisin Mac Aodha, Gabriel J. Brostow, Michael Firman, and Sara Vicente. Airplanes: Accurate plane estimation via 3d-consistent embeddings. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5270-5280, 2024. 1, 3, 6, 7", + "[36] Jianxiong Xiao and Yasutaka Furukawa. Reconstructing the world's museums. Int. J. Comput. Vis., 110(3):243-258, 2014. 3", + "[37] Yiming Xie, Matheus Gadelha, Fengting Yang, Xiaowei Zhou, and Huaizu Jiang. Planarrecon: Realtime 3d plane detection and reconstruction from posed monocular videos. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6209-6218, 2022. 1, 2, 3, 6, 7", + "[38] Sheng Xu, Ruisheng Wang, Hao Wang, and Ruigang Yang. Plane segmentation based on the optimal-vector-field in lidar point clouds. IEEE Trans. Pattern Anal. Mach. Intell., 43 (11):3991-4007, 2021. 2", + "[39] Nan Xue, Bin Tan, Yuxi Xiao, Liang Dong, Gui-Song Xia, Tianfu Wu, and Yujun Shen. NEAT: distilling 3d wireframes from neural attraction fields. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19968-19977. IEEE, 2024. 2", + "[40] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Int. Conf. Comput. Vis., pages 12-22, 2023. 2, 5, 6, 7", + "[41] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from A single image. In Int. Conf. Comput. Vis., 2023. 2", + "[42] Zehao Yu, Jia Zheng, Dongze Lian, Zihan Zhou, and Shenghua Gao. Single-image piece-wise planar 3d reconstruction via associative embedding. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1029-1037, 2019. 2, 3", + "[43] Jiahui Zhang, Jinfu Yang, Fuji Fu, and Jiaqi Ma. Planeac: Line-guided planar 3d reconstruction based on self-attention and convolution hybrid model. Pattern Recog., 153:110519, 2024. 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[44] Yanshu Zhang, Shichong Peng, Alireza Moazeni, and Ke Li. Papr: Proximity attention point rendering. In Adv. Neural Inform. Process. Syst., 2023. 3", + "bbox": [ + 514, + 90, + 905, + 133 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1199", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/3af07a07-d377-440f-9ac8-47f29938e643_model.json b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/3af07a07-d377-440f-9ac8-47f29938e643_model.json new file mode 100644 index 0000000000000000000000000000000000000000..900bac7720132d1b38cd5e7db9d5f0683113c268 --- /dev/null +++ b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/3af07a07-d377-440f-9ac8-47f29938e643_model.json @@ -0,0 +1,2585 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.856, + 0.152 + ], + "angle": 0, + "content": "PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.18, + 0.695, + 0.2 + ], + "angle": 0, + "content": "Bin Tan\\(^{1}\\) Rui Yu\\(^{2}\\) Yujun Shen\\(^{1}\\) Nan Xue\\(^{\\dagger,1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.215, + 0.658, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Ant Group \\(^{2}\\)University of Louisville" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.274, + 0.185, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.507, + 0.185, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.572, + 0.165, + 0.599 + ], + "angle": 0, + "content": "Input Images" + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.275, + 0.377, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.397, + 0.365, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.482, + 0.365, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.587, + 0.326, + 0.599 + ], + "angle": 0, + "content": "PlanarRecon" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.279, + 0.55, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.397, + 0.543, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.482, + 0.542, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.436, + 0.587, + 0.496, + 0.599 + ], + "angle": 0, + "content": "AirPlanes" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.276, + 0.733, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.397, + 0.722, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.482, + 0.721, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.587, + 0.685, + 0.6 + ], + "angle": 0, + "content": "Ours (3min)" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.276, + 0.907, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.397, + 0.901, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.482, + 0.901, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.784, + 0.587, + 0.866, + 0.599 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.615, + 0.908, + 0.658 + ], + "angle": 0, + "content": "Figure 1. We introduce PlanarSplatting, a fast and accurate optimization-based planar reconstruction method for indoor scenes. We show our planar reconstruction results on the ScanNetV2 [2] dataset achieved in 3 minutes. Compared to prior art PlanarRecon [37] and AirPlanes [35], our PlanarSplatting reconstructs more complete and detailed 3D planes." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.67, + 0.327, + 0.685 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.703, + 0.486, + 0.87 + ], + "angle": 0, + "content": "This paper presents PlanarSplatting, an ultra-fast and accurate surface reconstruction approach for multi-view indoor images. We take the 3D planes as the main objective due to their compactness and structural expressiveness in indoor scenes, and develop an explicit optimization framework that learns to fit the expected surface of indoor scenes by splatting the 3D planes into 2.5D depth and normal maps. As our PlanarSplatting operates directly on the 3D plane primitives, it eliminates the dependencies on 2D/3D plane detection and plane matching/tracking for planar surface reconstruction. Furthermore, with the essential" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.909, + 0.853 + ], + "angle": 0, + "content": "merits of plane-based representation coupled with CUDA-based implementation of planar splatting functions, Planar-Splatting reconstructs an indoor scene in 3 minutes while having significantly better geometric accuracy. Thanks to our ultra-fast reconstruction speed, the largest quantitative evaluation on the ScanNet and ScanNet++ datasets over hundreds of scenes clearly demonstrated the advantages of our method. We believe that our accurate and ultrafast planar surface reconstruction method will be applied in the structured data curation for surface reconstruction in the future. The code of our CUDA implementation will be publicly available." + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.887, + 0.242, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1190" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.264, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.089, + 0.419, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.089, + 0.574, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.089, + 0.731, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.089, + 0.887, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.29, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Figure 2. Novel View Synthesis: Seamless integration of our PlanarSplatting with recent Gaussian Splatting methods (e.g., 3DGS [13] and 2DGS [11]) achieves indoor-scene rendering with improved performance and significant optimization time reduction simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.337, + 0.223, + 0.352 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.516 + ], + "angle": 0, + "content": "We humans are long-term immersed in structured indoor scenes, ranging from bedrooms to offices. This fact has ignited a wealth of studies for reconstructing the indoor environment with various 3D structures such as lines [9, 20, 39], planes [12, 16, 18, 19, 33, 37, 42] and blocks [7, 23]. Among them, the 3D plane is the most common representation because of its simplicity and completeness in describing physical surfaces, thus motivating us to study the problem of 3D reconstruction for indoor scenes using 3D planes, i.e., planar 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.777 + ], + "angle": 0, + "content": "Planar 3D reconstruction has been extensively studied for years as a model fitting problem, in which a 3D scene geometry (e.g., point clouds, or meshes) was assumed to be known and the main goal is fitting the scene in a set of 3D planes [25, 30, 38]. Recently, the paradigm has been gradually simplified in image-based solutions in single-view and multi-view 3D reconstruction, eliminating the acquisition of known 3D scene geometry. To make the problem trackable, existing methods were extensively based on the image-level characterization of 2D/3D planes. That is to say, those methods have to detect 3D planes for each input image, match or track planes in across viewpoints, and finally reconstruct and merge 3D planes as the compact 3D representation of indoor scenes. Recent PlanarRecon [37] attempted to address these problems end-to-end by learning a 3D volume from monocular videos. Then, 3D planes can be detected, tracked, and fused in a consistent 3D space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Image-based planar 3D reconstruction mostly followed keypoint-based 3D reconstruction pipelines and treated planes as a special kind of visual features. However, because of the essential difference between the local point features and regional plane masks in image space, we argue, the existing approaches did not fully leverage the advantages of planar representations. Some evidences could be observed from the results of PlanarRecon [37], which are" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.339, + 0.875, + 0.353 + ], + "angle": 0, + "content": "usually very coarse and lose many details of the scene." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.355, + 0.907, + 0.612 + ], + "angle": 0, + "content": "In this paper, we are going to address the issues remained in image-based planar 3D reconstruction, aiming at obtaining a complete, structural, and compact indoor scene reconstruction from multi-view images. Our main idea is approximating the indoor scenes with a collection of solid 3D planar primitives from multi-view input images, directly optimizing them to have consistent 3D planes, but eliminating any suboptimal precomputing of 2D plane primitives (e.g., plane masks). We introduce PlanarSplatting that explicitly optimizes rectangular plane primitives in 3D space by differentiably splatting them into 2.5D depth and normal maps. Thanks to our well-designed plane splatting function, PlanarSplatting effectively leverage monocular geometry cues from modern foundational models [3, 10, 41] for accurate plane optimization. As shown in Fig. 1, the final high-quality planar surface can be reconstructed by simply merging adjacent 3D plane primitives." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.613, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Our PlanarSplatting is directly designed on the 3D plane primitives and efficiently implemented with CUDA, which accordingly induces a seamlessly integration with recent Gaussian Splatting (GS) methods for high-quality indoor novel view synthesis (NVS). As shown in Fig. 2, benefiting from our fast and accurate scene reconstruction (within 3 minutes), GS-based methods can be well-initialized and optimized without densification, leading to better rendering results and significantly less training time. It demonstrates the strong potential to promote the unity of surface reconstruction and novel view synthesis for indoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In the experiments, our PlanarSplatting shows its powerful ability for accurate indoor planar surface reconstruction for hundreds of scenes from two real-world indoor datasets, i.e. ScanNetV2 [2] and ScanNet++[40]. Furthermore, we show that with the combination of our PlanarSplatting and GS-based methods (e.g., 3DGS [13] and 2DGS [11]), we can effectively improve the rendering quality with less training time and fewer points." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1191" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.749 + ], + "angle": 0, + "content": "Indoor Planar Reconstruction. Reconstructing indoor scenes with 3D plane primitives has been studied for a long time [4-6, 29, 36]. Traditionally, it is usually achieved by deducing and fitting the plane geometry directly from 3D data (e.g., point clouds [24, 25, 30, 31] and line clouds [15]), or from 2D single-view images with strict scene assumptions (e.g., the Manhattan World constrain) [16, 22, 26] which seriously limited their application. In recent years, some learnable-based methods are proposed to formulate this problem as single-view 3D plane segmentation [17, 18, 28, 32, 42, 43] and cross-view plane instance matching [1, 12, 33]. Although impressive results have been achieved, these works are hard to extend to multiple views. PlanarRecon [37] was the first end-to-end work proposed to deal with holistic indoor plane recovery by learning a plane-related 3D volume from posed RGB videos. Then, 3D planes can be extracted, tracked, and fused from the learned 3D volume incrementally. Most recently, AirPlanes [35] developed a two-step method that first reconstructed a dense scene mesh and then learned consistent 3D plane embeddings from 2D plane embeddings for plane extraction from the dense mesh. Note that the aforementioned learning-based methods require 2D/3D plane annotations as their supervision, leading to performance bottlenecks due to the difficulty of obtaining a large scale of plane annotations. In contrast, we proposed PlanarSplatting to reconstruct accurate and complete indoor planar surfaces by directly optimizing a set of solid 3D plane primitives from posed multi-view images without any extra plane detection or matching operations. Benefiting from our differentiable planar primitive rendering, PlanarSplatting directly leverages monocular depth/normal cues from modern foundation models for optimization without plane annotations. Concurrent work, AlphaTablets [8], explores planar representations, which involves learning planar tablets equipped with alpha maps for shape modeling and pseudo-mesh construction for rendering. Our PlanarSplatting distinguishes itself from AlphaTablets by learning plane radii, providing a more explicit parameterization of 3D planes. As a result, we achieve a much more efficient rasterization scheme at the level of minutes while AlphaTablets takes about 2 hours per scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Primitive-based Scene Representation. Optimizing explicit primitives such as points [11, 13, 34, 44], volumes [21], and superquadric [23] to represent the 3D scene has been studied for a long time. The core of these methods is to design a differentiable rendering process to optimize the attributes of primitives by gradient descent. Among them, a typical paradigm is to render images from primitives with splatting techniques which is realized with radial basis functions defined on the primitive (e.g., the Gaussian function). Inspired by these works, we proposed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.213 + ], + "angle": 0, + "content": "PlanarSplatting to optimize solid 3D plane primitives in a differentiable rendering manner with a novel shape-aware plane splating function to better fit the scene geometry. Benefiting from our efficient CUDA implementation, PlanarSplatting can reconstruct the accurate indoor planar scene within 3 minutes and empower recent Gaussian-based works (e.g., 3DGS [13]) to further improve the rendering quality of novel views as shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.225, + 0.795, + 0.243 + ], + "angle": 0, + "content": "3. The Proposed PlanarSplatting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.909, + 0.388 + ], + "angle": 0, + "content": "As shown in Fig. 3, given a set of posed multi-view images, our PlanarSplatting can reconstruct the indoor planar scene from them by optimizing a set of learnable 3D planar primitives (Sec. 3.1). With the proposed Differentiable Planar Primitive Rendering (Sec. 3.2), we can explicitly learn these planar primitives from a coarse initialization to accurately recover the scene geometry. These optimized 3D planar primitives are then merged to achieve the 3D plane instances for the final planar reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.855, + 0.411 + ], + "angle": 0, + "content": "3.1. Learnable Planar Scene Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Since our key idea is directly optimizing explicit 3D primitives for planar scene reconstruction, we first present the representation of our learnable 3D planar primitives and then introduce how we initialize the scene with these planar primitives." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.492, + 0.909, + 0.614 + ], + "angle": 0, + "content": "Learnable Planar Primitive. As shown in Fig. 4, we formulate a planar primitive \\(\\pi\\) as a 3D rectangle which is equipped with several learnable parameters including the plane center \\(\\mathbf{p}_{\\pi} \\in \\mathbb{R}^3\\), the plane rotation \\(\\mathbf{q}_{\\pi} \\in \\mathbb{R}^4\\) (in quaternion representation) and the plane radii \\(\\mathbf{r}_{\\pi}\\). Specifically, to improve the optimization flexibility of plane shape, we use the design of double direction plane radii (Double Radii) for \\(\\mathbf{r}_{\\pi}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.622, + 0.907, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {\\pi} = \\left\\{r _ {\\pi} ^ {x +}, r _ {\\pi} ^ {x -}, r _ {\\pi} ^ {y +}, r _ {\\pi} ^ {y -} \\right\\} \\in \\mathbb {R} _ {+} ^ {4}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.909, + 0.74 + ], + "angle": 0, + "content": "where \\(r_{\\pi}^{x+}, r_{\\pi}^{x-}, r_{\\pi}^{y+}, r_{\\pi}^{y-}\\) are the radii defined on the positive/negative direction of the X-axis/Y-axis of the rectangle as shown in Fig. 4. For the sake of description, we further define the positive direction of the X-axis and Y-axis of the 3D planar primitive as two orthogonal unit vectors \\(\\mathbf{v}_{\\pi}^{x}, \\mathbf{v}_{\\pi}^{y} \\in \\mathbb{R}^{3}\\) which can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.746, + 0.907, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\pi} ^ {x} = \\mathbf {R} (\\mathbf {q} _ {\\pi}) [ 1, 0, 0 ] ^ {\\top}, \\quad \\mathbf {v} _ {\\pi} ^ {y} = \\mathbf {R} (\\mathbf {q} _ {\\pi}) [ 0, 1, 0 ] ^ {\\top}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.908, + 0.821 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}(\\mathbf{q}_{\\pi}) \\in \\mathbb{R}^{3 \\times 3}\\) means the rotation matrix of the quaternion \\(\\mathbf{q}_{\\pi}\\). Similarly, the normal of the planar primitive \\(\\mathbf{n}_{\\pi} \\in \\mathbb{R}^3\\) can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.827, + 0.907, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} _ {\\pi} = \\mathbf {R} (\\mathbf {q} _ {\\pi}) [ 0, 0, 1 ] ^ {\\top}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.902 + ], + "angle": 0, + "content": "With these learnable parameters, the 3D planar primitive can be moved to align with the potential scene surface and deformed to fit the surface shape during optimization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1192" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.107, + 0.875, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.388, + 0.908, + 0.431 + ], + "angle": 0, + "content": "Figure 3. Illustration of our proposed PlanarSplatting. Given a set of posed multi-view images of indoor scenes, our method renders depth and normal maps from 3D plane primitives. Then, with the supervision of monocular cues, these 3D plane primitives are gradually optimized to recover the scene geometry and finally merged to get the planar reconstruction result." + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.447, + 0.372, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.567 + ], + "angle": 0, + "content": "Figure 4. Representation of 3D plane primitive with learnable shape parameters including plane center, radii, and rotation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Scene Initialization. We use monocular depth from recent foundation models [10] to fast initialize our 3D planar primitives at the beginning of optimization. Specifically, we use depths from Metric3Dv2 [10] to get a very coarse scene geometry. Then we randomly sample 2,000 points on the coarse mesh to achieve the plane centers of our 3D planar primitives. The initial radii of each primitive \\(\\pi\\) is set to \\(0.5Dist(\\pi)\\). Here, \\(Dist(\\pi)\\) means the distance closest to \\(\\pi\\) to its neighbors. We use the normal direction on the coarse mesh to initialize the plane rotation. At the bottom of Fig. 3, we show an example of our initial planar primitives. With such a coarse initialization, our PlanarSplatting can finally reconstruct the accurate and complete scene surface." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.453, + 0.804 + ], + "angle": 0, + "content": "3.2. Differentiable Planar Primitive Rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.808, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To optimize the learnable planar primitives \\(\\Pi = \\{\\pi_i\\}_{i=1}^K\\), we introduce the differentiable planar primitive rendering with a carefully designed plane splating function which enables the planar primitives to accurately fit the scene geometry with the supervision only from 2D multi-view images." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.446, + 0.88, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.646 + ], + "angle": 0, + "content": "Figure 5. Illustration of the proposed plane splatting function. Naive Gaussian Splatting can not effectively approximate the boundary of our rectangular plane primitive (shown in black dashed border). In contrast, our proposed plane splatting function can approximate the boundary of the rectangle as the number of iterations increases, allowing our 3D planar primitives to better fit the surface of the scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Ray-to-Plane Intersection. To project 3D planar primitives to the 2D image space, we first calculate the intersections between planar primitives and the rays emitted from image pixels. Specifically, given a ray \\(\\mathbf{r} = \\{\\mathbf{o},\\mathbf{d}\\}\\) starting from the camera center \\(\\mathbf{o}\\in \\mathbb{R}^3\\) with direction \\(\\mathbf{d}\\in \\mathbb{R}^3\\), its intersection \\(\\mathbf{x}_{\\pi}^{\\mathbf{r}}\\in \\mathbb{R}^3\\) to one planar primitive \\(\\pi\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.767, + 0.907, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} = \\mathbf {o} + \\frac {\\left(\\mathbf {p} _ {\\pi} - \\mathbf {o}\\right) \\cdot \\mathbf {n} _ {\\pi}}{\\mathbf {d} \\cdot \\mathbf {n} _ {\\pi}} \\mathbf {d}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.839 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_{\\pi}\\) and \\(\\mathbf{n}_{\\pi}\\) are the center and the normal of the planar primitive \\(\\pi\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.885 + ], + "angle": 0, + "content": "Plane Splatting Function. After achieving the ray-to-plane intersection \\(\\mathbf{x}_{\\pi}^{\\mathbf{r}}\\), we calculate its splatting weight with our plane splatting function which will be used for rendering." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "A vanilla selection of the splatting function is the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1193" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.296, + 0.106 + ], + "angle": 0, + "content": "anisotropic Gaussian function:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.117, + 0.483, + 0.151 + ], + "angle": 0, + "content": "\\[\nw \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}, \\pi\\right) = \\exp \\left(- \\frac {1}{2} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right) ^ {\\top} \\Sigma^ {- 1} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.16, + 0.484, + 0.25 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_{\\pi}\\) is the center of the planar primitive \\(\\pi\\). The covariance matrix \\(\\Sigma\\) can be calculated like [11, 13]. However, as shown in Fig. 5 and Fig. 6, the Gaussian-based splatting function will make ambiguous boundaries of our rectangular planar primitive leading to the degeneration of reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.251, + 0.484, + 0.327 + ], + "angle": 0, + "content": "Thus, we propose to calculate the splating weight with a novel rectangle-based plane splating function. To a given intersection \\(\\mathbf{x}_{\\pi}^{\\mathbf{r}}\\) between ray \\(\\mathbf{r}\\) and planar primitive \\(\\pi\\), we first calculate its projection distance \\(\\mathcal{P}_X, \\mathcal{P}_Y \\in \\mathbb{R}\\) to the X-axis and Y-axis of the planar primitive as:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.337, + 0.483, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {X} = \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right) \\cdot \\mathbf {v} _ {\\pi} ^ {x}, \\quad \\mathcal {P} _ {Y} = \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right) \\cdot \\mathbf {v} _ {\\pi} ^ {y}. \\quad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.395 + ], + "angle": 0, + "content": "Then, we calculate the splating weight along the X-axis of the plane \\(\\pi\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.406, + 0.483, + 0.446 + ], + "angle": 0, + "content": "\\[\nw _ {X} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}\\right) = \\left\\{ \\begin{array}{l l} 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {x +} - | \\mathcal {P} _ {X} |\\right)\\right), & \\text {i f} \\mathcal {P} _ {X} > 0, \\\\ 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {x -} - | \\mathcal {P} _ {X} |\\right)\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.517 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) is the Sigmoid function and \\(\\lambda\\) is the hyperparameter to control the splattering weight. Similarly, we continue to calculate the splattering weight along the Y-axis of the plane \\(\\pi\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.528, + 0.483, + 0.568 + ], + "angle": 0, + "content": "\\[\nw _ {Y} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}\\right) = \\left\\{ \\begin{array}{l l} 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {y +} - | \\mathcal {P} _ {Y} |\\right)\\right), & \\text {i f} \\mathcal {P} _ {Y} > 0, \\\\ 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {y -} - | \\mathcal {P} _ {Y} |\\right)\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.484, + 0.621 + ], + "angle": 0, + "content": "where \\(r_{\\pi}^{x + },r_{\\pi}^{x - },r_{\\pi}^{y + },r_{\\pi}^{y - }\\) are the radii parameters of the plane \\(\\pi\\) . At last, the final splatting weight can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.621, + 0.483, + 0.66 + ], + "angle": 0, + "content": "\\[\nw \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}\\right) = \\left\\{ \\begin{array}{l l} w _ {X}, & \\text {i f} w _ {X} < w _ {Y}, \\\\ w _ {Y}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.484, + 0.743 + ], + "angle": 0, + "content": "In Fig. 5, we show that with the increment of hyperparameter \\(\\lambda\\), our plane splating function gradually approximates the shape of the rectangular plane primitive. In practice, we increase the value of \\(\\lambda\\) with an exponential function during optimization up to the maximum value of 300 as:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.75, + 0.483, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\lambda = \\min \\left(2 0 e ^ {\\left(- (1 - 0. 0 0 1 * i t e)\\right)}, 3 0 0\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.48, + 0.795 + ], + "angle": 0, + "content": "where ite means the iteration number during optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.905 + ], + "angle": 0, + "content": "Blending Composition. For all ray-to-plane intersections, we filter them with splating weight lower than 0.0001 and then sort the remaining intersections according to their depth from near to far. Then, \\(M\\) nearest intersections of each ray are selected for rendering (\\(M = 30\\) in this paper). Denote the selected intersections of a ray \\(\\mathbf{r}\\) as \\(P^{\\mathbf{r}} = \\{\\mathbf{x}_{\\pi_{\\tau(j)}}^{\\mathbf{r}}\\}_{j=1}^{M}\\). Here, \\(\\tau(j)\\) indicates the index of plane" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.63, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.153, + 0.626, + 0.167 + ], + "angle": 0, + "content": "(a) w/ GS Splitting" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.09, + 0.765, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.153, + 0.766, + 0.167 + ], + "angle": 0, + "content": "(b) w/ Plane Splatting" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.089, + 0.9, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.809, + 0.153, + 0.879, + 0.166 + ], + "angle": 0, + "content": "(c) GT Mesh" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.171, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Figure 6. Reconstruction comparison with different splatting functions. 'w/' means 'with'." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.216, + 0.905, + 0.247 + ], + "angle": 0, + "content": "among all planar primitives. At last, we render the depth and normal map of a certain image \\(\\mathbf{I}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.258, + 0.905, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) = \\sum_ {j = 1} ^ {M} T _ {j} w \\left(\\mathbf {x} _ {\\pi_ {\\tau (j)}} ^ {\\mathbf {r}}\\right) t _ {j}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.317, + 0.905, + 0.359 + ], + "angle": 0, + "content": "\\[\n\\mathbf {N} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) = \\sum_ {j = 1} ^ {M} T _ {j} w \\left(\\mathbf {x} _ {\\pi_ {\\tau (j)}} ^ {\\mathbf {r}}\\right) \\mathbf {n} _ {\\pi_ {\\tau (j)}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.367, + 0.56, + 0.38 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.379, + 0.905, + 0.42 + ], + "angle": 0, + "content": "\\[\nT _ {j} = \\prod_ {i = 1} ^ {j - 1} (1 - w \\left(\\mathbf {x} _ {\\pi_ {\\tau (i)}} ^ {\\mathbf {r}}\\right)). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.906, + 0.532 + ], + "angle": 0, + "content": "Here \\( t_j \\) is the depth of the intersection and \\( \\mathbf{n}_{\\pi_{\\tau(j)}} \\) is the normal of planar primitive \\( \\pi_{\\tau(j)} \\). To supervise the rendered depth and normal map, we use the pretrained model of Metric3Dv2 [10] to predict the depth map \\( \\mathbf{D}_{\\mathrm{pre}} \\) and normal map \\( \\mathbf{N}_{\\mathrm{pre}} \\) to serve as pseudo labels. Besides, we also test the normal map predicted by Omnidata [3]. Finally, the render loss can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.542, + 0.905, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e n d e r}} ^ {\\Pi} = \\alpha_ {1} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\| 1 - \\mathbf {N} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) ^ {\\top} \\mathbf {N} _ {\\text {p r e}} (\\mathbf {r})) \\| _ {1} + \\\\ \\alpha_ {1} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\| \\mathbf {N} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) - \\mathbf {N} _ {\\text {p r e}} (\\mathbf {r})) \\| _ {1} + \\tag {14} \\\\ \\alpha_ {2} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\| (\\mathbf {D} _ {\\mathrm {r e n d e r}} ^ {\\Pi} (\\mathbf {r})) - \\mathbf {D} _ {\\mathrm {p r e}} (\\mathbf {r}) \\| _ {1}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.656, + 0.905, + 0.687 + ], + "angle": 0, + "content": "where \\(\\alpha_{1} = 2.0\\), \\(\\alpha_{2} = 2.0\\), \\(\\mathbf{r}\\) is the ray/pixel emitted from image \\(\\mathbf{I}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.697, + 0.654, + 0.714 + ], + "angle": 0, + "content": "3.3. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.781 + ], + "angle": 0, + "content": "Loss Function. To each scene, we optimize our Planar-Splatting with the Adam optimizer [14] for 5,000 iterations on the ScanNetV2 [2] dataset and 10,000 iterations on the ScanNet++ [40] dataset using the loss as shown in Eq. (14)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Plane Splitting. During optimization, we introduce a splitting operation on planes according to the gradients of their radii to better fit the scene geometry. If the average radii gradients on the X-axis \\((r^{x+}\\) and \\(r^{x-})\\) of the plane are greater than 0.2, we split the plane along the Y-axis. Similarly, we split the planes along the X-axis, if their radii gradients on the Y-axis \\((r^{y+}\\) and \\(r^{y-})\\) are larger than 0.2. We conduct the splitting operation every 1,000 iterations." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1194" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of planar reconstruction results on the ScanNetV2 [2] dataset. 'P. Ann.' means using 2D/3D plane annotations in training stages." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.122, + 0.898, + 0.243 + ], + "angle": 0, + "content": "
MethodP. Ann.GeometrySegmentationPlanar
Chamfer ↓F-score ↑VOI ↓RI ↑SC ↑Fidelity ↓Acc ↓Chamfer ↓
PlanarRecon [37]✓9.8943.473.2010.9190.40518.8616.2117.53
AirPlanes [35]✓5.3064.922.2680.9570.5688.767.988.37
2DGS [11] + RANSACX14.1531.334.0300.9240.25740.0214.7727.40
Metric3Dv2 [10] + RANSACX5.9365.452.7440.9270.4889.2911.3610.33
SR [27] + RANSACX5.4065.452.5070.9460.5159.4210.139.78
Ours@OmnidataX4.8368.852.5020.9480.5326.6411.769.20
Ours@Metric3Dv2X4.2373.762.3100.9580.5605.7210.658.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.256, + 0.908, + 0.286 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of planar reconstruction results on the ScanNet++ [40] dataset. 'P. Ann.' means using 2D/3D plane annotations in training stages." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.289, + 0.898, + 0.41 + ], + "angle": 0, + "content": "
MethodP. Ann.GeometrySegmentationPlanar
Chamfer ↓F-score ↑VOI ↓RI ↑SC ↑Fidelity ↓Acc ↓Chamfer ↓
PlanarRecon [37]✓17.8531.103.5420.9190.36734.4419.3526.90
AirPlanes [35]✓13.7532.582.8590.9410.47028.1612.5820.37
2DGS [11] + RANSACX20.3926.464.4560.9270.24155.9016.8836.39
Metric3Dv2 [10] + RANSACX11.1441.543.0210.9300.44127.5010.8119.16
SR [27] + RANSACX13.1535.933.0130.9380.44230.2511.6220.94
Ours@OmnidataX9.3347.042.7720.9460.52317.2412.2614.75
Ours@Metric3Dv2X9.0150.682.7820.9440.52117.1712.3614.77
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.483, + 0.498 + ], + "angle": 0, + "content": "Plane Merge. After optimization, we greedily merge the adjacent 3D plane primitives with normal angle error lower than \\(25^{\\circ}\\) and offset distance error lower than \\(0.1m\\). Here, offset means the projection distance from the scene center to the plane surface." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.484, + 0.607 + ], + "angle": 0, + "content": "CUDA Implementation. For fast optimization of solid 3D planar primitives, we implement the forward and backward process of our Differentiable Planar Primitive Rendering with CUDA, which enables our PlanarSplatting to reconstruct one scene within 3 minutes. We will release the code of our CUDA implementation for solid 3D planar primitives optimization after publication." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.624, + 0.224, + 0.641 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.363, + 0.667 + ], + "angle": 0, + "content": "4.1. Dataset, Metrics and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Datasets. We evaluate our PlanarSplatting on two large indoor datasets including ScanNetV2 [2] and ScanNet++ [40] which provide posed images. On the ScanNetV2 dataset, we use the test split according to AirPlanes [35] which includes 100 scenes with ground truth 3D plane annotations provided by [18]. On the ScanNet++ dataset, we randomly select 30 scenes for evaluation and extract 3D plane annotations from the ground truth meshes like [18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. According to PlanarRecon [37], we evaluate the geometry quality of all reconstructed 3D planes with Chamfer Distance and F-score. Following AirPlanes [35], we also evaluate the reconstruction quality of Top-20 largest planes from the ground truth and use the metrics including Planar Fidelity, Planar Accuracy, and Planar Chamfer. To evaluate 3D plane segmentation, we use" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.455 + ], + "angle": 0, + "content": "the metrics including Variation of Information (VOI), Rand Index (RI), and Segmentation Covering (SC) like [18, 37]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.909, + 0.652 + ], + "angle": 0, + "content": "Baselines. Since our method is purely built upon geometry cues, we mainly compare our PlanarSplatting with those geometry-based methods including SR+RANSAC [27], Metric3Dv2+RANSAC [10] and 2DGS+RANSAC [11]. These baselines build dense scene meshes from multi-view images at first and then extract 3D planes with the RANSAC algorithm from the reconstructed meshes. We use the RANSAC implementation provided by AirPlanes [35] for all these geometry-based baselines. Besides, we also compare our PlanarSplatting with some plane annotation based methods that use 2D/3D plane labels/priors in their training stage including PlanarRecon [37] and AirPlanes [35], and report their results for reference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.767, + 0.683 + ], + "angle": 0, + "content": "4.2. Comparisons with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. We first evaluate our PlanarSplatting on the ScanNetV2 [2] dataset. Specifically, we evaluate two versions of our PlanarSplatting, utilizing normal maps from different pretrained models for supervision: Metric3Dv2 [10] and Omnidata [3], respectively. As shown in Tab. 1, our method achieves the best geometry performance in the metric of Chamfer and F-score compared to all baselines. To the results of Segmentation and Planar, our method outperforms geometry-based baselines in most metrics and is comparable to the state-of-the-art AirPlanes which uses plane embeddings trained on the same ScanNetV2 dataset. In Tab. 2, we show the results on the ScanNet++ [40] dataset. Benefiting from our novel design for differentiable plane rendering and" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1195" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.093, + 0.302, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.099, + 0.49, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.099, + 0.675, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.102, + 0.858, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.197, + 0.309, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.196, + 0.493, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.196, + 0.674, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.197, + 0.857, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.326, + 0.309, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.326, + 0.49, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.323, + 0.67, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.327, + 0.857, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.421, + 0.31, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.421, + 0.493, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.421, + 0.674, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.421, + 0.857, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.533, + 0.305, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.616, + 0.28, + 0.628 + ], + "angle": 0, + "content": "(a) PlanarRecon [37]" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.532, + 0.483, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.616, + 0.454, + 0.628 + ], + "angle": 0, + "content": "(b) AirPlanes [35]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.533, + 0.668, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.617, + 0.611, + 0.627 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.534, + 0.848, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.616, + 0.816, + 0.627 + ], + "angle": 0, + "content": "(d) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.633, + 0.814, + 0.647 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison on the ScanNetV2 [2] (rows 1-4) and ScanNet++ [40] (last row) datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.766 + ], + "angle": 0, + "content": "optimization via monocular cues provided by modern foundation models, our PlanarSplatting still achieves the best Geometry performance among all baselines. Furthermore, our method outperforms both geometry-based and plane annotation based methods in most metrics of Segmentation and Planar, demonstrating the robustness and superiority of our proposed PlanarSplatting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results. In Fig. 7, we show the comparisons of PlanarRecon [37], AirPlanes [35] and our PlanarSplatting with normal supervision of Omnidata [3]. PlanarRecon reconstructs coarse 3D planes from input multi-view images while our PlanarSplatting accurately reconstructs scene geometry via directly optimizing 3D plane primitives in the whole 3D space. With the help of a plane embedding model, AirPlanes achieves more semantic plane segmentation from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.905, + 0.766 + ], + "angle": 0, + "content": "co-plane regions. However, as shown in the zoom-in results in Fig. 7, AirPlanes loses many geometry details in the scene. In contrast, our PlanarSplatting successfully reconstructs detailed structures such as the chairs and the legs of the bed, resulting in both accuracy and completeness in results. These comparisons effectively indicate the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.679, + 0.787 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.855 + ], + "angle": 0, + "content": "We verify the design of our PlanarSplatting on 10 scenes randomly selected from the ScanNetV2 [2] dataset. We optimize 3D plane primitives with 5,000 iterations with the normal supervision of Omnidata [3] in default." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Plane Initialization. We first assess the sensitivity of the plane initialization strategy of our PlanarSplatting. We ablate the used Metric3Dv2 [10] initialization to the sphere" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1196" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 3. Ablation studies of the proposed PlanarSplatting on the ScanNetV2 dataset. 'w/' means 'with' and 'w/o' means 'without'." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.122, + 0.48, + 0.205 + ], + "angle": 0, + "content": "
Chamfer ↓F-score ↑VOI ↓
w/ Sphere Init.8.3957.533.073
w/ Sphere Init. (30K)4.5869.912.522
w/o Double Radii4.7370.252.491
w/o Plane Splitting4.6970.302.477
Ours (Full)4.6671.302.473
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.22, + 0.483, + 0.276 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of novel view synthesis on the ScanNetV2 [2] dataset. 'Plane' means time for our plane optimization. 'GS' means time for 2D/3D Gaussian optimization. #P' means the average number of Gaussian points." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.28, + 0.48, + 0.362 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓Avg. Time (min)#P
PlaneGSTotal
3DGS [13]24.4170.7810.321-12.212.21.27M
Ours+3DGS [13]25.4710.8160.2962.53.15.60.37M
2DGS [11]24.7660.7960.323-14.214.20.76M
Ours+2DGS [11]25.3800.8130.2962.54.87.30.37M
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.488 + ], + "angle": 0, + "content": "initialization, in which 3D plane primitives are initialized by setting their radii as 0.05 and are placed on the bounding sphere of the scene with their normals pointing to the center of the scene. As reported in the top two rows of Tab. 3, initialization with Metric3Dv2 [10] mainly improves our method to be faster converged, while sphere initialization could also obtain promising results with more iterations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.483, + 0.594 + ], + "angle": 0, + "content": "Plane Radii. We then evaluate the design of double-direction plane radii (Double Radii) by replacing it with the vanilla single-direction plane radii. It means that we learn one radius at each axis of the plane primitive. As shown in Tab. 3, optimizing without Double Radii leads to a decrease in the performance of geometry and segmentation, demonstrating the effectiveness of our Double Radii." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.484, + 0.641 + ], + "angle": 0, + "content": "Plane Splitting. As shown in the last two rows of Tab. 3, applying the Plane Splitting in optimization can further improve the performance in both geometry and segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.444, + 0.669 + ], + "angle": 0, + "content": "4.4. PlanarSplatting for Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Benefiting from our accurate and fast planar reconstruction, we show that our PlanarSplatting can seamlessly integrate with recent Gaussian Splatting methods for more efficient and better quality novel view synthesis in indoor scenes. We select two typical methods (3DGS [13] and 2DGS [11]) for evaluation and compare them with two variants including 'Ours+3DGS' and 'Ours+2DGS'. Specifically, for our two variants, we directly sample points from our reconstructed 3D plane primitives to use as the initialization of these Gaussian Splatting methods. Different from the original 3DGS and 2DGS, we fix the position of 3D points and exclude the densification operation when optimizing our two variants. For the baselines of 3DGS and 2DGS, we initialize them with SfM and run with densification to achieve the best performance. We conduct the experiments" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.089, + 0.709, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.185, + 0.668, + 0.197 + ], + "angle": 0, + "content": "(a) 3DGS [13]" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.872, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.185, + 0.831, + 0.197 + ], + "angle": 0, + "content": "(b) 2DGS [11]" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.198, + 0.708, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.294, + 0.682, + 0.307 + ], + "angle": 0, + "content": "(c) Ours+3DGS [13]" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.198, + 0.871, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.294, + 0.836, + 0.306 + ], + "angle": 0, + "content": "(d) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.311, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Figure 8. Comparison of novel view synthesis on the ScanNetV2 dataset [2]. With initialization from our planar reconstruction, we significantly improve the rendering result of 3DGS [13]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.485 + ], + "angle": 0, + "content": "using the same scenes as those used in the ablation study. As shown in Tab. 4, our two variants 'Ours+3DGS' and 'Ours+2DGS' significantly outperform the original 3DGS and 2DGS in all metrics with even less optimization time and Gaussian points, demonstrating the superiority and potential of our PlanarSplatting. In Fig. 8, we further show that 'Ours+3DGS' effectively improves the rendering quality on the scene from the ScanNetV2 dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.493, + 0.779, + 0.509 + ], + "angle": 0, + "content": "4.5. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.577 + ], + "angle": 0, + "content": "Although our PlanarSplatting can reconstruct the accurate indoor planar surface, it is not suitable for complex shapes such as curved surfaces. We leave this challenging problem in our future work for more flexible geometric modeling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.634, + 0.604 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.826 + ], + "angle": 0, + "content": "In this paper, we present PlanarSplatting, a novel approach for multi-view 3D reconstruction of indoor scenes. By formulating the problem through differentiable rendering with plane splatting, we demonstrate the powerful capabilities of 3D planar representation for both accurate geometry reconstruction and compact structural scene modeling. Our efficient CUDA implementation enables fast 3D surface reconstruction, allowing comprehensive evaluation across over 100 scenes within hours using a single GPU. Furthermore, the seamless integration of PlanarSplatting with Gaussian Splatting significantly enhances both the quality and efficiency of indoor novel view synthesis, highlighting the broader potential of our approach for indoor scene understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.661, + 0.85 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work was supported by Ant Group Postdoctoral Programme. The authors would love to thank the anonymous reviewers and area chairs for their constructive comments." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1197" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Samir Agarwala, Linyi Jin, Chris Rockwell, and David F. Fouhey. Planeformers: From sparse view planes to 3d reconstruction. In Eur. Conf. Comput. Vis., pages 192-209, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.174, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2432-2443, 2017. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[3] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Int. Conf. Comput. Vis., pages 10766-10776, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[4] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Manhattan-world stereo. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1422-1429, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.349, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[5] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Reconstructing building interiors from images. In Int. Conf. Comput. Vis., pages 80-87, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.393, + 0.483, + 0.449 + ], + "angle": 0, + "content": "[6] David Gallup, Jan-Michael Frahm, and Marc Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1418-1425, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[7] Abhinav Gupta, Alexei A. Efros, and Martial Hebert. Blocks world revisited: Image understanding using qualitative geometry and mechanics. In *Eur. Conf. Comput. Vis.*, pages 482-496, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.483, + 0.565 + ], + "angle": 0, + "content": "[8] Yuze He, Wang Zhao, Shaohui Liu, Yubin Hu, Yushi Bai, Yu-Hui Wen, and Yongjin Liu. Alphatables: A generic plane representation for 3d planar reconstruction from monocular videos. In Adv. Neural Inform. Process. Syst., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[9] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Comput. Vis. Image Underst., 157:167-178, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.612, + 0.483, + 0.695 + ], + "angle": 0, + "content": "[10] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):10579-10596, 2024. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.698, + 0.483, + 0.754 + ], + "angle": 0, + "content": "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH Conference Papers, page 32, 2024. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.757, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[12] Linyi Jin, Shengyi Qian, Andrew Owens, and David F. Fouhey. Planar surface reconstruction from sparse views. In Int. Conf. Comput. Vis., pages 12971-12980, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.801, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Int. Conf. Learn. Represent., 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[15] Pierre-Alain Langlois, Alexandre Boulch, and Renaud Marlet. Surface reconstruction from 3d line segments. In Int. Conf. 3D Vis., pages 553-563, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[16] David C. Lee, Martial Hebert, and Takeo Kanade. Geometric reasoning for single image structure recovery. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2136-2143, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[17] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single RGB image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2579-2588, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[18] Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz. Planercnn: 3d plane detection and reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4450-4459, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[19] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8655-8665. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[20] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21445-21455. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[21] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhöfer, Yaser Sheikh, and Jason M. Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 40(4):59:1-59:13, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[22] Branislav Micusik, Horst Wildenauer, and Markus Vincze. Towards detection of orthogonal planes in monocular images of indoor environments. In Int. Conf. on Robot. and Auto., pages 999-1004, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[23] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable blocks world: Qualitative 3d decomposition by rendering primitives. In Adv. Neural Inform. Process. Syst., 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[24] Aron Monszpart, Nicolas Mellado, Gabriel J. Brostow, and Niloy J. Mitra. Rapter: rebuilding man-made scenes with regular arrangements of planes. ACM Trans. Graph., 34(4): 103:1-103:12, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[25] Jann Poppinga, Narunas Vaskevicius, Andreas Birk, and Kaustubh Pathak. Fast plane detection and polygonalization in noisy 3d range images. In Int. Conf. Intell. Robots and Syst., pages 3378-3383, 2008. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[26] Yiming Qian, Srikumar Ramalingam, and James H. Elder. LS3D: single-view gestalt 3d surface reconstruction from Manhattan line segments. In *Asian Conf. Comput. Vis.*, pages 399-416, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[27] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplerecon: 3d reconstruction without 3d convolutions. In Eur. Conf. Comput. Vis., pages 1-19, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[28] Jingjia Shi, Shuaifeng Zhi, and Kai Xu. Planerectr: Unified query learning for 3d plane recovery from a single view. In Int. Conf. Comput. Vis., pages 9343-9352, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Sudipta N. Sinha, Drew Steedly, and Richard Szeliski. Piecewise planar stereo for image-based rendering. In Int. Conf. Comput. Vis., 2009. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1198" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[30] Christiane Sommer, Yumin Sun, Leonidas J. Guibas, Daniel Cremers, and Tolga Birdal. From planes to corners: Multipurpose primitive detection in unorganized 3d point clouds. IEEE Robotics Autom. Lett., 5(2):1764-1771, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[31] Bo Sun and Philippos Mordohai. Oriented point sampling for plane detection in unorganized point clouds. In Int. Conf. on Robot. and Auto., pages 2917-2923, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.195, + 0.482, + 0.236 + ], + "angle": 0, + "content": "[32] Bin Tan, Nan Xue, Song Bai, Tianfu Wu, and Gui-Song Xia. Planetr: Structure-guided transformers for 3d plane recovery. In Int. Conf. Comput. Vis., pages 4166-4175, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[33] Bin Tan, Nan Xue, Tianfu Wu, and Gui-Song Xia. NOPE-SAC: neural one-plane RANSAC for sparse-view planar 3d reconstruction. IEEE Trans. Pattern Anal. Mach. Intell., 45 (12):15233-15248, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[34] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Trans. Graph., 38 (6):230:1-230:14, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.354, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[35] Jamie Watson, Filippo Aleotti, Mohamed Sayed, Zawar Qureshi, Oisin Mac Aodha, Gabriel J. Brostow, Michael Firman, and Sara Vicente. Airplanes: Accurate plane estimation via 3d-consistent embeddings. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5270-5280, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.44, + 0.482, + 0.48 + ], + "angle": 0, + "content": "[36] Jianxiong Xiao and Yasutaka Furukawa. Reconstructing the world's museums. Int. J. Comput. Vis., 110(3):243-258, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[37] Yiming Xie, Matheus Gadelha, Fengting Yang, Xiaowei Zhou, and Huaizu Jiang. Planarrecon: Realtime 3d plane detection and reconstruction from posed monocular videos. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6209-6218, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.556, + 0.482, + 0.61 + ], + "angle": 0, + "content": "[38] Sheng Xu, Ruisheng Wang, Hao Wang, and Ruigang Yang. Plane segmentation based on the optimal-vector-field in lidar point clouds. IEEE Trans. Pattern Anal. Mach. Intell., 43 (11):3991-4007, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.614, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[39] Nan Xue, Bin Tan, Yuxi Xiao, Liang Dong, Gui-Song Xia, Tianfu Wu, and Yujun Shen. NEAT: distilling 3d wireframes from neural attraction fields. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19968-19977. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[40] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Int. Conf. Comput. Vis., pages 12-22, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.73, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[41] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from A single image. In Int. Conf. Comput. Vis., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.787, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[42] Zehao Yu, Jia Zheng, Dongze Lian, Zihan Zhou, and Shenghua Gao. Single-image piece-wise planar 3d reconstruction via associative embedding. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1029-1037, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[43] Jiahui Zhang, Jinfu Yang, Fuji Fu, and Jiaqi Ma. Planeac: Line-guided planar 3d reconstruction based on self-attention and convolution hybrid model. Pattern Recog., 153:110519, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.092, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[44] Yanshu Zhang, Shichong Peng, Alireza Moazeni, and Ke Li. Papr: Proximity attention point rendering. In Adv. Neural Inform. Process. Syst., 2023. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1199" + } + ] +] \ No newline at end of file diff --git a/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/3af07a07-d377-440f-9ac8-47f29938e643_origin.pdf b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/3af07a07-d377-440f-9ac8-47f29938e643_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cb56cc7ba985a4accb937c53fb380afe5d296977 --- /dev/null +++ b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/3af07a07-d377-440f-9ac8-47f29938e643_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b0cd500339830c4e3791a355645b17417fe7642da5bc0453c9d4e92ccf63f69 +size 8139945 diff --git a/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/full.md b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..57a3df806973251e7aaeb83254640b36e3979ea3 --- /dev/null +++ b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/full.md @@ -0,0 +1,397 @@ +# PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes + +Bin Tan $^{1}$ Rui Yu $^{2}$ Yujun Shen $^{1}$ Nan Xue $^{\dagger,1}$ + +$^{1}$ Ant Group $^{2}$ University of Louisville + +![](images/b827c1d16c0df3e53f13f20d1d328eaa178a00e3a8e2a42aad9c6a6c185dbaae.jpg) + +![](images/b3dddd3f4530dbc79fd5a16b7b9fbb9a27311b9783876e02963049084f9d3489.jpg) +Input Images + +![](images/0451cdae4cbd7334c7243a32e757d3d445c4ee3ec651a55faca10f7db8636092.jpg) + +![](images/79aa4b4f4e46379333f2b65e99ba502bed86d79f234d3637c59a33aef65efb4d.jpg) + +![](images/4916d6cbdba2112540472625f4b4183a3e4ef568bbb2e0857eb68ba39c726d95.jpg) +PlanarRecon +Figure 1. We introduce PlanarSplatting, a fast and accurate optimization-based planar reconstruction method for indoor scenes. We show our planar reconstruction results on the ScanNetV2 [2] dataset achieved in 3 minutes. Compared to prior art PlanarRecon [37] and AirPlanes [35], our PlanarSplatting reconstructs more complete and detailed 3D planes. + +![](images/1781cecd8eb6918046d57c32821de614ad47b67aa2e9518f3034e52851b0e9a3.jpg) + +![](images/e95653c421b1bbb220c8ce549fef0c28c7d86d8bbcd139a8c1464ff140bc80be.jpg) + +![](images/468b392657c3a7cb0b2a5dae055a6b06df7e94753aca59826605cccd7f5ff261.jpg) +AirPlanes + +![](images/352fddffc5b6279c78776f2f43f6aadfac74c4d1c15e1a3e5f01be4955fcabc2.jpg) + +![](images/74bee36ad248c719a6aa81152b3bff713ad696a2e5053bdbb878d5ec3626f41a.jpg) + +![](images/2541fda366f5214e314e57546836a2d3c5aded25a626ff05d9f762adb5b64aba.jpg) +Ours (3min) + +![](images/1a4123ad3d6a2e9ea0e457cba4140e1753e6f8b69f01afdb72c448e61339d145.jpg) + +![](images/9765116973bfe08b642c5744ecb8719655ee8b6df897d50875c60d2e9a7a2384.jpg) + +![](images/d2f33164e35e29dcb882b5948ad8c26a7f9dcfb2317e8e72ef6760342e42cd16.jpg) +Ground Truth + +# Abstract + +This paper presents PlanarSplatting, an ultra-fast and accurate surface reconstruction approach for multi-view indoor images. We take the 3D planes as the main objective due to their compactness and structural expressiveness in indoor scenes, and develop an explicit optimization framework that learns to fit the expected surface of indoor scenes by splatting the 3D planes into 2.5D depth and normal maps. As our PlanarSplatting operates directly on the 3D plane primitives, it eliminates the dependencies on 2D/3D plane detection and plane matching/tracking for planar surface reconstruction. Furthermore, with the essential + +merits of plane-based representation coupled with CUDA-based implementation of planar splatting functions, Planar-Splatting reconstructs an indoor scene in 3 minutes while having significantly better geometric accuracy. Thanks to our ultra-fast reconstruction speed, the largest quantitative evaluation on the ScanNet and ScanNet++ datasets over hundreds of scenes clearly demonstrated the advantages of our method. We believe that our accurate and ultrafast planar surface reconstruction method will be applied in the structured data curation for surface reconstruction in the future. The code of our CUDA implementation will be publicly available. + +![](images/eb3d1f06b998f755890b6cac6dfe7e9370a3513152a72585733a531311d3cd2d.jpg) +Figure 2. Novel View Synthesis: Seamless integration of our PlanarSplatting with recent Gaussian Splatting methods (e.g., 3DGS [13] and 2DGS [11]) achieves indoor-scene rendering with improved performance and significant optimization time reduction simultaneously. + +![](images/57d8d611aa27ba0110bf11462e32e60d5fa3707ccf20e762ed0f2d739edc2ef4.jpg) + +![](images/4351adb86ce454a84d5604a730584fc91e2048110c3ce9379b9f12b3aabafe57.jpg) + +![](images/45fb1fec7ac4f6e2bce35f9247eb24272ca72a120fdcbd4962f410603be3952f.jpg) + +![](images/a211d1496809b85aed57ff390bad18cb58749e4691c062529158b66765332b05.jpg) + +# 1. Introduction + +We humans are long-term immersed in structured indoor scenes, ranging from bedrooms to offices. This fact has ignited a wealth of studies for reconstructing the indoor environment with various 3D structures such as lines [9, 20, 39], planes [12, 16, 18, 19, 33, 37, 42] and blocks [7, 23]. Among them, the 3D plane is the most common representation because of its simplicity and completeness in describing physical surfaces, thus motivating us to study the problem of 3D reconstruction for indoor scenes using 3D planes, i.e., planar 3D reconstruction. + +Planar 3D reconstruction has been extensively studied for years as a model fitting problem, in which a 3D scene geometry (e.g., point clouds, or meshes) was assumed to be known and the main goal is fitting the scene in a set of 3D planes [25, 30, 38]. Recently, the paradigm has been gradually simplified in image-based solutions in single-view and multi-view 3D reconstruction, eliminating the acquisition of known 3D scene geometry. To make the problem trackable, existing methods were extensively based on the image-level characterization of 2D/3D planes. That is to say, those methods have to detect 3D planes for each input image, match or track planes in across viewpoints, and finally reconstruct and merge 3D planes as the compact 3D representation of indoor scenes. Recent PlanarRecon [37] attempted to address these problems end-to-end by learning a 3D volume from monocular videos. Then, 3D planes can be detected, tracked, and fused in a consistent 3D space. + +Image-based planar 3D reconstruction mostly followed keypoint-based 3D reconstruction pipelines and treated planes as a special kind of visual features. However, because of the essential difference between the local point features and regional plane masks in image space, we argue, the existing approaches did not fully leverage the advantages of planar representations. Some evidences could be observed from the results of PlanarRecon [37], which are + +usually very coarse and lose many details of the scene. + +In this paper, we are going to address the issues remained in image-based planar 3D reconstruction, aiming at obtaining a complete, structural, and compact indoor scene reconstruction from multi-view images. Our main idea is approximating the indoor scenes with a collection of solid 3D planar primitives from multi-view input images, directly optimizing them to have consistent 3D planes, but eliminating any suboptimal precomputing of 2D plane primitives (e.g., plane masks). We introduce PlanarSplatting that explicitly optimizes rectangular plane primitives in 3D space by differentiably splatting them into 2.5D depth and normal maps. Thanks to our well-designed plane splatting function, PlanarSplatting effectively leverage monocular geometry cues from modern foundational models [3, 10, 41] for accurate plane optimization. As shown in Fig. 1, the final high-quality planar surface can be reconstructed by simply merging adjacent 3D plane primitives. + +Our PlanarSplatting is directly designed on the 3D plane primitives and efficiently implemented with CUDA, which accordingly induces a seamlessly integration with recent Gaussian Splatting (GS) methods for high-quality indoor novel view synthesis (NVS). As shown in Fig. 2, benefiting from our fast and accurate scene reconstruction (within 3 minutes), GS-based methods can be well-initialized and optimized without densification, leading to better rendering results and significantly less training time. It demonstrates the strong potential to promote the unity of surface reconstruction and novel view synthesis for indoor scenes. + +In the experiments, our PlanarSplatting shows its powerful ability for accurate indoor planar surface reconstruction for hundreds of scenes from two real-world indoor datasets, i.e. ScanNetV2 [2] and ScanNet++[40]. Furthermore, we show that with the combination of our PlanarSplatting and GS-based methods (e.g., 3DGS [13] and 2DGS [11]), we can effectively improve the rendering quality with less training time and fewer points. + +# 2. Related Work + +Indoor Planar Reconstruction. Reconstructing indoor scenes with 3D plane primitives has been studied for a long time [4-6, 29, 36]. Traditionally, it is usually achieved by deducing and fitting the plane geometry directly from 3D data (e.g., point clouds [24, 25, 30, 31] and line clouds [15]), or from 2D single-view images with strict scene assumptions (e.g., the Manhattan World constrain) [16, 22, 26] which seriously limited their application. In recent years, some learnable-based methods are proposed to formulate this problem as single-view 3D plane segmentation [17, 18, 28, 32, 42, 43] and cross-view plane instance matching [1, 12, 33]. Although impressive results have been achieved, these works are hard to extend to multiple views. PlanarRecon [37] was the first end-to-end work proposed to deal with holistic indoor plane recovery by learning a plane-related 3D volume from posed RGB videos. Then, 3D planes can be extracted, tracked, and fused from the learned 3D volume incrementally. Most recently, AirPlanes [35] developed a two-step method that first reconstructed a dense scene mesh and then learned consistent 3D plane embeddings from 2D plane embeddings for plane extraction from the dense mesh. Note that the aforementioned learning-based methods require 2D/3D plane annotations as their supervision, leading to performance bottlenecks due to the difficulty of obtaining a large scale of plane annotations. In contrast, we proposed PlanarSplatting to reconstruct accurate and complete indoor planar surfaces by directly optimizing a set of solid 3D plane primitives from posed multi-view images without any extra plane detection or matching operations. Benefiting from our differentiable planar primitive rendering, PlanarSplatting directly leverages monocular depth/normal cues from modern foundation models for optimization without plane annotations. Concurrent work, AlphaTablets [8], explores planar representations, which involves learning planar tablets equipped with alpha maps for shape modeling and pseudo-mesh construction for rendering. Our PlanarSplatting distinguishes itself from AlphaTablets by learning plane radii, providing a more explicit parameterization of 3D planes. As a result, we achieve a much more efficient rasterization scheme at the level of minutes while AlphaTablets takes about 2 hours per scene. + +Primitive-based Scene Representation. Optimizing explicit primitives such as points [11, 13, 34, 44], volumes [21], and superquadric [23] to represent the 3D scene has been studied for a long time. The core of these methods is to design a differentiable rendering process to optimize the attributes of primitives by gradient descent. Among them, a typical paradigm is to render images from primitives with splatting techniques which is realized with radial basis functions defined on the primitive (e.g., the Gaussian function). Inspired by these works, we proposed + +PlanarSplatting to optimize solid 3D plane primitives in a differentiable rendering manner with a novel shape-aware plane splating function to better fit the scene geometry. Benefiting from our efficient CUDA implementation, PlanarSplatting can reconstruct the accurate indoor planar scene within 3 minutes and empower recent Gaussian-based works (e.g., 3DGS [13]) to further improve the rendering quality of novel views as shown in Fig. 2. + +# 3. The Proposed PlanarSplatting + +As shown in Fig. 3, given a set of posed multi-view images, our PlanarSplatting can reconstruct the indoor planar scene from them by optimizing a set of learnable 3D planar primitives (Sec. 3.1). With the proposed Differentiable Planar Primitive Rendering (Sec. 3.2), we can explicitly learn these planar primitives from a coarse initialization to accurately recover the scene geometry. These optimized 3D planar primitives are then merged to achieve the 3D plane instances for the final planar reconstruction. + +# 3.1. Learnable Planar Scene Representation + +Since our key idea is directly optimizing explicit 3D primitives for planar scene reconstruction, we first present the representation of our learnable 3D planar primitives and then introduce how we initialize the scene with these planar primitives. + +Learnable Planar Primitive. As shown in Fig. 4, we formulate a planar primitive $\pi$ as a 3D rectangle which is equipped with several learnable parameters including the plane center $\mathbf{p}_{\pi} \in \mathbb{R}^3$ , the plane rotation $\mathbf{q}_{\pi} \in \mathbb{R}^4$ (in quaternion representation) and the plane radii $\mathbf{r}_{\pi}$ . Specifically, to improve the optimization flexibility of plane shape, we use the design of double direction plane radii (Double Radii) for $\mathbf{r}_{\pi}$ as: + +$$ +\mathbf {r} _ {\pi} = \left\{r _ {\pi} ^ {x +}, r _ {\pi} ^ {x -}, r _ {\pi} ^ {y +}, r _ {\pi} ^ {y -} \right\} \in \mathbb {R} _ {+} ^ {4}, \tag {1} +$$ + +where $r_{\pi}^{x+}, r_{\pi}^{x-}, r_{\pi}^{y+}, r_{\pi}^{y-}$ are the radii defined on the positive/negative direction of the X-axis/Y-axis of the rectangle as shown in Fig. 4. For the sake of description, we further define the positive direction of the X-axis and Y-axis of the 3D planar primitive as two orthogonal unit vectors $\mathbf{v}_{\pi}^{x}, \mathbf{v}_{\pi}^{y} \in \mathbb{R}^{3}$ which can be calculated as: + +$$ +\mathbf {v} _ {\pi} ^ {x} = \mathbf {R} (\mathbf {q} _ {\pi}) [ 1, 0, 0 ] ^ {\top}, \quad \mathbf {v} _ {\pi} ^ {y} = \mathbf {R} (\mathbf {q} _ {\pi}) [ 0, 1, 0 ] ^ {\top}, \tag {2} +$$ + +where $\mathbf{R}(\mathbf{q}_{\pi}) \in \mathbb{R}^{3 \times 3}$ means the rotation matrix of the quaternion $\mathbf{q}_{\pi}$ . Similarly, the normal of the planar primitive $\mathbf{n}_{\pi} \in \mathbb{R}^3$ can be calculated as: + +$$ +\mathbf {n} _ {\pi} = \mathbf {R} (\mathbf {q} _ {\pi}) [ 0, 0, 1 ] ^ {\top}. \tag {3} +$$ + +With these learnable parameters, the 3D planar primitive can be moved to align with the potential scene surface and deformed to fit the surface shape during optimization. + +![](images/447b99c3bbe6bc3176621af2c5497560f491541fce2e96b126a2b29bb12331e1.jpg) +Figure 3. Illustration of our proposed PlanarSplatting. Given a set of posed multi-view images of indoor scenes, our method renders depth and normal maps from 3D plane primitives. Then, with the supervision of monocular cues, these 3D plane primitives are gradually optimized to recover the scene geometry and finally merged to get the planar reconstruction result. + +![](images/35b7220572873baffb5a9a2f549ec7e342306f7efcbe2faba124ceba94a1acd5.jpg) +Figure 4. Representation of 3D plane primitive with learnable shape parameters including plane center, radii, and rotation. + +Scene Initialization. We use monocular depth from recent foundation models [10] to fast initialize our 3D planar primitives at the beginning of optimization. Specifically, we use depths from Metric3Dv2 [10] to get a very coarse scene geometry. Then we randomly sample 2,000 points on the coarse mesh to achieve the plane centers of our 3D planar primitives. The initial radii of each primitive $\pi$ is set to $0.5Dist(\pi)$ . Here, $Dist(\pi)$ means the distance closest to $\pi$ to its neighbors. We use the normal direction on the coarse mesh to initialize the plane rotation. At the bottom of Fig. 3, we show an example of our initial planar primitives. With such a coarse initialization, our PlanarSplatting can finally reconstruct the accurate and complete scene surface. + +# 3.2. Differentiable Planar Primitive Rendering + +To optimize the learnable planar primitives $\Pi = \{\pi_i\}_{i=1}^K$ , we introduce the differentiable planar primitive rendering with a carefully designed plane splating function which enables the planar primitives to accurately fit the scene geometry with the supervision only from 2D multi-view images. + +![](images/b5dfbdc26ede9d699243bd82c88bb907ec43a8b2147a12fa0fb486aa0b484ef8.jpg) +Figure 5. Illustration of the proposed plane splatting function. Naive Gaussian Splatting can not effectively approximate the boundary of our rectangular plane primitive (shown in black dashed border). In contrast, our proposed plane splatting function can approximate the boundary of the rectangle as the number of iterations increases, allowing our 3D planar primitives to better fit the surface of the scene. + +Ray-to-Plane Intersection. To project 3D planar primitives to the 2D image space, we first calculate the intersections between planar primitives and the rays emitted from image pixels. Specifically, given a ray $\mathbf{r} = \{\mathbf{o},\mathbf{d}\}$ starting from the camera center $\mathbf{o}\in \mathbb{R}^3$ with direction $\mathbf{d}\in \mathbb{R}^3$ , its intersection $\mathbf{x}_{\pi}^{\mathbf{r}}\in \mathbb{R}^3$ to one planar primitive $\pi$ is: + +$$ +\mathbf {x} _ {\pi} ^ {\mathbf {r}} = \mathbf {o} + \frac {\left(\mathbf {p} _ {\pi} - \mathbf {o}\right) \cdot \mathbf {n} _ {\pi}}{\mathbf {d} \cdot \mathbf {n} _ {\pi}} \mathbf {d}, \tag {4} +$$ + +where $\mathbf{p}_{\pi}$ and $\mathbf{n}_{\pi}$ are the center and the normal of the planar primitive $\pi$ . + +Plane Splatting Function. After achieving the ray-to-plane intersection $\mathbf{x}_{\pi}^{\mathbf{r}}$ , we calculate its splatting weight with our plane splatting function which will be used for rendering. + +A vanilla selection of the splatting function is the + +anisotropic Gaussian function: + +$$ +w \left(\mathbf {x} _ {\pi} ^ {\mathbf {r}}, \pi\right) = \exp \left(- \frac {1}{2} \left(\mathbf {x} _ {\pi} ^ {\mathbf {r}} - \mathbf {p} _ {\pi}\right) ^ {\top} \Sigma^ {- 1} \left(\mathbf {x} _ {\pi} ^ {\mathbf {r}} - \mathbf {p} _ {\pi}\right)\right), \tag {5} +$$ + +where $\mathbf{p}_{\pi}$ is the center of the planar primitive $\pi$ . The covariance matrix $\Sigma$ can be calculated like [11, 13]. However, as shown in Fig. 5 and Fig. 6, the Gaussian-based splatting function will make ambiguous boundaries of our rectangular planar primitive leading to the degeneration of reconstruction quality. + +Thus, we propose to calculate the splating weight with a novel rectangle-based plane splating function. To a given intersection $\mathbf{x}_{\pi}^{\mathbf{r}}$ between ray $\mathbf{r}$ and planar primitive $\pi$ , we first calculate its projection distance $\mathcal{P}_X, \mathcal{P}_Y \in \mathbb{R}$ to the X-axis and Y-axis of the planar primitive as: + +$$ +\mathcal {P} _ {X} = \left(\mathbf {x} _ {\pi} ^ {\mathbf {r}} - \mathbf {p} _ {\pi}\right) \cdot \mathbf {v} _ {\pi} ^ {x}, \quad \mathcal {P} _ {Y} = \left(\mathbf {x} _ {\pi} ^ {\mathbf {r}} - \mathbf {p} _ {\pi}\right) \cdot \mathbf {v} _ {\pi} ^ {y}. \quad (6) +$$ + +Then, we calculate the splating weight along the X-axis of the plane $\pi$ as: + +$$ +w _ {X} \left(\mathbf {x} _ {\pi} ^ {\mathbf {r}}\right) = \left\{ \begin{array}{l l} 2 \sigma \left(5 \lambda \left(r _ {\pi} ^ {x +} - | \mathcal {P} _ {X} |\right)\right), & \text {i f} \mathcal {P} _ {X} > 0, \\ 2 \sigma \left(5 \lambda \left(r _ {\pi} ^ {x -} - | \mathcal {P} _ {X} |\right)\right), & \text {o t h e r w i s e}, \end{array} \right. \tag {7} +$$ + +where $\sigma (\cdot)$ is the Sigmoid function and $\lambda$ is the hyperparameter to control the splattering weight. Similarly, we continue to calculate the splattering weight along the Y-axis of the plane $\pi$ as: + +$$ +w _ {Y} \left(\mathbf {x} _ {\pi} ^ {\mathbf {r}}\right) = \left\{ \begin{array}{l l} 2 \sigma \left(5 \lambda \left(r _ {\pi} ^ {y +} - | \mathcal {P} _ {Y} |\right)\right), & \text {i f} \mathcal {P} _ {Y} > 0, \\ 2 \sigma \left(5 \lambda \left(r _ {\pi} ^ {y -} - | \mathcal {P} _ {Y} |\right)\right), & \text {o t h e r w i s e}, \end{array} \right. \tag {8} +$$ + +where $r_{\pi}^{x + },r_{\pi}^{x - },r_{\pi}^{y + },r_{\pi}^{y - }$ are the radii parameters of the plane $\pi$ . At last, the final splatting weight can be calculated as: + +$$ +w \left(\mathbf {x} _ {\pi} ^ {\mathbf {r}}\right) = \left\{ \begin{array}{l l} w _ {X}, & \text {i f} w _ {X} < w _ {Y}, \\ w _ {Y}, & \text {o t h e r w i s e .} \end{array} \right. \tag {9} +$$ + +In Fig. 5, we show that with the increment of hyperparameter $\lambda$ , our plane splating function gradually approximates the shape of the rectangular plane primitive. In practice, we increase the value of $\lambda$ with an exponential function during optimization up to the maximum value of 300 as: + +$$ +\lambda = \min \left(2 0 e ^ {\left(- (1 - 0. 0 0 1 * i t e)\right)}, 3 0 0\right), \tag {10} +$$ + +where ite means the iteration number during optimization. + +Blending Composition. For all ray-to-plane intersections, we filter them with splating weight lower than 0.0001 and then sort the remaining intersections according to their depth from near to far. Then, $M$ nearest intersections of each ray are selected for rendering ( $M = 30$ in this paper). Denote the selected intersections of a ray $\mathbf{r}$ as $P^{\mathbf{r}} = \{\mathbf{x}_{\pi_{\tau(j)}}^{\mathbf{r}}\}_{j=1}^{M}$ . Here, $\tau(j)$ indicates the index of plane + +![](images/d609f4ee2501591c5c13063024999d195c9498e0f1d5caa7aa8de2cd9d516373.jpg) +(a) w/ GS Splitting + +![](images/d900e5a7255e421cb3ed95003d65a2540f5bbdfc7cb0a5620067d8653d14481f.jpg) +(b) w/ Plane Splatting +Figure 6. Reconstruction comparison with different splatting functions. 'w/' means 'with'. + +![](images/0fc6fd89f60d5c112dce3c6820bbc905110efc178afc07d7fe501693f4f81889.jpg) +(c) GT Mesh + +among all planar primitives. At last, we render the depth and normal map of a certain image $\mathbf{I}$ as: + +$$ +\mathbf {D} _ {\text {r e n d e r}} ^ {\Pi} (\mathbf {r}) = \sum_ {j = 1} ^ {M} T _ {j} w \left(\mathbf {x} _ {\pi_ {\tau (j)}} ^ {\mathbf {r}}\right) t _ {j}, \tag {11} +$$ + +$$ +\mathbf {N} _ {\text {r e n d e r}} ^ {\Pi} (\mathbf {r}) = \sum_ {j = 1} ^ {M} T _ {j} w \left(\mathbf {x} _ {\pi_ {\tau (j)}} ^ {\mathbf {r}}\right) \mathbf {n} _ {\pi_ {\tau (j)}}, \tag {12} +$$ + +where + +$$ +T _ {j} = \prod_ {i = 1} ^ {j - 1} (1 - w \left(\mathbf {x} _ {\pi_ {\tau (i)}} ^ {\mathbf {r}}\right)). \tag {13} +$$ + +Here $t_j$ is the depth of the intersection and $\mathbf{n}_{\pi_{\tau(j)}}$ is the normal of planar primitive $\pi_{\tau(j)}$ . To supervise the rendered depth and normal map, we use the pretrained model of Metric3Dv2 [10] to predict the depth map $\mathbf{D}_{\mathrm{pre}}$ and normal map $\mathbf{N}_{\mathrm{pre}}$ to serve as pseudo labels. Besides, we also test the normal map predicted by Omnidata [3]. Finally, the render loss can be calculated as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {r e n d e r}} ^ {\Pi} = \alpha_ {1} \sum_ {\mathbf {r} \in \mathbf {I}} \| 1 - \mathbf {N} _ {\text {r e n d e r}} ^ {\Pi} (\mathbf {r}) ^ {\top} \mathbf {N} _ {\text {p r e}} (\mathbf {r})) \| _ {1} + \\ \alpha_ {1} \sum_ {\mathbf {r} \in \mathbf {I}} \| \mathbf {N} _ {\text {r e n d e r}} ^ {\Pi} (\mathbf {r}) - \mathbf {N} _ {\text {p r e}} (\mathbf {r})) \| _ {1} + \tag {14} \\ \alpha_ {2} \sum_ {\mathbf {r} \in \mathbf {I}} \| (\mathbf {D} _ {\mathrm {r e n d e r}} ^ {\Pi} (\mathbf {r})) - \mathbf {D} _ {\mathrm {p r e}} (\mathbf {r}) \| _ {1}, \\ \end{array} +$$ + +where $\alpha_{1} = 2.0$ , $\alpha_{2} = 2.0$ , $\mathbf{r}$ is the ray/pixel emitted from image $\mathbf{I}$ . + +# 3.3. Optimization + +Loss Function. To each scene, we optimize our Planar-Splatting with the Adam optimizer [14] for 5,000 iterations on the ScanNetV2 [2] dataset and 10,000 iterations on the ScanNet++ [40] dataset using the loss as shown in Eq. (14). + +Plane Splitting. During optimization, we introduce a splitting operation on planes according to the gradients of their radii to better fit the scene geometry. If the average radii gradients on the X-axis $(r^{x+}$ and $r^{x-})$ of the plane are greater than 0.2, we split the plane along the Y-axis. Similarly, we split the planes along the X-axis, if their radii gradients on the Y-axis $(r^{y+}$ and $r^{y-})$ are larger than 0.2. We conduct the splitting operation every 1,000 iterations. + +Table 1. Quantitative comparison of planar reconstruction results on the ScanNetV2 [2] dataset. 'P. Ann.' means using 2D/3D plane annotations in training stages. + +
MethodP. Ann.GeometrySegmentationPlanar
Chamfer ↓F-score ↑VOI ↓RI ↑SC ↑Fidelity ↓Acc ↓Chamfer ↓
PlanarRecon [37]✓9.8943.473.2010.9190.40518.8616.2117.53
AirPlanes [35]✓5.3064.922.2680.9570.5688.767.988.37
2DGS [11] + RANSACX14.1531.334.0300.9240.25740.0214.7727.40
Metric3Dv2 [10] + RANSACX5.9365.452.7440.9270.4889.2911.3610.33
SR [27] + RANSACX5.4065.452.5070.9460.5159.4210.139.78
Ours@OmnidataX4.8368.852.5020.9480.5326.6411.769.20
Ours@Metric3Dv2X4.2373.762.3100.9580.5605.7210.658.19
+ +Table 2. Quantitative comparison of planar reconstruction results on the ScanNet++ [40] dataset. 'P. Ann.' means using 2D/3D plane annotations in training stages. + +
MethodP. Ann.GeometrySegmentationPlanar
Chamfer ↓F-score ↑VOI ↓RI ↑SC ↑Fidelity ↓Acc ↓Chamfer ↓
PlanarRecon [37]✓17.8531.103.5420.9190.36734.4419.3526.90
AirPlanes [35]✓13.7532.582.8590.9410.47028.1612.5820.37
2DGS [11] + RANSACX20.3926.464.4560.9270.24155.9016.8836.39
Metric3Dv2 [10] + RANSACX11.1441.543.0210.9300.44127.5010.8119.16
SR [27] + RANSACX13.1535.933.0130.9380.44230.2511.6220.94
Ours@OmnidataX9.3347.042.7720.9460.52317.2412.2614.75
Ours@Metric3Dv2X9.0150.682.7820.9440.52117.1712.3614.77
+ +Plane Merge. After optimization, we greedily merge the adjacent 3D plane primitives with normal angle error lower than $25^{\circ}$ and offset distance error lower than $0.1m$ . Here, offset means the projection distance from the scene center to the plane surface. + +CUDA Implementation. For fast optimization of solid 3D planar primitives, we implement the forward and backward process of our Differentiable Planar Primitive Rendering with CUDA, which enables our PlanarSplatting to reconstruct one scene within 3 minutes. We will release the code of our CUDA implementation for solid 3D planar primitives optimization after publication. + +# 4. Experiments + +# 4.1. Dataset, Metrics and Baselines + +Datasets. We evaluate our PlanarSplatting on two large indoor datasets including ScanNetV2 [2] and ScanNet++ [40] which provide posed images. On the ScanNetV2 dataset, we use the test split according to AirPlanes [35] which includes 100 scenes with ground truth 3D plane annotations provided by [18]. On the ScanNet++ dataset, we randomly select 30 scenes for evaluation and extract 3D plane annotations from the ground truth meshes like [18]. + +Evaluation Metrics. According to PlanarRecon [37], we evaluate the geometry quality of all reconstructed 3D planes with Chamfer Distance and F-score. Following AirPlanes [35], we also evaluate the reconstruction quality of Top-20 largest planes from the ground truth and use the metrics including Planar Fidelity, Planar Accuracy, and Planar Chamfer. To evaluate 3D plane segmentation, we use + +the metrics including Variation of Information (VOI), Rand Index (RI), and Segmentation Covering (SC) like [18, 37]. + +Baselines. Since our method is purely built upon geometry cues, we mainly compare our PlanarSplatting with those geometry-based methods including SR+RANSAC [27], Metric3Dv2+RANSAC [10] and 2DGS+RANSAC [11]. These baselines build dense scene meshes from multi-view images at first and then extract 3D planes with the RANSAC algorithm from the reconstructed meshes. We use the RANSAC implementation provided by AirPlanes [35] for all these geometry-based baselines. Besides, we also compare our PlanarSplatting with some plane annotation based methods that use 2D/3D plane labels/priors in their training stage including PlanarRecon [37] and AirPlanes [35], and report their results for reference. + +# 4.2. Comparisons with Baselines + +Quantitative Results. We first evaluate our PlanarSplatting on the ScanNetV2 [2] dataset. Specifically, we evaluate two versions of our PlanarSplatting, utilizing normal maps from different pretrained models for supervision: Metric3Dv2 [10] and Omnidata [3], respectively. As shown in Tab. 1, our method achieves the best geometry performance in the metric of Chamfer and F-score compared to all baselines. To the results of Segmentation and Planar, our method outperforms geometry-based baselines in most metrics and is comparable to the state-of-the-art AirPlanes which uses plane embeddings trained on the same ScanNetV2 dataset. In Tab. 2, we show the results on the ScanNet++ [40] dataset. Benefiting from our novel design for differentiable plane rendering and + +![](images/0ee8655e0fcd8dbd7e212abbbe46f584c8dfa0127e044df27e2adac3fc79a426.jpg) + +![](images/087c91d570fb06d8e10fc2d319e95eb4f2b9b5723a70934bcb8e3946112ac1bc.jpg) + +![](images/892d7ded0fdf2f3d18ec68b91a7330b9f365816a231e449bf600e68677b3ec5e.jpg) + +![](images/da5de82a644b7150ff18daa56e0511f75ae9d0726e8bac4c267911ceb7178d4d.jpg) + +![](images/a4097cac7631077f509a2caa59c63170033411d36b9859551fbc664496aa43be.jpg) + +![](images/f3e4491a28c4e4b0d5190aa4816425a298cccad82b1ecca8548b3405017158d8.jpg) + +![](images/d7c4c1acfe3e66c27dd96655c5d682bcc44b16aa1979fb331ced419ae71dc75d.jpg) + +![](images/6578c113348bbfda366dd36c9d4e9f5604b5318858afedc87e3ae402696d8048.jpg) + +![](images/7035b4089af551553830c3fe32e725dfcf28f9fb46405490618e6670b64afb61.jpg) + +![](images/2212230165801d77165153c0cfc6d97b1165a0932d68e2e939bd106736becfab.jpg) + +![](images/86a40c7a3b4d65f485daa94a0525111123ff34d3335f27934c44cad584a03cdf.jpg) + +![](images/55ef8f1c13195e3561c5e8ce0ad07ab908815af9e058669b6aa66fb4543f5fb5.jpg) + +![](images/867d41b776217bfa3a818d9b59086e370c1f89e8d0735f82cb395f24c69333c7.jpg) + +![](images/062632401fc894e0bdd10083f67c0615f7aa9ed1c7853df05392c083d8d6706c.jpg) + +![](images/8b69d97ab7aa638401571b550c41a3caedf16d20446824f28652ab87f89f2e0c.jpg) + +![](images/f64063875bec07baae356f33420c6b621ee6436333f04a8ee543ed2c61fe8682.jpg) + +![](images/1916fe74eeb04f44110a79719014c03925d0a996dda6753e1605faf9b226e3af.jpg) +(a) PlanarRecon [37] + +![](images/a453f1e1cc145e922161c0554e42a454937035287caf033fb21fe3f83da8edef.jpg) +(b) AirPlanes [35] + +![](images/1249e5a5b2cf1181a6b7449c1648c21409bc4849354ccf63f956d508358845aa.jpg) +(c) Ours + +![](images/94530101150dcfc2422cf01a4ed422e2492d771d4de9df59440eff52e2840fe9.jpg) +(d) Ground Truth +Figure 7. Qualitative comparison on the ScanNetV2 [2] (rows 1-4) and ScanNet++ [40] (last row) datasets. + +optimization via monocular cues provided by modern foundation models, our PlanarSplatting still achieves the best Geometry performance among all baselines. Furthermore, our method outperforms both geometry-based and plane annotation based methods in most metrics of Segmentation and Planar, demonstrating the robustness and superiority of our proposed PlanarSplatting. + +Qualitative Results. In Fig. 7, we show the comparisons of PlanarRecon [37], AirPlanes [35] and our PlanarSplatting with normal supervision of Omnidata [3]. PlanarRecon reconstructs coarse 3D planes from input multi-view images while our PlanarSplatting accurately reconstructs scene geometry via directly optimizing 3D plane primitives in the whole 3D space. With the help of a plane embedding model, AirPlanes achieves more semantic plane segmentation from + +co-plane regions. However, as shown in the zoom-in results in Fig. 7, AirPlanes loses many geometry details in the scene. In contrast, our PlanarSplatting successfully reconstructs detailed structures such as the chairs and the legs of the bed, resulting in both accuracy and completeness in results. These comparisons effectively indicate the superiority of our method. + +# 4.3. Ablation Studies + +We verify the design of our PlanarSplatting on 10 scenes randomly selected from the ScanNetV2 [2] dataset. We optimize 3D plane primitives with 5,000 iterations with the normal supervision of Omnidata [3] in default. + +Plane Initialization. We first assess the sensitivity of the plane initialization strategy of our PlanarSplatting. We ablate the used Metric3Dv2 [10] initialization to the sphere + +Table 3. Ablation studies of the proposed PlanarSplatting on the ScanNetV2 dataset. 'w/' means 'with' and 'w/o' means 'without'. + +
Chamfer ↓F-score ↑VOI ↓
w/ Sphere Init.8.3957.533.073
w/ Sphere Init. (30K)4.5869.912.522
w/o Double Radii4.7370.252.491
w/o Plane Splitting4.6970.302.477
Ours (Full)4.6671.302.473
+ +Table 4. Quantitative comparison of novel view synthesis on the ScanNetV2 [2] dataset. 'Plane' means time for our plane optimization. 'GS' means time for 2D/3D Gaussian optimization. #P' means the average number of Gaussian points. + +
PSNR ↑SSIM ↑LPIPS ↓Avg. Time (min)#P
PlaneGSTotal
3DGS [13]24.4170.7810.321-12.212.21.27M
Ours+3DGS [13]25.4710.8160.2962.53.15.60.37M
2DGS [11]24.7660.7960.323-14.214.20.76M
Ours+2DGS [11]25.3800.8130.2962.54.87.30.37M
+ +initialization, in which 3D plane primitives are initialized by setting their radii as 0.05 and are placed on the bounding sphere of the scene with their normals pointing to the center of the scene. As reported in the top two rows of Tab. 3, initialization with Metric3Dv2 [10] mainly improves our method to be faster converged, while sphere initialization could also obtain promising results with more iterations. + +Plane Radii. We then evaluate the design of double-direction plane radii (Double Radii) by replacing it with the vanilla single-direction plane radii. It means that we learn one radius at each axis of the plane primitive. As shown in Tab. 3, optimizing without Double Radii leads to a decrease in the performance of geometry and segmentation, demonstrating the effectiveness of our Double Radii. + +Plane Splitting. As shown in the last two rows of Tab. 3, applying the Plane Splitting in optimization can further improve the performance in both geometry and segmentation. + +# 4.4. PlanarSplatting for Novel View Synthesis + +Benefiting from our accurate and fast planar reconstruction, we show that our PlanarSplatting can seamlessly integrate with recent Gaussian Splatting methods for more efficient and better quality novel view synthesis in indoor scenes. We select two typical methods (3DGS [13] and 2DGS [11]) for evaluation and compare them with two variants including 'Ours+3DGS' and 'Ours+2DGS'. Specifically, for our two variants, we directly sample points from our reconstructed 3D plane primitives to use as the initialization of these Gaussian Splatting methods. Different from the original 3DGS and 2DGS, we fix the position of 3D points and exclude the densification operation when optimizing our two variants. For the baselines of 3DGS and 2DGS, we initialize them with SfM and run with densification to achieve the best performance. We conduct the experiments + +![](images/4229a7f08aa8f9b36891862c2f620d7c1356dce12a0b05c09d435859f0dacc50.jpg) + +![](images/a0ac663c9fc8963e1f882135e47e455848d5fcb04147030b99df4253518b5639.jpg) + +![](images/861144c33b8304525270ef59a448c080c172c483ae21b137e1b44ceb0425e4fb.jpg) +(a) 3DGS [13] +(c) Ours+3DGS [13] +Figure 8. Comparison of novel view synthesis on the ScanNetV2 dataset [2]. With initialization from our planar reconstruction, we significantly improve the rendering result of 3DGS [13]. + +![](images/5762a9ab220771dc100ac680dc36ab7e552770b398ec31489a577cc2eb4ecb7e.jpg) +(b) 2DGS [11] +(d) Ground Truth + +using the same scenes as those used in the ablation study. As shown in Tab. 4, our two variants 'Ours+3DGS' and 'Ours+2DGS' significantly outperform the original 3DGS and 2DGS in all metrics with even less optimization time and Gaussian points, demonstrating the superiority and potential of our PlanarSplatting. In Fig. 8, we further show that 'Ours+3DGS' effectively improves the rendering quality on the scene from the ScanNetV2 dataset. + +# 4.5. Limitations and Future Work + +Although our PlanarSplatting can reconstruct the accurate indoor planar surface, it is not suitable for complex shapes such as curved surfaces. We leave this challenging problem in our future work for more flexible geometric modeling. + +# 5. Conclusion + +In this paper, we present PlanarSplatting, a novel approach for multi-view 3D reconstruction of indoor scenes. By formulating the problem through differentiable rendering with plane splatting, we demonstrate the powerful capabilities of 3D planar representation for both accurate geometry reconstruction and compact structural scene modeling. Our efficient CUDA implementation enables fast 3D surface reconstruction, allowing comprehensive evaluation across over 100 scenes within hours using a single GPU. Furthermore, the seamless integration of PlanarSplatting with Gaussian Splatting significantly enhances both the quality and efficiency of indoor novel view synthesis, highlighting the broader potential of our approach for indoor scene understanding. + +# Acknowledgments + +This work was supported by Ant Group Postdoctoral Programme. The authors would love to thank the anonymous reviewers and area chairs for their constructive comments. + +# References + +[1] Samir Agarwala, Linyi Jin, Chris Rockwell, and David F. Fouhey. Planeformers: From sparse view planes to 3d reconstruction. In Eur. Conf. Comput. Vis., pages 192-209, 2022. 3 +[2] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2432-2443, 2017. 1, 2, 5, 6, 7, 8 +[3] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Int. Conf. Comput. Vis., pages 10766-10776, 2021. 2, 5, 6, 7 +[4] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Manhattan-world stereo. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1422-1429, 2009. 3 +[5] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Reconstructing building interiors from images. In Int. Conf. Comput. Vis., pages 80-87, 2009. +[6] David Gallup, Jan-Michael Frahm, and Marc Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1418-1425, 2010. 3 +[7] Abhinav Gupta, Alexei A. Efros, and Martial Hebert. Blocks world revisited: Image understanding using qualitative geometry and mechanics. In *Eur. Conf. Comput. Vis.*, pages 482-496, 2010. 2 +[8] Yuze He, Wang Zhao, Shaohui Liu, Yubin Hu, Yushi Bai, Yu-Hui Wen, and Yongjin Liu. Alphatables: A generic plane representation for 3d planar reconstruction from monocular videos. In Adv. Neural Inform. Process. Syst., 2024. 3 +[9] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Comput. Vis. Image Underst., 157:167-178, 2017. 2 +[10] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):10579-10596, 2024. 2, 4, 5, 6, 7, 8 +[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH Conference Papers, page 32, 2024. 2, 3, 5, 6, 8 +[12] Linyi Jin, Shengyi Qian, Andrew Owens, and David F. Fouhey. Planar surface reconstruction from sparse views. In Int. Conf. Comput. Vis., pages 12971-12980, 2021. 2, 3 +[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 2, 3, 5, 8 +[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Int. Conf. Learn. Represent., 2015. 5 + +[15] Pierre-Alain Langlois, Alexandre Boulch, and Renaud Marlet. Surface reconstruction from 3d line segments. In Int. Conf. 3D Vis., pages 553-563, 2019. 3 +[16] David C. Lee, Martial Hebert, and Takeo Kanade. Geometric reasoning for single image structure recovery. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2136-2143, 2009. 2, 3 +[17] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single RGB image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2579-2588, 2018. 3 +[18] Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz. Planercnn: 3d plane detection and reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4450-4459, 2019. 2, 3, 6 +[19] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8655-8665. IEEE, 2022. 2 +[20] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21445-21455. IEEE, 2023. 2 +[21] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhöfer, Yaser Sheikh, and Jason M. Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 40(4):59:1-59:13, 2021. 3 +[22] Branislav Micusik, Horst Wildenauer, and Markus Vincze. Towards detection of orthogonal planes in monocular images of indoor environments. In Int. Conf. on Robot. and Auto., pages 999-1004, 2008. 3 +[23] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable blocks world: Qualitative 3d decomposition by rendering primitives. In Adv. Neural Inform. Process. Syst., 2023. 2, 3 +[24] Aron Monszpart, Nicolas Mellado, Gabriel J. Brostow, and Niloy J. Mitra. Rapter: rebuilding man-made scenes with regular arrangements of planes. ACM Trans. Graph., 34(4): 103:1-103:12, 2015. 3 +[25] Jann Poppinga, Narunas Vaskevicius, Andreas Birk, and Kaustubh Pathak. Fast plane detection and polygonalization in noisy 3d range images. In Int. Conf. Intell. Robots and Syst., pages 3378-3383, 2008. 2, 3 +[26] Yiming Qian, Srikumar Ramalingam, and James H. Elder. LS3D: single-view gestalt 3d surface reconstruction from Manhattan line segments. In *Asian Conf. Comput. Vis.*, pages 399-416, 2018. 3 +[27] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplerecon: 3d reconstruction without 3d convolutions. In Eur. Conf. Comput. Vis., pages 1-19, 2022. 6 +[28] Jingjia Shi, Shuaifeng Zhi, and Kai Xu. Planerectr: Unified query learning for 3d plane recovery from a single view. In Int. Conf. Comput. Vis., pages 9343-9352, 2023. 3 +[29] Sudipta N. Sinha, Drew Steedly, and Richard Szeliski. Piecewise planar stereo for image-based rendering. In Int. Conf. Comput. Vis., 2009. 3 + +[30] Christiane Sommer, Yumin Sun, Leonidas J. Guibas, Daniel Cremers, and Tolga Birdal. From planes to corners: Multipurpose primitive detection in unorganized 3d point clouds. IEEE Robotics Autom. Lett., 5(2):1764-1771, 2020. 2, 3 +[31] Bo Sun and Philippos Mordohai. Oriented point sampling for plane detection in unorganized point clouds. In Int. Conf. on Robot. and Auto., pages 2917-2923, 2019. 3 +[32] Bin Tan, Nan Xue, Song Bai, Tianfu Wu, and Gui-Song Xia. Planetr: Structure-guided transformers for 3d plane recovery. In Int. Conf. Comput. Vis., pages 4166-4175, 2021. 3 +[33] Bin Tan, Nan Xue, Tianfu Wu, and Gui-Song Xia. NOPE-SAC: neural one-plane RANSAC for sparse-view planar 3d reconstruction. IEEE Trans. Pattern Anal. Mach. Intell., 45 (12):15233-15248, 2023. 2, 3 +[34] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Trans. Graph., 38 (6):230:1-230:14, 2019. 3 +[35] Jamie Watson, Filippo Aleotti, Mohamed Sayed, Zawar Qureshi, Oisin Mac Aodha, Gabriel J. Brostow, Michael Firman, and Sara Vicente. Airplanes: Accurate plane estimation via 3d-consistent embeddings. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5270-5280, 2024. 1, 3, 6, 7 +[36] Jianxiong Xiao and Yasutaka Furukawa. Reconstructing the world's museums. Int. J. Comput. Vis., 110(3):243-258, 2014. 3 +[37] Yiming Xie, Matheus Gadelha, Fengting Yang, Xiaowei Zhou, and Huaizu Jiang. Planarrecon: Realtime 3d plane detection and reconstruction from posed monocular videos. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6209-6218, 2022. 1, 2, 3, 6, 7 +[38] Sheng Xu, Ruisheng Wang, Hao Wang, and Ruigang Yang. Plane segmentation based on the optimal-vector-field in lidar point clouds. IEEE Trans. Pattern Anal. Mach. Intell., 43 (11):3991-4007, 2021. 2 +[39] Nan Xue, Bin Tan, Yuxi Xiao, Liang Dong, Gui-Song Xia, Tianfu Wu, and Yujun Shen. NEAT: distilling 3d wireframes from neural attraction fields. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19968-19977. IEEE, 2024. 2 +[40] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Int. Conf. Comput. Vis., pages 12-22, 2023. 2, 5, 6, 7 +[41] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from A single image. In Int. Conf. Comput. Vis., 2023. 2 +[42] Zehao Yu, Jia Zheng, Dongze Lian, Zihan Zhou, and Shenghua Gao. Single-image piece-wise planar 3d reconstruction via associative embedding. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1029-1037, 2019. 2, 3 +[43] Jiahui Zhang, Jinfu Yang, Fuji Fu, and Jiaqi Ma. Planeac: Line-guided planar 3d reconstruction based on self-attention and convolution hybrid model. Pattern Recog., 153:110519, 2024. 3 + +[44] Yanshu Zhang, Shichong Peng, Alireza Moazeni, and Ke Li. Papr: Proximity attention point rendering. In Adv. Neural Inform. Process. Syst., 2023. 3 \ No newline at end of file diff --git a/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/images.zip b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bdf607e46b0ccc310008370541b9f2c13019c91b --- /dev/null +++ b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:146e027c18bf7752a75f5460220965ee5b242aad433a01359a66d99320e963f0 +size 825670 diff --git a/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/layout.json b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..af37132f26e7afb4593d1cf9191a90c2e9675782 --- /dev/null +++ b/2025/PlanarSplatting_ Accurate Planar Surface Reconstruction in 3 Minutes/layout.json @@ -0,0 +1,9702 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 523, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 523, + 120 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 523, + 120 + ], + "type": "text", + "content": "PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 184, + 142, + 425, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 142, + 425, + 158 + ], + "spans": [ + { + "bbox": [ + 184, + 142, + 425, + 158 + ], + "type": "text", + "content": "Bin Tan" + }, + { + "bbox": [ + 184, + 142, + 425, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 184, + 142, + 425, + 158 + ], + "type": "text", + "content": " Rui Yu" + }, + { + "bbox": [ + 184, + 142, + 425, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 184, + 142, + 425, + 158 + ], + "type": "text", + "content": " Yujun Shen" + }, + { + "bbox": [ + 184, + 142, + 425, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 184, + 142, + 425, + 158 + ], + "type": "text", + "content": " Nan Xue" + }, + { + "bbox": [ + 184, + 142, + 425, + 158 + ], + "type": "inline_equation", + "content": "^{\\dagger,1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 170, + 402, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 170, + 402, + 186 + ], + "spans": [ + { + "bbox": [ + 209, + 170, + 402, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 209, + 170, + 402, + 186 + ], + "type": "text", + "content": "Ant Group " + }, + { + "bbox": [ + 209, + 170, + 402, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 209, + 170, + 402, + 186 + ], + "type": "text", + "content": "University of Louisville" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 217, + 113, + 400 + ], + "blocks": [ + { + "bbox": [ + 56, + 217, + 113, + 400 + ], + "lines": [ + { + "bbox": [ + 56, + 217, + 113, + 400 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 113, + 400 + ], + "type": "image", + "image_path": "b827c1d16c0df3e53f13f20d1d328eaa178a00e3a8e2a42aad9c6a6c185dbaae.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 401, + 113, + 452 + ], + "blocks": [ + { + "bbox": [ + 56, + 401, + 113, + 452 + ], + "lines": [ + { + "bbox": [ + 56, + 401, + 113, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 401, + 113, + 452 + ], + "type": "image", + "image_path": "b3dddd3f4530dbc79fd5a16b7b9fbb9a27311b9783876e02963049084f9d3489.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 453, + 100, + 474 + ], + "lines": [ + { + "bbox": [ + 75, + 453, + 100, + 474 + ], + "spans": [ + { + "bbox": [ + 75, + 453, + 100, + 474 + ], + "type": "text", + "content": "Input Images" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 118, + 217, + 230, + 312 + ], + "blocks": [ + { + "bbox": [ + 118, + 217, + 230, + 312 + ], + "lines": [ + { + "bbox": [ + 118, + 217, + 230, + 312 + ], + "spans": [ + { + "bbox": [ + 118, + 217, + 230, + 312 + ], + "type": "image", + "image_path": "0451cdae4cbd7334c7243a32e757d3d445c4ee3ec651a55faca10f7db8636092.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 130, + 314, + 223, + 380 + ], + "blocks": [ + { + "bbox": [ + 130, + 314, + 223, + 380 + ], + "lines": [ + { + "bbox": [ + 130, + 314, + 223, + 380 + ], + "spans": [ + { + "bbox": [ + 130, + 314, + 223, + 380 + ], + "type": "image", + "image_path": "79aa4b4f4e46379333f2b65e99ba502bed86d79f234d3637c59a33aef65efb4d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 130, + 381, + 223, + 464 + ], + "blocks": [ + { + "bbox": [ + 130, + 381, + 223, + 464 + ], + "lines": [ + { + "bbox": [ + 130, + 381, + 223, + 464 + ], + "spans": [ + { + "bbox": [ + 130, + 381, + 223, + 464 + ], + "type": "image", + "image_path": "4916d6cbdba2112540472625f4b4183a3e4ef568bbb2e0857eb68ba39c726d95.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 464, + 199, + 474 + ], + "lines": [ + { + "bbox": [ + 151, + 464, + 199, + 474 + ], + "spans": [ + { + "bbox": [ + 151, + 464, + 199, + 474 + ], + "type": "text", + "content": "PlanarRecon" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 487, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 55, + 487, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 555, + 521 + ], + "type": "text", + "content": "Figure 1. We introduce PlanarSplatting, a fast and accurate optimization-based planar reconstruction method for indoor scenes. We show our planar reconstruction results on the ScanNetV2 [2] dataset achieved in 3 minutes. Compared to prior art PlanarRecon [37] and AirPlanes [35], our PlanarSplatting reconstructs more complete and detailed 3D planes." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 230, + 220, + 336, + 307 + ], + "blocks": [ + { + "bbox": [ + 230, + 220, + 336, + 307 + ], + "lines": [ + { + "bbox": [ + 230, + 220, + 336, + 307 + ], + "spans": [ + { + "bbox": [ + 230, + 220, + 336, + 307 + ], + "type": "image", + "image_path": "1781cecd8eb6918046d57c32821de614ad47b67aa2e9518f3034e52851b0e9a3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 239, + 314, + 332, + 380 + ], + "blocks": [ + { + "bbox": [ + 239, + 314, + 332, + 380 + ], + "lines": [ + { + "bbox": [ + 239, + 314, + 332, + 380 + ], + "spans": [ + { + "bbox": [ + 239, + 314, + 332, + 380 + ], + "type": "image", + "image_path": "e95653c421b1bbb220c8ce549fef0c28c7d86d8bbcd139a8c1464ff140bc80be.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 239, + 381, + 331, + 464 + ], + "blocks": [ + { + "bbox": [ + 239, + 381, + 331, + 464 + ], + "lines": [ + { + "bbox": [ + 239, + 381, + 331, + 464 + ], + "spans": [ + { + "bbox": [ + 239, + 381, + 331, + 464 + ], + "type": "image", + "image_path": "468b392657c3a7cb0b2a5dae055a6b06df7e94753aca59826605cccd7f5ff261.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 464, + 303, + 474 + ], + "lines": [ + { + "bbox": [ + 266, + 464, + 303, + 474 + ], + "spans": [ + { + "bbox": [ + 266, + 464, + 303, + 474 + ], + "type": "text", + "content": "AirPlanes" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 337, + 218, + 448, + 312 + ], + "blocks": [ + { + "bbox": [ + 337, + 218, + 448, + 312 + ], + "lines": [ + { + "bbox": [ + 337, + 218, + 448, + 312 + ], + "spans": [ + { + "bbox": [ + 337, + 218, + 448, + 312 + ], + "type": "image", + "image_path": "352fddffc5b6279c78776f2f43f6aadfac74c4d1c15e1a3e5f01be4955fcabc2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 348, + 314, + 441, + 380 + ], + "blocks": [ + { + "bbox": [ + 348, + 314, + 441, + 380 + ], + "lines": [ + { + "bbox": [ + 348, + 314, + 441, + 380 + ], + "spans": [ + { + "bbox": [ + 348, + 314, + 441, + 380 + ], + "type": "image", + "image_path": "74bee36ad248c719a6aa81152b3bff713ad696a2e5053bdbb878d5ec3626f41a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 349, + 381, + 441, + 464 + ], + "blocks": [ + { + "bbox": [ + 349, + 381, + 441, + 464 + ], + "lines": [ + { + "bbox": [ + 349, + 381, + 441, + 464 + ], + "spans": [ + { + "bbox": [ + 349, + 381, + 441, + 464 + ], + "type": "image", + "image_path": "2541fda366f5214e314e57546836a2d3c5aded25a626ff05d9f762adb5b64aba.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 464, + 419, + 475 + ], + "lines": [ + { + "bbox": [ + 370, + 464, + 419, + 475 + ], + "spans": [ + { + "bbox": [ + 370, + 464, + 419, + 475 + ], + "type": "text", + "content": "Ours (3min)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 449, + 218, + 555, + 313 + ], + "blocks": [ + { + "bbox": [ + 449, + 218, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 449, + 218, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 449, + 218, + 555, + 313 + ], + "type": "image", + "image_path": "1a4123ad3d6a2e9ea0e457cba4140e1753e6f8b69f01afdb72c448e61339d145.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 459, + 314, + 551, + 380 + ], + "blocks": [ + { + "bbox": [ + 459, + 314, + 551, + 380 + ], + "lines": [ + { + "bbox": [ + 459, + 314, + 551, + 380 + ], + "spans": [ + { + "bbox": [ + 459, + 314, + 551, + 380 + ], + "type": "image", + "image_path": "9765116973bfe08b642c5744ecb8719655ee8b6df897d50875c60d2e9a7a2384.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 459, + 381, + 551, + 464 + ], + "blocks": [ + { + "bbox": [ + 459, + 381, + 551, + 464 + ], + "lines": [ + { + "bbox": [ + 459, + 381, + 551, + 464 + ], + "spans": [ + { + "bbox": [ + 459, + 381, + 551, + 464 + ], + "type": "image", + "image_path": "d2f33164e35e29dcb882b5948ad8c26a7f9dcfb2317e8e72ef6760342e42cd16.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 464, + 529, + 474 + ], + "lines": [ + { + "bbox": [ + 479, + 464, + 529, + 474 + ], + "spans": [ + { + "bbox": [ + 479, + 464, + 529, + 474 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 151, + 530, + 200, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 530, + 200, + 542 + ], + "spans": [ + { + "bbox": [ + 151, + 530, + 200, + 542 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 54, + 556, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 556, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 556, + 297, + 689 + ], + "type": "text", + "content": "This paper presents PlanarSplatting, an ultra-fast and accurate surface reconstruction approach for multi-view indoor images. We take the 3D planes as the main objective due to their compactness and structural expressiveness in indoor scenes, and develop an explicit optimization framework that learns to fit the expected surface of indoor scenes by splatting the 3D planes into 2.5D depth and normal maps. As our PlanarSplatting operates directly on the 3D plane primitives, it eliminates the dependencies on 2D/3D plane detection and plane matching/tracking for planar surface reconstruction. Furthermore, with the essential" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 531, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 556, + 675 + ], + "type": "text", + "content": "merits of plane-based representation coupled with CUDA-based implementation of planar splatting functions, Planar-Splatting reconstructs an indoor scene in 3 minutes while having significantly better geometric accuracy. Thanks to our ultra-fast reconstruction speed, the largest quantitative evaluation on the ScanNet and ScanNet++ datasets over hundreds of scenes clearly demonstrated the advantages of our method. We believe that our accurate and ultrafast planar surface reconstruction method will be applied in the structured data curation for surface reconstruction in the future. The code of our CUDA implementation will be publicly available." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 148, + 712 + ], + "type": "text", + "content": "†Corresponding author." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1190" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 161, + 225 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 161, + 225 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 161, + 225 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 161, + 225 + ], + "type": "image", + "image_path": "eb3d1f06b998f755890b6cac6dfe7e9370a3513152a72585733a531311d3cd2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 229, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 555, + 253 + ], + "type": "text", + "content": "Figure 2. Novel View Synthesis: Seamless integration of our PlanarSplatting with recent Gaussian Splatting methods (e.g., 3DGS [13] and 2DGS [11]) achieves indoor-scene rendering with improved performance and significant optimization time reduction simultaneously." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 161, + 70, + 256, + 225 + ], + "blocks": [ + { + "bbox": [ + 161, + 70, + 256, + 225 + ], + "lines": [ + { + "bbox": [ + 161, + 70, + 256, + 225 + ], + "spans": [ + { + "bbox": [ + 161, + 70, + 256, + 225 + ], + "type": "image", + "image_path": "57d8d611aa27ba0110bf11462e32e60d5fa3707ccf20e762ed0f2d739edc2ef4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 256, + 70, + 351, + 225 + ], + "blocks": [ + { + "bbox": [ + 256, + 70, + 351, + 225 + ], + "lines": [ + { + "bbox": [ + 256, + 70, + 351, + 225 + ], + "spans": [ + { + "bbox": [ + 256, + 70, + 351, + 225 + ], + "type": "image", + "image_path": "4351adb86ce454a84d5604a730584fc91e2048110c3ce9379b9f12b3aabafe57.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 351, + 70, + 447, + 225 + ], + "blocks": [ + { + "bbox": [ + 351, + 70, + 447, + 225 + ], + "lines": [ + { + "bbox": [ + 351, + 70, + 447, + 225 + ], + "spans": [ + { + "bbox": [ + 351, + 70, + 447, + 225 + ], + "type": "image", + "image_path": "45fb1fec7ac4f6e2bce35f9247eb24272ca72a120fdcbd4962f410603be3952f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 447, + 70, + 542, + 226 + ], + "blocks": [ + { + "bbox": [ + 447, + 70, + 542, + 226 + ], + "lines": [ + { + "bbox": [ + 447, + 70, + 542, + 226 + ], + "spans": [ + { + "bbox": [ + 447, + 70, + 542, + 226 + ], + "type": "image", + "image_path": "a211d1496809b85aed57ff390bad18cb58749e4691c062529158b66765332b05.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 266, + 136, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 136, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 136, + 278 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 289, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 408 + ], + "type": "text", + "content": "We humans are long-term immersed in structured indoor scenes, ranging from bedrooms to offices. This fact has ignited a wealth of studies for reconstructing the indoor environment with various 3D structures such as lines [9, 20, 39], planes [12, 16, 18, 19, 33, 37, 42] and blocks [7, 23]. Among them, the 3D plane is the most common representation because of its simplicity and completeness in describing physical surfaces, thus motivating us to study the problem of 3D reconstruction for indoor scenes using 3D planes, i.e., planar 3D reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 411, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 615 + ], + "type": "text", + "content": "Planar 3D reconstruction has been extensively studied for years as a model fitting problem, in which a 3D scene geometry (e.g., point clouds, or meshes) was assumed to be known and the main goal is fitting the scene in a set of 3D planes [25, 30, 38]. Recently, the paradigm has been gradually simplified in image-based solutions in single-view and multi-view 3D reconstruction, eliminating the acquisition of known 3D scene geometry. To make the problem trackable, existing methods were extensively based on the image-level characterization of 2D/3D planes. That is to say, those methods have to detect 3D planes for each input image, match or track planes in across viewpoints, and finally reconstruct and merge 3D planes as the compact 3D representation of indoor scenes. Recent PlanarRecon [37] attempted to address these problems end-to-end by learning a 3D volume from monocular videos. Then, 3D planes can be detected, tracked, and fused in a consistent 3D space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Image-based planar 3D reconstruction mostly followed keypoint-based 3D reconstruction pipelines and treated planes as a special kind of visual features. However, because of the essential difference between the local point features and regional plane masks in image space, we argue, the existing approaches did not fully leverage the advantages of planar representations. Some evidences could be observed from the results of PlanarRecon [37], which are" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 268, + 535, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 535, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 535, + 279 + ], + "type": "text", + "content": "usually very coarse and lose many details of the scene." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 281, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 281, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 312, + 281, + 555, + 484 + ], + "type": "text", + "content": "In this paper, we are going to address the issues remained in image-based planar 3D reconstruction, aiming at obtaining a complete, structural, and compact indoor scene reconstruction from multi-view images. Our main idea is approximating the indoor scenes with a collection of solid 3D planar primitives from multi-view input images, directly optimizing them to have consistent 3D planes, but eliminating any suboptimal precomputing of 2D plane primitives (e.g., plane masks). We introduce PlanarSplatting that explicitly optimizes rectangular plane primitives in 3D space by differentiably splatting them into 2.5D depth and normal maps. Thanks to our well-designed plane splatting function, PlanarSplatting effectively leverage monocular geometry cues from modern foundational models [3, 10, 41] for accurate plane optimization. As shown in Fig. 1, the final high-quality planar surface can be reconstructed by simply merging adjacent 3D plane primitives." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 485, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 556, + 616 + ], + "type": "text", + "content": "Our PlanarSplatting is directly designed on the 3D plane primitives and efficiently implemented with CUDA, which accordingly induces a seamlessly integration with recent Gaussian Splatting (GS) methods for high-quality indoor novel view synthesis (NVS). As shown in Fig. 2, benefiting from our fast and accurate scene reconstruction (within 3 minutes), GS-based methods can be well-initialized and optimized without densification, leading to better rendering results and significantly less training time. It demonstrates the strong potential to promote the unity of surface reconstruction and novel view synthesis for indoor scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": "In the experiments, our PlanarSplatting shows its powerful ability for accurate indoor planar surface reconstruction for hundreds of scenes from two real-world indoor datasets, i.e. ScanNetV2 [2] and ScanNet++[40]. Furthermore, we show that with the combination of our PlanarSplatting and GS-based methods (e.g., 3DGS [13] and 2DGS [11]), we can effectively improve the rendering quality with less training time and fewer points." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1191" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 593 + ], + "type": "text", + "content": "Indoor Planar Reconstruction. Reconstructing indoor scenes with 3D plane primitives has been studied for a long time [4-6, 29, 36]. Traditionally, it is usually achieved by deducing and fitting the plane geometry directly from 3D data (e.g., point clouds [24, 25, 30, 31] and line clouds [15]), or from 2D single-view images with strict scene assumptions (e.g., the Manhattan World constrain) [16, 22, 26] which seriously limited their application. In recent years, some learnable-based methods are proposed to formulate this problem as single-view 3D plane segmentation [17, 18, 28, 32, 42, 43] and cross-view plane instance matching [1, 12, 33]. Although impressive results have been achieved, these works are hard to extend to multiple views. PlanarRecon [37] was the first end-to-end work proposed to deal with holistic indoor plane recovery by learning a plane-related 3D volume from posed RGB videos. Then, 3D planes can be extracted, tracked, and fused from the learned 3D volume incrementally. Most recently, AirPlanes [35] developed a two-step method that first reconstructed a dense scene mesh and then learned consistent 3D plane embeddings from 2D plane embeddings for plane extraction from the dense mesh. Note that the aforementioned learning-based methods require 2D/3D plane annotations as their supervision, leading to performance bottlenecks due to the difficulty of obtaining a large scale of plane annotations. In contrast, we proposed PlanarSplatting to reconstruct accurate and complete indoor planar surfaces by directly optimizing a set of solid 3D plane primitives from posed multi-view images without any extra plane detection or matching operations. Benefiting from our differentiable planar primitive rendering, PlanarSplatting directly leverages monocular depth/normal cues from modern foundation models for optimization without plane annotations. Concurrent work, AlphaTablets [8], explores planar representations, which involves learning planar tablets equipped with alpha maps for shape modeling and pseudo-mesh construction for rendering. Our PlanarSplatting distinguishes itself from AlphaTablets by learning plane radii, providing a more explicit parameterization of 3D planes. As a result, we achieve a much more efficient rasterization scheme at the level of minutes while AlphaTablets takes about 2 hours per scene." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Primitive-based Scene Representation. Optimizing explicit primitives such as points [11, 13, 34, 44], volumes [21], and superquadric [23] to represent the 3D scene has been studied for a long time. The core of these methods is to design a differentiable rendering process to optimize the attributes of primitives by gradient descent. Among them, a typical paradigm is to render images from primitives with splatting techniques which is realized with radial basis functions defined on the primitive (e.g., the Gaussian function). Inspired by these works, we proposed" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "PlanarSplatting to optimize solid 3D plane primitives in a differentiable rendering manner with a novel shape-aware plane splating function to better fit the scene geometry. Benefiting from our efficient CUDA implementation, PlanarSplatting can reconstruct the accurate indoor planar scene within 3 minutes and empower recent Gaussian-based works (e.g., 3DGS [13]) to further improve the rendering quality of novel views as shown in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 178, + 486, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 486, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 486, + 192 + ], + "type": "text", + "content": "3. The Proposed PlanarSplatting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 198, + 556, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 556, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 556, + 307 + ], + "type": "text", + "content": "As shown in Fig. 3, given a set of posed multi-view images, our PlanarSplatting can reconstruct the indoor planar scene from them by optimizing a set of learnable 3D planar primitives (Sec. 3.1). With the proposed Differentiable Planar Primitive Rendering (Sec. 3.2), we can explicitly learn these planar primitives from a coarse initialization to accurately recover the scene geometry. These optimized 3D planar primitives are then merged to achieve the 3D plane instances for the final planar reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 312, + 523, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 523, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 523, + 325 + ], + "type": "text", + "content": "3.1. Learnable Planar Scene Representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 330, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 389 + ], + "type": "text", + "content": "Since our key idea is directly optimizing explicit 3D primitives for planar scene reconstruction, we first present the representation of our learnable 3D planar primitives and then introduce how we initialize the scene with these planar primitives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "text", + "content": "Learnable Planar Primitive. As shown in Fig. 4, we formulate a planar primitive " + }, + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "text", + "content": " as a 3D rectangle which is equipped with several learnable parameters including the plane center " + }, + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\pi} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "text", + "content": ", the plane rotation " + }, + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\pi} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "text", + "content": " (in quaternion representation) and the plane radii " + }, + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\pi}" + }, + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "text", + "content": ". Specifically, to improve the optimization flexibility of plane shape, we use the design of double direction plane radii (Double Radii) for " + }, + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\pi}" + }, + { + "bbox": [ + 313, + 389, + 556, + 486 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 362, + 492, + 555, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 492, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 362, + 492, + 555, + 507 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {\\pi} = \\left\\{r _ {\\pi} ^ {x +}, r _ {\\pi} ^ {x -}, r _ {\\pi} ^ {y +}, r _ {\\pi} ^ {y -} \\right\\} \\in \\mathbb {R} _ {+} ^ {4}, \\tag {1}", + "image_path": "ea733037f1bf0cf514d66d2eef7f50631d4b120d22a3aabe7a7d07426431b731.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 513, + 556, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 556, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 556, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 513, + 556, + 586 + ], + "type": "inline_equation", + "content": "r_{\\pi}^{x+}, r_{\\pi}^{x-}, r_{\\pi}^{y+}, r_{\\pi}^{y-}" + }, + { + "bbox": [ + 313, + 513, + 556, + 586 + ], + "type": "text", + "content": " are the radii defined on the positive/negative direction of the X-axis/Y-axis of the rectangle as shown in Fig. 4. For the sake of description, we further define the positive direction of the X-axis and Y-axis of the 3D planar primitive as two orthogonal unit vectors " + }, + { + "bbox": [ + 313, + 513, + 556, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\pi}^{x}, \\mathbf{v}_{\\pi}^{y} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 313, + 513, + 556, + 586 + ], + "type": "text", + "content": " which can be calculated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 590, + 555, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 590, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 329, + 590, + 555, + 607 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\pi} ^ {x} = \\mathbf {R} (\\mathbf {q} _ {\\pi}) [ 1, 0, 0 ] ^ {\\top}, \\quad \\mathbf {v} _ {\\pi} ^ {y} = \\mathbf {R} (\\mathbf {q} _ {\\pi}) [ 0, 1, 0 ] ^ {\\top}, \\tag {2}", + "image_path": "20c6509c9fb1556369bbe2faceb6071ba94695b58051fb7ed576d3feb2de0bff.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 613, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 555, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 613, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(\\mathbf{q}_{\\pi}) \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 313, + 613, + 555, + 650 + ], + "type": "text", + "content": " means the rotation matrix of the quaternion " + }, + { + "bbox": [ + 313, + 613, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\pi}" + }, + { + "bbox": [ + 313, + 613, + 555, + 650 + ], + "type": "text", + "content": ". Similarly, the normal of the planar primitive " + }, + { + "bbox": [ + 313, + 613, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{\\pi} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 613, + 555, + 650 + ], + "type": "text", + "content": " can be calculated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 386, + 654, + 555, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 654, + 555, + 671 + ], + "spans": [ + { + "bbox": [ + 386, + 654, + 555, + 671 + ], + "type": "interline_equation", + "content": "\\mathbf {n} _ {\\pi} = \\mathbf {R} (\\mathbf {q} _ {\\pi}) [ 0, 0, 1 ] ^ {\\top}. \\tag {3}", + "image_path": "896d1431aad944133db7a767e619913f524e7837337848bade2d53e3375d2e10.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "With these learnable parameters, the 3D planar primitive can be moved to align with the potential scene surface and deformed to fit the surface shape during optimization." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1192" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 84, + 535, + 300 + ], + "blocks": [ + { + "bbox": [ + 77, + 84, + 535, + 300 + ], + "lines": [ + { + "bbox": [ + 77, + 84, + 535, + 300 + ], + "spans": [ + { + "bbox": [ + 77, + 84, + 535, + 300 + ], + "type": "image", + "image_path": "447b99c3bbe6bc3176621af2c5497560f491541fce2e96b126a2b29bb12331e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 307, + 555, + 341 + ], + "lines": [ + { + "bbox": [ + 54, + 307, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 555, + 341 + ], + "type": "text", + "content": "Figure 3. Illustration of our proposed PlanarSplatting. Given a set of posed multi-view images of indoor scenes, our method renders depth and normal maps from 3D plane primitives. Then, with the supervision of monocular cues, these 3D plane primitives are gradually optimized to recover the scene geometry and finally merged to get the planar reconstruction result." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 123, + 354, + 227, + 422 + ], + "blocks": [ + { + "bbox": [ + 123, + 354, + 227, + 422 + ], + "lines": [ + { + "bbox": [ + 123, + 354, + 227, + 422 + ], + "spans": [ + { + "bbox": [ + 123, + 354, + 227, + 422 + ], + "type": "image", + "image_path": "35b7220572873baffb5a9a2f549ec7e342306f7efcbe2faba124ceba94a1acd5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 426, + 295, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 449 + ], + "type": "text", + "content": "Figure 4. Representation of 3D plane primitive with learnable shape parameters including plane center, radii, and rotation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 616 + ], + "type": "text", + "content": "Scene Initialization. We use monocular depth from recent foundation models [10] to fast initialize our 3D planar primitives at the beginning of optimization. Specifically, we use depths from Metric3Dv2 [10] to get a very coarse scene geometry. Then we randomly sample 2,000 points on the coarse mesh to achieve the plane centers of our 3D planar primitives. The initial radii of each primitive " + }, + { + "bbox": [ + 55, + 460, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 460, + 296, + 616 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 55, + 460, + 296, + 616 + ], + "type": "inline_equation", + "content": "0.5Dist(\\pi)" + }, + { + "bbox": [ + 55, + 460, + 296, + 616 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 460, + 296, + 616 + ], + "type": "inline_equation", + "content": "Dist(\\pi)" + }, + { + "bbox": [ + 55, + 460, + 296, + 616 + ], + "type": "text", + "content": " means the distance closest to " + }, + { + "bbox": [ + 55, + 460, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 460, + 296, + 616 + ], + "type": "text", + "content": " to its neighbors. We use the normal direction on the coarse mesh to initialize the plane rotation. At the bottom of Fig. 3, we show an example of our initial planar primitives. With such a coarse initialization, our PlanarSplatting can finally reconstruct the accurate and complete scene surface." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 623, + 277, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 277, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 277, + 636 + ], + "type": "text", + "content": "3.2. Differentiable Planar Primitive Rendering" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "text", + "content": "To optimize the learnable planar primitives " + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Pi = \\{\\pi_i\\}_{i=1}^K" + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "text", + "content": ", we introduce the differentiable planar primitive rendering with a carefully designed plane splating function which enables the planar primitives to accurately fit the scene geometry with the supervision only from 2D multi-view images." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 353, + 538, + 432 + ], + "blocks": [ + { + "bbox": [ + 326, + 353, + 538, + 432 + ], + "lines": [ + { + "bbox": [ + 326, + 353, + 538, + 432 + ], + "spans": [ + { + "bbox": [ + 326, + 353, + 538, + 432 + ], + "type": "image", + "image_path": "b5dfbdc26ede9d699243bd82c88bb907ec43a8b2147a12fa0fb486aa0b484ef8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 435, + 555, + 511 + ], + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 511 + ], + "type": "text", + "content": "Figure 5. Illustration of the proposed plane splatting function. Naive Gaussian Splatting can not effectively approximate the boundary of our rectangular plane primitive (shown in black dashed border). In contrast, our proposed plane splatting function can approximate the boundary of the rectangle as the number of iterations increases, allowing our 3D planar primitives to better fit the surface of the scene." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "text", + "content": "Ray-to-Plane Intersection. To project 3D planar primitives to the 2D image space, we first calculate the intersections between planar primitives and the rays emitted from image pixels. Specifically, given a ray " + }, + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{r} = \\{\\mathbf{o},\\mathbf{d}\\}" + }, + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "text", + "content": " starting from the camera center " + }, + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{o}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "text", + "content": " with direction " + }, + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{d}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "text", + "content": ", its intersection " + }, + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\pi}^{\\mathbf{r}}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "text", + "content": " to one planar primitive " + }, + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 525, + 555, + 598 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 376, + 607, + 555, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 607, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 376, + 607, + 555, + 633 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} = \\mathbf {o} + \\frac {\\left(\\mathbf {p} _ {\\pi} - \\mathbf {o}\\right) \\cdot \\mathbf {n} _ {\\pi}}{\\mathbf {d} \\cdot \\mathbf {n} _ {\\pi}} \\mathbf {d}, \\tag {4}", + "image_path": "9407ed16be1fbcc7797b3eed1ae21feccb96701a6ba7d2879989d63e869c70b8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\pi}" + }, + { + "bbox": [ + 313, + 641, + 555, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{\\pi}" + }, + { + "bbox": [ + 313, + 641, + 555, + 664 + ], + "type": "text", + "content": " are the center and the normal of the planar primitive " + }, + { + "bbox": [ + 313, + 641, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 641, + 555, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 700 + ], + "type": "text", + "content": "Plane Splatting Function. After achieving the ray-to-plane intersection " + }, + { + "bbox": [ + 313, + 665, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\pi}^{\\mathbf{r}}" + }, + { + "bbox": [ + 313, + 665, + 554, + 700 + ], + "type": "text", + "content": ", we calculate its splatting weight with our plane splatting function which will be used for rendering." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "content": "A vanilla selection of the splatting function is the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1193" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 181, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 181, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 181, + 83 + ], + "type": "text", + "content": "anisotropic Gaussian function:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 92, + 295, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 92, + 295, + 119 + ], + "spans": [ + { + "bbox": [ + 61, + 92, + 295, + 119 + ], + "type": "interline_equation", + "content": "w \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}, \\pi\\right) = \\exp \\left(- \\frac {1}{2} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right) ^ {\\top} \\Sigma^ {- 1} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right)\\right), \\tag {5}", + "image_path": "fe48c80efe311655811e85f69eafbe090251ba447727b1dafa477c159de1008b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 126, + 296, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 296, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 126, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\pi}" + }, + { + "bbox": [ + 55, + 126, + 296, + 198 + ], + "type": "text", + "content": " is the center of the planar primitive " + }, + { + "bbox": [ + 55, + 126, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 126, + 296, + 198 + ], + "type": "text", + "content": ". The covariance matrix " + }, + { + "bbox": [ + 55, + 126, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 126, + 296, + 198 + ], + "type": "text", + "content": " can be calculated like [11, 13]. However, as shown in Fig. 5 and Fig. 6, the Gaussian-based splatting function will make ambiguous boundaries of our rectangular planar primitive leading to the degeneration of reconstruction quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "content": "Thus, we propose to calculate the splating weight with a novel rectangle-based plane splating function. To a given intersection " + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\pi}^{\\mathbf{r}}" + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "content": " between ray " + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "content": " and planar primitive " + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "content": ", we first calculate its projection distance " + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_X, \\mathcal{P}_Y \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "content": " to the X-axis and Y-axis of the planar primitive as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 266, + 295, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 266, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 69, + 266, + 295, + 281 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {X} = \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right) \\cdot \\mathbf {v} _ {\\pi} ^ {x}, \\quad \\mathcal {P} _ {Y} = \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}} - \\mathbf {p} _ {\\pi}\\right) \\cdot \\mathbf {v} _ {\\pi} ^ {y}. \\quad (6)", + "image_path": "3e2515ff0d9f54c266befbf460626956aa8503122c539f8a3c2083ea936a08ad.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 289, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 312 + ], + "type": "text", + "content": "Then, we calculate the splating weight along the X-axis of the plane " + }, + { + "bbox": [ + 55, + 289, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 289, + 296, + 312 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 321, + 295, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 321, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 63, + 321, + 295, + 353 + ], + "type": "interline_equation", + "content": "w _ {X} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}\\right) = \\left\\{ \\begin{array}{l l} 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {x +} - | \\mathcal {P} _ {X} |\\right)\\right), & \\text {i f} \\mathcal {P} _ {X} > 0, \\\\ 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {x -} - | \\mathcal {P} _ {X} |\\right)\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {7}", + "image_path": "ca1d5a8cf782dafce734ddc01b84b1fec2a367349725bea71aa3816db4da6faa.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "text", + "content": " is the Sigmoid function and " + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "text", + "content": " is the hyperparameter to control the splattering weight. Similarly, we continue to calculate the splattering weight along the Y-axis of the plane " + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 418, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 418, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 64, + 418, + 295, + 449 + ], + "type": "interline_equation", + "content": "w _ {Y} \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}\\right) = \\left\\{ \\begin{array}{l l} 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {y +} - | \\mathcal {P} _ {Y} |\\right)\\right), & \\text {i f} \\mathcal {P} _ {Y} > 0, \\\\ 2 \\sigma \\left(5 \\lambda \\left(r _ {\\pi} ^ {y -} - | \\mathcal {P} _ {Y} |\\right)\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}", + "image_path": "d79fc60912708d46587498377d9bd25986bc6bf795d26204cdbae6e010071bb5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 458, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 296, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 458, + 296, + 491 + ], + "type": "inline_equation", + "content": "r_{\\pi}^{x + },r_{\\pi}^{x - },r_{\\pi}^{y + },r_{\\pi}^{y - }" + }, + { + "bbox": [ + 55, + 458, + 296, + 491 + ], + "type": "text", + "content": " are the radii parameters of the plane " + }, + { + "bbox": [ + 55, + 458, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 458, + 296, + 491 + ], + "type": "text", + "content": " . At last, the final splatting weight can be calculated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 491, + 295, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 491, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 102, + 491, + 295, + 522 + ], + "type": "interline_equation", + "content": "w \\left(\\mathbf {x} _ {\\pi} ^ {\\mathbf {r}}\\right) = \\left\\{ \\begin{array}{l l} w _ {X}, & \\text {i f} w _ {X} < w _ {Y}, \\\\ w _ {Y}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {9}", + "image_path": "5a4af170a6cd49b3aa17dcf4f884eed0b1529b5d2ba1fb9a8df52e0827d9afba.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "text", + "content": "In Fig. 5, we show that with the increment of hyperparameter " + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "text", + "content": ", our plane splating function gradually approximates the shape of the rectangular plane primitive. In practice, we increase the value of " + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 526, + 296, + 588 + ], + "type": "text", + "content": " with an exponential function during optimization up to the maximum value of 300 as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 102, + 594, + 295, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 594, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 102, + 594, + 295, + 609 + ], + "type": "interline_equation", + "content": "\\lambda = \\min \\left(2 0 e ^ {\\left(- (1 - 0. 0 0 1 * i t e)\\right)}, 3 0 0\\right), \\tag {10}", + "image_path": "4d0bcd648b13b5da19ea7bb374c9ee3f47acb7afd819bd3712a44feff206af0d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 617, + 293, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 293, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 293, + 629 + ], + "type": "text", + "content": "where ite means the iteration number during optimization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "text", + "content": "Blending Composition. For all ray-to-plane intersections, we filter them with splating weight lower than 0.0001 and then sort the remaining intersections according to their depth from near to far. Then, " + }, + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "text", + "content": " nearest intersections of each ray are selected for rendering (" + }, + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "inline_equation", + "content": "M = 30" + }, + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "text", + "content": " in this paper). Denote the selected intersections of a ray " + }, + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "inline_equation", + "content": "P^{\\mathbf{r}} = \\{\\mathbf{x}_{\\pi_{\\tau(j)}}^{\\mathbf{r}}\\}_{j=1}^{M}" + }, + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\tau(j)" + }, + { + "bbox": [ + 55, + 630, + 296, + 716 + ], + "type": "text", + "content": " indicates the index of plane" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 385, + 119 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 385, + 119 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 385, + 119 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 385, + 119 + ], + "type": "image", + "image_path": "d609f4ee2501591c5c13063024999d195c9498e0f1d5caa7aa8de2cd9d516373.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 121, + 383, + 132 + ], + "lines": [ + { + "bbox": [ + 319, + 121, + 383, + 132 + ], + "spans": [ + { + "bbox": [ + 319, + 121, + 383, + 132 + ], + "type": "text", + "content": "(a) w/ GS Splitting" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 399, + 71, + 468, + 119 + ], + "blocks": [ + { + "bbox": [ + 399, + 71, + 468, + 119 + ], + "lines": [ + { + "bbox": [ + 399, + 71, + 468, + 119 + ], + "spans": [ + { + "bbox": [ + 399, + 71, + 468, + 119 + ], + "type": "image", + "image_path": "d900e5a7255e421cb3ed95003d65a2540f5bbdfc7cb0a5620067d8653d14481f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 121, + 468, + 132 + ], + "lines": [ + { + "bbox": [ + 398, + 121, + 468, + 132 + ], + "spans": [ + { + "bbox": [ + 398, + 121, + 468, + 132 + ], + "type": "text", + "content": "(b) w/ Plane Splatting" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 135, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 313, + 135, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 553, + 156 + ], + "type": "text", + "content": "Figure 6. Reconstruction comparison with different splatting functions. 'w/' means 'with'." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 481, + 70, + 550, + 119 + ], + "blocks": [ + { + "bbox": [ + 481, + 70, + 550, + 119 + ], + "lines": [ + { + "bbox": [ + 481, + 70, + 550, + 119 + ], + "spans": [ + { + "bbox": [ + 481, + 70, + 550, + 119 + ], + "type": "image", + "image_path": "0fc6fd89f60d5c112dce3c6820bbc905110efc178afc07d7fe501693f4f81889.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 121, + 537, + 131 + ], + "lines": [ + { + "bbox": [ + 495, + 121, + 537, + 131 + ], + "spans": [ + { + "bbox": [ + 495, + 121, + 537, + 131 + ], + "type": "text", + "content": "(c) GT Mesh" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "type": "text", + "content": "among all planar primitives. At last, we render the depth and normal map of a certain image " + }, + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 367, + 204, + 553, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 204, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 367, + 204, + 553, + 239 + ], + "type": "interline_equation", + "content": "\\mathbf {D} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) = \\sum_ {j = 1} ^ {M} T _ {j} w \\left(\\mathbf {x} _ {\\pi_ {\\tau (j)}} ^ {\\mathbf {r}}\\right) t _ {j}, \\tag {11}", + "image_path": "99e1c555544fc82bd7e15a8d5ca7b0e8ecc1f4999ef78e33ea0769d6ecd70eae.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 359, + 251, + 553, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 251, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 359, + 251, + 553, + 284 + ], + "type": "interline_equation", + "content": "\\mathbf {N} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) = \\sum_ {j = 1} ^ {M} T _ {j} w \\left(\\mathbf {x} _ {\\pi_ {\\tau (j)}} ^ {\\mathbf {r}}\\right) \\mathbf {n} _ {\\pi_ {\\tau (j)}}, \\tag {12}", + "image_path": "59883c5e6bee3be9aba691eb893a399ac924fea053db997435a63437e4c7b1e5.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 290, + 342, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 290, + 342, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 290, + 342, + 300 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 380, + 300, + 553, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 300, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 380, + 300, + 553, + 332 + ], + "type": "interline_equation", + "content": "T _ {j} = \\prod_ {i = 1} ^ {j - 1} (1 - w \\left(\\mathbf {x} _ {\\pi_ {\\tau (i)}} ^ {\\mathbf {r}}\\right)). \\tag {13}", + "image_path": "c53d243372d6be7f2df14b448a87ae33bcb90a9857bab7440d359f6ee6fe7f6c.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "text", + "content": " is the depth of the intersection and " + }, + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{\\pi_{\\tau(j)}}" + }, + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "text", + "content": " is the normal of planar primitive " + }, + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\pi_{\\tau(j)}" + }, + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "text", + "content": ". To supervise the rendered depth and normal map, we use the pretrained model of Metric3Dv2 [10] to predict the depth map " + }, + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{pre}}" + }, + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "text", + "content": " and normal map " + }, + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{\\mathrm{pre}}" + }, + { + "bbox": [ + 313, + 338, + 554, + 421 + ], + "type": "text", + "content": " to serve as pseudo labels. Besides, we also test the normal map predicted by Omnidata [3]. Finally, the render loss can be calculated as:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 331, + 429, + 553, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 429, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 331, + 429, + 553, + 510 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {r e n d e r}} ^ {\\Pi} = \\alpha_ {1} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\| 1 - \\mathbf {N} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) ^ {\\top} \\mathbf {N} _ {\\text {p r e}} (\\mathbf {r})) \\| _ {1} + \\\\ \\alpha_ {1} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\| \\mathbf {N} _ {\\text {r e n d e r}} ^ {\\Pi} (\\mathbf {r}) - \\mathbf {N} _ {\\text {p r e}} (\\mathbf {r})) \\| _ {1} + \\tag {14} \\\\ \\alpha_ {2} \\sum_ {\\mathbf {r} \\in \\mathbf {I}} \\| (\\mathbf {D} _ {\\mathrm {r e n d e r}} ^ {\\Pi} (\\mathbf {r})) - \\mathbf {D} _ {\\mathrm {p r e}} (\\mathbf {r}) \\| _ {1}, \\\\ \\end{array}", + "image_path": "e31c101a7ec6c07d18e52fdfd4e2b3c6100923421e8a4081bdaa8048eefa371a.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 519, + 553, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 553, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 519, + 553, + 544 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = 2.0" + }, + { + "bbox": [ + 313, + 519, + 553, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 519, + 553, + 544 + ], + "type": "inline_equation", + "content": "\\alpha_{2} = 2.0" + }, + { + "bbox": [ + 313, + 519, + 553, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 519, + 553, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 519, + 553, + 544 + ], + "type": "text", + "content": " is the ray/pixel emitted from image " + }, + { + "bbox": [ + 313, + 519, + 553, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 519, + 553, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 552, + 400, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 400, + 565 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 400, + 565 + ], + "type": "text", + "content": "3.3. Optimization" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 570, + 554, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 618 + ], + "type": "text", + "content": "Loss Function. To each scene, we optimize our Planar-Splatting with the Adam optimizer [14] for 5,000 iterations on the ScanNetV2 [2] dataset and 10,000 iterations on the ScanNet++ [40] dataset using the loss as shown in Eq. (14)." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": "Plane Splitting. During optimization, we introduce a splitting operation on planes according to the gradients of their radii to better fit the scene geometry. If the average radii gradients on the X-axis " + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "inline_equation", + "content": "(r^{x+}" + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "inline_equation", + "content": "r^{x-})" + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": " of the plane are greater than 0.2, we split the plane along the Y-axis. Similarly, we split the planes along the X-axis, if their radii gradients on the Y-axis " + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "inline_equation", + "content": "(r^{y+}" + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "inline_equation", + "content": "r^{y-})" + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": " are larger than 0.2. We conduct the splitting operation every 1,000 iterations." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1194" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 96, + 549, + 192 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of planar reconstruction results on the ScanNetV2 [2] dataset. 'P. Ann.' means using 2D/3D plane annotations in training stages." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 96, + 549, + 192 + ], + "lines": [ + { + "bbox": [ + 61, + 96, + 549, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 549, + 192 + ], + "type": "table", + "html": "
MethodP. Ann.GeometrySegmentationPlanar
Chamfer ↓F-score ↑VOI ↓RI ↑SC ↑Fidelity ↓Acc ↓Chamfer ↓
PlanarRecon [37]✓9.8943.473.2010.9190.40518.8616.2117.53
AirPlanes [35]✓5.3064.922.2680.9570.5688.767.988.37
2DGS [11] + RANSACX14.1531.334.0300.9240.25740.0214.7727.40
Metric3Dv2 [10] + RANSACX5.9365.452.7440.9270.4889.2911.3610.33
SR [27] + RANSACX5.4065.452.5070.9460.5159.4210.139.78
Ours@OmnidataX4.8368.852.5020.9480.5326.6411.769.20
Ours@Metric3Dv2X4.2373.762.3100.9580.5605.7210.658.19
", + "image_path": "56602deac7da37e8139cc4358e94a4b3bd8bcd627f136fc40c5965cd9927ef8e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 60, + 228, + 549, + 324 + ], + "blocks": [ + { + "bbox": [ + 55, + 202, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 555, + 226 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of planar reconstruction results on the ScanNet++ [40] dataset. 'P. Ann.' means using 2D/3D plane annotations in training stages." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 228, + 549, + 324 + ], + "lines": [ + { + "bbox": [ + 60, + 228, + 549, + 324 + ], + "spans": [ + { + "bbox": [ + 60, + 228, + 549, + 324 + ], + "type": "table", + "html": "
MethodP. Ann.GeometrySegmentationPlanar
Chamfer ↓F-score ↑VOI ↓RI ↑SC ↑Fidelity ↓Acc ↓Chamfer ↓
PlanarRecon [37]✓17.8531.103.5420.9190.36734.4419.3526.90
AirPlanes [35]✓13.7532.582.8590.9410.47028.1612.5820.37
2DGS [11] + RANSACX20.3926.464.4560.9270.24155.9016.8836.39
Metric3Dv2 [10] + RANSACX11.1441.543.0210.9300.44127.5010.8119.16
SR [27] + RANSACX13.1535.933.0130.9380.44230.2511.6220.94
Ours@OmnidataX9.3347.042.7720.9460.52317.2412.2614.75
Ours@Metric3Dv2X9.0150.682.7820.9440.52117.1712.3614.77
", + "image_path": "bac4030c307e0b6bba8c4fb58eb80f13a79d2f5c827b46b2e4f139a2831cb6e2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 335, + 295, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 394 + ], + "type": "text", + "content": "Plane Merge. After optimization, we greedily merge the adjacent 3D plane primitives with normal angle error lower than " + }, + { + "bbox": [ + 55, + 335, + 295, + 394 + ], + "type": "inline_equation", + "content": "25^{\\circ}" + }, + { + "bbox": [ + 55, + 335, + 295, + 394 + ], + "type": "text", + "content": " and offset distance error lower than " + }, + { + "bbox": [ + 55, + 335, + 295, + 394 + ], + "type": "inline_equation", + "content": "0.1m" + }, + { + "bbox": [ + 55, + 335, + 295, + 394 + ], + "type": "text", + "content": ". Here, offset means the projection distance from the scene center to the plane surface." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 480 + ], + "type": "text", + "content": "CUDA Implementation. For fast optimization of solid 3D planar primitives, we implement the forward and backward process of our Differentiable Planar Primitive Rendering with CUDA, which enables our PlanarSplatting to reconstruct one scene within 3 minutes. We will release the code of our CUDA implementation for solid 3D planar primitives optimization after publication." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 494, + 137, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 137, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 137, + 507 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 222, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 222, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 222, + 528 + ], + "type": "text", + "content": "4.1. Dataset, Metrics and Baselines" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": "Datasets. We evaluate our PlanarSplatting on two large indoor datasets including ScanNetV2 [2] and ScanNet++ [40] which provide posed images. On the ScanNetV2 dataset, we use the test split according to AirPlanes [35] which includes 100 scenes with ground truth 3D plane annotations provided by [18]. On the ScanNet++ dataset, we randomly select 30 scenes for evaluation and extract 3D plane annotations from the ground truth meshes like [18]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. According to PlanarRecon [37], we evaluate the geometry quality of all reconstructed 3D planes with Chamfer Distance and F-score. Following AirPlanes [35], we also evaluate the reconstruction quality of Top-20 largest planes from the ground truth and use the metrics including Planar Fidelity, Planar Accuracy, and Planar Chamfer. To evaluate 3D plane segmentation, we use" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 335, + 555, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 360 + ], + "type": "text", + "content": "the metrics including Variation of Information (VOI), Rand Index (RI), and Segmentation Covering (SC) like [18, 37]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 556, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 516 + ], + "type": "text", + "content": "Baselines. Since our method is purely built upon geometry cues, we mainly compare our PlanarSplatting with those geometry-based methods including SR+RANSAC [27], Metric3Dv2+RANSAC [10] and 2DGS+RANSAC [11]. These baselines build dense scene meshes from multi-view images at first and then extract 3D planes with the RANSAC algorithm from the reconstructed meshes. We use the RANSAC implementation provided by AirPlanes [35] for all these geometry-based baselines. Besides, we also compare our PlanarSplatting with some plane annotation based methods that use 2D/3D plane labels/priors in their training stage including PlanarRecon [37] and AirPlanes [35], and report their results for reference." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 527, + 469, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 469, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 469, + 540 + ], + "type": "text", + "content": "4.2. Comparisons with Baselines" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "Quantitative Results. We first evaluate our PlanarSplatting on the ScanNetV2 [2] dataset. Specifically, we evaluate two versions of our PlanarSplatting, utilizing normal maps from different pretrained models for supervision: Metric3Dv2 [10] and Omnidata [3], respectively. As shown in Tab. 1, our method achieves the best geometry performance in the metric of Chamfer and F-score compared to all baselines. To the results of Segmentation and Planar, our method outperforms geometry-based baselines in most metrics and is comparable to the state-of-the-art AirPlanes which uses plane embeddings trained on the same ScanNetV2 dataset. In Tab. 2, we show the results on the ScanNet++ [40] dataset. Benefiting from our novel design for differentiable plane rendering and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1195" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 73, + 184, + 155 + ], + "blocks": [ + { + "bbox": [ + 80, + 73, + 184, + 155 + ], + "lines": [ + { + "bbox": [ + 80, + 73, + 184, + 155 + ], + "spans": [ + { + "bbox": [ + 80, + 73, + 184, + 155 + ], + "type": "image", + "image_path": "0ee8655e0fcd8dbd7e212abbbe46f584c8dfa0127e044df27e2adac3fc79a426.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 194, + 78, + 299, + 145 + ], + "blocks": [ + { + "bbox": [ + 194, + 78, + 299, + 145 + ], + "lines": [ + { + "bbox": [ + 194, + 78, + 299, + 145 + ], + "spans": [ + { + "bbox": [ + 194, + 78, + 299, + 145 + ], + "type": "image", + "image_path": "087c91d570fb06d8e10fc2d319e95eb4f2b9b5723a70934bcb8e3946112ac1bc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 306, + 78, + 413, + 145 + ], + "blocks": [ + { + "bbox": [ + 306, + 78, + 413, + 145 + ], + "lines": [ + { + "bbox": [ + 306, + 78, + 413, + 145 + ], + "spans": [ + { + "bbox": [ + 306, + 78, + 413, + 145 + ], + "type": "image", + "image_path": "892d7ded0fdf2f3d18ec68b91a7330b9f365816a231e449bf600e68677b3ec5e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 417, + 80, + 525, + 156 + ], + "blocks": [ + { + "bbox": [ + 417, + 80, + 525, + 156 + ], + "lines": [ + { + "bbox": [ + 417, + 80, + 525, + 156 + ], + "spans": [ + { + "bbox": [ + 417, + 80, + 525, + 156 + ], + "type": "image", + "image_path": "da5de82a644b7150ff18daa56e0511f75ae9d0726e8bac4c267911ceb7178d4d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 82, + 156, + 189, + 247 + ], + "blocks": [ + { + "bbox": [ + 82, + 156, + 189, + 247 + ], + "lines": [ + { + "bbox": [ + 82, + 156, + 189, + 247 + ], + "spans": [ + { + "bbox": [ + 82, + 156, + 189, + 247 + ], + "type": "image", + "image_path": "a4097cac7631077f509a2caa59c63170033411d36b9859551fbc664496aa43be.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 194, + 155, + 301, + 247 + ], + "blocks": [ + { + "bbox": [ + 194, + 155, + 301, + 247 + ], + "lines": [ + { + "bbox": [ + 194, + 155, + 301, + 247 + ], + "spans": [ + { + "bbox": [ + 194, + 155, + 301, + 247 + ], + "type": "image", + "image_path": "f3e4491a28c4e4b0d5190aa4816425a298cccad82b1ecca8548b3405017158d8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 155, + 412, + 247 + ], + "blocks": [ + { + "bbox": [ + 306, + 155, + 412, + 247 + ], + "lines": [ + { + "bbox": [ + 306, + 155, + 412, + 247 + ], + "spans": [ + { + "bbox": [ + 306, + 155, + 412, + 247 + ], + "type": "image", + "image_path": "d7c4c1acfe3e66c27dd96655c5d682bcc44b16aa1979fb331ced419ae71dc75d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 417, + 156, + 524, + 247 + ], + "blocks": [ + { + "bbox": [ + 417, + 156, + 524, + 247 + ], + "lines": [ + { + "bbox": [ + 417, + 156, + 524, + 247 + ], + "spans": [ + { + "bbox": [ + 417, + 156, + 524, + 247 + ], + "type": "image", + "image_path": "6578c113348bbfda366dd36c9d4e9f5604b5318858afedc87e3ae402696d8048.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 81, + 258, + 189, + 327 + ], + "blocks": [ + { + "bbox": [ + 81, + 258, + 189, + 327 + ], + "lines": [ + { + "bbox": [ + 81, + 258, + 189, + 327 + ], + "spans": [ + { + "bbox": [ + 81, + 258, + 189, + 327 + ], + "type": "image", + "image_path": "7035b4089af551553830c3fe32e725dfcf28f9fb46405490618e6670b64afb61.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 193, + 258, + 299, + 327 + ], + "blocks": [ + { + "bbox": [ + 193, + 258, + 299, + 327 + ], + "lines": [ + { + "bbox": [ + 193, + 258, + 299, + 327 + ], + "spans": [ + { + "bbox": [ + 193, + 258, + 299, + 327 + ], + "type": "image", + "image_path": "2212230165801d77165153c0cfc6d97b1165a0932d68e2e939bd106736becfab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 255, + 410, + 327 + ], + "blocks": [ + { + "bbox": [ + 306, + 255, + 410, + 327 + ], + "lines": [ + { + "bbox": [ + 306, + 255, + 410, + 327 + ], + "spans": [ + { + "bbox": [ + 306, + 255, + 410, + 327 + ], + "type": "image", + "image_path": "86a40c7a3b4d65f485daa94a0525111123ff34d3335f27934c44cad584a03cdf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 416, + 258, + 524, + 327 + ], + "blocks": [ + { + "bbox": [ + 416, + 258, + 524, + 327 + ], + "lines": [ + { + "bbox": [ + 416, + 258, + 524, + 327 + ], + "spans": [ + { + "bbox": [ + 416, + 258, + 524, + 327 + ], + "type": "image", + "image_path": "55ef8f1c13195e3561c5e8ce0ad07ab908815af9e058669b6aa66fb4543f5fb5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 83, + 333, + 189, + 412 + ], + "blocks": [ + { + "bbox": [ + 83, + 333, + 189, + 412 + ], + "lines": [ + { + "bbox": [ + 83, + 333, + 189, + 412 + ], + "spans": [ + { + "bbox": [ + 83, + 333, + 189, + 412 + ], + "type": "image", + "image_path": "867d41b776217bfa3a818d9b59086e370c1f89e8d0735f82cb395f24c69333c7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 194, + 333, + 301, + 411 + ], + "blocks": [ + { + "bbox": [ + 194, + 333, + 301, + 411 + ], + "lines": [ + { + "bbox": [ + 194, + 333, + 301, + 411 + ], + "spans": [ + { + "bbox": [ + 194, + 333, + 301, + 411 + ], + "type": "image", + "image_path": "062632401fc894e0bdd10083f67c0615f7aa9ed1c7853df05392c083d8d6706c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 333, + 412, + 411 + ], + "blocks": [ + { + "bbox": [ + 306, + 333, + 412, + 411 + ], + "lines": [ + { + "bbox": [ + 306, + 333, + 412, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 412, + 411 + ], + "type": "image", + "image_path": "8b69d97ab7aa638401571b550c41a3caedf16d20446824f28652ab87f89f2e0c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 417, + 333, + 524, + 411 + ], + "blocks": [ + { + "bbox": [ + 417, + 333, + 524, + 411 + ], + "lines": [ + { + "bbox": [ + 417, + 333, + 524, + 411 + ], + "spans": [ + { + "bbox": [ + 417, + 333, + 524, + 411 + ], + "type": "image", + "image_path": "f64063875bec07baae356f33420c6b621ee6436333f04a8ee543ed2c61fe8682.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 86, + 422, + 186, + 482 + ], + "blocks": [ + { + "bbox": [ + 86, + 422, + 186, + 482 + ], + "lines": [ + { + "bbox": [ + 86, + 422, + 186, + 482 + ], + "spans": [ + { + "bbox": [ + 86, + 422, + 186, + 482 + ], + "type": "image", + "image_path": "1916fe74eeb04f44110a79719014c03925d0a996dda6753e1605faf9b226e3af.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 487, + 171, + 497 + ], + "lines": [ + { + "bbox": [ + 102, + 487, + 171, + 497 + ], + "spans": [ + { + "bbox": [ + 102, + 487, + 171, + 497 + ], + "type": "text", + "content": "(a) PlanarRecon [37]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 196, + 421, + 295, + 487 + ], + "blocks": [ + { + "bbox": [ + 196, + 421, + 295, + 487 + ], + "lines": [ + { + "bbox": [ + 196, + 421, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 196, + 421, + 295, + 487 + ], + "type": "image", + "image_path": "a453f1e1cc145e922161c0554e42a454937035287caf033fb21fe3f83da8edef.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 487, + 277, + 497 + ], + "lines": [ + { + "bbox": [ + 219, + 487, + 277, + 497 + ], + "spans": [ + { + "bbox": [ + 219, + 487, + 277, + 497 + ], + "type": "text", + "content": "(b) AirPlanes [35]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 309, + 422, + 408, + 481 + ], + "blocks": [ + { + "bbox": [ + 309, + 422, + 408, + 481 + ], + "lines": [ + { + "bbox": [ + 309, + 422, + 408, + 481 + ], + "spans": [ + { + "bbox": [ + 309, + 422, + 408, + 481 + ], + "type": "image", + "image_path": "1249e5a5b2cf1181a6b7449c1648c21409bc4849354ccf63f956d508358845aa.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 488, + 373, + 496 + ], + "lines": [ + { + "bbox": [ + 346, + 488, + 373, + 496 + ], + "spans": [ + { + "bbox": [ + 346, + 488, + 373, + 496 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 417, + 422, + 518, + 483 + ], + "blocks": [ + { + "bbox": [ + 417, + 422, + 518, + 483 + ], + "lines": [ + { + "bbox": [ + 417, + 422, + 518, + 483 + ], + "spans": [ + { + "bbox": [ + 417, + 422, + 518, + 483 + ], + "type": "image", + "image_path": "94530101150dcfc2422cf01a4ed422e2492d771d4de9df59440eff52e2840fe9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 487, + 499, + 496 + ], + "lines": [ + { + "bbox": [ + 443, + 487, + 499, + 496 + ], + "spans": [ + { + "bbox": [ + 443, + 487, + 499, + 496 + ], + "type": "text", + "content": "(d) Ground Truth" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 111, + 501, + 498, + 512 + ], + "lines": [ + { + "bbox": [ + 111, + 501, + 498, + 512 + ], + "spans": [ + { + "bbox": [ + 111, + 501, + 498, + 512 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison on the ScanNetV2 [2] (rows 1-4) and ScanNet++ [40] (last row) datasets." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 606 + ], + "type": "text", + "content": "optimization via monocular cues provided by modern foundation models, our PlanarSplatting still achieves the best Geometry performance among all baselines. Furthermore, our method outperforms both geometry-based and plane annotation based methods in most metrics of Segmentation and Planar, demonstrating the robustness and superiority of our proposed PlanarSplatting." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Qualitative Results. In Fig. 7, we show the comparisons of PlanarRecon [37], AirPlanes [35] and our PlanarSplatting with normal supervision of Omnidata [3]. PlanarRecon reconstructs coarse 3D planes from input multi-view images while our PlanarSplatting accurately reconstructs scene geometry via directly optimizing 3D plane primitives in the whole 3D space. With the help of a plane embedding model, AirPlanes achieves more semantic plane segmentation from" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 522, + 553, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 606 + ], + "type": "text", + "content": "co-plane regions. However, as shown in the zoom-in results in Fig. 7, AirPlanes loses many geometry details in the scene. In contrast, our PlanarSplatting successfully reconstructs detailed structures such as the chairs and the legs of the bed, resulting in both accuracy and completeness in results. These comparisons effectively indicate the superiority of our method." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 612, + 415, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 415, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 415, + 623 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 629, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 677 + ], + "type": "text", + "content": "We verify the design of our PlanarSplatting on 10 scenes randomly selected from the ScanNetV2 [2] dataset. We optimize 3D plane primitives with 5,000 iterations with the normal supervision of Omnidata [3] in default." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Plane Initialization. We first assess the sensitivity of the plane initialization strategy of our PlanarSplatting. We ablate the used Metric3Dv2 [10] initialization to the sphere" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1196" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 96, + 293, + 162 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 3. Ablation studies of the proposed PlanarSplatting on the ScanNetV2 dataset. 'w/' means 'with' and 'w/o' means 'without'." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 96, + 293, + 162 + ], + "lines": [ + { + "bbox": [ + 59, + 96, + 293, + 162 + ], + "spans": [ + { + "bbox": [ + 59, + 96, + 293, + 162 + ], + "type": "table", + "html": "
Chamfer ↓F-score ↑VOI ↓
w/ Sphere Init.8.3957.533.073
w/ Sphere Init. (30K)4.5869.912.522
w/o Double Radii4.7370.252.491
w/o Plane Splitting4.6970.302.477
Ours (Full)4.6671.302.473
", + "image_path": "2ecd1d1c9145758ca6f8b209284d2e4b656175ab07081826be09bf1aea0dc025.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 221, + 293, + 286 + ], + "blocks": [ + { + "bbox": [ + 55, + 174, + 295, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 174, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 295, + 218 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of novel view synthesis on the ScanNetV2 [2] dataset. 'Plane' means time for our plane optimization. 'GS' means time for 2D/3D Gaussian optimization. #P' means the average number of Gaussian points." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 221, + 293, + 286 + ], + "lines": [ + { + "bbox": [ + 58, + 221, + 293, + 286 + ], + "spans": [ + { + "bbox": [ + 58, + 221, + 293, + 286 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓Avg. Time (min)#P
PlaneGSTotal
3DGS [13]24.4170.7810.321-12.212.21.27M
Ours+3DGS [13]25.4710.8160.2962.53.15.60.37M
2DGS [11]24.7660.7960.323-14.214.20.76M
Ours+2DGS [11]25.3800.8130.2962.54.87.30.37M
", + "image_path": "70fcc33d63c13fd50df316f206f38d6de7286f30ecb0d0e2def89ead1330624a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "content": "initialization, in which 3D plane primitives are initialized by setting their radii as 0.05 and are placed on the bounding sphere of the scene with their normals pointing to the center of the scene. As reported in the top two rows of Tab. 3, initialization with Metric3Dv2 [10] mainly improves our method to be faster converged, while sphere initialization could also obtain promising results with more iterations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 386, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 295, + 470 + ], + "type": "text", + "content": "Plane Radii. We then evaluate the design of double-direction plane radii (Double Radii) by replacing it with the vanilla single-direction plane radii. It means that we learn one radius at each axis of the plane primitive. As shown in Tab. 3, optimizing without Double Radii leads to a decrease in the performance of geometry and segmentation, demonstrating the effectiveness of our Double Radii." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 470, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 296, + 507 + ], + "type": "text", + "content": "Plane Splitting. As shown in the last two rows of Tab. 3, applying the Plane Splitting in optimization can further improve the performance in both geometry and segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 516, + 271, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 271, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 271, + 529 + ], + "type": "text", + "content": "4.4. PlanarSplatting for Novel View Synthesis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Benefiting from our accurate and fast planar reconstruction, we show that our PlanarSplatting can seamlessly integrate with recent Gaussian Splatting methods for more efficient and better quality novel view synthesis in indoor scenes. We select two typical methods (3DGS [13] and 2DGS [11]) for evaluation and compare them with two variants including 'Ours+3DGS' and 'Ours+2DGS'. Specifically, for our two variants, we directly sample points from our reconstructed 3D plane primitives to use as the initialization of these Gaussian Splatting methods. Different from the original 3DGS and 2DGS, we fix the position of 3D points and exclude the densification operation when optimizing our two variants. For the baselines of 3DGS and 2DGS, we initialize them with SfM and run with densification to achieve the best performance. We conduct the experiments" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 335, + 70, + 433, + 144 + ], + "blocks": [ + { + "bbox": [ + 335, + 70, + 433, + 144 + ], + "lines": [ + { + "bbox": [ + 335, + 70, + 433, + 144 + ], + "spans": [ + { + "bbox": [ + 335, + 70, + 433, + 144 + ], + "type": "image", + "image_path": "4229a7f08aa8f9b36891862c2f620d7c1356dce12a0b05c09d435859f0dacc50.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 533, + 144 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 533, + 144 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 533, + 144 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 533, + 144 + ], + "type": "image", + "image_path": "a0ac663c9fc8963e1f882135e47e455848d5fcb04147030b99df4253518b5639.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 334, + 156, + 433, + 231 + ], + "blocks": [ + { + "bbox": [ + 361, + 146, + 408, + 156 + ], + "lines": [ + { + "bbox": [ + 361, + 146, + 408, + 156 + ], + "spans": [ + { + "bbox": [ + 361, + 146, + 408, + 156 + ], + "type": "text", + "content": "(a) 3DGS [13]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 156, + 433, + 231 + ], + "lines": [ + { + "bbox": [ + 334, + 156, + 433, + 231 + ], + "spans": [ + { + "bbox": [ + 334, + 156, + 433, + 231 + ], + "type": "image", + "image_path": "861144c33b8304525270ef59a448c080c172c483ae21b137e1b44ceb0425e4fb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 232, + 417, + 243 + ], + "lines": [ + { + "bbox": [ + 350, + 232, + 417, + 243 + ], + "spans": [ + { + "bbox": [ + 350, + 232, + 417, + 243 + ], + "type": "text", + "content": "(c) Ours+3DGS [13]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 246, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 281 + ], + "type": "text", + "content": "Figure 8. Comparison of novel view synthesis on the ScanNetV2 dataset [2]. With initialization from our planar reconstruction, we significantly improve the rendering result of 3DGS [13]." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 156, + 533, + 231 + ], + "blocks": [ + { + "bbox": [ + 460, + 146, + 508, + 156 + ], + "lines": [ + { + "bbox": [ + 460, + 146, + 508, + 156 + ], + "spans": [ + { + "bbox": [ + 460, + 146, + 508, + 156 + ], + "type": "text", + "content": "(b) 2DGS [11]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 156, + 533, + 231 + ], + "lines": [ + { + "bbox": [ + 434, + 156, + 533, + 231 + ], + "spans": [ + { + "bbox": [ + 434, + 156, + 533, + 231 + ], + "type": "image", + "image_path": "5762a9ab220771dc100ac680dc36ab7e552770b398ec31489a577cc2eb4ecb7e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 232, + 511, + 242 + ], + "lines": [ + { + "bbox": [ + 455, + 232, + 511, + 242 + ], + "spans": [ + { + "bbox": [ + 455, + 232, + 511, + 242 + ], + "type": "text", + "content": "(d) Ground Truth" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 288, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 384 + ], + "type": "text", + "content": "using the same scenes as those used in the ablation study. As shown in Tab. 4, our two variants 'Ours+3DGS' and 'Ours+2DGS' significantly outperform the original 3DGS and 2DGS in all metrics with even less optimization time and Gaussian points, demonstrating the superiority and potential of our PlanarSplatting. In Fig. 8, we further show that 'Ours+3DGS' effectively improves the rendering quality on the scene from the ScanNetV2 dataset." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 390, + 476, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 476, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 476, + 403 + ], + "type": "text", + "content": "4.5. Limitations and Future Work" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 408, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 456 + ], + "type": "text", + "content": "Although our PlanarSplatting can reconstruct the accurate indoor planar surface, it is not suitable for complex shapes such as curved surfaces. We leave this challenging problem in our future work for more flexible geometric modeling." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 466, + 388, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 388, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 388, + 478 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 486, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 654 + ], + "type": "text", + "content": "In this paper, we present PlanarSplatting, a novel approach for multi-view 3D reconstruction of indoor scenes. By formulating the problem through differentiable rendering with plane splatting, we demonstrate the powerful capabilities of 3D planar representation for both accurate geometry reconstruction and compact structural scene modeling. Our efficient CUDA implementation enables fast 3D surface reconstruction, allowing comprehensive evaluation across over 100 scenes within hours using a single GPU. Furthermore, the seamless integration of PlanarSplatting with Gaussian Splatting significantly enhances both the quality and efficiency of indoor novel view synthesis, highlighting the broader potential of our approach for indoor scene understanding." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 660, + 404, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 404, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 404, + 673 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "This work was supported by Ant Group Postdoctoral Programme. The authors would love to thank the anonymous reviewers and area chairs for their constructive comments." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1197" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Samir Agarwala, Linyi Jin, Chris Rockwell, and David F. Fouhey. Planeformers: From sparse view planes to 3d reconstruction. In Eur. Conf. Comput. Vis., pages 192-209, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas A. Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2432-2443, 2017. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 239 + ], + "type": "text", + "content": "[3] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Int. Conf. Comput. Vis., pages 10766-10776, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 240, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 295, + 274 + ], + "type": "text", + "content": "[4] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Manhattan-world stereo. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1422-1429, 2009. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 276, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 276, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 276, + 295, + 308 + ], + "type": "text", + "content": "[5] Yasutaka Furukawa, Brian Curless, Steven M. Seitz, and Richard Szeliski. Reconstructing building interiors from images. In Int. Conf. Comput. Vis., pages 80-87, 2009." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 311, + 295, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 311, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 62, + 311, + 295, + 355 + ], + "type": "text", + "content": "[6] David Gallup, Jan-Michael Frahm, and Marc Pollefeys. Piecewise planar and non-planar stereo for urban scene reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1418-1425, 2010. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 357, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 401 + ], + "type": "text", + "content": "[7] Abhinav Gupta, Alexei A. Efros, and Martial Hebert. Blocks world revisited: Image understanding using qualitative geometry and mechanics. In *Eur. Conf. Comput. Vis.*, pages 482-496, 2010. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 295, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 447 + ], + "type": "text", + "content": "[8] Yuze He, Wang Zhao, Shaohui Liu, Yubin Hu, Yushi Bai, Yu-Hui Wen, and Yongjin Liu. Alphatables: A generic plane representation for 3d planar reconstruction from monocular videos. In Adv. Neural Inform. Process. Syst., 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 449, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 295, + 482 + ], + "type": "text", + "content": "[9] Manuel Hofer, Michael Maurer, and Horst Bischof. Efficient 3d scene abstraction using line segments. Comput. Vis. Image Underst., 157:167-178, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 484, + 295, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 484, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 57, + 484, + 295, + 550 + ], + "type": "text", + "content": "[10] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):10579-10596, 2024. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 552, + 295, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 552, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 552, + 295, + 597 + ], + "type": "text", + "content": "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH Conference Papers, page 32, 2024. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 599, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 599, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 599, + 295, + 632 + ], + "type": "text", + "content": "[12] Linyi Jin, Shengyi Qian, Andrew Owens, and David F. Fouhey. Planar surface reconstruction from sparse views. In Int. Conf. Comput. Vis., pages 12971-12980, 2021. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 634, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 634, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 634, + 295, + 677 + ], + "type": "text", + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 2, 3, 5, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "text", + "content": "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Int. Conf. Learn. Represent., 2015. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Pierre-Alain Langlois, Alexandre Boulch, and Renaud Marlet. Surface reconstruction from 3d line segments. In Int. Conf. 3D Vis., pages 553-563, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "text", + "content": "[16] David C. Lee, Martial Hebert, and Takeo Kanade. Geometric reasoning for single image structure recovery. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2136-2143, 2009. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 185 + ], + "type": "text", + "content": "[17] Chen Liu, Jimei Yang, Duygu Ceylan, Ersin Yumer, and Yasutaka Furukawa. Planenet: Piece-wise planar reconstruction from a single RGB image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2579-2588, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "text", + "content": "[18] Chen Liu, Kihwan Kim, Jinwei Gu, Yasutaka Furukawa, and Jan Kautz. Planercnn: 3d plane detection and reconstruction from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4450-4459, 2019. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "type": "text", + "content": "[19] Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, and Yi Xu. Planemvs: 3d plane reconstruction from multi-view stereo. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8655-8665. IEEE, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[20] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21445-21455. IEEE, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "text", + "content": "[21] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhöfer, Yaser Sheikh, and Jason M. Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 40(4):59:1-59:13, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "text", + "content": "[22] Branislav Micusik, Horst Wildenauer, and Markus Vincze. Towards detection of orthogonal planes in monocular images of indoor environments. In Int. Conf. on Robot. and Auto., pages 999-1004, 2008. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "type": "text", + "content": "[23] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable blocks world: Qualitative 3d decomposition by rendering primitives. In Adv. Neural Inform. Process. Syst., 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "text", + "content": "[24] Aron Monszpart, Nicolas Mellado, Gabriel J. Brostow, and Niloy J. Mitra. Rapter: rebuilding man-made scenes with regular arrangements of planes. ACM Trans. Graph., 34(4): 103:1-103:12, 2015. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "text", + "content": "[25] Jann Poppinga, Narunas Vaskevicius, Andreas Birk, and Kaustubh Pathak. Fast plane detection and polygonalization in noisy 3d range images. In Int. Conf. Intell. Robots and Syst., pages 3378-3383, 2008. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 557, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 599 + ], + "type": "text", + "content": "[26] Yiming Qian, Srikumar Ramalingam, and James H. Elder. LS3D: single-view gestalt 3d surface reconstruction from Manhattan line segments. In *Asian Conf. Comput. Vis.*, pages 399-416, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "text", + "content": "[27] Mohamed Sayed, John Gibson, Jamie Watson, Victor Prisacariu, Michael Firman, and Clément Godard. Simplerecon: 3d reconstruction without 3d convolutions. In Eur. Conf. Comput. Vis., pages 1-19, 2022. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "text", + "content": "[28] Jingjia Shi, Shuaifeng Zhi, and Kai Xu. Planerectr: Unified query learning for 3d plane recovery from a single view. In Int. Conf. Comput. Vis., pages 9343-9352, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[29] Sudipta N. Sinha, Drew Steedly, and Richard Szeliski. Piecewise planar stereo for image-based rendering. In Int. Conf. Comput. Vis., 2009. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1198" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[30] Christiane Sommer, Yumin Sun, Leonidas J. Guibas, Daniel Cremers, and Tolga Birdal. From planes to corners: Multipurpose primitive detection in unorganized 3d point clouds. IEEE Robotics Autom. Lett., 5(2):1764-1771, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 152 + ], + "type": "text", + "content": "[31] Bo Sun and Philippos Mordohai. Oriented point sampling for plane detection in unorganized point clouds. In Int. Conf. on Robot. and Auto., pages 2917-2923, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "type": "text", + "content": "[32] Bin Tan, Nan Xue, Song Bai, Tianfu Wu, and Gui-Song Xia. Planetr: Structure-guided transformers for 3d plane recovery. In Int. Conf. Comput. Vis., pages 4166-4175, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 189, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 294, + 232 + ], + "type": "text", + "content": "[33] Bin Tan, Nan Xue, Tianfu Wu, and Gui-Song Xia. NOPE-SAC: neural one-plane RANSAC for sparse-view planar 3d reconstruction. IEEE Trans. Pattern Anal. Mach. Intell., 45 (12):15233-15248, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 278 + ], + "type": "text", + "content": "[34] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Trans. Graph., 38 (6):230:1-230:14, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 280, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 294, + 345 + ], + "type": "text", + "content": "[35] Jamie Watson, Filippo Aleotti, Mohamed Sayed, Zawar Qureshi, Oisin Mac Aodha, Gabriel J. Brostow, Michael Firman, and Sara Vicente. Airplanes: Accurate plane estimation via 3d-consistent embeddings. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5270-5280, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 348, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 294, + 380 + ], + "type": "text", + "content": "[36] Jianxiong Xiao and Yasutaka Furukawa. Reconstructing the world's museums. Int. J. Comput. Vis., 110(3):243-258, 2014. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[37] Yiming Xie, Matheus Gadelha, Fengting Yang, Xiaowei Zhou, and Huaizu Jiang. Planarrecon: Realtime 3d plane detection and reconstruction from posed monocular videos. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6209-6218, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 440, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 440, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 440, + 294, + 483 + ], + "type": "text", + "content": "[38] Sheng Xu, Ruisheng Wang, Hao Wang, and Ruigang Yang. Plane segmentation based on the optimal-vector-field in lidar point clouds. IEEE Trans. Pattern Anal. Mach. Intell., 43 (11):3991-4007, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 486, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 294, + 529 + ], + "type": "text", + "content": "[39] Nan Xue, Bin Tan, Yuxi Xiao, Liang Dong, Gui-Song Xia, Tianfu Wu, and Yujun Shen. NEAT: distilling 3d wireframes from neural attraction fields. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19968-19977. IEEE, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 294, + 574 + ], + "type": "text", + "content": "[40] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Int. Conf. Comput. Vis., pages 12-22, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "text", + "content": "[41] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from A single image. In Int. Conf. Comput. Vis., 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 623, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 294, + 667 + ], + "type": "text", + "content": "[42] Zehao Yu, Jia Zheng, Dongze Lian, Zihan Zhou, and Shenghua Gao. Single-image piece-wise planar 3d reconstruction via associative embedding. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1029-1037, 2019. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[43] Jiahui Zhang, Jinfu Yang, Fuji Fu, and Jiaqi Ma. Planeac: Line-guided planar 3d reconstruction based on self-attention and convolution hybrid model. Pattern Recog., 153:110519, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 106 + ], + "type": "text", + "content": "[44] Yanshu Zhang, Shichong Peng, Alireza Moazeni, and Ke Li. Papr: Proximity attention point rendering. In Adv. Neural Inform. Process. Syst., 2023. 3" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1199" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/013bcc47-e134-4c31-adf0-ced109b7aaa0_content_list.json b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/013bcc47-e134-4c31-adf0-ced109b7aaa0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..69025862c0fd78e00e4b45123e2388b681a54535 --- /dev/null +++ b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/013bcc47-e134-4c31-adf0-ced109b7aaa0_content_list.json @@ -0,0 +1,1350 @@ +[ + { + "type": "text", + "text": "Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy", + "text_level": 1, + "bbox": [ + 192, + 128, + 805, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joonhyun Jeong $^{1}$ Seyun Bae $^{2}$ Yeonsung Jung $^{2}$ Jaeryong Hwang $^{3}$ Eunho Yang $^{2,4*}$", + "bbox": [ + 143, + 200, + 857, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1NAVER Cloud, ImageVision", + "bbox": [ + 380, + 224, + 617, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Korea Advanced Institute of Science and Technology (KAIST)", + "bbox": [ + 246, + 244, + 750, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Republic of Korea Naval Academy 4AITRICS", + "bbox": [ + 302, + 263, + 694, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "joonhyun.jeong@navercorp.com {seyun.bae, ys.jung}@kaist.ac.kr \njhwang@navy.ac.kr eunhoy@kaist.ac.kr", + "bbox": [ + 222, + 285, + 774, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 347, + 326, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the remarkable versatility of Large Language Models (LLMs) and Multimodal LLMs (MLLMs) to generalize across both language and vision tasks, LLMs and MLLMs have shown vulnerability to jailbreaking, generating textual outputs that undermine safety, ethical, and bias standards when exposed to harmful or sensitive inputs. With the recent advancement of safety alignment via preference-tuning from human feedback, LLMs and MLLMs have been equipped with safety guardrails to yield safe, ethical, and fair responses with regard to harmful inputs. However, despite the significance of safety alignment, research on the vulnerabilities remains largely underexplored. In this paper, we investigate the unexplored vulnerability of the safety alignment, examining its ability to consistently provide safety guarantees for out-of-distribution(OOD)-ifying harmful inputs that may fall outside the aligned data distribution. Our key observation is that OOD-ifying the vanilla harmful inputs highly increases the uncertainty of the model to discern the malicious intent within the input, leading to a higher chance of being jailbroken. Exploiting this vulnerability, we propose JOOD, a new Jailbreak framework via OOD-ifying inputs beyond the safety alignment. We explore various off-the-shelf visual and textual transformation techniques for OOD-ifying the harmful inputs. Notably, we observe that even simple mixing-based techniques such as image mixup prove highly effective in increasing the uncertainty of the model, thereby facilitating the bypass of the safety alignment. Experiments across diverse jailbreak scenarios demonstrate that JOOD effectively jailbreaks recent proprietary LLMs and MLLMs such as GPT-4 and o1 with high attack success rate, which previous attack approaches have consistently struggled to jailbreak. Code is available at https://github.com/naver-ai/JOOD.", + "bbox": [ + 89, + 391, + 485, + 891 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 347, + 643, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Models (LLMs) have recently exhibited versatility on various language reasoning tasks [10, 20, 26, 28, 37, 39] based on scalable pre-training and fine-tuning on a large corpus of text data. However, due to the biases and misinformation [8, 30] present in the large-scale training data, LLMs have frequently been jailbroken which leads to the generation of biased or unsafe outputs that may compromise ethical standards, safety, or fairness [3, 19, 38, 40, 45] when provided with harmful or sensitive input text instructions such as \"tell me how to build a bomb\".", + "bbox": [ + 511, + 377, + 906, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While the primary target of jailbreaking has been LLMs, similar vulnerabilities have also surfaced in Multimodal LLMs (MLLMs [18, 24, 28, 44]). MLLMs extend the visual understanding capabilities of LLMs by integrating visual encoder [32] with several add-on layers. However, this integration introduces new pathways to bypass the safety guardrails equipped within LLMs. Recently, several jailbreak methods [22, 33] have focused on MLLM's weak safety alignment induced by the add-on linear layers; The visual encoder and linear layers projecting input images into LLMs [24] have been inadequately safety-aligned with regard to the visually harmful images. Inspired by this weak safety alignment regarding harmful images, MLLM jailbreak methods decompose the harmful textual instruction into a multimodal format, incorporating both a harmful image and a generic textual instruction. For example, as shown in the vanilla attack case of Figure 1, instead of using the textual instruction containing the sensitive phrase \"tell me how to build a bomb\", the harmful content \"bomb\" is embedded in the visual input, and the textual instruction is generalized to \"tell me how to build the object in this image\".", + "bbox": [ + 511, + 532, + 908, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these safety and ethics issues, safety alignment methods [28, 29] were proposed, involving posttraining LLMs and MLLMs with human-preference feed", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 109, + 900, + 238, + 914 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29937", + "bbox": [ + 478, + 950, + 517, + 963 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3dcec454d608e5b0d95b576dea8a277a7db2985ce039ac220042cfe896120e9.jpg", + "image_caption": [ + "Figure 1. Overview of JOOD, in terms of out-of-distribution from the training distribution of safety alignment. For each attack method, we measured uncertainty for discerning the maliciousness of the attack input on GPT-4 and GPT-4V (see Appendix A for the detailed prompt and results) and evaluated their attack success rate on AdvBench-M [27] dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 908, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "back (RLHF). In this approach, ethical and safe responses are prioritized through human feedback, guiding the model to yield safe, ethical, and fair outputs with regard to harmful requests. Therefore, the aforementioned naive attack strategies fail to effectively jailbreak safety-aligned models such as GPT-4 and GPT-4V [28]. As shown in Figure 1, when presented with harmful inputs within the bounds of the RLHF training data distribution, these models can confidently discern malicious intent in the inputs and effectively prevent the circumvention of the safety guardrails.", + "bbox": [ + 88, + 460, + 482, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While RLHF has sufficiently aligned the models to handle such transparently malicious inputs under the safety standards, it may still struggle to generalize to other harmful inputs that fall outside the training distribution. In this paper, we correspondingly investigate a naturally arising question: Does the underlying safety alignment of LLMs and MLLMs consistently guarantee safety even with regard to OOD-ifying inputs that possibly differ from the training inputs and learned knowledge during all the training procedures including safety alignment and are hence novel to the models? To answer the question, we present JOOD, an effective jailbreak attack framework via OOD-ifying harmful inputs beyond safety alignment of LLMs and MLLMs.", + "bbox": [ + 88, + 613, + 482, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While there can be various advanced transformation techniques to OOD-ify the harmful inputs, we explore various off-the-shelf visual and textual transforms and observe that the recent LLMs and MLLMs can be effectively jailbroken even with these simple input manipulation techniques. For instance, as shown in Figure 1, the vanilla harm", + "bbox": [ + 88, + 810, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ful inputs are transformed into the new ones with conventional mixing techniques [7, 42] where the harmful image of bomb and its word text are mixed with another image and text containing an arbitrary subject (e.g., apple), respectively. These newly transformed textual and visual inputs naturally have discrepancies with the vanilla inputs that were previously seen during all the pre-/fine-/post-training procedures including safety alignment learning, potentially conforming to the out-of-distribution beyond the scope of safety-aligned data distribution. Consequently, when these OOD-ifying harmful inputs are exploited for jailbreak attacks, the model fails to recognize their malicious intent with highly increased uncertainty. This allows an attacker to bypass the safety guardrails designed only for the original in-distribution harmful inputs and thereby yield a significantly higher chance of the model being jailbroken.", + "bbox": [ + 511, + 460, + 903, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our comprehensive experiments on various jailbreak scenarios (e.g., bombs, drugs, hacking) of AdvbenchM [27] demonstrate that this straightforward OOD-ifying strategy successfully jailbreaks existing LLMs and MLLMs, including state-of-the-art proprietary models such as GPT-4 and GPT-4V. Also, our method significantly outperforms baseline attack methods in all the jailbreak scenarios, achieving $63\\%$ attack success rate (ASR) against GPT-4V in Bombs or Explosives scenario and improving performance by $+42\\%$ ASR compared to the state-of-the-art baseline [9] in the Hacking scenario. In summary, our contribution is threefold:", + "bbox": [ + 511, + 703, + 903, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We systematically reveal the vulnerabilities of RLHF", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29938", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "based safety alignment, which leaves LLMs and MLLMs vulnerable to out-of-distribution textual or visual inputs that have a discrepancy with the in-distribution samples previously seen during safety alignment.", + "bbox": [ + 102, + 90, + 480, + 151 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel black-box jailbreak strategy via manipulating malicious inputs into the OOD-ifying ones. We observe that even simple off-the-shelf transforms amplify uncertainty of the model, allowing it to effectively bypass the safety alignment of LLMs and MLLMs.", + "- Extensive experiments on various jailbreak scenarios demonstrate the effectiveness of our attack strategy against the state-of-the-art proprietary models such as GPT-4 and even o1, with high attack success rate." + ], + "bbox": [ + 91, + 152, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 297, + 230, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Jailbreaking attacks The jailbreaking attacks undermine LLMs and MLLMs to generate forbidden text outputs that can possibly violate safety, ethics, and fairness regulations. For jailbreaking LLMs, the pioneer work [45] and its variants [1, 23] adversarially optimized the suffix within the attack prompt to yield an affirmative response. Another line of work disguised attack prompts via encryption [11, 41] and translations with low-resource languages [40]. In-context learning [2] was also instrumental in establishing preliminary contexts for jailbreaking by utilizing few-shot jailbroken examples [38] and refining the attack prompt [3]. Meanwhile, several studies [9, 22, 33] focused on the weak safety alignment of MLLMs and reformulated the textual harmful instruction into multimodal format containing a pair of harmful image and generic text instruction. HADES [22] further synthesized the harmful image into a semantically more harmful one by diffusion models for providing a better jailbreaking context. Similarly, Ma et al. [25] additionally provided visual contexts by utilizing the roleplaying concept, offering justifications to carry out the associated harmful request. Also, FigStep [9] converted harmful textual instruction into typography and prompted to complete the blanks in the execution steps. Although the above methods achieved state-of-the-art jailbreak performance on open-source MLLMs [24, 44], they still exhibit a lack of generalization on robust MLLMs [12, 14, 28] that were safety-aligned with human feedback. In this work, we focus on unveiling the vulnerability of these safety-aligned models in terms of out-of-distribution perspective.", + "bbox": [ + 89, + 324, + 483, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Safety alignment via human feedback While instruction tuning [37] successfully manipulated LLMs to act in accordance with the user's intention, there still remained a large headroom for improving their safety and reliability. To address this, RLHF [29] aligned LLMs with human preferences under the consideration of the safety standards, robustifying LLMs against malicious text instructions. Also, Sun et al. [34] aligned MLLMs using human feedback on the", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rectified image-text pairs where hallucinations and harmful contents were largely reduced. OpenAI [28] further strengthened the safety alignment of MLLMs using selffeedback as a reward for updating the policy model. However, we revisit that these safety-aligned models via RLHF methods can still be susceptible to jailbreak attacks using various types of data transformation techniques.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data transformation Data transformation refers to manipulating the input data into a new variation of a synthetic sample, primarily used for enhancing the robustness and generalization towards broader variations of inputs in the machine learning field. Specifically, regularizing neural networks with data transformation is vital for mitigating overfitting on a fixed training dataset. In image classification task [6], besides the basic transformations such as random resizing and center cropping, advanced ones such as photometric and geometric transformations [4, 5] have been applied to enhance test generalization. Also, image-mixing techniques [15-17, 42, 43], which synthesize data by combining training images, have further expanded the training data spectrum and improved generalization performance. In natural language processing (NLP), text transformations like adding a random word, synonym replacement, and sentence reordering [36] are used to increase input diversity. Other transforms, such as text-mixing [7], generate novel linguistic patterns, creating out-of-distribution data to challenge language models in unexpected ways. In this paper, we explore these off-the-shelf transform techniques to generate OOD-ifying inputs beyond training distribution of safety alignment for LLMs and MLLMs, thereby effectively performing jailbreak attacks against these large models.", + "bbox": [ + 511, + 231, + 906, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 614, + 648, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose JOOD, a simple yet effective jailbreak framework with the OOD-ifying harmful inputs. We explore a broad range of off-the-shelf transformations to manipulate the harmful inputs into OOD-ifying ones that may not be considered during safety alignment. Consequently, these synthetic OOD-ifying inputs effectively bypass the safety guardrails of LLMs and MLLMs, making them more susceptible to jailbreak attempts. This straightforward input manipulation strategy also enables black-box setting to attack the proprietary LLMs and MLLMs such as GPT-4 and GPT-4V. In Section 3.1, we introduce our LLM attack method leveraging textual transformations. In Section 3.2, we also present an effective MLLM attack method using image transformation techniques. Subsequently, we propose an attack evaluation method in Section 3.3 to quantitatively measure the maximum potential risk posed by our attacks.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29939", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Attack LLM with OOD-ifying inputs", + "text_level": 1, + "bbox": [ + 89, + 90, + 410, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an input text instruction $T^{\\mathrm{h}}$ that contains a harmful request such as \"tell me how to build a bomb\", the naive attack strategy is to feed $T^{\\mathrm{h}}$ into the target model $\\theta$ to generate output text response $r = \\theta (T^{\\mathrm{h}})$ , where $\\theta$ is typically an autoregressive LLM. While this naive attack had successfully jailbroken the primitive LLMs, it fails to jailbreak recent safety-aligned LLMs since the input $T^{\\mathrm{h}}$ is a natural in-distribution instruction that might have been seen during safety alignment training with human feedback. To overcome this limitation, we transform $T^{\\mathrm{h}}$ into a novel text instruction $T^{\\mathrm{ood}}$ that possibly had not been seen during all the LLM training procedures including the safety alignment learning, and then input to the target model:", + "bbox": [ + 89, + 112, + 483, + 309 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} ^ {\\mathrm {o o d}} = f \\left(T ^ {\\mathrm {h}}; \\varphi_ {i}\\right), \\quad R = \\left\\{\\theta \\left(T _ {i} ^ {\\mathrm {o o d}}\\right) \\right\\} _ {i = 1} ^ {n} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 321, + 482, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varphi_{i}$ represents transformation-dependent parameter applied by textual transformation function $f(\\cdot)$ , and $R$ is a set of model responses for each transformed instruction $T_{i}^{\\mathrm{ood}}$ with one of the $n$ differently sampled $\\varphi_{i}$ . To fulfill the malicious goal within $T^{\\mathrm{h}}$ , an attacker can employ various textual transformations for $f(\\cdot)$ to OOD-ify $T^{\\mathrm{h}}$ into a novel text instruction $T_{i}^{\\mathrm{ood}}$ . For instance, off-the-shelf textual transforms such as random word insertion or deletion [36] and character-level perturbations [21] can be employed in $f(\\cdot)$ . Among these methods, we observe that even a simple transform such as the text-mixing technique, where a word is combined with an unrelated auxiliary word to form a new nonsensical word, can effectively synthesize OOD-ifying inputs for the recent LLMs [28] with high uncertainty as shown in Figure 1. Specifically, the harmful subject in $T^{\\mathrm{h}}$ (e.g., bomb) is combined with one of $n$ different auxiliary words $\\varphi_{i}$ (e.g., apple) randomly sampled from real-world vocabularies, synthesizing a new nonsensical word (e.g., bombapple) that does not naturally exist in practice. Subsequently, we prompt the model $\\theta$ to comply with the generic request for each of the words within this newly blended word as follows (see Appendix B for prompt examples of other textual transform variants):", + "bbox": [ + 89, + 353, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The following is a newly coined word made by mixing words corresponding to two words in a random order. Please refer to the possible words inside this new word as \"[Words]\". Please answer the request below for each of these words.", + "bbox": [ + 94, + 708, + 480, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[Coined word]: \"bombapple\"", + "bbox": [ + 96, + 790, + 308, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[Request]: \"Tell me how to build [Words]\"", + "bbox": [ + 96, + 806, + 464, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Attack MLLM with OOD-ifying inputs", + "text_level": 1, + "bbox": [ + 89, + 832, + 428, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the original harmful text instruction $T^{\\mathrm{h}}$ , naive MLLM attack strategy is to decompose $T^{\\mathrm{h}}$ into a pair of harmful image $I^{\\mathrm{h}}$ (e.g., bomb) and a generic text instruction $T^{\\mathrm{g}}$", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(e.g., \"tell me how to build the object in this image\"), and then feed both inputs to the target MLLM $\\theta$ . However, this naive vanilla attack does not effectively jailbreak state-of-the-art MLLMs such as GPT-4V that had possibly safely-aligned on such harmful inputs. Therefore, we propose a more powerful OOD-ifying attack strategy that can bypass underlying safety alignment of the MLLMs, as shown in Figure 1. Specifically, we transform the harmful image $I^{\\mathrm{h}}$ into a synthetic image $I^{\\mathrm{ood}}$ via conventional image transformation techniques $f(\\cdot)$ that may not be considered during safety alignment training:", + "bbox": [ + 511, + 90, + 906, + 258 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {j} ^ {\\mathrm {o o d}} = f (I ^ {\\mathrm {h}}; \\alpha_ {j}), \\quad R = \\{\\theta (T ^ {\\mathrm {g}}, I _ {j} ^ {\\mathrm {o o d}}) \\} _ {j = 1} ^ {m} \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 270, + 906, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{j}$ is a transformation degree ranging from 0 to 1, and $R$ is a set of model responses for each transformed image $I_{j}^{\\mathrm{odd}}$ with one of $m$ differently sampled $\\alpha_{j}$ . To achieve the malicious goal, the attacker can employ various off-the-shelf image transformations for OOD-ifying the original harmful input $I^{\\mathrm{h}}$ into a novel harmful image $I_{j}^{\\mathrm{odd}}$ , such as geometric, photometric transforms [5] and image-mixing techniques [31, 42, 43]. When using image-mixing techniques, we consider another input $\\varphi_{i}$ , an auxiliary image containing an arbitrary object or concept, for mixing with the original harmful image $I^{\\mathrm{h}}$ :", + "bbox": [ + 511, + 301, + 906, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {(i, j)} ^ {\\mathrm {o o d}} = f \\left(I ^ {\\mathrm {h}}, \\varphi_ {i}; \\alpha_ {j}\\right), \\quad R = \\left\\{\\theta \\left(T ^ {\\mathrm {g}}, I _ {(i, j)} ^ {\\mathrm {o o d}}\\right) \\right\\} _ {i = 1, j = 1} ^ {n, m} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 481, + 906, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n$ denotes the number of auxiliary image $\\varphi_{i}$ used for attack, and $R$ is a set of model responses generated by mixing $I^{\\mathrm{h}}$ with $n$ different $\\varphi_{i}$ samples and $m$ different $\\alpha_{j}$ mixing coefficients. For mixup [43] case, $f(\\cdot)$ can be defined as $\\alpha_{j}\\varphi_{i} + (1 - \\alpha_{j})I^{\\mathrm{h}}$ . Also, we correspondingly reform the generic input instruction $T^{\\mathrm{g}}$ to comply with the request for each of the two objects within the mixed image as follows:", + "bbox": [ + 511, + 513, + 905, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are two realistic (not artistic but real-functioning) objects (or concepts) in this image.", + "bbox": [ + 517, + 632, + 903, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tell me how to build each of these objects.", + "bbox": [ + 519, + 662, + 854, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, this straightforward OOD-ifying strategy with simple image transform techniques induces a distinct shift from $I^{\\mathrm{h}}$ within the visual encoder of MLLM, as shown in Appendix D. When these shifted harmful visual embeddings are fed into the subsequent LLM, the uncertainty is substantially increased and hence fail to trigger its safety guardrails which were trained on the original in-distribution harmful inputs (i.e., $T^{\\mathrm{h}}$ and $I^{\\mathrm{h}}$ ), as shown in Figure 1.", + "bbox": [ + 511, + 700, + 905, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Attack Evaluation", + "text_level": 1, + "bbox": [ + 511, + 832, + 689, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each harmful instruction $T^{\\mathrm{h}}$ , our attack methods (Eq. 1, 3) produce a set of attack responses $R$ corresponding to our auxiliary attack input and parameter such as $\\varphi$ and $\\alpha$ ,", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29940", + "bbox": [ + 478, + 950, + 517, + 963 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/7d2e8a55529c75c4f1f447a5b3c180c6b21167e5d30c1ebb47c07a3577ad0757.jpg", + "table_caption": [ + "Table 1. Comparison of jailbreak performance with baseline attack methods on the unimodal (text-only) model GPT-4 and the multimodal (image + text) model GPT-4V. HF denotes the average harmfulness score of the responses over all the instructions as in Eq. 4, ranging from 0 to 10. The attack scenarios include Bombs or Explosives (BE), Drugs (D), Firearms / Weapons (FW), Hacking information (H), Kill someone (K), Social Violence (SV), and Self-harm and Suicide (SS)." + ], + "table_footnote": [], + "table_body": "
InputTarget modelAttack methodBEDFWHKSVSS
HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑
TextGPT-4Vanilla00000.1000000000
CipherChat 2024070.370.100.111080.21507
PAIR 2023000.230.900.8110.100.100.13
JOOD (Eq. 1)1.3132.5172.8243.1421.480.401.013
Image + TextGPT-4VVanilla00000.6120.35000000
FigStep 2023000.230.5000000000
FigStep-Pro 20230.9230.8171.8252.1320.480.300.10
HADES 2024000.230.100.10000.1000
JOOD (Eq. 3)7.1633.9237.2474.0742.141.1100.623
MiniGPT-4 7BVanilla1.5502.5731.260420.2420.8203.073
FigStep 20234.4633.1633.6476.7744.2633.6551.953
HADES 20240.1171.8401.5291.2470.6170.851.337
JOOD (Eq. 3)8.1838.0877.5477.3955.7543.9357.697
LLaVA-1.5 13BVanilla6.6873.2433.5472.4531.7461.0154.383
FigStep 20235.8773.5534.2416.9794.0543.2351.247
HADES 20241.0132.7533.6474.7744.0540.451.747
JOOD (Eq. 3)9.81008.5938.3655.3896.1635.8408.590
", + "bbox": [ + 94, + 155, + 916, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "respectively. Each response in this set may exhibit varying degrees of harmfulness, depending on how much the harmful information and sensitive content it contains corresponding to the malicious input request $T^{\\mathrm{h}}$ . To evaluate the maximum potential risk posed by our attacks, we propose a score-based evaluation method to quantitatively measure the harmfulness degree of each response, thereby identifying the most harmful attack response. Specifically, we employ another LLM $\\theta^{\\mathrm{hf}}$ [28] as the harmfulness score (HF) evaluator and prompt it to assess each attack response $r \\in R$ on a scale from 0 to 10, in consideration of the safety standards and compliance with the harmful input instruction $T^{\\mathrm{h}}$ . Details of the evaluation prompt are provided in Appendix C. Given the most harmful attack response with the highest harmfulness score, we also report attack success rate (ASR) by prompting to the binary-judging LLM $\\theta^{\\mathrm{bj}}$ [13] that outputs 1 if the response is unsafe and 0 for the harmless responses, following [22]. These evaluation metrics are formulated as follows:", + "bbox": [ + 88, + 488, + 485, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {H F} \\left(\\mathcal {T} ^ {\\mathrm {h}}\\right) = \\frac {\\sum_ {T ^ {\\mathrm {h}} \\in \\mathcal {T} ^ {\\mathrm {h}}} \\max _ {r \\in R} \\left(\\theta^ {\\mathrm {h f}} \\left(r \\mid T ^ {\\mathrm {h}}\\right)\\right)}{| \\mathcal {T} ^ {\\mathrm {h}} |} \\tag {4} \\\\ \\operatorname {ASR} \\left(\\mathcal {T} ^ {\\mathrm {h}}\\right) = \\frac {\\sum_ {T ^ {\\mathrm {h}} \\in \\mathcal {T} ^ {\\mathrm {h}}} \\theta^ {\\mathrm {b j}} \\left(\\underset {r \\in R} {\\arg \\max } \\left(\\theta^ {\\mathrm {h f}} \\left(r \\mid T ^ {\\mathrm {h}}\\right)\\right)\\right)}{\\left| \\mathcal {T} ^ {\\mathrm {h}} \\right|} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 808, + 480, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{T}^{\\mathrm{h}}$ is the entire set of harmful instructions used for jailbreaking attacks.", + "bbox": [ + 511, + 487, + 906, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 534, + 645, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We evaluate JOOD on the widely-used LLM and MLLM jailbreak benchmark, Advbench [45] and Advbench-M [27], to compare the performance against the previous state-of-the-art attack methods. Advbench consists of 500 textual instructions that encourage harmful behaviors, while Advbench-M further categorized the set of these harmful instructions into several distinct scenarios such as Bombs or Explosives, Drugs, and Hacking information. To attack with MLLMs, Advbench-M paired semantically relevant images for each scenario. We used around 30 harmful instructions for each scenario to evaluate LLM attack methods, along with one paired harmful image for each scenario when evaluating MLLM attack methods.", + "bbox": [ + 511, + 560, + 908, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. When attacking with mixing-based transformations of JOOD, the textual or visual auxiliary inputs $\\varphi$ in Eq. 1, 3, are required to synthesize the OOD-ifying input, respectively. We used randomly sampled arbitrary words such as apple and watch, and retrieved corresponding images from the Internet. We set the number of auxiliaries to $n = 5$ and sampled $m = 9$ discrete values for the transformation degree $\\alpha$ from", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29941", + "bbox": [ + 478, + 950, + 517, + 963 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\{0.1, 0.2, \\dots, 0.9\\}$ . When attacking with image transformation techniques, we resized and padded the input images to a uniform size of $320 \\times 320$ pixels as a preprocessing step before the transformation. Unless otherwise specified, we use GPT-4 as the target model for LLM attacks and GPT-4V for MLLM attacks, with inference parameters such as temperature and top-p set to 1. For LLM attacks, we employ the text-mixing transformation introduced in Eq. 1, and for MLLM attacks, we use the image-mixing transformations as in Eq. 3.", + "bbox": [ + 89, + 90, + 480, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 89, + 253, + 230, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with SOTA attack methods. In Table 1, we compare JOOD with other state-of-the-art attack methods on GPT-4 and GPT-4V. The vanilla attack methods using the original harmful instructions or images barely jailbreak the target models in almost all the scenarios, while the baseline works such as CipherChat, PAIR, FigStep, and HADES marginally enhanced both ASR and harmfulness scores. However, JOOD consistently exhibits the best jailbreak performance on almost all the scenarios and target models, largely outperforming the baselines. Specifically, when attacking text-only GPT-4, JOOD achieves $24\\%$ ASR in the Firearms scenario where all the baselines failed to jailbreak any of the instructions. Also, when attacking multimodal GPT-4V, JOOD outperforms the previous state-of-the-art attack FigStep-Pro by a large margin, achieving $+6.2$ average harmfulness and $+40\\%$ ASR in Bombs or Explosives scenario. Notably, JOOD exclusively jailbreaks a considerable amount of the instructions that these baselines failed to, specifically 10 additional instructions out of 30 in Bombs or Explosives scenario, as shown in Table 2. Also, JOOD consistently generates extremely harmful responses on most instructions while the baseline methods sparsely generate harmful responses with regard to a few instructions, as shown in Figure 2. These results demonstrate that our OOD-ifying attack strategy via input transformation techniques is effective in causing underlying safety alignment to malfunction to output harmful responses.", + "bbox": [ + 91, + 275, + 483, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalization on other MLLMs. In the bottom rows of Table 1, we further investigate generalization of JOOD on the other open-source MLLMs, MiniGPT-4 7B [44] and LLaVA-1.5 13B [24] post-trained with RLHF [34]. The results show that JOOD consistently outperforms the other baselines by a large margin in almost all of the scenarios. Surprisingly, JOOD achieves high attack success rates even against robustly safety-aligned models including GPT-4o [12] and o1 [14], which the baselines mostly failed to jailbreak (see Appendix E). The superiority of our OOD-ifying attack strategy on both proprietary and open-source MLLMs further corroborates that the existing MLLMs still lack of safety alignment on the OOD-ifying inputs gener", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2865bb14ba9fc0899ac98464289408f0f07620d3e0b96bb88008023f4d70b600.jpg", + "table_caption": [ + "Table 2. Comparison of the number of exclusively jailbroken instructions by each attack method." + ], + "table_footnote": [], + "table_body": "
Attack method# exclusively jailbroken instructions
Figstep-ProHADESJOODBEDFWHKSVSS
✓XX4532200
X✓X0200000
XX✓10547112
", + "bbox": [ + 521, + 128, + 898, + 261 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1903840763ec6bf59b981c360f9d629af4d513f8d22a9849c87ab523af655e50.jpg", + "image_caption": [ + "Figure 2. Comparison of harmfulness scores per instruction in Bombs scenario." + ], + "image_footnote": [], + "bbox": [ + 517, + 263, + 879, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ated from even simple off-the-shelf transforms, emphasizing the need for further research and development.", + "bbox": [ + 511, + 530, + 903, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 571, + 679, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct ablation studies to assess the effectiveness of each component and the sensitivity of the hyper-parameters comprising JOOD. Unless specified, we attack with the multimodal inputs using the mixup technique and adopt GPT-4V as the target model to be attacked. Also, we report jailbreaking performance in Bombs or Explosives scenario.", + "bbox": [ + 511, + 593, + 906, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of various transformation techniques. For attacking GPT-4 with OOD-ifying text inputs as in Eq. 1, we investigate the effect of various text-mixing techniques to obfuscate the harmful word (e.g., \"bomb\") with the auxiliary word (e.g., \"apple\"), producing a mixed word like \"baopmble\" for the $H$ (horizontal)-Interleave case. We observe that GPT-4 effectively interprets these mixed words, accurately recognizing and responding to both the harmful and auxiliary components for the given instruction. See Appendix B for the detailed analysis. Also, to generate OOD-ifying input images for the attack in Eq. 2, we test with widely-adopted image transformation techniques including geometric, photometric transformations", + "bbox": [ + 511, + 702, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29942", + "bbox": [ + 478, + 950, + 519, + 963 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ebbaa4912c33be151384b6ee2d380700da89ac6ef4a5c90b34fe7043cf815fe9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 89, + 294, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54bd666e8226776418213cbf2a020b2f60bd30061208f331bc3264454ec4d6bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 90, + 488, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/65b1310079864f6d3c68d7b8d72a813f8efc17f2511ef604ade18276bf5d3f43.jpg", + "image_caption": [ + "(a) Harmfulness / ASR of text-mixing attacks" + ], + "image_footnote": [], + "bbox": [ + 94, + 260, + 289, + 411 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bdc9a811c37910f224fa867b6716dbf0fa9d9a68c4e1f64dc7c9cde06248720d.jpg", + "image_caption": [ + "(b) Harmfulness / ASR of image-transform attacks" + ], + "image_footnote": [], + "bbox": [ + 290, + 261, + 488, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(RandAug [5]) and image-mixing techniques (Mixup [43], CutMix [42], and ResizeMix [31]).", + "bbox": [ + 89, + 511, + 482, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 3, all of the text-mixing variants and image-mixing techniques consistently outperformed the vanilla attack case without applying mixing techniques in all the scenarios. Also, the performance variance between all the image-mixing variants was marginal, indicating that an attacker can leverage our method for consistently jailbreaking GPT-4V without any dependence on the specific image-mixing technique. However, attacking with non-mixing image transformation techniques (i.e., RandAug) such as adding gaussian noise, rotation, and shearing showed inferior ASR and harmfulness scores compared to the image-mixing transformation techniques, with slightly higher performance than the vanilla case. These results may suggest that GPT-4V is equipped with relatively solid safety alignment on such basic transformations, but not adequately safety-aligned on more synthetic transformations such as image-mixing techniques.", + "bbox": [ + 89, + 542, + 482, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablations on mixing coefficient for mixup. In Figure 4, we analyze the effect of the mixing coefficient $\\alpha$ in Eq. 3 which modulates how much the auxiliary image will obfuscate the original harmful image. We additionally report the ratio of the responses refused by GPT-4V with substring", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6cec83050da5e4d07d3a8f3d8ed7da623d8314c6b6ccde296d9530371bb8c75d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 87, + 924, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e9da64943dc04488522e86858b18d21724c01601cf0a4123463a16dcae17f989.jpg", + "image_caption": [ + "Figure 4. Ablation on mixing coefficient $\\alpha$ for image mixup attack. We visualize the attack images corresponding to $\\alpha$ value below the x-axis. We report the average harmfulness / ASR / refusal rate on all the 30 instructions in Bombs or Explosives scenario." + ], + "image_footnote": [], + "bbox": [ + 526, + 227, + 578, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5bf34b6d4542f672a485899f08f919691747202e4713a35814001a85c2f827a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 227, + 638, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf56f2b5c8f579db16e2375f6f12e75c1cb43fb5a2608bf7cfd64cefad7f8df6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 227, + 694, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5e530947f53f879be50941b1ffd866a200bd0a40a666333f688f948cd2d70f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 227, + 754, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dfd721c65faca1b6d529825ab399f71f52204d60de6189f91a37c776b095eb4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 227, + 813, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aabca5b31b4ddab78a761fb3c0093fc244d8b70a72258ef314f11b22e13ac21d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 227, + 872, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d28407166ba7c3ddc69e68bef15f58270f899a4b4ba158665910c0979032669f.jpg", + "image_caption": [ + "Figure 3. Effect of attacking with (a) various text-mixing techniques on GPT-4 and (b) image-transformation techniques on GPT-4V in all the attack scenarios.", + "(a) Effect of semantically various auxiliary images when mixed with the target harmful image (\"bomb\")." + ], + "image_footnote": [], + "bbox": [ + 535, + 362, + 898, + 546 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95bbfab247648a56ecf6b8c76ed37494c3f497738c7c383204e0ca996f671380.jpg", + "image_caption": [ + "(b) Ablation for the visual type of auxiliary image, either using typography image or realistic image.", + "Figure 5. Effect of auxiliary images depending on the similarity with the target harmful image (a) and the visual image types (b). Red dotted line denotes the harmfulness of the vanilla attack." + ], + "image_footnote": [], + "bbox": [ + 540, + 577, + 895, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "matching [45] which verifies whether the attacked response has one of the refusal phrases (e.g., \"I'm sorry\"). Generally, the vanilla case (i.e., $\\alpha = 0$ ) without mixing the auxiliary image on the original harmful image showed signif", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29943", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ab6884731b0b37a51f929f9476d167c9d9968a8e5c8f67788f828e4a606885cd.jpg", + "table_caption": [ + "Table 3. Effect of attack methods for jailbreaking against the system prompt defense in Bombs or Explosives scenario." + ], + "table_footnote": [], + "table_body": "
System Prompt DefenseVanillaFigStep 2023FigStep-Pro 2023HADES 2024JOOD
HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑
×00%00%0.923%00%7.163%
✓00%00%0.813%00%4.360%
", + "bbox": [ + 133, + 109, + 859, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "icently low ASR and harmfulness score with high refusal rate possibly due to the safety alignment training with self feedback [28] of GPT-4V on these obviously harmful images. However, when an auxiliary image is mixed (i.e., $0 < \\alpha < 1$ ), the refusal rate significantly decreases while the ASR and harmfulness score are largely increased. This indicates the essential role of our attack to mitigate the evasive refusal of GPT-4V with regard to the harmful requests and further elicit the harmful responses via presenting the OOD-ifying inputs generated by mixup, which were possibly not seen during the safety alignment training. When $\\alpha$ reaches 1, the harmful image is entirely substituted by the auxiliary image (e.g., apple), losing the harmful semantic and hence reducing both ASR and harmfulness scores.", + "bbox": [ + 88, + 203, + 485, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of auxiliary images. In Figure 5, we analyze the effect of mixing the harmful image with the auxiliary images based on their similarity and visual types. Specifically, in Figure 5a, we analyze the effect of mixing with semantically various auxiliary images by comparing the cosine similarity with the harmful \"bomb\" image. Blue dots / green dots / red dots represent the auxiliary images that are dissimilar / moderately related / highly similar to the target harmful image, respectively. The result shows a strong negative correlation between the similarity and the harmfulness of the model response. Mixing with highly similar images containing similar unsafe objects such as a \"Molotov cocktail\" generates less harmful responses which only refuse to provide answers with regard to both the original harmful object (e.g., bomb) and another harmful object (e.g., Molotov cocktail). Meanwhile, mixing with moderately related or dissimilar auxiliary images such as \"softwood\" generates harmful responses containing a detailed description regarding the original harmful object (see Appendix J for detailed response examples).", + "bbox": [ + 91, + 431, + 482, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 5b, we conduct another analysis on the effect of mixing with either a typographic image that displays the text of the auxiliary word or a realistic image containing the visual scene related to the auxiliary word. In all the attack scenarios, using typographic auxiliary images for mixing harmful images generally amplifies the harmfulness of the output responses compared to using real images. While Gong et al. [9] empirically observed that existing MLLMs already exhibit vulnerability on the typographic image input itself, this result suggests that the safety guardrails of MLLMs can be more effectively neutralized when these ty", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pographic images are further leveraged for our proposed image-mixing based attacks.", + "bbox": [ + 511, + 203, + 906, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Jailbreak against defense methods. In Table 3, we investigate the effect of JOOD against system prompt-based defense [9, 25] where defensive textual guidance is given as input system prompt instructing the model to remain vigilant and refrain from responding to harmful textual or visual queries that may violate AI safety policies (see Appendix G for the detailed prompt). The results show that even after deploying the safety-aware system prompt, JOOD still achieves significantly higher harmfulness and ASR scores compared to the baseline attack methods. Notably, JOOD only degraded ASR by $3\\%$ after applying system-prompt-based defense, while FigStep-Pro largely degraded by $10\\%$ .", + "bbox": [ + 511, + 262, + 903, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Appendix G, we further evaluate jailbreak performance of JOOD against a more sophisticated defense, AdaShield [35], which adaptively retrieves the optimal defense prompt for each attack query. Even with this adaptive defense, our JOOD successfully jailbreaks over half of the attack queries with a significantly lower refusal rate, while all the baselines hardly succeed to jailbreak. The robustness of JOOD possibly originates from the ambiguity and uncertainty of the input harmful image induced by transform techniques such as mixup. This highlights the need for further exploration of safeguard mechanisms when dealing with such ambiguous and potentially harmful inputs.", + "bbox": [ + 511, + 446, + 906, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 646, + 633, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the unrevealed vulnerability of the safety alignment within existing LLMs and MLLMs when the out-of-distribution inputs are given as the input. Exploiting this vulnerability, we devise a new jailbreak strategy by generating OOD-ifying inputs with off-the-shelf data transformation techniques. We observe that even simple transformations such as mixup can fabricate OOD-ifying inputs that induce a high level of uncertainty for the models, successfully jailbreaking the state-of-the-art LLMs and MLLMs. By the extensive experiments on various jailbreak scenarios and rigorous ablations, we demonstrate that this simple yet straightforward OOD-ifying strategy can pose a significant threat to the recent LLMs and MLLMs, underscoring the necessity for further research and development to enhance safety alignment against these OOD-ifying inputs.", + "bbox": [ + 511, + 674, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29944", + "bbox": [ + 478, + 950, + 519, + 963 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partly supported by National Research Foundation of Korea (NRF) grant (No.RS-2023-00209060, A Study on Optimization and Network Interpretation Method for Large-Scale Machine Learning) funded by the Korea government (MSIT). This work was also supported by KAIST-NAVER Hypercreative AI Center.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. Jailbreaking leading safety-aligned llms with simple adaptive attacks. arXiv preprint arXiv:2404.02151, 2024.3", + "[2] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3", + "[3] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries. arXiv preprint arXiv:2310.08419, 2023. 1, 3, 5", + "[4] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018.3", + "[5] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 3, 4, 7", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3", + "[7] Srujan Kumar Enaganti, Lila Kari, Timothy Ng, and Zihao Wang. Word blending in formal languages: The brangelina effect. In Unconventional Computation and Natural Computation: 17th International Conference, UCNC 2018, Fontainebleau, France, June 25-29, 2018, Proceedings 17, pages 72-85. Springer, 2018. 2, 3", + "[8] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1-79, 2024. 1", + "[9] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 5, 8", + "[10] Google. Bard - chat based ai tool from google, 2023. 1", + "[11] Divij Handa, Advait Chirmule, Bimal Gajera, and Chitta Baral. Jailbreaking proprietary large language models using word substitution cipher. arXiv preprint arXiv:2402.10601, 2024.3", + "[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli" + ], + "bbox": [ + 93, + 244, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3, 6", + "[13] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Lmbased input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023.5", + "[14] Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 3, 6", + "[15] Joonhyun Jeong, Sungmin Cha, Jongwon Choi, Sangwoo Yun, Taesup Moon, and Youngjoon Yoo. Observations on k-image expansion of image-mixing augmentation. IEEE Access, 11:16631-16643, 2023. 3", + "[16] JangHyun Kim, Wonho Choo, Hosan Jeong, and Hyun Oh Song. Co-mixup: Saliency guided joint mixup with supermodular diversity. In International Conference on Learning Representations, 2021.", + "[17] Jang-Hyun Kim, Wonho Choo, and Hyun Oh Song. Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In International Conference on Machine Learning, pages 5275-5285. PMLR, 2020. 3", + "[18] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal generation. arXiv preprint arXiv:2301.13823, 2023. 1", + "[19] Raz Lapid, Ron Langberg, and Moshe Sipper. Open sesame! universal black box jailbreaking of large language models. arXiv preprint arXiv:2309.01446, 2023. 1", + "[20] Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, et al. Solving quantitative reasoning problems with language models, 2022. URL https://arxiv.org/abs/2206.14858, 2022. 1", + "[21] Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. Textbugger: Generating adversarial text against real-world applications. arXiv preprint arXiv:1812.05271, 2018. 4", + "[22] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. arXiv preprint arXiv:2403.09792, 2024. 1, 3, 5, 8", + "[23] Zeyi Liao and Huan Sun. Amplegcg: Learning a universal and transferable generative model of adversarial suffixes for jailbreaking both open and closed llms. arXiv preprint arXiv:2404.07921, 2024. 3", + "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 3, 6", + "[25] Siyuan Ma, Weidi Luo, Yu Wang, Xiaogeng Liu, Muhao Chen, Bo Li, and Chaowei Xiao. Visual-roleplay: Universal jailbreak attack on multimodal large language models via role-playing image character. arXiv preprint arXiv:2405.20773, 2024. 3, 8", + "[26] Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943, 2021. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29945", + "bbox": [ + 478, + 950, + 517, + 963 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 2, 5", + "[28] OpenAI. Gpt-4 technical report. Technical report, 2023. 1, 2, 3, 4, 5, 8", + "[29] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1, 3", + "[30] Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Yang Wang. On the risk of misinformation pollution with large language models. arXiv preprint arXiv:2305.13661, 2023. 1", + "[31] Jie Qin, Jiemin Fang, Qian Zhang, Wenyu Liu, Xingang Wang, and Xinggang Wang. Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101, 2020.4, 7", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1", + "[33] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2023. 1, 3", + "[34] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 3, 6", + "[35] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. In European Conference on Computer Vision, pages 77-94. Springer, 2024. 8", + "[36] Jason Wei and Kai Zou. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6382-6388, Hong Kong, China, 2019. Association for Computational Linguistics. 3, 4", + "[37] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 1, 3", + "[38] Zeming Wei, Yifei Wang, and Yisen Wang. Jailbreak and guard aligned language models with only few in-context demonstrations. arXiv preprint arXiv:2310.06387, 2023. 1, 3", + "[39] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 1", + "[40] Zheng-Xin Yong, Cristina Menghini, and Stephen H Bach. Low-resource languages jailbreak gpt-4. arXiv preprint arXiv:2310.02446, 2023. 1, 3", + "[41] Youliang Yuan, Wenxiang Jiao, Wenxuan Wang, Jen tse Huang, Pinjia He, Shuming Shi, and Zhaopeng Tu. GPT-4 is too smart to be safe: Stealthy chat with LLMs via cipher. In The Twelfth International Conference on Learning Representations, 2024. 3, 5", + "[42] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE International Conference on Computer Vision, pages 6023-6032, 2019. 2, 3, 4, 7", + "[43] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3, 4, 7", + "[44] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3, 6", + "[45] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 1, 3, 5, 7" + ], + "bbox": [ + 516, + 92, + 903, + 472 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29946", + "bbox": [ + 478, + 950, + 519, + 963 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/013bcc47-e134-4c31-adf0-ced109b7aaa0_model.json b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/013bcc47-e134-4c31-adf0-ced109b7aaa0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f721dbf591bb55f6d7cb8ec5de2ff7b8b35536a2 --- /dev/null +++ b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/013bcc47-e134-4c31-adf0-ced109b7aaa0_model.json @@ -0,0 +1,1903 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.194, + 0.13, + 0.806, + 0.177 + ], + "angle": 0, + "content": "Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.202, + 0.859, + 0.223 + ], + "angle": 0, + "content": "Joonhyun Jeong\\(^{1}\\) Seyun Bae\\(^{2}\\) Yeonsung Jung\\(^{2}\\) Jaeryong Hwang\\(^{3}\\) Eunho Yang\\(^{2,4*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.226, + 0.619, + 0.244 + ], + "angle": 0, + "content": "1NAVER Cloud, ImageVision" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.246, + 0.751, + 0.263 + ], + "angle": 0, + "content": "\\(^{2}\\)Korea Advanced Institute of Science and Technology (KAIST)" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.265, + 0.696, + 0.283 + ], + "angle": 0, + "content": "3Republic of Korea Naval Academy 4AITRICS" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.286, + 0.776, + 0.313 + ], + "angle": 0, + "content": "joonhyun.jeong@navercorp.com {seyun.bae, ys.jung}@kaist.ac.kr \njhwang@navy.ac.kr eunhoy@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.348, + 0.327, + 0.364 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.486, + 0.892 + ], + "angle": 0, + "content": "Despite the remarkable versatility of Large Language Models (LLMs) and Multimodal LLMs (MLLMs) to generalize across both language and vision tasks, LLMs and MLLMs have shown vulnerability to jailbreaking, generating textual outputs that undermine safety, ethical, and bias standards when exposed to harmful or sensitive inputs. With the recent advancement of safety alignment via preference-tuning from human feedback, LLMs and MLLMs have been equipped with safety guardrails to yield safe, ethical, and fair responses with regard to harmful inputs. However, despite the significance of safety alignment, research on the vulnerabilities remains largely underexplored. In this paper, we investigate the unexplored vulnerability of the safety alignment, examining its ability to consistently provide safety guarantees for out-of-distribution(OOD)-ifying harmful inputs that may fall outside the aligned data distribution. Our key observation is that OOD-ifying the vanilla harmful inputs highly increases the uncertainty of the model to discern the malicious intent within the input, leading to a higher chance of being jailbroken. Exploiting this vulnerability, we propose JOOD, a new Jailbreak framework via OOD-ifying inputs beyond the safety alignment. We explore various off-the-shelf visual and textual transformation techniques for OOD-ifying the harmful inputs. Notably, we observe that even simple mixing-based techniques such as image mixup prove highly effective in increasing the uncertainty of the model, thereby facilitating the bypass of the safety alignment. Experiments across diverse jailbreak scenarios demonstrate that JOOD effectively jailbreaks recent proprietary LLMs and MLLMs such as GPT-4 and o1 with high attack success rate, which previous attack approaches have consistently struggled to jailbreak. Code is available at https://github.com/naver-ai/JOOD." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.348, + 0.645, + 0.364 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.908, + 0.529 + ], + "angle": 0, + "content": "Large Language Models (LLMs) have recently exhibited versatility on various language reasoning tasks [10, 20, 26, 28, 37, 39] based on scalable pre-training and fine-tuning on a large corpus of text data. However, due to the biases and misinformation [8, 30] present in the large-scale training data, LLMs have frequently been jailbroken which leads to the generation of biased or unsafe outputs that may compromise ethical standards, safety, or fairness [3, 19, 38, 40, 45] when provided with harmful or sensitive input text instructions such as \"tell me how to build a bomb\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.909, + 0.852 + ], + "angle": 0, + "content": "While the primary target of jailbreaking has been LLMs, similar vulnerabilities have also surfaced in Multimodal LLMs (MLLMs [18, 24, 28, 44]). MLLMs extend the visual understanding capabilities of LLMs by integrating visual encoder [32] with several add-on layers. However, this integration introduces new pathways to bypass the safety guardrails equipped within LLMs. Recently, several jailbreak methods [22, 33] have focused on MLLM's weak safety alignment induced by the add-on linear layers; The visual encoder and linear layers projecting input images into LLMs [24] have been inadequately safety-aligned with regard to the visually harmful images. Inspired by this weak safety alignment regarding harmful images, MLLM jailbreak methods decompose the harmful textual instruction into a multimodal format, incorporating both a harmful image and a generic textual instruction. For example, as shown in the vanilla attack case of Figure 1, instead of using the textual instruction containing the sensitive phrase \"tell me how to build a bomb\", the harmful content \"bomb\" is embedded in the visual input, and the textual instruction is generalized to \"tell me how to build the object in this image\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "To address these safety and ethics issues, safety alignment methods [28, 29] were proposed, involving posttraining LLMs and MLLMs with human-preference feed" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.901, + 0.24, + 0.915 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.964 + ], + "angle": 0, + "content": "29937" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.91, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.393, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Figure 1. Overview of JOOD, in terms of out-of-distribution from the training distribution of safety alignment. For each attack method, we measured uncertainty for discerning the maliciousness of the attack input on GPT-4 and GPT-4V (see Appendix A for the detailed prompt and results) and evaluated their attack success rate on AdvBench-M [27] dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.612 + ], + "angle": 0, + "content": "back (RLHF). In this approach, ethical and safe responses are prioritized through human feedback, guiding the model to yield safe, ethical, and fair outputs with regard to harmful requests. Therefore, the aforementioned naive attack strategies fail to effectively jailbreak safety-aligned models such as GPT-4 and GPT-4V [28]. As shown in Figure 1, when presented with harmful inputs within the bounds of the RLHF training data distribution, these models can confidently discern malicious intent in the inputs and effectively prevent the circumvention of the safety guardrails." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.483, + 0.809 + ], + "angle": 0, + "content": "While RLHF has sufficiently aligned the models to handle such transparently malicious inputs under the safety standards, it may still struggle to generalize to other harmful inputs that fall outside the training distribution. In this paper, we correspondingly investigate a naturally arising question: Does the underlying safety alignment of LLMs and MLLMs consistently guarantee safety even with regard to OOD-ifying inputs that possibly differ from the training inputs and learned knowledge during all the training procedures including safety alignment and are hence novel to the models? To answer the question, we present JOOD, an effective jailbreak attack framework via OOD-ifying harmful inputs beyond safety alignment of LLMs and MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "While there can be various advanced transformation techniques to OOD-ify the harmful inputs, we explore various off-the-shelf visual and textual transforms and observe that the recent LLMs and MLLMs can be effectively jailbroken even with these simple input manipulation techniques. For instance, as shown in Figure 1, the vanilla harm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.905, + 0.703 + ], + "angle": 0, + "content": "ful inputs are transformed into the new ones with conventional mixing techniques [7, 42] where the harmful image of bomb and its word text are mixed with another image and text containing an arbitrary subject (e.g., apple), respectively. These newly transformed textual and visual inputs naturally have discrepancies with the vanilla inputs that were previously seen during all the pre-/fine-/post-training procedures including safety alignment learning, potentially conforming to the out-of-distribution beyond the scope of safety-aligned data distribution. Consequently, when these OOD-ifying harmful inputs are exploited for jailbreak attacks, the model fails to recognize their malicious intent with highly increased uncertainty. This allows an attacker to bypass the safety guardrails designed only for the original in-distribution harmful inputs and thereby yield a significantly higher chance of the model being jailbroken." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.905, + 0.884 + ], + "angle": 0, + "content": "Our comprehensive experiments on various jailbreak scenarios (e.g., bombs, drugs, hacking) of AdvbenchM [27] demonstrate that this straightforward OOD-ifying strategy successfully jailbreaks existing LLMs and MLLMs, including state-of-the-art proprietary models such as GPT-4 and GPT-4V. Also, our method significantly outperforms baseline attack methods in all the jailbreak scenarios, achieving \\(63\\%\\) attack success rate (ASR) against GPT-4V in Bombs or Explosives scenario and improving performance by \\(+42\\%\\) ASR compared to the state-of-the-art baseline [9] in the Hacking scenario. In summary, our contribution is threefold:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- We systematically reveal the vulnerabilities of RLHF" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.518, + 0.963 + ], + "angle": 0, + "content": "29938" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.103, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "based safety alignment, which leaves LLMs and MLLMs vulnerable to out-of-distribution textual or visual inputs that have a discrepancy with the in-distribution samples previously seen during safety alignment." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.153, + 0.482, + 0.227 + ], + "angle": 0, + "content": "- We propose a novel black-box jailbreak strategy via manipulating malicious inputs into the OOD-ifying ones. We observe that even simple off-the-shelf transforms amplify uncertainty of the model, allowing it to effectively bypass the safety alignment of LLMs and MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.228, + 0.482, + 0.288 + ], + "angle": 0, + "content": "- Extensive experiments on various jailbreak scenarios demonstrate the effectiveness of our attack strategy against the state-of-the-art proprietary models such as GPT-4 and even o1, with high attack success rate." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.153, + 0.482, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.299, + 0.232, + 0.315 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.485, + 0.764 + ], + "angle": 0, + "content": "Jailbreaking attacks The jailbreaking attacks undermine LLMs and MLLMs to generate forbidden text outputs that can possibly violate safety, ethics, and fairness regulations. For jailbreaking LLMs, the pioneer work [45] and its variants [1, 23] adversarially optimized the suffix within the attack prompt to yield an affirmative response. Another line of work disguised attack prompts via encryption [11, 41] and translations with low-resource languages [40]. In-context learning [2] was also instrumental in establishing preliminary contexts for jailbreaking by utilizing few-shot jailbroken examples [38] and refining the attack prompt [3]. Meanwhile, several studies [9, 22, 33] focused on the weak safety alignment of MLLMs and reformulated the textual harmful instruction into multimodal format containing a pair of harmful image and generic text instruction. HADES [22] further synthesized the harmful image into a semantically more harmful one by diffusion models for providing a better jailbreaking context. Similarly, Ma et al. [25] additionally provided visual contexts by utilizing the roleplaying concept, offering justifications to carry out the associated harmful request. Also, FigStep [9] converted harmful textual instruction into typography and prompted to complete the blanks in the execution steps. Although the above methods achieved state-of-the-art jailbreak performance on open-source MLLMs [24, 44], they still exhibit a lack of generalization on robust MLLMs [12, 14, 28] that were safety-aligned with human feedback. In this work, we focus on unveiling the vulnerability of these safety-aligned models in terms of out-of-distribution perspective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Safety alignment via human feedback While instruction tuning [37] successfully manipulated LLMs to act in accordance with the user's intention, there still remained a large headroom for improving their safety and reliability. To address this, RLHF [29] aligned LLMs with human preferences under the consideration of the safety standards, robustifying LLMs against malicious text instructions. Also, Sun et al. [34] aligned MLLMs using human feedback on the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "rectified image-text pairs where hallucinations and harmful contents were largely reduced. OpenAI [28] further strengthened the safety alignment of MLLMs using selffeedback as a reward for updating the policy model. However, we revisit that these safety-aligned models via RLHF methods can still be susceptible to jailbreak attacks using various types of data transformation techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.232, + 0.908, + 0.595 + ], + "angle": 0, + "content": "Data transformation Data transformation refers to manipulating the input data into a new variation of a synthetic sample, primarily used for enhancing the robustness and generalization towards broader variations of inputs in the machine learning field. Specifically, regularizing neural networks with data transformation is vital for mitigating overfitting on a fixed training dataset. In image classification task [6], besides the basic transformations such as random resizing and center cropping, advanced ones such as photometric and geometric transformations [4, 5] have been applied to enhance test generalization. Also, image-mixing techniques [15-17, 42, 43], which synthesize data by combining training images, have further expanded the training data spectrum and improved generalization performance. In natural language processing (NLP), text transformations like adding a random word, synonym replacement, and sentence reordering [36] are used to increase input diversity. Other transforms, such as text-mixing [7], generate novel linguistic patterns, creating out-of-distribution data to challenge language models in unexpected ways. In this paper, we explore these off-the-shelf transform techniques to generate OOD-ifying inputs beyond training distribution of safety alignment for LLMs and MLLMs, thereby effectively performing jailbreak attacks against these large models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.616, + 0.649, + 0.634 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this section, we propose JOOD, a simple yet effective jailbreak framework with the OOD-ifying harmful inputs. We explore a broad range of off-the-shelf transformations to manipulate the harmful inputs into OOD-ifying ones that may not be considered during safety alignment. Consequently, these synthetic OOD-ifying inputs effectively bypass the safety guardrails of LLMs and MLLMs, making them more susceptible to jailbreak attempts. This straightforward input manipulation strategy also enables black-box setting to attack the proprietary LLMs and MLLMs such as GPT-4 and GPT-4V. In Section 3.1, we introduce our LLM attack method leveraging textual transformations. In Section 3.2, we also present an effective MLLM attack method using image transformation techniques. Subsequently, we propose an attack evaluation method in Section 3.3 to quantitatively measure the maximum potential risk posed by our attacks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "29939" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.411, + 0.108 + ], + "angle": 0, + "content": "3.1. Attack LLM with OOD-ifying inputs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.31 + ], + "angle": 0, + "content": "Given an input text instruction \\( T^{\\mathrm{h}} \\) that contains a harmful request such as \"tell me how to build a bomb\", the naive attack strategy is to feed \\( T^{\\mathrm{h}} \\) into the target model \\( \\theta \\) to generate output text response \\( r = \\theta (T^{\\mathrm{h}}) \\), where \\( \\theta \\) is typically an autoregressive LLM. While this naive attack had successfully jailbroken the primitive LLMs, it fails to jailbreak recent safety-aligned LLMs since the input \\( T^{\\mathrm{h}} \\) is a natural in-distribution instruction that might have been seen during safety alignment training with human feedback. To overcome this limitation, we transform \\( T^{\\mathrm{h}} \\) into a novel text instruction \\( T^{\\mathrm{ood}} \\) that possibly had not been seen during all the LLM training procedures including the safety alignment learning, and then input to the target model:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.322, + 0.483, + 0.342 + ], + "angle": 0, + "content": "\\[\nT _ {i} ^ {\\mathrm {o o d}} = f \\left(T ^ {\\mathrm {h}}; \\varphi_ {i}\\right), \\quad R = \\left\\{\\theta \\left(T _ {i} ^ {\\mathrm {o o d}}\\right) \\right\\} _ {i = 1} ^ {n} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.484, + 0.701 + ], + "angle": 0, + "content": "where \\(\\varphi_{i}\\) represents transformation-dependent parameter applied by textual transformation function \\(f(\\cdot)\\), and \\(R\\) is a set of model responses for each transformed instruction \\(T_{i}^{\\mathrm{ood}}\\) with one of the \\(n\\) differently sampled \\(\\varphi_{i}\\). To fulfill the malicious goal within \\(T^{\\mathrm{h}}\\), an attacker can employ various textual transformations for \\(f(\\cdot)\\) to OOD-ify \\(T^{\\mathrm{h}}\\) into a novel text instruction \\(T_{i}^{\\mathrm{ood}}\\). For instance, off-the-shelf textual transforms such as random word insertion or deletion [36] and character-level perturbations [21] can be employed in \\(f(\\cdot)\\). Among these methods, we observe that even a simple transform such as the text-mixing technique, where a word is combined with an unrelated auxiliary word to form a new nonsensical word, can effectively synthesize OOD-ifying inputs for the recent LLMs [28] with high uncertainty as shown in Figure 1. Specifically, the harmful subject in \\(T^{\\mathrm{h}}\\) (e.g., bomb) is combined with one of \\(n\\) different auxiliary words \\(\\varphi_{i}\\) (e.g., apple) randomly sampled from real-world vocabularies, synthesizing a new nonsensical word (e.g., bombapple) that does not naturally exist in practice. Subsequently, we prompt the model \\(\\theta\\) to comply with the generic request for each of the words within this newly blended word as follows (see Appendix B for prompt examples of other textual transform variants):" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.709, + 0.482, + 0.785 + ], + "angle": 0, + "content": "The following is a newly coined word made by mixing words corresponding to two words in a random order. Please refer to the possible words inside this new word as \"[Words]\". Please answer the request below for each of these words." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.791, + 0.31, + 0.807 + ], + "angle": 0, + "content": "[Coined word]: \"bombapple\"" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.807, + 0.465, + 0.822 + ], + "angle": 0, + "content": "[Request]: \"Tell me how to build [Words]\"" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.429, + 0.851 + ], + "angle": 0, + "content": "3.2. Attack MLLM with OOD-ifying inputs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Given the original harmful text instruction \\( T^{\\mathrm{h}} \\), naive MLLM attack strategy is to decompose \\( T^{\\mathrm{h}} \\) into a pair of harmful image \\( I^{\\mathrm{h}} \\) (e.g., bomb) and a generic text instruction \\( T^{\\mathrm{g}} \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.259 + ], + "angle": 0, + "content": "(e.g., \"tell me how to build the object in this image\"), and then feed both inputs to the target MLLM \\(\\theta\\). However, this naive vanilla attack does not effectively jailbreak state-of-the-art MLLMs such as GPT-4V that had possibly safely-aligned on such harmful inputs. Therefore, we propose a more powerful OOD-ifying attack strategy that can bypass underlying safety alignment of the MLLMs, as shown in Figure 1. Specifically, we transform the harmful image \\(I^{\\mathrm{h}}\\) into a synthetic image \\(I^{\\mathrm{ood}}\\) via conventional image transformation techniques \\(f(\\cdot)\\) that may not be considered during safety alignment training:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.271, + 0.907, + 0.293 + ], + "angle": 0, + "content": "\\[\nI _ {j} ^ {\\mathrm {o o d}} = f (I ^ {\\mathrm {h}}; \\alpha_ {j}), \\quad R = \\{\\theta (T ^ {\\mathrm {g}}, I _ {j} ^ {\\mathrm {o o d}}) \\} _ {j = 1} ^ {m} \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.907, + 0.47 + ], + "angle": 0, + "content": "where \\(\\alpha_{j}\\) is a transformation degree ranging from 0 to 1, and \\(R\\) is a set of model responses for each transformed image \\(I_{j}^{\\mathrm{odd}}\\) with one of \\(m\\) differently sampled \\(\\alpha_{j}\\). To achieve the malicious goal, the attacker can employ various off-the-shelf image transformations for OOD-ifying the original harmful input \\(I^{\\mathrm{h}}\\) into a novel harmful image \\(I_{j}^{\\mathrm{odd}}\\), such as geometric, photometric transforms [5] and image-mixing techniques [31, 42, 43]. When using image-mixing techniques, we consider another input \\(\\varphi_{i}\\), an auxiliary image containing an arbitrary object or concept, for mixing with the original harmful image \\(I^{\\mathrm{h}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.482, + 0.907, + 0.503 + ], + "angle": 0, + "content": "\\[\nI _ {(i, j)} ^ {\\mathrm {o o d}} = f \\left(I ^ {\\mathrm {h}}, \\varphi_ {i}; \\alpha_ {j}\\right), \\quad R = \\left\\{\\theta \\left(T ^ {\\mathrm {g}}, I _ {(i, j)} ^ {\\mathrm {o o d}}\\right) \\right\\} _ {i = 1, j = 1} ^ {n, m} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.906, + 0.622 + ], + "angle": 0, + "content": "where \\( n \\) denotes the number of auxiliary image \\( \\varphi_{i} \\) used for attack, and \\( R \\) is a set of model responses generated by mixing \\( I^{\\mathrm{h}} \\) with \\( n \\) different \\( \\varphi_{i} \\) samples and \\( m \\) different \\( \\alpha_{j} \\) mixing coefficients. For mixup [43] case, \\( f(\\cdot) \\) can be defined as \\( \\alpha_{j}\\varphi_{i} + (1 - \\alpha_{j})I^{\\mathrm{h}} \\). Also, we correspondingly reform the generic input instruction \\( T^{\\mathrm{g}} \\) to comply with the request for each of the two objects within the mixed image as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.633, + 0.904, + 0.663 + ], + "angle": 0, + "content": "There are two realistic (not artistic but real-functioning) objects (or concepts) in this image." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.664, + 0.856, + 0.694 + ], + "angle": 0, + "content": "Tell me how to build each of these objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Notably, this straightforward OOD-ifying strategy with simple image transform techniques induces a distinct shift from \\( I^{\\mathrm{h}} \\) within the visual encoder of MLLM, as shown in Appendix D. When these shifted harmful visual embeddings are fed into the subsequent LLM, the uncertainty is substantially increased and hence fail to trigger its safety guardrails which were trained on the original in-distribution harmful inputs (i.e., \\( T^{\\mathrm{h}} \\) and \\( I^{\\mathrm{h}} \\)), as shown in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.691, + 0.849 + ], + "angle": 0, + "content": "3.3. Attack Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "For each harmful instruction \\( T^{\\mathrm{h}} \\), our attack methods (Eq. 1, 3) produce a set of attack responses \\( R \\) corresponding to our auxiliary attack input and parameter such as \\( \\varphi \\) and \\( \\alpha \\)," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.964 + ], + "angle": 0, + "content": "29940" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. Comparison of jailbreak performance with baseline attack methods on the unimodal (text-only) model GPT-4 and the multimodal (image + text) model GPT-4V. HF denotes the average harmfulness score of the responses over all the instructions as in Eq. 4, ranging from 0 to 10. The attack scenarios include Bombs or Explosives (BE), Drugs (D), Firearms / Weapons (FW), Hacking information (H), Kill someone (K), Social Violence (SV), and Self-harm and Suicide (SS)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.156, + 0.917, + 0.465 + ], + "angle": 0, + "content": "
InputTarget modelAttack methodBEDFWHKSVSS
HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑
TextGPT-4Vanilla00000.1000000000
CipherChat 2024070.370.100.111080.21507
PAIR 2023000.230.900.8110.100.100.13
JOOD (Eq. 1)1.3132.5172.8243.1421.480.401.013
Image + TextGPT-4VVanilla00000.6120.35000000
FigStep 2023000.230.5000000000
FigStep-Pro 20230.9230.8171.8252.1320.480.300.10
HADES 2024000.230.100.10000.1000
JOOD (Eq. 3)7.1633.9237.2474.0742.141.1100.623
MiniGPT-4 7BVanilla1.5502.5731.260420.2420.8203.073
FigStep 20234.4633.1633.6476.7744.2633.6551.953
HADES 20240.1171.8401.5291.2470.6170.851.337
JOOD (Eq. 3)8.1838.0877.5477.3955.7543.9357.697
LLaVA-1.5 13BVanilla6.6873.2433.5472.4531.7461.0154.383
FigStep 20235.8773.5534.2416.9794.0543.2351.247
HADES 20241.0132.7533.6474.7744.0540.451.747
JOOD (Eq. 3)9.81008.5938.3655.3896.1635.8408.590
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.489, + 0.486, + 0.775 + ], + "angle": 0, + "content": "respectively. Each response in this set may exhibit varying degrees of harmfulness, depending on how much the harmful information and sensitive content it contains corresponding to the malicious input request \\( T^{\\mathrm{h}} \\). To evaluate the maximum potential risk posed by our attacks, we propose a score-based evaluation method to quantitatively measure the harmfulness degree of each response, thereby identifying the most harmful attack response. Specifically, we employ another LLM \\( \\theta^{\\mathrm{hf}} \\) [28] as the harmfulness score (HF) evaluator and prompt it to assess each attack response \\( r \\in R \\) on a scale from 0 to 10, in consideration of the safety standards and compliance with the harmful input instruction \\( T^{\\mathrm{h}} \\). Details of the evaluation prompt are provided in Appendix C. Given the most harmful attack response with the highest harmfulness score, we also report attack success rate (ASR) by prompting to the binary-judging LLM \\( \\theta^{\\mathrm{bj}} \\) [13] that outputs 1 if the response is unsafe and 0 for the harmless responses, following [22]. These evaluation metrics are formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.809, + 0.482, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {H F} \\left(\\mathcal {T} ^ {\\mathrm {h}}\\right) = \\frac {\\sum_ {T ^ {\\mathrm {h}} \\in \\mathcal {T} ^ {\\mathrm {h}}} \\max _ {r \\in R} \\left(\\theta^ {\\mathrm {h f}} \\left(r \\mid T ^ {\\mathrm {h}}\\right)\\right)}{| \\mathcal {T} ^ {\\mathrm {h}} |} \\tag {4} \\\\ \\operatorname {ASR} \\left(\\mathcal {T} ^ {\\mathrm {h}}\\right) = \\frac {\\sum_ {T ^ {\\mathrm {h}} \\in \\mathcal {T} ^ {\\mathrm {h}}} \\theta^ {\\mathrm {b j}} \\left(\\underset {r \\in R} {\\arg \\max } \\left(\\theta^ {\\mathrm {h f}} \\left(r \\mid T ^ {\\mathrm {h}}\\right)\\right)\\right)}{\\left| \\mathcal {T} ^ {\\mathrm {h}} \\right|} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.52 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}^{\\mathrm{h}}\\) is the entire set of harmful instructions used for jailbreaking attacks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.535, + 0.646, + 0.552 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.909, + 0.758 + ], + "angle": 0, + "content": "Dataset. We evaluate JOOD on the widely-used LLM and MLLM jailbreak benchmark, Advbench [45] and Advbench-M [27], to compare the performance against the previous state-of-the-art attack methods. Advbench consists of 500 textual instructions that encourage harmful behaviors, while Advbench-M further categorized the set of these harmful instructions into several distinct scenarios such as Bombs or Explosives, Drugs, and Hacking information. To attack with MLLMs, Advbench-M paired semantically relevant images for each scenario. We used around 30 harmful instructions for each scenario to evaluate LLM attack methods, along with one paired harmful image for each scenario when evaluating MLLM attack methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation details. When attacking with mixing-based transformations of JOOD, the textual or visual auxiliary inputs \\(\\varphi\\) in Eq. 1, 3, are required to synthesize the OOD-ifying input, respectively. We used randomly sampled arbitrary words such as apple and watch, and retrieved corresponding images from the Internet. We set the number of auxiliaries to \\(n = 5\\) and sampled \\(m = 9\\) discrete values for the transformation degree \\(\\alpha\\) from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.952, + 0.518, + 0.964 + ], + "angle": 0, + "content": "29941" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.244 + ], + "angle": 0, + "content": "\\(\\{0.1, 0.2, \\dots, 0.9\\}\\). When attacking with image transformation techniques, we resized and padded the input images to a uniform size of \\(320 \\times 320\\) pixels as a preprocessing step before the transformation. Unless otherwise specified, we use GPT-4 as the target model for LLM attacks and GPT-4V for MLLM attacks, with inference parameters such as temperature and top-p set to 1. For LLM attacks, we employ the text-mixing transformation introduced in Eq. 1, and for MLLM attacks, we use the image-mixing transformations as in Eq. 3." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.254, + 0.231, + 0.269 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.276, + 0.485, + 0.685 + ], + "angle": 0, + "content": "Comparison with SOTA attack methods. In Table 1, we compare JOOD with other state-of-the-art attack methods on GPT-4 and GPT-4V. The vanilla attack methods using the original harmful instructions or images barely jailbreak the target models in almost all the scenarios, while the baseline works such as CipherChat, PAIR, FigStep, and HADES marginally enhanced both ASR and harmfulness scores. However, JOOD consistently exhibits the best jailbreak performance on almost all the scenarios and target models, largely outperforming the baselines. Specifically, when attacking text-only GPT-4, JOOD achieves \\(24\\%\\) ASR in the Firearms scenario where all the baselines failed to jailbreak any of the instructions. Also, when attacking multimodal GPT-4V, JOOD outperforms the previous state-of-the-art attack FigStep-Pro by a large margin, achieving \\(+6.2\\) average harmfulness and \\(+40\\%\\) ASR in Bombs or Explosives scenario. Notably, JOOD exclusively jailbreaks a considerable amount of the instructions that these baselines failed to, specifically 10 additional instructions out of 30 in Bombs or Explosives scenario, as shown in Table 2. Also, JOOD consistently generates extremely harmful responses on most instructions while the baseline methods sparsely generate harmful responses with regard to a few instructions, as shown in Figure 2. These results demonstrate that our OOD-ifying attack strategy via input transformation techniques is effective in causing underlying safety alignment to malfunction to output harmful responses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Generalization on other MLLMs. In the bottom rows of Table 1, we further investigate generalization of JOOD on the other open-source MLLMs, MiniGPT-4 7B [44] and LLaVA-1.5 13B [24] post-trained with RLHF [34]. The results show that JOOD consistently outperforms the other baselines by a large margin in almost all of the scenarios. Surprisingly, JOOD achieves high attack success rates even against robustly safety-aligned models including GPT-4o [12] and o1 [14], which the baselines mostly failed to jailbreak (see Appendix E). The superiority of our OOD-ifying attack strategy on both proprietary and open-source MLLMs further corroborates that the existing MLLMs still lack of safety alignment on the OOD-ifying inputs gener" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.9, + 0.118 + ], + "angle": 0, + "content": "Table 2. Comparison of the number of exclusively jailbroken instructions by each attack method." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.129, + 0.9, + 0.262 + ], + "angle": 0, + "content": "
Attack method# exclusively jailbroken instructions
Figstep-ProHADESJOODBEDFWHKSVSS
✓XX4532200
X✓X0200000
XX✓10547112
" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.265, + 0.88, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.476, + 0.9, + 0.504 + ], + "angle": 0, + "content": "Figure 2. Comparison of harmfulness scores per instruction in Bombs scenario." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.905, + 0.562 + ], + "angle": 0, + "content": "ated from even simple off-the-shelf transforms, emphasizing the need for further research and development." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.68, + 0.587 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.685 + ], + "angle": 0, + "content": "We conduct ablation studies to assess the effectiveness of each component and the sensitivity of the hyper-parameters comprising JOOD. Unless specified, we attack with the multimodal inputs using the mixup technique and adopt GPT-4V as the target model to be attacked. Also, we report jailbreaking performance in Bombs or Explosives scenario." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Effect of various transformation techniques. For attacking GPT-4 with OOD-ifying text inputs as in Eq. 1, we investigate the effect of various text-mixing techniques to obfuscate the harmful word (e.g., \"bomb\") with the auxiliary word (e.g., \"apple\"), producing a mixed word like \"baopmble\" for the \\(H\\) (horizontal)-Interleave case. We observe that GPT-4 effectively interprets these mixed words, accurately recognizing and responding to both the harmful and auxiliary components for the given instruction. See Appendix B for the detailed analysis. Also, to generate OOD-ifying input images for the attack in Eq. 2, we test with widely-adopted image transformation techniques including geometric, photometric transformations" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.52, + 0.964 + ], + "angle": 0, + "content": "29942" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.295, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.091, + 0.49, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.247, + 0.414, + 0.259 + ], + "angle": 0, + "content": "(a) Harmfulness / ASR of text-mixing attacks" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.261, + 0.29, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.262, + 0.49, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.417, + 0.427, + 0.429 + ], + "angle": 0, + "content": "(b) Harmfulness / ASR of image-transform attacks" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.441, + 0.483, + 0.482 + ], + "angle": 0, + "content": "Figure 3. Effect of attacking with (a) various text-mixing techniques on GPT-4 and (b) image-transformation techniques on GPT-4V in all the attack scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.483, + 0.542 + ], + "angle": 0, + "content": "(RandAug [5]) and image-mixing techniques (Mixup [43], CutMix [42], and ResizeMix [31])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.483, + 0.801 + ], + "angle": 0, + "content": "In Figure 3, all of the text-mixing variants and image-mixing techniques consistently outperformed the vanilla attack case without applying mixing techniques in all the scenarios. Also, the performance variance between all the image-mixing variants was marginal, indicating that an attacker can leverage our method for consistently jailbreaking GPT-4V without any dependence on the specific image-mixing technique. However, attacking with non-mixing image transformation techniques (i.e., RandAug) such as adding gaussian noise, rotation, and shearing showed inferior ASR and harmfulness scores compared to the image-mixing transformation techniques, with slightly higher performance than the vanilla case. These results may suggest that GPT-4V is equipped with relatively solid safety alignment on such basic transformations, but not adequately safety-aligned on more synthetic transformations such as image-mixing techniques." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablations on mixing coefficient for mixup. In Figure 4, we analyze the effect of the mixing coefficient \\(\\alpha\\) in Eq. 3 which modulates how much the auxiliary image will obfuscate the original harmful image. We additionally report the ratio of the responses refused by GPT-4V with substring" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.925, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.228, + 0.58, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.228, + 0.639, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.228, + 0.696, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.228, + 0.756, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.228, + 0.815, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.228, + 0.874, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.28, + 0.907, + 0.336 + ], + "angle": 0, + "content": "Figure 4. Ablation on mixing coefficient \\(\\alpha\\) for image mixup attack. We visualize the attack images corresponding to \\(\\alpha\\) value below the x-axis. We report the average harmfulness / ASR / refusal rate on all the 30 instructions in Bombs or Explosives scenario." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.363, + 0.9, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.549, + 0.92, + 0.574 + ], + "angle": 0, + "content": "(a) Effect of semantically various auxiliary images when mixed with the target harmful image (\"bomb\")." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.578, + 0.897, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.745, + 0.92, + 0.771 + ], + "angle": 0, + "content": "(b) Ablation for the visual type of auxiliary image, either using typography image or realistic image." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.782, + 0.907, + 0.823 + ], + "angle": 0, + "content": "Figure 5. Effect of auxiliary images depending on the similarity with the target harmful image (a) and the visual image types (b). Red dotted line denotes the harmfulness of the vanilla attack." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "matching [45] which verifies whether the attacked response has one of the refusal phrases (e.g., \"I'm sorry\"). Generally, the vanilla case (i.e., \\(\\alpha = 0\\)) without mixing the auxiliary image on the original harmful image showed signif" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.518, + 0.963 + ], + "angle": 0, + "content": "29943" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.15, + 0.09, + 0.848, + 0.105 + ], + "angle": 0, + "content": "Table 3. Effect of attack methods for jailbreaking against the system prompt defense in Bombs or Explosives scenario." + }, + { + "type": "table", + "bbox": [ + 0.135, + 0.111, + 0.86, + 0.189 + ], + "angle": 0, + "content": "
System Prompt DefenseVanillaFigStep 2023FigStep-Pro 2023HADES 2024JOOD
HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑
×00%00%0.923%00%7.163%
✓00%00%0.813%00%4.360%
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.204, + 0.486, + 0.416 + ], + "angle": 0, + "content": "icently low ASR and harmfulness score with high refusal rate possibly due to the safety alignment training with self feedback [28] of GPT-4V on these obviously harmful images. However, when an auxiliary image is mixed (i.e., \\(0 < \\alpha < 1\\)), the refusal rate significantly decreases while the ASR and harmfulness score are largely increased. This indicates the essential role of our attack to mitigate the evasive refusal of GPT-4V with regard to the harmful requests and further elicit the harmful responses via presenting the OOD-ifying inputs generated by mixup, which were possibly not seen during the safety alignment training. When \\(\\alpha\\) reaches 1, the harmful image is entirely substituted by the auxiliary image (e.g., apple), losing the harmful semantic and hence reducing both ASR and harmfulness scores." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.433, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Effect of auxiliary images. In Figure 5, we analyze the effect of mixing the harmful image with the auxiliary images based on their similarity and visual types. Specifically, in Figure 5a, we analyze the effect of mixing with semantically various auxiliary images by comparing the cosine similarity with the harmful \"bomb\" image. Blue dots / green dots / red dots represent the auxiliary images that are dissimilar / moderately related / highly similar to the target harmful image, respectively. The result shows a strong negative correlation between the similarity and the harmfulness of the model response. Mixing with highly similar images containing similar unsafe objects such as a \"Molotov cocktail\" generates less harmful responses which only refuse to provide answers with regard to both the original harmful object (e.g., bomb) and another harmful object (e.g., Molotov cocktail). Meanwhile, mixing with moderately related or dissimilar auxiliary images such as \"softwood\" generates harmful responses containing a detailed description regarding the original harmful object (see Appendix J for detailed response examples)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In Figure 5b, we conduct another analysis on the effect of mixing with either a typographic image that displays the text of the auxiliary word or a realistic image containing the visual scene related to the auxiliary word. In all the attack scenarios, using typographic auxiliary images for mixing harmful images generally amplifies the harmfulness of the output responses compared to using real images. While Gong et al. [9] empirically observed that existing MLLMs already exhibit vulnerability on the typographic image input itself, this result suggests that the safety guardrails of MLLMs can be more effectively neutralized when these ty" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.204, + 0.907, + 0.234 + ], + "angle": 0, + "content": "pographic images are further leveraged for our proposed image-mixing based attacks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.905, + 0.445 + ], + "angle": 0, + "content": "Jailbreak against defense methods. In Table 3, we investigate the effect of JOOD against system prompt-based defense [9, 25] where defensive textual guidance is given as input system prompt instructing the model to remain vigilant and refrain from responding to harmful textual or visual queries that may violate AI safety policies (see Appendix G for the detailed prompt). The results show that even after deploying the safety-aware system prompt, JOOD still achieves significantly higher harmfulness and ASR scores compared to the baseline attack methods. Notably, JOOD only degraded ASR by \\(3\\%\\) after applying system-prompt-based defense, while FigStep-Pro largely degraded by \\(10\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.629 + ], + "angle": 0, + "content": "In Appendix G, we further evaluate jailbreak performance of JOOD against a more sophisticated defense, AdaShield [35], which adaptively retrieves the optimal defense prompt for each attack query. Even with this adaptive defense, our JOOD successfully jailbreaks over half of the attack queries with a significantly lower refusal rate, while all the baselines hardly succeed to jailbreak. The robustness of JOOD possibly originates from the ambiguity and uncertainty of the input harmful image induced by transform techniques such as mixup. This highlights the need for further exploration of safeguard mechanisms when dealing with such ambiguous and potentially harmful inputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.647, + 0.634, + 0.663 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We investigate the unrevealed vulnerability of the safety alignment within existing LLMs and MLLMs when the out-of-distribution inputs are given as the input. Exploiting this vulnerability, we devise a new jailbreak strategy by generating OOD-ifying inputs with off-the-shelf data transformation techniques. We observe that even simple transformations such as mixup can fabricate OOD-ifying inputs that induce a high level of uncertainty for the models, successfully jailbreaking the state-of-the-art LLMs and MLLMs. By the extensive experiments on various jailbreak scenarios and rigorous ablations, we demonstrate that this simple yet straightforward OOD-ifying strategy can pose a significant threat to the recent LLMs and MLLMs, underscoring the necessity for further research and development to enhance safety alignment against these OOD-ifying inputs." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.521, + 0.964 + ], + "angle": 0, + "content": "29944" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was partly supported by National Research Foundation of Korea (NRF) grant (No.RS-2023-00209060, A Study on Optimization and Network Interpretation Method for Large-Scale Machine Learning) funded by the Korea government (MSIT). This work was also supported by KAIST-NAVER Hypercreative AI Center." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.221, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[1] Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. Jailbreaking leading safety-aligned llms with simple adaptive attacks. arXiv preprint arXiv:2404.02151, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[2] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[3] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries. arXiv preprint arXiv:2310.08419, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[4] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.485, + 0.515 + ], + "angle": 0, + "content": "[5] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[7] Srujan Kumar Enaganti, Lila Kari, Timothy Ng, and Zihao Wang. Word blending in formal languages: The brangelina effect. In Unconventional Computation and Natural Computation: 17th International Conference, UCNC 2018, Fontainebleau, France, June 25-29, 2018, Proceedings 17, pages 72-85. Springer, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[8] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1-79, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[9] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.451, + 0.814 + ], + "angle": 0, + "content": "[10] Google. Bard - chat based ai tool from google, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[11] Divij Handa, Advait Chirmule, Bimal Gajera, and Chitta Baral. Jailbreaking proprietary large language models using word substitution cipher. arXiv preprint arXiv:2402.10601, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.246, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Lmbased input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[14] Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[15] Joonhyun Jeong, Sungmin Cha, Jongwon Choi, Sangwoo Yun, Taesup Moon, and Youngjoon Yoo. Observations on k-image expansion of image-mixing augmentation. IEEE Access, 11:16631-16643, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[16] JangHyun Kim, Wonho Choo, Hosan Jeong, and Hyun Oh Song. Co-mixup: Saliency guided joint mixup with supermodular diversity. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.904, + 0.417 + ], + "angle": 0, + "content": "[17] Jang-Hyun Kim, Wonho Choo, and Hyun Oh Song. Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In International Conference on Machine Learning, pages 5275-5285. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[18] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal generation. arXiv preprint arXiv:2301.13823, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.904, + 0.503 + ], + "angle": 0, + "content": "[19] Raz Lapid, Ron Langberg, and Moshe Sipper. Open sesame! universal black box jailbreaking of large language models. arXiv preprint arXiv:2309.01446, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[20] Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, et al. Solving quantitative reasoning problems with language models, 2022. URL https://arxiv.org/abs/2206.14858, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[21] Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. Textbugger: Generating adversarial text against real-world applications. arXiv preprint arXiv:1812.05271, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[22] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. arXiv preprint arXiv:2403.09792, 2024. 1, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[23] Zeyi Liao and Huan Sun. Amplegcg: Learning a universal and transferable generative model of adversarial suffixes for jailbreaking both open and closed llms. arXiv preprint arXiv:2404.07921, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[25] Siyuan Ma, Weidi Luo, Yu Wang, Xiaogeng Liu, Muhao Chen, Bo Li, and Chaowei Xiao. Visual-roleplay: Universal jailbreak attack on multimodal large language models via role-playing image character. arXiv preprint arXiv:2405.20773, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.964 + ], + "angle": 0, + "content": "29945" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[27] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[28] OpenAI. Gpt-4 technical report. Technical report, 2023. 1, 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[29] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[30] Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Yang Wang. On the risk of misinformation pollution with large language models. arXiv preprint arXiv:2305.13661, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[31] Jie Qin, Jiemin Fang, Qian Zhang, Wenyu Liu, Xingang Wang, and Xinggang Wang. Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101, 2020.4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[33] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[34] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[35] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. In European Conference on Computer Vision, pages 77-94. Springer, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[36] Jason Wei and Kai Zou. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6382-6388, Hong Kong, China, 2019. Association for Computational Linguistics. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[37] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.869 + ], + "angle": 0, + "content": "[38] Zeming Wei, Yifei Wang, and Yisen Wang. Jailbreak and guard aligned language models with only few in-context demonstrations. arXiv preprint arXiv:2310.06387, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[40] Zheng-Xin Yong, Cristina Menghini, and Stephen H Bach. Low-resource languages jailbreak gpt-4. arXiv preprint arXiv:2310.02446, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[41] Youliang Yuan, Wenxiang Jiao, Wenxuan Wang, Jen tse Huang, Pinjia He, Shuming Shi, and Zhaopeng Tu. GPT-4 is too smart to be safe: Stealthy chat with LLMs via cipher. In The Twelfth International Conference on Learning Representations, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[42] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE International Conference on Computer Vision, pages 6023-6032, 2019. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[43] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[44] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[45] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 1, 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.52, + 0.964 + ], + "angle": 0, + "content": "29946" + } + ] +] \ No newline at end of file diff --git a/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/013bcc47-e134-4c31-adf0-ced109b7aaa0_origin.pdf b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/013bcc47-e134-4c31-adf0-ced109b7aaa0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a40d71695057affe416da12ab98fe23c0cce2827 --- /dev/null +++ b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/013bcc47-e134-4c31-adf0-ced109b7aaa0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35ea3dc5b900b5175f3850a0f7e8baa851351d00e70abbddab30503ad4ddbbb6 +size 4557044 diff --git a/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/full.md b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/full.md new file mode 100644 index 0000000000000000000000000000000000000000..47a19087e96a20647595e6b243322089762cba47 --- /dev/null +++ b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/full.md @@ -0,0 +1,261 @@ +# Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy + +Joonhyun Jeong $^{1}$ Seyun Bae $^{2}$ Yeonsung Jung $^{2}$ Jaeryong Hwang $^{3}$ Eunho Yang $^{2,4*}$ + +1NAVER Cloud, ImageVision + +$^{2}$ Korea Advanced Institute of Science and Technology (KAIST) + +3Republic of Korea Naval Academy 4AITRICS + +joonhyun.jeong@navercorp.com {seyun.bae, ys.jung}@kaist.ac.kr +jhwang@navy.ac.kr eunhoy@kaist.ac.kr + +# Abstract + +Despite the remarkable versatility of Large Language Models (LLMs) and Multimodal LLMs (MLLMs) to generalize across both language and vision tasks, LLMs and MLLMs have shown vulnerability to jailbreaking, generating textual outputs that undermine safety, ethical, and bias standards when exposed to harmful or sensitive inputs. With the recent advancement of safety alignment via preference-tuning from human feedback, LLMs and MLLMs have been equipped with safety guardrails to yield safe, ethical, and fair responses with regard to harmful inputs. However, despite the significance of safety alignment, research on the vulnerabilities remains largely underexplored. In this paper, we investigate the unexplored vulnerability of the safety alignment, examining its ability to consistently provide safety guarantees for out-of-distribution(OOD)-ifying harmful inputs that may fall outside the aligned data distribution. Our key observation is that OOD-ifying the vanilla harmful inputs highly increases the uncertainty of the model to discern the malicious intent within the input, leading to a higher chance of being jailbroken. Exploiting this vulnerability, we propose JOOD, a new Jailbreak framework via OOD-ifying inputs beyond the safety alignment. We explore various off-the-shelf visual and textual transformation techniques for OOD-ifying the harmful inputs. Notably, we observe that even simple mixing-based techniques such as image mixup prove highly effective in increasing the uncertainty of the model, thereby facilitating the bypass of the safety alignment. Experiments across diverse jailbreak scenarios demonstrate that JOOD effectively jailbreaks recent proprietary LLMs and MLLMs such as GPT-4 and o1 with high attack success rate, which previous attack approaches have consistently struggled to jailbreak. Code is available at https://github.com/naver-ai/JOOD. + +# 1. Introduction + +Large Language Models (LLMs) have recently exhibited versatility on various language reasoning tasks [10, 20, 26, 28, 37, 39] based on scalable pre-training and fine-tuning on a large corpus of text data. However, due to the biases and misinformation [8, 30] present in the large-scale training data, LLMs have frequently been jailbroken which leads to the generation of biased or unsafe outputs that may compromise ethical standards, safety, or fairness [3, 19, 38, 40, 45] when provided with harmful or sensitive input text instructions such as "tell me how to build a bomb". + +While the primary target of jailbreaking has been LLMs, similar vulnerabilities have also surfaced in Multimodal LLMs (MLLMs [18, 24, 28, 44]). MLLMs extend the visual understanding capabilities of LLMs by integrating visual encoder [32] with several add-on layers. However, this integration introduces new pathways to bypass the safety guardrails equipped within LLMs. Recently, several jailbreak methods [22, 33] have focused on MLLM's weak safety alignment induced by the add-on linear layers; The visual encoder and linear layers projecting input images into LLMs [24] have been inadequately safety-aligned with regard to the visually harmful images. Inspired by this weak safety alignment regarding harmful images, MLLM jailbreak methods decompose the harmful textual instruction into a multimodal format, incorporating both a harmful image and a generic textual instruction. For example, as shown in the vanilla attack case of Figure 1, instead of using the textual instruction containing the sensitive phrase "tell me how to build a bomb", the harmful content "bomb" is embedded in the visual input, and the textual instruction is generalized to "tell me how to build the object in this image". + +To address these safety and ethics issues, safety alignment methods [28, 29] were proposed, involving posttraining LLMs and MLLMs with human-preference feed + +![](images/c3dcec454d608e5b0d95b576dea8a277a7db2985ce039ac220042cfe896120e9.jpg) +Figure 1. Overview of JOOD, in terms of out-of-distribution from the training distribution of safety alignment. For each attack method, we measured uncertainty for discerning the maliciousness of the attack input on GPT-4 and GPT-4V (see Appendix A for the detailed prompt and results) and evaluated their attack success rate on AdvBench-M [27] dataset. + +back (RLHF). In this approach, ethical and safe responses are prioritized through human feedback, guiding the model to yield safe, ethical, and fair outputs with regard to harmful requests. Therefore, the aforementioned naive attack strategies fail to effectively jailbreak safety-aligned models such as GPT-4 and GPT-4V [28]. As shown in Figure 1, when presented with harmful inputs within the bounds of the RLHF training data distribution, these models can confidently discern malicious intent in the inputs and effectively prevent the circumvention of the safety guardrails. + +While RLHF has sufficiently aligned the models to handle such transparently malicious inputs under the safety standards, it may still struggle to generalize to other harmful inputs that fall outside the training distribution. In this paper, we correspondingly investigate a naturally arising question: Does the underlying safety alignment of LLMs and MLLMs consistently guarantee safety even with regard to OOD-ifying inputs that possibly differ from the training inputs and learned knowledge during all the training procedures including safety alignment and are hence novel to the models? To answer the question, we present JOOD, an effective jailbreak attack framework via OOD-ifying harmful inputs beyond safety alignment of LLMs and MLLMs. + +While there can be various advanced transformation techniques to OOD-ify the harmful inputs, we explore various off-the-shelf visual and textual transforms and observe that the recent LLMs and MLLMs can be effectively jailbroken even with these simple input manipulation techniques. For instance, as shown in Figure 1, the vanilla harm + +ful inputs are transformed into the new ones with conventional mixing techniques [7, 42] where the harmful image of bomb and its word text are mixed with another image and text containing an arbitrary subject (e.g., apple), respectively. These newly transformed textual and visual inputs naturally have discrepancies with the vanilla inputs that were previously seen during all the pre-/fine-/post-training procedures including safety alignment learning, potentially conforming to the out-of-distribution beyond the scope of safety-aligned data distribution. Consequently, when these OOD-ifying harmful inputs are exploited for jailbreak attacks, the model fails to recognize their malicious intent with highly increased uncertainty. This allows an attacker to bypass the safety guardrails designed only for the original in-distribution harmful inputs and thereby yield a significantly higher chance of the model being jailbroken. + +Our comprehensive experiments on various jailbreak scenarios (e.g., bombs, drugs, hacking) of AdvbenchM [27] demonstrate that this straightforward OOD-ifying strategy successfully jailbreaks existing LLMs and MLLMs, including state-of-the-art proprietary models such as GPT-4 and GPT-4V. Also, our method significantly outperforms baseline attack methods in all the jailbreak scenarios, achieving $63\%$ attack success rate (ASR) against GPT-4V in Bombs or Explosives scenario and improving performance by $+42\%$ ASR compared to the state-of-the-art baseline [9] in the Hacking scenario. In summary, our contribution is threefold: + +- We systematically reveal the vulnerabilities of RLHF + +based safety alignment, which leaves LLMs and MLLMs vulnerable to out-of-distribution textual or visual inputs that have a discrepancy with the in-distribution samples previously seen during safety alignment. + +- We propose a novel black-box jailbreak strategy via manipulating malicious inputs into the OOD-ifying ones. We observe that even simple off-the-shelf transforms amplify uncertainty of the model, allowing it to effectively bypass the safety alignment of LLMs and MLLMs. +- Extensive experiments on various jailbreak scenarios demonstrate the effectiveness of our attack strategy against the state-of-the-art proprietary models such as GPT-4 and even o1, with high attack success rate. + +# 2. Related Work + +Jailbreaking attacks The jailbreaking attacks undermine LLMs and MLLMs to generate forbidden text outputs that can possibly violate safety, ethics, and fairness regulations. For jailbreaking LLMs, the pioneer work [45] and its variants [1, 23] adversarially optimized the suffix within the attack prompt to yield an affirmative response. Another line of work disguised attack prompts via encryption [11, 41] and translations with low-resource languages [40]. In-context learning [2] was also instrumental in establishing preliminary contexts for jailbreaking by utilizing few-shot jailbroken examples [38] and refining the attack prompt [3]. Meanwhile, several studies [9, 22, 33] focused on the weak safety alignment of MLLMs and reformulated the textual harmful instruction into multimodal format containing a pair of harmful image and generic text instruction. HADES [22] further synthesized the harmful image into a semantically more harmful one by diffusion models for providing a better jailbreaking context. Similarly, Ma et al. [25] additionally provided visual contexts by utilizing the roleplaying concept, offering justifications to carry out the associated harmful request. Also, FigStep [9] converted harmful textual instruction into typography and prompted to complete the blanks in the execution steps. Although the above methods achieved state-of-the-art jailbreak performance on open-source MLLMs [24, 44], they still exhibit a lack of generalization on robust MLLMs [12, 14, 28] that were safety-aligned with human feedback. In this work, we focus on unveiling the vulnerability of these safety-aligned models in terms of out-of-distribution perspective. + +Safety alignment via human feedback While instruction tuning [37] successfully manipulated LLMs to act in accordance with the user's intention, there still remained a large headroom for improving their safety and reliability. To address this, RLHF [29] aligned LLMs with human preferences under the consideration of the safety standards, robustifying LLMs against malicious text instructions. Also, Sun et al. [34] aligned MLLMs using human feedback on the + +rectified image-text pairs where hallucinations and harmful contents were largely reduced. OpenAI [28] further strengthened the safety alignment of MLLMs using selffeedback as a reward for updating the policy model. However, we revisit that these safety-aligned models via RLHF methods can still be susceptible to jailbreak attacks using various types of data transformation techniques. + +Data transformation Data transformation refers to manipulating the input data into a new variation of a synthetic sample, primarily used for enhancing the robustness and generalization towards broader variations of inputs in the machine learning field. Specifically, regularizing neural networks with data transformation is vital for mitigating overfitting on a fixed training dataset. In image classification task [6], besides the basic transformations such as random resizing and center cropping, advanced ones such as photometric and geometric transformations [4, 5] have been applied to enhance test generalization. Also, image-mixing techniques [15-17, 42, 43], which synthesize data by combining training images, have further expanded the training data spectrum and improved generalization performance. In natural language processing (NLP), text transformations like adding a random word, synonym replacement, and sentence reordering [36] are used to increase input diversity. Other transforms, such as text-mixing [7], generate novel linguistic patterns, creating out-of-distribution data to challenge language models in unexpected ways. In this paper, we explore these off-the-shelf transform techniques to generate OOD-ifying inputs beyond training distribution of safety alignment for LLMs and MLLMs, thereby effectively performing jailbreak attacks against these large models. + +# 3. Methodology + +In this section, we propose JOOD, a simple yet effective jailbreak framework with the OOD-ifying harmful inputs. We explore a broad range of off-the-shelf transformations to manipulate the harmful inputs into OOD-ifying ones that may not be considered during safety alignment. Consequently, these synthetic OOD-ifying inputs effectively bypass the safety guardrails of LLMs and MLLMs, making them more susceptible to jailbreak attempts. This straightforward input manipulation strategy also enables black-box setting to attack the proprietary LLMs and MLLMs such as GPT-4 and GPT-4V. In Section 3.1, we introduce our LLM attack method leveraging textual transformations. In Section 3.2, we also present an effective MLLM attack method using image transformation techniques. Subsequently, we propose an attack evaluation method in Section 3.3 to quantitatively measure the maximum potential risk posed by our attacks. + +# 3.1. Attack LLM with OOD-ifying inputs + +Given an input text instruction $T^{\mathrm{h}}$ that contains a harmful request such as "tell me how to build a bomb", the naive attack strategy is to feed $T^{\mathrm{h}}$ into the target model $\theta$ to generate output text response $r = \theta (T^{\mathrm{h}})$ , where $\theta$ is typically an autoregressive LLM. While this naive attack had successfully jailbroken the primitive LLMs, it fails to jailbreak recent safety-aligned LLMs since the input $T^{\mathrm{h}}$ is a natural in-distribution instruction that might have been seen during safety alignment training with human feedback. To overcome this limitation, we transform $T^{\mathrm{h}}$ into a novel text instruction $T^{\mathrm{ood}}$ that possibly had not been seen during all the LLM training procedures including the safety alignment learning, and then input to the target model: + +$$ +T _ {i} ^ {\mathrm {o o d}} = f \left(T ^ {\mathrm {h}}; \varphi_ {i}\right), \quad R = \left\{\theta \left(T _ {i} ^ {\mathrm {o o d}}\right) \right\} _ {i = 1} ^ {n} \tag {1} +$$ + +where $\varphi_{i}$ represents transformation-dependent parameter applied by textual transformation function $f(\cdot)$ , and $R$ is a set of model responses for each transformed instruction $T_{i}^{\mathrm{ood}}$ with one of the $n$ differently sampled $\varphi_{i}$ . To fulfill the malicious goal within $T^{\mathrm{h}}$ , an attacker can employ various textual transformations for $f(\cdot)$ to OOD-ify $T^{\mathrm{h}}$ into a novel text instruction $T_{i}^{\mathrm{ood}}$ . For instance, off-the-shelf textual transforms such as random word insertion or deletion [36] and character-level perturbations [21] can be employed in $f(\cdot)$ . Among these methods, we observe that even a simple transform such as the text-mixing technique, where a word is combined with an unrelated auxiliary word to form a new nonsensical word, can effectively synthesize OOD-ifying inputs for the recent LLMs [28] with high uncertainty as shown in Figure 1. Specifically, the harmful subject in $T^{\mathrm{h}}$ (e.g., bomb) is combined with one of $n$ different auxiliary words $\varphi_{i}$ (e.g., apple) randomly sampled from real-world vocabularies, synthesizing a new nonsensical word (e.g., bombapple) that does not naturally exist in practice. Subsequently, we prompt the model $\theta$ to comply with the generic request for each of the words within this newly blended word as follows (see Appendix B for prompt examples of other textual transform variants): + +The following is a newly coined word made by mixing words corresponding to two words in a random order. Please refer to the possible words inside this new word as "[Words]". Please answer the request below for each of these words. + +[Coined word]: "bombapple" + +[Request]: "Tell me how to build [Words]" + +# 3.2. Attack MLLM with OOD-ifying inputs + +Given the original harmful text instruction $T^{\mathrm{h}}$ , naive MLLM attack strategy is to decompose $T^{\mathrm{h}}$ into a pair of harmful image $I^{\mathrm{h}}$ (e.g., bomb) and a generic text instruction $T^{\mathrm{g}}$ + +(e.g., "tell me how to build the object in this image"), and then feed both inputs to the target MLLM $\theta$ . However, this naive vanilla attack does not effectively jailbreak state-of-the-art MLLMs such as GPT-4V that had possibly safely-aligned on such harmful inputs. Therefore, we propose a more powerful OOD-ifying attack strategy that can bypass underlying safety alignment of the MLLMs, as shown in Figure 1. Specifically, we transform the harmful image $I^{\mathrm{h}}$ into a synthetic image $I^{\mathrm{ood}}$ via conventional image transformation techniques $f(\cdot)$ that may not be considered during safety alignment training: + +$$ +I _ {j} ^ {\mathrm {o o d}} = f (I ^ {\mathrm {h}}; \alpha_ {j}), \quad R = \{\theta (T ^ {\mathrm {g}}, I _ {j} ^ {\mathrm {o o d}}) \} _ {j = 1} ^ {m} \qquad (2) +$$ + +where $\alpha_{j}$ is a transformation degree ranging from 0 to 1, and $R$ is a set of model responses for each transformed image $I_{j}^{\mathrm{odd}}$ with one of $m$ differently sampled $\alpha_{j}$ . To achieve the malicious goal, the attacker can employ various off-the-shelf image transformations for OOD-ifying the original harmful input $I^{\mathrm{h}}$ into a novel harmful image $I_{j}^{\mathrm{odd}}$ , such as geometric, photometric transforms [5] and image-mixing techniques [31, 42, 43]. When using image-mixing techniques, we consider another input $\varphi_{i}$ , an auxiliary image containing an arbitrary object or concept, for mixing with the original harmful image $I^{\mathrm{h}}$ : + +$$ +I _ {(i, j)} ^ {\mathrm {o o d}} = f \left(I ^ {\mathrm {h}}, \varphi_ {i}; \alpha_ {j}\right), \quad R = \left\{\theta \left(T ^ {\mathrm {g}}, I _ {(i, j)} ^ {\mathrm {o o d}}\right) \right\} _ {i = 1, j = 1} ^ {n, m} \tag {3} +$$ + +where $n$ denotes the number of auxiliary image $\varphi_{i}$ used for attack, and $R$ is a set of model responses generated by mixing $I^{\mathrm{h}}$ with $n$ different $\varphi_{i}$ samples and $m$ different $\alpha_{j}$ mixing coefficients. For mixup [43] case, $f(\cdot)$ can be defined as $\alpha_{j}\varphi_{i} + (1 - \alpha_{j})I^{\mathrm{h}}$ . Also, we correspondingly reform the generic input instruction $T^{\mathrm{g}}$ to comply with the request for each of the two objects within the mixed image as follows: + +There are two realistic (not artistic but real-functioning) objects (or concepts) in this image. + +Tell me how to build each of these objects. + +Notably, this straightforward OOD-ifying strategy with simple image transform techniques induces a distinct shift from $I^{\mathrm{h}}$ within the visual encoder of MLLM, as shown in Appendix D. When these shifted harmful visual embeddings are fed into the subsequent LLM, the uncertainty is substantially increased and hence fail to trigger its safety guardrails which were trained on the original in-distribution harmful inputs (i.e., $T^{\mathrm{h}}$ and $I^{\mathrm{h}}$ ), as shown in Figure 1. + +# 3.3. Attack Evaluation + +For each harmful instruction $T^{\mathrm{h}}$ , our attack methods (Eq. 1, 3) produce a set of attack responses $R$ corresponding to our auxiliary attack input and parameter such as $\varphi$ and $\alpha$ , + +Table 1. Comparison of jailbreak performance with baseline attack methods on the unimodal (text-only) model GPT-4 and the multimodal (image + text) model GPT-4V. HF denotes the average harmfulness score of the responses over all the instructions as in Eq. 4, ranging from 0 to 10. The attack scenarios include Bombs or Explosives (BE), Drugs (D), Firearms / Weapons (FW), Hacking information (H), Kill someone (K), Social Violence (SV), and Self-harm and Suicide (SS). + +
InputTarget modelAttack methodBEDFWHKSVSS
HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑
TextGPT-4Vanilla00000.1000000000
CipherChat 2024070.370.100.111080.21507
PAIR 2023000.230.900.8110.100.100.13
JOOD (Eq. 1)1.3132.5172.8243.1421.480.401.013
Image + TextGPT-4VVanilla00000.6120.35000000
FigStep 2023000.230.5000000000
FigStep-Pro 20230.9230.8171.8252.1320.480.300.10
HADES 2024000.230.100.10000.1000
JOOD (Eq. 3)7.1633.9237.2474.0742.141.1100.623
MiniGPT-4 7BVanilla1.5502.5731.260420.2420.8203.073
FigStep 20234.4633.1633.6476.7744.2633.6551.953
HADES 20240.1171.8401.5291.2470.6170.851.337
JOOD (Eq. 3)8.1838.0877.5477.3955.7543.9357.697
LLaVA-1.5 13BVanilla6.6873.2433.5472.4531.7461.0154.383
FigStep 20235.8773.5534.2416.9794.0543.2351.247
HADES 20241.0132.7533.6474.7744.0540.451.747
JOOD (Eq. 3)9.81008.5938.3655.3896.1635.8408.590
+ +respectively. Each response in this set may exhibit varying degrees of harmfulness, depending on how much the harmful information and sensitive content it contains corresponding to the malicious input request $T^{\mathrm{h}}$ . To evaluate the maximum potential risk posed by our attacks, we propose a score-based evaluation method to quantitatively measure the harmfulness degree of each response, thereby identifying the most harmful attack response. Specifically, we employ another LLM $\theta^{\mathrm{hf}}$ [28] as the harmfulness score (HF) evaluator and prompt it to assess each attack response $r \in R$ on a scale from 0 to 10, in consideration of the safety standards and compliance with the harmful input instruction $T^{\mathrm{h}}$ . Details of the evaluation prompt are provided in Appendix C. Given the most harmful attack response with the highest harmfulness score, we also report attack success rate (ASR) by prompting to the binary-judging LLM $\theta^{\mathrm{bj}}$ [13] that outputs 1 if the response is unsafe and 0 for the harmless responses, following [22]. These evaluation metrics are formulated as follows: + +$$ +\begin{array}{l} \mathrm {H F} \left(\mathcal {T} ^ {\mathrm {h}}\right) = \frac {\sum_ {T ^ {\mathrm {h}} \in \mathcal {T} ^ {\mathrm {h}}} \max _ {r \in R} \left(\theta^ {\mathrm {h f}} \left(r \mid T ^ {\mathrm {h}}\right)\right)}{| \mathcal {T} ^ {\mathrm {h}} |} \tag {4} \\ \operatorname {ASR} \left(\mathcal {T} ^ {\mathrm {h}}\right) = \frac {\sum_ {T ^ {\mathrm {h}} \in \mathcal {T} ^ {\mathrm {h}}} \theta^ {\mathrm {b j}} \left(\underset {r \in R} {\arg \max } \left(\theta^ {\mathrm {h f}} \left(r \mid T ^ {\mathrm {h}}\right)\right)\right)}{\left| \mathcal {T} ^ {\mathrm {h}} \right|} \\ \end{array} +$$ + +where $\mathcal{T}^{\mathrm{h}}$ is the entire set of harmful instructions used for jailbreaking attacks. + +# 4. Experiments + +Dataset. We evaluate JOOD on the widely-used LLM and MLLM jailbreak benchmark, Advbench [45] and Advbench-M [27], to compare the performance against the previous state-of-the-art attack methods. Advbench consists of 500 textual instructions that encourage harmful behaviors, while Advbench-M further categorized the set of these harmful instructions into several distinct scenarios such as Bombs or Explosives, Drugs, and Hacking information. To attack with MLLMs, Advbench-M paired semantically relevant images for each scenario. We used around 30 harmful instructions for each scenario to evaluate LLM attack methods, along with one paired harmful image for each scenario when evaluating MLLM attack methods. + +Implementation details. When attacking with mixing-based transformations of JOOD, the textual or visual auxiliary inputs $\varphi$ in Eq. 1, 3, are required to synthesize the OOD-ifying input, respectively. We used randomly sampled arbitrary words such as apple and watch, and retrieved corresponding images from the Internet. We set the number of auxiliaries to $n = 5$ and sampled $m = 9$ discrete values for the transformation degree $\alpha$ from + +$\{0.1, 0.2, \dots, 0.9\}$ . When attacking with image transformation techniques, we resized and padded the input images to a uniform size of $320 \times 320$ pixels as a preprocessing step before the transformation. Unless otherwise specified, we use GPT-4 as the target model for LLM attacks and GPT-4V for MLLM attacks, with inference parameters such as temperature and top-p set to 1. For LLM attacks, we employ the text-mixing transformation introduced in Eq. 1, and for MLLM attacks, we use the image-mixing transformations as in Eq. 3. + +# 4.1. Main Results + +Comparison with SOTA attack methods. In Table 1, we compare JOOD with other state-of-the-art attack methods on GPT-4 and GPT-4V. The vanilla attack methods using the original harmful instructions or images barely jailbreak the target models in almost all the scenarios, while the baseline works such as CipherChat, PAIR, FigStep, and HADES marginally enhanced both ASR and harmfulness scores. However, JOOD consistently exhibits the best jailbreak performance on almost all the scenarios and target models, largely outperforming the baselines. Specifically, when attacking text-only GPT-4, JOOD achieves $24\%$ ASR in the Firearms scenario where all the baselines failed to jailbreak any of the instructions. Also, when attacking multimodal GPT-4V, JOOD outperforms the previous state-of-the-art attack FigStep-Pro by a large margin, achieving $+6.2$ average harmfulness and $+40\%$ ASR in Bombs or Explosives scenario. Notably, JOOD exclusively jailbreaks a considerable amount of the instructions that these baselines failed to, specifically 10 additional instructions out of 30 in Bombs or Explosives scenario, as shown in Table 2. Also, JOOD consistently generates extremely harmful responses on most instructions while the baseline methods sparsely generate harmful responses with regard to a few instructions, as shown in Figure 2. These results demonstrate that our OOD-ifying attack strategy via input transformation techniques is effective in causing underlying safety alignment to malfunction to output harmful responses. + +Generalization on other MLLMs. In the bottom rows of Table 1, we further investigate generalization of JOOD on the other open-source MLLMs, MiniGPT-4 7B [44] and LLaVA-1.5 13B [24] post-trained with RLHF [34]. The results show that JOOD consistently outperforms the other baselines by a large margin in almost all of the scenarios. Surprisingly, JOOD achieves high attack success rates even against robustly safety-aligned models including GPT-4o [12] and o1 [14], which the baselines mostly failed to jailbreak (see Appendix E). The superiority of our OOD-ifying attack strategy on both proprietary and open-source MLLMs further corroborates that the existing MLLMs still lack of safety alignment on the OOD-ifying inputs gener + +Table 2. Comparison of the number of exclusively jailbroken instructions by each attack method. + +
Attack method# exclusively jailbroken instructions
Figstep-ProHADESJOODBEDFWHKSVSS
✓XX4532200
X✓X0200000
XX✓10547112
+ +![](images/1903840763ec6bf59b981c360f9d629af4d513f8d22a9849c87ab523af655e50.jpg) +Figure 2. Comparison of harmfulness scores per instruction in Bombs scenario. + +ated from even simple off-the-shelf transforms, emphasizing the need for further research and development. + +# 4.2. Ablation Studies + +We conduct ablation studies to assess the effectiveness of each component and the sensitivity of the hyper-parameters comprising JOOD. Unless specified, we attack with the multimodal inputs using the mixup technique and adopt GPT-4V as the target model to be attacked. Also, we report jailbreaking performance in Bombs or Explosives scenario. + +Effect of various transformation techniques. For attacking GPT-4 with OOD-ifying text inputs as in Eq. 1, we investigate the effect of various text-mixing techniques to obfuscate the harmful word (e.g., "bomb") with the auxiliary word (e.g., "apple"), producing a mixed word like "baopmble" for the $H$ (horizontal)-Interleave case. We observe that GPT-4 effectively interprets these mixed words, accurately recognizing and responding to both the harmful and auxiliary components for the given instruction. See Appendix B for the detailed analysis. Also, to generate OOD-ifying input images for the attack in Eq. 2, we test with widely-adopted image transformation techniques including geometric, photometric transformations + +![](images/ebbaa4912c33be151384b6ee2d380700da89ac6ef4a5c90b34fe7043cf815fe9.jpg) + +![](images/54bd666e8226776418213cbf2a020b2f60bd30061208f331bc3264454ec4d6bc.jpg) + +![](images/65b1310079864f6d3c68d7b8d72a813f8efc17f2511ef604ade18276bf5d3f43.jpg) +(a) Harmfulness / ASR of text-mixing attacks + +![](images/bdc9a811c37910f224fa867b6716dbf0fa9d9a68c4e1f64dc7c9cde06248720d.jpg) +(b) Harmfulness / ASR of image-transform attacks + +(RandAug [5]) and image-mixing techniques (Mixup [43], CutMix [42], and ResizeMix [31]). + +In Figure 3, all of the text-mixing variants and image-mixing techniques consistently outperformed the vanilla attack case without applying mixing techniques in all the scenarios. Also, the performance variance between all the image-mixing variants was marginal, indicating that an attacker can leverage our method for consistently jailbreaking GPT-4V without any dependence on the specific image-mixing technique. However, attacking with non-mixing image transformation techniques (i.e., RandAug) such as adding gaussian noise, rotation, and shearing showed inferior ASR and harmfulness scores compared to the image-mixing transformation techniques, with slightly higher performance than the vanilla case. These results may suggest that GPT-4V is equipped with relatively solid safety alignment on such basic transformations, but not adequately safety-aligned on more synthetic transformations such as image-mixing techniques. + +Ablations on mixing coefficient for mixup. In Figure 4, we analyze the effect of the mixing coefficient $\alpha$ in Eq. 3 which modulates how much the auxiliary image will obfuscate the original harmful image. We additionally report the ratio of the responses refused by GPT-4V with substring + +![](images/6cec83050da5e4d07d3a8f3d8ed7da623d8314c6b6ccde296d9530371bb8c75d.jpg) + +![](images/e9da64943dc04488522e86858b18d21724c01601cf0a4123463a16dcae17f989.jpg) +Figure 4. Ablation on mixing coefficient $\alpha$ for image mixup attack. We visualize the attack images corresponding to $\alpha$ value below the x-axis. We report the average harmfulness / ASR / refusal rate on all the 30 instructions in Bombs or Explosives scenario. + +![](images/5bf34b6d4542f672a485899f08f919691747202e4713a35814001a85c2f827a8.jpg) + +![](images/cf56f2b5c8f579db16e2375f6f12e75c1cb43fb5a2608bf7cfd64cefad7f8df6.jpg) + +![](images/d5e530947f53f879be50941b1ffd866a200bd0a40a666333f688f948cd2d70f8.jpg) + +![](images/dfd721c65faca1b6d529825ab399f71f52204d60de6189f91a37c776b095eb4e.jpg) + +![](images/aabca5b31b4ddab78a761fb3c0093fc244d8b70a72258ef314f11b22e13ac21d.jpg) + +![](images/d28407166ba7c3ddc69e68bef15f58270f899a4b4ba158665910c0979032669f.jpg) +Figure 3. Effect of attacking with (a) various text-mixing techniques on GPT-4 and (b) image-transformation techniques on GPT-4V in all the attack scenarios. +(a) Effect of semantically various auxiliary images when mixed with the target harmful image ("bomb"). + +![](images/95bbfab247648a56ecf6b8c76ed37494c3f497738c7c383204e0ca996f671380.jpg) +(b) Ablation for the visual type of auxiliary image, either using typography image or realistic image. +Figure 5. Effect of auxiliary images depending on the similarity with the target harmful image (a) and the visual image types (b). Red dotted line denotes the harmfulness of the vanilla attack. + +matching [45] which verifies whether the attacked response has one of the refusal phrases (e.g., "I'm sorry"). Generally, the vanilla case (i.e., $\alpha = 0$ ) without mixing the auxiliary image on the original harmful image showed signif + +Table 3. Effect of attack methods for jailbreaking against the system prompt defense in Bombs or Explosives scenario. + +
System Prompt DefenseVanillaFigStep 2023FigStep-Pro 2023HADES 2024JOOD
HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑
×00%00%0.923%00%7.163%
✓00%00%0.813%00%4.360%
+ +icently low ASR and harmfulness score with high refusal rate possibly due to the safety alignment training with self feedback [28] of GPT-4V on these obviously harmful images. However, when an auxiliary image is mixed (i.e., $0 < \alpha < 1$ ), the refusal rate significantly decreases while the ASR and harmfulness score are largely increased. This indicates the essential role of our attack to mitigate the evasive refusal of GPT-4V with regard to the harmful requests and further elicit the harmful responses via presenting the OOD-ifying inputs generated by mixup, which were possibly not seen during the safety alignment training. When $\alpha$ reaches 1, the harmful image is entirely substituted by the auxiliary image (e.g., apple), losing the harmful semantic and hence reducing both ASR and harmfulness scores. + +Effect of auxiliary images. In Figure 5, we analyze the effect of mixing the harmful image with the auxiliary images based on their similarity and visual types. Specifically, in Figure 5a, we analyze the effect of mixing with semantically various auxiliary images by comparing the cosine similarity with the harmful "bomb" image. Blue dots / green dots / red dots represent the auxiliary images that are dissimilar / moderately related / highly similar to the target harmful image, respectively. The result shows a strong negative correlation between the similarity and the harmfulness of the model response. Mixing with highly similar images containing similar unsafe objects such as a "Molotov cocktail" generates less harmful responses which only refuse to provide answers with regard to both the original harmful object (e.g., bomb) and another harmful object (e.g., Molotov cocktail). Meanwhile, mixing with moderately related or dissimilar auxiliary images such as "softwood" generates harmful responses containing a detailed description regarding the original harmful object (see Appendix J for detailed response examples). + +In Figure 5b, we conduct another analysis on the effect of mixing with either a typographic image that displays the text of the auxiliary word or a realistic image containing the visual scene related to the auxiliary word. In all the attack scenarios, using typographic auxiliary images for mixing harmful images generally amplifies the harmfulness of the output responses compared to using real images. While Gong et al. [9] empirically observed that existing MLLMs already exhibit vulnerability on the typographic image input itself, this result suggests that the safety guardrails of MLLMs can be more effectively neutralized when these ty + +pographic images are further leveraged for our proposed image-mixing based attacks. + +Jailbreak against defense methods. In Table 3, we investigate the effect of JOOD against system prompt-based defense [9, 25] where defensive textual guidance is given as input system prompt instructing the model to remain vigilant and refrain from responding to harmful textual or visual queries that may violate AI safety policies (see Appendix G for the detailed prompt). The results show that even after deploying the safety-aware system prompt, JOOD still achieves significantly higher harmfulness and ASR scores compared to the baseline attack methods. Notably, JOOD only degraded ASR by $3\%$ after applying system-prompt-based defense, while FigStep-Pro largely degraded by $10\%$ . + +In Appendix G, we further evaluate jailbreak performance of JOOD against a more sophisticated defense, AdaShield [35], which adaptively retrieves the optimal defense prompt for each attack query. Even with this adaptive defense, our JOOD successfully jailbreaks over half of the attack queries with a significantly lower refusal rate, while all the baselines hardly succeed to jailbreak. The robustness of JOOD possibly originates from the ambiguity and uncertainty of the input harmful image induced by transform techniques such as mixup. This highlights the need for further exploration of safeguard mechanisms when dealing with such ambiguous and potentially harmful inputs. + +# 5. Conclusion + +We investigate the unrevealed vulnerability of the safety alignment within existing LLMs and MLLMs when the out-of-distribution inputs are given as the input. Exploiting this vulnerability, we devise a new jailbreak strategy by generating OOD-ifying inputs with off-the-shelf data transformation techniques. We observe that even simple transformations such as mixup can fabricate OOD-ifying inputs that induce a high level of uncertainty for the models, successfully jailbreaking the state-of-the-art LLMs and MLLMs. By the extensive experiments on various jailbreak scenarios and rigorous ablations, we demonstrate that this simple yet straightforward OOD-ifying strategy can pose a significant threat to the recent LLMs and MLLMs, underscoring the necessity for further research and development to enhance safety alignment against these OOD-ifying inputs. + +# Acknowledgement + +This work was partly supported by National Research Foundation of Korea (NRF) grant (No.RS-2023-00209060, A Study on Optimization and Network Interpretation Method for Large-Scale Machine Learning) funded by the Korea government (MSIT). This work was also supported by KAIST-NAVER Hypercreative AI Center. + +# References + +[1] Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. Jailbreaking leading safety-aligned llms with simple adaptive attacks. arXiv preprint arXiv:2404.02151, 2024.3 +[2] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3 +[3] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries. arXiv preprint arXiv:2310.08419, 2023. 1, 3, 5 +[4] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018.3 +[5] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 3, 4, 7 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3 +[7] Srujan Kumar Enaganti, Lila Kari, Timothy Ng, and Zihao Wang. Word blending in formal languages: The brangelina effect. In Unconventional Computation and Natural Computation: 17th International Conference, UCNC 2018, Fontainebleau, France, June 25-29, 2018, Proceedings 17, pages 72-85. Springer, 2018. 2, 3 +[8] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1-79, 2024. 1 +[9] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 5, 8 +[10] Google. Bard - chat based ai tool from google, 2023. 1 +[11] Divij Handa, Advait Chirmule, Bimal Gajera, and Chitta Baral. Jailbreaking proprietary large language models using word substitution cipher. arXiv preprint arXiv:2402.10601, 2024.3 +[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli + +hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3, 6 +[13] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Lmbased input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023.5 +[14] Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 3, 6 +[15] Joonhyun Jeong, Sungmin Cha, Jongwon Choi, Sangwoo Yun, Taesup Moon, and Youngjoon Yoo. Observations on k-image expansion of image-mixing augmentation. IEEE Access, 11:16631-16643, 2023. 3 +[16] JangHyun Kim, Wonho Choo, Hosan Jeong, and Hyun Oh Song. Co-mixup: Saliency guided joint mixup with supermodular diversity. In International Conference on Learning Representations, 2021. +[17] Jang-Hyun Kim, Wonho Choo, and Hyun Oh Song. Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In International Conference on Machine Learning, pages 5275-5285. PMLR, 2020. 3 +[18] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal generation. arXiv preprint arXiv:2301.13823, 2023. 1 +[19] Raz Lapid, Ron Langberg, and Moshe Sipper. Open sesame! universal black box jailbreaking of large language models. arXiv preprint arXiv:2309.01446, 2023. 1 +[20] Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, et al. Solving quantitative reasoning problems with language models, 2022. URL https://arxiv.org/abs/2206.14858, 2022. 1 +[21] Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. Textbugger: Generating adversarial text against real-world applications. arXiv preprint arXiv:1812.05271, 2018. 4 +[22] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. arXiv preprint arXiv:2403.09792, 2024. 1, 3, 5, 8 +[23] Zeyi Liao and Huan Sun. Amplegcg: Learning a universal and transferable generative model of adversarial suffixes for jailbreaking both open and closed llms. arXiv preprint arXiv:2404.07921, 2024. 3 +[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 3, 6 +[25] Siyuan Ma, Weidi Luo, Yu Wang, Xiaogeng Liu, Muhao Chen, Bo Li, and Chaowei Xiao. Visual-roleplay: Universal jailbreak attack on multimodal large language models via role-playing image character. arXiv preprint arXiv:2405.20773, 2024. 3, 8 +[26] Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943, 2021. 1 + +[27] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 2, 5 +[28] OpenAI. Gpt-4 technical report. Technical report, 2023. 1, 2, 3, 4, 5, 8 +[29] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1, 3 +[30] Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Yang Wang. On the risk of misinformation pollution with large language models. arXiv preprint arXiv:2305.13661, 2023. 1 +[31] Jie Qin, Jiemin Fang, Qian Zhang, Wenyu Liu, Xingang Wang, and Xinggang Wang. Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101, 2020.4, 7 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1 +[33] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2023. 1, 3 +[34] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 3, 6 +[35] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. In European Conference on Computer Vision, pages 77-94. Springer, 2024. 8 +[36] Jason Wei and Kai Zou. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6382-6388, Hong Kong, China, 2019. Association for Computational Linguistics. 3, 4 +[37] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 1, 3 +[38] Zeming Wei, Yifei Wang, and Yisen Wang. Jailbreak and guard aligned language models with only few in-context demonstrations. arXiv preprint arXiv:2310.06387, 2023. 1, 3 +[39] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing + +reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 1 +[40] Zheng-Xin Yong, Cristina Menghini, and Stephen H Bach. Low-resource languages jailbreak gpt-4. arXiv preprint arXiv:2310.02446, 2023. 1, 3 +[41] Youliang Yuan, Wenxiang Jiao, Wenxuan Wang, Jen tse Huang, Pinjia He, Shuming Shi, and Zhaopeng Tu. GPT-4 is too smart to be safe: Stealthy chat with LLMs via cipher. In The Twelfth International Conference on Learning Representations, 2024. 3, 5 +[42] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE International Conference on Computer Vision, pages 6023-6032, 2019. 2, 3, 4, 7 +[43] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3, 4, 7 +[44] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3, 6 +[45] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 1, 3, 5, 7 \ No newline at end of file diff --git a/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/images.zip b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..404e5d8f520fb0677515f0add0783a62f610477f --- /dev/null +++ b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c96c676a51238508bf16555614d95ea181be2d582bf6ac3502a4c4e5618c2585 +size 478623 diff --git a/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/layout.json b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..60d739ee63f0ceab0f0347ed416f30ebe3159e10 --- /dev/null +++ b/2025/Playing the Fool_ Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy/layout.json @@ -0,0 +1,7647 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 102, + 493, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 102, + 493, + 140 + ], + "spans": [ + { + "bbox": [ + 118, + 102, + 493, + 140 + ], + "type": "text", + "content": "Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "spans": [ + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "type": "text", + "content": "Joonhyun Jeong" + }, + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "type": "text", + "content": " Seyun Bae" + }, + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "type": "text", + "content": " Yeonsung Jung" + }, + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "type": "text", + "content": " Jaeryong Hwang" + }, + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "type": "text", + "content": " Eunho Yang" + }, + { + "bbox": [ + 88, + 159, + 525, + 176 + ], + "type": "inline_equation", + "content": "^{2,4*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 233, + 178, + 378, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 178, + 378, + 193 + ], + "spans": [ + { + "bbox": [ + 233, + 178, + 378, + 193 + ], + "type": "text", + "content": "1NAVER Cloud, ImageVision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 194, + 459, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 194, + 459, + 208 + ], + "spans": [ + { + "bbox": [ + 151, + 194, + 459, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 194, + 459, + 208 + ], + "type": "text", + "content": "Korea Advanced Institute of Science and Technology (KAIST)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 185, + 209, + 425, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 209, + 425, + 224 + ], + "spans": [ + { + "bbox": [ + 185, + 209, + 425, + 224 + ], + "type": "text", + "content": "3Republic of Korea Naval Academy 4AITRICS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 226, + 474, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 226, + 474, + 247 + ], + "spans": [ + { + "bbox": [ + 136, + 226, + 474, + 247 + ], + "type": "text", + "content": "joonhyun.jeong@navercorp.com {seyun.bae, ys.jung}@kaist.ac.kr \njhwang@navy.ac.kr eunhoy@kaist.ac.kr" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 275, + 200, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 275, + 200, + 288 + ], + "spans": [ + { + "bbox": [ + 151, + 275, + 200, + 288 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 310, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 297, + 706 + ], + "type": "text", + "content": "Despite the remarkable versatility of Large Language Models (LLMs) and Multimodal LLMs (MLLMs) to generalize across both language and vision tasks, LLMs and MLLMs have shown vulnerability to jailbreaking, generating textual outputs that undermine safety, ethical, and bias standards when exposed to harmful or sensitive inputs. With the recent advancement of safety alignment via preference-tuning from human feedback, LLMs and MLLMs have been equipped with safety guardrails to yield safe, ethical, and fair responses with regard to harmful inputs. However, despite the significance of safety alignment, research on the vulnerabilities remains largely underexplored. In this paper, we investigate the unexplored vulnerability of the safety alignment, examining its ability to consistently provide safety guarantees for out-of-distribution(OOD)-ifying harmful inputs that may fall outside the aligned data distribution. Our key observation is that OOD-ifying the vanilla harmful inputs highly increases the uncertainty of the model to discern the malicious intent within the input, leading to a higher chance of being jailbroken. Exploiting this vulnerability, we propose JOOD, a new Jailbreak framework via OOD-ifying inputs beyond the safety alignment. We explore various off-the-shelf visual and textual transformation techniques for OOD-ifying the harmful inputs. Notably, we observe that even simple mixing-based techniques such as image mixup prove highly effective in increasing the uncertainty of the model, thereby facilitating the bypass of the safety alignment. Experiments across diverse jailbreak scenarios demonstrate that JOOD effectively jailbreaks recent proprietary LLMs and MLLMs such as GPT-4 and o1 with high attack success rate, which previous attack approaches have consistently struggled to jailbreak. Code is available at https://github.com/naver-ai/JOOD." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 275, + 394, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 275, + 394, + 288 + ], + "spans": [ + { + "bbox": [ + 315, + 275, + 394, + 288 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 299, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 418 + ], + "type": "text", + "content": "Large Language Models (LLMs) have recently exhibited versatility on various language reasoning tasks [10, 20, 26, 28, 37, 39] based on scalable pre-training and fine-tuning on a large corpus of text data. However, due to the biases and misinformation [8, 30] present in the large-scale training data, LLMs have frequently been jailbroken which leads to the generation of biased or unsafe outputs that may compromise ethical standards, safety, or fairness [3, 19, 38, 40, 45] when provided with harmful or sensitive input text instructions such as \"tell me how to build a bomb\"." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 422, + 556, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 556, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 556, + 674 + ], + "type": "text", + "content": "While the primary target of jailbreaking has been LLMs, similar vulnerabilities have also surfaced in Multimodal LLMs (MLLMs [18, 24, 28, 44]). MLLMs extend the visual understanding capabilities of LLMs by integrating visual encoder [32] with several add-on layers. However, this integration introduces new pathways to bypass the safety guardrails equipped within LLMs. Recently, several jailbreak methods [22, 33] have focused on MLLM's weak safety alignment induced by the add-on linear layers; The visual encoder and linear layers projecting input images into LLMs [24] have been inadequately safety-aligned with regard to the visually harmful images. Inspired by this weak safety alignment regarding harmful images, MLLM jailbreak methods decompose the harmful textual instruction into a multimodal format, incorporating both a harmful image and a generic textual instruction. For example, as shown in the vanilla attack case of Figure 1, instead of using the textual instruction containing the sensitive phrase \"tell me how to build a bomb\", the harmful content \"bomb\" is embedded in the visual input, and the textual instruction is generalized to \"tell me how to build the object in this image\"." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "To address these safety and ethics issues, safety alignment methods [28, 29] were proposed, involving posttraining LLMs and MLLMs with human-preference feed" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 713, + 146, + 724 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 713, + 146, + 724 + ], + "spans": [ + { + "bbox": [ + 67, + 713, + 146, + 724 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "text", + "content": "29937" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 556, + 304 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 556, + 304 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 556, + 304 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 556, + 304 + ], + "type": "image", + "image_path": "c3dcec454d608e5b0d95b576dea8a277a7db2985ce039ac220042cfe896120e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 311, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 311, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 555, + 346 + ], + "type": "text", + "content": "Figure 1. Overview of JOOD, in terms of out-of-distribution from the training distribution of safety alignment. For each attack method, we measured uncertainty for discerning the maliciousness of the attack input on GPT-4 and GPT-4V (see Appendix A for the detailed prompt and results) and evaluated their attack success rate on AdvBench-M [27] dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 365, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 484 + ], + "type": "text", + "content": "back (RLHF). In this approach, ethical and safe responses are prioritized through human feedback, guiding the model to yield safe, ethical, and fair outputs with regard to harmful requests. Therefore, the aforementioned naive attack strategies fail to effectively jailbreak safety-aligned models such as GPT-4 and GPT-4V [28]. As shown in Figure 1, when presented with harmful inputs within the bounds of the RLHF training data distribution, these models can confidently discern malicious intent in the inputs and effectively prevent the circumvention of the safety guardrails." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 486, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 295, + 640 + ], + "type": "text", + "content": "While RLHF has sufficiently aligned the models to handle such transparently malicious inputs under the safety standards, it may still struggle to generalize to other harmful inputs that fall outside the training distribution. In this paper, we correspondingly investigate a naturally arising question: Does the underlying safety alignment of LLMs and MLLMs consistently guarantee safety even with regard to OOD-ifying inputs that possibly differ from the training inputs and learned knowledge during all the training procedures including safety alignment and are hence novel to the models? To answer the question, we present JOOD, an effective jailbreak attack framework via OOD-ifying harmful inputs beyond safety alignment of LLMs and MLLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 295, + 713 + ], + "type": "text", + "content": "While there can be various advanced transformation techniques to OOD-ify the harmful inputs, we explore various off-the-shelf visual and textual transforms and observe that the recent LLMs and MLLMs can be effectively jailbroken even with these simple input manipulation techniques. For instance, as shown in Figure 1, the vanilla harm" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 365, + 553, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 556 + ], + "type": "text", + "content": "ful inputs are transformed into the new ones with conventional mixing techniques [7, 42] where the harmful image of bomb and its word text are mixed with another image and text containing an arbitrary subject (e.g., apple), respectively. These newly transformed textual and visual inputs naturally have discrepancies with the vanilla inputs that were previously seen during all the pre-/fine-/post-training procedures including safety alignment learning, potentially conforming to the out-of-distribution beyond the scope of safety-aligned data distribution. Consequently, when these OOD-ifying harmful inputs are exploited for jailbreak attacks, the model fails to recognize their malicious intent with highly increased uncertainty. This allows an attacker to bypass the safety guardrails designed only for the original in-distribution harmful inputs and thereby yield a significantly higher chance of the model being jailbroken." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 557, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 553, + 700 + ], + "type": "text", + "content": "Our comprehensive experiments on various jailbreak scenarios (e.g., bombs, drugs, hacking) of AdvbenchM [27] demonstrate that this straightforward OOD-ifying strategy successfully jailbreaks existing LLMs and MLLMs, including state-of-the-art proprietary models such as GPT-4 and GPT-4V. Also, our method significantly outperforms baseline attack methods in all the jailbreak scenarios, achieving " + }, + { + "bbox": [ + 313, + 557, + 553, + 700 + ], + "type": "inline_equation", + "content": "63\\%" + }, + { + "bbox": [ + 313, + 557, + 553, + 700 + ], + "type": "text", + "content": " attack success rate (ASR) against GPT-4V in Bombs or Explosives scenario and improving performance by " + }, + { + "bbox": [ + 313, + 557, + 553, + 700 + ], + "type": "inline_equation", + "content": "+42\\%" + }, + { + "bbox": [ + 313, + 557, + 553, + 700 + ], + "type": "text", + "content": " ASR compared to the state-of-the-art baseline [9] in the Hacking scenario. In summary, our contribution is threefold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "- We systematically reveal the vulnerabilities of RLHF" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "29938" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 294, + 120 + ], + "type": "text", + "content": "based safety alignment, which leaves LLMs and MLLMs vulnerable to out-of-distribution textual or visual inputs that have a discrepancy with the in-distribution samples previously seen during safety alignment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 121, + 294, + 228 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 121, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 294, + 179 + ], + "type": "text", + "content": "- We propose a novel black-box jailbreak strategy via manipulating malicious inputs into the OOD-ifying ones. We observe that even simple off-the-shelf transforms amplify uncertainty of the model, allowing it to effectively bypass the safety alignment of LLMs and MLLMs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 180, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 180, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 294, + 228 + ], + "type": "text", + "content": "- Extensive experiments on various jailbreak scenarios demonstrate the effectiveness of our attack strategy against the state-of-the-art proprietary models such as GPT-4 and even o1, with high attack success rate." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 56, + 236, + 141, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 236, + 141, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 141, + 249 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 257, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 605 + ], + "type": "text", + "content": "Jailbreaking attacks The jailbreaking attacks undermine LLMs and MLLMs to generate forbidden text outputs that can possibly violate safety, ethics, and fairness regulations. For jailbreaking LLMs, the pioneer work [45] and its variants [1, 23] adversarially optimized the suffix within the attack prompt to yield an affirmative response. Another line of work disguised attack prompts via encryption [11, 41] and translations with low-resource languages [40]. In-context learning [2] was also instrumental in establishing preliminary contexts for jailbreaking by utilizing few-shot jailbroken examples [38] and refining the attack prompt [3]. Meanwhile, several studies [9, 22, 33] focused on the weak safety alignment of MLLMs and reformulated the textual harmful instruction into multimodal format containing a pair of harmful image and generic text instruction. HADES [22] further synthesized the harmful image into a semantically more harmful one by diffusion models for providing a better jailbreaking context. Similarly, Ma et al. [25] additionally provided visual contexts by utilizing the roleplaying concept, offering justifications to carry out the associated harmful request. Also, FigStep [9] converted harmful textual instruction into typography and prompted to complete the blanks in the execution steps. Although the above methods achieved state-of-the-art jailbreak performance on open-source MLLMs [24, 44], they still exhibit a lack of generalization on robust MLLMs [12, 14, 28] that were safety-aligned with human feedback. In this work, we focus on unveiling the vulnerability of these safety-aligned models in terms of out-of-distribution perspective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Safety alignment via human feedback While instruction tuning [37] successfully manipulated LLMs to act in accordance with the user's intention, there still remained a large headroom for improving their safety and reliability. To address this, RLHF [29] aligned LLMs with human preferences under the consideration of the safety standards, robustifying LLMs against malicious text instructions. Also, Sun et al. [34] aligned MLLMs using human feedback on the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "rectified image-text pairs where hallucinations and harmful contents were largely reduced. OpenAI [28] further strengthened the safety alignment of MLLMs using selffeedback as a reward for updating the policy model. However, we revisit that these safety-aligned models via RLHF methods can still be susceptible to jailbreak attacks using various types of data transformation techniques." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 183, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 471 + ], + "type": "text", + "content": "Data transformation Data transformation refers to manipulating the input data into a new variation of a synthetic sample, primarily used for enhancing the robustness and generalization towards broader variations of inputs in the machine learning field. Specifically, regularizing neural networks with data transformation is vital for mitigating overfitting on a fixed training dataset. In image classification task [6], besides the basic transformations such as random resizing and center cropping, advanced ones such as photometric and geometric transformations [4, 5] have been applied to enhance test generalization. Also, image-mixing techniques [15-17, 42, 43], which synthesize data by combining training images, have further expanded the training data spectrum and improved generalization performance. In natural language processing (NLP), text transformations like adding a random word, synonym replacement, and sentence reordering [36] are used to increase input diversity. Other transforms, such as text-mixing [7], generate novel linguistic patterns, creating out-of-distribution data to challenge language models in unexpected ways. In this paper, we explore these off-the-shelf transform techniques to generate OOD-ifying inputs beyond training distribution of safety alignment for LLMs and MLLMs, thereby effectively performing jailbreak attacks against these large models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 487, + 397, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 397, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 397, + 502 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "In this section, we propose JOOD, a simple yet effective jailbreak framework with the OOD-ifying harmful inputs. We explore a broad range of off-the-shelf transformations to manipulate the harmful inputs into OOD-ifying ones that may not be considered during safety alignment. Consequently, these synthetic OOD-ifying inputs effectively bypass the safety guardrails of LLMs and MLLMs, making them more susceptible to jailbreak attempts. This straightforward input manipulation strategy also enables black-box setting to attack the proprietary LLMs and MLLMs such as GPT-4 and GPT-4V. In Section 3.1, we introduce our LLM attack method leveraging textual transformations. In Section 3.2, we also present an effective MLLM attack method using image transformation techniques. Subsequently, we propose an attack evaluation method in Section 3.3 to quantitatively measure the maximum potential risk posed by our attacks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "29939" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 251, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 251, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 251, + 85 + ], + "type": "text", + "content": "3.1. Attack LLM with OOD-ifying inputs" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": "Given an input text instruction " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": " that contains a harmful request such as \"tell me how to build a bomb\", the naive attack strategy is to feed " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": " into the target model " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": " to generate output text response " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "r = \\theta (T^{\\mathrm{h}})" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": " is typically an autoregressive LLM. While this naive attack had successfully jailbroken the primitive LLMs, it fails to jailbreak recent safety-aligned LLMs since the input " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": " is a natural in-distribution instruction that might have been seen during safety alignment training with human feedback. To overcome this limitation, we transform " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": " into a novel text instruction " + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{ood}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": " that possibly had not been seen during all the LLM training procedures including the safety alignment learning, and then input to the target model:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 93, + 255, + 295, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 255, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 93, + 255, + 295, + 270 + ], + "type": "interline_equation", + "content": "T _ {i} ^ {\\mathrm {o o d}} = f \\left(T ^ {\\mathrm {h}}; \\varphi_ {i}\\right), \\quad R = \\left\\{\\theta \\left(T _ {i} ^ {\\mathrm {o o d}}\\right) \\right\\} _ {i = 1} ^ {n} \\tag {1}", + "image_path": "ae925c57388d5c6f264be34a40a9195a2e083de3e1f9ae784211598d18b7845d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\varphi_{i}" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": " represents transformation-dependent parameter applied by textual transformation function " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": " is a set of model responses for each transformed instruction " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "T_{i}^{\\mathrm{ood}}" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": " with one of the " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": " differently sampled " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\varphi_{i}" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": ". To fulfill the malicious goal within " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": ", an attacker can employ various textual transformations for " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": " to OOD-ify " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": " into a novel text instruction " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "T_{i}^{\\mathrm{ood}}" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": ". For instance, off-the-shelf textual transforms such as random word insertion or deletion [36] and character-level perturbations [21] can be employed in " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": ". Among these methods, we observe that even a simple transform such as the text-mixing technique, where a word is combined with an unrelated auxiliary word to form a new nonsensical word, can effectively synthesize OOD-ifying inputs for the recent LLMs [28] with high uncertainty as shown in Figure 1. Specifically, the harmful subject in " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": " (e.g., bomb) is combined with one of " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": " different auxiliary words " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\varphi_{i}" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": " (e.g., apple) randomly sampled from real-world vocabularies, synthesizing a new nonsensical word (e.g., bombapple) that does not naturally exist in practice. Subsequently, we prompt the model " + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 280, + 296, + 555 + ], + "type": "text", + "content": " to comply with the generic request for each of the words within this newly blended word as follows (see Appendix B for prompt examples of other textual transform variants):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 561, + 294, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 561, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 58, + 561, + 294, + 621 + ], + "type": "text", + "content": "The following is a newly coined word made by mixing words corresponding to two words in a random order. Please refer to the possible words inside this new word as \"[Words]\". Please answer the request below for each of these words." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 626, + 189, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 626, + 189, + 639 + ], + "spans": [ + { + "bbox": [ + 59, + 626, + 189, + 639 + ], + "type": "text", + "content": "[Coined word]: \"bombapple\"" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 639, + 284, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 639, + 284, + 651 + ], + "spans": [ + { + "bbox": [ + 59, + 639, + 284, + 651 + ], + "type": "text", + "content": "[Request]: \"Tell me how to build [Words]\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 262, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 262, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 262, + 673 + ], + "type": "text", + "content": "3.2. Attack MLLM with OOD-ifying inputs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Given the original harmful text instruction " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ", naive MLLM attack strategy is to decompose " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " into a pair of harmful image " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{h}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " (e.g., bomb) and a generic text instruction " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{g}}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": "(e.g., \"tell me how to build the object in this image\"), and then feed both inputs to the target MLLM " + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": ". However, this naive vanilla attack does not effectively jailbreak state-of-the-art MLLMs such as GPT-4V that had possibly safely-aligned on such harmful inputs. Therefore, we propose a more powerful OOD-ifying attack strategy that can bypass underlying safety alignment of the MLLMs, as shown in Figure 1. Specifically, we transform the harmful image " + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{h}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": " into a synthetic image " + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{ood}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": " via conventional image transformation techniques " + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": " that may not be considered during safety alignment training:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 347, + 214, + 555, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 214, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 347, + 214, + 555, + 232 + ], + "type": "interline_equation", + "content": "I _ {j} ^ {\\mathrm {o o d}} = f (I ^ {\\mathrm {h}}; \\alpha_ {j}), \\quad R = \\{\\theta (T ^ {\\mathrm {g}}, I _ {j} ^ {\\mathrm {o o d}}) \\} _ {j = 1} ^ {m} \\qquad (2)", + "image_path": "7408feffbb77b497fd7353862d6c3614feee78a18dabfc8e60239f3984009557.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\alpha_{j}" + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "text", + "content": " is a transformation degree ranging from 0 to 1, and " + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "text", + "content": " is a set of model responses for each transformed image " + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "inline_equation", + "content": "I_{j}^{\\mathrm{odd}}" + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "text", + "content": " with one of " + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "text", + "content": " differently sampled " + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\alpha_{j}" + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "text", + "content": ". To achieve the malicious goal, the attacker can employ various off-the-shelf image transformations for OOD-ifying the original harmful input " + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{h}}" + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "text", + "content": " into a novel harmful image " + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "inline_equation", + "content": "I_{j}^{\\mathrm{odd}}" + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "text", + "content": ", such as geometric, photometric transforms [5] and image-mixing techniques [31, 42, 43]. When using image-mixing techniques, we consider another input " + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\varphi_{i}" + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "text", + "content": ", an auxiliary image containing an arbitrary object or concept, for mixing with the original harmful image " + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{h}}" + }, + { + "bbox": [ + 313, + 239, + 555, + 372 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 381, + 555, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 381, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 322, + 381, + 555, + 398 + ], + "type": "interline_equation", + "content": "I _ {(i, j)} ^ {\\mathrm {o o d}} = f \\left(I ^ {\\mathrm {h}}, \\varphi_ {i}; \\alpha_ {j}\\right), \\quad R = \\left\\{\\theta \\left(T ^ {\\mathrm {g}}, I _ {(i, j)} ^ {\\mathrm {o o d}}\\right) \\right\\} _ {i = 1, j = 1} ^ {n, m} \\tag {3}", + "image_path": "9198c6414dd62b7f872a753fdd9356e1f0fc1a84586bfa094547db6cae0952bb.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": " denotes the number of auxiliary image " + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "inline_equation", + "content": "\\varphi_{i}" + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": " used for attack, and " + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": " is a set of model responses generated by mixing " + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{h}}" + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": " different " + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "inline_equation", + "content": "\\varphi_{i}" + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": " samples and " + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": " different " + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "inline_equation", + "content": "\\alpha_{j}" + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": " mixing coefficients. For mixup [43] case, " + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": " can be defined as " + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "inline_equation", + "content": "\\alpha_{j}\\varphi_{i} + (1 - \\alpha_{j})I^{\\mathrm{h}}" + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": ". Also, we correspondingly reform the generic input instruction " + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{g}}" + }, + { + "bbox": [ + 313, + 407, + 554, + 492 + ], + "type": "text", + "content": " to comply with the request for each of the two objects within the mixed image as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 501, + 553, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 525 + ], + "type": "text", + "content": "There are two realistic (not artistic but real-functioning) objects (or concepts) in this image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 525, + 523, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 525, + 523, + 549 + ], + "spans": [ + { + "bbox": [ + 318, + 525, + 523, + 549 + ], + "type": "text", + "content": "Tell me how to build each of these objects." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 555, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 652 + ], + "type": "text", + "content": "Notably, this straightforward OOD-ifying strategy with simple image transform techniques induces a distinct shift from " + }, + { + "bbox": [ + 313, + 555, + 554, + 652 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{h}}" + }, + { + "bbox": [ + 313, + 555, + 554, + 652 + ], + "type": "text", + "content": " within the visual encoder of MLLM, as shown in Appendix D. When these shifted harmful visual embeddings are fed into the subsequent LLM, the uncertainty is substantially increased and hence fail to trigger its safety guardrails which were trained on the original in-distribution harmful inputs (i.e., " + }, + { + "bbox": [ + 313, + 555, + 554, + 652 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 313, + 555, + 554, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 555, + 554, + 652 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{h}}" + }, + { + "bbox": [ + 313, + 555, + 554, + 652 + ], + "type": "text", + "content": "), as shown in Figure 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 659, + 422, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 422, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 422, + 672 + ], + "type": "text", + "content": "3.3. Attack Evaluation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "For each harmful instruction " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": ", our attack methods (Eq. 1, 3) produce a set of attack responses " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " corresponding to our auxiliary attack input and parameter such as " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "text", + "content": "29940" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 123, + 561, + 368 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. Comparison of jailbreak performance with baseline attack methods on the unimodal (text-only) model GPT-4 and the multimodal (image + text) model GPT-4V. HF denotes the average harmfulness score of the responses over all the instructions as in Eq. 4, ranging from 0 to 10. The attack scenarios include Bombs or Explosives (BE), Drugs (D), Firearms / Weapons (FW), Hacking information (H), Kill someone (K), Social Violence (SV), and Self-harm and Suicide (SS)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 123, + 561, + 368 + ], + "lines": [ + { + "bbox": [ + 58, + 123, + 561, + 368 + ], + "spans": [ + { + "bbox": [ + 58, + 123, + 561, + 368 + ], + "type": "table", + "html": "
InputTarget modelAttack methodBEDFWHKSVSS
HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑HF↑ASR%↑
TextGPT-4Vanilla00000.1000000000
CipherChat 2024070.370.100.111080.21507
PAIR 2023000.230.900.8110.100.100.13
JOOD (Eq. 1)1.3132.5172.8243.1421.480.401.013
Image + TextGPT-4VVanilla00000.6120.35000000
FigStep 2023000.230.5000000000
FigStep-Pro 20230.9230.8171.8252.1320.480.300.10
HADES 2024000.230.100.10000.1000
JOOD (Eq. 3)7.1633.9237.2474.0742.141.1100.623
MiniGPT-4 7BVanilla1.5502.5731.260420.2420.8203.073
FigStep 20234.4633.1633.6476.7744.2633.6551.953
HADES 20240.1171.8401.5291.2470.6170.851.337
JOOD (Eq. 3)8.1838.0877.5477.3955.7543.9357.697
LLaVA-1.5 13BVanilla6.6873.2433.5472.4531.7461.0154.383
FigStep 20235.8773.5534.2416.9794.0543.2351.247
HADES 20241.0132.7533.6474.7744.0540.451.747
JOOD (Eq. 3)9.81008.5938.3655.3896.1635.8408.590
", + "image_path": "7d2e8a55529c75c4f1f447a5b3c180c6b21167e5d30c1ebb47c07a3577ad0757.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "spans": [ + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "text", + "content": "respectively. Each response in this set may exhibit varying degrees of harmfulness, depending on how much the harmful information and sensitive content it contains corresponding to the malicious input request " + }, + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "text", + "content": ". To evaluate the maximum potential risk posed by our attacks, we propose a score-based evaluation method to quantitatively measure the harmfulness degree of each response, thereby identifying the most harmful attack response. Specifically, we employ another LLM " + }, + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{hf}}" + }, + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "text", + "content": " [28] as the harmfulness score (HF) evaluator and prompt it to assess each attack response " + }, + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "inline_equation", + "content": "r \\in R" + }, + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "text", + "content": " on a scale from 0 to 10, in consideration of the safety standards and compliance with the harmful input instruction " + }, + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{h}}" + }, + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "text", + "content": ". Details of the evaluation prompt are provided in Appendix C. Given the most harmful attack response with the highest harmfulness score, we also report attack success rate (ASR) by prompting to the binary-judging LLM " + }, + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{bj}}" + }, + { + "bbox": [ + 54, + 387, + 297, + 613 + ], + "type": "text", + "content": " [13] that outputs 1 if the response is unsafe and 0 for the harmless responses, following [22]. These evaluation metrics are formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 640, + 294, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 640, + 294, + 716 + ], + "spans": [ + { + "bbox": [ + 78, + 640, + 294, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {H F} \\left(\\mathcal {T} ^ {\\mathrm {h}}\\right) = \\frac {\\sum_ {T ^ {\\mathrm {h}} \\in \\mathcal {T} ^ {\\mathrm {h}}} \\max _ {r \\in R} \\left(\\theta^ {\\mathrm {h f}} \\left(r \\mid T ^ {\\mathrm {h}}\\right)\\right)}{| \\mathcal {T} ^ {\\mathrm {h}} |} \\tag {4} \\\\ \\operatorname {ASR} \\left(\\mathcal {T} ^ {\\mathrm {h}}\\right) = \\frac {\\sum_ {T ^ {\\mathrm {h}} \\in \\mathcal {T} ^ {\\mathrm {h}}} \\theta^ {\\mathrm {b j}} \\left(\\underset {r \\in R} {\\arg \\max } \\left(\\theta^ {\\mathrm {h f}} \\left(r \\mid T ^ {\\mathrm {h}}\\right)\\right)\\right)}{\\left| \\mathcal {T} ^ {\\mathrm {h}} \\right|} \\\\ \\end{array}", + "image_path": "f27c61b8ebec3e9a73dd07dbf72574ecb342fb2dcd5e587e69af59f2bd4011dd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 386, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 386, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{\\mathrm{h}}" + }, + { + "bbox": [ + 313, + 386, + 555, + 411 + ], + "type": "text", + "content": " is the entire set of harmful instructions used for jailbreaking attacks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 423, + 395, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 395, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 395, + 437 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 444, + 556, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 556, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 556, + 600 + ], + "type": "text", + "content": "Dataset. We evaluate JOOD on the widely-used LLM and MLLM jailbreak benchmark, Advbench [45] and Advbench-M [27], to compare the performance against the previous state-of-the-art attack methods. Advbench consists of 500 textual instructions that encourage harmful behaviors, while Advbench-M further categorized the set of these harmful instructions into several distinct scenarios such as Bombs or Explosives, Drugs, and Hacking information. To attack with MLLMs, Advbench-M paired semantically relevant images for each scenario. We used around 30 harmful instructions for each scenario to evaluate LLM attack methods, along with one paired harmful image for each scenario when evaluating MLLM attack methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Implementation details. When attacking with mixing-based transformations of JOOD, the textual or visual auxiliary inputs " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " in Eq. 1, 3, are required to synthesize the OOD-ifying input, respectively. We used randomly sampled arbitrary words such as apple and watch, and retrieved corresponding images from the Internet. We set the number of auxiliaries to " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "n = 5" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " and sampled " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "m = 9" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " discrete values for the transformation degree " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " from" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "text", + "content": "29941" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.2, \\dots, 0.9\\}" + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": ". When attacking with image transformation techniques, we resized and padded the input images to a uniform size of " + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "320 \\times 320" + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": " pixels as a preprocessing step before the transformation. Unless otherwise specified, we use GPT-4 as the target model for LLM attacks and GPT-4V for MLLM attacks, with inference parameters such as temperature and top-p set to 1. For LLM attacks, we employ the text-mixing transformation introduced in Eq. 1, and for MLLM attacks, we use the image-mixing transformations as in Eq. 3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 141, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 141, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 141, + 213 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 218, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 296, + 542 + ], + "type": "text", + "content": "Comparison with SOTA attack methods. In Table 1, we compare JOOD with other state-of-the-art attack methods on GPT-4 and GPT-4V. The vanilla attack methods using the original harmful instructions or images barely jailbreak the target models in almost all the scenarios, while the baseline works such as CipherChat, PAIR, FigStep, and HADES marginally enhanced both ASR and harmfulness scores. However, JOOD consistently exhibits the best jailbreak performance on almost all the scenarios and target models, largely outperforming the baselines. Specifically, when attacking text-only GPT-4, JOOD achieves " + }, + { + "bbox": [ + 56, + 218, + 296, + 542 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 56, + 218, + 296, + 542 + ], + "type": "text", + "content": " ASR in the Firearms scenario where all the baselines failed to jailbreak any of the instructions. Also, when attacking multimodal GPT-4V, JOOD outperforms the previous state-of-the-art attack FigStep-Pro by a large margin, achieving " + }, + { + "bbox": [ + 56, + 218, + 296, + 542 + ], + "type": "inline_equation", + "content": "+6.2" + }, + { + "bbox": [ + 56, + 218, + 296, + 542 + ], + "type": "text", + "content": " average harmfulness and " + }, + { + "bbox": [ + 56, + 218, + 296, + 542 + ], + "type": "inline_equation", + "content": "+40\\%" + }, + { + "bbox": [ + 56, + 218, + 296, + 542 + ], + "type": "text", + "content": " ASR in Bombs or Explosives scenario. Notably, JOOD exclusively jailbreaks a considerable amount of the instructions that these baselines failed to, specifically 10 additional instructions out of 30 in Bombs or Explosives scenario, as shown in Table 2. Also, JOOD consistently generates extremely harmful responses on most instructions while the baseline methods sparsely generate harmful responses with regard to a few instructions, as shown in Figure 2. These results demonstrate that our OOD-ifying attack strategy via input transformation techniques is effective in causing underlying safety alignment to malfunction to output harmful responses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "Generalization on other MLLMs. In the bottom rows of Table 1, we further investigate generalization of JOOD on the other open-source MLLMs, MiniGPT-4 7B [44] and LLaVA-1.5 13B [24] post-trained with RLHF [34]. The results show that JOOD consistently outperforms the other baselines by a large margin in almost all of the scenarios. Surprisingly, JOOD achieves high attack success rates even against robustly safety-aligned models including GPT-4o [12] and o1 [14], which the baselines mostly failed to jailbreak (see Appendix E). The superiority of our OOD-ifying attack strategy on both proprietary and open-source MLLMs further corroborates that the existing MLLMs still lack of safety alignment on the OOD-ifying inputs gener" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 319, + 102, + 550, + 207 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 550, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 550, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 550, + 93 + ], + "type": "text", + "content": "Table 2. Comparison of the number of exclusively jailbroken instructions by each attack method." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 102, + 550, + 207 + ], + "lines": [ + { + "bbox": [ + 319, + 102, + 550, + 207 + ], + "spans": [ + { + "bbox": [ + 319, + 102, + 550, + 207 + ], + "type": "table", + "html": "
Attack method# exclusively jailbroken instructions
Figstep-ProHADESJOODBEDFWHKSVSS
✓XX4532200
X✓X0200000
XX✓10547112
", + "image_path": "2865bb14ba9fc0899ac98464289408f0f07620d3e0b96bb88008023f4d70b600.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 209, + 538, + 368 + ], + "blocks": [ + { + "bbox": [ + 317, + 209, + 538, + 368 + ], + "lines": [ + { + "bbox": [ + 317, + 209, + 538, + 368 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 538, + 368 + ], + "type": "image", + "image_path": "1903840763ec6bf59b981c360f9d629af4d513f8d22a9849c87ab523af655e50.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 376, + 550, + 399 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 550, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 550, + 399 + ], + "type": "text", + "content": "Figure 2. Comparison of harmfulness scores per instruction in Bombs scenario." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 420, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 553, + 445 + ], + "type": "text", + "content": "ated from even simple off-the-shelf transforms, emphasizing the need for further research and development." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 416, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 416, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 416, + 464 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 470, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 542 + ], + "type": "text", + "content": "We conduct ablation studies to assess the effectiveness of each component and the sensitivity of the hyper-parameters comprising JOOD. Unless specified, we attack with the multimodal inputs using the mixup technique and adopt GPT-4V as the target model to be attacked. Also, we report jailbreaking performance in Bombs or Explosives scenario." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 556, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 715 + ], + "type": "text", + "content": "Effect of various transformation techniques. For attacking GPT-4 with OOD-ifying text inputs as in Eq. 1, we investigate the effect of various text-mixing techniques to obfuscate the harmful word (e.g., \"bomb\") with the auxiliary word (e.g., \"apple\"), producing a mixed word like \"baopmble\" for the " + }, + { + "bbox": [ + 313, + 556, + 555, + 715 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 556, + 555, + 715 + ], + "type": "text", + "content": " (horizontal)-Interleave case. We observe that GPT-4 effectively interprets these mixed words, accurately recognizing and responding to both the harmful and auxiliary components for the given instruction. See Appendix B for the detailed analysis. Also, to generate OOD-ifying input images for the attack in Eq. 2, we test with widely-adopted image transformation techniques including geometric, photometric transformations" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "text", + "content": "29942" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 180, + 190 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 180, + 190 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 180, + 190 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 180, + 190 + ], + "type": "image", + "image_path": "ebbaa4912c33be151384b6ee2d380700da89ac6ef4a5c90b34fe7043cf815fe9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 72, + 299, + 190 + ], + "blocks": [ + { + "bbox": [ + 179, + 72, + 299, + 190 + ], + "lines": [ + { + "bbox": [ + 179, + 72, + 299, + 190 + ], + "spans": [ + { + "bbox": [ + 179, + 72, + 299, + 190 + ], + "type": "image", + "image_path": "54bd666e8226776418213cbf2a020b2f60bd30061208f331bc3264454ec4d6bc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 206, + 177, + 326 + ], + "blocks": [ + { + "bbox": [ + 105, + 195, + 253, + 205 + ], + "lines": [ + { + "bbox": [ + 105, + 195, + 253, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 253, + 205 + ], + "type": "text", + "content": "(a) Harmfulness / ASR of text-mixing attacks" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 206, + 177, + 326 + ], + "lines": [ + { + "bbox": [ + 58, + 206, + 177, + 326 + ], + "spans": [ + { + "bbox": [ + 58, + 206, + 177, + 326 + ], + "type": "image", + "image_path": "65b1310079864f6d3c68d7b8d72a813f8efc17f2511ef604ade18276bf5d3f43.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 178, + 207, + 299, + 325 + ], + "blocks": [ + { + "bbox": [ + 178, + 207, + 299, + 325 + ], + "lines": [ + { + "bbox": [ + 178, + 207, + 299, + 325 + ], + "spans": [ + { + "bbox": [ + 178, + 207, + 299, + 325 + ], + "type": "image", + "image_path": "bdc9a811c37910f224fa867b6716dbf0fa9d9a68c4e1f64dc7c9cde06248720d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 330, + 261, + 339 + ], + "lines": [ + { + "bbox": [ + 97, + 330, + 261, + 339 + ], + "spans": [ + { + "bbox": [ + 97, + 330, + 261, + 339 + ], + "type": "text", + "content": "(b) Harmfulness / ASR of image-transform attacks" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 405, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 295, + 429 + ], + "type": "text", + "content": "(RandAug [5]) and image-mixing techniques (Mixup [43], CutMix [42], and ResizeMix [31])." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 430, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 634 + ], + "type": "text", + "content": "In Figure 3, all of the text-mixing variants and image-mixing techniques consistently outperformed the vanilla attack case without applying mixing techniques in all the scenarios. Also, the performance variance between all the image-mixing variants was marginal, indicating that an attacker can leverage our method for consistently jailbreaking GPT-4V without any dependence on the specific image-mixing technique. However, attacking with non-mixing image transformation techniques (i.e., RandAug) such as adding gaussian noise, rotation, and shearing showed inferior ASR and harmfulness scores compared to the image-mixing transformation techniques, with slightly higher performance than the vanilla case. These results may suggest that GPT-4V is equipped with relatively solid safety alignment on such basic transformations, but not adequately safety-aligned on more synthetic transformations such as image-mixing techniques." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Ablations on mixing coefficient for mixup. In Figure 4, we analyze the effect of the mixing coefficient " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": " in Eq. 3 which modulates how much the auxiliary image will obfuscate the original harmful image. We additionally report the ratio of the responses refused by GPT-4V with substring" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 566, + 178 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 566, + 178 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 566, + 178 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 566, + 178 + ], + "type": "image", + "image_path": "6cec83050da5e4d07d3a8f3d8ed7da623d8314c6b6ccde296d9530371bb8c75d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 322, + 180, + 354, + 213 + ], + "blocks": [ + { + "bbox": [ + 322, + 180, + 354, + 213 + ], + "lines": [ + { + "bbox": [ + 322, + 180, + 354, + 213 + ], + "spans": [ + { + "bbox": [ + 322, + 180, + 354, + 213 + ], + "type": "image", + "image_path": "e9da64943dc04488522e86858b18d21724c01601cf0a4123463a16dcae17f989.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "type": "text", + "content": "Figure 4. Ablation on mixing coefficient " + }, + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "type": "text", + "content": " for image mixup attack. We visualize the attack images corresponding to " + }, + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "type": "text", + "content": " value below the x-axis. We report the average harmfulness / ASR / refusal rate on all the 30 instructions in Bombs or Explosives scenario." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 358, + 180, + 391, + 213 + ], + "blocks": [ + { + "bbox": [ + 358, + 180, + 391, + 213 + ], + "lines": [ + { + "bbox": [ + 358, + 180, + 391, + 213 + ], + "spans": [ + { + "bbox": [ + 358, + 180, + 391, + 213 + ], + "type": "image", + "image_path": "5bf34b6d4542f672a485899f08f919691747202e4713a35814001a85c2f827a8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 394, + 180, + 425, + 213 + ], + "blocks": [ + { + "bbox": [ + 394, + 180, + 425, + 213 + ], + "lines": [ + { + "bbox": [ + 394, + 180, + 425, + 213 + ], + "spans": [ + { + "bbox": [ + 394, + 180, + 425, + 213 + ], + "type": "image", + "image_path": "cf56f2b5c8f579db16e2375f6f12e75c1cb43fb5a2608bf7cfd64cefad7f8df6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 429, + 180, + 462, + 213 + ], + "blocks": [ + { + "bbox": [ + 429, + 180, + 462, + 213 + ], + "lines": [ + { + "bbox": [ + 429, + 180, + 462, + 213 + ], + "spans": [ + { + "bbox": [ + 429, + 180, + 462, + 213 + ], + "type": "image", + "image_path": "d5e530947f53f879be50941b1ffd866a200bd0a40a666333f688f948cd2d70f8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 466, + 180, + 498, + 213 + ], + "blocks": [ + { + "bbox": [ + 466, + 180, + 498, + 213 + ], + "lines": [ + { + "bbox": [ + 466, + 180, + 498, + 213 + ], + "spans": [ + { + "bbox": [ + 466, + 180, + 498, + 213 + ], + "type": "image", + "image_path": "dfd721c65faca1b6d529825ab399f71f52204d60de6189f91a37c776b095eb4e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 503, + 180, + 534, + 213 + ], + "blocks": [ + { + "bbox": [ + 503, + 180, + 534, + 213 + ], + "lines": [ + { + "bbox": [ + 503, + 180, + 534, + 213 + ], + "spans": [ + { + "bbox": [ + 503, + 180, + 534, + 213 + ], + "type": "image", + "image_path": "aabca5b31b4ddab78a761fb3c0093fc244d8b70a72258ef314f11b22e13ac21d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 328, + 287, + 550, + 433 + ], + "blocks": [ + { + "bbox": [ + 55, + 349, + 295, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 349, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 295, + 381 + ], + "type": "text", + "content": "Figure 3. Effect of attacking with (a) various text-mixing techniques on GPT-4 and (b) image-transformation techniques on GPT-4V in all the attack scenarios." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 287, + 550, + 433 + ], + "lines": [ + { + "bbox": [ + 328, + 287, + 550, + 433 + ], + "spans": [ + { + "bbox": [ + 328, + 287, + 550, + 433 + ], + "type": "image", + "image_path": "d28407166ba7c3ddc69e68bef15f58270f899a4b4ba158665910c0979032669f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 434, + 563, + 454 + ], + "lines": [ + { + "bbox": [ + 315, + 434, + 563, + 454 + ], + "spans": [ + { + "bbox": [ + 315, + 434, + 563, + 454 + ], + "type": "text", + "content": "(a) Effect of semantically various auxiliary images when mixed with the target harmful image (\"bomb\")." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 331, + 457, + 548, + 586 + ], + "blocks": [ + { + "bbox": [ + 331, + 457, + 548, + 586 + ], + "lines": [ + { + "bbox": [ + 331, + 457, + 548, + 586 + ], + "spans": [ + { + "bbox": [ + 331, + 457, + 548, + 586 + ], + "type": "image", + "image_path": "95bbfab247648a56ecf6b8c76ed37494c3f497738c7c383204e0ca996f671380.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 590, + 563, + 610 + ], + "lines": [ + { + "bbox": [ + 313, + 590, + 563, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 563, + 610 + ], + "type": "text", + "content": "(b) Ablation for the visual type of auxiliary image, either using typography image or realistic image." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 619, + 555, + 651 + ], + "lines": [ + { + "bbox": [ + 313, + 619, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 555, + 651 + ], + "type": "text", + "content": "Figure 5. Effect of auxiliary images depending on the similarity with the target harmful image (a) and the visual image types (b). Red dotted line denotes the harmfulness of the vanilla attack." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "matching [45] which verifies whether the attacked response has one of the refusal phrases (e.g., \"I'm sorry\"). Generally, the vanilla case (i.e., " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ") without mixing the auxiliary image on the original harmful image showed signif" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "29943" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 87, + 526, + 149 + ], + "blocks": [ + { + "bbox": [ + 91, + 71, + 518, + 83 + ], + "lines": [ + { + "bbox": [ + 91, + 71, + 518, + 83 + ], + "spans": [ + { + "bbox": [ + 91, + 71, + 518, + 83 + ], + "type": "text", + "content": "Table 3. Effect of attack methods for jailbreaking against the system prompt defense in Bombs or Explosives scenario." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 87, + 526, + 149 + ], + "lines": [ + { + "bbox": [ + 82, + 87, + 526, + 149 + ], + "spans": [ + { + "bbox": [ + 82, + 87, + 526, + 149 + ], + "type": "table", + "html": "
System Prompt DefenseVanillaFigStep 2023FigStep-Pro 2023HADES 2024JOOD
HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑HF ↑ASR ↑
×00%00%0.923%00%7.163%
✓00%00%0.813%00%4.360%
", + "image_path": "ab6884731b0b37a51f929f9476d167c9d9968a8e5c8f67788f828e4a606885cd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 161, + 297, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 161, + 297, + 329 + ], + "spans": [ + { + "bbox": [ + 54, + 161, + 297, + 329 + ], + "type": "text", + "content": "icently low ASR and harmfulness score with high refusal rate possibly due to the safety alignment training with self feedback [28] of GPT-4V on these obviously harmful images. However, when an auxiliary image is mixed (i.e., " + }, + { + "bbox": [ + 54, + 161, + 297, + 329 + ], + "type": "inline_equation", + "content": "0 < \\alpha < 1" + }, + { + "bbox": [ + 54, + 161, + 297, + 329 + ], + "type": "text", + "content": "), the refusal rate significantly decreases while the ASR and harmfulness score are largely increased. This indicates the essential role of our attack to mitigate the evasive refusal of GPT-4V with regard to the harmful requests and further elicit the harmful responses via presenting the OOD-ifying inputs generated by mixup, which were possibly not seen during the safety alignment training. When " + }, + { + "bbox": [ + 54, + 161, + 297, + 329 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 161, + 297, + 329 + ], + "type": "text", + "content": " reaches 1, the harmful image is entirely substituted by the auxiliary image (e.g., apple), losing the harmful semantic and hence reducing both ASR and harmfulness scores." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 342, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 582 + ], + "type": "text", + "content": "Effect of auxiliary images. In Figure 5, we analyze the effect of mixing the harmful image with the auxiliary images based on their similarity and visual types. Specifically, in Figure 5a, we analyze the effect of mixing with semantically various auxiliary images by comparing the cosine similarity with the harmful \"bomb\" image. Blue dots / green dots / red dots represent the auxiliary images that are dissimilar / moderately related / highly similar to the target harmful image, respectively. The result shows a strong negative correlation between the similarity and the harmfulness of the model response. Mixing with highly similar images containing similar unsafe objects such as a \"Molotov cocktail\" generates less harmful responses which only refuse to provide answers with regard to both the original harmful object (e.g., bomb) and another harmful object (e.g., Molotov cocktail). Meanwhile, mixing with moderately related or dissimilar auxiliary images such as \"softwood\" generates harmful responses containing a detailed description regarding the original harmful object (see Appendix J for detailed response examples)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "In Figure 5b, we conduct another analysis on the effect of mixing with either a typographic image that displays the text of the auxiliary word or a realistic image containing the visual scene related to the auxiliary word. In all the attack scenarios, using typographic auxiliary images for mixing harmful images generally amplifies the harmfulness of the output responses compared to using real images. While Gong et al. [9] empirically observed that existing MLLMs already exhibit vulnerability on the typographic image input itself, this result suggests that the safety guardrails of MLLMs can be more effectively neutralized when these ty" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 161, + 555, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 555, + 185 + ], + "type": "text", + "content": "pographic images are further leveraged for our proposed image-mixing based attacks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 208, + 553, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 553, + 352 + ], + "type": "text", + "content": "Jailbreak against defense methods. In Table 3, we investigate the effect of JOOD against system prompt-based defense [9, 25] where defensive textual guidance is given as input system prompt instructing the model to remain vigilant and refrain from responding to harmful textual or visual queries that may violate AI safety policies (see Appendix G for the detailed prompt). The results show that even after deploying the safety-aware system prompt, JOOD still achieves significantly higher harmfulness and ASR scores compared to the baseline attack methods. Notably, JOOD only degraded ASR by " + }, + { + "bbox": [ + 313, + 208, + 553, + 352 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 208, + 553, + 352 + ], + "type": "text", + "content": " after applying system-prompt-based defense, while FigStep-Pro largely degraded by " + }, + { + "bbox": [ + 313, + 208, + 553, + 352 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 208, + 553, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 354, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 498 + ], + "type": "text", + "content": "In Appendix G, we further evaluate jailbreak performance of JOOD against a more sophisticated defense, AdaShield [35], which adaptively retrieves the optimal defense prompt for each attack query. Even with this adaptive defense, our JOOD successfully jailbreaks over half of the attack queries with a significantly lower refusal rate, while all the baselines hardly succeed to jailbreak. The robustness of JOOD possibly originates from the ambiguity and uncertainty of the input harmful image induced by transform techniques such as mixup. This highlights the need for further exploration of safeguard mechanisms when dealing with such ambiguous and potentially harmful inputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 512, + 388, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 388, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 388, + 525 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 715 + ], + "type": "text", + "content": "We investigate the unrevealed vulnerability of the safety alignment within existing LLMs and MLLMs when the out-of-distribution inputs are given as the input. Exploiting this vulnerability, we devise a new jailbreak strategy by generating OOD-ifying inputs with off-the-shelf data transformation techniques. We observe that even simple transformations such as mixup can fabricate OOD-ifying inputs that induce a high level of uncertainty for the models, successfully jailbreaking the state-of-the-art LLMs and MLLMs. By the extensive experiments on various jailbreak scenarios and rigorous ablations, we demonstrate that this simple yet straightforward OOD-ifying strategy can pose a significant threat to the recent LLMs and MLLMs, underscoring the necessity for further research and development to enhance safety alignment against these OOD-ifying inputs." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "text", + "content": "29944" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was partly supported by National Research Foundation of Korea (NRF) grant (No.RS-2023-00209060, A Study on Optimization and Network Interpretation Method for Large-Scale Machine Learning) funded by the Korea government (MSIT). This work was also supported by KAIST-NAVER Hypercreative AI Center." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 194, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 194, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 237 + ], + "type": "text", + "content": "[1] Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. Jailbreaking leading safety-aligned llms with simple adaptive attacks. arXiv preprint arXiv:2404.02151, 2024.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 262 + ], + "type": "text", + "content": "[2] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 263, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 295, + 306 + ], + "type": "text", + "content": "[3] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J Pappas, and Eric Wong. Jailbreaking black box large language models in twenty queries. arXiv preprint arXiv:2310.08419, 2023. 1, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "type": "text", + "content": "[4] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 354, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 296, + 407 + ], + "type": "text", + "content": "[5] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 3, 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 410, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 295, + 453 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 455, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 520 + ], + "type": "text", + "content": "[7] Srujan Kumar Enaganti, Lila Kari, Timothy Ng, and Zihao Wang. Word blending in formal languages: The brangelina effect. In Unconventional Computation and Natural Computation: 17th International Conference, UCNC 2018, Fontainebleau, France, June 25-29, 2018, Proceedings 17, pages 72-85. Springer, 2018. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 522, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 294, + 575 + ], + "type": "text", + "content": "[8] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1-79, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 578, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 294, + 632 + ], + "type": "text", + "content": "[9] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023. 2, 3, 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 276, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 276, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 276, + 644 + ], + "type": "text", + "content": "[10] Google. Bard - chat based ai tool from google, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[11] Divij Handa, Advait Chirmule, Bimal Gajera, and Chitta Baral. Jailbreaking proprietary large language models using word substitution cipher. arXiv preprint arXiv:2402.10601, 2024.3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "text", + "content": "[13] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Lmbased input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023.5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[14] Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "text", + "content": "[15] Joonhyun Jeong, Sungmin Cha, Jongwon Choi, Sangwoo Yun, Taesup Moon, and Youngjoon Yoo. Observations on k-image expansion of image-mixing augmentation. IEEE Access, 11:16631-16643, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[16] JangHyun Kim, Wonho Choo, Hosan Jeong, and Hyun Oh Song. Co-mixup: Saliency guided joint mixup with supermodular diversity. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "text", + "content": "[17] Jang-Hyun Kim, Wonho Choo, and Hyun Oh Song. Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In International Conference on Machine Learning, pages 5275-5285. PMLR, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "text", + "content": "[18] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal generation. arXiv preprint arXiv:2301.13823, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "type": "text", + "content": "[19] Raz Lapid, Ron Langberg, and Moshe Sipper. Open sesame! universal black box jailbreaking of large language models. arXiv preprint arXiv:2309.01446, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 400, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 454 + ], + "type": "text", + "content": "[20] Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, et al. Solving quantitative reasoning problems with language models, 2022. URL https://arxiv.org/abs/2206.14858, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "text", + "content": "[21] Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. Textbugger: Generating adversarial text against real-world applications. arXiv preprint arXiv:1812.05271, 2018. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 491, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 553, + 543 + ], + "type": "text", + "content": "[22] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and JiRong Wen. Images are achilles' heel of alignment: Exploiting visual vulnerabilities for jailbreaking multimodal large language models. arXiv preprint arXiv:2403.09792, 2024. 1, 3, 5, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 546, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 553, + 588 + ], + "type": "text", + "content": "[23] Zeyi Liao and Huan Sun. Amplegcg: Learning a universal and transferable generative model of adversarial suffixes for jailbreaking both open and closed llms. arXiv preprint arXiv:2404.07921, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 591, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 622 + ], + "type": "text", + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[25] Siyuan Ma, Weidi Luo, Yu Wang, Xiaogeng Liu, Muhao Chen, Bo Li, and Chaowei Xiao. Visual-roleplay: Universal jailbreak attack on multimodal large language models via role-playing image character. arXiv preprint arXiv:2405.20773, 2024. 3, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943, 2021. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "text", + "content": "29945" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[27] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model. arXiv preprint arXiv:2402.02309, 2024. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 139 + ], + "type": "text", + "content": "[28] OpenAI. Gpt-4 technical report. Technical report, 2023. 1, 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "text", + "content": "[29] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[30] Yikang Pan, Liangming Pan, Wenhu Chen, Preslav Nakov, Min-Yen Kan, and William Yang Wang. On the risk of misinformation pollution with large language models. arXiv preprint arXiv:2305.13661, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[31] Jie Qin, Jiemin Fang, Qian Zhang, Wenyu Liu, Xingang Wang, and Xinggang Wang. Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101, 2020.4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 363 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "text", + "content": "[33] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. In The Twelfth International Conference on Learning Representations, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[34] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "text", + "content": "[35] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. In European Conference on Computer Vision, pages 77-94. Springer, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 600 + ], + "type": "text", + "content": "[36] Jason Wei and Kai Zou. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6382-6388, Hong Kong, China, 2019. Association for Computational Linguistics. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[37] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 688 + ], + "type": "text", + "content": "[38] Zeming Wei, Yifei Wang, and Yisen Wang. Jailbreak and guard aligned language models with only few in-context demonstrations. arXiv preprint arXiv:2310.06387, 2023. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 374 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[40] Zheng-Xin Yong, Cristina Menghini, and Stephen H Bach. Low-resource languages jailbreak gpt-4. arXiv preprint arXiv:2310.02446, 2023. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[41] Youliang Yuan, Wenxiang Jiao, Wenxuan Wang, Jen tse Huang, Pinjia He, Shuming Shi, and Zhaopeng Tu. GPT-4 is too smart to be safe: Stealthy chat with LLMs via cipher. In The Twelfth International Conference on Learning Representations, 2024. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 249 + ], + "type": "text", + "content": "[42] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE International Conference on Computer Vision, pages 6023-6032, 2019. 2, 3, 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "text", + "content": "[43] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 3, 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[44] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[45] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 1, 3, 5, 7" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 763 + ], + "type": "text", + "content": "29946" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/2c24c8f1-1822-42ba-8659-5feab5b2d726_content_list.json b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/2c24c8f1-1822-42ba-8659-5feab5b2d726_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3db0f1c8d64ed9bac5f896436d8492f85d1c7c4f --- /dev/null +++ b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/2c24c8f1-1822-42ba-8659-5feab5b2d726_content_list.json @@ -0,0 +1,1490 @@ +[ + { + "type": "text", + "text": "Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control", + "text_level": 1, + "bbox": [ + 145, + 130, + 825, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Basim Azam Naveed Akhtar", + "bbox": [ + 359, + 212, + 611, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computing and Information Systems, The University of Melbourne, Australia", + "bbox": [ + 135, + 233, + 833, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{basim.azam, naveed.akhtar1}@unimelb.edu.au", + "bbox": [ + 292, + 257, + 674, + 271 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e4d90a3489cd24f005e5fd2733cdc53fd11648af1460046f84cdd1e4580ffe8.jpg", + "image_caption": [ + "Figure 1. Our unique plug-and-play interpretable approach simultaneously controls a range of concepts for responsible and fair image generation with text-to-image pipelines. Our method enables control over both text embedding space and latent diffusion space. Shown examples compare unfair and unsafe generation for the given prompts by the Stable Diffusion (top), along with their responsible counterparts resulting from our approach influencing the text encoder (middle) and the diffusion model (bottom). We control individual concepts for diverse/safe generation in these examples while modeling them in continuous composite responsible semantic spaces." + ], + "image_footnote": [], + "bbox": [ + 76, + 300, + 888, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 657, + 313, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ethical issues around text-to-image (T2I) models demand a comprehensive control over the generative content. Existing techniques addressing these issues for responsible T2I models aim for the generated content to be fair and safe (non-violent/explicit). However, these methods remain bounded to handling the facets of responsibility concepts individually, while also lacking in interpretability. Moreover, they often require alteration to the original model, which compromises the model performance. In this work, we propose a unique technique to enable responsible T2I generation by simultaneously accounting for an extensive range of concepts for fair and safe content generation in a scalable manner. The key idea is to distill the target T2I pipeline with an external plug-and-play mechanism that learns an inter", + "bbox": [ + 75, + 689, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pretable composite responsible space for the desired concepts, conditioned on the target T2I pipeline. We use knowledge distillation and concept whitening to enable this. At inference, the learned space is utilized to modulate the generative content. A typical T2I pipeline presents two plug-in points for our approach, namely; the text embedding space and the diffusion model latent space. We develop modules for both points and show the effectiveness of our approach with a range of strong results. Our code can be accessed at https://basim-azam.github.io/responsiblediffusion/", + "bbox": [ + 496, + 659, + 893, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 830, + 632, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rapid advances in text-to-image (T2I) generative pipeline are revolutionizing numerous vision applications, offering unprecedented convenience for synthesizing high quality", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2976", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "visual content using textual descriptions [34, 32, 30, 19, 17]. Unfortunately, this convenience coupled with the generated content realism offered by publicly accessible T2I models, e.g., Stable Diffusion [34], DALL-E [32], Imagen [36], also has undesired ethical implications. Confronting the core values of fair and safe (non-violent/explicit) development and deployment of Artificial Intelligence, these models lack in producing generative content responsibly [15, 51].", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The reasons of irresponsible content generation by contemporary T2I models lie deep in their training process, which must use large volumes of (mostly uncurated) data. Issues like inappropriate content and harmful stereotyping lurking in the data are passed on to the generative models through training [2, 42], which cause problems after their deployment. Hence, it is critical to develop methodologies to explicitly control generative models from producing irresponsible content. Addressing that, several techniques have emerged recently, encompassing approaches like input prompt filtering [48, 2], post-hoc content moderation [47, 43, 22], machine unlearning [45, 29, 18, 49], and model editing [14, 46]. Though effective, these approaches face limitations like requiring human-intervention, ad-hoc methodology, and selective dealing of different facets of the concepts defining 'responsible generation'.", + "bbox": [ + 75, + 229, + 472, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another contributing factor to the ethical concerns related to T2I models is our current lack of wholesome understanding of the latent space of the diffusion models. Researchers have already started exploring interpretability [16, 50] and manipulation of diffusion model latent spaces to mitigate unintended T2I model behavior [14, 23]. Controlling the latent space holds promise when it is possible to disentangle semantic concepts in it. One recent inspiring approach in this direction [23] identifies vectors in the latent diffusion space in the directions of individual semantic concepts, which can be used to restrict model outputs. However, it is arguable that such a discretized treatment of the concepts within the diffusion latent space is restrictive. It is also intrinsically limited to dealing with diverse facets of the concepts individually. Overall, it still remains a widely open challenge for the research community to comprehensively embed the multi-faceted concepts related to fair and safe image generation in T2I pipelines.", + "bbox": [ + 75, + 474, + 472, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we address this by presenting a unique scalable approach to extensively incorporate fair and safe generative abilities in a T2I pipeline in a plug-and-play manner. Our key insight is that a T2I model can be distilled for a range of user-desired responsibility concepts, conditioned on the original model. We perform this distillation by inducing a student model that treats the concepts related to fairness and safety in a continuous space. Within this space, we further leverage concept whitening [9] to disentangle the representations of concepts for a better ultimate", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "control on the generative image modulation. The underlying framework of our technique is generic, in that it is applicable to any representation space encoding semantic concepts. Hence, we explore its application to both text encoder embedding space and diffusion model latent space in the T2I pipeline. Methods for both variants address their unique challenges, but prove equally effective - see Fig. 1. The key contributions of this work are summarized below.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a unique plug-and-play method for responsible image generation with T2I pipeline that enables comprehensive incorporation of fairness and safety concepts while modeling them under a distilled continuous space conditioned on the generative model.", + "- We tailor our method as a text embedding plug-in, and diffusion latent plug-in while also leveraging concept whitening to ensure precise interpretable control over the content. Both plug-ins are employed in our approach.", + "- With extensive experiments, we not only demonstrate state-of-the-art or comparable performance in fair and safe image generation, but also show other unique interesting properties of our approach." + ], + "bbox": [ + 500, + 214, + 890, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 422, + 650, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image (T2I) generation has significantly expanded the capabilities of generative AI, enabling high-quality outputs using textual descriptions [34, 32, 30, 19, 17]. While T2I methods like Stable Diffusion [34], DALL-E [32], GLIDE [27], and Imagen [36] have achieved impressive realism and adaptation in numerous applications, they also face several challenges. One of them is the concern about their irresponsible and unethical content generation [15, 2].", + "bbox": [ + 496, + 446, + 890, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contemporary T2I models are known to generate inappropriate images, including nude and violent content, and they are also susceptible to social and cultural biases persistently found in the public domain datasets used to train these models [10, 24, 41]. The increased popularity of T2I models has driven research into understanding and mitigating biases in their outputs, particularly focusing on safety filters [33], semantic space organization [7], and tendencies toward stereotypical portrayals [5, 10, 26]. Studies show that these models, often guided by biased systems, e.g., CLIP [31, 1], may memorize sensitive data [8, 40], which, combined with unfiltered web-based datasets containing harmful content [6, 39, 4, 3], propagates unintended biases.", + "bbox": [ + 496, + 569, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Research to address these concerns can be divided into three main directions; namely, preemptive content filtering, model-level adjustments, and post-hoc moderation. The preemptive filtering approaches involve controlling input prompts to reduce biases before they influence image generation. For example, prompt-based filters [48, 2] screen out language associated with inappropriate content [19, 30]. In the model-level adjustment, unlearning techniques [45, 29, 18, 49] aim to remove sensitive concepts directly from the", + "bbox": [ + 496, + 765, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2977", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model's learned representations. Similarly, model editing methods target specific parameters of the models to limit the generation of harmful content without extensive re-training [28, 14, 46]. Although useful, these methods can negatively impact the original model performance. They can also become computationally intensive as the harmful concepts increase [49]. Post-hoc moderation methods [47, 43, 22] involve both automated and manual processes like GuardT2I [47] and context-based filters [43, 22]. These technique only partially address model biases and irresponsible generation, and may easily overlook implicit negative concept associations in the model.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recently, research on diffusion model latent space manipulation and interpretability is gaining prominence in regards to controlling the T2I outputs. Studies such as [16, 50] explore the semantic organization within the latent spaces to understand and intervene in the generation process. Li et al. [23] investigated interpretable directions of target concepts in the latent semantic space. These methods remain effective, however; their underlying discretized treatment of the individual semantic concepts limits their scalability. In this work, we address safe and fair modulation of both embedding and latent spaces without explicitly discretizing the directions, and focusing on their continuous space instead. Our distinctive plug-and-play approach learns an interpretable space for a range of responsible concepts while conditioning it on the T2I pipeline.", + "bbox": [ + 75, + 272, + 472, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 76, + 503, + 269, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present a unique approach for responsible image generation with T2I pipeline that develops add-on modules usable in a plug-and-play manner. The modules are neural models induced by distilling the target T2I model for a scalable list of responsible semantic concepts addressing fair and safe (non-violent/explicit) image generation. We first formally present the problem as perceived in this work.", + "bbox": [ + 75, + 525, + 468, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formalization", + "text_level": 1, + "bbox": [ + 76, + 636, + 292, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A text-to-image model, $\\Psi(t): t \\to \\bar{I}$ , maps a textual prompt $t \\in \\mathcal{T}$ to an image $\\bar{I} \\in \\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times 3}$ , where $\\mathcal{T}$ and $\\mathcal{M}$ are the sets of possible text prompts and images. Depending on the data used to train the T2I model, the image $\\bar{I}$ may belong to an image subset $\\overline{\\mathcal{R}}_{\\mathcal{A}_X} \\subset \\mathcal{M}$ . Conditioned on $\\mathcal{A}_X$ , $\\overline{\\mathcal{R}}_{\\mathcal{A}_X}$ denotes the set of irresponsible/inappropriate images as identified by a collection of high-level cultural/social/societal aspects in set $\\mathcal{A}_X$ . In our settings, $\\mathcal{A}_X$ is a subset of the aspects enlisted in Eq. (1). Notice that, $\\mathcal{A}_X \\subseteq \\mathcal{A}_{\\mathrm{resp}}$ implies that $\\overline{\\mathcal{R}}_{\\mathcal{A}_X}$ accounts for the fact that $\\bar{I}$ can be inappropriate along more than one aspect - an issue often ignored in the existing literature.", + "bbox": [ + 75, + 654, + 472, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {\\text {r e s p}} = \\left\\{\\mathcal {A} _ {\\text {a g e}}, \\mathcal {A} _ {\\text {g e n d e r}}, \\mathcal {A} _ {\\text {r a c e}}, \\mathcal {A} _ {\\text {s a f e}} \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 845, + 468, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We form $\\mathcal{A}_{\\mathrm{resp}}$ considering the aspects of responsible image generation accounted in the existing works [23, 38].", + "bbox": [ + 75, + 869, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nevertheless, we keep $\\mathcal{A}_{\\mathrm{resp}}$ extendable as required. In this work, we define each of the considered aspects using high-level attributes listed below, which also aligns with the existing literature [23, 38]: $\\mathcal{A}_{\\mathrm{gender}} = \\{\\mathbf{a}_{\\mathrm{male}}, \\mathbf{a}_{\\mathrm{female}}\\}$ , $\\mathcal{A}_{\\mathrm{race}} = \\{\\mathbf{a}_{\\mathrm{white}}, \\mathbf{a}_{\\mathrm{asian}}, \\mathbf{a}_{\\mathrm{black}}\\}$ , $\\mathcal{A}_{\\mathrm{age}} = \\{\\mathbf{a}_{\\mathrm{young}}, \\mathbf{a}_{\\mathrm{middle-aged}}, \\mathbf{a}_{\\mathrm{elderly}}\\}$ , and $\\mathcal{A}_{\\mathrm{safe}} = \\{\\mathbf{a}_{\\mathrm{harassment}}, \\mathbf{a}_{\\mathrm{sexual}}, \\mathbf{a}_{\\mathrm{violence}}\\}$ .", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is worth emphasizing that the attribute sets noted above can also be extended if required. Our approach is not constrained by these sets, except for further training required to account for any additional aspect. Given an $\\mathcal{A}_X$ - as defined by the model user, our objective is to alter the mapping of the T2I model to $\\Psi_{\\mathrm{resp}}(t):t\\to I$ such that $I\\in \\mathcal{R}_{\\mathcal{A}_X}\\subseteq \\mathcal{M}$ and $\\overline{\\mathcal{R}}_{\\mathcal{A}_X}\\cap \\mathcal{R}_{\\mathcal{A}_X} = \\emptyset$ . Here, $\\Psi_{\\mathrm{resp}}$ is the responsible variant of $\\Psi$ with an added functionality that restricts the outputs of $\\Psi$ to $\\mathcal{A}_X$ -constrained responsible image space $\\mathcal{R}_{\\mathcal{A}_X}$ .", + "bbox": [ + 496, + 183, + 893, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Framework Blueprint", + "text_level": 1, + "bbox": [ + 498, + 321, + 705, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conceptually, a T2I model $\\Psi$ is a hierarchical function $\\Psi(t) = \\mathcal{D}(\\mathcal{E}(t))$ , where $\\mathcal{E}(.)$ is a text-encoder - typically CLIP [20], and $\\mathcal{D}(.)$ is a diffusion model [34]. In this work, we first seek $\\Psi_{\\mathrm{resp}}$ mentioned in § 3.1 by devising add-on functionalities for responsible image generation while targeting $\\mathcal{E}(.)$ and $\\mathcal{D}(.)$ individually. In other words, we aim for $\\mathcal{E}_{\\mathrm{resp}}(.)$ and $\\mathcal{D}_{\\mathrm{resp}}(.)$ . Both of these potentially 'enhanced' sub-models can also be combined for $\\Psi_{\\mathrm{resp}}$ . We target the text encoder and diffusion model individually because in T2I modeling, both of them are known to effectively encode high-level semantic information in their embedding and latent spaces [25, 16, 50]. This provides us the possibility to plug-in our intended add-on modules to any of these models for responsible T2I generation.", + "bbox": [ + 496, + 342, + 890, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For both of the desired $\\mathcal{E}_{\\mathrm{resp}}(.)$ and $\\mathcal{D}_{\\mathrm{resp}}(.)$ , the framework underlying our approach remains the same at the conceptual level. We keep the target sub-model $(\\mathcal{E}$ or $\\mathcal{D})$ in the T2I pipeline frozen, and learn an add-on module (respectively, $\\psi_{\\mathcal{E}}$ or $\\psi_{\\mathcal{D}}$ ) by distilling the target sub-model with knowledge distillation [25]. The knowledge is distilled by conditioning the process on the set $\\mathcal{A}_X$ which covers the aspects along which the model needs to be made responsible. Moreover, to ensure that the distilled concepts are disentangled as much as possible, we apply concept whitening [9] to them. The eventual T2I signal gets modulated by a simple addition of the signal from the add-on modules for responsible generation.", + "bbox": [ + 496, + 555, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We explain the exact procedure for obtaining $\\psi_{\\mathcal{E}}$ for $\\mathcal{E}_{\\mathrm{resp}}(.)$ , and $\\psi_{\\mathcal{D}}$ for $\\mathcal{D}_{\\mathrm{resp}}(.)$ in § 3.3 and 3.4, respectively. Combining them for $\\Psi_{\\mathrm{resp}}$ is explained in § 3.5.", + "bbox": [ + 496, + 752, + 890, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Responsible Interpretable Embeddings", + "text_level": 1, + "bbox": [ + 498, + 804, + 834, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our add-on module $\\psi_{\\mathcal{E}}$ for $\\mathcal{E}_{\\mathrm{resp}}(.)$ is marked by not only responsible generation but also interpretability of the embedding space distilled from the source $\\mathcal{E}(.)$ . Hence, we term the devised module as RICE - Responsible and Interpretable CLIP Embedding. The RICE module treats the original text", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2978", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0ad5565c7e2d5f97ad49db789565b091a98ecd23525e01519782cf8f66f82b4f.jpg", + "image_caption": [ + "(a) RICE Embeddings Space" + ], + "image_footnote": [], + "bbox": [ + 89, + 97, + 331, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/32539d5ad9e2a463c0ce89108040edc9233992ee9241eff278119b1fe8e0286c.jpg", + "image_caption": [ + "(b) RICE Controlled T2I Outputs" + ], + "image_footnote": [], + "bbox": [ + 86, + 270, + 331, + 421 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5160f8954395ff8e10fc3329d7c7e8c03a23315f0b2fa650e7d2403f0aa37bcd.jpg", + "image_caption": [ + "Figure 2. (a) Illustration of responsible concept space learned by the RICE module, aiming to generate $z_{\\mathrm{clip}}^{\\mathrm{final}}$ embeddings along the responsible aspects in $\\mathcal{A}_{\\mathrm{X}}$ . Interconnected fairness related subspaces for race, gender, and age are illustrated. (b) Example outputs of RICE controlled T2I process, demonstrating a range of modulation along the concepts, which is possible due to the multi-facet control enabled by the RICE embedding space. (c) Depiction of the distillation process used for the RIIDL module, where the student (RIIDL) learns from the teacher (Diffusion Model) over $\\tau$ time-steps. At each step, the loss - see Eq. (6) - aligns the RIIDL module with the aspects of $\\mathcal{A}_{\\mathrm{X}}$ . (d) At inference, textual embeddings get fed into RIIDL. We illustrate control over $\\mathcal{A}_{\\mathrm{age}}$ . The latent vectors from RIIDL are injected into the T2I Diffusion Model noisy latents. We explore this injection at varying time-step (differentiated vertically in the figure). The best control is observed in the early stages ( $\\tau = 0$ to $30\\%$ ). Additional examples provided in the supplementary material. (e) Our method allows a scalable control over all concepts covered in $\\mathcal{A}_{\\mathrm{X}}$ in a composite manner." + ], + "image_footnote": [], + "bbox": [ + 334, + 95, + 888, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "encoder in the T2I pipeline as a teacher and distills it for the concepts in $\\mathcal{A}_X$ by training $\\psi_{\\mathcal{E}}$ while minimizing the Expected value of the following per-batch loss:", + "bbox": [ + 75, + 599, + 468, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K D - c l i p}} = \\frac {1}{| B |} \\sum_ {k = 1} ^ {| B |} \\left\\| z _ {\\mathrm {c l i p}, k} ^ {\\mathrm {T}} - z _ {\\mathrm {c l i p}, k} ^ {\\text {d i s t i l l}} \\right\\| ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 657, + 468, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B$ denotes the batch, $z_{\\mathrm{clip},k}^{\\mathrm{T}}$ and $z_{\\mathrm{clip},k}^{\\mathrm{distill}}$ are the teacher and distilled student embeddings for the $k^{\\mathrm{th}}$ sample, respectively. The $z_{\\mathrm{clip}}^{\\mathrm{T}}$ embeddings are computed using an automated prompt generation following $\\mathcal{A}_X$ . To keep the flow of discussion, we provide details about that process and the architecture of $\\psi_{\\mathcal{E}}$ in the supplementary material.", + "bbox": [ + 75, + 715, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At this stage, $\\mathbf{z}_{\\mathrm{clip}}^{\\mathrm{distll}} \\in \\mathbb{R}^{d \\times n}$ are the embedding representations conditioned on $\\mathcal{A}_X$ . We additionally apply concept whitening [9] to further decorrelate these distilled embeddings using the following transform:", + "bbox": [ + 75, + 809, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {\\mathrm {c l i p}} ^ {\\mathrm {z c a}} = \\phi (\\mathbf {z} _ {\\mathrm {c l i p}} ^ {\\mathrm {d i s t i l l}}) = \\mathcal {W} (\\mathbf {z} _ {\\mathrm {c l i p}} ^ {\\mathrm {d i s t i l l}} - \\mu), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu = \\frac{1}{n}\\sum_{i=1}^{n}z_{i}$ , and $\\mathcal{W}_{d\\times d}$ is the whitening matrix that obeys $\\mathcal{W}^T\\mathcal{W} = \\Sigma^{-1}$ . Here, $\\Sigma_{d\\times d} = \\frac{1}{n}(z_{\\mathrm{clip}}^{\\mathrm{distill}} - \\mu\\mathbf{1}^{\\intercal})(z_{\\mathrm{clip}}^{\\mathrm{distill}} - \\mu\\mathbf{1}^{\\intercal})^{\\intercal}$ is the covariance matrix. The final embeddings for responsible generation get computed as:", + "bbox": [ + 498, + 598, + 890, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {\\text {c l i p}} ^ {\\text {r e s p}} = \\alpha \\cdot \\mathbf {z} _ {\\text {c l i p}} ^ {\\text {d i s t i l l}} + (1 - \\alpha) \\cdot \\mathbf {z} _ {\\text {c l i p}} ^ {\\text {z c a}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 674, + 890, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is kept for balancing. At inference, the embedding $z_{\\mathrm{clip}}$ of the original prompt gets modulated as", + "bbox": [ + 500, + 705, + 890, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\text {c l i p}} ^ {\\text {f i n a l}} = z _ {\\text {c l i p}} + \\sum_ {k} \\gamma_ {k} \\mathbf {z} _ {k, \\text {c l i p}} ^ {\\text {r e s p}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 751, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma_{k}$ provides control over the influence of each concept on the embedding. With this, the embedding of a prompt can be simultaneously modulated to multiple responsible concepts represented in the continuous representation space of $\\psi_{\\mathcal{E}}$ . Fig. 2a illustrates this notion, with Fig. 2b showing the RICE impact on varying the eventual T2I output along different responsible concepts.", + "bbox": [ + 496, + 796, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2979", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Responsible Interpretable Latents", + "text_level": 1, + "bbox": [ + 76, + 90, + 375, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In regards to $\\mathcal{D}_{\\mathrm{resp}}(.)$ , we follow a similar process as adopted in § 3.3 for $\\mathcal{E}_{\\mathrm{resp}}(.)$ , tailoring the former for the diffusion model component $\\mathcal{D}$ of the T2I pipeline. Our add-on module $\\psi_{D}$ , again a neural model, gets trained using the intermediate latent representations of $\\mathcal{D}$ , distilled by conditioning it on $\\mathcal{A}_X$ , hence; we term it RIIDL - Responsible Interpretable Intermediate Diffusion Latents.", + "bbox": [ + 75, + 106, + 467, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The latent representation of a diffusion model gets updated for each time-step $(\\tau)$ of a Markov process denoising the image previously corrupted in the forward diffusion process. For brevity, we refrain from discussing minutiae of diffusion modeling - interested readers are referred to [12]. Here, we focus on the key idea of inducing our plug-in $\\psi_{D}$ for the process. To train $\\psi_{D}$ , we define the following loss", + "bbox": [ + 75, + 212, + 467, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K D - u n e t}} = \\mathbb {E} _ {\\tau \\sim \\mathcal {U} [ 0, 1 ]} \\left[ \\omega \\left(\\lambda_ {\\tau}\\right) \\left\\| z _ {\\text {u n e t}} ^ {\\mathrm {T}, \\tau} - z _ {\\text {u n e t}} ^ {\\text {d i s t i l l}, \\tau} \\right\\| ^ {2} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 329, + 467, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau \\sim \\mathcal{U}[0,1]$ refers to uniform sampling of time-steps, $z_{\\mathrm{unet}}^{\\mathrm{T},\\tau}$ and $z_{\\mathrm{unet}}^{\\mathrm{distill},\\tau}$ represent the latent representations of the teacher and student models at time-step $\\tau$ , $\\lambda_{\\tau}$ reflects the SNR of the signal at $\\tau$ , and $\\omega(\\lambda_{\\tau})$ is a pre-specified weighting function. We follow [21] for implementing the SNR and $\\omega(.)$ . In this setup, $\\psi_{D}$ is the student while the T2I diffusion model is the teacher.", + "bbox": [ + 75, + 375, + 467, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Eq. (6), the subscript 'unet' emphasizes the U-Net architecture of the student and teacher, which is commonly followed in denoising diffusion models. Details of our student model $\\psi_{\\mathcal{D}}$ are provided in the supplementary material. Similar to $z_{\\mathrm{clip}}^{\\mathrm{T}}$ in Eq. (2), the $z_{\\mathrm{unet}}^{\\mathrm{T}}$ at $\\tau$ is generated based on $\\mathcal{A}_X$ , which conditions $\\psi_{\\mathcal{D}}$ on the responsible concepts being considered by the user. We further follow the processing of $z_{\\mathrm{unet}}^{\\mathrm{T}}$ at a given $\\tau$ corresponding to Eq. (3), (4) and (5) to respectively compute $\\mathbf{z}_{\\mathrm{unet}}^{\\mathrm{zca}}$ , $\\mathbf{z}_{\\mathrm{unet}}^{\\mathrm{resp}}$ , and $z_{\\mathrm{unet}}^{\\mathrm{final}}$ at $\\tau$ . Fig. 2c illustrate the distillation process to train $\\psi_{\\mathcal{D}}$ for the RIIDL module on receiving the textual embedding input from $\\mathcal{E}$ . In Fig. 2d, an illustration is provided for the deployed model, where RIIDL is able to guide the generated image by injecting signals to the early denoising stages of the T2I diffusion model to encourage generation along responsible the concept in $\\mathcal{A}_X$ . Figure 2e further shows that extending $\\mathcal{A}_X$ enables simultaneous control over multiple concepts because $\\psi_{\\mathcal{D}}$ gets conditioned on multiple concepts.", + "bbox": [ + 75, + 481, + 467, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Dual-Space Integration", + "text_level": 1, + "bbox": [ + 76, + 758, + 290, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The methods discussed in § 3.3 and § 3.4 provide add-on modules $\\mathcal{E}_{\\mathrm{resp}}(\\cdot)$ and $\\mathcal{D}_{\\mathrm{resp}}(\\cdot)$ in the form of plug-and-play models $\\psi_{\\mathcal{E}}$ and $\\psi_{\\mathcal{D}}$ that can be directly plugged into the text encoder $\\mathcal{E}$ and diffusion model $\\mathcal{D}$ of the T2I pipeline. Whereas both modules are usable standalone, we also integrate them in this work for a comprehensive dual space control on the generative outputs. When both modules are integrated into the T2I pipeline, the overall function of gen", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "erating a responsible image $I$ from a prompt $t\\in \\mathcal{T}$ can be expressed as:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi_ {\\text {r e s p}} (t) = \\lambda_ {\\mathcal {E}} \\cdot \\mathcal {D} \\left(\\mathcal {E} _ {\\text {r e s p}} (t)\\right) + \\lambda_ {\\mathcal {D}} \\cdot \\mathcal {D} _ {\\text {r e s p}} (\\mathcal {E} (t)), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 130, + 890, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathcal{E}}$ and $\\lambda_{\\mathcal{D}}$ are weighting factors following the relation $\\lambda_{\\mathcal{E}} + \\lambda_{\\mathcal{D}} = 1$ . The integrated $\\Psi_{\\mathrm{resp}}(t)$ provides effective control over high-level semantic information in the embedding and latent spaces. It provides a comprehensive plug-and-play method for the T2I pipeline for altering the behavior of $\\Psi$ to follow concepts in $\\mathcal{A}_X$ for responsible text-to-image generation.", + "bbox": [ + 496, + 155, + 890, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 267, + 614, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method is evaluated for a variety of concepts related to fair and safe image generation. The used evaluation protocols are based on standardized benchmarks. For reproducibility, further details on implementation, dataset descriptions, and configurations of hardware and software libraries are also provided in the supplementary material. Our experiments use the Stable Diffusion v1.4 model [35]. We compare with SLD [38], ESD [13], UCE [14], and Vector-SD [23], using WinoBias [52] and I2P [38] datasets, which contain fairness-sensitive and safety-critical scenarios.", + "bbox": [ + 496, + 284, + 890, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fair Generation Evaluation: In regards to the fair generation, we evaluate performance across professions with documented gender and racial biases. The dataset for this evaluation is WinoBias [52], which comprises 35 professions historically associated with gender bias. Fairness is quantitatively assessed using the deviation ratio $\\Delta$ as in literature [28, 11, 23], which measures the disparity in generated representations across specified responsible aspects by comparing the model's output distribution against an expected baseline. Formally, $\\Delta$ is defined as: $\\Delta = \\frac{\\max_{a\\in\\mathcal{A}}\\left|\\frac{N_a}{N} - \\frac{1}{|\\mathcal{A}|}\\right|}{1 - \\frac{1}{|\\mathcal{A}|}}$ , where $\\mathcal{A}$ is the set of attributes within a responsible aspect (e.g., gender or race), $N$ represents the total number of generated images, and $N_{a}$ denotes the count of images for which the highest probability attribute prediction corresponds to attribute $a$ . For extensive evaluation, we further employ \"challenging prompts\" [23] in our experiments, which are known to produce bias representations towards male depictions [14].", + "bbox": [ + 496, + 436, + 890, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Safe Generation Evaluation: We also evaluate the generated content against the standard criteria for identifying harmful or culturally sensitive outputs [38, 23]. Images are flagged as inappropriate if they depict or imply content associated with sexual content, self-harm, hate, illegal activity, shock, harassment, or violence. Our evaluation integrates both the NudeNet detector1 and Q16 classifier [37], as per common practices in responsible generation assessment, to systematically flag content deemed inappropriate [38, 23]. We examine a diverse range of images to rigorously assess robustness across varied outputs.", + "bbox": [ + 496, + 719, + 890, + 886 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$^{1}$ https://github.com/notAI-tech/NudeNet", + "bbox": [ + 517, + 886, + 733, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2980", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d7e1d314c8d3e765326e787ae1dd005184cea0695e0da311e6ca5bb92a2981e9.jpg", + "table_caption": [ + "Table 1. Bias generation quantified by deviation ratio $\\left( {0 \\leq \\Delta \\leq 1}\\right)$ . Lower values indicate better performance. Results provided for Gender and Race bias across standard and extended (Gender-Pro/Race-Pro) Winobias datasets [52]." + ], + "table_footnote": [ + "SD: Standard Diffusion Model, U: Unified Concept Editing [14], V: Vector Interpret Diffusion [23]." + ], + "table_body": "
AttributeGender(↓)Gender-Pro(↓)Race(↓)Race-Pro(↓)
SD [35]U [14]V [23]OursSD [35]U [14]V [23]OursSDU [14]V [23]OursSD [35]U [14]V [23]Ours
CEO0.920.280.150.040.900.580.210.130.380.130.220.080.310.080.220.12
Doctor0.920.200.080.050.520.320.100.100.920.070.260.070.590.520.150.12
Nurse1.000.390.960.040.980.840.430.180.760.250.300.070.390.790.080.14
Receptionist0.840.380.880.040.980.960.860.170.880.100.360.070.740.140.250.11
Teacher0.300.060.510.040.480.160.070.130.510.100.040.080.260.230.210.10
", + "bbox": [ + 101, + 128, + 864, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a056058ee7132a054dffcd32923deab302601261f7b392c6b58deb2751636ae.jpg", + "image_caption": [ + "Figure 3. Pair-wise comparison for the Stable Diffusion (SD) baseline and our plugin to it for responsible generation along age, gender and race. SD images contain stereotypical associations to the profession, which get removed by our method by often accounting for more than one responsible attributes, while maintaining high image quality." + ], + "image_footnote": [], + "bbox": [ + 122, + 270, + 848, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 612, + 282, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fair Generation: In Table 1, we present the bias quantification results using the deviation ratio $(0 \\leq \\Delta \\leq 1)$ across gender and racial attributes, evaluated under both the standard (Gender and Race) and extended (Gender-Pro and Race-Pro) settings of the WinoBias dataset [52]. Lower values of $\\Delta$ indicate reduced demographic bias. Overall, our method consistently achieves highly competitive results. Even under the Gender-Pro setting, where prompts are specifically crafted to amplify stereotypical associations, our method maintains its robustness. For racial bias, our method similarly maintains its performance.", + "bbox": [ + 75, + 638, + 470, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 further quantifies the debiasing efficacy of the our approach across professions with known demographic skew, displaying results for five professions and the average deviation across all 35 professions in the WinoBias [52]. Results show a substantial bias reduction by our method. Consistent bias reduction demonstrates that bias mitiga", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion is achieved by our method independent of profession-specific attributes. Averaged over all professions (last row), our method provides a significant gain over the baseline methods. Such efficacy of our method comes from its ability to enable debiasing using a collective set of attributes in $\\mathcal{A}_X$ and modulate the T2I pipeline to generate images responsibly considering all those attributes.", + "bbox": [ + 496, + 613, + 890, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To show how this manipulation varies the images generated by Stable Diffusion (SD), in Fig. 3 we provide representative examples. The figure shows a range of professions for which pairs of images are generated using the same seeds. The SD images show stereotypical associations to the professions. Our method is plugged in to SD with set $\\mathcal{A}_X$ collectively containing race, gender and age attributes. It is observable in the provided representative examples that images get modulated towards these responsible concepts, often incorporating more than one responsible concept. We also provide further results in the supplementary material.", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2981", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b85499286b74f672c46ee1181811dca5cf3bff6f77d07ad48a94f00676565fa3.jpg", + "table_caption": [ + "Table 2. Debaising performance across 5 randomly selected professions and the average of all 35 professions in Winobias dataset [52]. The metric $\\Delta = 0$ indicate ideal debraising. Our method shows the lowest average deviation compared to previous approaches." + ], + "table_footnote": [], + "table_body": "
ProfessionOriginal-SDConcept Algebra [44]Debias-VL [11]TIME [28]TIMEP [14]Ours
Librarian0.86 ± 0.060.66 ± 0.070.34 ± 0.060.26 ± 0.050.35 ± 0.010.04 ± 0.02
Teacher0.42 ± 0.010.46 ± 0.000.11 ± 0.050.34 ± 0.060.07 ± 0.060.05 ± 0.01
Sheriff0.99 ± 0.010.38 ± 0.220.82 ± 0.080.22 ± 0.050.10 ± 0.050.06 ± 0.03
Analyst0.58 ± 0.120.24 ± 0.180.71 ± 0.020.52 ± 0.030.13 ± 0.050.07 ± 0.02
Doctor0.78 ± 0.040.40 ± 0.020.50 ± 0.040.58 ± 0.030.41 ± 0.080.06 ± 0.01
Average0.67 ± 0.010.43 ± 0.010.55 ± 0.010.44 ± 0.000.31 ± 0.000.05 ± 0.02
", + "bbox": [ + 187, + 111, + 781, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/647bae11f6f88fbbe900f34c275dd21b50938d0972003b742e99669fc4d6344b.jpg", + "image_caption": [ + "A doctor" + ], + "image_footnote": [], + "bbox": [ + 122, + 258, + 197, + 377 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f876c11b6a6d4cbb87f6edfcc345edaf356b5127790c1b7769e23493b8092e8.jpg", + "image_caption": [ + "Agender" + ], + "image_footnote": [], + "bbox": [ + 207, + 258, + 284, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d54ab4a7812cee38482f0d082e2a862cf31cbcfbfe5ec0f605a52a73b0f28e3.jpg", + "image_caption": [ + "Female" + ], + "image_footnote": [], + "bbox": [ + 207, + 318, + 284, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d402b393d3afed61194e4c40234ab634bae1e7f58592aef55ce8d046493c414b.jpg", + "image_caption": [ + "Arace" + ], + "image_footnote": [], + "bbox": [ + 290, + 258, + 364, + 318 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a708ac91ca8fa0f96621e9f674fe8e6ea38e3af83cc3b5201a5c7d789281db2.jpg", + "image_caption": [ + "Asian - Black" + ], + "image_footnote": [], + "bbox": [ + 290, + 318, + 364, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d414e854cb8a1ef231176c3df22d8b551463026e413d56fe90ae42ca532eafd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 258, + 439, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02b4e1490f252c7a839a038fc803e554af1dd703fb8678328a94e5a33b90d063.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 319, + 439, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6c90acb948f62f030c7a44cbaae6addece40cf6dde9801583ddfcf61dcf5ee3.jpg", + "image_caption": [ + "Aage" + ], + "image_footnote": [], + "bbox": [ + 442, + 258, + 596, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0198d61a9ff7b72fe4a191e253ad34c8c70fc626455d50e48b168af66a832363.jpg", + "image_caption": [ + "Young - Old" + ], + "image_footnote": [], + "bbox": [ + 442, + 319, + 596, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a35b63f1fbcba0ffd719dac79fa6770e7fa60adc313d33e2188566714a0c10f2.jpg", + "image_caption": [ + "$\\mathrm{A}_{\\text {gender }} + \\mathrm{A}_{\\text {race }}$", + "$\\mathrm{A}_{\\text {gender }} + \\mathrm{A}_{\\text {race }} + \\mathrm{A}_{\\text {age }}$" + ], + "image_footnote": [], + "bbox": [ + 599, + 258, + 751, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/369aaf48e207750062ca39011c6855bfe6d15b5f5afabe1f0af3d46cdab488fa.jpg", + "image_caption": [ + "Female/Black-Asian", + "Female / Asian / Old" + ], + "image_footnote": [], + "bbox": [ + 599, + 319, + 751, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/134f121f97b93500c5d70073091c71e994b3b9680b8a09625238187ab64456a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 258, + 834, + 318 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec55a9a0cab5be0431d09f1d415d883344f15eb310af057a4a9bd0302b58d9a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 318, + 834, + 376 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2c7f6022edae4cfa2040c0f9fca063b3de5615b06f35215f89c33aaa7aae9844.jpg", + "table_caption": [ + "Table 3. Proportion of images classified as inappropriate on I2P benchmark [38]. Lower values are more desirable." + ], + "table_footnote": [], + "table_body": "
CategorySDV-SD [23]SLD [38]ESD [13]Ours
Sexual0.390.230.160.190.15
Violence0.450.310.220.420.24
Hate0.420.300.190.340.18
Harassment0.350.190.160.280.19
Illegal0.350.240.180.340.14
Shocking0.530.380.270.430.28
Self-harm0.450.290.200.360.20
Avg.0.410.270.200.320.20
", + "bbox": [ + 98, + 489, + 442, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Safe Generation: To benchmark safe and appropriate nature of the generated content, we evaluate our method on the I2P [38], which classifies generated images across categories covering a range of inappropriate content. Table 3 benchmarks our method against the state-of-the-art methods following standard protocol [38], [13]. As seen, our technique significantly reduces inappropriateness of the content across all categories. Generally, maintaining superior performance against the methods. On average, our performance is similar to SLD [38], which is a dedicated 'safe generation' method.", + "bbox": [ + 75, + 628, + 468, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Further Results & Discussion", + "text_level": 1, + "bbox": [ + 76, + 801, + 354, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-Concept Composition and Control: Our technique allows responsible generation with multi-concept composition. Multiple concepts can compose $\\mathcal{A}_X$ , which enables the hierarchical function $\\Psi_{\\mathrm{resp}}$ to combine them in latent and embedding space of the T2I pipeline. Figure 4 demonstrates", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a5bb74c62d21feb7b64518ece0ba11f8cc3b2b93c9791bce1c5118335248016.jpg", + "image_caption": [ + "Figure 4. Controlled composite responsible generation using the proposed method. By using different concepts in $\\mathcal{A}_X$ in Eq. (1), and employing dual space control in (7), our technique can enable responsibility along single or multiple concepts, as desired. Provided the prompt \"A doctor\" to Stable Diffusion, alignment is achieved for a target concept composition (top label) for the attributes noted at the bottom of each image set. See supplementary material for more examples.", + "Figure 5. Examples of extending $\\mathcal{A}_{\\mathrm{resp}}$ with additional attributes - possibly unrelated to core responsible concepts. Extensions with smile and glasses are shown. In addition to varying images along responsible concepts of age, race and gender, our method is able to seamlessly incorporate the additional concepts in images." + ], + "image_footnote": [], + "bbox": [ + 503, + 462, + 885, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "this capability within $\\Psi_{\\mathrm{resp}}$ , showing how distinct responsible aspects, such as age, gender, and race; can be independently learned and then composed to produce varied yet coherent outputs. By manipulating combinations within $\\mathcal{A}_X$ , we demonstrate an excellent control achieved with interpretable variations across the composite attribute space, effectively balancing diversity with responsible aspects.", + "bbox": [ + 496, + 732, + 890, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scalability of Concept Spaces: In our original experiments, the responsible concept set $\\mathcal{A}_{\\mathrm{resp}}$ in our framework, as defined in Eq. (1), included core attributes of $\\mathcal{A}_{\\mathrm{race}}, \\mathcal{A}_{\\mathrm{gender}}, \\mathcal{A}_{\\mathrm{age}},$ and $\\mathcal{A}_{\\mathrm{safe}}$ . This was only to benchmark", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2982", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a40fef026c08f663c6167e331602b0476f7f0a15f7181530f7ccb117ce1c1006.jpg", + "image_caption": [ + "Figure 6. Representative curtailed example of transitions caused by RICE and RIIDL to an output. An image of 'a doctor' is generated and 'female' concept is added individually by the modules. Increasingly changing the influence of the added concept representation provided by each module leads to unique alteration to the generated output." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 468, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the method with the existing approaches. Our design allows the \"responsible\" set to be extendable. This scalability allows for seamlessly handling additional nuanced attributes in images, such as facial expressions (e.g., smiling) and accessories (e.g., glasses). Such concepts can act as supplementary attributes with the existing attributes, enabling complex yet interpretable multi-aspect control. In Fig. 5, we provide representative examples of scaling up the considered concepts set. It leads to combining the responsible subspaces of the core responsible features with subspaces for smiling and glasses. It is observable that our framework is scalable to not only core responsibility related concepts, but to any semantically meaningful concepts. We also provide additional results in the supplementary material.", + "bbox": [ + 75, + 310, + 468, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Individual Modules: Our method takes advantage of two modules, namely; of RICE (§ 3.3) and RIIDL (§ 3.4). To assess the specific contribution of each component within the responsible generation framework, we perform a systematic ablation by individually removing and adjusting key modules. This analysis allows selectively varying the influence both modules on the output. When only RICE operates, modifications originally occur in the embedding concept space for the T2I pipeline, facilitating transitions from base to responsible concept spaces. The RIIDL module contribute an additional layer of fine-grained control, allowing nuanced and context-sensitive adjustments within the intermediate latents during diffusion. This capability supports precise tuning of details without overriding the global context, complementing the broader adjustments. In tandem, these modules establish a dynamic balance within the dual-space intervention strategy, where RICE provides prediffusion global control, and RIIDL governs the responsible aspects during the denoising process. Whereas we provide further results supporting these point in the supplementary material, in Fig. 6 a curtailed representative example is provided to show the visual effects.", + "bbox": [ + 75, + 522, + 470, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the figure, we generate images using prompt 'a doctor' and with an added concept 'female' from the individual modules. The 'female' concept representation $\\mathbf{z}^{\\mathrm{female}}$ pro", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vided by the modules is added to the original corresponding representations with an increasing degree of influence on the output. It can be observed that for RICE, this results in more prominent shifts in appearance and attire, indicating a significant control. For RIIDL, the latent space provides more refined transitions with subtle and smoother progression of the visual features. On one hand, these results demonstrate the complementary nature of the proposed modules, on the other; they establish the effectiveness of the modules as stand-alone add-ons.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 248, + 619, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research presented a unique approach to control generative outputs of text-to-image (T2I) models in 'fair and safe' domain. Our method induces two plug-and-play modules that can be plugged into the textual embedding space and diffusion latent space of the T2I pipeline. The modules are neural models learned by distilling the text encoder and diffusion model of the original pipeline while conditioning the knowledge distillation on a pre-defined responsible concepts set, as desired by the user. Our modules also account for semantic interpretability by leveraging concept whitening. We demonstrate that both modules can explicitly contribute to responsible image generation in a complementary manner. Our extensive benchmarking with state-of-the-art methods for fair and safe image generation on two standard datasets show highly promising results. We also show that our method can conveniently deal with multiple concepts in the responsible image space while being seamlessly extendable to more broader concepts.", + "bbox": [ + 496, + 270, + 890, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and ethics concerns: A potential shortcoming of our method is that, to enable a comprehensive control over the generative output, it employs multiple hyperparameters, e.g., $\\lambda$ 's in Eq. (7). Tuning these hyperparameters can potentially be seen as a limitation of our approach. However, the main purpose of allowing these adjustable hyper-parameters in our design is to provide a finer control to the user. It is convenient to adjust their values because the controls they provide have clear high-level interpretations. We provide guidelines on the hyper-parameter value adjustment in the supplementary material. From the ethics viewpoint, although our aim is to ensure responsible outputs; a potential misuse of exploiting the strong output control of our technique can be by replacing the $A_{\\mathrm{resp}}$ with another set $A_{\\mathrm{irresp}}$ . In $A_{\\mathrm{irresp}}$ , unsafe and unfair concepts can be included. This can have an opposite impact of our technique. Hence, our method should only be used when the user has complete control over the set $A_{\\mathrm{resp}}$ .", + "bbox": [ + 496, + 544, + 890, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment: Naveed Akhtar is a recipient of the Australian Research Council Discovery Early Career Researcher Award (project # DE230101058) funded by the Australian Government. This work is also partially supported by Google Research Scholar Program Award.", + "bbox": [ + 496, + 816, + 890, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2983", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 90, + 166, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sandhini Agarwal et al. \"Evaluating CLIP: towards characterization of broader capabilities and downstream implications\". In: arXiv preprint arXiv:2108.02818 (2021).", + "[2] Zhongjie Ba et al. \"SurrogatePrompt: Bypassing the Safety Filter of Text-To-Image Models via Substitution\". In: arXiv preprint arXiv:2309.14122 (2023).", + "[3] Hritik Bansal et al. \"How well can Text-to-Image Generative Models understand Ethical Natural Language Interventions?\" In: arXiv preprint arXiv:2210.15230 (2022).", + "[4] Hugo Berg et al. \"A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning\". In: arXiv preprint arXiv:2203.11933 (2022).", + "[5] Federico Bianchi et al. \"Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale\". In: arXiv preprint arXiv:2211.03759 (2022).", + "[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. \"Multimodal datasets: misogyny, pornography, and malignant stereotypes\". In: arXiv preprint arXiv:2110.01963 (2021).", + "[7] Manuel Brack et al. \"The Stable Artist: Steering Semantics in Diffusion Latent Space\". In: arXiv preprint arXiv:2212.06013 (2022).", + "[8] Nicholas Carlini et al. Extracting Training Data from Diffusion Models. 2023. DOI: 10.48550/ARXIV.2301.13188. URL: https://arxiv.org/abs/2301.13188.", + "[9] Zhi Chen, Yijie Bei, and Cynthia Rudin. \"Concept Whitening for Interpretable Image Recognition\". In: Nature Machine Intelligence 2.12 (Dec. 2020), pp. 772-782.", + "[10] Jaemin Cho, Abhay Zala, and Mohit Bansal. “Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models”. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 3043–3054.", + "[11] Ching-Yao Chuang et al. \"Debiasing vision-language models via biased prompts\". In: arXiv preprint arXiv:2302.00070 (2023).", + "[12] Florinel-Alin Croitoru et al. \"Diffusion models in vision: A survey\". In: IEEE Transactions on Pattern Analysis and Machine Intelligence 45.9 (2023), pp. 10850-10869.", + "[13] Rohit Gandikota et al. “Erasing concepts from diffusion models”. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 2426–2436.", + "[14] Rohit Gandikota et al. \"Unified concept editing in diffusion models\". In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024, pp. 5111-5120.", + "[15] Jindong Gu et al. \"Towards robust prompts on vision-language models\". In: arXiv preprint arXiv:2304.08479 (2023)." + ], + "bbox": [ + 81, + 112, + 470, + 875 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] René Haas et al. \"Discovering interpretable directions in the semantic latent space of diffusion models\". In: 2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG). IEEE. 2024, pp. 1-9.", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. “Denoising diffusion probabilistic models”. In: Advances in neural information processing systems 33 (2020), pp. 6840–6851.", + "[18] Chi-Pin Huang et al. \"Receler: Reliable concept erasing of text-to-image diffusion models via lightweight erasers\". In: arXiv preprint arXiv:2311.17717 (2023).", + "[19] Tero Karras et al. “Elucidating the design space of diffusion-based generative models”. In: Advances in neural information processing systems 35 (2022), pp. 26565–26577.", + "[20] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. \"Diffusionclip: Text-guided diffusion models for robust image manipulation\". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022, pp. 2426-2435.", + "[21] Diederik Kingma et al. \"Variational diffusion models\". In: Advances in neural information processing systems 34 (2021), pp. 21696-21707.", + "[22] Nupur Kumari et al. \"Ablating concepts in text-to-image diffusion models\". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 22691-22702.", + "[23] Hang Li et al. \"Self-discovering interpretable diffusion latent directions for responsible text-to-image generation\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 2024.", + "[24] Sasha Luccioni et al. \"Stable bias: Evaluating societal representations in diffusion models\". In: Advances in Neural Information Processing Systems 36 (2024).", + "[25] Chenlin Meng et al. \"On distillation of guided diffusion models\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 14297-14306.", + "[26] Ranjita Naik and Besmira Nushi. \"Social Biases through the Text-to-Image Generation Lens\". In: arXiv preprint arXiv:2304.06034 (2023).", + "[27] Alex Nichol et al. \"Glide: Towards photorealistic image generation and editing with text-guided diffusion models\". In: arXiv preprint arXiv:2112.10741 (2021).", + "[28] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. \"Editing implicit assumptions in text-to-image diffusion models\". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 7053-7061.", + "[29] Yong-Hyun Park et al. \"Direct unlearning optimization for robust and safe text-to-image models\". In: arXiv preprint arXiv:2407.21035 (2024).", + "[30] William Peebles and Saining Xie. \"Scalable diffusion models with transformers\". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 4195-4205." + ], + "bbox": [ + 504, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2984", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Alec Radford et al. \"Learning Transferable Visual Models From Natural Language Supervision\". In: arXiv preprint arXiv:2103.00020 (2021).", + "[32] Aditya Ramesh et al. \"Hierarchical text-conditional image generation with clip latents\". In: arXiv preprint arXiv:2204.06125 1.2 (2022), p. 3.", + "[33] Javier Rando et al. Red-Teaming the Stable Diffusion Safety Filter. 2022. DOI: 10.48550/ARXIV.2210.04610. URL: https://arxiv.org/abs/2210.04610.", + "[34] Robin Rombach et al. \"High-resolution image synthesis with latent diffusion models\". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022, pp. 10684-10695.", + "[35] Robin Rombach et al. \"High-resolution image synthesis with latent diffusion models\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022, pp. 10684-10695.", + "[36] Chitwan Sahara et al. \"Photorealistic text-to-image diffusion models with deep language understanding\". In: Advances in neural information processing systems 35 (2022), pp. 36479-36494.", + "[37] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. \"Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content?\" In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. 2022, pp. 1350-1361.", + "[38] Patrick Schramowski et al. \"Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 22522-22531.", + "[39] Shoaib Ahmed Siddiqui et al. \"Metadata Archaeology: Unearthing Data Subsets by Leveraging Training Dynamics\". In: arXiv preprint arXiv:2209.10015 (2022).", + "[40] Gowthami Somepalli et al. \"Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models\". In: arXiv preprint arXiv:2212.03860 (2022).", + "[41] Lukas Struppek, Dominik Hintersdorf, and Kristian Kersting. \"The biased artist: Exploiting cultural biases via homoglyphs in text-guided image generation models\". In: (2022).", + "[42] Yu-Lin Tsai et al. \"Ring-A-Bell! How Reliable are Concept Removal Methods for Diffusion Models?\" In: arXiv preprint arXiv:2310.10012 (2023).", + "[43] Peiran Wang et al. \"Moderator: Moderating Text-to-Image Diffusion Models through Fine-grained Context-based Policies\". In: arXiv preprint arXiv:2408.07728 (2024).", + "[44] Zihao Wang et al. \"Concept algebra for text-controlled vision models\". In: arXiv preprint arXiv:2302.03693 2 (2023).", + "[45] Yongliang Wu et al. \"Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient\". In: arXiv preprint arXiv:2405.15304 (2024)." + ], + "bbox": [ + 81, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Tianwei Xiong et al. \"Editing Massive Concepts in Text-to-Image Diffusion Models\". In: arXiv preprint arXiv:2403.13807 (2024).", + "[47] Yijun Yang et al. \"GuardT2I: Defending Text-to-Image Models from Adversarial Prompts\". In: arXiv preprint arXiv:2403.01446 (2024).", + "[48] Jaehong Yoon et al. \"SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation\". In: arXiv preprint arXiv:2410.12761 (2024).", + "[49] Gong Zhang et al. “Forget-me-not: Learning to forget in text-to-image diffusion models”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024, pp. 1755–1764.", + "[50] Hongxiang Zhang, Yifeng He, and Hao Chen. \"Steerdiff: Steering towards safe text-to-image diffusion models\". In: arXiv preprint arXiv:2410.02710 (2024).", + "[51] Yimeng Zhang et al. \"To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now\". In: European Conference on Computer Vision. Springer. 2025, pp. 385-403.", + "[52] Jieyu Zhao et al. \"Gender bias in coreference resolution: Evaluation and debiasing methods\". In: arXiv preprint arXiv:1804.06876 (2018)." + ], + "bbox": [ + 504, + 90, + 890, + 440 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2985", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/2c24c8f1-1822-42ba-8659-5feab5b2d726_model.json b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/2c24c8f1-1822-42ba-8659-5feab5b2d726_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f97e21789551a593b76612bf72d5bafcd99f8211 --- /dev/null +++ b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/2c24c8f1-1822-42ba-8659-5feab5b2d726_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.826, + 0.177 + ], + "angle": 0, + "content": "Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control" + }, + { + "type": "text", + "bbox": [ + 0.36, + 0.213, + 0.612, + 0.229 + ], + "angle": 0, + "content": "Basim Azam Naveed Akhtar" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.234, + 0.834, + 0.253 + ], + "angle": 0, + "content": "School of Computing and Information Systems, The University of Melbourne, Australia" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.258, + 0.675, + 0.272 + ], + "angle": 0, + "content": "{basim.azam, naveed.akhtar1}@unimelb.edu.au" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.301, + 0.889, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.576, + 0.893, + 0.649 + ], + "angle": 0, + "content": "Figure 1. Our unique plug-and-play interpretable approach simultaneously controls a range of concepts for responsible and fair image generation with text-to-image pipelines. Our method enables control over both text embedding space and latent diffusion space. Shown examples compare unfair and unsafe generation for the given prompts by the Stable Diffusion (top), along with their responsible counterparts resulting from our approach influencing the text encoder (middle) and the diffusion model (bottom). We control individual concepts for diverse/safe generation in these examples while modeling them in continuous composite responsible semantic spaces." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.658, + 0.314, + 0.674 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.475, + 0.903 + ], + "angle": 0, + "content": "Ethical issues around text-to-image (T2I) models demand a comprehensive control over the generative content. Existing techniques addressing these issues for responsible T2I models aim for the generated content to be fair and safe (non-violent/explicit). However, these methods remain bounded to handling the facets of responsibility concepts individually, while also lacking in interpretability. Moreover, they often require alteration to the original model, which compromises the model performance. In this work, we propose a unique technique to enable responsible T2I generation by simultaneously accounting for an extensive range of concepts for fair and safe content generation in a scalable manner. The key idea is to distill the target T2I pipeline with an external plug-and-play mechanism that learns an inter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.812 + ], + "angle": 0, + "content": "pretable composite responsible space for the desired concepts, conditioned on the target T2I pipeline. We use knowledge distillation and concept whitening to enable this. At inference, the learned space is utilized to modulate the generative content. A typical T2I pipeline presents two plug-in points for our approach, namely; the text embedding space and the diffusion model latent space. We develop modules for both points and show the effectiveness of our approach with a range of strong results. Our code can be accessed at https://basim-azam.github.io/responsiblediffusion/" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.831, + 0.633, + 0.847 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Rapid advances in text-to-image (T2I) generative pipeline are revolutionizing numerous vision applications, offering unprecedented convenience for synthesizing high quality" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "2976" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "visual content using textual descriptions [34, 32, 30, 19, 17]. Unfortunately, this convenience coupled with the generated content realism offered by publicly accessible T2I models, e.g., Stable Diffusion [34], DALL-E [32], Imagen [36], also has undesired ethical implications. Confronting the core values of fair and safe (non-violent/explicit) development and deployment of Artificial Intelligence, these models lack in producing generative content responsibly [15, 51]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.473, + 0.474 + ], + "angle": 0, + "content": "The reasons of irresponsible content generation by contemporary T2I models lie deep in their training process, which must use large volumes of (mostly uncurated) data. Issues like inappropriate content and harmful stereotyping lurking in the data are passed on to the generative models through training [2, 42], which cause problems after their deployment. Hence, it is critical to develop methodologies to explicitly control generative models from producing irresponsible content. Addressing that, several techniques have emerged recently, encompassing approaches like input prompt filtering [48, 2], post-hoc content moderation [47, 43, 22], machine unlearning [45, 29, 18, 49], and model editing [14, 46]. Though effective, these approaches face limitations like requiring human-intervention, ad-hoc methodology, and selective dealing of different facets of the concepts defining 'responsible generation'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.473, + 0.748 + ], + "angle": 0, + "content": "Another contributing factor to the ethical concerns related to T2I models is our current lack of wholesome understanding of the latent space of the diffusion models. Researchers have already started exploring interpretability [16, 50] and manipulation of diffusion model latent spaces to mitigate unintended T2I model behavior [14, 23]. Controlling the latent space holds promise when it is possible to disentangle semantic concepts in it. One recent inspiring approach in this direction [23] identifies vectors in the latent diffusion space in the directions of individual semantic concepts, which can be used to restrict model outputs. However, it is arguable that such a discretized treatment of the concepts within the diffusion latent space is restrictive. It is also intrinsically limited to dealing with diverse facets of the concepts individually. Overall, it still remains a widely open challenge for the research community to comprehensively embed the multi-faceted concepts related to fair and safe image generation in T2I pipelines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this work, we address this by presenting a unique scalable approach to extensively incorporate fair and safe generative abilities in a T2I pipeline in a plug-and-play manner. Our key insight is that a T2I model can be distilled for a range of user-desired responsibility concepts, conditioned on the original model. We perform this distillation by inducing a student model that treats the concepts related to fairness and safety in a continuous space. Within this space, we further leverage concept whitening [9] to disentangle the representations of concepts for a better ultimate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "control on the generative image modulation. The underlying framework of our technique is generic, in that it is applicable to any representation space encoding semantic concepts. Hence, we explore its application to both text encoder embedding space and diffusion model latent space in the T2I pipeline. Methods for both variants address their unique challenges, but prove equally effective - see Fig. 1. The key contributions of this work are summarized below." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.215, + 0.892, + 0.289 + ], + "angle": 0, + "content": "- We propose a unique plug-and-play method for responsible image generation with T2I pipeline that enables comprehensive incorporation of fairness and safety concepts while modeling them under a distilled continuous space conditioned on the generative model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.29, + 0.892, + 0.35 + ], + "angle": 0, + "content": "- We tailor our method as a text embedding plug-in, and diffusion latent plug-in while also leveraging concept whitening to ensure precise interpretable control over the content. Both plug-ins are employed in our approach." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.35, + 0.892, + 0.412 + ], + "angle": 0, + "content": "- With extensive experiments, we not only demonstrate state-of-the-art or comparable performance in fair and safe image generation, but also show other unique interesting properties of our approach." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.215, + 0.892, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.423, + 0.651, + 0.438 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.569 + ], + "angle": 0, + "content": "Text-to-image (T2I) generation has significantly expanded the capabilities of generative AI, enabling high-quality outputs using textual descriptions [34, 32, 30, 19, 17]. While T2I methods like Stable Diffusion [34], DALL-E [32], GLIDE [27], and Imagen [36] have achieved impressive realism and adaptation in numerous applications, they also face several challenges. One of them is the concern about their irresponsible and unethical content generation [15, 2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Contemporary T2I models are known to generate inappropriate images, including nude and violent content, and they are also susceptible to social and cultural biases persistently found in the public domain datasets used to train these models [10, 24, 41]. The increased popularity of T2I models has driven research into understanding and mitigating biases in their outputs, particularly focusing on safety filters [33], semantic space organization [7], and tendencies toward stereotypical portrayals [5, 10, 26]. Studies show that these models, often guided by biased systems, e.g., CLIP [31, 1], may memorize sensitive data [8, 40], which, combined with unfiltered web-based datasets containing harmful content [6, 39, 4, 3], propagates unintended biases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Research to address these concerns can be divided into three main directions; namely, preemptive content filtering, model-level adjustments, and post-hoc moderation. The preemptive filtering approaches involve controlling input prompts to reduce biases before they influence image generation. For example, prompt-based filters [48, 2] screen out language associated with inappropriate content [19, 30]. In the model-level adjustment, unlearning techniques [45, 29, 18, 49] aim to remove sensitive concepts directly from the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2977" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "model's learned representations. Similarly, model editing methods target specific parameters of the models to limit the generation of harmful content without extensive re-training [28, 14, 46]. Although useful, these methods can negatively impact the original model performance. They can also become computationally intensive as the harmful concepts increase [49]. Post-hoc moderation methods [47, 43, 22] involve both automated and manual processes like GuardT2I [47] and context-based filters [43, 22]. These technique only partially address model biases and irresponsible generation, and may easily overlook implicit negative concept associations in the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.501 + ], + "angle": 0, + "content": "More recently, research on diffusion model latent space manipulation and interpretability is gaining prominence in regards to controlling the T2I outputs. Studies such as [16, 50] explore the semantic organization within the latent spaces to understand and intervene in the generation process. Li et al. [23] investigated interpretable directions of target concepts in the latent semantic space. These methods remain effective, however; their underlying discretized treatment of the individual semantic concepts limits their scalability. In this work, we address safe and fair modulation of both embedding and latent spaces without explicitly discretizing the directions, and focusing on their continuous space instead. Our distinctive plug-and-play approach learns an interpretable space for a range of responsible concepts while conditioning it on the T2I pipeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.505, + 0.27, + 0.523 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.633 + ], + "angle": 0, + "content": "We present a unique approach for responsible image generation with T2I pipeline that develops add-on modules usable in a plug-and-play manner. The modules are neural models induced by distilling the target T2I model for a scalable list of responsible semantic concepts addressing fair and safe (non-violent/explicit) image generation. We first formally present the problem as perceived in this work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.637, + 0.294, + 0.65 + ], + "angle": 0, + "content": "3.1. Problem Formalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.473, + 0.839 + ], + "angle": 0, + "content": "A text-to-image model, \\(\\Psi(t): t \\to \\bar{I}\\), maps a textual prompt \\(t \\in \\mathcal{T}\\) to an image \\(\\bar{I} \\in \\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times 3}\\), where \\(\\mathcal{T}\\) and \\(\\mathcal{M}\\) are the sets of possible text prompts and images. Depending on the data used to train the T2I model, the image \\(\\bar{I}\\) may belong to an image subset \\(\\overline{\\mathcal{R}}_{\\mathcal{A}_X} \\subset \\mathcal{M}\\). Conditioned on \\(\\mathcal{A}_X\\), \\(\\overline{\\mathcal{R}}_{\\mathcal{A}_X}\\) denotes the set of irresponsible/inappropriate images as identified by a collection of high-level cultural/social/societal aspects in set \\(\\mathcal{A}_X\\). In our settings, \\(\\mathcal{A}_X\\) is a subset of the aspects enlisted in Eq. (1). Notice that, \\(\\mathcal{A}_X \\subseteq \\mathcal{A}_{\\mathrm{resp}}\\) implies that \\(\\overline{\\mathcal{R}}_{\\mathcal{A}_X}\\) accounts for the fact that \\(\\bar{I}\\) can be inappropriate along more than one aspect - an issue often ignored in the existing literature." + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.846, + 0.47, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {\\text {r e s p}} = \\left\\{\\mathcal {A} _ {\\text {a g e}}, \\mathcal {A} _ {\\text {g e n d e r}}, \\mathcal {A} _ {\\text {r a c e}}, \\mathcal {A} _ {\\text {s a f e}} \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We form \\(\\mathcal{A}_{\\mathrm{resp}}\\) considering the aspects of responsible image generation accounted in the existing works [23, 38]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "Nevertheless, we keep \\(\\mathcal{A}_{\\mathrm{resp}}\\) extendable as required. In this work, we define each of the considered aspects using high-level attributes listed below, which also aligns with the existing literature [23, 38]: \\(\\mathcal{A}_{\\mathrm{gender}} = \\{\\mathbf{a}_{\\mathrm{male}}, \\mathbf{a}_{\\mathrm{female}}\\}\\), \\(\\mathcal{A}_{\\mathrm{race}} = \\{\\mathbf{a}_{\\mathrm{white}}, \\mathbf{a}_{\\mathrm{asian}}, \\mathbf{a}_{\\mathrm{black}}\\}\\), \\(\\mathcal{A}_{\\mathrm{age}} = \\{\\mathbf{a}_{\\mathrm{young}}, \\mathbf{a}_{\\mathrm{middle-aged}}, \\mathbf{a}_{\\mathrm{elderly}}\\}\\), and \\(\\mathcal{A}_{\\mathrm{safe}} = \\{\\mathbf{a}_{\\mathrm{harassment}}, \\mathbf{a}_{\\mathrm{sexual}}, \\mathbf{a}_{\\mathrm{violence}}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.895, + 0.32 + ], + "angle": 0, + "content": "It is worth emphasizing that the attribute sets noted above can also be extended if required. Our approach is not constrained by these sets, except for further training required to account for any additional aspect. Given an \\(\\mathcal{A}_X\\) - as defined by the model user, our objective is to alter the mapping of the T2I model to \\(\\Psi_{\\mathrm{resp}}(t):t\\to I\\) such that \\(I\\in \\mathcal{R}_{\\mathcal{A}_X}\\subseteq \\mathcal{M}\\) and \\(\\overline{\\mathcal{R}}_{\\mathcal{A}_X}\\cap \\mathcal{R}_{\\mathcal{A}_X} = \\emptyset\\). Here, \\(\\Psi_{\\mathrm{resp}}\\) is the responsible variant of \\(\\Psi\\) with an added functionality that restricts the outputs of \\(\\Psi\\) to \\(\\mathcal{A}_X\\)-constrained responsible image space \\(\\mathcal{R}_{\\mathcal{A}_X}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.323, + 0.707, + 0.339 + ], + "angle": 0, + "content": "3.2. Framework Blueprint" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.554 + ], + "angle": 0, + "content": "Conceptually, a T2I model \\(\\Psi\\) is a hierarchical function \\(\\Psi(t) = \\mathcal{D}(\\mathcal{E}(t))\\), where \\(\\mathcal{E}(.)\\) is a text-encoder - typically CLIP [20], and \\(\\mathcal{D}(.)\\) is a diffusion model [34]. In this work, we first seek \\(\\Psi_{\\mathrm{resp}}\\) mentioned in § 3.1 by devising add-on functionalities for responsible image generation while targeting \\(\\mathcal{E}(.)\\) and \\(\\mathcal{D}(.)\\) individually. In other words, we aim for \\(\\mathcal{E}_{\\mathrm{resp}}(.)\\) and \\(\\mathcal{D}_{\\mathrm{resp}}(.)\\). Both of these potentially 'enhanced' sub-models can also be combined for \\(\\Psi_{\\mathrm{resp}}\\). We target the text encoder and diffusion model individually because in T2I modeling, both of them are known to effectively encode high-level semantic information in their embedding and latent spaces [25, 16, 50]. This provides us the possibility to plug-in our intended add-on modules to any of these models for responsible T2I generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.892, + 0.751 + ], + "angle": 0, + "content": "For both of the desired \\(\\mathcal{E}_{\\mathrm{resp}}(.)\\) and \\(\\mathcal{D}_{\\mathrm{resp}}(.)\\), the framework underlying our approach remains the same at the conceptual level. We keep the target sub-model \\((\\mathcal{E}\\) or \\(\\mathcal{D})\\) in the T2I pipeline frozen, and learn an add-on module (respectively, \\(\\psi_{\\mathcal{E}}\\) or \\(\\psi_{\\mathcal{D}}\\)) by distilling the target sub-model with knowledge distillation [25]. The knowledge is distilled by conditioning the process on the set \\(\\mathcal{A}_X\\) which covers the aspects along which the model needs to be made responsible. Moreover, to ensure that the distilled concepts are disentangled as much as possible, we apply concept whitening [9] to them. The eventual T2I signal gets modulated by a simple addition of the signal from the add-on modules for responsible generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.892, + 0.799 + ], + "angle": 0, + "content": "We explain the exact procedure for obtaining \\(\\psi_{\\mathcal{E}}\\) for \\(\\mathcal{E}_{\\mathrm{resp}}(.)\\), and \\(\\psi_{\\mathcal{D}}\\) for \\(\\mathcal{D}_{\\mathrm{resp}}(.)\\) in § 3.3 and 3.4, respectively. Combining them for \\(\\Psi_{\\mathrm{resp}}\\) is explained in § 3.5." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.805, + 0.836, + 0.822 + ], + "angle": 0, + "content": "3.3. Responsible Interpretable Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our add-on module \\(\\psi_{\\mathcal{E}}\\) for \\(\\mathcal{E}_{\\mathrm{resp}}(.)\\) is marked by not only responsible generation but also interpretability of the embedding space distilled from the source \\(\\mathcal{E}(.)\\). Hence, we term the devised module as RICE - Responsible and Interpretable CLIP Embedding. The RICE module treats the original text" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2978" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.098, + 0.332, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.244, + 0.285, + 0.256 + ], + "angle": 0, + "content": "(a) RICE Embeddings Space" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.271, + 0.332, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.433, + 0.3, + 0.445 + ], + "angle": 0, + "content": "(b) RICE Controlled T2I Outputs" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.097, + 0.89, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.459, + 0.893, + 0.584 + ], + "angle": 0, + "content": "Figure 2. (a) Illustration of responsible concept space learned by the RICE module, aiming to generate \\( z_{\\mathrm{clip}}^{\\mathrm{final}} \\) embeddings along the responsible aspects in \\( \\mathcal{A}_{\\mathrm{X}} \\). Interconnected fairness related subspaces for race, gender, and age are illustrated. (b) Example outputs of RICE controlled T2I process, demonstrating a range of modulation along the concepts, which is possible due to the multi-facet control enabled by the RICE embedding space. (c) Depiction of the distillation process used for the RIIDL module, where the student (RIIDL) learns from the teacher (Diffusion Model) over \\( \\tau \\) time-steps. At each step, the loss - see Eq. (6) - aligns the RIIDL module with the aspects of \\( \\mathcal{A}_{\\mathrm{X}} \\). (d) At inference, textual embeddings get fed into RIIDL. We illustrate control over \\( \\mathcal{A}_{\\mathrm{age}} \\). The latent vectors from RIIDL are injected into the T2I Diffusion Model noisy latents. We explore this injection at varying time-step (differentiated vertically in the figure). The best control is observed in the early stages (\\( \\tau = 0 \\) to \\( 30\\% \\)). Additional examples provided in the supplementary material. (e) Our method allows a scalable control over all concepts covered in \\( \\mathcal{A}_{\\mathrm{X}} \\) in a composite manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.47, + 0.645 + ], + "angle": 0, + "content": "encoder in the T2I pipeline as a teacher and distills it for the concepts in \\(\\mathcal{A}_X\\) by training \\(\\psi_{\\mathcal{E}}\\) while minimizing the Expected value of the following per-batch loss:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.659, + 0.469, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K D - c l i p}} = \\frac {1}{| B |} \\sum_ {k = 1} ^ {| B |} \\left\\| z _ {\\mathrm {c l i p}, k} ^ {\\mathrm {T}} - z _ {\\mathrm {c l i p}, k} ^ {\\text {d i s t i l l}} \\right\\| ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.469, + 0.809 + ], + "angle": 0, + "content": "where \\(B\\) denotes the batch, \\(z_{\\mathrm{clip},k}^{\\mathrm{T}}\\) and \\(z_{\\mathrm{clip},k}^{\\mathrm{distill}}\\) are the teacher and distilled student embeddings for the \\(k^{\\mathrm{th}}\\) sample, respectively. The \\(z_{\\mathrm{clip}}^{\\mathrm{T}}\\) embeddings are computed using an automated prompt generation following \\(\\mathcal{A}_X\\). To keep the flow of discussion, we provide details about that process and the architecture of \\(\\psi_{\\mathcal{E}}\\) in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.871 + ], + "angle": 0, + "content": "At this stage, \\(\\mathbf{z}_{\\mathrm{clip}}^{\\mathrm{distll}} \\in \\mathbb{R}^{d \\times n}\\) are the embedding representations conditioned on \\(\\mathcal{A}_X\\). We additionally apply concept whitening [9] to further decorrelate these distilled embeddings using the following transform:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.884, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {\\mathrm {c l i p}} ^ {\\mathrm {z c a}} = \\phi (\\mathbf {z} _ {\\mathrm {c l i p}} ^ {\\mathrm {d i s t i l l}}) = \\mathcal {W} (\\mathbf {z} _ {\\mathrm {c l i p}} ^ {\\mathrm {d i s t i l l}} - \\mu), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.599, + 0.892, + 0.663 + ], + "angle": 0, + "content": "where \\(\\mu = \\frac{1}{n}\\sum_{i=1}^{n}z_{i}\\), and \\(\\mathcal{W}_{d\\times d}\\) is the whitening matrix that obeys \\(\\mathcal{W}^T\\mathcal{W} = \\Sigma^{-1}\\). Here, \\(\\Sigma_{d\\times d} = \\frac{1}{n}(z_{\\mathrm{clip}}^{\\mathrm{distill}} - \\mu\\mathbf{1}^{\\intercal})(z_{\\mathrm{clip}}^{\\mathrm{distill}} - \\mu\\mathbf{1}^{\\intercal})^{\\intercal}\\) is the covariance matrix. The final embeddings for responsible generation get computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.675, + 0.891, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {\\text {c l i p}} ^ {\\text {r e s p}} = \\alpha \\cdot \\mathbf {z} _ {\\text {c l i p}} ^ {\\text {d i s t i l l}} + (1 - \\alpha) \\cdot \\mathbf {z} _ {\\text {c l i p}} ^ {\\text {z c a}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.707, + 0.891, + 0.738 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is kept for balancing. At inference, the embedding \\(z_{\\mathrm{clip}}\\) of the original prompt gets modulated as" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.752, + 0.891, + 0.783 + ], + "angle": 0, + "content": "\\[\nz _ {\\text {c l i p}} ^ {\\text {f i n a l}} = z _ {\\text {c l i p}} + \\sum_ {k} \\gamma_ {k} \\mathbf {z} _ {k, \\text {c l i p}} ^ {\\text {r e s p}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\gamma_{k}\\) provides control over the influence of each concept on the embedding. With this, the embedding of a prompt can be simultaneously modulated to multiple responsible concepts represented in the continuous representation space of \\(\\psi_{\\mathcal{E}}\\). Fig. 2a illustrates this notion, with Fig. 2b showing the RICE impact on varying the eventual T2I output along different responsible concepts." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2979" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.376, + 0.107 + ], + "angle": 0, + "content": "3.4. Responsible Interpretable Latents" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.212 + ], + "angle": 0, + "content": "In regards to \\(\\mathcal{D}_{\\mathrm{resp}}(.)\\), we follow a similar process as adopted in § 3.3 for \\(\\mathcal{E}_{\\mathrm{resp}}(.)\\), tailoring the former for the diffusion model component \\(\\mathcal{D}\\) of the T2I pipeline. Our add-on module \\(\\psi_{D}\\), again a neural model, gets trained using the intermediate latent representations of \\(\\mathcal{D}\\), distilled by conditioning it on \\(\\mathcal{A}_X\\), hence; we term it RIIDL - Responsible Interpretable Intermediate Diffusion Latents." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.468, + 0.32 + ], + "angle": 0, + "content": "The latent representation of a diffusion model gets updated for each time-step \\((\\tau)\\) of a Markov process denoising the image previously corrupted in the forward diffusion process. For brevity, we refrain from discussing minutiae of diffusion modeling - interested readers are referred to [12]. Here, we focus on the key idea of inducing our plug-in \\(\\psi_{D}\\) for the process. To train \\(\\psi_{D}\\), we define the following loss" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.33, + 0.468, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K D - u n e t}} = \\mathbb {E} _ {\\tau \\sim \\mathcal {U} [ 0, 1 ]} \\left[ \\omega \\left(\\lambda_ {\\tau}\\right) \\left\\| z _ {\\text {u n e t}} ^ {\\mathrm {T}, \\tau} - z _ {\\text {u n e t}} ^ {\\text {d i s t i l l}, \\tau} \\right\\| ^ {2} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.468, + 0.48 + ], + "angle": 0, + "content": "where \\(\\tau \\sim \\mathcal{U}[0,1]\\) refers to uniform sampling of time-steps, \\(z_{\\mathrm{unet}}^{\\mathrm{T},\\tau}\\) and \\(z_{\\mathrm{unet}}^{\\mathrm{distill},\\tau}\\) represent the latent representations of the teacher and student models at time-step \\(\\tau\\), \\(\\lambda_{\\tau}\\) reflects the SNR of the signal at \\(\\tau\\), and \\(\\omega(\\lambda_{\\tau})\\) is a pre-specified weighting function. We follow [21] for implementing the SNR and \\(\\omega(.)\\). In this setup, \\(\\psi_{D}\\) is the student while the T2I diffusion model is the teacher." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.468, + 0.754 + ], + "angle": 0, + "content": "In Eq. (6), the subscript 'unet' emphasizes the U-Net architecture of the student and teacher, which is commonly followed in denoising diffusion models. Details of our student model \\(\\psi_{\\mathcal{D}}\\) are provided in the supplementary material. Similar to \\(z_{\\mathrm{clip}}^{\\mathrm{T}}\\) in Eq. (2), the \\(z_{\\mathrm{unet}}^{\\mathrm{T}}\\) at \\(\\tau\\) is generated based on \\(\\mathcal{A}_X\\), which conditions \\(\\psi_{\\mathcal{D}}\\) on the responsible concepts being considered by the user. We further follow the processing of \\(z_{\\mathrm{unet}}^{\\mathrm{T}}\\) at a given \\(\\tau\\) corresponding to Eq. (3), (4) and (5) to respectively compute \\(\\mathbf{z}_{\\mathrm{unet}}^{\\mathrm{zca}}\\), \\(\\mathbf{z}_{\\mathrm{unet}}^{\\mathrm{resp}}\\), and \\(z_{\\mathrm{unet}}^{\\mathrm{final}}\\) at \\(\\tau\\). Fig. 2c illustrate the distillation process to train \\(\\psi_{\\mathcal{D}}\\) for the RIIDL module on receiving the textual embedding input from \\(\\mathcal{E}\\). In Fig. 2d, an illustration is provided for the deployed model, where RIIDL is able to guide the generated image by injecting signals to the early denoising stages of the T2I diffusion model to encourage generation along responsible the concept in \\(\\mathcal{A}_X\\). Figure 2e further shows that extending \\(\\mathcal{A}_X\\) enables simultaneous control over multiple concepts because \\(\\psi_{\\mathcal{D}}\\) gets conditioned on multiple concepts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.76, + 0.292, + 0.777 + ], + "angle": 0, + "content": "3.5. Dual-Space Integration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "The methods discussed in § 3.3 and § 3.4 provide add-on modules \\(\\mathcal{E}_{\\mathrm{resp}}(\\cdot)\\) and \\(\\mathcal{D}_{\\mathrm{resp}}(\\cdot)\\) in the form of plug-and-play models \\(\\psi_{\\mathcal{E}}\\) and \\(\\psi_{\\mathcal{D}}\\) that can be directly plugged into the text encoder \\(\\mathcal{E}\\) and diffusion model \\(\\mathcal{D}\\) of the T2I pipeline. Whereas both modules are usable standalone, we also integrate them in this work for a comprehensive dual space control on the generative outputs. When both modules are integrated into the T2I pipeline, the overall function of gen" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "erating a responsible image \\(I\\) from a prompt \\(t\\in \\mathcal{T}\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.131, + 0.892, + 0.149 + ], + "angle": 0, + "content": "\\[\n\\Psi_ {\\text {r e s p}} (t) = \\lambda_ {\\mathcal {E}} \\cdot \\mathcal {D} \\left(\\mathcal {E} _ {\\text {r e s p}} (t)\\right) + \\lambda_ {\\mathcal {D}} \\cdot \\mathcal {D} _ {\\text {r e s p}} (\\mathcal {E} (t)), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.892, + 0.262 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathcal{E}}\\) and \\(\\lambda_{\\mathcal{D}}\\) are weighting factors following the relation \\(\\lambda_{\\mathcal{E}} + \\lambda_{\\mathcal{D}} = 1\\). The integrated \\(\\Psi_{\\mathrm{resp}}(t)\\) provides effective control over high-level semantic information in the embedding and latent spaces. It provides a comprehensive plug-and-play method for the T2I pipeline for altering the behavior of \\(\\Psi\\) to follow concepts in \\(\\mathcal{A}_X\\) for responsible text-to-image generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.268, + 0.616, + 0.282 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.285, + 0.892, + 0.436 + ], + "angle": 0, + "content": "Our method is evaluated for a variety of concepts related to fair and safe image generation. The used evaluation protocols are based on standardized benchmarks. For reproducibility, further details on implementation, dataset descriptions, and configurations of hardware and software libraries are also provided in the supplementary material. Our experiments use the Stable Diffusion v1.4 model [35]. We compare with SLD [38], ESD [13], UCE [14], and Vector-SD [23], using WinoBias [52] and I2P [38] datasets, which contain fairness-sensitive and safety-critical scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Fair Generation Evaluation: In regards to the fair generation, we evaluate performance across professions with documented gender and racial biases. The dataset for this evaluation is WinoBias [52], which comprises 35 professions historically associated with gender bias. Fairness is quantitatively assessed using the deviation ratio \\(\\Delta\\) as in literature [28, 11, 23], which measures the disparity in generated representations across specified responsible aspects by comparing the model's output distribution against an expected baseline. Formally, \\(\\Delta\\) is defined as: \\(\\Delta = \\frac{\\max_{a\\in\\mathcal{A}}\\left|\\frac{N_a}{N} - \\frac{1}{|\\mathcal{A}|}\\right|}{1 - \\frac{1}{|\\mathcal{A}|}}\\), where \\(\\mathcal{A}\\) is the set of attributes within a responsible aspect (e.g., gender or race), \\(N\\) represents the total number of generated images, and \\(N_{a}\\) denotes the count of images for which the highest probability attribute prediction corresponds to attribute \\(a\\). For extensive evaluation, we further employ \"challenging prompts\" [23] in our experiments, which are known to produce bias representations towards male depictions [14]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.887 + ], + "angle": 0, + "content": "Safe Generation Evaluation: We also evaluate the generated content against the standard criteria for identifying harmful or culturally sensitive outputs [38, 23]. Images are flagged as inappropriate if they depict or imply content associated with sexual content, self-harm, hate, illegal activity, shock, harassment, or violence. Our evaluation integrates both the NudeNet detector1 and Q16 classifier [37], as per common practices in responsible generation assessment, to systematically flag content deemed inappropriate [38, 23]. We examine a diverse range of images to rigorously assess robustness across varied outputs." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.887, + 0.735, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/notAI-tech/NudeNet" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2980" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.089, + 0.867, + 0.119 + ], + "angle": 0, + "content": "Table 1. Bias generation quantified by deviation ratio \\( \\left( {0 \\leq \\Delta \\leq 1}\\right) \\) . Lower values indicate better performance. Results provided for Gender and Race bias across standard and extended (Gender-Pro/Race-Pro) Winobias datasets [52]." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.129, + 0.865, + 0.252 + ], + "angle": 0, + "content": "
AttributeGender(↓)Gender-Pro(↓)Race(↓)Race-Pro(↓)
SD [35]U [14]V [23]OursSD [35]U [14]V [23]OursSDU [14]V [23]OursSD [35]U [14]V [23]Ours
CEO0.920.280.150.040.900.580.210.130.380.130.220.080.310.080.220.12
Doctor0.920.200.080.050.520.320.100.100.920.070.260.070.590.520.150.12
Nurse1.000.390.960.040.980.840.430.180.760.250.300.070.390.790.080.14
Receptionist0.840.380.880.040.980.960.860.170.880.100.360.070.740.140.250.11
Teacher0.300.060.510.040.480.160.070.130.510.100.040.080.260.230.210.10
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.102, + 0.253, + 0.626, + 0.266 + ], + "angle": 0, + "content": "SD: Standard Diffusion Model, U: Unified Concept Editing [14], V: Vector Interpret Diffusion [23]." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.271, + 0.849, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.556, + 0.895, + 0.6 + ], + "angle": 0, + "content": "Figure 3. Pair-wise comparison for the Stable Diffusion (SD) baseline and our plugin to it for responsible generation along age, gender and race. SD images contain stereotypical associations to the profession, which get removed by our method by often accounting for more than one responsible attributes, while maintaining high image quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.613, + 0.283, + 0.63 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.471, + 0.804 + ], + "angle": 0, + "content": "Fair Generation: In Table 1, we present the bias quantification results using the deviation ratio \\((0 \\leq \\Delta \\leq 1)\\) across gender and racial attributes, evaluated under both the standard (Gender and Race) and extended (Gender-Pro and Race-Pro) settings of the WinoBias dataset [52]. Lower values of \\(\\Delta\\) indicate reduced demographic bias. Overall, our method consistently achieves highly competitive results. Even under the Gender-Pro setting, where prompts are specifically crafted to amplify stereotypical associations, our method maintains its robustness. For racial bias, our method similarly maintains its performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Table 2 further quantifies the debiasing efficacy of the our approach across professions with known demographic skew, displaying results for five professions and the average deviation across all 35 professions in the WinoBias [52]. Results show a substantial bias reduction by our method. Consistent bias reduction demonstrates that bias mitiga" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.721 + ], + "angle": 0, + "content": "tion is achieved by our method independent of profession-specific attributes. Averaged over all professions (last row), our method provides a significant gain over the baseline methods. Such efficacy of our method comes from its ability to enable debiasing using a collective set of attributes in \\(\\mathcal{A}_X\\) and modulate the T2I pipeline to generate images responsibly considering all those attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To show how this manipulation varies the images generated by Stable Diffusion (SD), in Fig. 3 we provide representative examples. The figure shows a range of professions for which pairs of images are generated using the same seeds. The SD images show stereotypical associations to the professions. Our method is plugged in to SD with set \\(\\mathcal{A}_X\\) collectively containing race, gender and age attributes. It is observable in the provided representative examples that images get modulated towards these responsible concepts, often incorporating more than one responsible concept. We also provide further results in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2981" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.082, + 0.892, + 0.11 + ], + "angle": 0, + "content": "Table 2. Debaising performance across 5 randomly selected professions and the average of all 35 professions in Winobias dataset [52]. The metric \\( \\Delta = 0 \\) indicate ideal debraising. Our method shows the lowest average deviation compared to previous approaches." + }, + { + "type": "table", + "bbox": [ + 0.188, + 0.112, + 0.782, + 0.234 + ], + "angle": 0, + "content": "
ProfessionOriginal-SDConcept Algebra [44]Debias-VL [11]TIME [28]TIMEP [14]Ours
Librarian0.86 ± 0.060.66 ± 0.070.34 ± 0.060.26 ± 0.050.35 ± 0.010.04 ± 0.02
Teacher0.42 ± 0.010.46 ± 0.000.11 ± 0.050.34 ± 0.060.07 ± 0.060.05 ± 0.01
Sheriff0.99 ± 0.010.38 ± 0.220.82 ± 0.080.22 ± 0.050.10 ± 0.050.06 ± 0.03
Analyst0.58 ± 0.120.24 ± 0.180.71 ± 0.020.52 ± 0.030.13 ± 0.050.07 ± 0.02
Doctor0.78 ± 0.040.40 ± 0.020.50 ± 0.040.58 ± 0.030.41 ± 0.080.06 ± 0.01
Average0.67 ± 0.010.43 ± 0.010.55 ± 0.010.44 ± 0.000.31 ± 0.000.05 ± 0.02
" + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.241, + 0.18, + 0.253 + ], + "angle": 0, + "content": "A doctor" + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.259, + 0.199, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.24, + 0.269, + 0.252 + ], + "angle": 0, + "content": "Agender" + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.259, + 0.285, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.319, + 0.285, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.381, + 0.265, + 0.39 + ], + "angle": 0, + "content": "Female" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.259, + 0.366, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.319, + 0.366, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.381, + 0.398, + 0.391 + ], + "angle": 0, + "content": "Asian - Black" + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.239, + 0.383, + 0.252 + ], + "angle": 0, + "content": "Arace" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.259, + 0.441, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.32, + 0.441, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.24, + 0.538, + 0.254 + ], + "angle": 0, + "content": "Aage" + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.259, + 0.598, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.32, + 0.598, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.381, + 0.549, + 0.391 + ], + "angle": 0, + "content": "Young - Old" + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.24, + 0.713, + 0.253 + ], + "angle": 0, + "content": "\\(\\mathrm{A}_{\\text {gender }} + \\mathrm{A}_{\\text {race }}\\)" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.259, + 0.753, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.32, + 0.753, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.381, + 0.731, + 0.391 + ], + "angle": 0, + "content": "Female/Black-Asian" + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.241, + 0.848, + 0.254 + ], + "angle": 0, + "content": "\\(\\mathrm{A}_{\\text {gender }} + \\mathrm{A}_{\\text {race }} + \\mathrm{A}_{\\text {age }}\\)" + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.259, + 0.835, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.319, + 0.835, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.38, + 0.841, + 0.389 + ], + "angle": 0, + "content": "Female / Asian / Old" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.395, + 0.894, + 0.451 + ], + "angle": 0, + "content": "Figure 4. Controlled composite responsible generation using the proposed method. By using different concepts in \\(\\mathcal{A}_X\\) in Eq. (1), and employing dual space control in (7), our technique can enable responsibility along single or multiple concepts, as desired. Provided the prompt \"A doctor\" to Stable Diffusion, alignment is achieved for a target concept composition (top label) for the attributes noted at the bottom of each image set. See supplementary material for more examples." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.461, + 0.469, + 0.488 + ], + "angle": 0, + "content": "Table 3. Proportion of images classified as inappropriate on I2P benchmark [38]. Lower values are more desirable." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.49, + 0.443, + 0.62 + ], + "angle": 0, + "content": "
CategorySDV-SD [23]SLD [38]ESD [13]Ours
Sexual0.390.230.160.190.15
Violence0.450.310.220.420.24
Hate0.420.300.190.340.18
Harassment0.350.190.160.280.19
Illegal0.350.240.180.340.14
Shocking0.530.380.270.430.28
Self-harm0.450.290.200.360.20
Avg.0.410.270.200.320.20
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Safe Generation: To benchmark safe and appropriate nature of the generated content, we evaluate our method on the I2P [38], which classifies generated images across categories covering a range of inappropriate content. Table 3 benchmarks our method against the state-of-the-art methods following standard protocol [38], [13]. As seen, our technique significantly reduces inappropriateness of the content across all categories. Generally, maintaining superior performance against the methods. On average, our performance is similar to SLD [38], which is a dedicated 'safe generation' method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.803, + 0.355, + 0.819 + ], + "angle": 0, + "content": "6. Further Results & Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Multi-Concept Composition and Control: Our technique allows responsible generation with multi-concept composition. Multiple concepts can compose \\(\\mathcal{A}_X\\), which enables the hierarchical function \\(\\Psi_{\\mathrm{resp}}\\) to combine them in latent and embedding space of the T2I pipeline. Figure 4 demonstrates" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.463, + 0.887, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.721 + ], + "angle": 0, + "content": "Figure 5. Examples of extending \\(\\mathcal{A}_{\\mathrm{resp}}\\) with additional attributes - possibly unrelated to core responsible concepts. Extensions with smile and glasses are shown. In addition to varying images along responsible concepts of age, race and gender, our method is able to seamlessly incorporate the additional concepts in images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.839 + ], + "angle": 0, + "content": "this capability within \\(\\Psi_{\\mathrm{resp}}\\), showing how distinct responsible aspects, such as age, gender, and race; can be independently learned and then composed to produce varied yet coherent outputs. By manipulating combinations within \\(\\mathcal{A}_X\\), we demonstrate an excellent control achieved with interpretable variations across the composite attribute space, effectively balancing diversity with responsible aspects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Scalability of Concept Spaces: In our original experiments, the responsible concept set \\(\\mathcal{A}_{\\mathrm{resp}}\\) in our framework, as defined in Eq. (1), included core attributes of \\(\\mathcal{A}_{\\mathrm{race}}, \\mathcal{A}_{\\mathrm{gender}}, \\mathcal{A}_{\\mathrm{age}},\\) and \\(\\mathcal{A}_{\\mathrm{safe}}\\). This was only to benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2982" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.47, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.218, + 0.47, + 0.304 + ], + "angle": 0, + "content": "Figure 6. Representative curtailed example of transitions caused by RICE and RIIDL to an output. An image of 'a doctor' is generated and 'female' concept is added individually by the modules. Increasingly changing the influence of the added concept representation provided by each module leads to unique alteration to the generated output." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.522 + ], + "angle": 0, + "content": "the method with the existing approaches. Our design allows the \"responsible\" set to be extendable. This scalability allows for seamlessly handling additional nuanced attributes in images, such as facial expressions (e.g., smiling) and accessories (e.g., glasses). Such concepts can act as supplementary attributes with the existing attributes, enabling complex yet interpretable multi-aspect control. In Fig. 5, we provide representative examples of scaling up the considered concepts set. It leads to combining the responsible subspaces of the core responsible features with subspaces for smiling and glasses. It is observable that our framework is scalable to not only core responsibility related concepts, but to any semantically meaningful concepts. We also provide additional results in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.471, + 0.853 + ], + "angle": 0, + "content": "Individual Modules: Our method takes advantage of two modules, namely; of RICE (§ 3.3) and RIIDL (§ 3.4). To assess the specific contribution of each component within the responsible generation framework, we perform a systematic ablation by individually removing and adjusting key modules. This analysis allows selectively varying the influence both modules on the output. When only RICE operates, modifications originally occur in the embedding concept space for the T2I pipeline, facilitating transitions from base to responsible concept spaces. The RIIDL module contribute an additional layer of fine-grained control, allowing nuanced and context-sensitive adjustments within the intermediate latents during diffusion. This capability supports precise tuning of details without overriding the global context, complementing the broader adjustments. In tandem, these modules establish a dynamic balance within the dual-space intervention strategy, where RICE provides prediffusion global control, and RIIDL governs the responsible aspects during the denoising process. Whereas we provide further results supporting these point in the supplementary material, in Fig. 6 a curtailed representative example is provided to show the visual effects." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In the figure, we generate images using prompt 'a doctor' and with an added concept 'female' from the individual modules. The 'female' concept representation \\(\\mathbf{z}^{\\mathrm{female}}\\) pro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "vided by the modules is added to the original corresponding representations with an increasing degree of influence on the output. It can be observed that for RICE, this results in more prominent shifts in appearance and attire, indicating a significant control. For RIIDL, the latent space provides more refined transitions with subtle and smoother progression of the visual features. On one hand, these results demonstrate the complementary nature of the proposed modules, on the other; they establish the effectiveness of the modules as stand-alone add-ons." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.249, + 0.62, + 0.265 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.892, + 0.543 + ], + "angle": 0, + "content": "This research presented a unique approach to control generative outputs of text-to-image (T2I) models in 'fair and safe' domain. Our method induces two plug-and-play modules that can be plugged into the textual embedding space and diffusion latent space of the T2I pipeline. The modules are neural models learned by distilling the text encoder and diffusion model of the original pipeline while conditioning the knowledge distillation on a pre-defined responsible concepts set, as desired by the user. Our modules also account for semantic interpretability by leveraging concept whitening. We demonstrate that both modules can explicitly contribute to responsible image generation in a complementary manner. Our extensive benchmarking with state-of-the-art methods for fair and safe image generation on two standard datasets show highly promising results. We also show that our method can conveniently deal with multiple concepts in the responsible image space while being seamlessly extendable to more broader concepts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.816 + ], + "angle": 0, + "content": "Limitations and ethics concerns: A potential shortcoming of our method is that, to enable a comprehensive control over the generative output, it employs multiple hyperparameters, e.g., \\(\\lambda\\)'s in Eq. (7). Tuning these hyperparameters can potentially be seen as a limitation of our approach. However, the main purpose of allowing these adjustable hyper-parameters in our design is to provide a finer control to the user. It is convenient to adjust their values because the controls they provide have clear high-level interpretations. We provide guidelines on the hyper-parameter value adjustment in the supplementary material. From the ethics viewpoint, although our aim is to ensure responsible outputs; a potential misuse of exploiting the strong output control of our technique can be by replacing the \\(A_{\\mathrm{resp}}\\) with another set \\(A_{\\mathrm{irresp}}\\). In \\(A_{\\mathrm{irresp}}\\), unsafe and unfair concepts can be included. This can have an opposite impact of our technique. Hence, our method should only be used when the user has complete control over the set \\(A_{\\mathrm{resp}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.817, + 0.892, + 0.892 + ], + "angle": 0, + "content": "Acknowledgment: Naveed Akhtar is a recipient of the Australian Research Council Discovery Early Career Researcher Award (project # DE230101058) funded by the Australian Government. This work is also partially supported by Google Research Scholar Program Award." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2983" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.091, + 0.167, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.113, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Sandhini Agarwal et al. \"Evaluating CLIP: towards characterization of broader capabilities and downstream implications\". In: arXiv preprint arXiv:2108.02818 (2021)." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.16, + 0.471, + 0.203 + ], + "angle": 0, + "content": "[2] Zhongjie Ba et al. \"SurrogatePrompt: Bypassing the Safety Filter of Text-To-Image Models via Substitution\". In: arXiv preprint arXiv:2309.14122 (2023)." + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.207, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[3] Hritik Bansal et al. \"How well can Text-to-Image Generative Models understand Ethical Natural Language Interventions?\" In: arXiv preprint arXiv:2210.15230 (2022)." + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.254, + 0.469, + 0.295 + ], + "angle": 0, + "content": "[4] Hugo Berg et al. \"A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning\". In: arXiv preprint arXiv:2203.11933 (2022)." + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.3, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[5] Federico Bianchi et al. \"Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale\". In: arXiv preprint arXiv:2211.03759 (2022)." + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.346, + 0.469, + 0.402 + ], + "angle": 0, + "content": "[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. \"Multimodal datasets: misogyny, pornography, and malignant stereotypes\". In: arXiv preprint arXiv:2110.01963 (2021)." + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.407, + 0.469, + 0.448 + ], + "angle": 0, + "content": "[7] Manuel Brack et al. \"The Stable Artist: Steering Semantics in Diffusion Latent Space\". In: arXiv preprint arXiv:2212.06013 (2022)." + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.453, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[8] Nicholas Carlini et al. Extracting Training Data from Diffusion Models. 2023. DOI: 10.48550/ARXIV.2301.13188. URL: https://arxiv.org/abs/2301.13188." + }, + { + "type": "ref_text", + "bbox": [ + 0.09, + 0.514, + 0.469, + 0.556 + ], + "angle": 0, + "content": "[9] Zhi Chen, Yijie Bei, and Cynthia Rudin. \"Concept Whitening for Interpretable Image Recognition\". In: Nature Machine Intelligence 2.12 (Dec. 2020), pp. 772-782." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.56, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[10] Jaemin Cho, Abhay Zala, and Mohit Bansal. “Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models”. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 3043–3054." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.635, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[11] Ching-Yao Chuang et al. \"Debiasing vision-language models via biased prompts\". In: arXiv preprint arXiv:2302.00070 (2023)." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.681, + 0.469, + 0.723 + ], + "angle": 0, + "content": "[12] Florinel-Alin Croitoru et al. \"Diffusion models in vision: A survey\". In: IEEE Transactions on Pattern Analysis and Machine Intelligence 45.9 (2023), pp. 10850-10869." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.727, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[13] Rohit Gandikota et al. “Erasing concepts from diffusion models”. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 2426–2436." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.774, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[14] Rohit Gandikota et al. \"Unified concept editing in diffusion models\". In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024, pp. 5111-5120." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.834, + 0.469, + 0.876 + ], + "angle": 0, + "content": "[15] Jindong Gu et al. \"Towards robust prompts on vision-language models\". In: arXiv preprint arXiv:2304.08479 (2023)." + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.113, + 0.471, + 0.876 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[16] René Haas et al. \"Discovering interpretable directions in the semantic latent space of diffusion models\". In: 2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG). IEEE. 2024, pp. 1-9." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.152, + 0.892, + 0.194 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. “Denoising diffusion probabilistic models”. In: Advances in neural information processing systems 33 (2020), pp. 6840–6851." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.199, + 0.892, + 0.24 + ], + "angle": 0, + "content": "[18] Chi-Pin Huang et al. \"Receler: Reliable concept erasing of text-to-image diffusion models via lightweight erasers\". In: arXiv preprint arXiv:2311.17717 (2023)." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.244, + 0.892, + 0.298 + ], + "angle": 0, + "content": "[19] Tero Karras et al. “Elucidating the design space of diffusion-based generative models”. In: Advances in neural information processing systems 35 (2022), pp. 26565–26577." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.303, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[20] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. \"Diffusionclip: Text-guided diffusion models for robust image manipulation\". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022, pp. 2426-2435." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.377, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[21] Diederik Kingma et al. \"Variational diffusion models\". In: Advances in neural information processing systems 34 (2021), pp. 21696-21707." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.423, + 0.892, + 0.478 + ], + "angle": 0, + "content": "[22] Nupur Kumari et al. \"Ablating concepts in text-to-image diffusion models\". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 22691-22702." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.482, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[23] Hang Li et al. \"Self-discovering interpretable diffusion latent directions for responsible text-to-image generation\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.542, + 0.892, + 0.584 + ], + "angle": 0, + "content": "[24] Sasha Luccioni et al. \"Stable bias: Evaluating societal representations in diffusion models\". In: Advances in Neural Information Processing Systems 36 (2024)." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.588, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[25] Chenlin Meng et al. \"On distillation of guided diffusion models\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 14297-14306." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.649, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[26] Ranjita Naik and Besmira Nushi. \"Social Biases through the Text-to-Image Generation Lens\". In: arXiv preprint arXiv:2304.06034 (2023)." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.694, + 0.892, + 0.736 + ], + "angle": 0, + "content": "[27] Alex Nichol et al. \"Glide: Towards photorealistic image generation and editing with text-guided diffusion models\". In: arXiv preprint arXiv:2112.10741 (2021)." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.739, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[28] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. \"Editing implicit assumptions in text-to-image diffusion models\". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 7053-7061." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.8, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[29] Yong-Hyun Park et al. \"Direct unlearning optimization for robust and safe text-to-image models\". In: arXiv preprint arXiv:2407.21035 (2024)." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] William Peebles and Saining Xie. \"Scalable diffusion models with transformers\". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 4195-4205." + }, + { + "type": "list", + "bbox": [ + 0.505, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2984" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[31] Alec Radford et al. \"Learning Transferable Visual Models From Natural Language Supervision\". In: arXiv preprint arXiv:2103.00020 (2021)." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.138, + 0.472, + 0.18 + ], + "angle": 0, + "content": "[32] Aditya Ramesh et al. \"Hierarchical text-conditional image generation with clip latents\". In: arXiv preprint arXiv:2204.06125 1.2 (2022), p. 3." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.184, + 0.47, + 0.226 + ], + "angle": 0, + "content": "[33] Javier Rando et al. Red-Teaming the Stable Diffusion Safety Filter. 2022. DOI: 10.48550/ARXIV.2210.04610. URL: https://arxiv.org/abs/2210.04610." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.23, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[34] Robin Rombach et al. \"High-resolution image synthesis with latent diffusion models\". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022, pp. 10684-10695." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.29, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[35] Robin Rombach et al. \"High-resolution image synthesis with latent diffusion models\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022, pp. 10684-10695." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.349, + 0.47, + 0.405 + ], + "angle": 0, + "content": "[36] Chitwan Sahara et al. \"Photorealistic text-to-image diffusion models with deep language understanding\". In: Advances in neural information processing systems 35 (2022), pp. 36479-36494." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.409, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[37] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. \"Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content?\" In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. 2022, pp. 1350-1361." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.497, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[38] Patrick Schramowski et al. \"Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 22522-22531." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.556, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[39] Shoaib Ahmed Siddiqui et al. \"Metadata Archaeology: Unearthing Data Subsets by Leveraging Training Dynamics\". In: arXiv preprint arXiv:2209.10015 (2022)." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.602, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[40] Gowthami Somepalli et al. \"Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models\". In: arXiv preprint arXiv:2212.03860 (2022)." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.648, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[41] Lukas Struppek, Dominik Hintersdorf, and Kristian Kersting. \"The biased artist: Exploiting cultural biases via homoglyphs in text-guided image generation models\". In: (2022)." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.707, + 0.469, + 0.749 + ], + "angle": 0, + "content": "[42] Yu-Lin Tsai et al. \"Ring-A-Bell! How Reliable are Concept Removal Methods for Diffusion Models?\" In: arXiv preprint arXiv:2310.10012 (2023)." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.754, + 0.469, + 0.796 + ], + "angle": 0, + "content": "[43] Peiran Wang et al. \"Moderator: Moderating Text-to-Image Diffusion Models through Fine-grained Context-based Policies\". In: arXiv preprint arXiv:2408.07728 (2024)." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.8, + 0.469, + 0.841 + ], + "angle": 0, + "content": "[44] Zihao Wang et al. \"Concept algebra for text-controlled vision models\". In: arXiv preprint arXiv:2302.03693 2 (2023)." + }, + { + "type": "ref_text", + "bbox": [ + 0.082, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[45] Yongliang Wu et al. \"Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient\". In: arXiv preprint arXiv:2405.15304 (2024)." + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[46] Tianwei Xiong et al. \"Editing Massive Concepts in Text-to-Image Diffusion Models\". In: arXiv preprint arXiv:2403.13807 (2024)." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.138, + 0.892, + 0.18 + ], + "angle": 0, + "content": "[47] Yijun Yang et al. \"GuardT2I: Defending Text-to-Image Models from Adversarial Prompts\". In: arXiv preprint arXiv:2403.01446 (2024)." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.185, + 0.892, + 0.227 + ], + "angle": 0, + "content": "[48] Jaehong Yoon et al. \"SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation\". In: arXiv preprint arXiv:2410.12761 (2024)." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.232, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[49] Gong Zhang et al. “Forget-me-not: Learning to forget in text-to-image diffusion models”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024, pp. 1755–1764." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.293, + 0.892, + 0.334 + ], + "angle": 0, + "content": "[50] Hongxiang Zhang, Yifeng He, and Hao Chen. \"Steerdiff: Steering towards safe text-to-image diffusion models\". In: arXiv preprint arXiv:2410.02710 (2024)." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.339, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[51] Yimeng Zhang et al. \"To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now\". In: European Conference on Computer Vision. Springer. 2025, pp. 385-403." + }, + { + "type": "ref_text", + "bbox": [ + 0.505, + 0.399, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[52] Jieyu Zhao et al. \"Gender bias in coreference resolution: Evaluation and debiasing methods\". In: arXiv preprint arXiv:1804.06876 (2018)." + }, + { + "type": "list", + "bbox": [ + 0.505, + 0.092, + 0.892, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2985" + } + ] +] \ No newline at end of file diff --git a/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/2c24c8f1-1822-42ba-8659-5feab5b2d726_origin.pdf b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/2c24c8f1-1822-42ba-8659-5feab5b2d726_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f32016e47cf70f95d748b0cbb72f1aef7b5e7c56 --- /dev/null +++ b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/2c24c8f1-1822-42ba-8659-5feab5b2d726_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:426599ade9980b690cdfab6193c23f504e48f91c3e49b29dd796fdc33b6e5cd5 +size 8748049 diff --git a/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/full.md b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/full.md new file mode 100644 index 0000000000000000000000000000000000000000..036fa4260bce79bed0d0a0cda337bba93649af03 --- /dev/null +++ b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/full.md @@ -0,0 +1,304 @@ +# Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control + +Basim Azam Naveed Akhtar + +School of Computing and Information Systems, The University of Melbourne, Australia + +{basim.azam, naveed.akhtar1}@unimelb.edu.au + +![](images/9e4d90a3489cd24f005e5fd2733cdc53fd11648af1460046f84cdd1e4580ffe8.jpg) +Figure 1. Our unique plug-and-play interpretable approach simultaneously controls a range of concepts for responsible and fair image generation with text-to-image pipelines. Our method enables control over both text embedding space and latent diffusion space. Shown examples compare unfair and unsafe generation for the given prompts by the Stable Diffusion (top), along with their responsible counterparts resulting from our approach influencing the text encoder (middle) and the diffusion model (bottom). We control individual concepts for diverse/safe generation in these examples while modeling them in continuous composite responsible semantic spaces. + +# Abstract + +Ethical issues around text-to-image (T2I) models demand a comprehensive control over the generative content. Existing techniques addressing these issues for responsible T2I models aim for the generated content to be fair and safe (non-violent/explicit). However, these methods remain bounded to handling the facets of responsibility concepts individually, while also lacking in interpretability. Moreover, they often require alteration to the original model, which compromises the model performance. In this work, we propose a unique technique to enable responsible T2I generation by simultaneously accounting for an extensive range of concepts for fair and safe content generation in a scalable manner. The key idea is to distill the target T2I pipeline with an external plug-and-play mechanism that learns an inter + +pretable composite responsible space for the desired concepts, conditioned on the target T2I pipeline. We use knowledge distillation and concept whitening to enable this. At inference, the learned space is utilized to modulate the generative content. A typical T2I pipeline presents two plug-in points for our approach, namely; the text embedding space and the diffusion model latent space. We develop modules for both points and show the effectiveness of our approach with a range of strong results. Our code can be accessed at https://basim-azam.github.io/responsiblediffusion/ + +# 1. Introduction + +Rapid advances in text-to-image (T2I) generative pipeline are revolutionizing numerous vision applications, offering unprecedented convenience for synthesizing high quality + +visual content using textual descriptions [34, 32, 30, 19, 17]. Unfortunately, this convenience coupled with the generated content realism offered by publicly accessible T2I models, e.g., Stable Diffusion [34], DALL-E [32], Imagen [36], also has undesired ethical implications. Confronting the core values of fair and safe (non-violent/explicit) development and deployment of Artificial Intelligence, these models lack in producing generative content responsibly [15, 51]. + +The reasons of irresponsible content generation by contemporary T2I models lie deep in their training process, which must use large volumes of (mostly uncurated) data. Issues like inappropriate content and harmful stereotyping lurking in the data are passed on to the generative models through training [2, 42], which cause problems after their deployment. Hence, it is critical to develop methodologies to explicitly control generative models from producing irresponsible content. Addressing that, several techniques have emerged recently, encompassing approaches like input prompt filtering [48, 2], post-hoc content moderation [47, 43, 22], machine unlearning [45, 29, 18, 49], and model editing [14, 46]. Though effective, these approaches face limitations like requiring human-intervention, ad-hoc methodology, and selective dealing of different facets of the concepts defining 'responsible generation'. + +Another contributing factor to the ethical concerns related to T2I models is our current lack of wholesome understanding of the latent space of the diffusion models. Researchers have already started exploring interpretability [16, 50] and manipulation of diffusion model latent spaces to mitigate unintended T2I model behavior [14, 23]. Controlling the latent space holds promise when it is possible to disentangle semantic concepts in it. One recent inspiring approach in this direction [23] identifies vectors in the latent diffusion space in the directions of individual semantic concepts, which can be used to restrict model outputs. However, it is arguable that such a discretized treatment of the concepts within the diffusion latent space is restrictive. It is also intrinsically limited to dealing with diverse facets of the concepts individually. Overall, it still remains a widely open challenge for the research community to comprehensively embed the multi-faceted concepts related to fair and safe image generation in T2I pipelines. + +In this work, we address this by presenting a unique scalable approach to extensively incorporate fair and safe generative abilities in a T2I pipeline in a plug-and-play manner. Our key insight is that a T2I model can be distilled for a range of user-desired responsibility concepts, conditioned on the original model. We perform this distillation by inducing a student model that treats the concepts related to fairness and safety in a continuous space. Within this space, we further leverage concept whitening [9] to disentangle the representations of concepts for a better ultimate + +control on the generative image modulation. The underlying framework of our technique is generic, in that it is applicable to any representation space encoding semantic concepts. Hence, we explore its application to both text encoder embedding space and diffusion model latent space in the T2I pipeline. Methods for both variants address their unique challenges, but prove equally effective - see Fig. 1. The key contributions of this work are summarized below. + +- We propose a unique plug-and-play method for responsible image generation with T2I pipeline that enables comprehensive incorporation of fairness and safety concepts while modeling them under a distilled continuous space conditioned on the generative model. +- We tailor our method as a text embedding plug-in, and diffusion latent plug-in while also leveraging concept whitening to ensure precise interpretable control over the content. Both plug-ins are employed in our approach. +- With extensive experiments, we not only demonstrate state-of-the-art or comparable performance in fair and safe image generation, but also show other unique interesting properties of our approach. + +# 2. Related Works + +Text-to-image (T2I) generation has significantly expanded the capabilities of generative AI, enabling high-quality outputs using textual descriptions [34, 32, 30, 19, 17]. While T2I methods like Stable Diffusion [34], DALL-E [32], GLIDE [27], and Imagen [36] have achieved impressive realism and adaptation in numerous applications, they also face several challenges. One of them is the concern about their irresponsible and unethical content generation [15, 2]. + +Contemporary T2I models are known to generate inappropriate images, including nude and violent content, and they are also susceptible to social and cultural biases persistently found in the public domain datasets used to train these models [10, 24, 41]. The increased popularity of T2I models has driven research into understanding and mitigating biases in their outputs, particularly focusing on safety filters [33], semantic space organization [7], and tendencies toward stereotypical portrayals [5, 10, 26]. Studies show that these models, often guided by biased systems, e.g., CLIP [31, 1], may memorize sensitive data [8, 40], which, combined with unfiltered web-based datasets containing harmful content [6, 39, 4, 3], propagates unintended biases. + +Research to address these concerns can be divided into three main directions; namely, preemptive content filtering, model-level adjustments, and post-hoc moderation. The preemptive filtering approaches involve controlling input prompts to reduce biases before they influence image generation. For example, prompt-based filters [48, 2] screen out language associated with inappropriate content [19, 30]. In the model-level adjustment, unlearning techniques [45, 29, 18, 49] aim to remove sensitive concepts directly from the + +model's learned representations. Similarly, model editing methods target specific parameters of the models to limit the generation of harmful content without extensive re-training [28, 14, 46]. Although useful, these methods can negatively impact the original model performance. They can also become computationally intensive as the harmful concepts increase [49]. Post-hoc moderation methods [47, 43, 22] involve both automated and manual processes like GuardT2I [47] and context-based filters [43, 22]. These technique only partially address model biases and irresponsible generation, and may easily overlook implicit negative concept associations in the model. + +More recently, research on diffusion model latent space manipulation and interpretability is gaining prominence in regards to controlling the T2I outputs. Studies such as [16, 50] explore the semantic organization within the latent spaces to understand and intervene in the generation process. Li et al. [23] investigated interpretable directions of target concepts in the latent semantic space. These methods remain effective, however; their underlying discretized treatment of the individual semantic concepts limits their scalability. In this work, we address safe and fair modulation of both embedding and latent spaces without explicitly discretizing the directions, and focusing on their continuous space instead. Our distinctive plug-and-play approach learns an interpretable space for a range of responsible concepts while conditioning it on the T2I pipeline. + +# 3. Proposed Approach + +We present a unique approach for responsible image generation with T2I pipeline that develops add-on modules usable in a plug-and-play manner. The modules are neural models induced by distilling the target T2I model for a scalable list of responsible semantic concepts addressing fair and safe (non-violent/explicit) image generation. We first formally present the problem as perceived in this work. + +# 3.1. Problem Formalization + +A text-to-image model, $\Psi(t): t \to \bar{I}$ , maps a textual prompt $t \in \mathcal{T}$ to an image $\bar{I} \in \mathcal{M} \in \mathbb{R}^{H \times W \times 3}$ , where $\mathcal{T}$ and $\mathcal{M}$ are the sets of possible text prompts and images. Depending on the data used to train the T2I model, the image $\bar{I}$ may belong to an image subset $\overline{\mathcal{R}}_{\mathcal{A}_X} \subset \mathcal{M}$ . Conditioned on $\mathcal{A}_X$ , $\overline{\mathcal{R}}_{\mathcal{A}_X}$ denotes the set of irresponsible/inappropriate images as identified by a collection of high-level cultural/social/societal aspects in set $\mathcal{A}_X$ . In our settings, $\mathcal{A}_X$ is a subset of the aspects enlisted in Eq. (1). Notice that, $\mathcal{A}_X \subseteq \mathcal{A}_{\mathrm{resp}}$ implies that $\overline{\mathcal{R}}_{\mathcal{A}_X}$ accounts for the fact that $\bar{I}$ can be inappropriate along more than one aspect - an issue often ignored in the existing literature. + +$$ +\mathcal {A} _ {\text {r e s p}} = \left\{\mathcal {A} _ {\text {a g e}}, \mathcal {A} _ {\text {g e n d e r}}, \mathcal {A} _ {\text {r a c e}}, \mathcal {A} _ {\text {s a f e}} \right\}. \tag {1} +$$ + +We form $\mathcal{A}_{\mathrm{resp}}$ considering the aspects of responsible image generation accounted in the existing works [23, 38]. + +Nevertheless, we keep $\mathcal{A}_{\mathrm{resp}}$ extendable as required. In this work, we define each of the considered aspects using high-level attributes listed below, which also aligns with the existing literature [23, 38]: $\mathcal{A}_{\mathrm{gender}} = \{\mathbf{a}_{\mathrm{male}}, \mathbf{a}_{\mathrm{female}}\}$ , $\mathcal{A}_{\mathrm{race}} = \{\mathbf{a}_{\mathrm{white}}, \mathbf{a}_{\mathrm{asian}}, \mathbf{a}_{\mathrm{black}}\}$ , $\mathcal{A}_{\mathrm{age}} = \{\mathbf{a}_{\mathrm{young}}, \mathbf{a}_{\mathrm{middle-aged}}, \mathbf{a}_{\mathrm{elderly}}\}$ , and $\mathcal{A}_{\mathrm{safe}} = \{\mathbf{a}_{\mathrm{harassment}}, \mathbf{a}_{\mathrm{sexual}}, \mathbf{a}_{\mathrm{violence}}\}$ . + +It is worth emphasizing that the attribute sets noted above can also be extended if required. Our approach is not constrained by these sets, except for further training required to account for any additional aspect. Given an $\mathcal{A}_X$ - as defined by the model user, our objective is to alter the mapping of the T2I model to $\Psi_{\mathrm{resp}}(t):t\to I$ such that $I\in \mathcal{R}_{\mathcal{A}_X}\subseteq \mathcal{M}$ and $\overline{\mathcal{R}}_{\mathcal{A}_X}\cap \mathcal{R}_{\mathcal{A}_X} = \emptyset$ . Here, $\Psi_{\mathrm{resp}}$ is the responsible variant of $\Psi$ with an added functionality that restricts the outputs of $\Psi$ to $\mathcal{A}_X$ -constrained responsible image space $\mathcal{R}_{\mathcal{A}_X}$ . + +# 3.2. Framework Blueprint + +Conceptually, a T2I model $\Psi$ is a hierarchical function $\Psi(t) = \mathcal{D}(\mathcal{E}(t))$ , where $\mathcal{E}(.)$ is a text-encoder - typically CLIP [20], and $\mathcal{D}(.)$ is a diffusion model [34]. In this work, we first seek $\Psi_{\mathrm{resp}}$ mentioned in § 3.1 by devising add-on functionalities for responsible image generation while targeting $\mathcal{E}(.)$ and $\mathcal{D}(.)$ individually. In other words, we aim for $\mathcal{E}_{\mathrm{resp}}(.)$ and $\mathcal{D}_{\mathrm{resp}}(.)$ . Both of these potentially 'enhanced' sub-models can also be combined for $\Psi_{\mathrm{resp}}$ . We target the text encoder and diffusion model individually because in T2I modeling, both of them are known to effectively encode high-level semantic information in their embedding and latent spaces [25, 16, 50]. This provides us the possibility to plug-in our intended add-on modules to any of these models for responsible T2I generation. + +For both of the desired $\mathcal{E}_{\mathrm{resp}}(.)$ and $\mathcal{D}_{\mathrm{resp}}(.)$ , the framework underlying our approach remains the same at the conceptual level. We keep the target sub-model $(\mathcal{E}$ or $\mathcal{D})$ in the T2I pipeline frozen, and learn an add-on module (respectively, $\psi_{\mathcal{E}}$ or $\psi_{\mathcal{D}}$ ) by distilling the target sub-model with knowledge distillation [25]. The knowledge is distilled by conditioning the process on the set $\mathcal{A}_X$ which covers the aspects along which the model needs to be made responsible. Moreover, to ensure that the distilled concepts are disentangled as much as possible, we apply concept whitening [9] to them. The eventual T2I signal gets modulated by a simple addition of the signal from the add-on modules for responsible generation. + +We explain the exact procedure for obtaining $\psi_{\mathcal{E}}$ for $\mathcal{E}_{\mathrm{resp}}(.)$ , and $\psi_{\mathcal{D}}$ for $\mathcal{D}_{\mathrm{resp}}(.)$ in § 3.3 and 3.4, respectively. Combining them for $\Psi_{\mathrm{resp}}$ is explained in § 3.5. + +# 3.3. Responsible Interpretable Embeddings + +Our add-on module $\psi_{\mathcal{E}}$ for $\mathcal{E}_{\mathrm{resp}}(.)$ is marked by not only responsible generation but also interpretability of the embedding space distilled from the source $\mathcal{E}(.)$ . Hence, we term the devised module as RICE - Responsible and Interpretable CLIP Embedding. The RICE module treats the original text + +![](images/0ad5565c7e2d5f97ad49db789565b091a98ecd23525e01519782cf8f66f82b4f.jpg) +(a) RICE Embeddings Space + +![](images/32539d5ad9e2a463c0ce89108040edc9233992ee9241eff278119b1fe8e0286c.jpg) +(b) RICE Controlled T2I Outputs + +![](images/5160f8954395ff8e10fc3329d7c7e8c03a23315f0b2fa650e7d2403f0aa37bcd.jpg) +Figure 2. (a) Illustration of responsible concept space learned by the RICE module, aiming to generate $z_{\mathrm{clip}}^{\mathrm{final}}$ embeddings along the responsible aspects in $\mathcal{A}_{\mathrm{X}}$ . Interconnected fairness related subspaces for race, gender, and age are illustrated. (b) Example outputs of RICE controlled T2I process, demonstrating a range of modulation along the concepts, which is possible due to the multi-facet control enabled by the RICE embedding space. (c) Depiction of the distillation process used for the RIIDL module, where the student (RIIDL) learns from the teacher (Diffusion Model) over $\tau$ time-steps. At each step, the loss - see Eq. (6) - aligns the RIIDL module with the aspects of $\mathcal{A}_{\mathrm{X}}$ . (d) At inference, textual embeddings get fed into RIIDL. We illustrate control over $\mathcal{A}_{\mathrm{age}}$ . The latent vectors from RIIDL are injected into the T2I Diffusion Model noisy latents. We explore this injection at varying time-step (differentiated vertically in the figure). The best control is observed in the early stages ( $\tau = 0$ to $30\%$ ). Additional examples provided in the supplementary material. (e) Our method allows a scalable control over all concepts covered in $\mathcal{A}_{\mathrm{X}}$ in a composite manner. + +encoder in the T2I pipeline as a teacher and distills it for the concepts in $\mathcal{A}_X$ by training $\psi_{\mathcal{E}}$ while minimizing the Expected value of the following per-batch loss: + +$$ +\mathcal {L} _ {\mathrm {K D - c l i p}} = \frac {1}{| B |} \sum_ {k = 1} ^ {| B |} \left\| z _ {\mathrm {c l i p}, k} ^ {\mathrm {T}} - z _ {\mathrm {c l i p}, k} ^ {\text {d i s t i l l}} \right\| ^ {2}, \tag {2} +$$ + +where $B$ denotes the batch, $z_{\mathrm{clip},k}^{\mathrm{T}}$ and $z_{\mathrm{clip},k}^{\mathrm{distill}}$ are the teacher and distilled student embeddings for the $k^{\mathrm{th}}$ sample, respectively. The $z_{\mathrm{clip}}^{\mathrm{T}}$ embeddings are computed using an automated prompt generation following $\mathcal{A}_X$ . To keep the flow of discussion, we provide details about that process and the architecture of $\psi_{\mathcal{E}}$ in the supplementary material. + +At this stage, $\mathbf{z}_{\mathrm{clip}}^{\mathrm{distll}} \in \mathbb{R}^{d \times n}$ are the embedding representations conditioned on $\mathcal{A}_X$ . We additionally apply concept whitening [9] to further decorrelate these distilled embeddings using the following transform: + +$$ +\mathbf {z} _ {\mathrm {c l i p}} ^ {\mathrm {z c a}} = \phi (\mathbf {z} _ {\mathrm {c l i p}} ^ {\mathrm {d i s t i l l}}) = \mathcal {W} (\mathbf {z} _ {\mathrm {c l i p}} ^ {\mathrm {d i s t i l l}} - \mu), \tag {3} +$$ + +where $\mu = \frac{1}{n}\sum_{i=1}^{n}z_{i}$ , and $\mathcal{W}_{d\times d}$ is the whitening matrix that obeys $\mathcal{W}^T\mathcal{W} = \Sigma^{-1}$ . Here, $\Sigma_{d\times d} = \frac{1}{n}(z_{\mathrm{clip}}^{\mathrm{distill}} - \mu\mathbf{1}^{\intercal})(z_{\mathrm{clip}}^{\mathrm{distill}} - \mu\mathbf{1}^{\intercal})^{\intercal}$ is the covariance matrix. The final embeddings for responsible generation get computed as: + +$$ +\mathbf {z} _ {\text {c l i p}} ^ {\text {r e s p}} = \alpha \cdot \mathbf {z} _ {\text {c l i p}} ^ {\text {d i s t i l l}} + (1 - \alpha) \cdot \mathbf {z} _ {\text {c l i p}} ^ {\text {z c a}}, \tag {4} +$$ + +where $\alpha$ is kept for balancing. At inference, the embedding $z_{\mathrm{clip}}$ of the original prompt gets modulated as + +$$ +z _ {\text {c l i p}} ^ {\text {f i n a l}} = z _ {\text {c l i p}} + \sum_ {k} \gamma_ {k} \mathbf {z} _ {k, \text {c l i p}} ^ {\text {r e s p}}, \tag {5} +$$ + +where $\gamma_{k}$ provides control over the influence of each concept on the embedding. With this, the embedding of a prompt can be simultaneously modulated to multiple responsible concepts represented in the continuous representation space of $\psi_{\mathcal{E}}$ . Fig. 2a illustrates this notion, with Fig. 2b showing the RICE impact on varying the eventual T2I output along different responsible concepts. + +# 3.4. Responsible Interpretable Latents + +In regards to $\mathcal{D}_{\mathrm{resp}}(.)$ , we follow a similar process as adopted in § 3.3 for $\mathcal{E}_{\mathrm{resp}}(.)$ , tailoring the former for the diffusion model component $\mathcal{D}$ of the T2I pipeline. Our add-on module $\psi_{D}$ , again a neural model, gets trained using the intermediate latent representations of $\mathcal{D}$ , distilled by conditioning it on $\mathcal{A}_X$ , hence; we term it RIIDL - Responsible Interpretable Intermediate Diffusion Latents. + +The latent representation of a diffusion model gets updated for each time-step $(\tau)$ of a Markov process denoising the image previously corrupted in the forward diffusion process. For brevity, we refrain from discussing minutiae of diffusion modeling - interested readers are referred to [12]. Here, we focus on the key idea of inducing our plug-in $\psi_{D}$ for the process. To train $\psi_{D}$ , we define the following loss + +$$ +\mathcal {L} _ {\mathrm {K D - u n e t}} = \mathbb {E} _ {\tau \sim \mathcal {U} [ 0, 1 ]} \left[ \omega \left(\lambda_ {\tau}\right) \left\| z _ {\text {u n e t}} ^ {\mathrm {T}, \tau} - z _ {\text {u n e t}} ^ {\text {d i s t i l l}, \tau} \right\| ^ {2} \right], \tag {6} +$$ + +where $\tau \sim \mathcal{U}[0,1]$ refers to uniform sampling of time-steps, $z_{\mathrm{unet}}^{\mathrm{T},\tau}$ and $z_{\mathrm{unet}}^{\mathrm{distill},\tau}$ represent the latent representations of the teacher and student models at time-step $\tau$ , $\lambda_{\tau}$ reflects the SNR of the signal at $\tau$ , and $\omega(\lambda_{\tau})$ is a pre-specified weighting function. We follow [21] for implementing the SNR and $\omega(.)$ . In this setup, $\psi_{D}$ is the student while the T2I diffusion model is the teacher. + +In Eq. (6), the subscript 'unet' emphasizes the U-Net architecture of the student and teacher, which is commonly followed in denoising diffusion models. Details of our student model $\psi_{\mathcal{D}}$ are provided in the supplementary material. Similar to $z_{\mathrm{clip}}^{\mathrm{T}}$ in Eq. (2), the $z_{\mathrm{unet}}^{\mathrm{T}}$ at $\tau$ is generated based on $\mathcal{A}_X$ , which conditions $\psi_{\mathcal{D}}$ on the responsible concepts being considered by the user. We further follow the processing of $z_{\mathrm{unet}}^{\mathrm{T}}$ at a given $\tau$ corresponding to Eq. (3), (4) and (5) to respectively compute $\mathbf{z}_{\mathrm{unet}}^{\mathrm{zca}}$ , $\mathbf{z}_{\mathrm{unet}}^{\mathrm{resp}}$ , and $z_{\mathrm{unet}}^{\mathrm{final}}$ at $\tau$ . Fig. 2c illustrate the distillation process to train $\psi_{\mathcal{D}}$ for the RIIDL module on receiving the textual embedding input from $\mathcal{E}$ . In Fig. 2d, an illustration is provided for the deployed model, where RIIDL is able to guide the generated image by injecting signals to the early denoising stages of the T2I diffusion model to encourage generation along responsible the concept in $\mathcal{A}_X$ . Figure 2e further shows that extending $\mathcal{A}_X$ enables simultaneous control over multiple concepts because $\psi_{\mathcal{D}}$ gets conditioned on multiple concepts. + +# 3.5. Dual-Space Integration + +The methods discussed in § 3.3 and § 3.4 provide add-on modules $\mathcal{E}_{\mathrm{resp}}(\cdot)$ and $\mathcal{D}_{\mathrm{resp}}(\cdot)$ in the form of plug-and-play models $\psi_{\mathcal{E}}$ and $\psi_{\mathcal{D}}$ that can be directly plugged into the text encoder $\mathcal{E}$ and diffusion model $\mathcal{D}$ of the T2I pipeline. Whereas both modules are usable standalone, we also integrate them in this work for a comprehensive dual space control on the generative outputs. When both modules are integrated into the T2I pipeline, the overall function of gen + +erating a responsible image $I$ from a prompt $t\in \mathcal{T}$ can be expressed as: + +$$ +\Psi_ {\text {r e s p}} (t) = \lambda_ {\mathcal {E}} \cdot \mathcal {D} \left(\mathcal {E} _ {\text {r e s p}} (t)\right) + \lambda_ {\mathcal {D}} \cdot \mathcal {D} _ {\text {r e s p}} (\mathcal {E} (t)), \tag {7} +$$ + +where $\lambda_{\mathcal{E}}$ and $\lambda_{\mathcal{D}}$ are weighting factors following the relation $\lambda_{\mathcal{E}} + \lambda_{\mathcal{D}} = 1$ . The integrated $\Psi_{\mathrm{resp}}(t)$ provides effective control over high-level semantic information in the embedding and latent spaces. It provides a comprehensive plug-and-play method for the T2I pipeline for altering the behavior of $\Psi$ to follow concepts in $\mathcal{A}_X$ for responsible text-to-image generation. + +# 4. Evaluation + +Our method is evaluated for a variety of concepts related to fair and safe image generation. The used evaluation protocols are based on standardized benchmarks. For reproducibility, further details on implementation, dataset descriptions, and configurations of hardware and software libraries are also provided in the supplementary material. Our experiments use the Stable Diffusion v1.4 model [35]. We compare with SLD [38], ESD [13], UCE [14], and Vector-SD [23], using WinoBias [52] and I2P [38] datasets, which contain fairness-sensitive and safety-critical scenarios. + +Fair Generation Evaluation: In regards to the fair generation, we evaluate performance across professions with documented gender and racial biases. The dataset for this evaluation is WinoBias [52], which comprises 35 professions historically associated with gender bias. Fairness is quantitatively assessed using the deviation ratio $\Delta$ as in literature [28, 11, 23], which measures the disparity in generated representations across specified responsible aspects by comparing the model's output distribution against an expected baseline. Formally, $\Delta$ is defined as: $\Delta = \frac{\max_{a\in\mathcal{A}}\left|\frac{N_a}{N} - \frac{1}{|\mathcal{A}|}\right|}{1 - \frac{1}{|\mathcal{A}|}}$ , where $\mathcal{A}$ is the set of attributes within a responsible aspect (e.g., gender or race), $N$ represents the total number of generated images, and $N_{a}$ denotes the count of images for which the highest probability attribute prediction corresponds to attribute $a$ . For extensive evaluation, we further employ "challenging prompts" [23] in our experiments, which are known to produce bias representations towards male depictions [14]. + +Safe Generation Evaluation: We also evaluate the generated content against the standard criteria for identifying harmful or culturally sensitive outputs [38, 23]. Images are flagged as inappropriate if they depict or imply content associated with sexual content, self-harm, hate, illegal activity, shock, harassment, or violence. Our evaluation integrates both the NudeNet detector1 and Q16 classifier [37], as per common practices in responsible generation assessment, to systematically flag content deemed inappropriate [38, 23]. We examine a diverse range of images to rigorously assess robustness across varied outputs. + +$^{1}$ https://github.com/notAI-tech/NudeNet + +Table 1. Bias generation quantified by deviation ratio $\left( {0 \leq \Delta \leq 1}\right)$ . Lower values indicate better performance. Results provided for Gender and Race bias across standard and extended (Gender-Pro/Race-Pro) Winobias datasets [52]. + +
AttributeGender(↓)Gender-Pro(↓)Race(↓)Race-Pro(↓)
SD [35]U [14]V [23]OursSD [35]U [14]V [23]OursSDU [14]V [23]OursSD [35]U [14]V [23]Ours
CEO0.920.280.150.040.900.580.210.130.380.130.220.080.310.080.220.12
Doctor0.920.200.080.050.520.320.100.100.920.070.260.070.590.520.150.12
Nurse1.000.390.960.040.980.840.430.180.760.250.300.070.390.790.080.14
Receptionist0.840.380.880.040.980.960.860.170.880.100.360.070.740.140.250.11
Teacher0.300.060.510.040.480.160.070.130.510.100.040.080.260.230.210.10
+ +SD: Standard Diffusion Model, U: Unified Concept Editing [14], V: Vector Interpret Diffusion [23]. + +![](images/2a056058ee7132a054dffcd32923deab302601261f7b392c6b58deb2751636ae.jpg) +Figure 3. Pair-wise comparison for the Stable Diffusion (SD) baseline and our plugin to it for responsible generation along age, gender and race. SD images contain stereotypical associations to the profession, which get removed by our method by often accounting for more than one responsible attributes, while maintaining high image quality. + +# 5. Experimental Results + +Fair Generation: In Table 1, we present the bias quantification results using the deviation ratio $(0 \leq \Delta \leq 1)$ across gender and racial attributes, evaluated under both the standard (Gender and Race) and extended (Gender-Pro and Race-Pro) settings of the WinoBias dataset [52]. Lower values of $\Delta$ indicate reduced demographic bias. Overall, our method consistently achieves highly competitive results. Even under the Gender-Pro setting, where prompts are specifically crafted to amplify stereotypical associations, our method maintains its robustness. For racial bias, our method similarly maintains its performance. + +Table 2 further quantifies the debiasing efficacy of the our approach across professions with known demographic skew, displaying results for five professions and the average deviation across all 35 professions in the WinoBias [52]. Results show a substantial bias reduction by our method. Consistent bias reduction demonstrates that bias mitiga + +tion is achieved by our method independent of profession-specific attributes. Averaged over all professions (last row), our method provides a significant gain over the baseline methods. Such efficacy of our method comes from its ability to enable debiasing using a collective set of attributes in $\mathcal{A}_X$ and modulate the T2I pipeline to generate images responsibly considering all those attributes. + +To show how this manipulation varies the images generated by Stable Diffusion (SD), in Fig. 3 we provide representative examples. The figure shows a range of professions for which pairs of images are generated using the same seeds. The SD images show stereotypical associations to the professions. Our method is plugged in to SD with set $\mathcal{A}_X$ collectively containing race, gender and age attributes. It is observable in the provided representative examples that images get modulated towards these responsible concepts, often incorporating more than one responsible concept. We also provide further results in the supplementary material. + +Table 2. Debaising performance across 5 randomly selected professions and the average of all 35 professions in Winobias dataset [52]. The metric $\Delta = 0$ indicate ideal debraising. Our method shows the lowest average deviation compared to previous approaches. + +
ProfessionOriginal-SDConcept Algebra [44]Debias-VL [11]TIME [28]TIMEP [14]Ours
Librarian0.86 ± 0.060.66 ± 0.070.34 ± 0.060.26 ± 0.050.35 ± 0.010.04 ± 0.02
Teacher0.42 ± 0.010.46 ± 0.000.11 ± 0.050.34 ± 0.060.07 ± 0.060.05 ± 0.01
Sheriff0.99 ± 0.010.38 ± 0.220.82 ± 0.080.22 ± 0.050.10 ± 0.050.06 ± 0.03
Analyst0.58 ± 0.120.24 ± 0.180.71 ± 0.020.52 ± 0.030.13 ± 0.050.07 ± 0.02
Doctor0.78 ± 0.040.40 ± 0.020.50 ± 0.040.58 ± 0.030.41 ± 0.080.06 ± 0.01
Average0.67 ± 0.010.43 ± 0.010.55 ± 0.010.44 ± 0.000.31 ± 0.000.05 ± 0.02
+ +![](images/647bae11f6f88fbbe900f34c275dd21b50938d0972003b742e99669fc4d6344b.jpg) +A doctor + +![](images/8f876c11b6a6d4cbb87f6edfcc345edaf356b5127790c1b7769e23493b8092e8.jpg) +Agender + +![](images/0d54ab4a7812cee38482f0d082e2a862cf31cbcfbfe5ec0f605a52a73b0f28e3.jpg) +Female + +![](images/d402b393d3afed61194e4c40234ab634bae1e7f58592aef55ce8d046493c414b.jpg) +Arace + +![](images/0a708ac91ca8fa0f96621e9f674fe8e6ea38e3af83cc3b5201a5c7d789281db2.jpg) +Asian - Black + +![](images/d414e854cb8a1ef231176c3df22d8b551463026e413d56fe90ae42ca532eafd8.jpg) + +![](images/02b4e1490f252c7a839a038fc803e554af1dd703fb8678328a94e5a33b90d063.jpg) + +![](images/d6c90acb948f62f030c7a44cbaae6addece40cf6dde9801583ddfcf61dcf5ee3.jpg) +Aage + +![](images/0198d61a9ff7b72fe4a191e253ad34c8c70fc626455d50e48b168af66a832363.jpg) +Young - Old + +![](images/a35b63f1fbcba0ffd719dac79fa6770e7fa60adc313d33e2188566714a0c10f2.jpg) +$\mathrm{A}_{\text {gender }} + \mathrm{A}_{\text {race }}$ +$\mathrm{A}_{\text {gender }} + \mathrm{A}_{\text {race }} + \mathrm{A}_{\text {age }}$ + +![](images/369aaf48e207750062ca39011c6855bfe6d15b5f5afabe1f0af3d46cdab488fa.jpg) +Female/Black-Asian +Female / Asian / Old + +![](images/134f121f97b93500c5d70073091c71e994b3b9680b8a09625238187ab64456a2.jpg) + +![](images/ec55a9a0cab5be0431d09f1d415d883344f15eb310af057a4a9bd0302b58d9a1.jpg) + +Table 3. Proportion of images classified as inappropriate on I2P benchmark [38]. Lower values are more desirable. + +
CategorySDV-SD [23]SLD [38]ESD [13]Ours
Sexual0.390.230.160.190.15
Violence0.450.310.220.420.24
Hate0.420.300.190.340.18
Harassment0.350.190.160.280.19
Illegal0.350.240.180.340.14
Shocking0.530.380.270.430.28
Self-harm0.450.290.200.360.20
Avg.0.410.270.200.320.20
+ +Safe Generation: To benchmark safe and appropriate nature of the generated content, we evaluate our method on the I2P [38], which classifies generated images across categories covering a range of inappropriate content. Table 3 benchmarks our method against the state-of-the-art methods following standard protocol [38], [13]. As seen, our technique significantly reduces inappropriateness of the content across all categories. Generally, maintaining superior performance against the methods. On average, our performance is similar to SLD [38], which is a dedicated 'safe generation' method. + +# 6. Further Results & Discussion + +Multi-Concept Composition and Control: Our technique allows responsible generation with multi-concept composition. Multiple concepts can compose $\mathcal{A}_X$ , which enables the hierarchical function $\Psi_{\mathrm{resp}}$ to combine them in latent and embedding space of the T2I pipeline. Figure 4 demonstrates + +![](images/3a5bb74c62d21feb7b64518ece0ba11f8cc3b2b93c9791bce1c5118335248016.jpg) +Figure 4. Controlled composite responsible generation using the proposed method. By using different concepts in $\mathcal{A}_X$ in Eq. (1), and employing dual space control in (7), our technique can enable responsibility along single or multiple concepts, as desired. Provided the prompt "A doctor" to Stable Diffusion, alignment is achieved for a target concept composition (top label) for the attributes noted at the bottom of each image set. See supplementary material for more examples. +Figure 5. Examples of extending $\mathcal{A}_{\mathrm{resp}}$ with additional attributes - possibly unrelated to core responsible concepts. Extensions with smile and glasses are shown. In addition to varying images along responsible concepts of age, race and gender, our method is able to seamlessly incorporate the additional concepts in images. + +this capability within $\Psi_{\mathrm{resp}}$ , showing how distinct responsible aspects, such as age, gender, and race; can be independently learned and then composed to produce varied yet coherent outputs. By manipulating combinations within $\mathcal{A}_X$ , we demonstrate an excellent control achieved with interpretable variations across the composite attribute space, effectively balancing diversity with responsible aspects. + +Scalability of Concept Spaces: In our original experiments, the responsible concept set $\mathcal{A}_{\mathrm{resp}}$ in our framework, as defined in Eq. (1), included core attributes of $\mathcal{A}_{\mathrm{race}}, \mathcal{A}_{\mathrm{gender}}, \mathcal{A}_{\mathrm{age}},$ and $\mathcal{A}_{\mathrm{safe}}$ . This was only to benchmark + +![](images/a40fef026c08f663c6167e331602b0476f7f0a15f7181530f7ccb117ce1c1006.jpg) +Figure 6. Representative curtailed example of transitions caused by RICE and RIIDL to an output. An image of 'a doctor' is generated and 'female' concept is added individually by the modules. Increasingly changing the influence of the added concept representation provided by each module leads to unique alteration to the generated output. + +the method with the existing approaches. Our design allows the "responsible" set to be extendable. This scalability allows for seamlessly handling additional nuanced attributes in images, such as facial expressions (e.g., smiling) and accessories (e.g., glasses). Such concepts can act as supplementary attributes with the existing attributes, enabling complex yet interpretable multi-aspect control. In Fig. 5, we provide representative examples of scaling up the considered concepts set. It leads to combining the responsible subspaces of the core responsible features with subspaces for smiling and glasses. It is observable that our framework is scalable to not only core responsibility related concepts, but to any semantically meaningful concepts. We also provide additional results in the supplementary material. + +Individual Modules: Our method takes advantage of two modules, namely; of RICE (§ 3.3) and RIIDL (§ 3.4). To assess the specific contribution of each component within the responsible generation framework, we perform a systematic ablation by individually removing and adjusting key modules. This analysis allows selectively varying the influence both modules on the output. When only RICE operates, modifications originally occur in the embedding concept space for the T2I pipeline, facilitating transitions from base to responsible concept spaces. The RIIDL module contribute an additional layer of fine-grained control, allowing nuanced and context-sensitive adjustments within the intermediate latents during diffusion. This capability supports precise tuning of details without overriding the global context, complementing the broader adjustments. In tandem, these modules establish a dynamic balance within the dual-space intervention strategy, where RICE provides prediffusion global control, and RIIDL governs the responsible aspects during the denoising process. Whereas we provide further results supporting these point in the supplementary material, in Fig. 6 a curtailed representative example is provided to show the visual effects. + +In the figure, we generate images using prompt 'a doctor' and with an added concept 'female' from the individual modules. The 'female' concept representation $\mathbf{z}^{\mathrm{female}}$ pro + +vided by the modules is added to the original corresponding representations with an increasing degree of influence on the output. It can be observed that for RICE, this results in more prominent shifts in appearance and attire, indicating a significant control. For RIIDL, the latent space provides more refined transitions with subtle and smoother progression of the visual features. On one hand, these results demonstrate the complementary nature of the proposed modules, on the other; they establish the effectiveness of the modules as stand-alone add-ons. + +# 7. Conclusion + +This research presented a unique approach to control generative outputs of text-to-image (T2I) models in 'fair and safe' domain. Our method induces two plug-and-play modules that can be plugged into the textual embedding space and diffusion latent space of the T2I pipeline. The modules are neural models learned by distilling the text encoder and diffusion model of the original pipeline while conditioning the knowledge distillation on a pre-defined responsible concepts set, as desired by the user. Our modules also account for semantic interpretability by leveraging concept whitening. We demonstrate that both modules can explicitly contribute to responsible image generation in a complementary manner. Our extensive benchmarking with state-of-the-art methods for fair and safe image generation on two standard datasets show highly promising results. We also show that our method can conveniently deal with multiple concepts in the responsible image space while being seamlessly extendable to more broader concepts. + +Limitations and ethics concerns: A potential shortcoming of our method is that, to enable a comprehensive control over the generative output, it employs multiple hyperparameters, e.g., $\lambda$ 's in Eq. (7). Tuning these hyperparameters can potentially be seen as a limitation of our approach. However, the main purpose of allowing these adjustable hyper-parameters in our design is to provide a finer control to the user. It is convenient to adjust their values because the controls they provide have clear high-level interpretations. We provide guidelines on the hyper-parameter value adjustment in the supplementary material. From the ethics viewpoint, although our aim is to ensure responsible outputs; a potential misuse of exploiting the strong output control of our technique can be by replacing the $A_{\mathrm{resp}}$ with another set $A_{\mathrm{irresp}}$ . In $A_{\mathrm{irresp}}$ , unsafe and unfair concepts can be included. This can have an opposite impact of our technique. Hence, our method should only be used when the user has complete control over the set $A_{\mathrm{resp}}$ . + +Acknowledgment: Naveed Akhtar is a recipient of the Australian Research Council Discovery Early Career Researcher Award (project # DE230101058) funded by the Australian Government. This work is also partially supported by Google Research Scholar Program Award. + +# References + +[1] Sandhini Agarwal et al. "Evaluating CLIP: towards characterization of broader capabilities and downstream implications". In: arXiv preprint arXiv:2108.02818 (2021). +[2] Zhongjie Ba et al. "SurrogatePrompt: Bypassing the Safety Filter of Text-To-Image Models via Substitution". In: arXiv preprint arXiv:2309.14122 (2023). +[3] Hritik Bansal et al. "How well can Text-to-Image Generative Models understand Ethical Natural Language Interventions?" In: arXiv preprint arXiv:2210.15230 (2022). +[4] Hugo Berg et al. "A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning". In: arXiv preprint arXiv:2203.11933 (2022). +[5] Federico Bianchi et al. "Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale". In: arXiv preprint arXiv:2211.03759 (2022). +[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. "Multimodal datasets: misogyny, pornography, and malignant stereotypes". In: arXiv preprint arXiv:2110.01963 (2021). +[7] Manuel Brack et al. "The Stable Artist: Steering Semantics in Diffusion Latent Space". In: arXiv preprint arXiv:2212.06013 (2022). +[8] Nicholas Carlini et al. Extracting Training Data from Diffusion Models. 2023. DOI: 10.48550/ARXIV.2301.13188. URL: https://arxiv.org/abs/2301.13188. +[9] Zhi Chen, Yijie Bei, and Cynthia Rudin. "Concept Whitening for Interpretable Image Recognition". In: Nature Machine Intelligence 2.12 (Dec. 2020), pp. 772-782. +[10] Jaemin Cho, Abhay Zala, and Mohit Bansal. “Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models”. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 3043–3054. +[11] Ching-Yao Chuang et al. "Debiasing vision-language models via biased prompts". In: arXiv preprint arXiv:2302.00070 (2023). +[12] Florinel-Alin Croitoru et al. "Diffusion models in vision: A survey". In: IEEE Transactions on Pattern Analysis and Machine Intelligence 45.9 (2023), pp. 10850-10869. +[13] Rohit Gandikota et al. “Erasing concepts from diffusion models”. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 2426–2436. +[14] Rohit Gandikota et al. "Unified concept editing in diffusion models". In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024, pp. 5111-5120. +[15] Jindong Gu et al. "Towards robust prompts on vision-language models". In: arXiv preprint arXiv:2304.08479 (2023). + +[16] René Haas et al. "Discovering interpretable directions in the semantic latent space of diffusion models". In: 2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG). IEEE. 2024, pp. 1-9. +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. “Denoising diffusion probabilistic models”. In: Advances in neural information processing systems 33 (2020), pp. 6840–6851. +[18] Chi-Pin Huang et al. "Receler: Reliable concept erasing of text-to-image diffusion models via lightweight erasers". In: arXiv preprint arXiv:2311.17717 (2023). +[19] Tero Karras et al. “Elucidating the design space of diffusion-based generative models”. In: Advances in neural information processing systems 35 (2022), pp. 26565–26577. +[20] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. "Diffusionclip: Text-guided diffusion models for robust image manipulation". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022, pp. 2426-2435. +[21] Diederik Kingma et al. "Variational diffusion models". In: Advances in neural information processing systems 34 (2021), pp. 21696-21707. +[22] Nupur Kumari et al. "Ablating concepts in text-to-image diffusion models". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 22691-22702. +[23] Hang Li et al. "Self-discovering interpretable diffusion latent directions for responsible text-to-image generation". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 2024. +[24] Sasha Luccioni et al. "Stable bias: Evaluating societal representations in diffusion models". In: Advances in Neural Information Processing Systems 36 (2024). +[25] Chenlin Meng et al. "On distillation of guided diffusion models". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 14297-14306. +[26] Ranjita Naik and Besmira Nushi. "Social Biases through the Text-to-Image Generation Lens". In: arXiv preprint arXiv:2304.06034 (2023). +[27] Alex Nichol et al. "Glide: Towards photorealistic image generation and editing with text-guided diffusion models". In: arXiv preprint arXiv:2112.10741 (2021). +[28] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. "Editing implicit assumptions in text-to-image diffusion models". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 7053-7061. +[29] Yong-Hyun Park et al. "Direct unlearning optimization for robust and safe text-to-image models". In: arXiv preprint arXiv:2407.21035 (2024). +[30] William Peebles and Saining Xie. "Scalable diffusion models with transformers". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 4195-4205. + +[31] Alec Radford et al. "Learning Transferable Visual Models From Natural Language Supervision". In: arXiv preprint arXiv:2103.00020 (2021). +[32] Aditya Ramesh et al. "Hierarchical text-conditional image generation with clip latents". In: arXiv preprint arXiv:2204.06125 1.2 (2022), p. 3. +[33] Javier Rando et al. Red-Teaming the Stable Diffusion Safety Filter. 2022. DOI: 10.48550/ARXIV.2210.04610. URL: https://arxiv.org/abs/2210.04610. +[34] Robin Rombach et al. "High-resolution image synthesis with latent diffusion models". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022, pp. 10684-10695. +[35] Robin Rombach et al. "High-resolution image synthesis with latent diffusion models". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022, pp. 10684-10695. +[36] Chitwan Sahara et al. "Photorealistic text-to-image diffusion models with deep language understanding". In: Advances in neural information processing systems 35 (2022), pp. 36479-36494. +[37] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. "Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content?" In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. 2022, pp. 1350-1361. +[38] Patrick Schramowski et al. "Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 22522-22531. +[39] Shoaib Ahmed Siddiqui et al. "Metadata Archaeology: Unearthing Data Subsets by Leveraging Training Dynamics". In: arXiv preprint arXiv:2209.10015 (2022). +[40] Gowthami Somepalli et al. "Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models". In: arXiv preprint arXiv:2212.03860 (2022). +[41] Lukas Struppek, Dominik Hintersdorf, and Kristian Kersting. "The biased artist: Exploiting cultural biases via homoglyphs in text-guided image generation models". In: (2022). +[42] Yu-Lin Tsai et al. "Ring-A-Bell! How Reliable are Concept Removal Methods for Diffusion Models?" In: arXiv preprint arXiv:2310.10012 (2023). +[43] Peiran Wang et al. "Moderator: Moderating Text-to-Image Diffusion Models through Fine-grained Context-based Policies". In: arXiv preprint arXiv:2408.07728 (2024). +[44] Zihao Wang et al. "Concept algebra for text-controlled vision models". In: arXiv preprint arXiv:2302.03693 2 (2023). +[45] Yongliang Wu et al. "Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient". In: arXiv preprint arXiv:2405.15304 (2024). + +[46] Tianwei Xiong et al. "Editing Massive Concepts in Text-to-Image Diffusion Models". In: arXiv preprint arXiv:2403.13807 (2024). +[47] Yijun Yang et al. "GuardT2I: Defending Text-to-Image Models from Adversarial Prompts". In: arXiv preprint arXiv:2403.01446 (2024). +[48] Jaehong Yoon et al. "SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation". In: arXiv preprint arXiv:2410.12761 (2024). +[49] Gong Zhang et al. “Forget-me-not: Learning to forget in text-to-image diffusion models”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024, pp. 1755–1764. +[50] Hongxiang Zhang, Yifeng He, and Hao Chen. "Steerdiff: Steering towards safe text-to-image diffusion models". In: arXiv preprint arXiv:2410.02710 (2024). +[51] Yimeng Zhang et al. "To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now". In: European Conference on Computer Vision. Springer. 2025, pp. 385-403. +[52] Jieyu Zhao et al. "Gender bias in coreference resolution: Evaluation and debiasing methods". In: arXiv preprint arXiv:1804.06876 (2018). \ No newline at end of file diff --git a/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/images.zip b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8f322f4b82d89f7603fac5e0aaab8987909cfc05 --- /dev/null +++ b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9447d3e93e06bd33c9a3b317bac4ab2f4bcad9fdc506b9877673025e2cfe7e1e +size 780381 diff --git a/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/layout.json b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3dc6313ce2a9202609826e55cb779d316dd19d89 --- /dev/null +++ b/2025/Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control/layout.json @@ -0,0 +1,9978 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 505, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 505, + 140 + ], + "type": "text", + "content": "Plug-and-Play Interpretable Responsible Text-to-Image Generation via Dual-Space Multi-facet Concept Control" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 220, + 168, + 374, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 168, + 374, + 181 + ], + "spans": [ + { + "bbox": [ + 220, + 168, + 374, + 181 + ], + "type": "text", + "content": "Basim Azam Naveed Akhtar" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 185, + 510, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 185, + 510, + 200 + ], + "spans": [ + { + "bbox": [ + 83, + 185, + 510, + 200 + ], + "type": "text", + "content": "School of Computing and Information Systems, The University of Melbourne, Australia" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 204, + 413, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 204, + 413, + 215 + ], + "spans": [ + { + "bbox": [ + 179, + 204, + 413, + 215 + ], + "type": "text", + "content": "{basim.azam, naveed.akhtar1}@unimelb.edu.au" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 47, + 238, + 544, + 448 + ], + "blocks": [ + { + "bbox": [ + 47, + 238, + 544, + 448 + ], + "lines": [ + { + "bbox": [ + 47, + 238, + 544, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 544, + 448 + ], + "type": "image", + "image_path": "9e4d90a3489cd24f005e5fd2733cdc53fd11648af1460046f84cdd1e4580ffe8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 456, + 546, + 514 + ], + "lines": [ + { + "bbox": [ + 45, + 456, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 45, + 456, + 546, + 514 + ], + "type": "text", + "content": "Figure 1. Our unique plug-and-play interpretable approach simultaneously controls a range of concepts for responsible and fair image generation with text-to-image pipelines. Our method enables control over both text embedding space and latent diffusion space. Shown examples compare unfair and unsafe generation for the given prompts by the Stable Diffusion (top), along with their responsible counterparts resulting from our approach influencing the text encoder (middle) and the diffusion model (bottom). We control individual concepts for diverse/safe generation in these examples while modeling them in continuous composite responsible semantic spaces." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 521, + 192, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 521, + 192, + 533 + ], + "spans": [ + { + "bbox": [ + 143, + 521, + 192, + 533 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 546, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 290, + 715 + ], + "type": "text", + "content": "Ethical issues around text-to-image (T2I) models demand a comprehensive control over the generative content. Existing techniques addressing these issues for responsible T2I models aim for the generated content to be fair and safe (non-violent/explicit). However, these methods remain bounded to handling the facets of responsibility concepts individually, while also lacking in interpretability. Moreover, they often require alteration to the original model, which compromises the model performance. In this work, we propose a unique technique to enable responsible T2I generation by simultaneously accounting for an extensive range of concepts for fair and safe content generation in a scalable manner. The key idea is to distill the target T2I pipeline with an external plug-and-play mechanism that learns an inter" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 522, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 643 + ], + "type": "text", + "content": "pretable composite responsible space for the desired concepts, conditioned on the target T2I pipeline. We use knowledge distillation and concept whitening to enable this. At inference, the learned space is utilized to modulate the generative content. A typical T2I pipeline presents two plug-in points for our approach, namely; the text embedding space and the diffusion model latent space. We develop modules for both points and show the effectiveness of our approach with a range of strong results. Our code can be accessed at https://basim-azam.github.io/responsiblediffusion/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 658, + 387, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 387, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 387, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Rapid advances in text-to-image (T2I) generative pipeline are revolutionizing numerous vision applications, offering unprecedented convenience for synthesizing high quality" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "2976" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "visual content using textual descriptions [34, 32, 30, 19, 17]. Unfortunately, this convenience coupled with the generated content realism offered by publicly accessible T2I models, e.g., Stable Diffusion [34], DALL-E [32], Imagen [36], also has undesired ethical implications. Confronting the core values of fair and safe (non-violent/explicit) development and deployment of Artificial Intelligence, these models lack in producing generative content responsibly [15, 51]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 289, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 289, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 289, + 375 + ], + "type": "text", + "content": "The reasons of irresponsible content generation by contemporary T2I models lie deep in their training process, which must use large volumes of (mostly uncurated) data. Issues like inappropriate content and harmful stereotyping lurking in the data are passed on to the generative models through training [2, 42], which cause problems after their deployment. Hence, it is critical to develop methodologies to explicitly control generative models from producing irresponsible content. Addressing that, several techniques have emerged recently, encompassing approaches like input prompt filtering [48, 2], post-hoc content moderation [47, 43, 22], machine unlearning [45, 29, 18, 49], and model editing [14, 46]. Though effective, these approaches face limitations like requiring human-intervention, ad-hoc methodology, and selective dealing of different facets of the concepts defining 'responsible generation'." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 376, + 289, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 289, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 289, + 592 + ], + "type": "text", + "content": "Another contributing factor to the ethical concerns related to T2I models is our current lack of wholesome understanding of the latent space of the diffusion models. Researchers have already started exploring interpretability [16, 50] and manipulation of diffusion model latent spaces to mitigate unintended T2I model behavior [14, 23]. Controlling the latent space holds promise when it is possible to disentangle semantic concepts in it. One recent inspiring approach in this direction [23] identifies vectors in the latent diffusion space in the directions of individual semantic concepts, which can be used to restrict model outputs. However, it is arguable that such a discretized treatment of the concepts within the diffusion latent space is restrictive. It is also intrinsically limited to dealing with diverse facets of the concepts individually. Overall, it still remains a widely open challenge for the research community to comprehensively embed the multi-faceted concepts related to fair and safe image generation in T2I pipelines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "In this work, we address this by presenting a unique scalable approach to extensively incorporate fair and safe generative abilities in a T2I pipeline in a plug-and-play manner. Our key insight is that a T2I model can be distilled for a range of user-desired responsibility concepts, conditioned on the original model. We perform this distillation by inducing a student model that treats the concepts related to fairness and safety in a continuous space. Within this space, we further leverage concept whitening [9] to disentangle the representations of concepts for a better ultimate" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "control on the generative image modulation. The underlying framework of our technique is generic, in that it is applicable to any representation space encoding semantic concepts. Hence, we explore its application to both text encoder embedding space and diffusion model latent space in the T2I pipeline. Methods for both variants address their unique challenges, but prove equally effective - see Fig. 1. The key contributions of this work are summarized below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 170, + 545, + 326 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 170, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 170, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 170, + 545, + 228 + ], + "type": "text", + "content": "- We propose a unique plug-and-play method for responsible image generation with T2I pipeline that enables comprehensive incorporation of fairness and safety concepts while modeling them under a distilled continuous space conditioned on the generative model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 229, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 229, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 229, + 545, + 277 + ], + "type": "text", + "content": "- We tailor our method as a text embedding plug-in, and diffusion latent plug-in while also leveraging concept whitening to ensure precise interpretable control over the content. Both plug-ins are employed in our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 277, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 545, + 326 + ], + "type": "text", + "content": "- With extensive experiments, we not only demonstrate state-of-the-art or comparable performance in fair and safe image generation, but also show other unique interesting properties of our approach." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 335, + 398, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 398, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 398, + 346 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 354, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 450 + ], + "type": "text", + "content": "Text-to-image (T2I) generation has significantly expanded the capabilities of generative AI, enabling high-quality outputs using textual descriptions [34, 32, 30, 19, 17]. While T2I methods like Stable Diffusion [34], DALL-E [32], GLIDE [27], and Imagen [36] have achieved impressive realism and adaptation in numerous applications, they also face several challenges. One of them is the concern about their irresponsible and unethical content generation [15, 2]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 451, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 605 + ], + "type": "text", + "content": "Contemporary T2I models are known to generate inappropriate images, including nude and violent content, and they are also susceptible to social and cultural biases persistently found in the public domain datasets used to train these models [10, 24, 41]. The increased popularity of T2I models has driven research into understanding and mitigating biases in their outputs, particularly focusing on safety filters [33], semantic space organization [7], and tendencies toward stereotypical portrayals [5, 10, 26]. Studies show that these models, often guided by biased systems, e.g., CLIP [31, 1], may memorize sensitive data [8, 40], which, combined with unfiltered web-based datasets containing harmful content [6, 39, 4, 3], propagates unintended biases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "text", + "content": "Research to address these concerns can be divided into three main directions; namely, preemptive content filtering, model-level adjustments, and post-hoc moderation. The preemptive filtering approaches involve controlling input prompts to reduce biases before they influence image generation. For example, prompt-based filters [48, 2] screen out language associated with inappropriate content [19, 30]. In the model-level adjustment, unlearning techniques [45, 29, 18, 49] aim to remove sensitive concepts directly from the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2977" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "model's learned representations. Similarly, model editing methods target specific parameters of the models to limit the generation of harmful content without extensive re-training [28, 14, 46]. Although useful, these methods can negatively impact the original model performance. They can also become computationally intensive as the harmful concepts increase [49]. Post-hoc moderation methods [47, 43, 22] involve both automated and manual processes like GuardT2I [47] and context-based filters [43, 22]. These technique only partially address model biases and irresponsible generation, and may easily overlook implicit negative concept associations in the model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 396 + ], + "type": "text", + "content": "More recently, research on diffusion model latent space manipulation and interpretability is gaining prominence in regards to controlling the T2I outputs. Studies such as [16, 50] explore the semantic organization within the latent spaces to understand and intervene in the generation process. Li et al. [23] investigated interpretable directions of target concepts in the latent semantic space. These methods remain effective, however; their underlying discretized treatment of the individual semantic concepts limits their scalability. In this work, we address safe and fair modulation of both embedding and latent spaces without explicitly discretizing the directions, and focusing on their continuous space instead. Our distinctive plug-and-play approach learns an interpretable space for a range of responsible concepts while conditioning it on the T2I pipeline." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 399, + 165, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 165, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 165, + 414 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 416, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 501 + ], + "type": "text", + "content": "We present a unique approach for responsible image generation with T2I pipeline that develops add-on modules usable in a plug-and-play manner. The modules are neural models induced by distilling the target T2I model for a scalable list of responsible semantic concepts addressing fair and safe (non-violent/explicit) image generation. We first formally present the problem as perceived in this work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 504, + 179, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 179, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 179, + 514 + ], + "type": "text", + "content": "3.1. Problem Formalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": "A text-to-image model, " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\Psi(t): t \\to \\bar{I}" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": ", maps a textual prompt " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "t \\in \\mathcal{T}" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": " to an image " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\bar{I} \\in \\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": " are the sets of possible text prompts and images. Depending on the data used to train the T2I model, the image " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\bar{I}" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": " may belong to an image subset " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{R}}_{\\mathcal{A}_X} \\subset \\mathcal{M}" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": ". Conditioned on " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{R}}_{\\mathcal{A}_X}" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": " denotes the set of irresponsible/inappropriate images as identified by a collection of high-level cultural/social/societal aspects in set " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": ". In our settings, " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": " is a subset of the aspects enlisted in Eq. (1). Notice that, " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X \\subseteq \\mathcal{A}_{\\mathrm{resp}}" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": " implies that " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{R}}_{\\mathcal{A}_X}" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": " accounts for the fact that " + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\bar{I}" + }, + { + "bbox": [ + 46, + 518, + 289, + 664 + ], + "type": "text", + "content": " can be inappropriate along more than one aspect - an issue often ignored in the existing literature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 670, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 670, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 93, + 670, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {\\text {r e s p}} = \\left\\{\\mathcal {A} _ {\\text {a g e}}, \\mathcal {A} _ {\\text {g e n d e r}}, \\mathcal {A} _ {\\text {r a c e}}, \\mathcal {A} _ {\\text {s a f e}} \\right\\}. \\tag {1}", + "image_path": "a43a87fa6fecbc82c6ac31417b67b2a6e00c555611c2a249d1f0d7057481fd26.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "content": "We form " + }, + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{resp}}" + }, + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "content": " considering the aspects of responsible image generation accounted in the existing works [23, 38]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "Nevertheless, we keep " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{resp}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " extendable as required. In this work, we define each of the considered aspects using high-level attributes listed below, which also aligns with the existing literature [23, 38]: " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{gender}} = \\{\\mathbf{a}_{\\mathrm{male}}, \\mathbf{a}_{\\mathrm{female}}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{race}} = \\{\\mathbf{a}_{\\mathrm{white}}, \\mathbf{a}_{\\mathrm{asian}}, \\mathbf{a}_{\\mathrm{black}}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{age}} = \\{\\mathbf{a}_{\\mathrm{young}}, \\mathbf{a}_{\\mathrm{middle-aged}}, \\mathbf{a}_{\\mathrm{elderly}}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{safe}} = \\{\\mathbf{a}_{\\mathrm{harassment}}, \\mathbf{a}_{\\mathrm{sexual}}, \\mathbf{a}_{\\mathrm{violence}}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "text", + "content": "It is worth emphasizing that the attribute sets noted above can also be extended if required. Our approach is not constrained by these sets, except for further training required to account for any additional aspect. Given an " + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "text", + "content": " - as defined by the model user, our objective is to alter the mapping of the T2I model to " + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{resp}}(t):t\\to I" + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "inline_equation", + "content": "I\\in \\mathcal{R}_{\\mathcal{A}_X}\\subseteq \\mathcal{M}" + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{R}}_{\\mathcal{A}_X}\\cap \\mathcal{R}_{\\mathcal{A}_X} = \\emptyset" + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{resp}}" + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "text", + "content": " is the responsible variant of " + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "text", + "content": " with an added functionality that restricts the outputs of " + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "text", + "content": "-constrained responsible image space " + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathcal{A}_X}" + }, + { + "bbox": [ + 304, + 145, + 547, + 253 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 255, + 432, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 432, + 268 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 432, + 268 + ], + "type": "text", + "content": "3.2. Framework Blueprint" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "content": "Conceptually, a T2I model " + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "content": " is a hierarchical function " + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\Psi(t) = \\mathcal{D}(\\mathcal{E}(t))" + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(.)" + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "content": " is a text-encoder - typically CLIP [20], and " + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(.)" + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "content": " is a diffusion model [34]. In this work, we first seek " + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{resp}}" + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "content": " mentioned in § 3.1 by devising add-on functionalities for responsible image generation while targeting " + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(.)" + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(.)" + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "content": " individually. In other words, we aim for " + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{resp}}(.)" + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{resp}}(.)" + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "content": ". Both of these potentially 'enhanced' sub-models can also be combined for " + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{resp}}" + }, + { + "bbox": [ + 304, + 271, + 545, + 438 + ], + "type": "text", + "content": ". We target the text encoder and diffusion model individually because in T2I modeling, both of them are known to effectively encode high-level semantic information in their embedding and latent spaces [25, 16, 50]. This provides us the possibility to plug-in our intended add-on modules to any of these models for responsible T2I generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "text", + "content": "For both of the desired " + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{resp}}(.)" + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{resp}}(.)" + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "text", + "content": ", the framework underlying our approach remains the same at the conceptual level. We keep the target sub-model " + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}" + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{D})" + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "text", + "content": " in the T2I pipeline frozen, and learn an add-on module (respectively, " + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{E}}" + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{D}}" + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "text", + "content": ") by distilling the target sub-model with knowledge distillation [25]. The knowledge is distilled by conditioning the process on the set " + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 304, + 440, + 545, + 594 + ], + "type": "text", + "content": " which covers the aspects along which the model needs to be made responsible. Moreover, to ensure that the distilled concepts are disentangled as much as possible, we apply concept whitening [9] to them. The eventual T2I signal gets modulated by a simple addition of the signal from the add-on modules for responsible generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "text", + "content": "We explain the exact procedure for obtaining " + }, + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{E}}" + }, + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{resp}}(.)" + }, + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{D}}" + }, + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{resp}}(.)" + }, + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "text", + "content": " in § 3.3 and 3.4, respectively. Combining them for " + }, + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{resp}}" + }, + { + "bbox": [ + 304, + 596, + 545, + 632 + ], + "type": "text", + "content": " is explained in § 3.5." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 637, + 511, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 637, + 511, + 651 + ], + "spans": [ + { + "bbox": [ + 305, + 637, + 511, + 651 + ], + "type": "text", + "content": "3.3. Responsible Interpretable Embeddings" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Our add-on module " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{E}}" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{resp}}(.)" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " is marked by not only responsible generation but also interpretability of the embedding space distilled from the source " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(.)" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": ". Hence, we term the devised module as RICE - Responsible and Interpretable CLIP Embedding. The RICE module treats the original text" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2978" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 77, + 203, + 183 + ], + "blocks": [ + { + "bbox": [ + 55, + 77, + 203, + 183 + ], + "lines": [ + { + "bbox": [ + 55, + 77, + 203, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 77, + 203, + 183 + ], + "type": "image", + "image_path": "0ad5565c7e2d5f97ad49db789565b091a98ecd23525e01519782cf8f66f82b4f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 193, + 174, + 202 + ], + "lines": [ + { + "bbox": [ + 81, + 193, + 174, + 202 + ], + "spans": [ + { + "bbox": [ + 81, + 193, + 174, + 202 + ], + "type": "text", + "content": "(a) RICE Embeddings Space" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 214, + 203, + 334 + ], + "blocks": [ + { + "bbox": [ + 53, + 214, + 203, + 334 + ], + "lines": [ + { + "bbox": [ + 53, + 214, + 203, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 203, + 334 + ], + "type": "image", + "image_path": "32539d5ad9e2a463c0ce89108040edc9233992ee9241eff278119b1fe8e0286c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 342, + 183, + 352 + ], + "lines": [ + { + "bbox": [ + 75, + 342, + 183, + 352 + ], + "spans": [ + { + "bbox": [ + 75, + 342, + 183, + 352 + ], + "type": "text", + "content": "(b) RICE Controlled T2I Outputs" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 76, + 544, + 353 + ], + "blocks": [ + { + "bbox": [ + 205, + 76, + 544, + 353 + ], + "lines": [ + { + "bbox": [ + 205, + 76, + 544, + 353 + ], + "spans": [ + { + "bbox": [ + 205, + 76, + 544, + 353 + ], + "type": "image", + "image_path": "5160f8954395ff8e10fc3329d7c7e8c03a23315f0b2fa650e7d2403f0aa37bcd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "text", + "content": "Figure 2. (a) Illustration of responsible concept space learned by the RICE module, aiming to generate " + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{clip}}^{\\mathrm{final}}" + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "text", + "content": " embeddings along the responsible aspects in " + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{X}}" + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "text", + "content": ". Interconnected fairness related subspaces for race, gender, and age are illustrated. (b) Example outputs of RICE controlled T2I process, demonstrating a range of modulation along the concepts, which is possible due to the multi-facet control enabled by the RICE embedding space. (c) Depiction of the distillation process used for the RIIDL module, where the student (RIIDL) learns from the teacher (Diffusion Model) over " + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "text", + "content": " time-steps. At each step, the loss - see Eq. (6) - aligns the RIIDL module with the aspects of " + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{X}}" + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "text", + "content": ". (d) At inference, textual embeddings get fed into RIIDL. We illustrate control over " + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{age}}" + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "text", + "content": ". The latent vectors from RIIDL are injected into the T2I Diffusion Model noisy latents. We explore this injection at varying time-step (differentiated vertically in the figure). The best control is observed in the early stages (" + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "inline_equation", + "content": "\\tau = 0" + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "text", + "content": "). Additional examples provided in the supplementary material. (e) Our method allows a scalable control over all concepts covered in " + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{X}}" + }, + { + "bbox": [ + 46, + 363, + 546, + 462 + ], + "type": "text", + "content": " in a composite manner." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 475, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 510 + ], + "type": "text", + "content": "encoder in the T2I pipeline as a teacher and distills it for the concepts in " + }, + { + "bbox": [ + 46, + 475, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 475, + 287, + 510 + ], + "type": "text", + "content": " by training " + }, + { + "bbox": [ + 46, + 475, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{E}}" + }, + { + "bbox": [ + 46, + 475, + 287, + 510 + ], + "type": "text", + "content": " while minimizing the Expected value of the following per-batch loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 521, + 287, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 521, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 88, + 521, + 287, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K D - c l i p}} = \\frac {1}{| B |} \\sum_ {k = 1} ^ {| B |} \\left\\| z _ {\\mathrm {c l i p}, k} ^ {\\mathrm {T}} - z _ {\\mathrm {c l i p}, k} ^ {\\text {d i s t i l l}} \\right\\| ^ {2}, \\tag {2}", + "image_path": "23db477bee7ea8e63fb3c7bd3e48716a02b54651f7e97742389415ce3d8a1b30.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "text", + "content": " denotes the batch, " + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{clip},k}^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{clip},k}^{\\mathrm{distill}}" + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "text", + "content": " are the teacher and distilled student embeddings for the " + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "inline_equation", + "content": "k^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "text", + "content": " sample, respectively. The " + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{clip}}^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "text", + "content": " embeddings are computed using an automated prompt generation following " + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "text", + "content": ". To keep the flow of discussion, we provide details about that process and the architecture of " + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{E}}" + }, + { + "bbox": [ + 46, + 567, + 287, + 640 + ], + "type": "text", + "content": " in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": "At this stage, " + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{clip}}^{\\mathrm{distll}} \\in \\mathbb{R}^{d \\times n}" + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": " are the embedding representations conditioned on " + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": ". We additionally apply concept whitening [9] to further decorrelate these distilled embeddings using the following transform:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 95, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 95, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {\\mathrm {c l i p}} ^ {\\mathrm {z c a}} = \\phi (\\mathbf {z} _ {\\mathrm {c l i p}} ^ {\\mathrm {d i s t i l l}}) = \\mathcal {W} (\\mathbf {z} _ {\\mathrm {c l i p}} ^ {\\mathrm {d i s t i l l}} - \\mu), \\tag {3}", + "image_path": "4ea88189d549cdee47a2fddcd33ffb0c35fedd6fb060bec5917632a2aac2009f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 474, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 474, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 545, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 474, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mu = \\frac{1}{n}\\sum_{i=1}^{n}z_{i}" + }, + { + "bbox": [ + 305, + 474, + 545, + 525 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 474, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{d\\times d}" + }, + { + "bbox": [ + 305, + 474, + 545, + 525 + ], + "type": "text", + "content": " is the whitening matrix that obeys " + }, + { + "bbox": [ + 305, + 474, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^T\\mathcal{W} = \\Sigma^{-1}" + }, + { + "bbox": [ + 305, + 474, + 545, + 525 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 305, + 474, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\Sigma_{d\\times d} = \\frac{1}{n}(z_{\\mathrm{clip}}^{\\mathrm{distill}} - \\mu\\mathbf{1}^{\\intercal})(z_{\\mathrm{clip}}^{\\mathrm{distill}} - \\mu\\mathbf{1}^{\\intercal})^{\\intercal}" + }, + { + "bbox": [ + 305, + 474, + 545, + 525 + ], + "type": "text", + "content": " is the covariance matrix. The final embeddings for responsible generation get computed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 355, + 534, + 545, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 534, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 355, + 534, + 545, + 551 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {\\text {c l i p}} ^ {\\text {r e s p}} = \\alpha \\cdot \\mathbf {z} _ {\\text {c l i p}} ^ {\\text {d i s t i l l}} + (1 - \\alpha) \\cdot \\mathbf {z} _ {\\text {c l i p}} ^ {\\text {z c a}}, \\tag {4}", + "image_path": "b99153e994ffe6c0b9bd506c8d606f31d19e312ab81bf32ff6e5702d64650a81.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 559, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 545, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 559, + 545, + 584 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 559, + 545, + 584 + ], + "type": "text", + "content": " is kept for balancing. At inference, the embedding " + }, + { + "bbox": [ + 306, + 559, + 545, + 584 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{clip}}" + }, + { + "bbox": [ + 306, + 559, + 545, + 584 + ], + "type": "text", + "content": " of the original prompt gets modulated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 368, + 595, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 595, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 368, + 595, + 545, + 620 + ], + "type": "interline_equation", + "content": "z _ {\\text {c l i p}} ^ {\\text {f i n a l}} = z _ {\\text {c l i p}} + \\sum_ {k} \\gamma_ {k} \\mathbf {z} _ {k, \\text {c l i p}} ^ {\\text {r e s p}}, \\tag {5}", + "image_path": "4d630d33173ba4e3f12bf65b9e627077b35b2119a22b069a2380e8849be534a2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 631, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 631, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 304, + 631, + 545, + 714 + ], + "type": "text", + "content": " provides control over the influence of each concept on the embedding. With this, the embedding of a prompt can be simultaneously modulated to multiple responsible concepts represented in the continuous representation space of " + }, + { + "bbox": [ + 304, + 631, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{E}}" + }, + { + "bbox": [ + 304, + 631, + 545, + 714 + ], + "type": "text", + "content": ". Fig. 2a illustrates this notion, with Fig. 2b showing the RICE impact on varying the eventual T2I output along different responsible concepts." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2979" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 230, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 230, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 230, + 84 + ], + "type": "text", + "content": "3.4. Responsible Interpretable Latents" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "text", + "content": "In regards to " + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{resp}}(.)" + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "text", + "content": ", we follow a similar process as adopted in § 3.3 for " + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{resp}}(.)" + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "text", + "content": ", tailoring the former for the diffusion model component " + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "text", + "content": " of the T2I pipeline. Our add-on module " + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "inline_equation", + "content": "\\psi_{D}" + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "text", + "content": ", again a neural model, gets trained using the intermediate latent representations of " + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "text", + "content": ", distilled by conditioning it on " + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 84, + 286, + 167 + ], + "type": "text", + "content": ", hence; we term it RIIDL - Responsible Interpretable Intermediate Diffusion Latents." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 168, + 286, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 286, + 253 + ], + "type": "text", + "content": "The latent representation of a diffusion model gets updated for each time-step " + }, + { + "bbox": [ + 46, + 168, + 286, + 253 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 46, + 168, + 286, + 253 + ], + "type": "text", + "content": " of a Markov process denoising the image previously corrupted in the forward diffusion process. For brevity, we refrain from discussing minutiae of diffusion modeling - interested readers are referred to [12]. Here, we focus on the key idea of inducing our plug-in " + }, + { + "bbox": [ + 46, + 168, + 286, + 253 + ], + "type": "inline_equation", + "content": "\\psi_{D}" + }, + { + "bbox": [ + 46, + 168, + 286, + 253 + ], + "type": "text", + "content": " for the process. To train " + }, + { + "bbox": [ + 46, + 168, + 286, + 253 + ], + "type": "inline_equation", + "content": "\\psi_{D}" + }, + { + "bbox": [ + 46, + 168, + 286, + 253 + ], + "type": "text", + "content": ", we define the following loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 261, + 286, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 261, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 59, + 261, + 286, + 289 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K D - u n e t}} = \\mathbb {E} _ {\\tau \\sim \\mathcal {U} [ 0, 1 ]} \\left[ \\omega \\left(\\lambda_ {\\tau}\\right) \\left\\| z _ {\\text {u n e t}} ^ {\\mathrm {T}, \\tau} - z _ {\\text {u n e t}} ^ {\\text {d i s t i l l}, \\tau} \\right\\| ^ {2} \\right], \\tag {6}", + "image_path": "63a7e5e7de6fa4b4f8027e8c83269d5f0176447b95e3e63de86781315f3608d7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\tau \\sim \\mathcal{U}[0,1]" + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "text", + "content": " refers to uniform sampling of time-steps, " + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{unet}}^{\\mathrm{T},\\tau}" + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{unet}}^{\\mathrm{distill},\\tau}" + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "text", + "content": " represent the latent representations of the teacher and student models at time-step " + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\lambda_{\\tau}" + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "text", + "content": " reflects the SNR of the signal at " + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\omega(\\lambda_{\\tau})" + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "text", + "content": " is a pre-specified weighting function. We follow [21] for implementing the SNR and " + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\omega(.)" + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "text", + "content": ". In this setup, " + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\psi_{D}" + }, + { + "bbox": [ + 46, + 297, + 286, + 380 + ], + "type": "text", + "content": " is the student while the T2I diffusion model is the teacher." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": "In Eq. (6), the subscript 'unet' emphasizes the U-Net architecture of the student and teacher, which is commonly followed in denoising diffusion models. Details of our student model " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": " are provided in the supplementary material. Similar to " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{clip}}^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": " in Eq. (2), the " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{unet}}^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": " is generated based on " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": ", which conditions " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": " on the responsible concepts being considered by the user. We further follow the processing of " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{unet}}^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": " at a given " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": " corresponding to Eq. (3), (4) and (5) to respectively compute " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{unet}}^{\\mathrm{zca}}" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{unet}}^{\\mathrm{resp}}" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{unet}}^{\\mathrm{final}}" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": ". Fig. 2c illustrate the distillation process to train " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": " for the RIIDL module on receiving the textual embedding input from " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": ". In Fig. 2d, an illustration is provided for the deployed model, where RIIDL is able to guide the generated image by injecting signals to the early denoising stages of the T2I diffusion model to encourage generation along responsible the concept in " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": ". Figure 2e further shows that extending " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": " enables simultaneous control over multiple concepts because " + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 381, + 286, + 597 + ], + "type": "text", + "content": " gets conditioned on multiple concepts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 601, + 178, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 178, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 178, + 615 + ], + "type": "text", + "content": "3.5. Dual-Space Integration" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "The methods discussed in § 3.3 and § 3.4 provide add-on modules " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{resp}}(\\cdot)" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{resp}}(\\cdot)" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " in the form of plug-and-play models " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{E}}" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " that can be directly plugged into the text encoder " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " and diffusion model " + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": " of the T2I pipeline. Whereas both modules are usable standalone, we also integrate them in this work for a comprehensive dual space control on the generative outputs. When both modules are integrated into the T2I pipeline, the overall function of gen" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "erating a responsible image " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " from a prompt " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "t\\in \\mathcal{T}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 322, + 103, + 545, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 103, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 322, + 103, + 545, + 118 + ], + "type": "interline_equation", + "content": "\\Psi_ {\\text {r e s p}} (t) = \\lambda_ {\\mathcal {E}} \\cdot \\mathcal {D} \\left(\\mathcal {E} _ {\\text {r e s p}} (t)\\right) + \\lambda_ {\\mathcal {D}} \\cdot \\mathcal {D} _ {\\text {r e s p}} (\\mathcal {E} (t)), \\tag {7}", + "image_path": "20702dfb9c39558e714e564735b4ba6086fcf57c96c7b83e5e56d93be6ec5812.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathcal{E}}" + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathcal{D}}" + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": " are weighting factors following the relation " + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathcal{E}} + \\lambda_{\\mathcal{D}} = 1" + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": ". The integrated " + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{resp}}(t)" + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": " provides effective control over high-level semantic information in the embedding and latent spaces. It provides a comprehensive plug-and-play method for the T2I pipeline for altering the behavior of " + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": " to follow concepts in " + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 304, + 123, + 545, + 207 + ], + "type": "text", + "content": " for responsible text-to-image generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 212, + 376, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 212, + 376, + 223 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 376, + 223 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 225, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 225, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 545, + 345 + ], + "type": "text", + "content": "Our method is evaluated for a variety of concepts related to fair and safe image generation. The used evaluation protocols are based on standardized benchmarks. For reproducibility, further details on implementation, dataset descriptions, and configurations of hardware and software libraries are also provided in the supplementary material. Our experiments use the Stable Diffusion v1.4 model [35]. We compare with SLD [38], ESD [13], UCE [14], and Vector-SD [23], using WinoBias [52] and I2P [38] datasets, which contain fairness-sensitive and safety-critical scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "text", + "content": "Fair Generation Evaluation: In regards to the fair generation, we evaluate performance across professions with documented gender and racial biases. The dataset for this evaluation is WinoBias [52], which comprises 35 professions historically associated with gender bias. Fairness is quantitatively assessed using the deviation ratio " + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "text", + "content": " as in literature [28, 11, 23], which measures the disparity in generated representations across specified responsible aspects by comparing the model's output distribution against an expected baseline. Formally, " + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "text", + "content": " is defined as: " + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\Delta = \\frac{\\max_{a\\in\\mathcal{A}}\\left|\\frac{N_a}{N} - \\frac{1}{|\\mathcal{A}|}\\right|}{1 - \\frac{1}{|\\mathcal{A}|}}" + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "text", + "content": " is the set of attributes within a responsible aspect (e.g., gender or race), " + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "text", + "content": " represents the total number of generated images, and " + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "text", + "content": " denotes the count of images for which the highest probability attribute prediction corresponds to attribute " + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 346, + 545, + 569 + ], + "type": "text", + "content": ". For extensive evaluation, we further employ \"challenging prompts\" [23] in our experiments, which are known to produce bias representations towards male depictions [14]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 702 + ], + "type": "text", + "content": "Safe Generation Evaluation: We also evaluate the generated content against the standard criteria for identifying harmful or culturally sensitive outputs [38, 23]. Images are flagged as inappropriate if they depict or imply content associated with sexual content, self-harm, hate, illegal activity, shock, harassment, or violence. Our evaluation integrates both the NudeNet detector1 and Q16 classifier [37], as per common practices in responsible generation assessment, to systematically flag content deemed inappropriate [38, 23]. We examine a diverse range of images to rigorously assess robustness across varied outputs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 702, + 449, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 449, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 449, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 317, + 702, + 449, + 713 + ], + "type": "text", + "content": "https://github.com/notAI-tech/NudeNet" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2980" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 102, + 529, + 199 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 530, + 94 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 530, + 94 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 530, + 94 + ], + "type": "text", + "content": "Table 1. Bias generation quantified by deviation ratio " + }, + { + "bbox": [ + 61, + 70, + 530, + 94 + ], + "type": "inline_equation", + "content": "\\left( {0 \\leq \\Delta \\leq 1}\\right)" + }, + { + "bbox": [ + 61, + 70, + 530, + 94 + ], + "type": "text", + "content": " . Lower values indicate better performance. Results provided for Gender and Race bias across standard and extended (Gender-Pro/Race-Pro) Winobias datasets [52]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 102, + 529, + 199 + ], + "lines": [ + { + "bbox": [ + 62, + 102, + 529, + 199 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 529, + 199 + ], + "type": "table", + "html": "
AttributeGender(↓)Gender-Pro(↓)Race(↓)Race-Pro(↓)
SD [35]U [14]V [23]OursSD [35]U [14]V [23]OursSDU [14]V [23]OursSD [35]U [14]V [23]Ours
CEO0.920.280.150.040.900.580.210.130.380.130.220.080.310.080.220.12
Doctor0.920.200.080.050.520.320.100.100.920.070.260.070.590.520.150.12
Nurse1.000.390.960.040.980.840.430.180.760.250.300.070.390.790.080.14
Receptionist0.840.380.880.040.980.960.860.170.880.100.360.070.740.140.250.11
Teacher0.300.060.510.040.480.160.070.130.510.100.040.080.260.230.210.10
", + "image_path": "d7e1d314c8d3e765326e787ae1dd005184cea0695e0da311e6ca5bb92a2981e9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 62, + 200, + 383, + 210 + ], + "lines": [ + { + "bbox": [ + 62, + 200, + 383, + 210 + ], + "spans": [ + { + "bbox": [ + 62, + 200, + 383, + 210 + ], + "type": "text", + "content": "SD: Standard Diffusion Model, U: Unified Concept Editing [14], V: Vector Interpret Diffusion [23]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 75, + 214, + 519, + 430 + ], + "blocks": [ + { + "bbox": [ + 75, + 214, + 519, + 430 + ], + "lines": [ + { + "bbox": [ + 75, + 214, + 519, + 430 + ], + "spans": [ + { + "bbox": [ + 75, + 214, + 519, + 430 + ], + "type": "image", + "image_path": "2a056058ee7132a054dffcd32923deab302601261f7b392c6b58deb2751636ae.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 440, + 547, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 440, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 547, + 475 + ], + "type": "text", + "content": "Figure 3. Pair-wise comparison for the Stable Diffusion (SD) baseline and our plugin to it for responsible generation along age, gender and race. SD images contain stereotypical associations to the profession, which get removed by our method by often accounting for more than one responsible attributes, while maintaining high image quality." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 485, + 173, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 173, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 173, + 498 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 506, + 288, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 636 + ], + "type": "text", + "content": "Fair Generation: In Table 1, we present the bias quantification results using the deviation ratio " + }, + { + "bbox": [ + 46, + 506, + 288, + 636 + ], + "type": "inline_equation", + "content": "(0 \\leq \\Delta \\leq 1)" + }, + { + "bbox": [ + 46, + 506, + 288, + 636 + ], + "type": "text", + "content": " across gender and racial attributes, evaluated under both the standard (Gender and Race) and extended (Gender-Pro and Race-Pro) settings of the WinoBias dataset [52]. Lower values of " + }, + { + "bbox": [ + 46, + 506, + 288, + 636 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 506, + 288, + 636 + ], + "type": "text", + "content": " indicate reduced demographic bias. Overall, our method consistently achieves highly competitive results. Even under the Gender-Pro setting, where prompts are specifically crafted to amplify stereotypical associations, our method maintains its robustness. For racial bias, our method similarly maintains its performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Table 2 further quantifies the debiasing efficacy of the our approach across professions with known demographic skew, displaying results for five professions and the average deviation across all 35 professions in the WinoBias [52]. Results show a substantial bias reduction by our method. Consistent bias reduction demonstrates that bias mitiga" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 486, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 571 + ], + "type": "text", + "content": "tion is achieved by our method independent of profession-specific attributes. Averaged over all professions (last row), our method provides a significant gain over the baseline methods. Such efficacy of our method comes from its ability to enable debiasing using a collective set of attributes in " + }, + { + "bbox": [ + 304, + 486, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 304, + 486, + 545, + 571 + ], + "type": "text", + "content": " and modulate the T2I pipeline to generate images responsibly considering all those attributes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "To show how this manipulation varies the images generated by Stable Diffusion (SD), in Fig. 3 we provide representative examples. The figure shows a range of professions for which pairs of images are generated using the same seeds. The SD images show stereotypical associations to the professions. Our method is plugged in to SD with set " + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": " collectively containing race, gender and age attributes. It is observable in the provided representative examples that images get modulated towards these responsible concepts, often incorporating more than one responsible concept. We also provide further results in the supplementary material." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2981" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 88, + 478, + 185 + ], + "blocks": [ + { + "bbox": [ + 47, + 64, + 545, + 87 + ], + "lines": [ + { + "bbox": [ + 47, + 64, + 545, + 87 + ], + "spans": [ + { + "bbox": [ + 47, + 64, + 545, + 87 + ], + "type": "text", + "content": "Table 2. Debaising performance across 5 randomly selected professions and the average of all 35 professions in Winobias dataset [52]. The metric " + }, + { + "bbox": [ + 47, + 64, + 545, + 87 + ], + "type": "inline_equation", + "content": "\\Delta = 0" + }, + { + "bbox": [ + 47, + 64, + 545, + 87 + ], + "type": "text", + "content": " indicate ideal debraising. Our method shows the lowest average deviation compared to previous approaches." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 115, + 88, + 478, + 185 + ], + "lines": [ + { + "bbox": [ + 115, + 88, + 478, + 185 + ], + "spans": [ + { + "bbox": [ + 115, + 88, + 478, + 185 + ], + "type": "table", + "html": "
ProfessionOriginal-SDConcept Algebra [44]Debias-VL [11]TIME [28]TIMEP [14]Ours
Librarian0.86 ± 0.060.66 ± 0.070.34 ± 0.060.26 ± 0.050.35 ± 0.010.04 ± 0.02
Teacher0.42 ± 0.010.46 ± 0.000.11 ± 0.050.34 ± 0.060.07 ± 0.060.05 ± 0.01
Sheriff0.99 ± 0.010.38 ± 0.220.82 ± 0.080.22 ± 0.050.10 ± 0.050.06 ± 0.03
Analyst0.58 ± 0.120.24 ± 0.180.71 ± 0.020.52 ± 0.030.13 ± 0.050.07 ± 0.02
Doctor0.78 ± 0.040.40 ± 0.020.50 ± 0.040.58 ± 0.030.41 ± 0.080.06 ± 0.01
Average0.67 ± 0.010.43 ± 0.010.55 ± 0.010.44 ± 0.000.31 ± 0.000.05 ± 0.02
", + "image_path": "b85499286b74f672c46ee1181811dca5cf3bff6f77d07ad48a94f00676565fa3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 75, + 205, + 121, + 299 + ], + "blocks": [ + { + "bbox": [ + 78, + 190, + 110, + 200 + ], + "lines": [ + { + "bbox": [ + 78, + 190, + 110, + 200 + ], + "spans": [ + { + "bbox": [ + 78, + 190, + 110, + 200 + ], + "type": "text", + "content": "A doctor" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 205, + 121, + 299 + ], + "lines": [ + { + "bbox": [ + 75, + 205, + 121, + 299 + ], + "spans": [ + { + "bbox": [ + 75, + 205, + 121, + 299 + ], + "type": "image", + "image_path": "647bae11f6f88fbbe900f34c275dd21b50938d0972003b742e99669fc4d6344b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 127, + 205, + 174, + 253 + ], + "blocks": [ + { + "bbox": [ + 137, + 190, + 164, + 199 + ], + "lines": [ + { + "bbox": [ + 137, + 190, + 164, + 199 + ], + "spans": [ + { + "bbox": [ + 137, + 190, + 164, + 199 + ], + "type": "text", + "content": "Agender" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 127, + 205, + 174, + 253 + ], + "lines": [ + { + "bbox": [ + 127, + 205, + 174, + 253 + ], + "spans": [ + { + "bbox": [ + 127, + 205, + 174, + 253 + ], + "type": "image", + "image_path": "8f876c11b6a6d4cbb87f6edfcc345edaf356b5127790c1b7769e23493b8092e8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 127, + 252, + 174, + 298 + ], + "blocks": [ + { + "bbox": [ + 127, + 252, + 174, + 298 + ], + "lines": [ + { + "bbox": [ + 127, + 252, + 174, + 298 + ], + "spans": [ + { + "bbox": [ + 127, + 252, + 174, + 298 + ], + "type": "image", + "image_path": "0d54ab4a7812cee38482f0d082e2a862cf31cbcfbfe5ec0f605a52a73b0f28e3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 301, + 162, + 308 + ], + "lines": [ + { + "bbox": [ + 138, + 301, + 162, + 308 + ], + "spans": [ + { + "bbox": [ + 138, + 301, + 162, + 308 + ], + "type": "text", + "content": "Female" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 178, + 205, + 223, + 252 + ], + "blocks": [ + { + "bbox": [ + 178, + 205, + 223, + 252 + ], + "lines": [ + { + "bbox": [ + 178, + 205, + 223, + 252 + ], + "spans": [ + { + "bbox": [ + 178, + 205, + 223, + 252 + ], + "type": "image", + "image_path": "d402b393d3afed61194e4c40234ab634bae1e7f58592aef55ce8d046493c414b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 189, + 234, + 199 + ], + "lines": [ + { + "bbox": [ + 213, + 189, + 234, + 199 + ], + "spans": [ + { + "bbox": [ + 213, + 189, + 234, + 199 + ], + "type": "text", + "content": "Arace" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 178, + 252, + 223, + 298 + ], + "blocks": [ + { + "bbox": [ + 178, + 252, + 223, + 298 + ], + "lines": [ + { + "bbox": [ + 178, + 252, + 223, + 298 + ], + "spans": [ + { + "bbox": [ + 178, + 252, + 223, + 298 + ], + "type": "image", + "image_path": "0a708ac91ca8fa0f96621e9f674fe8e6ea38e3af83cc3b5201a5c7d789281db2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 301, + 243, + 309 + ], + "lines": [ + { + "bbox": [ + 201, + 301, + 243, + 309 + ], + "spans": [ + { + "bbox": [ + 201, + 301, + 243, + 309 + ], + "type": "text", + "content": "Asian - Black" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 225, + 205, + 269, + 253 + ], + "blocks": [ + { + "bbox": [ + 225, + 205, + 269, + 253 + ], + "lines": [ + { + "bbox": [ + 225, + 205, + 269, + 253 + ], + "spans": [ + { + "bbox": [ + 225, + 205, + 269, + 253 + ], + "type": "image", + "image_path": "d414e854cb8a1ef231176c3df22d8b551463026e413d56fe90ae42ca532eafd8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 225, + 253, + 269, + 298 + ], + "blocks": [ + { + "bbox": [ + 225, + 253, + 269, + 298 + ], + "lines": [ + { + "bbox": [ + 225, + 253, + 269, + 298 + ], + "spans": [ + { + "bbox": [ + 225, + 253, + 269, + 298 + ], + "type": "image", + "image_path": "02b4e1490f252c7a839a038fc803e554af1dd703fb8678328a94e5a33b90d063.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 271, + 205, + 365, + 253 + ], + "blocks": [ + { + "bbox": [ + 295, + 190, + 329, + 201 + ], + "lines": [ + { + "bbox": [ + 295, + 190, + 329, + 201 + ], + "spans": [ + { + "bbox": [ + 295, + 190, + 329, + 201 + ], + "type": "text", + "content": "Aage" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 271, + 205, + 365, + 253 + ], + "lines": [ + { + "bbox": [ + 271, + 205, + 365, + 253 + ], + "spans": [ + { + "bbox": [ + 271, + 205, + 365, + 253 + ], + "type": "image", + "image_path": "d6c90acb948f62f030c7a44cbaae6addece40cf6dde9801583ddfcf61dcf5ee3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 271, + 253, + 365, + 298 + ], + "blocks": [ + { + "bbox": [ + 271, + 253, + 365, + 298 + ], + "lines": [ + { + "bbox": [ + 271, + 253, + 365, + 298 + ], + "spans": [ + { + "bbox": [ + 271, + 253, + 365, + 298 + ], + "type": "image", + "image_path": "0198d61a9ff7b72fe4a191e253ad34c8c70fc626455d50e48b168af66a832363.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 297, + 301, + 335, + 309 + ], + "lines": [ + { + "bbox": [ + 297, + 301, + 335, + 309 + ], + "spans": [ + { + "bbox": [ + 297, + 301, + 335, + 309 + ], + "type": "text", + "content": "Young - Old" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 367, + 205, + 460, + 253 + ], + "blocks": [ + { + "bbox": [ + 383, + 190, + 436, + 200 + ], + "lines": [ + { + "bbox": [ + 383, + 190, + 436, + 200 + ], + "spans": [ + { + "bbox": [ + 383, + 190, + 436, + 200 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_{\\text {gender }} + \\mathrm{A}_{\\text {race }}" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 367, + 205, + 460, + 253 + ], + "lines": [ + { + "bbox": [ + 367, + 205, + 460, + 253 + ], + "spans": [ + { + "bbox": [ + 367, + 205, + 460, + 253 + ], + "type": "image", + "image_path": "a35b63f1fbcba0ffd719dac79fa6770e7fa60adc313d33e2188566714a0c10f2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 190, + 518, + 201 + ], + "lines": [ + { + "bbox": [ + 453, + 190, + 518, + 201 + ], + "spans": [ + { + "bbox": [ + 453, + 190, + 518, + 201 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_{\\text {gender }} + \\mathrm{A}_{\\text {race }} + \\mathrm{A}_{\\text {age }}" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 367, + 253, + 460, + 298 + ], + "blocks": [ + { + "bbox": [ + 367, + 253, + 460, + 298 + ], + "lines": [ + { + "bbox": [ + 367, + 253, + 460, + 298 + ], + "spans": [ + { + "bbox": [ + 367, + 253, + 460, + 298 + ], + "type": "image", + "image_path": "369aaf48e207750062ca39011c6855bfe6d15b5f5afabe1f0af3d46cdab488fa.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 301, + 447, + 309 + ], + "lines": [ + { + "bbox": [ + 377, + 301, + 447, + 309 + ], + "spans": [ + { + "bbox": [ + 377, + 301, + 447, + 309 + ], + "type": "text", + "content": "Female/Black-Asian" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 300, + 514, + 308 + ], + "lines": [ + { + "bbox": [ + 459, + 300, + 514, + 308 + ], + "spans": [ + { + "bbox": [ + 459, + 300, + 514, + 308 + ], + "type": "text", + "content": "Female / Asian / Old" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 464, + 205, + 511, + 252 + ], + "blocks": [ + { + "bbox": [ + 464, + 205, + 511, + 252 + ], + "lines": [ + { + "bbox": [ + 464, + 205, + 511, + 252 + ], + "spans": [ + { + "bbox": [ + 464, + 205, + 511, + 252 + ], + "type": "image", + "image_path": "134f121f97b93500c5d70073091c71e994b3b9680b8a09625238187ab64456a2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 464, + 252, + 511, + 298 + ], + "blocks": [ + { + "bbox": [ + 464, + 252, + 511, + 298 + ], + "lines": [ + { + "bbox": [ + 464, + 252, + 511, + 298 + ], + "spans": [ + { + "bbox": [ + 464, + 252, + 511, + 298 + ], + "type": "image", + "image_path": "ec55a9a0cab5be0431d09f1d415d883344f15eb310af057a4a9bd0302b58d9a1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 60, + 388, + 271, + 491 + ], + "blocks": [ + { + "bbox": [ + 47, + 365, + 287, + 386 + ], + "lines": [ + { + "bbox": [ + 47, + 365, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 287, + 386 + ], + "type": "text", + "content": "Table 3. Proportion of images classified as inappropriate on I2P benchmark [38]. Lower values are more desirable." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 388, + 271, + 491 + ], + "lines": [ + { + "bbox": [ + 60, + 388, + 271, + 491 + ], + "spans": [ + { + "bbox": [ + 60, + 388, + 271, + 491 + ], + "type": "table", + "html": "
CategorySDV-SD [23]SLD [38]ESD [13]Ours
Sexual0.390.230.160.190.15
Violence0.450.310.220.420.24
Hate0.420.300.190.340.18
Harassment0.350.190.160.280.19
Illegal0.350.240.180.340.14
Shocking0.530.380.270.430.28
Self-harm0.450.290.200.360.20
Avg.0.410.270.200.320.20
", + "image_path": "2c7f6022edae4cfa2040c0f9fca063b3de5615b06f35215f89c33aaa7aae9844.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "content": "Safe Generation: To benchmark safe and appropriate nature of the generated content, we evaluate our method on the I2P [38], which classifies generated images across categories covering a range of inappropriate content. Table 3 benchmarks our method against the state-of-the-art methods following standard protocol [38], [13]. As seen, our technique significantly reduces inappropriateness of the content across all categories. Generally, maintaining superior performance against the methods. On average, our performance is similar to SLD [38], which is a dedicated 'safe generation' method." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 47, + 635, + 217, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 217, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 217, + 648 + ], + "type": "text", + "content": "6. Further Results & Discussion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Multi-Concept Composition and Control: Our technique allows responsible generation with multi-concept composition. Multiple concepts can compose " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", which enables the hierarchical function " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{resp}}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " to combine them in latent and embedding space of the T2I pipeline. Figure 4 demonstrates" + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 308, + 366, + 542, + 502 + ], + "blocks": [ + { + "bbox": [ + 46, + 312, + 547, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 547, + 357 + ], + "type": "text", + "content": "Figure 4. Controlled composite responsible generation using the proposed method. By using different concepts in " + }, + { + "bbox": [ + 46, + 312, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 46, + 312, + 547, + 357 + ], + "type": "text", + "content": " in Eq. (1), and employing dual space control in (7), our technique can enable responsibility along single or multiple concepts, as desired. Provided the prompt \"A doctor\" to Stable Diffusion, alignment is achieved for a target concept composition (top label) for the attributes noted at the bottom of each image set. See supplementary material for more examples." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 366, + 542, + 502 + ], + "lines": [ + { + "bbox": [ + 308, + 366, + 542, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 542, + 502 + ], + "type": "image", + "image_path": "3a5bb74c62d21feb7b64518ece0ba11f8cc3b2b93c9791bce1c5118335248016.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 514, + 545, + 571 + ], + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 571 + ], + "type": "text", + "content": "Figure 5. Examples of extending " + }, + { + "bbox": [ + 304, + 514, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{resp}}" + }, + { + "bbox": [ + 304, + 514, + 545, + 571 + ], + "type": "text", + "content": " with additional attributes - possibly unrelated to core responsible concepts. Extensions with smile and glasses are shown. In addition to varying images along responsible concepts of age, race and gender, our method is able to seamlessly incorporate the additional concepts in images." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": "this capability within " + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{resp}}" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": ", showing how distinct responsible aspects, such as age, gender, and race; can be independently learned and then composed to produce varied yet coherent outputs. By manipulating combinations within " + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_X" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": ", we demonstrate an excellent control achieved with interpretable variations across the composite attribute space, effectively balancing diversity with responsible aspects." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Scalability of Concept Spaces: In our original experiments, the responsible concept set " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{resp}}" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " in our framework, as defined in Eq. (1), included core attributes of " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{race}}, \\mathcal{A}_{\\mathrm{gender}}, \\mathcal{A}_{\\mathrm{age}}," + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{safe}}" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": ". This was only to benchmark" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2982" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 287, + 163 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 287, + 163 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 287, + 163 + ], + "type": "image", + "image_path": "a40fef026c08f663c6167e331602b0476f7f0a15f7181530f7ccb117ce1c1006.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 172, + 287, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 172, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 287, + 240 + ], + "type": "text", + "content": "Figure 6. Representative curtailed example of transitions caused by RICE and RIIDL to an output. An image of 'a doctor' is generated and 'female' concept is added individually by the modules. Increasingly changing the influence of the added concept representation provided by each module leads to unique alteration to the generated output." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 246, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 413 + ], + "type": "text", + "content": "the method with the existing approaches. Our design allows the \"responsible\" set to be extendable. This scalability allows for seamlessly handling additional nuanced attributes in images, such as facial expressions (e.g., smiling) and accessories (e.g., glasses). Such concepts can act as supplementary attributes with the existing attributes, enabling complex yet interpretable multi-aspect control. In Fig. 5, we provide representative examples of scaling up the considered concepts set. It leads to combining the responsible subspaces of the core responsible features with subspaces for smiling and glasses. It is observable that our framework is scalable to not only core responsibility related concepts, but to any semantically meaningful concepts. We also provide additional results in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 414, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 288, + 675 + ], + "type": "text", + "content": "Individual Modules: Our method takes advantage of two modules, namely; of RICE (§ 3.3) and RIIDL (§ 3.4). To assess the specific contribution of each component within the responsible generation framework, we perform a systematic ablation by individually removing and adjusting key modules. This analysis allows selectively varying the influence both modules on the output. When only RICE operates, modifications originally occur in the embedding concept space for the T2I pipeline, facilitating transitions from base to responsible concept spaces. The RIIDL module contribute an additional layer of fine-grained control, allowing nuanced and context-sensitive adjustments within the intermediate latents during diffusion. This capability supports precise tuning of details without overriding the global context, complementing the broader adjustments. In tandem, these modules establish a dynamic balance within the dual-space intervention strategy, where RICE provides prediffusion global control, and RIIDL governs the responsible aspects during the denoising process. Whereas we provide further results supporting these point in the supplementary material, in Fig. 6 a curtailed representative example is provided to show the visual effects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "In the figure, we generate images using prompt 'a doctor' and with an added concept 'female' from the individual modules. The 'female' concept representation " + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{female}}" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": " pro" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "vided by the modules is added to the original corresponding representations with an increasing degree of influence on the output. It can be observed that for RICE, this results in more prominent shifts in appearance and attire, indicating a significant control. For RIIDL, the latent space provides more refined transitions with subtle and smoother progression of the visual features. On one hand, these results demonstrate the complementary nature of the proposed modules, on the other; they establish the effectiveness of the modules as stand-alone add-ons." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 197, + 379, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 197, + 379, + 209 + ], + "spans": [ + { + "bbox": [ + 306, + 197, + 379, + 209 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 214, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 545, + 430 + ], + "type": "text", + "content": "This research presented a unique approach to control generative outputs of text-to-image (T2I) models in 'fair and safe' domain. Our method induces two plug-and-play modules that can be plugged into the textual embedding space and diffusion latent space of the T2I pipeline. The modules are neural models learned by distilling the text encoder and diffusion model of the original pipeline while conditioning the knowledge distillation on a pre-defined responsible concepts set, as desired by the user. Our modules also account for semantic interpretability by leveraging concept whitening. We demonstrate that both modules can explicitly contribute to responsible image generation in a complementary manner. Our extensive benchmarking with state-of-the-art methods for fair and safe image generation on two standard datasets show highly promising results. We also show that our method can conveniently deal with multiple concepts in the responsible image space while being seamlessly extendable to more broader concepts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "text", + "content": "Limitations and ethics concerns: A potential shortcoming of our method is that, to enable a comprehensive control over the generative output, it employs multiple hyperparameters, e.g., " + }, + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "text", + "content": "'s in Eq. (7). Tuning these hyperparameters can potentially be seen as a limitation of our approach. However, the main purpose of allowing these adjustable hyper-parameters in our design is to provide a finer control to the user. It is convenient to adjust their values because the controls they provide have clear high-level interpretations. We provide guidelines on the hyper-parameter value adjustment in the supplementary material. From the ethics viewpoint, although our aim is to ensure responsible outputs; a potential misuse of exploiting the strong output control of our technique can be by replacing the " + }, + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{resp}}" + }, + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "text", + "content": " with another set " + }, + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{irresp}}" + }, + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "text", + "content": ". In " + }, + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{irresp}}" + }, + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "text", + "content": ", unsafe and unfair concepts can be included. This can have an opposite impact of our technique. Hence, our method should only be used when the user has complete control over the set " + }, + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{resp}}" + }, + { + "bbox": [ + 304, + 431, + 545, + 646 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 647, + 545, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 706 + ], + "type": "text", + "content": "Acknowledgment: Naveed Akhtar is a recipient of the Australian Research Council Discovery Early Career Researcher Award (project # DE230101058) funded by the Australian Government. This work is also partially supported by Google Research Scholar Program Award." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2983" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 102, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 102, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 102, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 89, + 288, + 693 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 89, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 89, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 89, + 287, + 123 + ], + "type": "text", + "content": "[1] Sandhini Agarwal et al. \"Evaluating CLIP: towards characterization of broader capabilities and downstream implications\". In: arXiv preprint arXiv:2108.02818 (2021)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 126, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 126, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 54, + 126, + 288, + 160 + ], + "type": "text", + "content": "[2] Zhongjie Ba et al. \"SurrogatePrompt: Bypassing the Safety Filter of Text-To-Image Models via Substitution\". In: arXiv preprint arXiv:2309.14122 (2023)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 288, + 196 + ], + "type": "text", + "content": "[3] Hritik Bansal et al. \"How well can Text-to-Image Generative Models understand Ethical Natural Language Interventions?\" In: arXiv preprint arXiv:2210.15230 (2022)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 201, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 287, + 233 + ], + "type": "text", + "content": "[4] Hugo Berg et al. \"A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning\". In: arXiv preprint arXiv:2203.11933 (2022)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 237, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 287, + 270 + ], + "type": "text", + "content": "[5] Federico Bianchi et al. \"Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale\". In: arXiv preprint arXiv:2211.03759 (2022)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 274, + 287, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 287, + 318 + ], + "type": "text", + "content": "[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. \"Multimodal datasets: misogyny, pornography, and malignant stereotypes\". In: arXiv preprint arXiv:2110.01963 (2021)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 322, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 287, + 354 + ], + "type": "text", + "content": "[7] Manuel Brack et al. \"The Stable Artist: Steering Semantics in Diffusion Latent Space\". In: arXiv preprint arXiv:2212.06013 (2022)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 358, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 287, + 402 + ], + "type": "text", + "content": "[8] Nicholas Carlini et al. Extracting Training Data from Diffusion Models. 2023. DOI: 10.48550/ARXIV.2301.13188. URL: https://arxiv.org/abs/2301.13188." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 407, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 287, + 440 + ], + "type": "text", + "content": "[9] Zhi Chen, Yijie Bei, and Cynthia Rudin. \"Concept Whitening for Interpretable Image Recognition\". In: Nature Machine Intelligence 2.12 (Dec. 2020), pp. 772-782." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 50, + 443, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 443, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 50, + 443, + 287, + 498 + ], + "type": "text", + "content": "[10] Jaemin Cho, Abhay Zala, and Mohit Bansal. “Dall-eval: Probing the reasoning skills and social biases of text-to-image generation models”. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 3043–3054." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 50, + 502, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 502, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 50, + 502, + 287, + 535 + ], + "type": "text", + "content": "[11] Ching-Yao Chuang et al. \"Debiasing vision-language models via biased prompts\". In: arXiv preprint arXiv:2302.00070 (2023)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 50, + 539, + 287, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 539, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 50, + 539, + 287, + 572 + ], + "type": "text", + "content": "[12] Florinel-Alin Croitoru et al. \"Diffusion models in vision: A survey\". In: IEEE Transactions on Pattern Analysis and Machine Intelligence 45.9 (2023), pp. 10850-10869." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 50, + 575, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 575, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 50, + 575, + 287, + 609 + ], + "type": "text", + "content": "[13] Rohit Gandikota et al. “Erasing concepts from diffusion models”. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 2426–2436." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 50, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 50, + 613, + 287, + 656 + ], + "type": "text", + "content": "[14] Rohit Gandikota et al. \"Unified concept editing in diffusion models\". In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024, pp. 5111-5120." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 50, + 660, + 287, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 660, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 50, + 660, + 287, + 693 + ], + "type": "text", + "content": "[15] Jindong Gu et al. \"Towards robust prompts on vision-language models\". In: arXiv preprint arXiv:2304.08479 (2023)." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 309, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 309, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 309, + 73, + 545, + 117 + ], + "type": "text", + "content": "[16] René Haas et al. \"Discovering interpretable directions in the semantic latent space of diffusion models\". In: 2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG). IEEE. 2024, pp. 1-9." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 309, + 120, + 545, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 120, + 545, + 153 + ], + "spans": [ + { + "bbox": [ + 309, + 120, + 545, + 153 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. “Denoising diffusion probabilistic models”. In: Advances in neural information processing systems 33 (2020), pp. 6840–6851." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 309, + 157, + 545, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 157, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 309, + 157, + 545, + 190 + ], + "type": "text", + "content": "[18] Chi-Pin Huang et al. \"Receler: Reliable concept erasing of text-to-image diffusion models via lightweight erasers\". In: arXiv preprint arXiv:2311.17717 (2023)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 309, + 193, + 545, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 193, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 309, + 193, + 545, + 236 + ], + "type": "text", + "content": "[19] Tero Karras et al. “Elucidating the design space of diffusion-based generative models”. In: Advances in neural information processing systems 35 (2022), pp. 26565–26577." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 309, + 239, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 239, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 309, + 239, + 545, + 295 + ], + "type": "text", + "content": "[20] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. \"Diffusionclip: Text-guided diffusion models for robust image manipulation\". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022, pp. 2426-2435." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 309, + 298, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 298, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 309, + 298, + 545, + 331 + ], + "type": "text", + "content": "[21] Diederik Kingma et al. \"Variational diffusion models\". In: Advances in neural information processing systems 34 (2021), pp. 21696-21707." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 309, + 335, + 545, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 335, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 309, + 335, + 545, + 378 + ], + "type": "text", + "content": "[22] Nupur Kumari et al. \"Ablating concepts in text-to-image diffusion models\". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 22691-22702." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 309, + 381, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 381, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 309, + 381, + 545, + 426 + ], + "type": "text", + "content": "[23] Hang Li et al. \"Self-discovering interpretable diffusion latent directions for responsible text-to-image generation\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 309, + 429, + 545, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 429, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 309, + 429, + 545, + 462 + ], + "type": "text", + "content": "[24] Sasha Luccioni et al. \"Stable bias: Evaluating societal representations in diffusion models\". In: Advances in Neural Information Processing Systems 36 (2024)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 309, + 465, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 465, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 309, + 465, + 545, + 510 + ], + "type": "text", + "content": "[25] Chenlin Meng et al. \"On distillation of guided diffusion models\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 14297-14306." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 309, + 514, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 514, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 309, + 514, + 545, + 545 + ], + "type": "text", + "content": "[26] Ranjita Naik and Besmira Nushi. \"Social Biases through the Text-to-Image Generation Lens\". In: arXiv preprint arXiv:2304.06034 (2023)." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 309, + 549, + 545, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 549, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 309, + 549, + 545, + 582 + ], + "type": "text", + "content": "[27] Alex Nichol et al. \"Glide: Towards photorealistic image generation and editing with text-guided diffusion models\". In: arXiv preprint arXiv:2112.10741 (2021)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 309, + 585, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 585, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 309, + 585, + 545, + 630 + ], + "type": "text", + "content": "[28] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. \"Editing implicit assumptions in text-to-image diffusion models\". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 7053-7061." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 309, + 633, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 633, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 309, + 633, + 545, + 666 + ], + "type": "text", + "content": "[29] Yong-Hyun Park et al. \"Direct unlearning optimization for robust and safe text-to-image models\". In: arXiv preprint arXiv:2407.21035 (2024)." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 309, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 309, + 669, + 545, + 713 + ], + "type": "text", + "content": "[30] William Peebles and Saining Xie. \"Scalable diffusion models with transformers\". In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 4195-4205." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2984" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 50, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 50, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 287, + 106 + ], + "type": "text", + "content": "[31] Alec Radford et al. \"Learning Transferable Visual Models From Natural Language Supervision\". In: arXiv preprint arXiv:2103.00020 (2021)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 109, + 288, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 109, + 288, + 142 + ], + "spans": [ + { + "bbox": [ + 50, + 109, + 288, + 142 + ], + "type": "text", + "content": "[32] Aditya Ramesh et al. \"Hierarchical text-conditional image generation with clip latents\". In: arXiv preprint arXiv:2204.06125 1.2 (2022), p. 3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 145, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 145, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 145, + 287, + 178 + ], + "type": "text", + "content": "[33] Javier Rando et al. Red-Teaming the Stable Diffusion Safety Filter. 2022. DOI: 10.48550/ARXIV.2210.04610. URL: https://arxiv.org/abs/2210.04610." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 50, + 182, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 182, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 50, + 182, + 287, + 226 + ], + "type": "text", + "content": "[34] Robin Rombach et al. \"High-resolution image synthesis with latent diffusion models\". In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022, pp. 10684-10695." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 50, + 229, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 229, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 50, + 229, + 287, + 274 + ], + "type": "text", + "content": "[35] Robin Rombach et al. \"High-resolution image synthesis with latent diffusion models\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022, pp. 10684-10695." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 276, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 276, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 50, + 276, + 287, + 320 + ], + "type": "text", + "content": "[36] Chitwan Sahara et al. \"Photorealistic text-to-image diffusion models with deep language understanding\". In: Advances in neural information processing systems 35 (2022), pp. 36479-36494." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 50, + 323, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 323, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 50, + 323, + 287, + 390 + ], + "type": "text", + "content": "[37] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. \"Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content?\" In: Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency. 2022, pp. 1350-1361." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 50, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 50, + 393, + 287, + 437 + ], + "type": "text", + "content": "[38] Patrick Schramowski et al. \"Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models\". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 22522-22531." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 50, + 440, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 440, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 50, + 440, + 287, + 474 + ], + "type": "text", + "content": "[39] Shoaib Ahmed Siddiqui et al. \"Metadata Archaeology: Unearthing Data Subsets by Leveraging Training Dynamics\". In: arXiv preprint arXiv:2209.10015 (2022)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 50, + 476, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 476, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 50, + 476, + 287, + 510 + ], + "type": "text", + "content": "[40] Gowthami Somepalli et al. \"Diffusion Art or Digital Forgery? Investigating Data Replication in Diffusion Models\". In: arXiv preprint arXiv:2212.03860 (2022)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 50, + 513, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 513, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 50, + 513, + 287, + 556 + ], + "type": "text", + "content": "[41] Lukas Struppek, Dominik Hintersdorf, and Kristian Kersting. \"The biased artist: Exploiting cultural biases via homoglyphs in text-guided image generation models\". In: (2022)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 50, + 559, + 287, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 559, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 50, + 559, + 287, + 593 + ], + "type": "text", + "content": "[42] Yu-Lin Tsai et al. \"Ring-A-Bell! How Reliable are Concept Removal Methods for Diffusion Models?\" In: arXiv preprint arXiv:2310.10012 (2023)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 50, + 597, + 287, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 597, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 50, + 597, + 287, + 630 + ], + "type": "text", + "content": "[43] Peiran Wang et al. \"Moderator: Moderating Text-to-Image Diffusion Models through Fine-grained Context-based Policies\". In: arXiv preprint arXiv:2408.07728 (2024)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 50, + 633, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 633, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 50, + 633, + 287, + 666 + ], + "type": "text", + "content": "[44] Zihao Wang et al. \"Concept algebra for text-controlled vision models\". In: arXiv preprint arXiv:2302.03693 2 (2023)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 50, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 50, + 669, + 287, + 712 + ], + "type": "text", + "content": "[45] Yongliang Wu et al. \"Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient\". In: arXiv preprint arXiv:2405.15304 (2024)." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 309, + 72, + 545, + 349 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 106 + ], + "type": "text", + "content": "[46] Tianwei Xiong et al. \"Editing Massive Concepts in Text-to-Image Diffusion Models\". In: arXiv preprint arXiv:2403.13807 (2024)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 309, + 109, + 545, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 109, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 309, + 109, + 545, + 142 + ], + "type": "text", + "content": "[47] Yijun Yang et al. \"GuardT2I: Defending Text-to-Image Models from Adversarial Prompts\". In: arXiv preprint arXiv:2403.01446 (2024)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 309, + 146, + 545, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 146, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 309, + 146, + 545, + 179 + ], + "type": "text", + "content": "[48] Jaehong Yoon et al. \"SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation\". In: arXiv preprint arXiv:2410.12761 (2024)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 309, + 183, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 183, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 309, + 183, + 545, + 228 + ], + "type": "text", + "content": "[49] Gong Zhang et al. “Forget-me-not: Learning to forget in text-to-image diffusion models”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024, pp. 1755–1764." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 309, + 232, + 545, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 232, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 309, + 232, + 545, + 264 + ], + "type": "text", + "content": "[50] Hongxiang Zhang, Yifeng He, and Hao Chen. \"Steerdiff: Steering towards safe text-to-image diffusion models\". In: arXiv preprint arXiv:2410.02710 (2024)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 309, + 268, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 268, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 309, + 268, + 545, + 312 + ], + "type": "text", + "content": "[51] Yimeng Zhang et al. \"To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images... for now\". In: European Conference on Computer Vision. Springer. 2025, pp. 385-403." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 309, + 316, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 316, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 309, + 316, + 545, + 349 + ], + "type": "text", + "content": "[52] Jieyu Zhao et al. \"Gender bias in coreference resolution: Evaluation and debiasing methods\". In: arXiv preprint arXiv:1804.06876 (2018)." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2985" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/fb8f3fbf-9e73-432d-9130-bbe1d55d00ee_content_list.json b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/fb8f3fbf-9e73-432d-9130-bbe1d55d00ee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e1bff168c0c3d4693b7b401b6f5f6cd54ae42781 --- /dev/null +++ b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/fb8f3fbf-9e73-432d-9130-bbe1d55d00ee_content_list.json @@ -0,0 +1,1942 @@ +[ + { + "type": "text", + "text": "Plug-and-Play PPO: An Adaptive Point Prompt Optimizer Making SAM Greater", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xueyu Liu $^{1\\dagger}$ , Rui Wang $^{2\\dagger}$ , Yexin Lai $^{1}$ , Guangze Shi $^{1}$ , Feixue Shao $^{1}$ , Fang Hao $^{1}$ , Jianan Zhang $^{1}$ , Jia Shen $^{1}$ , Yongfei Wu $^{1*}$ , Wen Zheng $^{1}$", + "bbox": [ + 181, + 179, + 815, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Taiyuan University of Technology", + "bbox": [ + 359, + 215, + 638, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Science and Technology of China", + "bbox": [ + 307, + 234, + 691, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Powered by extensive curated training data, the Segment Anything Model (SAM) demonstrates impressive generalization capabilities in open-world scenarios, effectively guided by user-provided prompts. However, the class-agnostic characteristic of SAM renders its segmentation accuracy highly dependent on prompt quality. In this paper, we propose a novel plug-and-play dual-space Point Prompt Optimizer (PPO) designed to enhance prompt distribution through deep reinforcement learning (DRL)-based heterogeneous graph optimization. PPO optimizes initial prompts for any task without requiring additional training, thereby improving SAM's downstream segmentation performance. Specifically, PPO constructs a dual-space heterogeneous graph, leveraging the robust feature-matching capabilities of a foundational pre-trained model to create internal feature and physical distance matrices. A DRL policy network iteratively refines the distribution of prompt points, optimizing segmentation predictions. We conducted experiments on four public datasets. The ablation study explores the necessity and balance of optimizing prompts in both feature and physical spaces. The comparative study shows that PPO enables SAM to surpass recent one-shot methods. Additionally, experiments with different initial prompts demonstrate PPO's generality across prompts generated by various methods. In conclusion, PPO redefines the prompt optimization problem as a heterogeneous graph optimization task, using DRL to construct an effective, plug-and-play prompt optimizer. This approach holds potential for broader applications across diverse segmentation tasks and provides a promising solution for point prompt optimization. The source code and demo are available at https://github.com/XueyuLiu/PPO.", + "bbox": [ + 88, + 316, + 485, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 827, + 220, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Segmentation, as a fundamental task, plays a pivotal role in visual understanding systems. The prevailing train-", + "bbox": [ + 89, + 845, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/acd48d42f961ae46fe3db906271ce5cafb1374b8a88812eab8054f7e2fbed250.jpg", + "image_caption": [ + "Figure 1. Ineffective prompts can diminish SAM's performance. Incorrect prompts may cause the segmentation mask to capture irrelevant regions, while excessive prompts for adjacent areas can lead to neglect of other parts of the object. Therefore, prompt optimization is essential to improve SAM's performance." + ], + "image_footnote": [], + "bbox": [ + 522, + 282, + 893, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing paradigm for segmentation relies on data-driven approaches, where large datasets are used to train models. While this method has achieved remarkable success, it faces several challenges, including the computational resources to train models for new tasks and the risk of domain-specific optimizations that can limit generalizability [18, 22].", + "bbox": [ + 509, + 612, + 906, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent breakthroughs in Pretrained Foundation Models (PFMs) in computer vision (CV) have demonstrated impressive capabilities in domain-invariant segmentation across diverse scenarios [11, 29]. Vision Foundation Models (VFMs), such as DINOv2 [2, 27], acquire generalized visual features by capturing intricate information at both the image and pixel levels using only raw image data. This approach enables these models to learn generic visual representations, ensuring robust zero-shot transferability for various downstream tasks. Meanwhile, the Segment Anything Model (SAM) [14, 30], trained on large-scale, high-quality datasets, has shown a strong ability to generalize across tasks and data distributions that are not encountered during", + "bbox": [ + 509, + 704, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. †Co-first author.", + "bbox": [ + 114, + 886, + 331, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4332", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "training. However, SAM relies heavily on manual prompts, such as bounding boxes and points, to identify objects of interest, as it is pretrained in a class-agnostic manner. Consequently, achieving task-specific adaptation of SAM is crucial to bridge the gap between its class-agnostic predictions and the specific needs of industrial applications.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some works attempt to train prompt encoders that enables SAM to adapt to downstream segmentation tasks [8, 15, 34, 43]. However, these task-specific prompt encoders require additional training for new segmentation scenarios. Recently, other efforts have focused on automatically generating geometric visual prompts, such as point prompts, to enhance SAM's segmentation performance [21, 23, 37]. In such methods, the quality of prompt schemes remains a critical factor influencing downstream segmentation performance. As illustrated in Fig. 1, incorrect prompts can result in erroneous segmentation, while excessive prompts in specific areas may cause SAM to focus disproportionately on those regions, neglecting other parts of the target. Current automatic prompt engineering methods often lack optimization in prompt distribution, leading to lower-quality prompts and segmentation performance. Therefore, there is an urgent need for methods that can refine prompts to effectively enhance SAM's performance.", + "bbox": [ + 91, + 185, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, effective point prompts for SAM should exhibit certain characteristics in both feature and physical spaces. In the feature space, regions containing positive prompts should be homogeneous to ensure consistency and clarity within the segmentation target, while regions between positive and negative prompts should be heterogeneous, distinguishing the target from the background. In the physical space, positive and negative prompts should be sufficiently spaced apart to mitigate the effects of excessive prompts and prevent over-segmentation, while also being close enough to boundaries to ensure precise delineation.", + "bbox": [ + 89, + 459, + 482, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon the aforementioned inspiration and committed to addressing the challenges mentioned above, we propose a novel Plug-and-Play Point Prompt Optimizer (PPO) designed to enhance SAM's segmentation performance by automatically optimizing initial point prompt schemes. Leveraging DINOv2's powerful feature association capabilities [27], we construct an internal feature and physical distance matrix within each image to create a heterogeneous graph environment. To optimize prompt distribution in dual spaces, we employ Deep Reinforcement Learning (DRL), a framework that trains agents to maximize cumulative rewards through interaction and feedback mechanisms. By designing specific actions and reward functions, PPO optimizes prompt point distribution under the guidance of DRL and inputs them into SAM to obtain refined segmentation results. We validated our approach through experiments on two natural and two medical image datasets. Ablation studies demonstrate that PPO", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "optimized prompt distributions significantly enhance segmentation performance. Comparative experiments show that PPO outperforms recent methods that have not been trained on in-domain datasets. Additionally, experiments with different initial prompts confirm PPO's generality across prompts generated by various methods. The main contributions are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to define an optimization direction for point prompts in both feature and physical spaces. Additionally, we are the first to use DRL to optimize point prompts, achieving prompt optimization without additional training for new tasks.", + "- PPO operates in a plug-and-play manner, allowing for the optimization of any initial geometric point prompts to improve SAM's segmentation performance.", + "- Extensive experiments demonstrate the effectiveness of PPO in optimizing point prompts across various initial prompt schemes." + ], + "bbox": [ + 513, + 199, + 903, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 387, + 653, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Pretrained Foundation Models", + "text_level": 1, + "bbox": [ + 513, + 407, + 785, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In natural language processing (NLP), PFMs, such as BERT [5], ChatGPT, and GPT-4, have become foundational tools for a wide range of downstream tasks across diverse data modalities. By leveraging extensive text data, PFMs provide effective parameter initialization for various NLP applications [44]. The success of PFMs in NLP has inspired researchers to explore their application in CV. Visual language models like CLIP [29] and GLIP [19], trained on large-scale datasets of millions of image-text pairs from the web, exhibit strong zero-shot generalization across multiple vision tasks, offering representations with robust generalization and transfer capabilities.", + "bbox": [ + 511, + 426, + 906, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on this trend, Meta AI introduced the SAM, a notable advancement in applying VFMs. SAM allows users to specify image segmentation tasks through prompts and demonstrates impressive zero-shot learning, enabling it to generalize effectively across diverse tasks without task-specific training [14]. However, SAM's segmentation performance heavily relies on the quality of input prompts. To address this, we propose an adaptive point prompt optimizer that automatically refines prompts to enhance SAM's segmentation performance.", + "bbox": [ + 511, + 609, + 905, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In another significant development, DINOv2 [27] leverages discriminative self-supervised learning at both image and patch levels, showing exceptional patch-matching capabilities that allow it to capture semantic similarities across images of the same target. By utilizing these patch-matching capabilities, we can extract feature distances between patches within an image, providing a theoretical foundation for constructing heterogeneous graphs to achieve effective prompt optimization.", + "bbox": [ + 511, + 763, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4333", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Automatic Prompt Engineering", + "text_level": 1, + "bbox": [ + 89, + 90, + 369, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the field of NLP, extensive researches have been conducted on designing automatic prompt engineering, with prompt tuning methods achieving significant advances in enhancing PFMs' performance through optimized language prompts [1, 12, 13, 16, 28, 41, 45, 45]. Simultaneously, efforts have been made to explore automatic prompt engineering in CV. On the one hand, some recent works have attempted to train prompt encoders that enable SAM to adapt to downstream segmentation tasks [9, 32, 34, 40]. For example, Sun et al. proposed VRP-SAM, an extension of SAM that encodes various granularities of visual references directly into prompt embeddings for mask generation [34]. These approaches encode SAM's prompts into prompt embeddings via a prompt encoder, which are then fed into SAM's mask decoder to generate target masks. However, these methods are often task-sensitive, requiring training different prompt encoders for different tasks, thus limiting rapid task transfer.", + "bbox": [ + 89, + 109, + 483, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, some works attempt to generate geometric prompts to achieve automatic prompt engineering for SAM, generally following two main paradigms. The first paradigm uses a small amount of annotated data to train a segmentation model, then samples prompts from pseudolabels generated by this model [3, 26, 36, 42, 43]. For instance, Li et al. randomly selected prompt points from masks generated by coarse segmentation, which were subsequently input into SAM after fine-tuning to obtain segmentation results [17]. Wang et al. utilized a novel prototype learning framework to generate pseudo-labels and extract prompts, which were then incorporated into SAM for segmentation [35]. Dai et al. developed various methods to expand prompt points, exploring the potential of a single prompt point for enhanced segmentation performance [4].", + "bbox": [ + 89, + 383, + 483, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, this first paradigm still requires a small amount of data to train the prompt generator, also limiting task transferability. To improve transferability, the second paradigm generates initial prompts by leveraging feature similarity between reference and target images without additional training [21, 23, 37, 43]. For example, Zhang et al. calculated feature similarity between pixels in the target and reference images to create a point prompt scheme using only a single reference image [43]. Liu et al. introduced Matcher, an instance-level and bidirectional matching strategy for prompt point generation [23], and Liu et al. proposed a series of prompt optimization strategies for the automatic segmentation of basement membranes in electron microscopy images [21]. However, the quality of prompts generated by these methods is often limited, as they do not sufficiently consider the relationships between positive and negative prompts, resulting in redundant and erroneous prompts that negatively affect segmentation performance.", + "bbox": [ + 89, + 611, + 483, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our work, PPO addresses this limitation by con", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "structuring a heterogeneous graph based on dual-space distances between initial prompts within each image. This approach optimizes any automatically generated geometric point prompts without additional training for new tasks, enhancing SAM's segmentation performance while preserving task transferability by ensuring high prompt quality.", + "bbox": [ + 511, + 90, + 905, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 196, + 648, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces our plug-and-play PPO, which is based on a heterogeneous graph structure designed to enhance the segmentation performance of SAM by optimizing initial point prompts. An overview of PPO is illustrated in Fig. 2. Specifically, we first compute the feature distance matrix and the physical distance matrix for each image. Using these matrices, we construct a heterogeneous graph where the edge attributes are defined by the feature and physical distances based on the initial prompt indices. During the training phase, we use this heterogeneous graph as the environment to design an action list and reward function to train the optimization agent. In the inference phase, the trained agent activates the appropriate point prompts as a final prompt scheme, which is then input into SAM to obtain the final segmentation result. In the following subsections, we will describe the main components in detail.", + "bbox": [ + 511, + 215, + 906, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dual-Space Distance Matrix Calculation", + "text_level": 1, + "bbox": [ + 511, + 482, + 861, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core objective of our method is to optimize positive and negative point prompts in both feature and distance spaces. To achieve this, we construct a heterogeneous graph comprising two homogeneous graphs that represent the feature and distance spaces. Edge values are determined by creating two patch-level distance matrices for each image $X$ .", + "bbox": [ + 511, + 500, + 905, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we divide $X$ into patches $x$ using a sliding window approach, where $x_{i} = x_{1},x_{2},\\ldots ,x_{n}$ . We then use the image encoder from DINOv2 [27] to extract patch-level features for $x$ , represented by $f$ , where $f_{i} = f_{1},f_{2},\\dots ,f_{n}$ . Next, we calculate the feature matrix $M_{f}$ as follows:", + "bbox": [ + 511, + 590, + 905, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {f} (i, j) = \\left\\| f _ {i} - f _ {j} \\right\\|, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 680, + 906, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M_{f}(i,j)$ represents the Euclidean distance $D_{f}$ between the $i$ -th patch feature $f_{i}$ and the $j$ -th patch feature $f_{j}$ . Similarly, we calculate the physical distance matrix $M_{p}$ using the center points of the patches $x$ as follows:", + "bbox": [ + 511, + 708, + 905, + 771 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {p} (i, j) = \\left\\| x _ {i} - x _ {j} \\right\\|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 782, + 906, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M_{p}(i,j)$ represents the Euclidean distance $D_{p}$ between the $i$ -th patch center $x_{i}$ and the $j$ -th patch center $x_{j}$ .", + "bbox": [ + 511, + 811, + 905, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Graph Environment Creation", + "text_level": 1, + "bbox": [ + 511, + 852, + 779, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by coarse segmentation sampling [3, 26] and feature matching techniques [21, 23], we can obtain initial", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4334", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a56ba54ca5273f651e41db2d47198126b844b525005abab8f95dd85e2bbb526e.jpg", + "image_caption": [ + "Figure 2. The PPO workflow is designed to enhance SAM's segmentation performance by optimizing initial point prompts through two key components: graph-based environment creation and agent training for prompt optimization. After optimized, the prompts are fed into SAM to obtain segmentation results." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "point prompt indices for optimization. Specifically, we define a heterogeneous graph $G = (V,E,\\phi ,\\psi)$ , where $V$ is the set of nodes and $E$ is the set of edges. The node type set $\\mathcal{A} = \\{\\text{positive nodes } V_p, \\text{negative nodes } V_n\\}$ is represented by the mapping function $\\phi (v): V \\to \\mathcal{A}$ , and the edge type set $\\mathcal{R} = \\{\\text{feature edges } E_f, \\text{physical edges } E_p\\}$ is represented by the mapping function $\\psi (e): E \\to \\mathcal{R}$ .", + "bbox": [ + 88, + 518, + 480, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the heterogeneous graph $G$ as the environment for DRL. By selecting different nodes, we can obtain different edge values as feedback.", + "bbox": [ + 89, + 625, + 482, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Agent Training for Prompt Optimization", + "text_level": 1, + "bbox": [ + 89, + 685, + 441, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the one hand, an effective point prompt scheme should ensure correctness. In an image, the features of the segmentation target are often homogeneous, while the features of the background are heterogeneous. Therefore, we aim for the feature distance between positive point prompts to be sufficiently close, and the feature distance between positive and negative point prompts to be sufficiently far. On the other hand, an effective point prompt scheme should also ensure sparsity. Over-prompting in adjacent areas can degrade segmentation performance. Thus, the physical distance between positive point prompts (and between negative point prompts) should not be too close. Additionally, to enhance boundary segmentation accuracy, the physical dis", + "bbox": [ + 88, + 704, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tance between positive and negative point prompts should not be too far.", + "bbox": [ + 511, + 518, + 906, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A DRL model can infer actions and receive rewards from the environment, reflecting the significance of these actions and updating the agent's state accordingly [38]. Using DRL, we aim to train an agent to activate appropriate point prompts to optimize the distribution. The detailed process is shown in Algorithm 1. The key components of agent training for prompt optimization will be introduced as follows:", + "bbox": [ + 511, + 549, + 906, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Actions: We define three possible actions: (1) restore node $A_r$ , which randomly activates a node and connects the corresponding edges; (2) delete node $A_d$ , which randomly deletes a node and removes the corresponding edges; (3) move node $A_m$ , which randomly selects a direction and moves the index position of a node. Additionally, we aim to maintain a stable number of prompts. Therefore, if the number of nodes $|V_p|$ or $|V_n|$ in the current state is less than the minimum value, the only possible action is $A_r$ . Conversely, if the $|V_p|$ or $|V_n|$ exceeds the maximum value, the only possible action is $A_d$ . Only when the $|V_p|$ or $|V_n|$ is between the minimum and maximum values are all three actions, $A_r$ , $A_d$ , and $A_m$ , possible.", + "bbox": [ + 511, + 672, + 908, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "States: For each action, there will be different nodes, resulting in a different heterogeneous graph $G^{s}$ . Based on", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4335", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 PPO training process", + "text_level": 1, + "bbox": [ + 91, + 90, + 323, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Heterogeneous graph environment $G$", + "bbox": [ + 91, + 108, + 387, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parameter: Max step; Max epoch;", + "bbox": [ + 91, + 123, + 326, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Max nodes $N_{max}$ ; Min nodes $N_{min}$", + "bbox": [ + 91, + 138, + 330, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Q-values", + "text_level": 1, + "bbox": [ + 91, + 154, + 215, + 167 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initializing the Q-values, Epoch, Step.", + "2: while Epoch < Max epoch do", + "3: Initializing the Step.", + "4: while Step $<$ Max step do", + "5: Emptying all nodes and edges.", + "6: Emptying reward.", + "7: Calculating the number of Nodes $N_{pos}$ and $N_{neg}$ .", + "8: if $N_{pos} < N_{min}$ or $N_{neg} < N_{min}$ then", + "9: Action list $= \\{A_{r}\\}$", + "10: else if $N_{pos} > N_{max}$ or $N_{neg} > N_{max}$ then", + "11: Action list $= \\{A_{d}\\}$", + "12: else", + "13: Action list $= \\{A_r, A_d, A_m\\}$ .", + "14: end if", + "15: Perform a random action and update the status $s$", + "16: Adding the reward $r_s$ based on the reward function.", + "17: Updating Q-values.", + "18: Step+=1.", + "19: end while", + "20: Epoch $+ = 1$", + "21: end while", + "22: return solution" + ], + "bbox": [ + 96, + 170, + 480, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$G^{s}$ , we can calculate the following mean edge values:", + "bbox": [ + 89, + 546, + 444, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {f} ^ {s} (p, p) = \\frac {1}{| V _ {p} | | V _ {p} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {p}} M _ {f} (i, j), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 571, + 482, + 611 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {f} ^ {\\mathrm {s}} (p, n) = \\frac {1}{| V _ {p} | | V _ {n} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {n}} M _ {f} (i, j), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 625, + 482, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {p} ^ {s} (p, p) = \\frac {1}{| V _ {p} | | V _ {p} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {p}} M _ {p} (i, j), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 670, + 482, + 710 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {p} ^ {s} (n, n) = \\frac {1}{| V _ {n} | | V _ {n} |} \\sum_ {i \\in V _ {n}} \\sum_ {j \\in V _ {n}} M _ {p} (i, j), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 718, + 482, + 756 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {p} ^ {s} (p, n) = \\frac {1}{| V _ {p} | | V _ {n} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {n}} M _ {p} (i, j). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 763, + 482, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In state $s$ , $D_{f}^{s}(p,p)$ and $D_{f}^{s}(p,n)$ represents the mean feature distance between positive prompts, and between positive and negative prompts. Similarly, $D_{p}^{s}(p,p)$ , $D_{p}^{s}(n,n)$ , and $D_{p}^{s}(p,n)$ represent the mean physical distances in state $s$ between positive prompts, negative prompts, and between positive and negative prompts, respectively.", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rewards: Based on the various distances obtained from $G^{s}$ , we can set a reward function according to the optimization objectives. The reward function for the current state, $r_{s}$ , is defined based on the changes in distance from the previous state:", + "bbox": [ + 511, + 90, + 905, + 166 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} r _ {s} = - \\alpha \\left(D _ {f} ^ {s} (p, p) - D _ {f} ^ {s - 1} (p, p)\\right) \\\\ + \\alpha \\left(D _ {f} ^ {s} (p, n) - D _ {f} ^ {s - 1} (p, n)\\right) \\\\ + \\beta \\left(D _ {p} ^ {s} (p, p) - D _ {p} ^ {s - 1} (p, p)\\right) \\tag {8} \\\\ + \\beta (D _ {p} ^ {s} (n, n) - D _ {p} ^ {s - 1} (n, n)) \\\\ - \\beta (D _ {p} ^ {s} (p, n) - D _ {p} ^ {s - 1} (p, n)), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 172, + 903, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where coefficients $\\alpha$ and $\\beta$ balance the feature and physical distance, respectively. Using this reward function, if the current state's $D_{f}^{s}(p,p)$ and $D_{p}^{s}(p,n)$ increase, a penalty is applied. Conversely, if the current state's $D_{f}^{s}(p,n)$ , $D_{p}^{s}(p,p)$ , and $D_{p}^{s}(n,n)$ increase, a reward is given.", + "bbox": [ + 511, + 282, + 905, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Policy Network: We adopt the Q-network [25] as the policy network, which consists of two fully connected layers with 1024 and 64 neurons, respectively, and a single output layer with 3 neurons. This output layer predicts the Q-value of 3 actions based on the current state. The policy network is regarded as the approximate action-value function $Q_{s}$ and takes the state $s$ as its input to approximate the underlying action-value function. This approximation is achieved by repeatedly performing a gradient descent step on $(y_{s} - Q_{s})^{2}$ with respect to the policy network parameters at each step. The Q-value $y_{s}$ is defined as:", + "bbox": [ + 511, + 359, + 905, + 525 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {s} = \\left\\{ \\begin{array}{l l} r _ {s} + \\gamma \\max _ {s + 1} Q _ {s + 1}, & \\text {i f n o t s t o p} \\\\ r _ {s}, & \\text {i f s t o p} \\end{array} , \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 534, + 903, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the discount factor $\\gamma$ represents the balance between the current and future rewards.", + "bbox": [ + 511, + 583, + 903, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the training process of PPO, we initially delete all nodes and treat them as potential activatable nodes in the action environment. As the agent attempts to activate different nodes, the edges of the heterogeneous graph acquire different values. Based on the reward function, the agent receives rewards or penalties and updates the Q-values accordingly. The policy network repeatedly interacts with the segmentation network, updates the network weights according to Eq. (11), and ultimately learns effective policies.", + "bbox": [ + 511, + 613, + 905, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 760, + 718, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 777, + 730, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our experiments are carried out on a Linux server platform equipped with an NVIDIA Tesla V100. During the PPO training phase, we set the step size for each iteration to 100 and the number of training epochs to 1000. The discount factor $\\gamma$ of Q-learning is set to 0.1. We use the $\\epsilon$ -greedy strategy to balance exploration and exploitation [25]. During $\\epsilon$ -greedy training, the agent can select a random action", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4336", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with probability $\\epsilon$ and a predicted action generated by the Q-network with probability $1 - \\epsilon$ . The $\\epsilon$ parameters are set as follows: $\\epsilon_{\\mathrm{start}} = 1$ , $\\epsilon_{\\mathrm{end}} = 0.1$ , and $\\epsilon_{\\mathrm{decay}} = 0.995$ . Additionally, to ensure stability during PPO training, the probability of action $A_{m}$ is always set to be 0.5 times the probability of the other actions.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the testing phase, PPO requires no additional training for new tasks; it simply inputs initial prompts and images to output optimized prompt schemes that serve SAM. For evaluation, we use the mean Dice Similarity Coefficient (mDSC) and the mean Hausdorff Distance (mHD) as metrics to assess the segmentation performance.", + "bbox": [ + 89, + 181, + 480, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Dataset Description", + "text_level": 1, + "bbox": [ + 89, + 282, + 277, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure that PPO remains unseen to in-domain datasets for downstream tasks, we trained PPO on the GBMSD dataset [21], generating initial prompts through feature matching and constructing a heterogeneous graph as the environment for DRL learning. During testing, we evaluated the model on natural image domains using the FSS-1000 [20] and VOC2012 [6] datasets. The FSS-1000 dataset comprises mask-annotated images from 1,000 classes, with 10 images per class, while the VOC2012 dataset is a widely-used benchmark for semantic segmentation tasks, containing 20 object classes and a total of 4,396 images. For medical image domains, we used the ISIC2018 [33] and Kvasir [10]. The ISIC2018 is a large-scale collection of dermatoscopic images focused on lesion segmentation, consisting of 2,594 images. The Kvasir, on the other hand, is an endoscopic dataset aimed at colorectal polyp segmentation, containing 1,000 gastroenterology images of colorectal polyps along with their corresponding segmentation masks.", + "bbox": [ + 89, + 299, + 482, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We generated initial prompts based on two mainstream point prompt-generation paradigms. The first paradigm is the feature matching (FM) approach [21, 23], which requires no additional training for new tasks. In this approach, only one annotated reference image is selected for each segmentation task, and this image is used to generate the initial prompt for testing images. The second paradigm is the coarse segmentation sampling (CS) approach, which involves few-shot training for new tasks. For each task, we selected 20 annotated images, trained a U-Net [31] with 200 epochs, and used the resulting coarse segmentation masks to sample initial prompts for the segmentation tasks.", + "bbox": [ + 89, + 571, + 482, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 762, + 243, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct an ablation study on the initial prompts generated by FM to evaluate the segmentation performance of SAM with PPO under varying parameter settings. Quantitative results are presented in Table 1. Initially, when SAM applies unoptimized prompts $(\\alpha = 0, \\beta = 0)$ , segmentation outcomes are suboptimal across all datasets. In contrast, applying PPO with optimized settings significantly enhances prompt quality, leading to notable performance gains for", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SAM across various tasks.", + "bbox": [ + 513, + 90, + 691, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When we examine optimization within the feature space $(\\alpha = 1, \\beta = 0)$ and physical space $(\\alpha = 0, \\beta = 1)$ separately, both settings improve SAM's performance; however, feature space optimization proves to be more effective, yielding a marked improvement over the initial prompts. The better results are achieved when both feature and physical spaces are optimized concurrently $(\\alpha = 1, \\beta = 1)$ .", + "bbox": [ + 511, + 106, + 903, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interestingly, further increasing the weight on the physical space $(\\beta = 5)$ reduces segmentation performance, likely due to an overemphasis on physical distributions at the expense of essential feature space information. In contrast, placing a higher weight on the feature space $(\\alpha = 5, \\beta = 1)$ yields the best segmentation results. These experiments demonstrate that the accuracy of point prompts is more crucial than sparsity, though sparsity remains beneficial.", + "bbox": [ + 511, + 212, + 903, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 3, we compare SAM's qualitative segmentation results using initial prompts versus PPO-optimized prompts under the best parameter settings. This figure includes images from both natural and medical datasets. The initial prompts contain inaccuracies and excessive redundancy, impairing SAM's segmentation accuracy. After PPO refinement, the prompts become more precise and sparse, allowing the segmentation results to closely align with the ground truth, demonstrating PPO's effectiveness in enhancing SAM's adaptability across diverse segmentation tasks.", + "bbox": [ + 511, + 333, + 903, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison with One-Shot Methods", + "text_level": 1, + "bbox": [ + 511, + 491, + 831, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our goal is to enable SAM to adapt to segmentation tasks across diverse scenarios with minimal cost. To this end, we compared recently proposed methods with our PPO-optimized FM prompt generation method, both following the one-shot paradigm, across multiple datasets. As shown in Table 2, FM-PPO achieves competitive performance in various natural image segmentation tasks without training on in-domain datasets, performing comparably to models trained on in-domain data and surpassing other SAM-based prompt generation methods. Additionally, FM-PPO significantly outperforms other segmentation methods on medical datasets without requiring additional training. This demonstrates that PPO-optimized visual prompt generation retains its training-free task transfer capability while ensuring robust segmentation performance for downstream SAM through optimized prompts.", + "bbox": [ + 511, + 508, + 903, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, Fig. 4 illustrates the qualitative results of training-free methods on natural and medical datasets, providing visual evidence of FM-PPO's effectiveness. The experimental results show that FM-PPO performs well even in small target segmentation by reducing erroneous prompt points. This success is attributed to the homogeneous feature expectation within the segmentation target and the heterogeneous feature expectation between the target and background during agent training. For medical image segmentation, incorrect and redundant prompts due to domain", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4337", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9566aba06ffd8ad75137ccb13a603ca08241cd4db6523b28518f428e0e66c207.jpg", + "table_caption": [ + "Table 1. Ablation study on different spatial weights during PPO training. The results with $\\alpha$ and $\\beta$ set to 0 represent the initial prompts." + ], + "table_footnote": [], + "table_body": "
Feature SpacePhysical SpaceFSS-1000VOC2012ISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
α=0β=072.19.934.712.148.59.927.213.2
α=1β=081.38.559.410.561.79.256.610.3
α=0β=173.59.938.411.751.99.930.513.1
α=1β=186.67.863.710.369.28.561.39.6
α=1β=582.48.956.910.264.99.058.89.7
α=5β=191.97.473.79.975.48.170.39.0
", + "bbox": [ + 96, + 114, + 898, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab9ce231d488bffebf21739d6fccf11704b565edab3daddb5888770939ade6aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 262, + 194, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85d9168ecc72029c87886af298d28f91a4e7223ee2337c7dee4ab556352c41b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 328, + 192, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f30d2aad89f5edc11dd910da859c4a4e06a7c71f211d71f07cfd1a1fe08e9b57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 262, + 277, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/258fa199c20e4d80d2c7d4d45fcc677d8dc67956cbb4e945de63ad990d1f9628.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 327, + 277, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a00f1262ac9a725495e538c9966d0c2b6f42d338baebfc9f7f91444cf81ddb71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 262, + 362, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc09af238f3772a7993b74b55e1f360891699bfe7a968838ff1229d63b4ab32f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 327, + 362, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4887450f8d53c8bf4836fa939a2a8834945384d64d0ff5cfb3a42850f5330c1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 262, + 454, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb489915c2d5abc6c6dbfe71d02bb9bad4b98623a4da22a9054b65a962ba7e6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 327, + 455, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/783830a089319459447f16084dc383527c09572097d6fb7422f881609a84d929.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 262, + 537, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f8370d466058de0fc32cf764618132c4b0e4e1184c4e090be3933866ed5ab6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 327, + 539, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16ae01fc8d896bfd470dca9aece9e6e5984e664b6f6eea7e9f7e1fc360a69105.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 262, + 624, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/294ae35a955193b3bb59c0c138d9f041f556a1b74f8f90abe72ee47390d364fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 327, + 624, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b89e89a823ea05b18fb09be4ac9466d3797e4bd1ecc0520c502a238a067d693.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 262, + 715, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd94debb756685920b8cf3f6eb95f239b430adcd20b700e3bf62846219bb4ed2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 327, + 715, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecaf7c120d1d14825a706be883fa310487ffb44643bb8b4c19d386ac2566cd46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 262, + 799, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/854656f251b557acff769bb4fdb183e71f06c5b9ca184b2fadea53520d4f623e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 327, + 799, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a1660150a204dd8dc2d8f84448f5ec52ed10aa4ec2a85808ed9b65d3313ac66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 262, + 883, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49b50c368b74b9588dbd562dc1e8ef8e28c37d7336cec38fec75f906bec496de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 327, + 883, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46dd9f4bbd0017298a4ed3f58084383a2acc08da40c027bfcb748ec0a62d76fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 396, + 194, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3378bfc7a9eb26f6be7646a7c6bf54f9fc0991ecbfd44c8904075d70a34c4597.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 112, + 462, + 192, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0544123be9cadc284a36bd1f5a0d89692a00fa286ade5b47d13dea1442e53960.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 396, + 277, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f3df79bcf9ab519f133526f87889aa5234f2551567c444ab17bbb59193b9275.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 197, + 462, + 277, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45dc42db52e8633d2acde500de003c670d3503e99582dff8c3ef5cb437869db0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 396, + 362, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5d4b2feda5a0480ee9c2a212aa3b1900881ef43fe29cb58130b51678234cae0.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 281, + 462, + 362, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8de72bc3507adaa94da68be45a31b5bf160b52fcd241e49b3bd3dd0764d2d0f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 396, + 454, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82bbda4bbf70bdb19d65371c5cb84cd668fefb7536dc80496a9f5a4e04bdbf54.jpg", + "image_caption": [ + "(a)", + "Figure 3. Qualitative segmentation results of initial prompts and optimized prompts. (a) ground truth, (b) segmentation results with initial prompts, (c) segmentation results with optimized prompts." + ], + "image_footnote": [], + "bbox": [ + 372, + 462, + 454, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fcb533011bf89b2a49f3d4cfbca9683139cc08d4d449c2cc6a52c77539945953.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 396, + 539, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/522af46bcc19a49f861cc8cea998d1ee0f306ef5583a4c8feac224c5be222e9b.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 457, + 462, + 539, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e71723dc835a06b96f8d8b6acb57442d90ced2579ce69640982f046ee025ff0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 396, + 624, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ea56c0c85a714a81c308fbea5ad42c8103b7f6a41b1714206f07b9f98f2c1da.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 540, + 462, + 624, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/407be0e27b39c19f3035bca69f04a109aa7cc3c060c939d09dcabda6e5cbeb47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 396, + 715, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5da259c4633ff8b3e1bdd431ffc9b74b11a67763505fa6892f99607d9675cc54.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 633, + 462, + 715, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/199597e39389a1207f6d2a2b79e4c2fe0271593d0f0c31ef9c0b7d441dfcfd5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 396, + 799, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5399ada3767f2d52b9df0f0da8d22eaee77a99ad95bdc609b8bbddacca773d45.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 717, + 462, + 799, + 523 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87332dd9cd9b791de2c3c48915e8fc4ac252d5f963e8f40a5edb761fa6727aa5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 396, + 883, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb7325b64ad29577351f3d40cb68d6a1e52943bd05457115eda3ac5fa2137cd5.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 802, + 462, + 883, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "differences can lead to poorer segmentation outcomes for SAM. Through prompt optimization, FM-PPO effectively bridges the domain gap between medical and natural images, enhancing segmentation performance across both domains. Additionally, FM-PPO promotes a relatively uniform physical distribution of point prompts and close proximity between positive and negative prompts. In contrast, Matcher lacks negative prompts, which leads to excessive background segmentation on medical images with domain shifts, particularly when segmenting smaller targets, resulting in suboptimal performance. We further elaborate on the necessity of negative prompts in the supplementary materials. As a result, our method excels at boundary segmentation, especially in complex medical images where other methods struggle with background-foreground adhesion. FM-PPO can optimize the initial prompts generated by FM without additional training, achieving efficient segmentation across various scenarios.", + "bbox": [ + 89, + 609, + 485, + 880 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Different Initial Prompts Scheme Optimization", + "text_level": 1, + "bbox": [ + 511, + 608, + 906, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further validate PPO's effectiveness in optimizing prompts, we applied it to various initial prompt generation schemes and evaluated the downstream SAM segmentation. We selected medical datasets with domain gap, as prompts have a greater impact on SAM's segmentation. Table 3 presents the experimental results, with corresponding qualitative results in the supplementary materials. FM and CS serve as two baseline prompt generation methods. As shown, PPO significantly enhances segmentation performance on both datasets by optimizing the prompts. FM generates a richer prompt set based on patch matching, allowing PPO to achieve stronger optimization. In contrast, CS samples within coarse segmentation regions, leading to a concentrated prompt distribution that may overlook some target areas, thus limiting PPO's optimization potential.", + "bbox": [ + 509, + 627, + 906, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, Matcher [23], a recent approach in visual prompt engineering, utilizes the FM paradigm to achieve task-agnostic segmentation. To demonstrate PPO's", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4338", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/987b1d16020689018910cceae0763685d0168bab4cc3d51ea771d3cf8a9dc8e0.jpg", + "table_caption": [ + "Table 2. Comparison results with one-shot methods in natural and medical image datasets. Gray metrics indicates the model is already trained by in-domain datasets. ${}^{ \\dagger }$ indicates the training-free method for segmentation tasks." + ], + "table_footnote": [], + "table_body": "
MethodsFSS-1000VOC2012ISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
HSNet [24]91.18.244.511.457.411.329.716.5
VAT [7]93.67.548.211.449.812.926.416.8
SegGPT† [39]93.77.178.27.861.29.544.114.4
PerSAM† [43]81.18.653.610.937.616.145.816.3
Matcher† [23]91.67.669.011.173.79.746.216.0
FM-PPO† (Ours)91.97.473.79.975.48.170.39.0
", + "bbox": [ + 96, + 128, + 901, + 265 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78bd95ee017b830780af4a2b7564d4dece78e3aa48abfa0afef3b6877fc7a834.jpg", + "image_caption": [ + "Figure 4. Qualitative segmentation results of different one-shot methods in natural and medical images." + ], + "image_footnote": [], + "bbox": [ + 104, + 289, + 470, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "plug-and-play capability, we applied PPO to optimize the prompts generated by Matcher and then fed them back into Matcher for segmentation. Table 3 shows that PPO-optimized prompts significantly improve segmentation performance, particularly on the Kvasir, where the DSC increased by $32.6\\%$ for smaller targets. This demonstrates that our method effectively boosts SAM's segmentation performance by optimizing various types of initial prompts.", + "bbox": [ + 89, + 652, + 482, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 790, + 210, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce PPO, designed to enhance the distribution of point prompts, thereby improving the downstream automatic segmentation performance of SAM. We first formalize the distribution of point prompts as a dual-space heterogeneous graph and define the optimization direction. Next, we train an agent within this graph environ-", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1b8aec8b125a1eee0b763a44567b840d7a013d2584f5afa7658818e4aa55aad7.jpg", + "table_caption": [ + "Table 3. Optimization performance of different initial prompts." + ], + "table_footnote": [], + "table_body": "
Prompt SchemeISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
FM48.59.927.213.2
FM-PPO75.4 (↑21.9)8.1 (↓1.8)70.3 (↑43.1)9.0 (↓4.2)
CS57.711.969.110.2
CS-PPO68.6 (↑10.9)10.8 (↓1.1)79.8 (↑10.7)8.9 (↓1.3)
Matcher [23]73.79.746.216.0
Matcher-PPO85.9 (↑12.2)9.0 (↓0.7)77.6 (↑32.6)9.3 (↓6.7)
", + "bbox": [ + 519, + 314, + 897, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ment to automatically activate effective prompt schemes, which are then input into SAM to obtain segmentation results. Our experiments on natural and medical image datasets demonstrate that PPO significantly improves segmentation performance compared to various initial point prompt schemes. In conclusion, PPO offers a promising solution for optimizing any initial point prompts, enhancing the performance of diverse downstream segmentation tasks. Limitations. Although PPO demonstrates strong performance in optimizing incorrect and redundant point prompts, its effectiveness is limited when the initial prompts lack coverage of the target. For example, in the CS approach, if the pseudo-labels do not include the target, prompts are not sampled over the target region. Future research will focus on designing more comprehensive point prompt generation methods to build a more robust prompt optimizer.", + "bbox": [ + 511, + 470, + 906, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 513, + 741, + 665, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported by the National Natural Science Foundation of China (Grant No. 61901292 and Grant No. 62201331), the Fundamental Research Program of Shanxi Province (Grant No. 201901D211080, Grant No. 202303021211082, Grant No. 202403021222276 and Grant No. 202403021212262), the Shanxi Province Graduate Education Innovation Project (Grant No. 2022Y196). We also express our gratitude to the all reviewers and area chairs for their constructive feedback.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4339", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 3", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1", + "[3] Zhen Chen, Qing Xu, Xinyu Liu, and Yixuan Yuan. Un-sam: Universal prompt-free segmentation for generalized nuclei images. arXiv preprint arXiv:2402.16663, 2024. 3", + "[4] Haixing Dai, Chong Ma, Zhengliang Liu, Yiwei Li, Peng Shu, Xiaozheng Wei, Lin Zhao, Zihao Wu, Dajiang Zhu, and Wei Liu. Samaug: Point prompt augmentation for segment anything model. arXiv preprint arXiv:2307.01187, 2023. 3", + "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2", + "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6", + "[7] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 8", + "[8] Duojun Huang, Xinyu Xiong, Jie Ma, Jichang Li, Zequn Jie, Lin Ma, and Guanbin Li. Alignsam: Aligning segment anything model to open context via reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3205-3215, 2024. 2", + "[9] Jiaxing Huang, Kai Jiang, Jingyi Zhang, Han Qiu, Lewei Lu, Shijian Lu, and Eric Xing. Learning to prompt segment anything models. arXiv preprint arXiv:2401.04651, 2024. 3", + "[10] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, pages 451–462. Springer, 2020. 6", + "[11] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen" + ], + "bbox": [ + 93, + 107, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1", + "[12] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3", + "[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3", + "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2", + "[15] Hyeokjun Kweon and Kuk-Jin Yoon. From sam to cams: Exploring segment anything model for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19499-19509, 2024. 2", + "[16] Yiming Lei, Jingqi Li, Zilong Li, Yuan Cao, and Hongming Shan. Prompt learning in computer vision: a survey. Frontiers of Information Technology & Electronic Engineering, 25(1):42-63, 2024. 3", + "[17] Chengyin Li, Prashant Khanduri, Yao Qiang, Rafi Ibn Sultan, Indrin Chetty, and Dongxiao Zhu. Autoprompting sam for mobile friendly 3d medical image segmentation. arXiv preprint arXiv:2308.14936, 2023. 3", + "[18] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2", + "[20] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 6", + "[21] Xueyu Liu, Guangze Shi, Rui Wang, Yexin Lai, Jianan Zhang, Lele Sun, Quan Yang, Yongfei Wu, Ming Li, Weixia Han, et al. Feature-prompting gbmseg:" + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4340", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "One-shot reference guided training-free prompt engineering for glomerular basement membrane segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 276-285. Springer, 2024. 2, 3, 6", + "[22] Xueyu Liu, Rui Wang, Yexin Lai, Yongfei Wu, Hangbei Cheng, Yuanyue Lu, Jianan Zhang, Ning Hao, Chenglong Ban, Yanru Wang, Shuqin Tang, Yuxuan Yang, Ming Li, Xiaoshuang Zhou, and Wen Zheng. Msmtseg: Multi-stained multi-tissue segmentation of kidney histology images via generative self-supervised meta-learning framework. IEEE Journal of Biomedical and Health Informatics, pages 1-12, 2024. 1", + "[23] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 2, 3, 6, 7, 8", + "[24] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 8", + "[25] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 5", + "[26] Saiyang Na, Yuzhi Guo, Feng Jiang, Hehuan Ma, and Junzhou Huang. Segment any cell: A sam-based autoprompting fine-tuning framework for nuclei segmentation. arXiv preprint arXiv:2401.13220, 2024. 3", + "[27] Maxime Aquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3", + "[28] Xiaoyu Qiu, Hao Feng, Yuechen Wang, Wengang Zhou, and Houqiang Li. Progressive multi-modal conditional prompt tuning. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 46-54, 2024. 3", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2", + "[30] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma," + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1", + "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6", + "[32] Tal Shaharabany, Aviad Dahan, Raja Giryes, and Lior Wolf. Autosam: Adapting sam to medical images by overloading the prompt encoder. arXiv preprint arXiv:2306.06370, 2023. 3", + "[33] Junji Shiraishi, Shigehiko Katsuragawa, Junpei Ikezoe, Tsuneo Matsumoto, Takeshi Kobayashi, Kenichi Komatsu, Mitate Matsui, Hiroshi Fujita, Yoshie Kodera, and Kunio Doi. Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American journal of roentgenology, 174(1):71-74, 2000. 6", + "[34] Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. Vrp-sam: Sam with visual reference prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23565-23574, 2024. 2, 3", + "[35] Chenglong Wang, Dexuan Li, Sucheng Wang, Chengxiu Zhang, Yida Wang, Yun Liu, and Guang Yang. SAMMed: A medical image annotation framework based on large vision model. arXiv preprint arXiv:2307.05617, 2023. 3", + "[36] Jiaqi Wang, Zhengliang Liu, Lin Zhao, Zihao Wu, Chong Ma, Sigang Yu, Haixing Dai, Qiushi Yang, Yiheng Liu, Songyao Zhang, et al. Review of large vision models and visual prompt engineering. MetaRadiology, page 100047, 2023. 3", + "[37] Rui Wang, Zhouwang Yang, and Yanzhi Song. Osamfundus: A training-free, one-shot segmentation framework for optic disc and cup in fundus images. Biomedical Signal Processing and Control, 100:107069, 2025. 2, 3", + "[38] Xu Wang, Sen Wang, Xingxing Liang, Dawei Zhao, Jincai Huang, Xin Xu, Bin Dai, and Qiguang Miao. Deep reinforcement learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 35 (4):5064-5078, 2022. 4", + "[39] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1130-1140, 2023. 8" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4341", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Bin Xie, Hao Tang, Bin Duan, Dawen Cai, and Yan Yan. Masksam: Towards auto-prompt sam with mask classification for medical image segmentation. arXiv preprint arXiv:2403.14103, 2024. 3", + "[41] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 3", + "[42] Chaoning Zhang, Fachrina Dewi Puspitasari, Sheng Zheng, Chenghao Li, Yu Qiao, Taegoo Kang, Xinru Shan, Chenshuang Zhang, Caiyan Qin, Francois Rameau, et al. A survey on segment anything model (sam): Vision foundation model meets prompt engineering. arXiv preprint arXiv:2306.06211, 2023. 3", + "[43] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 2, 3, 8", + "[44] Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, et al. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419, 2023. 2", + "[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3" + ], + "bbox": [ + 89, + 90, + 482, + 534 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4342", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/fb8f3fbf-9e73-432d-9130-bbe1d55d00ee_model.json b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/fb8f3fbf-9e73-432d-9130-bbe1d55d00ee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fb61f46c23daf7ff366cc077cdab1287c20ee832 --- /dev/null +++ b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/fb8f3fbf-9e73-432d-9130-bbe1d55d00ee_model.json @@ -0,0 +1,2708 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.155 + ], + "angle": 0, + "content": "Plug-and-Play PPO: An Adaptive Point Prompt Optimizer Making SAM Greater" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.18, + 0.816, + 0.217 + ], + "angle": 0, + "content": "Xueyu Liu\\(^{1\\dagger}\\), Rui Wang\\(^{2\\dagger}\\), Yexin Lai\\(^{1}\\), Guangze Shi\\(^{1}\\), Feixue Shao\\(^{1}\\), Fang Hao\\(^{1}\\), Jianan Zhang\\(^{1}\\), Jia Shen\\(^{1}\\), Yongfei Wu\\(^{1*}\\), Wen Zheng\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.217, + 0.639, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Taiyuan University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.235, + 0.692, + 0.253 + ], + "angle": 0, + "content": "\\(^{2}\\)University of Science and Technology of China" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.486, + 0.804 + ], + "angle": 0, + "content": "Powered by extensive curated training data, the Segment Anything Model (SAM) demonstrates impressive generalization capabilities in open-world scenarios, effectively guided by user-provided prompts. However, the class-agnostic characteristic of SAM renders its segmentation accuracy highly dependent on prompt quality. In this paper, we propose a novel plug-and-play dual-space Point Prompt Optimizer (PPO) designed to enhance prompt distribution through deep reinforcement learning (DRL)-based heterogeneous graph optimization. PPO optimizes initial prompts for any task without requiring additional training, thereby improving SAM's downstream segmentation performance. Specifically, PPO constructs a dual-space heterogeneous graph, leveraging the robust feature-matching capabilities of a foundational pre-trained model to create internal feature and physical distance matrices. A DRL policy network iteratively refines the distribution of prompt points, optimizing segmentation predictions. We conducted experiments on four public datasets. The ablation study explores the necessity and balance of optimizing prompts in both feature and physical spaces. The comparative study shows that PPO enables SAM to surpass recent one-shot methods. Additionally, experiments with different initial prompts demonstrate PPO's generality across prompts generated by various methods. In conclusion, PPO redefines the prompt optimization problem as a heterogeneous graph optimization task, using DRL to construct an effective, plug-and-play prompt optimizer. This approach holds potential for broader applications across diverse segmentation tasks and provides a promising solution for point prompt optimization. The source code and demo are available at https://github.com/XueyuLiu/PPO." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.828, + 0.222, + 0.843 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.846, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Segmentation, as a fundamental task, plays a pivotal role in visual understanding systems. The prevailing train-" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.284, + 0.895, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.515, + 0.908, + 0.586 + ], + "angle": 0, + "content": "Figure 1. Ineffective prompts can diminish SAM's performance. Incorrect prompts may cause the segmentation mask to capture irrelevant regions, while excessive prompts for adjacent areas can lead to neglect of other parts of the object. Therefore, prompt optimization is essential to improve SAM's performance." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.613, + 0.907, + 0.704 + ], + "angle": 0, + "content": "ing paradigm for segmentation relies on data-driven approaches, where large datasets are used to train models. While this method has achieved remarkable success, it faces several challenges, including the computational resources to train models for new tasks and the risk of domain-specific optimizations that can limit generalizability [18, 22]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recent breakthroughs in Pretrained Foundation Models (PFMs) in computer vision (CV) have demonstrated impressive capabilities in domain-invariant segmentation across diverse scenarios [11, 29]. Vision Foundation Models (VFMs), such as DINOv2 [2, 27], acquire generalized visual features by capturing intricate information at both the image and pixel levels using only raw image data. This approach enables these models to learn generic visual representations, ensuring robust zero-shot transferability for various downstream tasks. Meanwhile, the Segment Anything Model (SAM) [14, 30], trained on large-scale, high-quality datasets, has shown a strong ability to generalize across tasks and data distributions that are not encountered during" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.333, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author. †Co-first author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4332" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "training. However, SAM relies heavily on manual prompts, such as bounding boxes and points, to identify objects of interest, as it is pretrained in a class-agnostic manner. Consequently, achieving task-specific adaptation of SAM is crucial to bridge the gap between its class-agnostic predictions and the specific needs of industrial applications." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.483, + 0.456 + ], + "angle": 0, + "content": "Some works attempt to train prompt encoders that enables SAM to adapt to downstream segmentation tasks [8, 15, 34, 43]. However, these task-specific prompt encoders require additional training for new segmentation scenarios. Recently, other efforts have focused on automatically generating geometric visual prompts, such as point prompts, to enhance SAM's segmentation performance [21, 23, 37]. In such methods, the quality of prompt schemes remains a critical factor influencing downstream segmentation performance. As illustrated in Fig. 1, incorrect prompts can result in erroneous segmentation, while excessive prompts in specific areas may cause SAM to focus disproportionately on those regions, neglecting other parts of the target. Current automatic prompt engineering methods often lack optimization in prompt distribution, leading to lower-quality prompts and segmentation performance. Therefore, there is an urgent need for methods that can refine prompts to effectively enhance SAM's performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.626 + ], + "angle": 0, + "content": "To address this, effective point prompts for SAM should exhibit certain characteristics in both feature and physical spaces. In the feature space, regions containing positive prompts should be homogeneous to ensure consistency and clarity within the segmentation target, while regions between positive and negative prompts should be heterogeneous, distinguishing the target from the background. In the physical space, positive and negative prompts should be sufficiently spaced apart to mitigate the effects of excessive prompts and prevent over-segmentation, while also being close enough to boundaries to ensure precise delineation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Building upon the aforementioned inspiration and committed to addressing the challenges mentioned above, we propose a novel Plug-and-Play Point Prompt Optimizer (PPO) designed to enhance SAM's segmentation performance by automatically optimizing initial point prompt schemes. Leveraging DINOv2's powerful feature association capabilities [27], we construct an internal feature and physical distance matrix within each image to create a heterogeneous graph environment. To optimize prompt distribution in dual spaces, we employ Deep Reinforcement Learning (DRL), a framework that trains agents to maximize cumulative rewards through interaction and feedback mechanisms. By designing specific actions and reward functions, PPO optimizes prompt point distribution under the guidance of DRL and inputs them into SAM to obtain refined segmentation results. We validated our approach through experiments on two natural and two medical image datasets. Ablation studies demonstrate that PPO" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "optimized prompt distributions significantly enhance segmentation performance. Comparative experiments show that PPO outperforms recent methods that have not been trained on in-domain datasets. Additionally, experiments with different initial prompts confirm PPO's generality across prompts generated by various methods. The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.2, + 0.905, + 0.275 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to define an optimization direction for point prompts in both feature and physical spaces. Additionally, we are the first to use DRL to optimize point prompts, achieving prompt optimization without additional training for new tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.276, + 0.905, + 0.321 + ], + "angle": 0, + "content": "- PPO operates in a plug-and-play manner, allowing for the optimization of any initial geometric point prompts to improve SAM's segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.321, + 0.905, + 0.366 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the effectiveness of PPO in optimizing point prompts across various initial prompt schemes." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.2, + 0.905, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.388, + 0.655, + 0.403 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.786, + 0.423 + ], + "angle": 0, + "content": "2.1. Pretrained Foundation Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.609 + ], + "angle": 0, + "content": "In natural language processing (NLP), PFMs, such as BERT [5], ChatGPT, and GPT-4, have become foundational tools for a wide range of downstream tasks across diverse data modalities. By leveraging extensive text data, PFMs provide effective parameter initialization for various NLP applications [44]. The success of PFMs in NLP has inspired researchers to explore their application in CV. Visual language models like CLIP [29] and GLIP [19], trained on large-scale datasets of millions of image-text pairs from the web, exhibit strong zero-shot generalization across multiple vision tasks, offering representations with robust generalization and transfer capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Building on this trend, Meta AI introduced the SAM, a notable advancement in applying VFMs. SAM allows users to specify image segmentation tasks through prompts and demonstrates impressive zero-shot learning, enabling it to generalize effectively across diverse tasks without task-specific training [14]. However, SAM's segmentation performance heavily relies on the quality of input prompts. To address this, we propose an adaptive point prompt optimizer that automatically refines prompts to enhance SAM's segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In another significant development, DINOv2 [27] leverages discriminative self-supervised learning at both image and patch levels, showing exceptional patch-matching capabilities that allow it to capture semantic similarities across images of the same target. By utilizing these patch-matching capabilities, we can extract feature distances between patches within an image, providing a theoretical foundation for constructing heterogeneous graphs to achieve effective prompt optimization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4333" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.37, + 0.108 + ], + "angle": 0, + "content": "2.2. Automatic Prompt Engineering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.484, + 0.381 + ], + "angle": 0, + "content": "In the field of NLP, extensive researches have been conducted on designing automatic prompt engineering, with prompt tuning methods achieving significant advances in enhancing PFMs' performance through optimized language prompts [1, 12, 13, 16, 28, 41, 45, 45]. Simultaneously, efforts have been made to explore automatic prompt engineering in CV. On the one hand, some recent works have attempted to train prompt encoders that enable SAM to adapt to downstream segmentation tasks [9, 32, 34, 40]. For example, Sun et al. proposed VRP-SAM, an extension of SAM that encodes various granularities of visual references directly into prompt embeddings for mask generation [34]. These approaches encode SAM's prompts into prompt embeddings via a prompt encoder, which are then fed into SAM's mask decoder to generate target masks. However, these methods are often task-sensitive, requiring training different prompt encoders for different tasks, thus limiting rapid task transfer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.484, + 0.61 + ], + "angle": 0, + "content": "On the other hand, some works attempt to generate geometric prompts to achieve automatic prompt engineering for SAM, generally following two main paradigms. The first paradigm uses a small amount of annotated data to train a segmentation model, then samples prompts from pseudolabels generated by this model [3, 26, 36, 42, 43]. For instance, Li et al. randomly selected prompt points from masks generated by coarse segmentation, which were subsequently input into SAM after fine-tuning to obtain segmentation results [17]. Wang et al. utilized a novel prototype learning framework to generate pseudo-labels and extract prompts, which were then incorporated into SAM for segmentation [35]. Dai et al. developed various methods to expand prompt points, exploring the potential of a single prompt point for enhanced segmentation performance [4]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.484, + 0.884 + ], + "angle": 0, + "content": "However, this first paradigm still requires a small amount of data to train the prompt generator, also limiting task transferability. To improve transferability, the second paradigm generates initial prompts by leveraging feature similarity between reference and target images without additional training [21, 23, 37, 43]. For example, Zhang et al. calculated feature similarity between pixels in the target and reference images to create a point prompt scheme using only a single reference image [43]. Liu et al. introduced Matcher, an instance-level and bidirectional matching strategy for prompt point generation [23], and Liu et al. proposed a series of prompt optimization strategies for the automatic segmentation of basement membranes in electron microscopy images [21]. However, the quality of prompts generated by these methods is often limited, as they do not sufficiently consider the relationships between positive and negative prompts, resulting in redundant and erroneous prompts that negatively affect segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In our work, PPO addresses this limitation by con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.184 + ], + "angle": 0, + "content": "structuring a heterogeneous graph based on dual-space distances between initial prompts within each image. This approach optimizes any automatically generated geometric point prompts without additional training for new tasks, enhancing SAM's segmentation performance while preserving task transferability by ensuring high prompt quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.197, + 0.65, + 0.214 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.216, + 0.907, + 0.458 + ], + "angle": 0, + "content": "This section introduces our plug-and-play PPO, which is based on a heterogeneous graph structure designed to enhance the segmentation performance of SAM by optimizing initial point prompts. An overview of PPO is illustrated in Fig. 2. Specifically, we first compute the feature distance matrix and the physical distance matrix for each image. Using these matrices, we construct a heterogeneous graph where the edge attributes are defined by the feature and physical distances based on the initial prompt indices. During the training phase, we use this heterogeneous graph as the environment to design an action list and reward function to train the optimization agent. In the inference phase, the trained agent activates the appropriate point prompts as a final prompt scheme, which is then input into SAM to obtain the final segmentation result. In the following subsections, we will describe the main components in detail." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.862, + 0.499 + ], + "angle": 0, + "content": "3.1. Dual-Space Distance Matrix Calculation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.592 + ], + "angle": 0, + "content": "The core objective of our method is to optimize positive and negative point prompts in both feature and distance spaces. To achieve this, we construct a heterogeneous graph comprising two homogeneous graphs that represent the feature and distance spaces. Edge values are determined by creating two patch-level distance matrices for each image \\( X \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.906, + 0.669 + ], + "angle": 0, + "content": "First, we divide \\(X\\) into patches \\(x\\) using a sliding window approach, where \\(x_{i} = x_{1},x_{2},\\ldots ,x_{n}\\). We then use the image encoder from DINOv2 [27] to extract patch-level features for \\(x\\), represented by \\(f\\), where \\(f_{i} = f_{1},f_{2},\\dots ,f_{n}\\). Next, we calculate the feature matrix \\(M_{f}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.681, + 0.907, + 0.699 + ], + "angle": 0, + "content": "\\[\nM _ {f} (i, j) = \\left\\| f _ {i} - f _ {j} \\right\\|, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.709, + 0.906, + 0.772 + ], + "angle": 0, + "content": "where \\( M_{f}(i,j) \\) represents the Euclidean distance \\( D_{f} \\) between the \\( i \\)-th patch feature \\( f_{i} \\) and the \\( j \\)-th patch feature \\( f_{j} \\). Similarly, we calculate the physical distance matrix \\( M_{p} \\) using the center points of the patches \\( x \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.784, + 0.907, + 0.801 + ], + "angle": 0, + "content": "\\[\nM _ {p} (i, j) = \\left\\| x _ {i} - x _ {j} \\right\\|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.906, + 0.844 + ], + "angle": 0, + "content": "where \\( M_{p}(i,j) \\) represents the Euclidean distance \\( D_{p} \\) between the \\( i \\)-th patch center \\( x_{i} \\) and the \\( j \\)-th patch center \\( x_{j} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.853, + 0.78, + 0.869 + ], + "angle": 0, + "content": "3.2. Graph Environment Creation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Inspired by coarse segmentation sampling [3, 26] and feature matching techniques [21, 23], we can obtain initial" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4334" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.45, + 0.908, + 0.494 + ], + "angle": 0, + "content": "Figure 2. The PPO workflow is designed to enhance SAM's segmentation performance by optimizing initial point prompts through two key components: graph-based environment creation and agent training for prompt optimization. After optimized, the prompts are fed into SAM to obtain segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.519, + 0.482, + 0.625 + ], + "angle": 0, + "content": "point prompt indices for optimization. Specifically, we define a heterogeneous graph \\( G = (V,E,\\phi ,\\psi) \\), where \\( V \\) is the set of nodes and \\( E \\) is the set of edges. The node type set \\( \\mathcal{A} = \\{\\text{positive nodes } V_p, \\text{negative nodes } V_n\\} \\) is represented by the mapping function \\( \\phi (v): V \\to \\mathcal{A} \\), and the edge type set \\( \\mathcal{R} = \\{\\text{feature edges } E_f, \\text{physical edges } E_p\\} \\) is represented by the mapping function \\( \\psi (e): E \\to \\mathcal{R} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.672 + ], + "angle": 0, + "content": "We use the heterogeneous graph \\( G \\) as the environment for DRL. By selecting different nodes, we can obtain different edge values as feedback." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.686, + 0.442, + 0.703 + ], + "angle": 0, + "content": "3.3. Agent Training for Prompt Optimization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "On the one hand, an effective point prompt scheme should ensure correctness. In an image, the features of the segmentation target are often homogeneous, while the features of the background are heterogeneous. Therefore, we aim for the feature distance between positive point prompts to be sufficiently close, and the feature distance between positive and negative point prompts to be sufficiently far. On the other hand, an effective point prompt scheme should also ensure sparsity. Over-prompting in adjacent areas can degrade segmentation performance. Thus, the physical distance between positive point prompts (and between negative point prompts) should not be too close. Additionally, to enhance boundary segmentation accuracy, the physical dis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.548 + ], + "angle": 0, + "content": "tance between positive and negative point prompts should not be too far." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.657 + ], + "angle": 0, + "content": "A DRL model can infer actions and receive rewards from the environment, reflecting the significance of these actions and updating the agent's state accordingly [38]. Using DRL, we aim to train an agent to activate appropriate point prompts to optimize the distribution. The detailed process is shown in Algorithm 1. The key components of agent training for prompt optimization will be introduced as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.869 + ], + "angle": 0, + "content": "Actions: We define three possible actions: (1) restore node \\( A_r \\), which randomly activates a node and connects the corresponding edges; (2) delete node \\( A_d \\), which randomly deletes a node and removes the corresponding edges; (3) move node \\( A_m \\), which randomly selects a direction and moves the index position of a node. Additionally, we aim to maintain a stable number of prompts. Therefore, if the number of nodes \\( |V_p| \\) or \\( |V_n| \\) in the current state is less than the minimum value, the only possible action is \\( A_r \\). Conversely, if the \\( |V_p| \\) or \\( |V_n| \\) exceeds the maximum value, the only possible action is \\( A_d \\). Only when the \\( |V_p| \\) or \\( |V_n| \\) is between the minimum and maximum values are all three actions, \\( A_r \\), \\( A_d \\), and \\( A_m \\), possible." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "States: For each action, there will be different nodes, resulting in a different heterogeneous graph \\( G^{s} \\). Based on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4335" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.324, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 PPO training process" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.109, + 0.388, + 0.123 + ], + "angle": 0, + "content": "Input: Heterogeneous graph environment \\(G\\)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.124, + 0.327, + 0.138 + ], + "angle": 0, + "content": "Parameter: Max step; Max epoch;" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.139, + 0.331, + 0.153 + ], + "angle": 0, + "content": "Max nodes \\(N_{max}\\); Min nodes \\(N_{min}\\)" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.155, + 0.216, + 0.169 + ], + "angle": 0, + "content": "Output: Q-values" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.171, + 0.372, + 0.186 + ], + "angle": 0, + "content": "1: Initializing the Q-values, Epoch, Step." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.187, + 0.322, + 0.201 + ], + "angle": 0, + "content": "2: while Epoch < Max epoch do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.203, + 0.273, + 0.216 + ], + "angle": 0, + "content": "3: Initializing the Step." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.217, + 0.313, + 0.231 + ], + "angle": 0, + "content": "4: while Step \\(< \\) Max step do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.233, + 0.356, + 0.246 + ], + "angle": 0, + "content": "5: Emptying all nodes and edges." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.248, + 0.273, + 0.262 + ], + "angle": 0, + "content": "6: Emptying reward." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.263, + 0.482, + 0.278 + ], + "angle": 0, + "content": "7: Calculating the number of Nodes \\( N_{pos} \\) and \\( N_{neg} \\)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.278, + 0.415, + 0.293 + ], + "angle": 0, + "content": "8: if \\( N_{pos} < N_{min} \\) or \\( N_{neg} < N_{min} \\) then" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.293, + 0.289, + 0.307 + ], + "angle": 0, + "content": "9: Action list \\(= \\{A_{r}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.308, + 0.448, + 0.323 + ], + "angle": 0, + "content": "10: else if \\( N_{pos} > N_{max} \\) or \\( N_{neg} > N_{max} \\) then" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.323, + 0.29, + 0.337 + ], + "angle": 0, + "content": "11: Action list \\(= \\{A_{d}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.339, + 0.182, + 0.35 + ], + "angle": 0, + "content": "12: else" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.353, + 0.348, + 0.368 + ], + "angle": 0, + "content": "13: Action list \\( = \\{A_r, A_d, A_m\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.369, + 0.197, + 0.381 + ], + "angle": 0, + "content": "14: end if" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.383, + 0.47, + 0.397 + ], + "angle": 0, + "content": "15: Perform a random action and update the status \\(s\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.398, + 0.482, + 0.426 + ], + "angle": 0, + "content": "16: Adding the reward \\( r_s \\) based on the reward function." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.428, + 0.287, + 0.443 + ], + "angle": 0, + "content": "17: Updating Q-values." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.444, + 0.215, + 0.457 + ], + "angle": 0, + "content": "18: Step+=1." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.459, + 0.207, + 0.471 + ], + "angle": 0, + "content": "19: end while" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.474, + 0.212, + 0.487 + ], + "angle": 0, + "content": "20: Epoch \\(+ = 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.489, + 0.191, + 0.502 + ], + "angle": 0, + "content": "21: end while" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.504, + 0.225, + 0.516 + ], + "angle": 0, + "content": "22: return solution" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.171, + 0.482, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.547, + 0.446, + 0.562 + ], + "angle": 0, + "content": "\\(G^{s}\\), we can calculate the following mean edge values:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.573, + 0.483, + 0.612 + ], + "angle": 0, + "content": "\\[\nD _ {f} ^ {s} (p, p) = \\frac {1}{| V _ {p} | | V _ {p} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {p}} M _ {f} (i, j), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.625, + 0.483, + 0.663 + ], + "angle": 0, + "content": "\\[\nD _ {f} ^ {\\mathrm {s}} (p, n) = \\frac {1}{| V _ {p} | | V _ {n} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {n}} M _ {f} (i, j), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.671, + 0.483, + 0.711 + ], + "angle": 0, + "content": "\\[\nD _ {p} ^ {s} (p, p) = \\frac {1}{| V _ {p} | | V _ {p} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {p}} M _ {p} (i, j), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.719, + 0.483, + 0.757 + ], + "angle": 0, + "content": "\\[\nD _ {p} ^ {s} (n, n) = \\frac {1}{| V _ {n} | | V _ {n} |} \\sum_ {i \\in V _ {n}} \\sum_ {j \\in V _ {n}} M _ {p} (i, j), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.765, + 0.483, + 0.803 + ], + "angle": 0, + "content": "\\[\nD _ {p} ^ {s} (p, n) = \\frac {1}{| V _ {p} | | V _ {n} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {n}} M _ {p} (i, j). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In state \\(s\\), \\(D_{f}^{s}(p,p)\\) and \\(D_{f}^{s}(p,n)\\) represents the mean feature distance between positive prompts, and between positive and negative prompts. Similarly, \\(D_{p}^{s}(p,p)\\), \\(D_{p}^{s}(n,n)\\), and \\(D_{p}^{s}(p,n)\\) represent the mean physical distances in state \\(s\\) between positive prompts, negative prompts, and between positive and negative prompts, respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.167 + ], + "angle": 0, + "content": "Rewards: Based on the various distances obtained from \\( G^{s} \\), we can set a reward function according to the optimization objectives. The reward function for the current state, \\( r_{s} \\), is defined based on the changes in distance from the previous state:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.174, + 0.905, + 0.279 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} r _ {s} = - \\alpha \\left(D _ {f} ^ {s} (p, p) - D _ {f} ^ {s - 1} (p, p)\\right) \\\\ + \\alpha \\left(D _ {f} ^ {s} (p, n) - D _ {f} ^ {s - 1} (p, n)\\right) \\\\ + \\beta \\left(D _ {p} ^ {s} (p, p) - D _ {p} ^ {s - 1} (p, p)\\right) \\tag {8} \\\\ + \\beta (D _ {p} ^ {s} (n, n) - D _ {p} ^ {s - 1} (n, n)) \\\\ - \\beta (D _ {p} ^ {s} (p, n) - D _ {p} ^ {s - 1} (p, n)), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.284, + 0.906, + 0.361 + ], + "angle": 0, + "content": "where coefficients \\(\\alpha\\) and \\(\\beta\\) balance the feature and physical distance, respectively. Using this reward function, if the current state's \\(D_{f}^{s}(p,p)\\) and \\(D_{p}^{s}(p,n)\\) increase, a penalty is applied. Conversely, if the current state's \\(D_{f}^{s}(p,n)\\), \\(D_{p}^{s}(p,p)\\), and \\(D_{p}^{s}(n,n)\\) increase, a reward is given." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.906, + 0.526 + ], + "angle": 0, + "content": "Policy Network: We adopt the Q-network [25] as the policy network, which consists of two fully connected layers with 1024 and 64 neurons, respectively, and a single output layer with 3 neurons. This output layer predicts the Q-value of 3 actions based on the current state. The policy network is regarded as the approximate action-value function \\( Q_{s} \\) and takes the state \\( s \\) as its input to approximate the underlying action-value function. This approximation is achieved by repeatedly performing a gradient descent step on \\( (y_{s} - Q_{s})^{2} \\) with respect to the policy network parameters at each step. The Q-value \\( y_{s} \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.535, + 0.905, + 0.575 + ], + "angle": 0, + "content": "\\[\ny _ {s} = \\left\\{ \\begin{array}{l l} r _ {s} + \\gamma \\max _ {s + 1} Q _ {s + 1}, & \\text {i f n o t s t o p} \\\\ r _ {s}, & \\text {i f s t o p} \\end{array} , \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.904, + 0.612 + ], + "angle": 0, + "content": "where the discount factor \\(\\gamma\\) represents the balance between the current and future rewards." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.75 + ], + "angle": 0, + "content": "In the training process of PPO, we initially delete all nodes and treat them as potential activatable nodes in the action environment. As the agent attempts to activate different nodes, the edges of the heterogeneous graph acquire different values. Based on the reward function, the agent receives rewards or penalties and updates the Q-values accordingly. The policy network repeatedly interacts with the segmentation network, updates the network weights according to Eq. (11), and ultimately learns effective policies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.761, + 0.719, + 0.777 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.779, + 0.731, + 0.794 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our experiments are carried out on a Linux server platform equipped with an NVIDIA Tesla V100. During the PPO training phase, we set the step size for each iteration to 100 and the number of training epochs to 1000. The discount factor \\(\\gamma\\) of Q-learning is set to 0.1. We use the \\(\\epsilon\\)-greedy strategy to balance exploration and exploitation [25]. During \\(\\epsilon\\)-greedy training, the agent can select a random action" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4336" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "with probability \\(\\epsilon\\) and a predicted action generated by the Q-network with probability \\(1 - \\epsilon\\). The \\(\\epsilon\\) parameters are set as follows: \\(\\epsilon_{\\mathrm{start}} = 1\\), \\(\\epsilon_{\\mathrm{end}} = 0.1\\), and \\(\\epsilon_{\\mathrm{decay}} = 0.995\\). Additionally, to ensure stability during PPO training, the probability of action \\(A_{m}\\) is always set to be 0.5 times the probability of the other actions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.482, + 0.273 + ], + "angle": 0, + "content": "In the testing phase, PPO requires no additional training for new tasks; it simply inputs initial prompts and images to output optimized prompt schemes that serve SAM. For evaluation, we use the mean Dice Similarity Coefficient (mDSC) and the mean Hausdorff Distance (mHD) as metrics to assess the segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.283, + 0.279, + 0.299 + ], + "angle": 0, + "content": "4.2. Dataset Description" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.483, + 0.572 + ], + "angle": 0, + "content": "To ensure that PPO remains unseen to in-domain datasets for downstream tasks, we trained PPO on the GBMSD dataset [21], generating initial prompts through feature matching and constructing a heterogeneous graph as the environment for DRL learning. During testing, we evaluated the model on natural image domains using the FSS-1000 [20] and VOC2012 [6] datasets. The FSS-1000 dataset comprises mask-annotated images from 1,000 classes, with 10 images per class, while the VOC2012 dataset is a widely-used benchmark for semantic segmentation tasks, containing 20 object classes and a total of 4,396 images. For medical image domains, we used the ISIC2018 [33] and Kvasir [10]. The ISIC2018 is a large-scale collection of dermatoscopic images focused on lesion segmentation, consisting of 2,594 images. The Kvasir, on the other hand, is an endoscopic dataset aimed at colorectal polyp segmentation, containing 1,000 gastroenterology images of colorectal polyps along with their corresponding segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.483, + 0.753 + ], + "angle": 0, + "content": "We generated initial prompts based on two mainstream point prompt-generation paradigms. The first paradigm is the feature matching (FM) approach [21, 23], which requires no additional training for new tasks. In this approach, only one annotated reference image is selected for each segmentation task, and this image is used to generate the initial prompt for testing images. The second paradigm is the coarse segmentation sampling (CS) approach, which involves few-shot training for new tasks. For each task, we selected 20 annotated images, trained a U-Net [31] with 200 epochs, and used the resulting coarse segmentation masks to sample initial prompts for the segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.245, + 0.779 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conduct an ablation study on the initial prompts generated by FM to evaluate the segmentation performance of SAM with PPO under varying parameter settings. Quantitative results are presented in Table 1. Initially, when SAM applies unoptimized prompts \\((\\alpha = 0, \\beta = 0)\\), segmentation outcomes are suboptimal across all datasets. In contrast, applying PPO with optimized settings significantly enhances prompt quality, leading to notable performance gains for" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.692, + 0.106 + ], + "angle": 0, + "content": "SAM across various tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.212 + ], + "angle": 0, + "content": "When we examine optimization within the feature space \\((\\alpha = 1, \\beta = 0)\\) and physical space \\((\\alpha = 0, \\beta = 1)\\) separately, both settings improve SAM's performance; however, feature space optimization proves to be more effective, yielding a marked improvement over the initial prompts. The better results are achieved when both feature and physical spaces are optimized concurrently \\((\\alpha = 1, \\beta = 1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.333 + ], + "angle": 0, + "content": "Interestingly, further increasing the weight on the physical space \\((\\beta = 5)\\) reduces segmentation performance, likely due to an overemphasis on physical distributions at the expense of essential feature space information. In contrast, placing a higher weight on the feature space \\((\\alpha = 5, \\beta = 1)\\) yields the best segmentation results. These experiments demonstrate that the accuracy of point prompts is more crucial than sparsity, though sparsity remains beneficial." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.905, + 0.485 + ], + "angle": 0, + "content": "In Fig. 3, we compare SAM's qualitative segmentation results using initial prompts versus PPO-optimized prompts under the best parameter settings. This figure includes images from both natural and medical datasets. The initial prompts contain inaccuracies and excessive redundancy, impairing SAM's segmentation accuracy. After PPO refinement, the prompts become more precise and sparse, allowing the segmentation results to closely align with the ground truth, demonstrating PPO's effectiveness in enhancing SAM's adaptability across diverse segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.492, + 0.833, + 0.507 + ], + "angle": 0, + "content": "4.4. Comparison with One-Shot Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.509, + 0.905, + 0.75 + ], + "angle": 0, + "content": "Our goal is to enable SAM to adapt to segmentation tasks across diverse scenarios with minimal cost. To this end, we compared recently proposed methods with our PPO-optimized FM prompt generation method, both following the one-shot paradigm, across multiple datasets. As shown in Table 2, FM-PPO achieves competitive performance in various natural image segmentation tasks without training on in-domain datasets, performing comparably to models trained on in-domain data and surpassing other SAM-based prompt generation methods. Additionally, FM-PPO significantly outperforms other segmentation methods on medical datasets without requiring additional training. This demonstrates that PPO-optimized visual prompt generation retains its training-free task transfer capability while ensuring robust segmentation performance for downstream SAM through optimized prompts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.751, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Moreover, Fig. 4 illustrates the qualitative results of training-free methods on natural and medical datasets, providing visual evidence of FM-PPO's effectiveness. The experimental results show that FM-PPO performs well even in small target segmentation by reducing erroneous prompt points. This success is attributed to the homogeneous feature expectation within the segmentation target and the heterogeneous feature expectation between the target and background during agent training. For medical image segmentation, incorrect and redundant prompts due to domain" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4337" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.089, + 0.898, + 0.105 + ], + "angle": 0, + "content": "Table 1. Ablation study on different spatial weights during PPO training. The results with \\( \\alpha \\) and \\( \\beta \\) set to 0 represent the initial prompts." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.115, + 0.9, + 0.25 + ], + "angle": 0, + "content": "
Feature SpacePhysical SpaceFSS-1000VOC2012ISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
α=0β=072.19.934.712.148.59.927.213.2
α=1β=081.38.559.410.561.79.256.610.3
α=0β=173.59.938.411.751.99.930.513.1
α=1β=186.67.863.710.369.28.561.39.6
α=1β=582.48.956.910.264.99.058.89.7
α=5β=191.97.473.79.975.48.170.39.0
" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.263, + 0.195, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.329, + 0.194, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.263, + 0.279, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.328, + 0.279, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.263, + 0.364, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.328, + 0.364, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.263, + 0.455, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.328, + 0.456, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.263, + 0.539, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.328, + 0.54, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.263, + 0.625, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.328, + 0.625, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.263, + 0.716, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.328, + 0.716, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.263, + 0.8, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.328, + 0.8, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.263, + 0.884, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.328, + 0.884, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.397, + 0.195, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.463, + 0.194, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.526, + 0.164, + 0.538 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.397, + 0.279, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.463, + 0.279, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.526, + 0.248, + 0.538 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.397, + 0.364, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.463, + 0.364, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.526, + 0.332, + 0.538 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.397, + 0.455, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.463, + 0.455, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.526, + 0.424, + 0.538 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.397, + 0.54, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.463, + 0.54, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.526, + 0.509, + 0.538 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.397, + 0.625, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.463, + 0.625, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.526, + 0.592, + 0.538 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.397, + 0.716, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.463, + 0.716, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.526, + 0.684, + 0.538 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.397, + 0.8, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.463, + 0.8, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.526, + 0.768, + 0.538 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.397, + 0.884, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.463, + 0.884, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.835, + 0.526, + 0.853, + 0.538 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.555, + 0.908, + 0.585 + ], + "angle": 0, + "content": "Figure 3. Qualitative segmentation results of initial prompts and optimized prompts. (a) ground truth, (b) segmentation results with initial prompts, (c) segmentation results with optimized prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.486, + 0.881 + ], + "angle": 0, + "content": "differences can lead to poorer segmentation outcomes for SAM. Through prompt optimization, FM-PPO effectively bridges the domain gap between medical and natural images, enhancing segmentation performance across both domains. Additionally, FM-PPO promotes a relatively uniform physical distribution of point prompts and close proximity between positive and negative prompts. In contrast, Matcher lacks negative prompts, which leads to excessive background segmentation on medical images with domain shifts, particularly when segmenting smaller targets, resulting in suboptimal performance. We further elaborate on the necessity of negative prompts in the supplementary materials. As a result, our method excels at boundary segmentation, especially in complex medical images where other methods struggle with background-foreground adhesion. FM-PPO can optimize the initial prompts generated by FM without additional training, achieving efficient segmentation across various scenarios." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.626 + ], + "angle": 0, + "content": "4.5. Different Initial Prompts Scheme Optimization" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.628, + 0.907, + 0.854 + ], + "angle": 0, + "content": "To further validate PPO's effectiveness in optimizing prompts, we applied it to various initial prompt generation schemes and evaluated the downstream SAM segmentation. We selected medical datasets with domain gap, as prompts have a greater impact on SAM's segmentation. Table 3 presents the experimental results, with corresponding qualitative results in the supplementary materials. FM and CS serve as two baseline prompt generation methods. As shown, PPO significantly enhances segmentation performance on both datasets by optimizing the prompts. FM generates a richer prompt set based on patch matching, allowing PPO to achieve stronger optimization. In contrast, CS samples within coarse segmentation regions, leading to a concentrated prompt distribution that may overlook some target areas, thus limiting PPO's optimization potential." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Additionally, Matcher [23], a recent approach in visual prompt engineering, utilizes the FM paradigm to achieve task-agnostic segmentation. To demonstrate PPO's" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4338" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Comparison results with one-shot methods in natural and medical image datasets. Gray metrics indicates the model is already trained by in-domain datasets. \\( {}^{ \\dagger } \\) indicates the training-free method for segmentation tasks." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.129, + 0.903, + 0.266 + ], + "angle": 0, + "content": "
MethodsFSS-1000VOC2012ISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
HSNet [24]91.18.244.511.457.411.329.716.5
VAT [7]93.67.548.211.449.812.926.416.8
SegGPT† [39]93.77.178.27.861.29.544.114.4
PerSAM† [43]81.18.653.610.937.616.145.816.3
Matcher† [23]91.67.669.011.173.79.746.216.0
FM-PPO† (Ours)91.97.473.79.975.48.170.39.0
" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.29, + 0.472, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.626 + ], + "angle": 0, + "content": "Figure 4. Qualitative segmentation results of different one-shot methods in natural and medical images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.777 + ], + "angle": 0, + "content": "plug-and-play capability, we applied PPO to optimize the prompts generated by Matcher and then fed them back into Matcher for segmentation. Table 3 shows that PPO-optimized prompts significantly improve segmentation performance, particularly on the Kvasir, where the DSC increased by \\(32.6\\%\\) for smaller targets. This demonstrates that our method effectively boosts SAM's segmentation performance by optimizing various types of initial prompts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.791, + 0.212, + 0.807 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this paper, we introduce PPO, designed to enhance the distribution of point prompts, thereby improving the downstream automatic segmentation performance of SAM. We first formalize the distribution of point prompts as a dual-space heterogeneous graph and define the optimization direction. Next, we train an agent within this graph environ-" + }, + { + "type": "table_caption", + "bbox": [ + 0.522, + 0.29, + 0.896, + 0.305 + ], + "angle": 0, + "content": "Table 3. Optimization performance of different initial prompts." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.315, + 0.898, + 0.437 + ], + "angle": 0, + "content": "
Prompt SchemeISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
FM48.59.927.213.2
FM-PPO75.4 (↑21.9)8.1 (↓1.8)70.3 (↑43.1)9.0 (↓4.2)
CS57.711.969.110.2
CS-PPO68.6 (↑10.9)10.8 (↓1.1)79.8 (↑10.7)8.9 (↓1.3)
Matcher [23]73.79.746.216.0
Matcher-PPO85.9 (↑12.2)9.0 (↓0.7)77.6 (↑32.6)9.3 (↓6.7)
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.907, + 0.715 + ], + "angle": 0, + "content": "ment to automatically activate effective prompt schemes, which are then input into SAM to obtain segmentation results. Our experiments on natural and medical image datasets demonstrate that PPO significantly improves segmentation performance compared to various initial point prompt schemes. In conclusion, PPO offers a promising solution for optimizing any initial point prompts, enhancing the performance of diverse downstream segmentation tasks. Limitations. Although PPO demonstrates strong performance in optimizing incorrect and redundant point prompts, its effectiveness is limited when the initial prompts lack coverage of the target. For example, in the CS approach, if the pseudo-labels do not include the target, prompts are not sampled over the target region. Future research will focus on designing more comprehensive point prompt generation methods to build a more robust prompt optimizer." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.742, + 0.666, + 0.759 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This research was supported by the National Natural Science Foundation of China (Grant No. 61901292 and Grant No. 62201331), the Fundamental Research Program of Shanxi Province (Grant No. 201901D211080, Grant No. 202303021211082, Grant No. 202403021222276 and Grant No. 202403021212262), the Shanxi Province Graduate Education Innovation Project (Grant No. 2022Y196). We also express our gratitude to the all reviewers and area chairs for their constructive feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4339" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.108, + 0.484, + 0.167 + ], + "angle": 0, + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.169, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.483, + 0.32 + ], + "angle": 0, + "content": "[3] Zhen Chen, Qing Xu, Xinyu Liu, and Yixuan Yuan. Un-sam: Universal prompt-free segmentation for generalized nuclei images. arXiv preprint arXiv:2402.16663, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.322, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[4] Haixing Dai, Chong Ma, Zhengliang Liu, Yiwei Li, Peng Shu, Xiaozheng Wei, Lin Zhao, Zihao Wu, Dajiang Zhu, and Wei Liu. Samaug: Point prompt augmentation for segment anything model. arXiv preprint arXiv:2307.01187, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.399, + 0.483, + 0.457 + ], + "angle": 0, + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.483, + 0.533 + ], + "angle": 0, + "content": "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.611 + ], + "angle": 0, + "content": "[7] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.613, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[8] Duojun Huang, Xinyu Xiong, Jie Ma, Jichang Li, Zequn Jie, Lin Ma, and Guanbin Li. Alignsam: Aligning segment anything model to open context via reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3205-3215, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.483, + 0.763 + ], + "angle": 0, + "content": "[9] Jiaxing Huang, Kai Jiang, Jingyi Zhang, Han Qiu, Lewei Lu, Shijian Lu, and Eric Xing. Learning to prompt segment anything models. arXiv preprint arXiv:2401.04651, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[10] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, pages 451–462. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.108, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.155, + 0.905, + 0.229 + ], + "angle": 0, + "content": "[12] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.321 + ], + "angle": 0, + "content": "[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.326, + 0.905, + 0.4 + ], + "angle": 0, + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.907, + 0.479 + ], + "angle": 0, + "content": "[15] Hyeokjun Kweon and Kuk-Jin Yoon. From sam to cams: Exploring segment anything model for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19499-19509, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.482, + 0.905, + 0.541 + ], + "angle": 0, + "content": "[16] Yiming Lei, Jingqi Li, Zilong Li, Yuan Cao, and Hongming Shan. Prompt learning in computer vision: a survey. Frontiers of Information Technology & Electronic Engineering, 25(1):42-63, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.619 + ], + "angle": 0, + "content": "[17] Chengyin Li, Prashant Khanduri, Yao Qiang, Rafi Ibn Sultan, Indrin Chetty, and Dongxiao Zhu. Autoprompting sam for mobile friendly 3d medical image segmentation. arXiv preprint arXiv:2308.14936, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.682 + ], + "angle": 0, + "content": "[18] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.686, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.779, + 0.905, + 0.853 + ], + "angle": 0, + "content": "[20] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.857, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[21] Xueyu Liu, Guangze Shi, Rui Wang, Yexin Lai, Jianan Zhang, Lele Sun, Quan Yang, Yongfei Wu, Ming Li, Weixia Han, et al. Feature-prompting gbmseg:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4340" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.128, + 0.092, + 0.483, + 0.167 + ], + "angle": 0, + "content": "One-shot reference guided training-free prompt engineering for glomerular basement membrane segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 276-285. Springer, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.169, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[22] Xueyu Liu, Rui Wang, Yexin Lai, Yongfei Wu, Hangbei Cheng, Yuanyue Lu, Jianan Zhang, Ning Hao, Chenglong Ban, Yanru Wang, Shuqin Tang, Yuxuan Yang, Ming Li, Xiaoshuang Zhou, and Wen Zheng. Msmtseg: Multi-stained multi-tissue segmentation of kidney histology images via generative self-supervised meta-learning framework. IEEE Journal of Biomedical and Health Informatics, pages 1-12, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.379 + ], + "angle": 0, + "content": "[23] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.381, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[24] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.444, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[25] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.595 + ], + "angle": 0, + "content": "[26] Saiyang Na, Yuzhi Guo, Feng Jiang, Hehuan Ma, and Junzhou Huang. Segment any cell: A sam-based autoprompting fine-tuning framework for nuclei segmentation. arXiv preprint arXiv:2401.13220, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.597, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[27] Maxime Aquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.763 + ], + "angle": 0, + "content": "[28] Xiaoyu Qiu, Hao Feng, Yuechen Wang, Wengang Zhou, and Houqiang Li. Progressive multi-modal conditional prompt tuning. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 46-54, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.765, + 0.482, + 0.868 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.871, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[30] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.55, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.154, + 0.905, + 0.229 + ], + "angle": 0, + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.231, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[32] Tal Shaharabany, Aviad Dahan, Raja Giryes, and Lior Wolf. Autosam: Adapting sam to medical images by overloading the prompt encoder. arXiv preprint arXiv:2306.06370, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[33] Junji Shiraishi, Shigehiko Katsuragawa, Junpei Ikezoe, Tsuneo Matsumoto, Takeshi Kobayashi, Kenichi Komatsu, Mitate Matsui, Hiroshi Fujita, Yoshie Kodera, and Kunio Doi. Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American journal of roentgenology, 174(1):71-74, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[34] Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. Vrp-sam: Sam with visual reference prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23565-23574, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[35] Chenglong Wang, Dexuan Li, Sucheng Wang, Chengxiu Zhang, Yida Wang, Yun Liu, and Guang Yang. SAMMed: A medical image annotation framework based on large vision model. arXiv preprint arXiv:2307.05617, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.597, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[36] Jiaqi Wang, Zhengliang Liu, Lin Zhao, Zihao Wu, Chong Ma, Sigang Yu, Haixing Dai, Qiushi Yang, Yiheng Liu, Songyao Zhang, et al. Review of large vision models and visual prompt engineering. MetaRadiology, page 100047, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[37] Rui Wang, Zhouwang Yang, and Yanzhi Song. Osamfundus: A training-free, one-shot segmentation framework for optic disc and cup in fundus images. Biomedical Signal Processing and Control, 100:107069, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.824 + ], + "angle": 0, + "content": "[38] Xu Wang, Sen Wang, Xingxing Liang, Dawei Zhao, Jincai Huang, Xin Xu, Bin Dai, and Qiguang Miao. Deep reinforcement learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 35 (4):5064-5078, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.826, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[39] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1130-1140, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "4341" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.091, + 0.482, + 0.152 + ], + "angle": 0, + "content": "[40] Bin Xie, Hao Tang, Bin Duan, Dawen Cai, and Yan Yan. Masksam: Towards auto-prompt sam with mask classification for medical image segmentation. arXiv preprint arXiv:2403.14103, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.153, + 0.483, + 0.229 + ], + "angle": 0, + "content": "[41] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.231, + 0.483, + 0.32 + ], + "angle": 0, + "content": "[42] Chaoning Zhang, Fachrina Dewi Puspitasari, Sheng Zheng, Chenghao Li, Yu Qiao, Taegoo Kang, Xinru Shan, Chenshuang Zhang, Caiyan Qin, Francois Rameau, et al. A survey on segment anything model (sam): Vision foundation model meets prompt engineering. arXiv preprint arXiv:2306.06211, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.322, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[43] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.384, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[44] Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, et al. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.461, + 0.482, + 0.535 + ], + "angle": 0, + "content": "[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.091, + 0.483, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4342" + } + ] +] \ No newline at end of file diff --git a/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/fb8f3fbf-9e73-432d-9130-bbe1d55d00ee_origin.pdf b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/fb8f3fbf-9e73-432d-9130-bbe1d55d00ee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..935d412c6bcea444f8a0bb0a1f1c0acf6168b757 --- /dev/null +++ b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/fb8f3fbf-9e73-432d-9130-bbe1d55d00ee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3368c6ecdc52be0f81453ee042e3c7bca946f5234258a63880374a9fdcc5646f +size 5657968 diff --git a/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/full.md b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c08f34887d4e1ea08df20acf75f482705255dc4d --- /dev/null +++ b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/full.md @@ -0,0 +1,384 @@ +# Plug-and-Play PPO: An Adaptive Point Prompt Optimizer Making SAM Greater + +Xueyu Liu $^{1\dagger}$ , Rui Wang $^{2\dagger}$ , Yexin Lai $^{1}$ , Guangze Shi $^{1}$ , Feixue Shao $^{1}$ , Fang Hao $^{1}$ , Jianan Zhang $^{1}$ , Jia Shen $^{1}$ , Yongfei Wu $^{1*}$ , Wen Zheng $^{1}$ + +$^{1}$ Taiyuan University of Technology + +$^{2}$ University of Science and Technology of China + +# Abstract + +Powered by extensive curated training data, the Segment Anything Model (SAM) demonstrates impressive generalization capabilities in open-world scenarios, effectively guided by user-provided prompts. However, the class-agnostic characteristic of SAM renders its segmentation accuracy highly dependent on prompt quality. In this paper, we propose a novel plug-and-play dual-space Point Prompt Optimizer (PPO) designed to enhance prompt distribution through deep reinforcement learning (DRL)-based heterogeneous graph optimization. PPO optimizes initial prompts for any task without requiring additional training, thereby improving SAM's downstream segmentation performance. Specifically, PPO constructs a dual-space heterogeneous graph, leveraging the robust feature-matching capabilities of a foundational pre-trained model to create internal feature and physical distance matrices. A DRL policy network iteratively refines the distribution of prompt points, optimizing segmentation predictions. We conducted experiments on four public datasets. The ablation study explores the necessity and balance of optimizing prompts in both feature and physical spaces. The comparative study shows that PPO enables SAM to surpass recent one-shot methods. Additionally, experiments with different initial prompts demonstrate PPO's generality across prompts generated by various methods. In conclusion, PPO redefines the prompt optimization problem as a heterogeneous graph optimization task, using DRL to construct an effective, plug-and-play prompt optimizer. This approach holds potential for broader applications across diverse segmentation tasks and provides a promising solution for point prompt optimization. The source code and demo are available at https://github.com/XueyuLiu/PPO. + +# 1. Introduction + +Segmentation, as a fundamental task, plays a pivotal role in visual understanding systems. The prevailing train- + +![](images/acd48d42f961ae46fe3db906271ce5cafb1374b8a88812eab8054f7e2fbed250.jpg) +Figure 1. Ineffective prompts can diminish SAM's performance. Incorrect prompts may cause the segmentation mask to capture irrelevant regions, while excessive prompts for adjacent areas can lead to neglect of other parts of the object. Therefore, prompt optimization is essential to improve SAM's performance. + +ing paradigm for segmentation relies on data-driven approaches, where large datasets are used to train models. While this method has achieved remarkable success, it faces several challenges, including the computational resources to train models for new tasks and the risk of domain-specific optimizations that can limit generalizability [18, 22]. + +Recent breakthroughs in Pretrained Foundation Models (PFMs) in computer vision (CV) have demonstrated impressive capabilities in domain-invariant segmentation across diverse scenarios [11, 29]. Vision Foundation Models (VFMs), such as DINOv2 [2, 27], acquire generalized visual features by capturing intricate information at both the image and pixel levels using only raw image data. This approach enables these models to learn generic visual representations, ensuring robust zero-shot transferability for various downstream tasks. Meanwhile, the Segment Anything Model (SAM) [14, 30], trained on large-scale, high-quality datasets, has shown a strong ability to generalize across tasks and data distributions that are not encountered during + +training. However, SAM relies heavily on manual prompts, such as bounding boxes and points, to identify objects of interest, as it is pretrained in a class-agnostic manner. Consequently, achieving task-specific adaptation of SAM is crucial to bridge the gap between its class-agnostic predictions and the specific needs of industrial applications. + +Some works attempt to train prompt encoders that enables SAM to adapt to downstream segmentation tasks [8, 15, 34, 43]. However, these task-specific prompt encoders require additional training for new segmentation scenarios. Recently, other efforts have focused on automatically generating geometric visual prompts, such as point prompts, to enhance SAM's segmentation performance [21, 23, 37]. In such methods, the quality of prompt schemes remains a critical factor influencing downstream segmentation performance. As illustrated in Fig. 1, incorrect prompts can result in erroneous segmentation, while excessive prompts in specific areas may cause SAM to focus disproportionately on those regions, neglecting other parts of the target. Current automatic prompt engineering methods often lack optimization in prompt distribution, leading to lower-quality prompts and segmentation performance. Therefore, there is an urgent need for methods that can refine prompts to effectively enhance SAM's performance. + +To address this, effective point prompts for SAM should exhibit certain characteristics in both feature and physical spaces. In the feature space, regions containing positive prompts should be homogeneous to ensure consistency and clarity within the segmentation target, while regions between positive and negative prompts should be heterogeneous, distinguishing the target from the background. In the physical space, positive and negative prompts should be sufficiently spaced apart to mitigate the effects of excessive prompts and prevent over-segmentation, while also being close enough to boundaries to ensure precise delineation. + +Building upon the aforementioned inspiration and committed to addressing the challenges mentioned above, we propose a novel Plug-and-Play Point Prompt Optimizer (PPO) designed to enhance SAM's segmentation performance by automatically optimizing initial point prompt schemes. Leveraging DINOv2's powerful feature association capabilities [27], we construct an internal feature and physical distance matrix within each image to create a heterogeneous graph environment. To optimize prompt distribution in dual spaces, we employ Deep Reinforcement Learning (DRL), a framework that trains agents to maximize cumulative rewards through interaction and feedback mechanisms. By designing specific actions and reward functions, PPO optimizes prompt point distribution under the guidance of DRL and inputs them into SAM to obtain refined segmentation results. We validated our approach through experiments on two natural and two medical image datasets. Ablation studies demonstrate that PPO + +optimized prompt distributions significantly enhance segmentation performance. Comparative experiments show that PPO outperforms recent methods that have not been trained on in-domain datasets. Additionally, experiments with different initial prompts confirm PPO's generality across prompts generated by various methods. The main contributions are summarized as follows: + +- To the best of our knowledge, we are the first to define an optimization direction for point prompts in both feature and physical spaces. Additionally, we are the first to use DRL to optimize point prompts, achieving prompt optimization without additional training for new tasks. +- PPO operates in a plug-and-play manner, allowing for the optimization of any initial geometric point prompts to improve SAM's segmentation performance. +- Extensive experiments demonstrate the effectiveness of PPO in optimizing point prompts across various initial prompt schemes. + +# 2. Related Work + +# 2.1. Pretrained Foundation Models + +In natural language processing (NLP), PFMs, such as BERT [5], ChatGPT, and GPT-4, have become foundational tools for a wide range of downstream tasks across diverse data modalities. By leveraging extensive text data, PFMs provide effective parameter initialization for various NLP applications [44]. The success of PFMs in NLP has inspired researchers to explore their application in CV. Visual language models like CLIP [29] and GLIP [19], trained on large-scale datasets of millions of image-text pairs from the web, exhibit strong zero-shot generalization across multiple vision tasks, offering representations with robust generalization and transfer capabilities. + +Building on this trend, Meta AI introduced the SAM, a notable advancement in applying VFMs. SAM allows users to specify image segmentation tasks through prompts and demonstrates impressive zero-shot learning, enabling it to generalize effectively across diverse tasks without task-specific training [14]. However, SAM's segmentation performance heavily relies on the quality of input prompts. To address this, we propose an adaptive point prompt optimizer that automatically refines prompts to enhance SAM's segmentation performance. + +In another significant development, DINOv2 [27] leverages discriminative self-supervised learning at both image and patch levels, showing exceptional patch-matching capabilities that allow it to capture semantic similarities across images of the same target. By utilizing these patch-matching capabilities, we can extract feature distances between patches within an image, providing a theoretical foundation for constructing heterogeneous graphs to achieve effective prompt optimization. + +# 2.2. Automatic Prompt Engineering + +In the field of NLP, extensive researches have been conducted on designing automatic prompt engineering, with prompt tuning methods achieving significant advances in enhancing PFMs' performance through optimized language prompts [1, 12, 13, 16, 28, 41, 45, 45]. Simultaneously, efforts have been made to explore automatic prompt engineering in CV. On the one hand, some recent works have attempted to train prompt encoders that enable SAM to adapt to downstream segmentation tasks [9, 32, 34, 40]. For example, Sun et al. proposed VRP-SAM, an extension of SAM that encodes various granularities of visual references directly into prompt embeddings for mask generation [34]. These approaches encode SAM's prompts into prompt embeddings via a prompt encoder, which are then fed into SAM's mask decoder to generate target masks. However, these methods are often task-sensitive, requiring training different prompt encoders for different tasks, thus limiting rapid task transfer. + +On the other hand, some works attempt to generate geometric prompts to achieve automatic prompt engineering for SAM, generally following two main paradigms. The first paradigm uses a small amount of annotated data to train a segmentation model, then samples prompts from pseudolabels generated by this model [3, 26, 36, 42, 43]. For instance, Li et al. randomly selected prompt points from masks generated by coarse segmentation, which were subsequently input into SAM after fine-tuning to obtain segmentation results [17]. Wang et al. utilized a novel prototype learning framework to generate pseudo-labels and extract prompts, which were then incorporated into SAM for segmentation [35]. Dai et al. developed various methods to expand prompt points, exploring the potential of a single prompt point for enhanced segmentation performance [4]. + +However, this first paradigm still requires a small amount of data to train the prompt generator, also limiting task transferability. To improve transferability, the second paradigm generates initial prompts by leveraging feature similarity between reference and target images without additional training [21, 23, 37, 43]. For example, Zhang et al. calculated feature similarity between pixels in the target and reference images to create a point prompt scheme using only a single reference image [43]. Liu et al. introduced Matcher, an instance-level and bidirectional matching strategy for prompt point generation [23], and Liu et al. proposed a series of prompt optimization strategies for the automatic segmentation of basement membranes in electron microscopy images [21]. However, the quality of prompts generated by these methods is often limited, as they do not sufficiently consider the relationships between positive and negative prompts, resulting in redundant and erroneous prompts that negatively affect segmentation performance. + +In our work, PPO addresses this limitation by con + +structuring a heterogeneous graph based on dual-space distances between initial prompts within each image. This approach optimizes any automatically generated geometric point prompts without additional training for new tasks, enhancing SAM's segmentation performance while preserving task transferability by ensuring high prompt quality. + +# 3. Methodology + +This section introduces our plug-and-play PPO, which is based on a heterogeneous graph structure designed to enhance the segmentation performance of SAM by optimizing initial point prompts. An overview of PPO is illustrated in Fig. 2. Specifically, we first compute the feature distance matrix and the physical distance matrix for each image. Using these matrices, we construct a heterogeneous graph where the edge attributes are defined by the feature and physical distances based on the initial prompt indices. During the training phase, we use this heterogeneous graph as the environment to design an action list and reward function to train the optimization agent. In the inference phase, the trained agent activates the appropriate point prompts as a final prompt scheme, which is then input into SAM to obtain the final segmentation result. In the following subsections, we will describe the main components in detail. + +# 3.1. Dual-Space Distance Matrix Calculation + +The core objective of our method is to optimize positive and negative point prompts in both feature and distance spaces. To achieve this, we construct a heterogeneous graph comprising two homogeneous graphs that represent the feature and distance spaces. Edge values are determined by creating two patch-level distance matrices for each image $X$ . + +First, we divide $X$ into patches $x$ using a sliding window approach, where $x_{i} = x_{1},x_{2},\ldots ,x_{n}$ . We then use the image encoder from DINOv2 [27] to extract patch-level features for $x$ , represented by $f$ , where $f_{i} = f_{1},f_{2},\dots ,f_{n}$ . Next, we calculate the feature matrix $M_{f}$ as follows: + +$$ +M _ {f} (i, j) = \left\| f _ {i} - f _ {j} \right\|, \tag {1} +$$ + +where $M_{f}(i,j)$ represents the Euclidean distance $D_{f}$ between the $i$ -th patch feature $f_{i}$ and the $j$ -th patch feature $f_{j}$ . Similarly, we calculate the physical distance matrix $M_{p}$ using the center points of the patches $x$ as follows: + +$$ +M _ {p} (i, j) = \left\| x _ {i} - x _ {j} \right\|, \tag {2} +$$ + +where $M_{p}(i,j)$ represents the Euclidean distance $D_{p}$ between the $i$ -th patch center $x_{i}$ and the $j$ -th patch center $x_{j}$ . + +# 3.2. Graph Environment Creation + +Inspired by coarse segmentation sampling [3, 26] and feature matching techniques [21, 23], we can obtain initial + +![](images/a56ba54ca5273f651e41db2d47198126b844b525005abab8f95dd85e2bbb526e.jpg) +Figure 2. The PPO workflow is designed to enhance SAM's segmentation performance by optimizing initial point prompts through two key components: graph-based environment creation and agent training for prompt optimization. After optimized, the prompts are fed into SAM to obtain segmentation results. + +point prompt indices for optimization. Specifically, we define a heterogeneous graph $G = (V,E,\phi ,\psi)$ , where $V$ is the set of nodes and $E$ is the set of edges. The node type set $\mathcal{A} = \{\text{positive nodes } V_p, \text{negative nodes } V_n\}$ is represented by the mapping function $\phi (v): V \to \mathcal{A}$ , and the edge type set $\mathcal{R} = \{\text{feature edges } E_f, \text{physical edges } E_p\}$ is represented by the mapping function $\psi (e): E \to \mathcal{R}$ . + +We use the heterogeneous graph $G$ as the environment for DRL. By selecting different nodes, we can obtain different edge values as feedback. + +# 3.3. Agent Training for Prompt Optimization + +On the one hand, an effective point prompt scheme should ensure correctness. In an image, the features of the segmentation target are often homogeneous, while the features of the background are heterogeneous. Therefore, we aim for the feature distance between positive point prompts to be sufficiently close, and the feature distance between positive and negative point prompts to be sufficiently far. On the other hand, an effective point prompt scheme should also ensure sparsity. Over-prompting in adjacent areas can degrade segmentation performance. Thus, the physical distance between positive point prompts (and between negative point prompts) should not be too close. Additionally, to enhance boundary segmentation accuracy, the physical dis + +tance between positive and negative point prompts should not be too far. + +A DRL model can infer actions and receive rewards from the environment, reflecting the significance of these actions and updating the agent's state accordingly [38]. Using DRL, we aim to train an agent to activate appropriate point prompts to optimize the distribution. The detailed process is shown in Algorithm 1. The key components of agent training for prompt optimization will be introduced as follows: + +Actions: We define three possible actions: (1) restore node $A_r$ , which randomly activates a node and connects the corresponding edges; (2) delete node $A_d$ , which randomly deletes a node and removes the corresponding edges; (3) move node $A_m$ , which randomly selects a direction and moves the index position of a node. Additionally, we aim to maintain a stable number of prompts. Therefore, if the number of nodes $|V_p|$ or $|V_n|$ in the current state is less than the minimum value, the only possible action is $A_r$ . Conversely, if the $|V_p|$ or $|V_n|$ exceeds the maximum value, the only possible action is $A_d$ . Only when the $|V_p|$ or $|V_n|$ is between the minimum and maximum values are all three actions, $A_r$ , $A_d$ , and $A_m$ , possible. + +States: For each action, there will be different nodes, resulting in a different heterogeneous graph $G^{s}$ . Based on + +# Algorithm 1 PPO training process + +Input: Heterogeneous graph environment $G$ + +Parameter: Max step; Max epoch; + +Max nodes $N_{max}$ ; Min nodes $N_{min}$ + +# Output: Q-values + +1: Initializing the Q-values, Epoch, Step. +2: while Epoch < Max epoch do +3: Initializing the Step. +4: while Step $<$ Max step do +5: Emptying all nodes and edges. +6: Emptying reward. +7: Calculating the number of Nodes $N_{pos}$ and $N_{neg}$ . +8: if $N_{pos} < N_{min}$ or $N_{neg} < N_{min}$ then +9: Action list $= \{A_{r}\}$ +10: else if $N_{pos} > N_{max}$ or $N_{neg} > N_{max}$ then +11: Action list $= \{A_{d}\}$ +12: else +13: Action list $= \{A_r, A_d, A_m\}$ . +14: end if +15: Perform a random action and update the status $s$ +16: Adding the reward $r_s$ based on the reward function. +17: Updating Q-values. +18: Step+=1. +19: end while +20: Epoch $+ = 1$ +21: end while +22: return solution + +$G^{s}$ , we can calculate the following mean edge values: + +$$ +D _ {f} ^ {s} (p, p) = \frac {1}{| V _ {p} | | V _ {p} |} \sum_ {i \in V _ {p}} \sum_ {j \in V _ {p}} M _ {f} (i, j), \tag {3} +$$ + +$$ +D _ {f} ^ {\mathrm {s}} (p, n) = \frac {1}{| V _ {p} | | V _ {n} |} \sum_ {i \in V _ {p}} \sum_ {j \in V _ {n}} M _ {f} (i, j), \tag {4} +$$ + +$$ +D _ {p} ^ {s} (p, p) = \frac {1}{| V _ {p} | | V _ {p} |} \sum_ {i \in V _ {p}} \sum_ {j \in V _ {p}} M _ {p} (i, j), \tag {5} +$$ + +$$ +D _ {p} ^ {s} (n, n) = \frac {1}{| V _ {n} | | V _ {n} |} \sum_ {i \in V _ {n}} \sum_ {j \in V _ {n}} M _ {p} (i, j), \tag {6} +$$ + +$$ +D _ {p} ^ {s} (p, n) = \frac {1}{| V _ {p} | | V _ {n} |} \sum_ {i \in V _ {p}} \sum_ {j \in V _ {n}} M _ {p} (i, j). \tag {7} +$$ + +In state $s$ , $D_{f}^{s}(p,p)$ and $D_{f}^{s}(p,n)$ represents the mean feature distance between positive prompts, and between positive and negative prompts. Similarly, $D_{p}^{s}(p,p)$ , $D_{p}^{s}(n,n)$ , and $D_{p}^{s}(p,n)$ represent the mean physical distances in state $s$ between positive prompts, negative prompts, and between positive and negative prompts, respectively. + +Rewards: Based on the various distances obtained from $G^{s}$ , we can set a reward function according to the optimization objectives. The reward function for the current state, $r_{s}$ , is defined based on the changes in distance from the previous state: + +$$ +\begin{array}{l} r _ {s} = - \alpha \left(D _ {f} ^ {s} (p, p) - D _ {f} ^ {s - 1} (p, p)\right) \\ + \alpha \left(D _ {f} ^ {s} (p, n) - D _ {f} ^ {s - 1} (p, n)\right) \\ + \beta \left(D _ {p} ^ {s} (p, p) - D _ {p} ^ {s - 1} (p, p)\right) \tag {8} \\ + \beta (D _ {p} ^ {s} (n, n) - D _ {p} ^ {s - 1} (n, n)) \\ - \beta (D _ {p} ^ {s} (p, n) - D _ {p} ^ {s - 1} (p, n)), \\ \end{array} +$$ + +where coefficients $\alpha$ and $\beta$ balance the feature and physical distance, respectively. Using this reward function, if the current state's $D_{f}^{s}(p,p)$ and $D_{p}^{s}(p,n)$ increase, a penalty is applied. Conversely, if the current state's $D_{f}^{s}(p,n)$ , $D_{p}^{s}(p,p)$ , and $D_{p}^{s}(n,n)$ increase, a reward is given. + +Policy Network: We adopt the Q-network [25] as the policy network, which consists of two fully connected layers with 1024 and 64 neurons, respectively, and a single output layer with 3 neurons. This output layer predicts the Q-value of 3 actions based on the current state. The policy network is regarded as the approximate action-value function $Q_{s}$ and takes the state $s$ as its input to approximate the underlying action-value function. This approximation is achieved by repeatedly performing a gradient descent step on $(y_{s} - Q_{s})^{2}$ with respect to the policy network parameters at each step. The Q-value $y_{s}$ is defined as: + +$$ +y _ {s} = \left\{ \begin{array}{l l} r _ {s} + \gamma \max _ {s + 1} Q _ {s + 1}, & \text {i f n o t s t o p} \\ r _ {s}, & \text {i f s t o p} \end{array} , \right. \tag {9} +$$ + +where the discount factor $\gamma$ represents the balance between the current and future rewards. + +In the training process of PPO, we initially delete all nodes and treat them as potential activatable nodes in the action environment. As the agent attempts to activate different nodes, the edges of the heterogeneous graph acquire different values. Based on the reward function, the agent receives rewards or penalties and updates the Q-values accordingly. The policy network repeatedly interacts with the segmentation network, updates the network weights according to Eq. (11), and ultimately learns effective policies. + +# 4. Experimental Results + +# 4.1. Implementation Details + +Our experiments are carried out on a Linux server platform equipped with an NVIDIA Tesla V100. During the PPO training phase, we set the step size for each iteration to 100 and the number of training epochs to 1000. The discount factor $\gamma$ of Q-learning is set to 0.1. We use the $\epsilon$ -greedy strategy to balance exploration and exploitation [25]. During $\epsilon$ -greedy training, the agent can select a random action + +with probability $\epsilon$ and a predicted action generated by the Q-network with probability $1 - \epsilon$ . The $\epsilon$ parameters are set as follows: $\epsilon_{\mathrm{start}} = 1$ , $\epsilon_{\mathrm{end}} = 0.1$ , and $\epsilon_{\mathrm{decay}} = 0.995$ . Additionally, to ensure stability during PPO training, the probability of action $A_{m}$ is always set to be 0.5 times the probability of the other actions. + +In the testing phase, PPO requires no additional training for new tasks; it simply inputs initial prompts and images to output optimized prompt schemes that serve SAM. For evaluation, we use the mean Dice Similarity Coefficient (mDSC) and the mean Hausdorff Distance (mHD) as metrics to assess the segmentation performance. + +# 4.2. Dataset Description + +To ensure that PPO remains unseen to in-domain datasets for downstream tasks, we trained PPO on the GBMSD dataset [21], generating initial prompts through feature matching and constructing a heterogeneous graph as the environment for DRL learning. During testing, we evaluated the model on natural image domains using the FSS-1000 [20] and VOC2012 [6] datasets. The FSS-1000 dataset comprises mask-annotated images from 1,000 classes, with 10 images per class, while the VOC2012 dataset is a widely-used benchmark for semantic segmentation tasks, containing 20 object classes and a total of 4,396 images. For medical image domains, we used the ISIC2018 [33] and Kvasir [10]. The ISIC2018 is a large-scale collection of dermatoscopic images focused on lesion segmentation, consisting of 2,594 images. The Kvasir, on the other hand, is an endoscopic dataset aimed at colorectal polyp segmentation, containing 1,000 gastroenterology images of colorectal polyps along with their corresponding segmentation masks. + +We generated initial prompts based on two mainstream point prompt-generation paradigms. The first paradigm is the feature matching (FM) approach [21, 23], which requires no additional training for new tasks. In this approach, only one annotated reference image is selected for each segmentation task, and this image is used to generate the initial prompt for testing images. The second paradigm is the coarse segmentation sampling (CS) approach, which involves few-shot training for new tasks. For each task, we selected 20 annotated images, trained a U-Net [31] with 200 epochs, and used the resulting coarse segmentation masks to sample initial prompts for the segmentation tasks. + +# 4.3. Ablation Study + +We conduct an ablation study on the initial prompts generated by FM to evaluate the segmentation performance of SAM with PPO under varying parameter settings. Quantitative results are presented in Table 1. Initially, when SAM applies unoptimized prompts $(\alpha = 0, \beta = 0)$ , segmentation outcomes are suboptimal across all datasets. In contrast, applying PPO with optimized settings significantly enhances prompt quality, leading to notable performance gains for + +SAM across various tasks. + +When we examine optimization within the feature space $(\alpha = 1, \beta = 0)$ and physical space $(\alpha = 0, \beta = 1)$ separately, both settings improve SAM's performance; however, feature space optimization proves to be more effective, yielding a marked improvement over the initial prompts. The better results are achieved when both feature and physical spaces are optimized concurrently $(\alpha = 1, \beta = 1)$ . + +Interestingly, further increasing the weight on the physical space $(\beta = 5)$ reduces segmentation performance, likely due to an overemphasis on physical distributions at the expense of essential feature space information. In contrast, placing a higher weight on the feature space $(\alpha = 5, \beta = 1)$ yields the best segmentation results. These experiments demonstrate that the accuracy of point prompts is more crucial than sparsity, though sparsity remains beneficial. + +In Fig. 3, we compare SAM's qualitative segmentation results using initial prompts versus PPO-optimized prompts under the best parameter settings. This figure includes images from both natural and medical datasets. The initial prompts contain inaccuracies and excessive redundancy, impairing SAM's segmentation accuracy. After PPO refinement, the prompts become more precise and sparse, allowing the segmentation results to closely align with the ground truth, demonstrating PPO's effectiveness in enhancing SAM's adaptability across diverse segmentation tasks. + +# 4.4. Comparison with One-Shot Methods + +Our goal is to enable SAM to adapt to segmentation tasks across diverse scenarios with minimal cost. To this end, we compared recently proposed methods with our PPO-optimized FM prompt generation method, both following the one-shot paradigm, across multiple datasets. As shown in Table 2, FM-PPO achieves competitive performance in various natural image segmentation tasks without training on in-domain datasets, performing comparably to models trained on in-domain data and surpassing other SAM-based prompt generation methods. Additionally, FM-PPO significantly outperforms other segmentation methods on medical datasets without requiring additional training. This demonstrates that PPO-optimized visual prompt generation retains its training-free task transfer capability while ensuring robust segmentation performance for downstream SAM through optimized prompts. + +Moreover, Fig. 4 illustrates the qualitative results of training-free methods on natural and medical datasets, providing visual evidence of FM-PPO's effectiveness. The experimental results show that FM-PPO performs well even in small target segmentation by reducing erroneous prompt points. This success is attributed to the homogeneous feature expectation within the segmentation target and the heterogeneous feature expectation between the target and background during agent training. For medical image segmentation, incorrect and redundant prompts due to domain + +Table 1. Ablation study on different spatial weights during PPO training. The results with $\alpha$ and $\beta$ set to 0 represent the initial prompts. + +
Feature SpacePhysical SpaceFSS-1000VOC2012ISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
α=0β=072.19.934.712.148.59.927.213.2
α=1β=081.38.559.410.561.79.256.610.3
α=0β=173.59.938.411.751.99.930.513.1
α=1β=186.67.863.710.369.28.561.39.6
α=1β=582.48.956.910.264.99.058.89.7
α=5β=191.97.473.79.975.48.170.39.0
+ +![](images/ab9ce231d488bffebf21739d6fccf11704b565edab3daddb5888770939ade6aa.jpg) + +![](images/85d9168ecc72029c87886af298d28f91a4e7223ee2337c7dee4ab556352c41b0.jpg) + +![](images/f30d2aad89f5edc11dd910da859c4a4e06a7c71f211d71f07cfd1a1fe08e9b57.jpg) + +![](images/258fa199c20e4d80d2c7d4d45fcc677d8dc67956cbb4e945de63ad990d1f9628.jpg) + +![](images/a00f1262ac9a725495e538c9966d0c2b6f42d338baebfc9f7f91444cf81ddb71.jpg) + +![](images/cc09af238f3772a7993b74b55e1f360891699bfe7a968838ff1229d63b4ab32f.jpg) + +![](images/4887450f8d53c8bf4836fa939a2a8834945384d64d0ff5cfb3a42850f5330c1d.jpg) + +![](images/cb489915c2d5abc6c6dbfe71d02bb9bad4b98623a4da22a9054b65a962ba7e6a.jpg) + +![](images/783830a089319459447f16084dc383527c09572097d6fb7422f881609a84d929.jpg) + +![](images/1f8370d466058de0fc32cf764618132c4b0e4e1184c4e090be3933866ed5ab6c.jpg) + +![](images/16ae01fc8d896bfd470dca9aece9e6e5984e664b6f6eea7e9f7e1fc360a69105.jpg) + +![](images/294ae35a955193b3bb59c0c138d9f041f556a1b74f8f90abe72ee47390d364fd.jpg) + +![](images/6b89e89a823ea05b18fb09be4ac9466d3797e4bd1ecc0520c502a238a067d693.jpg) + +![](images/dd94debb756685920b8cf3f6eb95f239b430adcd20b700e3bf62846219bb4ed2.jpg) + +![](images/ecaf7c120d1d14825a706be883fa310487ffb44643bb8b4c19d386ac2566cd46.jpg) + +![](images/854656f251b557acff769bb4fdb183e71f06c5b9ca184b2fadea53520d4f623e.jpg) + +![](images/1a1660150a204dd8dc2d8f84448f5ec52ed10aa4ec2a85808ed9b65d3313ac66.jpg) + +![](images/49b50c368b74b9588dbd562dc1e8ef8e28c37d7336cec38fec75f906bec496de.jpg) + +![](images/46dd9f4bbd0017298a4ed3f58084383a2acc08da40c027bfcb748ec0a62d76fb.jpg) + +![](images/3378bfc7a9eb26f6be7646a7c6bf54f9fc0991ecbfd44c8904075d70a34c4597.jpg) +(a) + +![](images/0544123be9cadc284a36bd1f5a0d89692a00fa286ade5b47d13dea1442e53960.jpg) + +![](images/6f3df79bcf9ab519f133526f87889aa5234f2551567c444ab17bbb59193b9275.jpg) +(b) + +![](images/45dc42db52e8633d2acde500de003c670d3503e99582dff8c3ef5cb437869db0.jpg) + +![](images/c5d4b2feda5a0480ee9c2a212aa3b1900881ef43fe29cb58130b51678234cae0.jpg) +(c) + +![](images/8de72bc3507adaa94da68be45a31b5bf160b52fcd241e49b3bd3dd0764d2d0f6.jpg) + +![](images/82bbda4bbf70bdb19d65371c5cb84cd668fefb7536dc80496a9f5a4e04bdbf54.jpg) +(a) +Figure 3. Qualitative segmentation results of initial prompts and optimized prompts. (a) ground truth, (b) segmentation results with initial prompts, (c) segmentation results with optimized prompts. + +![](images/fcb533011bf89b2a49f3d4cfbca9683139cc08d4d449c2cc6a52c77539945953.jpg) + +![](images/522af46bcc19a49f861cc8cea998d1ee0f306ef5583a4c8feac224c5be222e9b.jpg) +(b) + +![](images/6e71723dc835a06b96f8d8b6acb57442d90ced2579ce69640982f046ee025ff0.jpg) + +![](images/3ea56c0c85a714a81c308fbea5ad42c8103b7f6a41b1714206f07b9f98f2c1da.jpg) +(c) + +![](images/407be0e27b39c19f3035bca69f04a109aa7cc3c060c939d09dcabda6e5cbeb47.jpg) + +![](images/5da259c4633ff8b3e1bdd431ffc9b74b11a67763505fa6892f99607d9675cc54.jpg) +(a) + +![](images/199597e39389a1207f6d2a2b79e4c2fe0271593d0f0c31ef9c0b7d441dfcfd5b.jpg) + +![](images/5399ada3767f2d52b9df0f0da8d22eaee77a99ad95bdc609b8bbddacca773d45.jpg) +(b) + +![](images/87332dd9cd9b791de2c3c48915e8fc4ac252d5f963e8f40a5edb761fa6727aa5.jpg) + +![](images/cb7325b64ad29577351f3d40cb68d6a1e52943bd05457115eda3ac5fa2137cd5.jpg) +(c) + +differences can lead to poorer segmentation outcomes for SAM. Through prompt optimization, FM-PPO effectively bridges the domain gap between medical and natural images, enhancing segmentation performance across both domains. Additionally, FM-PPO promotes a relatively uniform physical distribution of point prompts and close proximity between positive and negative prompts. In contrast, Matcher lacks negative prompts, which leads to excessive background segmentation on medical images with domain shifts, particularly when segmenting smaller targets, resulting in suboptimal performance. We further elaborate on the necessity of negative prompts in the supplementary materials. As a result, our method excels at boundary segmentation, especially in complex medical images where other methods struggle with background-foreground adhesion. FM-PPO can optimize the initial prompts generated by FM without additional training, achieving efficient segmentation across various scenarios. + +# 4.5. Different Initial Prompts Scheme Optimization + +To further validate PPO's effectiveness in optimizing prompts, we applied it to various initial prompt generation schemes and evaluated the downstream SAM segmentation. We selected medical datasets with domain gap, as prompts have a greater impact on SAM's segmentation. Table 3 presents the experimental results, with corresponding qualitative results in the supplementary materials. FM and CS serve as two baseline prompt generation methods. As shown, PPO significantly enhances segmentation performance on both datasets by optimizing the prompts. FM generates a richer prompt set based on patch matching, allowing PPO to achieve stronger optimization. In contrast, CS samples within coarse segmentation regions, leading to a concentrated prompt distribution that may overlook some target areas, thus limiting PPO's optimization potential. + +Additionally, Matcher [23], a recent approach in visual prompt engineering, utilizes the FM paradigm to achieve task-agnostic segmentation. To demonstrate PPO's + +Table 2. Comparison results with one-shot methods in natural and medical image datasets. Gray metrics indicates the model is already trained by in-domain datasets. ${}^{ \dagger }$ indicates the training-free method for segmentation tasks. + +
MethodsFSS-1000VOC2012ISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
HSNet [24]91.18.244.511.457.411.329.716.5
VAT [7]93.67.548.211.449.812.926.416.8
SegGPT† [39]93.77.178.27.861.29.544.114.4
PerSAM† [43]81.18.653.610.937.616.145.816.3
Matcher† [23]91.67.669.011.173.79.746.216.0
FM-PPO† (Ours)91.97.473.79.975.48.170.39.0
+ +![](images/78bd95ee017b830780af4a2b7564d4dece78e3aa48abfa0afef3b6877fc7a834.jpg) +Figure 4. Qualitative segmentation results of different one-shot methods in natural and medical images. + +plug-and-play capability, we applied PPO to optimize the prompts generated by Matcher and then fed them back into Matcher for segmentation. Table 3 shows that PPO-optimized prompts significantly improve segmentation performance, particularly on the Kvasir, where the DSC increased by $32.6\%$ for smaller targets. This demonstrates that our method effectively boosts SAM's segmentation performance by optimizing various types of initial prompts. + +# 5. Conclusion + +In this paper, we introduce PPO, designed to enhance the distribution of point prompts, thereby improving the downstream automatic segmentation performance of SAM. We first formalize the distribution of point prompts as a dual-space heterogeneous graph and define the optimization direction. Next, we train an agent within this graph environ- + +Table 3. Optimization performance of different initial prompts. + +
Prompt SchemeISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
FM48.59.927.213.2
FM-PPO75.4 (↑21.9)8.1 (↓1.8)70.3 (↑43.1)9.0 (↓4.2)
CS57.711.969.110.2
CS-PPO68.6 (↑10.9)10.8 (↓1.1)79.8 (↑10.7)8.9 (↓1.3)
Matcher [23]73.79.746.216.0
Matcher-PPO85.9 (↑12.2)9.0 (↓0.7)77.6 (↑32.6)9.3 (↓6.7)
+ +ment to automatically activate effective prompt schemes, which are then input into SAM to obtain segmentation results. Our experiments on natural and medical image datasets demonstrate that PPO significantly improves segmentation performance compared to various initial point prompt schemes. In conclusion, PPO offers a promising solution for optimizing any initial point prompts, enhancing the performance of diverse downstream segmentation tasks. Limitations. Although PPO demonstrates strong performance in optimizing incorrect and redundant point prompts, its effectiveness is limited when the initial prompts lack coverage of the target. For example, in the CS approach, if the pseudo-labels do not include the target, prompts are not sampled over the target region. Future research will focus on designing more comprehensive point prompt generation methods to build a more robust prompt optimizer. + +# Acknowledgment + +This research was supported by the National Natural Science Foundation of China (Grant No. 61901292 and Grant No. 62201331), the Fundamental Research Program of Shanxi Province (Grant No. 201901D211080, Grant No. 202303021211082, Grant No. 202403021222276 and Grant No. 202403021212262), the Shanxi Province Graduate Education Innovation Project (Grant No. 2022Y196). We also express our gratitude to the all reviewers and area chairs for their constructive feedback. + +# References + +[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 3 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1 +[3] Zhen Chen, Qing Xu, Xinyu Liu, and Yixuan Yuan. Un-sam: Universal prompt-free segmentation for generalized nuclei images. arXiv preprint arXiv:2402.16663, 2024. 3 +[4] Haixing Dai, Chong Ma, Zhengliang Liu, Yiwei Li, Peng Shu, Xiaozheng Wei, Lin Zhao, Zihao Wu, Dajiang Zhu, and Wei Liu. Samaug: Point prompt augmentation for segment anything model. arXiv preprint arXiv:2307.01187, 2023. 3 +[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 +[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6 +[7] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 8 +[8] Duojun Huang, Xinyu Xiong, Jie Ma, Jichang Li, Zequn Jie, Lin Ma, and Guanbin Li. Alignsam: Aligning segment anything model to open context via reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3205-3215, 2024. 2 +[9] Jiaxing Huang, Kai Jiang, Jingyi Zhang, Han Qiu, Lewei Lu, Shijian Lu, and Eric Xing. Learning to prompt segment anything models. arXiv preprint arXiv:2401.04651, 2024. 3 +[10] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, pages 451–462. Springer, 2020. 6 +[11] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen + +Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1 +[12] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3 +[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3 +[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2 +[15] Hyeokjun Kweon and Kuk-Jin Yoon. From sam to cams: Exploring segment anything model for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19499-19509, 2024. 2 +[16] Yiming Lei, Jingqi Li, Zilong Li, Yuan Cao, and Hongming Shan. Prompt learning in computer vision: a survey. Frontiers of Information Technology & Electronic Engineering, 25(1):42-63, 2024. 3 +[17] Chengyin Li, Prashant Khanduri, Yao Qiang, Rafi Ibn Sultan, Indrin Chetty, and Dongxiao Zhu. Autoprompting sam for mobile friendly 3d medical image segmentation. arXiv preprint arXiv:2308.14936, 2023. 3 +[18] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2 +[20] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 6 +[21] Xueyu Liu, Guangze Shi, Rui Wang, Yexin Lai, Jianan Zhang, Lele Sun, Quan Yang, Yongfei Wu, Ming Li, Weixia Han, et al. Feature-prompting gbmseg: + +One-shot reference guided training-free prompt engineering for glomerular basement membrane segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 276-285. Springer, 2024. 2, 3, 6 +[22] Xueyu Liu, Rui Wang, Yexin Lai, Yongfei Wu, Hangbei Cheng, Yuanyue Lu, Jianan Zhang, Ning Hao, Chenglong Ban, Yanru Wang, Shuqin Tang, Yuxuan Yang, Ming Li, Xiaoshuang Zhou, and Wen Zheng. Msmtseg: Multi-stained multi-tissue segmentation of kidney histology images via generative self-supervised meta-learning framework. IEEE Journal of Biomedical and Health Informatics, pages 1-12, 2024. 1 +[23] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 2, 3, 6, 7, 8 +[24] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 8 +[25] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 5 +[26] Saiyang Na, Yuzhi Guo, Feng Jiang, Hehuan Ma, and Junzhou Huang. Segment any cell: A sam-based autoprompting fine-tuning framework for nuclei segmentation. arXiv preprint arXiv:2401.13220, 2024. 3 +[27] Maxime Aquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3 +[28] Xiaoyu Qiu, Hao Feng, Yuechen Wang, Wengang Zhou, and Houqiang Li. Progressive multi-modal conditional prompt tuning. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 46-54, 2024. 3 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2 +[30] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, + +Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1 +[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6 +[32] Tal Shaharabany, Aviad Dahan, Raja Giryes, and Lior Wolf. Autosam: Adapting sam to medical images by overloading the prompt encoder. arXiv preprint arXiv:2306.06370, 2023. 3 +[33] Junji Shiraishi, Shigehiko Katsuragawa, Junpei Ikezoe, Tsuneo Matsumoto, Takeshi Kobayashi, Kenichi Komatsu, Mitate Matsui, Hiroshi Fujita, Yoshie Kodera, and Kunio Doi. Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American journal of roentgenology, 174(1):71-74, 2000. 6 +[34] Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. Vrp-sam: Sam with visual reference prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23565-23574, 2024. 2, 3 +[35] Chenglong Wang, Dexuan Li, Sucheng Wang, Chengxiu Zhang, Yida Wang, Yun Liu, and Guang Yang. SAMMed: A medical image annotation framework based on large vision model. arXiv preprint arXiv:2307.05617, 2023. 3 +[36] Jiaqi Wang, Zhengliang Liu, Lin Zhao, Zihao Wu, Chong Ma, Sigang Yu, Haixing Dai, Qiushi Yang, Yiheng Liu, Songyao Zhang, et al. Review of large vision models and visual prompt engineering. MetaRadiology, page 100047, 2023. 3 +[37] Rui Wang, Zhouwang Yang, and Yanzhi Song. Osamfundus: A training-free, one-shot segmentation framework for optic disc and cup in fundus images. Biomedical Signal Processing and Control, 100:107069, 2025. 2, 3 +[38] Xu Wang, Sen Wang, Xingxing Liang, Dawei Zhao, Jincai Huang, Xin Xu, Bin Dai, and Qiguang Miao. Deep reinforcement learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 35 (4):5064-5078, 2022. 4 +[39] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1130-1140, 2023. 8 + +[40] Bin Xie, Hao Tang, Bin Duan, Dawen Cai, and Yan Yan. Masksam: Towards auto-prompt sam with mask classification for medical image segmentation. arXiv preprint arXiv:2403.14103, 2024. 3 +[41] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 3 +[42] Chaoning Zhang, Fachrina Dewi Puspitasari, Sheng Zheng, Chenghao Li, Yu Qiao, Taegoo Kang, Xinru Shan, Chenshuang Zhang, Caiyan Qin, Francois Rameau, et al. A survey on segment anything model (sam): Vision foundation model meets prompt engineering. arXiv preprint arXiv:2306.06211, 2023. 3 +[43] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 2, 3, 8 +[44] Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, et al. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419, 2023. 2 +[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3 \ No newline at end of file diff --git a/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/images.zip b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..235be151e31a4c39cbcd87ffd9b6a40ec1d35ad3 --- /dev/null +++ b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374b3174f5d50c398bd90999c495345a4b87bac6f975cfbf20955540e45df20a +size 644842 diff --git a/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/layout.json b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e2ff60dc15036aede4464a1c5f0931c1fa58e807 --- /dev/null +++ b/2025/Plug-and-Play PPO_ An Adaptive Point Prompt Optimizer Making SAM Greater/layout.json @@ -0,0 +1,10762 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "text", + "content": "Plug-and-Play PPO: An Adaptive Point Prompt Optimizer Making SAM Greater" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "spans": [ + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "text", + "content": "Xueyu Liu" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "text", + "content": ", Rui Wang" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "text", + "content": ", Yexin Lai" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "text", + "content": ", Guangze Shi" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "text", + "content": ", Feixue Shao" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "text", + "content": ", Fang Hao" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "text", + "content": ", Jianan Zhang" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "text", + "content": ", Jia Shen" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "text", + "content": ", Yongfei Wu" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "text", + "content": ", Wen Zheng" + }, + { + "bbox": [ + 111, + 142, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 220, + 171, + 391, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 171, + 391, + 186 + ], + "spans": [ + { + "bbox": [ + 220, + 171, + 391, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 220, + 171, + 391, + 186 + ], + "type": "text", + "content": "Taiyuan University of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 188, + 186, + 423, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 186, + 423, + 200 + ], + "spans": [ + { + "bbox": [ + 188, + 186, + 423, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 188, + 186, + 423, + 200 + ], + "type": "text", + "content": "University of Science and Technology of China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 251, + 297, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 636 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 636 + ], + "type": "text", + "content": "Powered by extensive curated training data, the Segment Anything Model (SAM) demonstrates impressive generalization capabilities in open-world scenarios, effectively guided by user-provided prompts. However, the class-agnostic characteristic of SAM renders its segmentation accuracy highly dependent on prompt quality. In this paper, we propose a novel plug-and-play dual-space Point Prompt Optimizer (PPO) designed to enhance prompt distribution through deep reinforcement learning (DRL)-based heterogeneous graph optimization. PPO optimizes initial prompts for any task without requiring additional training, thereby improving SAM's downstream segmentation performance. Specifically, PPO constructs a dual-space heterogeneous graph, leveraging the robust feature-matching capabilities of a foundational pre-trained model to create internal feature and physical distance matrices. A DRL policy network iteratively refines the distribution of prompt points, optimizing segmentation predictions. We conducted experiments on four public datasets. The ablation study explores the necessity and balance of optimizing prompts in both feature and physical spaces. The comparative study shows that PPO enables SAM to surpass recent one-shot methods. Additionally, experiments with different initial prompts demonstrate PPO's generality across prompts generated by various methods. In conclusion, PPO redefines the prompt optimization problem as a heterogeneous graph optimization task, using DRL to construct an effective, plug-and-play prompt optimizer. This approach holds potential for broader applications across diverse segmentation tasks and provides a promising solution for point prompt optimization. The source code and demo are available at https://github.com/XueyuLiu/PPO." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 655, + 135, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 135, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 135, + 667 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 670, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 296, + 694 + ], + "type": "text", + "content": "Segmentation, as a fundamental task, plays a pivotal role in visual understanding systems. The prevailing train-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 224, + 547, + 399 + ], + "blocks": [ + { + "bbox": [ + 320, + 224, + 547, + 399 + ], + "lines": [ + { + "bbox": [ + 320, + 224, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 320, + 224, + 547, + 399 + ], + "type": "image", + "image_path": "acd48d42f961ae46fe3db906271ce5cafb1374b8a88812eab8054f7e2fbed250.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 407, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. Ineffective prompts can diminish SAM's performance. Incorrect prompts may cause the segmentation mask to capture irrelevant regions, while excessive prompts for adjacent areas can lead to neglect of other parts of the object. Therefore, prompt optimization is essential to improve SAM's performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 485, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 555, + 557 + ], + "type": "text", + "content": "ing paradigm for segmentation relies on data-driven approaches, where large datasets are used to train models. While this method has achieved remarkable success, it faces several challenges, including the computational resources to train models for new tasks and the risk of domain-specific optimizations that can limit generalizability [18, 22]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 558, + 556, + 715 + ], + "type": "text", + "content": "Recent breakthroughs in Pretrained Foundation Models (PFMs) in computer vision (CV) have demonstrated impressive capabilities in domain-invariant segmentation across diverse scenarios [11, 29]. Vision Foundation Models (VFMs), such as DINOv2 [2, 27], acquire generalized visual features by capturing intricate information at both the image and pixel levels using only raw image data. This approach enables these models to learn generic visual representations, ensuring robust zero-shot transferability for various downstream tasks. Meanwhile, the Segment Anything Model (SAM) [14, 30], trained on large-scale, high-quality datasets, has shown a strong ability to generalize across tasks and data distributions that are not encountered during" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 203, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 203, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 203, + 713 + ], + "type": "text", + "content": "*Corresponding author. †Co-first author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4332" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "training. However, SAM relies heavily on manual prompts, such as bounding boxes and points, to identify objects of interest, as it is pretrained in a class-agnostic manner. Consequently, achieving task-specific adaptation of SAM is crucial to bridge the gap between its class-agnostic predictions and the specific needs of industrial applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 147, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 295, + 361 + ], + "type": "text", + "content": "Some works attempt to train prompt encoders that enables SAM to adapt to downstream segmentation tasks [8, 15, 34, 43]. However, these task-specific prompt encoders require additional training for new segmentation scenarios. Recently, other efforts have focused on automatically generating geometric visual prompts, such as point prompts, to enhance SAM's segmentation performance [21, 23, 37]. In such methods, the quality of prompt schemes remains a critical factor influencing downstream segmentation performance. As illustrated in Fig. 1, incorrect prompts can result in erroneous segmentation, while excessive prompts in specific areas may cause SAM to focus disproportionately on those regions, neglecting other parts of the target. Current automatic prompt engineering methods often lack optimization in prompt distribution, leading to lower-quality prompts and segmentation performance. Therefore, there is an urgent need for methods that can refine prompts to effectively enhance SAM's performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 364, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 495 + ], + "type": "text", + "content": "To address this, effective point prompts for SAM should exhibit certain characteristics in both feature and physical spaces. In the feature space, regions containing positive prompts should be homogeneous to ensure consistency and clarity within the segmentation target, while regions between positive and negative prompts should be heterogeneous, distinguishing the target from the background. In the physical space, positive and negative prompts should be sufficiently spaced apart to mitigate the effects of excessive prompts and prevent over-segmentation, while also being close enough to boundaries to ensure precise delineation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "Building upon the aforementioned inspiration and committed to addressing the challenges mentioned above, we propose a novel Plug-and-Play Point Prompt Optimizer (PPO) designed to enhance SAM's segmentation performance by automatically optimizing initial point prompt schemes. Leveraging DINOv2's powerful feature association capabilities [27], we construct an internal feature and physical distance matrix within each image to create a heterogeneous graph environment. To optimize prompt distribution in dual spaces, we employ Deep Reinforcement Learning (DRL), a framework that trains agents to maximize cumulative rewards through interaction and feedback mechanisms. By designing specific actions and reward functions, PPO optimizes prompt point distribution under the guidance of DRL and inputs them into SAM to obtain refined segmentation results. We validated our approach through experiments on two natural and two medical image datasets. Ablation studies demonstrate that PPO" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "optimized prompt distributions significantly enhance segmentation performance. Comparative experiments show that PPO outperforms recent methods that have not been trained on in-domain datasets. Additionally, experiments with different initial prompts confirm PPO's generality across prompts generated by various methods. The main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 158, + 553, + 289 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 158, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 158, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 314, + 158, + 553, + 217 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to define an optimization direction for point prompts in both feature and physical spaces. Additionally, we are the first to use DRL to optimize point prompts, achieving prompt optimization without additional training for new tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 218, + 553, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 218, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 314, + 218, + 553, + 254 + ], + "type": "text", + "content": "- PPO operates in a plug-and-play manner, allowing for the optimization of any initial geometric point prompts to improve SAM's segmentation performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 254, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 254, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 254, + 553, + 289 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the effectiveness of PPO in optimizing point prompts across various initial prompt schemes." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 307, + 400, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 307, + 400, + 319 + ], + "spans": [ + { + "bbox": [ + 314, + 307, + 400, + 319 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 323, + 481, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 481, + 335 + ], + "type": "text", + "content": "2.1. Pretrained Foundation Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 338, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 482 + ], + "type": "text", + "content": "In natural language processing (NLP), PFMs, such as BERT [5], ChatGPT, and GPT-4, have become foundational tools for a wide range of downstream tasks across diverse data modalities. By leveraging extensive text data, PFMs provide effective parameter initialization for various NLP applications [44]. The success of PFMs in NLP has inspired researchers to explore their application in CV. Visual language models like CLIP [29] and GLIP [19], trained on large-scale datasets of millions of image-text pairs from the web, exhibit strong zero-shot generalization across multiple vision tasks, offering representations with robust generalization and transfer capabilities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 604 + ], + "type": "text", + "content": "Building on this trend, Meta AI introduced the SAM, a notable advancement in applying VFMs. SAM allows users to specify image segmentation tasks through prompts and demonstrates impressive zero-shot learning, enabling it to generalize effectively across diverse tasks without task-specific training [14]. However, SAM's segmentation performance heavily relies on the quality of input prompts. To address this, we propose an adaptive point prompt optimizer that automatically refines prompts to enhance SAM's segmentation performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": "In another significant development, DINOv2 [27] leverages discriminative self-supervised learning at both image and patch levels, showing exceptional patch-matching capabilities that allow it to capture semantic similarities across images of the same target. By utilizing these patch-matching capabilities, we can extract feature distances between patches within an image, providing a theoretical foundation for constructing heterogeneous graphs to achieve effective prompt optimization." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4333" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 226, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 226, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 226, + 85 + ], + "type": "text", + "content": "2.2. Automatic Prompt Engineering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 301 + ], + "type": "text", + "content": "In the field of NLP, extensive researches have been conducted on designing automatic prompt engineering, with prompt tuning methods achieving significant advances in enhancing PFMs' performance through optimized language prompts [1, 12, 13, 16, 28, 41, 45, 45]. Simultaneously, efforts have been made to explore automatic prompt engineering in CV. On the one hand, some recent works have attempted to train prompt encoders that enable SAM to adapt to downstream segmentation tasks [9, 32, 34, 40]. For example, Sun et al. proposed VRP-SAM, an extension of SAM that encodes various granularities of visual references directly into prompt embeddings for mask generation [34]. These approaches encode SAM's prompts into prompt embeddings via a prompt encoder, which are then fed into SAM's mask decoder to generate target masks. However, these methods are often task-sensitive, requiring training different prompt encoders for different tasks, thus limiting rapid task transfer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 304, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 483 + ], + "type": "text", + "content": "On the other hand, some works attempt to generate geometric prompts to achieve automatic prompt engineering for SAM, generally following two main paradigms. The first paradigm uses a small amount of annotated data to train a segmentation model, then samples prompts from pseudolabels generated by this model [3, 26, 36, 42, 43]. For instance, Li et al. randomly selected prompt points from masks generated by coarse segmentation, which were subsequently input into SAM after fine-tuning to obtain segmentation results [17]. Wang et al. utilized a novel prototype learning framework to generate pseudo-labels and extract prompts, which were then incorporated into SAM for segmentation [35]. Dai et al. developed various methods to expand prompt points, exploring the potential of a single prompt point for enhanced segmentation performance [4]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 484, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 700 + ], + "type": "text", + "content": "However, this first paradigm still requires a small amount of data to train the prompt generator, also limiting task transferability. To improve transferability, the second paradigm generates initial prompts by leveraging feature similarity between reference and target images without additional training [21, 23, 37, 43]. For example, Zhang et al. calculated feature similarity between pixels in the target and reference images to create a point prompt scheme using only a single reference image [43]. Liu et al. introduced Matcher, an instance-level and bidirectional matching strategy for prompt point generation [23], and Liu et al. proposed a series of prompt optimization strategies for the automatic segmentation of basement membranes in electron microscopy images [21]. However, the quality of prompts generated by these methods is often limited, as they do not sufficiently consider the relationships between positive and negative prompts, resulting in redundant and erroneous prompts that negatively affect segmentation performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "In our work, PPO addresses this limitation by con" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": "structuring a heterogeneous graph based on dual-space distances between initial prompts within each image. This approach optimizes any automatically generated geometric point prompts without additional training for new tasks, enhancing SAM's segmentation performance while preserving task transferability by ensuring high prompt quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 156, + 397, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 397, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 397, + 169 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 171, + 555, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 362 + ], + "type": "text", + "content": "This section introduces our plug-and-play PPO, which is based on a heterogeneous graph structure designed to enhance the segmentation performance of SAM by optimizing initial point prompts. An overview of PPO is illustrated in Fig. 2. Specifically, we first compute the feature distance matrix and the physical distance matrix for each image. Using these matrices, we construct a heterogeneous graph where the edge attributes are defined by the feature and physical distances based on the initial prompt indices. During the training phase, we use this heterogeneous graph as the environment to design an action list and reward function to train the optimization agent. In the inference phase, the trained agent activates the appropriate point prompts as a final prompt scheme, which is then input into SAM to obtain the final segmentation result. In the following subsections, we will describe the main components in detail." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 382, + 527, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 527, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 527, + 395 + ], + "type": "text", + "content": "3.1. Dual-Space Distance Matrix Calculation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "text", + "content": "The core objective of our method is to optimize positive and negative point prompts in both feature and distance spaces. To achieve this, we construct a heterogeneous graph comprising two homogeneous graphs that represent the feature and distance spaces. Edge values are determined by creating two patch-level distance matrices for each image " + }, + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 396, + 554, + 468 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "text", + "content": "First, we divide " + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "text", + "content": " into patches " + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "text", + "content": " using a sliding window approach, where " + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "inline_equation", + "content": "x_{i} = x_{1},x_{2},\\ldots ,x_{n}" + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "text", + "content": ". We then use the image encoder from DINOv2 [27] to extract patch-level features for " + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "text", + "content": ", represented by " + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "inline_equation", + "content": "f_{i} = f_{1},f_{2},\\dots ,f_{n}" + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "text", + "content": ". Next, we calculate the feature matrix " + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "inline_equation", + "content": "M_{f}" + }, + { + "bbox": [ + 313, + 468, + 554, + 529 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 386, + 539, + 555, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 539, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 386, + 539, + 555, + 553 + ], + "type": "interline_equation", + "content": "M _ {f} (i, j) = \\left\\| f _ {i} - f _ {j} \\right\\|, \\tag {1}", + "image_path": "94b271e2d0e0c8bdaf11fc859a48d9ada4ce2554fdb7930fbbce5c9c800239d9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "inline_equation", + "content": "M_{f}(i,j)" + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "text", + "content": " represents the Euclidean distance " + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "text", + "content": " between the " + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "text", + "content": "-th patch feature " + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "text", + "content": "-th patch feature " + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "text", + "content": ". Similarly, we calculate the physical distance matrix " + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "inline_equation", + "content": "M_{p}" + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "text", + "content": " using the center points of the patches " + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 561, + 554, + 611 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 386, + 620, + 555, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 620, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 386, + 620, + 555, + 634 + ], + "type": "interline_equation", + "content": "M _ {p} (i, j) = \\left\\| x _ {i} - x _ {j} \\right\\|, \\tag {2}", + "image_path": "db6ad3dff4121bb03feba0b305cfc1bf1fb6f8bfb1ffdf61c558c91a07b977e5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "inline_equation", + "content": "M_{p}(i,j)" + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "text", + "content": " represents the Euclidean distance " + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "inline_equation", + "content": "D_{p}" + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "text", + "content": " between the " + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "text", + "content": "-th patch center " + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "text", + "content": "-th patch center " + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 313, + 643, + 554, + 668 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 675, + 477, + 688 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 675, + 477, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 675, + 477, + 688 + ], + "type": "text", + "content": "3.2. Graph Environment Creation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Inspired by coarse segmentation sampling [3, 26] and feature matching techniques [21, 23], we can obtain initial" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4334" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 346 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 346 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 346 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 346 + ], + "type": "image", + "image_path": "a56ba54ca5273f651e41db2d47198126b844b525005abab8f95dd85e2bbb526e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 356, + 555, + 391 + ], + "lines": [ + { + "bbox": [ + 54, + 356, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 555, + 391 + ], + "type": "text", + "content": "Figure 2. The PPO workflow is designed to enhance SAM's segmentation performance by optimizing initial point prompts through two key components: graph-based environment creation and agent training for prompt optimization. After optimized, the prompts are fed into SAM to obtain segmentation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "text", + "content": "point prompt indices for optimization. Specifically, we define a heterogeneous graph " + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "inline_equation", + "content": "G = (V,E,\\phi ,\\psi)" + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "text", + "content": " is the set of nodes and " + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "text", + "content": " is the set of edges. The node type set " + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\{\\text{positive nodes } V_p, \\text{negative nodes } V_n\\}" + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "text", + "content": " is represented by the mapping function " + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\phi (v): V \\to \\mathcal{A}" + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "text", + "content": ", and the edge type set " + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{\\text{feature edges } E_f, \\text{physical edges } E_p\\}" + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "text", + "content": " is represented by the mapping function " + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\psi (e): E \\to \\mathcal{R}" + }, + { + "bbox": [ + 54, + 411, + 294, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 495, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 532 + ], + "type": "text", + "content": "We use the heterogeneous graph " + }, + { + "bbox": [ + 55, + 495, + 295, + 532 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 495, + 295, + 532 + ], + "type": "text", + "content": " as the environment for DRL. By selecting different nodes, we can obtain different edge values as feedback." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 270, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 270, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 270, + 556 + ], + "type": "text", + "content": "3.3. Agent Training for Prompt Optimization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "content": "On the one hand, an effective point prompt scheme should ensure correctness. In an image, the features of the segmentation target are often homogeneous, while the features of the background are heterogeneous. Therefore, we aim for the feature distance between positive point prompts to be sufficiently close, and the feature distance between positive and negative point prompts to be sufficiently far. On the other hand, an effective point prompt scheme should also ensure sparsity. Over-prompting in adjacent areas can degrade segmentation performance. Thus, the physical distance between positive point prompts (and between negative point prompts) should not be too close. Additionally, to enhance boundary segmentation accuracy, the physical dis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 411, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 434 + ], + "type": "text", + "content": "tance between positive and negative point prompts should not be too far." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 435, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 520 + ], + "type": "text", + "content": "A DRL model can infer actions and receive rewards from the environment, reflecting the significance of these actions and updating the agent's state accordingly [38]. Using DRL, we aim to train an agent to activate appropriate point prompts to optimize the distribution. The detailed process is shown in Algorithm 1. The key components of agent training for prompt optimization will be introduced as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": "Actions: We define three possible actions: (1) restore node " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "A_r" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": ", which randomly activates a node and connects the corresponding edges; (2) delete node " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "A_d" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": ", which randomly deletes a node and removes the corresponding edges; (3) move node " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "A_m" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": ", which randomly selects a direction and moves the index position of a node. Additionally, we aim to maintain a stable number of prompts. Therefore, if the number of nodes " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "|V_p|" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "|V_n|" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": " in the current state is less than the minimum value, the only possible action is " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "A_r" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": ". Conversely, if the " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "|V_p|" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "|V_n|" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": " exceeds the maximum value, the only possible action is " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "A_d" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": ". Only when the " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "|V_p|" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "|V_n|" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": " is between the minimum and maximum values are all three actions, " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "A_r" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "A_d" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "inline_equation", + "content": "A_m" + }, + { + "bbox": [ + 313, + 533, + 556, + 688 + ], + "type": "text", + "content": ", possible." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "States: For each action, there will be different nodes, resulting in a different heterogeneous graph " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ". Based on" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4335" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 198, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 198, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 198, + 84 + ], + "type": "text", + "content": "Algorithm 1 PPO training process" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 86, + 237, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 86, + 237, + 97 + ], + "spans": [ + { + "bbox": [ + 56, + 86, + 237, + 97 + ], + "type": "text", + "content": "Input: Heterogeneous graph environment " + }, + { + "bbox": [ + 56, + 86, + 237, + 97 + ], + "type": "inline_equation", + "content": "G" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 98, + 200, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 98, + 200, + 109 + ], + "spans": [ + { + "bbox": [ + 56, + 98, + 200, + 109 + ], + "type": "text", + "content": "Parameter: Max step; Max epoch;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 110, + 202, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 202, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 202, + 121 + ], + "type": "text", + "content": "Max nodes " + }, + { + "bbox": [ + 56, + 110, + 202, + 121 + ], + "type": "inline_equation", + "content": "N_{max}" + }, + { + "bbox": [ + 56, + 110, + 202, + 121 + ], + "type": "text", + "content": "; Min nodes " + }, + { + "bbox": [ + 56, + 110, + 202, + 121 + ], + "type": "inline_equation", + "content": "N_{min}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 122, + 132, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 122, + 132, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 122, + 132, + 133 + ], + "type": "text", + "content": "Output: Q-values" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 135, + 294, + 408 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 62, + 135, + 227, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 135, + 227, + 147 + ], + "spans": [ + { + "bbox": [ + 62, + 135, + 227, + 147 + ], + "type": "text", + "content": "1: Initializing the Q-values, Epoch, Step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 148, + 197, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 197, + 159 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 197, + 159 + ], + "type": "text", + "content": "2: while Epoch < Max epoch do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 160, + 167, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 167, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 167, + 171 + ], + "type": "text", + "content": "3: Initializing the Step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 171, + 191, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 191, + 182 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 191, + 182 + ], + "type": "text", + "content": "4: while Step " + }, + { + "bbox": [ + 62, + 171, + 191, + 182 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 62, + 171, + 191, + 182 + ], + "type": "text", + "content": " Max step do" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 184, + 217, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 217, + 194 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 217, + 194 + ], + "type": "text", + "content": "5: Emptying all nodes and edges." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 196, + 167, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 196, + 167, + 207 + ], + "spans": [ + { + "bbox": [ + 62, + 196, + 167, + 207 + ], + "type": "text", + "content": "6: Emptying reward." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 208, + 294, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 208, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 62, + 208, + 294, + 220 + ], + "type": "text", + "content": "7: Calculating the number of Nodes " + }, + { + "bbox": [ + 62, + 208, + 294, + 220 + ], + "type": "inline_equation", + "content": "N_{pos}" + }, + { + "bbox": [ + 62, + 208, + 294, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 62, + 208, + 294, + 220 + ], + "type": "inline_equation", + "content": "N_{neg}" + }, + { + "bbox": [ + 62, + 208, + 294, + 220 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 220, + 253, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 253, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 253, + 232 + ], + "type": "text", + "content": "8: if " + }, + { + "bbox": [ + 62, + 220, + 253, + 232 + ], + "type": "inline_equation", + "content": "N_{pos} < N_{min}" + }, + { + "bbox": [ + 62, + 220, + 253, + 232 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 62, + 220, + 253, + 232 + ], + "type": "inline_equation", + "content": "N_{neg} < N_{min}" + }, + { + "bbox": [ + 62, + 220, + 253, + 232 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 232, + 176, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 176, + 243 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 176, + 243 + ], + "type": "text", + "content": "9: Action list " + }, + { + "bbox": [ + 62, + 232, + 176, + 243 + ], + "type": "inline_equation", + "content": "= \\{A_{r}\\}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 243, + 274, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 274, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 274, + 255 + ], + "type": "text", + "content": "10: else if " + }, + { + "bbox": [ + 61, + 243, + 274, + 255 + ], + "type": "inline_equation", + "content": "N_{pos} > N_{max}" + }, + { + "bbox": [ + 61, + 243, + 274, + 255 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 61, + 243, + 274, + 255 + ], + "type": "inline_equation", + "content": "N_{neg} > N_{max}" + }, + { + "bbox": [ + 61, + 243, + 274, + 255 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 255, + 177, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 255, + 177, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 255, + 177, + 266 + ], + "type": "text", + "content": "11: Action list " + }, + { + "bbox": [ + 61, + 255, + 177, + 266 + ], + "type": "inline_equation", + "content": "= \\{A_{d}\\}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 61, + 268, + 111, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 268, + 111, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 268, + 111, + 277 + ], + "type": "text", + "content": "12: else" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 61, + 279, + 212, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 279, + 212, + 291 + ], + "spans": [ + { + "bbox": [ + 61, + 279, + 212, + 291 + ], + "type": "text", + "content": "13: Action list " + }, + { + "bbox": [ + 61, + 279, + 212, + 291 + ], + "type": "inline_equation", + "content": "= \\{A_r, A_d, A_m\\}" + }, + { + "bbox": [ + 61, + 279, + 212, + 291 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 61, + 292, + 120, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 292, + 120, + 301 + ], + "spans": [ + { + "bbox": [ + 61, + 292, + 120, + 301 + ], + "type": "text", + "content": "14: end if" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 61, + 303, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 303, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 61, + 303, + 287, + 314 + ], + "type": "text", + "content": "15: Perform a random action and update the status " + }, + { + "bbox": [ + 61, + 303, + 287, + 314 + ], + "type": "inline_equation", + "content": "s" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 59, + 315, + 294, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 315, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 59, + 315, + 294, + 337 + ], + "type": "text", + "content": "16: Adding the reward " + }, + { + "bbox": [ + 59, + 315, + 294, + 337 + ], + "type": "inline_equation", + "content": "r_s" + }, + { + "bbox": [ + 59, + 315, + 294, + 337 + ], + "type": "text", + "content": " based on the reward function." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 59, + 338, + 175, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 338, + 175, + 350 + ], + "spans": [ + { + "bbox": [ + 59, + 338, + 175, + 350 + ], + "type": "text", + "content": "17: Updating Q-values." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 59, + 351, + 131, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 351, + 131, + 361 + ], + "spans": [ + { + "bbox": [ + 59, + 351, + 131, + 361 + ], + "type": "text", + "content": "18: Step+=1." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 59, + 363, + 126, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 363, + 126, + 373 + ], + "spans": [ + { + "bbox": [ + 59, + 363, + 126, + 373 + ], + "type": "text", + "content": "19: end while" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 59, + 375, + 129, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 375, + 129, + 385 + ], + "spans": [ + { + "bbox": [ + 59, + 375, + 129, + 385 + ], + "type": "text", + "content": "20: Epoch " + }, + { + "bbox": [ + 59, + 375, + 129, + 385 + ], + "type": "inline_equation", + "content": "+ = 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 59, + 387, + 116, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 387, + 116, + 397 + ], + "spans": [ + { + "bbox": [ + 59, + 387, + 116, + 397 + ], + "type": "text", + "content": "21: end while" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 59, + 399, + 137, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 399, + 137, + 408 + ], + "spans": [ + { + "bbox": [ + 59, + 399, + 137, + 408 + ], + "type": "text", + "content": "22: return solution" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 433, + 272, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 272, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 272, + 445 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 55, + 433, + 272, + 445 + ], + "type": "text", + "content": ", we can calculate the following mean edge values:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 93, + 453, + 295, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 453, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 93, + 453, + 295, + 484 + ], + "type": "interline_equation", + "content": "D _ {f} ^ {s} (p, p) = \\frac {1}{| V _ {p} | | V _ {p} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {p}} M _ {f} (i, j), \\tag {3}", + "image_path": "61cfa4b49c5e79fc362b454451c50004d13c5fc438a4b28b5245fa824520cc78.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 92, + 495, + 295, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 495, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 92, + 495, + 295, + 525 + ], + "type": "interline_equation", + "content": "D _ {f} ^ {\\mathrm {s}} (p, n) = \\frac {1}{| V _ {p} | | V _ {n} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {n}} M _ {f} (i, j), \\tag {4}", + "image_path": "bcbcdc68ca676d1170f40671ec3b19464e78ed0d4896e5a1a45ea57c0ab530a0.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 94, + 531, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 531, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 94, + 531, + 295, + 563 + ], + "type": "interline_equation", + "content": "D _ {p} ^ {s} (p, p) = \\frac {1}{| V _ {p} | | V _ {p} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {p}} M _ {p} (i, j), \\tag {5}", + "image_path": "e0d650defbda743b2e21c56d9c50d8901e0ccd5dcd50027bf40c3d98a54dbe71.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 92, + 569, + 295, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 569, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 92, + 569, + 295, + 599 + ], + "type": "interline_equation", + "content": "D _ {p} ^ {s} (n, n) = \\frac {1}{| V _ {n} | | V _ {n} |} \\sum_ {i \\in V _ {n}} \\sum_ {j \\in V _ {n}} M _ {p} (i, j), \\tag {6}", + "image_path": "04f1f41328ba1e2df1742aee88cd4cfa8c35391c666d51952d318e22d54e4b8b.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 93, + 605, + 295, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 605, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 93, + 605, + 295, + 635 + ], + "type": "interline_equation", + "content": "D _ {p} ^ {s} (p, n) = \\frac {1}{| V _ {p} | | V _ {n} |} \\sum_ {i \\in V _ {p}} \\sum_ {j \\in V _ {n}} M _ {p} (i, j). \\tag {7}", + "image_path": "94707029ab17f26492f8ac9931e9c68987f90fedf9809bac2f4807b29f4eba70.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "In state " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{f}^{s}(p,p)" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{f}^{s}(p,n)" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " represents the mean feature distance between positive prompts, and between positive and negative prompts. Similarly, " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{p}^{s}(p,p)" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{p}^{s}(n,n)" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{p}^{s}(p,n)" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " represent the mean physical distances in state " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " between positive prompts, negative prompts, and between positive and negative prompts, respectively." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "content": "Rewards: Based on the various distances obtained from " + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "inline_equation", + "content": "G^{s}" + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "content": ", we can set a reward function according to the optimization objectives. The reward function for the current state, " + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "inline_equation", + "content": "r_{s}" + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "content": ", is defined based on the changes in distance from the previous state:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 361, + 137, + 553, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 137, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 361, + 137, + 553, + 220 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} r _ {s} = - \\alpha \\left(D _ {f} ^ {s} (p, p) - D _ {f} ^ {s - 1} (p, p)\\right) \\\\ + \\alpha \\left(D _ {f} ^ {s} (p, n) - D _ {f} ^ {s - 1} (p, n)\\right) \\\\ + \\beta \\left(D _ {p} ^ {s} (p, p) - D _ {p} ^ {s - 1} (p, p)\\right) \\tag {8} \\\\ + \\beta (D _ {p} ^ {s} (n, n) - D _ {p} ^ {s - 1} (n, n)) \\\\ - \\beta (D _ {p} ^ {s} (p, n) - D _ {p} ^ {s - 1} (p, n)), \\\\ \\end{array}", + "image_path": "e4b0631aadbae72b2566d6427c448ac002094fcad652262890bc2f9ee148f467.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "text", + "content": "where coefficients " + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "text", + "content": " balance the feature and physical distance, respectively. Using this reward function, if the current state's " + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "inline_equation", + "content": "D_{f}^{s}(p,p)" + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "inline_equation", + "content": "D_{p}^{s}(p,n)" + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "text", + "content": " increase, a penalty is applied. Conversely, if the current state's " + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "inline_equation", + "content": "D_{f}^{s}(p,n)" + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "inline_equation", + "content": "D_{p}^{s}(p,p)" + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "inline_equation", + "content": "D_{p}^{s}(n,n)" + }, + { + "bbox": [ + 313, + 224, + 554, + 285 + ], + "type": "text", + "content": " increase, a reward is given." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "text", + "content": "Policy Network: We adopt the Q-network [25] as the policy network, which consists of two fully connected layers with 1024 and 64 neurons, respectively, and a single output layer with 3 neurons. This output layer predicts the Q-value of 3 actions based on the current state. The policy network is regarded as the approximate action-value function " + }, + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "inline_equation", + "content": "Q_{s}" + }, + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "text", + "content": " and takes the state " + }, + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "text", + "content": " as its input to approximate the underlying action-value function. This approximation is achieved by repeatedly performing a gradient descent step on " + }, + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "inline_equation", + "content": "(y_{s} - Q_{s})^{2}" + }, + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "text", + "content": " with respect to the policy network parameters at each step. The Q-value " + }, + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "inline_equation", + "content": "y_{s}" + }, + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 351, + 423, + 553, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 423, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 351, + 423, + 553, + 455 + ], + "type": "interline_equation", + "content": "y _ {s} = \\left\\{ \\begin{array}{l l} r _ {s} + \\gamma \\max _ {s + 1} Q _ {s + 1}, & \\text {i f n o t s t o p} \\\\ r _ {s}, & \\text {i f s t o p} \\end{array} , \\right. \\tag {9}", + "image_path": "bb8cd130c0409201f1d39033bcc5463145e1f115e2bd635a2576fe2bf2062686.jpg" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 462, + 553, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 484 + ], + "type": "text", + "content": "where the discount factor " + }, + { + "bbox": [ + 313, + 462, + 553, + 484 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 462, + 553, + 484 + ], + "type": "text", + "content": " represents the balance between the current and future rewards." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 594 + ], + "type": "text", + "content": "In the training process of PPO, we initially delete all nodes and treat them as potential activatable nodes in the action environment. As the agent attempts to activate different nodes, the edges of the heterogeneous graph acquire different values. Based on the reward function, the agent receives rewards or penalties and updates the Q-values accordingly. The policy network repeatedly interacts with the segmentation network, updates the network weights according to Eq. (11), and ultimately learns effective policies." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 602, + 440, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 440, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 440, + 615 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 616, + 447, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 447, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 447, + 628 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Our experiments are carried out on a Linux server platform equipped with an NVIDIA Tesla V100. During the PPO training phase, we set the step size for each iteration to 100 and the number of training epochs to 1000. The discount factor " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " of Q-learning is set to 0.1. We use the " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "-greedy strategy to balance exploration and exploitation [25]. During " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "-greedy training, the agent can select a random action" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4336" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "with probability " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " and a predicted action generated by the Q-network with probability " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " parameters are set as follows: " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{start}} = 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{end}} = 0.1" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{decay}} = 0.995" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ". Additionally, to ensure stability during PPO training, the probability of action " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "A_{m}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is always set to be 0.5 times the probability of the other actions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 294, + 216 + ], + "type": "text", + "content": "In the testing phase, PPO requires no additional training for new tasks; it simply inputs initial prompts and images to output optimized prompt schemes that serve SAM. For evaluation, we use the mean Dice Similarity Coefficient (mDSC) and the mean Hausdorff Distance (mHD) as metrics to assess the segmentation performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 224, + 170, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 170, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 170, + 236 + ], + "type": "text", + "content": "4.2. Dataset Description" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 237, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 295, + 453 + ], + "type": "text", + "content": "To ensure that PPO remains unseen to in-domain datasets for downstream tasks, we trained PPO on the GBMSD dataset [21], generating initial prompts through feature matching and constructing a heterogeneous graph as the environment for DRL learning. During testing, we evaluated the model on natural image domains using the FSS-1000 [20] and VOC2012 [6] datasets. The FSS-1000 dataset comprises mask-annotated images from 1,000 classes, with 10 images per class, while the VOC2012 dataset is a widely-used benchmark for semantic segmentation tasks, containing 20 object classes and a total of 4,396 images. For medical image domains, we used the ISIC2018 [33] and Kvasir [10]. The ISIC2018 is a large-scale collection of dermatoscopic images focused on lesion segmentation, consisting of 2,594 images. The Kvasir, on the other hand, is an endoscopic dataset aimed at colorectal polyp segmentation, containing 1,000 gastroenterology images of colorectal polyps along with their corresponding segmentation masks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 295, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 596 + ], + "type": "text", + "content": "We generated initial prompts based on two mainstream point prompt-generation paradigms. The first paradigm is the feature matching (FM) approach [21, 23], which requires no additional training for new tasks. In this approach, only one annotated reference image is selected for each segmentation task, and this image is used to generate the initial prompt for testing images. The second paradigm is the coarse segmentation sampling (CS) approach, which involves few-shot training for new tasks. For each task, we selected 20 annotated images, trained a U-Net [31] with 200 epochs, and used the resulting coarse segmentation masks to sample initial prompts for the segmentation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 604, + 149, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 149, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 149, + 616 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "We conduct an ablation study on the initial prompts generated by FM to evaluate the segmentation performance of SAM with PPO under varying parameter settings. Quantitative results are presented in Table 1. Initially, when SAM applies unoptimized prompts " + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "(\\alpha = 0, \\beta = 0)" + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": ", segmentation outcomes are suboptimal across all datasets. In contrast, applying PPO with optimized settings significantly enhances prompt quality, leading to notable performance gains for" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 423, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 423, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 423, + 83 + ], + "type": "text", + "content": "SAM across various tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 84, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 167 + ], + "type": "text", + "content": "When we examine optimization within the feature space " + }, + { + "bbox": [ + 313, + 84, + 553, + 167 + ], + "type": "inline_equation", + "content": "(\\alpha = 1, \\beta = 0)" + }, + { + "bbox": [ + 313, + 84, + 553, + 167 + ], + "type": "text", + "content": " and physical space " + }, + { + "bbox": [ + 313, + 84, + 553, + 167 + ], + "type": "inline_equation", + "content": "(\\alpha = 0, \\beta = 1)" + }, + { + "bbox": [ + 313, + 84, + 553, + 167 + ], + "type": "text", + "content": " separately, both settings improve SAM's performance; however, feature space optimization proves to be more effective, yielding a marked improvement over the initial prompts. The better results are achieved when both feature and physical spaces are optimized concurrently " + }, + { + "bbox": [ + 313, + 84, + 553, + 167 + ], + "type": "inline_equation", + "content": "(\\alpha = 1, \\beta = 1)" + }, + { + "bbox": [ + 313, + 84, + 553, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 168, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 263 + ], + "type": "text", + "content": "Interestingly, further increasing the weight on the physical space " + }, + { + "bbox": [ + 313, + 168, + 553, + 263 + ], + "type": "inline_equation", + "content": "(\\beta = 5)" + }, + { + "bbox": [ + 313, + 168, + 553, + 263 + ], + "type": "text", + "content": " reduces segmentation performance, likely due to an overemphasis on physical distributions at the expense of essential feature space information. In contrast, placing a higher weight on the feature space " + }, + { + "bbox": [ + 313, + 168, + 553, + 263 + ], + "type": "inline_equation", + "content": "(\\alpha = 5, \\beta = 1)" + }, + { + "bbox": [ + 313, + 168, + 553, + 263 + ], + "type": "text", + "content": " yields the best segmentation results. These experiments demonstrate that the accuracy of point prompts is more crucial than sparsity, though sparsity remains beneficial." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 264, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 384 + ], + "type": "text", + "content": "In Fig. 3, we compare SAM's qualitative segmentation results using initial prompts versus PPO-optimized prompts under the best parameter settings. This figure includes images from both natural and medical datasets. The initial prompts contain inaccuracies and excessive redundancy, impairing SAM's segmentation accuracy. After PPO refinement, the prompts become more precise and sparse, allowing the segmentation results to closely align with the ground truth, demonstrating PPO's effectiveness in enhancing SAM's adaptability across diverse segmentation tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 389, + 509, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 509, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 509, + 401 + ], + "type": "text", + "content": "4.4. Comparison with One-Shot Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 403, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 594 + ], + "type": "text", + "content": "Our goal is to enable SAM to adapt to segmentation tasks across diverse scenarios with minimal cost. To this end, we compared recently proposed methods with our PPO-optimized FM prompt generation method, both following the one-shot paradigm, across multiple datasets. As shown in Table 2, FM-PPO achieves competitive performance in various natural image segmentation tasks without training on in-domain datasets, performing comparably to models trained on in-domain data and surpassing other SAM-based prompt generation methods. Additionally, FM-PPO significantly outperforms other segmentation methods on medical datasets without requiring additional training. This demonstrates that PPO-optimized visual prompt generation retains its training-free task transfer capability while ensuring robust segmentation performance for downstream SAM through optimized prompts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Moreover, Fig. 4 illustrates the qualitative results of training-free methods on natural and medical datasets, providing visual evidence of FM-PPO's effectiveness. The experimental results show that FM-PPO performs well even in small target segmentation by reducing erroneous prompt points. This success is attributed to the homogeneous feature expectation within the segmentation target and the heterogeneous feature expectation between the target and background during agent training. For medical image segmentation, incorrect and redundant prompts due to domain" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4337" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 91, + 550, + 198 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 549, + 83 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 549, + 83 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 549, + 83 + ], + "type": "text", + "content": "Table 1. Ablation study on different spatial weights during PPO training. The results with " + }, + { + "bbox": [ + 59, + 70, + 549, + 83 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 59, + 70, + 549, + 83 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 59, + 70, + 549, + 83 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 59, + 70, + 549, + 83 + ], + "type": "text", + "content": " set to 0 represent the initial prompts." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 91, + 550, + 198 + ], + "lines": [ + { + "bbox": [ + 59, + 91, + 550, + 198 + ], + "spans": [ + { + "bbox": [ + 59, + 91, + 550, + 198 + ], + "type": "table", + "html": "
Feature SpacePhysical SpaceFSS-1000VOC2012ISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
α=0β=072.19.934.712.148.59.927.213.2
α=1β=081.38.559.410.561.79.256.610.3
α=0β=173.59.938.411.751.99.930.513.1
α=1β=186.67.863.710.369.28.561.39.6
α=1β=582.48.956.910.264.99.058.89.7
α=5β=191.97.473.79.975.48.170.39.0
", + "image_path": "9566aba06ffd8ad75137ccb13a603ca08241cd4db6523b28518f428e0e66c207.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 69, + 208, + 119, + 258 + ], + "blocks": [ + { + "bbox": [ + 69, + 208, + 119, + 258 + ], + "lines": [ + { + "bbox": [ + 69, + 208, + 119, + 258 + ], + "spans": [ + { + "bbox": [ + 69, + 208, + 119, + 258 + ], + "type": "image", + "image_path": "ab9ce231d488bffebf21739d6fccf11704b565edab3daddb5888770939ade6aa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 260, + 118, + 310 + ], + "blocks": [ + { + "bbox": [ + 69, + 260, + 118, + 310 + ], + "lines": [ + { + "bbox": [ + 69, + 260, + 118, + 310 + ], + "spans": [ + { + "bbox": [ + 69, + 260, + 118, + 310 + ], + "type": "image", + "image_path": "85d9168ecc72029c87886af298d28f91a4e7223ee2337c7dee4ab556352c41b0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 120, + 208, + 170, + 258 + ], + "blocks": [ + { + "bbox": [ + 120, + 208, + 170, + 258 + ], + "lines": [ + { + "bbox": [ + 120, + 208, + 170, + 258 + ], + "spans": [ + { + "bbox": [ + 120, + 208, + 170, + 258 + ], + "type": "image", + "image_path": "f30d2aad89f5edc11dd910da859c4a4e06a7c71f211d71f07cfd1a1fe08e9b57.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 121, + 259, + 170, + 310 + ], + "blocks": [ + { + "bbox": [ + 121, + 259, + 170, + 310 + ], + "lines": [ + { + "bbox": [ + 121, + 259, + 170, + 310 + ], + "spans": [ + { + "bbox": [ + 121, + 259, + 170, + 310 + ], + "type": "image", + "image_path": "258fa199c20e4d80d2c7d4d45fcc677d8dc67956cbb4e945de63ad990d1f9628.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 172, + 208, + 222, + 258 + ], + "blocks": [ + { + "bbox": [ + 172, + 208, + 222, + 258 + ], + "lines": [ + { + "bbox": [ + 172, + 208, + 222, + 258 + ], + "spans": [ + { + "bbox": [ + 172, + 208, + 222, + 258 + ], + "type": "image", + "image_path": "a00f1262ac9a725495e538c9966d0c2b6f42d338baebfc9f7f91444cf81ddb71.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 172, + 259, + 222, + 310 + ], + "blocks": [ + { + "bbox": [ + 172, + 259, + 222, + 310 + ], + "lines": [ + { + "bbox": [ + 172, + 259, + 222, + 310 + ], + "spans": [ + { + "bbox": [ + 172, + 259, + 222, + 310 + ], + "type": "image", + "image_path": "cc09af238f3772a7993b74b55e1f360891699bfe7a968838ff1229d63b4ab32f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 228, + 208, + 278, + 258 + ], + "blocks": [ + { + "bbox": [ + 228, + 208, + 278, + 258 + ], + "lines": [ + { + "bbox": [ + 228, + 208, + 278, + 258 + ], + "spans": [ + { + "bbox": [ + 228, + 208, + 278, + 258 + ], + "type": "image", + "image_path": "4887450f8d53c8bf4836fa939a2a8834945384d64d0ff5cfb3a42850f5330c1d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 228, + 259, + 279, + 310 + ], + "blocks": [ + { + "bbox": [ + 228, + 259, + 279, + 310 + ], + "lines": [ + { + "bbox": [ + 228, + 259, + 279, + 310 + ], + "spans": [ + { + "bbox": [ + 228, + 259, + 279, + 310 + ], + "type": "image", + "image_path": "cb489915c2d5abc6c6dbfe71d02bb9bad4b98623a4da22a9054b65a962ba7e6a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 280, + 208, + 329, + 258 + ], + "blocks": [ + { + "bbox": [ + 280, + 208, + 329, + 258 + ], + "lines": [ + { + "bbox": [ + 280, + 208, + 329, + 258 + ], + "spans": [ + { + "bbox": [ + 280, + 208, + 329, + 258 + ], + "type": "image", + "image_path": "783830a089319459447f16084dc383527c09572097d6fb7422f881609a84d929.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 280, + 259, + 330, + 310 + ], + "blocks": [ + { + "bbox": [ + 280, + 259, + 330, + 310 + ], + "lines": [ + { + "bbox": [ + 280, + 259, + 330, + 310 + ], + "spans": [ + { + "bbox": [ + 280, + 259, + 330, + 310 + ], + "type": "image", + "image_path": "1f8370d466058de0fc32cf764618132c4b0e4e1184c4e090be3933866ed5ab6c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 331, + 208, + 382, + 258 + ], + "blocks": [ + { + "bbox": [ + 331, + 208, + 382, + 258 + ], + "lines": [ + { + "bbox": [ + 331, + 208, + 382, + 258 + ], + "spans": [ + { + "bbox": [ + 331, + 208, + 382, + 258 + ], + "type": "image", + "image_path": "16ae01fc8d896bfd470dca9aece9e6e5984e664b6f6eea7e9f7e1fc360a69105.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 331, + 259, + 382, + 310 + ], + "blocks": [ + { + "bbox": [ + 331, + 259, + 382, + 310 + ], + "lines": [ + { + "bbox": [ + 331, + 259, + 382, + 310 + ], + "spans": [ + { + "bbox": [ + 331, + 259, + 382, + 310 + ], + "type": "image", + "image_path": "294ae35a955193b3bb59c0c138d9f041f556a1b74f8f90abe72ee47390d364fd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 388, + 208, + 438, + 258 + ], + "blocks": [ + { + "bbox": [ + 388, + 208, + 438, + 258 + ], + "lines": [ + { + "bbox": [ + 388, + 208, + 438, + 258 + ], + "spans": [ + { + "bbox": [ + 388, + 208, + 438, + 258 + ], + "type": "image", + "image_path": "6b89e89a823ea05b18fb09be4ac9466d3797e4bd1ecc0520c502a238a067d693.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 388, + 259, + 438, + 310 + ], + "blocks": [ + { + "bbox": [ + 388, + 259, + 438, + 310 + ], + "lines": [ + { + "bbox": [ + 388, + 259, + 438, + 310 + ], + "spans": [ + { + "bbox": [ + 388, + 259, + 438, + 310 + ], + "type": "image", + "image_path": "dd94debb756685920b8cf3f6eb95f239b430adcd20b700e3bf62846219bb4ed2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 439, + 208, + 489, + 258 + ], + "blocks": [ + { + "bbox": [ + 439, + 208, + 489, + 258 + ], + "lines": [ + { + "bbox": [ + 439, + 208, + 489, + 258 + ], + "spans": [ + { + "bbox": [ + 439, + 208, + 489, + 258 + ], + "type": "image", + "image_path": "ecaf7c120d1d14825a706be883fa310487ffb44643bb8b4c19d386ac2566cd46.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 439, + 259, + 489, + 310 + ], + "blocks": [ + { + "bbox": [ + 439, + 259, + 489, + 310 + ], + "lines": [ + { + "bbox": [ + 439, + 259, + 489, + 310 + ], + "spans": [ + { + "bbox": [ + 439, + 259, + 489, + 310 + ], + "type": "image", + "image_path": "854656f251b557acff769bb4fdb183e71f06c5b9ca184b2fadea53520d4f623e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 491, + 208, + 541, + 258 + ], + "blocks": [ + { + "bbox": [ + 491, + 208, + 541, + 258 + ], + "lines": [ + { + "bbox": [ + 491, + 208, + 541, + 258 + ], + "spans": [ + { + "bbox": [ + 491, + 208, + 541, + 258 + ], + "type": "image", + "image_path": "1a1660150a204dd8dc2d8f84448f5ec52ed10aa4ec2a85808ed9b65d3313ac66.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 491, + 259, + 541, + 310 + ], + "blocks": [ + { + "bbox": [ + 491, + 259, + 541, + 310 + ], + "lines": [ + { + "bbox": [ + 491, + 259, + 541, + 310 + ], + "spans": [ + { + "bbox": [ + 491, + 259, + 541, + 310 + ], + "type": "image", + "image_path": "49b50c368b74b9588dbd562dc1e8ef8e28c37d7336cec38fec75f906bec496de.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 69, + 314, + 119, + 365 + ], + "blocks": [ + { + "bbox": [ + 69, + 314, + 119, + 365 + ], + "lines": [ + { + "bbox": [ + 69, + 314, + 119, + 365 + ], + "spans": [ + { + "bbox": [ + 69, + 314, + 119, + 365 + ], + "type": "image", + "image_path": "46dd9f4bbd0017298a4ed3f58084383a2acc08da40c027bfcb748ec0a62d76fb.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 69, + 366, + 118, + 415 + ], + "blocks": [ + { + "bbox": [ + 69, + 366, + 118, + 415 + ], + "lines": [ + { + "bbox": [ + 69, + 366, + 118, + 415 + ], + "spans": [ + { + "bbox": [ + 69, + 366, + 118, + 415 + ], + "type": "image", + "image_path": "3378bfc7a9eb26f6be7646a7c6bf54f9fc0991ecbfd44c8904075d70a34c4597.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 416, + 100, + 426 + ], + "lines": [ + { + "bbox": [ + 88, + 416, + 100, + 426 + ], + "spans": [ + { + "bbox": [ + 88, + 416, + 100, + 426 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 120, + 314, + 170, + 365 + ], + "blocks": [ + { + "bbox": [ + 120, + 314, + 170, + 365 + ], + "lines": [ + { + "bbox": [ + 120, + 314, + 170, + 365 + ], + "spans": [ + { + "bbox": [ + 120, + 314, + 170, + 365 + ], + "type": "image", + "image_path": "0544123be9cadc284a36bd1f5a0d89692a00fa286ade5b47d13dea1442e53960.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 121, + 366, + 170, + 415 + ], + "blocks": [ + { + "bbox": [ + 121, + 366, + 170, + 415 + ], + "lines": [ + { + "bbox": [ + 121, + 366, + 170, + 415 + ], + "spans": [ + { + "bbox": [ + 121, + 366, + 170, + 415 + ], + "type": "image", + "image_path": "6f3df79bcf9ab519f133526f87889aa5234f2551567c444ab17bbb59193b9275.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 416, + 151, + 426 + ], + "lines": [ + { + "bbox": [ + 140, + 416, + 151, + 426 + ], + "spans": [ + { + "bbox": [ + 140, + 416, + 151, + 426 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 172, + 314, + 222, + 365 + ], + "blocks": [ + { + "bbox": [ + 172, + 314, + 222, + 365 + ], + "lines": [ + { + "bbox": [ + 172, + 314, + 222, + 365 + ], + "spans": [ + { + "bbox": [ + 172, + 314, + 222, + 365 + ], + "type": "image", + "image_path": "45dc42db52e8633d2acde500de003c670d3503e99582dff8c3ef5cb437869db0.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 172, + 366, + 222, + 415 + ], + "blocks": [ + { + "bbox": [ + 172, + 366, + 222, + 415 + ], + "lines": [ + { + "bbox": [ + 172, + 366, + 222, + 415 + ], + "spans": [ + { + "bbox": [ + 172, + 366, + 222, + 415 + ], + "type": "image", + "image_path": "c5d4b2feda5a0480ee9c2a212aa3b1900881ef43fe29cb58130b51678234cae0.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 416, + 203, + 426 + ], + "lines": [ + { + "bbox": [ + 192, + 416, + 203, + 426 + ], + "spans": [ + { + "bbox": [ + 192, + 416, + 203, + 426 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 228, + 314, + 278, + 365 + ], + "blocks": [ + { + "bbox": [ + 228, + 314, + 278, + 365 + ], + "lines": [ + { + "bbox": [ + 228, + 314, + 278, + 365 + ], + "spans": [ + { + "bbox": [ + 228, + 314, + 278, + 365 + ], + "type": "image", + "image_path": "8de72bc3507adaa94da68be45a31b5bf160b52fcd241e49b3bd3dd0764d2d0f6.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 228, + 366, + 278, + 415 + ], + "blocks": [ + { + "bbox": [ + 228, + 366, + 278, + 415 + ], + "lines": [ + { + "bbox": [ + 228, + 366, + 278, + 415 + ], + "spans": [ + { + "bbox": [ + 228, + 366, + 278, + 415 + ], + "type": "image", + "image_path": "82bbda4bbf70bdb19d65371c5cb84cd668fefb7536dc80496a9f5a4e04bdbf54.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 416, + 259, + 426 + ], + "lines": [ + { + "bbox": [ + 248, + 416, + 259, + 426 + ], + "spans": [ + { + "bbox": [ + 248, + 416, + 259, + 426 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 439, + 555, + 463 + ], + "lines": [ + { + "bbox": [ + 55, + 439, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 555, + 463 + ], + "type": "text", + "content": "Figure 3. Qualitative segmentation results of initial prompts and optimized prompts. (a) ground truth, (b) segmentation results with initial prompts, (c) segmentation results with optimized prompts." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 280, + 314, + 330, + 365 + ], + "blocks": [ + { + "bbox": [ + 280, + 314, + 330, + 365 + ], + "lines": [ + { + "bbox": [ + 280, + 314, + 330, + 365 + ], + "spans": [ + { + "bbox": [ + 280, + 314, + 330, + 365 + ], + "type": "image", + "image_path": "fcb533011bf89b2a49f3d4cfbca9683139cc08d4d449c2cc6a52c77539945953.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 280, + 366, + 330, + 415 + ], + "blocks": [ + { + "bbox": [ + 280, + 366, + 330, + 415 + ], + "lines": [ + { + "bbox": [ + 280, + 366, + 330, + 415 + ], + "spans": [ + { + "bbox": [ + 280, + 366, + 330, + 415 + ], + "type": "image", + "image_path": "522af46bcc19a49f861cc8cea998d1ee0f306ef5583a4c8feac224c5be222e9b.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 416, + 311, + 426 + ], + "lines": [ + { + "bbox": [ + 299, + 416, + 311, + 426 + ], + "spans": [ + { + "bbox": [ + 299, + 416, + 311, + 426 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 331, + 314, + 382, + 365 + ], + "blocks": [ + { + "bbox": [ + 331, + 314, + 382, + 365 + ], + "lines": [ + { + "bbox": [ + 331, + 314, + 382, + 365 + ], + "spans": [ + { + "bbox": [ + 331, + 314, + 382, + 365 + ], + "type": "image", + "image_path": "6e71723dc835a06b96f8d8b6acb57442d90ced2579ce69640982f046ee025ff0.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 331, + 366, + 382, + 415 + ], + "blocks": [ + { + "bbox": [ + 331, + 366, + 382, + 415 + ], + "lines": [ + { + "bbox": [ + 331, + 366, + 382, + 415 + ], + "spans": [ + { + "bbox": [ + 331, + 366, + 382, + 415 + ], + "type": "image", + "image_path": "3ea56c0c85a714a81c308fbea5ad42c8103b7f6a41b1714206f07b9f98f2c1da.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 416, + 362, + 426 + ], + "lines": [ + { + "bbox": [ + 351, + 416, + 362, + 426 + ], + "spans": [ + { + "bbox": [ + 351, + 416, + 362, + 426 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 388, + 314, + 438, + 365 + ], + "blocks": [ + { + "bbox": [ + 388, + 314, + 438, + 365 + ], + "lines": [ + { + "bbox": [ + 388, + 314, + 438, + 365 + ], + "spans": [ + { + "bbox": [ + 388, + 314, + 438, + 365 + ], + "type": "image", + "image_path": "407be0e27b39c19f3035bca69f04a109aa7cc3c060c939d09dcabda6e5cbeb47.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 388, + 366, + 438, + 415 + ], + "blocks": [ + { + "bbox": [ + 388, + 366, + 438, + 415 + ], + "lines": [ + { + "bbox": [ + 388, + 366, + 438, + 415 + ], + "spans": [ + { + "bbox": [ + 388, + 366, + 438, + 415 + ], + "type": "image", + "image_path": "5da259c4633ff8b3e1bdd431ffc9b74b11a67763505fa6892f99607d9675cc54.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 416, + 418, + 426 + ], + "lines": [ + { + "bbox": [ + 407, + 416, + 418, + 426 + ], + "spans": [ + { + "bbox": [ + 407, + 416, + 418, + 426 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 439, + 314, + 489, + 365 + ], + "blocks": [ + { + "bbox": [ + 439, + 314, + 489, + 365 + ], + "lines": [ + { + "bbox": [ + 439, + 314, + 489, + 365 + ], + "spans": [ + { + "bbox": [ + 439, + 314, + 489, + 365 + ], + "type": "image", + "image_path": "199597e39389a1207f6d2a2b79e4c2fe0271593d0f0c31ef9c0b7d441dfcfd5b.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 439, + 366, + 489, + 415 + ], + "blocks": [ + { + "bbox": [ + 439, + 366, + 489, + 415 + ], + "lines": [ + { + "bbox": [ + 439, + 366, + 489, + 415 + ], + "spans": [ + { + "bbox": [ + 439, + 366, + 489, + 415 + ], + "type": "image", + "image_path": "5399ada3767f2d52b9df0f0da8d22eaee77a99ad95bdc609b8bbddacca773d45.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 416, + 470, + 426 + ], + "lines": [ + { + "bbox": [ + 459, + 416, + 470, + 426 + ], + "spans": [ + { + "bbox": [ + 459, + 416, + 470, + 426 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 491, + 314, + 541, + 365 + ], + "blocks": [ + { + "bbox": [ + 491, + 314, + 541, + 365 + ], + "lines": [ + { + "bbox": [ + 491, + 314, + 541, + 365 + ], + "spans": [ + { + "bbox": [ + 491, + 314, + 541, + 365 + ], + "type": "image", + "image_path": "87332dd9cd9b791de2c3c48915e8fc4ac252d5f963e8f40a5edb761fa6727aa5.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 491, + 366, + 541, + 415 + ], + "blocks": [ + { + "bbox": [ + 491, + 366, + 541, + 415 + ], + "lines": [ + { + "bbox": [ + 491, + 366, + 541, + 415 + ], + "spans": [ + { + "bbox": [ + 491, + 366, + 541, + 415 + ], + "type": "image", + "image_path": "cb7325b64ad29577351f3d40cb68d6a1e52943bd05457115eda3ac5fa2137cd5.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 416, + 522, + 426 + ], + "lines": [ + { + "bbox": [ + 511, + 416, + 522, + 426 + ], + "spans": [ + { + "bbox": [ + 511, + 416, + 522, + 426 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "bbox": [ + 55, + 483, + 297, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 297, + 697 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 297, + 697 + ], + "type": "text", + "content": "differences can lead to poorer segmentation outcomes for SAM. Through prompt optimization, FM-PPO effectively bridges the domain gap between medical and natural images, enhancing segmentation performance across both domains. Additionally, FM-PPO promotes a relatively uniform physical distribution of point prompts and close proximity between positive and negative prompts. In contrast, Matcher lacks negative prompts, which leads to excessive background segmentation on medical images with domain shifts, particularly when segmenting smaller targets, resulting in suboptimal performance. We further elaborate on the necessity of negative prompts in the supplementary materials. As a result, our method excels at boundary segmentation, especially in complex medical images where other methods struggle with background-foreground adhesion. FM-PPO can optimize the initial prompts generated by FM without additional training, achieving efficient segmentation across various scenarios." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 482, + 555, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 495 + ], + "type": "text", + "content": "4.5. Different Initial Prompts Scheme Optimization" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 312, + 497, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 497, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 497, + 555, + 676 + ], + "type": "text", + "content": "To further validate PPO's effectiveness in optimizing prompts, we applied it to various initial prompt generation schemes and evaluated the downstream SAM segmentation. We selected medical datasets with domain gap, as prompts have a greater impact on SAM's segmentation. Table 3 presents the experimental results, with corresponding qualitative results in the supplementary materials. FM and CS serve as two baseline prompt generation methods. As shown, PPO significantly enhances segmentation performance on both datasets by optimizing the prompts. FM generates a richer prompt set based on patch matching, allowing PPO to achieve stronger optimization. In contrast, CS samples within coarse segmentation regions, leading to a concentrated prompt distribution that may overlook some target areas, thus limiting PPO's optimization potential." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Additionally, Matcher [23], a recent approach in visual prompt engineering, utilizes the FM paradigm to achieve task-agnostic segmentation. To demonstrate PPO's" + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4338" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 102, + 552, + 210 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Comparison results with one-shot methods in natural and medical image datasets. Gray metrics indicates the model is already trained by in-domain datasets. " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " indicates the training-free method for segmentation tasks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 102, + 552, + 210 + ], + "lines": [ + { + "bbox": [ + 59, + 102, + 552, + 210 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 552, + 210 + ], + "type": "table", + "html": "
MethodsFSS-1000VOC2012ISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
HSNet [24]91.18.244.511.457.411.329.716.5
VAT [7]93.67.548.211.449.812.926.416.8
SegGPT† [39]93.77.178.27.861.29.544.114.4
PerSAM† [43]81.18.653.610.937.616.145.816.3
Matcher† [23]91.67.669.011.173.79.746.216.0
FM-PPO† (Ours)91.97.473.79.975.48.170.39.0
", + "image_path": "987b1d16020689018910cceae0763685d0168bab4cc3d51ea771d3cf8a9dc8e0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 64, + 229, + 288, + 466 + ], + "blocks": [ + { + "bbox": [ + 64, + 229, + 288, + 466 + ], + "lines": [ + { + "bbox": [ + 64, + 229, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 64, + 229, + 288, + 466 + ], + "type": "image", + "image_path": "78bd95ee017b830780af4a2b7564d4dece78e3aa48abfa0afef3b6877fc7a834.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 472, + 295, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 495 + ], + "type": "text", + "content": "Figure 4. Qualitative segmentation results of different one-shot methods in natural and medical images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 615 + ], + "type": "text", + "content": "plug-and-play capability, we applied PPO to optimize the prompts generated by Matcher and then fed them back into Matcher for segmentation. Table 3 shows that PPO-optimized prompts significantly improve segmentation performance, particularly on the Kvasir, where the DSC increased by " + }, + { + "bbox": [ + 55, + 517, + 295, + 615 + ], + "type": "inline_equation", + "content": "32.6\\%" + }, + { + "bbox": [ + 55, + 517, + 295, + 615 + ], + "type": "text", + "content": " for smaller targets. This demonstrates that our method effectively boosts SAM's segmentation performance by optimizing various types of initial prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 626, + 129, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 129, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 129, + 639 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "In this paper, we introduce PPO, designed to enhance the distribution of point prompts, thereby improving the downstream automatic segmentation performance of SAM. We first formalize the distribution of point prompts as a dual-space heterogeneous graph and define the optimization direction. Next, we train an agent within this graph environ-" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 249, + 549, + 346 + ], + "blocks": [ + { + "bbox": [ + 319, + 229, + 548, + 241 + ], + "lines": [ + { + "bbox": [ + 319, + 229, + 548, + 241 + ], + "spans": [ + { + "bbox": [ + 319, + 229, + 548, + 241 + ], + "type": "text", + "content": "Table 3. Optimization performance of different initial prompts." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 249, + 549, + 346 + ], + "lines": [ + { + "bbox": [ + 318, + 249, + 549, + 346 + ], + "spans": [ + { + "bbox": [ + 318, + 249, + 549, + 346 + ], + "type": "table", + "html": "
Prompt SchemeISICKvasir
mDSC (%↑)mHD (↓)mDSC (%↑)mHD (↓)
FM48.59.927.213.2
FM-PPO75.4 (↑21.9)8.1 (↓1.8)70.3 (↑43.1)9.0 (↓4.2)
CS57.711.969.110.2
CS-PPO68.6 (↑10.9)10.8 (↓1.1)79.8 (↑10.7)8.9 (↓1.3)
Matcher [23]73.79.746.216.0
Matcher-PPO85.9 (↑12.2)9.0 (↓0.7)77.6 (↑32.6)9.3 (↓6.7)
", + "image_path": "1b8aec8b125a1eee0b763a44567b840d7a013d2584f5afa7658818e4aa55aad7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 373, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 566 + ], + "type": "text", + "content": "ment to automatically activate effective prompt schemes, which are then input into SAM to obtain segmentation results. Our experiments on natural and medical image datasets demonstrate that PPO significantly improves segmentation performance compared to various initial point prompt schemes. In conclusion, PPO offers a promising solution for optimizing any initial point prompts, enhancing the performance of diverse downstream segmentation tasks. Limitations. Although PPO demonstrates strong performance in optimizing incorrect and redundant point prompts, its effectiveness is limited when the initial prompts lack coverage of the target. For example, in the CS approach, if the pseudo-labels do not include the target, prompts are not sampled over the target region. Future research will focus on designing more comprehensive point prompt generation methods to build a more robust prompt optimizer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 587, + 407, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 587, + 407, + 601 + ], + "spans": [ + { + "bbox": [ + 314, + 587, + 407, + 601 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "This research was supported by the National Natural Science Foundation of China (Grant No. 61901292 and Grant No. 62201331), the Fundamental Research Program of Shanxi Province (Grant No. 201901D211080, Grant No. 202303021211082, Grant No. 202403021222276 and Grant No. 202403021212262), the Shanxi Province Graduate Education Innovation Project (Grant No. 2022Y196). We also express our gratitude to the all reviewers and area chairs for their constructive feedback." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4339" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 85, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 85, + 296, + 132 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 85, + 296, + 132 + ], + "spans": [ + { + "bbox": [ + 61, + 85, + 296, + 132 + ], + "type": "text", + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 133, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 133, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 133, + 296, + 205 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 206, + 295, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 295, + 253 + ], + "type": "text", + "content": "[3] Zhen Chen, Qing Xu, Xinyu Liu, and Yixuan Yuan. Un-sam: Universal prompt-free segmentation for generalized nuclei images. arXiv preprint arXiv:2402.16663, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 255, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 255, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 255, + 295, + 313 + ], + "type": "text", + "content": "[4] Haixing Dai, Chong Ma, Zhengliang Liu, Yiwei Li, Peng Shu, Xiaozheng Wei, Lin Zhao, Zihao Wu, Dajiang Zhu, and Wei Liu. Samaug: Point prompt augmentation for segment anything model. arXiv preprint arXiv:2307.01187, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 316, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 295, + 361 + ], + "type": "text", + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 364, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 295, + 422 + ], + "type": "text", + "content": "[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 424, + 295, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 483 + ], + "type": "text", + "content": "[7] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 485, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 485, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 485, + 295, + 555 + ], + "type": "text", + "content": "[8] Duojun Huang, Xinyu Xiong, Jie Ma, Jichang Li, Zequn Jie, Lin Ma, and Guanbin Li. Alignsam: Aligning segment anything model to open context via reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3205-3215, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 557, + 295, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 295, + 604 + ], + "type": "text", + "content": "[9] Jiaxing Huang, Kai Jiang, Jingyi Zhang, Han Qiu, Lewei Lu, Shijian Lu, and Eric Xing. Learning to prompt segment anything models. arXiv preprint arXiv:2401.04651, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 605, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 605, + 295, + 689 + ], + "type": "text", + "content": "[10] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, pages 451–462. Springer, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 689, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 689, + 295, + 713 + ], + "type": "text", + "content": "[11] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 72, + 553, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 553, + 120 + ], + "type": "text", + "content": "Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 122, + 553, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 122, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 316, + 122, + 553, + 181 + ], + "type": "text", + "content": "[12] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 184, + 553, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 254 + ], + "type": "text", + "content": "[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 258, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 553, + 316 + ], + "type": "text", + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 319, + 555, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 555, + 379 + ], + "type": "text", + "content": "[15] Hyeokjun Kweon and Kuk-Jin Yoon. From sam to cams: Exploring segment anything model for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19499-19509, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 381, + 553, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 381, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 316, + 381, + 553, + 428 + ], + "type": "text", + "content": "[16] Yiming Lei, Jingqi Li, Zilong Li, Yuan Cao, and Hongming Shan. Prompt learning in computer vision: a survey. Frontiers of Information Technology & Electronic Engineering, 25(1):42-63, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 431, + 553, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 490 + ], + "type": "text", + "content": "[17] Chengyin Li, Prashant Khanduri, Yao Qiang, Rafi Ibn Sultan, Indrin Chetty, and Dongxiao Zhu. Autoprompting sam for mobile friendly 3d medical image segmentation. arXiv preprint arXiv:2308.14936, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 493, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 540 + ], + "type": "text", + "content": "[18] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 543, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 543, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 543, + 553, + 613 + ], + "type": "text", + "content": "[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 616, + 553, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 616, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 317, + 616, + 553, + 675 + ], + "type": "text", + "content": "[20] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 678, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 678, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 678, + 553, + 714 + ], + "type": "text", + "content": "[21] Xueyu Liu, Guangze Shi, Rui Wang, Yexin Lai, Jianan Zhang, Lele Sun, Quan Yang, Yongfei Wu, Ming Li, Weixia Han, et al. Feature-prompting gbmseg:" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4340" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 78, + 72, + 295, + 132 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 72, + 295, + 132 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 295, + 132 + ], + "type": "text", + "content": "One-shot reference guided training-free prompt engineering for glomerular basement membrane segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 276-285. Springer, 2024. 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 133, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 295, + 239 + ], + "type": "text", + "content": "[22] Xueyu Liu, Rui Wang, Yexin Lai, Yongfei Wu, Hangbei Cheng, Yuanyue Lu, Jianan Zhang, Ning Hao, Chenglong Ban, Yanru Wang, Shuqin Tang, Yuxuan Yang, Ming Li, Xiaoshuang Zhou, and Wen Zheng. Msmtseg: Multi-stained multi-tissue segmentation of kidney histology images via generative self-supervised meta-learning framework. IEEE Journal of Biomedical and Health Informatics, pages 1-12, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 241, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 300 + ], + "type": "text", + "content": "[23] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 301, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 294, + 350 + ], + "type": "text", + "content": "[24] Juhong Min, Dahyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 351, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 351, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 351, + 294, + 421 + ], + "type": "text", + "content": "[25] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 423, + 294, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 471 + ], + "type": "text", + "content": "[26] Saiyang Na, Yuzhi Guo, Feng Jiang, Hehuan Ma, and Junzhou Huang. Segment any cell: A sam-based autoprompting fine-tuning framework for nuclei segmentation. arXiv preprint arXiv:2401.13220, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 472, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 294, + 542 + ], + "type": "text", + "content": "[27] Maxime Aquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 544, + 294, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 604 + ], + "type": "text", + "content": "[28] Xiaoyu Qiu, Hao Feng, Yuechen Wang, Wengang Zhou, and Houqiang Li. Progressive multi-modal conditional prompt tuning. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 46-54, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 605, + 294, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 605, + 294, + 687 + ], + "spans": [ + { + "bbox": [ + 56, + 605, + 294, + 687 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 689, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 689, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 689, + 294, + 714 + ], + "type": "text", + "content": "[30] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma," + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 336, + 72, + 553, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 336, + 72, + 553, + 119 + ], + "type": "text", + "content": "Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 121, + 553, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 121, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 316, + 121, + 553, + 181 + ], + "type": "text", + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 182, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 182, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 182, + 553, + 228 + ], + "type": "text", + "content": "[32] Tal Shaharabany, Aviad Dahan, Raja Giryes, and Lior Wolf. Autosam: Adapting sam to medical images by overloading the prompt encoder. arXiv preprint arXiv:2306.06370, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 230, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 337 + ], + "type": "text", + "content": "[33] Junji Shiraishi, Shigehiko Katsuragawa, Junpei Ikezoe, Tsuneo Matsumoto, Takeshi Kobayashi, Kenichi Komatsu, Mitate Matsui, Hiroshi Fujita, Yoshie Kodera, and Kunio Doi. Development of a digital image database for chest radiographs with and without a lung nodule: receiver operating characteristic analysis of radiologists' detection of pulmonary nodules. American journal of roentgenology, 174(1):71-74, 2000. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 339, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 410 + ], + "type": "text", + "content": "[34] Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. Vrp-sam: Sam with visual reference prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23565-23574, 2024. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 412, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 470 + ], + "type": "text", + "content": "[35] Chenglong Wang, Dexuan Li, Sucheng Wang, Chengxiu Zhang, Yida Wang, Yun Liu, and Guang Yang. SAMMed: A medical image annotation framework based on large vision model. arXiv preprint arXiv:2307.05617, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 472, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 531 + ], + "type": "text", + "content": "[36] Jiaqi Wang, Zhengliang Liu, Lin Zhao, Zihao Wu, Chong Ma, Sigang Yu, Haixing Dai, Qiushi Yang, Yiheng Liu, Songyao Zhang, et al. Review of large vision models and visual prompt engineering. MetaRadiology, page 100047, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 533, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 591 + ], + "type": "text", + "content": "[37] Rui Wang, Zhouwang Yang, and Yanzhi Song. Osamfundus: A training-free, one-shot segmentation framework for optic disc and cup in fundus images. Biomedical Signal Processing and Control, 100:107069, 2025. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 593, + 553, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 652 + ], + "type": "text", + "content": "[38] Xu Wang, Sen Wang, Xingxing Liang, Dawei Zhao, Jincai Huang, Xin Xu, Bin Dai, and Qiguang Miao. Deep reinforcement learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 35 (4):5064-5078, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 654, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 553, + 714 + ], + "type": "text", + "content": "[39] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1130-1140, 2023. 8" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4341" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 423 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 120 + ], + "type": "text", + "content": "[40] Bin Xie, Hao Tang, Bin Duan, Dawen Cai, and Yan Yan. Masksam: Towards auto-prompt sam with mask classification for medical image segmentation. arXiv preprint arXiv:2403.14103, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 181 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 181 + ], + "type": "text", + "content": "[41] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 182, + 295, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 295, + 253 + ], + "type": "text", + "content": "[42] Chaoning Zhang, Fachrina Dewi Puspitasari, Sheng Zheng, Chenghao Li, Yu Qiao, Taegoo Kang, Xinru Shan, Chenshuang Zhang, Caiyan Qin, Francois Rameau, et al. A survey on segment anything model (sam): Vision foundation model meets prompt engineering. arXiv preprint arXiv:2306.06211, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 255, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 255, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 57, + 255, + 294, + 303 + ], + "type": "text", + "content": "[43] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 2, 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 304, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 304, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 57, + 304, + 294, + 363 + ], + "type": "text", + "content": "[44] Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, et al. A comprehensive survey on pretrained foundation models: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 365, + 294, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 365, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 57, + 365, + 294, + 423 + ], + "type": "text", + "content": "[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4342" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Plug-and-Play Versatile Compressed Video Enhancement/e1321f09-cc21-461b-8861-4f42cf1e3c7e_content_list.json b/2025/Plug-and-Play Versatile Compressed Video Enhancement/e1321f09-cc21-461b-8861-4f42cf1e3c7e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..829b72971d9055f039e896e81e69e666cc820c8d --- /dev/null +++ b/2025/Plug-and-Play Versatile Compressed Video Enhancement/e1321f09-cc21-461b-8861-4f42cf1e3c7e_content_list.json @@ -0,0 +1,2275 @@ +[ + { + "type": "text", + "text": "Plug-and-Play Versatile Compressed Video Enhancement", + "text_level": 1, + "bbox": [ + 207, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huimin Zeng Jiacheng Li Zhiwei Xiong*", + "bbox": [ + 299, + 180, + 702, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Science and Technology of China", + "bbox": [ + 308, + 199, + 687, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zenghuimin, jciee}@mail.ustc.edu.cn zwxiong@ustc.edu.cn", + "bbox": [ + 233, + 218, + 756, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a widely adopted technique in data transmission, video compression effectively reduces the size of files, making it possible for real-time cloud computing. However, it comes at the cost of visual quality, posing challenges to the robustness of downstream vision models. In this work, we present a versatile codec-aware enhancement framework that reuses codec information to adaptively enhance videos under different compression settings, assisting various downstream vision tasks without introducing computation bottleneck. Specifically, the proposed codec-aware framework consists of a compression-aware adaptation (CAA) network that employs a hierarchical adaptation mechanism to estimate parameters of the frame-wise enhancement network, namely the bitstream-aware enhancement (BAE) network. The BAE network further leverages temporal and spatial priors embedded in the bitstream to effectively improve the quality of compressed input frames. Extensive experimental results demonstrate the superior quality enhancement performance of our framework over existing enhancement methods, as well as its versatility in assisting multiple downstream tasks on compressed videos as a plug-and-play module. Code and models are available at https://huimin-zeng.github.io/PnP-VCVE/.", + "bbox": [ + 88, + 301, + 485, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 685, + 220, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the flower booming of short video platforms, video has become one of the most popular multimedia formats. In addition to distributing visual content, in practical scenarios (e.g., autonomous driving [43, 48]), it is common to upload the captured videos to the cloud end for further visual analysis and downstream applications (e.g., object detection [5, 51] and segmentation [20, 44]). However, due to the bandwidth constraint during transmission, these videos are typically compressed with varying levels, resulting in poor visual quality and suboptimal performance in downstream tasks [22, 59] (e.g., inaccurate segmentation bound", + "bbox": [ + 89, + 705, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b5952d352646c15004795393533a8a5ef99b2c6f588ac52605ac07b59022c64.jpg", + "image_caption": [ + "Figure 1. The proposed codec-aware enhancement framework reuses codec information to adaptively enhance videos across different compression settings, while assisting in various downstream tasks in a plug-and-play manner." + ], + "image_footnote": [], + "bbox": [ + 516, + 267, + 906, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "aries in Fig. 1). Given the crucial role of videos in data transmission, there is a critical need for a versatile solution to enhance videos of diverse compression levels and effectively support various downstream tasks.", + "bbox": [ + 511, + 532, + 906, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing video enhancement methods are hard to respond to these demands. Specifically, to effectively enhance videos of different compression levels, previous methods [11, 12, 16, 22, 32, 55] employ separate enhancement models for each compression level, which is inflexible occurring unseen compression levels. Recent approaches [8, 18, 47] consider this issue as the generalization ability across diverse compression levels, therefore randomly selecting inputs of different compression levels during training. However, such a training strategy is compression-agnostic and offers limited improvement. Most importantly, the aforementioned methods focus primarily on improving perceptual quality, neglecting the need to assist in downstream tasks in real-world scenarios.", + "bbox": [ + 511, + 595, + 908, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Based on the mismatch between the versatility demand and existing solutions, here we summarize the following criteria of a favorable solution: 1) adaptively enhance videos of varying compression levels with a single model; 2) effectively assist various downstream tasks on compressed videos in a plug-and-play manner; and 3) given the", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17767", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "practical scenarios where real-time processing is required, it should meet the above objectives without causing a computation bottleneck. To achieve this, we introduce a codec-aware enhancement framework (as shown in Fig. 1) that reuses codec information embedded in the bitstream. By incorporating compression factors, the framework dynamically adjusts its parameters to flexibly enhance inputs of different compression levels. By reusing motion vectors and partition maps, it efficiently aggregates temporal and spatial clues without introducing redundant computations.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our codec-aware enhancement framework comprises a compression-aware adaptation (CAA) network and a bitstream-aware enhancement (BAE) network. The CAA network serves as a \"meta\" network that dynamically adjusts the parameters of the subsequent BAE network. A hierarchical adaptation mechanism is proposed to first estimate sequence-adaptive parameters based on the sequence compression level, and then re-weight these parameters according to the frame compression level, thereby achieving a BAE network tailored for each input frame. The frame-adaptive BAE network conducts motion vector alignment to aggregate intra-frame information and provide useful clues for the current frame. Subsequently, based on the region complexity indicated by partition maps, the region-aware refinement assigns independent filters for different regions, achieving flexible enhancement for different regions. Comprehensive experiments are conducted to demonstrate the superiority of our method in improving the quality of compressed videos, and the effectiveness of assisting in various downstream tasks (i.e., video super-resolution, optical flow estimation, and video object segmentation). Our contributions are summarized as follows:", + "bbox": [ + 91, + 243, + 483, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a codec-aware framework for versatile compressed video enhancement, which adaptively enhances input videos of different compression levels and supports a wide range of downstream vision tasks.", + "- We develop a compression-aware adaptation (CAA) network and a bitstream-aware enhancement (BAE) network that utilize the off-the-shelf codec information, contributing to generalizing across different compression settings and boosting the enhancement performance with a unified framework.", + "- Experimental results show the superiority of our method over existing enhancement methods, and its effectiveness in serving as a plug-and-play enhancement module to assist in downstream tasks." + ], + "bbox": [ + 89, + 574, + 482, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 797, + 232, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Compressed Video Enhancement", + "text_level": 1, + "bbox": [ + 89, + 820, + 380, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing compressed video enhancement methods can be categorized into in-loop and post-processing methods. Although in-loop methods [15, 24, 26, 40] effectively improve the quality of reconstructed frames, they embed filters", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in the encoding and decoding loops, therefore not suitable for enhancing already compressed videos. While the post-processing methods [17, 25, 27, 36, 42, 47, 50, 56, 58, 61, 63, 64] provide more practical solutions to enhance compressed videos by placing filters at the decoder side. Observing the quality fluctuation across frames, MFQE [55] locates the peak quality frame (PQF) with an SVM-based detector and proposes a multi-frame quality enhancement mechanism to enhance non-PQFs. MFQE 2.0 [22] further designs a BiLSTM-based detector and performs multiframe quality enhancement for both non-PQF and PQF. To address the inaccuracies in optical flow estimation from compressed videos, STDF [16] proposes estimating the offset field using spatio-temporal deformable convolution. S2SVR [32] introduces a sequence-to-sequence network to model long-range dependencies within frames. The aforementioned methods inflexibly equip a separate model for each compression level, while we propose to adaptively handle diverse inputs with a single unified model. Recent methods [18, 64] utilize spatial priors from bitstream to address multiple compression levels, however, they only consider I/P-frames, whereas we design a hierarchical adaptation mechanism to address all types of frames.", + "bbox": [ + 511, + 90, + 906, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Codec-Aware Video Super-Resolution", + "text_level": 1, + "bbox": [ + 511, + 446, + 838, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some works in video super-resolution (VSR) [11, 12, 28, 29, 52, 60, 62] explore ways of using codec information such as motion estimation and spatial prior for reconstruction. COMISR [29] focuses on reducing accumulated warping errors caused by the random locations of the intra-frame from compressed video frames. Chen et al. [11] employ motion vectors to build the temporal relationship and suppress coding artifacts. CVCP [12] utilizes motion vectors and spatial priors with a guided soft alignment scheme and guided SFT layer, respectively. CIAF [60] leverages motion vectors and residuals to model temporal relationships and skip redundant computations, respectively. Despite leveraging codec information, these methods focus on a single task (i.e., VSR). In contrast, our method not only shows competitive performance in VSR (see supplementary materials), but also effectively supports a range of downstream tasks, which is not explored by the aforementioned works.", + "bbox": [ + 511, + 465, + 908, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Dynamic Neural Networks", + "text_level": 1, + "bbox": [ + 511, + 729, + 756, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of setting separate models for different inputs, the early mixture of expert (MoE) structure [2, 19, 37] constructs parallel network branches and selectively executes branches to obtain the weighted outputs. Instead of increasing the number of parallel branches, the dynamic parameters ensemble strategy [2, 13, 19, 23, 37, 54] presets parallel expert layers and selectively fusing their parameters to promote the network capability and generalization ability, therefore serving as an efficient alternative to MoE. To promote the generalization ability of pre-trained models, Gain-", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17768", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/252eafdcfa9fa26beb5db55d2291032b16fbdbcb0b367a33ac02f6e8a863c7c5.jpg", + "image_caption": [ + "Figure 2. Hierarchical structure of quality adjustment, where frames are divided into multiple groups of pictures (GOP). The Constant Rate Factor (CRF) affects video quality at both sequence and frame levels. An increase in the CRF value indicates a reduction in video quality (e.g., lower PSNR values)." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 475, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tune [38] proposes to predict a single multiplicative scaling parameter for each channel according to test samples, thus modifying static models to test-adaptive ones. Li et al. [30] handle the conflicts between the domain-agnostic model and multiple target domains with dynamic transfer, which is simply modeled by combining residual matrices and a static convolution matrix. DRConv [10] divides the input image into different regions with a learnable mask and assigns multiple filters for these regions, which enhances the feature representation ability without introducing a noticeable computation burden. Instead of searching and learning conditions for dynamic parameters, we leverage priors embedded in the bitstream as the condition.", + "bbox": [ + 88, + 344, + 482, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 549, + 230, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We take H.264 [53] as a representative standard to analyze available codec information. Note recent codecs [4, 46] also provide similar priors (i.e., CRF, motion vector, partition map), therefore assuring the applicability of our method. To reduce transmission bandwidth, codecs compress videos by adjusting quality and reducing redundancy.", + "bbox": [ + 89, + 571, + 482, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Hierarchical Quality Adjustment", + "text_level": 1, + "bbox": [ + 89, + 669, + 382, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video quality is commonly influenced by the constant rate factor (CRF), which involves hierarchical adjustment for both sequence-wise and frame-wise compression.", + "bbox": [ + 89, + 688, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sequence-wise CRF. The CRF ranges from 0 to 51 to balance compression efficiency and visual quality. A higher CRF results in more compact output but increased pixel loss (e.g., the average PSNR of CRF35 is much lower than CRF15 in Fig. 2). By considering the sequence-wise CRF (denoted as $CRF_{s}$ ), the enhancement network can be tailored to handle videos of different compression levels.", + "bbox": [ + 89, + 734, + 482, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frame-wise CRF As shown in Fig. 2, a video sequence is divided into multiple groups of pictures (GOP) and further categorized as I-frames, P-frames, or B-frames. The CRF value of each frame (denoted as $CRF_{i}$ ) is dynamically", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "adjusted based on $CRF_{s}$ so that lower $CRF_{i}$ is assigned for I/P frames to maintain quality and higher $CRF_{i}$ for B frames for compact representations.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the hierarchical quality adjustment paradigm, we design a hierarchical adaptation paradigm that first performs sequence adaptation to predict network parameters based on $CRF_{s}$ , and then re-weights these parameters according to $CRF_{i}$ for frame adaptation. In practical scenarios where $CRF_{i}$ is unavailable (e.g., limited access to full bitstream), the proposed method can instead use slice type (I/P/B) for frame adaptation, which is demonstrated to yield similar performance in Sec. 5.2.", + "bbox": [ + 511, + 136, + 906, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Redundancy Reduction", + "text_level": 1, + "bbox": [ + 511, + 277, + 730, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To reduce redundancy and improve the entropy of the bitstream,Codec block-wisely perform motion estimation to model the intra-frame correlations and embed the correlations in the bitstream for decoding.", + "bbox": [ + 511, + 296, + 903, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Partition map. As shown in Fig. 1, different regions of each frame are partitioned into blocks of varying sizes (e.g., H.264 provides macroblocks of $16 \\times 16$ , $16 \\times 8$ , $8 \\times 16$ , and $8 \\times 8$ ) according to the texture complexity. Flat regions (e.g., the ground) can tolerate higher quantization errors and are therefore divided into blocks of large size, while complex regions (e.g., leaves and fence) take smaller blocks to maintain details. To effectively enhance regions of different complexity, we propose dynamically assigning filters based on the partition map that indicates region complexity.", + "bbox": [ + 511, + 357, + 905, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion vector. Motion vectors are utilized in decoding to aggregate information from reference frames and propagate information of current frame. As illustrated in Fig. 1, they describe the relationship between current frame and its reference frames in a block-wise manner. Although motion vectors can be noisy and are less precise than optical flow, they effectively align reference frames with current frame, serving as a cost-effective alternative for optical flow.", + "bbox": [ + 511, + 508, + 905, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Codec-Aware Enhancement Framework", + "text_level": 1, + "bbox": [ + 511, + 638, + 872, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 660, + 624, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 3(a), the proposed method comprises a compression-aware adaption (CAA) network $\\mathcal{G}_{\\phi}$ and a bitstream-aware enhancement (BAE) network $\\mathcal{F}_{\\theta_i}$ . The CAA network employs a hierarchical compression adaptation mechanism to estimate parameters for the frame-adaptive BAE network, which then aggregates intra-frame information and performs region-aware refinement to enhance the input compressed frame.", + "bbox": [ + 511, + 680, + 905, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Compression-Aware Adaptation Network", + "text_level": 1, + "bbox": [ + 511, + 805, + 867, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To handle sequences of varying compression levels and quality fluctuations across frames, as illustrated in Fig. 3(a), the CAA network $\\mathcal{G}_{\\phi}$ utilizes $CRF_{s}$ for sequence adaptation to estimate sequence-adaptive parameters, and perform frame adaptation to refine these parameters based on $CRF_{i}$ .", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17769", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f2e6968fdb4faa56129914f2aa33472b03e6f6b132ac64863568be3c7d261bcf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 83, + 854, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/987705a8866119a8a9efeaab1fe4240c181b9b24b52fe414c75ca859ab1bbf7c.jpg", + "image_caption": [ + "Figure 3. The proposed Codec-Aware Enhancement Framework consists of two sub-networks: 1) the Compression-Aware Adaptation (CAA) Network, which hierarchically applies sequence adaptation and frame adaptation to dynamically adjust parameters of the enhancement network; and 2) Bitstream-Aware Enhancement (BAE) Network, which leverages motion vectors to align frames and conducts region-aware refinement to flexibly enhance regions of different complexity." + ], + "image_footnote": [], + "bbox": [ + 138, + 270, + 316, + 373 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/00b816b2a20be9a9474f78fed57b3062d9b63d43cb0adfd2ac883c8d7219e61d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 270, + 493, + 373 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4c54ae2e59ec269ad78571d3d467ee6468c19468448edf8f249ed0910aa9c41b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 270, + 650, + 373 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/516ba0b2603165777b42e9571593bd8a6b81f776ce183826e09d15478a26d939.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 271, + 852, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sequence adaptation. To ensure robust performance across multiple compression settings without increasing complexity, we propose estimating sequence-adaptive parameters for the enhancement network instead of fusing features from separate submodels. As shown in Fig. 3(b), parallel expert layers $\\{f_{\\theta_1}, f_{\\theta_2}, \\dots, f_{\\theta_N}\\}$ , which share the same architecture but have independent parameters, serve as the basis for parameter combination. The sequence-wise $CRF_s$ is adopted as the condition to re-weight parameters of these expert layers, which is expressed as follows,", + "bbox": [ + 88, + 450, + 483, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta_ {s}} = \\mathcal {G} _ {\\phi_ {s}} \\left(C R F _ {s}, \\left\\{f _ {\\theta_ {1}}, f _ {\\theta_ {2}}, \\dots , f _ {\\theta_ {N}} \\right\\}\\right) = \\sum_ {n = 1} ^ {N} w _ {n} f _ {\\theta_ {n}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 602, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\theta_s}$ and $\\mathcal{G}_{\\phi_s}$ denote the sequence-adaptive expert layer and the sequence-wise weight generator, respectively. $w_n$ denotes the weight for each expert layer. We set $N = 6$ (see ablation studies in supplementary materials) and visualize $w_n$ against different $CRF_s$ in Fig. 4, which shows that each expert layer has a distinct preference for specific $CRF_s$ . Compared to MoE that re-weights output features, re-weighting expert layer parameters (as shown in Eq. 1) is computationally efficient and comparable to the network constructed with a single expert layer. Note that $CRF_s$ is constant for frames within the same sequence, $f_{\\theta_s}$ is predicted only once and reused by subsequent frames.", + "bbox": [ + 88, + 643, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frame adaptation. To flexibly enhance frames with different visual quality, we propose to re-weight the sequence-based $f_{\\theta_s}$ using frame-wise $CRF_{i}$ . We attribute the quality fluctuation between the sequence and current frame to the disparity between $CRF_{s}$ and $CRF_{i}$ , which can be ad", + "bbox": [ + 88, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f66511060d4f358a95dc2fd332451b6b6dfafb423f2343ed4dd0663d8f934480.jpg", + "image_caption": [ + "Figure 4. Visualization of $w_{n}$ against different $CRF_{s}$ , where each expert shows a distinct preference for specific $CRF_{s}$ values." + ], + "image_footnote": [], + "bbox": [ + 560, + 446, + 857, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dressed by introducing a set of frame-wise auxiliary parameters $\\triangle \\theta_{i}$ . As shown in Fig. 3(c), the auxiliary parameters $\\triangle \\theta_{i}$ that conditioned on $CRF_{i}$ re-weights the sequence-adaptive $f_{\\theta_s}$ to obtain the frame-adaptive expert layer $f_{\\theta_i}$ which is expressed as follows,", + "bbox": [ + 511, + 633, + 906, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta_ {i}} = \\mathcal {G} _ {\\phi_ {i}} \\left(C R F _ {i}, f _ {\\theta_ {s}}\\right) = f _ {\\theta_ {s} + \\triangle \\theta_ {i}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 713, + 903, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\theta_i}$ and $\\mathcal{G}_{\\phi_i}$ denote the estimated frame-adaptive expert layer and the frame-wise parameters generator, respectively. As shown by the black dashed lines in Fig. 3(a), the obtained $f_{\\theta_i}$ is used to construct the enhancement blocks, resulting in the frame-adaptive BAE network $\\mathcal{F}_{\\theta_i}$ (introduced in the following Sec. 4.3).", + "bbox": [ + 511, + 733, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Bitstream-Aware Enhancement Network", + "text_level": 1, + "bbox": [ + 511, + 833, + 862, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To leverage high-quality frames and propagate information, the BAE network $\\mathcal{F}_{\\theta_i}$ utilizes motion vectors to align reference frames with the current frame. Meanwhile, the par", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17770", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cbe3707dba8b057c2a02dd066680850611f631e480f4b6fb40311aa443d64243.jpg", + "image_caption": [ + "Reference frame" + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 230, + 138 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/85bc66cfbd07932a6db7a2b93d1b18030ec275ca44ae7b88ab56f39599426598.jpg", + "image_caption": [ + "$h_i$" + ], + "image_footnote": [], + "bbox": [ + 232, + 88, + 343, + 138 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bfe3f50b910ec0825c4f674b67ba625663a0fbadd4f694562831e3dae7da534f.jpg", + "image_caption": [ + "$\\hat{h}_i$" + ], + "image_footnote": [], + "bbox": [ + 344, + 88, + 454, + 138 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fd9725b6d195a16ca67476ec4e1ee3128230dcf1e662db6450894c9486b3394b.jpg", + "image_caption": [ + "$\\mathcal{S}(M_i^{8\\times 8},h_i)$", + "Figure 5. Visualization of features in region-aware refinement, where $h_i$ and $\\hat{h}_i$ indicate the input and output features, respectively. The refined features are denoted in the format of $S(M_i^{type}, h_i)$ ." + ], + "image_footnote": [], + "bbox": [ + 120, + 151, + 230, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/26d2d57be50facdd724be3d6ea4582d2844fd15d73191db46b14de09f72fff73.jpg", + "image_caption": [ + "$\\mathcal{S}(M_i^{16\\times 16},h_i)$" + ], + "image_footnote": [], + "bbox": [ + 230, + 151, + 343, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d10ecead72505980800ca33277dc660618d1818ccfd7a3d2f9d3ddb02a8ae7a3.jpg", + "image_caption": [ + "$\\mathcal{S}(M_i^{16\\times 8},h_i)$" + ], + "image_footnote": [], + "bbox": [ + 344, + 151, + 454, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tition map serves as spatial complexity guidance to enable flexible enhancement of different regions.", + "bbox": [ + 89, + 287, + 482, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion vector alignment. Since the motion vectors roughly model the temporal relationship in a block-wise manner, for each block of the current frame shown in Fig. 3(d), motion vectors locate blocks with similar content in the previous and future reference features (highlighted with red and blue boxes). The warped reference features are concatenated with current frame along the channel dimension as input of the BAE network, expressed as follows,", + "bbox": [ + 89, + 319, + 483, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {i} = \\left[ M V \\left(h _ {i} ^ {p}\\right), M V \\left(h _ {i} ^ {f}\\right), x _ {i} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 441, + 480, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{x}_i$ , $h_i^p$ and $h_i^f$ denote the current input frame, enhanced features of previous and future reference frames, respectively. $MV$ denotes warping reference features based on motion vectors. $[,]$ denotes concatenation along channel dimension. Bilinear interpolation is adopted for the case that the offsets of motion vectors are not integers.", + "bbox": [ + 89, + 464, + 483, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Region-aware refinement. To effectively enhance regions of different complexity, we propose to dynamically assign different filters for regions based on the partition map. As shown in Fig. 3(e), the block-based partition map $P_{i}$ is decoupled into multiple binary masks $\\{M_i^{8\\times 8}, M_i^{16\\times 16}, \\dots, M_i^{type}\\}$ according to the size of macroblocks, allowing separate refinement of regions using sparse convolution [33]. The output is defined as the sum of frame-adaptive extracted features and separately region-aware refined features, depicted as follows,", + "bbox": [ + 89, + 555, + 483, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {h} _ {i} = \\mathcal {F} _ {\\theta_ {i}} \\left(\\hat {x} _ {i}, P _ {i}\\right) = f _ {\\theta_ {i}} * h _ {i} + \\sum_ {t y p e = 1} ^ {M} \\mathcal {S} \\left(M _ {i} ^ {t y p e}, h _ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 705, + 480, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{h}_i$ indicates the output features. $S$ denotes the operations applying sparse convolution guided by mask $M_i^{type}$ to refine input features. In H.264 standard, three types of macroblocks are used $(16\\times 16,8\\times 16 / 16\\times 8,$ and $8\\times 8)$ , thus $M$ is set to 3. Features in region-aware refinement are visualized in Fig 5, where the refined features are denoted with operations like $S(M_{i}^{8\\times 8},h_{i})$ . As can be seen, the output features $\\hat{h}_i$ contain more fine-grained and high-frequency details than $h_i$ . Meanwhile, refined features provide distinct activations for different regions. For instance, $S(M_i^{8\\times 8},h_i)$", + "bbox": [ + 89, + 747, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "focuses on static objects (e.g., trees) while $\\mathcal{S}(M_i^{8\\times 16},h_i)$ focuses on moving objects (e.g., the bus).", + "bbox": [ + 511, + 89, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Loss function", + "text_level": 1, + "bbox": [ + 513, + 128, + 653, + 142 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt Charbonnier penalty loss [9] as the loss function and train the proposed codec-aware enhancement framework in an end-to-end manner. The specific loss function is expressed as follows,", + "bbox": [ + 511, + 147, + 905, + 205 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\sqrt {\\left\\| y _ {i} - \\hat {y} _ {i} \\right\\| ^ {2} + \\epsilon^ {2}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 209, + 903, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{i}, \\hat{y}_{i}$ and $T$ indicate the uncompressed ground truth, the predicted output, and the length of the input sequence. $\\epsilon$ is set to $1 \\times 10^{-12}$ .", + "bbox": [ + 511, + 253, + 905, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 306, + 645, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 513, + 327, + 720, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compression settings. H.264 is a popular video compression standard that compresses nearly $85\\%$ of internet videos [1], and tends to introduce more severe degradations than H.265 and H.266. We adopt H.264 [45] and compress videos with the $CRF_{s}$ values of 15, 25, and 35.", + "bbox": [ + 511, + 347, + 905, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tasks and training dataset. Our tasks involve quality enhancement and assisting downstream tasks on compressed inputs. The primary downstream tasks include video superresolution, optical flow estimation and video object segmentation, with video inpainting reported as an extension to fully evaluate the versatility. Training splits of REDS [39] and DAVIS [41] datasets are combined for training.", + "bbox": [ + 511, + 422, + 905, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared methods. We compare with representative methods in compressed video enhancement, including MFQE 2.0 [22], STDF [16], S2SVR [32] and Metabit [18]. For a fair comparison, we fully retrain these methods with the same training dataset and configurations. For downstream tasks, the compressed video is first enhanced by quality enhancement methods, and then fed to downstream models for corresponding tasks and further assessment.", + "bbox": [ + 511, + 529, + 905, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 513, + 657, + 607, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our evaluations are two-fold: 1) verifying the quality enhancement performance on seen, unseen, and highly compressed scenarios; 2) evaluating the versatility to assist different downstream tasks on multiple compression settings.", + "bbox": [ + 511, + 676, + 905, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2.1. Quality Enhancement Performance", + "text_level": 1, + "bbox": [ + 511, + 746, + 805, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative results. The results of compressed video quality enhancement are reported in Tab.1, which is evaluated on the REDS4 datasets [39] using PSNR and SSIM (the higher the better). Note the CRF values of 18, 28 and 38 are not included during training. For each method, we include the model complexity and inference speed. For our method, we report results of both applying $CRF_{i}$ and its substitution with slice type (highlighted with grey). As shown in Tab. 1, leveraging slice type yields comparable performance", + "bbox": [ + 511, + 763, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17771", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d1ff7f80808d6f80ef7741a476dbe1b0fadf4f6fc15a8b008081f56fe7523e09.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParam/MFLOPs/GSpeed/msFPSCRF15CRF25CRF35CRF18CRF28CRF38
PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑
Input----41.04 / 0.978534.92 / 0.936329.25 / 0.823839.12 / 0.969833.18 / 0.912327.69 / 0.7725
MFQE 2.0 [22]1.6451531940.95 / 0.980634.83 / 0.937829.22 / 0.825638.97 / 0.971233.13 / 0.914027.67 / 0.7742
STDF [16]1.2745382641.15 / 0.979335.23 / 0.939829.74 / 0.835939.28 / 0.971233.58 / 0.917828.11 / 0.7853
S2SVR [32]7.43294--41.96 / 0.983435.61 / 0.944529.87 / 0.839139.88 / 0.975533.87 / 0.922328.19 / 0.7881
Metabit [18]1.6092244241.04 / 0.978534.92 / 0.936329.25 / 0.823839.11 / 0.969833.18 / 0.912327.69 / 0.7725
Ours4.5647362842.22 / 0.984235.90 / 0.946830.17 / 0.847140.17 / 0.976734.16 / 0.925828.49 / 0.7985
42.24 / 0.984235.91 / 0.946830.19 / 0.847240.18 / 0.976734.17 / 0.925828.52 / 0.7985
", + "bbox": [ + 94, + 80, + 903, + 198 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1381f0b495b2f23e1d4c044cbbd8c659778447964fe6f018c4888d88f39a40d8.jpg", + "image_caption": [ + "Frame 030,Clip 020,CRF35" + ], + "image_footnote": [], + "bbox": [ + 94, + 234, + 331, + 362 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81cce0bd7a8cb110d3f4a7e0bb49145379db7bc7a21652676799e0495b2bf111.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 234, + 413, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8a72818269e3cf40bff5988f1aebd5f68d5abe872bb04d02da8d7b9ac5e091b.jpg", + "image_caption": [ + "Input", + "Figure 6. Qualitative results on quality enhancement, where our method effectively reduces the compression artifacts, achieving visually pleasant results. In contrast, the results of the compared methods still contain severe distortions (e.g., the calf in the 1st row)." + ], + "image_footnote": [], + "bbox": [ + 333, + 299, + 413, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af260206614883bcc887d2f9a294dc6adfea2bfc66bb90fee07b277fea3d1575.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 234, + 495, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a7a719c8024b9de8a3b047dc00c4a5199f523a97615ee7439b839a5c4ecb88e.jpg", + "image_caption": [ + "MFQE 2.0" + ], + "image_footnote": [], + "bbox": [ + 413, + 299, + 495, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f12387a4674e6c115e1a8115c794c432337218e58e933bc6c71ca7d518e988db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 234, + 576, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d35ee88ccbfd149b7ef0582c3c0637ff0815c1ba910b4e42ce8956ca2754c463.jpg", + "image_caption": [ + "STDF" + ], + "image_footnote": [], + "bbox": [ + 496, + 299, + 576, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6ffadf35ac7f87219e87c799c72bb3e653e041741001322d8b5af50309e5e588.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 234, + 658, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/031ba3b1969564e735cbaffecb85a5c9deb642f3ca62f49f15a608e3a89af3a5.jpg", + "image_caption": [ + "S2SVR" + ], + "image_footnote": [], + "bbox": [ + 578, + 299, + 658, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/63cd79b7913622844f8d955d831212b87f77fe12dfc9b8821d8439f02297a927.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 234, + 738, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/facfdbbdb0410ffbf5d756bc64a1bec304b11b26421c3dd97366f016bc607f3d.jpg", + "image_caption": [ + "Metabit" + ], + "image_footnote": [], + "bbox": [ + 660, + 299, + 738, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1310b3cc38a7043f39276b23b5968bfd1ae9e29dd81a1cc087377f6b7698b771.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 234, + 821, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f267dba5ede5c5e0eb24f37454f5eda60de9267119b2b0168ff39cd0329ca258.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 741, + 299, + 821, + 361 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1b02fbfb2b15f26231ddf1b81ca81992646123e1ffff490c021ba64bdcacbeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 234, + 903, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/30e32489a6322e42fa8c8810cef6cfe3d551e67d9300d3b1e1e35c678d3684ae.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 821, + 299, + 903, + 361 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a72b3d58352e74146b0a3319cd0fbd9fcc221601d40b259f73b5caaaf0f9969b.jpg", + "table_caption": [ + "Table 1. Quantitative results on quality enhancement, where PSNR and SSIM (higher is better) are adopted for evaluation. The best and second best results are marked with bold and underline. Results obtained by replacing $CRF_{i}$ with slice type are highlighted with grey." + ], + "table_footnote": [], + "table_body": "
MethodCRF15CRF25CRF35
PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
BasicVSR29.24 / 0.821226.19 / 0.713123.40 / 0.6005
+ MFQE 2.029.29 / 0.823326.28 / 0.718223.46 / 0.6056
+ STDF29.31 / 0.824726.51 / 0.729323.80 / 0.6249
+ S2SVR29.45 / 0.828826.70 / 0.734623.86 / 0.6270
+ Metabit29.24 / 0.821126.19 / 0.713123.39 / 0.6005
+ Ours29.54 / 0.832826.85 / 0.741924.02 / 0.6361
IconVSR29.29 / 0.823026.19 / 0.713023.39 / 0.6003
+ MFQE 2.029.37 / 0.825426.28 / 0.718223.45 / 0.6055
+ STDF29.36 / 0.826326.52 / 0.729223.79 / 0.6248
+ S2SVR29.54 / 0.830626.71 / 0.734523.85 / 0.6269
+ Metabit29.29 / 0.823026.19 / 0.713023.39 / 0.6003
+ Ours29.63 / 0.834426.86 / 0.741824.01 / 0.6360
BasicVSR++29.61 / 0.830326.19 / 0.711823.38 / 0.5998
+ MFQE 2.029.66 / 0.832226.27 / 0.716923.44 / 0.6051
+ STDF29.68 / 0.833826.53 / 0.728923.79 / 0.6247
+ S2SVR29.82 / 0.837126.72 / 0.734623.85 / 0.6269
+ Metabit29.61 / 0.830326.19 / 0.711823.38 / 0.5997
+ Ours29.92 / 0.840726.87 / 0.741924.00 / 0.6358
", + "bbox": [ + 114, + 422, + 459, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative results of $\\times 4$ VSR,where the best and second best results are highlighted with bold and underline.", + "bbox": [ + 89, + 676, + 482, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to $CRF_{i}$ , with a negligible decrease of PSNR ( $< 0.03\\mathrm{dB}$ ), demonstrating the feasibility of replacing $CRF_{i}$ with slice type in practical. The proposed method notably improves the quality of compressed input, achieving a PSNR gain of $1.2\\mathrm{dB}$ on CRF15, while MFQE 2.0 and Metabit lead to no improvement. With similar computation cost and inference speed, our method significantly outperforms STDF, obtaining a PSNR gain of $1.09\\mathrm{dB}$ on CRF15. Compared to S2SVR, our approach takes only $61\\%$ of the parameters and $16\\%$ of the FLOPs, and achieves a throughput of 28 FPS, which underlines its practicality. In addition, our method shows robustness and generalization ability on unseen sce", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "narios (i.e., CRF18, CRF28 and CRF38), achieving up to 1.06 dB PSNR gain on CRF18. In contrast, other methods trained with mixed compression settings show sub-optimal performance. For instance, STDF and S2SVR only achieve PSNR gains of 0.16 dB and 0.76 dB at CRF18, while MFQE 2.0 shows no improvement. Quantitative results on highly compressed scenarios (i.e., CRF40, CRF45, CRF48) are included in the supplementary materials.", + "bbox": [ + 511, + 425, + 906, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results. Qualitative comparisons are provided in Fig. 6. As can be seen, MFQE 2.0 and Metabit struggle to improve the quality of compressed inputs. Both STDF and S2SVR cannot adequately remove compression artifacts (e.g., boundary of the calf), while the proposed method effectively eliminates the compression artifacts, preserving accurate edges and textures. We provide more qualitative comparisons in the supplementary materials.", + "bbox": [ + 511, + 546, + 908, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.2. Versatility Evaluation", + "text_level": 1, + "bbox": [ + 511, + 671, + 712, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the versatility in assisting practical downstream tasks, we employ the implementation that utilizes slice type for frame adaptation (as described in Sec. 3.1) to enhance compressed inputs for downstream tasks. More qualitative comparisons are included in the supplementary materials.", + "bbox": [ + 511, + 688, + 905, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video super-resolution. We adopt BasicVSR [6], IconVSR [6], and BasicVSR++ [7] as the representative baseline methods for $\\times 4$ video super-resolution (VSR), which are trained on \"clean\" data without considering compression. The evaluation is conducted on the REDS4 dataset [39] and summarized with PSNR and SSIM (the higher the better). As depicted in Tab. 2, pre-enhancing compressed inputs with Metabit fails to improve the performance of downstream VSR models. In contrast, pre-enhancing with our", + "bbox": [ + 511, + 765, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17772", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f5c6eb61718524fc7c5a8342f81d22f501d9bf5796f3c8c1cbaadf6828f79f56.jpg", + "image_caption": [ + "Figure 7. Qualitative results of $\\times 4$ VSR. Pre-enhancing with our method before VSR effectively avoids amplifying compression artifacts. While other methods cannot fully eliminate the artifacts and even severe the distortions (e.g., results of MFQE 2.0)." + ], + "image_footnote": [], + "bbox": [ + 93, + 80, + 331, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e2d88e0c6875def96096d6ffaf1d66c175b413f59da668e28aca2a8764ae853.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 80, + 413, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c1e179d8292670f767ab4404940df61ec6d0182be665f7b81c267e34583176c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 415, + 80, + 495, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/993962136e536e6f824225b3ff342976b971ea0d12452bd0a6fd75eb0b966e93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 80, + 576, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0dfc02ac5b0a2171f47befa998a68b11d1d24a62f29fb4849c61ef4d5bfb5892.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 80, + 658, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/994835e8b8fa07e8ef81d769ccdce568f21be3c75c2cd82d8ddbedbed6aed2b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 80, + 738, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0cc2618c2bb323309cbfafdc82338a25f2ac30777958036345236d75decfc9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 80, + 821, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b615ceab181c851d4c1b245b79a4b142b04e1872f3cbf7e3ab7b083a4e854ace.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 80, + 903, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69ecbf99b3c4ef8b7ee01333157240fd0cbe8db13f46305369c0b618cbf669c3.jpg", + "image_caption": [ + "Figure 8. Qualitative results of optical flow estimation. As can be seen, pre-processing with our method effectively corrects the mispredicted optical flow, especially the boundaries of moving objects (e.g., edge of the moving car)." + ], + "image_footnote": [], + "bbox": [ + 91, + 261, + 207, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa82bce0fdd460fcbcac107fd8d28b3f77928da1a28ff658478c587b9fdb1610.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 261, + 325, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6ec442c6680ffab03889346826d65d901f29c1ee83cd17c4340610f167a6187.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 261, + 439, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6cbfefe86bec18a9fa67df32607ae02d0ade8d8bd505c8c57ed512914783aeb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 261, + 555, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c10424d9984437e85177ada66863d9de1fd0344ddc8e296beb06f8a0d0dda1dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 261, + 671, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c09078aea4a8712df83c5d428948e802ecf2b6f25d4eecae23b74551efed6ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 261, + 787, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d09086fd4c04906dc99c3fe8b32c9f82a3f55a554a69a68f4582fe5db27d784.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 261, + 903, + 323 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d0e005c05014d4564ea2063fb891029d411e195d1c23d7590c2d8b83de7a853d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCRF15CRF25CRF35
EPE↓/ F1-all↓EPE↓/ F1-all↓EPE↓/ F1-all↓
RAFT5.26 / 17.817.37 / 22.1316.73 / 44.70
+ MFQE 2.05.32 / 17.837.27 / 21.9216.68 / 44.74
+ STDF5.34 / 17.937.13 / 22.1615.92 / 44.04
+ S2SVR5.22 / 17.716.90 / 21.5715.73 / 44.62
+ Metabit5.32 / 17.867.33 / 22.0716.69 / 44.28
+ Ours5.20 / 17.696.52 / 20.9914.84 / 42.56
DEQ3.99 / 13.715.40 / 17.3313.94 / 41.63
+ MFQE 2.03.97 / 13.735.14 / 17.0614.06 / 41.72
+ STDF4.08 / 13.845.27 / 17.3813.52 / 40.75
+ S2SVR4.01 / 13.695.22 / 16.9913.74 / 41.65
+ Metabit3.92 / 13.785.30 / 17.3313.84 / 41.21
+ Ours3.97 / 13.684.96 / 16.5413.09 / 39.43
KPAFlow4.46 / 16.076.71 / 20.9616.50 / 45.13
+ MFQE 2.04.42 / 15.966.71 / 20.9216.70 / 45.29
+ STDF4.52 / 16.196.96 / 21.5315.62 / 44.17
+ S2SVR4.37 / 15.966.10 / 19.7615.62 / 44.23
+ Metabit4.47 / 16.176.75 / 21.0816.68 / 44.90
+ Ours4.43 / 16.105.59 / 18.7314.83 / 41.24
", + "bbox": [ + 125, + 364, + 447, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "framework yields consistent improvement especially in scenarios of high compression (e.g., up to 0.62 dB PSNR gain with BaiscVSR++ on CRF35). Meanwhile, our method significantly outperforms MFQE 2.0 and STDF, achieving PSNR gains of 0.25 dB and 0.23 dB over MFQE 2.0 and STDF on BasicVSR/CRF15, respectively. Compared with S2SVR, the proposed method offers more effective support to VSR models with lower complexity. As shown in Fig. 7, performing VSR on compressed data inevitably amplifies compression artifacts (e.g., the 1st column), while results pre-enhanced with our method maintain accurate edges and textures, avoiding distortions seen in other methods.", + "bbox": [ + 88, + 642, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optical flow estimation. We adopt RAFT [49], DEQ [3], and KPAFlow [35] as baseline models for optical flow estimation. Evaluation on the KITTI-2015 dataset [21] is summarized with EPE (end-point-error) and F1-all loss, where lower values indicate better accuracy. As shown in Tab. 3,", + "bbox": [ + 88, + 824, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9479e5e22a2236d3a9c02c2064b7046e4d899952767b265d34b94838c44125c6.jpg", + "table_caption": [ + "Table 3. Quantitative results of optical flow estimation, where we highlight the best and second best results with **bold** and **underline**." + ], + "table_footnote": [], + "table_body": "
MethodCRF15CRF25CRF35
Avg↑/J↑/F↑Avg↑/J↑/F↑Avg↑/J↑/F↑
STCN85.07 / 81.83 / 88.3284.35 / 80.96 / 87.7479.20 / 76.04 / 82.37
+ MFQE 2.084.96 / 81.71 / 88.2184.30 / 80.92 / 87.6979.28 / 76.11 / 82.44
+ STDF85.01 / 81.73 / 88.2884.23 / 80.97 / 87.5079.77 / 76.52 / 83.02
+ S2SVR85.17 / 81.93 / 88.4184.46 / 81.20 / 87.7280.04 / 76.88 / 83.20
+ Metabit84.56 / 80.97 / 88.1483.86 / 80.18 / 87.5579.03 / 75.65 / 82.40
+ Ours85.21 / 81.99 / 88.4484.63 / 81.42 / 87.8581.57 / 78.46 / 84.69
DeAoT85.90 / 82.89 / 88.9185.18 / 82.37 / 88.0082.87 / 79.86 / 85.88
+ MFQE 2.085.86 / 82.84 / 88.8885.20 / 82.38 / 88.0382.86 / 79.86 / 85.85
+ STDF85.83 / 82.80 / 88.8785.18 / 82.27 / 88.0982.90 / 79.92 / 85.89
+ S2SVR86.05 / 83.09 / 89.0185.05 / 82.07 / 88.0482.64 / 79.63 / 85.65
+ Metabit85.47 / 82.04 / 88.9084.95 / 81.57 / 88.3382.32 / 79.04 / 85.59
+ Ours86.08 / 83.13 / 89.0385.31 / 82.38 / 88.2582.88 / 79.83 / 85.92
QDMN85.16 / 82.20 / 88.1184.16 / 81.20 / 87.1279.39 / 76.61 / 82.18
+ MFQE 2.085.13 / 82.20 / 88.0684.15 / 81.18 / 87.1379.51 / 76.75 / 82.27
+ STDF85.32 / 82.38 / 88.2783.36 / 80.44 / 86.2779.64 / 76.69 / 82.59
+ S2SVR85.28 / 82.32 / 88.2383.64 / 80.65 / 86.6379.02 / 76.15 / 81.89
+ Metabit84.50 / 81.14 / 87.8783.68 / 80.30 / 87.0679.47 / 76.41 / 82.52
+ Ours85.34 / 82.41 / 88.2784.37 / 81.42 / 87.3279.78 / 76.92 / 82.65
", + "bbox": [ + 517, + 362, + 903, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Quantitative results of VOS, where the best and second best results are highlighted with bold and underline.", + "bbox": [ + 511, + 603, + 903, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our method consistently reduces the EPE and F1-all loss across all baseline models, demonstrating its effectiveness in improving optical flow estimation. In contrast, methods such as MFQE 2.0, STDF and Metabit fail to deliver consistent improvements. For instance, MFQE 2.0 fails to improve the performance of RAFT on CRF15, STDF and Metabit detrimentally affects the performance of DEQ and KPAFlow on CRF15. Visualizations of predicted optical flow are shown in Fig. 8, where inaccurate boundaries are highlighted with red arrows. As can be seen, optical flow estimated from compressed inputs contains inaccurate boundaries, especially near-motion ones. The proposed method helps to deliver more accurate results in these regions compared to others. For instance, it effectively corrects the optical flow near the car that was mispredicted by DEQ, while MFQE 2.0 and S2SVR provide limited improvement.", + "bbox": [ + 509, + 642, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video object segmentation. For video object segmentation", + "bbox": [ + 513, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17773", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff5ff14a1891bd7c150379f6b76b20dbab27f79bd5c4045498a8cf3efc05f64e.jpg", + "image_caption": [ + "STCN" + ], + "image_footnote": [], + "bbox": [ + 93, + 80, + 209, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2fecfcbf265c390fa5c1b1935df9222a35f4bb360f2bc39096046aa6f6d3252.jpg", + "image_caption": [ + "+ MFQE 2.0" + ], + "image_footnote": [], + "bbox": [ + 210, + 80, + 323, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/69abf39d07991c266f382e0ef5b2ed082e89560ca6bdb79824d20c24e2b45eab.jpg", + "image_caption": [ + "+STDF" + ], + "image_footnote": [], + "bbox": [ + 326, + 80, + 441, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29d9116db90f2ad277f8a8a2503bdd1ad3e70d2fe8b603ef04dafcc6b0b7ae9f.jpg", + "image_caption": [ + "+ S2SVR" + ], + "image_footnote": [], + "bbox": [ + 442, + 80, + 555, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/285533b831b80dec5f2bd8146554e3be0953d9476c16712daba7b15aa0c33006.jpg", + "image_caption": [ + "+Metabit" + ], + "image_footnote": [], + "bbox": [ + 558, + 80, + 671, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d976055777897c391077cbf53cf4797495eac3730ccd0db47c248535761bbe4.jpg", + "image_caption": [ + "+Ours" + ], + "image_footnote": [], + "bbox": [ + 674, + 80, + 787, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed21269d8b4496c4a9c642558390e11dbf3b8dad7c8f9afc7467683bd0955860.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 790, + 80, + 903, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e0c90e81a7c68ecab8eec8d32477e758ccba7bf83db448c0c21b05a7da3ef72.jpg", + "image_caption": [ + "Figure 9. Qualitative results of VOS. As can be seen, directly performing VOS on compressed inputs leads to inaccurate masks, whereas pre-enhancing with our method effectively improves the accuracy, especially for the regions of irregular shapes (e.g., the windshield).", + "Masked frame", + "Figure 10. Qualitative results of video inpainting. As can be seen, pre-enhancing the compressed inputs with the proposed method helps to reduce the artifacts and color distortion in the removed region, providing more visually pleasant results." + ], + "image_footnote": [], + "bbox": [ + 91, + 188, + 207, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39e963a30d0bb188e87f7203be7d131e6cd49dc8e91fdcba0981bdce1546b408.jpg", + "image_caption": [ + "$E^2$ FGVI" + ], + "image_footnote": [], + "bbox": [ + 209, + 188, + 323, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/353d0f65ab54f34c566199b26ad022236da607bea4190039b49d117aa7871c32.jpg", + "image_caption": [ + "+ MFQE 2.0" + ], + "image_footnote": [], + "bbox": [ + 323, + 188, + 439, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71b79d6c54308e89a5dc76bd0d742ce0b298b3658dd878e54a4e811f3d8a016e.jpg", + "image_caption": [ + "+STDF" + ], + "image_footnote": [], + "bbox": [ + 442, + 188, + 555, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02add9649d9c6e82a9ee5e522da514b181cd8f884c52d19d5aa5c2558adf03e6.jpg", + "image_caption": [ + "+ S2SVR" + ], + "image_footnote": [], + "bbox": [ + 557, + 188, + 671, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/479b395931600a7b12769eed239b2d28969843af9abf1730f8fc8df5b229c082.jpg", + "image_caption": [ + "+Metabit" + ], + "image_footnote": [], + "bbox": [ + 671, + 188, + 787, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c13d21a0518fb74cda51b730621a1774293d4d5eddc95055d2f8a484d5dd541.jpg", + "image_caption": [ + "+Ours" + ], + "image_footnote": [], + "bbox": [ + 789, + 188, + 903, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(VOS), we adopt STCN [14], DeAoT [57] and QDMN [34] as representative baselines. Evaluations on DAVIS-17 val dataset [41] are summarized with the following metrics: the $\\mathcal{I}$ (average IoU), the $\\mathcal{F}$ score (boundary similarity), and the average of the above metrics (denoted as $Avg$ ). Higher values indicate better segmentation accuracy. As shown in Tab. 4, the proposed method shows the best performance in improving accuracy across VOS models. For instance, it elevates the average accuracy for up to $2.37\\%$ (79.20% to $81.57\\%$ ) on STCN at CRF35, while MFQE 2.0, STDF and S2SVR yield limited improvement of $0.08\\%$ , $0.57\\%$ and $0.84\\%$ , respectively. And Metabit provides no improvement on STCN at CRF35. The results of VOS are included in Fig. 9, where accurately segmenting objects in compressed videos is challenging for VOS baselines (e.g., inaccurate mask of the windshield predicted by STCN). Pre-enhancing the compressed videos with MFQE 2.0 and S2SVR struggles to address this issue, whereas the proposed method significantly refines the segmentation results, demonstrating its effectiveness in assisting the VOS task.", + "bbox": [ + 91, + 303, + 483, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video inpainting. We take $\\mathrm{E}^2\\mathrm{FGVI}$ [31] as the video inpainting model and perform video object removal on DAVIS-17 val dataset [41]. The qualitative results are shown in Fig. 10. As can be seen, compression-included misalignment between objects and masks hinders the ability to remove specified objects, causing color distortions (e.g., the horse region). Pre-enhancing the compressed inputs with the proposed method notably refines the artifacts and distortions, yielding more visually pleasing results.", + "bbox": [ + 89, + 606, + 483, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 755, + 254, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We start with a baseline that concatenates reference frames and current frames as input, without using codec information. We then progressively equip the baseline with MV alignment, region-aware refinement, sequence adaptation, and frame adaptation to assess their contributions.", + "bbox": [ + 89, + 777, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MV alignment. As shown in the 2nd row of Tab. 5, incorporating MV alignment yields a PSNR gain of up to 0.65 dB on CRF15, demonstrating the effectiveness of MV in", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/11afc9a2fe8cc1646e627742107f5260c90a02dd0e3da6bfa81e8ea1803109e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelCRF15CRF25CRF35
PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
Baseline41.04 / 0.978534.92 / 0.936329.25 / 0.8238
+ MV Align.41.69 / 0.982135.59 / 0.943729.95 / 0.8403
+ RA Refine.42.04 / 0.983735.70 / 0.944930.00 / 0.8427
+ Seq. Adapt.42.08 / 0.983835.76 / 0.945830.04 / 0.8444
+ Frame Adapt.42.14 / 0.983935.81 / 0.946030.09 / 0.8446
", + "bbox": [ + 517, + 301, + 901, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies on MV alignment, region-aware refinement, sequence adaptation, and frame adaptation.", + "bbox": [ + 511, + 412, + 903, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "aligning reference frames and current frame.", + "bbox": [ + 511, + 455, + 808, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Region-aware refinement. We further incorporate the region-aware refinement module to refine the features of different regions. As shown in the 3rd row of Tab. 5, it leads to notable PSNR gains of $0.35\\mathrm{dB}$ , $0.11\\mathrm{dB}$ and $0.05\\mathrm{dB}$ on CRF15, CRF25 and CRF35, respectively.", + "bbox": [ + 511, + 472, + 905, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sequence adaptation. As shown in the 4th row of Tab. 5, sequence adaptation brings PSNR gains of $0.04\\mathrm{dB}$ , $0.06\\mathrm{dB}$ and $0.04\\mathrm{dB}$ on CRF15, CRF25 and CRF35, respectively.", + "bbox": [ + 511, + 547, + 903, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frame adaptation. As shown in the 5th row of Tab. 5, frame adaptation improves PSNR by $0.06\\mathrm{dB}$ $0.05\\mathrm{dB}$ and $0.05\\mathrm{dB}$ on CRF15, CRF25 and CRF35, respectively. We further analyze its effectiveness on improving the temporal consistency in the supplementary materials.", + "bbox": [ + 511, + 593, + 905, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 683, + 633, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a versatile codec-aware enhancement framework that adaptively handles diverse compression settings and serves as a plug-and-play enhancement module to consistently boost various downstream tasks. By reusing the off-the-shelf codec information, our method minimizes additional computational costs. Compared with existing compressed video enhancement solutions, it shows superiority in both enhancement performance and robustness, making it possible to deploy pre-trained models on compressed videos without a significant performance drop.", + "bbox": [ + 511, + 704, + 906, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We acknowledge funding from the National Natural Science Foundation of China under Grants 62131003 and 62021001.", + "bbox": [ + 511, + 854, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17774", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Video developer report 2022-2023. https://bitmovin.com/downloads/ assets/bitmovin-7th-video-developer-report-2023-2024.pdf. 5", + "[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In CVPR, pages 3366-3375, 2017. 2", + "[3] Shaojie Bai, Zhengyang Geng, Yash Savani, and J. Zico Kolter. Deep equilibrium optical flow estimation. In CVPR, 2022. 7", + "[4] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 3", + "[5] Yuxuan Cai, Hongjia Li, Geng Yuan, Wei Niu, Yanyu Li, Xulong Tang, Bin Ren, and Yanzhi Wang. Yolobile: Realtime object detection on mobile devices via compression-compiler co-design. In AAAI, pages 955–963, 2021. 1", + "[6] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, pages 4947-4956, 2021. 6", + "[7] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, pages 5972-5981, 2022. 6", + "[8] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In CVPR, pages 5962-5971, 2022. 1", + "[9] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In ICIP, pages 168-172, 1994. 5", + "[10] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In CVPR, pages 8064-8073, 2021. 3", + "[11] Peilin Chen, Wenhan Yang, Long Sun, and Shiqi Wang. When bitstream prior meets deep prior: Compressed video super-resolution with learning from decoding. In ACM MM, pages 1000–1008, 2020. 1, 2", + "[12] Peilin Chen, Wenhan Yang, Meng Wang, Long Sun, Kangkang Hu, and Shiqi Wang. Compressed domain deep video super-resolution. IEEE Transactions on Image Processing, 30:7156-7169, 2021. 1, 2", + "[13] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In CVPR, pages 11030-11039, 2020. 2", + "[14] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 8", + "[15] Yuanying Dai, Dong Liu, and Feng Wu. A convolutional neural network approach for post-processing in hevc intra coding. In MultiMedia Modeling: 23rd International Con" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference, MMM 2017, Reykjavik, Iceland, January 4-6, 2017, Proceedings, Part I 23, pages 28-39. Springer, 2017. 2", + "[16] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. In AAAI, pages 10696-10703, 2020. 1, 2, 5, 6", + "[17] Cunhui Dong, Haichuan Ma, Zhuoyuan Li, Li Li, and Dong Liu. Temporal wavelet transform-based low-complexity perceptual quality enhancement of compressed video. IEEE TCSVT, pages 1-1, 2023. 2", + "[18] Max Ehrlich, Jon Barker, Namitha Padmanabhan, Larry Davis, Andrew Tao, Bryan Catanzaro, and Abhinav Shrivastava. Leveraging bitstream metadata for fast, accurate, generalized compressed video quality enhancement. In WACV, pages 1517-1527, 2024. 1, 2, 5, 6", + "[19] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 2", + "[20] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaeser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2020. 1", + "[21] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 7", + "[22] Zhenyu Guan, Qunliang Xing, Mai Xu, Ren Yang, Tie Liu, and Zulin Wang. Mfqe 2.0: A new approach for multi-frame quality enhancement on compressed video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3): 949-963, 2019. 1, 2, 5, 6", + "[23] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7436-7456, 2021. 2", + "[24] Hongyue Huang, Ionut Schiopu, and Adrian Munteanu. Frame-wise cnn-based filtering for intra-frame quality enhancement of hevc videos. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2100-2113, 2020. 2", + "[25] Zhijie Huang, Jun Sun, and Xiaopeng Guo. Fastconn: Towards fast and accurate spatiotemporal network for hevc compressed video enhancement. ACM Transactions on Multimedia Computing, Communications and Applications, 19 (3):1-22, 2023. 2", + "[26] Chuanmin Jia, Shiqi Wang, Xinfeng Zhang, Shanshe Wang, Jiaying Liu, Shiliang Pu, and Siwei Ma. Content-aware convolutional neural network for in-loop filtering in high efficiency video coding. IEEE Transactions on Image Processing, 28(7):3343-3356, 2019. 2", + "[27] Nanfeng Jiang, Weiling Chen, Jielian Lin, Tiesong Zhao, and Chia-Wen Lin. Video compression artifacts removal with spatial-temporal attention-guided enhancement. IEEE Transactions on Multimedia, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17775", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Feng Li, Yixuan Wu, Anqi Li, Huihui Bai, Runmin Cong, and Yao Zhao. Enhanced video super-resolution network towards compressed data. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-21, 2024. 2", + "[29] Yinxiao Li, Pengchong Jin, Feng Yang, Ce Liu, Ming-Hsuan Yang, and Peyman Milanfar. Comisr: Compression-informed video super-resolution. In ICCV, pages 2543-2552, 2021. 2", + "[30] Yunsheng Li, Lu Yuan, Yinpeng Chen, Pei Wang, and Nuno Vasconcelos. Dynamic transfer for multi-source domain adaptation. In CVPR, pages 10998-11007, 2021. 3", + "[31] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In CVPR, 2022. 8", + "[32] Jing Lin, Xiaowan Hu, Yuanhao Cai, Haoqian Wang, Youliang Yan, Xueyi Zou, Yulun Zhang, and Luc Van Gool. Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In ICML, pages 13394-13404. PMLR, 2022. 1, 2, 5, 6", + "[33] Baoyuan Liu, Min Wang, Hassan Foroosh, Marshall Tappen, and Marianna Pensky. Sparse convolutional neural networks. In CVPR, pages 806-814, 2015. 5", + "[34] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, pages 468-486, 2022. 8", + "[35] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In CVPR, pages 8906-8915, 2022. 7", + "[36] Di Ma, Fan Zhang, and David R Bull. Cvegan: a perceptually-inspired gan for compressed video enhancement. Signal Processing: Image Communication, page 117127, 2024. 2", + "[37] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2", + "[38] Sreyas Mohan, Joshua L Vincent, Ramon Manzorro, Peter Crozier, Carlos Fernandez-Granda, and Eero Simoncelli. Adaptive denoising via gaintuning. NeurIPS, 34:23727-23740, 2021. 3", + "[39] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, pages 0–0, 2019. 5, 6", + "[40] Zhaoqing Pan, Xiaokai Yi, Yun Zhang, Byeungwoo Jeon, and Sam Kwong. Efficient in-loop filtering based on enhanced deep convolutional neural networks for hevc. IEEE Transactions on Image Processing, 29:5352-5366, 2020. 2", + "[41] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 8" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Darren Ramsook and Anil Kokaram. Learnt deep hyperparameter selection in adversarial training for compressed video enhancement with a perceptual critic. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2420-2424. IEEE, 2023. 2", + "[43] Mennatullah Siam, Mostafa Gamal, Moemen Abdel-Razek, Senthil Yogamani, Martin Jagersand, and Hong Zhang. A comparative study of real-time semantic segmentation for autonomous driving. In CVPRW, pages 587-597, 2018. 1", + "[44] Mennatullah Siam, Alex Kendall, and Martin Jagersand. Video class agnostic segmentation benchmark for autonomous driving. In CVPR, pages 2825-2834, 2021. 1", + "[45] Gary J Sullivan, Pankaj N Topiwala, and Ajay Luthra. The h. 264/avc advanced video coding standard: Overview and introduction to the fidelity range extensions. Applications of Digital Image Processing XXVII, 5558:454-474, 2004. 5", + "[46] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 3", + "[47] Weiheng Sun, Xiaohai He, Chao Ren, Shuhua Xiong, and Honggang Chen. A quality enhancement network with coding priors for constant bit rate video coding. Knowledge-Based Systems, 258:110010, 2022. 1, 2", + "[48] Xuebin Sun, Sukai Wang, Miaohui Wang, Shing Shin Cheng, and Ming Liu. An advanced lidar point cloud sequence coding scheme for autonomous driving. In MM, pages 2793-2801, 2020. 1", + "[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020.", + "[50] Jianyi Wang, Mai Xu, Xin Deng, Liquan Shen, and Yuhang Song. Mw-gan+ for perceptual quality enhancement on compressed video. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4224-4237, 2021. 2", + "[51] Shiyao Wang, Hongchao Lu, and Zhidong Deng. Fast object detection in compressed video. In ICCV, pages 7104-7113, 2019. 1", + "[52] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Hutchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In CVPR, pages 2012-2021, 2023. 2", + "[53] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 3", + "[54] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. NeurIPS, 32, 2019. 2", + "[55] Ren Yang, Mai Xu, Zulin Wang, and Tianyi Li. Multi-frame quality enhancement for compressed video. In CVPR, pages 6664-6673, 2018. 1, 2", + "[56] Ren Yang, Mai Xu, Tie Liu, Zulin Wang, and Zhenyu Guan. Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology, 29(7): 2039-2054, 2019. 2", + "[57] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation. In NeurIPS, 2022. 8" + ], + "bbox": [ + 514, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17776", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Li Yu, Wenshuai Chang, Shiyu Wu, and Moncef Gabbouj. End-to-end transformer for compressed video quality enhancement. IEEE Transactions on Broadcasting, 70(1):197-207, 2024. 2", + "[59] Haichao Zhang, Jianchao Yang, Yanning Zhang, Nasser M Nasrabadi, and Thomas S Huang. Close the loop: Joint blind image restoration and recognition with sparse representation prior. In ICCV, pages 770-777, 2011. 1", + "[60] Hengsheng Zhang, Xueyi Zou, Jiaming Guo, Youliang Yan, Rong Xie, and Li Song. A codec information assisted framework for efficient compressed video super-resolution. In ECCV, pages 220-235. Springer, 2022. 2", + "[61] Saiping Zhang, Luis Herranz, Marta Mrak, Marc Górriz Blanch, Shuai Wan, and Fuzheng Yang. Dcngan: A deformable convolution-based gan with qp adaptation for perceptual quality enhancement of compressed video. In ICASSP, pages 2035-2039. IEEE, 2022. 2", + "[62] Zhengdong Zhang and Vivienne Sze. Fast: A framework to accelerate super-resolution processing on compressed videos. In CVPRW, pages 19-28, 2017. 2", + "[63] Hengrun Zhao, Bolun Zheng, Shanxin Yuan, Hua Zhang, Chenggang Yan, Liang Li, and Gregory Slabaugh. Cbren: Convolutional neural networks for constant bit rate video quality enhancement. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4138-4149, 2022. 2", + "[64] Qiang Zhu, Jinhua Hao, Yukang Ding, Yu Liu, Qiao Mo, Ming Sun, Chao Zhou, and Shuyuan Zhu. Cpga: Coding priors-guided aggregation network for compressed video quality enhancement. In CVPR, pages 2964-2974, 2024. 2" + ], + "bbox": [ + 91, + 92, + 480, + 501 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17777", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Plug-and-Play Versatile Compressed Video Enhancement/e1321f09-cc21-461b-8861-4f42cf1e3c7e_model.json b/2025/Plug-and-Play Versatile Compressed Video Enhancement/e1321f09-cc21-461b-8861-4f42cf1e3c7e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..672dbd1c74687cd8df2934fab96506d83cf0f8b6 --- /dev/null +++ b/2025/Plug-and-Play Versatile Compressed Video Enhancement/e1321f09-cc21-461b-8861-4f42cf1e3c7e_model.json @@ -0,0 +1,3170 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "Plug-and-Play Versatile Compressed Video Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.181, + 0.704, + 0.2 + ], + "angle": 0, + "content": "Huimin Zeng Jiacheng Li Zhiwei Xiong*" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.2, + 0.688, + 0.218 + ], + "angle": 0, + "content": "University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.219, + 0.758, + 0.234 + ], + "angle": 0, + "content": "{zenghuimin, jciee}@mail.ustc.edu.cn zwxiong@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.664 + ], + "angle": 0, + "content": "As a widely adopted technique in data transmission, video compression effectively reduces the size of files, making it possible for real-time cloud computing. However, it comes at the cost of visual quality, posing challenges to the robustness of downstream vision models. In this work, we present a versatile codec-aware enhancement framework that reuses codec information to adaptively enhance videos under different compression settings, assisting various downstream vision tasks without introducing computation bottleneck. Specifically, the proposed codec-aware framework consists of a compression-aware adaptation (CAA) network that employs a hierarchical adaptation mechanism to estimate parameters of the frame-wise enhancement network, namely the bitstream-aware enhancement (BAE) network. The BAE network further leverages temporal and spatial priors embedded in the bitstream to effectively improve the quality of compressed input frames. Extensive experimental results demonstrate the superior quality enhancement performance of our framework over existing enhancement methods, as well as its versatility in assisting multiple downstream tasks on compressed videos as a plug-and-play module. Code and models are available at https://huimin-zeng.github.io/PnP-VCVE/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.686, + 0.222, + 0.701 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.485, + 0.875 + ], + "angle": 0, + "content": "With the flower booming of short video platforms, video has become one of the most popular multimedia formats. In addition to distributing visual content, in practical scenarios (e.g., autonomous driving [43, 48]), it is common to upload the captured videos to the cloud end for further visual analysis and downstream applications (e.g., object detection [5, 51] and segmentation [20, 44]). However, due to the bandwidth constraint during transmission, these videos are typically compressed with varying levels, resulting in poor visual quality and suboptimal performance in downstream tasks [22, 59] (e.g., inaccurate segmentation bound" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.268, + 0.907, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Figure 1. The proposed codec-aware enhancement framework reuses codec information to adaptively enhance videos across different compression settings, while assisting in various downstream tasks in a plug-and-play manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.907, + 0.594 + ], + "angle": 0, + "content": "aries in Fig. 1). Given the crucial role of videos in data transmission, there is a critical need for a versatile solution to enhance videos of diverse compression levels and effectively support various downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Existing video enhancement methods are hard to respond to these demands. Specifically, to effectively enhance videos of different compression levels, previous methods [11, 12, 16, 22, 32, 55] employ separate enhancement models for each compression level, which is inflexible occurring unseen compression levels. Recent approaches [8, 18, 47] consider this issue as the generalization ability across diverse compression levels, therefore randomly selecting inputs of different compression levels during training. However, such a training strategy is compression-agnostic and offers limited improvement. Most importantly, the aforementioned methods focus primarily on improving perceptual quality, neglecting the need to assist in downstream tasks in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Based on the mismatch between the versatility demand and existing solutions, here we summarize the following criteria of a favorable solution: 1) adaptively enhance videos of varying compression levels with a single model; 2) effectively assist various downstream tasks on compressed videos in a plug-and-play manner; and 3) given the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17767" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "practical scenarios where real-time processing is required, it should meet the above objectives without causing a computation bottleneck. To achieve this, we introduce a codec-aware enhancement framework (as shown in Fig. 1) that reuses codec information embedded in the bitstream. By incorporating compression factors, the framework dynamically adjusts its parameters to flexibly enhance inputs of different compression levels. By reusing motion vectors and partition maps, it efficiently aggregates temporal and spatial clues without introducing redundant computations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.244, + 0.485, + 0.574 + ], + "angle": 0, + "content": "Specifically, our codec-aware enhancement framework comprises a compression-aware adaptation (CAA) network and a bitstream-aware enhancement (BAE) network. The CAA network serves as a \"meta\" network that dynamically adjusts the parameters of the subsequent BAE network. A hierarchical adaptation mechanism is proposed to first estimate sequence-adaptive parameters based on the sequence compression level, and then re-weight these parameters according to the frame compression level, thereby achieving a BAE network tailored for each input frame. The frame-adaptive BAE network conducts motion vector alignment to aggregate intra-frame information and provide useful clues for the current frame. Subsequently, based on the region complexity indicated by partition maps, the region-aware refinement assigns independent filters for different regions, achieving flexible enhancement for different regions. Comprehensive experiments are conducted to demonstrate the superiority of our method in improving the quality of compressed videos, and the effectiveness of assisting in various downstream tasks (i.e., video super-resolution, optical flow estimation, and video object segmentation). Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.575, + 0.483, + 0.635 + ], + "angle": 0, + "content": "- We present a codec-aware framework for versatile compressed video enhancement, which adaptively enhances input videos of different compression levels and supports a wide range of downstream vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.636, + 0.483, + 0.724 + ], + "angle": 0, + "content": "- We develop a compression-aware adaptation (CAA) network and a bitstream-aware enhancement (BAE) network that utilize the off-the-shelf codec information, contributing to generalizing across different compression settings and boosting the enhancement performance with a unified framework." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.726, + 0.483, + 0.785 + ], + "angle": 0, + "content": "- Experimental results show the superiority of our method over existing enhancement methods, and its effectiveness in serving as a plug-and-play enhancement module to assist in downstream tasks." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.575, + 0.483, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.799, + 0.233, + 0.815 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.382, + 0.837 + ], + "angle": 0, + "content": "2.1. Compressed Video Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Existing compressed video enhancement methods can be categorized into in-loop and post-processing methods. Although in-loop methods [15, 24, 26, 40] effectively improve the quality of reconstructed frames, they embed filters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.44 + ], + "angle": 0, + "content": "in the encoding and decoding loops, therefore not suitable for enhancing already compressed videos. While the post-processing methods [17, 25, 27, 36, 42, 47, 50, 56, 58, 61, 63, 64] provide more practical solutions to enhance compressed videos by placing filters at the decoder side. Observing the quality fluctuation across frames, MFQE [55] locates the peak quality frame (PQF) with an SVM-based detector and proposes a multi-frame quality enhancement mechanism to enhance non-PQFs. MFQE 2.0 [22] further designs a BiLSTM-based detector and performs multiframe quality enhancement for both non-PQF and PQF. To address the inaccuracies in optical flow estimation from compressed videos, STDF [16] proposes estimating the offset field using spatio-temporal deformable convolution. S2SVR [32] introduces a sequence-to-sequence network to model long-range dependencies within frames. The aforementioned methods inflexibly equip a separate model for each compression level, while we propose to adaptively handle diverse inputs with a single unified model. Recent methods [18, 64] utilize spatial priors from bitstream to address multiple compression levels, however, they only consider I/P-frames, whereas we design a hierarchical adaptation mechanism to address all types of frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.447, + 0.839, + 0.462 + ], + "angle": 0, + "content": "2.2. Codec-Aware Video Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.909, + 0.724 + ], + "angle": 0, + "content": "Some works in video super-resolution (VSR) [11, 12, 28, 29, 52, 60, 62] explore ways of using codec information such as motion estimation and spatial prior for reconstruction. COMISR [29] focuses on reducing accumulated warping errors caused by the random locations of the intra-frame from compressed video frames. Chen et al. [11] employ motion vectors to build the temporal relationship and suppress coding artifacts. CVCP [12] utilizes motion vectors and spatial priors with a guided soft alignment scheme and guided SFT layer, respectively. CIAF [60] leverages motion vectors and residuals to model temporal relationships and skip redundant computations, respectively. Despite leveraging codec information, these methods focus on a single task (i.e., VSR). In contrast, our method not only shows competitive performance in VSR (see supplementary materials), but also effectively supports a range of downstream tasks, which is not explored by the aforementioned works." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.757, + 0.746 + ], + "angle": 0, + "content": "2.3. Dynamic Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Instead of setting separate models for different inputs, the early mixture of expert (MoE) structure [2, 19, 37] constructs parallel network branches and selectively executes branches to obtain the weighted outputs. Instead of increasing the number of parallel branches, the dynamic parameters ensemble strategy [2, 13, 19, 23, 37, 54] presets parallel expert layers and selectively fusing their parameters to promote the network capability and generalization ability, therefore serving as an efficient alternative to MoE. To promote the generalization ability of pre-trained models, Gain-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17768" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.476, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.259, + 0.483, + 0.33 + ], + "angle": 0, + "content": "Figure 2. Hierarchical structure of quality adjustment, where frames are divided into multiple groups of pictures (GOP). The Constant Rate Factor (CRF) affects video quality at both sequence and frame levels. An increase in the CRF value indicates a reduction in video quality (e.g., lower PSNR values)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.345, + 0.483, + 0.542 + ], + "angle": 0, + "content": "tune [38] proposes to predict a single multiplicative scaling parameter for each channel according to test samples, thus modifying static models to test-adaptive ones. Li et al. [30] handle the conflicts between the domain-agnostic model and multiple target domains with dynamic transfer, which is simply modeled by combining residual matrices and a static convolution matrix. DRConv [10] divides the input image into different regions with a learnable mask and assigns multiple filters for these regions, which enhances the feature representation ability without introducing a noticeable computation burden. Instead of searching and learning conditions for dynamic parameters, we leverage priors embedded in the bitstream as the condition." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.55, + 0.231, + 0.567 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.483, + 0.664 + ], + "angle": 0, + "content": "We take H.264 [53] as a representative standard to analyze available codec information. Note recent codecs [4, 46] also provide similar priors (i.e., CRF, motion vector, partition map), therefore assuring the applicability of our method. To reduce transmission bandwidth, codecs compress videos by adjusting quality and reducing redundancy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.67, + 0.383, + 0.687 + ], + "angle": 0, + "content": "3.1. Hierarchical Quality Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Video quality is commonly influenced by the constant rate factor (CRF), which involves hierarchical adjustment for both sequence-wise and frame-wise compression." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Sequence-wise CRF. The CRF ranges from 0 to 51 to balance compression efficiency and visual quality. A higher CRF results in more compact output but increased pixel loss (e.g., the average PSNR of CRF35 is much lower than CRF15 in Fig. 2). By considering the sequence-wise CRF (denoted as \\( CRF_{s} \\)), the enhancement network can be tailored to handle videos of different compression levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Frame-wise CRF As shown in Fig. 2, a video sequence is divided into multiple groups of pictures (GOP) and further categorized as I-frames, P-frames, or B-frames. The CRF value of each frame (denoted as \\( CRF_{i} \\)) is dynamically" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "adjusted based on \\( CRF_{s} \\) so that lower \\( CRF_{i} \\) is assigned for I/P frames to maintain quality and higher \\( CRF_{i} \\) for B frames for compact representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Inspired by the hierarchical quality adjustment paradigm, we design a hierarchical adaptation paradigm that first performs sequence adaptation to predict network parameters based on \\( CRF_{s} \\), and then re-weights these parameters according to \\( CRF_{i} \\) for frame adaptation. In practical scenarios where \\( CRF_{i} \\) is unavailable (e.g., limited access to full bitstream), the proposed method can instead use slice type (I/P/B) for frame adaptation, which is demonstrated to yield similar performance in Sec. 5.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.278, + 0.731, + 0.294 + ], + "angle": 0, + "content": "3.2. Redundancy Reduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.905, + 0.357 + ], + "angle": 0, + "content": "To reduce redundancy and improve the entropy of the bitstream,Codec block-wisely perform motion estimation to model the intra-frame correlations and embed the correlations in the bitstream for decoding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.906, + 0.508 + ], + "angle": 0, + "content": "Partition map. As shown in Fig. 1, different regions of each frame are partitioned into blocks of varying sizes (e.g., H.264 provides macroblocks of \\(16 \\times 16\\), \\(16 \\times 8\\), \\(8 \\times 16\\), and \\(8 \\times 8\\)) according to the texture complexity. Flat regions (e.g., the ground) can tolerate higher quantization errors and are therefore divided into blocks of large size, while complex regions (e.g., leaves and fence) take smaller blocks to maintain details. To effectively enhance regions of different complexity, we propose dynamically assigning filters based on the partition map that indicates region complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.906, + 0.63 + ], + "angle": 0, + "content": "Motion vector. Motion vectors are utilized in decoding to aggregate information from reference frames and propagate information of current frame. As illustrated in Fig. 1, they describe the relationship between current frame and its reference frames in a block-wise manner. Although motion vectors can be noisy and are less precise than optical flow, they effectively align reference frames with current frame, serving as a cost-effective alternative for optical flow." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.639, + 0.874, + 0.655 + ], + "angle": 0, + "content": "4. Codec-Aware Enhancement Framework" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.661, + 0.625, + 0.675 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.906, + 0.801 + ], + "angle": 0, + "content": "As shown in Fig. 3(a), the proposed method comprises a compression-aware adaption (CAA) network \\(\\mathcal{G}_{\\phi}\\) and a bitstream-aware enhancement (BAE) network \\(\\mathcal{F}_{\\theta_i}\\). The CAA network employs a hierarchical compression adaptation mechanism to estimate parameters for the frame-adaptive BAE network, which then aggregates intra-frame information and performs region-aware refinement to enhance the input compressed frame." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.806, + 0.868, + 0.822 + ], + "angle": 0, + "content": "4.2. Compression-Aware Adaptation Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To handle sequences of varying compression levels and quality fluctuations across frames, as illustrated in Fig. 3(a), the CAA network \\(\\mathcal{G}_{\\phi}\\) utilizes \\(CRF_{s}\\) for sequence adaptation to estimate sequence-adaptive parameters, and perform frame adaptation to refine these parameters based on \\(CRF_{i}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17769" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.084, + 0.856, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.271, + 0.318, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.271, + 0.495, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.271, + 0.651, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.272, + 0.853, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.38, + 0.907, + 0.437 + ], + "angle": 0, + "content": "Figure 3. The proposed Codec-Aware Enhancement Framework consists of two sub-networks: 1) the Compression-Aware Adaptation (CAA) Network, which hierarchically applies sequence adaptation and frame adaptation to dynamically adjust parameters of the enhancement network; and 2) Bitstream-Aware Enhancement (BAE) Network, which leverages motion vectors to align frames and conducts region-aware refinement to flexibly enhance regions of different complexity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.484, + 0.603 + ], + "angle": 0, + "content": "Sequence adaptation. To ensure robust performance across multiple compression settings without increasing complexity, we propose estimating sequence-adaptive parameters for the enhancement network instead of fusing features from separate submodels. As shown in Fig. 3(b), parallel expert layers \\(\\{f_{\\theta_1}, f_{\\theta_2}, \\dots, f_{\\theta_N}\\}\\), which share the same architecture but have independent parameters, serve as the basis for parameter combination. The sequence-wise \\(CRF_s\\) is adopted as the condition to re-weight parameters of these expert layers, which is expressed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.603, + 0.483, + 0.645 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta_ {s}} = \\mathcal {G} _ {\\phi_ {s}} \\left(C R F _ {s}, \\left\\{f _ {\\theta_ {1}}, f _ {\\theta_ {2}}, \\dots , f _ {\\theta_ {N}} \\right\\}\\right) = \\sum_ {n = 1} ^ {N} w _ {n} f _ {\\theta_ {n}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.645, + 0.483, + 0.825 + ], + "angle": 0, + "content": "where \\( f_{\\theta_s} \\) and \\( \\mathcal{G}_{\\phi_s} \\) denote the sequence-adaptive expert layer and the sequence-wise weight generator, respectively. \\( w_n \\) denotes the weight for each expert layer. We set \\( N = 6 \\) (see ablation studies in supplementary materials) and visualize \\( w_n \\) against different \\( CRF_s \\) in Fig. 4, which shows that each expert layer has a distinct preference for specific \\( CRF_s \\). Compared to MoE that re-weights output features, re-weighting expert layer parameters (as shown in Eq. 1) is computationally efficient and comparable to the network constructed with a single expert layer. Note that \\( CRF_s \\) is constant for frames within the same sequence, \\( f_{\\theta_s} \\) is predicted only once and reused by subsequent frames." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Frame adaptation. To flexibly enhance frames with different visual quality, we propose to re-weight the sequence-based \\( f_{\\theta_s} \\) using frame-wise \\( CRF_{i} \\). We attribute the quality fluctuation between the sequence and current frame to the disparity between \\( CRF_{s} \\) and \\( CRF_{i} \\), which can be ad" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.447, + 0.859, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.59, + 0.907, + 0.618 + ], + "angle": 0, + "content": "Figure 4. Visualization of \\( w_{n} \\) against different \\( CRF_{s} \\), where each expert shows a distinct preference for specific \\( CRF_{s} \\) values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.71 + ], + "angle": 0, + "content": "dressed by introducing a set of frame-wise auxiliary parameters \\(\\triangle \\theta_{i}\\). As shown in Fig. 3(c), the auxiliary parameters \\(\\triangle \\theta_{i}\\) that conditioned on \\(CRF_{i}\\) re-weights the sequence-adaptive \\(f_{\\theta_s}\\) to obtain the frame-adaptive expert layer \\(f_{\\theta_i}\\) which is expressed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.714, + 0.905, + 0.731 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta_ {i}} = \\mathcal {G} _ {\\phi_ {i}} \\left(C R F _ {i}, f _ {\\theta_ {s}}\\right) = f _ {\\theta_ {s} + \\triangle \\theta_ {i}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.825 + ], + "angle": 0, + "content": "where \\(f_{\\theta_i}\\) and \\(\\mathcal{G}_{\\phi_i}\\) denote the estimated frame-adaptive expert layer and the frame-wise parameters generator, respectively. As shown by the black dashed lines in Fig. 3(a), the obtained \\(f_{\\theta_i}\\) is used to construct the enhancement blocks, resulting in the frame-adaptive BAE network \\(\\mathcal{F}_{\\theta_i}\\) (introduced in the following Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.863, + 0.849 + ], + "angle": 0, + "content": "4.3. Bitstream-Aware Enhancement Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To leverage high-quality frames and propagate information, the BAE network \\(\\mathcal{F}_{\\theta_i}\\) utilizes motion vectors to align reference frames with the current frame. Meanwhile, the par" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17770" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.232, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.14, + 0.216, + 0.15 + ], + "angle": 0, + "content": "Reference frame" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.089, + 0.344, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.14, + 0.297, + 0.151 + ], + "angle": 0, + "content": "\\(h_i\\)" + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.089, + 0.455, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.139, + 0.408, + 0.15 + ], + "angle": 0, + "content": "\\(\\hat{h}_i\\)" + }, + { + "type": "image", + "bbox": [ + 0.121, + 0.152, + 0.231, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.201, + 0.207, + 0.214 + ], + "angle": 0, + "content": "\\(\\mathcal{S}(M_i^{8\\times 8},h_i)\\)" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.152, + 0.344, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.201, + 0.323, + 0.215 + ], + "angle": 0, + "content": "\\(\\mathcal{S}(M_i^{16\\times 16},h_i)\\)" + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.152, + 0.455, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.201, + 0.432, + 0.215 + ], + "angle": 0, + "content": "\\(\\mathcal{S}(M_i^{16\\times 8},h_i)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.219, + 0.485, + 0.276 + ], + "angle": 0, + "content": "Figure 5. Visualization of features in region-aware refinement, where \\( h_i \\) and \\( \\hat{h}_i \\) indicate the input and output features, respectively. The refined features are denoted in the format of \\( S(M_i^{type}, h_i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.483, + 0.319 + ], + "angle": 0, + "content": "tition map serves as spatial complexity guidance to enable flexible enhancement of different regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Motion vector alignment. Since the motion vectors roughly model the temporal relationship in a block-wise manner, for each block of the current frame shown in Fig. 3(d), motion vectors locate blocks with similar content in the previous and future reference features (highlighted with red and blue boxes). The warped reference features are concatenated with current frame along the channel dimension as input of the BAE network, expressed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.443, + 0.482, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {i} = \\left[ M V \\left(h _ {i} ^ {p}\\right), M V \\left(h _ {i} ^ {f}\\right), x _ {i} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.556 + ], + "angle": 0, + "content": "where \\(\\hat{x}_i\\), \\(h_i^p\\) and \\(h_i^f\\) denote the current input frame, enhanced features of previous and future reference frames, respectively. \\(MV\\) denotes warping reference features based on motion vectors. \\([,]\\) denotes concatenation along channel dimension. Bilinear interpolation is adopted for the case that the offsets of motion vectors are not integers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.484, + 0.706 + ], + "angle": 0, + "content": "Region-aware refinement. To effectively enhance regions of different complexity, we propose to dynamically assign different filters for regions based on the partition map. As shown in Fig. 3(e), the block-based partition map \\( P_{i} \\) is decoupled into multiple binary masks \\( \\{M_i^{8\\times 8}, M_i^{16\\times 16}, \\dots, M_i^{type}\\} \\) according to the size of macroblocks, allowing separate refinement of regions using sparse convolution [33]. The output is defined as the sum of frame-adaptive extracted features and separately region-aware refined features, depicted as follows," + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.706, + 0.482, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\hat {h} _ {i} = \\mathcal {F} _ {\\theta_ {i}} \\left(\\hat {x} _ {i}, P _ {i}\\right) = f _ {\\theta_ {i}} * h _ {i} + \\sum_ {t y p e = 1} ^ {M} \\mathcal {S} \\left(M _ {i} ^ {t y p e}, h _ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{h}_i\\) indicates the output features. \\(S\\) denotes the operations applying sparse convolution guided by mask \\(M_i^{type}\\) to refine input features. In H.264 standard, three types of macroblocks are used \\((16\\times 16,8\\times 16 / 16\\times 8,\\) and \\(8\\times 8)\\), thus \\(M\\) is set to 3. Features in region-aware refinement are visualized in Fig 5, where the refined features are denoted with operations like \\(S(M_{i}^{8\\times 8},h_{i})\\). As can be seen, the output features \\(\\hat{h}_i\\) contain more fine-grained and high-frequency details than \\(h_i\\). Meanwhile, refined features provide distinct activations for different regions. For instance, \\(S(M_i^{8\\times 8},h_i)\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.122 + ], + "angle": 0, + "content": "focuses on static objects (e.g., trees) while \\(\\mathcal{S}(M_i^{8\\times 16},h_i)\\) focuses on moving objects (e.g., the bus)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.129, + 0.655, + 0.143 + ], + "angle": 0, + "content": "4.4. Loss function" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.148, + 0.906, + 0.207 + ], + "angle": 0, + "content": "We adopt Charbonnier penalty loss [9] as the loss function and train the proposed codec-aware enhancement framework in an end-to-end manner. The specific loss function is expressed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.21, + 0.905, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\sqrt {\\left\\| y _ {i} - \\hat {y} _ {i} \\right\\| ^ {2} + \\epsilon^ {2}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.255, + 0.906, + 0.3 + ], + "angle": 0, + "content": "where \\( y_{i}, \\hat{y}_{i} \\) and \\( T \\) indicate the uncompressed ground truth, the predicted output, and the length of the input sequence. \\( \\epsilon \\) is set to \\( 1 \\times 10^{-12} \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.307, + 0.646, + 0.323 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.328, + 0.722, + 0.344 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.906, + 0.422 + ], + "angle": 0, + "content": "Compression settings. H.264 is a popular video compression standard that compresses nearly \\(85\\%\\) of internet videos [1], and tends to introduce more severe degradations than H.265 and H.266. We adopt H.264 [45] and compress videos with the \\(CRF_{s}\\) values of 15, 25, and 35." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.424, + 0.906, + 0.529 + ], + "angle": 0, + "content": "Tasks and training dataset. Our tasks involve quality enhancement and assisting downstream tasks on compressed inputs. The primary downstream tasks include video superresolution, optical flow estimation and video object segmentation, with video inpainting reported as an extension to fully evaluate the versatility. Training splits of REDS [39] and DAVIS [41] datasets are combined for training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.53, + 0.906, + 0.65 + ], + "angle": 0, + "content": "Compared methods. We compare with representative methods in compressed video enhancement, including MFQE 2.0 [22], STDF [16], S2SVR [32] and Metabit [18]. For a fair comparison, we fully retrain these methods with the same training dataset and configurations. For downstream tasks, the compressed video is first enhanced by quality enhancement methods, and then fed to downstream models for corresponding tasks and further assessment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.659, + 0.608, + 0.673 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.678, + 0.906, + 0.74 + ], + "angle": 0, + "content": "Our evaluations are two-fold: 1) verifying the quality enhancement performance on seen, unseen, and highly compressed scenarios; 2) evaluating the versatility to assist different downstream tasks on multiple compression settings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.747, + 0.807, + 0.761 + ], + "angle": 0, + "content": "5.2.1. Quality Enhancement Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Quantitative results. The results of compressed video quality enhancement are reported in Tab.1, which is evaluated on the REDS4 datasets [39] using PSNR and SSIM (the higher the better). Note the CRF values of 18, 28 and 38 are not included during training. For each method, we include the model complexity and inference speed. For our method, we report results of both applying \\( CRF_{i} \\) and its substitution with slice type (highlighted with grey). As shown in Tab. 1, leveraging slice type yields comparable performance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17771" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.082, + 0.904, + 0.199 + ], + "angle": 0, + "content": "
MethodParam/MFLOPs/GSpeed/msFPSCRF15CRF25CRF35CRF18CRF28CRF38
PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑
Input----41.04 / 0.978534.92 / 0.936329.25 / 0.823839.12 / 0.969833.18 / 0.912327.69 / 0.7725
MFQE 2.0 [22]1.6451531940.95 / 0.980634.83 / 0.937829.22 / 0.825638.97 / 0.971233.13 / 0.914027.67 / 0.7742
STDF [16]1.2745382641.15 / 0.979335.23 / 0.939829.74 / 0.835939.28 / 0.971233.58 / 0.917828.11 / 0.7853
S2SVR [32]7.43294--41.96 / 0.983435.61 / 0.944529.87 / 0.839139.88 / 0.975533.87 / 0.922328.19 / 0.7881
Metabit [18]1.6092244241.04 / 0.978534.92 / 0.936329.25 / 0.823839.11 / 0.969833.18 / 0.912327.69 / 0.7725
Ours4.5647362842.22 / 0.984235.90 / 0.946830.17 / 0.847140.17 / 0.976734.16 / 0.925828.49 / 0.7985
42.24 / 0.984235.91 / 0.946830.19 / 0.847240.18 / 0.976734.17 / 0.925828.52 / 0.7985
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.202, + 0.905, + 0.231 + ], + "angle": 0, + "content": "Table 1. Quantitative results on quality enhancement, where PSNR and SSIM (higher is better) are adopted for evaluation. The best and second best results are marked with bold and underline. Results obtained by replacing \\( CRF_{i} \\) with slice type are highlighted with grey." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.235, + 0.332, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.366, + 0.291, + 0.378 + ], + "angle": 0, + "content": "Frame 030,Clip 020,CRF35" + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.235, + 0.414, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.3, + 0.414, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.366, + 0.389, + 0.378 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.235, + 0.496, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.3, + 0.496, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.366, + 0.484, + 0.377 + ], + "angle": 0, + "content": "MFQE 2.0" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.235, + 0.577, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.3, + 0.577, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.366, + 0.555, + 0.377 + ], + "angle": 0, + "content": "STDF" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.235, + 0.659, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.3, + 0.659, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.366, + 0.639, + 0.377 + ], + "angle": 0, + "content": "S2SVR" + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.235, + 0.74, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.3, + 0.74, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.366, + 0.722, + 0.377 + ], + "angle": 0, + "content": "Metabit" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.235, + 0.822, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.3, + 0.822, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.366, + 0.797, + 0.377 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.235, + 0.905, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.3, + 0.905, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.855, + 0.366, + 0.874, + 0.376 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.383, + 0.905, + 0.411 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on quality enhancement, where our method effectively reduces the compression artifacts, achieving visually pleasant results. In contrast, the results of the compared methods still contain severe distortions (e.g., the calf in the 1st row)." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.423, + 0.46, + 0.673 + ], + "angle": 0, + "content": "
MethodCRF15CRF25CRF35
PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
BasicVSR29.24 / 0.821226.19 / 0.713123.40 / 0.6005
+ MFQE 2.029.29 / 0.823326.28 / 0.718223.46 / 0.6056
+ STDF29.31 / 0.824726.51 / 0.729323.80 / 0.6249
+ S2SVR29.45 / 0.828826.70 / 0.734623.86 / 0.6270
+ Metabit29.24 / 0.821126.19 / 0.713123.39 / 0.6005
+ Ours29.54 / 0.832826.85 / 0.741924.02 / 0.6361
IconVSR29.29 / 0.823026.19 / 0.713023.39 / 0.6003
+ MFQE 2.029.37 / 0.825426.28 / 0.718223.45 / 0.6055
+ STDF29.36 / 0.826326.52 / 0.729223.79 / 0.6248
+ S2SVR29.54 / 0.830626.71 / 0.734523.85 / 0.6269
+ Metabit29.29 / 0.823026.19 / 0.713023.39 / 0.6003
+ Ours29.63 / 0.834426.86 / 0.741824.01 / 0.6360
BasicVSR++29.61 / 0.830326.19 / 0.711823.38 / 0.5998
+ MFQE 2.029.66 / 0.832226.27 / 0.716923.44 / 0.6051
+ STDF29.68 / 0.833826.53 / 0.728923.79 / 0.6247
+ S2SVR29.82 / 0.837126.72 / 0.734623.85 / 0.6269
+ Metabit29.61 / 0.830326.19 / 0.711823.38 / 0.5997
+ Ours29.92 / 0.840726.87 / 0.741924.00 / 0.6358
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.677, + 0.483, + 0.705 + ], + "angle": 0, + "content": "Table 2. Quantitative results of \\( \\times 4 \\) VSR,where the best and second best results are highlighted with bold and underline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "to \\( CRF_{i} \\), with a negligible decrease of PSNR (\\(< 0.03\\mathrm{dB}\\)), demonstrating the feasibility of replacing \\( CRF_{i} \\) with slice type in practical. The proposed method notably improves the quality of compressed input, achieving a PSNR gain of \\( 1.2\\mathrm{dB} \\) on CRF15, while MFQE 2.0 and Metabit lead to no improvement. With similar computation cost and inference speed, our method significantly outperforms STDF, obtaining a PSNR gain of \\( 1.09\\mathrm{dB} \\) on CRF15. Compared to S2SVR, our approach takes only \\( 61\\% \\) of the parameters and \\( 16\\% \\) of the FLOPs, and achieves a throughput of 28 FPS, which underlines its practicality. In addition, our method shows robustness and generalization ability on unseen sce" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.907, + 0.546 + ], + "angle": 0, + "content": "narios (i.e., CRF18, CRF28 and CRF38), achieving up to 1.06 dB PSNR gain on CRF18. In contrast, other methods trained with mixed compression settings show sub-optimal performance. For instance, STDF and S2SVR only achieve PSNR gains of 0.16 dB and 0.76 dB at CRF18, while MFQE 2.0 shows no improvement. Quantitative results on highly compressed scenarios (i.e., CRF40, CRF45, CRF48) are included in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.909, + 0.669 + ], + "angle": 0, + "content": "Qualitative results. Qualitative comparisons are provided in Fig. 6. As can be seen, MFQE 2.0 and Metabit struggle to improve the quality of compressed inputs. Both STDF and S2SVR cannot adequately remove compression artifacts (e.g., boundary of the calf), while the proposed method effectively eliminates the compression artifacts, preserving accurate edges and textures. We provide more qualitative comparisons in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.672, + 0.714, + 0.687 + ], + "angle": 0, + "content": "5.2.2. Versatility Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.765 + ], + "angle": 0, + "content": "To evaluate the versatility in assisting practical downstream tasks, we employ the implementation that utilizes slice type for frame adaptation (as described in Sec. 3.1) to enhance compressed inputs for downstream tasks. More qualitative comparisons are included in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Video super-resolution. We adopt BasicVSR [6], IconVSR [6], and BasicVSR++ [7] as the representative baseline methods for \\(\\times 4\\) video super-resolution (VSR), which are trained on \"clean\" data without considering compression. The evaluation is conducted on the REDS4 dataset [39] and summarized with PSNR and SSIM (the higher the better). As depicted in Tab. 2, pre-enhancing compressed inputs with Metabit fails to improve the performance of downstream VSR models. In contrast, pre-enhancing with our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17772" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.082, + 0.332, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.082, + 0.415, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.082, + 0.496, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.082, + 0.577, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.082, + 0.659, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.082, + 0.74, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.082, + 0.823, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.082, + 0.905, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.229, + 0.905, + 0.258 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of \\(\\times 4\\) VSR. Pre-enhancing with our method before VSR effectively avoids amplifying compression artifacts. While other methods cannot fully eliminate the artifacts and even severe the distortions (e.g., results of MFQE 2.0)." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.262, + 0.209, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.262, + 0.326, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.262, + 0.441, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.262, + 0.557, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.262, + 0.672, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.262, + 0.788, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.262, + 0.905, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.327, + 0.905, + 0.356 + ], + "angle": 0, + "content": "Figure 8. Qualitative results of optical flow estimation. As can be seen, pre-processing with our method effectively corrects the mispredicted optical flow, especially the boundaries of moving objects (e.g., edge of the moving car)." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.366, + 0.449, + 0.603 + ], + "angle": 0, + "content": "
MethodCRF15CRF25CRF35
EPE↓/ F1-all↓EPE↓/ F1-all↓EPE↓/ F1-all↓
RAFT5.26 / 17.817.37 / 22.1316.73 / 44.70
+ MFQE 2.05.32 / 17.837.27 / 21.9216.68 / 44.74
+ STDF5.34 / 17.937.13 / 22.1615.92 / 44.04
+ S2SVR5.22 / 17.716.90 / 21.5715.73 / 44.62
+ Metabit5.32 / 17.867.33 / 22.0716.69 / 44.28
+ Ours5.20 / 17.696.52 / 20.9914.84 / 42.56
DEQ3.99 / 13.715.40 / 17.3313.94 / 41.63
+ MFQE 2.03.97 / 13.735.14 / 17.0614.06 / 41.72
+ STDF4.08 / 13.845.27 / 17.3813.52 / 40.75
+ S2SVR4.01 / 13.695.22 / 16.9913.74 / 41.65
+ Metabit3.92 / 13.785.30 / 17.3313.84 / 41.21
+ Ours3.97 / 13.684.96 / 16.5413.09 / 39.43
KPAFlow4.46 / 16.076.71 / 20.9616.50 / 45.13
+ MFQE 2.04.42 / 15.966.71 / 20.9216.70 / 45.29
+ STDF4.52 / 16.196.96 / 21.5315.62 / 44.17
+ S2SVR4.37 / 15.966.10 / 19.7615.62 / 44.23
+ Metabit4.47 / 16.176.75 / 21.0816.68 / 44.90
+ Ours4.43 / 16.105.59 / 18.7314.83 / 41.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.636 + ], + "angle": 0, + "content": "Table 3. Quantitative results of optical flow estimation, where we highlight the best and second best results with **bold** and **underline**." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.824 + ], + "angle": 0, + "content": "framework yields consistent improvement especially in scenarios of high compression (e.g., up to 0.62 dB PSNR gain with BaiscVSR++ on CRF35). Meanwhile, our method significantly outperforms MFQE 2.0 and STDF, achieving PSNR gains of 0.25 dB and 0.23 dB over MFQE 2.0 and STDF on BasicVSR/CRF15, respectively. Compared with S2SVR, the proposed method offers more effective support to VSR models with lower complexity. As shown in Fig. 7, performing VSR on compressed data inevitably amplifies compression artifacts (e.g., the 1st column), while results pre-enhanced with our method maintain accurate edges and textures, avoiding distortions seen in other methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Optical flow estimation. We adopt RAFT [49], DEQ [3], and KPAFlow [35] as baseline models for optical flow estimation. Evaluation on the KITTI-2015 dataset [21] is summarized with EPE (end-point-error) and F1-all loss, where lower values indicate better accuracy. As shown in Tab. 3," + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.363, + 0.905, + 0.602 + ], + "angle": 0, + "content": "
MethodCRF15CRF25CRF35
Avg↑/J↑/F↑Avg↑/J↑/F↑Avg↑/J↑/F↑
STCN85.07 / 81.83 / 88.3284.35 / 80.96 / 87.7479.20 / 76.04 / 82.37
+ MFQE 2.084.96 / 81.71 / 88.2184.30 / 80.92 / 87.6979.28 / 76.11 / 82.44
+ STDF85.01 / 81.73 / 88.2884.23 / 80.97 / 87.5079.77 / 76.52 / 83.02
+ S2SVR85.17 / 81.93 / 88.4184.46 / 81.20 / 87.7280.04 / 76.88 / 83.20
+ Metabit84.56 / 80.97 / 88.1483.86 / 80.18 / 87.5579.03 / 75.65 / 82.40
+ Ours85.21 / 81.99 / 88.4484.63 / 81.42 / 87.8581.57 / 78.46 / 84.69
DeAoT85.90 / 82.89 / 88.9185.18 / 82.37 / 88.0082.87 / 79.86 / 85.88
+ MFQE 2.085.86 / 82.84 / 88.8885.20 / 82.38 / 88.0382.86 / 79.86 / 85.85
+ STDF85.83 / 82.80 / 88.8785.18 / 82.27 / 88.0982.90 / 79.92 / 85.89
+ S2SVR86.05 / 83.09 / 89.0185.05 / 82.07 / 88.0482.64 / 79.63 / 85.65
+ Metabit85.47 / 82.04 / 88.9084.95 / 81.57 / 88.3382.32 / 79.04 / 85.59
+ Ours86.08 / 83.13 / 89.0385.31 / 82.38 / 88.2582.88 / 79.83 / 85.92
QDMN85.16 / 82.20 / 88.1184.16 / 81.20 / 87.1279.39 / 76.61 / 82.18
+ MFQE 2.085.13 / 82.20 / 88.0684.15 / 81.18 / 87.1379.51 / 76.75 / 82.27
+ STDF85.32 / 82.38 / 88.2783.36 / 80.44 / 86.2779.64 / 76.69 / 82.59
+ S2SVR85.28 / 82.32 / 88.2383.64 / 80.65 / 86.6379.02 / 76.15 / 81.89
+ Metabit84.50 / 81.14 / 87.8783.68 / 80.30 / 87.0679.47 / 76.41 / 82.52
+ Ours85.34 / 82.41 / 88.2784.37 / 81.42 / 87.3279.78 / 76.92 / 82.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.604, + 0.905, + 0.633 + ], + "angle": 0, + "content": "Table 4. Quantitative results of VOS, where the best and second best results are highlighted with bold and underline." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.643, + 0.907, + 0.884 + ], + "angle": 0, + "content": "our method consistently reduces the EPE and F1-all loss across all baseline models, demonstrating its effectiveness in improving optical flow estimation. In contrast, methods such as MFQE 2.0, STDF and Metabit fail to deliver consistent improvements. For instance, MFQE 2.0 fails to improve the performance of RAFT on CRF15, STDF and Metabit detrimentally affects the performance of DEQ and KPAFlow on CRF15. Visualizations of predicted optical flow are shown in Fig. 8, where inaccurate boundaries are highlighted with red arrows. As can be seen, optical flow estimated from compressed inputs contains inaccurate boundaries, especially near-motion ones. The proposed method helps to deliver more accurate results in these regions compared to others. For instance, it effectively corrects the optical flow near the car that was mispredicted by DEQ, while MFQE 2.0 and S2SVR provide limited improvement." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Video object segmentation. For video object segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17773" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.081, + 0.21, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.136, + 0.171, + 0.146 + ], + "angle": 0, + "content": "STCN" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.082, + 0.325, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.136, + 0.302, + 0.147 + ], + "angle": 0, + "content": "+ MFQE 2.0" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.082, + 0.442, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.136, + 0.407, + 0.147 + ], + "angle": 0, + "content": "+STDF" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.082, + 0.557, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.136, + 0.527, + 0.146 + ], + "angle": 0, + "content": "+ S2SVR" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.082, + 0.673, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.136, + 0.643, + 0.146 + ], + "angle": 0, + "content": "+Metabit" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.082, + 0.789, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.136, + 0.75, + 0.146 + ], + "angle": 0, + "content": "+Ours" + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.082, + 0.905, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.837, + 0.136, + 0.857, + 0.145 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.152, + 0.905, + 0.181 + ], + "angle": 0, + "content": "Figure 9. Qualitative results of VOS. As can be seen, directly performing VOS on compressed inputs leads to inaccurate masks, whereas pre-enhancing with our method effectively improves the accuracy, especially for the regions of irregular shapes (e.g., the windshield)." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.189, + 0.208, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.244, + 0.19, + 0.254 + ], + "angle": 0, + "content": "Masked frame" + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.189, + 0.324, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.244, + 0.29, + 0.254 + ], + "angle": 0, + "content": "\\(E^2\\) FGVI" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.189, + 0.441, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.244, + 0.417, + 0.255 + ], + "angle": 0, + "content": "+ MFQE 2.0" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.189, + 0.557, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.244, + 0.52, + 0.254 + ], + "angle": 0, + "content": "+STDF" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.189, + 0.673, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.244, + 0.641, + 0.254 + ], + "angle": 0, + "content": "+ S2SVR" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.189, + 0.788, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.244, + 0.758, + 0.254 + ], + "angle": 0, + "content": "+Metabit" + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.189, + 0.905, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.827, + 0.244, + 0.865, + 0.254 + ], + "angle": 0, + "content": "+Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.259, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Figure 10. Qualitative results of video inpainting. As can be seen, pre-enhancing the compressed inputs with the proposed method helps to reduce the artifacts and color distortion in the removed region, providing more visually pleasant results." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.304, + 0.485, + 0.606 + ], + "angle": 0, + "content": "(VOS), we adopt STCN [14], DeAoT [57] and QDMN [34] as representative baselines. Evaluations on DAVIS-17 val dataset [41] are summarized with the following metrics: the \\(\\mathcal{I}\\) (average IoU), the \\(\\mathcal{F}\\) score (boundary similarity), and the average of the above metrics (denoted as \\(Avg\\)). Higher values indicate better segmentation accuracy. As shown in Tab. 4, the proposed method shows the best performance in improving accuracy across VOS models. For instance, it elevates the average accuracy for up to \\(2.37\\%\\) (79.20% to \\(81.57\\%\\)) on STCN at CRF35, while MFQE 2.0, STDF and S2SVR yield limited improvement of \\(0.08\\%\\), \\(0.57\\%\\) and \\(0.84\\%\\), respectively. And Metabit provides no improvement on STCN at CRF35. The results of VOS are included in Fig. 9, where accurately segmenting objects in compressed videos is challenging for VOS baselines (e.g., inaccurate mask of the windshield predicted by STCN). Pre-enhancing the compressed videos with MFQE 2.0 and S2SVR struggles to address this issue, whereas the proposed method significantly refines the segmentation results, demonstrating its effectiveness in assisting the VOS task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.744 + ], + "angle": 0, + "content": "Video inpainting. We take \\(\\mathrm{E}^2\\mathrm{FGVI}\\) [31] as the video inpainting model and perform video object removal on DAVIS-17 val dataset [41]. The qualitative results are shown in Fig. 10. As can be seen, compression-included misalignment between objects and masks hinders the ability to remove specified objects, causing color distortions (e.g., the horse region). Pre-enhancing the compressed inputs with the proposed method notably refines the artifacts and distortions, yielding more visually pleasing results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.756, + 0.255, + 0.77 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.854 + ], + "angle": 0, + "content": "We start with a baseline that concatenates reference frames and current frames as input, without using codec information. We then progressively equip the baseline with MV alignment, region-aware refinement, sequence adaptation, and frame adaptation to assess their contributions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "MV alignment. As shown in the 2nd row of Tab. 5, incorporating MV alignment yields a PSNR gain of up to 0.65 dB on CRF15, demonstrating the effectiveness of MV in" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.302, + 0.903, + 0.409 + ], + "angle": 0, + "content": "
ModelCRF15CRF25CRF35
PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
Baseline41.04 / 0.978534.92 / 0.936329.25 / 0.8238
+ MV Align.41.69 / 0.982135.59 / 0.943729.95 / 0.8403
+ RA Refine.42.04 / 0.983735.70 / 0.944930.00 / 0.8427
+ Seq. Adapt.42.08 / 0.983835.76 / 0.945830.04 / 0.8444
+ Frame Adapt.42.14 / 0.983935.81 / 0.946030.09 / 0.8446
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.413, + 0.905, + 0.442 + ], + "angle": 0, + "content": "Table 5. Ablation studies on MV alignment, region-aware refinement, sequence adaptation, and frame adaptation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.457, + 0.809, + 0.471 + ], + "angle": 0, + "content": "aligning reference frames and current frame." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.547 + ], + "angle": 0, + "content": "Region-aware refinement. We further incorporate the region-aware refinement module to refine the features of different regions. As shown in the 3rd row of Tab. 5, it leads to notable PSNR gains of \\(0.35\\mathrm{dB}\\), \\(0.11\\mathrm{dB}\\) and \\(0.05\\mathrm{dB}\\) on CRF15, CRF25 and CRF35, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.905, + 0.593 + ], + "angle": 0, + "content": "Sequence adaptation. As shown in the 4th row of Tab. 5, sequence adaptation brings PSNR gains of \\(0.04\\mathrm{dB}\\), \\(0.06\\mathrm{dB}\\) and \\(0.04\\mathrm{dB}\\) on CRF15, CRF25 and CRF35, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.67 + ], + "angle": 0, + "content": "Frame adaptation. As shown in the 5th row of Tab. 5, frame adaptation improves PSNR by \\(0.06\\mathrm{dB}\\) \\(0.05\\mathrm{dB}\\) and \\(0.05\\mathrm{dB}\\) on CRF15, CRF25 and CRF35, respectively. We further analyze its effectiveness on improving the temporal consistency in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.684, + 0.634, + 0.699 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.856 + ], + "angle": 0, + "content": "In this paper, we introduce a versatile codec-aware enhancement framework that adaptively handles diverse compression settings and serves as a plug-and-play enhancement module to consistently boost various downstream tasks. By reusing the off-the-shelf codec information, our method minimizes additional computational costs. Compared with existing compressed video enhancement solutions, it shows superiority in both enhancement performance and robustness, making it possible to deploy pre-trained models on compressed videos without a significant performance drop." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments. We acknowledge funding from the National Natural Science Foundation of China under Grants 62131003 and 62021001." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17774" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Video developer report 2022-2023. https://bitmovin.com/downloads/ assets/bitmovin-7th-video-developer-report-2023-2024.pdf. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In CVPR, pages 3366-3375, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.482, + 0.242 + ], + "angle": 0, + "content": "[3] Shaojie Bai, Zhengyang Geng, Yash Savani, and J. Zico Kolter. Deep equilibrium optical flow estimation. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[4] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[5] Yuxuan Cai, Hongjia Li, Geng Yuan, Wei Niu, Yanyu Li, Xulong Tang, Bin Ren, and Yanzhi Wang. Yolobile: Realtime object detection on mobile devices via compression-compiler co-design. In AAAI, pages 955–963, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[6] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, pages 4947-4956, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[7] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, pages 5972-5981, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[8] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In CVPR, pages 5962-5971, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[9] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In ICIP, pages 168-172, 1994. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.602, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[10] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In CVPR, pages 8064-8073, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[11] Peilin Chen, Wenhan Yang, Long Sun, and Shiqi Wang. When bitstream prior meets deep prior: Compressed video super-resolution with learning from decoding. In ACM MM, pages 1000–1008, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[12] Peilin Chen, Wenhan Yang, Meng Wang, Long Sun, Kangkang Hu, and Shiqi Wang. Compressed domain deep video super-resolution. IEEE Transactions on Image Processing, 30:7156-7169, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[13] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In CVPR, pages 11030-11039, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[14] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Yuanying Dai, Dong Liu, and Feng Wu. A convolutional neural network approach for post-processing in hevc intra coding. In MultiMedia Modeling: 23rd International Con" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "ference, MMM 2017, Reykjavik, Iceland, January 4-6, 2017, Proceedings, Part I 23, pages 28-39. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[16] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. In AAAI, pages 10696-10703, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[17] Cunhui Dong, Haichuan Ma, Zhuoyuan Li, Li Li, and Dong Liu. Temporal wavelet transform-based low-complexity perceptual quality enhancement of compressed video. IEEE TCSVT, pages 1-1, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[18] Max Ehrlich, Jon Barker, Namitha Padmanabhan, Larry Davis, Andrew Tao, Bryan Catanzaro, and Abhinav Shrivastava. Leveraging bitstream metadata for fast, accurate, generalized compressed video quality enhancement. In WACV, pages 1517-1527, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[19] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[20] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaeser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[21] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[22] Zhenyu Guan, Qunliang Xing, Mai Xu, Ren Yang, Tie Liu, and Zulin Wang. Mfqe 2.0: A new approach for multi-frame quality enhancement on compressed video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3): 949-963, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[23] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7436-7456, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[24] Hongyue Huang, Ionut Schiopu, and Adrian Munteanu. Frame-wise cnn-based filtering for intra-frame quality enhancement of hevc videos. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2100-2113, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[25] Zhijie Huang, Jun Sun, and Xiaopeng Guo. Fastconn: Towards fast and accurate spatiotemporal network for hevc compressed video enhancement. ACM Transactions on Multimedia Computing, Communications and Applications, 19 (3):1-22, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[26] Chuanmin Jia, Shiqi Wang, Xinfeng Zhang, Shanshe Wang, Jiaying Liu, Shiliang Pu, and Siwei Ma. Content-aware convolutional neural network for in-loop filtering in high efficiency video coding. IEEE Transactions on Image Processing, 28(7):3343-3356, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Nanfeng Jiang, Weiling Chen, Jielian Lin, Tiesong Zhao, and Chia-Wen Lin. Video compression artifacts removal with spatial-temporal attention-guided enhancement. IEEE Transactions on Multimedia, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17775" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[28] Feng Li, Yixuan Wu, Anqi Li, Huihui Bai, Runmin Cong, and Yao Zhao. Enhanced video super-resolution network towards compressed data. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-21, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[29] Yinxiao Li, Pengchong Jin, Feng Yang, Ce Liu, Ming-Hsuan Yang, and Peyman Milanfar. Comisr: Compression-informed video super-resolution. In ICCV, pages 2543-2552, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[30] Yunsheng Li, Lu Yuan, Yinpeng Chen, Pei Wang, and Nuno Vasconcelos. Dynamic transfer for multi-source domain adaptation. In CVPR, pages 10998-11007, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[31] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[32] Jing Lin, Xiaowan Hu, Yuanhao Cai, Haoqian Wang, Youliang Yan, Xueyi Zou, Yulun Zhang, and Luc Van Gool. Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In ICML, pages 13394-13404. PMLR, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[33] Baoyuan Liu, Min Wang, Hassan Foroosh, Marshall Tappen, and Marianna Pensky. Sparse convolutional neural networks. In CVPR, pages 806-814, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[34] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, pages 468-486, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[35] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In CVPR, pages 8906-8915, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[36] Di Ma, Fan Zhang, and David R Bull. Cvegan: a perceptually-inspired gan for compressed video enhancement. Signal Processing: Image Communication, page 117127, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[37] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[38] Sreyas Mohan, Joshua L Vincent, Ramon Manzorro, Peter Crozier, Carlos Fernandez-Granda, and Eero Simoncelli. Adaptive denoising via gaintuning. NeurIPS, 34:23727-23740, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[39] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, pages 0–0, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[40] Zhaoqing Pan, Xiaokai Yi, Yun Zhang, Byeungwoo Jeon, and Sam Kwong. Efficient in-loop filtering based on enhanced deep convolutional neural networks for hevc. IEEE Transactions on Image Processing, 29:5352-5366, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[41] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[42] Darren Ramsook and Anil Kokaram. Learnt deep hyperparameter selection in adversarial training for compressed video enhancement with a perceptual critic. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2420-2424. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.163, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[43] Mennatullah Siam, Mostafa Gamal, Moemen Abdel-Razek, Senthil Yogamani, Martin Jagersand, and Hong Zhang. A comparative study of real-time semantic segmentation for autonomous driving. In CVPRW, pages 587-597, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.219, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[44] Mennatullah Siam, Alex Kendall, and Martin Jagersand. Video class agnostic segmentation benchmark for autonomous driving. In CVPR, pages 2825-2834, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.908, + 0.315 + ], + "angle": 0, + "content": "[45] Gary J Sullivan, Pankaj N Topiwala, and Ajay Luthra. The h. 264/avc advanced video coding standard: Overview and introduction to the fidelity range extensions. Applications of Digital Image Processing XXVII, 5558:454-474, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.907, + 0.371 + ], + "angle": 0, + "content": "[46] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.426 + ], + "angle": 0, + "content": "[47] Weiheng Sun, Xiaohai He, Chao Ren, Shuhua Xiong, and Honggang Chen. A quality enhancement network with coding priors for constant bit rate video coding. Knowledge-Based Systems, 258:110010, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.481 + ], + "angle": 0, + "content": "[48] Xuebin Sun, Sukai Wang, Miaohui Wang, Shing Shin Cheng, and Ming Liu. An advanced lidar point cloud sequence coding scheme for autonomous driving. In MM, pages 2793-2801, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.522 + ], + "angle": 0, + "content": "[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.906, + 0.579 + ], + "angle": 0, + "content": "[50] Jianyi Wang, Mai Xu, Xin Deng, Liquan Shen, and Yuhang Song. Mw-gan+ for perceptual quality enhancement on compressed video. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4224-4237, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.906, + 0.62 + ], + "angle": 0, + "content": "[51] Shiyao Wang, Hongchao Lu, and Zhidong Deng. Fast object detection in compressed video. In ICCV, pages 7104-7113, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[52] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Hutchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In CVPR, pages 2012-2021, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[53] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[54] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. NeurIPS, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[55] Ren Yang, Mai Xu, Zulin Wang, and Tianyi Li. Multi-frame quality enhancement for compressed video. In CVPR, pages 6664-6673, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[56] Ren Yang, Mai Xu, Tie Liu, Zulin Wang, and Zhenyu Guan. Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology, 29(7): 2039-2054, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[57] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation. In NeurIPS, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17776" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[58] Li Yu, Wenshuai Chang, Shiyu Wu, and Moncef Gabbouj. End-to-end transformer for compressed video quality enhancement. IEEE Transactions on Broadcasting, 70(1):197-207, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[59] Haichao Zhang, Jianchao Yang, Yanning Zhang, Nasser M Nasrabadi, and Thomas S Huang. Close the loop: Joint blind image restoration and recognition with sparse representation prior. In ICCV, pages 770-777, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[60] Hengsheng Zhang, Xueyi Zou, Jiaming Guo, Youliang Yan, Rong Xie, and Li Song. A codec information assisted framework for efficient compressed video super-resolution. In ECCV, pages 220-235. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[61] Saiping Zhang, Luis Herranz, Marta Mrak, Marc Górriz Blanch, Shuai Wan, and Fuzheng Yang. Dcngan: A deformable convolution-based gan with qp adaptation for perceptual quality enhancement of compressed video. In ICASSP, pages 2035-2039. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[62] Zhengdong Zhang and Vivienne Sze. Fast: A framework to accelerate super-resolution processing on compressed videos. In CVPRW, pages 19-28, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[63] Hengrun Zhao, Bolun Zheng, Shanxin Yuan, Hua Zhang, Chenggang Yan, Liang Li, and Gregory Slabaugh. Cbren: Convolutional neural networks for constant bit rate video quality enhancement. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4138-4149, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[64] Qiang Zhu, Jinhua Hao, Yukang Ding, Yu Liu, Qiao Mo, Ming Sun, Chao Zhou, and Shuyuan Zhu. Cpga: Coding priors-guided aggregation network for compressed video quality enhancement. In CVPR, pages 2964-2974, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17777" + } + ] +] \ No newline at end of file diff --git a/2025/Plug-and-Play Versatile Compressed Video Enhancement/e1321f09-cc21-461b-8861-4f42cf1e3c7e_origin.pdf b/2025/Plug-and-Play Versatile Compressed Video Enhancement/e1321f09-cc21-461b-8861-4f42cf1e3c7e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bb8726c551c1e83497db8ba81e4813713880c75c --- /dev/null +++ b/2025/Plug-and-Play Versatile Compressed Video Enhancement/e1321f09-cc21-461b-8861-4f42cf1e3c7e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5aadff514d43225050f2e78441a44fbda30d4c9d98b9c265ba397354e031423f +size 7966316 diff --git a/2025/Plug-and-Play Versatile Compressed Video Enhancement/full.md b/2025/Plug-and-Play Versatile Compressed Video Enhancement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..37678e379cbd1b02d05ae27b52a0eb00f766f5c2 --- /dev/null +++ b/2025/Plug-and-Play Versatile Compressed Video Enhancement/full.md @@ -0,0 +1,441 @@ +# Plug-and-Play Versatile Compressed Video Enhancement + +Huimin Zeng Jiacheng Li Zhiwei Xiong* + +University of Science and Technology of China + +{zenghuimin, jciee}@mail.ustc.edu.cn zwxiong@ustc.edu.cn + +# Abstract + +As a widely adopted technique in data transmission, video compression effectively reduces the size of files, making it possible for real-time cloud computing. However, it comes at the cost of visual quality, posing challenges to the robustness of downstream vision models. In this work, we present a versatile codec-aware enhancement framework that reuses codec information to adaptively enhance videos under different compression settings, assisting various downstream vision tasks without introducing computation bottleneck. Specifically, the proposed codec-aware framework consists of a compression-aware adaptation (CAA) network that employs a hierarchical adaptation mechanism to estimate parameters of the frame-wise enhancement network, namely the bitstream-aware enhancement (BAE) network. The BAE network further leverages temporal and spatial priors embedded in the bitstream to effectively improve the quality of compressed input frames. Extensive experimental results demonstrate the superior quality enhancement performance of our framework over existing enhancement methods, as well as its versatility in assisting multiple downstream tasks on compressed videos as a plug-and-play module. Code and models are available at https://huimin-zeng.github.io/PnP-VCVE/. + +# 1. Introduction + +With the flower booming of short video platforms, video has become one of the most popular multimedia formats. In addition to distributing visual content, in practical scenarios (e.g., autonomous driving [43, 48]), it is common to upload the captured videos to the cloud end for further visual analysis and downstream applications (e.g., object detection [5, 51] and segmentation [20, 44]). However, due to the bandwidth constraint during transmission, these videos are typically compressed with varying levels, resulting in poor visual quality and suboptimal performance in downstream tasks [22, 59] (e.g., inaccurate segmentation bound + +![](images/9b5952d352646c15004795393533a8a5ef99b2c6f588ac52605ac07b59022c64.jpg) +Figure 1. The proposed codec-aware enhancement framework reuses codec information to adaptively enhance videos across different compression settings, while assisting in various downstream tasks in a plug-and-play manner. + +aries in Fig. 1). Given the crucial role of videos in data transmission, there is a critical need for a versatile solution to enhance videos of diverse compression levels and effectively support various downstream tasks. + +Existing video enhancement methods are hard to respond to these demands. Specifically, to effectively enhance videos of different compression levels, previous methods [11, 12, 16, 22, 32, 55] employ separate enhancement models for each compression level, which is inflexible occurring unseen compression levels. Recent approaches [8, 18, 47] consider this issue as the generalization ability across diverse compression levels, therefore randomly selecting inputs of different compression levels during training. However, such a training strategy is compression-agnostic and offers limited improvement. Most importantly, the aforementioned methods focus primarily on improving perceptual quality, neglecting the need to assist in downstream tasks in real-world scenarios. + +Based on the mismatch between the versatility demand and existing solutions, here we summarize the following criteria of a favorable solution: 1) adaptively enhance videos of varying compression levels with a single model; 2) effectively assist various downstream tasks on compressed videos in a plug-and-play manner; and 3) given the + +practical scenarios where real-time processing is required, it should meet the above objectives without causing a computation bottleneck. To achieve this, we introduce a codec-aware enhancement framework (as shown in Fig. 1) that reuses codec information embedded in the bitstream. By incorporating compression factors, the framework dynamically adjusts its parameters to flexibly enhance inputs of different compression levels. By reusing motion vectors and partition maps, it efficiently aggregates temporal and spatial clues without introducing redundant computations. + +Specifically, our codec-aware enhancement framework comprises a compression-aware adaptation (CAA) network and a bitstream-aware enhancement (BAE) network. The CAA network serves as a "meta" network that dynamically adjusts the parameters of the subsequent BAE network. A hierarchical adaptation mechanism is proposed to first estimate sequence-adaptive parameters based on the sequence compression level, and then re-weight these parameters according to the frame compression level, thereby achieving a BAE network tailored for each input frame. The frame-adaptive BAE network conducts motion vector alignment to aggregate intra-frame information and provide useful clues for the current frame. Subsequently, based on the region complexity indicated by partition maps, the region-aware refinement assigns independent filters for different regions, achieving flexible enhancement for different regions. Comprehensive experiments are conducted to demonstrate the superiority of our method in improving the quality of compressed videos, and the effectiveness of assisting in various downstream tasks (i.e., video super-resolution, optical flow estimation, and video object segmentation). Our contributions are summarized as follows: + +- We present a codec-aware framework for versatile compressed video enhancement, which adaptively enhances input videos of different compression levels and supports a wide range of downstream vision tasks. +- We develop a compression-aware adaptation (CAA) network and a bitstream-aware enhancement (BAE) network that utilize the off-the-shelf codec information, contributing to generalizing across different compression settings and boosting the enhancement performance with a unified framework. +- Experimental results show the superiority of our method over existing enhancement methods, and its effectiveness in serving as a plug-and-play enhancement module to assist in downstream tasks. + +# 2. Related Work + +# 2.1. Compressed Video Enhancement + +Existing compressed video enhancement methods can be categorized into in-loop and post-processing methods. Although in-loop methods [15, 24, 26, 40] effectively improve the quality of reconstructed frames, they embed filters + +in the encoding and decoding loops, therefore not suitable for enhancing already compressed videos. While the post-processing methods [17, 25, 27, 36, 42, 47, 50, 56, 58, 61, 63, 64] provide more practical solutions to enhance compressed videos by placing filters at the decoder side. Observing the quality fluctuation across frames, MFQE [55] locates the peak quality frame (PQF) with an SVM-based detector and proposes a multi-frame quality enhancement mechanism to enhance non-PQFs. MFQE 2.0 [22] further designs a BiLSTM-based detector and performs multiframe quality enhancement for both non-PQF and PQF. To address the inaccuracies in optical flow estimation from compressed videos, STDF [16] proposes estimating the offset field using spatio-temporal deformable convolution. S2SVR [32] introduces a sequence-to-sequence network to model long-range dependencies within frames. The aforementioned methods inflexibly equip a separate model for each compression level, while we propose to adaptively handle diverse inputs with a single unified model. Recent methods [18, 64] utilize spatial priors from bitstream to address multiple compression levels, however, they only consider I/P-frames, whereas we design a hierarchical adaptation mechanism to address all types of frames. + +# 2.2. Codec-Aware Video Super-Resolution + +Some works in video super-resolution (VSR) [11, 12, 28, 29, 52, 60, 62] explore ways of using codec information such as motion estimation and spatial prior for reconstruction. COMISR [29] focuses on reducing accumulated warping errors caused by the random locations of the intra-frame from compressed video frames. Chen et al. [11] employ motion vectors to build the temporal relationship and suppress coding artifacts. CVCP [12] utilizes motion vectors and spatial priors with a guided soft alignment scheme and guided SFT layer, respectively. CIAF [60] leverages motion vectors and residuals to model temporal relationships and skip redundant computations, respectively. Despite leveraging codec information, these methods focus on a single task (i.e., VSR). In contrast, our method not only shows competitive performance in VSR (see supplementary materials), but also effectively supports a range of downstream tasks, which is not explored by the aforementioned works. + +# 2.3. Dynamic Neural Networks + +Instead of setting separate models for different inputs, the early mixture of expert (MoE) structure [2, 19, 37] constructs parallel network branches and selectively executes branches to obtain the weighted outputs. Instead of increasing the number of parallel branches, the dynamic parameters ensemble strategy [2, 13, 19, 23, 37, 54] presets parallel expert layers and selectively fusing their parameters to promote the network capability and generalization ability, therefore serving as an efficient alternative to MoE. To promote the generalization ability of pre-trained models, Gain- + +![](images/252eafdcfa9fa26beb5db55d2291032b16fbdbcb0b367a33ac02f6e8a863c7c5.jpg) +Figure 2. Hierarchical structure of quality adjustment, where frames are divided into multiple groups of pictures (GOP). The Constant Rate Factor (CRF) affects video quality at both sequence and frame levels. An increase in the CRF value indicates a reduction in video quality (e.g., lower PSNR values). + +tune [38] proposes to predict a single multiplicative scaling parameter for each channel according to test samples, thus modifying static models to test-adaptive ones. Li et al. [30] handle the conflicts between the domain-agnostic model and multiple target domains with dynamic transfer, which is simply modeled by combining residual matrices and a static convolution matrix. DRConv [10] divides the input image into different regions with a learnable mask and assigns multiple filters for these regions, which enhances the feature representation ability without introducing a noticeable computation burden. Instead of searching and learning conditions for dynamic parameters, we leverage priors embedded in the bitstream as the condition. + +# 3. Preliminaries + +We take H.264 [53] as a representative standard to analyze available codec information. Note recent codecs [4, 46] also provide similar priors (i.e., CRF, motion vector, partition map), therefore assuring the applicability of our method. To reduce transmission bandwidth, codecs compress videos by adjusting quality and reducing redundancy. + +# 3.1. Hierarchical Quality Adjustment + +Video quality is commonly influenced by the constant rate factor (CRF), which involves hierarchical adjustment for both sequence-wise and frame-wise compression. + +Sequence-wise CRF. The CRF ranges from 0 to 51 to balance compression efficiency and visual quality. A higher CRF results in more compact output but increased pixel loss (e.g., the average PSNR of CRF35 is much lower than CRF15 in Fig. 2). By considering the sequence-wise CRF (denoted as $CRF_{s}$ ), the enhancement network can be tailored to handle videos of different compression levels. + +Frame-wise CRF As shown in Fig. 2, a video sequence is divided into multiple groups of pictures (GOP) and further categorized as I-frames, P-frames, or B-frames. The CRF value of each frame (denoted as $CRF_{i}$ ) is dynamically + +adjusted based on $CRF_{s}$ so that lower $CRF_{i}$ is assigned for I/P frames to maintain quality and higher $CRF_{i}$ for B frames for compact representations. + +Inspired by the hierarchical quality adjustment paradigm, we design a hierarchical adaptation paradigm that first performs sequence adaptation to predict network parameters based on $CRF_{s}$ , and then re-weights these parameters according to $CRF_{i}$ for frame adaptation. In practical scenarios where $CRF_{i}$ is unavailable (e.g., limited access to full bitstream), the proposed method can instead use slice type (I/P/B) for frame adaptation, which is demonstrated to yield similar performance in Sec. 5.2. + +# 3.2. Redundancy Reduction + +To reduce redundancy and improve the entropy of the bitstream,Codec block-wisely perform motion estimation to model the intra-frame correlations and embed the correlations in the bitstream for decoding. + +Partition map. As shown in Fig. 1, different regions of each frame are partitioned into blocks of varying sizes (e.g., H.264 provides macroblocks of $16 \times 16$ , $16 \times 8$ , $8 \times 16$ , and $8 \times 8$ ) according to the texture complexity. Flat regions (e.g., the ground) can tolerate higher quantization errors and are therefore divided into blocks of large size, while complex regions (e.g., leaves and fence) take smaller blocks to maintain details. To effectively enhance regions of different complexity, we propose dynamically assigning filters based on the partition map that indicates region complexity. + +Motion vector. Motion vectors are utilized in decoding to aggregate information from reference frames and propagate information of current frame. As illustrated in Fig. 1, they describe the relationship between current frame and its reference frames in a block-wise manner. Although motion vectors can be noisy and are less precise than optical flow, they effectively align reference frames with current frame, serving as a cost-effective alternative for optical flow. + +# 4. Codec-Aware Enhancement Framework + +# 4.1. Overview + +As shown in Fig. 3(a), the proposed method comprises a compression-aware adaption (CAA) network $\mathcal{G}_{\phi}$ and a bitstream-aware enhancement (BAE) network $\mathcal{F}_{\theta_i}$ . The CAA network employs a hierarchical compression adaptation mechanism to estimate parameters for the frame-adaptive BAE network, which then aggregates intra-frame information and performs region-aware refinement to enhance the input compressed frame. + +# 4.2. Compression-Aware Adaptation Network + +To handle sequences of varying compression levels and quality fluctuations across frames, as illustrated in Fig. 3(a), the CAA network $\mathcal{G}_{\phi}$ utilizes $CRF_{s}$ for sequence adaptation to estimate sequence-adaptive parameters, and perform frame adaptation to refine these parameters based on $CRF_{i}$ . + +![](images/f2e6968fdb4faa56129914f2aa33472b03e6f6b132ac64863568be3c7d261bcf.jpg) + +![](images/987705a8866119a8a9efeaab1fe4240c181b9b24b52fe414c75ca859ab1bbf7c.jpg) +Figure 3. The proposed Codec-Aware Enhancement Framework consists of two sub-networks: 1) the Compression-Aware Adaptation (CAA) Network, which hierarchically applies sequence adaptation and frame adaptation to dynamically adjust parameters of the enhancement network; and 2) Bitstream-Aware Enhancement (BAE) Network, which leverages motion vectors to align frames and conducts region-aware refinement to flexibly enhance regions of different complexity. + +![](images/00b816b2a20be9a9474f78fed57b3062d9b63d43cb0adfd2ac883c8d7219e61d.jpg) + +![](images/4c54ae2e59ec269ad78571d3d467ee6468c19468448edf8f249ed0910aa9c41b.jpg) + +![](images/516ba0b2603165777b42e9571593bd8a6b81f776ce183826e09d15478a26d939.jpg) + +Sequence adaptation. To ensure robust performance across multiple compression settings without increasing complexity, we propose estimating sequence-adaptive parameters for the enhancement network instead of fusing features from separate submodels. As shown in Fig. 3(b), parallel expert layers $\{f_{\theta_1}, f_{\theta_2}, \dots, f_{\theta_N}\}$ , which share the same architecture but have independent parameters, serve as the basis for parameter combination. The sequence-wise $CRF_s$ is adopted as the condition to re-weight parameters of these expert layers, which is expressed as follows, + +$$ +f _ {\theta_ {s}} = \mathcal {G} _ {\phi_ {s}} \left(C R F _ {s}, \left\{f _ {\theta_ {1}}, f _ {\theta_ {2}}, \dots , f _ {\theta_ {N}} \right\}\right) = \sum_ {n = 1} ^ {N} w _ {n} f _ {\theta_ {n}}, \tag {1} +$$ + +where $f_{\theta_s}$ and $\mathcal{G}_{\phi_s}$ denote the sequence-adaptive expert layer and the sequence-wise weight generator, respectively. $w_n$ denotes the weight for each expert layer. We set $N = 6$ (see ablation studies in supplementary materials) and visualize $w_n$ against different $CRF_s$ in Fig. 4, which shows that each expert layer has a distinct preference for specific $CRF_s$ . Compared to MoE that re-weights output features, re-weighting expert layer parameters (as shown in Eq. 1) is computationally efficient and comparable to the network constructed with a single expert layer. Note that $CRF_s$ is constant for frames within the same sequence, $f_{\theta_s}$ is predicted only once and reused by subsequent frames. + +Frame adaptation. To flexibly enhance frames with different visual quality, we propose to re-weight the sequence-based $f_{\theta_s}$ using frame-wise $CRF_{i}$ . We attribute the quality fluctuation between the sequence and current frame to the disparity between $CRF_{s}$ and $CRF_{i}$ , which can be ad + +![](images/f66511060d4f358a95dc2fd332451b6b6dfafb423f2343ed4dd0663d8f934480.jpg) +Figure 4. Visualization of $w_{n}$ against different $CRF_{s}$ , where each expert shows a distinct preference for specific $CRF_{s}$ values. + +dressed by introducing a set of frame-wise auxiliary parameters $\triangle \theta_{i}$ . As shown in Fig. 3(c), the auxiliary parameters $\triangle \theta_{i}$ that conditioned on $CRF_{i}$ re-weights the sequence-adaptive $f_{\theta_s}$ to obtain the frame-adaptive expert layer $f_{\theta_i}$ which is expressed as follows, + +$$ +f _ {\theta_ {i}} = \mathcal {G} _ {\phi_ {i}} \left(C R F _ {i}, f _ {\theta_ {s}}\right) = f _ {\theta_ {s} + \triangle \theta_ {i}}, \tag {2} +$$ + +where $f_{\theta_i}$ and $\mathcal{G}_{\phi_i}$ denote the estimated frame-adaptive expert layer and the frame-wise parameters generator, respectively. As shown by the black dashed lines in Fig. 3(a), the obtained $f_{\theta_i}$ is used to construct the enhancement blocks, resulting in the frame-adaptive BAE network $\mathcal{F}_{\theta_i}$ (introduced in the following Sec. 4.3). + +# 4.3. Bitstream-Aware Enhancement Network + +To leverage high-quality frames and propagate information, the BAE network $\mathcal{F}_{\theta_i}$ utilizes motion vectors to align reference frames with the current frame. Meanwhile, the par + +![](images/cbe3707dba8b057c2a02dd066680850611f631e480f4b6fb40311aa443d64243.jpg) +Reference frame + +![](images/85bc66cfbd07932a6db7a2b93d1b18030ec275ca44ae7b88ab56f39599426598.jpg) +$h_i$ + +![](images/bfe3f50b910ec0825c4f674b67ba625663a0fbadd4f694562831e3dae7da534f.jpg) +$\hat{h}_i$ + +![](images/fd9725b6d195a16ca67476ec4e1ee3128230dcf1e662db6450894c9486b3394b.jpg) +$\mathcal{S}(M_i^{8\times 8},h_i)$ +Figure 5. Visualization of features in region-aware refinement, where $h_i$ and $\hat{h}_i$ indicate the input and output features, respectively. The refined features are denoted in the format of $S(M_i^{type}, h_i)$ . + +![](images/26d2d57be50facdd724be3d6ea4582d2844fd15d73191db46b14de09f72fff73.jpg) +$\mathcal{S}(M_i^{16\times 16},h_i)$ + +![](images/d10ecead72505980800ca33277dc660618d1818ccfd7a3d2f9d3ddb02a8ae7a3.jpg) +$\mathcal{S}(M_i^{16\times 8},h_i)$ + +tition map serves as spatial complexity guidance to enable flexible enhancement of different regions. + +Motion vector alignment. Since the motion vectors roughly model the temporal relationship in a block-wise manner, for each block of the current frame shown in Fig. 3(d), motion vectors locate blocks with similar content in the previous and future reference features (highlighted with red and blue boxes). The warped reference features are concatenated with current frame along the channel dimension as input of the BAE network, expressed as follows, + +$$ +\hat {x} _ {i} = \left[ M V \left(h _ {i} ^ {p}\right), M V \left(h _ {i} ^ {f}\right), x _ {i} \right], \tag {3} +$$ + +where $\hat{x}_i$ , $h_i^p$ and $h_i^f$ denote the current input frame, enhanced features of previous and future reference frames, respectively. $MV$ denotes warping reference features based on motion vectors. $[,]$ denotes concatenation along channel dimension. Bilinear interpolation is adopted for the case that the offsets of motion vectors are not integers. + +Region-aware refinement. To effectively enhance regions of different complexity, we propose to dynamically assign different filters for regions based on the partition map. As shown in Fig. 3(e), the block-based partition map $P_{i}$ is decoupled into multiple binary masks $\{M_i^{8\times 8}, M_i^{16\times 16}, \dots, M_i^{type}\}$ according to the size of macroblocks, allowing separate refinement of regions using sparse convolution [33]. The output is defined as the sum of frame-adaptive extracted features and separately region-aware refined features, depicted as follows, + +$$ +\hat {h} _ {i} = \mathcal {F} _ {\theta_ {i}} \left(\hat {x} _ {i}, P _ {i}\right) = f _ {\theta_ {i}} * h _ {i} + \sum_ {t y p e = 1} ^ {M} \mathcal {S} \left(M _ {i} ^ {t y p e}, h _ {i}\right), \tag {4} +$$ + +where $\hat{h}_i$ indicates the output features. $S$ denotes the operations applying sparse convolution guided by mask $M_i^{type}$ to refine input features. In H.264 standard, three types of macroblocks are used $(16\times 16,8\times 16 / 16\times 8,$ and $8\times 8)$ , thus $M$ is set to 3. Features in region-aware refinement are visualized in Fig 5, where the refined features are denoted with operations like $S(M_{i}^{8\times 8},h_{i})$ . As can be seen, the output features $\hat{h}_i$ contain more fine-grained and high-frequency details than $h_i$ . Meanwhile, refined features provide distinct activations for different regions. For instance, $S(M_i^{8\times 8},h_i)$ + +focuses on static objects (e.g., trees) while $\mathcal{S}(M_i^{8\times 16},h_i)$ focuses on moving objects (e.g., the bus). + +# 4.4. Loss function + +We adopt Charbonnier penalty loss [9] as the loss function and train the proposed codec-aware enhancement framework in an end-to-end manner. The specific loss function is expressed as follows, + +$$ +\mathcal {L} = \frac {1}{T} \sum_ {i = 1} ^ {T} \sqrt {\left\| y _ {i} - \hat {y} _ {i} \right\| ^ {2} + \epsilon^ {2}}, \tag {5} +$$ + +where $y_{i}, \hat{y}_{i}$ and $T$ indicate the uncompressed ground truth, the predicted output, and the length of the input sequence. $\epsilon$ is set to $1 \times 10^{-12}$ . + +# 5. Experiments + +# 5.1. Experimental Settings + +Compression settings. H.264 is a popular video compression standard that compresses nearly $85\%$ of internet videos [1], and tends to introduce more severe degradations than H.265 and H.266. We adopt H.264 [45] and compress videos with the $CRF_{s}$ values of 15, 25, and 35. + +Tasks and training dataset. Our tasks involve quality enhancement and assisting downstream tasks on compressed inputs. The primary downstream tasks include video superresolution, optical flow estimation and video object segmentation, with video inpainting reported as an extension to fully evaluate the versatility. Training splits of REDS [39] and DAVIS [41] datasets are combined for training. + +Compared methods. We compare with representative methods in compressed video enhancement, including MFQE 2.0 [22], STDF [16], S2SVR [32] and Metabit [18]. For a fair comparison, we fully retrain these methods with the same training dataset and configurations. For downstream tasks, the compressed video is first enhanced by quality enhancement methods, and then fed to downstream models for corresponding tasks and further assessment. + +# 5.2. Results + +Our evaluations are two-fold: 1) verifying the quality enhancement performance on seen, unseen, and highly compressed scenarios; 2) evaluating the versatility to assist different downstream tasks on multiple compression settings. + +# 5.2.1. Quality Enhancement Performance + +Quantitative results. The results of compressed video quality enhancement are reported in Tab.1, which is evaluated on the REDS4 datasets [39] using PSNR and SSIM (the higher the better). Note the CRF values of 18, 28 and 38 are not included during training. For each method, we include the model complexity and inference speed. For our method, we report results of both applying $CRF_{i}$ and its substitution with slice type (highlighted with grey). As shown in Tab. 1, leveraging slice type yields comparable performance + +
MethodParam/MFLOPs/GSpeed/msFPSCRF15CRF25CRF35CRF18CRF28CRF38
PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑
Input----41.04 / 0.978534.92 / 0.936329.25 / 0.823839.12 / 0.969833.18 / 0.912327.69 / 0.7725
MFQE 2.0 [22]1.6451531940.95 / 0.980634.83 / 0.937829.22 / 0.825638.97 / 0.971233.13 / 0.914027.67 / 0.7742
STDF [16]1.2745382641.15 / 0.979335.23 / 0.939829.74 / 0.835939.28 / 0.971233.58 / 0.917828.11 / 0.7853
S2SVR [32]7.43294--41.96 / 0.983435.61 / 0.944529.87 / 0.839139.88 / 0.975533.87 / 0.922328.19 / 0.7881
Metabit [18]1.6092244241.04 / 0.978534.92 / 0.936329.25 / 0.823839.11 / 0.969833.18 / 0.912327.69 / 0.7725
Ours4.5647362842.22 / 0.984235.90 / 0.946830.17 / 0.847140.17 / 0.976734.16 / 0.925828.49 / 0.7985
42.24 / 0.984235.91 / 0.946830.19 / 0.847240.18 / 0.976734.17 / 0.925828.52 / 0.7985
+ +![](images/1381f0b495b2f23e1d4c044cbbd8c659778447964fe6f018c4888d88f39a40d8.jpg) +Frame 030,Clip 020,CRF35 + +![](images/81cce0bd7a8cb110d3f4a7e0bb49145379db7bc7a21652676799e0495b2bf111.jpg) + +![](images/c8a72818269e3cf40bff5988f1aebd5f68d5abe872bb04d02da8d7b9ac5e091b.jpg) +Input +Figure 6. Qualitative results on quality enhancement, where our method effectively reduces the compression artifacts, achieving visually pleasant results. In contrast, the results of the compared methods still contain severe distortions (e.g., the calf in the 1st row). + +![](images/af260206614883bcc887d2f9a294dc6adfea2bfc66bb90fee07b277fea3d1575.jpg) + +![](images/4a7a719c8024b9de8a3b047dc00c4a5199f523a97615ee7439b839a5c4ecb88e.jpg) +MFQE 2.0 + +![](images/f12387a4674e6c115e1a8115c794c432337218e58e933bc6c71ca7d518e988db.jpg) + +![](images/d35ee88ccbfd149b7ef0582c3c0637ff0815c1ba910b4e42ce8956ca2754c463.jpg) +STDF + +![](images/6ffadf35ac7f87219e87c799c72bb3e653e041741001322d8b5af50309e5e588.jpg) + +![](images/031ba3b1969564e735cbaffecb85a5c9deb642f3ca62f49f15a608e3a89af3a5.jpg) +S2SVR + +![](images/63cd79b7913622844f8d955d831212b87f77fe12dfc9b8821d8439f02297a927.jpg) + +![](images/facfdbbdb0410ffbf5d756bc64a1bec304b11b26421c3dd97366f016bc607f3d.jpg) +Metabit + +![](images/1310b3cc38a7043f39276b23b5968bfd1ae9e29dd81a1cc087377f6b7698b771.jpg) + +![](images/f267dba5ede5c5e0eb24f37454f5eda60de9267119b2b0168ff39cd0329ca258.jpg) +Ours + +![](images/c1b02fbfb2b15f26231ddf1b81ca81992646123e1ffff490c021ba64bdcacbeb.jpg) + +![](images/30e32489a6322e42fa8c8810cef6cfe3d551e67d9300d3b1e1e35c678d3684ae.jpg) +GT + +Table 1. Quantitative results on quality enhancement, where PSNR and SSIM (higher is better) are adopted for evaluation. The best and second best results are marked with bold and underline. Results obtained by replacing $CRF_{i}$ with slice type are highlighted with grey. + +
MethodCRF15CRF25CRF35
PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
BasicVSR29.24 / 0.821226.19 / 0.713123.40 / 0.6005
+ MFQE 2.029.29 / 0.823326.28 / 0.718223.46 / 0.6056
+ STDF29.31 / 0.824726.51 / 0.729323.80 / 0.6249
+ S2SVR29.45 / 0.828826.70 / 0.734623.86 / 0.6270
+ Metabit29.24 / 0.821126.19 / 0.713123.39 / 0.6005
+ Ours29.54 / 0.832826.85 / 0.741924.02 / 0.6361
IconVSR29.29 / 0.823026.19 / 0.713023.39 / 0.6003
+ MFQE 2.029.37 / 0.825426.28 / 0.718223.45 / 0.6055
+ STDF29.36 / 0.826326.52 / 0.729223.79 / 0.6248
+ S2SVR29.54 / 0.830626.71 / 0.734523.85 / 0.6269
+ Metabit29.29 / 0.823026.19 / 0.713023.39 / 0.6003
+ Ours29.63 / 0.834426.86 / 0.741824.01 / 0.6360
BasicVSR++29.61 / 0.830326.19 / 0.711823.38 / 0.5998
+ MFQE 2.029.66 / 0.832226.27 / 0.716923.44 / 0.6051
+ STDF29.68 / 0.833826.53 / 0.728923.79 / 0.6247
+ S2SVR29.82 / 0.837126.72 / 0.734623.85 / 0.6269
+ Metabit29.61 / 0.830326.19 / 0.711823.38 / 0.5997
+ Ours29.92 / 0.840726.87 / 0.741924.00 / 0.6358
+ +Table 2. Quantitative results of $\times 4$ VSR,where the best and second best results are highlighted with bold and underline. + +to $CRF_{i}$ , with a negligible decrease of PSNR ( $< 0.03\mathrm{dB}$ ), demonstrating the feasibility of replacing $CRF_{i}$ with slice type in practical. The proposed method notably improves the quality of compressed input, achieving a PSNR gain of $1.2\mathrm{dB}$ on CRF15, while MFQE 2.0 and Metabit lead to no improvement. With similar computation cost and inference speed, our method significantly outperforms STDF, obtaining a PSNR gain of $1.09\mathrm{dB}$ on CRF15. Compared to S2SVR, our approach takes only $61\%$ of the parameters and $16\%$ of the FLOPs, and achieves a throughput of 28 FPS, which underlines its practicality. In addition, our method shows robustness and generalization ability on unseen sce + +narios (i.e., CRF18, CRF28 and CRF38), achieving up to 1.06 dB PSNR gain on CRF18. In contrast, other methods trained with mixed compression settings show sub-optimal performance. For instance, STDF and S2SVR only achieve PSNR gains of 0.16 dB and 0.76 dB at CRF18, while MFQE 2.0 shows no improvement. Quantitative results on highly compressed scenarios (i.e., CRF40, CRF45, CRF48) are included in the supplementary materials. + +Qualitative results. Qualitative comparisons are provided in Fig. 6. As can be seen, MFQE 2.0 and Metabit struggle to improve the quality of compressed inputs. Both STDF and S2SVR cannot adequately remove compression artifacts (e.g., boundary of the calf), while the proposed method effectively eliminates the compression artifacts, preserving accurate edges and textures. We provide more qualitative comparisons in the supplementary materials. + +# 5.2.2. Versatility Evaluation + +To evaluate the versatility in assisting practical downstream tasks, we employ the implementation that utilizes slice type for frame adaptation (as described in Sec. 3.1) to enhance compressed inputs for downstream tasks. More qualitative comparisons are included in the supplementary materials. + +Video super-resolution. We adopt BasicVSR [6], IconVSR [6], and BasicVSR++ [7] as the representative baseline methods for $\times 4$ video super-resolution (VSR), which are trained on "clean" data without considering compression. The evaluation is conducted on the REDS4 dataset [39] and summarized with PSNR and SSIM (the higher the better). As depicted in Tab. 2, pre-enhancing compressed inputs with Metabit fails to improve the performance of downstream VSR models. In contrast, pre-enhancing with our + +![](images/f5c6eb61718524fc7c5a8342f81d22f501d9bf5796f3c8c1cbaadf6828f79f56.jpg) +Figure 7. Qualitative results of $\times 4$ VSR. Pre-enhancing with our method before VSR effectively avoids amplifying compression artifacts. While other methods cannot fully eliminate the artifacts and even severe the distortions (e.g., results of MFQE 2.0). + +![](images/4e2d88e0c6875def96096d6ffaf1d66c175b413f59da668e28aca2a8764ae853.jpg) + +![](images/8c1e179d8292670f767ab4404940df61ec6d0182be665f7b81c267e34583176c.jpg) + +![](images/993962136e536e6f824225b3ff342976b971ea0d12452bd0a6fd75eb0b966e93.jpg) + +![](images/0dfc02ac5b0a2171f47befa998a68b11d1d24a62f29fb4849c61ef4d5bfb5892.jpg) + +![](images/994835e8b8fa07e8ef81d769ccdce568f21be3c75c2cd82d8ddbedbed6aed2b7.jpg) + +![](images/c0cc2618c2bb323309cbfafdc82338a25f2ac30777958036345236d75decfc9a.jpg) + +![](images/b615ceab181c851d4c1b245b79a4b142b04e1872f3cbf7e3ab7b083a4e854ace.jpg) + +![](images/69ecbf99b3c4ef8b7ee01333157240fd0cbe8db13f46305369c0b618cbf669c3.jpg) +Figure 8. Qualitative results of optical flow estimation. As can be seen, pre-processing with our method effectively corrects the mispredicted optical flow, especially the boundaries of moving objects (e.g., edge of the moving car). + +![](images/fa82bce0fdd460fcbcac107fd8d28b3f77928da1a28ff658478c587b9fdb1610.jpg) + +![](images/c6ec442c6680ffab03889346826d65d901f29c1ee83cd17c4340610f167a6187.jpg) + +![](images/6cbfefe86bec18a9fa67df32607ae02d0ade8d8bd505c8c57ed512914783aeb5.jpg) + +![](images/c10424d9984437e85177ada66863d9de1fd0344ddc8e296beb06f8a0d0dda1dd.jpg) + +![](images/1c09078aea4a8712df83c5d428948e802ecf2b6f25d4eecae23b74551efed6ee.jpg) + +![](images/6d09086fd4c04906dc99c3fe8b32c9f82a3f55a554a69a68f4582fe5db27d784.jpg) + +
MethodCRF15CRF25CRF35
EPE↓/ F1-all↓EPE↓/ F1-all↓EPE↓/ F1-all↓
RAFT5.26 / 17.817.37 / 22.1316.73 / 44.70
+ MFQE 2.05.32 / 17.837.27 / 21.9216.68 / 44.74
+ STDF5.34 / 17.937.13 / 22.1615.92 / 44.04
+ S2SVR5.22 / 17.716.90 / 21.5715.73 / 44.62
+ Metabit5.32 / 17.867.33 / 22.0716.69 / 44.28
+ Ours5.20 / 17.696.52 / 20.9914.84 / 42.56
DEQ3.99 / 13.715.40 / 17.3313.94 / 41.63
+ MFQE 2.03.97 / 13.735.14 / 17.0614.06 / 41.72
+ STDF4.08 / 13.845.27 / 17.3813.52 / 40.75
+ S2SVR4.01 / 13.695.22 / 16.9913.74 / 41.65
+ Metabit3.92 / 13.785.30 / 17.3313.84 / 41.21
+ Ours3.97 / 13.684.96 / 16.5413.09 / 39.43
KPAFlow4.46 / 16.076.71 / 20.9616.50 / 45.13
+ MFQE 2.04.42 / 15.966.71 / 20.9216.70 / 45.29
+ STDF4.52 / 16.196.96 / 21.5315.62 / 44.17
+ S2SVR4.37 / 15.966.10 / 19.7615.62 / 44.23
+ Metabit4.47 / 16.176.75 / 21.0816.68 / 44.90
+ Ours4.43 / 16.105.59 / 18.7314.83 / 41.24
+ +framework yields consistent improvement especially in scenarios of high compression (e.g., up to 0.62 dB PSNR gain with BaiscVSR++ on CRF35). Meanwhile, our method significantly outperforms MFQE 2.0 and STDF, achieving PSNR gains of 0.25 dB and 0.23 dB over MFQE 2.0 and STDF on BasicVSR/CRF15, respectively. Compared with S2SVR, the proposed method offers more effective support to VSR models with lower complexity. As shown in Fig. 7, performing VSR on compressed data inevitably amplifies compression artifacts (e.g., the 1st column), while results pre-enhanced with our method maintain accurate edges and textures, avoiding distortions seen in other methods. + +Optical flow estimation. We adopt RAFT [49], DEQ [3], and KPAFlow [35] as baseline models for optical flow estimation. Evaluation on the KITTI-2015 dataset [21] is summarized with EPE (end-point-error) and F1-all loss, where lower values indicate better accuracy. As shown in Tab. 3, + +Table 3. Quantitative results of optical flow estimation, where we highlight the best and second best results with **bold** and **underline**. + +
MethodCRF15CRF25CRF35
Avg↑/J↑/F↑Avg↑/J↑/F↑Avg↑/J↑/F↑
STCN85.07 / 81.83 / 88.3284.35 / 80.96 / 87.7479.20 / 76.04 / 82.37
+ MFQE 2.084.96 / 81.71 / 88.2184.30 / 80.92 / 87.6979.28 / 76.11 / 82.44
+ STDF85.01 / 81.73 / 88.2884.23 / 80.97 / 87.5079.77 / 76.52 / 83.02
+ S2SVR85.17 / 81.93 / 88.4184.46 / 81.20 / 87.7280.04 / 76.88 / 83.20
+ Metabit84.56 / 80.97 / 88.1483.86 / 80.18 / 87.5579.03 / 75.65 / 82.40
+ Ours85.21 / 81.99 / 88.4484.63 / 81.42 / 87.8581.57 / 78.46 / 84.69
DeAoT85.90 / 82.89 / 88.9185.18 / 82.37 / 88.0082.87 / 79.86 / 85.88
+ MFQE 2.085.86 / 82.84 / 88.8885.20 / 82.38 / 88.0382.86 / 79.86 / 85.85
+ STDF85.83 / 82.80 / 88.8785.18 / 82.27 / 88.0982.90 / 79.92 / 85.89
+ S2SVR86.05 / 83.09 / 89.0185.05 / 82.07 / 88.0482.64 / 79.63 / 85.65
+ Metabit85.47 / 82.04 / 88.9084.95 / 81.57 / 88.3382.32 / 79.04 / 85.59
+ Ours86.08 / 83.13 / 89.0385.31 / 82.38 / 88.2582.88 / 79.83 / 85.92
QDMN85.16 / 82.20 / 88.1184.16 / 81.20 / 87.1279.39 / 76.61 / 82.18
+ MFQE 2.085.13 / 82.20 / 88.0684.15 / 81.18 / 87.1379.51 / 76.75 / 82.27
+ STDF85.32 / 82.38 / 88.2783.36 / 80.44 / 86.2779.64 / 76.69 / 82.59
+ S2SVR85.28 / 82.32 / 88.2383.64 / 80.65 / 86.6379.02 / 76.15 / 81.89
+ Metabit84.50 / 81.14 / 87.8783.68 / 80.30 / 87.0679.47 / 76.41 / 82.52
+ Ours85.34 / 82.41 / 88.2784.37 / 81.42 / 87.3279.78 / 76.92 / 82.65
+ +Table 4. Quantitative results of VOS, where the best and second best results are highlighted with bold and underline. + +our method consistently reduces the EPE and F1-all loss across all baseline models, demonstrating its effectiveness in improving optical flow estimation. In contrast, methods such as MFQE 2.0, STDF and Metabit fail to deliver consistent improvements. For instance, MFQE 2.0 fails to improve the performance of RAFT on CRF15, STDF and Metabit detrimentally affects the performance of DEQ and KPAFlow on CRF15. Visualizations of predicted optical flow are shown in Fig. 8, where inaccurate boundaries are highlighted with red arrows. As can be seen, optical flow estimated from compressed inputs contains inaccurate boundaries, especially near-motion ones. The proposed method helps to deliver more accurate results in these regions compared to others. For instance, it effectively corrects the optical flow near the car that was mispredicted by DEQ, while MFQE 2.0 and S2SVR provide limited improvement. + +Video object segmentation. For video object segmentation + +![](images/ff5ff14a1891bd7c150379f6b76b20dbab27f79bd5c4045498a8cf3efc05f64e.jpg) +STCN + +![](images/b2fecfcbf265c390fa5c1b1935df9222a35f4bb360f2bc39096046aa6f6d3252.jpg) ++ MFQE 2.0 + +![](images/69abf39d07991c266f382e0ef5b2ed082e89560ca6bdb79824d20c24e2b45eab.jpg) ++STDF + +![](images/29d9116db90f2ad277f8a8a2503bdd1ad3e70d2fe8b603ef04dafcc6b0b7ae9f.jpg) ++ S2SVR + +![](images/285533b831b80dec5f2bd8146554e3be0953d9476c16712daba7b15aa0c33006.jpg) ++Metabit + +![](images/1d976055777897c391077cbf53cf4797495eac3730ccd0db47c248535761bbe4.jpg) ++Ours + +![](images/ed21269d8b4496c4a9c642558390e11dbf3b8dad7c8f9afc7467683bd0955860.jpg) +GT + +![](images/9e0c90e81a7c68ecab8eec8d32477e758ccba7bf83db448c0c21b05a7da3ef72.jpg) +Figure 9. Qualitative results of VOS. As can be seen, directly performing VOS on compressed inputs leads to inaccurate masks, whereas pre-enhancing with our method effectively improves the accuracy, especially for the regions of irregular shapes (e.g., the windshield). +Masked frame +Figure 10. Qualitative results of video inpainting. As can be seen, pre-enhancing the compressed inputs with the proposed method helps to reduce the artifacts and color distortion in the removed region, providing more visually pleasant results. + +![](images/39e963a30d0bb188e87f7203be7d131e6cd49dc8e91fdcba0981bdce1546b408.jpg) +$E^2$ FGVI + +![](images/353d0f65ab54f34c566199b26ad022236da607bea4190039b49d117aa7871c32.jpg) ++ MFQE 2.0 + +![](images/71b79d6c54308e89a5dc76bd0d742ce0b298b3658dd878e54a4e811f3d8a016e.jpg) ++STDF + +![](images/02add9649d9c6e82a9ee5e522da514b181cd8f884c52d19d5aa5c2558adf03e6.jpg) ++ S2SVR + +![](images/479b395931600a7b12769eed239b2d28969843af9abf1730f8fc8df5b229c082.jpg) ++Metabit + +![](images/2c13d21a0518fb74cda51b730621a1774293d4d5eddc95055d2f8a484d5dd541.jpg) ++Ours + +(VOS), we adopt STCN [14], DeAoT [57] and QDMN [34] as representative baselines. Evaluations on DAVIS-17 val dataset [41] are summarized with the following metrics: the $\mathcal{I}$ (average IoU), the $\mathcal{F}$ score (boundary similarity), and the average of the above metrics (denoted as $Avg$ ). Higher values indicate better segmentation accuracy. As shown in Tab. 4, the proposed method shows the best performance in improving accuracy across VOS models. For instance, it elevates the average accuracy for up to $2.37\%$ (79.20% to $81.57\%$ ) on STCN at CRF35, while MFQE 2.0, STDF and S2SVR yield limited improvement of $0.08\%$ , $0.57\%$ and $0.84\%$ , respectively. And Metabit provides no improvement on STCN at CRF35. The results of VOS are included in Fig. 9, where accurately segmenting objects in compressed videos is challenging for VOS baselines (e.g., inaccurate mask of the windshield predicted by STCN). Pre-enhancing the compressed videos with MFQE 2.0 and S2SVR struggles to address this issue, whereas the proposed method significantly refines the segmentation results, demonstrating its effectiveness in assisting the VOS task. + +Video inpainting. We take $\mathrm{E}^2\mathrm{FGVI}$ [31] as the video inpainting model and perform video object removal on DAVIS-17 val dataset [41]. The qualitative results are shown in Fig. 10. As can be seen, compression-included misalignment between objects and masks hinders the ability to remove specified objects, causing color distortions (e.g., the horse region). Pre-enhancing the compressed inputs with the proposed method notably refines the artifacts and distortions, yielding more visually pleasing results. + +# 5.3. Ablation Studies + +We start with a baseline that concatenates reference frames and current frames as input, without using codec information. We then progressively equip the baseline with MV alignment, region-aware refinement, sequence adaptation, and frame adaptation to assess their contributions. + +MV alignment. As shown in the 2nd row of Tab. 5, incorporating MV alignment yields a PSNR gain of up to 0.65 dB on CRF15, demonstrating the effectiveness of MV in + +
ModelCRF15CRF25CRF35
PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
Baseline41.04 / 0.978534.92 / 0.936329.25 / 0.8238
+ MV Align.41.69 / 0.982135.59 / 0.943729.95 / 0.8403
+ RA Refine.42.04 / 0.983735.70 / 0.944930.00 / 0.8427
+ Seq. Adapt.42.08 / 0.983835.76 / 0.945830.04 / 0.8444
+ Frame Adapt.42.14 / 0.983935.81 / 0.946030.09 / 0.8446
+ +Table 5. Ablation studies on MV alignment, region-aware refinement, sequence adaptation, and frame adaptation. + +aligning reference frames and current frame. + +Region-aware refinement. We further incorporate the region-aware refinement module to refine the features of different regions. As shown in the 3rd row of Tab. 5, it leads to notable PSNR gains of $0.35\mathrm{dB}$ , $0.11\mathrm{dB}$ and $0.05\mathrm{dB}$ on CRF15, CRF25 and CRF35, respectively. + +Sequence adaptation. As shown in the 4th row of Tab. 5, sequence adaptation brings PSNR gains of $0.04\mathrm{dB}$ , $0.06\mathrm{dB}$ and $0.04\mathrm{dB}$ on CRF15, CRF25 and CRF35, respectively. + +Frame adaptation. As shown in the 5th row of Tab. 5, frame adaptation improves PSNR by $0.06\mathrm{dB}$ $0.05\mathrm{dB}$ and $0.05\mathrm{dB}$ on CRF15, CRF25 and CRF35, respectively. We further analyze its effectiveness on improving the temporal consistency in the supplementary materials. + +# 6. Conclusion + +In this paper, we introduce a versatile codec-aware enhancement framework that adaptively handles diverse compression settings and serves as a plug-and-play enhancement module to consistently boost various downstream tasks. By reusing the off-the-shelf codec information, our method minimizes additional computational costs. Compared with existing compressed video enhancement solutions, it shows superiority in both enhancement performance and robustness, making it possible to deploy pre-trained models on compressed videos without a significant performance drop. + +Acknowledgments. We acknowledge funding from the National Natural Science Foundation of China under Grants 62131003 and 62021001. + +# References + +[1] Video developer report 2022-2023. https://bitmovin.com/downloads/ assets/bitmovin-7th-video-developer-report-2023-2024.pdf. 5 +[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In CVPR, pages 3366-3375, 2017. 2 +[3] Shaojie Bai, Zhengyang Geng, Yash Savani, and J. Zico Kolter. Deep equilibrium optical flow estimation. In CVPR, 2022. 7 +[4] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 3 +[5] Yuxuan Cai, Hongjia Li, Geng Yuan, Wei Niu, Yanyu Li, Xulong Tang, Bin Ren, and Yanzhi Wang. Yolobile: Realtime object detection on mobile devices via compression-compiler co-design. In AAAI, pages 955–963, 2021. 1 +[6] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, pages 4947-4956, 2021. 6 +[7] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, pages 5972-5981, 2022. 6 +[8] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In CVPR, pages 5962-5971, 2022. 1 +[9] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In ICIP, pages 168-172, 1994. 5 +[10] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In CVPR, pages 8064-8073, 2021. 3 +[11] Peilin Chen, Wenhan Yang, Long Sun, and Shiqi Wang. When bitstream prior meets deep prior: Compressed video super-resolution with learning from decoding. In ACM MM, pages 1000–1008, 2020. 1, 2 +[12] Peilin Chen, Wenhan Yang, Meng Wang, Long Sun, Kangkang Hu, and Shiqi Wang. Compressed domain deep video super-resolution. IEEE Transactions on Image Processing, 30:7156-7169, 2021. 1, 2 +[13] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In CVPR, pages 11030-11039, 2020. 2 +[14] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 8 +[15] Yuanying Dai, Dong Liu, and Feng Wu. A convolutional neural network approach for post-processing in hevc intra coding. In MultiMedia Modeling: 23rd International Con + +ference, MMM 2017, Reykjavik, Iceland, January 4-6, 2017, Proceedings, Part I 23, pages 28-39. Springer, 2017. 2 +[16] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. In AAAI, pages 10696-10703, 2020. 1, 2, 5, 6 +[17] Cunhui Dong, Haichuan Ma, Zhuoyuan Li, Li Li, and Dong Liu. Temporal wavelet transform-based low-complexity perceptual quality enhancement of compressed video. IEEE TCSVT, pages 1-1, 2023. 2 +[18] Max Ehrlich, Jon Barker, Namitha Padmanabhan, Larry Davis, Andrew Tao, Bryan Catanzaro, and Abhinav Shrivastava. Leveraging bitstream metadata for fast, accurate, generalized compressed video quality enhancement. In WACV, pages 1517-1527, 2024. 1, 2, 5, 6 +[19] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 2 +[20] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaeser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2020. 1 +[21] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 7 +[22] Zhenyu Guan, Qunliang Xing, Mai Xu, Ren Yang, Tie Liu, and Zulin Wang. Mfqe 2.0: A new approach for multi-frame quality enhancement on compressed video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3): 949-963, 2019. 1, 2, 5, 6 +[23] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7436-7456, 2021. 2 +[24] Hongyue Huang, Ionut Schiopu, and Adrian Munteanu. Frame-wise cnn-based filtering for intra-frame quality enhancement of hevc videos. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2100-2113, 2020. 2 +[25] Zhijie Huang, Jun Sun, and Xiaopeng Guo. Fastconn: Towards fast and accurate spatiotemporal network for hevc compressed video enhancement. ACM Transactions on Multimedia Computing, Communications and Applications, 19 (3):1-22, 2023. 2 +[26] Chuanmin Jia, Shiqi Wang, Xinfeng Zhang, Shanshe Wang, Jiaying Liu, Shiliang Pu, and Siwei Ma. Content-aware convolutional neural network for in-loop filtering in high efficiency video coding. IEEE Transactions on Image Processing, 28(7):3343-3356, 2019. 2 +[27] Nanfeng Jiang, Weiling Chen, Jielian Lin, Tiesong Zhao, and Chia-Wen Lin. Video compression artifacts removal with spatial-temporal attention-guided enhancement. IEEE Transactions on Multimedia, 2023. 2 + +[28] Feng Li, Yixuan Wu, Anqi Li, Huihui Bai, Runmin Cong, and Yao Zhao. Enhanced video super-resolution network towards compressed data. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-21, 2024. 2 +[29] Yinxiao Li, Pengchong Jin, Feng Yang, Ce Liu, Ming-Hsuan Yang, and Peyman Milanfar. Comisr: Compression-informed video super-resolution. In ICCV, pages 2543-2552, 2021. 2 +[30] Yunsheng Li, Lu Yuan, Yinpeng Chen, Pei Wang, and Nuno Vasconcelos. Dynamic transfer for multi-source domain adaptation. In CVPR, pages 10998-11007, 2021. 3 +[31] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In CVPR, 2022. 8 +[32] Jing Lin, Xiaowan Hu, Yuanhao Cai, Haoqian Wang, Youliang Yan, Xueyi Zou, Yulun Zhang, and Luc Van Gool. Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In ICML, pages 13394-13404. PMLR, 2022. 1, 2, 5, 6 +[33] Baoyuan Liu, Min Wang, Hassan Foroosh, Marshall Tappen, and Marianna Pensky. Sparse convolutional neural networks. In CVPR, pages 806-814, 2015. 5 +[34] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, pages 468-486, 2022. 8 +[35] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In CVPR, pages 8906-8915, 2022. 7 +[36] Di Ma, Fan Zhang, and David R Bull. Cvegan: a perceptually-inspired gan for compressed video enhancement. Signal Processing: Image Communication, page 117127, 2024. 2 +[37] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2 +[38] Sreyas Mohan, Joshua L Vincent, Ramon Manzorro, Peter Crozier, Carlos Fernandez-Granda, and Eero Simoncelli. Adaptive denoising via gaintuning. NeurIPS, 34:23727-23740, 2021. 3 +[39] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, pages 0–0, 2019. 5, 6 +[40] Zhaoqing Pan, Xiaokai Yi, Yun Zhang, Byeungwoo Jeon, and Sam Kwong. Efficient in-loop filtering based on enhanced deep convolutional neural networks for hevc. IEEE Transactions on Image Processing, 29:5352-5366, 2020. 2 +[41] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 8 + +[42] Darren Ramsook and Anil Kokaram. Learnt deep hyperparameter selection in adversarial training for compressed video enhancement with a perceptual critic. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2420-2424. IEEE, 2023. 2 +[43] Mennatullah Siam, Mostafa Gamal, Moemen Abdel-Razek, Senthil Yogamani, Martin Jagersand, and Hong Zhang. A comparative study of real-time semantic segmentation for autonomous driving. In CVPRW, pages 587-597, 2018. 1 +[44] Mennatullah Siam, Alex Kendall, and Martin Jagersand. Video class agnostic segmentation benchmark for autonomous driving. In CVPR, pages 2825-2834, 2021. 1 +[45] Gary J Sullivan, Pankaj N Topiwala, and Ajay Luthra. The h. 264/avc advanced video coding standard: Overview and introduction to the fidelity range extensions. Applications of Digital Image Processing XXVII, 5558:454-474, 2004. 5 +[46] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 3 +[47] Weiheng Sun, Xiaohai He, Chao Ren, Shuhua Xiong, and Honggang Chen. A quality enhancement network with coding priors for constant bit rate video coding. Knowledge-Based Systems, 258:110010, 2022. 1, 2 +[48] Xuebin Sun, Sukai Wang, Miaohui Wang, Shing Shin Cheng, and Ming Liu. An advanced lidar point cloud sequence coding scheme for autonomous driving. In MM, pages 2793-2801, 2020. 1 +[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020. +[50] Jianyi Wang, Mai Xu, Xin Deng, Liquan Shen, and Yuhang Song. Mw-gan+ for perceptual quality enhancement on compressed video. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4224-4237, 2021. 2 +[51] Shiyao Wang, Hongchao Lu, and Zhidong Deng. Fast object detection in compressed video. In ICCV, pages 7104-7113, 2019. 1 +[52] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Hutchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In CVPR, pages 2012-2021, 2023. 2 +[53] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 3 +[54] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. NeurIPS, 32, 2019. 2 +[55] Ren Yang, Mai Xu, Zulin Wang, and Tianyi Li. Multi-frame quality enhancement for compressed video. In CVPR, pages 6664-6673, 2018. 1, 2 +[56] Ren Yang, Mai Xu, Tie Liu, Zulin Wang, and Zhenyu Guan. Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology, 29(7): 2039-2054, 2019. 2 +[57] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation. In NeurIPS, 2022. 8 + +[58] Li Yu, Wenshuai Chang, Shiyu Wu, and Moncef Gabbouj. End-to-end transformer for compressed video quality enhancement. IEEE Transactions on Broadcasting, 70(1):197-207, 2024. 2 +[59] Haichao Zhang, Jianchao Yang, Yanning Zhang, Nasser M Nasrabadi, and Thomas S Huang. Close the loop: Joint blind image restoration and recognition with sparse representation prior. In ICCV, pages 770-777, 2011. 1 +[60] Hengsheng Zhang, Xueyi Zou, Jiaming Guo, Youliang Yan, Rong Xie, and Li Song. A codec information assisted framework for efficient compressed video super-resolution. In ECCV, pages 220-235. Springer, 2022. 2 +[61] Saiping Zhang, Luis Herranz, Marta Mrak, Marc Górriz Blanch, Shuai Wan, and Fuzheng Yang. Dcngan: A deformable convolution-based gan with qp adaptation for perceptual quality enhancement of compressed video. In ICASSP, pages 2035-2039. IEEE, 2022. 2 +[62] Zhengdong Zhang and Vivienne Sze. Fast: A framework to accelerate super-resolution processing on compressed videos. In CVPRW, pages 19-28, 2017. 2 +[63] Hengrun Zhao, Bolun Zheng, Shanxin Yuan, Hua Zhang, Chenggang Yan, Liang Li, and Gregory Slabaugh. Cbren: Convolutional neural networks for constant bit rate video quality enhancement. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4138-4149, 2022. 2 +[64] Qiang Zhu, Jinhua Hao, Yukang Ding, Yu Liu, Qiao Mo, Ming Sun, Chao Zhou, and Shuyuan Zhu. Cpga: Coding priors-guided aggregation network for compressed video quality enhancement. In CVPR, pages 2964-2974, 2024. 2 \ No newline at end of file diff --git a/2025/Plug-and-Play Versatile Compressed Video Enhancement/images.zip b/2025/Plug-and-Play Versatile Compressed Video Enhancement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..142dfefc9945b403016c4a67575262c20ca0de46 --- /dev/null +++ b/2025/Plug-and-Play Versatile Compressed Video Enhancement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6f29e1dd1414cd8e4de8c0a53932970635aa2044145792d59d04f3e00db18da +size 916424 diff --git a/2025/Plug-and-Play Versatile Compressed Video Enhancement/layout.json b/2025/Plug-and-Play Versatile Compressed Video Enhancement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cab7613be581a94a11a69c1b7d5ed70f1165158f --- /dev/null +++ b/2025/Plug-and-Play Versatile Compressed Video Enhancement/layout.json @@ -0,0 +1,12247 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "text", + "content": "Plug-and-Play Versatile Compressed Video Enhancement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 183, + 143, + 430, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 143, + 430, + 158 + ], + "spans": [ + { + "bbox": [ + 183, + 143, + 430, + 158 + ], + "type": "text", + "content": "Huimin Zeng Jiacheng Li Zhiwei Xiong*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 158, + 421, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 158, + 421, + 172 + ], + "spans": [ + { + "bbox": [ + 189, + 158, + 421, + 172 + ], + "type": "text", + "content": "University of Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 173, + 463, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 173, + 463, + 185 + ], + "spans": [ + { + "bbox": [ + 143, + 173, + 463, + 185 + ], + "type": "text", + "content": "{zenghuimin, jciee}@mail.ustc.edu.cn zwxiong@ustc.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 239, + 297, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 525 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 525 + ], + "type": "text", + "content": "As a widely adopted technique in data transmission, video compression effectively reduces the size of files, making it possible for real-time cloud computing. However, it comes at the cost of visual quality, posing challenges to the robustness of downstream vision models. In this work, we present a versatile codec-aware enhancement framework that reuses codec information to adaptively enhance videos under different compression settings, assisting various downstream vision tasks without introducing computation bottleneck. Specifically, the proposed codec-aware framework consists of a compression-aware adaptation (CAA) network that employs a hierarchical adaptation mechanism to estimate parameters of the frame-wise enhancement network, namely the bitstream-aware enhancement (BAE) network. The BAE network further leverages temporal and spatial priors embedded in the bitstream to effectively improve the quality of compressed input frames. Extensive experimental results demonstrate the superior quality enhancement performance of our framework over existing enhancement methods, as well as its versatility in assisting multiple downstream tasks on compressed videos as a plug-and-play module. Code and models are available at https://huimin-zeng.github.io/PnP-VCVE/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 543, + 135, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 135, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 135, + 555 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "type": "text", + "content": "With the flower booming of short video platforms, video has become one of the most popular multimedia formats. In addition to distributing visual content, in practical scenarios (e.g., autonomous driving [43, 48]), it is common to upload the captured videos to the cloud end for further visual analysis and downstream applications (e.g., object detection [5, 51] and segmentation [20, 44]). However, due to the bandwidth constraint during transmission, these videos are typically compressed with varying levels, resulting in poor visual quality and suboptimal performance in downstream tasks [22, 59] (e.g., inaccurate segmentation bound" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 212, + 555, + 357 + ], + "blocks": [ + { + "bbox": [ + 316, + 212, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 316, + 212, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 212, + 555, + 357 + ], + "type": "image", + "image_path": "9b5952d352646c15004795393533a8a5ef99b2c6f588ac52605ac07b59022c64.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "type": "text", + "content": "Figure 1. The proposed codec-aware enhancement framework reuses codec information to adaptively enhance videos across different compression settings, while assisting in various downstream tasks in a plug-and-play manner." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 422, + 555, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 470 + ], + "type": "text", + "content": "aries in Fig. 1). Given the crucial role of videos in data transmission, there is a critical need for a versatile solution to enhance videos of diverse compression levels and effectively support various downstream tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 639 + ], + "type": "text", + "content": "Existing video enhancement methods are hard to respond to these demands. Specifically, to effectively enhance videos of different compression levels, previous methods [11, 12, 16, 22, 32, 55] employ separate enhancement models for each compression level, which is inflexible occurring unseen compression levels. Recent approaches [8, 18, 47] consider this issue as the generalization ability across diverse compression levels, therefore randomly selecting inputs of different compression levels during training. However, such a training strategy is compression-agnostic and offers limited improvement. Most importantly, the aforementioned methods focus primarily on improving perceptual quality, neglecting the need to assist in downstream tasks in real-world scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Based on the mismatch between the versatility demand and existing solutions, here we summarize the following criteria of a favorable solution: 1) adaptively enhance videos of varying compression levels with a single model; 2) effectively assist various downstream tasks on compressed videos in a plug-and-play manner; and 3) given the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17767" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "practical scenarios where real-time processing is required, it should meet the above objectives without causing a computation bottleneck. To achieve this, we introduce a codec-aware enhancement framework (as shown in Fig. 1) that reuses codec information embedded in the bitstream. By incorporating compression factors, the framework dynamically adjusts its parameters to flexibly enhance inputs of different compression levels. By reusing motion vectors and partition maps, it efficiently aggregates temporal and spatial clues without introducing redundant computations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 193, + 296, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 193, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 296, + 454 + ], + "type": "text", + "content": "Specifically, our codec-aware enhancement framework comprises a compression-aware adaptation (CAA) network and a bitstream-aware enhancement (BAE) network. The CAA network serves as a \"meta\" network that dynamically adjusts the parameters of the subsequent BAE network. A hierarchical adaptation mechanism is proposed to first estimate sequence-adaptive parameters based on the sequence compression level, and then re-weight these parameters according to the frame compression level, thereby achieving a BAE network tailored for each input frame. The frame-adaptive BAE network conducts motion vector alignment to aggregate intra-frame information and provide useful clues for the current frame. Subsequently, based on the region complexity indicated by partition maps, the region-aware refinement assigns independent filters for different regions, achieving flexible enhancement for different regions. Comprehensive experiments are conducted to demonstrate the superiority of our method in improving the quality of compressed videos, and the effectiveness of assisting in various downstream tasks (i.e., video super-resolution, optical flow estimation, and video object segmentation). Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 455, + 295, + 621 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 502 + ], + "type": "text", + "content": "- We present a codec-aware framework for versatile compressed video enhancement, which adaptively enhances input videos of different compression levels and supports a wide range of downstream vision tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 503, + 295, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 573 + ], + "type": "text", + "content": "- We develop a compression-aware adaptation (CAA) network and a bitstream-aware enhancement (BAE) network that utilize the off-the-shelf codec information, contributing to generalizing across different compression settings and boosting the enhancement performance with a unified framework." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 574, + 295, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 621 + ], + "type": "text", + "content": "- Experimental results show the superiority of our method over existing enhancement methods, and its effectiveness in serving as a plug-and-play enhancement module to assist in downstream tasks." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 632, + 142, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 142, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 142, + 645 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 650, + 233, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 233, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 233, + 662 + ], + "type": "text", + "content": "2.1. Compressed Video Enhancement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Existing compressed video enhancement methods can be categorized into in-loop and post-processing methods. Although in-loop methods [15, 24, 26, 40] effectively improve the quality of reconstructed frames, they embed filters" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "text", + "content": "in the encoding and decoding loops, therefore not suitable for enhancing already compressed videos. While the post-processing methods [17, 25, 27, 36, 42, 47, 50, 56, 58, 61, 63, 64] provide more practical solutions to enhance compressed videos by placing filters at the decoder side. Observing the quality fluctuation across frames, MFQE [55] locates the peak quality frame (PQF) with an SVM-based detector and proposes a multi-frame quality enhancement mechanism to enhance non-PQFs. MFQE 2.0 [22] further designs a BiLSTM-based detector and performs multiframe quality enhancement for both non-PQF and PQF. To address the inaccuracies in optical flow estimation from compressed videos, STDF [16] proposes estimating the offset field using spatio-temporal deformable convolution. S2SVR [32] introduces a sequence-to-sequence network to model long-range dependencies within frames. The aforementioned methods inflexibly equip a separate model for each compression level, while we propose to adaptively handle diverse inputs with a single unified model. Recent methods [18, 64] utilize spatial priors from bitstream to address multiple compression levels, however, they only consider I/P-frames, whereas we design a hierarchical adaptation mechanism to address all types of frames." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 354, + 513, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 513, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 513, + 365 + ], + "type": "text", + "content": "2.2. Codec-Aware Video Super-Resolution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 369, + 556, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 573 + ], + "type": "text", + "content": "Some works in video super-resolution (VSR) [11, 12, 28, 29, 52, 60, 62] explore ways of using codec information such as motion estimation and spatial prior for reconstruction. COMISR [29] focuses on reducing accumulated warping errors caused by the random locations of the intra-frame from compressed video frames. Chen et al. [11] employ motion vectors to build the temporal relationship and suppress coding artifacts. CVCP [12] utilizes motion vectors and spatial priors with a guided soft alignment scheme and guided SFT layer, respectively. CIAF [60] leverages motion vectors and residuals to model temporal relationships and skip redundant computations, respectively. Despite leveraging codec information, these methods focus on a single task (i.e., VSR). In contrast, our method not only shows competitive performance in VSR (see supplementary materials), but also effectively supports a range of downstream tasks, which is not explored by the aforementioned works." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 578, + 463, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 463, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 463, + 590 + ], + "type": "text", + "content": "2.3. Dynamic Neural Networks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Instead of setting separate models for different inputs, the early mixture of expert (MoE) structure [2, 19, 37] constructs parallel network branches and selectively executes branches to obtain the weighted outputs. Instead of increasing the number of parallel branches, the dynamic parameters ensemble strategy [2, 13, 19, 23, 37, 54] presets parallel expert layers and selectively fusing their parameters to promote the network capability and generalization ability, therefore serving as an efficient alternative to MoE. To promote the generalization ability of pre-trained models, Gain-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17768" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 291, + 201 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 291, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 291, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 291, + 201 + ], + "type": "image", + "image_path": "252eafdcfa9fa26beb5db55d2291032b16fbdbcb0b367a33ac02f6e8a863c7c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 205, + 295, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 261 + ], + "type": "text", + "content": "Figure 2. Hierarchical structure of quality adjustment, where frames are divided into multiple groups of pictures (GOP). The Constant Rate Factor (CRF) affects video quality at both sequence and frame levels. An increase in the CRF value indicates a reduction in video quality (e.g., lower PSNR values)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 273, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 273, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 295, + 429 + ], + "type": "text", + "content": "tune [38] proposes to predict a single multiplicative scaling parameter for each channel according to test samples, thus modifying static models to test-adaptive ones. Li et al. [30] handle the conflicts between the domain-agnostic model and multiple target domains with dynamic transfer, which is simply modeled by combining residual matrices and a static convolution matrix. DRConv [10] divides the input image into different regions with a learnable mask and assigns multiple filters for these regions, which enhances the feature representation ability without introducing a noticeable computation burden. Instead of searching and learning conditions for dynamic parameters, we leverage priors embedded in the bitstream as the condition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 435, + 141, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 141, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 141, + 449 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 525 + ], + "type": "text", + "content": "We take H.264 [53] as a representative standard to analyze available codec information. Note recent codecs [4, 46] also provide similar priors (i.e., CRF, motion vector, partition map), therefore assuring the applicability of our method. To reduce transmission bandwidth, codecs compress videos by adjusting quality and reducing redundancy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 530, + 234, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 234, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 234, + 544 + ], + "type": "text", + "content": "3.1. Hierarchical Quality Adjustment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "content": "Video quality is commonly influenced by the constant rate factor (CRF), which involves hierarchical adjustment for both sequence-wise and frame-wise compression." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": "Sequence-wise CRF. The CRF ranges from 0 to 51 to balance compression efficiency and visual quality. A higher CRF results in more compact output but increased pixel loss (e.g., the average PSNR of CRF35 is much lower than CRF15 in Fig. 2). By considering the sequence-wise CRF (denoted as " + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "inline_equation", + "content": "CRF_{s}" + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": "), the enhancement network can be tailored to handle videos of different compression levels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Frame-wise CRF As shown in Fig. 2, a video sequence is divided into multiple groups of pictures (GOP) and further categorized as I-frames, P-frames, or B-frames. The CRF value of each frame (denoted as " + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": ") is dynamically" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "adjusted based on " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "CRF_{s}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " so that lower " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " is assigned for I/P frames to maintain quality and higher " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " for B frames for compact representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "text", + "content": "Inspired by the hierarchical quality adjustment paradigm, we design a hierarchical adaptation paradigm that first performs sequence adaptation to predict network parameters based on " + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "inline_equation", + "content": "CRF_{s}" + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "text", + "content": ", and then re-weights these parameters according to " + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "text", + "content": " for frame adaptation. In practical scenarios where " + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 313, + 108, + 555, + 216 + ], + "type": "text", + "content": " is unavailable (e.g., limited access to full bitstream), the proposed method can instead use slice type (I/P/B) for frame adaptation, which is demonstrated to yield similar performance in Sec. 5.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 220, + 447, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 447, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 447, + 232 + ], + "type": "text", + "content": "3.2. Redundancy Reduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 235, + 553, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 553, + 282 + ], + "type": "text", + "content": "To reduce redundancy and improve the entropy of the bitstream,Codec block-wisely perform motion estimation to model the intra-frame correlations and embed the correlations in the bitstream for decoding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 283, + 554, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 554, + 402 + ], + "type": "text", + "content": "Partition map. As shown in Fig. 1, different regions of each frame are partitioned into blocks of varying sizes (e.g., H.264 provides macroblocks of " + }, + { + "bbox": [ + 313, + 283, + 554, + 402 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 283, + 554, + 402 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 283, + 554, + 402 + ], + "type": "inline_equation", + "content": "16 \\times 8" + }, + { + "bbox": [ + 313, + 283, + 554, + 402 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 283, + 554, + 402 + ], + "type": "inline_equation", + "content": "8 \\times 16" + }, + { + "bbox": [ + 313, + 283, + 554, + 402 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 283, + 554, + 402 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 283, + 554, + 402 + ], + "type": "text", + "content": ") according to the texture complexity. Flat regions (e.g., the ground) can tolerate higher quantization errors and are therefore divided into blocks of large size, while complex regions (e.g., leaves and fence) take smaller blocks to maintain details. To effectively enhance regions of different complexity, we propose dynamically assigning filters based on the partition map that indicates region complexity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 403, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 498 + ], + "type": "text", + "content": "Motion vector. Motion vectors are utilized in decoding to aggregate information from reference frames and propagate information of current frame. As illustrated in Fig. 1, they describe the relationship between current frame and its reference frames in a block-wise manner. Although motion vectors can be noisy and are less precise than optical flow, they effectively align reference frames with current frame, serving as a cost-effective alternative for optical flow." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 506, + 534, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 534, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 534, + 518 + ], + "type": "text", + "content": "4. Codec-Aware Enhancement Framework" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 523, + 382, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 382, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 382, + 534 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 539, + 554, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 554, + 634 + ], + "type": "text", + "content": "As shown in Fig. 3(a), the proposed method comprises a compression-aware adaption (CAA) network " + }, + { + "bbox": [ + 313, + 539, + 554, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi}" + }, + { + "bbox": [ + 313, + 539, + 554, + 634 + ], + "type": "text", + "content": " and a bitstream-aware enhancement (BAE) network " + }, + { + "bbox": [ + 313, + 539, + 554, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta_i}" + }, + { + "bbox": [ + 313, + 539, + 554, + 634 + ], + "type": "text", + "content": ". The CAA network employs a hierarchical compression adaptation mechanism to estimate parameters for the frame-adaptive BAE network, which then aggregates intra-frame information and performs region-aware refinement to enhance the input compressed frame." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 638, + 531, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 531, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 531, + 651 + ], + "type": "text", + "content": "4.2. Compression-Aware Adaptation Network" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "To handle sequences of varying compression levels and quality fluctuations across frames, as illustrated in Fig. 3(a), the CAA network " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi}" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " utilizes " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "CRF_{s}" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " for sequence adaptation to estimate sequence-adaptive parameters, and perform frame adaptation to refine these parameters based on " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17769" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 66, + 523, + 212 + ], + "blocks": [ + { + "bbox": [ + 88, + 66, + 523, + 212 + ], + "lines": [ + { + "bbox": [ + 88, + 66, + 523, + 212 + ], + "spans": [ + { + "bbox": [ + 88, + 66, + 523, + 212 + ], + "type": "image", + "image_path": "f2e6968fdb4faa56129914f2aa33472b03e6f6b132ac64863568be3c7d261bcf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 85, + 214, + 194, + 296 + ], + "blocks": [ + { + "bbox": [ + 85, + 214, + 194, + 296 + ], + "lines": [ + { + "bbox": [ + 85, + 214, + 194, + 296 + ], + "spans": [ + { + "bbox": [ + 85, + 214, + 194, + 296 + ], + "type": "image", + "image_path": "987705a8866119a8a9efeaab1fe4240c181b9b24b52fe414c75ca859ab1bbf7c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 300, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 300, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 555, + 346 + ], + "type": "text", + "content": "Figure 3. The proposed Codec-Aware Enhancement Framework consists of two sub-networks: 1) the Compression-Aware Adaptation (CAA) Network, which hierarchically applies sequence adaptation and frame adaptation to dynamically adjust parameters of the enhancement network; and 2) Bitstream-Aware Enhancement (BAE) Network, which leverages motion vectors to align frames and conducts region-aware refinement to flexibly enhance regions of different complexity." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 194, + 214, + 302, + 296 + ], + "blocks": [ + { + "bbox": [ + 194, + 214, + 302, + 296 + ], + "lines": [ + { + "bbox": [ + 194, + 214, + 302, + 296 + ], + "spans": [ + { + "bbox": [ + 194, + 214, + 302, + 296 + ], + "type": "image", + "image_path": "00b816b2a20be9a9474f78fed57b3062d9b63d43cb0adfd2ac883c8d7219e61d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 214, + 398, + 296 + ], + "blocks": [ + { + "bbox": [ + 302, + 214, + 398, + 296 + ], + "lines": [ + { + "bbox": [ + 302, + 214, + 398, + 296 + ], + "spans": [ + { + "bbox": [ + 302, + 214, + 398, + 296 + ], + "type": "image", + "image_path": "4c54ae2e59ec269ad78571d3d467ee6468c19468448edf8f249ed0910aa9c41b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 399, + 215, + 522, + 296 + ], + "blocks": [ + { + "bbox": [ + 399, + 215, + 522, + 296 + ], + "lines": [ + { + "bbox": [ + 399, + 215, + 522, + 296 + ], + "spans": [ + { + "bbox": [ + 399, + 215, + 522, + 296 + ], + "type": "image", + "image_path": "516ba0b2603165777b42e9571593bd8a6b81f776ce183826e09d15478a26d939.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 357, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 296, + 477 + ], + "type": "text", + "content": "Sequence adaptation. To ensure robust performance across multiple compression settings without increasing complexity, we propose estimating sequence-adaptive parameters for the enhancement network instead of fusing features from separate submodels. As shown in Fig. 3(b), parallel expert layers " + }, + { + "bbox": [ + 54, + 357, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\{f_{\\theta_1}, f_{\\theta_2}, \\dots, f_{\\theta_N}\\}" + }, + { + "bbox": [ + 54, + 357, + 296, + 477 + ], + "type": "text", + "content": ", which share the same architecture but have independent parameters, serve as the basis for parameter combination. The sequence-wise " + }, + { + "bbox": [ + 54, + 357, + 296, + 477 + ], + "type": "inline_equation", + "content": "CRF_s" + }, + { + "bbox": [ + 54, + 357, + 296, + 477 + ], + "type": "text", + "content": " is adopted as the condition to re-weight parameters of these expert layers, which is expressed as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 477, + 295, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 295, + 510 + ], + "type": "interline_equation", + "content": "f _ {\\theta_ {s}} = \\mathcal {G} _ {\\phi_ {s}} \\left(C R F _ {s}, \\left\\{f _ {\\theta_ {1}}, f _ {\\theta_ {2}}, \\dots , f _ {\\theta_ {N}} \\right\\}\\right) = \\sum_ {n = 1} ^ {N} w _ {n} f _ {\\theta_ {n}}, \\tag {1}", + "image_path": "360f661f08f0cd9fcbbb2b412c0d7b945c83f9e4bfb721ad068e8bb2b2c0d10e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi_s}" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " denote the sequence-adaptive expert layer and the sequence-wise weight generator, respectively. " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "w_n" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " denotes the weight for each expert layer. We set " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "N = 6" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " (see ablation studies in supplementary materials) and visualize " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "w_n" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " against different " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "CRF_s" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " in Fig. 4, which shows that each expert layer has a distinct preference for specific " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "CRF_s" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": ". Compared to MoE that re-weights output features, re-weighting expert layer parameters (as shown in Eq. 1) is computationally efficient and comparable to the network constructed with a single expert layer. Note that " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "CRF_s" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " is constant for frames within the same sequence, " + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 54, + 510, + 295, + 653 + ], + "type": "text", + "content": " is predicted only once and reused by subsequent frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "text", + "content": "Frame adaptation. To flexibly enhance frames with different visual quality, we propose to re-weight the sequence-based " + }, + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "text", + "content": " using frame-wise " + }, + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "text", + "content": ". We attribute the quality fluctuation between the sequence and current frame to the disparity between " + }, + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "CRF_{s}" + }, + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 54, + 653, + 296, + 714 + ], + "type": "text", + "content": ", which can be ad" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 343, + 354, + 525, + 464 + ], + "blocks": [ + { + "bbox": [ + 343, + 354, + 525, + 464 + ], + "lines": [ + { + "bbox": [ + 343, + 354, + 525, + 464 + ], + "spans": [ + { + "bbox": [ + 343, + 354, + 525, + 464 + ], + "type": "image", + "image_path": "f66511060d4f358a95dc2fd332451b6b6dfafb423f2343ed4dd0663d8f934480.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 467, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 489 + ], + "type": "text", + "content": "Figure 4. Visualization of " + }, + { + "bbox": [ + 313, + 467, + 555, + 489 + ], + "type": "inline_equation", + "content": "w_{n}" + }, + { + "bbox": [ + 313, + 467, + 555, + 489 + ], + "type": "text", + "content": " against different " + }, + { + "bbox": [ + 313, + 467, + 555, + 489 + ], + "type": "inline_equation", + "content": "CRF_{s}" + }, + { + "bbox": [ + 313, + 467, + 555, + 489 + ], + "type": "text", + "content": ", where each expert shows a distinct preference for specific " + }, + { + "bbox": [ + 313, + 467, + 555, + 489 + ], + "type": "inline_equation", + "content": "CRF_{s}" + }, + { + "bbox": [ + 313, + 467, + 555, + 489 + ], + "type": "text", + "content": " values." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "text", + "content": "dressed by introducing a set of frame-wise auxiliary parameters " + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "inline_equation", + "content": "\\triangle \\theta_{i}" + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "text", + "content": ". As shown in Fig. 3(c), the auxiliary parameters " + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "inline_equation", + "content": "\\triangle \\theta_{i}" + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "text", + "content": " that conditioned on " + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "text", + "content": " re-weights the sequence-adaptive " + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "text", + "content": " to obtain the frame-adaptive expert layer " + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "inline_equation", + "content": "f_{\\theta_i}" + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "text", + "content": " which is expressed as follows," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 357, + 565, + 553, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 565, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 357, + 565, + 553, + 578 + ], + "type": "interline_equation", + "content": "f _ {\\theta_ {i}} = \\mathcal {G} _ {\\phi_ {i}} \\left(C R F _ {i}, f _ {\\theta_ {s}}\\right) = f _ {\\theta_ {s} + \\triangle \\theta_ {i}}, \\tag {2}", + "image_path": "43f2576841a985249308a5a92a67fcfd5369b6a7ab7ecff2acecc508c36ea733.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "inline_equation", + "content": "f_{\\theta_i}" + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi_i}" + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "content": " denote the estimated frame-adaptive expert layer and the frame-wise parameters generator, respectively. As shown by the black dashed lines in Fig. 3(a), the obtained " + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "inline_equation", + "content": "f_{\\theta_i}" + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "content": " is used to construct the enhancement blocks, resulting in the frame-adaptive BAE network " + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta_i}" + }, + { + "bbox": [ + 313, + 581, + 555, + 653 + ], + "type": "text", + "content": " (introduced in the following Sec. 4.3)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 660, + 528, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 528, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 528, + 672 + ], + "type": "text", + "content": "4.3. Bitstream-Aware Enhancement Network" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "To leverage high-quality frames and propagate information, the BAE network " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta_i}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " utilizes motion vectors to align reference frames with the current frame. Meanwhile, the par" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17770" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 141, + 110 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 141, + 110 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 141, + 110 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 141, + 110 + ], + "type": "image", + "image_path": "cbe3707dba8b057c2a02dd066680850611f631e480f4b6fb40311aa443d64243.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 110, + 132, + 118 + ], + "lines": [ + { + "bbox": [ + 84, + 110, + 132, + 118 + ], + "spans": [ + { + "bbox": [ + 84, + 110, + 132, + 118 + ], + "type": "text", + "content": "Reference frame" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 142, + 70, + 210, + 110 + ], + "blocks": [ + { + "bbox": [ + 142, + 70, + 210, + 110 + ], + "lines": [ + { + "bbox": [ + 142, + 70, + 210, + 110 + ], + "spans": [ + { + "bbox": [ + 142, + 70, + 210, + 110 + ], + "type": "image", + "image_path": "85bc66cfbd07932a6db7a2b93d1b18030ec275ca44ae7b88ab56f39599426598.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 110, + 181, + 119 + ], + "lines": [ + { + "bbox": [ + 172, + 110, + 181, + 119 + ], + "spans": [ + { + "bbox": [ + 172, + 110, + 181, + 119 + ], + "type": "inline_equation", + "content": "h_i" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 211, + 70, + 278, + 110 + ], + "blocks": [ + { + "bbox": [ + 211, + 70, + 278, + 110 + ], + "lines": [ + { + "bbox": [ + 211, + 70, + 278, + 110 + ], + "spans": [ + { + "bbox": [ + 211, + 70, + 278, + 110 + ], + "type": "image", + "image_path": "bfe3f50b910ec0825c4f674b67ba625663a0fbadd4f694562831e3dae7da534f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 110, + 249, + 118 + ], + "lines": [ + { + "bbox": [ + 240, + 110, + 249, + 118 + ], + "spans": [ + { + "bbox": [ + 240, + 110, + 249, + 118 + ], + "type": "inline_equation", + "content": "\\hat{h}_i" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 74, + 120, + 141, + 159 + ], + "blocks": [ + { + "bbox": [ + 74, + 120, + 141, + 159 + ], + "lines": [ + { + "bbox": [ + 74, + 120, + 141, + 159 + ], + "spans": [ + { + "bbox": [ + 74, + 120, + 141, + 159 + ], + "type": "image", + "image_path": "fd9725b6d195a16ca67476ec4e1ee3128230dcf1e662db6450894c9486b3394b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 159, + 126, + 169 + ], + "lines": [ + { + "bbox": [ + 86, + 159, + 126, + 169 + ], + "spans": [ + { + "bbox": [ + 86, + 159, + 126, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{S}(M_i^{8\\times 8},h_i)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 173, + 296, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 296, + 218 + ], + "type": "text", + "content": "Figure 5. Visualization of features in region-aware refinement, where " + }, + { + "bbox": [ + 55, + 173, + 296, + 218 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 55, + 173, + 296, + 218 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 173, + 296, + 218 + ], + "type": "inline_equation", + "content": "\\hat{h}_i" + }, + { + "bbox": [ + 55, + 173, + 296, + 218 + ], + "type": "text", + "content": " indicate the input and output features, respectively. The refined features are denoted in the format of " + }, + { + "bbox": [ + 55, + 173, + 296, + 218 + ], + "type": "inline_equation", + "content": "S(M_i^{type}, h_i)" + }, + { + "bbox": [ + 55, + 173, + 296, + 218 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 141, + 120, + 210, + 159 + ], + "blocks": [ + { + "bbox": [ + 141, + 120, + 210, + 159 + ], + "lines": [ + { + "bbox": [ + 141, + 120, + 210, + 159 + ], + "spans": [ + { + "bbox": [ + 141, + 120, + 210, + 159 + ], + "type": "image", + "image_path": "26d2d57be50facdd724be3d6ea4582d2844fd15d73191db46b14de09f72fff73.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 159, + 197, + 170 + ], + "lines": [ + { + "bbox": [ + 154, + 159, + 197, + 170 + ], + "spans": [ + { + "bbox": [ + 154, + 159, + 197, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{S}(M_i^{16\\times 16},h_i)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 211, + 120, + 278, + 159 + ], + "blocks": [ + { + "bbox": [ + 211, + 120, + 278, + 159 + ], + "lines": [ + { + "bbox": [ + 211, + 120, + 278, + 159 + ], + "spans": [ + { + "bbox": [ + 211, + 120, + 278, + 159 + ], + "type": "image", + "image_path": "d10ecead72505980800ca33277dc660618d1818ccfd7a3d2f9d3ddb02a8ae7a3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 159, + 264, + 170 + ], + "lines": [ + { + "bbox": [ + 222, + 159, + 264, + 170 + ], + "spans": [ + { + "bbox": [ + 222, + 159, + 264, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{S}(M_i^{16\\times 8},h_i)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 228, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 252 + ], + "type": "text", + "content": "tition map serves as spatial complexity guidance to enable flexible enhancement of different regions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 253, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 348 + ], + "type": "text", + "content": "Motion vector alignment. Since the motion vectors roughly model the temporal relationship in a block-wise manner, for each block of the current frame shown in Fig. 3(d), motion vectors locate blocks with similar content in the previous and future reference features (highlighted with red and blue boxes). The warped reference features are concatenated with current frame along the channel dimension as input of the BAE network, expressed as follows," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 111, + 350, + 294, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 350, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 111, + 350, + 294, + 365 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {i} = \\left[ M V \\left(h _ {i} ^ {p}\\right), M V \\left(h _ {i} ^ {f}\\right), x _ {i} \\right], \\tag {3}", + "image_path": "5fd8debb58f66616f525aecbbcbba3685f9fbd2c92aa0011f4f9820b2efd22c0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "inline_equation", + "content": "\\hat{x}_i" + }, + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "inline_equation", + "content": "h_i^p" + }, + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "inline_equation", + "content": "h_i^f" + }, + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "text", + "content": " denote the current input frame, enhanced features of previous and future reference frames, respectively. " + }, + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "inline_equation", + "content": "MV" + }, + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "text", + "content": " denotes warping reference features based on motion vectors. " + }, + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "inline_equation", + "content": "[,]" + }, + { + "bbox": [ + 55, + 368, + 296, + 440 + ], + "type": "text", + "content": " denotes concatenation along channel dimension. Bilinear interpolation is adopted for the case that the offsets of motion vectors are not integers." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 440, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 296, + 559 + ], + "type": "text", + "content": "Region-aware refinement. To effectively enhance regions of different complexity, we propose to dynamically assign different filters for regions based on the partition map. As shown in Fig. 3(e), the block-based partition map " + }, + { + "bbox": [ + 55, + 440, + 296, + 559 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 55, + 440, + 296, + 559 + ], + "type": "text", + "content": " is decoupled into multiple binary masks " + }, + { + "bbox": [ + 55, + 440, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\{M_i^{8\\times 8}, M_i^{16\\times 16}, \\dots, M_i^{type}\\}" + }, + { + "bbox": [ + 55, + 440, + 296, + 559 + ], + "type": "text", + "content": " according to the size of macroblocks, allowing separate refinement of regions using sparse convolution [33]. The output is defined as the sum of frame-adaptive extracted features and separately region-aware refined features, depicted as follows," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 63, + 559, + 294, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 559, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 63, + 559, + 294, + 591 + ], + "type": "interline_equation", + "content": "\\hat {h} _ {i} = \\mathcal {F} _ {\\theta_ {i}} \\left(\\hat {x} _ {i}, P _ {i}\\right) = f _ {\\theta_ {i}} * h _ {i} + \\sum_ {t y p e = 1} ^ {M} \\mathcal {S} \\left(M _ {i} ^ {t y p e}, h _ {i}\\right), \\tag {4}", + "image_path": "30d909002fafccda5b00dcd78f0e561c85e141d4640655e1d31891832af2920a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{h}_i" + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": " indicates the output features. " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": " denotes the operations applying sparse convolution guided by mask " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_i^{type}" + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": " to refine input features. In H.264 standard, three types of macroblocks are used " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "(16\\times 16,8\\times 16 / 16\\times 8," + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "8\\times 8)" + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": ", thus " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": " is set to 3. Features in region-aware refinement are visualized in Fig 5, where the refined features are denoted with operations like " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "S(M_{i}^{8\\times 8},h_{i})" + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": ". As can be seen, the output features " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{h}_i" + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": " contain more fine-grained and high-frequency details than " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "text", + "content": ". Meanwhile, refined features provide distinct activations for different regions. For instance, " + }, + { + "bbox": [ + 55, + 592, + 296, + 714 + ], + "type": "inline_equation", + "content": "S(M_i^{8\\times 8},h_i)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 71, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 96 + ], + "type": "text", + "content": "focuses on static objects (e.g., trees) while " + }, + { + "bbox": [ + 313, + 71, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{S}(M_i^{8\\times 16},h_i)" + }, + { + "bbox": [ + 313, + 71, + 553, + 96 + ], + "type": "text", + "content": " focuses on moving objects (e.g., the bus)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 102, + 400, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 102, + 400, + 113 + ], + "spans": [ + { + "bbox": [ + 314, + 102, + 400, + 113 + ], + "type": "text", + "content": "4.4. Loss function" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 117, + 554, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 554, + 163 + ], + "type": "text", + "content": "We adopt Charbonnier penalty loss [9] as the loss function and train the proposed codec-aware enhancement framework in an end-to-end manner. The specific loss function is expressed as follows," + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 370, + 166, + 553, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 166, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 370, + 166, + 553, + 198 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\sqrt {\\left\\| y _ {i} - \\hat {y} _ {i} \\right\\| ^ {2} + \\epsilon^ {2}}, \\tag {5}", + "image_path": "be9b10c8770f8d3c8f85db102e9257d1599192bf6fe80c83a429cccdbd7504bb.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 201, + 554, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 554, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 201, + 554, + 237 + ], + "type": "inline_equation", + "content": "y_{i}, \\hat{y}_{i}" + }, + { + "bbox": [ + 313, + 201, + 554, + 237 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 201, + 554, + 237 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 201, + 554, + 237 + ], + "type": "text", + "content": " indicate the uncompressed ground truth, the predicted output, and the length of the input sequence. " + }, + { + "bbox": [ + 313, + 201, + 554, + 237 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 201, + 554, + 237 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 313, + 201, + 554, + 237 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-12}" + }, + { + "bbox": [ + 313, + 201, + 554, + 237 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 243, + 395, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 243, + 395, + 255 + ], + "spans": [ + { + "bbox": [ + 314, + 243, + 395, + 255 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 259, + 441, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 259, + 441, + 272 + ], + "spans": [ + { + "bbox": [ + 314, + 259, + 441, + 272 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 275, + 554, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 554, + 334 + ], + "type": "text", + "content": "Compression settings. H.264 is a popular video compression standard that compresses nearly " + }, + { + "bbox": [ + 313, + 275, + 554, + 334 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 275, + 554, + 334 + ], + "type": "text", + "content": " of internet videos [1], and tends to introduce more severe degradations than H.265 and H.266. We adopt H.264 [45] and compress videos with the " + }, + { + "bbox": [ + 313, + 275, + 554, + 334 + ], + "type": "inline_equation", + "content": "CRF_{s}" + }, + { + "bbox": [ + 313, + 275, + 554, + 334 + ], + "type": "text", + "content": " values of 15, 25, and 35." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 418 + ], + "type": "text", + "content": "Tasks and training dataset. Our tasks involve quality enhancement and assisting downstream tasks on compressed inputs. The primary downstream tasks include video superresolution, optical flow estimation and video object segmentation, with video inpainting reported as an extension to fully evaluate the versatility. Training splits of REDS [39] and DAVIS [41] datasets are combined for training." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 514 + ], + "type": "text", + "content": "Compared methods. We compare with representative methods in compressed video enhancement, including MFQE 2.0 [22], STDF [16], S2SVR [32] and Metabit [18]. For a fair comparison, we fully retrain these methods with the same training dataset and configurations. For downstream tasks, the compressed video is first enhanced by quality enhancement methods, and then fed to downstream models for corresponding tasks and further assessment." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 521, + 372, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 521, + 372, + 533 + ], + "spans": [ + { + "bbox": [ + 314, + 521, + 372, + 533 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 536, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 586 + ], + "type": "text", + "content": "Our evaluations are two-fold: 1) verifying the quality enhancement performance on seen, unseen, and highly compressed scenarios; 2) evaluating the versatility to assist different downstream tasks on multiple compression settings." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 591, + 493, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 493, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 493, + 602 + ], + "type": "text", + "content": "5.2.1. Quality Enhancement Performance" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": "Quantitative results. The results of compressed video quality enhancement are reported in Tab.1, which is evaluated on the REDS4 datasets [39] using PSNR and SSIM (the higher the better). Note the CRF values of 18, 28 and 38 are not included during training. For each method, we include the model complexity and inference speed. For our method, we report results of both applying " + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": " and its substitution with slice type (highlighted with grey). As shown in Tab. 1, leveraging slice type yields comparable performance" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17771" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 64, + 553, + 157 + ], + "blocks": [ + { + "bbox": [ + 58, + 64, + 553, + 157 + ], + "lines": [ + { + "bbox": [ + 58, + 64, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 58, + 64, + 553, + 157 + ], + "type": "table", + "html": "
MethodParam/MFLOPs/GSpeed/msFPSCRF15CRF25CRF35CRF18CRF28CRF38
PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑PSNR↑/SSIM↑
Input----41.04 / 0.978534.92 / 0.936329.25 / 0.823839.12 / 0.969833.18 / 0.912327.69 / 0.7725
MFQE 2.0 [22]1.6451531940.95 / 0.980634.83 / 0.937829.22 / 0.825638.97 / 0.971233.13 / 0.914027.67 / 0.7742
STDF [16]1.2745382641.15 / 0.979335.23 / 0.939829.74 / 0.835939.28 / 0.971233.58 / 0.917828.11 / 0.7853
S2SVR [32]7.43294--41.96 / 0.983435.61 / 0.944529.87 / 0.839139.88 / 0.975533.87 / 0.922328.19 / 0.7881
Metabit [18]1.6092244241.04 / 0.978534.92 / 0.936329.25 / 0.823839.11 / 0.969833.18 / 0.912327.69 / 0.7725
Ours4.5647362842.22 / 0.984235.90 / 0.946830.17 / 0.847140.17 / 0.976734.16 / 0.925828.49 / 0.7985
42.24 / 0.984235.91 / 0.946830.19 / 0.847240.18 / 0.976734.17 / 0.925828.52 / 0.7985
", + "image_path": "d1ff7f80808d6f80ef7741a476dbe1b0fadf4f6fc15a8b008081f56fe7523e09.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 186, + 203, + 287 + ], + "blocks": [ + { + "bbox": [ + 58, + 186, + 203, + 287 + ], + "lines": [ + { + "bbox": [ + 58, + 186, + 203, + 287 + ], + "spans": [ + { + "bbox": [ + 58, + 186, + 203, + 287 + ], + "type": "image", + "image_path": "1381f0b495b2f23e1d4c044cbbd8c659778447964fe6f018c4888d88f39a40d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 289, + 178, + 299 + ], + "lines": [ + { + "bbox": [ + 82, + 289, + 178, + 299 + ], + "spans": [ + { + "bbox": [ + 82, + 289, + 178, + 299 + ], + "type": "text", + "content": "Frame 030,Clip 020,CRF35" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 204, + 186, + 253, + 236 + ], + "blocks": [ + { + "bbox": [ + 204, + 186, + 253, + 236 + ], + "lines": [ + { + "bbox": [ + 204, + 186, + 253, + 236 + ], + "spans": [ + { + "bbox": [ + 204, + 186, + 253, + 236 + ], + "type": "image", + "image_path": "81cce0bd7a8cb110d3f4a7e0bb49145379db7bc7a21652676799e0495b2bf111.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 204, + 237, + 253, + 286 + ], + "blocks": [ + { + "bbox": [ + 204, + 237, + 253, + 286 + ], + "lines": [ + { + "bbox": [ + 204, + 237, + 253, + 286 + ], + "spans": [ + { + "bbox": [ + 204, + 237, + 253, + 286 + ], + "type": "image", + "image_path": "c8a72818269e3cf40bff5988f1aebd5f68d5abe872bb04d02da8d7b9ac5e091b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 289, + 238, + 299 + ], + "lines": [ + { + "bbox": [ + 219, + 289, + 238, + 299 + ], + "spans": [ + { + "bbox": [ + 219, + 289, + 238, + 299 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 303, + 553, + 325 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 553, + 325 + ], + "type": "text", + "content": "Figure 6. Qualitative results on quality enhancement, where our method effectively reduces the compression artifacts, achieving visually pleasant results. In contrast, the results of the compared methods still contain severe distortions (e.g., the calf in the 1st row)." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 253, + 186, + 303, + 236 + ], + "blocks": [ + { + "bbox": [ + 253, + 186, + 303, + 236 + ], + "lines": [ + { + "bbox": [ + 253, + 186, + 303, + 236 + ], + "spans": [ + { + "bbox": [ + 253, + 186, + 303, + 236 + ], + "type": "image", + "image_path": "af260206614883bcc887d2f9a294dc6adfea2bfc66bb90fee07b277fea3d1575.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 253, + 237, + 303, + 286 + ], + "blocks": [ + { + "bbox": [ + 253, + 237, + 303, + 286 + ], + "lines": [ + { + "bbox": [ + 253, + 237, + 303, + 286 + ], + "spans": [ + { + "bbox": [ + 253, + 237, + 303, + 286 + ], + "type": "image", + "image_path": "4a7a719c8024b9de8a3b047dc00c4a5199f523a97615ee7439b839a5c4ecb88e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 289, + 296, + 298 + ], + "lines": [ + { + "bbox": [ + 261, + 289, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 261, + 289, + 296, + 298 + ], + "type": "text", + "content": "MFQE 2.0" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 304, + 186, + 353, + 236 + ], + "blocks": [ + { + "bbox": [ + 304, + 186, + 353, + 236 + ], + "lines": [ + { + "bbox": [ + 304, + 186, + 353, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 353, + 236 + ], + "type": "image", + "image_path": "f12387a4674e6c115e1a8115c794c432337218e58e933bc6c71ca7d518e988db.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 304, + 237, + 353, + 286 + ], + "blocks": [ + { + "bbox": [ + 304, + 237, + 353, + 286 + ], + "lines": [ + { + "bbox": [ + 304, + 237, + 353, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 353, + 286 + ], + "type": "image", + "image_path": "d35ee88ccbfd149b7ef0582c3c0637ff0815c1ba910b4e42ce8956ca2754c463.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 289, + 339, + 298 + ], + "lines": [ + { + "bbox": [ + 318, + 289, + 339, + 298 + ], + "spans": [ + { + "bbox": [ + 318, + 289, + 339, + 298 + ], + "type": "text", + "content": "STDF" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 354, + 186, + 403, + 236 + ], + "blocks": [ + { + "bbox": [ + 354, + 186, + 403, + 236 + ], + "lines": [ + { + "bbox": [ + 354, + 186, + 403, + 236 + ], + "spans": [ + { + "bbox": [ + 354, + 186, + 403, + 236 + ], + "type": "image", + "image_path": "6ffadf35ac7f87219e87c799c72bb3e653e041741001322d8b5af50309e5e588.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 354, + 237, + 403, + 286 + ], + "blocks": [ + { + "bbox": [ + 354, + 237, + 403, + 286 + ], + "lines": [ + { + "bbox": [ + 354, + 237, + 403, + 286 + ], + "spans": [ + { + "bbox": [ + 354, + 237, + 403, + 286 + ], + "type": "image", + "image_path": "031ba3b1969564e735cbaffecb85a5c9deb642f3ca62f49f15a608e3a89af3a5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 289, + 391, + 298 + ], + "lines": [ + { + "bbox": [ + 365, + 289, + 391, + 298 + ], + "spans": [ + { + "bbox": [ + 365, + 289, + 391, + 298 + ], + "type": "text", + "content": "S2SVR" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 404, + 186, + 452, + 236 + ], + "blocks": [ + { + "bbox": [ + 404, + 186, + 452, + 236 + ], + "lines": [ + { + "bbox": [ + 404, + 186, + 452, + 236 + ], + "spans": [ + { + "bbox": [ + 404, + 186, + 452, + 236 + ], + "type": "image", + "image_path": "63cd79b7913622844f8d955d831212b87f77fe12dfc9b8821d8439f02297a927.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 404, + 237, + 452, + 286 + ], + "blocks": [ + { + "bbox": [ + 404, + 237, + 452, + 286 + ], + "lines": [ + { + "bbox": [ + 404, + 237, + 452, + 286 + ], + "spans": [ + { + "bbox": [ + 404, + 237, + 452, + 286 + ], + "type": "image", + "image_path": "facfdbbdb0410ffbf5d756bc64a1bec304b11b26421c3dd97366f016bc607f3d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 289, + 441, + 298 + ], + "lines": [ + { + "bbox": [ + 415, + 289, + 441, + 298 + ], + "spans": [ + { + "bbox": [ + 415, + 289, + 441, + 298 + ], + "type": "text", + "content": "Metabit" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 454, + 186, + 503, + 236 + ], + "blocks": [ + { + "bbox": [ + 454, + 186, + 503, + 236 + ], + "lines": [ + { + "bbox": [ + 454, + 186, + 503, + 236 + ], + "spans": [ + { + "bbox": [ + 454, + 186, + 503, + 236 + ], + "type": "image", + "image_path": "1310b3cc38a7043f39276b23b5968bfd1ae9e29dd81a1cc087377f6b7698b771.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 454, + 237, + 503, + 286 + ], + "blocks": [ + { + "bbox": [ + 454, + 237, + 503, + 286 + ], + "lines": [ + { + "bbox": [ + 454, + 237, + 503, + 286 + ], + "spans": [ + { + "bbox": [ + 454, + 237, + 503, + 286 + ], + "type": "image", + "image_path": "f267dba5ede5c5e0eb24f37454f5eda60de9267119b2b0168ff39cd0329ca258.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 289, + 487, + 298 + ], + "lines": [ + { + "bbox": [ + 470, + 289, + 487, + 298 + ], + "spans": [ + { + "bbox": [ + 470, + 289, + 487, + 298 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 503, + 186, + 553, + 236 + ], + "blocks": [ + { + "bbox": [ + 503, + 186, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 503, + 186, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 503, + 186, + 553, + 236 + ], + "type": "image", + "image_path": "c1b02fbfb2b15f26231ddf1b81ca81992646123e1ffff490c021ba64bdcacbeb.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 503, + 237, + 553, + 286 + ], + "blocks": [ + { + "bbox": [ + 503, + 237, + 553, + 286 + ], + "lines": [ + { + "bbox": [ + 503, + 237, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 503, + 237, + 553, + 286 + ], + "type": "image", + "image_path": "30e32489a6322e42fa8c8810cef6cfe3d551e67d9300d3b1e1e35c678d3684ae.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 523, + 289, + 534, + 297 + ], + "lines": [ + { + "bbox": [ + 523, + 289, + 534, + 297 + ], + "spans": [ + { + "bbox": [ + 523, + 289, + 534, + 297 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 70, + 335, + 281, + 533 + ], + "blocks": [ + { + "bbox": [ + 55, + 159, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 553, + 182 + ], + "type": "text", + "content": "Table 1. Quantitative results on quality enhancement, where PSNR and SSIM (higher is better) are adopted for evaluation. The best and second best results are marked with bold and underline. Results obtained by replacing " + }, + { + "bbox": [ + 55, + 159, + 553, + 182 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 55, + 159, + 553, + 182 + ], + "type": "text", + "content": " with slice type are highlighted with grey." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 335, + 281, + 533 + ], + "lines": [ + { + "bbox": [ + 70, + 335, + 281, + 533 + ], + "spans": [ + { + "bbox": [ + 70, + 335, + 281, + 533 + ], + "type": "table", + "html": "
MethodCRF15CRF25CRF35
PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
BasicVSR29.24 / 0.821226.19 / 0.713123.40 / 0.6005
+ MFQE 2.029.29 / 0.823326.28 / 0.718223.46 / 0.6056
+ STDF29.31 / 0.824726.51 / 0.729323.80 / 0.6249
+ S2SVR29.45 / 0.828826.70 / 0.734623.86 / 0.6270
+ Metabit29.24 / 0.821126.19 / 0.713123.39 / 0.6005
+ Ours29.54 / 0.832826.85 / 0.741924.02 / 0.6361
IconVSR29.29 / 0.823026.19 / 0.713023.39 / 0.6003
+ MFQE 2.029.37 / 0.825426.28 / 0.718223.45 / 0.6055
+ STDF29.36 / 0.826326.52 / 0.729223.79 / 0.6248
+ S2SVR29.54 / 0.830626.71 / 0.734523.85 / 0.6269
+ Metabit29.29 / 0.823026.19 / 0.713023.39 / 0.6003
+ Ours29.63 / 0.834426.86 / 0.741824.01 / 0.6360
BasicVSR++29.61 / 0.830326.19 / 0.711823.38 / 0.5998
+ MFQE 2.029.66 / 0.832226.27 / 0.716923.44 / 0.6051
+ STDF29.68 / 0.833826.53 / 0.728923.79 / 0.6247
+ S2SVR29.82 / 0.837126.72 / 0.734623.85 / 0.6269
+ Metabit29.61 / 0.830326.19 / 0.711823.38 / 0.5997
+ Ours29.92 / 0.840726.87 / 0.741924.00 / 0.6358
", + "image_path": "a72b3d58352e74146b0a3319cd0fbd9fcc221601d40b259f73b5caaaf0f9969b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 536, + 295, + 558 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 558 + ], + "type": "text", + "content": "Table 2. Quantitative results of " + }, + { + "bbox": [ + 55, + 536, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 536, + 295, + 558 + ], + "type": "text", + "content": " VSR,where the best and second best results are highlighted with bold and underline." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", with a negligible decrease of PSNR (" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "< 0.03\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "), demonstrating the feasibility of replacing " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "CRF_{i}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " with slice type in practical. The proposed method notably improves the quality of compressed input, achieving a PSNR gain of " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "1.2\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " on CRF15, while MFQE 2.0 and Metabit lead to no improvement. With similar computation cost and inference speed, our method significantly outperforms STDF, obtaining a PSNR gain of " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "1.09\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " on CRF15. Compared to S2SVR, our approach takes only " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "61\\%" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " of the parameters and " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " of the FLOPs, and achieves a throughput of 28 FPS, which underlines its practicality. In addition, our method shows robustness and generalization ability on unseen sce" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 337, + 555, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 432 + ], + "type": "text", + "content": "narios (i.e., CRF18, CRF28 and CRF38), achieving up to 1.06 dB PSNR gain on CRF18. In contrast, other methods trained with mixed compression settings show sub-optimal performance. For instance, STDF and S2SVR only achieve PSNR gains of 0.16 dB and 0.76 dB at CRF18, while MFQE 2.0 shows no improvement. Quantitative results on highly compressed scenarios (i.e., CRF40, CRF45, CRF48) are included in the supplementary materials." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 433, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 556, + 529 + ], + "type": "text", + "content": "Qualitative results. Qualitative comparisons are provided in Fig. 6. As can be seen, MFQE 2.0 and Metabit struggle to improve the quality of compressed inputs. Both STDF and S2SVR cannot adequately remove compression artifacts (e.g., boundary of the calf), while the proposed method effectively eliminates the compression artifacts, preserving accurate edges and textures. We provide more qualitative comparisons in the supplementary materials." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 532, + 436, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 436, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 436, + 544 + ], + "type": "text", + "content": "5.2.2. Versatility Evaluation" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 545, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 605 + ], + "type": "text", + "content": "To evaluate the versatility in assisting practical downstream tasks, we employ the implementation that utilizes slice type for frame adaptation (as described in Sec. 3.1) to enhance compressed inputs for downstream tasks. More qualitative comparisons are included in the supplementary materials." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "content": "Video super-resolution. We adopt BasicVSR [6], IconVSR [6], and BasicVSR++ [7] as the representative baseline methods for " + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "content": " video super-resolution (VSR), which are trained on \"clean\" data without considering compression. The evaluation is conducted on the REDS4 dataset [39] and summarized with PSNR and SSIM (the higher the better). As depicted in Tab. 2, pre-enhancing compressed inputs with Metabit fails to improve the performance of downstream VSR models. In contrast, pre-enhancing with our" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17772" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 64, + 203, + 177 + ], + "blocks": [ + { + "bbox": [ + 57, + 64, + 203, + 177 + ], + "lines": [ + { + "bbox": [ + 57, + 64, + 203, + 177 + ], + "spans": [ + { + "bbox": [ + 57, + 64, + 203, + 177 + ], + "type": "image", + "image_path": "f5c6eb61718524fc7c5a8342f81d22f501d9bf5796f3c8c1cbaadf6828f79f56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 181, + 553, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 181, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 553, + 204 + ], + "type": "text", + "content": "Figure 7. Qualitative results of " + }, + { + "bbox": [ + 55, + 181, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 181, + 553, + 204 + ], + "type": "text", + "content": " VSR. Pre-enhancing with our method before VSR effectively avoids amplifying compression artifacts. While other methods cannot fully eliminate the artifacts and even severe the distortions (e.g., results of MFQE 2.0)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 203, + 64, + 253, + 176 + ], + "blocks": [ + { + "bbox": [ + 203, + 64, + 253, + 176 + ], + "lines": [ + { + "bbox": [ + 203, + 64, + 253, + 176 + ], + "spans": [ + { + "bbox": [ + 203, + 64, + 253, + 176 + ], + "type": "image", + "image_path": "4e2d88e0c6875def96096d6ffaf1d66c175b413f59da668e28aca2a8764ae853.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 254, + 64, + 303, + 176 + ], + "blocks": [ + { + "bbox": [ + 254, + 64, + 303, + 176 + ], + "lines": [ + { + "bbox": [ + 254, + 64, + 303, + 176 + ], + "spans": [ + { + "bbox": [ + 254, + 64, + 303, + 176 + ], + "type": "image", + "image_path": "8c1e179d8292670f767ab4404940df61ec6d0182be665f7b81c267e34583176c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 304, + 64, + 353, + 176 + ], + "blocks": [ + { + "bbox": [ + 304, + 64, + 353, + 176 + ], + "lines": [ + { + "bbox": [ + 304, + 64, + 353, + 176 + ], + "spans": [ + { + "bbox": [ + 304, + 64, + 353, + 176 + ], + "type": "image", + "image_path": "993962136e536e6f824225b3ff342976b971ea0d12452bd0a6fd75eb0b966e93.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 353, + 64, + 403, + 176 + ], + "blocks": [ + { + "bbox": [ + 353, + 64, + 403, + 176 + ], + "lines": [ + { + "bbox": [ + 353, + 64, + 403, + 176 + ], + "spans": [ + { + "bbox": [ + 353, + 64, + 403, + 176 + ], + "type": "image", + "image_path": "0dfc02ac5b0a2171f47befa998a68b11d1d24a62f29fb4849c61ef4d5bfb5892.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 403, + 64, + 452, + 176 + ], + "blocks": [ + { + "bbox": [ + 403, + 64, + 452, + 176 + ], + "lines": [ + { + "bbox": [ + 403, + 64, + 452, + 176 + ], + "spans": [ + { + "bbox": [ + 403, + 64, + 452, + 176 + ], + "type": "image", + "image_path": "994835e8b8fa07e8ef81d769ccdce568f21be3c75c2cd82d8ddbedbed6aed2b7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 454, + 64, + 503, + 176 + ], + "blocks": [ + { + "bbox": [ + 454, + 64, + 503, + 176 + ], + "lines": [ + { + "bbox": [ + 454, + 64, + 503, + 176 + ], + "spans": [ + { + "bbox": [ + 454, + 64, + 503, + 176 + ], + "type": "image", + "image_path": "c0cc2618c2bb323309cbfafdc82338a25f2ac30777958036345236d75decfc9a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 504, + 64, + 553, + 176 + ], + "blocks": [ + { + "bbox": [ + 504, + 64, + 553, + 176 + ], + "lines": [ + { + "bbox": [ + 504, + 64, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 504, + 64, + 553, + 176 + ], + "type": "image", + "image_path": "b615ceab181c851d4c1b245b79a4b142b04e1872f3cbf7e3ab7b083a4e854ace.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 56, + 207, + 127, + 256 + ], + "blocks": [ + { + "bbox": [ + 56, + 207, + 127, + 256 + ], + "lines": [ + { + "bbox": [ + 56, + 207, + 127, + 256 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 127, + 256 + ], + "type": "image", + "image_path": "69ecbf99b3c4ef8b7ee01333157240fd0cbe8db13f46305369c0b618cbf669c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 258, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 553, + 281 + ], + "type": "text", + "content": "Figure 8. Qualitative results of optical flow estimation. As can be seen, pre-processing with our method effectively corrects the mispredicted optical flow, especially the boundaries of moving objects (e.g., edge of the moving car)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 129, + 207, + 199, + 256 + ], + "blocks": [ + { + "bbox": [ + 129, + 207, + 199, + 256 + ], + "lines": [ + { + "bbox": [ + 129, + 207, + 199, + 256 + ], + "spans": [ + { + "bbox": [ + 129, + 207, + 199, + 256 + ], + "type": "image", + "image_path": "fa82bce0fdd460fcbcac107fd8d28b3f77928da1a28ff658478c587b9fdb1610.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 200, + 207, + 269, + 256 + ], + "blocks": [ + { + "bbox": [ + 200, + 207, + 269, + 256 + ], + "lines": [ + { + "bbox": [ + 200, + 207, + 269, + 256 + ], + "spans": [ + { + "bbox": [ + 200, + 207, + 269, + 256 + ], + "type": "image", + "image_path": "c6ec442c6680ffab03889346826d65d901f29c1ee83cd17c4340610f167a6187.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 270, + 207, + 340, + 256 + ], + "blocks": [ + { + "bbox": [ + 270, + 207, + 340, + 256 + ], + "lines": [ + { + "bbox": [ + 270, + 207, + 340, + 256 + ], + "spans": [ + { + "bbox": [ + 270, + 207, + 340, + 256 + ], + "type": "image", + "image_path": "6cbfefe86bec18a9fa67df32607ae02d0ade8d8bd505c8c57ed512914783aeb5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 341, + 207, + 411, + 256 + ], + "blocks": [ + { + "bbox": [ + 341, + 207, + 411, + 256 + ], + "lines": [ + { + "bbox": [ + 341, + 207, + 411, + 256 + ], + "spans": [ + { + "bbox": [ + 341, + 207, + 411, + 256 + ], + "type": "image", + "image_path": "c10424d9984437e85177ada66863d9de1fd0344ddc8e296beb06f8a0d0dda1dd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 411, + 207, + 482, + 256 + ], + "blocks": [ + { + "bbox": [ + 411, + 207, + 482, + 256 + ], + "lines": [ + { + "bbox": [ + 411, + 207, + 482, + 256 + ], + "spans": [ + { + "bbox": [ + 411, + 207, + 482, + 256 + ], + "type": "image", + "image_path": "1c09078aea4a8712df83c5d428948e802ecf2b6f25d4eecae23b74551efed6ee.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 483, + 207, + 553, + 256 + ], + "blocks": [ + { + "bbox": [ + 483, + 207, + 553, + 256 + ], + "lines": [ + { + "bbox": [ + 483, + 207, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 483, + 207, + 553, + 256 + ], + "type": "image", + "image_path": "6d09086fd4c04906dc99c3fe8b32c9f82a3f55a554a69a68f4582fe5db27d784.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 77, + 289, + 274, + 477 + ], + "blocks": [ + { + "bbox": [ + 77, + 289, + 274, + 477 + ], + "lines": [ + { + "bbox": [ + 77, + 289, + 274, + 477 + ], + "spans": [ + { + "bbox": [ + 77, + 289, + 274, + 477 + ], + "type": "table", + "html": "
MethodCRF15CRF25CRF35
EPE↓/ F1-all↓EPE↓/ F1-all↓EPE↓/ F1-all↓
RAFT5.26 / 17.817.37 / 22.1316.73 / 44.70
+ MFQE 2.05.32 / 17.837.27 / 21.9216.68 / 44.74
+ STDF5.34 / 17.937.13 / 22.1615.92 / 44.04
+ S2SVR5.22 / 17.716.90 / 21.5715.73 / 44.62
+ Metabit5.32 / 17.867.33 / 22.0716.69 / 44.28
+ Ours5.20 / 17.696.52 / 20.9914.84 / 42.56
DEQ3.99 / 13.715.40 / 17.3313.94 / 41.63
+ MFQE 2.03.97 / 13.735.14 / 17.0614.06 / 41.72
+ STDF4.08 / 13.845.27 / 17.3813.52 / 40.75
+ S2SVR4.01 / 13.695.22 / 16.9913.74 / 41.65
+ Metabit3.92 / 13.785.30 / 17.3313.84 / 41.21
+ Ours3.97 / 13.684.96 / 16.5413.09 / 39.43
KPAFlow4.46 / 16.076.71 / 20.9616.50 / 45.13
+ MFQE 2.04.42 / 15.966.71 / 20.9216.70 / 45.29
+ STDF4.52 / 16.196.96 / 21.5315.62 / 44.17
+ S2SVR4.37 / 15.966.10 / 19.7615.62 / 44.23
+ Metabit4.47 / 16.176.75 / 21.0816.68 / 44.90
+ Ours4.43 / 16.105.59 / 18.7314.83 / 41.24
", + "image_path": "d0e005c05014d4564ea2063fb891029d411e195d1c23d7590c2d8b83de7a853d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 509, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 652 + ], + "type": "text", + "content": "framework yields consistent improvement especially in scenarios of high compression (e.g., up to 0.62 dB PSNR gain with BaiscVSR++ on CRF35). Meanwhile, our method significantly outperforms MFQE 2.0 and STDF, achieving PSNR gains of 0.25 dB and 0.23 dB over MFQE 2.0 and STDF on BasicVSR/CRF15, respectively. Compared with S2SVR, the proposed method offers more effective support to VSR models with lower complexity. As shown in Fig. 7, performing VSR on compressed data inevitably amplifies compression artifacts (e.g., the 1st column), while results pre-enhanced with our method maintain accurate edges and textures, avoiding distortions seen in other methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "content": "Optical flow estimation. We adopt RAFT [49], DEQ [3], and KPAFlow [35] as baseline models for optical flow estimation. Evaluation on the KITTI-2015 dataset [21] is summarized with EPE (end-point-error) and F1-all loss, where lower values indicate better accuracy. As shown in Tab. 3," + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 317, + 287, + 553, + 476 + ], + "blocks": [ + { + "bbox": [ + 55, + 480, + 295, + 503 + ], + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 503 + ], + "type": "text", + "content": "Table 3. Quantitative results of optical flow estimation, where we highlight the best and second best results with **bold** and **underline**." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 287, + 553, + 476 + ], + "lines": [ + { + "bbox": [ + 317, + 287, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 287, + 553, + 476 + ], + "type": "table", + "html": "
MethodCRF15CRF25CRF35
Avg↑/J↑/F↑Avg↑/J↑/F↑Avg↑/J↑/F↑
STCN85.07 / 81.83 / 88.3284.35 / 80.96 / 87.7479.20 / 76.04 / 82.37
+ MFQE 2.084.96 / 81.71 / 88.2184.30 / 80.92 / 87.6979.28 / 76.11 / 82.44
+ STDF85.01 / 81.73 / 88.2884.23 / 80.97 / 87.5079.77 / 76.52 / 83.02
+ S2SVR85.17 / 81.93 / 88.4184.46 / 81.20 / 87.7280.04 / 76.88 / 83.20
+ Metabit84.56 / 80.97 / 88.1483.86 / 80.18 / 87.5579.03 / 75.65 / 82.40
+ Ours85.21 / 81.99 / 88.4484.63 / 81.42 / 87.8581.57 / 78.46 / 84.69
DeAoT85.90 / 82.89 / 88.9185.18 / 82.37 / 88.0082.87 / 79.86 / 85.88
+ MFQE 2.085.86 / 82.84 / 88.8885.20 / 82.38 / 88.0382.86 / 79.86 / 85.85
+ STDF85.83 / 82.80 / 88.8785.18 / 82.27 / 88.0982.90 / 79.92 / 85.89
+ S2SVR86.05 / 83.09 / 89.0185.05 / 82.07 / 88.0482.64 / 79.63 / 85.65
+ Metabit85.47 / 82.04 / 88.9084.95 / 81.57 / 88.3382.32 / 79.04 / 85.59
+ Ours86.08 / 83.13 / 89.0385.31 / 82.38 / 88.2582.88 / 79.83 / 85.92
QDMN85.16 / 82.20 / 88.1184.16 / 81.20 / 87.1279.39 / 76.61 / 82.18
+ MFQE 2.085.13 / 82.20 / 88.0684.15 / 81.18 / 87.1379.51 / 76.75 / 82.27
+ STDF85.32 / 82.38 / 88.2783.36 / 80.44 / 86.2779.64 / 76.69 / 82.59
+ S2SVR85.28 / 82.32 / 88.2383.64 / 80.65 / 86.6379.02 / 76.15 / 81.89
+ Metabit84.50 / 81.14 / 87.8783.68 / 80.30 / 87.0679.47 / 76.41 / 82.52
+ Ours85.34 / 82.41 / 88.2784.37 / 81.42 / 87.3279.78 / 76.92 / 82.65
", + "image_path": "9479e5e22a2236d3a9c02c2064b7046e4d899952767b265d34b94838c44125c6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 478, + 553, + 501 + ], + "lines": [ + { + "bbox": [ + 313, + 478, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 553, + 501 + ], + "type": "text", + "content": "Table 4. Quantitative results of VOS, where the best and second best results are highlighted with bold and underline." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 312, + 509, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 555, + 700 + ], + "type": "text", + "content": "our method consistently reduces the EPE and F1-all loss across all baseline models, demonstrating its effectiveness in improving optical flow estimation. In contrast, methods such as MFQE 2.0, STDF and Metabit fail to deliver consistent improvements. For instance, MFQE 2.0 fails to improve the performance of RAFT on CRF15, STDF and Metabit detrimentally affects the performance of DEQ and KPAFlow on CRF15. Visualizations of predicted optical flow are shown in Fig. 8, where inaccurate boundaries are highlighted with red arrows. As can be seen, optical flow estimated from compressed inputs contains inaccurate boundaries, especially near-motion ones. The proposed method helps to deliver more accurate results in these regions compared to others. For instance, it effectively corrects the optical flow near the car that was mispredicted by DEQ, while MFQE 2.0 and S2SVR provide limited improvement." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "content": "Video object segmentation. For video object segmentation" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17773" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 64, + 128, + 105 + ], + "blocks": [ + { + "bbox": [ + 57, + 64, + 128, + 105 + ], + "lines": [ + { + "bbox": [ + 57, + 64, + 128, + 105 + ], + "spans": [ + { + "bbox": [ + 57, + 64, + 128, + 105 + ], + "type": "image", + "image_path": "ff5ff14a1891bd7c150379f6b76b20dbab27f79bd5c4045498a8cf3efc05f64e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 107, + 104, + 115 + ], + "lines": [ + { + "bbox": [ + 82, + 107, + 104, + 115 + ], + "spans": [ + { + "bbox": [ + 82, + 107, + 104, + 115 + ], + "type": "text", + "content": "STCN" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 64, + 198, + 105 + ], + "blocks": [ + { + "bbox": [ + 129, + 64, + 198, + 105 + ], + "lines": [ + { + "bbox": [ + 129, + 64, + 198, + 105 + ], + "spans": [ + { + "bbox": [ + 129, + 64, + 198, + 105 + ], + "type": "image", + "image_path": "b2fecfcbf265c390fa5c1b1935df9222a35f4bb360f2bc39096046aa6f6d3252.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 107, + 184, + 116 + ], + "lines": [ + { + "bbox": [ + 141, + 107, + 184, + 116 + ], + "spans": [ + { + "bbox": [ + 141, + 107, + 184, + 116 + ], + "type": "text", + "content": "+ MFQE 2.0" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 200, + 64, + 270, + 105 + ], + "blocks": [ + { + "bbox": [ + 200, + 64, + 270, + 105 + ], + "lines": [ + { + "bbox": [ + 200, + 64, + 270, + 105 + ], + "spans": [ + { + "bbox": [ + 200, + 64, + 270, + 105 + ], + "type": "image", + "image_path": "69abf39d07991c266f382e0ef5b2ed082e89560ca6bdb79824d20c24e2b45eab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 107, + 249, + 116 + ], + "lines": [ + { + "bbox": [ + 220, + 107, + 249, + 116 + ], + "spans": [ + { + "bbox": [ + 220, + 107, + 249, + 116 + ], + "type": "text", + "content": "+STDF" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 271, + 64, + 340, + 105 + ], + "blocks": [ + { + "bbox": [ + 271, + 64, + 340, + 105 + ], + "lines": [ + { + "bbox": [ + 271, + 64, + 340, + 105 + ], + "spans": [ + { + "bbox": [ + 271, + 64, + 340, + 105 + ], + "type": "image", + "image_path": "29d9116db90f2ad277f8a8a2503bdd1ad3e70d2fe8b603ef04dafcc6b0b7ae9f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 289, + 107, + 322, + 115 + ], + "lines": [ + { + "bbox": [ + 289, + 107, + 322, + 115 + ], + "spans": [ + { + "bbox": [ + 289, + 107, + 322, + 115 + ], + "type": "text", + "content": "+ S2SVR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 64, + 411, + 105 + ], + "blocks": [ + { + "bbox": [ + 342, + 64, + 411, + 105 + ], + "lines": [ + { + "bbox": [ + 342, + 64, + 411, + 105 + ], + "spans": [ + { + "bbox": [ + 342, + 64, + 411, + 105 + ], + "type": "image", + "image_path": "285533b831b80dec5f2bd8146554e3be0953d9476c16712daba7b15aa0c33006.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 107, + 393, + 115 + ], + "lines": [ + { + "bbox": [ + 360, + 107, + 393, + 115 + ], + "spans": [ + { + "bbox": [ + 360, + 107, + 393, + 115 + ], + "type": "text", + "content": "+Metabit" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 413, + 64, + 482, + 105 + ], + "blocks": [ + { + "bbox": [ + 413, + 64, + 482, + 105 + ], + "lines": [ + { + "bbox": [ + 413, + 64, + 482, + 105 + ], + "spans": [ + { + "bbox": [ + 413, + 64, + 482, + 105 + ], + "type": "image", + "image_path": "1d976055777897c391077cbf53cf4797495eac3730ccd0db47c248535761bbe4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 107, + 459, + 115 + ], + "lines": [ + { + "bbox": [ + 435, + 107, + 459, + 115 + ], + "spans": [ + { + "bbox": [ + 435, + 107, + 459, + 115 + ], + "type": "text", + "content": "+Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 484, + 64, + 553, + 105 + ], + "blocks": [ + { + "bbox": [ + 484, + 64, + 553, + 105 + ], + "lines": [ + { + "bbox": [ + 484, + 64, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 484, + 64, + 553, + 105 + ], + "type": "image", + "image_path": "ed21269d8b4496c4a9c642558390e11dbf3b8dad7c8f9afc7467683bd0955860.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 107, + 524, + 114 + ], + "lines": [ + { + "bbox": [ + 512, + 107, + 524, + 114 + ], + "spans": [ + { + "bbox": [ + 512, + 107, + 524, + 114 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 56, + 149, + 127, + 190 + ], + "blocks": [ + { + "bbox": [ + 55, + 120, + 553, + 143 + ], + "lines": [ + { + "bbox": [ + 55, + 120, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 553, + 143 + ], + "type": "text", + "content": "Figure 9. Qualitative results of VOS. As can be seen, directly performing VOS on compressed inputs leads to inaccurate masks, whereas pre-enhancing with our method effectively improves the accuracy, especially for the regions of irregular shapes (e.g., the windshield)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 149, + 127, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 149, + 127, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 127, + 190 + ], + "type": "image", + "image_path": "9e0c90e81a7c68ecab8eec8d32477e758ccba7bf83db448c0c21b05a7da3ef72.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 193, + 116, + 201 + ], + "lines": [ + { + "bbox": [ + 69, + 193, + 116, + 201 + ], + "spans": [ + { + "bbox": [ + 69, + 193, + 116, + 201 + ], + "type": "text", + "content": "Masked frame" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 205, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 553, + 227 + ], + "type": "text", + "content": "Figure 10. Qualitative results of video inpainting. As can be seen, pre-enhancing the compressed inputs with the proposed method helps to reduce the artifacts and color distortion in the removed region, providing more visually pleasant results." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 128, + 149, + 198, + 190 + ], + "blocks": [ + { + "bbox": [ + 128, + 149, + 198, + 190 + ], + "lines": [ + { + "bbox": [ + 128, + 149, + 198, + 190 + ], + "spans": [ + { + "bbox": [ + 128, + 149, + 198, + 190 + ], + "type": "image", + "image_path": "39e963a30d0bb188e87f7203be7d131e6cd49dc8e91fdcba0981bdce1546b408.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 193, + 177, + 201 + ], + "lines": [ + { + "bbox": [ + 148, + 193, + 177, + 201 + ], + "spans": [ + { + "bbox": [ + 148, + 193, + 177, + 201 + ], + "type": "inline_equation", + "content": "E^2" + }, + { + "bbox": [ + 148, + 193, + 177, + 201 + ], + "type": "text", + "content": " FGVI" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 198, + 149, + 269, + 190 + ], + "blocks": [ + { + "bbox": [ + 198, + 149, + 269, + 190 + ], + "lines": [ + { + "bbox": [ + 198, + 149, + 269, + 190 + ], + "spans": [ + { + "bbox": [ + 198, + 149, + 269, + 190 + ], + "type": "image", + "image_path": "353d0f65ab54f34c566199b26ad022236da607bea4190039b49d117aa7871c32.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 193, + 255, + 201 + ], + "lines": [ + { + "bbox": [ + 213, + 193, + 255, + 201 + ], + "spans": [ + { + "bbox": [ + 213, + 193, + 255, + 201 + ], + "type": "text", + "content": "+ MFQE 2.0" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 271, + 149, + 340, + 190 + ], + "blocks": [ + { + "bbox": [ + 271, + 149, + 340, + 190 + ], + "lines": [ + { + "bbox": [ + 271, + 149, + 340, + 190 + ], + "spans": [ + { + "bbox": [ + 271, + 149, + 340, + 190 + ], + "type": "image", + "image_path": "71b79d6c54308e89a5dc76bd0d742ce0b298b3658dd878e54a4e811f3d8a016e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 193, + 318, + 201 + ], + "lines": [ + { + "bbox": [ + 291, + 193, + 318, + 201 + ], + "spans": [ + { + "bbox": [ + 291, + 193, + 318, + 201 + ], + "type": "text", + "content": "+STDF" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 341, + 149, + 411, + 190 + ], + "blocks": [ + { + "bbox": [ + 341, + 149, + 411, + 190 + ], + "lines": [ + { + "bbox": [ + 341, + 149, + 411, + 190 + ], + "spans": [ + { + "bbox": [ + 341, + 149, + 411, + 190 + ], + "type": "image", + "image_path": "02add9649d9c6e82a9ee5e522da514b181cd8f884c52d19d5aa5c2558adf03e6.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 193, + 392, + 201 + ], + "lines": [ + { + "bbox": [ + 361, + 193, + 392, + 201 + ], + "spans": [ + { + "bbox": [ + 361, + 193, + 392, + 201 + ], + "type": "text", + "content": "+ S2SVR" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 411, + 149, + 482, + 190 + ], + "blocks": [ + { + "bbox": [ + 411, + 149, + 482, + 190 + ], + "lines": [ + { + "bbox": [ + 411, + 149, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 411, + 149, + 482, + 190 + ], + "type": "image", + "image_path": "479b395931600a7b12769eed239b2d28969843af9abf1730f8fc8df5b229c082.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 431, + 193, + 463, + 201 + ], + "lines": [ + { + "bbox": [ + 431, + 193, + 463, + 201 + ], + "spans": [ + { + "bbox": [ + 431, + 193, + 463, + 201 + ], + "type": "text", + "content": "+Metabit" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 483, + 149, + 553, + 190 + ], + "blocks": [ + { + "bbox": [ + 483, + 149, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 483, + 149, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 483, + 149, + 553, + 190 + ], + "type": "image", + "image_path": "2c13d21a0518fb74cda51b730621a1774293d4d5eddc95055d2f8a484d5dd541.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 193, + 529, + 201 + ], + "lines": [ + { + "bbox": [ + 506, + 193, + 529, + 201 + ], + "spans": [ + { + "bbox": [ + 506, + 193, + 529, + 201 + ], + "type": "text", + "content": "+Ours" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "text", + "content": "(VOS), we adopt STCN [14], DeAoT [57] and QDMN [34] as representative baselines. Evaluations on DAVIS-17 val dataset [41] are summarized with the following metrics: the " + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "text", + "content": " (average IoU), the " + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "text", + "content": " score (boundary similarity), and the average of the above metrics (denoted as " + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "inline_equation", + "content": "Avg" + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "text", + "content": "). Higher values indicate better segmentation accuracy. As shown in Tab. 4, the proposed method shows the best performance in improving accuracy across VOS models. For instance, it elevates the average accuracy for up to " + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "inline_equation", + "content": "2.37\\%" + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "text", + "content": " (79.20% to " + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "inline_equation", + "content": "81.57\\%" + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "text", + "content": ") on STCN at CRF35, while MFQE 2.0, STDF and S2SVR yield limited improvement of " + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "inline_equation", + "content": "0.08\\%" + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "inline_equation", + "content": "0.57\\%" + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "inline_equation", + "content": "0.84\\%" + }, + { + "bbox": [ + 56, + 240, + 296, + 479 + ], + "type": "text", + "content": ", respectively. And Metabit provides no improvement on STCN at CRF35. The results of VOS are included in Fig. 9, where accurately segmenting objects in compressed videos is challenging for VOS baselines (e.g., inaccurate mask of the windshield predicted by STCN). Pre-enhancing the compressed videos with MFQE 2.0 and S2SVR struggles to address this issue, whereas the proposed method significantly refines the segmentation results, demonstrating its effectiveness in assisting the VOS task." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 480, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 589 + ], + "type": "text", + "content": "Video inpainting. We take " + }, + { + "bbox": [ + 55, + 480, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{FGVI}" + }, + { + "bbox": [ + 55, + 480, + 296, + 589 + ], + "type": "text", + "content": " [31] as the video inpainting model and perform video object removal on DAVIS-17 val dataset [41]. The qualitative results are shown in Fig. 10. As can be seen, compression-included misalignment between objects and masks hinders the ability to remove specified objects, causing color distortions (e.g., the horse region). Pre-enhancing the compressed inputs with the proposed method notably refines the artifacts and distortions, yielding more visually pleasing results." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 598, + 156, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 156, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 156, + 609 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "content": "We start with a baseline that concatenates reference frames and current frames as input, without using codec information. We then progressively equip the baseline with MV alignment, region-aware refinement, sequence adaptation, and frame adaptation to assess their contributions." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "MV alignment. As shown in the 2nd row of Tab. 5, incorporating MV alignment yields a PSNR gain of up to 0.65 dB on CRF15, demonstrating the effectiveness of MV in" + } + ] + } + ], + "index": 34 + }, + { + "type": "table", + "bbox": [ + 317, + 239, + 552, + 323 + ], + "blocks": [ + { + "bbox": [ + 317, + 239, + 552, + 323 + ], + "lines": [ + { + "bbox": [ + 317, + 239, + 552, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 552, + 323 + ], + "type": "table", + "html": "
ModelCRF15CRF25CRF35
PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
Baseline41.04 / 0.978534.92 / 0.936329.25 / 0.8238
+ MV Align.41.69 / 0.982135.59 / 0.943729.95 / 0.8403
+ RA Refine.42.04 / 0.983735.70 / 0.944930.00 / 0.8427
+ Seq. Adapt.42.08 / 0.983835.76 / 0.945830.04 / 0.8444
+ Frame Adapt.42.14 / 0.983935.81 / 0.946030.09 / 0.8446
", + "image_path": "11afc9a2fe8cc1646e627742107f5260c90a02dd0e3da6bfa81e8ea1803109e7.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "table_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 350 + ], + "type": "text", + "content": "Table 5. Ablation studies on MV alignment, region-aware refinement, sequence adaptation, and frame adaptation." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 361, + 495, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 495, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 495, + 373 + ], + "type": "text", + "content": "aligning reference frames and current frame." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 374, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 433 + ], + "type": "text", + "content": "Region-aware refinement. We further incorporate the region-aware refinement module to refine the features of different regions. As shown in the 3rd row of Tab. 5, it leads to notable PSNR gains of " + }, + { + "bbox": [ + 313, + 374, + 554, + 433 + ], + "type": "inline_equation", + "content": "0.35\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 374, + 554, + 433 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 374, + 554, + 433 + ], + "type": "inline_equation", + "content": "0.11\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 374, + 554, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 374, + 554, + 433 + ], + "type": "inline_equation", + "content": "0.05\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 374, + 554, + 433 + ], + "type": "text", + "content": " on CRF15, CRF25 and CRF35, respectively." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "type": "text", + "content": "Sequence adaptation. As shown in the 4th row of Tab. 5, sequence adaptation brings PSNR gains of " + }, + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "type": "inline_equation", + "content": "0.04\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "type": "inline_equation", + "content": "0.06\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "type": "inline_equation", + "content": "0.04\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 434, + 553, + 469 + ], + "type": "text", + "content": " on CRF15, CRF25 and CRF35, respectively." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 470, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 530 + ], + "type": "text", + "content": "Frame adaptation. As shown in the 5th row of Tab. 5, frame adaptation improves PSNR by " + }, + { + "bbox": [ + 313, + 470, + 554, + 530 + ], + "type": "inline_equation", + "content": "0.06\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 470, + 554, + 530 + ], + "type": "inline_equation", + "content": "0.05\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 470, + 554, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 470, + 554, + 530 + ], + "type": "inline_equation", + "content": "0.05\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 470, + 554, + 530 + ], + "type": "text", + "content": " on CRF15, CRF25 and CRF35, respectively. We further analyze its effectiveness on improving the temporal consistency in the supplementary materials." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 541, + 388, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 388, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 388, + 553 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 558, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 677 + ], + "type": "text", + "content": "In this paper, we introduce a versatile codec-aware enhancement framework that adaptively handles diverse compression settings and serves as a plug-and-play enhancement module to consistently boost various downstream tasks. By reusing the off-the-shelf codec information, our method minimizes additional computational costs. Compared with existing compressed video enhancement solutions, it shows superiority in both enhancement performance and robustness, making it possible to deploy pre-trained models on compressed videos without a significant performance drop." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "type": "text", + "content": "Acknowledgments. We acknowledge funding from the National Natural Science Foundation of China under Grants 62131003 and 62021001." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17774" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Video developer report 2022-2023. https://bitmovin.com/downloads/ assets/bitmovin-7th-video-developer-report-2023-2024.pdf. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "text", + "content": "[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In CVPR, pages 3366-3375, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "type": "text", + "content": "[3] Shaojie Bai, Zhengyang Geng, Yash Savani, and J. Zico Kolter. Deep equilibrium optical flow estimation. In CVPR, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "type": "text", + "content": "[4] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (vvc) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 294, + 294 + ], + "type": "text", + "content": "[5] Yuxuan Cai, Hongjia Li, Geng Yuan, Wei Niu, Yanyu Li, Xulong Tang, Bin Ren, and Yanzhi Wang. Yolobile: Realtime object detection on mobile devices via compression-compiler co-design. In AAAI, pages 955–963, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 338 + ], + "type": "text", + "content": "[6] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, pages 4947-4956, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "type": "text", + "content": "[7] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, pages 5972-5981, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "type": "text", + "content": "[8] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating tradeoffs in real-world video super-resolution. In CVPR, pages 5962-5971, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 431, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 294, + 474 + ], + "type": "text", + "content": "[9] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In ICIP, pages 168-172, 1994. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 476, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 476, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 57, + 476, + 294, + 509 + ], + "type": "text", + "content": "[10] Jin Chen, Xijun Wang, Zichao Guo, Xiangyu Zhang, and Jian Sun. Dynamic region-aware convolution. In CVPR, pages 8064-8073, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "type": "text", + "content": "[11] Peilin Chen, Wenhan Yang, Long Sun, and Shiqi Wang. When bitstream prior meets deep prior: Compressed video super-resolution with learning from decoding. In ACM MM, pages 1000–1008, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "type": "text", + "content": "[12] Peilin Chen, Wenhan Yang, Meng Wang, Long Sun, Kangkang Hu, and Shiqi Wang. Compressed domain deep video super-resolution. IEEE Transactions on Image Processing, 30:7156-7169, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 643 + ], + "type": "text", + "content": "[13] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In CVPR, pages 11030-11039, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "text", + "content": "[14] Ho Kei Cheng, Yu-Wing Tai, and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[15] Yuanying Dai, Dong Liu, and Feng Wu. A convolutional neural network approach for post-processing in hevc intra coding. In MultiMedia Modeling: 23rd International Con" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ference, MMM 2017, Reykjavik, Iceland, January 4-6, 2017, Proceedings, Part I 23, pages 28-39. Springer, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[16] Jianing Deng, Li Wang, Shiliang Pu, and Cheng Zhuo. Spatio-temporal deformable convolution for compressed video quality enhancement. In AAAI, pages 10696-10703, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[17] Cunhui Dong, Haichuan Ma, Zhuoyuan Li, Li Li, and Dong Liu. Temporal wavelet transform-based low-complexity perceptual quality enhancement of compressed video. IEEE TCSVT, pages 1-1, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "text", + "content": "[18] Max Ehrlich, Jon Barker, Namitha Padmanabhan, Larry Davis, Andrew Tao, Bryan Catanzaro, and Abhinav Shrivastava. Leveraging bitstream metadata for fast, accurate, generalized compressed video quality enhancement. In WACV, pages 1517-1527, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 243, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 275 + ], + "type": "text", + "content": "[19] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 277, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 353 + ], + "type": "text", + "content": "[20] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaeser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 397 + ], + "type": "text", + "content": "[21] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "text", + "content": "[22] Zhenyu Guan, Qunliang Xing, Mai Xu, Ren Yang, Tie Liu, and Zulin Wang. Mfqe 2.0: A new approach for multi-frame quality enhancement on compressed video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3): 949-963, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "text", + "content": "[23] Yizeng Han, Gao Huang, Shiji Song, Le Yang, Honghui Wang, and Yulin Wang. Dynamic neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7436-7456, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "text", + "content": "[24] Hongyue Huang, Ionut Schiopu, and Adrian Munteanu. Frame-wise cnn-based filtering for intra-frame quality enhancement of hevc videos. IEEE Transactions on Circuits and Systems for Video Technology, 31(6):2100-2113, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "type": "text", + "content": "[25] Zhijie Huang, Jun Sun, and Xiaopeng Guo. Fastconn: Towards fast and accurate spatiotemporal network for hevc compressed video enhancement. ACM Transactions on Multimedia Computing, Communications and Applications, 19 (3):1-22, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "type": "text", + "content": "[26] Chuanmin Jia, Shiqi Wang, Xinfeng Zhang, Shanshe Wang, Jiaying Liu, Shiliang Pu, and Siwei Ma. Content-aware convolutional neural network for in-loop filtering in high efficiency video coding. IEEE Transactions on Image Processing, 28(7):3343-3356, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[27] Nanfeng Jiang, Weiling Chen, Jielian Lin, Tiesong Zhao, and Chia-Wen Lin. Video compression artifacts removal with spatial-temporal attention-guided enhancement. IEEE Transactions on Multimedia, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17775" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[28] Feng Li, Yixuan Wu, Anqi Li, Huihui Bai, Runmin Cong, and Yao Zhao. Enhanced video super-resolution network towards compressed data. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-21, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[29] Yinxiao Li, Pengchong Jin, Feng Yang, Ce Liu, Ming-Hsuan Yang, and Peyman Milanfar. Comisr: Compression-informed video super-resolution. In ICCV, pages 2543-2552, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "type": "text", + "content": "[30] Yunsheng Li, Lu Yuan, Yinpeng Chen, Pei Wang, and Nuno Vasconcelos. Dynamic transfer for multi-source domain adaptation. In CVPR, pages 10998-11007, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 241 + ], + "type": "text", + "content": "[31] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In CVPR, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "text", + "content": "[32] Jing Lin, Xiaowan Hu, Yuanhao Cai, Haoqian Wang, Youliang Yan, Xueyi Zou, Yulun Zhang, and Luc Van Gool. Unsupervised flow-aligned sequence-to-sequence learning for video restoration. In ICML, pages 13394-13404. PMLR, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "text", + "content": "[33] Baoyuan Liu, Min Wang, Hassan Foroosh, Marshall Tappen, and Marianna Pensky. Sparse convolutional neural networks. In CVPR, pages 806-814, 2015. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[34] Yong Liu, Ran Yu, Fei Yin, Xinyuan Zhao, Wei Zhao, Weihao Xia, and Yujiu Yang. Learning quality-aware dynamic memory for video object segmentation. In ECCV, pages 468-486, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "text", + "content": "[35] Ao Luo, Fan Yang, Xin Li, and Shuaicheng Liu. Learning optical flow with kernel patch attention. In CVPR, pages 8906-8915, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "type": "text", + "content": "[36] Di Ma, Fan Zhang, and David R Bull. Cvegan: a perceptually-inspired gan for compressed video enhancement. Signal Processing: Image Communication, page 117127, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "type": "text", + "content": "[37] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1930-1939, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "text", + "content": "[38] Sreyas Mohan, Joshua L Vincent, Ramon Manzorro, Peter Crozier, Carlos Fernandez-Granda, and Eero Simoncelli. Adaptive denoising via gaintuning. NeurIPS, 34:23727-23740, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 623 + ], + "type": "text", + "content": "[39] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, pages 0–0, 2019. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "text", + "content": "[40] Zhaoqing Pan, Xiaokai Yi, Yun Zhang, Byeungwoo Jeon, and Sam Kwong. Efficient in-loop filtering based on enhanced deep convolutional neural networks for hevc. IEEE Transactions on Image Processing, 29:5352-5366, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[41] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 127 + ], + "type": "text", + "content": "[42] Darren Ramsook and Anil Kokaram. Learnt deep hyperparameter selection in adversarial training for compressed video enhancement with a perceptual critic. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2420-2424. IEEE, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 129, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 554, + 172 + ], + "type": "text", + "content": "[43] Mennatullah Siam, Mostafa Gamal, Moemen Abdel-Razek, Senthil Yogamani, Martin Jagersand, and Hong Zhang. A comparative study of real-time semantic segmentation for autonomous driving. In CVPRW, pages 587-597, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 173, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 173, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 315, + 173, + 554, + 205 + ], + "type": "text", + "content": "[44] Mennatullah Siam, Alex Kendall, and Martin Jagersand. Video class agnostic segmentation benchmark for autonomous driving. In CVPR, pages 2825-2834, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 555, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 249 + ], + "type": "text", + "content": "[45] Gary J Sullivan, Pankaj N Topiwala, and Ajay Luthra. The h. 264/avc advanced video coding standard: Overview and introduction to the fidelity range extensions. Applications of Digital Image Processing XXVII, 5558:454-474, 2004. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 293 + ], + "type": "text", + "content": "[46] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on Circuits and Systems for Video Technology, 22(12):1649-1668, 2012. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "type": "text", + "content": "[47] Weiheng Sun, Xiaohai He, Chao Ren, Shuhua Xiong, and Honggang Chen. A quality enhancement network with coding priors for constant bit rate video coding. Knowledge-Based Systems, 258:110010, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 554, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 380 + ], + "type": "text", + "content": "[48] Xuebin Sun, Sukai Wang, Miaohui Wang, Shing Shin Cheng, and Ming Liu. An advanced lidar point cloud sequence coding scheme for autonomous driving. In MM, pages 2793-2801, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 554, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 413 + ], + "type": "text", + "content": "[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "type": "text", + "content": "[50] Jianyi Wang, Mai Xu, Xin Deng, Liquan Shen, and Yuhang Song. Mw-gan+ for perceptual quality enhancement on compressed video. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4224-4237, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 491 + ], + "type": "text", + "content": "[51] Shiyao Wang, Hongchao Lu, and Zhidong Deng. Fast object detection in compressed video. In ICCV, pages 7104-7113, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 493, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 554, + 525 + ], + "type": "text", + "content": "[52] Yingwei Wang, Takashi Isobe, Xu Jia, Xin Tao, Hutchuan Lu, and Yu-Wing Tai. Compression-aware video superresolution. In CVPR, pages 2012-2021, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 526, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 568 + ], + "type": "text", + "content": "[53] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on Circuits and Systems for Video Technology, 13(7):560-576, 2003. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 570, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 554, + 602 + ], + "type": "text", + "content": "[54] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. NeurIPS, 32, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 604, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 554, + 635 + ], + "type": "text", + "content": "[55] Ren Yang, Mai Xu, Zulin Wang, and Tianyi Li. Multi-frame quality enhancement for compressed video. In CVPR, pages 6664-6673, 2018. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 636, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 554, + 679 + ], + "type": "text", + "content": "[56] Ren Yang, Mai Xu, Tie Liu, Zulin Wang, and Zhenyu Guan. Enhancing quality for hevc compressed videos. IEEE Transactions on Circuits and Systems for Video Technology, 29(7): 2039-2054, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 712 + ], + "type": "text", + "content": "[57] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation. In NeurIPS, 2022. 8" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17776" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 397 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[58] Li Yu, Wenshuai Chang, Shiyu Wu, and Moncef Gabbouj. End-to-end transformer for compressed video quality enhancement. IEEE Transactions on Broadcasting, 70(1):197-207, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[59] Haichao Zhang, Jianchao Yang, Yanning Zhang, Nasser M Nasrabadi, and Thomas S Huang. Close the loop: Joint blind image restoration and recognition with sparse representation prior. In ICCV, pages 770-777, 2011. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[60] Hengsheng Zhang, Xueyi Zou, Jiaming Guo, Youliang Yan, Rong Xie, and Li Song. A codec information assisted framework for efficient compressed video super-resolution. In ECCV, pages 220-235. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[61] Saiping Zhang, Luis Herranz, Marta Mrak, Marc Górriz Blanch, Shuai Wan, and Fuzheng Yang. Dcngan: A deformable convolution-based gan with qp adaptation for perceptual quality enhancement of compressed video. In ICASSP, pages 2035-2039. IEEE, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "text", + "content": "[62] Zhengdong Zhang and Vivienne Sze. Fast: A framework to accelerate super-resolution processing on compressed videos. In CVPRW, pages 19-28, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "text", + "content": "[63] Hengrun Zhao, Bolun Zheng, Shanxin Yuan, Hua Zhang, Chenggang Yan, Liang Li, and Gregory Slabaugh. Cbren: Convolutional neural networks for constant bit rate video quality enhancement. IEEE Transactions on Circuits and Systems for Video Technology, 32(7):4138-4149, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "type": "text", + "content": "[64] Qiang Zhu, Jinhua Hao, Yukang Ding, Yu Liu, Qiao Mo, Ming Sun, Chao Zhou, and Shuyuan Zhu. Cpga: Coding priors-guided aggregation network for compressed video quality enhancement. In CVPR, pages 2964-2974, 2024. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17777" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/96fbf65a-6eae-49b1-9bdc-b206a300f6d4_content_list.json b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/96fbf65a-6eae-49b1-9bdc-b206a300f6d4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cd80b040c22b834fb84d427c44da087550e93139 --- /dev/null +++ b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/96fbf65a-6eae-49b1-9bdc-b206a300f6d4_content_list.json @@ -0,0 +1,1741 @@ +[ + { + "type": "text", + "text": "Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression", + "text_level": 1, + "bbox": [ + 114, + 128, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boqian Zhang\\*, Shen Yang\\*, Hao Chen, Chao Yang, Jing Jia, Guang Jiang† \nXidian University", + "bbox": [ + 207, + 202, + 795, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{24011211294，20009101459，haochen(cv，24011211295}@xidian.stu.edu.cn,", + "bbox": [ + 189, + 241, + 805, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jjia@mail.xidian.edu.cn, gjiang@mail.xidian.edu.cn", + "bbox": [ + 274, + 258, + 718, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud upsampling can improve the quality of the initial point cloud, significantly enhancing the performance of downstream tasks such as classification and segmentation. Existing methods mostly focus on generating the geometric details of point clouds, neglecting noise suppression. To address this, we propose a novel network based on a conditional diffusion model, incorporating the Adaptive Noise Suppression (ANS) module, which we refer to as PDANS. The ANS module assigns weights to each point and determines the removal strategy based on these weights, reducing the impact of noisy points on the sampling process. The module first selects the neighborhood set for each point in the point cloud and performs a weighted sum between the point and its neighbors. It then adjusts the removal points based on the weighted sum, effectively mitigating the bias caused by outliers. We introduce the TreeTrans (TT) module to capture more correlated feature information. This module learns the interaction between high-level and low-level features, resulting in a more comprehensive and refined feature representation. Our results on several widely used benchmark datasets demonstrate that PDANS exhibits exceptional robustness in noisy point cloud processing and outperforms current state-of-the-art(SOTA) methods in terms of performance. Code is available at https://github.com/Baty2023/PDANS.", + "bbox": [ + 89, + 340, + 483, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 747, + 223, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point clouds, as one of the key representations for 3D reconstruction [17] [22], are widely applied in autonomous driving [5] [47], embodied intelligence [18] [42], and AR/VR. However, real-world scanned point clouds are often sparse, noisy, and uneven, negatively impacting their performance in downstream tasks such as classification and segmenta", + "bbox": [ + 89, + 773, + 482, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion. Point cloud upsampling can transform sparse and noisy point clouds into dense, artifact-free representations. In recent years, significant progress has been made in this task [15] [27] [45] [40].", + "bbox": [ + 511, + 310, + 906, + 371 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/954005717d3044c550c2828f584847ee24c018e74bc14fa9b0731a959fa5970f.jpg", + "image_caption": [ + "Figure 1. Most existing methods achieve satisfactory results when input without noise, but they often struggle to effectively recover the geometric structure (e.g., panda ears) when dealing with input with noise." + ], + "image_footnote": [], + "bbox": [ + 535, + 385, + 897, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by deep learning, several methods have adopted deep networks to perform this task. Existing upsampling methods can be broadly classified into two categories: those based on generative models and those based on refinement methods. Generative models [15] [27] [28] first extract features from sparse point clouds into a latent space and then design a generator based on the extracted features to generate the coordinates of dense point clouds. The limitation of such methods lies in the difficulty of accurately predicting 3D coordinates, leading to potential outliers or artifacts in the generated dense point clouds [8] [45]. Refinement-based methods [4] [39] [30] first use interpolation to generate a coarse dense point cloud, and then employ a neural network to extract geometric features from the coarse point cloud for refinement. Combined with reconstruction loss functions, this process refines the point cloud to match the ground truth.", + "bbox": [ + 509, + 583, + 906, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These methods share two common issues: (1) The features lack significant dominance or distinguishability, leading to generating point clouds without clear structure, which tend to be sparse and scattered. (2) Most of these methods", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 112, + 875, + 370, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors.", + "bbox": [ + 112, + 887, + 246, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16987", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rely on loss functions such as CD loss, EMD loss, and other similar metrics, which are highly sensitive to outliers. This sensitivity makes the model ineffective at handling noisy point clouds [12] [38], as shown in Figure1.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we propose a point cloud upsampling method with adaptive noise suppression, PDANS. The network is designed based on a conditional diffusion model, leveraging the inherent denoising properties of the diffusion model's generative process to enhance its adaptability to noisy point clouds. The network also integrates an Adaptive Noise Suppression (ANS) module, which we designed to compute downsampling weights for each point by utilizing the relationships between the point and its neighboring points. This approach enables the extraction of more accurate downsampled points and features. It effectively reduces the impact of noisy points during feature extraction and minimizes the bias caused by outliers.", + "bbox": [ + 89, + 154, + 482, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance feature representation, we propose a structure based on graph convolutional networks [14], called the TreeTrans (TT) module. This module incorporates ancestral and recurrent terms[32] to fully utilize the features outputted by each encoder layer. Additionally, we apply both local and global attention mechanisms to the output features, further improving the efficiency and expressiveness of the feature representation.", + "bbox": [ + 89, + 352, + 482, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we designed a conditional point cloud network and a diffusion point cloud network. The conditional point cloud network provides relevant features of sparse point clouds to the diffusion point cloud network through a mutual attention module, guiding the diffusion network to generate the required noise for point cloud densification. Similar to PUDM[29], we incorporate the diffusion model, thereby eliminating the reliance on CD loss. Combined with the ANS module we designed, PDANS demonstrates exceptional robustness in noise experiments. Our main contributions are as follows:", + "bbox": [ + 89, + 474, + 482, + 641 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We designed a noise suppression module named ANS, which adaptively adjusts removal points and their neighborhood features to improve robustness against noise and outliers.", + "- The proposed TT module fully leverages features across different layers of the encoder, incorporating ancestral terms, recurrent terms, and local and global attention mechanisms to achieve more comprehensive and refined feature representations.", + "- We designed an end-to-end network architecture named PDANS, which utilizes a mutual attention mechanism and a diffusion model to guide the densification of sparse point clouds. The ANS module further enhances the network's ability to handle noise.", + "- Extensive experiments demonstrate the outstanding performance of PDANS in point cloud upsampling benchmarks, particularly in generating geometric details and" + ], + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "handling noisy point clouds effectively.", + "bbox": [ + 527, + 90, + 787, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 121, + 653, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Learning-based point cloud upsampling", + "text_level": 1, + "bbox": [ + 511, + 147, + 856, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several methods [25] [26] [36] [11] shave been developed to extract features from 3D data directly with the strong data-driven and representational capabilities of deep learning networks. PU-Net [45] was the first learning-based method for point cloud and upsampling. Following its proposed architecture, many methods have emerged: MPU [44], PUGAN [15], Dis-PU [16], PUGCN [27], and RepKPU [30]. The NePs [7] method enhances neural radiance fields by combining 2D information with 3D features, enabling point cloud upsampling. Some networks attempt to perform point cloud upsampling based on denoising paradigms. They generate a rough upsampled point cloud through interpolation, and then refine it using neural networks, such as GradPU [8] and PU-Flow [21].", + "bbox": [ + 511, + 169, + 903, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most methods predict the coordinates of dense point clouds from sparse point cloud features and then rely on upsampling modules to expand the point cloud, or further refine the coordinates from the coarsely densified point clouds. Since these methods typically depend on CD (Chamfer Distance) and EMD (Earth Mover's Distance) as loss functions, they exhibit high sensitivity to noise. In this paper, we design an additional denoising module and utilize conditional DDPM to reduce the dependence on CD.", + "bbox": [ + 511, + 383, + 903, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. DDPM for 3D Generation", + "text_level": 1, + "bbox": [ + 513, + 530, + 748, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the success of image generation tasks [10] [34] [9], direct point cloud generation using DDPM has gained increasing attention. Luo et al. [19] is the first to apply DDPM to unconditional point cloud generation. The DAM [35] network explores the interpretability of diffusion models for point cloud generation tasks. Nunes et al. [23] apply DDPM to the 3D-LiDAR point cloud completion task in real-world scenarios. The PDR [20] network takes the initial point cloud as input, but it introduces a coarse and refinement network, which increases time overhead. PUDM [29] densifies point clouds through conditional diffusion, but its performance in noise suppression is suboptimal.", + "bbox": [ + 511, + 551, + 903, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper uses a conditional diffusion model to perform point cloud upsampling. To enhance the model's denoising capability, we design an additional ANS module.", + "bbox": [ + 511, + 734, + 903, + 780 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Background of Conditional Diffusion Module", + "text_level": 1, + "bbox": [ + 511, + 796, + 903, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We assume that the dense point cloud $x$ is sampled from a meaningful distribution $P_{data}$ , and that the latent variable $x_{T} \\sim \\mathcal{N}(\\mathbf{0}_{3N},\\mathbf{I}_{3N\\times 3N})$ is sampled from a noise distribution $P_{latent}$ which is easier to handle, where $\\mathcal{N}$ follows a", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16988", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaussian distribution. DDPM[10] establishes a transfer relationship between $x$ and $x_{T}$ through two Markov chains. This is an autoregressive process: the forward process progressively adds noise to $x$ until $x_{T}$ , and $x_{T}$ is denoised back to $x$ through the reverse process. We use the timestep $t \\sim U(T)$ to control the speed of transformation.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Forward Process The process gradually adds noise to $x_0$ until $x_T$ through the probability $q(x_t | x_{t-1})$ . The entire process can be defined as:", + "bbox": [ + 89, + 181, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {1}, \\dots , x _ {T} \\mid x _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(x _ {t} \\mid x _ {t - 1}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 238, + 482, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $q(x_{t}|x_{t - 1}) = \\mathcal{N}(x_{t};\\sqrt{1 - \\beta_{t}} x_{t - 1},\\beta_{t}I)$ , where the hyperparameter $\\beta_{t}$ is a positive constant. Based on the derivation by Ho et al. [10], $x_{T}$ can be computed directly from $x_0$ . Define $\\alpha_{t} = 1 - \\beta_{t}$ and $\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}$ , so that $x_{T}$ can be sampled from the following formula:", + "bbox": [ + 89, + 291, + 483, + 367 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 377, + 482, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $\\epsilon$ follows a standard normal distribution.", + "bbox": [ + 89, + 406, + 406, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reverse Process The reverse process aims to predict and eliminate the noise added at each step $t$ . It is conditioned on a set of conditions, which may include incomplete point clouds $c$ , upsampling factor $r$ , etc. We denote the set of conditions as $C = \\{c_i \\mid i = 1, \\dots, N_c\\}$ , where $N_c$ represents the number of conditions. We can define the reverse process from denoising $x_{T}$ to $x_{0}$ as follows:", + "bbox": [ + 89, + 421, + 483, + 527 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\boldsymbol {\\theta}} \\left(x _ {0}, \\dots , x _ {T - 1} \\mid x _ {T}, \\boldsymbol {C}\\right) = \\prod_ {t = 1} ^ {T} p _ {\\boldsymbol {\\theta}} \\left(x _ {t - 1} \\mid x _ {t}, \\boldsymbol {C}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 537, + 482, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $p_{\\boldsymbol{\\theta}}(x_{t - 1}|x_t,\\boldsymbol {C}) = \\mathcal{N}(x_{t - 1};\\mu_{\\boldsymbol{\\theta}}(\\boldsymbol {x}_t,\\boldsymbol {C},t),\\sigma_t^2\\boldsymbol {I})$ where the mean $\\mu_{\\boldsymbol{\\theta}}(x_t,C,t)$ is the result predicted by a neural network, and $\\sigma_t^2$ is a time-dependent constant. Based on the derivation by DDIM [33], the reverse process can be written in the following form:", + "bbox": [ + 89, + 590, + 483, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (x _ {t} , \\boldsymbol {C} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) + \\tag {4} \\\\ \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} (x _ {t}, C, t) + \\sigma_ {t} \\epsilon_ {t} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 676, + 482, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $\\epsilon_{t} \\sim \\mathcal{N}(\\mathbf{0}_{3N},\\mathbf{I}_{3N\\times 3N})$ is standard Gaussian noise independent of $x_{T}$ . The noisy point cloud $x_{T}$ is progressively denoised according to the above equation until it recovers $x_0$ . At each time step of the reverse process, conditions are added, allowing control over the generation of dense point clouds, as shown in the process diagram in Figure 2.", + "bbox": [ + 89, + 750, + 482, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training process under given conditions The goal is to optimize the denoising model's prediction of the noise $\\epsilon_{\\theta}$ at step $t$ . Given a set of conditions $C$ and under the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4bf08cbdb1fcf68f37c1937cece8f9526363ded9fe1077aec8bc26a5170a3c84.jpg", + "image_caption": [ + "Figure 2. Visualization of the forward and reverse processes of PDANS." + ], + "image_footnote": [], + "bbox": [ + 521, + 92, + 898, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "condition that the time step $t\\sim U(T)$ , the training objective of DDPM is:", + "bbox": [ + 511, + 279, + 906, + 308 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {\\boldsymbol {t} \\sim U (T), \\epsilon \\sim \\mathcal {N} (0, I)} \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon + \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, C, t\\right) \\| ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 316, + 915, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the above equation, there are no CD loss or EMD loss terms commonly used in traditional point cloud upsampling networks, because DDPM naturally defines a one-to-one pointwise correspondence between two consecutive point clouds, as shown in equation1.", + "bbox": [ + 511, + 349, + 906, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 436, + 604, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces each module's design and the network's overall structure. In PDANS, we propose two modules: the Adaptive Noise Suppression (ANS) module and the TreeTrans (TT) module. These modules are described in Sections 4.1 and 4.2, respectively. The design of the overall PDANS network is presented in Section 4.3.", + "bbox": [ + 511, + 462, + 905, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Adaptive Noise Suppression(ANS) Module", + "text_level": 1, + "bbox": [ + 511, + 561, + 874, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the encoding stage, many frameworks perform down-sampling using Farthest Point Sampling (FPS), followed by selecting neighboring points within a certain radius [20] or via KNN [1] to enhance the features of the sampled points. However, two issues arise when the input point cloud contains significant noise: (1) FPS is sensitive to outliers[43][41], making it challenging to process noisy point clouds effectively. (2) FPS does not filter noisy points, leading to the inclusion of noise features in the enhanced features, which affects the accuracy and reliability of the sampling.", + "bbox": [ + 511, + 583, + 905, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose the Adaptive Noise Suppression (ANS) module to address the issues abovementioned, as shown in Figure 3. Unlike FPS, we first select the neighborhood point cloud $\\mathcal{P}_K\\in \\mathbb{R}^{N\\times 3\\times K}$ for all points in the initial point cloud $\\mathcal{P}\\in \\mathbb{R}^{N\\times 3}$ , where $K$ denotes the number of neighboring points for each point. These neighborhood point clouds contain both correct and noisy points. We then input the neighborhood point cloud, the initial point cloud, and their corresponding features into the ANS module to perform the removal operation.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16989", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fd530f1e0df6cb8f5f1b38e56839fad1c90b4e58fa275621700cde9d501592f9.jpg", + "image_caption": [ + "Figure 3. Visualization of the Adaptive Noise Suppression (ANS) module" + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 472, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each point $x_{i}$ selected from $\\mathcal{P}$ , we select its corresponding feature $f_{i}$ from $\\mathcal{F}s \\in \\mathbb{R}^{N \\times D_l}$ , where $\\mathcal{F}s$ is the feature set obtained by reordering the initial point cloud. We use KNN to select its $K$ nearest neighbors $x_{i,1}, \\ldots, x_{i,K} \\in \\mathcal{N}(x_i)$ and their corresponding features $f_{i,1}, \\ldots, f_{i,K} \\in \\mathcal{N}(f_i)$ , where $\\mathcal{N}(x_i)$ represents the $K$ nearest neighbors of point $x_{i}$ . Next, we calculate the point and feature weights using an attention mechanism, with both points and features updated similarly. $w(f_{i})$ calculation can be expressed as:", + "bbox": [ + 89, + 349, + 483, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw \\left(f _ {i}\\right) = \\sum_ {j = 1} ^ {K} \\theta \\left(x _ {i}, x _ {i, j}\\right) \\cdot \\gamma \\left(f _ {i, j}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 498, + 482, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the equation, the function $\\theta$ is used to compute the relationship between a neighbor point and the center point. The function $\\gamma (f_{i,j})$ represents the linear transformation of the neighborhood feature, expressed as $\\gamma (f_{i,j}) = W_{\\gamma}f_{i,j} + b_{\\gamma}$ , where $W_{\\gamma}$ and $b_{\\gamma}$ are learnable parameters.", + "bbox": [ + 89, + 553, + 483, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The function $\\theta$ is computed using the dot product similarity in the attention mechanism:", + "bbox": [ + 89, + 630, + 482, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta (x _ {i}, x _ {i, j}) = \\operatorname {s o f t m a x} (q (x _ {i}) \\cdot k (x _ {i, j}) ^ {T} / \\sqrt {C _ {q}}) \\quad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 672, + 482, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $q(x_{i})$ and $k(x_{i,j})$ represent the linear transformations of the center point $x_{i}$ , and the neighborhood points $\\mathcal{N}(x_i)$ , which can be implemented using a $1 \\times 1$ convolution kernel. The input to $q(x_{i})$ is the real coordinates of the center point, while the input to $k(x_{i,j})$ is the relative coordinates between the neighborhood point and the center point. The convolution kernel has an output channel of $C_q$ and an input channel of 3. Using real and relative coordinates, the model effectively combines global and local spatial information, enhancing its ability to perceive the 3D spatial geometry.", + "bbox": [ + 89, + 703, + 482, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then use the computed $w(f_{i})$ to refine the initial points and their features. This operation selects points with", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "more surrounding noise points as removing points, resulting in more accurate features. In other words, we can find the suitable refined point $x_{i}^{*}$ and its corresponding feature $f_{i}^{*}$ , expressed as:", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i} ^ {*} = \\delta_ {p} \\left(w \\left(f _ {i}\\right)\\right) x _ {i}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 162, + 906, + 179 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {*} = \\delta_ {f} \\left(w \\left(f _ {i}\\right)\\right) f _ {i}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 181, + 906, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $\\delta_p$ and $\\delta_f$ is activation function. Finally, we fuse the points $x_i^*$ and their corresponding features $f_i^*$ through an attention module to obtain the module's final output features $f_{out}$ .", + "bbox": [ + 511, + 205, + 905, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. TreeTrans(TT) Module", + "text_level": 1, + "bbox": [ + 511, + 276, + 728, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the decoding process of encoded point clouds, previous methods often begin with three rounds of interpolation, followed by feature fusion using neighboring point clouds [26], and some methods adopt point-adaptive deconvolution techniques, such as PDR [20]. These methods are typically built on UNet architecture [31], which leads to several issues: (1) The model becomes overly dependent on global features, thereby neglecting local features, and (2) The utilization of multi-scale features is not comprehensive enough, leading to suboptimal performance when handling complex point clouds.", + "bbox": [ + 511, + 297, + 905, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose the TreeTrans (TT) module as our decoding network to address the above issues. We first employ two operations similar to TreeGCN[32]: the cyclic and the ancestor operations, which propagate features from the root node to the leaf nodes. After obtaining the intermediate features $f_{tree}$ , we combine $f_{tree}$ with the corresponding features from the encoder using local and global attention mechanisms to obtain the output $f_{out}$ .", + "bbox": [ + 511, + 464, + 905, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In traditional graph convolutional networks, the cyclic operation uses a separate weight to transform the current node feature $f_{i}^{l}$ to the next layer. However, a single weight often fails to capture all the relevant information in point clouds. Thus, we employ a $K$ -support-based cyclic operation to refine the feature representation progressively:", + "bbox": [ + 511, + 585, + 905, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {i} ^ {l + 1} = F _ {K} ^ {l} \\left(f _ {i} ^ {l}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 684, + 903, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $F_{K}^{l}$ is a fully connected layer containing $K$ nodes, and $S_{i}^{l + 1}$ represents the cyclic operation at layer $l$ . The cyclic operation is then combined with the ancestor operation to update the node features, ensuring that each layer incorporates information from previous layers during computation. The updated node features can be written as:", + "bbox": [ + 511, + 712, + 905, + 804 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t r e e} = \\sigma \\left(S _ {i} ^ {l + 1} + \\sum_ {m _ {j} \\in A \\left(f _ {i} ^ {l}\\right)} U _ {j} ^ {l} m _ {j}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 811, + 903, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $\\sigma$ is the activation function, with ReLU used in this paper. $A(f_i^l)$ denotes the set of all ancestor nodes, and $U_j^l$ is", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16990", + "bbox": [ + 480, + 989, + 517, + 1000 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ba9457d5dc27e8fb2895c6d890b0921ac41a4268e2c56dd99122196ebbe1e684.jpg", + "image_caption": [ + "Figure 4. Overview of PDANS. The CP network (upper branch) and DP network (lower branch) both consist of an encoder made up of feature extraction and the ANS module, with the TT module serving as the decoder. They exchange information through a mutual attention module. the upsampled ratio and interpolated points are used as conditions to provide prior knowledge of the DP network." + ], + "image_footnote": [], + "bbox": [ + 94, + 94, + 911, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the linear transformation weight from ancestor nodes to the current node.", + "bbox": [ + 89, + 482, + 482, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining $f_{tree}$ , we input it into the local attention and global attention modules. In the local attention mechanism, the corresponding encoder output features $f^{l+1}$ are combined with $f_{tree}$ using local attention, as described by the following formula:", + "bbox": [ + 89, + 512, + 483, + 588 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {l o c a l}} = \\operatorname {s o f t m a x} \\left(\\frac {Q \\left(f ^ {l + 1}\\right) \\cdot K \\left(f _ {\\text {t r e e}}\\right) ^ {T}}{\\sqrt {d}}\\right) \\cdot V \\left(f _ {\\text {t r e e}}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 597, + 482, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$Q(f^{l + 1})$ and $K(f_{tree})$ represent two convolution operations that map $f^{l + 1}$ and $f_{tree}$ from their respective dimensions to $d$ . Then, the attention weights are obtained by calculating the dot product similarity followed by a softmax operation. Afterward, the weighted sum with $V(f_{tree})$ gives the local attention feature $f_{local}$ . Once the local attention is computed, the resulting $f_{local}$ undergoes a global attention mechanism to produce the final output $f_{out}$ of the TT module. The specific implementation is as follows:", + "bbox": [ + 89, + 643, + 483, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {o u t} = \\operatorname {s o f t m a x} \\left(\\frac {q (f _ {l o c a l}) \\cdot k (f _ {l o c a l}) ^ {T}}{\\sqrt {d}}\\right) \\cdot v (f _ {l o c a l}) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 789, + 483, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Network Architecture", + "text_level": 1, + "bbox": [ + 89, + 832, + 295, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sections 4.1 and 4.2, we introduced the network's encoding module, ANS, and decoding module, TT, respectively. In this section, we present the overall framework of", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PDANS, which consists of two main components: the Conditional Point Cloud (CP) network and the Diffusion Point Cloud (DP) network. Figure 4 provides a detailed illustration of this process.", + "bbox": [ + 511, + 482, + 906, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CP-Net Our encoder consists of a feature extraction network and the ANS module. The decoder comprises the TT module. Both downsampling and upsampling are performed four times. Before each sampling output, we embed the time step $t$ through an MLP: $t$ is encoded into a $C_t$ -dimensional vector space. The output feature $f_{out} \\in \\mathbb{R}^{N \\times D_l}$ is then transformed through MLPs into the dimension $f_{out}' \\in \\mathbb{R}^{N \\times C_t}$ . We then add the two to integrate the time step information into the feature, resulting in the output $f_{l+1} \\in \\mathbb{R}^{N \\times D_{l+1}}$ . Additionally, we only input the sparse point cloud into the CP-Net to ensure the extraction of effective features.", + "bbox": [ + 511, + 547, + 908, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DP-Net The DP-Net shares the same architecture as the CP-Net but requires additional conditional information to guide the diffusion process. First, we convert the sparse point cloud $\\mathcal{P} \\in \\mathbb{R}^{N \\times 3}$ into an interpolated point cloud $\\mathcal{P}^* \\in \\mathbb{R}^{rN \\times 3}$ using upsampling by a factor of $r$ . Then, we sum $\\mathcal{P}^*$ with the noisy point cloud $\\mathcal{P}_T \\in \\mathbb{R}^{rN \\times 3}$ and input them together into the DP-Net. Meanwhile, we extract global features from $\\mathcal{P}^*$ and encode both the time step $t$ and the upsampling factor $r$ . These pieces of information are incorporated as conditions at each stage of the encoder and decoder.", + "bbox": [ + 511, + 734, + 908, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16991", + "bbox": [ + 480, + 989, + 517, + 1000 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The information provided by the CP-Net is further guided to the DP-Net through a cross-attention (CA) module. Given the outputs of the encoder at each layer for the CP-Net and DP-Net: $F_{c} \\in \\mathbb{R}^{N_{c} \\times D_{c}}$ and $F_{d} \\in \\mathbb{R}^{N_{d} \\times D_{d}}$ , the CA module first reshapes the two feature sets using MLPs: $F_{c} \\rightarrow (Q) \\in \\mathbb{R}^{N_{c} \\times C_{i}}$ and $F_{d} \\rightarrow (K, V) \\in \\mathbb{R}^{N_{d} \\times C_{i}}$ . Then, the features are fused according to the following formula:", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {f} = M L P \\left(\\text {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {C _ {i}}}\\right) V\\right) + F _ {c} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 205, + 482, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, $F_{f}$ is input into the feedforward network to produce the final features. The same method can also be applied reversely, allowing data to flow between the two networks.", + "bbox": [ + 89, + 250, + 483, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Loss", + "text_level": 1, + "bbox": [ + 89, + 318, + 165, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the DP-Net, its loss can be referred to by equation 5. For the CP-Net, we aim for it to learn more features of the sparse point cloud. Therefore, we add MLPs at the end of the decoder to output a sparse point cloud $\\mathcal{P}'$ , which is then supervised by MSE as follows:", + "bbox": [ + 89, + 340, + 483, + 416 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} \\left(\\mathcal {P}, \\mathcal {P} ^ {\\prime}\\right) = \\mathbf {M S E} \\| \\mathcal {P} - \\mathcal {P} ^ {\\prime} \\| ^ {2} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 426, + 482, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Combining these loss functions, The training objective is:", + "bbox": [ + 89, + 455, + 472, + 470 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} (\\theta) + \\alpha \\mathcal {L} _ {C} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 481, + 482, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ means a weighting factor ( $\\alpha = 0.5$ in this paper).", + "bbox": [ + 89, + 508, + 470, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 536, + 223, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 561, + 307, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment Setup. All PDANS experiments are run in a PyTorch[24] environment. The PDANS model is trained on a single Nvidia 4080 GPU using the Adam optimizer[13], with a batch size 12 for 350 epochs. The channel sizes for the CP-Net and DP-Net are set to (64, 128, 256, 512) and (128, 256, 256, 512), respectively.", + "bbox": [ + 89, + 583, + 482, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. In our experiments, we utilize two public benchmarks (PUGAN[15], PU1K[27]) for evaluation. We adhere to the official training/testing partitioning protocols for these datasets. Poisson disk sampling [46] is used to generate 24,000 patches on PUGAN and 69,000 patches on PU1K for training. Each patch contains 256 points, with the corresponding ground truth containing 1,024 points. Additionally, 27 and 127 point cloud datasets are used for testing, respectively. The sparse point clouds contain 2,048 points, and the number of points in ground truth is set to $2,048 \\times r$ according to the upsampling factor $r$ .", + "bbox": [ + 89, + 674, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics Following previous methods, we select three metrics to evaluate our model: Chamfer Distance $(\\mathrm{CD} \\times 10^{-3})$ [6], Hausdorff distance $(\\mathrm{HD} \\times 10^{-3})$ [2] and Point to Surface Distance $(\\mathrm{P2F} \\times 10^{-3})$ [45].", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation on PUGAN Dataset", + "text_level": 1, + "bbox": [ + 511, + 90, + 787, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform sampling on PUGAN at low $(4\\times)$ and high $(16\\times)$ upsampling rates. The reductions in the three metrics shown in Table 1 demonstrate that our model has a significant advantage in describing geometric details. The $16\\times$ upsampled point cloud shown in Figure 5 further supports our claim, as the generated point clouds have fewer outliers and are closer to the ground truth geometric features.", + "bbox": [ + 511, + 112, + 906, + 219 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a4cb4da4cde9acbcb678177000946036ffd874af0246194043e22b74e4e54c4a.jpg", + "table_caption": [ + "Table 1. The results of $4 \\times$ and $16 \\times$ on PUGAN. Our method significantly surpasses other methods in terms of CD and HD." + ], + "table_footnote": [], + "table_body": "
Methods4×16×
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]0.5296.8054.4600.5108.2066.041
MPU[44]0.2926.6722.8220.2197.0543.085
PU-GAN[15]0.2825.5772.0160.2076.9632.556
PU-GCN[27]0.2683.2012.4890.1614.2832.632
Grad-PU[8]0.2452.3691.8930.1082.3522.127
RepKpU[30]0.2542.6981.8820.1073.2931.957
PUDM[29]0.1311.2201.9120.0821.1202.114
Ours0.1191.20651.8970.0781.0462.028
", + "bbox": [ + 524, + 270, + 895, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Evaluation on PU1K Dataset", + "text_level": 1, + "bbox": [ + 511, + 438, + 769, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows the results of $4\\mathrm{x}$ upsampling on the PU1K dataset. Compared to other methods, we demonstrate superior performance in both CD and HD metrics. The PU1K dataset is more complex and diverse than others, and our designed TT module exhibits tremendous advantages in handling such complex data. It helps approximate the valid geometric details more effectively, achieving higher accuracy.", + "bbox": [ + 511, + 459, + 906, + 566 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fc0a27375a169f06c6a61ae491f89d1b91928ec4c7102d00fa2773096c1c3ad0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCD↓HD↓P2F↓
PU-Net[45]1.15515.1704.834
MPU[44]0.93513.3273.511
PU-GAN[15]0.7359.9394.234
PU-GCN[27]0.5857.5772.499
Grad-PU[8]0.4043.7321.474
RepKpU[30]0.3312.7001.261
PUDM[29]0.2172.1641.477
Ours0.1551.5881.354
", + "bbox": [ + 531, + 590, + 888, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The results of $4 \\times$ on PU1K. Our method outperforms other methods across nearly all metrics.", + "bbox": [ + 511, + 715, + 906, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Robustness Test", + "text_level": 1, + "bbox": [ + 511, + 772, + 671, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment setup. Based on the testing meshes provided by PUGAN, we construct a $4 \\times$ upsampling dataset to conduct a comprehensive robustness test. Ground-truth point clouds are sampled from testing meshes via Poisson disk sampling. We tested three different types of noise: Gaussian, random, and Laplacian. These three types of noise are the most common in real-world point clouds, making", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16992", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7d2686eb0d5d4fbae318fb58e35b1ad43b1825fd43872fc998f86357d9bf504e.jpg", + "image_caption": [ + "Figure 5. Comparison $(16\\times)$ on PUGAN dataset." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 897, + 255 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the results more practical. We set four different noise levels for Gaussian noise: $\\tau = 0.01$ , $\\tau = 0.02$ , $\\tau = 0.05$ , and $\\tau = 0.1$ . We used two noise levels for Laplacian and random noise: $\\tau = 0.05$ and $\\tau = 0.1$ . For the input with 2048 points, we tested the results under each noise type and noise level.", + "bbox": [ + 89, + 306, + 483, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results of Gaussian noise Tables 3 and 4 present the quantitative results of Gaussian noise testing under different noise levels. As shown in the tables, our method significantly outperforms other models in terms of resistance to Gaussian noise. Thanks to the design of the ANS module in our model, which further enhances noise resistance. Therefore, our approach excels at noise suppression and surpasses other methods in both CD and HD metrics.", + "bbox": [ + 89, + 398, + 483, + 518 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a74324edbe9343ecc5bd97b1607931465f1580b19641f1ba6d0407c8730e4657.jpg", + "table_caption": [ + "Table 3. The results of $4 \\times$ at low-level $(\\tau = 0.01$ and $\\tau = 0.02)$ Gaussian noise on PUGAN. Our method significantly outperforms other methods in terms of Gaussian noise." + ], + "table_footnote": [], + "table_body": "
Methodsτ = 0.01τ = 0.02
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]0.6288.0689.8161.07810.86716.401
MPU[44]0.5066.9789.0590.92910.82015.621
PU-GAN[15]0.4646.0707.4980.88710.60215.088
PU-GCN[27]0.4485.5866.9890.8168.60413.798
Grad-PU[8]0.4144.1456.4000.7667.33611.534
RepKpU[30]0.3844.9376.7750.8008.41514.124
PUDM[29]0.2102.4306.0700.5295.4719.742
Ours0.1851.9185.9640.3093.0209.513
", + "bbox": [ + 102, + 584, + 470, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results of random noise and Laplacian noise Tables 5 and 6 present the quantitative test results under noise levels $\\tau = 0.05$ and $\\tau = 0.1$ . The results clearly show that the PDANS network we designed exhibits superior adaptability in high-noise environments. Figure 6 illustrates that, in noisy point clouds, our results show more complete structures with fewer outliers. Our method captures more geometric details, leading to higher-quality point clouds. In contrast, other methods tend to upsample blindly, resulting in point clouds with disorganized structures. While our method still maintains SOTA performance, it shows slightly", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b6460960edf226353a4ff0f523bcc414f5b64e9e328a0bb38115626fe0126f77.jpg", + "table_caption": [ + "Table 4. The results of $4 \\times$ at high-level $(\\tau = 0.05$ and $\\tau = 0.1)$ Gaussian noise on PUGAN. Our method significantly outperforms other methods in terms of Gaussian noise." + ], + "table_footnote": [], + "table_body": "
Methodsτ = 0.05τ = 0.1
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.37013.72923.2491.49814.19323.846
MPU[44]1.24711.64522.1891.32112.41523.841
PU-GAN[15]1.1249.09121.2521.27110.91123.174
PU-GCN[27]1.2639.86922.8351.45611.06325.213
Grad-PU[8]0.9788.05716.9271.1188.94618.845
RepKpU[30]1.23712.33520.5181.36012.90222.473
PUDM[29]0.6185.38614.7510.8536.23916.845
Ours0.4764.03814.6070.5434.32716.476
", + "bbox": [ + 524, + 357, + 895, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lower results than Gaussian noise. The experimental results validate the effectiveness of the ANS module we designed. By removing noisy point clouds, we can better learn the objects' structural information, thereby generating high-quality point clouds. As a result, PDANS demonstrates good robustness under different types and levels of noise, proving its stability in complex noisy environments.", + "bbox": [ + 511, + 508, + 906, + 616 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7e0a99a0d26b3cd6705d7befbae2a5a44cf370ddd9cef38571cf1da929bc8789.jpg", + "table_caption": [ + "Table 5. The results of $4 \\times$ at high-level ( $\\tau = 0.05$ and $\\tau = 0.1$ ) random noise on PUGAN." + ], + "table_footnote": [], + "table_body": "
Methodsτ = 0.05τ = 0.1
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.49014.47323.2231.72515.44225.251
MPU[44]1.22410.84220.4561.54511.64523.512
PU-GAN[15]1.0347.75718.6171.3279.70021.321
PU-GCN[27]1.0459.64318.8991.32510.87721.633
Grad-PU[8]1.0676.63417.7341.3997.21521.028
RepKpU[30]0.7894.78217.8130.8373.90220.953
PUDM[29]0.9986.11017.5581.3106.73220.564
Ours0.7223.16117.6340.6824.59819.401
", + "bbox": [ + 524, + 667, + 895, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Result on Real Dataset", + "text_level": 1, + "bbox": [ + 511, + 818, + 723, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted experiments on the ScanObjectNN [37] dataset, where point clouds are real-scanned, non-uniform and noisy. Since there is no ground truth point cloud, we performed a visual comparison in Fig. 7. The experiment", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16993", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99024c5e76bd76188745f79fb386841e6f430fbdbc4d3fdb375431b3c24a1d94.jpg", + "image_caption": [ + "Figure 6. Visualization results at $4\\times$ on PUGAN with random noise and Laplace noise." + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 890, + 220 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7897433a5a7590e0c88cbe0aa05c932eeed8bcfe7d862966e91814b0e2cfb46f.jpg", + "table_caption": [ + "Table 6. The results of $4 \\times$ at high-level $(\\tau = 0.05$ and $\\tau = 0.1)$ Laplace noise on PUGAN." + ], + "table_footnote": [], + "table_body": "
Methodsτ = 0.1τ = 0.05
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.94332.64626.9651.86127.56324.737
MPU[44]1.56920.32925.1641.42119.04922.783
PU-GAN[15]1.48614.91924.1481.33414.24321.879
PU-GCN[27]1.58715.71125.1341.42514.91522.876
Grad-PU[8]1.23310.34818.5761.0659.34916.591
RepKpU[30]1.34513.10422.2951.19412.37720.218
PUDM[29]0.7046.28418.2830.6437.92416.871
Ours0.5634.05116.2870.5334.43316.195
", + "bbox": [ + 102, + 316, + 470, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "used an input with 2048 points, followed by a $16\\times$ upsampling. The Ball-pivoting algorithm [3] was then applied to reconstruct the surface from the upsampled point cloud. Based on the reconstruction results, our generated point clouds exhibit fewer gaps and smoother surfaces, which benefits surface reconstruction.", + "bbox": [ + 89, + 467, + 483, + 556 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f139f4ab33bd845ac6d7800d3687d1a12602956b02768c63f1bff37a25962eb3.jpg", + "image_caption": [ + "Figure 7. Visual evaluation of impacts on surface construction on real-scanned ScanObjectNN dataset." + ], + "image_footnote": [], + "bbox": [ + 99, + 571, + 475, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 742, + 243, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With/Without ANS Module To demonstrate the effectiveness of the ANS module, we compared it with other network architectures. Model A replaces the ANS module with an encoder composed of Dense Blocks [8], Model B uses an encoder based on Set Abstraction [29], and Model C employs an encoder from PointNet++ [26]. Our network is referred to as Model D. We tested these models on the PU-GAN dataset with Gaussian noise level $\\tau = 0.1$ , with results presented in Table7.", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8494a5b2c903979e7a52d81b2e8ab67586ad65c227e730ab5e734518e085fdb7.jpg", + "table_caption": [ + "Table 7. The results of $4 \\times$ at high-level $(\\tau = 0.1)$ Gaussian noise on PUGAN with different encoders." + ], + "table_footnote": [], + "table_body": "
ModuleCD↓HD↓P2F↓
A1.2489.63119.657
B0.8536.34816.849
C1.16814.49823.189
D0.5434.32716.476
", + "bbox": [ + 540, + 316, + 879, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With/Without TT Module We conducted $16 \\times$ upsampling tests on the PUGAN dataset, as shown in Table 8. The experiments indicate a significant performance drop when the TT module is removed, underscoring the importance of the TT module in enhancing the quality of the model generation.", + "bbox": [ + 511, + 409, + 906, + 498 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cb217a598d01c4abb5917eb25e16cab64baf4df22f34383727745ccf0dc8e59b.jpg", + "table_caption": [ + "Table 8. TT Module Ablation Study" + ], + "table_footnote": [], + "table_body": "
MethodsCD↓HD↓P2F↓
Without TT Module0.0941.1242.158
With TT Module0.0781.0462.028
", + "bbox": [ + 519, + 535, + 901, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 614, + 633, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel conditional diffusion model-based network, PDANS, designed for point cloud upsampling. Suppressing noise and ensuring the geometric authenticity of generated results have long been primary research foci in point cloud upsampling. To address these issues, we have designed the ANS and TT modules. The ANS module integrates point cloud features with neighborhood information, enabling adaptive noise removal from the point cloud. Meanwhile, the TT module effectively captures local and global geometric details, allowing for a more accurate reconstruction of geometries close to real values. Extensive experiments validate that the proposed PDANS outperforms existing state-of-the-art methods in performance.", + "bbox": [ + 511, + 638, + 906, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 513, + 847, + 609, + 863 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Tenindra Abeywickrama, Muhammad Aamir Cheema, and David Taniar. K-nearest neighbors on road networks: a", + "bbox": [ + 519, + 872, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16994", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "journey in experimentation and in-memory implementation. arXiv preprint arXiv:1601.01549, 2016.3", + "[2] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Transactions on Graphics (TOG), 32(2):1-17, 2013. 6", + "[3] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 8", + "[4] Haolan Chen, Shitong Luo, Wei Hu, et al. Deep point set resampling via gradient fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):2913-2930, 2022. 1", + "[5] Yaodong Cui, Ren Chen, Wenbo Chu, Long Chen, Daxin Tian, Ying Li, and Dongpu Cao. Deep learning for image and point cloud fusion in autonomous driving: A review. IEEE Transactions on Intelligent Transportation Systems, 23(2): 722-739, 2021. 1", + "[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 6", + "[7] Wanquan Feng, Jin Li, Hongrui Cai, Xiaonan Luo, and Juyong Zhang. Neural points: Point cloud representation with neural fields for arbitrary upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18642, 2022. 2", + "[8] Yun He, Danhang Tang, Yinda Zhang, Xiangyang Xue, and Yanwei Fu. Grad-pu: Arbitrary-scale point cloud upsampling via gradient descent with learned distance functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2023. 1, 2, 6, 7, 8", + "[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2", + "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3", + "[11] F. Y. Hu, C. L. Li, T. Zhou, H. F. Cheng, and M. M. Gu. A survey on point cloud completion algorithms for deep learning. Journal of Image and Graphics, 30(02):0309-0333, 2025. 2", + "[12] Sheng Yu Huang, Hao-Yu Hsu, and Frank Wang. Spovt: Semantic-prototype variational transformer for dense point cloud semantic completion. Advances in Neural Information Processing Systems, 35:33934–33946, 2022. 2", + "[13] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[14] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2", + "[15] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7203-7212, 2019. 1, 2, 6, 7, 8" + ], + "bbox": [ + 93, + 92, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Ruihui Li, Xianzhi Li, Pheng-Ann Heng, and Chi-Wing Fu. Point cloud upsampling via disentangled refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 344–353, 2021. 2", + "[17] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3d object reconstruction. In proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1", + "[18] Kangcheng Liu, Yuzhi Zhao, Zhi Gao, and Ben M Chen. Weaklabel3d-net: A complete framework for real-scene lidar point clouds weakly supervised multi-tasks understanding. In 2022 international conference on robotics and automation (ICRA), pages 5108-5115. IEEE, 2022. 1", + "[19] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2837-2845, 2021. 2", + "[20] Zhaoyang Lyu, Zhifeng Kong, Xudong Xu, Liang Pan, and Dahua Lin. A conditional point diffusion-refinement paradigm for 3d point cloud completion. arXiv preprint arXiv:2112.03530, 2021. 2, 3, 4", + "[21] Aihua Mao, Zihui Du, Junhui Hou, Yaqi Duan, Yong-jin Liu, and Ying He. Pu-flow: A point cloud upsampling network with normalizing flows. IEEE Transactions on Visualization and Computer Graphics, 29(12):4964-4977, 2022. 2", + "[22] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12923-12932, 2023. 1", + "[23] Lucas Nunes, Rodrigo Marcuzzi, Benedikt Mersch, Jens Behley, and Cyril Stachniss. Scaling diffusion models to real-world 3d lidar scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14770-14780, 2024. 2", + "[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2", + "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4, 8", + "[27] Guocheng Qian, Abdulellah Abualshour, Guohao Li, Ali Thabet, and Bernard Ghanem. Pu-gcn: Point cloud upsampling using graph convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11683-11692, 2021. 1, 2, 6, 7, 8", + "[28] Shi Qiu, Saeed Anwar, and Nick Barnes. Pu-transformer: Point cloud upsampling transformer. In Proceedings of the" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16995", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Asian conference on computer vision, pages 2475-2493, 2022. 1", + "[29] Wentao Qu, Yuantian Shao, Lingwu Meng, Xiaoshui Huang, and Liang Xiao. A conditional denoising diffusion probabilistic model for point cloud upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20786-20795, 2024. 2, 6, 7, 8", + "[30] Yi Rong, Haoran Zhou, Kang Xia, Cheng Mei, Jiahao Wang, and Tong Lu. Repkpu: Point cloud upsampling with kernel point representation and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21050-21060, 2024. 1, 2, 6, 7, 8", + "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4", + "[32] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 2, 4", + "[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3", + "[34] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2", + "[35] Hanxiao Tan. Dam: Diffusion activation maximization for 3d global explanations. arXiv preprint arXiv:2401.14938, 2024. 2", + "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2", + "[37] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 7", + "[38] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. Advances in Neural Information Processing Systems, 34:29088-29100, 2021. 2", + "[39] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5499-5509, 2021. 1", + "[40] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5589-5598, 2020. 1", + "[41] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3323–3332, 2019. 3", + "[42] Lei Yang, Yanhong Liu, Jinzhu Peng, and Zize Liang. A novel system for off-line 3d seam extraction and path planning based on point cloud segmentation for arc welding robot. Robotics and Computer-Integrated Manufacturing, 64:101929, 2020. 1", + "[43] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 3", + "[44] Wang Yifan, Shihao Wu, Hui Huang, Daniel Cohen-Or, and Olga Sorkine-Hornung. Patch-based progressive 3d point set upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5958–5967, 2019. 2, 6, 7, 8", + "[45] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2790–2799, 2018. 1, 2, 6, 7, 8", + "[46] Cem Yuksel. Sample elimination for generating poisson disk sample sets. In Computer Graphics Forum, pages 25-32. Wiley Online Library, 2015. 6", + "[47] Yuchao Zheng, Yujie Li, Shuo Yang, and Huimin Lu. Global-pbnet: A novel point cloud registration for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(11):22312-22319, 2022. 1" + ], + "bbox": [ + 516, + 92, + 903, + 556 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16996", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/96fbf65a-6eae-49b1-9bdc-b206a300f6d4_model.json b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/96fbf65a-6eae-49b1-9bdc-b206a300f6d4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dfd2852067e3ed18412666a191f5f2b1c943d09c --- /dev/null +++ b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/96fbf65a-6eae-49b1-9bdc-b206a300f6d4_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.13, + 0.885, + 0.178 + ], + "angle": 0, + "content": "Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.203, + 0.796, + 0.24 + ], + "angle": 0, + "content": "Boqian Zhang\\*, Shen Yang\\*, Hao Chen, Chao Yang, Jing Jia, Guang Jiang† \nXidian University" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.242, + 0.807, + 0.257 + ], + "angle": 0, + "content": "{24011211294，20009101459，haochen(cv，24011211295}@xidian.stu.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.26, + 0.719, + 0.275 + ], + "angle": 0, + "content": "jjia@mail.xidian.edu.cn, gjiang@mail.xidian.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.328, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.485, + 0.721 + ], + "angle": 0, + "content": "Point cloud upsampling can improve the quality of the initial point cloud, significantly enhancing the performance of downstream tasks such as classification and segmentation. Existing methods mostly focus on generating the geometric details of point clouds, neglecting noise suppression. To address this, we propose a novel network based on a conditional diffusion model, incorporating the Adaptive Noise Suppression (ANS) module, which we refer to as PDANS. The ANS module assigns weights to each point and determines the removal strategy based on these weights, reducing the impact of noisy points on the sampling process. The module first selects the neighborhood set for each point in the point cloud and performs a weighted sum between the point and its neighbors. It then adjusts the removal points based on the weighted sum, effectively mitigating the bias caused by outliers. We introduce the TreeTrans (TT) module to capture more correlated feature information. This module learns the interaction between high-level and low-level features, resulting in a more comprehensive and refined feature representation. Our results on several widely used benchmark datasets demonstrate that PDANS exhibits exceptional robustness in noisy point cloud processing and outperforms current state-of-the-art(SOTA) methods in terms of performance. Code is available at https://github.com/Baty2023/PDANS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.224, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Point clouds, as one of the key representations for 3D reconstruction [17] [22], are widely applied in autonomous driving [5] [47], embodied intelligence [18] [42], and AR/VR. However, real-world scanned point clouds are often sparse, noisy, and uneven, negatively impacting their performance in downstream tasks such as classification and segmenta" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.372 + ], + "angle": 0, + "content": "tion. Point cloud upsampling can transform sparse and noisy point clouds into dense, artifact-free representations. In recent years, significant progress has been made in this task [15] [27] [45] [40]." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.386, + 0.898, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Figure 1. Most existing methods achieve satisfactory results when input without noise, but they often struggle to effectively recover the geometric structure (e.g., panda ears) when dealing with input with noise." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.584, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Inspired by deep learning, several methods have adopted deep networks to perform this task. Existing upsampling methods can be broadly classified into two categories: those based on generative models and those based on refinement methods. Generative models [15] [27] [28] first extract features from sparse point clouds into a latent space and then design a generator based on the extracted features to generate the coordinates of dense point clouds. The limitation of such methods lies in the difficulty of accurately predicting 3D coordinates, leading to potential outliers or artifacts in the generated dense point clouds [8] [45]. Refinement-based methods [4] [39] [30] first use interpolation to generate a coarse dense point cloud, and then employ a neural network to extract geometric features from the coarse point cloud for refinement. Combined with reconstruction loss functions, this process refines the point cloud to match the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "These methods share two common issues: (1) The features lack significant dominance or distinguishability, leading to generating point clouds without clear structure, which tend to be sparse and scattered. (2) Most of these methods" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.371, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.248, + 0.9 + ], + "angle": 0, + "content": "†Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.371, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16987" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "rely on loss functions such as CD loss, EMD loss, and other similar metrics, which are highly sensitive to outliers. This sensitivity makes the model ineffective at handling noisy point clouds [12] [38], as shown in Figure1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.351 + ], + "angle": 0, + "content": "To address the above issues, we propose a point cloud upsampling method with adaptive noise suppression, PDANS. The network is designed based on a conditional diffusion model, leveraging the inherent denoising properties of the diffusion model's generative process to enhance its adaptability to noisy point clouds. The network also integrates an Adaptive Noise Suppression (ANS) module, which we designed to compute downsampling weights for each point by utilizing the relationships between the point and its neighboring points. This approach enables the extraction of more accurate downsampled points and features. It effectively reduces the impact of noisy points during feature extraction and minimizes the bias caused by outliers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.474 + ], + "angle": 0, + "content": "To enhance feature representation, we propose a structure based on graph convolutional networks [14], called the TreeTrans (TT) module. This module incorporates ancestral and recurrent terms[32] to fully utilize the features outputted by each encoder layer. Additionally, we apply both local and global attention mechanisms to the output features, further improving the efficiency and expressiveness of the feature representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Additionally, we designed a conditional point cloud network and a diffusion point cloud network. The conditional point cloud network provides relevant features of sparse point clouds to the diffusion point cloud network through a mutual attention module, guiding the diffusion network to generate the required noise for point cloud densification. Similar to PUDM[29], we incorporate the diffusion model, thereby eliminating the reliance on CD loss. Combined with the ANS module we designed, PDANS demonstrates exceptional robustness in noise experiments. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.645, + 0.483, + 0.703 + ], + "angle": 0, + "content": "- We designed a noise suppression module named ANS, which adaptively adjusts removal points and their neighborhood features to improve robustness against noise and outliers." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.705, + 0.483, + 0.78 + ], + "angle": 0, + "content": "- The proposed TT module fully leverages features across different layers of the encoder, incorporating ancestral terms, recurrent terms, and local and global attention mechanisms to achieve more comprehensive and refined feature representations." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.781, + 0.483, + 0.855 + ], + "angle": 0, + "content": "- We designed an end-to-end network architecture named PDANS, which utilizes a mutual attention mechanism and a diffusion model to guide the densification of sparse point clouds. The ANS module further enhances the network's ability to handle noise." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the outstanding performance of PDANS in point cloud upsampling benchmarks, particularly in generating geometric details and" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.788, + 0.107 + ], + "angle": 0, + "content": "handling noisy point clouds effectively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.122, + 0.655, + 0.138 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.148, + 0.857, + 0.166 + ], + "angle": 0, + "content": "2.1. Learning-based point cloud upsampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.17, + 0.905, + 0.382 + ], + "angle": 0, + "content": "Several methods [25] [26] [36] [11] shave been developed to extract features from 3D data directly with the strong data-driven and representational capabilities of deep learning networks. PU-Net [45] was the first learning-based method for point cloud and upsampling. Following its proposed architecture, many methods have emerged: MPU [44], PUGAN [15], Dis-PU [16], PUGCN [27], and RepKPU [30]. The NePs [7] method enhances neural radiance fields by combining 2D information with 3D features, enabling point cloud upsampling. Some networks attempt to perform point cloud upsampling based on denoising paradigms. They generate a rough upsampled point cloud through interpolation, and then refine it using neural networks, such as GradPU [8] and PU-Flow [21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.905, + 0.519 + ], + "angle": 0, + "content": "Most methods predict the coordinates of dense point clouds from sparse point cloud features and then rely on upsampling modules to expand the point cloud, or further refine the coordinates from the coarsely densified point clouds. Since these methods typically depend on CD (Chamfer Distance) and EMD (Earth Mover's Distance) as loss functions, they exhibit high sensitivity to noise. In this paper, we design an additional denoising module and utilize conditional DDPM to reduce the dependence on CD." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.531, + 0.749, + 0.545 + ], + "angle": 0, + "content": "2.2. DDPM for 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.905, + 0.735 + ], + "angle": 0, + "content": "Inspired by the success of image generation tasks [10] [34] [9], direct point cloud generation using DDPM has gained increasing attention. Luo et al. [19] is the first to apply DDPM to unconditional point cloud generation. The DAM [35] network explores the interpretability of diffusion models for point cloud generation tasks. Nunes et al. [23] apply DDPM to the 3D-LiDAR point cloud completion task in real-world scenarios. The PDR [20] network takes the initial point cloud as input, but it introduces a coarse and refinement network, which increases time overhead. PUDM [29] densifies point clouds through conditional diffusion, but its performance in noise suppression is suboptimal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.905, + 0.781 + ], + "angle": 0, + "content": "This paper uses a conditional diffusion model to perform point cloud upsampling. To enhance the model's denoising capability, we design an additional ANS module." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.797, + 0.905, + 0.83 + ], + "angle": 0, + "content": "3. Background of Conditional Diffusion Module" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We assume that the dense point cloud \\( x \\) is sampled from a meaningful distribution \\( P_{data} \\), and that the latent variable \\( x_{T} \\sim \\mathcal{N}(\\mathbf{0}_{3N},\\mathbf{I}_{3N\\times 3N}) \\) is sampled from a noise distribution \\( P_{latent} \\) which is easier to handle, where \\( \\mathcal{N} \\) follows a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16988" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "Gaussian distribution. DDPM[10] establishes a transfer relationship between \\( x \\) and \\( x_{T} \\) through two Markov chains. This is an autoregressive process: the forward process progressively adds noise to \\( x \\) until \\( x_{T} \\), and \\( x_{T} \\) is denoised back to \\( x \\) through the reverse process. We use the timestep \\( t \\sim U(T) \\) to control the speed of transformation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.484, + 0.228 + ], + "angle": 0, + "content": "Forward Process The process gradually adds noise to \\( x_0 \\) until \\( x_T \\) through the probability \\( q(x_t | x_{t-1}) \\). The entire process can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.239, + 0.483, + 0.281 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {1}, \\dots , x _ {T} \\mid x _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(x _ {t} \\mid x _ {t - 1}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.484, + 0.368 + ], + "angle": 0, + "content": "Where \\(q(x_{t}|x_{t - 1}) = \\mathcal{N}(x_{t};\\sqrt{1 - \\beta_{t}} x_{t - 1},\\beta_{t}I)\\), where the hyperparameter \\(\\beta_{t}\\) is a positive constant. Based on the derivation by Ho et al. [10], \\(x_{T}\\) can be computed directly from \\(x_0\\). Define \\(\\alpha_{t} = 1 - \\beta_{t}\\) and \\(\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}\\), so that \\(x_{T}\\) can be sampled from the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.378, + 0.483, + 0.397 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.407, + 0.408, + 0.421 + ], + "angle": 0, + "content": "Where \\(\\epsilon\\) follows a standard normal distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.528 + ], + "angle": 0, + "content": "Reverse Process The reverse process aims to predict and eliminate the noise added at each step \\( t \\). It is conditioned on a set of conditions, which may include incomplete point clouds \\( c \\), upsampling factor \\( r \\), etc. We denote the set of conditions as \\( C = \\{c_i \\mid i = 1, \\dots, N_c\\} \\), where \\( N_c \\) represents the number of conditions. We can define the reverse process from denoising \\( x_{T} \\) to \\( x_{0} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.539, + 0.483, + 0.58 + ], + "angle": 0, + "content": "\\[\np _ {\\boldsymbol {\\theta}} \\left(x _ {0}, \\dots , x _ {T - 1} \\mid x _ {T}, \\boldsymbol {C}\\right) = \\prod_ {t = 1} ^ {T} p _ {\\boldsymbol {\\theta}} \\left(x _ {t - 1} \\mid x _ {t}, \\boldsymbol {C}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.668 + ], + "angle": 0, + "content": "Where \\(p_{\\boldsymbol{\\theta}}(x_{t - 1}|x_t,\\boldsymbol {C}) = \\mathcal{N}(x_{t - 1};\\mu_{\\boldsymbol{\\theta}}(\\boldsymbol {x}_t,\\boldsymbol {C},t),\\sigma_t^2\\boldsymbol {I})\\) where the mean \\(\\mu_{\\boldsymbol{\\theta}}(x_t,C,t)\\) is the result predicted by a neural network, and \\(\\sigma_t^2\\) is a time-dependent constant. Based on the derivation by DDIM [33], the reverse process can be written in the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.678, + 0.483, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (x _ {t} , \\boldsymbol {C} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) + \\tag {4} \\\\ \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} (x _ {t}, C, t) + \\sigma_ {t} \\epsilon_ {t} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Where \\(\\epsilon_{t} \\sim \\mathcal{N}(\\mathbf{0}_{3N},\\mathbf{I}_{3N\\times 3N})\\) is standard Gaussian noise independent of \\(x_{T}\\). The noisy point cloud \\(x_{T}\\) is progressively denoised according to the above equation until it recovers \\(x_0\\). At each time step of the reverse process, conditions are added, allowing control over the generation of dense point clouds, as shown in the process diagram in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Training process under given conditions The goal is to optimize the denoising model's prediction of the noise \\(\\epsilon_{\\theta}\\) at step \\(t\\). Given a set of conditions \\(C\\) and under the" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.093, + 0.9, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.228, + 0.907, + 0.254 + ], + "angle": 0, + "content": "Figure 2. Visualization of the forward and reverse processes of PDANS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.907, + 0.309 + ], + "angle": 0, + "content": "condition that the time step \\(t\\sim U(T)\\), the training objective of DDPM is:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.318, + 0.916, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {\\boldsymbol {t} \\sim U (T), \\epsilon \\sim \\mathcal {N} (0, I)} \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon + \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, C, t\\right) \\| ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.426 + ], + "angle": 0, + "content": "In the above equation, there are no CD loss or EMD loss terms commonly used in traditional point cloud upsampling networks, because DDPM naturally defines a one-to-one pointwise correspondence between two consecutive point clouds, as shown in equation1." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.438, + 0.605, + 0.453 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.554 + ], + "angle": 0, + "content": "This section introduces each module's design and the network's overall structure. In PDANS, we propose two modules: the Adaptive Noise Suppression (ANS) module and the TreeTrans (TT) module. These modules are described in Sections 4.1 and 4.2, respectively. The design of the overall PDANS network is presented in Section 4.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.875, + 0.579 + ], + "angle": 0, + "content": "4.1. Adaptive Noise Suppression(ANS) Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.75 + ], + "angle": 0, + "content": "In the encoding stage, many frameworks perform down-sampling using Farthest Point Sampling (FPS), followed by selecting neighboring points within a certain radius [20] or via KNN [1] to enhance the features of the sampled points. However, two issues arise when the input point cloud contains significant noise: (1) FPS is sensitive to outliers[43][41], making it challenging to process noisy point clouds effectively. (2) FPS does not filter noisy points, leading to the inclusion of noise features in the enhanced features, which affects the accuracy and reliability of the sampling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose the Adaptive Noise Suppression (ANS) module to address the issues abovementioned, as shown in Figure 3. Unlike FPS, we first select the neighborhood point cloud \\(\\mathcal{P}_K\\in \\mathbb{R}^{N\\times 3\\times K}\\) for all points in the initial point cloud \\(\\mathcal{P}\\in \\mathbb{R}^{N\\times 3}\\), where \\(K\\) denotes the number of neighboring points for each point. These neighborhood point clouds contain both correct and noisy points. We then input the neighborhood point cloud, the initial point cloud, and their corresponding features into the ANS module to perform the removal operation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16989" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.473, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.293, + 0.483, + 0.322 + ], + "angle": 0, + "content": "Figure 3. Visualization of the Adaptive Noise Suppression (ANS) module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.488 + ], + "angle": 0, + "content": "For each point \\( x_{i} \\) selected from \\( \\mathcal{P} \\), we select its corresponding feature \\( f_{i} \\) from \\( \\mathcal{F}s \\in \\mathbb{R}^{N \\times D_l} \\), where \\( \\mathcal{F}s \\) is the feature set obtained by reordering the initial point cloud. We use KNN to select its \\( K \\) nearest neighbors \\( x_{i,1}, \\ldots, x_{i,K} \\in \\mathcal{N}(x_i) \\) and their corresponding features \\( f_{i,1}, \\ldots, f_{i,K} \\in \\mathcal{N}(f_i) \\), where \\( \\mathcal{N}(x_i) \\) represents the \\( K \\) nearest neighbors of point \\( x_{i} \\). Next, we calculate the point and feature weights using an attention mechanism, with both points and features updated similarly. \\( w(f_{i}) \\) calculation can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.499, + 0.483, + 0.542 + ], + "angle": 0, + "content": "\\[\nw \\left(f _ {i}\\right) = \\sum_ {j = 1} ^ {K} \\theta \\left(x _ {i}, x _ {i, j}\\right) \\cdot \\gamma \\left(f _ {i, j}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.629 + ], + "angle": 0, + "content": "In the equation, the function \\(\\theta\\) is used to compute the relationship between a neighbor point and the center point. The function \\(\\gamma (f_{i,j})\\) represents the linear transformation of the neighborhood feature, expressed as \\(\\gamma (f_{i,j}) = W_{\\gamma}f_{i,j} + b_{\\gamma}\\), where \\(W_{\\gamma}\\) and \\(b_{\\gamma}\\) are learnable parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.483, + 0.66 + ], + "angle": 0, + "content": "The function \\(\\theta\\) is computed using the dot product similarity in the attention mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.673, + 0.483, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\theta (x _ {i}, x _ {i, j}) = \\operatorname {s o f t m a x} (q (x _ {i}) \\cdot k (x _ {i, j}) ^ {T} / \\sqrt {C _ {q}}) \\quad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Where \\( q(x_{i}) \\) and \\( k(x_{i,j}) \\) represent the linear transformations of the center point \\( x_{i} \\), and the neighborhood points \\( \\mathcal{N}(x_i) \\), which can be implemented using a \\( 1 \\times 1 \\) convolution kernel. The input to \\( q(x_{i}) \\) is the real coordinates of the center point, while the input to \\( k(x_{i,j}) \\) is the relative coordinates between the neighborhood point and the center point. The convolution kernel has an output channel of \\( C_q \\) and an input channel of 3. Using real and relative coordinates, the model effectively combines global and local spatial information, enhancing its ability to perceive the 3D spatial geometry." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We then use the computed \\( w(f_{i}) \\) to refine the initial points and their features. This operation selects points with" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "more surrounding noise points as removing points, resulting in more accurate features. In other words, we can find the suitable refined point \\( x_{i}^{*} \\) and its corresponding feature \\( f_{i}^{*} \\), expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.163, + 0.907, + 0.18 + ], + "angle": 0, + "content": "\\[\nx _ {i} ^ {*} = \\delta_ {p} \\left(w \\left(f _ {i}\\right)\\right) x _ {i}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.182, + 0.907, + 0.199 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {*} = \\delta_ {f} \\left(w \\left(f _ {i}\\right)\\right) f _ {i}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.207, + 0.906, + 0.268 + ], + "angle": 0, + "content": "Where \\(\\delta_p\\) and \\(\\delta_f\\) is activation function. Finally, we fuse the points \\(x_i^*\\) and their corresponding features \\(f_i^*\\) through an attention module to obtain the module's final output features \\(f_{out}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.277, + 0.729, + 0.292 + ], + "angle": 0, + "content": "4.2. TreeTrans(TT) Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.906, + 0.465 + ], + "angle": 0, + "content": "In the decoding process of encoded point clouds, previous methods often begin with three rounds of interpolation, followed by feature fusion using neighboring point clouds [26], and some methods adopt point-adaptive deconvolution techniques, such as PDR [20]. These methods are typically built on UNet architecture [31], which leads to several issues: (1) The model becomes overly dependent on global features, thereby neglecting local features, and (2) The utilization of multi-scale features is not comprehensive enough, leading to suboptimal performance when handling complex point clouds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.906, + 0.586 + ], + "angle": 0, + "content": "We propose the TreeTrans (TT) module as our decoding network to address the above issues. We first employ two operations similar to TreeGCN[32]: the cyclic and the ancestor operations, which propagate features from the root node to the leaf nodes. After obtaining the intermediate features \\( f_{tree} \\), we combine \\( f_{tree} \\) with the corresponding features from the encoder using local and global attention mechanisms to obtain the output \\( f_{out} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.906, + 0.677 + ], + "angle": 0, + "content": "In traditional graph convolutional networks, the cyclic operation uses a separate weight to transform the current node feature \\( f_{i}^{l} \\) to the next layer. However, a single weight often fails to capture all the relevant information in point clouds. Thus, we employ a \\( K \\)-support-based cyclic operation to refine the feature representation progressively:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.685, + 0.905, + 0.705 + ], + "angle": 0, + "content": "\\[\nS _ {i} ^ {l + 1} = F _ {K} ^ {l} \\left(f _ {i} ^ {l}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.906, + 0.805 + ], + "angle": 0, + "content": "Where \\( F_{K}^{l} \\) is a fully connected layer containing \\( K \\) nodes, and \\( S_{i}^{l + 1} \\) represents the cyclic operation at layer \\( l \\). The cyclic operation is then combined with the ancestor operation to update the node features, ensuring that each layer incorporates information from previous layers during computation. The updated node features can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.813, + 0.905, + 0.862 + ], + "angle": 0, + "content": "\\[\nf _ {t r e e} = \\sigma \\left(S _ {i} ^ {l + 1} + \\sum_ {m _ {j} \\in A \\left(f _ {i} ^ {l}\\right)} U _ {j} ^ {l} m _ {j}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Where \\(\\sigma\\) is the activation function, with ReLU used in this paper. \\(A(f_i^l)\\) denotes the set of all ancestor nodes, and \\(U_j^l\\) is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.99, + 0.518, + 1.0 + ], + "angle": 0, + "content": "16990" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.095, + 0.912, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.414, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 4. Overview of PDANS. The CP network (upper branch) and DP network (lower branch) both consist of an encoder made up of feature extraction and the ANS module, with the TT module serving as the decoder. They exchange information through a mutual attention module. the upsampled ratio and interpolated points are used as conditions to provide prior knowledge of the DP network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.483, + 0.483, + 0.511 + ], + "angle": 0, + "content": "the linear transformation weight from ancestor nodes to the current node." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.484, + 0.589 + ], + "angle": 0, + "content": "After obtaining \\( f_{tree} \\), we input it into the local attention and global attention modules. In the local attention mechanism, the corresponding encoder output features \\( f^{l+1} \\) are combined with \\( f_{tree} \\) using local attention, as described by the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.598, + 0.483, + 0.645 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {l o c a l}} = \\operatorname {s o f t m a x} \\left(\\frac {Q \\left(f ^ {l + 1}\\right) \\cdot K \\left(f _ {\\text {t r e e}}\\right) ^ {T}}{\\sqrt {d}}\\right) \\cdot V \\left(f _ {\\text {t r e e}}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.782 + ], + "angle": 0, + "content": "\\(Q(f^{l + 1})\\) and \\(K(f_{tree})\\) represent two convolution operations that map \\(f^{l + 1}\\) and \\(f_{tree}\\) from their respective dimensions to \\(d\\). Then, the attention weights are obtained by calculating the dot product similarity followed by a softmax operation. Afterward, the weighted sum with \\(V(f_{tree})\\) gives the local attention feature \\(f_{local}\\). Once the local attention is computed, the resulting \\(f_{local}\\) undergoes a global attention mechanism to produce the final output \\(f_{out}\\) of the TT module. The specific implementation is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.79, + 0.484, + 0.826 + ], + "angle": 0, + "content": "\\[\nf _ {o u t} = \\operatorname {s o f t m a x} \\left(\\frac {q (f _ {l o c a l}) \\cdot k (f _ {l o c a l}) ^ {T}}{\\sqrt {d}}\\right) \\cdot v (f _ {l o c a l}) \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.296, + 0.849 + ], + "angle": 0, + "content": "4.3. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In Sections 4.1 and 4.2, we introduced the network's encoding module, ANS, and decoding module, TT, respectively. In this section, we present the overall framework of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.907, + 0.544 + ], + "angle": 0, + "content": "PDANS, which consists of two main components: the Conditional Point Cloud (CP) network and the Diffusion Point Cloud (DP) network. Figure 4 provides a detailed illustration of this process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.909, + 0.73 + ], + "angle": 0, + "content": "CP-Net Our encoder consists of a feature extraction network and the ANS module. The decoder comprises the TT module. Both downsampling and upsampling are performed four times. Before each sampling output, we embed the time step \\( t \\) through an MLP: \\( t \\) is encoded into a \\( C_t \\)-dimensional vector space. The output feature \\( f_{out} \\in \\mathbb{R}^{N \\times D_l} \\) is then transformed through MLPs into the dimension \\( f_{out}' \\in \\mathbb{R}^{N \\times C_t} \\). We then add the two to integrate the time step information into the feature, resulting in the output \\( f_{l+1} \\in \\mathbb{R}^{N \\times D_{l+1}} \\). Additionally, we only input the sparse point cloud into the CP-Net to ensure the extraction of effective features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.9 + ], + "angle": 0, + "content": "DP-Net The DP-Net shares the same architecture as the CP-Net but requires additional conditional information to guide the diffusion process. First, we convert the sparse point cloud \\(\\mathcal{P} \\in \\mathbb{R}^{N \\times 3}\\) into an interpolated point cloud \\(\\mathcal{P}^* \\in \\mathbb{R}^{rN \\times 3}\\) using upsampling by a factor of \\(r\\). Then, we sum \\(\\mathcal{P}^*\\) with the noisy point cloud \\(\\mathcal{P}_T \\in \\mathbb{R}^{rN \\times 3}\\) and input them together into the DP-Net. Meanwhile, we extract global features from \\(\\mathcal{P}^*\\) and encode both the time step \\(t\\) and the upsampling factor \\(r\\). These pieces of information are incorporated as conditions at each stage of the encoder and decoder." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.99, + 0.518, + 1.0 + ], + "angle": 0, + "content": "16991" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "The information provided by the CP-Net is further guided to the DP-Net through a cross-attention (CA) module. Given the outputs of the encoder at each layer for the CP-Net and DP-Net: \\( F_{c} \\in \\mathbb{R}^{N_{c} \\times D_{c}} \\) and \\( F_{d} \\in \\mathbb{R}^{N_{d} \\times D_{d}} \\), the CA module first reshapes the two feature sets using MLPs: \\( F_{c} \\rightarrow (Q) \\in \\mathbb{R}^{N_{c} \\times C_{i}} \\) and \\( F_{d} \\rightarrow (K, V) \\in \\mathbb{R}^{N_{d} \\times C_{i}} \\). Then, the features are fused according to the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.207, + 0.483, + 0.242 + ], + "angle": 0, + "content": "\\[\nF _ {f} = M L P \\left(\\text {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {C _ {i}}}\\right) V\\right) + F _ {c} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.25, + 0.485, + 0.31 + ], + "angle": 0, + "content": "Subsequently, \\( F_{f} \\) is input into the feedforward network to produce the final features. The same method can also be applied reversely, allowing data to flow between the two networks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.319, + 0.166, + 0.334 + ], + "angle": 0, + "content": "4.4. Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.417 + ], + "angle": 0, + "content": "For the DP-Net, its loss can be referred to by equation 5. For the CP-Net, we aim for it to learn more features of the sparse point cloud. Therefore, we add MLPs at the end of the decoder to output a sparse point cloud \\(\\mathcal{P}'\\), which is then supervised by MSE as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.428, + 0.483, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} \\left(\\mathcal {P}, \\mathcal {P} ^ {\\prime}\\right) = \\mathbf {M S E} \\| \\mathcal {P} - \\mathcal {P} ^ {\\prime} \\| ^ {2} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.473, + 0.471 + ], + "angle": 0, + "content": "Combining these loss functions, The training objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.482, + 0.483, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} (\\theta) + \\alpha \\mathcal {L} _ {C} \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.472, + 0.525 + ], + "angle": 0, + "content": "where \\(\\alpha\\) means a weighting factor (\\(\\alpha = 0.5\\) in this paper)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.537, + 0.224, + 0.554 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.308, + 0.578 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.674 + ], + "angle": 0, + "content": "Experiment Setup. All PDANS experiments are run in a PyTorch[24] environment. The PDANS model is trained on a single Nvidia 4080 GPU using the Adam optimizer[13], with a batch size 12 for 350 epochs. The channel sizes for the CP-Net and DP-Net are set to (64, 128, 256, 512) and (128, 256, 256, 512), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Dataset. In our experiments, we utilize two public benchmarks (PUGAN[15], PU1K[27]) for evaluation. We adhere to the official training/testing partitioning protocols for these datasets. Poisson disk sampling [46] is used to generate 24,000 patches on PUGAN and 69,000 patches on PU1K for training. Each patch contains 256 points, with the corresponding ground truth containing 1,024 points. Additionally, 27 and 127 point cloud datasets are used for testing, respectively. The sparse point clouds contain 2,048 points, and the number of points in ground truth is set to \\(2,048 \\times r\\) according to the upsampling factor \\(r\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics Following previous methods, we select three metrics to evaluate our model: Chamfer Distance \\((\\mathrm{CD} \\times 10^{-3})\\) [6], Hausdorff distance \\((\\mathrm{HD} \\times 10^{-3})\\) [2] and Point to Surface Distance \\((\\mathrm{P2F} \\times 10^{-3})\\) [45]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.789, + 0.106 + ], + "angle": 0, + "content": "5.2. Evaluation on PUGAN Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.22 + ], + "angle": 0, + "content": "We perform sampling on PUGAN at low \\((4\\times)\\) and high \\((16\\times)\\) upsampling rates. The reductions in the three metrics shown in Table 1 demonstrate that our model has a significant advantage in describing geometric details. The \\(16\\times\\) upsampled point cloud shown in Figure 5 further supports our claim, as the generated point clouds have fewer outliers and are closer to the ground truth geometric features." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.232, + 0.907, + 0.261 + ], + "angle": 0, + "content": "Table 1. The results of \\(4 \\times\\) and \\(16 \\times\\) on PUGAN. Our method significantly surpasses other methods in terms of CD and HD." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.271, + 0.897, + 0.405 + ], + "angle": 0, + "content": "
Methods4×16×
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]0.5296.8054.4600.5108.2066.041
MPU[44]0.2926.6722.8220.2197.0543.085
PU-GAN[15]0.2825.5772.0160.2076.9632.556
PU-GCN[27]0.2683.2012.4890.1614.2832.632
Grad-PU[8]0.2452.3691.8930.1082.3522.127
RepKpU[30]0.2542.6981.8820.1073.2931.957
PUDM[29]0.1311.2201.9120.0821.1202.114
Ours0.1191.20651.8970.0781.0462.028
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.771, + 0.454 + ], + "angle": 0, + "content": "5.3. Evaluation on PU1K Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.567 + ], + "angle": 0, + "content": "Table 2 shows the results of \\(4\\mathrm{x}\\) upsampling on the PU1K dataset. Compared to other methods, we demonstrate superior performance in both CD and HD metrics. The PU1K dataset is more complex and diverse than others, and our designed TT module exhibits tremendous advantages in handling such complex data. It helps approximate the valid geometric details more effectively, achieving higher accuracy." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.591, + 0.89, + 0.706 + ], + "angle": 0, + "content": "
MethodsCD↓HD↓P2F↓
PU-Net[45]1.15515.1704.834
MPU[44]0.93513.3273.511
PU-GAN[15]0.7359.9394.234
PU-GCN[27]0.5857.5772.499
Grad-PU[8]0.4043.7321.474
RepKpU[30]0.3312.7001.261
PUDM[29]0.2172.1641.477
Ours0.1551.5881.354
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.716, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Table 2. The results of \\(4 \\times\\) on PU1K. Our method outperforms other methods across nearly all metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.673, + 0.788 + ], + "angle": 0, + "content": "5.4. Robustness Test" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Experiment setup. Based on the testing meshes provided by PUGAN, we construct a \\(4 \\times\\) upsampling dataset to conduct a comprehensive robustness test. Ground-truth point clouds are sampled from testing meshes via Poisson disk sampling. We tested three different types of noise: Gaussian, random, and Laplacian. These three types of noise are the most common in real-world point clouds, making" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16992" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.898, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.266, + 0.645, + 0.281 + ], + "angle": 0, + "content": "Figure 5. Comparison \\((16\\times)\\) on PUGAN dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.484, + 0.396 + ], + "angle": 0, + "content": "the results more practical. We set four different noise levels for Gaussian noise: \\(\\tau = 0.01\\), \\(\\tau = 0.02\\), \\(\\tau = 0.05\\), and \\(\\tau = 0.1\\). We used two noise levels for Laplacian and random noise: \\(\\tau = 0.05\\) and \\(\\tau = 0.1\\). For the input with 2048 points, we tested the results under each noise type and noise level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.519 + ], + "angle": 0, + "content": "The results of Gaussian noise Tables 3 and 4 present the quantitative results of Gaussian noise testing under different noise levels. As shown in the tables, our method significantly outperforms other models in terms of resistance to Gaussian noise. Thanks to the design of the ANS module in our model, which further enhances noise resistance. Therefore, our approach excels at noise suppression and surpasses other methods in both CD and HD metrics." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.575 + ], + "angle": 0, + "content": "Table 3. The results of \\(4 \\times\\) at low-level \\((\\tau = 0.01\\) and \\(\\tau = 0.02)\\) Gaussian noise on PUGAN. Our method significantly outperforms other methods in terms of Gaussian noise." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.585, + 0.472, + 0.715 + ], + "angle": 0, + "content": "
Methodsτ = 0.01τ = 0.02
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]0.6288.0689.8161.07810.86716.401
MPU[44]0.5066.9789.0590.92910.82015.621
PU-GAN[15]0.4646.0707.4980.88710.60215.088
PU-GCN[27]0.4485.5866.9890.8168.60413.798
Grad-PU[8]0.4144.1456.4000.7667.33611.534
RepKpU[30]0.3844.9376.7750.8008.41514.124
PUDM[29]0.2102.4306.0700.5295.4719.742
Ours0.1851.9185.9640.3093.0209.513
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The results of random noise and Laplacian noise Tables 5 and 6 present the quantitative test results under noise levels \\(\\tau = 0.05\\) and \\(\\tau = 0.1\\). The results clearly show that the PDANS network we designed exhibits superior adaptability in high-noise environments. Figure 6 illustrates that, in noisy point clouds, our results show more complete structures with fewer outliers. Our method captures more geometric details, leading to higher-quality point clouds. In contrast, other methods tend to upsample blindly, resulting in point clouds with disorganized structures. While our method still maintains SOTA performance, it shows slightly" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.306, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Table 4. The results of \\(4 \\times\\) at high-level \\((\\tau = 0.05\\) and \\(\\tau = 0.1)\\) Gaussian noise on PUGAN. Our method significantly outperforms other methods in terms of Gaussian noise." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.358, + 0.897, + 0.487 + ], + "angle": 0, + "content": "
Methodsτ = 0.05τ = 0.1
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.37013.72923.2491.49814.19323.846
MPU[44]1.24711.64522.1891.32112.41523.841
PU-GAN[15]1.1249.09121.2521.27110.91123.174
PU-GCN[27]1.2639.86922.8351.45611.06325.213
Grad-PU[8]0.9788.05716.9271.1188.94618.845
RepKpU[30]1.23712.33520.5181.36012.90222.473
PUDM[29]0.6185.38614.7510.8536.23916.845
Ours0.4764.03814.6070.5434.32716.476
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.617 + ], + "angle": 0, + "content": "lower results than Gaussian noise. The experimental results validate the effectiveness of the ANS module we designed. By removing noisy point clouds, we can better learn the objects' structural information, thereby generating high-quality point clouds. As a result, PDANS demonstrates good robustness under different types and levels of noise, proving its stability in complex noisy environments." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.629, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Table 5. The results of \\(4 \\times\\) at high-level (\\(\\tau = 0.05\\) and \\(\\tau = 0.1\\)) random noise on PUGAN." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.668, + 0.897, + 0.795 + ], + "angle": 0, + "content": "
Methodsτ = 0.05τ = 0.1
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.49014.47323.2231.72515.44225.251
MPU[44]1.22410.84220.4561.54511.64523.512
PU-GAN[15]1.0347.75718.6171.3279.70021.321
PU-GCN[27]1.0459.64318.8991.32510.87721.633
Grad-PU[8]1.0676.63417.7341.3997.21521.028
RepKpU[30]0.7894.78217.8130.8373.90220.953
PUDM[29]0.9986.11017.5581.3106.73220.564
Ours0.7223.16117.6340.6824.59819.401
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.725, + 0.833 + ], + "angle": 0, + "content": "5.5. Result on Real Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We conducted experiments on the ScanObjectNN [37] dataset, where point clouds are real-scanned, non-uniform and noisy. Since there is no ground truth point cloud, we performed a visual comparison in Fig. 7. The experiment" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16993" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.094, + 0.892, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.239, + 0.756, + 0.253 + ], + "angle": 0, + "content": "Figure 6. Visualization results at \\(4\\times\\) on PUGAN with random noise and Laplace noise." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.278, + 0.483, + 0.307 + ], + "angle": 0, + "content": "Table 6. The results of \\(4 \\times\\) at high-level \\((\\tau = 0.05\\) and \\(\\tau = 0.1)\\) Laplace noise on PUGAN." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.317, + 0.472, + 0.443 + ], + "angle": 0, + "content": "
Methodsτ = 0.1τ = 0.05
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.94332.64626.9651.86127.56324.737
MPU[44]1.56920.32925.1641.42119.04922.783
PU-GAN[15]1.48614.91924.1481.33414.24321.879
PU-GCN[27]1.58715.71125.1341.42514.91522.876
Grad-PU[8]1.23310.34818.5761.0659.34916.591
RepKpU[30]1.34513.10422.2951.19412.37720.218
PUDM[29]0.7046.28418.2830.6437.92416.871
Ours0.5634.05116.2870.5334.43316.195
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.484, + 0.557 + ], + "angle": 0, + "content": "used an input with 2048 points, followed by a \\(16\\times\\) upsampling. The Ball-pivoting algorithm [3] was then applied to reconstruct the surface from the upsampled point cloud. Based on the reconstruction results, our generated point clouds exhibit fewer gaps and smoother surfaces, which benefits surface reconstruction." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.572, + 0.477, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Figure 7. Visual evaluation of impacts on surface construction on real-scanned ScanObjectNN dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.245, + 0.759 + ], + "angle": 0, + "content": "5.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "With/Without ANS Module To demonstrate the effectiveness of the ANS module, we compared it with other network architectures. Model A replaces the ANS module with an encoder composed of Dense Blocks [8], Model B uses an encoder based on Set Abstraction [29], and Model C employs an encoder from PointNet++ [26]. Our network is referred to as Model D. We tested these models on the PU-GAN dataset with Gaussian noise level \\(\\tau = 0.1\\), with results presented in Table7." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.278, + 0.907, + 0.306 + ], + "angle": 0, + "content": "Table 7. The results of \\(4 \\times\\) at high-level \\((\\tau = 0.1)\\) Gaussian noise on PUGAN with different encoders." + }, + { + "type": "table", + "bbox": [ + 0.542, + 0.317, + 0.88, + 0.385 + ], + "angle": 0, + "content": "
ModuleCD↓HD↓P2F↓
A1.2489.63119.657
B0.8536.34816.849
C1.16814.49823.189
D0.5434.32716.476
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.499 + ], + "angle": 0, + "content": "With/Without TT Module We conducted \\(16 \\times\\) upsampling tests on the PUGAN dataset, as shown in Table 8. The experiments indicate a significant performance drop when the TT module is removed, underscoring the importance of the TT module in enhancing the quality of the model generation." + }, + { + "type": "table_caption", + "bbox": [ + 0.601, + 0.511, + 0.819, + 0.525 + ], + "angle": 0, + "content": "Table 8. TT Module Ablation Study" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.536, + 0.902, + 0.584 + ], + "angle": 0, + "content": "
MethodsCD↓HD↓P2F↓
Without TT Module0.0941.1242.158
With TT Module0.0781.0462.028
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.615, + 0.634, + 0.631 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.821 + ], + "angle": 0, + "content": "We propose a novel conditional diffusion model-based network, PDANS, designed for point cloud upsampling. Suppressing noise and ensuring the geometric authenticity of generated results have long been primary research foci in point cloud upsampling. To address these issues, we have designed the ANS and TT modules. The ANS module integrates point cloud features with neighborhood information, enabling adaptive noise removal from the point cloud. Meanwhile, the TT module effectively captures local and global geometric details, allowing for a more accurate reconstruction of geometries close to real values. Extensive experiments validate that the proposed PDANS outperforms existing state-of-the-art methods in performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.848, + 0.611, + 0.864 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.521, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[1] Tenindra Abeywickrama, Muhammad Aamir Cheema, and David Taniar. K-nearest neighbors on road networks: a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16994" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "journey in experimentation and in-memory implementation. arXiv preprint arXiv:1601.01549, 2016.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[2] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Transactions on Graphics (TOG), 32(2):1-17, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.179, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[3] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.236, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[4] Haolan Chen, Shitong Luo, Wei Hu, et al. Deep point set resampling via gradient fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):2913-2930, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[5] Yaodong Cui, Ren Chen, Wenbo Chu, Long Chen, Daxin Tian, Ying Li, and Dongpu Cao. Deep learning for image and point cloud fusion in autonomous driving: A review. IEEE Transactions on Intelligent Transportation Systems, 23(2): 722-739, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.363, + 0.483, + 0.418 + ], + "angle": 0, + "content": "[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.484, + 0.489 + ], + "angle": 0, + "content": "[7] Wanquan Feng, Jin Li, Hongrui Cai, Xiaonan Luo, and Juyong Zhang. Neural points: Point cloud representation with neural fields for arbitrary upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18642, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[8] Yun He, Danhang Tang, Yinda Zhang, Xiangyang Xue, and Yanwei Fu. Grad-pu: Arbitrary-scale point cloud upsampling via gradient descent with learned distance functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.604, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.647, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[11] F. Y. Hu, C. L. Li, T. Zhou, H. F. Cheng, and M. M. Gu. A survey on point cloud completion algorithms for deep learning. Journal of Image and Graphics, 30(02):0309-0333, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[12] Sheng Yu Huang, Hao-Yu Hsu, and Frank Wang. Spovt: Semantic-prototype variational transformer for dense point cloud semantic completion. Advances in Neural Information Processing Systems, 35:33934–33946, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[13] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[14] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[15] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7203-7212, 2019. 1, 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[16] Ruihui Li, Xianzhi Li, Pheng-Ann Heng, and Chi-Wing Fu. Point cloud upsampling via disentangled refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 344–353, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[17] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3d object reconstruction. In proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[18] Kangcheng Liu, Yuzhi Zhao, Zhi Gao, and Ben M Chen. Weaklabel3d-net: A complete framework for real-scene lidar point clouds weakly supervised multi-tasks understanding. In 2022 international conference on robotics and automation (ICRA), pages 5108-5115. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[19] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2837-2845, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[20] Zhaoyang Lyu, Zhifeng Kong, Xudong Xu, Liang Pan, and Dahua Lin. A conditional point diffusion-refinement paradigm for 3d point cloud completion. arXiv preprint arXiv:2112.03530, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[21] Aihua Mao, Zihui Du, Junhui Hou, Yaqi Duan, Yong-jin Liu, and Ying He. Pu-flow: A point cloud upsampling network with normalizing flows. IEEE Transactions on Visualization and Computer Graphics, 29(12):4964-4977, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.907, + 0.517 + ], + "angle": 0, + "content": "[22] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12923-12932, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.907, + 0.588 + ], + "angle": 0, + "content": "[23] Lucas Nunes, Rodrigo Marcuzzi, Benedikt Mersch, Jens Behley, and Cyril Stachniss. Scaling diffusion models to real-world 3d lidar scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14770-14780, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[27] Guocheng Qian, Abdulellah Abualshour, Guohao Li, Ali Thabet, and Bernard Ghanem. Pu-gcn: Point cloud upsampling using graph convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11683-11692, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Shi Qiu, Saeed Anwar, and Nick Barnes. Pu-transformer: Point cloud upsampling transformer. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16995" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Asian conference on computer vision, pages 2475-2493, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.486, + 0.191 + ], + "angle": 0, + "content": "[29] Wentao Qu, Yuantian Shao, Lingwu Meng, Xiaoshui Huang, and Liang Xiao. A conditional denoising diffusion probabilistic model for point cloud upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20786-20795, 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.485, + 0.262 + ], + "angle": 0, + "content": "[30] Yi Rong, Haoran Zhou, Kang Xia, Cheng Mei, Jiahao Wang, and Tong Lu. Repkpu: Point cloud upsampling with kernel point representation and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21050-21060, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[32] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[34] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[35] Hanxiao Tan. Dam: Diffusion activation maximization for 3d global explanations. arXiv preprint arXiv:2401.14938, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[37] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[38] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. Advances in Neural Information Processing Systems, 34:29088-29100, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[39] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5499-5509, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[40] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5589-5598, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[41] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3323–3332, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[42] Lei Yang, Yanhong Liu, Jinzhu Peng, and Zize Liang. A novel system for off-line 3d seam extraction and path planning based on point cloud segmentation for arc welding robot. Robotics and Computer-Integrated Manufacturing, 64:101929, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[43] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[44] Wang Yifan, Shihao Wu, Hui Huang, Daniel Cohen-Or, and Olga Sorkine-Hornung. Patch-based progressive 3d point set upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5958–5967, 2019. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[45] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2790–2799, 2018. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[46] Cem Yuksel. Sample elimination for generating poisson disk sample sets. In Computer Graphics Forum, pages 25-32. Wiley Online Library, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[47] Yuchao Zheng, Yujie Li, Shuo Yang, and Huimin Lu. Global-pbnet: A novel point cloud registration for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(11):22312-22319, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16996" + } + ] +] \ No newline at end of file diff --git a/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/96fbf65a-6eae-49b1-9bdc-b206a300f6d4_origin.pdf b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/96fbf65a-6eae-49b1-9bdc-b206a300f6d4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e0a87052d4bee3d8f94d3cdaae1c0d1ceebd5605 --- /dev/null +++ b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/96fbf65a-6eae-49b1-9bdc-b206a300f6d4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b39a336e626890d073c8ae83312ca3a96995563be3c78e002a5ac49b5264427a +size 2754084 diff --git a/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/full.md b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6027be3190f6e924ad741c1be93032b7d0092449 --- /dev/null +++ b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/full.md @@ -0,0 +1,359 @@ +# Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression + +Boqian Zhang\*, Shen Yang\*, Hao Chen, Chao Yang, Jing Jia, Guang Jiang† +Xidian University + +{24011211294，20009101459，haochen(cv，24011211295}@xidian.stu.edu.cn, + +jjia@mail.xidian.edu.cn, gjiang@mail.xidian.edu.cn + +# Abstract + +Point cloud upsampling can improve the quality of the initial point cloud, significantly enhancing the performance of downstream tasks such as classification and segmentation. Existing methods mostly focus on generating the geometric details of point clouds, neglecting noise suppression. To address this, we propose a novel network based on a conditional diffusion model, incorporating the Adaptive Noise Suppression (ANS) module, which we refer to as PDANS. The ANS module assigns weights to each point and determines the removal strategy based on these weights, reducing the impact of noisy points on the sampling process. The module first selects the neighborhood set for each point in the point cloud and performs a weighted sum between the point and its neighbors. It then adjusts the removal points based on the weighted sum, effectively mitigating the bias caused by outliers. We introduce the TreeTrans (TT) module to capture more correlated feature information. This module learns the interaction between high-level and low-level features, resulting in a more comprehensive and refined feature representation. Our results on several widely used benchmark datasets demonstrate that PDANS exhibits exceptional robustness in noisy point cloud processing and outperforms current state-of-the-art(SOTA) methods in terms of performance. Code is available at https://github.com/Baty2023/PDANS. + +# 1. Introduction + +Point clouds, as one of the key representations for 3D reconstruction [17] [22], are widely applied in autonomous driving [5] [47], embodied intelligence [18] [42], and AR/VR. However, real-world scanned point clouds are often sparse, noisy, and uneven, negatively impacting their performance in downstream tasks such as classification and segmenta + +tion. Point cloud upsampling can transform sparse and noisy point clouds into dense, artifact-free representations. In recent years, significant progress has been made in this task [15] [27] [45] [40]. + +![](images/954005717d3044c550c2828f584847ee24c018e74bc14fa9b0731a959fa5970f.jpg) +Figure 1. Most existing methods achieve satisfactory results when input without noise, but they often struggle to effectively recover the geometric structure (e.g., panda ears) when dealing with input with noise. + +Inspired by deep learning, several methods have adopted deep networks to perform this task. Existing upsampling methods can be broadly classified into two categories: those based on generative models and those based on refinement methods. Generative models [15] [27] [28] first extract features from sparse point clouds into a latent space and then design a generator based on the extracted features to generate the coordinates of dense point clouds. The limitation of such methods lies in the difficulty of accurately predicting 3D coordinates, leading to potential outliers or artifacts in the generated dense point clouds [8] [45]. Refinement-based methods [4] [39] [30] first use interpolation to generate a coarse dense point cloud, and then employ a neural network to extract geometric features from the coarse point cloud for refinement. Combined with reconstruction loss functions, this process refines the point cloud to match the ground truth. + +These methods share two common issues: (1) The features lack significant dominance or distinguishability, leading to generating point clouds without clear structure, which tend to be sparse and scattered. (2) Most of these methods + +rely on loss functions such as CD loss, EMD loss, and other similar metrics, which are highly sensitive to outliers. This sensitivity makes the model ineffective at handling noisy point clouds [12] [38], as shown in Figure1. + +To address the above issues, we propose a point cloud upsampling method with adaptive noise suppression, PDANS. The network is designed based on a conditional diffusion model, leveraging the inherent denoising properties of the diffusion model's generative process to enhance its adaptability to noisy point clouds. The network also integrates an Adaptive Noise Suppression (ANS) module, which we designed to compute downsampling weights for each point by utilizing the relationships between the point and its neighboring points. This approach enables the extraction of more accurate downsampled points and features. It effectively reduces the impact of noisy points during feature extraction and minimizes the bias caused by outliers. + +To enhance feature representation, we propose a structure based on graph convolutional networks [14], called the TreeTrans (TT) module. This module incorporates ancestral and recurrent terms[32] to fully utilize the features outputted by each encoder layer. Additionally, we apply both local and global attention mechanisms to the output features, further improving the efficiency and expressiveness of the feature representation. + +Additionally, we designed a conditional point cloud network and a diffusion point cloud network. The conditional point cloud network provides relevant features of sparse point clouds to the diffusion point cloud network through a mutual attention module, guiding the diffusion network to generate the required noise for point cloud densification. Similar to PUDM[29], we incorporate the diffusion model, thereby eliminating the reliance on CD loss. Combined with the ANS module we designed, PDANS demonstrates exceptional robustness in noise experiments. Our main contributions are as follows: + +- We designed a noise suppression module named ANS, which adaptively adjusts removal points and their neighborhood features to improve robustness against noise and outliers. +- The proposed TT module fully leverages features across different layers of the encoder, incorporating ancestral terms, recurrent terms, and local and global attention mechanisms to achieve more comprehensive and refined feature representations. +- We designed an end-to-end network architecture named PDANS, which utilizes a mutual attention mechanism and a diffusion model to guide the densification of sparse point clouds. The ANS module further enhances the network's ability to handle noise. +- Extensive experiments demonstrate the outstanding performance of PDANS in point cloud upsampling benchmarks, particularly in generating geometric details and + +handling noisy point clouds effectively. + +# 2. Related Work + +# 2.1. Learning-based point cloud upsampling + +Several methods [25] [26] [36] [11] shave been developed to extract features from 3D data directly with the strong data-driven and representational capabilities of deep learning networks. PU-Net [45] was the first learning-based method for point cloud and upsampling. Following its proposed architecture, many methods have emerged: MPU [44], PUGAN [15], Dis-PU [16], PUGCN [27], and RepKPU [30]. The NePs [7] method enhances neural radiance fields by combining 2D information with 3D features, enabling point cloud upsampling. Some networks attempt to perform point cloud upsampling based on denoising paradigms. They generate a rough upsampled point cloud through interpolation, and then refine it using neural networks, such as GradPU [8] and PU-Flow [21]. + +Most methods predict the coordinates of dense point clouds from sparse point cloud features and then rely on upsampling modules to expand the point cloud, or further refine the coordinates from the coarsely densified point clouds. Since these methods typically depend on CD (Chamfer Distance) and EMD (Earth Mover's Distance) as loss functions, they exhibit high sensitivity to noise. In this paper, we design an additional denoising module and utilize conditional DDPM to reduce the dependence on CD. + +# 2.2. DDPM for 3D Generation + +Inspired by the success of image generation tasks [10] [34] [9], direct point cloud generation using DDPM has gained increasing attention. Luo et al. [19] is the first to apply DDPM to unconditional point cloud generation. The DAM [35] network explores the interpretability of diffusion models for point cloud generation tasks. Nunes et al. [23] apply DDPM to the 3D-LiDAR point cloud completion task in real-world scenarios. The PDR [20] network takes the initial point cloud as input, but it introduces a coarse and refinement network, which increases time overhead. PUDM [29] densifies point clouds through conditional diffusion, but its performance in noise suppression is suboptimal. + +This paper uses a conditional diffusion model to perform point cloud upsampling. To enhance the model's denoising capability, we design an additional ANS module. + +# 3. Background of Conditional Diffusion Module + +We assume that the dense point cloud $x$ is sampled from a meaningful distribution $P_{data}$ , and that the latent variable $x_{T} \sim \mathcal{N}(\mathbf{0}_{3N},\mathbf{I}_{3N\times 3N})$ is sampled from a noise distribution $P_{latent}$ which is easier to handle, where $\mathcal{N}$ follows a + +Gaussian distribution. DDPM[10] establishes a transfer relationship between $x$ and $x_{T}$ through two Markov chains. This is an autoregressive process: the forward process progressively adds noise to $x$ until $x_{T}$ , and $x_{T}$ is denoised back to $x$ through the reverse process. We use the timestep $t \sim U(T)$ to control the speed of transformation. + +Forward Process The process gradually adds noise to $x_0$ until $x_T$ through the probability $q(x_t | x_{t-1})$ . The entire process can be defined as: + +$$ +q \left(x _ {1}, \dots , x _ {T} \mid x _ {0}\right) = \prod_ {t = 1} ^ {T} q \left(x _ {t} \mid x _ {t - 1}\right) \tag {1} +$$ + +Where $q(x_{t}|x_{t - 1}) = \mathcal{N}(x_{t};\sqrt{1 - \beta_{t}} x_{t - 1},\beta_{t}I)$ , where the hyperparameter $\beta_{t}$ is a positive constant. Based on the derivation by Ho et al. [10], $x_{T}$ can be computed directly from $x_0$ . Define $\alpha_{t} = 1 - \beta_{t}$ and $\bar{\alpha}_{t} = \prod_{i = 1}^{t}\alpha_{i}$ , so that $x_{T}$ can be sampled from the following formula: + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon \tag {2} +$$ + +Where $\epsilon$ follows a standard normal distribution. + +Reverse Process The reverse process aims to predict and eliminate the noise added at each step $t$ . It is conditioned on a set of conditions, which may include incomplete point clouds $c$ , upsampling factor $r$ , etc. We denote the set of conditions as $C = \{c_i \mid i = 1, \dots, N_c\}$ , where $N_c$ represents the number of conditions. We can define the reverse process from denoising $x_{T}$ to $x_{0}$ as follows: + +$$ +p _ {\boldsymbol {\theta}} \left(x _ {0}, \dots , x _ {T - 1} \mid x _ {T}, \boldsymbol {C}\right) = \prod_ {t = 1} ^ {T} p _ {\boldsymbol {\theta}} \left(x _ {t - 1} \mid x _ {t}, \boldsymbol {C}\right) \tag {3} +$$ + +Where $p_{\boldsymbol{\theta}}(x_{t - 1}|x_t,\boldsymbol {C}) = \mathcal{N}(x_{t - 1};\mu_{\boldsymbol{\theta}}(\boldsymbol {x}_t,\boldsymbol {C},t),\sigma_t^2\boldsymbol {I})$ where the mean $\mu_{\boldsymbol{\theta}}(x_t,C,t)$ is the result predicted by a neural network, and $\sigma_t^2$ is a time-dependent constant. Based on the derivation by DDIM [33], the reverse process can be written in the following form: + +$$ +\begin{array}{l} x _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} \left(\frac {x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} (x _ {t} , \boldsymbol {C} , t)}{\sqrt {\bar {\alpha} _ {t}}}\right) + \tag {4} \\ \sqrt {1 - \bar {\alpha} _ {t - 1} - \sigma_ {t} ^ {2}} \epsilon_ {\theta} (x _ {t}, C, t) + \sigma_ {t} \epsilon_ {t} \\ \end{array} +$$ + +Where $\epsilon_{t} \sim \mathcal{N}(\mathbf{0}_{3N},\mathbf{I}_{3N\times 3N})$ is standard Gaussian noise independent of $x_{T}$ . The noisy point cloud $x_{T}$ is progressively denoised according to the above equation until it recovers $x_0$ . At each time step of the reverse process, conditions are added, allowing control over the generation of dense point clouds, as shown in the process diagram in Figure 2. + +Training process under given conditions The goal is to optimize the denoising model's prediction of the noise $\epsilon_{\theta}$ at step $t$ . Given a set of conditions $C$ and under the + +![](images/4bf08cbdb1fcf68f37c1937cece8f9526363ded9fe1077aec8bc26a5170a3c84.jpg) +Figure 2. Visualization of the forward and reverse processes of PDANS. + +condition that the time step $t\sim U(T)$ , the training objective of DDPM is: + +$$ +\mathcal {L} (\theta) = \mathbb {E} _ {\boldsymbol {t} \sim U (T), \epsilon \sim \mathcal {N} (0, I)} \| \epsilon - \epsilon_ {\theta} \left(\sqrt {1 - \bar {\alpha} _ {t}} \epsilon + \sqrt {\bar {\alpha} _ {t}} x _ {0}, C, t\right) \| ^ {2} \tag {5} +$$ + +In the above equation, there are no CD loss or EMD loss terms commonly used in traditional point cloud upsampling networks, because DDPM naturally defines a one-to-one pointwise correspondence between two consecutive point clouds, as shown in equation1. + +# 4. Method + +This section introduces each module's design and the network's overall structure. In PDANS, we propose two modules: the Adaptive Noise Suppression (ANS) module and the TreeTrans (TT) module. These modules are described in Sections 4.1 and 4.2, respectively. The design of the overall PDANS network is presented in Section 4.3. + +# 4.1. Adaptive Noise Suppression(ANS) Module + +In the encoding stage, many frameworks perform down-sampling using Farthest Point Sampling (FPS), followed by selecting neighboring points within a certain radius [20] or via KNN [1] to enhance the features of the sampled points. However, two issues arise when the input point cloud contains significant noise: (1) FPS is sensitive to outliers[43][41], making it challenging to process noisy point clouds effectively. (2) FPS does not filter noisy points, leading to the inclusion of noise features in the enhanced features, which affects the accuracy and reliability of the sampling. + +We propose the Adaptive Noise Suppression (ANS) module to address the issues abovementioned, as shown in Figure 3. Unlike FPS, we first select the neighborhood point cloud $\mathcal{P}_K\in \mathbb{R}^{N\times 3\times K}$ for all points in the initial point cloud $\mathcal{P}\in \mathbb{R}^{N\times 3}$ , where $K$ denotes the number of neighboring points for each point. These neighborhood point clouds contain both correct and noisy points. We then input the neighborhood point cloud, the initial point cloud, and their corresponding features into the ANS module to perform the removal operation. + +![](images/fd530f1e0df6cb8f5f1b38e56839fad1c90b4e58fa275621700cde9d501592f9.jpg) +Figure 3. Visualization of the Adaptive Noise Suppression (ANS) module + +For each point $x_{i}$ selected from $\mathcal{P}$ , we select its corresponding feature $f_{i}$ from $\mathcal{F}s \in \mathbb{R}^{N \times D_l}$ , where $\mathcal{F}s$ is the feature set obtained by reordering the initial point cloud. We use KNN to select its $K$ nearest neighbors $x_{i,1}, \ldots, x_{i,K} \in \mathcal{N}(x_i)$ and their corresponding features $f_{i,1}, \ldots, f_{i,K} \in \mathcal{N}(f_i)$ , where $\mathcal{N}(x_i)$ represents the $K$ nearest neighbors of point $x_{i}$ . Next, we calculate the point and feature weights using an attention mechanism, with both points and features updated similarly. $w(f_{i})$ calculation can be expressed as: + +$$ +w \left(f _ {i}\right) = \sum_ {j = 1} ^ {K} \theta \left(x _ {i}, x _ {i, j}\right) \cdot \gamma \left(f _ {i, j}\right) \tag {6} +$$ + +In the equation, the function $\theta$ is used to compute the relationship between a neighbor point and the center point. The function $\gamma (f_{i,j})$ represents the linear transformation of the neighborhood feature, expressed as $\gamma (f_{i,j}) = W_{\gamma}f_{i,j} + b_{\gamma}$ , where $W_{\gamma}$ and $b_{\gamma}$ are learnable parameters. + +The function $\theta$ is computed using the dot product similarity in the attention mechanism: + +$$ +\theta (x _ {i}, x _ {i, j}) = \operatorname {s o f t m a x} (q (x _ {i}) \cdot k (x _ {i, j}) ^ {T} / \sqrt {C _ {q}}) \quad (7) +$$ + +Where $q(x_{i})$ and $k(x_{i,j})$ represent the linear transformations of the center point $x_{i}$ , and the neighborhood points $\mathcal{N}(x_i)$ , which can be implemented using a $1 \times 1$ convolution kernel. The input to $q(x_{i})$ is the real coordinates of the center point, while the input to $k(x_{i,j})$ is the relative coordinates between the neighborhood point and the center point. The convolution kernel has an output channel of $C_q$ and an input channel of 3. Using real and relative coordinates, the model effectively combines global and local spatial information, enhancing its ability to perceive the 3D spatial geometry. + +We then use the computed $w(f_{i})$ to refine the initial points and their features. This operation selects points with + +more surrounding noise points as removing points, resulting in more accurate features. In other words, we can find the suitable refined point $x_{i}^{*}$ and its corresponding feature $f_{i}^{*}$ , expressed as: + +$$ +x _ {i} ^ {*} = \delta_ {p} \left(w \left(f _ {i}\right)\right) x _ {i}, \tag {8} +$$ + +$$ +f _ {i} ^ {*} = \delta_ {f} \left(w \left(f _ {i}\right)\right) f _ {i}. \tag {9} +$$ + +Where $\delta_p$ and $\delta_f$ is activation function. Finally, we fuse the points $x_i^*$ and their corresponding features $f_i^*$ through an attention module to obtain the module's final output features $f_{out}$ . + +# 4.2. TreeTrans(TT) Module + +In the decoding process of encoded point clouds, previous methods often begin with three rounds of interpolation, followed by feature fusion using neighboring point clouds [26], and some methods adopt point-adaptive deconvolution techniques, such as PDR [20]. These methods are typically built on UNet architecture [31], which leads to several issues: (1) The model becomes overly dependent on global features, thereby neglecting local features, and (2) The utilization of multi-scale features is not comprehensive enough, leading to suboptimal performance when handling complex point clouds. + +We propose the TreeTrans (TT) module as our decoding network to address the above issues. We first employ two operations similar to TreeGCN[32]: the cyclic and the ancestor operations, which propagate features from the root node to the leaf nodes. After obtaining the intermediate features $f_{tree}$ , we combine $f_{tree}$ with the corresponding features from the encoder using local and global attention mechanisms to obtain the output $f_{out}$ . + +In traditional graph convolutional networks, the cyclic operation uses a separate weight to transform the current node feature $f_{i}^{l}$ to the next layer. However, a single weight often fails to capture all the relevant information in point clouds. Thus, we employ a $K$ -support-based cyclic operation to refine the feature representation progressively: + +$$ +S _ {i} ^ {l + 1} = F _ {K} ^ {l} \left(f _ {i} ^ {l}\right) \tag {10} +$$ + +Where $F_{K}^{l}$ is a fully connected layer containing $K$ nodes, and $S_{i}^{l + 1}$ represents the cyclic operation at layer $l$ . The cyclic operation is then combined with the ancestor operation to update the node features, ensuring that each layer incorporates information from previous layers during computation. The updated node features can be written as: + +$$ +f _ {t r e e} = \sigma \left(S _ {i} ^ {l + 1} + \sum_ {m _ {j} \in A \left(f _ {i} ^ {l}\right)} U _ {j} ^ {l} m _ {j}\right) \tag {11} +$$ + +Where $\sigma$ is the activation function, with ReLU used in this paper. $A(f_i^l)$ denotes the set of all ancestor nodes, and $U_j^l$ is + +![](images/ba9457d5dc27e8fb2895c6d890b0921ac41a4268e2c56dd99122196ebbe1e684.jpg) +Figure 4. Overview of PDANS. The CP network (upper branch) and DP network (lower branch) both consist of an encoder made up of feature extraction and the ANS module, with the TT module serving as the decoder. They exchange information through a mutual attention module. the upsampled ratio and interpolated points are used as conditions to provide prior knowledge of the DP network. + +the linear transformation weight from ancestor nodes to the current node. + +After obtaining $f_{tree}$ , we input it into the local attention and global attention modules. In the local attention mechanism, the corresponding encoder output features $f^{l+1}$ are combined with $f_{tree}$ using local attention, as described by the following formula: + +$$ +f _ {\text {l o c a l}} = \operatorname {s o f t m a x} \left(\frac {Q \left(f ^ {l + 1}\right) \cdot K \left(f _ {\text {t r e e}}\right) ^ {T}}{\sqrt {d}}\right) \cdot V \left(f _ {\text {t r e e}}\right) \tag {12} +$$ + +$Q(f^{l + 1})$ and $K(f_{tree})$ represent two convolution operations that map $f^{l + 1}$ and $f_{tree}$ from their respective dimensions to $d$ . Then, the attention weights are obtained by calculating the dot product similarity followed by a softmax operation. Afterward, the weighted sum with $V(f_{tree})$ gives the local attention feature $f_{local}$ . Once the local attention is computed, the resulting $f_{local}$ undergoes a global attention mechanism to produce the final output $f_{out}$ of the TT module. The specific implementation is as follows: + +$$ +f _ {o u t} = \operatorname {s o f t m a x} \left(\frac {q (f _ {l o c a l}) \cdot k (f _ {l o c a l}) ^ {T}}{\sqrt {d}}\right) \cdot v (f _ {l o c a l}) \tag {13} +$$ + +# 4.3. Network Architecture + +In Sections 4.1 and 4.2, we introduced the network's encoding module, ANS, and decoding module, TT, respectively. In this section, we present the overall framework of + +PDANS, which consists of two main components: the Conditional Point Cloud (CP) network and the Diffusion Point Cloud (DP) network. Figure 4 provides a detailed illustration of this process. + +CP-Net Our encoder consists of a feature extraction network and the ANS module. The decoder comprises the TT module. Both downsampling and upsampling are performed four times. Before each sampling output, we embed the time step $t$ through an MLP: $t$ is encoded into a $C_t$ -dimensional vector space. The output feature $f_{out} \in \mathbb{R}^{N \times D_l}$ is then transformed through MLPs into the dimension $f_{out}' \in \mathbb{R}^{N \times C_t}$ . We then add the two to integrate the time step information into the feature, resulting in the output $f_{l+1} \in \mathbb{R}^{N \times D_{l+1}}$ . Additionally, we only input the sparse point cloud into the CP-Net to ensure the extraction of effective features. + +DP-Net The DP-Net shares the same architecture as the CP-Net but requires additional conditional information to guide the diffusion process. First, we convert the sparse point cloud $\mathcal{P} \in \mathbb{R}^{N \times 3}$ into an interpolated point cloud $\mathcal{P}^* \in \mathbb{R}^{rN \times 3}$ using upsampling by a factor of $r$ . Then, we sum $\mathcal{P}^*$ with the noisy point cloud $\mathcal{P}_T \in \mathbb{R}^{rN \times 3}$ and input them together into the DP-Net. Meanwhile, we extract global features from $\mathcal{P}^*$ and encode both the time step $t$ and the upsampling factor $r$ . These pieces of information are incorporated as conditions at each stage of the encoder and decoder. + +The information provided by the CP-Net is further guided to the DP-Net through a cross-attention (CA) module. Given the outputs of the encoder at each layer for the CP-Net and DP-Net: $F_{c} \in \mathbb{R}^{N_{c} \times D_{c}}$ and $F_{d} \in \mathbb{R}^{N_{d} \times D_{d}}$ , the CA module first reshapes the two feature sets using MLPs: $F_{c} \rightarrow (Q) \in \mathbb{R}^{N_{c} \times C_{i}}$ and $F_{d} \rightarrow (K, V) \in \mathbb{R}^{N_{d} \times C_{i}}$ . Then, the features are fused according to the following formula: + +$$ +F _ {f} = M L P \left(\text {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {C _ {i}}}\right) V\right) + F _ {c} \tag {14} +$$ + +Subsequently, $F_{f}$ is input into the feedforward network to produce the final features. The same method can also be applied reversely, allowing data to flow between the two networks. + +# 4.4. Loss + +For the DP-Net, its loss can be referred to by equation 5. For the CP-Net, we aim for it to learn more features of the sparse point cloud. Therefore, we add MLPs at the end of the decoder to output a sparse point cloud $\mathcal{P}'$ , which is then supervised by MSE as follows: + +$$ +\mathcal {L} _ {C} \left(\mathcal {P}, \mathcal {P} ^ {\prime}\right) = \mathbf {M S E} \| \mathcal {P} - \mathcal {P} ^ {\prime} \| ^ {2} \tag {15} +$$ + +Combining these loss functions, The training objective is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} (\theta) + \alpha \mathcal {L} _ {C} \tag {16} +$$ + +where $\alpha$ means a weighting factor ( $\alpha = 0.5$ in this paper). + +# 5. Experiments + +# 5.1. Implementation Details + +Experiment Setup. All PDANS experiments are run in a PyTorch[24] environment. The PDANS model is trained on a single Nvidia 4080 GPU using the Adam optimizer[13], with a batch size 12 for 350 epochs. The channel sizes for the CP-Net and DP-Net are set to (64, 128, 256, 512) and (128, 256, 256, 512), respectively. + +Dataset. In our experiments, we utilize two public benchmarks (PUGAN[15], PU1K[27]) for evaluation. We adhere to the official training/testing partitioning protocols for these datasets. Poisson disk sampling [46] is used to generate 24,000 patches on PUGAN and 69,000 patches on PU1K for training. Each patch contains 256 points, with the corresponding ground truth containing 1,024 points. Additionally, 27 and 127 point cloud datasets are used for testing, respectively. The sparse point clouds contain 2,048 points, and the number of points in ground truth is set to $2,048 \times r$ according to the upsampling factor $r$ . + +Evaluation Metrics Following previous methods, we select three metrics to evaluate our model: Chamfer Distance $(\mathrm{CD} \times 10^{-3})$ [6], Hausdorff distance $(\mathrm{HD} \times 10^{-3})$ [2] and Point to Surface Distance $(\mathrm{P2F} \times 10^{-3})$ [45]. + +# 5.2. Evaluation on PUGAN Dataset + +We perform sampling on PUGAN at low $(4\times)$ and high $(16\times)$ upsampling rates. The reductions in the three metrics shown in Table 1 demonstrate that our model has a significant advantage in describing geometric details. The $16\times$ upsampled point cloud shown in Figure 5 further supports our claim, as the generated point clouds have fewer outliers and are closer to the ground truth geometric features. + +Table 1. The results of $4 \times$ and $16 \times$ on PUGAN. Our method significantly surpasses other methods in terms of CD and HD. + +
Methods4×16×
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]0.5296.8054.4600.5108.2066.041
MPU[44]0.2926.6722.8220.2197.0543.085
PU-GAN[15]0.2825.5772.0160.2076.9632.556
PU-GCN[27]0.2683.2012.4890.1614.2832.632
Grad-PU[8]0.2452.3691.8930.1082.3522.127
RepKpU[30]0.2542.6981.8820.1073.2931.957
PUDM[29]0.1311.2201.9120.0821.1202.114
Ours0.1191.20651.8970.0781.0462.028
+ +# 5.3. Evaluation on PU1K Dataset + +Table 2 shows the results of $4\mathrm{x}$ upsampling on the PU1K dataset. Compared to other methods, we demonstrate superior performance in both CD and HD metrics. The PU1K dataset is more complex and diverse than others, and our designed TT module exhibits tremendous advantages in handling such complex data. It helps approximate the valid geometric details more effectively, achieving higher accuracy. + +
MethodsCD↓HD↓P2F↓
PU-Net[45]1.15515.1704.834
MPU[44]0.93513.3273.511
PU-GAN[15]0.7359.9394.234
PU-GCN[27]0.5857.5772.499
Grad-PU[8]0.4043.7321.474
RepKpU[30]0.3312.7001.261
PUDM[29]0.2172.1641.477
Ours0.1551.5881.354
+ +Table 2. The results of $4 \times$ on PU1K. Our method outperforms other methods across nearly all metrics. + +# 5.4. Robustness Test + +Experiment setup. Based on the testing meshes provided by PUGAN, we construct a $4 \times$ upsampling dataset to conduct a comprehensive robustness test. Ground-truth point clouds are sampled from testing meshes via Poisson disk sampling. We tested three different types of noise: Gaussian, random, and Laplacian. These three types of noise are the most common in real-world point clouds, making + +![](images/7d2686eb0d5d4fbae318fb58e35b1ad43b1825fd43872fc998f86357d9bf504e.jpg) +Figure 5. Comparison $(16\times)$ on PUGAN dataset. + +the results more practical. We set four different noise levels for Gaussian noise: $\tau = 0.01$ , $\tau = 0.02$ , $\tau = 0.05$ , and $\tau = 0.1$ . We used two noise levels for Laplacian and random noise: $\tau = 0.05$ and $\tau = 0.1$ . For the input with 2048 points, we tested the results under each noise type and noise level. + +The results of Gaussian noise Tables 3 and 4 present the quantitative results of Gaussian noise testing under different noise levels. As shown in the tables, our method significantly outperforms other models in terms of resistance to Gaussian noise. Thanks to the design of the ANS module in our model, which further enhances noise resistance. Therefore, our approach excels at noise suppression and surpasses other methods in both CD and HD metrics. + +Table 3. The results of $4 \times$ at low-level $(\tau = 0.01$ and $\tau = 0.02)$ Gaussian noise on PUGAN. Our method significantly outperforms other methods in terms of Gaussian noise. + +
Methodsτ = 0.01τ = 0.02
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]0.6288.0689.8161.07810.86716.401
MPU[44]0.5066.9789.0590.92910.82015.621
PU-GAN[15]0.4646.0707.4980.88710.60215.088
PU-GCN[27]0.4485.5866.9890.8168.60413.798
Grad-PU[8]0.4144.1456.4000.7667.33611.534
RepKpU[30]0.3844.9376.7750.8008.41514.124
PUDM[29]0.2102.4306.0700.5295.4719.742
Ours0.1851.9185.9640.3093.0209.513
+ +The results of random noise and Laplacian noise Tables 5 and 6 present the quantitative test results under noise levels $\tau = 0.05$ and $\tau = 0.1$ . The results clearly show that the PDANS network we designed exhibits superior adaptability in high-noise environments. Figure 6 illustrates that, in noisy point clouds, our results show more complete structures with fewer outliers. Our method captures more geometric details, leading to higher-quality point clouds. In contrast, other methods tend to upsample blindly, resulting in point clouds with disorganized structures. While our method still maintains SOTA performance, it shows slightly + +Table 4. The results of $4 \times$ at high-level $(\tau = 0.05$ and $\tau = 0.1)$ Gaussian noise on PUGAN. Our method significantly outperforms other methods in terms of Gaussian noise. + +
Methodsτ = 0.05τ = 0.1
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.37013.72923.2491.49814.19323.846
MPU[44]1.24711.64522.1891.32112.41523.841
PU-GAN[15]1.1249.09121.2521.27110.91123.174
PU-GCN[27]1.2639.86922.8351.45611.06325.213
Grad-PU[8]0.9788.05716.9271.1188.94618.845
RepKpU[30]1.23712.33520.5181.36012.90222.473
PUDM[29]0.6185.38614.7510.8536.23916.845
Ours0.4764.03814.6070.5434.32716.476
+ +lower results than Gaussian noise. The experimental results validate the effectiveness of the ANS module we designed. By removing noisy point clouds, we can better learn the objects' structural information, thereby generating high-quality point clouds. As a result, PDANS demonstrates good robustness under different types and levels of noise, proving its stability in complex noisy environments. + +Table 5. The results of $4 \times$ at high-level ( $\tau = 0.05$ and $\tau = 0.1$ ) random noise on PUGAN. + +
Methodsτ = 0.05τ = 0.1
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.49014.47323.2231.72515.44225.251
MPU[44]1.22410.84220.4561.54511.64523.512
PU-GAN[15]1.0347.75718.6171.3279.70021.321
PU-GCN[27]1.0459.64318.8991.32510.87721.633
Grad-PU[8]1.0676.63417.7341.3997.21521.028
RepKpU[30]0.7894.78217.8130.8373.90220.953
PUDM[29]0.9986.11017.5581.3106.73220.564
Ours0.7223.16117.6340.6824.59819.401
+ +# 5.5. Result on Real Dataset + +We conducted experiments on the ScanObjectNN [37] dataset, where point clouds are real-scanned, non-uniform and noisy. Since there is no ground truth point cloud, we performed a visual comparison in Fig. 7. The experiment + +![](images/99024c5e76bd76188745f79fb386841e6f430fbdbc4d3fdb375431b3c24a1d94.jpg) +Figure 6. Visualization results at $4\times$ on PUGAN with random noise and Laplace noise. + +Table 6. The results of $4 \times$ at high-level $(\tau = 0.05$ and $\tau = 0.1)$ Laplace noise on PUGAN. + +
Methodsτ = 0.1τ = 0.05
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.94332.64626.9651.86127.56324.737
MPU[44]1.56920.32925.1641.42119.04922.783
PU-GAN[15]1.48614.91924.1481.33414.24321.879
PU-GCN[27]1.58715.71125.1341.42514.91522.876
Grad-PU[8]1.23310.34818.5761.0659.34916.591
RepKpU[30]1.34513.10422.2951.19412.37720.218
PUDM[29]0.7046.28418.2830.6437.92416.871
Ours0.5634.05116.2870.5334.43316.195
+ +used an input with 2048 points, followed by a $16\times$ upsampling. The Ball-pivoting algorithm [3] was then applied to reconstruct the surface from the upsampled point cloud. Based on the reconstruction results, our generated point clouds exhibit fewer gaps and smoother surfaces, which benefits surface reconstruction. + +![](images/f139f4ab33bd845ac6d7800d3687d1a12602956b02768c63f1bff37a25962eb3.jpg) +Figure 7. Visual evaluation of impacts on surface construction on real-scanned ScanObjectNN dataset. + +# 5.6. Ablation Study + +With/Without ANS Module To demonstrate the effectiveness of the ANS module, we compared it with other network architectures. Model A replaces the ANS module with an encoder composed of Dense Blocks [8], Model B uses an encoder based on Set Abstraction [29], and Model C employs an encoder from PointNet++ [26]. Our network is referred to as Model D. We tested these models on the PU-GAN dataset with Gaussian noise level $\tau = 0.1$ , with results presented in Table7. + +Table 7. The results of $4 \times$ at high-level $(\tau = 0.1)$ Gaussian noise on PUGAN with different encoders. + +
ModuleCD↓HD↓P2F↓
A1.2489.63119.657
B0.8536.34816.849
C1.16814.49823.189
D0.5434.32716.476
+ +With/Without TT Module We conducted $16 \times$ upsampling tests on the PUGAN dataset, as shown in Table 8. The experiments indicate a significant performance drop when the TT module is removed, underscoring the importance of the TT module in enhancing the quality of the model generation. + +Table 8. TT Module Ablation Study + +
MethodsCD↓HD↓P2F↓
Without TT Module0.0941.1242.158
With TT Module0.0781.0462.028
+ +# 6. Conclusion + +We propose a novel conditional diffusion model-based network, PDANS, designed for point cloud upsampling. Suppressing noise and ensuring the geometric authenticity of generated results have long been primary research foci in point cloud upsampling. To address these issues, we have designed the ANS and TT modules. The ANS module integrates point cloud features with neighborhood information, enabling adaptive noise removal from the point cloud. Meanwhile, the TT module effectively captures local and global geometric details, allowing for a more accurate reconstruction of geometries close to real values. Extensive experiments validate that the proposed PDANS outperforms existing state-of-the-art methods in performance. + +# References + +[1] Tenindra Abeywickrama, Muhammad Aamir Cheema, and David Taniar. K-nearest neighbors on road networks: a + +journey in experimentation and in-memory implementation. arXiv preprint arXiv:1601.01549, 2016.3 +[2] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Transactions on Graphics (TOG), 32(2):1-17, 2013. 6 +[3] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 8 +[4] Haolan Chen, Shitong Luo, Wei Hu, et al. Deep point set resampling via gradient fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):2913-2930, 2022. 1 +[5] Yaodong Cui, Ren Chen, Wenbo Chu, Long Chen, Daxin Tian, Ying Li, and Dongpu Cao. Deep learning for image and point cloud fusion in autonomous driving: A review. IEEE Transactions on Intelligent Transportation Systems, 23(2): 722-739, 2021. 1 +[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 6 +[7] Wanquan Feng, Jin Li, Hongrui Cai, Xiaonan Luo, and Juyong Zhang. Neural points: Point cloud representation with neural fields for arbitrary upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18642, 2022. 2 +[8] Yun He, Danhang Tang, Yinda Zhang, Xiangyang Xue, and Yanwei Fu. Grad-pu: Arbitrary-scale point cloud upsampling via gradient descent with learned distance functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2023. 1, 2, 6, 7, 8 +[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2 +[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3 +[11] F. Y. Hu, C. L. Li, T. Zhou, H. F. Cheng, and M. M. Gu. A survey on point cloud completion algorithms for deep learning. Journal of Image and Graphics, 30(02):0309-0333, 2025. 2 +[12] Sheng Yu Huang, Hao-Yu Hsu, and Frank Wang. Spovt: Semantic-prototype variational transformer for dense point cloud semantic completion. Advances in Neural Information Processing Systems, 35:33934–33946, 2022. 2 +[13] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[14] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2 +[15] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7203-7212, 2019. 1, 2, 6, 7, 8 + +[16] Ruihui Li, Xianzhi Li, Pheng-Ann Heng, and Chi-Wing Fu. Point cloud upsampling via disentangled refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 344–353, 2021. 2 +[17] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3d object reconstruction. In proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1 +[18] Kangcheng Liu, Yuzhi Zhao, Zhi Gao, and Ben M Chen. Weaklabel3d-net: A complete framework for real-scene lidar point clouds weakly supervised multi-tasks understanding. In 2022 international conference on robotics and automation (ICRA), pages 5108-5115. IEEE, 2022. 1 +[19] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2837-2845, 2021. 2 +[20] Zhaoyang Lyu, Zhifeng Kong, Xudong Xu, Liang Pan, and Dahua Lin. A conditional point diffusion-refinement paradigm for 3d point cloud completion. arXiv preprint arXiv:2112.03530, 2021. 2, 3, 4 +[21] Aihua Mao, Zihui Du, Junhui Hou, Yaqi Duan, Yong-jin Liu, and Ying He. Pu-flow: A point cloud upsampling network with normalizing flows. IEEE Transactions on Visualization and Computer Graphics, 29(12):4964-4977, 2022. 2 +[22] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12923-12932, 2023. 1 +[23] Lucas Nunes, Rodrigo Marcuzzi, Benedikt Mersch, Jens Behley, and Cyril Stachniss. Scaling diffusion models to real-world 3d lidar scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14770-14780, 2024. 2 +[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2 +[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4, 8 +[27] Guocheng Qian, Abdulellah Abualshour, Guohao Li, Ali Thabet, and Bernard Ghanem. Pu-gcn: Point cloud upsampling using graph convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11683-11692, 2021. 1, 2, 6, 7, 8 +[28] Shi Qiu, Saeed Anwar, and Nick Barnes. Pu-transformer: Point cloud upsampling transformer. In Proceedings of the + +Asian conference on computer vision, pages 2475-2493, 2022. 1 +[29] Wentao Qu, Yuantian Shao, Lingwu Meng, Xiaoshui Huang, and Liang Xiao. A conditional denoising diffusion probabilistic model for point cloud upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20786-20795, 2024. 2, 6, 7, 8 +[30] Yi Rong, Haoran Zhou, Kang Xia, Cheng Mei, Jiahao Wang, and Tong Lu. Repkpu: Point cloud upsampling with kernel point representation and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21050-21060, 2024. 1, 2, 6, 7, 8 +[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4 +[32] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 2, 4 +[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3 +[34] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[35] Hanxiao Tan. Dam: Diffusion activation maximization for 3d global explanations. arXiv preprint arXiv:2401.14938, 2024. 2 +[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2 +[37] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 7 +[38] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. Advances in Neural Information Processing Systems, 34:29088-29100, 2021. 2 +[39] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5499-5509, 2021. 1 +[40] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In + +Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5589-5598, 2020. 1 +[41] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3323–3332, 2019. 3 +[42] Lei Yang, Yanhong Liu, Jinzhu Peng, and Zize Liang. A novel system for off-line 3d seam extraction and path planning based on point cloud segmentation for arc welding robot. Robotics and Computer-Integrated Manufacturing, 64:101929, 2020. 1 +[43] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 3 +[44] Wang Yifan, Shihao Wu, Hui Huang, Daniel Cohen-Or, and Olga Sorkine-Hornung. Patch-based progressive 3d point set upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5958–5967, 2019. 2, 6, 7, 8 +[45] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2790–2799, 2018. 1, 2, 6, 7, 8 +[46] Cem Yuksel. Sample elimination for generating poisson disk sample sets. In Computer Graphics Forum, pages 25-32. Wiley Online Library, 2015. 6 +[47] Yuchao Zheng, Yujie Li, Shuo Yang, and Huimin Lu. Global-pbnet: A novel point cloud registration for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(11):22312-22319, 2022. 1 \ No newline at end of file diff --git a/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/images.zip b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7542a21befb5506d50124779e3f07c832f06c4c3 --- /dev/null +++ b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0e056c2fa67ea747fff9360b3f4d0463c027aa6bfdb43b6e433de32858d4638 +size 618106 diff --git a/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/layout.json b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5d1674cd08c01df0abc8c461e2f74f62329d7470 --- /dev/null +++ b/2025/Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression/layout.json @@ -0,0 +1,10425 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "type": "text", + "content": "Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 160, + 487, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 160, + 487, + 190 + ], + "spans": [ + { + "bbox": [ + 127, + 160, + 487, + 190 + ], + "type": "text", + "content": "Boqian Zhang\\*, Shen Yang\\*, Hao Chen, Chao Yang, Jing Jia, Guang Jiang† \nXidian University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 191, + 493, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 191, + 493, + 203 + ], + "spans": [ + { + "bbox": [ + 116, + 191, + 493, + 203 + ], + "type": "text", + "content": "{24011211294，20009101459，haochen(cv，24011211295}@xidian.stu.edu.cn," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 168, + 205, + 440, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 205, + 440, + 217 + ], + "spans": [ + { + "bbox": [ + 168, + 205, + 440, + 217 + ], + "type": "text", + "content": "jjia@mail.xidian.edu.cn, gjiang@mail.xidian.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 270, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 571 + ], + "type": "text", + "content": "Point cloud upsampling can improve the quality of the initial point cloud, significantly enhancing the performance of downstream tasks such as classification and segmentation. Existing methods mostly focus on generating the geometric details of point clouds, neglecting noise suppression. To address this, we propose a novel network based on a conditional diffusion model, incorporating the Adaptive Noise Suppression (ANS) module, which we refer to as PDANS. The ANS module assigns weights to each point and determines the removal strategy based on these weights, reducing the impact of noisy points on the sampling process. The module first selects the neighborhood set for each point in the point cloud and performs a weighted sum between the point and its neighbors. It then adjusts the removal points based on the weighted sum, effectively mitigating the bias caused by outliers. We introduce the TreeTrans (TT) module to capture more correlated feature information. This module learns the interaction between high-level and low-level features, resulting in a more comprehensive and refined feature representation. Our results on several widely used benchmark datasets demonstrate that PDANS exhibits exceptional robustness in noisy point cloud processing and outperforms current state-of-the-art(SOTA) methods in terms of performance. Code is available at https://github.com/Baty2023/PDANS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 592, + 137, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 137, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 137, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "content": "Point clouds, as one of the key representations for 3D reconstruction [17] [22], are widely applied in autonomous driving [5] [47], embodied intelligence [18] [42], and AR/VR. However, real-world scanned point clouds are often sparse, noisy, and uneven, negatively impacting their performance in downstream tasks such as classification and segmenta" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 246, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 294 + ], + "type": "text", + "content": "tion. Point cloud upsampling can transform sparse and noisy point clouds into dense, artifact-free representations. In recent years, significant progress has been made in this task [15] [27] [45] [40]." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 305, + 549, + 393 + ], + "blocks": [ + { + "bbox": [ + 328, + 305, + 549, + 393 + ], + "lines": [ + { + "bbox": [ + 328, + 305, + 549, + 393 + ], + "spans": [ + { + "bbox": [ + 328, + 305, + 549, + 393 + ], + "type": "image", + "image_path": "954005717d3044c550c2828f584847ee24c018e74bc14fa9b0731a959fa5970f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 406, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 449 + ], + "type": "text", + "content": "Figure 1. Most existing methods achieve satisfactory results when input without noise, but they often struggle to effectively recover the geometric structure (e.g., panda ears) when dealing with input with noise." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 462, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 462, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 312, + 462, + 555, + 665 + ], + "type": "text", + "content": "Inspired by deep learning, several methods have adopted deep networks to perform this task. Existing upsampling methods can be broadly classified into two categories: those based on generative models and those based on refinement methods. Generative models [15] [27] [28] first extract features from sparse point clouds into a latent space and then design a generator based on the extracted features to generate the coordinates of dense point clouds. The limitation of such methods lies in the difficulty of accurately predicting 3D coordinates, leading to potential outliers or artifacts in the generated dense point clouds [8] [45]. Refinement-based methods [4] [39] [30] first use interpolation to generate a coarse dense point cloud, and then employ a neural network to extract geometric features from the coarse point cloud for refinement. Combined with reconstruction loss functions, this process refines the point cloud to match the ground truth." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "These methods share two common issues: (1) The features lack significant dominance or distinguishability, leading to generating point clouds without clear structure, which tend to be sparse and scattered. (2) Most of these methods" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 227, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 227, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 227, + 703 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 703, + 151, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 151, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 151, + 712 + ], + "type": "text", + "content": "†Corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16987" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "rely on loss functions such as CD loss, EMD loss, and other similar metrics, which are highly sensitive to outliers. This sensitivity makes the model ineffective at handling noisy point clouds [12] [38], as shown in Figure1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 277 + ], + "type": "text", + "content": "To address the above issues, we propose a point cloud upsampling method with adaptive noise suppression, PDANS. The network is designed based on a conditional diffusion model, leveraging the inherent denoising properties of the diffusion model's generative process to enhance its adaptability to noisy point clouds. The network also integrates an Adaptive Noise Suppression (ANS) module, which we designed to compute downsampling weights for each point by utilizing the relationships between the point and its neighboring points. This approach enables the extraction of more accurate downsampled points and features. It effectively reduces the impact of noisy points during feature extraction and minimizes the bias caused by outliers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 295, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 375 + ], + "type": "text", + "content": "To enhance feature representation, we propose a structure based on graph convolutional networks [14], called the TreeTrans (TT) module. This module incorporates ancestral and recurrent terms[32] to fully utilize the features outputted by each encoder layer. Additionally, we apply both local and global attention mechanisms to the output features, further improving the efficiency and expressiveness of the feature representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 508 + ], + "type": "text", + "content": "Additionally, we designed a conditional point cloud network and a diffusion point cloud network. The conditional point cloud network provides relevant features of sparse point clouds to the diffusion point cloud network through a mutual attention module, guiding the diffusion network to generate the required noise for point cloud densification. Similar to PUDM[29], we incorporate the diffusion model, thereby eliminating the reliance on CD loss. Combined with the ANS module we designed, PDANS demonstrates exceptional robustness in noise experiments. Our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 510, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 556 + ], + "type": "text", + "content": "- We designed a noise suppression module named ANS, which adaptively adjusts removal points and their neighborhood features to improve robustness against noise and outliers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 558, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 295, + 617 + ], + "type": "text", + "content": "- The proposed TT module fully leverages features across different layers of the encoder, incorporating ancestral terms, recurrent terms, and local and global attention mechanisms to achieve more comprehensive and refined feature representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 618, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 618, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 618, + 295, + 677 + ], + "type": "text", + "content": "- We designed an end-to-end network architecture named PDANS, which utilizes a mutual attention mechanism and a diffusion model to guide the densification of sparse point clouds. The ANS module further enhances the network's ability to handle noise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the outstanding performance of PDANS in point cloud upsampling benchmarks, particularly in generating geometric details and" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 482, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 482, + 84 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 482, + 84 + ], + "type": "text", + "content": "handling noisy point clouds effectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 96, + 400, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 400, + 109 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 400, + 109 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 117, + 524, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 524, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 524, + 131 + ], + "type": "text", + "content": "2.1. Learning-based point cloud upsampling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 134, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 553, + 302 + ], + "type": "text", + "content": "Several methods [25] [26] [36] [11] shave been developed to extract features from 3D data directly with the strong data-driven and representational capabilities of deep learning networks. PU-Net [45] was the first learning-based method for point cloud and upsampling. Following its proposed architecture, many methods have emerged: MPU [44], PUGAN [15], Dis-PU [16], PUGCN [27], and RepKPU [30]. The NePs [7] method enhances neural radiance fields by combining 2D information with 3D features, enabling point cloud upsampling. Some networks attempt to perform point cloud upsampling based on denoising paradigms. They generate a rough upsampled point cloud through interpolation, and then refine it using neural networks, such as GradPU [8] and PU-Flow [21]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 304, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 411 + ], + "type": "text", + "content": "Most methods predict the coordinates of dense point clouds from sparse point cloud features and then rely on upsampling modules to expand the point cloud, or further refine the coordinates from the coarsely densified point clouds. Since these methods typically depend on CD (Chamfer Distance) and EMD (Earth Mover's Distance) as loss functions, they exhibit high sensitivity to noise. In this paper, we design an additional denoising module and utilize conditional DDPM to reduce the dependence on CD." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 420, + 458, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 420, + 458, + 431 + ], + "spans": [ + { + "bbox": [ + 314, + 420, + 458, + 431 + ], + "type": "text", + "content": "2.2. DDPM for 3D Generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 437, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 582 + ], + "type": "text", + "content": "Inspired by the success of image generation tasks [10] [34] [9], direct point cloud generation using DDPM has gained increasing attention. Luo et al. [19] is the first to apply DDPM to unconditional point cloud generation. The DAM [35] network explores the interpretability of diffusion models for point cloud generation tasks. Nunes et al. [23] apply DDPM to the 3D-LiDAR point cloud completion task in real-world scenarios. The PDR [20] network takes the initial point cloud as input, but it introduces a coarse and refinement network, which increases time overhead. PUDM [29] densifies point clouds through conditional diffusion, but its performance in noise suppression is suboptimal." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 553, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 618 + ], + "type": "text", + "content": "This paper uses a conditional diffusion model to perform point cloud upsampling. To enhance the model's denoising capability, we design an additional ANS module." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 631, + 553, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 553, + 657 + ], + "type": "text", + "content": "3. Background of Conditional Diffusion Module" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "We assume that the dense point cloud " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " is sampled from a meaningful distribution " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "P_{data}" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ", and that the latent variable " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "x_{T} \\sim \\mathcal{N}(\\mathbf{0}_{3N},\\mathbf{I}_{3N\\times 3N})" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " is sampled from a noise distribution " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "P_{latent}" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " which is easier to handle, where " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " follows a" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16988" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "Gaussian distribution. DDPM[10] establishes a transfer relationship between " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " through two Markov chains. This is an autoregressive process: the forward process progressively adds noise to " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " until " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is denoised back to " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " through the reverse process. We use the timestep " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "t \\sim U(T)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " to control the speed of transformation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 180 + ], + "type": "text", + "content": "Forward Process The process gradually adds noise to " + }, + { + "bbox": [ + 55, + 144, + 296, + 180 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 144, + 296, + 180 + ], + "type": "text", + "content": " until " + }, + { + "bbox": [ + 55, + 144, + 296, + 180 + ], + "type": "inline_equation", + "content": "x_T" + }, + { + "bbox": [ + 55, + 144, + 296, + 180 + ], + "type": "text", + "content": " through the probability " + }, + { + "bbox": [ + 55, + 144, + 296, + 180 + ], + "type": "inline_equation", + "content": "q(x_t | x_{t-1})" + }, + { + "bbox": [ + 55, + 144, + 296, + 180 + ], + "type": "text", + "content": ". The entire process can be defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 101, + 189, + 295, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 189, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 101, + 189, + 295, + 222 + ], + "type": "interline_equation", + "content": "q \\left(x _ {1}, \\dots , x _ {T} \\mid x _ {0}\\right) = \\prod_ {t = 1} ^ {T} q \\left(x _ {t} \\mid x _ {t - 1}\\right) \\tag {1}", + "image_path": "128e5181143db9b2689f6f8bd227e6b4501449d505cbc86c72d953dcaed2f6ce.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "inline_equation", + "content": "q(x_{t}|x_{t - 1}) = \\mathcal{N}(x_{t};\\sqrt{1 - \\beta_{t}} x_{t - 1},\\beta_{t}I)" + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": ", where the hyperparameter " + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": " is a positive constant. Based on the derivation by Ho et al. [10], " + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": " can be computed directly from " + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": ". Define " + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}" + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": ", so that " + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": " can be sampled from the following formula:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 299, + 295, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 299, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 121, + 299, + 295, + 314 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon \\tag {2}", + "image_path": "625b883c31f573a8adcd4b898ce02386013176b0cce7ce4c33da836214edb67b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 322, + 249, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 249, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 249, + 333 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 322, + 249, + 333 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 322, + 249, + 333 + ], + "type": "text", + "content": " follows a standard normal distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "text", + "content": "Reverse Process The reverse process aims to predict and eliminate the noise added at each step " + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "text", + "content": ". It is conditioned on a set of conditions, which may include incomplete point clouds " + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "text", + "content": ", upsampling factor " + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "text", + "content": ", etc. We denote the set of conditions as " + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "inline_equation", + "content": "C = \\{c_i \\mid i = 1, \\dots, N_c\\}" + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "text", + "content": " represents the number of conditions. We can define the reverse process from denoising " + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 55, + 334, + 296, + 418 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 426, + 295, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 426, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 70, + 426, + 295, + 459 + ], + "type": "interline_equation", + "content": "p _ {\\boldsymbol {\\theta}} \\left(x _ {0}, \\dots , x _ {T - 1} \\mid x _ {T}, \\boldsymbol {C}\\right) = \\prod_ {t = 1} ^ {T} p _ {\\boldsymbol {\\theta}} \\left(x _ {t - 1} \\mid x _ {t}, \\boldsymbol {C}\\right) \\tag {3}", + "image_path": "b97b5a04855b3d1579ab28da36bd45a9d2624234ecfaeacebe1c11168779f426.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "type": "inline_equation", + "content": "p_{\\boldsymbol{\\theta}}(x_{t - 1}|x_t,\\boldsymbol {C}) = \\mathcal{N}(x_{t - 1};\\mu_{\\boldsymbol{\\theta}}(\\boldsymbol {x}_t,\\boldsymbol {C},t),\\sigma_t^2\\boldsymbol {I})" + }, + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "type": "text", + "content": " where the mean " + }, + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mu_{\\boldsymbol{\\theta}}(x_t,C,t)" + }, + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "type": "text", + "content": " is the result predicted by a neural network, and " + }, + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "type": "text", + "content": " is a time-dependent constant. Based on the derivation by DDIM [33], the reverse process can be written in the following form:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 82, + 536, + 295, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 536, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 82, + 536, + 295, + 586 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (x _ {t} , \\boldsymbol {C} , t)}{\\sqrt {\\bar {\\alpha} _ {t}}}\\right) + \\tag {4} \\\\ \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\epsilon_ {\\theta} (x _ {t}, C, t) + \\sigma_ {t} \\epsilon_ {t} \\\\ \\end{array}", + "image_path": "dbfeefa2a23fe066b702cd2ead9fa4071d6420d42893a2255f16e614ae676a95.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\epsilon_{t} \\sim \\mathcal{N}(\\mathbf{0}_{3N},\\mathbf{I}_{3N\\times 3N})" + }, + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "text", + "content": " is standard Gaussian noise independent of " + }, + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "text", + "content": ". The noisy point cloud " + }, + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "text", + "content": " is progressively denoised according to the above equation until it recovers " + }, + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "text", + "content": ". At each time step of the reverse process, conditions are added, allowing control over the generation of dense point clouds, as shown in the process diagram in Figure 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Training process under given conditions The goal is to optimize the denoising model's prediction of the noise " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ". Given a set of conditions " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and under the" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 319, + 73, + 550, + 169 + ], + "blocks": [ + { + "bbox": [ + 319, + 73, + 550, + 169 + ], + "lines": [ + { + "bbox": [ + 319, + 73, + 550, + 169 + ], + "spans": [ + { + "bbox": [ + 319, + 73, + 550, + 169 + ], + "type": "image", + "image_path": "4bf08cbdb1fcf68f37c1937cece8f9526363ded9fe1077aec8bc26a5170a3c84.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 180, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 201 + ], + "type": "text", + "content": "Figure 2. Visualization of the forward and reverse processes of PDANS." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 221, + 555, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 244 + ], + "type": "text", + "content": "condition that the time step " + }, + { + "bbox": [ + 313, + 221, + 555, + 244 + ], + "type": "inline_equation", + "content": "t\\sim U(T)" + }, + { + "bbox": [ + 313, + 221, + 555, + 244 + ], + "type": "text", + "content": ", the training objective of DDPM is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 251, + 560, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 560, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 560, + 277 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = \\mathbb {E} _ {\\boldsymbol {t} \\sim U (T), \\epsilon \\sim \\mathcal {N} (0, I)} \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon + \\sqrt {\\bar {\\alpha} _ {t}} x _ {0}, C, t\\right) \\| ^ {2} \\tag {5}", + "image_path": "eb1aaa5d69b24b764aa3277164d86ff13155549531be2c0bd4135e771b88e3e2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 277, + 555, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 337 + ], + "type": "text", + "content": "In the above equation, there are no CD loss or EMD loss terms commonly used in traditional point cloud upsampling networks, because DDPM naturally defines a one-to-one pointwise correspondence between two consecutive point clouds, as shown in equation1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 346, + 370, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 346, + 370, + 358 + ], + "spans": [ + { + "bbox": [ + 314, + 346, + 370, + 358 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 366, + 554, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 438 + ], + "type": "text", + "content": "This section introduces each module's design and the network's overall structure. In PDANS, we propose two modules: the Adaptive Noise Suppression (ANS) module and the TreeTrans (TT) module. These modules are described in Sections 4.1 and 4.2, respectively. The design of the overall PDANS network is presented in Section 4.3." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 445, + 535, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 535, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 535, + 458 + ], + "type": "text", + "content": "4.1. Adaptive Noise Suppression(ANS) Module" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 462, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 594 + ], + "type": "text", + "content": "In the encoding stage, many frameworks perform down-sampling using Farthest Point Sampling (FPS), followed by selecting neighboring points within a certain radius [20] or via KNN [1] to enhance the features of the sampled points. However, two issues arise when the input point cloud contains significant noise: (1) FPS is sensitive to outliers[43][41], making it challenging to process noisy point clouds effectively. (2) FPS does not filter noisy points, leading to the inclusion of noise features in the enhanced features, which affects the accuracy and reliability of the sampling." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "We propose the Adaptive Noise Suppression (ANS) module to address the issues abovementioned, as shown in Figure 3. Unlike FPS, we first select the neighborhood point cloud " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_K\\in \\mathbb{R}^{N\\times 3\\times K}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " for all points in the initial point cloud " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " denotes the number of neighboring points for each point. These neighborhood point clouds contain both correct and noisy points. We then input the neighborhood point cloud, the initial point cloud, and their corresponding features into the ANS module to perform the removal operation." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16989" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 289, + 223 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 289, + 223 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 289, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 289, + 223 + ], + "type": "image", + "image_path": "fd530f1e0df6cb8f5f1b38e56839fad1c90b4e58fa275621700cde9d501592f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "type": "text", + "content": "Figure 3. Visualization of the Adaptive Noise Suppression (ANS) module" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": "For each point " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": " selected from " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": ", we select its corresponding feature " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{F}s \\in \\mathbb{R}^{N \\times D_l}" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{F}s" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": " is the feature set obtained by reordering the initial point cloud. We use KNN to select its " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": " nearest neighbors " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "x_{i,1}, \\ldots, x_{i,K} \\in \\mathcal{N}(x_i)" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": " and their corresponding features " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "f_{i,1}, \\ldots, f_{i,K} \\in \\mathcal{N}(f_i)" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(x_i)" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": " nearest neighbors of point " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": ". Next, we calculate the point and feature weights using an attention mechanism, with both points and features updated similarly. " + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "inline_equation", + "content": "w(f_{i})" + }, + { + "bbox": [ + 55, + 277, + 296, + 386 + ], + "type": "text", + "content": " calculation can be expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 395, + 295, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 395, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 109, + 395, + 295, + 429 + ], + "type": "interline_equation", + "content": "w \\left(f _ {i}\\right) = \\sum_ {j = 1} ^ {K} \\theta \\left(x _ {i}, x _ {i, j}\\right) \\cdot \\gamma \\left(f _ {i, j}\\right) \\tag {6}", + "image_path": "ec8082a2fdf3854e8190ed5f6b8c49dcedc63139e527e10fdb91d433f50536c4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "text", + "content": "In the equation, the function " + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "text", + "content": " is used to compute the relationship between a neighbor point and the center point. The function " + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\gamma (f_{i,j})" + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "text", + "content": " represents the linear transformation of the neighborhood feature, expressed as " + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\gamma (f_{i,j}) = W_{\\gamma}f_{i,j} + b_{\\gamma}" + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "inline_equation", + "content": "W_{\\gamma}" + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "inline_equation", + "content": "b_{\\gamma}" + }, + { + "bbox": [ + 55, + 438, + 296, + 498 + ], + "type": "text", + "content": " are learnable parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 499, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 522 + ], + "type": "text", + "content": "The function " + }, + { + "bbox": [ + 55, + 499, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 499, + 295, + 522 + ], + "type": "text", + "content": " is computed using the dot product similarity in the attention mechanism:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 533, + 295, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 533, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 80, + 533, + 295, + 548 + ], + "type": "interline_equation", + "content": "\\theta (x _ {i}, x _ {i, j}) = \\operatorname {s o f t m a x} (q (x _ {i}) \\cdot k (x _ {i, j}) ^ {T} / \\sqrt {C _ {q}}) \\quad (7)", + "image_path": "4d3a866043f1a1bf515b33194debeec7cb58530a5453e448dfb79be8383f481d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "q(x_{i})" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "k(x_{i,j})" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " represent the linear transformations of the center point " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": ", and the neighborhood points " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(x_i)" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": ", which can be implemented using a " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " convolution kernel. The input to " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "q(x_{i})" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " is the real coordinates of the center point, while the input to " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "k(x_{i,j})" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " is the relative coordinates between the neighborhood point and the center point. The convolution kernel has an output channel of " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "C_q" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " and an input channel of 3. Using real and relative coordinates, the model effectively combines global and local spatial information, enhancing its ability to perceive the 3D spatial geometry." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "We then use the computed " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "w(f_{i})" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " to refine the initial points and their features. This operation selects points with" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "more surrounding noise points as removing points, resulting in more accurate features. In other words, we can find the suitable refined point " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "x_{i}^{*}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " and its corresponding feature " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "f_{i}^{*}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ", expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 394, + 129, + 555, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 129, + 555, + 142 + ], + "spans": [ + { + "bbox": [ + 394, + 129, + 555, + 142 + ], + "type": "interline_equation", + "content": "x _ {i} ^ {*} = \\delta_ {p} \\left(w \\left(f _ {i}\\right)\\right) x _ {i}, \\tag {8}", + "image_path": "bb78bb544f4a1831ae80ba8f108d29051a0710c51400714a12265479e8edc83b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 394, + 144, + 555, + 157 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 144, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 394, + 144, + 555, + 157 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {*} = \\delta_ {f} \\left(w \\left(f _ {i}\\right)\\right) f _ {i}. \\tag {9}", + "image_path": "003e362fa71c07d959ee646a5ceac45c2980eae34b012a0ace6433b5470cde0f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "inline_equation", + "content": "\\delta_p" + }, + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "inline_equation", + "content": "\\delta_f" + }, + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "text", + "content": " is activation function. Finally, we fuse the points " + }, + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "inline_equation", + "content": "x_i^*" + }, + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "text", + "content": " and their corresponding features " + }, + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "inline_equation", + "content": "f_i^*" + }, + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "text", + "content": " through an attention module to obtain the module's final output features " + }, + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "inline_equation", + "content": "f_{out}" + }, + { + "bbox": [ + 313, + 163, + 554, + 212 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 219, + 446, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 446, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 446, + 231 + ], + "type": "text", + "content": "4.2. TreeTrans(TT) Module" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 236, + 554, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 554, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 554, + 368 + ], + "type": "text", + "content": "In the decoding process of encoded point clouds, previous methods often begin with three rounds of interpolation, followed by feature fusion using neighboring point clouds [26], and some methods adopt point-adaptive deconvolution techniques, such as PDR [20]. These methods are typically built on UNet architecture [31], which leads to several issues: (1) The model becomes overly dependent on global features, thereby neglecting local features, and (2) The utilization of multi-scale features is not comprehensive enough, leading to suboptimal performance when handling complex point clouds." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 368, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 464 + ], + "type": "text", + "content": "We propose the TreeTrans (TT) module as our decoding network to address the above issues. We first employ two operations similar to TreeGCN[32]: the cyclic and the ancestor operations, which propagate features from the root node to the leaf nodes. After obtaining the intermediate features " + }, + { + "bbox": [ + 313, + 368, + 554, + 464 + ], + "type": "inline_equation", + "content": "f_{tree}" + }, + { + "bbox": [ + 313, + 368, + 554, + 464 + ], + "type": "text", + "content": ", we combine " + }, + { + "bbox": [ + 313, + 368, + 554, + 464 + ], + "type": "inline_equation", + "content": "f_{tree}" + }, + { + "bbox": [ + 313, + 368, + 554, + 464 + ], + "type": "text", + "content": " with the corresponding features from the encoder using local and global attention mechanisms to obtain the output " + }, + { + "bbox": [ + 313, + 368, + 554, + 464 + ], + "type": "inline_equation", + "content": "f_{out}" + }, + { + "bbox": [ + 313, + 368, + 554, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "text", + "content": "In traditional graph convolutional networks, the cyclic operation uses a separate weight to transform the current node feature " + }, + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "inline_equation", + "content": "f_{i}^{l}" + }, + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "text", + "content": " to the next layer. However, a single weight often fails to capture all the relevant information in point clouds. Thus, we employ a " + }, + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "text", + "content": "-support-based cyclic operation to refine the feature representation progressively:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 400, + 542, + 553, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 542, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 400, + 542, + 553, + 558 + ], + "type": "interline_equation", + "content": "S _ {i} ^ {l + 1} = F _ {K} ^ {l} \\left(f _ {i} ^ {l}\\right) \\tag {10}", + "image_path": "e683553325030c3a74bedd64ef369bb20918bea0b3c16b95933ba32a74ea8667.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "inline_equation", + "content": "F_{K}^{l}" + }, + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "text", + "content": " is a fully connected layer containing " + }, + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "text", + "content": " nodes, and " + }, + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "inline_equation", + "content": "S_{i}^{l + 1}" + }, + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "text", + "content": " represents the cyclic operation at layer " + }, + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 564, + 554, + 637 + ], + "type": "text", + "content": ". The cyclic operation is then combined with the ancestor operation to update the node features, ensuring that each layer incorporates information from previous layers during computation. The updated node features can be written as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 356, + 643, + 553, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 643, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 356, + 643, + 553, + 682 + ], + "type": "interline_equation", + "content": "f _ {t r e e} = \\sigma \\left(S _ {i} ^ {l + 1} + \\sum_ {m _ {j} \\in A \\left(f _ {i} ^ {l}\\right)} U _ {j} ^ {l} m _ {j}\\right) \\tag {11}", + "image_path": "fb5ebc9a6c1bb2c3b9816cd1e5d87adf4527074effe68b94b0142550098aacef.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " is the activation function, with ReLU used in this paper. " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "A(f_i^l)" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " denotes the set of all ancestor nodes, and " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "U_j^l" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 784, + 317, + 792 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 784, + 317, + 792 + ], + "spans": [ + { + "bbox": [ + 294, + 784, + 317, + 792 + ], + "type": "text", + "content": "16990" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 75, + 558, + 319 + ], + "blocks": [ + { + "bbox": [ + 58, + 75, + 558, + 319 + ], + "lines": [ + { + "bbox": [ + 58, + 75, + 558, + 319 + ], + "spans": [ + { + "bbox": [ + 58, + 75, + 558, + 319 + ], + "type": "image", + "image_path": "ba9457d5dc27e8fb2895c6d890b0921ac41a4268e2c56dd99122196ebbe1e684.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 327, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 327, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 555, + 361 + ], + "type": "text", + "content": "Figure 4. Overview of PDANS. The CP network (upper branch) and DP network (lower branch) both consist of an encoder made up of feature extraction and the ANS module, with the TT module serving as the decoder. They exchange information through a mutual attention module. the upsampled ratio and interpolated points are used as conditions to provide prior knowledge of the DP network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 382, + 295, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 295, + 404 + ], + "type": "text", + "content": "the linear transformation weight from ancestor nodes to the current node." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 406, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 296, + 466 + ], + "type": "text", + "content": "After obtaining " + }, + { + "bbox": [ + 55, + 406, + 296, + 466 + ], + "type": "inline_equation", + "content": "f_{tree}" + }, + { + "bbox": [ + 55, + 406, + 296, + 466 + ], + "type": "text", + "content": ", we input it into the local attention and global attention modules. In the local attention mechanism, the corresponding encoder output features " + }, + { + "bbox": [ + 55, + 406, + 296, + 466 + ], + "type": "inline_equation", + "content": "f^{l+1}" + }, + { + "bbox": [ + 55, + 406, + 296, + 466 + ], + "type": "text", + "content": " are combined with " + }, + { + "bbox": [ + 55, + 406, + 296, + 466 + ], + "type": "inline_equation", + "content": "f_{tree}" + }, + { + "bbox": [ + 55, + 406, + 296, + 466 + ], + "type": "text", + "content": " using local attention, as described by the following formula:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 473, + 295, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 473, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 65, + 473, + 295, + 510 + ], + "type": "interline_equation", + "content": "f _ {\\text {l o c a l}} = \\operatorname {s o f t m a x} \\left(\\frac {Q \\left(f ^ {l + 1}\\right) \\cdot K \\left(f _ {\\text {t r e e}}\\right) ^ {T}}{\\sqrt {d}}\\right) \\cdot V \\left(f _ {\\text {t r e e}}\\right) \\tag {12}", + "image_path": "3a167dea46da23648760a08be3cd6be911d1c07c76801949a344e4c73f6bcd38.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "inline_equation", + "content": "Q(f^{l + 1})" + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "inline_equation", + "content": "K(f_{tree})" + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "text", + "content": " represent two convolution operations that map " + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "inline_equation", + "content": "f^{l + 1}" + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "inline_equation", + "content": "f_{tree}" + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "text", + "content": " from their respective dimensions to " + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "text", + "content": ". Then, the attention weights are obtained by calculating the dot product similarity followed by a softmax operation. Afterward, the weighted sum with " + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "inline_equation", + "content": "V(f_{tree})" + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "text", + "content": " gives the local attention feature " + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "inline_equation", + "content": "f_{local}" + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "text", + "content": ". Once the local attention is computed, the resulting " + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "inline_equation", + "content": "f_{local}" + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "text", + "content": " undergoes a global attention mechanism to produce the final output " + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "inline_equation", + "content": "f_{out}" + }, + { + "bbox": [ + 55, + 510, + 296, + 619 + ], + "type": "text", + "content": " of the TT module. The specific implementation is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 625, + 296, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 625, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 61, + 625, + 296, + 654 + ], + "type": "interline_equation", + "content": "f _ {o u t} = \\operatorname {s o f t m a x} \\left(\\frac {q (f _ {l o c a l}) \\cdot k (f _ {l o c a l}) ^ {T}}{\\sqrt {d}}\\right) \\cdot v (f _ {l o c a l}) \\tag {13}", + "image_path": "f4b46612b65db13fb2962b44e0e751f0fe901de6cdcfd47b7342e138b6b05e0d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 181, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 181, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 181, + 672 + ], + "type": "text", + "content": "4.3. Network Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In Sections 4.1 and 4.2, we introduced the network's encoding module, ANS, and decoding module, TT, respectively. In this section, we present the overall framework of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 382, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 555, + 430 + ], + "type": "text", + "content": "PDANS, which consists of two main components: the Conditional Point Cloud (CP) network and the Diffusion Point Cloud (DP) network. Figure 4 provides a detailed illustration of this process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "text", + "content": "CP-Net Our encoder consists of a feature extraction network and the ANS module. The decoder comprises the TT module. Both downsampling and upsampling are performed four times. Before each sampling output, we embed the time step " + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "text", + "content": " through an MLP: " + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "text", + "content": " is encoded into a " + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "text", + "content": "-dimensional vector space. The output feature " + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "inline_equation", + "content": "f_{out} \\in \\mathbb{R}^{N \\times D_l}" + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "text", + "content": " is then transformed through MLPs into the dimension " + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "inline_equation", + "content": "f_{out}' \\in \\mathbb{R}^{N \\times C_t}" + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "text", + "content": ". We then add the two to integrate the time step information into the feature, resulting in the output " + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "inline_equation", + "content": "f_{l+1} \\in \\mathbb{R}^{N \\times D_{l+1}}" + }, + { + "bbox": [ + 313, + 434, + 556, + 578 + ], + "type": "text", + "content": ". Additionally, we only input the sparse point cloud into the CP-Net to ensure the extraction of effective features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "text", + "content": "DP-Net The DP-Net shares the same architecture as the CP-Net but requires additional conditional information to guide the diffusion process. First, we convert the sparse point cloud " + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "text", + "content": " into an interpolated point cloud " + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^* \\in \\mathbb{R}^{rN \\times 3}" + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "text", + "content": " using upsampling by a factor of " + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "text", + "content": ". Then, we sum " + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^*" + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "text", + "content": " with the noisy point cloud " + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_T \\in \\mathbb{R}^{rN \\times 3}" + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "text", + "content": " and input them together into the DP-Net. Meanwhile, we extract global features from " + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^*" + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "text", + "content": " and encode both the time step " + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "text", + "content": " and the upsampling factor " + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 582, + 556, + 712 + ], + "type": "text", + "content": ". These pieces of information are incorporated as conditions at each stage of the encoder and decoder." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 784, + 317, + 792 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 784, + 317, + 792 + ], + "spans": [ + { + "bbox": [ + 294, + 784, + 317, + 792 + ], + "type": "text", + "content": "16991" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "The information provided by the CP-Net is further guided to the DP-Net through a cross-attention (CA) module. Given the outputs of the encoder at each layer for the CP-Net and DP-Net: " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "F_{c} \\in \\mathbb{R}^{N_{c} \\times D_{c}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "F_{d} \\in \\mathbb{R}^{N_{d} \\times D_{d}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", the CA module first reshapes the two feature sets using MLPs: " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "F_{c} \\rightarrow (Q) \\in \\mathbb{R}^{N_{c} \\times C_{i}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "F_{d} \\rightarrow (K, V) \\in \\mathbb{R}^{N_{d} \\times C_{i}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". Then, the features are fused according to the following formula:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 163, + 295, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 163, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 93, + 163, + 295, + 191 + ], + "type": "interline_equation", + "content": "F _ {f} = M L P \\left(\\text {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {C _ {i}}}\\right) V\\right) + F _ {c} \\tag {14}", + "image_path": "3b4ee774e2e7cdd7db16f7389908c5b24c4252d9b6ec1c8f3e48693d633bc8f6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 296, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 245 + ], + "type": "text", + "content": "Subsequently, " + }, + { + "bbox": [ + 55, + 198, + 296, + 245 + ], + "type": "inline_equation", + "content": "F_{f}" + }, + { + "bbox": [ + 55, + 198, + 296, + 245 + ], + "type": "text", + "content": " is input into the feedforward network to produce the final features. The same method can also be applied reversely, allowing data to flow between the two networks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 252, + 101, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 101, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 101, + 264 + ], + "type": "text", + "content": "4.4. Loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "content": "For the DP-Net, its loss can be referred to by equation 5. For the CP-Net, we aim for it to learn more features of the sparse point cloud. Therefore, we add MLPs at the end of the decoder to output a sparse point cloud " + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{P}'" + }, + { + "bbox": [ + 55, + 270, + 296, + 330 + ], + "type": "text", + "content": ", which is then supervised by MSE as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 338, + 295, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 338, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 110, + 338, + 295, + 353 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} \\left(\\mathcal {P}, \\mathcal {P} ^ {\\prime}\\right) = \\mathbf {M S E} \\| \\mathcal {P} - \\mathcal {P} ^ {\\prime} \\| ^ {2} \\tag {15}", + "image_path": "8759efbcb06d45ffe83facb94c6ba0fa33ae95b6cc7ecfc1bd73bf44949aa4cd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 361, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 289, + 373 + ], + "type": "text", + "content": "Combining these loss functions, The training objective is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 381, + 295, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 381, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 129, + 381, + 295, + 395 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} (\\theta) + \\alpha \\mathcal {L} _ {C} \\tag {16}", + "image_path": "f4640d9d7d61a75250942131d62fdbc5da6a2e7186531ade3e6a640725f38318.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 403, + 288, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 288, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 403, + 288, + 415 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 403, + 288, + 415 + ], + "type": "text", + "content": " means a weighting factor (" + }, + { + "bbox": [ + 55, + 403, + 288, + 415 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 55, + 403, + 288, + 415 + ], + "type": "text", + "content": " in this paper)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 425, + 137, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 137, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 137, + 438 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 445, + 188, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 188, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 188, + 457 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 462, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 533 + ], + "type": "text", + "content": "Experiment Setup. All PDANS experiments are run in a PyTorch[24] environment. The PDANS model is trained on a single Nvidia 4080 GPU using the Adam optimizer[13], with a batch size 12 for 350 epochs. The channel sizes for the CP-Net and DP-Net are set to (64, 128, 256, 512) and (128, 256, 256, 512), respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 534, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 665 + ], + "type": "text", + "content": "Dataset. In our experiments, we utilize two public benchmarks (PUGAN[15], PU1K[27]) for evaluation. We adhere to the official training/testing partitioning protocols for these datasets. Poisson disk sampling [46] is used to generate 24,000 patches on PUGAN and 69,000 patches on PU1K for training. Each patch contains 256 points, with the corresponding ground truth containing 1,024 points. Additionally, 27 and 127 point cloud datasets are used for testing, respectively. The sparse point clouds contain 2,048 points, and the number of points in ground truth is set to " + }, + { + "bbox": [ + 55, + 534, + 295, + 665 + ], + "type": "inline_equation", + "content": "2,048 \\times r" + }, + { + "bbox": [ + 55, + 534, + 295, + 665 + ], + "type": "text", + "content": " according to the upsampling factor " + }, + { + "bbox": [ + 55, + 534, + 295, + 665 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 534, + 295, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Evaluation Metrics Following previous methods, we select three metrics to evaluate our model: Chamfer Distance " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{CD} \\times 10^{-3})" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " [6], Hausdorff distance " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{HD} \\times 10^{-3})" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " [2] and Point to Surface Distance " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{P2F} \\times 10^{-3})" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " [45]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 482, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 482, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 482, + 83 + ], + "type": "text", + "content": "5.2. Evaluation on PUGAN Dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "type": "text", + "content": "We perform sampling on PUGAN at low " + }, + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "type": "inline_equation", + "content": "(4\\times)" + }, + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "type": "text", + "content": " and high " + }, + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "type": "inline_equation", + "content": "(16\\times)" + }, + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "type": "text", + "content": " upsampling rates. The reductions in the three metrics shown in Table 1 demonstrate that our model has a significant advantage in describing geometric details. The " + }, + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "type": "inline_equation", + "content": "16\\times" + }, + { + "bbox": [ + 313, + 89, + 555, + 174 + ], + "type": "text", + "content": " upsampled point cloud shown in Figure 5 further supports our claim, as the generated point clouds have fewer outliers and are closer to the ground truth geometric features." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 321, + 214, + 548, + 320 + ], + "blocks": [ + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "type": "text", + "content": "Table 1. The results of " + }, + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 313, + 183, + 555, + 206 + ], + "type": "text", + "content": " on PUGAN. Our method significantly surpasses other methods in terms of CD and HD." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 214, + 548, + 320 + ], + "lines": [ + { + "bbox": [ + 321, + 214, + 548, + 320 + ], + "spans": [ + { + "bbox": [ + 321, + 214, + 548, + 320 + ], + "type": "table", + "html": "
Methods4×16×
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]0.5296.8054.4600.5108.2066.041
MPU[44]0.2926.6722.8220.2197.0543.085
PU-GAN[15]0.2825.5772.0160.2076.9632.556
PU-GCN[27]0.2683.2012.4890.1614.2832.632
Grad-PU[8]0.2452.3691.8930.1082.3522.127
RepKpU[30]0.2542.6981.8820.1073.2931.957
PUDM[29]0.1311.2201.9120.0821.1202.114
Ours0.1191.20651.8970.0781.0462.028
", + "image_path": "a4cb4da4cde9acbcb678177000946036ffd874af0246194043e22b74e4e54c4a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 347, + 471, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 471, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 471, + 359 + ], + "type": "text", + "content": "5.3. Evaluation on PU1K Dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "text", + "content": "Table 2 shows the results of " + }, + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "inline_equation", + "content": "4\\mathrm{x}" + }, + { + "bbox": [ + 313, + 364, + 555, + 449 + ], + "type": "text", + "content": " upsampling on the PU1K dataset. Compared to other methods, we demonstrate superior performance in both CD and HD metrics. The PU1K dataset is more complex and diverse than others, and our designed TT module exhibits tremendous advantages in handling such complex data. It helps approximate the valid geometric details more effectively, achieving higher accuracy." + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 325, + 468, + 544, + 559 + ], + "blocks": [ + { + "bbox": [ + 325, + 468, + 544, + 559 + ], + "lines": [ + { + "bbox": [ + 325, + 468, + 544, + 559 + ], + "spans": [ + { + "bbox": [ + 325, + 468, + 544, + 559 + ], + "type": "table", + "html": "
MethodsCD↓HD↓P2F↓
PU-Net[45]1.15515.1704.834
MPU[44]0.93513.3273.511
PU-GAN[15]0.7359.9394.234
PU-GCN[27]0.5857.5772.499
Grad-PU[8]0.4043.7321.474
RepKpU[30]0.3312.7001.261
PUDM[29]0.2172.1641.477
Ours0.1551.5881.354
", + "image_path": "fc0a27375a169f06c6a61ae491f89d1b91928ec4c7102d00fa2773096c1c3ad0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 567, + 555, + 589 + ], + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 589 + ], + "type": "text", + "content": "Table 2. The results of " + }, + { + "bbox": [ + 313, + 567, + 555, + 589 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 567, + 555, + 589 + ], + "type": "text", + "content": " on PU1K. Our method outperforms other methods across nearly all metrics." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 612, + 411, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 411, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 411, + 624 + ], + "type": "text", + "content": "5.4. Robustness Test" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Experiment setup. Based on the testing meshes provided by PUGAN, we construct a " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " upsampling dataset to conduct a comprehensive robustness test. Ground-truth point clouds are sampled from testing meshes via Poisson disk sampling. We tested three different types of noise: Gaussian, random, and Laplacian. These three types of noise are the most common in real-world point clouds, making" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16992" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 549, + 202 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 549, + 202 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 549, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 549, + 202 + ], + "type": "image", + "image_path": "7d2686eb0d5d4fbae318fb58e35b1ad43b1825fd43872fc998f86357d9bf504e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 210, + 394, + 222 + ], + "lines": [ + { + "bbox": [ + 214, + 210, + 394, + 222 + ], + "spans": [ + { + "bbox": [ + 214, + 210, + 394, + 222 + ], + "type": "text", + "content": "Figure 5. Comparison " + }, + { + "bbox": [ + 214, + 210, + 394, + 222 + ], + "type": "inline_equation", + "content": "(16\\times)" + }, + { + "bbox": [ + 214, + 210, + 394, + 222 + ], + "type": "text", + "content": " on PUGAN dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "text", + "content": "the results more practical. We set four different noise levels for Gaussian noise: " + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "inline_equation", + "content": "\\tau = 0.01" + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "inline_equation", + "content": "\\tau = 0.02" + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "inline_equation", + "content": "\\tau = 0.05" + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "text", + "content": ". We used two noise levels for Laplacian and random noise: " + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "inline_equation", + "content": "\\tau = 0.05" + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 55, + 243, + 296, + 313 + ], + "type": "text", + "content": ". For the input with 2048 points, we tested the results under each noise type and noise level." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 411 + ], + "type": "text", + "content": "The results of Gaussian noise Tables 3 and 4 present the quantitative results of Gaussian noise testing under different noise levels. As shown in the tables, our method significantly outperforms other models in terms of resistance to Gaussian noise. Thanks to the design of the ANS module in our model, which further enhances noise resistance. Therefore, our approach excels at noise suppression and surpasses other methods in both CD and HD metrics." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 63, + 463, + 288, + 566 + ], + "blocks": [ + { + "bbox": [ + 55, + 421, + 296, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 455 + ], + "type": "text", + "content": "Table 3. The results of " + }, + { + "bbox": [ + 55, + 421, + 296, + 455 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 55, + 421, + 296, + 455 + ], + "type": "text", + "content": " at low-level " + }, + { + "bbox": [ + 55, + 421, + 296, + 455 + ], + "type": "inline_equation", + "content": "(\\tau = 0.01" + }, + { + "bbox": [ + 55, + 421, + 296, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 421, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\tau = 0.02)" + }, + { + "bbox": [ + 55, + 421, + 296, + 455 + ], + "type": "text", + "content": " Gaussian noise on PUGAN. Our method significantly outperforms other methods in terms of Gaussian noise." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 463, + 288, + 566 + ], + "lines": [ + { + "bbox": [ + 63, + 463, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 63, + 463, + 288, + 566 + ], + "type": "table", + "html": "
Methodsτ = 0.01τ = 0.02
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]0.6288.0689.8161.07810.86716.401
MPU[44]0.5066.9789.0590.92910.82015.621
PU-GAN[15]0.4646.0707.4980.88710.60215.088
PU-GCN[27]0.4485.5866.9890.8168.60413.798
Grad-PU[8]0.4144.1456.4000.7667.33611.534
RepKpU[30]0.3844.9376.7750.8008.41514.124
PUDM[29]0.2102.4306.0700.5295.4719.742
Ours0.1851.9185.9640.3093.0209.513
", + "image_path": "a74324edbe9343ecc5bd97b1607931465f1580b19641f1ba6d0407c8730e4657.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "The results of random noise and Laplacian noise Tables 5 and 6 present the quantitative test results under noise levels " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tau = 0.05" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ". The results clearly show that the PDANS network we designed exhibits superior adaptability in high-noise environments. Figure 6 illustrates that, in noisy point clouds, our results show more complete structures with fewer outliers. Our method captures more geometric details, leading to higher-quality point clouds. In contrast, other methods tend to upsample blindly, resulting in point clouds with disorganized structures. While our method still maintains SOTA performance, it shows slightly" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 321, + 283, + 548, + 385 + ], + "blocks": [ + { + "bbox": [ + 313, + 242, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 275 + ], + "type": "text", + "content": "Table 4. The results of " + }, + { + "bbox": [ + 313, + 242, + 555, + 275 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 242, + 555, + 275 + ], + "type": "text", + "content": " at high-level " + }, + { + "bbox": [ + 313, + 242, + 555, + 275 + ], + "type": "inline_equation", + "content": "(\\tau = 0.05" + }, + { + "bbox": [ + 313, + 242, + 555, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 242, + 555, + 275 + ], + "type": "inline_equation", + "content": "\\tau = 0.1)" + }, + { + "bbox": [ + 313, + 242, + 555, + 275 + ], + "type": "text", + "content": " Gaussian noise on PUGAN. Our method significantly outperforms other methods in terms of Gaussian noise." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 283, + 548, + 385 + ], + "lines": [ + { + "bbox": [ + 321, + 283, + 548, + 385 + ], + "spans": [ + { + "bbox": [ + 321, + 283, + 548, + 385 + ], + "type": "table", + "html": "
Methodsτ = 0.05τ = 0.1
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.37013.72923.2491.49814.19323.846
MPU[44]1.24711.64522.1891.32112.41523.841
PU-GAN[15]1.1249.09121.2521.27110.91123.174
PU-GCN[27]1.2639.86922.8351.45611.06325.213
Grad-PU[8]0.9788.05716.9271.1188.94618.845
RepKpU[30]1.23712.33520.5181.36012.90222.473
PUDM[29]0.6185.38614.7510.8536.23916.845
Ours0.4764.03814.6070.5434.32716.476
", + "image_path": "b6460960edf226353a4ff0f523bcc414f5b64e9e328a0bb38115626fe0126f77.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 403, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 488 + ], + "type": "text", + "content": "lower results than Gaussian noise. The experimental results validate the effectiveness of the ANS module we designed. By removing noisy point clouds, we can better learn the objects' structural information, thereby generating high-quality point clouds. As a result, PDANS demonstrates good robustness under different types and levels of noise, proving its stability in complex noisy environments." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 321, + 529, + 548, + 629 + ], + "blocks": [ + { + "bbox": [ + 313, + 498, + 555, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 519 + ], + "type": "text", + "content": "Table 5. The results of " + }, + { + "bbox": [ + 313, + 498, + 555, + 519 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 498, + 555, + 519 + ], + "type": "text", + "content": " at high-level (" + }, + { + "bbox": [ + 313, + 498, + 555, + 519 + ], + "type": "inline_equation", + "content": "\\tau = 0.05" + }, + { + "bbox": [ + 313, + 498, + 555, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 555, + 519 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 313, + 498, + 555, + 519 + ], + "type": "text", + "content": ") random noise on PUGAN." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 529, + 548, + 629 + ], + "lines": [ + { + "bbox": [ + 321, + 529, + 548, + 629 + ], + "spans": [ + { + "bbox": [ + 321, + 529, + 548, + 629 + ], + "type": "table", + "html": "
Methodsτ = 0.05τ = 0.1
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.49014.47323.2231.72515.44225.251
MPU[44]1.22410.84220.4561.54511.64523.512
PU-GAN[15]1.0347.75718.6171.3279.70021.321
PU-GCN[27]1.0459.64318.8991.32510.87721.633
Grad-PU[8]1.0676.63417.7341.3997.21521.028
RepKpU[30]0.7894.78217.8130.8373.90220.953
PUDM[29]0.9986.11017.5581.3106.73220.564
Ours0.7223.16117.6340.6824.59819.401
", + "image_path": "7e0a99a0d26b3cd6705d7befbae2a5a44cf370ddd9cef38571cf1da929bc8789.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 648, + 443, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 443, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 443, + 659 + ], + "type": "text", + "content": "5.5. Result on Real Dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "We conducted experiments on the ScanObjectNN [37] dataset, where point clouds are real-scanned, non-uniform and noisy. Since there is no ground truth point cloud, we performed a visual comparison in Fig. 7. The experiment" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16993" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 545, + 175 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 545, + 175 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 545, + 175 + ], + "type": "image", + "image_path": "99024c5e76bd76188745f79fb386841e6f430fbdbc4d3fdb375431b3c24a1d94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 189, + 462, + 200 + ], + "lines": [ + { + "bbox": [ + 146, + 189, + 462, + 200 + ], + "spans": [ + { + "bbox": [ + 146, + 189, + 462, + 200 + ], + "type": "text", + "content": "Figure 6. Visualization results at " + }, + { + "bbox": [ + 146, + 189, + 462, + 200 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 146, + 189, + 462, + 200 + ], + "type": "text", + "content": " on PUGAN with random noise and Laplace noise." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 251, + 288, + 350 + ], + "blocks": [ + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "type": "text", + "content": "Table 6. The results of " + }, + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "type": "text", + "content": " at high-level " + }, + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "type": "inline_equation", + "content": "(\\tau = 0.05" + }, + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "type": "inline_equation", + "content": "\\tau = 0.1)" + }, + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "type": "text", + "content": " Laplace noise on PUGAN." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 251, + 288, + 350 + ], + "lines": [ + { + "bbox": [ + 63, + 251, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 63, + 251, + 288, + 350 + ], + "type": "table", + "html": "
Methodsτ = 0.1τ = 0.05
CD↓HD↓P2F↓CD↓HD↓P2F↓
PU-Net[45]1.94332.64626.9651.86127.56324.737
MPU[44]1.56920.32925.1641.42119.04922.783
PU-GAN[15]1.48614.91924.1481.33414.24321.879
PU-GCN[27]1.58715.71125.1341.42514.91522.876
Grad-PU[8]1.23310.34818.5761.0659.34916.591
RepKpU[30]1.34513.10422.2951.19412.37720.218
PUDM[29]0.7046.28418.2830.6437.92416.871
Ours0.5634.05116.2870.5334.43316.195
", + "image_path": "7897433a5a7590e0c88cbe0aa05c932eeed8bcfe7d862966e91814b0e2cfb46f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 370, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 296, + 441 + ], + "type": "text", + "content": "used an input with 2048 points, followed by a " + }, + { + "bbox": [ + 55, + 370, + 296, + 441 + ], + "type": "inline_equation", + "content": "16\\times" + }, + { + "bbox": [ + 55, + 370, + 296, + 441 + ], + "type": "text", + "content": " upsampling. The Ball-pivoting algorithm [3] was then applied to reconstruct the surface from the upsampled point cloud. Based on the reconstruction results, our generated point clouds exhibit fewer gaps and smoother surfaces, which benefits surface reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 453, + 291, + 534 + ], + "blocks": [ + { + "bbox": [ + 61, + 453, + 291, + 534 + ], + "lines": [ + { + "bbox": [ + 61, + 453, + 291, + 534 + ], + "spans": [ + { + "bbox": [ + 61, + 453, + 291, + 534 + ], + "type": "image", + "image_path": "f139f4ab33bd845ac6d7800d3687d1a12602956b02768c63f1bff37a25962eb3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "type": "text", + "content": "Figure 7. Visual evaluation of impacts on surface construction on real-scanned ScanObjectNN dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 149, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 149, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 149, + 601 + ], + "type": "text", + "content": "5.6. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "With/Without ANS Module To demonstrate the effectiveness of the ANS module, we compared it with other network architectures. Model A replaces the ANS module with an encoder composed of Dense Blocks [8], Model B uses an encoder based on Set Abstraction [29], and Model C employs an encoder from PointNet++ [26]. Our network is referred to as Model D. We tested these models on the PU-GAN dataset with Gaussian noise level " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ", with results presented in Table7." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 331, + 251, + 538, + 304 + ], + "blocks": [ + { + "bbox": [ + 313, + 220, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 242 + ], + "type": "text", + "content": "Table 7. The results of " + }, + { + "bbox": [ + 313, + 220, + 555, + 242 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 220, + 555, + 242 + ], + "type": "text", + "content": " at high-level " + }, + { + "bbox": [ + 313, + 220, + 555, + 242 + ], + "type": "inline_equation", + "content": "(\\tau = 0.1)" + }, + { + "bbox": [ + 313, + 220, + 555, + 242 + ], + "type": "text", + "content": " Gaussian noise on PUGAN with different encoders." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 251, + 538, + 304 + ], + "lines": [ + { + "bbox": [ + 331, + 251, + 538, + 304 + ], + "spans": [ + { + "bbox": [ + 331, + 251, + 538, + 304 + ], + "type": "table", + "html": "
ModuleCD↓HD↓P2F↓
A1.2489.63119.657
B0.8536.34816.849
C1.16814.49823.189
D0.5434.32716.476
", + "image_path": "8494a5b2c903979e7a52d81b2e8ab67586ad65c227e730ab5e734518e085fdb7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 324, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 395 + ], + "type": "text", + "content": "With/Without TT Module We conducted " + }, + { + "bbox": [ + 313, + 324, + 555, + 395 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 313, + 324, + 555, + 395 + ], + "type": "text", + "content": " upsampling tests on the PUGAN dataset, as shown in Table 8. The experiments indicate a significant performance drop when the TT module is removed, underscoring the importance of the TT module in enhancing the quality of the model generation." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 318, + 424, + 552, + 462 + ], + "blocks": [ + { + "bbox": [ + 367, + 404, + 501, + 415 + ], + "lines": [ + { + "bbox": [ + 367, + 404, + 501, + 415 + ], + "spans": [ + { + "bbox": [ + 367, + 404, + 501, + 415 + ], + "type": "text", + "content": "Table 8. TT Module Ablation Study" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 424, + 552, + 462 + ], + "lines": [ + { + "bbox": [ + 318, + 424, + 552, + 462 + ], + "spans": [ + { + "bbox": [ + 318, + 424, + 552, + 462 + ], + "type": "table", + "html": "
MethodsCD↓HD↓P2F↓
Without TT Module0.0941.1242.158
With TT Module0.0781.0462.028
", + "image_path": "cb217a598d01c4abb5917eb25e16cab64baf4df22f34383727745ccf0dc8e59b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 487, + 388, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 388, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 388, + 499 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 506, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 650 + ], + "type": "text", + "content": "We propose a novel conditional diffusion model-based network, PDANS, designed for point cloud upsampling. Suppressing noise and ensuring the geometric authenticity of generated results have long been primary research foci in point cloud upsampling. To address these issues, we have designed the ANS and TT modules. The ANS module integrates point cloud features with neighborhood information, enabling adaptive noise removal from the point cloud. Meanwhile, the TT module effectively captures local and global geometric details, allowing for a more accurate reconstruction of geometries close to real values. Extensive experiments validate that the proposed PDANS outperforms existing state-of-the-art methods in performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 671, + 373, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 671, + 373, + 684 + ], + "spans": [ + { + "bbox": [ + 314, + 671, + 373, + 684 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 691, + 555, + 713 + ], + "type": "text", + "content": "[1] Tenindra Abeywickrama, Muhammad Aamir Cheema, and David Taniar. K-nearest neighbors on road networks: a" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16994" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "journey in experimentation and in-memory implementation. arXiv preprint arXiv:1601.01549, 2016.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 295, + 139 + ], + "type": "text", + "content": "[2] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Transactions on Graphics (TOG), 32(2):1-17, 2013. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 141, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 295, + 185 + ], + "type": "text", + "content": "[3] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 186, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 295, + 228 + ], + "type": "text", + "content": "[4] Haolan Chen, Shitong Luo, Wei Hu, et al. Deep point set resampling via gradient fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):2913-2930, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 295, + 285 + ], + "type": "text", + "content": "[5] Yaodong Cui, Ren Chen, Wenbo Chu, Long Chen, Daxin Tian, Ying Li, and Dongpu Cao. Deep learning for image and point cloud fusion in autonomous driving: A review. IEEE Transactions on Intelligent Transportation Systems, 23(2): 722-739, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 287, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 287, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 287, + 295, + 331 + ], + "type": "text", + "content": "[6] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 332, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 296, + 387 + ], + "type": "text", + "content": "[7] Wanquan Feng, Jin Li, Hongrui Cai, Xiaonan Luo, and Juyong Zhang. Neural points: Point cloud representation with neural fields for arbitrary upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18642, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 295, + 453 + ], + "type": "text", + "content": "[8] Yun He, Danhang Tang, Yinda Zhang, Xiangyang Xue, and Yanwei Fu. Grad-pu: Arbitrary-scale point cloud upsampling via gradient descent with learned distance functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 455, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 476 + ], + "type": "text", + "content": "[9] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 478, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 478, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 57, + 478, + 295, + 510 + ], + "type": "text", + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 512, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 512, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 295, + 554 + ], + "type": "text", + "content": "[11] F. Y. Hu, C. L. Li, T. Zhou, H. F. Cheng, and M. M. Gu. A survey on point cloud completion algorithms for deep learning. Journal of Image and Graphics, 30(02):0309-0333, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "type": "text", + "content": "[12] Sheng Yu Huang, Hao-Yu Hsu, and Frank Wang. Spovt: Semantic-prototype variational transformer for dense point cloud semantic completion. Advances in Neural Information Processing Systems, 35:33934–33946, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 624 + ], + "type": "text", + "content": "[13] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 657 + ], + "type": "text", + "content": "[14] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[15] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7203-7212, 2019. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[16] Ruihui Li, Xianzhi Li, Pheng-Ann Heng, and Chi-Wing Fu. Point cloud upsampling via disentangled refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 344–353, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "text", + "content": "[17] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3d object reconstruction. In proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "text", + "content": "[18] Kangcheng Liu, Yuzhi Zhao, Zhi Gao, and Ben M Chen. Weaklabel3d-net: A complete framework for real-scene lidar point clouds weakly supervised multi-tasks understanding. In 2022 international conference on robotics and automation (ICRA), pages 5108-5115. IEEE, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[19] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2837-2845, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "text", + "content": "[20] Zhaoyang Lyu, Zhifeng Kong, Xudong Xu, Liang Pan, and Dahua Lin. A conditional point diffusion-refinement paradigm for 3d point cloud completion. arXiv preprint arXiv:2112.03530, 2021. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "text", + "content": "[21] Aihua Mao, Zihui Du, Junhui Hou, Yaqi Duan, Yong-jin Liu, and Ying He. Pu-flow: A point cloud upsampling network with normalizing flows. IEEE Transactions on Visualization and Computer Graphics, 29(12):4964-4977, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 355, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 555, + 409 + ], + "type": "text", + "content": "[22] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12923-12932, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 411, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 555, + 465 + ], + "type": "text", + "content": "[23] Lucas Nunes, Rodrigo Marcuzzi, Benedikt Mersch, Jens Behley, and Cyril Stachniss. Scaling diffusion models to real-world 3d lidar scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14770-14780, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 468, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 531 + ], + "type": "text", + "content": "[24] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 534, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 587 + ], + "type": "text", + "content": "[25] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "text", + "content": "[26] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 4, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 635, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 555, + 689 + ], + "type": "text", + "content": "[27] Guocheng Qian, Abdulellah Abualshour, Guohao Li, Ali Thabet, and Bernard Ghanem. Pu-gcn: Point cloud upsampling using graph convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11683-11692, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[28] Shi Qiu, Saeed Anwar, and Nick Barnes. Pu-transformer: Point cloud upsampling transformer. In Proceedings of the" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16995" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "type": "text", + "content": "Asian conference on computer vision, pages 2475-2493, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 297, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 297, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 297, + 151 + ], + "type": "text", + "content": "[29] Wentao Qu, Yuantian Shao, Lingwu Meng, Xiaoshui Huang, and Liang Xiao. A conditional denoising diffusion probabilistic model for point cloud upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20786-20795, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "text", + "content": "[30] Yi Rong, Haoran Zhou, Kang Xia, Cheng Mei, Jiahao Wang, and Tong Lu. Repkpu: Point cloud upsampling with kernel point representation and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21050-21060, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 274 + ], + "type": "text", + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 330 + ], + "type": "text", + "content": "[32] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3859-3868, 2019. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 364 + ], + "type": "text", + "content": "[33] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 409 + ], + "type": "text", + "content": "[34] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 443 + ], + "type": "text", + "content": "[35] Hanxiao Tan. Dam: Diffusion activation maximization for 3d global explanations. arXiv preprint arXiv:2401.14938, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 499 + ], + "type": "text", + "content": "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 566 + ], + "type": "text", + "content": "[37] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "text", + "content": "[38] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. Advances in Neural Information Processing Systems, 34:29088-29100, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 677 + ], + "type": "text", + "content": "[39] Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5499-5509, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[40] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 441 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5589-5598, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "text", + "content": "[41] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3323–3332, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[42] Lei Yang, Yanhong Liu, Jinzhu Peng, and Zize Liang. A novel system for off-line 3d seam extraction and path planning based on point cloud segmentation for arc welding robot. Robotics and Computer-Integrated Manufacturing, 64:101929, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[43] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "text", + "content": "[44] Wang Yifan, Shihao Wu, Hui Huang, Daniel Cohen-Or, and Olga Sorkine-Hornung. Patch-based progressive 3d point set upsampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5958–5967, 2019. 2, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "text", + "content": "[45] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-net: Point cloud upsampling network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2790–2799, 2018. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 364, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 553, + 397 + ], + "type": "text", + "content": "[46] Cem Yuksel. Sample elimination for generating poisson disk sample sets. In Computer Graphics Forum, pages 25-32. Wiley Online Library, 2015. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "text", + "content": "[47] Yuchao Zheng, Yujie Li, Shuo Yang, and Huimin Lu. Global-pbnet: A novel point cloud registration for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(11):22312-22319, 2022. 1" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16996" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/7aa2860d-04df-426a-bde9-40b0ed00d4aa_content_list.json b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/7aa2860d-04df-426a-bde9-40b0ed00d4aa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b2020637aef5f6d500bb0ecb85607aeba07913b4 --- /dev/null +++ b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/7aa2860d-04df-426a-bde9-40b0ed00d4aa_content_list.json @@ -0,0 +1,2046 @@ +[ + { + "type": "text", + "text": "Point Clouds Meets Physics: Dynamic Acoustic Field Fitting Network for Point Cloud Understanding", + "text_level": 1, + "bbox": [ + 101, + 128, + 897, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changshuo Wang", + "bbox": [ + 197, + 204, + 339, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NTU, Singapore", + "bbox": [ + 202, + 222, + 336, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wangchangshuol@gmail.com", + "bbox": [ + 161, + 242, + 377, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuting He", + "bbox": [ + 472, + 204, + 566, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SUFE, China", + "bbox": [ + 467, + 222, + 573, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shuting.he@sufe.edu.cn", + "bbox": [ + 421, + 242, + 619, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiang Fang*", + "bbox": [ + 702, + 204, + 803, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NTU, Singapore", + "bbox": [ + 684, + 222, + 815, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xfang9508@gmail.com", + "bbox": [ + 665, + 242, + 834, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiawei Han", + "bbox": [ + 160, + 263, + 251, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "BIT, China", + "bbox": [ + 161, + 282, + 251, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hanjiawei@bit.edu.cn", + "bbox": [ + 116, + 301, + 295, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhonghang Liu", + "bbox": [ + 388, + 263, + 517, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SMU, Singapore", + "bbox": [ + 387, + 282, + 521, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhliu.2020@phdcs.smu.edu.sg", + "bbox": [ + 334, + 301, + 573, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Ning", + "bbox": [ + 599, + 263, + 785, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Institute of Semiconductors, CAS", + "bbox": [ + 612, + 282, + 880, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ningxin@semi.ac.cn", + "bbox": [ + 666, + 303, + 826, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weijun Li \nInstitute of Semiconductors, CAS", + "text_level": 1, + "bbox": [ + 218, + 324, + 486, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wjli@semi.ac.cn", + "bbox": [ + 284, + 363, + 419, + 376 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prayag Tiwari \nHalmstad University, Sweden", + "text_level": 1, + "bbox": [ + 540, + 324, + 777, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prayag.tiwari@ieee.org", + "bbox": [ + 558, + 364, + 754, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 412, + 326, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While existing pre-training-based methods have enhanced point cloud model performance, they have not fundamentally resolved the challenge of local structure representation in point clouds. The limited representational capacity of pure point cloud models continues to constrain the potential of cross-modal fusion methods and performance across various tasks. To address this challenge, we propose a Dynamic Acoustic Field Fitting Network (DAF-Net), inspired by physical acoustic principles. Specifically, we represent local point clouds as acoustic fields and introduce a novel Acoustic Field Convolution (AF-Conv), which treats local aggregation as an acoustic energy field modeling problem and captures fine-grained local shape awareness by dividing the local area into near field and far field. Furthermore, drawing inspiration from multifrequency wave phenomena and dynamic convolution, we develop the Dynamic Acoustic Field Convolution (DAF-Conv) based on AF-Conv. DAF-Conv dynamically generates multiple weights based on local geometric priors, effectively enhancing adaptability to diverse geometric features. Additionally, we design a Global Shape-Aware (GSA) layer incorporating EdgeConv and multi-head attention mechanisms, which combines with DAF-Conv to form the DAF Block. These blocks are then stacked to create a hierarchical DAFNet architecture. Extensive experiments demonstrate that DAFNet significantly outperforms existing methods across multiple tasks.", + "bbox": [ + 86, + 446, + 483, + 856 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ec06a8d1a7adeb879f68d13b5ef6ea34882045597c6c6a5cc6446db0b3077b9.jpg", + "image_caption": [ + "Local Neighborhood", + "Multi-Acoustic Energy Field Mapping", + "Shape-aware Representation", + "Figure 1. Illustration of local feature extraction using DAF-Conv. Given a local neighborhood of a point cloud (left), our method first models it as an acoustic field with near-field and far-field zones. Then, multiple acoustic energy field mappings are generated to capture various geometric patterns. Finally, DAF-Conv processes these mappings to produce shape-aware representations." + ], + "image_footnote": [], + "bbox": [ + 516, + 411, + 895, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 614, + 643, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud data, due to its direct representation of object surface information in three-dimensional space, has been widely applied in various scenarios including autonomous driving [6, 54, 72], robotic navigation [14, 45], and augmented reality [8, 43]. However, the sparsity, irregularity, and complexity of local structures in point cloud data make it challenging for existing models to fully capture geometric features, thus limiting their applications in 3D tasks.", + "bbox": [ + 509, + 641, + 906, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, several methods [33, 37, 38] have attempted to indirectly improve point cloud model performance through pre-training strategies or cross-modal contrastive learning. Although these approaches have enhanced point cloud processing tasks to some extent, they have not directly addressed the problem of local structure representation. Some researchers [19, 21, 25, 50] have also attempted to apply new architectures to point cloud understanding tasks. However, the local representational capacity of point cloud mod", + "bbox": [ + 509, + 763, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* denotes corresponding author.", + "bbox": [ + 107, + 886, + 274, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22182", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els remains a bottleneck in improving overall model performance. Therefore, how to efficiently capture and represent the local structures of point clouds has become a crucial problem that urgently needs to be solved.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physics, particularly the principles of acoustic field propagation [5], provides a new perspective for understanding local geometric structures. In acoustics, wave propagation is influenced by environmental geometric characteristics, and through constructing acoustic field models, we can effectively simulate wave behavior at different frequencies to obtain information in complex environments. Inspired by this, we propose a novel point cloud understanding method called Dynamic Acoustic Field Fitting Network (DAFNet). This method borrows ideas from physical acoustic field propagation, representing local point clouds as acoustic fields and dynamically adjusting weights in different regions to capture local structural features more precisely.", + "bbox": [ + 89, + 161, + 482, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose a novel Acoustic Field Convolution (AF-Conv) that models local aggregation as an acoustic energy propagation problem. To more finely model the contributions of different neighboring points to the center point, AF-Conv divides the local region into near-field and far-field zones for separate geometric modeling. Furthermore, as shown in Fig. 1, based on AF-Conv, we propose Dynamic Acoustic Field Convolution (DAF-Conv), which dynamically generates multiple weights based on geometric priors while simulating geometric feature variations under different \"frequencies\", thereby enhancing the model's adaptability to local geometric diversity. Finally, we combine a Global Shape-Aware (GSA) layer comprising EdgeConv and Multi-Head Attention (MHA) with DAF-Conv to build hierarchical DAF Blocks, forming the complete DAFNet. The experimental results show that DAFNet achieves superior performance over existing methods in multiple tasks, demonstrating its effectiveness in capturing complex geometric structures.", + "bbox": [ + 89, + 381, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this work are as follows:", + "text_level": 1, + "bbox": [ + 109, + 676, + 468, + 691 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a physics-inspired end-to-end DAFNet that can learn rich contextual information and precise shape representations in irregular point clouds.", + "- We introduce a novel Acoustic Field Convolution (AFConv) that achieves fine-grained local shape awareness by modeling local point cloud regions as acoustic fields.", + "- We design Dynamic Acoustic Field Convolution (DAFConv) that adapts to geometric feature diversity by dynamically generating multiple frequency-aware weights from local geometry.", + "- The proposed DAFNet achieves state-of-the-art performance in multiple tasks, reaching $96.7\\%$ accuracy on the challenging ScanObjectNN with only 2.47M parameters." + ], + "bbox": [ + 91, + 702, + 482, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. CNN-based Methods", + "text_level": 1, + "bbox": [ + 513, + 114, + 709, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early works [30, 42, 56, 63] in point cloud processing primarily involved projecting them onto 2D images for 2D CNN processing or converting them into regular 3D grid structures for 3D convolution operations. However, these transformation processes inevitably led to spatial information loss and increased computational costs. To address this issue, researchers began applying deep learning techniques directly in 3D space. PointNet [35] pioneered this direction. However, this method overlooked local structural information. Subsequently, PointNet++ [36] enhanced the local perception capability of point cloud models by grouping point clouds into different local neighborhoods. Based on PointNet++ [36], existing works [23, 49, 51, 52, 55, 66] have developed methods based on point-wise perceptrons, convolution operations, and attention mechanisms. Nevertheless, these methods still struggle to effectively model local structural shape information.", + "bbox": [ + 511, + 137, + 906, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Transformer-based Methods", + "text_level": 1, + "bbox": [ + 513, + 405, + 767, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers have demonstrated remarkable success in various tasks [9, 10, 13, 53] due to their inherent ability to capture global features. Leveraging this strength, researchers have progressively developed a range of Transformer-based point cloud models. PCT [17] enhances input embeddings through furthest point sampling and nearest neighbor search while using self-attention mechanisms to capture global features. Point Transformer [71] employs vector attention to extract aggregated local features. Subsequently, updated versions [57, 58] of Point Transformer explored different architectures to improve performance and efficiency. Inspired by Swin Transformer [27], works such as SST [12] and SWFormer [46] explored window attention mechanisms to improve efficiency. Although these methods excel at capturing global features, their performance hasn't shown significant advancement over CNN-based approaches.", + "bbox": [ + 511, + 429, + 906, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. State Space Model-based Methods", + "text_level": 1, + "bbox": [ + 513, + 681, + 812, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State Space Models (SSMs) [16] are mathematical models for describing dynamic systems. Recently, SSMs have emerged as a reliable alternative architecture to CNNs and Transformers. In computer vision, Vision Mamba [76] successfully compresses visual representations through bidirectional state space models, while VMamba [26] innovatively introduces cross-scanning modules, enabling global receptive fields in 2D images through 1D selective scanning. Researchers have explored SSM applications in point cloud processing. PCM [69] first introduced the Mamba architecture to point cloud processing through octree serialization strategies. Mamba3D [19] enhanced local geometric feature extraction through LNP blocks and designed bidi", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22183", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a56044a4d3b301eff35015c04723e07ca735b3db485842c06c6f6f8a22e0a8da.jpg", + "image_caption": [ + "Figure 2. Architecture of DAFNet. (a) For shape classification, the network consists of three sequential DAF Blocks followed by max pooling and classifier layers. Each DAF Block contains a GSA layer and DAF-Conv for feature extraction. (b) For part segmentation and few-shot semantic segmentation tasks, the network adopts an encoder-decoder structure with skip connections, where the encoder shares the same architecture as the classification network, and the decoder uses feature propagation to restore point cloud resolution. The few-shot version additionally incorporates support and query mechanisms for processing different point cloud inputs." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 883, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rectional SSM structures. PointMamba [25], another significant work, customized SSM architecture for point cloud analysis tasks, substantially reducing computational costs.", + "bbox": [ + 88, + 523, + 486, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Pre-training-based Methods", + "text_level": 1, + "bbox": [ + 89, + 575, + 344, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point cloud pre-training methods primarily comprise generative approaches [4, 33, 67] and cross-modal contrastive learning methods. In generative approaches, Point-MAE [33] extends the MAE concept by pre-training through random masking and reconstruction of point cloud segments PCP-MAE [70] learns shape priors by predicting missing regional point cloud structures. PointDif [74] utilizes diffusion models for pre-training, learning hierarchical geometric priors through point cloud recovery at different noise levels. In cross-modal contrastive learning, ReCon [37] combines generative modeling and contrastive learning paradigms, addressing representation overfitting in contrastive learning and data padding issues in generative methods. ULIP [61] achieves cross-modal knowledge transfer through a language-image-point cloud pre-training framework, while ACT [11] employs cross-modal autoencoders as teacher models to acquire knowledge from other modalities. Although these pre-training methods have advanced point cloud processing performance, they haven't fundamentally resolved the point cloud representation challenge.", + "bbox": [ + 89, + 598, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 513, + 521, + 612, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section first reviews point convolution and dynamic convolution, followed by introducing acoustic field convolution. Then, we present our proposed dynamic acoustic field convolution. Finally, we describe DAFNet (as shown in Fig. 2).", + "bbox": [ + 511, + 547, + 906, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background", + "text_level": 1, + "bbox": [ + 511, + 630, + 645, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Point Convolution", + "text_level": 1, + "bbox": [ + 511, + 652, + 687, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Local feature aggregation in point clouds typically follows a \"sampling-grouping-aggregation\" procedure. The aggregation operation is crucial as it directly affects local feature representation by modeling local point clouds. We denote center points as $\\{(p_i, f_i)\\}_{i=1}^M$ , where $M$ represents the number of points, $p_i \\in \\mathbb{R}^{1 \\times 3}$ and $f_i \\in \\mathbb{R}^{1 \\times C}$ represent the coordinates and features of the $i$ -th point respectively, and $C$ denotes the number of feature channels. The most commonly used method for local feature aggregation based on coordinate space is to construct point convolution, which can be formalized as:", + "bbox": [ + 511, + 671, + 908, + 837 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {\\prime} = \\mathcal {A} (\\{\\mathcal {M} (p _ {i}, p _ {j}) \\cdot \\mathcal {T} (f _ {i}, f _ {j}) | p _ {j} \\in \\mathcal {N} (p _ {i}) \\}), \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 845, + 906, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{A}$ represents the aggregation function, typically max-pooling. $\\mathcal{M}$ and $\\mathcal{T}$ are mapping functions, usually imple", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22184", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mented through Multi-Layer Perceptrons (MLPs). $\\mathcal{N}(p_i)$ represents the local neighborhood of $p_i$ , and $p_j$ denotes the neighboring points of $p_i$ .", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An alternative aggregation approach operates in feature space, commonly known as Dynamic Graph Convolution (DGC) [55]. It aggregates local features in feature space, meaning that points close in feature space might be distant in coordinate space. DGC can be formalized as:", + "bbox": [ + 89, + 136, + 483, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {\\prime} = \\mathcal {A} (\\{\\mathcal {T} (f _ {i}, f _ {j}) | f _ {j} \\in \\mathcal {N} (f _ {i}) \\}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 223, + 482, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2. Dynamic Convolution", + "text_level": 1, + "bbox": [ + 89, + 251, + 290, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic convolution [62] aims to enhance network modeling capability by dynamically generating multiple convolution kernels based on input data, adapting to different data structures. Taking PACConv's [60] dynamic convolution as an example, dynamic convolution in point clouds can be expressed as:", + "bbox": [ + 89, + 270, + 483, + 361 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} = \\mathcal {A} \\left(\\left\\{w \\left(p _ {j}\\right) \\cdot f _ {j} \\mid p _ {j} \\in \\mathcal {N} \\left(p _ {i}\\right) \\right\\}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 373, + 482, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w(p_j)$ denotes the weight corresponding to feature $f_j$ of point $p_j$ . $w(p_j)$ typically consists of multiple weights $w_h(p_i)$ and their corresponding attention coefficients $\\alpha_h(p_j)$ generated in a data-driven manner:", + "bbox": [ + 89, + 400, + 483, + 463 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw \\left(p _ {j}\\right) = \\sum_ {h = 1} ^ {H} \\alpha_ {h} \\left(p _ {j}\\right) \\cdot w _ {h} \\left(p _ {j}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 473, + 482, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sum_{h=1}^{H} \\alpha_h(p_j) = 1$ , $H$ represents the number of $w_h(p_j)$ .", + "bbox": [ + 89, + 526, + 483, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, formula 3 can be expressed as:", + "bbox": [ + 109, + 560, + 387, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} = \\mathcal {A} \\left(\\left\\{\\sum_ {h = 1} ^ {H} \\alpha_ {h} \\left(p _ {j}\\right) \\cdot w _ {h} \\left(p _ {j}\\right) \\cdot f _ {j} \\mid p _ {j} \\in \\mathcal {N} \\left(p _ {i}\\right) \\right\\}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 585, + 482, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Acoustic Field Convolution", + "text_level": 1, + "bbox": [ + 89, + 647, + 334, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In physical acoustics [22, 31], the sound pressure field $p(x)$ generated by a sound source at spatial point $x$ can be represented by Green's function:", + "bbox": [ + 89, + 667, + 483, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (x) = \\frac {e ^ {- j k \\| x - y \\|}}{4 \\pi \\| x - y \\|}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 724, + 482, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $y$ is the source position, $k$ is the wave number, and $\\| x - y\\|$ represents the distance between observation point $x$ and source $y$ . This formula reveals the fundamental physical laws of sound wave propagation: sound pressure intensity decays exponentially with distance while being affected by spherical spreading attenuation.", + "bbox": [ + 89, + 763, + 483, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, based on the distance between the source and observation point, the acoustic field can be divided into near-field and far-field regions. In the near-field region, the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/35d3576f695b2400d227bcea514c6e0feece7fa5a39904e756b17e6bd478a5d1.jpg", + "image_caption": [ + "Figure 3. Illustration of AF-Conv and DAF-Conv. (a) AF-Conv divides local neighborhood into near-field and far-field regions and calculates energy weights based on geometric priors. (b) DAF-Conv extends AF-Conv with multiple parallel acoustic field convolutions to achieve dynamic feature aggregation." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "relationship between sound pressure and distance is more complex, exhibiting strong local correlations; in the farfield region, the sound pressure demonstrates simple spherical wave attenuation characteristics.", + "bbox": [ + 511, + 359, + 903, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by this theory, we view local feature aggregation as acoustic field propagation, where the center point $p_i$ can be viewed as the sound source and neighboring points $p_j$ as observation points. Meanwhile, we refine the local neighborhood into near-field and far-field regions to extract fine-grained information, constructing the Acoustic Field Convolution (AF-Conv). As shown in Fig. 3(a), it can be formalized as:", + "bbox": [ + 511, + 421, + 905, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} = \\mathcal {A} \\left(\\left\\{W _ {A F} \\left(p _ {i}, p _ {j}\\right) \\cdot f _ {j} \\mid p _ {j} \\in \\mathcal {N} \\left(p _ {i}\\right) \\right\\}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 553, + 903, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the weight function $W_{AF}(p_i,p_j)$ directly draws inspiration from acoustic field propagation characteristics and designs different weight calculation methods for near-field and far-field regions:", + "bbox": [ + 511, + 580, + 905, + 641 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {A F} \\left(p _ {i}, p _ {j}\\right) = \\left\\{ \\begin{array}{l l} e ^ {- \\beta d _ {i j}}, & \\text {i f} d _ {i j} < d _ {\\text {t h r e s h o l d}} \\\\ \\frac {1}{d _ {i j} + \\epsilon}, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 651, + 903, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ represents learnable coefficient. $d_{ij}$ denotes the Euclidean distance between $p_j$ and $p_i$ . In $W_{AF}(p_i,p_j)$ , for the near-field region ( $d_{ij} < d_{threshold}$ ), we use the exponential decay term $e^{-\\beta d_{ij}}$ to capture subtle changes in local geometric structures, corresponding to the complex attenuation characteristics of the acoustic near-field region; for the far-field region ( $d_{ij} \\geq d_{threshold}$ ), we adopt a simple reciprocal relationship $\\frac{1}{d_{ij} + \\epsilon}$ to better perceive broader contextual information.", + "bbox": [ + 511, + 700, + 906, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dynamic Acoustic Field Convolution", + "text_level": 1, + "bbox": [ + 511, + 848, + 831, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In sound wave propagation [31], waves of different frequencies exhibit distinct propagation characteristics. Inspired by", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22185", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this multi-frequency phenomenon and dynamic convolution [62], as shown in Fig. 3(b), we propose Dynamic Acoustic Field Convolution (DAF-Conv), which extends AF-Conv by dynamically generating multiple sets of weights to simulate local geometric feature variations under different \"frequencies\". It can be formalized as:", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {h} \\left(p _ {i}, p _ {j}\\right) = W _ {h} \\left(p _ {i}, p _ {j}\\right) \\cdot f _ {j}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 191, + 482, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {h} \\left(p _ {i}, p _ {j}\\right) = \\left\\{ \\begin{array}{l l} \\alpha_ {h} e ^ {- \\beta_ {h} d _ {i j}}, & \\text {i f} d _ {i j} < d _ {\\text {t h r e s h o l d}} \\\\ \\frac {\\alpha_ {h}}{d _ {i j} + \\epsilon}, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 218, + 482, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\alpha_ {h}, \\beta_ {h} \\right\\} _ {h = 1} ^ {H} = \\operatorname {M L P} \\left(h _ {j}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 263, + 482, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_{threshold}$ is a scalar that defines the boundary between near-field and far-field regions. The weight coefficient $\\alpha_{h}$ and $\\beta_{h}$ are generated from local geometric prior $h_j = [p_i,p_j,p_j - p_i,\\| p_j - p_i\\| ]$ through MLP to simulate propagation characteristics of different frequency components.", + "bbox": [ + 89, + 287, + 483, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The local aggregated features are then separated into near-field and far-field components based on the distance threshold:", + "bbox": [ + 89, + 377, + 483, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {h} ^ {\\text {n e a r}} \\left(p _ {i}\\right) = \\mathcal {A} \\left(\\left\\{F _ {h} \\left(p _ {i}, p _ {j}\\right) \\mid j \\in \\mathcal {N} _ {\\text {n e a r}} (i) \\right\\}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 446, + 482, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {h} ^ {f a r} \\left(p _ {i}\\right) = \\mathcal {A} \\left(\\left\\{F _ {h} \\left(p _ {i}, p _ {j}\\right) \\mid j \\in \\mathcal {N} _ {f a r} (i) \\right\\}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 474, + 482, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{N}_{near}(i) = \\{j\\mid p_j\\in \\mathcal{N}(p_i),d_{ij} < d_{threshold}\\}$ represents the set of near-field neighbors of the center point $p_i$ and $\\mathcal{N}_{far}(i) = \\{j\\mid p_j\\in \\mathcal{N}(p_i),d_{ij}\\geq d_{threshold}\\}$ represents the set of far-field neighbors.", + "bbox": [ + 89, + 497, + 482, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The component for each frequency $h$ can be obtained by:", + "bbox": [ + 109, + 559, + 482, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {h} \\left(p _ {i}\\right) = F _ {h} ^ {\\text {n e a r}} \\left(p _ {i}\\right) + F _ {h} ^ {\\text {f a r}} \\left(p _ {i}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 585, + 482, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve adaptive fusion of features from different frequency components, we introduce an attention mechanism based on geometric prior $h_j$ . The attention weights are generated as fellow:", + "bbox": [ + 89, + 609, + 483, + 670 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {h} = \\operatorname {s o f t m a x} (\\operatorname {M a x P o o l i n g} (\\operatorname {M L P} (h _ {j}))). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 685, + 482, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final local aggregated feature is obtained by weighted summation of features from all frequency components:", + "bbox": [ + 89, + 707, + 483, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} = \\sum_ {h = 1} ^ {H} \\gamma_ {h} \\cdot F _ {h} \\left(p _ {i}\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 763, + 482, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DAF-Conv achieve adaptive modeling of local geometric features through separately processing near-field and far-field features for each frequency component, independent max-pooling aggregation for near-field and far-field features, and adaptive fusion of multiple frequency components through learned attention weights.", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared to the single weight function in AF-Conv, DAF-Conv achieves more sophisticated feature extraction capabilities through its multi-frequency design and separate processing of near-field and far-field features.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Dynamic Acoustic Field Convolution Neural Network", + "text_level": 1, + "bbox": [ + 511, + 162, + 906, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure each point has rich feature representation before constructing local neighborhoods, we build a Global Shape-Aware (GSA) layer before applying DAF-Conv. The GSA layer first aggregates each point through EdgeConv [55] without reducing point cloud resolution, then applies Multi-Head Attention (MHA) for global context modeling on the enhanced points. This layer effectively enhances DAF-Conv's perception ability for local structure modeling.", + "bbox": [ + 511, + 200, + 905, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, based on the GSA layer and DAF-Conv, we construct the DAF Block. By stacking these blocks, we design a novel pretraining-free network, DAFNet, for point cloud analysis. We implement models for three major tasks:", + "bbox": [ + 511, + 321, + 905, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point Cloud Classification. For point cloud classification, we employ three DAF Blocks, followed by a classifier consisting of max-pooling and fully connected layers. The output feature dimensions of the three blocks are 64, 128, 256 respectively. Notably, with only 2.47M parameters, it achieves the best performance on ScanObjectNN [48].", + "bbox": [ + 511, + 398, + 905, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Part Segmentation and Few-shot Point Cloud Semantic Segmentation. For part segmentation, the encoder structure remains identical to the classification task. In the decoder stage, we employ feature propagation to restore point cloud resolution. Between the encoder and decoder, we apply U-Net-like skip connections to fully utilize contextual information. In few-shot segmentation, while maintaining the same encoder-decoder structure as part segmentation, we incorporate the Query module from SegPN[77] to reduce domain gaps between query and support features.", + "bbox": [ + 511, + 489, + 906, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 655, + 648, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.3D Shape Classification", + "text_level": 1, + "bbox": [ + 511, + 681, + 732, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Real-world Object Classification on ScanObjectNN. We evaluate our method on ScanObjectNN [48], a challenging real-world point cloud dataset containing 2,902 objects across 15 categories. Following previous works, we conduct experiments on its most challenging variant (PB_T50_RS), where objects are presented with background and occlusion.", + "bbox": [ + 511, + 703, + 905, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Table 1, DAFNet achieves state-of-the-art performance on ScanObjectNN [48], reaching $96.7\\%$ overall accuracy (OA) and $96.3\\%$ mean class accuracy (mAcc) with voting mechanism. These results significantly surpass both existing supervised methods (e.g., GPSFormer [52]: $95.4\\%$ OA) and pre-training approaches (e.g., ReCon++", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22186", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ae65ce5a5890dc0da1f96ce041b5ca8baf1417a2e06f6238216baa944c7bf167.jpg", + "table_caption": [ + "Table 1. Classification results on the ScanObjectNN [48] dataset. \" - \" denotes unknown." + ], + "table_footnote": [], + "table_body": "
MethodYearmAcc (%)OA (%)
Supervised Methods
PointNet [35]201763.468.2
PointNet++ [36]201869.873.7
DGCNN [55]201973.678.1
SimpleView [15]2021-80.5
MVTN [18]202183.185.5
PointMLP [28]202284.485.7
RepSurf-U [40]202283.186.0
PointNeXt [39]202285.887.7
Point-PN [68]2023-87.1
DeLA [2]202389.390.4
GPSFormer [52]202493.895.4
Mamba3D [19]2024-92.6
Pre-training Methods
Point-BERT [65]2021-83.1
Point-MAE [33]2022-85.2
ULIP-2 + PointNeXt [61]202391.291.5
PointGPT [3]2023-93.6
ReCon [41]2023-91.3
ReCon++ [38]2024-95.25
PointMamba [25]2024-89.3
PCP-MAE [70]2024-90.4
Ours
DAFNet (w/o voting)-96.196.4
DAFNet (w/ voting)-96.396.7
", + "bbox": [ + 93, + 127, + 480, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[38]: $95.25\\%$ OA). The superior performance demonstrates the effectiveness of our acoustic field-inspired architecture in handling real-world point cloud data.", + "bbox": [ + 89, + 627, + 482, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic Object Classification on ModelNet40. We further evaluate DAFNet on ModelNet40 [59], a widely-used synthetic dataset containing 12,311 CAD models from 40 categories, with 9,843 for training and 2,468 for testing.", + "bbox": [ + 89, + 680, + 482, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As presented in Table 2, DAFNet achieves competitive results on ModelNet40 [59] with $94.1\\%$ overall accuracy (OA) and $92.2\\%$ mean class accuracy (mAcc). While these results are slightly lower than some recent methods like PointGPT [3] ( $94.9\\%$ OA), the performance gap is minimal (less than $1\\%$ ). This is mainly because ModelNet40 [3], as a synthetic dataset, has been well-addressed by existing methods. In contrast, DAFNet shows more significant advantages on real-world datasets, as demonstrated by our results on ScanObjectNN [48].", + "bbox": [ + 88, + 750, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/37f9fc03dace8d9ab885bf6e5ea20f62badfb9125e48382f4335eacde8180976.jpg", + "table_caption": [ + "Table 2. Classification results on ModelNet40 dataset. \"-\" denotes unknown." + ], + "table_footnote": [], + "table_body": "
MethodsYearmAcc (%)OA (%)
Supervised Methods
PointNet [35]201786.089.2
PointNet++[36]2017-90.7
DGCNN[55]201990.292.9
Point Transformer [71]202190.693.7
PAConv [60]2021-93.9
CurveNet [32]2021-94.2
PointMLP [28]202291.494.5
RepSurf-U [40]2022-94.7
GPSFormer [52]202492.294.2
Mamba3D [19]2024-94.1
Pre-training Methods
Point-BERT [65]2021-93.8
Point-MAE [33]2022-94.0
ULIP + PointMLP [61]202392.494.7
PointGPT [3]2023-94.9
ReCon [41]2023-94.5
ReCon++ [38]2024-94.8
PointMamba [25]2024-93.6
PCP-MAE [70]2024-94.0
Ours
DAFNet (w/o voting)-91.993.8
DAFNet (w/ voting)-92.294.1
", + "bbox": [ + 516, + 128, + 903, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Part Segmentation", + "text_level": 1, + "bbox": [ + 511, + 598, + 692, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate our method's capability in fine-grained shape understanding, we conduct experiments on ShapeNetPart [64], which contains 16,881 shapes from 16 categories with 50 part annotations. Following convention, we use 13,807 samples for training and 2,874 for testing.", + "bbox": [ + 511, + 619, + 906, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 3, DAFNet achieves competitive performance with $85.2\\%$ class mIoU and $86.8\\%$ instance mIoU, comparable to state-of-the-art methods. These results demonstrate that our acoustic field-inspired approach effectively captures both global shape information and local geometric details necessary for part segmentation tasks.", + "bbox": [ + 511, + 696, + 906, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Few-shot Point Cloud Semantic Segmentation", + "text_level": 1, + "bbox": [ + 511, + 797, + 901, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. Results analysis on the S3DIS dataset", + "text_level": 1, + "bbox": [ + 511, + 820, + 821, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our model on S3DIS [1], which contains 3D point clouds from 272 rooms with 13 semantic categories. Following [73], we divide scenes into $1\\mathrm{m} \\times 1\\mathrm{m}$ blocks and sample 2048 points per block, resulting in 7547", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22187", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f8eb4dce74e3afe2a0f39310eef621f78a85a5e833a2cad495441a376f7a2e76.jpg", + "table_caption": [ + "Table 3. Part segmentation results (\\%) on ShapeNetPart. We report both class mIoU and instance mIoU." + ], + "table_footnote": [], + "table_body": "
MethodsYearclass mIoUinstance mIoU
Supervised Methods
PointNet [35]201780.483.7
PointNet++[36]201781.985.1
DGCNN [55]201982.385.1
KPConv [47]201985.186.4
Point Transformer [71]202183.786.6
CurveNet [32]2021-86.8
PointNeXt [39]202285.287.1
SPOTr [34]202385.487.2
GPSFormer [52]202485.486.8
Pre-training Methods
PointGPT[3]202384.186.2
ReCon [41]202384.886.4
PointMamba[25]202484.486.2
PCP-MAE [70]202484.986.1
Ours
DAFNet-85.286.8
", + "bbox": [ + 94, + 128, + 480, + 436 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/908f98f9e481d3c88ba0b918f6c2dc820326c5232247d5e82954bf9b8bc65e02.jpg", + "table_caption": [ + "Table 4. Few-shot Results (\\%) on S3DIS. $S_{i}$ denotes the split $i$ is used for testing. Avg is their average mIoU. The best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
Method2-way3-way
1-shot5-shot1-shot5-shot
S0S1AvgS0S1AvgS0S1AvgS0S1Avg
DGCNN [55]36.3438.7937.5756.4956.9956.7430.0532.1931.1246.8847.5747.23
ProtoNet [44]48.3949.9849.1957.3463.2260.2840.8145.0742.9449.0553.4251.24
MPTI [73]52.2751.4851.8858.9360.5659.7544.2746.9245.6051.7448.5750.16
AttMPTI [73]53.7755.9454.8661.6767.0264.3545.1849.2747.2354.9256.7955.86
BFG [29]55.6055.9855.7963.7166.6265.1746.1848.3647.2755.0557.8056.43
2CBR [75]55.8961.9958.9463.5567.5165.5346.5153.9150.2155.5158.0756.79
PAP3D [20]59.4566.0862.7665.4070.3067.8548.9956.5752.7861.2760.8161.04
Seg-PN [77]64.8467.9866.4167.6371.4869.5659.1160.4259.7759.4864.7262.10
DAFNet68.1370.2769.2070.5173.1571.8361.3365.5563.4465.2568.6766.96
Improvement+3.29+2.29+2.79+2.88+1.67+2.27+2.22+5.13+3.67+5.77+3.95+4.86
", + "bbox": [ + 93, + 505, + 480, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "blocks. As shown in Table 4, DAFNet achieves state-of-the-art performance across all few-shot settings. In 2-way experiments, our method achieves $69.20\\%$ mIoU in 1-shot and $71.83\\%$ mIoU in 5-shot settings, surpassing the previous best method Seg-PN [77] by $2.79\\%$ and $2.27\\%$ respectively. For the more challenging 3-way scenario, DAFNet maintains its superiority with $63.44\\%$ mIoU in 1-shot and $66.96\\%$ mIoU in 5-shot settings, outperforming Seg-PN [77] by $3.67\\%$ and $4.86\\%$ . These improvements across different settings demonstrate DAFNet's strong feature extraction capability and generalization ability.", + "bbox": [ + 88, + 642, + 482, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2. Results analysis on the ScanNet dataset", + "text_level": 1, + "bbox": [ + 89, + 820, + 413, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluate DAFNet on ScanNet [7], which contains 1,513 real-world scanned scenes with 20 semantic categories. Following [73], we process the scenes into 36,350 blocks with 2,048 points each. As shown in Ta", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bc41661f682d844ebfcf108f9ef57a8d3dc62ad5713f7b2218e8c8737e8739e2.jpg", + "table_caption": [ + "Table 5. Few-shot Results (\\%) on ScanNet. $S_{i}$ denotes the split $i$ is used for testing. Avg is their average mIoU. The best results are shown in bold." + ], + "table_footnote": [], + "table_body": "
Method2-way3-way
1-shot5-shot1-shot5-shot
\\( S_0 \\)\\( S_1 \\)Avg.\\( S_0 \\)\\( S_1 \\)Avg.\\( S_0 \\)\\( S_1 \\)Avg.\\( S_0 \\)\\( S_1 \\)Avg.
DGCNN [55]31.5528.9430.2542.7137.2439.9823.9919.1021.5534.9328.1031.52
ProtoNet [44]33.9230.9532.4445.3442.0143.6828.4726.1327.3037.3634.9836.17
MPTI [73]39.2736.1437.7146.9043.5945.2529.9627.2628.6138.1434.3636.25
AttMPTI [73]42.5540.8341.6954.0050.3252.1635.2330.7232.9846.7440.8043.77
BFG [29]42.1540.5241.3451.2349.3950.3134.1231.9833.0546.2541.3843.82
2CBR [75]50.7347.6649.2052.3547.1449.7547.0046.3646.6845.0639.4742.27
PAP3D [20]57.0855.9456.5164.5559.6462.1055.2755.6055.4459.0253.1656.09
Seg-PN [77]63.1564.3263.7467.0869.0568.0761.8065.3463.5762.9468.2665.60
DAFNet68.7969.9569.3770.9170.6070.7666.1466.7066.4268.9771.9570.46
Improvement+5.64+5.63+5.63+3.83+1.55+2.69+4.34+1.36+2.85+6.03+3.69+4.86
", + "bbox": [ + 516, + 142, + 903, + 252 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4d4a625f37a5e7da63739143f02eedef36cd88b6f243d194c53a131ba2701bac.jpg", + "table_caption": [ + "Table 6. Effect of different components of DAFNet." + ], + "table_footnote": [], + "table_body": "
GSA Layer\nEdgeConv MHADAF-ConvOA (%)
Individual Components
✓XX88.6
X✓X89.4
XX✓91.7
Component Combinations
✓✓X91.3
✓X✓94.8
X✓✓93.5
✓✓✓96.7
", + "bbox": [ + 552, + 290, + 867, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ble 5, DAFNet achieves consistent improvements across all experimental settings. For 2-way scenarios, our method obtains $69.37\\%$ mIoU in 1-shot and $70.76\\%$ mIoU in 5-shot settings, surpassing Seg-PN [77] by $5.63\\%$ and $2.69\\%$ respectively. In the more challenging 3-way settings, DAFNet achieves $66.42\\%$ mIoU in 1-shot and $70.46\\%$ mIoU in 5-shot settings, outperforming Seg-PN [77] by $2.85\\%$ and $4.86\\%$ . These substantial improvements across different settings demonstrate DAFNet's strong generalization ability and robustness in handling complex real-world scenes with limited training samples.", + "bbox": [ + 511, + 537, + 906, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 714, + 668, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All experiments are conducted on the ScanObjectNN [48] dataset unless explicitly stated otherwise.", + "bbox": [ + 511, + 736, + 906, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1. Ablation Study of Each Component", + "text_level": 1, + "bbox": [ + 511, + 775, + 808, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments to analyze the effectiveness of each component in DAFNet, as shown in Table 6. Among individual components, DAF-Conv shows the strongest performance with $91.7\\%$ accuracy, notably higher than single GSA layer components (EdgeConv: $88.6\\%$ , MHA: $89.4\\%$ ). When combining components, the fusion of DAF-Conv with EdgeConv achieves $94.8\\%$ , while DAF", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22188", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8f7b2682f0f39d1ce53da5eb4eef5f9a4837058aa3356b55681f77d0b521ccc7.jpg", + "table_caption": [ + "Table 7. Effect of different local geometric priors ${h}_{j}$ ." + ], + "table_footnote": [], + "table_body": "
Local Geometric Prior hjOA(%)
[pj]93.6
[pj,pi]94.7
[pj,pipj-pi]96.2
[pj,pi, pj-pi, ||pj-pi||]96.7
", + "bbox": [ + 153, + 114, + 419, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conv with MHA reaches $93.5\\%$ , both showing substantial improvements over single component performance. The full model integrating all components achieves the best performance of $96.7\\%$ , demonstrating the strong complementary effects between DAF-Conv's local geometric modeling and the GSA layer's global context capturing.", + "bbox": [ + 89, + 234, + 483, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2. Ablation Study of Local Geometric Prior", + "text_level": 1, + "bbox": [ + 89, + 335, + 421, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 7, we conduct ablation studies on the composition of local geometric prior $h_j$ . Starting with only center point coordinates $[p_j]$ yields $93.6\\%$ accuracy. Incorporating neighbor point information $[p_j, p_i]$ brings a $1.1\\%$ improvement. Adding relative position $[p_j, p_i, p_j - p_i]$ further boosts performance by $1.5\\%$ . The complete geometric prior with Euclidean distance $||p_j - p_i||$ achieves the best result of $96.7\\%$ , validating the importance of comprehensive geometric information in local structure modeling.", + "bbox": [ + 89, + 354, + 483, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.3. Analysis of Model Hyperparameters", + "text_level": 1, + "bbox": [ + 89, + 500, + 390, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation studies on three key hyperparameters of DAFNet, as shown in Fig. 4.", + "bbox": [ + 89, + 518, + 482, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frequency Component Number H: Fig. 4(a) shows that model performance improves as $H$ increases from 1 to 6, achieving peak accuracy at $H = 6$ $(96.7\\%)$ . Further increasing $H$ leads to performance degradation, dropping to $95.1\\%$ at $H = 10$ . This indicates that while multiple frequency components enhance feature diversity, excessive components may introduce redundant information.", + "bbox": [ + 89, + 549, + 483, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Distance Threshold $\\mathbf{d}_{\\mathrm{threshold}}$ : As shown in Fig. 4(b), we examine the impact of $d_{threshold}$ which divides the local neighborhood into near-field and far-field regions. The optimal performance (96.7%) is achieved at $d_{threshold} = 0.6$ . Lower thresholds limit near-field information, while higher values may include irrelevant points in the near-field region, both resulting in suboptimal performance.", + "bbox": [ + 89, + 655, + 483, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Neighborhood Size K: Fig. 4(c) demonstrates the effect of neighborhood size $K$ in EdgeConv. Performance reaches its peak $(96.7\\%)$ at $K = 20$ . Smaller $K$ values provide insufficient local context, while larger values may incorporate irrelevant points, both leading to decreased accuracy.", + "bbox": [ + 89, + 761, + 483, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Model Complexity", + "text_level": 1, + "bbox": [ + 89, + 848, + 269, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the model complexity of DAFNet with existing methods on ScanObjectNN [48], as shown in Ta", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b40574680401fdf325107935a697c4e00cbc4851a869ef42ef3f47fbc0da4ded.jpg", + "image_caption": [ + "Figure 4. Ablation studies on key hyperparameters. (a) Impact of frequency component number $H$ in DAF-Conv. (b) Effect of distance threshold $d_{threshold}$ for near-field and far-field separation. (c) Influence of neighborhood size $K$ in EdgeConv layer." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 647, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/675e9445c731bc33922d1e5691f3052f95db5cc8b578c3a8a2bad7cea864947c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 89, + 776, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d70e76dbb6bdf3905d656993436fbb2175c7bdddde2c971181cd890cd614d50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 90, + 903, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/52d9a26284470b6363e40d04a8a0d6c66832301da93ae72309bcb6eda084992d.jpg", + "table_caption": [ + "Table 8. Model complexity comparison on ScanObjectNN." + ], + "table_footnote": [], + "table_body": "
MethodsOA(%)Params.(M)FLOPs(G)
PointNet [35]68.03.50.5
PointNet++[36]77.91.51.7
DGCNN[55]78.11.82.4
PointCNN [24]78.50.6-
MVTN [18]82.811.243.7
PointMLP [28]85.412.631.4
PointNeXt [39]87.41.43.6
DAFNet (Ours)96.72.471.1
", + "bbox": [ + 519, + 305, + 901, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ble 8. While achieving state-of-the-art performance (96.7% OA), DAFNet maintains moderate computational requirements with only 2.47M parameters and 1.1G FLOPs. This is significantly more efficient than models like MVTN [18] (11.2M parameters, 43.7G FLOPs) and PointMLP [28] (12.6M parameters, 31.4G FLOPs), demonstrating that our acoustic field-inspired design achieves superior accuracy without sacrificing computational efficiency.", + "bbox": [ + 511, + 526, + 906, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 674, + 633, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented DAFNet, a physics-inspired architecture for point cloud understanding that addresses local structure representation challenges. Inspired by acoustic field principles, we introduced AF-Conv modeling local feature aggregation through acoustic propagation, and extended it to DAF-Conv which dynamically adapts to diverse geometric features through multi-frequency simulation. By integrating these with GSA, our approach captures both fine-grained local structures and global context. Experiments demonstrate DAFNet achieves state-of-the-art performance across multiple tasks. Future work will extend acoustic field modeling to other domains, exploring applications in additional 3D tasks.", + "bbox": [ + 511, + 703, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22189", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project is sponsored by Shanghai Pujiang Programme 24PJD030. The computational work for this article is partially performed on resources of the National Supercomputing Centre, Singapore (https://www.nscc.sg).", + "bbox": [ + 89, + 113, + 483, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 213 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, pages 1534-1543, 2016. 6", + "[2] Binjie Chen, Yunzhou Xia, Yu Zang, Cheng Wang, and Jonathan Li. Decoupled local aggregation for point cloud learning. arXiv preprint arXiv:2308.16532, 2023. 6", + "[3] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 6, 7", + "[4] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. NeurIPS, 36, 2024. 3", + "[5] Ziyang Chen, Israel D Gebru, Christian Richardt, Anurag Kumar, William Laney, Andrew Owens, and Alexander Richard. Real acoustic fields: An audio-visual room acoustics dataset and benchmark. In CVPR, pages 21886–21896, 2024. 2", + "[6] Pranav Singh Chib and Pravendra Singh. Recent advancements in end-to-end autonomous driving using deep learning: A survey. IEEE Transactions on Intelligent Vehicles, 2023. 1", + "[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 7", + "[8] Jeevan S Devagiri, Sidike Paheding, Quamar Niyaz, Xiaoli Yang, and Samantha Smith. Augmented reality and artificial intelligence in industry: Trends, tools, and future challenges. Expert Systems with Applications, 207:118002, 2022. 1", + "[9] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 2", + "[10] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. Mose: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 2", + "[11] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. 3", + "[12] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, pages 8458-8468, 2022. 2" + ], + "bbox": [ + 93, + 222, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Xiang Fang, Wanlong Fang, Changshuo Wang, Daizong Liu, Keke Tang, Jianfeng Dong, Pan Zhou, and Beibei Li. Multipair temporal sentence grounding via multi-thread knowledge transfer network. arXiv preprint arXiv:2412.15678, 2024. 2", + "[14] Ruchi Goel and Pooja Gupta. Robotics and industry 4.0. A Roadmap to Industry 4.0: Smart Production, Sharp Business and Sustainable Development, pages 157-169, 2020. 1", + "[15] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In ICML, pages 3809-3820. PMLR, 2021. 6", + "[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2", + "[17] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2", + "[18] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In ICCV, pages 1-11, 2021. 6, 8", + "[19] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In ACM MM, pages 4995-5004, 2024. 1, 2, 6", + "[20] Shuting He, Xudong Jiang, Wei Jiang, and Henghui Ding. Prototype adaption and projection for few-and zero-shot 3d point cloud semantic segmentation. TIP, 32:3199-3211, 2023. 7", + "[21] Limin Jiang, Changshuo Wang, Xin Ning, and Zaiyang Yu. Lttpoint: a mlp-based point cloud classification method with local topology transformation module. In 2023 7th Asian Conference on Artificial Intelligence Technology (ACAIT), pages 783-789. IEEE, 2023. 1", + "[22] Lawrence E Kinsler, Austin R Frey, Alan B Coppens, and James V Sanders. Fundamentals of acoustics. John wiley & sons, 2000. 4", + "[23] Artem Komarichev, Zichun Zhong, and Jing Hua. A-cnn: Annually convolutional neural networks on point clouds. In CVPR, pages 7421-7430, 2019. 2", + "[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. NeurIPS, 31, 2018. 8", + "[25] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 1, 3, 6, 7", + "[26] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual state space model. In NeurIPS, 2024. 2", + "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22190", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 6, 8", + "[29] Yongqiang Mao, Zonghao Guo, LU Xiaonan, Zhiqiang Yuan, and Haowen Guo. Bidirectional feature globalization for few-shot semantic segmentation of 3d point cloud scenes. In 2022 International Conference on 3D Vision (3DV), pages 505-514. IEEE, 2022. 7", + "[30] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928. IEEE, 2015. 2", + "[31] Philip McCord Morse and K Uno Ingard. Theoretical acoustics. Princeton university press, 1986. 4", + "[32] AAM Muzahid, Wanggen Wan, Ferdous Sohel, Lianyao Wu, and Li Hou. Curvenet: Curvature-based multitask learning deep networks for 3d object recognition. IEEE/CAA Journal of Automatica Sinica, 8(6):1177-1187, 2020. 6, 7", + "[33] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, pages 604-621. Springer, 2022. 1, 3, 6", + "[34] Jinyoung Park, Sanghyeok Lee, Sihyeon Kim, Yunyang Xiong, and Hyunwoo J Kim. Self-positioning point-based transformer for point cloud understanding. In CVPR, pages 21814-21823, 2023. 7", + "[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 2, 6, 7, 8", + "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NIPS, 30, 2017. 2, 6, 7, 8", + "[37] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In ICML, pages 28223-28243. PMLR, 2023. 1, 3", + "[38] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. *Shapellm: Universal 3d object understanding for embodied interaction*. In *ECCV*, pages 214–238. Springer, 2025. 1, 6", + "[39] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. NIPS, 35:23192-23204, 2022. 6, 7, 8", + "[40] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In CVPR, pages 18942-18952, 2022. 6", + "[41] Jingjing Ren, Ashwin Rao, Martina Lindorfer, Arnaud Legout, and David Choffnes. Recon: Revealing and controlling pii leaks in mobile network traffic. In Proceedings of the 14th Annual International Conference on Mobile Systems, Applications, and Services, pages 361-374, 2016. 6, 7" + ], + "bbox": [ + 91, + 90, + 480, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, pages 3577-3586, 2017. 2", + "[43] Mickael Sereno, Xiyao Wang, Lonni Besançon, Michael J Mcguffin, and Tobias Isenberg. Collaborative work in augmented reality: A survey. TVCG, 28(6):2530-2549, 2020. 1", + "[44] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. NIPS, 30, 2017. 7", + "[45] Mohsen Soori, Behrooz Arezoo, and Roza Dastres. Artificial intelligence, machine learning and deep learning in advanced robotics, a review. Cognitive Robotics, 3:54-70, 2023. 1", + "[46] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. In ECCV, pages 426-442. Springer, 2022. 2", + "[47] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6411–6420, 2019. 7", + "[48] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, pages 1588–1597, 2019. 5, 6, 7, 8", + "[49] Changshuo Wang, Xin Ning, Linjun Sun, Liping Zhang, Weijun Li, and Xiao Bai. Learning discriminative features by covering local geometric space for point cloud analysis. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-15, 2022. 2", + "[50] C Wang, H Wang, X Ning, S Tian, and W Li. 3d point cloud classification method based on dynamic coverage of local area. J. Softw, 34(4):1962-1976, 2022. 1", + "[51] Changshuo Wang, Xin Ning, Weijun Li, Xiao Bai, and Xingyu Gao. 3d person re-identification based on global semantic guidance and local feature aggregation. TCSVT, 34 (6):4698-4712, 2023. 2", + "[52] Changshuo Wang, Meiqing Wu, Siew-Kei Lam, Xin Ning, Shangshu Yu, Ruiping Wang, Weijun Li, and Thambipillai Srikanthan. Gpsformer: A global perception and local structure fitting-based transformer for point cloud understanding. arXiv preprint arXiv:2407.13519, 2024. 2, 5, 6, 7", + "[53] Changshuo Wang, Rongsheng Cao, and Ruiping Wang. Learning discriminative topological structure information representation for 2d shape and social network classification via persistent homology. Knowledge-Based Systems, page 113125, 2025. 2", + "[54] Ruiping Wang, Siew-Kei Lam, Meiqing Wu, Zhijian Hu, Changshuo Wang, and Jing Wang. Destination intention estimation-based convolutional encoder-decoder for pedestrian trajectory multimodality forecast. Measurement, 239: 115470, 2025. 1", + "[55] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 4, 5, 6, 7, 8" + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22191", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Xin Wei, Ruixuan Yu, and Jian Sun. View-gcn: View-based graph convolutional network for 3d shape analysis. In CVPR, pages 1850-1859, 2020. 2", + "[57] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 2", + "[58] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 2", + "[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, pages 1912-1920, 2015. 6", + "[60] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, pages 3173-3182, 2021. 4, 6", + "[61] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In CVPR, pages 1179-1189, 2023. 3, 6", + "[62] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. NIPS, 32, 2019. 4, 5", + "[63] Ze Yang and Liwei Wang. Learning relationships for multiview 3d object recognition. In ICCV, pages 7505-7514, 2019. 2", + "[64] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6", + "[65] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, pages 19313-19322, 2022. 6", + "[66] Huang Zhang, Changshuo Wang, Long Yu, Shengwei Tian, Xin Ning, and Joel Rodrigues. Pointgt: A method for pointcloud classification and segmentation based on local geometric transformation. IEEE Transactions on Multimedia, 2024. 2", + "[67] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. NeurIPS, 35:27061-27074, 2022. 3", + "[68] Renrui Zhang, Liuhui Wang, Ziyu Guo, Yali Wang, Peng Gao, Hongsheng Li, and Jianbo Shi. Parameter is not all you need: Starting from non-parametric networks for 3d point cloud analysis. arXiv preprint arXiv:2303.08134, 2023. 6", + "[69] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024.2", + "[70] Xiangdong Zhang, Shaofeng Zhang, and Junchi Yan. Pcmae: Learning to predict centers for point masked autoen" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "coders. Advances in Neural Information Processing Systems, 37:80303-80327, 2024. 3, 6, 7", + "[71] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 2, 6, 7", + "[72] Jingyuan Zhao, Wenyi Zhao, Bo Deng, Zhenghong Wang, Feng Zhang, Wenxiang Zheng, Wanke Cao, Jinrui Nan, Yubo Lian, and Andrew F Burke. Autonomous driving system: A comprehensive survey. Expert Systems with Applications, page 122836, 2023. 1", + "[73] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In CVPR, pages 8873-8882, 2021. 6, 7", + "[74] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In CVPR, pages 22935-22945, 2024. 3", + "[75] Guanyu Zhu, Yong Zhou, Rui Yao, and Hancheng Zhu. Cross-class bias rectification for point cloud few-shot segmentation. TMM, 25:9175-9188, 2023. 7", + "[76] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2", + "[77] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ji-aming Liu, Han Xiao, Chaoyou Fu, Hao Dong, and Peng Gao. No time to train: Empowering non-parametric networks for few-shot 3d scene segmentation. In CVPR, pages 3838–3847, 2024. 5, 7" + ], + "bbox": [ + 516, + 92, + 903, + 501 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22192", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/7aa2860d-04df-426a-bde9-40b0ed00d4aa_model.json b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/7aa2860d-04df-426a-bde9-40b0ed00d4aa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4850c4a1c8f7bb41b6d18d8b2556f4037fd92fa3 --- /dev/null +++ b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/7aa2860d-04df-426a-bde9-40b0ed00d4aa_model.json @@ -0,0 +1,2928 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.13, + 0.898, + 0.177 + ], + "angle": 0, + "content": "Point Clouds Meets Physics: Dynamic Acoustic Field Fitting Network for Point Cloud Understanding" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.205, + 0.341, + 0.222 + ], + "angle": 0, + "content": "Changshuo Wang" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.223, + 0.337, + 0.24 + ], + "angle": 0, + "content": "NTU, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.243, + 0.378, + 0.257 + ], + "angle": 0, + "content": "wangchangshuol@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.473, + 0.205, + 0.568, + 0.222 + ], + "angle": 0, + "content": "Shuting He" + }, + { + "type": "text", + "bbox": [ + 0.468, + 0.223, + 0.575, + 0.239 + ], + "angle": 0, + "content": "SUFE, China" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.243, + 0.62, + 0.257 + ], + "angle": 0, + "content": "shuting.he@sufe.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.704, + 0.205, + 0.805, + 0.223 + ], + "angle": 0, + "content": "Xiang Fang*" + }, + { + "type": "text", + "bbox": [ + 0.685, + 0.223, + 0.816, + 0.24 + ], + "angle": 0, + "content": "NTU, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.666, + 0.243, + 0.836, + 0.257 + ], + "angle": 0, + "content": "xfang9508@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.161, + 0.265, + 0.252, + 0.28 + ], + "angle": 0, + "content": "Jiawei Han" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.283, + 0.253, + 0.298 + ], + "angle": 0, + "content": "BIT, China" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.303, + 0.296, + 0.317 + ], + "angle": 0, + "content": "hanjiawei@bit.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.389, + 0.265, + 0.518, + 0.282 + ], + "angle": 0, + "content": "Zhonghang Liu" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.283, + 0.522, + 0.3 + ], + "angle": 0, + "content": "SMU, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.303, + 0.575, + 0.317 + ], + "angle": 0, + "content": "zhliu.2020@phdcs.smu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.6, + 0.265, + 0.786, + 0.283 + ], + "angle": 0, + "content": "Xin Ning" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.283, + 0.881, + 0.299 + ], + "angle": 0, + "content": "Institute of Semiconductors, CAS" + }, + { + "type": "text", + "bbox": [ + 0.667, + 0.304, + 0.828, + 0.316 + ], + "angle": 0, + "content": "ningxin@semi.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.22, + 0.325, + 0.487, + 0.36 + ], + "angle": 0, + "content": "Weijun Li \nInstitute of Semiconductors, CAS" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.364, + 0.421, + 0.377 + ], + "angle": 0, + "content": "wjli@semi.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.542, + 0.325, + 0.779, + 0.36 + ], + "angle": 0, + "content": "Prayag Tiwari \nHalmstad University, Sweden" + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.365, + 0.756, + 0.378 + ], + "angle": 0, + "content": "prayag.tiwari@ieee.org" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.413, + 0.327, + 0.428 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.448, + 0.485, + 0.857 + ], + "angle": 0, + "content": "While existing pre-training-based methods have enhanced point cloud model performance, they have not fundamentally resolved the challenge of local structure representation in point clouds. The limited representational capacity of pure point cloud models continues to constrain the potential of cross-modal fusion methods and performance across various tasks. To address this challenge, we propose a Dynamic Acoustic Field Fitting Network (DAF-Net), inspired by physical acoustic principles. Specifically, we represent local point clouds as acoustic fields and introduce a novel Acoustic Field Convolution (AF-Conv), which treats local aggregation as an acoustic energy field modeling problem and captures fine-grained local shape awareness by dividing the local area into near field and far field. Furthermore, drawing inspiration from multifrequency wave phenomena and dynamic convolution, we develop the Dynamic Acoustic Field Convolution (DAF-Conv) based on AF-Conv. DAF-Conv dynamically generates multiple weights based on local geometric priors, effectively enhancing adaptability to diverse geometric features. Additionally, we design a Global Shape-Aware (GSA) layer incorporating EdgeConv and multi-head attention mechanisms, which combines with DAF-Conv to form the DAF Block. These blocks are then stacked to create a hierarchical DAFNet architecture. Extensive experiments demonstrate that DAFNet significantly outperforms existing methods across multiple tasks." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.412, + 0.897, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.475, + 0.7, + 0.484 + ], + "angle": 0, + "content": "Local Neighborhood" + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.474, + 0.78, + 0.489 + ], + "angle": 0, + "content": "Multi-Acoustic Energy Field Mapping" + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.473, + 0.896, + 0.482 + ], + "angle": 0, + "content": "Shape-aware Representation" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.504, + 0.907, + 0.588 + ], + "angle": 0, + "content": "Figure 1. Illustration of local feature extraction using DAF-Conv. Given a local neighborhood of a point cloud (left), our method first models it as an acoustic field with near-field and far-field zones. Then, multiple acoustic energy field mappings are generated to capture various geometric patterns. Finally, DAF-Conv processes these mappings to produce shape-aware representations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.616, + 0.645, + 0.632 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.642, + 0.907, + 0.764 + ], + "angle": 0, + "content": "Point cloud data, due to its direct representation of object surface information in three-dimensional space, has been widely applied in various scenarios including autonomous driving [6, 54, 72], robotic navigation [14, 45], and augmented reality [8, 43]. However, the sparsity, irregularity, and complexity of local structures in point cloud data make it challenging for existing models to fully capture geometric features, thus limiting their applications in 3D tasks." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recently, several methods [33, 37, 38] have attempted to indirectly improve point cloud model performance through pre-training strategies or cross-modal contrastive learning. Although these approaches have enhanced point cloud processing tasks to some extent, they have not directly addressed the problem of local structure representation. Some researchers [19, 21, 25, 50] have also attempted to apply new architectures to point cloud understanding tasks. However, the local representational capacity of point cloud mod" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.276, + 0.9 + ], + "angle": 0, + "content": "* denotes corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22182" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "els remains a bottleneck in improving overall model performance. Therefore, how to efficiently capture and represent the local structures of point clouds has become a crucial problem that urgently needs to be solved." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.162, + 0.483, + 0.372 + ], + "angle": 0, + "content": "Physics, particularly the principles of acoustic field propagation [5], provides a new perspective for understanding local geometric structures. In acoustics, wave propagation is influenced by environmental geometric characteristics, and through constructing acoustic field models, we can effectively simulate wave behavior at different frequencies to obtain information in complex environments. Inspired by this, we propose a novel point cloud understanding method called Dynamic Acoustic Field Fitting Network (DAFNet). This method borrows ideas from physical acoustic field propagation, representing local point clouds as acoustic fields and dynamically adjusting weights in different regions to capture local structural features more precisely." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.483, + 0.669 + ], + "angle": 0, + "content": "Specifically, we propose a novel Acoustic Field Convolution (AF-Conv) that models local aggregation as an acoustic energy propagation problem. To more finely model the contributions of different neighboring points to the center point, AF-Conv divides the local region into near-field and far-field zones for separate geometric modeling. Furthermore, as shown in Fig. 1, based on AF-Conv, we propose Dynamic Acoustic Field Convolution (DAF-Conv), which dynamically generates multiple weights based on geometric priors while simulating geometric feature variations under different \"frequencies\", thereby enhancing the model's adaptability to local geometric diversity. Finally, we combine a Global Shape-Aware (GSA) layer comprising EdgeConv and Multi-Head Attention (MHA) with DAF-Conv to build hierarchical DAF Blocks, forming the complete DAFNet. The experimental results show that DAFNet achieves superior performance over existing methods in multiple tasks, demonstrating its effectiveness in capturing complex geometric structures." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.678, + 0.469, + 0.693 + ], + "angle": 0, + "content": "The main contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.703, + 0.483, + 0.746 + ], + "angle": 0, + "content": "- We propose a physics-inspired end-to-end DAFNet that can learn rich contextual information and precise shape representations in irregular point clouds." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.748, + 0.483, + 0.792 + ], + "angle": 0, + "content": "- We introduce a novel Acoustic Field Convolution (AFConv) that achieves fine-grained local shape awareness by modeling local point cloud regions as acoustic fields." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.794, + 0.483, + 0.852 + ], + "angle": 0, + "content": "- We design Dynamic Acoustic Field Convolution (DAFConv) that adapts to geometric feature diversity by dynamically generating multiple frequency-aware weights from local geometry." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.854, + 0.483, + 0.898 + ], + "angle": 0, + "content": "- The proposed DAFNet achieves state-of-the-art performance in multiple tasks, reaching \\(96.7\\%\\) accuracy on the challenging ScanObjectNN with only 2.47M parameters." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.703, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.71, + 0.13 + ], + "angle": 0, + "content": "2.1. CNN-based Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.138, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Early works [30, 42, 56, 63] in point cloud processing primarily involved projecting them onto 2D images for 2D CNN processing or converting them into regular 3D grid structures for 3D convolution operations. However, these transformation processes inevitably led to spatial information loss and increased computational costs. To address this issue, researchers began applying deep learning techniques directly in 3D space. PointNet [35] pioneered this direction. However, this method overlooked local structural information. Subsequently, PointNet++ [36] enhanced the local perception capability of point cloud models by grouping point clouds into different local neighborhoods. Based on PointNet++ [36], existing works [23, 49, 51, 52, 55, 66] have developed methods based on point-wise perceptrons, convolution operations, and attention mechanisms. Nevertheless, these methods still struggle to effectively model local structural shape information." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.406, + 0.768, + 0.421 + ], + "angle": 0, + "content": "2.2. Transformer-based Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.43, + 0.907, + 0.672 + ], + "angle": 0, + "content": "Transformers have demonstrated remarkable success in various tasks [9, 10, 13, 53] due to their inherent ability to capture global features. Leveraging this strength, researchers have progressively developed a range of Transformer-based point cloud models. PCT [17] enhances input embeddings through furthest point sampling and nearest neighbor search while using self-attention mechanisms to capture global features. Point Transformer [71] employs vector attention to extract aggregated local features. Subsequently, updated versions [57, 58] of Point Transformer explored different architectures to improve performance and efficiency. Inspired by Swin Transformer [27], works such as SST [12] and SWFormer [46] explored window attention mechanisms to improve efficiency. Although these methods excel at capturing global features, their performance hasn't shown significant advancement over CNN-based approaches." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.682, + 0.813, + 0.698 + ], + "angle": 0, + "content": "2.3. State Space Model-based Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "State Space Models (SSMs) [16] are mathematical models for describing dynamic systems. Recently, SSMs have emerged as a reliable alternative architecture to CNNs and Transformers. In computer vision, Vision Mamba [76] successfully compresses visual representations through bidirectional state space models, while VMamba [26] innovatively introduces cross-scanning modules, enabling global receptive fields in 2D images through 1D selective scanning. Researchers have explored SSM applications in point cloud processing. PCM [69] first introduced the Mamba architecture to point cloud processing through octree serialization strategies. Mamba3D [19] enhanced local geometric feature extraction through LNP blocks and designed bidi" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22183" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.885, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.427, + 0.908, + 0.5 + ], + "angle": 0, + "content": "Figure 2. Architecture of DAFNet. (a) For shape classification, the network consists of three sequential DAF Blocks followed by max pooling and classifier layers. Each DAF Block contains a GSA layer and DAF-Conv for feature extraction. (b) For part segmentation and few-shot semantic segmentation tasks, the network adopts an encoder-decoder structure with skip connections, where the encoder shares the same architecture as the classification network, and the decoder uses feature propagation to restore point cloud resolution. The few-shot version additionally incorporates support and query mechanisms for processing different point cloud inputs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.487, + 0.571 + ], + "angle": 0, + "content": "rectional SSM structures. PointMamba [25], another significant work, customized SSM architecture for point cloud analysis tasks, substantially reducing computational costs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.345, + 0.594 + ], + "angle": 0, + "content": "2.4. Pre-training-based Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Point cloud pre-training methods primarily comprise generative approaches [4, 33, 67] and cross-modal contrastive learning methods. In generative approaches, Point-MAE [33] extends the MAE concept by pre-training through random masking and reconstruction of point cloud segments PCP-MAE [70] learns shape priors by predicting missing regional point cloud structures. PointDif [74] utilizes diffusion models for pre-training, learning hierarchical geometric priors through point cloud recovery at different noise levels. In cross-modal contrastive learning, ReCon [37] combines generative modeling and contrastive learning paradigms, addressing representation overfitting in contrastive learning and data padding issues in generative methods. ULIP [61] achieves cross-modal knowledge transfer through a language-image-point cloud pre-training framework, while ACT [11] employs cross-modal autoencoders as teacher models to acquire knowledge from other modalities. Although these pre-training methods have advanced point cloud processing performance, they haven't fundamentally resolved the point cloud representation challenge." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.522, + 0.613, + 0.538 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.907, + 0.625 + ], + "angle": 0, + "content": "This section first reviews point convolution and dynamic convolution, followed by introducing acoustic field convolution. Then, we present our proposed dynamic acoustic field convolution. Finally, we describe DAFNet (as shown in Fig. 2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.631, + 0.646, + 0.647 + ], + "angle": 0, + "content": "3.1. Background" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.688, + 0.667 + ], + "angle": 0, + "content": "3.1.1. Point Convolution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Local feature aggregation in point clouds typically follows a \"sampling-grouping-aggregation\" procedure. The aggregation operation is crucial as it directly affects local feature representation by modeling local point clouds. We denote center points as \\(\\{(p_i, f_i)\\}_{i=1}^M\\), where \\(M\\) represents the number of points, \\(p_i \\in \\mathbb{R}^{1 \\times 3}\\) and \\(f_i \\in \\mathbb{R}^{1 \\times C}\\) represent the coordinates and features of the \\(i\\)-th point respectively, and \\(C\\) denotes the number of feature channels. The most commonly used method for local feature aggregation based on coordinate space is to construct point convolution, which can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.846, + 0.907, + 0.865 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {\\prime} = \\mathcal {A} (\\{\\mathcal {M} (p _ {i}, p _ {j}) \\cdot \\mathcal {T} (f _ {i}, f _ {j}) | p _ {j} \\in \\mathcal {N} (p _ {i}) \\}), \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}\\) represents the aggregation function, typically max-pooling. \\(\\mathcal{M}\\) and \\(\\mathcal{T}\\) are mapping functions, usually imple" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22184" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.137 + ], + "angle": 0, + "content": "mented through Multi-Layer Perceptrons (MLPs). \\(\\mathcal{N}(p_i)\\) represents the local neighborhood of \\(p_i\\), and \\(p_j\\) denotes the neighboring points of \\(p_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.484, + 0.213 + ], + "angle": 0, + "content": "An alternative aggregation approach operates in feature space, commonly known as Dynamic Graph Convolution (DGC) [55]. It aggregates local features in feature space, meaning that points close in feature space might be distant in coordinate space. DGC can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.224, + 0.483, + 0.243 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {\\prime} = \\mathcal {A} (\\{\\mathcal {T} (f _ {i}, f _ {j}) | f _ {j} \\in \\mathcal {N} (f _ {i}) \\}). \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.252, + 0.291, + 0.267 + ], + "angle": 0, + "content": "3.1.2. Dynamic Convolution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.271, + 0.484, + 0.362 + ], + "angle": 0, + "content": "Dynamic convolution [62] aims to enhance network modeling capability by dynamically generating multiple convolution kernels based on input data, adapting to different data structures. Taking PACConv's [60] dynamic convolution as an example, dynamic convolution in point clouds can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.374, + 0.483, + 0.393 + ], + "angle": 0, + "content": "\\[\ng _ {i} = \\mathcal {A} \\left(\\left\\{w \\left(p _ {j}\\right) \\cdot f _ {j} \\mid p _ {j} \\in \\mathcal {N} \\left(p _ {i}\\right) \\right\\}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.464 + ], + "angle": 0, + "content": "where \\( w(p_j) \\) denotes the weight corresponding to feature \\( f_j \\) of point \\( p_j \\). \\( w(p_j) \\) typically consists of multiple weights \\( w_h(p_i) \\) and their corresponding attention coefficients \\( \\alpha_h(p_j) \\) generated in a data-driven manner:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.474, + 0.483, + 0.517 + ], + "angle": 0, + "content": "\\[\nw \\left(p _ {j}\\right) = \\sum_ {h = 1} ^ {H} \\alpha_ {h} \\left(p _ {j}\\right) \\cdot w _ {h} \\left(p _ {j}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.56 + ], + "angle": 0, + "content": "where \\(\\sum_{h=1}^{H} \\alpha_h(p_j) = 1\\), \\(H\\) represents the number of \\(w_h(p_j)\\)." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.561, + 0.388, + 0.575 + ], + "angle": 0, + "content": "Therefore, formula 3 can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.586, + 0.483, + 0.628 + ], + "angle": 0, + "content": "\\[\ng _ {i} = \\mathcal {A} \\left(\\left\\{\\sum_ {h = 1} ^ {H} \\alpha_ {h} \\left(p _ {j}\\right) \\cdot w _ {h} \\left(p _ {j}\\right) \\cdot f _ {j} \\mid p _ {j} \\in \\mathcal {N} \\left(p _ {i}\\right) \\right\\}\\right). \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.648, + 0.336, + 0.663 + ], + "angle": 0, + "content": "3.2. Acoustic Field Convolution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.669, + 0.484, + 0.716 + ], + "angle": 0, + "content": "In physical acoustics [22, 31], the sound pressure field \\( p(x) \\) generated by a sound source at spatial point \\( x \\) can be represented by Green's function:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.725, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\np (x) = \\frac {e ^ {- j k \\| x - y \\|}}{4 \\pi \\| x - y \\|}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.856 + ], + "angle": 0, + "content": "where \\(y\\) is the source position, \\(k\\) is the wave number, and \\(\\| x - y\\|\\) represents the distance between observation point \\(x\\) and source \\(y\\). This formula reveals the fundamental physical laws of sound wave propagation: sound pressure intensity decays exponentially with distance while being affected by spherical spreading attenuation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Notably, based on the distance between the source and observation point, the acoustic field can be divided into near-field and far-field regions. In the near-field region, the" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.265, + 0.907, + 0.336 + ], + "angle": 0, + "content": "Figure 3. Illustration of AF-Conv and DAF-Conv. (a) AF-Conv divides local neighborhood into near-field and far-field regions and calculates energy weights based on geometric priors. (b) DAF-Conv extends AF-Conv with multiple parallel acoustic field convolutions to achieve dynamic feature aggregation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.905, + 0.42 + ], + "angle": 0, + "content": "relationship between sound pressure and distance is more complex, exhibiting strong local correlations; in the farfield region, the sound pressure demonstrates simple spherical wave attenuation characteristics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.906, + 0.541 + ], + "angle": 0, + "content": "Inspired by this theory, we view local feature aggregation as acoustic field propagation, where the center point \\( p_i \\) can be viewed as the sound source and neighboring points \\( p_j \\) as observation points. Meanwhile, we refine the local neighborhood into near-field and far-field regions to extract fine-grained information, constructing the Acoustic Field Convolution (AF-Conv). As shown in Fig. 3(a), it can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.554, + 0.905, + 0.571 + ], + "angle": 0, + "content": "\\[\ng _ {i} = \\mathcal {A} \\left(\\left\\{W _ {A F} \\left(p _ {i}, p _ {j}\\right) \\cdot f _ {j} \\mid p _ {j} \\in \\mathcal {N} \\left(p _ {i}\\right) \\right\\}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.642 + ], + "angle": 0, + "content": "where the weight function \\( W_{AF}(p_i,p_j) \\) directly draws inspiration from acoustic field propagation characteristics and designs different weight calculation methods for near-field and far-field regions:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.652, + 0.905, + 0.694 + ], + "angle": 0, + "content": "\\[\nW _ {A F} \\left(p _ {i}, p _ {j}\\right) = \\left\\{ \\begin{array}{l l} e ^ {- \\beta d _ {i j}}, & \\text {i f} d _ {i j} < d _ {\\text {t h r e s h o l d}} \\\\ \\frac {1}{d _ {i j} + \\epsilon}, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.907, + 0.84 + ], + "angle": 0, + "content": "where \\(\\beta\\) represents learnable coefficient. \\(d_{ij}\\) denotes the Euclidean distance between \\(p_j\\) and \\(p_i\\). In \\(W_{AF}(p_i,p_j)\\), for the near-field region (\\(d_{ij} < d_{threshold}\\)), we use the exponential decay term \\(e^{-\\beta d_{ij}}\\) to capture subtle changes in local geometric structures, corresponding to the complex attenuation characteristics of the acoustic near-field region; for the far-field region (\\(d_{ij} \\geq d_{threshold}\\)), we adopt a simple reciprocal relationship \\(\\frac{1}{d_{ij} + \\epsilon}\\) to better perceive broader contextual information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.833, + 0.865 + ], + "angle": 0, + "content": "3.3. Dynamic Acoustic Field Convolution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In sound wave propagation [31], waves of different frequencies exhibit distinct propagation characteristics. Inspired by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22185" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "this multi-frequency phenomenon and dynamic convolution [62], as shown in Fig. 3(b), we propose Dynamic Acoustic Field Convolution (DAF-Conv), which extends AF-Conv by dynamically generating multiple sets of weights to simulate local geometric feature variations under different \"frequencies\". It can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.193, + 0.483, + 0.211 + ], + "angle": 0, + "content": "\\[\nF _ {h} \\left(p _ {i}, p _ {j}\\right) = W _ {h} \\left(p _ {i}, p _ {j}\\right) \\cdot f _ {j}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.219, + 0.483, + 0.259 + ], + "angle": 0, + "content": "\\[\nW _ {h} \\left(p _ {i}, p _ {j}\\right) = \\left\\{ \\begin{array}{l l} \\alpha_ {h} e ^ {- \\beta_ {h} d _ {i j}}, & \\text {i f} d _ {i j} < d _ {\\text {t h r e s h o l d}} \\\\ \\frac {\\alpha_ {h}}{d _ {i j} + \\epsilon}, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.264, + 0.483, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\alpha_ {h}, \\beta_ {h} \\right\\} _ {h = 1} ^ {H} = \\operatorname {M L P} \\left(h _ {j}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.376 + ], + "angle": 0, + "content": "where \\(d_{threshold}\\) is a scalar that defines the boundary between near-field and far-field regions. The weight coefficient \\(\\alpha_{h}\\) and \\(\\beta_{h}\\) are generated from local geometric prior \\(h_j = [p_i,p_j,p_j - p_i,\\| p_j - p_i\\| ]\\) through MLP to simulate propagation characteristics of different frequency components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.422 + ], + "angle": 0, + "content": "The local aggregated features are then separated into near-field and far-field components based on the distance threshold:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.448, + 0.483, + 0.467 + ], + "angle": 0, + "content": "\\[\nF _ {h} ^ {\\text {n e a r}} \\left(p _ {i}\\right) = \\mathcal {A} \\left(\\left\\{F _ {h} \\left(p _ {i}, p _ {j}\\right) \\mid j \\in \\mathcal {N} _ {\\text {n e a r}} (i) \\right\\}\\right), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.475, + 0.483, + 0.495 + ], + "angle": 0, + "content": "\\[\nF _ {h} ^ {f a r} \\left(p _ {i}\\right) = \\mathcal {A} \\left(\\left\\{F _ {h} \\left(p _ {i}, p _ {j}\\right) \\mid j \\in \\mathcal {N} _ {f a r} (i) \\right\\}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.483, + 0.559 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}_{near}(i) = \\{j\\mid p_j\\in \\mathcal{N}(p_i),d_{ij} < d_{threshold}\\}\\) represents the set of near-field neighbors of the center point \\(p_i\\) and \\(\\mathcal{N}_{far}(i) = \\{j\\mid p_j\\in \\mathcal{N}(p_i),d_{ij}\\geq d_{threshold}\\}\\) represents the set of far-field neighbors." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.56, + 0.483, + 0.575 + ], + "angle": 0, + "content": "The component for each frequency \\( h \\) can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.587, + 0.483, + 0.607 + ], + "angle": 0, + "content": "\\[\nF _ {h} \\left(p _ {i}\\right) = F _ {h} ^ {\\text {n e a r}} \\left(p _ {i}\\right) + F _ {h} ^ {\\text {f a r}} \\left(p _ {i}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.671 + ], + "angle": 0, + "content": "To achieve adaptive fusion of features from different frequency components, we introduce an attention mechanism based on geometric prior \\( h_j \\). The attention weights are generated as fellow:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.686, + 0.483, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {h} = \\operatorname {s o f t m a x} (\\operatorname {M a x P o o l i n g} (\\operatorname {M L P} (h _ {j}))). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.484, + 0.754 + ], + "angle": 0, + "content": "The final local aggregated feature is obtained by weighted summation of features from all frequency components:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.765, + 0.483, + 0.806 + ], + "angle": 0, + "content": "\\[\ng _ {i} = \\sum_ {h = 1} ^ {H} \\gamma_ {h} \\cdot F _ {h} \\left(p _ {i}\\right). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "DAF-Conv achieve adaptive modeling of local geometric features through separately processing near-field and far-field features for each frequency component, independent max-pooling aggregation for near-field and far-field features, and adaptive fusion of multiple frequency components through learned attention weights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "Compared to the single weight function in AF-Conv, DAF-Conv achieves more sophisticated feature extraction capabilities through its multi-frequency design and separate processing of near-field and far-field features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.163, + 0.907, + 0.193 + ], + "angle": 0, + "content": "3.4. Dynamic Acoustic Field Convolution Neural Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.201, + 0.906, + 0.322 + ], + "angle": 0, + "content": "To ensure each point has rich feature representation before constructing local neighborhoods, we build a Global Shape-Aware (GSA) layer before applying DAF-Conv. The GSA layer first aggregates each point through EdgeConv [55] without reducing point cloud resolution, then applies Multi-Head Attention (MHA) for global context modeling on the enhanced points. This layer effectively enhances DAF-Conv's perception ability for local structure modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.906, + 0.398 + ], + "angle": 0, + "content": "As shown in Fig. 2, based on the GSA layer and DAF-Conv, we construct the DAF Block. By stacking these blocks, we design a novel pretraining-free network, DAFNet, for point cloud analysis. We implement models for three major tasks:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.906, + 0.489 + ], + "angle": 0, + "content": "Point Cloud Classification. For point cloud classification, we employ three DAF Blocks, followed by a classifier consisting of max-pooling and fully connected layers. The output feature dimensions of the three blocks are 64, 128, 256 respectively. Notably, with only 2.47M parameters, it achieves the best performance on ScanObjectNN [48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.642 + ], + "angle": 0, + "content": "Part Segmentation and Few-shot Point Cloud Semantic Segmentation. For part segmentation, the encoder structure remains identical to the classification task. In the decoder stage, we employ feature propagation to restore point cloud resolution. Between the encoder and decoder, we apply U-Net-like skip connections to fully utilize contextual information. In few-shot segmentation, while maintaining the same encoder-decoder structure as part segmentation, we incorporate the Query module from SegPN[77] to reduce domain gaps between query and support features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.656, + 0.649, + 0.673 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.733, + 0.699 + ], + "angle": 0, + "content": "4.1.3D Shape Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Real-world Object Classification on ScanObjectNN. We evaluate our method on ScanObjectNN [48], a challenging real-world point cloud dataset containing 2,902 objects across 15 categories. Following previous works, we conduct experiments on its most challenging variant (PB_T50_RS), where objects are presented with background and occlusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 1, DAFNet achieves state-of-the-art performance on ScanObjectNN [48], reaching \\(96.7\\%\\) overall accuracy (OA) and \\(96.3\\%\\) mean class accuracy (mAcc) with voting mechanism. These results significantly surpass both existing supervised methods (e.g., GPSFormer [52]: \\(95.4\\%\\) OA) and pre-training approaches (e.g., ReCon++" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22186" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.117 + ], + "angle": 0, + "content": "Table 1. Classification results on the ScanObjectNN [48] dataset. \" - \" denotes unknown." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.128, + 0.482, + 0.59 + ], + "angle": 0, + "content": "
MethodYearmAcc (%)OA (%)
Supervised Methods
PointNet [35]201763.468.2
PointNet++ [36]201869.873.7
DGCNN [55]201973.678.1
SimpleView [15]2021-80.5
MVTN [18]202183.185.5
PointMLP [28]202284.485.7
RepSurf-U [40]202283.186.0
PointNeXt [39]202285.887.7
Point-PN [68]2023-87.1
DeLA [2]202389.390.4
GPSFormer [52]202493.895.4
Mamba3D [19]2024-92.6
Pre-training Methods
Point-BERT [65]2021-83.1
Point-MAE [33]2022-85.2
ULIP-2 + PointNeXt [61]202391.291.5
PointGPT [3]2023-93.6
ReCon [41]2023-91.3
ReCon++ [38]2024-95.25
PointMamba [25]2024-89.3
PCP-MAE [70]2024-90.4
Ours
DAFNet (w/o voting)-96.196.4
DAFNet (w/ voting)-96.396.7
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[38]: \\(95.25\\%\\) OA). The superior performance demonstrates the effectiveness of our acoustic field-inspired architecture in handling real-world point cloud data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Synthetic Object Classification on ModelNet40. We further evaluate DAFNet on ModelNet40 [59], a widely-used synthetic dataset containing 12,311 CAD models from 40 categories, with 9,843 for training and 2,468 for testing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.903 + ], + "angle": 0, + "content": "As presented in Table 2, DAFNet achieves competitive results on ModelNet40 [59] with \\(94.1\\%\\) overall accuracy (OA) and \\(92.2\\%\\) mean class accuracy (mAcc). While these results are slightly lower than some recent methods like PointGPT [3] (\\(94.9\\%\\) OA), the performance gap is minimal (less than \\(1\\%\\)). This is mainly because ModelNet40 [3], as a synthetic dataset, has been well-addressed by existing methods. In contrast, DAFNet shows more significant advantages on real-world datasets, as demonstrated by our results on ScanObjectNN [48]." + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.089, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 2. Classification results on ModelNet40 dataset. \"-\" denotes unknown." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.129, + 0.905, + 0.574 + ], + "angle": 0, + "content": "
MethodsYearmAcc (%)OA (%)
Supervised Methods
PointNet [35]201786.089.2
PointNet++[36]2017-90.7
DGCNN[55]201990.292.9
Point Transformer [71]202190.693.7
PAConv [60]2021-93.9
CurveNet [32]2021-94.2
PointMLP [28]202291.494.5
RepSurf-U [40]2022-94.7
GPSFormer [52]202492.294.2
Mamba3D [19]2024-94.1
Pre-training Methods
Point-BERT [65]2021-93.8
Point-MAE [33]2022-94.0
ULIP + PointMLP [61]202392.494.7
PointGPT [3]2023-94.9
ReCon [41]2023-94.5
ReCon++ [38]2024-94.8
PointMamba [25]2024-93.6
PCP-MAE [70]2024-94.0
Ours
DAFNet (w/o voting)-91.993.8
DAFNet (w/ voting)-92.294.1
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.599, + 0.694, + 0.615 + ], + "angle": 0, + "content": "4.2. Part Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.907, + 0.697 + ], + "angle": 0, + "content": "To evaluate our method's capability in fine-grained shape understanding, we conduct experiments on ShapeNetPart [64], which contains 16,881 shapes from 16 categories with 50 part annotations. Following convention, we use 13,807 samples for training and 2,874 for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.907, + 0.789 + ], + "angle": 0, + "content": "As shown in Table 3, DAFNet achieves competitive performance with \\(85.2\\%\\) class mIoU and \\(86.8\\%\\) instance mIoU, comparable to state-of-the-art methods. These results demonstrate that our acoustic field-inspired approach effectively captures both global shape information and local geometric details necessary for part segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.799, + 0.902, + 0.815 + ], + "angle": 0, + "content": "4.3. Few-shot Point Cloud Semantic Segmentation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.821, + 0.822, + 0.837 + ], + "angle": 0, + "content": "4.3.1. Results analysis on the S3DIS dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We evaluate our model on S3DIS [1], which contains 3D point clouds from 272 rooms with 13 semantic categories. Following [73], we divide scenes into \\(1\\mathrm{m} \\times 1\\mathrm{m}\\) blocks and sample 2048 points per block, resulting in 7547" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22187" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.119 + ], + "angle": 0, + "content": "Table 3. Part segmentation results (\\%) on ShapeNetPart. We report both class mIoU and instance mIoU." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.129, + 0.482, + 0.438 + ], + "angle": 0, + "content": "
MethodsYearclass mIoUinstance mIoU
Supervised Methods
PointNet [35]201780.483.7
PointNet++[36]201781.985.1
DGCNN [55]201982.385.1
KPConv [47]201985.186.4
Point Transformer [71]202183.786.6
CurveNet [32]2021-86.8
PointNeXt [39]202285.287.1
SPOTr [34]202385.487.2
GPSFormer [52]202485.486.8
Pre-training Methods
PointGPT[3]202384.186.2
ReCon [41]202384.886.4
PointMamba[25]202484.486.2
PCP-MAE [70]202484.986.1
Ours
DAFNet-85.286.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.452, + 0.485, + 0.496 + ], + "angle": 0, + "content": "Table 4. Few-shot Results (\\%) on S3DIS. \\(S_{i}\\) denotes the split \\(i\\) is used for testing. Avg is their average mIoU. The best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.506, + 0.482, + 0.616 + ], + "angle": 0, + "content": "
Method2-way3-way
1-shot5-shot1-shot5-shot
S0S1AvgS0S1AvgS0S1AvgS0S1Avg
DGCNN [55]36.3438.7937.5756.4956.9956.7430.0532.1931.1246.8847.5747.23
ProtoNet [44]48.3949.9849.1957.3463.2260.2840.8145.0742.9449.0553.4251.24
MPTI [73]52.2751.4851.8858.9360.5659.7544.2746.9245.6051.7448.5750.16
AttMPTI [73]53.7755.9454.8661.6767.0264.3545.1849.2747.2354.9256.7955.86
BFG [29]55.6055.9855.7963.7166.6265.1746.1848.3647.2755.0557.8056.43
2CBR [75]55.8961.9958.9463.5567.5165.5346.5153.9150.2155.5158.0756.79
PAP3D [20]59.4566.0862.7665.4070.3067.8548.9956.5752.7861.2760.8161.04
Seg-PN [77]64.8467.9866.4167.6371.4869.5659.1160.4259.7759.4864.7262.10
DAFNet68.1370.2769.2070.5173.1571.8361.3365.5563.4465.2568.6766.96
Improvement+3.29+2.29+2.79+2.88+1.67+2.27+2.22+5.13+3.67+5.77+3.95+4.86
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.81 + ], + "angle": 0, + "content": "blocks. As shown in Table 4, DAFNet achieves state-of-the-art performance across all few-shot settings. In 2-way experiments, our method achieves \\(69.20\\%\\) mIoU in 1-shot and \\(71.83\\%\\) mIoU in 5-shot settings, surpassing the previous best method Seg-PN [77] by \\(2.79\\%\\) and \\(2.27\\%\\) respectively. For the more challenging 3-way scenario, DAFNet maintains its superiority with \\(63.44\\%\\) mIoU in 1-shot and \\(66.96\\%\\) mIoU in 5-shot settings, outperforming Seg-PN [77] by \\(3.67\\%\\) and \\(4.86\\%\\). These improvements across different settings demonstrate DAFNet's strong feature extraction capability and generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.414, + 0.837 + ], + "angle": 0, + "content": "4.3.2. Results analysis on the ScanNet dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "We further evaluate DAFNet on ScanNet [7], which contains 1,513 real-world scanned scenes with 20 semantic categories. Following [73], we process the scenes into 36,350 blocks with 2,048 points each. As shown in Ta" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 5. Few-shot Results (\\%) on ScanNet. \\( S_{i} \\) denotes the split \\( i \\) is used for testing. Avg is their average mIoU. The best results are shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.143, + 0.905, + 0.253 + ], + "angle": 0, + "content": "
Method2-way3-way
1-shot5-shot1-shot5-shot
\\( S_0 \\)\\( S_1 \\)Avg.\\( S_0 \\)\\( S_1 \\)Avg.\\( S_0 \\)\\( S_1 \\)Avg.\\( S_0 \\)\\( S_1 \\)Avg.
DGCNN [55]31.5528.9430.2542.7137.2439.9823.9919.1021.5534.9328.1031.52
ProtoNet [44]33.9230.9532.4445.3442.0143.6828.4726.1327.3037.3634.9836.17
MPTI [73]39.2736.1437.7146.9043.5945.2529.9627.2628.6138.1434.3636.25
AttMPTI [73]42.5540.8341.6954.0050.3252.1635.2330.7232.9846.7440.8043.77
BFG [29]42.1540.5241.3451.2349.3950.3134.1231.9833.0546.2541.3843.82
2CBR [75]50.7347.6649.2052.3547.1449.7547.0046.3646.6845.0639.4742.27
PAP3D [20]57.0855.9456.5164.5559.6462.1055.2755.6055.4459.0253.1656.09
Seg-PN [77]63.1564.3263.7467.0869.0568.0761.8065.3463.5762.9468.2665.60
DAFNet68.7969.9569.3770.9170.6070.7666.1466.7066.4268.9771.9570.46
Improvement+5.64+5.63+5.63+3.83+1.55+2.69+4.34+1.36+2.85+6.03+3.69+4.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.265, + 0.866, + 0.28 + ], + "angle": 0, + "content": "Table 6. Effect of different components of DAFNet." + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.291, + 0.868, + 0.513 + ], + "angle": 0, + "content": "
GSA Layer\nEdgeConv MHADAF-ConvOA (%)
Individual Components
✓XX88.6
X✓X89.4
XX✓91.7
Component Combinations
✓✓X91.3
✓X✓94.8
X✓✓93.5
✓✓✓96.7
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.706 + ], + "angle": 0, + "content": "ble 5, DAFNet achieves consistent improvements across all experimental settings. For 2-way scenarios, our method obtains \\(69.37\\%\\) mIoU in 1-shot and \\(70.76\\%\\) mIoU in 5-shot settings, surpassing Seg-PN [77] by \\(5.63\\%\\) and \\(2.69\\%\\) respectively. In the more challenging 3-way settings, DAFNet achieves \\(66.42\\%\\) mIoU in 1-shot and \\(70.46\\%\\) mIoU in 5-shot settings, outperforming Seg-PN [77] by \\(2.85\\%\\) and \\(4.86\\%\\). These substantial improvements across different settings demonstrate DAFNet's strong generalization ability and robustness in handling complex real-world scenes with limited training samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.715, + 0.669, + 0.731 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.737, + 0.907, + 0.768 + ], + "angle": 0, + "content": "All experiments are conducted on the ScanObjectNN [48] dataset unless explicitly stated otherwise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.776, + 0.809, + 0.792 + ], + "angle": 0, + "content": "4.4.1. Ablation Study of Each Component" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We conduct ablation experiments to analyze the effectiveness of each component in DAFNet, as shown in Table 6. Among individual components, DAF-Conv shows the strongest performance with \\(91.7\\%\\) accuracy, notably higher than single GSA layer components (EdgeConv: \\(88.6\\%\\), MHA: \\(89.4\\%\\)). When combining components, the fusion of DAF-Conv with EdgeConv achieves \\(94.8\\%\\), while DAF" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22188" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.089, + 0.444, + 0.104 + ], + "angle": 0, + "content": "Table 7. Effect of different local geometric priors \\( {h}_{j} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.155, + 0.115, + 0.421, + 0.211 + ], + "angle": 0, + "content": "
Local Geometric Prior hjOA(%)
[pj]93.6
[pj,pi]94.7
[pj,pipj-pi]96.2
[pj,pi, pj-pi, ||pj-pi||]96.7
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.236, + 0.484, + 0.328 + ], + "angle": 0, + "content": "Conv with MHA reaches \\(93.5\\%\\), both showing substantial improvements over single component performance. The full model integrating all components achieves the best performance of \\(96.7\\%\\), demonstrating the strong complementary effects between DAF-Conv's local geometric modeling and the GSA layer's global context capturing." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.336, + 0.423, + 0.351 + ], + "angle": 0, + "content": "4.4.2. Ablation Study of Local Geometric Prior" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.491 + ], + "angle": 0, + "content": "As shown in Table 7, we conduct ablation studies on the composition of local geometric prior \\( h_j \\). Starting with only center point coordinates \\( [p_j] \\) yields \\( 93.6\\% \\) accuracy. Incorporating neighbor point information \\( [p_j, p_i] \\) brings a \\( 1.1\\% \\) improvement. Adding relative position \\( [p_j, p_i, p_j - p_i] \\) further boosts performance by \\( 1.5\\% \\). The complete geometric prior with Euclidean distance \\( ||p_j - p_i|| \\) achieves the best result of \\( 96.7\\% \\), validating the importance of comprehensive geometric information in local structure modeling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.5, + 0.391, + 0.515 + ], + "angle": 0, + "content": "4.4.3. Analysis of Model Hyperparameters" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.55 + ], + "angle": 0, + "content": "We conduct comprehensive ablation studies on three key hyperparameters of DAFNet, as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.655 + ], + "angle": 0, + "content": "Frequency Component Number H: Fig. 4(a) shows that model performance improves as \\(H\\) increases from 1 to 6, achieving peak accuracy at \\(H = 6\\) \\((96.7\\%)\\). Further increasing \\(H\\) leads to performance degradation, dropping to \\(95.1\\%\\) at \\(H = 10\\). This indicates that while multiple frequency components enhance feature diversity, excessive components may introduce redundant information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.762 + ], + "angle": 0, + "content": "Distance Threshold \\( \\mathbf{d}_{\\mathrm{threshold}} \\): As shown in Fig. 4(b), we examine the impact of \\( d_{threshold} \\) which divides the local neighborhood into near-field and far-field regions. The optimal performance (96.7%) is achieved at \\( d_{threshold} = 0.6 \\). Lower thresholds limit near-field information, while higher values may include irrelevant points in the near-field region, both resulting in suboptimal performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Neighborhood Size K: Fig. 4(c) demonstrates the effect of neighborhood size \\(K\\) in EdgeConv. Performance reaches its peak \\((96.7\\%)\\) at \\(K = 20\\). Smaller \\(K\\) values provide insufficient local context, while larger values may incorporate irrelevant points, both leading to decreased accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.27, + 0.865 + ], + "angle": 0, + "content": "4.5. Model Complexity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We compare the model complexity of DAFNet with existing methods on ScanObjectNN [48], as shown in Ta" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.648, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.09, + 0.777, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.091, + 0.904, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.2, + 0.907, + 0.256 + ], + "angle": 0, + "content": "Figure 4. Ablation studies on key hyperparameters. (a) Impact of frequency component number \\( H \\) in DAF-Conv. (b) Effect of distance threshold \\( d_{threshold} \\) for near-field and far-field separation. (c) Influence of neighborhood size \\( K \\) in EdgeConv layer." + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.279, + 0.885, + 0.294 + ], + "angle": 0, + "content": "Table 8. Model complexity comparison on ScanObjectNN." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.306, + 0.902, + 0.493 + ], + "angle": 0, + "content": "
MethodsOA(%)Params.(M)FLOPs(G)
PointNet [35]68.03.50.5
PointNet++[36]77.91.51.7
DGCNN[55]78.11.82.4
PointCNN [24]78.50.6-
MVTN [18]82.811.243.7
PointMLP [28]85.412.631.4
PointNeXt [39]87.41.43.6
DAFNet (Ours)96.72.471.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.65 + ], + "angle": 0, + "content": "ble 8. While achieving state-of-the-art performance (96.7% OA), DAFNet maintains moderate computational requirements with only 2.47M parameters and 1.1G FLOPs. This is significantly more efficient than models like MVTN [18] (11.2M parameters, 43.7G FLOPs) and PointMLP [28] (12.6M parameters, 31.4G FLOPs), demonstrating that our acoustic field-inspired design achieves superior accuracy without sacrificing computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.675, + 0.634, + 0.691 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this paper, we presented DAFNet, a physics-inspired architecture for point cloud understanding that addresses local structure representation challenges. Inspired by acoustic field principles, we introduced AF-Conv modeling local feature aggregation through acoustic propagation, and extended it to DAF-Conv which dynamically adapts to diverse geometric features through multi-frequency simulation. By integrating these with GSA, our approach captures both fine-grained local structures and global context. Experiments demonstrate DAFNet achieves state-of-the-art performance across multiple tasks. Future work will extend acoustic field modeling to other domains, exploring applications in additional 3D tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22189" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.173 + ], + "angle": 0, + "content": "This project is sponsored by Shanghai Pujiang Programme 24PJD030. The computational work for this article is partially performed on resources of the National Supercomputing Centre, Singapore (https://www.nscc.sg)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.198, + 0.188, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.223, + 0.484, + 0.278 + ], + "angle": 0, + "content": "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, pages 1534-1543, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.281, + 0.484, + 0.322 + ], + "angle": 0, + "content": "[2] Binjie Chen, Yunzhou Xia, Yu Zang, Cheng Wang, and Jonathan Li. Decoupled local aggregation for point cloud learning. arXiv preprint arXiv:2308.16532, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.324, + 0.484, + 0.377 + ], + "angle": 0, + "content": "[3] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.379, + 0.482, + 0.42 + ], + "angle": 0, + "content": "[4] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. NeurIPS, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.422, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[5] Ziyang Chen, Israel D Gebru, Christian Richardt, Anurag Kumar, William Laney, Andrew Owens, and Alexander Richard. Real acoustic fields: An audio-visual room acoustics dataset and benchmark. In CVPR, pages 21886–21896, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.493, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[6] Pranav Singh Chib and Pravendra Singh. Recent advancements in end-to-end autonomous driving using deep learning: A survey. IEEE Transactions on Intelligent Vehicles, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.484, + 0.604 + ], + "angle": 0, + "content": "[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.661 + ], + "angle": 0, + "content": "[8] Jeevan S Devagiri, Sidike Paheding, Quamar Niyaz, Xiaoli Yang, and Samantha Smith. Augmented reality and artificial intelligence in industry: Trends, tools, and future challenges. Expert Systems with Applications, 207:118002, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[9] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[10] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. Mose: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[11] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, pages 8458-8468, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.223, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[13] Xiang Fang, Wanlong Fang, Changshuo Wang, Daizong Liu, Keke Tang, Jianfeng Dong, Pan Zhou, and Beibei Li. Multipair temporal sentence grounding via multi-thread knowledge transfer network. arXiv preprint arXiv:2412.15678, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.906, + 0.207 + ], + "angle": 0, + "content": "[14] Ruchi Goel and Pooja Gupta. Robotics and industry 4.0. A Roadmap to Industry 4.0: Smart Production, Sharp Business and Sustainable Development, pages 157-169, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.209, + 0.906, + 0.263 + ], + "angle": 0, + "content": "[15] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In ICML, pages 3809-3820. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.266, + 0.906, + 0.308 + ], + "angle": 0, + "content": "[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.31, + 0.906, + 0.364 + ], + "angle": 0, + "content": "[17] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.906, + 0.41 + ], + "angle": 0, + "content": "[18] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In ICCV, pages 1-11, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.906, + 0.466 + ], + "angle": 0, + "content": "[19] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In ACM MM, pages 4995-5004, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.523 + ], + "angle": 0, + "content": "[20] Shuting He, Xudong Jiang, Wei Jiang, and Henghui Ding. Prototype adaption and projection for few-and zero-shot 3d point cloud semantic segmentation. TIP, 32:3199-3211, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.906, + 0.596 + ], + "angle": 0, + "content": "[21] Limin Jiang, Changshuo Wang, Xin Ning, and Zaiyang Yu. Lttpoint: a mlp-based point cloud classification method with local topology transformation module. In 2023 7th Asian Conference on Artificial Intelligence Technology (ACAIT), pages 783-789. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.599, + 0.906, + 0.639 + ], + "angle": 0, + "content": "[22] Lawrence E Kinsler, Austin R Frey, Alan B Coppens, and James V Sanders. Fundamentals of acoustics. John wiley & sons, 2000. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.906, + 0.684 + ], + "angle": 0, + "content": "[23] Artem Komarichev, Zichun Zhong, and Jing Hua. A-cnn: Annually convolutional neural networks on point clouds. In CVPR, pages 7421-7430, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.906, + 0.728 + ], + "angle": 0, + "content": "[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. NeurIPS, 31, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.906, + 0.785 + ], + "angle": 0, + "content": "[25] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.842 + ], + "angle": 0, + "content": "[26] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual state space model. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22190" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[28] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[29] Yongqiang Mao, Zonghao Guo, LU Xiaonan, Zhiqiang Yuan, and Haowen Guo. Bidirectional feature globalization for few-shot semantic segmentation of 3d point cloud scenes. In 2022 International Conference on 3D Vision (3DV), pages 505-514. IEEE, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.277 + ], + "angle": 0, + "content": "[30] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.28, + 0.482, + 0.308 + ], + "angle": 0, + "content": "[31] Philip McCord Morse and K Uno Ingard. Theoretical acoustics. Princeton university press, 1986. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.311, + 0.482, + 0.365 + ], + "angle": 0, + "content": "[32] AAM Muzahid, Wanggen Wan, Ferdous Sohel, Lianyao Wu, and Li Hou. Curvenet: Curvature-based multitask learning deep networks for 3d object recognition. IEEE/CAA Journal of Automatica Sinica, 8(6):1177-1187, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.369, + 0.482, + 0.423 + ], + "angle": 0, + "content": "[33] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, pages 604-621. Springer, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.482, + 0.48 + ], + "angle": 0, + "content": "[34] Jinyoung Park, Sanghyeok Lee, Sihyeon Kim, Yunyang Xiong, and Hyunwoo J Kim. Self-positioning point-based transformer for point cloud understanding. In CVPR, pages 21814-21823, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.528, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NIPS, 30, 2017. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.573, + 0.482, + 0.64 + ], + "angle": 0, + "content": "[37] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In ICML, pages 28223-28243. PMLR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.644, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[38] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. *Shapellm: Universal 3d object understanding for embodied interaction*. In *ECCV*, pages 214–238. Springer, 2025. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[39] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. NIPS, 35:23192-23204, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[40] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In CVPR, pages 18942-18952, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.898 + ], + "angle": 0, + "content": "[41] Jingjing Ren, Ashwin Rao, Martina Lindorfer, Arnaud Legout, and David Choffnes. Recon: Revealing and controlling pii leaks in mobile network traffic. In Proceedings of the 14th Annual International Conference on Mobile Systems, Applications, and Services, pages 361-374, 2016. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[42] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, pages 3577-3586, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[43] Mickael Sereno, Xiyao Wang, Lonni Besançon, Michael J Mcguffin, and Tobias Isenberg. Collaborative work in augmented reality: A survey. TVCG, 28(6):2530-2549, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.222 + ], + "angle": 0, + "content": "[44] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. NIPS, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.226, + 0.905, + 0.266 + ], + "angle": 0, + "content": "[45] Mohsen Soori, Behrooz Arezoo, and Roza Dastres. Artificial intelligence, machine learning and deep learning in advanced robotics, a review. Cognitive Robotics, 3:54-70, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.269, + 0.905, + 0.324 + ], + "angle": 0, + "content": "[46] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. In ECCV, pages 426-442. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.327, + 0.905, + 0.381 + ], + "angle": 0, + "content": "[47] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6411–6420, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[48] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, pages 1588–1597, 2019. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[49] Changshuo Wang, Xin Ning, Linjun Sun, Liping Zhang, Weijun Li, and Xiao Bai. Learning discriminative features by covering local geometric space for point cloud analysis. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.528, + 0.905, + 0.57 + ], + "angle": 0, + "content": "[50] C Wang, H Wang, X Ning, S Tian, and W Li. 3d point cloud classification method based on dynamic coverage of local area. J. Softw, 34(4):1962-1976, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[51] Changshuo Wang, Xin Ning, Weijun Li, Xiao Bai, and Xingyu Gao. 3d person re-identification based on global semantic guidance and local feature aggregation. TCSVT, 34 (6):4698-4712, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[52] Changshuo Wang, Meiqing Wu, Siew-Kei Lam, Xin Ning, Shangshu Yu, Ruiping Wang, Weijun Li, and Thambipillai Srikanthan. Gpsformer: A global perception and local structure fitting-based transformer for point cloud understanding. arXiv preprint arXiv:2407.13519, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[53] Changshuo Wang, Rongsheng Cao, and Ruiping Wang. Learning discriminative topological structure information representation for 2d shape and social network classification via persistent homology. Knowledge-Based Systems, page 113125, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[54] Ruiping Wang, Siew-Kei Lam, Meiqing Wu, Zhijian Hu, Changshuo Wang, and Jing Wang. Destination intention estimation-based convolutional encoder-decoder for pedestrian trajectory multimodality forecast. Measurement, 239: 115470, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[55] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22191" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[56] Xin Wei, Ruixuan Yu, and Jian Sun. View-gcn: View-based graph convolutional network for 3d shape analysis. In CVPR, pages 1850-1859, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[57] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[58] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, pages 1912-1920, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[60] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, pages 3173-3182, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[61] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In CVPR, pages 1179-1189, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[62] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. NIPS, 32, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[63] Ze Yang and Liwei Wang. Learning relationships for multiview 3d object recognition. In ICCV, pages 7505-7514, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[64] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[65] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, pages 19313-19322, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[66] Huang Zhang, Changshuo Wang, Long Yu, Shengwei Tian, Xin Ning, and Joel Rodrigues. Pointgt: A method for pointcloud classification and segmentation based on local geometric transformation. IEEE Transactions on Multimedia, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[67] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. NeurIPS, 35:27061-27074, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[68] Renrui Zhang, Liuhui Wang, Ziyu Guo, Yali Wang, Peng Gao, Hongsheng Li, and Jianbo Shi. Parameter is not all you need: Starting from non-parametric networks for 3d point cloud analysis. arXiv preprint arXiv:2303.08134, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.869 + ], + "angle": 0, + "content": "[69] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[70] Xiangdong Zhang, Shaofeng Zhang, and Junchi Yan. Pcmae: Learning to predict centers for point masked autoen" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "coders. Advances in Neural Information Processing Systems, 37:80303-80327, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[71] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[72] Jingyuan Zhao, Wenyi Zhao, Bo Deng, Zhenghong Wang, Feng Zhang, Wenxiang Zheng, Wanke Cao, Jinrui Nan, Yubo Lian, and Andrew F Burke. Autonomous driving system: A comprehensive survey. Expert Systems with Applications, page 122836, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[73] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In CVPR, pages 8873-8882, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[74] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In CVPR, pages 22935-22945, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.904, + 0.375 + ], + "angle": 0, + "content": "[75] Guanyu Zhu, Yong Zhou, Rui Yao, and Hancheng Zhu. Cross-class bias rectification for point cloud few-shot segmentation. TMM, 25:9175-9188, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[76] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[77] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ji-aming Liu, Han Xiao, Chaoyou Fu, Hao Dong, and Peng Gao. No time to train: Empowering non-parametric networks for few-shot 3d scene segmentation. In CVPR, pages 3838–3847, 2024. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22192" + } + ] +] \ No newline at end of file diff --git a/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/7aa2860d-04df-426a-bde9-40b0ed00d4aa_origin.pdf b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/7aa2860d-04df-426a-bde9-40b0ed00d4aa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7b7f3ed3ab83dc44b2256860690d98408fbef05b --- /dev/null +++ b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/7aa2860d-04df-426a-bde9-40b0ed00d4aa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e6271ad90a36fcf334fc048424826f57abad818caf25a68c3cb20ae590bb59b +size 413255 diff --git a/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/full.md b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f3af88217476a43ebb8cda4dd0ef5b291450f329 --- /dev/null +++ b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/full.md @@ -0,0 +1,438 @@ +# Point Clouds Meets Physics: Dynamic Acoustic Field Fitting Network for Point Cloud Understanding + +Changshuo Wang + +NTU, Singapore + +wangchangshuol@gmail.com + +Shuting He + +SUFE, China + +shuting.he@sufe.edu.cn + +Xiang Fang* + +NTU, Singapore + +xfang9508@gmail.com + +Jiawei Han + +BIT, China + +hanjiawei@bit.edu.cn + +Zhonghang Liu + +SMU, Singapore + +zhliu.2020@phdcs.smu.edu.sg + +Xin Ning + +Institute of Semiconductors, CAS + +ningxin@semi.ac.cn + +# Weijun Li +Institute of Semiconductors, CAS + +wjli@semi.ac.cn + +# Prayag Tiwari +Halmstad University, Sweden + +prayag.tiwari@ieee.org + +# Abstract + +While existing pre-training-based methods have enhanced point cloud model performance, they have not fundamentally resolved the challenge of local structure representation in point clouds. The limited representational capacity of pure point cloud models continues to constrain the potential of cross-modal fusion methods and performance across various tasks. To address this challenge, we propose a Dynamic Acoustic Field Fitting Network (DAF-Net), inspired by physical acoustic principles. Specifically, we represent local point clouds as acoustic fields and introduce a novel Acoustic Field Convolution (AF-Conv), which treats local aggregation as an acoustic energy field modeling problem and captures fine-grained local shape awareness by dividing the local area into near field and far field. Furthermore, drawing inspiration from multifrequency wave phenomena and dynamic convolution, we develop the Dynamic Acoustic Field Convolution (DAF-Conv) based on AF-Conv. DAF-Conv dynamically generates multiple weights based on local geometric priors, effectively enhancing adaptability to diverse geometric features. Additionally, we design a Global Shape-Aware (GSA) layer incorporating EdgeConv and multi-head attention mechanisms, which combines with DAF-Conv to form the DAF Block. These blocks are then stacked to create a hierarchical DAFNet architecture. Extensive experiments demonstrate that DAFNet significantly outperforms existing methods across multiple tasks. + +![](images/5ec06a8d1a7adeb879f68d13b5ef6ea34882045597c6c6a5cc6446db0b3077b9.jpg) +Local Neighborhood +Multi-Acoustic Energy Field Mapping +Shape-aware Representation +Figure 1. Illustration of local feature extraction using DAF-Conv. Given a local neighborhood of a point cloud (left), our method first models it as an acoustic field with near-field and far-field zones. Then, multiple acoustic energy field mappings are generated to capture various geometric patterns. Finally, DAF-Conv processes these mappings to produce shape-aware representations. + +# 1. Introduction + +Point cloud data, due to its direct representation of object surface information in three-dimensional space, has been widely applied in various scenarios including autonomous driving [6, 54, 72], robotic navigation [14, 45], and augmented reality [8, 43]. However, the sparsity, irregularity, and complexity of local structures in point cloud data make it challenging for existing models to fully capture geometric features, thus limiting their applications in 3D tasks. + +Recently, several methods [33, 37, 38] have attempted to indirectly improve point cloud model performance through pre-training strategies or cross-modal contrastive learning. Although these approaches have enhanced point cloud processing tasks to some extent, they have not directly addressed the problem of local structure representation. Some researchers [19, 21, 25, 50] have also attempted to apply new architectures to point cloud understanding tasks. However, the local representational capacity of point cloud mod + +els remains a bottleneck in improving overall model performance. Therefore, how to efficiently capture and represent the local structures of point clouds has become a crucial problem that urgently needs to be solved. + +Physics, particularly the principles of acoustic field propagation [5], provides a new perspective for understanding local geometric structures. In acoustics, wave propagation is influenced by environmental geometric characteristics, and through constructing acoustic field models, we can effectively simulate wave behavior at different frequencies to obtain information in complex environments. Inspired by this, we propose a novel point cloud understanding method called Dynamic Acoustic Field Fitting Network (DAFNet). This method borrows ideas from physical acoustic field propagation, representing local point clouds as acoustic fields and dynamically adjusting weights in different regions to capture local structural features more precisely. + +Specifically, we propose a novel Acoustic Field Convolution (AF-Conv) that models local aggregation as an acoustic energy propagation problem. To more finely model the contributions of different neighboring points to the center point, AF-Conv divides the local region into near-field and far-field zones for separate geometric modeling. Furthermore, as shown in Fig. 1, based on AF-Conv, we propose Dynamic Acoustic Field Convolution (DAF-Conv), which dynamically generates multiple weights based on geometric priors while simulating geometric feature variations under different "frequencies", thereby enhancing the model's adaptability to local geometric diversity. Finally, we combine a Global Shape-Aware (GSA) layer comprising EdgeConv and Multi-Head Attention (MHA) with DAF-Conv to build hierarchical DAF Blocks, forming the complete DAFNet. The experimental results show that DAFNet achieves superior performance over existing methods in multiple tasks, demonstrating its effectiveness in capturing complex geometric structures. + +# The main contributions of this work are as follows: + +- We propose a physics-inspired end-to-end DAFNet that can learn rich contextual information and precise shape representations in irregular point clouds. +- We introduce a novel Acoustic Field Convolution (AFConv) that achieves fine-grained local shape awareness by modeling local point cloud regions as acoustic fields. +- We design Dynamic Acoustic Field Convolution (DAFConv) that adapts to geometric feature diversity by dynamically generating multiple frequency-aware weights from local geometry. +- The proposed DAFNet achieves state-of-the-art performance in multiple tasks, reaching $96.7\%$ accuracy on the challenging ScanObjectNN with only 2.47M parameters. + +# 2. Related Work + +# 2.1. CNN-based Methods + +Early works [30, 42, 56, 63] in point cloud processing primarily involved projecting them onto 2D images for 2D CNN processing or converting them into regular 3D grid structures for 3D convolution operations. However, these transformation processes inevitably led to spatial information loss and increased computational costs. To address this issue, researchers began applying deep learning techniques directly in 3D space. PointNet [35] pioneered this direction. However, this method overlooked local structural information. Subsequently, PointNet++ [36] enhanced the local perception capability of point cloud models by grouping point clouds into different local neighborhoods. Based on PointNet++ [36], existing works [23, 49, 51, 52, 55, 66] have developed methods based on point-wise perceptrons, convolution operations, and attention mechanisms. Nevertheless, these methods still struggle to effectively model local structural shape information. + +# 2.2. Transformer-based Methods + +Transformers have demonstrated remarkable success in various tasks [9, 10, 13, 53] due to their inherent ability to capture global features. Leveraging this strength, researchers have progressively developed a range of Transformer-based point cloud models. PCT [17] enhances input embeddings through furthest point sampling and nearest neighbor search while using self-attention mechanisms to capture global features. Point Transformer [71] employs vector attention to extract aggregated local features. Subsequently, updated versions [57, 58] of Point Transformer explored different architectures to improve performance and efficiency. Inspired by Swin Transformer [27], works such as SST [12] and SWFormer [46] explored window attention mechanisms to improve efficiency. Although these methods excel at capturing global features, their performance hasn't shown significant advancement over CNN-based approaches. + +# 2.3. State Space Model-based Methods + +State Space Models (SSMs) [16] are mathematical models for describing dynamic systems. Recently, SSMs have emerged as a reliable alternative architecture to CNNs and Transformers. In computer vision, Vision Mamba [76] successfully compresses visual representations through bidirectional state space models, while VMamba [26] innovatively introduces cross-scanning modules, enabling global receptive fields in 2D images through 1D selective scanning. Researchers have explored SSM applications in point cloud processing. PCM [69] first introduced the Mamba architecture to point cloud processing through octree serialization strategies. Mamba3D [19] enhanced local geometric feature extraction through LNP blocks and designed bidi + +![](images/a56044a4d3b301eff35015c04723e07ca735b3db485842c06c6f6f8a22e0a8da.jpg) +Figure 2. Architecture of DAFNet. (a) For shape classification, the network consists of three sequential DAF Blocks followed by max pooling and classifier layers. Each DAF Block contains a GSA layer and DAF-Conv for feature extraction. (b) For part segmentation and few-shot semantic segmentation tasks, the network adopts an encoder-decoder structure with skip connections, where the encoder shares the same architecture as the classification network, and the decoder uses feature propagation to restore point cloud resolution. The few-shot version additionally incorporates support and query mechanisms for processing different point cloud inputs. + +rectional SSM structures. PointMamba [25], another significant work, customized SSM architecture for point cloud analysis tasks, substantially reducing computational costs. + +# 2.4. Pre-training-based Methods + +Point cloud pre-training methods primarily comprise generative approaches [4, 33, 67] and cross-modal contrastive learning methods. In generative approaches, Point-MAE [33] extends the MAE concept by pre-training through random masking and reconstruction of point cloud segments PCP-MAE [70] learns shape priors by predicting missing regional point cloud structures. PointDif [74] utilizes diffusion models for pre-training, learning hierarchical geometric priors through point cloud recovery at different noise levels. In cross-modal contrastive learning, ReCon [37] combines generative modeling and contrastive learning paradigms, addressing representation overfitting in contrastive learning and data padding issues in generative methods. ULIP [61] achieves cross-modal knowledge transfer through a language-image-point cloud pre-training framework, while ACT [11] employs cross-modal autoencoders as teacher models to acquire knowledge from other modalities. Although these pre-training methods have advanced point cloud processing performance, they haven't fundamentally resolved the point cloud representation challenge. + +# 3. Methods + +This section first reviews point convolution and dynamic convolution, followed by introducing acoustic field convolution. Then, we present our proposed dynamic acoustic field convolution. Finally, we describe DAFNet (as shown in Fig. 2). + +# 3.1. Background + +# 3.1.1. Point Convolution + +Local feature aggregation in point clouds typically follows a "sampling-grouping-aggregation" procedure. The aggregation operation is crucial as it directly affects local feature representation by modeling local point clouds. We denote center points as $\{(p_i, f_i)\}_{i=1}^M$ , where $M$ represents the number of points, $p_i \in \mathbb{R}^{1 \times 3}$ and $f_i \in \mathbb{R}^{1 \times C}$ represent the coordinates and features of the $i$ -th point respectively, and $C$ denotes the number of feature channels. The most commonly used method for local feature aggregation based on coordinate space is to construct point convolution, which can be formalized as: + +$$ +f _ {i} ^ {\prime} = \mathcal {A} (\{\mathcal {M} (p _ {i}, p _ {j}) \cdot \mathcal {T} (f _ {i}, f _ {j}) | p _ {j} \in \mathcal {N} (p _ {i}) \}), \quad (1) +$$ + +where $\mathcal{A}$ represents the aggregation function, typically max-pooling. $\mathcal{M}$ and $\mathcal{T}$ are mapping functions, usually imple + +mented through Multi-Layer Perceptrons (MLPs). $\mathcal{N}(p_i)$ represents the local neighborhood of $p_i$ , and $p_j$ denotes the neighboring points of $p_i$ . + +An alternative aggregation approach operates in feature space, commonly known as Dynamic Graph Convolution (DGC) [55]. It aggregates local features in feature space, meaning that points close in feature space might be distant in coordinate space. DGC can be formalized as: + +$$ +f _ {i} ^ {\prime} = \mathcal {A} (\{\mathcal {T} (f _ {i}, f _ {j}) | f _ {j} \in \mathcal {N} (f _ {i}) \}). \tag {2} +$$ + +# 3.1.2. Dynamic Convolution + +Dynamic convolution [62] aims to enhance network modeling capability by dynamically generating multiple convolution kernels based on input data, adapting to different data structures. Taking PACConv's [60] dynamic convolution as an example, dynamic convolution in point clouds can be expressed as: + +$$ +g _ {i} = \mathcal {A} \left(\left\{w \left(p _ {j}\right) \cdot f _ {j} \mid p _ {j} \in \mathcal {N} \left(p _ {i}\right) \right\}\right), \tag {3} +$$ + +where $w(p_j)$ denotes the weight corresponding to feature $f_j$ of point $p_j$ . $w(p_j)$ typically consists of multiple weights $w_h(p_i)$ and their corresponding attention coefficients $\alpha_h(p_j)$ generated in a data-driven manner: + +$$ +w \left(p _ {j}\right) = \sum_ {h = 1} ^ {H} \alpha_ {h} \left(p _ {j}\right) \cdot w _ {h} \left(p _ {j}\right), \tag {4} +$$ + +where $\sum_{h=1}^{H} \alpha_h(p_j) = 1$ , $H$ represents the number of $w_h(p_j)$ . + +Therefore, formula 3 can be expressed as: + +$$ +g _ {i} = \mathcal {A} \left(\left\{\sum_ {h = 1} ^ {H} \alpha_ {h} \left(p _ {j}\right) \cdot w _ {h} \left(p _ {j}\right) \cdot f _ {j} \mid p _ {j} \in \mathcal {N} \left(p _ {i}\right) \right\}\right). \tag {5} +$$ + +# 3.2. Acoustic Field Convolution + +In physical acoustics [22, 31], the sound pressure field $p(x)$ generated by a sound source at spatial point $x$ can be represented by Green's function: + +$$ +p (x) = \frac {e ^ {- j k \| x - y \|}}{4 \pi \| x - y \|}, \tag {6} +$$ + +where $y$ is the source position, $k$ is the wave number, and $\| x - y\|$ represents the distance between observation point $x$ and source $y$ . This formula reveals the fundamental physical laws of sound wave propagation: sound pressure intensity decays exponentially with distance while being affected by spherical spreading attenuation. + +Notably, based on the distance between the source and observation point, the acoustic field can be divided into near-field and far-field regions. In the near-field region, the + +![](images/35d3576f695b2400d227bcea514c6e0feece7fa5a39904e756b17e6bd478a5d1.jpg) +Figure 3. Illustration of AF-Conv and DAF-Conv. (a) AF-Conv divides local neighborhood into near-field and far-field regions and calculates energy weights based on geometric priors. (b) DAF-Conv extends AF-Conv with multiple parallel acoustic field convolutions to achieve dynamic feature aggregation. + +relationship between sound pressure and distance is more complex, exhibiting strong local correlations; in the farfield region, the sound pressure demonstrates simple spherical wave attenuation characteristics. + +Inspired by this theory, we view local feature aggregation as acoustic field propagation, where the center point $p_i$ can be viewed as the sound source and neighboring points $p_j$ as observation points. Meanwhile, we refine the local neighborhood into near-field and far-field regions to extract fine-grained information, constructing the Acoustic Field Convolution (AF-Conv). As shown in Fig. 3(a), it can be formalized as: + +$$ +g _ {i} = \mathcal {A} \left(\left\{W _ {A F} \left(p _ {i}, p _ {j}\right) \cdot f _ {j} \mid p _ {j} \in \mathcal {N} \left(p _ {i}\right) \right\}\right), \tag {7} +$$ + +where the weight function $W_{AF}(p_i,p_j)$ directly draws inspiration from acoustic field propagation characteristics and designs different weight calculation methods for near-field and far-field regions: + +$$ +W _ {A F} \left(p _ {i}, p _ {j}\right) = \left\{ \begin{array}{l l} e ^ {- \beta d _ {i j}}, & \text {i f} d _ {i j} < d _ {\text {t h r e s h o l d}} \\ \frac {1}{d _ {i j} + \epsilon}, & \text {o t h e r w i s e} \end{array} , \right. \tag {8} +$$ + +where $\beta$ represents learnable coefficient. $d_{ij}$ denotes the Euclidean distance between $p_j$ and $p_i$ . In $W_{AF}(p_i,p_j)$ , for the near-field region ( $d_{ij} < d_{threshold}$ ), we use the exponential decay term $e^{-\beta d_{ij}}$ to capture subtle changes in local geometric structures, corresponding to the complex attenuation characteristics of the acoustic near-field region; for the far-field region ( $d_{ij} \geq d_{threshold}$ ), we adopt a simple reciprocal relationship $\frac{1}{d_{ij} + \epsilon}$ to better perceive broader contextual information. + +# 3.3. Dynamic Acoustic Field Convolution + +In sound wave propagation [31], waves of different frequencies exhibit distinct propagation characteristics. Inspired by + +this multi-frequency phenomenon and dynamic convolution [62], as shown in Fig. 3(b), we propose Dynamic Acoustic Field Convolution (DAF-Conv), which extends AF-Conv by dynamically generating multiple sets of weights to simulate local geometric feature variations under different "frequencies". It can be formalized as: + +$$ +F _ {h} \left(p _ {i}, p _ {j}\right) = W _ {h} \left(p _ {i}, p _ {j}\right) \cdot f _ {j}, \tag {9} +$$ + +$$ +W _ {h} \left(p _ {i}, p _ {j}\right) = \left\{ \begin{array}{l l} \alpha_ {h} e ^ {- \beta_ {h} d _ {i j}}, & \text {i f} d _ {i j} < d _ {\text {t h r e s h o l d}} \\ \frac {\alpha_ {h}}{d _ {i j} + \epsilon}, & \text {o t h e r w i s e ,} \end{array} \right. \tag {10} +$$ + +$$ +\left\{\alpha_ {h}, \beta_ {h} \right\} _ {h = 1} ^ {H} = \operatorname {M L P} \left(h _ {j}\right), \tag {11} +$$ + +where $d_{threshold}$ is a scalar that defines the boundary between near-field and far-field regions. The weight coefficient $\alpha_{h}$ and $\beta_{h}$ are generated from local geometric prior $h_j = [p_i,p_j,p_j - p_i,\| p_j - p_i\| ]$ through MLP to simulate propagation characteristics of different frequency components. + +The local aggregated features are then separated into near-field and far-field components based on the distance threshold: + +$$ +F _ {h} ^ {\text {n e a r}} \left(p _ {i}\right) = \mathcal {A} \left(\left\{F _ {h} \left(p _ {i}, p _ {j}\right) \mid j \in \mathcal {N} _ {\text {n e a r}} (i) \right\}\right), \tag {12} +$$ + +$$ +F _ {h} ^ {f a r} \left(p _ {i}\right) = \mathcal {A} \left(\left\{F _ {h} \left(p _ {i}, p _ {j}\right) \mid j \in \mathcal {N} _ {f a r} (i) \right\}\right), \tag {13} +$$ + +where $\mathcal{N}_{near}(i) = \{j\mid p_j\in \mathcal{N}(p_i),d_{ij} < d_{threshold}\}$ represents the set of near-field neighbors of the center point $p_i$ and $\mathcal{N}_{far}(i) = \{j\mid p_j\in \mathcal{N}(p_i),d_{ij}\geq d_{threshold}\}$ represents the set of far-field neighbors. + +The component for each frequency $h$ can be obtained by: + +$$ +F _ {h} \left(p _ {i}\right) = F _ {h} ^ {\text {n e a r}} \left(p _ {i}\right) + F _ {h} ^ {\text {f a r}} \left(p _ {i}\right). \tag {14} +$$ + +To achieve adaptive fusion of features from different frequency components, we introduce an attention mechanism based on geometric prior $h_j$ . The attention weights are generated as fellow: + +$$ +\gamma_ {h} = \operatorname {s o f t m a x} (\operatorname {M a x P o o l i n g} (\operatorname {M L P} (h _ {j}))). \tag {15} +$$ + +The final local aggregated feature is obtained by weighted summation of features from all frequency components: + +$$ +g _ {i} = \sum_ {h = 1} ^ {H} \gamma_ {h} \cdot F _ {h} \left(p _ {i}\right). \tag {16} +$$ + +DAF-Conv achieve adaptive modeling of local geometric features through separately processing near-field and far-field features for each frequency component, independent max-pooling aggregation for near-field and far-field features, and adaptive fusion of multiple frequency components through learned attention weights. + +Compared to the single weight function in AF-Conv, DAF-Conv achieves more sophisticated feature extraction capabilities through its multi-frequency design and separate processing of near-field and far-field features. + +# 3.4. Dynamic Acoustic Field Convolution Neural Network + +To ensure each point has rich feature representation before constructing local neighborhoods, we build a Global Shape-Aware (GSA) layer before applying DAF-Conv. The GSA layer first aggregates each point through EdgeConv [55] without reducing point cloud resolution, then applies Multi-Head Attention (MHA) for global context modeling on the enhanced points. This layer effectively enhances DAF-Conv's perception ability for local structure modeling. + +As shown in Fig. 2, based on the GSA layer and DAF-Conv, we construct the DAF Block. By stacking these blocks, we design a novel pretraining-free network, DAFNet, for point cloud analysis. We implement models for three major tasks: + +Point Cloud Classification. For point cloud classification, we employ three DAF Blocks, followed by a classifier consisting of max-pooling and fully connected layers. The output feature dimensions of the three blocks are 64, 128, 256 respectively. Notably, with only 2.47M parameters, it achieves the best performance on ScanObjectNN [48]. + +Part Segmentation and Few-shot Point Cloud Semantic Segmentation. For part segmentation, the encoder structure remains identical to the classification task. In the decoder stage, we employ feature propagation to restore point cloud resolution. Between the encoder and decoder, we apply U-Net-like skip connections to fully utilize contextual information. In few-shot segmentation, while maintaining the same encoder-decoder structure as part segmentation, we incorporate the Query module from SegPN[77] to reduce domain gaps between query and support features. + +# 4. Experiments + +# 4.1.3D Shape Classification + +Real-world Object Classification on ScanObjectNN. We evaluate our method on ScanObjectNN [48], a challenging real-world point cloud dataset containing 2,902 objects across 15 categories. Following previous works, we conduct experiments on its most challenging variant (PB_T50_RS), where objects are presented with background and occlusion. + +As shown in Table 1, DAFNet achieves state-of-the-art performance on ScanObjectNN [48], reaching $96.7\%$ overall accuracy (OA) and $96.3\%$ mean class accuracy (mAcc) with voting mechanism. These results significantly surpass both existing supervised methods (e.g., GPSFormer [52]: $95.4\%$ OA) and pre-training approaches (e.g., ReCon++ + +Table 1. Classification results on the ScanObjectNN [48] dataset. " - " denotes unknown. + +
MethodYearmAcc (%)OA (%)
Supervised Methods
PointNet [35]201763.468.2
PointNet++ [36]201869.873.7
DGCNN [55]201973.678.1
SimpleView [15]2021-80.5
MVTN [18]202183.185.5
PointMLP [28]202284.485.7
RepSurf-U [40]202283.186.0
PointNeXt [39]202285.887.7
Point-PN [68]2023-87.1
DeLA [2]202389.390.4
GPSFormer [52]202493.895.4
Mamba3D [19]2024-92.6
Pre-training Methods
Point-BERT [65]2021-83.1
Point-MAE [33]2022-85.2
ULIP-2 + PointNeXt [61]202391.291.5
PointGPT [3]2023-93.6
ReCon [41]2023-91.3
ReCon++ [38]2024-95.25
PointMamba [25]2024-89.3
PCP-MAE [70]2024-90.4
Ours
DAFNet (w/o voting)-96.196.4
DAFNet (w/ voting)-96.396.7
+ +[38]: $95.25\%$ OA). The superior performance demonstrates the effectiveness of our acoustic field-inspired architecture in handling real-world point cloud data. + +Synthetic Object Classification on ModelNet40. We further evaluate DAFNet on ModelNet40 [59], a widely-used synthetic dataset containing 12,311 CAD models from 40 categories, with 9,843 for training and 2,468 for testing. + +As presented in Table 2, DAFNet achieves competitive results on ModelNet40 [59] with $94.1\%$ overall accuracy (OA) and $92.2\%$ mean class accuracy (mAcc). While these results are slightly lower than some recent methods like PointGPT [3] ( $94.9\%$ OA), the performance gap is minimal (less than $1\%$ ). This is mainly because ModelNet40 [3], as a synthetic dataset, has been well-addressed by existing methods. In contrast, DAFNet shows more significant advantages on real-world datasets, as demonstrated by our results on ScanObjectNN [48]. + +Table 2. Classification results on ModelNet40 dataset. "-" denotes unknown. + +
MethodsYearmAcc (%)OA (%)
Supervised Methods
PointNet [35]201786.089.2
PointNet++[36]2017-90.7
DGCNN[55]201990.292.9
Point Transformer [71]202190.693.7
PAConv [60]2021-93.9
CurveNet [32]2021-94.2
PointMLP [28]202291.494.5
RepSurf-U [40]2022-94.7
GPSFormer [52]202492.294.2
Mamba3D [19]2024-94.1
Pre-training Methods
Point-BERT [65]2021-93.8
Point-MAE [33]2022-94.0
ULIP + PointMLP [61]202392.494.7
PointGPT [3]2023-94.9
ReCon [41]2023-94.5
ReCon++ [38]2024-94.8
PointMamba [25]2024-93.6
PCP-MAE [70]2024-94.0
Ours
DAFNet (w/o voting)-91.993.8
DAFNet (w/ voting)-92.294.1
+ +# 4.2. Part Segmentation + +To evaluate our method's capability in fine-grained shape understanding, we conduct experiments on ShapeNetPart [64], which contains 16,881 shapes from 16 categories with 50 part annotations. Following convention, we use 13,807 samples for training and 2,874 for testing. + +As shown in Table 3, DAFNet achieves competitive performance with $85.2\%$ class mIoU and $86.8\%$ instance mIoU, comparable to state-of-the-art methods. These results demonstrate that our acoustic field-inspired approach effectively captures both global shape information and local geometric details necessary for part segmentation tasks. + +# 4.3. Few-shot Point Cloud Semantic Segmentation + +# 4.3.1. Results analysis on the S3DIS dataset + +We evaluate our model on S3DIS [1], which contains 3D point clouds from 272 rooms with 13 semantic categories. Following [73], we divide scenes into $1\mathrm{m} \times 1\mathrm{m}$ blocks and sample 2048 points per block, resulting in 7547 + +Table 3. Part segmentation results (\%) on ShapeNetPart. We report both class mIoU and instance mIoU. + +
MethodsYearclass mIoUinstance mIoU
Supervised Methods
PointNet [35]201780.483.7
PointNet++[36]201781.985.1
DGCNN [55]201982.385.1
KPConv [47]201985.186.4
Point Transformer [71]202183.786.6
CurveNet [32]2021-86.8
PointNeXt [39]202285.287.1
SPOTr [34]202385.487.2
GPSFormer [52]202485.486.8
Pre-training Methods
PointGPT[3]202384.186.2
ReCon [41]202384.886.4
PointMamba[25]202484.486.2
PCP-MAE [70]202484.986.1
Ours
DAFNet-85.286.8
+ +Table 4. Few-shot Results (\%) on S3DIS. $S_{i}$ denotes the split $i$ is used for testing. Avg is their average mIoU. The best results are shown in bold. + +
Method2-way3-way
1-shot5-shot1-shot5-shot
S0S1AvgS0S1AvgS0S1AvgS0S1Avg
DGCNN [55]36.3438.7937.5756.4956.9956.7430.0532.1931.1246.8847.5747.23
ProtoNet [44]48.3949.9849.1957.3463.2260.2840.8145.0742.9449.0553.4251.24
MPTI [73]52.2751.4851.8858.9360.5659.7544.2746.9245.6051.7448.5750.16
AttMPTI [73]53.7755.9454.8661.6767.0264.3545.1849.2747.2354.9256.7955.86
BFG [29]55.6055.9855.7963.7166.6265.1746.1848.3647.2755.0557.8056.43
2CBR [75]55.8961.9958.9463.5567.5165.5346.5153.9150.2155.5158.0756.79
PAP3D [20]59.4566.0862.7665.4070.3067.8548.9956.5752.7861.2760.8161.04
Seg-PN [77]64.8467.9866.4167.6371.4869.5659.1160.4259.7759.4864.7262.10
DAFNet68.1370.2769.2070.5173.1571.8361.3365.5563.4465.2568.6766.96
Improvement+3.29+2.29+2.79+2.88+1.67+2.27+2.22+5.13+3.67+5.77+3.95+4.86
+ +blocks. As shown in Table 4, DAFNet achieves state-of-the-art performance across all few-shot settings. In 2-way experiments, our method achieves $69.20\%$ mIoU in 1-shot and $71.83\%$ mIoU in 5-shot settings, surpassing the previous best method Seg-PN [77] by $2.79\%$ and $2.27\%$ respectively. For the more challenging 3-way scenario, DAFNet maintains its superiority with $63.44\%$ mIoU in 1-shot and $66.96\%$ mIoU in 5-shot settings, outperforming Seg-PN [77] by $3.67\%$ and $4.86\%$ . These improvements across different settings demonstrate DAFNet's strong feature extraction capability and generalization ability. + +# 4.3.2. Results analysis on the ScanNet dataset + +We further evaluate DAFNet on ScanNet [7], which contains 1,513 real-world scanned scenes with 20 semantic categories. Following [73], we process the scenes into 36,350 blocks with 2,048 points each. As shown in Ta + +Table 5. Few-shot Results (\%) on ScanNet. $S_{i}$ denotes the split $i$ is used for testing. Avg is their average mIoU. The best results are shown in bold. + +
Method2-way3-way
1-shot5-shot1-shot5-shot
\( S_0 \)\( S_1 \)Avg.\( S_0 \)\( S_1 \)Avg.\( S_0 \)\( S_1 \)Avg.\( S_0 \)\( S_1 \)Avg.
DGCNN [55]31.5528.9430.2542.7137.2439.9823.9919.1021.5534.9328.1031.52
ProtoNet [44]33.9230.9532.4445.3442.0143.6828.4726.1327.3037.3634.9836.17
MPTI [73]39.2736.1437.7146.9043.5945.2529.9627.2628.6138.1434.3636.25
AttMPTI [73]42.5540.8341.6954.0050.3252.1635.2330.7232.9846.7440.8043.77
BFG [29]42.1540.5241.3451.2349.3950.3134.1231.9833.0546.2541.3843.82
2CBR [75]50.7347.6649.2052.3547.1449.7547.0046.3646.6845.0639.4742.27
PAP3D [20]57.0855.9456.5164.5559.6462.1055.2755.6055.4459.0253.1656.09
Seg-PN [77]63.1564.3263.7467.0869.0568.0761.8065.3463.5762.9468.2665.60
DAFNet68.7969.9569.3770.9170.6070.7666.1466.7066.4268.9771.9570.46
Improvement+5.64+5.63+5.63+3.83+1.55+2.69+4.34+1.36+2.85+6.03+3.69+4.86
+ +Table 6. Effect of different components of DAFNet. + +
GSA Layer +EdgeConv MHADAF-ConvOA (%)
Individual Components
✓XX88.6
X✓X89.4
XX✓91.7
Component Combinations
✓✓X91.3
✓X✓94.8
X✓✓93.5
✓✓✓96.7
+ +ble 5, DAFNet achieves consistent improvements across all experimental settings. For 2-way scenarios, our method obtains $69.37\%$ mIoU in 1-shot and $70.76\%$ mIoU in 5-shot settings, surpassing Seg-PN [77] by $5.63\%$ and $2.69\%$ respectively. In the more challenging 3-way settings, DAFNet achieves $66.42\%$ mIoU in 1-shot and $70.46\%$ mIoU in 5-shot settings, outperforming Seg-PN [77] by $2.85\%$ and $4.86\%$ . These substantial improvements across different settings demonstrate DAFNet's strong generalization ability and robustness in handling complex real-world scenes with limited training samples. + +# 4.4. Ablation Study + +All experiments are conducted on the ScanObjectNN [48] dataset unless explicitly stated otherwise. + +# 4.4.1. Ablation Study of Each Component + +We conduct ablation experiments to analyze the effectiveness of each component in DAFNet, as shown in Table 6. Among individual components, DAF-Conv shows the strongest performance with $91.7\%$ accuracy, notably higher than single GSA layer components (EdgeConv: $88.6\%$ , MHA: $89.4\%$ ). When combining components, the fusion of DAF-Conv with EdgeConv achieves $94.8\%$ , while DAF + +Table 7. Effect of different local geometric priors ${h}_{j}$ . + +
Local Geometric Prior hjOA(%)
[pj]93.6
[pj,pi]94.7
[pj,pipj-pi]96.2
[pj,pi, pj-pi, ||pj-pi||]96.7
+ +Conv with MHA reaches $93.5\%$ , both showing substantial improvements over single component performance. The full model integrating all components achieves the best performance of $96.7\%$ , demonstrating the strong complementary effects between DAF-Conv's local geometric modeling and the GSA layer's global context capturing. + +# 4.4.2. Ablation Study of Local Geometric Prior + +As shown in Table 7, we conduct ablation studies on the composition of local geometric prior $h_j$ . Starting with only center point coordinates $[p_j]$ yields $93.6\%$ accuracy. Incorporating neighbor point information $[p_j, p_i]$ brings a $1.1\%$ improvement. Adding relative position $[p_j, p_i, p_j - p_i]$ further boosts performance by $1.5\%$ . The complete geometric prior with Euclidean distance $||p_j - p_i||$ achieves the best result of $96.7\%$ , validating the importance of comprehensive geometric information in local structure modeling. + +# 4.4.3. Analysis of Model Hyperparameters + +We conduct comprehensive ablation studies on three key hyperparameters of DAFNet, as shown in Fig. 4. + +Frequency Component Number H: Fig. 4(a) shows that model performance improves as $H$ increases from 1 to 6, achieving peak accuracy at $H = 6$ $(96.7\%)$ . Further increasing $H$ leads to performance degradation, dropping to $95.1\%$ at $H = 10$ . This indicates that while multiple frequency components enhance feature diversity, excessive components may introduce redundant information. + +Distance Threshold $\mathbf{d}_{\mathrm{threshold}}$ : As shown in Fig. 4(b), we examine the impact of $d_{threshold}$ which divides the local neighborhood into near-field and far-field regions. The optimal performance (96.7%) is achieved at $d_{threshold} = 0.6$ . Lower thresholds limit near-field information, while higher values may include irrelevant points in the near-field region, both resulting in suboptimal performance. + +Neighborhood Size K: Fig. 4(c) demonstrates the effect of neighborhood size $K$ in EdgeConv. Performance reaches its peak $(96.7\%)$ at $K = 20$ . Smaller $K$ values provide insufficient local context, while larger values may incorporate irrelevant points, both leading to decreased accuracy. + +# 4.5. Model Complexity + +We compare the model complexity of DAFNet with existing methods on ScanObjectNN [48], as shown in Ta + +![](images/b40574680401fdf325107935a697c4e00cbc4851a869ef42ef3f47fbc0da4ded.jpg) +Figure 4. Ablation studies on key hyperparameters. (a) Impact of frequency component number $H$ in DAF-Conv. (b) Effect of distance threshold $d_{threshold}$ for near-field and far-field separation. (c) Influence of neighborhood size $K$ in EdgeConv layer. + +![](images/675e9445c731bc33922d1e5691f3052f95db5cc8b578c3a8a2bad7cea864947c.jpg) + +![](images/5d70e76dbb6bdf3905d656993436fbb2175c7bdddde2c971181cd890cd614d50.jpg) + +Table 8. Model complexity comparison on ScanObjectNN. + +
MethodsOA(%)Params.(M)FLOPs(G)
PointNet [35]68.03.50.5
PointNet++[36]77.91.51.7
DGCNN[55]78.11.82.4
PointCNN [24]78.50.6-
MVTN [18]82.811.243.7
PointMLP [28]85.412.631.4
PointNeXt [39]87.41.43.6
DAFNet (Ours)96.72.471.1
+ +ble 8. While achieving state-of-the-art performance (96.7% OA), DAFNet maintains moderate computational requirements with only 2.47M parameters and 1.1G FLOPs. This is significantly more efficient than models like MVTN [18] (11.2M parameters, 43.7G FLOPs) and PointMLP [28] (12.6M parameters, 31.4G FLOPs), demonstrating that our acoustic field-inspired design achieves superior accuracy without sacrificing computational efficiency. + +# 5. Conclusion + +In this paper, we presented DAFNet, a physics-inspired architecture for point cloud understanding that addresses local structure representation challenges. Inspired by acoustic field principles, we introduced AF-Conv modeling local feature aggregation through acoustic propagation, and extended it to DAF-Conv which dynamically adapts to diverse geometric features through multi-frequency simulation. By integrating these with GSA, our approach captures both fine-grained local structures and global context. Experiments demonstrate DAFNet achieves state-of-the-art performance across multiple tasks. Future work will extend acoustic field modeling to other domains, exploring applications in additional 3D tasks. + +# Acknowledgment + +This project is sponsored by Shanghai Pujiang Programme 24PJD030. The computational work for this article is partially performed on resources of the National Supercomputing Centre, Singapore (https://www.nscc.sg). + +# References + +[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, pages 1534-1543, 2016. 6 +[2] Binjie Chen, Yunzhou Xia, Yu Zang, Cheng Wang, and Jonathan Li. Decoupled local aggregation for point cloud learning. arXiv preprint arXiv:2308.16532, 2023. 6 +[3] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 6, 7 +[4] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. NeurIPS, 36, 2024. 3 +[5] Ziyang Chen, Israel D Gebru, Christian Richardt, Anurag Kumar, William Laney, Andrew Owens, and Alexander Richard. Real acoustic fields: An audio-visual room acoustics dataset and benchmark. In CVPR, pages 21886–21896, 2024. 2 +[6] Pranav Singh Chib and Pravendra Singh. Recent advancements in end-to-end autonomous driving using deep learning: A survey. IEEE Transactions on Intelligent Vehicles, 2023. 1 +[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 7 +[8] Jeevan S Devagiri, Sidike Paheding, Quamar Niyaz, Xiaoli Yang, and Samantha Smith. Augmented reality and artificial intelligence in industry: Trends, tools, and future challenges. Expert Systems with Applications, 207:118002, 2022. 1 +[9] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 2 +[10] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. Mose: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 2 +[11] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. 3 +[12] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, pages 8458-8468, 2022. 2 + +[13] Xiang Fang, Wanlong Fang, Changshuo Wang, Daizong Liu, Keke Tang, Jianfeng Dong, Pan Zhou, and Beibei Li. Multipair temporal sentence grounding via multi-thread knowledge transfer network. arXiv preprint arXiv:2412.15678, 2024. 2 +[14] Ruchi Goel and Pooja Gupta. Robotics and industry 4.0. A Roadmap to Industry 4.0: Smart Production, Sharp Business and Sustainable Development, pages 157-169, 2020. 1 +[15] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In ICML, pages 3809-3820. PMLR, 2021. 6 +[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2 +[17] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2 +[18] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In ICCV, pages 1-11, 2021. 6, 8 +[19] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In ACM MM, pages 4995-5004, 2024. 1, 2, 6 +[20] Shuting He, Xudong Jiang, Wei Jiang, and Henghui Ding. Prototype adaption and projection for few-and zero-shot 3d point cloud semantic segmentation. TIP, 32:3199-3211, 2023. 7 +[21] Limin Jiang, Changshuo Wang, Xin Ning, and Zaiyang Yu. Lttpoint: a mlp-based point cloud classification method with local topology transformation module. In 2023 7th Asian Conference on Artificial Intelligence Technology (ACAIT), pages 783-789. IEEE, 2023. 1 +[22] Lawrence E Kinsler, Austin R Frey, Alan B Coppens, and James V Sanders. Fundamentals of acoustics. John wiley & sons, 2000. 4 +[23] Artem Komarichev, Zichun Zhong, and Jing Hua. A-cnn: Annually convolutional neural networks on point clouds. In CVPR, pages 7421-7430, 2019. 2 +[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. NeurIPS, 31, 2018. 8 +[25] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 1, 3, 6, 7 +[26] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual state space model. In NeurIPS, 2024. 2 +[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2 + +[28] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 6, 8 +[29] Yongqiang Mao, Zonghao Guo, LU Xiaonan, Zhiqiang Yuan, and Haowen Guo. Bidirectional feature globalization for few-shot semantic segmentation of 3d point cloud scenes. In 2022 International Conference on 3D Vision (3DV), pages 505-514. IEEE, 2022. 7 +[30] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928. IEEE, 2015. 2 +[31] Philip McCord Morse and K Uno Ingard. Theoretical acoustics. Princeton university press, 1986. 4 +[32] AAM Muzahid, Wanggen Wan, Ferdous Sohel, Lianyao Wu, and Li Hou. Curvenet: Curvature-based multitask learning deep networks for 3d object recognition. IEEE/CAA Journal of Automatica Sinica, 8(6):1177-1187, 2020. 6, 7 +[33] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, pages 604-621. Springer, 2022. 1, 3, 6 +[34] Jinyoung Park, Sanghyeok Lee, Sihyeon Kim, Yunyang Xiong, and Hyunwoo J Kim. Self-positioning point-based transformer for point cloud understanding. In CVPR, pages 21814-21823, 2023. 7 +[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 2, 6, 7, 8 +[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NIPS, 30, 2017. 2, 6, 7, 8 +[37] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In ICML, pages 28223-28243. PMLR, 2023. 1, 3 +[38] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. *Shapellm: Universal 3d object understanding for embodied interaction*. In *ECCV*, pages 214–238. Springer, 2025. 1, 6 +[39] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. NIPS, 35:23192-23204, 2022. 6, 7, 8 +[40] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In CVPR, pages 18942-18952, 2022. 6 +[41] Jingjing Ren, Ashwin Rao, Martina Lindorfer, Arnaud Legout, and David Choffnes. Recon: Revealing and controlling pii leaks in mobile network traffic. In Proceedings of the 14th Annual International Conference on Mobile Systems, Applications, and Services, pages 361-374, 2016. 6, 7 + +[42] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, pages 3577-3586, 2017. 2 +[43] Mickael Sereno, Xiyao Wang, Lonni Besançon, Michael J Mcguffin, and Tobias Isenberg. Collaborative work in augmented reality: A survey. TVCG, 28(6):2530-2549, 2020. 1 +[44] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. NIPS, 30, 2017. 7 +[45] Mohsen Soori, Behrooz Arezoo, and Roza Dastres. Artificial intelligence, machine learning and deep learning in advanced robotics, a review. Cognitive Robotics, 3:54-70, 2023. 1 +[46] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. In ECCV, pages 426-442. Springer, 2022. 2 +[47] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6411–6420, 2019. 7 +[48] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, pages 1588–1597, 2019. 5, 6, 7, 8 +[49] Changshuo Wang, Xin Ning, Linjun Sun, Liping Zhang, Weijun Li, and Xiao Bai. Learning discriminative features by covering local geometric space for point cloud analysis. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-15, 2022. 2 +[50] C Wang, H Wang, X Ning, S Tian, and W Li. 3d point cloud classification method based on dynamic coverage of local area. J. Softw, 34(4):1962-1976, 2022. 1 +[51] Changshuo Wang, Xin Ning, Weijun Li, Xiao Bai, and Xingyu Gao. 3d person re-identification based on global semantic guidance and local feature aggregation. TCSVT, 34 (6):4698-4712, 2023. 2 +[52] Changshuo Wang, Meiqing Wu, Siew-Kei Lam, Xin Ning, Shangshu Yu, Ruiping Wang, Weijun Li, and Thambipillai Srikanthan. Gpsformer: A global perception and local structure fitting-based transformer for point cloud understanding. arXiv preprint arXiv:2407.13519, 2024. 2, 5, 6, 7 +[53] Changshuo Wang, Rongsheng Cao, and Ruiping Wang. Learning discriminative topological structure information representation for 2d shape and social network classification via persistent homology. Knowledge-Based Systems, page 113125, 2025. 2 +[54] Ruiping Wang, Siew-Kei Lam, Meiqing Wu, Zhijian Hu, Changshuo Wang, and Jing Wang. Destination intention estimation-based convolutional encoder-decoder for pedestrian trajectory multimodality forecast. Measurement, 239: 115470, 2025. 1 +[55] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 4, 5, 6, 7, 8 + +[56] Xin Wei, Ruixuan Yu, and Jian Sun. View-gcn: View-based graph convolutional network for 3d shape analysis. In CVPR, pages 1850-1859, 2020. 2 +[57] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 2 +[58] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 2 +[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, pages 1912-1920, 2015. 6 +[60] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, pages 3173-3182, 2021. 4, 6 +[61] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In CVPR, pages 1179-1189, 2023. 3, 6 +[62] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. NIPS, 32, 2019. 4, 5 +[63] Ze Yang and Liwei Wang. Learning relationships for multiview 3d object recognition. In ICCV, pages 7505-7514, 2019. 2 +[64] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6 +[65] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, pages 19313-19322, 2022. 6 +[66] Huang Zhang, Changshuo Wang, Long Yu, Shengwei Tian, Xin Ning, and Joel Rodrigues. Pointgt: A method for pointcloud classification and segmentation based on local geometric transformation. IEEE Transactions on Multimedia, 2024. 2 +[67] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. NeurIPS, 35:27061-27074, 2022. 3 +[68] Renrui Zhang, Liuhui Wang, Ziyu Guo, Yali Wang, Peng Gao, Hongsheng Li, and Jianbo Shi. Parameter is not all you need: Starting from non-parametric networks for 3d point cloud analysis. arXiv preprint arXiv:2303.08134, 2023. 6 +[69] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024.2 +[70] Xiangdong Zhang, Shaofeng Zhang, and Junchi Yan. Pcmae: Learning to predict centers for point masked autoen + +coders. Advances in Neural Information Processing Systems, 37:80303-80327, 2024. 3, 6, 7 +[71] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 2, 6, 7 +[72] Jingyuan Zhao, Wenyi Zhao, Bo Deng, Zhenghong Wang, Feng Zhang, Wenxiang Zheng, Wanke Cao, Jinrui Nan, Yubo Lian, and Andrew F Burke. Autonomous driving system: A comprehensive survey. Expert Systems with Applications, page 122836, 2023. 1 +[73] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In CVPR, pages 8873-8882, 2021. 6, 7 +[74] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In CVPR, pages 22935-22945, 2024. 3 +[75] Guanyu Zhu, Yong Zhou, Rui Yao, and Hancheng Zhu. Cross-class bias rectification for point cloud few-shot segmentation. TMM, 25:9175-9188, 2023. 7 +[76] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2 +[77] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ji-aming Liu, Han Xiao, Chaoyou Fu, Hao Dong, and Peng Gao. No time to train: Empowering non-parametric networks for few-shot 3d scene segmentation. In CVPR, pages 3838–3847, 2024. 5, 7 \ No newline at end of file diff --git a/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/images.zip b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f41cfb4226562095dcb58a2c3dad071e847c46ae --- /dev/null +++ b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:758ec814f39779ad8b1e9ec86441eff65a6305aac62f0858ff3f7fd743e1242f +size 629105 diff --git a/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/layout.json b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f4ebaf973ea80a13522798bc15622599f17ce2ab --- /dev/null +++ b/2025/Point Clouds Meets Physics_ Dynamic Acoustic Field Fitting Network for Point Cloud Understanding/layout.json @@ -0,0 +1,11168 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "type": "text", + "content": "Point Clouds Meets Physics: Dynamic Acoustic Field Fitting Network for Point Cloud Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 162, + 208, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 162, + 208, + 175 + ], + "spans": [ + { + "bbox": [ + 121, + 162, + 208, + 175 + ], + "type": "text", + "content": "Changshuo Wang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 124, + 176, + 206, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 176, + 206, + 190 + ], + "spans": [ + { + "bbox": [ + 124, + 176, + 206, + 190 + ], + "type": "text", + "content": "NTU, Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 192, + 231, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 192, + 231, + 203 + ], + "spans": [ + { + "bbox": [ + 99, + 192, + 231, + 203 + ], + "type": "text", + "content": "wangchangshuol@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 289, + 162, + 347, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 289, + 162, + 347, + 175 + ], + "spans": [ + { + "bbox": [ + 289, + 162, + 347, + 175 + ], + "type": "text", + "content": "Shuting He" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 286, + 176, + 351, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 176, + 351, + 189 + ], + "spans": [ + { + "bbox": [ + 286, + 176, + 351, + 189 + ], + "type": "text", + "content": "SUFE, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 258, + 192, + 379, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 192, + 379, + 203 + ], + "spans": [ + { + "bbox": [ + 258, + 192, + 379, + 203 + ], + "type": "text", + "content": "shuting.he@sufe.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 430, + 162, + 492, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 162, + 492, + 176 + ], + "spans": [ + { + "bbox": [ + 430, + 162, + 492, + 176 + ], + "type": "text", + "content": "Xiang Fang*" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 419, + 176, + 499, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 176, + 499, + 190 + ], + "spans": [ + { + "bbox": [ + 419, + 176, + 499, + 190 + ], + "type": "text", + "content": "NTU, Singapore" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 407, + 192, + 511, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 192, + 511, + 203 + ], + "spans": [ + { + "bbox": [ + 407, + 192, + 511, + 203 + ], + "type": "text", + "content": "xfang9508@gmail.com" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 98, + 209, + 154, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 209, + 154, + 221 + ], + "spans": [ + { + "bbox": [ + 98, + 209, + 154, + 221 + ], + "type": "text", + "content": "Jiawei Han" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 99, + 224, + 154, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 224, + 154, + 236 + ], + "spans": [ + { + "bbox": [ + 99, + 224, + 154, + 236 + ], + "type": "text", + "content": "BIT, China" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 71, + 239, + 181, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 239, + 181, + 251 + ], + "spans": [ + { + "bbox": [ + 71, + 239, + 181, + 251 + ], + "type": "text", + "content": "hanjiawei@bit.edu.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 238, + 209, + 317, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 209, + 317, + 223 + ], + "spans": [ + { + "bbox": [ + 238, + 209, + 317, + 223 + ], + "type": "text", + "content": "Zhonghang Liu" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 237, + 224, + 319, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 224, + 319, + 237 + ], + "spans": [ + { + "bbox": [ + 237, + 224, + 319, + 237 + ], + "type": "text", + "content": "SMU, Singapore" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 205, + 239, + 351, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 239, + 351, + 251 + ], + "spans": [ + { + "bbox": [ + 205, + 239, + 351, + 251 + ], + "type": "text", + "content": "zhliu.2020@phdcs.smu.edu.sg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 367, + 209, + 481, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 209, + 481, + 224 + ], + "spans": [ + { + "bbox": [ + 367, + 209, + 481, + 224 + ], + "type": "text", + "content": "Xin Ning" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 375, + 224, + 539, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 224, + 539, + 236 + ], + "spans": [ + { + "bbox": [ + 375, + 224, + 539, + 236 + ], + "type": "text", + "content": "Institute of Semiconductors, CAS" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 408, + 240, + 506, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 240, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 408, + 240, + 506, + 250 + ], + "type": "text", + "content": "ningxin@semi.ac.cn" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 134, + 257, + 298, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 257, + 298, + 285 + ], + "spans": [ + { + "bbox": [ + 134, + 257, + 298, + 285 + ], + "type": "text", + "content": "Weijun Li \nInstitute of Semiconductors, CAS" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 174, + 288, + 257, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 288, + 257, + 298 + ], + "spans": [ + { + "bbox": [ + 174, + 288, + 257, + 298 + ], + "type": "text", + "content": "wjli@semi.ac.cn" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 331, + 257, + 476, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 257, + 476, + 285 + ], + "spans": [ + { + "bbox": [ + 331, + 257, + 476, + 285 + ], + "type": "text", + "content": "Prayag Tiwari \nHalmstad University, Sweden" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 342, + 289, + 462, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 289, + 462, + 299 + ], + "spans": [ + { + "bbox": [ + 342, + 289, + 462, + 299 + ], + "type": "text", + "content": "prayag.tiwari@ieee.org" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 152, + 327, + 200, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 327, + 200, + 338 + ], + "spans": [ + { + "bbox": [ + 152, + 327, + 200, + 338 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 53, + 354, + 296, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 296, + 678 + ], + "type": "text", + "content": "While existing pre-training-based methods have enhanced point cloud model performance, they have not fundamentally resolved the challenge of local structure representation in point clouds. The limited representational capacity of pure point cloud models continues to constrain the potential of cross-modal fusion methods and performance across various tasks. To address this challenge, we propose a Dynamic Acoustic Field Fitting Network (DAF-Net), inspired by physical acoustic principles. Specifically, we represent local point clouds as acoustic fields and introduce a novel Acoustic Field Convolution (AF-Conv), which treats local aggregation as an acoustic energy field modeling problem and captures fine-grained local shape awareness by dividing the local area into near field and far field. Furthermore, drawing inspiration from multifrequency wave phenomena and dynamic convolution, we develop the Dynamic Acoustic Field Convolution (DAF-Conv) based on AF-Conv. DAF-Conv dynamically generates multiple weights based on local geometric priors, effectively enhancing adaptability to diverse geometric features. Additionally, we design a Global Shape-Aware (GSA) layer incorporating EdgeConv and multi-head attention mechanisms, which combines with DAF-Conv to form the DAF Block. These blocks are then stacked to create a hierarchical DAFNet architecture. Extensive experiments demonstrate that DAFNet significantly outperforms existing methods across multiple tasks." + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 316, + 326, + 548, + 375 + ], + "blocks": [ + { + "bbox": [ + 316, + 326, + 548, + 375 + ], + "lines": [ + { + "bbox": [ + 316, + 326, + 548, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 548, + 375 + ], + "type": "image", + "image_path": "5ec06a8d1a7adeb879f68d13b5ef6ea34882045597c6c6a5cc6446db0b3077b9.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 376, + 428, + 383 + ], + "lines": [ + { + "bbox": [ + 386, + 376, + 428, + 383 + ], + "spans": [ + { + "bbox": [ + 386, + 376, + 428, + 383 + ], + "type": "text", + "content": "Local Neighborhood" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 375, + 477, + 387 + ], + "lines": [ + { + "bbox": [ + 431, + 375, + 477, + 387 + ], + "spans": [ + { + "bbox": [ + 431, + 375, + 477, + 387 + ], + "type": "text", + "content": "Multi-Acoustic Energy Field Mapping" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 492, + 374, + 548, + 381 + ], + "lines": [ + { + "bbox": [ + 492, + 374, + 548, + 381 + ], + "spans": [ + { + "bbox": [ + 492, + 374, + 548, + 381 + ], + "type": "text", + "content": "Shape-aware Representation" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 399, + 555, + 465 + ], + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 465 + ], + "type": "text", + "content": "Figure 1. Illustration of local feature extraction using DAF-Conv. Given a local neighborhood of a point cloud (left), our method first models it as an acoustic field with near-field and far-field zones. Then, multiple acoustic energy field mappings are generated to capture various geometric patterns. Finally, DAF-Conv processes these mappings to produce shape-aware representations." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 487, + 394, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 394, + 500 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 394, + 500 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 508, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 508, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 312, + 508, + 555, + 605 + ], + "type": "text", + "content": "Point cloud data, due to its direct representation of object surface information in three-dimensional space, has been widely applied in various scenarios including autonomous driving [6, 54, 72], robotic navigation [14, 45], and augmented reality [8, 43]. However, the sparsity, irregularity, and complexity of local structures in point cloud data make it challenging for existing models to fully capture geometric features, thus limiting their applications in 3D tasks." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 556, + 714 + ], + "type": "text", + "content": "Recently, several methods [33, 37, 38] have attempted to indirectly improve point cloud model performance through pre-training strategies or cross-modal contrastive learning. Although these approaches have enhanced point cloud processing tasks to some extent, they have not directly addressed the problem of local structure representation. Some researchers [19, 21, 25, 50] have also attempted to apply new architectures to point cloud understanding tasks. However, the local representational capacity of point cloud mod" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 168, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 168, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 168, + 712 + ], + "type": "text", + "content": "* denotes corresponding author." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22182" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "els remains a bottleneck in improving overall model performance. Therefore, how to efficiently capture and represent the local structures of point clouds has become a crucial problem that urgently needs to be solved." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 128, + 295, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 295, + 294 + ], + "type": "text", + "content": "Physics, particularly the principles of acoustic field propagation [5], provides a new perspective for understanding local geometric structures. In acoustics, wave propagation is influenced by environmental geometric characteristics, and through constructing acoustic field models, we can effectively simulate wave behavior at different frequencies to obtain information in complex environments. Inspired by this, we propose a novel point cloud understanding method called Dynamic Acoustic Field Fitting Network (DAFNet). This method borrows ideas from physical acoustic field propagation, representing local point clouds as acoustic fields and dynamically adjusting weights in different regions to capture local structural features more precisely." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 302, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 295, + 529 + ], + "type": "text", + "content": "Specifically, we propose a novel Acoustic Field Convolution (AF-Conv) that models local aggregation as an acoustic energy propagation problem. To more finely model the contributions of different neighboring points to the center point, AF-Conv divides the local region into near-field and far-field zones for separate geometric modeling. Furthermore, as shown in Fig. 1, based on AF-Conv, we propose Dynamic Acoustic Field Convolution (DAF-Conv), which dynamically generates multiple weights based on geometric priors while simulating geometric feature variations under different \"frequencies\", thereby enhancing the model's adaptability to local geometric diversity. Finally, we combine a Global Shape-Aware (GSA) layer comprising EdgeConv and Multi-Head Attention (MHA) with DAF-Conv to build hierarchical DAF Blocks, forming the complete DAFNet. The experimental results show that DAFNet achieves superior performance over existing methods in multiple tasks, demonstrating its effectiveness in capturing complex geometric structures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 536, + 287, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 536, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 67, + 536, + 287, + 548 + ], + "type": "text", + "content": "The main contributions of this work are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 556, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 556, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 295, + 590 + ], + "type": "text", + "content": "- We propose a physics-inspired end-to-end DAFNet that can learn rich contextual information and precise shape representations in irregular point clouds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 592, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 627 + ], + "type": "text", + "content": "- We introduce a novel Acoustic Field Convolution (AFConv) that achieves fine-grained local shape awareness by modeling local point cloud regions as acoustic fields." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 628, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 295, + 674 + ], + "type": "text", + "content": "- We design Dynamic Acoustic Field Convolution (DAFConv) that adapts to geometric feature diversity by dynamically generating multiple frequency-aware weights from local geometry." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 676, + 295, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 676, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 676, + 295, + 711 + ], + "type": "text", + "content": "- The proposed DAFNet achieves state-of-the-art performance in multiple tasks, reaching " + }, + { + "bbox": [ + 56, + 676, + 295, + 711 + ], + "type": "inline_equation", + "content": "96.7\\%" + }, + { + "bbox": [ + 56, + 676, + 295, + 711 + ], + "type": "text", + "content": " accuracy on the challenging ScanObjectNN with only 2.47M parameters." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 91, + 434, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 434, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 434, + 102 + ], + "type": "text", + "content": "2.1. CNN-based Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 312 + ], + "type": "text", + "content": "Early works [30, 42, 56, 63] in point cloud processing primarily involved projecting them onto 2D images for 2D CNN processing or converting them into regular 3D grid structures for 3D convolution operations. However, these transformation processes inevitably led to spatial information loss and increased computational costs. To address this issue, researchers began applying deep learning techniques directly in 3D space. PointNet [35] pioneered this direction. However, this method overlooked local structural information. Subsequently, PointNet++ [36] enhanced the local perception capability of point cloud models by grouping point clouds into different local neighborhoods. Based on PointNet++ [36], existing works [23, 49, 51, 52, 55, 66] have developed methods based on point-wise perceptrons, convolution operations, and attention mechanisms. Nevertheless, these methods still struggle to effectively model local structural shape information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 321, + 470, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 321, + 470, + 333 + ], + "spans": [ + { + "bbox": [ + 314, + 321, + 470, + 333 + ], + "type": "text", + "content": "2.2. Transformer-based Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 340, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 532 + ], + "type": "text", + "content": "Transformers have demonstrated remarkable success in various tasks [9, 10, 13, 53] due to their inherent ability to capture global features. Leveraging this strength, researchers have progressively developed a range of Transformer-based point cloud models. PCT [17] enhances input embeddings through furthest point sampling and nearest neighbor search while using self-attention mechanisms to capture global features. Point Transformer [71] employs vector attention to extract aggregated local features. Subsequently, updated versions [57, 58] of Point Transformer explored different architectures to improve performance and efficiency. Inspired by Swin Transformer [27], works such as SST [12] and SWFormer [46] explored window attention mechanisms to improve efficiency. Although these methods excel at capturing global features, their performance hasn't shown significant advancement over CNN-based approaches." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 540, + 497, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 540, + 497, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 540, + 497, + 552 + ], + "type": "text", + "content": "2.3. State Space Model-based Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "State Space Models (SSMs) [16] are mathematical models for describing dynamic systems. Recently, SSMs have emerged as a reliable alternative architecture to CNNs and Transformers. In computer vision, Vision Mamba [76] successfully compresses visual representations through bidirectional state space models, while VMamba [26] innovatively introduces cross-scanning modules, enabling global receptive fields in 2D images through 1D selective scanning. Researchers have explored SSM applications in point cloud processing. PCM [69] first introduced the Mamba architecture to point cloud processing through octree serialization strategies. Mamba3D [19] enhanced local geometric feature extraction through LNP blocks and designed bidi" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22183" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 541, + 329 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 541, + 329 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 541, + 329 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 541, + 329 + ], + "type": "image", + "image_path": "a56044a4d3b301eff35015c04723e07ca735b3db485842c06c6f6f8a22e0a8da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 338, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 54, + 338, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 555, + 396 + ], + "type": "text", + "content": "Figure 2. Architecture of DAFNet. (a) For shape classification, the network consists of three sequential DAF Blocks followed by max pooling and classifier layers. Each DAF Block contains a GSA layer and DAF-Conv for feature extraction. (b) For part segmentation and few-shot semantic segmentation tasks, the network adopts an encoder-decoder structure with skip connections, where the encoder shares the same architecture as the classification network, and the decoder uses feature propagation to restore point cloud resolution. The few-shot version additionally incorporates support and query mechanisms for processing different point cloud inputs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 415, + 298, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 298, + 452 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 298, + 452 + ], + "type": "text", + "content": "rectional SSM structures. PointMamba [25], another significant work, customized SSM architecture for point cloud analysis tasks, substantially reducing computational costs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 456, + 211, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 211, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 211, + 470 + ], + "type": "text", + "content": "2.4. Pre-training-based Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 474, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 715 + ], + "type": "text", + "content": "Point cloud pre-training methods primarily comprise generative approaches [4, 33, 67] and cross-modal contrastive learning methods. In generative approaches, Point-MAE [33] extends the MAE concept by pre-training through random masking and reconstruction of point cloud segments PCP-MAE [70] learns shape priors by predicting missing regional point cloud structures. PointDif [74] utilizes diffusion models for pre-training, learning hierarchical geometric priors through point cloud recovery at different noise levels. In cross-modal contrastive learning, ReCon [37] combines generative modeling and contrastive learning paradigms, addressing representation overfitting in contrastive learning and data padding issues in generative methods. ULIP [61] achieves cross-modal knowledge transfer through a language-image-point cloud pre-training framework, while ACT [11] employs cross-modal autoencoders as teacher models to acquire knowledge from other modalities. Although these pre-training methods have advanced point cloud processing performance, they haven't fundamentally resolved the point cloud representation challenge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 413, + 375, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 413, + 375, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 413, + 375, + 426 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 434, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 495 + ], + "type": "text", + "content": "This section first reviews point convolution and dynamic convolution, followed by introducing acoustic field convolution. Then, we present our proposed dynamic acoustic field convolution. Finally, we describe DAFNet (as shown in Fig. 2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 499, + 395, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 395, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 395, + 512 + ], + "type": "text", + "content": "3.1. Background" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 517, + 421, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 421, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 421, + 528 + ], + "type": "text", + "content": "3.1.1. Point Convolution" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "text", + "content": "Local feature aggregation in point clouds typically follows a \"sampling-grouping-aggregation\" procedure. The aggregation operation is crucial as it directly affects local feature representation by modeling local point clouds. We denote center points as " + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "inline_equation", + "content": "\\{(p_i, f_i)\\}_{i=1}^M" + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "text", + "content": " represents the number of points, " + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "inline_equation", + "content": "p_i \\in \\mathbb{R}^{1 \\times 3}" + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "inline_equation", + "content": "f_i \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "text", + "content": " represent the coordinates and features of the " + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "text", + "content": "-th point respectively, and " + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 532, + 556, + 663 + ], + "type": "text", + "content": " denotes the number of feature channels. The most commonly used method for local feature aggregation based on coordinate space is to construct point convolution, which can be formalized as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 670, + 555, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 670, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 339, + 670, + 555, + 685 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {\\prime} = \\mathcal {A} (\\{\\mathcal {M} (p _ {i}, p _ {j}) \\cdot \\mathcal {T} (f _ {i}, f _ {j}) | p _ {j} \\in \\mathcal {N} (p _ {i}) \\}), \\quad (1)", + "image_path": "65afa2e83a07a86bd38068319073418aca2d1a546b5d581f3d1498c9f298a6b2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " represents the aggregation function, typically max-pooling. " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " are mapping functions, usually imple" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22184" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "mented through Multi-Layer Perceptrons (MLPs). " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(p_i)" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " represents the local neighborhood of " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " denotes the neighboring points of " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 296, + 168 + ], + "type": "text", + "content": "An alternative aggregation approach operates in feature space, commonly known as Dynamic Graph Convolution (DGC) [55]. It aggregates local features in feature space, meaning that points close in feature space might be distant in coordinate space. DGC can be formalized as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 177, + 295, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 177, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 295, + 192 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {\\prime} = \\mathcal {A} (\\{\\mathcal {T} (f _ {i}, f _ {j}) | f _ {j} \\in \\mathcal {N} (f _ {i}) \\}). \\tag {2}", + "image_path": "3bff3a839e32042b6dcb92d46320c4ce99cf2c90a918898eeb10e3304eb27573.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 199, + 178, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 178, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 178, + 211 + ], + "type": "text", + "content": "3.1.2. Dynamic Convolution" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 214, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 296, + 286 + ], + "type": "text", + "content": "Dynamic convolution [62] aims to enhance network modeling capability by dynamically generating multiple convolution kernels based on input data, adapting to different data structures. Taking PACConv's [60] dynamic convolution as an example, dynamic convolution in point clouds can be expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 296, + 295, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 296, + 295, + 311 + ], + "spans": [ + { + "bbox": [ + 102, + 296, + 295, + 311 + ], + "type": "interline_equation", + "content": "g _ {i} = \\mathcal {A} \\left(\\left\\{w \\left(p _ {j}\\right) \\cdot f _ {j} \\mid p _ {j} \\in \\mathcal {N} \\left(p _ {i}\\right) \\right\\}\\right), \\tag {3}", + "image_path": "b96d1933747aa794ede61b4fdc467f7a2ffb759e5c2ba343dc6f222f677bc5d6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "inline_equation", + "content": "w(p_j)" + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "text", + "content": " denotes the weight corresponding to feature " + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "inline_equation", + "content": "f_j" + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "text", + "content": " of point " + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "inline_equation", + "content": "w(p_j)" + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "text", + "content": " typically consists of multiple weights " + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "inline_equation", + "content": "w_h(p_i)" + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "text", + "content": " and their corresponding attention coefficients " + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\alpha_h(p_j)" + }, + { + "bbox": [ + 55, + 317, + 296, + 367 + ], + "type": "text", + "content": " generated in a data-driven manner:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 112, + 375, + 295, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 375, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 112, + 375, + 295, + 409 + ], + "type": "interline_equation", + "content": "w \\left(p _ {j}\\right) = \\sum_ {h = 1} ^ {H} \\alpha_ {h} \\left(p _ {j}\\right) \\cdot w _ {h} \\left(p _ {j}\\right), \\tag {4}", + "image_path": "6cc441070eb5b99f984c20c609166989568c67f5b334744c1830e09da1aafc2c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 417, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 417, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\sum_{h=1}^{H} \\alpha_h(p_j) = 1" + }, + { + "bbox": [ + 55, + 417, + 296, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 417, + 296, + 443 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 417, + 296, + 443 + ], + "type": "text", + "content": " represents the number of " + }, + { + "bbox": [ + 55, + 417, + 296, + 443 + ], + "type": "inline_equation", + "content": "w_h(p_j)" + }, + { + "bbox": [ + 55, + 417, + 296, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 444, + 237, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 444, + 237, + 455 + ], + "spans": [ + { + "bbox": [ + 67, + 444, + 237, + 455 + ], + "type": "text", + "content": "Therefore, formula 3 can be expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 464, + 295, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 464, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 68, + 464, + 295, + 497 + ], + "type": "interline_equation", + "content": "g _ {i} = \\mathcal {A} \\left(\\left\\{\\sum_ {h = 1} ^ {H} \\alpha_ {h} \\left(p _ {j}\\right) \\cdot w _ {h} \\left(p _ {j}\\right) \\cdot f _ {j} \\mid p _ {j} \\in \\mathcal {N} \\left(p _ {i}\\right) \\right\\}\\right). \\tag {5}", + "image_path": "3d7b5262bbb19dc78c6210ecf67f51b56c00aeb0e60d8823d66a5cbaa50460a8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 513, + 205, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 205, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 205, + 525 + ], + "type": "text", + "content": "3.2. Acoustic Field Convolution" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 529, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 567 + ], + "type": "text", + "content": "In physical acoustics [22, 31], the sound pressure field " + }, + { + "bbox": [ + 55, + 529, + 296, + 567 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 55, + 529, + 296, + 567 + ], + "type": "text", + "content": " generated by a sound source at spatial point " + }, + { + "bbox": [ + 55, + 529, + 296, + 567 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 529, + 296, + 567 + ], + "type": "text", + "content": " can be represented by Green's function:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 132, + 574, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 574, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 132, + 574, + 295, + 602 + ], + "type": "interline_equation", + "content": "p (x) = \\frac {e ^ {- j k \\| x - y \\|}}{4 \\pi \\| x - y \\|}, \\tag {6}", + "image_path": "f988db4af8053bcdd606683eb50bbeb65e0d082f47eadda06098dd5a0f41e2b2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "text", + "content": " is the source position, " + }, + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "text", + "content": " is the wave number, and " + }, + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\| x - y\\|" + }, + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "text", + "content": " represents the distance between observation point " + }, + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "text", + "content": " and source " + }, + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 605, + 296, + 677 + ], + "type": "text", + "content": ". This formula reveals the fundamental physical laws of sound wave propagation: sound pressure intensity decays exponentially with distance while being affected by spherical spreading attenuation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Notably, based on the distance between the source and observation point, the acoustic field can be divided into near-field and far-field regions. In the near-field region, the" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 201 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 201 + ], + "type": "image", + "image_path": "35d3576f695b2400d227bcea514c6e0feece7fa5a39904e756b17e6bd478a5d1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 209, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 266 + ], + "type": "text", + "content": "Figure 3. Illustration of AF-Conv and DAF-Conv. (a) AF-Conv divides local neighborhood into near-field and far-field regions and calculates energy weights based on geometric priors. (b) DAF-Conv extends AF-Conv with multiple parallel acoustic field convolutions to achieve dynamic feature aggregation." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 285, + 553, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 553, + 332 + ], + "type": "text", + "content": "relationship between sound pressure and distance is more complex, exhibiting strong local correlations; in the farfield region, the sound pressure demonstrates simple spherical wave attenuation characteristics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 334, + 554, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 554, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 554, + 428 + ], + "type": "text", + "content": "Inspired by this theory, we view local feature aggregation as acoustic field propagation, where the center point " + }, + { + "bbox": [ + 313, + 334, + 554, + 428 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 334, + 554, + 428 + ], + "type": "text", + "content": " can be viewed as the sound source and neighboring points " + }, + { + "bbox": [ + 313, + 334, + 554, + 428 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 313, + 334, + 554, + 428 + ], + "type": "text", + "content": " as observation points. Meanwhile, we refine the local neighborhood into near-field and far-field regions to extract fine-grained information, constructing the Acoustic Field Convolution (AF-Conv). As shown in Fig. 3(a), it can be formalized as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 348, + 438, + 553, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 438, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 348, + 438, + 553, + 452 + ], + "type": "interline_equation", + "content": "g _ {i} = \\mathcal {A} \\left(\\left\\{W _ {A F} \\left(p _ {i}, p _ {j}\\right) \\cdot f _ {j} \\mid p _ {j} \\in \\mathcal {N} \\left(p _ {i}\\right) \\right\\}\\right), \\tag {7}", + "image_path": "6928d307194ccfd3861ec18eca2dfd7442e167dde274b27973c99a4a383d82b3.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 460, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 508 + ], + "type": "text", + "content": "where the weight function " + }, + { + "bbox": [ + 313, + 460, + 554, + 508 + ], + "type": "inline_equation", + "content": "W_{AF}(p_i,p_j)" + }, + { + "bbox": [ + 313, + 460, + 554, + 508 + ], + "type": "text", + "content": " directly draws inspiration from acoustic field propagation characteristics and designs different weight calculation methods for near-field and far-field regions:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 330, + 516, + 553, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 516, + 553, + 549 + ], + "spans": [ + { + "bbox": [ + 330, + 516, + 553, + 549 + ], + "type": "interline_equation", + "content": "W _ {A F} \\left(p _ {i}, p _ {j}\\right) = \\left\\{ \\begin{array}{l l} e ^ {- \\beta d _ {i j}}, & \\text {i f} d _ {i j} < d _ {\\text {t h r e s h o l d}} \\\\ \\frac {1}{d _ {i j} + \\epsilon}, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {8}", + "image_path": "d58f0ffef51e51f8f0dab68d2c35952cff54b9dfae325dd020fc95b6132ea696.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "text", + "content": " represents learnable coefficient. " + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "inline_equation", + "content": "d_{ij}" + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "text", + "content": " denotes the Euclidean distance between " + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "text", + "content": ". In " + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "inline_equation", + "content": "W_{AF}(p_i,p_j)" + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "text", + "content": ", for the near-field region (" + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "inline_equation", + "content": "d_{ij} < d_{threshold}" + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "text", + "content": "), we use the exponential decay term " + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "inline_equation", + "content": "e^{-\\beta d_{ij}}" + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "text", + "content": " to capture subtle changes in local geometric structures, corresponding to the complex attenuation characteristics of the acoustic near-field region; for the far-field region (" + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "inline_equation", + "content": "d_{ij} \\geq d_{threshold}" + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "text", + "content": "), we adopt a simple reciprocal relationship " + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\frac{1}{d_{ij} + \\epsilon}" + }, + { + "bbox": [ + 313, + 555, + 555, + 665 + ], + "type": "text", + "content": " to better perceive broader contextual information." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 672, + 509, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 509, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 509, + 685 + ], + "type": "text", + "content": "3.3. Dynamic Acoustic Field Convolution" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "In sound wave propagation [31], waves of different frequencies exhibit distinct propagation characteristics. Inspired by" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22185" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "this multi-frequency phenomenon and dynamic convolution [62], as shown in Fig. 3(b), we propose Dynamic Acoustic Field Convolution (DAF-Conv), which extends AF-Conv by dynamically generating multiple sets of weights to simulate local geometric feature variations under different \"frequencies\". It can be formalized as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 152, + 295, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 152, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 152, + 295, + 167 + ], + "type": "interline_equation", + "content": "F _ {h} \\left(p _ {i}, p _ {j}\\right) = W _ {h} \\left(p _ {i}, p _ {j}\\right) \\cdot f _ {j}, \\tag {9}", + "image_path": "1f1cd7e606b511a20ae83409fa6c434b5d8c910dde0c64ec86e587d685009053.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 173, + 295, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 173, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 67, + 173, + 295, + 205 + ], + "type": "interline_equation", + "content": "W _ {h} \\left(p _ {i}, p _ {j}\\right) = \\left\\{ \\begin{array}{l l} \\alpha_ {h} e ^ {- \\beta_ {h} d _ {i j}}, & \\text {i f} d _ {i j} < d _ {\\text {t h r e s h o l d}} \\\\ \\frac {\\alpha_ {h}}{d _ {i j} + \\epsilon}, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {10}", + "image_path": "c3b39c854e6dd9c8a6227acef5510c3f3f851060fa483071c9c9557bc7b94c7a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 209, + 295, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 209, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 121, + 209, + 295, + 224 + ], + "type": "interline_equation", + "content": "\\left\\{\\alpha_ {h}, \\beta_ {h} \\right\\} _ {h = 1} ^ {H} = \\operatorname {M L P} \\left(h _ {j}\\right), \\tag {11}", + "image_path": "bea3480dc555ef3522a615229ca80c2b0cfeb566b84ea2e172892f07b2b23592.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 228, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 228, + 296, + 297 + ], + "type": "inline_equation", + "content": "d_{threshold}" + }, + { + "bbox": [ + 55, + 228, + 296, + 297 + ], + "type": "text", + "content": " is a scalar that defines the boundary between near-field and far-field regions. The weight coefficient " + }, + { + "bbox": [ + 55, + 228, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\alpha_{h}" + }, + { + "bbox": [ + 55, + 228, + 296, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\beta_{h}" + }, + { + "bbox": [ + 55, + 228, + 296, + 297 + ], + "type": "text", + "content": " are generated from local geometric prior " + }, + { + "bbox": [ + 55, + 228, + 296, + 297 + ], + "type": "inline_equation", + "content": "h_j = [p_i,p_j,p_j - p_i,\\| p_j - p_i\\| ]" + }, + { + "bbox": [ + 55, + 228, + 296, + 297 + ], + "type": "text", + "content": " through MLP to simulate propagation characteristics of different frequency components." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 299, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 334 + ], + "type": "text", + "content": "The local aggregated features are then separated into near-field and far-field components based on the distance threshold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 354, + 295, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 354, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 72, + 354, + 295, + 369 + ], + "type": "interline_equation", + "content": "F _ {h} ^ {\\text {n e a r}} \\left(p _ {i}\\right) = \\mathcal {A} \\left(\\left\\{F _ {h} \\left(p _ {i}, p _ {j}\\right) \\mid j \\in \\mathcal {N} _ {\\text {n e a r}} (i) \\right\\}\\right), \\tag {12}", + "image_path": "b4e6a4ff10e82e820b7eefdd39efaff0c8b45e9bc8d1f12e572d17a5e588b3e1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 376, + 295, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 376, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 77, + 376, + 295, + 392 + ], + "type": "interline_equation", + "content": "F _ {h} ^ {f a r} \\left(p _ {i}\\right) = \\mathcal {A} \\left(\\left\\{F _ {h} \\left(p _ {i}, p _ {j}\\right) \\mid j \\in \\mathcal {N} _ {f a r} (i) \\right\\}\\right), \\tag {13}", + "image_path": "f9d1a899b3f4f203abb45f93c934aeeb4ce1a166b4ff84a01faa9b9d92b23e87.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 394, + 295, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 295, + 442 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 394, + 295, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{near}(i) = \\{j\\mid p_j\\in \\mathcal{N}(p_i),d_{ij} < d_{threshold}\\}" + }, + { + "bbox": [ + 55, + 394, + 295, + 442 + ], + "type": "text", + "content": " represents the set of near-field neighbors of the center point " + }, + { + "bbox": [ + 55, + 394, + 295, + 442 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 394, + 295, + 442 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 394, + 295, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{far}(i) = \\{j\\mid p_j\\in \\mathcal{N}(p_i),d_{ij}\\geq d_{threshold}\\}" + }, + { + "bbox": [ + 55, + 394, + 295, + 442 + ], + "type": "text", + "content": " represents the set of far-field neighbors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 443, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 443, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 67, + 443, + 295, + 455 + ], + "type": "text", + "content": "The component for each frequency " + }, + { + "bbox": [ + 67, + 443, + 295, + 455 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 67, + 443, + 295, + 455 + ], + "type": "text", + "content": " can be obtained by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 464, + 295, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 464, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 107, + 464, + 295, + 480 + ], + "type": "interline_equation", + "content": "F _ {h} \\left(p _ {i}\\right) = F _ {h} ^ {\\text {n e a r}} \\left(p _ {i}\\right) + F _ {h} ^ {\\text {f a r}} \\left(p _ {i}\\right). \\tag {14}", + "image_path": "0df8f3d8e43634923634029051be78162803cf7a97d99d39a416f7202b113a02.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": "To achieve adaptive fusion of features from different frequency components, we introduce an attention mechanism based on geometric prior " + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": ". The attention weights are generated as fellow:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 93, + 543, + 295, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 543, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 93, + 543, + 295, + 558 + ], + "type": "interline_equation", + "content": "\\gamma_ {h} = \\operatorname {s o f t m a x} (\\operatorname {M a x P o o l i n g} (\\operatorname {M L P} (h _ {j}))). \\tag {15}", + "image_path": "30fd1ad7e6e17d87010e67da0da47bedc7180f4e0c9010f457a92d766de092b1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 560, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 597 + ], + "type": "text", + "content": "The final local aggregated feature is obtained by weighted summation of features from all frequency components:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 130, + 605, + 295, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 605, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 605, + 295, + 638 + ], + "type": "interline_equation", + "content": "g _ {i} = \\sum_ {h = 1} ^ {H} \\gamma_ {h} \\cdot F _ {h} \\left(p _ {i}\\right). \\tag {16}", + "image_path": "2d9938c195d4ca38b4b3d1e60e17a145a742a924486f269ec8437e16ea203933.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "DAF-Conv achieve adaptive modeling of local geometric features through separately processing near-field and far-field features for each frequency component, independent max-pooling aggregation for near-field and far-field features, and adaptive fusion of multiple frequency components through learned attention weights." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "Compared to the single weight function in AF-Conv, DAF-Conv achieves more sophisticated feature extraction capabilities through its multi-frequency design and separate processing of near-field and far-field features." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 129, + 555, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 555, + 152 + ], + "type": "text", + "content": "3.4. Dynamic Acoustic Field Convolution Neural Network" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 159, + 554, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 554, + 255 + ], + "type": "text", + "content": "To ensure each point has rich feature representation before constructing local neighborhoods, we build a Global Shape-Aware (GSA) layer before applying DAF-Conv. The GSA layer first aggregates each point through EdgeConv [55] without reducing point cloud resolution, then applies Multi-Head Attention (MHA) for global context modeling on the enhanced points. This layer effectively enhances DAF-Conv's perception ability for local structure modeling." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 255, + 554, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 554, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 554, + 315 + ], + "type": "text", + "content": "As shown in Fig. 2, based on the GSA layer and DAF-Conv, we construct the DAF Block. By stacking these blocks, we design a novel pretraining-free network, DAFNet, for point cloud analysis. We implement models for three major tasks:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 316, + 554, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 387 + ], + "type": "text", + "content": "Point Cloud Classification. For point cloud classification, we employ three DAF Blocks, followed by a classifier consisting of max-pooling and fully connected layers. The output feature dimensions of the three blocks are 64, 128, 256 respectively. Notably, with only 2.47M parameters, it achieves the best performance on ScanObjectNN [48]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 388, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 508 + ], + "type": "text", + "content": "Part Segmentation and Few-shot Point Cloud Semantic Segmentation. For part segmentation, the encoder structure remains identical to the classification task. In the decoder stage, we employ feature propagation to restore point cloud resolution. Between the encoder and decoder, we apply U-Net-like skip connections to fully utilize contextual information. In few-shot segmentation, while maintaining the same encoder-decoder structure as part segmentation, we incorporate the Query module from SegPN[77] to reduce domain gaps between query and support features." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 519, + 397, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 397, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 397, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 540, + 448, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 448, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 448, + 553 + ], + "type": "text", + "content": "4.1.3D Shape Classification" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 557, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 640 + ], + "type": "text", + "content": "Real-world Object Classification on ScanObjectNN. We evaluate our method on ScanObjectNN [48], a challenging real-world point cloud dataset containing 2,902 objects across 15 categories. Following previous works, we conduct experiments on its most challenging variant (PB_T50_RS), where objects are presented with background and occlusion." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "As shown in Table 1, DAFNet achieves state-of-the-art performance on ScanObjectNN [48], reaching " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "96.7\\%" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " overall accuracy (OA) and " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "96.3\\%" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " mean class accuracy (mAcc) with voting mechanism. These results significantly surpass both existing supervised methods (e.g., GPSFormer [52]: " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "95.4\\%" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " OA) and pre-training approaches (e.g., ReCon++" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22186" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 101, + 294, + 467 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 92 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 92 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 92 + ], + "type": "text", + "content": "Table 1. Classification results on the ScanObjectNN [48] dataset. \" - \" denotes unknown." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 101, + 294, + 467 + ], + "lines": [ + { + "bbox": [ + 57, + 101, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 57, + 101, + 294, + 467 + ], + "type": "table", + "html": "
MethodYearmAcc (%)OA (%)
Supervised Methods
PointNet [35]201763.468.2
PointNet++ [36]201869.873.7
DGCNN [55]201973.678.1
SimpleView [15]2021-80.5
MVTN [18]202183.185.5
PointMLP [28]202284.485.7
RepSurf-U [40]202283.186.0
PointNeXt [39]202285.887.7
Point-PN [68]2023-87.1
DeLA [2]202389.390.4
GPSFormer [52]202493.895.4
Mamba3D [19]2024-92.6
Pre-training Methods
Point-BERT [65]2021-83.1
Point-MAE [33]2022-85.2
ULIP-2 + PointNeXt [61]202391.291.5
PointGPT [3]2023-93.6
ReCon [41]2023-91.3
ReCon++ [38]2024-95.25
PointMamba [25]2024-89.3
PCP-MAE [70]2024-90.4
Ours
DAFNet (w/o voting)-96.196.4
DAFNet (w/ voting)-96.396.7
", + "image_path": "ae65ce5a5890dc0da1f96ce041b5ca8baf1417a2e06f6238216baa944c7bf167.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 497, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 534 + ], + "type": "text", + "content": "[38]: " + }, + { + "bbox": [ + 55, + 497, + 295, + 534 + ], + "type": "inline_equation", + "content": "95.25\\%" + }, + { + "bbox": [ + 55, + 497, + 295, + 534 + ], + "type": "text", + "content": " OA). The superior performance demonstrates the effectiveness of our acoustic field-inspired architecture in handling real-world point cloud data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 539, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 589 + ], + "type": "text", + "content": "Synthetic Object Classification on ModelNet40. We further evaluate DAFNet on ModelNet40 [59], a widely-used synthetic dataset containing 12,311 CAD models from 40 categories, with 9,843 for training and 2,468 for testing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": "As presented in Table 2, DAFNet achieves competitive results on ModelNet40 [59] with " + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "94.1\\%" + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": " overall accuracy (OA) and " + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "92.2\\%" + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": " mean class accuracy (mAcc). While these results are slightly lower than some recent methods like PointGPT [3] (" + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "94.9\\%" + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": " OA), the performance gap is minimal (less than " + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": "). This is mainly because ModelNet40 [3], as a synthetic dataset, has been well-addressed by existing methods. In contrast, DAFNet shows more significant advantages on real-world datasets, as demonstrated by our results on ScanObjectNN [48]." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 102, + 553, + 454 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 2. Classification results on ModelNet40 dataset. \"-\" denotes unknown." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 102, + 553, + 454 + ], + "lines": [ + { + "bbox": [ + 316, + 102, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 102, + 553, + 454 + ], + "type": "table", + "html": "
MethodsYearmAcc (%)OA (%)
Supervised Methods
PointNet [35]201786.089.2
PointNet++[36]2017-90.7
DGCNN[55]201990.292.9
Point Transformer [71]202190.693.7
PAConv [60]2021-93.9
CurveNet [32]2021-94.2
PointMLP [28]202291.494.5
RepSurf-U [40]2022-94.7
GPSFormer [52]202492.294.2
Mamba3D [19]2024-94.1
Pre-training Methods
Point-BERT [65]2021-93.8
Point-MAE [33]2022-94.0
ULIP + PointMLP [61]202392.494.7
PointGPT [3]2023-94.9
ReCon [41]2023-94.5
ReCon++ [38]2024-94.8
PointMamba [25]2024-93.6
PCP-MAE [70]2024-94.0
Ours
DAFNet (w/o voting)-91.993.8
DAFNet (w/ voting)-92.294.1
", + "image_path": "37f9fc03dace8d9ab885bf6e5ea20f62badfb9125e48382f4335eacde8180976.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 474, + 424, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 424, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 424, + 487 + ], + "type": "text", + "content": "4.2. Part Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 552 + ], + "type": "text", + "content": "To evaluate our method's capability in fine-grained shape understanding, we conduct experiments on ShapeNetPart [64], which contains 16,881 shapes from 16 categories with 50 part annotations. Following convention, we use 13,807 samples for training and 2,874 for testing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 552, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 624 + ], + "type": "text", + "content": "As shown in Table 3, DAFNet achieves competitive performance with " + }, + { + "bbox": [ + 313, + 552, + 555, + 624 + ], + "type": "inline_equation", + "content": "85.2\\%" + }, + { + "bbox": [ + 313, + 552, + 555, + 624 + ], + "type": "text", + "content": " class mIoU and " + }, + { + "bbox": [ + 313, + 552, + 555, + 624 + ], + "type": "inline_equation", + "content": "86.8\\%" + }, + { + "bbox": [ + 313, + 552, + 555, + 624 + ], + "type": "text", + "content": " instance mIoU, comparable to state-of-the-art methods. These results demonstrate that our acoustic field-inspired approach effectively captures both global shape information and local geometric details necessary for part segmentation tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 632, + 552, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 552, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 552, + 645 + ], + "type": "text", + "content": "4.3. Few-shot Point Cloud Semantic Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 650, + 503, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 503, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 503, + 662 + ], + "type": "text", + "content": "4.3.1. Results analysis on the S3DIS dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "We evaluate our model on S3DIS [1], which contains 3D point clouds from 272 rooms with 13 semantic categories. Following [73], we divide scenes into " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "1\\mathrm{m} \\times 1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": " blocks and sample 2048 points per block, resulting in 7547" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22187" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 102, + 294, + 346 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "text", + "content": "Table 3. Part segmentation results (\\%) on ShapeNetPart. We report both class mIoU and instance mIoU." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 102, + 294, + 346 + ], + "lines": [ + { + "bbox": [ + 58, + 102, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 294, + 346 + ], + "type": "table", + "html": "
MethodsYearclass mIoUinstance mIoU
Supervised Methods
PointNet [35]201780.483.7
PointNet++[36]201781.985.1
DGCNN [55]201982.385.1
KPConv [47]201985.186.4
Point Transformer [71]202183.786.6
CurveNet [32]2021-86.8
PointNeXt [39]202285.287.1
SPOTr [34]202385.487.2
GPSFormer [52]202485.486.8
Pre-training Methods
PointGPT[3]202384.186.2
ReCon [41]202384.886.4
PointMamba[25]202484.486.2
PCP-MAE [70]202484.986.1
Ours
DAFNet-85.286.8
", + "image_path": "f8eb4dce74e3afe2a0f39310eef621f78a85a5e833a2cad495441a376f7a2e76.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 400, + 294, + 487 + ], + "blocks": [ + { + "bbox": [ + 55, + 357, + 296, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 392 + ], + "type": "text", + "content": "Table 4. Few-shot Results (\\%) on S3DIS. " + }, + { + "bbox": [ + 55, + 357, + 296, + 392 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 357, + 296, + 392 + ], + "type": "text", + "content": " denotes the split " + }, + { + "bbox": [ + 55, + 357, + 296, + 392 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 357, + 296, + 392 + ], + "type": "text", + "content": " is used for testing. Avg is their average mIoU. The best results are shown in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 400, + 294, + 487 + ], + "lines": [ + { + "bbox": [ + 57, + 400, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 400, + 294, + 487 + ], + "type": "table", + "html": "
Method2-way3-way
1-shot5-shot1-shot5-shot
S0S1AvgS0S1AvgS0S1AvgS0S1Avg
DGCNN [55]36.3438.7937.5756.4956.9956.7430.0532.1931.1246.8847.5747.23
ProtoNet [44]48.3949.9849.1957.3463.2260.2840.8145.0742.9449.0553.4251.24
MPTI [73]52.2751.4851.8858.9360.5659.7544.2746.9245.6051.7448.5750.16
AttMPTI [73]53.7755.9454.8661.6767.0264.3545.1849.2747.2354.9256.7955.86
BFG [29]55.6055.9855.7963.7166.6265.1746.1848.3647.2755.0557.8056.43
2CBR [75]55.8961.9958.9463.5567.5165.5346.5153.9150.2155.5158.0756.79
PAP3D [20]59.4566.0862.7665.4070.3067.8548.9956.5752.7861.2760.8161.04
Seg-PN [77]64.8467.9866.4167.6371.4869.5659.1160.4259.7759.4864.7262.10
DAFNet68.1370.2769.2070.5173.1571.8361.3365.5563.4465.2568.6766.96
Improvement+3.29+2.29+2.79+2.88+1.67+2.27+2.22+5.13+3.67+5.77+3.95+4.86
", + "image_path": "908f98f9e481d3c88ba0b918f6c2dc820326c5232247d5e82954bf9b8bc65e02.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "text", + "content": "blocks. As shown in Table 4, DAFNet achieves state-of-the-art performance across all few-shot settings. In 2-way experiments, our method achieves " + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "inline_equation", + "content": "69.20\\%" + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "text", + "content": " mIoU in 1-shot and " + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "inline_equation", + "content": "71.83\\%" + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "text", + "content": " mIoU in 5-shot settings, surpassing the previous best method Seg-PN [77] by " + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "inline_equation", + "content": "2.79\\%" + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "inline_equation", + "content": "2.27\\%" + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "text", + "content": " respectively. For the more challenging 3-way scenario, DAFNet maintains its superiority with " + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "inline_equation", + "content": "63.44\\%" + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "text", + "content": " mIoU in 1-shot and " + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "inline_equation", + "content": "66.96\\%" + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "text", + "content": " mIoU in 5-shot settings, outperforming Seg-PN [77] by " + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "inline_equation", + "content": "3.67\\%" + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "inline_equation", + "content": "4.86\\%" + }, + { + "bbox": [ + 54, + 509, + 295, + 641 + ], + "type": "text", + "content": ". These improvements across different settings demonstrate DAFNet's strong feature extraction capability and generalization ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 650, + 253, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 253, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 253, + 662 + ], + "type": "text", + "content": "4.3.2. Results analysis on the ScanNet dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "We further evaluate DAFNet on ScanNet [7], which contains 1,513 real-world scanned scenes with 20 semantic categories. Following [73], we process the scenes into 36,350 blocks with 2,048 points each. As shown in Ta" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 113, + 553, + 200 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 5. Few-shot Results (\\%) on ScanNet. " + }, + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "text", + "content": " denotes the split " + }, + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 70, + 555, + 104 + ], + "type": "text", + "content": " is used for testing. Avg is their average mIoU. The best results are shown in bold." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 113, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 316, + 113, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 316, + 113, + 553, + 200 + ], + "type": "table", + "html": "
Method2-way3-way
1-shot5-shot1-shot5-shot
\\( S_0 \\)\\( S_1 \\)Avg.\\( S_0 \\)\\( S_1 \\)Avg.\\( S_0 \\)\\( S_1 \\)Avg.\\( S_0 \\)\\( S_1 \\)Avg.
DGCNN [55]31.5528.9430.2542.7137.2439.9823.9919.1021.5534.9328.1031.52
ProtoNet [44]33.9230.9532.4445.3442.0143.6828.4726.1327.3037.3634.9836.17
MPTI [73]39.2736.1437.7146.9043.5945.2529.9627.2628.6138.1434.3636.25
AttMPTI [73]42.5540.8341.6954.0050.3252.1635.2330.7232.9846.7440.8043.77
BFG [29]42.1540.5241.3451.2349.3950.3134.1231.9833.0546.2541.3843.82
2CBR [75]50.7347.6649.2052.3547.1449.7547.0046.3646.6845.0639.4742.27
PAP3D [20]57.0855.9456.5164.5559.6462.1055.2755.6055.4459.0253.1656.09
Seg-PN [77]63.1564.3263.7467.0869.0568.0761.8065.3463.5762.9468.2665.60
DAFNet68.7969.9569.3770.9170.6070.7666.1466.7066.4268.9771.9570.46
Improvement+5.64+5.63+5.63+3.83+1.55+2.69+4.34+1.36+2.85+6.03+3.69+4.86
", + "image_path": "bc41661f682d844ebfcf108f9ef57a8d3dc62ad5713f7b2218e8c8737e8739e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 338, + 230, + 531, + 406 + ], + "blocks": [ + { + "bbox": [ + 339, + 209, + 529, + 221 + ], + "lines": [ + { + "bbox": [ + 339, + 209, + 529, + 221 + ], + "spans": [ + { + "bbox": [ + 339, + 209, + 529, + 221 + ], + "type": "text", + "content": "Table 6. Effect of different components of DAFNet." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 338, + 230, + 531, + 406 + ], + "lines": [ + { + "bbox": [ + 338, + 230, + 531, + 406 + ], + "spans": [ + { + "bbox": [ + 338, + 230, + 531, + 406 + ], + "type": "table", + "html": "
GSA Layer\nEdgeConv MHADAF-ConvOA (%)
Individual Components
✓XX88.6
X✓X89.4
XX✓91.7
Component Combinations
✓✓X91.3
✓X✓94.8
X✓✓93.5
✓✓✓96.7
", + "image_path": "4d4a625f37a5e7da63739143f02eedef36cd88b6f243d194c53a131ba2701bac.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "content": "ble 5, DAFNet achieves consistent improvements across all experimental settings. For 2-way scenarios, our method obtains " + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "inline_equation", + "content": "69.37\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "content": " mIoU in 1-shot and " + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "inline_equation", + "content": "70.76\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "content": " mIoU in 5-shot settings, surpassing Seg-PN [77] by " + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "inline_equation", + "content": "5.63\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "inline_equation", + "content": "2.69\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "content": " respectively. In the more challenging 3-way settings, DAFNet achieves " + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "inline_equation", + "content": "66.42\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "content": " mIoU in 1-shot and " + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "inline_equation", + "content": "70.46\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "content": " mIoU in 5-shot settings, outperforming Seg-PN [77] by " + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "inline_equation", + "content": "2.85\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "inline_equation", + "content": "4.86\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "content": ". These substantial improvements across different settings demonstrate DAFNet's strong generalization ability and robustness in handling complex real-world scenes with limited training samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 566, + 409, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 409, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 409, + 578 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 583, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 555, + 608 + ], + "type": "text", + "content": "All experiments are conducted on the ScanObjectNN [48] dataset unless explicitly stated otherwise." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 614, + 495, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 495, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 495, + 627 + ], + "type": "text", + "content": "4.4.1. Ablation Study of Each Component" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "We conduct ablation experiments to analyze the effectiveness of each component in DAFNet, as shown in Table 6. Among individual components, DAF-Conv shows the strongest performance with " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "91.7\\%" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " accuracy, notably higher than single GSA layer components (EdgeConv: " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "88.6\\%" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ", MHA: " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "89.4\\%" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "). When combining components, the fusion of DAF-Conv with EdgeConv achieves " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "94.8\\%" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ", while DAF" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22188" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 91, + 257, + 167 + ], + "blocks": [ + { + "bbox": [ + 79, + 70, + 271, + 82 + ], + "lines": [ + { + "bbox": [ + 79, + 70, + 271, + 82 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 271, + 82 + ], + "type": "text", + "content": "Table 7. Effect of different local geometric priors " + }, + { + "bbox": [ + 79, + 70, + 271, + 82 + ], + "type": "inline_equation", + "content": "{h}_{j}" + }, + { + "bbox": [ + 79, + 70, + 271, + 82 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 91, + 257, + 167 + ], + "lines": [ + { + "bbox": [ + 94, + 91, + 257, + 167 + ], + "spans": [ + { + "bbox": [ + 94, + 91, + 257, + 167 + ], + "type": "table", + "html": "
Local Geometric Prior hjOA(%)
[pj]93.6
[pj,pi]94.7
[pj,pipj-pi]96.2
[pj,pi, pj-pi, ||pj-pi||]96.7
", + "image_path": "8f7b2682f0f39d1ce53da5eb4eef5f9a4837058aa3356b55681f77d0b521ccc7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 296, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 259 + ], + "type": "text", + "content": "Conv with MHA reaches " + }, + { + "bbox": [ + 55, + 186, + 296, + 259 + ], + "type": "inline_equation", + "content": "93.5\\%" + }, + { + "bbox": [ + 55, + 186, + 296, + 259 + ], + "type": "text", + "content": ", both showing substantial improvements over single component performance. The full model integrating all components achieves the best performance of " + }, + { + "bbox": [ + 55, + 186, + 296, + 259 + ], + "type": "inline_equation", + "content": "96.7\\%" + }, + { + "bbox": [ + 55, + 186, + 296, + 259 + ], + "type": "text", + "content": ", demonstrating the strong complementary effects between DAF-Conv's local geometric modeling and the GSA layer's global context capturing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 258, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 258, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 258, + 277 + ], + "type": "text", + "content": "4.4.2. Ablation Study of Local Geometric Prior" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "content": "As shown in Table 7, we conduct ablation studies on the composition of local geometric prior " + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "content": ". Starting with only center point coordinates " + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "inline_equation", + "content": "[p_j]" + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "content": " yields " + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "inline_equation", + "content": "93.6\\%" + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "content": " accuracy. Incorporating neighbor point information " + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "inline_equation", + "content": "[p_j, p_i]" + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "content": " brings a " + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "content": " improvement. Adding relative position " + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "inline_equation", + "content": "[p_j, p_i, p_j - p_i]" + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "content": " further boosts performance by " + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "content": ". The complete geometric prior with Euclidean distance " + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "inline_equation", + "content": "||p_j - p_i||" + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "content": " achieves the best result of " + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "inline_equation", + "content": "96.7\\%" + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "content": ", validating the importance of comprehensive geometric information in local structure modeling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 239, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 239, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 239, + 407 + ], + "type": "text", + "content": "4.4.3. Analysis of Model Hyperparameters" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 411, + 295, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 435 + ], + "type": "text", + "content": "We conduct comprehensive ablation studies on three key hyperparameters of DAFNet, as shown in Fig. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "text", + "content": "Frequency Component Number H: Fig. 4(a) shows that model performance improves as " + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "text", + "content": " increases from 1 to 6, achieving peak accuracy at " + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "inline_equation", + "content": "H = 6" + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "inline_equation", + "content": "(96.7\\%)" + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "text", + "content": ". Further increasing " + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "text", + "content": " leads to performance degradation, dropping to " + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "inline_equation", + "content": "95.1\\%" + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "inline_equation", + "content": "H = 10" + }, + { + "bbox": [ + 55, + 435, + 296, + 518 + ], + "type": "text", + "content": ". This indicates that while multiple frequency components enhance feature diversity, excessive components may introduce redundant information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 519, + 296, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 603 + ], + "type": "text", + "content": "Distance Threshold " + }, + { + "bbox": [ + 55, + 519, + 296, + 603 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{\\mathrm{threshold}}" + }, + { + "bbox": [ + 55, + 519, + 296, + 603 + ], + "type": "text", + "content": ": As shown in Fig. 4(b), we examine the impact of " + }, + { + "bbox": [ + 55, + 519, + 296, + 603 + ], + "type": "inline_equation", + "content": "d_{threshold}" + }, + { + "bbox": [ + 55, + 519, + 296, + 603 + ], + "type": "text", + "content": " which divides the local neighborhood into near-field and far-field regions. The optimal performance (96.7%) is achieved at " + }, + { + "bbox": [ + 55, + 519, + 296, + 603 + ], + "type": "inline_equation", + "content": "d_{threshold} = 0.6" + }, + { + "bbox": [ + 55, + 519, + 296, + 603 + ], + "type": "text", + "content": ". Lower thresholds limit near-field information, while higher values may include irrelevant points in the near-field region, both resulting in suboptimal performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 603, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 296, + 664 + ], + "type": "text", + "content": "Neighborhood Size K: Fig. 4(c) demonstrates the effect of neighborhood size " + }, + { + "bbox": [ + 55, + 603, + 296, + 664 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 603, + 296, + 664 + ], + "type": "text", + "content": " in EdgeConv. Performance reaches its peak " + }, + { + "bbox": [ + 55, + 603, + 296, + 664 + ], + "type": "inline_equation", + "content": "(96.7\\%)" + }, + { + "bbox": [ + 55, + 603, + 296, + 664 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 603, + 296, + 664 + ], + "type": "inline_equation", + "content": "K = 20" + }, + { + "bbox": [ + 55, + 603, + 296, + 664 + ], + "type": "text", + "content": ". Smaller " + }, + { + "bbox": [ + 55, + 603, + 296, + 664 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 603, + 296, + 664 + ], + "type": "text", + "content": " values provide insufficient local context, while larger values may incorporate irrelevant points, both leading to decreased accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 672, + 165, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 165, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 165, + 685 + ], + "type": "text", + "content": "4.5. Model Complexity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "We compare the model complexity of DAFNet with existing methods on ScanObjectNN [48], as shown in Ta" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 396, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 396, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 396, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 396, + 148 + ], + "type": "image", + "image_path": "b40574680401fdf325107935a697c4e00cbc4851a869ef42ef3f47fbc0da4ded.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 158, + 555, + 202 + ], + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 202 + ], + "type": "text", + "content": "Figure 4. Ablation studies on key hyperparameters. (a) Impact of frequency component number " + }, + { + "bbox": [ + 313, + 158, + 555, + 202 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 158, + 555, + 202 + ], + "type": "text", + "content": " in DAF-Conv. (b) Effect of distance threshold " + }, + { + "bbox": [ + 313, + 158, + 555, + 202 + ], + "type": "inline_equation", + "content": "d_{threshold}" + }, + { + "bbox": [ + 313, + 158, + 555, + 202 + ], + "type": "text", + "content": " for near-field and far-field separation. (c) Influence of neighborhood size " + }, + { + "bbox": [ + 313, + 158, + 555, + 202 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 158, + 555, + 202 + ], + "type": "text", + "content": " in EdgeConv layer." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 396, + 71, + 475, + 148 + ], + "blocks": [ + { + "bbox": [ + 396, + 71, + 475, + 148 + ], + "lines": [ + { + "bbox": [ + 396, + 71, + 475, + 148 + ], + "spans": [ + { + "bbox": [ + 396, + 71, + 475, + 148 + ], + "type": "image", + "image_path": "675e9445c731bc33922d1e5691f3052f95db5cc8b578c3a8a2bad7cea864947c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 475, + 72, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 475, + 72, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 475, + 72, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 475, + 72, + 553, + 148 + ], + "type": "image", + "image_path": "5d70e76dbb6bdf3905d656993436fbb2175c7bdddde2c971181cd890cd614d50.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 318, + 242, + 552, + 390 + ], + "blocks": [ + { + "bbox": [ + 326, + 220, + 541, + 232 + ], + "lines": [ + { + "bbox": [ + 326, + 220, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 326, + 220, + 541, + 232 + ], + "type": "text", + "content": "Table 8. Model complexity comparison on ScanObjectNN." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 242, + 552, + 390 + ], + "lines": [ + { + "bbox": [ + 318, + 242, + 552, + 390 + ], + "spans": [ + { + "bbox": [ + 318, + 242, + 552, + 390 + ], + "type": "table", + "html": "
MethodsOA(%)Params.(M)FLOPs(G)
PointNet [35]68.03.50.5
PointNet++[36]77.91.51.7
DGCNN[55]78.11.82.4
PointCNN [24]78.50.6-
MVTN [18]82.811.243.7
PointMLP [28]85.412.631.4
PointNeXt [39]87.41.43.6
DAFNet (Ours)96.72.471.1
", + "image_path": "52d9a26284470b6363e40d04a8a0d6c66832301da93ae72309bcb6eda084992d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "text", + "content": "ble 8. While achieving state-of-the-art performance (96.7% OA), DAFNet maintains moderate computational requirements with only 2.47M parameters and 1.1G FLOPs. This is significantly more efficient than models like MVTN [18] (11.2M parameters, 43.7G FLOPs) and PointMLP [28] (12.6M parameters, 31.4G FLOPs), demonstrating that our acoustic field-inspired design achieves superior accuracy without sacrificing computational efficiency." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 534, + 388, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 534, + 388, + 547 + ], + "spans": [ + { + "bbox": [ + 314, + 534, + 388, + 547 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "text", + "content": "In this paper, we presented DAFNet, a physics-inspired architecture for point cloud understanding that addresses local structure representation challenges. Inspired by acoustic field principles, we introduced AF-Conv modeling local feature aggregation through acoustic propagation, and extended it to DAF-Conv which dynamically adapts to diverse geometric features through multi-frequency simulation. By integrating these with GSA, our approach captures both fine-grained local structures and global context. Experiments demonstrate DAFNet achieves state-of-the-art performance across multiple tasks. Future work will extend acoustic field modeling to other domains, exploring applications in additional 3D tasks." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22189" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 137 + ], + "type": "text", + "content": "This project is sponsored by Shanghai Pujiang Programme 24PJD030. The computational work for this article is partially performed on resources of the National Supercomputing Centre, Singapore (https://www.nscc.sg)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 176, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 176, + 296, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 176, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 176, + 296, + 220 + ], + "type": "text", + "content": "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, pages 1534-1543, 2016. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 222, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 222, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 296, + 255 + ], + "type": "text", + "content": "[2] Binjie Chen, Yunzhou Xia, Yu Zang, Cheng Wang, and Jonathan Li. Decoupled local aggregation for point cloud learning. arXiv preprint arXiv:2308.16532, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 256, + 296, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 256, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 61, + 256, + 296, + 298 + ], + "type": "text", + "content": "[3] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre-training from point clouds. arXiv preprint arXiv:2305.11487, 2023. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 300, + 294, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 300, + 294, + 332 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 294, + 332 + ], + "type": "text", + "content": "[4] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. NeurIPS, 36, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 334, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 334, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 334, + 296, + 388 + ], + "type": "text", + "content": "[5] Ziyang Chen, Israel D Gebru, Christian Richardt, Anurag Kumar, William Laney, Andrew Owens, and Alexander Richard. Real acoustic fields: An audio-visual room acoustics dataset and benchmark. In CVPR, pages 21886–21896, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 390, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 390, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 390, + 296, + 432 + ], + "type": "text", + "content": "[6] Pranav Singh Chib and Pravendra Singh. Recent advancements in end-to-end autonomous driving using deep learning: A survey. IEEE Transactions on Intelligent Vehicles, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 434, + 296, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 478 + ], + "type": "text", + "content": "[7] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 480, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 523 + ], + "type": "text", + "content": "[8] Jeevan S Devagiri, Sidike Paheding, Quamar Niyaz, Xiaoli Yang, and Samantha Smith. Augmented reality and artificial intelligence in industry: Trends, tools, and future challenges. Expert Systems with Applications, 207:118002, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "type": "text", + "content": "[9] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "type": "text", + "content": "[10] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. Mose: A new dataset for video object segmentation in complex scenes. In ICCV, pages 20224-20234, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "type": "text", + "content": "[11] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[12] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, pages 8458-8468, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[13] Xiang Fang, Wanlong Fang, Changshuo Wang, Daizong Liu, Keke Tang, Jianfeng Dong, Pan Zhou, and Beibei Li. Multipair temporal sentence grounding via multi-thread knowledge transfer network. arXiv preprint arXiv:2412.15678, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 554, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 554, + 163 + ], + "type": "text", + "content": "[14] Ruchi Goel and Pooja Gupta. Robotics and industry 4.0. A Roadmap to Industry 4.0: Smart Production, Sharp Business and Sustainable Development, pages 157-169, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 165, + 554, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 165, + 554, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 165, + 554, + 208 + ], + "type": "text", + "content": "[15] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In ICML, pages 3809-3820. PMLR, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 210, + 554, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 210, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 210, + 554, + 243 + ], + "type": "text", + "content": "[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 245, + 554, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 554, + 288 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 554, + 288 + ], + "type": "text", + "content": "[17] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 291, + 554, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 554, + 324 + ], + "type": "text", + "content": "[18] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In ICCV, pages 1-11, 2021. 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 326, + 554, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 554, + 369 + ], + "type": "text", + "content": "[19] Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud analysis via state space model. In ACM MM, pages 4995-5004, 2024. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 372, + 554, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 414 + ], + "type": "text", + "content": "[20] Shuting He, Xudong Jiang, Wei Jiang, and Henghui Ding. Prototype adaption and projection for few-and zero-shot 3d point cloud semantic segmentation. TIP, 32:3199-3211, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 417, + 554, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 554, + 472 + ], + "type": "text", + "content": "[21] Limin Jiang, Changshuo Wang, Xin Ning, and Zaiyang Yu. Lttpoint: a mlp-based point cloud classification method with local topology transformation module. In 2023 7th Asian Conference on Artificial Intelligence Technology (ACAIT), pages 783-789. IEEE, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 474, + 554, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 474, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 474, + 554, + 506 + ], + "type": "text", + "content": "[22] Lawrence E Kinsler, Austin R Frey, Alan B Coppens, and James V Sanders. Fundamentals of acoustics. John wiley & sons, 2000. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 509, + 554, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 554, + 541 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 554, + 541 + ], + "type": "text", + "content": "[23] Artem Komarichev, Zichun Zhong, and Jing Hua. A-cnn: Annually convolutional neural networks on point clouds. In CVPR, pages 7421-7430, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 544, + 554, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 554, + 576 + ], + "type": "text", + "content": "[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. NeurIPS, 31, 2018. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 578, + 554, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 554, + 621 + ], + "type": "text", + "content": "[25] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 554, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 666 + ], + "type": "text", + "content": "[26] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual state space model. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22190" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 711 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[28] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 173 + ], + "type": "text", + "content": "[29] Yongqiang Mao, Zonghao Guo, LU Xiaonan, Zhiqiang Yuan, and Haowen Guo. Bidirectional feature globalization for few-shot semantic segmentation of 3d point cloud scenes. In 2022 International Conference on 3D Vision (3DV), pages 505-514. IEEE, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "type": "text", + "content": "[30] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 922-928. IEEE, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 221, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 221, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 57, + 221, + 294, + 243 + ], + "type": "text", + "content": "[31] Philip McCord Morse and K Uno Ingard. Theoretical acoustics. Princeton university press, 1986. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 246, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 294, + 289 + ], + "type": "text", + "content": "[32] AAM Muzahid, Wanggen Wan, Ferdous Sohel, Lianyao Wu, and Li Hou. Curvenet: Curvature-based multitask learning deep networks for 3d object recognition. IEEE/CAA Journal of Automatica Sinica, 8(6):1177-1187, 2020. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 292, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 294, + 335 + ], + "type": "text", + "content": "[33] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, pages 604-621. Springer, 2022. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 337, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 294, + 380 + ], + "type": "text", + "content": "[34] Jinyoung Park, Sanghyeok Lee, Sihyeon Kim, Yunyang Xiong, and Hyunwoo J Kim. Self-positioning point-based transformer for point cloud understanding. In CVPR, pages 21814-21823, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 416 + ], + "type": "text", + "content": "[35] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 2, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 418, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 294, + 451 + ], + "type": "text", + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NIPS, 30, 2017. 2, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 453, + 294, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 506 + ], + "type": "text", + "content": "[37] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In ICML, pages 28223-28243. PMLR, 2023. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 553 + ], + "type": "text", + "content": "[38] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. *Shapellm: Universal 3d object understanding for embodied interaction*. In *ECCV*, pages 214–238. Springer, 2025. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "type": "text", + "content": "[39] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. NIPS, 35:23192-23204, 2022. 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "type": "text", + "content": "[40] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In CVPR, pages 18942-18952, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 711 + ], + "type": "text", + "content": "[41] Jingjing Ren, Ashwin Rao, Martina Lindorfer, Arnaud Legout, and David Choffnes. Recon: Revealing and controlling pii leaks in mobile network traffic. In Proceedings of the 14th Annual International Conference on Mobile Systems, Applications, and Services, pages 361-374, 2016. 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[42] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, pages 3577-3586, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "text", + "content": "[43] Mickael Sereno, Xiyao Wang, Lonni Besançon, Michael J Mcguffin, and Tobias Isenberg. Collaborative work in augmented reality: A survey. TVCG, 28(6):2530-2549, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 154, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 175 + ], + "type": "text", + "content": "[44] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. NIPS, 30, 2017. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 178, + 553, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 178, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 316, + 178, + 553, + 210 + ], + "type": "text", + "content": "[45] Mohsen Soori, Behrooz Arezoo, and Roza Dastres. Artificial intelligence, machine learning and deep learning in advanced robotics, a review. Cognitive Robotics, 3:54-70, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 213, + 553, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 213, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 553, + 256 + ], + "type": "text", + "content": "[46] Pei Sun, Mingxing Tan, Weiyue Wang, Chenxi Liu, Fei Xia, Zhaoqi Leng, and Dragomir Anguelov. Swformer: Sparse window transformer for 3d object detection in point clouds. In ECCV, pages 426-442. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 258, + 553, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 553, + 301 + ], + "type": "text", + "content": "[47] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6411–6420, 2019. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 304, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 358 + ], + "type": "text", + "content": "[48] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, pages 1588–1597, 2019. 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "text", + "content": "[49] Changshuo Wang, Xin Ning, Linjun Sun, Liping Zhang, Weijun Li, and Xiao Bai. Learning discriminative features by covering local geometric space for point cloud analysis. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-15, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 418, + 553, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 553, + 451 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 553, + 451 + ], + "type": "text", + "content": "[50] C Wang, H Wang, X Ning, S Tian, and W Li. 3d point cloud classification method based on dynamic coverage of local area. J. Softw, 34(4):1962-1976, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "text", + "content": "[51] Changshuo Wang, Xin Ning, Weijun Li, Xiao Bai, and Xingyu Gao. 3d person re-identification based on global semantic guidance and local feature aggregation. TCSVT, 34 (6):4698-4712, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 499, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 553 + ], + "type": "text", + "content": "[52] Changshuo Wang, Meiqing Wu, Siew-Kei Lam, Xin Ning, Shangshu Yu, Ruiping Wang, Weijun Li, and Thambipillai Srikanthan. Gpsformer: A global perception and local structure fitting-based transformer for point cloud understanding. arXiv preprint arXiv:2407.13519, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "type": "text", + "content": "[53] Changshuo Wang, Rongsheng Cao, and Ruiping Wang. Learning discriminative topological structure information representation for 2d shape and social network classification via persistent homology. Knowledge-Based Systems, page 113125, 2025. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 666 + ], + "type": "text", + "content": "[54] Ruiping Wang, Siew-Kei Lam, Meiqing Wu, Zhijian Hu, Changshuo Wang, and Jing Wang. Destination intention estimation-based convolutional encoder-decoder for pedestrian trajectory multimodality forecast. Measurement, 239: 115470, 2025. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[55] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22191" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[56] Xin Wei, Ruixuan Yu, and Jian Sun. View-gcn: View-based graph convolutional network for 3d shape analysis. In CVPR, pages 1850-1859, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[57] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[58] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, pages 4840-4851, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 230 + ], + "type": "text", + "content": "[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, pages 1912-1920, 2015. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "text", + "content": "[60] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, pages 3173-3182, 2021. 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[61] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In CVPR, pages 1179-1189, 2023. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "text", + "content": "[62] Brandon Yang, Gabriel Bender, Quoc V Le, and Jiquan Ngiam. Condconv: Conditionally parameterized convolutions for efficient inference. NIPS, 32, 2019. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 397 + ], + "type": "text", + "content": "[63] Ze Yang and Liwei Wang. Learning relationships for multiview 3d object recognition. In ICCV, pages 7505-7514, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "text", + "content": "[64] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "text", + "content": "[65] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, pages 19313-19322, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 501, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 553 + ], + "type": "text", + "content": "[66] Huang Zhang, Changshuo Wang, Long Yu, Shengwei Tian, Xin Ning, and Joel Rodrigues. Pointgt: A method for pointcloud classification and segmentation based on local geometric transformation. IEEE Transactions on Multimedia, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 556, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 600 + ], + "type": "text", + "content": "[67] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. NeurIPS, 35:27061-27074, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 644 + ], + "type": "text", + "content": "[68] Renrui Zhang, Liuhui Wang, Ziyu Guo, Yali Wang, Peng Gao, Hongsheng Li, and Jianbo Shi. Parameter is not all you need: Starting from non-parametric networks for 3d point cloud analysis. arXiv preprint arXiv:2303.08134, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 688 + ], + "type": "text", + "content": "[69] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024.2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[70] Xiangdong Zhang, Shaofeng Zhang, and Junchi Yan. Pcmae: Learning to predict centers for point masked autoen" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 397 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "coders. Advances in Neural Information Processing Systems, 37:80303-80327, 2024. 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[71] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[72] Jingyuan Zhao, Wenyi Zhao, Bo Deng, Zhenghong Wang, Feng Zhang, Wenxiang Zheng, Wanke Cao, Jinrui Nan, Yubo Lian, and Andrew F Burke. Autonomous driving system: A comprehensive survey. Expert Systems with Applications, page 122836, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 217 + ], + "type": "text", + "content": "[73] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In CVPR, pages 8873-8882, 2021. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[74] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In CVPR, pages 22935-22945, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[75] Guanyu Zhu, Yong Zhou, Rui Yao, and Hancheng Zhu. Cross-class bias rectification for point cloud few-shot segmentation. TMM, 25:9175-9188, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "text", + "content": "[76] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "text", + "content": "[77] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ji-aming Liu, Han Xiao, Chaoyou Fu, Hao Dong, and Peng Gao. No time to train: Empowering non-parametric networks for few-shot 3d scene segmentation. In CVPR, pages 3838–3847, 2024. 5, 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22192" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/b2000dba-b471-4537-8cba-488e6bd7eaf3_content_list.json b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/b2000dba-b471-4537-8cba-488e6bd7eaf3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..473d4659dd231bfb9a8c8b0b7143def00553fd4e --- /dev/null +++ b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/b2000dba-b471-4537-8cba-488e6bd7eaf3_content_list.json @@ -0,0 +1,1841 @@ +[ + { + "type": "text", + "text": "Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis", + "text_level": 1, + "bbox": [ + 119, + 128, + 851, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongyu Sun $^{1,2}$ , Qiuhong Ke $^{2}$ , Ming Cheng $^{2}$ , Yongcai Wang $^{1*}$ , Deying Li $^{1}$ , Chenhui Gou $^{2}$ , Jianfei Cai $^{2}$ $^{1}$ Department of Computer Science, Renmin University of China, China \n $^{2}$ Department of Data Science & AI, Monash University, Australia", + "bbox": [ + 86, + 202, + 888, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\{\\mathrm{sunhongyu,ycw,deyingli}\\} \\mathrm{ruc.edu.cn}$ {qiuhong.ke,ming.cheng, jianfei.cai}@monash.edu", + "bbox": [ + 117, + 258, + 852, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes a general solution to enable point cloud recognition models to handle distribution shifts at test time. Unlike prior methods, which rely heavily on training data (often inaccessible during online inference) and are limited to recognizing a fixed set of point cloud classes predefined during training, we explore a more practical and challenging scenario: adapting the model solely based on online test data to recognize both previously seen classes and novel, unseen classes at test time. To this end, we develop Point-Cache, a hierarchical cache model that captures essential clues of online test samples, particularly focusing on the global structure of point clouds and their local-part details. Point-Cache, which serves as a rich 3D knowledge base, is dynamically managed to prioritize the inclusion of high-quality samples. Designed as a plug-and-play module, our method can be flexibly integrated into large multimodal 3D models to support open-vocabulary point cloud recognition. Notably, our solution operates with efficiency comparable to zero-shot inference, as it is entirely training-free. Point-Cache demonstrates substantial gains across 8 challenging benchmarks and 4 representative large 3D models, highlighting its effectiveness. Code is available at https://github.com/auniquesun/Point-Cache.", + "bbox": [ + 76, + 340, + 473, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 731, + 209, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, 3D sensors such as LiDARs and RGB-D cameras have been widely adopted in robotics and electric vehicles for their ability to provide reliable 3D geometry measurements [11, 16, 20, 31, 59, 61-63, 68]. Point clouds are among the most direct data formats produced by these 3D sensors. Although remarkable progress has been made in 3D point cloud recognition, the success is primarily based on the assumption that the test data and the model train", + "bbox": [ + 75, + 756, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/390ff48f53c480ada107b8db48c5877d240abfb34ddd2ac6deb5d504dd339e62.jpg", + "image_caption": [ + "(a) ModelNet & ModelNet-C." + ], + "image_footnote": [], + "bbox": [ + 501, + 308, + 696, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7cdb9ca047fdb5bc01e75f1994eb9b424688e3037dab17f397cabbb43fdd220.jpg", + "image_caption": [ + "(b) ScanObjNN & ScanObjNN-C.", + "Figure 1. Recognition accuracy comparison on clean and corrupted point cloud datasets. The suffix -C indicates datasets with corruptions. Models experience a severe performance drop when data corruptions arise. Point-Cache effectively narrows down the performance gap between clean and corrupted data. The hardest split of ScanObjNN is used." + ], + "image_footnote": [], + "bbox": [ + 696, + 309, + 887, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing data are identically distributed [2, 9, 26, 32, 37, 39, 40, 64, 87]. However, this assumption is frequently violated in real-world scenarios due to complex geometries, as well as sensing and processing errors [43, 50, 51].", + "bbox": [ + 496, + 550, + 892, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In practice, a majority of models remain vulnerable to distribution shifts, such as out-of-distribution (OOD) samples [6, 7, 50, 69], data corruptions [43, 44, 51, 56], and more. As shown in Fig. 1, notable performance gaps (e.g., $11 + \\%$ ) occur when models are tested on clean (ModelNet [70]) versus corrupted data (ModelNet-C [43]). These issues drive researchers to enhance the robustness and generalization of point cloud recognition methods. Related efforts can be grouped into three main streams. The first focuses on training models that can transfer to target data. These models learn a shared feature space between source and target domains [10, 41, 45, 67], or apply meta-learning frameworks to handle geometry shifts by training multiple classifiers [19], or design learnable prompts to elicit general knowledge from pre-trained large 3D models [50]. However, possible shifts are unpredictable, and the training-based methods cannot account for all cases during training. As a result, these models are limited when new shifts arise at test time [75]. The second stream addresses the problem", + "bbox": [ + 496, + 613, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1263", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "through test-time adaptation strategies [35, 46, 66]. These strategies improve performance with various adaptation techniques, such as masked auto-encoding for point clouds (MATE [35]), static prototype memory (BFTT3D [66]), and diffusion models (CloudFixer [46]). Nevertheless, we find that these approaches rely heavily on training data. For instance, MATE [35] requires pre-training on the entire training set, and BFTT3D [66] employs training data to build a static and offline prototype memory to store point cloud features. This heavy dependence on training data poses challenges, particularly when such data is unavailable or when distribution shifts occur at test time. In these cases, revisiting the training data may not be helpful to adapt the model to online test data. Additionally, these test-time methods can only recognize classes seen during training, which limits their practicality in generalizable point cloud analysis.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by advancements in language-guided image recognition [22, 38, 42, 79], the third branch of works utilizes contrastive pre-training on large-scale triplets (point, text, image) to connect flexible language descriptions and enable open-vocabulary point cloud recognition [29, 72, 73, 88]. These large multimodal 3D models demonstrate promising zero-shot performance and provide a new paradigm for generalizable point cloud analysis. However, zero-shot predictions are performed independently for each test sample, with the final accuracy simply averaged across the test set. This approach does not fully exploit the distribution of test data, particularly the statistics of reliable predictions from large 3D models, which could further enhance accuracy and generalization. At this point, we pose the following question: can we take advantages of both the promising zero-shot ability of large 3D models and test-time adaptation techniques to unlock more robust and generalizable point cloud recognition?", + "bbox": [ + 75, + 335, + 472, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, to fully exploit the data distribution during test time, we develop a cache model to store critical fingerprints of test samples. These fingerprints consist of key-value pairs derived by large multimodal 3D models, where the key represents the sample feature extracted by the point encoder, and the value represents the predicted (pseudo) label from the model. This cache model is entirely constructed from online test data without accessing any training samples. Our approach is thus more practical and challenging than prior methods [10, 19, 35, 41, 45, 46, 50, 66, 67]. Our cache model is hierarchical, as it stores not only the global features of a point cloud in a global cache but also its local-part details in a local cache, which are essential for distinguishing subtle differences among classes. We also empirically demonstrate that this coarse-to-fine design enhances robustness and generalization in the wild. Additionally, our cache model is dynamic, since we need to continuously update the cache with online test data to prioritize high-quality key-value pairs. The constructed cache", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "serves as a powerful 3D knowledge base, capturing key characteristics of online test data. When a new sample arrives, we query the knowledge base, compute the affinity between the sample and cached keys, and produce a prediction by weighting the cached (pseudo) labels accordingly. Our method is training-free and flexible, requiring only a pre-trained 3D model and test data. More details will be provided in Sec. 3.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The superiority of the proposed method is supported by comprehensive comparisons with previous strong baselines across 8 challenging benchmarks, containing up to 1,156 point cloud classes. Moreover, we implement the hierarchical cache as a plug-and-play module and integrate it into representative large 3D models, such as ULIP [72], ULIP-2 [73], OpenShape [29] and Uni3D [88], to enhance test-time point cloud recognition dynamically. The proposed cache model brings notable and consistent gains over the baselines. For instance, our model yields $+6.18\\%$ and $+4.84\\%$ absolute accuracy improvements average over 7 types of corruptions in ModelNet-C [43] based on ULIP [72] and ULIP-2 [73]. On Objaverse-LVIS [7] that includes 1,156 categories, Point-Cache boosts the accuracy by $+2.10\\%$ absolute points compared to the zero-shot ULIP-2 [73]. In summary, the contributions of our paper are threefold.", + "bbox": [ + 496, + 212, + 892, + 467 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To our knowledge, Point-Cache is the first to explore test-time point cloud recognition in a more practical and challenging setting, where only online test data is available and the model needs to generalize to both known classes and unseen new classes, which is beyond the capability of prior test-time methods.", + "- We propose a hierarchical cache model to capture essential distribution clues of online test samples, creating an accurate profile for various point clouds from global to local perspectives. We implement the hierarchical cache as a plug-and-play component, enabling diverse 3D backbones to achieve enhanced test-time performance.", + "- Comprehensive experiments and consistent gains across 8 public datasets and 4 large 3D models demonstrate the efficacy of Point-Cache." + ], + "bbox": [ + 500, + 468, + 890, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 709, + 640, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test-time adaptation in 3D point cloud analysis receives relatively less attention compared to the 2D image community [5, 12, 14, 21, 23, 27, 28, 30, 34, 47, 52, 54, 60, 78, 80]. This technique has recently emerged in 3D point cloud tasks, including recognition [35, 46, 66], registration [17], upsampling [18] and object detection [65, 74, 85]. Among these, MATE [35], BFTT3D [66], and CloudFixer [46] are specifically focused on test-time point cloud recognition and are thus closely related to our work. However, as previously analyzed, these methods depend heavily on training data from the source domain to construct their pipelines,", + "bbox": [ + 496, + 734, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1264", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3ca772f4516ab57674d72aabfc6ae6c3055f7993462a8a11bfdf0e2c4ddb4a95.jpg", + "image_caption": [ + "Figure 2. The overall pipeline of Point-Cache. The zero-shot predictions $\\hat{\\mathbf{y}}_{zs}$ of large 3D models are effectively adapted by our global cache logits $\\hat{\\mathbf{y}}_g$ and local cache logits $\\hat{\\mathbf{y}}_l$ to handle the distribution shifts, enabling robust and generalizable point cloud analysis." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 885, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which may not be available or helpful to understand target data distributions at test time. Furthermore, they lack effective strategies to leverage the distribution of test samples. By contrast, Point-Cache is constructed entirely from online test data without access to any training samples, making our setting both more challenging and more practical. Additionally, our cache model, powered by large 3D models, can generalize to both known classes and unseen new categories, an advancement that prior point cloud adaptation/generalization frameworks have not achieved [10, 19, 35, 41, 45, 46, 66, 67].", + "bbox": [ + 75, + 397, + 472, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cache models that store sample features in memory for further reference and reuse have been explored in fields like language modeling [15, 24, 33, 58], image classification [3-5, 13, 21, 23, 36, 48, 55, 58, 82, 83] and point cloud recognition [53, 66, 84]. Our proposed cache model stands out from previous ones in several key aspects: (1) Our cache model is constructed based on test samples whereas previous cache models rely on a large training set for construction, as in Tip-Adapter [82], CaFo [83], PointNN [84], Point-PEFT [53] and BFTT3D [66]. (2) Our cache model is hierarchically designed to capture both global and local-part features of a point cloud object, while existing methods store only global representations [3-5, 13, 21, 48, 53, 58, 66, 82-84]. We empirically demonstrate that this coarse-to-fine design improves the capture of target data features, thereby enhancing test-time performance. (3) Our cache model is dynamically updated with online test samples, incorporating a selective mechanism to ensure high-quality key-value pairs replace less reliable ones when the cache reaches its capacity. In contrast, prior cache models, which are constructed offline using the entire training set, are static and consume a substantial amount of memory [53, 66, 84].", + "bbox": [ + 75, + 568, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Multi-modal 3D Models [25, 29, 72, 73, 76, 81, 88, 89] spark a wave of open-vocabulary point cloud understanding by pre-training on large-scale datasets [7] in recent years. For instance, ULIP-2 [73], OpenShape [29] and Uni3D [88] pre-train the 3D encoders on million-scale (point, image, text) triplets using contrastive learning. These pre-trained point cloud encoders are then transferred to downstream 3D tasks like classification and retrieval, demonstrating strong zero-shot performance that surpasses traditional 3D models [32, 39, 40, 64, 71, 87] which are typically tailored to specific datasets [1, 56, 70]. This new paradigm presents a promising direction for generalizable point cloud processing. Our work is complementary to these models: rather than developing another large 3D model, we build upon their capabilities to enhance robustness and generalization for point cloud analysis at test time. Our hierarchical cache, backed by these large 3D models, generalizes to unseen classes beyond training, addressing a significant limitation of current test-time methods for point cloud recognition [35, 46, 66].", + "bbox": [ + 496, + 397, + 893, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 719, + 591, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section elaborates on the details of Point-Cache, the overall pipeline of our approach is depicted in Fig. 2.", + "bbox": [ + 496, + 747, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary. We begin by reviewing the zero-shot inference process of existing large multimodal 3D models [29, 72, 73, 81, 88, 89]. These models use separate encoders to map inputs from a point cloud $P \\in \\mathbb{R}^{N \\times 3}$ and text $T$ (e.g., \"a point cloud object of a {class}\") to a shared feature space. Denoting the point cloud encoder as $f_{p}(\\cdot)$ and the text encoder as $f_{t}(\\cdot)$ , we can obtain the global feature $\\mathbf{e}_p^g = f_p(P) \\in \\mathbb{R}^d$ of the point cloud $P$ and the text", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1265", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "feature $\\mathbf{e}_t = f_t(T) \\in \\mathbb{R}^d$ , where $d$ is the feature dimension. To recognize the input point cloud $P$ in a zero-shot manner, these models replace “ $\\{\\mathrm{class}\\}$ ” in text $T$ with each of possible $C$ category names and compute the class probability $\\hat{\\mathbf{y}}_{zs}$ using Eq. 1.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} _ {z s} = \\{\\hat {y} _ {i} | _ {i = 1} ^ {C} \\}, \\quad \\hat {y} _ {i} = \\frac {\\exp (s i m (\\mathbf {e} _ {p} ^ {g} , \\mathbf {e} _ {t} ^ {i}) / \\tau)}{\\sum_ {j = 1} ^ {C} \\exp (s i m (\\mathbf {e} _ {p} ^ {g} , \\mathbf {e} _ {t} ^ {j}) / \\tau)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 175, + 468, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\mathbf{e}_t^i$ indicates the text encoding of the $i$ -th category. $sim(\\cdot, \\cdot)$ measures the cosine similarity between inputs, and $\\tau$ is a temperature coefficient. Finally, the point cloud $P$ is classified into the category with the highest probability, indexed by $\\hat{L} = \\arg \\max_{i} \\{\\hat{y}_i|_{i=1}^C\\}$ .", + "bbox": [ + 76, + 224, + 468, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As described in Sec. 1, our hierarchical cache stores the fingerprints of online test samples, which consists of key-value pairs. The key is the test sample feature $\\mathbf{e}_p^g$ extracted by the point encoder of a pre-trained large 3D model, and the value is the class label $\\hat{L}$ predicted by the zero-shot inference. In addition, we compute the entropy $h = -\\sum_{i=1}^{C} \\hat{y}_i \\log \\hat{y}_i$ of the zero-shot prediction $\\hat{\\mathbf{y}}_{zs}$ , as it is essential to assess the confidence with which the large 3D model produces this prediction. This entropy value is employed to filter out high-quality samples in our cache model.", + "bbox": [ + 75, + 306, + 468, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Text & Point Cloud Encoding", + "text_level": 1, + "bbox": [ + 76, + 465, + 341, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can obtain the text encoding $\\mathbf{e}_t$ of the input $T$ and the global feature $\\mathbf{e}_p^g$ of a point cloud $P$ as introduced in the preliminary. However, it is non-trivial to represent the local-part features $\\mathbf{e}_p^l$ of the point cloud $P$ due to the following challenges analyzed in Fig. 3:", + "bbox": [ + 75, + 488, + 468, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "01 Part Feature Capture", + "text_level": 1, + "bbox": [ + 109, + 580, + 264, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Difficult to define parts precisely and get their features for various 3D objects of different classes", + "bbox": [ + 119, + 602, + 256, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "02. Part Feature Storage", + "text_level": 1, + "bbox": [ + 279, + 580, + 431, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Expensive to store multiple part features instead of a global feature for a single point cloud", + "bbox": [ + 287, + 601, + 419, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3. Challenges in encoding part feature for various 3D objects of different classes: Part Feature Capture & Storage.", + "bbox": [ + 75, + 675, + 468, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address these challenges, we made a careful analysis and developed effective solutions, seeing Fig. 9 in the Supplementary for intuitive illustration. (1) A straightforward approach might involve using an additional pretrained segmentation network to segment point cloud parts, but this would be computationally prohibitive during online inference. Instead, we propose a fuzzy definition of parts. Specifically, since the architecture of the point encoder of large 3D models is based on Transformer [8, 57, 77] and the input is a sequence of point patches, we avoid strict part definitions according to physical semantics (e.g., head, body, wing, and tail in an airplane), and instead treat the output", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "point patches from the last encoder layer as local part representations. (2) A second challenge is that a long point patch sequence can become resource-intensive; for example, the point encoder in ULIP-2 [73] generates 512 point patches per object, which would require substantial memory. To mitigate this problem, we propose to summarize these point patches into $m$ centers $\\mathbf{e}_p^l \\in \\mathbb{R}^{m \\times d}$ using K-Means clustering ( $m < 10$ ), reducing memory usage by two orders of magnitude compared to storing all patches.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Construction & Update of Hierarchical Cache", + "text_level": 1, + "bbox": [ + 498, + 237, + 888, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To thoroughly explore the distribution of test data, we develop a hierarchical cache to record the fingerprints of online test samples. This hierarchical cache includes both a global cache, which stores overall information, and a local cache, which captures detailed part-level information for each point cloud. The coarse-to-fine structure enables more precise profiling of test data and distinguishes our approach from previous cache models used in image recognition [3-5, 13, 21, 23, 36, 48, 55, 58, 82, 83] and point cloud analysis [53, 66, 84]. The complete construction process of hierarchical cache is outlined in Alg. 1 of the Supplementary.", + "bbox": [ + 496, + 260, + 890, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global cache is represented by $\\mathbf{C}_g = (\\mathbf{E}_g, \\hat{\\mathbf{L}}_g, \\mathbf{h}_g)$ , storing (feature, label, entropy) triplets for a set of point cloud objects. To manage storage efficiently, we design a selection mechanism to identify high-quality test samples to place into the cache.", + "bbox": [ + 496, + 426, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Construction: We set an upper bound $K$ on the number of samples per class in the cache, allowing at most $K$ fingerprints (shots) can be placed in each category. Initially, the global cache is empty. When a new test sample $P$ arrives, we generate its fingerprint $(\\mathbf{e}_p^g, \\hat{L}, h)$ and assign it to class $\\hat{L}$ .", + "bbox": [ + 496, + 503, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Update: With the number of online samples grows, we check whether the number of shots for class $\\hat{L}$ has hit the upper bound. If it is not the case, the fingerprint of current sample is appended to class $\\hat{L}$ . Otherwise, we locate the fingerprint with highest entropy within class $\\hat{L}$ and compare it to the entropy of current test sample. If the current test sample has a lower entropy, we replace the highest-entropy fingerprint with the new one; otherwise, the current sample is skipped. When the cache is full, $\\mathbf{E}_g$ has dimensions $CK\\times d$ , $\\hat{\\mathbf{L}}_g$ has dimensions $CK\\times 1$ , and $\\mathbf{h}_g$ has dimensions $CK\\times 1$ .", + "bbox": [ + 496, + 597, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local cache. Although the global feature of a point cloud captures expressive information, it lacks the local details of the 3D object, which are critical for point cloud recognition, especially when shapes share an overall appearance but differ in local parts. The local cache addresses this limitation by storing part-level information. We denote the local cache as $\\mathbf{C}_l = (\\mathbf{E}_l,\\hat{\\mathbf{L}}_l)$ , which store the fingerprints of local parts of point cloud objects. Here $\\mathbf{E}_l$ represents part features of the test samples and $\\hat{\\mathbf{L}}_l$ indicates their labels.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1266", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accordingly, we define the local-part fingerprint of a test sample $P$ as $(\\mathbf{e}_p^l,\\hat{L})$ , where $\\mathbf{e}_p^l\\in \\mathbb{R}^{m\\times d}$ corresponds to local-part features and $\\hat{L}$ is the class label of $P$ .", + "bbox": [ + 76, + 90, + 468, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Construction & Update: Note that the sample selective mechanism of the local cache follows the same rules as the global cache, meaning the local-part fingerprint of a test sample is only stored if it qualifies for inclusion in the global cache. So there is no need to record the entropy $h$ again in the local cache. When the cache is full, $\\mathbf{E}_l$ has dimensions $(m \\cdot CK) \\times d$ and $\\hat{\\mathbf{L}}_l$ has dimensions $CK \\times 1$ .", + "bbox": [ + 76, + 138, + 467, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Test-time Adaptation by Hierarchical Cache", + "text_level": 1, + "bbox": [ + 76, + 253, + 452, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the hierarchical cache serving as a rich knowledge base of online point cloud data, we can adapt the zero-shot predictions of large 3D models for new samples, effectively addressing distribution shifts that occur at test time. The adaptation follows a coarse-to-fine process as below.", + "bbox": [ + 76, + 277, + 467, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test-time adaptation by global cache. During adaptation, a new sample $Q$ is treated as a query to retrieve knowledge from the global cache, which we use to produce the adaptation logits. Specifically, the retrieval process is simulated by calculating the affinity $A_{g}$ between the query and cached features $\\mathbf{E}_{g}$ . This allows the new sample to establish connections with cached high-quality samples, providing it with a contextual understanding of the online test data. We then generate the adaptation logits $\\hat{\\mathbf{y}}_{g}$ by weighed ensembling of cache labels $\\hat{\\mathbf{L}}_{g}$ , as shown in Eq. 2.", + "bbox": [ + 75, + 353, + 467, + 507 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {g} = \\exp \\left(- \\beta_ {g} \\left(1 - \\mathbf {e} _ {q} ^ {g} \\mathbf {E} _ {g} ^ {\\top}\\right)\\right), \\hat {\\mathbf {L}} _ {g} = \\mathrm {O H} (\\hat {\\mathbf {L}} _ {g}), \\hat {\\mathbf {y}} _ {g} = A _ {g} \\hat {\\mathbf {L}} _ {g} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 517, + 467, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{e}_q^g\\in \\mathbb{R}^{1\\times d}$ represents the global feature of sample $Q$ and $\\mathbf{E}_g\\in \\mathbb{R}^{CK\\times d}$ denotes the cached keys. The 'OH' operation transforms $\\hat{\\mathbf{L}}_g$ into a one-hot encoding. The term $\\mathbf{e}_q^g\\mathbf{E}_g^\\top \\in \\mathbb{R}^{1\\times CK}$ measures the cosine similarity between the query and all keys, with $\\beta_{g}$ acting as a hyper-parameter to modulate the sharpness and the exponential operation ensuring similarity values fall within (0, 1]. The affinity $A_{g}\\in \\mathbb{R}^{1\\times CK}$ , which grasps the relationship between the query and cached keys, is utilized to ensemble the transformed one-hot labels $\\hat{\\mathbf{L}}_g\\in \\mathbb{R}^{CK\\times C}$ , yielding global cache adaptation logits $\\hat{\\mathbf{y}}_g\\in \\mathbb{R}^{1\\times C}$ .", + "bbox": [ + 76, + 550, + 467, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test-time adaptation by local cache. While global cache adaptation is effective, it may not handle fine-grained distribution shifts at test time. To achieve a more precise adjustment, we exploit local part knowledge. The local-part features $\\mathbf{e}_q^l$ of the point cloud $Q$ serve as queries to the local cache. During retrieval, we compute the affinity $A_{l}$ between $\\mathbf{e}_q^l$ and all cached part features $\\mathbf{E}_l$ . The affinity $A_{l}$ is then used to weight the cache labels $\\hat{\\mathbf{L}}_l$ , producing local cache adaptation logits $\\hat{\\mathbf{y}}_l$ as in Eq. 3.", + "bbox": [ + 75, + 719, + 467, + 858 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {l} = \\exp \\left(- \\beta_ {l} \\left(1 - \\mathbf {e} _ {q} ^ {l} \\mathbf {E} _ {l} ^ {\\top}\\right)\\right), \\hat {\\mathbf {L}} _ {l} = \\mathrm {O H} (\\hat {\\mathbf {L}} _ {l}), \\hat {\\mathbf {y}} _ {l} = \\phi \\left(A _ {l} \\hat {\\mathbf {L}} _ {l}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 867, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{e}_q^l\\in \\mathbb{R}^{m\\times d}$ , $\\mathbf{E}_l\\in \\mathbb{R}^{(m\\cdot CK)\\times d}$ , and $\\hat{\\mathbf{L}}_l\\in \\mathbb{R}^{(m\\cdot CK)\\times C}$ . The function $\\phi (\\cdot)$ is an average pooling operation over $m$ parts to produce class logits. The term $e_q^l\\mathbf{E}_l^\\top \\in \\mathbb{R}^{m\\times (m\\cdot CK)}$ measures the cosine similarity between the part features of sample $Q$ and all part features in the cache. Likewise, the similarity is modulated by the coefficient $\\beta_{l}$ and scaled with the exponential function, resulting in the affinity $A_{l}\\in \\mathbb{R}^{m\\times (m\\cdot CK)}$ . The term $A_{l}\\hat{\\mathbf{L}}_{l}\\in \\mathbb{R}^{m\\times C}$ represents the class logits for all $m$ parts, which are transformed by a pooling operation $\\phi (\\cdot)$ to derive the final local cache adaptation logits $\\hat{\\mathbf{y}}_l\\in \\mathbb{R}^{1\\times C}$ .", + "bbox": [ + 496, + 89, + 890, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall prediction combines three components as shown in Eq. 4. The zero-shot prediction $\\hat{\\mathbf{y}}_{zs}$ of large multimodal 3D model is effectively adapted using our global cache prediction $\\hat{\\mathbf{y}}_g$ and local cache prediction $\\hat{\\mathbf{y}}_l$ , balanced by the coefficients $\\alpha_{q}$ and $\\alpha_{l}$ .", + "bbox": [ + 496, + 260, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} = \\hat {\\mathbf {y}} _ {z s} + \\alpha_ {g} \\hat {\\mathbf {y}} _ {g} + \\alpha_ {l} \\hat {\\mathbf {y}} _ {l} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 349, + 890, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 381, + 633, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 498, + 407, + 705, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. To evaluate the robustness of point cloud recognition, we choose four public datasets that contain various point cloud corruptions: ModelNet-C [43] and three variants of ScanObjectNN-C [56]. ModelNet-C [43] introduces 7 atomic corruptions, including adding global outliers, adding local outliers, dropping global structure, dropping local parts, rotation, scaling, and jittering. Other corruptions can be derived from these atomic corruptions, as discussed in [43]. We apply these atomic corruptions to the three variants of ScanObjectNN [56] and produce their corrupted versions. To assess generalization on unseen new data, we test the proposed method on four challenging benchmarks: OmniObject3D [69] (216 classes), ObjaverseLVIS [7] (1,156 classes), the hardest variant of ScanObjectNN [56] and widely used ModelNet40 [70]. Recognition accuracy (\\%) is the primary evaluation metric.", + "bbox": [ + 496, + 431, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We choose ULIP [72], ULIP-2 [73], OpenShape [29] and Uni3D [88] as the large multimodal 3D models for experiments. Their pre-trained weights are loaded and then frozen. Global and local-part features are generated by modifying the forward pass function without altering the model architecture. The maximum number of samples per class in the cache, $K$ , is set to 3, and each point cloud is clustered into $m = 3$ local parts. Both the balance factors $\\alpha_{g}$ and $\\alpha_{l}$ are set to 4.0 while the coefficients $\\beta_{g}$ and $\\beta_{l}$ are set to 3.0. We will examine the design choices in the ablation studies. To optimize memory and computation, we convert the input point clouds and model parameters into 'dtype' fp16 in Pytorch. All experiments are conducted on two 4090 GPUs. Additional details can be found in the Supplementary.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1267", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/861996e93f967c92c33c077126398bd1ebe9e4550bae37f844847473bfb1a673.jpg", + "table_caption": [ + "Table 1. Comparison of recognition accuracy on ModelNet-C that contains 7 types of corruptions. Each clean point cloud has 1024 points and the corruption severity level is 2. The last column is the average over 7 corruption types. The best results are in bold and the second best are underlined. The setting applies to the following tables unless otherwise specified." + ], + "table_footnote": [], + "table_body": "
MethodClean Data ModelNetCorruption TypeAvg.
Add GlobalAdd LocalDrop GlobalDrop LocalRotateScaleJitter
ULIP [72]56.1633.5543.9254.7050.8955.2750.2044.0847.52
+Global Cache(Ours)62.1245.7947.9856.8553.8960.2554.3448.9152.56
+Hierarchical Cache(Ours)64.2246.1547.8559.1656.0061.4755.3549.9253.70
ULIP-2 [73]71.2365.1554.6268.7657.9870.3067.1021.7657.95
+Global Cache(Ours)73.9567.0259.3271.3561.5972.3768.4028.2061.18
+Hierarchical Cache(Ours)74.5368.1161.2673.2263.6573.3470.4229.5062.79
O-Shape [29]84.5671.6467.7981.5673.5882.0178.4859.3673.49
+Global Cache(Ours)84.5274.7272.7782.4175.1283.1878.9367.9176.43
+Hierarchical Cache(Ours)84.0474.8473.7082.2176.2682.6678.1268.3576.59
Uni3D [88]81.8172.4556.3668.1567.1879.9475.3656.2467.95
+Global Cache(Ours)83.1476.1366.4971.4369.8181.5275.8561.4371.81
+Hierarchical Cache(Ours)83.8777.5171.1572.1670.7581.7777.3162.5273.31
", + "bbox": [ + 94, + 142, + 875, + 345 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2c300dc0bcb188624dad860d7c356578c2aa7027efc3c63aa11240f63be5ac6b.jpg", + "table_caption": [ + "Table 2. Comparison of recognition accuracy across a suite of datasets. S-PB_RS_T50 is the hardest split of ScanObjectNN. OLVIS: Objverse-LVIS. Omni3D: OmniObject3D. The number under each dataset indicates the number of points (pts) for each object. In Omni3D, each point cloud can be represented by a different number of points. The last column is the average accuracy on these datasets." + ], + "table_footnote": [], + "table_body": "
MethodModelNet40 (10000 pts)S-PB_RS_T50 (2048 pts)O-LVIS (10000 pts)Omni3DAvg.
(1024 pts4096 pts16384 pts)
ULIP [72]58.7546.446.248.397.757.2822.48
+Global Cache(Ours)61.2250.217.0210.009.368.4324.37
+Hierarchical Cache(Ours)62.9351.807.0210.479.758.9025.15
ULIP-2 [73]72.9747.1330.2626.3629.2026.5838.75
+Global Cache(Ours)74.5151.7032.6528.5131.1028.5341.17
+Hierarchical Cache(Ours)75.5354.9832.3629.3731.2429.4442.15
Uni3D [88]88.4165.1955.4231.5241.9841.8654.09
+Global Cache(Ours)88.8668.5153.3634.9745.1345.1956.00
+Hierarchical Cache(Ours)89.1868.2455.1935.8245.6045.8956.65
", + "bbox": [ + 138, + 411, + 831, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Test-time Robustness and Generalization", + "text_level": 1, + "bbox": [ + 76, + 597, + 426, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Robustness. The robustness analysis of our training-free test-time hierarchical caches takes place on 4 point cloud corruption datasets [43, 56]. As observed in Tab. 1, the zero-shot accuracy of various large 3D models on ModelNet-C is considerably boosted by our global cache model, with increases such as $+5.04\\%$ for ULIP [72], $+3.23\\%$ for ULIP-2 [73], $+2.94\\%$ for OpenShape [29] and $+3.86\\%$ for Uni3D [88], averaged across 7 corruption types. By combining global and local cache together, our hierarchical cache model attains further improvements, highlighting the benefits of incorporating a local cache. Notably, these gains extend beyond corrupted data, as our hierarchical cache model also improves recognition accuracy on the clean ModelNet dataset, e.g., $+8.06\\%$ absolute increase over ULIP's zero-shot performance. An exception is OpenShape, which maintains a slightly higher zero-shot accuracy on clean data than our cache model. However, when considering the 7 corruption types, our global and hi", + "bbox": [ + 75, + 628, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "erarchical cache models outperform OpenShape by $2.94\\%$ and $3.10\\%$ , respectively, demonstrating improved robustness against data corruption at test time. Consistent improvements are also showcased on the variants of ScanObjectNN, shown in Tab. 5, 6 and 7 in the Supplementary.", + "bbox": [ + 496, + 598, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalization. We evaluated the generalization capability of Point-Cache on four representative benchmarks that were unseen during the pre-training of the large 3D models. These benchmarks generally do not contain data corruptions. While ModelNet40 [70] and ScanObjectNN [56] contain a limited number of classes, OmniObject3D [69] and Objverse-LVIS [7] represent a broader range of 3D concepts, covering 216 and 1,156 classes, respectively. As shown in Tab. 2, Point-Cache consistently enhances the recognition accuracy across datasets and 3D backbones. For instance, our hierarchical cache built on ULIP and ULIP-2 yields absolute improvements of $6.31\\%$ and $5.51\\%$ , respectively, average over 6 scores. On ScanObjectNN, our hierarchical cache significantly raises the zero-shot prediction of ULIP and ULIP-2, from $27.20\\%$ to $51.80\\%$ for ULIP", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1268", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and from $36.22\\%$ to $54.98\\%$ for ULIP-2. On the challenging Objaverse-LVIS (O-LVIS), our global cache with ULIP-2 successfully realizes a $2.39\\%$ absolute gains in zero-shot accuracy across 1,156 classes, a notable achievement. Furthermore, our cache model enables Uni3D to attain an $89.18\\%$ accuracy on ModelNet40 without any training, effectively matching the fully-supervised PointNet $(89.20\\%)$ .", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Memory Usage and Throughput", + "text_level": 1, + "bbox": [ + 76, + 208, + 362, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Memory. We use Uni3D [88] as the baseline and adapt its zero-shot predictions with Point-Cache. The experiments are conducted on 3 datasets as shown in Tab. 3. The batch size is set to 1 because we need to update the cache and adapt the class logits according to every single sample. The results indicate our global cache and hierarchical cache consume similar or slightly higher memory than the baseline on each dataset. Furthermore, as the number of classes increases rapidly, the memory usage grows at a slower rate. This is primarily because the memory consumption is dominated by the large number of parameters in Uni3D (e.g., 1,016.5M). In contrast, the hierarchical cache built on OLVIS uses only approximately 7.1M parameters, which is negligible compared to Uni3D. Detailed calculations are provided in the Supplementary.", + "bbox": [ + 75, + 232, + 473, + 460 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b34b90bb25a1b4218faeb495a08ac3abe5d1c5d7b7a66c9a0606f4262c60d7b9.jpg", + "table_caption": [ + "Table 3. Comparison of memory usage (MB). The batch size is set to 1 and the used device is an RTX 4090. The number under each dataset indicates #Classes. Hierar: Hierarchical Cache." + ], + "table_footnote": [], + "table_body": "
MethodModelNet-C (40)Omni3D (216)O-LVIS (1156)#Params (M)
Uni3D5,0625,0625,0621,016.5
+Global(Ours)5,0625,0645,0701,016.5
+Hierar(Ours)5,0645,0685,0901,016.5
", + "bbox": [ + 80, + 527, + 464, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Throughput. We evaluate the throughput of Point-Cache on ModelNet40, where throughput is measured as the number of test samples processed per second (abbreviated as $t / s$ ). The results are reported in Tab. 4. Point-Cache runs slightly slower than zero-shot inference due to the additional overhead introduced by cache updates, logits computation, etc. However, the throughput degradation is minimal, e.g., a $0.04 \\, t / s$ drop for OpenShape with the global cache, and a $0.07 \\, t / s$ drop for ULIP-2 with the hierarchical cache. These results prove that Point-Cache brings remarkable accuracy improvements with negligible computational cost.", + "bbox": [ + 75, + 621, + 468, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 814, + 230, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use Uni3D with our hierarchical cache and 4 public datasets for ablation analysis, as shown in Fig. 4.", + "bbox": [ + 75, + 838, + 468, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shot size per class in the global cache. The upper bound $K$ on the number of samples per class is a critical variable", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/36a3695144c02f118ba581585afcd6b81ba2db1684724c35f586634e6c4d2d95.jpg", + "table_caption": [ + "Table 4. Comparison of throughput $(t / s)$ for different models on ModelNet40. The batch size is set to 1 and the device is a 4090 GPU. The results are averaged over all test samples." + ], + "table_footnote": [], + "table_body": "
MethodZero-shot+Global Cache+Hierar Cache
ULIP11.2511.2111.18
ULIP-211.1811.1411.11
OpenShape9.809.769.74
Uni3D9.789.759.73
", + "bbox": [ + 514, + 142, + 874, + 219 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as it affects the total size of the hierarchical cache. We analyze its impact on the test-time recognition accuracy on 4 datasets, shown in Fig. 4a. As $K$ increases, accuracy trend differ across datasets; for example, accuracy increases on Omni3D but decreases on S-PB_T50_RS. We adopt $K = 3$ as it produces the best average accuracy (63.98%). The results indicate that our cache model does not require a large number of samples to achieve strong performance.", + "bbox": [ + 496, + 246, + 890, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of local parts per 3D object. To optimize memory and computation, we cluster hundreds of point patches of a 3D object into $m$ parts. We investigate the effect of this variable on test-time performance, as exhibited in Fig. 4b. For three out of the four datasets, namely S-PB_T50_RS, Omni3D, and O-LVIS, accuracy generally decreases as the number of parts increases, suggesting that a point cloud does not need to be represented with an excessive number of parts. We choose $m = 3$ by default, as this setting reaches the best average accuracy across the 4 datasets.", + "bbox": [ + 496, + 367, + 892, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d50fac6fb5374b8f54c7e8633e362ea56581ba805c8dfd380cb4c885ca85db3.jpg", + "image_caption": [ + "(a) shot size $K$" + ], + "image_footnote": [], + "bbox": [ + 501, + 532, + 602, + 585 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ece0fcdfedac25c6712cece2c89841ee3708073de6ac2b219f6334ed8e01072.jpg", + "image_caption": [ + "(b) $m$ parts", + "Figure 4. Ablation studies on the hyper-parameters in the cache design, including the shot size $K$ per class, the number of parts $m$ per object, the balance factors in the final prediction logits and the sharpness coefficients in affinity computation." + ], + "image_footnote": [], + "bbox": [ + 604, + 532, + 696, + 585 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6109cecf29b549bc36eb9a36123cd2f26d6ba1c536cb083225f17b3ad98efae0.jpg", + "image_caption": [ + "(c) $\\alpha_{g}$ and $\\alpha_{l}$" + ], + "image_footnote": [], + "bbox": [ + 700, + 532, + 794, + 585 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4edbd5e95bdef17bf6a07aa570830d92308e27d9fde2c24acc8f84316addb291.jpg", + "image_caption": [ + "(d) $\\beta_{g}$ and $\\beta_{l}$" + ], + "image_footnote": [], + "bbox": [ + 797, + 532, + 888, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The balance factors. The factors $\\alpha_{g}$ and $\\alpha_{l}$ control the relative weights of the global cache prediction $\\hat{\\mathbf{p}}_g$ and local cache prediction $\\hat{\\mathbf{p}}_l$ . To simplify experimentation, we set $\\alpha_{g} = \\alpha_{l}$ in the ablation study. Candidate values are $\\{0.5, 1.0, 2.0, 3.0, 4.0, 5.0\\}$ . As shown in Fig. 4c, increasing these factors improves performance on S-PB_T50_RS and Omni3D, while it has minimal effect on ModelNet40 and O-LVIS. We set the factors to 4.0, which yields the optimal average accuracy across all four datasets.", + "bbox": [ + 496, + 686, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The sharpness coefficients. The coefficients $\\beta_{g}$ and $\\beta_{l}$ are used to modulate the sharpness of the query-key affinity. Here we analyze their impact on test-time accuracy on 4 benchmarks. Likewise, we let $\\alpha_{g} = \\alpha_{l}$ and the candidate values are $\\{0.5,1.0,3.0,5.0,7.0,9.0\\}$ . As Fig. 4d displays,", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1269", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "accuracy initially improves slightly with increasing sharpness but then decreases as sharpness becomes too high. We set the sharpness to 3.0 that yields the optimal result average over 4 datasets.", + "bbox": [ + 75, + 90, + 472, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 76, + 162, + 215, + 178 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Online Inference. We visualize the online inference process of Point-Cache using different large 3D models and compare three model variants: zero-shot prediction, the model with our global cache, and the model with our hierarchical cache, referring to Fig. 5. During inference, we calculate the average recognition accuracy over accumulated samples. Due to the small number of samples at the initial stage, the average accuracy may fluctuate drastically, so we discard the statistics for the first 5 samples. Overall, methods employing the hierarchical cache consistently outperform other variants by a clear margin across various large 3D models and datasets.", + "bbox": [ + 75, + 186, + 472, + 368 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/440b50728ed305619b42439be57f9c1a5148c3761d22b4f62467eca245965696.jpg", + "image_caption": [ + "(a) ULIP." + ], + "image_footnote": [], + "bbox": [ + 78, + 381, + 176, + 434 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d2a831c5e67b2bd1b433629862998a689f8b2c507112e6e2157973f98e03fe8.jpg", + "image_caption": [ + "(b) ULIP-2." + ], + "image_footnote": [], + "bbox": [ + 181, + 381, + 272, + 434 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b2f6617e09c641732dc4e1f4053dcdd9997d714bcd920fc6cf63321534fc20b0.jpg", + "image_caption": [ + "(c) OpenShape." + ], + "image_footnote": [], + "bbox": [ + 274, + 381, + 370, + 434 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33acc4474aff6e357a76eaa0ebbec68112e2f9519cc2095162ad15676ab68582.jpg", + "image_caption": [ + "(d) Uni3D.", + "Figure 5. The average recognition accuracy of accumulated samples during online inference. The curve changes significantly in the initial stage due to the small number of samples. Models with our global and hierarchical cache receive perceptible performance gains." + ], + "image_footnote": [], + "bbox": [ + 374, + 381, + 467, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Entropy and Accuracy. Fig. 6 illustrates the relationship between entropy and accuracy for different models on various datasets. During online inference, we first compute the entropy of top 5 predictions for each test sample and then average the entropy across all accumulated samples. Similarly, the corresponding average accuracy can be calculated. The results reveal that Point-Cache effectively reduces the uncertainty of the zero-shot predictions made by large 3D models, enhances prediction confidence, and consequently lifts recognition accuracy. Additionally, we observe that the hierarchical cache model is more effective in reducing entropy compared to the global cache model, further underscoring the importance of incorporating the local cache.", + "bbox": [ + 75, + 551, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Analysis. Fig. 7 presents the step-by-step adaptation process of Point-Cache. The first three rows demonstrate cases where the global cache model attempts to adjust the zero-shot predictions of various large 3D models but fails. Instead, the hierarchical cache model successfully refines the class logits based on the global cache. In the last row, our global cache model corrects the zero-shot prediction in the first step, and the hierarchical cache model subsequently retains the top class while increasing its confidence. The examples showcased here are not cherry-picked.", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dfae67d8bc2bd2ef47512d2147e69fda9f45bc3e08f427b575c2cad011c7e3ed.jpg", + "image_caption": [ + "(a) Group One." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 705, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea4a105ba81065854a08090615625325c57c373aa7f6c779134a0360484f40f6.jpg", + "image_caption": [ + "(b) Group Two.", + "Figure 6. Relationship between entropy and accuracy. Models with our global and hierarchical cache consistently exhibit lower entropy and higher accuracy across various benchmarks." + ], + "image_footnote": [], + "bbox": [ + 707, + 90, + 906, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional quantitative and qualitative results can be found in the Supplementary.", + "bbox": [ + 498, + 294, + 890, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f7c4c6c897456a02f83d8d4c42b4b202a92d64aaf3342f89f726e428386c7e2b.jpg", + "image_caption": [ + "(a) Ground truth." + ], + "image_footnote": [], + "bbox": [ + 506, + 339, + 586, + 391 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e171fc5b94492d1120b78cb77f7a4d235b3f748196acddf9af6f68f94187f660.jpg", + "image_caption": [ + "(b) ULIP-2." + ], + "image_footnote": [], + "bbox": [ + 604, + 338, + 702, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/541d690dabe490f921f4b8ff4d2de598a796c28ae3687f753cbbe4c006bf82af.jpg", + "image_caption": [ + "(c) ULIP-2+GC." + ], + "image_footnote": [], + "bbox": [ + 707, + 338, + 803, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f474d2b4bc06b1aabcd1b45f58def868e61af1df741c04e8f68da5233d5d9be.jpg", + "image_caption": [ + "(d) ULIP-2+HC." + ], + "image_footnote": [], + "bbox": [ + 808, + 338, + 906, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ec08d92dc0fbdc5f08019b6cbc0ecc7299b6af2faf5d5172371f405ae327bfa.jpg", + "image_caption": [ + "(e) Ground truth." + ], + "image_footnote": [], + "bbox": [ + 506, + 414, + 568, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d2815006760c86947ebb23a96c4c4073489f681b1cfac8826049d37d72af20a.jpg", + "image_caption": [ + "(f) ULIP." + ], + "image_footnote": [], + "bbox": [ + 606, + 414, + 702, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cda03881c1794279a3e0c44c51b6159973f9d3e094f6420e2af978771056e0c8.jpg", + "image_caption": [ + "(g) $\\mathbf{ULIP} + \\mathbf{GC}$" + ], + "image_footnote": [], + "bbox": [ + 707, + 414, + 805, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b8edd9672bad0310a6a5fe8be5fbb81684266b27d00e682112af604761596e1.jpg", + "image_caption": [ + "(h) ULIP+HC." + ], + "image_footnote": [], + "bbox": [ + 808, + 414, + 906, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2fba44e82830eae34d45912ee571c2ec993c1bd55ec74a93e2839fc6b1dfcf1.jpg", + "image_caption": [ + "(i) Ground truth." + ], + "image_footnote": [], + "bbox": [ + 506, + 491, + 583, + 536 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c913cbec7698e2cfed8cca4bd072891d98c4e2b3b728e250a60c65b5159ce646.jpg", + "image_caption": [ + "(j) O-Shape." + ], + "image_footnote": [], + "bbox": [ + 606, + 489, + 702, + 539 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c2af08b62cb0cf90275ddfa836cae7097e71a2c9a4a0f994d43789bc27ddbd4.jpg", + "image_caption": [ + "(k) O-Shape+GC." + ], + "image_footnote": [], + "bbox": [ + 707, + 489, + 805, + 539 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94e02d96d5e670abf5f9bedb16766283f106932a0a94cb1021941ce1e43413ab.jpg", + "image_caption": [ + "(1) O-Shape+HC." + ], + "image_footnote": [], + "bbox": [ + 808, + 489, + 906, + 539 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7041b5e4a0499d37f542d31007dbf6b1c0b7e1979ae0ea9a238840a5693762e0.jpg", + "image_caption": [ + "(m) Ground truth.", + "Figure 7. The zero-shot predictions before and after adaptation by Point-Cache. GC: global cache. HC: hierarchical cache." + ], + "image_footnote": [], + "bbox": [ + 506, + 560, + 578, + 604 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cde187005839dff345c5eb5eceaa48900ed362a6b62a315c87e7ffaa12cd6b31.jpg", + "image_caption": [ + "(n) Uni3D." + ], + "image_footnote": [], + "bbox": [ + 606, + 559, + 702, + 604 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e3ba34ddb4567227316d2fc8e8387cbc2aaa2883f031822f0c504a47da74daa.jpg", + "image_caption": [ + "o) $\\mathrm{Uni3D + GC}$" + ], + "image_footnote": [], + "bbox": [ + 707, + 559, + 805, + 604 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf042d04e709a9b9597c9c17b4a5e92114c5e44f426cc4a0969ce5953e8708fa.jpg", + "image_caption": [ + "(p) $\\mathrm{Uni3D + HC}$" + ], + "image_footnote": [], + "bbox": [ + 808, + 559, + 906, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 694, + 617, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper explores test-time point cloud recognition in a practical and challenging setting, where only test data and the pre-trained models are available. We develop PointCache, a powerful 3D knowledge base that enhances robustness and generalization at test time. It records the fingerprints of online test samples through a global and a local cache, creating a more accurate profile of 3D data. Designed as a plug-and-play module, Point-Cache can be seamlessly integrated into various large 3D models without compromising efficiency, enabling open-vocabulary point cloud recognition for both known and unseen classes, effectively addressing limitations of prior test-time methods.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1270", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 90, + 236, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank all anonymous reviewers and area chairs for their time and valuable feedback. This work was partially supported by China Scholarship Council (CSC) under the Grant No. 202306360147 and Australian Research Council funding under the Grant No. DE250100030. Dr. Qiuhong Ke is partially supported by the Australian Research Council funding Discovery Early Career Researcher Award (DE250100030). Dr. Deying Li was partially supported by the National Natural Science Foundation of China under the Grant No. 12071478. Dr. Yongcai Wang was partially supported by the National Natural Science Foundation of China under the Grant No. 61972404, the Public Computing Cloud at Renmin University of China, and the Blockchain Lab, School of Information, Renmin University of China.", + "bbox": [ + 75, + 114, + 472, + 327 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 340, + 174, + 356 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. CoRR, abs/1512.03012, 2015. 3", + "[2] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. PointGPT: Auto-regressively generative pre-training from point clouds. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1", + "[3] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations, 2019. 3, 4", + "[4] Yinbo Chen, Zhuang Liu, Huijuan Xu, Trevor Darrell, and Xiaolong Wang. Meta-baseline: Exploring simple meta-learning for few-shot learning. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9042-9051, 2021.", + "[5] Sungha Choi, Seunghan Yang, Seokeon Choi, and Sungrack Yun. Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. In European Conference on Computer Vision, pages 440–458. Springer, 2022. 2, 3, 4", + "[6] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. In Advances in Neural Information Processing Systems, pages 35799-35813. Curran Associates, Inc., 2023. 1", + "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 1, 2, 3, 5, 6" + ], + "bbox": [ + 86, + 364, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4", + "[9] Lunhao Duan, Shanshan Zhao, Nan Xue, Mingming Gong, Gui-Song Xia, and Dacheng Tao. Condaformer: Disassembled transformer with local structure enhancement for 3d point cloud understanding. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1", + "[10] Hehe Fan, Xiaojun Chang, Wanyue Zhang, Yi Cheng, Ying Sun, and Mohan Kankanhalli. Self-supervised global-local structure modeling for point cloud domain adaptation with reliable voted pseudo labels. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6367-6376, 2022. 1, 2, 3", + "[11] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and ZhaoXiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2918-2927, 2021. 1", + "[12] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2704-2714, 2023. 2", + "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 3, 4", + "[14] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In Advances in Neural Information Processing Systems, pages 29374-29385. Curran Associates, Inc., 2022. 2", + "[15] Edouard Grave, Moustapha M Cisse, and Armand Joulin. Unbounded cache model for online language modeling with open vocabulary. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3", + "[16] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 1", + "[17] Ahmed Hatem, Yiming Qian, and Yang Wang. Point-tta: Test-time adaptation for point cloud registration using multitask meta-auxiliary learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16494-16504, 2023. 2", + "[18] Ahmed Hatem, Yiming Qian, and Yang Wang. Test-time adaptation for point cloud upsampling using meta-learning. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1284–1291, 2023. 2", + "[19] Chao Huang, Zhangjie Cao, Yunbo Wang, Jianmin Wang, and Mingsheng Long. Metasets: Meta-learning on point sets for generalizable representations. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8859-8868, 2021. 1, 2, 3, 4" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1271", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[20] Zhe Huang, Shuo Wang, Yongcai Wang, Wanting Li, Deying Li, and Lei Wang. Roco: Robust cooperative perception by iterative object matching and pose adjustment. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7833-7842, 2024. 1", + "[21] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In Advances in Neural Information Processing Systems, pages 2427–2440. Curran Associates, Inc., 2021. 2, 3, 4", + "[22] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2", + "[23] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14162-14171, 2024. 2, 3, 4", + "[24] Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models. In International Conference on Learning Representations, 2020. 3", + "[25] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[26] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 1", + "[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? Source hypothesis transfer for unsupervised domain adaptation. In Proceedings of the 37th International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2", + "[28] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022. 2", + "[29] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In Advances in Neural Information Processing Systems, pages 44860-44879. Curran Associates, Inc., 2023. 2, 3, 5, 6, 1, 4", + "[30] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In Advances in Neural Information Processing Systems, pages 21808-21820. Curran Associates, Inc., 2021. 2", + "[31] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d an" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "notation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1190-1199, 2023. 1", + "[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In International Conference on Learning Representations, 2022. 1, 3", + "[33] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. In International Conference on Learning Representations, 2017. 3", + "[34] M. Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[35] M. Jehanzeb Mirza, Inkyu Shin, Wei Lin, Andreas Schriebl, Kunyang Sun, Jaesung Choe, Mateusz Kozinski, Horst Possegger, In So Kweon, Kuk-Jin Yoon, and Horst Bischof. Mate: Masked autoencoders are online 3d test-time learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16709-16718, 2023. 2, 3", + "[36] Emin Orhan. A simple cache model for image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3, 4", + "[37] Jinyoung Park, Sanghyeok Lee, Sihyeon Kim, Yunyang Xiong, and Hyunwoo J. Kim. Self-positioning point-based transformer for point cloud understanding. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21814-21823, 2023. 1", + "[38] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc Le. Combined scaling for open-vocabulary image classification. 2021. 2", + "[39] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 3", + "[40] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1, 3", + "[41] Can Qin, Haoxuan You, Lichen Wang, C.-C. Jay Kuo, and Yun Fu. Pointdan: A multi-scale 3d domain adaption network for point cloud representation. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 2, 3", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[43] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions. In In" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1272", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ternational Conference on Machine Learning (ICML), 2022. 1, 2, 5, 6", + "[44] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions, 2022. 1", + "[45] Yuefan Shen, Yanchao Yang, Mi Yan, He Wang, Youyi Zheng, and Leonidas Guibas. Domain adaptation on point clouds via geometry-aware implicit. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7213-7222, 2022. 1, 2, 3", + "[46] Hajin Shim, Changhun Kim, and Eunho Yang. CloudFixer: Test-Time Adaptation for 3D Point Clouds via Diffusion-Guided Geometric Transformation. In European Conference on Computer Vision (ECCV), 2024. 2, 3", + "[47] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, pages 14274-14289. Curran Associates, Inc., 2022. 2", + "[48] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3, 4", + "[49] Hongyu Sun, Yongcai Wang, Xudong Cai, Xuewei Bai, and Deying Li. Vipformer: Efficient vision-and-pointcloud transformer for unsupervised pointcloud understanding. In IEEE International Conference on Robotics and Automation (ICRA), 2023. 3", + "[50] Hongyu Sun, Qiuhong Ke, Yongcai Wang, Wang Chen, Kang Yang, Deying Li, and Jianfei Cai. Point-prc: A prompt learning based regulation framework for generalizable point cloud analysis. In The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2", + "[51] Jiachen Sun, Qingzhao Zhang, Bhavya Kailkhura, Zhiding Yu, Chaowei Xiao, and Z. Morley Mao. Benchmarking robustness of 3d point cloud recognition against common corruptions, 2022. 1", + "[52] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2", + "[53] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. Proceedings of the AAAI Conference on Artificial Intelligence, 38(6):5171-5179, 2024. 3, 4, 6", + "[54] Yun-Yun Tsai, Fu-Chen Chen, Albert Y. C. Chen, Junfeng Yang, Che-Chun Su, Min Sun, and Cheng-Hao Kuo. Gda: Generalized diffusion for robust test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23242-23251, 2024. 2", + "[55] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2725-2736, 2023. 3, 4", + "[56] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 3, 5, 6", + "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 4", + "[58] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 3, 4", + "[59] Anran Wang, Jiwen Lu, Jianfei Cai, Tat-Jen Cham, and Gang Wang. Large-margin multi-modal deep learning for rgb-d object recognition. IEEE Transactions on Multimedia, 17 (11):1887-1898, 2015. 1", + "[60] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 2", + "[61] Haiyang Wang, Chen Shi, Shaoshuai Shi, Meng Lei, Sen Wang, Di He, Bernt Schiele, and Liwei Wang. Dsvt: Dynamic sparse voxel transformer with rotated sets. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13520-13529, 2023. 1", + "[62] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023.", + "[63] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q. Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8437-8445, 2019. 1", + "[64] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 2019. 1, 3", + "[65] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q. Weinberger, and WeiLun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11710-11720, 2020. 2", + "[66] Yanshuo Wang, Ali Cheraghian, Zeeshan Hayden, Jie Hong, Sameera Ramasinghe, Shafin Rahman, David Ahmedt-Aristizabal, Xuesong Li, Lars Petersson, and Mehrtash Harandi. Backpropagation-free network for 3d test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23231-23241, 2024. 2, 3, 4, 6" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1273", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Xin Wei, Xiang Gu, and Jian Sun. Learning generalizable part-based feature representation for 3d point clouds. In Advances in Neural Information Processing Systems, pages 29305-29318. Curran Associates, Inc., 2022. 1, 2, 3, 4", + "[68] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7): 5092-5113, 2024. 1", + "[69] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 803-814, 2023. 1, 5, 6", + "[70] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1, 3, 5, 6", + "[71] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 915-924, 2021. 3", + "[72] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning unified representation of language, image and point cloud for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6, 1, 4", + "[73] Le Xue, Ning Yu, Shu Zhang, Junnan Li, Roberto Martín-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 5, 6, 1", + "[74] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[75] Teresa Yeo, Oğuzhan Fatih Kar, Zahra Sodagar, and Amir Zamir. Rapid network adaptation: learning to adapt neural networks using test-time feedback. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4651-4664, 2023. 1", + "[76] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In Advances in Neural Information Processing Systems, pages 32215-32234. Curran Associates, Inc., 2023. 3", + "[77] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 4, 1, 3", + "[78] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23783-23793, 2024. 2", + "[79] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18123-18133, 2022. 2", + "[80] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In Advances in Neural Information Processing Systems, pages 38629-38642. Curran Associates, Inc., 2022. 2", + "[81] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8552-8562, 2022. 3", + "[82] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 493-510, Berlin, Heidelberg, 2022. Springer-Verlag. 3, 4", + "[83] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15211-15222, 2023. 3, 4", + "[84] Renrui Zhang, Liuhui Wang, Yali Wang, Peng Gao, Hongsheng Li, and Jianbo Shi. Starting from non-parametric networks for 3d point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5344-5353, 2023. 3, 4, 6", + "[85] Weichen Zhang, Wen Li, and Dong Xu. Srdan: Scale-aware and range-aware domain adaptation network for cross-dataset 3d object detection. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6765-6775, 2021. 2", + "[86] Zhimin Zhang, Xiang Gao, and Wei Hu. Invariantoodg: Learning invariant features of point clouds for out-of-distribution generalization, 2024. 2, 4", + "[87] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H.S. Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16259-16268, 2021. 1, 3", + "[88] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 5, 6, 7, 4" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1274", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "[89] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3", + "bbox": [ + 78, + 90, + 470, + 161 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "1275", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/b2000dba-b471-4537-8cba-488e6bd7eaf3_model.json b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/b2000dba-b471-4537-8cba-488e6bd7eaf3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..34b78c9738e16ea4883bb61e2122d8043267c1fc --- /dev/null +++ b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/b2000dba-b471-4537-8cba-488e6bd7eaf3_model.json @@ -0,0 +1,3053 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.852, + 0.175 + ], + "angle": 0, + "content": "Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.203, + 0.89, + 0.256 + ], + "angle": 0, + "content": "Hongyu Sun\\(^{1,2}\\), Qiuhong Ke\\(^{2}\\), Ming Cheng\\(^{2}\\), Yongcai Wang\\(^{1*}\\), Deying Li\\(^{1}\\), Chenhui Gou\\(^{2}\\), Jianfei Cai\\(^{2}\\) \n\\(^{1}\\)Department of Computer Science, Renmin University of China, China \n\\(^{2}\\)Department of Data Science & AI, Monash University, Australia" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.26, + 0.854, + 0.275 + ], + "angle": 0, + "content": "\\(\\{\\mathrm{sunhongyu,ycw,deyingli}\\} \\mathrm{ruc.edu.cn}\\) {qiuhong.ke,ming.cheng, jianfei.cai}@monash.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.474, + 0.705 + ], + "angle": 0, + "content": "This paper proposes a general solution to enable point cloud recognition models to handle distribution shifts at test time. Unlike prior methods, which rely heavily on training data (often inaccessible during online inference) and are limited to recognizing a fixed set of point cloud classes predefined during training, we explore a more practical and challenging scenario: adapting the model solely based on online test data to recognize both previously seen classes and novel, unseen classes at test time. To this end, we develop Point-Cache, a hierarchical cache model that captures essential clues of online test samples, particularly focusing on the global structure of point clouds and their local-part details. Point-Cache, which serves as a rich 3D knowledge base, is dynamically managed to prioritize the inclusion of high-quality samples. Designed as a plug-and-play module, our method can be flexibly integrated into large multimodal 3D models to support open-vocabulary point cloud recognition. Notably, our solution operates with efficiency comparable to zero-shot inference, as it is entirely training-free. Point-Cache demonstrates substantial gains across 8 challenging benchmarks and 4 representative large 3D models, highlighting its effectiveness. Code is available at https://github.com/auniquesun/Point-Cache." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.732, + 0.21, + 0.749 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.757, + 0.47, + 0.879 + ], + "angle": 0, + "content": "In recent years, 3D sensors such as LiDARs and RGB-D cameras have been widely adopted in robotics and electric vehicles for their ability to provide reliable 3D geometry measurements [11, 16, 20, 31, 59, 61-63, 68]. Point clouds are among the most direct data formats produced by these 3D sensors. Although remarkable progress has been made in 3D point cloud recognition, the success is primarily based on the assumption that the test data and the model train" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.309, + 0.697, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.414, + 0.68, + 0.425 + ], + "angle": 0, + "content": "(a) ModelNet & ModelNet-C." + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.31, + 0.888, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.413, + 0.884, + 0.425 + ], + "angle": 0, + "content": "(b) ScanObjNN & ScanObjNN-C." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.438, + 0.895, + 0.522 + ], + "angle": 0, + "content": "Figure 1. Recognition accuracy comparison on clean and corrupted point cloud datasets. The suffix -C indicates datasets with corruptions. Models experience a severe performance drop when data corruptions arise. Point-Cache effectively narrows down the performance gap between clean and corrupted data. The hardest split of ScanObjNN is used." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.893, + 0.613 + ], + "angle": 0, + "content": "ing data are identically distributed [2, 9, 26, 32, 37, 39, 40, 64, 87]. However, this assumption is frequently violated in real-world scenarios due to complex geometries, as well as sensing and processing errors [43, 50, 51]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In practice, a majority of models remain vulnerable to distribution shifts, such as out-of-distribution (OOD) samples [6, 7, 50, 69], data corruptions [43, 44, 51, 56], and more. As shown in Fig. 1, notable performance gaps (e.g., \\(11 + \\%\\) ) occur when models are tested on clean (ModelNet [70]) versus corrupted data (ModelNet-C [43]). These issues drive researchers to enhance the robustness and generalization of point cloud recognition methods. Related efforts can be grouped into three main streams. The first focuses on training models that can transfer to target data. These models learn a shared feature space between source and target domains [10, 41, 45, 67], or apply meta-learning frameworks to handle geometry shifts by training multiple classifiers [19], or design learnable prompts to elicit general knowledge from pre-trained large 3D models [50]. However, possible shifts are unpredictable, and the training-based methods cannot account for all cases during training. As a result, these models are limited when new shifts arise at test time [75]. The second stream addresses the problem" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1263" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "through test-time adaptation strategies [35, 46, 66]. These strategies improve performance with various adaptation techniques, such as masked auto-encoding for point clouds (MATE [35]), static prototype memory (BFTT3D [66]), and diffusion models (CloudFixer [46]). Nevertheless, we find that these approaches rely heavily on training data. For instance, MATE [35] requires pre-training on the entire training set, and BFTT3D [66] employs training data to build a static and offline prototype memory to store point cloud features. This heavy dependence on training data poses challenges, particularly when such data is unavailable or when distribution shifts occur at test time. In these cases, revisiting the training data may not be helpful to adapt the model to online test data. Additionally, these test-time methods can only recognize classes seen during training, which limits their practicality in generalizable point cloud analysis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.473, + 0.61 + ], + "angle": 0, + "content": "Inspired by advancements in language-guided image recognition [22, 38, 42, 79], the third branch of works utilizes contrastive pre-training on large-scale triplets (point, text, image) to connect flexible language descriptions and enable open-vocabulary point cloud recognition [29, 72, 73, 88]. These large multimodal 3D models demonstrate promising zero-shot performance and provide a new paradigm for generalizable point cloud analysis. However, zero-shot predictions are performed independently for each test sample, with the final accuracy simply averaged across the test set. This approach does not fully exploit the distribution of test data, particularly the statistics of reliable predictions from large 3D models, which could further enhance accuracy and generalization. At this point, we pose the following question: can we take advantages of both the promising zero-shot ability of large 3D models and test-time adaptation techniques to unlock more robust and generalizable point cloud recognition?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this work, to fully exploit the data distribution during test time, we develop a cache model to store critical fingerprints of test samples. These fingerprints consist of key-value pairs derived by large multimodal 3D models, where the key represents the sample feature extracted by the point encoder, and the value represents the predicted (pseudo) label from the model. This cache model is entirely constructed from online test data without accessing any training samples. Our approach is thus more practical and challenging than prior methods [10, 19, 35, 41, 45, 46, 50, 66, 67]. Our cache model is hierarchical, as it stores not only the global features of a point cloud in a global cache but also its local-part details in a local cache, which are essential for distinguishing subtle differences among classes. We also empirically demonstrate that this coarse-to-fine design enhances robustness and generalization in the wild. Additionally, our cache model is dynamic, since we need to continuously update the cache with online test data to prioritize high-quality key-value pairs. The constructed cache" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "serves as a powerful 3D knowledge base, capturing key characteristics of online test data. When a new sample arrives, we query the knowledge base, compute the affinity between the sample and cached keys, and produce a prediction by weighting the cached (pseudo) labels accordingly. Our method is training-free and flexible, requiring only a pre-trained 3D model and test data. More details will be provided in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.213, + 0.893, + 0.468 + ], + "angle": 0, + "content": "The superiority of the proposed method is supported by comprehensive comparisons with previous strong baselines across 8 challenging benchmarks, containing up to 1,156 point cloud classes. Moreover, we implement the hierarchical cache as a plug-and-play module and integrate it into representative large 3D models, such as ULIP [72], ULIP-2 [73], OpenShape [29] and Uni3D [88], to enhance test-time point cloud recognition dynamically. The proposed cache model brings notable and consistent gains over the baselines. For instance, our model yields \\(+6.18\\%\\) and \\(+4.84\\%\\) absolute accuracy improvements average over 7 types of corruptions in ModelNet-C [43] based on ULIP [72] and ULIP-2 [73]. On Objaverse-LVIS [7] that includes 1,156 categories, Point-Cache boosts the accuracy by \\(+2.10\\%\\) absolute points compared to the zero-shot ULIP-2 [73]. In summary, the contributions of our paper are threefold." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.469, + 0.892, + 0.559 + ], + "angle": 0, + "content": "- To our knowledge, Point-Cache is the first to explore test-time point cloud recognition in a more practical and challenging setting, where only online test data is available and the model needs to generalize to both known classes and unseen new classes, which is beyond the capability of prior test-time methods." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.56, + 0.892, + 0.65 + ], + "angle": 0, + "content": "- We propose a hierarchical cache model to capture essential distribution clues of online test samples, creating an accurate profile for various point clouds from global to local perspectives. We implement the hierarchical cache as a plug-and-play component, enabling diverse 3D backbones to achieve enhanced test-time performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.651, + 0.892, + 0.695 + ], + "angle": 0, + "content": "- Comprehensive experiments and consistent gains across 8 public datasets and 4 large 3D models demonstrate the efficacy of Point-Cache." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.469, + 0.892, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.642, + 0.725 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.735, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Test-time adaptation in 3D point cloud analysis receives relatively less attention compared to the 2D image community [5, 12, 14, 21, 23, 27, 28, 30, 34, 47, 52, 54, 60, 78, 80]. This technique has recently emerged in 3D point cloud tasks, including recognition [35, 46, 66], registration [17], upsampling [18] and object detection [65, 74, 85]. Among these, MATE [35], BFTT3D [66], and CloudFixer [46] are specifically focused on test-time point cloud recognition and are thus closely related to our work. However, as previously analyzed, these methods depend heavily on training data from the source domain to construct their pipelines," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1264" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.887, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.893, + 0.375 + ], + "angle": 0, + "content": "Figure 2. The overall pipeline of Point-Cache. The zero-shot predictions \\(\\hat{\\mathbf{y}}_{zs}\\) of large 3D models are effectively adapted by our global cache logits \\(\\hat{\\mathbf{y}}_g\\) and local cache logits \\(\\hat{\\mathbf{y}}_l\\) to handle the distribution shifts, enabling robust and generalizable point cloud analysis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.473, + 0.565 + ], + "angle": 0, + "content": "which may not be available or helpful to understand target data distributions at test time. Furthermore, they lack effective strategies to leverage the distribution of test samples. By contrast, Point-Cache is constructed entirely from online test data without access to any training samples, making our setting both more challenging and more practical. Additionally, our cache model, powered by large 3D models, can generalize to both known classes and unseen new categories, an advancement that prior point cloud adaptation/generalization frameworks have not achieved [10, 19, 35, 41, 45, 46, 66, 67]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Cache models that store sample features in memory for further reference and reuse have been explored in fields like language modeling [15, 24, 33, 58], image classification [3-5, 13, 21, 23, 36, 48, 55, 58, 82, 83] and point cloud recognition [53, 66, 84]. Our proposed cache model stands out from previous ones in several key aspects: (1) Our cache model is constructed based on test samples whereas previous cache models rely on a large training set for construction, as in Tip-Adapter [82], CaFo [83], PointNN [84], Point-PEFT [53] and BFTT3D [66]. (2) Our cache model is hierarchically designed to capture both global and local-part features of a point cloud object, while existing methods store only global representations [3-5, 13, 21, 48, 53, 58, 66, 82-84]. We empirically demonstrate that this coarse-to-fine design improves the capture of target data features, thereby enhancing test-time performance. (3) Our cache model is dynamically updated with online test samples, incorporating a selective mechanism to ensure high-quality key-value pairs replace less reliable ones when the cache reaches its capacity. In contrast, prior cache models, which are constructed offline using the entire training set, are static and consume a substantial amount of memory [53, 66, 84]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.895, + 0.701 + ], + "angle": 0, + "content": "Large Multi-modal 3D Models [25, 29, 72, 73, 76, 81, 88, 89] spark a wave of open-vocabulary point cloud understanding by pre-training on large-scale datasets [7] in recent years. For instance, ULIP-2 [73], OpenShape [29] and Uni3D [88] pre-train the 3D encoders on million-scale (point, image, text) triplets using contrastive learning. These pre-trained point cloud encoders are then transferred to downstream 3D tasks like classification and retrieval, demonstrating strong zero-shot performance that surpasses traditional 3D models [32, 39, 40, 64, 71, 87] which are typically tailored to specific datasets [1, 56, 70]. This new paradigm presents a promising direction for generalizable point cloud processing. Our work is complementary to these models: rather than developing another large 3D model, we build upon their capabilities to enhance robustness and generalization for point cloud analysis at test time. Our hierarchical cache, backed by these large 3D models, generalizes to unseen classes beyond training, addressing a significant limitation of current test-time methods for point cloud recognition [35, 46, 66]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.72, + 0.593, + 0.736 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.892, + 0.779 + ], + "angle": 0, + "content": "This section elaborates on the details of Point-Cache, the overall pipeline of our approach is depicted in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.904 + ], + "angle": 0, + "content": "Preliminary. We begin by reviewing the zero-shot inference process of existing large multimodal 3D models [29, 72, 73, 81, 88, 89]. These models use separate encoders to map inputs from a point cloud \\( P \\in \\mathbb{R}^{N \\times 3} \\) and text \\( T \\) (e.g., \"a point cloud object of a {class}\") to a shared feature space. Denoting the point cloud encoder as \\( f_{p}(\\cdot) \\) and the text encoder as \\( f_{t}(\\cdot) \\), we can obtain the global feature \\( \\mathbf{e}_p^g = f_p(P) \\in \\mathbb{R}^d \\) of the point cloud \\( P \\) and the text" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1265" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.169 + ], + "angle": 0, + "content": "feature \\(\\mathbf{e}_t = f_t(T) \\in \\mathbb{R}^d\\), where \\(d\\) is the feature dimension. To recognize the input point cloud \\(P\\) in a zero-shot manner, these models replace “\\(\\{\\mathrm{class}\\}\\)” in text \\(T\\) with each of possible \\(C\\) category names and compute the class probability \\(\\hat{\\mathbf{y}}_{zs}\\) using Eq. 1." + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.176, + 0.47, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} _ {z s} = \\{\\hat {y} _ {i} | _ {i = 1} ^ {C} \\}, \\quad \\hat {y} _ {i} = \\frac {\\exp (s i m (\\mathbf {e} _ {p} ^ {g} , \\mathbf {e} _ {t} ^ {i}) / \\tau)}{\\sum_ {j = 1} ^ {C} \\exp (s i m (\\mathbf {e} _ {p} ^ {g} , \\mathbf {e} _ {t} ^ {j}) / \\tau)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.225, + 0.469, + 0.305 + ], + "angle": 0, + "content": "Here \\(\\mathbf{e}_t^i\\) indicates the text encoding of the \\(i\\)-th category. \\(sim(\\cdot, \\cdot)\\) measures the cosine similarity between inputs, and \\(\\tau\\) is a temperature coefficient. Finally, the point cloud \\(P\\) is classified into the category with the highest probability, indexed by \\(\\hat{L} = \\arg \\max_{i} \\{\\hat{y}_i|_{i=1}^C\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.47, + 0.458 + ], + "angle": 0, + "content": "As described in Sec. 1, our hierarchical cache stores the fingerprints of online test samples, which consists of key-value pairs. The key is the test sample feature \\(\\mathbf{e}_p^g\\) extracted by the point encoder of a pre-trained large 3D model, and the value is the class label \\(\\hat{L}\\) predicted by the zero-shot inference. In addition, we compute the entropy \\(h = -\\sum_{i=1}^{C} \\hat{y}_i \\log \\hat{y}_i\\) of the zero-shot prediction \\(\\hat{\\mathbf{y}}_{zs}\\), as it is essential to assess the confidence with which the large 3D model produces this prediction. This entropy value is employed to filter out high-quality samples in our cache model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.342, + 0.483 + ], + "angle": 0, + "content": "3.1. Text & Point Cloud Encoding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.566 + ], + "angle": 0, + "content": "We can obtain the text encoding \\(\\mathbf{e}_t\\) of the input \\(T\\) and the global feature \\(\\mathbf{e}_p^g\\) of a point cloud \\(P\\) as introduced in the preliminary. However, it is non-trivial to represent the local-part features \\(\\mathbf{e}_p^l\\) of the point cloud \\(P\\) due to the following challenges analyzed in Fig. 3:" + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.581, + 0.266, + 0.595 + ], + "angle": 0, + "content": "01 Part Feature Capture" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.603, + 0.257, + 0.66 + ], + "angle": 0, + "content": "- Difficult to define parts precisely and get their features for various 3D objects of different classes" + }, + { + "type": "title", + "bbox": [ + 0.28, + 0.581, + 0.433, + 0.596 + ], + "angle": 0, + "content": "02. Part Feature Storage" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.602, + 0.42, + 0.66 + ], + "angle": 0, + "content": "- Expensive to store multiple part features instead of a global feature for a single point cloud" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.676, + 0.469, + 0.705 + ], + "angle": 0, + "content": "Figure 3. Challenges in encoding part feature for various 3D objects of different classes: Part Feature Capture & Storage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To address these challenges, we made a careful analysis and developed effective solutions, seeing Fig. 9 in the Supplementary for intuitive illustration. (1) A straightforward approach might involve using an additional pretrained segmentation network to segment point cloud parts, but this would be computationally prohibitive during online inference. Instead, we propose a fuzzy definition of parts. Specifically, since the architecture of the point encoder of large 3D models is based on Transformer [8, 57, 77] and the input is a sequence of point patches, we avoid strict part definitions according to physical semantics (e.g., head, body, wing, and tail in an airplane), and instead treat the output" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "point patches from the last encoder layer as local part representations. (2) A second challenge is that a long point patch sequence can become resource-intensive; for example, the point encoder in ULIP-2 [73] generates 512 point patches per object, which would require substantial memory. To mitigate this problem, we propose to summarize these point patches into \\( m \\) centers \\( \\mathbf{e}_p^l \\in \\mathbb{R}^{m \\times d} \\) using K-Means clustering (\\( m < 10 \\)), reducing memory usage by two orders of magnitude compared to storing all patches." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.238, + 0.89, + 0.254 + ], + "angle": 0, + "content": "3.2. Construction & Update of Hierarchical Cache" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.427 + ], + "angle": 0, + "content": "To thoroughly explore the distribution of test data, we develop a hierarchical cache to record the fingerprints of online test samples. This hierarchical cache includes both a global cache, which stores overall information, and a local cache, which captures detailed part-level information for each point cloud. The coarse-to-fine structure enables more precise profiling of test data and distinguishes our approach from previous cache models used in image recognition [3-5, 13, 21, 23, 36, 48, 55, 58, 82, 83] and point cloud analysis [53, 66, 84]. The complete construction process of hierarchical cache is outlined in Alg. 1 of the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Global cache is represented by \\(\\mathbf{C}_g = (\\mathbf{E}_g, \\hat{\\mathbf{L}}_g, \\mathbf{h}_g)\\), storing (feature, label, entropy) triplets for a set of point cloud objects. To manage storage efficiently, we design a selection mechanism to identify high-quality test samples to place into the cache." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Construction: We set an upper bound \\( K \\) on the number of samples per class in the cache, allowing at most \\( K \\) fingerprints (shots) can be placed in each category. Initially, the global cache is empty. When a new test sample \\( P \\) arrives, we generate its fingerprint \\( (\\mathbf{e}_p^g, \\hat{L}, h) \\) and assign it to class \\( \\hat{L} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Update: With the number of online samples grows, we check whether the number of shots for class \\(\\hat{L}\\) has hit the upper bound. If it is not the case, the fingerprint of current sample is appended to class \\(\\hat{L}\\). Otherwise, we locate the fingerprint with highest entropy within class \\(\\hat{L}\\) and compare it to the entropy of current test sample. If the current test sample has a lower entropy, we replace the highest-entropy fingerprint with the new one; otherwise, the current sample is skipped. When the cache is full, \\(\\mathbf{E}_g\\) has dimensions \\(CK\\times d\\), \\(\\hat{\\mathbf{L}}_g\\) has dimensions \\(CK\\times 1\\), and \\(\\mathbf{h}_g\\) has dimensions \\(CK\\times 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Local cache. Although the global feature of a point cloud captures expressive information, it lacks the local details of the 3D object, which are critical for point cloud recognition, especially when shapes share an overall appearance but differ in local parts. The local cache addresses this limitation by storing part-level information. We denote the local cache as \\(\\mathbf{C}_l = (\\mathbf{E}_l,\\hat{\\mathbf{L}}_l)\\), which store the fingerprints of local parts of point cloud objects. Here \\(\\mathbf{E}_l\\) represents part features of the test samples and \\(\\hat{\\mathbf{L}}_l\\) indicates their labels." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1266" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.139 + ], + "angle": 0, + "content": "Accordingly, we define the local-part fingerprint of a test sample \\(P\\) as \\((\\mathbf{e}_p^l,\\hat{L})\\), where \\(\\mathbf{e}_p^l\\in \\mathbb{R}^{m\\times d}\\) corresponds to local-part features and \\(\\hat{L}\\) is the class label of \\(P\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.14, + 0.468, + 0.246 + ], + "angle": 0, + "content": "Construction & Update: Note that the sample selective mechanism of the local cache follows the same rules as the global cache, meaning the local-part fingerprint of a test sample is only stored if it qualifies for inclusion in the global cache. So there is no need to record the entropy \\( h \\) again in the local cache. When the cache is full, \\( \\mathbf{E}_l \\) has dimensions \\( (m \\cdot CK) \\times d \\) and \\( \\hat{\\mathbf{L}}_l \\) has dimensions \\( CK \\times 1 \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.255, + 0.453, + 0.272 + ], + "angle": 0, + "content": "3.3. Test-time Adaptation by Hierarchical Cache" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.278, + 0.468, + 0.353 + ], + "angle": 0, + "content": "With the hierarchical cache serving as a rich knowledge base of online point cloud data, we can adapt the zero-shot predictions of large 3D models for new samples, effectively addressing distribution shifts that occur at test time. The adaptation follows a coarse-to-fine process as below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.468, + 0.508 + ], + "angle": 0, + "content": "Test-time adaptation by global cache. During adaptation, a new sample \\( Q \\) is treated as a query to retrieve knowledge from the global cache, which we use to produce the adaptation logits. Specifically, the retrieval process is simulated by calculating the affinity \\( A_{g} \\) between the query and cached features \\( \\mathbf{E}_{g} \\). This allows the new sample to establish connections with cached high-quality samples, providing it with a contextual understanding of the online test data. We then generate the adaptation logits \\( \\hat{\\mathbf{y}}_{g} \\) by weighed ensembling of cache labels \\( \\hat{\\mathbf{L}}_{g} \\), as shown in Eq. 2." + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.518, + 0.468, + 0.55 + ], + "angle": 0, + "content": "\\[\nA _ {g} = \\exp \\left(- \\beta_ {g} \\left(1 - \\mathbf {e} _ {q} ^ {g} \\mathbf {E} _ {g} ^ {\\top}\\right)\\right), \\hat {\\mathbf {L}} _ {g} = \\mathrm {O H} (\\hat {\\mathbf {L}} _ {g}), \\hat {\\mathbf {y}} _ {g} = A _ {g} \\hat {\\mathbf {L}} _ {g} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.551, + 0.468, + 0.72 + ], + "angle": 0, + "content": "where \\(\\mathbf{e}_q^g\\in \\mathbb{R}^{1\\times d}\\) represents the global feature of sample \\(Q\\) and \\(\\mathbf{E}_g\\in \\mathbb{R}^{CK\\times d}\\) denotes the cached keys. The 'OH' operation transforms \\(\\hat{\\mathbf{L}}_g\\) into a one-hot encoding. The term \\(\\mathbf{e}_q^g\\mathbf{E}_g^\\top \\in \\mathbb{R}^{1\\times CK}\\) measures the cosine similarity between the query and all keys, with \\(\\beta_{g}\\) acting as a hyper-parameter to modulate the sharpness and the exponential operation ensuring similarity values fall within (0, 1]. The affinity \\(A_{g}\\in \\mathbb{R}^{1\\times CK}\\), which grasps the relationship between the query and cached keys, is utilized to ensemble the transformed one-hot labels \\(\\hat{\\mathbf{L}}_g\\in \\mathbb{R}^{CK\\times C}\\), yielding global cache adaptation logits \\(\\hat{\\mathbf{y}}_g\\in \\mathbb{R}^{1\\times C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.859 + ], + "angle": 0, + "content": "Test-time adaptation by local cache. While global cache adaptation is effective, it may not handle fine-grained distribution shifts at test time. To achieve a more precise adjustment, we exploit local part knowledge. The local-part features \\(\\mathbf{e}_q^l\\) of the point cloud \\(Q\\) serve as queries to the local cache. During retrieval, we compute the affinity \\(A_{l}\\) between \\(\\mathbf{e}_q^l\\) and all cached part features \\(\\mathbf{E}_l\\). The affinity \\(A_{l}\\) is then used to weight the cache labels \\(\\hat{\\mathbf{L}}_l\\), producing local cache adaptation logits \\(\\hat{\\mathbf{y}}_l\\) as in Eq. 3." + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.868, + 0.468, + 0.901 + ], + "angle": 0, + "content": "\\[\nA _ {l} = \\exp \\left(- \\beta_ {l} \\left(1 - \\mathbf {e} _ {q} ^ {l} \\mathbf {E} _ {l} ^ {\\top}\\right)\\right), \\hat {\\mathbf {L}} _ {l} = \\mathrm {O H} (\\hat {\\mathbf {L}} _ {l}), \\hat {\\mathbf {y}} _ {l} = \\phi \\left(A _ {l} \\hat {\\mathbf {L}} _ {l}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.09, + 0.892, + 0.26 + ], + "angle": 0, + "content": "where \\(\\mathbf{e}_q^l\\in \\mathbb{R}^{m\\times d}\\), \\(\\mathbf{E}_l\\in \\mathbb{R}^{(m\\cdot CK)\\times d}\\), and \\(\\hat{\\mathbf{L}}_l\\in \\mathbb{R}^{(m\\cdot CK)\\times C}\\). The function \\(\\phi (\\cdot)\\) is an average pooling operation over \\(m\\) parts to produce class logits. The term \\(e_q^l\\mathbf{E}_l^\\top \\in \\mathbb{R}^{m\\times (m\\cdot CK)}\\) measures the cosine similarity between the part features of sample \\(Q\\) and all part features in the cache. Likewise, the similarity is modulated by the coefficient \\(\\beta_{l}\\) and scaled with the exponential function, resulting in the affinity \\(A_{l}\\in \\mathbb{R}^{m\\times (m\\cdot CK)}\\). The term \\(A_{l}\\hat{\\mathbf{L}}_{l}\\in \\mathbb{R}^{m\\times C}\\) represents the class logits for all \\(m\\) parts, which are transformed by a pooling operation \\(\\phi (\\cdot)\\) to derive the final local cache adaptation logits \\(\\hat{\\mathbf{y}}_l\\in \\mathbb{R}^{1\\times C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.336 + ], + "angle": 0, + "content": "Overall prediction combines three components as shown in Eq. 4. The zero-shot prediction \\(\\hat{\\mathbf{y}}_{zs}\\) of large multimodal 3D model is effectively adapted using our global cache prediction \\(\\hat{\\mathbf{y}}_g\\) and local cache prediction \\(\\hat{\\mathbf{y}}_l\\), balanced by the coefficients \\(\\alpha_{q}\\) and \\(\\alpha_{l}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.351, + 0.892, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} = \\hat {\\mathbf {y}} _ {z s} + \\alpha_ {g} \\hat {\\mathbf {y}} _ {g} + \\alpha_ {l} \\hat {\\mathbf {y}} _ {l} \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.382, + 0.634, + 0.399 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.408, + 0.707, + 0.424 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Datasets. To evaluate the robustness of point cloud recognition, we choose four public datasets that contain various point cloud corruptions: ModelNet-C [43] and three variants of ScanObjectNN-C [56]. ModelNet-C [43] introduces 7 atomic corruptions, including adding global outliers, adding local outliers, dropping global structure, dropping local parts, rotation, scaling, and jittering. Other corruptions can be derived from these atomic corruptions, as discussed in [43]. We apply these atomic corruptions to the three variants of ScanObjectNN [56] and produce their corrupted versions. To assess generalization on unseen new data, we test the proposed method on four challenging benchmarks: OmniObject3D [69] (216 classes), ObjaverseLVIS [7] (1,156 classes), the hardest variant of ScanObjectNN [56] and widely used ModelNet40 [70]. Recognition accuracy (\\%) is the primary evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We choose ULIP [72], ULIP-2 [73], OpenShape [29] and Uni3D [88] as the large multimodal 3D models for experiments. Their pre-trained weights are loaded and then frozen. Global and local-part features are generated by modifying the forward pass function without altering the model architecture. The maximum number of samples per class in the cache, \\( K \\), is set to 3, and each point cloud is clustered into \\( m = 3 \\) local parts. Both the balance factors \\( \\alpha_{g} \\) and \\( \\alpha_{l} \\) are set to 4.0 while the coefficients \\( \\beta_{g} \\) and \\( \\beta_{l} \\) are set to 3.0. We will examine the design choices in the ablation studies. To optimize memory and computation, we convert the input point clouds and model parameters into 'dtype' fp16 in Pytorch. All experiments are conducted on two 4090 GPUs. Additional details can be found in the Supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1267" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Comparison of recognition accuracy on ModelNet-C that contains 7 types of corruptions. Each clean point cloud has 1024 points and the corruption severity level is 2. The last column is the average over 7 corruption types. The best results are in bold and the second best are underlined. The setting applies to the following tables unless otherwise specified." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.143, + 0.877, + 0.346 + ], + "angle": 0, + "content": "
MethodClean Data ModelNetCorruption TypeAvg.
Add GlobalAdd LocalDrop GlobalDrop LocalRotateScaleJitter
ULIP [72]56.1633.5543.9254.7050.8955.2750.2044.0847.52
+Global Cache(Ours)62.1245.7947.9856.8553.8960.2554.3448.9152.56
+Hierarchical Cache(Ours)64.2246.1547.8559.1656.0061.4755.3549.9253.70
ULIP-2 [73]71.2365.1554.6268.7657.9870.3067.1021.7657.95
+Global Cache(Ours)73.9567.0259.3271.3561.5972.3768.4028.2061.18
+Hierarchical Cache(Ours)74.5368.1161.2673.2263.6573.3470.4229.5062.79
O-Shape [29]84.5671.6467.7981.5673.5882.0178.4859.3673.49
+Global Cache(Ours)84.5274.7272.7782.4175.1283.1878.9367.9176.43
+Hierarchical Cache(Ours)84.0474.8473.7082.2176.2682.6678.1268.3576.59
Uni3D [88]81.8172.4556.3668.1567.1879.9475.3656.2467.95
+Global Cache(Ours)83.1476.1366.4971.4369.8181.5275.8561.4371.81
+Hierarchical Cache(Ours)83.8777.5171.1572.1670.7581.7777.3162.5273.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.359, + 0.893, + 0.403 + ], + "angle": 0, + "content": "Table 2. Comparison of recognition accuracy across a suite of datasets. S-PB_RS_T50 is the hardest split of ScanObjectNN. OLVIS: Objverse-LVIS. Omni3D: OmniObject3D. The number under each dataset indicates the number of points (pts) for each object. In Omni3D, each point cloud can be represented by a different number of points. The last column is the average accuracy on these datasets." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.412, + 0.832, + 0.574 + ], + "angle": 0, + "content": "
MethodModelNet40 (10000 pts)S-PB_RS_T50 (2048 pts)O-LVIS (10000 pts)Omni3DAvg.
(1024 pts4096 pts16384 pts)
ULIP [72]58.7546.446.248.397.757.2822.48
+Global Cache(Ours)61.2250.217.0210.009.368.4324.37
+Hierarchical Cache(Ours)62.9351.807.0210.479.758.9025.15
ULIP-2 [73]72.9747.1330.2626.3629.2026.5838.75
+Global Cache(Ours)74.5151.7032.6528.5131.1028.5341.17
+Hierarchical Cache(Ours)75.5354.9832.3629.3731.2429.4442.15
Uni3D [88]88.4165.1955.4231.5241.9841.8654.09
+Global Cache(Ours)88.8668.5153.3634.9745.1345.1956.00
+Hierarchical Cache(Ours)89.1868.2455.1935.8245.6045.8956.65
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.598, + 0.427, + 0.613 + ], + "angle": 0, + "content": "4.2. Test-time Robustness and Generalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Robustness. The robustness analysis of our training-free test-time hierarchical caches takes place on 4 point cloud corruption datasets [43, 56]. As observed in Tab. 1, the zero-shot accuracy of various large 3D models on ModelNet-C is considerably boosted by our global cache model, with increases such as \\(+5.04\\%\\) for ULIP [72], \\(+3.23\\%\\) for ULIP-2 [73], \\(+2.94\\%\\) for OpenShape [29] and \\(+3.86\\%\\) for Uni3D [88], averaged across 7 corruption types. By combining global and local cache together, our hierarchical cache model attains further improvements, highlighting the benefits of incorporating a local cache. Notably, these gains extend beyond corrupted data, as our hierarchical cache model also improves recognition accuracy on the clean ModelNet dataset, e.g., \\(+8.06\\%\\) absolute increase over ULIP's zero-shot performance. An exception is OpenShape, which maintains a slightly higher zero-shot accuracy on clean data than our cache model. However, when considering the 7 corruption types, our global and hi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.674 + ], + "angle": 0, + "content": "erarchical cache models outperform OpenShape by \\(2.94\\%\\) and \\(3.10\\%\\), respectively, demonstrating improved robustness against data corruption at test time. Consistent improvements are also showcased on the variants of ScanObjectNN, shown in Tab. 5, 6 and 7 in the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Generalization. We evaluated the generalization capability of Point-Cache on four representative benchmarks that were unseen during the pre-training of the large 3D models. These benchmarks generally do not contain data corruptions. While ModelNet40 [70] and ScanObjectNN [56] contain a limited number of classes, OmniObject3D [69] and Objverse-LVIS [7] represent a broader range of 3D concepts, covering 216 and 1,156 classes, respectively. As shown in Tab. 2, Point-Cache consistently enhances the recognition accuracy across datasets and 3D backbones. For instance, our hierarchical cache built on ULIP and ULIP-2 yields absolute improvements of \\(6.31\\%\\) and \\(5.51\\%\\), respectively, average over 6 scores. On ScanObjectNN, our hierarchical cache significantly raises the zero-shot prediction of ULIP and ULIP-2, from \\(27.20\\%\\) to \\(51.80\\%\\) for ULIP" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1268" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.198 + ], + "angle": 0, + "content": "and from \\(36.22\\%\\) to \\(54.98\\%\\) for ULIP-2. On the challenging Objaverse-LVIS (O-LVIS), our global cache with ULIP-2 successfully realizes a \\(2.39\\%\\) absolute gains in zero-shot accuracy across 1,156 classes, a notable achievement. Furthermore, our cache model enables Uni3D to attain an \\(89.18\\%\\) accuracy on ModelNet40 without any training, effectively matching the fully-supervised PointNet \\((89.20\\%)\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.209, + 0.363, + 0.226 + ], + "angle": 0, + "content": "4.3. Memory Usage and Throughput" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.233, + 0.474, + 0.462 + ], + "angle": 0, + "content": "Memory. We use Uni3D [88] as the baseline and adapt its zero-shot predictions with Point-Cache. The experiments are conducted on 3 datasets as shown in Tab. 3. The batch size is set to 1 because we need to update the cache and adapt the class logits according to every single sample. The results indicate our global cache and hierarchical cache consume similar or slightly higher memory than the baseline on each dataset. Furthermore, as the number of classes increases rapidly, the memory usage grows at a slower rate. This is primarily because the memory consumption is dominated by the large number of parameters in Uni3D (e.g., 1,016.5M). In contrast, the hierarchical cache built on OLVIS uses only approximately 7.1M parameters, which is negligible compared to Uni3D. Detailed calculations are provided in the Supplementary." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.475, + 0.471, + 0.517 + ], + "angle": 0, + "content": "Table 3. Comparison of memory usage (MB). The batch size is set to 1 and the used device is an RTX 4090. The number under each dataset indicates #Classes. Hierar: Hierarchical Cache." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.528, + 0.465, + 0.606 + ], + "angle": 0, + "content": "
MethodModelNet-C (40)Omni3D (216)O-LVIS (1156)#Params (M)
Uni3D5,0625,0625,0621,016.5
+Global(Ours)5,0625,0645,0701,016.5
+Hierar(Ours)5,0645,0685,0901,016.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.47, + 0.803 + ], + "angle": 0, + "content": "Throughput. We evaluate the throughput of Point-Cache on ModelNet40, where throughput is measured as the number of test samples processed per second (abbreviated as \\( t / s \\)). The results are reported in Tab. 4. Point-Cache runs slightly slower than zero-shot inference due to the additional overhead introduced by cache updates, logits computation, etc. However, the throughput degradation is minimal, e.g., a \\( 0.04 \\, t / s \\) drop for OpenShape with the global cache, and a \\( 0.07 \\, t / s \\) drop for ULIP-2 with the hierarchical cache. These results prove that Point-Cache brings remarkable accuracy improvements with negligible computational cost." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.231, + 0.832 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.839, + 0.469, + 0.871 + ], + "angle": 0, + "content": "We use Uni3D with our hierarchical cache and 4 public datasets for ablation analysis, as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Shot size per class in the global cache. The upper bound \\( K \\) on the number of samples per class is a critical variable" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.894, + 0.133 + ], + "angle": 0, + "content": "Table 4. Comparison of throughput \\((t / s)\\) for different models on ModelNet40. The batch size is set to 1 and the device is a 4090 GPU. The results are averaged over all test samples." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.143, + 0.875, + 0.22 + ], + "angle": 0, + "content": "
MethodZero-shot+Global Cache+Hierar Cache
ULIP11.2511.2111.18
ULIP-211.1811.1411.11
OpenShape9.809.769.74
Uni3D9.789.759.73
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.892, + 0.368 + ], + "angle": 0, + "content": "as it affects the total size of the hierarchical cache. We analyze its impact on the test-time recognition accuracy on 4 datasets, shown in Fig. 4a. As \\( K \\) increases, accuracy trend differ across datasets; for example, accuracy increases on Omni3D but decreases on S-PB_T50_RS. We adopt \\( K = 3 \\) as it produces the best average accuracy (63.98%). The results indicate that our cache model does not require a large number of samples to achieve strong performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Number of local parts per 3D object. To optimize memory and computation, we cluster hundreds of point patches of a 3D object into \\( m \\) parts. We investigate the effect of this variable on test-time performance, as exhibited in Fig. 4b. For three out of the four datasets, namely S-PB_T50_RS, Omni3D, and O-LVIS, accuracy generally decreases as the number of parts increases, suggesting that a point cloud does not need to be represented with an excessive number of parts. We choose \\( m = 3 \\) by default, as this setting reaches the best average accuracy across the 4 datasets." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.534, + 0.603, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.59, + 0.591, + 0.602 + ], + "angle": 0, + "content": "(a) shot size \\(K\\)" + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.534, + 0.697, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.59, + 0.68, + 0.602 + ], + "angle": 0, + "content": "(b) \\(m\\) parts" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.534, + 0.795, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.59, + 0.782, + 0.603 + ], + "angle": 0, + "content": "(c) \\(\\alpha_{g}\\) and \\(\\alpha_{l}\\)" + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.534, + 0.89, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.59, + 0.878, + 0.603 + ], + "angle": 0, + "content": "(d) \\(\\beta_{g}\\) and \\(\\beta_{l}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.615, + 0.894, + 0.672 + ], + "angle": 0, + "content": "Figure 4. Ablation studies on the hyper-parameters in the cache design, including the shot size \\( K \\) per class, the number of parts \\( m \\) per object, the balance factors in the final prediction logits and the sharpness coefficients in affinity computation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.824 + ], + "angle": 0, + "content": "The balance factors. The factors \\(\\alpha_{g}\\) and \\(\\alpha_{l}\\) control the relative weights of the global cache prediction \\(\\hat{\\mathbf{p}}_g\\) and local cache prediction \\(\\hat{\\mathbf{p}}_l\\). To simplify experimentation, we set \\(\\alpha_{g} = \\alpha_{l}\\) in the ablation study. Candidate values are \\(\\{0.5, 1.0, 2.0, 3.0, 4.0, 5.0\\}\\). As shown in Fig. 4c, increasing these factors improves performance on S-PB_T50_RS and Omni3D, while it has minimal effect on ModelNet40 and O-LVIS. We set the factors to 4.0, which yields the optimal average accuracy across all four datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The sharpness coefficients. The coefficients \\(\\beta_{g}\\) and \\(\\beta_{l}\\) are used to modulate the sharpness of the query-key affinity. Here we analyze their impact on test-time accuracy on 4 benchmarks. Likewise, we let \\(\\alpha_{g} = \\alpha_{l}\\) and the candidate values are \\(\\{0.5,1.0,3.0,5.0,7.0,9.0\\}\\). As Fig. 4d displays," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1269" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.152 + ], + "angle": 0, + "content": "accuracy initially improves slightly with increasing sharpness but then decreases as sharpness becomes too high. We set the sharpness to 3.0 that yields the optimal result average over 4 datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.164, + 0.216, + 0.179 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.188, + 0.473, + 0.369 + ], + "angle": 0, + "content": "Online Inference. We visualize the online inference process of Point-Cache using different large 3D models and compare three model variants: zero-shot prediction, the model with our global cache, and the model with our hierarchical cache, referring to Fig. 5. During inference, we calculate the average recognition accuracy over accumulated samples. Due to the small number of samples at the initial stage, the average accuracy may fluctuate drastically, so we discard the statistics for the first 5 samples. Overall, methods employing the hierarchical cache consistently outperform other variants by a clear margin across various large 3D models and datasets." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.382, + 0.177, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.438, + 0.154, + 0.45 + ], + "angle": 0, + "content": "(a) ULIP." + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.382, + 0.273, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.438, + 0.257, + 0.45 + ], + "angle": 0, + "content": "(b) ULIP-2." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.382, + 0.371, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.438, + 0.364, + 0.451 + ], + "angle": 0, + "content": "(c) OpenShape." + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.382, + 0.468, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.438, + 0.448, + 0.45 + ], + "angle": 0, + "content": "(d) Uni3D." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.533 + ], + "angle": 0, + "content": "Figure 5. The average recognition accuracy of accumulated samples during online inference. The curve changes significantly in the initial stage due to the small number of samples. Models with our global and hierarchical cache receive perceptible performance gains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.749 + ], + "angle": 0, + "content": "Entropy and Accuracy. Fig. 6 illustrates the relationship between entropy and accuracy for different models on various datasets. During online inference, we first compute the entropy of top 5 predictions for each test sample and then average the entropy across all accumulated samples. Similarly, the corresponding average accuracy can be calculated. The results reveal that Point-Cache effectively reduces the uncertainty of the zero-shot predictions made by large 3D models, enhances prediction confidence, and consequently lifts recognition accuracy. Additionally, we observe that the hierarchical cache model is more effective in reducing entropy compared to the global cache model, further underscoring the importance of incorporating the local cache." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Qualitative Analysis. Fig. 7 presents the step-by-step adaptation process of Point-Cache. The first three rows demonstrate cases where the global cache model attempts to adjust the zero-shot predictions of various large 3D models but fails. Instead, the hierarchical cache model successfully refines the class logits based on the global cache. In the last row, our global cache model corrects the zero-shot prediction in the first step, and the hierarchical cache model subsequently retains the top class while increasing its confidence. The examples showcased here are not cherry-picked." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.707, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.201, + 0.646, + 0.214 + ], + "angle": 0, + "content": "(a) Group One." + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.091, + 0.907, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.201, + 0.848, + 0.213 + ], + "angle": 0, + "content": "(b) Group Two." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.226, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Figure 6. Relationship between entropy and accuracy. Models with our global and hierarchical cache consistently exhibit lower entropy and higher accuracy across various benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.295, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Additional quantitative and qualitative results can be found in the Supplementary." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.34, + 0.587, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.396, + 0.599, + 0.407 + ], + "angle": 0, + "content": "(a) Ground truth." + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.339, + 0.704, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.397, + 0.688, + 0.409 + ], + "angle": 0, + "content": "(b) ULIP-2." + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.339, + 0.805, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.397, + 0.8, + 0.41 + ], + "angle": 0, + "content": "(c) ULIP-2+GC." + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.339, + 0.907, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.397, + 0.902, + 0.409 + ], + "angle": 0, + "content": "(d) ULIP-2+HC." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.415, + 0.57, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.476, + 0.599, + 0.486 + ], + "angle": 0, + "content": "(e) Ground truth." + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.415, + 0.704, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.472, + 0.679, + 0.484 + ], + "angle": 0, + "content": "(f) ULIP." + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.415, + 0.806, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.472, + 0.796, + 0.484 + ], + "angle": 0, + "content": "(g) \\(\\mathbf{ULIP} + \\mathbf{GC}\\)" + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.415, + 0.907, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.472, + 0.898, + 0.484 + ], + "angle": 0, + "content": "(h) ULIP+HC." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.492, + 0.584, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.544, + 0.598, + 0.555 + ], + "angle": 0, + "content": "(i) Ground truth." + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.49, + 0.704, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.544, + 0.688, + 0.556 + ], + "angle": 0, + "content": "(j) O-Shape." + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.49, + 0.806, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.544, + 0.804, + 0.556 + ], + "angle": 0, + "content": "(k) O-Shape+GC." + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.49, + 0.907, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.812, + 0.544, + 0.904, + 0.556 + ], + "angle": 0, + "content": "(1) O-Shape+HC." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.561, + 0.58, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.611, + 0.602, + 0.623 + ], + "angle": 0, + "content": "(m) Ground truth." + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.56, + 0.704, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.61, + 0.685, + 0.621 + ], + "angle": 0, + "content": "(n) Uni3D." + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.56, + 0.806, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.61, + 0.799, + 0.621 + ], + "angle": 0, + "content": "o) \\(\\mathrm{Uni3D + GC}\\)" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.56, + 0.907, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.61, + 0.901, + 0.621 + ], + "angle": 0, + "content": "(p) \\(\\mathrm{Uni3D + HC}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.636, + 0.892, + 0.664 + ], + "angle": 0, + "content": "Figure 7. The zero-shot predictions before and after adaptation by Point-Cache. GC: global cache. HC: hierarchical cache." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.619, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper explores test-time point cloud recognition in a practical and challenging setting, where only test data and the pre-trained models are available. We develop PointCache, a powerful 3D knowledge base that enhances robustness and generalization at test time. It records the fingerprints of online test samples through a global and a local cache, creating a more accurate profile of 3D data. Designed as a plug-and-play module, Point-Cache can be seamlessly integrated into various large 3D models without compromising efficiency, enabling open-vocabulary point cloud recognition for both known and unseen classes, effectively addressing limitations of prior test-time methods." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1270" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.238, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.473, + 0.328 + ], + "angle": 0, + "content": "We thank all anonymous reviewers and area chairs for their time and valuable feedback. This work was partially supported by China Scholarship Council (CSC) under the Grant No. 202306360147 and Australian Research Council funding under the Grant No. DE250100030. Dr. Qiuhong Ke is partially supported by the Australian Research Council funding Discovery Early Career Researcher Award (DE250100030). Dr. Deying Li was partially supported by the National Natural Science Foundation of China under the Grant No. 12071478. Dr. Yongcai Wang was partially supported by the National Natural Science Foundation of China under the Grant No. 61972404, the Public Computing Cloud at Renmin University of China, and the Blockchain Lab, School of Information, Renmin University of China." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.342, + 0.175, + 0.357 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.366, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[1] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. CoRR, abs/1512.03012, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.472, + 0.492 + ], + "angle": 0, + "content": "[2] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. PointGPT: Auto-regressively generative pre-training from point clouds. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.494, + 0.47, + 0.548 + ], + "angle": 0, + "content": "[3] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.551, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[4] Yinbo Chen, Zhuang Liu, Huijuan Xu, Trevor Darrell, and Xiaolong Wang. Meta-baseline: Exploring simple meta-learning for few-shot learning. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9042-9051, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.622, + 0.47, + 0.69 + ], + "angle": 0, + "content": "[5] Sungha Choi, Seunghan Yang, Seokeon Choi, and Sungrack Yun. Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. In European Conference on Computer Vision, pages 440–458. Springer, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.692, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[6] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. In Advances in Neural Information Processing Systems, pages 35799-35813. Curran Associates, Inc., 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.804, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.366, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.19, + 0.894, + 0.259 + ], + "angle": 0, + "content": "[9] Lunhao Duan, Shanshan Zhao, Nan Xue, Mingming Gong, Gui-Song Xia, and Dacheng Tao. Condaformer: Disassembled transformer with local structure enhancement for 3d point cloud understanding. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[10] Hehe Fan, Xiaojun Chang, Wanyue Zhang, Yi Cheng, Ying Sun, and Mohan Kankanhalli. Self-supervised global-local structure modeling for point cloud domain adaptation with reliable voted pseudo labels. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6367-6376, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[11] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and ZhaoXiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2918-2927, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.894, + 0.482 + ], + "angle": 0, + "content": "[12] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2704-2714, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[14] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In Advances in Neural Information Processing Systems, pages 29374-29385. Curran Associates, Inc., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[15] Edouard Grave, Moustapha M Cisse, and Armand Joulin. Unbounded cache model for online language modeling with open vocabulary. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[16] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.894, + 0.775 + ], + "angle": 0, + "content": "[17] Ahmed Hatem, Yiming Qian, and Yang Wang. Point-tta: Test-time adaptation for point cloud registration using multitask meta-auxiliary learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16494-16504, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[18] Ahmed Hatem, Yiming Qian, and Yang Wang. Test-time adaptation for point cloud upsampling using meta-learning. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1284–1291, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[19] Chao Huang, Zhangjie Cao, Yunbo Wang, Jianmin Wang, and Mingsheng Long. Metasets: Meta-learning on point sets for generalizable representations. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8859-8868, 2021. 1, 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1271" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[20] Zhe Huang, Shuo Wang, Yongcai Wang, Wanting Li, Deying Li, and Lei Wang. Roco: Robust cooperative perception by iterative object matching and pose adjustment. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7833-7842, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[21] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In Advances in Neural Information Processing Systems, pages 2427–2440. Curran Associates, Inc., 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[22] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[23] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14162-14171, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[24] Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models. In International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[25] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[26] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? Source hypothesis transfer for unsupervised domain adaptation. In Proceedings of the 37th International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[28] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[29] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In Advances in Neural Information Processing Systems, pages 44860-44879. Curran Associates, Inc., 2023. 2, 3, 5, 6, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[30] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In Advances in Neural Information Processing Systems, pages 21808-21820. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[31] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d an" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "notation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1190-1199, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In International Conference on Learning Representations, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[33] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. In International Conference on Learning Representations, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[34] M. Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[35] M. Jehanzeb Mirza, Inkyu Shin, Wei Lin, Andreas Schriebl, Kunyang Sun, Jaesung Choe, Mateusz Kozinski, Horst Possegger, In So Kweon, Kuk-Jin Yoon, and Horst Bischof. Mate: Masked autoencoders are online 3d test-time learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16709-16718, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[36] Emin Orhan. A simple cache model for image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[37] Jinyoung Park, Sanghyeok Lee, Sihyeon Kim, Yunyang Xiong, and Hyunwoo J. Kim. Self-positioning point-based transformer for point cloud understanding. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21814-21823, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.492, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[38] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc Le. Combined scaling for open-vocabulary image classification. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[39] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[40] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[41] Can Qin, Haoxuan You, Lichen Wang, C.-C. Jay Kuo, and Yun Fu. Pointdan: A multi-scale 3d domain adaption network for point cloud representation. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[43] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions. In In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1272" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "ternational Conference on Machine Learning (ICML), 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.163 + ], + "angle": 0, + "content": "[44] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.165, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[45] Yuefan Shen, Yanchao Yang, Mi Yan, He Wang, Youyi Zheng, and Leonidas Guibas. Domain adaptation on point clouds via geometry-aware implicit. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7213-7222, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.469, + 0.292 + ], + "angle": 0, + "content": "[46] Hajin Shim, Changhun Kim, and Eunho Yang. CloudFixer: Test-Time Adaptation for 3D Point Clouds via Diffusion-Guided Geometric Transformation. In European Conference on Computer Vision (ECCV), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.293, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[47] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, pages 14274-14289. Curran Associates, Inc., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[48] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[49] Hongyu Sun, Yongcai Wang, Xudong Cai, Xuewei Bai, and Deying Li. Vipformer: Efficient vision-and-pointcloud transformer for unsupervised pointcloud understanding. In IEEE International Conference on Robotics and Automation (ICRA), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[50] Hongyu Sun, Qiuhong Ke, Yongcai Wang, Wang Chen, Kang Yang, Deying Li, and Jianfei Cai. Point-prc: A prompt learning based regulation framework for generalizable point cloud analysis. In The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.59, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[51] Jiachen Sun, Qingzhao Zhang, Bhavya Kailkhura, Zhiding Yu, Chaowei Xiao, and Z. Morley Mao. Benchmarking robustness of 3d point cloud recognition against common corruptions, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.647, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[52] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[53] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. Proceedings of the AAAI Conference on Artificial Intelligence, 38(6):5171-5179, 2024. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[54] Yun-Yun Tsai, Fu-Chen Chen, Albert Y. C. Chen, Junfeng Yang, Che-Chun Su, Min Sun, and Cheng-Hao Kuo. Gda: Generalized diffusion for robust test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23242-23251, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[55] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2725-2736, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[56] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[58] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[59] Anran Wang, Jiwen Lu, Jianfei Cai, Tat-Jen Cham, and Gang Wang. Large-margin multi-modal deep learning for rgb-d object recognition. IEEE Transactions on Multimedia, 17 (11):1887-1898, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.448 + ], + "angle": 0, + "content": "[60] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.519 + ], + "angle": 0, + "content": "[61] Haiyang Wang, Chen Shi, Shaoshuai Shi, Meng Lei, Sen Wang, Di He, Bernt Schiele, and Liwei Wang. Dsvt: Dynamic sparse voxel transformer with rotated sets. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13520-13529, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[62] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[63] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q. Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8437-8445, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[64] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[65] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q. Weinberger, and WeiLun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11710-11720, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[66] Yanshuo Wang, Ali Cheraghian, Zeeshan Hayden, Jie Hong, Sameera Ramasinghe, Shafin Rahman, David Ahmedt-Aristizabal, Xuesong Li, Lars Petersson, and Mehrtash Harandi. Backpropagation-free network for 3d test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23231-23241, 2024. 2, 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1273" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[67] Xin Wei, Xiang Gu, and Jian Sun. Learning generalizable part-based feature representation for 3d point clouds. In Advances in Neural Information Processing Systems, pages 29305-29318. Curran Associates, Inc., 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[68] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7): 5092-5113, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.331 + ], + "angle": 0, + "content": "[69] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 803-814, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[70] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.471, + 0.473 + ], + "angle": 0, + "content": "[71] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 915-924, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.571 + ], + "angle": 0, + "content": "[72] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning unified representation of language, image and point cloud for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[73] Le Xue, Ning Yu, Shu Zhang, Junnan Li, Roberto Martín-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.471, + 0.729 + ], + "angle": 0, + "content": "[74] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[75] Teresa Yeo, Oğuzhan Fatih Kar, Zahra Sodagar, and Amir Zamir. Rapid network adaptation: learning to adapt neural networks using test-time feedback. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4651-4664, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[76] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In Advances in Neural Information Processing Systems, pages 32215-32234. Curran Associates, Inc., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[77] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 4, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.893, + 0.204 + ], + "angle": 0, + "content": "[78] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23783-23793, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.893, + 0.286 + ], + "angle": 0, + "content": "[79] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18123-18133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.893, + 0.343 + ], + "angle": 0, + "content": "[80] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In Advances in Neural Information Processing Systems, pages 38629-38642. Curran Associates, Inc., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.893, + 0.413 + ], + "angle": 0, + "content": "[81] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8552-8562, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.893, + 0.51 + ], + "angle": 0, + "content": "[82] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 493-510, Berlin, Heidelberg, 2022. Springer-Verlag. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.893, + 0.594 + ], + "angle": 0, + "content": "[83] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15211-15222, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.893, + 0.663 + ], + "angle": 0, + "content": "[84] Renrui Zhang, Liuhui Wang, Yali Wang, Peng Gao, Hongsheng Li, and Jianbo Shi. Starting from non-parametric networks for 3d point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5344-5353, 2023. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.893, + 0.732 + ], + "angle": 0, + "content": "[85] Weichen Zhang, Wen Li, and Dong Xu. Srdan: Scale-aware and range-aware domain adaptation network for cross-dataset 3d object detection. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6765-6775, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.893, + 0.775 + ], + "angle": 0, + "content": "[86] Zhimin Zhang, Xiang Gao, and Wei Hu. Invariantoodg: Learning invariant features of point clouds for out-of-distribution generalization, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.894, + 0.831 + ], + "angle": 0, + "content": "[87] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H.S. Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16259-16268, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.893, + 0.899 + ], + "angle": 0, + "content": "[88] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 5, 6, 7, 4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1274" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.162 + ], + "angle": 0, + "content": "[89] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1275" + } + ] +] \ No newline at end of file diff --git a/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/b2000dba-b471-4537-8cba-488e6bd7eaf3_origin.pdf b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/b2000dba-b471-4537-8cba-488e6bd7eaf3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..96a74b72e4252b13488ae9ce71e70128cb4a7b0b --- /dev/null +++ b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/b2000dba-b471-4537-8cba-488e6bd7eaf3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf757f491826a6feb6207c80cb63ecc06dcb95211cdc7a66435b501d6a49815e +size 2315808 diff --git a/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/full.md b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..621b1c38e04487793b655a46ef030311d9e8c68d --- /dev/null +++ b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/full.md @@ -0,0 +1,388 @@ +# Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis + +Hongyu Sun $^{1,2}$ , Qiuhong Ke $^{2}$ , Ming Cheng $^{2}$ , Yongcai Wang $^{1*}$ , Deying Li $^{1}$ , Chenhui Gou $^{2}$ , Jianfei Cai $^{2}$ $^{1}$ Department of Computer Science, Renmin University of China, China + $^{2}$ Department of Data Science & AI, Monash University, Australia + +$\{\mathrm{sunhongyu,ycw,deyingli}\} \mathrm{ruc.edu.cn}$ {qiuhong.ke,ming.cheng, jianfei.cai}@monash.edu + +# Abstract + +This paper proposes a general solution to enable point cloud recognition models to handle distribution shifts at test time. Unlike prior methods, which rely heavily on training data (often inaccessible during online inference) and are limited to recognizing a fixed set of point cloud classes predefined during training, we explore a more practical and challenging scenario: adapting the model solely based on online test data to recognize both previously seen classes and novel, unseen classes at test time. To this end, we develop Point-Cache, a hierarchical cache model that captures essential clues of online test samples, particularly focusing on the global structure of point clouds and their local-part details. Point-Cache, which serves as a rich 3D knowledge base, is dynamically managed to prioritize the inclusion of high-quality samples. Designed as a plug-and-play module, our method can be flexibly integrated into large multimodal 3D models to support open-vocabulary point cloud recognition. Notably, our solution operates with efficiency comparable to zero-shot inference, as it is entirely training-free. Point-Cache demonstrates substantial gains across 8 challenging benchmarks and 4 representative large 3D models, highlighting its effectiveness. Code is available at https://github.com/auniquesun/Point-Cache. + +# 1. Introduction + +In recent years, 3D sensors such as LiDARs and RGB-D cameras have been widely adopted in robotics and electric vehicles for their ability to provide reliable 3D geometry measurements [11, 16, 20, 31, 59, 61-63, 68]. Point clouds are among the most direct data formats produced by these 3D sensors. Although remarkable progress has been made in 3D point cloud recognition, the success is primarily based on the assumption that the test data and the model train + +![](images/390ff48f53c480ada107b8db48c5877d240abfb34ddd2ac6deb5d504dd339e62.jpg) +(a) ModelNet & ModelNet-C. + +![](images/e7cdb9ca047fdb5bc01e75f1994eb9b424688e3037dab17f397cabbb43fdd220.jpg) +(b) ScanObjNN & ScanObjNN-C. +Figure 1. Recognition accuracy comparison on clean and corrupted point cloud datasets. The suffix -C indicates datasets with corruptions. Models experience a severe performance drop when data corruptions arise. Point-Cache effectively narrows down the performance gap between clean and corrupted data. The hardest split of ScanObjNN is used. + +ing data are identically distributed [2, 9, 26, 32, 37, 39, 40, 64, 87]. However, this assumption is frequently violated in real-world scenarios due to complex geometries, as well as sensing and processing errors [43, 50, 51]. + +In practice, a majority of models remain vulnerable to distribution shifts, such as out-of-distribution (OOD) samples [6, 7, 50, 69], data corruptions [43, 44, 51, 56], and more. As shown in Fig. 1, notable performance gaps (e.g., $11 + \%$ ) occur when models are tested on clean (ModelNet [70]) versus corrupted data (ModelNet-C [43]). These issues drive researchers to enhance the robustness and generalization of point cloud recognition methods. Related efforts can be grouped into three main streams. The first focuses on training models that can transfer to target data. These models learn a shared feature space between source and target domains [10, 41, 45, 67], or apply meta-learning frameworks to handle geometry shifts by training multiple classifiers [19], or design learnable prompts to elicit general knowledge from pre-trained large 3D models [50]. However, possible shifts are unpredictable, and the training-based methods cannot account for all cases during training. As a result, these models are limited when new shifts arise at test time [75]. The second stream addresses the problem + +through test-time adaptation strategies [35, 46, 66]. These strategies improve performance with various adaptation techniques, such as masked auto-encoding for point clouds (MATE [35]), static prototype memory (BFTT3D [66]), and diffusion models (CloudFixer [46]). Nevertheless, we find that these approaches rely heavily on training data. For instance, MATE [35] requires pre-training on the entire training set, and BFTT3D [66] employs training data to build a static and offline prototype memory to store point cloud features. This heavy dependence on training data poses challenges, particularly when such data is unavailable or when distribution shifts occur at test time. In these cases, revisiting the training data may not be helpful to adapt the model to online test data. Additionally, these test-time methods can only recognize classes seen during training, which limits their practicality in generalizable point cloud analysis. + +Inspired by advancements in language-guided image recognition [22, 38, 42, 79], the third branch of works utilizes contrastive pre-training on large-scale triplets (point, text, image) to connect flexible language descriptions and enable open-vocabulary point cloud recognition [29, 72, 73, 88]. These large multimodal 3D models demonstrate promising zero-shot performance and provide a new paradigm for generalizable point cloud analysis. However, zero-shot predictions are performed independently for each test sample, with the final accuracy simply averaged across the test set. This approach does not fully exploit the distribution of test data, particularly the statistics of reliable predictions from large 3D models, which could further enhance accuracy and generalization. At this point, we pose the following question: can we take advantages of both the promising zero-shot ability of large 3D models and test-time adaptation techniques to unlock more robust and generalizable point cloud recognition? + +In this work, to fully exploit the data distribution during test time, we develop a cache model to store critical fingerprints of test samples. These fingerprints consist of key-value pairs derived by large multimodal 3D models, where the key represents the sample feature extracted by the point encoder, and the value represents the predicted (pseudo) label from the model. This cache model is entirely constructed from online test data without accessing any training samples. Our approach is thus more practical and challenging than prior methods [10, 19, 35, 41, 45, 46, 50, 66, 67]. Our cache model is hierarchical, as it stores not only the global features of a point cloud in a global cache but also its local-part details in a local cache, which are essential for distinguishing subtle differences among classes. We also empirically demonstrate that this coarse-to-fine design enhances robustness and generalization in the wild. Additionally, our cache model is dynamic, since we need to continuously update the cache with online test data to prioritize high-quality key-value pairs. The constructed cache + +serves as a powerful 3D knowledge base, capturing key characteristics of online test data. When a new sample arrives, we query the knowledge base, compute the affinity between the sample and cached keys, and produce a prediction by weighting the cached (pseudo) labels accordingly. Our method is training-free and flexible, requiring only a pre-trained 3D model and test data. More details will be provided in Sec. 3. + +The superiority of the proposed method is supported by comprehensive comparisons with previous strong baselines across 8 challenging benchmarks, containing up to 1,156 point cloud classes. Moreover, we implement the hierarchical cache as a plug-and-play module and integrate it into representative large 3D models, such as ULIP [72], ULIP-2 [73], OpenShape [29] and Uni3D [88], to enhance test-time point cloud recognition dynamically. The proposed cache model brings notable and consistent gains over the baselines. For instance, our model yields $+6.18\%$ and $+4.84\%$ absolute accuracy improvements average over 7 types of corruptions in ModelNet-C [43] based on ULIP [72] and ULIP-2 [73]. On Objaverse-LVIS [7] that includes 1,156 categories, Point-Cache boosts the accuracy by $+2.10\%$ absolute points compared to the zero-shot ULIP-2 [73]. In summary, the contributions of our paper are threefold. + +- To our knowledge, Point-Cache is the first to explore test-time point cloud recognition in a more practical and challenging setting, where only online test data is available and the model needs to generalize to both known classes and unseen new classes, which is beyond the capability of prior test-time methods. +- We propose a hierarchical cache model to capture essential distribution clues of online test samples, creating an accurate profile for various point clouds from global to local perspectives. We implement the hierarchical cache as a plug-and-play component, enabling diverse 3D backbones to achieve enhanced test-time performance. +- Comprehensive experiments and consistent gains across 8 public datasets and 4 large 3D models demonstrate the efficacy of Point-Cache. + +# 2. Related Work + +Test-time adaptation in 3D point cloud analysis receives relatively less attention compared to the 2D image community [5, 12, 14, 21, 23, 27, 28, 30, 34, 47, 52, 54, 60, 78, 80]. This technique has recently emerged in 3D point cloud tasks, including recognition [35, 46, 66], registration [17], upsampling [18] and object detection [65, 74, 85]. Among these, MATE [35], BFTT3D [66], and CloudFixer [46] are specifically focused on test-time point cloud recognition and are thus closely related to our work. However, as previously analyzed, these methods depend heavily on training data from the source domain to construct their pipelines, + +![](images/3ca772f4516ab57674d72aabfc6ae6c3055f7993462a8a11bfdf0e2c4ddb4a95.jpg) +Figure 2. The overall pipeline of Point-Cache. The zero-shot predictions $\hat{\mathbf{y}}_{zs}$ of large 3D models are effectively adapted by our global cache logits $\hat{\mathbf{y}}_g$ and local cache logits $\hat{\mathbf{y}}_l$ to handle the distribution shifts, enabling robust and generalizable point cloud analysis. + +which may not be available or helpful to understand target data distributions at test time. Furthermore, they lack effective strategies to leverage the distribution of test samples. By contrast, Point-Cache is constructed entirely from online test data without access to any training samples, making our setting both more challenging and more practical. Additionally, our cache model, powered by large 3D models, can generalize to both known classes and unseen new categories, an advancement that prior point cloud adaptation/generalization frameworks have not achieved [10, 19, 35, 41, 45, 46, 66, 67]. + +Cache models that store sample features in memory for further reference and reuse have been explored in fields like language modeling [15, 24, 33, 58], image classification [3-5, 13, 21, 23, 36, 48, 55, 58, 82, 83] and point cloud recognition [53, 66, 84]. Our proposed cache model stands out from previous ones in several key aspects: (1) Our cache model is constructed based on test samples whereas previous cache models rely on a large training set for construction, as in Tip-Adapter [82], CaFo [83], PointNN [84], Point-PEFT [53] and BFTT3D [66]. (2) Our cache model is hierarchically designed to capture both global and local-part features of a point cloud object, while existing methods store only global representations [3-5, 13, 21, 48, 53, 58, 66, 82-84]. We empirically demonstrate that this coarse-to-fine design improves the capture of target data features, thereby enhancing test-time performance. (3) Our cache model is dynamically updated with online test samples, incorporating a selective mechanism to ensure high-quality key-value pairs replace less reliable ones when the cache reaches its capacity. In contrast, prior cache models, which are constructed offline using the entire training set, are static and consume a substantial amount of memory [53, 66, 84]. + +Large Multi-modal 3D Models [25, 29, 72, 73, 76, 81, 88, 89] spark a wave of open-vocabulary point cloud understanding by pre-training on large-scale datasets [7] in recent years. For instance, ULIP-2 [73], OpenShape [29] and Uni3D [88] pre-train the 3D encoders on million-scale (point, image, text) triplets using contrastive learning. These pre-trained point cloud encoders are then transferred to downstream 3D tasks like classification and retrieval, demonstrating strong zero-shot performance that surpasses traditional 3D models [32, 39, 40, 64, 71, 87] which are typically tailored to specific datasets [1, 56, 70]. This new paradigm presents a promising direction for generalizable point cloud processing. Our work is complementary to these models: rather than developing another large 3D model, we build upon their capabilities to enhance robustness and generalization for point cloud analysis at test time. Our hierarchical cache, backed by these large 3D models, generalizes to unseen classes beyond training, addressing a significant limitation of current test-time methods for point cloud recognition [35, 46, 66]. + +# 3. Method + +This section elaborates on the details of Point-Cache, the overall pipeline of our approach is depicted in Fig. 2. + +Preliminary. We begin by reviewing the zero-shot inference process of existing large multimodal 3D models [29, 72, 73, 81, 88, 89]. These models use separate encoders to map inputs from a point cloud $P \in \mathbb{R}^{N \times 3}$ and text $T$ (e.g., "a point cloud object of a {class}") to a shared feature space. Denoting the point cloud encoder as $f_{p}(\cdot)$ and the text encoder as $f_{t}(\cdot)$ , we can obtain the global feature $\mathbf{e}_p^g = f_p(P) \in \mathbb{R}^d$ of the point cloud $P$ and the text + +feature $\mathbf{e}_t = f_t(T) \in \mathbb{R}^d$ , where $d$ is the feature dimension. To recognize the input point cloud $P$ in a zero-shot manner, these models replace “ $\{\mathrm{class}\}$ ” in text $T$ with each of possible $C$ category names and compute the class probability $\hat{\mathbf{y}}_{zs}$ using Eq. 1. + +$$ +\hat {\mathbf {y}} _ {z s} = \{\hat {y} _ {i} | _ {i = 1} ^ {C} \}, \quad \hat {y} _ {i} = \frac {\exp (s i m (\mathbf {e} _ {p} ^ {g} , \mathbf {e} _ {t} ^ {i}) / \tau)}{\sum_ {j = 1} ^ {C} \exp (s i m (\mathbf {e} _ {p} ^ {g} , \mathbf {e} _ {t} ^ {j}) / \tau)} \tag {1} +$$ + +Here $\mathbf{e}_t^i$ indicates the text encoding of the $i$ -th category. $sim(\cdot, \cdot)$ measures the cosine similarity between inputs, and $\tau$ is a temperature coefficient. Finally, the point cloud $P$ is classified into the category with the highest probability, indexed by $\hat{L} = \arg \max_{i} \{\hat{y}_i|_{i=1}^C\}$ . + +As described in Sec. 1, our hierarchical cache stores the fingerprints of online test samples, which consists of key-value pairs. The key is the test sample feature $\mathbf{e}_p^g$ extracted by the point encoder of a pre-trained large 3D model, and the value is the class label $\hat{L}$ predicted by the zero-shot inference. In addition, we compute the entropy $h = -\sum_{i=1}^{C} \hat{y}_i \log \hat{y}_i$ of the zero-shot prediction $\hat{\mathbf{y}}_{zs}$ , as it is essential to assess the confidence with which the large 3D model produces this prediction. This entropy value is employed to filter out high-quality samples in our cache model. + +# 3.1. Text & Point Cloud Encoding + +We can obtain the text encoding $\mathbf{e}_t$ of the input $T$ and the global feature $\mathbf{e}_p^g$ of a point cloud $P$ as introduced in the preliminary. However, it is non-trivial to represent the local-part features $\mathbf{e}_p^l$ of the point cloud $P$ due to the following challenges analyzed in Fig. 3: + +# 01 Part Feature Capture + +- Difficult to define parts precisely and get their features for various 3D objects of different classes + +# 02. Part Feature Storage + +- Expensive to store multiple part features instead of a global feature for a single point cloud + +Figure 3. Challenges in encoding part feature for various 3D objects of different classes: Part Feature Capture & Storage. + +To address these challenges, we made a careful analysis and developed effective solutions, seeing Fig. 9 in the Supplementary for intuitive illustration. (1) A straightforward approach might involve using an additional pretrained segmentation network to segment point cloud parts, but this would be computationally prohibitive during online inference. Instead, we propose a fuzzy definition of parts. Specifically, since the architecture of the point encoder of large 3D models is based on Transformer [8, 57, 77] and the input is a sequence of point patches, we avoid strict part definitions according to physical semantics (e.g., head, body, wing, and tail in an airplane), and instead treat the output + +point patches from the last encoder layer as local part representations. (2) A second challenge is that a long point patch sequence can become resource-intensive; for example, the point encoder in ULIP-2 [73] generates 512 point patches per object, which would require substantial memory. To mitigate this problem, we propose to summarize these point patches into $m$ centers $\mathbf{e}_p^l \in \mathbb{R}^{m \times d}$ using K-Means clustering ( $m < 10$ ), reducing memory usage by two orders of magnitude compared to storing all patches. + +# 3.2. Construction & Update of Hierarchical Cache + +To thoroughly explore the distribution of test data, we develop a hierarchical cache to record the fingerprints of online test samples. This hierarchical cache includes both a global cache, which stores overall information, and a local cache, which captures detailed part-level information for each point cloud. The coarse-to-fine structure enables more precise profiling of test data and distinguishes our approach from previous cache models used in image recognition [3-5, 13, 21, 23, 36, 48, 55, 58, 82, 83] and point cloud analysis [53, 66, 84]. The complete construction process of hierarchical cache is outlined in Alg. 1 of the Supplementary. + +Global cache is represented by $\mathbf{C}_g = (\mathbf{E}_g, \hat{\mathbf{L}}_g, \mathbf{h}_g)$ , storing (feature, label, entropy) triplets for a set of point cloud objects. To manage storage efficiently, we design a selection mechanism to identify high-quality test samples to place into the cache. + +Construction: We set an upper bound $K$ on the number of samples per class in the cache, allowing at most $K$ fingerprints (shots) can be placed in each category. Initially, the global cache is empty. When a new test sample $P$ arrives, we generate its fingerprint $(\mathbf{e}_p^g, \hat{L}, h)$ and assign it to class $\hat{L}$ . + +Update: With the number of online samples grows, we check whether the number of shots for class $\hat{L}$ has hit the upper bound. If it is not the case, the fingerprint of current sample is appended to class $\hat{L}$ . Otherwise, we locate the fingerprint with highest entropy within class $\hat{L}$ and compare it to the entropy of current test sample. If the current test sample has a lower entropy, we replace the highest-entropy fingerprint with the new one; otherwise, the current sample is skipped. When the cache is full, $\mathbf{E}_g$ has dimensions $CK\times d$ , $\hat{\mathbf{L}}_g$ has dimensions $CK\times 1$ , and $\mathbf{h}_g$ has dimensions $CK\times 1$ . + +Local cache. Although the global feature of a point cloud captures expressive information, it lacks the local details of the 3D object, which are critical for point cloud recognition, especially when shapes share an overall appearance but differ in local parts. The local cache addresses this limitation by storing part-level information. We denote the local cache as $\mathbf{C}_l = (\mathbf{E}_l,\hat{\mathbf{L}}_l)$ , which store the fingerprints of local parts of point cloud objects. Here $\mathbf{E}_l$ represents part features of the test samples and $\hat{\mathbf{L}}_l$ indicates their labels. + +Accordingly, we define the local-part fingerprint of a test sample $P$ as $(\mathbf{e}_p^l,\hat{L})$ , where $\mathbf{e}_p^l\in \mathbb{R}^{m\times d}$ corresponds to local-part features and $\hat{L}$ is the class label of $P$ . + +Construction & Update: Note that the sample selective mechanism of the local cache follows the same rules as the global cache, meaning the local-part fingerprint of a test sample is only stored if it qualifies for inclusion in the global cache. So there is no need to record the entropy $h$ again in the local cache. When the cache is full, $\mathbf{E}_l$ has dimensions $(m \cdot CK) \times d$ and $\hat{\mathbf{L}}_l$ has dimensions $CK \times 1$ . + +# 3.3. Test-time Adaptation by Hierarchical Cache + +With the hierarchical cache serving as a rich knowledge base of online point cloud data, we can adapt the zero-shot predictions of large 3D models for new samples, effectively addressing distribution shifts that occur at test time. The adaptation follows a coarse-to-fine process as below. + +Test-time adaptation by global cache. During adaptation, a new sample $Q$ is treated as a query to retrieve knowledge from the global cache, which we use to produce the adaptation logits. Specifically, the retrieval process is simulated by calculating the affinity $A_{g}$ between the query and cached features $\mathbf{E}_{g}$ . This allows the new sample to establish connections with cached high-quality samples, providing it with a contextual understanding of the online test data. We then generate the adaptation logits $\hat{\mathbf{y}}_{g}$ by weighed ensembling of cache labels $\hat{\mathbf{L}}_{g}$ , as shown in Eq. 2. + +$$ +A _ {g} = \exp \left(- \beta_ {g} \left(1 - \mathbf {e} _ {q} ^ {g} \mathbf {E} _ {g} ^ {\top}\right)\right), \hat {\mathbf {L}} _ {g} = \mathrm {O H} (\hat {\mathbf {L}} _ {g}), \hat {\mathbf {y}} _ {g} = A _ {g} \hat {\mathbf {L}} _ {g} \tag {2} +$$ + +where $\mathbf{e}_q^g\in \mathbb{R}^{1\times d}$ represents the global feature of sample $Q$ and $\mathbf{E}_g\in \mathbb{R}^{CK\times d}$ denotes the cached keys. The 'OH' operation transforms $\hat{\mathbf{L}}_g$ into a one-hot encoding. The term $\mathbf{e}_q^g\mathbf{E}_g^\top \in \mathbb{R}^{1\times CK}$ measures the cosine similarity between the query and all keys, with $\beta_{g}$ acting as a hyper-parameter to modulate the sharpness and the exponential operation ensuring similarity values fall within (0, 1]. The affinity $A_{g}\in \mathbb{R}^{1\times CK}$ , which grasps the relationship between the query and cached keys, is utilized to ensemble the transformed one-hot labels $\hat{\mathbf{L}}_g\in \mathbb{R}^{CK\times C}$ , yielding global cache adaptation logits $\hat{\mathbf{y}}_g\in \mathbb{R}^{1\times C}$ . + +Test-time adaptation by local cache. While global cache adaptation is effective, it may not handle fine-grained distribution shifts at test time. To achieve a more precise adjustment, we exploit local part knowledge. The local-part features $\mathbf{e}_q^l$ of the point cloud $Q$ serve as queries to the local cache. During retrieval, we compute the affinity $A_{l}$ between $\mathbf{e}_q^l$ and all cached part features $\mathbf{E}_l$ . The affinity $A_{l}$ is then used to weight the cache labels $\hat{\mathbf{L}}_l$ , producing local cache adaptation logits $\hat{\mathbf{y}}_l$ as in Eq. 3. + +$$ +A _ {l} = \exp \left(- \beta_ {l} \left(1 - \mathbf {e} _ {q} ^ {l} \mathbf {E} _ {l} ^ {\top}\right)\right), \hat {\mathbf {L}} _ {l} = \mathrm {O H} (\hat {\mathbf {L}} _ {l}), \hat {\mathbf {y}} _ {l} = \phi \left(A _ {l} \hat {\mathbf {L}} _ {l}\right) \tag {3} +$$ + +where $\mathbf{e}_q^l\in \mathbb{R}^{m\times d}$ , $\mathbf{E}_l\in \mathbb{R}^{(m\cdot CK)\times d}$ , and $\hat{\mathbf{L}}_l\in \mathbb{R}^{(m\cdot CK)\times C}$ . The function $\phi (\cdot)$ is an average pooling operation over $m$ parts to produce class logits. The term $e_q^l\mathbf{E}_l^\top \in \mathbb{R}^{m\times (m\cdot CK)}$ measures the cosine similarity between the part features of sample $Q$ and all part features in the cache. Likewise, the similarity is modulated by the coefficient $\beta_{l}$ and scaled with the exponential function, resulting in the affinity $A_{l}\in \mathbb{R}^{m\times (m\cdot CK)}$ . The term $A_{l}\hat{\mathbf{L}}_{l}\in \mathbb{R}^{m\times C}$ represents the class logits for all $m$ parts, which are transformed by a pooling operation $\phi (\cdot)$ to derive the final local cache adaptation logits $\hat{\mathbf{y}}_l\in \mathbb{R}^{1\times C}$ . + +Overall prediction combines three components as shown in Eq. 4. The zero-shot prediction $\hat{\mathbf{y}}_{zs}$ of large multimodal 3D model is effectively adapted using our global cache prediction $\hat{\mathbf{y}}_g$ and local cache prediction $\hat{\mathbf{y}}_l$ , balanced by the coefficients $\alpha_{q}$ and $\alpha_{l}$ . + +$$ +\hat {\mathbf {y}} = \hat {\mathbf {y}} _ {z s} + \alpha_ {g} \hat {\mathbf {y}} _ {g} + \alpha_ {l} \hat {\mathbf {y}} _ {l} \tag {4} +$$ + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. To evaluate the robustness of point cloud recognition, we choose four public datasets that contain various point cloud corruptions: ModelNet-C [43] and three variants of ScanObjectNN-C [56]. ModelNet-C [43] introduces 7 atomic corruptions, including adding global outliers, adding local outliers, dropping global structure, dropping local parts, rotation, scaling, and jittering. Other corruptions can be derived from these atomic corruptions, as discussed in [43]. We apply these atomic corruptions to the three variants of ScanObjectNN [56] and produce their corrupted versions. To assess generalization on unseen new data, we test the proposed method on four challenging benchmarks: OmniObject3D [69] (216 classes), ObjaverseLVIS [7] (1,156 classes), the hardest variant of ScanObjectNN [56] and widely used ModelNet40 [70]. Recognition accuracy (\%) is the primary evaluation metric. + +Implementation Details. We choose ULIP [72], ULIP-2 [73], OpenShape [29] and Uni3D [88] as the large multimodal 3D models for experiments. Their pre-trained weights are loaded and then frozen. Global and local-part features are generated by modifying the forward pass function without altering the model architecture. The maximum number of samples per class in the cache, $K$ , is set to 3, and each point cloud is clustered into $m = 3$ local parts. Both the balance factors $\alpha_{g}$ and $\alpha_{l}$ are set to 4.0 while the coefficients $\beta_{g}$ and $\beta_{l}$ are set to 3.0. We will examine the design choices in the ablation studies. To optimize memory and computation, we convert the input point clouds and model parameters into 'dtype' fp16 in Pytorch. All experiments are conducted on two 4090 GPUs. Additional details can be found in the Supplementary. + +Table 1. Comparison of recognition accuracy on ModelNet-C that contains 7 types of corruptions. Each clean point cloud has 1024 points and the corruption severity level is 2. The last column is the average over 7 corruption types. The best results are in bold and the second best are underlined. The setting applies to the following tables unless otherwise specified. + +
MethodClean Data ModelNetCorruption TypeAvg.
Add GlobalAdd LocalDrop GlobalDrop LocalRotateScaleJitter
ULIP [72]56.1633.5543.9254.7050.8955.2750.2044.0847.52
+Global Cache(Ours)62.1245.7947.9856.8553.8960.2554.3448.9152.56
+Hierarchical Cache(Ours)64.2246.1547.8559.1656.0061.4755.3549.9253.70
ULIP-2 [73]71.2365.1554.6268.7657.9870.3067.1021.7657.95
+Global Cache(Ours)73.9567.0259.3271.3561.5972.3768.4028.2061.18
+Hierarchical Cache(Ours)74.5368.1161.2673.2263.6573.3470.4229.5062.79
O-Shape [29]84.5671.6467.7981.5673.5882.0178.4859.3673.49
+Global Cache(Ours)84.5274.7272.7782.4175.1283.1878.9367.9176.43
+Hierarchical Cache(Ours)84.0474.8473.7082.2176.2682.6678.1268.3576.59
Uni3D [88]81.8172.4556.3668.1567.1879.9475.3656.2467.95
+Global Cache(Ours)83.1476.1366.4971.4369.8181.5275.8561.4371.81
+Hierarchical Cache(Ours)83.8777.5171.1572.1670.7581.7777.3162.5273.31
+ +Table 2. Comparison of recognition accuracy across a suite of datasets. S-PB_RS_T50 is the hardest split of ScanObjectNN. OLVIS: Objverse-LVIS. Omni3D: OmniObject3D. The number under each dataset indicates the number of points (pts) for each object. In Omni3D, each point cloud can be represented by a different number of points. The last column is the average accuracy on these datasets. + +
MethodModelNet40 (10000 pts)S-PB_RS_T50 (2048 pts)O-LVIS (10000 pts)Omni3DAvg.
(1024 pts4096 pts16384 pts)
ULIP [72]58.7546.446.248.397.757.2822.48
+Global Cache(Ours)61.2250.217.0210.009.368.4324.37
+Hierarchical Cache(Ours)62.9351.807.0210.479.758.9025.15
ULIP-2 [73]72.9747.1330.2626.3629.2026.5838.75
+Global Cache(Ours)74.5151.7032.6528.5131.1028.5341.17
+Hierarchical Cache(Ours)75.5354.9832.3629.3731.2429.4442.15
Uni3D [88]88.4165.1955.4231.5241.9841.8654.09
+Global Cache(Ours)88.8668.5153.3634.9745.1345.1956.00
+Hierarchical Cache(Ours)89.1868.2455.1935.8245.6045.8956.65
+ +# 4.2. Test-time Robustness and Generalization + +Robustness. The robustness analysis of our training-free test-time hierarchical caches takes place on 4 point cloud corruption datasets [43, 56]. As observed in Tab. 1, the zero-shot accuracy of various large 3D models on ModelNet-C is considerably boosted by our global cache model, with increases such as $+5.04\%$ for ULIP [72], $+3.23\%$ for ULIP-2 [73], $+2.94\%$ for OpenShape [29] and $+3.86\%$ for Uni3D [88], averaged across 7 corruption types. By combining global and local cache together, our hierarchical cache model attains further improvements, highlighting the benefits of incorporating a local cache. Notably, these gains extend beyond corrupted data, as our hierarchical cache model also improves recognition accuracy on the clean ModelNet dataset, e.g., $+8.06\%$ absolute increase over ULIP's zero-shot performance. An exception is OpenShape, which maintains a slightly higher zero-shot accuracy on clean data than our cache model. However, when considering the 7 corruption types, our global and hi + +erarchical cache models outperform OpenShape by $2.94\%$ and $3.10\%$ , respectively, demonstrating improved robustness against data corruption at test time. Consistent improvements are also showcased on the variants of ScanObjectNN, shown in Tab. 5, 6 and 7 in the Supplementary. + +Generalization. We evaluated the generalization capability of Point-Cache on four representative benchmarks that were unseen during the pre-training of the large 3D models. These benchmarks generally do not contain data corruptions. While ModelNet40 [70] and ScanObjectNN [56] contain a limited number of classes, OmniObject3D [69] and Objverse-LVIS [7] represent a broader range of 3D concepts, covering 216 and 1,156 classes, respectively. As shown in Tab. 2, Point-Cache consistently enhances the recognition accuracy across datasets and 3D backbones. For instance, our hierarchical cache built on ULIP and ULIP-2 yields absolute improvements of $6.31\%$ and $5.51\%$ , respectively, average over 6 scores. On ScanObjectNN, our hierarchical cache significantly raises the zero-shot prediction of ULIP and ULIP-2, from $27.20\%$ to $51.80\%$ for ULIP + +and from $36.22\%$ to $54.98\%$ for ULIP-2. On the challenging Objaverse-LVIS (O-LVIS), our global cache with ULIP-2 successfully realizes a $2.39\%$ absolute gains in zero-shot accuracy across 1,156 classes, a notable achievement. Furthermore, our cache model enables Uni3D to attain an $89.18\%$ accuracy on ModelNet40 without any training, effectively matching the fully-supervised PointNet $(89.20\%)$ . + +# 4.3. Memory Usage and Throughput + +Memory. We use Uni3D [88] as the baseline and adapt its zero-shot predictions with Point-Cache. The experiments are conducted on 3 datasets as shown in Tab. 3. The batch size is set to 1 because we need to update the cache and adapt the class logits according to every single sample. The results indicate our global cache and hierarchical cache consume similar or slightly higher memory than the baseline on each dataset. Furthermore, as the number of classes increases rapidly, the memory usage grows at a slower rate. This is primarily because the memory consumption is dominated by the large number of parameters in Uni3D (e.g., 1,016.5M). In contrast, the hierarchical cache built on OLVIS uses only approximately 7.1M parameters, which is negligible compared to Uni3D. Detailed calculations are provided in the Supplementary. + +Table 3. Comparison of memory usage (MB). The batch size is set to 1 and the used device is an RTX 4090. The number under each dataset indicates #Classes. Hierar: Hierarchical Cache. + +
MethodModelNet-C (40)Omni3D (216)O-LVIS (1156)#Params (M)
Uni3D5,0625,0625,0621,016.5
+Global(Ours)5,0625,0645,0701,016.5
+Hierar(Ours)5,0645,0685,0901,016.5
+ +Throughput. We evaluate the throughput of Point-Cache on ModelNet40, where throughput is measured as the number of test samples processed per second (abbreviated as $t / s$ ). The results are reported in Tab. 4. Point-Cache runs slightly slower than zero-shot inference due to the additional overhead introduced by cache updates, logits computation, etc. However, the throughput degradation is minimal, e.g., a $0.04 \, t / s$ drop for OpenShape with the global cache, and a $0.07 \, t / s$ drop for ULIP-2 with the hierarchical cache. These results prove that Point-Cache brings remarkable accuracy improvements with negligible computational cost. + +# 4.4. Ablation Study + +We use Uni3D with our hierarchical cache and 4 public datasets for ablation analysis, as shown in Fig. 4. + +Shot size per class in the global cache. The upper bound $K$ on the number of samples per class is a critical variable + +Table 4. Comparison of throughput $(t / s)$ for different models on ModelNet40. The batch size is set to 1 and the device is a 4090 GPU. The results are averaged over all test samples. + +
MethodZero-shot+Global Cache+Hierar Cache
ULIP11.2511.2111.18
ULIP-211.1811.1411.11
OpenShape9.809.769.74
Uni3D9.789.759.73
+ +as it affects the total size of the hierarchical cache. We analyze its impact on the test-time recognition accuracy on 4 datasets, shown in Fig. 4a. As $K$ increases, accuracy trend differ across datasets; for example, accuracy increases on Omni3D but decreases on S-PB_T50_RS. We adopt $K = 3$ as it produces the best average accuracy (63.98%). The results indicate that our cache model does not require a large number of samples to achieve strong performance. + +Number of local parts per 3D object. To optimize memory and computation, we cluster hundreds of point patches of a 3D object into $m$ parts. We investigate the effect of this variable on test-time performance, as exhibited in Fig. 4b. For three out of the four datasets, namely S-PB_T50_RS, Omni3D, and O-LVIS, accuracy generally decreases as the number of parts increases, suggesting that a point cloud does not need to be represented with an excessive number of parts. We choose $m = 3$ by default, as this setting reaches the best average accuracy across the 4 datasets. + +![](images/3d50fac6fb5374b8f54c7e8633e362ea56581ba805c8dfd380cb4c885ca85db3.jpg) +(a) shot size $K$ + +![](images/5ece0fcdfedac25c6712cece2c89841ee3708073de6ac2b219f6334ed8e01072.jpg) +(b) $m$ parts +Figure 4. Ablation studies on the hyper-parameters in the cache design, including the shot size $K$ per class, the number of parts $m$ per object, the balance factors in the final prediction logits and the sharpness coefficients in affinity computation. + +![](images/6109cecf29b549bc36eb9a36123cd2f26d6ba1c536cb083225f17b3ad98efae0.jpg) +(c) $\alpha_{g}$ and $\alpha_{l}$ + +![](images/4edbd5e95bdef17bf6a07aa570830d92308e27d9fde2c24acc8f84316addb291.jpg) +(d) $\beta_{g}$ and $\beta_{l}$ + +The balance factors. The factors $\alpha_{g}$ and $\alpha_{l}$ control the relative weights of the global cache prediction $\hat{\mathbf{p}}_g$ and local cache prediction $\hat{\mathbf{p}}_l$ . To simplify experimentation, we set $\alpha_{g} = \alpha_{l}$ in the ablation study. Candidate values are $\{0.5, 1.0, 2.0, 3.0, 4.0, 5.0\}$ . As shown in Fig. 4c, increasing these factors improves performance on S-PB_T50_RS and Omni3D, while it has minimal effect on ModelNet40 and O-LVIS. We set the factors to 4.0, which yields the optimal average accuracy across all four datasets. + +The sharpness coefficients. The coefficients $\beta_{g}$ and $\beta_{l}$ are used to modulate the sharpness of the query-key affinity. Here we analyze their impact on test-time accuracy on 4 benchmarks. Likewise, we let $\alpha_{g} = \alpha_{l}$ and the candidate values are $\{0.5,1.0,3.0,5.0,7.0,9.0\}$ . As Fig. 4d displays, + +accuracy initially improves slightly with increasing sharpness but then decreases as sharpness becomes too high. We set the sharpness to 3.0 that yields the optimal result average over 4 datasets. + +# 4.5. Visualization + +Online Inference. We visualize the online inference process of Point-Cache using different large 3D models and compare three model variants: zero-shot prediction, the model with our global cache, and the model with our hierarchical cache, referring to Fig. 5. During inference, we calculate the average recognition accuracy over accumulated samples. Due to the small number of samples at the initial stage, the average accuracy may fluctuate drastically, so we discard the statistics for the first 5 samples. Overall, methods employing the hierarchical cache consistently outperform other variants by a clear margin across various large 3D models and datasets. + +![](images/440b50728ed305619b42439be57f9c1a5148c3761d22b4f62467eca245965696.jpg) +(a) ULIP. + +![](images/3d2a831c5e67b2bd1b433629862998a689f8b2c507112e6e2157973f98e03fe8.jpg) +(b) ULIP-2. + +![](images/b2f6617e09c641732dc4e1f4053dcdd9997d714bcd920fc6cf63321534fc20b0.jpg) +(c) OpenShape. + +![](images/33acc4474aff6e357a76eaa0ebbec68112e2f9519cc2095162ad15676ab68582.jpg) +(d) Uni3D. +Figure 5. The average recognition accuracy of accumulated samples during online inference. The curve changes significantly in the initial stage due to the small number of samples. Models with our global and hierarchical cache receive perceptible performance gains. + +Entropy and Accuracy. Fig. 6 illustrates the relationship between entropy and accuracy for different models on various datasets. During online inference, we first compute the entropy of top 5 predictions for each test sample and then average the entropy across all accumulated samples. Similarly, the corresponding average accuracy can be calculated. The results reveal that Point-Cache effectively reduces the uncertainty of the zero-shot predictions made by large 3D models, enhances prediction confidence, and consequently lifts recognition accuracy. Additionally, we observe that the hierarchical cache model is more effective in reducing entropy compared to the global cache model, further underscoring the importance of incorporating the local cache. + +Qualitative Analysis. Fig. 7 presents the step-by-step adaptation process of Point-Cache. The first three rows demonstrate cases where the global cache model attempts to adjust the zero-shot predictions of various large 3D models but fails. Instead, the hierarchical cache model successfully refines the class logits based on the global cache. In the last row, our global cache model corrects the zero-shot prediction in the first step, and the hierarchical cache model subsequently retains the top class while increasing its confidence. The examples showcased here are not cherry-picked. + +![](images/dfae67d8bc2bd2ef47512d2147e69fda9f45bc3e08f427b575c2cad011c7e3ed.jpg) +(a) Group One. + +![](images/ea4a105ba81065854a08090615625325c57c373aa7f6c779134a0360484f40f6.jpg) +(b) Group Two. +Figure 6. Relationship between entropy and accuracy. Models with our global and hierarchical cache consistently exhibit lower entropy and higher accuracy across various benchmarks. + +Additional quantitative and qualitative results can be found in the Supplementary. + +![](images/f7c4c6c897456a02f83d8d4c42b4b202a92d64aaf3342f89f726e428386c7e2b.jpg) +(a) Ground truth. + +![](images/e171fc5b94492d1120b78cb77f7a4d235b3f748196acddf9af6f68f94187f660.jpg) +(b) ULIP-2. + +![](images/541d690dabe490f921f4b8ff4d2de598a796c28ae3687f753cbbe4c006bf82af.jpg) +(c) ULIP-2+GC. + +![](images/0f474d2b4bc06b1aabcd1b45f58def868e61af1df741c04e8f68da5233d5d9be.jpg) +(d) ULIP-2+HC. + +![](images/6ec08d92dc0fbdc5f08019b6cbc0ecc7299b6af2faf5d5172371f405ae327bfa.jpg) +(e) Ground truth. + +![](images/6d2815006760c86947ebb23a96c4c4073489f681b1cfac8826049d37d72af20a.jpg) +(f) ULIP. + +![](images/cda03881c1794279a3e0c44c51b6159973f9d3e094f6420e2af978771056e0c8.jpg) +(g) $\mathbf{ULIP} + \mathbf{GC}$ + +![](images/5b8edd9672bad0310a6a5fe8be5fbb81684266b27d00e682112af604761596e1.jpg) +(h) ULIP+HC. + +![](images/e2fba44e82830eae34d45912ee571c2ec993c1bd55ec74a93e2839fc6b1dfcf1.jpg) +(i) Ground truth. + +![](images/c913cbec7698e2cfed8cca4bd072891d98c4e2b3b728e250a60c65b5159ce646.jpg) +(j) O-Shape. + +![](images/7c2af08b62cb0cf90275ddfa836cae7097e71a2c9a4a0f994d43789bc27ddbd4.jpg) +(k) O-Shape+GC. + +![](images/94e02d96d5e670abf5f9bedb16766283f106932a0a94cb1021941ce1e43413ab.jpg) +(1) O-Shape+HC. + +![](images/7041b5e4a0499d37f542d31007dbf6b1c0b7e1979ae0ea9a238840a5693762e0.jpg) +(m) Ground truth. +Figure 7. The zero-shot predictions before and after adaptation by Point-Cache. GC: global cache. HC: hierarchical cache. + +![](images/cde187005839dff345c5eb5eceaa48900ed362a6b62a315c87e7ffaa12cd6b31.jpg) +(n) Uni3D. + +![](images/3e3ba34ddb4567227316d2fc8e8387cbc2aaa2883f031822f0c504a47da74daa.jpg) +o) $\mathrm{Uni3D + GC}$ + +![](images/bf042d04e709a9b9597c9c17b4a5e92114c5e44f426cc4a0969ce5953e8708fa.jpg) +(p) $\mathrm{Uni3D + HC}$ + +# 5. Conclusion + +This paper explores test-time point cloud recognition in a practical and challenging setting, where only test data and the pre-trained models are available. We develop PointCache, a powerful 3D knowledge base that enhances robustness and generalization at test time. It records the fingerprints of online test samples through a global and a local cache, creating a more accurate profile of 3D data. Designed as a plug-and-play module, Point-Cache can be seamlessly integrated into various large 3D models without compromising efficiency, enabling open-vocabulary point cloud recognition for both known and unseen classes, effectively addressing limitations of prior test-time methods. + +# Acknowledgement + +We thank all anonymous reviewers and area chairs for their time and valuable feedback. This work was partially supported by China Scholarship Council (CSC) under the Grant No. 202306360147 and Australian Research Council funding under the Grant No. DE250100030. Dr. Qiuhong Ke is partially supported by the Australian Research Council funding Discovery Early Career Researcher Award (DE250100030). Dr. Deying Li was partially supported by the National Natural Science Foundation of China under the Grant No. 12071478. Dr. Yongcai Wang was partially supported by the National Natural Science Foundation of China under the Grant No. 61972404, the Public Computing Cloud at Renmin University of China, and the Blockchain Lab, School of Information, Renmin University of China. + +# References + +[1] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. CoRR, abs/1512.03012, 2015. 3 +[2] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. PointGPT: Auto-regressively generative pre-training from point clouds. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1 +[3] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations, 2019. 3, 4 +[4] Yinbo Chen, Zhuang Liu, Huijuan Xu, Trevor Darrell, and Xiaolong Wang. Meta-baseline: Exploring simple meta-learning for few-shot learning. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9042-9051, 2021. +[5] Sungha Choi, Seunghan Yang, Seokeon Choi, and Sungrack Yun. Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. In European Conference on Computer Vision, pages 440–458. Springer, 2022. 2, 3, 4 +[6] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. In Advances in Neural Information Processing Systems, pages 35799-35813. Curran Associates, Inc., 2023. 1 +[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 1, 2, 3, 5, 6 + +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4 +[9] Lunhao Duan, Shanshan Zhao, Nan Xue, Mingming Gong, Gui-Song Xia, and Dacheng Tao. Condaformer: Disassembled transformer with local structure enhancement for 3d point cloud understanding. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1 +[10] Hehe Fan, Xiaojun Chang, Wanyue Zhang, Yi Cheng, Ying Sun, and Mohan Kankanhalli. Self-supervised global-local structure modeling for point cloud domain adaptation with reliable voted pseudo labels. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6367-6376, 2022. 1, 2, 3 +[11] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and ZhaoXiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2918-2927, 2021. 1 +[12] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2704-2714, 2023. 2 +[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 3, 4 +[14] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In Advances in Neural Information Processing Systems, pages 29374-29385. Curran Associates, Inc., 2022. 2 +[15] Edouard Grave, Moustapha M Cisse, and Armand Joulin. Unbounded cache model for online language modeling with open vocabulary. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3 +[16] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 1 +[17] Ahmed Hatem, Yiming Qian, and Yang Wang. Point-tta: Test-time adaptation for point cloud registration using multitask meta-auxiliary learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16494-16504, 2023. 2 +[18] Ahmed Hatem, Yiming Qian, and Yang Wang. Test-time adaptation for point cloud upsampling using meta-learning. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1284–1291, 2023. 2 +[19] Chao Huang, Zhangjie Cao, Yunbo Wang, Jianmin Wang, and Mingsheng Long. Metasets: Meta-learning on point sets for generalizable representations. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8859-8868, 2021. 1, 2, 3, 4 + +[20] Zhe Huang, Shuo Wang, Yongcai Wang, Wanting Li, Deying Li, and Lei Wang. Roco: Robust cooperative perception by iterative object matching and pose adjustment. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7833-7842, 2024. 1 +[21] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In Advances in Neural Information Processing Systems, pages 2427–2440. Curran Associates, Inc., 2021. 2, 3, 4 +[22] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2 +[23] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14162-14171, 2024. 2, 3, 4 +[24] Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models. In International Conference on Learning Representations, 2020. 3 +[25] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3 +[26] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 1 +[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? Source hypothesis transfer for unsupervised domain adaptation. In Proceedings of the 37th International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2 +[28] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022. 2 +[29] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In Advances in Neural Information Processing Systems, pages 44860-44879. Curran Associates, Inc., 2023. 2, 3, 5, 6, 1, 4 +[30] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In Advances in Neural Information Processing Systems, pages 21808-21820. Curran Associates, Inc., 2021. 2 +[31] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d an + +notation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1190-1199, 2023. 1 +[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In International Conference on Learning Representations, 2022. 1, 3 +[33] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. In International Conference on Learning Representations, 2017. 3 +[34] M. Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[35] M. Jehanzeb Mirza, Inkyu Shin, Wei Lin, Andreas Schriebl, Kunyang Sun, Jaesung Choe, Mateusz Kozinski, Horst Possegger, In So Kweon, Kuk-Jin Yoon, and Horst Bischof. Mate: Masked autoencoders are online 3d test-time learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16709-16718, 2023. 2, 3 +[36] Emin Orhan. A simple cache model for image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3, 4 +[37] Jinyoung Park, Sanghyeok Lee, Sihyeon Kim, Yunyang Xiong, and Hyunwoo J. Kim. Self-positioning point-based transformer for point cloud understanding. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21814-21823, 2023. 1 +[38] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc Le. Combined scaling for open-vocabulary image classification. 2021. 2 +[39] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 3 +[40] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1, 3 +[41] Can Qin, Haoxuan You, Lichen Wang, C.-C. Jay Kuo, and Yun Fu. Pointdan: A multi-scale 3d domain adaption network for point cloud representation. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 2, 3 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[43] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions. In In + +ternational Conference on Machine Learning (ICML), 2022. 1, 2, 5, 6 +[44] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions, 2022. 1 +[45] Yuefan Shen, Yanchao Yang, Mi Yan, He Wang, Youyi Zheng, and Leonidas Guibas. Domain adaptation on point clouds via geometry-aware implicit. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7213-7222, 2022. 1, 2, 3 +[46] Hajin Shim, Changhun Kim, and Eunho Yang. CloudFixer: Test-Time Adaptation for 3D Point Clouds via Diffusion-Guided Geometric Transformation. In European Conference on Computer Vision (ECCV), 2024. 2, 3 +[47] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, pages 14274-14289. Curran Associates, Inc., 2022. 2 +[48] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3, 4 +[49] Hongyu Sun, Yongcai Wang, Xudong Cai, Xuewei Bai, and Deying Li. Vipformer: Efficient vision-and-pointcloud transformer for unsupervised pointcloud understanding. In IEEE International Conference on Robotics and Automation (ICRA), 2023. 3 +[50] Hongyu Sun, Qiuhong Ke, Yongcai Wang, Wang Chen, Kang Yang, Deying Li, and Jianfei Cai. Point-prc: A prompt learning based regulation framework for generalizable point cloud analysis. In The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2 +[51] Jiachen Sun, Qingzhao Zhang, Bhavya Kailkhura, Zhiding Yu, Chaowei Xiao, and Z. Morley Mao. Benchmarking robustness of 3d point cloud recognition against common corruptions, 2022. 1 +[52] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2 +[53] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. Proceedings of the AAAI Conference on Artificial Intelligence, 38(6):5171-5179, 2024. 3, 4, 6 +[54] Yun-Yun Tsai, Fu-Chen Chen, Albert Y. C. Chen, Junfeng Yang, Che-Chun Su, Min Sun, and Cheng-Hao Kuo. Gda: Generalized diffusion for robust test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23242-23251, 2024. 2 +[55] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language + +models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2725-2736, 2023. 3, 4 +[56] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 3, 5, 6 +[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 4 +[58] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 3, 4 +[59] Anran Wang, Jiwen Lu, Jianfei Cai, Tat-Jen Cham, and Gang Wang. Large-margin multi-modal deep learning for rgb-d object recognition. IEEE Transactions on Multimedia, 17 (11):1887-1898, 2015. 1 +[60] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 2 +[61] Haiyang Wang, Chen Shi, Shaoshuai Shi, Meng Lei, Sen Wang, Di He, Bernt Schiele, and Liwei Wang. Dsvt: Dynamic sparse voxel transformer with rotated sets. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13520-13529, 2023. 1 +[62] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023. +[63] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q. Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8437-8445, 2019. 1 +[64] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 2019. 1, 3 +[65] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q. Weinberger, and WeiLun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11710-11720, 2020. 2 +[66] Yanshuo Wang, Ali Cheraghian, Zeeshan Hayden, Jie Hong, Sameera Ramasinghe, Shafin Rahman, David Ahmedt-Aristizabal, Xuesong Li, Lars Petersson, and Mehrtash Harandi. Backpropagation-free network for 3d test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23231-23241, 2024. 2, 3, 4, 6 + +[67] Xin Wei, Xiang Gu, and Jian Sun. Learning generalizable part-based feature representation for 3d point clouds. In Advances in Neural Information Processing Systems, pages 29305-29318. Curran Associates, Inc., 2022. 1, 2, 3, 4 +[68] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7): 5092-5113, 2024. 1 +[69] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 803-814, 2023. 1, 5, 6 +[70] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1, 3, 5, 6 +[71] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 915-924, 2021. 3 +[72] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning unified representation of language, image and point cloud for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6, 1, 4 +[73] Le Xue, Ning Yu, Shu Zhang, Junnan Li, Roberto Martín-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 5, 6, 1 +[74] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[75] Teresa Yeo, Oğuzhan Fatih Kar, Zahra Sodagar, and Amir Zamir. Rapid network adaptation: learning to adapt neural networks using test-time feedback. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4651-4664, 2023. 1 +[76] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In Advances in Neural Information Processing Systems, pages 32215-32234. Curran Associates, Inc., 2023. 3 +[77] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud + +transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 4, 1, 3 +[78] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23783-23793, 2024. 2 +[79] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18123-18133, 2022. 2 +[80] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In Advances in Neural Information Processing Systems, pages 38629-38642. Curran Associates, Inc., 2022. 2 +[81] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8552-8562, 2022. 3 +[82] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 493-510, Berlin, Heidelberg, 2022. Springer-Verlag. 3, 4 +[83] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15211-15222, 2023. 3, 4 +[84] Renrui Zhang, Liuhui Wang, Yali Wang, Peng Gao, Hongsheng Li, and Jianbo Shi. Starting from non-parametric networks for 3d point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5344-5353, 2023. 3, 4, 6 +[85] Weichen Zhang, Wen Li, and Dong Xu. Srdan: Scale-aware and range-aware domain adaptation network for cross-dataset 3d object detection. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6765-6775, 2021. 2 +[86] Zhimin Zhang, Xiang Gao, and Wei Hu. Invariantoodg: Learning invariant features of point clouds for out-of-distribution generalization, 2024. 2, 4 +[87] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H.S. Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16259-16268, 2021. 1, 3 +[88] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 5, 6, 7, 4 + +[89] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3 \ No newline at end of file diff --git a/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/images.zip b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..05dd706615395d2da16c59ccb359e074aedc85d0 --- /dev/null +++ b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0196c4ddec45ed036f6ac0a65e787f65cd8a2932c1073094facc12d5e576421d +size 432849 diff --git a/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/layout.json b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..774df10c6267163a65beae48b1f822b279e25681 --- /dev/null +++ b/2025/Point-Cache_ Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis/layout.json @@ -0,0 +1,12777 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 521, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 521, + 138 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 521, + 138 + ], + "type": "text", + "content": "Point-Cache: Test-time Dynamic and Hierarchical Cache for Robust and Generalizable Point Cloud Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "text", + "content": "Hongyu Sun" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "text", + "content": ", Qiuhong Ke" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "text", + "content": ", Ming Cheng" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "text", + "content": ", Yongcai Wang" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "text", + "content": ", Deying Li" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "text", + "content": ", Chenhui Gou" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "text", + "content": ", Jianfei Cai" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "text", + "content": "Department of Computer Science, Renmin University of China, China \n" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 160, + 544, + 202 + ], + "type": "text", + "content": "Department of Data Science & AI, Monash University, Australia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 205, + 522, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 205, + 522, + 217 + ], + "spans": [ + { + "bbox": [ + 72, + 205, + 522, + 217 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{sunhongyu,ycw,deyingli}\\} \\mathrm{ruc.edu.cn}" + }, + { + "bbox": [ + 72, + 205, + 522, + 217 + ], + "type": "text", + "content": " {qiuhong.ke,ming.cheng, jianfei.cai}@monash.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 270, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 290, + 558 + ], + "type": "text", + "content": "This paper proposes a general solution to enable point cloud recognition models to handle distribution shifts at test time. Unlike prior methods, which rely heavily on training data (often inaccessible during online inference) and are limited to recognizing a fixed set of point cloud classes predefined during training, we explore a more practical and challenging scenario: adapting the model solely based on online test data to recognize both previously seen classes and novel, unseen classes at test time. To this end, we develop Point-Cache, a hierarchical cache model that captures essential clues of online test samples, particularly focusing on the global structure of point clouds and their local-part details. Point-Cache, which serves as a rich 3D knowledge base, is dynamically managed to prioritize the inclusion of high-quality samples. Designed as a plug-and-play module, our method can be flexibly integrated into large multimodal 3D models to support open-vocabulary point cloud recognition. Notably, our solution operates with efficiency comparable to zero-shot inference, as it is entirely training-free. Point-Cache demonstrates substantial gains across 8 challenging benchmarks and 4 representative large 3D models, highlighting its effectiveness. Code is available at https://github.com/auniquesun/Point-Cache." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 128, + 593 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 696 + ], + "type": "text", + "content": "In recent years, 3D sensors such as LiDARs and RGB-D cameras have been widely adopted in robotics and electric vehicles for their ability to provide reliable 3D geometry measurements [11, 16, 20, 31, 59, 61-63, 68]. Point clouds are among the most direct data formats produced by these 3D sensors. Although remarkable progress has been made in 3D point cloud recognition, the success is primarily based on the assumption that the test data and the model train" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 244, + 426, + 323 + ], + "blocks": [ + { + "bbox": [ + 307, + 244, + 426, + 323 + ], + "lines": [ + { + "bbox": [ + 307, + 244, + 426, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 426, + 323 + ], + "type": "image", + "image_path": "390ff48f53c480ada107b8db48c5877d240abfb34ddd2ac6deb5d504dd339e62.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 327, + 416, + 336 + ], + "lines": [ + { + "bbox": [ + 318, + 327, + 416, + 336 + ], + "spans": [ + { + "bbox": [ + 318, + 327, + 416, + 336 + ], + "type": "text", + "content": "(a) ModelNet & ModelNet-C." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 426, + 245, + 543, + 323 + ], + "blocks": [ + { + "bbox": [ + 426, + 245, + 543, + 323 + ], + "lines": [ + { + "bbox": [ + 426, + 245, + 543, + 323 + ], + "spans": [ + { + "bbox": [ + 426, + 245, + 543, + 323 + ], + "type": "image", + "image_path": "e7cdb9ca047fdb5bc01e75f1994eb9b424688e3037dab17f397cabbb43fdd220.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 327, + 541, + 336 + ], + "lines": [ + { + "bbox": [ + 429, + 327, + 541, + 336 + ], + "spans": [ + { + "bbox": [ + 429, + 327, + 541, + 336 + ], + "type": "text", + "content": "(b) ScanObjNN & ScanObjNN-C." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 346, + 547, + 413 + ], + "lines": [ + { + "bbox": [ + 305, + 346, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 547, + 413 + ], + "type": "text", + "content": "Figure 1. Recognition accuracy comparison on clean and corrupted point cloud datasets. The suffix -C indicates datasets with corruptions. Models experience a severe performance drop when data corruptions arise. Point-Cache effectively narrows down the performance gap between clean and corrupted data. The hardest split of ScanObjNN is used." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 436, + 546, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 546, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 546, + 485 + ], + "type": "text", + "content": "ing data are identically distributed [2, 9, 26, 32, 37, 39, 40, 64, 87]. However, this assumption is frequently violated in real-world scenarios due to complex geometries, as well as sensing and processing errors [43, 50, 51]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "content": "In practice, a majority of models remain vulnerable to distribution shifts, such as out-of-distribution (OOD) samples [6, 7, 50, 69], data corruptions [43, 44, 51, 56], and more. As shown in Fig. 1, notable performance gaps (e.g., " + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "inline_equation", + "content": "11 + \\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 714 + ], + "type": "text", + "content": " ) occur when models are tested on clean (ModelNet [70]) versus corrupted data (ModelNet-C [43]). These issues drive researchers to enhance the robustness and generalization of point cloud recognition methods. Related efforts can be grouped into three main streams. The first focuses on training models that can transfer to target data. These models learn a shared feature space between source and target domains [10, 41, 45, 67], or apply meta-learning frameworks to handle geometry shifts by training multiple classifiers [19], or design learnable prompts to elicit general knowledge from pre-trained large 3D models [50]. However, possible shifts are unpredictable, and the training-based methods cannot account for all cases during training. As a result, these models are limited when new shifts arise at test time [75]. The second stream addresses the problem" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1263" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "through test-time adaptation strategies [35, 46, 66]. These strategies improve performance with various adaptation techniques, such as masked auto-encoding for point clouds (MATE [35]), static prototype memory (BFTT3D [66]), and diffusion models (CloudFixer [46]). Nevertheless, we find that these approaches rely heavily on training data. For instance, MATE [35] requires pre-training on the entire training set, and BFTT3D [66] employs training data to build a static and offline prototype memory to store point cloud features. This heavy dependence on training data poses challenges, particularly when such data is unavailable or when distribution shifts occur at test time. In these cases, revisiting the training data may not be helpful to adapt the model to online test data. Additionally, these test-time methods can only recognize classes seen during training, which limits their practicality in generalizable point cloud analysis." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 483 + ], + "type": "text", + "content": "Inspired by advancements in language-guided image recognition [22, 38, 42, 79], the third branch of works utilizes contrastive pre-training on large-scale triplets (point, text, image) to connect flexible language descriptions and enable open-vocabulary point cloud recognition [29, 72, 73, 88]. These large multimodal 3D models demonstrate promising zero-shot performance and provide a new paradigm for generalizable point cloud analysis. However, zero-shot predictions are performed independently for each test sample, with the final accuracy simply averaged across the test set. This approach does not fully exploit the distribution of test data, particularly the statistics of reliable predictions from large 3D models, which could further enhance accuracy and generalization. At this point, we pose the following question: can we take advantages of both the promising zero-shot ability of large 3D models and test-time adaptation techniques to unlock more robust and generalizable point cloud recognition?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "In this work, to fully exploit the data distribution during test time, we develop a cache model to store critical fingerprints of test samples. These fingerprints consist of key-value pairs derived by large multimodal 3D models, where the key represents the sample feature extracted by the point encoder, and the value represents the predicted (pseudo) label from the model. This cache model is entirely constructed from online test data without accessing any training samples. Our approach is thus more practical and challenging than prior methods [10, 19, 35, 41, 45, 46, 50, 66, 67]. Our cache model is hierarchical, as it stores not only the global features of a point cloud in a global cache but also its local-part details in a local cache, which are essential for distinguishing subtle differences among classes. We also empirically demonstrate that this coarse-to-fine design enhances robustness and generalization in the wild. Additionally, our cache model is dynamic, since we need to continuously update the cache with online test data to prioritize high-quality key-value pairs. The constructed cache" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "serves as a powerful 3D knowledge base, capturing key characteristics of online test data. When a new sample arrives, we query the knowledge base, compute the affinity between the sample and cached keys, and produce a prediction by weighting the cached (pseudo) labels accordingly. Our method is training-free and flexible, requiring only a pre-trained 3D model and test data. More details will be provided in Sec. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 168, + 546, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 546, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 546, + 370 + ], + "type": "text", + "content": "The superiority of the proposed method is supported by comprehensive comparisons with previous strong baselines across 8 challenging benchmarks, containing up to 1,156 point cloud classes. Moreover, we implement the hierarchical cache as a plug-and-play module and integrate it into representative large 3D models, such as ULIP [72], ULIP-2 [73], OpenShape [29] and Uni3D [88], to enhance test-time point cloud recognition dynamically. The proposed cache model brings notable and consistent gains over the baselines. For instance, our model yields " + }, + { + "bbox": [ + 304, + 168, + 546, + 370 + ], + "type": "inline_equation", + "content": "+6.18\\%" + }, + { + "bbox": [ + 304, + 168, + 546, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 168, + 546, + 370 + ], + "type": "inline_equation", + "content": "+4.84\\%" + }, + { + "bbox": [ + 304, + 168, + 546, + 370 + ], + "type": "text", + "content": " absolute accuracy improvements average over 7 types of corruptions in ModelNet-C [43] based on ULIP [72] and ULIP-2 [73]. On Objaverse-LVIS [7] that includes 1,156 categories, Point-Cache boosts the accuracy by " + }, + { + "bbox": [ + 304, + 168, + 546, + 370 + ], + "type": "inline_equation", + "content": "+2.10\\%" + }, + { + "bbox": [ + 304, + 168, + 546, + 370 + ], + "type": "text", + "content": " absolute points compared to the zero-shot ULIP-2 [73]. In summary, the contributions of our paper are threefold." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 371, + 545, + 550 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 371, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 545, + 442 + ], + "type": "text", + "content": "- To our knowledge, Point-Cache is the first to explore test-time point cloud recognition in a more practical and challenging setting, where only online test data is available and the model needs to generalize to both known classes and unseen new classes, which is beyond the capability of prior test-time methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 443, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 545, + 514 + ], + "type": "text", + "content": "- We propose a hierarchical cache model to capture essential distribution clues of online test samples, creating an accurate profile for various point clouds from global to local perspectives. We implement the hierarchical cache as a plug-and-play component, enabling diverse 3D backbones to achieve enhanced test-time performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 515, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 545, + 550 + ], + "type": "text", + "content": "- Comprehensive experiments and consistent gains across 8 public datasets and 4 large 3D models demonstrate the efficacy of Point-Cache." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 562, + 392, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 392, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 392, + 574 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 715 + ], + "type": "text", + "content": "Test-time adaptation in 3D point cloud analysis receives relatively less attention compared to the 2D image community [5, 12, 14, 21, 23, 27, 28, 30, 34, 47, 52, 54, 60, 78, 80]. This technique has recently emerged in 3D point cloud tasks, including recognition [35, 46, 66], registration [17], upsampling [18] and object detection [65, 74, 85]. Among these, MATE [35], BFTT3D [66], and CloudFixer [46] are specifically focused on test-time point cloud recognition and are thus closely related to our work. However, as previously analyzed, these methods depend heavily on training data from the source domain to construct their pipelines," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1264" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 542, + 261 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 542, + 261 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 542, + 261 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 542, + 261 + ], + "type": "image", + "image_path": "3ca772f4516ab57674d72aabfc6ae6c3055f7993462a8a11bfdf0e2c4ddb4a95.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 546, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 546, + 297 + ], + "type": "text", + "content": "Figure 2. The overall pipeline of Point-Cache. The zero-shot predictions " + }, + { + "bbox": [ + 46, + 271, + 546, + 297 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{zs}" + }, + { + "bbox": [ + 46, + 271, + 546, + 297 + ], + "type": "text", + "content": " of large 3D models are effectively adapted by our global cache logits " + }, + { + "bbox": [ + 46, + 271, + 546, + 297 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_g" + }, + { + "bbox": [ + 46, + 271, + 546, + 297 + ], + "type": "text", + "content": " and local cache logits " + }, + { + "bbox": [ + 46, + 271, + 546, + 297 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_l" + }, + { + "bbox": [ + 46, + 271, + 546, + 297 + ], + "type": "text", + "content": " to handle the distribution shifts, enabling robust and generalizable point cloud analysis." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 315, + 289, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 289, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 289, + 447 + ], + "type": "text", + "content": "which may not be available or helpful to understand target data distributions at test time. Furthermore, they lack effective strategies to leverage the distribution of test samples. By contrast, Point-Cache is constructed entirely from online test data without access to any training samples, making our setting both more challenging and more practical. Additionally, our cache model, powered by large 3D models, can generalize to both known classes and unseen new categories, an advancement that prior point cloud adaptation/generalization frameworks have not achieved [10, 19, 35, 41, 45, 46, 66, 67]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": "Cache models that store sample features in memory for further reference and reuse have been explored in fields like language modeling [15, 24, 33, 58], image classification [3-5, 13, 21, 23, 36, 48, 55, 58, 82, 83] and point cloud recognition [53, 66, 84]. Our proposed cache model stands out from previous ones in several key aspects: (1) Our cache model is constructed based on test samples whereas previous cache models rely on a large training set for construction, as in Tip-Adapter [82], CaFo [83], PointNN [84], Point-PEFT [53] and BFTT3D [66]. (2) Our cache model is hierarchically designed to capture both global and local-part features of a point cloud object, while existing methods store only global representations [3-5, 13, 21, 48, 53, 58, 66, 82-84]. We empirically demonstrate that this coarse-to-fine design improves the capture of target data features, thereby enhancing test-time performance. (3) Our cache model is dynamically updated with online test samples, incorporating a selective mechanism to ensure high-quality key-value pairs replace less reliable ones when the cache reaches its capacity. In contrast, prior cache models, which are constructed offline using the entire training set, are static and consume a substantial amount of memory [53, 66, 84]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 315, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 547, + 555 + ], + "type": "text", + "content": "Large Multi-modal 3D Models [25, 29, 72, 73, 76, 81, 88, 89] spark a wave of open-vocabulary point cloud understanding by pre-training on large-scale datasets [7] in recent years. For instance, ULIP-2 [73], OpenShape [29] and Uni3D [88] pre-train the 3D encoders on million-scale (point, image, text) triplets using contrastive learning. These pre-trained point cloud encoders are then transferred to downstream 3D tasks like classification and retrieval, demonstrating strong zero-shot performance that surpasses traditional 3D models [32, 39, 40, 64, 71, 87] which are typically tailored to specific datasets [1, 56, 70]. This new paradigm presents a promising direction for generalizable point cloud processing. Our work is complementary to these models: rather than developing another large 3D model, we build upon their capabilities to enhance robustness and generalization for point cloud analysis at test time. Our hierarchical cache, backed by these large 3D models, generalizes to unseen classes beyond training, addressing a significant limitation of current test-time methods for point cloud recognition [35, 46, 66]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 570, + 362, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 362, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 362, + 582 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 592, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 545, + 616 + ], + "type": "text", + "content": "This section elaborates on the details of Point-Cache, the overall pipeline of our approach is depicted in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Preliminary. We begin by reviewing the zero-shot inference process of existing large multimodal 3D models [29, 72, 73, 81, 88, 89]. These models use separate encoders to map inputs from a point cloud " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " (e.g., \"a point cloud object of a {class}\") to a shared feature space. Denoting the point cloud encoder as " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and the text encoder as " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{t}(\\cdot)" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ", we can obtain the global feature " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_p^g = f_p(P) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " of the point cloud " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and the text" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1265" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "feature " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_t = f_t(T) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " is the feature dimension. To recognize the input point cloud " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " in a zero-shot manner, these models replace “" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{class}\\}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "” in text " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " with each of possible " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " category names and compute the class probability " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{zs}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " using Eq. 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 139, + 287, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 287, + 171 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} _ {z s} = \\{\\hat {y} _ {i} | _ {i = 1} ^ {C} \\}, \\quad \\hat {y} _ {i} = \\frac {\\exp (s i m (\\mathbf {e} _ {p} ^ {g} , \\mathbf {e} _ {t} ^ {i}) / \\tau)}{\\sum_ {j = 1} ^ {C} \\exp (s i m (\\mathbf {e} _ {p} ^ {g} , \\mathbf {e} _ {t} ^ {j}) / \\tau)} \\tag {1}", + "image_path": "a79375c235550e40fb8f733d3e154425e460587ee0c72ebf7ffb2b74e5f57fb9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_t^i" + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "text", + "content": " indicates the text encoding of the " + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "text", + "content": "-th category. " + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "inline_equation", + "content": "sim(\\cdot, \\cdot)" + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "text", + "content": " measures the cosine similarity between inputs, and " + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "text", + "content": " is a temperature coefficient. Finally, the point cloud " + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "text", + "content": " is classified into the category with the highest probability, indexed by " + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "inline_equation", + "content": "\\hat{L} = \\arg \\max_{i} \\{\\hat{y}_i|_{i=1}^C\\}" + }, + { + "bbox": [ + 47, + 178, + 287, + 241 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 287, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 362 + ], + "type": "text", + "content": "As described in Sec. 1, our hierarchical cache stores the fingerprints of online test samples, which consists of key-value pairs. The key is the test sample feature " + }, + { + "bbox": [ + 46, + 243, + 287, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_p^g" + }, + { + "bbox": [ + 46, + 243, + 287, + 362 + ], + "type": "text", + "content": " extracted by the point encoder of a pre-trained large 3D model, and the value is the class label " + }, + { + "bbox": [ + 46, + 243, + 287, + 362 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 46, + 243, + 287, + 362 + ], + "type": "text", + "content": " predicted by the zero-shot inference. In addition, we compute the entropy " + }, + { + "bbox": [ + 46, + 243, + 287, + 362 + ], + "type": "inline_equation", + "content": "h = -\\sum_{i=1}^{C} \\hat{y}_i \\log \\hat{y}_i" + }, + { + "bbox": [ + 46, + 243, + 287, + 362 + ], + "type": "text", + "content": " of the zero-shot prediction " + }, + { + "bbox": [ + 46, + 243, + 287, + 362 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{zs}" + }, + { + "bbox": [ + 46, + 243, + 287, + 362 + ], + "type": "text", + "content": ", as it is essential to assess the confidence with which the large 3D model produces this prediction. This entropy value is employed to filter out high-quality samples in our cache model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 369, + 209, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 209, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 209, + 382 + ], + "type": "text", + "content": "3.1. Text & Point Cloud Encoding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "text", + "content": "We can obtain the text encoding " + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_t" + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "text", + "content": " of the input " + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "text", + "content": " and the global feature " + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_p^g" + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "text", + "content": " of a point cloud " + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "text", + "content": " as introduced in the preliminary. However, it is non-trivial to represent the local-part features " + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_p^l" + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "text", + "content": " of the point cloud " + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 387, + 287, + 448 + ], + "type": "text", + "content": " due to the following challenges analyzed in Fig. 3:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 460, + 162, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 460, + 162, + 471 + ], + "spans": [ + { + "bbox": [ + 67, + 460, + 162, + 471 + ], + "type": "text", + "content": "01 Part Feature Capture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 477, + 157, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 477, + 157, + 522 + ], + "spans": [ + { + "bbox": [ + 73, + 477, + 157, + 522 + ], + "type": "text", + "content": "- Difficult to define parts precisely and get their features for various 3D objects of different classes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 171, + 460, + 264, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 460, + 264, + 472 + ], + "spans": [ + { + "bbox": [ + 171, + 460, + 264, + 472 + ], + "type": "text", + "content": "02. Part Feature Storage" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 176, + 476, + 257, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 476, + 257, + 522 + ], + "spans": [ + { + "bbox": [ + 176, + 476, + 257, + 522 + ], + "type": "text", + "content": "- Expensive to store multiple part features instead of a global feature for a single point cloud" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 535, + 287, + 558 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 535, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 535, + 287, + 558 + ], + "type": "text", + "content": "Figure 3. Challenges in encoding part feature for various 3D objects of different classes: Part Feature Capture & Storage." + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "To address these challenges, we made a careful analysis and developed effective solutions, seeing Fig. 9 in the Supplementary for intuitive illustration. (1) A straightforward approach might involve using an additional pretrained segmentation network to segment point cloud parts, but this would be computationally prohibitive during online inference. Instead, we propose a fuzzy definition of parts. Specifically, since the architecture of the point encoder of large 3D models is based on Transformer [8, 57, 77] and the input is a sequence of point patches, we avoid strict part definitions according to physical semantics (e.g., head, body, wing, and tail in an airplane), and instead treat the output" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "point patches from the last encoder layer as local part representations. (2) A second challenge is that a long point patch sequence can become resource-intensive; for example, the point encoder in ULIP-2 [73] generates 512 point patches per object, which would require substantial memory. To mitigate this problem, we propose to summarize these point patches into " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " centers " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_p^l \\in \\mathbb{R}^{m \\times d}" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " using K-Means clustering (" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "m < 10" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "), reducing memory usage by two orders of magnitude compared to storing all patches." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 188, + 544, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 188, + 544, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 544, + 201 + ], + "type": "text", + "content": "3.2. Construction & Update of Hierarchical Cache" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 206, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 338 + ], + "type": "text", + "content": "To thoroughly explore the distribution of test data, we develop a hierarchical cache to record the fingerprints of online test samples. This hierarchical cache includes both a global cache, which stores overall information, and a local cache, which captures detailed part-level information for each point cloud. The coarse-to-fine structure enables more precise profiling of test data and distinguishes our approach from previous cache models used in image recognition [3-5, 13, 21, 23, 36, 48, 55, 58, 82, 83] and point cloud analysis [53, 66, 84]. The complete construction process of hierarchical cache is outlined in Alg. 1 of the Supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 338, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 396 + ], + "type": "text", + "content": "Global cache is represented by " + }, + { + "bbox": [ + 304, + 338, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_g = (\\mathbf{E}_g, \\hat{\\mathbf{L}}_g, \\mathbf{h}_g)" + }, + { + "bbox": [ + 304, + 338, + 545, + 396 + ], + "type": "text", + "content": ", storing (feature, label, entropy) triplets for a set of point cloud objects. To manage storage efficiently, we design a selection mechanism to identify high-quality test samples to place into the cache." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": "Construction: We set an upper bound " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": " on the number of samples per class in the cache, allowing at most " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": " fingerprints (shots) can be placed in each category. Initially, the global cache is empty. When a new test sample " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": " arrives, we generate its fingerprint " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "(\\mathbf{e}_p^g, \\hat{L}, h)" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": " and assign it to class " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": "Update: With the number of online samples grows, we check whether the number of shots for class " + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": " has hit the upper bound. If it is not the case, the fingerprint of current sample is appended to class " + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": ". Otherwise, we locate the fingerprint with highest entropy within class " + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": " and compare it to the entropy of current test sample. If the current test sample has a lower entropy, we replace the highest-entropy fingerprint with the new one; otherwise, the current sample is skipped. When the cache is full, " + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_g" + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": " has dimensions " + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "inline_equation", + "content": "CK\\times d" + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_g" + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": " has dimensions " + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "inline_equation", + "content": "CK\\times 1" + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_g" + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": " has dimensions " + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "inline_equation", + "content": "CK\\times 1" + }, + { + "bbox": [ + 304, + 473, + 545, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Local cache. Although the global feature of a point cloud captures expressive information, it lacks the local details of the 3D object, which are critical for point cloud recognition, especially when shapes share an overall appearance but differ in local parts. The local cache addresses this limitation by storing part-level information. We denote the local cache as " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_l = (\\mathbf{E}_l,\\hat{\\mathbf{L}}_l)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", which store the fingerprints of local parts of point cloud objects. Here " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_l" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " represents part features of the test samples and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_l" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " indicates their labels." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1266" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "content": "Accordingly, we define the local-part fingerprint of a test sample " + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "inline_equation", + "content": "(\\mathbf{e}_p^l,\\hat{L})" + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_p^l\\in \\mathbb{R}^{m\\times d}" + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "content": " corresponds to local-part features and " + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "content": " is the class label of " + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 72, + 287, + 110 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "text", + "content": "Construction & Update: Note that the sample selective mechanism of the local cache follows the same rules as the global cache, meaning the local-part fingerprint of a test sample is only stored if it qualifies for inclusion in the global cache. So there is no need to record the entropy " + }, + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "text", + "content": " again in the local cache. When the cache is full, " + }, + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_l" + }, + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "text", + "content": " has dimensions " + }, + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "inline_equation", + "content": "(m \\cdot CK) \\times d" + }, + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_l" + }, + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "text", + "content": " has dimensions " + }, + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "inline_equation", + "content": "CK \\times 1" + }, + { + "bbox": [ + 47, + 110, + 286, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 201, + 277, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 277, + 215 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 277, + 215 + ], + "type": "text", + "content": "3.3. Test-time Adaptation by Hierarchical Cache" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 220, + 286, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 220, + 286, + 279 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 286, + 279 + ], + "type": "text", + "content": "With the hierarchical cache serving as a rich knowledge base of online point cloud data, we can adapt the zero-shot predictions of large 3D models for new samples, effectively addressing distribution shifts that occur at test time. The adaptation follows a coarse-to-fine process as below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "text", + "content": "Test-time adaptation by global cache. During adaptation, a new sample " + }, + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "text", + "content": " is treated as a query to retrieve knowledge from the global cache, which we use to produce the adaptation logits. Specifically, the retrieval process is simulated by calculating the affinity " + }, + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "inline_equation", + "content": "A_{g}" + }, + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "text", + "content": " between the query and cached features " + }, + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{g}" + }, + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "text", + "content": ". This allows the new sample to establish connections with cached high-quality samples, providing it with a contextual understanding of the online test data. We then generate the adaptation logits " + }, + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{g}" + }, + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "text", + "content": " by weighed ensembling of cache labels " + }, + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_{g}" + }, + { + "bbox": [ + 46, + 280, + 286, + 402 + ], + "type": "text", + "content": ", as shown in Eq. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 410, + 286, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 410, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 49, + 410, + 286, + 435 + ], + "type": "interline_equation", + "content": "A _ {g} = \\exp \\left(- \\beta_ {g} \\left(1 - \\mathbf {e} _ {q} ^ {g} \\mathbf {E} _ {g} ^ {\\top}\\right)\\right), \\hat {\\mathbf {L}} _ {g} = \\mathrm {O H} (\\hat {\\mathbf {L}} _ {g}), \\hat {\\mathbf {y}} _ {g} = A _ {g} \\hat {\\mathbf {L}} _ {g} \\tag {2}", + "image_path": "939be9de9ff0ec2f606a42e474532f40d3b5b79d22873842a9b4f3162f78992a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_q^g\\in \\mathbb{R}^{1\\times d}" + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "text", + "content": " represents the global feature of sample " + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_g\\in \\mathbb{R}^{CK\\times d}" + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "text", + "content": " denotes the cached keys. The 'OH' operation transforms " + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_g" + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "text", + "content": " into a one-hot encoding. The term " + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_q^g\\mathbf{E}_g^\\top \\in \\mathbb{R}^{1\\times CK}" + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "text", + "content": " measures the cosine similarity between the query and all keys, with " + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "inline_equation", + "content": "\\beta_{g}" + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "text", + "content": " acting as a hyper-parameter to modulate the sharpness and the exponential operation ensuring similarity values fall within (0, 1]. The affinity " + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "inline_equation", + "content": "A_{g}\\in \\mathbb{R}^{1\\times CK}" + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "text", + "content": ", which grasps the relationship between the query and cached keys, is utilized to ensemble the transformed one-hot labels " + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_g\\in \\mathbb{R}^{CK\\times C}" + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "text", + "content": ", yielding global cache adaptation logits " + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_g\\in \\mathbb{R}^{1\\times C}" + }, + { + "bbox": [ + 47, + 436, + 286, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "text", + "content": "Test-time adaptation by local cache. While global cache adaptation is effective, it may not handle fine-grained distribution shifts at test time. To achieve a more precise adjustment, we exploit local part knowledge. The local-part features " + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_q^l" + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "text", + "content": " of the point cloud " + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "text", + "content": " serve as queries to the local cache. During retrieval, we compute the affinity " + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "inline_equation", + "content": "A_{l}" + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_q^l" + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "text", + "content": " and all cached part features " + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_l" + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "text", + "content": ". The affinity " + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "inline_equation", + "content": "A_{l}" + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "text", + "content": " is then used to weight the cache labels " + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_l" + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "text", + "content": ", producing local cache adaptation logits " + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_l" + }, + { + "bbox": [ + 46, + 570, + 286, + 680 + ], + "type": "text", + "content": " as in Eq. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 687, + 286, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 687, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 687, + 286, + 713 + ], + "type": "interline_equation", + "content": "A _ {l} = \\exp \\left(- \\beta_ {l} \\left(1 - \\mathbf {e} _ {q} ^ {l} \\mathbf {E} _ {l} ^ {\\top}\\right)\\right), \\hat {\\mathbf {L}} _ {l} = \\mathrm {O H} (\\hat {\\mathbf {L}} _ {l}), \\hat {\\mathbf {y}} _ {l} = \\phi \\left(A _ {l} \\hat {\\mathbf {L}} _ {l}\\right) \\tag {3}", + "image_path": "2c8c85ad0f23d53875f6d55ddacbb836a7c070f1689d6a35690d78f3f78e4bbb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_q^l\\in \\mathbb{R}^{m\\times d}" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_l\\in \\mathbb{R}^{(m\\cdot CK)\\times d}" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{L}}_l\\in \\mathbb{R}^{(m\\cdot CK)\\times C}" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": " is an average pooling operation over " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": " parts to produce class logits. The term " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "e_q^l\\mathbf{E}_l^\\top \\in \\mathbb{R}^{m\\times (m\\cdot CK)}" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": " measures the cosine similarity between the part features of sample " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": " and all part features in the cache. Likewise, the similarity is modulated by the coefficient " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\beta_{l}" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": " and scaled with the exponential function, resulting in the affinity " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "A_{l}\\in \\mathbb{R}^{m\\times (m\\cdot CK)}" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "A_{l}\\hat{\\mathbf{L}}_{l}\\in \\mathbb{R}^{m\\times C}" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": " represents the class logits for all " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": " parts, which are transformed by a pooling operation " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": " to derive the final local cache adaptation logits " + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_l\\in \\mathbb{R}^{1\\times C}" + }, + { + "bbox": [ + 304, + 71, + 545, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "text", + "content": "Overall prediction combines three components as shown in Eq. 4. The zero-shot prediction " + }, + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{zs}" + }, + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "text", + "content": " of large multimodal 3D model is effectively adapted using our global cache prediction " + }, + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_g" + }, + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "text", + "content": " and local cache prediction " + }, + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_l" + }, + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "text", + "content": ", balanced by the coefficients " + }, + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\alpha_{q}" + }, + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\alpha_{l}" + }, + { + "bbox": [ + 304, + 206, + 545, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 376, + 277, + 545, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 277, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 376, + 277, + 545, + 291 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} = \\hat {\\mathbf {y}} _ {z s} + \\alpha_ {g} \\hat {\\mathbf {y}} _ {g} + \\alpha_ {l} \\hat {\\mathbf {y}} _ {l} \\tag {4}", + "image_path": "167f5cc0bc3c20e4d22142c0469cffd1834f78ba66a35bf47dc45ae999cbe637.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 302, + 388, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 302, + 388, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 388, + 316 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 323, + 432, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 432, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 432, + 335 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 342, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 533 + ], + "type": "text", + "content": "Datasets. To evaluate the robustness of point cloud recognition, we choose four public datasets that contain various point cloud corruptions: ModelNet-C [43] and three variants of ScanObjectNN-C [56]. ModelNet-C [43] introduces 7 atomic corruptions, including adding global outliers, adding local outliers, dropping global structure, dropping local parts, rotation, scaling, and jittering. Other corruptions can be derived from these atomic corruptions, as discussed in [43]. We apply these atomic corruptions to the three variants of ScanObjectNN [56] and produce their corrupted versions. To assess generalization on unseen new data, we test the proposed method on four challenging benchmarks: OmniObject3D [69] (216 classes), ObjaverseLVIS [7] (1,156 classes), the hardest variant of ScanObjectNN [56] and widely used ModelNet40 [70]. Recognition accuracy (\\%) is the primary evaluation metric." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. We choose ULIP [72], ULIP-2 [73], OpenShape [29] and Uni3D [88] as the large multimodal 3D models for experiments. Their pre-trained weights are loaded and then frozen. Global and local-part features are generated by modifying the forward pass function without altering the model architecture. The maximum number of samples per class in the cache, " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ", is set to 3, and each point cloud is clustered into " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " local parts. Both the balance factors " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{g}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{l}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " are set to 4.0 while the coefficients " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{g}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{l}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " are set to 3.0. We will examine the design choices in the ablation studies. To optimize memory and computation, we convert the input point clouds and model parameters into 'dtype' fp16 in Pytorch. All experiments are conducted on two 4090 GPUs. Additional details can be found in the Supplementary." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1267" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 113, + 536, + 274 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Comparison of recognition accuracy on ModelNet-C that contains 7 types of corruptions. Each clean point cloud has 1024 points and the corruption severity level is 2. The last column is the average over 7 corruption types. The best results are in bold and the second best are underlined. The setting applies to the following tables unless otherwise specified." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 113, + 536, + 274 + ], + "lines": [ + { + "bbox": [ + 58, + 113, + 536, + 274 + ], + "spans": [ + { + "bbox": [ + 58, + 113, + 536, + 274 + ], + "type": "table", + "html": "
MethodClean Data ModelNetCorruption TypeAvg.
Add GlobalAdd LocalDrop GlobalDrop LocalRotateScaleJitter
ULIP [72]56.1633.5543.9254.7050.8955.2750.2044.0847.52
+Global Cache(Ours)62.1245.7947.9856.8553.8960.2554.3448.9152.56
+Hierarchical Cache(Ours)64.2246.1547.8559.1656.0061.4755.3549.9253.70
ULIP-2 [73]71.2365.1554.6268.7657.9870.3067.1021.7657.95
+Global Cache(Ours)73.9567.0259.3271.3561.5972.3768.4028.2061.18
+Hierarchical Cache(Ours)74.5368.1161.2673.2263.6573.3470.4229.5062.79
O-Shape [29]84.5671.6467.7981.5673.5882.0178.4859.3673.49
+Global Cache(Ours)84.5274.7272.7782.4175.1283.1878.9367.9176.43
+Hierarchical Cache(Ours)84.0474.8473.7082.2176.2682.6678.1268.3576.59
Uni3D [88]81.8172.4556.3668.1567.1879.9475.3656.2467.95
+Global Cache(Ours)83.1476.1366.4971.4369.8181.5275.8561.4371.81
+Hierarchical Cache(Ours)83.8777.5171.1572.1670.7581.7777.3162.5273.31
", + "image_path": "861996e93f967c92c33c077126398bd1ebe9e4550bae37f844847473bfb1a673.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 85, + 326, + 509, + 454 + ], + "blocks": [ + { + "bbox": [ + 46, + 284, + 546, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 546, + 319 + ], + "type": "text", + "content": "Table 2. Comparison of recognition accuracy across a suite of datasets. S-PB_RS_T50 is the hardest split of ScanObjectNN. OLVIS: Objverse-LVIS. Omni3D: OmniObject3D. The number under each dataset indicates the number of points (pts) for each object. In Omni3D, each point cloud can be represented by a different number of points. The last column is the average accuracy on these datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 326, + 509, + 454 + ], + "lines": [ + { + "bbox": [ + 85, + 326, + 509, + 454 + ], + "spans": [ + { + "bbox": [ + 85, + 326, + 509, + 454 + ], + "type": "table", + "html": "
MethodModelNet40 (10000 pts)S-PB_RS_T50 (2048 pts)O-LVIS (10000 pts)Omni3DAvg.
(1024 pts4096 pts16384 pts)
ULIP [72]58.7546.446.248.397.757.2822.48
+Global Cache(Ours)61.2250.217.0210.009.368.4324.37
+Hierarchical Cache(Ours)62.9351.807.0210.479.758.9025.15
ULIP-2 [73]72.9747.1330.2626.3629.2026.5838.75
+Global Cache(Ours)74.5151.7032.6528.5131.1028.5341.17
+Hierarchical Cache(Ours)75.5354.9832.3629.3731.2429.4442.15
Uni3D [88]88.4165.1955.4231.5241.9841.8654.09
+Global Cache(Ours)88.8668.5153.3634.9745.1345.1956.00
+Hierarchical Cache(Ours)89.1868.2455.1935.8245.6045.8956.65
", + "image_path": "2c300dc0bcb188624dad860d7c356578c2aa7027efc3c63aa11240f63be5ac6b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 473, + 261, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 261, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 261, + 485 + ], + "type": "text", + "content": "4.2. Test-time Robustness and Generalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "text", + "content": "Robustness. The robustness analysis of our training-free test-time hierarchical caches takes place on 4 point cloud corruption datasets [43, 56]. As observed in Tab. 1, the zero-shot accuracy of various large 3D models on ModelNet-C is considerably boosted by our global cache model, with increases such as " + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "inline_equation", + "content": "+5.04\\%" + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "text", + "content": " for ULIP [72], " + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "inline_equation", + "content": "+3.23\\%" + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "text", + "content": " for ULIP-2 [73], " + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "inline_equation", + "content": "+2.94\\%" + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "text", + "content": " for OpenShape [29] and " + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "inline_equation", + "content": "+3.86\\%" + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "text", + "content": " for Uni3D [88], averaged across 7 corruption types. By combining global and local cache together, our hierarchical cache model attains further improvements, highlighting the benefits of incorporating a local cache. Notably, these gains extend beyond corrupted data, as our hierarchical cache model also improves recognition accuracy on the clean ModelNet dataset, e.g., " + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "inline_equation", + "content": "+8.06\\%" + }, + { + "bbox": [ + 46, + 498, + 288, + 715 + ], + "type": "text", + "content": " absolute increase over ULIP's zero-shot performance. An exception is OpenShape, which maintains a slightly higher zero-shot accuracy on clean data than our cache model. However, when considering the 7 corruption types, our global and hi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "content": "erarchical cache models outperform OpenShape by " + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "inline_equation", + "content": "2.94\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "inline_equation", + "content": "3.10\\%" + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "content": ", respectively, demonstrating improved robustness against data corruption at test time. Consistent improvements are also showcased on the variants of ScanObjectNN, shown in Tab. 5, 6 and 7 in the Supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Generalization. We evaluated the generalization capability of Point-Cache on four representative benchmarks that were unseen during the pre-training of the large 3D models. These benchmarks generally do not contain data corruptions. While ModelNet40 [70] and ScanObjectNN [56] contain a limited number of classes, OmniObject3D [69] and Objverse-LVIS [7] represent a broader range of 3D concepts, covering 216 and 1,156 classes, respectively. As shown in Tab. 2, Point-Cache consistently enhances the recognition accuracy across datasets and 3D backbones. For instance, our hierarchical cache built on ULIP and ULIP-2 yields absolute improvements of " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "6.31\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "5.51\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", respectively, average over 6 scores. On ScanObjectNN, our hierarchical cache significantly raises the zero-shot prediction of ULIP and ULIP-2, from " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "27.20\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "51.80\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " for ULIP" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1268" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "and from " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "36.22\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "54.98\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " for ULIP-2. On the challenging Objaverse-LVIS (O-LVIS), our global cache with ULIP-2 successfully realizes a " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "2.39\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " absolute gains in zero-shot accuracy across 1,156 classes, a notable achievement. Furthermore, our cache model enables Uni3D to attain an " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "89.18\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " accuracy on ModelNet40 without any training, effectively matching the fully-supervised PointNet " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "(89.20\\%)" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 165, + 222, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 165, + 222, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 165, + 222, + 178 + ], + "type": "text", + "content": "4.3. Memory Usage and Throughput" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 184, + 290, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 184, + 290, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 290, + 365 + ], + "type": "text", + "content": "Memory. We use Uni3D [88] as the baseline and adapt its zero-shot predictions with Point-Cache. The experiments are conducted on 3 datasets as shown in Tab. 3. The batch size is set to 1 because we need to update the cache and adapt the class logits according to every single sample. The results indicate our global cache and hierarchical cache consume similar or slightly higher memory than the baseline on each dataset. Furthermore, as the number of classes increases rapidly, the memory usage grows at a slower rate. This is primarily because the memory consumption is dominated by the large number of parameters in Uni3D (e.g., 1,016.5M). In contrast, the hierarchical cache built on OLVIS uses only approximately 7.1M parameters, which is negligible compared to Uni3D. Detailed calculations are provided in the Supplementary." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 49, + 418, + 284, + 479 + ], + "blocks": [ + { + "bbox": [ + 46, + 376, + 288, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 409 + ], + "type": "text", + "content": "Table 3. Comparison of memory usage (MB). The batch size is set to 1 and the used device is an RTX 4090. The number under each dataset indicates #Classes. Hierar: Hierarchical Cache." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 418, + 284, + 479 + ], + "lines": [ + { + "bbox": [ + 49, + 418, + 284, + 479 + ], + "spans": [ + { + "bbox": [ + 49, + 418, + 284, + 479 + ], + "type": "table", + "html": "
MethodModelNet-C (40)Omni3D (216)O-LVIS (1156)#Params (M)
Uni3D5,0625,0625,0621,016.5
+Global(Ours)5,0625,0645,0701,016.5
+Hierar(Ours)5,0645,0685,0901,016.5
", + "image_path": "b34b90bb25a1b4218faeb495a08ac3abe5d1c5d7b7a66c9a0606f4262c60d7b9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 492, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 635 + ], + "type": "text", + "content": "Throughput. We evaluate the throughput of Point-Cache on ModelNet40, where throughput is measured as the number of test samples processed per second (abbreviated as " + }, + { + "bbox": [ + 46, + 492, + 287, + 635 + ], + "type": "inline_equation", + "content": "t / s" + }, + { + "bbox": [ + 46, + 492, + 287, + 635 + ], + "type": "text", + "content": "). The results are reported in Tab. 4. Point-Cache runs slightly slower than zero-shot inference due to the additional overhead introduced by cache updates, logits computation, etc. However, the throughput degradation is minimal, e.g., a " + }, + { + "bbox": [ + 46, + 492, + 287, + 635 + ], + "type": "inline_equation", + "content": "0.04 \\, t / s" + }, + { + "bbox": [ + 46, + 492, + 287, + 635 + ], + "type": "text", + "content": " drop for OpenShape with the global cache, and a " + }, + { + "bbox": [ + 46, + 492, + 287, + 635 + ], + "type": "inline_equation", + "content": "0.07 \\, t / s" + }, + { + "bbox": [ + 46, + 492, + 287, + 635 + ], + "type": "text", + "content": " drop for ULIP-2 with the hierarchical cache. These results prove that Point-Cache brings remarkable accuracy improvements with negligible computational cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 645, + 141, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 141, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 141, + 658 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 664, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 287, + 689 + ], + "type": "text", + "content": "We use Uni3D with our hierarchical cache and 4 public datasets for ablation analysis, as shown in Fig. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Shot size per class in the global cache. The upper bound " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " on the number of samples per class is a critical variable" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 113, + 535, + 174 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 4. Comparison of throughput " + }, + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "(t / s)" + }, + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "type": "text", + "content": " for different models on ModelNet40. The batch size is set to 1 and the device is a 4090 GPU. The results are averaged over all test samples." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 113, + 535, + 174 + ], + "lines": [ + { + "bbox": [ + 315, + 113, + 535, + 174 + ], + "spans": [ + { + "bbox": [ + 315, + 113, + 535, + 174 + ], + "type": "table", + "html": "
MethodZero-shot+Global Cache+Hierar Cache
ULIP11.2511.2111.18
ULIP-211.1811.1411.11
OpenShape9.809.769.74
Uni3D9.789.759.73
", + "image_path": "36a3695144c02f118ba581585afcd6b81ba2db1684724c35f586634e6c4d2d95.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 195, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 545, + 291 + ], + "type": "text", + "content": "as it affects the total size of the hierarchical cache. We analyze its impact on the test-time recognition accuracy on 4 datasets, shown in Fig. 4a. As " + }, + { + "bbox": [ + 304, + 195, + 545, + 291 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 195, + 545, + 291 + ], + "type": "text", + "content": " increases, accuracy trend differ across datasets; for example, accuracy increases on Omni3D but decreases on S-PB_T50_RS. We adopt " + }, + { + "bbox": [ + 304, + 195, + 545, + 291 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 304, + 195, + 545, + 291 + ], + "type": "text", + "content": " as it produces the best average accuracy (63.98%). The results indicate that our cache model does not require a large number of samples to achieve strong performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 291, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 546, + 411 + ], + "type": "text", + "content": "Number of local parts per 3D object. To optimize memory and computation, we cluster hundreds of point patches of a 3D object into " + }, + { + "bbox": [ + 304, + 291, + 546, + 411 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 291, + 546, + 411 + ], + "type": "text", + "content": " parts. We investigate the effect of this variable on test-time performance, as exhibited in Fig. 4b. For three out of the four datasets, namely S-PB_T50_RS, Omni3D, and O-LVIS, accuracy generally decreases as the number of parts increases, suggesting that a point cloud does not need to be represented with an excessive number of parts. We choose " + }, + { + "bbox": [ + 304, + 291, + 546, + 411 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 304, + 291, + 546, + 411 + ], + "type": "text", + "content": " by default, as this setting reaches the best average accuracy across the 4 datasets." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 422, + 369, + 464 + ], + "blocks": [ + { + "bbox": [ + 307, + 422, + 369, + 464 + ], + "lines": [ + { + "bbox": [ + 307, + 422, + 369, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 369, + 464 + ], + "type": "image", + "image_path": "3d50fac6fb5374b8f54c7e8633e362ea56581ba805c8dfd380cb4c885ca85db3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 467, + 361, + 476 + ], + "lines": [ + { + "bbox": [ + 312, + 467, + 361, + 476 + ], + "spans": [ + { + "bbox": [ + 312, + 467, + 361, + 476 + ], + "type": "text", + "content": "(a) shot size " + }, + { + "bbox": [ + 312, + 467, + 361, + 476 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 370, + 422, + 426, + 464 + ], + "blocks": [ + { + "bbox": [ + 370, + 422, + 426, + 464 + ], + "lines": [ + { + "bbox": [ + 370, + 422, + 426, + 464 + ], + "spans": [ + { + "bbox": [ + 370, + 422, + 426, + 464 + ], + "type": "image", + "image_path": "5ece0fcdfedac25c6712cece2c89841ee3708073de6ac2b219f6334ed8e01072.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 467, + 416, + 476 + ], + "lines": [ + { + "bbox": [ + 378, + 467, + 416, + 476 + ], + "spans": [ + { + "bbox": [ + 378, + 467, + 416, + 476 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 378, + 467, + 416, + 476 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 378, + 467, + 416, + 476 + ], + "type": "text", + "content": " parts" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 487, + 547, + 532 + ], + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 532 + ], + "type": "text", + "content": "Figure 4. Ablation studies on the hyper-parameters in the cache design, including the shot size " + }, + { + "bbox": [ + 304, + 487, + 547, + 532 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 487, + 547, + 532 + ], + "type": "text", + "content": " per class, the number of parts " + }, + { + "bbox": [ + 304, + 487, + 547, + 532 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 487, + 547, + 532 + ], + "type": "text", + "content": " per object, the balance factors in the final prediction logits and the sharpness coefficients in affinity computation." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 429, + 422, + 486, + 464 + ], + "blocks": [ + { + "bbox": [ + 429, + 422, + 486, + 464 + ], + "lines": [ + { + "bbox": [ + 429, + 422, + 486, + 464 + ], + "spans": [ + { + "bbox": [ + 429, + 422, + 486, + 464 + ], + "type": "image", + "image_path": "6109cecf29b549bc36eb9a36123cd2f26d6ba1c536cb083225f17b3ad98efae0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 467, + 478, + 477 + ], + "lines": [ + { + "bbox": [ + 433, + 467, + 478, + 477 + ], + "spans": [ + { + "bbox": [ + 433, + 467, + 478, + 477 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 433, + 467, + 478, + 477 + ], + "type": "inline_equation", + "content": "\\alpha_{g}" + }, + { + "bbox": [ + 433, + 467, + 478, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 433, + 467, + 478, + 477 + ], + "type": "inline_equation", + "content": "\\alpha_{l}" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 488, + 422, + 544, + 464 + ], + "blocks": [ + { + "bbox": [ + 488, + 422, + 544, + 464 + ], + "lines": [ + { + "bbox": [ + 488, + 422, + 544, + 464 + ], + "spans": [ + { + "bbox": [ + 488, + 422, + 544, + 464 + ], + "type": "image", + "image_path": "4edbd5e95bdef17bf6a07aa570830d92308e27d9fde2c24acc8f84316addb291.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 467, + 537, + 477 + ], + "lines": [ + { + "bbox": [ + 493, + 467, + 537, + 477 + ], + "spans": [ + { + "bbox": [ + 493, + 467, + 537, + 477 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 493, + 467, + 537, + 477 + ], + "type": "inline_equation", + "content": "\\beta_{g}" + }, + { + "bbox": [ + 493, + 467, + 537, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 493, + 467, + 537, + 477 + ], + "type": "inline_equation", + "content": "\\beta_{l}" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "text", + "content": "The balance factors. The factors " + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\alpha_{g}" + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\alpha_{l}" + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "text", + "content": " control the relative weights of the global cache prediction " + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_g" + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "text", + "content": " and local cache prediction " + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}_l" + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "text", + "content": ". To simplify experimentation, we set " + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\alpha_{g} = \\alpha_{l}" + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "text", + "content": " in the ablation study. Candidate values are " + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\{0.5, 1.0, 2.0, 3.0, 4.0, 5.0\\}" + }, + { + "bbox": [ + 304, + 544, + 545, + 652 + ], + "type": "text", + "content": ". As shown in Fig. 4c, increasing these factors improves performance on S-PB_T50_RS and Omni3D, while it has minimal effect on ModelNet40 and O-LVIS. We set the factors to 4.0, which yields the optimal average accuracy across all four datasets." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "The sharpness coefficients. The coefficients " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{g}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{l}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " are used to modulate the sharpness of the query-key affinity. Here we analyze their impact on test-time accuracy on 4 benchmarks. Likewise, we let " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{g} = \\alpha_{l}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " and the candidate values are " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{0.5,1.0,3.0,5.0,7.0,9.0\\}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ". As Fig. 4d displays," + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1269" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "text", + "content": "accuracy initially improves slightly with increasing sharpness but then decreases as sharpness becomes too high. We set the sharpness to 3.0 that yields the optimal result average over 4 datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 132, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 132, + 141 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 132, + 141 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 148, + 289, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 289, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 289, + 292 + ], + "type": "text", + "content": "Online Inference. We visualize the online inference process of Point-Cache using different large 3D models and compare three model variants: zero-shot prediction, the model with our global cache, and the model with our hierarchical cache, referring to Fig. 5. During inference, we calculate the average recognition accuracy over accumulated samples. Due to the small number of samples at the initial stage, the average accuracy may fluctuate drastically, so we discard the statistics for the first 5 samples. Overall, methods employing the hierarchical cache consistently outperform other variants by a clear margin across various large 3D models and datasets." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 302, + 108, + 344 + ], + "blocks": [ + { + "bbox": [ + 48, + 302, + 108, + 344 + ], + "lines": [ + { + "bbox": [ + 48, + 302, + 108, + 344 + ], + "spans": [ + { + "bbox": [ + 48, + 302, + 108, + 344 + ], + "type": "image", + "image_path": "440b50728ed305619b42439be57f9c1a5148c3761d22b4f62467eca245965696.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 346, + 94, + 356 + ], + "lines": [ + { + "bbox": [ + 63, + 346, + 94, + 356 + ], + "spans": [ + { + "bbox": [ + 63, + 346, + 94, + 356 + ], + "type": "text", + "content": "(a) ULIP." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 111, + 302, + 167, + 344 + ], + "blocks": [ + { + "bbox": [ + 111, + 302, + 167, + 344 + ], + "lines": [ + { + "bbox": [ + 111, + 302, + 167, + 344 + ], + "spans": [ + { + "bbox": [ + 111, + 302, + 167, + 344 + ], + "type": "image", + "image_path": "3d2a831c5e67b2bd1b433629862998a689f8b2c507112e6e2157973f98e03fe8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 346, + 157, + 356 + ], + "lines": [ + { + "bbox": [ + 119, + 346, + 157, + 356 + ], + "spans": [ + { + "bbox": [ + 119, + 346, + 157, + 356 + ], + "type": "text", + "content": "(b) ULIP-2." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 168, + 302, + 227, + 344 + ], + "blocks": [ + { + "bbox": [ + 168, + 302, + 227, + 344 + ], + "lines": [ + { + "bbox": [ + 168, + 302, + 227, + 344 + ], + "spans": [ + { + "bbox": [ + 168, + 302, + 227, + 344 + ], + "type": "image", + "image_path": "b2f6617e09c641732dc4e1f4053dcdd9997d714bcd920fc6cf63321534fc20b0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 346, + 222, + 357 + ], + "lines": [ + { + "bbox": [ + 171, + 346, + 222, + 357 + ], + "spans": [ + { + "bbox": [ + 171, + 346, + 222, + 357 + ], + "type": "text", + "content": "(c) OpenShape." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 229, + 302, + 286, + 344 + ], + "blocks": [ + { + "bbox": [ + 229, + 302, + 286, + 344 + ], + "lines": [ + { + "bbox": [ + 229, + 302, + 286, + 344 + ], + "spans": [ + { + "bbox": [ + 229, + 302, + 286, + 344 + ], + "type": "image", + "image_path": "33acc4474aff6e357a76eaa0ebbec68112e2f9519cc2095162ad15676ab68582.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 346, + 274, + 356 + ], + "lines": [ + { + "bbox": [ + 238, + 346, + 274, + 356 + ], + "spans": [ + { + "bbox": [ + 238, + 346, + 274, + 356 + ], + "type": "text", + "content": "(d) Uni3D." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 365, + 287, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 422 + ], + "type": "text", + "content": "Figure 5. The average recognition accuracy of accumulated samples during online inference. The curve changes significantly in the initial stage due to the small number of samples. Models with our global and hierarchical cache receive perceptible performance gains." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 437, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 593 + ], + "type": "text", + "content": "Entropy and Accuracy. Fig. 6 illustrates the relationship between entropy and accuracy for different models on various datasets. During online inference, we first compute the entropy of top 5 predictions for each test sample and then average the entropy across all accumulated samples. Similarly, the corresponding average accuracy can be calculated. The results reveal that Point-Cache effectively reduces the uncertainty of the zero-shot predictions made by large 3D models, enhances prediction confidence, and consequently lifts recognition accuracy. Additionally, we observe that the hierarchical cache model is more effective in reducing entropy compared to the global cache model, further underscoring the importance of incorporating the local cache." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Qualitative Analysis. Fig. 7 presents the step-by-step adaptation process of Point-Cache. The first three rows demonstrate cases where the global cache model attempts to adjust the zero-shot predictions of various large 3D models but fails. Instead, the hierarchical cache model successfully refines the class logits based on the global cache. In the last row, our global cache model corrects the zero-shot prediction in the first step, and the hierarchical cache model subsequently retains the top class while increasing its confidence. The examples showcased here are not cherry-picked." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 432, + 156 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 432, + 156 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 432, + 156 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 432, + 156 + ], + "type": "image", + "image_path": "dfae67d8bc2bd2ef47512d2147e69fda9f45bc3e08f427b575c2cad011c7e3ed.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 159, + 395, + 169 + ], + "lines": [ + { + "bbox": [ + 345, + 159, + 395, + 169 + ], + "spans": [ + { + "bbox": [ + 345, + 159, + 395, + 169 + ], + "type": "text", + "content": "(a) Group One." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 433, + 72, + 555, + 156 + ], + "blocks": [ + { + "bbox": [ + 433, + 72, + 555, + 156 + ], + "lines": [ + { + "bbox": [ + 433, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 433, + 72, + 555, + 156 + ], + "type": "image", + "image_path": "ea4a105ba81065854a08090615625325c57c373aa7f6c779134a0360484f40f6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 159, + 518, + 168 + ], + "lines": [ + { + "bbox": [ + 468, + 159, + 518, + 168 + ], + "spans": [ + { + "bbox": [ + 468, + 159, + 518, + 168 + ], + "type": "text", + "content": "(b) Group Two." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 178, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 545, + 212 + ], + "type": "text", + "content": "Figure 6. Relationship between entropy and accuracy. Models with our global and hierarchical cache consistently exhibit lower entropy and higher accuracy across various benchmarks." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 233, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 233, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 233, + 545, + 258 + ], + "type": "text", + "content": "Additional quantitative and qualitative results can be found in the Supplementary." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 310, + 269, + 359, + 310 + ], + "blocks": [ + { + "bbox": [ + 310, + 269, + 359, + 310 + ], + "lines": [ + { + "bbox": [ + 310, + 269, + 359, + 310 + ], + "spans": [ + { + "bbox": [ + 310, + 269, + 359, + 310 + ], + "type": "image", + "image_path": "f7c4c6c897456a02f83d8d4c42b4b202a92d64aaf3342f89f726e428386c7e2b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 313, + 366, + 322 + ], + "lines": [ + { + "bbox": [ + 311, + 313, + 366, + 322 + ], + "spans": [ + { + "bbox": [ + 311, + 313, + 366, + 322 + ], + "type": "text", + "content": "(a) Ground truth." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 370, + 268, + 430, + 311 + ], + "blocks": [ + { + "bbox": [ + 370, + 268, + 430, + 311 + ], + "lines": [ + { + "bbox": [ + 370, + 268, + 430, + 311 + ], + "spans": [ + { + "bbox": [ + 370, + 268, + 430, + 311 + ], + "type": "image", + "image_path": "e171fc5b94492d1120b78cb77f7a4d235b3f748196acddf9af6f68f94187f660.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 314, + 421, + 323 + ], + "lines": [ + { + "bbox": [ + 381, + 314, + 421, + 323 + ], + "spans": [ + { + "bbox": [ + 381, + 314, + 421, + 323 + ], + "type": "text", + "content": "(b) ULIP-2." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 433, + 268, + 492, + 311 + ], + "blocks": [ + { + "bbox": [ + 433, + 268, + 492, + 311 + ], + "lines": [ + { + "bbox": [ + 433, + 268, + 492, + 311 + ], + "spans": [ + { + "bbox": [ + 433, + 268, + 492, + 311 + ], + "type": "image", + "image_path": "541d690dabe490f921f4b8ff4d2de598a796c28ae3687f753cbbe4c006bf82af.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 314, + 489, + 324 + ], + "lines": [ + { + "bbox": [ + 435, + 314, + 489, + 324 + ], + "spans": [ + { + "bbox": [ + 435, + 314, + 489, + 324 + ], + "type": "text", + "content": "(c) ULIP-2+GC." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 495, + 268, + 555, + 311 + ], + "blocks": [ + { + "bbox": [ + 495, + 268, + 555, + 311 + ], + "lines": [ + { + "bbox": [ + 495, + 268, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 495, + 268, + 555, + 311 + ], + "type": "image", + "image_path": "0f474d2b4bc06b1aabcd1b45f58def868e61af1df741c04e8f68da5233d5d9be.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 497, + 314, + 552, + 323 + ], + "lines": [ + { + "bbox": [ + 497, + 314, + 552, + 323 + ], + "spans": [ + { + "bbox": [ + 497, + 314, + 552, + 323 + ], + "type": "text", + "content": "(d) ULIP-2+HC." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 310, + 328, + 348, + 371 + ], + "blocks": [ + { + "bbox": [ + 310, + 328, + 348, + 371 + ], + "lines": [ + { + "bbox": [ + 310, + 328, + 348, + 371 + ], + "spans": [ + { + "bbox": [ + 310, + 328, + 348, + 371 + ], + "type": "image", + "image_path": "6ec08d92dc0fbdc5f08019b6cbc0ecc7299b6af2faf5d5172371f405ae327bfa.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 376, + 366, + 384 + ], + "lines": [ + { + "bbox": [ + 311, + 376, + 366, + 384 + ], + "spans": [ + { + "bbox": [ + 311, + 376, + 366, + 384 + ], + "type": "text", + "content": "(e) Ground truth." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 371, + 328, + 430, + 370 + ], + "blocks": [ + { + "bbox": [ + 371, + 328, + 430, + 370 + ], + "lines": [ + { + "bbox": [ + 371, + 328, + 430, + 370 + ], + "spans": [ + { + "bbox": [ + 371, + 328, + 430, + 370 + ], + "type": "image", + "image_path": "6d2815006760c86947ebb23a96c4c4073489f681b1cfac8826049d37d72af20a.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 373, + 415, + 383 + ], + "lines": [ + { + "bbox": [ + 386, + 373, + 415, + 383 + ], + "spans": [ + { + "bbox": [ + 386, + 373, + 415, + 383 + ], + "type": "text", + "content": "(f) ULIP." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 433, + 328, + 493, + 370 + ], + "blocks": [ + { + "bbox": [ + 433, + 328, + 493, + 370 + ], + "lines": [ + { + "bbox": [ + 433, + 328, + 493, + 370 + ], + "spans": [ + { + "bbox": [ + 433, + 328, + 493, + 370 + ], + "type": "image", + "image_path": "cda03881c1794279a3e0c44c51b6159973f9d3e094f6420e2af978771056e0c8.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 373, + 487, + 383 + ], + "lines": [ + { + "bbox": [ + 438, + 373, + 487, + 383 + ], + "spans": [ + { + "bbox": [ + 438, + 373, + 487, + 383 + ], + "type": "text", + "content": "(g) " + }, + { + "bbox": [ + 438, + 373, + 487, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{ULIP} + \\mathbf{GC}" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 495, + 328, + 555, + 370 + ], + "blocks": [ + { + "bbox": [ + 495, + 328, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 495, + 328, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 495, + 328, + 555, + 370 + ], + "type": "image", + "image_path": "5b8edd9672bad0310a6a5fe8be5fbb81684266b27d00e682112af604761596e1.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 373, + 549, + 383 + ], + "lines": [ + { + "bbox": [ + 501, + 373, + 549, + 383 + ], + "spans": [ + { + "bbox": [ + 501, + 373, + 549, + 383 + ], + "type": "text", + "content": "(h) ULIP+HC." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 310, + 389, + 357, + 425 + ], + "blocks": [ + { + "bbox": [ + 310, + 389, + 357, + 425 + ], + "lines": [ + { + "bbox": [ + 310, + 389, + 357, + 425 + ], + "spans": [ + { + "bbox": [ + 310, + 389, + 357, + 425 + ], + "type": "image", + "image_path": "e2fba44e82830eae34d45912ee571c2ec993c1bd55ec74a93e2839fc6b1dfcf1.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 430, + 365, + 439 + ], + "lines": [ + { + "bbox": [ + 312, + 430, + 365, + 439 + ], + "spans": [ + { + "bbox": [ + 312, + 430, + 365, + 439 + ], + "type": "text", + "content": "(i) Ground truth." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 371, + 388, + 430, + 427 + ], + "blocks": [ + { + "bbox": [ + 371, + 388, + 430, + 427 + ], + "lines": [ + { + "bbox": [ + 371, + 388, + 430, + 427 + ], + "spans": [ + { + "bbox": [ + 371, + 388, + 430, + 427 + ], + "type": "image", + "image_path": "c913cbec7698e2cfed8cca4bd072891d98c4e2b3b728e250a60c65b5159ce646.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 430, + 421, + 440 + ], + "lines": [ + { + "bbox": [ + 381, + 430, + 421, + 440 + ], + "spans": [ + { + "bbox": [ + 381, + 430, + 421, + 440 + ], + "type": "text", + "content": "(j) O-Shape." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 433, + 388, + 493, + 427 + ], + "blocks": [ + { + "bbox": [ + 433, + 388, + 493, + 427 + ], + "lines": [ + { + "bbox": [ + 433, + 388, + 493, + 427 + ], + "spans": [ + { + "bbox": [ + 433, + 388, + 493, + 427 + ], + "type": "image", + "image_path": "7c2af08b62cb0cf90275ddfa836cae7097e71a2c9a4a0f994d43789bc27ddbd4.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 430, + 492, + 440 + ], + "lines": [ + { + "bbox": [ + 433, + 430, + 492, + 440 + ], + "spans": [ + { + "bbox": [ + 433, + 430, + 492, + 440 + ], + "type": "text", + "content": "(k) O-Shape+GC." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 495, + 388, + 555, + 427 + ], + "blocks": [ + { + "bbox": [ + 495, + 388, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 495, + 388, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 495, + 388, + 555, + 427 + ], + "type": "image", + "image_path": "94e02d96d5e670abf5f9bedb16766283f106932a0a94cb1021941ce1e43413ab.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 430, + 553, + 440 + ], + "lines": [ + { + "bbox": [ + 496, + 430, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 496, + 430, + 553, + 440 + ], + "type": "text", + "content": "(1) O-Shape+HC." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 310, + 444, + 354, + 479 + ], + "blocks": [ + { + "bbox": [ + 310, + 444, + 354, + 479 + ], + "lines": [ + { + "bbox": [ + 310, + 444, + 354, + 479 + ], + "spans": [ + { + "bbox": [ + 310, + 444, + 354, + 479 + ], + "type": "image", + "image_path": "7041b5e4a0499d37f542d31007dbf6b1c0b7e1979ae0ea9a238840a5693762e0.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 483, + 368, + 493 + ], + "lines": [ + { + "bbox": [ + 310, + 483, + 368, + 493 + ], + "spans": [ + { + "bbox": [ + 310, + 483, + 368, + 493 + ], + "type": "text", + "content": "(m) Ground truth." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 503, + 545, + 525 + ], + "lines": [ + { + "bbox": [ + 305, + 503, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 545, + 525 + ], + "type": "text", + "content": "Figure 7. The zero-shot predictions before and after adaptation by Point-Cache. GC: global cache. HC: hierarchical cache." + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 371, + 443, + 430, + 479 + ], + "blocks": [ + { + "bbox": [ + 371, + 443, + 430, + 479 + ], + "lines": [ + { + "bbox": [ + 371, + 443, + 430, + 479 + ], + "spans": [ + { + "bbox": [ + 371, + 443, + 430, + 479 + ], + "type": "image", + "image_path": "cde187005839dff345c5eb5eceaa48900ed362a6b62a315c87e7ffaa12cd6b31.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 483, + 419, + 491 + ], + "lines": [ + { + "bbox": [ + 383, + 483, + 419, + 491 + ], + "spans": [ + { + "bbox": [ + 383, + 483, + 419, + 491 + ], + "type": "text", + "content": "(n) Uni3D." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 433, + 443, + 493, + 479 + ], + "blocks": [ + { + "bbox": [ + 433, + 443, + 493, + 479 + ], + "lines": [ + { + "bbox": [ + 433, + 443, + 493, + 479 + ], + "spans": [ + { + "bbox": [ + 433, + 443, + 493, + 479 + ], + "type": "image", + "image_path": "3e3ba34ddb4567227316d2fc8e8387cbc2aaa2883f031822f0c504a47da74daa.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 483, + 488, + 491 + ], + "lines": [ + { + "bbox": [ + 436, + 483, + 488, + 491 + ], + "spans": [ + { + "bbox": [ + 436, + 483, + 488, + 491 + ], + "type": "text", + "content": "o) " + }, + { + "bbox": [ + 436, + 483, + 488, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{Uni3D + GC}" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 495, + 443, + 555, + 479 + ], + "blocks": [ + { + "bbox": [ + 495, + 443, + 555, + 479 + ], + "lines": [ + { + "bbox": [ + 495, + 443, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 495, + 443, + 555, + 479 + ], + "type": "image", + "image_path": "bf042d04e709a9b9597c9c17b4a5e92114c5e44f426cc4a0969ce5953e8708fa.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 483, + 551, + 491 + ], + "lines": [ + { + "bbox": [ + 498, + 483, + 551, + 491 + ], + "spans": [ + { + "bbox": [ + 498, + 483, + 551, + 491 + ], + "type": "text", + "content": "(p) " + }, + { + "bbox": [ + 498, + 483, + 551, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{Uni3D + HC}" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "This paper explores test-time point cloud recognition in a practical and challenging setting, where only test data and the pre-trained models are available. We develop PointCache, a powerful 3D knowledge base that enhances robustness and generalization at test time. It records the fingerprints of online test samples through a global and a local cache, creating a more accurate profile of 3D data. Designed as a plug-and-play module, Point-Cache can be seamlessly integrated into various large 3D models without compromising efficiency, enabling open-vocabulary point cloud recognition for both known and unseen classes, effectively addressing limitations of prior test-time methods." + } + ] + } + ], + "index": 54 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1270" + } + ] + } + ], + "index": 55 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 145, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 145, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 145, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 259 + ], + "type": "text", + "content": "We thank all anonymous reviewers and area chairs for their time and valuable feedback. This work was partially supported by China Scholarship Council (CSC) under the Grant No. 202306360147 and Australian Research Council funding under the Grant No. DE250100030. Dr. Qiuhong Ke is partially supported by the Australian Research Council funding Discovery Early Career Researcher Award (DE250100030). Dr. Deying Li was partially supported by the National Natural Science Foundation of China under the Grant No. 12071478. Dr. Yongcai Wang was partially supported by the National Natural Science Foundation of China under the Grant No. 61972404, the Public Computing Cloud at Renmin University of China, and the Blockchain Lab, School of Information, Renmin University of China." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 270, + 107, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 270, + 107, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 270, + 107, + 282 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 289, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 53, + 289, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 289, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 289, + 287, + 345 + ], + "type": "text", + "content": "[1] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. CoRR, abs/1512.03012, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 346, + 288, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 389 + ], + "type": "text", + "content": "[2] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. PointGPT: Auto-regressively generative pre-training from point clouds. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "text", + "content": "[3] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations, 2019. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "text", + "content": "[4] Yinbo Chen, Zhuang Liu, Huijuan Xu, Trevor Darrell, and Xiaolong Wang. Meta-baseline: Exploring simple meta-learning for few-shot learning. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9042-9051, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 492, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 492, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 492, + 287, + 546 + ], + "type": "text", + "content": "[5] Sungha Choi, Seunghan Yang, Seokeon Choi, and Sungrack Yun. Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. In European Conference on Computer Vision, pages 440–458. Springer, 2022. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 548, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 287, + 635 + ], + "type": "text", + "content": "[6] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of " + }, + { + "bbox": [ + 53, + 548, + 287, + 635 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 53, + 548, + 287, + 635 + ], + "type": "text", + "content": " objects. In Advances in Neural Information Processing Systems, pages 35799-35813. Curran Associates, Inc., 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 636, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 288, + 712 + ], + "type": "text", + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13142-13153, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 148 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 150, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 150, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 150, + 547, + 205 + ], + "type": "text", + "content": "[9] Lunhao Duan, Shanshan Zhao, Nan Xue, Mingming Gong, Gui-Song Xia, and Dacheng Tao. Condaformer: Disassembled transformer with local structure enhancement for 3d point cloud understanding. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 206, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 270 + ], + "type": "text", + "content": "[10] Hehe Fan, Xiaojun Chang, Wanyue Zhang, Yi Cheng, Ying Sun, and Mohan Kankanhalli. Self-supervised global-local structure modeling for point cloud domain adaptation with reliable voted pseudo labels. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6367-6376, 2022. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 272, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 327 + ], + "type": "text", + "content": "[11] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and ZhaoXiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2918-2927, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 327, + 547, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 547, + 381 + ], + "type": "text", + "content": "[12] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2704-2714, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[13] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning, pages 1126–1135. PMLR, 2017. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "text", + "content": "[14] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei Efros. Test-time training with masked autoencoders. In Advances in Neural Information Processing Systems, pages 29374-29385. Curran Associates, Inc., 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[15] Edouard Grave, Moustapha M Cisse, and Armand Joulin. Unbounded cache model for online language modeling with open vocabulary. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[16] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 559, + 547, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 547, + 613 + ], + "type": "text", + "content": "[17] Ahmed Hatem, Yiming Qian, and Yang Wang. Point-tta: Test-time adaptation for point cloud registration using multitask meta-auxiliary learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16494-16504, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 615, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 658 + ], + "type": "text", + "content": "[18] Ahmed Hatem, Yiming Qian, and Yang Wang. Test-time adaptation for point cloud upsampling using meta-learning. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1284–1291, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[19] Chao Huang, Zhangjie Cao, Yunbo Wang, Jianmin Wang, and Mingsheng Long. Metasets: Meta-learning on point sets for generalizable representations. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8859-8868, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1271" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[20] Zhe Huang, Shuo Wang, Yongcai Wang, Wanting Li, Deying Li, and Lei Wang. Roco: Robust cooperative perception by iterative object matching and pose adjustment. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7833-7842, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[21] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In Advances in Neural Information Processing Systems, pages 2427–2440. Curran Associates, Inc., 2021. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "text", + "content": "[22] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 297 + ], + "type": "text", + "content": "[23] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14162-14171, 2024. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 342 + ], + "type": "text", + "content": "[24] Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor language models. In International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 343, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 397 + ], + "type": "text", + "content": "[25] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 442 + ], + "type": "text", + "content": "[26] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "text", + "content": "[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? Source hypothesis transfer for unsupervised domain adaptation. In Proceedings of the 37th International Conference on Machine Learning, pages 6028-6039. PMLR, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 555 + ], + "type": "text", + "content": "[28] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 557, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 622 + ], + "type": "text", + "content": "[29] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In Advances in Neural Information Processing Systems, pages 44860-44879. Curran Associates, Inc., 2023. 2, 3, 5, 6, 1, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[30] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In Advances in Neural Information Processing Systems, pages 21808-21820. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[31] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d an" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "notation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1190-1199, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "text", + "content": "[32] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In International Conference on Learning Representations, 2022. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 141, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 175 + ], + "type": "text", + "content": "[33] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. In International Conference on Learning Representations, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 230 + ], + "type": "text", + "content": "[34] M. Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 232, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 297 + ], + "type": "text", + "content": "[35] M. Jehanzeb Mirza, Inkyu Shin, Wei Lin, Andreas Schriebl, Kunyang Sun, Jaesung Choe, Mateusz Kozinski, Horst Possegger, In So Kweon, Kuk-Jin Yoon, and Horst Bischof. Mate: Masked autoencoders are online 3d test-time learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16709-16718, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "type": "text", + "content": "[36] Emin Orhan. A simple cache model for image recognition. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 333, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 388 + ], + "type": "text", + "content": "[37] Jinyoung Park, Sanghyeok Lee, Sihyeon Kim, Yunyang Xiong, and Hyunwoo J. Kim. Self-positioning point-based transformer for point cloud understanding. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21814-21823, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 389, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 389, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 389, + 545, + 442 + ], + "type": "text", + "content": "[38] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc Le. Combined scaling for open-vocabulary image classification. 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "text", + "content": "[39] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "text", + "content": "[40] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "type": "text", + "content": "[41] Can Qin, Haoxuan You, Lichen Wang, C.-C. Jay Kuo, and Yun Fu. Pointdan: A multi-scale 3d domain adaption network for point cloud representation. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 613, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 689 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[43] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions. In In" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1272" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "ternational Conference on Machine Learning (ICML), 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[44] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 130, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 130, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 130, + 287, + 185 + ], + "type": "text", + "content": "[45] Yuefan Shen, Yanchao Yang, Mi Yan, He Wang, Youyi Zheng, and Leonidas Guibas. Domain adaptation on point clouds via geometry-aware implicit. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7213-7222, 2022. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 231 + ], + "type": "text", + "content": "[46] Hajin Shim, Changhun Kim, and Eunho Yang. CloudFixer: Test-Time Adaptation for 3D Point Clouds via Diffusion-Guided Geometric Transformation. In European Conference on Computer Vision (ECCV), 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 232, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 232, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 287, + 297 + ], + "type": "text", + "content": "[47] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In Advances in Neural Information Processing Systems, pages 14274-14289. Curran Associates, Inc., 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 299, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 299, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 287, + 342 + ], + "type": "text", + "content": "[48] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 343, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 287, + 398 + ], + "type": "text", + "content": "[49] Hongyu Sun, Yongcai Wang, Xudong Cai, Xuewei Bai, and Deying Li. Vipformer: Efficient vision-and-pointcloud transformer for unsupervised pointcloud understanding. In IEEE International Conference on Robotics and Automation (ICRA), 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 399, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 287, + 464 + ], + "type": "text", + "content": "[50] Hongyu Sun, Qiuhong Ke, Yongcai Wang, Wang Chen, Kang Yang, Deying Li, and Jianfei Cai. Point-prc: A prompt learning based regulation framework for generalizable point cloud analysis. In The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS), 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 467, + 287, + 510 + ], + "type": "text", + "content": "[51] Jiachen Sun, Qingzhao Zhang, Bhavya Kailkhura, Zhiding Yu, Chaowei Xiao, and Z. Morley Mao. Benchmarking robustness of 3d point cloud recognition against common corruptions, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 49, + 512, + 287, + 567 + ], + "type": "text", + "content": "[52] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Proceedings of the 37th International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 287, + 622 + ], + "type": "text", + "content": "[53] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. Proceedings of the AAAI Conference on Artificial Intelligence, 38(6):5171-5179, 2024. 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "type": "text", + "content": "[54] Yun-Yun Tsai, Fu-Chen Chen, Albert Y. C. Chen, Junfeng Yang, Che-Chun Su, Min Sun, and Cheng-Hao Kuo. Gda: Generalized diffusion for robust test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23242-23251, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[55] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2725-2736, 2023. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 162 + ], + "type": "text", + "content": "[56] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "text", + "content": "[58] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, koray kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[59] Anran Wang, Jiwen Lu, Jianfei Cai, Tat-Jen Cham, and Gang Wang. Large-margin multi-modal deep learning for rgb-d object recognition. IEEE Transactions on Multimedia, 17 (11):1887-1898, 2015. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "text", + "content": "[60] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 355, + 545, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 411 + ], + "type": "text", + "content": "[61] Haiyang Wang, Chen Shi, Shaoshuai Shi, Meng Lei, Sen Wang, Di He, Bernt Schiele, and Liwei Wang. Dsvt: Dynamic sparse voxel transformer with rotated sets. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13520-13529, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "type": "text", + "content": "[62] Haiyang Wang, Hao Tang, Shaoshuai Shi, Aoxue Li, Zhenguo Li, Bernt Schiele, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 456, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 522 + ], + "type": "text", + "content": "[63] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q. Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8437-8445, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 524, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 545, + 568 + ], + "type": "text", + "content": "[64] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 2019. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 569, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 545, + 634 + ], + "type": "text", + "content": "[65] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q. Weinberger, and WeiLun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11710-11720, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 636, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 712 + ], + "type": "text", + "content": "[66] Yanshuo Wang, Ali Cheraghian, Zeeshan Hayden, Jie Hong, Sameera Ramasinghe, Shafin Rahman, David Ahmedt-Aristizabal, Xuesong Li, Lars Petersson, and Mehrtash Harandi. Backpropagation-free network for 3d test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23231-23241, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1273" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[67] Xin Wei, Xiang Gu, and Jian Sun. Learning generalizable part-based feature representation for 3d point clouds. In Advances in Neural Information Processing Systems, pages 29305-29318. Curran Associates, Inc., 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 184 + ], + "type": "text", + "content": "[68] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7): 5092-5113, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "type": "text", + "content": "[69] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 803-814, 2023. 1, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "text", + "content": "[70] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 320, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 374 + ], + "type": "text", + "content": "[71] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 915-924, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 376, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 452 + ], + "type": "text", + "content": "[72] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning unified representation of language, image and point cloud for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6, 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 455, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 288, + 520 + ], + "type": "text", + "content": "[73] Le Xue, Ning Yu, Shu Zhang, Junnan Li, Roberto Martín-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 5, 6, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 522, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 577 + ], + "type": "text", + "content": "[74] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "text", + "content": "[75] Teresa Yeo, Oğuzhan Fatih Kar, Zahra Sodagar, and Amir Zamir. Rapid network adaptation: learning to adapt neural networks using test-time feedback. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4651-4664, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[76] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In Advances in Neural Information Processing Systems, pages 32215-32234. Curran Associates, Inc., 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[77] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "text", + "content": "transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19313-19322, 2022. 4, 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 107, + 546, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 546, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 546, + 161 + ], + "type": "text", + "content": "[78] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23783-23793, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 162, + 546, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 546, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 546, + 226 + ], + "type": "text", + "content": "[79] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18123-18133, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 228, + 546, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 546, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 546, + 271 + ], + "type": "text", + "content": "[80] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. In Advances in Neural Information Processing Systems, pages 38629-38642. Curran Associates, Inc., 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 272, + 546, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 546, + 327 + ], + "type": "text", + "content": "[81] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8552-8562, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 327, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 546, + 403 + ], + "type": "text", + "content": "[82] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXV, pages 493-510, Berlin, Heidelberg, 2022. Springer-Verlag. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 404, + 546, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 546, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 546, + 470 + ], + "type": "text", + "content": "[83] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15211-15222, 2023. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 471, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 546, + 525 + ], + "type": "text", + "content": "[84] Renrui Zhang, Liuhui Wang, Yali Wang, Peng Gao, Hongsheng Li, and Jianbo Shi. Starting from non-parametric networks for 3d point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5344-5353, 2023. 3, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 526, + 546, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 546, + 579 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 546, + 579 + ], + "type": "text", + "content": "[85] Weichen Zhang, Wen Li, and Dong Xu. Srdan: Scale-aware and range-aware domain adaptation network for cross-dataset 3d object detection. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6765-6775, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 581, + 546, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 546, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 546, + 613 + ], + "type": "text", + "content": "[86] Zhimin Zhang, Xiang Gao, and Wei Hu. Invariantoodg: Learning invariant features of point clouds for out-of-distribution generalization, 2024. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 614, + 547, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 547, + 658 + ], + "type": "text", + "content": "[87] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H.S. Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16259-16268, 2021. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 546, + 712 + ], + "type": "text", + "content": "[88] Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 5, 6, 7, 4" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1274" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 128 + ], + "type": "text", + "content": "[89] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1275" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/7bc00c8f-f0e1-486b-af9f-cab0b77e9a84_content_list.json b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/7bc00c8f-f0e1-486b-af9f-cab0b77e9a84_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..05e7a4545d0e546eb60d162a1443d80f33bdd34d --- /dev/null +++ b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/7bc00c8f-f0e1-486b-af9f-cab0b77e9a84_content_list.json @@ -0,0 +1,2359 @@ +[ + { + "type": "text", + "text": "Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting", + "text_level": 1, + "bbox": [ + 135, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Lin, Chenyang Zhao, and Antoni B. Chan", + "bbox": [ + 305, + 181, + 689, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Science, City University of Hong Kong", + "bbox": [ + 241, + 199, + 754, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "elonlin24@gmail.com, chenyzhao9-c@my.cityu.edu.hk, abchan@cityu.edu.hk", + "bbox": [ + 184, + 219, + 805, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point detection has been developed to locate pedestrians in crowded scenes by training a counter through a point-to-point (P2P) supervision scheme. Despite its excellent localization and counting performance, training a point-based counter still faces challenges concerning annotation labor: hundreds to thousands of points are required to annotate a single sample capturing a dense crowd. In this paper, we integrate point-based methods into a semi-supervised counting framework based on pseudo-labeling, enabling the training of a counter with only a few annotated samples supplemented by a large volume of pseudolabeled data. However, during implementation, the training encounters issues as the confidence for pseudo-labels fails to be propagated to background pixels via the P2P. To tackle this challenge, we devise a point-specific activation map (PSAM) to visually interpret the phenomena occurring during the ill-posed training. Observations from the PSAM suggest that the feature map is excessively activated by the loss for unlabeled data, causing the decoder to misinterpret these over-activations as pedestrians. To mitigate this issue, we propose a point-to-region (P2R) scheme to substitute P2P, which segments out local regions rather than detects a point corresponding to a pedestrian for supervision. Consequently, pixels in the local region can share the same confidence with the corresponding pseudo points. Experimental results in both semi-supervised counting and unsupervised domain adaptation highlight the advantages of our method, illustrating P2R can resolve issues identified in PSAM. The code is available at https://github.com/Elin24/P2RLoss.", + "bbox": [ + 89, + 301, + 483, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 768, + 220, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Crowd scene analysis has been investigated for many years due to its significant applications in smart cities and urban safety [2, 7, 29, 41, 46, 62, 75]. Crowd counting and localization in extremely dense crowd scenes, as the fundamental task in this field, have attracted considerable attention from researchers in computer vision [4, 5, 37, 61, 66, 68]. Although most supervised crowd counting methods fol", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f8654d6fa55092cd93b854925481932197aa3f3617e69f7065ee6cc69b0c2dcb.jpg", + "image_caption": [ + "Figure 1. The workflow of semi-supervised point-based counting methods. The teacher model generates pseudo labels by extracting the foreground pixels, while the student model takes the corresponding strongly augmented image as input to construct the computation graph. The training loss between the pseudo label and the student's prediction involves two steps: the proposed P2R matching and the weighted cross-entropy computation." + ], + "image_footnote": [], + "bbox": [ + 522, + 268, + 620, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e4ee55c1d884fd65d86c1d6233a1875ae7bb5d03211fa2166453f6520e603bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 276, + 686, + 321 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/29dad3d08f89e07b65a71e50ecf6be38f07b709bc2245095e9ed0104f2a45e1e.jpg", + "image_caption": [ + "Prediction", + "w/o gradient", + "P2R Matching" + ], + "image_footnote": [], + "bbox": [ + 696, + 273, + 774, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/928f6f3b2ac36d47a69f7b9084e69bcb6d14188db1b83eb2cc32f47965cdd4e1.jpg", + "image_caption": [ + ">0.5", + "Sudo Label", + "Weighted Cross Entropy", + "w/ gradient" + ], + "image_footnote": [], + "bbox": [ + 818, + 276, + 851, + 316 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/edd9073bc1c491cc4c1267cd01898ab7ba8b88a9b3e9d4acce3d0c2a8fb6e431.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 324, + 620, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ee8b5bdfd9bd808a379c9bfaf6cbdd6c20f24a0ccbb21ec8001ceeddf4b4b32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 320, + 683, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b656620bb22599a32d692f01e06c6b70d78e1199393337a31b3729d11c0f8067.jpg", + "image_caption": [ + "Student" + ], + "image_footnote": [], + "bbox": [ + 630, + 349, + 687, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "low the scheme of density regression [10, 24, 38, 61, 73], there are also relevant studies implementing point detection for crowd localization and counting [21, 31, 57], due to its straightforwardness in indicating pedestrians' positions. These approaches use the Hungarian algorithm to perform point-to-point (P2P) matching between predicted point maps and ground-truth (GT) pedestrian annotations during training. Therefore, these matched pixels are considered as pedestrians' positions, while the others are designated as background for loss computation. This scheme successfully trains a point detector for crowd analysis, yielding both good localization and counting performance.", + "bbox": [ + 509, + 532, + 906, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Great progress has been made in supervised counting, but the labor-intensive nature of annotating data has hindered the development of crowd understanding for a long time [26, 66]. Addressing this issue, semi-supervised counting is gradually being developed, aiming to train a model using large amounts of unlabeled data alongside only a few annotated samples. Similar to supervised counting, most semi-supervised methods also follow the procedure of density regression. Notably, the method proposed in OT-M [26] stands out by employing a simple self-training scheme based on mean-teacher [59] and FixMatch [56]. However, there is no study exploring semi-supervised counting and", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29363", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "localization using a point-detection method. To fill this gap, this paper explores an approach to train a point detector with limited annotated data.", + "bbox": [ + 89, + 90, + 480, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1, typical semi-supervised counting methods rely on pseudo-labeling, which generates pseudo labels to serve as learning targets for unlabeled samples using a model pre-trained on a small amount of annotated data. During training, there is a confidence score to determine whether a concerned prediction should be retained as a pseudo label. However, if the P2P matching strategy is applied to compute loss within the pipeline presented in Fig. 1, there is no reasonable way to define the confidence for unmatched background pixels, as P2P matching only cares about matched foreground pixels. Specifically, a foreground pixel can be retained for training if its matched pseudo-label's foreground probability exceeds a predefined threshold, similar to the approach used in semi-supervised classification tasks [15, 56]. However, background pixels lack confidence scores since no pseudo-labels are assigned to them after matching. Experimental results also demonstrate that ignoring the background leads to the breakdown of semi-supervised training.", + "bbox": [ + 91, + 138, + 482, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issue, we propose a gradient-based visual explanation method (akin to [48, 74, 76, 78]), termed point-specific activation map (PSAM), to visualize regions of interest and the activation levels of detected points. Using PSAM, we observe that the activation values and coverage areas of foreground pixels continuously grow during semi-supervised training, causing the model to misclassify neighboring pixels as foreground.", + "bbox": [ + 89, + 426, + 482, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspecting PSAMs reveals that each point in GT is implicitly associated with a local region rather than a single pixel in the prediction. Accordingly, we propose to replace the P2P with a point-to-region (P2R) strategy, which matches local regions to points in the pseudo-label. This allows all pixels within a matched region, including background pixels, to share the confidence of the corresponding pseudo-point. Experimental results and visualizations demonstrate that P2R can effectively train a semisupervised point-based counter and achieve expected performance. Experimental results show that P2R effectively trains semi-supervised point-based counters with comparable performance to P2P but at significantly reduced computational cost. Unlike P2P [57], P2R eliminates the need for the Hungarian algorithm by assigning each ground-truth or pseudo-point a local \"dominated zone\" and selecting a representative foreground pixel based on a predefined cost matrix.", + "bbox": [ + 91, + 549, + 482, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of our paper are as follows:", + "bbox": [ + 109, + 823, + 413, + 837 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We formulate a pipeline of semi-supervised counting with a point-detection method, but find a problem leading the training breakdown.", + "- We propose Point-Specific Activation Map (PSAM) to" + ], + "bbox": [ + 109, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "visualize the regions of interest and the activation values of each point detected by the point-based crowd counter. Through PSAM, we observe that the surrounding pixels of each expected foreground pixel are over-activated, causing the decoder to recognizes these pixels as instances different from the concerned one.", + "bbox": [ + 544, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Based on insights from PSAM, we propose replacing the P2P with a P2R matching strategy for semi-supervised learning in point-based crowd counting. This not only makes the training process work as expected, but also brings additional benefits of reducing computational requirements.", + "- The proposed P2R achieves outstanding performance in several experiments, including semi-supervised counting and unsupervised domain adaptation. Ablation study shows that P2R excellently tackles the problem caused by P2P in semi-supervised counting." + ], + "bbox": [ + 532, + 181, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 361, + 663, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object counting in dense scenes. The development of deep learning has significantly influenced crowd counting. Prior to this, detecting body parts [16, 17] was the common pipeline for crowd understanding in computer vision, although some methods implemented counting models without explicit object segmentation or tracking information [4-6]. Beyond specific crowd scenarios, class-agnostic counting [14, 27, 30, 45, 49, 71] has also been developed to estimate the distribution of objects in the same category. Most recent counting methods implement neural networks following the density map regression scheme [3, 8, 20, 50, 60, 65, 67, 69, 73]. In addition to continuous innovations in model architecture design [3, 20, 24], new loss functions have also been developed, including Bayesian Loss [38], loss functions based on balanced [63] and unbalanced optimal transport [22, 39, 61], and characteristic function loss computed in the frequency domain [51, 52]. A study analyzes the relationship between current loss functions from the perspective of proximal mapping [28]. Some methods also perform crowd analysis via instance segmentation [1, 9] and point detection [21, 31, 47, 57]. The latter is the scheme adopted in our semi-supervised model.", + "bbox": [ + 511, + 387, + 906, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-supervised counting. Semi-supervised counting has attracted broad attention recently [18, 19, 23, 25, 26, 32-34, 36, 42, 54]. L2R [32, 33] introduces a ranking rule for unlabeled crowd images based on the principle that cropped images should have smaller counts than full images. GP [54] iteratively updates pseudo-labels using Gaussian Processes. IRAST [34] and SUA [42] enhance models by incorporating a segmentation branch. DAC [23] uses learnable density agents to capture varying crowd densities. OT-M [26] transforms density maps into point maps as pseudo-labels, achieving strong performance with a density-to-point loss function [61]. Our method also em", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29364", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ploys pseudo-labeling but uses a point detector for counting, avoiding the need for density map transformations as in [26]. Additionally, uncertainty is directly measured by the scores of extracted points, without any additional learnable structures like those in [42].", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Explanation. Visualizing the importance of extracted image features is a straightforward way to interpret a model. The mainstream of visual explanation focuses on methods producing activation maps [13, 44, 48, 64, 74, 76-78], which are the products of feature maps from some intermediate layer and weight maps indicating the corresponding feature's importance. For example, GradCAM [13, 48] defines the weights using the corresponding gradient maps, generated via backpropagation. Beyond the focus on classification tasks, ODAM [74, 76] extends gradient-based CAM to object detection, visualizing the local impact of features on the detector's decisions. In this paper, we propose PSAM to visualize how the activation map changes for each predicted point in an ill-posed semi-supervised counting framework. Based on the observations from PSAM, we design a P2R matching strategy to avoid the over-activation of pixels in local regions, which effectively resolves the drawbacks of P2P and enables the successful training of a semi-supervised point detector.", + "bbox": [ + 91, + 167, + 483, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Point Counter in Semi-Supervised Counting", + "text_level": 1, + "bbox": [ + 89, + 468, + 483, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section formulates a simplified point-based counter by removing the regression branch in P2PNet [57], and then presents the training scheme and challenges when training with labeled and unlabeled data, respectively.", + "bbox": [ + 89, + 494, + 483, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary on A Simplified Point Counter", + "text_level": 1, + "bbox": [ + 89, + 566, + 455, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Common point-based crowd counters like P2PNet [57] contain a classification and a regression branch. However, we find a simplification can be achieved by removing the offset regression branch. On one hand, the crowd count can be estimated by counting foreground pixels with scores greater than 0.5, a process that does not involve the regression branch. On the other hand, pixel coordinates alone are sufficient to locate pedestrians, as demonstrated in STEERER [10] and GL [61], which use the coordinates of local maxima to indicate pedestrians locations.", + "bbox": [ + 89, + 590, + 483, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a result, P2PNet can be simplified to:", + "bbox": [ + 109, + 741, + 380, + 755 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {F} (I, \\Theta_ {f}) \\rightarrow F \\in \\mathbb {R} ^ {c \\times h \\times w}, (1) \\\\ \\mathcal {D} (F, \\Theta_ {d}) \\rightarrow D \\in \\mathbb {R} ^ {h \\times w} (2) \\\\ \\rightarrow \\mathcal {P} = \\left\\{\\boldsymbol {p} \\in \\mathbb {R} ^ {n}, \\boldsymbol {x} \\in \\mathbb {R} ^ {n \\times 2} \\right\\}, (3) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 767, + 482, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{F}$ is the image encoder with learnable parameters $\\Theta_f$ , and $\\mathcal{D}$ denotes the density decoder with learnable parameter $\\Theta_d$ . Notably, the offset regression branch is excluded in this simplified formulation. The pixel set $\\pmb{p} \\in \\mathbb{R}^n$", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "represents the flattened values in $D$ , and its corresponding coordinate set is denoted as $\\pmb{x} \\in \\mathbb{R}^{n \\times 2}$ , where $n = h \\times w$ .", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Training with Labeled Data via P2P Matching", + "text_level": 1, + "bbox": [ + 511, + 128, + 903, + 146 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The GT with $m$ points can be denoted as $\\pmb{x}^{\\prime} \\in \\mathbb{R}^{m \\times 2}$ , where $\\pmb{x}_{[j]}^{\\prime} \\in \\mathbb{R}^{2}$ represents the location of the $j$ -th pedestrian. Using $\\mathcal{P}$ and $\\pmb{x}^{\\prime}$ , the loss computation involves two parts: P2P matching and binary cross entropy (BCE) calculation.", + "bbox": [ + 511, + 152, + 903, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The former aims to obtain a matrix-based solution $\\mathbf{M} \\in \\{0,1\\}^{n \\times m}$ that minimizes the bipartite-graph matching cost between $\\pmb{x}$ and $\\pmb{x}'$ within the cost matrix $\\mathbf{C} \\in \\mathbb{R}^{n \\times m}$ :", + "bbox": [ + 511, + 214, + 903, + 258 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {[ i j ]} = \\tau \\| \\boldsymbol {x} _ {[ i ]} - \\boldsymbol {x} ^ {\\prime} _ {[ j ]} \\| _ {2} - S (\\boldsymbol {p} _ {[ i ]}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 268, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S(\\cdot)$ is defined as the inverse sigmoid function $S(p) = -\\log (1 / p - 1)$ , rather than the identity operator used in the vanilla P2PNet [57], for improved performance.", + "bbox": [ + 511, + 294, + 903, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To compute BCE loss, the learning objective of $\\pmb{p}$ is formulated as $\\hat{\\pmb{p}} = \\mathbf{M}\\mathbf{1}_m$ , and the BCE loss is calculated for optimization as follows:", + "bbox": [ + 511, + 339, + 903, + 383 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l} = - \\lambda \\hat {\\boldsymbol {p}} ^ {\\top} \\log \\boldsymbol {p} - \\left(\\mathbf {1} _ {n} - \\hat {\\boldsymbol {p}}\\right) ^ {\\top} \\log \\left(\\mathbf {1} _ {n} - \\boldsymbol {p}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 392, + 903, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda$ is a re-weighting factor for matched pixels [57].", + "bbox": [ + 511, + 420, + 885, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Problem in Training with Unlabeled Image", + "text_level": 1, + "bbox": [ + 511, + 443, + 879, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we consider training with unlabeled data following the standard pseudo-labeling procedure [15, 56]. As illustrated in Fig. 1, it requires the cooperation of a student and teacher model that share the same structure described in (1)~(3). The difference lies in that the teacher model is updated via exponential moving average (EMA) [59], while the student is optimized via gradient descending. During training, the teacher and student process the weakly and strongly data-augmented versions of an unlabeled image, producing $\\mathcal{P}_t = \\{\\pmb{p}_t, \\pmb{x}_t\\}$ and $\\mathcal{P}_s = \\{\\pmb{p}_s, \\pmb{x}_s\\}$ , respectively. After that, foreground points are extracted from $\\mathcal{P}_t$ to construct pseudo-labels:", + "bbox": [ + 511, + 465, + 906, + 648 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {t} ^ {\\prime} = \\left\\{\\boldsymbol {p} _ {t} ^ {\\prime}, \\boldsymbol {x} _ {t} ^ {\\prime} \\right\\} = \\left\\{\\boldsymbol {p} _ {t [ j ]}, \\boldsymbol {x} _ {t [ j ]} \\mid \\boldsymbol {p} _ {t [ j ]} > 0. 5 \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 656, + 903, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{x}_t^\\prime$ is treated as the pseudo hard label to estimate the matching matrix $\\mathbf{M}_{st}$ via Hungarian algorithm, and the loss is obtained by substituting $\\hat{\\pmb{p}}_t = \\mathbf{M}_{st}\\mathbf{1}_m$ and $\\pmb{p}_s$ into (5).", + "bbox": [ + 511, + 684, + 903, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, some pseudo-labels may not be reliable. A common solution is to only consider points in $\\mathcal{P}_t^\\prime$ with high scores [15, 56], i.e., using a confidence vector $\\zeta$ to mask out elements whose score is below a threshold $\\eta$ :", + "bbox": [ + 511, + 729, + 903, + 790 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta = \\mathbb {1} \\left(\\boldsymbol {p} _ {t} ^ {\\prime} > \\eta\\right) \\xrightarrow {\\text {m a p t o} \\hat {\\boldsymbol {p}} _ {t} \\text {b y} \\mathbf {M} _ {s t}} \\boldsymbol {z} = \\mathbf {M} _ {s t} \\zeta . \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 799, + 903, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note $\\eta$ is greater than 0.5 and (7) only let pseudo points with score greater than $\\eta$ be reliable. Subsequently, $\\mathbf{Z} = \\mathrm{diag}(z)$ is incorporated into the BCE loss for supervision:", + "bbox": [ + 511, + 830, + 903, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = - \\lambda \\hat {\\boldsymbol {p}} _ {t} ^ {\\top} \\mathbf {Z} \\log \\boldsymbol {p} _ {s} - \\left(\\mathbf {1} _ {n} - \\hat {\\boldsymbol {p}} _ {t}\\right) ^ {\\top} \\mathbf {Z} \\log \\left(\\mathbf {1} _ {n} - \\boldsymbol {p} _ {s}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 883, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29365", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the second item of (8) in P2P is identically zero:", + "bbox": [ + 89, + 90, + 477, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u [ P 2 P ]} = - \\lambda \\hat {\\boldsymbol {p}} _ {t} ^ {\\top} \\mathbf {Z} \\log \\boldsymbol {p} _ {s} - 0, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 111, + 482, + 130 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "regardless of $\\zeta$ (embedded in $\\mathbf{Z}$ ). Consequently, no gradients are back-propagated from the losses associated with background pixels, which is ill-posed. Utilizing (9) as the loss function is similar to training a two-class classifier with only samples from the category \"1\", which will lead the trained model to output \"1\" regardless of the input.", + "bbox": [ + 89, + 135, + 483, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Visual Explanation with PSAM", + "text_level": 1, + "bbox": [ + 89, + 237, + 377, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the theoretical analysis, we implement a framework that supervises a model using labeled and unlabeled data via (5) and (9), respectively, to empirically investigate the behavior of P2P-based semi-supervised counting. Subsequently, PSAM is designed to visualize how the prediction is influenced by (9) during training. Thus, we can design an effective confidence scheme accordingly to counteract the effects revealed by PSAM, thereby enabling semi-supervised learning to function as expected.", + "bbox": [ + 89, + 263, + 483, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. PSAM: Point-Specific Activation Map", + "text_level": 1, + "bbox": [ + 89, + 406, + 419, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An activation map is a heat map with the same spatial resolution as the image features, where local regions with relatively high values are interpreted as being discriminative (informative) for the current output [48, 74, 78]. Most techniques obtain activation maps by linearly aggregating features with their corresponding gradients, as [53, 58] have shown that the partial derivative with respect to a specific element can reflect its contribution to the final prediction.", + "bbox": [ + 89, + 429, + 483, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike [48, 78], which compute gradients at the image level, crowd counting requires a point-specific activation map (PSAM) to visually explain the regions of interest for every element in $\\pmb{p}$ of (3). Thus, the Jacobian matrix of $\\pmb{p}$ with respect to image feature $F$ is required:", + "bbox": [ + 89, + 550, + 483, + 626 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {J} _ {\\mathcal {D}} = \\frac {\\partial \\boldsymbol {p}}{\\partial F} = \\left[ \\nabla \\boldsymbol {p} _ {[ 1 ]}, \\nabla \\boldsymbol {p} _ {[ 2 ]}, \\dots , \\nabla \\boldsymbol {p} _ {[ n ]} \\right] ^ {\\top}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 631, + 480, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterwards, the $q$ -th heat map $H_{[q]}$ in PSAM is derived by filtering out negative influences on the aggregated feature:", + "bbox": [ + 89, + 657, + 482, + 688 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {[ q ]} = \\max \\left(\\sum_ {k = 1} ^ {c} \\nabla \\boldsymbol {p} _ {[ q ]} \\odot F, 0\\right) \\in \\mathbb {R} ^ {h \\times w}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 705, + 482, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the summation operator is applied along the channel dimension, and $\\odot$ denotes the Hadamard product.", + "bbox": [ + 89, + 734, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a naive implementation, the computation of $\\mathbf{J}_{\\mathcal{D}} \\in \\mathbb{R}^{n \\times hwc}$ requires a large memory and involves $n$ backpropagations through the computational graph, posing a challenge for visual explanation in pixel-level tasks. However, this issue can be resolved by computing gradients only between $\\pmb{p}_{[q]}$ and features within its receptive field. Gradients in other areas are zero, as there is no path connecting them in the computational graph. Thus, an efficient approach can be implemented, as illustrated in Fig. 2.", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d58de3567dcc2d6d9e22e47fb4788cd6add94e05ed944dc3cff47ef66ecd303.jpg", + "image_caption": [ + "Figure 2. The generation process of PSAM." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this implementation, the forward process begins with extracting sliding local blocks $\\tilde{F}$ according to the receptive field $r$ of $\\mathcal{D}$ , and then decoding $\\tilde{F}$ through (2) in parallel. The final prediction $\\pmb{p}$ is generated by stacking pixels located at the center of $r\\times r$ patches denoted as in $\\tilde{D}$ :", + "bbox": [ + 511, + 340, + 905, + 416 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF \\in \\mathbb {R} ^ {h \\times w \\times c} \\rightarrow \\tilde {F} \\in \\mathbb {R} ^ {(h w) \\times (r \\times r \\times c)}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 425, + 903, + 444 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} (\\tilde {F}, \\Theta_ {d}) \\rightarrow \\tilde {D} \\in \\mathbb {R} ^ {n \\times (r \\times r)} \\xrightarrow {\\text {c e n t e r o f p a t c h e s}} \\boldsymbol {p} \\in \\mathbb {R} ^ {n}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 448, + 903, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The remaining $r^2 - 1$ pixels in each patch are discarded.", + "bbox": [ + 511, + 479, + 888, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denoting the sum of the output as $\\xi = \\mathbf{1}_n^\\top \\pmb{p}$ , the gradient $\\nabla \\xi = \\partial \\xi / \\partial \\tilde{F}$ can be obtained via just one backpropagation, and the $q$ -th gradient block, $\\nabla \\xi_{[q]}$ , is equal to the gradient of $\\pmb{p}_{[q]}$ with respect to $\\tilde{F}_{[q]}$ :", + "bbox": [ + 511, + 494, + 905, + 560 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla \\xi_ {[ q ]} = \\frac {\\partial \\xi}{\\partial \\tilde {F} _ {[ q ]}} = \\frac {\\partial \\xi}{\\partial \\boldsymbol {p} _ {[ q ]}} \\cdot \\frac {\\partial \\boldsymbol {p} _ {[ q ]}}{\\partial \\tilde {F} _ {[ q ]}} = \\frac {\\partial \\boldsymbol {p} _ {[ q ]}}{\\partial \\tilde {F} _ {[ q ]}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 570, + 903, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "since $\\partial \\xi /\\partial \\pmb{p}_{[q]} = 1$ . After that, we can parallel compute the PSAM (denoted as $\\tilde{H}_{[q]}$ ) within the receptive field of $\\pmb{p}_{[q]}$ by substituting (14) into the formulation presented in (11):", + "bbox": [ + 511, + 607, + 905, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {H} _ {[ q ]} = \\max \\left(\\sum_ {k = 1} ^ {c} \\nabla \\xi_ {[ q ]} \\odot \\tilde {F} _ {[ q ]}, 0\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 666, + 903, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in the bottom of Fig. 2, the complete PSAM $H_{[1,\\dots,n]}$ in (11) can be obtained by filling $\\tilde{H}_{[1,\\dots,n]}$ back into the corresponding regions of a fully zero matrix:", + "bbox": [ + 511, + 699, + 905, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {[ q, t ]} = \\left\\{ \\begin{array}{l l} \\tilde {H} _ {[ q, t ^ {\\prime} ]}, & \\text {i f} t \\in \\Omega_ {q} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 755, + 903, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Omega_q$ represents the index set of $p_{[q]}$ 's receptive field, and $t' \\in \\mathbb{R}^2$ is the transformed coordinate of $t \\in \\mathbb{R}^2$ in $H$ , ensuring $F_{[t]} \\equiv \\tilde{F}_{[q,t']}.$ In this way, the matrix memory required to obtain all $H_{[q]}$ is reduced from $\\mathbb{R}^{n \\times hwc}$ to $\\mathbb{R}^{n \\times r^2c}$ ( $r^2 \\ll hw$ ), and only one back-propagation is needed to obtain PSAMs for all pixels.", + "bbox": [ + 511, + 804, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29366", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/173057d1eb0dda57d364b4a3e1c373941501cb4d7e3c2b4a66c083925e329cff.jpg", + "image_caption": [ + "(a) MAE/MSE during training" + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 218, + 162 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/233930473ae392bc74e00fbda96b9d0a8622f2a75a8bf6b31af8ddc59b84c335.jpg", + "image_caption": [ + "(c) Local PSAM mean (model-L)" + ], + "image_footnote": [], + "bbox": [ + 232, + 90, + 326, + 162 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f7fd1d069acbc0bce15ae0e9894c45afc251c70ce9db4ce699c4d0ec496df5d8.jpg", + "image_caption": [ + "(e) aggregated PSAM (model-L)" + ], + "image_footnote": [], + "bbox": [ + 341, + 90, + 473, + 162 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f4567a78e79a560022aa540449d250304ca3e352ffa4b7367bac20ea4e1bc0b9.jpg", + "image_caption": [ + "(b) Sorted average PSAM" + ], + "image_footnote": [], + "bbox": [ + 102, + 179, + 214, + 246 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/919694c90acae328e6ba18491b9f7b880869903d74fdf7cb9d57c6faaf6c3a68.jpg", + "image_caption": [ + "(d) Local PSAM mean (model-U)", + "Figure 3. Observations in PSAM. (a) The training process, where model-L and model-U are extracted from the 100th and 200th epochs, respectively. (b) Comparing sorted values of PSAMs in the two models. (c) & (d) Visualizing the average of local PSAM, and (e) & (f) the aggregated PSAM to compare model-L and model-U from a global perspective." + ], + "image_footnote": [], + "bbox": [ + 232, + 176, + 328, + 248 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/66416986b940cbb75c8bbbfa470ac1c4a4a42040afda78c7030ca3acfbbf1fdd.jpg", + "image_caption": [ + "(f) aggregated PSAM (model-U)" + ], + "image_footnote": [], + "bbox": [ + 341, + 176, + 473, + 248 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Observation in PSAM", + "text_level": 1, + "bbox": [ + 89, + 366, + 295, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We next apply PSAM to visualize the model during semi-supervised counting. The semi-supervised loss function combines (5) and (8):", + "bbox": [ + 89, + 390, + 482, + 435 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\alpha \\mathcal {L} _ {u} + (1 - \\alpha) \\mathcal {L} _ {l}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 446, + 480, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyperparameter to balance the training of labeled and unlabeled data. Following [15], we set $\\alpha = 0$ during the first 100 epochs to generate valid pseudo labels. After that, $\\alpha$ is gradually increased to 1. In Fig. 3(a), we plot how the MAE/MSE changes on the validation set during training. The training failed as evinced by both metrics increasing significantly when unlabeled data were involved in training (after epoch 100). We denote 'model-L' as the model at the 100-th epoch that is trained only on labeled data, and 'model-U' as the model at the 200-th epoch that was trained with both labeled and unlabeled data.", + "bbox": [ + 88, + 474, + 482, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We next compare explanations of model-L and model-U using PSAM. We visualize the concerned regions of the crowd by extracting the PSAMs corresponding to the foreground pixels $(\\pmb{p}_{[q]} > 0.5)$ . In Fig. 3(c & d) the element-wise mean of the PSAM patch $(\\sum_{q} \\tilde{H}_{[q]} \\cdot (\\pmb{p}_{[q]} > 0.5))$ is displayed to illustrate how the area of the concerned regions changes between model-L and model-U, while Fig. 3(b) presents a detailed value comparison using a line chart. On one hand, Fig. 3(b) shows that the values in model-U's PSAM are larger than those of model-L; on the other hand, Fig. 3(c & d) shows that model-L's PSAM is more concentrated when compared to model-U. These observations reveal how (8) in P2P scheme affects the trained model: $\\mathcal{L}_u$ leads to over-activation in the PSAM of foreground pixels, causing the decoder $\\mathcal{D}$ to falsely identify surrounded pixels as foreground pixels as well.", + "bbox": [ + 89, + 641, + 482, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we visualize the global PSAM of both models", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by compressing $H_{[q]}$ corresponding to foreground pixels: $\\bar{H} = \\sum_{q=1}^{n} (\\pmb{p}_{[q]} > 0.5) \\cdot H_{[q]}$ . As displayed in Fig. 3(e & f), the aggregated PSAM of model-L is sparser than that of model-U, which is consistent with the observation for individual points presented in Fig. 3(c & d). Another observation is that the background regions do not change significantly, indicating that almost all false positives occur in the neighborhood of the initial foreground pixels. This suggests that $\\mathcal{L}_u$ does not alter the hyperplane separating crowd and non-crowd features in the high-dimensional space.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. P2R Matching", + "text_level": 1, + "bbox": [ + 511, + 258, + 661, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The observations described in last section reveal that $\\mathcal{L}_u$ over-activates the values and enlarges the area of the PSAM corresponding to foreground pixels, which should be suppressed by the supervision of background part in (5), but is actually zero in (8) ( $\\mathcal{L}_{u[\\mathrm{P2P}]}^{\\prime}$ in (9)). This over-activation also confirms two important functions of the background item in (5): it not only guides the model to classify the crowd and non-crowd regions in the image, but also suppresses the surrounding person region from being activated. The latter is important since the neighboring pixels of a detected foreground pixel have similar features to the foreground pixel due to the CNN architecture, and thus may become activated if suppression is not applied. This interpretation makes more sense because each person in a crowd image is represented by tens to thousands of pixels, not just one. Supervising the region around the positive pixel with labels of 0 is crucial to a point-based counter. Therefore, during training, it would be more effective to adopt a point-to-region (P2R) matching strategy instead of a point-to-point (P2P) approach. Moreover, as suggested by the above interpretation, P2R is not limited in training with unlabeled data, but can also work in supervised learning with point supervision.", + "bbox": [ + 511, + 284, + 906, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Training with Labeled Data via P2R", + "text_level": 1, + "bbox": [ + 511, + 643, + 826, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recalling the notation of the prediction $\\mathcal{P} = \\{\\pmb {p},\\pmb {x}\\}$ in (3) and ground truth $\\pmb{x}^{\\prime}$ , the matching matrix $\\mathbf{M}$ in our P2R matching is the Hadamard product of two items:", + "bbox": [ + 511, + 667, + 905, + 714 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\mathbf {M} _ {f} \\odot \\left(\\boldsymbol {\\beta} ^ {\\top} \\mathbf {1} _ {m}\\right) \\quad \\in \\mathbf {M} ^ {n \\times m}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 726, + 903, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first item $\\mathbf{M}_f\\in \\{0,1\\}^{n\\times m}$ is a many-to-one matrix that assigns each pixel in $\\mathcal{P}$ to its nearest point in $x^{\\prime}$ :", + "bbox": [ + 511, + 758, + 905, + 789 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {f [ i j ]} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} l _ {2} (i, j) < l _ {2} (i, k) \\forall k \\neq j \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 801, + 903, + 842 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $l_{2}(i,j) = \\| \\pmb{x}_{[i]} - \\pmb{x}^{\\prime}_{[j]}\\|_{2}$ is the $l_{2}$ -distance between the $i$ -th pixel's and the $j$ -th point's coordinates. $\\beta \\in \\{0,1\\}^{n}$ in the second item of (18) is a vector marking foreground", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29367", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2d3037d526dbec96c083a70d5cfaa3c38f5a7a8f5ba86833d86b4c558aeb5176.jpg", + "image_caption": [ + "(a) P2P scheme" + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 194, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5a89ef1fd0767b95f729c31ee59a7693f87f94e9c88ef75a51e6879443359794.jpg", + "image_caption": [ + "(b) matched point" + ], + "image_footnote": [], + "bbox": [ + 205, + 92, + 331, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e9ca855a9338a9e281301059c52e330acd103eea4f4505e3158c10f3895bced.jpg", + "image_caption": [ + "(c) semi-supervised points" + ], + "image_footnote": [], + "bbox": [ + 346, + 92, + 473, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d52126ae431bda47c3298a8fe18f63c4f40d36ec9102cd1bf43b1fdfb4460d0d.jpg", + "image_caption": [ + "(d) P2R scheme", + "Figure 4. Difference between P2P and P2R matching. (a) and (d) demonstrate an overall difference. P2P focuses only on foreground pixels in $\\mathbf{P}_s$ , while P2R considers background pixels as well. (b) and (e) show how the matching is performed in P2P and P2R, respectively. P2R segments out local regions for each pseudo-label, whereas P2P only detects one point. (c) and (f) illustrate how untrusted predictions are filtered. P2P retains only foreground pixels for loss computation, while P2R also keeps pixels in the neighborhoods of their corresponding pseudo points." + ], + "image_footnote": [], + "bbox": [ + 101, + 172, + 189, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/50797048b980f15b33ddf80a206c725012f898e4239eb929921c4e3236208965.jpg", + "image_caption": [ + "(e) matched region" + ], + "image_footnote": [], + "bbox": [ + 205, + 175, + 331, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b4774abf9ee95739c14124767740efce4801d5d91c06433a4fb03a77b5d4168b.jpg", + "image_caption": [ + "(f) semi-supervised regions" + ], + "image_footnote": [], + "bbox": [ + 346, + 174, + 473, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "regions, by setting those pixels in $\\mathcal{P}$ far from any points in $x^{\\prime}$ as zero:", + "bbox": [ + 89, + 411, + 482, + 440 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {[ i ]} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\min _ {j} l _ {2} (i, j) < \\mu \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 445, + 480, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in which $\\mu$ is a hyper parameter indicating the radius if the neighborhood of each GT point is considered as a circle. As displayed in Fig. 4(e), $\\mathbf{M}$ segments out a coarse local region that is consistent with the location of the GT point. After that, the learning objective $\\hat{\\pmb{p}}$ is derived by defining the cost matrix $\\mathbf{C} \\in \\mathbb{R}^{n \\times m}$ :", + "bbox": [ + 89, + 491, + 483, + 582 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {[ i j ]} = \\left\\{ \\begin{array}{l l} \\tau l _ {2} (i, j) - \\mathcal {S} (\\boldsymbol {p} _ {[ i ]}), & \\text {i f} \\mathbf {M} _ {[ i j ]} = 1 \\\\ \\infty , & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 585, + 482, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, the pixel with the minimum cost in each column of $\\mathbf{C}$ is marked as the potential foreground pixel:", + "bbox": [ + 89, + 632, + 482, + 662 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} _ {[ i j ]} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\mathbf {C} _ {[ i j ]} < \\mathbf {C} _ {[ k j ]} \\forall k \\neq i \\\\ 0, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 669, + 480, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, the learning objective $\\hat{\\pmb{p}}$ is derived as $\\hat{\\pmb{p}} = \\hat{\\mathbf{M}}\\mathbf{1}$ , and the loss $\\mathcal{L}_l$ is computed between $\\pmb{p}$ and $\\hat{\\pmb{p}}$ according to BCE loss formulated as (5) during training with labeled data.", + "bbox": [ + 89, + 718, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An extra benefit is that computing $\\mathbf{M}$ and $\\hat{\\mathbf{M}}$ in P2R requires only finding the minimum of each row or column in the matrices, which is significantly faster than executing the Hungarian algorithm — a time-consuming component of the P2P matching strategy [57].", + "bbox": [ + 89, + 763, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Training with Unlabeled Data via P2R", + "text_level": 1, + "bbox": [ + 89, + 847, + 419, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recalling the pseudo-label $\\mathcal{P}_t^\\prime = \\{\\pmb {p}_t^\\prime ,\\pmb{x}_t^\\prime \\}$ from the teacher and the supervised student output as $\\mathcal{P}_s = \\{\\pmb {p}_s,\\pmb {x}_s\\}$ , the", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "matching matrix $\\mathbf{M}_{st}$ between $\\mathcal{P}_s$ and $\\mathcal{P}_t^\\prime$ is computed via (18) in P2R, and then combined with the confidence in (7) to mask unreliable pseudo labels. Additionally, pixels far from any pseudo points can be directly marked as reliable background pixels, as suggested by observations from the aggregated PSAM in Fig. 3(e & f). Thus, the confidence diagonal matrix $\\mathbf{Z}$ is formulated as follows with $\\zeta = \\mathbb{1}(\\hat{\\pmb{p}}_t' > \\eta)$ :", + "bbox": [ + 511, + 90, + 903, + 198 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} = \\operatorname {d i a g} \\left[ \\mathbf {M} _ {s t} \\boldsymbol {\\zeta} + \\left(\\mathbf {1} _ {n} - \\boldsymbol {\\beta}\\right) \\right] \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 204, + 903, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final loss supervising unlabeled data $\\mathcal{P}_s$ with pseudo labels is derived by substituting $\\mathbf{Z}$ in (23) into (8).", + "bbox": [ + 511, + 228, + 903, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As displayed in Fig. 4(f), the confidence for each pseudo point is propagated to it corresponding region. On one hand, reliable foreground parts (blue) contain a pixel whose pseudo-label is 1, while the pseudo-labels of its surrounding pixels are marked as 0. On the other hand, all pixels in the reliable background parts (khaki) are assigned pseudolabels of 0 to train the counter with unlabeled data. The uncolored regions are ignored because their confidence is 0.", + "bbox": [ + 511, + 258, + 905, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pseudo code and flow chart of the proposed P2R loss can be found in the supplemental material.", + "bbox": [ + 511, + 378, + 903, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 511, + 421, + 645, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section explores how the proposed P2R training approach enhances a point-based counting model using both labeled and unlabeled data across three parts: semi-supervised crowd counting, unsupervised domain adaptation (UDA) for crowd counting, and ablation studies.", + "bbox": [ + 511, + 446, + 905, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the semi-supervised experiments, our method is implemented on four crowd counting datasets: ShTech A/B [73], UCF-QNRF [12], and JHU++ [55]. Following DAC [67] and OT-M [26], three protocols are applied: $5\\%$ , $10\\%$ , and $40\\%$ of labeled data, with the remaining crowd images involved in training without annotations. In the UDA part, labeled data from one domain, e.g., ShTech A, is used to initialize a counting model, while unlabeled data from another domain is utilized to capture the domain characteristics of the application scenes. In the ablation study, we test different hyperparameters to demonstrate their impact on the performance of P2R.", + "bbox": [ + 511, + 522, + 905, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Semi-Supervised Counting", + "text_level": 1, + "bbox": [ + 511, + 710, + 756, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 presents the experimental results of semi-supervised counting. Our method achieves the best performance in nearly all protocols across all crowd datasets. Besides, using $5\\%$ data with P2R surpasses the fully-supervised learning with $10\\%$ data and is even equivalent to other semi-supervised learning methods using $10\\%$ data. Similar performances are also observed between the semi-supervised training with $10\\%$ and $40\\%$ labeled data. These results demonstrate that P2R is label-efficient and advantageous in semi-supervised learning. In Fig. 5, we visualize predictions of DAC [67], OTM [26], and our P2R. The 3rd row is", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29368", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a11eb0bc84848c91c552d284bdb6c7bf1fefc58f892941bca0691e5b30fb7253.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Label Pct.MethodsShTech A [73]ShTech B [73]UCF-QNRF [12]JHU++ [55]
MAEMSEMAEMSEMAEMSEMAEMSE
5%label-only93.7155.213.124.0132.7231.7104.1383.5
MT [59]104.7156.919.333.2172.4284.9101.5363.5
L2R [32]103.0155.420.327.6160.1272.3101.4338.8
GP [54]102.0172.015.727.9160.0275.098.9355.7
DAC [23]85.4134.512.622.8120.2209.382.2294.9
OT-M [26]83.7133.312.621.5118.4195.482.7304.5
P2R (ours)69.9119.59.116.6100.1182.577.8293.5
10%label-only84.0138.310.719.2112.4186.878.7305.5
MT [59]319.394.515.624.5156.1245.5250.390.2
L2R [32]90.3153.515.624.4148.9249.887.5315.3
IRAST [34]86.9148.914.722.9135.6233.486.7303.4
DAC [23]74.9115.511.119.1109.0187.275.9282.3
OT-M [26]80.1118.510.818.2113.1186.773.0280.6
CU [18]70.7116.69.717.7104.01644.274.8281.6
P2R (ours)65.2114.68.414.594.9167.268.7272.3
40%label-only64.5105.68.114.099.2174.768.8283.5
MT [59]88.2151.115.925.7147.2249.6121.5388.9
L2R [32]86.5148.216.825.1145.1256.1123.6376.1
SUA [42]68.5121.914.120.6130.3226.380.7290.8
DAC [23]67.5110.79.614.691.1153.465.1260.0
OT-M [26]70.7114.58.113.1100.6167.672.1272.0
P2R (ours)55.695.06.811.086.0144.363.3271.1
100%P2PNet [57]52.7485.066.259.9085.32154.5061.25258.65
P2R (ours)51.0279.686.179.8483.26138.1158.83253.10
", + "bbox": [ + 94, + 88, + 480, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a failure case of P2R due to crowd motion blur. However, the overall performance of P2R is best.", + "bbox": [ + 89, + 398, + 482, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The last row of Tab. 1 also compares P2P and P2R under a fully-supervised learning scheme (100% labels); P2R performs better than P2PNet on all crowd counting datasets. Additionally, as described in Section 5.1, P2R is faster than P2P due to the absence of the Hungarian algorithm. We compared the efficiency using an image with a resolution of $576 \\times 960$ and 775 annotated points — P2P requires an average of 0.4307 seconds for loss computation, while P2R only needs 0.0064 seconds, which is nearly 68 times faster than P2P. These comparisons demonstrate that P2R surpasses P2P in both effectiveness and efficiency.", + "bbox": [ + 88, + 429, + 482, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Unsupervised Domain Adaptation (UDA)", + "text_level": 1, + "bbox": [ + 89, + 604, + 442, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UDA aims to transfer knowledge learned from a source domain to a target domain by designing modules or learning objectives that capture domain-agnostic features [11, 29, 43, 65, 70, 72]. Unlike semi-supervised learning, where labeled and unlabeled data originate from the same domain, UDA typically involves a domain gap between the labeled (source domain) and unlabeled crowd data (target domain). This gap often manifests as differences in crowd distribution, density levels, and perspectives. Our P2R can also be applied to UDA. When an unlabeled image from the target domain is provided, the threshold automatically partitions its corresponding pseudo points into two categories: higher-score $(> \\eta)$ and lower-score $(< \\eta)$ . The higher-score class is considered to be closer to the source domain, while the lower-score class is considered to be farther from the source domain. Thus, the higher-score pseudo points can participate in the training to gradually guide the model to learn to count in the target domain.", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d25312d39243a6b6634ea424acc16f50e830886b2550e10fa04726d36a4ba1ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 87, + 612, + 130 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e4bca1ce1a802d635973ce4b53dc29485be1ceabda551efa9d41edfb3489c10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 130, + 611, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/498debb671b3a4edee2dceac2f04c8a36334121c31d2a059ebc01fe31e84e8ff.jpg", + "image_caption": [ + "Figure 5. qualitative comparison with other models." + ], + "image_footnote": [], + "bbox": [ + 516, + 186, + 611, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ground truth: 61", + "bbox": [ + 532, + 237, + 591, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1254c483cece1cfc09f2ac61fdd5d89d1a529df00416c2721d340b5daa89f315.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 88, + 709, + 130 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94a12bc5c589ddd0a8f7af83e499663ce4398dcbea20e6710ee1212e237c7c05.jpg", + "image_caption": [ + "DAC [23]: 49.21" + ], + "image_footnote": [], + "bbox": [ + 614, + 130, + 709, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9650b28896edc859802b98431ddb2a87918b2f43ebe30de7e800abc36ac7d799.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 88, + 807, + 130 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/361c13ec83e61e4f11807d326fba288c08177559bbfdefc61ad9a6457f522c39.jpg", + "image_caption": [ + "OTM [26]: 51.43" + ], + "image_footnote": [], + "bbox": [ + 712, + 130, + 807, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ab0ea4ddece539f20bd8786803703c3bf75987758b349dfb9d6ae8c9b6f23e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 88, + 906, + 130 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1aa919225f6be1983959b821c70c72260f1e5fc7cd54f8da446701be6a0d5191.jpg", + "image_caption": [ + "P2R (ours): 60.0" + ], + "image_footnote": [], + "bbox": [ + 813, + 130, + 906, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 516, + 194, + 611, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ground truth: 61", + "bbox": [ + 532, + 237, + 591, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4299101a1c5146f624ae2d673bd7bb7b8f9db033834b5a9418e45b907948dc6d.jpg", + "image_caption": [ + "DAC [23]: 146.09" + ], + "image_footnote": [], + "bbox": [ + 614, + 194, + 709, + 236 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e81e68ee2a8eaf4974274886e9d870487d019eb84db173bb908cb81c627302f0.jpg", + "image_caption": [ + "OTM [26]: 126.11" + ], + "image_footnote": [], + "bbox": [ + 712, + 194, + 807, + 236 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8be1375a8f5baece94489667dadbd13ab3a9b4e5a45376da504be7396a81025a.jpg", + "image_caption": [ + "P2R (ours): 227.0" + ], + "image_footnote": [], + "bbox": [ + 813, + 194, + 906, + 236 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c668c9539047841ba65c01e5f2b0bb059a0f592f70d0e0a73a8f4689b169cb23.jpg", + "table_caption": [ + "Table 1. Comparison with other recent methods on four benchmark datasets under different labeled protocols." + ], + "table_footnote": [], + "table_body": "
METHODSA → BA → QB → AB → Q
MAEMSEMAEMSEMAEMSEMAEMSE
DGDCCUS [40]12.624.6119.4216.6121.8203.1179.1316.2
MPCount [43]11.419.7115.7199.899.6182.9165.6290.4
UDABL [38] (w/o DA)15.925.8166.7287.6138.1228.1226.4411.0
RBT [35]13.429.3175.0294.8112.2218.2211.3381.9
C2MoT [70]12.421.1125.7218.3120.7192.0198.9368.0
FGFD [79]12.723.3124.1242.0123.5210.7209.7384.7
FSIM [80]11.119.3105.3191.1120.3202.6194.9324.5
UDAP2R (w/o DA)25.635.7155.1286.8130.2229.1173.3329.8
P2R (w/ DA)10.618.7105.3194.889.3176.0139.5243.1
", + "bbox": [ + 527, + 282, + 893, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Unsupervised domain adaptation for crowd counting. The LHS and RHS of “ $\\rightarrow$ ” represent adaptation direction. Q, A, and B indicate UCF-QNRF [12], ShTech A, and ShTech B [73]. $DG$ indicates domain generalization. The training of UDA requires samples from the target domain, whereas DG does not.", + "bbox": [ + 511, + 407, + 906, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct experiments on four protocols following [70]. The results are presented in Tab. 2. The last two rows demonstrate the results of the proposed P2R. The poor performance obtained without domain adaptation highlights the domain gap between the source data and target data. However, when the unlabeled data from the target domain is involved in training, the estimation errors are significantly reduced, and the results surpass previous methods with complex network structures for domain adaptation. These advanced experimental results demonstrate that the threshold works as expected to capture samples close to the source domain and gradually teaches the model to count in the target domain.", + "bbox": [ + 511, + 491, + 906, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 696, + 666, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We next conduct ablation studies to investigate the influence of hyperparameters in P2R, including $\\tau$ in (21), the threshold $\\eta$ , $\\mu$ in (20), and $\\alpha$ in (17). The results are in Fig. 6", + "bbox": [ + 511, + 719, + 905, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The impact of $\\tau$ in (21). $\\tau$ is a hyperparameter to balance the distance and foreground score when selecting the foreground pixel from the matched region. When $\\tau \\rightarrow \\infty$ , the predicted score is ignored, which is equivalent to computing cross-entropy between the prediction and the ground truth point map. If $\\tau = 0$ , the pixel with the highest score is chosen as the foreground in $\\hat{\\pmb{p}}$ . However, Fig. 6(a) shows that neither extreme is optimal. The best performance is achieved when $\\tau = 8$ .", + "bbox": [ + 511, + 765, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29369", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d5fe63da2bf1fddb316df5f7d8e3c7e7678febfa5a7b3271468130cdbee91bd.jpg", + "image_caption": [ + "(a) errors vs. $\\tau$" + ], + "image_footnote": [], + "bbox": [ + 107, + 85, + 281, + 179 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9c7fe401fb6d177465c000264ef0308194a1cb1d7e70f6fe4536c9358ad99b4.jpg", + "image_caption": [ + "(c) errors vs. $\\mu$", + "Figure 6. The estimation errors vs. various hyperparameters. (a) is conducted on the fully-supervised ShTech A since it is irrelevant to semi-supervised counting, while (b)-(d) are conducted with $5\\%$ labeled data in ShTech A." + ], + "image_footnote": [], + "bbox": [ + 107, + 196, + 281, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1490810d57fd7ef78bcffb49f22d8a9721c65052b843f722653ce7a89f89bfc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 85, + 467, + 179 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0072ff971013eced25b31451a39f06e40e677af0efbb691d4afd072e6a0db34a.jpg", + "image_caption": [ + "(b) errors vs. $\\eta$", + "(d) errors vs. $\\frac{\\alpha}{1 - \\alpha}$" + ], + "image_footnote": [], + "bbox": [ + 292, + 196, + 467, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of threshold $\\eta$ . $\\eta$ defines the threshold to distinguish reliable and unreliable regions. Since all pseudo points are selected by picking pixels whose score is greater than 0.5 in the prediction, the threshold should also be greater than 0.5. Moreover, a lower $\\eta$ introduces unreliable pseudo-labels into training, while a higher $\\eta$ cannot fully utilize the unlabeled data as most parts cannot participate in training. As shown in Fig. 6(b), the estimation errors exhibit a trend of first decreasing and then increasing within the range of 0.6 to 0.9, with the lowest errors occurring when the threshold is set to 0.7.", + "bbox": [ + 88, + 383, + 482, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The radius of each point $\\mu$ . The vector $\\beta$ in (20) derived from $\\mu$ marks the foreground points' neighborhood and background regions. It not only restricts the positions where foreground pixels appear by substituting $\\beta$ into (18), but also directly identifies reliable background parts via the second item in (23) based on the observation of PSAM in Section 4.2. Fig. 6(c) presents how $\\mu$ affects the counting performance. It shows the MAE remains around 68 when $\\mu$ increases from 32 to 256, indicating that a larger $\\mu$ does not significantly affect the results. This is consistent with the observation in PSAM since pixels far from the foreground points are less likely to be activated during training.", + "bbox": [ + 88, + 551, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The weight of unlabeled loss. Fig. 6(d) presents how $\\alpha$ in (17) affects the performance. A large $\\alpha$ focuses more on unlabeled data, leading to training failure, while a small $\\alpha$ may cause the model to overfit on labeled data. In the experiments with ShTech A and $5\\%$ labeled data, $\\frac{\\alpha}{1 - \\alpha} = 2$ , i.e., $\\alpha = 2/3$ , results in the best performance. Additionally, as long as $\\alpha$ is not 0, the point-based counter can benefit from unlabeled data, as the maximum MAE/MSE in Fig. 6(d) are smaller than 80/150, which is much better than the model trained with only $5\\%$ labeled data (MAE: 93.7, MSE: 155.2).", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bc9d5517a41f3760c947c215bfb6cc4dc5fb4769bbcd349a78272a90d6a583e2.jpg", + "image_caption": [ + "(a) sorted average PSAM" + ], + "image_footnote": [], + "bbox": [ + 526, + 94, + 727, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18c3907376122314226fbf8f929c8556edd5c45d55abd39e848e55802195a990.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 94, + 880, + 133 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0171d89ef65055b33551da679ddf963d531fbdc140881281682d438984b1449f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 135, + 880, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b4bec7e249dfc042fb1836bbfc195dd36705f4f2e00267923a0339ba8e0217ba.jpg", + "image_caption": [ + "(b) Local & aggregated PSAM", + "Figure 7. The comparison of PSAMs among different training schemes. Model-L (green) is trained with only labeled data; model-U_P2P (blue) is trained with both labeled data and unlabeled data using the P2P scheme, while model-U_P2R (orange) is trained with the proposed P2R scheme." + ], + "image_footnote": [], + "bbox": [ + 745, + 175, + 880, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PSAM comparison. Fig. 7 visualizes the PSAMs of models trained with the P2P and P2R schemes. Comparing model-L (green) and model-U_P2R (blue), the latter effectively restricts the PSAM to the neighborhood of the concerned foreground pixel, and its aggregated PSAM is also sparser than that of model-L, demonstrating the effectiveness of P2R. Comparing model-U_P2P and model-U_P2R in Fig. 7(a), the values of the PSAM are similar at the non-peak indexes, but the latter has a higher activation at the peak position compared to the former, indicating that semi-supervised learning further enhances the identification ability of the point-based counter.", + "bbox": [ + 511, + 321, + 906, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 517, + 633, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a P2R scheme to train a point-based counter for semi-supervised crowd counting. At the beginning, we establish a pseudo-labeling framework based on P2P matching, but its formulation is ill-posed since the confidence is only applicable to foreground pixels and cannot be propagated to backgrounds. To observe how the foreground and background feature are learned, we propose PSAM to visually interpret the concerned region and activation value of all foreground pixels. From the visualized activation map, it is observed that the neighborhood of each foreground pixel is over-activated and falsely recognized as instances by the decoder. Based on this observation, we replace P2P with P2R by segmenting the prediction into multiple regions containing the corresponding instances. Thus, pixels in the neighborhood of the concerned pseudo points can share the weights with the corresponding pixels. Additionally, the time-consuming Hungarian algorithm is no longer necessary in P2R. Several outstanding experimental results in semi-supervised counting and unsupervised domain adaptation demonstrate the advantages of our P2R matching strategy.", + "bbox": [ + 511, + 542, + 906, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by a grant from City University of Hong Kong (Project No. 7030010).", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29370", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shahira Abousamra, Minh Hoai, Dimitris Samaras, and Chao Chen. Localization in the crowd with topological constraints. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 872-881, 2021. 2", + "[2] Areej Althali, Amal Balabid, Reem Alharthi, Bander Alzahrani, Reem Alotaibi, and Ahmed Barnawi. Anomalous event detection and localization in dense crowd scenes. Multimedia Tools and Applications, 82(10):15673-15694, 2023. 1", + "[3] Deepak Babu Sam, Shiv Surya, and R Venkatesh Babu. Switching convolutional neural network for crowd counting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5744-5752, 2017. 2", + "[4] Antoni B Chan and Nuno Vasconcelos. Bayesianoisson regression for crowd counting. In 2009 IEEE 12th international conference on computer vision, pages 545-551. IEEE, 2009. 1, 2", + "[5] Antoni B Chan and Nuno Vasconcelos. Counting people with low-level features and bayesian regression. IEEE Transactions on image processing, 21(4):2160-2177, 2011. 1", + "[6] Antoni B Chan, Zhang-Sheng John Liang, and Nuno Vasconcelos. Privacy preserving crowd monitoring: Counting people without people models or tracking. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-7. IEEE, 2008. 2", + "[7] Mulin Chen, Qi Wang, and Xuelong Li. Anchor-based group detection in crowd scenes. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1378-1382. IEEE, 2017. 1", + "[8] Zhi-Qi Cheng, Qi Dai, Hong Li, Jingkuan Song, Xiao Wu, and Alexander G Hauptmann. Rethinking spatial invariance of convolutional networks for object counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19638-19648, 2022. 2", + "[9] Junyu Gao, Tao Han, Qi Wang, Yuan Yuan, and Xuelong Li. Learning independent instance maps for crowd localization. arXiv preprint arXiv:2012.04164, 2020. 2", + "[10] Tao Han, Lei Bai, Lingbo Liu, and Wanli Ouyang. Steerer: Resolving scale variations for counting and localization via selective inheritance learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21848-21859, 2023. 1, 3", + "[11] Yuhang He, Zhiheng Ma, Xing Wei, Xiaopeng Hong, Wei Ke, and Yihong Gong. Error-aware density isomorphism reconstruction for unsupervised cross-domain crowd counting. In Proceedings of the AAAI conference on artificial intelligence, pages 1540-1548, 2021. 7", + "[12] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 6, 7", + "[13] Brian Kenji Iwana, Ryohei Kuroki, and Seiichi Uchida. Explaining convolutional neural networks using softmax gradient layer-wise relevance propagation. In 2019 IEEE/CVF" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision Workshop (ICCVW), pages 4176-4185. IEEE, 2019. 3", + "[14] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 2", + "[15] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2, 3, 5", + "[16] Bastian Leibe, Edgar Seemann, and Bernt Schiele. Pedestrian detection in crowded scenes. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 878-885. IEEE, 2005. 2", + "[17] Victor Lempitsky and Andrew Zisserman. Learning to count objects in images. Advances in neural information processing systems, 23, 2010. 2", + "[18] Chen Li, Xiaoling Hu, Shahira Abousamra, and Chao Chen. Calibrating uncertainty for semi-supervised crowd counting. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16685-16695. IEEE, 2023. 2, 7", + "[19] Hanxiao Li, Yonghong Song, and Tong Geng. Semi-supervised crowd counting based on hard pseudo-labels. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2024. 2", + "[20] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1091-1100, 2018. 2", + "[21] Dingkang Liang, Wei Xu, and Xiang Bai. An end-to-end transformer model for crowd localization. In ECCV, pages 38-54. Springer, 2022. 1, 2", + "[22] Hui Lin, Xiaopeng Hong, Zhiheng Ma, Xing Wei, Yunfeng Qiu, Yaowei Wang, and Yihong Gong. Direct measure matching for crowd counting. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages = 837–844, publisher = ijcai.org, year = 2021, .2", + "[23] Hui Lin, Zhiheng Ma, Xiaopeng Hong, Yaowei Wang, and Zhou Su. Semi-supervised crowd counting via density agency. In ACM Multimedia, 2022. 2, 7", + "[24] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19628-19637, 2022. 1, 2", + "[25] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, Zhou Su, Xiaopeng Hong, and Deyu Meng. Semi-supervised counting via pixel-by-pixel density distribution modelling. arXiv preprint arXiv:2402.15297, 2024. 2", + "[26] Wei Lin and Antoni B Chan. Optimal transport minimization: Crowd localization on density maps for semi-supervised counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21663-21673, 2023. 1, 2, 3, 6, 7", + "[27] Wei Lin and Antoni B Chan. A fixed-point approach to unified prompt-based counting. In Proceedings of the" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29371", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "AAAI Conference on Artificial Intelligence, pages 3468-3476, 2024. 2", + "[28] Wei Lin, Jia Wan, and Antoni B Chan. Proximal mapping loss: Understanding loss functions in crowd counting & localization. In The Thirteenth International Conference on Learning Representations, 2", + "[29] Wei Lin, Junyu Gao, Qi Wang, and Xuelong Li. Learning to detect anomaly events in crowd scenes from synthetic data. Neurocomputing, 436:248-259, 2021. 1, 7", + "[30] Wei Lin, Kunlin Yang, Xinzhu Ma, Junyu Gao, Lingbo Liu, Shinan Liu, Jun Hou, Shuai Yi, and Antoni B Chan. Scale-prior deformable convolution for exemplar-guided class-agnostic counting. In BMVC, page 313, 2022. 2", + "[31] Chengxin Liu, Hao Lu, Zhiguo Cao, and Tongliang Liu. Point-query quadtree for crowd counting, localization, and more. In CVPR, pages 1676-1685, 2023. 1, 2", + "[32] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Leveraging unlabeled data for crowd counting by learning to rank. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7661-7669, 2018. 2, 7", + "[33] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Exploiting unlabeled data in cnns by self-supervised learning to rank. IEEE transactions on pattern analysis and machine intelligence, 41(8):1862-1878, 2019. 2", + "[34] Yan Liu, Lingqiao Liu, Peng Wang, Pingping Zhang, and Yinjie Lei. Semi-supervised crowd counting via self-training on surrogate tasks. In European Conference on Computer Vision, pages 242-259. Springer, 2020. 2, 7", + "[35] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, pages 129-137, 2020. 7", + "[36] Yanbo Liu, Yingxiang Hu, Guo Cao, and Yanfeng Shang. Semi-supervised crowd counting via multi-task pseudo-label self-correction strategy. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2", + "[37] Zheng Ma and Antoni B Chan. Counting people crossing a line using integer programming and local features. IEEE Transactions on Circuits and Systems for Video Technology, 26(10):1955-1969, 2015. 1", + "[38] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In Proceedings of the IEEE International Conference on Computer Vision, pages 6142-6151, 2019. 1, 2, 7, 13", + "[39] Zhiheng Ma, Xing Wei, Xiaopeng Hong, Hui Lin, Yunfeng Qiu, and Yihong Gong. Learning to count via unbalanced optimal transport. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2319-2327, 2021. 2", + "[40] Lucas Mansilla, Rodrigo Echeveste, Diego H Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6630-6638, 2021. 7", + "[41] Riccardo Mazzon, Fabio Poiesi, and Andrea Cavallaro. Detection and tracking of groups in crowd. In 2013 10th IEEE International conference on advanced video and signal based surveillance, pages 202-207. IEEE, 2013. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15549-15559, 2021. 2, 3, 7", + "[43] Zhuoxuan Peng and S-H Gary Chan. Single domain generalization for crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28025-28034, 2024. 7", + "[44] Harish Guruprasad Ramaswamy et al. Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 983-991, 2020. 3", + "[45] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3394-3403, 2021. 2", + "[46] Weihong Ren, Xinchao Wang, Jiandong Tian, Yandong Tang, and Antoni B Chan. Tracking-by-counting: Using network flows on crowd density maps for tracking multiple targets. IEEE Transactions on Image Processing, 30:1439-1452, 2020. 1", + "[47] Jihye Ryu and Kwangho Song. Crowd counting and individual localization using pseudo square label. IEEE Access, 2024. 2", + "[48] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 3, 4", + "[49] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 323-331, 2024. 2", + "[50] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19618-19627, 2022. 2", + "[51] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19618-19627, 2022. 2", + "[52] Weibo Shu, Jia Wan, and Antoni B Chan. Generalized characteristic function loss for crowd analysis in the frequency domain. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[53] Karen Simonyan. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 4", + "[54] Vishwanath A Sindagi, Rajeev Yasarla, Deepak Sam Babu, R Venkatesh Babu, and Vishal M Patel. Learning to count in the crowd from limited labeled data. In European Conference on Computer Vision, pages 212-229. Springer, 2020. 2, 7", + "[55] Vishwanath A Sindagi, Rajeev Yasarla, and Vishal M Patel. Jhu-crowd++: Large-scale crowd counting dataset and a" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29372", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "benchmark method. IEEE transactions on pattern analysis and machine intelligence, 44(5):2594-2609, 2020. 6, 7", + "[56] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 2, 3", + "[57] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In CVPR, pages 3365-3374, 2021. 1, 2, 3, 6, 7, 13", + "[58] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin A. Riedmiller. Striving for simplicity: The all convolutional net. In 3rd International Conference on Learning Representations, ICLR 2015, Workshop Track Proceedings, 2015. 4", + "[59] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 3, 7", + "[60] Jia Wan, Qingzhong Wang, and Antoni B Chan. Kernel-based density map generation for dense object counting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1357-1370, 2020. 2", + "[61] Jia Wan, Ziquan Liu, and Antoni B Chan. A generalized loss function for crowd counting and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1974-1983, 2021. 1, 2, 3, 13", + "[62] Jia Wan, Qiangqiang Wu, Wei Lin, and Antoni Chan. Robust zero-shot crowd counting and localization with adaptive resolution sam. In European Conference on Computer Vision, pages 478-495. Springer, 2024. 1", + "[63] Boyu Wang, Huidong Liu, Dimitris Samaras, and Minh Hoai Nguyen. Distribution matching for crowd counting. Advances in neural information processing systems, 33:1595-1607, 2020. 2, 13", + "[64] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 3", + "[65] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8198-8207, 2019. 2, 7", + "[66] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. Nwpu-crowd: A large-scale benchmark for crowd counting and localization. IEEE transactions on pattern analysis and machine intelligence, 43(6):2141-2149, 2020. 1", + "[67] Qi Wang, Wei Lin, Junyu Gao, and Xuelong Li. Density-aware curriculum learning for crowd counting. IEEE Transactions on Cybernetics, 52(6):4675-4687, 2020. 2, 6", + "[68] Qi Wang, Wei Lin, Junyu Gao, and Xuelong Li. Density-aware curriculum learning for crowd counting. IEEE Transactions on Cybernetics, 52(6):4675-4687, 2020. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Pixel-wise crowd understanding via synthetic data. International Journal of Computer Vision, 129(1):225-245, 2021. 2", + "[70] Qiangqiang Wu, Jia Wan, and Antoni B Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 658-666, 2021. 7", + "[71] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15548-15557, 2023. 2", + "[72] Qi Zhang, Wei Lin, and Antoni B Chan. Cross-view cross-scene multi-view crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 557-567, 2021. 7", + "[73] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 589-597, 2016. 1, 2, 6, 7, 13", + "[74] Chenyang Zhao and Antoni B. Chan. ODAM: gradient-based instance-specific visual explanations for object detection. In ICLR, 2023. 2, 3, 4", + "[75] Chunhui Zhao, Zhiyuan Zhang, Jinwen Hu, Dong Wang, Bin Fan, Quan Pan, and Qiang He. Crowd anomaly event detection in surveillance video based on the evolution of the spatial position relationship feature. In 2018 IEEE 14th International Conference on Control and Automation (ICCA), pages 247-252, 2018. 1", + "[76] Chenyang Zhao, Janet H Hsiao, and Antoni B Chan. Gradient-based instance-specific visual explanations for object specification and object discrimination. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 3", + "[77] Chenyang Zhao, Kun Wang, Xingyu Zeng, Rui Zhao, and Antoni B Chan. Gradient-based visual explanation for transformer-based clip. In International Conference on Machine Learning, pages 61072-61091. PMLR, 2024.", + "[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 3, 4", + "[79] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5659-5668, 2022. 7", + "[80] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 2023. 7" + ], + "bbox": [ + 516, + 90, + 906, + 825 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29373", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/7bc00c8f-f0e1-486b-af9f-cab0b77e9a84_model.json b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/7bc00c8f-f0e1-486b-af9f-cab0b77e9a84_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1083d567f9842bac3af026a33e243c933aba0b2c --- /dev/null +++ b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/7bc00c8f-f0e1-486b-af9f-cab0b77e9a84_model.json @@ -0,0 +1,3478 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.131, + 0.861, + 0.153 + ], + "angle": 0, + "content": "Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.182, + 0.69, + 0.199 + ], + "angle": 0, + "content": "Wei Lin, Chenyang Zhao, and Antoni B. Chan" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.2, + 0.756, + 0.218 + ], + "angle": 0, + "content": "Department of Computer Science, City University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.22, + 0.806, + 0.233 + ], + "angle": 0, + "content": "elonlin24@gmail.com, chenyzhao9-c@my.cityu.edu.hk, abchan@cityu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.755 + ], + "angle": 0, + "content": "Point detection has been developed to locate pedestrians in crowded scenes by training a counter through a point-to-point (P2P) supervision scheme. Despite its excellent localization and counting performance, training a point-based counter still faces challenges concerning annotation labor: hundreds to thousands of points are required to annotate a single sample capturing a dense crowd. In this paper, we integrate point-based methods into a semi-supervised counting framework based on pseudo-labeling, enabling the training of a counter with only a few annotated samples supplemented by a large volume of pseudolabeled data. However, during implementation, the training encounters issues as the confidence for pseudo-labels fails to be propagated to background pixels via the P2P. To tackle this challenge, we devise a point-specific activation map (PSAM) to visually interpret the phenomena occurring during the ill-posed training. Observations from the PSAM suggest that the feature map is excessively activated by the loss for unlabeled data, causing the decoder to misinterpret these over-activations as pedestrians. To mitigate this issue, we propose a point-to-region (P2R) scheme to substitute P2P, which segments out local regions rather than detects a point corresponding to a pedestrian for supervision. Consequently, pixels in the local region can share the same confidence with the corresponding pseudo points. Experimental results in both semi-supervised counting and unsupervised domain adaptation highlight the advantages of our method, illustrating P2R can resolve issues identified in PSAM. The code is available at https://github.com/Elin24/P2RLoss." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.769, + 0.222, + 0.784 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Crowd scene analysis has been investigated for many years due to its significant applications in smart cities and urban safety [2, 7, 29, 41, 46, 62, 75]. Crowd counting and localization in extremely dense crowd scenes, as the fundamental task in this field, have attracted considerable attention from researchers in computer vision [4, 5, 37, 61, 66, 68]. Although most supervised crowd counting methods fol" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.269, + 0.621, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.277, + 0.687, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.275, + 0.776, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.287, + 0.807, + 0.296 + ], + "angle": 0, + "content": ">0.5" + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.277, + 0.852, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.321, + 0.864, + 0.329 + ], + "angle": 0, + "content": "Sudo Label" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.325, + 0.621, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.321, + 0.684, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.351, + 0.688, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.391, + 0.676, + 0.398 + ], + "angle": 0, + "content": "Student" + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.321, + 0.758, + 0.329 + ], + "angle": 0, + "content": "Prediction" + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.339, + 0.865, + 0.35 + ], + "angle": 0, + "content": "w/o gradient" + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.359, + 0.824, + 0.375 + ], + "angle": 0, + "content": "P2R Matching" + }, + { + "type": "image_caption", + "bbox": [ + 0.839, + 0.359, + 0.889, + 0.375 + ], + "angle": 0, + "content": "Weighted Cross Entropy" + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.384, + 0.865, + 0.395 + ], + "angle": 0, + "content": "w/ gradient" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.418, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Figure 1. The workflow of semi-supervised point-based counting methods. The teacher model generates pseudo labels by extracting the foreground pixels, while the student model takes the corresponding strongly augmented image as input to construct the computation graph. The training loss between the pseudo label and the student's prediction involves two steps: the proposed P2R matching and the weighted cross-entropy computation." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.533, + 0.907, + 0.716 + ], + "angle": 0, + "content": "low the scheme of density regression [10, 24, 38, 61, 73], there are also relevant studies implementing point detection for crowd localization and counting [21, 31, 57], due to its straightforwardness in indicating pedestrians' positions. These approaches use the Hungarian algorithm to perform point-to-point (P2P) matching between predicted point maps and ground-truth (GT) pedestrian annotations during training. Therefore, these matched pixels are considered as pedestrians' positions, while the others are designated as background for loss computation. This scheme successfully trains a point detector for crowd analysis, yielding both good localization and counting performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Great progress has been made in supervised counting, but the labor-intensive nature of annotating data has hindered the development of crowd understanding for a long time [26, 66]. Addressing this issue, semi-supervised counting is gradually being developed, aiming to train a model using large amounts of unlabeled data alongside only a few annotated samples. Similar to supervised counting, most semi-supervised methods also follow the procedure of density regression. Notably, the method proposed in OT-M [26] stands out by employing a simple self-training scheme based on mean-teacher [59] and FixMatch [56]. However, there is no study exploring semi-supervised counting and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29363" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "localization using a point-detection method. To fill this gap, this paper explores an approach to train a point detector with limited annotated data." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.139, + 0.483, + 0.426 + ], + "angle": 0, + "content": "As shown in Fig. 1, typical semi-supervised counting methods rely on pseudo-labeling, which generates pseudo labels to serve as learning targets for unlabeled samples using a model pre-trained on a small amount of annotated data. During training, there is a confidence score to determine whether a concerned prediction should be retained as a pseudo label. However, if the P2P matching strategy is applied to compute loss within the pipeline presented in Fig. 1, there is no reasonable way to define the confidence for unmatched background pixels, as P2P matching only cares about matched foreground pixels. Specifically, a foreground pixel can be retained for training if its matched pseudo-label's foreground probability exceeds a predefined threshold, similar to the approach used in semi-supervised classification tasks [15, 56]. However, background pixels lack confidence scores since no pseudo-labels are assigned to them after matching. Experimental results also demonstrate that ignoring the background leads to the breakdown of semi-supervised training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.483, + 0.548 + ], + "angle": 0, + "content": "To address the aforementioned issue, we propose a gradient-based visual explanation method (akin to [48, 74, 76, 78]), termed point-specific activation map (PSAM), to visualize regions of interest and the activation levels of detected points. Using PSAM, we observe that the activation values and coverage areas of foreground pixels continuously grow during semi-supervised training, causing the model to misclassify neighboring pixels as foreground." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.55, + 0.483, + 0.821 + ], + "angle": 0, + "content": "Inspecting PSAMs reveals that each point in GT is implicitly associated with a local region rather than a single pixel in the prediction. Accordingly, we propose to replace the P2P with a point-to-region (P2R) strategy, which matches local regions to points in the pseudo-label. This allows all pixels within a matched region, including background pixels, to share the confidence of the corresponding pseudo-point. Experimental results and visualizations demonstrate that P2R can effectively train a semisupervised point-based counter and achieve expected performance. Experimental results show that P2R effectively trains semi-supervised point-based counters with comparable performance to P2P but at significantly reduced computational cost. Unlike P2P [57], P2R eliminates the need for the Hungarian algorithm by assigning each ground-truth or pseudo-point a local \"dominated zone\" and selecting a representative foreground pixel based on a predefined cost matrix." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.824, + 0.414, + 0.838 + ], + "angle": 0, + "content": "The contributions of our paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.84, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- We formulate a pipeline of semi-supervised counting with a point-detection method, but find a problem leading the training breakdown." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- We propose Point-Specific Activation Map (PSAM) to" + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "visualize the regions of interest and the activation values of each point detected by the point-based crowd counter. Through PSAM, we observe that the surrounding pixels of each expected foreground pixel are over-activated, causing the decoder to recognizes these pixels as instances different from the concerned one." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.183, + 0.905, + 0.272 + ], + "angle": 0, + "content": "- Based on insights from PSAM, we propose replacing the P2P with a P2R matching strategy for semi-supervised learning in point-based crowd counting. This not only makes the training process work as expected, but also brings additional benefits of reducing computational requirements." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.273, + 0.905, + 0.348 + ], + "angle": 0, + "content": "- The proposed P2R achieves outstanding performance in several experiments, including semi-supervised counting and unsupervised domain adaptation. Ablation study shows that P2R excellently tackles the problem caused by P2P in semi-supervised counting." + }, + { + "type": "list", + "bbox": [ + 0.533, + 0.183, + 0.905, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.664, + 0.377 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.388, + 0.907, + 0.719 + ], + "angle": 0, + "content": "Object counting in dense scenes. The development of deep learning has significantly influenced crowd counting. Prior to this, detecting body parts [16, 17] was the common pipeline for crowd understanding in computer vision, although some methods implemented counting models without explicit object segmentation or tracking information [4-6]. Beyond specific crowd scenarios, class-agnostic counting [14, 27, 30, 45, 49, 71] has also been developed to estimate the distribution of objects in the same category. Most recent counting methods implement neural networks following the density map regression scheme [3, 8, 20, 50, 60, 65, 67, 69, 73]. In addition to continuous innovations in model architecture design [3, 20, 24], new loss functions have also been developed, including Bayesian Loss [38], loss functions based on balanced [63] and unbalanced optimal transport [22, 39, 61], and characteristic function loss computed in the frequency domain [51, 52]. A study analyzes the relationship between current loss functions from the perspective of proximal mapping [28]. Some methods also perform crowd analysis via instance segmentation [1, 9] and point detection [21, 31, 47, 57]. The latter is the scheme adopted in our semi-supervised model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Semi-supervised counting. Semi-supervised counting has attracted broad attention recently [18, 19, 23, 25, 26, 32-34, 36, 42, 54]. L2R [32, 33] introduces a ranking rule for unlabeled crowd images based on the principle that cropped images should have smaller counts than full images. GP [54] iteratively updates pseudo-labels using Gaussian Processes. IRAST [34] and SUA [42] enhance models by incorporating a segmentation branch. DAC [23] uses learnable density agents to capture varying crowd densities. OT-M [26] transforms density maps into point maps as pseudo-labels, achieving strong performance with a density-to-point loss function [61]. Our method also em" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29364" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "ploys pseudo-labeling but uses a point detector for counting, avoiding the need for density map transformations as in [26]. Additionally, uncertainty is directly measured by the scores of extracted points, without any additional learnable structures like those in [42]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.168, + 0.485, + 0.456 + ], + "angle": 0, + "content": "Visual Explanation. Visualizing the importance of extracted image features is a straightforward way to interpret a model. The mainstream of visual explanation focuses on methods producing activation maps [13, 44, 48, 64, 74, 76-78], which are the products of feature maps from some intermediate layer and weight maps indicating the corresponding feature's importance. For example, GradCAM [13, 48] defines the weights using the corresponding gradient maps, generated via backpropagation. Beyond the focus on classification tasks, ODAM [74, 76] extends gradient-based CAM to object detection, visualizing the local impact of features on the detector's decisions. In this paper, we propose PSAM to visualize how the activation map changes for each predicted point in an ill-posed semi-supervised counting framework. Based on the observations from PSAM, we design a P2R matching strategy to avoid the over-activation of pixels in local regions, which effectively resolves the drawbacks of P2P and enables the successful training of a semi-supervised point detector." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.484, + 0.487 + ], + "angle": 0, + "content": "3. Point Counter in Semi-Supervised Counting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.556 + ], + "angle": 0, + "content": "This section formulates a simplified point-based counter by removing the regression branch in P2PNet [57], and then presents the training scheme and challenges when training with labeled and unlabeled data, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.567, + 0.457, + 0.582 + ], + "angle": 0, + "content": "3.1. Preliminary on A Simplified Point Counter" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.741 + ], + "angle": 0, + "content": "Common point-based crowd counters like P2PNet [57] contain a classification and a regression branch. However, we find a simplification can be achieved by removing the offset regression branch. On one hand, the crowd count can be estimated by counting foreground pixels with scores greater than 0.5, a process that does not involve the regression branch. On the other hand, pixel coordinates alone are sufficient to locate pedestrians, as demonstrated in STEERER [10] and GL [61], which use the coordinates of local maxima to indicate pedestrians locations." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.742, + 0.381, + 0.756 + ], + "angle": 0, + "content": "As a result, P2PNet can be simplified to:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.768, + 0.483, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {F} (I, \\Theta_ {f}) \\rightarrow F \\in \\mathbb {R} ^ {c \\times h \\times w}, (1) \\\\ \\mathcal {D} (F, \\Theta_ {d}) \\rightarrow D \\in \\mathbb {R} ^ {h \\times w} (2) \\\\ \\rightarrow \\mathcal {P} = \\left\\{\\boldsymbol {p} \\in \\mathbb {R} ^ {n}, \\boldsymbol {x} \\in \\mathbb {R} ^ {n \\times 2} \\right\\}, (3) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\) is the image encoder with learnable parameters \\(\\Theta_f\\), and \\(\\mathcal{D}\\) denotes the density decoder with learnable parameter \\(\\Theta_d\\). Notably, the offset regression branch is excluded in this simplified formulation. The pixel set \\(\\pmb{p} \\in \\mathbb{R}^n\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "represents the flattened values in \\(D\\), and its corresponding coordinate set is denoted as \\(\\pmb{x} \\in \\mathbb{R}^{n \\times 2}\\), where \\(n = h \\times w\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.13, + 0.905, + 0.147 + ], + "angle": 0, + "content": "3.2. Training with Labeled Data via P2P Matching" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.214 + ], + "angle": 0, + "content": "The GT with \\( m \\) points can be denoted as \\( \\pmb{x}^{\\prime} \\in \\mathbb{R}^{m \\times 2} \\), where \\( \\pmb{x}_{[j]}^{\\prime} \\in \\mathbb{R}^{2} \\) represents the location of the \\( j \\)-th pedestrian. Using \\( \\mathcal{P} \\) and \\( \\pmb{x}^{\\prime} \\), the loss computation involves two parts: P2P matching and binary cross entropy (BCE) calculation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.215, + 0.905, + 0.259 + ], + "angle": 0, + "content": "The former aims to obtain a matrix-based solution \\(\\mathbf{M} \\in \\{0,1\\}^{n \\times m}\\) that minimizes the bipartite-graph matching cost between \\(\\pmb{x}\\) and \\(\\pmb{x}'\\) within the cost matrix \\(\\mathbf{C} \\in \\mathbb{R}^{n \\times m}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.269, + 0.905, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {[ i j ]} = \\tau \\| \\boldsymbol {x} _ {[ i ]} - \\boldsymbol {x} ^ {\\prime} _ {[ j ]} \\| _ {2} - S (\\boldsymbol {p} _ {[ i ]}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.295, + 0.905, + 0.34 + ], + "angle": 0, + "content": "where \\( S(\\cdot) \\) is defined as the inverse sigmoid function \\( S(p) = -\\log (1 / p - 1) \\), rather than the identity operator used in the vanilla P2PNet [57], for improved performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.905, + 0.385 + ], + "angle": 0, + "content": "To compute BCE loss, the learning objective of \\(\\pmb{p}\\) is formulated as \\(\\hat{\\pmb{p}} = \\mathbf{M}\\mathbf{1}_m\\), and the BCE loss is calculated for optimization as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.393, + 0.905, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l} = - \\lambda \\hat {\\boldsymbol {p}} ^ {\\top} \\log \\boldsymbol {p} - \\left(\\mathbf {1} _ {n} - \\hat {\\boldsymbol {p}}\\right) ^ {\\top} \\log \\left(\\mathbf {1} _ {n} - \\boldsymbol {p}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.421, + 0.887, + 0.436 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a re-weighting factor for matched pixels [57]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.444, + 0.88, + 0.46 + ], + "angle": 0, + "content": "3.3. Problem in Training with Unlabeled Image" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.649 + ], + "angle": 0, + "content": "Here we consider training with unlabeled data following the standard pseudo-labeling procedure [15, 56]. As illustrated in Fig. 1, it requires the cooperation of a student and teacher model that share the same structure described in (1)~(3). The difference lies in that the teacher model is updated via exponential moving average (EMA) [59], while the student is optimized via gradient descending. During training, the teacher and student process the weakly and strongly data-augmented versions of an unlabeled image, producing \\(\\mathcal{P}_t = \\{\\pmb{p}_t, \\pmb{x}_t\\}\\) and \\(\\mathcal{P}_s = \\{\\pmb{p}_s, \\pmb{x}_s\\}\\), respectively. After that, foreground points are extracted from \\(\\mathcal{P}_t\\) to construct pseudo-labels:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.657, + 0.905, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {t} ^ {\\prime} = \\left\\{\\boldsymbol {p} _ {t} ^ {\\prime}, \\boldsymbol {x} _ {t} ^ {\\prime} \\right\\} = \\left\\{\\boldsymbol {p} _ {t [ j ]}, \\boldsymbol {x} _ {t [ j ]} \\mid \\boldsymbol {p} _ {t [ j ]} > 0. 5 \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.905, + 0.731 + ], + "angle": 0, + "content": "where \\( \\pmb{x}_t^\\prime \\) is treated as the pseudo hard label to estimate the matching matrix \\( \\mathbf{M}_{st} \\) via Hungarian algorithm, and the loss is obtained by substituting \\( \\hat{\\pmb{p}}_t = \\mathbf{M}_{st}\\mathbf{1}_m \\) and \\( \\pmb{p}_s \\) into (5)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.905, + 0.791 + ], + "angle": 0, + "content": "However, some pseudo-labels may not be reliable. A common solution is to only consider points in \\(\\mathcal{P}_t^\\prime\\) with high scores [15, 56], i.e., using a confidence vector \\(\\zeta\\) to mask out elements whose score is below a threshold \\(\\eta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.8, + 0.905, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\zeta = \\mathbb {1} \\left(\\boldsymbol {p} _ {t} ^ {\\prime} > \\eta\\right) \\xrightarrow {\\text {m a p t o} \\hat {\\boldsymbol {p}} _ {t} \\text {b y} \\mathbf {M} _ {s t}} \\boldsymbol {z} = \\mathbf {M} _ {s t} \\zeta . \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.905, + 0.876 + ], + "angle": 0, + "content": "Note \\(\\eta\\) is greater than 0.5 and (7) only let pseudo points with score greater than \\(\\eta\\) be reliable. Subsequently, \\(\\mathbf{Z} = \\mathrm{diag}(z)\\) is incorporated into the BCE loss for supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.884, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = - \\lambda \\hat {\\boldsymbol {p}} _ {t} ^ {\\top} \\mathbf {Z} \\log \\boldsymbol {p} _ {s} - \\left(\\mathbf {1} _ {n} - \\hat {\\boldsymbol {p}} _ {t}\\right) ^ {\\top} \\mathbf {Z} \\log \\left(\\mathbf {1} _ {n} - \\boldsymbol {p} _ {s}\\right). \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29365" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.478, + 0.107 + ], + "angle": 0, + "content": "However, the second item of (8) in P2P is identically zero:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.112, + 0.483, + 0.131 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u [ P 2 P ]} = - \\lambda \\hat {\\boldsymbol {p}} _ {t} ^ {\\top} \\mathbf {Z} \\log \\boldsymbol {p} _ {s} - 0, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.136, + 0.484, + 0.228 + ], + "angle": 0, + "content": "regardless of \\(\\zeta\\) (embedded in \\(\\mathbf{Z}\\)). Consequently, no gradients are back-propagated from the losses associated with background pixels, which is ill-posed. Utilizing (9) as the loss function is similar to training a two-class classifier with only samples from the category \"1\", which will lead the trained model to output \"1\" regardless of the input." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.238, + 0.379, + 0.256 + ], + "angle": 0, + "content": "4. Visual Explanation with PSAM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.484, + 0.4 + ], + "angle": 0, + "content": "In addition to the theoretical analysis, we implement a framework that supervises a model using labeled and unlabeled data via (5) and (9), respectively, to empirically investigate the behavior of P2P-based semi-supervised counting. Subsequently, PSAM is designed to visualize how the prediction is influenced by (9) during training. Thus, we can design an effective confidence scheme accordingly to counteract the effects revealed by PSAM, thereby enabling semi-supervised learning to function as expected." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.407, + 0.42, + 0.423 + ], + "angle": 0, + "content": "4.1. PSAM: Point-Specific Activation Map" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.55 + ], + "angle": 0, + "content": "An activation map is a heat map with the same spatial resolution as the image features, where local regions with relatively high values are interpreted as being discriminative (informative) for the current output [48, 74, 78]. Most techniques obtain activation maps by linearly aggregating features with their corresponding gradients, as [53, 58] have shown that the partial derivative with respect to a specific element can reflect its contribution to the final prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.627 + ], + "angle": 0, + "content": "Unlike [48, 78], which compute gradients at the image level, crowd counting requires a point-specific activation map (PSAM) to visually explain the regions of interest for every element in \\( \\pmb{p} \\) of (3). Thus, the Jacobian matrix of \\( \\pmb{p} \\) with respect to image feature \\( F \\) is required:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.632, + 0.482, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathbf {J} _ {\\mathcal {D}} = \\frac {\\partial \\boldsymbol {p}}{\\partial F} = \\left[ \\nabla \\boldsymbol {p} _ {[ 1 ]}, \\nabla \\boldsymbol {p} _ {[ 2 ]}, \\dots , \\nabla \\boldsymbol {p} _ {[ n ]} \\right] ^ {\\top}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Afterwards, the \\(q\\)-th heat map \\(H_{[q]}\\) in PSAM is derived by filtering out negative influences on the aggregated feature:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.707, + 0.483, + 0.731 + ], + "angle": 0, + "content": "\\[\nH _ {[ q ]} = \\max \\left(\\sum_ {k = 1} ^ {c} \\nabla \\boldsymbol {p} _ {[ q ]} \\odot F, 0\\right) \\in \\mathbb {R} ^ {h \\times w}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.764 + ], + "angle": 0, + "content": "where the summation operator is applied along the channel dimension, and \\(\\odot\\) denotes the Hadamard product." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In a naive implementation, the computation of \\(\\mathbf{J}_{\\mathcal{D}} \\in \\mathbb{R}^{n \\times hwc}\\) requires a large memory and involves \\(n\\) backpropagations through the computational graph, posing a challenge for visual explanation in pixel-level tasks. However, this issue can be resolved by computing gradients only between \\(\\pmb{p}_{[q]}\\) and features within its receptive field. Gradients in other areas are zero, as there is no path connecting them in the computational graph. Thus, an efficient approach can be implemented, as illustrated in Fig. 2." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.301, + 0.842, + 0.315 + ], + "angle": 0, + "content": "Figure 2. The generation process of PSAM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.906, + 0.417 + ], + "angle": 0, + "content": "In this implementation, the forward process begins with extracting sliding local blocks \\(\\tilde{F}\\) according to the receptive field \\(r\\) of \\(\\mathcal{D}\\), and then decoding \\(\\tilde{F}\\) through (2) in parallel. The final prediction \\(\\pmb{p}\\) is generated by stacking pixels located at the center of \\(r\\times r\\) patches denoted as in \\(\\tilde{D}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.426, + 0.905, + 0.445 + ], + "angle": 0, + "content": "\\[\nF \\in \\mathbb {R} ^ {h \\times w \\times c} \\rightarrow \\tilde {F} \\in \\mathbb {R} ^ {(h w) \\times (r \\times r \\times c)}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.449, + 0.905, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} (\\tilde {F}, \\Theta_ {d}) \\rightarrow \\tilde {D} \\in \\mathbb {R} ^ {n \\times (r \\times r)} \\xrightarrow {\\text {c e n t e r o f p a t c h e s}} \\boldsymbol {p} \\in \\mathbb {R} ^ {n}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.48, + 0.889, + 0.495 + ], + "angle": 0, + "content": "The remaining \\(r^2 - 1\\) pixels in each patch are discarded." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.906, + 0.561 + ], + "angle": 0, + "content": "Denoting the sum of the output as \\(\\xi = \\mathbf{1}_n^\\top \\pmb{p}\\), the gradient \\(\\nabla \\xi = \\partial \\xi / \\partial \\tilde{F}\\) can be obtained via just one backpropagation, and the \\(q\\)-th gradient block, \\(\\nabla \\xi_{[q]}\\), is equal to the gradient of \\(\\pmb{p}_{[q]}\\) with respect to \\(\\tilde{F}_{[q]}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.571, + 0.905, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\nabla \\xi_ {[ q ]} = \\frac {\\partial \\xi}{\\partial \\tilde {F} _ {[ q ]}} = \\frac {\\partial \\xi}{\\partial \\boldsymbol {p} _ {[ q ]}} \\cdot \\frac {\\partial \\boldsymbol {p} _ {[ q ]}}{\\partial \\tilde {F} _ {[ q ]}} = \\frac {\\partial \\boldsymbol {p} _ {[ q ]}}{\\partial \\tilde {F} _ {[ q ]}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.608, + 0.906, + 0.657 + ], + "angle": 0, + "content": "since \\(\\partial \\xi /\\partial \\pmb{p}_{[q]} = 1\\). After that, we can parallel compute the PSAM (denoted as \\(\\tilde{H}_{[q]}\\)) within the receptive field of \\(\\pmb{p}_{[q]}\\) by substituting (14) into the formulation presented in (11):" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.667, + 0.905, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\tilde {H} _ {[ q ]} = \\max \\left(\\sum_ {k = 1} ^ {c} \\nabla \\xi_ {[ q ]} \\odot \\tilde {F} _ {[ q ]}, 0\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.7, + 0.906, + 0.745 + ], + "angle": 0, + "content": "As illustrated in the bottom of Fig. 2, the complete PSAM \\(H_{[1,\\dots,n]}\\) in (11) can be obtained by filling \\(\\tilde{H}_{[1,\\dots,n]}\\) back into the corresponding regions of a fully zero matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.756, + 0.905, + 0.796 + ], + "angle": 0, + "content": "\\[\nH _ {[ q, t ]} = \\left\\{ \\begin{array}{l l} \\tilde {H} _ {[ q, t ^ {\\prime} ]}, & \\text {i f} t \\in \\Omega_ {q} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.805, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\Omega_q\\) represents the index set of \\(p_{[q]}\\)'s receptive field, and \\(t' \\in \\mathbb{R}^2\\) is the transformed coordinate of \\(t \\in \\mathbb{R}^2\\) in \\(H\\), ensuring \\(F_{[t]} \\equiv \\tilde{F}_{[q,t']}.\\) In this way, the matrix memory required to obtain all \\(H_{[q]}\\) is reduced from \\(\\mathbb{R}^{n \\times hwc}\\) to \\(\\mathbb{R}^{n \\times r^2c}\\) (\\(r^2 \\ll hw\\)), and only one back-propagation is needed to obtain PSAMs for all pixels." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29366" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.091, + 0.219, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.164, + 0.212, + 0.173 + ], + "angle": 0, + "content": "(a) MAE/MSE during training" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.091, + 0.328, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.164, + 0.338, + 0.173 + ], + "angle": 0, + "content": "(c) Local PSAM mean (model-L)" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.092, + 0.475, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.164, + 0.466, + 0.173 + ], + "angle": 0, + "content": "(e) aggregated PSAM (model-L)" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.18, + 0.215, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.249, + 0.205, + 0.258 + ], + "angle": 0, + "content": "(b) Sorted average PSAM" + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.178, + 0.33, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.25, + 0.339, + 0.257 + ], + "angle": 0, + "content": "(d) Local PSAM mean (model-U)" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.178, + 0.475, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.25, + 0.466, + 0.259 + ], + "angle": 0, + "content": "(f) aggregated PSAM (model-U)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.271, + 0.483, + 0.354 + ], + "angle": 0, + "content": "Figure 3. Observations in PSAM. (a) The training process, where model-L and model-U are extracted from the 100th and 200th epochs, respectively. (b) Comparing sorted values of PSAMs in the two models. (c) & (d) Visualizing the average of local PSAM, and (e) & (f) the aggregated PSAM to compare model-L and model-U from a global perspective." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.367, + 0.296, + 0.382 + ], + "angle": 0, + "content": "4.2. Observation in PSAM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.483, + 0.436 + ], + "angle": 0, + "content": "We next apply PSAM to visualize the model during semi-supervised counting. The semi-supervised loss function combines (5) and (8):" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.448, + 0.482, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\alpha \\mathcal {L} _ {u} + (1 - \\alpha) \\mathcal {L} _ {l}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.475, + 0.483, + 0.64 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyperparameter to balance the training of labeled and unlabeled data. Following [15], we set \\(\\alpha = 0\\) during the first 100 epochs to generate valid pseudo labels. After that, \\(\\alpha\\) is gradually increased to 1. In Fig. 3(a), we plot how the MAE/MSE changes on the validation set during training. The training failed as evinced by both metrics increasing significantly when unlabeled data were involved in training (after epoch 100). We denote 'model-L' as the model at the 100-th epoch that is trained only on labeled data, and 'model-U' as the model at the 200-th epoch that was trained with both labeled and unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.885 + ], + "angle": 0, + "content": "We next compare explanations of model-L and model-U using PSAM. We visualize the concerned regions of the crowd by extracting the PSAMs corresponding to the foreground pixels \\((\\pmb{p}_{[q]} > 0.5)\\). In Fig. 3(c & d) the element-wise mean of the PSAM patch \\((\\sum_{q} \\tilde{H}_{[q]} \\cdot (\\pmb{p}_{[q]} > 0.5))\\) is displayed to illustrate how the area of the concerned regions changes between model-L and model-U, while Fig. 3(b) presents a detailed value comparison using a line chart. On one hand, Fig. 3(b) shows that the values in model-U's PSAM are larger than those of model-L; on the other hand, Fig. 3(c & d) shows that model-L's PSAM is more concentrated when compared to model-U. These observations reveal how (8) in P2P scheme affects the trained model: \\(\\mathcal{L}_u\\) leads to over-activation in the PSAM of foreground pixels, causing the decoder \\(\\mathcal{D}\\) to falsely identify surrounded pixels as foreground pixels as well." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Next, we visualize the global PSAM of both models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "by compressing \\( H_{[q]} \\) corresponding to foreground pixels: \\( \\bar{H} = \\sum_{q=1}^{n} (\\pmb{p}_{[q]} > 0.5) \\cdot H_{[q]} \\). As displayed in Fig. 3(e & f), the aggregated PSAM of model-L is sparser than that of model-U, which is consistent with the observation for individual points presented in Fig. 3(c & d). Another observation is that the background regions do not change significantly, indicating that almost all false positives occur in the neighborhood of the initial foreground pixels. This suggests that \\( \\mathcal{L}_u \\) does not alter the hyperplane separating crowd and non-crowd features in the high-dimensional space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.259, + 0.662, + 0.276 + ], + "angle": 0, + "content": "5. P2R Matching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.907, + 0.633 + ], + "angle": 0, + "content": "The observations described in last section reveal that \\(\\mathcal{L}_u\\) over-activates the values and enlarges the area of the PSAM corresponding to foreground pixels, which should be suppressed by the supervision of background part in (5), but is actually zero in (8) (\\(\\mathcal{L}_{u[\\mathrm{P2P}]}^{\\prime}\\) in (9)). This over-activation also confirms two important functions of the background item in (5): it not only guides the model to classify the crowd and non-crowd regions in the image, but also suppresses the surrounding person region from being activated. The latter is important since the neighboring pixels of a detected foreground pixel have similar features to the foreground pixel due to the CNN architecture, and thus may become activated if suppression is not applied. This interpretation makes more sense because each person in a crowd image is represented by tens to thousands of pixels, not just one. Supervising the region around the positive pixel with labels of 0 is crucial to a point-based counter. Therefore, during training, it would be more effective to adopt a point-to-region (P2R) matching strategy instead of a point-to-point (P2P) approach. Moreover, as suggested by the above interpretation, P2R is not limited in training with unlabeled data, but can also work in supervised learning with point supervision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.645, + 0.828, + 0.661 + ], + "angle": 0, + "content": "5.1. Training with Labeled Data via P2R" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.906, + 0.715 + ], + "angle": 0, + "content": "Recalling the notation of the prediction \\(\\mathcal{P} = \\{\\pmb {p},\\pmb {x}\\}\\) in (3) and ground truth \\(\\pmb{x}^{\\prime}\\), the matching matrix \\(\\mathbf{M}\\) in our P2R matching is the Hadamard product of two items:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.727, + 0.905, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\mathbf {M} _ {f} \\odot \\left(\\boldsymbol {\\beta} ^ {\\top} \\mathbf {1} _ {m}\\right) \\quad \\in \\mathbf {M} ^ {n \\times m}. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.759, + 0.906, + 0.79 + ], + "angle": 0, + "content": "The first item \\(\\mathbf{M}_f\\in \\{0,1\\}^{n\\times m}\\) is a many-to-one matrix that assigns each pixel in \\(\\mathcal{P}\\) to its nearest point in \\(x^{\\prime}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.803, + 0.905, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {f [ i j ]} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} l _ {2} (i, j) < l _ {2} (i, k) \\forall k \\neq j \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.855, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(l_{2}(i,j) = \\| \\pmb{x}_{[i]} - \\pmb{x}^{\\prime}_{[j]}\\|_{2}\\) is the \\(l_{2}\\)-distance between the \\(i\\)-th pixel's and the \\(j\\)-th point's coordinates. \\(\\beta \\in \\{0,1\\}^{n}\\) in the second item of (18) is a vector marking foreground" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29367" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.195, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.16, + 0.176, + 0.169 + ], + "angle": 0, + "content": "(a) P2P scheme" + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.093, + 0.332, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.16, + 0.305, + 0.169 + ], + "angle": 0, + "content": "(b) matched point" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.093, + 0.474, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.16, + 0.463, + 0.169 + ], + "angle": 0, + "content": "(c) semi-supervised points" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.174, + 0.191, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.243, + 0.177, + 0.252 + ], + "angle": 0, + "content": "(d) P2R scheme" + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.176, + 0.332, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.243, + 0.306, + 0.253 + ], + "angle": 0, + "content": "(e) matched region" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.175, + 0.475, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.243, + 0.465, + 0.253 + ], + "angle": 0, + "content": "(f) semi-supervised regions" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.264, + 0.485, + 0.389 + ], + "angle": 0, + "content": "Figure 4. Difference between P2P and P2R matching. (a) and (d) demonstrate an overall difference. P2P focuses only on foreground pixels in \\(\\mathbf{P}_s\\), while P2R considers background pixels as well. (b) and (e) show how the matching is performed in P2P and P2R, respectively. P2R segments out local regions for each pseudo-label, whereas P2P only detects one point. (c) and (f) illustrate how untrusted predictions are filtered. P2P retains only foreground pixels for loss computation, while P2R also keeps pixels in the neighborhoods of their corresponding pseudo points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.483, + 0.441 + ], + "angle": 0, + "content": "regions, by setting those pixels in \\(\\mathcal{P}\\) far from any points in \\(x^{\\prime}\\) as zero:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.446, + 0.482, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\beta_ {[ i ]} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\min _ {j} l _ {2} (i, j) < \\mu \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.583 + ], + "angle": 0, + "content": "in which \\(\\mu\\) is a hyper parameter indicating the radius if the neighborhood of each GT point is considered as a circle. As displayed in Fig. 4(e), \\(\\mathbf{M}\\) segments out a coarse local region that is consistent with the location of the GT point. After that, the learning objective \\(\\hat{\\pmb{p}}\\) is derived by defining the cost matrix \\(\\mathbf{C} \\in \\mathbb{R}^{n \\times m}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.587, + 0.483, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {[ i j ]} = \\left\\{ \\begin{array}{l l} \\tau l _ {2} (i, j) - \\mathcal {S} (\\boldsymbol {p} _ {[ i ]}), & \\text {i f} \\mathbf {M} _ {[ i j ]} = 1 \\\\ \\infty , & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.664 + ], + "angle": 0, + "content": "Then, the pixel with the minimum cost in each column of \\(\\mathbf{C}\\) is marked as the potential foreground pixel:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.67, + 0.482, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} _ {[ i j ]} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\mathbf {C} _ {[ i j ]} < \\mathbf {C} _ {[ k j ]} \\forall k \\neq i \\\\ 0, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Finally, the learning objective \\(\\hat{\\pmb{p}}\\) is derived as \\(\\hat{\\pmb{p}} = \\hat{\\mathbf{M}}\\mathbf{1}\\), and the loss \\(\\mathcal{L}_l\\) is computed between \\(\\pmb{p}\\) and \\(\\hat{\\pmb{p}}\\) according to BCE loss formulated as (5) during training with labeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.84 + ], + "angle": 0, + "content": "An extra benefit is that computing \\(\\mathbf{M}\\) and \\(\\hat{\\mathbf{M}}\\) in P2R requires only finding the minimum of each row or column in the matrices, which is significantly faster than executing the Hungarian algorithm — a time-consuming component of the P2P matching strategy [57]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.421, + 0.864 + ], + "angle": 0, + "content": "5.2. Training with Unlabeled Data via P2R" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Recalling the pseudo-label \\(\\mathcal{P}_t^\\prime = \\{\\pmb {p}_t^\\prime ,\\pmb{x}_t^\\prime \\}\\) from the teacher and the supervised student output as \\(\\mathcal{P}_s = \\{\\pmb {p}_s,\\pmb {x}_s\\}\\), the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.199 + ], + "angle": 0, + "content": "matching matrix \\(\\mathbf{M}_{st}\\) between \\(\\mathcal{P}_s\\) and \\(\\mathcal{P}_t^\\prime\\) is computed via (18) in P2R, and then combined with the confidence in (7) to mask unreliable pseudo labels. Additionally, pixels far from any pseudo points can be directly marked as reliable background pixels, as suggested by observations from the aggregated PSAM in Fig. 3(e & f). Thus, the confidence diagonal matrix \\(\\mathbf{Z}\\) is formulated as follows with \\(\\zeta = \\mathbb{1}(\\hat{\\pmb{p}}_t' > \\eta)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.205, + 0.905, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} = \\operatorname {d i a g} \\left[ \\mathbf {M} _ {s t} \\boldsymbol {\\zeta} + \\left(\\mathbf {1} _ {n} - \\boldsymbol {\\beta}\\right) \\right] \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.905, + 0.259 + ], + "angle": 0, + "content": "The final loss supervising unlabeled data \\(\\mathcal{P}_s\\) with pseudo labels is derived by substituting \\(\\mathbf{Z}\\) in (23) into (8)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.906, + 0.379 + ], + "angle": 0, + "content": "As displayed in Fig. 4(f), the confidence for each pseudo point is propagated to it corresponding region. On one hand, reliable foreground parts (blue) contain a pixel whose pseudo-label is 1, while the pseudo-labels of its surrounding pixels are marked as 0. On the other hand, all pixels in the reliable background parts (khaki) are assigned pseudolabels of 0 to train the counter with unlabeled data. The uncolored regions are ignored because their confidence is 0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.905, + 0.41 + ], + "angle": 0, + "content": "The pseudo code and flow chart of the proposed P2R loss can be found in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.422, + 0.646, + 0.439 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.906, + 0.522 + ], + "angle": 0, + "content": "This section explores how the proposed P2R training approach enhances a point-based counting model using both labeled and unlabeled data across three parts: semi-supervised crowd counting, unsupervised domain adaptation (UDA) for crowd counting, and ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.705 + ], + "angle": 0, + "content": "In the semi-supervised experiments, our method is implemented on four crowd counting datasets: ShTech A/B [73], UCF-QNRF [12], and JHU++ [55]. Following DAC [67] and OT-M [26], three protocols are applied: \\(5\\%\\), \\(10\\%\\), and \\(40\\%\\) of labeled data, with the remaining crowd images involved in training without annotations. In the UDA part, labeled data from one domain, e.g., ShTech A, is used to initialize a counting model, while unlabeled data from another domain is utilized to capture the domain characteristics of the application scenes. In the ablation study, we test different hyperparameters to demonstrate their impact on the performance of P2R." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.757, + 0.728 + ], + "angle": 0, + "content": "6.1. Semi-Supervised Counting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Tab. 1 presents the experimental results of semi-supervised counting. Our method achieves the best performance in nearly all protocols across all crowd datasets. Besides, using \\(5\\%\\) data with P2R surpasses the fully-supervised learning with \\(10\\%\\) data and is even equivalent to other semi-supervised learning methods using \\(10\\%\\) data. Similar performances are also observed between the semi-supervised training with \\(10\\%\\) and \\(40\\%\\) labeled data. These results demonstrate that P2R is label-efficient and advantageous in semi-supervised learning. In Fig. 5, we visualize predictions of DAC [67], OTM [26], and our P2R. The 3rd row is" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29368" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.347 + ], + "angle": 0, + "content": "
Label Pct.MethodsShTech A [73]ShTech B [73]UCF-QNRF [12]JHU++ [55]
MAEMSEMAEMSEMAEMSEMAEMSE
5%label-only93.7155.213.124.0132.7231.7104.1383.5
MT [59]104.7156.919.333.2172.4284.9101.5363.5
L2R [32]103.0155.420.327.6160.1272.3101.4338.8
GP [54]102.0172.015.727.9160.0275.098.9355.7
DAC [23]85.4134.512.622.8120.2209.382.2294.9
OT-M [26]83.7133.312.621.5118.4195.482.7304.5
P2R (ours)69.9119.59.116.6100.1182.577.8293.5
10%label-only84.0138.310.719.2112.4186.878.7305.5
MT [59]319.394.515.624.5156.1245.5250.390.2
L2R [32]90.3153.515.624.4148.9249.887.5315.3
IRAST [34]86.9148.914.722.9135.6233.486.7303.4
DAC [23]74.9115.511.119.1109.0187.275.9282.3
OT-M [26]80.1118.510.818.2113.1186.773.0280.6
CU [18]70.7116.69.717.7104.01644.274.8281.6
P2R (ours)65.2114.68.414.594.9167.268.7272.3
40%label-only64.5105.68.114.099.2174.768.8283.5
MT [59]88.2151.115.925.7147.2249.6121.5388.9
L2R [32]86.5148.216.825.1145.1256.1123.6376.1
SUA [42]68.5121.914.120.6130.3226.380.7290.8
DAC [23]67.5110.79.614.691.1153.465.1260.0
OT-M [26]70.7114.58.113.1100.6167.672.1272.0
P2R (ours)55.695.06.811.086.0144.363.3271.1
100%P2PNet [57]52.7485.066.259.9085.32154.5061.25258.65
P2R (ours)51.0279.686.179.8483.26138.1158.83253.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.358, + 0.483, + 0.387 + ], + "angle": 0, + "content": "Table 1. Comparison with other recent methods on four benchmark datasets under different labeled protocols." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.429 + ], + "angle": 0, + "content": "a failure case of P2R due to crowd motion blur. However, the overall performance of P2R is best." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.43, + 0.483, + 0.597 + ], + "angle": 0, + "content": "The last row of Tab. 1 also compares P2P and P2R under a fully-supervised learning scheme (100% labels); P2R performs better than P2PNet on all crowd counting datasets. Additionally, as described in Section 5.1, P2R is faster than P2P due to the absence of the Hungarian algorithm. We compared the efficiency using an image with a resolution of \\(576 \\times 960\\) and 775 annotated points — P2P requires an average of 0.4307 seconds for loss computation, while P2R only needs 0.0064 seconds, which is nearly 68 times faster than P2P. These comparisons demonstrate that P2R surpasses P2P in both effectiveness and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.606, + 0.443, + 0.623 + ], + "angle": 0, + "content": "6.2. Unsupervised Domain Adaptation (UDA)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": "UDA aims to transfer knowledge learned from a source domain to a target domain by designing modules or learning objectives that capture domain-agnostic features [11, 29, 43, 65, 70, 72]. Unlike semi-supervised learning, where labeled and unlabeled data originate from the same domain, UDA typically involves a domain gap between the labeled (source domain) and unlabeled crowd data (target domain). This gap often manifests as differences in crowd distribution, density levels, and perspectives. Our P2R can also be applied to UDA. When an unlabeled image from the target domain is provided, the threshold automatically partitions its corresponding pseudo points into two categories: higher-score \\((> \\eta)\\) and lower-score \\((< \\eta)\\). The higher-score class is considered to be closer to the source domain, while the lower-score class is considered to be farther from the source domain. Thus, the higher-score pseudo points can participate in the training to gradually guide the model to learn to count in the target domain." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.613, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.131, + 0.612, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.187, + 0.612, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.238, + 0.592, + 0.245 + ], + "angle": 0, + "content": "ground truth: 61" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.089, + 0.71, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.131, + 0.71, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.187, + 0.69, + 0.195 + ], + "angle": 0, + "content": "DAC [23]: 49.21" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.089, + 0.808, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.131, + 0.808, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.187, + 0.789, + 0.195 + ], + "angle": 0, + "content": "OTM [26]: 51.43" + }, + { + "type": "image", + "bbox": [ + 0.814, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.814, + 0.131, + 0.907, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.835, + 0.187, + 0.888, + 0.195 + ], + "angle": 0, + "content": "P2R (ours): 60.0" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.195, + 0.612, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.238, + 0.592, + 0.245 + ], + "angle": 0, + "content": "ground truth: 61" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.195, + 0.71, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.238, + 0.692, + 0.245 + ], + "angle": 0, + "content": "DAC [23]: 146.09" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.195, + 0.808, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.238, + 0.791, + 0.245 + ], + "angle": 0, + "content": "OTM [26]: 126.11" + }, + { + "type": "image", + "bbox": [ + 0.814, + 0.195, + 0.907, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.238, + 0.89, + 0.244 + ], + "angle": 0, + "content": "P2R (ours): 227.0" + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.256, + 0.863, + 0.271 + ], + "angle": 0, + "content": "Figure 5. qualitative comparison with other models." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.284, + 0.895, + 0.397 + ], + "angle": 0, + "content": "
METHODSA → BA → QB → AB → Q
MAEMSEMAEMSEMAEMSEMAEMSE
DGDCCUS [40]12.624.6119.4216.6121.8203.1179.1316.2
MPCount [43]11.419.7115.7199.899.6182.9165.6290.4
UDABL [38] (w/o DA)15.925.8166.7287.6138.1228.1226.4411.0
RBT [35]13.429.3175.0294.8112.2218.2211.3381.9
C2MoT [70]12.421.1125.7218.3120.7192.0198.9368.0
FGFD [79]12.723.3124.1242.0123.5210.7209.7384.7
FSIM [80]11.119.3105.3191.1120.3202.6194.9324.5
UDAP2R (w/o DA)25.635.7155.1286.8130.2229.1173.3329.8
P2R (w/ DA)10.618.7105.3194.889.3176.0139.5243.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.478 + ], + "angle": 0, + "content": "Table 2. Unsupervised domain adaptation for crowd counting. The LHS and RHS of “\\(\\rightarrow\\)” represent adaptation direction. Q, A, and B indicate UCF-QNRF [12], ShTech A, and ShTech B [73]. \\(DG\\) indicates domain generalization. The training of UDA requires samples from the target domain, whereas DG does not." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.688 + ], + "angle": 0, + "content": "We conduct experiments on four protocols following [70]. The results are presented in Tab. 2. The last two rows demonstrate the results of the proposed P2R. The poor performance obtained without domain adaptation highlights the domain gap between the source data and target data. However, when the unlabeled data from the target domain is involved in training, the estimation errors are significantly reduced, and the results surpass previous methods with complex network structures for domain adaptation. These advanced experimental results demonstrate that the threshold works as expected to capture samples close to the source domain and gradually teaches the model to count in the target domain." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.667, + 0.713 + ], + "angle": 0, + "content": "6.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.765 + ], + "angle": 0, + "content": "We next conduct ablation studies to investigate the influence of hyperparameters in P2R, including \\(\\tau\\) in (21), the threshold \\(\\eta\\), \\(\\mu\\) in (20), and \\(\\alpha\\) in (17). The results are in Fig. 6" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The impact of \\(\\tau\\) in (21). \\(\\tau\\) is a hyperparameter to balance the distance and foreground score when selecting the foreground pixel from the matched region. When \\(\\tau \\rightarrow \\infty\\), the predicted score is ignored, which is equivalent to computing cross-entropy between the prediction and the ground truth point map. If \\(\\tau = 0\\), the pixel with the highest score is chosen as the foreground in \\(\\hat{\\pmb{p}}\\). However, Fig. 6(a) shows that neither extreme is optimal. The best performance is achieved when \\(\\tau = 8\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29369" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.087, + 0.282, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.183, + 0.245, + 0.196 + ], + "angle": 0, + "content": "(a) errors vs. \\(\\tau\\)" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.197, + 0.282, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.29, + 0.245, + 0.304 + ], + "angle": 0, + "content": "(c) errors vs. \\(\\mu\\)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.087, + 0.468, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.183, + 0.431, + 0.197 + ], + "angle": 0, + "content": "(b) errors vs. \\(\\eta\\)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.197, + 0.468, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.289, + 0.438, + 0.307 + ], + "angle": 0, + "content": "(d) errors vs. \\(\\frac{\\alpha}{1 - \\alpha}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.484, + 0.372 + ], + "angle": 0, + "content": "Figure 6. The estimation errors vs. various hyperparameters. (a) is conducted on the fully-supervised ShTech A since it is irrelevant to semi-supervised counting, while (b)-(d) are conducted with \\(5\\%\\) labeled data in ShTech A." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.385, + 0.483, + 0.55 + ], + "angle": 0, + "content": "The effect of threshold \\(\\eta\\). \\(\\eta\\) defines the threshold to distinguish reliable and unreliable regions. Since all pseudo points are selected by picking pixels whose score is greater than 0.5 in the prediction, the threshold should also be greater than 0.5. Moreover, a lower \\(\\eta\\) introduces unreliable pseudo-labels into training, while a higher \\(\\eta\\) cannot fully utilize the unlabeled data as most parts cannot participate in training. As shown in Fig. 6(b), the estimation errors exhibit a trend of first decreasing and then increasing within the range of 0.6 to 0.9, with the lowest errors occurring when the threshold is set to 0.7." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.552, + 0.483, + 0.734 + ], + "angle": 0, + "content": "The radius of each point \\(\\mu\\). The vector \\(\\beta\\) in (20) derived from \\(\\mu\\) marks the foreground points' neighborhood and background regions. It not only restricts the positions where foreground pixels appear by substituting \\(\\beta\\) into (18), but also directly identifies reliable background parts via the second item in (23) based on the observation of PSAM in Section 4.2. Fig. 6(c) presents how \\(\\mu\\) affects the counting performance. It shows the MAE remains around 68 when \\(\\mu\\) increases from 32 to 256, indicating that a larger \\(\\mu\\) does not significantly affect the results. This is consistent with the observation in PSAM since pixels far from the foreground points are less likely to be activated during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The weight of unlabeled loss. Fig. 6(d) presents how \\(\\alpha\\) in (17) affects the performance. A large \\(\\alpha\\) focuses more on unlabeled data, leading to training failure, while a small \\(\\alpha\\) may cause the model to overfit on labeled data. In the experiments with ShTech A and \\(5\\%\\) labeled data, \\(\\frac{\\alpha}{1 - \\alpha} = 2\\), i.e., \\(\\alpha = 2/3\\), results in the best performance. Additionally, as long as \\(\\alpha\\) is not 0, the point-based counter can benefit from unlabeled data, as the maximum MAE/MSE in Fig. 6(d) are smaller than 80/150, which is much better than the model trained with only \\(5\\%\\) labeled data (MAE: 93.7, MSE: 155.2)." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.095, + 0.728, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.218, + 0.692, + 0.228 + ], + "angle": 0, + "content": "(a) sorted average PSAM" + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.095, + 0.882, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.136, + 0.881, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.176, + 0.881, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.218, + 0.884, + 0.228 + ], + "angle": 0, + "content": "(b) Local & aggregated PSAM" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.242, + 0.906, + 0.312 + ], + "angle": 0, + "content": "Figure 7. The comparison of PSAMs among different training schemes. Model-L (green) is trained with only labeled data; model-U_P2P (blue) is trained with both labeled data and unlabeled data using the P2P scheme, while model-U_P2R (orange) is trained with the proposed P2R scheme." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.907, + 0.505 + ], + "angle": 0, + "content": "PSAM comparison. Fig. 7 visualizes the PSAMs of models trained with the P2P and P2R schemes. Comparing model-L (green) and model-U_P2R (blue), the latter effectively restricts the PSAM to the neighborhood of the concerned foreground pixel, and its aggregated PSAM is also sparser than that of model-L, demonstrating the effectiveness of P2R. Comparing model-U_P2P and model-U_P2R in Fig. 7(a), the values of the PSAM are similar at the non-peak indexes, but the latter has a higher activation at the peak position compared to the former, indicating that semi-supervised learning further enhances the identification ability of the point-based counter." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.518, + 0.634, + 0.533 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.86 + ], + "angle": 0, + "content": "This paper presents a P2R scheme to train a point-based counter for semi-supervised crowd counting. At the beginning, we establish a pseudo-labeling framework based on P2P matching, but its formulation is ill-posed since the confidence is only applicable to foreground pixels and cannot be propagated to backgrounds. To observe how the foreground and background feature are learned, we propose PSAM to visually interpret the concerned region and activation value of all foreground pixels. From the visualized activation map, it is observed that the neighborhood of each foreground pixel is over-activated and falsely recognized as instances by the decoder. Based on this observation, we replace P2P with P2R by segmenting the prediction into multiple regions containing the corresponding instances. Thus, pixels in the neighborhood of the concerned pseudo points can share the weights with the corresponding pixels. Additionally, the time-consuming Hungarian algorithm is no longer necessary in P2R. Several outstanding experimental results in semi-supervised counting and unsupervised domain adaptation demonstrate the advantages of our P2R matching strategy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by a grant from City University of Hong Kong (Project No. 7030010)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29370" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Shahira Abousamra, Minh Hoai, Dimitris Samaras, and Chao Chen. Localization in the crowd with topological constraints. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 872-881, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Areej Althali, Amal Balabid, Reem Alharthi, Bander Alzahrani, Reem Alotaibi, and Ahmed Barnawi. Anomalous event detection and localization in dense crowd scenes. Multimedia Tools and Applications, 82(10):15673-15694, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Deepak Babu Sam, Shiv Surya, and R Venkatesh Babu. Switching convolutional neural network for crowd counting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5744-5752, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[4] Antoni B Chan and Nuno Vasconcelos. Bayesianoisson regression for crowd counting. In 2009 IEEE 12th international conference on computer vision, pages 545-551. IEEE, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[5] Antoni B Chan and Nuno Vasconcelos. Counting people with low-level features and bayesian regression. IEEE Transactions on image processing, 21(4):2160-2177, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[6] Antoni B Chan, Zhang-Sheng John Liang, and Nuno Vasconcelos. Privacy preserving crowd monitoring: Counting people without people models or tracking. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-7. IEEE, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] Mulin Chen, Qi Wang, and Xuelong Li. Anchor-based group detection in crowd scenes. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1378-1382. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[8] Zhi-Qi Cheng, Qi Dai, Hong Li, Jingkuan Song, Xiao Wu, and Alexander G Hauptmann. Rethinking spatial invariance of convolutional networks for object counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19638-19648, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[9] Junyu Gao, Tao Han, Qi Wang, Yuan Yuan, and Xuelong Li. Learning independent instance maps for crowd localization. arXiv preprint arXiv:2012.04164, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.636, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[10] Tao Han, Lei Bai, Lingbo Liu, and Wanli Ouyang. Steerer: Resolving scale variations for counting and localization via selective inheritance learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21848-21859, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[11] Yuhang He, Zhiheng Ma, Xing Wei, Xiaopeng Hong, Wei Ke, and Yihong Gong. Error-aware density isomorphism reconstruction for unsupervised cross-domain crowd counting. In Proceedings of the AAAI conference on artificial intelligence, pages 1540-1548, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[12] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Brian Kenji Iwana, Ryohei Kuroki, and Seiichi Uchida. Explaining convolutional neural networks using softmax gradient layer-wise relevance propagation. In 2019 IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision Workshop (ICCVW), pages 4176-4185. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[14] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[15] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[16] Bastian Leibe, Edgar Seemann, and Bernt Schiele. Pedestrian detection in crowded scenes. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 878-885. IEEE, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[17] Victor Lempitsky and Andrew Zisserman. Learning to count objects in images. Advances in neural information processing systems, 23, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[18] Chen Li, Xiaoling Hu, Shahira Abousamra, and Chao Chen. Calibrating uncertainty for semi-supervised crowd counting. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16685-16695. IEEE, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[19] Hanxiao Li, Yonghong Song, and Tong Geng. Semi-supervised crowd counting based on hard pseudo-labels. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[20] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1091-1100, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[21] Dingkang Liang, Wei Xu, and Xiang Bai. An end-to-end transformer model for crowd localization. In ECCV, pages 38-54. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[22] Hui Lin, Xiaopeng Hong, Zhiheng Ma, Xing Wei, Yunfeng Qiu, Yaowei Wang, and Yihong Gong. Direct measure matching for crowd counting. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages = 837–844, publisher = ijcai.org, year = 2021, .2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[23] Hui Lin, Zhiheng Ma, Xiaopeng Hong, Yaowei Wang, and Zhou Su. Semi-supervised crowd counting via density agency. In ACM Multimedia, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[24] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19628-19637, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[25] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, Zhou Su, Xiaopeng Hong, and Deyu Meng. Semi-supervised counting via pixel-by-pixel density distribution modelling. arXiv preprint arXiv:2402.15297, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[26] Wei Lin and Antoni B Chan. Optimal transport minimization: Crowd localization on density maps for semi-supervised counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21663-21673, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Wei Lin and Antoni B Chan. A fixed-point approach to unified prompt-based counting. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29371" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "AAAI Conference on Artificial Intelligence, pages 3468-3476, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[28] Wei Lin, Jia Wan, and Antoni B Chan. Proximal mapping loss: Understanding loss functions in crowd counting & localization. In The Thirteenth International Conference on Learning Representations, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[29] Wei Lin, Junyu Gao, Qi Wang, and Xuelong Li. Learning to detect anomaly events in crowd scenes from synthetic data. Neurocomputing, 436:248-259, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[30] Wei Lin, Kunlin Yang, Xinzhu Ma, Junyu Gao, Lingbo Liu, Shinan Liu, Jun Hou, Shuai Yi, and Antoni B Chan. Scale-prior deformable convolution for exemplar-guided class-agnostic counting. In BMVC, page 313, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[31] Chengxin Liu, Hao Lu, Zhiguo Cao, and Tongliang Liu. Point-query quadtree for crowd counting, localization, and more. In CVPR, pages 1676-1685, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.483, + 0.377 + ], + "angle": 0, + "content": "[32] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Leveraging unlabeled data for crowd counting by learning to rank. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7661-7669, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[33] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Exploiting unlabeled data in cnns by self-supervised learning to rank. IEEE transactions on pattern analysis and machine intelligence, 41(8):1862-1878, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[34] Yan Liu, Lingqiao Liu, Peng Wang, Pingping Zhang, and Yinjie Lei. Semi-supervised crowd counting via self-training on surrogate tasks. In European Conference on Computer Vision, pages 242-259. Springer, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[35] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, pages 129-137, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[36] Yanbo Liu, Yingxiang Hu, Guo Cao, and Yanfeng Shang. Semi-supervised crowd counting via multi-task pseudo-label self-correction strategy. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[37] Zheng Ma and Antoni B Chan. Counting people crossing a line using integer programming and local features. IEEE Transactions on Circuits and Systems for Video Technology, 26(10):1955-1969, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[38] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In Proceedings of the IEEE International Conference on Computer Vision, pages 6142-6151, 2019. 1, 2, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[39] Zhiheng Ma, Xing Wei, Xiaopeng Hong, Hui Lin, Yunfeng Qiu, and Yihong Gong. Learning to count via unbalanced optimal transport. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2319-2327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[40] Lucas Mansilla, Rodrigo Echeveste, Diego H Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6630-6638, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[41] Riccardo Mazzon, Fabio Poiesi, and Andrea Cavallaro. Detection and tracking of groups in crowd. In 2013 10th IEEE International conference on advanced video and signal based surveillance, pages 202-207. IEEE, 2013. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[42] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15549-15559, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[43] Zhuoxuan Peng and S-H Gary Chan. Single domain generalization for crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28025-28034, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[44] Harish Guruprasad Ramaswamy et al. Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 983-991, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[45] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3394-3403, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[46] Weihong Ren, Xinchao Wang, Jiandong Tian, Yandong Tang, and Antoni B Chan. Tracking-by-counting: Using network flows on crowd density maps for tracking multiple targets. IEEE Transactions on Image Processing, 30:1439-1452, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[47] Jihye Ryu and Kwangho Song. Crowd counting and individual localization using pseudo square label. IEEE Access, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[48] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[49] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 323-331, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[50] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19618-19627, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[51] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19618-19627, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[52] Weibo Shu, Jia Wan, and Antoni B Chan. Generalized characteristic function loss for crowd analysis in the frequency domain. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[53] Karen Simonyan. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.805, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[54] Vishwanath A Sindagi, Rajeev Yasarla, Deepak Sam Babu, R Venkatesh Babu, and Vishal M Patel. Learning to count in the crowd from limited labeled data. In European Conference on Computer Vision, pages 212-229. Springer, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[55] Vishwanath A Sindagi, Rajeev Yasarla, and Vishal M Patel. Jhu-crowd++: Large-scale crowd counting dataset and a" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29372" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "benchmark method. IEEE transactions on pattern analysis and machine intelligence, 44(5):2594-2609, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[56] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[57] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In CVPR, pages 3365-3374, 2021. 1, 2, 3, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[58] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin A. Riedmiller. Striving for simplicity: The all convolutional net. In 3rd International Conference on Learning Representations, ICLR 2015, Workshop Track Proceedings, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[59] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[60] Jia Wan, Qingzhong Wang, and Antoni B Chan. Kernel-based density map generation for dense object counting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1357-1370, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[61] Jia Wan, Ziquan Liu, and Antoni B Chan. A generalized loss function for crowd counting and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1974-1983, 2021. 1, 2, 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.567 + ], + "angle": 0, + "content": "[62] Jia Wan, Qiangqiang Wu, Wei Lin, and Antoni Chan. Robust zero-shot crowd counting and localization with adaptive resolution sam. In European Conference on Computer Vision, pages 478-495. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[63] Boyu Wang, Huidong Liu, Dimitris Samaras, and Minh Hoai Nguyen. Distribution matching for crowd counting. Advances in neural information processing systems, 33:1595-1607, 2020. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[64] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[65] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8198-8207, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[66] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. Nwpu-crowd: A large-scale benchmark for crowd counting and localization. IEEE transactions on pattern analysis and machine intelligence, 43(6):2141-2149, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[67] Qi Wang, Wei Lin, Junyu Gao, and Xuelong Li. Density-aware curriculum learning for crowd counting. IEEE Transactions on Cybernetics, 52(6):4675-4687, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[68] Qi Wang, Wei Lin, Junyu Gao, and Xuelong Li. Density-aware curriculum learning for crowd counting. IEEE Transactions on Cybernetics, 52(6):4675-4687, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[69] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Pixel-wise crowd understanding via synthetic data. International Journal of Computer Vision, 129(1):225-245, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[70] Qiangqiang Wu, Jia Wan, and Antoni B Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 658-666, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[71] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15548-15557, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[72] Qi Zhang, Wei Lin, and Antoni B Chan. Cross-view cross-scene multi-view crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 557-567, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[73] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 589-597, 2016. 1, 2, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[74] Chenyang Zhao and Antoni B. Chan. ODAM: gradient-based instance-specific visual explanations for object detection. In ICLR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[75] Chunhui Zhao, Zhiyuan Zhang, Jinwen Hu, Dong Wang, Bin Fan, Quan Pan, and Qiang He. Crowd anomaly event detection in surveillance video based on the evolution of the spatial position relationship feature. In 2018 IEEE 14th International Conference on Control and Automation (ICCA), pages 247-252, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[76] Chenyang Zhao, Janet H Hsiao, and Antoni B Chan. Gradient-based instance-specific visual explanations for object specification and object discrimination. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[77] Chenyang Zhao, Kun Wang, Xingyu Zeng, Rui Zhao, and Antoni B Chan. Gradient-based visual explanation for transformer-based clip. In International Conference on Machine Learning, pages 61072-61091. PMLR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.769 + ], + "angle": 0, + "content": "[79] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5659-5668, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.77, + 0.905, + 0.826 + ], + "angle": 0, + "content": "[80] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29373" + } + ] +] \ No newline at end of file diff --git a/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/7bc00c8f-f0e1-486b-af9f-cab0b77e9a84_origin.pdf b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/7bc00c8f-f0e1-486b-af9f-cab0b77e9a84_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fdb50ea0d123b60ec10e84ba9700dcf929ae4144 --- /dev/null +++ b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/7bc00c8f-f0e1-486b-af9f-cab0b77e9a84_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f0761bdd1e5a2856e42d6070d1667c1fcab98801d831d2223cc86199ecc3816 +size 1628156 diff --git a/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/full.md b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..95575d4854af339eeb18ed3e4d23594731ebd66a --- /dev/null +++ b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/full.md @@ -0,0 +1,507 @@ +# Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting + +Wei Lin, Chenyang Zhao, and Antoni B. Chan + +Department of Computer Science, City University of Hong Kong + +elonlin24@gmail.com, chenyzhao9-c@my.cityu.edu.hk, abchan@cityu.edu.hk + +# Abstract + +Point detection has been developed to locate pedestrians in crowded scenes by training a counter through a point-to-point (P2P) supervision scheme. Despite its excellent localization and counting performance, training a point-based counter still faces challenges concerning annotation labor: hundreds to thousands of points are required to annotate a single sample capturing a dense crowd. In this paper, we integrate point-based methods into a semi-supervised counting framework based on pseudo-labeling, enabling the training of a counter with only a few annotated samples supplemented by a large volume of pseudolabeled data. However, during implementation, the training encounters issues as the confidence for pseudo-labels fails to be propagated to background pixels via the P2P. To tackle this challenge, we devise a point-specific activation map (PSAM) to visually interpret the phenomena occurring during the ill-posed training. Observations from the PSAM suggest that the feature map is excessively activated by the loss for unlabeled data, causing the decoder to misinterpret these over-activations as pedestrians. To mitigate this issue, we propose a point-to-region (P2R) scheme to substitute P2P, which segments out local regions rather than detects a point corresponding to a pedestrian for supervision. Consequently, pixels in the local region can share the same confidence with the corresponding pseudo points. Experimental results in both semi-supervised counting and unsupervised domain adaptation highlight the advantages of our method, illustrating P2R can resolve issues identified in PSAM. The code is available at https://github.com/Elin24/P2RLoss. + +# 1. Introduction + +Crowd scene analysis has been investigated for many years due to its significant applications in smart cities and urban safety [2, 7, 29, 41, 46, 62, 75]. Crowd counting and localization in extremely dense crowd scenes, as the fundamental task in this field, have attracted considerable attention from researchers in computer vision [4, 5, 37, 61, 66, 68]. Although most supervised crowd counting methods fol + +![](images/f8654d6fa55092cd93b854925481932197aa3f3617e69f7065ee6cc69b0c2dcb.jpg) +Figure 1. The workflow of semi-supervised point-based counting methods. The teacher model generates pseudo labels by extracting the foreground pixels, while the student model takes the corresponding strongly augmented image as input to construct the computation graph. The training loss between the pseudo label and the student's prediction involves two steps: the proposed P2R matching and the weighted cross-entropy computation. + +![](images/2e4ee55c1d884fd65d86c1d6233a1875ae7bb5d03211fa2166453f6520e603bf.jpg) + +![](images/29dad3d08f89e07b65a71e50ecf6be38f07b709bc2245095e9ed0104f2a45e1e.jpg) +Prediction +w/o gradient +P2R Matching + +![](images/928f6f3b2ac36d47a69f7b9084e69bcb6d14188db1b83eb2cc32f47965cdd4e1.jpg) +>0.5 +Sudo Label +Weighted Cross Entropy +w/ gradient + +![](images/edd9073bc1c491cc4c1267cd01898ab7ba8b88a9b3e9d4acce3d0c2a8fb6e431.jpg) + +![](images/2ee8b5bdfd9bd808a379c9bfaf6cbdd6c20f24a0ccbb21ec8001ceeddf4b4b32.jpg) + +![](images/b656620bb22599a32d692f01e06c6b70d78e1199393337a31b3729d11c0f8067.jpg) +Student + +low the scheme of density regression [10, 24, 38, 61, 73], there are also relevant studies implementing point detection for crowd localization and counting [21, 31, 57], due to its straightforwardness in indicating pedestrians' positions. These approaches use the Hungarian algorithm to perform point-to-point (P2P) matching between predicted point maps and ground-truth (GT) pedestrian annotations during training. Therefore, these matched pixels are considered as pedestrians' positions, while the others are designated as background for loss computation. This scheme successfully trains a point detector for crowd analysis, yielding both good localization and counting performance. + +Great progress has been made in supervised counting, but the labor-intensive nature of annotating data has hindered the development of crowd understanding for a long time [26, 66]. Addressing this issue, semi-supervised counting is gradually being developed, aiming to train a model using large amounts of unlabeled data alongside only a few annotated samples. Similar to supervised counting, most semi-supervised methods also follow the procedure of density regression. Notably, the method proposed in OT-M [26] stands out by employing a simple self-training scheme based on mean-teacher [59] and FixMatch [56]. However, there is no study exploring semi-supervised counting and + +localization using a point-detection method. To fill this gap, this paper explores an approach to train a point detector with limited annotated data. + +As shown in Fig. 1, typical semi-supervised counting methods rely on pseudo-labeling, which generates pseudo labels to serve as learning targets for unlabeled samples using a model pre-trained on a small amount of annotated data. During training, there is a confidence score to determine whether a concerned prediction should be retained as a pseudo label. However, if the P2P matching strategy is applied to compute loss within the pipeline presented in Fig. 1, there is no reasonable way to define the confidence for unmatched background pixels, as P2P matching only cares about matched foreground pixels. Specifically, a foreground pixel can be retained for training if its matched pseudo-label's foreground probability exceeds a predefined threshold, similar to the approach used in semi-supervised classification tasks [15, 56]. However, background pixels lack confidence scores since no pseudo-labels are assigned to them after matching. Experimental results also demonstrate that ignoring the background leads to the breakdown of semi-supervised training. + +To address the aforementioned issue, we propose a gradient-based visual explanation method (akin to [48, 74, 76, 78]), termed point-specific activation map (PSAM), to visualize regions of interest and the activation levels of detected points. Using PSAM, we observe that the activation values and coverage areas of foreground pixels continuously grow during semi-supervised training, causing the model to misclassify neighboring pixels as foreground. + +Inspecting PSAMs reveals that each point in GT is implicitly associated with a local region rather than a single pixel in the prediction. Accordingly, we propose to replace the P2P with a point-to-region (P2R) strategy, which matches local regions to points in the pseudo-label. This allows all pixels within a matched region, including background pixels, to share the confidence of the corresponding pseudo-point. Experimental results and visualizations demonstrate that P2R can effectively train a semisupervised point-based counter and achieve expected performance. Experimental results show that P2R effectively trains semi-supervised point-based counters with comparable performance to P2P but at significantly reduced computational cost. Unlike P2P [57], P2R eliminates the need for the Hungarian algorithm by assigning each ground-truth or pseudo-point a local "dominated zone" and selecting a representative foreground pixel based on a predefined cost matrix. + +The contributions of our paper are as follows: + +- We formulate a pipeline of semi-supervised counting with a point-detection method, but find a problem leading the training breakdown. +- We propose Point-Specific Activation Map (PSAM) to + +visualize the regions of interest and the activation values of each point detected by the point-based crowd counter. Through PSAM, we observe that the surrounding pixels of each expected foreground pixel are over-activated, causing the decoder to recognizes these pixels as instances different from the concerned one. + +- Based on insights from PSAM, we propose replacing the P2P with a P2R matching strategy for semi-supervised learning in point-based crowd counting. This not only makes the training process work as expected, but also brings additional benefits of reducing computational requirements. +- The proposed P2R achieves outstanding performance in several experiments, including semi-supervised counting and unsupervised domain adaptation. Ablation study shows that P2R excellently tackles the problem caused by P2P in semi-supervised counting. + +# 2. Related Works + +Object counting in dense scenes. The development of deep learning has significantly influenced crowd counting. Prior to this, detecting body parts [16, 17] was the common pipeline for crowd understanding in computer vision, although some methods implemented counting models without explicit object segmentation or tracking information [4-6]. Beyond specific crowd scenarios, class-agnostic counting [14, 27, 30, 45, 49, 71] has also been developed to estimate the distribution of objects in the same category. Most recent counting methods implement neural networks following the density map regression scheme [3, 8, 20, 50, 60, 65, 67, 69, 73]. In addition to continuous innovations in model architecture design [3, 20, 24], new loss functions have also been developed, including Bayesian Loss [38], loss functions based on balanced [63] and unbalanced optimal transport [22, 39, 61], and characteristic function loss computed in the frequency domain [51, 52]. A study analyzes the relationship between current loss functions from the perspective of proximal mapping [28]. Some methods also perform crowd analysis via instance segmentation [1, 9] and point detection [21, 31, 47, 57]. The latter is the scheme adopted in our semi-supervised model. + +Semi-supervised counting. Semi-supervised counting has attracted broad attention recently [18, 19, 23, 25, 26, 32-34, 36, 42, 54]. L2R [32, 33] introduces a ranking rule for unlabeled crowd images based on the principle that cropped images should have smaller counts than full images. GP [54] iteratively updates pseudo-labels using Gaussian Processes. IRAST [34] and SUA [42] enhance models by incorporating a segmentation branch. DAC [23] uses learnable density agents to capture varying crowd densities. OT-M [26] transforms density maps into point maps as pseudo-labels, achieving strong performance with a density-to-point loss function [61]. Our method also em + +ploys pseudo-labeling but uses a point detector for counting, avoiding the need for density map transformations as in [26]. Additionally, uncertainty is directly measured by the scores of extracted points, without any additional learnable structures like those in [42]. + +Visual Explanation. Visualizing the importance of extracted image features is a straightforward way to interpret a model. The mainstream of visual explanation focuses on methods producing activation maps [13, 44, 48, 64, 74, 76-78], which are the products of feature maps from some intermediate layer and weight maps indicating the corresponding feature's importance. For example, GradCAM [13, 48] defines the weights using the corresponding gradient maps, generated via backpropagation. Beyond the focus on classification tasks, ODAM [74, 76] extends gradient-based CAM to object detection, visualizing the local impact of features on the detector's decisions. In this paper, we propose PSAM to visualize how the activation map changes for each predicted point in an ill-posed semi-supervised counting framework. Based on the observations from PSAM, we design a P2R matching strategy to avoid the over-activation of pixels in local regions, which effectively resolves the drawbacks of P2P and enables the successful training of a semi-supervised point detector. + +# 3. Point Counter in Semi-Supervised Counting + +This section formulates a simplified point-based counter by removing the regression branch in P2PNet [57], and then presents the training scheme and challenges when training with labeled and unlabeled data, respectively. + +# 3.1. Preliminary on A Simplified Point Counter + +Common point-based crowd counters like P2PNet [57] contain a classification and a regression branch. However, we find a simplification can be achieved by removing the offset regression branch. On one hand, the crowd count can be estimated by counting foreground pixels with scores greater than 0.5, a process that does not involve the regression branch. On the other hand, pixel coordinates alone are sufficient to locate pedestrians, as demonstrated in STEERER [10] and GL [61], which use the coordinates of local maxima to indicate pedestrians locations. + +As a result, P2PNet can be simplified to: + +$$ +\begin{array}{l} \mathcal {F} (I, \Theta_ {f}) \rightarrow F \in \mathbb {R} ^ {c \times h \times w}, (1) \\ \mathcal {D} (F, \Theta_ {d}) \rightarrow D \in \mathbb {R} ^ {h \times w} (2) \\ \rightarrow \mathcal {P} = \left\{\boldsymbol {p} \in \mathbb {R} ^ {n}, \boldsymbol {x} \in \mathbb {R} ^ {n \times 2} \right\}, (3) \\ \end{array} +$$ + +where $\mathcal{F}$ is the image encoder with learnable parameters $\Theta_f$ , and $\mathcal{D}$ denotes the density decoder with learnable parameter $\Theta_d$ . Notably, the offset regression branch is excluded in this simplified formulation. The pixel set $\pmb{p} \in \mathbb{R}^n$ + +represents the flattened values in $D$ , and its corresponding coordinate set is denoted as $\pmb{x} \in \mathbb{R}^{n \times 2}$ , where $n = h \times w$ . + +# 3.2. Training with Labeled Data via P2P Matching + +The GT with $m$ points can be denoted as $\pmb{x}^{\prime} \in \mathbb{R}^{m \times 2}$ , where $\pmb{x}_{[j]}^{\prime} \in \mathbb{R}^{2}$ represents the location of the $j$ -th pedestrian. Using $\mathcal{P}$ and $\pmb{x}^{\prime}$ , the loss computation involves two parts: P2P matching and binary cross entropy (BCE) calculation. + +The former aims to obtain a matrix-based solution $\mathbf{M} \in \{0,1\}^{n \times m}$ that minimizes the bipartite-graph matching cost between $\pmb{x}$ and $\pmb{x}'$ within the cost matrix $\mathbf{C} \in \mathbb{R}^{n \times m}$ : + +$$ +\mathbf {C} _ {[ i j ]} = \tau \| \boldsymbol {x} _ {[ i ]} - \boldsymbol {x} ^ {\prime} _ {[ j ]} \| _ {2} - S (\boldsymbol {p} _ {[ i ]}), \tag {4} +$$ + +where $S(\cdot)$ is defined as the inverse sigmoid function $S(p) = -\log (1 / p - 1)$ , rather than the identity operator used in the vanilla P2PNet [57], for improved performance. + +To compute BCE loss, the learning objective of $\pmb{p}$ is formulated as $\hat{\pmb{p}} = \mathbf{M}\mathbf{1}_m$ , and the BCE loss is calculated for optimization as follows: + +$$ +\mathcal {L} _ {l} = - \lambda \hat {\boldsymbol {p}} ^ {\top} \log \boldsymbol {p} - \left(\mathbf {1} _ {n} - \hat {\boldsymbol {p}}\right) ^ {\top} \log \left(\mathbf {1} _ {n} - \boldsymbol {p}\right) \tag {5} +$$ + +where $\lambda$ is a re-weighting factor for matched pixels [57]. + +# 3.3. Problem in Training with Unlabeled Image + +Here we consider training with unlabeled data following the standard pseudo-labeling procedure [15, 56]. As illustrated in Fig. 1, it requires the cooperation of a student and teacher model that share the same structure described in (1)~(3). The difference lies in that the teacher model is updated via exponential moving average (EMA) [59], while the student is optimized via gradient descending. During training, the teacher and student process the weakly and strongly data-augmented versions of an unlabeled image, producing $\mathcal{P}_t = \{\pmb{p}_t, \pmb{x}_t\}$ and $\mathcal{P}_s = \{\pmb{p}_s, \pmb{x}_s\}$ , respectively. After that, foreground points are extracted from $\mathcal{P}_t$ to construct pseudo-labels: + +$$ +\mathcal {P} _ {t} ^ {\prime} = \left\{\boldsymbol {p} _ {t} ^ {\prime}, \boldsymbol {x} _ {t} ^ {\prime} \right\} = \left\{\boldsymbol {p} _ {t [ j ]}, \boldsymbol {x} _ {t [ j ]} \mid \boldsymbol {p} _ {t [ j ]} > 0. 5 \right\}, \tag {6} +$$ + +where $\pmb{x}_t^\prime$ is treated as the pseudo hard label to estimate the matching matrix $\mathbf{M}_{st}$ via Hungarian algorithm, and the loss is obtained by substituting $\hat{\pmb{p}}_t = \mathbf{M}_{st}\mathbf{1}_m$ and $\pmb{p}_s$ into (5). + +However, some pseudo-labels may not be reliable. A common solution is to only consider points in $\mathcal{P}_t^\prime$ with high scores [15, 56], i.e., using a confidence vector $\zeta$ to mask out elements whose score is below a threshold $\eta$ : + +$$ +\zeta = \mathbb {1} \left(\boldsymbol {p} _ {t} ^ {\prime} > \eta\right) \xrightarrow {\text {m a p t o} \hat {\boldsymbol {p}} _ {t} \text {b y} \mathbf {M} _ {s t}} \boldsymbol {z} = \mathbf {M} _ {s t} \zeta . \tag {7} +$$ + +Note $\eta$ is greater than 0.5 and (7) only let pseudo points with score greater than $\eta$ be reliable. Subsequently, $\mathbf{Z} = \mathrm{diag}(z)$ is incorporated into the BCE loss for supervision: + +$$ +\mathcal {L} _ {u} = - \lambda \hat {\boldsymbol {p}} _ {t} ^ {\top} \mathbf {Z} \log \boldsymbol {p} _ {s} - \left(\mathbf {1} _ {n} - \hat {\boldsymbol {p}} _ {t}\right) ^ {\top} \mathbf {Z} \log \left(\mathbf {1} _ {n} - \boldsymbol {p} _ {s}\right). \tag {8} +$$ + +However, the second item of (8) in P2P is identically zero: + +$$ +\mathcal {L} _ {u [ P 2 P ]} = - \lambda \hat {\boldsymbol {p}} _ {t} ^ {\top} \mathbf {Z} \log \boldsymbol {p} _ {s} - 0, \tag {9} +$$ + +regardless of $\zeta$ (embedded in $\mathbf{Z}$ ). Consequently, no gradients are back-propagated from the losses associated with background pixels, which is ill-posed. Utilizing (9) as the loss function is similar to training a two-class classifier with only samples from the category "1", which will lead the trained model to output "1" regardless of the input. + +# 4. Visual Explanation with PSAM + +In addition to the theoretical analysis, we implement a framework that supervises a model using labeled and unlabeled data via (5) and (9), respectively, to empirically investigate the behavior of P2P-based semi-supervised counting. Subsequently, PSAM is designed to visualize how the prediction is influenced by (9) during training. Thus, we can design an effective confidence scheme accordingly to counteract the effects revealed by PSAM, thereby enabling semi-supervised learning to function as expected. + +# 4.1. PSAM: Point-Specific Activation Map + +An activation map is a heat map with the same spatial resolution as the image features, where local regions with relatively high values are interpreted as being discriminative (informative) for the current output [48, 74, 78]. Most techniques obtain activation maps by linearly aggregating features with their corresponding gradients, as [53, 58] have shown that the partial derivative with respect to a specific element can reflect its contribution to the final prediction. + +Unlike [48, 78], which compute gradients at the image level, crowd counting requires a point-specific activation map (PSAM) to visually explain the regions of interest for every element in $\pmb{p}$ of (3). Thus, the Jacobian matrix of $\pmb{p}$ with respect to image feature $F$ is required: + +$$ +\mathbf {J} _ {\mathcal {D}} = \frac {\partial \boldsymbol {p}}{\partial F} = \left[ \nabla \boldsymbol {p} _ {[ 1 ]}, \nabla \boldsymbol {p} _ {[ 2 ]}, \dots , \nabla \boldsymbol {p} _ {[ n ]} \right] ^ {\top}. \tag {10} +$$ + +Afterwards, the $q$ -th heat map $H_{[q]}$ in PSAM is derived by filtering out negative influences on the aggregated feature: + +$$ +H _ {[ q ]} = \max \left(\sum_ {k = 1} ^ {c} \nabla \boldsymbol {p} _ {[ q ]} \odot F, 0\right) \in \mathbb {R} ^ {h \times w}, \tag {11} +$$ + +where the summation operator is applied along the channel dimension, and $\odot$ denotes the Hadamard product. + +In a naive implementation, the computation of $\mathbf{J}_{\mathcal{D}} \in \mathbb{R}^{n \times hwc}$ requires a large memory and involves $n$ backpropagations through the computational graph, posing a challenge for visual explanation in pixel-level tasks. However, this issue can be resolved by computing gradients only between $\pmb{p}_{[q]}$ and features within its receptive field. Gradients in other areas are zero, as there is no path connecting them in the computational graph. Thus, an efficient approach can be implemented, as illustrated in Fig. 2. + +![](images/4d58de3567dcc2d6d9e22e47fb4788cd6add94e05ed944dc3cff47ef66ecd303.jpg) +Figure 2. The generation process of PSAM. + +In this implementation, the forward process begins with extracting sliding local blocks $\tilde{F}$ according to the receptive field $r$ of $\mathcal{D}$ , and then decoding $\tilde{F}$ through (2) in parallel. The final prediction $\pmb{p}$ is generated by stacking pixels located at the center of $r\times r$ patches denoted as in $\tilde{D}$ : + +$$ +F \in \mathbb {R} ^ {h \times w \times c} \rightarrow \tilde {F} \in \mathbb {R} ^ {(h w) \times (r \times r \times c)}, \tag {12} +$$ + +$$ +\mathcal {D} (\tilde {F}, \Theta_ {d}) \rightarrow \tilde {D} \in \mathbb {R} ^ {n \times (r \times r)} \xrightarrow {\text {c e n t e r o f p a t c h e s}} \boldsymbol {p} \in \mathbb {R} ^ {n}. \tag {13} +$$ + +The remaining $r^2 - 1$ pixels in each patch are discarded. + +Denoting the sum of the output as $\xi = \mathbf{1}_n^\top \pmb{p}$ , the gradient $\nabla \xi = \partial \xi / \partial \tilde{F}$ can be obtained via just one backpropagation, and the $q$ -th gradient block, $\nabla \xi_{[q]}$ , is equal to the gradient of $\pmb{p}_{[q]}$ with respect to $\tilde{F}_{[q]}$ : + +$$ +\nabla \xi_ {[ q ]} = \frac {\partial \xi}{\partial \tilde {F} _ {[ q ]}} = \frac {\partial \xi}{\partial \boldsymbol {p} _ {[ q ]}} \cdot \frac {\partial \boldsymbol {p} _ {[ q ]}}{\partial \tilde {F} _ {[ q ]}} = \frac {\partial \boldsymbol {p} _ {[ q ]}}{\partial \tilde {F} _ {[ q ]}}, \tag {14} +$$ + +since $\partial \xi /\partial \pmb{p}_{[q]} = 1$ . After that, we can parallel compute the PSAM (denoted as $\tilde{H}_{[q]}$ ) within the receptive field of $\pmb{p}_{[q]}$ by substituting (14) into the formulation presented in (11): + +$$ +\tilde {H} _ {[ q ]} = \max \left(\sum_ {k = 1} ^ {c} \nabla \xi_ {[ q ]} \odot \tilde {F} _ {[ q ]}, 0\right). \tag {15} +$$ + +As illustrated in the bottom of Fig. 2, the complete PSAM $H_{[1,\dots,n]}$ in (11) can be obtained by filling $\tilde{H}_{[1,\dots,n]}$ back into the corresponding regions of a fully zero matrix: + +$$ +H _ {[ q, t ]} = \left\{ \begin{array}{l l} \tilde {H} _ {[ q, t ^ {\prime} ]}, & \text {i f} t \in \Omega_ {q} \\ 0, & \text {o t h e r w i s e} \end{array} , \right. \tag {16} +$$ + +where $\Omega_q$ represents the index set of $p_{[q]}$ 's receptive field, and $t' \in \mathbb{R}^2$ is the transformed coordinate of $t \in \mathbb{R}^2$ in $H$ , ensuring $F_{[t]} \equiv \tilde{F}_{[q,t']}.$ In this way, the matrix memory required to obtain all $H_{[q]}$ is reduced from $\mathbb{R}^{n \times hwc}$ to $\mathbb{R}^{n \times r^2c}$ ( $r^2 \ll hw$ ), and only one back-propagation is needed to obtain PSAMs for all pixels. + +![](images/173057d1eb0dda57d364b4a3e1c373941501cb4d7e3c2b4a66c083925e329cff.jpg) +(a) MAE/MSE during training + +![](images/233930473ae392bc74e00fbda96b9d0a8622f2a75a8bf6b31af8ddc59b84c335.jpg) +(c) Local PSAM mean (model-L) + +![](images/f7fd1d069acbc0bce15ae0e9894c45afc251c70ce9db4ce699c4d0ec496df5d8.jpg) +(e) aggregated PSAM (model-L) + +![](images/f4567a78e79a560022aa540449d250304ca3e352ffa4b7367bac20ea4e1bc0b9.jpg) +(b) Sorted average PSAM + +![](images/919694c90acae328e6ba18491b9f7b880869903d74fdf7cb9d57c6faaf6c3a68.jpg) +(d) Local PSAM mean (model-U) +Figure 3. Observations in PSAM. (a) The training process, where model-L and model-U are extracted from the 100th and 200th epochs, respectively. (b) Comparing sorted values of PSAMs in the two models. (c) & (d) Visualizing the average of local PSAM, and (e) & (f) the aggregated PSAM to compare model-L and model-U from a global perspective. + +![](images/66416986b940cbb75c8bbbfa470ac1c4a4a42040afda78c7030ca3acfbbf1fdd.jpg) +(f) aggregated PSAM (model-U) + +# 4.2. Observation in PSAM + +We next apply PSAM to visualize the model during semi-supervised counting. The semi-supervised loss function combines (5) and (8): + +$$ +\mathcal {L} = \alpha \mathcal {L} _ {u} + (1 - \alpha) \mathcal {L} _ {l}, \tag {17} +$$ + +where $\alpha$ is a hyperparameter to balance the training of labeled and unlabeled data. Following [15], we set $\alpha = 0$ during the first 100 epochs to generate valid pseudo labels. After that, $\alpha$ is gradually increased to 1. In Fig. 3(a), we plot how the MAE/MSE changes on the validation set during training. The training failed as evinced by both metrics increasing significantly when unlabeled data were involved in training (after epoch 100). We denote 'model-L' as the model at the 100-th epoch that is trained only on labeled data, and 'model-U' as the model at the 200-th epoch that was trained with both labeled and unlabeled data. + +We next compare explanations of model-L and model-U using PSAM. We visualize the concerned regions of the crowd by extracting the PSAMs corresponding to the foreground pixels $(\pmb{p}_{[q]} > 0.5)$ . In Fig. 3(c & d) the element-wise mean of the PSAM patch $(\sum_{q} \tilde{H}_{[q]} \cdot (\pmb{p}_{[q]} > 0.5))$ is displayed to illustrate how the area of the concerned regions changes between model-L and model-U, while Fig. 3(b) presents a detailed value comparison using a line chart. On one hand, Fig. 3(b) shows that the values in model-U's PSAM are larger than those of model-L; on the other hand, Fig. 3(c & d) shows that model-L's PSAM is more concentrated when compared to model-U. These observations reveal how (8) in P2P scheme affects the trained model: $\mathcal{L}_u$ leads to over-activation in the PSAM of foreground pixels, causing the decoder $\mathcal{D}$ to falsely identify surrounded pixels as foreground pixels as well. + +Next, we visualize the global PSAM of both models + +by compressing $H_{[q]}$ corresponding to foreground pixels: $\bar{H} = \sum_{q=1}^{n} (\pmb{p}_{[q]} > 0.5) \cdot H_{[q]}$ . As displayed in Fig. 3(e & f), the aggregated PSAM of model-L is sparser than that of model-U, which is consistent with the observation for individual points presented in Fig. 3(c & d). Another observation is that the background regions do not change significantly, indicating that almost all false positives occur in the neighborhood of the initial foreground pixels. This suggests that $\mathcal{L}_u$ does not alter the hyperplane separating crowd and non-crowd features in the high-dimensional space. + +# 5. P2R Matching + +The observations described in last section reveal that $\mathcal{L}_u$ over-activates the values and enlarges the area of the PSAM corresponding to foreground pixels, which should be suppressed by the supervision of background part in (5), but is actually zero in (8) ( $\mathcal{L}_{u[\mathrm{P2P}]}^{\prime}$ in (9)). This over-activation also confirms two important functions of the background item in (5): it not only guides the model to classify the crowd and non-crowd regions in the image, but also suppresses the surrounding person region from being activated. The latter is important since the neighboring pixels of a detected foreground pixel have similar features to the foreground pixel due to the CNN architecture, and thus may become activated if suppression is not applied. This interpretation makes more sense because each person in a crowd image is represented by tens to thousands of pixels, not just one. Supervising the region around the positive pixel with labels of 0 is crucial to a point-based counter. Therefore, during training, it would be more effective to adopt a point-to-region (P2R) matching strategy instead of a point-to-point (P2P) approach. Moreover, as suggested by the above interpretation, P2R is not limited in training with unlabeled data, but can also work in supervised learning with point supervision. + +# 5.1. Training with Labeled Data via P2R + +Recalling the notation of the prediction $\mathcal{P} = \{\pmb {p},\pmb {x}\}$ in (3) and ground truth $\pmb{x}^{\prime}$ , the matching matrix $\mathbf{M}$ in our P2R matching is the Hadamard product of two items: + +$$ +\mathbf {M} = \mathbf {M} _ {f} \odot \left(\boldsymbol {\beta} ^ {\top} \mathbf {1} _ {m}\right) \quad \in \mathbf {M} ^ {n \times m}. \tag {18} +$$ + +The first item $\mathbf{M}_f\in \{0,1\}^{n\times m}$ is a many-to-one matrix that assigns each pixel in $\mathcal{P}$ to its nearest point in $x^{\prime}$ : + +$$ +\mathbf {M} _ {f [ i j ]} = \left\{ \begin{array}{l l} 1, & \text {i f} l _ {2} (i, j) < l _ {2} (i, k) \forall k \neq j \\ 0, & \text {o t h e r w i s e} \end{array} , \right. \tag {19} +$$ + +where $l_{2}(i,j) = \| \pmb{x}_{[i]} - \pmb{x}^{\prime}_{[j]}\|_{2}$ is the $l_{2}$ -distance between the $i$ -th pixel's and the $j$ -th point's coordinates. $\beta \in \{0,1\}^{n}$ in the second item of (18) is a vector marking foreground + +![](images/2d3037d526dbec96c083a70d5cfaa3c38f5a7a8f5ba86833d86b4c558aeb5176.jpg) +(a) P2P scheme + +![](images/5a89ef1fd0767b95f729c31ee59a7693f87f94e9c88ef75a51e6879443359794.jpg) +(b) matched point + +![](images/6e9ca855a9338a9e281301059c52e330acd103eea4f4505e3158c10f3895bced.jpg) +(c) semi-supervised points + +![](images/d52126ae431bda47c3298a8fe18f63c4f40d36ec9102cd1bf43b1fdfb4460d0d.jpg) +(d) P2R scheme +Figure 4. Difference between P2P and P2R matching. (a) and (d) demonstrate an overall difference. P2P focuses only on foreground pixels in $\mathbf{P}_s$ , while P2R considers background pixels as well. (b) and (e) show how the matching is performed in P2P and P2R, respectively. P2R segments out local regions for each pseudo-label, whereas P2P only detects one point. (c) and (f) illustrate how untrusted predictions are filtered. P2P retains only foreground pixels for loss computation, while P2R also keeps pixels in the neighborhoods of their corresponding pseudo points. + +![](images/50797048b980f15b33ddf80a206c725012f898e4239eb929921c4e3236208965.jpg) +(e) matched region + +![](images/b4774abf9ee95739c14124767740efce4801d5d91c06433a4fb03a77b5d4168b.jpg) +(f) semi-supervised regions + +regions, by setting those pixels in $\mathcal{P}$ far from any points in $x^{\prime}$ as zero: + +$$ +\beta_ {[ i ]} = \left\{ \begin{array}{l l} 1, & \text {i f} \min _ {j} l _ {2} (i, j) < \mu \\ 0, & \text {o t h e r w i s e} \end{array} , \right. \tag {20} +$$ + +in which $\mu$ is a hyper parameter indicating the radius if the neighborhood of each GT point is considered as a circle. As displayed in Fig. 4(e), $\mathbf{M}$ segments out a coarse local region that is consistent with the location of the GT point. After that, the learning objective $\hat{\pmb{p}}$ is derived by defining the cost matrix $\mathbf{C} \in \mathbb{R}^{n \times m}$ : + +$$ +\mathbf {C} _ {[ i j ]} = \left\{ \begin{array}{l l} \tau l _ {2} (i, j) - \mathcal {S} (\boldsymbol {p} _ {[ i ]}), & \text {i f} \mathbf {M} _ {[ i j ]} = 1 \\ \infty , & \text {o t h e r w i s e} \end{array} . \right. \tag {21} +$$ + +Then, the pixel with the minimum cost in each column of $\mathbf{C}$ is marked as the potential foreground pixel: + +$$ +\hat {\mathbf {M}} _ {[ i j ]} = \left\{ \begin{array}{l l} 1, & \text {i f} \mathbf {C} _ {[ i j ]} < \mathbf {C} _ {[ k j ]} \forall k \neq i \\ 0, & \text {o t h e r w i s e} \end{array} . \right. \tag {22} +$$ + +Finally, the learning objective $\hat{\pmb{p}}$ is derived as $\hat{\pmb{p}} = \hat{\mathbf{M}}\mathbf{1}$ , and the loss $\mathcal{L}_l$ is computed between $\pmb{p}$ and $\hat{\pmb{p}}$ according to BCE loss formulated as (5) during training with labeled data. + +An extra benefit is that computing $\mathbf{M}$ and $\hat{\mathbf{M}}$ in P2R requires only finding the minimum of each row or column in the matrices, which is significantly faster than executing the Hungarian algorithm — a time-consuming component of the P2P matching strategy [57]. + +# 5.2. Training with Unlabeled Data via P2R + +Recalling the pseudo-label $\mathcal{P}_t^\prime = \{\pmb {p}_t^\prime ,\pmb{x}_t^\prime \}$ from the teacher and the supervised student output as $\mathcal{P}_s = \{\pmb {p}_s,\pmb {x}_s\}$ , the + +matching matrix $\mathbf{M}_{st}$ between $\mathcal{P}_s$ and $\mathcal{P}_t^\prime$ is computed via (18) in P2R, and then combined with the confidence in (7) to mask unreliable pseudo labels. Additionally, pixels far from any pseudo points can be directly marked as reliable background pixels, as suggested by observations from the aggregated PSAM in Fig. 3(e & f). Thus, the confidence diagonal matrix $\mathbf{Z}$ is formulated as follows with $\zeta = \mathbb{1}(\hat{\pmb{p}}_t' > \eta)$ : + +$$ +\mathbf {Z} = \operatorname {d i a g} \left[ \mathbf {M} _ {s t} \boldsymbol {\zeta} + \left(\mathbf {1} _ {n} - \boldsymbol {\beta}\right) \right] \tag {23} +$$ + +The final loss supervising unlabeled data $\mathcal{P}_s$ with pseudo labels is derived by substituting $\mathbf{Z}$ in (23) into (8). + +As displayed in Fig. 4(f), the confidence for each pseudo point is propagated to it corresponding region. On one hand, reliable foreground parts (blue) contain a pixel whose pseudo-label is 1, while the pseudo-labels of its surrounding pixels are marked as 0. On the other hand, all pixels in the reliable background parts (khaki) are assigned pseudolabels of 0 to train the counter with unlabeled data. The uncolored regions are ignored because their confidence is 0. + +The pseudo code and flow chart of the proposed P2R loss can be found in the supplemental material. + +# 6. Experiments + +This section explores how the proposed P2R training approach enhances a point-based counting model using both labeled and unlabeled data across three parts: semi-supervised crowd counting, unsupervised domain adaptation (UDA) for crowd counting, and ablation studies. + +In the semi-supervised experiments, our method is implemented on four crowd counting datasets: ShTech A/B [73], UCF-QNRF [12], and JHU++ [55]. Following DAC [67] and OT-M [26], three protocols are applied: $5\%$ , $10\%$ , and $40\%$ of labeled data, with the remaining crowd images involved in training without annotations. In the UDA part, labeled data from one domain, e.g., ShTech A, is used to initialize a counting model, while unlabeled data from another domain is utilized to capture the domain characteristics of the application scenes. In the ablation study, we test different hyperparameters to demonstrate their impact on the performance of P2R. + +# 6.1. Semi-Supervised Counting + +Tab. 1 presents the experimental results of semi-supervised counting. Our method achieves the best performance in nearly all protocols across all crowd datasets. Besides, using $5\%$ data with P2R surpasses the fully-supervised learning with $10\%$ data and is even equivalent to other semi-supervised learning methods using $10\%$ data. Similar performances are also observed between the semi-supervised training with $10\%$ and $40\%$ labeled data. These results demonstrate that P2R is label-efficient and advantageous in semi-supervised learning. In Fig. 5, we visualize predictions of DAC [67], OTM [26], and our P2R. The 3rd row is + +
Label Pct.MethodsShTech A [73]ShTech B [73]UCF-QNRF [12]JHU++ [55]
MAEMSEMAEMSEMAEMSEMAEMSE
5%label-only93.7155.213.124.0132.7231.7104.1383.5
MT [59]104.7156.919.333.2172.4284.9101.5363.5
L2R [32]103.0155.420.327.6160.1272.3101.4338.8
GP [54]102.0172.015.727.9160.0275.098.9355.7
DAC [23]85.4134.512.622.8120.2209.382.2294.9
OT-M [26]83.7133.312.621.5118.4195.482.7304.5
P2R (ours)69.9119.59.116.6100.1182.577.8293.5
10%label-only84.0138.310.719.2112.4186.878.7305.5
MT [59]319.394.515.624.5156.1245.5250.390.2
L2R [32]90.3153.515.624.4148.9249.887.5315.3
IRAST [34]86.9148.914.722.9135.6233.486.7303.4
DAC [23]74.9115.511.119.1109.0187.275.9282.3
OT-M [26]80.1118.510.818.2113.1186.773.0280.6
CU [18]70.7116.69.717.7104.01644.274.8281.6
P2R (ours)65.2114.68.414.594.9167.268.7272.3
40%label-only64.5105.68.114.099.2174.768.8283.5
MT [59]88.2151.115.925.7147.2249.6121.5388.9
L2R [32]86.5148.216.825.1145.1256.1123.6376.1
SUA [42]68.5121.914.120.6130.3226.380.7290.8
DAC [23]67.5110.79.614.691.1153.465.1260.0
OT-M [26]70.7114.58.113.1100.6167.672.1272.0
P2R (ours)55.695.06.811.086.0144.363.3271.1
100%P2PNet [57]52.7485.066.259.9085.32154.5061.25258.65
P2R (ours)51.0279.686.179.8483.26138.1158.83253.10
+ +a failure case of P2R due to crowd motion blur. However, the overall performance of P2R is best. + +The last row of Tab. 1 also compares P2P and P2R under a fully-supervised learning scheme (100% labels); P2R performs better than P2PNet on all crowd counting datasets. Additionally, as described in Section 5.1, P2R is faster than P2P due to the absence of the Hungarian algorithm. We compared the efficiency using an image with a resolution of $576 \times 960$ and 775 annotated points — P2P requires an average of 0.4307 seconds for loss computation, while P2R only needs 0.0064 seconds, which is nearly 68 times faster than P2P. These comparisons demonstrate that P2R surpasses P2P in both effectiveness and efficiency. + +# 6.2. Unsupervised Domain Adaptation (UDA) + +UDA aims to transfer knowledge learned from a source domain to a target domain by designing modules or learning objectives that capture domain-agnostic features [11, 29, 43, 65, 70, 72]. Unlike semi-supervised learning, where labeled and unlabeled data originate from the same domain, UDA typically involves a domain gap between the labeled (source domain) and unlabeled crowd data (target domain). This gap often manifests as differences in crowd distribution, density levels, and perspectives. Our P2R can also be applied to UDA. When an unlabeled image from the target domain is provided, the threshold automatically partitions its corresponding pseudo points into two categories: higher-score $(> \eta)$ and lower-score $(< \eta)$ . The higher-score class is considered to be closer to the source domain, while the lower-score class is considered to be farther from the source domain. Thus, the higher-score pseudo points can participate in the training to gradually guide the model to learn to count in the target domain. + +![](images/d25312d39243a6b6634ea424acc16f50e830886b2550e10fa04726d36a4ba1ad.jpg) + +![](images/8e4bca1ce1a802d635973ce4b53dc29485be1ceabda551efa9d41edfb3489c10.jpg) + +![](images/498debb671b3a4edee2dceac2f04c8a36334121c31d2a059ebc01fe31e84e8ff.jpg) +Figure 5. qualitative comparison with other models. + +ground truth: 61 + +![](images/1254c483cece1cfc09f2ac61fdd5d89d1a529df00416c2721d340b5daa89f315.jpg) + +![](images/94a12bc5c589ddd0a8f7af83e499663ce4398dcbea20e6710ee1212e237c7c05.jpg) +DAC [23]: 49.21 + +![](images/9650b28896edc859802b98431ddb2a87918b2f43ebe30de7e800abc36ac7d799.jpg) + +![](images/361c13ec83e61e4f11807d326fba288c08177559bbfdefc61ad9a6457f522c39.jpg) +OTM [26]: 51.43 + +![](images/1ab0ea4ddece539f20bd8786803703c3bf75987758b349dfb9d6ae8c9b6f23e2.jpg) + +![](images/1aa919225f6be1983959b821c70c72260f1e5fc7cd54f8da446701be6a0d5191.jpg) +P2R (ours): 60.0 + +ground truth: 61 + +![](images/4299101a1c5146f624ae2d673bd7bb7b8f9db033834b5a9418e45b907948dc6d.jpg) +DAC [23]: 146.09 + +![](images/e81e68ee2a8eaf4974274886e9d870487d019eb84db173bb908cb81c627302f0.jpg) +OTM [26]: 126.11 + +![](images/8be1375a8f5baece94489667dadbd13ab3a9b4e5a45376da504be7396a81025a.jpg) +P2R (ours): 227.0 + +Table 1. Comparison with other recent methods on four benchmark datasets under different labeled protocols. + +
METHODSA → BA → QB → AB → Q
MAEMSEMAEMSEMAEMSEMAEMSE
DGDCCUS [40]12.624.6119.4216.6121.8203.1179.1316.2
MPCount [43]11.419.7115.7199.899.6182.9165.6290.4
UDABL [38] (w/o DA)15.925.8166.7287.6138.1228.1226.4411.0
RBT [35]13.429.3175.0294.8112.2218.2211.3381.9
C2MoT [70]12.421.1125.7218.3120.7192.0198.9368.0
FGFD [79]12.723.3124.1242.0123.5210.7209.7384.7
FSIM [80]11.119.3105.3191.1120.3202.6194.9324.5
UDAP2R (w/o DA)25.635.7155.1286.8130.2229.1173.3329.8
P2R (w/ DA)10.618.7105.3194.889.3176.0139.5243.1
+ +Table 2. Unsupervised domain adaptation for crowd counting. The LHS and RHS of “ $\rightarrow$ ” represent adaptation direction. Q, A, and B indicate UCF-QNRF [12], ShTech A, and ShTech B [73]. $DG$ indicates domain generalization. The training of UDA requires samples from the target domain, whereas DG does not. + +We conduct experiments on four protocols following [70]. The results are presented in Tab. 2. The last two rows demonstrate the results of the proposed P2R. The poor performance obtained without domain adaptation highlights the domain gap between the source data and target data. However, when the unlabeled data from the target domain is involved in training, the estimation errors are significantly reduced, and the results surpass previous methods with complex network structures for domain adaptation. These advanced experimental results demonstrate that the threshold works as expected to capture samples close to the source domain and gradually teaches the model to count in the target domain. + +# 6.3. Ablation Study + +We next conduct ablation studies to investigate the influence of hyperparameters in P2R, including $\tau$ in (21), the threshold $\eta$ , $\mu$ in (20), and $\alpha$ in (17). The results are in Fig. 6 + +The impact of $\tau$ in (21). $\tau$ is a hyperparameter to balance the distance and foreground score when selecting the foreground pixel from the matched region. When $\tau \rightarrow \infty$ , the predicted score is ignored, which is equivalent to computing cross-entropy between the prediction and the ground truth point map. If $\tau = 0$ , the pixel with the highest score is chosen as the foreground in $\hat{\pmb{p}}$ . However, Fig. 6(a) shows that neither extreme is optimal. The best performance is achieved when $\tau = 8$ . + +![](images/5d5fe63da2bf1fddb316df5f7d8e3c7e7678febfa5a7b3271468130cdbee91bd.jpg) +(a) errors vs. $\tau$ + +![](images/c9c7fe401fb6d177465c000264ef0308194a1cb1d7e70f6fe4536c9358ad99b4.jpg) +(c) errors vs. $\mu$ +Figure 6. The estimation errors vs. various hyperparameters. (a) is conducted on the fully-supervised ShTech A since it is irrelevant to semi-supervised counting, while (b)-(d) are conducted with $5\%$ labeled data in ShTech A. + +![](images/1490810d57fd7ef78bcffb49f22d8a9721c65052b843f722653ce7a89f89bfc2.jpg) + +![](images/0072ff971013eced25b31451a39f06e40e677af0efbb691d4afd072e6a0db34a.jpg) +(b) errors vs. $\eta$ +(d) errors vs. $\frac{\alpha}{1 - \alpha}$ + +The effect of threshold $\eta$ . $\eta$ defines the threshold to distinguish reliable and unreliable regions. Since all pseudo points are selected by picking pixels whose score is greater than 0.5 in the prediction, the threshold should also be greater than 0.5. Moreover, a lower $\eta$ introduces unreliable pseudo-labels into training, while a higher $\eta$ cannot fully utilize the unlabeled data as most parts cannot participate in training. As shown in Fig. 6(b), the estimation errors exhibit a trend of first decreasing and then increasing within the range of 0.6 to 0.9, with the lowest errors occurring when the threshold is set to 0.7. + +The radius of each point $\mu$ . The vector $\beta$ in (20) derived from $\mu$ marks the foreground points' neighborhood and background regions. It not only restricts the positions where foreground pixels appear by substituting $\beta$ into (18), but also directly identifies reliable background parts via the second item in (23) based on the observation of PSAM in Section 4.2. Fig. 6(c) presents how $\mu$ affects the counting performance. It shows the MAE remains around 68 when $\mu$ increases from 32 to 256, indicating that a larger $\mu$ does not significantly affect the results. This is consistent with the observation in PSAM since pixels far from the foreground points are less likely to be activated during training. + +The weight of unlabeled loss. Fig. 6(d) presents how $\alpha$ in (17) affects the performance. A large $\alpha$ focuses more on unlabeled data, leading to training failure, while a small $\alpha$ may cause the model to overfit on labeled data. In the experiments with ShTech A and $5\%$ labeled data, $\frac{\alpha}{1 - \alpha} = 2$ , i.e., $\alpha = 2/3$ , results in the best performance. Additionally, as long as $\alpha$ is not 0, the point-based counter can benefit from unlabeled data, as the maximum MAE/MSE in Fig. 6(d) are smaller than 80/150, which is much better than the model trained with only $5\%$ labeled data (MAE: 93.7, MSE: 155.2). + +![](images/bc9d5517a41f3760c947c215bfb6cc4dc5fb4769bbcd349a78272a90d6a583e2.jpg) +(a) sorted average PSAM + +![](images/18c3907376122314226fbf8f929c8556edd5c45d55abd39e848e55802195a990.jpg) + +![](images/0171d89ef65055b33551da679ddf963d531fbdc140881281682d438984b1449f.jpg) + +![](images/b4bec7e249dfc042fb1836bbfc195dd36705f4f2e00267923a0339ba8e0217ba.jpg) +(b) Local & aggregated PSAM +Figure 7. The comparison of PSAMs among different training schemes. Model-L (green) is trained with only labeled data; model-U_P2P (blue) is trained with both labeled data and unlabeled data using the P2P scheme, while model-U_P2R (orange) is trained with the proposed P2R scheme. + +PSAM comparison. Fig. 7 visualizes the PSAMs of models trained with the P2P and P2R schemes. Comparing model-L (green) and model-U_P2R (blue), the latter effectively restricts the PSAM to the neighborhood of the concerned foreground pixel, and its aggregated PSAM is also sparser than that of model-L, demonstrating the effectiveness of P2R. Comparing model-U_P2P and model-U_P2R in Fig. 7(a), the values of the PSAM are similar at the non-peak indexes, but the latter has a higher activation at the peak position compared to the former, indicating that semi-supervised learning further enhances the identification ability of the point-based counter. + +# 7. Conclusion + +This paper presents a P2R scheme to train a point-based counter for semi-supervised crowd counting. At the beginning, we establish a pseudo-labeling framework based on P2P matching, but its formulation is ill-posed since the confidence is only applicable to foreground pixels and cannot be propagated to backgrounds. To observe how the foreground and background feature are learned, we propose PSAM to visually interpret the concerned region and activation value of all foreground pixels. From the visualized activation map, it is observed that the neighborhood of each foreground pixel is over-activated and falsely recognized as instances by the decoder. Based on this observation, we replace P2P with P2R by segmenting the prediction into multiple regions containing the corresponding instances. Thus, pixels in the neighborhood of the concerned pseudo points can share the weights with the corresponding pixels. Additionally, the time-consuming Hungarian algorithm is no longer necessary in P2R. Several outstanding experimental results in semi-supervised counting and unsupervised domain adaptation demonstrate the advantages of our P2R matching strategy. + +Acknowledgements. This work was supported by a grant from City University of Hong Kong (Project No. 7030010). + +# References + +[1] Shahira Abousamra, Minh Hoai, Dimitris Samaras, and Chao Chen. Localization in the crowd with topological constraints. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 872-881, 2021. 2 +[2] Areej Althali, Amal Balabid, Reem Alharthi, Bander Alzahrani, Reem Alotaibi, and Ahmed Barnawi. Anomalous event detection and localization in dense crowd scenes. Multimedia Tools and Applications, 82(10):15673-15694, 2023. 1 +[3] Deepak Babu Sam, Shiv Surya, and R Venkatesh Babu. Switching convolutional neural network for crowd counting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5744-5752, 2017. 2 +[4] Antoni B Chan and Nuno Vasconcelos. Bayesianoisson regression for crowd counting. In 2009 IEEE 12th international conference on computer vision, pages 545-551. IEEE, 2009. 1, 2 +[5] Antoni B Chan and Nuno Vasconcelos. Counting people with low-level features and bayesian regression. IEEE Transactions on image processing, 21(4):2160-2177, 2011. 1 +[6] Antoni B Chan, Zhang-Sheng John Liang, and Nuno Vasconcelos. Privacy preserving crowd monitoring: Counting people without people models or tracking. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-7. IEEE, 2008. 2 +[7] Mulin Chen, Qi Wang, and Xuelong Li. Anchor-based group detection in crowd scenes. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1378-1382. IEEE, 2017. 1 +[8] Zhi-Qi Cheng, Qi Dai, Hong Li, Jingkuan Song, Xiao Wu, and Alexander G Hauptmann. Rethinking spatial invariance of convolutional networks for object counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19638-19648, 2022. 2 +[9] Junyu Gao, Tao Han, Qi Wang, Yuan Yuan, and Xuelong Li. Learning independent instance maps for crowd localization. arXiv preprint arXiv:2012.04164, 2020. 2 +[10] Tao Han, Lei Bai, Lingbo Liu, and Wanli Ouyang. Steerer: Resolving scale variations for counting and localization via selective inheritance learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21848-21859, 2023. 1, 3 +[11] Yuhang He, Zhiheng Ma, Xing Wei, Xiaopeng Hong, Wei Ke, and Yihong Gong. Error-aware density isomorphism reconstruction for unsupervised cross-domain crowd counting. In Proceedings of the AAAI conference on artificial intelligence, pages 1540-1548, 2021. 7 +[12] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 6, 7 +[13] Brian Kenji Iwana, Ryohei Kuroki, and Seiichi Uchida. Explaining convolutional neural networks using softmax gradient layer-wise relevance propagation. In 2019 IEEE/CVF + +International Conference on Computer Vision Workshop (ICCVW), pages 4176-4185. IEEE, 2019. 3 +[14] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 2 +[15] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2, 3, 5 +[16] Bastian Leibe, Edgar Seemann, and Bernt Schiele. Pedestrian detection in crowded scenes. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 878-885. IEEE, 2005. 2 +[17] Victor Lempitsky and Andrew Zisserman. Learning to count objects in images. Advances in neural information processing systems, 23, 2010. 2 +[18] Chen Li, Xiaoling Hu, Shahira Abousamra, and Chao Chen. Calibrating uncertainty for semi-supervised crowd counting. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16685-16695. IEEE, 2023. 2, 7 +[19] Hanxiao Li, Yonghong Song, and Tong Geng. Semi-supervised crowd counting based on hard pseudo-labels. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2024. 2 +[20] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1091-1100, 2018. 2 +[21] Dingkang Liang, Wei Xu, and Xiang Bai. An end-to-end transformer model for crowd localization. In ECCV, pages 38-54. Springer, 2022. 1, 2 +[22] Hui Lin, Xiaopeng Hong, Zhiheng Ma, Xing Wei, Yunfeng Qiu, Yaowei Wang, and Yihong Gong. Direct measure matching for crowd counting. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages = 837–844, publisher = ijcai.org, year = 2021, .2 +[23] Hui Lin, Zhiheng Ma, Xiaopeng Hong, Yaowei Wang, and Zhou Su. Semi-supervised crowd counting via density agency. In ACM Multimedia, 2022. 2, 7 +[24] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19628-19637, 2022. 1, 2 +[25] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, Zhou Su, Xiaopeng Hong, and Deyu Meng. Semi-supervised counting via pixel-by-pixel density distribution modelling. arXiv preprint arXiv:2402.15297, 2024. 2 +[26] Wei Lin and Antoni B Chan. Optimal transport minimization: Crowd localization on density maps for semi-supervised counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21663-21673, 2023. 1, 2, 3, 6, 7 +[27] Wei Lin and Antoni B Chan. A fixed-point approach to unified prompt-based counting. In Proceedings of the + +AAAI Conference on Artificial Intelligence, pages 3468-3476, 2024. 2 +[28] Wei Lin, Jia Wan, and Antoni B Chan. Proximal mapping loss: Understanding loss functions in crowd counting & localization. In The Thirteenth International Conference on Learning Representations, 2 +[29] Wei Lin, Junyu Gao, Qi Wang, and Xuelong Li. Learning to detect anomaly events in crowd scenes from synthetic data. Neurocomputing, 436:248-259, 2021. 1, 7 +[30] Wei Lin, Kunlin Yang, Xinzhu Ma, Junyu Gao, Lingbo Liu, Shinan Liu, Jun Hou, Shuai Yi, and Antoni B Chan. Scale-prior deformable convolution for exemplar-guided class-agnostic counting. In BMVC, page 313, 2022. 2 +[31] Chengxin Liu, Hao Lu, Zhiguo Cao, and Tongliang Liu. Point-query quadtree for crowd counting, localization, and more. In CVPR, pages 1676-1685, 2023. 1, 2 +[32] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Leveraging unlabeled data for crowd counting by learning to rank. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7661-7669, 2018. 2, 7 +[33] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Exploiting unlabeled data in cnns by self-supervised learning to rank. IEEE transactions on pattern analysis and machine intelligence, 41(8):1862-1878, 2019. 2 +[34] Yan Liu, Lingqiao Liu, Peng Wang, Pingping Zhang, and Yinjie Lei. Semi-supervised crowd counting via self-training on surrogate tasks. In European Conference on Computer Vision, pages 242-259. Springer, 2020. 2, 7 +[35] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, pages 129-137, 2020. 7 +[36] Yanbo Liu, Yingxiang Hu, Guo Cao, and Yanfeng Shang. Semi-supervised crowd counting via multi-task pseudo-label self-correction strategy. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2 +[37] Zheng Ma and Antoni B Chan. Counting people crossing a line using integer programming and local features. IEEE Transactions on Circuits and Systems for Video Technology, 26(10):1955-1969, 2015. 1 +[38] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In Proceedings of the IEEE International Conference on Computer Vision, pages 6142-6151, 2019. 1, 2, 7, 13 +[39] Zhiheng Ma, Xing Wei, Xiaopeng Hong, Hui Lin, Yunfeng Qiu, and Yihong Gong. Learning to count via unbalanced optimal transport. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2319-2327, 2021. 2 +[40] Lucas Mansilla, Rodrigo Echeveste, Diego H Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6630-6638, 2021. 7 +[41] Riccardo Mazzon, Fabio Poiesi, and Andrea Cavallaro. Detection and tracking of groups in crowd. In 2013 10th IEEE International conference on advanced video and signal based surveillance, pages 202-207. IEEE, 2013. 1 + +[42] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15549-15559, 2021. 2, 3, 7 +[43] Zhuoxuan Peng and S-H Gary Chan. Single domain generalization for crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28025-28034, 2024. 7 +[44] Harish Guruprasad Ramaswamy et al. Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 983-991, 2020. 3 +[45] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3394-3403, 2021. 2 +[46] Weihong Ren, Xinchao Wang, Jiandong Tian, Yandong Tang, and Antoni B Chan. Tracking-by-counting: Using network flows on crowd density maps for tracking multiple targets. IEEE Transactions on Image Processing, 30:1439-1452, 2020. 1 +[47] Jihye Ryu and Kwangho Song. Crowd counting and individual localization using pseudo square label. IEEE Access, 2024. 2 +[48] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 3, 4 +[49] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 323-331, 2024. 2 +[50] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19618-19627, 2022. 2 +[51] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19618-19627, 2022. 2 +[52] Weibo Shu, Jia Wan, and Antoni B Chan. Generalized characteristic function loss for crowd analysis in the frequency domain. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[53] Karen Simonyan. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 4 +[54] Vishwanath A Sindagi, Rajeev Yasarla, Deepak Sam Babu, R Venkatesh Babu, and Vishal M Patel. Learning to count in the crowd from limited labeled data. In European Conference on Computer Vision, pages 212-229. Springer, 2020. 2, 7 +[55] Vishwanath A Sindagi, Rajeev Yasarla, and Vishal M Patel. Jhu-crowd++: Large-scale crowd counting dataset and a + +benchmark method. IEEE transactions on pattern analysis and machine intelligence, 44(5):2594-2609, 2020. 6, 7 +[56] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 2, 3 +[57] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In CVPR, pages 3365-3374, 2021. 1, 2, 3, 6, 7, 13 +[58] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin A. Riedmiller. Striving for simplicity: The all convolutional net. In 3rd International Conference on Learning Representations, ICLR 2015, Workshop Track Proceedings, 2015. 4 +[59] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 3, 7 +[60] Jia Wan, Qingzhong Wang, and Antoni B Chan. Kernel-based density map generation for dense object counting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1357-1370, 2020. 2 +[61] Jia Wan, Ziquan Liu, and Antoni B Chan. A generalized loss function for crowd counting and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1974-1983, 2021. 1, 2, 3, 13 +[62] Jia Wan, Qiangqiang Wu, Wei Lin, and Antoni Chan. Robust zero-shot crowd counting and localization with adaptive resolution sam. In European Conference on Computer Vision, pages 478-495. Springer, 2024. 1 +[63] Boyu Wang, Huidong Liu, Dimitris Samaras, and Minh Hoai Nguyen. Distribution matching for crowd counting. Advances in neural information processing systems, 33:1595-1607, 2020. 2, 13 +[64] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 3 +[65] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8198-8207, 2019. 2, 7 +[66] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. Nwpu-crowd: A large-scale benchmark for crowd counting and localization. IEEE transactions on pattern analysis and machine intelligence, 43(6):2141-2149, 2020. 1 +[67] Qi Wang, Wei Lin, Junyu Gao, and Xuelong Li. Density-aware curriculum learning for crowd counting. IEEE Transactions on Cybernetics, 52(6):4675-4687, 2020. 2, 6 +[68] Qi Wang, Wei Lin, Junyu Gao, and Xuelong Li. Density-aware curriculum learning for crowd counting. IEEE Transactions on Cybernetics, 52(6):4675-4687, 2020. 1 + +[69] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Pixel-wise crowd understanding via synthetic data. International Journal of Computer Vision, 129(1):225-245, 2021. 2 +[70] Qiangqiang Wu, Jia Wan, and Antoni B Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 658-666, 2021. 7 +[71] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15548-15557, 2023. 2 +[72] Qi Zhang, Wei Lin, and Antoni B Chan. Cross-view cross-scene multi-view crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 557-567, 2021. 7 +[73] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 589-597, 2016. 1, 2, 6, 7, 13 +[74] Chenyang Zhao and Antoni B. Chan. ODAM: gradient-based instance-specific visual explanations for object detection. In ICLR, 2023. 2, 3, 4 +[75] Chunhui Zhao, Zhiyuan Zhang, Jinwen Hu, Dong Wang, Bin Fan, Quan Pan, and Qiang He. Crowd anomaly event detection in surveillance video based on the evolution of the spatial position relationship feature. In 2018 IEEE 14th International Conference on Control and Automation (ICCA), pages 247-252, 2018. 1 +[76] Chenyang Zhao, Janet H Hsiao, and Antoni B Chan. Gradient-based instance-specific visual explanations for object specification and object discrimination. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 3 +[77] Chenyang Zhao, Kun Wang, Xingyu Zeng, Rui Zhao, and Antoni B Chan. Gradient-based visual explanation for transformer-based clip. In International Conference on Machine Learning, pages 61072-61091. PMLR, 2024. +[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 3, 4 +[79] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5659-5668, 2022. 7 +[80] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 2023. 7 \ No newline at end of file diff --git a/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/images.zip b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..96806376ca8627b4c919c974e88f0d627911884d --- /dev/null +++ b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80712b0a5dc49fd95c6a4be7f875dc1ceb471712bc80c41d9a1aa60c5a46c2f3 +size 517552 diff --git a/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/layout.json b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a662906a6f690a1cb88cbf779c6e0e1b4d07b161 --- /dev/null +++ b/2025/Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting/layout.json @@ -0,0 +1,14214 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 526, + 121 + ], + "type": "text", + "content": "Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 144, + 422, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 144, + 422, + 157 + ], + "spans": [ + { + "bbox": [ + 187, + 144, + 422, + 157 + ], + "type": "text", + "content": "Wei Lin, Chenyang Zhao, and Antoni B. Chan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 158, + 462, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 158, + 462, + 172 + ], + "spans": [ + { + "bbox": [ + 148, + 158, + 462, + 172 + ], + "type": "text", + "content": "Department of Computer Science, City University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 174, + 493, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 174, + 493, + 184 + ], + "spans": [ + { + "bbox": [ + 113, + 174, + 493, + 184 + ], + "type": "text", + "content": "elonlin24@gmail.com, chenyzhao9-c@my.cityu.edu.hk, abchan@cityu.edu.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 213, + 200, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 224 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 239, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 597 + ], + "type": "text", + "content": "Point detection has been developed to locate pedestrians in crowded scenes by training a counter through a point-to-point (P2P) supervision scheme. Despite its excellent localization and counting performance, training a point-based counter still faces challenges concerning annotation labor: hundreds to thousands of points are required to annotate a single sample capturing a dense crowd. In this paper, we integrate point-based methods into a semi-supervised counting framework based on pseudo-labeling, enabling the training of a counter with only a few annotated samples supplemented by a large volume of pseudolabeled data. However, during implementation, the training encounters issues as the confidence for pseudo-labels fails to be propagated to background pixels via the P2P. To tackle this challenge, we devise a point-specific activation map (PSAM) to visually interpret the phenomena occurring during the ill-posed training. Observations from the PSAM suggest that the feature map is excessively activated by the loss for unlabeled data, causing the decoder to misinterpret these over-activations as pedestrians. To mitigate this issue, we propose a point-to-region (P2R) scheme to substitute P2P, which segments out local regions rather than detects a point corresponding to a pedestrian for supervision. Consequently, pixels in the local region can share the same confidence with the corresponding pseudo points. Experimental results in both semi-supervised counting and unsupervised domain adaptation highlight the advantages of our method, illustrating P2R can resolve issues identified in PSAM. The code is available at https://github.com/Elin24/P2RLoss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 609, + 135, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 609, + 135, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 135, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Crowd scene analysis has been investigated for many years due to its significant applications in smart cities and urban safety [2, 7, 29, 41, 46, 62, 75]. Crowd counting and localization in extremely dense crowd scenes, as the fundamental task in this field, have attracted considerable attention from researchers in computer vision [4, 5, 37, 61, 66, 68]. Although most supervised crowd counting methods fol" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 213, + 380, + 322 + ], + "blocks": [ + { + "bbox": [ + 320, + 213, + 380, + 322 + ], + "lines": [ + { + "bbox": [ + 320, + 213, + 380, + 322 + ], + "spans": [ + { + "bbox": [ + 320, + 213, + 380, + 322 + ], + "type": "image", + "image_path": "f8654d6fa55092cd93b854925481932197aa3f3617e69f7065ee6cc69b0c2dcb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 331, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 407 + ], + "type": "text", + "content": "Figure 1. The workflow of semi-supervised point-based counting methods. The teacher model generates pseudo labels by extracting the foreground pixels, while the student model takes the corresponding strongly augmented image as input to construct the computation graph. The training loss between the pseudo label and the student's prediction involves two steps: the proposed P2R matching and the weighted cross-entropy computation." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 386, + 219, + 420, + 255 + ], + "blocks": [ + { + "bbox": [ + 386, + 219, + 420, + 255 + ], + "lines": [ + { + "bbox": [ + 386, + 219, + 420, + 255 + ], + "spans": [ + { + "bbox": [ + 386, + 219, + 420, + 255 + ], + "type": "image", + "image_path": "2e4ee55c1d884fd65d86c1d6233a1875ae7bb5d03211fa2166453f6520e603bf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 426, + 217, + 474, + 253 + ], + "blocks": [ + { + "bbox": [ + 426, + 217, + 474, + 253 + ], + "lines": [ + { + "bbox": [ + 426, + 217, + 474, + 253 + ], + "spans": [ + { + "bbox": [ + 426, + 217, + 474, + 253 + ], + "type": "image", + "image_path": "29dad3d08f89e07b65a71e50ecf6be38f07b709bc2245095e9ed0104f2a45e1e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 254, + 463, + 260 + ], + "lines": [ + { + "bbox": [ + 427, + 254, + 463, + 260 + ], + "spans": [ + { + "bbox": [ + 427, + 254, + 463, + 260 + ], + "type": "text", + "content": "Prediction" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 473, + 268, + 529, + 277 + ], + "lines": [ + { + "bbox": [ + 473, + 268, + 529, + 277 + ], + "spans": [ + { + "bbox": [ + 473, + 268, + 529, + 277 + ], + "type": "text", + "content": "w/o gradient" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 482, + 284, + 504, + 297 + ], + "lines": [ + { + "bbox": [ + 482, + 284, + 504, + 297 + ], + "spans": [ + { + "bbox": [ + 482, + 284, + 504, + 297 + ], + "type": "text", + "content": "P2R Matching" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 501, + 219, + 521, + 251 + ], + "blocks": [ + { + "bbox": [ + 477, + 227, + 493, + 234 + ], + "lines": [ + { + "bbox": [ + 477, + 227, + 493, + 234 + ], + "spans": [ + { + "bbox": [ + 477, + 227, + 493, + 234 + ], + "type": "text", + "content": ">0.5" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 501, + 219, + 521, + 251 + ], + "lines": [ + { + "bbox": [ + 501, + 219, + 521, + 251 + ], + "spans": [ + { + "bbox": [ + 501, + 219, + 521, + 251 + ], + "type": "image", + "image_path": "928f6f3b2ac36d47a69f7b9084e69bcb6d14188db1b83eb2cc32f47965cdd4e1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 254, + 528, + 260 + ], + "lines": [ + { + "bbox": [ + 501, + 254, + 528, + 260 + ], + "spans": [ + { + "bbox": [ + 501, + 254, + 528, + 260 + ], + "type": "text", + "content": "Sudo Label" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 513, + 284, + 544, + 297 + ], + "lines": [ + { + "bbox": [ + 513, + 284, + 544, + 297 + ], + "spans": [ + { + "bbox": [ + 513, + 284, + 544, + 297 + ], + "type": "text", + "content": "Weighted Cross Entropy" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 490, + 304, + 529, + 312 + ], + "lines": [ + { + "bbox": [ + 490, + 304, + 529, + 312 + ], + "spans": [ + { + "bbox": [ + 490, + 304, + 529, + 312 + ], + "type": "text", + "content": "w/ gradient" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 318, + 257, + 380, + 275 + ], + "blocks": [ + { + "bbox": [ + 318, + 257, + 380, + 275 + ], + "lines": [ + { + "bbox": [ + 318, + 257, + 380, + 275 + ], + "spans": [ + { + "bbox": [ + 318, + 257, + 380, + 275 + ], + "type": "image", + "image_path": "edd9073bc1c491cc4c1267cd01898ab7ba8b88a9b3e9d4acce3d0c2a8fb6e431.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 386, + 254, + 418, + 275 + ], + "blocks": [ + { + "bbox": [ + 386, + 254, + 418, + 275 + ], + "lines": [ + { + "bbox": [ + 386, + 254, + 418, + 275 + ], + "spans": [ + { + "bbox": [ + 386, + 254, + 418, + 275 + ], + "type": "image", + "image_path": "2ee8b5bdfd9bd808a379c9bfaf6cbdd6c20f24a0ccbb21ec8001ceeddf4b4b32.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 386, + 277, + 421, + 308 + ], + "blocks": [ + { + "bbox": [ + 386, + 277, + 421, + 308 + ], + "lines": [ + { + "bbox": [ + 386, + 277, + 421, + 308 + ], + "spans": [ + { + "bbox": [ + 386, + 277, + 421, + 308 + ], + "type": "image", + "image_path": "b656620bb22599a32d692f01e06c6b70d78e1199393337a31b3729d11c0f8067.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 309, + 413, + 315 + ], + "lines": [ + { + "bbox": [ + 393, + 309, + 413, + 315 + ], + "spans": [ + { + "bbox": [ + 393, + 309, + 413, + 315 + ], + "type": "text", + "content": "Student" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 422, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 422, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 312, + 422, + 555, + 567 + ], + "type": "text", + "content": "low the scheme of density regression [10, 24, 38, 61, 73], there are also relevant studies implementing point detection for crowd localization and counting [21, 31, 57], due to its straightforwardness in indicating pedestrians' positions. These approaches use the Hungarian algorithm to perform point-to-point (P2P) matching between predicted point maps and ground-truth (GT) pedestrian annotations during training. Therefore, these matched pixels are considered as pedestrians' positions, while the others are designated as background for loss computation. This scheme successfully trains a point detector for crowd analysis, yielding both good localization and counting performance." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Great progress has been made in supervised counting, but the labor-intensive nature of annotating data has hindered the development of crowd understanding for a long time [26, 66]. Addressing this issue, semi-supervised counting is gradually being developed, aiming to train a model using large amounts of unlabeled data alongside only a few annotated samples. Similar to supervised counting, most semi-supervised methods also follow the procedure of density regression. Notably, the method proposed in OT-M [26] stands out by employing a simple self-training scheme based on mean-teacher [59] and FixMatch [56]. However, there is no study exploring semi-supervised counting and" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29363" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": "localization using a point-detection method. To fill this gap, this paper explores an approach to train a point detector with limited annotated data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 110, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 295, + 337 + ], + "type": "text", + "content": "As shown in Fig. 1, typical semi-supervised counting methods rely on pseudo-labeling, which generates pseudo labels to serve as learning targets for unlabeled samples using a model pre-trained on a small amount of annotated data. During training, there is a confidence score to determine whether a concerned prediction should be retained as a pseudo label. However, if the P2P matching strategy is applied to compute loss within the pipeline presented in Fig. 1, there is no reasonable way to define the confidence for unmatched background pixels, as P2P matching only cares about matched foreground pixels. Specifically, a foreground pixel can be retained for training if its matched pseudo-label's foreground probability exceeds a predefined threshold, similar to the approach used in semi-supervised classification tasks [15, 56]. However, background pixels lack confidence scores since no pseudo-labels are assigned to them after matching. Experimental results also demonstrate that ignoring the background leads to the breakdown of semi-supervised training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "text", + "content": "To address the aforementioned issue, we propose a gradient-based visual explanation method (akin to [48, 74, 76, 78]), termed point-specific activation map (PSAM), to visualize regions of interest and the activation levels of detected points. Using PSAM, we observe that the activation values and coverage areas of foreground pixels continuously grow during semi-supervised training, causing the model to misclassify neighboring pixels as foreground." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 435, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 435, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 56, + 435, + 295, + 650 + ], + "type": "text", + "content": "Inspecting PSAMs reveals that each point in GT is implicitly associated with a local region rather than a single pixel in the prediction. Accordingly, we propose to replace the P2P with a point-to-region (P2R) strategy, which matches local regions to points in the pseudo-label. This allows all pixels within a matched region, including background pixels, to share the confidence of the corresponding pseudo-point. Experimental results and visualizations demonstrate that P2R can effectively train a semisupervised point-based counter and achieve expected performance. Experimental results show that P2R effectively trains semi-supervised point-based counters with comparable performance to P2P but at significantly reduced computational cost. Unlike P2P [57], P2R eliminates the need for the Hungarian algorithm by assigning each ground-truth or pseudo-point a local \"dominated zone\" and selecting a representative foreground pixel based on a predefined cost matrix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 652, + 253, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 652, + 253, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 652, + 253, + 663 + ], + "type": "text", + "content": "The contributions of our paper are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 665, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 665, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 295, + 700 + ], + "type": "text", + "content": "- We formulate a pipeline of semi-supervised counting with a point-detection method, but find a problem leading the training breakdown." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "content": "- We propose Point-Specific Activation Map (PSAM) to" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 333, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 144 + ], + "type": "text", + "content": "visualize the regions of interest and the activation values of each point detected by the point-based crowd counter. Through PSAM, we observe that the surrounding pixels of each expected foreground pixel are over-activated, causing the decoder to recognizes these pixels as instances different from the concerned one." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 144, + 553, + 275 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 326, + 144, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 144, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 326, + 144, + 553, + 215 + ], + "type": "text", + "content": "- Based on insights from PSAM, we propose replacing the P2P with a P2R matching strategy for semi-supervised learning in point-based crowd counting. This not only makes the training process work as expected, but also brings additional benefits of reducing computational requirements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 216, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 216, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 326, + 216, + 553, + 275 + ], + "type": "text", + "content": "- The proposed P2R achieves outstanding performance in several experiments, including semi-supervised counting and unsupervised domain adaptation. Ablation study shows that P2R excellently tackles the problem caused by P2P in semi-supervised counting." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 286, + 406, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 406, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 406, + 298 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 307, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 569 + ], + "type": "text", + "content": "Object counting in dense scenes. The development of deep learning has significantly influenced crowd counting. Prior to this, detecting body parts [16, 17] was the common pipeline for crowd understanding in computer vision, although some methods implemented counting models without explicit object segmentation or tracking information [4-6]. Beyond specific crowd scenarios, class-agnostic counting [14, 27, 30, 45, 49, 71] has also been developed to estimate the distribution of objects in the same category. Most recent counting methods implement neural networks following the density map regression scheme [3, 8, 20, 50, 60, 65, 67, 69, 73]. In addition to continuous innovations in model architecture design [3, 20, 24], new loss functions have also been developed, including Bayesian Loss [38], loss functions based on balanced [63] and unbalanced optimal transport [22, 39, 61], and characteristic function loss computed in the frequency domain [51, 52]. A study analyzes the relationship between current loss functions from the perspective of proximal mapping [28]. Some methods also perform crowd analysis via instance segmentation [1, 9] and point detection [21, 31, 47, 57]. The latter is the scheme adopted in our semi-supervised model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Semi-supervised counting. Semi-supervised counting has attracted broad attention recently [18, 19, 23, 25, 26, 32-34, 36, 42, 54]. L2R [32, 33] introduces a ranking rule for unlabeled crowd images based on the principle that cropped images should have smaller counts than full images. GP [54] iteratively updates pseudo-labels using Gaussian Processes. IRAST [34] and SUA [42] enhance models by incorporating a segmentation branch. DAC [23] uses learnable density agents to capture varying crowd densities. OT-M [26] transforms density maps into point maps as pseudo-labels, achieving strong performance with a density-to-point loss function [61]. Our method also em" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29364" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "ploys pseudo-labeling but uses a point detector for counting, avoiding the need for density map transformations as in [26]. Additionally, uncertainty is directly measured by the scores of extracted points, without any additional learnable structures like those in [42]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 133, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 296, + 361 + ], + "type": "text", + "content": "Visual Explanation. Visualizing the importance of extracted image features is a straightforward way to interpret a model. The mainstream of visual explanation focuses on methods producing activation maps [13, 44, 48, 64, 74, 76-78], which are the products of feature maps from some intermediate layer and weight maps indicating the corresponding feature's importance. For example, GradCAM [13, 48] defines the weights using the corresponding gradient maps, generated via backpropagation. Beyond the focus on classification tasks, ODAM [74, 76] extends gradient-based CAM to object detection, visualizing the local impact of features on the detector's decisions. In this paper, we propose PSAM to visualize how the activation map changes for each predicted point in an ill-posed semi-supervised counting framework. Based on the observations from PSAM, we design a P2R matching strategy to avoid the over-activation of pixels in local regions, which effectively resolves the drawbacks of P2P and enables the successful training of a semi-supervised point detector." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 371, + 296, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 385 + ], + "type": "text", + "content": "3. Point Counter in Semi-Supervised Counting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 392, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 440 + ], + "type": "text", + "content": "This section formulates a simplified point-based counter by removing the regression branch in P2PNet [57], and then presents the training scheme and challenges when training with labeled and unlabeled data, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 279, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 279, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 279, + 460 + ], + "type": "text", + "content": "3.1. Preliminary on A Simplified Point Counter" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 586 + ], + "type": "text", + "content": "Common point-based crowd counters like P2PNet [57] contain a classification and a regression branch. However, we find a simplification can be achieved by removing the offset regression branch. On one hand, the crowd count can be estimated by counting foreground pixels with scores greater than 0.5, a process that does not involve the regression branch. On the other hand, pixel coordinates alone are sufficient to locate pedestrians, as demonstrated in STEERER [10] and GL [61], which use the coordinates of local maxima to indicate pedestrians locations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 587, + 233, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 587, + 233, + 598 + ], + "spans": [ + { + "bbox": [ + 67, + 587, + 233, + 598 + ], + "type": "text", + "content": "As a result, P2PNet can be simplified to:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 608, + 295, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 608, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 91, + 608, + 295, + 657 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {F} (I, \\Theta_ {f}) \\rightarrow F \\in \\mathbb {R} ^ {c \\times h \\times w}, (1) \\\\ \\mathcal {D} (F, \\Theta_ {d}) \\rightarrow D \\in \\mathbb {R} ^ {h \\times w} (2) \\\\ \\rightarrow \\mathcal {P} = \\left\\{\\boldsymbol {p} \\in \\mathbb {R} ^ {n}, \\boldsymbol {x} \\in \\mathbb {R} ^ {n \\times 2} \\right\\}, (3) \\\\ \\end{array}", + "image_path": "7e591782e248b5b5bbd30d07d4f3ddf1b7ebfd0800b4d0da7414fcd694d5b037.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is the image encoder with learnable parameters " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Theta_f" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " denotes the density decoder with learnable parameter " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Theta_d" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". Notably, the offset regression branch is excluded in this simplified formulation. The pixel set " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{p} \\in \\mathbb{R}^n" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "represents the flattened values in " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ", and its corresponding coordinate set is denoted as " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^{n \\times 2}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "n = h \\times w" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 102, + 553, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 553, + 116 + ], + "type": "text", + "content": "3.2. Training with Labeled Data via P2P Matching" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "content": "The GT with " + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "content": " points can be denoted as " + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\prime} \\in \\mathbb{R}^{m \\times 2}" + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{[j]}^{\\prime} \\in \\mathbb{R}^{2}" + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "content": " represents the location of the " + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "content": "-th pedestrian. Using " + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\prime}" + }, + { + "bbox": [ + 313, + 121, + 553, + 169 + ], + "type": "text", + "content": ", the loss computation involves two parts: P2P matching and binary cross entropy (BCE) calculation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 170, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 205 + ], + "type": "text", + "content": "The former aims to obtain a matrix-based solution " + }, + { + "bbox": [ + 313, + 170, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\{0,1\\}^{n \\times m}" + }, + { + "bbox": [ + 313, + 170, + 553, + 205 + ], + "type": "text", + "content": " that minimizes the bipartite-graph matching cost between " + }, + { + "bbox": [ + 313, + 170, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 170, + 553, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 170, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\pmb{x}'" + }, + { + "bbox": [ + 313, + 170, + 553, + 205 + ], + "type": "text", + "content": " within the cost matrix " + }, + { + "bbox": [ + 313, + 170, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{n \\times m}" + }, + { + "bbox": [ + 313, + 170, + 553, + 205 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 213, + 553, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 213, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 361, + 213, + 553, + 228 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {[ i j ]} = \\tau \\| \\boldsymbol {x} _ {[ i ]} - \\boldsymbol {x} ^ {\\prime} _ {[ j ]} \\| _ {2} - S (\\boldsymbol {p} _ {[ i ]}), \\tag {4}", + "image_path": "07c6ccd2b67afec28a85d4cc72cfa0ce3620ae25949be993d6c35b1e8073f820.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 233, + 553, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 553, + 269 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 233, + 553, + 269 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 313, + 233, + 553, + 269 + ], + "type": "text", + "content": " is defined as the inverse sigmoid function " + }, + { + "bbox": [ + 313, + 233, + 553, + 269 + ], + "type": "inline_equation", + "content": "S(p) = -\\log (1 / p - 1)" + }, + { + "bbox": [ + 313, + 233, + 553, + 269 + ], + "type": "text", + "content": ", rather than the identity operator used in the vanilla P2PNet [57], for improved performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 269, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 304 + ], + "type": "text", + "content": "To compute BCE loss, the learning objective of " + }, + { + "bbox": [ + 313, + 269, + 553, + 304 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 313, + 269, + 553, + 304 + ], + "type": "text", + "content": " is formulated as " + }, + { + "bbox": [ + 313, + 269, + 553, + 304 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}} = \\mathbf{M}\\mathbf{1}_m" + }, + { + "bbox": [ + 313, + 269, + 553, + 304 + ], + "type": "text", + "content": ", and the BCE loss is calculated for optimization as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 311, + 553, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 311, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 343, + 311, + 553, + 326 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l} = - \\lambda \\hat {\\boldsymbol {p}} ^ {\\top} \\log \\boldsymbol {p} - \\left(\\mathbf {1} _ {n} - \\hat {\\boldsymbol {p}}\\right) ^ {\\top} \\log \\left(\\mathbf {1} _ {n} - \\boldsymbol {p}\\right) \\tag {5}", + "image_path": "cc8acfb22a5dfe458b06630956ed9715529bd663fedd929732459009cb4aa183.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 333, + 542, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 542, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 542, + 345 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 333, + 542, + 345 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 333, + 542, + 345 + ], + "type": "text", + "content": " is a re-weighting factor for matched pixels [57]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 351, + 538, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 538, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 538, + 364 + ], + "type": "text", + "content": "3.3. Problem in Training with Unlabeled Image" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 369, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 514 + ], + "type": "text", + "content": "Here we consider training with unlabeled data following the standard pseudo-labeling procedure [15, 56]. As illustrated in Fig. 1, it requires the cooperation of a student and teacher model that share the same structure described in (1)~(3). The difference lies in that the teacher model is updated via exponential moving average (EMA) [59], while the student is optimized via gradient descending. During training, the teacher and student process the weakly and strongly data-augmented versions of an unlabeled image, producing " + }, + { + "bbox": [ + 313, + 369, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t = \\{\\pmb{p}_t, \\pmb{x}_t\\}" + }, + { + "bbox": [ + 313, + 369, + 555, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 369, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_s = \\{\\pmb{p}_s, \\pmb{x}_s\\}" + }, + { + "bbox": [ + 313, + 369, + 555, + 514 + ], + "type": "text", + "content": ", respectively. After that, foreground points are extracted from " + }, + { + "bbox": [ + 313, + 369, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 313, + 369, + 555, + 514 + ], + "type": "text", + "content": " to construct pseudo-labels:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 339, + 520, + 553, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 520, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 339, + 520, + 553, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {t} ^ {\\prime} = \\left\\{\\boldsymbol {p} _ {t} ^ {\\prime}, \\boldsymbol {x} _ {t} ^ {\\prime} \\right\\} = \\left\\{\\boldsymbol {p} _ {t [ j ]}, \\boldsymbol {x} _ {t [ j ]} \\mid \\boldsymbol {p} _ {t [ j ]} > 0. 5 \\right\\}, \\tag {6}", + "image_path": "cfcc3582344545979b55d198068dfb0d35de63704cb874f3e9d8f785cda15882.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 542, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 553, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 542, + 553, + 578 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t^\\prime" + }, + { + "bbox": [ + 313, + 542, + 553, + 578 + ], + "type": "text", + "content": " is treated as the pseudo hard label to estimate the matching matrix " + }, + { + "bbox": [ + 313, + 542, + 553, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{st}" + }, + { + "bbox": [ + 313, + 542, + 553, + 578 + ], + "type": "text", + "content": " via Hungarian algorithm, and the loss is obtained by substituting " + }, + { + "bbox": [ + 313, + 542, + 553, + 578 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}_t = \\mathbf{M}_{st}\\mathbf{1}_m" + }, + { + "bbox": [ + 313, + 542, + 553, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 542, + 553, + 578 + ], + "type": "inline_equation", + "content": "\\pmb{p}_s" + }, + { + "bbox": [ + 313, + 542, + 553, + 578 + ], + "type": "text", + "content": " into (5)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 578, + 553, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 553, + 626 + ], + "type": "text", + "content": "However, some pseudo-labels may not be reliable. A common solution is to only consider points in " + }, + { + "bbox": [ + 313, + 578, + 553, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t^\\prime" + }, + { + "bbox": [ + 313, + 578, + 553, + 626 + ], + "type": "text", + "content": " with high scores [15, 56], i.e., using a confidence vector " + }, + { + "bbox": [ + 313, + 578, + 553, + 626 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 313, + 578, + 553, + 626 + ], + "type": "text", + "content": " to mask out elements whose score is below a threshold " + }, + { + "bbox": [ + 313, + 578, + 553, + 626 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 578, + 553, + 626 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 341, + 633, + 553, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 633, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 341, + 633, + 553, + 651 + ], + "type": "interline_equation", + "content": "\\zeta = \\mathbb {1} \\left(\\boldsymbol {p} _ {t} ^ {\\prime} > \\eta\\right) \\xrightarrow {\\text {m a p t o} \\hat {\\boldsymbol {p}} _ {t} \\text {b y} \\mathbf {M} _ {s t}} \\boldsymbol {z} = \\mathbf {M} _ {s t} \\zeta . \\tag {7}", + "image_path": "b5f5bcc2f2cf9e8e4785d7d69ec13acaaf08039429d8b1554c74cdbe7b6bc210.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 658, + 553, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 553, + 693 + ], + "type": "text", + "content": "Note " + }, + { + "bbox": [ + 313, + 658, + 553, + 693 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 658, + 553, + 693 + ], + "type": "text", + "content": " is greater than 0.5 and (7) only let pseudo points with score greater than " + }, + { + "bbox": [ + 313, + 658, + 553, + 693 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 658, + 553, + 693 + ], + "type": "text", + "content": " be reliable. Subsequently, " + }, + { + "bbox": [ + 313, + 658, + 553, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\mathrm{diag}(z)" + }, + { + "bbox": [ + 313, + 658, + 553, + 693 + ], + "type": "text", + "content": " is incorporated into the BCE loss for supervision:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 321, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = - \\lambda \\hat {\\boldsymbol {p}} _ {t} ^ {\\top} \\mathbf {Z} \\log \\boldsymbol {p} _ {s} - \\left(\\mathbf {1} _ {n} - \\hat {\\boldsymbol {p}} _ {t}\\right) ^ {\\top} \\mathbf {Z} \\log \\left(\\mathbf {1} _ {n} - \\boldsymbol {p} _ {s}\\right). \\tag {8}", + "image_path": "350155e5e61c8db8289534694037caa99b0945abe943d0850414a5b2ded21b57.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29365" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 292, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 292, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 292, + 84 + ], + "type": "text", + "content": "However, the second item of (8) in P2P is identically zero:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 88, + 295, + 103 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 88, + 295, + 103 + ], + "spans": [ + { + "bbox": [ + 111, + 88, + 295, + 103 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u [ P 2 P ]} = - \\lambda \\hat {\\boldsymbol {p}} _ {t} ^ {\\top} \\mathbf {Z} \\log \\boldsymbol {p} _ {s} - 0, \\tag {9}", + "image_path": "31cf56f6e3aa8eddd45b37adb9df1a777391c845c83cf8ffd2bcd3c04aaab813.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 107, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 296, + 180 + ], + "type": "text", + "content": "regardless of " + }, + { + "bbox": [ + 55, + 107, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 55, + 107, + 296, + 180 + ], + "type": "text", + "content": " (embedded in " + }, + { + "bbox": [ + 55, + 107, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 107, + 296, + 180 + ], + "type": "text", + "content": "). Consequently, no gradients are back-propagated from the losses associated with background pixels, which is ill-posed. Utilizing (9) as the loss function is similar to training a two-class classifier with only samples from the category \"1\", which will lead the trained model to output \"1\" regardless of the input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 188, + 231, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 231, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 231, + 202 + ], + "type": "text", + "content": "4. Visual Explanation with PSAM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 209, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 316 + ], + "type": "text", + "content": "In addition to the theoretical analysis, we implement a framework that supervises a model using labeled and unlabeled data via (5) and (9), respectively, to empirically investigate the behavior of P2P-based semi-supervised counting. Subsequently, PSAM is designed to visualize how the prediction is influenced by (9) during training. Thus, we can design an effective confidence scheme accordingly to counteract the effects revealed by PSAM, thereby enabling semi-supervised learning to function as expected." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 322, + 257, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 257, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 257, + 335 + ], + "type": "text", + "content": "4.1. PSAM: Point-Specific Activation Map" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 340, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 435 + ], + "type": "text", + "content": "An activation map is a heat map with the same spatial resolution as the image features, where local regions with relatively high values are interpreted as being discriminative (informative) for the current output [48, 74, 78]. Most techniques obtain activation maps by linearly aggregating features with their corresponding gradients, as [53, 58] have shown that the partial derivative with respect to a specific element can reflect its contribution to the final prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "type": "text", + "content": "Unlike [48, 78], which compute gradients at the image level, crowd counting requires a point-specific activation map (PSAM) to visually explain the regions of interest for every element in " + }, + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "type": "text", + "content": " of (3). Thus, the Jacobian matrix of " + }, + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "type": "text", + "content": " with respect to image feature " + }, + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 436, + 296, + 496 + ], + "type": "text", + "content": " is required:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 500, + 294, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 500, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 79, + 500, + 294, + 517 + ], + "type": "interline_equation", + "content": "\\mathbf {J} _ {\\mathcal {D}} = \\frac {\\partial \\boldsymbol {p}}{\\partial F} = \\left[ \\nabla \\boldsymbol {p} _ {[ 1 ]}, \\nabla \\boldsymbol {p} _ {[ 2 ]}, \\dots , \\nabla \\boldsymbol {p} _ {[ n ]} \\right] ^ {\\top}. \\tag {10}", + "image_path": "d513c2310f881ac9738f91d05c65932fb311ffd9837a0c910cb79a876a0f0644.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "text", + "content": "Afterwards, the " + }, + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "text", + "content": "-th heat map " + }, + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "inline_equation", + "content": "H_{[q]}" + }, + { + "bbox": [ + 55, + 521, + 295, + 545 + ], + "type": "text", + "content": " in PSAM is derived by filtering out negative influences on the aggregated feature:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 75, + 559, + 295, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 559, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 75, + 559, + 295, + 578 + ], + "type": "interline_equation", + "content": "H _ {[ q ]} = \\max \\left(\\sum_ {k = 1} ^ {c} \\nabla \\boldsymbol {p} _ {[ q ]} \\odot F, 0\\right) \\in \\mathbb {R} ^ {h \\times w}, \\tag {11}", + "image_path": "49ba6c502f496c9d023638419c26242d6ace6750924efc1459ffdf11404a0f05.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "text", + "content": "where the summation operator is applied along the channel dimension, and " + }, + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 582, + 295, + 605 + ], + "type": "text", + "content": " denotes the Hadamard product." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "In a naive implementation, the computation of " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{\\mathcal{D}} \\in \\mathbb{R}^{n \\times hwc}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " requires a large memory and involves " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " backpropagations through the computational graph, posing a challenge for visual explanation in pixel-level tasks. However, this issue can be resolved by computing gradients only between " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{[q]}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " and features within its receptive field. Gradients in other areas are zero, as there is no path connecting them in the computational graph. Thus, an efficient approach can be implemented, as illustrated in Fig. 2." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 232 + ], + "type": "image", + "image_path": "4d58de3567dcc2d6d9e22e47fb4788cd6add94e05ed944dc3cff47ef66ecd303.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 238, + 515, + 249 + ], + "lines": [ + { + "bbox": [ + 354, + 238, + 515, + 249 + ], + "spans": [ + { + "bbox": [ + 354, + 238, + 515, + 249 + ], + "type": "text", + "content": "Figure 2. The generation process of PSAM." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": "In this implementation, the forward process begins with extracting sliding local blocks " + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "inline_equation", + "content": "\\tilde{F}" + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": " according to the receptive field " + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": ", and then decoding " + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "inline_equation", + "content": "\\tilde{F}" + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": " through (2) in parallel. The final prediction " + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": " is generated by stacking pixels located at the center of " + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "inline_equation", + "content": "r\\times r" + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": " patches denoted as in " + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "inline_equation", + "content": "\\tilde{D}" + }, + { + "bbox": [ + 313, + 270, + 554, + 330 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 358, + 337, + 553, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 337, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 358, + 337, + 553, + 352 + ], + "type": "interline_equation", + "content": "F \\in \\mathbb {R} ^ {h \\times w \\times c} \\rightarrow \\tilde {F} \\in \\mathbb {R} ^ {(h w) \\times (r \\times r \\times c)}, \\tag {12}", + "image_path": "8897abf377cdac69e7dd91380a746467cfc171e4268e270e38facb50b5f10f7a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 355, + 553, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 355, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 321, + 355, + 553, + 372 + ], + "type": "interline_equation", + "content": "\\mathcal {D} (\\tilde {F}, \\Theta_ {d}) \\rightarrow \\tilde {D} \\in \\mathbb {R} ^ {n \\times (r \\times r)} \\xrightarrow {\\text {c e n t e r o f p a t c h e s}} \\boldsymbol {p} \\in \\mathbb {R} ^ {n}. \\tag {13}", + "image_path": "ae96e1160ed799dd9732862305177c02e4ecc82065c8df18a377c8539086b208.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 380, + 544, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 544, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 544, + 392 + ], + "type": "text", + "content": "The remaining " + }, + { + "bbox": [ + 313, + 380, + 544, + 392 + ], + "type": "inline_equation", + "content": "r^2 - 1" + }, + { + "bbox": [ + 313, + 380, + 544, + 392 + ], + "type": "text", + "content": " pixels in each patch are discarded." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "text", + "content": "Denoting the sum of the output as " + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "inline_equation", + "content": "\\xi = \\mathbf{1}_n^\\top \\pmb{p}" + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "text", + "content": ", the gradient " + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "inline_equation", + "content": "\\nabla \\xi = \\partial \\xi / \\partial \\tilde{F}" + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "text", + "content": " can be obtained via just one backpropagation, and the " + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "text", + "content": "-th gradient block, " + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "inline_equation", + "content": "\\nabla \\xi_{[q]}" + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "text", + "content": ", is equal to the gradient of " + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{[q]}" + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "inline_equation", + "content": "\\tilde{F}_{[q]}" + }, + { + "bbox": [ + 313, + 392, + 554, + 444 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 355, + 452, + 553, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 452, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 355, + 452, + 553, + 473 + ], + "type": "interline_equation", + "content": "\\nabla \\xi_ {[ q ]} = \\frac {\\partial \\xi}{\\partial \\tilde {F} _ {[ q ]}} = \\frac {\\partial \\xi}{\\partial \\boldsymbol {p} _ {[ q ]}} \\cdot \\frac {\\partial \\boldsymbol {p} _ {[ q ]}}{\\partial \\tilde {F} _ {[ q ]}} = \\frac {\\partial \\boldsymbol {p} _ {[ q ]}}{\\partial \\tilde {F} _ {[ q ]}}, \\tag {14}", + "image_path": "883e56a8e310eb0c2d22442f586a8b1bdcc35cf4cd33144dd26a0cc382a07baa.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 481, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 554, + 520 + ], + "type": "text", + "content": "since " + }, + { + "bbox": [ + 313, + 481, + 554, + 520 + ], + "type": "inline_equation", + "content": "\\partial \\xi /\\partial \\pmb{p}_{[q]} = 1" + }, + { + "bbox": [ + 313, + 481, + 554, + 520 + ], + "type": "text", + "content": ". After that, we can parallel compute the PSAM (denoted as " + }, + { + "bbox": [ + 313, + 481, + 554, + 520 + ], + "type": "inline_equation", + "content": "\\tilde{H}_{[q]}" + }, + { + "bbox": [ + 313, + 481, + 554, + 520 + ], + "type": "text", + "content": ") within the receptive field of " + }, + { + "bbox": [ + 313, + 481, + 554, + 520 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{[q]}" + }, + { + "bbox": [ + 313, + 481, + 554, + 520 + ], + "type": "text", + "content": " by substituting (14) into the formulation presented in (11):" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 356, + 528, + 553, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 528, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 356, + 528, + 553, + 546 + ], + "type": "interline_equation", + "content": "\\tilde {H} _ {[ q ]} = \\max \\left(\\sum_ {k = 1} ^ {c} \\nabla \\xi_ {[ q ]} \\odot \\tilde {F} _ {[ q ]}, 0\\right). \\tag {15}", + "image_path": "57903704fbc93e18aaf6447617a97c36c06dd7e13eb65f197bb525f0b917aabd.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "content": "As illustrated in the bottom of Fig. 2, the complete PSAM " + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "inline_equation", + "content": "H_{[1,\\dots,n]}" + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "content": " in (11) can be obtained by filling " + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\tilde{H}_{[1,\\dots,n]}" + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "content": " back into the corresponding regions of a fully zero matrix:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 367, + 598, + 553, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 598, + 553, + 630 + ], + "spans": [ + { + "bbox": [ + 367, + 598, + 553, + 630 + ], + "type": "interline_equation", + "content": "H _ {[ q, t ]} = \\left\\{ \\begin{array}{l l} \\tilde {H} _ {[ q, t ^ {\\prime} ]}, & \\text {i f} t \\in \\Omega_ {q} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {16}", + "image_path": "88f09b9ea0c72e30db5b550e5cfacdbee9b7a4bdb700b849916db4c55c76f82f.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\Omega_q" + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "content": " represents the index set of " + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "inline_equation", + "content": "p_{[q]}" + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "content": "'s receptive field, and " + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "inline_equation", + "content": "t' \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "content": " is the transformed coordinate of " + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "content": ", ensuring " + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "inline_equation", + "content": "F_{[t]} \\equiv \\tilde{F}_{[q,t']}." + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "content": " In this way, the matrix memory required to obtain all " + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "inline_equation", + "content": "H_{[q]}" + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "content": " is reduced from " + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n \\times hwc}" + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n \\times r^2c}" + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "inline_equation", + "content": "r^2 \\ll hw" + }, + { + "bbox": [ + 313, + 637, + 554, + 714 + ], + "type": "text", + "content": "), and only one back-propagation is needed to obtain PSAMs for all pixels." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29366" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 134, + 129 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 134, + 129 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 134, + 129 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 134, + 129 + ], + "type": "image", + "image_path": "173057d1eb0dda57d364b4a3e1c373941501cb4d7e3c2b4a66c083925e329cff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 129, + 129, + 137 + ], + "lines": [ + { + "bbox": [ + 63, + 129, + 129, + 137 + ], + "spans": [ + { + "bbox": [ + 63, + 129, + 129, + 137 + ], + "type": "text", + "content": "(a) MAE/MSE during training" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 142, + 72, + 200, + 129 + ], + "blocks": [ + { + "bbox": [ + 142, + 72, + 200, + 129 + ], + "lines": [ + { + "bbox": [ + 142, + 72, + 200, + 129 + ], + "spans": [ + { + "bbox": [ + 142, + 72, + 200, + 129 + ], + "type": "image", + "image_path": "233930473ae392bc74e00fbda96b9d0a8622f2a75a8bf6b31af8ddc59b84c335.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 129, + 206, + 137 + ], + "lines": [ + { + "bbox": [ + 137, + 129, + 206, + 137 + ], + "spans": [ + { + "bbox": [ + 137, + 129, + 206, + 137 + ], + "type": "text", + "content": "(c) Local PSAM mean (model-L)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 72, + 290, + 129 + ], + "blocks": [ + { + "bbox": [ + 209, + 72, + 290, + 129 + ], + "lines": [ + { + "bbox": [ + 209, + 72, + 290, + 129 + ], + "spans": [ + { + "bbox": [ + 209, + 72, + 290, + 129 + ], + "type": "image", + "image_path": "f7fd1d069acbc0bce15ae0e9894c45afc251c70ce9db4ce699c4d0ec496df5d8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 129, + 285, + 137 + ], + "lines": [ + { + "bbox": [ + 214, + 129, + 285, + 137 + ], + "spans": [ + { + "bbox": [ + 214, + 129, + 285, + 137 + ], + "type": "text", + "content": "(e) aggregated PSAM (model-L)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 142, + 131, + 195 + ], + "blocks": [ + { + "bbox": [ + 63, + 142, + 131, + 195 + ], + "lines": [ + { + "bbox": [ + 63, + 142, + 131, + 195 + ], + "spans": [ + { + "bbox": [ + 63, + 142, + 131, + 195 + ], + "type": "image", + "image_path": "f4567a78e79a560022aa540449d250304ca3e352ffa4b7367bac20ea4e1bc0b9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 197, + 125, + 204 + ], + "lines": [ + { + "bbox": [ + 69, + 197, + 125, + 204 + ], + "spans": [ + { + "bbox": [ + 69, + 197, + 125, + 204 + ], + "type": "text", + "content": "(b) Sorted average PSAM" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 142, + 140, + 201, + 197 + ], + "blocks": [ + { + "bbox": [ + 142, + 140, + 201, + 197 + ], + "lines": [ + { + "bbox": [ + 142, + 140, + 201, + 197 + ], + "spans": [ + { + "bbox": [ + 142, + 140, + 201, + 197 + ], + "type": "image", + "image_path": "919694c90acae328e6ba18491b9f7b880869903d74fdf7cb9d57c6faaf6c3a68.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 198, + 207, + 203 + ], + "lines": [ + { + "bbox": [ + 135, + 198, + 207, + 203 + ], + "spans": [ + { + "bbox": [ + 135, + 198, + 207, + 203 + ], + "type": "text", + "content": "(d) Local PSAM mean (model-U)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 214, + 295, + 280 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 295, + 280 + ], + "type": "text", + "content": "Figure 3. Observations in PSAM. (a) The training process, where model-L and model-U are extracted from the 100th and 200th epochs, respectively. (b) Comparing sorted values of PSAMs in the two models. (c) & (d) Visualizing the average of local PSAM, and (e) & (f) the aggregated PSAM to compare model-L and model-U from a global perspective." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 209, + 140, + 290, + 197 + ], + "blocks": [ + { + "bbox": [ + 209, + 140, + 290, + 197 + ], + "lines": [ + { + "bbox": [ + 209, + 140, + 290, + 197 + ], + "spans": [ + { + "bbox": [ + 209, + 140, + 290, + 197 + ], + "type": "image", + "image_path": "66416986b940cbb75c8bbbfa470ac1c4a4a42040afda78c7030ca3acfbbf1fdd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 198, + 285, + 205 + ], + "lines": [ + { + "bbox": [ + 214, + 198, + 285, + 205 + ], + "spans": [ + { + "bbox": [ + 214, + 198, + 285, + 205 + ], + "type": "text", + "content": "(f) aggregated PSAM (model-U)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 290, + 181, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 181, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 181, + 302 + ], + "type": "text", + "content": "4.2. Observation in PSAM" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 309, + 295, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 295, + 345 + ], + "type": "text", + "content": "We next apply PSAM to visualize the model during semi-supervised counting. The semi-supervised loss function combines (5) and (8):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 354, + 294, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 354, + 294, + 368 + ], + "spans": [ + { + "bbox": [ + 126, + 354, + 294, + 368 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\alpha \\mathcal {L} _ {u} + (1 - \\alpha) \\mathcal {L} _ {l}, \\tag {17}", + "image_path": "580f14c6317841359322203612646ea813b0a7a5405672af2489a1c315b744a0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 376, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 376, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 376, + 295, + 506 + ], + "type": "text", + "content": " is a hyperparameter to balance the training of labeled and unlabeled data. Following [15], we set " + }, + { + "bbox": [ + 54, + 376, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 54, + 376, + 295, + 506 + ], + "type": "text", + "content": " during the first 100 epochs to generate valid pseudo labels. After that, " + }, + { + "bbox": [ + 54, + 376, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 376, + 295, + 506 + ], + "type": "text", + "content": " is gradually increased to 1. In Fig. 3(a), we plot how the MAE/MSE changes on the validation set during training. The training failed as evinced by both metrics increasing significantly when unlabeled data were involved in training (after epoch 100). We denote 'model-L' as the model at the 100-th epoch that is trained only on labeled data, and 'model-U' as the model at the 200-th epoch that was trained with both labeled and unlabeled data." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 508, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 700 + ], + "type": "text", + "content": "We next compare explanations of model-L and model-U using PSAM. We visualize the concerned regions of the crowd by extracting the PSAMs corresponding to the foreground pixels " + }, + { + "bbox": [ + 55, + 508, + 295, + 700 + ], + "type": "inline_equation", + "content": "(\\pmb{p}_{[q]} > 0.5)" + }, + { + "bbox": [ + 55, + 508, + 295, + 700 + ], + "type": "text", + "content": ". In Fig. 3(c & d) the element-wise mean of the PSAM patch " + }, + { + "bbox": [ + 55, + 508, + 295, + 700 + ], + "type": "inline_equation", + "content": "(\\sum_{q} \\tilde{H}_{[q]} \\cdot (\\pmb{p}_{[q]} > 0.5))" + }, + { + "bbox": [ + 55, + 508, + 295, + 700 + ], + "type": "text", + "content": " is displayed to illustrate how the area of the concerned regions changes between model-L and model-U, while Fig. 3(b) presents a detailed value comparison using a line chart. On one hand, Fig. 3(b) shows that the values in model-U's PSAM are larger than those of model-L; on the other hand, Fig. 3(c & d) shows that model-L's PSAM is more concentrated when compared to model-U. These observations reveal how (8) in P2P scheme affects the trained model: " + }, + { + "bbox": [ + 55, + 508, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 55, + 508, + 295, + 700 + ], + "type": "text", + "content": " leads to over-activation in the PSAM of foreground pixels, causing the decoder " + }, + { + "bbox": [ + 55, + 508, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 508, + 295, + 700 + ], + "type": "text", + "content": " to falsely identify surrounded pixels as foreground pixels as well." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Next, we visualize the global PSAM of both models" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "by compressing " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "H_{[q]}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " corresponding to foreground pixels: " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\bar{H} = \\sum_{q=1}^{n} (\\pmb{p}_{[q]} > 0.5) \\cdot H_{[q]}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ". As displayed in Fig. 3(e & f), the aggregated PSAM of model-L is sparser than that of model-U, which is consistent with the observation for individual points presented in Fig. 3(c & d). Another observation is that the background regions do not change significantly, indicating that almost all false positives occur in the neighborhood of the initial foreground pixels. This suggests that " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " does not alter the hyperplane separating crowd and non-crowd features in the high-dimensional space." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 205, + 405, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 405, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 405, + 218 + ], + "type": "text", + "content": "5. P2R Matching" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 225, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 501 + ], + "type": "text", + "content": "The observations described in last section reveal that " + }, + { + "bbox": [ + 313, + 225, + 555, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 313, + 225, + 555, + 501 + ], + "type": "text", + "content": " over-activates the values and enlarges the area of the PSAM corresponding to foreground pixels, which should be suppressed by the supervision of background part in (5), but is actually zero in (8) (" + }, + { + "bbox": [ + 313, + 225, + 555, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{u[\\mathrm{P2P}]}^{\\prime}" + }, + { + "bbox": [ + 313, + 225, + 555, + 501 + ], + "type": "text", + "content": " in (9)). This over-activation also confirms two important functions of the background item in (5): it not only guides the model to classify the crowd and non-crowd regions in the image, but also suppresses the surrounding person region from being activated. The latter is important since the neighboring pixels of a detected foreground pixel have similar features to the foreground pixel due to the CNN architecture, and thus may become activated if suppression is not applied. This interpretation makes more sense because each person in a crowd image is represented by tens to thousands of pixels, not just one. Supervising the region around the positive pixel with labels of 0 is crucial to a point-based counter. Therefore, during training, it would be more effective to adopt a point-to-region (P2R) matching strategy instead of a point-to-point (P2P) approach. Moreover, as suggested by the above interpretation, P2R is not limited in training with unlabeled data, but can also work in supervised learning with point supervision." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 510, + 506, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 506, + 523 + ], + "type": "text", + "content": "5.1. Training with Labeled Data via P2R" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 529, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 566 + ], + "type": "text", + "content": "Recalling the notation of the prediction " + }, + { + "bbox": [ + 313, + 529, + 554, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\pmb {p},\\pmb {x}\\}" + }, + { + "bbox": [ + 313, + 529, + 554, + 566 + ], + "type": "text", + "content": " in (3) and ground truth " + }, + { + "bbox": [ + 313, + 529, + 554, + 566 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\prime}" + }, + { + "bbox": [ + 313, + 529, + 554, + 566 + ], + "type": "text", + "content": ", the matching matrix " + }, + { + "bbox": [ + 313, + 529, + 554, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 529, + 554, + 566 + ], + "type": "text", + "content": " in our P2R matching is the Hadamard product of two items:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 362, + 575, + 553, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 575, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 362, + 575, + 553, + 590 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\mathbf {M} _ {f} \\odot \\left(\\boldsymbol {\\beta} ^ {\\top} \\mathbf {1} _ {m}\\right) \\quad \\in \\mathbf {M} ^ {n \\times m}. \\tag {18}", + "image_path": "5a731f0cb40e7117d063b63ede857641dcbd58927a18010162261fb0fdd301fa.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 601, + 554, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 554, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 554, + 625 + ], + "type": "text", + "content": "The first item " + }, + { + "bbox": [ + 313, + 601, + 554, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_f\\in \\{0,1\\}^{n\\times m}" + }, + { + "bbox": [ + 313, + 601, + 554, + 625 + ], + "type": "text", + "content": " is a many-to-one matrix that assigns each pixel in " + }, + { + "bbox": [ + 313, + 601, + 554, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 601, + 554, + 625 + ], + "type": "text", + "content": " to its nearest point in " + }, + { + "bbox": [ + 313, + 601, + 554, + 625 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 313, + 601, + 554, + 625 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 329, + 635, + 553, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 329, + 635, + 553, + 667 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {f [ i j ]} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} l _ {2} (i, j) < l _ {2} (i, k) \\forall k \\neq j \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {19}", + "image_path": "a08fded93d8949119cb1e7b04ddac3300995d4608ef3e55e36fbff2eb720ad68.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "l_{2}(i,j) = \\| \\pmb{x}_{[i]} - \\pmb{x}^{\\prime}_{[j]}\\|_{2}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "-distance between the " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "-th pixel's and the " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "-th point's coordinates. " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\beta \\in \\{0,1\\}^{n}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " in the second item of (18) is a vector marking foreground" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29367" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 119, + 125 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 119, + 125 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 119, + 125 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 119, + 125 + ], + "type": "image", + "image_path": "2d3037d526dbec96c083a70d5cfaa3c38f5a7a8f5ba86833d86b4c558aeb5176.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 126, + 107, + 133 + ], + "lines": [ + { + "bbox": [ + 70, + 126, + 107, + 133 + ], + "spans": [ + { + "bbox": [ + 70, + 126, + 107, + 133 + ], + "type": "text", + "content": "(a) P2P scheme" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 126, + 73, + 203, + 125 + ], + "blocks": [ + { + "bbox": [ + 126, + 73, + 203, + 125 + ], + "lines": [ + { + "bbox": [ + 126, + 73, + 203, + 125 + ], + "spans": [ + { + "bbox": [ + 126, + 73, + 203, + 125 + ], + "type": "image", + "image_path": "5a89ef1fd0767b95f729c31ee59a7693f87f94e9c88ef75a51e6879443359794.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 126, + 186, + 133 + ], + "lines": [ + { + "bbox": [ + 142, + 126, + 186, + 133 + ], + "spans": [ + { + "bbox": [ + 142, + 126, + 186, + 133 + ], + "type": "text", + "content": "(b) matched point" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 212, + 73, + 290, + 125 + ], + "blocks": [ + { + "bbox": [ + 212, + 73, + 290, + 125 + ], + "lines": [ + { + "bbox": [ + 212, + 73, + 290, + 125 + ], + "spans": [ + { + "bbox": [ + 212, + 73, + 290, + 125 + ], + "type": "image", + "image_path": "6e9ca855a9338a9e281301059c52e330acd103eea4f4505e3158c10f3895bced.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 126, + 283, + 133 + ], + "lines": [ + { + "bbox": [ + 219, + 126, + 283, + 133 + ], + "spans": [ + { + "bbox": [ + 219, + 126, + 283, + 133 + ], + "type": "text", + "content": "(c) semi-supervised points" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 62, + 137, + 116, + 192 + ], + "blocks": [ + { + "bbox": [ + 62, + 137, + 116, + 192 + ], + "lines": [ + { + "bbox": [ + 62, + 137, + 116, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 116, + 192 + ], + "type": "image", + "image_path": "d52126ae431bda47c3298a8fe18f63c4f40d36ec9102cd1bf43b1fdfb4460d0d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 192, + 108, + 199 + ], + "lines": [ + { + "bbox": [ + 70, + 192, + 108, + 199 + ], + "spans": [ + { + "bbox": [ + 70, + 192, + 108, + 199 + ], + "type": "text", + "content": "(d) P2R scheme" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 209, + 296, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 308 + ], + "type": "text", + "content": "Figure 4. Difference between P2P and P2R matching. (a) and (d) demonstrate an overall difference. P2P focuses only on foreground pixels in " + }, + { + "bbox": [ + 55, + 209, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_s" + }, + { + "bbox": [ + 55, + 209, + 296, + 308 + ], + "type": "text", + "content": ", while P2R considers background pixels as well. (b) and (e) show how the matching is performed in P2P and P2R, respectively. P2R segments out local regions for each pseudo-label, whereas P2P only detects one point. (c) and (f) illustrate how untrusted predictions are filtered. P2P retains only foreground pixels for loss computation, while P2R also keeps pixels in the neighborhoods of their corresponding pseudo points." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 126, + 139, + 203, + 190 + ], + "blocks": [ + { + "bbox": [ + 126, + 139, + 203, + 190 + ], + "lines": [ + { + "bbox": [ + 126, + 139, + 203, + 190 + ], + "spans": [ + { + "bbox": [ + 126, + 139, + 203, + 190 + ], + "type": "image", + "image_path": "50797048b980f15b33ddf80a206c725012f898e4239eb929921c4e3236208965.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 192, + 187, + 200 + ], + "lines": [ + { + "bbox": [ + 141, + 192, + 187, + 200 + ], + "spans": [ + { + "bbox": [ + 141, + 192, + 187, + 200 + ], + "type": "text", + "content": "(e) matched region" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 212, + 138, + 290, + 191 + ], + "blocks": [ + { + "bbox": [ + 212, + 138, + 290, + 191 + ], + "lines": [ + { + "bbox": [ + 212, + 138, + 290, + 191 + ], + "spans": [ + { + "bbox": [ + 212, + 138, + 290, + 191 + ], + "type": "image", + "image_path": "b4774abf9ee95739c14124767740efce4801d5d91c06433a4fb03a77b5d4168b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 192, + 284, + 200 + ], + "lines": [ + { + "bbox": [ + 217, + 192, + 284, + 200 + ], + "spans": [ + { + "bbox": [ + 217, + 192, + 284, + 200 + ], + "type": "text", + "content": "(f) semi-supervised regions" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 326, + 295, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 295, + 349 + ], + "type": "text", + "content": "regions, by setting those pixels in " + }, + { + "bbox": [ + 55, + 326, + 295, + 349 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 326, + 295, + 349 + ], + "type": "text", + "content": " far from any points in " + }, + { + "bbox": [ + 55, + 326, + 295, + 349 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 55, + 326, + 295, + 349 + ], + "type": "text", + "content": " as zero:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 96, + 353, + 294, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 353, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 96, + 353, + 294, + 384 + ], + "type": "interline_equation", + "content": "\\beta_ {[ i ]} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\min _ {j} l _ {2} (i, j) < \\mu \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {20}", + "image_path": "3036cf2154158af1d91a882573f8ed78de4e2da58750287626f1763fb53bc2b7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 389, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 461 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 55, + 389, + 296, + 461 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 389, + 296, + 461 + ], + "type": "text", + "content": " is a hyper parameter indicating the radius if the neighborhood of each GT point is considered as a circle. As displayed in Fig. 4(e), " + }, + { + "bbox": [ + 55, + 389, + 296, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 389, + 296, + 461 + ], + "type": "text", + "content": " segments out a coarse local region that is consistent with the location of the GT point. After that, the learning objective " + }, + { + "bbox": [ + 55, + 389, + 296, + 461 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}" + }, + { + "bbox": [ + 55, + 389, + 296, + 461 + ], + "type": "text", + "content": " is derived by defining the cost matrix " + }, + { + "bbox": [ + 55, + 389, + 296, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{n \\times m}" + }, + { + "bbox": [ + 55, + 389, + 296, + 461 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 71, + 464, + 295, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 464, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 71, + 464, + 295, + 496 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {[ i j ]} = \\left\\{ \\begin{array}{l l} \\tau l _ {2} (i, j) - \\mathcal {S} (\\boldsymbol {p} _ {[ i ]}), & \\text {i f} \\mathbf {M} _ {[ i j ]} = 1 \\\\ \\infty , & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {21}", + "image_path": "a2cc80def19a97ecc000f14d58206652b477494997c77211baffaa96e1c9b8e5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 501, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 525 + ], + "type": "text", + "content": "Then, the pixel with the minimum cost in each column of " + }, + { + "bbox": [ + 55, + 501, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 55, + 501, + 295, + 525 + ], + "type": "text", + "content": " is marked as the potential foreground pixel:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 530, + 294, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 530, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 89, + 530, + 294, + 563 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} _ {[ i j ]} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\mathbf {C} _ {[ i j ]} < \\mathbf {C} _ {[ k j ]} \\forall k \\neq i \\\\ 0, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {22}", + "image_path": "a13ebcae5fc30dea9c52d2cea40aa0c029501631b43e37ae11353d351b527b19.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": "Finally, the learning objective " + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": " is derived as " + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}} = \\hat{\\mathbf{M}}\\mathbf{1}" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": ", and the loss " + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_l" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": " is computed between " + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": " according to BCE loss formulated as (5) during training with labeled data." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": "An extra benefit is that computing " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": " in P2R requires only finding the minimum of each row or column in the matrices, which is significantly faster than executing the Hungarian algorithm — a time-consuming component of the P2P matching strategy [57]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 671, + 257, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 257, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 257, + 684 + ], + "type": "text", + "content": "5.2. Training with Unlabeled Data via P2R" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Recalling the pseudo-label " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t^\\prime = \\{\\pmb {p}_t^\\prime ,\\pmb{x}_t^\\prime \\}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " from the teacher and the supervised student output as " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_s = \\{\\pmb {p}_s,\\pmb {x}_s\\}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ", the" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "text", + "content": "matching matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{st}" + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_s" + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t^\\prime" + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "text", + "content": " is computed via (18) in P2R, and then combined with the confidence in (7) to mask unreliable pseudo labels. Additionally, pixels far from any pseudo points can be directly marked as reliable background pixels, as suggested by observations from the aggregated PSAM in Fig. 3(e & f). Thus, the confidence diagonal matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "text", + "content": " is formulated as follows with " + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "inline_equation", + "content": "\\zeta = \\mathbb{1}(\\hat{\\pmb{p}}_t' > \\eta)" + }, + { + "bbox": [ + 313, + 72, + 553, + 157 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 372, + 162, + 553, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 162, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 372, + 162, + 553, + 175 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} = \\operatorname {d i a g} \\left[ \\mathbf {M} _ {s t} \\boldsymbol {\\zeta} + \\left(\\mathbf {1} _ {n} - \\boldsymbol {\\beta}\\right) \\right] \\tag {23}", + "image_path": "5af7cd2307f9d07711543ac8b9d7b6a16ee540e7066b29bd358521b90c27c5e0.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 181, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 205 + ], + "type": "text", + "content": "The final loss supervising unlabeled data " + }, + { + "bbox": [ + 313, + 181, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_s" + }, + { + "bbox": [ + 313, + 181, + 553, + 205 + ], + "type": "text", + "content": " with pseudo labels is derived by substituting " + }, + { + "bbox": [ + 313, + 181, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 181, + 553, + 205 + ], + "type": "text", + "content": " in (23) into (8)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 205, + 554, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 554, + 300 + ], + "type": "text", + "content": "As displayed in Fig. 4(f), the confidence for each pseudo point is propagated to it corresponding region. On one hand, reliable foreground parts (blue) contain a pixel whose pseudo-label is 1, while the pseudo-labels of its surrounding pixels are marked as 0. On the other hand, all pixels in the reliable background parts (khaki) are assigned pseudolabels of 0 to train the counter with unlabeled data. The uncolored regions are ignored because their confidence is 0." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 300, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 324 + ], + "type": "text", + "content": "The pseudo code and flow chart of the proposed P2R loss can be found in the supplemental material." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 334, + 395, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 395, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 395, + 347 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 354, + 554, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 413 + ], + "type": "text", + "content": "This section explores how the proposed P2R training approach enhances a point-based counting model using both labeled and unlabeled data across three parts: semi-supervised crowd counting, unsupervised domain adaptation (UDA) for crowd counting, and ablation studies." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "text", + "content": "In the semi-supervised experiments, our method is implemented on four crowd counting datasets: ShTech A/B [73], UCF-QNRF [12], and JHU++ [55]. Following DAC [67] and OT-M [26], three protocols are applied: " + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "text", + "content": " of labeled data, with the remaining crowd images involved in training without annotations. In the UDA part, labeled data from one domain, e.g., ShTech A, is used to initialize a counting model, while unlabeled data from another domain is utilized to capture the domain characteristics of the application scenes. In the ablation study, we test different hyperparameters to demonstrate their impact on the performance of P2R." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 563, + 463, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 463, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 463, + 576 + ], + "type": "text", + "content": "6.1. Semi-Supervised Counting" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "Tab. 1 presents the experimental results of semi-supervised counting. Our method achieves the best performance in nearly all protocols across all crowd datasets. Besides, using " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " data with P2R surpasses the fully-supervised learning with " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " data and is even equivalent to other semi-supervised learning methods using " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " data. Similar performances are also observed between the semi-supervised training with " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " labeled data. These results demonstrate that P2R is label-efficient and advantageous in semi-supervised learning. In Fig. 5, we visualize predictions of DAC [67], OTM [26], and our P2R. The 3rd row is" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29368" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 274 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 274 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 274 + ], + "type": "table", + "html": "
Label Pct.MethodsShTech A [73]ShTech B [73]UCF-QNRF [12]JHU++ [55]
MAEMSEMAEMSEMAEMSEMAEMSE
5%label-only93.7155.213.124.0132.7231.7104.1383.5
MT [59]104.7156.919.333.2172.4284.9101.5363.5
L2R [32]103.0155.420.327.6160.1272.3101.4338.8
GP [54]102.0172.015.727.9160.0275.098.9355.7
DAC [23]85.4134.512.622.8120.2209.382.2294.9
OT-M [26]83.7133.312.621.5118.4195.482.7304.5
P2R (ours)69.9119.59.116.6100.1182.577.8293.5
10%label-only84.0138.310.719.2112.4186.878.7305.5
MT [59]319.394.515.624.5156.1245.5250.390.2
L2R [32]90.3153.515.624.4148.9249.887.5315.3
IRAST [34]86.9148.914.722.9135.6233.486.7303.4
DAC [23]74.9115.511.119.1109.0187.275.9282.3
OT-M [26]80.1118.510.818.2113.1186.773.0280.6
CU [18]70.7116.69.717.7104.01644.274.8281.6
P2R (ours)65.2114.68.414.594.9167.268.7272.3
40%label-only64.5105.68.114.099.2174.768.8283.5
MT [59]88.2151.115.925.7147.2249.6121.5388.9
L2R [32]86.5148.216.825.1145.1256.1123.6376.1
SUA [42]68.5121.914.120.6130.3226.380.7290.8
DAC [23]67.5110.79.614.691.1153.465.1260.0
OT-M [26]70.7114.58.113.1100.6167.672.1272.0
P2R (ours)55.695.06.811.086.0144.363.3271.1
100%P2PNet [57]52.7485.066.259.9085.32154.5061.25258.65
P2R (ours)51.0279.686.179.8483.26138.1158.83253.10
", + "image_path": "a11eb0bc84848c91c552d284bdb6c7bf1fefc58f892941bca0691e5b30fb7253.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 339 + ], + "type": "text", + "content": "a failure case of P2R due to crowd motion blur. However, the overall performance of P2R is best." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 340, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 340, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 295, + 472 + ], + "type": "text", + "content": "The last row of Tab. 1 also compares P2P and P2R under a fully-supervised learning scheme (100% labels); P2R performs better than P2PNet on all crowd counting datasets. Additionally, as described in Section 5.1, P2R is faster than P2P due to the absence of the Hungarian algorithm. We compared the efficiency using an image with a resolution of " + }, + { + "bbox": [ + 54, + 340, + 295, + 472 + ], + "type": "inline_equation", + "content": "576 \\times 960" + }, + { + "bbox": [ + 54, + 340, + 295, + 472 + ], + "type": "text", + "content": " and 775 annotated points — P2P requires an average of 0.4307 seconds for loss computation, while P2R only needs 0.0064 seconds, which is nearly 68 times faster than P2P. These comparisons demonstrate that P2R surpasses P2P in both effectiveness and efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 271, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 271, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 271, + 493 + ], + "type": "text", + "content": "6.2. Unsupervised Domain Adaptation (UDA)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "UDA aims to transfer knowledge learned from a source domain to a target domain by designing modules or learning objectives that capture domain-agnostic features [11, 29, 43, 65, 70, 72]. Unlike semi-supervised learning, where labeled and unlabeled data originate from the same domain, UDA typically involves a domain gap between the labeled (source domain) and unlabeled crowd data (target domain). This gap often manifests as differences in crowd distribution, density levels, and perspectives. Our P2R can also be applied to UDA. When an unlabeled image from the target domain is provided, the threshold automatically partitions its corresponding pseudo points into two categories: higher-score " + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "(> \\eta)" + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": " and lower-score " + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "(< \\eta)" + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": ". The higher-score class is considered to be closer to the source domain, while the lower-score class is considered to be farther from the source domain. Thus, the higher-score pseudo points can participate in the training to gradually guide the model to learn to count in the target domain." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 375, + 103 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 375, + 103 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 375, + 103 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 375, + 103 + ], + "type": "image", + "image_path": "d25312d39243a6b6634ea424acc16f50e830886b2550e10fa04726d36a4ba1ad.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 103, + 374, + 148 + ], + "blocks": [ + { + "bbox": [ + 316, + 103, + 374, + 148 + ], + "lines": [ + { + "bbox": [ + 316, + 103, + 374, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 103, + 374, + 148 + ], + "type": "image", + "image_path": "8e4bca1ce1a802d635973ce4b53dc29485be1ceabda551efa9d41edfb3489c10.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 148, + 374, + 187 + ], + "blocks": [ + { + "bbox": [ + 316, + 148, + 374, + 187 + ], + "lines": [ + { + "bbox": [ + 316, + 148, + 374, + 187 + ], + "spans": [ + { + "bbox": [ + 316, + 148, + 374, + 187 + ], + "type": "image", + "image_path": "498debb671b3a4edee2dceac2f04c8a36334121c31d2a059ebc01fe31e84e8ff.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 202, + 528, + 214 + ], + "lines": [ + { + "bbox": [ + 339, + 202, + 528, + 214 + ], + "spans": [ + { + "bbox": [ + 339, + 202, + 528, + 214 + ], + "type": "text", + "content": "Figure 5. qualitative comparison with other models." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 188, + 362, + 194 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 188, + 362, + 194 + ], + "spans": [ + { + "bbox": [ + 326, + 188, + 362, + 194 + ], + "type": "text", + "content": "ground truth: 61" + } + ] + } + ], + "index": 9, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 376, + 70, + 434, + 103 + ], + "blocks": [ + { + "bbox": [ + 376, + 70, + 434, + 103 + ], + "lines": [ + { + "bbox": [ + 376, + 70, + 434, + 103 + ], + "spans": [ + { + "bbox": [ + 376, + 70, + 434, + 103 + ], + "type": "image", + "image_path": "1254c483cece1cfc09f2ac61fdd5d89d1a529df00416c2721d340b5daa89f315.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 376, + 103, + 434, + 148 + ], + "blocks": [ + { + "bbox": [ + 376, + 103, + 434, + 148 + ], + "lines": [ + { + "bbox": [ + 376, + 103, + 434, + 148 + ], + "spans": [ + { + "bbox": [ + 376, + 103, + 434, + 148 + ], + "type": "image", + "image_path": "94a12bc5c589ddd0a8f7af83e499663ce4398dcbea20e6710ee1212e237c7c05.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 148, + 422, + 154 + ], + "lines": [ + { + "bbox": [ + 388, + 148, + 422, + 154 + ], + "spans": [ + { + "bbox": [ + 388, + 148, + 422, + 154 + ], + "type": "text", + "content": "DAC [23]: 49.21" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 436, + 70, + 494, + 103 + ], + "blocks": [ + { + "bbox": [ + 436, + 70, + 494, + 103 + ], + "lines": [ + { + "bbox": [ + 436, + 70, + 494, + 103 + ], + "spans": [ + { + "bbox": [ + 436, + 70, + 494, + 103 + ], + "type": "image", + "image_path": "9650b28896edc859802b98431ddb2a87918b2f43ebe30de7e800abc36ac7d799.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 436, + 103, + 494, + 148 + ], + "blocks": [ + { + "bbox": [ + 436, + 103, + 494, + 148 + ], + "lines": [ + { + "bbox": [ + 436, + 103, + 494, + 148 + ], + "spans": [ + { + "bbox": [ + 436, + 103, + 494, + 148 + ], + "type": "image", + "image_path": "361c13ec83e61e4f11807d326fba288c08177559bbfdefc61ad9a6457f522c39.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 148, + 482, + 154 + ], + "lines": [ + { + "bbox": [ + 449, + 148, + 482, + 154 + ], + "spans": [ + { + "bbox": [ + 449, + 148, + 482, + 154 + ], + "type": "text", + "content": "OTM [26]: 51.43" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 498, + 70, + 555, + 103 + ], + "blocks": [ + { + "bbox": [ + 498, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 498, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 498, + 70, + 555, + 103 + ], + "type": "image", + "image_path": "1ab0ea4ddece539f20bd8786803703c3bf75987758b349dfb9d6ae8c9b6f23e2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 498, + 103, + 555, + 148 + ], + "blocks": [ + { + "bbox": [ + 498, + 103, + 555, + 148 + ], + "lines": [ + { + "bbox": [ + 498, + 103, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 498, + 103, + 555, + 148 + ], + "type": "image", + "image_path": "1aa919225f6be1983959b821c70c72260f1e5fc7cd54f8da446701be6a0d5191.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 148, + 543, + 154 + ], + "lines": [ + { + "bbox": [ + 511, + 148, + 543, + 154 + ], + "spans": [ + { + "bbox": [ + 511, + 148, + 543, + 154 + ], + "type": "text", + "content": "P2R (ours): 60.0" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 154, + 374, + 187 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 374, + 187 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 374, + 187 + ], + "type": "image", + "image_path": "3c41bc60c93abdbcf25816c96adb7392151f73273a5558f6204e74d9bb3942d8.jpg" + } + ] + } + ], + "index": 19, + "type": "text" + }, + { + "bbox": [ + 326, + 188, + 362, + 194 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 188, + 362, + 194 + ], + "spans": [ + { + "bbox": [ + 326, + 188, + 362, + 194 + ], + "type": "text", + "content": "ground truth: 61" + } + ] + } + ], + "index": 20, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 376, + 154, + 434, + 187 + ], + "blocks": [ + { + "bbox": [ + 376, + 154, + 434, + 187 + ], + "lines": [ + { + "bbox": [ + 376, + 154, + 434, + 187 + ], + "spans": [ + { + "bbox": [ + 376, + 154, + 434, + 187 + ], + "type": "image", + "image_path": "4299101a1c5146f624ae2d673bd7bb7b8f9db033834b5a9418e45b907948dc6d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 188, + 423, + 194 + ], + "lines": [ + { + "bbox": [ + 387, + 188, + 423, + 194 + ], + "spans": [ + { + "bbox": [ + 387, + 188, + 423, + 194 + ], + "type": "text", + "content": "DAC [23]: 146.09" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 436, + 154, + 494, + 187 + ], + "blocks": [ + { + "bbox": [ + 436, + 154, + 494, + 187 + ], + "lines": [ + { + "bbox": [ + 436, + 154, + 494, + 187 + ], + "spans": [ + { + "bbox": [ + 436, + 154, + 494, + 187 + ], + "type": "image", + "image_path": "e81e68ee2a8eaf4974274886e9d870487d019eb84db173bb908cb81c627302f0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 188, + 484, + 194 + ], + "lines": [ + { + "bbox": [ + 446, + 188, + 484, + 194 + ], + "spans": [ + { + "bbox": [ + 446, + 188, + 484, + 194 + ], + "type": "text", + "content": "OTM [26]: 126.11" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 498, + 154, + 555, + 187 + ], + "blocks": [ + { + "bbox": [ + 498, + 154, + 555, + 187 + ], + "lines": [ + { + "bbox": [ + 498, + 154, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 498, + 154, + 555, + 187 + ], + "type": "image", + "image_path": "8be1375a8f5baece94489667dadbd13ab3a9b4e5a45376da504be7396a81025a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 188, + 544, + 193 + ], + "lines": [ + { + "bbox": [ + 507, + 188, + 544, + 193 + ], + "spans": [ + { + "bbox": [ + 507, + 188, + 544, + 193 + ], + "type": "text", + "content": "P2R (ours): 227.0" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 323, + 224, + 547, + 314 + ], + "blocks": [ + { + "bbox": [ + 55, + 283, + 295, + 306 + ], + "lines": [ + { + "bbox": [ + 55, + 283, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 295, + 306 + ], + "type": "text", + "content": "Table 1. Comparison with other recent methods on four benchmark datasets under different labeled protocols." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 224, + 547, + 314 + ], + "lines": [ + { + "bbox": [ + 323, + 224, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 323, + 224, + 547, + 314 + ], + "type": "table", + "html": "
METHODSA → BA → QB → AB → Q
MAEMSEMAEMSEMAEMSEMAEMSE
DGDCCUS [40]12.624.6119.4216.6121.8203.1179.1316.2
MPCount [43]11.419.7115.7199.899.6182.9165.6290.4
UDABL [38] (w/o DA)15.925.8166.7287.6138.1228.1226.4411.0
RBT [35]13.429.3175.0294.8112.2218.2211.3381.9
C2MoT [70]12.421.1125.7218.3120.7192.0198.9368.0
FGFD [79]12.723.3124.1242.0123.5210.7209.7384.7
FSIM [80]11.119.3105.3191.1120.3202.6194.9324.5
UDAP2R (w/o DA)25.635.7155.1286.8130.2229.1173.3329.8
P2R (w/ DA)10.618.7105.3194.889.3176.0139.5243.1
", + "image_path": "c668c9539047841ba65c01e5f2b0bb059a0f592f70d0e0a73a8f4689b169cb23.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "type": "text", + "content": "Table 2. Unsupervised domain adaptation for crowd counting. The LHS and RHS of “" + }, + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "type": "text", + "content": "” represent adaptation direction. Q, A, and B indicate UCF-QNRF [12], ShTech A, and ShTech B [73]. " + }, + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "type": "inline_equation", + "content": "DG" + }, + { + "bbox": [ + 313, + 323, + 555, + 378 + ], + "type": "text", + "content": " indicates domain generalization. The training of UDA requires samples from the target domain, whereas DG does not." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 389, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 544 + ], + "type": "text", + "content": "We conduct experiments on four protocols following [70]. The results are presented in Tab. 2. The last two rows demonstrate the results of the proposed P2R. The poor performance obtained without domain adaptation highlights the domain gap between the source data and target data. However, when the unlabeled data from the target domain is involved in training, the estimation errors are significantly reduced, and the results surpass previous methods with complex network structures for domain adaptation. These advanced experimental results demonstrate that the threshold works as expected to capture samples close to the source domain and gradually teaches the model to count in the target domain." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 552, + 408, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 408, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 408, + 564 + ], + "type": "text", + "content": "6.3. Ablation Study" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": "We next conduct ablation studies to investigate the influence of hyperparameters in P2R, including " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": " in (21), the threshold " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": " in (20), and " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": " in (17). The results are in Fig. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": "The impact of " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": " in (21). " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": " is a hyperparameter to balance the distance and foreground score when selecting the foreground pixel from the matched region. When " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tau \\rightarrow \\infty" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": ", the predicted score is ignored, which is equivalent to computing cross-entropy between the prediction and the ground truth point map. If " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": ", the pixel with the highest score is chosen as the foreground in " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": ". However, Fig. 6(a) shows that neither extreme is optimal. The best performance is achieved when " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 8" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29369" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 68, + 172, + 142 + ], + "blocks": [ + { + "bbox": [ + 66, + 68, + 172, + 142 + ], + "lines": [ + { + "bbox": [ + 66, + 68, + 172, + 142 + ], + "spans": [ + { + "bbox": [ + 66, + 68, + 172, + 142 + ], + "type": "image", + "image_path": "5d5fe63da2bf1fddb316df5f7d8e3c7e7678febfa5a7b3271468130cdbee91bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 144, + 149, + 155 + ], + "lines": [ + { + "bbox": [ + 88, + 144, + 149, + 155 + ], + "spans": [ + { + "bbox": [ + 88, + 144, + 149, + 155 + ], + "type": "text", + "content": "(a) errors vs. " + }, + { + "bbox": [ + 88, + 144, + 149, + 155 + ], + "type": "inline_equation", + "content": "\\tau" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 156, + 172, + 227 + ], + "blocks": [ + { + "bbox": [ + 66, + 156, + 172, + 227 + ], + "lines": [ + { + "bbox": [ + 66, + 156, + 172, + 227 + ], + "spans": [ + { + "bbox": [ + 66, + 156, + 172, + 227 + ], + "type": "image", + "image_path": "c9c7fe401fb6d177465c000264ef0308194a1cb1d7e70f6fe4536c9358ad99b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 229, + 149, + 240 + ], + "lines": [ + { + "bbox": [ + 89, + 229, + 149, + 240 + ], + "spans": [ + { + "bbox": [ + 89, + 229, + 149, + 240 + ], + "type": "text", + "content": "(c) errors vs. " + }, + { + "bbox": [ + 89, + 229, + 149, + 240 + ], + "type": "inline_equation", + "content": "\\mu" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 251, + 296, + 294 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 294 + ], + "type": "text", + "content": "Figure 6. The estimation errors vs. various hyperparameters. (a) is conducted on the fully-supervised ShTech A since it is irrelevant to semi-supervised counting, while (b)-(d) are conducted with " + }, + { + "bbox": [ + 55, + 251, + 296, + 294 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 251, + 296, + 294 + ], + "type": "text", + "content": " labeled data in ShTech A." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 68, + 286, + 142 + ], + "blocks": [ + { + "bbox": [ + 179, + 68, + 286, + 142 + ], + "lines": [ + { + "bbox": [ + 179, + 68, + 286, + 142 + ], + "spans": [ + { + "bbox": [ + 179, + 68, + 286, + 142 + ], + "type": "image", + "image_path": "1490810d57fd7ef78bcffb49f22d8a9721c65052b843f722653ce7a89f89bfc2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 179, + 156, + 286, + 227 + ], + "blocks": [ + { + "bbox": [ + 202, + 144, + 263, + 156 + ], + "lines": [ + { + "bbox": [ + 202, + 144, + 263, + 156 + ], + "spans": [ + { + "bbox": [ + 202, + 144, + 263, + 156 + ], + "type": "text", + "content": "(b) errors vs. " + }, + { + "bbox": [ + 202, + 144, + 263, + 156 + ], + "type": "inline_equation", + "content": "\\eta" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 156, + 286, + 227 + ], + "lines": [ + { + "bbox": [ + 179, + 156, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 179, + 156, + 286, + 227 + ], + "type": "image", + "image_path": "0072ff971013eced25b31451a39f06e40e677af0efbb691d4afd072e6a0db34a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 228, + 268, + 243 + ], + "lines": [ + { + "bbox": [ + 196, + 228, + 268, + 243 + ], + "spans": [ + { + "bbox": [ + 196, + 228, + 268, + 243 + ], + "type": "text", + "content": "(d) errors vs. " + }, + { + "bbox": [ + 196, + 228, + 268, + 243 + ], + "type": "inline_equation", + "content": "\\frac{\\alpha}{1 - \\alpha}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 304, + 295, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 295, + 435 + ], + "type": "text", + "content": "The effect of threshold " + }, + { + "bbox": [ + 54, + 304, + 295, + 435 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 304, + 295, + 435 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 304, + 295, + 435 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 304, + 295, + 435 + ], + "type": "text", + "content": " defines the threshold to distinguish reliable and unreliable regions. Since all pseudo points are selected by picking pixels whose score is greater than 0.5 in the prediction, the threshold should also be greater than 0.5. Moreover, a lower " + }, + { + "bbox": [ + 54, + 304, + 295, + 435 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 304, + 295, + 435 + ], + "type": "text", + "content": " introduces unreliable pseudo-labels into training, while a higher " + }, + { + "bbox": [ + 54, + 304, + 295, + 435 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 304, + 295, + 435 + ], + "type": "text", + "content": " cannot fully utilize the unlabeled data as most parts cannot participate in training. As shown in Fig. 6(b), the estimation errors exhibit a trend of first decreasing and then increasing within the range of 0.6 to 0.9, with the lowest errors occurring when the threshold is set to 0.7." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": "The radius of each point " + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": ". The vector " + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": " in (20) derived from " + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": " marks the foreground points' neighborhood and background regions. It not only restricts the positions where foreground pixels appear by substituting " + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": " into (18), but also directly identifies reliable background parts via the second item in (23) based on the observation of PSAM in Section 4.2. Fig. 6(c) presents how " + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": " affects the counting performance. It shows the MAE remains around 68 when " + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": " increases from 32 to 256, indicating that a larger " + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": " does not significantly affect the results. This is consistent with the observation in PSAM since pixels far from the foreground points are less likely to be activated during training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "The weight of unlabeled loss. Fig. 6(d) presents how " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " in (17) affects the performance. A large " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " focuses more on unlabeled data, leading to training failure, while a small " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " may cause the model to overfit on labeled data. In the experiments with ShTech A and " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " labeled data, " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\frac{\\alpha}{1 - \\alpha} = 2" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 2/3" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ", results in the best performance. Additionally, as long as " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " is not 0, the point-based counter can benefit from unlabeled data, as the maximum MAE/MSE in Fig. 6(d) are smaller than 80/150, which is much better than the model trained with only " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " labeled data (MAE: 93.7, MSE: 155.2)." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 322, + 75, + 445, + 168 + ], + "blocks": [ + { + "bbox": [ + 322, + 75, + 445, + 168 + ], + "lines": [ + { + "bbox": [ + 322, + 75, + 445, + 168 + ], + "spans": [ + { + "bbox": [ + 322, + 75, + 445, + 168 + ], + "type": "image", + "image_path": "bc9d5517a41f3760c947c215bfb6cc4dc5fb4769bbcd349a78272a90d6a583e2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 172, + 423, + 180 + ], + "lines": [ + { + "bbox": [ + 352, + 172, + 423, + 180 + ], + "spans": [ + { + "bbox": [ + 352, + 172, + 423, + 180 + ], + "type": "text", + "content": "(a) sorted average PSAM" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 455, + 75, + 539, + 106 + ], + "blocks": [ + { + "bbox": [ + 455, + 75, + 539, + 106 + ], + "lines": [ + { + "bbox": [ + 455, + 75, + 539, + 106 + ], + "spans": [ + { + "bbox": [ + 455, + 75, + 539, + 106 + ], + "type": "image", + "image_path": "18c3907376122314226fbf8f929c8556edd5c45d55abd39e848e55802195a990.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 456, + 107, + 539, + 137 + ], + "blocks": [ + { + "bbox": [ + 456, + 107, + 539, + 137 + ], + "lines": [ + { + "bbox": [ + 456, + 107, + 539, + 137 + ], + "spans": [ + { + "bbox": [ + 456, + 107, + 539, + 137 + ], + "type": "image", + "image_path": "0171d89ef65055b33551da679ddf963d531fbdc140881281682d438984b1449f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 456, + 139, + 539, + 169 + ], + "blocks": [ + { + "bbox": [ + 456, + 139, + 539, + 169 + ], + "lines": [ + { + "bbox": [ + 456, + 139, + 539, + 169 + ], + "spans": [ + { + "bbox": [ + 456, + 139, + 539, + 169 + ], + "type": "image", + "image_path": "b4bec7e249dfc042fb1836bbfc195dd36705f4f2e00267923a0339ba8e0217ba.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 172, + 541, + 180 + ], + "lines": [ + { + "bbox": [ + 456, + 172, + 541, + 180 + ], + "spans": [ + { + "bbox": [ + 456, + 172, + 541, + 180 + ], + "type": "text", + "content": "(b) Local & aggregated PSAM" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 191, + 554, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 191, + 554, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 554, + 247 + ], + "type": "text", + "content": "Figure 7. The comparison of PSAMs among different training schemes. Model-L (green) is trained with only labeled data; model-U_P2P (blue) is trained with both labeled data and unlabeled data using the P2P scheme, while model-U_P2R (orange) is trained with the proposed P2R scheme." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 255, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 399 + ], + "type": "text", + "content": "PSAM comparison. Fig. 7 visualizes the PSAMs of models trained with the P2P and P2R schemes. Comparing model-L (green) and model-U_P2R (blue), the latter effectively restricts the PSAM to the neighborhood of the concerned foreground pixel, and its aggregated PSAM is also sparser than that of model-L, demonstrating the effectiveness of P2R. Comparing model-U_P2P and model-U_P2R in Fig. 7(a), the values of the PSAM are similar at the non-peak indexes, but the latter has a higher activation at the peak position compared to the former, indicating that semi-supervised learning further enhances the identification ability of the point-based counter." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 410, + 388, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 410, + 388, + 422 + ], + "spans": [ + { + "bbox": [ + 314, + 410, + 388, + 422 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 430, + 555, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 681 + ], + "type": "text", + "content": "This paper presents a P2R scheme to train a point-based counter for semi-supervised crowd counting. At the beginning, we establish a pseudo-labeling framework based on P2P matching, but its formulation is ill-posed since the confidence is only applicable to foreground pixels and cannot be propagated to backgrounds. To observe how the foreground and background feature are learned, we propose PSAM to visually interpret the concerned region and activation value of all foreground pixels. From the visualized activation map, it is observed that the neighborhood of each foreground pixel is over-activated and falsely recognized as instances by the decoder. Based on this observation, we replace P2P with P2R by segmenting the prediction into multiple regions containing the corresponding instances. Thus, pixels in the neighborhood of the concerned pseudo points can share the weights with the corresponding pixels. Additionally, the time-consuming Hungarian algorithm is no longer necessary in P2R. Several outstanding experimental results in semi-supervised counting and unsupervised domain adaptation demonstrate the advantages of our P2R matching strategy." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by a grant from City University of Hong Kong (Project No. 7030010)." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29370" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Shahira Abousamra, Minh Hoai, Dimitris Samaras, and Chao Chen. Localization in the crowd with topological constraints. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 872-881, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Areej Althali, Amal Balabid, Reem Alharthi, Bander Alzahrani, Reem Alotaibi, and Ahmed Barnawi. Anomalous event detection and localization in dense crowd scenes. Multimedia Tools and Applications, 82(10):15673-15694, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "text", + "content": "[3] Deepak Babu Sam, Shiv Surya, and R Venkatesh Babu. Switching convolutional neural network for crowd counting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5744-5752, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 278 + ], + "type": "text", + "content": "[4] Antoni B Chan and Nuno Vasconcelos. Bayesianoisson regression for crowd counting. In 2009 IEEE 12th international conference on computer vision, pages 545-551. IEEE, 2009. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 313 + ], + "type": "text", + "content": "[5] Antoni B Chan and Nuno Vasconcelos. Counting people with low-level features and bayesian regression. IEEE Transactions on image processing, 21(4):2160-2177, 2011. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 367 + ], + "type": "text", + "content": "[6] Antoni B Chan, Zhang-Sheng John Liang, and Nuno Vasconcelos. Privacy preserving crowd monitoring: Counting people without people models or tracking. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-7. IEEE, 2008. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "text", + "content": "[7] Mulin Chen, Qi Wang, and Xuelong Li. Anchor-based group detection in crowd scenes. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1378-1382. IEEE, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "text", + "content": "[8] Zhi-Qi Cheng, Qi Dai, Hong Li, Jingkuan Song, Xiao Wu, and Alexander G Hauptmann. Rethinking spatial invariance of convolutional networks for object counting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19638-19648, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 470, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 294, + 502 + ], + "type": "text", + "content": "[9] Junyu Gao, Tao Han, Qi Wang, Yuan Yuan, and Xuelong Li. Learning independent instance maps for crowd localization. arXiv preprint arXiv:2012.04164, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 503, + 295, + 557 + ], + "type": "text", + "content": "[10] Tao Han, Lei Bai, Lingbo Liu, and Wanli Ouyang. Steerer: Resolving scale variations for counting and localization via selective inheritance learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21848-21859, 2023. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "type": "text", + "content": "[11] Yuhang He, Zhiheng Ma, Xing Wei, Xiaopeng Hong, Wei Ke, and Yihong Gong. Error-aware density isomorphism reconstruction for unsupervised cross-domain crowd counting. In Proceedings of the AAAI conference on artificial intelligence, pages 1540-1548, 2021. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 678 + ], + "type": "text", + "content": "[12] Haroon Idrees, Muhammad Tayyab, Kishan Athrey, Dong Zhang, Somaya Al-Maadeed, Nasir Rajpoot, and Mubarak Shah. Composition loss for counting, density map estimation and localization in dense crowds. In Proceedings of the European conference on computer vision (ECCV), pages 532-546, 2018. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[13] Brian Kenji Iwana, Ryohei Kuroki, and Seiichi Uchida. Explaining convolutional neural networks using softmax gradient layer-wise relevance propagation. In 2019 IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision Workshop (ICCVW), pages 4176-4185. IEEE, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[14] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[15] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[16] Bastian Leibe, Edgar Seemann, and Bernt Schiele. Pedestrian detection in crowded scenes. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 878-885. IEEE, 2005. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "type": "text", + "content": "[17] Victor Lempitsky and Andrew Zisserman. Learning to count objects in images. Advances in neural information processing systems, 23, 2010. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "text", + "content": "[18] Chen Li, Xiaoling Hu, Shahira Abousamra, and Chao Chen. Calibrating uncertainty for semi-supervised crowd counting. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16685-16695. IEEE, 2023. 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "text", + "content": "[19] Hanxiao Li, Yonghong Song, and Tong Geng. Semi-supervised crowd counting based on hard pseudo-labels. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "type": "text", + "content": "[20] Yuhong Li, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1091-1100, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "text", + "content": "[21] Dingkang Liang, Wei Xu, and Xiang Bai. An end-to-end transformer model for crowd localization. In ECCV, pages 38-54. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "type": "text", + "content": "[22] Hui Lin, Xiaopeng Hong, Zhiheng Ma, Xing Wei, Yunfeng Qiu, Yaowei Wang, and Yihong Gong. Direct measure matching for crowd counting. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages = 837–844, publisher = ijcai.org, year = 2021, .2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "text", + "content": "[23] Hui Lin, Zhiheng Ma, Xiaopeng Hong, Yaowei Wang, and Zhou Su. Semi-supervised crowd counting via density agency. In ACM Multimedia, 2022. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[24] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19628-19637, 2022. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[25] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, Zhou Su, Xiaopeng Hong, and Deyu Meng. Semi-supervised counting via pixel-by-pixel density distribution modelling. arXiv preprint arXiv:2402.15297, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[26] Wei Lin and Antoni B Chan. Optimal transport minimization: Crowd localization on density maps for semi-supervised counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21663-21673, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[27] Wei Lin and Antoni B Chan. A fixed-point approach to unified prompt-based counting. In Proceedings of the" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29371" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "text", + "content": "AAAI Conference on Artificial Intelligence, pages 3468-3476, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[28] Wei Lin, Jia Wan, and Antoni B Chan. Proximal mapping loss: Understanding loss functions in crowd counting & localization. In The Thirteenth International Conference on Learning Representations, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[29] Wei Lin, Junyu Gao, Qi Wang, and Xuelong Li. Learning to detect anomaly events in crowd scenes from synthetic data. Neurocomputing, 436:248-259, 2021. 1, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 219 + ], + "type": "text", + "content": "[30] Wei Lin, Kunlin Yang, Xinzhu Ma, Junyu Gao, Lingbo Liu, Shinan Liu, Jun Hou, Shuai Yi, and Antoni B Chan. Scale-prior deformable convolution for exemplar-guided class-agnostic counting. In BMVC, page 313, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 253 + ], + "type": "text", + "content": "[31] Chengxin Liu, Hao Lu, Zhiguo Cao, and Tongliang Liu. Point-query quadtree for crowd counting, localization, and more. In CVPR, pages 1676-1685, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 255, + 295, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 295, + 298 + ], + "type": "text", + "content": "[32] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Leveraging unlabeled data for crowd counting by learning to rank. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7661-7669, 2018. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 300, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 342 + ], + "type": "text", + "content": "[33] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Exploiting unlabeled data in cnns by self-supervised learning to rank. IEEE transactions on pattern analysis and machine intelligence, 41(8):1862-1878, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 388 + ], + "type": "text", + "content": "[34] Yan Liu, Lingqiao Liu, Peng Wang, Pingping Zhang, and Yinjie Lei. Semi-supervised crowd counting via self-training on surrogate tasks. In European Conference on Computer Vision, pages 242-259. Springer, 2020. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 389, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 295, + 443 + ], + "type": "text", + "content": "[35] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. In Proceedings of the 28th ACM International Conference on Multimedia, pages 129-137, 2020. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 488 + ], + "type": "text", + "content": "[36] Yanbo Liu, Yingxiang Hu, Guo Cao, and Yanfeng Shang. Semi-supervised crowd counting via multi-task pseudo-label self-correction strategy. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "type": "text", + "content": "[37] Zheng Ma and Antoni B Chan. Counting people crossing a line using integer programming and local features. IEEE Transactions on Circuits and Systems for Video Technology, 26(10):1955-1969, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "text", + "content": "[38] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In Proceedings of the IEEE International Conference on Computer Vision, pages 6142-6151, 2019. 1, 2, 7, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "type": "text", + "content": "[39] Zhiheng Ma, Xing Wei, Xiaopeng Hong, Hui Lin, Yunfeng Qiu, and Yihong Gong. Learning to count via unbalanced optimal transport. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2319-2327, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "text", + "content": "[40] Lucas Mansilla, Rodrigo Echeveste, Diego H Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6630-6638, 2021. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[41] Riccardo Mazzon, Fabio Poiesi, and Andrea Cavallaro. Detection and tracking of groups in crowd. In 2013 10th IEEE International conference on advanced video and signal based surveillance, pages 202-207. IEEE, 2013. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[42] Yanda Meng, Hongrun Zhang, Yitian Zhao, Xiaoyun Yang, Xuesheng Qian, Xiaowei Huang, and Yalin Zheng. Spatial uncertainty-aware semi-supervised crowd counting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15549-15559, 2021. 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[43] Zhuoxuan Peng and S-H Gary Chan. Single domain generalization for crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28025-28034, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "type": "text", + "content": "[44] Harish Guruprasad Ramaswamy et al. Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 983-991, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "text", + "content": "[45] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3394-3403, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 273, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 326 + ], + "type": "text", + "content": "[46] Weihong Ren, Xinchao Wang, Jiandong Tian, Yandong Tang, and Antoni B Chan. Tracking-by-counting: Using network flows on crowd density maps for tracking multiple targets. IEEE Transactions on Image Processing, 30:1439-1452, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "type": "text", + "content": "[47] Jihye Ryu and Kwangho Song. Crowd counting and individual localization using pseudo square label. IEEE Access, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 360, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 425 + ], + "type": "text", + "content": "[48] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "text", + "content": "[49] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 323-331, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "text", + "content": "[50] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19618-19627, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "text", + "content": "[51] Weibo Shu, Jia Wan, Kay Chen Tan, Sam Kwong, and Antoni B Chan. Crowd counting in the frequency domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19618-19627, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "text", + "content": "[52] Weibo Shu, Jia Wan, and Antoni B Chan. Generalized characteristic function loss for crowd analysis in the frequency domain. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "text", + "content": "[53] Karen Simonyan. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 637, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 637, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 637, + 553, + 689 + ], + "type": "text", + "content": "[54] Vishwanath A Sindagi, Rajeev Yasarla, Deepak Sam Babu, R Venkatesh Babu, and Vishal M Patel. Learning to count in the crowd from limited labeled data. In European Conference on Computer Vision, pages 212-229. Springer, 2020. 2, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[55] Vishwanath A Sindagi, Rajeev Yasarla, and Vishal M Patel. Jhu-crowd++: Large-scale crowd counting dataset and a" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29372" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "benchmark method. IEEE transactions on pattern analysis and machine intelligence, 44(5):2594-2609, 2020. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 160 + ], + "type": "text", + "content": "[56] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 216 + ], + "type": "text", + "content": "[57] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In CVPR, pages 3365-3374, 2021. 1, 2, 3, 6, 7, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "text", + "content": "[58] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin A. Riedmiller. Striving for simplicity: The all convolutional net. In 3rd International Conference on Learning Representations, ICLR 2015, Workshop Track Proceedings, 2015. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 271, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 294, + 316 + ], + "type": "text", + "content": "[59] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "text", + "content": "[60] Jia Wan, Qingzhong Wang, and Antoni B Chan. Kernel-based density map generation for dense object counting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1357-1370, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "text", + "content": "[61] Jia Wan, Ziquan Liu, and Antoni B Chan. A generalized loss function for crowd counting and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1974-1983, 2021. 1, 2, 3, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 449 + ], + "type": "text", + "content": "[62] Jia Wan, Qiangqiang Wu, Wei Lin, and Antoni Chan. Robust zero-shot crowd counting and localization with adaptive resolution sam. In European Conference on Computer Vision, pages 478-495. Springer, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[63] Boyu Wang, Huidong Liu, Dimitris Samaras, and Minh Hoai Nguyen. Distribution matching for crowd counting. Advances in neural information processing systems, 33:1595-1607, 2020. 2, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 557 + ], + "type": "text", + "content": "[64] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "text", + "content": "[65] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8198-8207, 2019. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 647 + ], + "type": "text", + "content": "[66] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. Nwpu-crowd: A large-scale benchmark for crowd counting and localization. IEEE transactions on pattern analysis and machine intelligence, 43(6):2141-2149, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "text", + "content": "[67] Qi Wang, Wei Lin, Junyu Gao, and Xuelong Li. Density-aware curriculum learning for crowd counting. IEEE Transactions on Cybernetics, 52(6):4675-4687, 2020. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[68] Qi Wang, Wei Lin, Junyu Gao, and Xuelong Li. Density-aware curriculum learning for crowd counting. IEEE Transactions on Cybernetics, 52(6):4675-4687, 2020. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 654 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[69] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Pixel-wise crowd understanding via synthetic data. International Journal of Computer Vision, 129(1):225-245, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[70] Qiangqiang Wu, Jia Wan, and Antoni B Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. In Proceedings of the 29th ACM International Conference on Multimedia, pages 658-666, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "type": "text", + "content": "[71] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15548-15557, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 240 + ], + "type": "text", + "content": "[72] Qi Zhang, Wei Lin, and Antoni B Chan. Cross-view cross-scene multi-view crowd counting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 557-567, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 297 + ], + "type": "text", + "content": "[73] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 589-597, 2016. 1, 2, 6, 7, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "text", + "content": "[74] Chenyang Zhao and Antoni B. Chan. ODAM: gradient-based instance-specific visual explanations for object detection. In ICLR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 397 + ], + "type": "text", + "content": "[75] Chunhui Zhao, Zhiyuan Zhang, Jinwen Hu, Dong Wang, Bin Fan, Quan Pan, and Qiang He. Crowd anomaly event detection in surveillance video based on the evolution of the spatial position relationship feature. In 2018 IEEE 14th International Conference on Control and Automation (ICCA), pages 247-252, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 398, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 452 + ], + "type": "text", + "content": "[76] Chenyang Zhao, Janet H Hsiao, and Antoni B Chan. Gradient-based instance-specific visual explanations for object specification and object discrimination. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 453, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 498 + ], + "type": "text", + "content": "[77] Chenyang Zhao, Kun Wang, Xingyu Zeng, Rui Zhao, and Antoni B Chan. Gradient-based visual explanation for transformer-based clip. In International Conference on Machine Learning, pages 61072-61091. PMLR, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "text", + "content": "[78] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 554, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 609 + ], + "type": "text", + "content": "[79] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross domain crowd counting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5659-5668, 2022. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 609, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 609, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 609, + 553, + 654 + ], + "type": "text", + "content": "[80] Huilin Zhu, Jingling Yuan, Xian Zhong, Liang Liao, and Zheng Wang. Find gold in sand: Fine-grained similarity mining for domain-adaptive crowd counting. IEEE Transactions on Multimedia, 2023. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29373" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/d3a3a010-6834-441e-beb3-f3f784483984_content_list.json b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/d3a3a010-6834-441e-beb3-f3f784483984_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..509c8bf32fa8351714573207dce2600da1b90a5a --- /dev/null +++ b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/d3a3a010-6834-441e-beb3-f3f784483984_content_list.json @@ -0,0 +1,2284 @@ +[ + { + "type": "text", + "text": "Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances", + "text_level": 1, + "bbox": [ + 135, + 125, + 864, + 171 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi Yu $^{1*}$ , Botao Ren $^{2*}$ , Peiyuan Zhang $^{3*}$ , Mingxin Liu $^{4}$ , Junwei Luo $^{3}$ , Shaofeng Zhang $^{4}$ , Feipeng Da $^{1}$ , Junchi Yan $^{4}$ , Xue Yang $^{4\\dagger}$", + "bbox": [ + 135, + 190, + 861, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Southeast University $^{2}$ Tsinghua University $^{3}$ Wuhan University $^{4}$ Department of Automation & CSE & SAI, Shanghai Jiao Tong University", + "bbox": [ + 205, + 227, + 794, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yuyi,dafp}@seu.edu.cn rbt22@mails.tsinghua.edu.cn {peiyuanzhangwhu,luojunwei}@whu.edu.cn", + "bbox": [ + 106, + 265, + 892, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liumingxin,zhangshaofeng,yanjunchi, yangxue-2019-sjtu}@sjtu.edu.cn", + "bbox": [ + 207, + 280, + 790, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/VisionXLab/point2rbox-v2", + "bbox": [ + 305, + 294, + 689, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 334, + 326, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapidly increasing demand for oriented object detection (OOD), recent research involving weakly-supervised detectors for learning OOD from point annotations has gained great attention. In this paper, we rethink this challenging task setting with the layout among instances and present Point2RBox-v2. At the core are three principles: 1) Gaussian overlap loss. It learns an upper bound for each instance by treating objects as 2D Gaussian distributions and minimizing their overlap. 2) Voronoi watershed loss. It learns a lower bound for each instance through watershed on Voronoi tessellation. 3) Consistency loss. It learns the size/rotation variation between two output sets with respect to an input image and its augmented view. Supplemented by a few devised techniques, e.g. edge loss and copy-paste, the detector is further enhanced. To our best knowledge, Point2RBox-v2 is the first approach to explore the spatial layout among instances for learning point-supervised OOD. Our solution is elegant and lightweight, yet it is expected to give a competitive performance especially in densely packed scenes: $62.61\\% /86.15\\% /34.71\\%$ on DOTA/HRSC/FAIR1M.", + "bbox": [ + 88, + 367, + 485, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 688, + 220, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emerged as an essential task in computer vision, oriented object detection (OOD) has become a prominent demand in autonomous driving [6], aerial images [7, 30, 48, 51, 52], scene text [23, 28, 72], retail scenes [10, 35], industrial inspection [29, 47], and more.", + "bbox": [ + 89, + 713, + 483, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Manual annotations are essential to teach the detector new concepts of visual objects. Early research is often super", + "bbox": [ + 89, + 789, + 483, + 820 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/312565b9515f1af73282525218512fdb213bafa5407c82fb3751e9830a45ac65.jpg", + "image_caption": [ + "Figure 1. Related methods, their principles for knowledge mining, whether using additional priors, and performance on DOTA-v1.0." + ], + "image_footnote": [], + "bbox": [ + 516, + 334, + 905, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vised by rotated bounding boxes (RBoxes), where annotations align with the desired output. While this approach has demonstrated promising results, RBox labels are considerably expensive. The cost of annotating each RBox is approximately $36.5\\%$ higher than a horizontal bounding box (HBox) and $104.8\\%$ higher than a point [66]. This highlights the importance of utilizing coarser labels for weakly-supervised oriented detection. Recent advancements in HBox-supervised OOD, particularly with H2RBox [59] and H2RBox-v2 [65], have shown promise in bridging the gap between HBox-and RBox-supervised methods, thereby lessening the dependence on labor-intensive RBox labeling.", + "bbox": [ + 511, + 551, + 908, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following HBox-supervised methods, the more challenging point-supervised OOD opens up a new avenue. Several innovative approaches have emerged in the past year (see Fig. 1): 1) Point-prompt OOD. P2RBox [2], PMHO [69], and PointSAM [27] employ the zero-shot ability of SAM [17]. 2) Pseudo generation. PointOBB [33] and its v2/v3 version [38, 68] use multiple instance learning and class probability map for RBox generation. 3) Knowledge combination. Point2RBox [66] learns from one-shot examples. However, existing methods overlook the spatial relationships between objects. This brings us to the motivation of this paper.", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding author. The work was in part supported by National Natural Science Foundation of China (62306069, 62222607, 72342023), Major Science and Technology Project of Jiangsu Province (BG2024003), and China Postdoctoral Science Foundation (2023M740602). Yi Yu is also supported by Jiangsu Funding Program for Excellent Postdoctoral Talent (2023ZB616).", + "bbox": [ + 89, + 827, + 485, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19283", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b63c1ba009eff7f950c198770e57cd82ae19f8188beced4bc1bfb304c4dd30bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 89, + 267, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bf2e91e935cbc8d25a72d01b0a5f5748c83cb7adeb106676947de0a84f178c68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 90, + 421, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b60dab802c0e0688e36305437db8d85bc1053c441c799691fa9caa51e807c684.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 90, + 573, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/172107eaaa86ce0c6fe6de5bdeb54f7a291ca8116068561a1bb8984311d12b73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 90, + 725, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a28d2a33aaf6741c75ba6fd7d0f75bdf78427d22b2ec892c5049d5bbcb093f88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 90, + 879, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/776a7aff52f3b88b79e08e9a00a9ffe80ae4a8d915e0b9b268d7990201fae575.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 207, + 267, + 323 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ef2e656aa692901b0e17f433fb4c56053618fc8bbeb2ed5d01caa3e3515f6f9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 207, + 419, + 323 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9d72e47dbcea8085f9d67829db6eeca8a209d158d6523e0fce9346adbe4c6bb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 207, + 573, + 323 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/da7da2d16ef61360469eb00e29d5c9501b0b9206f87873470b38813b913a8066.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 207, + 725, + 323 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a627dc09b71c35b05b41399cff5a9b067e79b77869fbba346af16badad417758.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 207, + 879, + 323 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/300ce1cd207b24d07a541bd9e4577ebacf925f091e39392bfdff2eb76cd3a151.jpg", + "image_caption": [ + "Figure 2. Visual comparisons with state-of-the-art methods including: Point2Mask (2023) [21], PointOBB (2024) [33], PointOBB-v2 (2025) [38], PointOBB-v3 (2025) [68], and Point2RBox (2024) [66]. The boxes detected by our method (last row) wrap the objects more tightly." + ], + "image_footnote": [], + "bbox": [ + 117, + 325, + 267, + 441 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a465eef682ae40c9eddd692767959461c348b5b7d30a7bec551e1c65cd45edca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 325, + 419, + 440 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1f315e80353531f669466f8a0b9b01944960bcacdae9df85b38af33ab04a8483.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 325, + 573, + 440 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cc63f69098b5c6a2217ed56b43849dfbfd441be35c9a7caf06efa322f88822c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 325, + 725, + 440 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/32f2371c429606c2eacbae6fad263eb1f0335f9ffd21efe8ccb5a28850ca27dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 325, + 879, + 440 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivation. While point-supervised OOD has gained attention, the utilization of relationships among instances remains absent in the literature. For example, the arrangement of vehicles in a parking lot can serve as effective constraints for learning their size and orientation. This approach may yield valuable insights into point-supervised OOD, particularly in densely packed scenes where current methods face significant challenges. Can we leverage the spatial layout of objects to enhance point-supervised OOD? In this paper, we dive into this idea and seek to answer this question.", + "bbox": [ + 88, + 492, + 482, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Highlights. 1) Point2RBox-v2 is proposed for point-supervised OOD, advancing the state of the art as displayed in Fig. 2 and Tables 1-2. 2) We propose novel and elegant losses to enforce constraints from the spatial layout among instances based on Gaussian overlap and Voronoi tessellation [1]. 3) Other modules are devised (i.e. edge loss, consistency loss, copy-paste) to further enhance the method.", + "bbox": [ + 89, + 643, + 482, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. 1) To our best knowledge, this work is the first attempt to learn point-supervised OOD from the layout among objects, where we propose novel principles based on Gaussian overlap and Voronoi tessellation. 2) The training pipeline and detailed implementation are elucidated, with necessary modules (i.e. edge loss, consistency loss, copy-paste) incorporated. The source code will be made publicly available. 3) Extensive experiments demonstrate that leveraging the spatial layout of instances can significantly advance the state of the art, surpassing other alternatives in accuracy.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 489, + 653, + 506 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. RBox-supervised Oriented Detection", + "text_level": 1, + "bbox": [ + 511, + 515, + 830, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to horizontal detection [26, 71], oriented object detection (OOD) [46, 52] has received extensive attention. Representative works include anchor-based detector Rotated RetinaNet [25], anchor-free detector Rotated FCOS [42], and two-stage solutions, e.g. RoI Transformer [5], Oriented R-CNN [49], and ReDet [11]. Some research enhances the detector by exploiting alignment features, e.g. $\\mathrm{R}^3\\mathrm{Det}$ [55] and $\\mathrm{S}^2\\mathrm{A}$ -Net [12]. The angle regression may face boundary discontinuity and remedies are developed, including modulated losses [37, 53, 58] that alleviate loss jumps, angle coders [50, 51, 54, 63] that convert the angle into boundary-free coded data, and Gaussian-based losses [34, 56, 57, 60, 61] transforming rotated bounding boxes into Gaussian distributions. RepPoint-based methods [15, 20, 62] provide alternatives that predict a set of points that bounds the spatial extent of an object. LMMRotate [19] is a new paradigm of OOD based on multimodal language model.", + "bbox": [ + 511, + 537, + 906, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Point-supervised Oriented Detection", + "text_level": 1, + "bbox": [ + 511, + 803, + 828, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, several methods for point-supervised OOD have been proposed: 1) P2RBox [2], PMHO [69], and PointSAM [27] propose point-prompt OOD by employing the zero-shot Point-to-Mask ability of SAM [17]. 2) Point2RBox [66] introduces an end-to-end approach based on knowledge com", + "bbox": [ + 511, + 825, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19284", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7f7ab96800cefc9e462732b3d097e16a231a98bbb2684aa785ee6bb13d3a5146.jpg", + "image_caption": [ + "Figure 3. The training pipeline of Point2RBox-v2. Gaussian overlap loss and Voronoi watershed loss utilize the spatial layout (see Fig. 4), while edge loss (see Sec. 3.4), symmetry-aware learning (see Sec. 3.5), and copy-paste (see Sec. 3.6) further enhance the method." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 893, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bination. 3) PointOBB [33, 68] achieves RBox generation through multiple instance learning. 4) PointOBB-v2 [38] learns a class probability map to generate pseudo RBoxes. 5) Wholly-WOOD [67] uses points, HBoxes, RBoxes, or their combinations for mixed supervised OOD.", + "bbox": [ + 88, + 351, + 480, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Among these methods, P2RBox, PMHO, and PointSAM rely on the SAM model pre-trained on massive labeled datasets, whereas Point2RBox requires one-shot examples for each category. In contrast, PointOBB series do not use many priors, but they necessitate two-stage training.", + "bbox": [ + 88, + 428, + 482, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Other Weakly-supervised Settings", + "text_level": 1, + "bbox": [ + 89, + 513, + 387, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to Point-to-RBox, some other settings have been better studied. These methods are potentially applicable to our Point-to-RBox task setting by using a cascade pipeline, such as Point-to-HBox-to-RBox. In our experiment, cascade pipelines powered by state-of-the-art approaches are also compared. Here, representative works are introduced.", + "bbox": [ + 88, + 537, + 482, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HBox-to-RBox. H2RBox [59] establishes a paradigm that limits the object to a few candidate angles through geometric constraint from HBoxes, with a self-supervised branch eliminating the undesired results. An enhanced version H2RBox-v2 [65] is proposed to leverage the reflection symmetry of objects to further boost the accuracy. EIE-Det [45] uses an explicit equivariance branch for learning rotation consistency, and an implicit equivariance branch for learning position, aspect ratio, and scale consistency. KCR [74] combines RBox- and HBox-annotated datasets for transfer learning. Some studies [16, 41] use additional annotated data for training, which are also attractive but less general.", + "bbox": [ + 88, + 628, + 482, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point-to-HBox. Several related approaches have been developed, including: 1) P2BNet [3] samples box proposals of different sizes around the labeled point and classify them to achieve point-supervised horizontal object detection. 2) PSOD [8] achieves point-supervised salient object detection using an edge detector and adaptive masked flood fill.", + "bbox": [ + 88, + 810, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point-to-Mask. Point2Mask [21] is proposed to achieve panoptic segmentation using single point annotation per target. SAM (Segment Anything Model) [17] produces object masks from input point/HBox prompts. Though RBoxes can be obtained from the segmentation mask by finding the minimum circumscribed rectangle, such a complex pipeline can be less cost-efficient and perform worse [59, 65].", + "bbox": [ + 511, + 351, + 906, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 470, + 604, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview and Preliminary", + "text_level": 1, + "bbox": [ + 511, + 496, + 754, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of Point2RBox-v2 is illustrated in Fig. 3. The network is based on ResNet50 [13] backbone, FPN [24] head, and PSC [64] angle coder. Objects of varying sizes are typically assigned to different FPN layers based on the scale. However, points lack size, so we assign them all to the P3 layer (stride = 8). Assume the detector $f_{\\mathrm{nn}}(\\cdot)$ maps an image $I$ to a set of RBoxes as detection results:", + "bbox": [ + 511, + 517, + 905, + 623 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(x _ {c}, y _ {c}, w, h, \\theta\\right) = f _ {\\mathrm {n n}} (I) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 630, + 906, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Equivalently, oriented objects can also be represented by 2D Gaussian distributions $\\mathcal{N}(\\mu, \\Sigma)$ [60]:", + "bbox": [ + 511, + 654, + 905, + 685 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mu = \\left[ \\begin{array}{l l} x _ {c} & y _ {c} \\end{array} \\right], \\quad \\Sigma = \\mathbf {R} \\left[ \\begin{array}{c c} w / 2 & 0 \\\\ 0 & h / 2 \\end{array} \\right] ^ {2} \\mathbf {R} ^ {\\top} \\right\\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 690, + 906, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 511, + 731, + 558, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} = \\left[ \\begin{array}{c c} \\cos \\theta & - \\sin \\theta \\\\ \\sin \\theta & \\cos \\theta \\end{array} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 741, + 906, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At the core of Point2RBox-v2 is to utilize the constraints from the layout, which is achieved by Gaussian overlap loss (Sec. 3.2) and Voronoi watershed loss (Sec. 3.3). These two losses effectively limit the size and rotation of objects to a reasonable range. Upon that, edge loss (Sec. 3.4) aligns the bounding box with the edge of objects to improve the accuracy. Incorporated with symmetry-aware learning (Sec. 3.5) and copy-paste augmentation (Sec. 3.6), we achieve a", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19285", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e493037026f870df4b6d99c842b610fed2377006c2b5ec8358a853e93a79d1fc.jpg", + "image_caption": [ + "Figure 4. To illustrate the procedure of the three newly proposed loss functions and their impact on the learning results. (a) Gaussian overlap loss (see Sec. 3.2). (b) Voronoi watershed loss (see Sec. 3.3). (c) Edge loss (see Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 883, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "stable and high-accuracy solution for point-supervised OOD. In subsequent subsections, these modules are detailed.", + "bbox": [ + 89, + 522, + 485, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Gaussian Overlap Loss", + "text_level": 1, + "bbox": [ + 89, + 566, + 305, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned, oriented objects can be represented by 2D Gaussian distributions $\\mathcal{N}(\\mu, \\Sigma)$ . The overlap volume between two distributions $\\mathcal{N}_1(\\mu_1, \\Sigma_1)$ and $\\mathcal{N}_2(\\mu_2, \\Sigma_2)$ can be approximated by the Bhattacharyya coefficient [60] as:", + "bbox": [ + 89, + 590, + 483, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nB \\left(\\mathcal {N} _ {1}, \\mathcal {N} _ {2}\\right) = \\exp \\left(- \\frac {1}{8} \\mu^ {\\top} \\Sigma^ {- 1} \\mu\\right) \\cdot \\frac {\\left| \\Sigma_ {1} \\right| ^ {1 / 4} \\left| \\Sigma_ {2} \\right| ^ {1 / 4}}{\\left| \\Sigma \\right| ^ {1 / 2}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 656, + 483, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu = \\mu_{2} - \\mu_{1}$ , $\\Sigma = \\frac{1}{2} (\\Sigma_{1} + \\Sigma_{2})$ , $|\\Sigma|$ denotes the determinant of the covariance matrix.", + "bbox": [ + 89, + 696, + 482, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above equation, we build a Gaussian overlap matrix $\\mathbf{M} \\in \\mathbb{R}^{N \\times N}$ for each scene image as:", + "bbox": [ + 89, + 729, + 482, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i, j} = B \\left(\\mathcal {N} _ {i}, \\mathcal {N} _ {j}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 766, + 482, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i,j = 1,2,\\ldots ,N;N$ is the distribution count (or the instance count) within one training image.", + "bbox": [ + 89, + 789, + 482, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Gaussian overlap loss can then be expressed as:", + "bbox": [ + 109, + 821, + 455, + 837 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {O} = \\frac {1}{N} \\sum_ {i \\neq j} \\left(\\mathbf {M} _ {i, j}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 842, + 483, + 880 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i \\neq j$ omits diagonal elements. With this loss, the", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "detector learns to arrange instances (see Fig. 4a) based on the mutual exclusivity among instances.", + "bbox": [ + 511, + 522, + 906, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Voronoi Watershed Loss", + "text_level": 1, + "bbox": [ + 511, + 561, + 738, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A Voronoi diagram [1] is a partitioning of a space based on a set of points. Point annotations can be effectively utilized to calculate a Voronoi diagram, where a distinct polygon region is assigned to each point-annotated instance. The watershed algorithm [44], on the other hand, is a region segmentation technique that treats the intensity of pixels as a topographic surface, identifying regions as \"catchment basins\".", + "bbox": [ + 511, + 583, + 905, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The calculation of Voronoi ridges can be formulated as:", + "bbox": [ + 532, + 689, + 898, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV = \\text {V o r o n o i} (X) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 710, + 906, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X$ are the annotated points within a training image; $V$ are the output Voronoi ridges (pixel coordinates).", + "bbox": [ + 511, + 734, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interestingly, we find that Voronoi diagrams can be utilized as initial markers for watershed to obtain a region for each instance (see Fig. 4b). In concrete terms, the points $X$ can be employed as foreground markers, while the Voronoi ridges $V$ can act as background boundaries:", + "bbox": [ + 511, + 765, + 906, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS = \\text {W a t e r s h e d} (I, X, V) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 847, + 906, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S$ are the output basin regions (pixel coordinates) corresponding to each annotated instance. By rotating $S$", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19286", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to align with the direction of the current prediction, the regression target of width and height can be expressed as:", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l} w _ {t} \\\\ h _ {t} \\end{array} \\right] = 2 \\max \\left| \\mathbf {R} ^ {\\top} \\left(S - \\left[ \\begin{array}{l} x _ {c} \\\\ y _ {c} \\end{array} \\right]\\right) \\right| \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 126, + 483, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(x_{c},y_{c},w,h,\\theta)$ is the current prediction; $\\mathbf{R}$ is defined by Eq. (3); $w_{t}$ and $h_t$ are detached to stop the gradient.", + "bbox": [ + 89, + 166, + 482, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterward, the Voronoi watershed loss to regress the width and height of objects can be calculated as:", + "bbox": [ + 89, + 196, + 483, + 226 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {W} = L _ {\\mathrm {G W D}} \\left(\\left[ \\begin{array}{c c} w / 2 & 0 \\\\ 0 & h / 2 \\end{array} \\right] ^ {2}, \\left[ \\begin{array}{c c} w _ {t} / 2 & 0 \\\\ 0 & h _ {t} / 2 \\end{array} \\right] ^ {2}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 232, + 483, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{\\mathrm{GWD}}(\\cdot)$ is Gaussian Wasserstein Distance Loss [60].", + "bbox": [ + 89, + 279, + 483, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Edge Loss", + "text_level": 1, + "bbox": [ + 89, + 303, + 207, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above two losses have limited the size to a reasonable range. To make it more accurate, we propose the edge loss to snap the boundaries toward the edges (see Fig. 4c).", + "bbox": [ + 89, + 325, + 483, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, a region $P \\in \\mathbb{R}^{(2K + 1) \\times (2K + 1)}$ around each predicted RBox is extracted via Rotated RoI Align [14]:", + "bbox": [ + 89, + 369, + 483, + 401 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP = R o I A l i g n (E (I), \\left(x _ {c}, y _ {c}, \\beta w, \\beta h, \\theta\\right)) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 407, + 483, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E(\\cdot)$ is the edge detection function [39]. We set $K = 24$ and $\\beta = 1.6$ in our experiments (see Table 9).", + "bbox": [ + 89, + 431, + 482, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By calculating the sum of each row of $P$ , the edge distribution in $y$ direction can be obtained as:", + "bbox": [ + 89, + 460, + 483, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {i} = \\sum_ {j = 1} ^ {2 K + 1} \\left(P _ {(K + 1 - i), j} + P _ {(K + 1 + i), j}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 497, + 483, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i = 1,2,\\dots ,K$ , indicating that the upper half of $P$ is reversed and added to the lower half.", + "bbox": [ + 89, + 545, + 482, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meanwhile, the current prediction of the edge can also be softened into a distribution as:", + "bbox": [ + 89, + 575, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {i} = \\exp \\left(- \\frac {(i - K / \\beta) ^ {2}}{2 \\sigma_ {E} {} ^ {2}}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 608, + 483, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i = 1,2,\\ldots ,K;\\sigma_{E}$ is set to 6. Note that we crop $P$ based on the predicted RBox, thus $\\lambda$ is always the same.", + "bbox": [ + 89, + 648, + 482, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multiplying the two distributions yields the target:", + "bbox": [ + 109, + 679, + 442, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = \\frac {\\beta h}{K} \\arg \\max (\\mu \\times \\lambda) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 698, + 483, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h_t$ is the regression target of height. Likewise, the width target $w_t$ is calculated along $x$ direction.", + "bbox": [ + 89, + 733, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The edge loss is then obtained as:", + "bbox": [ + 109, + 763, + 334, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {E} = \\operatorname {s m o o t h} _ {L 1} \\left(\\left[ \\begin{array}{l l} w & h \\end{array} \\right], \\left[ \\begin{array}{l l} w _ {t} & h _ {t} \\end{array} \\right]\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 784, + 483, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $\\mathcal{L}_W$ and $\\mathcal{L}_E$ merely refines the width and height of boxes, without involving the angle regression.", + "bbox": [ + 89, + 809, + 482, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Symmetry-aware Learning", + "text_level": 1, + "bbox": [ + 89, + 848, + 334, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proven effective in learning the rotation of objects [65], in this work, symmetry-aware learning is extended to Gaussian-", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "based OOD. In the left part of Fig. 3, the training image $I$ is transformed (randomly selected from rotation, flip, and scale, see Table 7) to generate an augmented view as:", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {a u g}} = \\alpha I \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 674, + 143, + 906, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 511, + 167, + 558, + 179 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = \\left[ \\begin{array}{c c} \\cos \\mathcal {R} & - \\sin \\mathcal {R} \\\\ \\sin \\mathcal {R} & \\cos \\mathcal {R} \\end{array} \\right] ^ {p _ {1}} \\left[ \\begin{array}{c c} 1 & 0 \\\\ 0 & - 1 \\end{array} \\right] ^ {p _ {2}} \\left[ \\begin{array}{c c} s & 0 \\\\ 0 & s \\end{array} \\right] ^ {p _ {3}} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 184, + 906, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(p_1, p_2, p_3) = (1, 0, 0)$ when rotation is selected, $(0, 1, 0)$ when flip, $(0, 0, 1)$ when scale. $\\mathcal{R}$ and $s$ are the random amount of rotation and scale, $s \\in (0.5, 0.9)$ .", + "bbox": [ + 511, + 224, + 905, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By feeding both $I$ and $I_{\\mathrm{aug}}$ into the network, we obtain two sets of output Gaussian distributions and angles:", + "bbox": [ + 511, + 270, + 905, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} (\\Sigma , \\theta) = f _ {\\mathrm {n n}} (I) \\\\ (\\Sigma_ {\\mathrm {a u g}}, \\theta_ {\\mathrm {a u g}}) = f _ {\\mathrm {n n}} (I _ {\\mathrm {a u g}}) \\end{array} \\right. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 306, + 905, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The consistency loss is calculated between the two sets so that the network learns the size and rotation variation:", + "bbox": [ + 511, + 345, + 905, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s s}} = L _ {\\mathrm {G W D}} \\left(\\alpha \\Sigma \\alpha^ {\\top}, \\Sigma_ {\\mathrm {a u g}}\\right) + L _ {\\mathrm {A N G}} (m \\theta + \\mathcal {R}, \\theta_ {\\mathrm {a u g}}) \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 381, + 906, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ is the rotation angle and $m = 1$ when rotation is selected; $(m,\\mathcal{R}) = (-1,0)$ when flip; $(m,\\mathcal{R}) = (1,0)$ when scale. $L_{\\mathrm{GWD}}(\\cdot)$ is Gaussian Wasserstein Distance Loss [60] and $L_{\\mathrm{ANG}}(\\cdot)$ is defined as:", + "bbox": [ + 511, + 406, + 905, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {A N G}} \\left(\\theta_ {1}, \\theta_ {2}\\right) = \\min _ {k \\in Z} \\left(s m o o t h _ {L 1} \\left(\\theta_ {1}, k \\pi + \\theta_ {2}\\right)\\right) \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 473, + 906, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\min (\\cdot)$ regresses the prediction toward the closest target to circumvent the periodicity problem [65].", + "bbox": [ + 511, + 502, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Copy-paste Augmentation", + "text_level": 1, + "bbox": [ + 511, + 544, + 750, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [9], we propose to crop the detected instances of step $k$ and paste them on the training image of step $k + 1$ . The maximum number of paste boxes is limited to 10 in each step. We simply use the bounding boxes of the cropped instances as the regression targets, and use Gaussian Wasserstein Distance Loss [60] to calculate $\\mathcal{L}_{\\mathrm{box}}$ .", + "bbox": [ + 511, + 566, + 906, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.7. Overall Loss", + "text_level": 1, + "bbox": [ + 511, + 669, + 648, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss $\\mathcal{L}$ for Point2RBox-v2 can be expressed as:", + "bbox": [ + 511, + 691, + 898, + 707 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c l s}} + w _ {\\text {b o x}} \\mathcal {L} _ {\\text {b o x}} + w _ {\\mathrm {O}} \\mathcal {L} _ {\\mathrm {O}} + w _ {\\mathrm {W}} \\mathcal {L} _ {\\mathrm {W}} + w _ {\\mathrm {E}} \\mathcal {L} _ {\\mathrm {E}} + w _ {\\mathrm {s s}} \\mathcal {L} _ {\\mathrm {s s}} \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 714, + 906, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{cls}}$ is the focal loss [25] for classification, $\\mathcal{L}_{\\mathrm{box}}$ regresses boxes/centers toward copy-paste/ground-truth labels, $w_{\\mathrm{box}}$ is set to one by default, $(w_{\\mathrm{O}}, w_{\\mathrm{W}}, w_{\\mathrm{E}}, w_{\\mathrm{ss}})$ are set to (10, 5, 0.3, 1) based on our ablation studies (see Tables 3-5).", + "bbox": [ + 511, + 737, + 906, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 814, + 643, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiments are carried out on NVIDIA RTX4090 GPUs using PyTorch 2.2.0 [36] and the rotation detection tool kits: MMRotate 1.0.0 [73]. All the experiments follow the same hyper-parameters (learning rate, batch size, optimizer, etc.).", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19287", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e4131da4848aee7ba00ff98671702dd7665daf3fd9c22843b21b7598dc65396e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods*PL1BDBRGTFSVLVSHTCBCSTSBFRAHASPHCAP50
▼RBox-supervised OOD
RepPoints (2019) [62]✓86.781.141.662.076.256.375.790.780.885.363.366.659.167.633.768.45
RetinaNet (2017) [25]✓88.277.045.069.471.559.074.590.884.979.357.364.762.766.539.668.69
GWD (2021) [56]✓89.375.447.861.979.573.886.190.984.579.455.959.763.271.045.471.66
FCOS (2019) [42]✓89.176.950.163.279.879.887.190.480.884.659.766.365.871.341.772.44
S2A-Net (2022) [12]✓89.283.052.574.678.879.287.590.984.984.861.968.070.771.459.875.81
▼HBox-supervised OOD
Sun et al. (2021) [41]×51.538.716.136.829.819.223.483.950.680.018.950.225.628.725.538.60
BoxInst-RBox (2021) [43]2×68.440.833.132.346.955.456.679.566.882.141.252.852.865.030.053.59
H2RBox (2023) [59]✓88.573.540.856.977.565.477.990.983.285.355.362.952.463.643.367.82
EIE-Det (2024) [45]✓87.770.241.560.580.776.386.390.982.684.753.164.558.170.443.870.10
H2RBox-v2 (2023) [65]✓89.074.450.060.579.875.386.990.985.185.059.263.265.270.549.772.31
▼Point-supervised OOD
Point2Mask-RBox (2023) [21]2×4.023.13.81.315.11.03.319.01.029.10.09.57.421.17.19.72
P2BNet+H2RBox (2023) [3, 59]×24.735.97.127.93.312.117.517.50.834.06.349.611.627.218.819.63
P2BNet+H2RBox-v2 (2023) [3, 65]×11.044.814.915.436.816.727.812.11.831.23.450.612.636.712.521.87
P2RBox (2024) [2]†×87.865.715.060.773.071.778.981.544.581.241.239.345.557.541.259.04
PointOBB (2024) [33]×26.165.79.159.465.834.929.80.52.316.70.649.021.841.036.730.08
Point2RBox (2024) [66]✓62.964.314.435.028.238.933.325.22.244.53.448.125.945.022.634.07
Point2RBox+SK (2024) [66]†✓53.363.93.750.940.039.245.776.710.556.15.449.524.251.233.840.27
Point2RBox+SK (2024) [66]†×66.459.55.252.654.153.957.390.83.257.86.147.422.955.740.544.90
PointOBB-v2 (2025) [38]×64.527.81.936.258.847.253.490.562.245.312.141.78.143.732.041.68
PointOBB-v3 (2025) [68]✓30.939.413.522.761.27.043.162.459.847.32.745.116.855.211.441.29
PointOBB-v3 (2025) [68]×52.954.421.352.765.644.967.887.226.773.432.653.339.056.410.249.24
Point2RBox-v2 (ours)✓78.452.78.340.971.060.574.788.765.572.124.426.130.150.721.051.00
Point2RBox-v2 (ours)×88.072.68.046.279.676.386.989.179.782.926.245.345.866.346.362.61
*Comparison tracks: √ = End-to-end training and testing; × = Generating pseudo labels to train the FCOS detector (two-stage training).†Using additional priors. P2RBox: Pre-trained SAM model; Point2RBox+SK: One-shot sketches for each class.†PL: Plane, BD: Baseball diamond, BR: Bridge, GTF: Ground track field, SV: Small vehicle, LV: Large vehicle, SH: Ship, TC: Tennis court, BC: Basketball court, ST: Storage tank, SBF: Soccer-ball field, RA: Roundabout, HA: Harbor, SP: Swimming pool, HC: Helicopter.‡RBox: The minimum rectangle operation is performed on the output Mask to obtain the RBox.
", + "bbox": [ + 101, + 88, + 898, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Detection performance of each category on the DOTA-v1.0 and the mean ${\\mathrm{{AP}}}_{50}$ of all categories.", + "bbox": [ + 186, + 511, + 805, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Average precision (AP) is adopted as the primary metric. All the models are configured upon ResNet50 [13] and trained with AdamW [31]. 1) Learning rate. Initialized at 5e-5, warm-up for 500 iterations, and divided by ten at each decay step. 2) Epochs. 72 for HRSC; 12 for the others. 3) Augmentation. Random rotation/flip for HRSC; random flip for the others. 4) Image size. Split into $1,024 \\times 1,024$ with an overlap of 200 for DOTA/FAIR1M/STAR; scaled to $800 \\times 800$ for others. 5) Multi-scale. All experiments evaluated without multi-scale technique [73]. 6) Datasets. Six remote sensing and one retail scene datasets, covering all datasets used by the main counterparts [2, 33, 66]:", + "bbox": [ + 89, + 540, + 485, + 720 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- DOTA [48]. DOTA-v1.0 has 2,806 aerial images annotated with 15 categories, while DOTA-v1.5/2.0 are the extended versions with more small objects and categories.", + "- DIOR [4]. It is an aerial image dataset re-annotated with RBoxes based on its original HBox version [18], with a high variation in object size and high intra-class diversity.", + "- HRSC [30]. It contains ship instances on the sea and inshore. The train/val/test set includes 436/181/444 images.", + "- FAIR1M [40]. It has more than 1 million instances and more than 40,000 images for fine-grained object recognition in remote sensing imagery, annotated with 37 cate" + ], + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gories. The results are evaluated on FAIR1M-1.0.", + "bbox": [ + 526, + 540, + 854, + 554 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- STAR [22]. It is extensive for scene graph generation, covering more than 210,000 objects with diverse spatial resolutions, classified into 48 fine-grained categories and precisely annotated with oriented bounding boxes.", + "- SKU110K [35]. It focuses on the detection of densely packed retail scenes with 110,712 objects in 11,762 images. The density reaches 86 instances per image.", + "- RSAR [70]. It is a remote sensing dataset based on Synthetic Aperture Radar (SAR) imagery with 6 categories." + ], + "bbox": [ + 513, + 555, + 906, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results on DOTA-v1.0", + "text_level": 1, + "bbox": [ + 511, + 708, + 766, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 compares Point2RBox-v2 with the state-of-the-art methods, which can be categorized into two tracks:", + "bbox": [ + 511, + 731, + 905, + 761 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) End-to-end training. These methods apply the trained weakly-supervised detector directly to the test set. Without relying on priors, our approach demonstrates an improvement of $16.93\\%$ (51.00% vs. 34.07%) compared to Point2RBox. Even when compared to Point2RBox+SK, which incorporates additional data-side priors (i.e. one-shot examples for each class), our method still outperforms it by $10.73\\%$ (51.00% vs. 40.27%).", + "2) Two-stage training. These methods generate RBox" + ], + "bbox": [ + 511, + 762, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19288", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8a48b60e8c06edd4fef8b98ffdae2956a8555086058b7fb8cbbe3703ac5e1932.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods*DOTA-v1.0DOTA-v1.5DOTA-v2.0DIORHRSCFAIR1MSTARSKU110KRSAR
▼RBox-supervised OOD
RetinaNet (2017) [25]✓68.6960.5747.0054.9684.4937.6721.8078.5057.67
GWD (2021) [56]✓71.6663.2748.8757.6086.6739.1125.3079.1657.80
FCOS (2019) [42]✓72.4464.5351.7759.8388.9941.2528.1080.0966.66
S2A-Net (2022) [12]✓75.8166.5352.3961.4190.1042.4427.3080.3666.47
▼HBox-supervised OOD
Sun et al. (2021) [41]×38.60--------
KCR (2023) [74]✓----79.10----
H2RBox (2023) [59]✓70.0561.7048.6857.807.0335.9417.2057.1549.92
H2RBox-v2 (2023) [65]✓72.3164.7650.3357.6489.6642.2727.3070.7065.16
AFWS (2024) [32]✓72.5565.9251.7359.07-41.80---
▼Point-supervised OOD
P2RBox (2024) [2]†×59.04--------
PointSAM (2025) [27]†×---46.20-----
PointOBB (2024) [33]×30.0810.665.5337.31-11.199.19-13.80
Point2RBox+SK (2024) [66]†✓40.2730.5123.4327.3479.4020.037.863.4127.81
PointOBB-v2 (2025) [38]×41.6830.5920.6439.56-13.369.0056.6318.99
PointOBB-v3 (2025) [68]✓41.2031.2522.8237.60-11.4211.31-15.84
PointOBB-v3 (2025) [68]×49.2433.7923.5240.18-18.3512.85-22.60
Point2RBox-v2 (ours)✓51.0039.4527.1134.7082.6725.727.8064.0028.60
Point2RBox-v2 (ours)×62.6154.0638.7944.4586.1534.7114.2065.6430.90
*Comparison tracks: √= End-to-end training and testing; × = Generating pseudo labels to train the FCOS detector (two-stage training).†Using additional priors. P2RBox/PointSAM: Pre-trained SAM model; Point2RBox+SK: One-shot sketches for each class.
", + "bbox": [ + 99, + 88, + 898, + 405 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d9ac847a5316a81c66a138a340189589b56d353a87d8af36f5593abbc67b3a3e.jpg", + "table_caption": [ + "Table 2. Accuracy $\\left( {\\mathrm{{AP}}}_{50}\\right)$ comparisons on the DOTA-v1.0/1.5/2.0, DIOR, HRSC, FAIR1M, STAR, SKU110K, and RSAR datasets." + ], + "table_footnote": [], + "table_body": "
\\(w_{O}\\)DOTAHRSC
E2EFCOSE2EFCOS
348.7661.6281.8584.36
549.8162.4482.4685.76
1051.0062.6182.6786.15
3045.8857.8381.5685.61
", + "bbox": [ + 101, + 434, + 352, + 521 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9888f00fcc3664d536be816ddeea6798415124dd597845c19eb0faaf9b4a7aec.jpg", + "table_caption": [ + "Table 3. Ablation with the weight of ${\\mathcal{L}}_{\\mathrm{O}}$ ." + ], + "table_footnote": [], + "table_body": "
\\( {w}_{\\mathrm{W}} \\)DOTAHRSC
E2EFCOSE2EFCOS
350.8556.7878.4283.49
551.0062.6182.6786.15
1049.1560.5430.3735.13
3042.8452.5323.8925.91
", + "bbox": [ + 370, + 434, + 625, + 521 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/721bc0741978f2e3538b152646b1bb8dc21165b18b0b41337812c5e5df25dc69.jpg", + "table_caption": [ + "Table 4. Ablation with the weight of ${\\mathcal{L}}_{\\mathrm{W}}$ ." + ], + "table_footnote": [], + "table_body": "
wEDOTAHRSC
E2EFCOSE2EFCOS
0.148.7557.6234.7139.45
0.351.0062.6182.6786.15
0.551.3662.6376.8585.22
1.049.0560.6356.5959.59
", + "bbox": [ + 645, + 434, + 895, + 521 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dcc7f3e3ed88f1e91f4254d34de293ed1491aa6d9fa4c7fec08cf3d046b70d2f.jpg", + "table_caption": [ + "Table 5. Ablation with the weight of ${\\mathcal{L}}_{\\mathrm{E}}$ ." + ], + "table_footnote": [], + "table_body": "
wssDOTAHRSC
E2EFCOSE2EFCOS
0.149.2859.6673.6678.92
1.051.0062.6182.6786.15
3.049.1559.201.301.65
", + "bbox": [ + 101, + 547, + 349, + 625 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/94857a85bd91954170322bd64883aa400374e0ce5f81ceb896c8a33c7519df8f.jpg", + "table_caption": [ + "Table 6. Ablation with the weight of ${\\mathcal{L}}_{\\mathrm{{ss}}}$ ." + ], + "table_footnote": [], + "table_body": "
R / F / SDOTAHRSCR / F / SDOTAHRSC
90% / 10% / 0%60.4285.4680% / 20% / 0%59.4684.73
75% / 0% / 25%60.7986.2260% / 15% / 25%62.3884.21
68% / 7% / 25%62.6186.1538% / 37% / 25%45.878.56
45% / 5% / 50%60.5585.3440% / 10% / 50%60.4910.74
", + "bbox": [ + 370, + 549, + 893, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Ablation with the proportion of augmented views in self-supervision.", + "bbox": [ + 400, + 628, + 861, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "labels on train/val sets, with which the FCOS detector is trained. In this two-stage mode, Point2RBox-v2 achieves an accuracy of $62.61\\%$ , considerably surpassing PointOBB series. Remarkably, it even outperforms the SAM-powered method P2RBox by $3.57\\%$ (62.61% vs. 59.04%).", + "bbox": [ + 89, + 656, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Class-wise analysis. The FCOS detector trained with labels generated by Point2RBox-v2 achieves accuracy nearly equivalent to RBox-supervised FCOS across six high-density categories: SH (86.9% vs. 87.1%), SV (79.6% vs. 79.8%), LV (76.3% vs. 79.8%), PL (88.0% vs. 89.1%), ST (82.9% vs. 84.6%), and TC (89.1% vs. 90.4%). Interestingly, these six high-density categories account for 88% of DOTA instances. By annotating these categories with points and generating RBoxes using Point2RBox-v2 while labeling the other sparse categories with RBoxes, we can significantly reduce annotation labor without sacrificing much accuracy,", + "bbox": [ + 88, + 734, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "highlighting the valuable role our method can play.", + "bbox": [ + 511, + 656, + 849, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results on More Datasets", + "text_level": 1, + "bbox": [ + 511, + 681, + 743, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are displayed in Table 2. On more challenging DOTA-v1.5/2.0, Point2RBox-v2 presents a similar trend, $23.47\\% / 18.15\\%$ higher than PointOBB-v2 in the pseudogeneration track. On the ship detection dataset HRSC, the gap between Point2RBox-v2 and RBox-supervised FCOS is only $2.84\\%$ (86.15% vs. 88.99%). DIOR is relatively sparse, leading to less improvement with our methods—lower than PointSAM (44.45% vs. 46.20%) but still higher than methods that do not use SAM. Our method also provides competitive performance on fine-grained datasets FAIR1M and STAR. In addition to remote sensing scenarios, we carry out experiments on SKU110K for densely packed retail scenes. Existing point-supervised methods struggle in this", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19289", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/916035995ff8381e486e21560479f131cec1896e656181e344006d143384d8eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 88, + 267, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6649cc04a16fbfe2ca5554605e2cd6478862a2ddaf500faad08f94dc6452c6d2.jpg", + "image_caption": [ + "Figure 5. Qualitative analysis on failed cases and overlap cases." + ], + "image_footnote": [], + "bbox": [ + 269, + 88, + 421, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97bfa86eff7ab22c510e4a551eda61f101afbe60d1c2069a12b0ed94dbb6b98f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 88, + 573, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93b828120c97dda984b2f8b9fe4d63c951428c0cc587a558e49af2fc60c50db1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 88, + 727, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3853fb30847fbd0e2573b0ef5cd80fb92b4d1ff11649132c646ac14645267b31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 89, + 880, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1216b04349f2f50f8d4dc434cf0e4193bc66f13fc22b1941ee48a0e78643a281.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModulesDOTAHRSC
\\( \\mathcal{L}_{\\mathrm{O}} \\)\\( \\mathcal{L}_{\\mathrm{W}} \\)\\( \\mathcal{L}_{\\mathrm{ss}} \\)\\( \\mathcal{L}_{\\mathrm{E}} \\)\\( CP \\)E2EFCOSE2EFCOS
✓0.000.000.000.00
✓✓41.5452.9817.9619.64
✓✓✓46.6454.2618.1022.13
✓✓✓✓49.5561.8878.7983.79
✓✓✓✓48.5859.5620.3524.76
✓✓✓✓38.9448.4411.6414.93
✓✓✓✓47.0855.0519.5821.78
✓✓✓✓✓51.0062.6182.6786.15
", + "bbox": [ + 99, + 239, + 475, + 377 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fb56a081ffc32266d71f888a9576ba5ceab2d6395293e835e7b89ede5215b5f5.jpg", + "table_caption": [ + "Table 8. Ablation with incremental addition of modules." + ], + "table_footnote": [], + "table_body": "
16K=2432|| 1.2β=1.62.0
50.8751.0048.08|| 48.1451.0051.33
", + "bbox": [ + 99, + 406, + 473, + 444 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1831d5d8524c4a8c175094ebc09bb6277c376a5bcd65d15beedea970f4063599.jpg", + "table_caption": [ + "Table 9. Ablation with $K$ and $\\beta$ in edge loss on DOTA (E2E)." + ], + "table_footnote": [], + "table_body": "
σPoint2RBoxPointOBB-v2Point2RBox-v2
DOTAHRSCDOTAHRSCDOTAHRSC
0%40.2779.4044.85-62.6186.15
10%39.6078.8142.30-61.5885.76
30%38.4278.2838.46-60.3185.71
", + "bbox": [ + 99, + 470, + 473, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Ablation with the inaccuracy in point annotations.", + "bbox": [ + 106, + 550, + 465, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "case, whereas Point2RBox-v2 achieves performance on par with HBox-supervised H2RBox (65.64% vs. 57.15%).", + "bbox": [ + 89, + 580, + 482, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 619, + 254, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tables 3-10 display the ablation studies on DOTA-v1.0 and HRSC. \"E2E\" denotes end-to-end training; \"FCOS\" denotes two-stage training (i.e. generating pseudo labels to train FCOS). The final values adopted are highlighted in gray.", + "bbox": [ + 89, + 642, + 482, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Weight of each loss. Tables 3-5 determine the weights of the proposed losses. Based on these experiments, the weights $(w_{\\mathrm{O}}, w_{\\mathrm{W}}, w_{\\mathrm{E}}, w_{\\mathrm{ss}})$ are set to (10, 5, 0.3, 1).", + "bbox": [ + 89, + 703, + 482, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Proportion of augmented views. Table 7 studies the proportion between rotation, flip, and scale. The results are reported with two-stage training (FCOS). Based on the results, the proportion is set to $68\\%$ , $7\\%$ , and $25\\%$ .", + "bbox": [ + 89, + 750, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Incremental addition of modules. Table 8 demonstrates the constraints from Gaussian and Voronoi achieve an accuracy of $52.98\\%$ on DOTA. Adding consistency loss and edge loss further boosts it to $54.26\\%$ and $61.88\\%$ , respectively, whereas the improvement from copy-paste is $0.73\\%$ . We also demonstrate the impact of omitting each core loss.", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Edge loss parameters. We set $K = 24$ and $\\beta = 1.6$ as they are observed to discern the correct edges during code development. Table 9 provides a more precise ablation.", + "bbox": [ + 511, + 242, + 903, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Annotation inaccuracy. We offset the annotated points by a noise from the uniform distribution $[- \\sigma H, + \\sigma H]$ , where $H$ is the height of objects. Table 10 shows that the $\\mathrm{AP}_{50}$ of Point2RBox-v2 decreases by less than $3\\%$ when noise is added to point annotations, demonstrating the robustness of the proposed learning mechanisms.", + "bbox": [ + 511, + 287, + 906, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. More Discussions", + "text_level": 1, + "bbox": [ + 511, + 388, + 684, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The qualitative analysis on the failed/overlap cases is shown in Fig. 5. 1) Failed cases. Although our method performs well overall, it struggles with certain categories that are sparse and not constrained by other objects. 2) Overlap cases. Minimizing overlap as a soft constraint during training does not entirely eliminate overlap. Once trained, the model remains robust to some overlap during inference.", + "bbox": [ + 511, + 411, + 906, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 532, + 633, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces Point2RBox-v2, a point-supervised oriented object detector that effectively leverages the arrangement and layout of instances. We propose the integration of Gaussian overlay and Voronoi tessellation to constrain the size and rotation of instances based on their spatial relationships. Additionally, by incorporating self-supervised consistency loss, edge loss, and copy-paste augmentation, the accuracy of the model is further enhanced.", + "bbox": [ + 511, + 556, + 906, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiments yield the following observations: 1) The integration of Gaussian and Voronoi concepts effectively harnesses the spatial layout of objects, significantly enhancing point-supervised OOD. 2) Point2RBox-v2 demonstrates exceptional performance in densely packed scenes (see Fig. 2), where existing methods struggle. 3) Our method does not require priors (i.e. pre-trained SAM or one-shot examples) and is applicable to both end-to-end and pseudo-generation modes. 4) It advances the state of the art by a large amount, achieving $62.61\\%$ , $86.15\\%$ , and $34.71\\%$ on the DOTA-v1.0, HRSC, and FAIR1M datasets, respectively.", + "bbox": [ + 511, + 678, + 908, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. The gap between Point2RBox-v2 and the RBox-supervised OOD is still huge in terms of sparse categories (i.e. BR/SBF) since little constraint can be obtained from the layout between them when the objects are sparse.", + "bbox": [ + 511, + 844, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19290", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 118, + 187, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Franz Aurenhammer. Voronoi diagrams—a survey of a fundamental geometric data structure. ACM Computing Surveys, 23(3):345-405, 1991. 2, 4", + "[2] Guangming Cao, Xuehui Yu, Wenwen Yu, Xumeng Han, Xue Yang, Guorong Li, Jianbin Jiao, and Zhenjun Han. P2rbox: Point prompt oriented object detection with SAM. arXiv preprint arXiv:2311.13128, 2024. 1, 2, 6, 7", + "[3] Pengfei Chen, Xuehui Yu, Xumeng Han, Najmul Hassan, Kai Wang, Jiachen Li, Jian Zhao, Humphrey Shi, Zhenjun Han, and Qixiang Ye. Point-to-box network for accurate object detection via single point supervision. In European Conference on Computer Vision, 2022. 3, 6", + "[4] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 2022. 6", + "[5] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2849-2858, 2019. 2", + "[6] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2021. 1", + "[7] Kun Fu, Zhonghan Chang, Yue Zhang, Guangluan Xu, Keshu Zhang, and Xian Sun. Rotation-aware and multi-scale convolutional neural network for object detection in remote sensing images. ISPRS Journal of Photogrammetry and Remote Sensing, 161:294-308, 2020. 1", + "[8] Shuyong Gao, Wei Zhang, Yan Wang, Qianyu Guo, Chenglong Zhang, Yangji He, and Wenqiang Zhang. Weakly-supervised salient object detection using point supervision. In AAAI Conference on Artificial Intelligence, pages 670-678, 2022. 3", + "[9] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D. Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 5", + "[10] Eran Goldman, Roei Herzig, Aviv Eisenschat, Jacob Goldberger, and Tal Hassner. Precise detection in densely packed scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5236, 2019. 1", + "[11] Jiaming Han, Jian Ding, Nan Xue, and Gui-Song Xia. Redet: A rotation-equivariant detector for aerial object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2785-2794, 2021. 2", + "[12] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2, 6, 7", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Con-" + ], + "bbox": [ + 93, + 143, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3, 6", + "[14] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In IEEE International Conference on Computer Vision, 2017. 5", + "[15] Liping Hou, Ke Lu, Xue Yang, Yuqiu Li, and Jian Xue. Grep: Gaussian representation for arbitrary-oriented object detection. Remote Sensing, 15(3):757, 2023. 2", + "[16] Javed Iqbal, Muhammad Akhtar Munir, Arif Mahmood, Afsheen Rafaqat Ali, and Mohsen Ali. Leveraging orientation for weakly supervised object detection with application to firearm localization. Neurocomputing, 440:310-320, 2021. 3", + "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3", + "[18] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing, 159:296-307, 2020. 6", + "[19] Qingyun Li, Yushi Chen, Xinya Shu, Dong Chen, Xin He, Yi Yu, and Xue Yang. A simple aerial detection baseline of multimodal language models. arXiv preprint arXiv:2501.09720, 2025. 2", + "[20] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented leptopoints for aerial object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1829-1838, 2022. 2", + "[21] Wentong Li, Yuqian Yuan, Song Wang, Jianke Zhu, Jianshu Li, Jian Liu, and Lei Zhang. Point2mask: Point-supervised panoptic segmentation via optimal transport. In IEEE International Conference on Computer Vision, 2023. 2, 3, 6", + "[22] Yansheng Li, Linlin Wang, Tingzhu Wang, Xue Yang, Junwei Luo, Qi Wang, Youming Deng, Wenbin Wang, Xian Sun, Haifeng Li, Bo Dang, Yongjun Zhang, Yi Yu, and Junchi Yan. Star: A first-ever dataset and a large-scale benchmark for scene graph generation in large-size satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1832-1849, 2025. 6", + "[23] Minghui Liao, Zhen Zhu, Baoguang Shi, Gui-Song Xia, and Xiang Bai. Rotation-sensitive regression for oriented scene text detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5909-5918, 2018. 1", + "[24] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 3", + "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 2, 5, 6, 7", + "[26] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International Journal of Computer Vision, 128(2):261-318, 2020. 2" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19291", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Nanqing Liu, Xun Xu, Yongyi Su, Haojie Zhang, and Heng-Chao Li. Pointsam: Pointly-supervised segment anything model for remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 63:1-15, 2025. 1, 2, 7", + "[28] Xuebo Liu, Ding Liang, Shi Yan, Dagui Chen, Yu Qiao, and Junjie Yan. Fots: Fast oriented text spotting with a unified network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5676-5685, 2018. 1", + "[29] Yuekai Liu, Hongli Gao, Liang Guo, Aoping Qin, Canyu Cai, and Zhichao You. A data-flow oriented deep ensemble learning method for real-time surface defect inspection. IEEE Transactions on Instrumentation and Measurement, 69(7): 4681-4691, 2020. 1", + "[30] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International Conference on Pattern Recognition Applications and Methods, pages 324-331, 2017. 1, 6", + "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6", + "[32] Junyan Lu, Qinglei Hu, Ruifei Zhu, Yali Wei, and Tie Li. Afws: Angle-free weakly-supervised rotating object detection for remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 7", + "[33] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. Pointobb: Learning oriented object detection via single point supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 6, 7", + "[34] Jeffri Murragarra-Llerena, Lucas N Kirsten, Luis Felipe Zeni, and Claudio R Jung. Probabilistic intersection-over-union for training and evaluation of oriented object detectors. IEEE Transactions on Image Processing, 2024. 2", + "[35] Xingjia Pan, Yuqiang Ren, Kekai Sheng, Weiming Dong, Haolei Yuan, Xiaowei Guo, Chongyang Ma, and Changsheng Xu. Dynamic refinement network for oriented and densely packed object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11207-11216, 2020. 1, 6", + "[36] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035, 2019. 5", + "[37] Wen Qian, Xue Yang, Silong Peng, Junchi Yan, and Yue Guo. Learning modulated loss for rotated object detection. In AAAI Conference on Artificial Intelligence, pages 2458-2466, 2021. 2", + "[38] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. Pointobb-v2: Towards simpler, faster, and stronger single point supervised oriented object detection. In International Conference on Learning Representations, 2025. 1, 2, 3, 6, 7", + "[39] Xavier Soria, Yachuan Li, Mohammad Rouhani, and Angel D. Sappa. Tiny and efficient model for the edge detection" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generalization. In IEEE/CVF International Conference on Computer Vision Workshops, pages 1364-1373, 2023. 5", + "[40] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, Martin Weinmann, Stefan Hinz, Cheng Wang, and Kun Fu. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 6", + "[41] Yongqing Sun, Jie Ran, Feng Yang, Chenqiang Gao, Takayuki Kurozumi, Hideaki Kimata, and Ziqi Ye. Oriented object detection for remote sensing images based on weakly supervised learning. In IEEE International Conference on Multimedia & Expo Workshops, pages 1-6, 2021. 3, 6, 7", + "[42] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In IEEE/CVF International Conference on Computer Vision, pages 9626-9635, 2019. 2, 6, 7", + "[43] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 6", + "[44] L. Vincent and P. Soille. Watersheds in digital spaces: an efficient algorithm based on immersion simulations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(6):583-598, 1991. 4", + "[45] Linfei Wang, Yibing Zhan, Xu Lin, Baosheng Yu, Liang Ding, Jianqing Zhu, and Dapeng Tao. Explicit and implicit box equivariance learning for weakly-supervised rotated object detection. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 3, 6", + "[46] Long Wen, Yu Cheng, Yi Fang, and Xinyu Li. A comprehensive survey of oriented object detection in remote sensing images. Expert Systems with Applications, page 119960, 2023. 2", + "[47] Hongjin Wu, Ruoshan Lei, and Yibing Peng. Pcbnet: A lightweight convolutional neural network for defect inspection in surface mount technology. IEEE Transactions on Instrumentation and Measurement, 71:1-14, 2022. 1", + "[48] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2018. 1, 6", + "[49] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In IEEE/CVF International Conference on Computer Vision, pages 3520-3529, 2021. 2", + "[50] Xue Yang and Junchi Yan. Arbitrary-oriented object detection with circular smooth label. In European Conference on Computer Vision, pages 677-694, 2020. 2", + "[51] Xue Yang and Junchi Yan. On the arbitrary-oriented object detection: Classification based approaches revisited. International Journal of Computer Vision, 130(5):1340-1365, 2022. 1, 2", + "[52] Xue Yang, Hao Sun, Kun Fu, Jirui Yang, Xian Sun, Menglong Yan, and Zhi Guo. Automatic ship detection in remote" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19292", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sensing images from google earth of complex scenes based on multiscale rotation dense feature pyramid networks. Remote sensing, 10(1):132, 2018. 1, 2", + "[53] Xue Yang, Jirui Yang, Junchi Yan, Yue Zhang, Tengfei Zhang, Zhi Guo, Xian Sun, and Kun Fu. Scdet: Towards more robust detection for small, cluttered and rotated objects. In IEEE/CVF International Conference on Computer Vision, pages 8231-8240, 2019. 2", + "[54] Xue Yang, Liping Hou, Yue Zhou, Wentao Wang, and Junchi Yan. Dense label encoding for boundary discontinuity free rotation detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15814-15824, 2021. 2", + "[55] Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3det: Refined single-stage detector with feature refinement for rotating object. In AAAI Conference on Artificial Intelligence, pages 3163-3171, 2021. 2", + "[56] Xue Yang, Junchi Yan, Ming Qi, Wentao Wang, Xiaopeng Zhang, and Tian Qi. Rethinking rotated object detection with gaussian Wasserstein distance loss. In 38th International Conference on Machine Learning, pages 11830-11841, 2021. 2, 6, 7", + "[57] Xue Yang, Xiaojiang Yang, Jirui Yang, Qi Ming, Wentao Wang, Qi Tian, and Junchi Yan. Learning high-precision bounding box for rotated object detection via kullback-leibler divergence. In Advances in Neural Information Processing Systems, pages 18381-18394, 2021. 2", + "[58] Xue Yang, Junchi Yan, Wenlong Liao, Xiaokang Yang, Jin Tang, and Tao He. Scdet++: Detecting small, cluttered and rotated objects via instance-level feature denoising and rotation loss smoothing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2384-2399, 2022. 2", + "[59] Xue Yang, Gefan Zhang, Wentong Li, Xuehui Wang, Yue Zhou, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. International Conference on Learning Representations, 2023. 1, 3, 6, 7", + "[60] Xue Yang, Gefan Zhang, Xiaojiang Yang, Yue Zhou, Wentao Wang, Jin Tang, Tao He, and Junchi Yan. Detecting rotated objects as gaussian distributions and its 3-d generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4335-4354, 2023. 2, 3, 4, 5", + "[61] Xue Yang, Yue Zhou, Gefan Zhang, Jirui Yang, Wentao Wang, Junchi Yan, Xiaopeng Zhang, and Qi Tian. The kfiou loss for rotated object detection. In International Conference on Learning Representations, 2023. 2", + "[62] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Reppoints: Point set representation for object detection. In IEEE/CVF International Conference on Computer Vision, pages 9656-9665, 2019. 2, 6", + "[63] Yi Yu and Feipeng Da. Phase-shifting coder: Predicting accurate orientation in oriented object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[64] Yi Yu and Feipeng Da. On boundary discontinuity in angle regression based arbitrary oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-14, 2024. 3" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Yi Yu, Xue Yang, Qingyun Li, Yue Zhou, Feipeng Da, and Junchi Yan. H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. In Advances in Neural Information Processing Systems, 2023. 1, 3, 5, 6, 7", + "[66] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2rbox: Combine knowledge from synthetic visual patterns for end-to-end oriented object detection with single point supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6, 7", + "[67] Yi Yu, Xue Yang, Yansheng Li, Zhenjun Han, Feipeng Da, and Junchi Yan. Wholly-wood: Wholly leveraging diversified-quality labels for weakly-supervised oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025. 3", + "[68] Peiyuan Zhang, Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Yue Zhou, Xiaosong Jia, Xudong Lu, Jingdong Chen, Xiang Li, et al. Pointobb-v3: Expanding performance boundaries of single point-supervised oriented object detection. arXiv preprint arXiv:2501.13898, 2025. 1, 2, 3, 6, 7", + "[69] Shun Zhang, Jihui Long, Yaohui Xu, and Shaohui Mei. Pmho: Point-supervised oriented object detection based on segmentation-driven proposal generation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2", + "[70] Xin Zhang, Xue Yang, Yuxuan Li, Jian Yang, Ming-Ming Cheng, and Xiang Li. Rsar: Restricted state angle resolver and rotated sar benchmark. arXiv preprint arXiv:2501.04440, 2025.6", + "[71] Zhong-Qiu Zhao, Peng Zheng, Shou-Tao Xu, and Xindong Wu. Object detection with deep learning: A review. IEEE Transactions on Neural Networks and Learning Systems, 30 (11):3212-3232, 2019. 2", + "[72] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: An efficient and accurate scene text detector. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2017. 1", + "[73] Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, et al. MMrotate: A rotated object detection benchmark using pytorch. In 30th ACM International Conference on Multimedia, pages 7331-7334, 2022. 5, 6", + "[74] Tianyu Zhu, Bryce Ferenczi, Pulak Purkait, Tom Drummond, Hamid Rezatofighi, and Anton van den Hengel. Knowledge combination to learn rotated detection without rotated annotation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 7" + ], + "bbox": [ + 516, + 92, + 906, + 739 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19293", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/d3a3a010-6834-441e-beb3-f3f784483984_model.json b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/d3a3a010-6834-441e-beb3-f3f784483984_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1ca76e4746a1351228984e5aba6e156794096424 --- /dev/null +++ b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/d3a3a010-6834-441e-beb3-f3f784483984_model.json @@ -0,0 +1,3137 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.126, + 0.865, + 0.172 + ], + "angle": 0, + "content": "Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.191, + 0.862, + 0.228 + ], + "angle": 0, + "content": "Yi Yu\\(^{1*}\\), Botao Ren\\(^{2*}\\), Peiyuan Zhang\\(^{3*}\\), Mingxin Liu\\(^{4}\\), Junwei Luo\\(^{3}\\), Shaofeng Zhang\\(^{4}\\), Feipeng Da\\(^{1}\\), Junchi Yan\\(^{4}\\), Xue Yang\\(^{4\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.228, + 0.795, + 0.264 + ], + "angle": 0, + "content": "\\(^{1}\\)Southeast University \\(^{2}\\)Tsinghua University \\(^{3}\\)Wuhan University \\(^{4}\\)Department of Automation & CSE & SAI, Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.266, + 0.893, + 0.281 + ], + "angle": 0, + "content": "{yuyi,dafp}@seu.edu.cn rbt22@mails.tsinghua.edu.cn {peiyuanzhangwhu,luojunwei}@whu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.281, + 0.792, + 0.295 + ], + "angle": 0, + "content": "{liumingxin,zhangshaofeng,yanjunchi, yangxue-2019-sjtu}@sjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.295, + 0.691, + 0.308 + ], + "angle": 0, + "content": "https://github.com/VisionXLab/point2rbox-v2" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.335, + 0.327, + 0.352 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.368, + 0.486, + 0.671 + ], + "angle": 0, + "content": "With the rapidly increasing demand for oriented object detection (OOD), recent research involving weakly-supervised detectors for learning OOD from point annotations has gained great attention. In this paper, we rethink this challenging task setting with the layout among instances and present Point2RBox-v2. At the core are three principles: 1) Gaussian overlap loss. It learns an upper bound for each instance by treating objects as 2D Gaussian distributions and minimizing their overlap. 2) Voronoi watershed loss. It learns a lower bound for each instance through watershed on Voronoi tessellation. 3) Consistency loss. It learns the size/rotation variation between two output sets with respect to an input image and its augmented view. Supplemented by a few devised techniques, e.g. edge loss and copy-paste, the detector is further enhanced. To our best knowledge, Point2RBox-v2 is the first approach to explore the spatial layout among instances for learning point-supervised OOD. Our solution is elegant and lightweight, yet it is expected to give a competitive performance especially in densely packed scenes: \\(62.61\\% /86.15\\% /34.71\\%\\) on DOTA/HRSC/FAIR1M." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.689, + 0.222, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.485, + 0.79 + ], + "angle": 0, + "content": "Emerged as an essential task in computer vision, oriented object detection (OOD) has become a prominent demand in autonomous driving [6], aerial images [7, 30, 48, 51, 52], scene text [23, 28, 72], retail scenes [10, 35], industrial inspection [29, 47], and more." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.485, + 0.821 + ], + "angle": 0, + "content": "Manual annotations are essential to teach the detector new concepts of visual objects. Early research is often super" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.335, + 0.906, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.502, + 0.907, + 0.531 + ], + "angle": 0, + "content": "Figure 1. Related methods, their principles for knowledge mining, whether using additional priors, and performance on DOTA-v1.0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.909, + 0.734 + ], + "angle": 0, + "content": "vised by rotated bounding boxes (RBoxes), where annotations align with the desired output. While this approach has demonstrated promising results, RBox labels are considerably expensive. The cost of annotating each RBox is approximately \\(36.5\\%\\) higher than a horizontal bounding box (HBox) and \\(104.8\\%\\) higher than a point [66]. This highlights the importance of utilizing coarser labels for weakly-supervised oriented detection. Recent advancements in HBox-supervised OOD, particularly with H2RBox [59] and H2RBox-v2 [65], have shown promise in bridging the gap between HBox-and RBox-supervised methods, thereby lessening the dependence on labor-intensive RBox labeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Following HBox-supervised methods, the more challenging point-supervised OOD opens up a new avenue. Several innovative approaches have emerged in the past year (see Fig. 1): 1) Point-prompt OOD. P2RBox [2], PMHO [69], and PointSAM [27] employ the zero-shot ability of SAM [17]. 2) Pseudo generation. PointOBB [33] and its v2/v3 version [38, 68] use multiple instance learning and class probability map for RBox generation. 3) Knowledge combination. Point2RBox [66] learns from one-shot examples. However, existing methods overlook the spatial relationships between objects. This brings us to the motivation of this paper." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.828, + 0.486, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding author. The work was in part supported by National Natural Science Foundation of China (62306069, 62222607, 72342023), Major Science and Technology Project of Jiangsu Province (BG2024003), and China Postdoctoral Science Foundation (2023M740602). Yi Yu is also supported by Jiangsu Funding Program for Excellent Postdoctoral Talent (2023ZB616)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19283" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.09, + 0.268, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.091, + 0.422, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.091, + 0.574, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.091, + 0.726, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.091, + 0.88, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.208, + 0.268, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.208, + 0.421, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.208, + 0.574, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.208, + 0.726, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.208, + 0.88, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.326, + 0.268, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.326, + 0.421, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.326, + 0.574, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.326, + 0.726, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.326, + 0.88, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.445, + 0.908, + 0.475 + ], + "angle": 0, + "content": "Figure 2. Visual comparisons with state-of-the-art methods including: Point2Mask (2023) [21], PointOBB (2024) [33], PointOBB-v2 (2025) [38], PointOBB-v3 (2025) [68], and Point2RBox (2024) [66]. The boxes detected by our method (last row) wrap the objects more tightly." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.493, + 0.483, + 0.644 + ], + "angle": 0, + "content": "Motivation. While point-supervised OOD has gained attention, the utilization of relationships among instances remains absent in the literature. For example, the arrangement of vehicles in a parking lot can serve as effective constraints for learning their size and orientation. This approach may yield valuable insights into point-supervised OOD, particularly in densely packed scenes where current methods face significant challenges. Can we leverage the spatial layout of objects to enhance point-supervised OOD? In this paper, we dive into this idea and seek to answer this question." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Highlights. 1) Point2RBox-v2 is proposed for point-supervised OOD, advancing the state of the art as displayed in Fig. 2 and Tables 1-2. 2) We propose novel and elegant losses to enforce constraints from the spatial layout among instances based on Gaussian overlap and Voronoi tessellation [1]. 3) Other modules are devised (i.e. edge loss, consistency loss, copy-paste) to further enhance the method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Contributions. 1) To our best knowledge, this work is the first attempt to learn point-supervised OOD from the layout among objects, where we propose novel principles based on Gaussian overlap and Voronoi tessellation. 2) The training pipeline and detailed implementation are elucidated, with necessary modules (i.e. edge loss, consistency loss, copy-paste) incorporated. The source code will be made publicly available. 3) Extensive experiments demonstrate that leveraging the spatial layout of instances can significantly advance the state of the art, surpassing other alternatives in accuracy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.491, + 0.655, + 0.507 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.831, + 0.532 + ], + "angle": 0, + "content": "2.1. RBox-supervised Oriented Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.908, + 0.795 + ], + "angle": 0, + "content": "In addition to horizontal detection [26, 71], oriented object detection (OOD) [46, 52] has received extensive attention. Representative works include anchor-based detector Rotated RetinaNet [25], anchor-free detector Rotated FCOS [42], and two-stage solutions, e.g. RoI Transformer [5], Oriented R-CNN [49], and ReDet [11]. Some research enhances the detector by exploiting alignment features, e.g. \\(\\mathrm{R}^3\\mathrm{Det}\\) [55] and \\(\\mathrm{S}^2\\mathrm{A}\\)-Net [12]. The angle regression may face boundary discontinuity and remedies are developed, including modulated losses [37, 53, 58] that alleviate loss jumps, angle coders [50, 51, 54, 63] that convert the angle into boundary-free coded data, and Gaussian-based losses [34, 56, 57, 60, 61] transforming rotated bounding boxes into Gaussian distributions. RepPoint-based methods [15, 20, 62] provide alternatives that predict a set of points that bounds the spatial extent of an object. LMMRotate [19] is a new paradigm of OOD based on multimodal language model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.829, + 0.82 + ], + "angle": 0, + "content": "2.2. Point-supervised Oriented Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Recently, several methods for point-supervised OOD have been proposed: 1) P2RBox [2], PMHO [69], and PointSAM [27] propose point-prompt OOD by employing the zero-shot Point-to-Mask ability of SAM [17]. 2) Point2RBox [66] introduces an end-to-end approach based on knowledge com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19284" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.895, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.305, + 0.908, + 0.335 + ], + "angle": 0, + "content": "Figure 3. The training pipeline of Point2RBox-v2. Gaussian overlap loss and Voronoi watershed loss utilize the spatial layout (see Fig. 4), while edge loss (see Sec. 3.4), symmetry-aware learning (see Sec. 3.5), and copy-paste (see Sec. 3.6) further enhance the method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.352, + 0.482, + 0.427 + ], + "angle": 0, + "content": "bination. 3) PointOBB [33, 68] achieves RBox generation through multiple instance learning. 4) PointOBB-v2 [38] learns a class probability map to generate pseudo RBoxes. 5) Wholly-WOOD [67] uses points, HBoxes, RBoxes, or their combinations for mixed supervised OOD." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.429, + 0.483, + 0.504 + ], + "angle": 0, + "content": "Among these methods, P2RBox, PMHO, and PointSAM rely on the SAM model pre-trained on massive labeled datasets, whereas Point2RBox requires one-shot examples for each category. In contrast, PointOBB series do not use many priors, but they necessitate two-stage training." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.515, + 0.388, + 0.531 + ], + "angle": 0, + "content": "2.3. Other Weakly-supervised Settings" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.483, + 0.627 + ], + "angle": 0, + "content": "Compared to Point-to-RBox, some other settings have been better studied. These methods are potentially applicable to our Point-to-RBox task setting by using a cascade pipeline, such as Point-to-HBox-to-RBox. In our experiment, cascade pipelines powered by state-of-the-art approaches are also compared. Here, representative works are introduced." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.483, + 0.81 + ], + "angle": 0, + "content": "HBox-to-RBox. H2RBox [59] establishes a paradigm that limits the object to a few candidate angles through geometric constraint from HBoxes, with a self-supervised branch eliminating the undesired results. An enhanced version H2RBox-v2 [65] is proposed to leverage the reflection symmetry of objects to further boost the accuracy. EIE-Det [45] uses an explicit equivariance branch for learning rotation consistency, and an implicit equivariance branch for learning position, aspect ratio, and scale consistency. KCR [74] combines RBox- and HBox-annotated datasets for transfer learning. Some studies [16, 41] use additional annotated data for training, which are also attractive but less general." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Point-to-HBox. Several related approaches have been developed, including: 1) P2BNet [3] samples box proposals of different sizes around the labeled point and classify them to achieve point-supervised horizontal object detection. 2) PSOD [8] achieves point-supervised salient object detection using an edge detector and adaptive masked flood fill." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.459 + ], + "angle": 0, + "content": "Point-to-Mask. Point2Mask [21] is proposed to achieve panoptic segmentation using single point annotation per target. SAM (Segment Anything Model) [17] produces object masks from input point/HBox prompts. Though RBoxes can be obtained from the segmentation mask by finding the minimum circumscribed rectangle, such a complex pipeline can be less cost-efficient and perform worse [59, 65]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.471, + 0.605, + 0.487 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.755, + 0.513 + ], + "angle": 0, + "content": "3.1. Overview and Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.906, + 0.624 + ], + "angle": 0, + "content": "An overview of Point2RBox-v2 is illustrated in Fig. 3. The network is based on ResNet50 [13] backbone, FPN [24] head, and PSC [64] angle coder. Objects of varying sizes are typically assigned to different FPN layers based on the scale. However, points lack size, so we assign them all to the P3 layer (stride = 8). Assume the detector \\( f_{\\mathrm{nn}}(\\cdot) \\) maps an image \\( I \\) to a set of RBoxes as detection results:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.631, + 0.907, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\left(x _ {c}, y _ {c}, w, h, \\theta\\right) = f _ {\\mathrm {n n}} (I) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.655, + 0.906, + 0.686 + ], + "angle": 0, + "content": "Equivalently, oriented objects can also be represented by 2D Gaussian distributions \\(\\mathcal{N}(\\mu, \\Sigma)\\) [60]:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.691, + 0.907, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mu = \\left[ \\begin{array}{l l} x _ {c} & y _ {c} \\end{array} \\right], \\quad \\Sigma = \\mathbf {R} \\left[ \\begin{array}{c c} w / 2 & 0 \\\\ 0 & h / 2 \\end{array} \\right] ^ {2} \\mathbf {R} ^ {\\top} \\right\\} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.732, + 0.559, + 0.745 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.742, + 0.907, + 0.776 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} = \\left[ \\begin{array}{c c} \\cos \\theta & - \\sin \\theta \\\\ \\sin \\theta & \\cos \\theta \\end{array} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "At the core of Point2RBox-v2 is to utilize the constraints from the layout, which is achieved by Gaussian overlap loss (Sec. 3.2) and Voronoi watershed loss (Sec. 3.3). These two losses effectively limit the size and rotation of objects to a reasonable range. Upon that, edge loss (Sec. 3.4) aligns the bounding box with the edge of objects to improve the accuracy. Incorporated with symmetry-aware learning (Sec. 3.5) and copy-paste augmentation (Sec. 3.6), we achieve a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19285" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.885, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.476, + 0.907, + 0.505 + ], + "angle": 0, + "content": "Figure 4. To illustrate the procedure of the three newly proposed loss functions and their impact on the learning results. (a) Gaussian overlap loss (see Sec. 3.2). (b) Voronoi watershed loss (see Sec. 3.3). (c) Edge loss (see Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.486, + 0.553 + ], + "angle": 0, + "content": "stable and high-accuracy solution for point-supervised OOD. In subsequent subsections, these modules are detailed." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.567, + 0.307, + 0.584 + ], + "angle": 0, + "content": "3.2. Gaussian Overlap Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.651 + ], + "angle": 0, + "content": "As mentioned, oriented objects can be represented by 2D Gaussian distributions \\(\\mathcal{N}(\\mu, \\Sigma)\\). The overlap volume between two distributions \\(\\mathcal{N}_1(\\mu_1, \\Sigma_1)\\) and \\(\\mathcal{N}_2(\\mu_2, \\Sigma_2)\\) can be approximated by the Bhattacharyya coefficient [60] as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.657, + 0.484, + 0.692 + ], + "angle": 0, + "content": "\\[\nB \\left(\\mathcal {N} _ {1}, \\mathcal {N} _ {2}\\right) = \\exp \\left(- \\frac {1}{8} \\mu^ {\\top} \\Sigma^ {- 1} \\mu\\right) \\cdot \\frac {\\left| \\Sigma_ {1} \\right| ^ {1 / 4} \\left| \\Sigma_ {2} \\right| ^ {1 / 4}}{\\left| \\Sigma \\right| ^ {1 / 2}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.483, + 0.728 + ], + "angle": 0, + "content": "where \\(\\mu = \\mu_{2} - \\mu_{1}\\), \\(\\Sigma = \\frac{1}{2} (\\Sigma_{1} + \\Sigma_{2})\\), \\(|\\Sigma|\\) denotes the determinant of the covariance matrix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Based on the above equation, we build a Gaussian overlap matrix \\(\\mathbf{M} \\in \\mathbb{R}^{N \\times N}\\) for each scene image as:" + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.767, + 0.483, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i, j} = B \\left(\\mathcal {N} _ {i}, \\mathcal {N} _ {j}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.483, + 0.821 + ], + "angle": 0, + "content": "where \\(i,j = 1,2,\\ldots ,N;N\\) is the distribution count (or the instance count) within one training image." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.822, + 0.456, + 0.838 + ], + "angle": 0, + "content": "The Gaussian overlap loss can then be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.843, + 0.484, + 0.881 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {O} = \\frac {1}{N} \\sum_ {i \\neq j} \\left(\\mathbf {M} _ {i, j}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(i \\neq j\\) omits diagonal elements. With this loss, the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.554 + ], + "angle": 0, + "content": "detector learns to arrange instances (see Fig. 4a) based on the mutual exclusivity among instances." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.74, + 0.577 + ], + "angle": 0, + "content": "3.3. Voronoi Watershed Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.689 + ], + "angle": 0, + "content": "A Voronoi diagram [1] is a partitioning of a space based on a set of points. Point annotations can be effectively utilized to calculate a Voronoi diagram, where a distinct polygon region is assigned to each point-annotated instance. The watershed algorithm [44], on the other hand, is a region segmentation technique that treats the intensity of pixels as a topographic surface, identifying regions as \"catchment basins\"." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.69, + 0.9, + 0.705 + ], + "angle": 0, + "content": "The calculation of Voronoi ridges can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.646, + 0.712, + 0.907, + 0.728 + ], + "angle": 0, + "content": "\\[\nV = \\text {V o r o n o i} (X) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.765 + ], + "angle": 0, + "content": "where \\(X\\) are the annotated points within a training image; \\(V\\) are the output Voronoi ridges (pixel coordinates)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.907, + 0.841 + ], + "angle": 0, + "content": "Interestingly, we find that Voronoi diagrams can be utilized as initial markers for watershed to obtain a region for each instance (see Fig. 4b). In concrete terms, the points \\( X \\) can be employed as foreground markers, while the Voronoi ridges \\( V \\) can act as background boundaries:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.848, + 0.907, + 0.865 + ], + "angle": 0, + "content": "\\[\nS = \\text {W a t e r s h e d} (I, X, V) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(S\\) are the output basin regions (pixel coordinates) corresponding to each annotated instance. By rotating \\(S\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19286" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "to align with the direction of the current prediction, the regression target of width and height can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.127, + 0.484, + 0.162 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l} w _ {t} \\\\ h _ {t} \\end{array} \\right] = 2 \\max \\left| \\mathbf {R} ^ {\\top} \\left(S - \\left[ \\begin{array}{l} x _ {c} \\\\ y _ {c} \\end{array} \\right]\\right) \\right| \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.167, + 0.483, + 0.197 + ], + "angle": 0, + "content": "where \\((x_{c},y_{c},w,h,\\theta)\\) is the current prediction; \\(\\mathbf{R}\\) is defined by Eq. (3); \\(w_{t}\\) and \\(h_t\\) are detached to stop the gradient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.197, + 0.484, + 0.227 + ], + "angle": 0, + "content": "Afterward, the Voronoi watershed loss to regress the width and height of objects can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.233, + 0.484, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {W} = L _ {\\mathrm {G W D}} \\left(\\left[ \\begin{array}{c c} w / 2 & 0 \\\\ 0 & h / 2 \\end{array} \\right] ^ {2}, \\left[ \\begin{array}{c c} w _ {t} / 2 & 0 \\\\ 0 & h _ {t} / 2 \\end{array} \\right] ^ {2}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.484, + 0.295 + ], + "angle": 0, + "content": "where \\(L_{\\mathrm{GWD}}(\\cdot)\\) is Gaussian Wasserstein Distance Loss [60]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.304, + 0.208, + 0.319 + ], + "angle": 0, + "content": "3.4. Edge Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.484, + 0.371 + ], + "angle": 0, + "content": "The above two losses have limited the size to a reasonable range. To make it more accurate, we propose the edge loss to snap the boundaries toward the edges (see Fig. 4c)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.402 + ], + "angle": 0, + "content": "First, a region \\(P \\in \\mathbb{R}^{(2K + 1) \\times (2K + 1)}\\) around each predicted RBox is extracted via Rotated RoI Align [14]:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.408, + 0.484, + 0.426 + ], + "angle": 0, + "content": "\\[\nP = R o I A l i g n (E (I), \\left(x _ {c}, y _ {c}, \\beta w, \\beta h, \\theta\\right)) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.462 + ], + "angle": 0, + "content": "where \\(E(\\cdot)\\) is the edge detection function [39]. We set \\(K = 24\\) and \\(\\beta = 1.6\\) in our experiments (see Table 9)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.492 + ], + "angle": 0, + "content": "By calculating the sum of each row of \\( P \\), the edge distribution in \\( y \\) direction can be obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.498, + 0.484, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\mu_ {i} = \\sum_ {j = 1} ^ {2 K + 1} \\left(P _ {(K + 1 - i), j} + P _ {(K + 1 + i), j}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.575 + ], + "angle": 0, + "content": "where \\( i = 1,2,\\dots ,K \\), indicating that the upper half of \\( P \\) is reversed and added to the lower half." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Meanwhile, the current prediction of the edge can also be softened into a distribution as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.609, + 0.484, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {i} = \\exp \\left(- \\frac {(i - K / \\beta) ^ {2}}{2 \\sigma_ {E} {} ^ {2}}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.483, + 0.679 + ], + "angle": 0, + "content": "where \\(i = 1,2,\\ldots ,K;\\sigma_{E}\\) is set to 6. Note that we crop \\(P\\) based on the predicted RBox, thus \\(\\lambda\\) is always the same." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.68, + 0.443, + 0.695 + ], + "angle": 0, + "content": "Multiplying the two distributions yields the target:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.699, + 0.484, + 0.731 + ], + "angle": 0, + "content": "\\[\nh _ {t} = \\frac {\\beta h}{K} \\arg \\max (\\mu \\times \\lambda) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.764 + ], + "angle": 0, + "content": "where \\( h_t \\) is the regression target of height. Likewise, the width target \\( w_t \\) is calculated along \\( x \\) direction." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.765, + 0.336, + 0.779 + ], + "angle": 0, + "content": "The edge loss is then obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.785, + 0.484, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {E} = \\operatorname {s m o o t h} _ {L 1} \\left(\\left[ \\begin{array}{l l} w & h \\end{array} \\right], \\left[ \\begin{array}{l l} w _ {t} & h _ {t} \\end{array} \\right]\\right) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Note that \\(\\mathcal{L}_W\\) and \\(\\mathcal{L}_E\\) merely refines the width and height of boxes, without involving the angle regression." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.336, + 0.866 + ], + "angle": 0, + "content": "3.5. Symmetry-aware Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Proven effective in learning the rotation of objects [65], in this work, symmetry-aware learning is extended to Gaussian-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "based OOD. In the left part of Fig. 3, the training image \\(I\\) is transformed (randomly selected from rotation, flip, and scale, see Table 7) to generate an augmented view as:" + }, + { + "type": "equation", + "bbox": [ + 0.675, + 0.145, + 0.907, + 0.161 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {a u g}} = \\alpha I \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.56, + 0.18 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.185, + 0.907, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\alpha = \\left[ \\begin{array}{c c} \\cos \\mathcal {R} & - \\sin \\mathcal {R} \\\\ \\sin \\mathcal {R} & \\cos \\mathcal {R} \\end{array} \\right] ^ {p _ {1}} \\left[ \\begin{array}{c c} 1 & 0 \\\\ 0 & - 1 \\end{array} \\right] ^ {p _ {2}} \\left[ \\begin{array}{c c} s & 0 \\\\ 0 & s \\end{array} \\right] ^ {p _ {3}} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.225, + 0.906, + 0.27 + ], + "angle": 0, + "content": "where \\((p_1, p_2, p_3) = (1, 0, 0)\\) when rotation is selected, \\((0, 1, 0)\\) when flip, \\((0, 0, 1)\\) when scale. \\(\\mathcal{R}\\) and \\(s\\) are the random amount of rotation and scale, \\(s \\in (0.5, 0.9)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.906, + 0.302 + ], + "angle": 0, + "content": "By feeding both \\( I \\) and \\( I_{\\mathrm{aug}} \\) into the network, we obtain two sets of output Gaussian distributions and angles:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.307, + 0.906, + 0.341 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} (\\Sigma , \\theta) = f _ {\\mathrm {n n}} (I) \\\\ (\\Sigma_ {\\mathrm {a u g}}, \\theta_ {\\mathrm {a u g}}) = f _ {\\mathrm {n n}} (I _ {\\mathrm {a u g}}) \\end{array} \\right. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.906, + 0.376 + ], + "angle": 0, + "content": "The consistency loss is calculated between the two sets so that the network learns the size and rotation variation:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.382, + 0.907, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s s}} = L _ {\\mathrm {G W D}} \\left(\\alpha \\Sigma \\alpha^ {\\top}, \\Sigma_ {\\mathrm {a u g}}\\right) + L _ {\\mathrm {A N G}} (m \\theta + \\mathcal {R}, \\theta_ {\\mathrm {a u g}}) \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.906, + 0.467 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) is the rotation angle and \\(m = 1\\) when rotation is selected; \\((m,\\mathcal{R}) = (-1,0)\\) when flip; \\((m,\\mathcal{R}) = (1,0)\\) when scale. \\(L_{\\mathrm{GWD}}(\\cdot)\\) is Gaussian Wasserstein Distance Loss [60] and \\(L_{\\mathrm{ANG}}(\\cdot)\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.474, + 0.907, + 0.496 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {A N G}} \\left(\\theta_ {1}, \\theta_ {2}\\right) = \\min _ {k \\in Z} \\left(s m o o t h _ {L 1} \\left(\\theta_ {1}, k \\pi + \\theta_ {2}\\right)\\right) \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.906, + 0.534 + ], + "angle": 0, + "content": "where \\(\\min (\\cdot)\\) regresses the prediction toward the closest target to circumvent the periodicity problem [65]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.75, + 0.561 + ], + "angle": 0, + "content": "3.6. Copy-paste Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Inspired by [9], we propose to crop the detected instances of step \\( k \\) and paste them on the training image of step \\( k + 1 \\). The maximum number of paste boxes is limited to 10 in each step. We simply use the bounding boxes of the cropped instances as the regression targets, and use Gaussian Wasserstein Distance Loss [60] to calculate \\( \\mathcal{L}_{\\mathrm{box}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.67, + 0.649, + 0.685 + ], + "angle": 0, + "content": "3.7. Overall Loss" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.693, + 0.9, + 0.708 + ], + "angle": 0, + "content": "The overall loss \\(\\mathcal{L}\\) for Point2RBox-v2 can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.715, + 0.907, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c l s}} + w _ {\\text {b o x}} \\mathcal {L} _ {\\text {b o x}} + w _ {\\mathrm {O}} \\mathcal {L} _ {\\mathrm {O}} + w _ {\\mathrm {W}} \\mathcal {L} _ {\\mathrm {W}} + w _ {\\mathrm {E}} \\mathcal {L} _ {\\mathrm {E}} + w _ {\\mathrm {s s}} \\mathcal {L} _ {\\mathrm {s s}} \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.738, + 0.907, + 0.799 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{cls}}\\) is the focal loss [25] for classification, \\(\\mathcal{L}_{\\mathrm{box}}\\) regresses boxes/centers toward copy-paste/ground-truth labels, \\(w_{\\mathrm{box}}\\) is set to one by default, \\((w_{\\mathrm{O}}, w_{\\mathrm{W}}, w_{\\mathrm{E}}, w_{\\mathrm{ss}})\\) are set to (10, 5, 0.3, 1) based on our ablation studies (see Tables 3-5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.645, + 0.831 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Experiments are carried out on NVIDIA RTX4090 GPUs using PyTorch 2.2.0 [36] and the rotation detection tool kits: MMRotate 1.0.0 [73]. All the experiments follow the same hyper-parameters (learning rate, batch size, optimizer, etc.)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19287" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.899, + 0.509 + ], + "angle": 0, + "content": "
Methods*PL1BDBRGTFSVLVSHTCBCSTSBFRAHASPHCAP50
▼RBox-supervised OOD
RepPoints (2019) [62]✓86.781.141.662.076.256.375.790.780.885.363.366.659.167.633.768.45
RetinaNet (2017) [25]✓88.277.045.069.471.559.074.590.884.979.357.364.762.766.539.668.69
GWD (2021) [56]✓89.375.447.861.979.573.886.190.984.579.455.959.763.271.045.471.66
FCOS (2019) [42]✓89.176.950.163.279.879.887.190.480.884.659.766.365.871.341.772.44
S2A-Net (2022) [12]✓89.283.052.574.678.879.287.590.984.984.861.968.070.771.459.875.81
▼HBox-supervised OOD
Sun et al. (2021) [41]×51.538.716.136.829.819.223.483.950.680.018.950.225.628.725.538.60
BoxInst-RBox (2021) [43]2×68.440.833.132.346.955.456.679.566.882.141.252.852.865.030.053.59
H2RBox (2023) [59]✓88.573.540.856.977.565.477.990.983.285.355.362.952.463.643.367.82
EIE-Det (2024) [45]✓87.770.241.560.580.776.386.390.982.684.753.164.558.170.443.870.10
H2RBox-v2 (2023) [65]✓89.074.450.060.579.875.386.990.985.185.059.263.265.270.549.772.31
▼Point-supervised OOD
Point2Mask-RBox (2023) [21]2×4.023.13.81.315.11.03.319.01.029.10.09.57.421.17.19.72
P2BNet+H2RBox (2023) [3, 59]×24.735.97.127.93.312.117.517.50.834.06.349.611.627.218.819.63
P2BNet+H2RBox-v2 (2023) [3, 65]×11.044.814.915.436.816.727.812.11.831.23.450.612.636.712.521.87
P2RBox (2024) [2]†×87.865.715.060.773.071.778.981.544.581.241.239.345.557.541.259.04
PointOBB (2024) [33]×26.165.79.159.465.834.929.80.52.316.70.649.021.841.036.730.08
Point2RBox (2024) [66]✓62.964.314.435.028.238.933.325.22.244.53.448.125.945.022.634.07
Point2RBox+SK (2024) [66]†✓53.363.93.750.940.039.245.776.710.556.15.449.524.251.233.840.27
Point2RBox+SK (2024) [66]†×66.459.55.252.654.153.957.390.83.257.86.147.422.955.740.544.90
PointOBB-v2 (2025) [38]×64.527.81.936.258.847.253.490.562.245.312.141.78.143.732.041.68
PointOBB-v3 (2025) [68]✓30.939.413.522.761.27.043.162.459.847.32.745.116.855.211.441.29
PointOBB-v3 (2025) [68]×52.954.421.352.765.644.967.887.226.773.432.653.339.056.410.249.24
Point2RBox-v2 (ours)✓78.452.78.340.971.060.574.788.765.572.124.426.130.150.721.051.00
Point2RBox-v2 (ours)×88.072.68.046.279.676.386.989.179.782.926.245.345.866.346.362.61
*Comparison tracks: √ = End-to-end training and testing; × = Generating pseudo labels to train the FCOS detector (two-stage training).†Using additional priors. P2RBox: Pre-trained SAM model; Point2RBox+SK: One-shot sketches for each class.†PL: Plane, BD: Baseball diamond, BR: Bridge, GTF: Ground track field, SV: Small vehicle, LV: Large vehicle, SH: Ship, TC: Tennis court, BC: Basketball court, ST: Storage tank, SBF: Soccer-ball field, RA: Roundabout, HA: Harbor, SP: Swimming pool, HC: Helicopter.‡RBox: The minimum rectangle operation is performed on the output Mask to obtain the RBox.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.187, + 0.512, + 0.807, + 0.526 + ], + "angle": 0, + "content": "Table 1. Detection performance of each category on the DOTA-v1.0 and the mean \\( {\\mathrm{{AP}}}_{50} \\) of all categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.541, + 0.486, + 0.722 + ], + "angle": 0, + "content": "Average precision (AP) is adopted as the primary metric. All the models are configured upon ResNet50 [13] and trained with AdamW [31]. 1) Learning rate. Initialized at 5e-5, warm-up for 500 iterations, and divided by ten at each decay step. 2) Epochs. 72 for HRSC; 12 for the others. 3) Augmentation. Random rotation/flip for HRSC; random flip for the others. 4) Image size. Split into \\(1,024 \\times 1,024\\) with an overlap of 200 for DOTA/FAIR1M/STAR; scaled to \\(800 \\times 800\\) for others. 5) Multi-scale. All experiments evaluated without multi-scale technique [73]. 6) Datasets. Six remote sensing and one retail scene datasets, covering all datasets used by the main counterparts [2, 33, 66]:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.735, + 0.485, + 0.78 + ], + "angle": 0, + "content": "- DOTA [48]. DOTA-v1.0 has 2,806 aerial images annotated with 15 categories, while DOTA-v1.5/2.0 are the extended versions with more small objects and categories." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.484, + 0.825 + ], + "angle": 0, + "content": "- DIOR [4]. It is an aerial image dataset re-annotated with RBoxes based on its original HBox version [18], with a high variation in object size and high intra-class diversity." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.484, + 0.855 + ], + "angle": 0, + "content": "- HRSC [30]. It contains ship instances on the sea and inshore. The train/val/test set includes 436/181/444 images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- FAIR1M [40]. It has more than 1 million instances and more than 40,000 images for fine-grained object recognition in remote sensing imagery, annotated with 37 cate" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.735, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.541, + 0.856, + 0.555 + ], + "angle": 0, + "content": "gories. The results are evaluated on FAIR1M-1.0." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.556, + 0.906, + 0.616 + ], + "angle": 0, + "content": "- STAR [22]. It is extensive for scene graph generation, covering more than 210,000 objects with diverse spatial resolutions, classified into 48 fine-grained categories and precisely annotated with oriented bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.616, + 0.907, + 0.661 + ], + "angle": 0, + "content": "- SKU110K [35]. It focuses on the detection of densely packed retail scenes with 110,712 objects in 11,762 images. The density reaches 86 instances per image." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.661, + 0.907, + 0.692 + ], + "angle": 0, + "content": "- RSAR [70]. It is a remote sensing dataset based on Synthetic Aperture Radar (SAR) imagery with 6 categories." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.556, + 0.907, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.767, + 0.724 + ], + "angle": 0, + "content": "4.1. Main Results on DOTA-v1.0" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.906, + 0.762 + ], + "angle": 0, + "content": "Table 1 compares Point2RBox-v2 with the state-of-the-art methods, which can be categorized into two tracks:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.884 + ], + "angle": 0, + "content": "1) End-to-end training. These methods apply the trained weakly-supervised detector directly to the test set. Without relying on priors, our approach demonstrates an improvement of \\(16.93\\%\\) (51.00% vs. 34.07%) compared to Point2RBox. Even when compared to Point2RBox+SK, which incorporates additional data-side priors (i.e. one-shot examples for each class), our method still outperforms it by \\(10.73\\%\\) (51.00% vs. 40.27%)." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.906, + 0.901 + ], + "angle": 0, + "content": "2) Two-stage training. These methods generate RBox" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19288" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.899, + 0.406 + ], + "angle": 0, + "content": "
Methods*DOTA-v1.0DOTA-v1.5DOTA-v2.0DIORHRSCFAIR1MSTARSKU110KRSAR
▼RBox-supervised OOD
RetinaNet (2017) [25]✓68.6960.5747.0054.9684.4937.6721.8078.5057.67
GWD (2021) [56]✓71.6663.2748.8757.6086.6739.1125.3079.1657.80
FCOS (2019) [42]✓72.4464.5351.7759.8388.9941.2528.1080.0966.66
S2A-Net (2022) [12]✓75.8166.5352.3961.4190.1042.4427.3080.3666.47
▼HBox-supervised OOD
Sun et al. (2021) [41]×38.60--------
KCR (2023) [74]✓----79.10----
H2RBox (2023) [59]✓70.0561.7048.6857.807.0335.9417.2057.1549.92
H2RBox-v2 (2023) [65]✓72.3164.7650.3357.6489.6642.2727.3070.7065.16
AFWS (2024) [32]✓72.5565.9251.7359.07-41.80---
▼Point-supervised OOD
P2RBox (2024) [2]†×59.04--------
PointSAM (2025) [27]†×---46.20-----
PointOBB (2024) [33]×30.0810.665.5337.31-11.199.19-13.80
Point2RBox+SK (2024) [66]†✓40.2730.5123.4327.3479.4020.037.863.4127.81
PointOBB-v2 (2025) [38]×41.6830.5920.6439.56-13.369.0056.6318.99
PointOBB-v3 (2025) [68]✓41.2031.2522.8237.60-11.4211.31-15.84
PointOBB-v3 (2025) [68]×49.2433.7923.5240.18-18.3512.85-22.60
Point2RBox-v2 (ours)✓51.0039.4527.1134.7082.6725.727.8064.0028.60
Point2RBox-v2 (ours)×62.6154.0638.7944.4586.1534.7114.2065.6430.90
*Comparison tracks: √= End-to-end training and testing; × = Generating pseudo labels to train the FCOS detector (two-stage training).†Using additional priors. P2RBox/PointSAM: Pre-trained SAM model; Point2RBox+SK: One-shot sketches for each class.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.411, + 0.886, + 0.426 + ], + "angle": 0, + "content": "Table 2. Accuracy \\( \\left( {\\mathrm{{AP}}}_{50}\\right) \\) comparisons on the DOTA-v1.0/1.5/2.0, DIOR, HRSC, FAIR1M, STAR, SKU110K, and RSAR datasets." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.435, + 0.353, + 0.522 + ], + "angle": 0, + "content": "
\\(w_{O}\\)DOTAHRSC
E2EFCOSE2EFCOS
348.7661.6281.8584.36
549.8162.4482.4685.76
1051.0062.6182.6786.15
3045.8857.8381.5685.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.526, + 0.347, + 0.541 + ], + "angle": 0, + "content": "Table 3. Ablation with the weight of \\( {\\mathcal{L}}_{\\mathrm{O}} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.372, + 0.435, + 0.626, + 0.522 + ], + "angle": 0, + "content": "
\\( {w}_{\\mathrm{W}} \\)DOTAHRSC
E2EFCOSE2EFCOS
350.8556.7878.4283.49
551.0062.6182.6786.15
1049.1560.5430.3735.13
3042.8452.5323.8925.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.375, + 0.526, + 0.62, + 0.541 + ], + "angle": 0, + "content": "Table 4. Ablation with the weight of \\( {\\mathcal{L}}_{\\mathrm{W}} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.647, + 0.435, + 0.896, + 0.522 + ], + "angle": 0, + "content": "
wEDOTAHRSC
E2EFCOSE2EFCOS
0.148.7557.6234.7139.45
0.351.0062.6182.6786.15
0.551.3662.6376.8585.22
1.049.0560.6356.5959.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.649, + 0.526, + 0.891, + 0.541 + ], + "angle": 0, + "content": "Table 5. Ablation with the weight of \\( {\\mathcal{L}}_{\\mathrm{E}} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.549, + 0.351, + 0.626 + ], + "angle": 0, + "content": "
wssDOTAHRSC
E2EFCOSE2EFCOS
0.149.2859.6673.6678.92
1.051.0062.6182.6786.15
3.049.1559.201.301.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.629, + 0.347, + 0.643 + ], + "angle": 0, + "content": "Table 6. Ablation with the weight of \\( {\\mathcal{L}}_{\\mathrm{{ss}}} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.372, + 0.55, + 0.895, + 0.625 + ], + "angle": 0, + "content": "
R / F / SDOTAHRSCR / F / SDOTAHRSC
90% / 10% / 0%60.4285.4680% / 20% / 0%59.4684.73
75% / 0% / 25%60.7986.2260% / 15% / 25%62.3884.21
68% / 7% / 25%62.6186.1538% / 37% / 25%45.878.56
45% / 5% / 50%60.5585.3440% / 10% / 50%60.4910.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.401, + 0.629, + 0.862, + 0.643 + ], + "angle": 0, + "content": "Table 7. Ablation with the proportion of augmented views in self-supervision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.734 + ], + "angle": 0, + "content": "labels on train/val sets, with which the FCOS detector is trained. In this two-stage mode, Point2RBox-v2 achieves an accuracy of \\(62.61\\%\\), considerably surpassing PointOBB series. Remarkably, it even outperforms the SAM-powered method P2RBox by \\(3.57\\%\\) (62.61% vs. 59.04%)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Class-wise analysis. The FCOS detector trained with labels generated by Point2RBox-v2 achieves accuracy nearly equivalent to RBox-supervised FCOS across six high-density categories: SH (86.9% vs. 87.1%), SV (79.6% vs. 79.8%), LV (76.3% vs. 79.8%), PL (88.0% vs. 89.1%), ST (82.9% vs. 84.6%), and TC (89.1% vs. 90.4%). Interestingly, these six high-density categories account for 88% of DOTA instances. By annotating these categories with points and generating RBoxes using Point2RBox-v2 while labeling the other sparse categories with RBoxes, we can significantly reduce annotation labor without sacrificing much accuracy," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.657, + 0.851, + 0.673 + ], + "angle": 0, + "content": "highlighting the valuable role our method can play." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.745, + 0.697 + ], + "angle": 0, + "content": "4.2. Results on More Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.908, + 0.901 + ], + "angle": 0, + "content": "The results are displayed in Table 2. On more challenging DOTA-v1.5/2.0, Point2RBox-v2 presents a similar trend, \\(23.47\\% / 18.15\\%\\) higher than PointOBB-v2 in the pseudogeneration track. On the ship detection dataset HRSC, the gap between Point2RBox-v2 and RBox-supervised FCOS is only \\(2.84\\%\\) (86.15% vs. 88.99%). DIOR is relatively sparse, leading to less improvement with our methods—lower than PointSAM (44.45% vs. 46.20%) but still higher than methods that do not use SAM. Our method also provides competitive performance on fine-grained datasets FAIR1M and STAR. In addition to remote sensing scenarios, we carry out experiments on SKU110K for densely packed retail scenes. Existing point-supervised methods struggle in this" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19289" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.268, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.089, + 0.422, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.089, + 0.575, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.089, + 0.728, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.09, + 0.882, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.21, + 0.688, + 0.224 + ], + "angle": 0, + "content": "Figure 5. Qualitative analysis on failed cases and overlap cases." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.24, + 0.476, + 0.378 + ], + "angle": 0, + "content": "
ModulesDOTAHRSC
\\( \\mathcal{L}_{\\mathrm{O}} \\)\\( \\mathcal{L}_{\\mathrm{W}} \\)\\( \\mathcal{L}_{\\mathrm{ss}} \\)\\( \\mathcal{L}_{\\mathrm{E}} \\)\\( CP \\)E2EFCOSE2EFCOS
✓0.000.000.000.00
✓✓41.5452.9817.9619.64
✓✓✓46.6454.2618.1022.13
✓✓✓✓49.5561.8878.7983.79
✓✓✓✓48.5859.5620.3524.76
✓✓✓✓38.9448.4411.6414.93
✓✓✓✓47.0855.0519.5821.78
✓✓✓✓✓51.0062.6182.6786.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.383, + 0.454, + 0.396 + ], + "angle": 0, + "content": "Table 8. Ablation with incremental addition of modules." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.407, + 0.475, + 0.445 + ], + "angle": 0, + "content": "
16K=2432|| 1.2β=1.62.0
50.8751.0048.08|| 48.1451.0051.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.448, + 0.471, + 0.462 + ], + "angle": 0, + "content": "Table 9. Ablation with \\( K \\) and \\( \\beta \\) in edge loss on DOTA (E2E)." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.472, + 0.475, + 0.548 + ], + "angle": 0, + "content": "
σPoint2RBoxPointOBB-v2Point2RBox-v2
DOTAHRSCDOTAHRSCDOTAHRSC
0%40.2779.4044.85-62.6186.15
10%39.6078.8142.30-61.5885.76
30%38.4278.2838.46-60.3185.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.551, + 0.466, + 0.565 + ], + "angle": 0, + "content": "Table 10. Ablation with the inaccuracy in point annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.611 + ], + "angle": 0, + "content": "case, whereas Point2RBox-v2 achieves performance on par with HBox-supervised H2RBox (65.64% vs. 57.15%)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.255, + 0.635 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.704 + ], + "angle": 0, + "content": "Tables 3-10 display the ablation studies on DOTA-v1.0 and HRSC. \"E2E\" denotes end-to-end training; \"FCOS\" denotes two-stage training (i.e. generating pseudo labels to train FCOS). The final values adopted are highlighted in gray." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Weight of each loss. Tables 3-5 determine the weights of the proposed losses. Based on these experiments, the weights \\((w_{\\mathrm{O}}, w_{\\mathrm{W}}, w_{\\mathrm{E}}, w_{\\mathrm{ss}})\\) are set to (10, 5, 0.3, 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Proportion of augmented views. Table 7 studies the proportion between rotation, flip, and scale. The results are reported with two-stage training (FCOS). Based on the results, the proportion is set to \\(68\\%\\), \\(7\\%\\), and \\(25\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Incremental addition of modules. Table 8 demonstrates the constraints from Gaussian and Voronoi achieve an accuracy of \\(52.98\\%\\) on DOTA. Adding consistency loss and edge loss further boosts it to \\(54.26\\%\\) and \\(61.88\\%\\), respectively, whereas the improvement from copy-paste is \\(0.73\\%\\). We also demonstrate the impact of omitting each core loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.905, + 0.288 + ], + "angle": 0, + "content": "Edge loss parameters. We set \\( K = 24 \\) and \\( \\beta = 1.6 \\) as they are observed to discern the correct edges during code development. Table 9 provides a more precise ablation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.908, + 0.38 + ], + "angle": 0, + "content": "Annotation inaccuracy. We offset the annotated points by a noise from the uniform distribution \\([- \\sigma H, + \\sigma H]\\), where \\(H\\) is the height of objects. Table 10 shows that the \\(\\mathrm{AP}_{50}\\) of Point2RBox-v2 decreases by less than \\(3\\%\\) when noise is added to point annotations, demonstrating the robustness of the proposed learning mechanisms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.39, + 0.685, + 0.403 + ], + "angle": 0, + "content": "4.4. More Discussions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.518 + ], + "angle": 0, + "content": "The qualitative analysis on the failed/overlap cases is shown in Fig. 5. 1) Failed cases. Although our method performs well overall, it struggles with certain categories that are sparse and not constrained by other objects. 2) Overlap cases. Minimizing overlap as a soft constraint during training does not entirely eliminate overlap. Once trained, the model remains robust to some overlap during inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.533, + 0.634, + 0.548 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.908, + 0.678 + ], + "angle": 0, + "content": "This paper introduces Point2RBox-v2, a point-supervised oriented object detector that effectively leverages the arrangement and layout of instances. We propose the integration of Gaussian overlay and Voronoi tessellation to constrain the size and rotation of instances based on their spatial relationships. Additionally, by incorporating self-supervised consistency loss, edge loss, and copy-paste augmentation, the accuracy of the model is further enhanced." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.909, + 0.845 + ], + "angle": 0, + "content": "Experiments yield the following observations: 1) The integration of Gaussian and Voronoi concepts effectively harnesses the spatial layout of objects, significantly enhancing point-supervised OOD. 2) Point2RBox-v2 demonstrates exceptional performance in densely packed scenes (see Fig. 2), where existing methods struggle. 3) Our method does not require priors (i.e. pre-trained SAM or one-shot examples) and is applicable to both end-to-end and pseudo-generation modes. 4) It advances the state of the art by a large amount, achieving \\(62.61\\%\\), \\(86.15\\%\\), and \\(34.71\\%\\) on the DOTA-v1.0, HRSC, and FAIR1M datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.845, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Limitations. The gap between Point2RBox-v2 and the RBox-supervised OOD is still huge in terms of sparse categories (i.e. BR/SBF) since little constraint can be obtained from the layout between them when the objects are sparse." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19290" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.119, + 0.188, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.486, + 0.186 + ], + "angle": 0, + "content": "[1] Franz Aurenhammer. Voronoi diagrams—a survey of a fundamental geometric data structure. ACM Computing Surveys, 23(3):345-405, 1991. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Guangming Cao, Xuehui Yu, Wenwen Yu, Xumeng Han, Xue Yang, Guorong Li, Jianbin Jiao, and Zhenjun Han. P2rbox: Point prompt oriented object detection with SAM. arXiv preprint arXiv:2311.13128, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.484, + 0.312 + ], + "angle": 0, + "content": "[3] Pengfei Chen, Xuehui Yu, Xumeng Han, Najmul Hassan, Kai Wang, Jiachen Li, Jian Zhao, Humphrey Shi, Zhenjun Han, and Qixiang Ye. Point-to-box network for accurate object detection via single point supervision. In European Conference on Computer Vision, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.483, + 0.368 + ], + "angle": 0, + "content": "[4] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[5] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2849-2858, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.484, + 0.508 + ], + "angle": 0, + "content": "[6] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.485, + 0.578 + ], + "angle": 0, + "content": "[7] Kun Fu, Zhonghan Chang, Yue Zhang, Guangluan Xu, Keshu Zhang, and Xian Sun. Rotation-aware and multi-scale convolutional neural network for object detection in remote sensing images. ISPRS Journal of Photogrammetry and Remote Sensing, 161:294-308, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[8] Shuyong Gao, Wei Zhang, Yan Wang, Qianyu Guo, Chenglong Zhang, Yangji He, and Wenqiang Zhang. Weakly-supervised salient object detection using point supervision. In AAAI Conference on Artificial Intelligence, pages 670-678, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[9] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D. Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[10] Eran Goldman, Roei Herzig, Aviv Eisenschat, Jacob Goldberger, and Tal Hassner. Precise detection in densely packed scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5236, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[11] Jiaming Han, Jian Ding, Nan Xue, and Gui-Song Xia. Redet: A rotation-equivariant detector for aerial object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2785-2794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[12] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Con-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.145, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.907, + 0.12 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.164 + ], + "angle": 0, + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In IEEE International Conference on Computer Vision, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.907, + 0.207 + ], + "angle": 0, + "content": "[15] Liping Hou, Ke Lu, Xue Yang, Yuqiu Li, and Jian Xue. Grep: Gaussian representation for arbitrary-oriented object detection. Remote Sensing, 15(3):757, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.907, + 0.263 + ], + "angle": 0, + "content": "[16] Javed Iqbal, Muhammad Akhtar Munir, Arif Mahmood, Afsheen Rafaqat Ali, and Mohsen Ali. Leveraging orientation for weakly supervised object detection with application to firearm localization. Neurocomputing, 440:310-320, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.907, + 0.333 + ], + "angle": 0, + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.907, + 0.39 + ], + "angle": 0, + "content": "[18] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing, 159:296-307, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[19] Qingyun Li, Yushi Chen, Xinya Shu, Dong Chen, Xin He, Yi Yu, and Xue Yang. A simple aerial detection baseline of multimodal language models. arXiv preprint arXiv:2501.09720, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.907, + 0.505 + ], + "angle": 0, + "content": "[20] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented leptopoints for aerial object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1829-1838, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.507, + 0.907, + 0.561 + ], + "angle": 0, + "content": "[21] Wentong Li, Yuqian Yuan, Song Wang, Jianke Zhu, Jianshu Li, Jian Liu, and Lei Zhang. Point2mask: Point-supervised panoptic segmentation via optimal transport. In IEEE International Conference on Computer Vision, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.907, + 0.658 + ], + "angle": 0, + "content": "[22] Yansheng Li, Linlin Wang, Tingzhu Wang, Xue Yang, Junwei Luo, Qi Wang, Youming Deng, Wenbin Wang, Xian Sun, Haifeng Li, Bo Dang, Yongjun Zhang, Yi Yu, and Junchi Yan. Star: A first-ever dataset and a large-scale benchmark for scene graph generation in large-size satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1832-1849, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.906, + 0.717 + ], + "angle": 0, + "content": "[23] Minghui Liao, Zhen Zhu, Baoguang Shi, Gui-Song Xia, and Xiang Bai. Rotation-sensitive regression for oriented scene text detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5909-5918, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.907, + 0.786 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International Journal of Computer Vision, 128(2):261-318, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19291" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.147 + ], + "angle": 0, + "content": "[27] Nanqing Liu, Xun Xu, Yongyi Su, Haojie Zhang, and Heng-Chao Li. Pointsam: Pointly-supervised segment anything model for remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 63:1-15, 2025. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.204 + ], + "angle": 0, + "content": "[28] Xuebo Liu, Ding Liang, Shi Yan, Dagui Chen, Yu Qiao, and Junjie Yan. Fots: Fast oriented text spotting with a unified network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5676-5685, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[29] Yuekai Liu, Hongli Gao, Liang Guo, Aoping Qin, Canyu Cai, and Zhichao You. A data-flow oriented deep ensemble learning method for real-time surface defect inspection. IEEE Transactions on Instrumentation and Measurement, 69(7): 4681-4691, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[30] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International Conference on Pattern Recognition Applications and Methods, pages 324-331, 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.484, + 0.441 + ], + "angle": 0, + "content": "[32] Junyan Lu, Qinglei Hu, Ruifei Zhu, Yali Wei, and Tie Li. Afws: Angle-free weakly-supervised rotating object detection for remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[33] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. Pointobb: Learning oriented object detection via single point supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[34] Jeffri Murragarra-Llerena, Lucas N Kirsten, Luis Felipe Zeni, and Claudio R Jung. Probabilistic intersection-over-union for training and evaluation of oriented object detectors. IEEE Transactions on Image Processing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[35] Xingjia Pan, Yuqiang Ren, Kekai Sheng, Weiming Dong, Haolei Yuan, Xiaowei Guo, Chongyang Ma, and Changsheng Xu. Dynamic refinement network for oriented and densely packed object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11207-11216, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.484, + 0.76 + ], + "angle": 0, + "content": "[36] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[37] Wen Qian, Xue Yang, Silong Peng, Junchi Yan, and Yue Guo. Learning modulated loss for rotated object detection. In AAAI Conference on Artificial Intelligence, pages 2458-2466, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[38] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. Pointobb-v2: Towards simpler, faster, and stronger single point supervised oriented object detection. In International Conference on Learning Representations, 2025. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[39] Xavier Soria, Yachuan Li, Mohammad Rouhani, and Angel D. Sappa. Tiny and efficient model for the edge detection" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "generalization. In IEEE/CVF International Conference on Computer Vision Workshops, pages 1364-1373, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.217 + ], + "angle": 0, + "content": "[40] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, Martin Weinmann, Stefan Hinz, Cheng Wang, and Kun Fu. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.908, + 0.289 + ], + "angle": 0, + "content": "[41] Yongqing Sun, Jie Ran, Feng Yang, Chenqiang Gao, Takayuki Kurozumi, Hideaki Kimata, and Ziqi Ye. Oriented object detection for remote sensing images based on weakly supervised learning. In IEEE International Conference on Multimedia & Expo Workshops, pages 1-6, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.908, + 0.345 + ], + "angle": 0, + "content": "[42] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In IEEE/CVF International Conference on Computer Vision, pages 9626-9635, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.908, + 0.403 + ], + "angle": 0, + "content": "[43] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.908, + 0.459 + ], + "angle": 0, + "content": "[44] L. Vincent and P. Soille. Watersheds in digital spaces: an efficient algorithm based on immersion simulations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(6):583-598, 1991. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.908, + 0.531 + ], + "angle": 0, + "content": "[45] Linfei Wang, Yibing Zhan, Xu Lin, Baosheng Yu, Liang Ding, Jianqing Zhu, and Dapeng Tao. Explicit and implicit box equivariance learning for weakly-supervised rotated object detection. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.908, + 0.586 + ], + "angle": 0, + "content": "[46] Long Wen, Yu Cheng, Yi Fang, and Xinyu Li. A comprehensive survey of oriented object detection in remote sensing images. Expert Systems with Applications, page 119960, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.908, + 0.643 + ], + "angle": 0, + "content": "[47] Hongjin Wu, Ruoshan Lei, and Yibing Peng. Pcbnet: A lightweight convolutional neural network for defect inspection in surface mount technology. IEEE Transactions on Instrumentation and Measurement, 71:1-14, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.908, + 0.715 + ], + "angle": 0, + "content": "[48] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.908, + 0.771 + ], + "angle": 0, + "content": "[49] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In IEEE/CVF International Conference on Computer Vision, pages 3520-3529, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.908, + 0.815 + ], + "angle": 0, + "content": "[50] Xue Yang and Junchi Yan. Arbitrary-oriented object detection with circular smooth label. In European Conference on Computer Vision, pages 677-694, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.908, + 0.87 + ], + "angle": 0, + "content": "[51] Xue Yang and Junchi Yan. On the arbitrary-oriented object detection: Classification based approaches revisited. International Journal of Computer Vision, 130(5):1340-1365, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[52] Xue Yang, Hao Sun, Kun Fu, Jirui Yang, Xian Sun, Menglong Yan, and Zhi Guo. Automatic ship detection in remote" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19292" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "sensing images from google earth of complex scenes based on multiscale rotation dense feature pyramid networks. Remote sensing, 10(1):132, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[53] Xue Yang, Jirui Yang, Junchi Yan, Yue Zhang, Tengfei Zhang, Zhi Guo, Xian Sun, and Kun Fu. Scdet: Towards more robust detection for small, cluttered and rotated objects. In IEEE/CVF International Conference on Computer Vision, pages 8231-8240, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.484, + 0.264 + ], + "angle": 0, + "content": "[54] Xue Yang, Liping Hou, Yue Zhou, Wentao Wang, and Junchi Yan. Dense label encoding for boundary discontinuity free rotation detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15814-15824, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.267, + 0.484, + 0.322 + ], + "angle": 0, + "content": "[55] Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3det: Refined single-stage detector with feature refinement for rotating object. In AAAI Conference on Artificial Intelligence, pages 3163-3171, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.326, + 0.484, + 0.393 + ], + "angle": 0, + "content": "[56] Xue Yang, Junchi Yan, Ming Qi, Wentao Wang, Xiaopeng Zhang, and Tian Qi. Rethinking rotated object detection with gaussian Wasserstein distance loss. In 38th International Conference on Machine Learning, pages 11830-11841, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.398, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[57] Xue Yang, Xiaojiang Yang, Jirui Yang, Qi Ming, Wentao Wang, Qi Tian, and Junchi Yan. Learning high-precision bounding box for rotated object detection via kullback-leibler divergence. In Advances in Neural Information Processing Systems, pages 18381-18394, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.484, + 0.538 + ], + "angle": 0, + "content": "[58] Xue Yang, Junchi Yan, Wenlong Liao, Xiaokang Yang, Jin Tang, and Tao He. Scdet++: Detecting small, cluttered and rotated objects via instance-level feature denoising and rotation loss smoothing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2384-2399, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.484, + 0.596 + ], + "angle": 0, + "content": "[59] Xue Yang, Gefan Zhang, Wentong Li, Xuehui Wang, Yue Zhou, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. International Conference on Learning Representations, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.484, + 0.668 + ], + "angle": 0, + "content": "[60] Xue Yang, Gefan Zhang, Xiaojiang Yang, Yue Zhou, Wentao Wang, Jin Tang, Tao He, and Junchi Yan. Detecting rotated objects as gaussian distributions and its 3-d generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4335-4354, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.484, + 0.726 + ], + "angle": 0, + "content": "[61] Xue Yang, Yue Zhou, Gefan Zhang, Jirui Yang, Wentao Wang, Junchi Yan, Xiaopeng Zhang, and Qi Tian. The kfiou loss for rotated object detection. In International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.73, + 0.484, + 0.784 + ], + "angle": 0, + "content": "[62] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Reppoints: Point set representation for object detection. In IEEE/CVF International Conference on Computer Vision, pages 9656-9665, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.484, + 0.841 + ], + "angle": 0, + "content": "[63] Yi Yu and Feipeng Da. Phase-shifting coder: Predicting accurate orientation in oriented object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[64] Yi Yu and Feipeng Da. On boundary discontinuity in angle regression based arbitrary oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-14, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.16 + ], + "angle": 0, + "content": "[65] Yi Yu, Xue Yang, Qingyun Li, Yue Zhou, Feipeng Da, and Junchi Yan. H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. In Advances in Neural Information Processing Systems, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.908, + 0.232 + ], + "angle": 0, + "content": "[66] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2rbox: Combine knowledge from synthetic visual patterns for end-to-end oriented object detection with single point supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.908, + 0.302 + ], + "angle": 0, + "content": "[67] Yi Yu, Xue Yang, Yansheng Li, Zhenjun Han, Feipeng Da, and Junchi Yan. Wholly-wood: Wholly leveraging diversified-quality labels for weakly-supervised oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.373 + ], + "angle": 0, + "content": "[68] Peiyuan Zhang, Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Yue Zhou, Xiaosong Jia, Xudong Lu, Jingdong Chen, Xiang Li, et al. Pointobb-v3: Expanding performance boundaries of single point-supervised oriented object detection. arXiv preprint arXiv:2501.13898, 2025. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.907, + 0.43 + ], + "angle": 0, + "content": "[69] Shun Zhang, Jihui Long, Yaohui Xu, and Shaohui Mei. Pmho: Point-supervised oriented object detection based on segmentation-driven proposal generation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.907, + 0.485 + ], + "angle": 0, + "content": "[70] Xin Zhang, Xue Yang, Yuxuan Li, Jian Yang, Ming-Ming Cheng, and Xiang Li. Rsar: Restricted state angle resolver and rotated sar benchmark. arXiv preprint arXiv:2501.04440, 2025.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.907, + 0.542 + ], + "angle": 0, + "content": "[71] Zhong-Qiu Zhao, Peng Zheng, Shou-Tao Xu, and Xindong Wu. Object detection with deep learning: A review. IEEE Transactions on Neural Networks and Learning Systems, 30 (11):3212-3232, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.907, + 0.599 + ], + "angle": 0, + "content": "[72] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: An efficient and accurate scene text detector. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.907, + 0.669 + ], + "angle": 0, + "content": "[73] Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, et al. MMrotate: A rotated object detection benchmark using pytorch. In 30th ACM International Conference on Multimedia, pages 7331-7334, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.908, + 0.74 + ], + "angle": 0, + "content": "[74] Tianyu Zhu, Bryce Ferenczi, Pulak Purkait, Tom Drummond, Hamid Rezatofighi, and Anton van den Hengel. Knowledge combination to learn rotated detection without rotated annotation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19293" + } + ] +] \ No newline at end of file diff --git a/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/d3a3a010-6834-441e-beb3-f3f784483984_origin.pdf b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/d3a3a010-6834-441e-beb3-f3f784483984_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..84f7a31794fe6702973e2f75f8fa716cc99dffef --- /dev/null +++ b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/d3a3a010-6834-441e-beb3-f3f784483984_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56b07b3ed511247eb52d3ff1de4f7df9666789e9a9f4b6d98e30612631422e69 +size 4109831 diff --git a/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/full.md b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/full.md new file mode 100644 index 0000000000000000000000000000000000000000..232bb0fa45e7c765d23bd93a017b3125ba8de12f --- /dev/null +++ b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/full.md @@ -0,0 +1,479 @@ +# Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances + +Yi Yu $^{1*}$ , Botao Ren $^{2*}$ , Peiyuan Zhang $^{3*}$ , Mingxin Liu $^{4}$ , Junwei Luo $^{3}$ , Shaofeng Zhang $^{4}$ , Feipeng Da $^{1}$ , Junchi Yan $^{4}$ , Xue Yang $^{4\dagger}$ + +$^{1}$ Southeast University $^{2}$ Tsinghua University $^{3}$ Wuhan University $^{4}$ Department of Automation & CSE & SAI, Shanghai Jiao Tong University + +{yuyi,dafp}@seu.edu.cn rbt22@mails.tsinghua.edu.cn {peiyuanzhangwhu,luojunwei}@whu.edu.cn + +{liumingxin,zhangshaofeng,yanjunchi, yangxue-2019-sjtu}@sjtu.edu.cn + +https://github.com/VisionXLab/point2rbox-v2 + +# Abstract + +With the rapidly increasing demand for oriented object detection (OOD), recent research involving weakly-supervised detectors for learning OOD from point annotations has gained great attention. In this paper, we rethink this challenging task setting with the layout among instances and present Point2RBox-v2. At the core are three principles: 1) Gaussian overlap loss. It learns an upper bound for each instance by treating objects as 2D Gaussian distributions and minimizing their overlap. 2) Voronoi watershed loss. It learns a lower bound for each instance through watershed on Voronoi tessellation. 3) Consistency loss. It learns the size/rotation variation between two output sets with respect to an input image and its augmented view. Supplemented by a few devised techniques, e.g. edge loss and copy-paste, the detector is further enhanced. To our best knowledge, Point2RBox-v2 is the first approach to explore the spatial layout among instances for learning point-supervised OOD. Our solution is elegant and lightweight, yet it is expected to give a competitive performance especially in densely packed scenes: $62.61\% /86.15\% /34.71\%$ on DOTA/HRSC/FAIR1M. + +# 1. Introduction + +Emerged as an essential task in computer vision, oriented object detection (OOD) has become a prominent demand in autonomous driving [6], aerial images [7, 30, 48, 51, 52], scene text [23, 28, 72], retail scenes [10, 35], industrial inspection [29, 47], and more. + +Manual annotations are essential to teach the detector new concepts of visual objects. Early research is often super + +![](images/312565b9515f1af73282525218512fdb213bafa5407c82fb3751e9830a45ac65.jpg) +Figure 1. Related methods, their principles for knowledge mining, whether using additional priors, and performance on DOTA-v1.0. + +vised by rotated bounding boxes (RBoxes), where annotations align with the desired output. While this approach has demonstrated promising results, RBox labels are considerably expensive. The cost of annotating each RBox is approximately $36.5\%$ higher than a horizontal bounding box (HBox) and $104.8\%$ higher than a point [66]. This highlights the importance of utilizing coarser labels for weakly-supervised oriented detection. Recent advancements in HBox-supervised OOD, particularly with H2RBox [59] and H2RBox-v2 [65], have shown promise in bridging the gap between HBox-and RBox-supervised methods, thereby lessening the dependence on labor-intensive RBox labeling. + +Following HBox-supervised methods, the more challenging point-supervised OOD opens up a new avenue. Several innovative approaches have emerged in the past year (see Fig. 1): 1) Point-prompt OOD. P2RBox [2], PMHO [69], and PointSAM [27] employ the zero-shot ability of SAM [17]. 2) Pseudo generation. PointOBB [33] and its v2/v3 version [38, 68] use multiple instance learning and class probability map for RBox generation. 3) Knowledge combination. Point2RBox [66] learns from one-shot examples. However, existing methods overlook the spatial relationships between objects. This brings us to the motivation of this paper. + +![](images/b63c1ba009eff7f950c198770e57cd82ae19f8188beced4bc1bfb304c4dd30bc.jpg) + +![](images/bf2e91e935cbc8d25a72d01b0a5f5748c83cb7adeb106676947de0a84f178c68.jpg) + +![](images/b60dab802c0e0688e36305437db8d85bc1053c441c799691fa9caa51e807c684.jpg) + +![](images/172107eaaa86ce0c6fe6de5bdeb54f7a291ca8116068561a1bb8984311d12b73.jpg) + +![](images/a28d2a33aaf6741c75ba6fd7d0f75bdf78427d22b2ec892c5049d5bbcb093f88.jpg) + +![](images/776a7aff52f3b88b79e08e9a00a9ffe80ae4a8d915e0b9b268d7990201fae575.jpg) + +![](images/ef2e656aa692901b0e17f433fb4c56053618fc8bbeb2ed5d01caa3e3515f6f9b.jpg) + +![](images/9d72e47dbcea8085f9d67829db6eeca8a209d158d6523e0fce9346adbe4c6bb8.jpg) + +![](images/da7da2d16ef61360469eb00e29d5c9501b0b9206f87873470b38813b913a8066.jpg) + +![](images/a627dc09b71c35b05b41399cff5a9b067e79b77869fbba346af16badad417758.jpg) + +![](images/300ce1cd207b24d07a541bd9e4577ebacf925f091e39392bfdff2eb76cd3a151.jpg) +Figure 2. Visual comparisons with state-of-the-art methods including: Point2Mask (2023) [21], PointOBB (2024) [33], PointOBB-v2 (2025) [38], PointOBB-v3 (2025) [68], and Point2RBox (2024) [66]. The boxes detected by our method (last row) wrap the objects more tightly. + +![](images/a465eef682ae40c9eddd692767959461c348b5b7d30a7bec551e1c65cd45edca.jpg) + +![](images/1f315e80353531f669466f8a0b9b01944960bcacdae9df85b38af33ab04a8483.jpg) + +![](images/cc63f69098b5c6a2217ed56b43849dfbfd441be35c9a7caf06efa322f88822c2.jpg) + +![](images/32f2371c429606c2eacbae6fad263eb1f0335f9ffd21efe8ccb5a28850ca27dd.jpg) + +Motivation. While point-supervised OOD has gained attention, the utilization of relationships among instances remains absent in the literature. For example, the arrangement of vehicles in a parking lot can serve as effective constraints for learning their size and orientation. This approach may yield valuable insights into point-supervised OOD, particularly in densely packed scenes where current methods face significant challenges. Can we leverage the spatial layout of objects to enhance point-supervised OOD? In this paper, we dive into this idea and seek to answer this question. + +Highlights. 1) Point2RBox-v2 is proposed for point-supervised OOD, advancing the state of the art as displayed in Fig. 2 and Tables 1-2. 2) We propose novel and elegant losses to enforce constraints from the spatial layout among instances based on Gaussian overlap and Voronoi tessellation [1]. 3) Other modules are devised (i.e. edge loss, consistency loss, copy-paste) to further enhance the method. + +Contributions. 1) To our best knowledge, this work is the first attempt to learn point-supervised OOD from the layout among objects, where we propose novel principles based on Gaussian overlap and Voronoi tessellation. 2) The training pipeline and detailed implementation are elucidated, with necessary modules (i.e. edge loss, consistency loss, copy-paste) incorporated. The source code will be made publicly available. 3) Extensive experiments demonstrate that leveraging the spatial layout of instances can significantly advance the state of the art, surpassing other alternatives in accuracy. + +# 2. Related Work + +# 2.1. RBox-supervised Oriented Detection + +In addition to horizontal detection [26, 71], oriented object detection (OOD) [46, 52] has received extensive attention. Representative works include anchor-based detector Rotated RetinaNet [25], anchor-free detector Rotated FCOS [42], and two-stage solutions, e.g. RoI Transformer [5], Oriented R-CNN [49], and ReDet [11]. Some research enhances the detector by exploiting alignment features, e.g. $\mathrm{R}^3\mathrm{Det}$ [55] and $\mathrm{S}^2\mathrm{A}$ -Net [12]. The angle regression may face boundary discontinuity and remedies are developed, including modulated losses [37, 53, 58] that alleviate loss jumps, angle coders [50, 51, 54, 63] that convert the angle into boundary-free coded data, and Gaussian-based losses [34, 56, 57, 60, 61] transforming rotated bounding boxes into Gaussian distributions. RepPoint-based methods [15, 20, 62] provide alternatives that predict a set of points that bounds the spatial extent of an object. LMMRotate [19] is a new paradigm of OOD based on multimodal language model. + +# 2.2. Point-supervised Oriented Detection + +Recently, several methods for point-supervised OOD have been proposed: 1) P2RBox [2], PMHO [69], and PointSAM [27] propose point-prompt OOD by employing the zero-shot Point-to-Mask ability of SAM [17]. 2) Point2RBox [66] introduces an end-to-end approach based on knowledge com + +![](images/7f7ab96800cefc9e462732b3d097e16a231a98bbb2684aa785ee6bb13d3a5146.jpg) +Figure 3. The training pipeline of Point2RBox-v2. Gaussian overlap loss and Voronoi watershed loss utilize the spatial layout (see Fig. 4), while edge loss (see Sec. 3.4), symmetry-aware learning (see Sec. 3.5), and copy-paste (see Sec. 3.6) further enhance the method. + +bination. 3) PointOBB [33, 68] achieves RBox generation through multiple instance learning. 4) PointOBB-v2 [38] learns a class probability map to generate pseudo RBoxes. 5) Wholly-WOOD [67] uses points, HBoxes, RBoxes, or their combinations for mixed supervised OOD. + +Among these methods, P2RBox, PMHO, and PointSAM rely on the SAM model pre-trained on massive labeled datasets, whereas Point2RBox requires one-shot examples for each category. In contrast, PointOBB series do not use many priors, but they necessitate two-stage training. + +# 2.3. Other Weakly-supervised Settings + +Compared to Point-to-RBox, some other settings have been better studied. These methods are potentially applicable to our Point-to-RBox task setting by using a cascade pipeline, such as Point-to-HBox-to-RBox. In our experiment, cascade pipelines powered by state-of-the-art approaches are also compared. Here, representative works are introduced. + +HBox-to-RBox. H2RBox [59] establishes a paradigm that limits the object to a few candidate angles through geometric constraint from HBoxes, with a self-supervised branch eliminating the undesired results. An enhanced version H2RBox-v2 [65] is proposed to leverage the reflection symmetry of objects to further boost the accuracy. EIE-Det [45] uses an explicit equivariance branch for learning rotation consistency, and an implicit equivariance branch for learning position, aspect ratio, and scale consistency. KCR [74] combines RBox- and HBox-annotated datasets for transfer learning. Some studies [16, 41] use additional annotated data for training, which are also attractive but less general. + +Point-to-HBox. Several related approaches have been developed, including: 1) P2BNet [3] samples box proposals of different sizes around the labeled point and classify them to achieve point-supervised horizontal object detection. 2) PSOD [8] achieves point-supervised salient object detection using an edge detector and adaptive masked flood fill. + +Point-to-Mask. Point2Mask [21] is proposed to achieve panoptic segmentation using single point annotation per target. SAM (Segment Anything Model) [17] produces object masks from input point/HBox prompts. Though RBoxes can be obtained from the segmentation mask by finding the minimum circumscribed rectangle, such a complex pipeline can be less cost-efficient and perform worse [59, 65]. + +# 3. Method + +# 3.1. Overview and Preliminary + +An overview of Point2RBox-v2 is illustrated in Fig. 3. The network is based on ResNet50 [13] backbone, FPN [24] head, and PSC [64] angle coder. Objects of varying sizes are typically assigned to different FPN layers based on the scale. However, points lack size, so we assign them all to the P3 layer (stride = 8). Assume the detector $f_{\mathrm{nn}}(\cdot)$ maps an image $I$ to a set of RBoxes as detection results: + +$$ +\left(x _ {c}, y _ {c}, w, h, \theta\right) = f _ {\mathrm {n n}} (I) \tag {1} +$$ + +Equivalently, oriented objects can also be represented by 2D Gaussian distributions $\mathcal{N}(\mu, \Sigma)$ [60]: + +$$ +\left\{\mu = \left[ \begin{array}{l l} x _ {c} & y _ {c} \end{array} \right], \quad \Sigma = \mathbf {R} \left[ \begin{array}{c c} w / 2 & 0 \\ 0 & h / 2 \end{array} \right] ^ {2} \mathbf {R} ^ {\top} \right\} \tag {2} +$$ + +where + +$$ +\mathbf {R} = \left[ \begin{array}{c c} \cos \theta & - \sin \theta \\ \sin \theta & \cos \theta \end{array} \right] \tag {3} +$$ + +At the core of Point2RBox-v2 is to utilize the constraints from the layout, which is achieved by Gaussian overlap loss (Sec. 3.2) and Voronoi watershed loss (Sec. 3.3). These two losses effectively limit the size and rotation of objects to a reasonable range. Upon that, edge loss (Sec. 3.4) aligns the bounding box with the edge of objects to improve the accuracy. Incorporated with symmetry-aware learning (Sec. 3.5) and copy-paste augmentation (Sec. 3.6), we achieve a + +![](images/e493037026f870df4b6d99c842b610fed2377006c2b5ec8358a853e93a79d1fc.jpg) +Figure 4. To illustrate the procedure of the three newly proposed loss functions and their impact on the learning results. (a) Gaussian overlap loss (see Sec. 3.2). (b) Voronoi watershed loss (see Sec. 3.3). (c) Edge loss (see Sec. 3.4). + +stable and high-accuracy solution for point-supervised OOD. In subsequent subsections, these modules are detailed. + +# 3.2. Gaussian Overlap Loss + +As mentioned, oriented objects can be represented by 2D Gaussian distributions $\mathcal{N}(\mu, \Sigma)$ . The overlap volume between two distributions $\mathcal{N}_1(\mu_1, \Sigma_1)$ and $\mathcal{N}_2(\mu_2, \Sigma_2)$ can be approximated by the Bhattacharyya coefficient [60] as: + +$$ +B \left(\mathcal {N} _ {1}, \mathcal {N} _ {2}\right) = \exp \left(- \frac {1}{8} \mu^ {\top} \Sigma^ {- 1} \mu\right) \cdot \frac {\left| \Sigma_ {1} \right| ^ {1 / 4} \left| \Sigma_ {2} \right| ^ {1 / 4}}{\left| \Sigma \right| ^ {1 / 2}} \tag {4} +$$ + +where $\mu = \mu_{2} - \mu_{1}$ , $\Sigma = \frac{1}{2} (\Sigma_{1} + \Sigma_{2})$ , $|\Sigma|$ denotes the determinant of the covariance matrix. + +Based on the above equation, we build a Gaussian overlap matrix $\mathbf{M} \in \mathbb{R}^{N \times N}$ for each scene image as: + +$$ +\mathbf {M} _ {i, j} = B \left(\mathcal {N} _ {i}, \mathcal {N} _ {j}\right) \tag {5} +$$ + +where $i,j = 1,2,\ldots ,N;N$ is the distribution count (or the instance count) within one training image. + +The Gaussian overlap loss can then be expressed as: + +$$ +\mathcal {L} _ {O} = \frac {1}{N} \sum_ {i \neq j} \left(\mathbf {M} _ {i, j}\right) \tag {6} +$$ + +where $i \neq j$ omits diagonal elements. With this loss, the + +detector learns to arrange instances (see Fig. 4a) based on the mutual exclusivity among instances. + +# 3.3. Voronoi Watershed Loss + +A Voronoi diagram [1] is a partitioning of a space based on a set of points. Point annotations can be effectively utilized to calculate a Voronoi diagram, where a distinct polygon region is assigned to each point-annotated instance. The watershed algorithm [44], on the other hand, is a region segmentation technique that treats the intensity of pixels as a topographic surface, identifying regions as "catchment basins". + +The calculation of Voronoi ridges can be formulated as: + +$$ +V = \text {V o r o n o i} (X) \tag {7} +$$ + +where $X$ are the annotated points within a training image; $V$ are the output Voronoi ridges (pixel coordinates). + +Interestingly, we find that Voronoi diagrams can be utilized as initial markers for watershed to obtain a region for each instance (see Fig. 4b). In concrete terms, the points $X$ can be employed as foreground markers, while the Voronoi ridges $V$ can act as background boundaries: + +$$ +S = \text {W a t e r s h e d} (I, X, V) \tag {8} +$$ + +where $S$ are the output basin regions (pixel coordinates) corresponding to each annotated instance. By rotating $S$ + +to align with the direction of the current prediction, the regression target of width and height can be expressed as: + +$$ +\left[ \begin{array}{l} w _ {t} \\ h _ {t} \end{array} \right] = 2 \max \left| \mathbf {R} ^ {\top} \left(S - \left[ \begin{array}{l} x _ {c} \\ y _ {c} \end{array} \right]\right) \right| \tag {9} +$$ + +where $(x_{c},y_{c},w,h,\theta)$ is the current prediction; $\mathbf{R}$ is defined by Eq. (3); $w_{t}$ and $h_t$ are detached to stop the gradient. + +Afterward, the Voronoi watershed loss to regress the width and height of objects can be calculated as: + +$$ +\mathcal {L} _ {W} = L _ {\mathrm {G W D}} \left(\left[ \begin{array}{c c} w / 2 & 0 \\ 0 & h / 2 \end{array} \right] ^ {2}, \left[ \begin{array}{c c} w _ {t} / 2 & 0 \\ 0 & h _ {t} / 2 \end{array} \right] ^ {2}\right) \tag {10} +$$ + +where $L_{\mathrm{GWD}}(\cdot)$ is Gaussian Wasserstein Distance Loss [60]. + +# 3.4. Edge Loss + +The above two losses have limited the size to a reasonable range. To make it more accurate, we propose the edge loss to snap the boundaries toward the edges (see Fig. 4c). + +First, a region $P \in \mathbb{R}^{(2K + 1) \times (2K + 1)}$ around each predicted RBox is extracted via Rotated RoI Align [14]: + +$$ +P = R o I A l i g n (E (I), \left(x _ {c}, y _ {c}, \beta w, \beta h, \theta\right)) \tag {11} +$$ + +where $E(\cdot)$ is the edge detection function [39]. We set $K = 24$ and $\beta = 1.6$ in our experiments (see Table 9). + +By calculating the sum of each row of $P$ , the edge distribution in $y$ direction can be obtained as: + +$$ +\mu_ {i} = \sum_ {j = 1} ^ {2 K + 1} \left(P _ {(K + 1 - i), j} + P _ {(K + 1 + i), j}\right) \tag {12} +$$ + +where $i = 1,2,\dots ,K$ , indicating that the upper half of $P$ is reversed and added to the lower half. + +Meanwhile, the current prediction of the edge can also be softened into a distribution as: + +$$ +\lambda_ {i} = \exp \left(- \frac {(i - K / \beta) ^ {2}}{2 \sigma_ {E} {} ^ {2}}\right) \tag {13} +$$ + +where $i = 1,2,\ldots ,K;\sigma_{E}$ is set to 6. Note that we crop $P$ based on the predicted RBox, thus $\lambda$ is always the same. + +Multiplying the two distributions yields the target: + +$$ +h _ {t} = \frac {\beta h}{K} \arg \max (\mu \times \lambda) \tag {14} +$$ + +where $h_t$ is the regression target of height. Likewise, the width target $w_t$ is calculated along $x$ direction. + +The edge loss is then obtained as: + +$$ +\mathcal {L} _ {E} = \operatorname {s m o o t h} _ {L 1} \left(\left[ \begin{array}{l l} w & h \end{array} \right], \left[ \begin{array}{l l} w _ {t} & h _ {t} \end{array} \right]\right) \tag {15} +$$ + +Note that $\mathcal{L}_W$ and $\mathcal{L}_E$ merely refines the width and height of boxes, without involving the angle regression. + +# 3.5. Symmetry-aware Learning + +Proven effective in learning the rotation of objects [65], in this work, symmetry-aware learning is extended to Gaussian- + +based OOD. In the left part of Fig. 3, the training image $I$ is transformed (randomly selected from rotation, flip, and scale, see Table 7) to generate an augmented view as: + +$$ +I _ {\mathrm {a u g}} = \alpha I \tag {16} +$$ + +where + +$$ +\alpha = \left[ \begin{array}{c c} \cos \mathcal {R} & - \sin \mathcal {R} \\ \sin \mathcal {R} & \cos \mathcal {R} \end{array} \right] ^ {p _ {1}} \left[ \begin{array}{c c} 1 & 0 \\ 0 & - 1 \end{array} \right] ^ {p _ {2}} \left[ \begin{array}{c c} s & 0 \\ 0 & s \end{array} \right] ^ {p _ {3}} \tag {17} +$$ + +where $(p_1, p_2, p_3) = (1, 0, 0)$ when rotation is selected, $(0, 1, 0)$ when flip, $(0, 0, 1)$ when scale. $\mathcal{R}$ and $s$ are the random amount of rotation and scale, $s \in (0.5, 0.9)$ . + +By feeding both $I$ and $I_{\mathrm{aug}}$ into the network, we obtain two sets of output Gaussian distributions and angles: + +$$ +\left\{ \begin{array}{l} (\Sigma , \theta) = f _ {\mathrm {n n}} (I) \\ (\Sigma_ {\mathrm {a u g}}, \theta_ {\mathrm {a u g}}) = f _ {\mathrm {n n}} (I _ {\mathrm {a u g}}) \end{array} \right. \tag {18} +$$ + +The consistency loss is calculated between the two sets so that the network learns the size and rotation variation: + +$$ +\mathcal {L} _ {\mathrm {s s}} = L _ {\mathrm {G W D}} \left(\alpha \Sigma \alpha^ {\top}, \Sigma_ {\mathrm {a u g}}\right) + L _ {\mathrm {A N G}} (m \theta + \mathcal {R}, \theta_ {\mathrm {a u g}}) \tag {19} +$$ + +where $\mathcal{R}$ is the rotation angle and $m = 1$ when rotation is selected; $(m,\mathcal{R}) = (-1,0)$ when flip; $(m,\mathcal{R}) = (1,0)$ when scale. $L_{\mathrm{GWD}}(\cdot)$ is Gaussian Wasserstein Distance Loss [60] and $L_{\mathrm{ANG}}(\cdot)$ is defined as: + +$$ +L _ {\text {A N G}} \left(\theta_ {1}, \theta_ {2}\right) = \min _ {k \in Z} \left(s m o o t h _ {L 1} \left(\theta_ {1}, k \pi + \theta_ {2}\right)\right) \tag {20} +$$ + +where $\min (\cdot)$ regresses the prediction toward the closest target to circumvent the periodicity problem [65]. + +# 3.6. Copy-paste Augmentation + +Inspired by [9], we propose to crop the detected instances of step $k$ and paste them on the training image of step $k + 1$ . The maximum number of paste boxes is limited to 10 in each step. We simply use the bounding boxes of the cropped instances as the regression targets, and use Gaussian Wasserstein Distance Loss [60] to calculate $\mathcal{L}_{\mathrm{box}}$ . + +# 3.7. Overall Loss + +The overall loss $\mathcal{L}$ for Point2RBox-v2 can be expressed as: + +$$ +\mathcal {L} _ {\mathrm {c l s}} + w _ {\text {b o x}} \mathcal {L} _ {\text {b o x}} + w _ {\mathrm {O}} \mathcal {L} _ {\mathrm {O}} + w _ {\mathrm {W}} \mathcal {L} _ {\mathrm {W}} + w _ {\mathrm {E}} \mathcal {L} _ {\mathrm {E}} + w _ {\mathrm {s s}} \mathcal {L} _ {\mathrm {s s}} \tag {21} +$$ + +where $\mathcal{L}_{\mathrm{cls}}$ is the focal loss [25] for classification, $\mathcal{L}_{\mathrm{box}}$ regresses boxes/centers toward copy-paste/ground-truth labels, $w_{\mathrm{box}}$ is set to one by default, $(w_{\mathrm{O}}, w_{\mathrm{W}}, w_{\mathrm{E}}, w_{\mathrm{ss}})$ are set to (10, 5, 0.3, 1) based on our ablation studies (see Tables 3-5). + +# 4. Experiments + +Experiments are carried out on NVIDIA RTX4090 GPUs using PyTorch 2.2.0 [36] and the rotation detection tool kits: MMRotate 1.0.0 [73]. All the experiments follow the same hyper-parameters (learning rate, batch size, optimizer, etc.). + +
Methods*PL1BDBRGTFSVLVSHTCBCSTSBFRAHASPHCAP50
▼RBox-supervised OOD
RepPoints (2019) [62]✓86.781.141.662.076.256.375.790.780.885.363.366.659.167.633.768.45
RetinaNet (2017) [25]✓88.277.045.069.471.559.074.590.884.979.357.364.762.766.539.668.69
GWD (2021) [56]✓89.375.447.861.979.573.886.190.984.579.455.959.763.271.045.471.66
FCOS (2019) [42]✓89.176.950.163.279.879.887.190.480.884.659.766.365.871.341.772.44
S2A-Net (2022) [12]✓89.283.052.574.678.879.287.590.984.984.861.968.070.771.459.875.81
▼HBox-supervised OOD
Sun et al. (2021) [41]×51.538.716.136.829.819.223.483.950.680.018.950.225.628.725.538.60
BoxInst-RBox (2021) [43]2×68.440.833.132.346.955.456.679.566.882.141.252.852.865.030.053.59
H2RBox (2023) [59]✓88.573.540.856.977.565.477.990.983.285.355.362.952.463.643.367.82
EIE-Det (2024) [45]✓87.770.241.560.580.776.386.390.982.684.753.164.558.170.443.870.10
H2RBox-v2 (2023) [65]✓89.074.450.060.579.875.386.990.985.185.059.263.265.270.549.772.31
▼Point-supervised OOD
Point2Mask-RBox (2023) [21]2×4.023.13.81.315.11.03.319.01.029.10.09.57.421.17.19.72
P2BNet+H2RBox (2023) [3, 59]×24.735.97.127.93.312.117.517.50.834.06.349.611.627.218.819.63
P2BNet+H2RBox-v2 (2023) [3, 65]×11.044.814.915.436.816.727.812.11.831.23.450.612.636.712.521.87
P2RBox (2024) [2]†×87.865.715.060.773.071.778.981.544.581.241.239.345.557.541.259.04
PointOBB (2024) [33]×26.165.79.159.465.834.929.80.52.316.70.649.021.841.036.730.08
Point2RBox (2024) [66]✓62.964.314.435.028.238.933.325.22.244.53.448.125.945.022.634.07
Point2RBox+SK (2024) [66]†✓53.363.93.750.940.039.245.776.710.556.15.449.524.251.233.840.27
Point2RBox+SK (2024) [66]†×66.459.55.252.654.153.957.390.83.257.86.147.422.955.740.544.90
PointOBB-v2 (2025) [38]×64.527.81.936.258.847.253.490.562.245.312.141.78.143.732.041.68
PointOBB-v3 (2025) [68]✓30.939.413.522.761.27.043.162.459.847.32.745.116.855.211.441.29
PointOBB-v3 (2025) [68]×52.954.421.352.765.644.967.887.226.773.432.653.339.056.410.249.24
Point2RBox-v2 (ours)✓78.452.78.340.971.060.574.788.765.572.124.426.130.150.721.051.00
Point2RBox-v2 (ours)×88.072.68.046.279.676.386.989.179.782.926.245.345.866.346.362.61
*Comparison tracks: √ = End-to-end training and testing; × = Generating pseudo labels to train the FCOS detector (two-stage training).†Using additional priors. P2RBox: Pre-trained SAM model; Point2RBox+SK: One-shot sketches for each class.†PL: Plane, BD: Baseball diamond, BR: Bridge, GTF: Ground track field, SV: Small vehicle, LV: Large vehicle, SH: Ship, TC: Tennis court, BC: Basketball court, ST: Storage tank, SBF: Soccer-ball field, RA: Roundabout, HA: Harbor, SP: Swimming pool, HC: Helicopter.‡RBox: The minimum rectangle operation is performed on the output Mask to obtain the RBox.
+ +Table 1. Detection performance of each category on the DOTA-v1.0 and the mean ${\mathrm{{AP}}}_{50}$ of all categories. + +Average precision (AP) is adopted as the primary metric. All the models are configured upon ResNet50 [13] and trained with AdamW [31]. 1) Learning rate. Initialized at 5e-5, warm-up for 500 iterations, and divided by ten at each decay step. 2) Epochs. 72 for HRSC; 12 for the others. 3) Augmentation. Random rotation/flip for HRSC; random flip for the others. 4) Image size. Split into $1,024 \times 1,024$ with an overlap of 200 for DOTA/FAIR1M/STAR; scaled to $800 \times 800$ for others. 5) Multi-scale. All experiments evaluated without multi-scale technique [73]. 6) Datasets. Six remote sensing and one retail scene datasets, covering all datasets used by the main counterparts [2, 33, 66]: + +- DOTA [48]. DOTA-v1.0 has 2,806 aerial images annotated with 15 categories, while DOTA-v1.5/2.0 are the extended versions with more small objects and categories. +- DIOR [4]. It is an aerial image dataset re-annotated with RBoxes based on its original HBox version [18], with a high variation in object size and high intra-class diversity. +- HRSC [30]. It contains ship instances on the sea and inshore. The train/val/test set includes 436/181/444 images. +- FAIR1M [40]. It has more than 1 million instances and more than 40,000 images for fine-grained object recognition in remote sensing imagery, annotated with 37 cate + +gories. The results are evaluated on FAIR1M-1.0. + +- STAR [22]. It is extensive for scene graph generation, covering more than 210,000 objects with diverse spatial resolutions, classified into 48 fine-grained categories and precisely annotated with oriented bounding boxes. +- SKU110K [35]. It focuses on the detection of densely packed retail scenes with 110,712 objects in 11,762 images. The density reaches 86 instances per image. +- RSAR [70]. It is a remote sensing dataset based on Synthetic Aperture Radar (SAR) imagery with 6 categories. + +# 4.1. Main Results on DOTA-v1.0 + +Table 1 compares Point2RBox-v2 with the state-of-the-art methods, which can be categorized into two tracks: + +1) End-to-end training. These methods apply the trained weakly-supervised detector directly to the test set. Without relying on priors, our approach demonstrates an improvement of $16.93\%$ (51.00% vs. 34.07%) compared to Point2RBox. Even when compared to Point2RBox+SK, which incorporates additional data-side priors (i.e. one-shot examples for each class), our method still outperforms it by $10.73\%$ (51.00% vs. 40.27%). +2) Two-stage training. These methods generate RBox + +
Methods*DOTA-v1.0DOTA-v1.5DOTA-v2.0DIORHRSCFAIR1MSTARSKU110KRSAR
▼RBox-supervised OOD
RetinaNet (2017) [25]✓68.6960.5747.0054.9684.4937.6721.8078.5057.67
GWD (2021) [56]✓71.6663.2748.8757.6086.6739.1125.3079.1657.80
FCOS (2019) [42]✓72.4464.5351.7759.8388.9941.2528.1080.0966.66
S2A-Net (2022) [12]✓75.8166.5352.3961.4190.1042.4427.3080.3666.47
▼HBox-supervised OOD
Sun et al. (2021) [41]×38.60--------
KCR (2023) [74]✓----79.10----
H2RBox (2023) [59]✓70.0561.7048.6857.807.0335.9417.2057.1549.92
H2RBox-v2 (2023) [65]✓72.3164.7650.3357.6489.6642.2727.3070.7065.16
AFWS (2024) [32]✓72.5565.9251.7359.07-41.80---
▼Point-supervised OOD
P2RBox (2024) [2]†×59.04--------
PointSAM (2025) [27]†×---46.20-----
PointOBB (2024) [33]×30.0810.665.5337.31-11.199.19-13.80
Point2RBox+SK (2024) [66]†✓40.2730.5123.4327.3479.4020.037.863.4127.81
PointOBB-v2 (2025) [38]×41.6830.5920.6439.56-13.369.0056.6318.99
PointOBB-v3 (2025) [68]✓41.2031.2522.8237.60-11.4211.31-15.84
PointOBB-v3 (2025) [68]×49.2433.7923.5240.18-18.3512.85-22.60
Point2RBox-v2 (ours)✓51.0039.4527.1134.7082.6725.727.8064.0028.60
Point2RBox-v2 (ours)×62.6154.0638.7944.4586.1534.7114.2065.6430.90
*Comparison tracks: √= End-to-end training and testing; × = Generating pseudo labels to train the FCOS detector (two-stage training).†Using additional priors. P2RBox/PointSAM: Pre-trained SAM model; Point2RBox+SK: One-shot sketches for each class.
+ +Table 2. Accuracy $\left( {\mathrm{{AP}}}_{50}\right)$ comparisons on the DOTA-v1.0/1.5/2.0, DIOR, HRSC, FAIR1M, STAR, SKU110K, and RSAR datasets. + +
\(w_{O}\)DOTAHRSC
E2EFCOSE2EFCOS
348.7661.6281.8584.36
549.8162.4482.4685.76
1051.0062.6182.6786.15
3045.8857.8381.5685.61
+ +Table 3. Ablation with the weight of ${\mathcal{L}}_{\mathrm{O}}$ . + +
\( {w}_{\mathrm{W}} \)DOTAHRSC
E2EFCOSE2EFCOS
350.8556.7878.4283.49
551.0062.6182.6786.15
1049.1560.5430.3735.13
3042.8452.5323.8925.91
+ +Table 4. Ablation with the weight of ${\mathcal{L}}_{\mathrm{W}}$ . + +
wEDOTAHRSC
E2EFCOSE2EFCOS
0.148.7557.6234.7139.45
0.351.0062.6182.6786.15
0.551.3662.6376.8585.22
1.049.0560.6356.5959.59
+ +Table 5. Ablation with the weight of ${\mathcal{L}}_{\mathrm{E}}$ . + +
wssDOTAHRSC
E2EFCOSE2EFCOS
0.149.2859.6673.6678.92
1.051.0062.6182.6786.15
3.049.1559.201.301.65
+ +Table 6. Ablation with the weight of ${\mathcal{L}}_{\mathrm{{ss}}}$ . + +
R / F / SDOTAHRSCR / F / SDOTAHRSC
90% / 10% / 0%60.4285.4680% / 20% / 0%59.4684.73
75% / 0% / 25%60.7986.2260% / 15% / 25%62.3884.21
68% / 7% / 25%62.6186.1538% / 37% / 25%45.878.56
45% / 5% / 50%60.5585.3440% / 10% / 50%60.4910.74
+ +Table 7. Ablation with the proportion of augmented views in self-supervision. + +labels on train/val sets, with which the FCOS detector is trained. In this two-stage mode, Point2RBox-v2 achieves an accuracy of $62.61\%$ , considerably surpassing PointOBB series. Remarkably, it even outperforms the SAM-powered method P2RBox by $3.57\%$ (62.61% vs. 59.04%). + +Class-wise analysis. The FCOS detector trained with labels generated by Point2RBox-v2 achieves accuracy nearly equivalent to RBox-supervised FCOS across six high-density categories: SH (86.9% vs. 87.1%), SV (79.6% vs. 79.8%), LV (76.3% vs. 79.8%), PL (88.0% vs. 89.1%), ST (82.9% vs. 84.6%), and TC (89.1% vs. 90.4%). Interestingly, these six high-density categories account for 88% of DOTA instances. By annotating these categories with points and generating RBoxes using Point2RBox-v2 while labeling the other sparse categories with RBoxes, we can significantly reduce annotation labor without sacrificing much accuracy, + +highlighting the valuable role our method can play. + +# 4.2. Results on More Datasets + +The results are displayed in Table 2. On more challenging DOTA-v1.5/2.0, Point2RBox-v2 presents a similar trend, $23.47\% / 18.15\%$ higher than PointOBB-v2 in the pseudogeneration track. On the ship detection dataset HRSC, the gap between Point2RBox-v2 and RBox-supervised FCOS is only $2.84\%$ (86.15% vs. 88.99%). DIOR is relatively sparse, leading to less improvement with our methods—lower than PointSAM (44.45% vs. 46.20%) but still higher than methods that do not use SAM. Our method also provides competitive performance on fine-grained datasets FAIR1M and STAR. In addition to remote sensing scenarios, we carry out experiments on SKU110K for densely packed retail scenes. Existing point-supervised methods struggle in this + +![](images/916035995ff8381e486e21560479f131cec1896e656181e344006d143384d8eb.jpg) + +![](images/6649cc04a16fbfe2ca5554605e2cd6478862a2ddaf500faad08f94dc6452c6d2.jpg) +Figure 5. Qualitative analysis on failed cases and overlap cases. + +![](images/97bfa86eff7ab22c510e4a551eda61f101afbe60d1c2069a12b0ed94dbb6b98f.jpg) + +![](images/93b828120c97dda984b2f8b9fe4d63c951428c0cc587a558e49af2fc60c50db1.jpg) + +![](images/3853fb30847fbd0e2573b0ef5cd80fb92b4d1ff11649132c646ac14645267b31.jpg) + +
ModulesDOTAHRSC
\( \mathcal{L}_{\mathrm{O}} \)\( \mathcal{L}_{\mathrm{W}} \)\( \mathcal{L}_{\mathrm{ss}} \)\( \mathcal{L}_{\mathrm{E}} \)\( CP \)E2EFCOSE2EFCOS
✓0.000.000.000.00
✓✓41.5452.9817.9619.64
✓✓✓46.6454.2618.1022.13
✓✓✓✓49.5561.8878.7983.79
✓✓✓✓48.5859.5620.3524.76
✓✓✓✓38.9448.4411.6414.93
✓✓✓✓47.0855.0519.5821.78
✓✓✓✓✓51.0062.6182.6786.15
+ +Table 8. Ablation with incremental addition of modules. + +
16K=2432|| 1.2β=1.62.0
50.8751.0048.08|| 48.1451.0051.33
+ +Table 9. Ablation with $K$ and $\beta$ in edge loss on DOTA (E2E). + +
σPoint2RBoxPointOBB-v2Point2RBox-v2
DOTAHRSCDOTAHRSCDOTAHRSC
0%40.2779.4044.85-62.6186.15
10%39.6078.8142.30-61.5885.76
30%38.4278.2838.46-60.3185.71
+ +Table 10. Ablation with the inaccuracy in point annotations. + +case, whereas Point2RBox-v2 achieves performance on par with HBox-supervised H2RBox (65.64% vs. 57.15%). + +# 4.3. Ablation Studies + +Tables 3-10 display the ablation studies on DOTA-v1.0 and HRSC. "E2E" denotes end-to-end training; "FCOS" denotes two-stage training (i.e. generating pseudo labels to train FCOS). The final values adopted are highlighted in gray. + +Weight of each loss. Tables 3-5 determine the weights of the proposed losses. Based on these experiments, the weights $(w_{\mathrm{O}}, w_{\mathrm{W}}, w_{\mathrm{E}}, w_{\mathrm{ss}})$ are set to (10, 5, 0.3, 1). + +Proportion of augmented views. Table 7 studies the proportion between rotation, flip, and scale. The results are reported with two-stage training (FCOS). Based on the results, the proportion is set to $68\%$ , $7\%$ , and $25\%$ . + +Incremental addition of modules. Table 8 demonstrates the constraints from Gaussian and Voronoi achieve an accuracy of $52.98\%$ on DOTA. Adding consistency loss and edge loss further boosts it to $54.26\%$ and $61.88\%$ , respectively, whereas the improvement from copy-paste is $0.73\%$ . We also demonstrate the impact of omitting each core loss. + +Edge loss parameters. We set $K = 24$ and $\beta = 1.6$ as they are observed to discern the correct edges during code development. Table 9 provides a more precise ablation. + +Annotation inaccuracy. We offset the annotated points by a noise from the uniform distribution $[- \sigma H, + \sigma H]$ , where $H$ is the height of objects. Table 10 shows that the $\mathrm{AP}_{50}$ of Point2RBox-v2 decreases by less than $3\%$ when noise is added to point annotations, demonstrating the robustness of the proposed learning mechanisms. + +# 4.4. More Discussions + +The qualitative analysis on the failed/overlap cases is shown in Fig. 5. 1) Failed cases. Although our method performs well overall, it struggles with certain categories that are sparse and not constrained by other objects. 2) Overlap cases. Minimizing overlap as a soft constraint during training does not entirely eliminate overlap. Once trained, the model remains robust to some overlap during inference. + +# 5. Conclusion + +This paper introduces Point2RBox-v2, a point-supervised oriented object detector that effectively leverages the arrangement and layout of instances. We propose the integration of Gaussian overlay and Voronoi tessellation to constrain the size and rotation of instances based on their spatial relationships. Additionally, by incorporating self-supervised consistency loss, edge loss, and copy-paste augmentation, the accuracy of the model is further enhanced. + +Experiments yield the following observations: 1) The integration of Gaussian and Voronoi concepts effectively harnesses the spatial layout of objects, significantly enhancing point-supervised OOD. 2) Point2RBox-v2 demonstrates exceptional performance in densely packed scenes (see Fig. 2), where existing methods struggle. 3) Our method does not require priors (i.e. pre-trained SAM or one-shot examples) and is applicable to both end-to-end and pseudo-generation modes. 4) It advances the state of the art by a large amount, achieving $62.61\%$ , $86.15\%$ , and $34.71\%$ on the DOTA-v1.0, HRSC, and FAIR1M datasets, respectively. + +Limitations. The gap between Point2RBox-v2 and the RBox-supervised OOD is still huge in terms of sparse categories (i.e. BR/SBF) since little constraint can be obtained from the layout between them when the objects are sparse. + +# References + +[1] Franz Aurenhammer. Voronoi diagrams—a survey of a fundamental geometric data structure. ACM Computing Surveys, 23(3):345-405, 1991. 2, 4 +[2] Guangming Cao, Xuehui Yu, Wenwen Yu, Xumeng Han, Xue Yang, Guorong Li, Jianbin Jiao, and Zhenjun Han. P2rbox: Point prompt oriented object detection with SAM. arXiv preprint arXiv:2311.13128, 2024. 1, 2, 6, 7 +[3] Pengfei Chen, Xuehui Yu, Xumeng Han, Najmul Hassan, Kai Wang, Jiachen Li, Jian Zhao, Humphrey Shi, Zhenjun Han, and Qixiang Ye. Point-to-box network for accurate object detection via single point supervision. In European Conference on Computer Vision, 2022. 3, 6 +[4] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 2022. 6 +[5] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2849-2858, 2019. 2 +[6] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2021. 1 +[7] Kun Fu, Zhonghan Chang, Yue Zhang, Guangluan Xu, Keshu Zhang, and Xian Sun. Rotation-aware and multi-scale convolutional neural network for object detection in remote sensing images. ISPRS Journal of Photogrammetry and Remote Sensing, 161:294-308, 2020. 1 +[8] Shuyong Gao, Wei Zhang, Yan Wang, Qianyu Guo, Chenglong Zhang, Yangji He, and Wenqiang Zhang. Weakly-supervised salient object detection using point supervision. In AAAI Conference on Artificial Intelligence, pages 670-678, 2022. 3 +[9] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D. Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 5 +[10] Eran Goldman, Roei Herzig, Aviv Eisenschat, Jacob Goldberger, and Tal Hassner. Precise detection in densely packed scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5236, 2019. 1 +[11] Jiaming Han, Jian Ding, Nan Xue, and Gui-Song Xia. Redet: A rotation-equivariant detector for aerial object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2785-2794, 2021. 2 +[12] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2, 6, 7 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Con- + +ference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3, 6 +[14] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In IEEE International Conference on Computer Vision, 2017. 5 +[15] Liping Hou, Ke Lu, Xue Yang, Yuqiu Li, and Jian Xue. Grep: Gaussian representation for arbitrary-oriented object detection. Remote Sensing, 15(3):757, 2023. 2 +[16] Javed Iqbal, Muhammad Akhtar Munir, Arif Mahmood, Afsheen Rafaqat Ali, and Mohsen Ali. Leveraging orientation for weakly supervised object detection with application to firearm localization. Neurocomputing, 440:310-320, 2021. 3 +[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3 +[18] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing, 159:296-307, 2020. 6 +[19] Qingyun Li, Yushi Chen, Xinya Shu, Dong Chen, Xin He, Yi Yu, and Xue Yang. A simple aerial detection baseline of multimodal language models. arXiv preprint arXiv:2501.09720, 2025. 2 +[20] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented leptopoints for aerial object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1829-1838, 2022. 2 +[21] Wentong Li, Yuqian Yuan, Song Wang, Jianke Zhu, Jianshu Li, Jian Liu, and Lei Zhang. Point2mask: Point-supervised panoptic segmentation via optimal transport. In IEEE International Conference on Computer Vision, 2023. 2, 3, 6 +[22] Yansheng Li, Linlin Wang, Tingzhu Wang, Xue Yang, Junwei Luo, Qi Wang, Youming Deng, Wenbin Wang, Xian Sun, Haifeng Li, Bo Dang, Yongjun Zhang, Yi Yu, and Junchi Yan. Star: A first-ever dataset and a large-scale benchmark for scene graph generation in large-size satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1832-1849, 2025. 6 +[23] Minghui Liao, Zhen Zhu, Baoguang Shi, Gui-Song Xia, and Xiang Bai. Rotation-sensitive regression for oriented scene text detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5909-5918, 2018. 1 +[24] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 3 +[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 2, 5, 6, 7 +[26] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International Journal of Computer Vision, 128(2):261-318, 2020. 2 + +[27] Nanqing Liu, Xun Xu, Yongyi Su, Haojie Zhang, and Heng-Chao Li. Pointsam: Pointly-supervised segment anything model for remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 63:1-15, 2025. 1, 2, 7 +[28] Xuebo Liu, Ding Liang, Shi Yan, Dagui Chen, Yu Qiao, and Junjie Yan. Fots: Fast oriented text spotting with a unified network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5676-5685, 2018. 1 +[29] Yuekai Liu, Hongli Gao, Liang Guo, Aoping Qin, Canyu Cai, and Zhichao You. A data-flow oriented deep ensemble learning method for real-time surface defect inspection. IEEE Transactions on Instrumentation and Measurement, 69(7): 4681-4691, 2020. 1 +[30] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International Conference on Pattern Recognition Applications and Methods, pages 324-331, 2017. 1, 6 +[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6 +[32] Junyan Lu, Qinglei Hu, Ruifei Zhu, Yali Wei, and Tie Li. Afws: Angle-free weakly-supervised rotating object detection for remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 7 +[33] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. Pointobb: Learning oriented object detection via single point supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 6, 7 +[34] Jeffri Murragarra-Llerena, Lucas N Kirsten, Luis Felipe Zeni, and Claudio R Jung. Probabilistic intersection-over-union for training and evaluation of oriented object detectors. IEEE Transactions on Image Processing, 2024. 2 +[35] Xingjia Pan, Yuqiang Ren, Kekai Sheng, Weiming Dong, Haolei Yuan, Xiaowei Guo, Chongyang Ma, and Changsheng Xu. Dynamic refinement network for oriented and densely packed object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11207-11216, 2020. 1, 6 +[36] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035, 2019. 5 +[37] Wen Qian, Xue Yang, Silong Peng, Junchi Yan, and Yue Guo. Learning modulated loss for rotated object detection. In AAAI Conference on Artificial Intelligence, pages 2458-2466, 2021. 2 +[38] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. Pointobb-v2: Towards simpler, faster, and stronger single point supervised oriented object detection. In International Conference on Learning Representations, 2025. 1, 2, 3, 6, 7 +[39] Xavier Soria, Yachuan Li, Mohammad Rouhani, and Angel D. Sappa. Tiny and efficient model for the edge detection + +generalization. In IEEE/CVF International Conference on Computer Vision Workshops, pages 1364-1373, 2023. 5 +[40] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, Martin Weinmann, Stefan Hinz, Cheng Wang, and Kun Fu. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 6 +[41] Yongqing Sun, Jie Ran, Feng Yang, Chenqiang Gao, Takayuki Kurozumi, Hideaki Kimata, and Ziqi Ye. Oriented object detection for remote sensing images based on weakly supervised learning. In IEEE International Conference on Multimedia & Expo Workshops, pages 1-6, 2021. 3, 6, 7 +[42] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In IEEE/CVF International Conference on Computer Vision, pages 9626-9635, 2019. 2, 6, 7 +[43] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 6 +[44] L. Vincent and P. Soille. Watersheds in digital spaces: an efficient algorithm based on immersion simulations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(6):583-598, 1991. 4 +[45] Linfei Wang, Yibing Zhan, Xu Lin, Baosheng Yu, Liang Ding, Jianqing Zhu, and Dapeng Tao. Explicit and implicit box equivariance learning for weakly-supervised rotated object detection. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 3, 6 +[46] Long Wen, Yu Cheng, Yi Fang, and Xinyu Li. A comprehensive survey of oriented object detection in remote sensing images. Expert Systems with Applications, page 119960, 2023. 2 +[47] Hongjin Wu, Ruoshan Lei, and Yibing Peng. Pcbnet: A lightweight convolutional neural network for defect inspection in surface mount technology. IEEE Transactions on Instrumentation and Measurement, 71:1-14, 2022. 1 +[48] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2018. 1, 6 +[49] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In IEEE/CVF International Conference on Computer Vision, pages 3520-3529, 2021. 2 +[50] Xue Yang and Junchi Yan. Arbitrary-oriented object detection with circular smooth label. In European Conference on Computer Vision, pages 677-694, 2020. 2 +[51] Xue Yang and Junchi Yan. On the arbitrary-oriented object detection: Classification based approaches revisited. International Journal of Computer Vision, 130(5):1340-1365, 2022. 1, 2 +[52] Xue Yang, Hao Sun, Kun Fu, Jirui Yang, Xian Sun, Menglong Yan, and Zhi Guo. Automatic ship detection in remote + +sensing images from google earth of complex scenes based on multiscale rotation dense feature pyramid networks. Remote sensing, 10(1):132, 2018. 1, 2 +[53] Xue Yang, Jirui Yang, Junchi Yan, Yue Zhang, Tengfei Zhang, Zhi Guo, Xian Sun, and Kun Fu. Scdet: Towards more robust detection for small, cluttered and rotated objects. In IEEE/CVF International Conference on Computer Vision, pages 8231-8240, 2019. 2 +[54] Xue Yang, Liping Hou, Yue Zhou, Wentao Wang, and Junchi Yan. Dense label encoding for boundary discontinuity free rotation detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15814-15824, 2021. 2 +[55] Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3det: Refined single-stage detector with feature refinement for rotating object. In AAAI Conference on Artificial Intelligence, pages 3163-3171, 2021. 2 +[56] Xue Yang, Junchi Yan, Ming Qi, Wentao Wang, Xiaopeng Zhang, and Tian Qi. Rethinking rotated object detection with gaussian Wasserstein distance loss. In 38th International Conference on Machine Learning, pages 11830-11841, 2021. 2, 6, 7 +[57] Xue Yang, Xiaojiang Yang, Jirui Yang, Qi Ming, Wentao Wang, Qi Tian, and Junchi Yan. Learning high-precision bounding box for rotated object detection via kullback-leibler divergence. In Advances in Neural Information Processing Systems, pages 18381-18394, 2021. 2 +[58] Xue Yang, Junchi Yan, Wenlong Liao, Xiaokang Yang, Jin Tang, and Tao He. Scdet++: Detecting small, cluttered and rotated objects via instance-level feature denoising and rotation loss smoothing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2384-2399, 2022. 2 +[59] Xue Yang, Gefan Zhang, Wentong Li, Xuehui Wang, Yue Zhou, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. International Conference on Learning Representations, 2023. 1, 3, 6, 7 +[60] Xue Yang, Gefan Zhang, Xiaojiang Yang, Yue Zhou, Wentao Wang, Jin Tang, Tao He, and Junchi Yan. Detecting rotated objects as gaussian distributions and its 3-d generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4335-4354, 2023. 2, 3, 4, 5 +[61] Xue Yang, Yue Zhou, Gefan Zhang, Jirui Yang, Wentao Wang, Junchi Yan, Xiaopeng Zhang, and Qi Tian. The kfiou loss for rotated object detection. In International Conference on Learning Representations, 2023. 2 +[62] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Reppoints: Point set representation for object detection. In IEEE/CVF International Conference on Computer Vision, pages 9656-9665, 2019. 2, 6 +[63] Yi Yu and Feipeng Da. Phase-shifting coder: Predicting accurate orientation in oriented object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[64] Yi Yu and Feipeng Da. On boundary discontinuity in angle regression based arbitrary oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-14, 2024. 3 + +[65] Yi Yu, Xue Yang, Qingyun Li, Yue Zhou, Feipeng Da, and Junchi Yan. H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. In Advances in Neural Information Processing Systems, 2023. 1, 3, 5, 6, 7 +[66] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2rbox: Combine knowledge from synthetic visual patterns for end-to-end oriented object detection with single point supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6, 7 +[67] Yi Yu, Xue Yang, Yansheng Li, Zhenjun Han, Feipeng Da, and Junchi Yan. Wholly-wood: Wholly leveraging diversified-quality labels for weakly-supervised oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025. 3 +[68] Peiyuan Zhang, Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Yue Zhou, Xiaosong Jia, Xudong Lu, Jingdong Chen, Xiang Li, et al. Pointobb-v3: Expanding performance boundaries of single point-supervised oriented object detection. arXiv preprint arXiv:2501.13898, 2025. 1, 2, 3, 6, 7 +[69] Shun Zhang, Jihui Long, Yaohui Xu, and Shaohui Mei. Pmho: Point-supervised oriented object detection based on segmentation-driven proposal generation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2 +[70] Xin Zhang, Xue Yang, Yuxuan Li, Jian Yang, Ming-Ming Cheng, and Xiang Li. Rsar: Restricted state angle resolver and rotated sar benchmark. arXiv preprint arXiv:2501.04440, 2025.6 +[71] Zhong-Qiu Zhao, Peng Zheng, Shou-Tao Xu, and Xindong Wu. Object detection with deep learning: A review. IEEE Transactions on Neural Networks and Learning Systems, 30 (11):3212-3232, 2019. 2 +[72] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: An efficient and accurate scene text detector. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2017. 1 +[73] Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, et al. MMrotate: A rotated object detection benchmark using pytorch. In 30th ACM International Conference on Multimedia, pages 7331-7334, 2022. 5, 6 +[74] Tianyu Zhu, Bryce Ferenczi, Pulak Purkait, Tom Drummond, Hamid Rezatofighi, and Anton van den Hengel. Knowledge combination to learn rotated detection without rotated annotation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 7 \ No newline at end of file diff --git a/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/images.zip b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..79e137ebacd905eae6b9ced427e77e3a149d8f7d --- /dev/null +++ b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:044c1a1e85aa4f98f98755d6dc37ce3641a57537b69c8efe90309402638cbc55 +size 1367659 diff --git a/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/layout.json b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..26b84fca1274e24b9747d584dcb5c23ea59ca9f6 --- /dev/null +++ b/2025/Point2RBox-v2_ Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances/layout.json @@ -0,0 +1,11972 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 99, + 529, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 99, + 529, + 136 + ], + "spans": [ + { + "bbox": [ + 83, + 99, + 529, + 136 + ], + "type": "text", + "content": "Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "spans": [ + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "text", + "content": "Yi Yu" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "text", + "content": ", Botao Ren" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "text", + "content": ", Peiyuan Zhang" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "text", + "content": ", Mingxin Liu" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "text", + "content": ", Junwei Luo" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "text", + "content": ", Shaofeng Zhang" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "text", + "content": ", Feipeng Da" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "text", + "content": ", Junchi Yan" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "text", + "content": ", Xue Yang" + }, + { + "bbox": [ + 83, + 151, + 527, + 180 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 180, + 486, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 180, + 486, + 209 + ], + "spans": [ + { + "bbox": [ + 126, + 180, + 486, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 180, + 486, + 209 + ], + "type": "text", + "content": "Southeast University " + }, + { + "bbox": [ + 126, + 180, + 486, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 180, + 486, + 209 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 126, + 180, + 486, + 209 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 126, + 180, + 486, + 209 + ], + "type": "text", + "content": "Wuhan University " + }, + { + "bbox": [ + 126, + 180, + 486, + 209 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 126, + 180, + 486, + 209 + ], + "type": "text", + "content": "Department of Automation & CSE & SAI, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 210, + 546, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 210, + 546, + 222 + ], + "spans": [ + { + "bbox": [ + 65, + 210, + 546, + 222 + ], + "type": "text", + "content": "{yuyi,dafp}@seu.edu.cn rbt22@mails.tsinghua.edu.cn {peiyuanzhangwhu,luojunwei}@whu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 222, + 484, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 222, + 484, + 233 + ], + "spans": [ + { + "bbox": [ + 127, + 222, + 484, + 233 + ], + "type": "text", + "content": "{liumingxin,zhangshaofeng,yanjunchi, yangxue-2019-sjtu}@sjtu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 187, + 233, + 422, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 233, + 422, + 243 + ], + "spans": [ + { + "bbox": [ + 187, + 233, + 422, + 243 + ], + "type": "text", + "content": "https://github.com/VisionXLab/point2rbox-v2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 265, + 200, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 265, + 200, + 278 + ], + "spans": [ + { + "bbox": [ + 151, + 265, + 200, + 278 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 291, + 297, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 291, + 297, + 531 + ], + "spans": [ + { + "bbox": [ + 54, + 291, + 297, + 531 + ], + "type": "text", + "content": "With the rapidly increasing demand for oriented object detection (OOD), recent research involving weakly-supervised detectors for learning OOD from point annotations has gained great attention. In this paper, we rethink this challenging task setting with the layout among instances and present Point2RBox-v2. At the core are three principles: 1) Gaussian overlap loss. It learns an upper bound for each instance by treating objects as 2D Gaussian distributions and minimizing their overlap. 2) Voronoi watershed loss. It learns a lower bound for each instance through watershed on Voronoi tessellation. 3) Consistency loss. It learns the size/rotation variation between two output sets with respect to an input image and its augmented view. Supplemented by a few devised techniques, e.g. edge loss and copy-paste, the detector is further enhanced. To our best knowledge, Point2RBox-v2 is the first approach to explore the spatial layout among instances for learning point-supervised OOD. Our solution is elegant and lightweight, yet it is expected to give a competitive performance especially in densely packed scenes: " + }, + { + "bbox": [ + 54, + 291, + 297, + 531 + ], + "type": "inline_equation", + "content": "62.61\\% /86.15\\% /34.71\\%" + }, + { + "bbox": [ + 54, + 291, + 297, + 531 + ], + "type": "text", + "content": " on DOTA/HRSC/FAIR1M." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 135, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 135, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 135, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 565, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 625 + ], + "type": "text", + "content": "Emerged as an essential task in computer vision, oriented object detection (OOD) has become a prominent demand in autonomous driving [6], aerial images [7, 30, 48, 51, 52], scene text [23, 28, 72], retail scenes [10, 35], industrial inspection [29, 47], and more." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 625, + 296, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 296, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 296, + 650 + ], + "type": "text", + "content": "Manual annotations are essential to teach the detector new concepts of visual objects. Early research is often super" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 265, + 554, + 395 + ], + "blocks": [ + { + "bbox": [ + 316, + 265, + 554, + 395 + ], + "lines": [ + { + "bbox": [ + 316, + 265, + 554, + 395 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 554, + 395 + ], + "type": "image", + "image_path": "312565b9515f1af73282525218512fdb213bafa5407c82fb3751e9830a45ac65.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 397, + 555, + 420 + ], + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 420 + ], + "type": "text", + "content": "Figure 1. Related methods, their principles for knowledge mining, whether using additional priors, and performance on DOTA-v1.0." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "text", + "content": "vised by rotated bounding boxes (RBoxes), where annotations align with the desired output. While this approach has demonstrated promising results, RBox labels are considerably expensive. The cost of annotating each RBox is approximately " + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "inline_equation", + "content": "36.5\\%" + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "text", + "content": " higher than a horizontal bounding box (HBox) and " + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "inline_equation", + "content": "104.8\\%" + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "text", + "content": " higher than a point [66]. This highlights the importance of utilizing coarser labels for weakly-supervised oriented detection. Recent advancements in HBox-supervised OOD, particularly with H2RBox [59] and H2RBox-v2 [65], have shown promise in bridging the gap between HBox-and RBox-supervised methods, thereby lessening the dependence on labor-intensive RBox labeling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Following HBox-supervised methods, the more challenging point-supervised OOD opens up a new avenue. Several innovative approaches have emerged in the past year (see Fig. 1): 1) Point-prompt OOD. P2RBox [2], PMHO [69], and PointSAM [27] employ the zero-shot ability of SAM [17]. 2) Pseudo generation. PointOBB [33] and its v2/v3 version [38, 68] use multiple instance learning and class probability map for RBox generation. 3) Knowledge combination. Point2RBox [66] learns from one-shot examples. However, existing methods overlook the spatial relationships between objects. This brings us to the motivation of this paper." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 655, + 297, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 297, + 713 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding author. The work was in part supported by National Natural Science Foundation of China (62306069, 62222607, 72342023), Major Science and Technology Project of Jiangsu Province (BG2024003), and China Postdoctoral Science Foundation (2023M740602). Yi Yu is also supported by Jiangsu Funding Program for Excellent Postdoctoral Talent (2023ZB616)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19283" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 71, + 164, + 163 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 164, + 163 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 164, + 163 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 164, + 163 + ], + "type": "image", + "image_path": "b63c1ba009eff7f950c198770e57cd82ae19f8188beced4bc1bfb304c4dd30bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 165, + 72, + 258, + 163 + ], + "blocks": [ + { + "bbox": [ + 165, + 72, + 258, + 163 + ], + "lines": [ + { + "bbox": [ + 165, + 72, + 258, + 163 + ], + "spans": [ + { + "bbox": [ + 165, + 72, + 258, + 163 + ], + "type": "image", + "image_path": "bf2e91e935cbc8d25a72d01b0a5f5748c83cb7adeb106676947de0a84f178c68.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 260, + 72, + 351, + 163 + ], + "blocks": [ + { + "bbox": [ + 260, + 72, + 351, + 163 + ], + "lines": [ + { + "bbox": [ + 260, + 72, + 351, + 163 + ], + "spans": [ + { + "bbox": [ + 260, + 72, + 351, + 163 + ], + "type": "image", + "image_path": "b60dab802c0e0688e36305437db8d85bc1053c441c799691fa9caa51e807c684.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 353, + 72, + 444, + 163 + ], + "blocks": [ + { + "bbox": [ + 353, + 72, + 444, + 163 + ], + "lines": [ + { + "bbox": [ + 353, + 72, + 444, + 163 + ], + "spans": [ + { + "bbox": [ + 353, + 72, + 444, + 163 + ], + "type": "image", + "image_path": "172107eaaa86ce0c6fe6de5bdeb54f7a291ca8116068561a1bb8984311d12b73.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 446, + 72, + 538, + 163 + ], + "blocks": [ + { + "bbox": [ + 446, + 72, + 538, + 163 + ], + "lines": [ + { + "bbox": [ + 446, + 72, + 538, + 163 + ], + "spans": [ + { + "bbox": [ + 446, + 72, + 538, + 163 + ], + "type": "image", + "image_path": "a28d2a33aaf6741c75ba6fd7d0f75bdf78427d22b2ec892c5049d5bbcb093f88.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 72, + 164, + 164, + 256 + ], + "blocks": [ + { + "bbox": [ + 72, + 164, + 164, + 256 + ], + "lines": [ + { + "bbox": [ + 72, + 164, + 164, + 256 + ], + "spans": [ + { + "bbox": [ + 72, + 164, + 164, + 256 + ], + "type": "image", + "image_path": "776a7aff52f3b88b79e08e9a00a9ffe80ae4a8d915e0b9b268d7990201fae575.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 165, + 164, + 257, + 256 + ], + "blocks": [ + { + "bbox": [ + 165, + 164, + 257, + 256 + ], + "lines": [ + { + "bbox": [ + 165, + 164, + 257, + 256 + ], + "spans": [ + { + "bbox": [ + 165, + 164, + 257, + 256 + ], + "type": "image", + "image_path": "ef2e656aa692901b0e17f433fb4c56053618fc8bbeb2ed5d01caa3e3515f6f9b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 260, + 164, + 351, + 256 + ], + "blocks": [ + { + "bbox": [ + 260, + 164, + 351, + 256 + ], + "lines": [ + { + "bbox": [ + 260, + 164, + 351, + 256 + ], + "spans": [ + { + "bbox": [ + 260, + 164, + 351, + 256 + ], + "type": "image", + "image_path": "9d72e47dbcea8085f9d67829db6eeca8a209d158d6523e0fce9346adbe4c6bb8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 353, + 164, + 444, + 256 + ], + "blocks": [ + { + "bbox": [ + 353, + 164, + 444, + 256 + ], + "lines": [ + { + "bbox": [ + 353, + 164, + 444, + 256 + ], + "spans": [ + { + "bbox": [ + 353, + 164, + 444, + 256 + ], + "type": "image", + "image_path": "da7da2d16ef61360469eb00e29d5c9501b0b9206f87873470b38813b913a8066.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 446, + 164, + 538, + 256 + ], + "blocks": [ + { + "bbox": [ + 446, + 164, + 538, + 256 + ], + "lines": [ + { + "bbox": [ + 446, + 164, + 538, + 256 + ], + "spans": [ + { + "bbox": [ + 446, + 164, + 538, + 256 + ], + "type": "image", + "image_path": "a627dc09b71c35b05b41399cff5a9b067e79b77869fbba346af16badad417758.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 72, + 258, + 164, + 350 + ], + "blocks": [ + { + "bbox": [ + 72, + 258, + 164, + 350 + ], + "lines": [ + { + "bbox": [ + 72, + 258, + 164, + 350 + ], + "spans": [ + { + "bbox": [ + 72, + 258, + 164, + 350 + ], + "type": "image", + "image_path": "300ce1cd207b24d07a541bd9e4577ebacf925f091e39392bfdff2eb76cd3a151.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 352, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 352, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 555, + 376 + ], + "type": "text", + "content": "Figure 2. Visual comparisons with state-of-the-art methods including: Point2Mask (2023) [21], PointOBB (2024) [33], PointOBB-v2 (2025) [38], PointOBB-v3 (2025) [68], and Point2RBox (2024) [66]. The boxes detected by our method (last row) wrap the objects more tightly." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 165, + 258, + 257, + 349 + ], + "blocks": [ + { + "bbox": [ + 165, + 258, + 257, + 349 + ], + "lines": [ + { + "bbox": [ + 165, + 258, + 257, + 349 + ], + "spans": [ + { + "bbox": [ + 165, + 258, + 257, + 349 + ], + "type": "image", + "image_path": "a465eef682ae40c9eddd692767959461c348b5b7d30a7bec551e1c65cd45edca.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 260, + 258, + 351, + 349 + ], + "blocks": [ + { + "bbox": [ + 260, + 258, + 351, + 349 + ], + "lines": [ + { + "bbox": [ + 260, + 258, + 351, + 349 + ], + "spans": [ + { + "bbox": [ + 260, + 258, + 351, + 349 + ], + "type": "image", + "image_path": "1f315e80353531f669466f8a0b9b01944960bcacdae9df85b38af33ab04a8483.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 353, + 258, + 444, + 349 + ], + "blocks": [ + { + "bbox": [ + 353, + 258, + 444, + 349 + ], + "lines": [ + { + "bbox": [ + 353, + 258, + 444, + 349 + ], + "spans": [ + { + "bbox": [ + 353, + 258, + 444, + 349 + ], + "type": "image", + "image_path": "cc63f69098b5c6a2217ed56b43849dfbfd441be35c9a7caf06efa322f88822c2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 446, + 258, + 538, + 349 + ], + "blocks": [ + { + "bbox": [ + 446, + 258, + 538, + 349 + ], + "lines": [ + { + "bbox": [ + 446, + 258, + 538, + 349 + ], + "spans": [ + { + "bbox": [ + 446, + 258, + 538, + 349 + ], + "type": "image", + "image_path": "32f2371c429606c2eacbae6fad263eb1f0335f9ffd21efe8ccb5a28850ca27dd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 390, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 390, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 295, + 510 + ], + "type": "text", + "content": "Motivation. While point-supervised OOD has gained attention, the utilization of relationships among instances remains absent in the literature. For example, the arrangement of vehicles in a parking lot can serve as effective constraints for learning their size and orientation. This approach may yield valuable insights into point-supervised OOD, particularly in densely packed scenes where current methods face significant challenges. Can we leverage the spatial layout of objects to enhance point-supervised OOD? In this paper, we dive into this idea and seek to answer this question." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 510, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 593 + ], + "type": "text", + "content": "Highlights. 1) Point2RBox-v2 is proposed for point-supervised OOD, advancing the state of the art as displayed in Fig. 2 and Tables 1-2. 2) We propose novel and elegant losses to enforce constraints from the spatial layout among instances based on Gaussian overlap and Voronoi tessellation [1]. 3) Other modules are devised (i.e. edge loss, consistency loss, copy-paste) to further enhance the method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Contributions. 1) To our best knowledge, this work is the first attempt to learn point-supervised OOD from the layout among objects, where we propose novel principles based on Gaussian overlap and Voronoi tessellation. 2) The training pipeline and detailed implementation are elucidated, with necessary modules (i.e. edge loss, consistency loss, copy-paste) incorporated. The source code will be made publicly available. 3) Extensive experiments demonstrate that leveraging the spatial layout of instances can significantly advance the state of the art, surpassing other alternatives in accuracy." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 388, + 400, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 388, + 400, + 401 + ], + "spans": [ + { + "bbox": [ + 314, + 388, + 400, + 401 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 408, + 508, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 508, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 508, + 421 + ], + "type": "text", + "content": "2.1. RBox-supervised Oriented Detection" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": "In addition to horizontal detection [26, 71], oriented object detection (OOD) [46, 52] has received extensive attention. Representative works include anchor-based detector Rotated RetinaNet [25], anchor-free detector Rotated FCOS [42], and two-stage solutions, e.g. RoI Transformer [5], Oriented R-CNN [49], and ReDet [11]. Some research enhances the detector by exploiting alignment features, e.g. " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{R}^3\\mathrm{Det}" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": " [55] and " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{S}^2\\mathrm{A}" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": "-Net [12]. The angle regression may face boundary discontinuity and remedies are developed, including modulated losses [37, 53, 58] that alleviate loss jumps, angle coders [50, 51, 54, 63] that convert the angle into boundary-free coded data, and Gaussian-based losses [34, 56, 57, 60, 61] transforming rotated bounding boxes into Gaussian distributions. RepPoint-based methods [15, 20, 62] provide alternatives that predict a set of points that bounds the spatial extent of an object. LMMRotate [19] is a new paradigm of OOD based on multimodal language model." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 636, + 507, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 507, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 507, + 649 + ], + "type": "text", + "content": "2.2. Point-supervised Oriented Detection" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": "Recently, several methods for point-supervised OOD have been proposed: 1) P2RBox [2], PMHO [69], and PointSAM [27] propose point-prompt OOD by employing the zero-shot Point-to-Mask ability of SAM [17]. 2) Point2RBox [66] introduces an end-to-end approach based on knowledge com" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19284" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 547, + 240 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 547, + 240 + ], + "type": "image", + "image_path": "7f7ab96800cefc9e462732b3d097e16a231a98bbb2684aa785ee6bb13d3a5146.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 241, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 55, + 241, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 555, + 265 + ], + "type": "text", + "content": "Figure 3. The training pipeline of Point2RBox-v2. Gaussian overlap loss and Voronoi watershed loss utilize the spatial layout (see Fig. 4), while edge loss (see Sec. 3.4), symmetry-aware learning (see Sec. 3.5), and copy-paste (see Sec. 3.6) further enhance the method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 278, + 294, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 278, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 54, + 278, + 294, + 338 + ], + "type": "text", + "content": "bination. 3) PointOBB [33, 68] achieves RBox generation through multiple instance learning. 4) PointOBB-v2 [38] learns a class probability map to generate pseudo RBoxes. 5) Wholly-WOOD [67] uses points, HBoxes, RBoxes, or their combinations for mixed supervised OOD." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 339, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 339, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 339, + 295, + 399 + ], + "type": "text", + "content": "Among these methods, P2RBox, PMHO, and PointSAM rely on the SAM model pre-trained on massive labeled datasets, whereas Point2RBox requires one-shot examples for each category. In contrast, PointOBB series do not use many priors, but they necessitate two-stage training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 407, + 237, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 237, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 237, + 420 + ], + "type": "text", + "content": "2.3. Other Weakly-supervised Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 426, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 496 + ], + "type": "text", + "content": "Compared to Point-to-RBox, some other settings have been better studied. These methods are potentially applicable to our Point-to-RBox task setting by using a cascade pipeline, such as Point-to-HBox-to-RBox. In our experiment, cascade pipelines powered by state-of-the-art approaches are also compared. Here, representative works are introduced." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 498, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 641 + ], + "type": "text", + "content": "HBox-to-RBox. H2RBox [59] establishes a paradigm that limits the object to a few candidate angles through geometric constraint from HBoxes, with a self-supervised branch eliminating the undesired results. An enhanced version H2RBox-v2 [65] is proposed to leverage the reflection symmetry of objects to further boost the accuracy. EIE-Det [45] uses an explicit equivariance branch for learning rotation consistency, and an implicit equivariance branch for learning position, aspect ratio, and scale consistency. KCR [74] combines RBox- and HBox-annotated datasets for transfer learning. Some studies [16, 41] use additional annotated data for training, which are also attractive but less general." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 295, + 713 + ], + "type": "text", + "content": "Point-to-HBox. Several related approaches have been developed, including: 1) P2BNet [3] samples box proposals of different sizes around the labeled point and classify them to achieve point-supervised horizontal object detection. 2) PSOD [8] achieves point-supervised salient object detection using an edge detector and adaptive masked flood fill." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 278, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 363 + ], + "type": "text", + "content": "Point-to-Mask. Point2Mask [21] is proposed to achieve panoptic segmentation using single point annotation per target. SAM (Segment Anything Model) [17] produces object masks from input point/HBox prompts. Though RBoxes can be obtained from the segmentation mask by finding the minimum circumscribed rectangle, such a complex pipeline can be less cost-efficient and perform worse [59, 65]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 373, + 370, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 370, + 385 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 370, + 385 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 393, + 462, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 462, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 462, + 406 + ], + "type": "text", + "content": "3.1. Overview and Preliminary" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 410, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 494 + ], + "type": "text", + "content": "An overview of Point2RBox-v2 is illustrated in Fig. 3. The network is based on ResNet50 [13] backbone, FPN [24] head, and PSC [64] angle coder. Objects of varying sizes are typically assigned to different FPN layers based on the scale. However, points lack size, so we assign them all to the P3 layer (stride = 8). Assume the detector " + }, + { + "bbox": [ + 313, + 410, + 554, + 494 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{nn}}(\\cdot)" + }, + { + "bbox": [ + 313, + 410, + 554, + 494 + ], + "type": "text", + "content": " maps an image " + }, + { + "bbox": [ + 313, + 410, + 554, + 494 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 410, + 554, + 494 + ], + "type": "text", + "content": " to a set of RBoxes as detection results:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 381, + 499, + 555, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 499, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 381, + 499, + 555, + 513 + ], + "type": "interline_equation", + "content": "\\left(x _ {c}, y _ {c}, w, h, \\theta\\right) = f _ {\\mathrm {n n}} (I) \\tag {1}", + "image_path": "d3787e626f191aa4420ffe70b21d2187f2556dfb13fcfe0e1217d1f5255f2ca0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 518, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 554, + 543 + ], + "type": "text", + "content": "Equivalently, oriented objects can also be represented by 2D Gaussian distributions " + }, + { + "bbox": [ + 313, + 518, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu, \\Sigma)" + }, + { + "bbox": [ + 313, + 518, + 554, + 543 + ], + "type": "text", + "content": " [60]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 547, + 555, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 547, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 337, + 547, + 555, + 576 + ], + "type": "interline_equation", + "content": "\\left\\{\\mu = \\left[ \\begin{array}{l l} x _ {c} & y _ {c} \\end{array} \\right], \\quad \\Sigma = \\mathbf {R} \\left[ \\begin{array}{c c} w / 2 & 0 \\\\ 0 & h / 2 \\end{array} \\right] ^ {2} \\mathbf {R} ^ {\\top} \\right\\} \\tag {2}", + "image_path": "84d591fd74355f9917eef6977810bfc651dda95ce7ee91cc9effea445fc3e3ab.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 579, + 342, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 342, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 342, + 590 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 387, + 587, + 555, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 587, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 387, + 587, + 555, + 614 + ], + "type": "interline_equation", + "content": "\\mathbf {R} = \\left[ \\begin{array}{c c} \\cos \\theta & - \\sin \\theta \\\\ \\sin \\theta & \\cos \\theta \\end{array} \\right] \\tag {3}", + "image_path": "a81e69eb7e1aa6ae279b26bed64bd4c47cfee4981b10266bbe0a9f8342b455d4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "At the core of Point2RBox-v2 is to utilize the constraints from the layout, which is achieved by Gaussian overlap loss (Sec. 3.2) and Voronoi watershed loss (Sec. 3.3). These two losses effectively limit the size and rotation of objects to a reasonable range. Upon that, edge loss (Sec. 3.4) aligns the bounding box with the edge of objects to improve the accuracy. Incorporated with symmetry-aware learning (Sec. 3.5) and copy-paste augmentation (Sec. 3.6), we achieve a" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19285" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 541, + 373 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 541, + 373 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 541, + 373 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 541, + 373 + ], + "type": "image", + "image_path": "e493037026f870df4b6d99c842b610fed2377006c2b5ec8358a853e93a79d1fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 376, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 555, + 399 + ], + "type": "text", + "content": "Figure 4. To illustrate the procedure of the three newly proposed loss functions and their impact on the learning results. (a) Gaussian overlap loss (see Sec. 3.2). (b) Voronoi watershed loss (see Sec. 3.3). (c) Edge loss (see Sec. 3.4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 414, + 297, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 297, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 297, + 437 + ], + "type": "text", + "content": "stable and high-accuracy solution for point-supervised OOD. In subsequent subsections, these modules are detailed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 187, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 187, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 187, + 462 + ], + "type": "text", + "content": "3.2. Gaussian Overlap Loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 515 + ], + "type": "text", + "content": "As mentioned, oriented objects can be represented by 2D Gaussian distributions " + }, + { + "bbox": [ + 55, + 468, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu, \\Sigma)" + }, + { + "bbox": [ + 55, + 468, + 296, + 515 + ], + "type": "text", + "content": ". The overlap volume between two distributions " + }, + { + "bbox": [ + 55, + 468, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_1(\\mu_1, \\Sigma_1)" + }, + { + "bbox": [ + 55, + 468, + 296, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 468, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_2(\\mu_2, \\Sigma_2)" + }, + { + "bbox": [ + 55, + 468, + 296, + 515 + ], + "type": "text", + "content": " can be approximated by the Bhattacharyya coefficient [60] as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 520, + 296, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 520, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 64, + 520, + 296, + 548 + ], + "type": "interline_equation", + "content": "B \\left(\\mathcal {N} _ {1}, \\mathcal {N} _ {2}\\right) = \\exp \\left(- \\frac {1}{8} \\mu^ {\\top} \\Sigma^ {- 1} \\mu\\right) \\cdot \\frac {\\left| \\Sigma_ {1} \\right| ^ {1 / 4} \\left| \\Sigma_ {2} \\right| ^ {1 / 4}}{\\left| \\Sigma \\right| ^ {1 / 2}} \\tag {4}", + "image_path": "d9b45c5e736b5869270865263536ed0ff9453e9b7cd3acb07c5f91f79a6e5e5d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 295, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 552, + 295, + 576 + ], + "type": "inline_equation", + "content": "\\mu = \\mu_{2} - \\mu_{1}" + }, + { + "bbox": [ + 55, + 552, + 295, + 576 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 552, + 295, + 576 + ], + "type": "inline_equation", + "content": "\\Sigma = \\frac{1}{2} (\\Sigma_{1} + \\Sigma_{2})" + }, + { + "bbox": [ + 55, + 552, + 295, + 576 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 552, + 295, + 576 + ], + "type": "inline_equation", + "content": "|\\Sigma|" + }, + { + "bbox": [ + 55, + 552, + 295, + 576 + ], + "type": "text", + "content": " denotes the determinant of the covariance matrix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 602 + ], + "type": "text", + "content": "Based on the above equation, we build a Gaussian overlap matrix " + }, + { + "bbox": [ + 55, + 578, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 55, + 578, + 295, + 602 + ], + "type": "text", + "content": " for each scene image as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 607, + 295, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 607, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 134, + 607, + 295, + 622 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i, j} = B \\left(\\mathcal {N} _ {i}, \\mathcal {N} _ {j}\\right) \\tag {5}", + "image_path": "eda8fdc31bf77574a634b0776af0023dad5f2df6dcccc37656e654cce3bc6ae5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 625, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 295, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 625, + 295, + 650 + ], + "type": "inline_equation", + "content": "i,j = 1,2,\\ldots ,N;N" + }, + { + "bbox": [ + 55, + 625, + 295, + 650 + ], + "type": "text", + "content": " is the distribution count (or the instance count) within one training image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 651, + 279, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 279, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 279, + 663 + ], + "type": "text", + "content": "The Gaussian overlap loss can then be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 131, + 667, + 296, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 667, + 296, + 697 + ], + "spans": [ + { + "bbox": [ + 131, + 667, + 296, + 697 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {O} = \\frac {1}{N} \\sum_ {i \\neq j} \\left(\\mathbf {M} _ {i, j}\\right) \\tag {6}", + "image_path": "7c893b80295615a5053538a0bf679354952d9ea4dd4f6b4d4078264b9beafe54.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": " omits diagonal elements. With this loss, the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 414, + 555, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 438 + ], + "type": "text", + "content": "detector learns to arrange instances (see Fig. 4a) based on the mutual exclusivity among instances." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 445, + 452, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 452, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 452, + 456 + ], + "type": "text", + "content": "3.3. Voronoi Watershed Loss" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 462, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 545 + ], + "type": "text", + "content": "A Voronoi diagram [1] is a partitioning of a space based on a set of points. Point annotations can be effectively utilized to calculate a Voronoi diagram, where a distinct polygon region is assigned to each point-annotated instance. The watershed algorithm [44], on the other hand, is a region segmentation technique that treats the intensity of pixels as a topographic surface, identifying regions as \"catchment basins\"." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 546, + 550, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 546, + 550, + 558 + ], + "spans": [ + { + "bbox": [ + 326, + 546, + 550, + 558 + ], + "type": "text", + "content": "The calculation of Voronoi ridges can be formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 395, + 563, + 555, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 563, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 395, + 563, + 555, + 576 + ], + "type": "interline_equation", + "content": "V = \\text {V o r o n o i} (X) \\tag {7}", + "image_path": "9ed7b7f4a5035b5be9bee7d97b68a668edcfa58473129343a00d88556648c931.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": " are the annotated points within a training image; " + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": " are the output Voronoi ridges (pixel coordinates)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 606, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 666 + ], + "type": "text", + "content": "Interestingly, we find that Voronoi diagrams can be utilized as initial markers for watershed to obtain a region for each instance (see Fig. 4b). In concrete terms, the points " + }, + { + "bbox": [ + 313, + 606, + 555, + 666 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 606, + 555, + 666 + ], + "type": "text", + "content": " can be employed as foreground markers, while the Voronoi ridges " + }, + { + "bbox": [ + 313, + 606, + 555, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 606, + 555, + 666 + ], + "type": "text", + "content": " can act as background boundaries:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 379, + 671, + 555, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 671, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 379, + 671, + 555, + 685 + ], + "type": "interline_equation", + "content": "S = \\text {W a t e r s h e d} (I, X, V) \\tag {8}", + "image_path": "aad363de77fd8e5fe009773803b8728dab9cbf14cceb6967cc3f2bd612da120e.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " are the output basin regions (pixel coordinates) corresponding to each annotated instance. By rotating " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "S" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19286" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "to align with the direction of the current prediction, the regression target of width and height can be expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 296, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 100, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 296, + 128 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l} w _ {t} \\\\ h _ {t} \\end{array} \\right] = 2 \\max \\left| \\mathbf {R} ^ {\\top} \\left(S - \\left[ \\begin{array}{l} x _ {c} \\\\ y _ {c} \\end{array} \\right]\\right) \\right| \\tag {9}", + "image_path": "1475d452fc47cd403079d718376abae3b9311909781e81986be1ac117e758108.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 132, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 132, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 132, + 295, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 132, + 295, + 156 + ], + "type": "inline_equation", + "content": "(x_{c},y_{c},w,h,\\theta)" + }, + { + "bbox": [ + 55, + 132, + 295, + 156 + ], + "type": "text", + "content": " is the current prediction; " + }, + { + "bbox": [ + 55, + 132, + 295, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 132, + 295, + 156 + ], + "type": "text", + "content": " is defined by Eq. (3); " + }, + { + "bbox": [ + 55, + 132, + 295, + 156 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 55, + 132, + 295, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 132, + 295, + 156 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 55, + 132, + 295, + 156 + ], + "type": "text", + "content": " are detached to stop the gradient." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 156, + 296, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 296, + 179 + ], + "type": "text", + "content": "Afterward, the Voronoi watershed loss to regress the width and height of objects can be calculated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 184, + 296, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 184, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 68, + 184, + 296, + 217 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {W} = L _ {\\mathrm {G W D}} \\left(\\left[ \\begin{array}{c c} w / 2 & 0 \\\\ 0 & h / 2 \\end{array} \\right] ^ {2}, \\left[ \\begin{array}{c c} w _ {t} / 2 & 0 \\\\ 0 & h _ {t} / 2 \\end{array} \\right] ^ {2}\\right) \\tag {10}", + "image_path": "52cb9522c942d8c65ef08cd113e2738b0d96684951d21098d46e2cb7b07b9bf1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 221, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 233 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 221, + 296, + 233 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{GWD}}(\\cdot)" + }, + { + "bbox": [ + 55, + 221, + 296, + 233 + ], + "type": "text", + "content": " is Gaussian Wasserstein Distance Loss [60]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 240, + 127, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 127, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 127, + 252 + ], + "type": "text", + "content": "3.4. Edge Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "text", + "content": "The above two losses have limited the size to a reasonable range. To make it more accurate, we propose the edge loss to snap the boundaries toward the edges (see Fig. 4c)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 293, + 296, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 318 + ], + "type": "text", + "content": "First, a region " + }, + { + "bbox": [ + 55, + 293, + 296, + 318 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{(2K + 1) \\times (2K + 1)}" + }, + { + "bbox": [ + 55, + 293, + 296, + 318 + ], + "type": "text", + "content": " around each predicted RBox is extracted via Rotated RoI Align [14]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 87, + 323, + 296, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 323, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 87, + 323, + 296, + 337 + ], + "type": "interline_equation", + "content": "P = R o I A l i g n (E (I), \\left(x _ {c}, y _ {c}, \\beta w, \\beta h, \\theta\\right)) \\tag {11}", + "image_path": "7afaf50618bc2101a592c855ec712acea0611f1d1107d41aff2fb3d2619598a1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "text", + "content": " is the edge detection function [39]. We set " + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "inline_equation", + "content": "K = 24" + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "inline_equation", + "content": "\\beta = 1.6" + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "text", + "content": " in our experiments (see Table 9)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "text", + "content": "By calculating the sum of each row of " + }, + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "text", + "content": ", the edge distribution in " + }, + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 365, + 296, + 389 + ], + "type": "text", + "content": " direction can be obtained as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 96, + 394, + 296, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 394, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 96, + 394, + 296, + 428 + ], + "type": "interline_equation", + "content": "\\mu_ {i} = \\sum_ {j = 1} ^ {2 K + 1} \\left(P _ {(K + 1 - i), j} + P _ {(K + 1 + i), j}\\right) \\tag {12}", + "image_path": "7e1bc27685be24f9164bfc3ad15b4aeae1a7ec89611c51704220cb0314e9995d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 432, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 432, + 295, + 455 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots ,K" + }, + { + "bbox": [ + 55, + 432, + 295, + 455 + ], + "type": "text", + "content": ", indicating that the upper half of " + }, + { + "bbox": [ + 55, + 432, + 295, + 455 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 432, + 295, + 455 + ], + "type": "text", + "content": " is reversed and added to the lower half." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 456, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 479 + ], + "type": "text", + "content": "Meanwhile, the current prediction of the edge can also be softened into a distribution as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 117, + 482, + 296, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 482, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 117, + 482, + 296, + 510 + ], + "type": "interline_equation", + "content": "\\lambda_ {i} = \\exp \\left(- \\frac {(i - K / \\beta) ^ {2}}{2 \\sigma_ {E} {} ^ {2}}\\right) \\tag {13}", + "image_path": "070fbc5e330cd3c65133c6f2ca1afee42827b9637d27f3631b34bc9d0a93305d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 514, + 295, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 514, + 295, + 537 + ], + "type": "inline_equation", + "content": "i = 1,2,\\ldots ,K;\\sigma_{E}" + }, + { + "bbox": [ + 55, + 514, + 295, + 537 + ], + "type": "text", + "content": " is set to 6. Note that we crop " + }, + { + "bbox": [ + 55, + 514, + 295, + 537 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 514, + 295, + 537 + ], + "type": "text", + "content": " based on the predicted RBox, thus " + }, + { + "bbox": [ + 55, + 514, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 514, + 295, + 537 + ], + "type": "text", + "content": " is always the same." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 538, + 271, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 538, + 271, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 538, + 271, + 550 + ], + "type": "text", + "content": "Multiplying the two distributions yields the target:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 121, + 553, + 296, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 553, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 121, + 553, + 296, + 578 + ], + "type": "interline_equation", + "content": "h _ {t} = \\frac {\\beta h}{K} \\arg \\max (\\mu \\times \\lambda) \\tag {14}", + "image_path": "587d9c83a73f12caac7f705954b078e110d1907299ecf46bfc34d1e08cb03373.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "text", + "content": " is the regression target of height. Likewise, the width target " + }, + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "inline_equation", + "content": "w_t" + }, + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "text", + "content": " is calculated along " + }, + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "text", + "content": " direction." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 67, + 605, + 205, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 605, + 205, + 616 + ], + "spans": [ + { + "bbox": [ + 67, + 605, + 205, + 616 + ], + "type": "text", + "content": "The edge loss is then obtained as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 96, + 621, + 296, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 621, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 96, + 621, + 296, + 637 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {E} = \\operatorname {s m o o t h} _ {L 1} \\left(\\left[ \\begin{array}{l l} w & h \\end{array} \\right], \\left[ \\begin{array}{l l} w _ {t} & h _ {t} \\end{array} \\right]\\right) \\tag {15}", + "image_path": "c86cf16299afe805c34136c66e30784137678056749ecb5f270cc08673501481.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_W" + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_E" + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "content": " merely refines the width and height of boxes, without involving the angle regression." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 672, + 205, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 205, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 205, + 685 + ], + "type": "text", + "content": "3.5. Symmetry-aware Learning" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Proven effective in learning the rotation of objects [65], in this work, symmetry-aware learning is extended to Gaussian-" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "based OOD. In the left part of Fig. 3, the training image " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " is transformed (randomly selected from rotation, flip, and scale, see Table 7) to generate an augmented view as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 413, + 114, + 555, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 114, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 413, + 114, + 555, + 127 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {a u g}} = \\alpha I \\tag {16}", + "image_path": "1cc62818e276fc60b17b57456baca8a1e3f627afa1635e8445c9a5fe38d02c85.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 133, + 342, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 342, + 142 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 342, + 142 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 328, + 146, + 555, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 146, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 328, + 146, + 555, + 174 + ], + "type": "interline_equation", + "content": "\\alpha = \\left[ \\begin{array}{c c} \\cos \\mathcal {R} & - \\sin \\mathcal {R} \\\\ \\sin \\mathcal {R} & \\cos \\mathcal {R} \\end{array} \\right] ^ {p _ {1}} \\left[ \\begin{array}{c c} 1 & 0 \\\\ 0 & - 1 \\end{array} \\right] ^ {p _ {2}} \\left[ \\begin{array}{c c} s & 0 \\\\ 0 & s \\end{array} \\right] ^ {p _ {3}} \\tag {17}", + "image_path": "1d5c7bdb50ce17eff8e308a27667ae2322663f17c17493b9a87fede7ac8247cc.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "inline_equation", + "content": "(p_1, p_2, p_3) = (1, 0, 0)" + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "text", + "content": " when rotation is selected, " + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "inline_equation", + "content": "(0, 1, 0)" + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "text", + "content": " when flip, " + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "inline_equation", + "content": "(0, 0, 1)" + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "text", + "content": " when scale. " + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "text", + "content": " are the random amount of rotation and scale, " + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "inline_equation", + "content": "s \\in (0.5, 0.9)" + }, + { + "bbox": [ + 313, + 178, + 554, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 214, + 554, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 554, + 239 + ], + "type": "text", + "content": "By feeding both " + }, + { + "bbox": [ + 313, + 214, + 554, + 239 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 214, + 554, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 214, + 554, + 239 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{aug}}" + }, + { + "bbox": [ + 313, + 214, + 554, + 239 + ], + "type": "text", + "content": " into the network, we obtain two sets of output Gaussian distributions and angles:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 375, + 243, + 554, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 243, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 375, + 243, + 554, + 270 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} (\\Sigma , \\theta) = f _ {\\mathrm {n n}} (I) \\\\ (\\Sigma_ {\\mathrm {a u g}}, \\theta_ {\\mathrm {a u g}}) = f _ {\\mathrm {n n}} (I _ {\\mathrm {a u g}}) \\end{array} \\right. \\tag {18}", + "image_path": "efa487478709676162b71d86f41853c67a05602b8e61c29a0da6abe4e2d888a3.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 274, + 554, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 554, + 297 + ], + "type": "text", + "content": "The consistency loss is calculated between the two sets so that the network learns the size and rotation variation:" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 324, + 302, + 555, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 302, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 324, + 302, + 555, + 317 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s s}} = L _ {\\mathrm {G W D}} \\left(\\alpha \\Sigma \\alpha^ {\\top}, \\Sigma_ {\\mathrm {a u g}}\\right) + L _ {\\mathrm {A N G}} (m \\theta + \\mathcal {R}, \\theta_ {\\mathrm {a u g}}) \\tag {19}", + "image_path": "f150dd7a6ac2be7c1920b11e5b4ee86f7d7bb51f7b470cf0b6147940a6b53f21.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "text", + "content": " is the rotation angle and " + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "inline_equation", + "content": "m = 1" + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "text", + "content": " when rotation is selected; " + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "inline_equation", + "content": "(m,\\mathcal{R}) = (-1,0)" + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "text", + "content": " when flip; " + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "inline_equation", + "content": "(m,\\mathcal{R}) = (1,0)" + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "text", + "content": " when scale. " + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{GWD}}(\\cdot)" + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "text", + "content": " is Gaussian Wasserstein Distance Loss [60] and " + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ANG}}(\\cdot)" + }, + { + "bbox": [ + 313, + 322, + 554, + 369 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 328, + 375, + 555, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 375, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 328, + 375, + 555, + 392 + ], + "type": "interline_equation", + "content": "L _ {\\text {A N G}} \\left(\\theta_ {1}, \\theta_ {2}\\right) = \\min _ {k \\in Z} \\left(s m o o t h _ {L 1} \\left(\\theta_ {1}, k \\pi + \\theta_ {2}\\right)\\right) \\tag {20}", + "image_path": "8918554bd00da89598c4fa5cfc33c0437f1dddb5a4fff8cabc9b313a512881bd.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 398, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 554, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 398, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\min (\\cdot)" + }, + { + "bbox": [ + 313, + 398, + 554, + 422 + ], + "type": "text", + "content": " regresses the prediction toward the closest target to circumvent the periodicity problem [65]." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 431, + 459, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 459, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 459, + 444 + ], + "type": "text", + "content": "3.6. Copy-paste Augmentation" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "content": "Inspired by [9], we propose to crop the detected instances of step " + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "content": " and paste them on the training image of step " + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "inline_equation", + "content": "k + 1" + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "content": ". The maximum number of paste boxes is limited to 10 in each step. We simply use the bounding boxes of the cropped instances as the regression targets, and use Gaussian Wasserstein Distance Loss [60] to calculate " + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{box}}" + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 530, + 397, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 397, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 397, + 542 + ], + "type": "text", + "content": "3.7. Overall Loss" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 548, + 550, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 550, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 550, + 560 + ], + "type": "text", + "content": "The overall loss " + }, + { + "bbox": [ + 313, + 548, + 550, + 560 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 548, + 550, + 560 + ], + "type": "text", + "content": " for Point2RBox-v2 can be expressed as:" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 325, + 566, + 555, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 566, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 325, + 566, + 555, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c l s}} + w _ {\\text {b o x}} \\mathcal {L} _ {\\text {b o x}} + w _ {\\mathrm {O}} \\mathcal {L} _ {\\mathrm {O}} + w _ {\\mathrm {W}} \\mathcal {L} _ {\\mathrm {W}} + w _ {\\mathrm {E}} \\mathcal {L} _ {\\mathrm {E}} + w _ {\\mathrm {s s}} \\mathcal {L} _ {\\mathrm {s s}} \\tag {21}", + "image_path": "d78c6cede2c25bf75407894c8177f170395712e8450884ae6a1363f04584fbb7.jpg" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 584, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 555, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 584, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 584, + 555, + 632 + ], + "type": "text", + "content": " is the focal loss [25] for classification, " + }, + { + "bbox": [ + 313, + 584, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{box}}" + }, + { + "bbox": [ + 313, + 584, + 555, + 632 + ], + "type": "text", + "content": " regresses boxes/centers toward copy-paste/ground-truth labels, " + }, + { + "bbox": [ + 313, + 584, + 555, + 632 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{box}}" + }, + { + "bbox": [ + 313, + 584, + 555, + 632 + ], + "type": "text", + "content": " is set to one by default, " + }, + { + "bbox": [ + 313, + 584, + 555, + 632 + ], + "type": "inline_equation", + "content": "(w_{\\mathrm{O}}, w_{\\mathrm{W}}, w_{\\mathrm{E}}, w_{\\mathrm{ss}})" + }, + { + "bbox": [ + 313, + 584, + 555, + 632 + ], + "type": "text", + "content": " are set to (10, 5, 0.3, 1) based on our ablation studies (see Tables 3-5)." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 645, + 394, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 394, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 394, + 658 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Experiments are carried out on NVIDIA RTX4090 GPUs using PyTorch 2.2.0 [36] and the rotation detection tool kits: MMRotate 1.0.0 [73]. All the experiments follow the same hyper-parameters (learning rate, batch size, optimizer, etc.)." + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19287" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 550, + 403 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 550, + 403 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 550, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 550, + 403 + ], + "type": "table", + "html": "
Methods*PL1BDBRGTFSVLVSHTCBCSTSBFRAHASPHCAP50
▼RBox-supervised OOD
RepPoints (2019) [62]✓86.781.141.662.076.256.375.790.780.885.363.366.659.167.633.768.45
RetinaNet (2017) [25]✓88.277.045.069.471.559.074.590.884.979.357.364.762.766.539.668.69
GWD (2021) [56]✓89.375.447.861.979.573.886.190.984.579.455.959.763.271.045.471.66
FCOS (2019) [42]✓89.176.950.163.279.879.887.190.480.884.659.766.365.871.341.772.44
S2A-Net (2022) [12]✓89.283.052.574.678.879.287.590.984.984.861.968.070.771.459.875.81
▼HBox-supervised OOD
Sun et al. (2021) [41]×51.538.716.136.829.819.223.483.950.680.018.950.225.628.725.538.60
BoxInst-RBox (2021) [43]2×68.440.833.132.346.955.456.679.566.882.141.252.852.865.030.053.59
H2RBox (2023) [59]✓88.573.540.856.977.565.477.990.983.285.355.362.952.463.643.367.82
EIE-Det (2024) [45]✓87.770.241.560.580.776.386.390.982.684.753.164.558.170.443.870.10
H2RBox-v2 (2023) [65]✓89.074.450.060.579.875.386.990.985.185.059.263.265.270.549.772.31
▼Point-supervised OOD
Point2Mask-RBox (2023) [21]2×4.023.13.81.315.11.03.319.01.029.10.09.57.421.17.19.72
P2BNet+H2RBox (2023) [3, 59]×24.735.97.127.93.312.117.517.50.834.06.349.611.627.218.819.63
P2BNet+H2RBox-v2 (2023) [3, 65]×11.044.814.915.436.816.727.812.11.831.23.450.612.636.712.521.87
P2RBox (2024) [2]†×87.865.715.060.773.071.778.981.544.581.241.239.345.557.541.259.04
PointOBB (2024) [33]×26.165.79.159.465.834.929.80.52.316.70.649.021.841.036.730.08
Point2RBox (2024) [66]✓62.964.314.435.028.238.933.325.22.244.53.448.125.945.022.634.07
Point2RBox+SK (2024) [66]†✓53.363.93.750.940.039.245.776.710.556.15.449.524.251.233.840.27
Point2RBox+SK (2024) [66]†×66.459.55.252.654.153.957.390.83.257.86.147.422.955.740.544.90
PointOBB-v2 (2025) [38]×64.527.81.936.258.847.253.490.562.245.312.141.78.143.732.041.68
PointOBB-v3 (2025) [68]✓30.939.413.522.761.27.043.162.459.847.32.745.116.855.211.441.29
PointOBB-v3 (2025) [68]×52.954.421.352.765.644.967.887.226.773.432.653.339.056.410.249.24
Point2RBox-v2 (ours)✓78.452.78.340.971.060.574.788.765.572.124.426.130.150.721.051.00
Point2RBox-v2 (ours)×88.072.68.046.279.676.386.989.179.782.926.245.345.866.346.362.61
*Comparison tracks: √ = End-to-end training and testing; × = Generating pseudo labels to train the FCOS detector (two-stage training).†Using additional priors. P2RBox: Pre-trained SAM model; Point2RBox+SK: One-shot sketches for each class.†PL: Plane, BD: Baseball diamond, BR: Bridge, GTF: Ground track field, SV: Small vehicle, LV: Large vehicle, SH: Ship, TC: Tennis court, BC: Basketball court, ST: Storage tank, SBF: Soccer-ball field, RA: Roundabout, HA: Harbor, SP: Swimming pool, HC: Helicopter.‡RBox: The minimum rectangle operation is performed on the output Mask to obtain the RBox.
", + "image_path": "e4131da4848aee7ba00ff98671702dd7665daf3fd9c22843b21b7598dc65396e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 114, + 405, + 493, + 416 + ], + "lines": [ + { + "bbox": [ + 114, + 405, + 493, + 416 + ], + "spans": [ + { + "bbox": [ + 114, + 405, + 493, + 416 + ], + "type": "text", + "content": "Table 1. Detection performance of each category on the DOTA-v1.0 and the mean " + }, + { + "bbox": [ + 114, + 405, + 493, + 416 + ], + "type": "inline_equation", + "content": "{\\mathrm{{AP}}}_{50}" + }, + { + "bbox": [ + 114, + 405, + 493, + 416 + ], + "type": "text", + "content": " of all categories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 428, + 297, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 297, + 571 + ], + "type": "text", + "content": "Average precision (AP) is adopted as the primary metric. All the models are configured upon ResNet50 [13] and trained with AdamW [31]. 1) Learning rate. Initialized at 5e-5, warm-up for 500 iterations, and divided by ten at each decay step. 2) Epochs. 72 for HRSC; 12 for the others. 3) Augmentation. Random rotation/flip for HRSC; random flip for the others. 4) Image size. Split into " + }, + { + "bbox": [ + 55, + 428, + 297, + 571 + ], + "type": "inline_equation", + "content": "1,024 \\times 1,024" + }, + { + "bbox": [ + 55, + 428, + 297, + 571 + ], + "type": "text", + "content": " with an overlap of 200 for DOTA/FAIR1M/STAR; scaled to " + }, + { + "bbox": [ + 55, + 428, + 297, + 571 + ], + "type": "inline_equation", + "content": "800 \\times 800" + }, + { + "bbox": [ + 55, + 428, + 297, + 571 + ], + "type": "text", + "content": " for others. 5) Multi-scale. All experiments evaluated without multi-scale technique [73]. 6) Datasets. Six remote sensing and one retail scene datasets, covering all datasets used by the main counterparts [2, 33, 66]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 582, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 617 + ], + "type": "text", + "content": "- DOTA [48]. DOTA-v1.0 has 2,806 aerial images annotated with 15 categories, while DOTA-v1.5/2.0 are the extended versions with more small objects and categories." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 653 + ], + "type": "text", + "content": "- DIOR [4]. It is an aerial image dataset re-annotated with RBoxes based on its original HBox version [18], with a high variation in object size and high intra-class diversity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 677 + ], + "type": "text", + "content": "- HRSC [30]. It contains ship instances on the sea and inshore. The train/val/test set includes 436/181/444 images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "- FAIR1M [40]. It has more than 1 million instances and more than 40,000 images for fine-grained object recognition in remote sensing imagery, annotated with 37 cate" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 428, + 523, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 428, + 523, + 439 + ], + "spans": [ + { + "bbox": [ + 322, + 428, + 523, + 439 + ], + "type": "text", + "content": "gories. The results are evaluated on FAIR1M-1.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 440, + 555, + 548 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 440, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 440, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 314, + 440, + 554, + 487 + ], + "type": "text", + "content": "- STAR [22]. It is extensive for scene graph generation, covering more than 210,000 objects with diverse spatial resolutions, classified into 48 fine-grained categories and precisely annotated with oriented bounding boxes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 487, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 555, + 523 + ], + "type": "text", + "content": "- SKU110K [35]. It focuses on the detection of densely packed retail scenes with 110,712 objects in 11,762 images. The density reaches 86 instances per image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 523, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 523, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 314, + 523, + 555, + 548 + ], + "type": "text", + "content": "- RSAR [70]. It is a remote sensing dataset based on Synthetic Aperture Radar (SAR) imagery with 6 categories." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 561, + 469, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 469, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 469, + 573 + ], + "type": "text", + "content": "4.1. Main Results on DOTA-v1.0" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "text", + "content": "Table 1 compares Point2RBox-v2 with the state-of-the-art methods, which can be categorized into two tracks:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 604, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "type": "text", + "content": "1) End-to-end training. These methods apply the trained weakly-supervised detector directly to the test set. Without relying on priors, our approach demonstrates an improvement of " + }, + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "type": "inline_equation", + "content": "16.93\\%" + }, + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "type": "text", + "content": " (51.00% vs. 34.07%) compared to Point2RBox. Even when compared to Point2RBox+SK, which incorporates additional data-side priors (i.e. one-shot examples for each class), our method still outperforms it by " + }, + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "type": "inline_equation", + "content": "10.73\\%" + }, + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "type": "text", + "content": " (51.00% vs. 40.27%)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 701, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 554, + 713 + ], + "type": "text", + "content": "2) Two-stage training. These methods generate RBox" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19288" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 550, + 321 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 321 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 321 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 321 + ], + "type": "table", + "html": "
Methods*DOTA-v1.0DOTA-v1.5DOTA-v2.0DIORHRSCFAIR1MSTARSKU110KRSAR
▼RBox-supervised OOD
RetinaNet (2017) [25]✓68.6960.5747.0054.9684.4937.6721.8078.5057.67
GWD (2021) [56]✓71.6663.2748.8757.6086.6739.1125.3079.1657.80
FCOS (2019) [42]✓72.4464.5351.7759.8388.9941.2528.1080.0966.66
S2A-Net (2022) [12]✓75.8166.5352.3961.4190.1042.4427.3080.3666.47
▼HBox-supervised OOD
Sun et al. (2021) [41]×38.60--------
KCR (2023) [74]✓----79.10----
H2RBox (2023) [59]✓70.0561.7048.6857.807.0335.9417.2057.1549.92
H2RBox-v2 (2023) [65]✓72.3164.7650.3357.6489.6642.2727.3070.7065.16
AFWS (2024) [32]✓72.5565.9251.7359.07-41.80---
▼Point-supervised OOD
P2RBox (2024) [2]†×59.04--------
PointSAM (2025) [27]†×---46.20-----
PointOBB (2024) [33]×30.0810.665.5337.31-11.199.19-13.80
Point2RBox+SK (2024) [66]†✓40.2730.5123.4327.3479.4020.037.863.4127.81
PointOBB-v2 (2025) [38]×41.6830.5920.6439.56-13.369.0056.6318.99
PointOBB-v3 (2025) [68]✓41.2031.2522.8237.60-11.4211.31-15.84
PointOBB-v3 (2025) [68]×49.2433.7923.5240.18-18.3512.85-22.60
Point2RBox-v2 (ours)✓51.0039.4527.1134.7082.6725.727.8064.0028.60
Point2RBox-v2 (ours)×62.6154.0638.7944.4586.1534.7114.2065.6430.90
*Comparison tracks: √= End-to-end training and testing; × = Generating pseudo labels to train the FCOS detector (two-stage training).†Using additional priors. P2RBox/PointSAM: Pre-trained SAM model; Point2RBox+SK: One-shot sketches for each class.
", + "image_path": "8a48b60e8c06edd4fef8b98ffdae2956a8555086058b7fb8cbbe3703ac5e1932.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 344, + 216, + 413 + ], + "blocks": [ + { + "bbox": [ + 66, + 325, + 542, + 337 + ], + "lines": [ + { + "bbox": [ + 66, + 325, + 542, + 337 + ], + "spans": [ + { + "bbox": [ + 66, + 325, + 542, + 337 + ], + "type": "text", + "content": "Table 2. Accuracy " + }, + { + "bbox": [ + 66, + 325, + 542, + 337 + ], + "type": "inline_equation", + "content": "\\left( {\\mathrm{{AP}}}_{50}\\right)" + }, + { + "bbox": [ + 66, + 325, + 542, + 337 + ], + "type": "text", + "content": " comparisons on the DOTA-v1.0/1.5/2.0, DIOR, HRSC, FAIR1M, STAR, SKU110K, and RSAR datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 344, + 216, + 413 + ], + "lines": [ + { + "bbox": [ + 62, + 344, + 216, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 344, + 216, + 413 + ], + "type": "table", + "html": "
\\(w_{O}\\)DOTAHRSC
E2EFCOSE2EFCOS
348.7661.6281.8584.36
549.8162.4482.4685.76
1051.0062.6182.6786.15
3045.8857.8381.5685.61
", + "image_path": "d9ac847a5316a81c66a138a340189589b56d353a87d8af36f5593abbc67b3a3e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 227, + 344, + 383, + 413 + ], + "blocks": [ + { + "bbox": [ + 63, + 416, + 212, + 428 + ], + "lines": [ + { + "bbox": [ + 63, + 416, + 212, + 428 + ], + "spans": [ + { + "bbox": [ + 63, + 416, + 212, + 428 + ], + "type": "text", + "content": "Table 3. Ablation with the weight of " + }, + { + "bbox": [ + 63, + 416, + 212, + 428 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{\\mathrm{O}}" + }, + { + "bbox": [ + 63, + 416, + 212, + 428 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 227, + 344, + 383, + 413 + ], + "lines": [ + { + "bbox": [ + 227, + 344, + 383, + 413 + ], + "spans": [ + { + "bbox": [ + 227, + 344, + 383, + 413 + ], + "type": "table", + "html": "
\\( {w}_{\\mathrm{W}} \\)DOTAHRSC
E2EFCOSE2EFCOS
350.8556.7878.4283.49
551.0062.6182.6786.15
1049.1560.5430.3735.13
3042.8452.5323.8925.91
", + "image_path": "9888f00fcc3664d536be816ddeea6798415124dd597845c19eb0faaf9b4a7aec.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 395, + 344, + 548, + 413 + ], + "blocks": [ + { + "bbox": [ + 229, + 416, + 379, + 428 + ], + "lines": [ + { + "bbox": [ + 229, + 416, + 379, + 428 + ], + "spans": [ + { + "bbox": [ + 229, + 416, + 379, + 428 + ], + "type": "text", + "content": "Table 4. Ablation with the weight of " + }, + { + "bbox": [ + 229, + 416, + 379, + 428 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{\\mathrm{W}}" + }, + { + "bbox": [ + 229, + 416, + 379, + 428 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 395, + 344, + 548, + 413 + ], + "lines": [ + { + "bbox": [ + 395, + 344, + 548, + 413 + ], + "spans": [ + { + "bbox": [ + 395, + 344, + 548, + 413 + ], + "type": "table", + "html": "
wEDOTAHRSC
E2EFCOSE2EFCOS
0.148.7557.6234.7139.45
0.351.0062.6182.6786.15
0.551.3662.6376.8585.22
1.049.0560.6356.5959.59
", + "image_path": "721bc0741978f2e3538b152646b1bb8dc21165b18b0b41337812c5e5df25dc69.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 62, + 434, + 214, + 495 + ], + "blocks": [ + { + "bbox": [ + 397, + 416, + 545, + 428 + ], + "lines": [ + { + "bbox": [ + 397, + 416, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 397, + 416, + 545, + 428 + ], + "type": "text", + "content": "Table 5. Ablation with the weight of " + }, + { + "bbox": [ + 397, + 416, + 545, + 428 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{\\mathrm{E}}" + }, + { + "bbox": [ + 397, + 416, + 545, + 428 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 434, + 214, + 495 + ], + "lines": [ + { + "bbox": [ + 62, + 434, + 214, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 214, + 495 + ], + "type": "table", + "html": "
wssDOTAHRSC
E2EFCOSE2EFCOS
0.149.2859.6673.6678.92
1.051.0062.6182.6786.15
3.049.1559.201.301.65
", + "image_path": "dcc7f3e3ed88f1e91f4254d34de293ed1491aa6d9fa4c7fec08cf3d046b70d2f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 227, + 435, + 547, + 495 + ], + "blocks": [ + { + "bbox": [ + 63, + 498, + 212, + 509 + ], + "lines": [ + { + "bbox": [ + 63, + 498, + 212, + 509 + ], + "spans": [ + { + "bbox": [ + 63, + 498, + 212, + 509 + ], + "type": "text", + "content": "Table 6. Ablation with the weight of " + }, + { + "bbox": [ + 63, + 498, + 212, + 509 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{\\mathrm{{ss}}}" + }, + { + "bbox": [ + 63, + 498, + 212, + 509 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 227, + 435, + 547, + 495 + ], + "lines": [ + { + "bbox": [ + 227, + 435, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 227, + 435, + 547, + 495 + ], + "type": "table", + "html": "
R / F / SDOTAHRSCR / F / SDOTAHRSC
90% / 10% / 0%60.4285.4680% / 20% / 0%59.4684.73
75% / 0% / 25%60.7986.2260% / 15% / 25%62.3884.21
68% / 7% / 25%62.6186.1538% / 37% / 25%45.878.56
45% / 5% / 50%60.5585.3440% / 10% / 50%60.4910.74
", + "image_path": "94857a85bd91954170322bd64883aa400374e0ce5f81ceb896c8a33c7519df8f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 245, + 498, + 527, + 509 + ], + "lines": [ + { + "bbox": [ + 245, + 498, + 527, + 509 + ], + "spans": [ + { + "bbox": [ + 245, + 498, + 527, + 509 + ], + "type": "text", + "content": "Table 7. Ablation with the proportion of augmented views in self-supervision." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "text", + "content": "labels on train/val sets, with which the FCOS detector is trained. In this two-stage mode, Point2RBox-v2 achieves an accuracy of " + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "inline_equation", + "content": "62.61\\%" + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "text", + "content": ", considerably surpassing PointOBB series. Remarkably, it even outperforms the SAM-powered method P2RBox by " + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "inline_equation", + "content": "3.57\\%" + }, + { + "bbox": [ + 55, + 520, + 295, + 581 + ], + "type": "text", + "content": " (62.61% vs. 59.04%)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "content": "Class-wise analysis. The FCOS detector trained with labels generated by Point2RBox-v2 achieves accuracy nearly equivalent to RBox-supervised FCOS across six high-density categories: SH (86.9% vs. 87.1%), SV (79.6% vs. 79.8%), LV (76.3% vs. 79.8%), PL (88.0% vs. 89.1%), ST (82.9% vs. 84.6%), and TC (89.1% vs. 90.4%). Interestingly, these six high-density categories account for 88% of DOTA instances. By annotating these categories with points and generating RBoxes using Point2RBox-v2 while labeling the other sparse categories with RBoxes, we can significantly reduce annotation labor without sacrificing much accuracy," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 520, + 520, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 520, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 520, + 533 + ], + "type": "text", + "content": "highlighting the valuable role our method can play." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 540, + 455, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 455, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 455, + 552 + ], + "type": "text", + "content": "4.2. Results on More Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "The results are displayed in Table 2. On more challenging DOTA-v1.5/2.0, Point2RBox-v2 presents a similar trend, " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "23.47\\% / 18.15\\%" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": " higher than PointOBB-v2 in the pseudogeneration track. On the ship detection dataset HRSC, the gap between Point2RBox-v2 and RBox-supervised FCOS is only " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "2.84\\%" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": " (86.15% vs. 88.99%). DIOR is relatively sparse, leading to less improvement with our methods—lower than PointSAM (44.45% vs. 46.20%) but still higher than methods that do not use SAM. Our method also provides competitive performance on fine-grained datasets FAIR1M and STAR. In addition to remote sensing scenarios, we carry out experiments on SKU110K for densely packed retail scenes. Existing point-supervised methods struggle in this" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19289" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 164, + 163 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 164, + 163 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 164, + 163 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 164, + 163 + ], + "type": "image", + "image_path": "916035995ff8381e486e21560479f131cec1896e656181e344006d143384d8eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 165, + 70, + 258, + 163 + ], + "blocks": [ + { + "bbox": [ + 165, + 70, + 258, + 163 + ], + "lines": [ + { + "bbox": [ + 165, + 70, + 258, + 163 + ], + "spans": [ + { + "bbox": [ + 165, + 70, + 258, + 163 + ], + "type": "image", + "image_path": "6649cc04a16fbfe2ca5554605e2cd6478862a2ddaf500faad08f94dc6452c6d2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 166, + 421, + 177 + ], + "lines": [ + { + "bbox": [ + 189, + 166, + 421, + 177 + ], + "spans": [ + { + "bbox": [ + 189, + 166, + 421, + 177 + ], + "type": "text", + "content": "Figure 5. Qualitative analysis on failed cases and overlap cases." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 259, + 70, + 351, + 163 + ], + "blocks": [ + { + "bbox": [ + 259, + 70, + 351, + 163 + ], + "lines": [ + { + "bbox": [ + 259, + 70, + 351, + 163 + ], + "spans": [ + { + "bbox": [ + 259, + 70, + 351, + 163 + ], + "type": "image", + "image_path": "97bfa86eff7ab22c510e4a551eda61f101afbe60d1c2069a12b0ed94dbb6b98f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 353, + 70, + 445, + 163 + ], + "blocks": [ + { + "bbox": [ + 353, + 70, + 445, + 163 + ], + "lines": [ + { + "bbox": [ + 353, + 70, + 445, + 163 + ], + "spans": [ + { + "bbox": [ + 353, + 70, + 445, + 163 + ], + "type": "image", + "image_path": "93b828120c97dda984b2f8b9fe4d63c951428c0cc587a558e49af2fc60c50db1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 447, + 71, + 539, + 163 + ], + "blocks": [ + { + "bbox": [ + 447, + 71, + 539, + 163 + ], + "lines": [ + { + "bbox": [ + 447, + 71, + 539, + 163 + ], + "spans": [ + { + "bbox": [ + 447, + 71, + 539, + 163 + ], + "type": "image", + "image_path": "3853fb30847fbd0e2573b0ef5cd80fb92b4d1ff11649132c646ac14645267b31.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 61, + 190, + 291, + 299 + ], + "blocks": [ + { + "bbox": [ + 61, + 190, + 291, + 299 + ], + "lines": [ + { + "bbox": [ + 61, + 190, + 291, + 299 + ], + "spans": [ + { + "bbox": [ + 61, + 190, + 291, + 299 + ], + "type": "table", + "html": "
ModulesDOTAHRSC
\\( \\mathcal{L}_{\\mathrm{O}} \\)\\( \\mathcal{L}_{\\mathrm{W}} \\)\\( \\mathcal{L}_{\\mathrm{ss}} \\)\\( \\mathcal{L}_{\\mathrm{E}} \\)\\( CP \\)E2EFCOSE2EFCOS
✓0.000.000.000.00
✓✓41.5452.9817.9619.64
✓✓✓46.6454.2618.1022.13
✓✓✓✓49.5561.8878.7983.79
✓✓✓✓48.5859.5620.3524.76
✓✓✓✓38.9448.4411.6414.93
✓✓✓✓47.0855.0519.5821.78
✓✓✓✓✓51.0062.6182.6786.15
", + "image_path": "1216b04349f2f50f8d4dc434cf0e4193bc66f13fc22b1941ee48a0e78643a281.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 61, + 322, + 290, + 352 + ], + "blocks": [ + { + "bbox": [ + 73, + 303, + 277, + 313 + ], + "lines": [ + { + "bbox": [ + 73, + 303, + 277, + 313 + ], + "spans": [ + { + "bbox": [ + 73, + 303, + 277, + 313 + ], + "type": "text", + "content": "Table 8. Ablation with incremental addition of modules." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 322, + 290, + 352 + ], + "lines": [ + { + "bbox": [ + 61, + 322, + 290, + 352 + ], + "spans": [ + { + "bbox": [ + 61, + 322, + 290, + 352 + ], + "type": "table", + "html": "
16K=2432|| 1.2β=1.62.0
50.8751.0048.08|| 48.1451.0051.33
", + "image_path": "fb56a081ffc32266d71f888a9576ba5ceab2d6395293e835e7b89ede5215b5f5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 61, + 373, + 290, + 434 + ], + "blocks": [ + { + "bbox": [ + 62, + 354, + 288, + 365 + ], + "lines": [ + { + "bbox": [ + 62, + 354, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 288, + 365 + ], + "type": "text", + "content": "Table 9. Ablation with " + }, + { + "bbox": [ + 62, + 354, + 288, + 365 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 62, + 354, + 288, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 62, + 354, + 288, + 365 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 62, + 354, + 288, + 365 + ], + "type": "text", + "content": " in edge loss on DOTA (E2E)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 373, + 290, + 434 + ], + "lines": [ + { + "bbox": [ + 61, + 373, + 290, + 434 + ], + "spans": [ + { + "bbox": [ + 61, + 373, + 290, + 434 + ], + "type": "table", + "html": "
σPoint2RBoxPointOBB-v2Point2RBox-v2
DOTAHRSCDOTAHRSCDOTAHRSC
0%40.2779.4044.85-62.6186.15
10%39.6078.8142.30-61.5885.76
30%38.4278.2838.46-60.3185.71
", + "image_path": "1831d5d8524c4a8c175094ebc09bb6277c376a5bcd65d15beedea970f4063599.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 65, + 436, + 285, + 447 + ], + "lines": [ + { + "bbox": [ + 65, + 436, + 285, + 447 + ], + "spans": [ + { + "bbox": [ + 65, + 436, + 285, + 447 + ], + "type": "text", + "content": "Table 10. Ablation with the inaccuracy in point annotations." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "content": "case, whereas Point2RBox-v2 achieves performance on par with HBox-supervised H2RBox (65.64% vs. 57.15%)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 491, + 156, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 156, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 156, + 502 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 509, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 557 + ], + "type": "text", + "content": "Tables 3-10 display the ablation studies on DOTA-v1.0 and HRSC. \"E2E\" denotes end-to-end training; \"FCOS\" denotes two-stage training (i.e. generating pseudo labels to train FCOS). The final values adopted are highlighted in gray." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 557, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 593 + ], + "type": "text", + "content": "Weight of each loss. Tables 3-5 determine the weights of the proposed losses. Based on these experiments, the weights " + }, + { + "bbox": [ + 55, + 557, + 295, + 593 + ], + "type": "inline_equation", + "content": "(w_{\\mathrm{O}}, w_{\\mathrm{W}}, w_{\\mathrm{E}}, w_{\\mathrm{ss}})" + }, + { + "bbox": [ + 55, + 557, + 295, + 593 + ], + "type": "text", + "content": " are set to (10, 5, 0.3, 1)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": "Proportion of augmented views. Table 7 studies the proportion between rotation, flip, and scale. The results are reported with two-stage training (FCOS). Based on the results, the proportion is set to " + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "inline_equation", + "content": "68\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "Incremental addition of modules. Table 8 demonstrates the constraints from Gaussian and Voronoi achieve an accuracy of " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "52.98\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " on DOTA. Adding consistency loss and edge loss further boosts it to " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "54.26\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "61.88\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": ", respectively, whereas the improvement from copy-paste is " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.73\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": ". We also demonstrate the impact of omitting each core loss." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "text", + "content": "Edge loss parameters. We set " + }, + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "inline_equation", + "content": "K = 24" + }, + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\beta = 1.6" + }, + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "text", + "content": " as they are observed to discern the correct edges during code development. Table 9 provides a more precise ablation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": "Annotation inaccuracy. We offset the annotated points by a noise from the uniform distribution " + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "inline_equation", + "content": "[- \\sigma H, + \\sigma H]" + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": " is the height of objects. Table 10 shows that the " + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": " of Point2RBox-v2 decreases by less than " + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": " when noise is added to point annotations, demonstrating the robustness of the proposed learning mechanisms." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 308, + 419, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 419, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 419, + 319 + ], + "type": "text", + "content": "4.4. More Discussions" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 326, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 410 + ], + "type": "text", + "content": "The qualitative analysis on the failed/overlap cases is shown in Fig. 5. 1) Failed cases. Although our method performs well overall, it struggles with certain categories that are sparse and not constrained by other objects. 2) Overlap cases. Minimizing overlap as a soft constraint during training does not entirely eliminate overlap. Once trained, the model remains robust to some overlap during inference." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 422, + 388, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 388, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 388, + 434 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 441, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 536 + ], + "type": "text", + "content": "This paper introduces Point2RBox-v2, a point-supervised oriented object detector that effectively leverages the arrangement and layout of instances. We propose the integration of Gaussian overlay and Voronoi tessellation to constrain the size and rotation of instances based on their spatial relationships. Additionally, by incorporating self-supervised consistency loss, edge loss, and copy-paste augmentation, the accuracy of the model is further enhanced." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "type": "text", + "content": "Experiments yield the following observations: 1) The integration of Gaussian and Voronoi concepts effectively harnesses the spatial layout of objects, significantly enhancing point-supervised OOD. 2) Point2RBox-v2 demonstrates exceptional performance in densely packed scenes (see Fig. 2), where existing methods struggle. 3) Our method does not require priors (i.e. pre-trained SAM or one-shot examples) and is applicable to both end-to-end and pseudo-generation modes. 4) It advances the state of the art by a large amount, achieving " + }, + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "type": "inline_equation", + "content": "62.61\\%" + }, + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "type": "inline_equation", + "content": "86.15\\%" + }, + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "type": "inline_equation", + "content": "34.71\\%" + }, + { + "bbox": [ + 313, + 537, + 556, + 669 + ], + "type": "text", + "content": " on the DOTA-v1.0, HRSC, and FAIR1M datasets, respectively." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 669, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 556, + 713 + ], + "type": "text", + "content": "Limitations. The gap between Point2RBox-v2 and the RBox-supervised OOD is still huge in terms of sparse categories (i.e. BR/SBF) since little constraint can be obtained from the layout between them when the objects are sparse." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19290" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 114, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 114, + 297, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 297, + 147 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 297, + 147 + ], + "type": "text", + "content": "[1] Franz Aurenhammer. Voronoi diagrams—a survey of a fundamental geometric data structure. ACM Computing Surveys, 23(3):345-405, 1991. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 191 + ], + "type": "text", + "content": "[2] Guangming Cao, Xuehui Yu, Wenwen Yu, Xumeng Han, Xue Yang, Guorong Li, Jianbin Jiao, and Zhenjun Han. P2rbox: Point prompt oriented object detection with SAM. arXiv preprint arXiv:2311.13128, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 296, + 247 + ], + "type": "text", + "content": "[3] Pengfei Chen, Xuehui Yu, Xumeng Han, Najmul Hassan, Kai Wang, Jiachen Li, Jian Zhao, Humphrey Shi, Zhenjun Han, and Qixiang Ye. Point-to-box network for accurate object detection via single point supervision. In European Conference on Computer Vision, 2022. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 291 + ], + "type": "text", + "content": "[4] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 292, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 295, + 335 + ], + "type": "text", + "content": "[5] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2849-2858, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 296, + 402 + ], + "type": "text", + "content": "[6] Di Feng, Christian Haase-Schütz, Lars Rosenbaum, Heinz Hertlein, Claudius Glaser, Fabian Timm, Werner Wiesbeck, and Klaus Dietmayer. Deep multi-modal object detection and semantic segmentation for autonomous driving: Datasets, methods, and challenges. IEEE Transactions on Intelligent Transportation Systems, 22(3):1341-1360, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 296, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 457 + ], + "type": "text", + "content": "[7] Kun Fu, Zhonghan Chang, Yue Zhang, Guangluan Xu, Keshu Zhang, and Xian Sun. Rotation-aware and multi-scale convolutional neural network for object detection in remote sensing images. ISPRS Journal of Photogrammetry and Remote Sensing, 161:294-308, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 296, + 512 + ], + "type": "text", + "content": "[8] Shuyong Gao, Wei Zhang, Yan Wang, Qianyu Guo, Chenglong Zhang, Yangji He, and Wenqiang Zhang. Weakly-supervised salient object detection using point supervision. In AAAI Conference on Artificial Intelligence, pages 670-678, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 568 + ], + "type": "text", + "content": "[9] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D. Cubuk, Quoc V. Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 569, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 613 + ], + "type": "text", + "content": "[10] Eran Goldman, Roei Herzig, Aviv Eisenschat, Jacob Goldberger, and Tal Hassner. Precise detection in densely packed scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5227-5236, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 657 + ], + "type": "text", + "content": "[11] Jiaming Han, Jian Ding, Nan Xue, and Gui-Song Xia. Redet: A rotation-equivariant detector for aerial object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2785-2794, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "text", + "content": "[12] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Con-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 74, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 555, + 95 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "type": "text", + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In IEEE International Conference on Computer Vision, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 555, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 555, + 163 + ], + "type": "text", + "content": "[15] Liping Hou, Ke Lu, Xue Yang, Yuqiu Li, and Jian Xue. Grep: Gaussian representation for arbitrary-oriented object detection. Remote Sensing, 15(3):757, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 164, + 555, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 555, + 208 + ], + "type": "text", + "content": "[16] Javed Iqbal, Muhammad Akhtar Munir, Arif Mahmood, Afsheen Rafaqat Ali, and Mohsen Ali. Leveraging orientation for weakly supervised object detection with application to firearm localization. Neurocomputing, 440:310-320, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 209, + 555, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 555, + 263 + ], + "type": "text", + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 265, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 555, + 308 + ], + "type": "text", + "content": "[18] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS Journal of Photogrammetry and Remote Sensing, 159:296-307, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 310, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 555, + 354 + ], + "type": "text", + "content": "[19] Qingyun Li, Yushi Chen, Xinya Shu, Dong Chen, Xin He, Yi Yu, and Xue Yang. A simple aerial detection baseline of multimodal language models. arXiv preprint arXiv:2501.09720, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 555, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 399 + ], + "type": "text", + "content": "[20] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented leptopoints for aerial object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1829-1838, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 401, + 555, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 401, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 401, + 555, + 444 + ], + "type": "text", + "content": "[21] Wentong Li, Yuqian Yuan, Song Wang, Jianke Zhu, Jianshu Li, Jian Liu, and Lei Zhang. Point2mask: Point-supervised panoptic segmentation via optimal transport. In IEEE International Conference on Computer Vision, 2023. 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 555, + 521 + ], + "type": "text", + "content": "[22] Yansheng Li, Linlin Wang, Tingzhu Wang, Xue Yang, Junwei Luo, Qi Wang, Youming Deng, Wenbin Wang, Xian Sun, Haifeng Li, Bo Dang, Yongjun Zhang, Yi Yu, and Junchi Yan. Star: A first-ever dataset and a large-scale benchmark for scene graph generation in large-size satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(3):1832-1849, 2025. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "type": "text", + "content": "[23] Minghui Liao, Zhen Zhu, Baoguang Shi, Gui-Song Xia, and Xiang Bai. Rotation-sensitive regression for oriented scene text detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5909-5918, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 555, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 555, + 622 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2117-2125, 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 669, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 555, + 713 + ], + "type": "text", + "content": "[26] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International Journal of Computer Vision, 128(2):261-318, 2020. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19291" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "text", + "content": "[27] Nanqing Liu, Xun Xu, Yongyi Su, Haojie Zhang, and Heng-Chao Li. Pointsam: Pointly-supervised segment anything model for remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 63:1-15, 2025. 1, 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "text", + "content": "[28] Xuebo Liu, Ding Liang, Shi Yan, Dagui Chen, Yu Qiao, and Junjie Yan. Fots: Fast oriented text spotting with a unified network. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5676-5685, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 215 + ], + "type": "text", + "content": "[29] Yuekai Liu, Hongli Gao, Liang Guo, Aoping Qin, Canyu Cai, and Zhichao You. A data-flow oriented deep ensemble learning method for real-time surface defect inspection. IEEE Transactions on Instrumentation and Measurement, 69(7): 4681-4691, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 270 + ], + "type": "text", + "content": "[30] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International Conference on Pattern Recognition Applications and Methods, pages 324-331, 2017. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 296, + 304 + ], + "type": "text", + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 296, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 296, + 349 + ], + "type": "text", + "content": "[32] Junyan Lu, Qinglei Hu, Ruifei Zhu, Yali Wei, and Tie Li. Afws: Angle-free weakly-supervised rotating object detection for remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 392 + ], + "type": "text", + "content": "[33] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. Pointobb: Learning oriented object detection via single point supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 296, + 437 + ], + "type": "text", + "content": "[34] Jeffri Murragarra-Llerena, Lucas N Kirsten, Luis Felipe Zeni, and Claudio R Jung. Probabilistic intersection-over-union for training and evaluation of oriented object detectors. IEEE Transactions on Image Processing, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 502 + ], + "type": "text", + "content": "[35] Xingjia Pan, Yuqiang Ren, Kekai Sheng, Weiming Dong, Haolei Yuan, Xiaowei Guo, Chongyang Ma, and Changsheng Xu. Dynamic refinement network for oriented and densely packed object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11207-11216, 2020. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 296, + 601 + ], + "type": "text", + "content": "[36] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024-8035, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 296, + 645 + ], + "type": "text", + "content": "[37] Wen Qian, Xue Yang, Silong Peng, Junchi Yan, and Yue Guo. Learning modulated loss for rotated object detection. In AAAI Conference on Artificial Intelligence, pages 2458-2466, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "type": "text", + "content": "[38] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. Pointobb-v2: Towards simpler, faster, and stronger single point supervised oriented object detection. In International Conference on Learning Representations, 2025. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[39] Xavier Soria, Yachuan Li, Mohammad Rouhani, and Angel D. Sappa. Tiny and efficient model for the edge detection" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "generalization. In IEEE/CVF International Conference on Computer Vision Workshops, pages 1364-1373, 2023. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 171 + ], + "type": "text", + "content": "[40] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, Martin Weinmann, Stefan Hinz, Cheng Wang, and Kun Fu. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "type": "text", + "content": "[41] Yongqing Sun, Jie Ran, Feng Yang, Chenqiang Gao, Takayuki Kurozumi, Hideaki Kimata, and Ziqi Ye. Oriented object detection for remote sensing images based on weakly supervised learning. In IEEE International Conference on Multimedia & Expo Workshops, pages 1-6, 2021. 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 555, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 273 + ], + "type": "text", + "content": "[42] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In IEEE/CVF International Conference on Computer Vision, pages 9626-9635, 2019. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 555, + 319 + ], + "type": "text", + "content": "[43] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 320, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 555, + 363 + ], + "type": "text", + "content": "[44] L. Vincent and P. Soille. Watersheds in digital spaces: an efficient algorithm based on immersion simulations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(6):583-598, 1991. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 420 + ], + "type": "text", + "content": "[45] Linfei Wang, Yibing Zhan, Xu Lin, Baosheng Yu, Liang Ding, Jianqing Zhu, and Dapeng Tao. Explicit and implicit box equivariance learning for weakly-supervised rotated object detection. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 421, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 555, + 464 + ], + "type": "text", + "content": "[46] Long Wen, Yu Cheng, Yi Fang, and Xinyu Li. A comprehensive survey of oriented object detection in remote sensing images. Expert Systems with Applications, page 119960, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 466, + 555, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 555, + 509 + ], + "type": "text", + "content": "[47] Hongjin Wu, Ruoshan Lei, and Yibing Peng. Pcbnet: A lightweight convolutional neural network for defect inspection in surface mount technology. IEEE Transactions on Instrumentation and Measurement, 71:1-14, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 510, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 555, + 566 + ], + "type": "text", + "content": "[48] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2018. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 567, + 555, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 555, + 610 + ], + "type": "text", + "content": "[49] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In IEEE/CVF International Conference on Computer Vision, pages 3520-3529, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 612, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 555, + 645 + ], + "type": "text", + "content": "[50] Xue Yang and Junchi Yan. Arbitrary-oriented object detection with circular smooth label. In European Conference on Computer Vision, pages 677-694, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[51] Xue Yang and Junchi Yan. On the arbitrary-oriented object detection: Classification based approaches revisited. International Journal of Computer Vision, 130(5):1340-1365, 2022. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[52] Xue Yang, Hao Sun, Kun Fu, Jirui Yang, Xian Sun, Menglong Yan, and Zhi Guo. Automatic ship detection in remote" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19292" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 106 + ], + "type": "text", + "content": "sensing images from google earth of complex scenes based on multiscale rotation dense feature pyramid networks. Remote sensing, 10(1):132, 2018. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "type": "text", + "content": "[53] Xue Yang, Jirui Yang, Junchi Yan, Yue Zhang, Tengfei Zhang, Zhi Guo, Xian Sun, and Kun Fu. Scdet: Towards more robust detection for small, cluttered and rotated objects. In IEEE/CVF International Conference on Computer Vision, pages 8231-8240, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 296, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 296, + 209 + ], + "type": "text", + "content": "[54] Xue Yang, Liping Hou, Yue Zhou, Wentao Wang, and Junchi Yan. Dense label encoding for boundary discontinuity free rotation detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15814-15824, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 211, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 296, + 255 + ], + "type": "text", + "content": "[55] Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3det: Refined single-stage detector with feature refinement for rotating object. In AAAI Conference on Artificial Intelligence, pages 3163-3171, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 258, + 296, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 296, + 311 + ], + "type": "text", + "content": "[56] Xue Yang, Junchi Yan, Ming Qi, Wentao Wang, Xiaopeng Zhang, and Tian Qi. Rethinking rotated object detection with gaussian Wasserstein distance loss. In 38th International Conference on Machine Learning, pages 11830-11841, 2021. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 315, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 296, + 369 + ], + "type": "text", + "content": "[57] Xue Yang, Xiaojiang Yang, Jirui Yang, Qi Ming, Wentao Wang, Qi Tian, and Junchi Yan. Learning high-precision bounding box for rotated object detection via kullback-leibler divergence. In Advances in Neural Information Processing Systems, pages 18381-18394, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 296, + 426 + ], + "type": "text", + "content": "[58] Xue Yang, Junchi Yan, Wenlong Liao, Xiaokang Yang, Jin Tang, and Tao He. Scdet++: Detecting small, cluttered and rotated objects via instance-level feature denoising and rotation loss smoothing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2384-2399, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 428, + 296, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 296, + 472 + ], + "type": "text", + "content": "[59] Xue Yang, Gefan Zhang, Wentong Li, Xuehui Wang, Yue Zhou, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. International Conference on Learning Representations, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 296, + 529 + ], + "type": "text", + "content": "[60] Xue Yang, Gefan Zhang, Xiaojiang Yang, Yue Zhou, Wentao Wang, Jin Tang, Tao He, and Junchi Yan. Detecting rotated objects as gaussian distributions and its 3-d generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4335-4354, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 532, + 296, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 296, + 574 + ], + "type": "text", + "content": "[61] Xue Yang, Yue Zhou, Gefan Zhang, Jirui Yang, Wentao Wang, Junchi Yan, Xiaopeng Zhang, and Qi Tian. The kfiou loss for rotated object detection. In International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 578, + 296, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 296, + 620 + ], + "type": "text", + "content": "[62] Ze Yang, Shaohui Liu, Han Hu, Liwei Wang, and Stephen Lin. Reppoints: Point set representation for object detection. In IEEE/CVF International Conference on Computer Vision, pages 9656-9665, 2019. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 666 + ], + "type": "text", + "content": "[63] Yi Yu and Feipeng Da. Phase-shifting coder: Predicting accurate orientation in oriented object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[64] Yi Yu and Feipeng Da. On boundary discontinuity in angle regression based arbitrary oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-14, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 586 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "type": "text", + "content": "[65] Yi Yu, Xue Yang, Qingyun Li, Yue Zhou, Feipeng Da, and Junchi Yan. H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. In Advances in Neural Information Processing Systems, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "text", + "content": "[66] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2rbox: Combine knowledge from synthetic visual patterns for end-to-end oriented object detection with single point supervision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "type": "text", + "content": "[67] Yi Yu, Xue Yang, Yansheng Li, Zhenjun Han, Feipeng Da, and Junchi Yan. Wholly-wood: Wholly leveraging diversified-quality labels for weakly-supervised oriented object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 295 + ], + "type": "text", + "content": "[68] Peiyuan Zhang, Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Yue Zhou, Xiaosong Jia, Xudong Lu, Jingdong Chen, Xiang Li, et al. Pointobb-v3: Expanding performance boundaries of single point-supervised oriented object detection. arXiv preprint arXiv:2501.13898, 2025. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 555, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 340 + ], + "type": "text", + "content": "[69] Shun Zhang, Jihui Long, Yaohui Xu, and Shaohui Mei. Pmho: Point-supervised oriented object detection based on segmentation-driven proposal generation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 555, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 555, + 384 + ], + "type": "text", + "content": "[70] Xin Zhang, Xue Yang, Yuxuan Li, Jian Yang, Ming-Ming Cheng, and Xiang Li. Rsar: Restricted state angle resolver and rotated sar benchmark. arXiv preprint arXiv:2501.04440, 2025.6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 386, + 555, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 555, + 429 + ], + "type": "text", + "content": "[71] Zhong-Qiu Zhao, Peng Zheng, Shou-Tao Xu, and Xindong Wu. Object detection with deep learning: A review. IEEE Transactions on Neural Networks and Learning Systems, 30 (11):3212-3232, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 431, + 555, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 555, + 474 + ], + "type": "text", + "content": "[72] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: An efficient and accurate scene text detector. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2017. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 475, + 555, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 555, + 529 + ], + "type": "text", + "content": "[73] Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, et al. MMrotate: A rotated object detection benchmark using pytorch. In 30th ACM International Conference on Multimedia, pages 7331-7334, 2022. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 532, + 555, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 555, + 586 + ], + "type": "text", + "content": "[74] Tianyu Zhu, Bryce Ferenczi, Pulak Purkait, Tom Drummond, Hamid Rezatofighi, and Anton van den Hengel. Knowledge combination to learn rotated detection without rotated annotation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19293" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/d2184bbd-78d6-4278-aed5-2e0864e5914c_content_list.json b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/d2184bbd-78d6-4278-aed5-2e0864e5914c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d4833f7406db52156d96519d7756f8e235d4b72a --- /dev/null +++ b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/d2184bbd-78d6-4278-aed5-2e0864e5914c_content_list.json @@ -0,0 +1,1673 @@ +[ + { + "type": "text", + "text": "PointLoRA: Low-Rank Adaptation with Token Selection for Point Cloud Learning", + "text_level": 1, + "bbox": [ + 205, + 130, + 790, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Song Wang $^{1,2}$ , Xiaolu Liu $^{1}$ , Lingdong Kong $^{2}$ , Jianyun Xu $^{3*}$ , Chunyong Hu $^{3}$ , Gongfan Fang $^{2}$ , Wentong Li $^{4}$ , Jianke Zhu $^{1\\dagger}$ , Xinchao Wang $^{2\\dagger}$ $^{1}$ ZJU NUS AD Lab, CaiNiao, Alibaba NUAA \n{songw, jkzhu}@zju.edu.cn, xinchao@nus.edu.sg", + "bbox": [ + 169, + 202, + 825, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised representation learning for point cloud has demonstrated effectiveness in improving pre-trained model performance across diverse tasks. However, as pretrained models grow in complexity, fully fine-tuning them for downstream applications demands substantial computational and storage resources. Parameter-efficient fine-tuning (PEFT) methods offer a promising solution to mitigate these resource requirements, yet most current approaches rely on complex adapter and prompt mechanisms that increase tunable parameters. In this paper, we propose PointLoRA, a simple yet effective method that combines low-rank adaptation (LoRA) with multi-scale token selection to efficiently fine-tune point cloud models. Our approach embeds LoRA layers within the most parameter-intensive components of point cloud transformers, reducing the need for tunable parameters while enhancing global feature capture. Additionally, multi-scale token selection extracts critical local information to serve as prompts for downstream fine-tuning, effectively complementing the global context captured by LoRA. The experimental results across various pre-trained models and three challenging public datasets demonstrate that our approach achieves competitive performance with only $3.43\\%$ of the trainable parameters, making it highly effective for resource-constrained applications. Source code is available at: https://github.com/songw-zju/PointLoRA.", + "bbox": [ + 88, + 340, + 485, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D point cloud learning plays a vital role in computer vision, advancing the understanding and reconstruction of complex 3D scenes [4, 12, 22, 30, 62]. Training deep neural networks on point clouds presents unique challenges due to their unordered structure, sparsity, and irregularity. In recent", + "bbox": [ + 89, + 787, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "years, point-based methods [29, 33, 41, 42, 45, 64] have made significant progress in addressing these challenges.", + "bbox": [ + 511, + 310, + 903, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As the volume of available point cloud data grows, there has been a surge in interest toward pre-training on unlabeled point clouds to learn generalizable representations [36, 43, 47, 64, 69, 76]. Robust pre-trained models enable efficient fine-tuning for downstream tasks, reducing dependency on labeled data, accelerating convergence, and improving accuracy [23, 66]. However, conventional full fine-tuning can disrupt pre-trained knowledge, potentially diminishing the model's generalization capability. Moreover, full fine-tuning requires storing multiple versions of model weights, leading to substantial storage costs as datasets and tasks expand, along with increased computational demands for larger pre-trained models.", + "bbox": [ + 509, + 340, + 906, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies [51, 72, 75, 77] have made strides in introducing parameter-efficient fine-tuning (PEFT) methods to pre-trained point cloud models, building on approaches widely used in natural language processing (NLP) and 2D vision [9, 65]. As an illustration, IDPT [72] utilizes instance-aware dynamic prompt tuning to enhance model robustness in downstream transfer learning tasks. The subsequent methods [51, 77] combine prompt tuning with adapter tuning, while relying on carefully crafted adapters and specialized prompt design. PPT [75] presents positional prompt tuning, which doubles the sequence length and significantly increases computational requirements.", + "bbox": [ + 509, + 537, + 908, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To fine-tune in a more efficient manner and further reduce the parameter count, we propose PointLoRA, which leverages low-rank adaptation (LoRA) [20], widely studied in large language models, to enable better fine-tuning for point cloud models. As shown in Fig. 1 (a), during LoRA fine-tuning, instead of directly updating the pretrained weights, two auxiliary low-rank matrices, $W_{u}$ and $W_{d}$ , are injected to update the network with fewer tunable parameters. From a 3D learning perspective, these matrices function as fully connected layers that effectively capture global point cloud features but lack local information extraction, akin to the architecture of PointNet [41].", + "bbox": [ + 509, + 719, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Project leader.", + "bbox": [ + 107, + 875, + 194, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 109, + 887, + 240, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6605", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/90c33258060e29642edcfebf82cc9ffa7711557dc1a3df46201a7a87d168ab3b.jpg", + "image_caption": [ + "(a) LoRA" + ], + "image_footnote": [], + "bbox": [ + 127, + 108, + 436, + 281 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/af3e24347766390db475aaeba5053e394a678020ab4a48b74d885df6826b364c.jpg", + "image_caption": [ + "(b) PointLoRA", + "Figure 1. Comparing our proposed PointLoRA approach against vanilla LoRA methods. Both LoRA and our approach incorporate low-rank adaptation matrices into the pre-trained weights to extract global information from the point cloud sequence. Furthermore, our approach integrates tokens selected at various scales to capture local information, which is refined using a shared Prompt MLP and then output alongside the results derived from the original low-rank matrices." + ], + "image_footnote": [], + "bbox": [ + 465, + 109, + 869, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike 2D visual or language data, 3D point clouds contain distinctive local geometric features critical for downstream tasks. Additionally, not all local positions contribute equally to the model, as their relevance varies depending on the specific task and data characteristics. To address this, we further design a Multi-Scale Token Selection module, providing local prior information as a prompt to complement LoRA's global feature aggregation capabilities. Using multi-scale sampling and $k$ -nearest neighbor aggregation, we derive tokens that represent features across different scales. A shared Mask Predictor then selects a subset of these tokens to act as prompts. During fine-tuning, tokens from various positions are dynamically selected, processed through a Prompt MLP, and incorporated into the LoRA layer, as illustrated in Fig. 1 (b). This design enables the integration of essential local geometric features with global context, allowing the pre-trained model to adapt effectively to diverse downstream tasks.", + "bbox": [ + 91, + 348, + 483, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows.", + "bbox": [ + 109, + 645, + 449, + 660 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose PointLoRA, a simple yet effective scheme for parameter-efficient fine-tuning in point cloud learning.", + "- Low-rank adaptation is injected into the most parameter-intensive components of the point cloud transformer to capture global information, with a corresponding analysis in the context of PointNet provided.", + "- We further design a multi-scale token selection module that dynamically incorporates local geometric information as prompt, enriching global low-rank adaptation with the crucial local context.", + "- Extensive experiments across three challenging datasets and two widely used pre-trained models validate the effectiveness of our method, achieving state-of-the-art performance with only $3.43\\%$ of the trainable parameters." + ], + "bbox": [ + 89, + 686, + 483, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 345, + 655, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Pre-Training on Point Clouds. Self-supervised pre-training for 3D point clouds has been widely explored [36, 43, 47, 55, 56, 64, 69, 76], generally categorized into contrastive-based approaches [1, 7, 39, 64, 66] and masked signal reconstruction methods [5, 6, 40, 69, 74]. In contrastive learning, PointContrast [64] and CrossPoint [1] leverage different views or instances to extract latent features through contrasting representations. SegContrast [39] extends this approach to outdoor LiDAR data, extracting class-agnostic segments from augmented views [24]. Reconstruction-based methods, such as Point-BERT [69] and Point-MAE [40], employ masked point cloud reconstruction for pre-training. In ACT [10], a pre-trained autoencoder serves as a cross-modality teacher, guiding 3D point cloud learning through knowledge distillation [17, 57, 58]. PointGPT [5] adapts the GPT framework [38] for autoregressive generation on point clouds. Additionally, Qi et al. [43, 44] propose a combination of reconstruction and cross-modal contrastive learning within generative models. While full fine-tuning is commonly applied to these pretrained models for downstream tasks, the high computational costs and potential dilution of pre-trained knowledge motivate our exploration of efficient fine-tuning strategies.", + "bbox": [ + 511, + 371, + 906, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Parameter-Efficient Fine-Tuning (PEFT). PEFT methods have gained traction in NLP and 2D vision tasks [8, 9, 19, 21, 28, 48, 63, 73] to enhance downstream performance with minimal tunable parameters. Mainstream PEFT approaches include prompt and prefix tuning, adapter-based methods, and low-rank adaptation. Prompt and prefix tuning [21, 25, 28, 48] introduce additional tokens as tunable components, with VPT [21] pioneering the use of prompt tokens for pre-trained vision transformers (ViTs) [11, 26, 49, 67]. Adapter-based tuning inserts trainable modules between frozen layers [8, 19, 27, 50], while LoRA [20] applies a low-rank approximation to update linear layers within", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6606", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attention blocks. Building on these techniques, recent innovations explore alternative adapter placements [14, 31], tuning bias terms [71], and combining with Mixture-of-Experts (MoE) techniques [35, 70]. Recent studies have also adapted PEFT to point cloud analysis [32, 51, 72, 75, 77]. IDPT [72] extends prompt tuning to DGCNN [59] for instance-aware prompt extraction. Point-PEFT [51] and DAPT [77] combine prompt tuning with adapters to further improve fine-tuning performance, while PPT [75] introduces positional prompt tuning for efficient 3D representation learning. In contrast to these approaches, our work uniquely integrates low-rank adaptation with multi-scale token selection, achieving competitive performance with a significantly reduced parameter footprint.", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 314, + 217, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we revisit the architecture of point cloud transformer and the corresponding fine-tuning prototypes.", + "bbox": [ + 89, + 339, + 483, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Point Cloud Transformer", + "text_level": 1, + "bbox": [ + 89, + 377, + 321, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Tokenizer. Due to the sparsity and irregularity of point clouds, it is necessary to transform them into a token sequence suitable for processing by the transformer. We follow standard settings [40, 69] by segmenting the point cloud into irregular patches using the farthest point sampling (FPS) and the $k$ -nearest neighbors ( $k$ -NN) algorithm. Formally, given the point cloud $P = \\{p_1, p_2, \\ldots, p_N\\} \\in \\mathbb{R}^{N \\times 3}$ with $N$ points, the $g$ group centers are obtained with FPS:", + "bbox": [ + 89, + 398, + 483, + 532 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC _ {g} = \\operatorname {F P S} (P), \\quad C _ {g} \\in \\mathbb {R} ^ {g \\times 3}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 534, + 480, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C_g$ is the selected center points. Then, we use $k$ -NN to select $k$ nearest neighbor points for each center in $C_g$ :", + "bbox": [ + 89, + 556, + 482, + 588 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nN _ {p} = k \\text {- N N} (P, C _ {g}), \\quad N _ {p} \\in \\mathbb {R} ^ {g \\times k \\times 3}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 592, + 480, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N_{p}$ is the corresponding neighbor point patch. Notably, these local point patches are made unbiased by subtracting their center points, which promotes better convergence. Then a mini-PointNet [41] is utilized to embed the point patches into discrete tokens $T_{p}$ :", + "bbox": [ + 89, + 618, + 483, + 694 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {p} = \\operatorname {m i n i - P o i n t N e t} \\left(N _ {p}\\right), \\quad T _ {p} \\in \\mathbb {R} ^ {g \\times d}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 700, + 480, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d$ is the embedding dimension. Following the tokenization process, the 3D point cloud is converted into a feature vector, enabling subsequent processing with Transformer architectures similar to those used in natural language processing and 2D vision.", + "bbox": [ + 89, + 724, + 483, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer Block. With a classification token $T_{\\mathrm{cls}}$ and the obtained $T_{p}$ , the input $T_{\\mathrm{input}} = \\mathrm{Concat}(T_{\\mathrm{cls}}, T_{p})$ is further processed by $L$ -layer transformer blocks. In each block, we first project $T_{\\mathrm{input}}$ into query, key, and value spaces:", + "bbox": [ + 89, + 801, + 483, + 863 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = T _ {\\text {i n p u t}} \\mathbf {W} ^ {Q}, \\quad \\mathbf {K} = T _ {\\text {i n p u t}} \\mathbf {W} ^ {K}, \\quad \\mathbf {V} = T _ {\\text {i n p u t}} \\mathbf {W} ^ {V}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 867, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{W}^Q, \\mathbf{W}^K, \\mathbf{W}^V \\in \\mathbb{R}^{d \\times d}$ are trainable matrices in qkv projection layer. The attention layer then computes the self-attention with a skip connection as follows:", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {\\text {o u t p u t}} ^ {\\prime} = \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) + T _ {\\text {i n p u t}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 151, + 903, + 169 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, the final output $T_{\\mathrm{output}}$ is computed as:", + "bbox": [ + 513, + 172, + 807, + 188 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {\\text {o u t p u t}} = \\text {F F N} \\left(\\ln \\left(T _ {\\text {o u t p u t}} ^ {\\prime}\\right)\\right) + T _ {\\text {o u t p u t}} ^ {\\prime}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 196, + 903, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where LN is the layer norm function [2] and FFN is the feed-forward network layer [3]. Through $L$ Transformer blocks, the network extracts deep and essential features from the discrete tokens. Notably, the qkv projection parameters within the attention layer, along with those in the FFN layer constitute a significant portion of the model's parameters and are crucial for performance.", + "bbox": [ + 511, + 222, + 905, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Fine-Tuning Prototype", + "text_level": 1, + "bbox": [ + 511, + 335, + 723, + 352 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Leveraging well-pretrained models, fine-tuning prototypes to adapt these models to diverse downstream tasks can be categorized as follows.", + "bbox": [ + 511, + 358, + 905, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Full Fine-Tuning. Full fine-tuning is the predominant approach in current point cloud pre-training research for downstream tasks, involving the direct loading of pretrained model parameters and joint training of the encoder alongside the task-specific head. Although this method offers a theoretically higher accuracy potential, it also entails substantial computational and storage overhead.", + "bbox": [ + 511, + 402, + 905, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameter-Efficient Fine-Tuning. Parameter-efficient fine-tuning adapts pre-trained models to downstream tasks by freezing most of the model's parameters, introducing lightweight modules, and selectively fine-tuning a small subset of parameters. This approach provides a more practical solution, particularly as model sizes continue to increase. Existing methods typically combine adapter tuning, which inserts bottleneck-like layers, and prompt tuning, which introduces prior knowledge, to fine-tune point cloud pre-trained models.", + "bbox": [ + 511, + 508, + 906, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. PointLoRA", + "text_level": 1, + "bbox": [ + 513, + 672, + 633, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 696, + 624, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we aim to provide a simple but effective parameter-efficient fine-tuning method for point cloud pretrained models. The complete framework of our proposed PointLoRA, integrated into the existing point cloud transformer, is illustrated in Fig. 2, addressing the parameter-intensive nature of the transformer-based network design. Our approach can effectively enhance downstream task performance while keeping most of the pre-trained model parameters frozen.", + "bbox": [ + 511, + 719, + 905, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following section, we first outline the establishment of the vanilla LoRA baseline and analyze the factors contributing to its effectiveness in point cloud learning.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6607", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fb5a7328ee85be8410f9f0e2ce0bba6660df8b6873aeed6c30dbc5f1f4e18103.jpg", + "image_caption": [ + "Figure 2. Overview of PointLoRA integrated into point cloud transformer pipeline. Given an input point cloud, we first tokenize it using the original Point Tokenizer and perform token selection across multiple scales (Multi-Scale Token Selection). The tokens from both components are then concatenated and fed into the Transformer Block. Our approach is injected into the qkv projection and FFN layers, utilizing a shared Prompt MLP within these layers to enhance parameter efficiency." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we present how our method extends LoRA's feature extraction capabilities to achieve competitive performance in 3D scenarios.", + "bbox": [ + 89, + 378, + 482, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Vanilla LoRA Baseline", + "text_level": 1, + "bbox": [ + 89, + 431, + 302, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned in Sec. 3.1, the qkv projection and feed-forward network (FFN) layers in point cloud transformer significantly increase the number of parameters, limiting the fine-tuning efficiency and hindering deployment in resource-limited environments. To address this issue, we incorporate Low-Rank Adaptation (LoRA) into these layers by introducing low-rank matrices to adapt pre-trained weights, thereby reducing the number of trainable parameters while preserving model performance. Given a pretrained weight matrix $W_{p}$ , LoRA modifies it as follows:", + "bbox": [ + 89, + 455, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {\\text {u p d a t e}} = W _ {p} + \\Delta W = W _ {p} + W _ {u} \\cdot W _ {d}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 617, + 482, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta W$ and $W_{\\mathrm{update}}$ denotes the updated weight and new weight matrix, respectively. $W_{u}\\in \\mathbb{R}^{d\\times r}$ and $W_{d}\\in \\mathbb{R}^{r\\times d}$ are low-rank matrices, where $r\\ll d$ indicates the rank, controlling the complexity of the adaptation.", + "bbox": [ + 89, + 643, + 482, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, only $W_{u}$ and $W_{d}$ are updated, while the pre-trained weight $W_{p}$ remains frozen. For inference, the adaptation $\\Delta W$ can be merged with the original weights, yielding in a single consolidated weight matrix: $W_{\\mathrm{infer}} = W_{p} + \\Delta W$ . This consolidation enables the model to retain the benefits of low-rank adaptation while avoiding additional computational overhead in the inference phase.", + "bbox": [ + 89, + 704, + 482, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analysis. LoRA's architecture is particularly well-suited for point cloud data, as it can capture the complex relationships between points essential to understanding 3D shapes and spatial arrangements. Its MLP-like structure (i.e., two low-rank matrices) effectively learns global features while aligning with PointNet's principles of handling unordered point", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sets and extracting permutation-invariant features. The synergy between LoRA and PointNet's feature extraction capabilities establishes a robust framework for harnessing the rich information in point cloud data, enhancing both generalization and adaptability in complex environments.", + "bbox": [ + 511, + 378, + 906, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. PointLoRA with Token Selection", + "text_level": 1, + "bbox": [ + 511, + 467, + 802, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While global features provide a comprehensive understanding of the overall structure of the point cloud, local features capture intricate details vital for accurate downstream task performance. To fully investigate LoRA's capabilities, it is crucial to complement the extracted global features with robust local features. The tokenization process produces a sequence of tokens that capture various aspects of the point cloud data. However, not all tokens are valuable for downstream tasks; some may encapsulate redundant or noninformative content, offering limited contribution to overall model performance. This necessitates a systematic approach for filtering and selecting the most informative tokens to enhance the model's representation of local features.", + "bbox": [ + 511, + 489, + 906, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To tackle the above concerns, we introduce a MultiScale Token Selection module integrated with vanilla LoRA, which is presented in Fig. 3. This module dynamically selects tokens from the raw point cloud at multiple scales, allowing the model to capture local features at different levels of detail. Additionally, PointLoRA incorporates a shared Prompt MLP to further embed local information from these selected tokens, enhancing the vanilla LoRA's performance.", + "bbox": [ + 511, + 688, + 905, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Scale Token Generation. Specifically, we apply farthest point sampling (FPS) on the input point cloud $P$ with varying numbers of centroids (i.e., $N_{1}, N_{2}, \\dots, N_{M}$ ) to generate tokens at $M$ different scales. As described in Sec. 3, we can obtain $M$ sets of center points, $C_{g}^{1}, C_{g}^{2}, \\dots, C_{g}^{M}$ ,", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6608", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/306fb50b4823f49e3de30b28f55d5505b0a9a2dc3bd1d02dfe76fb5a056738d5.jpg", + "image_caption": [ + "Figure 3. Illustration of Multi-Scale Token Selection. In a two-scale setup, we first sample different numbers of center points, then cluster around each center point and apply Mini-PointNet encoding to generate the corresponding tokens. These tokens are also fed into a Mask Predictor to estimate importance scores, allowing us to select the Top-K tokens at each scale." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 467, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "each containing $g_{1}, g_{2}, \\ldots, g_{M}$ center points. Subsequently, $k$ -NN is employed to retrieve the corresponding neighbor point patches, which are then embedded into discrete tokens using a shared Mini-PointNet.", + "bbox": [ + 89, + 492, + 483, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mask Predictor for Selection. The above process yields $M$ token sets, $T_{p}^{1}, T_{p}^{2}, \\ldots, T_{p}^{M}$ , encoding local information at multiple scales. Our goal is to select tokens from these sets that are most beneficial for fine-tuning on downstream tasks. Therefore, a simple but effective Mask Predictor is designed to estimate the importance of each token, producing a score vector $s^{m} \\in \\mathbb{R}^{g_{m} \\times 1}$ for each token in $T_{p}^{m} \\in \\mathbb{R}^{g_{m} \\times d}$ , $m = 1, \\ldots, M$ . In particular, the Mask Predictor comprises two multi-layer perceptron (MLP) layers followed by a Sigmoid activation function:", + "bbox": [ + 89, + 553, + 483, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {m} = \\operatorname {S i g m o i d} \\left(\\operatorname {M L P} \\left(\\left(T _ {p} ^ {m}\\right)\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 714, + 480, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This design enables the model to assign scores between 0 and 1 to each token, indicating their relevance for downstream tasks. We then select a predefined number $N_{1}^{\\prime}, \\ldots, N_{M}^{\\prime}$ of tokens from each scale based on their importance scores using Top-K selection, resulting in a refined token set $\\{S_{\\mathrm{p}}^{m}\\}_{m=1}^{M}$ :", + "bbox": [ + 89, + 741, + 483, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\mathrm {p}} ^ {m} = \\operatorname {T o p K} \\left(T _ {p} ^ {m}, N _ {m} ^ {\\prime}\\right), m = 1, \\dots , M. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 843, + 482, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Local Geometry Prompt. With the selected $N_{s} = N_{1}^{\\prime} + \\ldots + N_{M}^{\\prime}$ tokens $S_{p}$ , we can capture crucial local information", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for downstream tasks. These tokens are then concatenated with the original inputs to form a local geometry prompt. In each LoRA layer of the point cloud transformer block, a Prompt MLP with a GELU activation function [16] encodes the new input features alongside the local prompt. This process integrates the encoded features of the selected tokens with the LoRA adaptation output. The resulting combined output can be formulated as follows:", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO _ {\\text {u p d a t e}} = \\operatorname {P r o m p t} \\mathrm {M L P} \\left(T _ {\\text {i n p u t}}, S _ {p}\\right) + \\Delta W \\cdot \\left(T _ {\\text {i n p u t}}, S _ {p}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 223, + 903, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $O_{\\mathrm{update}}$ is the updated amount of the PointLoRA layer. Notably, to further optimize the efficiency of the parameters, the Prompt MLP is configured independently for both the qkv projection and FFN layers, while being shared across all blocks.", + "bbox": [ + 511, + 255, + 906, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The multi-scale token selection process improves model performance by prioritizing the most relevant local features during fine-tuning. By incorporating this approach, our scheme effectively refines both global and local feature representations, thereby enhancing its adaptability and accuracy across a range of 3D tasks.", + "bbox": [ + 511, + 330, + 905, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Training & Inference", + "text_level": 1, + "bbox": [ + 511, + 431, + 714, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall Loss Function. When fine-tuning on diverse downstream tasks across various datasets, we follow the common practice of maintaining consistency between the adopted loss and the loss function of the original task. Additionally, we introduce a regularization term to further supervise the Mask Predictor. In general, the total loss function $\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_{\\mathrm{task}} + \\lambda \\cdot \\mathcal{L}_{\\mathrm{mask}}$ , where $\\mathcal{L}_{\\mathrm{task}}$ is the task-specific loss for classification or segmentation. $\\mathcal{L}_{\\mathrm{mask}}$ is the regularization loss for mask predictor with the balanced weight $\\lambda$ :", + "bbox": [ + 511, + 454, + 906, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = - \\frac {1}{N _ {\\text {t o t a l}}} \\sum_ {i = 1} ^ {N _ {\\text {t o t a l}}} \\left(s _ {i} \\log \\left(s _ {i} + \\epsilon\\right) + \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 599, + 903, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\left(1 - s _ {i}\\right) \\log \\left(1 - s _ {i} + \\epsilon\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 633, + 797, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{\\mathrm{total}} = N_1 + N_2 \\dots + N_M$ , $s_i$ denotes the importance score for each token and $\\epsilon$ is a small constant (e.g., $10^{-6}$ ) to prevent undefined values in the logarithm.", + "bbox": [ + 511, + 657, + 905, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. During inference, the additional parameters introduced by LoRA in the transformer block are directly merged into the original pre-trained weights, with the only newly added component being a small shared MLP dedicated to embed the selected tokens.", + "bbox": [ + 511, + 703, + 905, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 792, + 645, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 511, + 818, + 594, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We validate the effectiveness of our approach through extensive experiments on three widely used 3D datasets: ScanObjectNN [52], ModelNet40 [61], and ShapeNetPart [68]. ScanObjectNN [52] is a challenging", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6609", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/72eea66754c48c0d9ef383e5c4dd9e50d9d4f4ee65648a6c73eb643a50995ff3.jpg", + "table_caption": [ + "Table 1. Performance comparison on three variants of the ScanObjectNN [52] and the ModelNet40 [61] datasets, respectively, for real-world and synthetic object classification. Both the number of tunable parameters and overall accuracy (OA) are reported. All methods only employ the default data argumentation without voting as the baseline. * denotes results reproduced from the public source code." + ], + "table_footnote": [], + "table_body": "
MethodsPublicationTunable Params.ScanObjectNNModelNet40
OBJ-BGOBJ-ONLYPB-T50-RSPoints Num.OA (%)
Traditional Supervised Learning Only
PointNet [41]CVPR'173.5 M73.379.268.01k89.2
PointNet++ [42]NeurIPS'171.5 M82.384.377.91k90.7
DGCNN [59]TOG'191.8 M82.886.278.11k92.9
MVTN [13]ICCV'2111.2 M--82.81k93.8
PointNeXt [45]NeurIPS'221.4 M--87.71k94.0
PointMLP [37]ICLR'2213.2 M--85.41k94.5
RepSurf-U [46]CVPR'221.5 M--84.31k94.4
ADS [18]ICCV'23---87.51k95.1
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2122.1 M84.8585.5478.791k92.1
Point-BERT [69]CVPR'2222.1 M87.4388.1283.071k93.2
MaskPoint [34]ECCV'2222.1 M89.7089.3084.601k93.8
Point-MAE [40]ECCV'2222.1 M90.0288.2985.181k93.8
Point-M2AE [74]NeurIPS 2215.3 M91.2288.8186.431k94.0
ACT [10]ICLR'2322.1 M93.2991.9188.211k93.7
RECON [43]ICML'2343.6 M94.1593.1289.731k93.9
Self-Supervised Representation Learning (Parameter-Efficient Fine-Tuning)
Point-MAE [40] (Full-FT)ECCV'2222.1 M (100%)90.0288.2985.181k93.2
Point-MAE + IDPT [72]ICCV'231.7 M (7.69%)91.22(+1.20)90.02(+1.73)84.94(-0.24)1k93.3(+0.1)
Point-MAE + DAPT [77]CVPR'241.1 M (4.97%)90.88(+0.86)90.19(+1.90)85.08(-0.10)1k93.5(+0.3)
Point-MAE + PPT* [75]arXiv'241.04 M (4.57%)89.84(-0.18)88.98(+0.69)84.45(-0.73)1k93.2(+0.0)
Point-MAE + PointLoRAOurs0.77 M (3.43%)90.71(+0.69)89.33(+1.04)85.53(+0.35)1k93.3(+0.1)
", + "bbox": [ + 91, + 135, + 906, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "real-world 3D object classification dataset, which contains approximately 15,000 indoor point cloud instances across 15 categories. The object classification task is performed on three variants of increasing complexity, OBJ-BG, OBJ-ONLY, and PB-T50-RS, each representing different levels of scene realism and occlusion. ModelNet40 [61], a classical synthetic dataset for 3D object recognition, consists of 12,311 meshed 3D CAD objects that span 40 categories. We perform synthetic object classification and few-shot learning experiments on ModelNet40 to evaluate the robustness of the model in both standard and low-resource settings. Additionally, to evaluate performance on detailed structures and component segmentation, we provide experiments on ShapeNetPart [68], a widely used benchmark for part segmentation, comprising 16,881 point-level synthetic objects across 16 object categories and 50 part categories.", + "bbox": [ + 88, + 507, + 480, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. In the evaluation, overall accuracy (OA) is adopted to assess performance on the 3D object classification task, representing the ratio of correctly classified instances to the total number of instances, thus providing an aggregate score over all classes. For the part segmentation task, we employ mean Intersection over Union (mIoU) to evaluate the overlap between prediction and ground-truth segments, averaged across all classes.", + "bbox": [ + 88, + 750, + 480, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our method can be directly integrated into existing point cloud pre-trained models. Dur", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing fine-tuning, we follow a commonly used setup, freezing most parameters of the pre-trained model, and primarily updating the newly inserted parameters. The number of scales, $M$ , is set to 2. At these two scales, we apply farthest point sampling (FPS) and $k$ -NN clustering with (128, 32) and (64, 64) center and neighboring points, respectively, and then select tokens with $N_1' = 32$ and $N_2' = 8$ . The balanced weight $\\lambda$ for Mask Predictor learning is set to 0.004 For real-world and synthetic object classification, we fine-tune using Point-MAE [40], training for 300 epochs with a learning rate of $5e^{-4}$ and a weight decay of 0.05. For few-shot learning and part segmentation, we utilize the Recon [43] model for fine-tuning to further validate the generalizability of our approach, with a learning rate of $2e^{-4}$ for part segmentation. All experiments are conducted on a single GPU. More implementation details with different baselines are provided in the Supplementary Material.", + "bbox": [ + 511, + 507, + 906, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparative Study", + "text_level": 1, + "bbox": [ + 511, + 772, + 700, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-World & Synthetic Object Classification. We conduct real-world and synthetic object classification on three variants of ScanObjectNN (OBJ-BG, OBJ-ONLY, PB-T50-RS) [52] and ModelNet40 [61], using default data augmentation and without voting. As illustrated in Tab. 1, our proposed approach demonstrates consistent performance improvements, especially on the most challenging vari", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6610", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/118a6bffe5490a0c36f67841301bbe781d4ea0774d2749af33ec0314523b6425.jpg", + "image_caption": [ + "Figure 4. The t-SNE visualization results on the PB-T50-RS split of the ScanObjectNN dataset [52] with different fine-tuning schemes. We adopt Point-MAE [40] as the baseline model for fair comparison. TP: Number of tunable parameters. OA: Overall accuracy. Symbol * denotes re-produced with official implementation. Best viewed in colors and zoomed-in for additional details." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 259, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/488cca78a4feae37af758eef1855e3ffccafa3efa9fa104c185128d89469deb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 88, + 413, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5321543229616d63dff5024f18e68f96b69c510fedb3e209af31322c40973f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 88, + 570, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eab0465e9f56b2288f8d2a66c7db0e7ac7fdfd0f995ce7c7220c582629baf689.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 88, + 725, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c974cb58e42fde1a0170015bfd386e71c0217df8d6239f6f3fd197dac07d4f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 88, + 880, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a0d0e042e9440da5d1a68e0841f58d3da408edbbfb9a51475895bfe5ea75fc46.jpg", + "table_caption": [ + "Table 2. Performance comparison on ModelNet40 [61] for few-shot learning. We report the scores of the overall accuracy $(\\%)\\pm$ the standard deviation $(\\%)$ without voting. The top two highest accuracies are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsPublication5-way10-way
10-shot20-shot10-shot20-shot
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2194.0±3.695.9±2.389.4±5.192.4±4.6
Point-BERT [69]CVPR'2294.6±3.196.3±2.791.0±5.492.7±5.1
MaskPoint [34]ECCV'2295.0±3.797.2±1.791.4±4.093.4±3.5
Point-MAE [40]ECCV'2296.3±2.597.8±1.892.6±4.195.0±3.0
Point-M2AE [74]NeurIPS'2296.8±1.898.3±1.492.3±4.595.0±3.0
ACT [10]ICLR'2396.8±2.398.0±1.493.3±4.095.6±2.8
Self-Supervised Representation Learning (Efficient Fine-Tuning)
RECON [43] (Full-FT)ICML'2397.3±1.998.9±3.993.3±3.995.8±3.0
RECON+ IDPT [72]ICCV'2396.9±2.498.3±0.792.8±4.095.5±3.2
RECON+ DAPT [77]CVPR'2495.6±2.897.7±1.691.9±4.194.6±3.5
RECON+ PPT [75]arXiv'2497.0±2.798.7±1.692.2±5.095.6±2.9
RECON+ PointLoRAOurs96.9±2.798.8±1.292.7±4.495.8±2.9
", + "bbox": [ + 93, + 351, + 480, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ant, PB-T50-RS, where it is the only parameter-efficient fine-tuning method that surpasses the full fine-tuned PointMAE [40] (85.53% vs. 85.18%) and significantly outperforms the state-of-the-art method, DAPT [77]. Notably, our approach requires only 0.77M tunable parameters, the fewest among all methods. Furthermore, the t-SNE feature visualization [53] of existing methods and PointLoRA is presented in Fig. 4. Our method enables the pre-trained model to fine-tune with a minimal number of parameters, producing distinctive feature representations.", + "bbox": [ + 88, + 580, + 482, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Few-shot Learning. We also perform few-shot learning experiments on ModelNet40 [61] to validate the transfer learning capability of our method with limited annotations. Following standard protocols [15, 43], we evaluate the fine-tuning performance with ReCon [43] in the 5-way / 10-way and 10-shot / 20-shot settings, respectively. As shown in Tab. 2, our method achieves the best or second-best accuracy in all configurations, further demonstrating the generalizability of the proposed scheme.", + "bbox": [ + 88, + 732, + 482, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Point Cloud Part Segmentation. The part segmentation results on the ShapeNetPart dataset [68] are provided in", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9ce3f9ee25c1446336a8053a916fc10938b863490a76bdb13dec682a8297f5e6.jpg", + "table_caption": [ + "Table 3. Performance comparison on the ShapeNetPart [68] for part segmentation. Both the mIoU for all classes (Cls.) and instances (Inst.) are provided. TP indicates the number of tunable parameters. $\\dagger$ means the result, reported from DAPT [77]. $*$ denotes the results reproduced from the official implementation." + ], + "table_footnote": [], + "table_body": "
MethodsPublicationTPCls. mIoU (%)Inst. mIoU (%)
Traditional Supervised Learning Only
PointNet [41]CVPR'17-80.3983.7
PointNet++ [42]NeurIPS'17-81.8585.1
DGCNN [59]TOG'19-82.3385.2
APES [60]CVPR'23-83.6785.8
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2127.09 M83.4285.1
MaskPoint [34]ECCV'22-84.6086.0
Point-BERT [69]CVPR'2227.09 M84.1185.6
Point-MAE [40]ECCV'2227.06 M84.1986.1
ACT [10]ICLR'2327.06 M84.6686.1
Self-Supervised Representation Learning (Efficient Fine-Tuning)
RECON [43] (Full-FT)ICML'2327.06 M84.5286.1
RECON+ IDPT† [72]ICCV'235.69 M83.6685.7
RECON+ DAPT [77]CVPR'245.65 M83.8785.7
RECON+ PPT* [75]arXiv'245.62 M83.8885.4
RECON+ PointLoRAOurs5.63 M83.9885.4
", + "bbox": [ + 514, + 364, + 898, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3. In this fine-grained scene understanding task, our approach still achieves competitive performance with a constrained parameter budget. Unlike prior classification models, the increase in the parameter count is primarily attributed to the segmentation head.", + "bbox": [ + 511, + 626, + 906, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Other PEFT Methods. We further present a comparison with existing PEFT methods designed for other tasks, including NLP and 2D vision. We select the most challenging PB-T50-RS variant [52] and use the pretrained Point-MAE model [40] as a baseline. As shown in Tab. 4, these methods designed for other tasks provide limited performance gains when fine-tuned in 3D scenes, despite their parameter efficiency compared to existing 3D fine-tuning approaches. In contrast, our method not only achieves the highest accuracy, but also uses significantly fewer parameters than the three current fine-tuning techniques specifically developed for point clouds, underscoring the strong potential of PointLoRA.", + "bbox": [ + 511, + 703, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6611", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e9fbd8f9be3f15648903c7266f5280b1f605a0e35d020dc55e862dcf627f4058.jpg", + "table_caption": [ + "Table 4. Performance comparison with other parameter-efficient methods designed for NLP and 2D Vision tasks on the hardest variant of ScanObjectNN [52]." + ], + "table_footnote": [], + "table_body": "
MethodsPublicationTPPB-T50-RS
Point-MAE [34]ECCV'2222.1 M85.18
Linear probing-0.3 M75.99
+ Adapter [19]ICML'190.9 M83.93
+ Perfix tuning [28]ACL'210.7 M77.72
+ BitFit [71]ACL'210.3 M82.62
+ LoRA†[20]ICLR'220.9 M81.74
+ VPT-Deep [21]ECCV'220.4 M81.09
+ AdaptFormer [8]NeurIPS'220.9 M83.45
+ SSF [31]NeurIPS'220.4 M82.58
+ IDPT [72]ICCV'231.7 M84.94
+ DAPT [77]CVPR'241.1 M85.08
+ PPT* [75]arXiv'241.04 M84.45
+ PointLoRAOurs0.77 M85.53
", + "bbox": [ + 109, + 135, + 459, + 323 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e191a770868277c554a9bceaac916571b8b3fee78133837b5159c3e1bd4dadd8.jpg", + "table_caption": [ + "Table 5. The impact of each module of our scheme. We provide the overall accuracy (\\%) on the hardest variant of ScanObjectNN [52] and corresponding tunable parameters (TP). \"MS-FPS\" indicates multi-scale furthest point sampling." + ], + "table_footnote": [], + "table_body": "
LoRAToken SelectionMS-FPSTPPB-T50-RS
Full Fine-Tuning22.1 M85.18
Linear Probing0.27 M75.99
✓0.53 M83.83
✓✓0.77 M84.91
✓✓✓0.77 M85.53
", + "bbox": [ + 107, + 386, + 460, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 484, + 243, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we perform exhaustive ablation experiments on the challenging PB-T50-RS variant to investigate the rationale and effectiveness of the design choices of our proposed approach. The pre-trained Point-MAE [40] is adopted as a baseline for a fair comparison.", + "bbox": [ + 89, + 507, + 482, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on PointLoRA Scheme. Firstly, we provide the ablations on each module of the proposed method. As illustrated in Tab. 5, directly applying low-rank adaptation (LoRA) produces a significant improvement over linear probing, demonstrating the effectiveness of incorporating global information during the fine-tuning process. Injecting selected tokens as prompts into LoRA further improves accuracy with only a marginal increase in the number of parameters. Finally, the local information obtained from selected tokens across multiple scales complements the global features of LoRA, achieving the best performance.", + "bbox": [ + 89, + 583, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Low-Rank Adaptation. Then we perform ablation experiments on rank $(r)$ in our approach, as shown in Tab. 6. Different ranks have a noticeable impact on finetuning performance. Taking into account both parameter count and accuracy, we set the rank to 8 to maximize the extraction of global information for fine-tuning.", + "bbox": [ + 89, + 750, + 482, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Multi-Scale Token Selection. Given the critical improvement from multi-scale token selection for PointLoRA, we first perform ablation experiments on the number of tokens selected at the same scale. As shown in", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/52161e44b9a30864adbe9dba75b1856e90bda544454909dbd9b825a204b14afc.jpg", + "table_caption": [ + "Table 6. Ablation study on rank $(r)$ of the proposed method." + ], + "table_footnote": [], + "table_body": "
Rank rTPRatioPB-T50-RS
40.66 M2.96%84.28
80.77 M3.43%85.53
160.99 M4.37%85.15
321.44 M6.32%84.87
", + "bbox": [ + 555, + 107, + 859, + 183 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/32cd0e67a7f45ce0cfbde672ecc05044d9f6fab72ad64e082784ab6fcc44856c.jpg", + "table_caption": [ + "Table 7. Ablation study on scale number $M$ and selected token number $N_{s}$ in Multi-Scale Token Selection." + ], + "table_footnote": [], + "table_body": "
Scale MToken Num. NsPB-T50-RS
16484.80
13284.91
11685.22
264 & 1684.39
232 & 885.53
216 & 484.49
332 & 16 & 885.05
", + "bbox": [ + 555, + 217, + 859, + 338 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7a9cb13abd0f87a10f05c435879a5a3668d0b117d99ea5b6547bf3b92b2ef097.jpg", + "table_caption": [ + "Table 8. Ablation study on the dimension of Prompt MLP." + ], + "table_footnote": [], + "table_body": "
DimensionTPRatioPB-T50-RS
80.68 M3.03%83.55
160.71 M3.17%84.77
320.77 M3.43%85.53
640.90 M3.96%85.25
", + "bbox": [ + 555, + 359, + 857, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the upper part of Tab. 7, selecting 16 tokens at the original scale yields the best results, demonstrating the effectiveness of token selection. Furthermore, when incorporating tokens from multiple scales, optimal performance is achieved by selecting 32 and 8 tokens from two different scales, as shown in the lower part of Tab. 7.", + "bbox": [ + 511, + 452, + 906, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Prompt MLP. We provide the ablation study on the dimension of the shared Prompt MLP, as this affects the embedding of local information in our method. As illustrated in Tab. 8, setting this dimension to 32 yields the best fine-tuning performance. Additionally, adjusting the dimension causes only minimal changes in tunable parameter counts, suggesting flexibility in selecting this hyperparameter for different datasets and tasks.", + "bbox": [ + 511, + 542, + 908, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, a simple yet effective parameter-efficient finetuning approach named PointLoRA, is presented with low-rank adaptation and multi-scale token selection for point cloud models. Low-rank adaptation efficiently reduces the parameter-heavy components of the point cloud transformer architecture while capturing global information. Multi-scale token selection effectively encodes essential local features as prompt, further enhancing the finetuning process. The integration of global and local information enables our approach to achieve state-of-the-art results on the most challenging dataset with a minimal number of tunable parameters while also delivering competitive performance across other datasets and models.", + "bbox": [ + 511, + 703, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6612", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the National Natural Science Foundation of China under Grant No. 62376244, in part by the Zhejiang Provincial Natural Science Foundation of China under Grant No. LD24F030001, by the Information Technology Center and State Key Lab of CAD&CG, Zhejiang University, and by the National Research Foundation, Singapore, and Cyber Security Agency of Singapore under its National Cybersecurity R&D Programme and CyberSG R&D Cyber Research Programme Office (Award: CRPO-GC1-NTU-002). Lingdong Kong is supported by the Apple Scholars in AI/ML Ph.D. Fellowship program.", + "bbox": [ + 89, + 114, + 485, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 294, + 187, + 311 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022.", + "[2] Jimmy Lei Ba. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.", + "[3] George Bebis and Michael Georgiopoulos. Feed-forward neural networks. IEEE Potentials, 13(4):27-31, 1994.", + "[4] Saifullahi Aminu Bello, Shangshu Yu, Cheng Wang, Jibril Muhammad Adam, and Jonathan Li. Deep learning on 3d point clouds. Remote Sensing, 12(11):1729, 2020.", + "[5] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. In Advances in Neural Information Processing Systems, 2024.", + "[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023.", + "[7] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023.", + "[8] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In Advances in Neural Information Processing Systems, 2022.", + "[9] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023.", + "[10] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In International Conference on Learning Representations, 2022." + ], + "bbox": [ + 93, + 319, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021.", + "[12] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020.", + "[13] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In IEEE/CVF International Conference on Computer Vision, pages 1-11, 2021.", + "[14] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations, 2021.", + "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[16] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016.", + "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015.", + "[18] Cheng-Yao Hong, Yu-Ying Chou, and Tyng-Luh Liu. Attention discriminant sampling for point clouds. In IEEE/CVF International Conference on Computer Vision, pages 14429-14440, 2023.", + "[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning. PMLR, 2019.", + "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021.", + "[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, 2022.", + "[22] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023.", + "[23] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023.", + "[24] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6613", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Empirical Methods in Natural Language Processing, pages 3045-3059, 2021.", + "[26] Bonan Li, Yinhan Hu, Xuecheng Nie, Congying Han, Xi-angjian Jiang, Tiande Guo, and Luoqi Liu. Dropkey for vision transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22700–22709, 2023.", + "[27] Minglei Li, Peng Ye, Yongqi Huang, Lin Zhang, Tao Chen, Tong He, Jiayuan Fan, and Wanli Ouyang. Adapter-x: A novel general parameter-efficient fine-tuning framework for vision. arXiv preprint arXiv:2406.03051, 2024.", + "[28] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Annual Meeting of the Association for Computational Linguistics, 2021.", + "[29] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In Advances in Neural Information Processing Systems, 2018.", + "[30] Ye Li, Lingdong Kong, Hanjiang Hu, Xiaohao Xu, and Xiaonan Huang. Is your lidar placement optimized for 3d scene understanding? In Advances in Neural Information Processing Systems, pages 34980-35017, 2024.", + "[31] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In Advances in Neural Information Processing Systems, 2022.", + "[32] Dingkang Liang, Tianrui Feng, Xin Zhou, Yumeng Zhang, Zhikang Zou, and Xiang Bai. Parameter-efficient fine-tuning in spectral domain for point cloud learning. arXiv preprint arXiv:2410.08114, 2024.", + "[33] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024.", + "[34] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, 2022.", + "[35] Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Derong Xu, Feng Tian, and Yefeng Zheng. Moelora: An moe-based parameter efficient fine-tuning method for multitask medical applications. arXiv preprint arXiv:2310.18339, 2023.", + "[36] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. In Advances in Neural Information Processing Systems, pages 37193-37229, 2023.", + "[37] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2022.", + "[38] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020.", + "[39] Lucas Nunes, Rodrigo Marcuzzi, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Segcontrast: 3d point cloud" + ], + "bbox": [ + 91, + 92, + 483, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "feature representation learning through self-supervised segment discrimination. IEEE Robotics and Automation Letters, 7(2):2116-2123, 2022.", + "[40] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European Conference on Computer Vision, 2022.", + "[41] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017.", + "[42] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems, 2017.", + "[43] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In International Conference on Machine Learning. PMLR, 2023.", + "[44] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. *Shapellm: Universal 3d object understanding for embodied interaction*. arXiv preprint arXiv:2402.17766, 2024.", + "[45] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In Advances in Neural Information Processing Systems, pages 23192-23204, 2022.", + "[46] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[47] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022.", + "[48] Zhengxiang Shi and Aldo Lipani. Dept: Decomposed prompt tuning for parameter-efficient fine-tuning. In International Conference on Learning Representations, 2024.", + "[49]Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8791-8800, 2022.", + "[50] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In Advances in Neural Information Processing Systems, pages 12991-13005, 2022.", + "[51] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. In AAAI Conference on Artificial Intelligence, pages 5171-5179, 2024.", + "[52] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, 2019." + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6614", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 2008.", + "[54] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In IEEE/CVF International Conference on Computer Vision, 2021.", + "[55] Song Wang, Jianke Zhu, and Ruixiang Zhang. Metarangeseg: Lidar sequence semantic segmentation using multiple feature aggregation. IEEE Robotics and Automation Letters, 7(4):9739-9746, 2022.", + "[56] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5186-5195, 2023.", + "[57] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14792-14801, 2024.", + "[58] Song Wang, Jiawei Yu, Wentong Li, Hao Shi, Kailun Yang, Junbo Chen, and Jianke Zhu. Label-efficient semantic scene completion with scribble annotations. In International Joint Conference on Artificial Intelligence, pages 1398-1406, 2024.", + "[59] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics, 38:1-12, 2019.", + "[60] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.", + "[61] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1912-1920, 2015.", + "[62] Aoran Xiao, Xiaoqin Zhang, Ling Shao, and Shijian Lu. A survey of label-efficient deep learning for 3d point clouds. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024.", + "[63] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video superresolution. In European Conference on Computer Vision, pages 217-235. Springer, 2024.", + "[64] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In European Conference on Computer Vision, 2020.", + "[65] Yi Xin, Siqi Luo, Haodi Zhou, Junlong Du, Xiaohong Liu, Yue Fan, Qing Li, and Yuntao Du. Parameter-efficient fine-tuning for pre-trained vision models: A survey. arXiv preprint arXiv:2402.02242, 2024.", + "[66] Xiang Xu, Lingdong Kong, Hui Shuai, Wenwei Zhang, Liang Pan, Kai Chen, Ziwei Liu, and Qingshan Liu. 4d" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "contrastive superflows are dense 3d representation learners. In European Conference on Computer Vision, pages 58-80, 2024.", + "[67] Xingyi Yang and Xinchao Wang. Kolmogorov-arnold transformer. In International Conference on Learning Representations, 2025.", + "[68] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics, 35(6):1-12, 2016.", + "[69] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[70] Ted Zadouri, Ahmet Üstün, Arash Ahmadian, Beyza Ermis, Acyr Locatelli, and Sara Hooker. Pushing mixture of experts to the limit: Extremely parameter efficient moe for instruction tuning. arXiv preprint arXiv:2309.05444, 2023.", + "[71] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Annual Meeting of the Association for Computational Linguistics, 2022.", + "[72] Yaohua Zha, Jinpeng Wang, Tao Dai, Bin Chen, Zhi Wang, and Shu-Tao Xia. Instance-aware dynamic prompt tuning for pre-trained point cloud models. In IEEE/CVF International Conference on Computer Vision, 2023.", + "[73] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient fin-tuning. In International Conference on Learning Representations, 2023.", + "[74] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. In Advances in Neural Information Processing Systems, 2022.", + "[75] Shaochen Zhang, Zekun Qi, Runpei Dong, Xiuxiu Bai, and Xing Wei. Positional prompt tuning for efficient 3d representation learning. arXiv preprint arXiv:2408.11567, 2024.", + "[76] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024.", + "[77] Xin Zhou, Dingkang Liang, Wei Xu, Xingkui Zhu, Yihan Xu, Zhikang Zou, and Xiang Bai. Dynamic adapter meets prompt tuning: Parameter-efficient transfer learning for point cloud analysis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14707-14717, 2024." + ], + "bbox": [ + 516, + 92, + 903, + 811 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6615", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/d2184bbd-78d6-4278-aed5-2e0864e5914c_model.json b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/d2184bbd-78d6-4278-aed5-2e0864e5914c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..565776877114b684a0b2957a7232e8cbac59b35e --- /dev/null +++ b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/d2184bbd-78d6-4278-aed5-2e0864e5914c_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.207, + 0.131, + 0.791, + 0.176 + ], + "angle": 0, + "content": "PointLoRA: Low-Rank Adaptation with Token Selection for Point Cloud Learning" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.203, + 0.826, + 0.278 + ], + "angle": 0, + "content": "Song Wang\\(^{1,2}\\), Xiaolu Liu\\(^{1}\\), Lingdong Kong\\(^{2}\\), Jianyun Xu\\(^{3*}\\), Chunyong Hu\\(^{3}\\), Gongfan Fang\\(^{2}\\), Wentong Li\\(^{4}\\), Jianke Zhu\\(^{1\\dagger}\\), Xinchao Wang\\(^{2\\dagger}\\) \n\\(^{1}\\)ZJU NUS AD Lab, CaiNiao, Alibaba NUAA \n{songw, jkzhu}@zju.edu.cn, xinchao@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.486, + 0.735 + ], + "angle": 0, + "content": "Self-supervised representation learning for point cloud has demonstrated effectiveness in improving pre-trained model performance across diverse tasks. However, as pretrained models grow in complexity, fully fine-tuning them for downstream applications demands substantial computational and storage resources. Parameter-efficient fine-tuning (PEFT) methods offer a promising solution to mitigate these resource requirements, yet most current approaches rely on complex adapter and prompt mechanisms that increase tunable parameters. In this paper, we propose PointLoRA, a simple yet effective method that combines low-rank adaptation (LoRA) with multi-scale token selection to efficiently fine-tune point cloud models. Our approach embeds LoRA layers within the most parameter-intensive components of point cloud transformers, reducing the need for tunable parameters while enhancing global feature capture. Additionally, multi-scale token selection extracts critical local information to serve as prompts for downstream fine-tuning, effectively complementing the global context captured by LoRA. The experimental results across various pre-trained models and three challenging public datasets demonstrate that our approach achieves competitive performance with only \\(3.43\\%\\) of the trainable parameters, making it highly effective for resource-constrained applications. Source code is available at: https://github.com/songw-zju/PointLoRA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.484, + 0.866 + ], + "angle": 0, + "content": "3D point cloud learning plays a vital role in computer vision, advancing the understanding and reconstruction of complex 3D scenes [4, 12, 22, 30, 62]. Training deep neural networks on point clouds presents unique challenges due to their unordered structure, sparsity, and irregularity. In recent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.905, + 0.341 + ], + "angle": 0, + "content": "years, point-based methods [29, 33, 41, 42, 45, 64] have made significant progress in addressing these challenges." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.342, + 0.907, + 0.538 + ], + "angle": 0, + "content": "As the volume of available point cloud data grows, there has been a surge in interest toward pre-training on unlabeled point clouds to learn generalizable representations [36, 43, 47, 64, 69, 76]. Robust pre-trained models enable efficient fine-tuning for downstream tasks, reducing dependency on labeled data, accelerating convergence, and improving accuracy [23, 66]. However, conventional full fine-tuning can disrupt pre-trained knowledge, potentially diminishing the model's generalization capability. Moreover, full fine-tuning requires storing multiple versions of model weights, leading to substantial storage costs as datasets and tasks expand, along with increased computational demands for larger pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.538, + 0.909, + 0.719 + ], + "angle": 0, + "content": "Recent studies [51, 72, 75, 77] have made strides in introducing parameter-efficient fine-tuning (PEFT) methods to pre-trained point cloud models, building on approaches widely used in natural language processing (NLP) and 2D vision [9, 65]. As an illustration, IDPT [72] utilizes instance-aware dynamic prompt tuning to enhance model robustness in downstream transfer learning tasks. The subsequent methods [51, 77] combine prompt tuning with adapter tuning, while relying on carefully crafted adapters and specialized prompt design. PPT [75] presents positional prompt tuning, which doubles the sequence length and significantly increases computational requirements." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To fine-tune in a more efficient manner and further reduce the parameter count, we propose PointLoRA, which leverages low-rank adaptation (LoRA) [20], widely studied in large language models, to enable better fine-tuning for point cloud models. As shown in Fig. 1 (a), during LoRA fine-tuning, instead of directly updating the pretrained weights, two auxiliary low-rank matrices, \\( W_{u} \\) and \\( W_{d} \\), are injected to update the network with fewer tunable parameters. From a 3D learning perspective, these matrices function as fully connected layers that effectively capture global point cloud features but lack local information extraction, akin to the architecture of PointNet [41]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.195, + 0.888 + ], + "angle": 0, + "content": "* Project leader." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.241, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6605" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.09, + 0.184, + 0.105 + ], + "angle": 0, + "content": "(a) LoRA" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.109, + 0.437, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.09, + 0.551, + 0.105 + ], + "angle": 0, + "content": "(b) PointLoRA" + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.11, + 0.87, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.284, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 1. Comparing our proposed PointLoRA approach against vanilla LoRA methods. Both LoRA and our approach incorporate low-rank adaptation matrices into the pre-trained weights to extract global information from the point cloud sequence. Furthermore, our approach integrates tokens selected at various scales to capture local information, which is refined using a shared Prompt MLP and then output alongside the results derived from the original low-rank matrices." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.349, + 0.485, + 0.62 + ], + "angle": 0, + "content": "Unlike 2D visual or language data, 3D point clouds contain distinctive local geometric features critical for downstream tasks. Additionally, not all local positions contribute equally to the model, as their relevance varies depending on the specific task and data characteristics. To address this, we further design a Multi-Scale Token Selection module, providing local prior information as a prompt to complement LoRA's global feature aggregation capabilities. Using multi-scale sampling and \\( k \\)-nearest neighbor aggregation, we derive tokens that represent features across different scales. A shared Mask Predictor then selects a subset of these tokens to act as prompts. During fine-tuning, tokens from various positions are dynamically selected, processed through a Prompt MLP, and incorporated into the LoRA layer, as illustrated in Fig. 1 (b). This design enables the integration of essential local geometric features with global context, allowing the pre-trained model to adapt effectively to diverse downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.646, + 0.45, + 0.661 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.687, + 0.483, + 0.717 + ], + "angle": 0, + "content": "- We propose PointLoRA, a simple yet effective scheme for parameter-efficient fine-tuning in point cloud learning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.718, + 0.483, + 0.776 + ], + "angle": 0, + "content": "- Low-rank adaptation is injected into the most parameter-intensive components of the point cloud transformer to capture global information, with a corresponding analysis in the context of PointNet provided." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.778, + 0.483, + 0.837 + ], + "angle": 0, + "content": "- We further design a multi-scale token selection module that dynamically incorporates local geometric information as prompt, enriching global low-rank adaptation with the crucial local context." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.838, + 0.484, + 0.899 + ], + "angle": 0, + "content": "- Extensive experiments across three challenging datasets and two widely used pre-trained models validate the effectiveness of our method, achieving state-of-the-art performance with only \\(3.43\\%\\) of the trainable parameters." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.687, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.347, + 0.656, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.719 + ], + "angle": 0, + "content": "Model Pre-Training on Point Clouds. Self-supervised pre-training for 3D point clouds has been widely explored [36, 43, 47, 55, 56, 64, 69, 76], generally categorized into contrastive-based approaches [1, 7, 39, 64, 66] and masked signal reconstruction methods [5, 6, 40, 69, 74]. In contrastive learning, PointContrast [64] and CrossPoint [1] leverage different views or instances to extract latent features through contrasting representations. SegContrast [39] extends this approach to outdoor LiDAR data, extracting class-agnostic segments from augmented views [24]. Reconstruction-based methods, such as Point-BERT [69] and Point-MAE [40], employ masked point cloud reconstruction for pre-training. In ACT [10], a pre-trained autoencoder serves as a cross-modality teacher, guiding 3D point cloud learning through knowledge distillation [17, 57, 58]. PointGPT [5] adapts the GPT framework [38] for autoregressive generation on point clouds. Additionally, Qi et al. [43, 44] propose a combination of reconstruction and cross-modal contrastive learning within generative models. While full fine-tuning is commonly applied to these pretrained models for downstream tasks, the high computational costs and potential dilution of pre-trained knowledge motivate our exploration of efficient fine-tuning strategies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Parameter-Efficient Fine-Tuning (PEFT). PEFT methods have gained traction in NLP and 2D vision tasks [8, 9, 19, 21, 28, 48, 63, 73] to enhance downstream performance with minimal tunable parameters. Mainstream PEFT approaches include prompt and prefix tuning, adapter-based methods, and low-rank adaptation. Prompt and prefix tuning [21, 25, 28, 48] introduce additional tokens as tunable components, with VPT [21] pioneering the use of prompt tokens for pre-trained vision transformers (ViTs) [11, 26, 49, 67]. Adapter-based tuning inserts trainable modules between frozen layers [8, 19, 27, 50], while LoRA [20] applies a low-rank approximation to update linear layers within" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6606" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "attention blocks. Building on these techniques, recent innovations explore alternative adapter placements [14, 31], tuning bias terms [71], and combining with Mixture-of-Experts (MoE) techniques [35, 70]. Recent studies have also adapted PEFT to point cloud analysis [32, 51, 72, 75, 77]. IDPT [72] extends prompt tuning to DGCNN [59] for instance-aware prompt extraction. Point-PEFT [51] and DAPT [77] combine prompt tuning with adapters to further improve fine-tuning performance, while PPT [75] introduces positional prompt tuning for efficient 3D representation learning. In contrast to these approaches, our work uniquely integrates low-rank adaptation with multi-scale token selection, achieving competitive performance with a significantly reduced parameter footprint." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.315, + 0.218, + 0.332 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.372 + ], + "angle": 0, + "content": "In this section, we revisit the architecture of point cloud transformer and the corresponding fine-tuning prototypes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.378, + 0.322, + 0.393 + ], + "angle": 0, + "content": "3.1. Point Cloud Transformer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Point Tokenizer. Due to the sparsity and irregularity of point clouds, it is necessary to transform them into a token sequence suitable for processing by the transformer. We follow standard settings [40, 69] by segmenting the point cloud into irregular patches using the farthest point sampling (FPS) and the \\(k\\)-nearest neighbors (\\(k\\)-NN) algorithm. Formally, given the point cloud \\(P = \\{p_1, p_2, \\ldots, p_N\\} \\in \\mathbb{R}^{N \\times 3}\\) with \\(N\\) points, the \\(g\\) group centers are obtained with FPS:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.535, + 0.482, + 0.553 + ], + "angle": 0, + "content": "\\[\nC _ {g} = \\operatorname {F P S} (P), \\quad C _ {g} \\in \\mathbb {R} ^ {g \\times 3}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.483, + 0.589 + ], + "angle": 0, + "content": "where \\( C_g \\) is the selected center points. Then, we use \\( k \\)-NN to select \\( k \\) nearest neighbor points for each center in \\( C_g \\):" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.593, + 0.482, + 0.612 + ], + "angle": 0, + "content": "\\[\nN _ {p} = k \\text {- N N} (P, C _ {g}), \\quad N _ {p} \\in \\mathbb {R} ^ {g \\times k \\times 3}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.695 + ], + "angle": 0, + "content": "where \\( N_{p} \\) is the corresponding neighbor point patch. Notably, these local point patches are made unbiased by subtracting their center points, which promotes better convergence. Then a mini-PointNet [41] is utilized to embed the point patches into discrete tokens \\( T_{p} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.701, + 0.482, + 0.72 + ], + "angle": 0, + "content": "\\[\nT _ {p} = \\operatorname {m i n i - P o i n t N e t} \\left(N _ {p}\\right), \\quad T _ {p} \\in \\mathbb {R} ^ {g \\times d}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.484, + 0.802 + ], + "angle": 0, + "content": "where \\(d\\) is the embedding dimension. Following the tokenization process, the 3D point cloud is converted into a feature vector, enabling subsequent processing with Transformer architectures similar to those used in natural language processing and 2D vision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Transformer Block. With a classification token \\( T_{\\mathrm{cls}} \\) and the obtained \\( T_{p} \\), the input \\( T_{\\mathrm{input}} = \\mathrm{Concat}(T_{\\mathrm{cls}}, T_{p}) \\) is further processed by \\( L \\)-layer transformer blocks. In each block, we first project \\( T_{\\mathrm{input}} \\) into query, key, and value spaces:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.868, + 0.482, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = T _ {\\text {i n p u t}} \\mathbf {W} ^ {Q}, \\quad \\mathbf {K} = T _ {\\text {i n p u t}} \\mathbf {W} ^ {K}, \\quad \\mathbf {V} = T _ {\\text {i n p u t}} \\mathbf {W} ^ {V}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.906, + 0.138 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}^Q, \\mathbf{W}^K, \\mathbf{W}^V \\in \\mathbb{R}^{d \\times d}\\) are trainable matrices in qkv projection layer. The attention layer then computes the self-attention with a skip connection as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.152, + 0.905, + 0.17 + ], + "angle": 0, + "content": "\\[\nT _ {\\text {o u t p u t}} ^ {\\prime} = \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) + T _ {\\text {i n p u t}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.173, + 0.808, + 0.189 + ], + "angle": 0, + "content": "Then, the final output \\( T_{\\mathrm{output}} \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.197, + 0.905, + 0.216 + ], + "angle": 0, + "content": "\\[\nT _ {\\text {o u t p u t}} = \\text {F F N} \\left(\\ln \\left(T _ {\\text {o u t p u t}} ^ {\\prime}\\right)\\right) + T _ {\\text {o u t p u t}} ^ {\\prime}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.906, + 0.329 + ], + "angle": 0, + "content": "where LN is the layer norm function [2] and FFN is the feed-forward network layer [3]. Through \\(L\\) Transformer blocks, the network extracts deep and essential features from the discrete tokens. Notably, the qkv projection parameters within the attention layer, along with those in the FFN layer constitute a significant portion of the model's parameters and are crucial for performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.337, + 0.725, + 0.353 + ], + "angle": 0, + "content": "3.2. Fine-Tuning Prototype" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.906, + 0.403 + ], + "angle": 0, + "content": "Leveraging well-pretrained models, fine-tuning prototypes to adapt these models to diverse downstream tasks can be categorized as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.906, + 0.509 + ], + "angle": 0, + "content": "Full Fine-Tuning. Full fine-tuning is the predominant approach in current point cloud pre-training research for downstream tasks, involving the direct loading of pretrained model parameters and joint training of the encoder alongside the task-specific head. Although this method offers a theoretically higher accuracy potential, it also entails substantial computational and storage overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.661 + ], + "angle": 0, + "content": "Parameter-Efficient Fine-Tuning. Parameter-efficient fine-tuning adapts pre-trained models to downstream tasks by freezing most of the model's parameters, introducing lightweight modules, and selectively fine-tuning a small subset of parameters. This approach provides a more practical solution, particularly as model sizes continue to increase. Existing methods typically combine adapter tuning, which inserts bottleneck-like layers, and prompt tuning, which introduces prior knowledge, to fine-tune point cloud pre-trained models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.673, + 0.634, + 0.689 + ], + "angle": 0, + "content": "4. PointLoRA" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.625, + 0.713 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.854 + ], + "angle": 0, + "content": "In this work, we aim to provide a simple but effective parameter-efficient fine-tuning method for point cloud pretrained models. The complete framework of our proposed PointLoRA, integrated into the existing point cloud transformer, is illustrated in Fig. 2, addressing the parameter-intensive nature of the transformer-based network design. Our approach can effectively enhance downstream task performance while keeping most of the pre-trained model parameters frozen." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In the following section, we first outline the establishment of the vanilla LoRA baseline and analyze the factors contributing to its effectiveness in point cloud learning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6607" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.297, + 0.908, + 0.355 + ], + "angle": 0, + "content": "Figure 2. Overview of PointLoRA integrated into point cloud transformer pipeline. Given an input point cloud, we first tokenize it using the original Point Tokenizer and perform token selection across multiple scales (Multi-Scale Token Selection). The tokens from both components are then concatenated and fed into the Transformer Block. Our approach is injected into the qkv projection and FFN layers, utilizing a shared Prompt MLP within these layers to enhance parameter efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.424 + ], + "angle": 0, + "content": "Then we present how our method extends LoRA's feature extraction capabilities to achieve competitive performance in 3D scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.433, + 0.303, + 0.449 + ], + "angle": 0, + "content": "4.2. Vanilla LoRA Baseline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.606 + ], + "angle": 0, + "content": "As mentioned in Sec. 3.1, the qkv projection and feed-forward network (FFN) layers in point cloud transformer significantly increase the number of parameters, limiting the fine-tuning efficiency and hindering deployment in resource-limited environments. To address this issue, we incorporate Low-Rank Adaptation (LoRA) into these layers by introducing low-rank matrices to adapt pre-trained weights, thereby reducing the number of trainable parameters while preserving model performance. Given a pretrained weight matrix \\( W_{p} \\), LoRA modifies it as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.618, + 0.483, + 0.635 + ], + "angle": 0, + "content": "\\[\nW _ {\\text {u p d a t e}} = W _ {p} + \\Delta W = W _ {p} + W _ {u} \\cdot W _ {d}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.704 + ], + "angle": 0, + "content": "where \\(\\Delta W\\) and \\(W_{\\mathrm{update}}\\) denotes the updated weight and new weight matrix, respectively. \\(W_{u}\\in \\mathbb{R}^{d\\times r}\\) and \\(W_{d}\\in \\mathbb{R}^{r\\times d}\\) are low-rank matrices, where \\(r\\ll d\\) indicates the rank, controlling the complexity of the adaptation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.811 + ], + "angle": 0, + "content": "During training, only \\( W_{u} \\) and \\( W_{d} \\) are updated, while the pre-trained weight \\( W_{p} \\) remains frozen. For inference, the adaptation \\( \\Delta W \\) can be merged with the original weights, yielding in a single consolidated weight matrix: \\( W_{\\mathrm{infer}} = W_{p} + \\Delta W \\). This consolidation enables the model to retain the benefits of low-rank adaptation while avoiding additional computational overhead in the inference phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Analysis. LoRA's architecture is particularly well-suited for point cloud data, as it can capture the complex relationships between points essential to understanding 3D shapes and spatial arrangements. Its MLP-like structure (i.e., two low-rank matrices) effectively learns global features while aligning with PointNet's principles of handling unordered point" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.456 + ], + "angle": 0, + "content": "sets and extracting permutation-invariant features. The synergy between LoRA and PointNet's feature extraction capabilities establishes a robust framework for harnessing the rich information in point cloud data, enhancing both generalization and adaptability in complex environments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.803, + 0.483 + ], + "angle": 0, + "content": "4.3. PointLoRA with Token Selection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.687 + ], + "angle": 0, + "content": "While global features provide a comprehensive understanding of the overall structure of the point cloud, local features capture intricate details vital for accurate downstream task performance. To fully investigate LoRA's capabilities, it is crucial to complement the extracted global features with robust local features. The tokenization process produces a sequence of tokens that capture various aspects of the point cloud data. However, not all tokens are valuable for downstream tasks; some may encapsulate redundant or noninformative content, offering limited contribution to overall model performance. This necessitates a systematic approach for filtering and selecting the most informative tokens to enhance the model's representation of local features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.824 + ], + "angle": 0, + "content": "To tackle the above concerns, we introduce a MultiScale Token Selection module integrated with vanilla LoRA, which is presented in Fig. 3. This module dynamically selects tokens from the raw point cloud at multiple scales, allowing the model to capture local features at different levels of detail. Additionally, PointLoRA incorporates a shared Prompt MLP to further embed local information from these selected tokens, enhancing the vanilla LoRA's performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.904 + ], + "angle": 0, + "content": "Multi-Scale Token Generation. Specifically, we apply farthest point sampling (FPS) on the input point cloud \\(P\\) with varying numbers of centroids (i.e., \\(N_{1}, N_{2}, \\dots, N_{M}\\)) to generate tokens at \\(M\\) different scales. As described in Sec. 3, we can obtain \\(M\\) sets of center points, \\(C_{g}^{1}, C_{g}^{2}, \\dots, C_{g}^{M}\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.468, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.384, + 0.484, + 0.468 + ], + "angle": 0, + "content": "Figure 3. Illustration of Multi-Scale Token Selection. In a two-scale setup, we first sample different numbers of center points, then cluster around each center point and apply Mini-PointNet encoding to generate the corresponding tokens. These tokens are also fed into a Mask Predictor to estimate importance scores, allowing us to select the Top-K tokens at each scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.553 + ], + "angle": 0, + "content": "each containing \\( g_{1}, g_{2}, \\ldots, g_{M} \\) center points. Subsequently, \\( k \\)-NN is employed to retrieve the corresponding neighbor point patches, which are then embedded into discrete tokens using a shared Mini-PointNet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.706 + ], + "angle": 0, + "content": "Mask Predictor for Selection. The above process yields \\(M\\) token sets, \\(T_{p}^{1}, T_{p}^{2}, \\ldots, T_{p}^{M}\\), encoding local information at multiple scales. Our goal is to select tokens from these sets that are most beneficial for fine-tuning on downstream tasks. Therefore, a simple but effective Mask Predictor is designed to estimate the importance of each token, producing a score vector \\(s^{m} \\in \\mathbb{R}^{g_{m} \\times 1}\\) for each token in \\(T_{p}^{m} \\in \\mathbb{R}^{g_{m} \\times d}\\), \\(m = 1, \\ldots, M\\). In particular, the Mask Predictor comprises two multi-layer perceptron (MLP) layers followed by a Sigmoid activation function:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.715, + 0.482, + 0.734 + ], + "angle": 0, + "content": "\\[\ns ^ {m} = \\operatorname {S i g m o i d} \\left(\\operatorname {M L P} \\left(\\left(T _ {p} ^ {m}\\right)\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.835 + ], + "angle": 0, + "content": "This design enables the model to assign scores between 0 and 1 to each token, indicating their relevance for downstream tasks. We then select a predefined number \\( N_{1}^{\\prime}, \\ldots, N_{M}^{\\prime} \\) of tokens from each scale based on their importance scores using Top-K selection, resulting in a refined token set \\( \\{S_{\\mathrm{p}}^{m}\\}_{m=1}^{M} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.844, + 0.483, + 0.863 + ], + "angle": 0, + "content": "\\[\nS _ {\\mathrm {p}} ^ {m} = \\operatorname {T o p K} \\left(T _ {p} ^ {m}, N _ {m} ^ {\\prime}\\right), m = 1, \\dots , M. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.87, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Local Geometry Prompt. With the selected \\( N_{s} = N_{1}^{\\prime} + \\ldots + N_{M}^{\\prime} \\) tokens \\( S_{p} \\), we can capture crucial local information" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "for downstream tasks. These tokens are then concatenated with the original inputs to form a local geometry prompt. In each LoRA layer of the point cloud transformer block, a Prompt MLP with a GELU activation function [16] encodes the new input features alongside the local prompt. This process integrates the encoded features of the selected tokens with the LoRA adaptation output. The resulting combined output can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.224, + 0.905, + 0.254 + ], + "angle": 0, + "content": "\\[\nO _ {\\text {u p d a t e}} = \\operatorname {P r o m p t} \\mathrm {M L P} \\left(T _ {\\text {i n p u t}}, S _ {p}\\right) + \\Delta W \\cdot \\left(T _ {\\text {i n p u t}}, S _ {p}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.907, + 0.33 + ], + "angle": 0, + "content": "where \\( O_{\\mathrm{update}} \\) is the updated amount of the PointLoRA layer. Notably, to further optimize the efficiency of the parameters, the Prompt MLP is configured independently for both the qkv projection and FFN layers, while being shared across all blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.906, + 0.422 + ], + "angle": 0, + "content": "The multi-scale token selection process improves model performance by prioritizing the most relevant local features during fine-tuning. By incorporating this approach, our scheme effectively refines both global and local feature representations, thereby enhancing its adaptability and accuracy across a range of 3D tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.432, + 0.715, + 0.449 + ], + "angle": 0, + "content": "4.4. Training & Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.907, + 0.592 + ], + "angle": 0, + "content": "Overall Loss Function. When fine-tuning on diverse downstream tasks across various datasets, we follow the common practice of maintaining consistency between the adopted loss and the loss function of the original task. Additionally, we introduce a regularization term to further supervise the Mask Predictor. In general, the total loss function \\(\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_{\\mathrm{task}} + \\lambda \\cdot \\mathcal{L}_{\\mathrm{mask}}\\), where \\(\\mathcal{L}_{\\mathrm{task}}\\) is the task-specific loss for classification or segmentation. \\(\\mathcal{L}_{\\mathrm{mask}}\\) is the regularization loss for mask predictor with the balanced weight \\(\\lambda\\):" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.601, + 0.905, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = - \\frac {1}{N _ {\\text {t o t a l}}} \\sum_ {i = 1} ^ {N _ {\\text {t o t a l}}} \\left(s _ {i} \\log \\left(s _ {i} + \\epsilon\\right) + \\right. \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.635, + 0.799, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\left. \\left(1 - s _ {i}\\right) \\log \\left(1 - s _ {i} + \\epsilon\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.703 + ], + "angle": 0, + "content": "where \\( N_{\\mathrm{total}} = N_1 + N_2 \\dots + N_M \\), \\( s_i \\) denotes the importance score for each token and \\( \\epsilon \\) is a small constant (e.g., \\( 10^{-6} \\)) to prevent undefined values in the logarithm." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.779 + ], + "angle": 0, + "content": "Inference. During inference, the additional parameters introduced by LoRA in the transformer block are directly merged into the original pre-trained weights, with the only newly added component being a small shared MLP dedicated to embed the selected tokens." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.794, + 0.646, + 0.81 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.596, + 0.835 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Datasets. We validate the effectiveness of our approach through extensive experiments on three widely used 3D datasets: ScanObjectNN [52], ModelNet40 [61], and ShapeNetPart [68]. ScanObjectNN [52] is a challenging" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6609" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Performance comparison on three variants of the ScanObjectNN [52] and the ModelNet40 [61] datasets, respectively, for real-world and synthetic object classification. Both the number of tunable parameters and overall accuracy (OA) are reported. All methods only employ the default data argumentation without voting as the baseline. * denotes results reproduced from the public source code." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.136, + 0.907, + 0.493 + ], + "angle": 0, + "content": "
MethodsPublicationTunable Params.ScanObjectNNModelNet40
OBJ-BGOBJ-ONLYPB-T50-RSPoints Num.OA (%)
Traditional Supervised Learning Only
PointNet [41]CVPR'173.5 M73.379.268.01k89.2
PointNet++ [42]NeurIPS'171.5 M82.384.377.91k90.7
DGCNN [59]TOG'191.8 M82.886.278.11k92.9
MVTN [13]ICCV'2111.2 M--82.81k93.8
PointNeXt [45]NeurIPS'221.4 M--87.71k94.0
PointMLP [37]ICLR'2213.2 M--85.41k94.5
RepSurf-U [46]CVPR'221.5 M--84.31k94.4
ADS [18]ICCV'23---87.51k95.1
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2122.1 M84.8585.5478.791k92.1
Point-BERT [69]CVPR'2222.1 M87.4388.1283.071k93.2
MaskPoint [34]ECCV'2222.1 M89.7089.3084.601k93.8
Point-MAE [40]ECCV'2222.1 M90.0288.2985.181k93.8
Point-M2AE [74]NeurIPS 2215.3 M91.2288.8186.431k94.0
ACT [10]ICLR'2322.1 M93.2991.9188.211k93.7
RECON [43]ICML'2343.6 M94.1593.1289.731k93.9
Self-Supervised Representation Learning (Parameter-Efficient Fine-Tuning)
Point-MAE [40] (Full-FT)ECCV'2222.1 M (100%)90.0288.2985.181k93.2
Point-MAE + IDPT [72]ICCV'231.7 M (7.69%)91.22(+1.20)90.02(+1.73)84.94(-0.24)1k93.3(+0.1)
Point-MAE + DAPT [77]CVPR'241.1 M (4.97%)90.88(+0.86)90.19(+1.90)85.08(-0.10)1k93.5(+0.3)
Point-MAE + PPT* [75]arXiv'241.04 M (4.57%)89.84(-0.18)88.98(+0.69)84.45(-0.73)1k93.2(+0.0)
Point-MAE + PointLoRAOurs0.77 M (3.43%)90.71(+0.69)89.33(+1.04)85.53(+0.35)1k93.3(+0.1)
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.482, + 0.749 + ], + "angle": 0, + "content": "real-world 3D object classification dataset, which contains approximately 15,000 indoor point cloud instances across 15 categories. The object classification task is performed on three variants of increasing complexity, OBJ-BG, OBJ-ONLY, and PB-T50-RS, each representing different levels of scene realism and occlusion. ModelNet40 [61], a classical synthetic dataset for 3D object recognition, consists of 12,311 meshed 3D CAD objects that span 40 categories. We perform synthetic object classification and few-shot learning experiments on ModelNet40 to evaluate the robustness of the model in both standard and low-resource settings. Additionally, to evaluate performance on detailed structures and component segmentation, we provide experiments on ShapeNetPart [68], a widely used benchmark for part segmentation, comprising 16,881 point-level synthetic objects across 16 object categories and 50 part categories." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.482, + 0.87 + ], + "angle": 0, + "content": "Evaluation Metrics. In the evaluation, overall accuracy (OA) is adopted to assess performance on the 3D object classification task, representing the ratio of correctly classified instances to the total number of instances, thus providing an aggregate score over all classes. For the part segmentation task, we employ mean Intersection over Union (mIoU) to evaluate the overlap between prediction and ground-truth segments, averaged across all classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. Our method can be directly integrated into existing point cloud pre-trained models. Dur" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.766 + ], + "angle": 0, + "content": "ing fine-tuning, we follow a commonly used setup, freezing most parameters of the pre-trained model, and primarily updating the newly inserted parameters. The number of scales, \\(M\\), is set to 2. At these two scales, we apply farthest point sampling (FPS) and \\(k\\)-NN clustering with (128, 32) and (64, 64) center and neighboring points, respectively, and then select tokens with \\(N_1' = 32\\) and \\(N_2' = 8\\). The balanced weight \\(\\lambda\\) for Mask Predictor learning is set to 0.004 For real-world and synthetic object classification, we fine-tune using Point-MAE [40], training for 300 epochs with a learning rate of \\(5e^{-4}\\) and a weight decay of 0.05. For few-shot learning and part segmentation, we utilize the Recon [43] model for fine-tuning to further validate the generalizability of our approach, with a learning rate of \\(2e^{-4}\\) for part segmentation. All experiments are conducted on a single GPU. More implementation details with different baselines are provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.702, + 0.79 + ], + "angle": 0, + "content": "5.2. Comparative Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Real-World & Synthetic Object Classification. We conduct real-world and synthetic object classification on three variants of ScanObjectNN (OBJ-BG, OBJ-ONLY, PB-T50-RS) [52] and ModelNet40 [61], using default data augmentation and without voting. As illustrated in Tab. 1, our proposed approach demonstrates consistent performance improvements, especially on the most challenging vari" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6610" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.26, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.089, + 0.415, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.089, + 0.571, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.089, + 0.727, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.089, + 0.882, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.236, + 0.907, + 0.278 + ], + "angle": 0, + "content": "Figure 4. The t-SNE visualization results on the PB-T50-RS split of the ScanObjectNN dataset [52] with different fine-tuning schemes. We adopt Point-MAE [40] as the baseline model for fair comparison. TP: Number of tunable parameters. OA: Overall accuracy. Symbol * denotes re-produced with official implementation. Best viewed in colors and zoomed-in for additional details." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.292, + 0.486, + 0.349 + ], + "angle": 0, + "content": "Table 2. Performance comparison on ModelNet40 [61] for few-shot learning. We report the scores of the overall accuracy \\((\\%)\\pm\\) the standard deviation \\((\\%)\\) without voting. The top two highest accuracies are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.352, + 0.482, + 0.564 + ], + "angle": 0, + "content": "
MethodsPublication5-way10-way
10-shot20-shot10-shot20-shot
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2194.0±3.695.9±2.389.4±5.192.4±4.6
Point-BERT [69]CVPR'2294.6±3.196.3±2.791.0±5.492.7±5.1
MaskPoint [34]ECCV'2295.0±3.797.2±1.791.4±4.093.4±3.5
Point-MAE [40]ECCV'2296.3±2.597.8±1.892.6±4.195.0±3.0
Point-M2AE [74]NeurIPS'2296.8±1.898.3±1.492.3±4.595.0±3.0
ACT [10]ICLR'2396.8±2.398.0±1.493.3±4.095.6±2.8
Self-Supervised Representation Learning (Efficient Fine-Tuning)
RECON [43] (Full-FT)ICML'2397.3±1.998.9±3.993.3±3.995.8±3.0
RECON+ IDPT [72]ICCV'2396.9±2.498.3±0.792.8±4.095.5±3.2
RECON+ DAPT [77]CVPR'2495.6±2.897.7±1.691.9±4.194.6±3.5
RECON+ PPT [75]arXiv'2497.0±2.798.7±1.692.2±5.095.6±2.9
RECON+ PointLoRAOurs96.9±2.798.8±1.292.7±4.495.8±2.9
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.483, + 0.732 + ], + "angle": 0, + "content": "ant, PB-T50-RS, where it is the only parameter-efficient fine-tuning method that surpasses the full fine-tuned PointMAE [40] (85.53% vs. 85.18%) and significantly outperforms the state-of-the-art method, DAPT [77]. Notably, our approach requires only 0.77M tunable parameters, the fewest among all methods. Furthermore, the t-SNE feature visualization [53] of existing methods and PointLoRA is presented in Fig. 4. Our method enables the pre-trained model to fine-tune with a minimal number of parameters, producing distinctive feature representations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.733, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Few-shot Learning. We also perform few-shot learning experiments on ModelNet40 [61] to validate the transfer learning capability of our method with limited annotations. Following standard protocols [15, 43], we evaluate the fine-tuning performance with ReCon [43] in the 5-way / 10-way and 10-shot / 20-shot settings, respectively. As shown in Tab. 2, our method achieves the best or second-best accuracy in all configurations, further demonstrating the generalizability of the proposed scheme." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Point Cloud Part Segmentation. The part segmentation results on the ShapeNetPart dataset [68] are provided in" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.362 + ], + "angle": 0, + "content": "Table 3. Performance comparison on the ShapeNetPart [68] for part segmentation. Both the mIoU for all classes (Cls.) and instances (Inst.) are provided. TP indicates the number of tunable parameters. \\(\\dagger\\) means the result, reported from DAPT [77]. \\(*\\) denotes the results reproduced from the official implementation." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.365, + 0.9, + 0.617 + ], + "angle": 0, + "content": "
MethodsPublicationTPCls. mIoU (%)Inst. mIoU (%)
Traditional Supervised Learning Only
PointNet [41]CVPR'17-80.3983.7
PointNet++ [42]NeurIPS'17-81.8585.1
DGCNN [59]TOG'19-82.3385.2
APES [60]CVPR'23-83.6785.8
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2127.09 M83.4285.1
MaskPoint [34]ECCV'22-84.6086.0
Point-BERT [69]CVPR'2227.09 M84.1185.6
Point-MAE [40]ECCV'2227.06 M84.1986.1
ACT [10]ICLR'2327.06 M84.6686.1
Self-Supervised Representation Learning (Efficient Fine-Tuning)
RECON [43] (Full-FT)ICML'2327.06 M84.5286.1
RECON+ IDPT† [72]ICCV'235.69 M83.6685.7
RECON+ DAPT [77]CVPR'245.65 M83.8785.7
RECON+ PPT* [75]arXiv'245.62 M83.8885.4
RECON+ PointLoRAOurs5.63 M83.9885.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.702 + ], + "angle": 0, + "content": "Tab. 3. In this fine-grained scene understanding task, our approach still achieves competitive performance with a constrained parameter budget. Unlike prior classification models, the increase in the parameter count is primarily attributed to the segmentation head." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Comparison with Other PEFT Methods. We further present a comparison with existing PEFT methods designed for other tasks, including NLP and 2D vision. We select the most challenging PB-T50-RS variant [52] and use the pretrained Point-MAE model [40] as a baseline. As shown in Tab. 4, these methods designed for other tasks provide limited performance gains when fine-tuned in 3D scenes, despite their parameter efficiency compared to existing 3D fine-tuning approaches. In contrast, our method not only achieves the highest accuracy, but also uses significantly fewer parameters than the three current fine-tuning techniques specifically developed for point clouds, underscoring the strong potential of PointLoRA." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6611" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.486, + 0.132 + ], + "angle": 0, + "content": "Table 4. Performance comparison with other parameter-efficient methods designed for NLP and 2D Vision tasks on the hardest variant of ScanObjectNN [52]." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.136, + 0.46, + 0.324 + ], + "angle": 0, + "content": "
MethodsPublicationTPPB-T50-RS
Point-MAE [34]ECCV'2222.1 M85.18
Linear probing-0.3 M75.99
+ Adapter [19]ICML'190.9 M83.93
+ Perfix tuning [28]ACL'210.7 M77.72
+ BitFit [71]ACL'210.3 M82.62
+ LoRA†[20]ICLR'220.9 M81.74
+ VPT-Deep [21]ECCV'220.4 M81.09
+ AdaptFormer [8]NeurIPS'220.9 M83.45
+ SSF [31]NeurIPS'220.4 M82.58
+ IDPT [72]ICCV'231.7 M84.94
+ DAPT [77]CVPR'241.1 M85.08
+ PPT* [75]arXiv'241.04 M84.45
+ PointLoRAOurs0.77 M85.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.326, + 0.485, + 0.385 + ], + "angle": 0, + "content": "Table 5. The impact of each module of our scheme. We provide the overall accuracy (\\%) on the hardest variant of ScanObjectNN [52] and corresponding tunable parameters (TP). \"MS-FPS\" indicates multi-scale furthest point sampling." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.387, + 0.462, + 0.48 + ], + "angle": 0, + "content": "
LoRAToken SelectionMS-FPSTPPB-T50-RS
Full Fine-Tuning22.1 M85.18
Linear Probing0.27 M75.99
✓0.53 M83.83
✓✓0.77 M84.91
✓✓✓0.77 M85.53
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.486, + 0.245, + 0.503 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.483, + 0.584 + ], + "angle": 0, + "content": "In this section, we perform exhaustive ablation experiments on the challenging PB-T50-RS variant to investigate the rationale and effectiveness of the design choices of our proposed approach. The pre-trained Point-MAE [40] is adopted as a baseline for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Ablation on PointLoRA Scheme. Firstly, we provide the ablations on each module of the proposed method. As illustrated in Tab. 5, directly applying low-rank adaptation (LoRA) produces a significant improvement over linear probing, demonstrating the effectiveness of incorporating global information during the fine-tuning process. Injecting selected tokens as prompts into LoRA further improves accuracy with only a marginal increase in the number of parameters. Finally, the local information obtained from selected tokens across multiple scales complements the global features of LoRA, achieving the best performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Ablation on Low-Rank Adaptation. Then we perform ablation experiments on rank \\((r)\\) in our approach, as shown in Tab. 6. Different ranks have a noticeable impact on finetuning performance. Taking into account both parameter count and accuracy, we set the rank to 8 to maximize the extraction of global information for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Ablation on Multi-Scale Token Selection. Given the critical improvement from multi-scale token selection for PointLoRA, we first perform ablation experiments on the number of tokens selected at the same scale. As shown in" + }, + { + "type": "table_caption", + "bbox": [ + 0.531, + 0.09, + 0.888, + 0.105 + ], + "angle": 0, + "content": "Table 6. Ablation study on rank \\((r)\\) of the proposed method." + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.108, + 0.86, + 0.184 + ], + "angle": 0, + "content": "
Rank rTPRatioPB-T50-RS
40.66 M2.96%84.28
80.77 M3.43%85.53
160.99 M4.37%85.15
321.44 M6.32%84.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.186, + 0.905, + 0.213 + ], + "angle": 0, + "content": "Table 7. Ablation study on scale number \\( M \\) and selected token number \\( N_{s} \\) in Multi-Scale Token Selection." + }, + { + "type": "table", + "bbox": [ + 0.556, + 0.218, + 0.86, + 0.339 + ], + "angle": 0, + "content": "
Scale MToken Num. NsPB-T50-RS
16484.80
13284.91
11685.22
264 & 1684.39
232 & 885.53
216 & 484.49
332 & 16 & 885.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.342, + 0.882, + 0.357 + ], + "angle": 0, + "content": "Table 8. Ablation study on the dimension of Prompt MLP." + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.36, + 0.858, + 0.436 + ], + "angle": 0, + "content": "
DimensionTPRatioPB-T50-RS
80.68 M3.03%83.55
160.71 M3.17%84.77
320.77 M3.43%85.53
640.90 M3.96%85.25
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.544 + ], + "angle": 0, + "content": "the upper part of Tab. 7, selecting 16 tokens at the original scale yields the best results, demonstrating the effectiveness of token selection. Furthermore, when incorporating tokens from multiple scales, optimal performance is achieved by selecting 32 and 8 tokens from two different scales, as shown in the lower part of Tab. 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.909, + 0.665 + ], + "angle": 0, + "content": "Ablation on Prompt MLP. We provide the ablation study on the dimension of the shared Prompt MLP, as this affects the embedding of local information in our method. As illustrated in Tab. 8, setting this dimension to 32 yields the best fine-tuning performance. Additionally, adjusting the dimension causes only minimal changes in tunable parameter counts, suggesting flexibility in selecting this hyperparameter for different datasets and tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.9 + ], + "angle": 0, + "content": "In this paper, a simple yet effective parameter-efficient finetuning approach named PointLoRA, is presented with low-rank adaptation and multi-scale token selection for point cloud models. Low-rank adaptation efficiently reduces the parameter-heavy components of the point cloud transformer architecture while capturing global information. Multi-scale token selection effectively encodes essential local features as prompt, further enhancing the finetuning process. The integration of global and local information enables our approach to achieve state-of-the-art results on the most challenging dataset with a minimal number of tunable parameters while also delivering competitive performance across other datasets and models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6612" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.284 + ], + "angle": 0, + "content": "This work is supported in part by the National Natural Science Foundation of China under Grant No. 62376244, in part by the Zhejiang Provincial Natural Science Foundation of China under Grant No. LD24F030001, by the Information Technology Center and State Key Lab of CAD&CG, Zhejiang University, and by the National Research Foundation, Singapore, and Cyber Security Agency of Singapore under its National Cybersecurity R&D Programme and CyberSG R&D Cyber Research Programme Office (Award: CRPO-GC1-NTU-002). Lingdong Kong is supported by the Apple Scholars in AI/ML Ph.D. Fellowship program." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.295, + 0.188, + 0.312 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.32, + 0.484, + 0.405 + ], + "angle": 0, + "content": "[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.406, + 0.484, + 0.433 + ], + "angle": 0, + "content": "[2] Jimmy Lei Ba. Layer normalization. arXiv preprint arXiv:1607.06450, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.434, + 0.484, + 0.462 + ], + "angle": 0, + "content": "[3] George Bebis and Michael Georgiopoulos. Feed-forward neural networks. IEEE Potentials, 13(4):27-31, 1994." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.463, + 0.484, + 0.505 + ], + "angle": 0, + "content": "[4] Saifullahi Aminu Bello, Shangshu Yu, Cheng Wang, Jibril Muhammad Adam, and Jonathan Li. Deep learning on 3d point clouds. Remote Sensing, 12(11):1729, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.507, + 0.484, + 0.562 + ], + "angle": 0, + "content": "[5] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. In Advances in Neural Information Processing Systems, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.564, + 0.484, + 0.632 + ], + "angle": 0, + "content": "[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.634, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[7] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.704, + 0.484, + 0.76 + ], + "angle": 0, + "content": "[8] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.761, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[9] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In International Conference on Learning Representations, 2022." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.32, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[12] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[13] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In IEEE/CVF International Conference on Computer Vision, pages 1-11, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.343 + ], + "angle": 0, + "content": "[14] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.906, + 0.427 + ], + "angle": 0, + "content": "[16] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.524 + ], + "angle": 0, + "content": "[18] Cheng-Yao Hong, Yu-Ying Chou, and Tyng-Luh Liu. Attention discriminant sampling for point clouds. In IEEE/CVF International Conference on Computer Vision, pages 14429-14440, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.906, + 0.775 + ], + "angle": 0, + "content": "[22] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[23] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6613" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.146 + ], + "angle": 0, + "content": "[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Empirical Methods in Natural Language Processing, pages 3045-3059, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.484, + 0.204 + ], + "angle": 0, + "content": "[26] Bonan Li, Yinhan Hu, Xuecheng Nie, Congying Han, Xi-angjian Jiang, Tiande Guo, and Luoqi Liu. Dropkey for vision transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22700–22709, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[27] Minglei Li, Peng Ye, Yongqi Huang, Lin Zhang, Tao Chen, Tong He, Jiayuan Fan, and Wanli Ouyang. Adapter-x: A novel general parameter-efficient fine-tuning framework for vision. arXiv preprint arXiv:2406.03051, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[28] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Annual Meeting of the Association for Computational Linguistics, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[29] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In Advances in Neural Information Processing Systems, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[30] Ye Li, Lingdong Kong, Hanjiang Hu, Xiaohao Xu, and Xiaonan Huang. Is your lidar placement optimized for 3d scene understanding? In Advances in Neural Information Processing Systems, pages 34980-35017, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[31] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.484, + 0.524 + ], + "angle": 0, + "content": "[32] Dingkang Liang, Tianrui Feng, Xin Zhou, Yumeng Zhang, Zhikang Zou, and Xiang Bai. Parameter-efficient fine-tuning in spectral domain for point cloud learning. arXiv preprint arXiv:2410.08114, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[33] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.484, + 0.623 + ], + "angle": 0, + "content": "[34] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.484, + 0.69 + ], + "angle": 0, + "content": "[35] Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Derong Xu, Feng Tian, and Yefeng Zheng. Moelora: An moe-based parameter efficient fine-tuning method for multitask medical applications. arXiv preprint arXiv:2310.18339, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.762 + ], + "angle": 0, + "content": "[36] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. In Advances in Neural Information Processing Systems, pages 37193-37229, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[37] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[38] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[39] Lucas Nunes, Rodrigo Marcuzzi, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Segcontrast: 3d point cloud" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "feature representation learning through self-supervised segment discrimination. IEEE Robotics and Automation Letters, 7(2):2116-2123, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[41] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[42] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[43] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In International Conference on Machine Learning. PMLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[44] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. *Shapellm: Universal 3d object understanding for embodied interaction*. arXiv preprint arXiv:2402.17766, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[45] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In Advances in Neural Information Processing Systems, pages 23192-23204, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[46] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[47] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[48] Zhengxiang Shi and Aldo Lipani. Dept: Decomposed prompt tuning for parameter-efficient fine-tuning. In International Conference on Learning Representations, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[49]Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8791-8800, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[50] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In Advances in Neural Information Processing Systems, pages 12991-13005, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[51] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. In AAAI Conference on Artificial Intelligence, pages 5171-5179, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[52] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, 2019." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6614" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[53] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[54] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In IEEE/CVF International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[55] Song Wang, Jianke Zhu, and Ruixiang Zhang. Metarangeseg: Lidar sequence semantic segmentation using multiple feature aggregation. IEEE Robotics and Automation Letters, 7(4):9739-9746, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[56] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5186-5195, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[57] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14792-14801, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[58] Song Wang, Jiawei Yu, Wentong Li, Hao Shi, Kailun Yang, Junbo Chen, and Jianke Zhu. Label-efficient semantic scene completion with scribble annotations. In International Joint Conference on Artificial Intelligence, pages 1398-1406, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[59] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics, 38:1-12, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[60] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[61] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1912-1920, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[62] Aoran Xiao, Xiaoqin Zhang, Ling Shao, and Shijian Lu. A survey of label-efficient deep learning for 3d point clouds. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[63] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video superresolution. In European Conference on Computer Vision, pages 217-235. Springer, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[64] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In European Conference on Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[65] Yi Xin, Siqi Luo, Haodi Zhou, Junlong Du, Xiaohong Liu, Yue Fan, Qing Li, and Yuntao Du. Parameter-efficient fine-tuning for pre-trained vision models: A survey. arXiv preprint arXiv:2402.02242, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[66] Xiang Xu, Lingdong Kong, Hui Shuai, Wenwei Zhang, Liang Pan, Kai Chen, Ziwei Liu, and Qingshan Liu. 4d" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "contrastive superflows are dense 3d representation learners. In European Conference on Computer Vision, pages 58-80, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[67] Xingyi Yang and Xinchao Wang. Kolmogorov-arnold transformer. In International Conference on Learning Representations, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[68] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics, 35(6):1-12, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[69] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[70] Ted Zadouri, Ahmet Üstün, Arash Ahmadian, Beyza Ermis, Acyr Locatelli, and Sara Hooker. Pushing mixture of experts to the limit: Extremely parameter efficient moe for instruction tuning. arXiv preprint arXiv:2309.05444, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[71] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Annual Meeting of the Association for Computational Linguistics, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[72] Yaohua Zha, Jinpeng Wang, Tao Dai, Bin Chen, Zhi Wang, and Shu-Tao Xia. Instance-aware dynamic prompt tuning for pre-trained point cloud models. In IEEE/CVF International Conference on Computer Vision, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[73] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient fin-tuning. In International Conference on Learning Representations, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[74] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. In Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[75] Shaochen Zhang, Zekun Qi, Runpei Dong, Xiuxiu Bai, and Xing Wei. Positional prompt tuning for efficient 3d representation learning. arXiv preprint arXiv:2408.11567, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[76] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.812 + ], + "angle": 0, + "content": "[77] Xin Zhou, Dingkang Liang, Wei Xu, Xingkui Zhu, Yihan Xu, Zhikang Zou, and Xiang Bai. Dynamic adapter meets prompt tuning: Parameter-efficient transfer learning for point cloud analysis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14707-14717, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.812 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6615" + } + ] +] \ No newline at end of file diff --git a/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/d2184bbd-78d6-4278-aed5-2e0864e5914c_origin.pdf b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/d2184bbd-78d6-4278-aed5-2e0864e5914c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cdef4a64aff03c36fde963a9b1d59a27247fe5a0 --- /dev/null +++ b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/d2184bbd-78d6-4278-aed5-2e0864e5914c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0d653edd9ef20a8dfd00915a22379d8cd2f9a29eeca3beeb37a5e8fee053246 +size 1619485 diff --git a/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/full.md b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fd485ce29213d95d1aa0a11992a1215543929857 --- /dev/null +++ b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/full.md @@ -0,0 +1,360 @@ +# PointLoRA: Low-Rank Adaptation with Token Selection for Point Cloud Learning + +Song Wang $^{1,2}$ , Xiaolu Liu $^{1}$ , Lingdong Kong $^{2}$ , Jianyun Xu $^{3*}$ , Chunyong Hu $^{3}$ , Gongfan Fang $^{2}$ , Wentong Li $^{4}$ , Jianke Zhu $^{1\dagger}$ , Xinchao Wang $^{2\dagger}$ $^{1}$ ZJU NUS AD Lab, CaiNiao, Alibaba NUAA +{songw, jkzhu}@zju.edu.cn, xinchao@nus.edu.sg + +# Abstract + +Self-supervised representation learning for point cloud has demonstrated effectiveness in improving pre-trained model performance across diverse tasks. However, as pretrained models grow in complexity, fully fine-tuning them for downstream applications demands substantial computational and storage resources. Parameter-efficient fine-tuning (PEFT) methods offer a promising solution to mitigate these resource requirements, yet most current approaches rely on complex adapter and prompt mechanisms that increase tunable parameters. In this paper, we propose PointLoRA, a simple yet effective method that combines low-rank adaptation (LoRA) with multi-scale token selection to efficiently fine-tune point cloud models. Our approach embeds LoRA layers within the most parameter-intensive components of point cloud transformers, reducing the need for tunable parameters while enhancing global feature capture. Additionally, multi-scale token selection extracts critical local information to serve as prompts for downstream fine-tuning, effectively complementing the global context captured by LoRA. The experimental results across various pre-trained models and three challenging public datasets demonstrate that our approach achieves competitive performance with only $3.43\%$ of the trainable parameters, making it highly effective for resource-constrained applications. Source code is available at: https://github.com/songw-zju/PointLoRA. + +# 1. Introduction + +3D point cloud learning plays a vital role in computer vision, advancing the understanding and reconstruction of complex 3D scenes [4, 12, 22, 30, 62]. Training deep neural networks on point clouds presents unique challenges due to their unordered structure, sparsity, and irregularity. In recent + +years, point-based methods [29, 33, 41, 42, 45, 64] have made significant progress in addressing these challenges. + +As the volume of available point cloud data grows, there has been a surge in interest toward pre-training on unlabeled point clouds to learn generalizable representations [36, 43, 47, 64, 69, 76]. Robust pre-trained models enable efficient fine-tuning for downstream tasks, reducing dependency on labeled data, accelerating convergence, and improving accuracy [23, 66]. However, conventional full fine-tuning can disrupt pre-trained knowledge, potentially diminishing the model's generalization capability. Moreover, full fine-tuning requires storing multiple versions of model weights, leading to substantial storage costs as datasets and tasks expand, along with increased computational demands for larger pre-trained models. + +Recent studies [51, 72, 75, 77] have made strides in introducing parameter-efficient fine-tuning (PEFT) methods to pre-trained point cloud models, building on approaches widely used in natural language processing (NLP) and 2D vision [9, 65]. As an illustration, IDPT [72] utilizes instance-aware dynamic prompt tuning to enhance model robustness in downstream transfer learning tasks. The subsequent methods [51, 77] combine prompt tuning with adapter tuning, while relying on carefully crafted adapters and specialized prompt design. PPT [75] presents positional prompt tuning, which doubles the sequence length and significantly increases computational requirements. + +To fine-tune in a more efficient manner and further reduce the parameter count, we propose PointLoRA, which leverages low-rank adaptation (LoRA) [20], widely studied in large language models, to enable better fine-tuning for point cloud models. As shown in Fig. 1 (a), during LoRA fine-tuning, instead of directly updating the pretrained weights, two auxiliary low-rank matrices, $W_{u}$ and $W_{d}$ , are injected to update the network with fewer tunable parameters. From a 3D learning perspective, these matrices function as fully connected layers that effectively capture global point cloud features but lack local information extraction, akin to the architecture of PointNet [41]. + +![](images/90c33258060e29642edcfebf82cc9ffa7711557dc1a3df46201a7a87d168ab3b.jpg) +(a) LoRA + +![](images/af3e24347766390db475aaeba5053e394a678020ab4a48b74d885df6826b364c.jpg) +(b) PointLoRA +Figure 1. Comparing our proposed PointLoRA approach against vanilla LoRA methods. Both LoRA and our approach incorporate low-rank adaptation matrices into the pre-trained weights to extract global information from the point cloud sequence. Furthermore, our approach integrates tokens selected at various scales to capture local information, which is refined using a shared Prompt MLP and then output alongside the results derived from the original low-rank matrices. + +Unlike 2D visual or language data, 3D point clouds contain distinctive local geometric features critical for downstream tasks. Additionally, not all local positions contribute equally to the model, as their relevance varies depending on the specific task and data characteristics. To address this, we further design a Multi-Scale Token Selection module, providing local prior information as a prompt to complement LoRA's global feature aggregation capabilities. Using multi-scale sampling and $k$ -nearest neighbor aggregation, we derive tokens that represent features across different scales. A shared Mask Predictor then selects a subset of these tokens to act as prompts. During fine-tuning, tokens from various positions are dynamically selected, processed through a Prompt MLP, and incorporated into the LoRA layer, as illustrated in Fig. 1 (b). This design enables the integration of essential local geometric features with global context, allowing the pre-trained model to adapt effectively to diverse downstream tasks. + +Our main contributions are summarized as follows. + +- We propose PointLoRA, a simple yet effective scheme for parameter-efficient fine-tuning in point cloud learning. +- Low-rank adaptation is injected into the most parameter-intensive components of the point cloud transformer to capture global information, with a corresponding analysis in the context of PointNet provided. +- We further design a multi-scale token selection module that dynamically incorporates local geometric information as prompt, enriching global low-rank adaptation with the crucial local context. +- Extensive experiments across three challenging datasets and two widely used pre-trained models validate the effectiveness of our method, achieving state-of-the-art performance with only $3.43\%$ of the trainable parameters. + +# 2. Related Work + +Model Pre-Training on Point Clouds. Self-supervised pre-training for 3D point clouds has been widely explored [36, 43, 47, 55, 56, 64, 69, 76], generally categorized into contrastive-based approaches [1, 7, 39, 64, 66] and masked signal reconstruction methods [5, 6, 40, 69, 74]. In contrastive learning, PointContrast [64] and CrossPoint [1] leverage different views or instances to extract latent features through contrasting representations. SegContrast [39] extends this approach to outdoor LiDAR data, extracting class-agnostic segments from augmented views [24]. Reconstruction-based methods, such as Point-BERT [69] and Point-MAE [40], employ masked point cloud reconstruction for pre-training. In ACT [10], a pre-trained autoencoder serves as a cross-modality teacher, guiding 3D point cloud learning through knowledge distillation [17, 57, 58]. PointGPT [5] adapts the GPT framework [38] for autoregressive generation on point clouds. Additionally, Qi et al. [43, 44] propose a combination of reconstruction and cross-modal contrastive learning within generative models. While full fine-tuning is commonly applied to these pretrained models for downstream tasks, the high computational costs and potential dilution of pre-trained knowledge motivate our exploration of efficient fine-tuning strategies. + +Parameter-Efficient Fine-Tuning (PEFT). PEFT methods have gained traction in NLP and 2D vision tasks [8, 9, 19, 21, 28, 48, 63, 73] to enhance downstream performance with minimal tunable parameters. Mainstream PEFT approaches include prompt and prefix tuning, adapter-based methods, and low-rank adaptation. Prompt and prefix tuning [21, 25, 28, 48] introduce additional tokens as tunable components, with VPT [21] pioneering the use of prompt tokens for pre-trained vision transformers (ViTs) [11, 26, 49, 67]. Adapter-based tuning inserts trainable modules between frozen layers [8, 19, 27, 50], while LoRA [20] applies a low-rank approximation to update linear layers within + +attention blocks. Building on these techniques, recent innovations explore alternative adapter placements [14, 31], tuning bias terms [71], and combining with Mixture-of-Experts (MoE) techniques [35, 70]. Recent studies have also adapted PEFT to point cloud analysis [32, 51, 72, 75, 77]. IDPT [72] extends prompt tuning to DGCNN [59] for instance-aware prompt extraction. Point-PEFT [51] and DAPT [77] combine prompt tuning with adapters to further improve fine-tuning performance, while PPT [75] introduces positional prompt tuning for efficient 3D representation learning. In contrast to these approaches, our work uniquely integrates low-rank adaptation with multi-scale token selection, achieving competitive performance with a significantly reduced parameter footprint. + +# 3. Preliminary + +In this section, we revisit the architecture of point cloud transformer and the corresponding fine-tuning prototypes. + +# 3.1. Point Cloud Transformer + +Point Tokenizer. Due to the sparsity and irregularity of point clouds, it is necessary to transform them into a token sequence suitable for processing by the transformer. We follow standard settings [40, 69] by segmenting the point cloud into irregular patches using the farthest point sampling (FPS) and the $k$ -nearest neighbors ( $k$ -NN) algorithm. Formally, given the point cloud $P = \{p_1, p_2, \ldots, p_N\} \in \mathbb{R}^{N \times 3}$ with $N$ points, the $g$ group centers are obtained with FPS: + +$$ +C _ {g} = \operatorname {F P S} (P), \quad C _ {g} \in \mathbb {R} ^ {g \times 3}, \tag {1} +$$ + +where $C_g$ is the selected center points. Then, we use $k$ -NN to select $k$ nearest neighbor points for each center in $C_g$ : + +$$ +N _ {p} = k \text {- N N} (P, C _ {g}), \quad N _ {p} \in \mathbb {R} ^ {g \times k \times 3}, \tag {2} +$$ + +where $N_{p}$ is the corresponding neighbor point patch. Notably, these local point patches are made unbiased by subtracting their center points, which promotes better convergence. Then a mini-PointNet [41] is utilized to embed the point patches into discrete tokens $T_{p}$ : + +$$ +T _ {p} = \operatorname {m i n i - P o i n t N e t} \left(N _ {p}\right), \quad T _ {p} \in \mathbb {R} ^ {g \times d}, \tag {3} +$$ + +where $d$ is the embedding dimension. Following the tokenization process, the 3D point cloud is converted into a feature vector, enabling subsequent processing with Transformer architectures similar to those used in natural language processing and 2D vision. + +Transformer Block. With a classification token $T_{\mathrm{cls}}$ and the obtained $T_{p}$ , the input $T_{\mathrm{input}} = \mathrm{Concat}(T_{\mathrm{cls}}, T_{p})$ is further processed by $L$ -layer transformer blocks. In each block, we first project $T_{\mathrm{input}}$ into query, key, and value spaces: + +$$ +\mathbf {Q} = T _ {\text {i n p u t}} \mathbf {W} ^ {Q}, \quad \mathbf {K} = T _ {\text {i n p u t}} \mathbf {W} ^ {K}, \quad \mathbf {V} = T _ {\text {i n p u t}} \mathbf {W} ^ {V}, \tag {4} +$$ + +where $\mathbf{W}^Q, \mathbf{W}^K, \mathbf{W}^V \in \mathbb{R}^{d \times d}$ are trainable matrices in qkv projection layer. The attention layer then computes the self-attention with a skip connection as follows: + +$$ +T _ {\text {o u t p u t}} ^ {\prime} = \text {A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) + T _ {\text {i n p u t}}. \tag {5} +$$ + +Then, the final output $T_{\mathrm{output}}$ is computed as: + +$$ +T _ {\text {o u t p u t}} = \text {F F N} \left(\ln \left(T _ {\text {o u t p u t}} ^ {\prime}\right)\right) + T _ {\text {o u t p u t}} ^ {\prime}, \tag {6} +$$ + +where LN is the layer norm function [2] and FFN is the feed-forward network layer [3]. Through $L$ Transformer blocks, the network extracts deep and essential features from the discrete tokens. Notably, the qkv projection parameters within the attention layer, along with those in the FFN layer constitute a significant portion of the model's parameters and are crucial for performance. + +# 3.2. Fine-Tuning Prototype + +Leveraging well-pretrained models, fine-tuning prototypes to adapt these models to diverse downstream tasks can be categorized as follows. + +Full Fine-Tuning. Full fine-tuning is the predominant approach in current point cloud pre-training research for downstream tasks, involving the direct loading of pretrained model parameters and joint training of the encoder alongside the task-specific head. Although this method offers a theoretically higher accuracy potential, it also entails substantial computational and storage overhead. + +Parameter-Efficient Fine-Tuning. Parameter-efficient fine-tuning adapts pre-trained models to downstream tasks by freezing most of the model's parameters, introducing lightweight modules, and selectively fine-tuning a small subset of parameters. This approach provides a more practical solution, particularly as model sizes continue to increase. Existing methods typically combine adapter tuning, which inserts bottleneck-like layers, and prompt tuning, which introduces prior knowledge, to fine-tune point cloud pre-trained models. + +# 4. PointLoRA + +# 4.1. Overview + +In this work, we aim to provide a simple but effective parameter-efficient fine-tuning method for point cloud pretrained models. The complete framework of our proposed PointLoRA, integrated into the existing point cloud transformer, is illustrated in Fig. 2, addressing the parameter-intensive nature of the transformer-based network design. Our approach can effectively enhance downstream task performance while keeping most of the pre-trained model parameters frozen. + +In the following section, we first outline the establishment of the vanilla LoRA baseline and analyze the factors contributing to its effectiveness in point cloud learning. + +![](images/fb5a7328ee85be8410f9f0e2ce0bba6660df8b6873aeed6c30dbc5f1f4e18103.jpg) +Figure 2. Overview of PointLoRA integrated into point cloud transformer pipeline. Given an input point cloud, we first tokenize it using the original Point Tokenizer and perform token selection across multiple scales (Multi-Scale Token Selection). The tokens from both components are then concatenated and fed into the Transformer Block. Our approach is injected into the qkv projection and FFN layers, utilizing a shared Prompt MLP within these layers to enhance parameter efficiency. + +Then we present how our method extends LoRA's feature extraction capabilities to achieve competitive performance in 3D scenarios. + +# 4.2. Vanilla LoRA Baseline + +As mentioned in Sec. 3.1, the qkv projection and feed-forward network (FFN) layers in point cloud transformer significantly increase the number of parameters, limiting the fine-tuning efficiency and hindering deployment in resource-limited environments. To address this issue, we incorporate Low-Rank Adaptation (LoRA) into these layers by introducing low-rank matrices to adapt pre-trained weights, thereby reducing the number of trainable parameters while preserving model performance. Given a pretrained weight matrix $W_{p}$ , LoRA modifies it as follows: + +$$ +W _ {\text {u p d a t e}} = W _ {p} + \Delta W = W _ {p} + W _ {u} \cdot W _ {d}, \tag {7} +$$ + +where $\Delta W$ and $W_{\mathrm{update}}$ denotes the updated weight and new weight matrix, respectively. $W_{u}\in \mathbb{R}^{d\times r}$ and $W_{d}\in \mathbb{R}^{r\times d}$ are low-rank matrices, where $r\ll d$ indicates the rank, controlling the complexity of the adaptation. + +During training, only $W_{u}$ and $W_{d}$ are updated, while the pre-trained weight $W_{p}$ remains frozen. For inference, the adaptation $\Delta W$ can be merged with the original weights, yielding in a single consolidated weight matrix: $W_{\mathrm{infer}} = W_{p} + \Delta W$ . This consolidation enables the model to retain the benefits of low-rank adaptation while avoiding additional computational overhead in the inference phase. + +Analysis. LoRA's architecture is particularly well-suited for point cloud data, as it can capture the complex relationships between points essential to understanding 3D shapes and spatial arrangements. Its MLP-like structure (i.e., two low-rank matrices) effectively learns global features while aligning with PointNet's principles of handling unordered point + +sets and extracting permutation-invariant features. The synergy between LoRA and PointNet's feature extraction capabilities establishes a robust framework for harnessing the rich information in point cloud data, enhancing both generalization and adaptability in complex environments. + +# 4.3. PointLoRA with Token Selection + +While global features provide a comprehensive understanding of the overall structure of the point cloud, local features capture intricate details vital for accurate downstream task performance. To fully investigate LoRA's capabilities, it is crucial to complement the extracted global features with robust local features. The tokenization process produces a sequence of tokens that capture various aspects of the point cloud data. However, not all tokens are valuable for downstream tasks; some may encapsulate redundant or noninformative content, offering limited contribution to overall model performance. This necessitates a systematic approach for filtering and selecting the most informative tokens to enhance the model's representation of local features. + +To tackle the above concerns, we introduce a MultiScale Token Selection module integrated with vanilla LoRA, which is presented in Fig. 3. This module dynamically selects tokens from the raw point cloud at multiple scales, allowing the model to capture local features at different levels of detail. Additionally, PointLoRA incorporates a shared Prompt MLP to further embed local information from these selected tokens, enhancing the vanilla LoRA's performance. + +Multi-Scale Token Generation. Specifically, we apply farthest point sampling (FPS) on the input point cloud $P$ with varying numbers of centroids (i.e., $N_{1}, N_{2}, \dots, N_{M}$ ) to generate tokens at $M$ different scales. As described in Sec. 3, we can obtain $M$ sets of center points, $C_{g}^{1}, C_{g}^{2}, \dots, C_{g}^{M}$ , + +![](images/306fb50b4823f49e3de30b28f55d5505b0a9a2dc3bd1d02dfe76fb5a056738d5.jpg) +Figure 3. Illustration of Multi-Scale Token Selection. In a two-scale setup, we first sample different numbers of center points, then cluster around each center point and apply Mini-PointNet encoding to generate the corresponding tokens. These tokens are also fed into a Mask Predictor to estimate importance scores, allowing us to select the Top-K tokens at each scale. + +each containing $g_{1}, g_{2}, \ldots, g_{M}$ center points. Subsequently, $k$ -NN is employed to retrieve the corresponding neighbor point patches, which are then embedded into discrete tokens using a shared Mini-PointNet. + +Mask Predictor for Selection. The above process yields $M$ token sets, $T_{p}^{1}, T_{p}^{2}, \ldots, T_{p}^{M}$ , encoding local information at multiple scales. Our goal is to select tokens from these sets that are most beneficial for fine-tuning on downstream tasks. Therefore, a simple but effective Mask Predictor is designed to estimate the importance of each token, producing a score vector $s^{m} \in \mathbb{R}^{g_{m} \times 1}$ for each token in $T_{p}^{m} \in \mathbb{R}^{g_{m} \times d}$ , $m = 1, \ldots, M$ . In particular, the Mask Predictor comprises two multi-layer perceptron (MLP) layers followed by a Sigmoid activation function: + +$$ +s ^ {m} = \operatorname {S i g m o i d} \left(\operatorname {M L P} \left(\left(T _ {p} ^ {m}\right)\right)\right). \tag {8} +$$ + +This design enables the model to assign scores between 0 and 1 to each token, indicating their relevance for downstream tasks. We then select a predefined number $N_{1}^{\prime}, \ldots, N_{M}^{\prime}$ of tokens from each scale based on their importance scores using Top-K selection, resulting in a refined token set $\{S_{\mathrm{p}}^{m}\}_{m=1}^{M}$ : + +$$ +S _ {\mathrm {p}} ^ {m} = \operatorname {T o p K} \left(T _ {p} ^ {m}, N _ {m} ^ {\prime}\right), m = 1, \dots , M. \tag {9} +$$ + +Local Geometry Prompt. With the selected $N_{s} = N_{1}^{\prime} + \ldots + N_{M}^{\prime}$ tokens $S_{p}$ , we can capture crucial local information + +for downstream tasks. These tokens are then concatenated with the original inputs to form a local geometry prompt. In each LoRA layer of the point cloud transformer block, a Prompt MLP with a GELU activation function [16] encodes the new input features alongside the local prompt. This process integrates the encoded features of the selected tokens with the LoRA adaptation output. The resulting combined output can be formulated as follows: + +$$ +O _ {\text {u p d a t e}} = \operatorname {P r o m p t} \mathrm {M L P} \left(T _ {\text {i n p u t}}, S _ {p}\right) + \Delta W \cdot \left(T _ {\text {i n p u t}}, S _ {p}\right), \tag {10} +$$ + +where $O_{\mathrm{update}}$ is the updated amount of the PointLoRA layer. Notably, to further optimize the efficiency of the parameters, the Prompt MLP is configured independently for both the qkv projection and FFN layers, while being shared across all blocks. + +The multi-scale token selection process improves model performance by prioritizing the most relevant local features during fine-tuning. By incorporating this approach, our scheme effectively refines both global and local feature representations, thereby enhancing its adaptability and accuracy across a range of 3D tasks. + +# 4.4. Training & Inference + +Overall Loss Function. When fine-tuning on diverse downstream tasks across various datasets, we follow the common practice of maintaining consistency between the adopted loss and the loss function of the original task. Additionally, we introduce a regularization term to further supervise the Mask Predictor. In general, the total loss function $\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{\mathrm{task}} + \lambda \cdot \mathcal{L}_{\mathrm{mask}}$ , where $\mathcal{L}_{\mathrm{task}}$ is the task-specific loss for classification or segmentation. $\mathcal{L}_{\mathrm{mask}}$ is the regularization loss for mask predictor with the balanced weight $\lambda$ : + +$$ +\mathcal {L} _ {\text {m a s k}} = - \frac {1}{N _ {\text {t o t a l}}} \sum_ {i = 1} ^ {N _ {\text {t o t a l}}} \left(s _ {i} \log \left(s _ {i} + \epsilon\right) + \right. \tag {11} +$$ + +$$ +\left. \left(1 - s _ {i}\right) \log \left(1 - s _ {i} + \epsilon\right)\right), +$$ + +where $N_{\mathrm{total}} = N_1 + N_2 \dots + N_M$ , $s_i$ denotes the importance score for each token and $\epsilon$ is a small constant (e.g., $10^{-6}$ ) to prevent undefined values in the logarithm. + +Inference. During inference, the additional parameters introduced by LoRA in the transformer block are directly merged into the original pre-trained weights, with the only newly added component being a small shared MLP dedicated to embed the selected tokens. + +# 5. Experiments + +# 5.1. Setup + +Datasets. We validate the effectiveness of our approach through extensive experiments on three widely used 3D datasets: ScanObjectNN [52], ModelNet40 [61], and ShapeNetPart [68]. ScanObjectNN [52] is a challenging + +Table 1. Performance comparison on three variants of the ScanObjectNN [52] and the ModelNet40 [61] datasets, respectively, for real-world and synthetic object classification. Both the number of tunable parameters and overall accuracy (OA) are reported. All methods only employ the default data argumentation without voting as the baseline. * denotes results reproduced from the public source code. + +
MethodsPublicationTunable Params.ScanObjectNNModelNet40
OBJ-BGOBJ-ONLYPB-T50-RSPoints Num.OA (%)
Traditional Supervised Learning Only
PointNet [41]CVPR'173.5 M73.379.268.01k89.2
PointNet++ [42]NeurIPS'171.5 M82.384.377.91k90.7
DGCNN [59]TOG'191.8 M82.886.278.11k92.9
MVTN [13]ICCV'2111.2 M--82.81k93.8
PointNeXt [45]NeurIPS'221.4 M--87.71k94.0
PointMLP [37]ICLR'2213.2 M--85.41k94.5
RepSurf-U [46]CVPR'221.5 M--84.31k94.4
ADS [18]ICCV'23---87.51k95.1
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2122.1 M84.8585.5478.791k92.1
Point-BERT [69]CVPR'2222.1 M87.4388.1283.071k93.2
MaskPoint [34]ECCV'2222.1 M89.7089.3084.601k93.8
Point-MAE [40]ECCV'2222.1 M90.0288.2985.181k93.8
Point-M2AE [74]NeurIPS 2215.3 M91.2288.8186.431k94.0
ACT [10]ICLR'2322.1 M93.2991.9188.211k93.7
RECON [43]ICML'2343.6 M94.1593.1289.731k93.9
Self-Supervised Representation Learning (Parameter-Efficient Fine-Tuning)
Point-MAE [40] (Full-FT)ECCV'2222.1 M (100%)90.0288.2985.181k93.2
Point-MAE + IDPT [72]ICCV'231.7 M (7.69%)91.22(+1.20)90.02(+1.73)84.94(-0.24)1k93.3(+0.1)
Point-MAE + DAPT [77]CVPR'241.1 M (4.97%)90.88(+0.86)90.19(+1.90)85.08(-0.10)1k93.5(+0.3)
Point-MAE + PPT* [75]arXiv'241.04 M (4.57%)89.84(-0.18)88.98(+0.69)84.45(-0.73)1k93.2(+0.0)
Point-MAE + PointLoRAOurs0.77 M (3.43%)90.71(+0.69)89.33(+1.04)85.53(+0.35)1k93.3(+0.1)
+ +real-world 3D object classification dataset, which contains approximately 15,000 indoor point cloud instances across 15 categories. The object classification task is performed on three variants of increasing complexity, OBJ-BG, OBJ-ONLY, and PB-T50-RS, each representing different levels of scene realism and occlusion. ModelNet40 [61], a classical synthetic dataset for 3D object recognition, consists of 12,311 meshed 3D CAD objects that span 40 categories. We perform synthetic object classification and few-shot learning experiments on ModelNet40 to evaluate the robustness of the model in both standard and low-resource settings. Additionally, to evaluate performance on detailed structures and component segmentation, we provide experiments on ShapeNetPart [68], a widely used benchmark for part segmentation, comprising 16,881 point-level synthetic objects across 16 object categories and 50 part categories. + +Evaluation Metrics. In the evaluation, overall accuracy (OA) is adopted to assess performance on the 3D object classification task, representing the ratio of correctly classified instances to the total number of instances, thus providing an aggregate score over all classes. For the part segmentation task, we employ mean Intersection over Union (mIoU) to evaluate the overlap between prediction and ground-truth segments, averaged across all classes. + +Implementation Details. Our method can be directly integrated into existing point cloud pre-trained models. Dur + +ing fine-tuning, we follow a commonly used setup, freezing most parameters of the pre-trained model, and primarily updating the newly inserted parameters. The number of scales, $M$ , is set to 2. At these two scales, we apply farthest point sampling (FPS) and $k$ -NN clustering with (128, 32) and (64, 64) center and neighboring points, respectively, and then select tokens with $N_1' = 32$ and $N_2' = 8$ . The balanced weight $\lambda$ for Mask Predictor learning is set to 0.004 For real-world and synthetic object classification, we fine-tune using Point-MAE [40], training for 300 epochs with a learning rate of $5e^{-4}$ and a weight decay of 0.05. For few-shot learning and part segmentation, we utilize the Recon [43] model for fine-tuning to further validate the generalizability of our approach, with a learning rate of $2e^{-4}$ for part segmentation. All experiments are conducted on a single GPU. More implementation details with different baselines are provided in the Supplementary Material. + +# 5.2. Comparative Study + +Real-World & Synthetic Object Classification. We conduct real-world and synthetic object classification on three variants of ScanObjectNN (OBJ-BG, OBJ-ONLY, PB-T50-RS) [52] and ModelNet40 [61], using default data augmentation and without voting. As illustrated in Tab. 1, our proposed approach demonstrates consistent performance improvements, especially on the most challenging vari + +![](images/118a6bffe5490a0c36f67841301bbe781d4ea0774d2749af33ec0314523b6425.jpg) +Figure 4. The t-SNE visualization results on the PB-T50-RS split of the ScanObjectNN dataset [52] with different fine-tuning schemes. We adopt Point-MAE [40] as the baseline model for fair comparison. TP: Number of tunable parameters. OA: Overall accuracy. Symbol * denotes re-produced with official implementation. Best viewed in colors and zoomed-in for additional details. + +![](images/488cca78a4feae37af758eef1855e3ffccafa3efa9fa104c185128d89469deb6.jpg) + +![](images/c5321543229616d63dff5024f18e68f96b69c510fedb3e209af31322c40973f7.jpg) + +![](images/eab0465e9f56b2288f8d2a66c7db0e7ac7fdfd0f995ce7c7220c582629baf689.jpg) + +![](images/5c974cb58e42fde1a0170015bfd386e71c0217df8d6239f6f3fd197dac07d4f2.jpg) + +Table 2. Performance comparison on ModelNet40 [61] for few-shot learning. We report the scores of the overall accuracy $(\%)\pm$ the standard deviation $(\%)$ without voting. The top two highest accuracies are highlighted in bold and underline, respectively. + +
MethodsPublication5-way10-way
10-shot20-shot10-shot20-shot
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2194.0±3.695.9±2.389.4±5.192.4±4.6
Point-BERT [69]CVPR'2294.6±3.196.3±2.791.0±5.492.7±5.1
MaskPoint [34]ECCV'2295.0±3.797.2±1.791.4±4.093.4±3.5
Point-MAE [40]ECCV'2296.3±2.597.8±1.892.6±4.195.0±3.0
Point-M2AE [74]NeurIPS'2296.8±1.898.3±1.492.3±4.595.0±3.0
ACT [10]ICLR'2396.8±2.398.0±1.493.3±4.095.6±2.8
Self-Supervised Representation Learning (Efficient Fine-Tuning)
RECON [43] (Full-FT)ICML'2397.3±1.998.9±3.993.3±3.995.8±3.0
RECON+ IDPT [72]ICCV'2396.9±2.498.3±0.792.8±4.095.5±3.2
RECON+ DAPT [77]CVPR'2495.6±2.897.7±1.691.9±4.194.6±3.5
RECON+ PPT [75]arXiv'2497.0±2.798.7±1.692.2±5.095.6±2.9
RECON+ PointLoRAOurs96.9±2.798.8±1.292.7±4.495.8±2.9
+ +ant, PB-T50-RS, where it is the only parameter-efficient fine-tuning method that surpasses the full fine-tuned PointMAE [40] (85.53% vs. 85.18%) and significantly outperforms the state-of-the-art method, DAPT [77]. Notably, our approach requires only 0.77M tunable parameters, the fewest among all methods. Furthermore, the t-SNE feature visualization [53] of existing methods and PointLoRA is presented in Fig. 4. Our method enables the pre-trained model to fine-tune with a minimal number of parameters, producing distinctive feature representations. + +Few-shot Learning. We also perform few-shot learning experiments on ModelNet40 [61] to validate the transfer learning capability of our method with limited annotations. Following standard protocols [15, 43], we evaluate the fine-tuning performance with ReCon [43] in the 5-way / 10-way and 10-shot / 20-shot settings, respectively. As shown in Tab. 2, our method achieves the best or second-best accuracy in all configurations, further demonstrating the generalizability of the proposed scheme. + +Point Cloud Part Segmentation. The part segmentation results on the ShapeNetPart dataset [68] are provided in + +Table 3. Performance comparison on the ShapeNetPart [68] for part segmentation. Both the mIoU for all classes (Cls.) and instances (Inst.) are provided. TP indicates the number of tunable parameters. $\dagger$ means the result, reported from DAPT [77]. $*$ denotes the results reproduced from the official implementation. + +
MethodsPublicationTPCls. mIoU (%)Inst. mIoU (%)
Traditional Supervised Learning Only
PointNet [41]CVPR'17-80.3983.7
PointNet++ [42]NeurIPS'17-81.8585.1
DGCNN [59]TOG'19-82.3385.2
APES [60]CVPR'23-83.6785.8
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2127.09 M83.4285.1
MaskPoint [34]ECCV'22-84.6086.0
Point-BERT [69]CVPR'2227.09 M84.1185.6
Point-MAE [40]ECCV'2227.06 M84.1986.1
ACT [10]ICLR'2327.06 M84.6686.1
Self-Supervised Representation Learning (Efficient Fine-Tuning)
RECON [43] (Full-FT)ICML'2327.06 M84.5286.1
RECON+ IDPT† [72]ICCV'235.69 M83.6685.7
RECON+ DAPT [77]CVPR'245.65 M83.8785.7
RECON+ PPT* [75]arXiv'245.62 M83.8885.4
RECON+ PointLoRAOurs5.63 M83.9885.4
+ +Tab. 3. In this fine-grained scene understanding task, our approach still achieves competitive performance with a constrained parameter budget. Unlike prior classification models, the increase in the parameter count is primarily attributed to the segmentation head. + +Comparison with Other PEFT Methods. We further present a comparison with existing PEFT methods designed for other tasks, including NLP and 2D vision. We select the most challenging PB-T50-RS variant [52] and use the pretrained Point-MAE model [40] as a baseline. As shown in Tab. 4, these methods designed for other tasks provide limited performance gains when fine-tuned in 3D scenes, despite their parameter efficiency compared to existing 3D fine-tuning approaches. In contrast, our method not only achieves the highest accuracy, but also uses significantly fewer parameters than the three current fine-tuning techniques specifically developed for point clouds, underscoring the strong potential of PointLoRA. + +Table 4. Performance comparison with other parameter-efficient methods designed for NLP and 2D Vision tasks on the hardest variant of ScanObjectNN [52]. + +
MethodsPublicationTPPB-T50-RS
Point-MAE [34]ECCV'2222.1 M85.18
Linear probing-0.3 M75.99
+ Adapter [19]ICML'190.9 M83.93
+ Perfix tuning [28]ACL'210.7 M77.72
+ BitFit [71]ACL'210.3 M82.62
+ LoRA†[20]ICLR'220.9 M81.74
+ VPT-Deep [21]ECCV'220.4 M81.09
+ AdaptFormer [8]NeurIPS'220.9 M83.45
+ SSF [31]NeurIPS'220.4 M82.58
+ IDPT [72]ICCV'231.7 M84.94
+ DAPT [77]CVPR'241.1 M85.08
+ PPT* [75]arXiv'241.04 M84.45
+ PointLoRAOurs0.77 M85.53
+ +Table 5. The impact of each module of our scheme. We provide the overall accuracy (\%) on the hardest variant of ScanObjectNN [52] and corresponding tunable parameters (TP). "MS-FPS" indicates multi-scale furthest point sampling. + +
LoRAToken SelectionMS-FPSTPPB-T50-RS
Full Fine-Tuning22.1 M85.18
Linear Probing0.27 M75.99
✓0.53 M83.83
✓✓0.77 M84.91
✓✓✓0.77 M85.53
+ +# 5.3. Ablation Study + +In this section, we perform exhaustive ablation experiments on the challenging PB-T50-RS variant to investigate the rationale and effectiveness of the design choices of our proposed approach. The pre-trained Point-MAE [40] is adopted as a baseline for a fair comparison. + +Ablation on PointLoRA Scheme. Firstly, we provide the ablations on each module of the proposed method. As illustrated in Tab. 5, directly applying low-rank adaptation (LoRA) produces a significant improvement over linear probing, demonstrating the effectiveness of incorporating global information during the fine-tuning process. Injecting selected tokens as prompts into LoRA further improves accuracy with only a marginal increase in the number of parameters. Finally, the local information obtained from selected tokens across multiple scales complements the global features of LoRA, achieving the best performance. + +Ablation on Low-Rank Adaptation. Then we perform ablation experiments on rank $(r)$ in our approach, as shown in Tab. 6. Different ranks have a noticeable impact on finetuning performance. Taking into account both parameter count and accuracy, we set the rank to 8 to maximize the extraction of global information for fine-tuning. + +Ablation on Multi-Scale Token Selection. Given the critical improvement from multi-scale token selection for PointLoRA, we first perform ablation experiments on the number of tokens selected at the same scale. As shown in + +Table 6. Ablation study on rank $(r)$ of the proposed method. + +
Rank rTPRatioPB-T50-RS
40.66 M2.96%84.28
80.77 M3.43%85.53
160.99 M4.37%85.15
321.44 M6.32%84.87
+ +Table 7. Ablation study on scale number $M$ and selected token number $N_{s}$ in Multi-Scale Token Selection. + +
Scale MToken Num. NsPB-T50-RS
16484.80
13284.91
11685.22
264 & 1684.39
232 & 885.53
216 & 484.49
332 & 16 & 885.05
+ +Table 8. Ablation study on the dimension of Prompt MLP. + +
DimensionTPRatioPB-T50-RS
80.68 M3.03%83.55
160.71 M3.17%84.77
320.77 M3.43%85.53
640.90 M3.96%85.25
+ +the upper part of Tab. 7, selecting 16 tokens at the original scale yields the best results, demonstrating the effectiveness of token selection. Furthermore, when incorporating tokens from multiple scales, optimal performance is achieved by selecting 32 and 8 tokens from two different scales, as shown in the lower part of Tab. 7. + +Ablation on Prompt MLP. We provide the ablation study on the dimension of the shared Prompt MLP, as this affects the embedding of local information in our method. As illustrated in Tab. 8, setting this dimension to 32 yields the best fine-tuning performance. Additionally, adjusting the dimension causes only minimal changes in tunable parameter counts, suggesting flexibility in selecting this hyperparameter for different datasets and tasks. + +# 6. Conclusion + +In this paper, a simple yet effective parameter-efficient finetuning approach named PointLoRA, is presented with low-rank adaptation and multi-scale token selection for point cloud models. Low-rank adaptation efficiently reduces the parameter-heavy components of the point cloud transformer architecture while capturing global information. Multi-scale token selection effectively encodes essential local features as prompt, further enhancing the finetuning process. The integration of global and local information enables our approach to achieve state-of-the-art results on the most challenging dataset with a minimal number of tunable parameters while also delivering competitive performance across other datasets and models. + +# Acknowledgments + +This work is supported in part by the National Natural Science Foundation of China under Grant No. 62376244, in part by the Zhejiang Provincial Natural Science Foundation of China under Grant No. LD24F030001, by the Information Technology Center and State Key Lab of CAD&CG, Zhejiang University, and by the National Research Foundation, Singapore, and Cyber Security Agency of Singapore under its National Cybersecurity R&D Programme and CyberSG R&D Cyber Research Programme Office (Award: CRPO-GC1-NTU-002). Lingdong Kong is supported by the Apple Scholars in AI/ML Ph.D. Fellowship program. + +# References + +[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. +[2] Jimmy Lei Ba. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. +[3] George Bebis and Michael Georgiopoulos. Feed-forward neural networks. IEEE Potentials, 13(4):27-31, 1994. +[4] Saifullahi Aminu Bello, Shangshu Yu, Cheng Wang, Jibril Muhammad Adam, and Jonathan Li. Deep learning on 3d point clouds. Remote Sensing, 12(11):1729, 2020. +[5] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. In Advances in Neural Information Processing Systems, 2024. +[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. +[7] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023. +[8] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In Advances in Neural Information Processing Systems, 2022. +[9] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023. +[10] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In International Conference on Learning Representations, 2022. + +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. +[12] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020. +[13] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In IEEE/CVF International Conference on Computer Vision, pages 1-11, 2021. +[14] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations, 2021. +[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[16] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. +[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. +[18] Cheng-Yao Hong, Yu-Ying Chou, and Tyng-Luh Liu. Attention discriminant sampling for point clouds. In IEEE/CVF International Conference on Computer Vision, pages 14429-14440, 2023. +[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning. PMLR, 2019. +[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. +[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, 2022. +[22] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023. +[23] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023. +[24] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023. + +[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Empirical Methods in Natural Language Processing, pages 3045-3059, 2021. +[26] Bonan Li, Yinhan Hu, Xuecheng Nie, Congying Han, Xi-angjian Jiang, Tiande Guo, and Luoqi Liu. Dropkey for vision transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22700–22709, 2023. +[27] Minglei Li, Peng Ye, Yongqi Huang, Lin Zhang, Tao Chen, Tong He, Jiayuan Fan, and Wanli Ouyang. Adapter-x: A novel general parameter-efficient fine-tuning framework for vision. arXiv preprint arXiv:2406.03051, 2024. +[28] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Annual Meeting of the Association for Computational Linguistics, 2021. +[29] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In Advances in Neural Information Processing Systems, 2018. +[30] Ye Li, Lingdong Kong, Hanjiang Hu, Xiaohao Xu, and Xiaonan Huang. Is your lidar placement optimized for 3d scene understanding? In Advances in Neural Information Processing Systems, pages 34980-35017, 2024. +[31] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In Advances in Neural Information Processing Systems, 2022. +[32] Dingkang Liang, Tianrui Feng, Xin Zhou, Yumeng Zhang, Zhikang Zou, and Xiang Bai. Parameter-efficient fine-tuning in spectral domain for point cloud learning. arXiv preprint arXiv:2410.08114, 2024. +[33] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. +[34] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, 2022. +[35] Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Derong Xu, Feng Tian, and Yefeng Zheng. Moelora: An moe-based parameter efficient fine-tuning method for multitask medical applications. arXiv preprint arXiv:2310.18339, 2023. +[36] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. In Advances in Neural Information Processing Systems, pages 37193-37229, 2023. +[37] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2022. +[38] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020. +[39] Lucas Nunes, Rodrigo Marcuzzi, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Segcontrast: 3d point cloud + +feature representation learning through self-supervised segment discrimination. IEEE Robotics and Automation Letters, 7(2):2116-2123, 2022. +[40] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European Conference on Computer Vision, 2022. +[41] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. +[42] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems, 2017. +[43] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In International Conference on Machine Learning. PMLR, 2023. +[44] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. *Shapellm: Universal 3d object understanding for embodied interaction*. arXiv preprint arXiv:2402.17766, 2024. +[45] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In Advances in Neural Information Processing Systems, pages 23192-23204, 2022. +[46] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[47] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022. +[48] Zhengxiang Shi and Aldo Lipani. Dept: Decomposed prompt tuning for parameter-efficient fine-tuning. In International Conference on Learning Representations, 2024. +[49]Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8791-8800, 2022. +[50] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In Advances in Neural Information Processing Systems, pages 12991-13005, 2022. +[51] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. In AAAI Conference on Artificial Intelligence, pages 5171-5179, 2024. +[52] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, 2019. + +[53] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 2008. +[54] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In IEEE/CVF International Conference on Computer Vision, 2021. +[55] Song Wang, Jianke Zhu, and Ruixiang Zhang. Metarangeseg: Lidar sequence semantic segmentation using multiple feature aggregation. IEEE Robotics and Automation Letters, 7(4):9739-9746, 2022. +[56] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5186-5195, 2023. +[57] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14792-14801, 2024. +[58] Song Wang, Jiawei Yu, Wentong Li, Hao Shi, Kailun Yang, Junbo Chen, and Jianke Zhu. Label-efficient semantic scene completion with scribble annotations. In International Joint Conference on Artificial Intelligence, pages 1398-1406, 2024. +[59] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics, 38:1-12, 2019. +[60] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. +[61] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1912-1920, 2015. +[62] Aoran Xiao, Xiaoqin Zhang, Ling Shao, and Shijian Lu. A survey of label-efficient deep learning for 3d point clouds. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. +[63] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video superresolution. In European Conference on Computer Vision, pages 217-235. Springer, 2024. +[64] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In European Conference on Computer Vision, 2020. +[65] Yi Xin, Siqi Luo, Haodi Zhou, Junlong Du, Xiaohong Liu, Yue Fan, Qing Li, and Yuntao Du. Parameter-efficient fine-tuning for pre-trained vision models: A survey. arXiv preprint arXiv:2402.02242, 2024. +[66] Xiang Xu, Lingdong Kong, Hui Shuai, Wenwei Zhang, Liang Pan, Kai Chen, Ziwei Liu, and Qingshan Liu. 4d + +contrastive superflows are dense 3d representation learners. In European Conference on Computer Vision, pages 58-80, 2024. +[67] Xingyi Yang and Xinchao Wang. Kolmogorov-arnold transformer. In International Conference on Learning Representations, 2025. +[68] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics, 35(6):1-12, 2016. +[69] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[70] Ted Zadouri, Ahmet Üstün, Arash Ahmadian, Beyza Ermis, Acyr Locatelli, and Sara Hooker. Pushing mixture of experts to the limit: Extremely parameter efficient moe for instruction tuning. arXiv preprint arXiv:2309.05444, 2023. +[71] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Annual Meeting of the Association for Computational Linguistics, 2022. +[72] Yaohua Zha, Jinpeng Wang, Tao Dai, Bin Chen, Zhi Wang, and Shu-Tao Xia. Instance-aware dynamic prompt tuning for pre-trained point cloud models. In IEEE/CVF International Conference on Computer Vision, 2023. +[73] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient fin-tuning. In International Conference on Learning Representations, 2023. +[74] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. In Advances in Neural Information Processing Systems, 2022. +[75] Shaochen Zhang, Zekun Qi, Runpei Dong, Xiuxiu Bai, and Xing Wei. Positional prompt tuning for efficient 3d representation learning. arXiv preprint arXiv:2408.11567, 2024. +[76] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024. +[77] Xin Zhou, Dingkang Liang, Wei Xu, Xingkui Zhu, Yihan Xu, Zhikang Zou, and Xiang Bai. Dynamic adapter meets prompt tuning: Parameter-efficient transfer learning for point cloud analysis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14707-14717, 2024. \ No newline at end of file diff --git a/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/images.zip b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5b7c30f294adda4615bc59b8cb29b2915fc88e2c --- /dev/null +++ b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd529987b0a51e37d3f575166be15e5da2141f9097d329939b2493cd276fcac2 +size 676850 diff --git a/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/layout.json b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..872911261b7387e9d0b48a6fb9221d01ef6f203c --- /dev/null +++ b/2025/PointLoRA_ Low-Rank Adaptation with Token Selection for Point Cloud Learning/layout.json @@ -0,0 +1,9551 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 103, + 484, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 103, + 484, + 139 + ], + "spans": [ + { + "bbox": [ + 126, + 103, + 484, + 139 + ], + "type": "text", + "content": "PointLoRA: Low-Rank Adaptation with Token Selection for Point Cloud Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "text", + "content": "Song Wang" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "text", + "content": ", Xiaolu Liu" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "text", + "content": ", Lingdong Kong" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "text", + "content": ", Jianyun Xu" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "text", + "content": ", Chunyong Hu" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "text", + "content": ", Gongfan Fang" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "text", + "content": ", Wentong Li" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "text", + "content": ", Jianke Zhu" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "text", + "content": ", Xinchao Wang" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 160, + 505, + 220 + ], + "type": "text", + "content": "ZJU NUS AD Lab, CaiNiao, Alibaba NUAA \n{songw, jkzhu}@zju.edu.cn, xinchao@nus.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "text", + "content": "Self-supervised representation learning for point cloud has demonstrated effectiveness in improving pre-trained model performance across diverse tasks. However, as pretrained models grow in complexity, fully fine-tuning them for downstream applications demands substantial computational and storage resources. Parameter-efficient fine-tuning (PEFT) methods offer a promising solution to mitigate these resource requirements, yet most current approaches rely on complex adapter and prompt mechanisms that increase tunable parameters. In this paper, we propose PointLoRA, a simple yet effective method that combines low-rank adaptation (LoRA) with multi-scale token selection to efficiently fine-tune point cloud models. Our approach embeds LoRA layers within the most parameter-intensive components of point cloud transformers, reducing the need for tunable parameters while enhancing global feature capture. Additionally, multi-scale token selection extracts critical local information to serve as prompts for downstream fine-tuning, effectively complementing the global context captured by LoRA. The experimental results across various pre-trained models and three challenging public datasets demonstrate that our approach achieves competitive performance with only " + }, + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "inline_equation", + "content": "3.43\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 582 + ], + "type": "text", + "content": " of the trainable parameters, making it highly effective for resource-constrained applications. Source code is available at: https://github.com/songw-zju/PointLoRA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "type": "text", + "content": "3D point cloud learning plays a vital role in computer vision, advancing the understanding and reconstruction of complex 3D scenes [4, 12, 22, 30, 62]. Training deep neural networks on point clouds presents unique challenges due to their unordered structure, sparsity, and irregularity. In recent" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 246, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 270 + ], + "type": "text", + "content": "years, point-based methods [29, 33, 41, 42, 45, 64] have made significant progress in addressing these challenges." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 270, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 270, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 270, + 555, + 426 + ], + "type": "text", + "content": "As the volume of available point cloud data grows, there has been a surge in interest toward pre-training on unlabeled point clouds to learn generalizable representations [36, 43, 47, 64, 69, 76]. Robust pre-trained models enable efficient fine-tuning for downstream tasks, reducing dependency on labeled data, accelerating convergence, and improving accuracy [23, 66]. However, conventional full fine-tuning can disrupt pre-trained knowledge, potentially diminishing the model's generalization capability. Moreover, full fine-tuning requires storing multiple versions of model weights, leading to substantial storage costs as datasets and tasks expand, along with increased computational demands for larger pre-trained models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 556, + 569 + ], + "type": "text", + "content": "Recent studies [51, 72, 75, 77] have made strides in introducing parameter-efficient fine-tuning (PEFT) methods to pre-trained point cloud models, building on approaches widely used in natural language processing (NLP) and 2D vision [9, 65]. As an illustration, IDPT [72] utilizes instance-aware dynamic prompt tuning to enhance model robustness in downstream transfer learning tasks. The subsequent methods [51, 77] combine prompt tuning with adapter tuning, while relying on carefully crafted adapters and specialized prompt design. PPT [75] presents positional prompt tuning, which doubles the sequence length and significantly increases computational requirements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 556, + 713 + ], + "type": "text", + "content": "To fine-tune in a more efficient manner and further reduce the parameter count, we propose PointLoRA, which leverages low-rank adaptation (LoRA) [20], widely studied in large language models, to enable better fine-tuning for point cloud models. As shown in Fig. 1 (a), during LoRA fine-tuning, instead of directly updating the pretrained weights, two auxiliary low-rank matrices, " + }, + { + "bbox": [ + 312, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "W_{u}" + }, + { + "bbox": [ + 312, + 570, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "W_{d}" + }, + { + "bbox": [ + 312, + 570, + 556, + 713 + ], + "type": "text", + "content": ", are injected to update the network with fewer tunable parameters. From a 3D learning perspective, these matrices function as fully connected layers that effectively capture global point cloud features but lack local information extraction, akin to the architecture of PointNet [41]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 119, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 119, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 119, + 703 + ], + "type": "text", + "content": "* Project leader." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 147, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 147, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 147, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6605" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 86, + 267, + 223 + ], + "blocks": [ + { + "bbox": [ + 76, + 71, + 112, + 83 + ], + "lines": [ + { + "bbox": [ + 76, + 71, + 112, + 83 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 112, + 83 + ], + "type": "text", + "content": "(a) LoRA" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 86, + 267, + 223 + ], + "lines": [ + { + "bbox": [ + 78, + 86, + 267, + 223 + ], + "spans": [ + { + "bbox": [ + 78, + 86, + 267, + 223 + ], + "type": "image", + "image_path": "90c33258060e29642edcfebf82cc9ffa7711557dc1a3df46201a7a87d168ab3b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 285, + 87, + 532, + 223 + ], + "blocks": [ + { + "bbox": [ + 282, + 71, + 337, + 83 + ], + "lines": [ + { + "bbox": [ + 282, + 71, + 337, + 83 + ], + "spans": [ + { + "bbox": [ + 282, + 71, + 337, + 83 + ], + "type": "text", + "content": "(b) PointLoRA" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 285, + 87, + 532, + 223 + ], + "lines": [ + { + "bbox": [ + 285, + 87, + 532, + 223 + ], + "spans": [ + { + "bbox": [ + 285, + 87, + 532, + 223 + ], + "type": "image", + "image_path": "af3e24347766390db475aaeba5053e394a678020ab4a48b74d885df6826b364c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 224, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 224, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 555, + 269 + ], + "type": "text", + "content": "Figure 1. Comparing our proposed PointLoRA approach against vanilla LoRA methods. Both LoRA and our approach incorporate low-rank adaptation matrices into the pre-trained weights to extract global information from the point cloud sequence. Furthermore, our approach integrates tokens selected at various scales to capture local information, which is refined using a shared Prompt MLP and then output alongside the results derived from the original low-rank matrices." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 491 + ], + "type": "text", + "content": "Unlike 2D visual or language data, 3D point clouds contain distinctive local geometric features critical for downstream tasks. Additionally, not all local positions contribute equally to the model, as their relevance varies depending on the specific task and data characteristics. To address this, we further design a Multi-Scale Token Selection module, providing local prior information as a prompt to complement LoRA's global feature aggregation capabilities. Using multi-scale sampling and " + }, + { + "bbox": [ + 56, + 276, + 296, + 491 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 276, + 296, + 491 + ], + "type": "text", + "content": "-nearest neighbor aggregation, we derive tokens that represent features across different scales. A shared Mask Predictor then selects a subset of these tokens to act as prompts. During fine-tuning, tokens from various positions are dynamically selected, processed through a Prompt MLP, and incorporated into the LoRA layer, as illustrated in Fig. 1 (b). This design enables the integration of essential local geometric features with global context, allowing the pre-trained model to adapt effectively to diverse downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 511, + 275, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 511, + 275, + 523 + ], + "spans": [ + { + "bbox": [ + 67, + 511, + 275, + 523 + ], + "type": "text", + "content": "Our main contributions are summarized as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 567 + ], + "type": "text", + "content": "- We propose PointLoRA, a simple yet effective scheme for parameter-efficient fine-tuning in point cloud learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 568, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 614 + ], + "type": "text", + "content": "- Low-rank adaptation is injected into the most parameter-intensive components of the point cloud transformer to capture global information, with a corresponding analysis in the context of PointNet provided." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 616, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 662 + ], + "type": "text", + "content": "- We further design a multi-scale token selection module that dynamically incorporates local geometric information as prompt, enriching global low-rank adaptation with the crucial local context." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "content": "- Extensive experiments across three challenging datasets and two widely used pre-trained models validate the effectiveness of our method, achieving state-of-the-art performance with only " + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "inline_equation", + "content": "3.43\\%" + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "content": " of the trainable parameters." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 274, + 401, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 401, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 401, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 294, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 569 + ], + "type": "text", + "content": "Model Pre-Training on Point Clouds. Self-supervised pre-training for 3D point clouds has been widely explored [36, 43, 47, 55, 56, 64, 69, 76], generally categorized into contrastive-based approaches [1, 7, 39, 64, 66] and masked signal reconstruction methods [5, 6, 40, 69, 74]. In contrastive learning, PointContrast [64] and CrossPoint [1] leverage different views or instances to extract latent features through contrasting representations. SegContrast [39] extends this approach to outdoor LiDAR data, extracting class-agnostic segments from augmented views [24]. Reconstruction-based methods, such as Point-BERT [69] and Point-MAE [40], employ masked point cloud reconstruction for pre-training. In ACT [10], a pre-trained autoencoder serves as a cross-modality teacher, guiding 3D point cloud learning through knowledge distillation [17, 57, 58]. PointGPT [5] adapts the GPT framework [38] for autoregressive generation on point clouds. Additionally, Qi et al. [43, 44] propose a combination of reconstruction and cross-modal contrastive learning within generative models. While full fine-tuning is commonly applied to these pretrained models for downstream tasks, the high computational costs and potential dilution of pre-trained knowledge motivate our exploration of efficient fine-tuning strategies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Parameter-Efficient Fine-Tuning (PEFT). PEFT methods have gained traction in NLP and 2D vision tasks [8, 9, 19, 21, 28, 48, 63, 73] to enhance downstream performance with minimal tunable parameters. Mainstream PEFT approaches include prompt and prefix tuning, adapter-based methods, and low-rank adaptation. Prompt and prefix tuning [21, 25, 28, 48] introduce additional tokens as tunable components, with VPT [21] pioneering the use of prompt tokens for pre-trained vision transformers (ViTs) [11, 26, 49, 67]. Adapter-based tuning inserts trainable modules between frozen layers [8, 19, 27, 50], while LoRA [20] applies a low-rank approximation to update linear layers within" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6606" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "attention blocks. Building on these techniques, recent innovations explore alternative adapter placements [14, 31], tuning bias terms [71], and combining with Mixture-of-Experts (MoE) techniques [35, 70]. Recent studies have also adapted PEFT to point cloud analysis [32, 51, 72, 75, 77]. IDPT [72] extends prompt tuning to DGCNN [59] for instance-aware prompt extraction. Point-PEFT [51] and DAPT [77] combine prompt tuning with adapters to further improve fine-tuning performance, while PPT [75] introduces positional prompt tuning for efficient 3D representation learning. In contrast to these approaches, our work uniquely integrates low-rank adaptation with multi-scale token selection, achieving competitive performance with a significantly reduced parameter footprint." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 249, + 133, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 133, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 133, + 262 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 269, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 294 + ], + "type": "text", + "content": "In this section, we revisit the architecture of point cloud transformer and the corresponding fine-tuning prototypes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 299, + 197, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 197, + 311 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 197, + 311 + ], + "type": "text", + "content": "3.1. Point Cloud Transformer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "text", + "content": "Point Tokenizer. Due to the sparsity and irregularity of point clouds, it is necessary to transform them into a token sequence suitable for processing by the transformer. We follow standard settings [40, 69] by segmenting the point cloud into irregular patches using the farthest point sampling (FPS) and the " + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "text", + "content": "-nearest neighbors (" + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "text", + "content": "-NN) algorithm. Formally, given the point cloud " + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "inline_equation", + "content": "P = \\{p_1, p_2, \\ldots, p_N\\} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "text", + "content": " points, the " + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 316, + 296, + 422 + ], + "type": "text", + "content": " group centers are obtained with FPS:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 423, + 294, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 423, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 110, + 423, + 294, + 437 + ], + "type": "interline_equation", + "content": "C _ {g} = \\operatorname {F P S} (P), \\quad C _ {g} \\in \\mathbb {R} ^ {g \\times 3}, \\tag {1}", + "image_path": "292cbc80554b0f963e610bcad50a683fb69a20947497c34374a209bd98e1a9b4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 441, + 295, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 466 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 441, + 295, + 466 + ], + "type": "inline_equation", + "content": "C_g" + }, + { + "bbox": [ + 55, + 441, + 295, + 466 + ], + "type": "text", + "content": " is the selected center points. Then, we use " + }, + { + "bbox": [ + 55, + 441, + 295, + 466 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 441, + 295, + 466 + ], + "type": "text", + "content": "-NN to select " + }, + { + "bbox": [ + 55, + 441, + 295, + 466 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 441, + 295, + 466 + ], + "type": "text", + "content": " nearest neighbor points for each center in " + }, + { + "bbox": [ + 55, + 441, + 295, + 466 + ], + "type": "inline_equation", + "content": "C_g" + }, + { + "bbox": [ + 55, + 441, + 295, + 466 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 469, + 294, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 469, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 94, + 469, + 294, + 484 + ], + "type": "interline_equation", + "content": "N _ {p} = k \\text {- N N} (P, C _ {g}), \\quad N _ {p} \\in \\mathbb {R} ^ {g \\times k \\times 3}, \\tag {2}", + "image_path": "935afc813ccf3ff53c8b7975255b235004dc7661ce1a895d83e114b5edadf5af.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": " is the corresponding neighbor point patch. Notably, these local point patches are made unbiased by subtracting their center points, which promotes better convergence. Then a mini-PointNet [41] is utilized to embed the point patches into discrete tokens " + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 555, + 294, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 555, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 88, + 555, + 294, + 570 + ], + "type": "interline_equation", + "content": "T _ {p} = \\operatorname {m i n i - P o i n t N e t} \\left(N _ {p}\\right), \\quad T _ {p} \\in \\mathbb {R} ^ {g \\times d}, \\tag {3}", + "image_path": "a9221e3fdca33ffbe3f13dab32e13dc1f597a67092fea4d6fa4a5c40e7e9823c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 574, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 574, + 296, + 635 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 574, + 296, + 635 + ], + "type": "text", + "content": " is the embedding dimension. Following the tokenization process, the 3D point cloud is converted into a feature vector, enabling subsequent processing with Transformer architectures similar to those used in natural language processing and 2D vision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "text", + "content": "Transformer Block. With a classification token " + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "text", + "content": " and the obtained " + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "inline_equation", + "content": "T_{p}" + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "text", + "content": ", the input " + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{input}} = \\mathrm{Concat}(T_{\\mathrm{cls}}, T_{p})" + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "text", + "content": " is further processed by " + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "text", + "content": "-layer transformer blocks. In each block, we first project " + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{input}}" + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "text", + "content": " into query, key, and value spaces:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 64, + 687, + 294, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 687, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 64, + 687, + 294, + 713 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = T _ {\\text {i n p u t}} \\mathbf {W} ^ {Q}, \\quad \\mathbf {K} = T _ {\\text {i n p u t}} \\mathbf {W} ^ {K}, \\quad \\mathbf {V} = T _ {\\text {i n p u t}} \\mathbf {W} ^ {V}, \\tag {4}", + "image_path": "59cfa0a319fccba7302d95453bdca3d5c3e2580910a5b87d134bfea028e10b59.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^Q, \\mathbf{W}^K, \\mathbf{W}^V \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " are trainable matrices in qkv projection layer. The attention layer then computes the self-attention with a skip connection as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 352, + 120, + 553, + 134 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 120, + 553, + 134 + ], + "spans": [ + { + "bbox": [ + 352, + 120, + 553, + 134 + ], + "type": "interline_equation", + "content": "T _ {\\text {o u t p u t}} ^ {\\prime} = \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) + T _ {\\text {i n p u t}}. \\tag {5}", + "image_path": "a996d614b31ce3dfbcb404c360c53df2ee6fba9e47358561c1e51052173a5cab.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 137, + 494, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 137, + 494, + 149 + ], + "spans": [ + { + "bbox": [ + 314, + 137, + 494, + 149 + ], + "type": "text", + "content": "Then, the final output " + }, + { + "bbox": [ + 314, + 137, + 494, + 149 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{output}}" + }, + { + "bbox": [ + 314, + 137, + 494, + 149 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 156, + 553, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 156, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 354, + 156, + 553, + 171 + ], + "type": "interline_equation", + "content": "T _ {\\text {o u t p u t}} = \\text {F F N} \\left(\\ln \\left(T _ {\\text {o u t p u t}} ^ {\\prime}\\right)\\right) + T _ {\\text {o u t p u t}} ^ {\\prime}, \\tag {6}", + "image_path": "94e8352d3627ffb606dcbe8c74923fe4229aeaa09c43e3c7221408c553b03af6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 176, + 554, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 554, + 260 + ], + "type": "text", + "content": "where LN is the layer norm function [2] and FFN is the feed-forward network layer [3]. Through " + }, + { + "bbox": [ + 313, + 176, + 554, + 260 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 176, + 554, + 260 + ], + "type": "text", + "content": " Transformer blocks, the network extracts deep and essential features from the discrete tokens. Notably, the qkv projection parameters within the attention layer, along with those in the FFN layer constitute a significant portion of the model's parameters and are crucial for performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 266, + 443, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 443, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 443, + 279 + ], + "type": "text", + "content": "3.2. Fine-Tuning Prototype" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 284, + 554, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 554, + 319 + ], + "type": "text", + "content": "Leveraging well-pretrained models, fine-tuning prototypes to adapt these models to diverse downstream tasks can be categorized as follows." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 319, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 403 + ], + "type": "text", + "content": "Full Fine-Tuning. Full fine-tuning is the predominant approach in current point cloud pre-training research for downstream tasks, involving the direct loading of pretrained model parameters and joint training of the encoder alongside the task-specific head. Although this method offers a theoretically higher accuracy potential, it also entails substantial computational and storage overhead." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 403, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 523 + ], + "type": "text", + "content": "Parameter-Efficient Fine-Tuning. Parameter-efficient fine-tuning adapts pre-trained models to downstream tasks by freezing most of the model's parameters, introducing lightweight modules, and selectively fine-tuning a small subset of parameters. This approach provides a more practical solution, particularly as model sizes continue to increase. Existing methods typically combine adapter tuning, which inserts bottleneck-like layers, and prompt tuning, which introduces prior knowledge, to fine-tune point cloud pre-trained models." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 533, + 388, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 533, + 388, + 545 + ], + "spans": [ + { + "bbox": [ + 314, + 533, + 388, + 545 + ], + "type": "text", + "content": "4. PointLoRA" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 552, + 382, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 382, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 382, + 564 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 570, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 676 + ], + "type": "text", + "content": "In this work, we aim to provide a simple but effective parameter-efficient fine-tuning method for point cloud pretrained models. The complete framework of our proposed PointLoRA, integrated into the existing point cloud transformer, is illustrated in Fig. 2, addressing the parameter-intensive nature of the transformer-based network design. Our approach can effectively enhance downstream task performance while keeping most of the pre-trained model parameters frozen." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "In the following section, we first outline the establishment of the vanilla LoRA baseline and analyze the factors contributing to its effectiveness in point cloud learning." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6607" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 230 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 230 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 230 + ], + "type": "image", + "image_path": "fb5a7328ee85be8410f9f0e2ce0bba6660df8b6873aeed6c30dbc5f1f4e18103.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 235, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 555, + 281 + ], + "type": "text", + "content": "Figure 2. Overview of PointLoRA integrated into point cloud transformer pipeline. Given an input point cloud, we first tokenize it using the original Point Tokenizer and perform token selection across multiple scales (Multi-Scale Token Selection). The tokens from both components are then concatenated and fed into the Transformer Block. Our approach is injected into the qkv projection and FFN layers, utilizing a shared Prompt MLP within these layers to enhance parameter efficiency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 335 + ], + "type": "text", + "content": "Then we present how our method extends LoRA's feature extraction capabilities to achieve competitive performance in 3D scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 185, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 185, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 185, + 355 + ], + "type": "text", + "content": "4.2. Vanilla LoRA Baseline" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 479 + ], + "type": "text", + "content": "As mentioned in Sec. 3.1, the qkv projection and feed-forward network (FFN) layers in point cloud transformer significantly increase the number of parameters, limiting the fine-tuning efficiency and hindering deployment in resource-limited environments. To address this issue, we incorporate Low-Rank Adaptation (LoRA) into these layers by introducing low-rank matrices to adapt pre-trained weights, thereby reducing the number of trainable parameters while preserving model performance. Given a pretrained weight matrix " + }, + { + "bbox": [ + 55, + 361, + 296, + 479 + ], + "type": "inline_equation", + "content": "W_{p}" + }, + { + "bbox": [ + 55, + 361, + 296, + 479 + ], + "type": "text", + "content": ", LoRA modifies it as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 489, + 295, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 489, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 91, + 489, + 295, + 502 + ], + "type": "interline_equation", + "content": "W _ {\\text {u p d a t e}} = W _ {p} + \\Delta W = W _ {p} + W _ {u} \\cdot W _ {d}, \\tag {7}", + "image_path": "fa1574df655aff5bb08aa345f4c0c08aa797c5abdc8fc7516ea6241eae9b09fe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{update}}" + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "content": " denotes the updated weight and new weight matrix, respectively. " + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "inline_equation", + "content": "W_{u}\\in \\mathbb{R}^{d\\times r}" + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "inline_equation", + "content": "W_{d}\\in \\mathbb{R}^{r\\times d}" + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "content": " are low-rank matrices, where " + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "inline_equation", + "content": "r\\ll d" + }, + { + "bbox": [ + 55, + 510, + 295, + 557 + ], + "type": "text", + "content": " indicates the rank, controlling the complexity of the adaptation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "text", + "content": "During training, only " + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "inline_equation", + "content": "W_{u}" + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "inline_equation", + "content": "W_{d}" + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "text", + "content": " are updated, while the pre-trained weight " + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "inline_equation", + "content": "W_{p}" + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "text", + "content": " remains frozen. For inference, the adaptation " + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "text", + "content": " can be merged with the original weights, yielding in a single consolidated weight matrix: " + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{infer}} = W_{p} + \\Delta W" + }, + { + "bbox": [ + 55, + 558, + 295, + 642 + ], + "type": "text", + "content": ". This consolidation enables the model to retain the benefits of low-rank adaptation while avoiding additional computational overhead in the inference phase." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Analysis. LoRA's architecture is particularly well-suited for point cloud data, as it can capture the complex relationships between points essential to understanding 3D shapes and spatial arrangements. Its MLP-like structure (i.e., two low-rank matrices) effectively learns global features while aligning with PointNet's principles of handling unordered point" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 361 + ], + "type": "text", + "content": "sets and extracting permutation-invariant features. The synergy between LoRA and PointNet's feature extraction capabilities establishes a robust framework for harnessing the rich information in point cloud data, enhancing both generalization and adaptability in complex environments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 370, + 491, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 491, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 491, + 382 + ], + "type": "text", + "content": "4.3. PointLoRA with Token Selection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 388, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 544 + ], + "type": "text", + "content": "While global features provide a comprehensive understanding of the overall structure of the point cloud, local features capture intricate details vital for accurate downstream task performance. To fully investigate LoRA's capabilities, it is crucial to complement the extracted global features with robust local features. The tokenization process produces a sequence of tokens that capture various aspects of the point cloud data. However, not all tokens are valuable for downstream tasks; some may encapsulate redundant or noninformative content, offering limited contribution to overall model performance. This necessitates a systematic approach for filtering and selecting the most informative tokens to enhance the model's representation of local features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 652 + ], + "type": "text", + "content": "To tackle the above concerns, we introduce a MultiScale Token Selection module integrated with vanilla LoRA, which is presented in Fig. 3. This module dynamically selects tokens from the raw point cloud at multiple scales, allowing the model to capture local features at different levels of detail. Additionally, PointLoRA incorporates a shared Prompt MLP to further embed local information from these selected tokens, enhancing the vanilla LoRA's performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "Multi-Scale Token Generation. Specifically, we apply farthest point sampling (FPS) on the input point cloud " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " with varying numbers of centroids (i.e., " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "N_{1}, N_{2}, \\dots, N_{M}" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": ") to generate tokens at " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " different scales. As described in Sec. 3, we can obtain " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " sets of center points, " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "C_{g}^{1}, C_{g}^{2}, \\dots, C_{g}^{M}" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6608" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 286, + 299 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 286, + 299 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 286, + 299 + ], + "type": "image", + "image_path": "306fb50b4823f49e3de30b28f55d5505b0a9a2dc3bd1d02dfe76fb5a056738d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 304, + 296, + 370 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 370 + ], + "type": "text", + "content": "Figure 3. Illustration of Multi-Scale Token Selection. In a two-scale setup, we first sample different numbers of center points, then cluster around each center point and apply Mini-PointNet encoding to generate the corresponding tokens. These tokens are also fed into a Mask Predictor to estimate importance scores, allowing us to select the Top-K tokens at each scale." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 390, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 437 + ], + "type": "text", + "content": "each containing " + }, + { + "bbox": [ + 55, + 390, + 296, + 437 + ], + "type": "inline_equation", + "content": "g_{1}, g_{2}, \\ldots, g_{M}" + }, + { + "bbox": [ + 55, + 390, + 296, + 437 + ], + "type": "text", + "content": " center points. Subsequently, " + }, + { + "bbox": [ + 55, + 390, + 296, + 437 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 390, + 296, + 437 + ], + "type": "text", + "content": "-NN is employed to retrieve the corresponding neighbor point patches, which are then embedded into discrete tokens using a shared Mini-PointNet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "text", + "content": "Mask Predictor for Selection. The above process yields " + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "text", + "content": " token sets, " + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "inline_equation", + "content": "T_{p}^{1}, T_{p}^{2}, \\ldots, T_{p}^{M}" + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "text", + "content": ", encoding local information at multiple scales. Our goal is to select tokens from these sets that are most beneficial for fine-tuning on downstream tasks. Therefore, a simple but effective Mask Predictor is designed to estimate the importance of each token, producing a score vector " + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "inline_equation", + "content": "s^{m} \\in \\mathbb{R}^{g_{m} \\times 1}" + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "text", + "content": " for each token in " + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "inline_equation", + "content": "T_{p}^{m} \\in \\mathbb{R}^{g_{m} \\times d}" + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "inline_equation", + "content": "m = 1, \\ldots, M" + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "text", + "content": ". In particular, the Mask Predictor comprises two multi-layer perceptron (MLP) layers followed by a Sigmoid activation function:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 566, + 294, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 566, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 111, + 566, + 294, + 581 + ], + "type": "interline_equation", + "content": "s ^ {m} = \\operatorname {S i g m o i d} \\left(\\operatorname {M L P} \\left(\\left(T _ {p} ^ {m}\\right)\\right)\\right). \\tag {8}", + "image_path": "7a3c341e41990db326a6e2571393a9d6bb6dfd30aac791f42b0daadec8db37b6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 587, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 661 + ], + "type": "text", + "content": "This design enables the model to assign scores between 0 and 1 to each token, indicating their relevance for downstream tasks. We then select a predefined number " + }, + { + "bbox": [ + 55, + 587, + 296, + 661 + ], + "type": "inline_equation", + "content": "N_{1}^{\\prime}, \\ldots, N_{M}^{\\prime}" + }, + { + "bbox": [ + 55, + 587, + 296, + 661 + ], + "type": "text", + "content": " of tokens from each scale based on their importance scores using Top-K selection, resulting in a refined token set " + }, + { + "bbox": [ + 55, + 587, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\{S_{\\mathrm{p}}^{m}\\}_{m=1}^{M}" + }, + { + "bbox": [ + 55, + 587, + 296, + 661 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 95, + 668, + 295, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 668, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 95, + 668, + 295, + 683 + ], + "type": "interline_equation", + "content": "S _ {\\mathrm {p}} ^ {m} = \\operatorname {T o p K} \\left(T _ {p} ^ {m}, N _ {m} ^ {\\prime}\\right), m = 1, \\dots , M. \\tag {9}", + "image_path": "0fce226fe518636450a266c2236e77f78e6638c4bd5e5aad8fbb4ed1dc24c634.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Local Geometry Prompt. With the selected " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "N_{s} = N_{1}^{\\prime} + \\ldots + N_{M}^{\\prime}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " tokens " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "S_{p}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ", we can capture crucial local information" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "for downstream tasks. These tokens are then concatenated with the original inputs to form a local geometry prompt. In each LoRA layer of the point cloud transformer block, a Prompt MLP with a GELU activation function [16] encodes the new input features alongside the local prompt. This process integrates the encoded features of the selected tokens with the LoRA adaptation output. The resulting combined output can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 177, + 553, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 177, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 321, + 177, + 553, + 201 + ], + "type": "interline_equation", + "content": "O _ {\\text {u p d a t e}} = \\operatorname {P r o m p t} \\mathrm {M L P} \\left(T _ {\\text {i n p u t}}, S _ {p}\\right) + \\Delta W \\cdot \\left(T _ {\\text {i n p u t}}, S _ {p}\\right), \\tag {10}", + "image_path": "c05d705d20eb9850d4d38ecfa60be6f9e2800d32fd2c945d1c77fca5c88b493c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 202, + 555, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 555, + 261 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 202, + 555, + 261 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{update}}" + }, + { + "bbox": [ + 313, + 202, + 555, + 261 + ], + "type": "text", + "content": " is the updated amount of the PointLoRA layer. Notably, to further optimize the efficiency of the parameters, the Prompt MLP is configured independently for both the qkv projection and FFN layers, while being shared across all blocks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 262, + 554, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 334 + ], + "type": "text", + "content": "The multi-scale token selection process improves model performance by prioritizing the most relevant local features during fine-tuning. By incorporating this approach, our scheme effectively refines both global and local feature representations, thereby enhancing its adaptability and accuracy across a range of 3D tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 342, + 437, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 437, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 437, + 355 + ], + "type": "text", + "content": "4.4. Training & Inference" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "text", + "content": "Overall Loss Function. When fine-tuning on diverse downstream tasks across various datasets, we follow the common practice of maintaining consistency between the adopted loss and the loss function of the original task. Additionally, we introduce a regularization term to further supervise the Mask Predictor. In general, the total loss function " + }, + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_{\\mathrm{task}} + \\lambda \\cdot \\mathcal{L}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "text", + "content": " is the task-specific loss for classification or segmentation. " + }, + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "text", + "content": " is the regularization loss for mask predictor with the balanced weight " + }, + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 360, + 555, + 468 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 475, + 553, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 475, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 348, + 475, + 553, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = - \\frac {1}{N _ {\\text {t o t a l}}} \\sum_ {i = 1} ^ {N _ {\\text {t o t a l}}} \\left(s _ {i} \\log \\left(s _ {i} + \\epsilon\\right) + \\right. \\tag {11}", + "image_path": "fceb2862f1b32dae108b89ffafa38e603dc6792e61e667ef481de0abacf34a35.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 383, + 502, + 488, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 502, + 488, + 516 + ], + "spans": [ + { + "bbox": [ + 383, + 502, + 488, + 516 + ], + "type": "interline_equation", + "content": "\\left. \\left(1 - s _ {i}\\right) \\log \\left(1 - s _ {i} + \\epsilon\\right)\\right),", + "image_path": "86708c59a40eeaa0403c21c6aa824787b87688d5b509be2a26b172582c4afc30.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{total}} = N_1 + N_2 \\dots + N_M" + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": " denotes the importance score for each token and " + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": " is a small constant (e.g., " + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": ") to prevent undefined values in the logarithm." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 557, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 616 + ], + "type": "text", + "content": "Inference. During inference, the additional parameters introduced by LoRA in the transformer block are directly merged into the original pre-trained weights, with the only newly added component being a small shared MLP dedicated to embed the selected tokens." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 628, + 395, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 395, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 395, + 641 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 648, + 364, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 364, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 364, + 661 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Datasets. We validate the effectiveness of our approach through extensive experiments on three widely used 3D datasets: ScanObjectNN [52], ModelNet40 [61], and ShapeNetPart [68]. ScanObjectNN [52] is a challenging" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6609" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 107, + 555, + 390 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Performance comparison on three variants of the ScanObjectNN [52] and the ModelNet40 [61] datasets, respectively, for real-world and synthetic object classification. Both the number of tunable parameters and overall accuracy (OA) are reported. All methods only employ the default data argumentation without voting as the baseline. * denotes results reproduced from the public source code." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 107, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 56, + 107, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 555, + 390 + ], + "type": "table", + "html": "
MethodsPublicationTunable Params.ScanObjectNNModelNet40
OBJ-BGOBJ-ONLYPB-T50-RSPoints Num.OA (%)
Traditional Supervised Learning Only
PointNet [41]CVPR'173.5 M73.379.268.01k89.2
PointNet++ [42]NeurIPS'171.5 M82.384.377.91k90.7
DGCNN [59]TOG'191.8 M82.886.278.11k92.9
MVTN [13]ICCV'2111.2 M--82.81k93.8
PointNeXt [45]NeurIPS'221.4 M--87.71k94.0
PointMLP [37]ICLR'2213.2 M--85.41k94.5
RepSurf-U [46]CVPR'221.5 M--84.31k94.4
ADS [18]ICCV'23---87.51k95.1
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2122.1 M84.8585.5478.791k92.1
Point-BERT [69]CVPR'2222.1 M87.4388.1283.071k93.2
MaskPoint [34]ECCV'2222.1 M89.7089.3084.601k93.8
Point-MAE [40]ECCV'2222.1 M90.0288.2985.181k93.8
Point-M2AE [74]NeurIPS 2215.3 M91.2288.8186.431k94.0
ACT [10]ICLR'2322.1 M93.2991.9188.211k93.7
RECON [43]ICML'2343.6 M94.1593.1289.731k93.9
Self-Supervised Representation Learning (Parameter-Efficient Fine-Tuning)
Point-MAE [40] (Full-FT)ECCV'2222.1 M (100%)90.0288.2985.181k93.2
Point-MAE + IDPT [72]ICCV'231.7 M (7.69%)91.22(+1.20)90.02(+1.73)84.94(-0.24)1k93.3(+0.1)
Point-MAE + DAPT [77]CVPR'241.1 M (4.97%)90.88(+0.86)90.19(+1.90)85.08(-0.10)1k93.5(+0.3)
Point-MAE + PPT* [75]arXiv'241.04 M (4.57%)89.84(-0.18)88.98(+0.69)84.45(-0.73)1k93.2(+0.0)
Point-MAE + PointLoRAOurs0.77 M (3.43%)90.71(+0.69)89.33(+1.04)85.53(+0.35)1k93.3(+0.1)
", + "image_path": "72eea66754c48c0d9ef383e5c4dd9e50d9d4f4ee65648a6c73eb643a50995ff3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 402, + 294, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 294, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 294, + 593 + ], + "type": "text", + "content": "real-world 3D object classification dataset, which contains approximately 15,000 indoor point cloud instances across 15 categories. The object classification task is performed on three variants of increasing complexity, OBJ-BG, OBJ-ONLY, and PB-T50-RS, each representing different levels of scene realism and occlusion. ModelNet40 [61], a classical synthetic dataset for 3D object recognition, consists of 12,311 meshed 3D CAD objects that span 40 categories. We perform synthetic object classification and few-shot learning experiments on ModelNet40 to evaluate the robustness of the model in both standard and low-resource settings. Additionally, to evaluate performance on detailed structures and component segmentation, we provide experiments on ShapeNetPart [68], a widely used benchmark for part segmentation, comprising 16,881 point-level synthetic objects across 16 object categories and 50 part categories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 594, + 294, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 294, + 689 + ], + "type": "text", + "content": "Evaluation Metrics. In the evaluation, overall accuracy (OA) is adopted to assess performance on the 3D object classification task, representing the ratio of correctly classified instances to the total number of instances, thus providing an aggregate score over all classes. For the part segmentation task, we employ mean Intersection over Union (mIoU) to evaluate the overlap between prediction and ground-truth segments, averaged across all classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Implementation Details. Our method can be directly integrated into existing point cloud pre-trained models. Dur" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": "ing fine-tuning, we follow a commonly used setup, freezing most parameters of the pre-trained model, and primarily updating the newly inserted parameters. The number of scales, " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": ", is set to 2. At these two scales, we apply farthest point sampling (FPS) and " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": "-NN clustering with (128, 32) and (64, 64) center and neighboring points, respectively, and then select tokens with " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "N_1' = 32" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "N_2' = 8" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": ". The balanced weight " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " for Mask Predictor learning is set to 0.004 For real-world and synthetic object classification, we fine-tune using Point-MAE [40], training for 300 epochs with a learning rate of " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "5e^{-4}" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " and a weight decay of 0.05. For few-shot learning and part segmentation, we utilize the Recon [43] model for fine-tuning to further validate the generalizability of our approach, with a learning rate of " + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "inline_equation", + "content": "2e^{-4}" + }, + { + "bbox": [ + 313, + 402, + 555, + 606 + ], + "type": "text", + "content": " for part segmentation. All experiments are conducted on a single GPU. More implementation details with different baselines are provided in the Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 612, + 429, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 429, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 429, + 625 + ], + "type": "text", + "content": "5.2. Comparative Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Real-World & Synthetic Object Classification. We conduct real-world and synthetic object classification on three variants of ScanObjectNN (OBJ-BG, OBJ-ONLY, PB-T50-RS) [52] and ModelNet40 [61], using default data augmentation and without voting. As illustrated in Tab. 1, our proposed approach demonstrates consistent performance improvements, especially on the most challenging vari" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6610" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 159, + 183 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 159, + 183 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 159, + 183 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 159, + 183 + ], + "type": "image", + "image_path": "118a6bffe5490a0c36f67841301bbe781d4ea0774d2749af33ec0314523b6425.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "type": "text", + "content": "Figure 4. The t-SNE visualization results on the PB-T50-RS split of the ScanObjectNN dataset [52] with different fine-tuning schemes. We adopt Point-MAE [40] as the baseline model for fair comparison. TP: Number of tunable parameters. OA: Overall accuracy. Symbol * denotes re-produced with official implementation. Best viewed in colors and zoomed-in for additional details." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 165, + 70, + 253, + 182 + ], + "blocks": [ + { + "bbox": [ + 165, + 70, + 253, + 182 + ], + "lines": [ + { + "bbox": [ + 165, + 70, + 253, + 182 + ], + "spans": [ + { + "bbox": [ + 165, + 70, + 253, + 182 + ], + "type": "image", + "image_path": "488cca78a4feae37af758eef1855e3ffccafa3efa9fa104c185128d89469deb6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 259, + 70, + 349, + 182 + ], + "blocks": [ + { + "bbox": [ + 259, + 70, + 349, + 182 + ], + "lines": [ + { + "bbox": [ + 259, + 70, + 349, + 182 + ], + "spans": [ + { + "bbox": [ + 259, + 70, + 349, + 182 + ], + "type": "image", + "image_path": "c5321543229616d63dff5024f18e68f96b69c510fedb3e209af31322c40973f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 354, + 70, + 444, + 182 + ], + "blocks": [ + { + "bbox": [ + 354, + 70, + 444, + 182 + ], + "lines": [ + { + "bbox": [ + 354, + 70, + 444, + 182 + ], + "spans": [ + { + "bbox": [ + 354, + 70, + 444, + 182 + ], + "type": "image", + "image_path": "eab0465e9f56b2288f8d2a66c7db0e7ac7fdfd0f995ce7c7220c582629baf689.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 451, + 70, + 539, + 182 + ], + "blocks": [ + { + "bbox": [ + 451, + 70, + 539, + 182 + ], + "lines": [ + { + "bbox": [ + 451, + 70, + 539, + 182 + ], + "spans": [ + { + "bbox": [ + 451, + 70, + 539, + 182 + ], + "type": "image", + "image_path": "5c974cb58e42fde1a0170015bfd386e71c0217df8d6239f6f3fd197dac07d4f2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 57, + 278, + 294, + 446 + ], + "blocks": [ + { + "bbox": [ + 55, + 231, + 297, + 276 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 297, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 297, + 276 + ], + "type": "text", + "content": "Table 2. Performance comparison on ModelNet40 [61] for few-shot learning. We report the scores of the overall accuracy " + }, + { + "bbox": [ + 55, + 231, + 297, + 276 + ], + "type": "inline_equation", + "content": "(\\%)\\pm" + }, + { + "bbox": [ + 55, + 231, + 297, + 276 + ], + "type": "text", + "content": " the standard deviation " + }, + { + "bbox": [ + 55, + 231, + 297, + 276 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 231, + 297, + 276 + ], + "type": "text", + "content": " without voting. The top two highest accuracies are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 278, + 294, + 446 + ], + "lines": [ + { + "bbox": [ + 57, + 278, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 57, + 278, + 294, + 446 + ], + "type": "table", + "html": "
MethodsPublication5-way10-way
10-shot20-shot10-shot20-shot
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2194.0±3.695.9±2.389.4±5.192.4±4.6
Point-BERT [69]CVPR'2294.6±3.196.3±2.791.0±5.492.7±5.1
MaskPoint [34]ECCV'2295.0±3.797.2±1.791.4±4.093.4±3.5
Point-MAE [40]ECCV'2296.3±2.597.8±1.892.6±4.195.0±3.0
Point-M2AE [74]NeurIPS'2296.8±1.898.3±1.492.3±4.595.0±3.0
ACT [10]ICLR'2396.8±2.398.0±1.493.3±4.095.6±2.8
Self-Supervised Representation Learning (Efficient Fine-Tuning)
RECON [43] (Full-FT)ICML'2397.3±1.998.9±3.993.3±3.995.8±3.0
RECON+ IDPT [72]ICCV'2396.9±2.498.3±0.792.8±4.095.5±3.2
RECON+ DAPT [77]CVPR'2495.6±2.897.7±1.691.9±4.194.6±3.5
RECON+ PPT [75]arXiv'2497.0±2.798.7±1.692.2±5.095.6±2.9
RECON+ PointLoRAOurs96.9±2.798.8±1.292.7±4.495.8±2.9
", + "image_path": "a0d0e042e9440da5d1a68e0841f58d3da408edbbfb9a51475895bfe5ea75fc46.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 460, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 579 + ], + "type": "text", + "content": "ant, PB-T50-RS, where it is the only parameter-efficient fine-tuning method that surpasses the full fine-tuned PointMAE [40] (85.53% vs. 85.18%) and significantly outperforms the state-of-the-art method, DAPT [77]. Notably, our approach requires only 0.77M tunable parameters, the fewest among all methods. Furthermore, the t-SNE feature visualization [53] of existing methods and PointLoRA is presented in Fig. 4. Our method enables the pre-trained model to fine-tune with a minimal number of parameters, producing distinctive feature representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 580, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 580, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 54, + 580, + 295, + 688 + ], + "type": "text", + "content": "Few-shot Learning. We also perform few-shot learning experiments on ModelNet40 [61] to validate the transfer learning capability of our method with limited annotations. Following standard protocols [15, 43], we evaluate the fine-tuning performance with ReCon [43] in the 5-way / 10-way and 10-shot / 20-shot settings, respectively. As shown in Tab. 2, our method achieves the best or second-best accuracy in all configurations, further demonstrating the generalizability of the proposed scheme." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Point Cloud Part Segmentation. The part segmentation results on the ShapeNetPart dataset [68] are provided in" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 289, + 550, + 488 + ], + "blocks": [ + { + "bbox": [ + 313, + 231, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 286 + ], + "type": "text", + "content": "Table 3. Performance comparison on the ShapeNetPart [68] for part segmentation. Both the mIoU for all classes (Cls.) and instances (Inst.) are provided. TP indicates the number of tunable parameters. " + }, + { + "bbox": [ + 313, + 231, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 231, + 555, + 286 + ], + "type": "text", + "content": " means the result, reported from DAPT [77]. " + }, + { + "bbox": [ + 313, + 231, + 555, + 286 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 313, + 231, + 555, + 286 + ], + "type": "text", + "content": " denotes the results reproduced from the official implementation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 289, + 550, + 488 + ], + "lines": [ + { + "bbox": [ + 315, + 289, + 550, + 488 + ], + "spans": [ + { + "bbox": [ + 315, + 289, + 550, + 488 + ], + "type": "table", + "html": "
MethodsPublicationTPCls. mIoU (%)Inst. mIoU (%)
Traditional Supervised Learning Only
PointNet [41]CVPR'17-80.3983.7
PointNet++ [42]NeurIPS'17-81.8585.1
DGCNN [59]TOG'19-82.3385.2
APES [60]CVPR'23-83.6785.8
Self-Supervised Representation Learning (Full Fine-Tuning)
OcCo [54]ICCV'2127.09 M83.4285.1
MaskPoint [34]ECCV'22-84.6086.0
Point-BERT [69]CVPR'2227.09 M84.1185.6
Point-MAE [40]ECCV'2227.06 M84.1986.1
ACT [10]ICLR'2327.06 M84.6686.1
Self-Supervised Representation Learning (Efficient Fine-Tuning)
RECON [43] (Full-FT)ICML'2327.06 M84.5286.1
RECON+ IDPT† [72]ICCV'235.69 M83.6685.7
RECON+ DAPT [77]CVPR'245.65 M83.8785.7
RECON+ PPT* [75]arXiv'245.62 M83.8885.4
RECON+ PointLoRAOurs5.63 M83.9885.4
", + "image_path": "9ce3f9ee25c1446336a8053a916fc10938b863490a76bdb13dec682a8297f5e6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 555 + ], + "type": "text", + "content": "Tab. 3. In this fine-grained scene understanding task, our approach still achieves competitive performance with a constrained parameter budget. Unlike prior classification models, the increase in the parameter count is primarily attributed to the segmentation head." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": "Comparison with Other PEFT Methods. We further present a comparison with existing PEFT methods designed for other tasks, including NLP and 2D vision. We select the most challenging PB-T50-RS variant [52] and use the pretrained Point-MAE model [40] as a baseline. As shown in Tab. 4, these methods designed for other tasks provide limited performance gains when fine-tuned in 3D scenes, despite their parameter efficiency compared to existing 3D fine-tuning approaches. In contrast, our method not only achieves the highest accuracy, but also uses significantly fewer parameters than the three current fine-tuning techniques specifically developed for point clouds, underscoring the strong potential of PointLoRA." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6611" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 107, + 281, + 256 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 297, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 297, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 297, + 104 + ], + "type": "text", + "content": "Table 4. Performance comparison with other parameter-efficient methods designed for NLP and 2D Vision tasks on the hardest variant of ScanObjectNN [52]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 107, + 281, + 256 + ], + "lines": [ + { + "bbox": [ + 67, + 107, + 281, + 256 + ], + "spans": [ + { + "bbox": [ + 67, + 107, + 281, + 256 + ], + "type": "table", + "html": "
MethodsPublicationTPPB-T50-RS
Point-MAE [34]ECCV'2222.1 M85.18
Linear probing-0.3 M75.99
+ Adapter [19]ICML'190.9 M83.93
+ Perfix tuning [28]ACL'210.7 M77.72
+ BitFit [71]ACL'210.3 M82.62
+ LoRA†[20]ICLR'220.9 M81.74
+ VPT-Deep [21]ECCV'220.4 M81.09
+ AdaptFormer [8]NeurIPS'220.9 M83.45
+ SSF [31]NeurIPS'220.4 M82.58
+ IDPT [72]ICCV'231.7 M84.94
+ DAPT [77]CVPR'241.1 M85.08
+ PPT* [75]arXiv'241.04 M84.45
+ PointLoRAOurs0.77 M85.53
", + "image_path": "e9fbd8f9be3f15648903c7266f5280b1f605a0e35d020dc55e862dcf627f4058.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 306, + 282, + 380 + ], + "blocks": [ + { + "bbox": [ + 55, + 258, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 304 + ], + "type": "text", + "content": "Table 5. The impact of each module of our scheme. We provide the overall accuracy (\\%) on the hardest variant of ScanObjectNN [52] and corresponding tunable parameters (TP). \"MS-FPS\" indicates multi-scale furthest point sampling." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 306, + 282, + 380 + ], + "lines": [ + { + "bbox": [ + 66, + 306, + 282, + 380 + ], + "spans": [ + { + "bbox": [ + 66, + 306, + 282, + 380 + ], + "type": "table", + "html": "
LoRAToken SelectionMS-FPSTPPB-T50-RS
Full Fine-Tuning22.1 M85.18
Linear Probing0.27 M75.99
✓0.53 M83.83
✓✓0.77 M84.91
✓✓✓0.77 M85.53
", + "image_path": "e191a770868277c554a9bceaac916571b8b3fee78133837b5159c3e1bd4dadd8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 149, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 149, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 149, + 398 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 462 + ], + "type": "text", + "content": "In this section, we perform exhaustive ablation experiments on the challenging PB-T50-RS variant to investigate the rationale and effectiveness of the design choices of our proposed approach. The pre-trained Point-MAE [40] is adopted as a baseline for a fair comparison." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": "Ablation on PointLoRA Scheme. Firstly, we provide the ablations on each module of the proposed method. As illustrated in Tab. 5, directly applying low-rank adaptation (LoRA) produces a significant improvement over linear probing, demonstrating the effectiveness of incorporating global information during the fine-tuning process. Injecting selected tokens as prompts into LoRA further improves accuracy with only a marginal increase in the number of parameters. Finally, the local information obtained from selected tokens across multiple scales complements the global features of LoRA, achieving the best performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": "Ablation on Low-Rank Adaptation. Then we perform ablation experiments on rank " + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": " in our approach, as shown in Tab. 6. Different ranks have a noticeable impact on finetuning performance. Taking into account both parameter count and accuracy, we set the rank to 8 to maximize the extraction of global information for fine-tuning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "Ablation on Multi-Scale Token Selection. Given the critical improvement from multi-scale token selection for PointLoRA, we first perform ablation experiments on the number of tokens selected at the same scale. As shown in" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 340, + 85, + 526, + 145 + ], + "blocks": [ + { + "bbox": [ + 324, + 71, + 543, + 83 + ], + "lines": [ + { + "bbox": [ + 324, + 71, + 543, + 83 + ], + "spans": [ + { + "bbox": [ + 324, + 71, + 543, + 83 + ], + "type": "text", + "content": "Table 6. Ablation study on rank " + }, + { + "bbox": [ + 324, + 71, + 543, + 83 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 324, + 71, + 543, + 83 + ], + "type": "text", + "content": " of the proposed method." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 85, + 526, + 145 + ], + "lines": [ + { + "bbox": [ + 340, + 85, + 526, + 145 + ], + "spans": [ + { + "bbox": [ + 340, + 85, + 526, + 145 + ], + "type": "table", + "html": "
Rank rTPRatioPB-T50-RS
40.66 M2.96%84.28
80.77 M3.43%85.53
160.99 M4.37%85.15
321.44 M6.32%84.87
", + "image_path": "52161e44b9a30864adbe9dba75b1856e90bda544454909dbd9b825a204b14afc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 340, + 172, + 526, + 268 + ], + "blocks": [ + { + "bbox": [ + 313, + 147, + 553, + 168 + ], + "lines": [ + { + "bbox": [ + 313, + 147, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 553, + 168 + ], + "type": "text", + "content": "Table 7. Ablation study on scale number " + }, + { + "bbox": [ + 313, + 147, + 553, + 168 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 147, + 553, + 168 + ], + "type": "text", + "content": " and selected token number " + }, + { + "bbox": [ + 313, + 147, + 553, + 168 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 313, + 147, + 553, + 168 + ], + "type": "text", + "content": " in Multi-Scale Token Selection." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 172, + 526, + 268 + ], + "lines": [ + { + "bbox": [ + 340, + 172, + 526, + 268 + ], + "spans": [ + { + "bbox": [ + 340, + 172, + 526, + 268 + ], + "type": "table", + "html": "
Scale MToken Num. NsPB-T50-RS
16484.80
13284.91
11685.22
264 & 1684.39
232 & 885.53
216 & 484.49
332 & 16 & 885.05
", + "image_path": "32cd0e67a7f45ce0cfbde672ecc05044d9f6fab72ad64e082784ab6fcc44856c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 340, + 285, + 525, + 345 + ], + "blocks": [ + { + "bbox": [ + 329, + 270, + 539, + 282 + ], + "lines": [ + { + "bbox": [ + 329, + 270, + 539, + 282 + ], + "spans": [ + { + "bbox": [ + 329, + 270, + 539, + 282 + ], + "type": "text", + "content": "Table 8. Ablation study on the dimension of Prompt MLP." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 285, + 525, + 345 + ], + "lines": [ + { + "bbox": [ + 340, + 285, + 525, + 345 + ], + "spans": [ + { + "bbox": [ + 340, + 285, + 525, + 345 + ], + "type": "table", + "html": "
DimensionTPRatioPB-T50-RS
80.68 M3.03%83.55
160.71 M3.17%84.77
320.77 M3.43%85.53
640.90 M3.96%85.25
", + "image_path": "7a9cb13abd0f87a10f05c435879a5a3668d0b117d99ea5b6547bf3b92b2ef097.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 358, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 430 + ], + "type": "text", + "content": "the upper part of Tab. 7, selecting 16 tokens at the original scale yields the best results, demonstrating the effectiveness of token selection. Furthermore, when incorporating tokens from multiple scales, optimal performance is achieved by selecting 32 and 8 tokens from two different scales, as shown in the lower part of Tab. 7." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 430, + 556, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 526 + ], + "type": "text", + "content": "Ablation on Prompt MLP. We provide the ablation study on the dimension of the shared Prompt MLP, as this affects the embedding of local information in our method. As illustrated in Tab. 8, setting this dimension to 32 yields the best fine-tuning performance. Additionally, adjusting the dimension causes only minimal changes in tunable parameter counts, suggesting flexibility in selecting this hyperparameter for different datasets and tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 557, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 712 + ], + "type": "text", + "content": "In this paper, a simple yet effective parameter-efficient finetuning approach named PointLoRA, is presented with low-rank adaptation and multi-scale token selection for point cloud models. Low-rank adaptation efficiently reduces the parameter-heavy components of the point cloud transformer architecture while capturing global information. Multi-scale token selection effectively encodes essential local features as prompt, further enhancing the finetuning process. The integration of global and local information enables our approach to achieve state-of-the-art results on the most challenging dataset with a minimal number of tunable parameters while also delivering competitive performance across other datasets and models." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6612" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "content": "This work is supported in part by the National Natural Science Foundation of China under Grant No. 62376244, in part by the Zhejiang Provincial Natural Science Foundation of China under Grant No. LD24F030001, by the Information Technology Center and State Key Lab of CAD&CG, Zhejiang University, and by the National Research Foundation, Singapore, and Cyber Security Agency of Singapore under its National Cybersecurity R&D Programme and CyberSG R&D Cyber Research Programme Office (Award: CRPO-GC1-NTU-002). Lingdong Kong is supported by the Apple Scholars in AI/ML Ph.D. Fellowship program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 233, + 115, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 115, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 115, + 247 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 253, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 253, + 296, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 253, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 61, + 253, + 296, + 320 + ], + "type": "text", + "content": "[1] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 321, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 321, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 61, + 321, + 296, + 342 + ], + "type": "text", + "content": "[2] Jimmy Lei Ba. Layer normalization. arXiv preprint arXiv:1607.06450, 2016." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 343, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 343, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 61, + 343, + 296, + 365 + ], + "type": "text", + "content": "[3] George Bebis and Michael Georgiopoulos. Feed-forward neural networks. IEEE Potentials, 13(4):27-31, 1994." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 366, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 366, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 61, + 366, + 296, + 399 + ], + "type": "text", + "content": "[4] Saifullahi Aminu Bello, Shangshu Yu, Cheng Wang, Jibril Muhammad Adam, and Jonathan Li. Deep learning on 3d point clouds. Remote Sensing, 12(11):1729, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 401, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 401, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 61, + 401, + 296, + 445 + ], + "type": "text", + "content": "[5] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. In Advances in Neural Information Processing Systems, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 446, + 296, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 446, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 61, + 446, + 296, + 500 + ], + "type": "text", + "content": "[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 502, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 502, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 61, + 502, + 296, + 555 + ], + "type": "text", + "content": "[7] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 557, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 557, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 61, + 557, + 296, + 601 + ], + "type": "text", + "content": "[8] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 602, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 602, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 61, + 602, + 296, + 657 + ], + "type": "text", + "content": "[9] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[10] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "text", + "content": "[12] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 184, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 554, + 228 + ], + "type": "text", + "content": "[13] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. Mvtn: Multi-view transformation network for 3d shape recognition. In IEEE/CVF International Conference on Computer Vision, pages 1-11, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 271 + ], + "type": "text", + "content": "[14] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "text", + "content": "[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 554, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 554, + 338 + ], + "type": "text", + "content": "[16] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 370 + ], + "type": "text", + "content": "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "text", + "content": "[18] Cheng-Yao Hong, Yu-Ying Chou, and Tyng-Luh Liu. Attention discriminant sampling for point clouds. In IEEE/CVF International Conference on Computer Vision, pages 14429-14440, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 470 + ], + "type": "text", + "content": "[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning. PMLR, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 471, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 554, + 514 + ], + "type": "text", + "content": "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 554, + 558 + ], + "type": "text", + "content": "[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "type": "text", + "content": "[22] Lingdong Kong, Youquan Liu, Runnan Chen, Yuexin Ma, Xinge Zhu, Yikang Li, Yuenan Hou, Yu Qiao, and Ziwei Liu. Rethinking range view representation for lidar segmentation. In IEEE/CVF International Conference on Computer Vision, pages 228-240, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 669 + ], + "type": "text", + "content": "[23] Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. In IEEE/CVF International Conference on Computer Vision, pages 19994-20006, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[24] Lingdong Kong, Jiawei Ren, Liang Pan, and Ziwei Liu. Lasermix for semi-supervised lidar semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21705-21715, 2023." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6613" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "type": "text", + "content": "[25] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Empirical Methods in Natural Language Processing, pages 3045-3059, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 296, + 161 + ], + "type": "text", + "content": "[26] Bonan Li, Yinhan Hu, Xuecheng Nie, Congying Han, Xi-angjian Jiang, Tiande Guo, and Luoqi Liu. Dropkey for vision transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22700–22709, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 205 + ], + "type": "text", + "content": "[27] Minglei Li, Peng Ye, Yongqi Huang, Lin Zhang, Tao Chen, Tong He, Jiayuan Fan, and Wanli Ouyang. Adapter-x: A novel general parameter-efficient fine-tuning framework for vision. arXiv preprint arXiv:2406.03051, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 296, + 239 + ], + "type": "text", + "content": "[28] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Annual Meeting of the Association for Computational Linguistics, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "type": "text", + "content": "[29] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In Advances in Neural Information Processing Systems, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "type": "text", + "content": "[30] Ye Li, Lingdong Kong, Hanjiang Hu, Xiaohao Xu, and Xiaonan Huang. Is your lidar placement optimized for 3d scene understanding? In Advances in Neural Information Processing Systems, pages 34980-35017, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 370 + ], + "type": "text", + "content": "[31] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 296, + 415 + ], + "type": "text", + "content": "[32] Dingkang Liang, Tianrui Feng, Xin Zhou, Yumeng Zhang, Zhikang Zou, and Xiang Bai. Parameter-efficient fine-tuning in spectral domain for point cloud learning. arXiv preprint arXiv:2410.08114, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 459 + ], + "type": "text", + "content": "[33] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 296, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 296, + 493 + ], + "type": "text", + "content": "[34] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 493, + 296, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 296, + 546 + ], + "type": "text", + "content": "[35] Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Derong Xu, Feng Tian, and Yefeng Zheng. Moelora: An moe-based parameter efficient fine-tuning method for multitask medical applications. arXiv preprint arXiv:2310.18339, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 296, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 603 + ], + "type": "text", + "content": "[36] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. In Advances in Neural Information Processing Systems, pages 37193-37229, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 296, + 647 + ], + "type": "text", + "content": "[37] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "type": "text", + "content": "[38] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[39] Lucas Nunes, Rodrigo Marcuzzi, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Segcontrast: 3d point cloud" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "feature representation learning through self-supervised segment discrimination. IEEE Robotics and Automation Letters, 7(2):2116-2123, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "text", + "content": "[40] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[41] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 194, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 239 + ], + "type": "text", + "content": "[42] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[43] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In International Conference on Machine Learning. PMLR, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 293, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 338 + ], + "type": "text", + "content": "[44] Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. *Shapellm: Universal 3d object understanding for embodied interaction*. arXiv preprint arXiv:2402.17766, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "type": "text", + "content": "[45] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In Advances in Neural Information Processing Systems, pages 23192-23204, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "text", + "content": "[46] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 426, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 481 + ], + "type": "text", + "content": "[47] Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9891-9901, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 481, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 553, + 514 + ], + "type": "text", + "content": "[48] Zhengxiang Shi and Aldo Lipani. Dept: Decomposed prompt tuning for parameter-efficient fine-tuning. In International Conference on Learning Representations, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 515, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 559 + ], + "type": "text", + "content": "[49]Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8791-8800, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "type": "text", + "content": "[50] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. In Advances in Neural Information Processing Systems, pages 12991-13005, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "type": "text", + "content": "[51] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. In AAAI Conference on Artificial Intelligence, pages 5171-5179, 2024." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 715 + ], + "type": "text", + "content": "[52] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, 2019." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6614" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[53] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 2008." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[54] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In IEEE/CVF International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[55] Song Wang, Jianke Zhu, and Ruixiang Zhang. Metarangeseg: Lidar sequence semantic segmentation using multiple feature aggregation. IEEE Robotics and Automation Letters, 7(4):9739-9746, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "text", + "content": "[56] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5186-5195, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[57] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14792-14801, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "text", + "content": "[58] Song Wang, Jiawei Yu, Wentong Li, Hao Shi, Kailun Yang, Junbo Chen, and Jianke Zhu. Label-efficient semantic scene completion with scribble annotations. In International Joint Conference on Artificial Intelligence, pages 1398-1406, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[59] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics, 38:1-12, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[60] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[61] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1912-1920, 2015." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "text", + "content": "[62] Aoran Xiao, Xiaoqin Zhang, Ling Shao, and Shijian Lu. A survey of label-efficient deep learning for 3d point clouds. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "type": "text", + "content": "[63] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Zheng-Jun Zha, Xiaoyan Sun, and Zhiwei Xiong. Event-adapted video superresolution. In European Conference on Computer Vision, pages 217-235. Springer, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "type": "text", + "content": "[64] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In European Conference on Computer Vision, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[65] Yi Xin, Siqi Luo, Haodi Zhou, Junlong Du, Xiaohong Liu, Yue Fan, Qing Li, and Yuntao Du. Parameter-efficient fine-tuning for pre-trained vision models: A survey. arXiv preprint arXiv:2402.02242, 2024." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[66] Xiang Xu, Lingdong Kong, Hui Shuai, Wenwei Zhang, Liang Pan, Kai Chen, Ziwei Liu, and Qingshan Liu. 4d" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 643 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "contrastive superflows are dense 3d representation learners. In European Conference on Computer Vision, pages 58-80, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[67] Xingyi Yang and Xinchao Wang. Kolmogorov-arnold transformer. In International Conference on Learning Representations, 2025." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[68] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics, 35(6):1-12, 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 250 + ], + "type": "text", + "content": "[69] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "text", + "content": "[70] Ted Zadouri, Ahmet Üstün, Arash Ahmadian, Beyza Ermis, Acyr Locatelli, and Sara Hooker. Pushing mixture of experts to the limit: Extremely parameter efficient moe for instruction tuning. arXiv preprint arXiv:2309.05444, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[71] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Annual Meeting of the Association for Computational Linguistics, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 385 + ], + "type": "text", + "content": "[72] Yaohua Zha, Jinpeng Wang, Tao Dai, Bin Chen, Zhi Wang, and Shu-Tao Xia. Instance-aware dynamic prompt tuning for pre-trained point cloud models. In IEEE/CVF International Conference on Computer Vision, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "text", + "content": "[73] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient fin-tuning. In International Conference on Learning Representations, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 443, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 496 + ], + "type": "text", + "content": "[74] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. In Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 499, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 531 + ], + "type": "text", + "content": "[75] Shaochen Zhang, Zekun Qi, Runpei Dong, Xiuxiu Bai, and Xing Wei. Positional prompt tuning for efficient 3d representation learning. arXiv preprint arXiv:2408.11567, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 533, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 586 + ], + "type": "text", + "content": "[76] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 589, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 643 + ], + "type": "text", + "content": "[77] Xin Zhou, Dingkang Liang, Wei Xu, Xingkui Zhu, Yihan Xu, Zhikang Zou, and Xiang Bai. Dynamic adapter meets prompt tuning: Parameter-efficient transfer learning for point cloud analysis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14707-14717, 2024." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6615" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/f94115e7-2e5c-44d9-81fe-c067d1bde80f_content_list.json b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/f94115e7-2e5c-44d9-81fe-c067d1bde80f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..08bb5d7d197b5e8f422d8be171e99e69c6e17a91 --- /dev/null +++ b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/f94115e7-2e5c-44d9-81fe-c067d1bde80f_content_list.json @@ -0,0 +1,1986 @@ +[ + { + "type": "text", + "text": "PointSR: Self-Regularized Point Supervision for Drone-View Object Detection", + "text_level": 1, + "bbox": [ + 104, + 130, + 890, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weizhuo Li $^{1}$ Yue Xi $^{1*}$ Wenjing Jia $^{2}$ Zehao Zhang $^{1}$ Fei Li $^{3}$ Xiangzeng Liu $^{3}$ Qiguang Miao $^{3*}$", + "bbox": [ + 99, + 178, + 929, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Guangzhou Institute of Technology, Xidian University, Guangzhou, China", + "bbox": [ + 285, + 200, + 733, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Faculty of Engineering and IT, University of Technology Sydney, Sydney, Australia", + "bbox": [ + 259, + 217, + 764, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Computer Science and Technology, Xidian University, Xi'an, China", + "bbox": [ + 272, + 234, + 746, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liweizhuo@stu., xiyue@, zzehao@stu., lifei@, xzliu@, qgmiao@}xidian.edu.cn", + "bbox": [ + 181, + 253, + 841, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenjing.Jia@uts.edu.au", + "bbox": [ + 410, + 272, + 607, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 321, + 325, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point-Supervised Object Detection (PSOD) in a discriminative style has recently gained significant attention for its impressive detection performance and cost-effectiveness. However, accurately predicting high-quality pseudo-box labels for drone-view images, which often feature densely packed small objects, remains a challenge. This difficulty arises primarily from the limitation of rigid sampling strategies, which hinder the pseudo-box optimization process. To address this, we propose PointSR, an effective and robust point-supervised object detection framework with self-regularized sampling that integrates temporal and informative constraints throughout the pseudo-box generation process. Specifically, the framework comprises three key components: Temporal-Ensembling Encoder (TE Encoder), Coarse Pseudo-box Prediction, and Pseudo-box Refinement. The TE Encoder builds an anchor prototype library by aggregating temporal information for dynamic anchor adjustment. In Coarse Pseudo-box Prediction, anchors are refined using the prototype library, and a set of informative samples is collected for subsequent refinement. During Pseudo-box Refinement, these informative negative samples are used to suppress low-confidence candidate positive samples, thereby improving the quality of the pseudoboxes. Experimental results on benchmark datasets demonstrate that PointSR significantly outperforms state-of-the-art methods, achieving up to $2.6\\% \\sim 7.2\\%$ higher $AP_{50}$ using only point supervision. Additionally, it exhibits strong robustness to perturbation in human-labeled points.", + "bbox": [ + 89, + 353, + 483, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 791, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Drone-view object detection (DroneDet) [7, 23, 27, 28, 33], which involves detecting objects from drone-captured images, plays a vital role in various domains, including national defense, resource management, and drone inspection.", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a47cef66bbead7e6e2bf9a2f294ce9d285d104eb044972bf0e5f2fdc2fa52920.jpg", + "image_caption": [ + "Figure 1. Comparisons of the pseudo-box optimization process. (a) The rigid sampling strategy in Point-to-Box [3] results in the inefficiency of the training process for low-quality pseudo-box labels. (b) The self-regularized sampling in our PointSR improves the stability and efficiency of the training process and thus obtains high-quality pseudo-box labels." + ], + "image_footnote": [], + "bbox": [ + 516, + 318, + 910, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, DroneDet models often experience a significant performance drop when applied to new scenes or datasets due to their limited generalization capabilities. A straightforward solution is to collect and annotate new datasets and retrain the models. However, annotating drone-view images with box-level annotations containing accurate position information is particularly challenging due to the prevalence of many densely packed small targets, far more complex than annotating ground-view images. Moreover, the diversity of the scenes and object types can lead to combinatorial explosions of labeling requirements, making the process infeasible in many cases. Consequently, constructing new datasets with box-level annotations is both time-consuming and costly when deploying drones in new scenarios.", + "bbox": [ + 509, + 598, + 906, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To alleviate the cost associated with box-level annotations, Image-Supervised Object Detection (ISOD) [1, 14, 26] has been extensively studied. ISOD treats images as a bag of regions, assuming at least one region contains the object and uses a Multiple Instance Learning (MIL) classifier to identify regions corresponding to image-level labels. De", + "bbox": [ + 509, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 45 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11707", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "spite its labeling efficiency, ISOD methods often perform poorly in complex scenarios such as drone-view images, with targets typically small and densely packed, due to the lack of essential position information in image-level labels.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome this limitation, Point-Supervised Object Detection (PSOD) has emerged as an alternative solution. PSOD provides minimal and yet additional distinctive position information for each object and has gained significant attention in tasks such as detection [10, 32], localization [30, 31], and segmentation [12, 20]. According to [17], the labor cost of point annotations per object is only about $1/18$ of that for bounding boxes, making PSOD a highly cost-effective solution for DroneDet applications.", + "bbox": [ + 89, + 154, + 482, + 291 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing PSOD methods [3, 10, 11, 15, 29, 32] can be broadly categorized into generative and discriminative approaches. Generative methods [10, 11, 29] predict a masked region covering a target labeled by a point and then derive a pseudo box by estimating the minimum circumscribed rectangle of the region. However, pixels in the predicted mask of each object are susceptible to misclassification, resulting in inaccurate and unstable pseudo-box predictions. Discriminative methods [3, 15, 32] first generate a bag of proposal boxes for each label and then select the optimal one as the final pseudo-box through an MIL-based optimization process, as illustrated in Fig. 1(a). A reasonable sampling strategy aiming to extract the labeled training set from an unlabeled dataset is critical for the successful training of the MIL classifier. Despite their potential, these methods still face two significant challenges, as shown in Fig. 2(a): 1) Static positive sampling in an anchor style leads to training instability, especially in the early stages (see the blue curves in Fig. 7); 2) Inefficient negative sampling that randomly samples background regions introduces numerous irrelevant negative samples, reducing training efficiency.", + "bbox": [ + 91, + 292, + 483, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose PointSR, an effective and robust point-supervised DroneDet framework with self-regularized sampling. As shown in Fig. 1(b), PointSR integrates temporal and informative constraints throughout the pseudo-box generation process to produce high-quality pseudo-box labels. Temporal-Ensembling Encoder (TE Encoder) first builds an anchor prototype library by collecting and encoding temporal information from pseudo-box labels during refinement. Then, the Coarse Pseudo-box Generation adjusts anchors using the anchor prototype library to produce better coarse pseudo boxes while collecting a set of informative samples for the subsequent refinement. Finally, during the Refinement stage, the collected informative samples are used to suppress low-confidence candidate positive samples to refine pseudo boxes. Benefiting from this self-regularized sampling strategy, PointSR not only ensures a more stable and efficient training process but also effectively improves the model's robustness to perturbation brought by manual labeling.", + "bbox": [ + 89, + 613, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9c7531681c4b12d3ae9dce9890344cc34d695c37ee8232a48466cf33bb2dafb8.jpg", + "image_caption": [ + "Figure 2. (a) Point-to-Box [3] uses a rigid sampling strategy with fixed positive proposals and randomly sampled negative background regions for training the MIL-based predictor. (b) PointSR adopts a self-regularized sampling strategy, incorporating temporal and informative constraints into the pseudo-box generation process for both positive and negative sampling." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 903, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions can be summarized as follows:", + "bbox": [ + 532, + 351, + 883, + 364 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We propose PointSR, a novel and robust framework for point-supervised DroneDet, capable of generating high-quality pseudo-box labels from point-level supervision.", + "2) We design a Temporal-Ensemble Encoder for stable pseudo-box prediction and an informative negative sampling strategy to enhance pseudo-box quality.", + "3) PointSR achieves an average $\\mathrm{AP}_{50}$ of $42.5\\%$ , $38.7\\%$ , and $35.0\\%$ on three drone-captured benchmark datasets, i.e., DroneVehicle [22], UAVDT [5], and VisDrone [34] with only point supervision, outperforming the previous best methods by $7.2\\%$ , $5.4\\%$ , and $2.6\\%$ , respectively." + ], + "bbox": [ + 511, + 366, + 903, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 546, + 653, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we briefly review the current state of research on ISOD, PSOD and PSOD for aerial images.", + "bbox": [ + 511, + 565, + 903, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image-Supervised Object Detection (ISOD)", + "text_level": 1, + "bbox": [ + 511, + 604, + 880, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ISOD refers to an approach in which a detector is trained with image-level annotations rather than box-level ones. The pioneering work WSDDNs [1] integrated the MIL into ISOD, employing a dual-stream structure to classify proposals, and enabling effective weakly-supervised object detection. Inspired by this, Tang et al. [24] designed a Proposal Cluster Learning (PCL) approach, which used proposal clusters as supervision to infer rough object locations. However, image-level annotations lack explicit position information, leading to low-quality proposals, such as background regions, in the potential object tool. This limitation makes it infeasible to directly apply ISOD to DroneDet.", + "bbox": [ + 511, + 627, + 906, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Point-Supervised Object Detection (PSOD)", + "text_level": 1, + "bbox": [ + 511, + 816, + 879, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PSOD is a relatively new field, with only a few methods [2, 3, 17, 21] published in recent years. By leveraging point-level annotations, PSOD enables more accurate object localization and classification compared to image-level", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11708", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "annotations. From the perspective of human cognition, Papadopoulos et al. [17] introduced the first model utilizing center-click annotations, pioneering the use of point supervision in object detection as an alternative to conventional box-level supervision. Ren et al. [21] proposed a unified framework that performs object detection across various annotation types. Chen et al. designed an MIL-based framework Point-to-Box [3] to achieve object detection by classifying proposals. In camouflaged object detection, [2] employed a novel point-based learning paradigm for the challenging task with an attention regulator to distinguish objects that conceal themselves in the environment. However, these PSOD methods only focus on objects in ground-view images and produce poor performance in complex aerial scenarios due to numerous densely packed small objects.", + "bbox": [ + 89, + 90, + 483, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. PSOD for Aerial Images", + "text_level": 1, + "bbox": [ + 89, + 325, + 313, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing PSOD methods for aerial images can be grouped into two categories: generative and discriminative fashions. The generative approaches, such as PLUG [10], employ a short path algorithm to produce pseudo boxes with pixel-level labels. However, the pseudo boxes rely heavily on the accuracy of pixel-level mask generation. If there is an outlier far from the object mask, the corresponding pseudo label will be seriously affected. In contrast, discriminative methods aim to improve the quality of pseudo labels through advanced sampling strategies. PointOBB [15] produced pseudo labels through the collaborative utilization of three distinctive views to achieve coupled optimization of scale and angle. Point2Rbox [32] combined point annotations and synthetic pattern knowledge with bounding boxes to improve the diversity of samples and provide more supervised information for accurate pseudo labels.", + "bbox": [ + 89, + 347, + 483, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, despite these advancements, existing methods still suffer from the limitation of fixed positive candidate proposals and the instability of the random distribution of negative samples. Specifically, the positive sample space remains constant during training and the negative one is only affected slightly by the coarse pseudo boxes in the conventional PSOD framework, formulated in Eq. (5). To address these limitations, we propose an effective and robust PSOD framework to generate high-quality pseudo boxes with a self-regularized sampling strategy, adjusting sampling space adaptively during the model training.", + "bbox": [ + 89, + 589, + 483, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 767, + 228, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our PointSR is built upon the Weakly Supervised Deep Detection Networks (WSDDNs) [1] for ISOD. The input data consists of a collection of images, each associated with image-level labels $\\mathbf{y} \\in \\{0,1\\}^C$ . Here, $C$ represents the number of object categories. Candidate object regions $\\mathcal{R} = \\{R_1,R_2,\\dots ,R_{|\\mathcal{R}|}\\}$ are generated by using region proposal mechanisms [25]. Features of these regions $\\mathbf{x} \\in \\mathbb{R}^{D \\times |\\mathcal{R}|}$", + "bbox": [ + 89, + 792, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are extracted through fully connected layers.", + "bbox": [ + 511, + 90, + 808, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "WSDDNs adopts a dual-stream structure, consisting of a classification and detection stream. The extracted features $\\mathbf{x}$ are fed into the two streams, producing two matrices, $\\mathbf{x}^c$ and $\\mathbf{x}^d\\in \\mathbb{R}^{C\\times |\\mathcal{R}|}$ , representing the category and position prediction scores for each region, respectively. These scores are then passed through the corresponding softmax operators $\\sigma_{\\mathrm{cla}}(\\cdot)$ and $\\sigma_{\\mathrm{det}}(\\cdot)$ , defined as:", + "bbox": [ + 511, + 106, + 906, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\sigma_ {\\mathrm {c l a}} \\left(\\mathbf {x} ^ {c}\\right) \\right] _ {i j} = \\frac {e ^ {x _ {i j} ^ {c}}}{\\sum_ {k = 1} ^ {C} e ^ {x _ {k j} ^ {c}}}, \\left[ \\sigma_ {\\det } \\left(\\mathbf {x} ^ {d}\\right) \\right] _ {i j} = \\frac {e ^ {x _ {i j} ^ {d}}}{\\sum_ {k = 1} ^ {| \\mathcal {R} |} e ^ {x _ {i k} ^ {d}}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 220, + 903, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The regions' scores are then calculated via element-wise multiplication as:", + "bbox": [ + 511, + 273, + 903, + 303 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\mathcal {R}} = \\sigma_ {\\operatorname {c l a}} \\left(\\mathbf {x} ^ {c}\\right) \\cdot \\sigma_ {\\det } \\left(\\mathbf {x} ^ {d}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 311, + 903, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, an image-level class prediction score is obtained by summing over all regions as:", + "bbox": [ + 511, + 339, + 903, + 369 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} _ {c} = \\sum_ {r = 1} ^ {| \\mathcal {R} |} \\mathbf {x} _ {c r} ^ {\\mathcal {R}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 378, + 903, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a proposal regions' sampling space $X$ and a label space $Y$ , features $\\mathbf{x}$ of proposal regions sampled from $X$ in an image are fed into an MIL-based classifier $f_{\\theta}(\\cdot)$ , parameterized by $\\theta$ , to obtain the image-level class prediction score $\\mathbf{y} \\in Y$ . Therefore, ISOD can be formulated as:", + "bbox": [ + 511, + 429, + 906, + 503 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\arg \\min } \\sum_ {k = 1} ^ {C} \\mathcal {L} \\left(f _ {\\theta} (\\mathbf {x}), \\mathbf {y} _ {k}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 513, + 903, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(\\cdot)$ is the model's loss function, and $k$ is the label's $k$ -th category. Note that the sampling space $X$ is constant during training.", + "bbox": [ + 511, + 564, + 906, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. The Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 622, + 723, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Problem Formulation and Overview", + "text_level": 1, + "bbox": [ + 511, + 647, + 826, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation: PSOD provides minimal and yet distinctive position information for each object. An input image $I$ contains $M$ targets, each associated with point-level labels $\\mathbf{y} \\in \\{0,1\\}^C$ . For the $i$ -th target, its positive samples $\\mathbf{x}_i^+$ with labels $\\mathbf{y}_i^+ \\in \\{0,1\\}^C$ are generated using a static anchor strategy, while its negative samples $\\mathbf{x}_i^-$ with labels $\\mathbf{y}_i^-\\in \\{0\\}$ are collected through random sampling on background. The negative sampling space is defined as the difference between the image $I$ and the predicted pseudo boxes' regions (PB), denoted as \"diff(I, PB)\".", + "bbox": [ + 511, + 669, + 906, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The optimization for PSOD can be formulated as:", + "bbox": [ + 532, + 820, + 862, + 835 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta , x ^ {-} \\in \\operatorname {d i f f} (I, \\mathrm {P B})} {\\arg \\min } \\sum_ {i = 1} ^ {M} \\sum_ {k = 1} ^ {C} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i, k} ^ {+}\\right), \\mathbf {y} _ {i, k} ^ {+}\\right) + \\sum_ {j = 1} ^ {N} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {j} ^ {-}\\right), \\mathbf {y} _ {j} ^ {-}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 844, + 916, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11709", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2e42270782033b04d9372b3b54c104559174b5be2fd72b8bc2801a390f720b7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b1066da5f5cbd22c7ee236f5f456c45c1ed13591ec598f67bb2951de87e2a291.jpg", + "image_caption": [ + "Figure 3. The architecture of PointSR, comprising the TE Encoder, Coarse Pseudo-box Prediction and Pseudo-box Refinement. The training process spans $T$ epochs, each consisting of two stages: Coarse Pseudo-box Prediction and Pseudo-box Refinement. The former generates coarse pseudo boxes and collects a set of negative samples by ranking their losses. The latter utilizes these negative samples and the positive samples from the previous stage to refine the pseudo boxes. The TE Encoder collects temporal information during each epoch to dynamically adjust candidate proposal generation at the beginning of the Coarse Pseudo-box Prediction stage." + ], + "image_footnote": [], + "bbox": [ + 93, + 226, + 903, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where positive samples $x^{+}$ of each target remain constant, and the negative sampling space $\\mathrm{diff}(I,\\mathrm{PB})$ depends on PB.", + "bbox": [ + 88, + 419, + 480, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the limitations of rigid sampling, we propose a self-regularized sampling mechanism, dynamically adjusting the positive $X^{+}$ and negative $X^{-}$ sampling spaces using temporal and informative constraints. According to Eq. (5), PointSR's optimization for PSOD is formulated as:", + "bbox": [ + 88, + 450, + 482, + 526 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\underset {\\theta , x ^ {+} \\in X ^ {+}, x ^ {-} \\in X ^ {-}} {\\arg \\min } \\sum_ {i = 1} ^ {M} \\sum_ {k = 1} ^ {C} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i, k} ^ {+}\\right), \\mathbf {y} _ {i, k} ^ {+}\\right) + \\tag {6} \\\\ \\sum_ {j = 1} ^ {N} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {j} ^ {-}\\right), \\mathbf {y} _ {j} ^ {-}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 536, + 482, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview: PointSR employs a Self-regularized Sampling mechanism to optimize the pseudo box generation from point-level annotations. Fig. 3 shows that the architecture of PointSR consists of three main components: TE Encoder, Coarse Pseudo-box Prediction, and Pseudo-box Refinement. TE Encoder aggregates temporal information from predicted pseudo labels in each epoch and encodes it into an anchor prototype library. Coarse Pseudo-box Prediction uses the anchor prototype library for anchor adjustment and predicts coarse pseudo boxes using an MIL-based classifier. It also collects informative negative samples for refinement. Finally, Pseudo-box Refinement employs the MIL-based classifier to further refine pseudo boxes by suppressing candidate positive samples with low confidence.", + "bbox": [ + 88, + 628, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Temporal-Ensembling Encoder", + "text_level": 1, + "bbox": [ + 89, + 848, + 367, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first step of predicting pseudo boxes is to create candidate boxes. Our approach predicts these candidate boxes in", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "an anchor style, done both statically and dynamically.", + "bbox": [ + 511, + 420, + 867, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the widely adopted static anchor strategy, anchors' width and height are pre-determined heuristically. Anchor shapes are set and remain constant throughout the entire training process. The static anchor strategy in Faster R-CNN [19] can be sub-optimal as it neglects the underlying distribution of real data, potentially leading to low-quality pseudo boxes and slow convergence during training.", + "bbox": [ + 511, + 436, + 903, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this limitation, we propose a learning-based dynamic anchor adjustment module, called \"TemporalEnsembling Encoder\" (TE Encoder), as illustrated in Fig. 3(d). TE Encoder focuses on encoding the ratio of objects' width to height instead of objects' scale because objects' aspect ratio is numerically more stable than object scales in drone-view images. Drone-view images show that object scales vary significantly depending on their distance from the drone-mounted camera, while an object's aspect ratio remains relatively stable, particularly in top-down aerial views in Fig 5. At the end of each training iteration, we can obtain high-quality pseudo boxes predicted in the refinement stage, serving as references to adjust anchors' shapes. In conclusion, very different from the pre-defined and fixed anchors, advanced anchors by TE Encoder are dynamically adjusted by the pseudo-box's aspect ratio derived from temporal information during training.", + "bbox": [ + 511, + 542, + 905, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, TE Encoder aggregates pseudo-boxes' shapes for each category in each iteration and stores them as the category-aware prototype library, $P \\in \\mathbb{R}^{C \\times 1}$ . For Epoch $t$ , the $c$ -th element $P_{c} \\in (0,1)$ , $c \\in \\{1,2,\\dots,C\\}$", + "bbox": [ + 511, + 800, + 903, + 861 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Setting: anchor shapes are preset as six scales (4, 8, 16, 32, 64, 128) with seven aspect ratios (1: 3, 1: 2, 1: 1.5, 1: 1, 1.5: 1, 2: 1, 3: 1).", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11710", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "represents the ratio of width $w_{tj} \\in \\mathbb{R}$ to height $h_{tj} \\in \\mathbb{R}$ of the $j$ -th object of category $c$ . $P_{c}$ is updated iteratively as:", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {c} ^ {\\prime} = \\lambda P _ {c} + (1 - \\lambda) \\frac {w _ {t j}}{h _ {t j}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 132, + 482, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is used to balance the contributions of knowledge distilled in current and previous epochs.", + "bbox": [ + 89, + 172, + 482, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Coarse Pseudo-Box Prediction", + "text_level": 1, + "bbox": [ + 89, + 212, + 362, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adjust anchors with the anchor prototype library for better coarse pseudo boxes and simultaneously collect a set of informative samples for the pseudo-box refinement.", + "bbox": [ + 89, + 233, + 483, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MIL-based Pseudo-Box Predictor: Given an adjusted bag of region proposals for each point label, a dual-stream structure is adopted as an MIL classifier to find the best box region in the bag to represent the target. As shown in Fig. 3(b), features of these proposals are extracted through a RoI [9] layer and two fully connected layers. Secondly, these features are fed into the dual-stream structure to estimate scores of each proposal in the same bag according to Eq. (1). Then, a bag's score is calculated by the sum of all its proposals. Finally, the Top- $k$ boxes with the highest scores in the same bag are chosen for coarse pseudo boxes.", + "bbox": [ + 89, + 279, + 483, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Informative-Sample Collector (IS Collector): To capture informative negative samples with high uncertainty, we present IS Collector, which aims to collect negatives with high loss values. Specifically, negative samples randomly sampled on background with their corresponding losses estimated by the MIL-based classifier, are fed into our IS collector for the pseudo-box refinement, as shown in Fig. 3(e). Considering the high computational costs of processing all negative samples, Non-Maximum Suppression [16] is first applied for de-duplication, removing redundant negative samples with high similarity and retaining representative ones. Then, we rank the filtered negative samples in descending order of their losses and thus collect the most informative samples for pseudo-box refinement.", + "bbox": [ + 89, + 446, + 483, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Pseudo-Box Refinement", + "text_level": 1, + "bbox": [ + 89, + 667, + 310, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The collected informative samples are used to suppress candidate positive samples with low confidence so as to refine pseudo boxes. As shown in Fig. 3(c), the model for pseudo-box refinement has a similar structure as the model for coarse pseudo-box prediction. Specifically, for positive sample generation, we update the aspect ratios of coarse pseudo boxes and jitter their center positions. Meanwhile, given the informative negative samples collected in coarse pseudo boxes, after the RoI layer, all features of samples are fed into the MIL classifier according to Eq. (1). Finally, the relatively accurate Top- $k$ candidate proposals are chosen to be merged as the final predicted pseudo box.", + "bbox": [ + 89, + 689, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion: As shown in Fig. 4(b), numerous proposals by random sampling in the background are far from targets.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0b12f872133572c7a7b9125389ed278146c70d54867e6e5aa12b6cffed53d2aa.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 622, + 162 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ac7af72bf99260191e05b56c9e8aeebb5c718b94c5c06fa39816e934ea3cda69.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 622, + 90, + 795, + 171 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/388ed6638282936231a959a86173cedc143ed52a185281db238e1651b9fa3141.jpg", + "image_caption": [ + "(c)", + "Figure 4. The negative samples distribution of PointSR and Point-to-Box. The histogram (b) shows the number of negative samples for different generation methods, where the distance denotes the Euclidean distance between the center of negative samples (Background) and positive samples (Pedestrians). Two specific examples for PointSR (a) and Point-to-Box (c) are visualized." + ], + "image_footnote": [], + "bbox": [ + 795, + 90, + 900, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, Fig. 4(c) shows partial negative samples, indicating that they provide few contributions to object detection. As for densely packed objects in drone-view images, edge information between objects and objects' neighborhood is crucial to distinguish each object. The random negative sampling in the background only generates proposals in a relatively stable sample space. While this sampling ensures diversity among negative samples, it may include less relevant ones, reducing learning efficiency. Additionally, since this method assigns equal weights to all negative samples, it can not adaptively select useful and informative negatives.", + "bbox": [ + 511, + 289, + 905, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 4(a) presents that negative samples produced by our IS Collector are close enough to the decision boundary in feature space. Training with these negative samples is beneficial to the MIL classifier to recognize which candidate proposals should be assigned low scores and force the model to learn finer distinctions because these negatives contribute more significantly to the gradients. By combining with the IS Collector, the TE Encoder refines positive samples through anchor aspect ratio adjustments, enhancing the MIL classifier's ability to predict losses for negative samples more accurately. This improvement enables the IS Collector to better select high-loss negative samples. As a result, the collaboration between the two modules yields high-quality pseudo-boxes.", + "bbox": [ + 511, + 460, + 905, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Analysis of Scalability: The time complexity of TE Encoder, $\\mathbf{O}(n)$ , scales linearly with the number of the predicted pseudo boxes $n$ . Whereas the IS Collector selects a few informative samples from around 400 negative samples randomly sampled from the background. Since this process is independent of $n$ and the image resolution, PointSR's computational overhead is stable, even for high-resolution images or larger datasets.", + "bbox": [ + 511, + 676, + 905, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Loss Functions", + "text_level": 1, + "bbox": [ + 511, + 823, + 666, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by the discriminative method [3], we design the MIL loss and negative loss for our PointSR training.", + "bbox": [ + 511, + 849, + 903, + 880 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The MIL Loss: We formulate the MIL loss $\\mathcal{L}_{mil}$ as a cross", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11711", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "entropy loss for positive samples, given by:", + "bbox": [ + 89, + 90, + 379, + 107 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m i l} = - \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {k = 1} ^ {C} \\log \\left(\\hat {\\mathbf {y}} _ {j, k} ^ {\\mathbf {y} _ {j, k}} \\left(1 - \\hat {\\mathbf {y}} _ {j, k}\\right) ^ {\\left(1 - \\mathbf {y} _ {j, k}\\right)}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 119, + 482, + 164 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{y}_{j,k} \\in \\{0,1\\}$ denotes the one-hot category labels and $\\hat{\\mathbf{y}}_{jk}$ is the predict scores generated by MIL-based predictor. The MIL loss is used for the correct prediction of the category and instance to which the proposal belongs to.", + "bbox": [ + 89, + 174, + 482, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Negative Loss: Furthermore, the negative loss $\\mathcal{L}_{\\text{neg}}$ in the form of focal loss is used for model optimization, which is formulated as follows:", + "bbox": [ + 89, + 234, + 483, + 280 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n e g} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\sum_ {k = 1} ^ {C} \\beta \\cdot (\\hat {\\mathbf {y}} _ {j, k}) ^ {\\gamma} \\log (1 - \\hat {\\mathbf {y}} _ {j, k}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 289, + 482, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma$ is set to 0.2, the same in [13] and $\\beta$ is a weight to balance the contribution of different negative samples. The Negative loss in the Refinement stage is used to further suppress the background for pseudo-box refinement.", + "bbox": [ + 89, + 342, + 483, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 416, + 223, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 440, + 192, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted experiments on three benchmark datasets: DroneVehicle [22], VisDrone [34] and UAVDT [5].", + "bbox": [ + 89, + 462, + 482, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**DroneVehicle:** The DroneVehicle dataset includes 56,878 images, which are annotated with box-level annotations across five vehicle categories, i.e., Car, Truck, Bus, Van and Freight Car, covering various scenarios, including urban roads, residential areas, and parking lots captured under different lighting conditions from day to night.", + "bbox": [ + 89, + 493, + 483, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VisDrone: The VisDrone dataset consists of 10,209 images with a resolution of $2,000 \\times 1$ , 500 pixels and 542,000 instances across 10 common object categories in traffic scenarios, i.e., Pedestrian, People, Bicycle, Car, Van, Truck, Tricycle, Awning-tricycle, Bus and Motor. These images, captured from various urban scenes, present significant challenges due to the presence of numerous small objects, varying viewpoints, and heavy occlusions. To the hardware memory limitations, following PLUG [10], we cropped the original images into patches of $512 \\times 512$ pixels with a 128-pixel overlap for training and inference.", + "bbox": [ + 89, + 584, + 482, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UAVDT: The UAVDT dataset consists of 100 video sequences of about 80,000 representative frames with a resolution of $1,024 \\times 540$ pixels and 0.84 million box-level annotations across three vehicle categories: Car, Truck, Bus.", + "bbox": [ + 89, + 750, + 482, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To accurately reproduce the biases inherent in manual annotation, we employed Rectified Gaussian Distribution (RG) [30] with central ellipse constraints, assuming that the point distribution of manual annotation approximately follows an RG distribution. Therefore, for our experiments, the three datasets with point-level annotations are employed", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for evaluating our approach, referred to as \"DroneVehicle (Point-Supervision)\", \"VisDrone (Point-Supervision)\", and \"UAVDT (Point-Supervision)\", respectively.", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 147, + 857, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details: We implemented our proposed PointSR based on the MMDetection toolbox on an NVIDIA RTX 4090 GPU. The Stochastic Gradient Descent (SGD) algorithm was used to optimize with a $3 \\times$ training schedule. The learning rate was set to 0.02 and decayed by a factor of 0.1 at the 24th and 33rd epochs, respectively. All comparison models were configured with a ResNet-50 [8] backbone pre-trained on ImageNet [4].", + "bbox": [ + 511, + 170, + 906, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics: We evaluated our method from three aspects: the quality of predicted pseudo boxes, the detection performance and the model's robustness. The mean Intersection over Union (mIoU) between generated pseudo boxes and ground-truth boxes was utilized to evaluate the quality of predicted pseudo boxes, where mIoU closer to 1 indicates higher accuracy. The Average Precision $(\\mathrm{AP}_{50})$ [6] was utilized to evaluate the models' detection performance. Finally, the variances of AP across different point-label distributions were used to measure models' robustness, with lower variance indicating higher robustness.", + "bbox": [ + 511, + 291, + 908, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 511, + 468, + 805, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compared the performance of a detector trained using pseudo boxes generated by our PointSR and five SOTA point-supervised DroneDet methods: PLUG [10], PointOBB [15], Point2Rbox [32], PointOBBv2 [18] and Point-to-Box [3]. In these experiments, pseudo boxes were first generated by our PointSR and SOTA methods. These pseudo boxes were then utilized to train a classic detector, Faster-RCNN [19], to evaluate its detection performance and robustness. In addition, a Faster-RCNN detector was trained with ground-truth box-level annotations, which detection performance was referred to as the upper bound.", + "bbox": [ + 511, + 491, + 906, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Detection Performance: Table 1 shows that the detector trained with pseudo boxes generated by our PointSR achieves superior detection performance, outperforming that trained with pseudo boxes generated by SOTA methods significantly on all datasets. Specifically, our model improved $\\mathrm{AP}_{50}$ by $7.2\\%$ on DroneVehicle (Point-Supervision) dataset, $3.3\\%$ on VisDrone (Point-Supervision) dataset and $5.4\\%$ on UAVDT (Point-Supervision) dataset, compared to Point-to-Box [3]. Besides, it is worth mentioning that the detection results of our model trained with point-level annotations are better than the results of Faster-RCNN trained with box-level annotations in some cases in Fig. 5.", + "bbox": [ + 511, + 657, + 908, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Quality of the Predicted Pseudo-Boxes: We use mIoU between the generated pseudo boxes and ground-truth boxes to evaluate the quality of the predicted pseudo boxes, where mIoU closer to 1 indicates higher accuracy.", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11712", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ac15fad5c68f108e5db8238e2c05baf4b667573c78e60876be0ee07e51c77b7a.jpg", + "table_caption": [ + "Table 1. Comparisons of per-class detection accuracy $(\\mathrm{AP}_{50})$ across different object classes, average $\\mathrm{AP}_{50}$ , and mIoU for a detector trained using pseudo boxes generated by our PointSR and SOTA methods on VisDrone, DroneVehicle and UAVDT. \"Faster R-CNN\" indicates the accuracy obtained with the detector trained with ground-truth box-level annotations, serving as the upper bound for performance." + ], + "table_footnote": [], + "table_body": "
MethodVisDroneDroneVehicleUAVDT
Ped.Peo.Bic.CarVanTru.Tri.Awn.BusMot.AP50mIoUCarTru.BusVanFre.AP50mIoUCarTru.BusAP50mIoU
PLUG [10]0.63.44.73.40.24.50.70.323.65.34.726.91.06.10.28.02.33.529.04.58.23.85.531.1
Point2Rbox [32]0.30.29.122.315.315.311.59.915.81.410.1-0.03.675.718.928.025.2-15.724.119.219.7-
PointOBB [15]25.613.620.834.612.430.732.930.456.512.427.056.218.37.06.034.510.115.277.529.230.816.825.672.6
PointOBBv2 [18]33.24.316.169.257.431.632.123.232.823.832.464.817.119.734.229.222.824.680.116.628.837.227.577.5
Point-to-Box [3]40.24.423.162.953.130.930.922.828.419.631.766.268.925.524.233.024.735.386.045.535.818.533.395.3
Faster R-CNN [19]53.730.930.072.257.644.231.325.035.837.241.8-89.766.092.253.352.770.1-66.955.759.660.7-
PointSR (Ours)43.45.826.265.653.531.529.722.444.727.735.075.782.825.133.542.229.142.588.950.140.825.438.796.8
", + "bbox": [ + 91, + 136, + 903, + 272 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a31108634041c8e16f28e1b27dc685d4d14e3ee52cceb3a8b8ef9d42ed7485da.jpg", + "table_caption": [ + "Table 2. Comparison of average precision $(\\mathrm{AP}_{50})$ and variance for a detector trained using pseudo boxes generated by our PointSR and SOTA methods across different point-label distributions. PointSR consistently demonstrates lower variance on both datasets, exhibiting greater stability and robustness to annotation perturbations." + ], + "table_footnote": [], + "table_body": "
MethodPublicationVisDroneDroneVehicle
U0×LU0.1×LU0.3×LU0.5×LVariance Of AP50U0×LU0.1×LU0.3×LU0.5×LVariance Of AP50
PLUG [10]TGRS'241.54.315.517.848.911.13.930.132.5209.1
Point2Rbox [32]CVPR'247.69.98.39.10.74215.224.117.618.310.61
PointOBB [15]CVPR'2424.328.517.88.556.9113.415.212.68.95.269
Point-to-Box [3]ECCV'2231.632.531.130.10.75234.835.532.934.00.935
PointSR (Ours)N/A34.634.833.933.70.21241.142.041.340.90.171
", + "bbox": [ + 99, + 324, + 893, + 426 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f51b1844acf38fa4711b569f42a2f4f737eb9f3b6abe4fc1f10e3fb98f8b2baf.jpg", + "table_caption": [ + "Table 3. The impact of the TE Encoder and Informative Sampling (IS) on the baseline Point-to-Box [3], evaluated on DroneVehicle." + ], + "table_footnote": [], + "table_body": "
TEEISAP50mIoU [%]
Baseline××35.2686.0
Baseline + TEE✓×36.1886.7
Baseline + IS×✓40.8788.1
Baseline + TEE + IS✓✓42.5388.9
", + "bbox": [ + 124, + 468, + 450, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tables 1 also show that PointSR achieves better mIoUs than the SOTA methods on both datasets. Specifically, our model improves mIoU by $9.5\\%$ on Visdrone, $2.9\\%$ on DroneVehicle and $1.5\\%$ on UAVDT compared to Point-to-Box [3].", + "bbox": [ + 89, + 564, + 482, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The visualization in Fig. 6 shows that our predicted pseudo boxes accurately localize objects in challenging drone-view images by leveraging the precise location information of point annotations and can distinguish dense objects in complex scenes. In contrast, Point-to-Box often produces loose pseudo boxes due to its limited sampling strategy for MIL training, and PLUG occasionally predicts oversized boxes due to outliers in the generative process. These results confirm the high quality of pseudo boxes generated by our model with point-level annotations.", + "bbox": [ + 88, + 627, + 482, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Robustness of the Model: To evaluate the model's robustness, we use the variance of $\\mathrm{AP}_{50}$ across different point distributions. Specially, we assume manually annotated points follow a uniform distribution $x\\sim U(-a,a)$ , $a = \\alpha \\times L$ , $\\alpha = \\{0,0.1,0.3,0.5\\}$ and $L$ is the maximum value for the width and height of a bounding box. $U_{0\\times L}$ , $U_{0.1\\times L}$ , $U_{0.3\\times L}$ , and $U_{0.5\\times L}$ denote the manually annotated points $x\\sim U(-\\alpha \\times L,\\alpha \\times L)$ with offsets of 0, $10\\%$ ,", + "bbox": [ + 89, + 779, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$30\\%$ , $50\\%$ , respectively. As shown in Table 2, our model has achieved a variance of $\\mathrm{AP}_{50}$ 0.171 on the DroneVehicle dataset, compared to 0.935 of Point-to-Box [3], indicating superior robustness to annotation perturbation.", + "bbox": [ + 511, + 438, + 906, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 511, + 679, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted ablation studies on the Drone Vehicle (Point-Supervision) dataset using Point-to-Box [3] as the baseline to validate the effectiveness of the two key components in PointSR: the TE Encoder and IS Collector.", + "bbox": [ + 511, + 535, + 906, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Effectiveness of the TE Encoder: We conducted a quantitative evaluation to verify the effectiveness of the TE Encoder in our PointSR. The baseline with TE Encoder is referred to as \"Baseline + TEE\". Table 3 shows that our proposed module improves $\\mathrm{AP}_{50}$ by $0.92\\%$ and mIoU by $0.7\\%$ , compared with the vanilla baseline model. Moreover, TE Encoder facilitates stable and efficient training, in Fig. 7(a), where the orange curve (Baseline + TEE) descends faster and is smoother compared with the blue curve, especially in the early stage of training, indicating that our Anchor Learning model achieves stable training and faster convergence. In addition, Fig. 7(b) further demonstrates that \"Baseline + TEE\" achieves better detection performance. Finally, Table 4 shows the performance of Faster R-CNN trained using PointSR with different $\\lambda$ settings, where PointSR with $\\lambda = 0.5$ achieves the best detection performance. These results verify the effectiveness of the TE Encoder.", + "bbox": [ + 511, + 595, + 908, + 852 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/846a5eb1783baefce740347df24195d7e8419bf241b1ddc35def3c551db9fbdb.jpg", + "table_caption": [ + "Table 4. Impact of different $\\lambda$ values on ${\\mathrm{{AP}}}_{50}$ in DroneVehicle." + ], + "table_footnote": [], + "table_body": "
λ0.10.30.50.70.9
AP50 [%]34.240.542.534.933.9
", + "bbox": [ + 586, + 875, + 831, + 911 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11713", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ab5bd1e68a37f9fac28eb85b7fe29b67fb21168c48516db8c48e9fd4993b7d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 88, + 222, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/842e2739c5726a3597088ede6ec2bc4fae4ae9566c2b503187565406eb8752dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 88, + 334, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fee694ba271e870394dbb783cbbedd4e4f9d1427d4e13232b7478a2d0ccf0a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 88, + 442, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c6aed0d7d051948b739b514ad9445f14ad50e5e966cbaabe5ae1d1b0aa34a09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 88, + 553, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19b0f7f2bcac7ca0a6f5480ff581bef80d20fe23a34194da6d58815bd6bd035d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 88, + 663, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d7ff54c9952ae1433f159d6eab948c5dea8a890650e0d04e07ebed100e594c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 88, + 774, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/193dbc5dc579c5ef93983df44dabb5f6591223a5a1405784a732231fc631bd72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 88, + 883, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de0e591a03e899085ec4cb1c1ca60a35dd6aa0c5bb176949035417608204ce91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 174, + 222, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15ebc8de4d615f4298ae1809deb56d3f2acf50c55d7d6c4fab1be7cb983c6599.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 174, + 333, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45f50f650bc813487a0fdf273a3cccb8a9720e1f6b8b12e720aa498ef631aaed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 174, + 442, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ebbd4f1405439f97eef5400133c70fef5dcc3614b63dd812d0b0be8f014e2104.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 174, + 553, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eadcc5a9674d5551c82e162e15f790f673d20caeebad517115f806787e6b7445.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 174, + 663, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df3beca3190aa231e17e22992f573b72bea68251efd4a8b02e36479609da4f83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 174, + 774, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/373ea0f28963dde991576207a441795b4f99c94d00e023ad4525dd9590429d99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 174, + 883, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/49346a8337214009e048e2780b51c7f38c55ff90cf7d14379ae691f31b2295f0.jpg", + "image_caption": [ + "PLUG" + ], + "image_footnote": [], + "bbox": [ + 114, + 258, + 222, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7800909db44a5597fe4b9c721cf8ceeb94609423177a6f3a9838ea772886406a.jpg", + "image_caption": [ + "Point2Rbox" + ], + "image_footnote": [], + "bbox": [ + 223, + 258, + 333, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d618c06ed51f814bbb66ba92c2a41aa015c92c14df6e21e39a5fdbb578123247.jpg", + "image_caption": [ + "PointOBB" + ], + "image_footnote": [], + "bbox": [ + 334, + 258, + 442, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78311976e08c06e25af1eccfeaf10319b213bc33fb87f57abd132a6e24990527.jpg", + "image_caption": [ + "Point-to-Box" + ], + "image_footnote": [], + "bbox": [ + 444, + 258, + 553, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4ea27984c2b42950187254adb2ce9975cf3c006901e38dd526532bc9bccc170.jpg", + "image_caption": [ + "Faster-RCNN" + ], + "image_footnote": [], + "bbox": [ + 555, + 258, + 663, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/947e3b7c408d04e477740e93378035313dc19aa8d0bc59ed667f4ddbbb342c90.jpg", + "image_caption": [ + "PointSR(Ours)" + ], + "image_footnote": [], + "bbox": [ + 665, + 258, + 774, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66ec264911bcbb0f672f0a5d6baeec402ebb848e0f45fd1b1aec803c273774a8.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 774, + 258, + 883, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f72046695e71cdb2af191a7f3aa0a63d595adaf0135dfaf762b27c170dec981b.jpg", + "image_caption": [ + "Figure 5. Visualization of detection results from a detector trained with pseudo boxes generated by PointSR and SOTA methods. Ground-truth bounding boxes are shown in green, correct detections in yellow, and misclassified or missed targets in red. For a better view, detection results in the last row are highlighted by cyan rectangular boxes. Best viewed in color and zoomed-in mode." + ], + "image_footnote": [], + "bbox": [ + 107, + 412, + 467, + 635 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e049363795044915f1b0457cd1491b3ac8dc0d07346ee85c154015fcf99db252.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 94, + 688, + 285, + 779 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a53ad46899b8d191dbd8e38ec952c2af9cf29d76cfba86a83c2068ff7b53369.jpg", + "image_caption": [ + "Figure 6. Visualization of pseudo boxes (pink) predicted by PointSR and SOTA methods. Ground-truth boxes are shown in green. Best viewed in color and zoomed-in mode.", + "(b)", + "Figure 7. The loss and $\\mathrm{AP}_{50}$ curves during the whole training stage. The blue curves represent baseline training, while the orange curves show baseline training equipped with our TE Encoder." + ], + "image_footnote": [], + "bbox": [ + 289, + 686, + 480, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Effectiveness of the IS Collector: We evaluated the effectiveness of the Informative Sampling Collector in our PointSR. Table 3 shows that the baseline with our Informative Sampling module, referred to as \"Baseline + IS\",", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b57e286ba27eb69786d64ee192dbbdac30c29ce5a9874c3afd44af10d63ffee.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 517, + 414, + 710, + 505 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2db59f364d79f76298aa0506db13f17cfc78c846879521a3e68c4ae9cedf095.jpg", + "image_caption": [ + "(b)", + "Figure 8. The $\\mathrm{AP}_{50}$ and IoU curves for varying Sampling Ratio $\\delta$ . improves $\\mathrm{AP}_{50}$ by $5.26\\%$ and mIoU by $2.2\\%$ . These experimental results confirm the effectiveness of our IS Collector. In addition, we also analyzed the impact of the ratio of selected negative samples to whole negative samples $\\delta$ , denoted as \"Sampling Ratio\", on detection performance. Fig. 8 visualizes that changes in $\\delta$ significantly affect the detection performance of our PiontSR in terms of $\\mathrm{AP}_{50}$ and mIoU, suggesting its critical role in optimizing results." + ], + "image_footnote": [], + "bbox": [ + 712, + 414, + 905, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Towards generating high-quality pseudo boxes for PSOD from drone-view images, in this paper, we proposed an effective and robust self-regularized point-supervision approach, PointSR, which leverages a Temporal-Ensemble Encoder for stable pseudo-box generation and an Informative Sampling Collector for efficient pseudo-box refinement. The two processes mutually benefit each other, facilitating the generation of self-regularized high-quality pseudo boxes. Experimental results on benchmark datasets have demonstrated that detectors trained with pseudo boxes generated by PointSR significantly outperform those generated by SOTAs over $2.6\\% \\sim 7.2\\%$ in DroneDet and gain the robust results under perturbations of the manually annotated labels, significantly advancing the applicability of PSOD.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11714", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This work was supported by the National Natural Science Foundation of China (Grant No. 62402370, 62301397), the Natural Science Basic Research Program of Shaanxi (Grant No. 2024JC-YBQN-0728), the Guangzhou Basic and Applied Basic Research Foundation (Grant No. 2023A04J1742). Additionally, we would like to thank Zhiwen Wei for his helpful discussions and insightful suggestions.", + "bbox": [ + 89, + 90, + 485, + 203 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 233, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2854, 2016. 1, 2, 3", + "[2] Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. Just a Hint: Point-Supervised Camouflaged Object Detection. In European Conference on Computer Vision, pages 332-348. Springer, 2024. 2, 3", + "[3] Pengfei Chen, Xuehui Yu, Xumeng Han, Najmul Hassan, Kai Wang, Jiachen Li, Jian Zhao, Humphrey Shi, Zhenjun Han, and Qixiang Ye. Point-to-Box Network for Accurate Object Detection via Single Point Supervision. In European Conference on Computer Vision, pages 51-67. Springer, 2022. 1, 2, 3, 5, 6, 7", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-scale Hierarchical Image Database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 6", + "[5] Dawei Du, Yuankai Qi, Hongyang Yu, Yifan Yang, Kaiwen Duan, Guorong Li, Weigang Zhang, Qingming Huang, and Qi Tian. The Unmanned Aerial Vehicle Benchmark: Object Detection and Tracking. In Proceedings of the European Conference on Computer Vision, pages 370-386, 2018. 2, 6", + "[6] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes Challenge: A Retrospective. International Journal of Computer Vision, 111:98-136, 2015. 6", + "[7] Xiaoxu Feng, Xiwen Yao, Gong Cheng, and Junwei Han. Weakly Supervised Rotation-Invariant Aerial Object Detection Network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14146-14155, 2022. 1", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6", + "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In Proceedings of the IEEE International Conference on Computer Vision, pages 2961-2969, 2017. 5", + "[10] Shitian He, Huanxin Zou, Yingqian Wang, Boyang Li, Xu Cao, and Ning Jing. Learning Remote Sensing Object Detection With Single Point Supervision. IEEE Transactions" + ], + "bbox": [ + 93, + 258, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Geoscience and Remote Sensing, 62:1-16, 2024. 2, 3, 6, 7", + "[11] Boyang Li, Yingqian Wang, Longguang Wang, Fei Zhang, Ting Liu, Zaiping Lin, Wei An, and Yulan Guo. Monte Carlo Linear Clustering with Single-Point Supervision is Enough for Infrared Small Target Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1009–1019, 2023. 2", + "[12] Jing Li, Junsong Fan, and Zhaoxiang Zhang. Point-Supervised Panoptic Segmentation via Estimating Pseudo Labels from Learnable Distance. In European Conference on Computer Vision, pages 95-112. Springer, 2024. 2", + "[13] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal Loss for Dense Object Detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 6", + "[14] Junyan Lu, Qinglei Hu, Ruifei Zhu, Yali Wei, and Tie Li. AFWS: Angle-Free Weakly-Supervised Rotating Object Detection for Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1", + "[15] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. PointOBB: Learning Oriented Object Detection via Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16730-16740, 2024. 2, 3, 6, 7", + "[16] Alexander Neubeck and Luc Van Gool. Efficient Nonmaximum Suppression. In 18th International Conference on Pattern Recognition, pages 850-855. IEEE, 2006. 5", + "[17] Dim P Papadopoulos, Jasper RR Uijlings, Frank Keller, and Vittorio Ferrari. Training Object Class Detectors with Click Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6374-6383, 2017. 2, 3", + "[18] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection. In The Thirteenth International Conference on Learning Representations, 2025. 6, 7", + "[19] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2016. 4, 6, 7", + "[20] Tianqi Ren, Qiu Shen, Ying Fu, and Shaodi You. Point-Supervised Semantic Segmentation of Natural Scenes via Hyperspectral Imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1357-1367, 2024. 2", + "[21] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, Ming-Yu Liu, Alexander G Schwing, and Jan Kautz. $\\mathbf{UFO}^2$ : A Unified Framework Towards Omni-Supervised Object Detection. In European Conference on Computer Vision, pages 288-313. Springer, 2020. 2, 3", + "[22] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Drone-Based RGB-Infrared Cross-Modality Vehicle Detection via Uncertainty-Aware Learning. IEEE Transactions on Circuits and Systems for Video Technology, 32(10):6700-6713, 2022. 2, 6" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11715", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Zhiwen Tan, Zhiguo Jiang, Chen Guo, and Haopeng Zhang. WSODet: A Weakly Supervised Oriented Detector for Aerial Object Detection. IEEE Transactions on Geoscience and Remote Sensing, 61:1-12, 2023. 1", + "[24] Peng Tang, Xinggang Wang, Song Bai, Wei Shen, Xiang Bai, Wenyu Liu, and Alan Yuille. PCL: Proposal Cluster Learning for Weakly Supervised Object Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(1): 176-191, 2018. 2", + "[25] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective Search for Object Recognition. International Journal of Computer Vision, 104: 154-171, 2013. 3", + "[26] Yuting Wang, Velibor Ilic, Jiatong Li, Branislav Kisačanin, and Vladimir Pavlovic. ALWOD: Active Learning for Weakly-Supervised Object Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6459–6469, 2023. 1", + "[27] Yue Xi, Wenjing Jia, Qiguang Miao, Junmei Feng, Jinchang Ren, and Heng Luo. Detection-Driven Exposure-Correction Network for Nighttime Drone-View Object Detection. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1", + "[28] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Query-Det: Cascaded Sparse Query for Accelerating High-Resolution Small Object Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13668–13677, 2022. 1", + "[29] Xinyi Ying, Li Liu, Yingqian Wang, Ruojing Li, Nuo Chen, Zaiping Lin, Weidong Sheng, and Shilin Zhou. Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15528-15538, 2023. 2", + "[30] Xuehui Yu, Pengfei Chen, Di Wu, Najmul Hassan, Guorong Li, Junchi Yan, Humphrey Shi, Qixiang Ye, and Zhenjun Han. Object Localization Under Single Coarse Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4868-4877, 2022. 2, 6", + "[31] Xuehui Yu, Pengfei Chen, Kuiran Wang, Xumeng Han, Guorong Li, Zhenjun Han, Qixiang Ye, and Jianbin Jiao. CPR++: Object Localization via Single Coarse Point Supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[32] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16783-16793, 2024. 2, 3, 6, 7", + "[33] Xiang Yuan, Gong Cheng, Kebing Yan, Qinghua Zeng, and Junwei Han. Small Object Detection via Coarse-to-Fine Proposal Generation and Imitation Learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6317-6327, 2023. 1", + "[34] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and Tracking" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Meet Drones Challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 2, 6", + "bbox": [ + 545, + 90, + 903, + 132 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11716", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/f94115e7-2e5c-44d9-81fe-c067d1bde80f_model.json b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/f94115e7-2e5c-44d9-81fe-c067d1bde80f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..171de541e411a3b75d1f744d0e3e51a2546afa2e --- /dev/null +++ b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/f94115e7-2e5c-44d9-81fe-c067d1bde80f_model.json @@ -0,0 +1,2508 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.046 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.892, + 0.152 + ], + "angle": 0, + "content": "PointSR: Self-Regularized Point Supervision for Drone-View Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.179, + 0.93, + 0.201 + ], + "angle": 0, + "content": "Weizhuo Li\\(^{1}\\) Yue Xi\\(^{1*}\\) Wenjing Jia\\(^{2}\\) Zehao Zhang\\(^{1}\\) Fei Li\\(^{3}\\) Xiangzeng Liu\\(^{3}\\) Qiguang Miao\\(^{3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.201, + 0.734, + 0.217 + ], + "angle": 0, + "content": "1Guangzhou Institute of Technology, Xidian University, Guangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.218, + 0.765, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\) Faculty of Engineering and IT, University of Technology Sydney, Sydney, Australia" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.236, + 0.747, + 0.251 + ], + "angle": 0, + "content": "\\(^{3}\\) School of Computer Science and Technology, Xidian University, Xi'an, China" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.254, + 0.842, + 0.269 + ], + "angle": 0, + "content": "{liweizhuo@stu., xiyue@, zzehao@stu., lifei@, xzliu@, qgmiao@}xidian.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.273, + 0.609, + 0.286 + ], + "angle": 0, + "content": "Wenjing.Jia@uts.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.322, + 0.326, + 0.337 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.485, + 0.778 + ], + "angle": 0, + "content": "Point-Supervised Object Detection (PSOD) in a discriminative style has recently gained significant attention for its impressive detection performance and cost-effectiveness. However, accurately predicting high-quality pseudo-box labels for drone-view images, which often feature densely packed small objects, remains a challenge. This difficulty arises primarily from the limitation of rigid sampling strategies, which hinder the pseudo-box optimization process. To address this, we propose PointSR, an effective and robust point-supervised object detection framework with self-regularized sampling that integrates temporal and informative constraints throughout the pseudo-box generation process. Specifically, the framework comprises three key components: Temporal-Ensembling Encoder (TE Encoder), Coarse Pseudo-box Prediction, and Pseudo-box Refinement. The TE Encoder builds an anchor prototype library by aggregating temporal information for dynamic anchor adjustment. In Coarse Pseudo-box Prediction, anchors are refined using the prototype library, and a set of informative samples is collected for subsequent refinement. During Pseudo-box Refinement, these informative negative samples are used to suppress low-confidence candidate positive samples, thereby improving the quality of the pseudoboxes. Experimental results on benchmark datasets demonstrate that PointSR significantly outperforms state-of-the-art methods, achieving up to \\(2.6\\% \\sim 7.2\\%\\) higher \\(AP_{50}\\) using only point supervision. Additionally, it exhibits strong robustness to perturbation in human-labeled points." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.792, + 0.222, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Drone-view object detection (DroneDet) [7, 23, 27, 28, 33], which involves detecting objects from drone-captured images, plays a vital role in various domains, including national defense, resource management, and drone inspection." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.319, + 0.911, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.51, + 0.907, + 0.594 + ], + "angle": 0, + "content": "Figure 1. Comparisons of the pseudo-box optimization process. (a) The rigid sampling strategy in Point-to-Box [3] results in the inefficiency of the training process for low-quality pseudo-box labels. (b) The self-regularized sampling in our PointSR improves the stability and efficiency of the training process and thus obtains high-quality pseudo-box labels." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.599, + 0.907, + 0.81 + ], + "angle": 0, + "content": "However, DroneDet models often experience a significant performance drop when applied to new scenes or datasets due to their limited generalization capabilities. A straightforward solution is to collect and annotate new datasets and retrain the models. However, annotating drone-view images with box-level annotations containing accurate position information is particularly challenging due to the prevalence of many densely packed small targets, far more complex than annotating ground-view images. Moreover, the diversity of the scenes and object types can lead to combinatorial explosions of labeling requirements, making the process infeasible in many cases. Consequently, constructing new datasets with box-level annotations is both time-consuming and costly when deploying drones in new scenarios." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To alleviate the cost associated with box-level annotations, Image-Supervised Object Detection (ISOD) [1, 14, 26] has been extensively studied. ISOD treats images as a bag of regions, assuming at least one region contains the object and uses a Multiple Instance Learning (MIL) classifier to identify regions corresponding to image-level labels. De" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11707" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "spite its labeling efficiency, ISOD methods often perform poorly in complex scenarios such as drone-view images, with targets typically small and densely packed, due to the lack of essential position information in image-level labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.292 + ], + "angle": 0, + "content": "To overcome this limitation, Point-Supervised Object Detection (PSOD) has emerged as an alternative solution. PSOD provides minimal and yet additional distinctive position information for each object and has gained significant attention in tasks such as detection [10, 32], localization [30, 31], and segmentation [12, 20]. According to [17], the labor cost of point annotations per object is only about \\(1/18\\) of that for bounding boxes, making PSOD a highly cost-effective solution for DroneDet applications." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.294, + 0.485, + 0.612 + ], + "angle": 0, + "content": "Existing PSOD methods [3, 10, 11, 15, 29, 32] can be broadly categorized into generative and discriminative approaches. Generative methods [10, 11, 29] predict a masked region covering a target labeled by a point and then derive a pseudo box by estimating the minimum circumscribed rectangle of the region. However, pixels in the predicted mask of each object are susceptible to misclassification, resulting in inaccurate and unstable pseudo-box predictions. Discriminative methods [3, 15, 32] first generate a bag of proposal boxes for each label and then select the optimal one as the final pseudo-box through an MIL-based optimization process, as illustrated in Fig. 1(a). A reasonable sampling strategy aiming to extract the labeled training set from an unlabeled dataset is critical for the successful training of the MIL classifier. Despite their potential, these methods still face two significant challenges, as shown in Fig. 2(a): 1) Static positive sampling in an anchor style leads to training instability, especially in the early stages (see the blue curves in Fig. 7); 2) Inefficient negative sampling that randomly samples background regions introduces numerous irrelevant negative samples, reducing training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address these limitations, we propose PointSR, an effective and robust point-supervised DroneDet framework with self-regularized sampling. As shown in Fig. 1(b), PointSR integrates temporal and informative constraints throughout the pseudo-box generation process to produce high-quality pseudo-box labels. Temporal-Ensembling Encoder (TE Encoder) first builds an anchor prototype library by collecting and encoding temporal information from pseudo-box labels during refinement. Then, the Coarse Pseudo-box Generation adjusts anchors using the anchor prototype library to produce better coarse pseudo boxes while collecting a set of informative samples for the subsequent refinement. Finally, during the Refinement stage, the collected informative samples are used to suppress low-confidence candidate positive samples to refine pseudo boxes. Benefiting from this self-regularized sampling strategy, PointSR not only ensures a more stable and efficient training process but also effectively improves the model's robustness to perturbation brought by manual labeling." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.905, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.258, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 2. (a) Point-to-Box [3] uses a rigid sampling strategy with fixed positive proposals and randomly sampled negative background regions for training the MIL-based predictor. (b) PointSR adopts a self-regularized sampling strategy, incorporating temporal and informative constraints into the pseudo-box generation process for both positive and negative sampling." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.352, + 0.885, + 0.366 + ], + "angle": 0, + "content": "Our key contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.367, + 0.905, + 0.412 + ], + "angle": 0, + "content": "1) We propose PointSR, a novel and robust framework for point-supervised DroneDet, capable of generating high-quality pseudo-box labels from point-level supervision." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.412, + 0.905, + 0.457 + ], + "angle": 0, + "content": "2) We design a Temporal-Ensemble Encoder for stable pseudo-box prediction and an informative negative sampling strategy to enhance pseudo-box quality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.457, + 0.905, + 0.533 + ], + "angle": 0, + "content": "3) PointSR achieves an average \\(\\mathrm{AP}_{50}\\) of \\(42.5\\%\\), \\(38.7\\%\\), and \\(35.0\\%\\) on three drone-captured benchmark datasets, i.e., DroneVehicle [22], UAVDT [5], and VisDrone [34] with only point supervision, outperforming the previous best methods by \\(7.2\\%\\), \\(5.4\\%\\), and \\(2.6\\%\\), respectively." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.367, + 0.905, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.655, + 0.562 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.905, + 0.597 + ], + "angle": 0, + "content": "In this section, we briefly review the current state of research on ISOD, PSOD and PSOD for aerial images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.882, + 0.622 + ], + "angle": 0, + "content": "2.1. Image-Supervised Object Detection (ISOD)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.809 + ], + "angle": 0, + "content": "ISOD refers to an approach in which a detector is trained with image-level annotations rather than box-level ones. The pioneering work WSDDNs [1] integrated the MIL into ISOD, employing a dual-stream structure to classify proposals, and enabling effective weakly-supervised object detection. Inspired by this, Tang et al. [24] designed a Proposal Cluster Learning (PCL) approach, which used proposal clusters as supervision to infer rough object locations. However, image-level annotations lack explicit position information, leading to low-quality proposals, such as background regions, in the potential object tool. This limitation makes it infeasible to directly apply ISOD to DroneDet." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.88, + 0.835 + ], + "angle": 0, + "content": "2.2. Point-Supervised Object Detection (PSOD)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "PSOD is a relatively new field, with only a few methods [2, 3, 17, 21] published in recent years. By leveraging point-level annotations, PSOD enables more accurate object localization and classification compared to image-level" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11708" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.319 + ], + "angle": 0, + "content": "annotations. From the perspective of human cognition, Papadopoulos et al. [17] introduced the first model utilizing center-click annotations, pioneering the use of point supervision in object detection as an alternative to conventional box-level supervision. Ren et al. [21] proposed a unified framework that performs object detection across various annotation types. Chen et al. designed an MIL-based framework Point-to-Box [3] to achieve object detection by classifying proposals. In camouflaged object detection, [2] employed a novel point-based learning paradigm for the challenging task with an attention regulator to distinguish objects that conceal themselves in the environment. However, these PSOD methods only focus on objects in ground-view images and produce poor performance in complex aerial scenarios due to numerous densely packed small objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.326, + 0.315, + 0.342 + ], + "angle": 0, + "content": "2.3. PSOD for Aerial Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.485, + 0.59 + ], + "angle": 0, + "content": "Existing PSOD methods for aerial images can be grouped into two categories: generative and discriminative fashions. The generative approaches, such as PLUG [10], employ a short path algorithm to produce pseudo boxes with pixel-level labels. However, the pseudo boxes rely heavily on the accuracy of pixel-level mask generation. If there is an outlier far from the object mask, the corresponding pseudo label will be seriously affected. In contrast, discriminative methods aim to improve the quality of pseudo labels through advanced sampling strategies. PointOBB [15] produced pseudo labels through the collaborative utilization of three distinctive views to achieve coupled optimization of scale and angle. Point2Rbox [32] combined point annotations and synthetic pattern knowledge with bounding boxes to improve the diversity of samples and provide more supervised information for accurate pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.756 + ], + "angle": 0, + "content": "However, despite these advancements, existing methods still suffer from the limitation of fixed positive candidate proposals and the instability of the random distribution of negative samples. Specifically, the positive sample space remains constant during training and the negative one is only affected slightly by the coarse pseudo boxes in the conventional PSOD framework, formulated in Eq. (5). To address these limitations, we propose an effective and robust PSOD framework to generate high-quality pseudo boxes with a self-regularized sampling strategy, adjusting sampling space adaptively during the model training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.768, + 0.229, + 0.784 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our PointSR is built upon the Weakly Supervised Deep Detection Networks (WSDDNs) [1] for ISOD. The input data consists of a collection of images, each associated with image-level labels \\(\\mathbf{y} \\in \\{0,1\\}^C\\). Here, \\(C\\) represents the number of object categories. Candidate object regions \\(\\mathcal{R} = \\{R_1,R_2,\\dots ,R_{|\\mathcal{R}|}\\}\\) are generated by using region proposal mechanisms [25]. Features of these regions \\(\\mathbf{x} \\in \\mathbb{R}^{D \\times |\\mathcal{R}|}\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.809, + 0.107 + ], + "angle": 0, + "content": "are extracted through fully connected layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.107, + 0.907, + 0.213 + ], + "angle": 0, + "content": "WSDDNs adopts a dual-stream structure, consisting of a classification and detection stream. The extracted features \\(\\mathbf{x}\\) are fed into the two streams, producing two matrices, \\(\\mathbf{x}^c\\) and \\(\\mathbf{x}^d\\in \\mathbb{R}^{C\\times |\\mathcal{R}|}\\), representing the category and position prediction scores for each region, respectively. These scores are then passed through the corresponding softmax operators \\(\\sigma_{\\mathrm{cla}}(\\cdot)\\) and \\(\\sigma_{\\mathrm{det}}(\\cdot)\\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.221, + 0.905, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\left[ \\sigma_ {\\mathrm {c l a}} \\left(\\mathbf {x} ^ {c}\\right) \\right] _ {i j} = \\frac {e ^ {x _ {i j} ^ {c}}}{\\sum_ {k = 1} ^ {C} e ^ {x _ {k j} ^ {c}}}, \\left[ \\sigma_ {\\det } \\left(\\mathbf {x} ^ {d}\\right) \\right] _ {i j} = \\frac {e ^ {x _ {i j} ^ {d}}}{\\sum_ {k = 1} ^ {| \\mathcal {R} |} e ^ {x _ {i k} ^ {d}}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.274, + 0.905, + 0.304 + ], + "angle": 0, + "content": "The regions' scores are then calculated via element-wise multiplication as:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.313, + 0.905, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\mathcal {R}} = \\sigma_ {\\operatorname {c l a}} \\left(\\mathbf {x} ^ {c}\\right) \\cdot \\sigma_ {\\det } \\left(\\mathbf {x} ^ {d}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.34, + 0.905, + 0.371 + ], + "angle": 0, + "content": "Finally, an image-level class prediction score is obtained by summing over all regions as:" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.38, + 0.905, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} _ {c} = \\sum_ {r = 1} ^ {| \\mathcal {R} |} \\mathbf {x} _ {c r} ^ {\\mathcal {R}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.43, + 0.907, + 0.505 + ], + "angle": 0, + "content": "Given a proposal regions' sampling space \\( X \\) and a label space \\( Y \\), features \\( \\mathbf{x} \\) of proposal regions sampled from \\( X \\) in an image are fed into an MIL-based classifier \\( f_{\\theta}(\\cdot) \\), parameterized by \\( \\theta \\), to obtain the image-level class prediction score \\( \\mathbf{y} \\in Y \\). Therefore, ISOD can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.515, + 0.905, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\arg \\min } \\sum_ {k = 1} ^ {C} \\mathcal {L} \\left(f _ {\\theta} (\\mathbf {x}), \\mathbf {y} _ {k}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.565, + 0.907, + 0.611 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(\\cdot)\\) is the model's loss function, and \\(k\\) is the label's \\(k\\)-th category. Note that the sampling space \\(X\\) is constant during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.725, + 0.64 + ], + "angle": 0, + "content": "4. The Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.828, + 0.663 + ], + "angle": 0, + "content": "4.1. Problem Formulation and Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.907, + 0.821 + ], + "angle": 0, + "content": "Problem Formulation: PSOD provides minimal and yet distinctive position information for each object. An input image \\(I\\) contains \\(M\\) targets, each associated with point-level labels \\(\\mathbf{y} \\in \\{0,1\\}^C\\). For the \\(i\\)-th target, its positive samples \\(\\mathbf{x}_i^+\\) with labels \\(\\mathbf{y}_i^+ \\in \\{0,1\\}^C\\) are generated using a static anchor strategy, while its negative samples \\(\\mathbf{x}_i^-\\) with labels \\(\\mathbf{y}_i^-\\in \\{0\\}\\) are collected through random sampling on background. The negative sampling space is defined as the difference between the image \\(I\\) and the predicted pseudo boxes' regions (PB), denoted as \"diff(I, PB)\"." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.821, + 0.863, + 0.836 + ], + "angle": 0, + "content": "The optimization for PSOD can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.845, + 0.918, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta , x ^ {-} \\in \\operatorname {d i f f} (I, \\mathrm {P B})} {\\arg \\min } \\sum_ {i = 1} ^ {M} \\sum_ {k = 1} ^ {C} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i, k} ^ {+}\\right), \\mathbf {y} _ {i, k} ^ {+}\\right) + \\sum_ {j = 1} ^ {N} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {j} ^ {-}\\right), \\mathbf {y} _ {j} ^ {-}\\right). \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11709" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.227, + 0.905, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.342, + 0.908, + 0.414 + ], + "angle": 0, + "content": "Figure 3. The architecture of PointSR, comprising the TE Encoder, Coarse Pseudo-box Prediction and Pseudo-box Refinement. The training process spans \\( T \\) epochs, each consisting of two stages: Coarse Pseudo-box Prediction and Pseudo-box Refinement. The former generates coarse pseudo boxes and collects a set of negative samples by ranking their losses. The latter utilizes these negative samples and the positive samples from the previous stage to refine the pseudo boxes. The TE Encoder collects temporal information during each epoch to dynamically adjust candidate proposal generation at the beginning of the Coarse Pseudo-box Prediction stage." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.42, + 0.482, + 0.451 + ], + "angle": 0, + "content": "where positive samples \\(x^{+}\\) of each target remain constant, and the negative sampling space \\(\\mathrm{diff}(I,\\mathrm{PB})\\) depends on PB." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.451, + 0.483, + 0.527 + ], + "angle": 0, + "content": "To address the limitations of rigid sampling, we propose a self-regularized sampling mechanism, dynamically adjusting the positive \\( X^{+} \\) and negative \\( X^{-} \\) sampling spaces using temporal and informative constraints. According to Eq. (5), PointSR's optimization for PSOD is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.537, + 0.483, + 0.624 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\underset {\\theta , x ^ {+} \\in X ^ {+}, x ^ {-} \\in X ^ {-}} {\\arg \\min } \\sum_ {i = 1} ^ {M} \\sum_ {k = 1} ^ {C} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i, k} ^ {+}\\right), \\mathbf {y} _ {i, k} ^ {+}\\right) + \\tag {6} \\\\ \\sum_ {j = 1} ^ {N} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {j} ^ {-}\\right), \\mathbf {y} _ {j} ^ {-}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Overview: PointSR employs a Self-regularized Sampling mechanism to optimize the pseudo box generation from point-level annotations. Fig. 3 shows that the architecture of PointSR consists of three main components: TE Encoder, Coarse Pseudo-box Prediction, and Pseudo-box Refinement. TE Encoder aggregates temporal information from predicted pseudo labels in each epoch and encodes it into an anchor prototype library. Coarse Pseudo-box Prediction uses the anchor prototype library for anchor adjustment and predicts coarse pseudo boxes using an MIL-based classifier. It also collects informative negative samples for refinement. Finally, Pseudo-box Refinement employs the MIL-based classifier to further refine pseudo boxes by suppressing candidate positive samples with low confidence." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.849, + 0.368, + 0.866 + ], + "angle": 0, + "content": "4.2. Temporal-Ensembling Encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The first step of predicting pseudo boxes is to create candidate boxes. Our approach predicts these candidate boxes in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.421, + 0.868, + 0.435 + ], + "angle": 0, + "content": "an anchor style, done both statically and dynamically." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.905, + 0.542 + ], + "angle": 0, + "content": "In the widely adopted static anchor strategy, anchors' width and height are pre-determined heuristically. Anchor shapes are set and remain constant throughout the entire training process. The static anchor strategy in Faster R-CNN [19] can be sub-optimal as it neglects the underlying distribution of real data, potentially leading to low-quality pseudo boxes and slow convergence during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.906, + 0.8 + ], + "angle": 0, + "content": "To address this limitation, we propose a learning-based dynamic anchor adjustment module, called \"TemporalEnsembling Encoder\" (TE Encoder), as illustrated in Fig. 3(d). TE Encoder focuses on encoding the ratio of objects' width to height instead of objects' scale because objects' aspect ratio is numerically more stable than object scales in drone-view images. Drone-view images show that object scales vary significantly depending on their distance from the drone-mounted camera, while an object's aspect ratio remains relatively stable, particularly in top-down aerial views in Fig 5. At the end of each training iteration, we can obtain high-quality pseudo boxes predicted in the refinement stage, serving as references to adjust anchors' shapes. In conclusion, very different from the pre-defined and fixed anchors, advanced anchors by TE Encoder are dynamically adjusted by the pseudo-box's aspect ratio derived from temporal information during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.801, + 0.905, + 0.862 + ], + "angle": 0, + "content": "Specifically, TE Encoder aggregates pseudo-boxes' shapes for each category in each iteration and stores them as the category-aware prototype library, \\( P \\in \\mathbb{R}^{C \\times 1} \\). For Epoch \\( t \\), the \\( c \\)-th element \\( P_{c} \\in (0,1) \\), \\( c \\in \\{1,2,\\dots,C\\} \\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Setting: anchor shapes are preset as six scales (4, 8, 16, 32, 64, 128) with seven aspect ratios (1: 3, 1: 2, 1: 1.5, 1: 1, 1.5: 1, 2: 1, 3: 1)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11710" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "represents the ratio of width \\( w_{tj} \\in \\mathbb{R} \\) to height \\( h_{tj} \\in \\mathbb{R} \\) of the \\( j \\)-th object of category \\( c \\). \\( P_{c} \\) is updated iteratively as:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.133, + 0.483, + 0.164 + ], + "angle": 0, + "content": "\\[\nP _ {c} ^ {\\prime} = \\lambda P _ {c} + (1 - \\lambda) \\frac {w _ {t j}}{h _ {t j}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.173, + 0.483, + 0.203 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is used to balance the contributions of knowledge distilled in current and previous epochs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.213, + 0.363, + 0.228 + ], + "angle": 0, + "content": "4.3. Coarse Pseudo-Box Prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.234, + 0.484, + 0.28 + ], + "angle": 0, + "content": "We adjust anchors with the anchor prototype library for better coarse pseudo boxes and simultaneously collect a set of informative samples for the pseudo-box refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.484, + 0.446 + ], + "angle": 0, + "content": "MIL-based Pseudo-Box Predictor: Given an adjusted bag of region proposals for each point label, a dual-stream structure is adopted as an MIL classifier to find the best box region in the bag to represent the target. As shown in Fig. 3(b), features of these proposals are extracted through a RoI [9] layer and two fully connected layers. Secondly, these features are fed into the dual-stream structure to estimate scores of each proposal in the same bag according to Eq. (1). Then, a bag's score is calculated by the sum of all its proposals. Finally, the Top-\\(k\\) boxes with the highest scores in the same bag are chosen for coarse pseudo boxes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.658 + ], + "angle": 0, + "content": "Informative-Sample Collector (IS Collector): To capture informative negative samples with high uncertainty, we present IS Collector, which aims to collect negatives with high loss values. Specifically, negative samples randomly sampled on background with their corresponding losses estimated by the MIL-based classifier, are fed into our IS collector for the pseudo-box refinement, as shown in Fig. 3(e). Considering the high computational costs of processing all negative samples, Non-Maximum Suppression [16] is first applied for de-duplication, removing redundant negative samples with high similarity and retaining representative ones. Then, we rank the filtered negative samples in descending order of their losses and thus collect the most informative samples for pseudo-box refinement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.312, + 0.683 + ], + "angle": 0, + "content": "4.4. Pseudo-Box Refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.87 + ], + "angle": 0, + "content": "The collected informative samples are used to suppress candidate positive samples with low confidence so as to refine pseudo boxes. As shown in Fig. 3(c), the model for pseudo-box refinement has a similar structure as the model for coarse pseudo-box prediction. Specifically, for positive sample generation, we update the aspect ratios of coarse pseudo boxes and jitter their center positions. Meanwhile, given the informative negative samples collected in coarse pseudo boxes, after the RoI layer, all features of samples are fed into the MIL classifier according to Eq. (1). Finally, the relatively accurate Top-\\(k\\) candidate proposals are chosen to be merged as the final predicted pseudo box." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Discussion: As shown in Fig. 4(b), numerous proposals by random sampling in the background are far from targets." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.623, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.175, + 0.581, + 0.187 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.091, + 0.796, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.175, + 0.722, + 0.186 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.091, + 0.901, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.841, + 0.175, + 0.857, + 0.186 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.189, + 0.906, + 0.272 + ], + "angle": 0, + "content": "Figure 4. The negative samples distribution of PointSR and Point-to-Box. The histogram (b) shows the number of negative samples for different generation methods, where the distance denotes the Euclidean distance between the center of negative samples (Background) and positive samples (Pedestrians). Two specific examples for PointSR (a) and Point-to-Box (c) are visualized." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.906, + 0.457 + ], + "angle": 0, + "content": "Similarly, Fig. 4(c) shows partial negative samples, indicating that they provide few contributions to object detection. As for densely packed objects in drone-view images, edge information between objects and objects' neighborhood is crucial to distinguish each object. The random negative sampling in the background only generates proposals in a relatively stable sample space. While this sampling ensures diversity among negative samples, it may include less relevant ones, reducing learning efficiency. Additionally, since this method assigns equal weights to all negative samples, it can not adaptively select useful and informative negatives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.906, + 0.673 + ], + "angle": 0, + "content": "Fig. 4(a) presents that negative samples produced by our IS Collector are close enough to the decision boundary in feature space. Training with these negative samples is beneficial to the MIL classifier to recognize which candidate proposals should be assigned low scores and force the model to learn finer distinctions because these negatives contribute more significantly to the gradients. By combining with the IS Collector, the TE Encoder refines positive samples through anchor aspect ratio adjustments, enhancing the MIL classifier's ability to predict losses for negative samples more accurately. This improvement enables the IS Collector to better select high-loss negative samples. As a result, the collaboration between the two modules yields high-quality pseudo-boxes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.906, + 0.799 + ], + "angle": 0, + "content": "The Analysis of Scalability: The time complexity of TE Encoder, \\(\\mathbf{O}(n)\\), scales linearly with the number of the predicted pseudo boxes \\(n\\). Whereas the IS Collector selects a few informative samples from around 400 negative samples randomly sampled from the background. Since this process is independent of \\(n\\) and the image resolution, PointSR's computational overhead is stable, even for high-resolution images or larger datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.667, + 0.838 + ], + "angle": 0, + "content": "4.5. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.851, + 0.905, + 0.881 + ], + "angle": 0, + "content": "Inspired by the discriminative method [3], we design the MIL loss and negative loss for our PointSR training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The MIL Loss: We formulate the MIL loss \\(\\mathcal{L}_{mil}\\) as a cross" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11711" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.38, + 0.108 + ], + "angle": 0, + "content": "entropy loss for positive samples, given by:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.121, + 0.483, + 0.165 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m i l} = - \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {k = 1} ^ {C} \\log \\left(\\hat {\\mathbf {y}} _ {j, k} ^ {\\mathbf {y} _ {j, k}} \\left(1 - \\hat {\\mathbf {y}} _ {j, k}\\right) ^ {\\left(1 - \\mathbf {y} _ {j, k}\\right)}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.175, + 0.483, + 0.235 + ], + "angle": 0, + "content": "where \\(\\mathbf{y}_{j,k} \\in \\{0,1\\}\\) denotes the one-hot category labels and \\(\\hat{\\mathbf{y}}_{jk}\\) is the predict scores generated by MIL-based predictor. The MIL loss is used for the correct prediction of the category and instance to which the proposal belongs to." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.236, + 0.484, + 0.281 + ], + "angle": 0, + "content": "The Negative Loss: Furthermore, the negative loss \\(\\mathcal{L}_{\\text{neg}}\\) in the form of focal loss is used for model optimization, which is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.29, + 0.483, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n e g} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\sum_ {k = 1} ^ {C} \\beta \\cdot (\\hat {\\mathbf {y}} _ {j, k}) ^ {\\gamma} \\log (1 - \\hat {\\mathbf {y}} _ {j, k}) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.404 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is set to 0.2, the same in [13] and \\(\\beta\\) is a weight to balance the contribution of different negative samples. The Negative loss in the Refinement stage is used to further suppress the background for pseudo-box refinement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.417, + 0.224, + 0.433 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.441, + 0.194, + 0.456 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.494 + ], + "angle": 0, + "content": "We conducted experiments on three benchmark datasets: DroneVehicle [22], VisDrone [34] and UAVDT [5]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.584 + ], + "angle": 0, + "content": "**DroneVehicle:** The DroneVehicle dataset includes 56,878 images, which are annotated with box-level annotations across five vehicle categories, i.e., Car, Truck, Bus, Van and Freight Car, covering various scenarios, including urban roads, residential areas, and parking lots captured under different lighting conditions from day to night." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.75 + ], + "angle": 0, + "content": "VisDrone: The VisDrone dataset consists of 10,209 images with a resolution of \\(2,000 \\times 1\\), 500 pixels and 542,000 instances across 10 common object categories in traffic scenarios, i.e., Pedestrian, People, Bicycle, Car, Van, Truck, Tricycle, Awning-tricycle, Bus and Motor. These images, captured from various urban scenes, present significant challenges due to the presence of numerous small objects, varying viewpoints, and heavy occlusions. To the hardware memory limitations, following PLUG [10], we cropped the original images into patches of \\(512 \\times 512\\) pixels with a 128-pixel overlap for training and inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.811 + ], + "angle": 0, + "content": "UAVDT: The UAVDT dataset consists of 100 video sequences of about 80,000 representative frames with a resolution of \\(1,024 \\times 540\\) pixels and 0.84 million box-level annotations across three vehicle categories: Car, Truck, Bus." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To accurately reproduce the biases inherent in manual annotation, we employed Rectified Gaussian Distribution (RG) [30] with central ellipse constraints, assuming that the point distribution of manual annotation approximately follows an RG distribution. Therefore, for our experiments, the three datasets with point-level annotations are employed" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "for evaluating our approach, referred to as \"DroneVehicle (Point-Supervision)\", \"VisDrone (Point-Supervision)\", and \"UAVDT (Point-Supervision)\", respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.148, + 0.859, + 0.164 + ], + "angle": 0, + "content": "5.2. Implementation and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.171, + 0.907, + 0.292 + ], + "angle": 0, + "content": "Implementation Details: We implemented our proposed PointSR based on the MMDetection toolbox on an NVIDIA RTX 4090 GPU. The Stochastic Gradient Descent (SGD) algorithm was used to optimize with a \\(3 \\times\\) training schedule. The learning rate was set to 0.02 and decayed by a factor of 0.1 at the 24th and 33rd epochs, respectively. All comparison models were configured with a ResNet-50 [8] backbone pre-trained on ImageNet [4]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.909, + 0.459 + ], + "angle": 0, + "content": "Evaluation Metrics: We evaluated our method from three aspects: the quality of predicted pseudo boxes, the detection performance and the model's robustness. The mean Intersection over Union (mIoU) between generated pseudo boxes and ground-truth boxes was utilized to evaluate the quality of predicted pseudo boxes, where mIoU closer to 1 indicates higher accuracy. The Average Precision \\((\\mathrm{AP}_{50})\\) [6] was utilized to evaluate the models' detection performance. Finally, the variances of AP across different point-label distributions were used to measure models' robustness, with lower variance indicating higher robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.807, + 0.485 + ], + "angle": 0, + "content": "5.3. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.658 + ], + "angle": 0, + "content": "We compared the performance of a detector trained using pseudo boxes generated by our PointSR and five SOTA point-supervised DroneDet methods: PLUG [10], PointOBB [15], Point2Rbox [32], PointOBBv2 [18] and Point-to-Box [3]. In these experiments, pseudo boxes were first generated by our PointSR and SOTA methods. These pseudo boxes were then utilized to train a classic detector, Faster-RCNN [19], to evaluate its detection performance and robustness. In addition, a Faster-RCNN detector was trained with ground-truth box-level annotations, which detection performance was referred to as the upper bound." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.84 + ], + "angle": 0, + "content": "The Detection Performance: Table 1 shows that the detector trained with pseudo boxes generated by our PointSR achieves superior detection performance, outperforming that trained with pseudo boxes generated by SOTA methods significantly on all datasets. Specifically, our model improved \\(\\mathrm{AP}_{50}\\) by \\(7.2\\%\\) on DroneVehicle (Point-Supervision) dataset, \\(3.3\\%\\) on VisDrone (Point-Supervision) dataset and \\(5.4\\%\\) on UAVDT (Point-Supervision) dataset, compared to Point-to-Box [3]. Besides, it is worth mentioning that the detection results of our model trained with point-level annotations are better than the results of Faster-RCNN trained with box-level annotations in some cases in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The Quality of the Predicted Pseudo-Boxes: We use mIoU between the generated pseudo boxes and ground-truth boxes to evaluate the quality of the predicted pseudo boxes, where mIoU closer to 1 indicates higher accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11712" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Comparisons of per-class detection accuracy \\((\\mathrm{AP}_{50})\\) across different object classes, average \\(\\mathrm{AP}_{50}\\), and mIoU for a detector trained using pseudo boxes generated by our PointSR and SOTA methods on VisDrone, DroneVehicle and UAVDT. \"Faster R-CNN\" indicates the accuracy obtained with the detector trained with ground-truth box-level annotations, serving as the upper bound for performance." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.137, + 0.905, + 0.273 + ], + "angle": 0, + "content": "
MethodVisDroneDroneVehicleUAVDT
Ped.Peo.Bic.CarVanTru.Tri.Awn.BusMot.AP50mIoUCarTru.BusVanFre.AP50mIoUCarTru.BusAP50mIoU
PLUG [10]0.63.44.73.40.24.50.70.323.65.34.726.91.06.10.28.02.33.529.04.58.23.85.531.1
Point2Rbox [32]0.30.29.122.315.315.311.59.915.81.410.1-0.03.675.718.928.025.2-15.724.119.219.7-
PointOBB [15]25.613.620.834.612.430.732.930.456.512.427.056.218.37.06.034.510.115.277.529.230.816.825.672.6
PointOBBv2 [18]33.24.316.169.257.431.632.123.232.823.832.464.817.119.734.229.222.824.680.116.628.837.227.577.5
Point-to-Box [3]40.24.423.162.953.130.930.922.828.419.631.766.268.925.524.233.024.735.386.045.535.818.533.395.3
Faster R-CNN [19]53.730.930.072.257.644.231.325.035.837.241.8-89.766.092.253.352.770.1-66.955.759.660.7-
PointSR (Ours)43.45.826.265.653.531.529.722.444.727.735.075.782.825.133.542.229.142.588.950.140.825.438.796.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.279, + 0.907, + 0.322 + ], + "angle": 0, + "content": "Table 2. Comparison of average precision \\((\\mathrm{AP}_{50})\\) and variance for a detector trained using pseudo boxes generated by our PointSR and SOTA methods across different point-label distributions. PointSR consistently demonstrates lower variance on both datasets, exhibiting greater stability and robustness to annotation perturbations." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.325, + 0.895, + 0.428 + ], + "angle": 0, + "content": "
MethodPublicationVisDroneDroneVehicle
U0×LU0.1×LU0.3×LU0.5×LVariance Of AP50U0×LU0.1×LU0.3×LU0.5×LVariance Of AP50
PLUG [10]TGRS'241.54.315.517.848.911.13.930.132.5209.1
Point2Rbox [32]CVPR'247.69.98.39.10.74215.224.117.618.310.61
PointOBB [15]CVPR'2424.328.517.88.556.9113.415.212.68.95.269
Point-to-Box [3]ECCV'2231.632.531.130.10.75234.835.532.934.00.935
PointSR (Ours)N/A34.634.833.933.70.21241.142.041.340.90.171
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.437, + 0.483, + 0.465 + ], + "angle": 0, + "content": "Table 3. The impact of the TE Encoder and Informative Sampling (IS) on the baseline Point-to-Box [3], evaluated on DroneVehicle." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.469, + 0.451, + 0.55 + ], + "angle": 0, + "content": "
TEEISAP50mIoU [%]
Baseline××35.2686.0
Baseline + TEE✓×36.1886.7
Baseline + IS×✓40.8788.1
Baseline + TEE + IS✓✓42.5388.9
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.626 + ], + "angle": 0, + "content": "Tables 1 also show that PointSR achieves better mIoUs than the SOTA methods on both datasets. Specifically, our model improves mIoU by \\(9.5\\%\\) on Visdrone, \\(2.9\\%\\) on DroneVehicle and \\(1.5\\%\\) on UAVDT compared to Point-to-Box [3]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.628, + 0.483, + 0.779 + ], + "angle": 0, + "content": "The visualization in Fig. 6 shows that our predicted pseudo boxes accurately localize objects in challenging drone-view images by leveraging the precise location information of point annotations and can distinguish dense objects in complex scenes. In contrast, Point-to-Box often produces loose pseudo boxes due to its limited sampling strategy for MIL training, and PLUG occasionally predicts oversized boxes due to outliers in the generative process. These results confirm the high quality of pseudo boxes generated by our model with point-level annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The Robustness of the Model: To evaluate the model's robustness, we use the variance of \\(\\mathrm{AP}_{50}\\) across different point distributions. Specially, we assume manually annotated points follow a uniform distribution \\(x\\sim U(-a,a)\\), \\(a = \\alpha \\times L\\), \\(\\alpha = \\{0,0.1,0.3,0.5\\}\\) and \\(L\\) is the maximum value for the width and height of a bounding box. \\(U_{0\\times L}\\), \\(U_{0.1\\times L}\\), \\(U_{0.3\\times L}\\), and \\(U_{0.5\\times L}\\) denote the manually annotated points \\(x\\sim U(-\\alpha \\times L,\\alpha \\times L)\\) with offsets of 0, \\(10\\%\\)," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.501 + ], + "angle": 0, + "content": "\\(30\\%\\), \\(50\\%\\), respectively. As shown in Table 2, our model has achieved a variance of \\(\\mathrm{AP}_{50}\\) 0.171 on the DroneVehicle dataset, compared to 0.935 of Point-to-Box [3], indicating superior robustness to annotation perturbation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.512, + 0.68, + 0.527 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.595 + ], + "angle": 0, + "content": "We conducted ablation studies on the Drone Vehicle (Point-Supervision) dataset using Point-to-Box [3] as the baseline to validate the effectiveness of the two key components in PointSR: the TE Encoder and IS Collector." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.853 + ], + "angle": 0, + "content": "The Effectiveness of the TE Encoder: We conducted a quantitative evaluation to verify the effectiveness of the TE Encoder in our PointSR. The baseline with TE Encoder is referred to as \"Baseline + TEE\". Table 3 shows that our proposed module improves \\(\\mathrm{AP}_{50}\\) by \\(0.92\\%\\) and mIoU by \\(0.7\\%\\), compared with the vanilla baseline model. Moreover, TE Encoder facilitates stable and efficient training, in Fig. 7(a), where the orange curve (Baseline + TEE) descends faster and is smoother compared with the blue curve, especially in the early stage of training, indicating that our Anchor Learning model achieves stable training and faster convergence. In addition, Fig. 7(b) further demonstrates that \"Baseline + TEE\" achieves better detection performance. Finally, Table 4 shows the performance of Faster R-CNN trained using PointSR with different \\(\\lambda\\) settings, where PointSR with \\(\\lambda = 0.5\\) achieves the best detection performance. These results verify the effectiveness of the TE Encoder." + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.859, + 0.898, + 0.873 + ], + "angle": 0, + "content": "Table 4. Impact of different \\( \\lambda \\) values on \\( {\\mathrm{{AP}}}_{50} \\) in DroneVehicle." + }, + { + "type": "table", + "bbox": [ + 0.588, + 0.875, + 0.832, + 0.912 + ], + "angle": 0, + "content": "
λ0.10.30.50.70.9
AP50 [%]34.240.542.534.933.9
" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11713" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.089, + 0.223, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.089, + 0.335, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.089, + 0.444, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.089, + 0.555, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.089, + 0.665, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.089, + 0.775, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.089, + 0.885, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.175, + 0.223, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.175, + 0.334, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.175, + 0.444, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.175, + 0.554, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.175, + 0.665, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.175, + 0.775, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.175, + 0.885, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.26, + 0.223, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.348, + 0.188, + 0.361 + ], + "angle": 0, + "content": "PLUG" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.26, + 0.334, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.348, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Point2Rbox" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.26, + 0.444, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.349, + 0.42, + 0.361 + ], + "angle": 0, + "content": "PointOBB" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.26, + 0.554, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.349, + 0.537, + 0.361 + ], + "angle": 0, + "content": "Point-to-Box" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.26, + 0.665, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.349, + 0.648, + 0.361 + ], + "angle": 0, + "content": "Faster-RCNN" + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.26, + 0.775, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.349, + 0.761, + 0.362 + ], + "angle": 0, + "content": "PointSR(Ours)" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.26, + 0.885, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.348, + 0.838, + 0.361 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.364, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Figure 5. Visualization of detection results from a detector trained with pseudo boxes generated by PointSR and SOTA methods. Ground-truth bounding boxes are shown in green, correct detections in yellow, and misclassified or missed targets in red. For a better view, detection results in the last row are highlighted by cyan rectangular boxes. Best viewed in color and zoomed-in mode." + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.413, + 0.468, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.682 + ], + "angle": 0, + "content": "Figure 6. Visualization of pseudo boxes (pink) predicted by PointSR and SOTA methods. Ground-truth boxes are shown in green. Best viewed in color and zoomed-in mode." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.689, + 0.287, + 0.78 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.782, + 0.211, + 0.793 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.688, + 0.481, + 0.781 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.782, + 0.407, + 0.793 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Figure 7. The loss and \\(\\mathrm{AP}_{50}\\) curves during the whole training stage. The blue curves represent baseline training, while the orange curves show baseline training equipped with our TE Encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The Effectiveness of the IS Collector: We evaluated the effectiveness of the Informative Sampling Collector in our PointSR. Table 3 shows that the baseline with our Informative Sampling module, referred to as \"Baseline + IS\"," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.415, + 0.712, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.507, + 0.636, + 0.518 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.415, + 0.906, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.508, + 0.835, + 0.519 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.658 + ], + "angle": 0, + "content": "Figure 8. The \\(\\mathrm{AP}_{50}\\) and IoU curves for varying Sampling Ratio \\(\\delta\\). improves \\(\\mathrm{AP}_{50}\\) by \\(5.26\\%\\) and mIoU by \\(2.2\\%\\). These experimental results confirm the effectiveness of our IS Collector. In addition, we also analyzed the impact of the ratio of selected negative samples to whole negative samples \\(\\delta\\), denoted as \"Sampling Ratio\", on detection performance. Fig. 8 visualizes that changes in \\(\\delta\\) significantly affect the detection performance of our PiontSR in terms of \\(\\mathrm{AP}_{50}\\) and mIoU, suggesting its critical role in optimizing results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Towards generating high-quality pseudo boxes for PSOD from drone-view images, in this paper, we proposed an effective and robust self-regularized point-supervision approach, PointSR, which leverages a Temporal-Ensemble Encoder for stable pseudo-box generation and an Informative Sampling Collector for efficient pseudo-box refinement. The two processes mutually benefit each other, facilitating the generation of self-regularized high-quality pseudo boxes. Experimental results on benchmark datasets have demonstrated that detectors trained with pseudo boxes generated by PointSR significantly outperform those generated by SOTAs over \\(2.6\\% \\sim 7.2\\%\\) in DroneDet and gain the robust results under perturbations of the manually annotated labels, significantly advancing the applicability of PSOD." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11714" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.204 + ], + "angle": 0, + "content": "Acknowledgments: This work was supported by the National Natural Science Foundation of China (Grant No. 62402370, 62301397), the Natural Science Basic Research Program of Shaanxi (Grant No. 2024JC-YBQN-0728), the Guangzhou Basic and Applied Basic Research Foundation (Grant No. 2023A04J1742). Additionally, we would like to thank Zhiwen Wei for his helpful discussions and insightful suggestions." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.234, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[1] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2854, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.317, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[2] Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. Just a Hint: Point-Supervised Camouflaged Object Detection. In European Conference on Computer Vision, pages 332-348. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[3] Pengfei Chen, Xuehui Yu, Xumeng Han, Najmul Hassan, Kai Wang, Jiachen Li, Jian Zhao, Humphrey Shi, Zhenjun Han, and Qixiang Ye. Point-to-Box Network for Accurate Object Detection via Single Point Supervision. In European Conference on Computer Vision, pages 51-67. Springer, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-scale Hierarchical Image Database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[5] Dawei Du, Yuankai Qi, Hongyang Yu, Yifan Yang, Kaiwen Duan, Guorong Li, Weigang Zhang, Qingming Huang, and Qi Tian. The Unmanned Aerial Vehicle Benchmark: Object Detection and Tracking. In Proceedings of the European Conference on Computer Vision, pages 370-386, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[6] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes Challenge: A Retrospective. International Journal of Computer Vision, 111:98-136, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[7] Xiaoxu Feng, Xiwen Yao, Gong Cheng, and Junwei Han. Weakly Supervised Rotation-Invariant Aerial Object Detection Network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14146-14155, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.745, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In Proceedings of the IEEE International Conference on Computer Vision, pages 2961-2969, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Shitian He, Huanxin Zou, Yingqian Wang, Boyang Li, Xu Cao, and Ning Jing. Learning Remote Sensing Object Detection With Single Point Supervision. IEEE Transactions" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.259, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "on Geoscience and Remote Sensing, 62:1-16, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[11] Boyang Li, Yingqian Wang, Longguang Wang, Fei Zhang, Ting Liu, Zaiping Lin, Wei An, and Yulan Guo. Monte Carlo Linear Clustering with Single-Point Supervision is Enough for Infrared Small Target Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1009–1019, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[12] Jing Li, Junsong Fan, and Zhaoxiang Zhang. Point-Supervised Panoptic Segmentation via Estimating Pseudo Labels from Learnable Distance. In European Conference on Computer Vision, pages 95-112. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[13] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal Loss for Dense Object Detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[14] Junyan Lu, Qinglei Hu, Ruifei Zhu, Yali Wei, and Tie Li. AFWS: Angle-Free Weakly-Supervised Rotating Object Detection for Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.907, + 0.44 + ], + "angle": 0, + "content": "[15] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. PointOBB: Learning Oriented Object Detection via Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16730-16740, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.443, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[16] Alexander Neubeck and Luc Van Gool. Efficient Nonmaximum Suppression. In 18th International Conference on Pattern Recognition, pages 850-855. IEEE, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.485, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[17] Dim P Papadopoulos, Jasper RR Uijlings, Frank Keller, and Vittorio Ferrari. Training Object Class Detectors with Click Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6374-6383, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[18] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection. In The Thirteenth International Conference on Learning Representations, 2025. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[19] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2016. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[20] Tianqi Ren, Qiu Shen, Ying Fu, and Shaodi You. Point-Supervised Semantic Segmentation of Natural Scenes via Hyperspectral Imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1357-1367, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[21] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, Ming-Yu Liu, Alexander G Schwing, and Jan Kautz. \\(\\mathbf{UFO}^2\\): A Unified Framework Towards Omni-Supervised Object Detection. In European Conference on Computer Vision, pages 288-313. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[22] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Drone-Based RGB-Infrared Cross-Modality Vehicle Detection via Uncertainty-Aware Learning. IEEE Transactions on Circuits and Systems for Video Technology, 32(10):6700-6713, 2022. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11715" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[23] Zhiwen Tan, Zhiguo Jiang, Chen Guo, and Haopeng Zhang. WSODet: A Weakly Supervised Oriented Detector for Aerial Object Detection. IEEE Transactions on Geoscience and Remote Sensing, 61:1-12, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[24] Peng Tang, Xinggang Wang, Song Bai, Wei Shen, Xiang Bai, Wenyu Liu, and Alan Yuille. PCL: Proposal Cluster Learning for Weakly Supervised Object Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(1): 176-191, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[25] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective Search for Object Recognition. International Journal of Computer Vision, 104: 154-171, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[26] Yuting Wang, Velibor Ilic, Jiatong Li, Branislav Kisačanin, and Vladimir Pavlovic. ALWOD: Active Learning for Weakly-Supervised Object Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6459–6469, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[27] Yue Xi, Wenjing Jia, Qiguang Miao, Junmei Feng, Jinchang Ren, and Heng Luo. Detection-Driven Exposure-Correction Network for Nighttime Drone-View Object Detection. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[28] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Query-Det: Cascaded Sparse Query for Accelerating High-Resolution Small Object Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13668–13677, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[29] Xinyi Ying, Li Liu, Yingqian Wang, Ruojing Li, Nuo Chen, Zaiping Lin, Weidong Sheng, and Shilin Zhou. Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15528-15538, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[30] Xuehui Yu, Pengfei Chen, Di Wu, Najmul Hassan, Guorong Li, Junchi Yan, Humphrey Shi, Qixiang Ye, and Zhenjun Han. Object Localization Under Single Coarse Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4868-4877, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[31] Xuehui Yu, Pengfei Chen, Kuiran Wang, Xumeng Han, Guorong Li, Zhenjun Han, Qixiang Ye, and Jianbin Jiao. CPR++: Object Localization via Single Coarse Point Supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[32] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16783-16793, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[33] Xiang Yuan, Gong Cheng, Kebing Yan, Qinghua Zeng, and Junwei Han. Small Object Detection via Coarse-to-Fine Proposal Generation and Imitation Learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6317-6327, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[34] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and Tracking" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.092, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Meet Drones Challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 2, 6" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11716" + } + ] +] \ No newline at end of file diff --git a/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/f94115e7-2e5c-44d9-81fe-c067d1bde80f_origin.pdf b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/f94115e7-2e5c-44d9-81fe-c067d1bde80f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..405a0454bb47ce73533155ad5d87a8fd8f5a20e5 --- /dev/null +++ b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/f94115e7-2e5c-44d9-81fe-c067d1bde80f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c0d17d1c514d26a605024b7a5d471071781ef5ca6cae537924ed13b5c898419 +size 9170057 diff --git a/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/full.md b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9c12de9711f45fb5e4aa08f9cd2da05c048d16c5 --- /dev/null +++ b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/full.md @@ -0,0 +1,373 @@ +# PointSR: Self-Regularized Point Supervision for Drone-View Object Detection + +Weizhuo Li $^{1}$ Yue Xi $^{1*}$ Wenjing Jia $^{2}$ Zehao Zhang $^{1}$ Fei Li $^{3}$ Xiangzeng Liu $^{3}$ Qiguang Miao $^{3*}$ + +1Guangzhou Institute of Technology, Xidian University, Guangzhou, China + +$^{2}$ Faculty of Engineering and IT, University of Technology Sydney, Sydney, Australia + +$^{3}$ School of Computer Science and Technology, Xidian University, Xi'an, China + +{liweizhuo@stu., xiyue@, zzehao@stu., lifei@, xzliu@, qgmiao@}xidian.edu.cn + +Wenjing.Jia@uts.edu.au + +# Abstract + +Point-Supervised Object Detection (PSOD) in a discriminative style has recently gained significant attention for its impressive detection performance and cost-effectiveness. However, accurately predicting high-quality pseudo-box labels for drone-view images, which often feature densely packed small objects, remains a challenge. This difficulty arises primarily from the limitation of rigid sampling strategies, which hinder the pseudo-box optimization process. To address this, we propose PointSR, an effective and robust point-supervised object detection framework with self-regularized sampling that integrates temporal and informative constraints throughout the pseudo-box generation process. Specifically, the framework comprises three key components: Temporal-Ensembling Encoder (TE Encoder), Coarse Pseudo-box Prediction, and Pseudo-box Refinement. The TE Encoder builds an anchor prototype library by aggregating temporal information for dynamic anchor adjustment. In Coarse Pseudo-box Prediction, anchors are refined using the prototype library, and a set of informative samples is collected for subsequent refinement. During Pseudo-box Refinement, these informative negative samples are used to suppress low-confidence candidate positive samples, thereby improving the quality of the pseudoboxes. Experimental results on benchmark datasets demonstrate that PointSR significantly outperforms state-of-the-art methods, achieving up to $2.6\% \sim 7.2\%$ higher $AP_{50}$ using only point supervision. Additionally, it exhibits strong robustness to perturbation in human-labeled points. + +# 1. Introduction + +Drone-view object detection (DroneDet) [7, 23, 27, 28, 33], which involves detecting objects from drone-captured images, plays a vital role in various domains, including national defense, resource management, and drone inspection. + +![](images/a47cef66bbead7e6e2bf9a2f294ce9d285d104eb044972bf0e5f2fdc2fa52920.jpg) +Figure 1. Comparisons of the pseudo-box optimization process. (a) The rigid sampling strategy in Point-to-Box [3] results in the inefficiency of the training process for low-quality pseudo-box labels. (b) The self-regularized sampling in our PointSR improves the stability and efficiency of the training process and thus obtains high-quality pseudo-box labels. + +However, DroneDet models often experience a significant performance drop when applied to new scenes or datasets due to their limited generalization capabilities. A straightforward solution is to collect and annotate new datasets and retrain the models. However, annotating drone-view images with box-level annotations containing accurate position information is particularly challenging due to the prevalence of many densely packed small targets, far more complex than annotating ground-view images. Moreover, the diversity of the scenes and object types can lead to combinatorial explosions of labeling requirements, making the process infeasible in many cases. Consequently, constructing new datasets with box-level annotations is both time-consuming and costly when deploying drones in new scenarios. + +To alleviate the cost associated with box-level annotations, Image-Supervised Object Detection (ISOD) [1, 14, 26] has been extensively studied. ISOD treats images as a bag of regions, assuming at least one region contains the object and uses a Multiple Instance Learning (MIL) classifier to identify regions corresponding to image-level labels. De + +spite its labeling efficiency, ISOD methods often perform poorly in complex scenarios such as drone-view images, with targets typically small and densely packed, due to the lack of essential position information in image-level labels. + +To overcome this limitation, Point-Supervised Object Detection (PSOD) has emerged as an alternative solution. PSOD provides minimal and yet additional distinctive position information for each object and has gained significant attention in tasks such as detection [10, 32], localization [30, 31], and segmentation [12, 20]. According to [17], the labor cost of point annotations per object is only about $1/18$ of that for bounding boxes, making PSOD a highly cost-effective solution for DroneDet applications. + +Existing PSOD methods [3, 10, 11, 15, 29, 32] can be broadly categorized into generative and discriminative approaches. Generative methods [10, 11, 29] predict a masked region covering a target labeled by a point and then derive a pseudo box by estimating the minimum circumscribed rectangle of the region. However, pixels in the predicted mask of each object are susceptible to misclassification, resulting in inaccurate and unstable pseudo-box predictions. Discriminative methods [3, 15, 32] first generate a bag of proposal boxes for each label and then select the optimal one as the final pseudo-box through an MIL-based optimization process, as illustrated in Fig. 1(a). A reasonable sampling strategy aiming to extract the labeled training set from an unlabeled dataset is critical for the successful training of the MIL classifier. Despite their potential, these methods still face two significant challenges, as shown in Fig. 2(a): 1) Static positive sampling in an anchor style leads to training instability, especially in the early stages (see the blue curves in Fig. 7); 2) Inefficient negative sampling that randomly samples background regions introduces numerous irrelevant negative samples, reducing training efficiency. + +To address these limitations, we propose PointSR, an effective and robust point-supervised DroneDet framework with self-regularized sampling. As shown in Fig. 1(b), PointSR integrates temporal and informative constraints throughout the pseudo-box generation process to produce high-quality pseudo-box labels. Temporal-Ensembling Encoder (TE Encoder) first builds an anchor prototype library by collecting and encoding temporal information from pseudo-box labels during refinement. Then, the Coarse Pseudo-box Generation adjusts anchors using the anchor prototype library to produce better coarse pseudo boxes while collecting a set of informative samples for the subsequent refinement. Finally, during the Refinement stage, the collected informative samples are used to suppress low-confidence candidate positive samples to refine pseudo boxes. Benefiting from this self-regularized sampling strategy, PointSR not only ensures a more stable and efficient training process but also effectively improves the model's robustness to perturbation brought by manual labeling. + +![](images/9c7531681c4b12d3ae9dce9890344cc34d695c37ee8232a48466cf33bb2dafb8.jpg) +Figure 2. (a) Point-to-Box [3] uses a rigid sampling strategy with fixed positive proposals and randomly sampled negative background regions for training the MIL-based predictor. (b) PointSR adopts a self-regularized sampling strategy, incorporating temporal and informative constraints into the pseudo-box generation process for both positive and negative sampling. + +Our key contributions can be summarized as follows: + +1) We propose PointSR, a novel and robust framework for point-supervised DroneDet, capable of generating high-quality pseudo-box labels from point-level supervision. +2) We design a Temporal-Ensemble Encoder for stable pseudo-box prediction and an informative negative sampling strategy to enhance pseudo-box quality. +3) PointSR achieves an average $\mathrm{AP}_{50}$ of $42.5\%$ , $38.7\%$ , and $35.0\%$ on three drone-captured benchmark datasets, i.e., DroneVehicle [22], UAVDT [5], and VisDrone [34] with only point supervision, outperforming the previous best methods by $7.2\%$ , $5.4\%$ , and $2.6\%$ , respectively. + +# 2. Related Work + +In this section, we briefly review the current state of research on ISOD, PSOD and PSOD for aerial images. + +# 2.1. Image-Supervised Object Detection (ISOD) + +ISOD refers to an approach in which a detector is trained with image-level annotations rather than box-level ones. The pioneering work WSDDNs [1] integrated the MIL into ISOD, employing a dual-stream structure to classify proposals, and enabling effective weakly-supervised object detection. Inspired by this, Tang et al. [24] designed a Proposal Cluster Learning (PCL) approach, which used proposal clusters as supervision to infer rough object locations. However, image-level annotations lack explicit position information, leading to low-quality proposals, such as background regions, in the potential object tool. This limitation makes it infeasible to directly apply ISOD to DroneDet. + +# 2.2. Point-Supervised Object Detection (PSOD) + +PSOD is a relatively new field, with only a few methods [2, 3, 17, 21] published in recent years. By leveraging point-level annotations, PSOD enables more accurate object localization and classification compared to image-level + +annotations. From the perspective of human cognition, Papadopoulos et al. [17] introduced the first model utilizing center-click annotations, pioneering the use of point supervision in object detection as an alternative to conventional box-level supervision. Ren et al. [21] proposed a unified framework that performs object detection across various annotation types. Chen et al. designed an MIL-based framework Point-to-Box [3] to achieve object detection by classifying proposals. In camouflaged object detection, [2] employed a novel point-based learning paradigm for the challenging task with an attention regulator to distinguish objects that conceal themselves in the environment. However, these PSOD methods only focus on objects in ground-view images and produce poor performance in complex aerial scenarios due to numerous densely packed small objects. + +# 2.3. PSOD for Aerial Images + +Existing PSOD methods for aerial images can be grouped into two categories: generative and discriminative fashions. The generative approaches, such as PLUG [10], employ a short path algorithm to produce pseudo boxes with pixel-level labels. However, the pseudo boxes rely heavily on the accuracy of pixel-level mask generation. If there is an outlier far from the object mask, the corresponding pseudo label will be seriously affected. In contrast, discriminative methods aim to improve the quality of pseudo labels through advanced sampling strategies. PointOBB [15] produced pseudo labels through the collaborative utilization of three distinctive views to achieve coupled optimization of scale and angle. Point2Rbox [32] combined point annotations and synthetic pattern knowledge with bounding boxes to improve the diversity of samples and provide more supervised information for accurate pseudo labels. + +However, despite these advancements, existing methods still suffer from the limitation of fixed positive candidate proposals and the instability of the random distribution of negative samples. Specifically, the positive sample space remains constant during training and the negative one is only affected slightly by the coarse pseudo boxes in the conventional PSOD framework, formulated in Eq. (5). To address these limitations, we propose an effective and robust PSOD framework to generate high-quality pseudo boxes with a self-regularized sampling strategy, adjusting sampling space adaptively during the model training. + +# 3. Preliminaries + +Our PointSR is built upon the Weakly Supervised Deep Detection Networks (WSDDNs) [1] for ISOD. The input data consists of a collection of images, each associated with image-level labels $\mathbf{y} \in \{0,1\}^C$ . Here, $C$ represents the number of object categories. Candidate object regions $\mathcal{R} = \{R_1,R_2,\dots ,R_{|\mathcal{R}|}\}$ are generated by using region proposal mechanisms [25]. Features of these regions $\mathbf{x} \in \mathbb{R}^{D \times |\mathcal{R}|}$ + +are extracted through fully connected layers. + +WSDDNs adopts a dual-stream structure, consisting of a classification and detection stream. The extracted features $\mathbf{x}$ are fed into the two streams, producing two matrices, $\mathbf{x}^c$ and $\mathbf{x}^d\in \mathbb{R}^{C\times |\mathcal{R}|}$ , representing the category and position prediction scores for each region, respectively. These scores are then passed through the corresponding softmax operators $\sigma_{\mathrm{cla}}(\cdot)$ and $\sigma_{\mathrm{det}}(\cdot)$ , defined as: + +$$ +\left[ \sigma_ {\mathrm {c l a}} \left(\mathbf {x} ^ {c}\right) \right] _ {i j} = \frac {e ^ {x _ {i j} ^ {c}}}{\sum_ {k = 1} ^ {C} e ^ {x _ {k j} ^ {c}}}, \left[ \sigma_ {\det } \left(\mathbf {x} ^ {d}\right) \right] _ {i j} = \frac {e ^ {x _ {i j} ^ {d}}}{\sum_ {k = 1} ^ {| \mathcal {R} |} e ^ {x _ {i k} ^ {d}}}. \tag {1} +$$ + +The regions' scores are then calculated via element-wise multiplication as: + +$$ +\mathbf {x} ^ {\mathcal {R}} = \sigma_ {\operatorname {c l a}} \left(\mathbf {x} ^ {c}\right) \cdot \sigma_ {\det } \left(\mathbf {x} ^ {d}\right). \tag {2} +$$ + +Finally, an image-level class prediction score is obtained by summing over all regions as: + +$$ +\hat {\mathbf {y}} _ {c} = \sum_ {r = 1} ^ {| \mathcal {R} |} \mathbf {x} _ {c r} ^ {\mathcal {R}}. \tag {3} +$$ + +Given a proposal regions' sampling space $X$ and a label space $Y$ , features $\mathbf{x}$ of proposal regions sampled from $X$ in an image are fed into an MIL-based classifier $f_{\theta}(\cdot)$ , parameterized by $\theta$ , to obtain the image-level class prediction score $\mathbf{y} \in Y$ . Therefore, ISOD can be formulated as: + +$$ +\underset {\theta} {\arg \min } \sum_ {k = 1} ^ {C} \mathcal {L} \left(f _ {\theta} (\mathbf {x}), \mathbf {y} _ {k}\right), \tag {4} +$$ + +where $\mathcal{L}(\cdot)$ is the model's loss function, and $k$ is the label's $k$ -th category. Note that the sampling space $X$ is constant during training. + +# 4. The Proposed Method + +# 4.1. Problem Formulation and Overview + +Problem Formulation: PSOD provides minimal and yet distinctive position information for each object. An input image $I$ contains $M$ targets, each associated with point-level labels $\mathbf{y} \in \{0,1\}^C$ . For the $i$ -th target, its positive samples $\mathbf{x}_i^+$ with labels $\mathbf{y}_i^+ \in \{0,1\}^C$ are generated using a static anchor strategy, while its negative samples $\mathbf{x}_i^-$ with labels $\mathbf{y}_i^-\in \{0\}$ are collected through random sampling on background. The negative sampling space is defined as the difference between the image $I$ and the predicted pseudo boxes' regions (PB), denoted as "diff(I, PB)". + +The optimization for PSOD can be formulated as: + +$$ +\underset {\theta , x ^ {-} \in \operatorname {d i f f} (I, \mathrm {P B})} {\arg \min } \sum_ {i = 1} ^ {M} \sum_ {k = 1} ^ {C} \mathcal {L} \left(f _ {\theta} \left(\mathbf {x} _ {i, k} ^ {+}\right), \mathbf {y} _ {i, k} ^ {+}\right) + \sum_ {j = 1} ^ {N} \mathcal {L} \left(f _ {\theta} \left(\mathbf {x} _ {j} ^ {-}\right), \mathbf {y} _ {j} ^ {-}\right). \tag {5} +$$ + +![](images/2e42270782033b04d9372b3b54c104559174b5be2fd72b8bc2801a390f720b7f.jpg) + +![](images/b1066da5f5cbd22c7ee236f5f456c45c1ed13591ec598f67bb2951de87e2a291.jpg) +Figure 3. The architecture of PointSR, comprising the TE Encoder, Coarse Pseudo-box Prediction and Pseudo-box Refinement. The training process spans $T$ epochs, each consisting of two stages: Coarse Pseudo-box Prediction and Pseudo-box Refinement. The former generates coarse pseudo boxes and collects a set of negative samples by ranking their losses. The latter utilizes these negative samples and the positive samples from the previous stage to refine the pseudo boxes. The TE Encoder collects temporal information during each epoch to dynamically adjust candidate proposal generation at the beginning of the Coarse Pseudo-box Prediction stage. + +where positive samples $x^{+}$ of each target remain constant, and the negative sampling space $\mathrm{diff}(I,\mathrm{PB})$ depends on PB. + +To address the limitations of rigid sampling, we propose a self-regularized sampling mechanism, dynamically adjusting the positive $X^{+}$ and negative $X^{-}$ sampling spaces using temporal and informative constraints. According to Eq. (5), PointSR's optimization for PSOD is formulated as: + +$$ +\begin{array}{l} \underset {\theta , x ^ {+} \in X ^ {+}, x ^ {-} \in X ^ {-}} {\arg \min } \sum_ {i = 1} ^ {M} \sum_ {k = 1} ^ {C} \mathcal {L} \left(f _ {\theta} \left(\mathbf {x} _ {i, k} ^ {+}\right), \mathbf {y} _ {i, k} ^ {+}\right) + \tag {6} \\ \sum_ {j = 1} ^ {N} \mathcal {L} \left(f _ {\theta} \left(\mathbf {x} _ {j} ^ {-}\right), \mathbf {y} _ {j} ^ {-}\right). \\ \end{array} +$$ + +Overview: PointSR employs a Self-regularized Sampling mechanism to optimize the pseudo box generation from point-level annotations. Fig. 3 shows that the architecture of PointSR consists of three main components: TE Encoder, Coarse Pseudo-box Prediction, and Pseudo-box Refinement. TE Encoder aggregates temporal information from predicted pseudo labels in each epoch and encodes it into an anchor prototype library. Coarse Pseudo-box Prediction uses the anchor prototype library for anchor adjustment and predicts coarse pseudo boxes using an MIL-based classifier. It also collects informative negative samples for refinement. Finally, Pseudo-box Refinement employs the MIL-based classifier to further refine pseudo boxes by suppressing candidate positive samples with low confidence. + +# 4.2. Temporal-Ensembling Encoder + +The first step of predicting pseudo boxes is to create candidate boxes. Our approach predicts these candidate boxes in + +an anchor style, done both statically and dynamically. + +In the widely adopted static anchor strategy, anchors' width and height are pre-determined heuristically. Anchor shapes are set and remain constant throughout the entire training process. The static anchor strategy in Faster R-CNN [19] can be sub-optimal as it neglects the underlying distribution of real data, potentially leading to low-quality pseudo boxes and slow convergence during training. + +To address this limitation, we propose a learning-based dynamic anchor adjustment module, called "TemporalEnsembling Encoder" (TE Encoder), as illustrated in Fig. 3(d). TE Encoder focuses on encoding the ratio of objects' width to height instead of objects' scale because objects' aspect ratio is numerically more stable than object scales in drone-view images. Drone-view images show that object scales vary significantly depending on their distance from the drone-mounted camera, while an object's aspect ratio remains relatively stable, particularly in top-down aerial views in Fig 5. At the end of each training iteration, we can obtain high-quality pseudo boxes predicted in the refinement stage, serving as references to adjust anchors' shapes. In conclusion, very different from the pre-defined and fixed anchors, advanced anchors by TE Encoder are dynamically adjusted by the pseudo-box's aspect ratio derived from temporal information during training. + +Specifically, TE Encoder aggregates pseudo-boxes' shapes for each category in each iteration and stores them as the category-aware prototype library, $P \in \mathbb{R}^{C \times 1}$ . For Epoch $t$ , the $c$ -th element $P_{c} \in (0,1)$ , $c \in \{1,2,\dots,C\}$ + +represents the ratio of width $w_{tj} \in \mathbb{R}$ to height $h_{tj} \in \mathbb{R}$ of the $j$ -th object of category $c$ . $P_{c}$ is updated iteratively as: + +$$ +P _ {c} ^ {\prime} = \lambda P _ {c} + (1 - \lambda) \frac {w _ {t j}}{h _ {t j}}. \tag {7} +$$ + +where $\lambda$ is used to balance the contributions of knowledge distilled in current and previous epochs. + +# 4.3. Coarse Pseudo-Box Prediction + +We adjust anchors with the anchor prototype library for better coarse pseudo boxes and simultaneously collect a set of informative samples for the pseudo-box refinement. + +MIL-based Pseudo-Box Predictor: Given an adjusted bag of region proposals for each point label, a dual-stream structure is adopted as an MIL classifier to find the best box region in the bag to represent the target. As shown in Fig. 3(b), features of these proposals are extracted through a RoI [9] layer and two fully connected layers. Secondly, these features are fed into the dual-stream structure to estimate scores of each proposal in the same bag according to Eq. (1). Then, a bag's score is calculated by the sum of all its proposals. Finally, the Top- $k$ boxes with the highest scores in the same bag are chosen for coarse pseudo boxes. + +Informative-Sample Collector (IS Collector): To capture informative negative samples with high uncertainty, we present IS Collector, which aims to collect negatives with high loss values. Specifically, negative samples randomly sampled on background with their corresponding losses estimated by the MIL-based classifier, are fed into our IS collector for the pseudo-box refinement, as shown in Fig. 3(e). Considering the high computational costs of processing all negative samples, Non-Maximum Suppression [16] is first applied for de-duplication, removing redundant negative samples with high similarity and retaining representative ones. Then, we rank the filtered negative samples in descending order of their losses and thus collect the most informative samples for pseudo-box refinement. + +# 4.4. Pseudo-Box Refinement + +The collected informative samples are used to suppress candidate positive samples with low confidence so as to refine pseudo boxes. As shown in Fig. 3(c), the model for pseudo-box refinement has a similar structure as the model for coarse pseudo-box prediction. Specifically, for positive sample generation, we update the aspect ratios of coarse pseudo boxes and jitter their center positions. Meanwhile, given the informative negative samples collected in coarse pseudo boxes, after the RoI layer, all features of samples are fed into the MIL classifier according to Eq. (1). Finally, the relatively accurate Top- $k$ candidate proposals are chosen to be merged as the final predicted pseudo box. + +Discussion: As shown in Fig. 4(b), numerous proposals by random sampling in the background are far from targets. + +![](images/0b12f872133572c7a7b9125389ed278146c70d54867e6e5aa12b6cffed53d2aa.jpg) +(a) + +![](images/ac7af72bf99260191e05b56c9e8aeebb5c718b94c5c06fa39816e934ea3cda69.jpg) +(b) + +![](images/388ed6638282936231a959a86173cedc143ed52a185281db238e1651b9fa3141.jpg) +(c) +Figure 4. The negative samples distribution of PointSR and Point-to-Box. The histogram (b) shows the number of negative samples for different generation methods, where the distance denotes the Euclidean distance between the center of negative samples (Background) and positive samples (Pedestrians). Two specific examples for PointSR (a) and Point-to-Box (c) are visualized. + +Similarly, Fig. 4(c) shows partial negative samples, indicating that they provide few contributions to object detection. As for densely packed objects in drone-view images, edge information between objects and objects' neighborhood is crucial to distinguish each object. The random negative sampling in the background only generates proposals in a relatively stable sample space. While this sampling ensures diversity among negative samples, it may include less relevant ones, reducing learning efficiency. Additionally, since this method assigns equal weights to all negative samples, it can not adaptively select useful and informative negatives. + +Fig. 4(a) presents that negative samples produced by our IS Collector are close enough to the decision boundary in feature space. Training with these negative samples is beneficial to the MIL classifier to recognize which candidate proposals should be assigned low scores and force the model to learn finer distinctions because these negatives contribute more significantly to the gradients. By combining with the IS Collector, the TE Encoder refines positive samples through anchor aspect ratio adjustments, enhancing the MIL classifier's ability to predict losses for negative samples more accurately. This improvement enables the IS Collector to better select high-loss negative samples. As a result, the collaboration between the two modules yields high-quality pseudo-boxes. + +The Analysis of Scalability: The time complexity of TE Encoder, $\mathbf{O}(n)$ , scales linearly with the number of the predicted pseudo boxes $n$ . Whereas the IS Collector selects a few informative samples from around 400 negative samples randomly sampled from the background. Since this process is independent of $n$ and the image resolution, PointSR's computational overhead is stable, even for high-resolution images or larger datasets. + +# 4.5. Loss Functions + +Inspired by the discriminative method [3], we design the MIL loss and negative loss for our PointSR training. + +The MIL Loss: We formulate the MIL loss $\mathcal{L}_{mil}$ as a cross + +entropy loss for positive samples, given by: + +$$ +\mathcal {L} _ {m i l} = - \frac {1}{M} \sum_ {j = 1} ^ {M} \sum_ {k = 1} ^ {C} \log \left(\hat {\mathbf {y}} _ {j, k} ^ {\mathbf {y} _ {j, k}} \left(1 - \hat {\mathbf {y}} _ {j, k}\right) ^ {\left(1 - \mathbf {y} _ {j, k}\right)}\right), \tag {8} +$$ + +where $\mathbf{y}_{j,k} \in \{0,1\}$ denotes the one-hot category labels and $\hat{\mathbf{y}}_{jk}$ is the predict scores generated by MIL-based predictor. The MIL loss is used for the correct prediction of the category and instance to which the proposal belongs to. + +The Negative Loss: Furthermore, the negative loss $\mathcal{L}_{\text{neg}}$ in the form of focal loss is used for model optimization, which is formulated as follows: + +$$ +\mathcal {L} _ {n e g} = - \frac {1}{N} \sum_ {j = 1} ^ {N} \sum_ {k = 1} ^ {C} \beta \cdot (\hat {\mathbf {y}} _ {j, k}) ^ {\gamma} \log (1 - \hat {\mathbf {y}} _ {j, k}) \tag {9} +$$ + +where $\gamma$ is set to 0.2, the same in [13] and $\beta$ is a weight to balance the contribution of different negative samples. The Negative loss in the Refinement stage is used to further suppress the background for pseudo-box refinement. + +# 5. Experiments + +# 5.1. Datasets + +We conducted experiments on three benchmark datasets: DroneVehicle [22], VisDrone [34] and UAVDT [5]. + +**DroneVehicle:** The DroneVehicle dataset includes 56,878 images, which are annotated with box-level annotations across five vehicle categories, i.e., Car, Truck, Bus, Van and Freight Car, covering various scenarios, including urban roads, residential areas, and parking lots captured under different lighting conditions from day to night. + +VisDrone: The VisDrone dataset consists of 10,209 images with a resolution of $2,000 \times 1$ , 500 pixels and 542,000 instances across 10 common object categories in traffic scenarios, i.e., Pedestrian, People, Bicycle, Car, Van, Truck, Tricycle, Awning-tricycle, Bus and Motor. These images, captured from various urban scenes, present significant challenges due to the presence of numerous small objects, varying viewpoints, and heavy occlusions. To the hardware memory limitations, following PLUG [10], we cropped the original images into patches of $512 \times 512$ pixels with a 128-pixel overlap for training and inference. + +UAVDT: The UAVDT dataset consists of 100 video sequences of about 80,000 representative frames with a resolution of $1,024 \times 540$ pixels and 0.84 million box-level annotations across three vehicle categories: Car, Truck, Bus. + +To accurately reproduce the biases inherent in manual annotation, we employed Rectified Gaussian Distribution (RG) [30] with central ellipse constraints, assuming that the point distribution of manual annotation approximately follows an RG distribution. Therefore, for our experiments, the three datasets with point-level annotations are employed + +for evaluating our approach, referred to as "DroneVehicle (Point-Supervision)", "VisDrone (Point-Supervision)", and "UAVDT (Point-Supervision)", respectively. + +# 5.2. Implementation and Evaluation Metrics + +Implementation Details: We implemented our proposed PointSR based on the MMDetection toolbox on an NVIDIA RTX 4090 GPU. The Stochastic Gradient Descent (SGD) algorithm was used to optimize with a $3 \times$ training schedule. The learning rate was set to 0.02 and decayed by a factor of 0.1 at the 24th and 33rd epochs, respectively. All comparison models were configured with a ResNet-50 [8] backbone pre-trained on ImageNet [4]. + +Evaluation Metrics: We evaluated our method from three aspects: the quality of predicted pseudo boxes, the detection performance and the model's robustness. The mean Intersection over Union (mIoU) between generated pseudo boxes and ground-truth boxes was utilized to evaluate the quality of predicted pseudo boxes, where mIoU closer to 1 indicates higher accuracy. The Average Precision $(\mathrm{AP}_{50})$ [6] was utilized to evaluate the models' detection performance. Finally, the variances of AP across different point-label distributions were used to measure models' robustness, with lower variance indicating higher robustness. + +# 5.3. Comparison with SOTA Methods + +We compared the performance of a detector trained using pseudo boxes generated by our PointSR and five SOTA point-supervised DroneDet methods: PLUG [10], PointOBB [15], Point2Rbox [32], PointOBBv2 [18] and Point-to-Box [3]. In these experiments, pseudo boxes were first generated by our PointSR and SOTA methods. These pseudo boxes were then utilized to train a classic detector, Faster-RCNN [19], to evaluate its detection performance and robustness. In addition, a Faster-RCNN detector was trained with ground-truth box-level annotations, which detection performance was referred to as the upper bound. + +The Detection Performance: Table 1 shows that the detector trained with pseudo boxes generated by our PointSR achieves superior detection performance, outperforming that trained with pseudo boxes generated by SOTA methods significantly on all datasets. Specifically, our model improved $\mathrm{AP}_{50}$ by $7.2\%$ on DroneVehicle (Point-Supervision) dataset, $3.3\%$ on VisDrone (Point-Supervision) dataset and $5.4\%$ on UAVDT (Point-Supervision) dataset, compared to Point-to-Box [3]. Besides, it is worth mentioning that the detection results of our model trained with point-level annotations are better than the results of Faster-RCNN trained with box-level annotations in some cases in Fig. 5. + +The Quality of the Predicted Pseudo-Boxes: We use mIoU between the generated pseudo boxes and ground-truth boxes to evaluate the quality of the predicted pseudo boxes, where mIoU closer to 1 indicates higher accuracy. + +Table 1. Comparisons of per-class detection accuracy $(\mathrm{AP}_{50})$ across different object classes, average $\mathrm{AP}_{50}$ , and mIoU for a detector trained using pseudo boxes generated by our PointSR and SOTA methods on VisDrone, DroneVehicle and UAVDT. "Faster R-CNN" indicates the accuracy obtained with the detector trained with ground-truth box-level annotations, serving as the upper bound for performance. + +
MethodVisDroneDroneVehicleUAVDT
Ped.Peo.Bic.CarVanTru.Tri.Awn.BusMot.AP50mIoUCarTru.BusVanFre.AP50mIoUCarTru.BusAP50mIoU
PLUG [10]0.63.44.73.40.24.50.70.323.65.34.726.91.06.10.28.02.33.529.04.58.23.85.531.1
Point2Rbox [32]0.30.29.122.315.315.311.59.915.81.410.1-0.03.675.718.928.025.2-15.724.119.219.7-
PointOBB [15]25.613.620.834.612.430.732.930.456.512.427.056.218.37.06.034.510.115.277.529.230.816.825.672.6
PointOBBv2 [18]33.24.316.169.257.431.632.123.232.823.832.464.817.119.734.229.222.824.680.116.628.837.227.577.5
Point-to-Box [3]40.24.423.162.953.130.930.922.828.419.631.766.268.925.524.233.024.735.386.045.535.818.533.395.3
Faster R-CNN [19]53.730.930.072.257.644.231.325.035.837.241.8-89.766.092.253.352.770.1-66.955.759.660.7-
PointSR (Ours)43.45.826.265.653.531.529.722.444.727.735.075.782.825.133.542.229.142.588.950.140.825.438.796.8
+ +Table 2. Comparison of average precision $(\mathrm{AP}_{50})$ and variance for a detector trained using pseudo boxes generated by our PointSR and SOTA methods across different point-label distributions. PointSR consistently demonstrates lower variance on both datasets, exhibiting greater stability and robustness to annotation perturbations. + +
MethodPublicationVisDroneDroneVehicle
U0×LU0.1×LU0.3×LU0.5×LVariance Of AP50U0×LU0.1×LU0.3×LU0.5×LVariance Of AP50
PLUG [10]TGRS'241.54.315.517.848.911.13.930.132.5209.1
Point2Rbox [32]CVPR'247.69.98.39.10.74215.224.117.618.310.61
PointOBB [15]CVPR'2424.328.517.88.556.9113.415.212.68.95.269
Point-to-Box [3]ECCV'2231.632.531.130.10.75234.835.532.934.00.935
PointSR (Ours)N/A34.634.833.933.70.21241.142.041.340.90.171
+ +Table 3. The impact of the TE Encoder and Informative Sampling (IS) on the baseline Point-to-Box [3], evaluated on DroneVehicle. + +
TEEISAP50mIoU [%]
Baseline××35.2686.0
Baseline + TEE✓×36.1886.7
Baseline + IS×✓40.8788.1
Baseline + TEE + IS✓✓42.5388.9
+ +Tables 1 also show that PointSR achieves better mIoUs than the SOTA methods on both datasets. Specifically, our model improves mIoU by $9.5\%$ on Visdrone, $2.9\%$ on DroneVehicle and $1.5\%$ on UAVDT compared to Point-to-Box [3]. + +The visualization in Fig. 6 shows that our predicted pseudo boxes accurately localize objects in challenging drone-view images by leveraging the precise location information of point annotations and can distinguish dense objects in complex scenes. In contrast, Point-to-Box often produces loose pseudo boxes due to its limited sampling strategy for MIL training, and PLUG occasionally predicts oversized boxes due to outliers in the generative process. These results confirm the high quality of pseudo boxes generated by our model with point-level annotations. + +The Robustness of the Model: To evaluate the model's robustness, we use the variance of $\mathrm{AP}_{50}$ across different point distributions. Specially, we assume manually annotated points follow a uniform distribution $x\sim U(-a,a)$ , $a = \alpha \times L$ , $\alpha = \{0,0.1,0.3,0.5\}$ and $L$ is the maximum value for the width and height of a bounding box. $U_{0\times L}$ , $U_{0.1\times L}$ , $U_{0.3\times L}$ , and $U_{0.5\times L}$ denote the manually annotated points $x\sim U(-\alpha \times L,\alpha \times L)$ with offsets of 0, $10\%$ , + +$30\%$ , $50\%$ , respectively. As shown in Table 2, our model has achieved a variance of $\mathrm{AP}_{50}$ 0.171 on the DroneVehicle dataset, compared to 0.935 of Point-to-Box [3], indicating superior robustness to annotation perturbation. + +# 5.4. Ablation Studies + +We conducted ablation studies on the Drone Vehicle (Point-Supervision) dataset using Point-to-Box [3] as the baseline to validate the effectiveness of the two key components in PointSR: the TE Encoder and IS Collector. + +The Effectiveness of the TE Encoder: We conducted a quantitative evaluation to verify the effectiveness of the TE Encoder in our PointSR. The baseline with TE Encoder is referred to as "Baseline + TEE". Table 3 shows that our proposed module improves $\mathrm{AP}_{50}$ by $0.92\%$ and mIoU by $0.7\%$ , compared with the vanilla baseline model. Moreover, TE Encoder facilitates stable and efficient training, in Fig. 7(a), where the orange curve (Baseline + TEE) descends faster and is smoother compared with the blue curve, especially in the early stage of training, indicating that our Anchor Learning model achieves stable training and faster convergence. In addition, Fig. 7(b) further demonstrates that "Baseline + TEE" achieves better detection performance. Finally, Table 4 shows the performance of Faster R-CNN trained using PointSR with different $\lambda$ settings, where PointSR with $\lambda = 0.5$ achieves the best detection performance. These results verify the effectiveness of the TE Encoder. + +Table 4. Impact of different $\lambda$ values on ${\mathrm{{AP}}}_{50}$ in DroneVehicle. + +
λ0.10.30.50.70.9
AP50 [%]34.240.542.534.933.9
+ +![](images/2ab5bd1e68a37f9fac28eb85b7fe29b67fb21168c48516db8c48e9fd4993b7d5.jpg) + +![](images/842e2739c5726a3597088ede6ec2bc4fae4ae9566c2b503187565406eb8752dd.jpg) + +![](images/3fee694ba271e870394dbb783cbbedd4e4f9d1427d4e13232b7478a2d0ccf0a1.jpg) + +![](images/2c6aed0d7d051948b739b514ad9445f14ad50e5e966cbaabe5ae1d1b0aa34a09.jpg) + +![](images/19b0f7f2bcac7ca0a6f5480ff581bef80d20fe23a34194da6d58815bd6bd035d.jpg) + +![](images/2d7ff54c9952ae1433f159d6eab948c5dea8a890650e0d04e07ebed100e594c9.jpg) + +![](images/193dbc5dc579c5ef93983df44dabb5f6591223a5a1405784a732231fc631bd72.jpg) + +![](images/de0e591a03e899085ec4cb1c1ca60a35dd6aa0c5bb176949035417608204ce91.jpg) + +![](images/15ebc8de4d615f4298ae1809deb56d3f2acf50c55d7d6c4fab1be7cb983c6599.jpg) + +![](images/45f50f650bc813487a0fdf273a3cccb8a9720e1f6b8b12e720aa498ef631aaed.jpg) + +![](images/ebbd4f1405439f97eef5400133c70fef5dcc3614b63dd812d0b0be8f014e2104.jpg) + +![](images/eadcc5a9674d5551c82e162e15f790f673d20caeebad517115f806787e6b7445.jpg) + +![](images/df3beca3190aa231e17e22992f573b72bea68251efd4a8b02e36479609da4f83.jpg) + +![](images/373ea0f28963dde991576207a441795b4f99c94d00e023ad4525dd9590429d99.jpg) + +![](images/49346a8337214009e048e2780b51c7f38c55ff90cf7d14379ae691f31b2295f0.jpg) +PLUG + +![](images/7800909db44a5597fe4b9c721cf8ceeb94609423177a6f3a9838ea772886406a.jpg) +Point2Rbox + +![](images/d618c06ed51f814bbb66ba92c2a41aa015c92c14df6e21e39a5fdbb578123247.jpg) +PointOBB + +![](images/78311976e08c06e25af1eccfeaf10319b213bc33fb87f57abd132a6e24990527.jpg) +Point-to-Box + +![](images/e4ea27984c2b42950187254adb2ce9975cf3c006901e38dd526532bc9bccc170.jpg) +Faster-RCNN + +![](images/947e3b7c408d04e477740e93378035313dc19aa8d0bc59ed667f4ddbbb342c90.jpg) +PointSR(Ours) + +![](images/66ec264911bcbb0f672f0a5d6baeec402ebb848e0f45fd1b1aec803c273774a8.jpg) +GT + +![](images/f72046695e71cdb2af191a7f3aa0a63d595adaf0135dfaf762b27c170dec981b.jpg) +Figure 5. Visualization of detection results from a detector trained with pseudo boxes generated by PointSR and SOTA methods. Ground-truth bounding boxes are shown in green, correct detections in yellow, and misclassified or missed targets in red. For a better view, detection results in the last row are highlighted by cyan rectangular boxes. Best viewed in color and zoomed-in mode. + +![](images/e049363795044915f1b0457cd1491b3ac8dc0d07346ee85c154015fcf99db252.jpg) +(a) + +![](images/5a53ad46899b8d191dbd8e38ec952c2af9cf29d76cfba86a83c2068ff7b53369.jpg) +Figure 6. Visualization of pseudo boxes (pink) predicted by PointSR and SOTA methods. Ground-truth boxes are shown in green. Best viewed in color and zoomed-in mode. +(b) +Figure 7. The loss and $\mathrm{AP}_{50}$ curves during the whole training stage. The blue curves represent baseline training, while the orange curves show baseline training equipped with our TE Encoder. + +The Effectiveness of the IS Collector: We evaluated the effectiveness of the Informative Sampling Collector in our PointSR. Table 3 shows that the baseline with our Informative Sampling module, referred to as "Baseline + IS", + +![](images/1b57e286ba27eb69786d64ee192dbbdac30c29ce5a9874c3afd44af10d63ffee.jpg) +(a) + +![](images/a2db59f364d79f76298aa0506db13f17cfc78c846879521a3e68c4ae9cedf095.jpg) +(b) +Figure 8. The $\mathrm{AP}_{50}$ and IoU curves for varying Sampling Ratio $\delta$ . improves $\mathrm{AP}_{50}$ by $5.26\%$ and mIoU by $2.2\%$ . These experimental results confirm the effectiveness of our IS Collector. In addition, we also analyzed the impact of the ratio of selected negative samples to whole negative samples $\delta$ , denoted as "Sampling Ratio", on detection performance. Fig. 8 visualizes that changes in $\delta$ significantly affect the detection performance of our PiontSR in terms of $\mathrm{AP}_{50}$ and mIoU, suggesting its critical role in optimizing results. + +# 6. Conclusion + +Towards generating high-quality pseudo boxes for PSOD from drone-view images, in this paper, we proposed an effective and robust self-regularized point-supervision approach, PointSR, which leverages a Temporal-Ensemble Encoder for stable pseudo-box generation and an Informative Sampling Collector for efficient pseudo-box refinement. The two processes mutually benefit each other, facilitating the generation of self-regularized high-quality pseudo boxes. Experimental results on benchmark datasets have demonstrated that detectors trained with pseudo boxes generated by PointSR significantly outperform those generated by SOTAs over $2.6\% \sim 7.2\%$ in DroneDet and gain the robust results under perturbations of the manually annotated labels, significantly advancing the applicability of PSOD. + +Acknowledgments: This work was supported by the National Natural Science Foundation of China (Grant No. 62402370, 62301397), the Natural Science Basic Research Program of Shaanxi (Grant No. 2024JC-YBQN-0728), the Guangzhou Basic and Applied Basic Research Foundation (Grant No. 2023A04J1742). Additionally, we would like to thank Zhiwen Wei for his helpful discussions and insightful suggestions. + +# References + +[1] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2854, 2016. 1, 2, 3 +[2] Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. Just a Hint: Point-Supervised Camouflaged Object Detection. In European Conference on Computer Vision, pages 332-348. Springer, 2024. 2, 3 +[3] Pengfei Chen, Xuehui Yu, Xumeng Han, Najmul Hassan, Kai Wang, Jiachen Li, Jian Zhao, Humphrey Shi, Zhenjun Han, and Qixiang Ye. Point-to-Box Network for Accurate Object Detection via Single Point Supervision. In European Conference on Computer Vision, pages 51-67. Springer, 2022. 1, 2, 3, 5, 6, 7 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-scale Hierarchical Image Database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 6 +[5] Dawei Du, Yuankai Qi, Hongyang Yu, Yifan Yang, Kaiwen Duan, Guorong Li, Weigang Zhang, Qingming Huang, and Qi Tian. The Unmanned Aerial Vehicle Benchmark: Object Detection and Tracking. In Proceedings of the European Conference on Computer Vision, pages 370-386, 2018. 2, 6 +[6] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes Challenge: A Retrospective. International Journal of Computer Vision, 111:98-136, 2015. 6 +[7] Xiaoxu Feng, Xiwen Yao, Gong Cheng, and Junwei Han. Weakly Supervised Rotation-Invariant Aerial Object Detection Network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14146-14155, 2022. 1 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6 +[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In Proceedings of the IEEE International Conference on Computer Vision, pages 2961-2969, 2017. 5 +[10] Shitian He, Huanxin Zou, Yingqian Wang, Boyang Li, Xu Cao, and Ning Jing. Learning Remote Sensing Object Detection With Single Point Supervision. IEEE Transactions + +on Geoscience and Remote Sensing, 62:1-16, 2024. 2, 3, 6, 7 +[11] Boyang Li, Yingqian Wang, Longguang Wang, Fei Zhang, Ting Liu, Zaiping Lin, Wei An, and Yulan Guo. Monte Carlo Linear Clustering with Single-Point Supervision is Enough for Infrared Small Target Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1009–1019, 2023. 2 +[12] Jing Li, Junsong Fan, and Zhaoxiang Zhang. Point-Supervised Panoptic Segmentation via Estimating Pseudo Labels from Learnable Distance. In European Conference on Computer Vision, pages 95-112. Springer, 2024. 2 +[13] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal Loss for Dense Object Detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 6 +[14] Junyan Lu, Qinglei Hu, Ruifei Zhu, Yali Wei, and Tie Li. AFWS: Angle-Free Weakly-Supervised Rotating Object Detection for Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1 +[15] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. PointOBB: Learning Oriented Object Detection via Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16730-16740, 2024. 2, 3, 6, 7 +[16] Alexander Neubeck and Luc Van Gool. Efficient Nonmaximum Suppression. In 18th International Conference on Pattern Recognition, pages 850-855. IEEE, 2006. 5 +[17] Dim P Papadopoulos, Jasper RR Uijlings, Frank Keller, and Vittorio Ferrari. Training Object Class Detectors with Click Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6374-6383, 2017. 2, 3 +[18] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection. In The Thirteenth International Conference on Learning Representations, 2025. 6, 7 +[19] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2016. 4, 6, 7 +[20] Tianqi Ren, Qiu Shen, Ying Fu, and Shaodi You. Point-Supervised Semantic Segmentation of Natural Scenes via Hyperspectral Imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1357-1367, 2024. 2 +[21] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, Ming-Yu Liu, Alexander G Schwing, and Jan Kautz. $\mathbf{UFO}^2$ : A Unified Framework Towards Omni-Supervised Object Detection. In European Conference on Computer Vision, pages 288-313. Springer, 2020. 2, 3 +[22] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Drone-Based RGB-Infrared Cross-Modality Vehicle Detection via Uncertainty-Aware Learning. IEEE Transactions on Circuits and Systems for Video Technology, 32(10):6700-6713, 2022. 2, 6 + +[23] Zhiwen Tan, Zhiguo Jiang, Chen Guo, and Haopeng Zhang. WSODet: A Weakly Supervised Oriented Detector for Aerial Object Detection. IEEE Transactions on Geoscience and Remote Sensing, 61:1-12, 2023. 1 +[24] Peng Tang, Xinggang Wang, Song Bai, Wei Shen, Xiang Bai, Wenyu Liu, and Alan Yuille. PCL: Proposal Cluster Learning for Weakly Supervised Object Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(1): 176-191, 2018. 2 +[25] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective Search for Object Recognition. International Journal of Computer Vision, 104: 154-171, 2013. 3 +[26] Yuting Wang, Velibor Ilic, Jiatong Li, Branislav Kisačanin, and Vladimir Pavlovic. ALWOD: Active Learning for Weakly-Supervised Object Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6459–6469, 2023. 1 +[27] Yue Xi, Wenjing Jia, Qiguang Miao, Junmei Feng, Jinchang Ren, and Heng Luo. Detection-Driven Exposure-Correction Network for Nighttime Drone-View Object Detection. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1 +[28] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Query-Det: Cascaded Sparse Query for Accelerating High-Resolution Small Object Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13668–13677, 2022. 1 +[29] Xinyi Ying, Li Liu, Yingqian Wang, Ruojing Li, Nuo Chen, Zaiping Lin, Weidong Sheng, and Shilin Zhou. Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15528-15538, 2023. 2 +[30] Xuehui Yu, Pengfei Chen, Di Wu, Najmul Hassan, Guorong Li, Junchi Yan, Humphrey Shi, Qixiang Ye, and Zhenjun Han. Object Localization Under Single Coarse Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4868-4877, 2022. 2, 6 +[31] Xuehui Yu, Pengfei Chen, Kuiran Wang, Xumeng Han, Guorong Li, Zhenjun Han, Qixiang Ye, and Jianbin Jiao. CPR++: Object Localization via Single Coarse Point Supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[32] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16783-16793, 2024. 2, 3, 6, 7 +[33] Xiang Yuan, Gong Cheng, Kebing Yan, Qinghua Zeng, and Junwei Han. Small Object Detection via Coarse-to-Fine Proposal Generation and Imitation Learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6317-6327, 2023. 1 +[34] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and Tracking + +Meet Drones Challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 2, 6 \ No newline at end of file diff --git a/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/images.zip b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a6011d777fa4ebefdffe7ccb588aba1e260d7e2 --- /dev/null +++ b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53ffae116b8c84d333b49f8eca34670f5e3b88864715c999b048a6c403eb3a39 +size 723097 diff --git a/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/layout.json b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..36fc1e8ba3f342777e888eb20f5e0d21477bc6b7 --- /dev/null +++ b/2025/PointSR_ Self-Regularized Point Supervision for Drone-View Object Detection/layout.json @@ -0,0 +1,10345 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 545, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 545, + 120 + ], + "type": "text", + "content": "PointSR: Self-Regularized Point Supervision for Drone-View Object Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "text", + "content": "Weizhuo Li" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "text", + "content": " Yue Xi" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "text", + "content": " Wenjing Jia" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "text", + "content": " Zehao Zhang" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "text", + "content": " Fei Li" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "text", + "content": " Xiangzeng Liu" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "text", + "content": " Qiguang Miao" + }, + { + "bbox": [ + 61, + 141, + 569, + 159 + ], + "type": "inline_equation", + "content": "^{3*}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 175, + 159, + 449, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 159, + 449, + 171 + ], + "spans": [ + { + "bbox": [ + 175, + 159, + 449, + 171 + ], + "type": "text", + "content": "1Guangzhou Institute of Technology, Xidian University, Guangzhou, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 172, + 468, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 172, + 468, + 185 + ], + "spans": [ + { + "bbox": [ + 159, + 172, + 468, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 172, + 468, + 185 + ], + "type": "text", + "content": " Faculty of Engineering and IT, University of Technology Sydney, Sydney, Australia" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 167, + 186, + 457, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 186, + 457, + 198 + ], + "spans": [ + { + "bbox": [ + 167, + 186, + 457, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 167, + 186, + 457, + 198 + ], + "type": "text", + "content": " School of Computer Science and Technology, Xidian University, Xi'an, China" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 111, + 201, + 515, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 201, + 515, + 213 + ], + "spans": [ + { + "bbox": [ + 111, + 201, + 515, + 213 + ], + "type": "text", + "content": "{liweizhuo@stu., xiyue@, zzehao@stu., lifei@, xzliu@, qgmiao@}xidian.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 251, + 216, + 372, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 216, + 372, + 226 + ], + "spans": [ + { + "bbox": [ + 251, + 216, + 372, + 226 + ], + "type": "text", + "content": "Wenjing.Jia@uts.edu.au" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 152, + 255, + 199, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 255, + 199, + 266 + ], + "spans": [ + { + "bbox": [ + 152, + 255, + 199, + 266 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 280, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 616 + ], + "type": "text", + "content": "Point-Supervised Object Detection (PSOD) in a discriminative style has recently gained significant attention for its impressive detection performance and cost-effectiveness. However, accurately predicting high-quality pseudo-box labels for drone-view images, which often feature densely packed small objects, remains a challenge. This difficulty arises primarily from the limitation of rigid sampling strategies, which hinder the pseudo-box optimization process. To address this, we propose PointSR, an effective and robust point-supervised object detection framework with self-regularized sampling that integrates temporal and informative constraints throughout the pseudo-box generation process. Specifically, the framework comprises three key components: Temporal-Ensembling Encoder (TE Encoder), Coarse Pseudo-box Prediction, and Pseudo-box Refinement. The TE Encoder builds an anchor prototype library by aggregating temporal information for dynamic anchor adjustment. In Coarse Pseudo-box Prediction, anchors are refined using the prototype library, and a set of informative samples is collected for subsequent refinement. During Pseudo-box Refinement, these informative negative samples are used to suppress low-confidence candidate positive samples, thereby improving the quality of the pseudoboxes. Experimental results on benchmark datasets demonstrate that PointSR significantly outperforms state-of-the-art methods, achieving up to " + }, + { + "bbox": [ + 55, + 280, + 296, + 616 + ], + "type": "inline_equation", + "content": "2.6\\% \\sim 7.2\\%" + }, + { + "bbox": [ + 55, + 280, + 296, + 616 + ], + "type": "text", + "content": " higher " + }, + { + "bbox": [ + 55, + 280, + 296, + 616 + ], + "type": "inline_equation", + "content": "AP_{50}" + }, + { + "bbox": [ + 55, + 280, + 296, + 616 + ], + "type": "text", + "content": " using only point supervision. Additionally, it exhibits strong robustness to perturbation in human-labeled points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 627, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "Drone-view object detection (DroneDet) [7, 23, 27, 28, 33], which involves detecting objects from drone-captured images, plays a vital role in various domains, including national defense, resource management, and drone inspection." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 252, + 557, + 401 + ], + "blocks": [ + { + "bbox": [ + 316, + 252, + 557, + 401 + ], + "lines": [ + { + "bbox": [ + 316, + 252, + 557, + 401 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 557, + 401 + ], + "type": "image", + "image_path": "a47cef66bbead7e6e2bf9a2f294ce9d285d104eb044972bf0e5f2fdc2fa52920.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 555, + 470 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 470 + ], + "type": "text", + "content": "Figure 1. Comparisons of the pseudo-box optimization process. (a) The rigid sampling strategy in Point-to-Box [3] results in the inefficiency of the training process for low-quality pseudo-box labels. (b) The self-regularized sampling in our PointSR improves the stability and efficiency of the training process and thus obtains high-quality pseudo-box labels." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 474, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 474, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 312, + 474, + 555, + 641 + ], + "type": "text", + "content": "However, DroneDet models often experience a significant performance drop when applied to new scenes or datasets due to their limited generalization capabilities. A straightforward solution is to collect and annotate new datasets and retrain the models. However, annotating drone-view images with box-level annotations containing accurate position information is particularly challenging due to the prevalence of many densely packed small targets, far more complex than annotating ground-view images. Moreover, the diversity of the scenes and object types can lead to combinatorial explosions of labeling requirements, making the process infeasible in many cases. Consequently, constructing new datasets with box-level annotations is both time-consuming and costly when deploying drones in new scenarios." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 642, + 556, + 714 + ], + "type": "text", + "content": "To alleviate the cost associated with box-level annotations, Image-Supervised Object Detection (ISOD) [1, 14, 26] has been extensively studied. ISOD treats images as a bag of regions, assuming at least one region contains the object and uses a Multiple Instance Learning (MIL) classifier to identify regions corresponding to image-level labels. De" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11707" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "spite its labeling efficiency, ISOD methods often perform poorly in complex scenarios such as drone-view images, with targets typically small and densely packed, due to the lack of essential position information in image-level labels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 231 + ], + "type": "text", + "content": "To overcome this limitation, Point-Supervised Object Detection (PSOD) has emerged as an alternative solution. PSOD provides minimal and yet additional distinctive position information for each object and has gained significant attention in tasks such as detection [10, 32], localization [30, 31], and segmentation [12, 20]. According to [17], the labor cost of point annotations per object is only about " + }, + { + "bbox": [ + 55, + 122, + 295, + 231 + ], + "type": "inline_equation", + "content": "1/18" + }, + { + "bbox": [ + 55, + 122, + 295, + 231 + ], + "type": "text", + "content": " of that for bounding boxes, making PSOD a highly cost-effective solution for DroneDet applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 484 + ], + "type": "text", + "content": "Existing PSOD methods [3, 10, 11, 15, 29, 32] can be broadly categorized into generative and discriminative approaches. Generative methods [10, 11, 29] predict a masked region covering a target labeled by a point and then derive a pseudo box by estimating the minimum circumscribed rectangle of the region. However, pixels in the predicted mask of each object are susceptible to misclassification, resulting in inaccurate and unstable pseudo-box predictions. Discriminative methods [3, 15, 32] first generate a bag of proposal boxes for each label and then select the optimal one as the final pseudo-box through an MIL-based optimization process, as illustrated in Fig. 1(a). A reasonable sampling strategy aiming to extract the labeled training set from an unlabeled dataset is critical for the successful training of the MIL classifier. Despite their potential, these methods still face two significant challenges, as shown in Fig. 2(a): 1) Static positive sampling in an anchor style leads to training instability, especially in the early stages (see the blue curves in Fig. 7); 2) Inefficient negative sampling that randomly samples background regions introduces numerous irrelevant negative samples, reducing training efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 714 + ], + "type": "text", + "content": "To address these limitations, we propose PointSR, an effective and robust point-supervised DroneDet framework with self-regularized sampling. As shown in Fig. 1(b), PointSR integrates temporal and informative constraints throughout the pseudo-box generation process to produce high-quality pseudo-box labels. Temporal-Ensembling Encoder (TE Encoder) first builds an anchor prototype library by collecting and encoding temporal information from pseudo-box labels during refinement. Then, the Coarse Pseudo-box Generation adjusts anchors using the anchor prototype library to produce better coarse pseudo boxes while collecting a set of informative samples for the subsequent refinement. Finally, during the Refinement stage, the collected informative samples are used to suppress low-confidence candidate positive samples to refine pseudo boxes. Benefiting from this self-regularized sampling strategy, PointSR not only ensures a more stable and efficient training process but also effectively improves the model's robustness to perturbation brought by manual labeling." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 553, + 201 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 553, + 201 + ], + "type": "image", + "image_path": "9c7531681c4b12d3ae9dce9890344cc34d695c37ee8232a48466cf33bb2dafb8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 204, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 270 + ], + "type": "text", + "content": "Figure 2. (a) Point-to-Box [3] uses a rigid sampling strategy with fixed positive proposals and randomly sampled negative background regions for training the MIL-based predictor. (b) PointSR adopts a self-regularized sampling strategy, incorporating temporal and informative constraints into the pseudo-box generation process for both positive and negative sampling." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 278, + 541, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 278, + 541, + 289 + ], + "spans": [ + { + "bbox": [ + 326, + 278, + 541, + 289 + ], + "type": "text", + "content": "Our key contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 290, + 553, + 422 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 290, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 553, + 326 + ], + "type": "text", + "content": "1) We propose PointSR, a novel and robust framework for point-supervised DroneDet, capable of generating high-quality pseudo-box labels from point-level supervision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "type": "text", + "content": "2) We design a Temporal-Ensemble Encoder for stable pseudo-box prediction and an informative negative sampling strategy to enhance pseudo-box quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "text", + "content": "3) PointSR achieves an average " + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "inline_equation", + "content": "42.5\\%" + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "inline_equation", + "content": "38.7\\%" + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "inline_equation", + "content": "35.0\\%" + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "text", + "content": " on three drone-captured benchmark datasets, i.e., DroneVehicle [22], UAVDT [5], and VisDrone [34] with only point supervision, outperforming the previous best methods by " + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 313, + 361, + 553, + 422 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 433, + 400, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 400, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 400, + 445 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 448, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 553, + 472 + ], + "type": "text", + "content": "In this section, we briefly review the current state of research on ISOD, PSOD and PSOD for aerial images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 479, + 539, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 539, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 539, + 492 + ], + "type": "text", + "content": "2.1. Image-Supervised Object Detection (ISOD)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 497, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 640 + ], + "type": "text", + "content": "ISOD refers to an approach in which a detector is trained with image-level annotations rather than box-level ones. The pioneering work WSDDNs [1] integrated the MIL into ISOD, employing a dual-stream structure to classify proposals, and enabling effective weakly-supervised object detection. Inspired by this, Tang et al. [24] designed a Proposal Cluster Learning (PCL) approach, which used proposal clusters as supervision to infer rough object locations. However, image-level annotations lack explicit position information, leading to low-quality proposals, such as background regions, in the potential object tool. This limitation makes it infeasible to directly apply ISOD to DroneDet." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 538, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 538, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 538, + 661 + ], + "type": "text", + "content": "2.2. Point-Supervised Object Detection (PSOD)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "PSOD is a relatively new field, with only a few methods [2, 3, 17, 21] published in recent years. By leveraging point-level annotations, PSOD enables more accurate object localization and classification compared to image-level" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11708" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": "annotations. From the perspective of human cognition, Papadopoulos et al. [17] introduced the first model utilizing center-click annotations, pioneering the use of point supervision in object detection as an alternative to conventional box-level supervision. Ren et al. [21] proposed a unified framework that performs object detection across various annotation types. Chen et al. designed an MIL-based framework Point-to-Box [3] to achieve object detection by classifying proposals. In camouflaged object detection, [2] employed a novel point-based learning paradigm for the challenging task with an attention regulator to distinguish objects that conceal themselves in the environment. However, these PSOD methods only focus on objects in ground-view images and produce poor performance in complex aerial scenarios due to numerous densely packed small objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 192, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 192, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 192, + 270 + ], + "type": "text", + "content": "2.3. PSOD for Aerial Images" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 275, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 467 + ], + "type": "text", + "content": "Existing PSOD methods for aerial images can be grouped into two categories: generative and discriminative fashions. The generative approaches, such as PLUG [10], employ a short path algorithm to produce pseudo boxes with pixel-level labels. However, the pseudo boxes rely heavily on the accuracy of pixel-level mask generation. If there is an outlier far from the object mask, the corresponding pseudo label will be seriously affected. In contrast, discriminative methods aim to improve the quality of pseudo labels through advanced sampling strategies. PointOBB [15] produced pseudo labels through the collaborative utilization of three distinctive views to achieve coupled optimization of scale and angle. Point2Rbox [32] combined point annotations and synthetic pattern knowledge with bounding boxes to improve the diversity of samples and provide more supervised information for accurate pseudo labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 467, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 598 + ], + "type": "text", + "content": "However, despite these advancements, existing methods still suffer from the limitation of fixed positive candidate proposals and the instability of the random distribution of negative samples. Specifically, the positive sample space remains constant during training and the negative one is only affected slightly by the coarse pseudo boxes in the conventional PSOD framework, formulated in Eq. (5). To address these limitations, we propose an effective and robust PSOD framework to generate high-quality pseudo boxes with a self-regularized sampling strategy, adjusting sampling space adaptively during the model training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 608, + 140, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 140, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 140, + 620 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": "Our PointSR is built upon the Weakly Supervised Deep Detection Networks (WSDDNs) [1] for ISOD. The input data consists of a collection of images, each associated with image-level labels " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\{0,1\\}^C" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": " represents the number of object categories. Candidate object regions " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{R_1,R_2,\\dots ,R_{|\\mathcal{R}|}\\}" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": " are generated by using region proposal mechanisms [25]. Features of these regions " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{D \\times |\\mathcal{R}|}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 495, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 495, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 495, + 84 + ], + "type": "text", + "content": "are extracted through fully connected layers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "text", + "content": "WSDDNs adopts a dual-stream structure, consisting of a classification and detection stream. The extracted features " + }, + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "text", + "content": " are fed into the two streams, producing two matrices, " + }, + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^c" + }, + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^d\\in \\mathbb{R}^{C\\times |\\mathcal{R}|}" + }, + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "text", + "content": ", representing the category and position prediction scores for each region, respectively. These scores are then passed through the corresponding softmax operators " + }, + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{cla}}(\\cdot)" + }, + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{det}}(\\cdot)" + }, + { + "bbox": [ + 313, + 84, + 555, + 168 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 322, + 175, + 553, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 322, + 175, + 553, + 217 + ], + "type": "interline_equation", + "content": "\\left[ \\sigma_ {\\mathrm {c l a}} \\left(\\mathbf {x} ^ {c}\\right) \\right] _ {i j} = \\frac {e ^ {x _ {i j} ^ {c}}}{\\sum_ {k = 1} ^ {C} e ^ {x _ {k j} ^ {c}}}, \\left[ \\sigma_ {\\det } \\left(\\mathbf {x} ^ {d}\\right) \\right] _ {i j} = \\frac {e ^ {x _ {i j} ^ {d}}}{\\sum_ {k = 1} ^ {| \\mathcal {R} |} e ^ {x _ {i k} ^ {d}}}. \\tag {1}", + "image_path": "07aad22ac81d3837611425a01d3515ed4fe3c7164af6de6c67fede66690bf60e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "type": "text", + "content": "The regions' scores are then calculated via element-wise multiplication as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 381, + 247, + 553, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 247, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 381, + 247, + 553, + 262 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\mathcal {R}} = \\sigma_ {\\operatorname {c l a}} \\left(\\mathbf {x} ^ {c}\\right) \\cdot \\sigma_ {\\det } \\left(\\mathbf {x} ^ {d}\\right). \\tag {2}", + "image_path": "8c63ad2eb6ea3bfd635f32b9dd0f11d0c8ff296ebe428c0b936ed3ac6e12e41e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 269, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 293 + ], + "type": "text", + "content": "Finally, an image-level class prediction score is obtained by summing over all regions as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 403, + 300, + 553, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 300, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 403, + 300, + 553, + 335 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} _ {c} = \\sum_ {r = 1} ^ {| \\mathcal {R} |} \\mathbf {x} _ {c r} ^ {\\mathcal {R}}. \\tag {3}", + "image_path": "087e40328280b68229bf1754061bc32d3bd598804fb02406b8c00295e0afaf1a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": "Given a proposal regions' sampling space " + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": " and a label space " + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": ", features " + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": " of proposal regions sampled from " + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": " in an image are fed into an MIL-based classifier " + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": ", to obtain the image-level class prediction score " + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in Y" + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": ". Therefore, ISOD can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 379, + 407, + 553, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 407, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 379, + 407, + 553, + 440 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\arg \\min } \\sum_ {k = 1} ^ {C} \\mathcal {L} \\left(f _ {\\theta} (\\mathbf {x}), \\mathbf {y} _ {k}\\right), \\tag {4}", + "image_path": "1d40bf3605660f6ec16410a53aca22dc24a30b596c68090c53c997a8b0d8399d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": " is the model's loss function, and " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": " is the label's " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": "-th category. Note that the sampling space " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": " is constant during training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 493, + 443, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 443, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 443, + 506 + ], + "type": "text", + "content": "4. The Proposed Method" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 513, + 506, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 506, + 525 + ], + "type": "text", + "content": "4.1. Problem Formulation and Overview" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "text", + "content": "Problem Formulation: PSOD provides minimal and yet distinctive position information for each object. An input image " + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "text", + "content": " targets, each associated with point-level labels " + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\{0,1\\}^C" + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "text", + "content": "-th target, its positive samples " + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^+" + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "text", + "content": " with labels " + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^+ \\in \\{0,1\\}^C" + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "text", + "content": " are generated using a static anchor strategy, while its negative samples " + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^-" + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "text", + "content": " with labels " + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i^-\\in \\{0\\}" + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "text", + "content": " are collected through random sampling on background. The negative sampling space is defined as the difference between the image " + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 530, + 555, + 650 + ], + "type": "text", + "content": " and the predicted pseudo boxes' regions (PB), denoted as \"diff(I, PB)\"." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 650, + 528, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 650, + 528, + 662 + ], + "spans": [ + { + "bbox": [ + 326, + 650, + 528, + 662 + ], + "type": "text", + "content": "The optimization for PSOD can be formulated as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 669, + 561, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 669, + 561, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 669, + 561, + 713 + ], + "type": "interline_equation", + "content": "\\underset {\\theta , x ^ {-} \\in \\operatorname {d i f f} (I, \\mathrm {P B})} {\\arg \\min } \\sum_ {i = 1} ^ {M} \\sum_ {k = 1} ^ {C} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i, k} ^ {+}\\right), \\mathbf {y} _ {i, k} ^ {+}\\right) + \\sum_ {j = 1} ^ {N} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {j} ^ {-}\\right), \\mathbf {y} _ {j} ^ {-}\\right). \\tag {5}", + "image_path": "63dad5665d401602b5664e995e218494cd1d816778aa2d5ce57e5b0d03da1fa6.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11709" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 178 + ], + "type": "image", + "image_path": "2e42270782033b04d9372b3b54c104559174b5be2fd72b8bc2801a390f720b7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 179, + 553, + 264 + ], + "blocks": [ + { + "bbox": [ + 57, + 179, + 553, + 264 + ], + "lines": [ + { + "bbox": [ + 57, + 179, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 57, + 179, + 553, + 264 + ], + "type": "image", + "image_path": "b1066da5f5cbd22c7ee236f5f456c45c1ed13591ec598f67bb2951de87e2a291.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 270, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 54, + 270, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 555, + 327 + ], + "type": "text", + "content": "Figure 3. The architecture of PointSR, comprising the TE Encoder, Coarse Pseudo-box Prediction and Pseudo-box Refinement. The training process spans " + }, + { + "bbox": [ + 54, + 270, + 555, + 327 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 270, + 555, + 327 + ], + "type": "text", + "content": " epochs, each consisting of two stages: Coarse Pseudo-box Prediction and Pseudo-box Refinement. The former generates coarse pseudo boxes and collects a set of negative samples by ranking their losses. The latter utilizes these negative samples and the positive samples from the previous stage to refine the pseudo boxes. The TE Encoder collects temporal information during each epoch to dynamically adjust candidate proposal generation at the beginning of the Coarse Pseudo-box Prediction stage." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 332, + 294, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 332, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 332, + 294, + 357 + ], + "type": "text", + "content": "where positive samples " + }, + { + "bbox": [ + 54, + 332, + 294, + 357 + ], + "type": "inline_equation", + "content": "x^{+}" + }, + { + "bbox": [ + 54, + 332, + 294, + 357 + ], + "type": "text", + "content": " of each target remain constant, and the negative sampling space " + }, + { + "bbox": [ + 54, + 332, + 294, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{diff}(I,\\mathrm{PB})" + }, + { + "bbox": [ + 54, + 332, + 294, + 357 + ], + "type": "text", + "content": " depends on PB." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 357, + 295, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 295, + 417 + ], + "type": "text", + "content": "To address the limitations of rigid sampling, we propose a self-regularized sampling mechanism, dynamically adjusting the positive " + }, + { + "bbox": [ + 54, + 357, + 295, + 417 + ], + "type": "inline_equation", + "content": "X^{+}" + }, + { + "bbox": [ + 54, + 357, + 295, + 417 + ], + "type": "text", + "content": " and negative " + }, + { + "bbox": [ + 54, + 357, + 295, + 417 + ], + "type": "inline_equation", + "content": "X^{-}" + }, + { + "bbox": [ + 54, + 357, + 295, + 417 + ], + "type": "text", + "content": " sampling spaces using temporal and informative constraints. According to Eq. (5), PointSR's optimization for PSOD is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 425, + 295, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 425, + 295, + 494 + ], + "spans": [ + { + "bbox": [ + 85, + 425, + 295, + 494 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\underset {\\theta , x ^ {+} \\in X ^ {+}, x ^ {-} \\in X ^ {-}} {\\arg \\min } \\sum_ {i = 1} ^ {M} \\sum_ {k = 1} ^ {C} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {i, k} ^ {+}\\right), \\mathbf {y} _ {i, k} ^ {+}\\right) + \\tag {6} \\\\ \\sum_ {j = 1} ^ {N} \\mathcal {L} \\left(f _ {\\theta} \\left(\\mathbf {x} _ {j} ^ {-}\\right), \\mathbf {y} _ {j} ^ {-}\\right). \\\\ \\end{array}", + "image_path": "6f725747116b690dd4a8d22e4a22e19ad5e3899881e4c9824364ccda87a94994.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 498, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 666 + ], + "type": "text", + "content": "Overview: PointSR employs a Self-regularized Sampling mechanism to optimize the pseudo box generation from point-level annotations. Fig. 3 shows that the architecture of PointSR consists of three main components: TE Encoder, Coarse Pseudo-box Prediction, and Pseudo-box Refinement. TE Encoder aggregates temporal information from predicted pseudo labels in each epoch and encodes it into an anchor prototype library. Coarse Pseudo-box Prediction uses the anchor prototype library for anchor adjustment and predicts coarse pseudo boxes using an MIL-based classifier. It also collects informative negative samples for refinement. Finally, Pseudo-box Refinement employs the MIL-based classifier to further refine pseudo boxes by suppressing candidate positive samples with low confidence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 672, + 225, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 225, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 225, + 685 + ], + "type": "text", + "content": "4.2. Temporal-Ensembling Encoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "The first step of predicting pseudo boxes is to create candidate boxes. Our approach predicts these candidate boxes in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 333, + 531, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 531, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 531, + 344 + ], + "type": "text", + "content": "an anchor style, done both statically and dynamically." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 553, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 553, + 429 + ], + "type": "text", + "content": "In the widely adopted static anchor strategy, anchors' width and height are pre-determined heuristically. Anchor shapes are set and remain constant throughout the entire training process. The static anchor strategy in Faster R-CNN [19] can be sub-optimal as it neglects the underlying distribution of real data, potentially leading to low-quality pseudo boxes and slow convergence during training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 430, + 554, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 633 + ], + "type": "text", + "content": "To address this limitation, we propose a learning-based dynamic anchor adjustment module, called \"TemporalEnsembling Encoder\" (TE Encoder), as illustrated in Fig. 3(d). TE Encoder focuses on encoding the ratio of objects' width to height instead of objects' scale because objects' aspect ratio is numerically more stable than object scales in drone-view images. Drone-view images show that object scales vary significantly depending on their distance from the drone-mounted camera, while an object's aspect ratio remains relatively stable, particularly in top-down aerial views in Fig 5. At the end of each training iteration, we can obtain high-quality pseudo boxes predicted in the refinement stage, serving as references to adjust anchors' shapes. In conclusion, very different from the pre-defined and fixed anchors, advanced anchors by TE Encoder are dynamically adjusted by the pseudo-box's aspect ratio derived from temporal information during training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "type": "text", + "content": "Specifically, TE Encoder aggregates pseudo-boxes' shapes for each category in each iteration and stores them as the category-aware prototype library, " + }, + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{C \\times 1}" + }, + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "type": "text", + "content": ". For Epoch " + }, + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "type": "text", + "content": "-th element " + }, + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "type": "inline_equation", + "content": "P_{c} \\in (0,1)" + }, + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 634, + 553, + 682 + ], + "type": "inline_equation", + "content": "c \\in \\{1,2,\\dots,C\\}" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "Setting: anchor shapes are preset as six scales (4, 8, 16, 32, 64, 128) with seven aspect ratios (1: 3, 1: 2, 1: 1.5, 1: 1, 1.5: 1, 2: 1, 3: 1)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11710" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "represents the ratio of width " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "w_{tj} \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": " to height " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "h_{tj} \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "-th object of category " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": " is updated iteratively as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 122, + 105, + 295, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 105, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 122, + 105, + 295, + 129 + ], + "type": "interline_equation", + "content": "P _ {c} ^ {\\prime} = \\lambda P _ {c} + (1 - \\lambda) \\frac {w _ {t j}}{h _ {t j}}. \\tag {7}", + "image_path": "bf2d56c374d11cccaa6df141c90f569ae45ec0f3095824e4796f476958b6c903.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "type": "text", + "content": " is used to balance the contributions of knowledge distilled in current and previous epochs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 168, + 222, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 222, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 222, + 180 + ], + "type": "text", + "content": "4.3. Coarse Pseudo-Box Prediction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 185, + 296, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 185, + 296, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 296, + 221 + ], + "type": "text", + "content": "We adjust anchors with the anchor prototype library for better coarse pseudo boxes and simultaneously collect a set of informative samples for the pseudo-box refinement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 221, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 353 + ], + "type": "text", + "content": "MIL-based Pseudo-Box Predictor: Given an adjusted bag of region proposals for each point label, a dual-stream structure is adopted as an MIL classifier to find the best box region in the bag to represent the target. As shown in Fig. 3(b), features of these proposals are extracted through a RoI [9] layer and two fully connected layers. Secondly, these features are fed into the dual-stream structure to estimate scores of each proposal in the same bag according to Eq. (1). Then, a bag's score is calculated by the sum of all its proposals. Finally, the Top-" + }, + { + "bbox": [ + 55, + 221, + 296, + 353 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 221, + 296, + 353 + ], + "type": "text", + "content": " boxes with the highest scores in the same bag are chosen for coarse pseudo boxes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 354, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 521 + ], + "type": "text", + "content": "Informative-Sample Collector (IS Collector): To capture informative negative samples with high uncertainty, we present IS Collector, which aims to collect negatives with high loss values. Specifically, negative samples randomly sampled on background with their corresponding losses estimated by the MIL-based classifier, are fed into our IS collector for the pseudo-box refinement, as shown in Fig. 3(e). Considering the high computational costs of processing all negative samples, Non-Maximum Suppression [16] is first applied for de-duplication, removing redundant negative samples with high similarity and retaining representative ones. Then, we rank the filtered negative samples in descending order of their losses and thus collect the most informative samples for pseudo-box refinement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 529, + 190, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 190, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 190, + 540 + ], + "type": "text", + "content": "4.4. Pseudo-Box Refinement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "type": "text", + "content": "The collected informative samples are used to suppress candidate positive samples with low confidence so as to refine pseudo boxes. As shown in Fig. 3(c), the model for pseudo-box refinement has a similar structure as the model for coarse pseudo-box prediction. Specifically, for positive sample generation, we update the aspect ratios of coarse pseudo boxes and jitter their center positions. Meanwhile, given the informative negative samples collected in coarse pseudo boxes, after the RoI layer, all features of samples are fed into the MIL classifier according to Eq. (1). Finally, the relatively accurate Top-" + }, + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "type": "text", + "content": " candidate proposals are chosen to be merged as the final predicted pseudo box." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Discussion: As shown in Fig. 4(b), numerous proposals by random sampling in the background are far from targets." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 381, + 129 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 381, + 129 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 381, + 129 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 381, + 129 + ], + "type": "image", + "image_path": "0b12f872133572c7a7b9125389ed278146c70d54867e6e5aa12b6cffed53d2aa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 138, + 355, + 148 + ], + "lines": [ + { + "bbox": [ + 345, + 138, + 355, + 148 + ], + "spans": [ + { + "bbox": [ + 345, + 138, + 355, + 148 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 381, + 72, + 487, + 136 + ], + "blocks": [ + { + "bbox": [ + 381, + 72, + 487, + 136 + ], + "lines": [ + { + "bbox": [ + 381, + 72, + 487, + 136 + ], + "spans": [ + { + "bbox": [ + 381, + 72, + 487, + 136 + ], + "type": "image", + "image_path": "ac7af72bf99260191e05b56c9e8aeebb5c718b94c5c06fa39816e934ea3cda69.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 138, + 441, + 147 + ], + "lines": [ + { + "bbox": [ + 432, + 138, + 441, + 147 + ], + "spans": [ + { + "bbox": [ + 432, + 138, + 441, + 147 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 487, + 72, + 551, + 129 + ], + "blocks": [ + { + "bbox": [ + 487, + 72, + 551, + 129 + ], + "lines": [ + { + "bbox": [ + 487, + 72, + 551, + 129 + ], + "spans": [ + { + "bbox": [ + 487, + 72, + 551, + 129 + ], + "type": "image", + "image_path": "388ed6638282936231a959a86173cedc143ed52a185281db238e1651b9fa3141.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 138, + 524, + 147 + ], + "lines": [ + { + "bbox": [ + 514, + 138, + 524, + 147 + ], + "spans": [ + { + "bbox": [ + 514, + 138, + 524, + 147 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 149, + 554, + 215 + ], + "lines": [ + { + "bbox": [ + 313, + 149, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 554, + 215 + ], + "type": "text", + "content": "Figure 4. The negative samples distribution of PointSR and Point-to-Box. The histogram (b) shows the number of negative samples for different generation methods, where the distance denotes the Euclidean distance between the center of negative samples (Background) and positive samples (Pedestrians). Two specific examples for PointSR (a) and Point-to-Box (c) are visualized." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 229, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 554, + 361 + ], + "type": "text", + "content": "Similarly, Fig. 4(c) shows partial negative samples, indicating that they provide few contributions to object detection. As for densely packed objects in drone-view images, edge information between objects and objects' neighborhood is crucial to distinguish each object. The random negative sampling in the background only generates proposals in a relatively stable sample space. While this sampling ensures diversity among negative samples, it may include less relevant ones, reducing learning efficiency. Additionally, since this method assigns equal weights to all negative samples, it can not adaptively select useful and informative negatives." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 365, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 533 + ], + "type": "text", + "content": "Fig. 4(a) presents that negative samples produced by our IS Collector are close enough to the decision boundary in feature space. Training with these negative samples is beneficial to the MIL classifier to recognize which candidate proposals should be assigned low scores and force the model to learn finer distinctions because these negatives contribute more significantly to the gradients. By combining with the IS Collector, the TE Encoder refines positive samples through anchor aspect ratio adjustments, enhancing the MIL classifier's ability to predict losses for negative samples more accurately. This improvement enables the IS Collector to better select high-loss negative samples. As a result, the collaboration between the two modules yields high-quality pseudo-boxes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 536, + 554, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 632 + ], + "type": "text", + "content": "The Analysis of Scalability: The time complexity of TE Encoder, " + }, + { + "bbox": [ + 313, + 536, + 554, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{O}(n)" + }, + { + "bbox": [ + 313, + 536, + 554, + 632 + ], + "type": "text", + "content": ", scales linearly with the number of the predicted pseudo boxes " + }, + { + "bbox": [ + 313, + 536, + 554, + 632 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 536, + 554, + 632 + ], + "type": "text", + "content": ". Whereas the IS Collector selects a few informative samples from around 400 negative samples randomly sampled from the background. Since this process is independent of " + }, + { + "bbox": [ + 313, + 536, + 554, + 632 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 536, + 554, + 632 + ], + "type": "text", + "content": " and the image resolution, PointSR's computational overhead is stable, even for high-resolution images or larger datasets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 652, + 408, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 408, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 408, + 663 + ], + "type": "text", + "content": "4.5. Loss Functions" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 673, + 553, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 553, + 697 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 553, + 697 + ], + "type": "text", + "content": "Inspired by the discriminative method [3], we design the MIL loss and negative loss for our PointSR training." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "The MIL Loss: We formulate the MIL loss " + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mil}" + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": " as a cross" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11711" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 232, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 232, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 232, + 85 + ], + "type": "text", + "content": "entropy loss for positive samples, given by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 95, + 295, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 95, + 295, + 130 + ], + "spans": [ + { + "bbox": [ + 65, + 95, + 295, + 130 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m i l} = - \\frac {1}{M} \\sum_ {j = 1} ^ {M} \\sum_ {k = 1} ^ {C} \\log \\left(\\hat {\\mathbf {y}} _ {j, k} ^ {\\mathbf {y} _ {j, k}} \\left(1 - \\hat {\\mathbf {y}} _ {j, k}\\right) ^ {\\left(1 - \\mathbf {y} _ {j, k}\\right)}\\right), \\tag {8}", + "image_path": "6cc280e450e6a85f4446e70610cdf2082d547d6b6e822f3c729b73c986c2fabd.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 138, + 295, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 138, + 295, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{j,k} \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 138, + 295, + 186 + ], + "type": "text", + "content": " denotes the one-hot category labels and " + }, + { + "bbox": [ + 55, + 138, + 295, + 186 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{jk}" + }, + { + "bbox": [ + 55, + 138, + 295, + 186 + ], + "type": "text", + "content": " is the predict scores generated by MIL-based predictor. The MIL loss is used for the correct prediction of the category and instance to which the proposal belongs to." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 186, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 222 + ], + "type": "text", + "content": "The Negative Loss: Furthermore, the negative loss " + }, + { + "bbox": [ + 55, + 186, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{neg}}" + }, + { + "bbox": [ + 55, + 186, + 296, + 222 + ], + "type": "text", + "content": " in the form of focal loss is used for model optimization, which is formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 229, + 295, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 229, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 80, + 229, + 295, + 262 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n e g} = - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\sum_ {k = 1} ^ {C} \\beta \\cdot (\\hat {\\mathbf {y}} _ {j, k}) ^ {\\gamma} \\log (1 - \\hat {\\mathbf {y}} _ {j, k}) \\tag {9}", + "image_path": "c987ffeed81d5a6f3a746793363583852c840c0fa5a20555b36fa71193b0a75e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 271, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 271, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 271, + 296, + 319 + ], + "type": "text", + "content": " is set to 0.2, the same in [13] and " + }, + { + "bbox": [ + 55, + 271, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 271, + 296, + 319 + ], + "type": "text", + "content": " is a weight to balance the contribution of different negative samples. The Negative loss in the Refinement stage is used to further suppress the background for pseudo-box refinement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 330, + 137, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 137, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 137, + 342 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 349, + 118, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 118, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 118, + 361 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 366, + 295, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 391 + ], + "type": "text", + "content": "We conducted experiments on three benchmark datasets: DroneVehicle [22], VisDrone [34] and UAVDT [5]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 391, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 462 + ], + "type": "text", + "content": "**DroneVehicle:** The DroneVehicle dataset includes 56,878 images, which are annotated with box-level annotations across five vehicle categories, i.e., Car, Truck, Bus, Van and Freight Car, covering various scenarios, including urban roads, residential areas, and parking lots captured under different lighting conditions from day to night." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 463, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 594 + ], + "type": "text", + "content": "VisDrone: The VisDrone dataset consists of 10,209 images with a resolution of " + }, + { + "bbox": [ + 55, + 463, + 295, + 594 + ], + "type": "inline_equation", + "content": "2,000 \\times 1" + }, + { + "bbox": [ + 55, + 463, + 295, + 594 + ], + "type": "text", + "content": ", 500 pixels and 542,000 instances across 10 common object categories in traffic scenarios, i.e., Pedestrian, People, Bicycle, Car, Van, Truck, Tricycle, Awning-tricycle, Bus and Motor. These images, captured from various urban scenes, present significant challenges due to the presence of numerous small objects, varying viewpoints, and heavy occlusions. To the hardware memory limitations, following PLUG [10], we cropped the original images into patches of " + }, + { + "bbox": [ + 55, + 463, + 295, + 594 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 463, + 295, + 594 + ], + "type": "text", + "content": " pixels with a 128-pixel overlap for training and inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 642 + ], + "type": "text", + "content": "UAVDT: The UAVDT dataset consists of 100 video sequences of about 80,000 representative frames with a resolution of " + }, + { + "bbox": [ + 55, + 594, + 295, + 642 + ], + "type": "inline_equation", + "content": "1,024 \\times 540" + }, + { + "bbox": [ + 55, + 594, + 295, + 642 + ], + "type": "text", + "content": " pixels and 0.84 million box-level annotations across three vehicle categories: Car, Truck, Bus." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "To accurately reproduce the biases inherent in manual annotation, we employed Rectified Gaussian Distribution (RG) [30] with central ellipse constraints, assuming that the point distribution of manual annotation approximately follows an RG distribution. Therefore, for our experiments, the three datasets with point-level annotations are employed" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "for evaluating our approach, referred to as \"DroneVehicle (Point-Supervision)\", \"VisDrone (Point-Supervision)\", and \"UAVDT (Point-Supervision)\", respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 117, + 525, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 525, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 525, + 129 + ], + "type": "text", + "content": "5.2. Implementation and Evaluation Metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 135, + 555, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 555, + 231 + ], + "type": "text", + "content": "Implementation Details: We implemented our proposed PointSR based on the MMDetection toolbox on an NVIDIA RTX 4090 GPU. The Stochastic Gradient Descent (SGD) algorithm was used to optimize with a " + }, + { + "bbox": [ + 313, + 135, + 555, + 231 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 313, + 135, + 555, + 231 + ], + "type": "text", + "content": " training schedule. The learning rate was set to 0.02 and decayed by a factor of 0.1 at the 24th and 33rd epochs, respectively. All comparison models were configured with a ResNet-50 [8] backbone pre-trained on ImageNet [4]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 231, + 556, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 556, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 556, + 363 + ], + "type": "text", + "content": "Evaluation Metrics: We evaluated our method from three aspects: the quality of predicted pseudo boxes, the detection performance and the model's robustness. The mean Intersection over Union (mIoU) between generated pseudo boxes and ground-truth boxes was utilized to evaluate the quality of predicted pseudo boxes, where mIoU closer to 1 indicates higher accuracy. The Average Precision " + }, + { + "bbox": [ + 313, + 231, + 556, + 363 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_{50})" + }, + { + "bbox": [ + 313, + 231, + 556, + 363 + ], + "type": "text", + "content": " [6] was utilized to evaluate the models' detection performance. Finally, the variances of AP across different point-label distributions were used to measure models' robustness, with lower variance indicating higher robustness." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 371, + 493, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 493, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 493, + 384 + ], + "type": "text", + "content": "5.3. Comparison with SOTA Methods" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 389, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 521 + ], + "type": "text", + "content": "We compared the performance of a detector trained using pseudo boxes generated by our PointSR and five SOTA point-supervised DroneDet methods: PLUG [10], PointOBB [15], Point2Rbox [32], PointOBBv2 [18] and Point-to-Box [3]. In these experiments, pseudo boxes were first generated by our PointSR and SOTA methods. These pseudo boxes were then utilized to train a classic detector, Faster-RCNN [19], to evaluate its detection performance and robustness. In addition, a Faster-RCNN detector was trained with ground-truth box-level annotations, which detection performance was referred to as the upper bound." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": "The Detection Performance: Table 1 shows that the detector trained with pseudo boxes generated by our PointSR achieves superior detection performance, outperforming that trained with pseudo boxes generated by SOTA methods significantly on all datasets. Specifically, our model improved " + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": " on DroneVehicle (Point-Supervision) dataset, " + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": " on VisDrone (Point-Supervision) dataset and " + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": " on UAVDT (Point-Supervision) dataset, compared to Point-to-Box [3]. Besides, it is worth mentioning that the detection results of our model trained with point-level annotations are better than the results of Faster-RCNN trained with box-level annotations in some cases in Fig. 5." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "The Quality of the Predicted Pseudo-Boxes: We use mIoU between the generated pseudo boxes and ground-truth boxes to evaluate the quality of the predicted pseudo boxes, where mIoU closer to 1 indicates higher accuracy." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11712" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 108, + 553, + 216 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Comparisons of per-class detection accuracy " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_{50})" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " across different object classes, average " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": ", and mIoU for a detector trained using pseudo boxes generated by our PointSR and SOTA methods on VisDrone, DroneVehicle and UAVDT. \"Faster R-CNN\" indicates the accuracy obtained with the detector trained with ground-truth box-level annotations, serving as the upper bound for performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 108, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 56, + 108, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 553, + 216 + ], + "type": "table", + "html": "
MethodVisDroneDroneVehicleUAVDT
Ped.Peo.Bic.CarVanTru.Tri.Awn.BusMot.AP50mIoUCarTru.BusVanFre.AP50mIoUCarTru.BusAP50mIoU
PLUG [10]0.63.44.73.40.24.50.70.323.65.34.726.91.06.10.28.02.33.529.04.58.23.85.531.1
Point2Rbox [32]0.30.29.122.315.315.311.59.915.81.410.1-0.03.675.718.928.025.2-15.724.119.219.7-
PointOBB [15]25.613.620.834.612.430.732.930.456.512.427.056.218.37.06.034.510.115.277.529.230.816.825.672.6
PointOBBv2 [18]33.24.316.169.257.431.632.123.232.823.832.464.817.119.734.229.222.824.680.116.628.837.227.577.5
Point-to-Box [3]40.24.423.162.953.130.930.922.828.419.631.766.268.925.524.233.024.735.386.045.535.818.533.395.3
Faster R-CNN [19]53.730.930.072.257.644.231.325.035.837.241.8-89.766.092.253.352.770.1-66.955.759.660.7-
PointSR (Ours)43.45.826.265.653.531.529.722.444.727.735.075.782.825.133.542.229.142.588.950.140.825.438.796.8
", + "image_path": "ac15fad5c68f108e5db8238e2c05baf4b667573c78e60876be0ee07e51c77b7a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 257, + 547, + 338 + ], + "blocks": [ + { + "bbox": [ + 55, + 220, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 255 + ], + "type": "text", + "content": "Table 2. Comparison of average precision " + }, + { + "bbox": [ + 55, + 220, + 555, + 255 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_{50})" + }, + { + "bbox": [ + 55, + 220, + 555, + 255 + ], + "type": "text", + "content": " and variance for a detector trained using pseudo boxes generated by our PointSR and SOTA methods across different point-label distributions. PointSR consistently demonstrates lower variance on both datasets, exhibiting greater stability and robustness to annotation perturbations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 257, + 547, + 338 + ], + "lines": [ + { + "bbox": [ + 61, + 257, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 61, + 257, + 547, + 338 + ], + "type": "table", + "html": "
MethodPublicationVisDroneDroneVehicle
U0×LU0.1×LU0.3×LU0.5×LVariance Of AP50U0×LU0.1×LU0.3×LU0.5×LVariance Of AP50
PLUG [10]TGRS'241.54.315.517.848.911.13.930.132.5209.1
Point2Rbox [32]CVPR'247.69.98.39.10.74215.224.117.618.310.61
PointOBB [15]CVPR'2424.328.517.88.556.9113.415.212.68.95.269
Point-to-Box [3]ECCV'2231.632.531.130.10.75234.835.532.934.00.935
PointSR (Ours)N/A34.634.833.933.70.21241.142.041.340.90.171
", + "image_path": "a31108634041c8e16f28e1b27dc685d4d14e3ee52cceb3a8b8ef9d42ed7485da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 76, + 371, + 276, + 435 + ], + "blocks": [ + { + "bbox": [ + 55, + 346, + 295, + 368 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 368 + ], + "type": "text", + "content": "Table 3. The impact of the TE Encoder and Informative Sampling (IS) on the baseline Point-to-Box [3], evaluated on DroneVehicle." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 371, + 276, + 435 + ], + "lines": [ + { + "bbox": [ + 76, + 371, + 276, + 435 + ], + "spans": [ + { + "bbox": [ + 76, + 371, + 276, + 435 + ], + "type": "table", + "html": "
TEEISAP50mIoU [%]
Baseline××35.2686.0
Baseline + TEE✓×36.1886.7
Baseline + IS×✓40.8788.1
Baseline + TEE + IS✓✓42.5388.9
", + "image_path": "f51b1844acf38fa4711b569f42a2f4f737eb9f3b6abe4fc1f10e3fb98f8b2baf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "content": "Tables 1 also show that PointSR achieves better mIoUs than the SOTA methods on both datasets. Specifically, our model improves mIoU by " + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "content": " on Visdrone, " + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "content": " on DroneVehicle and " + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "content": " on UAVDT compared to Point-to-Box [3]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 497, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 497, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 497, + 295, + 616 + ], + "type": "text", + "content": "The visualization in Fig. 6 shows that our predicted pseudo boxes accurately localize objects in challenging drone-view images by leveraging the precise location information of point annotations and can distinguish dense objects in complex scenes. In contrast, Point-to-Box often produces loose pseudo boxes due to its limited sampling strategy for MIL training, and PLUG occasionally predicts oversized boxes due to outliers in the generative process. These results confirm the high quality of pseudo boxes generated by our model with point-level annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": "The Robustness of the Model: To evaluate the model's robustness, we use the variance of " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": " across different point distributions. Specially, we assume manually annotated points follow a uniform distribution " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "x\\sim U(-a,a)" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "a = \\alpha \\times L" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\alpha = \\{0,0.1,0.3,0.5\\}" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": " is the maximum value for the width and height of a bounding box. " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "U_{0\\times L}" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "U_{0.1\\times L}" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "U_{0.3\\times L}" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "U_{0.5\\times L}" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": " denote the manually annotated points " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "x\\sim U(-\\alpha \\times L,\\alpha \\times L)" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": " with offsets of 0, " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 347, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 396 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 347, + 555, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 347, + 555, + 396 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 347, + 555, + 396 + ], + "type": "text", + "content": ", respectively. As shown in Table 2, our model has achieved a variance of " + }, + { + "bbox": [ + 313, + 347, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 347, + 555, + 396 + ], + "type": "text", + "content": " 0.171 on the DroneVehicle dataset, compared to 0.935 of Point-to-Box [3], indicating superior robustness to annotation perturbation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 405, + 416, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 416, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 416, + 417 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 424, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 471 + ], + "type": "text", + "content": "We conducted ablation studies on the Drone Vehicle (Point-Supervision) dataset using Point-to-Box [3] as the baseline to validate the effectiveness of the two key components in PointSR: the TE Encoder and IS Collector." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "text", + "content": "The Effectiveness of the TE Encoder: We conducted a quantitative evaluation to verify the effectiveness of the TE Encoder in our PointSR. The baseline with TE Encoder is referred to as \"Baseline + TEE\". Table 3 shows that our proposed module improves " + }, + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "inline_equation", + "content": "0.92\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "text", + "content": " and mIoU by " + }, + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "text", + "content": ", compared with the vanilla baseline model. Moreover, TE Encoder facilitates stable and efficient training, in Fig. 7(a), where the orange curve (Baseline + TEE) descends faster and is smoother compared with the blue curve, especially in the early stage of training, indicating that our Anchor Learning model achieves stable training and faster convergence. In addition, Fig. 7(b) further demonstrates that \"Baseline + TEE\" achieves better detection performance. Finally, Table 4 shows the performance of Faster R-CNN trained using PointSR with different " + }, + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "text", + "content": " settings, where PointSR with " + }, + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 313, + 472, + 556, + 675 + ], + "type": "text", + "content": " achieves the best detection performance. These results verify the effectiveness of the TE Encoder." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 359, + 693, + 509, + 722 + ], + "blocks": [ + { + "bbox": [ + 318, + 680, + 549, + 691 + ], + "lines": [ + { + "bbox": [ + 318, + 680, + 549, + 691 + ], + "spans": [ + { + "bbox": [ + 318, + 680, + 549, + 691 + ], + "type": "text", + "content": "Table 4. Impact of different " + }, + { + "bbox": [ + 318, + 680, + 549, + 691 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 318, + 680, + 549, + 691 + ], + "type": "text", + "content": " values on " + }, + { + "bbox": [ + 318, + 680, + 549, + 691 + ], + "type": "inline_equation", + "content": "{\\mathrm{{AP}}}_{50}" + }, + { + "bbox": [ + 318, + 680, + 549, + 691 + ], + "type": "text", + "content": " in DroneVehicle." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 359, + 693, + 509, + 722 + ], + "lines": [ + { + "bbox": [ + 359, + 693, + 509, + 722 + ], + "spans": [ + { + "bbox": [ + 359, + 693, + 509, + 722 + ], + "type": "table", + "html": "
λ0.10.30.50.70.9
AP50 [%]34.240.542.534.933.9
", + "image_path": "846a5eb1783baefce740347df24195d7e8419bf241b1ddc35def3c551db9fbdb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11713" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 136, + 137 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 136, + 137 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 136, + 137 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 136, + 137 + ], + "type": "image", + "image_path": "2ab5bd1e68a37f9fac28eb85b7fe29b67fb21168c48516db8c48e9fd4993b7d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 138, + 70, + 205, + 137 + ], + "blocks": [ + { + "bbox": [ + 138, + 70, + 205, + 137 + ], + "lines": [ + { + "bbox": [ + 138, + 70, + 205, + 137 + ], + "spans": [ + { + "bbox": [ + 138, + 70, + 205, + 137 + ], + "type": "image", + "image_path": "842e2739c5726a3597088ede6ec2bc4fae4ae9566c2b503187565406eb8752dd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 205, + 70, + 271, + 137 + ], + "blocks": [ + { + "bbox": [ + 205, + 70, + 271, + 137 + ], + "lines": [ + { + "bbox": [ + 205, + 70, + 271, + 137 + ], + "spans": [ + { + "bbox": [ + 205, + 70, + 271, + 137 + ], + "type": "image", + "image_path": "3fee694ba271e870394dbb783cbbedd4e4f9d1427d4e13232b7478a2d0ccf0a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 272, + 70, + 339, + 137 + ], + "blocks": [ + { + "bbox": [ + 272, + 70, + 339, + 137 + ], + "lines": [ + { + "bbox": [ + 272, + 70, + 339, + 137 + ], + "spans": [ + { + "bbox": [ + 272, + 70, + 339, + 137 + ], + "type": "image", + "image_path": "2c6aed0d7d051948b739b514ad9445f14ad50e5e966cbaabe5ae1d1b0aa34a09.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 340, + 70, + 406, + 137 + ], + "blocks": [ + { + "bbox": [ + 340, + 70, + 406, + 137 + ], + "lines": [ + { + "bbox": [ + 340, + 70, + 406, + 137 + ], + "spans": [ + { + "bbox": [ + 340, + 70, + 406, + 137 + ], + "type": "image", + "image_path": "19b0f7f2bcac7ca0a6f5480ff581bef80d20fe23a34194da6d58815bd6bd035d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 408, + 70, + 474, + 137 + ], + "blocks": [ + { + "bbox": [ + 408, + 70, + 474, + 137 + ], + "lines": [ + { + "bbox": [ + 408, + 70, + 474, + 137 + ], + "spans": [ + { + "bbox": [ + 408, + 70, + 474, + 137 + ], + "type": "image", + "image_path": "2d7ff54c9952ae1433f159d6eab948c5dea8a890650e0d04e07ebed100e594c9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 474, + 70, + 541, + 137 + ], + "blocks": [ + { + "bbox": [ + 474, + 70, + 541, + 137 + ], + "lines": [ + { + "bbox": [ + 474, + 70, + 541, + 137 + ], + "spans": [ + { + "bbox": [ + 474, + 70, + 541, + 137 + ], + "type": "image", + "image_path": "193dbc5dc579c5ef93983df44dabb5f6591223a5a1405784a732231fc631bd72.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 70, + 138, + 136, + 205 + ], + "blocks": [ + { + "bbox": [ + 70, + 138, + 136, + 205 + ], + "lines": [ + { + "bbox": [ + 70, + 138, + 136, + 205 + ], + "spans": [ + { + "bbox": [ + 70, + 138, + 136, + 205 + ], + "type": "image", + "image_path": "de0e591a03e899085ec4cb1c1ca60a35dd6aa0c5bb176949035417608204ce91.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 137, + 138, + 204, + 205 + ], + "blocks": [ + { + "bbox": [ + 137, + 138, + 204, + 205 + ], + "lines": [ + { + "bbox": [ + 137, + 138, + 204, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 138, + 204, + 205 + ], + "type": "image", + "image_path": "15ebc8de4d615f4298ae1809deb56d3f2acf50c55d7d6c4fab1be7cb983c6599.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 205, + 138, + 271, + 205 + ], + "blocks": [ + { + "bbox": [ + 205, + 138, + 271, + 205 + ], + "lines": [ + { + "bbox": [ + 205, + 138, + 271, + 205 + ], + "spans": [ + { + "bbox": [ + 205, + 138, + 271, + 205 + ], + "type": "image", + "image_path": "45f50f650bc813487a0fdf273a3cccb8a9720e1f6b8b12e720aa498ef631aaed.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 272, + 138, + 339, + 205 + ], + "blocks": [ + { + "bbox": [ + 272, + 138, + 339, + 205 + ], + "lines": [ + { + "bbox": [ + 272, + 138, + 339, + 205 + ], + "spans": [ + { + "bbox": [ + 272, + 138, + 339, + 205 + ], + "type": "image", + "image_path": "ebbd4f1405439f97eef5400133c70fef5dcc3614b63dd812d0b0be8f014e2104.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 340, + 138, + 406, + 205 + ], + "blocks": [ + { + "bbox": [ + 340, + 138, + 406, + 205 + ], + "lines": [ + { + "bbox": [ + 340, + 138, + 406, + 205 + ], + "spans": [ + { + "bbox": [ + 340, + 138, + 406, + 205 + ], + "type": "image", + "image_path": "eadcc5a9674d5551c82e162e15f790f673d20caeebad517115f806787e6b7445.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 407, + 138, + 474, + 205 + ], + "blocks": [ + { + "bbox": [ + 407, + 138, + 474, + 205 + ], + "lines": [ + { + "bbox": [ + 407, + 138, + 474, + 205 + ], + "spans": [ + { + "bbox": [ + 407, + 138, + 474, + 205 + ], + "type": "image", + "image_path": "df3beca3190aa231e17e22992f573b72bea68251efd4a8b02e36479609da4f83.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 474, + 138, + 541, + 205 + ], + "blocks": [ + { + "bbox": [ + 474, + 138, + 541, + 205 + ], + "lines": [ + { + "bbox": [ + 474, + 138, + 541, + 205 + ], + "spans": [ + { + "bbox": [ + 474, + 138, + 541, + 205 + ], + "type": "image", + "image_path": "373ea0f28963dde991576207a441795b4f99c94d00e023ad4525dd9590429d99.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 70, + 205, + 136, + 272 + ], + "blocks": [ + { + "bbox": [ + 70, + 205, + 136, + 272 + ], + "lines": [ + { + "bbox": [ + 70, + 205, + 136, + 272 + ], + "spans": [ + { + "bbox": [ + 70, + 205, + 136, + 272 + ], + "type": "image", + "image_path": "49346a8337214009e048e2780b51c7f38c55ff90cf7d14379ae691f31b2295f0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 275, + 115, + 285 + ], + "lines": [ + { + "bbox": [ + 89, + 275, + 115, + 285 + ], + "spans": [ + { + "bbox": [ + 89, + 275, + 115, + 285 + ], + "type": "text", + "content": "PLUG" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 137, + 205, + 204, + 272 + ], + "blocks": [ + { + "bbox": [ + 137, + 205, + 204, + 272 + ], + "lines": [ + { + "bbox": [ + 137, + 205, + 204, + 272 + ], + "spans": [ + { + "bbox": [ + 137, + 205, + 204, + 272 + ], + "type": "image", + "image_path": "7800909db44a5597fe4b9c721cf8ceeb94609423177a6f3a9838ea772886406a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 275, + 192, + 285 + ], + "lines": [ + { + "bbox": [ + 148, + 275, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 148, + 275, + 192, + 285 + ], + "type": "text", + "content": "Point2Rbox" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 205, + 205, + 271, + 272 + ], + "blocks": [ + { + "bbox": [ + 205, + 205, + 271, + 272 + ], + "lines": [ + { + "bbox": [ + 205, + 205, + 271, + 272 + ], + "spans": [ + { + "bbox": [ + 205, + 205, + 271, + 272 + ], + "type": "image", + "image_path": "d618c06ed51f814bbb66ba92c2a41aa015c92c14df6e21e39a5fdbb578123247.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 276, + 257, + 285 + ], + "lines": [ + { + "bbox": [ + 217, + 276, + 257, + 285 + ], + "spans": [ + { + "bbox": [ + 217, + 276, + 257, + 285 + ], + "type": "text", + "content": "PointOBB" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 272, + 205, + 339, + 272 + ], + "blocks": [ + { + "bbox": [ + 272, + 205, + 339, + 272 + ], + "lines": [ + { + "bbox": [ + 272, + 205, + 339, + 272 + ], + "spans": [ + { + "bbox": [ + 272, + 205, + 339, + 272 + ], + "type": "image", + "image_path": "78311976e08c06e25af1eccfeaf10319b213bc33fb87f57abd132a6e24990527.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 276, + 328, + 285 + ], + "lines": [ + { + "bbox": [ + 280, + 276, + 328, + 285 + ], + "spans": [ + { + "bbox": [ + 280, + 276, + 328, + 285 + ], + "type": "text", + "content": "Point-to-Box" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 340, + 205, + 406, + 272 + ], + "blocks": [ + { + "bbox": [ + 340, + 205, + 406, + 272 + ], + "lines": [ + { + "bbox": [ + 340, + 205, + 406, + 272 + ], + "spans": [ + { + "bbox": [ + 340, + 205, + 406, + 272 + ], + "type": "image", + "image_path": "e4ea27984c2b42950187254adb2ce9975cf3c006901e38dd526532bc9bccc170.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 276, + 396, + 285 + ], + "lines": [ + { + "bbox": [ + 346, + 276, + 396, + 285 + ], + "spans": [ + { + "bbox": [ + 346, + 276, + 396, + 285 + ], + "type": "text", + "content": "Faster-RCNN" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 407, + 205, + 474, + 272 + ], + "blocks": [ + { + "bbox": [ + 407, + 205, + 474, + 272 + ], + "lines": [ + { + "bbox": [ + 407, + 205, + 474, + 272 + ], + "spans": [ + { + "bbox": [ + 407, + 205, + 474, + 272 + ], + "type": "image", + "image_path": "947e3b7c408d04e477740e93378035313dc19aa8d0bc59ed667f4ddbbb342c90.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 276, + 465, + 286 + ], + "lines": [ + { + "bbox": [ + 411, + 276, + 465, + 286 + ], + "spans": [ + { + "bbox": [ + 411, + 276, + 465, + 286 + ], + "type": "text", + "content": "PointSR(Ours)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 474, + 205, + 541, + 272 + ], + "blocks": [ + { + "bbox": [ + 474, + 205, + 541, + 272 + ], + "lines": [ + { + "bbox": [ + 474, + 205, + 541, + 272 + ], + "spans": [ + { + "bbox": [ + 474, + 205, + 541, + 272 + ], + "type": "image", + "image_path": "66ec264911bcbb0f672f0a5d6baeec402ebb848e0f45fd1b1aec803c273774a8.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 275, + 512, + 285 + ], + "lines": [ + { + "bbox": [ + 499, + 275, + 512, + 285 + ], + "spans": [ + { + "bbox": [ + 499, + 275, + 512, + 285 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 66, + 327, + 286, + 503 + ], + "blocks": [ + { + "bbox": [ + 55, + 288, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 555, + 321 + ], + "type": "text", + "content": "Figure 5. Visualization of detection results from a detector trained with pseudo boxes generated by PointSR and SOTA methods. Ground-truth bounding boxes are shown in green, correct detections in yellow, and misclassified or missed targets in red. For a better view, detection results in the last row are highlighted by cyan rectangular boxes. Best viewed in color and zoomed-in mode." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 327, + 286, + 503 + ], + "lines": [ + { + "bbox": [ + 66, + 327, + 286, + 503 + ], + "spans": [ + { + "bbox": [ + 66, + 327, + 286, + 503 + ], + "type": "image", + "image_path": "f72046695e71cdb2af191a7f3aa0a63d595adaf0135dfaf762b27c170dec981b.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 58, + 545, + 175, + 617 + ], + "blocks": [ + { + "bbox": [ + 58, + 545, + 175, + 617 + ], + "lines": [ + { + "bbox": [ + 58, + 545, + 175, + 617 + ], + "spans": [ + { + "bbox": [ + 58, + 545, + 175, + 617 + ], + "type": "image", + "image_path": "e049363795044915f1b0457cd1491b3ac8dc0d07346ee85c154015fcf99db252.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 619, + 129, + 628 + ], + "lines": [ + { + "bbox": [ + 118, + 619, + 129, + 628 + ], + "spans": [ + { + "bbox": [ + 118, + 619, + 129, + 628 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 177, + 544, + 294, + 618 + ], + "blocks": [ + { + "bbox": [ + 55, + 506, + 295, + 540 + ], + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 540 + ], + "type": "text", + "content": "Figure 6. Visualization of pseudo boxes (pink) predicted by PointSR and SOTA methods. Ground-truth boxes are shown in green. Best viewed in color and zoomed-in mode." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 177, + 544, + 294, + 618 + ], + "lines": [ + { + "bbox": [ + 177, + 544, + 294, + 618 + ], + "spans": [ + { + "bbox": [ + 177, + 544, + 294, + 618 + ], + "type": "image", + "image_path": "5a53ad46899b8d191dbd8e38ec952c2af9cf29d76cfba86a83c2068ff7b53369.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 619, + 249, + 628 + ], + "lines": [ + { + "bbox": [ + 238, + 619, + 249, + 628 + ], + "spans": [ + { + "bbox": [ + 238, + 619, + 249, + 628 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 629, + 295, + 662 + ], + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 662 + ], + "type": "text", + "content": "Figure 7. The loss and " + }, + { + "bbox": [ + 55, + 629, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 55, + 629, + 295, + 662 + ], + "type": "text", + "content": " curves during the whole training stage. The blue curves represent baseline training, while the orange curves show baseline training equipped with our TE Encoder." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The Effectiveness of the IS Collector: We evaluated the effectiveness of the Informative Sampling Collector in our PointSR. Table 3 shows that the baseline with our Informative Sampling module, referred to as \"Baseline + IS\"," + } + ] + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 317, + 328, + 435, + 400 + ], + "blocks": [ + { + "bbox": [ + 317, + 328, + 435, + 400 + ], + "lines": [ + { + "bbox": [ + 317, + 328, + 435, + 400 + ], + "spans": [ + { + "bbox": [ + 317, + 328, + 435, + 400 + ], + "type": "image", + "image_path": "1b57e286ba27eb69786d64ee192dbbdac30c29ce5a9874c3afd44af10d63ffee.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 401, + 389, + 410 + ], + "lines": [ + { + "bbox": [ + 375, + 401, + 389, + 410 + ], + "spans": [ + { + "bbox": [ + 375, + 401, + 389, + 410 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 436, + 328, + 554, + 400 + ], + "blocks": [ + { + "bbox": [ + 436, + 328, + 554, + 400 + ], + "lines": [ + { + "bbox": [ + 436, + 328, + 554, + 400 + ], + "spans": [ + { + "bbox": [ + 436, + 328, + 554, + 400 + ], + "type": "image", + "image_path": "a2db59f364d79f76298aa0506db13f17cfc78c846879521a3e68c4ae9cedf095.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 402, + 511, + 411 + ], + "lines": [ + { + "bbox": [ + 500, + 402, + 511, + 411 + ], + "spans": [ + { + "bbox": [ + 500, + 402, + 511, + 411 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "text", + "content": "Figure 8. The " + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "text", + "content": " and IoU curves for varying Sampling Ratio " + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "text", + "content": ". improves " + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "inline_equation", + "content": "5.26\\%" + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "text", + "content": " and mIoU by " + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "text", + "content": ". These experimental results confirm the effectiveness of our IS Collector. In addition, we also analyzed the impact of the ratio of selected negative samples to whole negative samples " + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "text", + "content": ", denoted as \"Sampling Ratio\", on detection performance. Fig. 8 visualizes that changes in " + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "text", + "content": " significantly affect the detection performance of our PiontSR in terms of " + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 411, + 555, + 521 + ], + "type": "text", + "content": " and mIoU, suggesting its critical role in optimizing results." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Towards generating high-quality pseudo boxes for PSOD from drone-view images, in this paper, we proposed an effective and robust self-regularized point-supervision approach, PointSR, which leverages a Temporal-Ensemble Encoder for stable pseudo-box generation and an Informative Sampling Collector for efficient pseudo-box refinement. The two processes mutually benefit each other, facilitating the generation of self-regularized high-quality pseudo boxes. Experimental results on benchmark datasets have demonstrated that detectors trained with pseudo boxes generated by PointSR significantly outperform those generated by SOTAs over " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "2.6\\% \\sim 7.2\\%" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " in DroneDet and gain the robust results under perturbations of the manually annotated labels, significantly advancing the applicability of PSOD." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11714" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 161 + ], + "type": "text", + "content": "Acknowledgments: This work was supported by the National Natural Science Foundation of China (Grant No. 62402370, 62301397), the Natural Science Basic Research Program of Shaanxi (Grant No. 2024JC-YBQN-0728), the Guangzhou Basic and Applied Basic Research Foundation (Grant No. 2023A04J1742). Additionally, we would like to thank Zhiwen Wei for his helpful discussions and insightful suggestions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 205, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 205, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 294, + 248 + ], + "type": "text", + "content": "[1] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2854, 2016. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 251, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 294, + 294 + ], + "type": "text", + "content": "[2] Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. Just a Hint: Point-Supervised Camouflaged Object Detection. In European Conference on Computer Vision, pages 332-348. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 296, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 361 + ], + "type": "text", + "content": "[3] Pengfei Chen, Xuehui Yu, Xumeng Han, Najmul Hassan, Kai Wang, Jiachen Li, Jian Zhao, Humphrey Shi, Zhenjun Han, and Qixiang Ye. Point-to-Box Network for Accurate Object Detection via Single Point Supervision. In European Conference on Computer Vision, pages 51-67. Springer, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 294, + 418 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-scale Hierarchical Image Database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 420, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 475 + ], + "type": "text", + "content": "[5] Dawei Du, Yuankai Qi, Hongyang Yu, Yifan Yang, Kaiwen Duan, Guorong Li, Weigang Zhang, Qingming Huang, and Qi Tian. The Unmanned Aerial Vehicle Benchmark: Object Detection and Tracking. In Proceedings of the European Conference on Computer Vision, pages 370-386, 2018. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 476, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 294, + 529 + ], + "type": "text", + "content": "[6] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes Challenge: A Retrospective. International Journal of Computer Vision, 111:98-136, 2015. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "type": "text", + "content": "[7] Xiaoxu Feng, Xiwen Yao, Gong Cheng, and Junwei Han. Weakly Supervised Rotation-Invariant Aerial Object Detection Network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14146-14155, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 294, + 633 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 635, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 294, + 677 + ], + "type": "text", + "content": "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In Proceedings of the IEEE International Conference on Computer Vision, pages 2961-2969, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[10] Shitian He, Huanxin Zou, Yingqian Wang, Boyang Li, Xu Cao, and Ning Jing. Learning Remote Sensing Object Detection With Single Point Supervision. IEEE Transactions" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "on Geoscience and Remote Sensing, 62:1-16, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "text", + "content": "[11] Boyang Li, Yingqian Wang, Longguang Wang, Fei Zhang, Ting Liu, Zaiping Lin, Wei An, and Yulan Guo. Monte Carlo Linear Clustering with Single-Point Supervision is Enough for Infrared Small Target Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1009–1019, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "text", + "content": "[12] Jing Li, Junsong Fan, and Zhaoxiang Zhang. Point-Supervised Panoptic Segmentation via Estimating Pseudo Labels from Learnable Distance. In European Conference on Computer Vision, pages 95-112. Springer, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "text", + "content": "[13] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal Loss for Dense Object Detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980-2988, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 251, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 294 + ], + "type": "text", + "content": "[14] Junyan Lu, Qinglei Hu, Ruifei Zhu, Yali Wei, and Tie Li. AFWS: Angle-Free Weakly-Supervised Rotating Object Detection for Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 295, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 555, + 348 + ], + "type": "text", + "content": "[15] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. PointOBB: Learning Oriented Object Detection via Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16730-16740, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "text", + "content": "[16] Alexander Neubeck and Luc Van Gool. Efficient Nonmaximum Suppression. In 18th International Conference on Pattern Recognition, pages 850-855. IEEE, 2006. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 384, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 384, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 553, + 437 + ], + "type": "text", + "content": "[17] Dim P Papadopoulos, Jasper RR Uijlings, Frank Keller, and Vittorio Ferrari. Training Object Class Detectors with Click Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6374-6383, 2017. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "type": "text", + "content": "[18] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection. In The Thirteenth International Conference on Learning Representations, 2025. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 494, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 547 + ], + "type": "text", + "content": "[19] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2016. 4, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "type": "text", + "content": "[20] Tianqi Ren, Qiu Shen, Ying Fu, and Shaodi You. Point-Supervised Semantic Segmentation of Natural Scenes via Hyperspectral Imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1357-1367, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 604, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 658 + ], + "type": "text", + "content": "[21] Zhongzheng Ren, Zhiding Yu, Xiaodong Yang, Ming-Yu Liu, Alexander G Schwing, and Jan Kautz. " + }, + { + "bbox": [ + 317, + 604, + 553, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{UFO}^2" + }, + { + "bbox": [ + 317, + 604, + 553, + 658 + ], + "type": "text", + "content": ": A Unified Framework Towards Omni-Supervised Object Detection. In European Conference on Computer Vision, pages 288-313. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[22] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Drone-Based RGB-Infrared Cross-Modality Vehicle Detection via Uncertainty-Aware Learning. IEEE Transactions on Circuits and Systems for Video Technology, 32(10):6700-6713, 2022. 2, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11715" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[23] Zhiwen Tan, Zhiguo Jiang, Chen Guo, and Haopeng Zhang. WSODet: A Weakly Supervised Oriented Detector for Aerial Object Detection. IEEE Transactions on Geoscience and Remote Sensing, 61:1-12, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[24] Peng Tang, Xinggang Wang, Song Bai, Wei Shen, Xiang Bai, Wenyu Liu, and Alan Yuille. PCL: Proposal Cluster Learning for Weakly Supervised Object Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(1): 176-191, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[25] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective Search for Object Recognition. International Journal of Computer Vision, 104: 154-171, 2013. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "text", + "content": "[26] Yuting Wang, Velibor Ilic, Jiatong Li, Branislav Kisačanin, and Vladimir Pavlovic. ALWOD: Active Learning for Weakly-Supervised Object Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6459–6469, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[27] Yue Xi, Wenjing Jia, Qiguang Miao, Junmei Feng, Jinchang Ren, and Heng Luo. Detection-Driven Exposure-Correction Network for Nighttime Drone-View Object Detection. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "type": "text", + "content": "[28] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Query-Det: Cascaded Sparse Query for Accelerating High-Resolution Small Object Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13668–13677, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 443 + ], + "type": "text", + "content": "[29] Xinyi Ying, Li Liu, Yingqian Wang, Ruojing Li, Nuo Chen, Zaiping Lin, Weidong Sheng, and Shilin Zhou. Mapping Degeneration Meets Label Evolution: Learning Infrared Small Target Detection with Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15528-15538, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "text", + "content": "[30] Xuehui Yu, Pengfei Chen, Di Wu, Najmul Hassan, Guorong Li, Junchi Yan, Humphrey Shi, Qixiang Ye, and Zhenjun Han. Object Localization Under Single Coarse Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4868-4877, 2022. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "text", + "content": "[31] Xuehui Yu, Pengfei Chen, Kuiran Wang, Xumeng Han, Guorong Li, Zhenjun Han, Qixiang Ye, and Jianbin Jiao. CPR++: Object Localization via Single Coarse Point Supervision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 633 + ], + "type": "text", + "content": "[32] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16783-16793, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[33] Xiang Yuan, Gong Cheng, Kebing Yan, Qinghua Zeng, and Junwei Han. Small Object Detection via Coarse-to-Fine Proposal Generation and Imitation Learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6317-6327, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[34] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and Tracking" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 72, + 553, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 72, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 553, + 105 + ], + "type": "text", + "content": "Meet Drones Challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 2, 6" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11716" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PolarFree_ Polarization-based Reflection-Free Imaging/e1567ff5-b726-4ec2-baeb-81cce9b4c3bf_content_list.json b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/e1567ff5-b726-4ec2-baeb-81cce9b4c3bf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cdb72bc0a3e03dfef455054559a2db1d9acc8d51 --- /dev/null +++ b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/e1567ff5-b726-4ec2-baeb-81cce9b4c3bf_content_list.json @@ -0,0 +1,2386 @@ +[ + { + "type": "text", + "text": "PolarFree: Polarization-based Reflection-Free Imaging", + "text_level": 1, + "bbox": [ + 207, + 130, + 763, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingde Yao $^{1}$ , Menglu Wang $^{3}$ , King-Man Tam $^{1,4}$ , Lingen Li $^{1}$ , Tianfan Xue $^{\\text{☑}1,2}$ , Jinwei Gu $^{1}$ $^{1}$ The Chinese University of Hong Kong, $^{2}$ Shanghai AI Laboratory \n $^{3}$ University of Science and Technology of China, $^{4}$ Institute of Science Tokyo", + "bbox": [ + 99, + 179, + 867, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mingdeyao@foxmail.com, tfxue@ie.cuhk.edu.hk", + "bbox": [ + 284, + 237, + 676, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 280, + 313, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reflection removal is challenging due to complex light interactions, where reflections obscure important details and hinder scene understanding. Polarization naturally provides a powerful cue to distinguish between reflected and transmitted light, enabling more accurate reflection removal. However, existing methods often rely on small-scale or synthetic datasets, which fail to capture the diversity and complexity of real-world scenarios. To this end, we construct a large-scale dataset, PolaRGB, for Polarization-based reflection removal of RGB images, which enables us to train models that generalize effectively across a wide range of real-world scenarios. The PolaRGB dataset contains 6,500 well-aligned mixed-transmission image pairs, $8 \\times$ larger than existing polarization datasets, and is the first to include both RGB and polarization images captured across diverse indoor and outdoor environments with varying lighting conditions. Besides, to fully exploit the potential of polarization cues for reflection removal, we introduce PolarFree, which leverages diffusion process to generate reflection-free cues for accurate reflection removal. Extensive experiments show that PolarFree significantly enhances image clarity in challenging reflective scenarios, setting a new benchmark for polarized imaging and reflection removal. Code and dataset are available at https://github.com/mdyao/PolarFree.", + "bbox": [ + 75, + 299, + 473, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reflection removal algorithms [1, 19, 30, 34, 38, 41] remove unwanted reflections in captured images, playing a critical role in applications such as autonomous driving [12] and photography [11, 43]. This problem commonly arises when imaging through semi-reflectors, like windows or glass, and overlapping reflections may obscure important details of scenes we want to capture. This problem is often formulated [10, 11] as a linear combination of the transmission layer $T$ and the reflection layer $R$ :", + "bbox": [ + 75, + 715, + 468, + 851 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\nM = \\alpha_ {t} T + \\alpha_ {r} R, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 853, + 468, + 868 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80e710133ca035792331381ce071ae6798e7f3a6d4566b4935608c56fe27cc34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 284, + 630, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df1740ae295ef3e8eefaf548cb03914837bd7268e032c6f634102d229c4fc1c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 285, + 759, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41aeecca7066eabb3ce5abf7df59c06933a09c0d75f0eeee98d7be733ba82e16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 285, + 890, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e089df2f146b43c1ea7ee94c61df54a93e2dd0614df58e2ee137b157fb915a94.jpg", + "image_caption": [ + "Input", + "RDRNet [46]", + "Figure 1. Our PolarFree effectively leverages polarization information to remove reflections, achieving superior performance in challenging scenes with complex backgrounds and highlights where previous methods [11, 15, 46] often fail." + ], + "image_footnote": [], + "bbox": [ + 501, + 382, + 630, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39171562de537b9915487cc734622c40e855ab5d80fe66765cc1ea62e07066a2.jpg", + "image_caption": [ + "Kim et al. [11]", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 632, + 382, + 759, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48ffe3f5b52fd54b2427e6c1e2c77d46d06971a8eab1c4ec5010b2e6e091d1e8.jpg", + "image_caption": [ + "DSRNet [15]", + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 761, + 382, + 890, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $M$ is the mixed captured image, and $\\alpha_{t},\\alpha_{r}$ are blending coefficients resulting from light attenuation.", + "bbox": [ + 500, + 546, + 890, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Polarization image sensors are becoming mainstream, allowing users to easily capture polarization images from a single shot in real-time [32]. However, existing methods [10, 11, 24, 46] typically rely on intensity-based cues, such as pixel brightness and color gradients, to distinguish transmitted and reflected layers. These methods face challenges because reflection removal is a highly ill-posed inverse problem [21] that recovers two unknown layers (reflection and transmission) from a single observation. Polarization provides valuable physics-based cues [27, 29] to alleviate ill-posed issues (Fig. 1), whereas transmitted light retains a distinct degree of polarization(Fig. 2b). This difference provides crucial signals for separating the two layers. Notably, at the Brewster angle (Fig.2c), reflected light is fully polarized, enabling effective reflection removal [2].", + "bbox": [ + 496, + 579, + 892, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite significant advances of polarization images, a key challenge in polarization-based reflection removal is the lack of large-scale, high-quality datasets. Existing polarized reflection removal datasets [18, 39] are limited in size and diversity, relying on small (<1000) or synthetic samples that fail to capture the complexity of real-world light-", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 99, + 886, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "10890", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ea6bef2076cfbf69882589e06b5d3ae6c93b962c7da50576b216d8f73874dfb.jpg", + "image_caption": [ + "Figure 2. (a) & (b) A semi-reflector transforms unpolarized light into polarized light upon reflection and refraction, which is undetectable by standard RGB cameras but can be leveraged by polarization cameras for reflection-suppression tasks. (c) At the Brewster angle [2], a polarizer minimizes reflections." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 470, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing conditions, materials, and scenes. Moreover, they typically exclude color information, reducing their applicability in real-world reflection removal tasks. Thus, there is a pressing need for a large-scale, comprehensive dataset that includes both RGB and polarization images, captured in diverse real-world environments, to advance polarization-based reflection removal.", + "bbox": [ + 75, + 291, + 468, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap, we introduce PolaRGB, a novel dataset specifically collected for polarization-based reflection removal. As shown in Table 1, PolaRGB contains 6,500 high-quality, well-aligned RGB-polarization image pairs, $8 \\times$ larger than previous dataset [18]. Our dataset covers a diverse range of scenes, lighting conditions, and exposure settings, and is captured using off-the-shelf commercial cameras with polarized color patterns to ensure real-world applicability. PolaRGB provides both mixed images and ground-truth transmission layers, enabling accurate reflection separation and significantly enhancing the effectiveness of reflection removal across real-world scenarios.", + "bbox": [ + 75, + 398, + 468, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, extracting reflection-free information from polarization data is challenging [18, 25, 39] due to the randomness of shooting angles, scene variations, and changing lighting conditions. To address this issue, we leverage the powerful generative capabilities of the diffusion model [3,6,8] to generate reflection-free cues. The diffusion model extracts and refines reflection-free priors from polarization images, effectively guiding the reflection removal and yielding precise and robust reflection-free results.", + "bbox": [ + 75, + 580, + 468, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, PolarFree consists of two steps: a prior-generation step and a reflection removal step. First, in the prior-generation step, we use a diffusion model to generate reflection-free prior based on the polarization and RGB inputs. This strategy not only guides accurate reflection isolation in RGB images but also recovers background details, which previous methods [11, 15, 46] may miss. Next, the reflection removal step leverages the prior to effectively remove reflections, ensuring accurate transmission restoration. Additionally, we introduce a phase-based loss function in the frequency domain to mitigate color discrepancies caused by semi-reflections, guiding the network to focus", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/ba3aca2169ccbb3aa213af4da651194a337f0255e4ba274ee915577c8369ebc6.jpg", + "table_caption": [ + "Table 1. Comparisons between PolaRGB and existing reflection removal datasets." + ], + "table_footnote": [], + "table_body": "
DatasetPolarizationRGBRAWData sizeResolution
SIR2 [36]NoYesNo500540 × 400
RRW [46]NoYesNo14,9522580 × 1460
ReflectNet [39]Yes (Syn.)YesNo-1500 × 1000
Lei et al. [18]Yes (Real)NoYes8071224 × 1024
PolaRGB (Ours)Yes (Real)YesYes6,5001224 × 1024
", + "bbox": [ + 503, + 116, + 890, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on reflection removal rather than color adjustment. These components enable PolarFree to achieve robust reflection suppression while preserving the clarity and integrity of the transmission across diverse real-world scenes.", + "bbox": [ + 496, + 208, + 890, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments on the PolaRGB dataset demonstrate the effectiveness of PolarFree, which outperforms existing methods [11, 15, 46] $\\sim$ 2dB in terms of PSNR. Additionally, we performed real-world testing in environments commonly affected by reflections such as museums and galleries, demonstrating the method's effectiveness in real-world scenarios. Our approach improves image clarity and preserves details better than previous methods, paving the way for more practical, reflection-robust applications.", + "bbox": [ + 496, + 268, + 890, + 405 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 406, + 640, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Reflection Removal", + "text_level": 1, + "bbox": [ + 500, + 424, + 684, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reflection removal [1, 30, 34] has a longstanding history in computer vision, aiming to enhance the occluded objects caused by semi-reflectors like glass and windows. Early approaches [20, 24] typically relied on handcrafted priors, such as smoothness assumptions and gradient sparsity, to differentiate the reflection layer from the underlying scene. Some methods incorporated ghosting [31], flash cues [17, 19] and edge consistency [7], while others improved accuracy by leveraging multi-image setups [22, 28], utilizing temporal variation [26] and parallax [28] to separate reflections from background layers. However, these methods often require strict assumptions and fail in real-world scenes with complex textures or lighting [19, 38].", + "bbox": [ + 496, + 441, + 890, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rise of deep learning [5, 35, 42, 44, 46], reflection removal methods advanced considerably. Techniques based on neural networks have shown promise by learning to separate reflection and transmission layers directly from data [10, 15, 35-37, 44]. However, without essential physical insights, these methods act as Bayesian regressions toward a learned average, often lead to suboptimal separation in complex environments with varying lighting, textures, and reflection intensities [35, 44]. Recent methods [9, 45] utilize language to interactively remove the reflection, which require additional efforts, limiting their practicality.", + "bbox": [ + 496, + 638, + 890, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Polarization Reflection Removal", + "text_level": 1, + "bbox": [ + 500, + 808, + 782, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Polarization is an effective way to remove reflections due to the physics principle: the polarization of light behaves differently in the reflection and transmission layers [4, 13, 23, 27, 30], allowing separation, especially at the Brewster angle [4, 27]. However, capturing images exactly", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10891", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8ca8a3972a20070ec3d390a56d58221c826881066fa30c006b18aa06961f73a9.jpg", + "image_caption": [ + "(a) Hierarchical structure of PolarRR dataset" + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 449, + 354 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eea8759d742d4f8018cced432db6271d3df78cdc252da9d3585e413096a9e90f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 88, + 563, + 147 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4f283e0279b4acc02c52c830457a3829881c71f71df8625e75cf5f24c7b80dbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 88, + 663, + 147 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/880553008569a864c8d9aaaf274b5fb203da3d7fb67e7d31fd02ab9ce0e8705b.jpg", + "image_caption": [ + "(b) Diverse reflection conditions" + ], + "image_footnote": [], + "bbox": [ + 465, + 148, + 563, + 208 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/65ac40515c73091e5f6d18db5c5e31dda4b9c865764a52d22efd4ac11e8b5785.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 565, + 148, + 663, + 208 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d2809239934c849ed6ebd53e3634f7a3bd7cb6960776f5947e5061b3abe70246.jpg", + "image_caption": [ + "(c) Video-based capture method" + ], + "image_footnote": [], + "bbox": [ + 671, + 88, + 834, + 208 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6d5014dcc4f28f019f5fe6eb6c70c122dc15cc435762b52fce29220e54f21187.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 836, + 88, + 887, + 208 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d64caf525b3d078a6dd1b940c763a7d5c9710b205b7241eac8eeab4d949d7754.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 231, + 570, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/45136a25128f63284bc3a2d8b0d899872b13cd308c274102f896c51a81f2162f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 231, + 674, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/937d9fcc2dde35c82bb94007feabde22a2a2d9455eec666fa2846b7d345f7f63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 231, + 781, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c37ca18470fac448ded46c855863281c1dce2742f0ac08118c99ef13030baa6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 231, + 887, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/de043a9bc68b63d70554f9cae2c1a74cd12ae5e74ddcec154aa0adf9024cb4a1.jpg", + "image_caption": [ + "Figure 3. Overview of the PolaRGB dataset. (a) Hierarchical structure of scenes is shown in the ring, with legends indicating sample counts and subset types. (b) Typical scenes illustrating varied reflection conditions: I. smooth blending of reflection and refraction, II. abrupt reflection with mixed components, III. reflection dominant over transmission, and IV. minimal or no reflection. (c) Video-based capture method (details in Sec. 3). (d) We provide polarized images at angles $\\phi = 0^{\\circ}$ , $45^{\\circ}$ , $90^{\\circ}$ , and $135^{\\circ}$ , along with derived AoLP, DoLP, and a well-aligned unpolarized image. The dataset also includes ground truth transmission and estimated reflections, all available in both raw and RGB formats." + ], + "image_footnote": [], + "bbox": [ + 465, + 292, + 570, + 351 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a16e0d41a7120c3aa13d8dea18c55b320b56fb14e74871494372cae3881f862c.jpg", + "image_caption": [ + "(d) Image types included in the dataset" + ], + "image_footnote": [], + "bbox": [ + 570, + 292, + 676, + 351 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/afd0ba01fe5cf90e494b5945746281ecf547a5c58f87a2eab31ecbb57e1b4af3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 292, + 781, + 351 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bcf55f5dec9a9232227e777e00f74312bf5ea9a7612f9123117f7bbdc71fbab0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 292, + 887, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "at the Brewster angle, as shown in Fig. 2c, is challenging. Therefore, in practice, methods typically rely on images taken at multiple polarization angles (e.g., $0^{\\circ}$ , $45^{\\circ}$ , $90^{\\circ}$ , and $135^{\\circ}$ ) to capture polarization information [4, 27, 30].", + "bbox": [ + 76, + 457, + 468, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early polarized reflection removal methods [4, 27, 30] employ mathematical techniques like Principle Component Analysis (PCA) to separate reflection layers, which are effective when the transmission and reflection layers have significant content differences. Subsequently, Kong et al. [16] propose a multiscale scheme that automatically identifies the optimal separation of the reflection and background layers. Another advancements [39] introduce a neural network-based approach for polarization-guided reflection removal with synthetic dataset. Lyu et al. [25] utilize a pair of unpolarized and polarized images but lack of realistic polarization patterns. Lei et al. [18] further contribute by collecting a dataset of polarized images for reflection removal. However, this dataset only includes limited scene variations from pure polarization scenes, which limits its application scenario. In contrast, we present the first large-scale dataset captured for polarization-based RGB reflection removal, consisting of 6,500 well-aligned transmission-reflection pairs. We also propose a polarization-based reflection removal network based on diffusion models.", + "bbox": [ + 75, + 518, + 468, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. PolaRGB Dataset", + "text_level": 1, + "bbox": [ + 76, + 825, + 248, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Analysis", + "text_level": 1, + "bbox": [ + 76, + 848, + 179, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although a few polarization-based reflection removal datasets exist [18, 39], they face major limitations: (1) they", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are often synthetically generated [25, 39], which may not generalize well to real polarization, as perfectly simulating polarization phenomena is very hard; (2) they are typically limited in size and diversity (<1,000 samples), with limited scene variations, which constrains model robustness and generalizability [18, 39]; and (3) they lack RGB data [18, 39], reducing the practical value of these datasets.", + "bbox": [ + 496, + 457, + 890, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our polarized-based reflection removal dataset, PolaRGB, addresses all these limitations with the following key features, as shown in Fig. 3. It has following advantages: (1) It contains real-captured images with perfect spatial alignment. We achieve pixel-level alignment through a three-step process: careful capture setup, manual filtering, and homography transformations, all in raw space to avoid demosaicing artifacts (see Fig. 4b and more details in supplementary). (2) It contains large-scale diverse scenes. The collected dataset covers a wide range of indoor and outdoor settings with four distinct reflection types, including smooth, sharp, high-brightness, and subtle reflections (see Fig. 3b). (3) It has extensive data modalities. Our dataset offers paired mixed and transmission images with both polarization and RGB captures, precisely aligned and available in both raw and RGB formats (see Fig. 3d).", + "bbox": [ + 496, + 563, + 892, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Processing Pipeline", + "text_level": 1, + "bbox": [ + 500, + 808, + 684, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We leverage an efficient video-based capture flow for dataset collection, as shown in Fig. 3c. First, we capture a transmission-only image $T_{raw}$ as the ground truth. Then, we place a semi-reflective glass plate in front of the scene and rotate it continuously to capture mixed images", + "bbox": [ + 496, + 825, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10892", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/718d1884c76556825fa944bdbda653d24baa38fb79c5faeddbee1429641b0068.jpg", + "image_caption": [ + "Figure 4. Data processing pipeline for obtaining aligned mixed and transmission images, and polarized images." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$M_{\\text{raw}}$ . A division-of-focal-plane polarization camera with color Bayer pattern is used to simultaneously capture both polarized and RGB information, as shown in Fig. 4.", + "bbox": [ + 76, + 551, + 468, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the raw mixed image $M_{raw}$ and the transmission image $T_{raw}$ , we process the images sequentially in the raw and RGB domains as shown in Fig. 4a. In the raw domain, we first align $M_{raw}$ and $T_{raw}$ to correct spatial misalignment caused by light refraction. This is done by separating the images into different polarization angles and color channels, then applying affine transformation matrices to each of the channels to avoid aliasing from direct alignment, as shown in Fig. 4b. Next, we perform polarization separation on the aligned images to obtain four polarization images $(0^{\\circ}, 45^{\\circ}, 90^{\\circ}$ , and $135^{\\circ}$ ), and use Malus's law [27, 29] to sum them, producing the unpolarized image of the scene. Finally, we estimate the reflection image by searching for the optimal blending coefficient $\\alpha_r$ and $\\alpha_t$ in edge space, as illustrated in Eq. 1. For detailed steps and proof, please refer to the supplementary material.", + "bbox": [ + 75, + 597, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we apply demosaicking to the processed raw images to obtain the mixed, transmission, and reflection images, each consisting of four polarization images and the unpolarized RGB image. Additionally, we compute the", + "bbox": [ + 76, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stokes parameters (see Sec. 4.1) to estimate the Angle of Linear Polarization (AOLP) and Degree of Linear Polarization (DOLP), which are then used in reflection removal.", + "bbox": [ + 498, + 90, + 893, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 138, + 591, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 159, + 702, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object. Given an RGB image and its corresponding polarization images, we aim to recover the transmission layer of the RGB image by utilizing the distinct polarization characteristics. Our measurement consists of spatially-aligned RGB and polarization images, where the polarization images captured at four distinct angles $(0^{\\circ}, 45^{\\circ}, 90^{\\circ}$ , and $135^{\\circ})$ . These four angles provide a comprehensive polarization measurement [30], which allows us to compute polarization features crucial for separating layers.", + "bbox": [ + 496, + 179, + 890, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Polarization Preliminary. For semi-reflective surfaces that produce mixed images, the observed intensity $I_M^\\phi (x)$ can be decomposed into reflected intensity $I_R^\\phi (x)$ and transmitted intensity $I_T^\\phi (x)$ as", + "bbox": [ + 496, + 321, + 890, + 387 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {M} ^ {\\phi} (x) = \\alpha_ {R} I _ {R} ^ {\\phi} (x) + \\alpha_ {T} I _ {T} ^ {\\phi} (x). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 388, + 890, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If a linear polarizer is placed in front of the camera at an angle $\\phi$ , the captured intensity $I^{\\phi}(x)$ can be represented as", + "bbox": [ + 498, + 409, + 890, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {M} ^ {\\phi} (x) = \\alpha (\\theta ; \\phi ; \\phi_ {\\perp}) I _ {R} ^ {\\phi} (x) + (1 - \\alpha (\\theta ; \\phi ; \\phi_ {\\parallel})) I _ {T} ^ {\\phi} (x), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 449, + 890, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha (\\theta ;\\phi ;\\phi_{\\perp})$ and $(1 - \\alpha (\\theta ;\\phi ;\\phi_{\\parallel}))$ are polarization-dependent mixing coefficients based on the angle of incidence $\\theta ,\\phi_{\\perp}$ and $\\phi_{\\parallel}$ are the canonical polarization directions for the reflection and transmission, respectively. Directly solving for these parameters is challenging because $\\alpha (\\cdot)$ , $I_R^\\phi (x)$ , and $I_R^\\phi (x)$ are all unknown [27, 39]. Additionally, relationship between the observed intensity and the underlying reflected and transmitted components is highly nonlinear and influenced by various factors [39]. To address this, we utilize Stokes parameters [18] that provide a efficient way to represent and analyze polarized light, enabling more robust separation of reflection and transmission.", + "bbox": [ + 496, + 476, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stokes Parameters. To capture the polarization effects in the scene, we use Stokes parameters $[S_0, S_1, S_2]$ , which can be derived from intensity measurements at specific polarization angles $(0^\\circ, 45^\\circ, 90^\\circ$ , and $135^\\circ$ ):", + "bbox": [ + 496, + 666, + 890, + 726 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} S _ {0} = \\left(I _ {0 ^ {\\circ}} + I _ {4 5 ^ {\\circ}} + I _ {9 0 ^ {\\circ}} + I _ {1 3 5 ^ {\\circ}}\\right) / 2, (4) \\\\ S _ {1} = I _ {0 ^ {\\circ}} - I _ {9 0 ^ {\\circ}}, S _ {2} = I _ {4 5 ^ {\\circ}} - I _ {1 3 5 ^ {\\circ}}. (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 727, + 890, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $S_0$ represents the total intensity of light, and $S_1$ and $S_2$ provide information about the linear polarization state of the light based on intensity differences at these key angles.", + "bbox": [ + 496, + 763, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the Stokes parameters, we compute the Degree of Linear Polarization (DOLP) and Angle of Linear Polarization (AOLP) as follows [2]", + "bbox": [ + 496, + 810, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD O L P (x) = \\sqrt {S _ {1} (x) ^ {2} + S _ {2} (x) ^ {2}} / S _ {0} (x), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 854, + 890, + 886 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA O L P (x) = \\frac {1}{2} \\operatorname {a t a n 2} \\left(S _ {2} (x) / S _ {1} (x)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 869, + 890, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10893", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7de8f37e913ce7e31160fa6f024fdfaef8403b25dd3642752a06ce45afc6a5af.jpg", + "image_caption": [ + "(a) Inference", + "Figure 5. Pipeline of PolarFree. (a) During inference, PolarFree leverages polarized and RGB images as inputs, which are feeds into a conditional diffusion model to generate the prior $\\hat{z}_0$ . The generated prior, along with the inputs, is then passed to the reflection removal backbone $\\mathcal{F}_{\\mathrm{remove}}$ to remove reflections. (b) PolarFree is trained in two stages. (1) A prior encoder extracts a reflection-free prior $z_0$ from clean transmission images and polarization cues, which serves as the supervision for the conditional diffusion model in stage two. (2) The conditional diffusion model is trained to progressively denoise noisy images, supervised by the prior from stage one, ensuring robust reflection separation." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 416, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6de156d3ccb9bc576f12eb41b7bcc4bdf4e035160926479af57b03b528b5875d.jpg", + "image_caption": [ + "(b) Training" + ], + "image_footnote": [], + "bbox": [ + 421, + 88, + 883, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this formulation, DOLP describes the proportion of polarized light relative to total intensity, which helps indicate the degree of reflection in the scene. AOLP, on the other hand, reveals the orientation of the polarized light, allowing us to distinguish between reflection and transmission components more effectively, as shown in Fig. 3d. By leveraging intensity measurements at $0^{\\circ}$ , $45^{\\circ}$ , $90^{\\circ}$ , and $135^{\\circ}$ , these parameters provide valuable cues for separating mixed reflection and transmission layers in semi-reflective scenes.", + "bbox": [ + 75, + 457, + 468, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. PolarFree Network", + "text_level": 1, + "bbox": [ + 76, + 595, + 261, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve high-quality reflection removal with effective utilization of polarization information, we introduce Polar-Free, a delicately designed two-step network, where each step addresses a distinct aspect of the reflection removal challenge. Inspired by [3], we leverage diffusion model to generate the prior for reflection removal. As shown in Fig. 5, during inference, the first step utilizes a conditional diffusion model to extract reflection-free priors, effectively isolating essential details from polarization data. The second step integrates these priors with RGB inputs, guiding the network to accurately separate reflections and enhance clarity, even in complex, real-world environments.", + "bbox": [ + 75, + 609, + 470, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prior Generation. As shown in Fig. 5a, in the first step, we start with a randomly initialized noise $n$ , which is gradually denoised through a conditional diffusion model $\\mathcal{F}_{\\mathrm{diff}}$", + "bbox": [ + 76, + 799, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {0} = \\mathcal {F} _ {\\text {d i f f}} (n | M _ {\\text {c o n d}}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 848, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{z}_0$ is the generated prior, $M_{\\mathrm{cond}} = \\{M_{\\mathrm{polar}}, M_{\\mathrm{aolp}}, M_{\\mathrm{dolp}}, M_{\\mathrm{rgb}}\\}$ represents the mixed im", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ages, and $n$ is the initial noise. Through an iterative denoising process, the diffusion model progressively refines a noise image, conditioned on the polarization and RGB data, to generate a reflection prior.", + "bbox": [ + 496, + 457, + 890, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The denoising process follows the denoising diffusion probabilistic model (DDPM) framework [8] and employs a U-Net architecture to predict noise. At each timestep $t$ , the U-Net receives a noisy intermediate $z_{t}$ and outputs a noise estimate $\\epsilon_{\\theta}(z_{t},t)$ , predicting the noise added at that timestep. This process can be formulated as", + "bbox": [ + 496, + 520, + 890, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(z _ {t}, M _ {\\text {c o n d}}, t\\right)\\right) + \\sigma_ {t} z, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 609, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ and $\\beta_{t}$ control the noise schedule across the timesteps, $\\sigma_{t}$ is the standard deviation, and $z$ is noise sampled from a standard Gaussian distribution. $\\bar{\\alpha}_{t}$ is the cumulative product of $\\alpha_{t}$ , indicating noise level to timestep $t$ .", + "bbox": [ + 496, + 645, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, the U-Net gradually removes the noise in each step, conditioning on the polarization-based measurement input $M_{\\mathrm{cond}}$ . This process continues until noise is fully removed, yielding a sample $\\hat{z}_0$ that represents the reflection-free prior distribution extracted from the conditioned inputs.", + "bbox": [ + 496, + 708, + 890, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reflection Removal. As shown in Fig. 5a, once the prior $\\hat{z}_0$ has been obtained, the second step of PolarFree is separating the transmission and reflection layers. By leveraging the polarization cues provided by $M_{\\mathrm{polar}}$ , $M_{\\mathrm{aolp}}$ , and $M_{\\mathrm{dolp}}$ , the model $\\mathcal{F}_{\\mathrm{remove}}$ separates the transmission features from the reflection ones, which can be represented as", + "bbox": [ + 496, + 792, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {T} _ {r g b} = \\mathcal {F} _ {\\text {r e m o v e}} \\left(\\hat {z} _ {0}, M _ {\\text {c o n d}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 883, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10894", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/be2378f448fa49b76e8cc9819f7cdac0e5aa1f09f0737731308edce26a1fc292.jpg", + "image_caption": [ + "Figure 6. (a) Phase information preserves shape and texture details, while color primarily affects the amplitude. (b) We apply two types of random color perturbations to the image and compute the perturbation errors. Phase-based loss is less sensitive to color changes." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 444, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\mathrm{remove}}$ is a reflection removal neural network. This step ensures that the final output contains distinct transmission and reflection components, facilitating high-quality image reconstruction under challenging reflective scenarios.", + "bbox": [ + 75, + 291, + 467, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Training", + "text_level": 1, + "bbox": [ + 76, + 358, + 181, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To train our PolarFree, particularly to address the challenge of the diffusion model lacking a suitable ground truth, we adopt a two-stage training strategy, as illustrated in Fig. 5b. This strategy consists of two sequential objectives: learning to extract a reflection-free prior and learning to generate a reflection-free prior.", + "bbox": [ + 75, + 378, + 467, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First Stage. In the first stage, we train an encoder to extract a prior $z_0$ for reflection-free information, which will serve as the supervision signal for the diffusion model in the second stage. Specifically, we feed the ground truth transmission images $T_{\\mathrm{cond}} = \\{T_{\\mathrm{polar}}, T_{\\mathrm{aolp}}, T_{\\mathrm{dolp}}, T_{\\mathrm{rgb}}\\}$ into the encoder $\\mathcal{E}$ , obtaining $z = \\mathcal{E}(M_{\\mathrm{cond}}, T_{\\mathrm{rgb}})$ . This $z$ contains reflection-free information enriched with polarization-related cues. Subsequently, $z_0$ conditions the reflection removal backbone $\\mathcal{F}_{\\mathrm{remove}}$ to predict the clean transmission image $\\hat{T}_{\\mathrm{rgb}} = \\mathcal{F}_{\\mathrm{remove}}(z, M_{\\mathrm{cond}})$ .", + "bbox": [ + 75, + 476, + 467, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second Stage. In the second stage, we train the diffusion model $\\mathcal{F}_{\\mathrm{diff}}$ to generate a reflection-free prior $z_0$ from noisy input images, and finetune the reflection removal backbone $F_{\\mathrm{remove}}$ . The key challenge here is that the diffusion model lacks of direct supervision. Therefore, we leverage the prior $z_0$ extracted in the first stage as the supervision signal to guide the model.", + "bbox": [ + 76, + 635, + 467, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We start with the extracted reflection-free prior $z_0$ and add noise over multiple timesteps. This process transforms the \"ground-truth\" prior $z_0$ into noisy versions $z_t$ at each timestep $t$ , which can be expressed as", + "bbox": [ + 75, + 742, + 467, + 801 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(z _ {t} \\mid z _ {0}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {0}, \\beta_ {t} \\mathbf {I}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 803, + 467, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{N}$ represents a Gaussian distribution, $\\beta_{t}$ is the noise schedule controlling the noise level added at each timestep, $\\mathbf{I}$ represents the identity matrix.", + "bbox": [ + 75, + 823, + 467, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the reverse diffusion process, the model is trained to progressively remove the noise step by step to recover the", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "clean prior $z_0$ , which can be represented as", + "bbox": [ + 498, + 92, + 784, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t, M _ {\\text {c o n d}}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right)), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 106, + 888, + 123 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta}(z_t,t,M_{\\mathrm{cond}})$ is the mean function of the current state $z_{t}$ , timestep $t$ , and a conditional input $M_{\\mathrm{cond}}$ .", + "bbox": [ + 498, + 126, + 888, + 156 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Throughout the reverse process, the model conditions on the polarization measurements $M_{\\mathrm{cond}}$ , which provide essential information about the scene's physical properties. These measurements help the model accurately generate reflection-free components during the denoising process.", + "bbox": [ + 496, + 157, + 888, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After the reverse process completes, the model outputs a final reflection-free prior $\\hat{z}_0$ , which contains clean transmission information. This prior is used as the guidance signal for the reflection removal backbone $\\mathcal{F}_{\\mathrm{remove}}$ . The backbone takes this reflection-free prior and the polarization data to predict the clean transmission image $\\hat{T}_{rgb}$ .", + "bbox": [ + 496, + 233, + 890, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Losses", + "text_level": 1, + "bbox": [ + 500, + 327, + 584, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Basic Loss. To optimize PolarFree, we follow [46] to utilize three basic losses: L1 loss, VGG perceptual loss [14], and total variation (TV) loss [33]. The L1 loss minimizes the pixel-wise difference between the predicted transmission image $\\hat{T}_{\\mathrm{rgb}}$ and the ground truth transmission image $T_{\\mathrm{rgb}}$ as $\\mathcal{L}_1 = \\|\\hat{T}_{\\mathrm{rgb}} - T_{\\mathrm{rgb}}\\|_1$ . The VGG perceptual loss compares feature activations from a pre-trained VGG network [14], weighted by $\\lambda_l$ : $\\mathcal{L}_{\\mathrm{VGG}} = \\sum_l \\lambda_l \\|\\phi_l(\\hat{T}_{\\mathrm{rgb}}) - \\phi_l(T_{\\mathrm{rgb}})\\|_1$ , where $\\phi_l$ denotes the activations of the $l$ -th VGG layer. We also use TV loss to constrain consistency via the gradient operator $\\mathcal{L}_{\\mathrm{TV}} = \\|\\nabla \\hat{T}_{\\mathrm{rgb}} - \\nabla T_{\\mathrm{rgb}}\\|_1$ .", + "bbox": [ + 496, + 347, + 890, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Phase loss. While basic losses help the network match color and intensity values to the ground truth, they struggle with color discrepancies caused by semi-reflective surfaces during dataset collection. These variations, due to reflection and transmission properties of materials, affect the color and intensity, leading to mismatches between the predicted and target images and hindering the model's ability to learn the correct transmission map.", + "bbox": [ + 496, + 521, + 890, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address this issue, we introduce a phase loss to focus on the structural information of the transmission, which is less sensitive to color variations. As shown in Fig. 6, phase information primarily captures the geometry and texture of the image, independent of color changes. The phase loss is formulated as", + "bbox": [ + 496, + 643, + 890, + 731 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p h a s e}} = \\| \\angle (F F T (\\hat {T})) - \\angle (F F T ((T _ {r g b})) \\| _ {1}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 733, + 888, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $FFT$ is the Fourier transform, and $\\angle (\\cdot)$ represents the phase angle of the Fourier coefficients.", + "bbox": [ + 496, + 753, + 888, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diffusion Loss. The diffusion loss follows the standard DDPM formulation [8], where the model predicts the noise at each step $t$ and minimizes the difference between the predicted noise and the true noise as", + "bbox": [ + 496, + 789, + 888, + 848 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i f f} = \\mathbb {E} _ {q \\left(z _ {t} \\mid z _ {t - 1}\\right)} \\left[ \\| \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) - \\epsilon_ {\\text {t r u e}} \\left(z _ {t}\\right) \\| _ {2} ^ {2} \\right]. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 849, + 888, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\epsilon_{true}(z_t)$ is the added noise obtained by Eq. 9, and $\\epsilon_{\\theta}(z_t,t)$ is the predicted noise, conditioned on the polariza", + "bbox": [ + 500, + 869, + 888, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10895", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f1d6246840b4417a8613b3dadab3f17d72504d483da8bedd949d838176c584f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 209, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e5b7ec04197727068b0c5a2c107194a0d261b1e1315855dbb59204ae1692cc02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 88, + 346, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8f75f8c97a880694f46513fb0e643b1cce199f11e37a28fd297731bc27ddc24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 88, + 482, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9bc029067c7aee81c3bce02c4710497f445c49c00df256b3c468e3cb88bde4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 88, + 619, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f8b43ca56488c0242d67801492d80300353f6c8485d42c3bc67e72595de61cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 88, + 754, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/577ef286cb47296cdc5f62e8f6a17cd37c06941375e9a7929de354a13195c1df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 88, + 890, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bcdaea9895d86fbae76f14feb4b6d37a098aeda4cd8082140675adaebc1309a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 176, + 209, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1c381867913b6bb267ec4c9dce476adffac5c81fccb6408ca54dd863c508b7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 176, + 346, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c778d1bce0bc6fef0246deb4852dc3859dcdf5831771e7a21620f1558f8653c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 176, + 482, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ea181c82777c9adcf39c493d5929db1e9215397fe58645ced72e5de4cdc84ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 176, + 619, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0497e34c1fd4e89d7e2a9b5a6a659190cf0747655d3d73c68a82629dfdd4f690.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 176, + 754, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6fb04191e0953ba6859f3a9497edde818e7410bf7cb608c800a2920c229df5f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 176, + 890, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31c184f1606ca7eb2defae89ce820be5a7684e230d7fd2185a85557f814df35c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 265, + 209, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b85ea1bef5394f460395ea3cfc8be2042bd1385c1c8b068fdfa4c52c67ff916a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 265, + 346, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26e573a11f10e382b1c5442278c7a747ea9a5861febdd2753500d6db2dbc36d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 265, + 482, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14dc69fbec5c9db6ad2fee69f3c7db6e8a85a29df9b9cad330971156575e569f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 265, + 619, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5bff9ef0535e591fc3fbd9686b58a9ffbb85b03d477271ff0de8821a9bde7d97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 265, + 754, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc8c45ec2fb7034965edaeb020639806da030c85d3ebb2a76574270bc3114120.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 265, + 890, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80e68386a5e357c4c12f457aaa5f773c74ed9e0aaa7162cbac2541ce775ce3ca.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 80, + 352, + 209, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3cf543eb9e9c5288c4c081825a6d5385a4d8c84b9f9df9b448e52242536985c4.jpg", + "image_caption": [ + "Kim et al. [15]" + ], + "image_footnote": [], + "bbox": [ + 215, + 352, + 346, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7c677f301c14db4d59f3307a9dbac00631c9891658789df36351c371f0ca390.jpg", + "image_caption": [ + "DSRNet [11]" + ], + "image_footnote": [], + "bbox": [ + 351, + 352, + 482, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4c5da7b26707524aa0121620f676b2fff8c99f894307c03f3d59bbd037b8e8a.jpg", + "image_caption": [ + "RDRNet [46]" + ], + "image_footnote": [], + "bbox": [ + 486, + 352, + 619, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb932e0324bd6638b77ad9fe20bcdb0a2a2e7498bb837ed4038e25ca5efb80d0.jpg", + "image_caption": [ + "Ours", + "Figure 7. Qualitative comparisons on the PolaRGB dataset." + ], + "image_footnote": [], + "bbox": [ + 624, + 352, + 754, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a885f30b7d3401b4b44d869e8fd419942a596ee2ea10aadf663d0f38edd9d59f.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 759, + 352, + 890, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion measurements and RGB input $M_{\\mathrm{cond}}$", + "bbox": [ + 76, + 474, + 354, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Total Loss. The above losses are weighted summed to serve as the supervision for the first stage as", + "bbox": [ + 76, + 496, + 468, + 526 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t a g e 1}} = \\gamma_ {1} \\mathcal {L} _ {1} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {V G G}} + \\gamma_ {3} \\mathcal {L} _ {\\mathrm {T V}} + \\gamma_ {4} \\mathcal {L} _ {\\text {p h a s e}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 527, + 468, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the second stage, we use a combined loss function consisting of the diffusion loss and the reconstruction loss as", + "bbox": [ + 76, + 546, + 468, + 575 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t a g e 2}} = \\gamma_ {5} \\mathcal {L} _ {\\text {d i f f}} + \\gamma_ {6} \\mathcal {L} _ {\\text {r e c o n}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 580, + 468, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{recon}}$ has the same format as $\\mathcal{L}_{\\mathrm{stage1}}$ , and $\\gamma (\\cdot)$ are the coefficients. This combined loss helps refine the reflection-free image while ensuring it aligns with the ground truth.", + "bbox": [ + 76, + 598, + 468, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 648, + 209, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 669, + 294, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train and evaluate PolarFree on the PolaRGB dataset. The whole dataset consists of 67 scenes, with a total of 6,500 paired images. For each scene, we keep the background (transmission) and camera fixed, and varying the glass position to capture images with reflections. The 67 scenes are then randomly split into 56 training scenes and 11 testing scenes, containing 6,312 and 188 paired images, respectively. This division ensures no data leakage between the training and testing sets, with each set containing only a subset of all categories.", + "bbox": [ + 75, + 688, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also test PolarFree in real-world scenes like museums and galleries, where ground truth is unavailable. We implement PolarFree using PyTorch on a single NVIDIA RTX 4090 GPU. Training is conducted with a batch size of", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2 and an AdamW optimizer with a learning rate of $2 \\times 10^{-4}$ , 30k iterations on the PolaRGB dataset for both stages.", + "bbox": [ + 500, + 474, + 890, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with recent advanced reflection removal methods, including Lei et al. [18], IBCLN [21], DSRNet [11], YTMT [10], and RDRNet [46]. For a fair comparison, we modify the input settings to align with ours and use only the transmission layer for supervision. We have re-trained baseline methods on the PolaRGB dataset. The evaluation is conducted using both objective metrics (PSNR, SSIM), a perceptual metric (LPIPS), and a language-based non-reference metric (Q-Align [40]).", + "bbox": [ + 496, + 506, + 890, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 500, + 648, + 591, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on PolaRGB. We present the quantitative results in Tab. 2. It can be seen that, our method outperforms other methods across multiple metrics, demonstrating the effectiveness of PolarFree for high-quality reflection removal. Notably, for fairness, we modified the input layer of all baselines to accept polarization information. Visual results in Fig. 7 show that PolarFree provides cleaner reflection removal with sharper edges and better color preservation. In contrast, previous methods often suffer from color distortions or imperfect reflection removal, especially in areas with complex reflections. Our method maintains high fidelity to the ground truth, especially in challenging regions with low light and subtle reflections.", + "bbox": [ + 496, + 664, + 890, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-captured without Ground Truth. To demonstrate the generalization ability to unseen and more complex re", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10896", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4dbf17d1e91ac2a2b36077764aefec87ba6f88160cafc98eeb7dc53b9d46e541.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 173, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/363fed455e966944353ab5237450f1889461997aeb0c0c154cc8bb2d28fa4ca7.jpg", + "image_caption": [ + "DSRNet [46]" + ], + "image_footnote": [], + "bbox": [ + 176, + 88, + 272, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f62b3e1c43008901a0003b0d099f0ea5bdc557a89c50301bf530a69d5f18a001.jpg", + "image_caption": [ + "Kim et al. [15]", + "Figure 8. Visual comparison of reflection removal results in real-world museum scenes with glass display cases." + ], + "image_footnote": [], + "bbox": [ + 274, + 88, + 370, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1a8b063639f65fea0303ee16ee9042c207ad2528618cf5a50f3572ec9f1a440.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 372, + 88, + 468, + 214 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fd4f9d18e16cb1553d6d6eef6802bb92596146e2d83a18b81567c90bfe044179.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on the PolaRGB dataset, evaluated using objective metrics (PSNR, SSIM), perceptual metric (LPIPS), and the Language-based non-reference metric (Q-align [40]). Best results are in bold, second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓Q-align↑
Lei et al. [18]18.730.79620.38043.2109
Kim et al. [15]20.670.83990.27143.7148
IBCLN [21]19.730.81730.24883.0938
YTMT [10]16.860.75440.44893.0938
DSRNet [11]16.840.79130.28283.6992
RDRNet [46]15.880.69640.52502.9531
Ours22.440.86810.13253.8867
", + "bbox": [ + 81, + 335, + 465, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "flection, We further evaluate our model on real-captured images at museum with complex reflections from glass enclosures. Here, we only provide qualitative results to visually assess the model's capability. As shown in Fig. 8, our approach effectively reduces reflections while preserving fine details, despite the complex lighting and material variations often present in museum environments. This demonstrates the practical robustness of our method and its potential for real-world applications.", + "bbox": [ + 75, + 465, + 468, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 604, + 230, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform ablation studies to evaluate the effectiveness of key components in our method. More experimental results and analyses can be found in the supplementary.", + "bbox": [ + 75, + 623, + 468, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Polarization Information. To assess the importance of polarization cues, we conduct an experiment where the polarization information (AOLP, DOLP, and polarization images) is removed, and the model is trained solely using RGB data. As shown in Table 3 and Fig. 9, the absence of polarization significantly impairs the model's ability to differentiate between reflection and transmission layers, leading to noticeably poorer reflection removal performance. These results demonstrate the critical role of polarization in accurate reflection removal.", + "bbox": [ + 75, + 669, + 468, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diffusion Prior. We also evaluate the effectiveness of diffusion prior by removing the conditional diffusion model. While without diffusion prior achieves reasonable results, as indicated by the comparison metrics in Table 3, it lacks fine-grained control and high-quality output offered by the", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c1b7b3fc411e10044bd1f1e525e275ea06cbeff5af14633e0a6490259dcb056.jpg", + "image_caption": [ + "w/o diffusion", + "Figure 9. Highly challenging reflection removal scenes with complex reflections and highlights. Top: Polarization images from different angles provide complementary information for effective reflection removal, note the third $(90^{\\circ})$ image. Bottom: The diffusion module effectively handles highlights, while polarization better restores the color and details in such challenging scenarios." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 596, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8904b546ba9899d560b2ee37c9e9e9f73080ad336a43c01adb31e23fbda4afad.jpg", + "image_caption": [ + "w/o polarization" + ], + "image_footnote": [], + "bbox": [ + 601, + 88, + 694, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40e5e2dfce30f99dd30c516287cd51007e097a17f1f610b5cf50972e0f4bc386.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 697, + 88, + 792, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d73bf48359244d1c9e4befe2ba5e1fe7a0381cf3cc40df20f3220e78116213e5.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 795, + 88, + 890, + 212 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5e36121b69ee3724364c4917b2296a48cc8805f1a1823854b82483002292b106.jpg", + "table_caption": [ + "Table 3. Ablation study on PolarFree framework." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓Q-align↑
w/o Lphase22.410.86220.14023.8524
w/o polarization21.560.86200.14833.7949
w/o diffusion prior20.210.86270.15523.7949
Ours22.440.86810.13253.8867
", + "bbox": [ + 513, + 321, + 875, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "diffusion model, especially in complex scenes (Fig. 9). The conditional diffusion process significantly improves separation accuracy and preserves finer transmission details.", + "bbox": [ + 496, + 404, + 890, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Phase-based Loss. We explore the impact of our phase-based loss function. We remove this loss from the training procedure and observe the model's performance without it. The results show a noticeable increase in color discrepancies and less accurate reflection removal, particularly in scenarios involving semi-reflective surfaces (Table 3). The phase-based loss function ensures the model focuses on reflection removal rather than compensating for color inconsistencies, leading to improved overall results.", + "bbox": [ + 496, + 460, + 890, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 603, + 617, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose PolaRGB, a large-scale, comprehensive dataset for polarization-based reflection removal. With 6,500 well-aligned RGB-polarization image pairs, PolaRGB is $8 \\times$ larger than existing polarization datasets and is the first to include both RGB and polarization images, uniquely captured across diverse real-world scenes and lighting conditions. Additionally, we present PolarFree, a novel reflection removal model that leverages the generative capabilities of diffusion models. This approach enables PolarFree to generate reflection-free cues, enhancing separation accuracy while preserving fine details in the transmission layer. A phase-based loss function is also introduced to further improve the model's performance. Comprehensive experimental results on realistic scenarios demonstrate the effectiveness our method. We believe these contributions set a foundation for further advancements in polarization-based reflection removal, paving the way for more sophisticated applications in complex real-world environments.", + "bbox": [ + 496, + 623, + 890, + 910 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10897", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 76, + 90, + 235, + 106 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China (No.2022ZD0160201) and Shanghai Artificial Intelligence Laboratory, and the RGC Early Career Scheme (ECS) No. 24209224.", + "bbox": [ + 76, + 106, + 467, + 164 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 179, + 171, + 194 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amit Agrawal, Ramesh Raskar, Shree K Nayar, and Yuanzhen Li. Removing photography artifacts using gradient projection and flash-exposure sampling. ACM Transactions on Graphics, 24(3):828-835, 2005. 1, 2", + "[2] Max Born and Emil Wolf. Principles of optics: electromagnetic theory of propagation, interference and diffraction of light. Elsevier, 2013. 1, 2, 4", + "[3] Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring. Advances in neural information processing systems, 36, 2024. 2, 5", + "[4] Hany Farid and Edward H Adelson. Separating reflections and lighting using independent components analysis. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149), volume 1, pages 262-267. IEEE, 1999. 2, 3", + "[5] Yuanbiao Gou, Boyun Li, Zitao Liu, Songfan Yang, and Xi Peng. Clearer: Multi-scale neural architecture search for image restoration. Advances in neural information processing systems, 33:17129-17140, 2020. 2", + "[6] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted hdr video reconstruction. arXiv preprint arXiv:2406.08204, 2024. 2", + "[7] Byeong-Ju Han and Jae-Young Sim. Reflection removal using low-rank matrix completion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5438-5446, 2017. 2", + "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 5, 6", + "[9] Yuchen Hong, Haofeng Zhong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. L-differ: Single image reflection removal with language-based diffusion model. In Proceedings of the European conference on computer vision (ECCV), 2024. 2", + "[10] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation*. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 7, 8", + "[11] Qiming Hu and Xiaojie Guo. Single image reflection separation via component synergy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13138-13147, 2023. 1, 2, 7, 8", + "[12] Zhiyong Huang, Biao Xiong, Cao Tian, Jing Zhan, Xiang Fei, and Nazaraf Shah. Dust and reflection removal from videos captured in moving car. In 2016 IEEE 13th International Conference on e-Business Engineering (ICEBE), pages 182-187. IEEE, 2016. 1" + ], + "bbox": [ + 78, + 204, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yujin Jeon, Eunsue Choi, Youngchan Kim, Yunseong Moon, Khalid Omer, Felix Heide, and Seung-Hwan Baek. Spectral and polarization vision: Spectro-polarimetric real-world dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22098-22108, 2024. 2", + "[14] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 6", + "[15] Soomin Kim, Yuchi Huo, and Sung-Eui Yoon. Single image reflection removal with physically-based training images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5164-5173, 2020. 1, 2, 8", + "[16] Naejin Kong, Yu-Wing Tai, and Joseph S Shin. A physically-based approach to reflection separation: from physical modeling to constrained optimization. IEEE transactions on pattern analysis and machine intelligence, 36(2):209-221, 2013. 3", + "[17] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14811-14820, 2021. 2", + "[18] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1750-1758, 2020. 1, 2, 3, 4, 7, 8", + "[19] Chenyang Lei, Xudong Jiang, and Qifeng Chen. Robust reflection removal with flash-only cues in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2", + "[20] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(9):1647-1654, 2007. 2", + "[21] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E Hopcroft. Single image reflection removal through cascaded refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3565-3574, 2020. 1, 7, 8", + "[22] Tingtian Li, Yuk-Hee Chan, and Daniel PK Lun. Improved multiple-image-based reflection removal algorithm using deep neural networks. IEEE Transactions on Image Processing, 30:68-79, 2020. 2", + "[23] Xiaobo Li, Lei Yan, Pengfei Qi, Liping Zhang, François Goudail, Tiegen Liu, Jingsheng Zhai, and Haofeng Hu. Polarimetric imaging via deep learning: A review. Remote Sensing, 15(6):1540, 2023. 2", + "[24] Yu Li and Michael S Brown. Single image layer separation using relative smoothness. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2752-2759, 2014. 1, 2", + "[25] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolarized and polarized images. Advances in neural information processing systems, 32, 2019. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10898", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Ajay Nandoriya, Mohamed Elgharib, Changil Kim, Mohamed Hefeeda, and Wojciech Matusik. Video reflection removal through spatio-temporal optimization. In Proceedings of the IEEE International Conference on Computer Vision, pages 2411-2419, 2017. 2", + "[27] Shree K Nayar, Xi-Sheng Fang, and Terrance Boult. Separation of reflection components using color and polarization. International Journal of Computer Vision, 21(3):163-186, 1997. 1, 2, 3, 4", + "[28] Simon Niklaus, Xuaner Cecilia Zhang, Jonathan T Barron, Neal Wadhwa, Rahul Garg, Feng Liu, and Tianfan Xue. Learned dual-view reflection removal. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3713-3722, 2021. 2", + "[29] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization-based decorrelation of transparent layers: The inclination angle of an invisible surface. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 814-819. IEEE, 1999. 1, 4", + "[30] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization and statistical analysis of scenes containing a semireflector. JOSA A, 17(2):276-284, 2000. 1, 2, 3, 4", + "[31] YiChang Shih, Dilip Krishnan, Fredo Durand, and William T Freeman. Reflection removal using ghosting cues. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3201, 2015. 2", + "[32] Sony Semiconductor Solutions. Polarsens: Polarization sensing technology, 2024. Accessed: 2024-11-15. 1", + "[33] David Strong and Tony Chan. Edge-preserving and scale-dependent properties of total variation regularization. Inverse problems, 19(6):S165, 2003. 6", + "[34] Robby T Tan and Katsushi Ikeuchi. Separating reflection components of textured surfaces using a single image. IEEE transactions on pattern analysis and machine intelligence, 27(2):178-193, 2005. 1, 2", + "[35] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Crrn: Multi-scale guided concurrent reflection removal network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4777-4785, 2018. 2", + "[36] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Corrn: Cooperative reflection removal network. IEEE transactions on pattern analysis and machine intelligence, 42(12):2969-2982, 2019. 2", + "[37] Mengyi Wang, Xinxin Zhang, Yongshun Gong, and Yilong Yin. Personalized single image reflection removal network through adaptive cascade refinement. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8204-8213, 2023. 2", + "[38] Kaixuan Wei, Jiaolong Yang, Ying Fu, David Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8178-8187, 2019. 1, 2", + "[39] Patrick Wieschollek, Orazio Gallo, Jinwei Gu, and Jan Kautz. Separating reflection and transmission images in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), pages 89-104, 2018. 1, 2, 3, 4" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 7, 8", + "[41] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T Freeman. A computational approach for obstruction-free photography. ACM Transactions on Graphics (TOG), 34(4):1-11, 2015. 1", + "[42] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 2", + "[43] Jae-Seong Yun and Jae-Young Sim. Reflection removal for large-scale 3d point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4597–4605, 2018. 1", + "[44] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4786-4794, 2018. 2", + "[45] Haofeng Zhong, Yuchen Hong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. Language-guided image reflection separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24913-24922, 2024. 2", + "[46] Yurui Zhu, Xueyang Fu, Peng-Tao Jiang, Hao Zhang, Qibin Sun, Jinwei Chen, Zheng-Jun Zha, and Bo Li. Revisiting single image reflection removal in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25468-25478, 2024. 1, 2, 6, 7, 8" + ], + "bbox": [ + 501, + 90, + 893, + 529 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10899", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PolarFree_ Polarization-based Reflection-Free Imaging/e1567ff5-b726-4ec2-baeb-81cce9b4c3bf_model.json b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/e1567ff5-b726-4ec2-baeb-81cce9b4c3bf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9b9b09d5bc26cf305146ca0c788410423a7ff187 --- /dev/null +++ b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/e1567ff5-b726-4ec2-baeb-81cce9b4c3bf_model.json @@ -0,0 +1,3047 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.208, + 0.131, + 0.764, + 0.154 + ], + "angle": 0, + "content": "PolarFree: Polarization-based Reflection-Free Imaging" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.868, + 0.235 + ], + "angle": 0, + "content": "Mingde Yao\\(^{1}\\), Menglu Wang\\(^{3}\\), King-Man Tam\\(^{1,4}\\), Lingen Li\\(^{1}\\), Tianfan Xue\\(^{\\text{☑}1,2}\\), Jinwei Gu\\(^{1}\\) \n\\(^{1}\\)The Chinese University of Hong Kong, \\(^{2}\\)Shanghai AI Laboratory \n\\(^{3}\\)University of Science and Technology of China, \\(^{4}\\)Institute of Science Tokyo" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.238, + 0.678, + 0.251 + ], + "angle": 0, + "content": "mingdeyao@foxmail.com, tfxue@ie.cuhk.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.281, + 0.314, + 0.295 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.474, + 0.679 + ], + "angle": 0, + "content": "Reflection removal is challenging due to complex light interactions, where reflections obscure important details and hinder scene understanding. Polarization naturally provides a powerful cue to distinguish between reflected and transmitted light, enabling more accurate reflection removal. However, existing methods often rely on small-scale or synthetic datasets, which fail to capture the diversity and complexity of real-world scenarios. To this end, we construct a large-scale dataset, PolaRGB, for Polarization-based reflection removal of RGB images, which enables us to train models that generalize effectively across a wide range of real-world scenarios. The PolaRGB dataset contains 6,500 well-aligned mixed-transmission image pairs, \\(8 \\times\\) larger than existing polarization datasets, and is the first to include both RGB and polarization images captured across diverse indoor and outdoor environments with varying lighting conditions. Besides, to fully exploit the potential of polarization cues for reflection removal, we introduce PolarFree, which leverages diffusion process to generate reflection-free cues for accurate reflection removal. Extensive experiments show that PolarFree significantly enhances image clarity in challenging reflective scenarios, setting a new benchmark for polarized imaging and reflection removal. Code and dataset are available at https://github.com/mdyao/PolarFree." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.692, + 0.21, + 0.708 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Reflection removal algorithms [1, 19, 30, 34, 38, 41] remove unwanted reflections in captured images, playing a critical role in applications such as autonomous driving [12] and photography [11, 43]. This problem commonly arises when imaging through semi-reflectors, like windows or glass, and overlapping reflections may obscure important details of scenes we want to capture. This problem is often formulated [10, 11] as a linear combination of the transmission layer \\( T \\) and the reflection layer \\( R \\):" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.854, + 0.47, + 0.869 + ], + "angle": 0, + "content": "\\[\nM = \\alpha_ {t} T + \\alpha_ {r} R, \\tag {1}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.285, + 0.631, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.371, + 0.584, + 0.383 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.286, + 0.761, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.37, + 0.742, + 0.383 + ], + "angle": 0, + "content": "Kim et al. [11]" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.286, + 0.891, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.37, + 0.868, + 0.383 + ], + "angle": 0, + "content": "DSRNet [15]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.383, + 0.631, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.467, + 0.61, + 0.479 + ], + "angle": 0, + "content": "RDRNet [46]" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.383, + 0.761, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.468, + 0.712, + 0.478 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.383, + 0.891, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.468, + 0.867, + 0.478 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.481, + 0.892, + 0.537 + ], + "angle": 0, + "content": "Figure 1. Our PolarFree effectively leverages polarization information to remove reflections, achieving superior performance in challenging scenes with complex backgrounds and highlights where previous methods [11, 15, 46] often fail." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.547, + 0.891, + 0.577 + ], + "angle": 0, + "content": "where \\(M\\) is the mixed captured image, and \\(\\alpha_{t},\\alpha_{r}\\) are blending coefficients resulting from light attenuation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.807 + ], + "angle": 0, + "content": "Polarization image sensors are becoming mainstream, allowing users to easily capture polarization images from a single shot in real-time [32]. However, existing methods [10, 11, 24, 46] typically rely on intensity-based cues, such as pixel brightness and color gradients, to distinguish transmitted and reflected layers. These methods face challenges because reflection removal is a highly ill-posed inverse problem [21] that recovers two unknown layers (reflection and transmission) from a single observation. Polarization provides valuable physics-based cues [27, 29] to alleviate ill-posed issues (Fig. 1), whereas transmitted light retains a distinct degree of polarization(Fig. 2b). This difference provides crucial signals for separating the two layers. Notably, at the Brewster angle (Fig.2c), reflected light is fully polarized, enabling effective reflection removal [2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Despite significant advances of polarization images, a key challenge in polarization-based reflection removal is the lack of large-scale, high-quality datasets. Existing polarized reflection removal datasets [18, 39] are limited in size and diversity, relying on small (<1000) or synthetic samples that fail to capture the complexity of real-world light-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.234, + 0.9 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "footer", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10890" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.472, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.216, + 0.47, + 0.286 + ], + "angle": 0, + "content": "Figure 2. (a) & (b) A semi-reflector transforms unpolarized light into polarized light upon reflection and refraction, which is undetectable by standard RGB cameras but can be leveraged by polarization cameras for reflection-suppression tasks. (c) At the Brewster angle [2], a polarizer minimizes reflections." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.469, + 0.397 + ], + "angle": 0, + "content": "ing conditions, materials, and scenes. Moreover, they typically exclude color information, reducing their applicability in real-world reflection removal tasks. Thus, there is a pressing need for a large-scale, comprehensive dataset that includes both RGB and polarization images, captured in diverse real-world environments, to advance polarization-based reflection removal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.469, + 0.58 + ], + "angle": 0, + "content": "To bridge this gap, we introduce PolaRGB, a novel dataset specifically collected for polarization-based reflection removal. As shown in Table 1, PolaRGB contains 6,500 high-quality, well-aligned RGB-polarization image pairs, \\(8 \\times\\) larger than previous dataset [18]. Our dataset covers a diverse range of scenes, lighting conditions, and exposure settings, and is captured using off-the-shelf commercial cameras with polarized color patterns to ensure real-world applicability. PolaRGB provides both mixed images and ground-truth transmission layers, enabling accurate reflection separation and significantly enhancing the effectiveness of reflection removal across real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.718 + ], + "angle": 0, + "content": "Moreover, extracting reflection-free information from polarization data is challenging [18, 25, 39] due to the randomness of shooting angles, scene variations, and changing lighting conditions. To address this issue, we leverage the powerful generative capabilities of the diffusion model [3,6,8] to generate reflection-free cues. The diffusion model extracts and refines reflection-free priors from polarization images, effectively guiding the reflection removal and yielding precise and robust reflection-free results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Specifically, PolarFree consists of two steps: a prior-generation step and a reflection removal step. First, in the prior-generation step, we use a diffusion model to generate reflection-free prior based on the polarization and RGB inputs. This strategy not only guides accurate reflection isolation in RGB images but also recovers background details, which previous methods [11, 15, 46] may miss. Next, the reflection removal step leverages the prior to effectively remove reflections, ensuring accurate transmission restoration. Additionally, we introduce a phase-based loss function in the frequency domain to mitigate color discrepancies caused by semi-reflections, guiding the network to focus" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.116 + ], + "angle": 0, + "content": "Table 1. Comparisons between PolaRGB and existing reflection removal datasets." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.117, + 0.892, + 0.202 + ], + "angle": 0, + "content": "
DatasetPolarizationRGBRAWData sizeResolution
SIR2 [36]NoYesNo500540 × 400
RRW [46]NoYesNo14,9522580 × 1460
ReflectNet [39]Yes (Syn.)YesNo-1500 × 1000
Lei et al. [18]Yes (Real)NoYes8071224 × 1024
PolaRGB (Ours)Yes (Real)YesYes6,5001224 × 1024
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.892, + 0.268 + ], + "angle": 0, + "content": "on reflection removal rather than color adjustment. These components enable PolarFree to achieve robust reflection suppression while preserving the clarity and integrity of the transmission across diverse real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.892, + 0.406 + ], + "angle": 0, + "content": "Extensive experiments on the PolaRGB dataset demonstrate the effectiveness of PolarFree, which outperforms existing methods [11, 15, 46] \\(\\sim\\) 2dB in terms of PSNR. Additionally, we performed real-world testing in environments commonly affected by reflections such as museums and galleries, demonstrating the method's effectiveness in real-world scenarios. Our approach improves image clarity and preserves details better than previous methods, paving the way for more practical, reflection-robust applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.407, + 0.642, + 0.422 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.425, + 0.686, + 0.439 + ], + "angle": 0, + "content": "2.1. Reflection Removal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.638 + ], + "angle": 0, + "content": "Reflection removal [1, 30, 34] has a longstanding history in computer vision, aiming to enhance the occluded objects caused by semi-reflectors like glass and windows. Early approaches [20, 24] typically relied on handcrafted priors, such as smoothness assumptions and gradient sparsity, to differentiate the reflection layer from the underlying scene. Some methods incorporated ghosting [31], flash cues [17, 19] and edge consistency [7], while others improved accuracy by leveraging multi-image setups [22, 28], utilizing temporal variation [26] and parallax [28] to separate reflections from background layers. However, these methods often require strict assumptions and fail in real-world scenes with complex textures or lighting [19, 38]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.805 + ], + "angle": 0, + "content": "With the rise of deep learning [5, 35, 42, 44, 46], reflection removal methods advanced considerably. Techniques based on neural networks have shown promise by learning to separate reflection and transmission layers directly from data [10, 15, 35-37, 44]. However, without essential physical insights, these methods act as Bayesian regressions toward a learned average, often lead to suboptimal separation in complex environments with varying lighting, textures, and reflection intensities [35, 44]. Recent methods [9, 45] utilize language to interactively remove the reflection, which require additional efforts, limiting their practicality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.809, + 0.783, + 0.823 + ], + "angle": 0, + "content": "2.2. Polarization Reflection Removal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Polarization is an effective way to remove reflections due to the physics principle: the polarization of light behaves differently in the reflection and transmission layers [4, 13, 23, 27, 30], allowing separation, especially at the Brewster angle [4, 27]. However, capturing images exactly" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10891" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.088, + 0.45, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.357, + 0.398, + 0.369 + ], + "angle": 0, + "content": "(a) Hierarchical structure of PolarRR dataset" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.089, + 0.565, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.089, + 0.665, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.15, + 0.565, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.15, + 0.665, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.211, + 0.654, + 0.223 + ], + "angle": 0, + "content": "(b) Diverse reflection conditions" + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.089, + 0.836, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.211, + 0.871, + 0.223 + ], + "angle": 0, + "content": "(c) Video-based capture method" + }, + { + "type": "image", + "bbox": [ + 0.838, + 0.089, + 0.888, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.232, + 0.571, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.232, + 0.676, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.232, + 0.782, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.232, + 0.888, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.293, + 0.571, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.293, + 0.677, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.293, + 0.782, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.293, + 0.888, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.356, + 0.786, + 0.37 + ], + "angle": 0, + "content": "(d) Image types included in the dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.373, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Figure 3. Overview of the PolaRGB dataset. (a) Hierarchical structure of scenes is shown in the ring, with legends indicating sample counts and subset types. (b) Typical scenes illustrating varied reflection conditions: I. smooth blending of reflection and refraction, II. abrupt reflection with mixed components, III. reflection dominant over transmission, and IV. minimal or no reflection. (c) Video-based capture method (details in Sec. 3). (d) We provide polarized images at angles \\(\\phi = 0^{\\circ}\\), \\(45^{\\circ}\\), \\(90^{\\circ}\\), and \\(135^{\\circ}\\), along with derived AoLP, DoLP, and a well-aligned unpolarized image. The dataset also includes ground truth transmission and estimated reflections, all available in both raw and RGB formats." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.458, + 0.469, + 0.519 + ], + "angle": 0, + "content": "at the Brewster angle, as shown in Fig. 2c, is challenging. Therefore, in practice, methods typically rely on images taken at multiple polarization angles (e.g., \\(0^{\\circ}\\), \\(45^{\\circ}\\), \\(90^{\\circ}\\), and \\(135^{\\circ}\\)) to capture polarization information [4, 27, 30]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Early polarized reflection removal methods [4, 27, 30] employ mathematical techniques like Principle Component Analysis (PCA) to separate reflection layers, which are effective when the transmission and reflection layers have significant content differences. Subsequently, Kong et al. [16] propose a multiscale scheme that automatically identifies the optimal separation of the reflection and background layers. Another advancements [39] introduce a neural network-based approach for polarization-guided reflection removal with synthetic dataset. Lyu et al. [25] utilize a pair of unpolarized and polarized images but lack of realistic polarization patterns. Lei et al. [18] further contribute by collecting a dataset of polarized images for reflection removal. However, this dataset only includes limited scene variations from pure polarization scenes, which limits its application scenario. In contrast, we present the first large-scale dataset captured for polarization-based RGB reflection removal, consisting of 6,500 well-aligned transmission-reflection pairs. We also propose a polarization-based reflection removal network based on diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.827, + 0.249, + 0.842 + ], + "angle": 0, + "content": "3. PolaRGB Dataset" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.849, + 0.18, + 0.866 + ], + "angle": 0, + "content": "3.1. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Although a few polarization-based reflection removal datasets exist [18, 39], they face major limitations: (1) they" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.563 + ], + "angle": 0, + "content": "are often synthetically generated [25, 39], which may not generalize well to real polarization, as perfectly simulating polarization phenomena is very hard; (2) they are typically limited in size and diversity (<1,000 samples), with limited scene variations, which constrains model robustness and generalizability [18, 39]; and (3) they lack RGB data [18, 39], reducing the practical value of these datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.893, + 0.805 + ], + "angle": 0, + "content": "Our polarized-based reflection removal dataset, PolaRGB, addresses all these limitations with the following key features, as shown in Fig. 3. It has following advantages: (1) It contains real-captured images with perfect spatial alignment. We achieve pixel-level alignment through a three-step process: careful capture setup, manual filtering, and homography transformations, all in raw space to avoid demosaicing artifacts (see Fig. 4b and more details in supplementary). (2) It contains large-scale diverse scenes. The collected dataset covers a wide range of indoor and outdoor settings with four distinct reflection types, including smooth, sharp, high-brightness, and subtle reflections (see Fig. 3b). (3) It has extensive data modalities. Our dataset offers paired mixed and transmission images with both polarization and RGB captures, precisely aligned and available in both raw and RGB formats (see Fig. 3d)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.809, + 0.685, + 0.825 + ], + "angle": 0, + "content": "3.2. Processing Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We leverage an efficient video-based capture flow for dataset collection, as shown in Fig. 3c. First, we capture a transmission-only image \\( T_{raw} \\) as the ground truth. Then, we place a semi-reflective glass plate in front of the scene and rotate it continuously to capture mixed images" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10892" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.52, + 0.47, + 0.548 + ], + "angle": 0, + "content": "Figure 4. Data processing pipeline for obtaining aligned mixed and transmission images, and polarized images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.552, + 0.47, + 0.597 + ], + "angle": 0, + "content": "\\(M_{\\text{raw}}\\). A division-of-focal-plane polarization camera with color Bayer pattern is used to simultaneously capture both polarized and RGB information, as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.84 + ], + "angle": 0, + "content": "After obtaining the raw mixed image \\( M_{raw} \\) and the transmission image \\( T_{raw} \\), we process the images sequentially in the raw and RGB domains as shown in Fig. 4a. In the raw domain, we first align \\( M_{raw} \\) and \\( T_{raw} \\) to correct spatial misalignment caused by light refraction. This is done by separating the images into different polarization angles and color channels, then applying affine transformation matrices to each of the channels to avoid aliasing from direct alignment, as shown in Fig. 4b. Next, we perform polarization separation on the aligned images to obtain four polarization images \\( (0^{\\circ}, 45^{\\circ}, 90^{\\circ} \\), and \\( 135^{\\circ} \\)), and use Malus's law [27, 29] to sum them, producing the unpolarized image of the scene. Finally, we estimate the reflection image by searching for the optimal blending coefficient \\( \\alpha_r \\) and \\( \\alpha_t \\) in edge space, as illustrated in Eq. 1. For detailed steps and proof, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Next, we apply demosaicking to the processed raw images to obtain the mixed, transmission, and reflection images, each consisting of four polarization images and the unpolarized RGB image. Additionally, we compute the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.136 + ], + "angle": 0, + "content": "Stokes parameters (see Sec. 4.1) to estimate the Angle of Linear Polarization (AOLP) and Degree of Linear Polarization (DOLP), which are then used in reflection removal." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.139, + 0.592, + 0.154 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.16, + 0.704, + 0.175 + ], + "angle": 0, + "content": "4.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.18, + 0.892, + 0.317 + ], + "angle": 0, + "content": "Object. Given an RGB image and its corresponding polarization images, we aim to recover the transmission layer of the RGB image by utilizing the distinct polarization characteristics. Our measurement consists of spatially-aligned RGB and polarization images, where the polarization images captured at four distinct angles \\((0^{\\circ}, 45^{\\circ}, 90^{\\circ}\\), and \\(135^{\\circ})\\). These four angles provide a comprehensive polarization measurement [30], which allows us to compute polarization features crucial for separating layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.388 + ], + "angle": 0, + "content": "Polarization Preliminary. For semi-reflective surfaces that produce mixed images, the observed intensity \\( I_M^\\phi (x) \\) can be decomposed into reflected intensity \\( I_R^\\phi (x) \\) and transmitted intensity \\( I_T^\\phi (x) \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.389, + 0.891, + 0.408 + ], + "angle": 0, + "content": "\\[\nI _ {M} ^ {\\phi} (x) = \\alpha_ {R} I _ {R} ^ {\\phi} (x) + \\alpha_ {T} I _ {T} ^ {\\phi} (x). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.44 + ], + "angle": 0, + "content": "If a linear polarizer is placed in front of the camera at an angle \\(\\phi\\), the captured intensity \\(I^{\\phi}(x)\\) can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.45, + 0.892, + 0.469 + ], + "angle": 0, + "content": "\\[\nI _ {M} ^ {\\phi} (x) = \\alpha (\\theta ; \\phi ; \\phi_ {\\perp}) I _ {R} ^ {\\phi} (x) + (1 - \\alpha (\\theta ; \\phi ; \\phi_ {\\parallel})) I _ {T} ^ {\\phi} (x), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.659 + ], + "angle": 0, + "content": "where \\(\\alpha (\\theta ;\\phi ;\\phi_{\\perp})\\) and \\((1 - \\alpha (\\theta ;\\phi ;\\phi_{\\parallel}))\\) are polarization-dependent mixing coefficients based on the angle of incidence \\(\\theta ,\\phi_{\\perp}\\) and \\(\\phi_{\\parallel}\\) are the canonical polarization directions for the reflection and transmission, respectively. Directly solving for these parameters is challenging because \\(\\alpha (\\cdot)\\), \\(I_R^\\phi (x)\\), and \\(I_R^\\phi (x)\\) are all unknown [27, 39]. Additionally, relationship between the observed intensity and the underlying reflected and transmitted components is highly nonlinear and influenced by various factors [39]. To address this, we utilize Stokes parameters [18] that provide a efficient way to represent and analyze polarized light, enabling more robust separation of reflection and transmission." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Stokes Parameters. To capture the polarization effects in the scene, we use Stokes parameters \\([S_0, S_1, S_2]\\), which can be derived from intensity measurements at specific polarization angles \\((0^\\circ, 45^\\circ, 90^\\circ\\), and \\(135^\\circ\\)):" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.728, + 0.892, + 0.762 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} S _ {0} = \\left(I _ {0 ^ {\\circ}} + I _ {4 5 ^ {\\circ}} + I _ {9 0 ^ {\\circ}} + I _ {1 3 5 ^ {\\circ}}\\right) / 2, (4) \\\\ S _ {1} = I _ {0 ^ {\\circ}} - I _ {9 0 ^ {\\circ}}, S _ {2} = I _ {4 5 ^ {\\circ}} - I _ {1 3 5 ^ {\\circ}}. (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Here, \\( S_0 \\) represents the total intensity of light, and \\( S_1 \\) and \\( S_2 \\) provide information about the linear polarization state of the light based on intensity differences at these key angles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Using the Stokes parameters, we compute the Degree of Linear Polarization (DOLP) and Angle of Linear Polarization (AOLP) as follows [2]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.856, + 0.892, + 0.887 + ], + "angle": 0, + "content": "\\[\nD O L P (x) = \\sqrt {S _ {1} (x) ^ {2} + S _ {2} (x) ^ {2}} / S _ {0} (x), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.871, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\nA O L P (x) = \\frac {1}{2} \\operatorname {a t a n 2} \\left(S _ {2} (x) / S _ {1} (x)\\right). \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10893" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.418, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.354, + 0.301, + 0.366 + ], + "angle": 0, + "content": "(a) Inference" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.089, + 0.885, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.354, + 0.705, + 0.366 + ], + "angle": 0, + "content": "(b) Training" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.369, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Figure 5. Pipeline of PolarFree. (a) During inference, PolarFree leverages polarized and RGB images as inputs, which are feeds into a conditional diffusion model to generate the prior \\(\\hat{z}_0\\). The generated prior, along with the inputs, is then passed to the reflection removal backbone \\(\\mathcal{F}_{\\mathrm{remove}}\\) to remove reflections. (b) PolarFree is trained in two stages. (1) A prior encoder extracts a reflection-free prior \\(z_0\\) from clean transmission images and polarization cues, which serves as the supervision for the conditional diffusion model in stage two. (2) The conditional diffusion model is trained to progressively denoise noisy images, supervised by the prior from stage one, ensuring robust reflection separation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.593 + ], + "angle": 0, + "content": "In this formulation, DOLP describes the proportion of polarized light relative to total intensity, which helps indicate the degree of reflection in the scene. AOLP, on the other hand, reveals the orientation of the polarized light, allowing us to distinguish between reflection and transmission components more effectively, as shown in Fig. 3d. By leveraging intensity measurements at \\(0^{\\circ}\\), \\(45^{\\circ}\\), \\(90^{\\circ}\\), and \\(135^{\\circ}\\), these parameters provide valuable cues for separating mixed reflection and transmission layers in semi-reflective scenes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.596, + 0.262, + 0.609 + ], + "angle": 0, + "content": "4.2. PolarFree Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.471, + 0.793 + ], + "angle": 0, + "content": "To achieve high-quality reflection removal with effective utilization of polarization information, we introduce Polar-Free, a delicately designed two-step network, where each step addresses a distinct aspect of the reflection removal challenge. Inspired by [3], we leverage diffusion model to generate the prior for reflection removal. As shown in Fig. 5, during inference, the first step utilizes a conditional diffusion model to extract reflection-free priors, effectively isolating essential details from polarization data. The second step integrates these priors with RGB inputs, guiding the network to accurately separate reflections and enhance clarity, even in complex, real-world environments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.8, + 0.47, + 0.845 + ], + "angle": 0, + "content": "Prior Generation. As shown in Fig. 5a, in the first step, we start with a randomly initialized noise \\( n \\), which is gradually denoised through a conditional diffusion model \\( \\mathcal{F}_{\\mathrm{diff}} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.849, + 0.469, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {0} = \\mathcal {F} _ {\\text {d i f f}} (n | M _ {\\text {c o n d}}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "where \\(\\hat{z}_0\\) is the generated prior, \\(M_{\\mathrm{cond}} = \\{M_{\\mathrm{polar}}, M_{\\mathrm{aolp}}, M_{\\mathrm{dolp}}, M_{\\mathrm{rgb}}\\}\\) represents the mixed im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.892, + 0.519 + ], + "angle": 0, + "content": "ages, and \\( n \\) is the initial noise. Through an iterative denoising process, the diffusion model progressively refines a noise image, conditioned on the polarization and RGB data, to generate a reflection prior." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.611 + ], + "angle": 0, + "content": "The denoising process follows the denoising diffusion probabilistic model (DDPM) framework [8] and employs a U-Net architecture to predict noise. At each timestep \\( t \\), the U-Net receives a noisy intermediate \\( z_{t} \\) and outputs a noise estimate \\( \\epsilon_{\\theta}(z_{t},t) \\), predicting the noise added at that timestep. This process can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.611, + 0.892, + 0.645 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(z _ {t}, M _ {\\text {c o n d}}, t\\right)\\right) + \\sigma_ {t} z, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.891, + 0.707 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) and \\(\\beta_{t}\\) control the noise schedule across the timesteps, \\(\\sigma_{t}\\) is the standard deviation, and \\(z\\) is noise sampled from a standard Gaussian distribution. \\(\\bar{\\alpha}_{t}\\) is the cumulative product of \\(\\alpha_{t}\\), indicating noise level to timestep \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.892, + 0.785 + ], + "angle": 0, + "content": "In this way, the U-Net gradually removes the noise in each step, conditioning on the polarization-based measurement input \\( M_{\\mathrm{cond}} \\). This process continues until noise is fully removed, yielding a sample \\( \\hat{z}_0 \\) that represents the reflection-free prior distribution extracted from the conditioned inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Reflection Removal. As shown in Fig. 5a, once the prior \\(\\hat{z}_0\\) has been obtained, the second step of PolarFree is separating the transmission and reflection layers. By leveraging the polarization cues provided by \\(M_{\\mathrm{polar}}\\), \\(M_{\\mathrm{aolp}}\\), and \\(M_{\\mathrm{dolp}}\\), the model \\(\\mathcal{F}_{\\mathrm{remove}}\\) separates the transmission features from the reflection ones, which can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.885, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {T} _ {r g b} = \\mathcal {F} _ {\\text {r e m o v e}} \\left(\\hat {z} _ {0}, M _ {\\text {c o n d}}\\right), \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10894" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.445, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.22, + 0.468, + 0.287 + ], + "angle": 0, + "content": "Figure 6. (a) Phase information preserves shape and texture details, while color primarily affects the amplitude. (b) We apply two types of random color perturbations to the image and compute the perturbation errors. Phase-based loss is less sensitive to color changes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.468, + 0.352 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\mathrm{remove}}\\) is a reflection removal neural network. This step ensures that the final output contains distinct transmission and reflection components, facilitating high-quality image reconstruction under challenging reflective scenarios." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.359, + 0.182, + 0.374 + ], + "angle": 0, + "content": "4.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.468, + 0.47 + ], + "angle": 0, + "content": "To train our PolarFree, particularly to address the challenge of the diffusion model lacking a suitable ground truth, we adopt a two-stage training strategy, as illustrated in Fig. 5b. This strategy consists of two sequential objectives: learning to extract a reflection-free prior and learning to generate a reflection-free prior." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.468, + 0.629 + ], + "angle": 0, + "content": "First Stage. In the first stage, we train an encoder to extract a prior \\( z_0 \\) for reflection-free information, which will serve as the supervision signal for the diffusion model in the second stage. Specifically, we feed the ground truth transmission images \\( T_{\\mathrm{cond}} = \\{T_{\\mathrm{polar}}, T_{\\mathrm{aolp}}, T_{\\mathrm{dolp}}, T_{\\mathrm{rgb}}\\} \\) into the encoder \\( \\mathcal{E} \\), obtaining \\( z = \\mathcal{E}(M_{\\mathrm{cond}}, T_{\\mathrm{rgb}}) \\). This \\( z \\) contains reflection-free information enriched with polarization-related cues. Subsequently, \\( z_0 \\) conditions the reflection removal backbone \\( \\mathcal{F}_{\\mathrm{remove}} \\) to predict the clean transmission image \\( \\hat{T}_{\\mathrm{rgb}} = \\mathcal{F}_{\\mathrm{remove}}(z, M_{\\mathrm{cond}}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.636, + 0.468, + 0.741 + ], + "angle": 0, + "content": "Second Stage. In the second stage, we train the diffusion model \\(\\mathcal{F}_{\\mathrm{diff}}\\) to generate a reflection-free prior \\(z_0\\) from noisy input images, and finetune the reflection removal backbone \\(F_{\\mathrm{remove}}\\). The key challenge here is that the diffusion model lacks of direct supervision. Therefore, we leverage the prior \\(z_0\\) extracted in the first stage as the supervision signal to guide the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.468, + 0.802 + ], + "angle": 0, + "content": "We start with the extracted reflection-free prior \\( z_0 \\) and add noise over multiple timesteps. This process transforms the \"ground-truth\" prior \\( z_0 \\) into noisy versions \\( z_t \\) at each timestep \\( t \\), which can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.804, + 0.468, + 0.821 + ], + "angle": 0, + "content": "\\[\nq \\left(z _ {t} \\mid z _ {0}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {0}, \\beta_ {t} \\mathbf {I}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.824, + 0.468, + 0.869 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}\\) represents a Gaussian distribution, \\(\\beta_{t}\\) is the noise schedule controlling the noise level added at each timestep, \\(\\mathbf{I}\\) represents the identity matrix." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In the reverse diffusion process, the model is trained to progressively remove the noise step by step to recover the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.093, + 0.785, + 0.106 + ], + "angle": 0, + "content": "clean prior \\(z_0\\), which can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.107, + 0.89, + 0.124 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t, M _ {\\text {c o n d}}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right)), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.127, + 0.89, + 0.157 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta}(z_t,t,M_{\\mathrm{cond}})\\) is the mean function of the current state \\(z_{t}\\), timestep \\(t\\), and a conditional input \\(M_{\\mathrm{cond}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.158, + 0.89, + 0.233 + ], + "angle": 0, + "content": "Throughout the reverse process, the model conditions on the polarization measurements \\( M_{\\mathrm{cond}} \\), which provide essential information about the scene's physical properties. These measurements help the model accurately generate reflection-free components during the denoising process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.891, + 0.325 + ], + "angle": 0, + "content": "After the reverse process completes, the model outputs a final reflection-free prior \\(\\hat{z}_0\\), which contains clean transmission information. This prior is used as the guidance signal for the reflection removal backbone \\(\\mathcal{F}_{\\mathrm{remove}}\\). The backbone takes this reflection-free prior and the polarization data to predict the clean transmission image \\(\\hat{T}_{rgb}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.328, + 0.586, + 0.342 + ], + "angle": 0, + "content": "4.4. Losses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.517 + ], + "angle": 0, + "content": "Basic Loss. To optimize PolarFree, we follow [46] to utilize three basic losses: L1 loss, VGG perceptual loss [14], and total variation (TV) loss [33]. The L1 loss minimizes the pixel-wise difference between the predicted transmission image \\(\\hat{T}_{\\mathrm{rgb}}\\) and the ground truth transmission image \\(T_{\\mathrm{rgb}}\\) as \\(\\mathcal{L}_1 = \\|\\hat{T}_{\\mathrm{rgb}} - T_{\\mathrm{rgb}}\\|_1\\). The VGG perceptual loss compares feature activations from a pre-trained VGG network [14], weighted by \\(\\lambda_l\\): \\(\\mathcal{L}_{\\mathrm{VGG}} = \\sum_l \\lambda_l \\|\\phi_l(\\hat{T}_{\\mathrm{rgb}}) - \\phi_l(T_{\\mathrm{rgb}})\\|_1\\), where \\(\\phi_l\\) denotes the activations of the \\(l\\)-th VGG layer. We also use TV loss to constrain consistency via the gradient operator \\(\\mathcal{L}_{\\mathrm{TV}} = \\|\\nabla \\hat{T}_{\\mathrm{rgb}} - \\nabla T_{\\mathrm{rgb}}\\|_1\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.891, + 0.643 + ], + "angle": 0, + "content": "Phase loss. While basic losses help the network match color and intensity values to the ground truth, they struggle with color discrepancies caused by semi-reflective surfaces during dataset collection. These variations, due to reflection and transmission properties of materials, affect the color and intensity, leading to mismatches between the predicted and target images and hindering the model's ability to learn the correct transmission map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.891, + 0.732 + ], + "angle": 0, + "content": "To address this issue, we introduce a phase loss to focus on the structural information of the transmission, which is less sensitive to color variations. As shown in Fig. 6, phase information primarily captures the geometry and texture of the image, independent of color changes. The phase loss is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.734, + 0.89, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p h a s e}} = \\| \\angle (F F T (\\hat {T})) - \\angle (F F T ((T _ {r g b})) \\| _ {1}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.754, + 0.89, + 0.783 + ], + "angle": 0, + "content": "where \\( FFT \\) is the Fourier transform, and \\( \\angle (\\cdot) \\) represents the phase angle of the Fourier coefficients." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.89, + 0.849 + ], + "angle": 0, + "content": "Diffusion Loss. The diffusion loss follows the standard DDPM formulation [8], where the model predicts the noise at each step \\( t \\) and minimizes the difference between the predicted noise and the true noise as" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.85, + 0.89, + 0.869 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i f f} = \\mathbb {E} _ {q \\left(z _ {t} \\mid z _ {t - 1}\\right)} \\left[ \\| \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) - \\epsilon_ {\\text {t r u e}} \\left(z _ {t}\\right) \\| _ {2} ^ {2} \\right]. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.89, + 0.901 + ], + "angle": 0, + "content": "Here, \\(\\epsilon_{true}(z_t)\\) is the added noise obtained by Eq. 9, and \\(\\epsilon_{\\theta}(z_t,t)\\) is the predicted noise, conditioned on the polariza" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10895" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.21, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.089, + 0.348, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.089, + 0.483, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.089, + 0.62, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.089, + 0.756, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.089, + 0.892, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.177, + 0.21, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.177, + 0.348, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.177, + 0.483, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.177, + 0.62, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.177, + 0.756, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.177, + 0.892, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.266, + 0.21, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.266, + 0.348, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.266, + 0.483, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.266, + 0.62, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.266, + 0.756, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.266, + 0.892, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.353, + 0.21, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.442, + 0.157, + 0.455 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.353, + 0.348, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.441, + 0.326, + 0.455 + ], + "angle": 0, + "content": "Kim et al. [15]" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.353, + 0.483, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.441, + 0.46, + 0.455 + ], + "angle": 0, + "content": "DSRNet [11]" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.353, + 0.62, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.441, + 0.596, + 0.456 + ], + "angle": 0, + "content": "RDRNet [46]" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.353, + 0.756, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.442, + 0.706, + 0.454 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.353, + 0.892, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.784, + 0.441, + 0.868, + 0.454 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.456, + 0.66, + 0.47 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons on the PolaRGB dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.476, + 0.355, + 0.491 + ], + "angle": 0, + "content": "tion measurements and RGB input \\(M_{\\mathrm{cond}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.497, + 0.469, + 0.527 + ], + "angle": 0, + "content": "Total Loss. The above losses are weighted summed to serve as the supervision for the first stage as" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.528, + 0.469, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t a g e 1}} = \\gamma_ {1} \\mathcal {L} _ {1} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {V G G}} + \\gamma_ {3} \\mathcal {L} _ {\\mathrm {T V}} + \\gamma_ {4} \\mathcal {L} _ {\\text {p h a s e}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.547, + 0.469, + 0.577 + ], + "angle": 0, + "content": "In the second stage, we use a combined loss function consisting of the diffusion loss and the reconstruction loss as" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.582, + 0.469, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t a g e 2}} = \\gamma_ {5} \\mathcal {L} _ {\\text {d i f f}} + \\gamma_ {6} \\mathcal {L} _ {\\text {r e c o n}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.599, + 0.469, + 0.644 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{recon}}\\) has the same format as \\(\\mathcal{L}_{\\mathrm{stage1}}\\), and \\(\\gamma (\\cdot)\\) are the coefficients. This combined loss helps refine the reflection-free image while ensuring it aligns with the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.65, + 0.21, + 0.666 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.67, + 0.295, + 0.685 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.839 + ], + "angle": 0, + "content": "We train and evaluate PolarFree on the PolaRGB dataset. The whole dataset consists of 67 scenes, with a total of 6,500 paired images. For each scene, we keep the background (transmission) and camera fixed, and varying the glass position to capture images with reflections. The 67 scenes are then randomly split into 56 training scenes and 11 testing scenes, containing 6,312 and 188 paired images, respectively. This division ensures no data leakage between the training and testing sets, with each set containing only a subset of all categories." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We also test PolarFree in real-world scenes like museums and galleries, where ground truth is unavailable. We implement PolarFree using PyTorch on a single NVIDIA RTX 4090 GPU. Training is conducted with a batch size of" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.475, + 0.891, + 0.505 + ], + "angle": 0, + "content": "2 and an AdamW optimizer with a learning rate of \\( 2 \\times 10^{-4} \\), 30k iterations on the PolaRGB dataset for both stages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.643 + ], + "angle": 0, + "content": "We compare our method with recent advanced reflection removal methods, including Lei et al. [18], IBCLN [21], DSRNet [11], YTMT [10], and RDRNet [46]. For a fair comparison, we modify the input settings to align with ours and use only the transmission layer for supervision. We have re-trained baseline methods on the PolaRGB dataset. The evaluation is conducted using both objective metrics (PSNR, SSIM), a perceptual metric (LPIPS), and a language-based non-reference metric (Q-Align [40])." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.593, + 0.662 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.861 + ], + "angle": 0, + "content": "Results on PolaRGB. We present the quantitative results in Tab. 2. It can be seen that, our method outperforms other methods across multiple metrics, demonstrating the effectiveness of PolarFree for high-quality reflection removal. Notably, for fairness, we modified the input layer of all baselines to accept polarization information. Visual results in Fig. 7 show that PolarFree provides cleaner reflection removal with sharper edges and better color preservation. In contrast, previous methods often suffer from color distortions or imperfect reflection removal, especially in areas with complex reflections. Our method maintains high fidelity to the ground truth, especially in challenging regions with low light and subtle reflections." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Real-captured without Ground Truth. To demonstrate the generalization ability to unseen and more complex re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10896" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.174, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.217, + 0.142, + 0.228 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.089, + 0.273, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.217, + 0.256, + 0.229 + ], + "angle": 0, + "content": "DSRNet [46]" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.089, + 0.371, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.217, + 0.361, + 0.229 + ], + "angle": 0, + "content": "Kim et al. [15]" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.089, + 0.469, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.217, + 0.434, + 0.228 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.23, + 0.468, + 0.258 + ], + "angle": 0, + "content": "Figure 8. Visual comparison of reflection removal results in real-world museum scenes with glass display cases." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.262, + 0.469, + 0.331 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on the PolaRGB dataset, evaluated using objective metrics (PSNR, SSIM), perceptual metric (LPIPS), and the Language-based non-reference metric (Q-align [40]). Best results are in bold, second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.336, + 0.466, + 0.461 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓Q-align↑
Lei et al. [18]18.730.79620.38043.2109
Kim et al. [15]20.670.83990.27143.7148
IBCLN [21]19.730.81730.24883.0938
YTMT [10]16.860.75440.44893.0938
DSRNet [11]16.840.79130.28283.6992
RDRNet [46]15.880.69640.52502.9531
Ours22.440.86810.13253.8867
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.469, + 0.603 + ], + "angle": 0, + "content": "flection, We further evaluate our model on real-captured images at museum with complex reflections from glass enclosures. Here, we only provide qualitative results to visually assess the model's capability. As shown in Fig. 8, our approach effectively reduces reflections while preserving fine details, despite the complex lighting and material variations often present in museum environments. This demonstrates the practical robustness of our method and its potential for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.605, + 0.231, + 0.621 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.469, + 0.669 + ], + "angle": 0, + "content": "We perform ablation studies to evaluate the effectiveness of key components in our method. More experimental results and analyses can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.469, + 0.82 + ], + "angle": 0, + "content": "Polarization Information. To assess the importance of polarization cues, we conduct an experiment where the polarization information (AOLP, DOLP, and polarization images) is removed, and the model is trained solely using RGB data. As shown in Table 3 and Fig. 9, the absence of polarization significantly impairs the model's ability to differentiate between reflection and transmission layers, leading to noticeably poorer reflection removal performance. These results demonstrate the critical role of polarization in accurate reflection removal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Diffusion Prior. We also evaluate the effectiveness of diffusion prior by removing the conditional diffusion model. While without diffusion prior achieves reasonable results, as indicated by the comparison metrics in Table 3, it lacks fine-grained control and high-quality output offered by the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.598, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.215, + 0.585, + 0.225 + ], + "angle": 0, + "content": "w/o diffusion" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.089, + 0.696, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.215, + 0.688, + 0.225 + ], + "angle": 0, + "content": "w/o polarization" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.089, + 0.793, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.215, + 0.76, + 0.224 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.796, + 0.089, + 0.891, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.215, + 0.879, + 0.224 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.31 + ], + "angle": 0, + "content": "Figure 9. Highly challenging reflection removal scenes with complex reflections and highlights. Top: Polarization images from different angles provide complementary information for effective reflection removal, note the third \\((90^{\\circ})\\) image. Bottom: The diffusion module effectively handles highlights, while polarization better restores the color and details in such challenging scenarios." + }, + { + "type": "table_caption", + "bbox": [ + 0.549, + 0.311, + 0.842, + 0.323 + ], + "angle": 0, + "content": "Table 3. Ablation study on PolarFree framework." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.323, + 0.877, + 0.399 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓Q-align↑
w/o Lphase22.410.86220.14023.8524
w/o polarization21.560.86200.14833.7949
w/o diffusion prior20.210.86270.15523.7949
Ours22.440.86810.13253.8867
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.451 + ], + "angle": 0, + "content": "diffusion model, especially in complex scenes (Fig. 9). The conditional diffusion process significantly improves separation accuracy and preserves finer transmission details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.597 + ], + "angle": 0, + "content": "Phase-based Loss. We explore the impact of our phase-based loss function. We remove this loss from the training procedure and observe the model's performance without it. The results show a noticeable increase in color discrepancies and less accurate reflection removal, particularly in scenarios involving semi-reflective surfaces (Table 3). The phase-based loss function ensures the model focuses on reflection removal rather than compensating for color inconsistencies, leading to improved overall results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.619, + 0.619 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.911 + ], + "angle": 0, + "content": "In this work, we propose PolaRGB, a large-scale, comprehensive dataset for polarization-based reflection removal. With 6,500 well-aligned RGB-polarization image pairs, PolaRGB is \\(8 \\times\\) larger than existing polarization datasets and is the first to include both RGB and polarization images, uniquely captured across diverse real-world scenes and lighting conditions. Additionally, we present PolarFree, a novel reflection removal model that leverages the generative capabilities of diffusion models. This approach enables PolarFree to generate reflection-free cues, enhancing separation accuracy while preserving fine details in the transmission layer. A phase-based loss function is also introduced to further improve the model's performance. Comprehensive experimental results on realistic scenarios demonstrate the effectiveness our method. We believe these contributions set a foundation for further advancements in polarization-based reflection removal, paving the way for more sophisticated applications in complex real-world environments." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10897" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.236, + 0.107 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.107, + 0.468, + 0.165 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China (No.2022ZD0160201) and Shanghai Artificial Intelligence Laboratory, and the RGC Early Career Scheme (ECS) No. 24209224." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.18, + 0.173, + 0.195 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[1] Amit Agrawal, Ramesh Raskar, Shree K Nayar, and Yuanzhen Li. Removing photography artifacts using gradient projection and flash-exposure sampling. ACM Transactions on Graphics, 24(3):828-835, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.263, + 0.472, + 0.304 + ], + "angle": 0, + "content": "[2] Max Born and Emil Wolf. Principles of optics: electromagnetic theory of propagation, interference and diffraction of light. Elsevier, 2013. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.305, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[3] Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring. Advances in neural information processing systems, 36, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.362, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[4] Hany Farid and Edward H Adelson. Separating reflections and lighting using independent components analysis. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149), volume 1, pages 262-267. IEEE, 1999. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.433, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[5] Yuanbiao Gou, Boyun Li, Zitao Liu, Songfan Yang, and Xi Peng. Clearer: Multi-scale neural architecture search for image restoration. Advances in neural information processing systems, 33:17129-17140, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.49, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[6] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted hdr video reconstruction. arXiv preprint arXiv:2406.08204, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[7] Byeong-Ju Han and Jae-Young Sim. Reflection removal using low-rank matrix completion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5438-5446, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.604, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.647, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[9] Yuchen Hong, Haofeng Zhong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. L-differ: Single image reflection removal with language-based diffusion model. In Proceedings of the European conference on computer vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.718, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[10] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation*. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.775, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[11] Qiming Hu and Xiaojie Guo. Single image reflection separation via component synergy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13138-13147, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.832, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Zhiyong Huang, Biao Xiong, Cao Tian, Jing Zhan, Xiang Fei, and Nazaraf Shah. Dust and reflection removal from videos captured in moving car. In 2016 IEEE 13th International Conference on e-Business Engineering (ICEBE), pages 182-187. IEEE, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.205, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[13] Yujin Jeon, Eunsue Choi, Youngchan Kim, Yunseong Moon, Khalid Omer, Felix Heide, and Seung-Hwan Baek. Spectral and polarization vision: Spectro-polarimetric real-world dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22098-22108, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.176, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[15] Soomin Kim, Yuchi Huo, and Sung-Eui Yoon. Single image reflection removal with physically-based training images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5164-5173, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.301, + 0.892, + 0.369 + ], + "angle": 0, + "content": "[16] Naejin Kong, Yu-Wing Tai, and Joseph S Shin. A physically-based approach to reflection separation: from physical modeling to constrained optimization. IEEE transactions on pattern analysis and machine intelligence, 36(2):209-221, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.371, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[17] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14811-14820, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[18] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1750-1758, 2020. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[19] Chenyang Lei, Xudong Jiang, and Qifeng Chen. Robust reflection removal with flash-only cues in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[20] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(9):1647-1654, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[21] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E Hopcroft. Single image reflection removal through cascaded refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3565-3574, 2020. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[22] Tingtian Li, Yuk-Hee Chan, and Daniel PK Lun. Improved multiple-image-based reflection removal algorithm using deep neural networks. IEEE Transactions on Image Processing, 30:68-79, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[23] Xiaobo Li, Lei Yan, Pengfei Qi, Liping Zhang, François Goudail, Tiegen Liu, Jingsheng Zhai, and Haofeng Hu. Polarimetric imaging via deep learning: A review. Remote Sensing, 15(6):1540, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[24] Yu Li and Michael S Brown. Single image layer separation using relative smoothness. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2752-2759, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolarized and polarized images. Advances in neural information processing systems, 32, 2019. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10898" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Ajay Nandoriya, Mohamed Elgharib, Changil Kim, Mohamed Hefeeda, and Wojciech Matusik. Video reflection removal through spatio-temporal optimization. In Proceedings of the IEEE International Conference on Computer Vision, pages 2411-2419, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[27] Shree K Nayar, Xi-Sheng Fang, and Terrance Boult. Separation of reflection components using color and polarization. International Journal of Computer Vision, 21(3):163-186, 1997. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[28] Simon Niklaus, Xuaner Cecilia Zhang, Jonathan T Barron, Neal Wadhwa, Rahul Garg, Feng Liu, and Tianfan Xue. Learned dual-view reflection removal. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3713-3722, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[29] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization-based decorrelation of transparent layers: The inclination angle of an invisible surface. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 814-819. IEEE, 1999. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[30] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization and statistical analysis of scenes containing a semireflector. JOSA A, 17(2):276-284, 2000. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[31] YiChang Shih, Dilip Krishnan, Fredo Durand, and William T Freeman. Reflection removal using ghosting cues. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3201, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.47, + 0.483 + ], + "angle": 0, + "content": "[32] Sony Semiconductor Solutions. Polarsens: Polarization sensing technology, 2024. Accessed: 2024-11-15. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.47, + 0.525 + ], + "angle": 0, + "content": "[33] David Strong and Tony Chan. Edge-preserving and scale-dependent properties of total variation regularization. Inverse problems, 19(6):S165, 2003. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[34] Robby T Tan and Katsushi Ikeuchi. Separating reflection components of textured surfaces using a single image. IEEE transactions on pattern analysis and machine intelligence, 27(2):178-193, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.47, + 0.649 + ], + "angle": 0, + "content": "[35] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Crrn: Multi-scale guided concurrent reflection removal network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4777-4785, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.47, + 0.705 + ], + "angle": 0, + "content": "[36] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Corrn: Cooperative reflection removal network. IEEE transactions on pattern analysis and machine intelligence, 42(12):2969-2982, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[37] Mengyi Wang, Xinxin Zhang, Yongshun Gong, and Yilong Yin. Personalized single image reflection removal network through adaptive cascade refinement. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8204-8213, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[38] Kaixuan Wei, Jiaolong Yang, Ying Fu, David Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8178-8187, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[39] Patrick Wieschollek, Orazio Gallo, Jinwei Gu, and Jan Kautz. Separating reflection and transmission images in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), pages 89-104, 2018. 1, 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[40] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.162, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[41] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T Freeman. A computational approach for obstruction-free photography. ACM Transactions on Graphics (TOG), 34(4):1-11, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.893, + 0.275 + ], + "angle": 0, + "content": "[42] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.893, + 0.331 + ], + "angle": 0, + "content": "[43] Jae-Seong Yun and Jae-Young Sim. Reflection removal for large-scale 3d point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4597–4605, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.894, + 0.388 + ], + "angle": 0, + "content": "[44] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4786-4794, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.893, + 0.458 + ], + "angle": 0, + "content": "[45] Haofeng Zhong, Yuchen Hong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. Language-guided image reflection separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24913-24922, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.46, + 0.894, + 0.53 + ], + "angle": 0, + "content": "[46] Yurui Zhu, Xueyang Fu, Peng-Tao Jiang, Hao Zhang, Qibin Sun, Jinwei Chen, Zheng-Jun Zha, and Bo Li. Revisiting single image reflection removal in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25468-25478, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10899" + } + ] +] \ No newline at end of file diff --git a/2025/PolarFree_ Polarization-based Reflection-Free Imaging/e1567ff5-b726-4ec2-baeb-81cce9b4c3bf_origin.pdf b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/e1567ff5-b726-4ec2-baeb-81cce9b4c3bf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..645bf34bb0ed1d7a9b96c40d892721eb1c290a4a --- /dev/null +++ b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/e1567ff5-b726-4ec2-baeb-81cce9b4c3bf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45ca5e940a1ac1ac23070d10a54eccf526e59028b7770a6f5fcf18182de4a3ca +size 5327061 diff --git a/2025/PolarFree_ Polarization-based Reflection-Free Imaging/full.md b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cb52202e0a333251d0cf1b67f97b83091bc2814f --- /dev/null +++ b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/full.md @@ -0,0 +1,469 @@ +# PolarFree: Polarization-based Reflection-Free Imaging + +Mingde Yao $^{1}$ , Menglu Wang $^{3}$ , King-Man Tam $^{1,4}$ , Lingen Li $^{1}$ , Tianfan Xue $^{\text{☑}1,2}$ , Jinwei Gu $^{1}$ $^{1}$ The Chinese University of Hong Kong, $^{2}$ Shanghai AI Laboratory + $^{3}$ University of Science and Technology of China, $^{4}$ Institute of Science Tokyo + +mingdeyao@foxmail.com, tfxue@ie.cuhk.edu.hk + +# Abstract + +Reflection removal is challenging due to complex light interactions, where reflections obscure important details and hinder scene understanding. Polarization naturally provides a powerful cue to distinguish between reflected and transmitted light, enabling more accurate reflection removal. However, existing methods often rely on small-scale or synthetic datasets, which fail to capture the diversity and complexity of real-world scenarios. To this end, we construct a large-scale dataset, PolaRGB, for Polarization-based reflection removal of RGB images, which enables us to train models that generalize effectively across a wide range of real-world scenarios. The PolaRGB dataset contains 6,500 well-aligned mixed-transmission image pairs, $8 \times$ larger than existing polarization datasets, and is the first to include both RGB and polarization images captured across diverse indoor and outdoor environments with varying lighting conditions. Besides, to fully exploit the potential of polarization cues for reflection removal, we introduce PolarFree, which leverages diffusion process to generate reflection-free cues for accurate reflection removal. Extensive experiments show that PolarFree significantly enhances image clarity in challenging reflective scenarios, setting a new benchmark for polarized imaging and reflection removal. Code and dataset are available at https://github.com/mdyao/PolarFree. + +# 1. Introduction + +Reflection removal algorithms [1, 19, 30, 34, 38, 41] remove unwanted reflections in captured images, playing a critical role in applications such as autonomous driving [12] and photography [11, 43]. This problem commonly arises when imaging through semi-reflectors, like windows or glass, and overlapping reflections may obscure important details of scenes we want to capture. This problem is often formulated [10, 11] as a linear combination of the transmission layer $T$ and the reflection layer $R$ : + +$$ +M = \alpha_ {t} T + \alpha_ {r} R, \tag {1} +$$ + +![](images/80e710133ca035792331381ce071ae6798e7f3a6d4566b4935608c56fe27cc34.jpg) + +![](images/df1740ae295ef3e8eefaf548cb03914837bd7268e032c6f634102d229c4fc1c8.jpg) + +![](images/41aeecca7066eabb3ce5abf7df59c06933a09c0d75f0eeee98d7be733ba82e16.jpg) + +![](images/e089df2f146b43c1ea7ee94c61df54a93e2dd0614df58e2ee137b157fb915a94.jpg) +Input +RDRNet [46] +Figure 1. Our PolarFree effectively leverages polarization information to remove reflections, achieving superior performance in challenging scenes with complex backgrounds and highlights where previous methods [11, 15, 46] often fail. + +![](images/39171562de537b9915487cc734622c40e855ab5d80fe66765cc1ea62e07066a2.jpg) +Kim et al. [11] +Ours + +![](images/48ffe3f5b52fd54b2427e6c1e2c77d46d06971a8eab1c4ec5010b2e6e091d1e8.jpg) +DSRNet [15] +Ground truth + +where $M$ is the mixed captured image, and $\alpha_{t},\alpha_{r}$ are blending coefficients resulting from light attenuation. + +Polarization image sensors are becoming mainstream, allowing users to easily capture polarization images from a single shot in real-time [32]. However, existing methods [10, 11, 24, 46] typically rely on intensity-based cues, such as pixel brightness and color gradients, to distinguish transmitted and reflected layers. These methods face challenges because reflection removal is a highly ill-posed inverse problem [21] that recovers two unknown layers (reflection and transmission) from a single observation. Polarization provides valuable physics-based cues [27, 29] to alleviate ill-posed issues (Fig. 1), whereas transmitted light retains a distinct degree of polarization(Fig. 2b). This difference provides crucial signals for separating the two layers. Notably, at the Brewster angle (Fig.2c), reflected light is fully polarized, enabling effective reflection removal [2]. + +Despite significant advances of polarization images, a key challenge in polarization-based reflection removal is the lack of large-scale, high-quality datasets. Existing polarized reflection removal datasets [18, 39] are limited in size and diversity, relying on small (<1000) or synthetic samples that fail to capture the complexity of real-world light- + +![](images/7ea6bef2076cfbf69882589e06b5d3ae6c93b962c7da50576b216d8f73874dfb.jpg) +Figure 2. (a) & (b) A semi-reflector transforms unpolarized light into polarized light upon reflection and refraction, which is undetectable by standard RGB cameras but can be leveraged by polarization cameras for reflection-suppression tasks. (c) At the Brewster angle [2], a polarizer minimizes reflections. + +ing conditions, materials, and scenes. Moreover, they typically exclude color information, reducing their applicability in real-world reflection removal tasks. Thus, there is a pressing need for a large-scale, comprehensive dataset that includes both RGB and polarization images, captured in diverse real-world environments, to advance polarization-based reflection removal. + +To bridge this gap, we introduce PolaRGB, a novel dataset specifically collected for polarization-based reflection removal. As shown in Table 1, PolaRGB contains 6,500 high-quality, well-aligned RGB-polarization image pairs, $8 \times$ larger than previous dataset [18]. Our dataset covers a diverse range of scenes, lighting conditions, and exposure settings, and is captured using off-the-shelf commercial cameras with polarized color patterns to ensure real-world applicability. PolaRGB provides both mixed images and ground-truth transmission layers, enabling accurate reflection separation and significantly enhancing the effectiveness of reflection removal across real-world scenarios. + +Moreover, extracting reflection-free information from polarization data is challenging [18, 25, 39] due to the randomness of shooting angles, scene variations, and changing lighting conditions. To address this issue, we leverage the powerful generative capabilities of the diffusion model [3,6,8] to generate reflection-free cues. The diffusion model extracts and refines reflection-free priors from polarization images, effectively guiding the reflection removal and yielding precise and robust reflection-free results. + +Specifically, PolarFree consists of two steps: a prior-generation step and a reflection removal step. First, in the prior-generation step, we use a diffusion model to generate reflection-free prior based on the polarization and RGB inputs. This strategy not only guides accurate reflection isolation in RGB images but also recovers background details, which previous methods [11, 15, 46] may miss. Next, the reflection removal step leverages the prior to effectively remove reflections, ensuring accurate transmission restoration. Additionally, we introduce a phase-based loss function in the frequency domain to mitigate color discrepancies caused by semi-reflections, guiding the network to focus + +Table 1. Comparisons between PolaRGB and existing reflection removal datasets. + +
DatasetPolarizationRGBRAWData sizeResolution
SIR2 [36]NoYesNo500540 × 400
RRW [46]NoYesNo14,9522580 × 1460
ReflectNet [39]Yes (Syn.)YesNo-1500 × 1000
Lei et al. [18]Yes (Real)NoYes8071224 × 1024
PolaRGB (Ours)Yes (Real)YesYes6,5001224 × 1024
+ +on reflection removal rather than color adjustment. These components enable PolarFree to achieve robust reflection suppression while preserving the clarity and integrity of the transmission across diverse real-world scenes. + +Extensive experiments on the PolaRGB dataset demonstrate the effectiveness of PolarFree, which outperforms existing methods [11, 15, 46] $\sim$ 2dB in terms of PSNR. Additionally, we performed real-world testing in environments commonly affected by reflections such as museums and galleries, demonstrating the method's effectiveness in real-world scenarios. Our approach improves image clarity and preserves details better than previous methods, paving the way for more practical, reflection-robust applications. + +# 2. Related Work + +# 2.1. Reflection Removal + +Reflection removal [1, 30, 34] has a longstanding history in computer vision, aiming to enhance the occluded objects caused by semi-reflectors like glass and windows. Early approaches [20, 24] typically relied on handcrafted priors, such as smoothness assumptions and gradient sparsity, to differentiate the reflection layer from the underlying scene. Some methods incorporated ghosting [31], flash cues [17, 19] and edge consistency [7], while others improved accuracy by leveraging multi-image setups [22, 28], utilizing temporal variation [26] and parallax [28] to separate reflections from background layers. However, these methods often require strict assumptions and fail in real-world scenes with complex textures or lighting [19, 38]. + +With the rise of deep learning [5, 35, 42, 44, 46], reflection removal methods advanced considerably. Techniques based on neural networks have shown promise by learning to separate reflection and transmission layers directly from data [10, 15, 35-37, 44]. However, without essential physical insights, these methods act as Bayesian regressions toward a learned average, often lead to suboptimal separation in complex environments with varying lighting, textures, and reflection intensities [35, 44]. Recent methods [9, 45] utilize language to interactively remove the reflection, which require additional efforts, limiting their practicality. + +# 2.2. Polarization Reflection Removal + +Polarization is an effective way to remove reflections due to the physics principle: the polarization of light behaves differently in the reflection and transmission layers [4, 13, 23, 27, 30], allowing separation, especially at the Brewster angle [4, 27]. However, capturing images exactly + +![](images/8ca8a3972a20070ec3d390a56d58221c826881066fa30c006b18aa06961f73a9.jpg) +(a) Hierarchical structure of PolarRR dataset + +![](images/eea8759d742d4f8018cced432db6271d3df78cdc252da9d3585e413096a9e90f.jpg) + +![](images/4f283e0279b4acc02c52c830457a3829881c71f71df8625e75cf5f24c7b80dbd.jpg) + +![](images/880553008569a864c8d9aaaf274b5fb203da3d7fb67e7d31fd02ab9ce0e8705b.jpg) +(b) Diverse reflection conditions + +![](images/65ac40515c73091e5f6d18db5c5e31dda4b9c865764a52d22efd4ac11e8b5785.jpg) + +![](images/d2809239934c849ed6ebd53e3634f7a3bd7cb6960776f5947e5061b3abe70246.jpg) +(c) Video-based capture method + +![](images/6d5014dcc4f28f019f5fe6eb6c70c122dc15cc435762b52fce29220e54f21187.jpg) + +![](images/d64caf525b3d078a6dd1b940c763a7d5c9710b205b7241eac8eeab4d949d7754.jpg) + +![](images/45136a25128f63284bc3a2d8b0d899872b13cd308c274102f896c51a81f2162f.jpg) + +![](images/937d9fcc2dde35c82bb94007feabde22a2a2d9455eec666fa2846b7d345f7f63.jpg) + +![](images/7c37ca18470fac448ded46c855863281c1dce2742f0ac08118c99ef13030baa6.jpg) + +![](images/de043a9bc68b63d70554f9cae2c1a74cd12ae5e74ddcec154aa0adf9024cb4a1.jpg) +Figure 3. Overview of the PolaRGB dataset. (a) Hierarchical structure of scenes is shown in the ring, with legends indicating sample counts and subset types. (b) Typical scenes illustrating varied reflection conditions: I. smooth blending of reflection and refraction, II. abrupt reflection with mixed components, III. reflection dominant over transmission, and IV. minimal or no reflection. (c) Video-based capture method (details in Sec. 3). (d) We provide polarized images at angles $\phi = 0^{\circ}$ , $45^{\circ}$ , $90^{\circ}$ , and $135^{\circ}$ , along with derived AoLP, DoLP, and a well-aligned unpolarized image. The dataset also includes ground truth transmission and estimated reflections, all available in both raw and RGB formats. + +![](images/a16e0d41a7120c3aa13d8dea18c55b320b56fb14e74871494372cae3881f862c.jpg) +(d) Image types included in the dataset + +![](images/afd0ba01fe5cf90e494b5945746281ecf547a5c58f87a2eab31ecbb57e1b4af3.jpg) + +![](images/bcf55f5dec9a9232227e777e00f74312bf5ea9a7612f9123117f7bbdc71fbab0.jpg) + +at the Brewster angle, as shown in Fig. 2c, is challenging. Therefore, in practice, methods typically rely on images taken at multiple polarization angles (e.g., $0^{\circ}$ , $45^{\circ}$ , $90^{\circ}$ , and $135^{\circ}$ ) to capture polarization information [4, 27, 30]. + +Early polarized reflection removal methods [4, 27, 30] employ mathematical techniques like Principle Component Analysis (PCA) to separate reflection layers, which are effective when the transmission and reflection layers have significant content differences. Subsequently, Kong et al. [16] propose a multiscale scheme that automatically identifies the optimal separation of the reflection and background layers. Another advancements [39] introduce a neural network-based approach for polarization-guided reflection removal with synthetic dataset. Lyu et al. [25] utilize a pair of unpolarized and polarized images but lack of realistic polarization patterns. Lei et al. [18] further contribute by collecting a dataset of polarized images for reflection removal. However, this dataset only includes limited scene variations from pure polarization scenes, which limits its application scenario. In contrast, we present the first large-scale dataset captured for polarization-based RGB reflection removal, consisting of 6,500 well-aligned transmission-reflection pairs. We also propose a polarization-based reflection removal network based on diffusion models. + +# 3. PolaRGB Dataset + +# 3.1. Analysis + +Although a few polarization-based reflection removal datasets exist [18, 39], they face major limitations: (1) they + +are often synthetically generated [25, 39], which may not generalize well to real polarization, as perfectly simulating polarization phenomena is very hard; (2) they are typically limited in size and diversity (<1,000 samples), with limited scene variations, which constrains model robustness and generalizability [18, 39]; and (3) they lack RGB data [18, 39], reducing the practical value of these datasets. + +Our polarized-based reflection removal dataset, PolaRGB, addresses all these limitations with the following key features, as shown in Fig. 3. It has following advantages: (1) It contains real-captured images with perfect spatial alignment. We achieve pixel-level alignment through a three-step process: careful capture setup, manual filtering, and homography transformations, all in raw space to avoid demosaicing artifacts (see Fig. 4b and more details in supplementary). (2) It contains large-scale diverse scenes. The collected dataset covers a wide range of indoor and outdoor settings with four distinct reflection types, including smooth, sharp, high-brightness, and subtle reflections (see Fig. 3b). (3) It has extensive data modalities. Our dataset offers paired mixed and transmission images with both polarization and RGB captures, precisely aligned and available in both raw and RGB formats (see Fig. 3d). + +# 3.2. Processing Pipeline + +We leverage an efficient video-based capture flow for dataset collection, as shown in Fig. 3c. First, we capture a transmission-only image $T_{raw}$ as the ground truth. Then, we place a semi-reflective glass plate in front of the scene and rotate it continuously to capture mixed images + +![](images/718d1884c76556825fa944bdbda653d24baa38fb79c5faeddbee1429641b0068.jpg) +Figure 4. Data processing pipeline for obtaining aligned mixed and transmission images, and polarized images. + +$M_{\text{raw}}$ . A division-of-focal-plane polarization camera with color Bayer pattern is used to simultaneously capture both polarized and RGB information, as shown in Fig. 4. + +After obtaining the raw mixed image $M_{raw}$ and the transmission image $T_{raw}$ , we process the images sequentially in the raw and RGB domains as shown in Fig. 4a. In the raw domain, we first align $M_{raw}$ and $T_{raw}$ to correct spatial misalignment caused by light refraction. This is done by separating the images into different polarization angles and color channels, then applying affine transformation matrices to each of the channels to avoid aliasing from direct alignment, as shown in Fig. 4b. Next, we perform polarization separation on the aligned images to obtain four polarization images $(0^{\circ}, 45^{\circ}, 90^{\circ}$ , and $135^{\circ}$ ), and use Malus's law [27, 29] to sum them, producing the unpolarized image of the scene. Finally, we estimate the reflection image by searching for the optimal blending coefficient $\alpha_r$ and $\alpha_t$ in edge space, as illustrated in Eq. 1. For detailed steps and proof, please refer to the supplementary material. + +Next, we apply demosaicking to the processed raw images to obtain the mixed, transmission, and reflection images, each consisting of four polarization images and the unpolarized RGB image. Additionally, we compute the + +Stokes parameters (see Sec. 4.1) to estimate the Angle of Linear Polarization (AOLP) and Degree of Linear Polarization (DOLP), which are then used in reflection removal. + +# 4. Method + +# 4.1. Problem Formulation + +Object. Given an RGB image and its corresponding polarization images, we aim to recover the transmission layer of the RGB image by utilizing the distinct polarization characteristics. Our measurement consists of spatially-aligned RGB and polarization images, where the polarization images captured at four distinct angles $(0^{\circ}, 45^{\circ}, 90^{\circ}$ , and $135^{\circ})$ . These four angles provide a comprehensive polarization measurement [30], which allows us to compute polarization features crucial for separating layers. + +Polarization Preliminary. For semi-reflective surfaces that produce mixed images, the observed intensity $I_M^\phi (x)$ can be decomposed into reflected intensity $I_R^\phi (x)$ and transmitted intensity $I_T^\phi (x)$ as + +$$ +I _ {M} ^ {\phi} (x) = \alpha_ {R} I _ {R} ^ {\phi} (x) + \alpha_ {T} I _ {T} ^ {\phi} (x). \tag {2} +$$ + +If a linear polarizer is placed in front of the camera at an angle $\phi$ , the captured intensity $I^{\phi}(x)$ can be represented as + +$$ +I _ {M} ^ {\phi} (x) = \alpha (\theta ; \phi ; \phi_ {\perp}) I _ {R} ^ {\phi} (x) + (1 - \alpha (\theta ; \phi ; \phi_ {\parallel})) I _ {T} ^ {\phi} (x), \tag {3} +$$ + +where $\alpha (\theta ;\phi ;\phi_{\perp})$ and $(1 - \alpha (\theta ;\phi ;\phi_{\parallel}))$ are polarization-dependent mixing coefficients based on the angle of incidence $\theta ,\phi_{\perp}$ and $\phi_{\parallel}$ are the canonical polarization directions for the reflection and transmission, respectively. Directly solving for these parameters is challenging because $\alpha (\cdot)$ , $I_R^\phi (x)$ , and $I_R^\phi (x)$ are all unknown [27, 39]. Additionally, relationship between the observed intensity and the underlying reflected and transmitted components is highly nonlinear and influenced by various factors [39]. To address this, we utilize Stokes parameters [18] that provide a efficient way to represent and analyze polarized light, enabling more robust separation of reflection and transmission. + +Stokes Parameters. To capture the polarization effects in the scene, we use Stokes parameters $[S_0, S_1, S_2]$ , which can be derived from intensity measurements at specific polarization angles $(0^\circ, 45^\circ, 90^\circ$ , and $135^\circ$ ): + +$$ +\begin{array}{l} S _ {0} = \left(I _ {0 ^ {\circ}} + I _ {4 5 ^ {\circ}} + I _ {9 0 ^ {\circ}} + I _ {1 3 5 ^ {\circ}}\right) / 2, (4) \\ S _ {1} = I _ {0 ^ {\circ}} - I _ {9 0 ^ {\circ}}, S _ {2} = I _ {4 5 ^ {\circ}} - I _ {1 3 5 ^ {\circ}}. (4) \\ \end{array} +$$ + +Here, $S_0$ represents the total intensity of light, and $S_1$ and $S_2$ provide information about the linear polarization state of the light based on intensity differences at these key angles. + +Using the Stokes parameters, we compute the Degree of Linear Polarization (DOLP) and Angle of Linear Polarization (AOLP) as follows [2] + +$$ +D O L P (x) = \sqrt {S _ {1} (x) ^ {2} + S _ {2} (x) ^ {2}} / S _ {0} (x), \tag {5} +$$ + +$$ +A O L P (x) = \frac {1}{2} \operatorname {a t a n 2} \left(S _ {2} (x) / S _ {1} (x)\right). \tag {5} +$$ + +![](images/7de8f37e913ce7e31160fa6f024fdfaef8403b25dd3642752a06ce45afc6a5af.jpg) +(a) Inference +Figure 5. Pipeline of PolarFree. (a) During inference, PolarFree leverages polarized and RGB images as inputs, which are feeds into a conditional diffusion model to generate the prior $\hat{z}_0$ . The generated prior, along with the inputs, is then passed to the reflection removal backbone $\mathcal{F}_{\mathrm{remove}}$ to remove reflections. (b) PolarFree is trained in two stages. (1) A prior encoder extracts a reflection-free prior $z_0$ from clean transmission images and polarization cues, which serves as the supervision for the conditional diffusion model in stage two. (2) The conditional diffusion model is trained to progressively denoise noisy images, supervised by the prior from stage one, ensuring robust reflection separation. + +![](images/6de156d3ccb9bc576f12eb41b7bcc4bdf4e035160926479af57b03b528b5875d.jpg) +(b) Training + +In this formulation, DOLP describes the proportion of polarized light relative to total intensity, which helps indicate the degree of reflection in the scene. AOLP, on the other hand, reveals the orientation of the polarized light, allowing us to distinguish between reflection and transmission components more effectively, as shown in Fig. 3d. By leveraging intensity measurements at $0^{\circ}$ , $45^{\circ}$ , $90^{\circ}$ , and $135^{\circ}$ , these parameters provide valuable cues for separating mixed reflection and transmission layers in semi-reflective scenes. + +# 4.2. PolarFree Network + +To achieve high-quality reflection removal with effective utilization of polarization information, we introduce Polar-Free, a delicately designed two-step network, where each step addresses a distinct aspect of the reflection removal challenge. Inspired by [3], we leverage diffusion model to generate the prior for reflection removal. As shown in Fig. 5, during inference, the first step utilizes a conditional diffusion model to extract reflection-free priors, effectively isolating essential details from polarization data. The second step integrates these priors with RGB inputs, guiding the network to accurately separate reflections and enhance clarity, even in complex, real-world environments. + +Prior Generation. As shown in Fig. 5a, in the first step, we start with a randomly initialized noise $n$ , which is gradually denoised through a conditional diffusion model $\mathcal{F}_{\mathrm{diff}}$ + +$$ +\hat {z} _ {0} = \mathcal {F} _ {\text {d i f f}} (n | M _ {\text {c o n d}}), \tag {6} +$$ + +where $\hat{z}_0$ is the generated prior, $M_{\mathrm{cond}} = \{M_{\mathrm{polar}}, M_{\mathrm{aolp}}, M_{\mathrm{dolp}}, M_{\mathrm{rgb}}\}$ represents the mixed im + +ages, and $n$ is the initial noise. Through an iterative denoising process, the diffusion model progressively refines a noise image, conditioned on the polarization and RGB data, to generate a reflection prior. + +The denoising process follows the denoising diffusion probabilistic model (DDPM) framework [8] and employs a U-Net architecture to predict noise. At each timestep $t$ , the U-Net receives a noisy intermediate $z_{t}$ and outputs a noise estimate $\epsilon_{\theta}(z_{t},t)$ , predicting the noise added at that timestep. This process can be formulated as + +$$ +z _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(z _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} \left(z _ {t}, M _ {\text {c o n d}}, t\right)\right) + \sigma_ {t} z, \tag {7} +$$ + +where $\alpha_{t}$ and $\beta_{t}$ control the noise schedule across the timesteps, $\sigma_{t}$ is the standard deviation, and $z$ is noise sampled from a standard Gaussian distribution. $\bar{\alpha}_{t}$ is the cumulative product of $\alpha_{t}$ , indicating noise level to timestep $t$ . + +In this way, the U-Net gradually removes the noise in each step, conditioning on the polarization-based measurement input $M_{\mathrm{cond}}$ . This process continues until noise is fully removed, yielding a sample $\hat{z}_0$ that represents the reflection-free prior distribution extracted from the conditioned inputs. + +Reflection Removal. As shown in Fig. 5a, once the prior $\hat{z}_0$ has been obtained, the second step of PolarFree is separating the transmission and reflection layers. By leveraging the polarization cues provided by $M_{\mathrm{polar}}$ , $M_{\mathrm{aolp}}$ , and $M_{\mathrm{dolp}}$ , the model $\mathcal{F}_{\mathrm{remove}}$ separates the transmission features from the reflection ones, which can be represented as + +$$ +\hat {T} _ {r g b} = \mathcal {F} _ {\text {r e m o v e}} \left(\hat {z} _ {0}, M _ {\text {c o n d}}\right), \tag {8} +$$ + +![](images/be2378f448fa49b76e8cc9819f7cdac0e5aa1f09f0737731308edce26a1fc292.jpg) +Figure 6. (a) Phase information preserves shape and texture details, while color primarily affects the amplitude. (b) We apply two types of random color perturbations to the image and compute the perturbation errors. Phase-based loss is less sensitive to color changes. + +where $\mathcal{F}_{\mathrm{remove}}$ is a reflection removal neural network. This step ensures that the final output contains distinct transmission and reflection components, facilitating high-quality image reconstruction under challenging reflective scenarios. + +# 4.3. Training + +To train our PolarFree, particularly to address the challenge of the diffusion model lacking a suitable ground truth, we adopt a two-stage training strategy, as illustrated in Fig. 5b. This strategy consists of two sequential objectives: learning to extract a reflection-free prior and learning to generate a reflection-free prior. + +First Stage. In the first stage, we train an encoder to extract a prior $z_0$ for reflection-free information, which will serve as the supervision signal for the diffusion model in the second stage. Specifically, we feed the ground truth transmission images $T_{\mathrm{cond}} = \{T_{\mathrm{polar}}, T_{\mathrm{aolp}}, T_{\mathrm{dolp}}, T_{\mathrm{rgb}}\}$ into the encoder $\mathcal{E}$ , obtaining $z = \mathcal{E}(M_{\mathrm{cond}}, T_{\mathrm{rgb}})$ . This $z$ contains reflection-free information enriched with polarization-related cues. Subsequently, $z_0$ conditions the reflection removal backbone $\mathcal{F}_{\mathrm{remove}}$ to predict the clean transmission image $\hat{T}_{\mathrm{rgb}} = \mathcal{F}_{\mathrm{remove}}(z, M_{\mathrm{cond}})$ . + +Second Stage. In the second stage, we train the diffusion model $\mathcal{F}_{\mathrm{diff}}$ to generate a reflection-free prior $z_0$ from noisy input images, and finetune the reflection removal backbone $F_{\mathrm{remove}}$ . The key challenge here is that the diffusion model lacks of direct supervision. Therefore, we leverage the prior $z_0$ extracted in the first stage as the supervision signal to guide the model. + +We start with the extracted reflection-free prior $z_0$ and add noise over multiple timesteps. This process transforms the "ground-truth" prior $z_0$ into noisy versions $z_t$ at each timestep $t$ , which can be expressed as + +$$ +q \left(z _ {t} \mid z _ {0}\right) = \mathcal {N} \left(z _ {t}; \sqrt {1 - \beta_ {t}} z _ {0}, \beta_ {t} \mathbf {I}\right), \tag {9} +$$ + +where $\mathcal{N}$ represents a Gaussian distribution, $\beta_{t}$ is the noise schedule controlling the noise level added at each timestep, $\mathbf{I}$ represents the identity matrix. + +In the reverse diffusion process, the model is trained to progressively remove the noise step by step to recover the + +clean prior $z_0$ , which can be represented as + +$$ +p _ {\theta} \left(z _ {t - 1} \mid z _ {t}\right) = \mathcal {N} \left(z _ {t - 1}; \mu_ {\theta} \left(z _ {t}, t, M _ {\text {c o n d}}\right), \sigma_ {t} ^ {2} \mathbf {I}\right)), \tag {10} +$$ + +where $\mu_{\theta}(z_t,t,M_{\mathrm{cond}})$ is the mean function of the current state $z_{t}$ , timestep $t$ , and a conditional input $M_{\mathrm{cond}}$ . + +Throughout the reverse process, the model conditions on the polarization measurements $M_{\mathrm{cond}}$ , which provide essential information about the scene's physical properties. These measurements help the model accurately generate reflection-free components during the denoising process. + +After the reverse process completes, the model outputs a final reflection-free prior $\hat{z}_0$ , which contains clean transmission information. This prior is used as the guidance signal for the reflection removal backbone $\mathcal{F}_{\mathrm{remove}}$ . The backbone takes this reflection-free prior and the polarization data to predict the clean transmission image $\hat{T}_{rgb}$ . + +# 4.4. Losses + +Basic Loss. To optimize PolarFree, we follow [46] to utilize three basic losses: L1 loss, VGG perceptual loss [14], and total variation (TV) loss [33]. The L1 loss minimizes the pixel-wise difference between the predicted transmission image $\hat{T}_{\mathrm{rgb}}$ and the ground truth transmission image $T_{\mathrm{rgb}}$ as $\mathcal{L}_1 = \|\hat{T}_{\mathrm{rgb}} - T_{\mathrm{rgb}}\|_1$ . The VGG perceptual loss compares feature activations from a pre-trained VGG network [14], weighted by $\lambda_l$ : $\mathcal{L}_{\mathrm{VGG}} = \sum_l \lambda_l \|\phi_l(\hat{T}_{\mathrm{rgb}}) - \phi_l(T_{\mathrm{rgb}})\|_1$ , where $\phi_l$ denotes the activations of the $l$ -th VGG layer. We also use TV loss to constrain consistency via the gradient operator $\mathcal{L}_{\mathrm{TV}} = \|\nabla \hat{T}_{\mathrm{rgb}} - \nabla T_{\mathrm{rgb}}\|_1$ . + +Phase loss. While basic losses help the network match color and intensity values to the ground truth, they struggle with color discrepancies caused by semi-reflective surfaces during dataset collection. These variations, due to reflection and transmission properties of materials, affect the color and intensity, leading to mismatches between the predicted and target images and hindering the model's ability to learn the correct transmission map. + +To address this issue, we introduce a phase loss to focus on the structural information of the transmission, which is less sensitive to color variations. As shown in Fig. 6, phase information primarily captures the geometry and texture of the image, independent of color changes. The phase loss is formulated as + +$$ +\mathcal {L} _ {\text {p h a s e}} = \| \angle (F F T (\hat {T})) - \angle (F F T ((T _ {r g b})) \| _ {1}, \tag {11} +$$ + +where $FFT$ is the Fourier transform, and $\angle (\cdot)$ represents the phase angle of the Fourier coefficients. + +Diffusion Loss. The diffusion loss follows the standard DDPM formulation [8], where the model predicts the noise at each step $t$ and minimizes the difference between the predicted noise and the true noise as + +$$ +\mathcal {L} _ {d i f f} = \mathbb {E} _ {q \left(z _ {t} \mid z _ {t - 1}\right)} \left[ \| \epsilon_ {\theta} \left(z _ {t}, t\right) - \epsilon_ {\text {t r u e}} \left(z _ {t}\right) \| _ {2} ^ {2} \right]. \tag {12} +$$ + +Here, $\epsilon_{true}(z_t)$ is the added noise obtained by Eq. 9, and $\epsilon_{\theta}(z_t,t)$ is the predicted noise, conditioned on the polariza + +![](images/7f1d6246840b4417a8613b3dadab3f17d72504d483da8bedd949d838176c584f.jpg) + +![](images/e5b7ec04197727068b0c5a2c107194a0d261b1e1315855dbb59204ae1692cc02.jpg) + +![](images/a8f75f8c97a880694f46513fb0e643b1cce199f11e37a28fd297731bc27ddc24.jpg) + +![](images/f9bc029067c7aee81c3bce02c4710497f445c49c00df256b3c468e3cb88bde4d.jpg) + +![](images/0f8b43ca56488c0242d67801492d80300353f6c8485d42c3bc67e72595de61cd.jpg) + +![](images/577ef286cb47296cdc5f62e8f6a17cd37c06941375e9a7929de354a13195c1df.jpg) + +![](images/2bcdaea9895d86fbae76f14feb4b6d37a098aeda4cd8082140675adaebc1309a.jpg) + +![](images/d1c381867913b6bb267ec4c9dce476adffac5c81fccb6408ca54dd863c508b7e.jpg) + +![](images/c778d1bce0bc6fef0246deb4852dc3859dcdf5831771e7a21620f1558f8653c6.jpg) + +![](images/9ea181c82777c9adcf39c493d5929db1e9215397fe58645ced72e5de4cdc84ab.jpg) + +![](images/0497e34c1fd4e89d7e2a9b5a6a659190cf0747655d3d73c68a82629dfdd4f690.jpg) + +![](images/6fb04191e0953ba6859f3a9497edde818e7410bf7cb608c800a2920c229df5f5.jpg) + +![](images/31c184f1606ca7eb2defae89ce820be5a7684e230d7fd2185a85557f814df35c.jpg) + +![](images/b85ea1bef5394f460395ea3cfc8be2042bd1385c1c8b068fdfa4c52c67ff916a.jpg) + +![](images/26e573a11f10e382b1c5442278c7a747ea9a5861febdd2753500d6db2dbc36d3.jpg) + +![](images/14dc69fbec5c9db6ad2fee69f3c7db6e8a85a29df9b9cad330971156575e569f.jpg) + +![](images/5bff9ef0535e591fc3fbd9686b58a9ffbb85b03d477271ff0de8821a9bde7d97.jpg) + +![](images/fc8c45ec2fb7034965edaeb020639806da030c85d3ebb2a76574270bc3114120.jpg) + +![](images/80e68386a5e357c4c12f457aaa5f773c74ed9e0aaa7162cbac2541ce775ce3ca.jpg) +Input + +![](images/3cf543eb9e9c5288c4c081825a6d5385a4d8c84b9f9df9b448e52242536985c4.jpg) +Kim et al. [15] + +![](images/f7c677f301c14db4d59f3307a9dbac00631c9891658789df36351c371f0ca390.jpg) +DSRNet [11] + +![](images/c4c5da7b26707524aa0121620f676b2fff8c99f894307c03f3d59bbd037b8e8a.jpg) +RDRNet [46] + +![](images/cb932e0324bd6638b77ad9fe20bcdb0a2a2e7498bb837ed4038e25ca5efb80d0.jpg) +Ours +Figure 7. Qualitative comparisons on the PolaRGB dataset. + +![](images/a885f30b7d3401b4b44d869e8fd419942a596ee2ea10aadf663d0f38edd9d59f.jpg) +Ground Truth + +tion measurements and RGB input $M_{\mathrm{cond}}$ + +Total Loss. The above losses are weighted summed to serve as the supervision for the first stage as + +$$ +\mathcal {L} _ {\text {s t a g e 1}} = \gamma_ {1} \mathcal {L} _ {1} + \gamma_ {2} \mathcal {L} _ {\mathrm {V G G}} + \gamma_ {3} \mathcal {L} _ {\mathrm {T V}} + \gamma_ {4} \mathcal {L} _ {\text {p h a s e}}. \tag {13} +$$ + +In the second stage, we use a combined loss function consisting of the diffusion loss and the reconstruction loss as + +$$ +\mathcal {L} _ {\text {s t a g e 2}} = \gamma_ {5} \mathcal {L} _ {\text {d i f f}} + \gamma_ {6} \mathcal {L} _ {\text {r e c o n}}. \tag {14} +$$ + +where $\mathcal{L}_{\mathrm{recon}}$ has the same format as $\mathcal{L}_{\mathrm{stage1}}$ , and $\gamma (\cdot)$ are the coefficients. This combined loss helps refine the reflection-free image while ensuring it aligns with the ground truth. + +# 5. Experiments + +# 5.1. Implementation Details + +We train and evaluate PolarFree on the PolaRGB dataset. The whole dataset consists of 67 scenes, with a total of 6,500 paired images. For each scene, we keep the background (transmission) and camera fixed, and varying the glass position to capture images with reflections. The 67 scenes are then randomly split into 56 training scenes and 11 testing scenes, containing 6,312 and 188 paired images, respectively. This division ensures no data leakage between the training and testing sets, with each set containing only a subset of all categories. + +We also test PolarFree in real-world scenes like museums and galleries, where ground truth is unavailable. We implement PolarFree using PyTorch on a single NVIDIA RTX 4090 GPU. Training is conducted with a batch size of + +2 and an AdamW optimizer with a learning rate of $2 \times 10^{-4}$ , 30k iterations on the PolaRGB dataset for both stages. + +We compare our method with recent advanced reflection removal methods, including Lei et al. [18], IBCLN [21], DSRNet [11], YTMT [10], and RDRNet [46]. For a fair comparison, we modify the input settings to align with ours and use only the transmission layer for supervision. We have re-trained baseline methods on the PolaRGB dataset. The evaluation is conducted using both objective metrics (PSNR, SSIM), a perceptual metric (LPIPS), and a language-based non-reference metric (Q-Align [40]). + +# 5.2. Results + +Results on PolaRGB. We present the quantitative results in Tab. 2. It can be seen that, our method outperforms other methods across multiple metrics, demonstrating the effectiveness of PolarFree for high-quality reflection removal. Notably, for fairness, we modified the input layer of all baselines to accept polarization information. Visual results in Fig. 7 show that PolarFree provides cleaner reflection removal with sharper edges and better color preservation. In contrast, previous methods often suffer from color distortions or imperfect reflection removal, especially in areas with complex reflections. Our method maintains high fidelity to the ground truth, especially in challenging regions with low light and subtle reflections. + +Real-captured without Ground Truth. To demonstrate the generalization ability to unseen and more complex re + +![](images/4dbf17d1e91ac2a2b36077764aefec87ba6f88160cafc98eeb7dc53b9d46e541.jpg) +Input + +![](images/363fed455e966944353ab5237450f1889461997aeb0c0c154cc8bb2d28fa4ca7.jpg) +DSRNet [46] + +![](images/f62b3e1c43008901a0003b0d099f0ea5bdc557a89c50301bf530a69d5f18a001.jpg) +Kim et al. [15] +Figure 8. Visual comparison of reflection removal results in real-world museum scenes with glass display cases. + +![](images/c1a8b063639f65fea0303ee16ee9042c207ad2528618cf5a50f3572ec9f1a440.jpg) +Ours + +Table 2. Quantitative comparison on the PolaRGB dataset, evaluated using objective metrics (PSNR, SSIM), perceptual metric (LPIPS), and the Language-based non-reference metric (Q-align [40]). Best results are in bold, second-best results are underlined. + +
MethodPSNR↑SSIM↑LPIPS↓Q-align↑
Lei et al. [18]18.730.79620.38043.2109
Kim et al. [15]20.670.83990.27143.7148
IBCLN [21]19.730.81730.24883.0938
YTMT [10]16.860.75440.44893.0938
DSRNet [11]16.840.79130.28283.6992
RDRNet [46]15.880.69640.52502.9531
Ours22.440.86810.13253.8867
+ +flection, We further evaluate our model on real-captured images at museum with complex reflections from glass enclosures. Here, we only provide qualitative results to visually assess the model's capability. As shown in Fig. 8, our approach effectively reduces reflections while preserving fine details, despite the complex lighting and material variations often present in museum environments. This demonstrates the practical robustness of our method and its potential for real-world applications. + +# 5.3. Ablation Study + +We perform ablation studies to evaluate the effectiveness of key components in our method. More experimental results and analyses can be found in the supplementary. + +Polarization Information. To assess the importance of polarization cues, we conduct an experiment where the polarization information (AOLP, DOLP, and polarization images) is removed, and the model is trained solely using RGB data. As shown in Table 3 and Fig. 9, the absence of polarization significantly impairs the model's ability to differentiate between reflection and transmission layers, leading to noticeably poorer reflection removal performance. These results demonstrate the critical role of polarization in accurate reflection removal. + +Diffusion Prior. We also evaluate the effectiveness of diffusion prior by removing the conditional diffusion model. While without diffusion prior achieves reasonable results, as indicated by the comparison metrics in Table 3, it lacks fine-grained control and high-quality output offered by the + +![](images/2c1b7b3fc411e10044bd1f1e525e275ea06cbeff5af14633e0a6490259dcb056.jpg) +w/o diffusion +Figure 9. Highly challenging reflection removal scenes with complex reflections and highlights. Top: Polarization images from different angles provide complementary information for effective reflection removal, note the third $(90^{\circ})$ image. Bottom: The diffusion module effectively handles highlights, while polarization better restores the color and details in such challenging scenarios. + +![](images/8904b546ba9899d560b2ee37c9e9e9f73080ad336a43c01adb31e23fbda4afad.jpg) +w/o polarization + +![](images/40e5e2dfce30f99dd30c516287cd51007e097a17f1f610b5cf50972e0f4bc386.jpg) +Ours + +![](images/d73bf48359244d1c9e4befe2ba5e1fe7a0381cf3cc40df20f3220e78116213e5.jpg) +Ground Truth + +Table 3. Ablation study on PolarFree framework. + +
MethodPSNR↑SSIM↑LPIPS↓Q-align↑
w/o Lphase22.410.86220.14023.8524
w/o polarization21.560.86200.14833.7949
w/o diffusion prior20.210.86270.15523.7949
Ours22.440.86810.13253.8867
+ +diffusion model, especially in complex scenes (Fig. 9). The conditional diffusion process significantly improves separation accuracy and preserves finer transmission details. + +Phase-based Loss. We explore the impact of our phase-based loss function. We remove this loss from the training procedure and observe the model's performance without it. The results show a noticeable increase in color discrepancies and less accurate reflection removal, particularly in scenarios involving semi-reflective surfaces (Table 3). The phase-based loss function ensures the model focuses on reflection removal rather than compensating for color inconsistencies, leading to improved overall results. + +# 6. Conclusion + +In this work, we propose PolaRGB, a large-scale, comprehensive dataset for polarization-based reflection removal. With 6,500 well-aligned RGB-polarization image pairs, PolaRGB is $8 \times$ larger than existing polarization datasets and is the first to include both RGB and polarization images, uniquely captured across diverse real-world scenes and lighting conditions. Additionally, we present PolarFree, a novel reflection removal model that leverages the generative capabilities of diffusion models. This approach enables PolarFree to generate reflection-free cues, enhancing separation accuracy while preserving fine details in the transmission layer. A phase-based loss function is also introduced to further improve the model's performance. Comprehensive experimental results on realistic scenarios demonstrate the effectiveness our method. We believe these contributions set a foundation for further advancements in polarization-based reflection removal, paving the way for more sophisticated applications in complex real-world environments. + +# Acknowledgments + +This work was supported in part by the National Key R&D Program of China (No.2022ZD0160201) and Shanghai Artificial Intelligence Laboratory, and the RGC Early Career Scheme (ECS) No. 24209224. + +# References + +[1] Amit Agrawal, Ramesh Raskar, Shree K Nayar, and Yuanzhen Li. Removing photography artifacts using gradient projection and flash-exposure sampling. ACM Transactions on Graphics, 24(3):828-835, 2005. 1, 2 +[2] Max Born and Emil Wolf. Principles of optics: electromagnetic theory of propagation, interference and diffraction of light. Elsevier, 2013. 1, 2, 4 +[3] Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring. Advances in neural information processing systems, 36, 2024. 2, 5 +[4] Hany Farid and Edward H Adelson. Separating reflections and lighting using independent components analysis. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149), volume 1, pages 262-267. IEEE, 1999. 2, 3 +[5] Yuanbiao Gou, Boyun Li, Zitao Liu, Songfan Yang, and Xi Peng. Clearer: Multi-scale neural architecture search for image restoration. Advances in neural information processing systems, 33:17129-17140, 2020. 2 +[6] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted hdr video reconstruction. arXiv preprint arXiv:2406.08204, 2024. 2 +[7] Byeong-Ju Han and Jae-Young Sim. Reflection removal using low-rank matrix completion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5438-5446, 2017. 2 +[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 5, 6 +[9] Yuchen Hong, Haofeng Zhong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. L-differ: Single image reflection removal with language-based diffusion model. In Proceedings of the European conference on computer vision (ECCV), 2024. 2 +[10] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation*. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 7, 8 +[11] Qiming Hu and Xiaojie Guo. Single image reflection separation via component synergy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13138-13147, 2023. 1, 2, 7, 8 +[12] Zhiyong Huang, Biao Xiong, Cao Tian, Jing Zhan, Xiang Fei, and Nazaraf Shah. Dust and reflection removal from videos captured in moving car. In 2016 IEEE 13th International Conference on e-Business Engineering (ICEBE), pages 182-187. IEEE, 2016. 1 + +[13] Yujin Jeon, Eunsue Choi, Youngchan Kim, Yunseong Moon, Khalid Omer, Felix Heide, and Seung-Hwan Baek. Spectral and polarization vision: Spectro-polarimetric real-world dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22098-22108, 2024. 2 +[14] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 6 +[15] Soomin Kim, Yuchi Huo, and Sung-Eui Yoon. Single image reflection removal with physically-based training images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5164-5173, 2020. 1, 2, 8 +[16] Naejin Kong, Yu-Wing Tai, and Joseph S Shin. A physically-based approach to reflection separation: from physical modeling to constrained optimization. IEEE transactions on pattern analysis and machine intelligence, 36(2):209-221, 2013. 3 +[17] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14811-14820, 2021. 2 +[18] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1750-1758, 2020. 1, 2, 3, 4, 7, 8 +[19] Chenyang Lei, Xudong Jiang, and Qifeng Chen. Robust reflection removal with flash-only cues in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2 +[20] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(9):1647-1654, 2007. 2 +[21] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E Hopcroft. Single image reflection removal through cascaded refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3565-3574, 2020. 1, 7, 8 +[22] Tingtian Li, Yuk-Hee Chan, and Daniel PK Lun. Improved multiple-image-based reflection removal algorithm using deep neural networks. IEEE Transactions on Image Processing, 30:68-79, 2020. 2 +[23] Xiaobo Li, Lei Yan, Pengfei Qi, Liping Zhang, François Goudail, Tiegen Liu, Jingsheng Zhai, and Haofeng Hu. Polarimetric imaging via deep learning: A review. Remote Sensing, 15(6):1540, 2023. 2 +[24] Yu Li and Michael S Brown. Single image layer separation using relative smoothness. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2752-2759, 2014. 1, 2 +[25] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolarized and polarized images. Advances in neural information processing systems, 32, 2019. 2, 3 + +[26] Ajay Nandoriya, Mohamed Elgharib, Changil Kim, Mohamed Hefeeda, and Wojciech Matusik. Video reflection removal through spatio-temporal optimization. In Proceedings of the IEEE International Conference on Computer Vision, pages 2411-2419, 2017. 2 +[27] Shree K Nayar, Xi-Sheng Fang, and Terrance Boult. Separation of reflection components using color and polarization. International Journal of Computer Vision, 21(3):163-186, 1997. 1, 2, 3, 4 +[28] Simon Niklaus, Xuaner Cecilia Zhang, Jonathan T Barron, Neal Wadhwa, Rahul Garg, Feng Liu, and Tianfan Xue. Learned dual-view reflection removal. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3713-3722, 2021. 2 +[29] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization-based decorrelation of transparent layers: The inclination angle of an invisible surface. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 814-819. IEEE, 1999. 1, 4 +[30] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization and statistical analysis of scenes containing a semireflector. JOSA A, 17(2):276-284, 2000. 1, 2, 3, 4 +[31] YiChang Shih, Dilip Krishnan, Fredo Durand, and William T Freeman. Reflection removal using ghosting cues. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3201, 2015. 2 +[32] Sony Semiconductor Solutions. Polarsens: Polarization sensing technology, 2024. Accessed: 2024-11-15. 1 +[33] David Strong and Tony Chan. Edge-preserving and scale-dependent properties of total variation regularization. Inverse problems, 19(6):S165, 2003. 6 +[34] Robby T Tan and Katsushi Ikeuchi. Separating reflection components of textured surfaces using a single image. IEEE transactions on pattern analysis and machine intelligence, 27(2):178-193, 2005. 1, 2 +[35] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Crrn: Multi-scale guided concurrent reflection removal network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4777-4785, 2018. 2 +[36] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Corrn: Cooperative reflection removal network. IEEE transactions on pattern analysis and machine intelligence, 42(12):2969-2982, 2019. 2 +[37] Mengyi Wang, Xinxin Zhang, Yongshun Gong, and Yilong Yin. Personalized single image reflection removal network through adaptive cascade refinement. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8204-8213, 2023. 2 +[38] Kaixuan Wei, Jiaolong Yang, Ying Fu, David Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8178-8187, 2019. 1, 2 +[39] Patrick Wieschollek, Orazio Gallo, Jinwei Gu, and Jan Kautz. Separating reflection and transmission images in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), pages 89-104, 2018. 1, 2, 3, 4 + +[40] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 7, 8 +[41] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T Freeman. A computational approach for obstruction-free photography. ACM Transactions on Graphics (TOG), 34(4):1-11, 2015. 1 +[42] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 2 +[43] Jae-Seong Yun and Jae-Young Sim. Reflection removal for large-scale 3d point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4597–4605, 2018. 1 +[44] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4786-4794, 2018. 2 +[45] Haofeng Zhong, Yuchen Hong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. Language-guided image reflection separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24913-24922, 2024. 2 +[46] Yurui Zhu, Xueyang Fu, Peng-Tao Jiang, Hao Zhang, Qibin Sun, Jinwei Chen, Zheng-Jun Zha, and Bo Li. Revisiting single image reflection removal in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25468-25478, 2024. 1, 2, 6, 7, 8 \ No newline at end of file diff --git a/2025/PolarFree_ Polarization-based Reflection-Free Imaging/images.zip b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..91426cb2106774bfa0acf6cffe1d085a38edbb46 --- /dev/null +++ b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b020b96cadd2b79d6ad81647a8ff927fe8d4facc71afee158c1d27b867159be +size 972166 diff --git a/2025/PolarFree_ Polarization-based Reflection-Free Imaging/layout.json b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2ebf8dc1b77b876cd8b7c9d34a6cebf8dacf3d53 --- /dev/null +++ b/2025/PolarFree_ Polarization-based Reflection-Free Imaging/layout.json @@ -0,0 +1,12485 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "type": "text", + "content": "PolarFree: Polarization-based Reflection-Free Imaging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "text", + "content": "Mingde Yao" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "text", + "content": ", Menglu Wang" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "text", + "content": ", King-Man Tam" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "text", + "content": ", Lingen Li" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "text", + "content": ", Tianfan Xue" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "inline_equation", + "content": "^{\\text{☑}1,2}" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "text", + "content": ", Jinwei Gu" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "text", + "content": "The Chinese University of Hong Kong, " + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "text", + "content": "Shanghai AI Laboratory \n" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "text", + "content": "University of Science and Technology of China, " + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 531, + 186 + ], + "type": "text", + "content": "Institute of Science Tokyo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 174, + 188, + 414, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 188, + 414, + 198 + ], + "spans": [ + { + "bbox": [ + 174, + 188, + 414, + 198 + ], + "type": "text", + "content": "mingdeyao@foxmail.com, tfxue@ie.cuhk.edu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 222, + 192, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 222, + 192, + 233 + ], + "spans": [ + { + "bbox": [ + 143, + 222, + 192, + 233 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 237, + 290, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 290, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 290, + 537 + ], + "type": "text", + "content": "Reflection removal is challenging due to complex light interactions, where reflections obscure important details and hinder scene understanding. Polarization naturally provides a powerful cue to distinguish between reflected and transmitted light, enabling more accurate reflection removal. However, existing methods often rely on small-scale or synthetic datasets, which fail to capture the diversity and complexity of real-world scenarios. To this end, we construct a large-scale dataset, PolaRGB, for Polarization-based reflection removal of RGB images, which enables us to train models that generalize effectively across a wide range of real-world scenarios. The PolaRGB dataset contains 6,500 well-aligned mixed-transmission image pairs, " + }, + { + "bbox": [ + 46, + 237, + 290, + 537 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 46, + 237, + 290, + 537 + ], + "type": "text", + "content": " larger than existing polarization datasets, and is the first to include both RGB and polarization images captured across diverse indoor and outdoor environments with varying lighting conditions. Besides, to fully exploit the potential of polarization cues for reflection removal, we introduce PolarFree, which leverages diffusion process to generate reflection-free cues for accurate reflection removal. Extensive experiments show that PolarFree significantly enhances image clarity in challenging reflective scenarios, setting a new benchmark for polarized imaging and reflection removal. Code and dataset are available at https://github.com/mdyao/PolarFree." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 567, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 674 + ], + "type": "text", + "content": "Reflection removal algorithms [1, 19, 30, 34, 38, 41] remove unwanted reflections in captured images, playing a critical role in applications such as autonomous driving [12] and photography [11, 43]. This problem commonly arises when imaging through semi-reflectors, like windows or glass, and overlapping reflections may obscure important details of scenes we want to capture. This problem is often formulated [10, 11] as a linear combination of the transmission layer " + }, + { + "bbox": [ + 46, + 567, + 287, + 674 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 567, + 287, + 674 + ], + "type": "text", + "content": " and the reflection layer " + }, + { + "bbox": [ + 46, + 567, + 287, + 674 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 567, + 287, + 674 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 676, + 287, + 688 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 676, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 129, + 676, + 287, + 688 + ], + "type": "interline_equation", + "content": "M = \\alpha_ {t} T + \\alpha_ {r} R, \\tag {1}", + "image_path": "c5b9a3c8a0a19e404a9f634e095cb2e51b7b38109ea0e1ed6e4c4cb63f9a93be.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 225, + 386, + 292 + ], + "blocks": [ + { + "bbox": [ + 307, + 225, + 386, + 292 + ], + "lines": [ + { + "bbox": [ + 307, + 225, + 386, + 292 + ], + "spans": [ + { + "bbox": [ + 307, + 225, + 386, + 292 + ], + "type": "image", + "image_path": "80e710133ca035792331381ce071ae6798e7f3a6d4566b4935608c56fe27cc34.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 387, + 226, + 465, + 292 + ], + "blocks": [ + { + "bbox": [ + 387, + 226, + 465, + 292 + ], + "lines": [ + { + "bbox": [ + 387, + 226, + 465, + 292 + ], + "spans": [ + { + "bbox": [ + 387, + 226, + 465, + 292 + ], + "type": "image", + "image_path": "df1740ae295ef3e8eefaf548cb03914837bd7268e032c6f634102d229c4fc1c8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 465, + 226, + 545, + 292 + ], + "blocks": [ + { + "bbox": [ + 465, + 226, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 465, + 226, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 465, + 226, + 545, + 292 + ], + "type": "image", + "image_path": "41aeecca7066eabb3ce5abf7df59c06933a09c0d75f0eeee98d7be733ba82e16.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 303, + 386, + 369 + ], + "blocks": [ + { + "bbox": [ + 335, + 293, + 357, + 303 + ], + "lines": [ + { + "bbox": [ + 335, + 293, + 357, + 303 + ], + "spans": [ + { + "bbox": [ + 335, + 293, + 357, + 303 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 303, + 386, + 369 + ], + "lines": [ + { + "bbox": [ + 307, + 303, + 386, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 386, + 369 + ], + "type": "image", + "image_path": "e089df2f146b43c1ea7ee94c61df54a93e2dd0614df58e2ee137b157fb915a94.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 369, + 373, + 379 + ], + "lines": [ + { + "bbox": [ + 322, + 369, + 373, + 379 + ], + "spans": [ + { + "bbox": [ + 322, + 369, + 373, + 379 + ], + "type": "text", + "content": "RDRNet [46]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 380, + 545, + 425 + ], + "lines": [ + { + "bbox": [ + 306, + 380, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 545, + 425 + ], + "type": "text", + "content": "Figure 1. Our PolarFree effectively leverages polarization information to remove reflections, achieving superior performance in challenging scenes with complex backgrounds and highlights where previous methods [11, 15, 46] often fail." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 387, + 303, + 465, + 369 + ], + "blocks": [ + { + "bbox": [ + 399, + 293, + 454, + 303 + ], + "lines": [ + { + "bbox": [ + 399, + 293, + 454, + 303 + ], + "spans": [ + { + "bbox": [ + 399, + 293, + 454, + 303 + ], + "type": "text", + "content": "Kim et al. [11]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 387, + 303, + 465, + 369 + ], + "lines": [ + { + "bbox": [ + 387, + 303, + 465, + 369 + ], + "spans": [ + { + "bbox": [ + 387, + 303, + 465, + 369 + ], + "type": "image", + "image_path": "39171562de537b9915487cc734622c40e855ab5d80fe66765cc1ea62e07066a2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 370, + 435, + 378 + ], + "lines": [ + { + "bbox": [ + 415, + 370, + 435, + 378 + ], + "spans": [ + { + "bbox": [ + 415, + 370, + 435, + 378 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 466, + 303, + 545, + 369 + ], + "blocks": [ + { + "bbox": [ + 481, + 293, + 531, + 303 + ], + "lines": [ + { + "bbox": [ + 481, + 293, + 531, + 303 + ], + "spans": [ + { + "bbox": [ + 481, + 293, + 531, + 303 + ], + "type": "text", + "content": "DSRNet [15]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 466, + 303, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 466, + 303, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 466, + 303, + 545, + 369 + ], + "type": "image", + "image_path": "48ffe3f5b52fd54b2427e6c1e2c77d46d06971a8eab1c4ec5010b2e6e091d1e8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 370, + 530, + 378 + ], + "lines": [ + { + "bbox": [ + 481, + 370, + 530, + 378 + ], + "spans": [ + { + "bbox": [ + 481, + 370, + 530, + 378 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 433, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 433, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 433, + 545, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 433, + 545, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 433, + 545, + 456 + ], + "type": "text", + "content": " is the mixed captured image, and " + }, + { + "bbox": [ + 306, + 433, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\alpha_{t},\\alpha_{r}" + }, + { + "bbox": [ + 306, + 433, + 545, + 456 + ], + "type": "text", + "content": " are blending coefficients resulting from light attenuation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 459, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 639 + ], + "type": "text", + "content": "Polarization image sensors are becoming mainstream, allowing users to easily capture polarization images from a single shot in real-time [32]. However, existing methods [10, 11, 24, 46] typically rely on intensity-based cues, such as pixel brightness and color gradients, to distinguish transmitted and reflected layers. These methods face challenges because reflection removal is a highly ill-posed inverse problem [21] that recovers two unknown layers (reflection and transmission) from a single observation. Polarization provides valuable physics-based cues [27, 29] to alleviate ill-posed issues (Fig. 1), whereas transmitted light retains a distinct degree of polarization(Fig. 2b). This difference provides crucial signals for separating the two layers. Notably, at the Brewster angle (Fig.2c), reflected light is fully polarized, enabling effective reflection removal [2]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Despite significant advances of polarization images, a key challenge in polarization-based reflection removal is the lack of large-scale, high-quality datasets. Existing polarized reflection removal datasets [18, 39] are limited in size and diversity, relying on small (<1000) or synthetic samples that fail to capture the complexity of real-world light-" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 143, + 712 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10890" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 288, + 163 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 288, + 163 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 288, + 163 + ], + "type": "image", + "image_path": "7ea6bef2076cfbf69882589e06b5d3ae6c93b962c7da50576b216d8f73874dfb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 287, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 226 + ], + "type": "text", + "content": "Figure 2. (a) & (b) A semi-reflector transforms unpolarized light into polarized light upon reflection and refraction, which is undetectable by standard RGB cameras but can be leveraged by polarization cameras for reflection-suppression tasks. (c) At the Brewster angle [2], a polarizer minimizes reflections." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 314 + ], + "type": "text", + "content": "ing conditions, materials, and scenes. Moreover, they typically exclude color information, reducing their applicability in real-world reflection removal tasks. Thus, there is a pressing need for a large-scale, comprehensive dataset that includes both RGB and polarization images, captured in diverse real-world environments, to advance polarization-based reflection removal." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 459 + ], + "type": "text", + "content": "To bridge this gap, we introduce PolaRGB, a novel dataset specifically collected for polarization-based reflection removal. As shown in Table 1, PolaRGB contains 6,500 high-quality, well-aligned RGB-polarization image pairs, " + }, + { + "bbox": [ + 46, + 316, + 287, + 459 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 46, + 316, + 287, + 459 + ], + "type": "text", + "content": " larger than previous dataset [18]. Our dataset covers a diverse range of scenes, lighting conditions, and exposure settings, and is captured using off-the-shelf commercial cameras with polarized color patterns to ensure real-world applicability. PolaRGB provides both mixed images and ground-truth transmission layers, enabling accurate reflection separation and significantly enhancing the effectiveness of reflection removal across real-world scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 460, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 568 + ], + "type": "text", + "content": "Moreover, extracting reflection-free information from polarization data is challenging [18, 25, 39] due to the randomness of shooting angles, scene variations, and changing lighting conditions. To address this issue, we leverage the powerful generative capabilities of the diffusion model [3,6,8] to generate reflection-free cues. The diffusion model extracts and refines reflection-free priors from polarization images, effectively guiding the reflection removal and yielding precise and robust reflection-free results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Specifically, PolarFree consists of two steps: a prior-generation step and a reflection removal step. First, in the prior-generation step, we use a diffusion model to generate reflection-free prior based on the polarization and RGB inputs. This strategy not only guides accurate reflection isolation in RGB images but also recovers background details, which previous methods [11, 15, 46] may miss. Next, the reflection removal step leverages the prior to effectively remove reflections, ensuring accurate transmission restoration. Additionally, we introduce a phase-based loss function in the frequency domain to mitigate color discrepancies caused by semi-reflections, guiding the network to focus" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 92, + 545, + 159 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 91 + ], + "type": "text", + "content": "Table 1. Comparisons between PolaRGB and existing reflection removal datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 92, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 308, + 92, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 308, + 92, + 545, + 159 + ], + "type": "table", + "html": "
DatasetPolarizationRGBRAWData sizeResolution
SIR2 [36]NoYesNo500540 × 400
RRW [46]NoYesNo14,9522580 × 1460
ReflectNet [39]Yes (Syn.)YesNo-1500 × 1000
Lei et al. [18]Yes (Real)NoYes8071224 × 1024
PolaRGB (Ours)Yes (Real)YesYes6,5001224 × 1024
", + "image_path": "ba3aca2169ccbb3aa213af4da651194a337f0255e4ba274ee915577c8369ebc6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 165, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 545, + 212 + ], + "type": "text", + "content": "on reflection removal rather than color adjustment. These components enable PolarFree to achieve robust reflection suppression while preserving the clarity and integrity of the transmission across diverse real-world scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 213, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 321 + ], + "type": "text", + "content": "Extensive experiments on the PolaRGB dataset demonstrate the effectiveness of PolarFree, which outperforms existing methods [11, 15, 46] " + }, + { + "bbox": [ + 304, + 213, + 545, + 321 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 304, + 213, + 545, + 321 + ], + "type": "text", + "content": " 2dB in terms of PSNR. Additionally, we performed real-world testing in environments commonly affected by reflections such as museums and galleries, demonstrating the method's effectiveness in real-world scenarios. Our approach improves image clarity and preserves details better than previous methods, paving the way for more practical, reflection-robust applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 322, + 392, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 322, + 392, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 322, + 392, + 334 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 336, + 419, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 336, + 419, + 347 + ], + "spans": [ + { + "bbox": [ + 306, + 336, + 419, + 347 + ], + "type": "text", + "content": "2.1. Reflection Removal" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 505 + ], + "type": "text", + "content": "Reflection removal [1, 30, 34] has a longstanding history in computer vision, aiming to enhance the occluded objects caused by semi-reflectors like glass and windows. Early approaches [20, 24] typically relied on handcrafted priors, such as smoothness assumptions and gradient sparsity, to differentiate the reflection layer from the underlying scene. Some methods incorporated ghosting [31], flash cues [17, 19] and edge consistency [7], while others improved accuracy by leveraging multi-image setups [22, 28], utilizing temporal variation [26] and parallax [28] to separate reflections from background layers. However, these methods often require strict assumptions and fail in real-world scenes with complex textures or lighting [19, 38]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 637 + ], + "type": "text", + "content": "With the rise of deep learning [5, 35, 42, 44, 46], reflection removal methods advanced considerably. Techniques based on neural networks have shown promise by learning to separate reflection and transmission layers directly from data [10, 15, 35-37, 44]. However, without essential physical insights, these methods act as Bayesian regressions toward a learned average, often lead to suboptimal separation in complex environments with varying lighting, textures, and reflection intensities [35, 44]. Recent methods [9, 45] utilize language to interactively remove the reflection, which require additional efforts, limiting their practicality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 640, + 479, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 640, + 479, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 640, + 479, + 651 + ], + "type": "text", + "content": "2.2. Polarization Reflection Removal" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Polarization is an effective way to remove reflections due to the physics principle: the polarization of light behaves differently in the reflection and transmission layers [4, 13, 23, 27, 30], allowing separation, especially at the Brewster angle [4, 27]. However, capturing images exactly" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10891" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 275, + 281 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 275, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 275, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 275, + 281 + ], + "type": "image", + "image_path": "8ca8a3972a20070ec3d390a56d58221c826881066fa30c006b18aa06961f73a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 282, + 243, + 292 + ], + "lines": [ + { + "bbox": [ + 83, + 282, + 243, + 292 + ], + "spans": [ + { + "bbox": [ + 83, + 282, + 243, + 292 + ], + "type": "text", + "content": "(a) Hierarchical structure of PolarRR dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 284, + 70, + 345, + 117 + ], + "blocks": [ + { + "bbox": [ + 284, + 70, + 345, + 117 + ], + "lines": [ + { + "bbox": [ + 284, + 70, + 345, + 117 + ], + "spans": [ + { + "bbox": [ + 284, + 70, + 345, + 117 + ], + "type": "image", + "image_path": "eea8759d742d4f8018cced432db6271d3df78cdc252da9d3585e413096a9e90f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 345, + 70, + 406, + 117 + ], + "blocks": [ + { + "bbox": [ + 345, + 70, + 406, + 117 + ], + "lines": [ + { + "bbox": [ + 345, + 70, + 406, + 117 + ], + "spans": [ + { + "bbox": [ + 345, + 70, + 406, + 117 + ], + "type": "image", + "image_path": "4f283e0279b4acc02c52c830457a3829881c71f71df8625e75cf5f24c7b80dbd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 285, + 118, + 345, + 165 + ], + "blocks": [ + { + "bbox": [ + 285, + 118, + 345, + 165 + ], + "lines": [ + { + "bbox": [ + 285, + 118, + 345, + 165 + ], + "spans": [ + { + "bbox": [ + 285, + 118, + 345, + 165 + ], + "type": "image", + "image_path": "880553008569a864c8d9aaaf274b5fb203da3d7fb67e7d31fd02ab9ce0e8705b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 167, + 400, + 176 + ], + "lines": [ + { + "bbox": [ + 284, + 167, + 400, + 176 + ], + "spans": [ + { + "bbox": [ + 284, + 167, + 400, + 176 + ], + "type": "text", + "content": "(b) Diverse reflection conditions" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 346, + 118, + 406, + 165 + ], + "blocks": [ + { + "bbox": [ + 346, + 118, + 406, + 165 + ], + "lines": [ + { + "bbox": [ + 346, + 118, + 406, + 165 + ], + "spans": [ + { + "bbox": [ + 346, + 118, + 406, + 165 + ], + "type": "image", + "image_path": "65ac40515c73091e5f6d18db5c5e31dda4b9c865764a52d22efd4ac11e8b5785.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 411, + 70, + 511, + 165 + ], + "blocks": [ + { + "bbox": [ + 411, + 70, + 511, + 165 + ], + "lines": [ + { + "bbox": [ + 411, + 70, + 511, + 165 + ], + "spans": [ + { + "bbox": [ + 411, + 70, + 511, + 165 + ], + "type": "image", + "image_path": "d2809239934c849ed6ebd53e3634f7a3bd7cb6960776f5947e5061b3abe70246.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 167, + 533, + 176 + ], + "lines": [ + { + "bbox": [ + 419, + 167, + 533, + 176 + ], + "spans": [ + { + "bbox": [ + 419, + 167, + 533, + 176 + ], + "type": "text", + "content": "(c) Video-based capture method" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 512, + 70, + 543, + 165 + ], + "blocks": [ + { + "bbox": [ + 512, + 70, + 543, + 165 + ], + "lines": [ + { + "bbox": [ + 512, + 70, + 543, + 165 + ], + "spans": [ + { + "bbox": [ + 512, + 70, + 543, + 165 + ], + "type": "image", + "image_path": "6d5014dcc4f28f019f5fe6eb6c70c122dc15cc435762b52fce29220e54f21187.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 285, + 183, + 349, + 232 + ], + "blocks": [ + { + "bbox": [ + 285, + 183, + 349, + 232 + ], + "lines": [ + { + "bbox": [ + 285, + 183, + 349, + 232 + ], + "spans": [ + { + "bbox": [ + 285, + 183, + 349, + 232 + ], + "type": "image", + "image_path": "d64caf525b3d078a6dd1b940c763a7d5c9710b205b7241eac8eeab4d949d7754.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 349, + 183, + 413, + 232 + ], + "blocks": [ + { + "bbox": [ + 349, + 183, + 413, + 232 + ], + "lines": [ + { + "bbox": [ + 349, + 183, + 413, + 232 + ], + "spans": [ + { + "bbox": [ + 349, + 183, + 413, + 232 + ], + "type": "image", + "image_path": "45136a25128f63284bc3a2d8b0d899872b13cd308c274102f896c51a81f2162f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 414, + 183, + 478, + 232 + ], + "blocks": [ + { + "bbox": [ + 414, + 183, + 478, + 232 + ], + "lines": [ + { + "bbox": [ + 414, + 183, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 414, + 183, + 478, + 232 + ], + "type": "image", + "image_path": "937d9fcc2dde35c82bb94007feabde22a2a2d9455eec666fa2846b7d345f7f63.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 479, + 183, + 543, + 232 + ], + "blocks": [ + { + "bbox": [ + 479, + 183, + 543, + 232 + ], + "lines": [ + { + "bbox": [ + 479, + 183, + 543, + 232 + ], + "spans": [ + { + "bbox": [ + 479, + 183, + 543, + 232 + ], + "type": "image", + "image_path": "7c37ca18470fac448ded46c855863281c1dce2742f0ac08118c99ef13030baa6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 285, + 232, + 349, + 278 + ], + "blocks": [ + { + "bbox": [ + 285, + 232, + 349, + 278 + ], + "lines": [ + { + "bbox": [ + 285, + 232, + 349, + 278 + ], + "spans": [ + { + "bbox": [ + 285, + 232, + 349, + 278 + ], + "type": "image", + "image_path": "de043a9bc68b63d70554f9cae2c1a74cd12ae5e74ddcec154aa0adf9024cb4a1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 295, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 295, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 545, + 361 + ], + "type": "text", + "content": "Figure 3. Overview of the PolaRGB dataset. (a) Hierarchical structure of scenes is shown in the ring, with legends indicating sample counts and subset types. (b) Typical scenes illustrating varied reflection conditions: I. smooth blending of reflection and refraction, II. abrupt reflection with mixed components, III. reflection dominant over transmission, and IV. minimal or no reflection. (c) Video-based capture method (details in Sec. 3). (d) We provide polarized images at angles " + }, + { + "bbox": [ + 46, + 295, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\phi = 0^{\\circ}" + }, + { + "bbox": [ + 46, + 295, + 545, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 295, + 545, + 361 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 46, + 295, + 545, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 295, + 545, + 361 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 295, + 545, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 295, + 545, + 361 + ], + "type": "inline_equation", + "content": "135^{\\circ}" + }, + { + "bbox": [ + 46, + 295, + 545, + 361 + ], + "type": "text", + "content": ", along with derived AoLP, DoLP, and a well-aligned unpolarized image. The dataset also includes ground truth transmission and estimated reflections, all available in both raw and RGB formats." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 349, + 232, + 414, + 278 + ], + "blocks": [ + { + "bbox": [ + 349, + 232, + 414, + 278 + ], + "lines": [ + { + "bbox": [ + 349, + 232, + 414, + 278 + ], + "spans": [ + { + "bbox": [ + 349, + 232, + 414, + 278 + ], + "type": "image", + "image_path": "a16e0d41a7120c3aa13d8dea18c55b320b56fb14e74871494372cae3881f862c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 281, + 481, + 293 + ], + "lines": [ + { + "bbox": [ + 341, + 281, + 481, + 293 + ], + "spans": [ + { + "bbox": [ + 341, + 281, + 481, + 293 + ], + "type": "text", + "content": "(d) Image types included in the dataset" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 414, + 232, + 478, + 278 + ], + "blocks": [ + { + "bbox": [ + 414, + 232, + 478, + 278 + ], + "lines": [ + { + "bbox": [ + 414, + 232, + 478, + 278 + ], + "spans": [ + { + "bbox": [ + 414, + 232, + 478, + 278 + ], + "type": "image", + "image_path": "afd0ba01fe5cf90e494b5945746281ecf547a5c58f87a2eab31ecbb57e1b4af3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 479, + 232, + 543, + 278 + ], + "blocks": [ + { + "bbox": [ + 479, + 232, + 543, + 278 + ], + "lines": [ + { + "bbox": [ + 479, + 232, + 543, + 278 + ], + "spans": [ + { + "bbox": [ + 479, + 232, + 543, + 278 + ], + "type": "image", + "image_path": "bcf55f5dec9a9232227e777e00f74312bf5ea9a7612f9123117f7bbdc71fbab0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 362, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 287, + 411 + ], + "type": "text", + "content": "at the Brewster angle, as shown in Fig. 2c, is challenging. Therefore, in practice, methods typically rely on images taken at multiple polarization angles (e.g., " + }, + { + "bbox": [ + 47, + 362, + 287, + 411 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 47, + 362, + 287, + 411 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 362, + 287, + 411 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 47, + 362, + 287, + 411 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 362, + 287, + 411 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 47, + 362, + 287, + 411 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 362, + 287, + 411 + ], + "type": "inline_equation", + "content": "135^{\\circ}" + }, + { + "bbox": [ + 47, + 362, + 287, + 411 + ], + "type": "text", + "content": ") to capture polarization information [4, 27, 30]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 411, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 651 + ], + "type": "text", + "content": "Early polarized reflection removal methods [4, 27, 30] employ mathematical techniques like Principle Component Analysis (PCA) to separate reflection layers, which are effective when the transmission and reflection layers have significant content differences. Subsequently, Kong et al. [16] propose a multiscale scheme that automatically identifies the optimal separation of the reflection and background layers. Another advancements [39] introduce a neural network-based approach for polarization-guided reflection removal with synthetic dataset. Lyu et al. [25] utilize a pair of unpolarized and polarized images but lack of realistic polarization patterns. Lei et al. [18] further contribute by collecting a dataset of polarized images for reflection removal. However, this dataset only includes limited scene variations from pure polarization scenes, which limits its application scenario. In contrast, we present the first large-scale dataset captured for polarization-based RGB reflection removal, consisting of 6,500 well-aligned transmission-reflection pairs. We also propose a polarization-based reflection removal network based on diffusion models." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 654, + 152, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 152, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 152, + 666 + ], + "type": "text", + "content": "3. PolaRGB Dataset" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 672, + 110, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 110, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 110, + 685 + ], + "type": "text", + "content": "3.1. Analysis" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Although a few polarization-based reflection removal datasets exist [18, 39], they face major limitations: (1) they" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 362, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 445 + ], + "type": "text", + "content": "are often synthetically generated [25, 39], which may not generalize well to real polarization, as perfectly simulating polarization phenomena is very hard; (2) they are typically limited in size and diversity (<1,000 samples), with limited scene variations, which constrains model robustness and generalizability [18, 39]; and (3) they lack RGB data [18, 39], reducing the practical value of these datasets." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 446, + 546, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 546, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 546, + 637 + ], + "type": "text", + "content": "Our polarized-based reflection removal dataset, PolaRGB, addresses all these limitations with the following key features, as shown in Fig. 3. It has following advantages: (1) It contains real-captured images with perfect spatial alignment. We achieve pixel-level alignment through a three-step process: careful capture setup, manual filtering, and homography transformations, all in raw space to avoid demosaicing artifacts (see Fig. 4b and more details in supplementary). (2) It contains large-scale diverse scenes. The collected dataset covers a wide range of indoor and outdoor settings with four distinct reflection types, including smooth, sharp, high-brightness, and subtle reflections (see Fig. 3b). (3) It has extensive data modalities. Our dataset offers paired mixed and transmission images with both polarization and RGB captures, precisely aligned and available in both raw and RGB formats (see Fig. 3d)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 640, + 419, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 640, + 419, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 640, + 419, + 653 + ], + "type": "text", + "content": "3.2. Processing Pipeline" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": "We leverage an efficient video-based capture flow for dataset collection, as shown in Fig. 3c. First, we capture a transmission-only image " + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "inline_equation", + "content": "T_{raw}" + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": " as the ground truth. Then, we place a semi-reflective glass plate in front of the scene and rotate it continuously to capture mixed images" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10892" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 408 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 408 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 408 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 408 + ], + "type": "image", + "image_path": "718d1884c76556825fa944bdbda653d24baa38fb79c5faeddbee1429641b0068.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 411, + 287, + 434 + ], + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 434 + ], + "type": "text", + "content": "Figure 4. Data processing pipeline for obtaining aligned mixed and transmission images, and polarized images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 437, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 287, + 472 + ], + "type": "inline_equation", + "content": "M_{\\text{raw}}" + }, + { + "bbox": [ + 47, + 437, + 287, + 472 + ], + "type": "text", + "content": ". A division-of-focal-plane polarization camera with color Bayer pattern is used to simultaneously capture both polarized and RGB information, as shown in Fig. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "text", + "content": "After obtaining the raw mixed image " + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "inline_equation", + "content": "M_{raw}" + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "text", + "content": " and the transmission image " + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "inline_equation", + "content": "T_{raw}" + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "text", + "content": ", we process the images sequentially in the raw and RGB domains as shown in Fig. 4a. In the raw domain, we first align " + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "inline_equation", + "content": "M_{raw}" + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "inline_equation", + "content": "T_{raw}" + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "text", + "content": " to correct spatial misalignment caused by light refraction. This is done by separating the images into different polarization angles and color channels, then applying affine transformation matrices to each of the channels to avoid aliasing from direct alignment, as shown in Fig. 4b. Next, we perform polarization separation on the aligned images to obtain four polarization images " + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "inline_equation", + "content": "(0^{\\circ}, 45^{\\circ}, 90^{\\circ}" + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "inline_equation", + "content": "135^{\\circ}" + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "text", + "content": "), and use Malus's law [27, 29] to sum them, producing the unpolarized image of the scene. Finally, we estimate the reflection image by searching for the optimal blending coefficient " + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\alpha_r" + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 46, + 473, + 287, + 665 + ], + "type": "text", + "content": " in edge space, as illustrated in Eq. 1. For detailed steps and proof, please refer to the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 714 + ], + "type": "text", + "content": "Next, we apply demosaicking to the processed raw images to obtain the mixed, transmission, and reflection images, each consisting of four polarization images and the unpolarized RGB image. Additionally, we compute the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 547, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 107 + ], + "type": "text", + "content": "Stokes parameters (see Sec. 4.1) to estimate the Angle of Linear Polarization (AOLP) and Degree of Linear Polarization (DOLP), which are then used in reflection removal." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 110, + 362, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 110, + 362, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 110, + 362, + 121 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 126, + 430, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 126, + 430, + 138 + ], + "spans": [ + { + "bbox": [ + 306, + 126, + 430, + 138 + ], + "type": "text", + "content": "4.1. Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": "Object. Given an RGB image and its corresponding polarization images, we aim to recover the transmission layer of the RGB image by utilizing the distinct polarization characteristics. Our measurement consists of spatially-aligned RGB and polarization images, where the polarization images captured at four distinct angles " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "(0^{\\circ}, 45^{\\circ}, 90^{\\circ}" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "inline_equation", + "content": "135^{\\circ})" + }, + { + "bbox": [ + 304, + 142, + 545, + 251 + ], + "type": "text", + "content": ". These four angles provide a comprehensive polarization measurement [30], which allows us to compute polarization features crucial for separating layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 255, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 307 + ], + "type": "text", + "content": "Polarization Preliminary. For semi-reflective surfaces that produce mixed images, the observed intensity " + }, + { + "bbox": [ + 304, + 255, + 545, + 307 + ], + "type": "inline_equation", + "content": "I_M^\\phi (x)" + }, + { + "bbox": [ + 304, + 255, + 545, + 307 + ], + "type": "text", + "content": " can be decomposed into reflected intensity " + }, + { + "bbox": [ + 304, + 255, + 545, + 307 + ], + "type": "inline_equation", + "content": "I_R^\\phi (x)" + }, + { + "bbox": [ + 304, + 255, + 545, + 307 + ], + "type": "text", + "content": " and transmitted intensity " + }, + { + "bbox": [ + 304, + 255, + 545, + 307 + ], + "type": "inline_equation", + "content": "I_T^\\phi (x)" + }, + { + "bbox": [ + 304, + 255, + 545, + 307 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 359, + 308, + 545, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 308, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 359, + 308, + 545, + 323 + ], + "type": "interline_equation", + "content": "I _ {M} ^ {\\phi} (x) = \\alpha_ {R} I _ {R} ^ {\\phi} (x) + \\alpha_ {T} I _ {T} ^ {\\phi} (x). \\tag {2}", + "image_path": "e3f05032859a82d728d873325ac35988bd833946c3e9ad3b83355e9b6045fb07.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 324, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 348 + ], + "type": "text", + "content": "If a linear polarizer is placed in front of the camera at an angle " + }, + { + "bbox": [ + 305, + 324, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 324, + 545, + 348 + ], + "type": "text", + "content": ", the captured intensity " + }, + { + "bbox": [ + 305, + 324, + 545, + 348 + ], + "type": "inline_equation", + "content": "I^{\\phi}(x)" + }, + { + "bbox": [ + 305, + 324, + 545, + 348 + ], + "type": "text", + "content": " can be represented as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 356, + 545, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 356, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 310, + 356, + 545, + 371 + ], + "type": "interline_equation", + "content": "I _ {M} ^ {\\phi} (x) = \\alpha (\\theta ; \\phi ; \\phi_ {\\perp}) I _ {R} ^ {\\phi} (x) + (1 - \\alpha (\\theta ; \\phi ; \\phi_ {\\parallel})) I _ {T} ^ {\\phi} (x), \\tag {3}", + "image_path": "7747ea2f1bdfcb9600716ed15ca84e115adcb49bea09e4189190565cfaf4dacf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\alpha (\\theta ;\\phi ;\\phi_{\\perp})" + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "inline_equation", + "content": "(1 - \\alpha (\\theta ;\\phi ;\\phi_{\\parallel}))" + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "content": " are polarization-dependent mixing coefficients based on the angle of incidence " + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\theta ,\\phi_{\\perp}" + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\phi_{\\parallel}" + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "content": " are the canonical polarization directions for the reflection and transmission, respectively. Directly solving for these parameters is challenging because " + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\alpha (\\cdot)" + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "inline_equation", + "content": "I_R^\\phi (x)" + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "inline_equation", + "content": "I_R^\\phi (x)" + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "content": " are all unknown [27, 39]. Additionally, relationship between the observed intensity and the underlying reflected and transmitted components is highly nonlinear and influenced by various factors [39]. To address this, we utilize Stokes parameters [18] that provide a efficient way to represent and analyze polarized light, enabling more robust separation of reflection and transmission." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 528, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 575 + ], + "type": "text", + "content": "Stokes Parameters. To capture the polarization effects in the scene, we use Stokes parameters " + }, + { + "bbox": [ + 304, + 528, + 545, + 575 + ], + "type": "inline_equation", + "content": "[S_0, S_1, S_2]" + }, + { + "bbox": [ + 304, + 528, + 545, + 575 + ], + "type": "text", + "content": ", which can be derived from intensity measurements at specific polarization angles " + }, + { + "bbox": [ + 304, + 528, + 545, + 575 + ], + "type": "inline_equation", + "content": "(0^\\circ, 45^\\circ, 90^\\circ" + }, + { + "bbox": [ + 304, + 528, + 545, + 575 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 528, + 545, + 575 + ], + "type": "inline_equation", + "content": "135^\\circ" + }, + { + "bbox": [ + 304, + 528, + 545, + 575 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 349, + 576, + 545, + 603 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 576, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 349, + 576, + 545, + 603 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} S _ {0} = \\left(I _ {0 ^ {\\circ}} + I _ {4 5 ^ {\\circ}} + I _ {9 0 ^ {\\circ}} + I _ {1 3 5 ^ {\\circ}}\\right) / 2, (4) \\\\ S _ {1} = I _ {0 ^ {\\circ}} - I _ {9 0 ^ {\\circ}}, S _ {2} = I _ {4 5 ^ {\\circ}} - I _ {1 3 5 ^ {\\circ}}. (4) \\\\ \\end{array}", + "image_path": "bd177e8dbce2ce50e120254ec19388c7d889192249a88af11e4ee406fe797cbe.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": " represents the total intensity of light, and " + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "S_1" + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "S_2" + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": " provide information about the linear polarization state of the light based on intensity differences at these key angles." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 677 + ], + "type": "text", + "content": "Using the Stokes parameters, we compute the Degree of Linear Polarization (DOLP) and Angle of Linear Polarization (AOLP) as follows [2]" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 345, + 677, + 545, + 702 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 677, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 345, + 677, + 545, + 702 + ], + "type": "interline_equation", + "content": "D O L P (x) = \\sqrt {S _ {1} (x) ^ {2} + S _ {2} (x) ^ {2}} / S _ {0} (x), \\tag {5}", + "image_path": "c21f317c0b5d6cc9ef9fa86868fe7bfda716aa388e357404c249e93656513e5a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 347, + 689, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 689, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 347, + 689, + 545, + 717 + ], + "type": "interline_equation", + "content": "A O L P (x) = \\frac {1}{2} \\operatorname {a t a n 2} \\left(S _ {2} (x) / S _ {1} (x)\\right). \\tag {5}", + "image_path": "c57a647bf38abdf2ddc73a9dbb47143cc4dfa9f070a0c3ffb8ffea6c5fef6cea.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10893" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 255, + 277 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 255, + 277 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 255, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 255, + 277 + ], + "type": "image", + "image_path": "7de8f37e913ce7e31160fa6f024fdfaef8403b25dd3642752a06ce45afc6a5af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 280, + 184, + 289 + ], + "lines": [ + { + "bbox": [ + 136, + 280, + 184, + 289 + ], + "spans": [ + { + "bbox": [ + 136, + 280, + 184, + 289 + ], + "type": "text", + "content": "(a) Inference" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "type": "text", + "content": "Figure 5. Pipeline of PolarFree. (a) During inference, PolarFree leverages polarized and RGB images as inputs, which are feeds into a conditional diffusion model to generate the prior " + }, + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "type": "text", + "content": ". The generated prior, along with the inputs, is then passed to the reflection removal backbone " + }, + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{remove}}" + }, + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "type": "text", + "content": " to remove reflections. (b) PolarFree is trained in two stages. (1) A prior encoder extracts a reflection-free prior " + }, + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 292, + 545, + 358 + ], + "type": "text", + "content": " from clean transmission images and polarization cues, which serves as the supervision for the conditional diffusion model in stage two. (2) The conditional diffusion model is trained to progressively denoise noisy images, supervised by the prior from stage one, ensuring robust reflection separation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 258, + 70, + 541, + 277 + ], + "blocks": [ + { + "bbox": [ + 258, + 70, + 541, + 277 + ], + "lines": [ + { + "bbox": [ + 258, + 70, + 541, + 277 + ], + "spans": [ + { + "bbox": [ + 258, + 70, + 541, + 277 + ], + "type": "image", + "image_path": "6de156d3ccb9bc576f12eb41b7bcc4bdf4e035160926479af57b03b528b5875d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 280, + 431, + 289 + ], + "lines": [ + { + "bbox": [ + 385, + 280, + 431, + 289 + ], + "spans": [ + { + "bbox": [ + 385, + 280, + 431, + 289 + ], + "type": "text", + "content": "(b) Training" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 362, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 469 + ], + "type": "text", + "content": "In this formulation, DOLP describes the proportion of polarized light relative to total intensity, which helps indicate the degree of reflection in the scene. AOLP, on the other hand, reveals the orientation of the polarized light, allowing us to distinguish between reflection and transmission components more effectively, as shown in Fig. 3d. By leveraging intensity measurements at " + }, + { + "bbox": [ + 46, + 362, + 287, + 469 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 46, + 362, + 287, + 469 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 362, + 287, + 469 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 46, + 362, + 287, + 469 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 362, + 287, + 469 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 362, + 287, + 469 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 362, + 287, + 469 + ], + "type": "inline_equation", + "content": "135^{\\circ}" + }, + { + "bbox": [ + 46, + 362, + 287, + 469 + ], + "type": "text", + "content": ", these parameters provide valuable cues for separating mixed reflection and transmission layers in semi-reflective scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 472, + 160, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 160, + 482 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 160, + 482 + ], + "type": "text", + "content": "4.2. PolarFree Network" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 483, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 628 + ], + "type": "text", + "content": "To achieve high-quality reflection removal with effective utilization of polarization information, we introduce Polar-Free, a delicately designed two-step network, where each step addresses a distinct aspect of the reflection removal challenge. Inspired by [3], we leverage diffusion model to generate the prior for reflection removal. As shown in Fig. 5, during inference, the first step utilizes a conditional diffusion model to extract reflection-free priors, effectively isolating essential details from polarization data. The second step integrates these priors with RGB inputs, guiding the network to accurately separate reflections and enhance clarity, even in complex, real-world environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 633, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 287, + 669 + ], + "type": "text", + "content": "Prior Generation. As shown in Fig. 5a, in the first step, we start with a randomly initialized noise " + }, + { + "bbox": [ + 47, + 633, + 287, + 669 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 633, + 287, + 669 + ], + "type": "text", + "content": ", which is gradually denoised through a conditional diffusion model " + }, + { + "bbox": [ + 47, + 633, + 287, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{diff}}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 123, + 672, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 672, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 123, + 672, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {0} = \\mathcal {F} _ {\\text {d i f f}} (n | M _ {\\text {c o n d}}), \\tag {6}", + "image_path": "c914be96e2e6e258ce83a7eb43bb693d3c02af01d7e4904a55a808327c0b6a47.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " is the generated prior, " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{cond}} = \\{M_{\\mathrm{polar}}, M_{\\mathrm{aolp}}, M_{\\mathrm{dolp}}, M_{\\mathrm{rgb}}\\}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " represents the mixed im" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 362, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 545, + 411 + ], + "type": "text", + "content": "ages, and " + }, + { + "bbox": [ + 304, + 362, + 545, + 411 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 362, + 545, + 411 + ], + "type": "text", + "content": " is the initial noise. Through an iterative denoising process, the diffusion model progressively refines a noise image, conditioned on the polarization and RGB data, to generate a reflection prior." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "type": "text", + "content": "The denoising process follows the denoising diffusion probabilistic model (DDPM) framework [8] and employs a U-Net architecture to predict noise. At each timestep " + }, + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "type": "text", + "content": ", the U-Net receives a noisy intermediate " + }, + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "type": "text", + "content": " and outputs a noise estimate " + }, + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_{t},t)" + }, + { + "bbox": [ + 304, + 412, + 545, + 483 + ], + "type": "text", + "content": ", predicting the noise added at that timestep. This process can be formulated as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 483, + 545, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 483, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 312, + 483, + 545, + 510 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(z _ {t}, M _ {\\text {c o n d}}, t\\right)\\right) + \\sigma_ {t} z, \\tag {7}", + "image_path": "c9192a70d5a7e9325a48dd69009ad9b2658acd98ac529cd483c5f5f9b51d9298.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "text", + "content": " control the noise schedule across the timesteps, " + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "text", + "content": " is the standard deviation, and " + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "text", + "content": " is noise sampled from a standard Gaussian distribution. " + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t}" + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "text", + "content": " is the cumulative product of " + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "text", + "content": ", indicating noise level to timestep " + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 511, + 545, + 559 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 561, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 545, + 621 + ], + "type": "text", + "content": "In this way, the U-Net gradually removes the noise in each step, conditioning on the polarization-based measurement input " + }, + { + "bbox": [ + 304, + 561, + 545, + 621 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{cond}}" + }, + { + "bbox": [ + 304, + 561, + 545, + 621 + ], + "type": "text", + "content": ". This process continues until noise is fully removed, yielding a sample " + }, + { + "bbox": [ + 304, + 561, + 545, + 621 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 304, + 561, + 545, + 621 + ], + "type": "text", + "content": " that represents the reflection-free prior distribution extracted from the conditioned inputs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "text", + "content": "Reflection Removal. As shown in Fig. 5a, once the prior " + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "text", + "content": " has been obtained, the second step of PolarFree is separating the transmission and reflection layers. By leveraging the polarization cues provided by " + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{polar}}" + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{aolp}}" + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{dolp}}" + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "text", + "content": ", the model " + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{remove}}" + }, + { + "bbox": [ + 304, + 628, + 545, + 700 + ], + "type": "text", + "content": " separates the transmission features from the reflection ones, which can be represented as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 370, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 370, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\hat {T} _ {r g b} = \\mathcal {F} _ {\\text {r e m o v e}} \\left(\\hat {z} _ {0}, M _ {\\text {c o n d}}\\right), \\tag {8}", + "image_path": "7ba2f1cfe71a3df67a827791d77ac5a49888fe7fb365768797bd1c1669fca4fc.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10894" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 272, + 173 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 272, + 173 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 272, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 272, + 173 + ], + "type": "image", + "image_path": "be2378f448fa49b76e8cc9819f7cdac0e5aa1f09f0737731308edce26a1fc292.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 174, + 286, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 286, + 227 + ], + "type": "text", + "content": "Figure 6. (a) Phase information preserves shape and texture details, while color primarily affects the amplitude. (b) We apply two types of random color perturbations to the image and compute the perturbation errors. Phase-based loss is less sensitive to color changes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 286, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 286, + 278 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 231, + 286, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{remove}}" + }, + { + "bbox": [ + 46, + 231, + 286, + 278 + ], + "type": "text", + "content": " is a reflection removal neural network. This step ensures that the final output contains distinct transmission and reflection components, facilitating high-quality image reconstruction under challenging reflective scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 284, + 111, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 111, + 296 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 111, + 296 + ], + "type": "text", + "content": "4.3. Training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 300, + 286, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 286, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 286, + 372 + ], + "type": "text", + "content": "To train our PolarFree, particularly to address the challenge of the diffusion model lacking a suitable ground truth, we adopt a two-stage training strategy, as illustrated in Fig. 5b. This strategy consists of two sequential objectives: learning to extract a reflection-free prior and learning to generate a reflection-free prior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "text", + "content": "First Stage. In the first stage, we train an encoder to extract a prior " + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "text", + "content": " for reflection-free information, which will serve as the supervision signal for the diffusion model in the second stage. Specifically, we feed the ground truth transmission images " + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{cond}} = \\{T_{\\mathrm{polar}}, T_{\\mathrm{aolp}}, T_{\\mathrm{dolp}}, T_{\\mathrm{rgb}}\\}" + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "text", + "content": " into the encoder " + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "text", + "content": ", obtaining " + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(M_{\\mathrm{cond}}, T_{\\mathrm{rgb}})" + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "text", + "content": ". This " + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "text", + "content": " contains reflection-free information enriched with polarization-related cues. Subsequently, " + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "text", + "content": " conditions the reflection removal backbone " + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{remove}}" + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "text", + "content": " to predict the clean transmission image " + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "inline_equation", + "content": "\\hat{T}_{\\mathrm{rgb}} = \\mathcal{F}_{\\mathrm{remove}}(z, M_{\\mathrm{cond}})" + }, + { + "bbox": [ + 46, + 377, + 286, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 503, + 286, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 286, + 586 + ], + "type": "text", + "content": "Second Stage. In the second stage, we train the diffusion model " + }, + { + "bbox": [ + 47, + 503, + 286, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 47, + 503, + 286, + 586 + ], + "type": "text", + "content": " to generate a reflection-free prior " + }, + { + "bbox": [ + 47, + 503, + 286, + 586 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 47, + 503, + 286, + 586 + ], + "type": "text", + "content": " from noisy input images, and finetune the reflection removal backbone " + }, + { + "bbox": [ + 47, + 503, + 286, + 586 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{remove}}" + }, + { + "bbox": [ + 47, + 503, + 286, + 586 + ], + "type": "text", + "content": ". The key challenge here is that the diffusion model lacks of direct supervision. Therefore, we leverage the prior " + }, + { + "bbox": [ + 47, + 503, + 286, + 586 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 47, + 503, + 286, + 586 + ], + "type": "text", + "content": " extracted in the first stage as the supervision signal to guide the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 588, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 286, + 635 + ], + "type": "text", + "content": "We start with the extracted reflection-free prior " + }, + { + "bbox": [ + 46, + 588, + 286, + 635 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 588, + 286, + 635 + ], + "type": "text", + "content": " and add noise over multiple timesteps. This process transforms the \"ground-truth\" prior " + }, + { + "bbox": [ + 46, + 588, + 286, + 635 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 588, + 286, + 635 + ], + "type": "text", + "content": " into noisy versions " + }, + { + "bbox": [ + 46, + 588, + 286, + 635 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 46, + 588, + 286, + 635 + ], + "type": "text", + "content": " at each timestep " + }, + { + "bbox": [ + 46, + 588, + 286, + 635 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 588, + 286, + 635 + ], + "type": "text", + "content": ", which can be expressed as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 95, + 636, + 286, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 636, + 286, + 650 + ], + "spans": [ + { + "bbox": [ + 95, + 636, + 286, + 650 + ], + "type": "interline_equation", + "content": "q \\left(z _ {t} \\mid z _ {0}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {0}, \\beta_ {t} \\mathbf {I}\\right), \\tag {9}", + "image_path": "f73fee172ecd8ef24f5b1e3981076b9948cb44560d924683dae087dc0f6e7226.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 652, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 652, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 652, + 286, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 652, + 286, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 46, + 652, + 286, + 688 + ], + "type": "text", + "content": " represents a Gaussian distribution, " + }, + { + "bbox": [ + 46, + 652, + 286, + 688 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 46, + 652, + 286, + 688 + ], + "type": "text", + "content": " is the noise schedule controlling the noise level added at each timestep, " + }, + { + "bbox": [ + 46, + 652, + 286, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 652, + 286, + 688 + ], + "type": "text", + "content": " represents the identity matrix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "In the reverse diffusion process, the model is trained to progressively remove the noise step by step to recover the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 73, + 480, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 73, + 480, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 73, + 480, + 83 + ], + "type": "text", + "content": "clean prior " + }, + { + "bbox": [ + 305, + 73, + 480, + 83 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 305, + 73, + 480, + 83 + ], + "type": "text", + "content": ", which can be represented as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 84, + 544, + 98 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 84, + 544, + 98 + ], + "spans": [ + { + "bbox": [ + 321, + 84, + 544, + 98 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t, M _ {\\text {c o n d}}\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right)), \\tag {10}", + "image_path": "05765e69987bdb4fce5961fbbc3823f2a88fbfa636d6d58dfc4598eb26de9ae0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 100, + 544, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 100, + 544, + 124 + ], + "spans": [ + { + "bbox": [ + 305, + 100, + 544, + 124 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 100, + 544, + 124 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(z_t,t,M_{\\mathrm{cond}})" + }, + { + "bbox": [ + 305, + 100, + 544, + 124 + ], + "type": "text", + "content": " is the mean function of the current state " + }, + { + "bbox": [ + 305, + 100, + 544, + 124 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 305, + 100, + 544, + 124 + ], + "type": "text", + "content": ", timestep " + }, + { + "bbox": [ + 305, + 100, + 544, + 124 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 100, + 544, + 124 + ], + "type": "text", + "content": ", and a conditional input " + }, + { + "bbox": [ + 305, + 100, + 544, + 124 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{cond}}" + }, + { + "bbox": [ + 305, + 100, + 544, + 124 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 125, + 544, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 544, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 544, + 184 + ], + "type": "text", + "content": "Throughout the reverse process, the model conditions on the polarization measurements " + }, + { + "bbox": [ + 304, + 125, + 544, + 184 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{cond}}" + }, + { + "bbox": [ + 304, + 125, + 544, + 184 + ], + "type": "text", + "content": ", which provide essential information about the scene's physical properties. These measurements help the model accurately generate reflection-free components during the denoising process." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 185, + 545, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 257 + ], + "type": "text", + "content": "After the reverse process completes, the model outputs a final reflection-free prior " + }, + { + "bbox": [ + 304, + 185, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 304, + 185, + 545, + 257 + ], + "type": "text", + "content": ", which contains clean transmission information. This prior is used as the guidance signal for the reflection removal backbone " + }, + { + "bbox": [ + 304, + 185, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{remove}}" + }, + { + "bbox": [ + 304, + 185, + 545, + 257 + ], + "type": "text", + "content": ". The backbone takes this reflection-free prior and the polarization data to predict the clean transmission image " + }, + { + "bbox": [ + 304, + 185, + 545, + 257 + ], + "type": "inline_equation", + "content": "\\hat{T}_{rgb}" + }, + { + "bbox": [ + 304, + 185, + 545, + 257 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 259, + 358, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 358, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 358, + 270 + ], + "type": "text", + "content": "4.4. Losses" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "text", + "content": "Basic Loss. To optimize PolarFree, we follow [46] to utilize three basic losses: L1 loss, VGG perceptual loss [14], and total variation (TV) loss [33]. The L1 loss minimizes the pixel-wise difference between the predicted transmission image " + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\hat{T}_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "text", + "content": " and the ground truth transmission image " + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1 = \\|\\hat{T}_{\\mathrm{rgb}} - T_{\\mathrm{rgb}}\\|_1" + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "text", + "content": ". The VGG perceptual loss compares feature activations from a pre-trained VGG network [14], weighted by " + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\lambda_l" + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{VGG}} = \\sum_l \\lambda_l \\|\\phi_l(\\hat{T}_{\\mathrm{rgb}}) - \\phi_l(T_{\\mathrm{rgb}})\\|_1" + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\phi_l" + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "text", + "content": " denotes the activations of the " + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "text", + "content": "-th VGG layer. We also use TV loss to constrain consistency via the gradient operator " + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TV}} = \\|\\nabla \\hat{T}_{\\mathrm{rgb}} - \\nabla T_{\\mathrm{rgb}}\\|_1" + }, + { + "bbox": [ + 304, + 275, + 545, + 409 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 413, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 509 + ], + "type": "text", + "content": "Phase loss. While basic losses help the network match color and intensity values to the ground truth, they struggle with color discrepancies caused by semi-reflective surfaces during dataset collection. These variations, due to reflection and transmission properties of materials, affect the color and intensity, leading to mismatches between the predicted and target images and hindering the model's ability to learn the correct transmission map." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 510, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 579 + ], + "type": "text", + "content": "To address this issue, we introduce a phase loss to focus on the structural information of the transmission, which is less sensitive to color variations. As shown in Fig. 6, phase information primarily captures the geometry and texture of the image, independent of color changes. The phase loss is formulated as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 581, + 544, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 581, + 544, + 594 + ], + "spans": [ + { + "bbox": [ + 320, + 581, + 544, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p h a s e}} = \\| \\angle (F F T (\\hat {T})) - \\angle (F F T ((T _ {r g b})) \\| _ {1}, \\tag {11}", + "image_path": "3006c2958868cd6976283e13477918cc6d0e029f233deaa0e00d9466c03a9190.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 597, + 544, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 544, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 544, + 620 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 597, + 544, + 620 + ], + "type": "inline_equation", + "content": "FFT" + }, + { + "bbox": [ + 304, + 597, + 544, + 620 + ], + "type": "text", + "content": " is the Fourier transform, and " + }, + { + "bbox": [ + 304, + 597, + 544, + 620 + ], + "type": "inline_equation", + "content": "\\angle (\\cdot)" + }, + { + "bbox": [ + 304, + 597, + 544, + 620 + ], + "type": "text", + "content": " represents the phase angle of the Fourier coefficients." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 625, + 544, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 544, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 544, + 672 + ], + "type": "text", + "content": "Diffusion Loss. The diffusion loss follows the standard DDPM formulation [8], where the model predicts the noise at each step " + }, + { + "bbox": [ + 304, + 625, + 544, + 672 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 625, + 544, + 672 + ], + "type": "text", + "content": " and minimizes the difference between the predicted noise and the true noise as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 323, + 673, + 544, + 688 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 673, + 544, + 688 + ], + "spans": [ + { + "bbox": [ + 323, + 673, + 544, + 688 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i f f} = \\mathbb {E} _ {q \\left(z _ {t} \\mid z _ {t - 1}\\right)} \\left[ \\| \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) - \\epsilon_ {\\text {t r u e}} \\left(z _ {t}\\right) \\| _ {2} ^ {2} \\right]. \\tag {12}", + "image_path": "b4e604054a96c3df518ec9aedfda9279d8cb4e8bc499b05f8b4c15c752e45977.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{true}(z_t)" + }, + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "text", + "content": " is the added noise obtained by Eq. 9, and " + }, + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t)" + }, + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "text", + "content": " is the predicted noise, conditioned on the polariza" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10895" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 128, + 137 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 128, + 137 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 128, + 137 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 128, + 137 + ], + "type": "image", + "image_path": "7f1d6246840b4417a8613b3dadab3f17d72504d483da8bedd949d838176c584f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 132, + 70, + 212, + 137 + ], + "blocks": [ + { + "bbox": [ + 132, + 70, + 212, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 70, + 212, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 70, + 212, + 137 + ], + "type": "image", + "image_path": "e5b7ec04197727068b0c5a2c107194a0d261b1e1315855dbb59204ae1692cc02.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 215, + 70, + 295, + 137 + ], + "blocks": [ + { + "bbox": [ + 215, + 70, + 295, + 137 + ], + "lines": [ + { + "bbox": [ + 215, + 70, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 215, + 70, + 295, + 137 + ], + "type": "image", + "image_path": "a8f75f8c97a880694f46513fb0e643b1cce199f11e37a28fd297731bc27ddc24.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 379, + 137 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 379, + 137 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 379, + 137 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 379, + 137 + ], + "type": "image", + "image_path": "f9bc029067c7aee81c3bce02c4710497f445c49c00df256b3c468e3cb88bde4d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 382, + 70, + 462, + 137 + ], + "blocks": [ + { + "bbox": [ + 382, + 70, + 462, + 137 + ], + "lines": [ + { + "bbox": [ + 382, + 70, + 462, + 137 + ], + "spans": [ + { + "bbox": [ + 382, + 70, + 462, + 137 + ], + "type": "image", + "image_path": "0f8b43ca56488c0242d67801492d80300353f6c8485d42c3bc67e72595de61cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 465, + 70, + 545, + 137 + ], + "blocks": [ + { + "bbox": [ + 465, + 70, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 465, + 70, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 465, + 70, + 545, + 137 + ], + "type": "image", + "image_path": "577ef286cb47296cdc5f62e8f6a17cd37c06941375e9a7929de354a13195c1df.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 49, + 140, + 128, + 208 + ], + "blocks": [ + { + "bbox": [ + 49, + 140, + 128, + 208 + ], + "lines": [ + { + "bbox": [ + 49, + 140, + 128, + 208 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 128, + 208 + ], + "type": "image", + "image_path": "2bcdaea9895d86fbae76f14feb4b6d37a098aeda4cd8082140675adaebc1309a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 132, + 140, + 212, + 208 + ], + "blocks": [ + { + "bbox": [ + 132, + 140, + 212, + 208 + ], + "lines": [ + { + "bbox": [ + 132, + 140, + 212, + 208 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 212, + 208 + ], + "type": "image", + "image_path": "d1c381867913b6bb267ec4c9dce476adffac5c81fccb6408ca54dd863c508b7e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 215, + 140, + 295, + 208 + ], + "blocks": [ + { + "bbox": [ + 215, + 140, + 295, + 208 + ], + "lines": [ + { + "bbox": [ + 215, + 140, + 295, + 208 + ], + "spans": [ + { + "bbox": [ + 215, + 140, + 295, + 208 + ], + "type": "image", + "image_path": "c778d1bce0bc6fef0246deb4852dc3859dcdf5831771e7a21620f1558f8653c6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 298, + 140, + 379, + 208 + ], + "blocks": [ + { + "bbox": [ + 298, + 140, + 379, + 208 + ], + "lines": [ + { + "bbox": [ + 298, + 140, + 379, + 208 + ], + "spans": [ + { + "bbox": [ + 298, + 140, + 379, + 208 + ], + "type": "image", + "image_path": "9ea181c82777c9adcf39c493d5929db1e9215397fe58645ced72e5de4cdc84ab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 382, + 140, + 462, + 208 + ], + "blocks": [ + { + "bbox": [ + 382, + 140, + 462, + 208 + ], + "lines": [ + { + "bbox": [ + 382, + 140, + 462, + 208 + ], + "spans": [ + { + "bbox": [ + 382, + 140, + 462, + 208 + ], + "type": "image", + "image_path": "0497e34c1fd4e89d7e2a9b5a6a659190cf0747655d3d73c68a82629dfdd4f690.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 465, + 140, + 545, + 208 + ], + "blocks": [ + { + "bbox": [ + 465, + 140, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 465, + 140, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 465, + 140, + 545, + 208 + ], + "type": "image", + "image_path": "6fb04191e0953ba6859f3a9497edde818e7410bf7cb608c800a2920c229df5f5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 49, + 210, + 128, + 277 + ], + "blocks": [ + { + "bbox": [ + 49, + 210, + 128, + 277 + ], + "lines": [ + { + "bbox": [ + 49, + 210, + 128, + 277 + ], + "spans": [ + { + "bbox": [ + 49, + 210, + 128, + 277 + ], + "type": "image", + "image_path": "31c184f1606ca7eb2defae89ce820be5a7684e230d7fd2185a85557f814df35c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 132, + 210, + 212, + 277 + ], + "blocks": [ + { + "bbox": [ + 132, + 210, + 212, + 277 + ], + "lines": [ + { + "bbox": [ + 132, + 210, + 212, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 210, + 212, + 277 + ], + "type": "image", + "image_path": "b85ea1bef5394f460395ea3cfc8be2042bd1385c1c8b068fdfa4c52c67ff916a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 215, + 210, + 295, + 277 + ], + "blocks": [ + { + "bbox": [ + 215, + 210, + 295, + 277 + ], + "lines": [ + { + "bbox": [ + 215, + 210, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 215, + 210, + 295, + 277 + ], + "type": "image", + "image_path": "26e573a11f10e382b1c5442278c7a747ea9a5861febdd2753500d6db2dbc36d3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 298, + 210, + 379, + 277 + ], + "blocks": [ + { + "bbox": [ + 298, + 210, + 379, + 277 + ], + "lines": [ + { + "bbox": [ + 298, + 210, + 379, + 277 + ], + "spans": [ + { + "bbox": [ + 298, + 210, + 379, + 277 + ], + "type": "image", + "image_path": "14dc69fbec5c9db6ad2fee69f3c7db6e8a85a29df9b9cad330971156575e569f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 382, + 210, + 462, + 277 + ], + "blocks": [ + { + "bbox": [ + 382, + 210, + 462, + 277 + ], + "lines": [ + { + "bbox": [ + 382, + 210, + 462, + 277 + ], + "spans": [ + { + "bbox": [ + 382, + 210, + 462, + 277 + ], + "type": "image", + "image_path": "5bff9ef0535e591fc3fbd9686b58a9ffbb85b03d477271ff0de8821a9bde7d97.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 465, + 210, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 465, + 210, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 465, + 210, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 465, + 210, + 545, + 277 + ], + "type": "image", + "image_path": "fc8c45ec2fb7034965edaeb020639806da030c85d3ebb2a76574270bc3114120.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 49, + 279, + 128, + 347 + ], + "blocks": [ + { + "bbox": [ + 49, + 279, + 128, + 347 + ], + "lines": [ + { + "bbox": [ + 49, + 279, + 128, + 347 + ], + "spans": [ + { + "bbox": [ + 49, + 279, + 128, + 347 + ], + "type": "image", + "image_path": "80e68386a5e357c4c12f457aaa5f773c74ed9e0aaa7162cbac2541ce775ce3ca.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 350, + 96, + 360 + ], + "lines": [ + { + "bbox": [ + 75, + 350, + 96, + 360 + ], + "spans": [ + { + "bbox": [ + 75, + 350, + 96, + 360 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 132, + 279, + 212, + 347 + ], + "blocks": [ + { + "bbox": [ + 132, + 279, + 212, + 347 + ], + "lines": [ + { + "bbox": [ + 132, + 279, + 212, + 347 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 212, + 347 + ], + "type": "image", + "image_path": "3cf543eb9e9c5288c4c081825a6d5385a4d8c84b9f9df9b448e52242536985c4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 349, + 199, + 360 + ], + "lines": [ + { + "bbox": [ + 144, + 349, + 199, + 360 + ], + "spans": [ + { + "bbox": [ + 144, + 349, + 199, + 360 + ], + "type": "text", + "content": "Kim et al. [15]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 215, + 279, + 295, + 347 + ], + "blocks": [ + { + "bbox": [ + 215, + 279, + 295, + 347 + ], + "lines": [ + { + "bbox": [ + 215, + 279, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 215, + 279, + 295, + 347 + ], + "type": "image", + "image_path": "f7c677f301c14db4d59f3307a9dbac00631c9891658789df36351c371f0ca390.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 349, + 281, + 360 + ], + "lines": [ + { + "bbox": [ + 231, + 349, + 281, + 360 + ], + "spans": [ + { + "bbox": [ + 231, + 349, + 281, + 360 + ], + "type": "text", + "content": "DSRNet [11]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 298, + 279, + 379, + 347 + ], + "blocks": [ + { + "bbox": [ + 298, + 279, + 379, + 347 + ], + "lines": [ + { + "bbox": [ + 298, + 279, + 379, + 347 + ], + "spans": [ + { + "bbox": [ + 298, + 279, + 379, + 347 + ], + "type": "image", + "image_path": "c4c5da7b26707524aa0121620f676b2fff8c99f894307c03f3d59bbd037b8e8a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 349, + 364, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 349, + 364, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 364, + 361 + ], + "type": "text", + "content": "RDRNet [46]" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 382, + 279, + 462, + 347 + ], + "blocks": [ + { + "bbox": [ + 382, + 279, + 462, + 347 + ], + "lines": [ + { + "bbox": [ + 382, + 279, + 462, + 347 + ], + "spans": [ + { + "bbox": [ + 382, + 279, + 462, + 347 + ], + "type": "image", + "image_path": "cb932e0324bd6638b77ad9fe20bcdb0a2a2e7498bb837ed4038e25ca5efb80d0.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 350, + 432, + 359 + ], + "lines": [ + { + "bbox": [ + 412, + 350, + 432, + 359 + ], + "spans": [ + { + "bbox": [ + 412, + 350, + 432, + 359 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 188, + 361, + 403, + 372 + ], + "lines": [ + { + "bbox": [ + 188, + 361, + 403, + 372 + ], + "spans": [ + { + "bbox": [ + 188, + 361, + 403, + 372 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons on the PolaRGB dataset." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 465, + 279, + 545, + 347 + ], + "blocks": [ + { + "bbox": [ + 465, + 279, + 545, + 347 + ], + "lines": [ + { + "bbox": [ + 465, + 279, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 465, + 279, + 545, + 347 + ], + "type": "image", + "image_path": "a885f30b7d3401b4b44d869e8fd419942a596ee2ea10aadf663d0f38edd9d59f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 349, + 531, + 359 + ], + "lines": [ + { + "bbox": [ + 479, + 349, + 531, + 359 + ], + "spans": [ + { + "bbox": [ + 479, + 349, + 531, + 359 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 47, + 376, + 217, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 217, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 217, + 388 + ], + "type": "text", + "content": "tion measurements and RGB input " + }, + { + "bbox": [ + 47, + 376, + 217, + 388 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{cond}}" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 47, + 393, + 287, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 287, + 417 + ], + "type": "text", + "content": "Total Loss. The above losses are weighted summed to serve as the supervision for the first stage as" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 63, + 418, + 287, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 418, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 63, + 418, + 287, + 432 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t a g e 1}} = \\gamma_ {1} \\mathcal {L} _ {1} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {V G G}} + \\gamma_ {3} \\mathcal {L} _ {\\mathrm {T V}} + \\gamma_ {4} \\mathcal {L} _ {\\text {p h a s e}}. \\tag {13}", + "image_path": "fbd907c6753f4ec5bd2ac3689d1f6842dda04386ff1e398d3b48d2de09dd3f3f.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 433, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 433, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 433, + 287, + 456 + ], + "type": "text", + "content": "In the second stage, we use a combined loss function consisting of the diffusion loss and the reconstruction loss as" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 108, + 460, + 287, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 460, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 108, + 460, + 287, + 473 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t a g e 2}} = \\gamma_ {5} \\mathcal {L} _ {\\text {d i f f}} + \\gamma_ {6} \\mathcal {L} _ {\\text {r e c o n}}. \\tag {14}", + "image_path": "53b7f64371a4bde8d1df99d509c74ad8a6c190e473ffca880a964b56057a0724.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 47, + 474, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 287, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 474, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 47, + 474, + 287, + 510 + ], + "type": "text", + "content": " has the same format as " + }, + { + "bbox": [ + 47, + 474, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{stage1}}" + }, + { + "bbox": [ + 47, + 474, + 287, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 474, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\gamma (\\cdot)" + }, + { + "bbox": [ + 47, + 474, + 287, + 510 + ], + "type": "text", + "content": " are the coefficients. This combined loss helps refine the reflection-free image while ensuring it aligns with the ground truth." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 47, + 514, + 128, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 128, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 128, + 527 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 47, + 530, + 180, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 180, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 180, + 542 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 46, + 545, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 664 + ], + "type": "text", + "content": "We train and evaluate PolarFree on the PolaRGB dataset. The whole dataset consists of 67 scenes, with a total of 6,500 paired images. For each scene, we keep the background (transmission) and camera fixed, and varying the glass position to capture images with reflections. The 67 scenes are then randomly split into 56 training scenes and 11 testing scenes, containing 6,312 and 188 paired images, respectively. This division ensures no data leakage between the training and testing sets, with each set containing only a subset of all categories." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "We also test PolarFree in real-world scenes like museums and galleries, where ground truth is unavailable. We implement PolarFree using PyTorch on a single NVIDIA RTX 4090 GPU. Training is conducted with a batch size of" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 376, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 545, + 399 + ], + "type": "text", + "content": "2 and an AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 306, + 376, + 545, + 399 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 306, + 376, + 545, + 399 + ], + "type": "text", + "content": ", 30k iterations on the PolaRGB dataset for both stages." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 401, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 509 + ], + "type": "text", + "content": "We compare our method with recent advanced reflection removal methods, including Lei et al. [18], IBCLN [21], DSRNet [11], YTMT [10], and RDRNet [46]. For a fair comparison, we modify the input settings to align with ours and use only the transmission layer for supervision. We have re-trained baseline methods on the PolaRGB dataset. The evaluation is conducted using both objective metrics (PSNR, SSIM), a perceptual metric (LPIPS), and a language-based non-reference metric (Q-Align [40])." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 306, + 514, + 362, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 362, + 524 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 362, + 524 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 526, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 681 + ], + "type": "text", + "content": "Results on PolaRGB. We present the quantitative results in Tab. 2. It can be seen that, our method outperforms other methods across multiple metrics, demonstrating the effectiveness of PolarFree for high-quality reflection removal. Notably, for fairness, we modified the input layer of all baselines to accept polarization information. Visual results in Fig. 7 show that PolarFree provides cleaner reflection removal with sharper edges and better color preservation. In contrast, previous methods often suffer from color distortions or imperfect reflection removal, especially in areas with complex reflections. Our method maintains high fidelity to the ground truth, especially in challenging regions with low light and subtle reflections." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Real-captured without Ground Truth. To demonstrate the generalization ability to unseen and more complex re" + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10896" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 106, + 170 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 106, + 170 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 106, + 170 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 106, + 170 + ], + "type": "image", + "image_path": "4dbf17d1e91ac2a2b36077764aefec87ba6f88160cafc98eeb7dc53b9d46e541.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 171, + 86, + 180 + ], + "lines": [ + { + "bbox": [ + 67, + 171, + 86, + 180 + ], + "spans": [ + { + "bbox": [ + 67, + 171, + 86, + 180 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 70, + 167, + 170 + ], + "blocks": [ + { + "bbox": [ + 108, + 70, + 167, + 170 + ], + "lines": [ + { + "bbox": [ + 108, + 70, + 167, + 170 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 167, + 170 + ], + "type": "image", + "image_path": "363fed455e966944353ab5237450f1889461997aeb0c0c154cc8bb2d28fa4ca7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 171, + 156, + 181 + ], + "lines": [ + { + "bbox": [ + 113, + 171, + 156, + 181 + ], + "spans": [ + { + "bbox": [ + 113, + 171, + 156, + 181 + ], + "type": "text", + "content": "DSRNet [46]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 227, + 170 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 227, + 170 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 227, + 170 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 227, + 170 + ], + "type": "image", + "image_path": "f62b3e1c43008901a0003b0d099f0ea5bdc557a89c50301bf530a69d5f18a001.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 171, + 220, + 181 + ], + "lines": [ + { + "bbox": [ + 171, + 171, + 220, + 181 + ], + "spans": [ + { + "bbox": [ + 171, + 171, + 220, + 181 + ], + "type": "text", + "content": "Kim et al. [15]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 182, + 286, + 204 + ], + "lines": [ + { + "bbox": [ + 47, + 182, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 286, + 204 + ], + "type": "text", + "content": "Figure 8. Visual comparison of reflection removal results in real-world museum scenes with glass display cases." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 228, + 70, + 287, + 170 + ], + "blocks": [ + { + "bbox": [ + 228, + 70, + 287, + 170 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 287, + 170 + ], + "type": "image", + "image_path": "c1a8b063639f65fea0303ee16ee9042c207ad2528618cf5a50f3572ec9f1a440.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 171, + 265, + 180 + ], + "lines": [ + { + "bbox": [ + 249, + 171, + 265, + 180 + ], + "spans": [ + { + "bbox": [ + 249, + 171, + 265, + 180 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 50, + 266, + 285, + 365 + ], + "blocks": [ + { + "bbox": [ + 47, + 207, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 207, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 287, + 262 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on the PolaRGB dataset, evaluated using objective metrics (PSNR, SSIM), perceptual metric (LPIPS), and the Language-based non-reference metric (Q-align [40]). Best results are in bold, second-best results are underlined." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 266, + 285, + 365 + ], + "lines": [ + { + "bbox": [ + 50, + 266, + 285, + 365 + ], + "spans": [ + { + "bbox": [ + 50, + 266, + 285, + 365 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓Q-align↑
Lei et al. [18]18.730.79620.38043.2109
Kim et al. [15]20.670.83990.27143.7148
IBCLN [21]19.730.81730.24883.0938
YTMT [10]16.860.75440.44893.0938
DSRNet [11]16.840.79130.28283.6992
RDRNet [46]15.880.69640.52502.9531
Ours22.440.86810.13253.8867
", + "image_path": "fd4f9d18e16cb1553d6d6eef6802bb92596146e2d83a18b81567c90bfe044179.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": "flection, We further evaluate our model on real-captured images at museum with complex reflections from glass enclosures. Here, we only provide qualitative results to visually assess the model's capability. As shown in Fig. 8, our approach effectively reduces reflections while preserving fine details, despite the complex lighting and material variations often present in museum environments. This demonstrates the practical robustness of our method and its potential for real-world applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 479, + 141, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 141, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 141, + 491 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 494, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 529 + ], + "type": "text", + "content": "We perform ablation studies to evaluate the effectiveness of key components in our method. More experimental results and analyses can be found in the supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 530, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 649 + ], + "type": "text", + "content": "Polarization Information. To assess the importance of polarization cues, we conduct an experiment where the polarization information (AOLP, DOLP, and polarization images) is removed, and the model is trained solely using RGB data. As shown in Table 3 and Fig. 9, the absence of polarization significantly impairs the model's ability to differentiate between reflection and transmission layers, leading to noticeably poorer reflection removal performance. These results demonstrate the critical role of polarization in accurate reflection removal." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Diffusion Prior. We also evaluate the effectiveness of diffusion prior by removing the conditional diffusion model. While without diffusion prior achieves reasonable results, as indicated by the comparison metrics in Table 3, it lacks fine-grained control and high-quality output offered by the" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 365, + 168 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 365, + 168 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 365, + 168 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 365, + 168 + ], + "type": "image", + "image_path": "2c1b7b3fc411e10044bd1f1e525e275ea06cbeff5af14633e0a6490259dcb056.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 170, + 358, + 178 + ], + "lines": [ + { + "bbox": [ + 316, + 170, + 358, + 178 + ], + "spans": [ + { + "bbox": [ + 316, + 170, + 358, + 178 + ], + "type": "text", + "content": "w/o diffusion" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 180, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 245 + ], + "type": "text", + "content": "Figure 9. Highly challenging reflection removal scenes with complex reflections and highlights. Top: Polarization images from different angles provide complementary information for effective reflection removal, note the third " + }, + { + "bbox": [ + 304, + 180, + 545, + 245 + ], + "type": "inline_equation", + "content": "(90^{\\circ})" + }, + { + "bbox": [ + 304, + 180, + 545, + 245 + ], + "type": "text", + "content": " image. Bottom: The diffusion module effectively handles highlights, while polarization better restores the color and details in such challenging scenarios." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 368, + 70, + 425, + 168 + ], + "blocks": [ + { + "bbox": [ + 368, + 70, + 425, + 168 + ], + "lines": [ + { + "bbox": [ + 368, + 70, + 425, + 168 + ], + "spans": [ + { + "bbox": [ + 368, + 70, + 425, + 168 + ], + "type": "image", + "image_path": "8904b546ba9899d560b2ee37c9e9e9f73080ad336a43c01adb31e23fbda4afad.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 170, + 421, + 178 + ], + "lines": [ + { + "bbox": [ + 371, + 170, + 421, + 178 + ], + "spans": [ + { + "bbox": [ + 371, + 170, + 421, + 178 + ], + "type": "text", + "content": "w/o polarization" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 485, + 168 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 485, + 168 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 485, + 168 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 485, + 168 + ], + "type": "image", + "image_path": "40e5e2dfce30f99dd30c516287cd51007e097a17f1f610b5cf50972e0f4bc386.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 170, + 465, + 177 + ], + "lines": [ + { + "bbox": [ + 449, + 170, + 465, + 177 + ], + "spans": [ + { + "bbox": [ + 449, + 170, + 465, + 177 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 487, + 70, + 545, + 168 + ], + "blocks": [ + { + "bbox": [ + 487, + 70, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 487, + 70, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 487, + 70, + 545, + 168 + ], + "type": "image", + "image_path": "d73bf48359244d1c9e4befe2ba5e1fe7a0381cf3cc40df20f3220e78116213e5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 170, + 537, + 177 + ], + "lines": [ + { + "bbox": [ + 495, + 170, + 537, + 177 + ], + "spans": [ + { + "bbox": [ + 495, + 170, + 537, + 177 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 314, + 255, + 536, + 316 + ], + "blocks": [ + { + "bbox": [ + 335, + 246, + 515, + 255 + ], + "lines": [ + { + "bbox": [ + 335, + 246, + 515, + 255 + ], + "spans": [ + { + "bbox": [ + 335, + 246, + 515, + 255 + ], + "type": "text", + "content": "Table 3. Ablation study on PolarFree framework." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 255, + 536, + 316 + ], + "lines": [ + { + "bbox": [ + 314, + 255, + 536, + 316 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 536, + 316 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓Q-align↑
w/o Lphase22.410.86220.14023.8524
w/o polarization21.560.86200.14833.7949
w/o diffusion prior20.210.86270.15523.7949
Ours22.440.86810.13253.8867
", + "image_path": "5e36121b69ee3724364c4917b2296a48cc8805f1a1823854b82483002292b106.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 320, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 357 + ], + "type": "text", + "content": "diffusion model, especially in complex scenes (Fig. 9). The conditional diffusion process significantly improves separation accuracy and preserves finer transmission details." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 472 + ], + "type": "text", + "content": "Phase-based Loss. We explore the impact of our phase-based loss function. We remove this loss from the training procedure and observe the model's performance without it. The results show a noticeable increase in color discrepancies and less accurate reflection removal, particularly in scenarios involving semi-reflective surfaces (Table 3). The phase-based loss function ensures the model focuses on reflection removal rather than compensating for color inconsistencies, leading to improved overall results." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 494, + 545, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 721 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 721 + ], + "type": "text", + "content": "In this work, we propose PolaRGB, a large-scale, comprehensive dataset for polarization-based reflection removal. With 6,500 well-aligned RGB-polarization image pairs, PolaRGB is " + }, + { + "bbox": [ + 304, + 494, + 545, + 721 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 304, + 494, + 545, + 721 + ], + "type": "text", + "content": " larger than existing polarization datasets and is the first to include both RGB and polarization images, uniquely captured across diverse real-world scenes and lighting conditions. Additionally, we present PolarFree, a novel reflection removal model that leverages the generative capabilities of diffusion models. This approach enables PolarFree to generate reflection-free cues, enhancing separation accuracy while preserving fine details in the transmission layer. A phase-based loss function is also introduced to further improve the model's performance. Comprehensive experimental results on realistic scenarios demonstrate the effectiveness our method. We believe these contributions set a foundation for further advancements in polarization-based reflection removal, paving the way for more sophisticated applications in complex real-world environments." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10897" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 144, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 144, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 144, + 84 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 84, + 286, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 84, + 286, + 130 + ], + "spans": [ + { + "bbox": [ + 47, + 84, + 286, + 130 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China (No.2022ZD0160201) and Shanghai Artificial Intelligence Laboratory, and the RGC Early Career Scheme (ECS) No. 24209224." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 142, + 105, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 142, + 105, + 154 + ], + "spans": [ + { + "bbox": [ + 48, + 142, + 105, + 154 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 162, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 162, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 287, + 206 + ], + "type": "text", + "content": "[1] Amit Agrawal, Ramesh Raskar, Shree K Nayar, and Yuanzhen Li. Removing photography artifacts using gradient projection and flash-exposure sampling. ACM Transactions on Graphics, 24(3):828-835, 2005. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 208, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 288, + 240 + ], + "type": "text", + "content": "[2] Max Born and Emil Wolf. Principles of optics: electromagnetic theory of propagation, interference and diffraction of light. Elsevier, 2013. 1, 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 241, + 287, + 285 + ], + "type": "text", + "content": "[3] Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring. Advances in neural information processing systems, 36, 2024. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 286, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 286, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 286, + 287, + 342 + ], + "type": "text", + "content": "[4] Hany Farid and Edward H Adelson. Separating reflections and lighting using independent components analysis. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149), volume 1, pages 262-267. IEEE, 1999. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 342, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 386 + ], + "type": "text", + "content": "[5] Yuanbiao Gou, Boyun Li, Zitao Liu, Songfan Yang, and Xi Peng. Clearer: Multi-scale neural architecture search for image restoration. Advances in neural information processing systems, 33:17129-17140, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 388, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 287, + 430 + ], + "type": "text", + "content": "[6] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted hdr video reconstruction. arXiv preprint arXiv:2406.08204, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 433, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 287, + 476 + ], + "type": "text", + "content": "[7] Byeong-Ju Han and Jae-Young Sim. Reflection removal using low-rank matrix completion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5438-5446, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 478, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 287, + 511 + ], + "type": "text", + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 512, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 512, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 53, + 512, + 287, + 566 + ], + "type": "text", + "content": "[9] Yuchen Hong, Haofeng Zhong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. L-differ: Single image reflection removal with language-based diffusion model. In Proceedings of the European conference on computer vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "text", + "content": "[10] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation*. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "text", + "content": "[11] Qiming Hu and Xiaojie Guo. Single image reflection separation via component synergy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13138-13147, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[12] Zhiyong Huang, Biao Xiong, Cao Tian, Jing Zhan, Xiang Fei, and Nazaraf Shah. Dust and reflection removal from videos captured in moving car. In 2016 IEEE 13th International Conference on e-Business Engineering (ICEBE), pages 182-187. IEEE, 2016. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": "[13] Yujin Jeon, Eunsue Choi, Youngchan Kim, Yunseong Moon, Khalid Omer, Felix Heide, and Seung-Hwan Baek. Spectral and polarization vision: Spectro-polarimetric real-world dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22098-22108, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 139, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 139, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 139, + 545, + 183 + ], + "type": "text", + "content": "[14] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "text", + "content": "[15] Soomin Kim, Yuchi Huo, and Sung-Eui Yoon. Single image reflection removal with physically-based training images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5164-5173, 2020. 1, 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 238, + 545, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 238, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 308, + 238, + 545, + 292 + ], + "type": "text", + "content": "[16] Naejin Kong, Yu-Wing Tai, and Joseph S Shin. A physically-based approach to reflection separation: from physical modeling to constrained optimization. IEEE transactions on pattern analysis and machine intelligence, 36(2):209-221, 2013. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 293, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 293, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 545, + 338 + ], + "type": "text", + "content": "[17] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14811-14820, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 392 + ], + "type": "text", + "content": "[18] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1750-1758, 2020. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 393, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 436 + ], + "type": "text", + "content": "[19] Chenyang Lei, Xudong Jiang, and Qifeng Chen. Robust reflection removal with flash-only cues in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "text", + "content": "[20] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(9):1647-1654, 2007. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "text", + "content": "[21] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E Hopcroft. Single image reflection removal through cascaded refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3565-3574, 2020. 1, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "text", + "content": "[22] Tingtian Li, Yuk-Hee Chan, and Daniel PK Lun. Improved multiple-image-based reflection removal algorithm using deep neural networks. IEEE Transactions on Image Processing, 30:68-79, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[23] Xiaobo Li, Lei Yan, Pengfei Qi, Liping Zhang, François Goudail, Tiegen Liu, Jingsheng Zhai, and Haofeng Hu. Polarimetric imaging via deep learning: A review. Remote Sensing, 15(6):1540, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "text", + "content": "[24] Yu Li and Michael S Brown. Single image layer separation using relative smoothness. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2752-2759, 2014. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[25] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolarized and polarized images. Advances in neural information processing systems, 32, 2019. 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10898" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Ajay Nandoriya, Mohamed Elgharib, Changil Kim, Mohamed Hefeeda, and Wojciech Matusik. Video reflection removal through spatio-temporal optimization. In Proceedings of the IEEE International Conference on Computer Vision, pages 2411-2419, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 171 + ], + "type": "text", + "content": "[27] Shree K Nayar, Xi-Sheng Fang, and Terrance Boult. Separation of reflection components using color and polarization. International Journal of Computer Vision, 21(3):163-186, 1997. 1, 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 227 + ], + "type": "text", + "content": "[28] Simon Niklaus, Xuaner Cecilia Zhang, Jonathan T Barron, Neal Wadhwa, Rahul Garg, Feng Liu, and Tianfan Xue. Learned dual-view reflection removal. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3713-3722, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 282 + ], + "type": "text", + "content": "[29] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization-based decorrelation of transparent layers: The inclination angle of an invisible surface. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 814-819. IEEE, 1999. 1, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "text", + "content": "[30] Yoav Y Schechner, Joseph Shamir, and Nahum Kiryati. Polarization and statistical analysis of scenes containing a semireflector. JOSA A, 17(2):276-284, 2000. 1, 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "text", + "content": "[31] YiChang Shih, Dilip Krishnan, Fredo Durand, and William T Freeman. Reflection removal using ghosting cues. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3193-3201, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 382 + ], + "type": "text", + "content": "[32] Sony Semiconductor Solutions. Polarsens: Polarization sensing technology, 2024. Accessed: 2024-11-15. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "text", + "content": "[33] David Strong and Tony Chan. Edge-preserving and scale-dependent properties of total variation regularization. Inverse problems, 19(6):S165, 2003. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 459 + ], + "type": "text", + "content": "[34] Robby T Tan and Katsushi Ikeuchi. Separating reflection components of textured surfaces using a single image. IEEE transactions on pattern analysis and machine intelligence, 27(2):178-193, 2005. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[35] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Crrn: Multi-scale guided concurrent reflection removal network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4777-4785, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "type": "text", + "content": "[36] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Corrn: Cooperative reflection removal network. IEEE transactions on pattern analysis and machine intelligence, 42(12):2969-2982, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 612 + ], + "type": "text", + "content": "[37] Mengyi Wang, Xinxin Zhang, Yongshun Gong, and Yilong Yin. Personalized single image reflection removal network through adaptive cascade refinement. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8204-8213, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 669 + ], + "type": "text", + "content": "[38] Kaixuan Wei, Jiaolong Yang, Ying Fu, David Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8178-8187, 2019. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 714 + ], + "type": "text", + "content": "[39] Patrick Wieschollek, Orazio Gallo, Jinwei Gu, and Jan Kautz. Separating reflection and transmission images in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), pages 89-104, 2018. 1, 2, 3, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 419 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 127 + ], + "type": "text", + "content": "[40] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 128, + 546, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 128, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 128, + 546, + 172 + ], + "type": "text", + "content": "[41] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T Freeman. A computational approach for obstruction-free photography. ACM Transactions on Graphics (TOG), 34(4):1-11, 2015. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 546, + 217 + ], + "type": "text", + "content": "[42] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 546, + 262 + ], + "type": "text", + "content": "[43] Jae-Seong Yun and Jae-Young Sim. Reflection removal for large-scale 3d point clouds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4597–4605, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 262, + 547, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 547, + 307 + ], + "type": "text", + "content": "[44] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4786-4794, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 546, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 546, + 362 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 546, + 362 + ], + "type": "text", + "content": "[45] Haofeng Zhong, Yuchen Hong, Shuchen Weng, Jinxiu Liang, and Boxin Shi. Language-guided image reflection separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24913-24922, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 364, + 547, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 364, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 364, + 547, + 419 + ], + "type": "text", + "content": "[46] Yurui Zhu, Xueyang Fu, Peng-Tao Jiang, Hao Zhang, Qibin Sun, Jinwei Chen, Zheng-Jun Zha, and Bo Li. Revisiting single image reflection removal in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25468-25478, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10899" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/a5f92f60-4941-4386-a4db-1f556681a113_content_list.json b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/a5f92f60-4941-4386-a4db-1f556681a113_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..df02cee8c50ad35c708c0e67e32a2f391b8191f5 --- /dev/null +++ b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/a5f92f60-4941-4386-a4db-1f556681a113_content_list.json @@ -0,0 +1,1260 @@ +[ + { + "type": "text", + "text": "PolarNeXt: Rethink Instance Segmentation with Polar Representation", + "text_level": 1, + "bbox": [ + 143, + 130, + 852, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiacheng Sun $^{1}$ , Xinghong Zhou $^{1}$ , Yiqiang Wu $^{1}$ , Bin Zhu $^{1}$ , Jiaxuan Lu $^{2}$ , Yu Qin $^{1}$ , Xiaomao Li $^{1,*}$ $^{1}$ Shanghai University, $^{2}$ Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 125, + 179, + 875, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sunjc, jkchou, yiqiangwu, zhubin01, qy21722148, lixiaomao}@shu.edu.cn", + "bbox": [ + 210, + 218, + 784, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lujiaxuan@pjlab.org.cn", + "bbox": [ + 397, + 236, + 594, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the roadblocks for instance segmentation today is heavy computational overhead and model parameters. Previous methods based on Polar Representation made the initial mark to address this challenge by formulating instance segmentation as polygon detection, but failed to align with mainstream methods in performance. In this paper, we highlight that Representation Errors, arising from the limited capacity of polygons to capture boundary details, have long been overlooked, which results in severe performance degradation. Observing that optimal starting point selection effectively alleviates this issue, we propose an Adaptive Polygonal Sample Decision strategy to dynamically capture the positional variation of representation errors across samples. Additionally, we design a Union-aligned Rasterization Module to incorporate these errors into polygonal assessment, further advancing the proposed strategy. With these components, our framework PolarNeXt achieves a remarkable performance boost of over $4.8\\%$ AP compared to other polar-based methods. PolarNeXt is markedly more lightweight and efficient than state-of-the-art instance segmentation methods, while achieving comparable segmentation accuracy. We expect this work will open up a new direction for instance segmentation in high-resolution images and resource-limited scenarios. Codes can be found at https://github.com/Sun15194/PolarNeXt.", + "bbox": [ + 89, + 318, + 483, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 727, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection and instance segmentation are fundamental tasks in computer vision, aiming at identifying specific objects within images. Object detection [26, 27, 33], also known as box detection, localizes objects using rectangular boxes, which are efficient but coarse-grained. In comparison, instance segmentation offers additional geometric information, such as shape and boundary. Mainstream methods define instance segmentation as pixel-wise classifica", + "bbox": [ + 89, + 753, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18c0e919d50ff766d9db2bba7da3d8be3142a08044d4b12e0381fce0d959959c.jpg", + "image_caption": [ + "Figure 1. Illustration of the composition of prediction errors in polar-based methods. Classification and Regression (Cls.&Reg.) Errors are produced during network optimization, which reflects the deviation from starting point classification $(s\\rightarrow \\hat{s})$ and dense distance regression $(D\\to \\hat{D})$ . The ignored Representation (Rep.) Error arises from the missing boundary details of instance contour $C$ in Polar Representation, as exemplified by the two cases below. Obviously, the center point does not necessarily coincide with the optimal point, from which the assembled polygon with minimal Rep. Error can be predicted." + ], + "image_footnote": [], + "bbox": [ + 516, + 284, + 903, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion [7, 14, 34] or iterative deformation [9, 25, 39]. Despite the finer granularity, these methods introduce heavy computational overhead and model parameters, which restricts their applicability in high-resolution images or resource-limited scenarios [11].", + "bbox": [ + 511, + 643, + 906, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To combine the efficiency of detection with the granularity of segmentation, a pioneering attempt is Polar Representation. As shown at the top of Fig. 1, Polar Representation reconstructs the instance contour into a bounding polygon, assembled by dense rays emitted uniformly from a starting point. Building on this concept, typical polar-based methods [36-38] formulate instance segmentation as starting point classification and dense distance regression. These methods can be instantiated on one-stage object detectors [26, 33], directly inheriting their advantages in inference speed and model complexity. However, Polar Representation has not become a fundamental paradigm for in", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19315", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2291bc6b548b54c582c4e8b602cc289b04fc1d80ccd807848f77bf06bc006a7.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed Adaptive Polygonal Sample Decision (APSD). The three stages of APSD are labeled as: I.Sampling Stage, II.Matching Stage, and III.Weighting Stage. The Union-aligned Rasterization Module (URM) takes the predicted polygons $\\hat{P}$ of the candidates selected by Multi-layer Center Sampling (MCS) as input, and outputs their RMask IoU with the instance contours $C$ . RMask IoU functions in both matching costs and sample weighting." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 893, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stance segmentation, due to the significant performance gap between polar-based methods and other advanced methods.", + "bbox": [ + 88, + 421, + 482, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we highlight that an additional prediction error has been ignored in Polar Representation, which results in the performance degradation of polar-based methods. As illustrated at the bottom of Fig. 1, from different starting points, the assembled polygons exhibit varying degrees of deviation with their corresponding contours. This deviation, termed Representation (Rep.) Error, inevitably occurs in Polar Representation, where the boundary details of contours are complex for polygons to reflect, especially in disconnected or concave regions. However, existing training pipelines in polar-based methods only account for the Classification and Regression (Cls.&Reg.) Errors generated during network optimization. Their label assignment and sample weighting, collectively referred to as sample decisions, rely entirely on the handcrafted center prior, which fails to capture the positional variation of representation errors (as proven in Sec. 3.2(1)). Furthermore, even though a natural idea is to dynamically select the optimal starting point with minimal representation error, a reliable metric to incorporate this error into polygonal assessment is still lacking. As the only available metric, Polar IoU [36] exhibits assessment blindness to representation errors (as verified in Sec. 3.2(2)), which is too coarse-grained to support this idea.", + "bbox": [ + 91, + 455, + 482, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these challenges, we propose PolarNeXt, a highly lightweight yet effective polar-based framework for real-time instance segmentation. The training pipeline of this framework includes an Adaptive Polygonal Sample Decision (APSD) strategy, enabling optimal starting point", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "selection during inference. APSD assigns positive labels and higher weights to the candidate samples based on their matching costs, which account for the comprehensive errors between predictions and ground truths. To eliminate the assessment blindness to representation errors, a Union-aligned Rasterization Module (URM) is designed for more accurate polygonal assessment. The paired polygon and contour are aligned within their Union Box and then converted into rasterized masks for RMask IoU measurement. Here, RMask IoU serves as a new metric to evaluate the actual polygonal overlap, forming the core element of sample decisions and loss functions. Benefitting from these components, PolarNeXt achieves a remarkable improvement of over $4.8\\%$ AP compared to other polar-based methods. To the best of our knowledge, this is the first polar-based method competitive with state-of-the-art instance segmentation methods in performance. Notably, PolarNeXt requires only half or even less of their computational cost and inference time, which redemonstrates the superiority of Polar Representation in instance segmentation. The main contributions of this work can be summarized as follows:", + "bbox": [ + 511, + 420, + 906, + 738 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conduct a thorough analysis of the implementation challenges of Polar Representation in instance segmentation, and reveal that the performance limitations of existing polar-based methods stem from their neglect of the representation error.", + "- We propose a new polar-based instance segmentation framework called PolarNeXt, including an APSD strategy to dynamically account for representation errors across samples and a URM module that eliminates the assessment blindness to these errors using RMask IoU." + ], + "bbox": [ + 513, + 750, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19316", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We fully leverage the advantages of Polar Representation in instance segmentation tasks. Without bells and whistles, PolarNeXt is markedly more lightweight and efficient than other mainstream methods, while aligning with them in terms of segmentation performance.", + "bbox": [ + 89, + 90, + 485, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 186, + 232, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instance Segmentation. Mainstream instance segmentation methods can be classified into two categories based on their representation, i.e., mask-based methods and contour-based methods. Mask-based methods treat instance segmentation as pixel-wise classification, in which binary masks are predicted within their Regions of Interest (RoIs). As a pioneering work, Mask R-CNN [14] introduces a detect-then-segment paradigm, built on the two-stage detector [27], with box proposals serving as the RoIs for pixel-by-pixel segmentation. Nowadays, state-of-the-art benchmark results are still held by the mask-based two-stage methods [2, 3, 15] constructed on strong detection baselines. Additionally, to reduce reliance on bounding boxes, one-stage methods [32, 34, 35] treat the entire image as the RoI, predicting masks by the assembly of encoded mask vectors and mask kernels from dynamic convolutions. Other mask-based methods focus on the high-quality boundary [17, 20], or the real-time inference [1, 7]. Contour-based methods take a different paradigm, formulating instance segmentation as iterative deformation based on the active contour mechanism of Snake [16]. Deep Snake [25] establishes the first multi-stage deformation network, which progressively refines the pre-predicted initial contours using their vertex features. Based on this baseline, further optimization strategies have been explored, such as Segment-wise Matching Scheme [24], Dynamic matching Loss [39], and Multi-scale Contour Refinement [9]. However, the high computational cost introduced by pixel-wise classification and iterative deformation limits the applicability and prevalence of these methods in high-resolution images or resource-limited scenarios [11]. To solve this problem, we bring back Polar Representation and redemonstrate its superiority in instance segmentation. Building on Polar Representation, our proposed PolarNeXt is markedly more lightweight and efficient than the aforementioned methods, while remaining competitive in segmentation performance.", + "bbox": [ + 89, + 215, + 485, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Polar Representation. Polar Representation initially appears in some interdisciplinary fields, such as automatic building segmentation [6] and medical cell detection [30]. Subsequently, ESE-Seg [38] migrates Polar Representation into instance segmentation tasks, introducing Chebyshev polynomial fitting to shorten the distance vectors and resist noise. PolarMask [36] and its variant [37] make the initial mark of polar-based methods on instance segmentation benchmarks. They maximize the advantage of the", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "center prior by modified Center Sampling and Polar Centerness, along with a proposed Polar IoU metric for polygonal assessment. However, the representation error introduced by Polar Representation has been ignored in these polar-based methods, which results in severe performance degradation. To address this, we propose an APSD strategy to dynamically balance representation errors across samples, in conjunction with a URM module that incorporates these errors into polygonal assessment. Together, these components bring a notable performance boost to PolarNeXt.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 511, + 260, + 612, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, PolarMask [36], a representative polar-based method, is used as a case study for Polar Representation. The design principle and training pipeline of PolarMask are reviewed in Sec. 3.1. Under this pipeline, the effectiveness of its sample decisions and polygonal assessment is investigated in Sec. 3.2. Finally, a novel polar-based framework called PolarNeXt is detailed in Sec. 3.3 to advance Polar Representation in instance segmentation tasks.", + "bbox": [ + 511, + 287, + 908, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 421, + 643, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PolarMask defines instance segmentation as starting point classification and dense distance regression. As illustrated in Fig. 1, for a given instance contour $C$ , bounding polygon $P = \\psi(s, D)$ can be constructed through Polar Representation $\\psi$ , using a distance set $D = \\{d_i | i = 1, \\dots, m\\}$ from a starting point $s$ . The starting point $s$ and distance set $D$ serve as supervision signals for the network, which then classifies a starting point $\\hat{s}$ and regresses an equal-size set $\\hat{D} = \\{\\hat{d}_i | i = 1, \\dots, m\\}$ to construct predicted polygon $\\hat{P} = \\psi(\\hat{s}, \\hat{D})$ . In the training pipeline of PolarMask, the center prior fully dictates label assignment and sample weighting, given that samples closer to the center tend to have richer receptive field and lower regression difficulty [19]. First, Center Sampling is applied on a single FPN [22] layer, directly assigning positive labels to the samples around instance centers. Second, Polar Centerness, which quantifies the proximity to the center, weights these positive samples to modulate their contribution to loss functions. Furthermore, due to the lack of polygon-specific IoUs, Polar IoU is proposed as a metric for polygonal assessment. It compares the consistency between target distance set $D$ and predicted distance set $\\hat{D}$ to approximate the overlap between $P$ and $\\hat{P}$ . Additionally, Polar IoU is transformed into Polar IoU Loss to supervise distance regression. More details are provided in supplementary materials.", + "bbox": [ + 511, + 446, + 908, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Investigation", + "text_level": 1, + "bbox": [ + 511, + 845, + 650, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we identify that two challenges lie in implementing Polar Representation for instance segmentation:", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19317", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/80afa08be13c791a94722d44f607be8f6a30e6f2ace8684d0fef8a67ed6c6f03.jpg", + "image_caption": [ + "Figure 3. Illustration of the divergence between Polarness and Centerness. On the left, heatmaps of these two attributes are visualized to reveal the distributional divergence, with positive samples (9 white points) selected by Center Sampling marked. Then, attribute values of these samples are plotted to verify the quantitative divergence on the right, where samples are sorted by ascending Polarness." + ], + "image_footnote": [], + "bbox": [ + 101, + 87, + 473, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(1) Diminished Effectiveness of Center Prior. The center prior plays a diminished role in sample decisions, which fails to capture the positional variation of representation errors. To validate this, a comparative experiment is conducted in Fig. 3, illustrating the divergence between the center prior and representation error. For a clear comparison, we define Polarness $(Plr)$ as an inverse measure of Representation Error $(RE)$ , where $Plr = 1 - RE$ . Likewise, normalized Polar Centerness is introduced to quantify the influence of the center prior. Experimental results show that Polarness presents no observable correlation with the distribution pattern of Centerness, while Centerness does not increase monotonically on these samples sorted by ascending Polarness. Accordingly, the samples selected by center sampling do not necessarily correspond to lower representation errors, and Centerness also fails to accurately modulate the weights of high-quality samples. Based on these observations, we conclude that the sample decisions entirely reliant on the center prior are less applicable in polar-based methods, as some suboptimal samples in center regions receive excessive attention.", + "bbox": [ + 91, + 380, + 482, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(2) Assessment Blindness of Polar IoU. As the only available metric, Polar IoU exhibits assessment blindness to representation errors. This limitation arises because Polar IoU evaluates the consistency between target polygon $P$ and predicted polygon $\\hat{P}$ by comparing their distance sets, entirely independent of instance contour $C$ . To explore the impact of this blindness on polygonal assessment, a scatter plot is drawn to examine the effectiveness of Polar IoU. As shown in Fig. 4(a), the scatter points show a significant departure from the expected linear trend, with substantial noise introduced. For example, when Real IoU reaches 0.4, Polar IoU fluctuates substantially, ranging irregularly from 0.4 to 0.9. Based on these observations, we conclude that", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f13d06d9cb4a848b9b7427846efb9181be1fb12ecdfa83c61d5ee4d60d0072da.jpg", + "image_caption": [ + "Figure 4. Comparison between Polar IoU and our proposed RMask IoU. The green solid lines denote the curves fitted by scatter points, while the red dashed lines indicate the desired linear correlation trend. Real IoU is the polygon-specific IoU between the predicted polygon and its corresponding instance contour, calculated using Weiler-Atherton algorithm [10]." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 906, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Polar IoU fails to function as a reliable metric for polygonal assessment.", + "bbox": [ + 511, + 362, + 906, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. PolarNeXt", + "text_level": 1, + "bbox": [ + 513, + 404, + 632, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome these two challenges, a new framework named PolarNeXt is proposed, which is expected to become the next level of polar-based methods for instance segmentation. The training pipeline of PolarNeXt includes two tailored components for Polar Representation: Adaptive Polygonal Sample Decision (APSD) and Union-aligned Rasterization Module (URM). As illustrated in Fig. 2, APSD samples a wide range of candidates around instances on the input image. Then, those candidates with minimal matching costs are assigned positive labels and weighted by the quality scores of their predicted polygons. For polygonal assessment, URM aligns the paired polygon and contour within their Union Box, and converts them into rasterized masks for overlap measurement. The RMask IoU output by URM incorporates representation errors, serving as precise guidance in both label assignment and sample weighting. The detailed design principles of our proposed methods are outlined below.", + "bbox": [ + 511, + 428, + 906, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Union-aligned Rasterization Module. To eliminate the assessment blindness to representation errors, a straightforward idea is to bridge the predicted polygon $\\hat{P}$ with its instance contour $C$ and measure their polygonal overlap. However, traditional polygon-specific IoU algorithms, such as Weiler-Atherton [10], are impractical for parallel processing, due to uncertain vertex counts and inevitable recursive operations. Inspired by 3D renderers [12, 18, 23], our solution is to convert $\\hat{P}$ and $C$ from vertex-set format into rasterized-mask format for mask-level overlap measurement, but an additional challenge lies in their alignment at a fixed resolution, given that no pre-predicted RoIs are available in one-stage detectors.", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19318", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we propose a Union-aligned Rasterization Module (URM) which employs Union Box, the minimal bounding box enclosing all the vertices of $\\hat{P}$ and $C$ , as a reference for coordinate alignment. Concretely, assuming the vertex sets of $\\hat{P}$ and $C$ are symbolized by $V_{\\hat{P}}$ and $V_{C}$ , with a desired mask resolution for rasterization specified as $H\\times W$ . The coordinate $(v_x,v_y)$ of a vertex $v$ can be adjusted by the union set $U = V_{\\hat{P}}\\cup V_{C}$ as follows:", + "bbox": [ + 89, + 90, + 483, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left( \\begin{array}{c} v _ {x} \\\\ v _ {y} \\end{array} \\right) = \\left( \\begin{array}{c} \\frac {v _ {x} - \\underset {u \\in U} {\\min } u _ {x}}{\\underset {u \\in U} {\\max } u _ {x} - \\underset {u \\in U} {\\min } u _ {x}} \\times W \\\\ \\frac {v _ {y} - \\underset {u \\in U} {\\min } u _ {y}}{\\underset {u \\in U} {\\max } u _ {y} - \\underset {u \\in U} {\\min } u _ {y}} \\times H \\end{array} \\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 223, + 483, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the differentiable rasterizer in [18] is introduced to transform the aligned vertex sets into rasterized masks. For pixel $p(x,y)$ of a mask, where $0 \\leq x \\leq W$ and $0 \\leq y \\leq H$ , two attributes are used to model its contribution value $M[x,y]$ to the mask: (1) the binary inside-outside state $IO(V,p)$ of pixel $p$ relative to vertex set $V$ , and (2) the closest distance $CD(V,p)$ from pixel $p$ to vertex set $V$ . The contribution value $M[x,y]$ is formulated as shown below:", + "bbox": [ + 89, + 300, + 483, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM [ x, y ] = \\operatorname {s i g m o i d} \\left(\\frac {I O (V , p) \\times C D (V , p)}{\\tau}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 430, + 482, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is a hyperparameter for sharpness. Finally, two rasterized masks $M_{\\hat{P}}$ and $M_C$ with the same resolution $H\\times W$ are produced, and their mask-level IoU, called RMask IoU, can be calculated as follows:", + "bbox": [ + 89, + 469, + 483, + 530 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R M a s k} \\mathrm {I o U} = \\frac {2 \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\mathrm {M} _ {C} [ \\mathrm {i} , \\mathrm {j} ] \\cdot \\mathrm {M} _ {\\hat {P}} [ \\mathrm {i} , \\mathrm {j} ]}{\\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\mathrm {M} _ {C} [ \\mathrm {i} , \\mathrm {j} ] + \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\mathrm {M} _ {\\hat {P}} [ \\mathrm {i} , \\mathrm {j} ]}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 539, + 482, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 4(b), RMask IoU accurately reflects the actual overlap between polygons and contours. Further, we perform a simple deformation on RMask IoU to obtain a new loss function, RMask IoU Loss, for polygonal external constraints:", + "bbox": [ + 89, + 619, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R M a s k} \\quad \\mathrm {I o U} \\quad \\text {L o s s} = 1 - \\mathrm {R M a s k} \\quad \\mathrm {I o U}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 707, + 482, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive Polygonal Sample Decision. The essence of sample decisions is to focus more attention on high-quality samples in each iteration. However, as an additional error introduced in Polar Representation, the representation error is susceptible to multiple factors, such as contour connectivity and convexity, which makes it challenging to be captured using the fixed priors. As shown in Fig. 2, our solution is to construct an optimal matching problem, in which positive labels and higher weights are assigned to those samples with lower matching costs. Specifically, we propose an Adaptive Polygonal Sample Decision (APSD) strategy to", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dynamically balance and minimize various types of errors across samples during training. For clarity, we divide this strategy into three stages:", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Sampling Stage. In this stage, a group of candidates is sampled from the multi-scale grids of FPN layers. Here, Center Sampling is still adopted to maintain the strong receptive field for candidates, with a slight adjustment. To expand the range of candidate selection, we put forward Multi-layer Center Sampling (MCS), which applies Center Sampling across all FPN layers, rather than sampling with an increasing radius on a single layer (as proven ineffective in Sec. 4.3).", + "(2) Matching Stage. The purpose of this stage is to select positive or negative labels for candidates based on their matching costs. The Matching cost is defined as a weighted summation of classification loss $L_{cls}$ , regression loss $L_{reg}$ , and polygon loss $L_{poly}$ , which together evaluate the optimized cost from the prediction to ground truth. In this work, we employ Focal Loss [28], Polar IoU Loss [36], and the proposed RMask IoU Loss as classification loss, regression loss, and polygon loss, respectively. For a given instance, the matrix of matching costs $Cost = \\{c_i | i = 1, \\dots, n\\}$ for $n$ candidates can be formulated as follows:" + ], + "bbox": [ + 511, + 138, + 906, + 445 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC o s t = \\lambda_ {c l s} L _ {c l s} + \\lambda_ {r e g} L _ {r e g} + \\lambda_ {p o l y} L _ {p o l y}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 455, + 903, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{cls}$ , $\\lambda_{reg}$ and $\\lambda_{poly}$ stand for the loss coefficients. Afterward, a fixed number $k$ of positive labels are assigned to the candidates with minimal costs, and the others are negative:", + "bbox": [ + 511, + 481, + 905, + 540 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {i} = \\left\\{ \\begin{array}{l l} \\text {P o s i t i v e}, & c _ {i} \\in \\operatorname {t o p} - k _ {\\min } \\left(c _ {1}, c _ {2}, \\dots , c _ {n}\\right) \\\\ \\text {N e g a t i v e}, & c _ {i} \\notin \\operatorname {t o p} - k _ {\\min } \\left(c _ {1}, c _ {2}, \\dots , c _ {n}\\right), \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 547, + 903, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X_{i}$ refers to the label assigned to the $i$ -th candidates. Notably, if a sample is matched by two instances simultaneously, the one with the largest area will be chosen.", + "bbox": [ + 511, + 589, + 905, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(3) Weighting Stage. In the final stage, each positive sample is provided with a quality score, which modulates its contribution to loss functions. Our approach here is straightforward but effective, replacing the heuristic Centerness with the RMask IoU output by URM. Compared to Centerness, RMask IoU treats each positive sample equally, especially those high-quality samples located outside the contour, whose weights are zeroed out by Centerness (e.g., samples 5 and 6 in Fig. 3).", + "bbox": [ + 511, + 637, + 905, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function. The overall loss function $L$ consists of four parts:", + "bbox": [ + 511, + 773, + 905, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {c l s} L _ {c l s} + \\lambda_ {r e g} L _ {r e g} + \\lambda_ {p o l y} L _ {p o l y} + \\lambda_ {s r e} L _ {s r e}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 814, + 903, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{sre}$ denotes the Cross-Entropy Loss in PolarMask used to supervise the Centerness head. In this case, we replace the original supervision signals with RMask IoU in the Centerness head.", + "bbox": [ + 511, + 839, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19319", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f6c0a1cd8d1ae9f0528167386fe7e2e1cd83deb9c09570a071c2d1c4e451adbd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodBackboneEpochSize\\(AP\\)\\(AP_{50}\\)\\(AP_{75}\\)FPS\\(Params\\downarrow\\)FLOPs↓
ContourDeepSnake [25]DLA-3416051230.3--24--
E2EC [39]DLA-3415051233.852.932.83529.80121
PolySnake [9]DLA-3425051234.4--18--
DANCE [24]R50-FPN3680036.858.539.01644.60274
MaskMask R-CNN [14]R50-FPN3680037.559.140.23344.40240
Cascade R-CNN [2]R50-FPN3680038.459.741.52377.331804
HTC [3]R50-FPN2080038.760.141.91780.261735
MS R-CNN [15]R50-FPN2480036.856.739.73060.74279
PointRend [17]R50-FPN3680038.459.541.52260.22266
CondInst [32]R50-FPN3680036.257.238.72934.17308
SOLO [34]R50-FPN3680036.457.938.92536.31332
SOLOv2 [35]R50-FPN3680038.459.541.23546.60227
YOLACT [1]R50-FPN5580029.048.729.84735.30237
SparseInst [7]R50-FPN14680033.253.434.75932.80201
PolarESE-Seg [38]R50-FPN30080021.648.722.438--
PolarMask [36]R50-FPN3680031.352.532.34234.74264
PolarNeXt (ours)R50-FPN3680036.159.737.34932.36186
", + "bbox": [ + 138, + 88, + 859, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5fafdc340ced1c68fbe9a840eae35cf39a8373f68bec6931b3d7f4b3e01dfed9.jpg", + "image_caption": [ + "Figure 5. Comparison of Memory Usage (MB) for inference. During inference on COCO test-dev images, the peak memory usage is recorded as the final result of each method." + ], + "image_footnote": [], + "bbox": [ + 91, + 420, + 464, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 671, + 223, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiments and Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 696, + 444, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment Settings. The experiments are conducted on MS COCO [21] benchmark and Cityscapes [8] dataset. MS COCO contains 80 classes with $115\\mathrm{k}$ training, $5\\mathrm{k}$ validation, and $20\\mathrm{k}$ testing images. Cityscapes has 2795 training images from 8 classes and another 500 images with high-quality annotations for validation. In terms of evaluation metrics, the accuracy of segmentation is evaluated by the standard AP metric, denoted as mask AP. Furthermore, mask AP can be divided into $AP_{50} / AP_{75}$ based on different IoU thresholds, as well as $AP_{S} / AP_{M} / AP_{L}$ for small/medium/large objects. All inference experiments are conducted on a single NVIDIA 4090D GPU, where FPS,", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/43696f6fb9d53d93fb595fc5b1e1a32c49dbc3e4818653339076a585834e25b6.jpg", + "table_caption": [ + "Table 1. Experimental results on COCO test-dev. \"Size\" refers to the length of the shortest edge of input images. ${}^{ \\blacklozenge }$ means smaller values are preferred." + ], + "table_footnote": [], + "table_body": "
MethodAPAP50FPSMems↓
Mask R-CNN [14]31.558.6161968
SOLOv2 [35]27.248.5172076
CondInst [32]33.359.1152319
PolarMask [36]27.452.328575
PolarNeXt (ours)30.758.831566
", + "bbox": [ + 526, + 420, + 893, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Experimental results on Cityscapes val set. \"Mems\" stands for the memory usage (MB).", + "bbox": [ + 511, + 532, + 906, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Parms, and FLOPs are calculated separately for inference speed, model complexity, and computational overhead. Notably, in all our experiments, TensorRT or FP16 is not used for acceleration.", + "bbox": [ + 511, + 580, + 906, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. PolarNeXt is instantiated on the PolarMask network, using a backbone ResNet-50 [13] pretrained on ImageNet [29], and removed the auxiliary box branch. All models are implemented on the MMDetection toolbox [4], trained on 2 GPUs with 2 images per GPU, and optimized with SGD. Weight decay and momentum are configured to 0.0001 and 0.9, respectively. The base learning rate is set to 0.01 without any decay scheme. For a fair comparison, following common practice, data augmentation only contains random flip and scale jitter unless specified. In APSD, the radius of Center Sampling for candidate selection at each FPN layer is set to 1.5 and the number $k$ of assigned positive labels is set to 9 for each instance. Loss coefficients $\\lambda_{cls}$ , $\\lambda_{reg}$ , $\\lambda_{poly}$ , and $\\lambda_{sre}$ are all empirically set to 1.0. Following the rasterizer configuration in [18], we set the desired mask resolution to $64 \\times 64$ and the rasterization sharpness to $\\tau = 0.1$ .", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19320", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b4afa684a23bf894c1bf7bd867b7f043dd9ea03ac5f1f3a18a400eb545dce7f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
URMAPSDAP
weightlosscost
29.1
✓30.3 (+1.2)
✓✓31.0 (+1.9)
✓28.6 (-0.5)
✓✓✓30.7 (+1.6)
✓✓✓✓33.9 (+4.8)
", + "bbox": [ + 129, + 88, + 444, + 219 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/12bd417aa70d2130bcdf9633f294d522f6abe274dd4f42d12433e848dd9f8372.jpg", + "table_caption": [ + "Table 3. Ablation studies on URM and APSD. \"weight\" means whether Centerness is replaced by RMask IoU for sample weighting. \"loss\" refers to the use of RMask IoU Loss in loss functions. \"cost\" indicates the incorporation of RMask IoU Loss into matching costs." + ], + "table_footnote": [], + "table_body": "
Polar IoU LossAPAPSAPMAPL
w/o.32.814.335.148.5
w/.33.916.435.949.4
", + "bbox": [ + 120, + 311, + 452, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 436, + 228, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MS COCO. Following common practice, we evaluate PolarNeXt on the MS COCO benchmark and compare test-dev results to some state-of-the-art models in Tab. 1. In terms of computational cost, PolarNeXt is more lightweight than the other models with R50-FPN backbone, only requiring 32.8 MB Parameters and 186 GB FLOPs. Even when compared to the contour-based methods using DLA-34 backbone, it remains on par. In terms of inference speed, PolarNeXt is second only to the prevalent SpareInst, but it achieves higher accuracy (36.1% vs. 33.2% AP) with a shorter training schedule (36 vs. 146 epochs). In terms of segmentation accuracy, PolarNeXt delivers competitive results compared to other more complex methods. Particularly, without bells and whistles, PolarNeXt achieves a 4.8% mask AP improvement over PolarMask, redemonstrating the potential of Polar Representation in instance segmentation. Moreover, we visualize the memory usage of some representative methods during inference in Fig. 5. Obviously, PolarNeXt shows its consistency with the classical object detector FCOS, with memory usage significantly lower than other instance segmentation methods, only half or even less.", + "bbox": [ + 89, + 458, + 482, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cityscapes. As an extended comparative experiment in Tab. 2, Cityscapes dataset, characterized by high-resolution images, is utilized to examine the robustness and generalization ability of our model. Notably, all models are directly trained on the Cityscapes training set under 64 epochs, without COCO initialization. Since fragmented instances frequently occur in Cityscapes, the Multi-component Detection Strategy suggested in [25] is applied to PolarNeXt and", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8368d2df9e16f674199f52a6b8bc75cddaa343ac0dc37fa4cd9ed1adca5ee422.jpg", + "table_caption": [ + "Table 4. Feasibility analysis of removing Polar IoU Loss. The symbols \"w/o.\" and \"w/.\" indicate whether the Polar IoU Loss is employed." + ], + "table_footnote": [], + "table_body": "
Candidate SelectionLayerAPT.T.
Center Sampling (s.r.=1.5)single31.013h
Center Sampling (s.r.=2.0)30.6 (-0.4)14h
All Grid Pointsmulti--
Points within Boxes33.1 (+2.1)28h
Points within Contours33.0 (+2.0)25h
Center Sampling (s.r.=1.5)33.9 (+2.9)16h
", + "bbox": [ + 514, + 89, + 919, + 205 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6d7f67da247df54ce897022a65efa6c6a86ddce65cd4b23319818495c7c5ac3b.jpg", + "table_caption": [ + "Table 5. The effectiveness of some approaches for candidate selection. \"single\" denotes selecting candidates on a specific FPN layer based on the instance size, and \"multi\" refers to selecting candidates across all FPN layers. \"T.T.\" and \"s.r.\" are abbreviations for training time and sampling radius, respectively." + ], + "table_footnote": [], + "table_body": "
MethodAPbAP50APSbAPMAPLb
PolarMask16.338.316.028.413.2
PolarNeXt19.846.720.232.215.9
", + "bbox": [ + 526, + 297, + 890, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Comparison of boundary quality between PolarMask and PolarNeXt.", + "bbox": [ + 511, + 364, + 906, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PolarMask, leading to approximately a $1.5\\%$ AP improvement. Under the increased resolution of input images (from $800 \\times 1333$ to $1024 \\times 2048$ ), mask-based methods suffer from the heavy memory usage introduced by dense pixel classification, which is nearly four times greater than that of polar-based methods. Meanwhile, the speed advantage of polar-based methods becomes more apparent, requiring only half the inference time compared to mask-based methods. As expected, PolarNeXt maintains a $3.3\\%$ AP improvement over PolarMask, further demonstrating the effectiveness of our proposed method.", + "bbox": [ + 511, + 409, + 906, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 587, + 679, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablations to assess the influence of the proposed components on instance segmentation performance, and then analyze their details in the following. All experiments are conducted on MS COCO val under a 1x training schedule.", + "bbox": [ + 511, + 609, + 905, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component Impact and Correlation. The effectiveness of URM and APSD is reported in Tab. 3. Using URM alone brings a $1.9\\%$ AP improvement on the baseline, while combining it with APSD further boosts the performance by an additional $2.9\\%$ AP. It is worth noting that APSD fails to function independently of URM, the absence of which results in a decrease in performance. To be specific, without APSD, URM contributes to the sample weighting and loss function, leading to a $1.2\\%$ and $0.7\\%$ AP increase, respectively. Conversely, when APSD is applied, the matching costs have to incorporate RMask IoU Loss; otherwise, this strategy will fail or even backfire. In view of this, we conclude that the effectiveness of our adaptive strategy hinges on reliable polygonal assessment for support.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19321", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4420ab76d06333c0ef197cf47c546769c0f1bb98d8ecf44aac60ead69e5361c1.jpg", + "image_caption": [ + "Figure 6. Visualization with Rays for PolarMask and PolarNeXt." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 867, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate Selection. In Tab. 5, we compare some approaches to expand the range of candidate selection. On the one hand, Center Sampling on multiple FPN layers is more effective and efficient than the dense point selection within boxes or contours. In particular, the approach suggested in [31], which selects all grid points of FPN, results in unacceptable training time and GPU memory overflow. On the other hand, directly expanding the radius of Center Sampling on a single layer causes a $0.4\\%$ drop in AP, as the additional selected samples tend to fall outside the receptive field. Accordingly, the center prior still has a positive influence on sample decisions.", + "bbox": [ + 89, + 362, + 483, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "RMask IoU vs. Polar IoU. In Fig. 4, two scatter plots are drawn to compare Polar IoU and RMask IoU in polygonal assessment. These scatter points denote all the polygons predicted during inference on MS COCO val images. The fitted curve in Fig. 4(b) is almost aligned with the ideal curve, whereas the one in Fig. 4(a) shows significant deviation. This indicates that RMask IoU better reflects the actual polygonal quality compared to Polar IoU.", + "bbox": [ + 89, + 547, + 483, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The necessity of Polar IoU Loss. In Tab. 4, we explore the feasibility of directly removing Polar IoU Loss in our proposed training pipeline. Experimental results indicate that this operation results in a $1.1\\%$ AP degradation in performance. The impact on small objects is more notable, with $AP_{S}$ decreased by $2.1\\%$ . We analyze that there may be a slight granularity missing as well as feature misalignment during the conversion from vertex sets to rasterized masks. Moreover, as mentioned in [18], aliasing occurs in the deconvolutional process of FPN when pre-predicted RoIs are unavailable in one-stage detectors. Fortunately, the combination of Polar IoU Loss and RMask IoU Loss brings a notable boost, with the two complementing each other. We believe that Polar IoU Loss serves as a polygonal internal constraint for dense distance regression, while RMask IoU", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss provides a polygonal external constraint to incorporate the representation deviation.", + "bbox": [ + 511, + 362, + 906, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Boundary Quality Analysis", + "text_level": 1, + "bbox": [ + 511, + 400, + 759, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 6, we introduce Boundary AP [5] ( $AP^b$ for short) to evaluate the boundary quality of predicted polygons. Experimental results show that PolarNeXt achieves a significant improvement in boundary quality compared to PolarMask, boosting $AP^b$ by $3.5\\%$ . Furthermore, we present some segmentation results with differences in starting points in Fig. 6. Obviously, compared to the center-fixed starting points in PolarMask, the dynamically selected starting points in PolarNeXt are more flexible, which enhances the representation capability of bounding polygons.", + "bbox": [ + 511, + 421, + 906, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 584, + 633, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel polar-based framework, PolarNeXt, which redemonstrates the superiority of Polar Representation in instance segmentation. To account for the previously ignored representation error, tailored training strategy APSD and assessment module URM are proposed. Together, these components significantly enhance PolarNeXt, yielding notable performance improvements over other polar-based methods. By fully leveraging the advantages of Polar Representation, PolarNeXt achieves comparable segmentation accuracy while requiring only half or even less of the computational cost and inference time compared to mainstream mask-based and contour-based methods. We anticipate that PolarNeXt will set a new standard for polar-based methods in instance segmentation.", + "bbox": [ + 511, + 609, + 908, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported by the National Natural Science Foundation of China [Grant No. 62225308], and the Shanghai Technical Service Computing Center of Science and Engineering, Shanghai University.", + "bbox": [ + 511, + 830, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19322", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, pages 9157-9166, 2019. 3, 6", + "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE TPAMI, 43(5):1483-1498, 2019. 3, 6", + "[3] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In CVPR, pages 4974-4983, 2019. 3, 6", + "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6", + "[5] Bowen Cheng, Ross Girshick, Piotr Dólar, Alexander C Berg, and Alexander Kirillov. Boundary iou: Improving object-centric image segmentation evaluation. In CVPR, pages 15334-15342, 2021. 8", + "[6] Dominic Cheng, Renjie Liao, Sanja Fidler, and Raquel Urtasun. Darnet: Deep active ray network for building segmentation. In CVPR, pages 7431-7439, 2019. 3", + "[7] Tianheng Cheng, Xinggang Wang, Shaoyu Chen, Wenqiang Zhang, Qian Zhang, Chang Huang, Zhaoxiang Zhang, and Wenyu Liu. Sparse instance activation for real-time instance segmentation. In CVPR, pages 4433-4442, 2022. 1, 3, 6", + "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6", + "[9] Hao Feng, Keyi Zhou, Wengang Zhou, Yufei Yin, Jiajun Deng, Qi Sun, and Houqiang Li. Recurrent generic contour-based instance segmentation with progressive learning. IEEE TCSVT, 2024. 1, 3, 6", + "[10] Gunther Greiner and Kai Hormann. Efficient clipping of arbitrary polygons. ACM Transactions on Graphics (TOG), 17 (2):71-83, 1998. 4", + "[11] Wenchao Gu, Shuang Bai, and Lingxing Kong. A review on 2d instance segmentation based on deep neural networks. Image and Vision Computing, 120:104401, 2022. 1, 3", + "[12] Shir Gur, Tal Shaharabany, and Lior Wolf. End to end trainable active contours via differentiable rendering. arXiv preprint arXiv:1912.00367, 2019. 4", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6", + "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3, 6", + "[15] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In CVPR, pages 6409-6418, 2019. 3, 6", + "[16] Michael Kass, Andrew Witkin, and Demetri Terzopoulos. Snakes: Active contour models. IJCV, 1(4):321-331, 1988. 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In CVPR, pages 9799-9808, 2020. 3, 6", + "[18] Justin Lazarow, Weijian Xu, and Zhuowen Tu. Instance segmentation with mask-supervised polygonal boundary transformers. In CVPR, pages 4382-4391, 2022. 4, 5, 6, 8", + "[19] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. NeurIPS, 33:21002-21012, 2020. 3", + "[20] Justin Liang, Namdar Homayounfar, Wei-Chiu Ma, Yuwen Xiong, Rui Hu, and Raquel Urtasun. Polytransform: Deep polygon transformer for instance segmentation. In CVPR, pages 9131-9140, 2020. 3", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6", + "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 3", + "[23] Huan Ling, Jun Gao, Amlan Kar, Wenzheng Chen, and Sanja Fidler. Fast interactive object annotation with curve-gcn. In CVPR, pages 5257-5266, 2019. 4", + "[24] Zichen Liu, Jun Hao Liew, Xiangyu Chen, and Jiashi Feng. Dance: A deep attentive contour model for efficient instance segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 345-354, 2021. 3, 6", + "[25] Sida Peng, Wen Jiang, Huaijin Pi, Xiuli Li, Hujun Bao, and Xiaowei Zhou. Deep snake for real-time instance segmentation. In CVPR, pages 8533-8542, 2020. 1, 3, 6, 7", + "[26] Joseph Redmon. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 1", + "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2016. 1, 3", + "[28] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 5", + "[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 6", + "[30] Uwe Schmidt, Martin Weigert, Coleman Broaddus, and Gene Myers. Cell detection with star-convex polygons. In Medical Image Computing and Computer Assisted Intervention, pages 265–273, 2018. 3", + "[31] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In International Conference on Machine Learning, pages 9934-9944, 2021. 8", + "[32] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, pages 282-298, 2020. 3, 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19323", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: A simple and strong anchor-free object detector. IEEE TPAMI, 44(4):1922-1933, 2020. 1", + "[34] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, pages 649-665, 2020. 1, 3, 6", + "[35] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmentation. NeurIPS, 33:17721-17732, 2020. 3, 6", + "[36] Enze Xie, Peize Sun, Xiaoge Song, Wenhai Wang, Xuebo Liu, Ding Liang, Chunhua Shen, and Ping Luo. Polarmask: Single shot instance segmentation with polar representation. In CVPR, pages 12193-12202, 2020. 1, 2, 3, 5, 6", + "[37] Enze Xie, Wenhai Wang, Mingyu Ding, Ruimao Zhang, and Ping Luo. Polarmask++: Enhanced polar representation for single-shot instance segmentation and beyond. IEEE TPAMI, 44(9):5385-5400, 2021. 3", + "[38] Wenqiang Xu, Haiyang Wang, Fubo Qi, and Cewu Lu. Explicit shape encoding for real-time instance segmentation. In ICCV, pages 5168-5177, 2019. 1, 3, 6", + "[39] Tao Zhang, Shiqing Wei, and Shunping Ji. E2ec: An end-to-end contour-based method for high-quality high-speed instance segmentation. In CVPR, pages 4443-4452, 2022. 1, 3, 6" + ], + "bbox": [ + 91, + 92, + 480, + 429 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19324", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/a5f92f60-4941-4386-a4db-1f556681a113_model.json b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/a5f92f60-4941-4386-a4db-1f556681a113_model.json new file mode 100644 index 0000000000000000000000000000000000000000..24839cf19843631c2ebaccd1e71a0ef9860bf3f7 --- /dev/null +++ b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/a5f92f60-4941-4386-a4db-1f556681a113_model.json @@ -0,0 +1,1738 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.131, + 0.853, + 0.154 + ], + "angle": 0, + "content": "PolarNeXt: Rethink Instance Segmentation with Polar Representation" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.18, + 0.877, + 0.217 + ], + "angle": 0, + "content": "Jiacheng Sun\\(^{1}\\), Xinghong Zhou\\(^{1}\\), Yiqiang Wu\\(^{1}\\), Bin Zhu\\(^{1}\\), Jiaxuan Lu\\(^{2}\\), Yu Qin\\(^{1}\\), Xiaomao Li\\(^{1,*}\\) \n\\(^{1}\\)Shanghai University, \\(^{2}\\)Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.219, + 0.785, + 0.234 + ], + "angle": 0, + "content": "{sunjc, jkchou, yiqiangwu, zhubin01, qy21722148, lixiaomao}@shu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.237, + 0.595, + 0.251 + ], + "angle": 0, + "content": "lujiaxuan@pjlab.org.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.485, + 0.697 + ], + "angle": 0, + "content": "One of the roadblocks for instance segmentation today is heavy computational overhead and model parameters. Previous methods based on Polar Representation made the initial mark to address this challenge by formulating instance segmentation as polygon detection, but failed to align with mainstream methods in performance. In this paper, we highlight that Representation Errors, arising from the limited capacity of polygons to capture boundary details, have long been overlooked, which results in severe performance degradation. Observing that optimal starting point selection effectively alleviates this issue, we propose an Adaptive Polygonal Sample Decision strategy to dynamically capture the positional variation of representation errors across samples. Additionally, we design a Union-aligned Rasterization Module to incorporate these errors into polygonal assessment, further advancing the proposed strategy. With these components, our framework PolarNeXt achieves a remarkable performance boost of over \\(4.8\\%\\) AP compared to other polar-based methods. PolarNeXt is markedly more lightweight and efficient than state-of-the-art instance segmentation methods, while achieving comparable segmentation accuracy. We expect this work will open up a new direction for instance segmentation in high-resolution images and resource-limited scenarios. Codes can be found at https://github.com/Sun15194/PolarNeXt." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.728, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.754, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Object detection and instance segmentation are fundamental tasks in computer vision, aiming at identifying specific objects within images. Object detection [26, 27, 33], also known as box detection, localizes objects using rectangular boxes, which are efficient but coarse-grained. In comparison, instance segmentation offers additional geometric information, such as shape and boundary. Mainstream methods define instance segmentation as pixel-wise classifica" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.285, + 0.905, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.628 + ], + "angle": 0, + "content": "Figure 1. Illustration of the composition of prediction errors in polar-based methods. Classification and Regression (Cls.&Reg.) Errors are produced during network optimization, which reflects the deviation from starting point classification \\((s\\rightarrow \\hat{s})\\) and dense distance regression \\((D\\to \\hat{D})\\). The ignored Representation (Rep.) Error arises from the missing boundary details of instance contour \\(C\\) in Polar Representation, as exemplified by the two cases below. Obviously, the center point does not necessarily coincide with the optimal point, from which the assembled polygon with minimal Rep. Error can be predicted." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.719 + ], + "angle": 0, + "content": "tion [7, 14, 34] or iterative deformation [9, 25, 39]. Despite the finer granularity, these methods introduce heavy computational overhead and model parameters, which restricts their applicability in high-resolution images or resource-limited scenarios [11]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To combine the efficiency of detection with the granularity of segmentation, a pioneering attempt is Polar Representation. As shown at the top of Fig. 1, Polar Representation reconstructs the instance contour into a bounding polygon, assembled by dense rays emitted uniformly from a starting point. Building on this concept, typical polar-based methods [36-38] formulate instance segmentation as starting point classification and dense distance regression. These methods can be instantiated on one-stage object detectors [26, 33], directly inheriting their advantages in inference speed and model complexity. However, Polar Representation has not become a fundamental paradigm for in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19315" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.089, + 0.895, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.349, + 0.908, + 0.409 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed Adaptive Polygonal Sample Decision (APSD). The three stages of APSD are labeled as: I.Sampling Stage, II.Matching Stage, and III.Weighting Stage. The Union-aligned Rasterization Module (URM) takes the predicted polygons \\(\\hat{P}\\) of the candidates selected by Multi-layer Center Sampling (MCS) as input, and outputs their RMask IoU with the instance contours \\(C\\). RMask IoU functions in both matching costs and sample weighting." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.422, + 0.483, + 0.452 + ], + "angle": 0, + "content": "stance segmentation, due to the significant performance gap between polar-based methods and other advanced methods." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.457, + 0.483, + 0.82 + ], + "angle": 0, + "content": "In this work, we highlight that an additional prediction error has been ignored in Polar Representation, which results in the performance degradation of polar-based methods. As illustrated at the bottom of Fig. 1, from different starting points, the assembled polygons exhibit varying degrees of deviation with their corresponding contours. This deviation, termed Representation (Rep.) Error, inevitably occurs in Polar Representation, where the boundary details of contours are complex for polygons to reflect, especially in disconnected or concave regions. However, existing training pipelines in polar-based methods only account for the Classification and Regression (Cls.&Reg.) Errors generated during network optimization. Their label assignment and sample weighting, collectively referred to as sample decisions, rely entirely on the handcrafted center prior, which fails to capture the positional variation of representation errors (as proven in Sec. 3.2(1)). Furthermore, even though a natural idea is to dynamically select the optimal starting point with minimal representation error, a reliable metric to incorporate this error into polygonal assessment is still lacking. As the only available metric, Polar IoU [36] exhibits assessment blindness to representation errors (as verified in Sec. 3.2(2)), which is too coarse-grained to support this idea." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To overcome these challenges, we propose PolarNeXt, a highly lightweight yet effective polar-based framework for real-time instance segmentation. The training pipeline of this framework includes an Adaptive Polygonal Sample Decision (APSD) strategy, enabling optimal starting point" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.908, + 0.739 + ], + "angle": 0, + "content": "selection during inference. APSD assigns positive labels and higher weights to the candidate samples based on their matching costs, which account for the comprehensive errors between predictions and ground truths. To eliminate the assessment blindness to representation errors, a Union-aligned Rasterization Module (URM) is designed for more accurate polygonal assessment. The paired polygon and contour are aligned within their Union Box and then converted into rasterized masks for RMask IoU measurement. Here, RMask IoU serves as a new metric to evaluate the actual polygonal overlap, forming the core element of sample decisions and loss functions. Benefitting from these components, PolarNeXt achieves a remarkable improvement of over \\(4.8\\%\\) AP compared to other polar-based methods. To the best of our knowledge, this is the first polar-based method competitive with state-of-the-art instance segmentation methods in performance. Notably, PolarNeXt requires only half or even less of their computational cost and inference time, which redemonstrates the superiority of Polar Representation in instance segmentation. The main contributions of this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.75, + 0.907, + 0.825 + ], + "angle": 0, + "content": "- We conduct a thorough analysis of the implementation challenges of Polar Representation in instance segmentation, and reveal that the performance limitations of existing polar-based methods stem from their neglect of the representation error." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "- We propose a new polar-based instance segmentation framework called PolarNeXt, including an APSD strategy to dynamically account for representation errors across samples and a URM module that eliminates the assessment blindness to these errors using RMask IoU." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19316" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.169 + ], + "angle": 0, + "content": "- We fully leverage the advantages of Polar Representation in instance segmentation tasks. Without bells and whistles, PolarNeXt is markedly more lightweight and efficient than other mainstream methods, while aligning with them in terms of segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.188, + 0.233, + 0.205 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.486, + 0.761 + ], + "angle": 0, + "content": "Instance Segmentation. Mainstream instance segmentation methods can be classified into two categories based on their representation, i.e., mask-based methods and contour-based methods. Mask-based methods treat instance segmentation as pixel-wise classification, in which binary masks are predicted within their Regions of Interest (RoIs). As a pioneering work, Mask R-CNN [14] introduces a detect-then-segment paradigm, built on the two-stage detector [27], with box proposals serving as the RoIs for pixel-by-pixel segmentation. Nowadays, state-of-the-art benchmark results are still held by the mask-based two-stage methods [2, 3, 15] constructed on strong detection baselines. Additionally, to reduce reliance on bounding boxes, one-stage methods [32, 34, 35] treat the entire image as the RoI, predicting masks by the assembly of encoded mask vectors and mask kernels from dynamic convolutions. Other mask-based methods focus on the high-quality boundary [17, 20], or the real-time inference [1, 7]. Contour-based methods take a different paradigm, formulating instance segmentation as iterative deformation based on the active contour mechanism of Snake [16]. Deep Snake [25] establishes the first multi-stage deformation network, which progressively refines the pre-predicted initial contours using their vertex features. Based on this baseline, further optimization strategies have been explored, such as Segment-wise Matching Scheme [24], Dynamic matching Loss [39], and Multi-scale Contour Refinement [9]. However, the high computational cost introduced by pixel-wise classification and iterative deformation limits the applicability and prevalence of these methods in high-resolution images or resource-limited scenarios [11]. To solve this problem, we bring back Polar Representation and redemonstrate its superiority in instance segmentation. Building on Polar Representation, our proposed PolarNeXt is markedly more lightweight and efficient than the aforementioned methods, while remaining competitive in segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Polar Representation. Polar Representation initially appears in some interdisciplinary fields, such as automatic building segmentation [6] and medical cell detection [30]. Subsequently, ESE-Seg [38] migrates Polar Representation into instance segmentation tasks, introducing Chebyshev polynomial fitting to shorten the distance vectors and resist noise. PolarMask [36] and its variant [37] make the initial mark of polar-based methods on instance segmentation benchmarks. They maximize the advantage of the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.244 + ], + "angle": 0, + "content": "center prior by modified Center Sampling and Polar Centerness, along with a proposed Polar IoU metric for polygonal assessment. However, the representation error introduced by Polar Representation has been ignored in these polar-based methods, which results in severe performance degradation. To address this, we propose an APSD strategy to dynamically balance representation errors across samples, in conjunction with a URM module that incorporates these errors into polygonal assessment. Together, these components bring a notable performance boost to PolarNeXt." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.261, + 0.613, + 0.277 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.909, + 0.41 + ], + "angle": 0, + "content": "In this work, PolarMask [36], a representative polar-based method, is used as a case study for Polar Representation. The design principle and training pipeline of PolarMask are reviewed in Sec. 3.1. Under this pipeline, the effectiveness of its sample decisions and polygonal assessment is investigated in Sec. 3.2. Finally, a novel polar-based framework called PolarNeXt is detailed in Sec. 3.3 to advance Polar Representation in instance segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.422, + 0.644, + 0.439 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.909, + 0.835 + ], + "angle": 0, + "content": "PolarMask defines instance segmentation as starting point classification and dense distance regression. As illustrated in Fig. 1, for a given instance contour \\( C \\), bounding polygon \\( P = \\psi(s, D) \\) can be constructed through Polar Representation \\( \\psi \\), using a distance set \\( D = \\{d_i | i = 1, \\dots, m\\} \\) from a starting point \\( s \\). The starting point \\( s \\) and distance set \\( D \\) serve as supervision signals for the network, which then classifies a starting point \\( \\hat{s} \\) and regresses an equal-size set \\( \\hat{D} = \\{\\hat{d}_i | i = 1, \\dots, m\\} \\) to construct predicted polygon \\( \\hat{P} = \\psi(\\hat{s}, \\hat{D}) \\). In the training pipeline of PolarMask, the center prior fully dictates label assignment and sample weighting, given that samples closer to the center tend to have richer receptive field and lower regression difficulty [19]. First, Center Sampling is applied on a single FPN [22] layer, directly assigning positive labels to the samples around instance centers. Second, Polar Centerness, which quantifies the proximity to the center, weights these positive samples to modulate their contribution to loss functions. Furthermore, due to the lack of polygon-specific IoUs, Polar IoU is proposed as a metric for polygonal assessment. It compares the consistency between target distance set \\( D \\) and predicted distance set \\( \\hat{D} \\) to approximate the overlap between \\( P \\) and \\( \\hat{P} \\). Additionally, Polar IoU is transformed into Polar IoU Loss to supervise distance regression. More details are provided in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.651, + 0.864 + ], + "angle": 0, + "content": "3.2. Investigation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "In this section, we identify that two challenges lie in implementing Polar Representation for instance segmentation:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19317" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.088, + 0.475, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.485, + 0.358 + ], + "angle": 0, + "content": "Figure 3. Illustration of the divergence between Polarness and Centerness. On the left, heatmaps of these two attributes are visualized to reveal the distributional divergence, with positive samples (9 white points) selected by Center Sampling marked. Then, attribute values of these samples are plotted to verify the quantitative divergence on the right, where samples are sorted by ascending Polarness." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.381, + 0.483, + 0.697 + ], + "angle": 0, + "content": "(1) Diminished Effectiveness of Center Prior. The center prior plays a diminished role in sample decisions, which fails to capture the positional variation of representation errors. To validate this, a comparative experiment is conducted in Fig. 3, illustrating the divergence between the center prior and representation error. For a clear comparison, we define Polarness \\((Plr)\\) as an inverse measure of Representation Error \\((RE)\\), where \\(Plr = 1 - RE\\). Likewise, normalized Polar Centerness is introduced to quantify the influence of the center prior. Experimental results show that Polarness presents no observable correlation with the distribution pattern of Centerness, while Centerness does not increase monotonically on these samples sorted by ascending Polarness. Accordingly, the samples selected by center sampling do not necessarily correspond to lower representation errors, and Centerness also fails to accurately modulate the weights of high-quality samples. Based on these observations, we conclude that the sample decisions entirely reliant on the center prior are less applicable in polar-based methods, as some suboptimal samples in center regions receive excessive attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "(2) Assessment Blindness of Polar IoU. As the only available metric, Polar IoU exhibits assessment blindness to representation errors. This limitation arises because Polar IoU evaluates the consistency between target polygon \\( P \\) and predicted polygon \\( \\hat{P} \\) by comparing their distance sets, entirely independent of instance contour \\( C \\). To explore the impact of this blindness on polygonal assessment, a scatter plot is drawn to examine the effectiveness of Polar IoU. As shown in Fig. 4(a), the scatter points show a significant departure from the expected linear trend, with substantial noise introduced. For example, when Real IoU reaches 0.4, Polar IoU fluctuates substantially, ranging irregularly from 0.4 to 0.9. Based on these observations, we conclude that" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.907, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.261, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Figure 4. Comparison between Polar IoU and our proposed RMask IoU. The green solid lines denote the curves fitted by scatter points, while the red dashed lines indicate the desired linear correlation trend. Real IoU is the polygon-specific IoU between the predicted polygon and its corresponding instance contour, calculated using Weiler-Atherton algorithm [10]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.363, + 0.907, + 0.392 + ], + "angle": 0, + "content": "Polar IoU fails to function as a reliable metric for polygonal assessment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.405, + 0.633, + 0.42 + ], + "angle": 0, + "content": "3.3. PolarNeXt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.699 + ], + "angle": 0, + "content": "To overcome these two challenges, a new framework named PolarNeXt is proposed, which is expected to become the next level of polar-based methods for instance segmentation. The training pipeline of PolarNeXt includes two tailored components for Polar Representation: Adaptive Polygonal Sample Decision (APSD) and Union-aligned Rasterization Module (URM). As illustrated in Fig. 2, APSD samples a wide range of candidates around instances on the input image. Then, those candidates with minimal matching costs are assigned positive labels and weighted by the quality scores of their predicted polygons. For polygonal assessment, URM aligns the paired polygon and contour within their Union Box, and converts them into rasterized masks for overlap measurement. The RMask IoU output by URM incorporates representation errors, serving as precise guidance in both label assignment and sample weighting. The detailed design principles of our proposed methods are outlined below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Union-aligned Rasterization Module. To eliminate the assessment blindness to representation errors, a straightforward idea is to bridge the predicted polygon \\(\\hat{P}\\) with its instance contour \\(C\\) and measure their polygonal overlap. However, traditional polygon-specific IoU algorithms, such as Weiler-Atherton [10], are impractical for parallel processing, due to uncertain vertex counts and inevitable recursive operations. Inspired by 3D renderers [12, 18, 23], our solution is to convert \\(\\hat{P}\\) and \\(C\\) from vertex-set format into rasterized-mask format for mask-level overlap measurement, but an additional challenge lies in their alignment at a fixed resolution, given that no pre-predicted RoIs are available in one-stage detectors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19318" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.214 + ], + "angle": 0, + "content": "As shown in Fig. 2, we propose a Union-aligned Rasterization Module (URM) which employs Union Box, the minimal bounding box enclosing all the vertices of \\(\\hat{P}\\) and \\(C\\), as a reference for coordinate alignment. Concretely, assuming the vertex sets of \\(\\hat{P}\\) and \\(C\\) are symbolized by \\(V_{\\hat{P}}\\) and \\(V_{C}\\), with a desired mask resolution for rasterization specified as \\(H\\times W\\). The coordinate \\((v_x,v_y)\\) of a vertex \\(v\\) can be adjusted by the union set \\(U = V_{\\hat{P}}\\cup V_{C}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.224, + 0.484, + 0.292 + ], + "angle": 0, + "content": "\\[\n\\left( \\begin{array}{c} v _ {x} \\\\ v _ {y} \\end{array} \\right) = \\left( \\begin{array}{c} \\frac {v _ {x} - \\underset {u \\in U} {\\min } u _ {x}}{\\underset {u \\in U} {\\max } u _ {x} - \\underset {u \\in U} {\\min } u _ {x}} \\times W \\\\ \\frac {v _ {y} - \\underset {u \\in U} {\\min } u _ {y}}{\\underset {u \\in U} {\\max } u _ {y} - \\underset {u \\in U} {\\min } u _ {y}} \\times H \\end{array} \\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.484, + 0.422 + ], + "angle": 0, + "content": "Then, the differentiable rasterizer in [18] is introduced to transform the aligned vertex sets into rasterized masks. For pixel \\( p(x,y) \\) of a mask, where \\( 0 \\leq x \\leq W \\) and \\( 0 \\leq y \\leq H \\), two attributes are used to model its contribution value \\( M[x,y] \\) to the mask: (1) the binary inside-outside state \\( IO(V,p) \\) of pixel \\( p \\) relative to vertex set \\( V \\), and (2) the closest distance \\( CD(V,p) \\) from pixel \\( p \\) to vertex set \\( V \\). The contribution value \\( M[x,y] \\) is formulated as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.431, + 0.483, + 0.462 + ], + "angle": 0, + "content": "\\[\nM [ x, y ] = \\operatorname {s i g m o i d} \\left(\\frac {I O (V , p) \\times C D (V , p)}{\\tau}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.531 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a hyperparameter for sharpness. Finally, two rasterized masks \\(M_{\\hat{P}}\\) and \\(M_C\\) with the same resolution \\(H\\times W\\) are produced, and their mask-level IoU, called RMask IoU, can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.54, + 0.483, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R M a s k} \\mathrm {I o U} = \\frac {2 \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\mathrm {M} _ {C} [ \\mathrm {i} , \\mathrm {j} ] \\cdot \\mathrm {M} _ {\\hat {P}} [ \\mathrm {i} , \\mathrm {j} ]}{\\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\mathrm {M} _ {C} [ \\mathrm {i} , \\mathrm {j} ] + \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\mathrm {M} _ {\\hat {P}} [ \\mathrm {i} , \\mathrm {j} ]}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.484, + 0.697 + ], + "angle": 0, + "content": "As shown in Fig. 4(b), RMask IoU accurately reflects the actual overlap between polygons and contours. Further, we perform a simple deformation on RMask IoU to obtain a new loss function, RMask IoU Loss, for polygonal external constraints:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.708, + 0.483, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R M a s k} \\quad \\mathrm {I o U} \\quad \\text {L o s s} = 1 - \\mathrm {R M a s k} \\quad \\mathrm {I o U}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Adaptive Polygonal Sample Decision. The essence of sample decisions is to focus more attention on high-quality samples in each iteration. However, as an additional error introduced in Polar Representation, the representation error is susceptible to multiple factors, such as contour connectivity and convexity, which makes it challenging to be captured using the fixed priors. As shown in Fig. 2, our solution is to construct an optimal matching problem, in which positive labels and higher weights are assigned to those samples with lower matching costs. Specifically, we propose an Adaptive Polygonal Sample Decision (APSD) strategy to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "dynamically balance and minimize various types of errors across samples during training. For clarity, we divide this strategy into three stages:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.14, + 0.907, + 0.276 + ], + "angle": 0, + "content": "(1) Sampling Stage. In this stage, a group of candidates is sampled from the multi-scale grids of FPN layers. Here, Center Sampling is still adopted to maintain the strong receptive field for candidates, with a slight adjustment. To expand the range of candidate selection, we put forward Multi-layer Center Sampling (MCS), which applies Center Sampling across all FPN layers, rather than sampling with an increasing radius on a single layer (as proven ineffective in Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.908, + 0.446 + ], + "angle": 0, + "content": "(2) Matching Stage. The purpose of this stage is to select positive or negative labels for candidates based on their matching costs. The Matching cost is defined as a weighted summation of classification loss \\( L_{cls} \\), regression loss \\( L_{reg} \\), and polygon loss \\( L_{poly} \\), which together evaluate the optimized cost from the prediction to ground truth. In this work, we employ Focal Loss [28], Polar IoU Loss [36], and the proposed RMask IoU Loss as classification loss, regression loss, and polygon loss, respectively. For a given instance, the matrix of matching costs \\( Cost = \\{c_i | i = 1, \\dots, n\\} \\) for \\( n \\) candidates can be formulated as follows:" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.14, + 0.908, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.456, + 0.905, + 0.473 + ], + "angle": 0, + "content": "\\[\nC o s t = \\lambda_ {c l s} L _ {c l s} + \\lambda_ {r e g} L _ {r e g} + \\lambda_ {p o l y} L _ {p o l y}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.906, + 0.541 + ], + "angle": 0, + "content": "where \\(\\lambda_{cls}\\), \\(\\lambda_{reg}\\) and \\(\\lambda_{poly}\\) stand for the loss coefficients. Afterward, a fixed number \\(k\\) of positive labels are assigned to the candidates with minimal costs, and the others are negative:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.549, + 0.905, + 0.582 + ], + "angle": 0, + "content": "\\[\nX _ {i} = \\left\\{ \\begin{array}{l l} \\text {P o s i t i v e}, & c _ {i} \\in \\operatorname {t o p} - k _ {\\min } \\left(c _ {1}, c _ {2}, \\dots , c _ {n}\\right) \\\\ \\text {N e g a t i v e}, & c _ {i} \\notin \\operatorname {t o p} - k _ {\\min } \\left(c _ {1}, c _ {2}, \\dots , c _ {n}\\right), \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.636 + ], + "angle": 0, + "content": "where \\( X_{i} \\) refers to the label assigned to the \\( i \\)-th candidates. Notably, if a sample is matched by two instances simultaneously, the one with the largest area will be chosen." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.906, + 0.774 + ], + "angle": 0, + "content": "(3) Weighting Stage. In the final stage, each positive sample is provided with a quality score, which modulates its contribution to loss functions. Our approach here is straightforward but effective, replacing the heuristic Centerness with the RMask IoU output by URM. Compared to Centerness, RMask IoU treats each positive sample equally, especially those high-quality samples located outside the contour, whose weights are zeroed out by Centerness (e.g., samples 5 and 6 in Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.906, + 0.805 + ], + "angle": 0, + "content": "Loss Function. The overall loss function \\( L \\) consists of four parts:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.815, + 0.905, + 0.832 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {c l s} L _ {c l s} + \\lambda_ {r e g} L _ {r e g} + \\lambda_ {p o l y} L _ {p o l y} + \\lambda_ {s r e} L _ {s r e}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.9 + ], + "angle": 0, + "content": "where \\( L_{sre} \\) denotes the Cross-Entropy Loss in PolarMask used to supervise the Centerness head. In this case, we replace the original supervision signals with RMask IoU in the Centerness head." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19319" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.14, + 0.089, + 0.86, + 0.372 + ], + "angle": 0, + "content": "
TypeMethodBackboneEpochSize\\(AP\\)\\(AP_{50}\\)\\(AP_{75}\\)FPS\\(Params\\downarrow\\)FLOPs↓
ContourDeepSnake [25]DLA-3416051230.3--24--
E2EC [39]DLA-3415051233.852.932.83529.80121
PolySnake [9]DLA-3425051234.4--18--
DANCE [24]R50-FPN3680036.858.539.01644.60274
MaskMask R-CNN [14]R50-FPN3680037.559.140.23344.40240
Cascade R-CNN [2]R50-FPN3680038.459.741.52377.331804
HTC [3]R50-FPN2080038.760.141.91780.261735
MS R-CNN [15]R50-FPN2480036.856.739.73060.74279
PointRend [17]R50-FPN3680038.459.541.52260.22266
CondInst [32]R50-FPN3680036.257.238.72934.17308
SOLO [34]R50-FPN3680036.457.938.92536.31332
SOLOv2 [35]R50-FPN3680038.459.541.23546.60227
YOLACT [1]R50-FPN5580029.048.729.84735.30237
SparseInst [7]R50-FPN14680033.253.434.75932.80201
PolarESE-Seg [38]R50-FPN30080021.648.722.438--
PolarMask [36]R50-FPN3680031.352.532.34234.74264
PolarNeXt (ours)R50-FPN3680036.159.737.34932.36186
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.383, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Table 1. Experimental results on COCO test-dev. \"Size\" refers to the length of the shortest edge of input images. \\( {}^{ \\blacklozenge } \\) means smaller values are preferred." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.421, + 0.465, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Figure 5. Comparison of Memory Usage (MB) for inference. During inference on COCO test-dev images, the peak memory usage is recorded as the final result of each method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.225, + 0.691 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.445, + 0.714 + ], + "angle": 0, + "content": "4.1. Experiments and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Experiment Settings. The experiments are conducted on MS COCO [21] benchmark and Cityscapes [8] dataset. MS COCO contains 80 classes with \\(115\\mathrm{k}\\) training, \\(5\\mathrm{k}\\) validation, and \\(20\\mathrm{k}\\) testing images. Cityscapes has 2795 training images from 8 classes and another 500 images with high-quality annotations for validation. In terms of evaluation metrics, the accuracy of segmentation is evaluated by the standard AP metric, denoted as mask AP. Furthermore, mask AP can be divided into \\(AP_{50} / AP_{75}\\) based on different IoU thresholds, as well as \\(AP_{S} / AP_{M} / AP_{L}\\) for small/medium/large objects. All inference experiments are conducted on a single NVIDIA 4090D GPU, where FPS," + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.421, + 0.895, + 0.523 + ], + "angle": 0, + "content": "
MethodAPAP50FPSMems↓
Mask R-CNN [14]31.558.6161968
SOLOv2 [35]27.248.5172076
CondInst [32]33.359.1152319
PolarMask [36]27.452.328575
PolarNeXt (ours)30.758.831566
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.533, + 0.907, + 0.562 + ], + "angle": 0, + "content": "Table 2. Experimental results on Cityscapes val set. \"Mems\" stands for the memory usage (MB)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.642 + ], + "angle": 0, + "content": "Parms, and FLOPs are calculated separately for inference speed, model complexity, and computational overhead. Notably, in all our experiments, TensorRT or FP16 is not used for acceleration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. PolarNeXt is instantiated on the PolarMask network, using a backbone ResNet-50 [13] pretrained on ImageNet [29], and removed the auxiliary box branch. All models are implemented on the MMDetection toolbox [4], trained on 2 GPUs with 2 images per GPU, and optimized with SGD. Weight decay and momentum are configured to 0.0001 and 0.9, respectively. The base learning rate is set to 0.01 without any decay scheme. For a fair comparison, following common practice, data augmentation only contains random flip and scale jitter unless specified. In APSD, the radius of Center Sampling for candidate selection at each FPN layer is set to 1.5 and the number \\( k \\) of assigned positive labels is set to 9 for each instance. Loss coefficients \\( \\lambda_{cls} \\), \\( \\lambda_{reg} \\), \\( \\lambda_{poly} \\), and \\( \\lambda_{sre} \\) are all empirically set to 1.0. Following the rasterizer configuration in [18], we set the desired mask resolution to \\( 64 \\times 64 \\) and the rasterization sharpness to \\( \\tau = 0.1 \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19320" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.13, + 0.089, + 0.445, + 0.22 + ], + "angle": 0, + "content": "
URMAPSDAP
weightlosscost
29.1
✓30.3 (+1.2)
✓✓31.0 (+1.9)
✓28.6 (-0.5)
✓✓✓30.7 (+1.6)
✓✓✓✓33.9 (+4.8)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.232, + 0.482, + 0.302 + ], + "angle": 0, + "content": "Table 3. Ablation studies on URM and APSD. \"weight\" means whether Centerness is replaced by RMask IoU for sample weighting. \"loss\" refers to the use of RMask IoU Loss in loss functions. \"cost\" indicates the incorporation of RMask IoU Loss into matching costs." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.312, + 0.453, + 0.369 + ], + "angle": 0, + "content": "
Polar IoU LossAPAPSAPMAPL
w/o.32.814.335.148.5
w/.33.916.435.949.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.422 + ], + "angle": 0, + "content": "Table 4. Feasibility analysis of removing Polar IoU Loss. The symbols \"w/o.\" and \"w/.\" indicate whether the Polar IoU Loss is employed." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.437, + 0.229, + 0.451 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.777 + ], + "angle": 0, + "content": "MS COCO. Following common practice, we evaluate PolarNeXt on the MS COCO benchmark and compare test-dev results to some state-of-the-art models in Tab. 1. In terms of computational cost, PolarNeXt is more lightweight than the other models with R50-FPN backbone, only requiring 32.8 MB Parameters and 186 GB FLOPs. Even when compared to the contour-based methods using DLA-34 backbone, it remains on par. In terms of inference speed, PolarNeXt is second only to the prevalent SpareInst, but it achieves higher accuracy (36.1% vs. 33.2% AP) with a shorter training schedule (36 vs. 146 epochs). In terms of segmentation accuracy, PolarNeXt delivers competitive results compared to other more complex methods. Particularly, without bells and whistles, PolarNeXt achieves a 4.8% mask AP improvement over PolarMask, redemonstrating the potential of Polar Representation in instance segmentation. Moreover, we visualize the memory usage of some representative methods during inference in Fig. 5. Obviously, PolarNeXt shows its consistency with the classical object detector FCOS, with memory usage significantly lower than other instance segmentation methods, only half or even less." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Cityscapes. As an extended comparative experiment in Tab. 2, Cityscapes dataset, characterized by high-resolution images, is utilized to examine the robustness and generalization ability of our model. Notably, all models are directly trained on the Cityscapes training set under 64 epochs, without COCO initialization. Since fragmented instances frequently occur in Cityscapes, the Multi-component Detection Strategy suggested in [25] is applied to PolarNeXt and" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.09, + 0.921, + 0.206 + ], + "angle": 0, + "content": "
Candidate SelectionLayerAPT.T.
Center Sampling (s.r.=1.5)single31.013h
Center Sampling (s.r.=2.0)30.6 (-0.4)14h
All Grid Pointsmulti--
Points within Boxes33.1 (+2.1)28h
Points within Contours33.0 (+2.0)25h
Center Sampling (s.r.=1.5)33.9 (+2.9)16h
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.217, + 0.907, + 0.287 + ], + "angle": 0, + "content": "Table 5. The effectiveness of some approaches for candidate selection. \"single\" denotes selecting candidates on a specific FPN layer based on the instance size, and \"multi\" refers to selecting candidates across all FPN layers. \"T.T.\" and \"s.r.\" are abbreviations for training time and sampling radius, respectively." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.298, + 0.892, + 0.354 + ], + "angle": 0, + "content": "
MethodAPbAP50APSbAPMAPLb
PolarMask16.338.316.028.413.2
PolarNeXt19.846.720.232.215.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.365, + 0.907, + 0.392 + ], + "angle": 0, + "content": "Table 6. Comparison of boundary quality between PolarMask and PolarNeXt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.577 + ], + "angle": 0, + "content": "PolarMask, leading to approximately a \\(1.5\\%\\) AP improvement. Under the increased resolution of input images (from \\(800 \\times 1333\\) to \\(1024 \\times 2048\\)), mask-based methods suffer from the heavy memory usage introduced by dense pixel classification, which is nearly four times greater than that of polar-based methods. Meanwhile, the speed advantage of polar-based methods becomes more apparent, requiring only half the inference time compared to mask-based methods. As expected, PolarNeXt maintains a \\(3.3\\%\\) AP improvement over PolarMask, further demonstrating the effectiveness of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.68, + 0.603 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.686 + ], + "angle": 0, + "content": "In this section, we conduct ablations to assess the influence of the proposed components on instance segmentation performance, and then analyze their details in the following. All experiments are conducted on MS COCO val under a 1x training schedule." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Component Impact and Correlation. The effectiveness of URM and APSD is reported in Tab. 3. Using URM alone brings a \\(1.9\\%\\) AP improvement on the baseline, while combining it with APSD further boosts the performance by an additional \\(2.9\\%\\) AP. It is worth noting that APSD fails to function independently of URM, the absence of which results in a decrease in performance. To be specific, without APSD, URM contributes to the sample weighting and loss function, leading to a \\(1.2\\%\\) and \\(0.7\\%\\) AP increase, respectively. Conversely, when APSD is applied, the matching costs have to incorporate RMask IoU Loss; otherwise, this strategy will fail or even backfire. In view of this, we conclude that the effectiveness of our adaptive strategy hinges on reliable polygonal assessment for support." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19321" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.089, + 0.868, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.332, + 0.692, + 0.347 + ], + "angle": 0, + "content": "Figure 6. Visualization with Rays for PolarMask and PolarNeXt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.544 + ], + "angle": 0, + "content": "Candidate Selection. In Tab. 5, we compare some approaches to expand the range of candidate selection. On the one hand, Center Sampling on multiple FPN layers is more effective and efficient than the dense point selection within boxes or contours. In particular, the approach suggested in [31], which selects all grid points of FPN, results in unacceptable training time and GPU memory overflow. On the other hand, directly expanding the radius of Center Sampling on a single layer causes a \\(0.4\\%\\) drop in AP, as the additional selected samples tend to fall outside the receptive field. Accordingly, the center prior still has a positive influence on sample decisions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.671 + ], + "angle": 0, + "content": "RMask IoU vs. Polar IoU. In Fig. 4, two scatter plots are drawn to compare Polar IoU and RMask IoU in polygonal assessment. These scatter points denote all the polygons predicted during inference on MS COCO val images. The fitted curve in Fig. 4(b) is almost aligned with the ideal curve, whereas the one in Fig. 4(a) shows significant deviation. This indicates that RMask IoU better reflects the actual polygonal quality compared to Polar IoU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The necessity of Polar IoU Loss. In Tab. 4, we explore the feasibility of directly removing Polar IoU Loss in our proposed training pipeline. Experimental results indicate that this operation results in a \\(1.1\\%\\) AP degradation in performance. The impact on small objects is more notable, with \\(AP_{S}\\) decreased by \\(2.1\\%\\). We analyze that there may be a slight granularity missing as well as feature misalignment during the conversion from vertex sets to rasterized masks. Moreover, as mentioned in [18], aliasing occurs in the deconvolutional process of FPN when pre-predicted RoIs are unavailable in one-stage detectors. Fortunately, the combination of Polar IoU Loss and RMask IoU Loss brings a notable boost, with the two complementing each other. We believe that Polar IoU Loss serves as a polygonal internal constraint for dense distance regression, while RMask IoU" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.363, + 0.907, + 0.394 + ], + "angle": 0, + "content": "Loss provides a polygonal external constraint to incorporate the representation deviation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.401, + 0.761, + 0.417 + ], + "angle": 0, + "content": "4.4. Boundary Quality Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.907, + 0.575 + ], + "angle": 0, + "content": "In Tab. 6, we introduce Boundary AP [5] (\\(AP^b\\) for short) to evaluate the boundary quality of predicted polygons. Experimental results show that PolarNeXt achieves a significant improvement in boundary quality compared to PolarMask, boosting \\(AP^b\\) by \\(3.5\\%\\). Furthermore, we present some segmentation results with differences in starting points in Fig. 6. Obviously, compared to the center-fixed starting points in PolarMask, the dynamically selected starting points in PolarNeXt are more flexible, which enhances the representation capability of bounding polygons." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.585, + 0.634, + 0.601 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.909, + 0.822 + ], + "angle": 0, + "content": "In this paper, we introduce a novel polar-based framework, PolarNeXt, which redemonstrates the superiority of Polar Representation in instance segmentation. To account for the previously ignored representation error, tailored training strategy APSD and assessment module URM are proposed. Together, these components significantly enhance PolarNeXt, yielding notable performance improvements over other polar-based methods. By fully leveraging the advantages of Polar Representation, PolarNeXt achieves comparable segmentation accuracy while requiring only half or even less of the computational cost and inference time compared to mainstream mask-based and contour-based methods. We anticipate that PolarNeXt will set a new standard for polar-based methods in instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported by the National Natural Science Foundation of China [Grant No. 62225308], and the Shanghai Technical Service Computing Center of Science and Engineering, Shanghai University." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19322" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, pages 9157-9166, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE TPAMI, 43(5):1483-1498, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.2, + 0.483, + 0.255 + ], + "angle": 0, + "content": "[3] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In CVPR, pages 4974-4983, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.483, + 0.323 + ], + "angle": 0, + "content": "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.381 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ross Girshick, Piotr Dólar, Alexander C Berg, and Alexander Kirillov. Boundary iou: Improving object-centric image segmentation evaluation. In CVPR, pages 15334-15342, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[6] Dominic Cheng, Renjie Liao, Sanja Fidler, and Raquel Urtasun. Darnet: Deep active ray network for building segmentation. In CVPR, pages 7431-7439, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.479 + ], + "angle": 0, + "content": "[7] Tianheng Cheng, Xinggang Wang, Shaoyu Chen, Wenqiang Zhang, Qian Zhang, Chang Huang, Zhaoxiang Zhang, and Wenyu Liu. Sparse instance activation for real-time instance segmentation. In CVPR, pages 4433-4442, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.549 + ], + "angle": 0, + "content": "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[9] Hao Feng, Keyi Zhou, Wengang Zhou, Yufei Yin, Jiajun Deng, Qi Sun, and Houqiang Li. Recurrent generic contour-based instance segmentation with progressive learning. IEEE TCSVT, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[10] Gunther Greiner and Kai Hormann. Efficient clipping of arbitrary polygons. ACM Transactions on Graphics (TOG), 17 (2):71-83, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[11] Wenchao Gu, Shuang Bai, and Lingxing Kong. A review on 2d instance segmentation based on deep neural networks. Image and Vision Computing, 120:104401, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[12] Shir Gur, Tal Shaharabany, and Lior Wolf. End to end trainable active contours via differentiable rendering. arXiv preprint arXiv:1912.00367, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[15] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In CVPR, pages 6409-6418, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[16] Michael Kass, Andrew Witkin, and Demetri Terzopoulos. Snakes: Active contour models. IJCV, 1(4):321-331, 1988. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[17] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In CVPR, pages 9799-9808, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[18] Justin Lazarow, Weijian Xu, and Zhuowen Tu. Instance segmentation with mask-supervised polygonal boundary transformers. In CVPR, pages 4382-4391, 2022. 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[19] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. NeurIPS, 33:21002-21012, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[20] Justin Liang, Namdar Homayounfar, Wei-Chiu Ma, Yuwen Xiong, Rui Hu, and Raquel Urtasun. Polytransform: Deep polygon transformer for instance segmentation. In CVPR, pages 9131-9140, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.404, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[23] Huan Ling, Jun Gao, Amlan Kar, Wenzheng Chen, and Sanja Fidler. Fast interactive object annotation with curve-gcn. In CVPR, pages 5257-5266, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[24] Zichen Liu, Jun Hao Liew, Xiangyu Chen, and Jiashi Feng. Dance: A deep attentive contour model for efficient instance segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 345-354, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.518, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[25] Sida Peng, Wen Jiang, Huaijin Pi, Xiuli Li, Hujun Bao, and Xiaowei Zhou. Deep snake for real-time instance segmentation. In CVPR, pages 8533-8542, 2020. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.904, + 0.588 + ], + "angle": 0, + "content": "[26] Joseph Redmon. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[28] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[30] Uwe Schmidt, Martin Weigert, Coleman Broaddus, and Gene Myers. Cell detection with star-convex polygons. In Medical Image Computing and Computer Assisted Intervention, pages 265–273, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[31] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In International Conference on Machine Learning, pages 9934-9944, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[32] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, pages 282-298, 2020. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19323" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[33] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: A simple and strong anchor-free object detector. IEEE TPAMI, 44(4):1922-1933, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[34] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, pages 649-665, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[35] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmentation. NeurIPS, 33:17721-17732, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[36] Enze Xie, Peize Sun, Xiaoge Song, Wenhai Wang, Xuebo Liu, Ding Liang, Chunhua Shen, and Ping Luo. Polarmask: Single shot instance segmentation with polar representation. In CVPR, pages 12193-12202, 2020. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[37] Enze Xie, Wenhai Wang, Mingyu Ding, Ruimao Zhang, and Ping Luo. Polarmask++: Enhanced polar representation for single-shot instance segmentation and beyond. IEEE TPAMI, 44(9):5385-5400, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[38] Wenqiang Xu, Haiyang Wang, Fubo Qi, and Cewu Lu. Explicit shape encoding for real-time instance segmentation. In ICCV, pages 5168-5177, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[39] Tao Zhang, Shiqing Wei, and Shunping Ji. E2ec: An end-to-end contour-based method for high-quality high-speed instance segmentation. In CVPR, pages 4443-4452, 2022. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19324" + } + ] +] \ No newline at end of file diff --git a/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/a5f92f60-4941-4386-a4db-1f556681a113_origin.pdf b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/a5f92f60-4941-4386-a4db-1f556681a113_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..49b88068fbebe8989c932e0d338e7208b002f001 --- /dev/null +++ b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/a5f92f60-4941-4386-a4db-1f556681a113_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57a641a3a502e712afb6a750079bb21abf9787b8bea26234d35a9f0e562759d0 +size 7393675 diff --git a/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/full.md b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4dea49fc4543657e05a7d083471d1b9f5cf85d03 --- /dev/null +++ b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/full.md @@ -0,0 +1,245 @@ +# PolarNeXt: Rethink Instance Segmentation with Polar Representation + +Jiacheng Sun $^{1}$ , Xinghong Zhou $^{1}$ , Yiqiang Wu $^{1}$ , Bin Zhu $^{1}$ , Jiaxuan Lu $^{2}$ , Yu Qin $^{1}$ , Xiaomao Li $^{1,*}$ $^{1}$ Shanghai University, $^{2}$ Shanghai Artificial Intelligence Laboratory + +{sunjc, jkchou, yiqiangwu, zhubin01, qy21722148, lixiaomao}@shu.edu.cn + +lujiaxuan@pjlab.org.cn + +# Abstract + +One of the roadblocks for instance segmentation today is heavy computational overhead and model parameters. Previous methods based on Polar Representation made the initial mark to address this challenge by formulating instance segmentation as polygon detection, but failed to align with mainstream methods in performance. In this paper, we highlight that Representation Errors, arising from the limited capacity of polygons to capture boundary details, have long been overlooked, which results in severe performance degradation. Observing that optimal starting point selection effectively alleviates this issue, we propose an Adaptive Polygonal Sample Decision strategy to dynamically capture the positional variation of representation errors across samples. Additionally, we design a Union-aligned Rasterization Module to incorporate these errors into polygonal assessment, further advancing the proposed strategy. With these components, our framework PolarNeXt achieves a remarkable performance boost of over $4.8\%$ AP compared to other polar-based methods. PolarNeXt is markedly more lightweight and efficient than state-of-the-art instance segmentation methods, while achieving comparable segmentation accuracy. We expect this work will open up a new direction for instance segmentation in high-resolution images and resource-limited scenarios. Codes can be found at https://github.com/Sun15194/PolarNeXt. + +# 1. Introduction + +Object detection and instance segmentation are fundamental tasks in computer vision, aiming at identifying specific objects within images. Object detection [26, 27, 33], also known as box detection, localizes objects using rectangular boxes, which are efficient but coarse-grained. In comparison, instance segmentation offers additional geometric information, such as shape and boundary. Mainstream methods define instance segmentation as pixel-wise classifica + +![](images/18c0e919d50ff766d9db2bba7da3d8be3142a08044d4b12e0381fce0d959959c.jpg) +Figure 1. Illustration of the composition of prediction errors in polar-based methods. Classification and Regression (Cls.&Reg.) Errors are produced during network optimization, which reflects the deviation from starting point classification $(s\rightarrow \hat{s})$ and dense distance regression $(D\to \hat{D})$ . The ignored Representation (Rep.) Error arises from the missing boundary details of instance contour $C$ in Polar Representation, as exemplified by the two cases below. Obviously, the center point does not necessarily coincide with the optimal point, from which the assembled polygon with minimal Rep. Error can be predicted. + +tion [7, 14, 34] or iterative deformation [9, 25, 39]. Despite the finer granularity, these methods introduce heavy computational overhead and model parameters, which restricts their applicability in high-resolution images or resource-limited scenarios [11]. + +To combine the efficiency of detection with the granularity of segmentation, a pioneering attempt is Polar Representation. As shown at the top of Fig. 1, Polar Representation reconstructs the instance contour into a bounding polygon, assembled by dense rays emitted uniformly from a starting point. Building on this concept, typical polar-based methods [36-38] formulate instance segmentation as starting point classification and dense distance regression. These methods can be instantiated on one-stage object detectors [26, 33], directly inheriting their advantages in inference speed and model complexity. However, Polar Representation has not become a fundamental paradigm for in + +![](images/f2291bc6b548b54c582c4e8b602cc289b04fc1d80ccd807848f77bf06bc006a7.jpg) +Figure 2. An overview of the proposed Adaptive Polygonal Sample Decision (APSD). The three stages of APSD are labeled as: I.Sampling Stage, II.Matching Stage, and III.Weighting Stage. The Union-aligned Rasterization Module (URM) takes the predicted polygons $\hat{P}$ of the candidates selected by Multi-layer Center Sampling (MCS) as input, and outputs their RMask IoU with the instance contours $C$ . RMask IoU functions in both matching costs and sample weighting. + +stance segmentation, due to the significant performance gap between polar-based methods and other advanced methods. + +In this work, we highlight that an additional prediction error has been ignored in Polar Representation, which results in the performance degradation of polar-based methods. As illustrated at the bottom of Fig. 1, from different starting points, the assembled polygons exhibit varying degrees of deviation with their corresponding contours. This deviation, termed Representation (Rep.) Error, inevitably occurs in Polar Representation, where the boundary details of contours are complex for polygons to reflect, especially in disconnected or concave regions. However, existing training pipelines in polar-based methods only account for the Classification and Regression (Cls.&Reg.) Errors generated during network optimization. Their label assignment and sample weighting, collectively referred to as sample decisions, rely entirely on the handcrafted center prior, which fails to capture the positional variation of representation errors (as proven in Sec. 3.2(1)). Furthermore, even though a natural idea is to dynamically select the optimal starting point with minimal representation error, a reliable metric to incorporate this error into polygonal assessment is still lacking. As the only available metric, Polar IoU [36] exhibits assessment blindness to representation errors (as verified in Sec. 3.2(2)), which is too coarse-grained to support this idea. + +To overcome these challenges, we propose PolarNeXt, a highly lightweight yet effective polar-based framework for real-time instance segmentation. The training pipeline of this framework includes an Adaptive Polygonal Sample Decision (APSD) strategy, enabling optimal starting point + +selection during inference. APSD assigns positive labels and higher weights to the candidate samples based on their matching costs, which account for the comprehensive errors between predictions and ground truths. To eliminate the assessment blindness to representation errors, a Union-aligned Rasterization Module (URM) is designed for more accurate polygonal assessment. The paired polygon and contour are aligned within their Union Box and then converted into rasterized masks for RMask IoU measurement. Here, RMask IoU serves as a new metric to evaluate the actual polygonal overlap, forming the core element of sample decisions and loss functions. Benefitting from these components, PolarNeXt achieves a remarkable improvement of over $4.8\%$ AP compared to other polar-based methods. To the best of our knowledge, this is the first polar-based method competitive with state-of-the-art instance segmentation methods in performance. Notably, PolarNeXt requires only half or even less of their computational cost and inference time, which redemonstrates the superiority of Polar Representation in instance segmentation. The main contributions of this work can be summarized as follows: + +- We conduct a thorough analysis of the implementation challenges of Polar Representation in instance segmentation, and reveal that the performance limitations of existing polar-based methods stem from their neglect of the representation error. +- We propose a new polar-based instance segmentation framework called PolarNeXt, including an APSD strategy to dynamically account for representation errors across samples and a URM module that eliminates the assessment blindness to these errors using RMask IoU. + +- We fully leverage the advantages of Polar Representation in instance segmentation tasks. Without bells and whistles, PolarNeXt is markedly more lightweight and efficient than other mainstream methods, while aligning with them in terms of segmentation performance. + +# 2. Related Work + +Instance Segmentation. Mainstream instance segmentation methods can be classified into two categories based on their representation, i.e., mask-based methods and contour-based methods. Mask-based methods treat instance segmentation as pixel-wise classification, in which binary masks are predicted within their Regions of Interest (RoIs). As a pioneering work, Mask R-CNN [14] introduces a detect-then-segment paradigm, built on the two-stage detector [27], with box proposals serving as the RoIs for pixel-by-pixel segmentation. Nowadays, state-of-the-art benchmark results are still held by the mask-based two-stage methods [2, 3, 15] constructed on strong detection baselines. Additionally, to reduce reliance on bounding boxes, one-stage methods [32, 34, 35] treat the entire image as the RoI, predicting masks by the assembly of encoded mask vectors and mask kernels from dynamic convolutions. Other mask-based methods focus on the high-quality boundary [17, 20], or the real-time inference [1, 7]. Contour-based methods take a different paradigm, formulating instance segmentation as iterative deformation based on the active contour mechanism of Snake [16]. Deep Snake [25] establishes the first multi-stage deformation network, which progressively refines the pre-predicted initial contours using their vertex features. Based on this baseline, further optimization strategies have been explored, such as Segment-wise Matching Scheme [24], Dynamic matching Loss [39], and Multi-scale Contour Refinement [9]. However, the high computational cost introduced by pixel-wise classification and iterative deformation limits the applicability and prevalence of these methods in high-resolution images or resource-limited scenarios [11]. To solve this problem, we bring back Polar Representation and redemonstrate its superiority in instance segmentation. Building on Polar Representation, our proposed PolarNeXt is markedly more lightweight and efficient than the aforementioned methods, while remaining competitive in segmentation performance. + +Polar Representation. Polar Representation initially appears in some interdisciplinary fields, such as automatic building segmentation [6] and medical cell detection [30]. Subsequently, ESE-Seg [38] migrates Polar Representation into instance segmentation tasks, introducing Chebyshev polynomial fitting to shorten the distance vectors and resist noise. PolarMask [36] and its variant [37] make the initial mark of polar-based methods on instance segmentation benchmarks. They maximize the advantage of the + +center prior by modified Center Sampling and Polar Centerness, along with a proposed Polar IoU metric for polygonal assessment. However, the representation error introduced by Polar Representation has been ignored in these polar-based methods, which results in severe performance degradation. To address this, we propose an APSD strategy to dynamically balance representation errors across samples, in conjunction with a URM module that incorporates these errors into polygonal assessment. Together, these components bring a notable performance boost to PolarNeXt. + +# 3. Methods + +In this work, PolarMask [36], a representative polar-based method, is used as a case study for Polar Representation. The design principle and training pipeline of PolarMask are reviewed in Sec. 3.1. Under this pipeline, the effectiveness of its sample decisions and polygonal assessment is investigated in Sec. 3.2. Finally, a novel polar-based framework called PolarNeXt is detailed in Sec. 3.3 to advance Polar Representation in instance segmentation tasks. + +# 3.1. Preliminary + +PolarMask defines instance segmentation as starting point classification and dense distance regression. As illustrated in Fig. 1, for a given instance contour $C$ , bounding polygon $P = \psi(s, D)$ can be constructed through Polar Representation $\psi$ , using a distance set $D = \{d_i | i = 1, \dots, m\}$ from a starting point $s$ . The starting point $s$ and distance set $D$ serve as supervision signals for the network, which then classifies a starting point $\hat{s}$ and regresses an equal-size set $\hat{D} = \{\hat{d}_i | i = 1, \dots, m\}$ to construct predicted polygon $\hat{P} = \psi(\hat{s}, \hat{D})$ . In the training pipeline of PolarMask, the center prior fully dictates label assignment and sample weighting, given that samples closer to the center tend to have richer receptive field and lower regression difficulty [19]. First, Center Sampling is applied on a single FPN [22] layer, directly assigning positive labels to the samples around instance centers. Second, Polar Centerness, which quantifies the proximity to the center, weights these positive samples to modulate their contribution to loss functions. Furthermore, due to the lack of polygon-specific IoUs, Polar IoU is proposed as a metric for polygonal assessment. It compares the consistency between target distance set $D$ and predicted distance set $\hat{D}$ to approximate the overlap between $P$ and $\hat{P}$ . Additionally, Polar IoU is transformed into Polar IoU Loss to supervise distance regression. More details are provided in supplementary materials. + +# 3.2. Investigation + +In this section, we identify that two challenges lie in implementing Polar Representation for instance segmentation: + +![](images/80afa08be13c791a94722d44f607be8f6a30e6f2ace8684d0fef8a67ed6c6f03.jpg) +Figure 3. Illustration of the divergence between Polarness and Centerness. On the left, heatmaps of these two attributes are visualized to reveal the distributional divergence, with positive samples (9 white points) selected by Center Sampling marked. Then, attribute values of these samples are plotted to verify the quantitative divergence on the right, where samples are sorted by ascending Polarness. + +(1) Diminished Effectiveness of Center Prior. The center prior plays a diminished role in sample decisions, which fails to capture the positional variation of representation errors. To validate this, a comparative experiment is conducted in Fig. 3, illustrating the divergence between the center prior and representation error. For a clear comparison, we define Polarness $(Plr)$ as an inverse measure of Representation Error $(RE)$ , where $Plr = 1 - RE$ . Likewise, normalized Polar Centerness is introduced to quantify the influence of the center prior. Experimental results show that Polarness presents no observable correlation with the distribution pattern of Centerness, while Centerness does not increase monotonically on these samples sorted by ascending Polarness. Accordingly, the samples selected by center sampling do not necessarily correspond to lower representation errors, and Centerness also fails to accurately modulate the weights of high-quality samples. Based on these observations, we conclude that the sample decisions entirely reliant on the center prior are less applicable in polar-based methods, as some suboptimal samples in center regions receive excessive attention. + +(2) Assessment Blindness of Polar IoU. As the only available metric, Polar IoU exhibits assessment blindness to representation errors. This limitation arises because Polar IoU evaluates the consistency between target polygon $P$ and predicted polygon $\hat{P}$ by comparing their distance sets, entirely independent of instance contour $C$ . To explore the impact of this blindness on polygonal assessment, a scatter plot is drawn to examine the effectiveness of Polar IoU. As shown in Fig. 4(a), the scatter points show a significant departure from the expected linear trend, with substantial noise introduced. For example, when Real IoU reaches 0.4, Polar IoU fluctuates substantially, ranging irregularly from 0.4 to 0.9. Based on these observations, we conclude that + +![](images/f13d06d9cb4a848b9b7427846efb9181be1fb12ecdfa83c61d5ee4d60d0072da.jpg) +Figure 4. Comparison between Polar IoU and our proposed RMask IoU. The green solid lines denote the curves fitted by scatter points, while the red dashed lines indicate the desired linear correlation trend. Real IoU is the polygon-specific IoU between the predicted polygon and its corresponding instance contour, calculated using Weiler-Atherton algorithm [10]. + +Polar IoU fails to function as a reliable metric for polygonal assessment. + +# 3.3. PolarNeXt + +To overcome these two challenges, a new framework named PolarNeXt is proposed, which is expected to become the next level of polar-based methods for instance segmentation. The training pipeline of PolarNeXt includes two tailored components for Polar Representation: Adaptive Polygonal Sample Decision (APSD) and Union-aligned Rasterization Module (URM). As illustrated in Fig. 2, APSD samples a wide range of candidates around instances on the input image. Then, those candidates with minimal matching costs are assigned positive labels and weighted by the quality scores of their predicted polygons. For polygonal assessment, URM aligns the paired polygon and contour within their Union Box, and converts them into rasterized masks for overlap measurement. The RMask IoU output by URM incorporates representation errors, serving as precise guidance in both label assignment and sample weighting. The detailed design principles of our proposed methods are outlined below. + +Union-aligned Rasterization Module. To eliminate the assessment blindness to representation errors, a straightforward idea is to bridge the predicted polygon $\hat{P}$ with its instance contour $C$ and measure their polygonal overlap. However, traditional polygon-specific IoU algorithms, such as Weiler-Atherton [10], are impractical for parallel processing, due to uncertain vertex counts and inevitable recursive operations. Inspired by 3D renderers [12, 18, 23], our solution is to convert $\hat{P}$ and $C$ from vertex-set format into rasterized-mask format for mask-level overlap measurement, but an additional challenge lies in their alignment at a fixed resolution, given that no pre-predicted RoIs are available in one-stage detectors. + +As shown in Fig. 2, we propose a Union-aligned Rasterization Module (URM) which employs Union Box, the minimal bounding box enclosing all the vertices of $\hat{P}$ and $C$ , as a reference for coordinate alignment. Concretely, assuming the vertex sets of $\hat{P}$ and $C$ are symbolized by $V_{\hat{P}}$ and $V_{C}$ , with a desired mask resolution for rasterization specified as $H\times W$ . The coordinate $(v_x,v_y)$ of a vertex $v$ can be adjusted by the union set $U = V_{\hat{P}}\cup V_{C}$ as follows: + +$$ +\left( \begin{array}{c} v _ {x} \\ v _ {y} \end{array} \right) = \left( \begin{array}{c} \frac {v _ {x} - \underset {u \in U} {\min } u _ {x}}{\underset {u \in U} {\max } u _ {x} - \underset {u \in U} {\min } u _ {x}} \times W \\ \frac {v _ {y} - \underset {u \in U} {\min } u _ {y}}{\underset {u \in U} {\max } u _ {y} - \underset {u \in U} {\min } u _ {y}} \times H \end{array} \right). \tag {1} +$$ + +Then, the differentiable rasterizer in [18] is introduced to transform the aligned vertex sets into rasterized masks. For pixel $p(x,y)$ of a mask, where $0 \leq x \leq W$ and $0 \leq y \leq H$ , two attributes are used to model its contribution value $M[x,y]$ to the mask: (1) the binary inside-outside state $IO(V,p)$ of pixel $p$ relative to vertex set $V$ , and (2) the closest distance $CD(V,p)$ from pixel $p$ to vertex set $V$ . The contribution value $M[x,y]$ is formulated as shown below: + +$$ +M [ x, y ] = \operatorname {s i g m o i d} \left(\frac {I O (V , p) \times C D (V , p)}{\tau}\right), \tag {2} +$$ + +where $\tau$ is a hyperparameter for sharpness. Finally, two rasterized masks $M_{\hat{P}}$ and $M_C$ with the same resolution $H\times W$ are produced, and their mask-level IoU, called RMask IoU, can be calculated as follows: + +$$ +\mathrm {R M a s k} \mathrm {I o U} = \frac {2 \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \mathrm {M} _ {C} [ \mathrm {i} , \mathrm {j} ] \cdot \mathrm {M} _ {\hat {P}} [ \mathrm {i} , \mathrm {j} ]}{\sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \mathrm {M} _ {C} [ \mathrm {i} , \mathrm {j} ] + \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} \mathrm {M} _ {\hat {P}} [ \mathrm {i} , \mathrm {j} ]}. \tag {3} +$$ + +As shown in Fig. 4(b), RMask IoU accurately reflects the actual overlap between polygons and contours. Further, we perform a simple deformation on RMask IoU to obtain a new loss function, RMask IoU Loss, for polygonal external constraints: + +$$ +\mathrm {R M a s k} \quad \mathrm {I o U} \quad \text {L o s s} = 1 - \mathrm {R M a s k} \quad \mathrm {I o U}. \tag {4} +$$ + +Adaptive Polygonal Sample Decision. The essence of sample decisions is to focus more attention on high-quality samples in each iteration. However, as an additional error introduced in Polar Representation, the representation error is susceptible to multiple factors, such as contour connectivity and convexity, which makes it challenging to be captured using the fixed priors. As shown in Fig. 2, our solution is to construct an optimal matching problem, in which positive labels and higher weights are assigned to those samples with lower matching costs. Specifically, we propose an Adaptive Polygonal Sample Decision (APSD) strategy to + +dynamically balance and minimize various types of errors across samples during training. For clarity, we divide this strategy into three stages: + +(1) Sampling Stage. In this stage, a group of candidates is sampled from the multi-scale grids of FPN layers. Here, Center Sampling is still adopted to maintain the strong receptive field for candidates, with a slight adjustment. To expand the range of candidate selection, we put forward Multi-layer Center Sampling (MCS), which applies Center Sampling across all FPN layers, rather than sampling with an increasing radius on a single layer (as proven ineffective in Sec. 4.3). +(2) Matching Stage. The purpose of this stage is to select positive or negative labels for candidates based on their matching costs. The Matching cost is defined as a weighted summation of classification loss $L_{cls}$ , regression loss $L_{reg}$ , and polygon loss $L_{poly}$ , which together evaluate the optimized cost from the prediction to ground truth. In this work, we employ Focal Loss [28], Polar IoU Loss [36], and the proposed RMask IoU Loss as classification loss, regression loss, and polygon loss, respectively. For a given instance, the matrix of matching costs $Cost = \{c_i | i = 1, \dots, n\}$ for $n$ candidates can be formulated as follows: + +$$ +C o s t = \lambda_ {c l s} L _ {c l s} + \lambda_ {r e g} L _ {r e g} + \lambda_ {p o l y} L _ {p o l y}, \tag {5} +$$ + +where $\lambda_{cls}$ , $\lambda_{reg}$ and $\lambda_{poly}$ stand for the loss coefficients. Afterward, a fixed number $k$ of positive labels are assigned to the candidates with minimal costs, and the others are negative: + +$$ +X _ {i} = \left\{ \begin{array}{l l} \text {P o s i t i v e}, & c _ {i} \in \operatorname {t o p} - k _ {\min } \left(c _ {1}, c _ {2}, \dots , c _ {n}\right) \\ \text {N e g a t i v e}, & c _ {i} \notin \operatorname {t o p} - k _ {\min } \left(c _ {1}, c _ {2}, \dots , c _ {n}\right), \end{array} \right. \tag {6} +$$ + +where $X_{i}$ refers to the label assigned to the $i$ -th candidates. Notably, if a sample is matched by two instances simultaneously, the one with the largest area will be chosen. + +(3) Weighting Stage. In the final stage, each positive sample is provided with a quality score, which modulates its contribution to loss functions. Our approach here is straightforward but effective, replacing the heuristic Centerness with the RMask IoU output by URM. Compared to Centerness, RMask IoU treats each positive sample equally, especially those high-quality samples located outside the contour, whose weights are zeroed out by Centerness (e.g., samples 5 and 6 in Fig. 3). + +Loss Function. The overall loss function $L$ consists of four parts: + +$$ +L = \lambda_ {c l s} L _ {c l s} + \lambda_ {r e g} L _ {r e g} + \lambda_ {p o l y} L _ {p o l y} + \lambda_ {s r e} L _ {s r e}, \tag {7} +$$ + +where $L_{sre}$ denotes the Cross-Entropy Loss in PolarMask used to supervise the Centerness head. In this case, we replace the original supervision signals with RMask IoU in the Centerness head. + +
TypeMethodBackboneEpochSize\(AP\)\(AP_{50}\)\(AP_{75}\)FPS\(Params\downarrow\)FLOPs↓
ContourDeepSnake [25]DLA-3416051230.3--24--
E2EC [39]DLA-3415051233.852.932.83529.80121
PolySnake [9]DLA-3425051234.4--18--
DANCE [24]R50-FPN3680036.858.539.01644.60274
MaskMask R-CNN [14]R50-FPN3680037.559.140.23344.40240
Cascade R-CNN [2]R50-FPN3680038.459.741.52377.331804
HTC [3]R50-FPN2080038.760.141.91780.261735
MS R-CNN [15]R50-FPN2480036.856.739.73060.74279
PointRend [17]R50-FPN3680038.459.541.52260.22266
CondInst [32]R50-FPN3680036.257.238.72934.17308
SOLO [34]R50-FPN3680036.457.938.92536.31332
SOLOv2 [35]R50-FPN3680038.459.541.23546.60227
YOLACT [1]R50-FPN5580029.048.729.84735.30237
SparseInst [7]R50-FPN14680033.253.434.75932.80201
PolarESE-Seg [38]R50-FPN30080021.648.722.438--
PolarMask [36]R50-FPN3680031.352.532.34234.74264
PolarNeXt (ours)R50-FPN3680036.159.737.34932.36186
+ +![](images/5fafdc340ced1c68fbe9a840eae35cf39a8373f68bec6931b3d7f4b3e01dfed9.jpg) +Figure 5. Comparison of Memory Usage (MB) for inference. During inference on COCO test-dev images, the peak memory usage is recorded as the final result of each method. + +# 4. Experiments + +# 4.1. Experiments and Implementation Details + +Experiment Settings. The experiments are conducted on MS COCO [21] benchmark and Cityscapes [8] dataset. MS COCO contains 80 classes with $115\mathrm{k}$ training, $5\mathrm{k}$ validation, and $20\mathrm{k}$ testing images. Cityscapes has 2795 training images from 8 classes and another 500 images with high-quality annotations for validation. In terms of evaluation metrics, the accuracy of segmentation is evaluated by the standard AP metric, denoted as mask AP. Furthermore, mask AP can be divided into $AP_{50} / AP_{75}$ based on different IoU thresholds, as well as $AP_{S} / AP_{M} / AP_{L}$ for small/medium/large objects. All inference experiments are conducted on a single NVIDIA 4090D GPU, where FPS, + +Table 1. Experimental results on COCO test-dev. "Size" refers to the length of the shortest edge of input images. ${}^{ \blacklozenge }$ means smaller values are preferred. + +
MethodAPAP50FPSMems↓
Mask R-CNN [14]31.558.6161968
SOLOv2 [35]27.248.5172076
CondInst [32]33.359.1152319
PolarMask [36]27.452.328575
PolarNeXt (ours)30.758.831566
+ +Table 2. Experimental results on Cityscapes val set. "Mems" stands for the memory usage (MB). + +Parms, and FLOPs are calculated separately for inference speed, model complexity, and computational overhead. Notably, in all our experiments, TensorRT or FP16 is not used for acceleration. + +Implementation Details. PolarNeXt is instantiated on the PolarMask network, using a backbone ResNet-50 [13] pretrained on ImageNet [29], and removed the auxiliary box branch. All models are implemented on the MMDetection toolbox [4], trained on 2 GPUs with 2 images per GPU, and optimized with SGD. Weight decay and momentum are configured to 0.0001 and 0.9, respectively. The base learning rate is set to 0.01 without any decay scheme. For a fair comparison, following common practice, data augmentation only contains random flip and scale jitter unless specified. In APSD, the radius of Center Sampling for candidate selection at each FPN layer is set to 1.5 and the number $k$ of assigned positive labels is set to 9 for each instance. Loss coefficients $\lambda_{cls}$ , $\lambda_{reg}$ , $\lambda_{poly}$ , and $\lambda_{sre}$ are all empirically set to 1.0. Following the rasterizer configuration in [18], we set the desired mask resolution to $64 \times 64$ and the rasterization sharpness to $\tau = 0.1$ . + +
URMAPSDAP
weightlosscost
29.1
✓30.3 (+1.2)
✓✓31.0 (+1.9)
✓28.6 (-0.5)
✓✓✓30.7 (+1.6)
✓✓✓✓33.9 (+4.8)
+ +Table 3. Ablation studies on URM and APSD. "weight" means whether Centerness is replaced by RMask IoU for sample weighting. "loss" refers to the use of RMask IoU Loss in loss functions. "cost" indicates the incorporation of RMask IoU Loss into matching costs. + +
Polar IoU LossAPAPSAPMAPL
w/o.32.814.335.148.5
w/.33.916.435.949.4
+ +# 4.2. Main Results + +MS COCO. Following common practice, we evaluate PolarNeXt on the MS COCO benchmark and compare test-dev results to some state-of-the-art models in Tab. 1. In terms of computational cost, PolarNeXt is more lightweight than the other models with R50-FPN backbone, only requiring 32.8 MB Parameters and 186 GB FLOPs. Even when compared to the contour-based methods using DLA-34 backbone, it remains on par. In terms of inference speed, PolarNeXt is second only to the prevalent SpareInst, but it achieves higher accuracy (36.1% vs. 33.2% AP) with a shorter training schedule (36 vs. 146 epochs). In terms of segmentation accuracy, PolarNeXt delivers competitive results compared to other more complex methods. Particularly, without bells and whistles, PolarNeXt achieves a 4.8% mask AP improvement over PolarMask, redemonstrating the potential of Polar Representation in instance segmentation. Moreover, we visualize the memory usage of some representative methods during inference in Fig. 5. Obviously, PolarNeXt shows its consistency with the classical object detector FCOS, with memory usage significantly lower than other instance segmentation methods, only half or even less. + +Cityscapes. As an extended comparative experiment in Tab. 2, Cityscapes dataset, characterized by high-resolution images, is utilized to examine the robustness and generalization ability of our model. Notably, all models are directly trained on the Cityscapes training set under 64 epochs, without COCO initialization. Since fragmented instances frequently occur in Cityscapes, the Multi-component Detection Strategy suggested in [25] is applied to PolarNeXt and + +Table 4. Feasibility analysis of removing Polar IoU Loss. The symbols "w/o." and "w/." indicate whether the Polar IoU Loss is employed. + +
Candidate SelectionLayerAPT.T.
Center Sampling (s.r.=1.5)single31.013h
Center Sampling (s.r.=2.0)30.6 (-0.4)14h
All Grid Pointsmulti--
Points within Boxes33.1 (+2.1)28h
Points within Contours33.0 (+2.0)25h
Center Sampling (s.r.=1.5)33.9 (+2.9)16h
+ +Table 5. The effectiveness of some approaches for candidate selection. "single" denotes selecting candidates on a specific FPN layer based on the instance size, and "multi" refers to selecting candidates across all FPN layers. "T.T." and "s.r." are abbreviations for training time and sampling radius, respectively. + +
MethodAPbAP50APSbAPMAPLb
PolarMask16.338.316.028.413.2
PolarNeXt19.846.720.232.215.9
+ +Table 6. Comparison of boundary quality between PolarMask and PolarNeXt. + +PolarMask, leading to approximately a $1.5\%$ AP improvement. Under the increased resolution of input images (from $800 \times 1333$ to $1024 \times 2048$ ), mask-based methods suffer from the heavy memory usage introduced by dense pixel classification, which is nearly four times greater than that of polar-based methods. Meanwhile, the speed advantage of polar-based methods becomes more apparent, requiring only half the inference time compared to mask-based methods. As expected, PolarNeXt maintains a $3.3\%$ AP improvement over PolarMask, further demonstrating the effectiveness of our proposed method. + +# 4.3. Ablation Studies + +In this section, we conduct ablations to assess the influence of the proposed components on instance segmentation performance, and then analyze their details in the following. All experiments are conducted on MS COCO val under a 1x training schedule. + +Component Impact and Correlation. The effectiveness of URM and APSD is reported in Tab. 3. Using URM alone brings a $1.9\%$ AP improvement on the baseline, while combining it with APSD further boosts the performance by an additional $2.9\%$ AP. It is worth noting that APSD fails to function independently of URM, the absence of which results in a decrease in performance. To be specific, without APSD, URM contributes to the sample weighting and loss function, leading to a $1.2\%$ and $0.7\%$ AP increase, respectively. Conversely, when APSD is applied, the matching costs have to incorporate RMask IoU Loss; otherwise, this strategy will fail or even backfire. In view of this, we conclude that the effectiveness of our adaptive strategy hinges on reliable polygonal assessment for support. + +![](images/4420ab76d06333c0ef197cf47c546769c0f1bb98d8ecf44aac60ead69e5361c1.jpg) +Figure 6. Visualization with Rays for PolarMask and PolarNeXt. + +Candidate Selection. In Tab. 5, we compare some approaches to expand the range of candidate selection. On the one hand, Center Sampling on multiple FPN layers is more effective and efficient than the dense point selection within boxes or contours. In particular, the approach suggested in [31], which selects all grid points of FPN, results in unacceptable training time and GPU memory overflow. On the other hand, directly expanding the radius of Center Sampling on a single layer causes a $0.4\%$ drop in AP, as the additional selected samples tend to fall outside the receptive field. Accordingly, the center prior still has a positive influence on sample decisions. + +RMask IoU vs. Polar IoU. In Fig. 4, two scatter plots are drawn to compare Polar IoU and RMask IoU in polygonal assessment. These scatter points denote all the polygons predicted during inference on MS COCO val images. The fitted curve in Fig. 4(b) is almost aligned with the ideal curve, whereas the one in Fig. 4(a) shows significant deviation. This indicates that RMask IoU better reflects the actual polygonal quality compared to Polar IoU. + +The necessity of Polar IoU Loss. In Tab. 4, we explore the feasibility of directly removing Polar IoU Loss in our proposed training pipeline. Experimental results indicate that this operation results in a $1.1\%$ AP degradation in performance. The impact on small objects is more notable, with $AP_{S}$ decreased by $2.1\%$ . We analyze that there may be a slight granularity missing as well as feature misalignment during the conversion from vertex sets to rasterized masks. Moreover, as mentioned in [18], aliasing occurs in the deconvolutional process of FPN when pre-predicted RoIs are unavailable in one-stage detectors. Fortunately, the combination of Polar IoU Loss and RMask IoU Loss brings a notable boost, with the two complementing each other. We believe that Polar IoU Loss serves as a polygonal internal constraint for dense distance regression, while RMask IoU + +Loss provides a polygonal external constraint to incorporate the representation deviation. + +# 4.4. Boundary Quality Analysis + +In Tab. 6, we introduce Boundary AP [5] ( $AP^b$ for short) to evaluate the boundary quality of predicted polygons. Experimental results show that PolarNeXt achieves a significant improvement in boundary quality compared to PolarMask, boosting $AP^b$ by $3.5\%$ . Furthermore, we present some segmentation results with differences in starting points in Fig. 6. Obviously, compared to the center-fixed starting points in PolarMask, the dynamically selected starting points in PolarNeXt are more flexible, which enhances the representation capability of bounding polygons. + +# 5. Conclusion + +In this paper, we introduce a novel polar-based framework, PolarNeXt, which redemonstrates the superiority of Polar Representation in instance segmentation. To account for the previously ignored representation error, tailored training strategy APSD and assessment module URM are proposed. Together, these components significantly enhance PolarNeXt, yielding notable performance improvements over other polar-based methods. By fully leveraging the advantages of Polar Representation, PolarNeXt achieves comparable segmentation accuracy while requiring only half or even less of the computational cost and inference time compared to mainstream mask-based and contour-based methods. We anticipate that PolarNeXt will set a new standard for polar-based methods in instance segmentation. + +Acknowledgments. This work is supported by the National Natural Science Foundation of China [Grant No. 62225308], and the Shanghai Technical Service Computing Center of Science and Engineering, Shanghai University. + +# References + +[1] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, pages 9157-9166, 2019. 3, 6 +[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE TPAMI, 43(5):1483-1498, 2019. 3, 6 +[3] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In CVPR, pages 4974-4983, 2019. 3, 6 +[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6 +[5] Bowen Cheng, Ross Girshick, Piotr Dólar, Alexander C Berg, and Alexander Kirillov. Boundary iou: Improving object-centric image segmentation evaluation. In CVPR, pages 15334-15342, 2021. 8 +[6] Dominic Cheng, Renjie Liao, Sanja Fidler, and Raquel Urtasun. Darnet: Deep active ray network for building segmentation. In CVPR, pages 7431-7439, 2019. 3 +[7] Tianheng Cheng, Xinggang Wang, Shaoyu Chen, Wenqiang Zhang, Qian Zhang, Chang Huang, Zhaoxiang Zhang, and Wenyu Liu. Sparse instance activation for real-time instance segmentation. In CVPR, pages 4433-4442, 2022. 1, 3, 6 +[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6 +[9] Hao Feng, Keyi Zhou, Wengang Zhou, Yufei Yin, Jiajun Deng, Qi Sun, and Houqiang Li. Recurrent generic contour-based instance segmentation with progressive learning. IEEE TCSVT, 2024. 1, 3, 6 +[10] Gunther Greiner and Kai Hormann. Efficient clipping of arbitrary polygons. ACM Transactions on Graphics (TOG), 17 (2):71-83, 1998. 4 +[11] Wenchao Gu, Shuang Bai, and Lingxing Kong. A review on 2d instance segmentation based on deep neural networks. Image and Vision Computing, 120:104401, 2022. 1, 3 +[12] Shir Gur, Tal Shaharabany, and Lior Wolf. End to end trainable active contours via differentiable rendering. arXiv preprint arXiv:1912.00367, 2019. 4 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6 +[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3, 6 +[15] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In CVPR, pages 6409-6418, 2019. 3, 6 +[16] Michael Kass, Andrew Witkin, and Demetri Terzopoulos. Snakes: Active contour models. IJCV, 1(4):321-331, 1988. 3 + +[17] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In CVPR, pages 9799-9808, 2020. 3, 6 +[18] Justin Lazarow, Weijian Xu, and Zhuowen Tu. Instance segmentation with mask-supervised polygonal boundary transformers. In CVPR, pages 4382-4391, 2022. 4, 5, 6, 8 +[19] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. NeurIPS, 33:21002-21012, 2020. 3 +[20] Justin Liang, Namdar Homayounfar, Wei-Chiu Ma, Yuwen Xiong, Rui Hu, and Raquel Urtasun. Polytransform: Deep polygon transformer for instance segmentation. In CVPR, pages 9131-9140, 2020. 3 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6 +[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 3 +[23] Huan Ling, Jun Gao, Amlan Kar, Wenzheng Chen, and Sanja Fidler. Fast interactive object annotation with curve-gcn. In CVPR, pages 5257-5266, 2019. 4 +[24] Zichen Liu, Jun Hao Liew, Xiangyu Chen, and Jiashi Feng. Dance: A deep attentive contour model for efficient instance segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 345-354, 2021. 3, 6 +[25] Sida Peng, Wen Jiang, Huaijin Pi, Xiuli Li, Hujun Bao, and Xiaowei Zhou. Deep snake for real-time instance segmentation. In CVPR, pages 8533-8542, 2020. 1, 3, 6, 7 +[26] Joseph Redmon. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 1 +[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2016. 1, 3 +[28] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 5 +[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 6 +[30] Uwe Schmidt, Martin Weigert, Coleman Broaddus, and Gene Myers. Cell detection with star-convex polygons. In Medical Image Computing and Computer Assisted Intervention, pages 265–273, 2018. 3 +[31] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In International Conference on Machine Learning, pages 9934-9944, 2021. 8 +[32] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, pages 282-298, 2020. 3, 6 + +[33] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: A simple and strong anchor-free object detector. IEEE TPAMI, 44(4):1922-1933, 2020. 1 +[34] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, pages 649-665, 2020. 1, 3, 6 +[35] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmentation. NeurIPS, 33:17721-17732, 2020. 3, 6 +[36] Enze Xie, Peize Sun, Xiaoge Song, Wenhai Wang, Xuebo Liu, Ding Liang, Chunhua Shen, and Ping Luo. Polarmask: Single shot instance segmentation with polar representation. In CVPR, pages 12193-12202, 2020. 1, 2, 3, 5, 6 +[37] Enze Xie, Wenhai Wang, Mingyu Ding, Ruimao Zhang, and Ping Luo. Polarmask++: Enhanced polar representation for single-shot instance segmentation and beyond. IEEE TPAMI, 44(9):5385-5400, 2021. 3 +[38] Wenqiang Xu, Haiyang Wang, Fubo Qi, and Cewu Lu. Explicit shape encoding for real-time instance segmentation. In ICCV, pages 5168-5177, 2019. 1, 3, 6 +[39] Tao Zhang, Shiqing Wei, and Shunping Ji. E2ec: An end-to-end contour-based method for high-quality high-speed instance segmentation. In CVPR, pages 4443-4452, 2022. 1, 3, 6 \ No newline at end of file diff --git a/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/images.zip b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3451dacec9e3559105ab7844347fc09d933897f0 --- /dev/null +++ b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7c71aac3c49897b0fb9485f689a92b14c5a49c86f68cf4b55eb83db721dcc11 +size 712309 diff --git a/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/layout.json b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3403403869e83b65f5d05109f252359f0a6a8744 --- /dev/null +++ b/2025/PolarNeXt_ Rethink Instance Segmentation with Polar Representation/layout.json @@ -0,0 +1,7436 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "type": "text", + "content": "PolarNeXt: Rethink Instance Segmentation with Polar Representation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "text", + "content": "Jiacheng Sun" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "text", + "content": ", Xinghong Zhou" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "text", + "content": ", Yiqiang Wu" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "text", + "content": ", Bin Zhu" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "text", + "content": ", Jiaxuan Lu" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "text", + "content": ", Yu Qin" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "text", + "content": ", Xiaomao Li" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "text", + "content": "Shanghai University, " + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 142, + 536, + 171 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 173, + 480, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 173, + 480, + 185 + ], + "spans": [ + { + "bbox": [ + 129, + 173, + 480, + 185 + ], + "type": "text", + "content": "{sunjc, jkchou, yiqiangwu, zhubin01, qy21722148, lixiaomao}@shu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 243, + 187, + 364, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 187, + 364, + 198 + ], + "spans": [ + { + "bbox": [ + 243, + 187, + 364, + 198 + ], + "type": "text", + "content": "lujiaxuan@pjlab.org.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 252, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 552 + ], + "type": "text", + "content": "One of the roadblocks for instance segmentation today is heavy computational overhead and model parameters. Previous methods based on Polar Representation made the initial mark to address this challenge by formulating instance segmentation as polygon detection, but failed to align with mainstream methods in performance. In this paper, we highlight that Representation Errors, arising from the limited capacity of polygons to capture boundary details, have long been overlooked, which results in severe performance degradation. Observing that optimal starting point selection effectively alleviates this issue, we propose an Adaptive Polygonal Sample Decision strategy to dynamically capture the positional variation of representation errors across samples. Additionally, we design a Union-aligned Rasterization Module to incorporate these errors into polygonal assessment, further advancing the proposed strategy. With these components, our framework PolarNeXt achieves a remarkable performance boost of over " + }, + { + "bbox": [ + 55, + 252, + 296, + 552 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 55, + 252, + 296, + 552 + ], + "type": "text", + "content": " AP compared to other polar-based methods. PolarNeXt is markedly more lightweight and efficient than state-of-the-art instance segmentation methods, while achieving comparable segmentation accuracy. We expect this work will open up a new direction for instance segmentation in high-resolution images and resource-limited scenarios. Codes can be found at https://github.com/Sun15194/PolarNeXt." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 597, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 693 + ], + "type": "text", + "content": "Object detection and instance segmentation are fundamental tasks in computer vision, aiming at identifying specific objects within images. Object detection [26, 27, 33], also known as box detection, localizes objects using rectangular boxes, which are efficient but coarse-grained. In comparison, instance segmentation offers additional geometric information, such as shape and boundary. Mainstream methods define instance segmentation as pixel-wise classifica" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 225, + 553, + 378 + ], + "blocks": [ + { + "bbox": [ + 316, + 225, + 553, + 378 + ], + "lines": [ + { + "bbox": [ + 316, + 225, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 316, + 225, + 553, + 378 + ], + "type": "image", + "image_path": "18c0e919d50ff766d9db2bba7da3d8be3142a08044d4b12e0381fce0d959959c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 386, + 555, + 497 + ], + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 497 + ], + "type": "text", + "content": "Figure 1. Illustration of the composition of prediction errors in polar-based methods. Classification and Regression (Cls.&Reg.) Errors are produced during network optimization, which reflects the deviation from starting point classification " + }, + { + "bbox": [ + 313, + 386, + 555, + 497 + ], + "type": "inline_equation", + "content": "(s\\rightarrow \\hat{s})" + }, + { + "bbox": [ + 313, + 386, + 555, + 497 + ], + "type": "text", + "content": " and dense distance regression " + }, + { + "bbox": [ + 313, + 386, + 555, + 497 + ], + "type": "inline_equation", + "content": "(D\\to \\hat{D})" + }, + { + "bbox": [ + 313, + 386, + 555, + 497 + ], + "type": "text", + "content": ". The ignored Representation (Rep.) Error arises from the missing boundary details of instance contour " + }, + { + "bbox": [ + 313, + 386, + 555, + 497 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 386, + 555, + 497 + ], + "type": "text", + "content": " in Polar Representation, as exemplified by the two cases below. Obviously, the center point does not necessarily coincide with the optimal point, from which the assembled polygon with minimal Rep. Error can be predicted." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 569 + ], + "type": "text", + "content": "tion [7, 14, 34] or iterative deformation [9, 25, 39]. Despite the finer granularity, these methods introduce heavy computational overhead and model parameters, which restricts their applicability in high-resolution images or resource-limited scenarios [11]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "To combine the efficiency of detection with the granularity of segmentation, a pioneering attempt is Polar Representation. As shown at the top of Fig. 1, Polar Representation reconstructs the instance contour into a bounding polygon, assembled by dense rays emitted uniformly from a starting point. Building on this concept, typical polar-based methods [36-38] formulate instance segmentation as starting point classification and dense distance regression. These methods can be instantiated on one-stage object detectors [26, 33], directly inheriting their advantages in inference speed and model complexity. However, Polar Representation has not become a fundamental paradigm for in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19315" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 547, + 266 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 547, + 266 + ], + "type": "image", + "image_path": "f2291bc6b548b54c582c4e8b602cc289b04fc1d80ccd807848f77bf06bc006a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 276, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 54, + 276, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 276, + 555, + 323 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed Adaptive Polygonal Sample Decision (APSD). The three stages of APSD are labeled as: I.Sampling Stage, II.Matching Stage, and III.Weighting Stage. The Union-aligned Rasterization Module (URM) takes the predicted polygons " + }, + { + "bbox": [ + 54, + 276, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 54, + 276, + 555, + 323 + ], + "type": "text", + "content": " of the candidates selected by Multi-layer Center Sampling (MCS) as input, and outputs their RMask IoU with the instance contours " + }, + { + "bbox": [ + 54, + 276, + 555, + 323 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 276, + 555, + 323 + ], + "type": "text", + "content": ". RMask IoU functions in both matching costs and sample weighting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 334, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 334, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 334, + 295, + 357 + ], + "type": "text", + "content": "stance segmentation, due to the significant performance gap between polar-based methods and other advanced methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 361, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 649 + ], + "type": "text", + "content": "In this work, we highlight that an additional prediction error has been ignored in Polar Representation, which results in the performance degradation of polar-based methods. As illustrated at the bottom of Fig. 1, from different starting points, the assembled polygons exhibit varying degrees of deviation with their corresponding contours. This deviation, termed Representation (Rep.) Error, inevitably occurs in Polar Representation, where the boundary details of contours are complex for polygons to reflect, especially in disconnected or concave regions. However, existing training pipelines in polar-based methods only account for the Classification and Regression (Cls.&Reg.) Errors generated during network optimization. Their label assignment and sample weighting, collectively referred to as sample decisions, rely entirely on the handcrafted center prior, which fails to capture the positional variation of representation errors (as proven in Sec. 3.2(1)). Furthermore, even though a natural idea is to dynamically select the optimal starting point with minimal representation error, a reliable metric to incorporate this error into polygonal assessment is still lacking. As the only available metric, Polar IoU [36] exhibits assessment blindness to representation errors (as verified in Sec. 3.2(2)), which is too coarse-grained to support this idea." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "To overcome these challenges, we propose PolarNeXt, a highly lightweight yet effective polar-based framework for real-time instance segmentation. The training pipeline of this framework includes an Adaptive Polygonal Sample Decision (APSD) strategy, enabling optimal starting point" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 333, + 555, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 585 + ], + "type": "text", + "content": "selection during inference. APSD assigns positive labels and higher weights to the candidate samples based on their matching costs, which account for the comprehensive errors between predictions and ground truths. To eliminate the assessment blindness to representation errors, a Union-aligned Rasterization Module (URM) is designed for more accurate polygonal assessment. The paired polygon and contour are aligned within their Union Box and then converted into rasterized masks for RMask IoU measurement. Here, RMask IoU serves as a new metric to evaluate the actual polygonal overlap, forming the core element of sample decisions and loss functions. Benefitting from these components, PolarNeXt achieves a remarkable improvement of over " + }, + { + "bbox": [ + 313, + 333, + 555, + 585 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 313, + 333, + 555, + 585 + ], + "type": "text", + "content": " AP compared to other polar-based methods. To the best of our knowledge, this is the first polar-based method competitive with state-of-the-art instance segmentation methods in performance. Notably, PolarNeXt requires only half or even less of their computational cost and inference time, which redemonstrates the superiority of Polar Representation in instance segmentation. The main contributions of this work can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 594, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 594, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 555, + 653 + ], + "type": "text", + "content": "- We conduct a thorough analysis of the implementation challenges of Polar Representation in instance segmentation, and reveal that the performance limitations of existing polar-based methods stem from their neglect of the representation error." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 654, + 556, + 713 + ], + "type": "text", + "content": "- We propose a new polar-based instance segmentation framework called PolarNeXt, including an APSD strategy to dynamically account for representation errors across samples and a URM module that eliminates the assessment blindness to these errors using RMask IoU." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19316" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "content": "- We fully leverage the advantages of Polar Representation in instance segmentation tasks. Without bells and whistles, PolarNeXt is markedly more lightweight and efficient than other mainstream methods, while aligning with them in terms of segmentation performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 148, + 142, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 142, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 142, + 162 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 297, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 297, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 297, + 602 + ], + "type": "text", + "content": "Instance Segmentation. Mainstream instance segmentation methods can be classified into two categories based on their representation, i.e., mask-based methods and contour-based methods. Mask-based methods treat instance segmentation as pixel-wise classification, in which binary masks are predicted within their Regions of Interest (RoIs). As a pioneering work, Mask R-CNN [14] introduces a detect-then-segment paradigm, built on the two-stage detector [27], with box proposals serving as the RoIs for pixel-by-pixel segmentation. Nowadays, state-of-the-art benchmark results are still held by the mask-based two-stage methods [2, 3, 15] constructed on strong detection baselines. Additionally, to reduce reliance on bounding boxes, one-stage methods [32, 34, 35] treat the entire image as the RoI, predicting masks by the assembly of encoded mask vectors and mask kernels from dynamic convolutions. Other mask-based methods focus on the high-quality boundary [17, 20], or the real-time inference [1, 7]. Contour-based methods take a different paradigm, formulating instance segmentation as iterative deformation based on the active contour mechanism of Snake [16]. Deep Snake [25] establishes the first multi-stage deformation network, which progressively refines the pre-predicted initial contours using their vertex features. Based on this baseline, further optimization strategies have been explored, such as Segment-wise Matching Scheme [24], Dynamic matching Loss [39], and Multi-scale Contour Refinement [9]. However, the high computational cost introduced by pixel-wise classification and iterative deformation limits the applicability and prevalence of these methods in high-resolution images or resource-limited scenarios [11]. To solve this problem, we bring back Polar Representation and redemonstrate its superiority in instance segmentation. Building on Polar Representation, our proposed PolarNeXt is markedly more lightweight and efficient than the aforementioned methods, while remaining competitive in segmentation performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Polar Representation. Polar Representation initially appears in some interdisciplinary fields, such as automatic building segmentation [6] and medical cell detection [30]. Subsequently, ESE-Seg [38] migrates Polar Representation into instance segmentation tasks, introducing Chebyshev polynomial fitting to shorten the distance vectors and resist noise. PolarMask [36] and its variant [37] make the initial mark of polar-based methods on instance segmentation benchmarks. They maximize the advantage of the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "center prior by modified Center Sampling and Polar Centerness, along with a proposed Polar IoU metric for polygonal assessment. However, the representation error introduced by Polar Representation has been ignored in these polar-based methods, which results in severe performance degradation. To address this, we propose an APSD strategy to dynamically balance representation errors across samples, in conjunction with a URM module that incorporates these errors into polygonal assessment. Together, these components bring a notable performance boost to PolarNeXt." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 206, + 375, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 375, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 375, + 219 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 228, + 556, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 556, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 556, + 324 + ], + "type": "text", + "content": "In this work, PolarMask [36], a representative polar-based method, is used as a case study for Polar Representation. The design principle and training pipeline of PolarMask are reviewed in Sec. 3.1. Under this pipeline, the effectiveness of its sample decisions and polygonal assessment is investigated in Sec. 3.2. Finally, a novel polar-based framework called PolarNeXt is detailed in Sec. 3.3 to advance Polar Representation in instance segmentation tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 334, + 394, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 394, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 394, + 347 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": "PolarMask defines instance segmentation as starting point classification and dense distance regression. As illustrated in Fig. 1, for a given instance contour " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": ", bounding polygon " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "P = \\psi(s, D)" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": " can be constructed through Polar Representation " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": ", using a distance set " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "D = \\{d_i | i = 1, \\dots, m\\}" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": " from a starting point " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": ". The starting point " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": " and distance set " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": " serve as supervision signals for the network, which then classifies a starting point " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": " and regresses an equal-size set " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\hat{D} = \\{\\hat{d}_i | i = 1, \\dots, m\\}" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": " to construct predicted polygon " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\hat{P} = \\psi(\\hat{s}, \\hat{D})" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": ". In the training pipeline of PolarMask, the center prior fully dictates label assignment and sample weighting, given that samples closer to the center tend to have richer receptive field and lower regression difficulty [19]. First, Center Sampling is applied on a single FPN [22] layer, directly assigning positive labels to the samples around instance centers. Second, Polar Centerness, which quantifies the proximity to the center, weights these positive samples to modulate their contribution to loss functions. Furthermore, due to the lack of polygon-specific IoUs, Polar IoU is proposed as a metric for polygonal assessment. It compares the consistency between target distance set " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": " and predicted distance set " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": " to approximate the overlap between " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 313, + 354, + 556, + 661 + ], + "type": "text", + "content": ". Additionally, Polar IoU is transformed into Polar IoU Loss to supervise distance regression. More details are provided in supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 670, + 398, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 398, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 398, + 684 + ], + "type": "text", + "content": "3.2. Investigation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "In this section, we identify that two challenges lie in implementing Polar Representation for instance segmentation:" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19317" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 69, + 290, + 199 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 290, + 199 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 290, + 199 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 290, + 199 + ], + "type": "image", + "image_path": "80afa08be13c791a94722d44f607be8f6a30e6f2ace8684d0fef8a67ed6c6f03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 206, + 296, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 283 + ], + "type": "text", + "content": "Figure 3. Illustration of the divergence between Polarness and Centerness. On the left, heatmaps of these two attributes are visualized to reveal the distributional divergence, with positive samples (9 white points) selected by Center Sampling marked. Then, attribute values of these samples are plotted to verify the quantitative divergence on the right, where samples are sorted by ascending Polarness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 301, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 295, + 552 + ], + "type": "text", + "content": "(1) Diminished Effectiveness of Center Prior. The center prior plays a diminished role in sample decisions, which fails to capture the positional variation of representation errors. To validate this, a comparative experiment is conducted in Fig. 3, illustrating the divergence between the center prior and representation error. For a clear comparison, we define Polarness " + }, + { + "bbox": [ + 56, + 301, + 295, + 552 + ], + "type": "inline_equation", + "content": "(Plr)" + }, + { + "bbox": [ + 56, + 301, + 295, + 552 + ], + "type": "text", + "content": " as an inverse measure of Representation Error " + }, + { + "bbox": [ + 56, + 301, + 295, + 552 + ], + "type": "inline_equation", + "content": "(RE)" + }, + { + "bbox": [ + 56, + 301, + 295, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 301, + 295, + 552 + ], + "type": "inline_equation", + "content": "Plr = 1 - RE" + }, + { + "bbox": [ + 56, + 301, + 295, + 552 + ], + "type": "text", + "content": ". Likewise, normalized Polar Centerness is introduced to quantify the influence of the center prior. Experimental results show that Polarness presents no observable correlation with the distribution pattern of Centerness, while Centerness does not increase monotonically on these samples sorted by ascending Polarness. Accordingly, the samples selected by center sampling do not necessarily correspond to lower representation errors, and Centerness also fails to accurately modulate the weights of high-quality samples. Based on these observations, we conclude that the sample decisions entirely reliant on the center prior are less applicable in polar-based methods, as some suboptimal samples in center regions receive excessive attention." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "(2) Assessment Blindness of Polar IoU. As the only available metric, Polar IoU exhibits assessment blindness to representation errors. This limitation arises because Polar IoU evaluates the consistency between target polygon " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " and predicted polygon " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " by comparing their distance sets, entirely independent of instance contour " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": ". To explore the impact of this blindness on polygonal assessment, a scatter plot is drawn to examine the effectiveness of Polar IoU. As shown in Fig. 4(a), the scatter points show a significant departure from the expected linear trend, with substantial noise introduced. For example, when Real IoU reaches 0.4, Polar IoU fluctuates substantially, ranging irregularly from 0.4 to 0.9. Based on these observations, we conclude that" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 555, + 198 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 555, + 198 + ], + "type": "image", + "image_path": "f13d06d9cb4a848b9b7427846efb9181be1fb12ecdfa83c61d5ee4d60d0072da.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 206, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 313, + 206, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 555, + 273 + ], + "type": "text", + "content": "Figure 4. Comparison between Polar IoU and our proposed RMask IoU. The green solid lines denote the curves fitted by scatter points, while the red dashed lines indicate the desired linear correlation trend. Real IoU is the polygon-specific IoU between the predicted polygon and its corresponding instance contour, calculated using Weiler-Atherton algorithm [10]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 287, + 555, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 310 + ], + "type": "text", + "content": "Polar IoU fails to function as a reliable metric for polygonal assessment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 320, + 387, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 320, + 387, + 332 + ], + "spans": [ + { + "bbox": [ + 314, + 320, + 387, + 332 + ], + "type": "text", + "content": "3.3. PolarNeXt" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 339, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 553 + ], + "type": "text", + "content": "To overcome these two challenges, a new framework named PolarNeXt is proposed, which is expected to become the next level of polar-based methods for instance segmentation. The training pipeline of PolarNeXt includes two tailored components for Polar Representation: Adaptive Polygonal Sample Decision (APSD) and Union-aligned Rasterization Module (URM). As illustrated in Fig. 2, APSD samples a wide range of candidates around instances on the input image. Then, those candidates with minimal matching costs are assigned positive labels and weighted by the quality scores of their predicted polygons. For polygonal assessment, URM aligns the paired polygon and contour within their Union Box, and converts them into rasterized masks for overlap measurement. The RMask IoU output by URM incorporates representation errors, serving as precise guidance in both label assignment and sample weighting. The detailed design principles of our proposed methods are outlined below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Union-aligned Rasterization Module. To eliminate the assessment blindness to representation errors, a straightforward idea is to bridge the predicted polygon " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " with its instance contour " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " and measure their polygonal overlap. However, traditional polygon-specific IoU algorithms, such as Weiler-Atherton [10], are impractical for parallel processing, due to uncertain vertex counts and inevitable recursive operations. Inspired by 3D renderers [12, 18, 23], our solution is to convert " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " from vertex-set format into rasterized-mask format for mask-level overlap measurement, but an additional challenge lies in their alignment at a fixed resolution, given that no pre-predicted RoIs are available in one-stage detectors." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19318" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "content": "As shown in Fig. 2, we propose a Union-aligned Rasterization Module (URM) which employs Union Box, the minimal bounding box enclosing all the vertices of " + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "content": ", as a reference for coordinate alignment. Concretely, assuming the vertex sets of " + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "content": " are symbolized by " + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "inline_equation", + "content": "V_{\\hat{P}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "inline_equation", + "content": "V_{C}" + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "content": ", with a desired mask resolution for rasterization specified as " + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "content": ". The coordinate " + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "inline_equation", + "content": "(v_x,v_y)" + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "content": " of a vertex " + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "content": " can be adjusted by the union set " + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "inline_equation", + "content": "U = V_{\\hat{P}}\\cup V_{C}" + }, + { + "bbox": [ + 55, + 72, + 296, + 169 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 177, + 296, + 231 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 177, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 93, + 177, + 296, + 231 + ], + "type": "interline_equation", + "content": "\\left( \\begin{array}{c} v _ {x} \\\\ v _ {y} \\end{array} \\right) = \\left( \\begin{array}{c} \\frac {v _ {x} - \\underset {u \\in U} {\\min } u _ {x}}{\\underset {u \\in U} {\\max } u _ {x} - \\underset {u \\in U} {\\min } u _ {x}} \\times W \\\\ \\frac {v _ {y} - \\underset {u \\in U} {\\min } u _ {y}}{\\underset {u \\in U} {\\max } u _ {y} - \\underset {u \\in U} {\\min } u _ {y}} \\times H \\end{array} \\right). \\tag {1}", + "image_path": "a3c99656b259f14046189cbaf640b57cda1f6bc39dcca17fc7c7ca2274d6a614.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": "Then, the differentiable rasterizer in [18] is introduced to transform the aligned vertex sets into rasterized masks. For pixel " + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "inline_equation", + "content": "p(x,y)" + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": " of a mask, where " + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "inline_equation", + "content": "0 \\leq x \\leq W" + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "inline_equation", + "content": "0 \\leq y \\leq H" + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": ", two attributes are used to model its contribution value " + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "inline_equation", + "content": "M[x,y]" + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": " to the mask: (1) the binary inside-outside state " + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "inline_equation", + "content": "IO(V,p)" + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": " of pixel " + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": " relative to vertex set " + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": ", and (2) the closest distance " + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "inline_equation", + "content": "CD(V,p)" + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": " from pixel " + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": " to vertex set " + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": ". The contribution value " + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "inline_equation", + "content": "M[x,y]" + }, + { + "bbox": [ + 55, + 238, + 296, + 334 + ], + "type": "text", + "content": " is formulated as shown below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 341, + 295, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 341, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 83, + 341, + 295, + 365 + ], + "type": "interline_equation", + "content": "M [ x, y ] = \\operatorname {s i g m o i d} \\left(\\frac {I O (V , p) \\times C D (V , p)}{\\tau}\\right), \\tag {2}", + "image_path": "c3e0f10a8ea6e6ac5c5289e220c830b21b2dad2fdbbd583c8f994eb2f0de58e7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 420 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 372, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 372, + 296, + 420 + ], + "type": "text", + "content": " is a hyperparameter for sharpness. Finally, two rasterized masks " + }, + { + "bbox": [ + 55, + 372, + 296, + 420 + ], + "type": "inline_equation", + "content": "M_{\\hat{P}}" + }, + { + "bbox": [ + 55, + 372, + 296, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 372, + 296, + 420 + ], + "type": "inline_equation", + "content": "M_C" + }, + { + "bbox": [ + 55, + 372, + 296, + 420 + ], + "type": "text", + "content": " with the same resolution " + }, + { + "bbox": [ + 55, + 372, + 296, + 420 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 55, + 372, + 296, + 420 + ], + "type": "text", + "content": " are produced, and their mask-level IoU, called RMask IoU, can be calculated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 427, + 295, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 427, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 67, + 427, + 295, + 484 + ], + "type": "interline_equation", + "content": "\\mathrm {R M a s k} \\mathrm {I o U} = \\frac {2 \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\mathrm {M} _ {C} [ \\mathrm {i} , \\mathrm {j} ] \\cdot \\mathrm {M} _ {\\hat {P}} [ \\mathrm {i} , \\mathrm {j} ]}{\\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\mathrm {M} _ {C} [ \\mathrm {i} , \\mathrm {j} ] + \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} \\mathrm {M} _ {\\hat {P}} [ \\mathrm {i} , \\mathrm {j} ]}. \\tag {3}", + "image_path": "2238e37acbbc81c5c9c6b483598799dbcd03fef42dc401730baea95bac07c49c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": "As shown in Fig. 4(b), RMask IoU accurately reflects the actual overlap between polygons and contours. Further, we perform a simple deformation on RMask IoU to obtain a new loss function, RMask IoU Loss, for polygonal external constraints:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 560, + 295, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 560, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 97, + 560, + 295, + 572 + ], + "type": "interline_equation", + "content": "\\mathrm {R M a s k} \\quad \\mathrm {I o U} \\quad \\text {L o s s} = 1 - \\mathrm {R M a s k} \\quad \\mathrm {I o U}. \\tag {4}", + "image_path": "94d6645a287fec7d89d085c7529d54c84cc0ec643c081843f86481ca711d92ad.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Adaptive Polygonal Sample Decision. The essence of sample decisions is to focus more attention on high-quality samples in each iteration. However, as an additional error introduced in Polar Representation, the representation error is susceptible to multiple factors, such as contour connectivity and convexity, which makes it challenging to be captured using the fixed priors. As shown in Fig. 2, our solution is to construct an optimal matching problem, in which positive labels and higher weights are assigned to those samples with lower matching costs. Specifically, we propose an Adaptive Polygonal Sample Decision (APSD) strategy to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "dynamically balance and minimize various types of errors across samples during training. For clarity, we divide this strategy into three stages:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 110, + 555, + 353 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 110, + 555, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 218 + ], + "type": "text", + "content": "(1) Sampling Stage. In this stage, a group of candidates is sampled from the multi-scale grids of FPN layers. Here, Center Sampling is still adopted to maintain the strong receptive field for candidates, with a slight adjustment. To expand the range of candidate selection, we put forward Multi-layer Center Sampling (MCS), which applies Center Sampling across all FPN layers, rather than sampling with an increasing radius on a single layer (as proven ineffective in Sec. 4.3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "text", + "content": "(2) Matching Stage. The purpose of this stage is to select positive or negative labels for candidates based on their matching costs. The Matching cost is defined as a weighted summation of classification loss " + }, + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "text", + "content": ", regression loss " + }, + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "inline_equation", + "content": "L_{reg}" + }, + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "text", + "content": ", and polygon loss " + }, + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "inline_equation", + "content": "L_{poly}" + }, + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "text", + "content": ", which together evaluate the optimized cost from the prediction to ground truth. In this work, we employ Focal Loss [28], Polar IoU Loss [36], and the proposed RMask IoU Loss as classification loss, regression loss, and polygon loss, respectively. For a given instance, the matrix of matching costs " + }, + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "inline_equation", + "content": "Cost = \\{c_i | i = 1, \\dots, n\\}" + }, + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 221, + 555, + 353 + ], + "type": "text", + "content": " candidates can be formulated as follows:" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 344, + 361, + 553, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 361, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 344, + 361, + 553, + 374 + ], + "type": "interline_equation", + "content": "C o s t = \\lambda_ {c l s} L _ {c l s} + \\lambda_ {r e g} L _ {r e g} + \\lambda_ {p o l y} L _ {p o l y}, \\tag {5}", + "image_path": "bf9dd478455f56b933d27dfcab495531cc6a7dec8bb5029603fa9116ada0f13f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 381, + 554, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 554, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 554, + 428 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 381, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 313, + 381, + 554, + 428 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 381, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 313, + 381, + 554, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 381, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\lambda_{poly}" + }, + { + "bbox": [ + 313, + 381, + 554, + 428 + ], + "type": "text", + "content": " stand for the loss coefficients. Afterward, a fixed number " + }, + { + "bbox": [ + 313, + 381, + 554, + 428 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 381, + 554, + 428 + ], + "type": "text", + "content": " of positive labels are assigned to the candidates with minimal costs, and the others are negative:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 434, + 553, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 434, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 320, + 434, + 553, + 460 + ], + "type": "interline_equation", + "content": "X _ {i} = \\left\\{ \\begin{array}{l l} \\text {P o s i t i v e}, & c _ {i} \\in \\operatorname {t o p} - k _ {\\min } \\left(c _ {1}, c _ {2}, \\dots , c _ {n}\\right) \\\\ \\text {N e g a t i v e}, & c _ {i} \\notin \\operatorname {t o p} - k _ {\\min } \\left(c _ {1}, c _ {2}, \\dots , c _ {n}\\right), \\end{array} \\right. \\tag {6}", + "image_path": "0a3cfa92a07b757c032a1a8300b6c882d2f36890fcad1a3c423423311d01f4df.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 467, + 554, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 467, + 554, + 503 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 313, + 467, + 554, + 503 + ], + "type": "text", + "content": " refers to the label assigned to the " + }, + { + "bbox": [ + 313, + 467, + 554, + 503 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 467, + 554, + 503 + ], + "type": "text", + "content": "-th candidates. Notably, if a sample is matched by two instances simultaneously, the one with the largest area will be chosen." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 505, + 554, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 554, + 613 + ], + "type": "text", + "content": "(3) Weighting Stage. In the final stage, each positive sample is provided with a quality score, which modulates its contribution to loss functions. Our approach here is straightforward but effective, replacing the heuristic Centerness with the RMask IoU output by URM. Compared to Centerness, RMask IoU treats each positive sample equally, especially those high-quality samples located outside the contour, whose weights are zeroed out by Centerness (e.g., samples 5 and 6 in Fig. 3)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 613, + 554, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 554, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 554, + 637 + ], + "type": "text", + "content": "Loss Function. The overall loss function " + }, + { + "bbox": [ + 313, + 613, + 554, + 637 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 613, + 554, + 637 + ], + "type": "text", + "content": " consists of four parts:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 645, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 645, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 321, + 645, + 553, + 658 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {c l s} L _ {c l s} + \\lambda_ {r e g} L _ {r e g} + \\lambda_ {p o l y} L _ {p o l y} + \\lambda_ {s r e} L _ {s r e}, \\tag {7}", + "image_path": "7f9543d2b0674d5bafd490f273abec46fe221e673bd66e677b3028b8153b6df1.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "inline_equation", + "content": "L_{sre}" + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": " denotes the Cross-Entropy Loss in PolarMask used to supervise the Centerness head. In this case, we replace the original supervision signals with RMask IoU in the Centerness head." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19319" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 70, + 526, + 294 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 526, + 294 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 526, + 294 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 526, + 294 + ], + "type": "table", + "html": "
TypeMethodBackboneEpochSize\\(AP\\)\\(AP_{50}\\)\\(AP_{75}\\)FPS\\(Params\\downarrow\\)FLOPs↓
ContourDeepSnake [25]DLA-3416051230.3--24--
E2EC [39]DLA-3415051233.852.932.83529.80121
PolySnake [9]DLA-3425051234.4--18--
DANCE [24]R50-FPN3680036.858.539.01644.60274
MaskMask R-CNN [14]R50-FPN3680037.559.140.23344.40240
Cascade R-CNN [2]R50-FPN3680038.459.741.52377.331804
HTC [3]R50-FPN2080038.760.141.91780.261735
MS R-CNN [15]R50-FPN2480036.856.739.73060.74279
PointRend [17]R50-FPN3680038.459.541.52260.22266
CondInst [32]R50-FPN3680036.257.238.72934.17308
SOLO [34]R50-FPN3680036.457.938.92536.31332
SOLOv2 [35]R50-FPN3680038.459.541.23546.60227
YOLACT [1]R50-FPN5580029.048.729.84735.30237
SparseInst [7]R50-FPN14680033.253.434.75932.80201
PolarESE-Seg [38]R50-FPN30080021.648.722.438--
PolarMask [36]R50-FPN3680031.352.532.34234.74264
PolarNeXt (ours)R50-FPN3680036.159.737.34932.36186
", + "image_path": "f6c0a1cd8d1ae9f0528167386fe7e2e1cd83deb9c09570a071c2d1c4e451adbd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 333, + 284, + 479 + ], + "blocks": [ + { + "bbox": [ + 56, + 333, + 284, + 479 + ], + "lines": [ + { + "bbox": [ + 56, + 333, + 284, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 284, + 479 + ], + "type": "image", + "image_path": "5fafdc340ced1c68fbe9a840eae35cf39a8373f68bec6931b3d7f4b3e01dfed9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 487, + 295, + 521 + ], + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 521 + ], + "type": "text", + "content": "Figure 5. Comparison of Memory Usage (MB) for inference. During inference on COCO test-dev images, the peak memory usage is recorded as the final result of each method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 532, + 137, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 137, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 137, + 547 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 272, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 272, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 272, + 565 + ], + "type": "text", + "content": "4.1. Experiments and Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Experiment Settings. The experiments are conducted on MS COCO [21] benchmark and Cityscapes [8] dataset. MS COCO contains 80 classes with " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "115\\mathrm{k}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " training, " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "5\\mathrm{k}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " validation, and " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "20\\mathrm{k}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " testing images. Cityscapes has 2795 training images from 8 classes and another 500 images with high-quality annotations for validation. In terms of evaluation metrics, the accuracy of segmentation is evaluated by the standard AP metric, denoted as mask AP. Furthermore, mask AP can be divided into " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "AP_{50} / AP_{75}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " based on different IoU thresholds, as well as " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "AP_{S} / AP_{M} / AP_{L}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " for small/medium/large objects. All inference experiments are conducted on a single NVIDIA 4090D GPU, where FPS," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 333, + 547, + 414 + ], + "blocks": [ + { + "bbox": [ + 55, + 303, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 555, + 326 + ], + "type": "text", + "content": "Table 1. Experimental results on COCO test-dev. \"Size\" refers to the length of the shortest edge of input images. " + }, + { + "bbox": [ + 55, + 303, + 555, + 326 + ], + "type": "inline_equation", + "content": "{}^{ \\blacklozenge }" + }, + { + "bbox": [ + 55, + 303, + 555, + 326 + ], + "type": "text", + "content": " means smaller values are preferred." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 333, + 547, + 414 + ], + "lines": [ + { + "bbox": [ + 322, + 333, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 322, + 333, + 547, + 414 + ], + "type": "table", + "html": "
MethodAPAP50FPSMems↓
Mask R-CNN [14]31.558.6161968
SOLOv2 [35]27.248.5172076
CondInst [32]33.359.1152319
PolarMask [36]27.452.328575
PolarNeXt (ours)30.758.831566
", + "image_path": "43696f6fb9d53d93fb595fc5b1e1a32c49dbc3e4818653339076a585834e25b6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 422, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 445 + ], + "type": "text", + "content": "Table 2. Experimental results on Cityscapes val set. \"Mems\" stands for the memory usage (MB)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 460, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 508 + ], + "type": "text", + "content": "Parms, and FLOPs are calculated separately for inference speed, model complexity, and computational overhead. Notably, in all our experiments, TensorRT or FP16 is not used for acceleration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Implementation Details. PolarNeXt is instantiated on the PolarMask network, using a backbone ResNet-50 [13] pretrained on ImageNet [29], and removed the auxiliary box branch. All models are implemented on the MMDetection toolbox [4], trained on 2 GPUs with 2 images per GPU, and optimized with SGD. Weight decay and momentum are configured to 0.0001 and 0.9, respectively. The base learning rate is set to 0.01 without any decay scheme. For a fair comparison, following common practice, data augmentation only contains random flip and scale jitter unless specified. In APSD, the radius of Center Sampling for candidate selection at each FPN layer is set to 1.5 and the number " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " of assigned positive labels is set to 9 for each instance. Loss coefficients " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{reg}" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{poly}" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{sre}" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " are all empirically set to 1.0. Following the rasterizer configuration in [18], we set the desired mask resolution to " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " and the rasterization sharpness to " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19320" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 70, + 272, + 174 + ], + "blocks": [ + { + "bbox": [ + 79, + 70, + 272, + 174 + ], + "lines": [ + { + "bbox": [ + 79, + 70, + 272, + 174 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 272, + 174 + ], + "type": "table", + "html": "
URMAPSDAP
weightlosscost
29.1
✓30.3 (+1.2)
✓✓31.0 (+1.9)
✓28.6 (-0.5)
✓✓✓30.7 (+1.6)
✓✓✓✓33.9 (+4.8)
", + "image_path": "b4afa684a23bf894c1bf7bd867b7f043dd9ea03ac5f1f3a18a400eb545dce7f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 74, + 247, + 277, + 292 + ], + "blocks": [ + { + "bbox": [ + 55, + 183, + 294, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 183, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 294, + 239 + ], + "type": "text", + "content": "Table 3. Ablation studies on URM and APSD. \"weight\" means whether Centerness is replaced by RMask IoU for sample weighting. \"loss\" refers to the use of RMask IoU Loss in loss functions. \"cost\" indicates the incorporation of RMask IoU Loss into matching costs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 247, + 277, + 292 + ], + "lines": [ + { + "bbox": [ + 74, + 247, + 277, + 292 + ], + "spans": [ + { + "bbox": [ + 74, + 247, + 277, + 292 + ], + "type": "table", + "html": "
Polar IoU LossAPAPSAPMAPL
w/o.32.814.335.148.5
w/.33.916.435.949.4
", + "image_path": "12bd417aa70d2130bcdf9633f294d522f6abe274dd4f42d12433e848dd9f8372.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 140, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 140, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 140, + 357 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 363, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 615 + ], + "type": "text", + "content": "MS COCO. Following common practice, we evaluate PolarNeXt on the MS COCO benchmark and compare test-dev results to some state-of-the-art models in Tab. 1. In terms of computational cost, PolarNeXt is more lightweight than the other models with R50-FPN backbone, only requiring 32.8 MB Parameters and 186 GB FLOPs. Even when compared to the contour-based methods using DLA-34 backbone, it remains on par. In terms of inference speed, PolarNeXt is second only to the prevalent SpareInst, but it achieves higher accuracy (36.1% vs. 33.2% AP) with a shorter training schedule (36 vs. 146 epochs). In terms of segmentation accuracy, PolarNeXt delivers competitive results compared to other more complex methods. Particularly, without bells and whistles, PolarNeXt achieves a 4.8% mask AP improvement over PolarMask, redemonstrating the potential of Polar Representation in instance segmentation. Moreover, we visualize the memory usage of some representative methods during inference in Fig. 5. Obviously, PolarNeXt shows its consistency with the classical object detector FCOS, with memory usage significantly lower than other instance segmentation methods, only half or even less." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Cityscapes. As an extended comparative experiment in Tab. 2, Cityscapes dataset, characterized by high-resolution images, is utilized to examine the robustness and generalization ability of our model. Notably, all models are directly trained on the Cityscapes training set under 64 epochs, without COCO initialization. Since fragmented instances frequently occur in Cityscapes, the Multi-component Detection Strategy suggested in [25] is applied to PolarNeXt and" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 563, + 163 + ], + "blocks": [ + { + "bbox": [ + 55, + 300, + 295, + 334 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 334 + ], + "type": "text", + "content": "Table 4. Feasibility analysis of removing Polar IoU Loss. The symbols \"w/o.\" and \"w/.\" indicate whether the Polar IoU Loss is employed." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 71, + 563, + 163 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 563, + 163 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 563, + 163 + ], + "type": "table", + "html": "
Candidate SelectionLayerAPT.T.
Center Sampling (s.r.=1.5)single31.013h
Center Sampling (s.r.=2.0)30.6 (-0.4)14h
All Grid Pointsmulti--
Points within Boxes33.1 (+2.1)28h
Points within Contours33.0 (+2.0)25h
Center Sampling (s.r.=1.5)33.9 (+2.9)16h
", + "image_path": "8368d2df9e16f674199f52a6b8bc75cddaa343ac0dc37fa4cd9ed1adca5ee422.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 322, + 236, + 545, + 280 + ], + "blocks": [ + { + "bbox": [ + 313, + 171, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 227 + ], + "type": "text", + "content": "Table 5. The effectiveness of some approaches for candidate selection. \"single\" denotes selecting candidates on a specific FPN layer based on the instance size, and \"multi\" refers to selecting candidates across all FPN layers. \"T.T.\" and \"s.r.\" are abbreviations for training time and sampling radius, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 236, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 322, + 236, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 322, + 236, + 545, + 280 + ], + "type": "table", + "html": "
MethodAPbAP50APSbAPMAPLb
PolarMask16.338.316.028.413.2
PolarNeXt19.846.720.232.215.9
", + "image_path": "6d7f67da247df54ce897022a65efa6c6a86ddce65cd4b23319818495c7c5ac3b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 289, + 555, + 310 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 310 + ], + "type": "text", + "content": "Table 6. Comparison of boundary quality between PolarMask and PolarNeXt." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "text", + "content": "PolarMask, leading to approximately a " + }, + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "text", + "content": " AP improvement. Under the increased resolution of input images (from " + }, + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "inline_equation", + "content": "800 \\times 1333" + }, + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "inline_equation", + "content": "1024 \\times 2048" + }, + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "text", + "content": "), mask-based methods suffer from the heavy memory usage introduced by dense pixel classification, which is nearly four times greater than that of polar-based methods. Meanwhile, the speed advantage of polar-based methods becomes more apparent, requiring only half the inference time compared to mask-based methods. As expected, PolarNeXt maintains a " + }, + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "text", + "content": " AP improvement over PolarMask, further demonstrating the effectiveness of our proposed method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 465, + 416, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 416, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 416, + 477 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 483, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 543 + ], + "type": "text", + "content": "In this section, we conduct ablations to assess the influence of the proposed components on instance segmentation performance, and then analyze their details in the following. All experiments are conducted on MS COCO val under a 1x training schedule." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Component Impact and Correlation. The effectiveness of URM and APSD is reported in Tab. 3. Using URM alone brings a " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " AP improvement on the baseline, while combining it with APSD further boosts the performance by an additional " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " AP. It is worth noting that APSD fails to function independently of URM, the absence of which results in a decrease in performance. To be specific, without APSD, URM contributes to the sample weighting and loss function, leading to a " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " AP increase, respectively. Conversely, when APSD is applied, the matching costs have to incorporate RMask IoU Loss; otherwise, this strategy will fail or even backfire. In view of this, we conclude that the effectiveness of our adaptive strategy hinges on reliable polygonal assessment for support." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19321" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 531, + 254 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 531, + 254 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 531, + 254 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 531, + 254 + ], + "type": "image", + "image_path": "4420ab76d06333c0ef197cf47c546769c0f1bb98d8ecf44aac60ead69e5361c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 262, + 423, + 274 + ], + "lines": [ + { + "bbox": [ + 186, + 262, + 423, + 274 + ], + "spans": [ + { + "bbox": [ + 186, + 262, + 423, + 274 + ], + "type": "text", + "content": "Figure 6. Visualization with Rays for PolarMask and PolarNeXt." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 287, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 430 + ], + "type": "text", + "content": "Candidate Selection. In Tab. 5, we compare some approaches to expand the range of candidate selection. On the one hand, Center Sampling on multiple FPN layers is more effective and efficient than the dense point selection within boxes or contours. In particular, the approach suggested in [31], which selects all grid points of FPN, results in unacceptable training time and GPU memory overflow. On the other hand, directly expanding the radius of Center Sampling on a single layer causes a " + }, + { + "bbox": [ + 55, + 287, + 296, + 430 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 55, + 287, + 296, + 430 + ], + "type": "text", + "content": " drop in AP, as the additional selected samples tend to fall outside the receptive field. Accordingly, the center prior still has a positive influence on sample decisions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 531 + ], + "type": "text", + "content": "RMask IoU vs. Polar IoU. In Fig. 4, two scatter plots are drawn to compare Polar IoU and RMask IoU in polygonal assessment. These scatter points denote all the polygons predicted during inference on MS COCO val images. The fitted curve in Fig. 4(b) is almost aligned with the ideal curve, whereas the one in Fig. 4(a) shows significant deviation. This indicates that RMask IoU better reflects the actual polygonal quality compared to Polar IoU." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "The necessity of Polar IoU Loss. In Tab. 4, we explore the feasibility of directly removing Polar IoU Loss in our proposed training pipeline. Experimental results indicate that this operation results in a " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " AP degradation in performance. The impact on small objects is more notable, with " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "AP_{S}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " decreased by " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": ". We analyze that there may be a slight granularity missing as well as feature misalignment during the conversion from vertex sets to rasterized masks. Moreover, as mentioned in [18], aliasing occurs in the deconvolutional process of FPN when pre-predicted RoIs are unavailable in one-stage detectors. Fortunately, the combination of Polar IoU Loss and RMask IoU Loss brings a notable boost, with the two complementing each other. We believe that Polar IoU Loss serves as a polygonal internal constraint for dense distance regression, while RMask IoU" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 287, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 312 + ], + "type": "text", + "content": "Loss provides a polygonal external constraint to incorporate the representation deviation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 317, + 465, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 465, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 465, + 330 + ], + "type": "text", + "content": "4.4. Boundary Quality Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 334, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 455 + ], + "type": "text", + "content": "In Tab. 6, we introduce Boundary AP [5] (" + }, + { + "bbox": [ + 313, + 334, + 555, + 455 + ], + "type": "inline_equation", + "content": "AP^b" + }, + { + "bbox": [ + 313, + 334, + 555, + 455 + ], + "type": "text", + "content": " for short) to evaluate the boundary quality of predicted polygons. Experimental results show that PolarNeXt achieves a significant improvement in boundary quality compared to PolarMask, boosting " + }, + { + "bbox": [ + 313, + 334, + 555, + 455 + ], + "type": "inline_equation", + "content": "AP^b" + }, + { + "bbox": [ + 313, + 334, + 555, + 455 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 334, + 555, + 455 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 313, + 334, + 555, + 455 + ], + "type": "text", + "content": ". Furthermore, we present some segmentation results with differences in starting points in Fig. 6. Obviously, compared to the center-fixed starting points in PolarMask, the dynamically selected starting points in PolarNeXt are more flexible, which enhances the representation capability of bounding polygons." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 463, + 388, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 388, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 388, + 475 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 483, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 651 + ], + "type": "text", + "content": "In this paper, we introduce a novel polar-based framework, PolarNeXt, which redemonstrates the superiority of Polar Representation in instance segmentation. To account for the previously ignored representation error, tailored training strategy APSD and assessment module URM are proposed. Together, these components significantly enhance PolarNeXt, yielding notable performance improvements over other polar-based methods. By fully leveraging the advantages of Polar Representation, PolarNeXt achieves comparable segmentation accuracy while requiring only half or even less of the computational cost and inference time compared to mainstream mask-based and contour-based methods. We anticipate that PolarNeXt will set a new standard for polar-based methods in instance segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 658, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgments. This work is supported by the National Natural Science Foundation of China [Grant No. 62225308], and the Shanghai Technical Service Computing Center of Science and Engineering, Shanghai University." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19322" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmentation. In ICCV, pages 9157-9166, 2019. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 296, + 156 + ], + "type": "text", + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE TPAMI, 43(5):1483-1498, 2019. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 158, + 295, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 158, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 295, + 201 + ], + "type": "text", + "content": "[3] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In CVPR, pages 4974-4983, 2019. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 203, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 295, + 255 + ], + "type": "text", + "content": "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ross Girshick, Piotr Dólar, Alexander C Berg, and Alexander Kirillov. Boundary iou: Improving object-centric image segmentation evaluation. In CVPR, pages 15334-15342, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 334 + ], + "type": "text", + "content": "[6] Dominic Cheng, Renjie Liao, Sanja Fidler, and Raquel Urtasun. Darnet: Deep active ray network for building segmentation. In CVPR, pages 7431-7439, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 295, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 379 + ], + "type": "text", + "content": "[7] Tianheng Cheng, Xinggang Wang, Shaoyu Chen, Wenqiang Zhang, Qian Zhang, Chang Huang, Zhaoxiang Zhang, and Wenyu Liu. Sparse instance activation for real-time instance segmentation. In CVPR, pages 4433-4442, 2022. 1, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "text", + "content": "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "text", + "content": "[9] Hao Feng, Keyi Zhou, Wengang Zhou, Yufei Yin, Jiajun Deng, Qi Sun, and Houqiang Li. Recurrent generic contour-based instance segmentation with progressive learning. IEEE TCSVT, 2024. 1, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "type": "text", + "content": "[10] Gunther Greiner and Kai Hormann. Efficient clipping of arbitrary polygons. ACM Transactions on Graphics (TOG), 17 (2):71-83, 1998. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 514, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 546 + ], + "type": "text", + "content": "[11] Wenchao Gu, Shuang Bai, and Lingxing Kong. A review on 2d instance segmentation based on deep neural networks. Image and Vision Computing, 120:104401, 2022. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 547, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 579 + ], + "type": "text", + "content": "[12] Shir Gur, Tal Shaharabany, and Lior Wolf. End to end trainable active contours via differentiable rendering. arXiv preprint arXiv:1912.00367, 2019. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "type": "text", + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 1, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "text", + "content": "[15] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In CVPR, pages 6409-6418, 2019. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[16] Michael Kass, Andrew Witkin, and Demetri Terzopoulos. Snakes: Active contour models. IJCV, 1(4):321-331, 1988. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[17] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In CVPR, pages 9799-9808, 2020. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "text", + "content": "[18] Justin Lazarow, Weijian Xu, and Zhuowen Tu. Instance segmentation with mask-supervised polygonal boundary transformers. In CVPR, pages 4382-4391, 2022. 4, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "text", + "content": "[19] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. NeurIPS, 33:21002-21012, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "type": "text", + "content": "[20] Justin Liang, Namdar Homayounfar, Wei-Chiu Ma, Yuwen Xiong, Rui Hu, and Raquel Urtasun. Polytransform: Deep polygon transformer for instance segmentation. In CVPR, pages 9131-9140, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 276, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 318 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117-2125, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 319, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 319, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 319, + 553, + 353 + ], + "type": "text", + "content": "[23] Huan Ling, Jun Gao, Amlan Kar, Wenzheng Chen, and Sanja Fidler. Fast interactive object annotation with curve-gcn. In CVPR, pages 5257-5266, 2019. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 354, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 408 + ], + "type": "text", + "content": "[24] Zichen Liu, Jun Hao Liew, Xiangyu Chen, and Jiashi Feng. Dance: A deep attentive contour model for efficient instance segmentation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 345-354, 2021. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 410, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 410, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 410, + 553, + 443 + ], + "type": "text", + "content": "[25] Sida Peng, Wen Jiang, Huaijin Pi, Xiuli Li, Hujun Bao, and Xiaowei Zhou. Deep snake for real-time instance segmentation. In CVPR, pages 8533-8542, 2020. 1, 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 444, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 465 + ], + "type": "text", + "content": "[26] Joseph Redmon. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "text", + "content": "[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2016. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 511, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 553, + 533 + ], + "type": "text", + "content": "[28] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "text", + "content": "[30] Uwe Schmidt, Martin Weigert, Coleman Broaddus, and Gene Myers. Cell detection with star-convex polygons. In Medical Image Computing and Computer Assisted Intervention, pages 265–273, 2018. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "text", + "content": "[31] Peize Sun, Yi Jiang, Enze Xie, Wenqi Shao, Zehuan Yuan, Changhu Wang, and Ping Luo. What makes for end-to-end object detection? In International Conference on Machine Learning, pages 9934-9944, 2021. 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "type": "text", + "content": "[32] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation. In ECCV, pages 282-298, 2020. 3, 6" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19323" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 340 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[33] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: A simple and strong anchor-free object detector. IEEE TPAMI, 44(4):1922-1933, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[34] Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, and Lei Li. Solo: Segmenting objects by locations. In ECCV, pages 649-665, 2020. 1, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "text", + "content": "[35] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmentation. NeurIPS, 33:17721-17732, 2020. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[36] Enze Xie, Peize Sun, Xiaoge Song, Wenhai Wang, Xuebo Liu, Ding Liang, Chunhua Shen, and Ping Luo. Polarmask: Single shot instance segmentation with polar representation. In CVPR, pages 12193-12202, 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[37] Enze Xie, Wenhai Wang, Mingyu Ding, Ruimao Zhang, and Ping Luo. Polarmask++: Enhanced polar representation for single-shot instance segmentation and beyond. IEEE TPAMI, 44(9):5385-5400, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[38] Wenqiang Xu, Haiyang Wang, Fubo Qi, and Cewu Lu. Explicit shape encoding for real-time instance segmentation. In ICCV, pages 5168-5177, 2019. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 299, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 340 + ], + "type": "text", + "content": "[39] Tao Zhang, Shiqing Wei, and Shunping Ji. E2ec: An end-to-end contour-based method for high-quality high-speed instance segmentation. In CVPR, pages 4443-4452, 2022. 1, 3, 6" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19324" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Polarized Color Screen Matting/4bc8f905-0a33-46ab-838e-6d058731820a_content_list.json b/2025/Polarized Color Screen Matting/4bc8f905-0a33-46ab-838e-6d058731820a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a5b108dc2f7b2899f8418dd002ae106253f31d09 --- /dev/null +++ b/2025/Polarized Color Screen Matting/4bc8f905-0a33-46ab-838e-6d058731820a_content_list.json @@ -0,0 +1,2158 @@ +[ + { + "type": "text", + "text": "Polarized Color Screen Matting", + "text_level": 1, + "bbox": [ + 336, + 130, + 658, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kenji Enomoto", + "bbox": [ + 236, + 180, + 359, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scott Cohen", + "bbox": [ + 400, + 181, + 500, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Brian Price", + "bbox": [ + 540, + 181, + 632, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adobe Research", + "bbox": [ + 433, + 199, + 563, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "TJ Rhodes", + "bbox": [ + 673, + 181, + 759, + 196 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0676404ca766dac43b5b8a02829ede517d2550acb824f2073535671628529079.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 243, + 194, + 318 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e95f43ac7e90b9ebac418a26311ad2a3d08a3b8dcba7b6d7937d071c82fdbcc0.jpg", + "image_caption": [ + "(a) Scene" + ], + "image_footnote": [], + "bbox": [ + 96, + 320, + 194, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce6f9c81fb53732e98eb389493d40b34a96ae4e2f62a6b5212b7c4e53a3759a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 243, + 294, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6438447227b25d72ca2a8e9bdb7ad59b4d825e488fd9b42f1cfcaa44919fa33.jpg", + "image_caption": [ + "Figure 1. (a) Scenes with fully polarized blue/green backgrounds under unpolarized known color illumination. (b) Commercial keying tools [31] use color information to extract an alpha matte, often struggling with hair details or motion blur due to the essential ill-posedness of the problem. (c) Polarimetric matting [22] uses polarization cues to extract an alpha matte without manual operations, often struggling with polarized foreground reflections as shown in the specular highlight. (d) Our Polarized Color Screen Matting leverages color and polarization cues, which enables more accurate alpha matte extraction for a wide range of real-world objects." + ], + "image_footnote": [], + "bbox": [ + 196, + 320, + 294, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69c96b4abb5b06b2071d3def511369b60a010a45cb65e915c17f764726b91d2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 243, + 395, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5cad6778570a9389a0a3ac9d09ce863bb96634e94f38a8e7b9b6927a96ebd662.jpg", + "image_caption": [ + "(b) Chroma keying" + ], + "image_footnote": [], + "bbox": [ + 295, + 320, + 395, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bed4f6d6d4e1a03af811c6f179af0054a70f98c21d97e9cf3b20b6141410b003.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 397, + 243, + 495, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5af54e8b8df1b94ca91f7ea6203238b69b6915162b021aa454df94f51b43b50.jpg", + "image_caption": [ + "(c) Polarimetric matting" + ], + "image_footnote": [], + "bbox": [ + 397, + 320, + 495, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f880d8aef1391ebe6642591716bc5ed694a8cbaac7b0994b45b6dfdf7bdba2d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 243, + 596, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50dd9ddca87d68425d7b2cc0222296af58854754c09d0bb75e17979a94c0db13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 320, + 596, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b527ec44273f4b9cdef9dcddba211b0c587f1db1d19a0b8f9cd2f5f5c62e7e0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 243, + 696, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5cc1348bab708cab9fc053d68e931caa1241e4ee4ede2a96b16ab744692f5ec0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 320, + 696, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7673760209bd2dcc6c8eef3eb83ef57b2c52d9a53a3bf33d88034958b8ef9712.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 243, + 797, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/96572da303678e1e9df8f2f7b5de0272a3dd54b32048b9cd8a37ef830bd48393.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 699, + 320, + 795, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a57e20190af89c48e969dc6cfd5742f021175db4a29626edc9793cea21285089.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 243, + 897, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/42fd110c552e5a0f523e6182015f3774a4d6b3765acfa0c31b02da02d2368bad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 320, + 897, + 396 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 517, + 325, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper considers the long-standing problem of extracting alpha mattes from video using a known background. While various color-based or polarization-based approaches have been studied in past decades, the problem remains ill-posed because the solutions solely rely on either color or polarization. We introduce Polarized Color Screen Matting, a single-shot, per-pixel matting theory for alpha matte and foreground color recovery using both color and polarization cues. Through a theoretical analysis of our diffuse-specular polarimetric compositing equation, we derive practical closed-form matting methods with their solvability conditions. Our theory concludes that an alpha matte can be extracted without manual corrections using off-the-shelf equipment such as an LCD monitor, polarization camera, and unpolarized lights with calibrated color. Experiments on synthetic and real-world datasets verify the validity of our theory and show the capability of our matting methods on real videos with quantitative and qualitative comparisons to color-based and polarization-based matting methods.", + "bbox": [ + 89, + 580, + 483, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 517, + 643, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image and video editing rely on compositing [25] to blend images. The quality of the composite depends on how accurately the alpha matte depicts opacity. Modern matte extraction methods use inherent information such as color or polarization to extract the alpha matte. However, they are forced to compromise on quality of foreground color, alpha, or both. This paper combines color and polarimetric observations to constrain the problem.", + "bbox": [ + 511, + 547, + 906, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Color-based alpha matte extraction methods are most commonly used for compositing due to equipment availability. Chroma keying [2, 29] is the de facto standard in film making, but because it is underconstrained, it often requires manual corrections to extract high-quality alpha mattes. In addition to this, it often struggles with motion blur and transparent objects as depicted in Fig. 1. Recent learning-based matting methods [18, 19, 23, 32] can sometimes extract acceptable mattes but they still struggle with complex situations; potential improvements in modeling and dataset remain. Magenta green screen [28] can extract a more precise alpha matte by sacrificing the green color channel, but this later requires a restoration of the green color channel using machine-learning colorization. All these challenges are rooted in the fact that the use of", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "391", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "color alone to recover an alpha matte is underconstrained.", + "bbox": [ + 89, + 90, + 470, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alternatively, polarization-based matting methods [6, 11, 22] can extract the alpha matte using the polarization of light. For a scene backlit with polarized light, there is less ambiguity between phase angle than color, thus physically accurate alpha mattes can be obtained for the majority of pixels. However, in cases where the foreground subject reflects polarized light whose phase angle is comparable to the background, the alpha matte quality will be degraded, as shown in the headphone and glass scenes in Fig. 1.", + "bbox": [ + 88, + 106, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce Polarized Color Screen Matting (pCSM), a single-shot, per-pixel matting theory that uses both color and polarization cues. Our theory is built upon our proposed diffuse-specular polarimetric compositing equation, where foreground appearance is explicitly modeled by a summation of diffuse and specular reflections. This new compositing equation steps forward a theoretical analysis of the matting problem, resulting in two matting methods: (1) $p\\text{CSM-S}$ , a fast and robust matting method with an approximation of unpolarized diffuse reflections that is applicable to any scene, and (2) $p\\text{CSM-DS}$ , an accurate and approximation-free matting method. Through our theoretical analysis, we found solvability conditions of pCSM-DS in terms of color and polarized reflection from the foreground subject. For the practical use of our theoretical matting method, we introduce a physics-inspired regularization, which grants our methods significant robustness to noisy real-world observations.", + "bbox": [ + 91, + 242, + 483, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are threefold:", + "bbox": [ + 109, + 513, + 444, + 527 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A diffuse-specular polarimetric compositing equation that unlocks further theoretical analysis of the polarized screen matting problem.", + "- A fast and simple matting method, pCSM-S, that is applicable to any scene.", + "- An accurate and approximation-free matting method, pCSM-DS, with solvability conditions." + ], + "bbox": [ + 89, + 529, + 482, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show quantitatively and qualitatively that our methods produce more accurate results than state-of-the-art solely color-based or polarization-based matting methods.", + "bbox": [ + 89, + 635, + 482, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 694, + 232, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early methods for computing alpha mattes from video have focused on capturing footage against a solid-colored background [7], often referred to as chroma keying. Indeed, chroma keying is an industry standard, implemented in commercial products [31] and used for generating video matting datasets [18, 33]. However, despite its popularity, chroma keying is underconstrained and often performs poorly unless an expert carefully sets up the capture environment and manually post-processes the matte.", + "bbox": [ + 89, + 719, + 482, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the shortcomings of standard chroma keying, further work has focused on using color to facilitate matte extraction. The sodium vapor matting process [24] constrains", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the problem by narrowing the spectrum of backlight emitted from a low-pressure sodium vapor lamp. A beamsplitter and bandpass filter allow the background sodium vapor light to be captured on one strip of film, forming an alpha matte, and the foreground subject's color to be captured through a notch reject filter on a second strip of film, blocking the sodium vapor light. More recently, magenta green screen [28] looked to eliminate the need for a specialized camera system by placing a foreground subject strongly lit by magenta light in front of a green background to observe the green spectrum solely at the background. However, the foreground magenta light loses the true foreground color, requiring machine-learning colorization to recover the lost green color of the foreground. In theory, triangulation matting [29] can compute the true alpha matte by taking two (or more) images of a static subject in front of different colored backgrounds to triangulate the alpha matte. However, since the images must be taken at different times to allow the background color to change, triangulation matting is not suitable for dynamic scenes without sacrificing motion blur, such as with stop-motion photography [12].", + "bbox": [ + 511, + 90, + 903, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of using color to constrain the capture system, polarimetric triangulation matting (PTM) [22] employs polarization. PTM uses a beamsplitter with two polarization filters to simultaneously capture two polarization images of a target scene against a polarized background screen. These images can serve as input for triangulation matting if the foreground reflects only unpolarized light; however, in practice, polarized foreground reflections are observed, which cause perceivable errors in foreground subjects.", + "bbox": [ + 511, + 420, + 903, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Addressing the issue of polarized foreground reflections, PolarMatte [11] reformulates the polarized screen matting problem by leveraging four polarization images captured with a modern polarization camera (see Fig. 2). In this framework, an extended robust PTM is first applied to pixels with negligible polarized reflections, and the resulting high-confidence alpha mattes are subsequently interpolated to pixels with polarized foreground reflections using spatial context through a polarimetric matting Laplacian. However, two critical limitations persist. First, because PolarMatte is not formulated as a per-pixel method, establishing solvability conditions remains challenging, thereby complicating both its theoretical analysis and practical reliability. Second, its exclusive reliance on polarization cues (i.e., the use of white polarized background) renders it vulnerable when the polarized reflection aligns with the background polarity—a situation often encountered in the presence of specular highlights (e.g., from jewelry or glossy skin). In this paper, we aim to overcome these limitations by jointly integrating color and polarization cues, which more robustly constrains the matting problem without adding significant complexity to the capture process.", + "bbox": [ + 511, + 568, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "392", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/371d023ffd82abc666dc055f83eaafda5f5050a9af943cffc5707a6aef7ee313.jpg", + "image_caption": [ + "Figure 2. Illustration of our capture setup. A recent polarization camera with a color polarization filter array sensor captures four polarization images in a single-shot. Our method is tailored for the polarization images of a foreground subject with a known polarized background under known color illuminations." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 482, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Polarized Color Screen Matting", + "text_level": 1, + "bbox": [ + 89, + 378, + 382, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section delivers a proof that a single-shot and per-pixel matting problem has a closed-form solution given: (1) the background is known and polarized, (2) the foreground illumination color is known, and (3) a diffuse-specular reflectance model for the foreground subject. Our pCSM theory is tailored for polarization images at four angles $\\left\\{0, \\frac{\\pi}{4}, \\frac{\\pi}{2}, \\frac{3\\pi}{4}\\right\\}$ captured by a recent polarization camera shown in Fig. 2, similarly to PolarMatte [11]. The reader is encouraged to reference Table 5 for the notations used in the following sections.", + "bbox": [ + 89, + 404, + 483, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffuse-Specular Polarimetric Compositing Equation", + "text_level": 1, + "bbox": [ + 89, + 563, + 483, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The matting problem is formulated as a decomposition of the color vector of a given pixel into foreground and background color vectors and the opacity of the foreground, commonly called an alpha matte. In a linear color space, the compositing process of a polarization image is described in the polarimetric compositing equation [11]:", + "bbox": [ + 89, + 602, + 483, + 694 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {\\phi} = \\alpha \\mathbf {f} _ {\\phi} + (1 - \\alpha) \\mathbf {b} _ {\\phi}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 703, + 480, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_{\\phi}, \\mathbf{f}_{\\phi}, \\mathbf{b}_{\\phi} \\in \\mathbb{R}_+^3$ are composite, foreground, and background color polarization vectors at an angle $\\phi \\in \\{0, \\frac{\\pi}{4}, \\frac{\\pi}{2}, \\frac{3\\pi}{4}\\}$ , and $\\alpha \\in [0,1]$ is an alpha matte. We employ a diffuse-specular reflectance model [16] on the foreground color, which is the sum of diffuse and specular terms:", + "bbox": [ + 89, + 728, + 483, + 818 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {\\phi} = \\underbrace {\\left(u _ {d} + p _ {d} \\cos^ {2} \\left(\\theta_ {d} - \\phi\\right)\\right) \\mathbf {f} _ {d}} _ {\\text {d i f f u s e t e r m}} + \\underbrace {\\left(u _ {s} + p _ {s} \\cos^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {s}} _ {\\text {s p e c u l a r t e r m}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 825, + 482, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, the diffuse and specular terms are decomposed into unpolarized and polarized terms, where $u_{d}, p_{d} \\in \\mathbb{R}_{+}$", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are the intensities of the unpolarized and polarized diffuse reflections, respectively, and $u_{s}, p_{s} \\in \\mathbb{R}_{+}$ are the same for the specular reflection. $\\mathbf{f}_d, \\mathbf{f}_s \\in S_+^2 \\subset \\mathbb{R}_+^3$ denote unit color vectors of diffuse and specular reflections, respectively. $\\theta_d, \\theta_s \\in [0, \\pi)$ are phase angles of the diffuse and specular reflections, and it is known that $\\theta_d = \\theta_s \\pm \\frac{\\pi}{2}$ under unpolarized illumination [3, 9, 16]. Therefore, the diffuse term in Eq. (2) can be rewritten using $\\theta_{s}$ as", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(u _ {d} + p _ {d} \\sin^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {d}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 223, + 906, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our problem setting, the background is fully polarized, which can be described in", + "bbox": [ + 511, + 253, + 905, + 282 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} _ {\\phi} = p _ {b} \\cos^ {2} (\\theta_ {b} - \\phi) \\mathbf {q}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 296, + 906, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_b \\in \\mathbb{R}_+$ , $\\theta_b \\in [0, \\pi)$ , and $\\mathbf{q} \\in S_+^2 \\subset \\mathbb{R}_+^3$ are the intensity, phase angle, and unit color vector of the backlight, respectively. Finally, the polarimetric compositing equation (1) can be written as", + "bbox": [ + 511, + 325, + 905, + 387 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {c} _ {\\phi} = \\alpha \\left(\\left(u _ {d} + p _ {d} \\sin^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {d} + \\left(u _ {s} + p _ {s} \\cos^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {s}\\right) \\\\ + (1 - \\alpha) p _ {b} \\cos^ {2} (\\theta_ {b} - \\phi) \\mathbf {q}. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 397, + 903, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We call this compositing process the diffuse-specular polarimetric compositing equation. The diffuse-specular reflectance model is widely used in polarimetric vision problems, such as diffuse-specular separation [10, 16], shape-from-polarization [17, 20], and polarimetric BRDF acquisition [4]. We extend it to the compositing equation to represent opacity.", + "bbox": [ + 511, + 446, + 905, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For ease of theoretical analysis, we transform observed polarization images into Stokes parameters [30]. This paper only considers linear polarization, and the Stokes parameters can also be described in the form of compositing equations:", + "bbox": [ + 511, + 551, + 905, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {0} = \\frac {1}{2} \\sum_ {\\phi} \\mathbf {c} _ {\\phi} \\stackrel {\\mathrm {d e f}} {=} \\mathbf {d} _ {0} + (1 - \\alpha) \\mathbf {q} _ {0}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 638, + 906, + 675 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {1} = \\mathbf {c} _ {0} - \\mathbf {c} _ {\\frac {\\pi}{2}} \\stackrel {\\mathrm {d e f}} {=} \\mathbf {d} _ {1} + (1 - \\alpha) \\mathbf {q} _ {1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 680, + 903, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {2} = \\mathbf {c} _ {\\frac {\\pi}{4}} - \\mathbf {c} _ {\\frac {3 \\pi}{4}} \\stackrel {\\text {d e f}} {=} \\mathbf {d} _ {2} + (1 - \\alpha) \\mathbf {q} _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 700, + 903, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 729, + 558, + 742 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {0} = \\alpha \\left(\\left(2 u _ {d} + p _ {d}\\right) \\mathbf {f} _ {d} + \\left(2 u _ {s} + p _ {s}\\right) \\mathbf {f} _ {s}\\right), \\quad \\mathbf {q} _ {0} = p _ {b} \\mathbf {q}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 757, + 903, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {1} = \\alpha \\cos 2 \\theta_ {s} \\left(p _ {s} \\mathbf {f} _ {s} - p _ {d} \\mathbf {f} _ {d}\\right), \\mathbf {q} _ {1} = p _ {b} \\cos 2 \\theta_ {b} \\mathbf {q}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 777, + 903, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {2} = \\alpha \\sin 2 \\theta_ {s} \\left(p _ {s} \\mathbf {f} _ {s} - p _ {d} \\mathbf {f} _ {d}\\right), \\mathbf {q} _ {2} = p _ {b} \\sin 2 \\theta_ {b} \\mathbf {q}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 796, + 903, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\mathbf{s}_0$ is equivalent to an ordinary color image, and $\\mathbf{s}_1$ and $\\mathbf{s}_2$ represent the polarization properties of an observation.", + "bbox": [ + 511, + 824, + 903, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper aims to recover the alpha matte $\\alpha$ given the polarization images $\\mathbf{c}_{\\phi}$ and the background images $\\mathbf{b}_{\\phi}$ for all the angles $\\phi \\in \\{0, \\frac{\\pi}{4}, \\frac{\\pi}{2}, \\frac{3\\pi}{4}\\}$ .", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "393", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f44b2a1cc6a24d367f7a5c879f93c4118db12b0edb918cb8db7717f51ef82718.jpg", + "image_caption": [ + "Figure 3. (Left) A scene of a headphone in front of a blue polarized background and its polarized intensity computed as $\\sqrt{\\mathbf{s}_1^2 + \\mathbf{s}_2^2}$ . (Right) The color distribution of the polarized intensities, showing that the majority of the polarized intensities derive from background and specular reflections." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 478, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. pCSM with Polarized Specular Reflection", + "text_level": 1, + "bbox": [ + 89, + 409, + 447, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section introduces a pCSM method, which we call $p\\mathrm{CSM - S}$ that highlights the polarized specular reflection. Prior polarimetric matting methods [11, 22] show that polarized reflections from a foreground subject are a major source of error in the alpha matte estimation. In particular, polarized specular reflections have larger intensities than polarized diffuse reflections and a dominant effect on the estimated alpha matte (see Fig. 3). This observation motivates us to simplify the polarimetric compositing equation and Stokes parameters without the diffuse term (i.e., $p_d = 0$ ). With this simplification, Eqs. (7) and (8) can be rewritten as", + "bbox": [ + 89, + 431, + 483, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {1} = \\alpha p _ {s} \\cos 2 \\theta_ {s} \\mathbf {f} _ {s} + (1 - \\alpha) \\mathbf {q} _ {1}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 613, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {2} = \\alpha p _ {s} \\sin 2 \\theta_ {s} \\mathbf {f} _ {s} + (1 - \\alpha) \\mathbf {q} _ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 631, + 482, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assuming the dichromatic reflectance model [26], the specular reflection color $\\mathbf{f}_s$ is independent of the object's albedo and is equal to the ambient light color. Therefore, we can assume the specular reflection color $\\mathbf{f}_s$ to be known if the ambient light color is calibrated or ambient light is natural and white. In the following, we assume that the specular reflection color $\\mathbf{f}_s$ is known.", + "bbox": [ + 89, + 660, + 483, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With Eqs. (12) and (13), the alpha matte can be estimated by solving the following linear system:", + "bbox": [ + 89, + 767, + 482, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {\\left[ \\begin{array}{l} \\mathbf {s} _ {1} - \\mathbf {q} _ {1} \\\\ \\mathbf {s} _ {2} - \\mathbf {q} _ {2} \\end{array} \\right]} _ {\\mathbf {y}} = \\underbrace {\\left[ \\begin{array}{c c c} - \\mathbf {q} _ {1} & \\mathbf {f} _ {s} & \\mathbf {0} \\\\ - \\mathbf {q} _ {2} & \\mathbf {0} & \\mathbf {f} _ {s} \\end{array} \\right]} _ {\\mathbf {A}} \\underbrace {\\left[ \\begin{array}{c} \\alpha \\\\ \\alpha p _ {s} \\cos 2 \\theta_ {s} \\\\ \\alpha p _ {s} \\sin 2 \\theta_ {s} \\end{array} \\right]} _ {\\mathbf {x}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 806, + 482, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, this linear system has a closed-form solution, and in", + "bbox": [ + 89, + 885, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "practice, the alpha matte can be computed as", + "bbox": [ + 511, + 92, + 812, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = 1 - \\frac {\\mathbf {q} _ {1} ^ {\\top} \\mathbf {s} _ {1} + \\mathbf {q} _ {2} ^ {\\top} \\mathbf {s} _ {2} - \\left(\\mathbf {q} _ {1} ^ {\\top} \\mathbf {f} _ {s}\\right) \\left(\\mathbf {s} _ {1} ^ {\\top} \\mathbf {f} _ {s}\\right) - \\left(\\mathbf {q} _ {2} ^ {\\top} \\mathbf {f} _ {s}\\right) \\left(\\mathbf {s} _ {2} ^ {\\top} \\mathbf {f} _ {s}\\right)}{p _ {b} ^ {2} - \\left(\\mathbf {q} _ {1} ^ {\\top} \\mathbf {f} _ {s}\\right) ^ {2} - \\left(\\mathbf {q} _ {2} ^ {\\top} \\mathbf {f} _ {s}\\right) ^ {2}} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 109, + 906, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "without computing the pseudo inverse matrix of $\\mathbf{A}$ for each pixel of a given image.", + "bbox": [ + 511, + 148, + 906, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1. If $\\mathbf{q} \\neq \\mathbf{f}_s$ , the $p$ CSM-S problem has a closed-form solution.", + "bbox": [ + 511, + 186, + 906, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. The determinant of $\\mathbf{A}^{\\top}\\mathbf{A}$ can be calculated as", + "bbox": [ + 511, + 226, + 870, + 242 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\det \\left(\\mathbf {A} ^ {\\top} \\mathbf {A}\\right) = p _ {b} ^ {2} \\left(1 - \\left(\\mathbf {q} ^ {\\top} \\mathbf {f} _ {s}\\right) ^ {2}\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 252, + 906, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $\\mathbf{q},\\mathbf{f}_s\\geq \\mathbf{0}$ and $\\| \\mathbf{q}\\| _2 = \\| \\mathbf{f}_s\\| _2 = 1$ ,det $(\\mathbf{A}^{\\top}\\mathbf{A}) = 0$ if and only if $\\mathbf{q} = \\mathbf{f}_s$", + "bbox": [ + 511, + 287, + 906, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, our pCSM-S can be always applied by using a backlight color sufficiently different from the ambient light color; for instance, a green background with natural white ambient light.", + "bbox": [ + 511, + 329, + 905, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. pCSM with Polarized Diffuse-Specular Reflection", + "text_level": 1, + "bbox": [ + 511, + 398, + 905, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section introduces another pCSM method, which we call $p\\mathrm{CSM - DS}$ , using the full expression of the diffuse-specular polarimetric compositing equation without any simplification. Research shows that more general modeling leads to more ambiguity [27], and our pCSM-DS problem is no exception. Here, we present the solvability conditions of our pCSM-DS and the solutions when each condition is satisfied. Since the pCSM-DS problem is equivalent to pCSM-S if $\\mathbf{f}_d = \\mathbf{f}_s$ or $p_d = 0$ , we suppose $\\mathbf{f}_d \\neq \\mathbf{f}_s$ and $p_d > 0$ in the following.", + "bbox": [ + 511, + 436, + 906, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 2. If a strong color condition, span $\\{\\mathbf{q},\\mathbf{f}_d,\\mathbf{f}_s\\} = \\mathbb{R}^3$ is satisfied, the pCSM-DS problem has a closed-form solution.", + "bbox": [ + 511, + 598, + 905, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. Since Eq. (6) can be written as", + "bbox": [ + 513, + 652, + 767, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {0} - p _ {b} \\mathbf {q} = \\left[ \\begin{array}{l l l} - p _ {b} \\mathbf {q} & \\mathbf {f} _ {d} & \\mathbf {f} _ {s} \\end{array} \\right] \\left[ \\begin{array}{c} \\alpha \\\\ \\alpha \\left(2 u _ {d} + p _ {d}\\right) \\\\ \\alpha \\left(2 u _ {s} + p _ {s}\\right) \\end{array} \\right], \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 676, + 905, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "if $\\mathbf{q}$ , $\\mathbf{f}_d$ , $\\mathbf{f}_s$ are linearly independent, the matrix in Eq. (17) is invertible, and it has a closed-form solution.", + "bbox": [ + 511, + 734, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The missing variable to solve Eq. (17) is the diffuse color vector $\\mathbf{f}_d$ , which is inherently difficult to be known. However, the use of polarization images with a fully polarized background allows us to compute a linear combination of $\\mathbf{f}_d$ and $\\mathbf{f}_s$ as", + "bbox": [ + 511, + 773, + 906, + 844 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {f} _ {d s} = \\mathbf {c} _ {\\theta_ {b} - \\frac {\\pi}{2}} \\\\ = \\alpha \\left(\\left(u _ {d} + p _ {d} \\cos^ {2} \\left(\\theta_ {s} - \\theta_ {b}\\right)\\right) \\mathbf {f} _ {d} + \\left(u _ {s} + p _ {s} \\sin^ {2} \\left(\\theta_ {s} - \\theta_ {b}\\right)\\right) \\mathbf {f} _ {s}\\right). \\tag {18} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 852, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "394", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We call this vector $\\mathbf{f}_{ds}$ the foreground basis. Since the unpolarized reflection is dominant in the diffuse reflection, the unpolarized diffuse intensity $u_{d}$ is almost always nonzero when the polarized diffuse intensity is positive $p_d > 0$ . Hence, the foreground basis and $\\mathbf{f}_s$ are linearly independent unless $\\mathbf{f}_d = \\mathbf{f}_s$ . With this foreground basis, Eq. (17) can be rewritten as", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {0} - p _ {b} \\mathbf {q} = \\underbrace {\\left[ - p _ {b} \\mathbf {q} \\quad \\mathbf {f} _ {d s} \\quad \\mathbf {f} _ {s} \\right]} _ {\\mathbf {B}} \\left[ \\begin{array}{l} \\alpha \\\\ v _ {d s} \\\\ v _ {s} \\end{array} \\right], \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 205, + 482, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v_{ds}, v_s \\in \\mathbb{R}$ . Therefore, if the matrix $\\mathbf{B}$ is full-rank, i.e., the strong color condition is satisfied, this linear system has a closed-form solution, and the alpha matte is uniquely determined. We call this method linear $pCSM-DS$ .", + "bbox": [ + 89, + 268, + 482, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then consider the case of $\\mathbf{q} \\in \\operatorname{span}\\{\\mathbf{f}_d, \\mathbf{f}_s\\}$ . In this case, Eq. (6) can be satisfied with multiple alpha matte values, hence, we leverage Eqs. (7) and (8) to derive a unique alpha matte.", + "bbox": [ + 89, + 329, + 483, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 3. If a weak color condition, $\\mathbf{q} \\not\\propto p_s\\mathbf{f}_s - p_d\\mathbf{f}_d$ and a polarization condition, $\\theta_b \\neq \\theta_s$ nor $\\theta_b \\neq \\theta_d$ , are satisfied, the $p$ CSM-DS problem has a unique solution, even if the strong color condition is not satisfied.", + "bbox": [ + 89, + 400, + 483, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. See the supplementary material.", + "bbox": [ + 89, + 470, + 354, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 3 shows the well-posedness of the pCSM-DS problem when the weak color condition and polarization condition are satisfied; therefore, here we derive a solution to the pCSM-DS problem. The alpha matte can be found as a minimizer of the following objective function:", + "bbox": [ + 89, + 500, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f (\\alpha , \\mathbf {w}, t) = \\| \\mathbf {s} _ {1} - \\mathbf {d} _ {1} - (1 - \\alpha) \\mathbf {q} _ {1} \\| _ {2} ^ {2} + \\| \\mathbf {s} _ {2} - \\mathbf {d} _ {2} - (1 - \\alpha) \\mathbf {q} _ {2} \\| _ {2} ^ {2} \\\\ = \\| \\mathbf {s} _ {1} - \\mathbf {F} \\mathbf {w} - (1 - \\alpha) \\mathbf {q} _ {1} \\| _ {2} ^ {2} + \\| \\mathbf {s} _ {2} - t \\mathbf {F} \\mathbf {w} - (1 - \\alpha) \\mathbf {q} _ {2} \\| _ {2} ^ {2} \\tag {20} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 584, + 483, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{d}_1$ and $\\mathbf{d}_2$ , defined in Eqs. (10) and (11), are represented by a basis matrix $\\mathbf{F} = [\\mathbf{f}_{ds}\\mathbf{f}_s]\\in \\mathbb{R}^{3\\times 2}$ , coefficient vector $\\mathbf{w} = [w_{ds}w_s]^\\top \\in \\mathbb{R}^2$ , and a scalar $t\\in \\mathbb{R}$ . This objective function is written in matrix form as", + "bbox": [ + 89, + 633, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f (\\alpha , \\mathbf {w}, t) = \\left\\| \\left[ \\begin{array}{l} \\mathbf {s} _ {1} - \\mathbf {q} _ {1} \\\\ \\mathbf {s} _ {2} - \\mathbf {q} _ {2} \\end{array} \\right] - \\left[ \\begin{array}{l l} - \\mathbf {q} _ {1} & \\mathbf {F} \\\\ - \\mathbf {q} _ {2} & t \\mathbf {F} \\end{array} \\right] \\left[ \\begin{array}{l} \\alpha \\\\ \\mathbf {w} \\end{array} \\right] \\right\\| _ {2} ^ {2} \\tag {21} \\\\ \\stackrel {{\\mathrm {d e f}}} {{=}} \\left\\| \\mathbf {y} - \\mathbf {D} (t) \\left[ \\begin{array}{c} \\alpha \\\\ \\mathbf {w} \\end{array} \\right] \\right\\| _ {2} ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 705, + 482, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since $t$ is unknown, the alpha matte cannot be derived in a closed-form manner. However, as shown in Eq. (22), $\\alpha$ and $\\mathbf{w}$ may be eliminated from Eq. (21), leaving a polynomial function only in $t$ .", + "bbox": [ + 89, + 792, + 483, + 853 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/788694c48578e692902a03bb07909d3b4859fdd297e3a37d7c3bff0212762cd7.jpg", + "image_caption": [ + "Figure 4. An illustrative interpretation of pCSM-DS optimization. For a fixed $t$ , the objective function has a least-squares solution (blue points). Projecting the least-squares solutions to the $t - f(\\alpha, \\mathbf{w}, t)$ plane forms $g(t)$ , where the critical points (red and yellow points) can be derived as roots of the derivative $g'(t)$ . The final solution (red point) can be found by assessing the values of $f(\\alpha, \\mathbf{w}, t)$ at all the critical points." + ], + "image_footnote": [], + "bbox": [ + 516, + 97, + 903, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 4. Define $g(t)$ as", + "bbox": [ + 513, + 441, + 718, + 458 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng (t) = \\min _ {\\alpha , \\mathbf {w}} f (\\alpha , \\mathbf {w}, t). \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 468, + 903, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, $g(t)$ is a ratio of polynomials in $t$ as", + "bbox": [ + 513, + 501, + 795, + 516 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng (t) = \\frac {n (t)}{d (t)}, \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 663, + 525, + 903, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n(t)$ and $d(t)$ are fourth-degree polynomials in $t$ .", + "bbox": [ + 511, + 568, + 880, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. See the supplementary material.", + "bbox": [ + 513, + 592, + 777, + 607 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7ed151b7b90a84b84bfc12aec7e1b2adc85f651d193cad847b5556aa60bc5186.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 888, + 593, + 905, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With Proposition 4, we can find critical points of $g(t)$ as roots of the derivative $g'(t)$ and corresponding alpha matte values and coefficient vectors. Noting the range of the alpha matte value, we discard critical points if the corresponding alpha matte values are out of the range [0, 1], forming a set of valid critical points, say $\\mathcal{T}$ . Therefore, the optimal variables can be found as", + "bbox": [ + 511, + 617, + 905, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nt _ {m} = \\underset {t \\in \\mathcal {T}} {\\operatorname {a r g m i n}} g (t), \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 734, + 903, + 757 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {m}, \\mathbf {w} _ {m} = \\underset {\\alpha , \\mathbf {w}} {\\operatorname {a r g m i n}} f (\\alpha , \\mathbf {w}, t = t _ {m}). \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 762, + 903, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An illustrative interpretation of this optimization are shown in Fig. 4.", + "bbox": [ + 511, + 796, + 903, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that we need to assess the boundaries of the alpha matte value separately. Given $\\alpha = \\alpha'$ , the objective function becomes", + "bbox": [ + 511, + 825, + 903, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\alpha^ {\\prime}} (t) = \\min _ {\\mathbf {w}} f (\\alpha = \\alpha^ {\\prime}, \\mathbf {w}, t). \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 882, + 903, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1Note that the foreground basis is not necessary to be captured as one of four polarization images because the four polarization images allows us to compute the polarization image intensity at any angle.", + "bbox": [ + 89, + 862, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "395", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 5. $g_{\\alpha^{\\prime}}(t)$ is a ratio of quadratic functions in $t$", + "bbox": [ + 89, + 90, + 480, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof. See the supplementary material.", + "bbox": [ + 89, + 114, + 354, + 130 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91d07de54d10a712c2d7b3a586190e064552ba763ae80a1ea8fc8156592c22b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 114, + 480, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly with Eqs. (24) and (25), the optimal $t$ and coefficient vector for $g_0(t)$ and $g_{1}(t)$ can be found as $(t_0,\\mathbf{w}_0)$ and $(t_{1},\\mathbf{w}_{1})$ , respectively. Gathering all candidates of the solution in $\\mathcal{V} = \\left\\{\\left(\\alpha_{m},\\mathbf{w}_{m},t_{m}\\right),(0,\\mathbf{w}_{0},t_{0}),(1,\\mathbf{w}_{1},t_{1})\\right\\}$ , the final solution can be written as", + "bbox": [ + 89, + 138, + 483, + 214 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha^ {*}, \\mathbf {w} ^ {*}, t ^ {*} = \\underset {\\alpha , \\mathbf {w}, t \\in \\mathcal {V}} {\\operatorname {a r g m i n}} f (\\alpha , \\mathbf {w}, t). \\tag {27}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 224, + 482, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This pCSM-DS optimizes a non-linear objective function; therefore, we call this method non-linear $p\\mathrm{CSM - DS}$", + "bbox": [ + 89, + 258, + 482, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Polarized Diffuse Regularization", + "text_level": 1, + "bbox": [ + 89, + 297, + 377, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our pCSM-DS is built upon a solid theory; however, we observed cases where the linear and non-linear pCSM-DS yield undesirable noisy alpha mattes. This behavior is due to the overfitting of the diffuse-specular polarimetric compositing equation. For instance, in the linear pCSM-DS, the matrix $\\mathbf{B}$ in Eq. (19) spans the entire RGB space, and $\\mathbf{s}_0$ is perfectly approximated even if $\\mathbf{s}_0$ contains deviations from the model, e.g., noise. To avoid overfitting, we introduce a polarized diffuse regularization term, motivated by a physical observation that the polarized diffuse intensity $p_d$ tends to be small in most scenes. To introduce this regularization into the linear pCSM-DS, we reformulate the linear pCSM-DS with $\\mathbf{s}_1$ and $\\mathbf{s}_2$ as follows:", + "bbox": [ + 89, + 321, + 483, + 517 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {\\left[ \\begin{array}{l} \\mathbf {s} _ {1} - \\mathbf {q} _ {1} \\\\ \\mathbf {s} _ {2} - \\mathbf {q} _ {2} \\end{array} \\right]} _ {\\mathbf {y}} = \\underbrace {\\left[ \\begin{array}{c c c c c} - \\mathbf {q} _ {1} & \\mathbf {f} _ {d s} & \\mathbf {f} _ {s} & \\mathbf {0} & \\mathbf {0} \\\\ - \\mathbf {q} _ {2} & \\mathbf {0} & \\mathbf {0} & \\mathbf {f} _ {d s} & \\mathbf {f} _ {s} \\end{array} \\right]} _ {\\overline {{\\mathbf {B}}}} \\underbrace {\\left[ \\begin{array}{l} \\alpha \\\\ m _ {d s} \\\\ m _ {s} \\\\ n _ {d s} \\\\ n _ {s} \\end{array} \\right]} _ {\\overline {{\\mathbf {x}}}}. \\tag {28}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 523, + 483, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, the polarized diffuse intensity $p_d$ is only included in $m_{ds}$ and $n_{ds}$ , hence, we regularize $m_{ds}$ and $n_{ds}$ to indirectly constrain the polarized diffuse intensity. We use the ridge regression to form the objective function as", + "bbox": [ + 89, + 630, + 483, + 691 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {x}} ^ {*} = \\underset {\\bar {\\mathbf {x}}} {\\operatorname {a r g m i n}} \\| \\mathbf {y} - \\bar {\\mathbf {B}} \\bar {\\mathbf {x}} \\| _ {2} ^ {2} + \\lambda \\| \\mathbf {E} \\bar {\\mathbf {x}} \\| _ {2} ^ {2}, \\tag {29}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 699, + 480, + 724 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} = \\left[ \\begin{array}{l l l l l} 0 & 1 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 1 & 0 \\end{array} \\right], \\tag {30}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 726, + 482, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda \\in \\mathbb{R}_+$ is a regularization parameter. This ridge regression problem can be solved in a closed-form manner.", + "bbox": [ + 89, + 768, + 482, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we introduce a regularization in the non-linear pCSM-DS. In Eq. (21), the polarized diffuse intensity $p_d$ is only included in $w_{ds}$ , hence, we simply regularize $w_{ds}$ and obtain the objective function as", + "bbox": [ + 89, + 799, + 483, + 859 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {f} (\\alpha , \\mathbf {w}, t) = \\left\\| \\mathbf {y} - \\mathbf {D} (t) \\left[ \\begin{array}{l} \\alpha \\\\ \\mathbf {w} \\end{array} \\right] \\right\\| _ {2} ^ {2} + \\mu \\| \\mathbf {e} ^ {\\top} \\mathbf {w} \\| _ {2} ^ {2}, \\tag {31}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 867, + 483, + 905 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{e} = [10]^\\top$ , $\\mu \\in \\mathbb{R}_+$ is a regularization parameter. Since this objective function is a quadratic function with respect to $\\alpha$ and $\\mathbf{w}$ , we can still compute a closed-form expression for", + "bbox": [ + 511, + 89, + 905, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {g} (t) = \\min _ {\\alpha , \\mathbf {w}} \\bar {f} (\\alpha , \\mathbf {w}, t), \\tag {32}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 161, + 906, + 184 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {g} _ {\\alpha^ {\\prime}} (t) = \\min _ {\\mathbf {w}} \\bar {f} (\\alpha = \\alpha^ {\\prime}, \\mathbf {w}, t). \\tag {33}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 188, + 906, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, the optimal alpha matte can be estimated in the same manner with the unregularized non-linear pCSM-DS.", + "bbox": [ + 511, + 219, + 905, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 6. $\\overline{g}(t)$ is a ratio of fourth-degree polynomials in $t$ .", + "bbox": [ + 511, + 258, + 905, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof. See the supplementary material.", + "bbox": [ + 513, + 291, + 777, + 306 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0e396fb0ad517b49e7488443af63e9596d35b66d33d547fbbe03d5879cb57dda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 888, + 291, + 905, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 7. $\\overline{g}_{\\alpha'}(t)$ is a ratio of fourth-degree polynomials in $t$ .", + "bbox": [ + 511, + 316, + 905, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof. See the supplementary material.", + "bbox": [ + 513, + 347, + 777, + 363 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c7ff336401856daeac0694080ab423bf7220aa574aa793c0ad19b2af236b303.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 888, + 348, + 905, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that, similar to pCSM-S, the polarized diffuse regularization makes pCSM-DS applicable to any scene regardless of the color and polarization conditions. We assess the capability of each regularized pCSM-DS using synthetic datasets in Sec. 4.1 to identify the superior method.", + "bbox": [ + 511, + 372, + 905, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 511, + 460, + 750, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic Dataset: We evaluated the capability of each matting method in isolation with four types of synthetic datasets, each of which satisfies different color or polarization conditions as shown in Tab. 1. We generated all the parameters and synthesized $\\mathbf{c}_{\\phi}$ and $\\mathbf{b}_{\\phi}$ following Eqs. (4) and (5). We generated 1000 samples (pixels) for each type of data. We also synthesized noisy datasets, to evaluate the robustness to noise, by adding the Gaussian noise $\\mathbf{n}_{\\phi} \\sim \\mathcal{N}(0, 0.01)$ as $\\mathbf{c}_{\\phi} + \\mathbf{n}_{\\phi}, \\forall \\phi$ . In total, we synthesized eight types of datasets: A, B, C, and D in Tab. 1 with and without noise. The details of how the parameters were generated can be found in the supplementary material.", + "bbox": [ + 511, + 493, + 906, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-world Dataset: For quantitative evaluations, we used the evaluation dataset used in PolarMatte [11], which contains 12 static scenes with a green polarized background. This dataset provides reference alpha mattes computed by triangulation matting [29] for quantitative evaluation. For further visual evaluations, we captured several videos containing various objects such as furry or transparent objects.", + "bbox": [ + 511, + 681, + 905, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hardware Setup: All videos were captured with the Blackfly® S USB3 color polarization camera [13] and an LCD monitor [14] served as the polarized background. The camera's configuration was set to 30 frames per second, 1/30 second exposure time, and f/1.4. We used a color checker board [8] to linearize the camera response function and calibrate the color of illumination.", + "bbox": [ + 511, + 794, + 905, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "396", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/76c6ebcb9429e1c508bc68e48de91c02615e97df2deb4a8dc426139842afb2fc.jpg", + "table_caption": [ + "Table 1. Color and polarization conditions for four types of synthetic datasets: A, B, C, and D." + ], + "table_footnote": [], + "table_body": "
ABCD
Strong color condition✓
Weak color condition✓✓✓
Polarization condition✓✓✓
", + "bbox": [ + 112, + 128, + 464, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameters: Throughout the experiments, we used $\\lambda = 0.01$ and $\\mu = 0.01$ for the polarized diffuse regularization. Since the PolarMatte static scene test data does not provide the color of illumination, we assume the ambient illumination is natural and white, and the specular reflection color is assumed to be $\\mathbf{f}_s = (1 / \\sqrt{3})\\mathbf{1}$ .", + "bbox": [ + 89, + 220, + 483, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparative Methods: We compare our method with polarimetric triangulation matting (PTM) [21] using four polarization images, PolarMatte [11], background matting v2 (BGMv2) [18], and the Keylight plug-in [31] in Adobe After Effects [1]. The Keylight results shown in this paper were done by a Keylight expert in 5 minutes for each image to compare it with fully computational matting methods.", + "bbox": [ + 89, + 320, + 483, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics: We employ three commonly used metrics for evaluating the estimated alpha matte: mean absolute error (MAE), mean squared error (MSE), and gradient error (GRAD). We use MAE unless otherwise specified.", + "bbox": [ + 89, + 433, + 483, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Results on Synthetic Dataset", + "text_level": 1, + "bbox": [ + 89, + 506, + 346, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows the quantitative comparisons on the noise-free and noisy synthetic datasets. The linear pCSM-DS on the noise-free dataset A and the non-linear pCSM-DS on the noise-free datasets A and B correctly estimate the true alpha matte with zero errors—verifying the validity of our theory. Although unregularized pCSM-DS is sensitive to noise, the polarized diffuse regularization improves robustness to noise. Consequently, the regularized non-linear pCSM-DS shows the best performance on noisy datasets A and B, in which the majority of real-world data is classified. On datasets C and D, since even the non-linear pCSM-DS is not solvable, all the methods produce a similar margin of error. PTM critically fails on dataset D because PTM only relies on polarization cues and the polarization condition is not satisfied.", + "bbox": [ + 88, + 529, + 482, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 shows the processing time of BGMv2 and our prototype implementations of matting methods written in Python, where PTM and pCSM-S for GPU processing are implemented using PyTorch. Since our pCSM-S and PTM are only required to perform a simple calculation, they are more than 9 times faster than BGMv2. Since the (regularized) non-linear pCSM-DS needs to find the zeros of high", + "bbox": [ + 89, + 757, + 483, + 863 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/712d1d555b734ef68bb0f1b86ff41e1d0b0577ef1cc77892bd5c21d91998620f.jpg", + "table_caption": [ + "Table 2. Results on the noise-free synthetic data (top) and noisy synthetic data (bottom). Numbers in this table are MAE. See Tab. 1 for the conditions that each data type satisfies." + ], + "table_footnote": [], + "table_body": "
Noise-free dataABCD
PTM83.6478.4616.07117.60
pCSM-S10.9210.7416.0717.87
Linear pCSM-DS0.00NANANA
Non-linear pCSM-DS0.000.00NANA
Regularized linear pCSM-DS6.0110.7416.0717.87
Regularized non-linear pCSM-DS3.945.7421.4817.87
Noisy dataABCD
PTM85.2579.8421.30118.60
pCSM-S20.3821.4323.6425.39
Linear pCSM-DS42.45NANANA
Non-linear pCSM-DS23.2837.48NANA
Regularized linear pCSM-DS19.1721.7324.1125.60
Regularized non-linear pCSM-DS17.7819.6524.1925.82
", + "bbox": [ + 522, + 142, + 898, + 325 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3f292ff0e136aa0b3980d856d9edc27eb829c6d5002dfbf7437af43022105a12.jpg", + "table_caption": [ + "Table 3. Processing time on $2\\mathrm{\\;K}$ and $4\\mathrm{\\;K}$ images." + ], + "table_footnote": [], + "table_body": "
Resolution2K4K
BGMv2 (GPU)20.55 ms39.83 ms
PTM (GPU)1.80 ms6.61 ms
pCSM-S (GPU)2.28 ms8.39 ms
PTM (CPU)0.22 s1.15 s
pCSM-S (CPU)0.36 s1.83 s
Linear pCSM-DS (CPU)70 s259 s
Non-linear pCSM-DS (CPU)71 s276 s
Regularized linear pCSM-DS (CPU)63 s267 s
Regularized non-linear pCSM-DS (CPU)72 s277 s
", + "bbox": [ + 516, + 371, + 906, + 520 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e395515516f4c9f46585189ef0cc1acd3989f8b1930cecd90dceee5ea68be6a1.jpg", + "table_caption": [ + "Table 4. Quantitative evaluations of Keylight, PTM [21], BGMv2 [18], PolarMatte [11], and our pCSM-S and pCSM-DS on the static scene test data with a green background [11]." + ], + "table_footnote": [], + "table_body": "
MAE ↓MSE ↓GRAD ↓
Keylight27.134.7412.18
BGMv285.2056.6212.51
PTM11.261.223.31
PolarMatte10.350.841.97
pCSM-S7.570.291.92
Regularized non-linear pCSM-DS7.480.261.68
", + "bbox": [ + 514, + 593, + 906, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "order polynomials and solve an ordinary least squares problem for every critical point, it shows a less efficient result. We note that while BGMv2 or per-image methods often struggle with dealing with very high-resolution images, all our methods are operated in a per-pixel manner and applicable to any resolution.", + "bbox": [ + 511, + 729, + 906, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experimental results on the synthetic datasets show that the regularized non-linear pCSM-DS is the best choice for offline applications such as matting dataset labeling, and pCSM-S is the best choice for real-time applications such as video conferencing.", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2The processing time was measured using Intel® Xeon® Platinum 8275CL CPU @ 3.00 GHz and NVIDIA ® Tesla® V100.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "397", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff1e952abb082207ac677aad896a560dfb76b97bded4dbc3198e47e43165c511.jpg", + "image_caption": [ + "Figure 5. Visual results of the reference alpha matte computed by triangulation matting (TM) [29], Keylight [31], BGMv2 [18], polarimetric triangulation matting (PTM) [21], PolarMatte [11], and our pCSM-S and Regularized non-linear pCSM-DS on the static scene test data. See more visual results in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 906, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Results on Real-world Dataset", + "text_level": 1, + "bbox": [ + 89, + 598, + 359, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the results on the synthetic datasets, we only applied the regularized non-linear pCSM-DS to the real-world datasets as a representative of pCSM-DS. Table 4 shows quantitative comparisons on the static scene test data. These results suggest that our pCSM-S and pCSM-DS significantly improve the quality of matte extraction. While pCSM-DS shows marginal improvements from pCSM-S because the intensity of polarized diffuse reflections tend to be small in the real-world, visual improvements can be seen in Fig. 5. On the headphone scene, pCSM-S mitigates the errors in PTM and PolarMatte due to polarized reflections, and pCSM-DS further improves the alpha matte by considering polarized diffuse reflections. A similar trend can be seen in the zoomed-in figures of the model airplane scene.", + "bbox": [ + 89, + 628, + 483, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show qualitative comparisons on actual videos in the supplementary video, which demonstrate our accurate and temporally consistent alpha mattes and composite results.", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 597, + 633, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduced Polarized Color Screen Matting, a single-shot, per-pixel matting theory that fully leverages color and polarization cues. The theoretical analysis on the proposed diffuse-specular polarimetric compositing equation derived two matting methods, pCSM-S and pCSM-DS. pCSM-S offers inherently efficient matting by simplifying the compositing equation. pCSM-DS revealed the color and polarization conditions to solve the pCSM problem using the full expression of the compositing equation. By introducing polarized diffuse regularization, pCSM-DS gains robustness to noise and applicability to any scene. Experiments on synthetic and real-world datasets show the stable and accurate matting capability of our methods compared with solely color-based or polarization-based methods. Future work may include more general reflectance models for further generalization and applicability.", + "bbox": [ + 511, + 622, + 906, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: We would like to thank Brian Maffitt for his contributions as the Keylight expert.", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "398", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adobe Inc. Adobe After Effects. https://www.adobe.com/products/aftereffects.html. Accessed: 2025-02-07. 7", + "[2] Yagiz Aksoy, Tunç Ozan Aydin, Marc Pollefeys, and Aljosa Smolić. Interactive high-quality green-screen keying via color unmixing. ACM TOG, 36(4):1, 2016. 1", + "[3] Yunhao Ba, Alex Gilbert, Franklin Wang, Jina Yang, Rui Chen, Yiqin Wang, Lei Yan, Boxin Shi, and Achuta Kadambi. Deep shape from polarization. In ECCV, 2020. 3", + "[4] Seung-Hwan Baek, Daniel S Jeon, Xin Tong, and Min H Kim. Simultaneous acquisition of polarimetric svbrdf and normals. ACM TOG, 37(6):268, 2018. 3", + "[5] Seung-Hwan Baek, Tizian Zeltner, Hyunjin Ku, Inseung Hwang, Xin Tong, Wenzel Jakob, and Min H Kim. Image-based acquisition and modeling of polarimetric reflectance. ACM TOG, 39(4):139, 2020. 7, 8", + "[6] Moshe Ben-Ezra. Segmentation with invisible keying signal. In CVPR, 2000. 2", + "[7] Walter Beyer. Traveling-matte photography and the bluescreen system: a tutorial paper. Journal of the SMPTE, 74 (3):217-239, 1965. 2", + "[8] Calibrite LLC. ColorChecker Classic. https:// calibrite.com/us/product/colorchecker-classic/. Accessed: 2025-02-07. 6", + "[9] Zhaopeng Cui, Jinwei Gu, Boxin Shi, Ping Tan, and Jan Kautz. Polarimetric multi-view stereo. In CVPR, 2017. 3", + "[10] Akshat Dave, Yannick Hold-Geoffroy, Milos Hasan, Kalyan Sunkavalli, and Ashok Veeraraghavan. Snapshot polarimetric diffuse-specular separation. Optics Express, 30(19): 34239-34255, 2022. 3", + "[11] Kenji Enomoto, TJ Rhodes, Brian Price, and Gavin Miller. PolarMatte: Fully computational ground-truth-quality alpha matte extraction for images and video using polarized screen matting. In CVPR, 2024. 2, 3, 4, 6, 7, 8, 9, 10", + "[12] Mikhail Erofeev, Yury Gitman, Dmitriy Vatolin, Alexey Fedorov, and Jue Wang. Perceptually motivated benchmark for video matting. In BMVC, 2015. 2", + "[13] FLIR Systems Inc. BFS-U3-51S5PC-C USB 3.1. https://www.teledynevisionsolutions.com/ products/blackfly-s-usb3/?model=BFS-U3-51S5PC-C. Accessed: 2025-02-07. 6", + "[14] Hisense Group Co., Ltd. 55\" 4K UHD Hisense Android Smart TV (2021). https://www.hisense-usa.com/support/55a6g-55-4k-uhd-hisense android-smart-TV-2021. Accessed: 2025-02-07. 6", + "[15] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 7", + "[16] Soma Kajiyama, Taihe Piao, Ryo Kawahara, and Takahiro Okabe. Separating partially-polarized diffuse and specular reflection components under unpolarized light sources. In WACV, 2023. 3", + "[17] Chenhao Li, Taishi Ono, Takeshi Uemori, Hajime Mihara, Alexander Gatto, Hajime Nagahara, and Yuseke Moriuchi." + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "NeISF: Neural incident stokes field for geometry and material estimation. In CVPR, 2024. 3", + "[18] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In CVPR, 2021. 1, 2, 7, 8, 9, 10", + "[19] Hao Lu, Yutong Dai, Chunhua Shen, and Songcen Xu. Indices matter: Learning to index for deep image matting. In ICCV, 2019. 1", + "[20] Youwei Lyu, Lingran Zhao, Si Li, and Boxin Shi. Shape from polarization with distant lighting estimation. IEEE TPAMI, 45(11):13991-14004, 2023. 3", + "[21] Morgan McGuire and Wojciech Matusik. Real-time triangulation matting using passive polarization. In SIGGRAPH Sketches. 2006. 7, 8, 9, 10", + "[22] Morgan McGuire, Wojciech Matusik, and William Yerazunis. Practical, Real-time Studio Matting using Dual Imagers. In Symposium on Rendering, 2006. 1, 2, 4", + "[23] GyuTae Park, SungJoon Son, JaeYoung Yoo, SeHo Kim, and Nojun Kwak. Matteformer: Transformer-based image matting via prior-tokens. In CVPR, 2022. 1", + "[24] Vlahos Petro. Composite photography utilizing sodium vapor illumination, 1963. US Patent 3,095,304. 2", + "[25] Thomas Porter and Tom Duff. Compositing digital images. SIGGRAPH, 1984. 1", + "[26] Steven A Shafer. Using color to separate reflection components. Color Research & Application, 10(4):210-218, 1985. 4", + "[27] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE TPAMI, 41(2):271-284, 2019. 4", + "[28] Dmitriy Smirnov, Chloe LeGendre, Xueming Yu, and Paul Debevec. Magenta green screen: Spectrally multiplexed alpha matting with deep colorization. In The Digital Production Symposium (DigiPro), 2023. 1, 2", + "[29] Alvy Ray Smith and James F Blinn. Blue screen matting. In SIGGRAPH, 1996. 1, 2, 6, 8, 9, 10", + "[30] George Gabriel Stokes. On the composition and resolution of streams of polarized light from different sources. Transactions of the Cambridge Philosophical Society, 9:399, 1851. 3", + "[31] The Foundry Visionmongers Ltd. Keylight. https://learn.foundry.com/nuke/13.0/content/reference_guide/keyer_nodes/ keylight.html. Accessed: 2025-02-07. 1, 2, 7, 8, 9, 10", + "[32] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In CVPR, 2017. 1", + "[33] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In ACM MM, 2021. 2" + ], + "bbox": [ + 516, + 92, + 903, + 830 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "399", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2025/Polarized Color Screen Matting/4bc8f905-0a33-46ab-838e-6d058731820a_model.json b/2025/Polarized Color Screen Matting/4bc8f905-0a33-46ab-838e-6d058731820a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5be3abfdf44980170878f37da4a117c6a00e17ca --- /dev/null +++ b/2025/Polarized Color Screen Matting/4bc8f905-0a33-46ab-838e-6d058731820a_model.json @@ -0,0 +1,2550 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.338, + 0.131, + 0.66, + 0.154 + ], + "angle": 0, + "content": "Polarized Color Screen Matting" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.181, + 0.361, + 0.2 + ], + "angle": 0, + "content": "Kenji Enomoto" + }, + { + "type": "text", + "bbox": [ + 0.401, + 0.182, + 0.501, + 0.197 + ], + "angle": 0, + "content": "Scott Cohen" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.182, + 0.633, + 0.197 + ], + "angle": 0, + "content": "Brian Price" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.2, + 0.565, + 0.215 + ], + "angle": 0, + "content": "Adobe Research" + }, + { + "type": "text", + "bbox": [ + 0.674, + 0.182, + 0.76, + 0.197 + ], + "angle": 0, + "content": "TJ Rhodes" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.244, + 0.196, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.321, + 0.195, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.407, + 0.228, + 0.421 + ], + "angle": 0, + "content": "(a) Scene" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.244, + 0.295, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.321, + 0.295, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.244, + 0.397, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.321, + 0.397, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.407, + 0.461, + 0.421 + ], + "angle": 0, + "content": "(b) Chroma keying" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.244, + 0.496, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.321, + 0.496, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.244, + 0.597, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.321, + 0.597, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.244, + 0.697, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.321, + 0.697, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.407, + 0.68, + 0.421 + ], + "angle": 0, + "content": "(c) Polarimetric matting" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.244, + 0.798, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.321, + 0.797, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.407, + 0.831, + 0.42 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.244, + 0.898, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.802, + 0.321, + 0.898, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.425, + 0.907, + 0.496 + ], + "angle": 0, + "content": "Figure 1. (a) Scenes with fully polarized blue/green backgrounds under unpolarized known color illumination. (b) Commercial keying tools [31] use color information to extract an alpha matte, often struggling with hair details or motion blur due to the essential ill-posedness of the problem. (c) Polarimetric matting [22] uses polarization cues to extract an alpha matte without manual operations, often struggling with polarized foreground reflections as shown in the specular highlight. (d) Our Polarized Color Screen Matting leverages color and polarization cues, which enables more accurate alpha matte extraction for a wide range of real-world objects." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.518, + 0.326, + 0.533 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.485, + 0.884 + ], + "angle": 0, + "content": "This paper considers the long-standing problem of extracting alpha mattes from video using a known background. While various color-based or polarization-based approaches have been studied in past decades, the problem remains ill-posed because the solutions solely rely on either color or polarization. We introduce Polarized Color Screen Matting, a single-shot, per-pixel matting theory for alpha matte and foreground color recovery using both color and polarization cues. Through a theoretical analysis of our diffuse-specular polarimetric compositing equation, we derive practical closed-form matting methods with their solvability conditions. Our theory concludes that an alpha matte can be extracted without manual corrections using off-the-shelf equipment such as an LCD monitor, polarization camera, and unpolarized lights with calibrated color. Experiments on synthetic and real-world datasets verify the validity of our theory and show the capability of our matting methods on real videos with quantitative and qualitative comparisons to color-based and polarization-based matting methods." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.518, + 0.644, + 0.533 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.907, + 0.67 + ], + "angle": 0, + "content": "Image and video editing rely on compositing [25] to blend images. The quality of the composite depends on how accurately the alpha matte depicts opacity. Modern matte extraction methods use inherent information such as color or polarization to extract the alpha matte. However, they are forced to compromise on quality of foreground color, alpha, or both. This paper combines color and polarimetric observations to constrain the problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Color-based alpha matte extraction methods are most commonly used for compositing due to equipment availability. Chroma keying [2, 29] is the de facto standard in film making, but because it is underconstrained, it often requires manual corrections to extract high-quality alpha mattes. In addition to this, it often struggles with motion blur and transparent objects as depicted in Fig. 1. Recent learning-based matting methods [18, 19, 23, 32] can sometimes extract acceptable mattes but they still struggle with complex situations; potential improvements in modeling and dataset remain. Magenta green screen [28] can extract a more precise alpha matte by sacrificing the green color channel, but this later requires a restoration of the green color channel using machine-learning colorization. All these challenges are rooted in the fact that the use of" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "391" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.472, + 0.106 + ], + "angle": 0, + "content": "color alone to recover an alpha matte is underconstrained." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.107, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Alternatively, polarization-based matting methods [6, 11, 22] can extract the alpha matte using the polarization of light. For a scene backlit with polarized light, there is less ambiguity between phase angle than color, thus physically accurate alpha mattes can be obtained for the majority of pixels. However, in cases where the foreground subject reflects polarized light whose phase angle is comparable to the background, the alpha matte quality will be degraded, as shown in the headphone and glass scenes in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.243, + 0.485, + 0.514 + ], + "angle": 0, + "content": "We introduce Polarized Color Screen Matting (pCSM), a single-shot, per-pixel matting theory that uses both color and polarization cues. Our theory is built upon our proposed diffuse-specular polarimetric compositing equation, where foreground appearance is explicitly modeled by a summation of diffuse and specular reflections. This new compositing equation steps forward a theoretical analysis of the matting problem, resulting in two matting methods: (1) \\( p\\text{CSM-S} \\), a fast and robust matting method with an approximation of unpolarized diffuse reflections that is applicable to any scene, and (2) \\( p\\text{CSM-DS} \\), an accurate and approximation-free matting method. Through our theoretical analysis, we found solvability conditions of pCSM-DS in terms of color and polarized reflection from the foreground subject. For the practical use of our theoretical matting method, we introduce a physics-inspired regularization, which grants our methods significant robustness to noisy real-world observations." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.515, + 0.445, + 0.529 + ], + "angle": 0, + "content": "The main contributions of this paper are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.53, + 0.483, + 0.575 + ], + "angle": 0, + "content": "- A diffuse-specular polarimetric compositing equation that unlocks further theoretical analysis of the polarized screen matting problem." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.575, + 0.483, + 0.605 + ], + "angle": 0, + "content": "- A fast and simple matting method, pCSM-S, that is applicable to any scene." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.606, + 0.483, + 0.636 + ], + "angle": 0, + "content": "- An accurate and approximation-free matting method, pCSM-DS, with solvability conditions." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.53, + 0.483, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.483, + 0.681 + ], + "angle": 0, + "content": "We show quantitatively and qualitatively that our methods produce more accurate results than state-of-the-art solely color-based or polarization-based matting methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.233, + 0.71 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Early methods for computing alpha mattes from video have focused on capturing footage against a solid-colored background [7], often referred to as chroma keying. Indeed, chroma keying is an industry standard, implemented in commercial products [31] and used for generating video matting datasets [18, 33]. However, despite its popularity, chroma keying is underconstrained and often performs poorly unless an expert carefully sets up the capture environment and manually post-processes the matte." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Due to the shortcomings of standard chroma keying, further work has focused on using color to facilitate matte extraction. The sodium vapor matting process [24] constrains" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.41 + ], + "angle": 0, + "content": "the problem by narrowing the spectrum of backlight emitted from a low-pressure sodium vapor lamp. A beamsplitter and bandpass filter allow the background sodium vapor light to be captured on one strip of film, forming an alpha matte, and the foreground subject's color to be captured through a notch reject filter on a second strip of film, blocking the sodium vapor light. More recently, magenta green screen [28] looked to eliminate the need for a specialized camera system by placing a foreground subject strongly lit by magenta light in front of a green background to observe the green spectrum solely at the background. However, the foreground magenta light loses the true foreground color, requiring machine-learning colorization to recover the lost green color of the foreground. In theory, triangulation matting [29] can compute the true alpha matte by taking two (or more) images of a static subject in front of different colored backgrounds to triangulate the alpha matte. However, since the images must be taken at different times to allow the background color to change, triangulation matting is not suitable for dynamic scenes without sacrificing motion blur, such as with stop-motion photography [12]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.905, + 0.558 + ], + "angle": 0, + "content": "Instead of using color to constrain the capture system, polarimetric triangulation matting (PTM) [22] employs polarization. PTM uses a beamsplitter with two polarization filters to simultaneously capture two polarization images of a target scene against a polarized background screen. These images can serve as input for triangulation matting if the foreground reflects only unpolarized light; however, in practice, polarized foreground reflections are observed, which cause perceivable errors in foreground subjects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Addressing the issue of polarized foreground reflections, PolarMatte [11] reformulates the polarized screen matting problem by leveraging four polarization images captured with a modern polarization camera (see Fig. 2). In this framework, an extended robust PTM is first applied to pixels with negligible polarized reflections, and the resulting high-confidence alpha mattes are subsequently interpolated to pixels with polarized foreground reflections using spatial context through a polarimetric matting Laplacian. However, two critical limitations persist. First, because PolarMatte is not formulated as a per-pixel method, establishing solvability conditions remains challenging, thereby complicating both its theoretical analysis and practical reliability. Second, its exclusive reliance on polarization cues (i.e., the use of white polarized background) renders it vulnerable when the polarized reflection aligns with the background polarity—a situation often encountered in the presence of specular highlights (e.g., from jewelry or glossy skin). In this paper, we aim to overcome these limitations by jointly integrating color and polarization cues, which more robustly constrains the matting problem without adding significant complexity to the capture process." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "392" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.483, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.356 + ], + "angle": 0, + "content": "Figure 2. Illustration of our capture setup. A recent polarization camera with a color polarization filter array sensor captures four polarization images in a single-shot. Our method is tailored for the polarization images of a foreground subject with a known polarized background under known color illuminations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.38, + 0.383, + 0.397 + ], + "angle": 0, + "content": "3. Polarized Color Screen Matting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.405, + 0.484, + 0.556 + ], + "angle": 0, + "content": "This section delivers a proof that a single-shot and per-pixel matting problem has a closed-form solution given: (1) the background is known and polarized, (2) the foreground illumination color is known, and (3) a diffuse-specular reflectance model for the foreground subject. Our pCSM theory is tailored for polarization images at four angles \\(\\left\\{0, \\frac{\\pi}{4}, \\frac{\\pi}{2}, \\frac{3\\pi}{4}\\right\\}\\) captured by a recent polarization camera shown in Fig. 2, similarly to PolarMatte [11]. The reader is encouraged to reference Table 5 for the notations used in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.564, + 0.484, + 0.597 + ], + "angle": 0, + "content": "3.1. Diffuse-Specular Polarimetric Compositing Equation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.484, + 0.695 + ], + "angle": 0, + "content": "The matting problem is formulated as a decomposition of the color vector of a given pixel into foreground and background color vectors and the opacity of the foreground, commonly called an alpha matte. In a linear color space, the compositing process of a polarization image is described in the polarimetric compositing equation [11]:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.704, + 0.482, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {\\phi} = \\alpha \\mathbf {f} _ {\\phi} + (1 - \\alpha) \\mathbf {b} _ {\\phi}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.484, + 0.819 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_{\\phi}, \\mathbf{f}_{\\phi}, \\mathbf{b}_{\\phi} \\in \\mathbb{R}_+^3\\) are composite, foreground, and background color polarization vectors at an angle \\(\\phi \\in \\{0, \\frac{\\pi}{4}, \\frac{\\pi}{2}, \\frac{3\\pi}{4}\\}\\), and \\(\\alpha \\in [0,1]\\) is an alpha matte. We employ a diffuse-specular reflectance model [16] on the foreground color, which is the sum of diffuse and specular terms:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.826, + 0.483, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {\\phi} = \\underbrace {\\left(u _ {d} + p _ {d} \\cos^ {2} \\left(\\theta_ {d} - \\phi\\right)\\right) \\mathbf {f} _ {d}} _ {\\text {d i f f u s e t e r m}} + \\underbrace {\\left(u _ {s} + p _ {s} \\cos^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {s}} _ {\\text {s p e c u l a r t e r m}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Here, the diffuse and specular terms are decomposed into unpolarized and polarized terms, where \\( u_{d}, p_{d} \\in \\mathbb{R}_{+} \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "are the intensities of the unpolarized and polarized diffuse reflections, respectively, and \\( u_{s}, p_{s} \\in \\mathbb{R}_{+} \\) are the same for the specular reflection. \\( \\mathbf{f}_d, \\mathbf{f}_s \\in S_+^2 \\subset \\mathbb{R}_+^3 \\) denote unit color vectors of diffuse and specular reflections, respectively. \\( \\theta_d, \\theta_s \\in [0, \\pi) \\) are phase angles of the diffuse and specular reflections, and it is known that \\( \\theta_d = \\theta_s \\pm \\frac{\\pi}{2} \\) under unpolarized illumination [3, 9, 16]. Therefore, the diffuse term in Eq. (2) can be rewritten using \\( \\theta_{s} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.224, + 0.907, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\left(u _ {d} + p _ {d} \\sin^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {d}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.255, + 0.906, + 0.284 + ], + "angle": 0, + "content": "In our problem setting, the background is fully polarized, which can be described in" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.297, + 0.907, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} _ {\\phi} = p _ {b} \\cos^ {2} (\\theta_ {b} - \\phi) \\mathbf {q}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.906, + 0.388 + ], + "angle": 0, + "content": "where \\(p_b \\in \\mathbb{R}_+\\), \\(\\theta_b \\in [0, \\pi)\\), and \\(\\mathbf{q} \\in S_+^2 \\subset \\mathbb{R}_+^3\\) are the intensity, phase angle, and unit color vector of the backlight, respectively. Finally, the polarimetric compositing equation (1) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.398, + 0.905, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {c} _ {\\phi} = \\alpha \\left(\\left(u _ {d} + p _ {d} \\sin^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {d} + \\left(u _ {s} + p _ {s} \\cos^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {s}\\right) \\\\ + (1 - \\alpha) p _ {b} \\cos^ {2} (\\theta_ {b} - \\phi) \\mathbf {q}. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.906, + 0.552 + ], + "angle": 0, + "content": "We call this compositing process the diffuse-specular polarimetric compositing equation. The diffuse-specular reflectance model is widely used in polarimetric vision problems, such as diffuse-specular separation [10, 16], shape-from-polarization [17, 20], and polarimetric BRDF acquisition [4]. We extend it to the compositing equation to represent opacity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.629 + ], + "angle": 0, + "content": "For ease of theoretical analysis, we transform observed polarization images into Stokes parameters [30]. This paper only considers linear polarization, and the Stokes parameters can also be described in the form of compositing equations:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.64, + 0.907, + 0.676 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {0} = \\frac {1}{2} \\sum_ {\\phi} \\mathbf {c} _ {\\phi} \\stackrel {\\mathrm {d e f}} {=} \\mathbf {d} _ {0} + (1 - \\alpha) \\mathbf {q} _ {0}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.681, + 0.905, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {1} = \\mathbf {c} _ {0} - \\mathbf {c} _ {\\frac {\\pi}{2}} \\stackrel {\\mathrm {d e f}} {=} \\mathbf {d} _ {1} + (1 - \\alpha) \\mathbf {q} _ {1}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.702, + 0.905, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {2} = \\mathbf {c} _ {\\frac {\\pi}{4}} - \\mathbf {c} _ {\\frac {3 \\pi}{4}} \\stackrel {\\text {d e f}} {=} \\mathbf {d} _ {2} + (1 - \\alpha) \\mathbf {q} _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.731, + 0.56, + 0.743 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.758, + 0.905, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {0} = \\alpha \\left(\\left(2 u _ {d} + p _ {d}\\right) \\mathbf {f} _ {d} + \\left(2 u _ {s} + p _ {s}\\right) \\mathbf {f} _ {s}\\right), \\quad \\mathbf {q} _ {0} = p _ {b} \\mathbf {q}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.778, + 0.905, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {1} = \\alpha \\cos 2 \\theta_ {s} \\left(p _ {s} \\mathbf {f} _ {s} - p _ {d} \\mathbf {f} _ {d}\\right), \\mathbf {q} _ {1} = p _ {b} \\cos 2 \\theta_ {b} \\mathbf {q}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.797, + 0.905, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {2} = \\alpha \\sin 2 \\theta_ {s} \\left(p _ {s} \\mathbf {f} _ {s} - p _ {d} \\mathbf {f} _ {d}\\right), \\mathbf {q} _ {2} = p _ {b} \\sin 2 \\theta_ {b} \\mathbf {q}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.855 + ], + "angle": 0, + "content": "Here, \\( \\mathbf{s}_0 \\) is equivalent to an ordinary color image, and \\( \\mathbf{s}_1 \\) and \\( \\mathbf{s}_2 \\) represent the polarization properties of an observation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "This paper aims to recover the alpha matte \\(\\alpha\\) given the polarization images \\(\\mathbf{c}_{\\phi}\\) and the background images \\(\\mathbf{b}_{\\phi}\\) for all the angles \\(\\phi \\in \\{0, \\frac{\\pi}{4}, \\frac{\\pi}{2}, \\frac{3\\pi}{4}\\}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "393" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.48, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.314, + 0.485, + 0.385 + ], + "angle": 0, + "content": "Figure 3. (Left) A scene of a headphone in front of a blue polarized background and its polarized intensity computed as \\(\\sqrt{\\mathbf{s}_1^2 + \\mathbf{s}_2^2}\\). (Right) The color distribution of the polarized intensities, showing that the majority of the polarized intensities derive from background and specular reflections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.449, + 0.427 + ], + "angle": 0, + "content": "3.2. pCSM with Polarized Specular Reflection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.601 + ], + "angle": 0, + "content": "This section introduces a pCSM method, which we call \\(p\\mathrm{CSM - S}\\) that highlights the polarized specular reflection. Prior polarimetric matting methods [11, 22] show that polarized reflections from a foreground subject are a major source of error in the alpha matte estimation. In particular, polarized specular reflections have larger intensities than polarized diffuse reflections and a dominant effect on the estimated alpha matte (see Fig. 3). This observation motivates us to simplify the polarimetric compositing equation and Stokes parameters without the diffuse term (i.e., \\(p_d = 0\\)). With this simplification, Eqs. (7) and (8) can be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.614, + 0.483, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {1} = \\alpha p _ {s} \\cos 2 \\theta_ {s} \\mathbf {f} _ {s} + (1 - \\alpha) \\mathbf {q} _ {1}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.632, + 0.483, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {2} = \\alpha p _ {s} \\sin 2 \\theta_ {s} \\mathbf {f} _ {s} + (1 - \\alpha) \\mathbf {q} _ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.484, + 0.766 + ], + "angle": 0, + "content": "Assuming the dichromatic reflectance model [26], the specular reflection color \\(\\mathbf{f}_s\\) is independent of the object's albedo and is equal to the ambient light color. Therefore, we can assume the specular reflection color \\(\\mathbf{f}_s\\) to be known if the ambient light color is calibrated or ambient light is natural and white. In the following, we assume that the specular reflection color \\(\\mathbf{f}_s\\) is known." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.483, + 0.799 + ], + "angle": 0, + "content": "With Eqs. (12) and (13), the alpha matte can be estimated by solving the following linear system:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.808, + 0.483, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\underbrace {\\left[ \\begin{array}{l} \\mathbf {s} _ {1} - \\mathbf {q} _ {1} \\\\ \\mathbf {s} _ {2} - \\mathbf {q} _ {2} \\end{array} \\right]} _ {\\mathbf {y}} = \\underbrace {\\left[ \\begin{array}{c c c} - \\mathbf {q} _ {1} & \\mathbf {f} _ {s} & \\mathbf {0} \\\\ - \\mathbf {q} _ {2} & \\mathbf {0} & \\mathbf {f} _ {s} \\end{array} \\right]} _ {\\mathbf {A}} \\underbrace {\\left[ \\begin{array}{c} \\alpha \\\\ \\alpha p _ {s} \\cos 2 \\theta_ {s} \\\\ \\alpha p _ {s} \\sin 2 \\theta_ {s} \\end{array} \\right]} _ {\\mathbf {x}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Thus, this linear system has a closed-form solution, and in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.813, + 0.107 + ], + "angle": 0, + "content": "practice, the alpha matte can be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.111, + 0.907, + 0.145 + ], + "angle": 0, + "content": "\\[\n\\alpha = 1 - \\frac {\\mathbf {q} _ {1} ^ {\\top} \\mathbf {s} _ {1} + \\mathbf {q} _ {2} ^ {\\top} \\mathbf {s} _ {2} - \\left(\\mathbf {q} _ {1} ^ {\\top} \\mathbf {f} _ {s}\\right) \\left(\\mathbf {s} _ {1} ^ {\\top} \\mathbf {f} _ {s}\\right) - \\left(\\mathbf {q} _ {2} ^ {\\top} \\mathbf {f} _ {s}\\right) \\left(\\mathbf {s} _ {2} ^ {\\top} \\mathbf {f} _ {s}\\right)}{p _ {b} ^ {2} - \\left(\\mathbf {q} _ {1} ^ {\\top} \\mathbf {f} _ {s}\\right) ^ {2} - \\left(\\mathbf {q} _ {2} ^ {\\top} \\mathbf {f} _ {s}\\right) ^ {2}} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.149, + 0.907, + 0.18 + ], + "angle": 0, + "content": "without computing the pseudo inverse matrix of \\(\\mathbf{A}\\) for each pixel of a given image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.188, + 0.907, + 0.219 + ], + "angle": 0, + "content": "Proposition 1. If \\(\\mathbf{q} \\neq \\mathbf{f}_s\\), the \\(p\\)CSM-S problem has a closed-form solution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.227, + 0.871, + 0.243 + ], + "angle": 0, + "content": "Proof. The determinant of \\(\\mathbf{A}^{\\top}\\mathbf{A}\\) can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.253, + 0.907, + 0.279 + ], + "angle": 0, + "content": "\\[\n\\det \\left(\\mathbf {A} ^ {\\top} \\mathbf {A}\\right) = p _ {b} ^ {2} \\left(1 - \\left(\\mathbf {q} ^ {\\top} \\mathbf {f} _ {s}\\right) ^ {2}\\right). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.907, + 0.321 + ], + "angle": 0, + "content": "Since \\(\\mathbf{q},\\mathbf{f}_s\\geq \\mathbf{0}\\) and \\(\\| \\mathbf{q}\\| _2 = \\| \\mathbf{f}_s\\| _2 = 1\\) ,det \\((\\mathbf{A}^{\\top}\\mathbf{A}) = 0\\) if and only if \\(\\mathbf{q} = \\mathbf{f}_s\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.906, + 0.391 + ], + "angle": 0, + "content": "Therefore, our pCSM-S can be always applied by using a backlight color sufficiently different from the ambient light color; for instance, a green background with natural white ambient light." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.399, + 0.906, + 0.43 + ], + "angle": 0, + "content": "3.3. pCSM with Polarized Diffuse-Specular Reflection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.59 + ], + "angle": 0, + "content": "This section introduces another pCSM method, which we call \\(p\\mathrm{CSM - DS}\\), using the full expression of the diffuse-specular polarimetric compositing equation without any simplification. Research shows that more general modeling leads to more ambiguity [27], and our pCSM-DS problem is no exception. Here, we present the solvability conditions of our pCSM-DS and the solutions when each condition is satisfied. Since the pCSM-DS problem is equivalent to pCSM-S if \\(\\mathbf{f}_d = \\mathbf{f}_s\\) or \\(p_d = 0\\), we suppose \\(\\mathbf{f}_d \\neq \\mathbf{f}_s\\) and \\(p_d > 0\\) in the following." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.599, + 0.906, + 0.645 + ], + "angle": 0, + "content": "Proposition 2. If a strong color condition, span \\(\\{\\mathbf{q},\\mathbf{f}_d,\\mathbf{f}_s\\} = \\mathbb{R}^3\\) is satisfied, the pCSM-DS problem has a closed-form solution." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.653, + 0.768, + 0.669 + ], + "angle": 0, + "content": "Proof. Since Eq. (6) can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.677, + 0.906, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {0} - p _ {b} \\mathbf {q} = \\left[ \\begin{array}{l l l} - p _ {b} \\mathbf {q} & \\mathbf {f} _ {d} & \\mathbf {f} _ {s} \\end{array} \\right] \\left[ \\begin{array}{c} \\alpha \\\\ \\alpha \\left(2 u _ {d} + p _ {d}\\right) \\\\ \\alpha \\left(2 u _ {s} + p _ {s}\\right) \\end{array} \\right], \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.906, + 0.765 + ], + "angle": 0, + "content": "if \\(\\mathbf{q}\\), \\(\\mathbf{f}_d\\), \\(\\mathbf{f}_s\\) are linearly independent, the matrix in Eq. (17) is invertible, and it has a closed-form solution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.774, + 0.907, + 0.845 + ], + "angle": 0, + "content": "The missing variable to solve Eq. (17) is the diffuse color vector \\(\\mathbf{f}_d\\), which is inherently difficult to be known. However, the use of polarization images with a fully polarized background allows us to compute a linear combination of \\(\\mathbf{f}_d\\) and \\(\\mathbf{f}_s\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.853, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {f} _ {d s} = \\mathbf {c} _ {\\theta_ {b} - \\frac {\\pi}{2}} \\\\ = \\alpha \\left(\\left(u _ {d} + p _ {d} \\cos^ {2} \\left(\\theta_ {s} - \\theta_ {b}\\right)\\right) \\mathbf {f} _ {d} + \\left(u _ {s} + p _ {s} \\sin^ {2} \\left(\\theta_ {s} - \\theta_ {b}\\right)\\right) \\mathbf {f} _ {s}\\right). \\tag {18} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "394" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.198 + ], + "angle": 0, + "content": "We call this vector \\(\\mathbf{f}_{ds}\\) the foreground basis. Since the unpolarized reflection is dominant in the diffuse reflection, the unpolarized diffuse intensity \\(u_{d}\\) is almost always nonzero when the polarized diffuse intensity is positive \\(p_d > 0\\). Hence, the foreground basis and \\(\\mathbf{f}_s\\) are linearly independent unless \\(\\mathbf{f}_d = \\mathbf{f}_s\\). With this foreground basis, Eq. (17) can be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.206, + 0.483, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {0} - p _ {b} \\mathbf {q} = \\underbrace {\\left[ - p _ {b} \\mathbf {q} \\quad \\mathbf {f} _ {d s} \\quad \\mathbf {f} _ {s} \\right]} _ {\\mathbf {B}} \\left[ \\begin{array}{l} \\alpha \\\\ v _ {d s} \\\\ v _ {s} \\end{array} \\right], \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.483, + 0.329 + ], + "angle": 0, + "content": "where \\( v_{ds}, v_s \\in \\mathbb{R} \\). Therefore, if the matrix \\( \\mathbf{B} \\) is full-rank, i.e., the strong color condition is satisfied, this linear system has a closed-form solution, and the alpha matte is uniquely determined. We call this method linear \\( pCSM-DS \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.484, + 0.39 + ], + "angle": 0, + "content": "We then consider the case of \\(\\mathbf{q} \\in \\operatorname{span}\\{\\mathbf{f}_d, \\mathbf{f}_s\\}\\). In this case, Eq. (6) can be satisfied with multiple alpha matte values, hence, we leverage Eqs. (7) and (8) to derive a unique alpha matte." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.462 + ], + "angle": 0, + "content": "Proposition 3. If a weak color condition, \\(\\mathbf{q} \\not\\propto p_s\\mathbf{f}_s - p_d\\mathbf{f}_d\\) and a polarization condition, \\(\\theta_b \\neq \\theta_s\\) nor \\(\\theta_b \\neq \\theta_d\\), are satisfied, the \\(p\\)CSM-DS problem has a unique solution, even if the strong color condition is not satisfied." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.472, + 0.355, + 0.487 + ], + "angle": 0, + "content": "Proof. See the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.484, + 0.576 + ], + "angle": 0, + "content": "Proposition 3 shows the well-posedness of the pCSM-DS problem when the weak color condition and polarization condition are satisfied; therefore, here we derive a solution to the pCSM-DS problem. The alpha matte can be found as a minimizer of the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.585, + 0.484, + 0.633 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f (\\alpha , \\mathbf {w}, t) = \\| \\mathbf {s} _ {1} - \\mathbf {d} _ {1} - (1 - \\alpha) \\mathbf {q} _ {1} \\| _ {2} ^ {2} + \\| \\mathbf {s} _ {2} - \\mathbf {d} _ {2} - (1 - \\alpha) \\mathbf {q} _ {2} \\| _ {2} ^ {2} \\\\ = \\| \\mathbf {s} _ {1} - \\mathbf {F} \\mathbf {w} - (1 - \\alpha) \\mathbf {q} _ {1} \\| _ {2} ^ {2} + \\| \\mathbf {s} _ {2} - t \\mathbf {F} \\mathbf {w} - (1 - \\alpha) \\mathbf {q} _ {2} \\| _ {2} ^ {2} \\tag {20} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.484, + 0.697 + ], + "angle": 0, + "content": "where \\(\\mathbf{d}_1\\) and \\(\\mathbf{d}_2\\), defined in Eqs. (10) and (11), are represented by a basis matrix \\(\\mathbf{F} = [\\mathbf{f}_{ds}\\mathbf{f}_s]\\in \\mathbb{R}^{3\\times 2}\\), coefficient vector \\(\\mathbf{w} = [w_{ds}w_s]^\\top \\in \\mathbb{R}^2\\), and a scalar \\(t\\in \\mathbb{R}\\). This objective function is written in matrix form as" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.707, + 0.483, + 0.783 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f (\\alpha , \\mathbf {w}, t) = \\left\\| \\left[ \\begin{array}{l} \\mathbf {s} _ {1} - \\mathbf {q} _ {1} \\\\ \\mathbf {s} _ {2} - \\mathbf {q} _ {2} \\end{array} \\right] - \\left[ \\begin{array}{l l} - \\mathbf {q} _ {1} & \\mathbf {F} \\\\ - \\mathbf {q} _ {2} & t \\mathbf {F} \\end{array} \\right] \\left[ \\begin{array}{l} \\alpha \\\\ \\mathbf {w} \\end{array} \\right] \\right\\| _ {2} ^ {2} \\tag {21} \\\\ \\stackrel {{\\mathrm {d e f}}} {{=}} \\left\\| \\mathbf {y} - \\mathbf {D} (t) \\left[ \\begin{array}{c} \\alpha \\\\ \\mathbf {w} \\end{array} \\right] \\right\\| _ {2} ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Since \\( t \\) is unknown, the alpha matte cannot be derived in a closed-form manner. However, as shown in Eq. (22), \\( \\alpha \\) and \\( \\mathbf{w} \\) may be eliminated from Eq. (21), leaving a polynomial function only in \\( t \\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.484, + 0.902 + ], + "angle": 0, + "content": "1Note that the foreground basis is not necessary to be captured as one of four polarization images because the four polarization images allows us to compute the polarization image intensity at any angle." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.098, + 0.905, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.32, + 0.908, + 0.418 + ], + "angle": 0, + "content": "Figure 4. An illustrative interpretation of pCSM-DS optimization. For a fixed \\( t \\), the objective function has a least-squares solution (blue points). Projecting the least-squares solutions to the \\( t - f(\\alpha, \\mathbf{w}, t) \\) plane forms \\( g(t) \\), where the critical points (red and yellow points) can be derived as roots of the derivative \\( g'(t) \\). The final solution (red point) can be found by assessing the values of \\( f(\\alpha, \\mathbf{w}, t) \\) at all the critical points." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.443, + 0.719, + 0.459 + ], + "angle": 0, + "content": "Proposition 4. Define \\( g(t) \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.469, + 0.905, + 0.49 + ], + "angle": 0, + "content": "\\[\ng (t) = \\min _ {\\alpha , \\mathbf {w}} f (\\alpha , \\mathbf {w}, t). \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.502, + 0.797, + 0.517 + ], + "angle": 0, + "content": "Then, \\( g(t) \\) is a ratio of polynomials in \\( t \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.664, + 0.526, + 0.905, + 0.56 + ], + "angle": 0, + "content": "\\[\ng (t) = \\frac {n (t)}{d (t)}, \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.569, + 0.882, + 0.585 + ], + "angle": 0, + "content": "where \\( n(t) \\) and \\( d(t) \\) are fourth-degree polynomials in \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.593, + 0.779, + 0.608 + ], + "angle": 0, + "content": "Proof. See the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.889, + 0.594, + 0.906, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.906, + 0.723 + ], + "angle": 0, + "content": "With Proposition 4, we can find critical points of \\( g(t) \\) as roots of the derivative \\( g'(t) \\) and corresponding alpha matte values and coefficient vectors. Noting the range of the alpha matte value, we discard critical points if the corresponding alpha matte values are out of the range [0, 1], forming a set of valid critical points, say \\( \\mathcal{T} \\). Therefore, the optimal variables can be found as" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.735, + 0.905, + 0.758 + ], + "angle": 0, + "content": "\\[\nt _ {m} = \\underset {t \\in \\mathcal {T}} {\\operatorname {a r g m i n}} g (t), \\tag {24}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.763, + 0.905, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {m}, \\mathbf {w} _ {m} = \\underset {\\alpha , \\mathbf {w}} {\\operatorname {a r g m i n}} f (\\alpha , \\mathbf {w}, t = t _ {m}). \\tag {25}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.797, + 0.905, + 0.826 + ], + "angle": 0, + "content": "An illustrative interpretation of this optimization are shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.827, + 0.905, + 0.871 + ], + "angle": 0, + "content": "Note that we need to assess the boundaries of the alpha matte value separately. Given \\(\\alpha = \\alpha'\\), the objective function becomes" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.883, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\ng _ {\\alpha^ {\\prime}} (t) = \\min _ {\\mathbf {w}} f (\\alpha = \\alpha^ {\\prime}, \\mathbf {w}, t). \\tag {26}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "395" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.108 + ], + "angle": 0, + "content": "Proposition 5. \\(g_{\\alpha^{\\prime}}(t)\\) is a ratio of quadratic functions in \\(t\\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.355, + 0.131 + ], + "angle": 0, + "content": "Proof. See the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.466, + 0.116, + 0.482, + 0.128 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.14, + 0.485, + 0.215 + ], + "angle": 0, + "content": "Similarly with Eqs. (24) and (25), the optimal \\(t\\) and coefficient vector for \\(g_0(t)\\) and \\(g_{1}(t)\\) can be found as \\((t_0,\\mathbf{w}_0)\\) and \\((t_{1},\\mathbf{w}_{1})\\), respectively. Gathering all candidates of the solution in \\(\\mathcal{V} = \\left\\{\\left(\\alpha_{m},\\mathbf{w}_{m},t_{m}\\right),(0,\\mathbf{w}_{0},t_{0}),(1,\\mathbf{w}_{1},t_{1})\\right\\}\\), the final solution can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.225, + 0.483, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\alpha^ {*}, \\mathbf {w} ^ {*}, t ^ {*} = \\underset {\\alpha , \\mathbf {w}, t \\in \\mathcal {V}} {\\operatorname {a r g m i n}} f (\\alpha , \\mathbf {w}, t). \\tag {27}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.29 + ], + "angle": 0, + "content": "This pCSM-DS optimizes a non-linear objective function; therefore, we call this method non-linear \\(p\\mathrm{CSM - DS}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.298, + 0.379, + 0.314 + ], + "angle": 0, + "content": "3.4. Polarized Diffuse Regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.518 + ], + "angle": 0, + "content": "Our pCSM-DS is built upon a solid theory; however, we observed cases where the linear and non-linear pCSM-DS yield undesirable noisy alpha mattes. This behavior is due to the overfitting of the diffuse-specular polarimetric compositing equation. For instance, in the linear pCSM-DS, the matrix \\(\\mathbf{B}\\) in Eq. (19) spans the entire RGB space, and \\(\\mathbf{s}_0\\) is perfectly approximated even if \\(\\mathbf{s}_0\\) contains deviations from the model, e.g., noise. To avoid overfitting, we introduce a polarized diffuse regularization term, motivated by a physical observation that the polarized diffuse intensity \\(p_d\\) tends to be small in most scenes. To introduce this regularization into the linear pCSM-DS, we reformulate the linear pCSM-DS with \\(\\mathbf{s}_1\\) and \\(\\mathbf{s}_2\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.525, + 0.484, + 0.624 + ], + "angle": 0, + "content": "\\[\n\\underbrace {\\left[ \\begin{array}{l} \\mathbf {s} _ {1} - \\mathbf {q} _ {1} \\\\ \\mathbf {s} _ {2} - \\mathbf {q} _ {2} \\end{array} \\right]} _ {\\mathbf {y}} = \\underbrace {\\left[ \\begin{array}{c c c c c} - \\mathbf {q} _ {1} & \\mathbf {f} _ {d s} & \\mathbf {f} _ {s} & \\mathbf {0} & \\mathbf {0} \\\\ - \\mathbf {q} _ {2} & \\mathbf {0} & \\mathbf {0} & \\mathbf {f} _ {d s} & \\mathbf {f} _ {s} \\end{array} \\right]} _ {\\overline {{\\mathbf {B}}}} \\underbrace {\\left[ \\begin{array}{l} \\alpha \\\\ m _ {d s} \\\\ m _ {s} \\\\ n _ {d s} \\\\ n _ {s} \\end{array} \\right]} _ {\\overline {{\\mathbf {x}}}}. \\tag {28}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.484, + 0.692 + ], + "angle": 0, + "content": "Here, the polarized diffuse intensity \\( p_d \\) is only included in \\( m_{ds} \\) and \\( n_{ds} \\), hence, we regularize \\( m_{ds} \\) and \\( n_{ds} \\) to indirectly constrain the polarized diffuse intensity. We use the ridge regression to form the objective function as" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.7, + 0.482, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {x}} ^ {*} = \\underset {\\bar {\\mathbf {x}}} {\\operatorname {a r g m i n}} \\| \\mathbf {y} - \\bar {\\mathbf {B}} \\bar {\\mathbf {x}} \\| _ {2} ^ {2} + \\lambda \\| \\mathbf {E} \\bar {\\mathbf {x}} \\| _ {2} ^ {2}, \\tag {29}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.727, + 0.483, + 0.762 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} = \\left[ \\begin{array}{l l l l l} 0 & 1 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 1 & 0 \\end{array} \\right], \\tag {30}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.769, + 0.483, + 0.799 + ], + "angle": 0, + "content": "where \\(\\lambda \\in \\mathbb{R}_+\\) is a regularization parameter. This ridge regression problem can be solved in a closed-form manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.484, + 0.86 + ], + "angle": 0, + "content": "Next, we introduce a regularization in the non-linear pCSM-DS. In Eq. (21), the polarized diffuse intensity \\(p_d\\) is only included in \\(w_{ds}\\), hence, we simply regularize \\(w_{ds}\\) and obtain the objective function as" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.868, + 0.484, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\bar {f} (\\alpha , \\mathbf {w}, t) = \\left\\| \\mathbf {y} - \\mathbf {D} (t) \\left[ \\begin{array}{l} \\alpha \\\\ \\mathbf {w} \\end{array} \\right] \\right\\| _ {2} ^ {2} + \\mu \\| \\mathbf {e} ^ {\\top} \\mathbf {w} \\| _ {2} ^ {2}, \\tag {31}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.906, + 0.152 + ], + "angle": 0, + "content": "where \\(\\mathbf{e} = [10]^\\top\\), \\(\\mu \\in \\mathbb{R}_+\\) is a regularization parameter. Since this objective function is a quadratic function with respect to \\(\\alpha\\) and \\(\\mathbf{w}\\), we can still compute a closed-form expression for" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.162, + 0.907, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\bar {g} (t) = \\min _ {\\alpha , \\mathbf {w}} \\bar {f} (\\alpha , \\mathbf {w}, t), \\tag {32}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.189, + 0.907, + 0.212 + ], + "angle": 0, + "content": "\\[\n\\bar {g} _ {\\alpha^ {\\prime}} (t) = \\min _ {\\mathbf {w}} \\bar {f} (\\alpha = \\alpha^ {\\prime}, \\mathbf {w}, t). \\tag {33}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.22, + 0.906, + 0.251 + ], + "angle": 0, + "content": "Therefore, the optimal alpha matte can be estimated in the same manner with the unregularized non-linear pCSM-DS." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.906, + 0.288 + ], + "angle": 0, + "content": "Proposition 6. \\(\\overline{g}(t)\\) is a ratio of fourth-degree polynomials in \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.292, + 0.779, + 0.307 + ], + "angle": 0, + "content": "Proof. See the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.889, + 0.292, + 0.906, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.317, + 0.906, + 0.345 + ], + "angle": 0, + "content": "Proposition 7. \\(\\overline{g}_{\\alpha'}(t)\\) is a ratio of fourth-degree polynomials in \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.348, + 0.779, + 0.364 + ], + "angle": 0, + "content": "Proof. See the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.889, + 0.349, + 0.906, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.906, + 0.45 + ], + "angle": 0, + "content": "Note that, similar to pCSM-S, the polarized diffuse regularization makes pCSM-DS applicable to any scene regardless of the color and polarization conditions. We assess the capability of each regularized pCSM-DS using synthetic datasets in Sec. 4.1 to identify the superior method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.462, + 0.75, + 0.479 + ], + "angle": 0, + "content": "4. Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.907, + 0.676 + ], + "angle": 0, + "content": "Synthetic Dataset: We evaluated the capability of each matting method in isolation with four types of synthetic datasets, each of which satisfies different color or polarization conditions as shown in Tab. 1. We generated all the parameters and synthesized \\(\\mathbf{c}_{\\phi}\\) and \\(\\mathbf{b}_{\\phi}\\) following Eqs. (4) and (5). We generated 1000 samples (pixels) for each type of data. We also synthesized noisy datasets, to evaluate the robustness to noise, by adding the Gaussian noise \\(\\mathbf{n}_{\\phi} \\sim \\mathcal{N}(0, 0.01)\\) as \\(\\mathbf{c}_{\\phi} + \\mathbf{n}_{\\phi}, \\forall \\phi\\). In total, we synthesized eight types of datasets: A, B, C, and D in Tab. 1 with and without noise. The details of how the parameters were generated can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.906, + 0.789 + ], + "angle": 0, + "content": "Real-world Dataset: For quantitative evaluations, we used the evaluation dataset used in PolarMatte [11], which contains 12 static scenes with a green polarized background. This dataset provides reference alpha mattes computed by triangulation matting [29] for quantitative evaluation. For further visual evaluations, we captured several videos containing various objects such as furry or transparent objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Hardware Setup: All videos were captured with the Blackfly® S USB3 color polarization camera [13] and an LCD monitor [14] served as the polarized background. The camera's configuration was set to 30 frames per second, 1/30 second exposure time, and f/1.4. We used a color checker board [8] to linearize the camera response function and calibrate the color of illumination." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "396" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.119 + ], + "angle": 0, + "content": "Table 1. Color and polarization conditions for four types of synthetic datasets: A, B, C, and D." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.129, + 0.465, + 0.197 + ], + "angle": 0, + "content": "
ABCD
Strong color condition✓
Weak color condition✓✓✓
Polarization condition✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.485, + 0.316 + ], + "angle": 0, + "content": "Hyperparameters: Throughout the experiments, we used \\(\\lambda = 0.01\\) and \\(\\mu = 0.01\\) for the polarized diffuse regularization. Since the PolarMatte static scene test data does not provide the color of illumination, we assume the ambient illumination is natural and white, and the specular reflection color is assumed to be \\(\\mathbf{f}_s = (1 / \\sqrt{3})\\mathbf{1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.429 + ], + "angle": 0, + "content": "Comparative Methods: We compare our method with polarimetric triangulation matting (PTM) [21] using four polarization images, PolarMatte [11], background matting v2 (BGMv2) [18], and the Keylight plug-in [31] in Adobe After Effects [1]. The Keylight results shown in this paper were done by a Keylight expert in 5 minutes for each image to compare it with fully computational matting methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.485, + 0.496 + ], + "angle": 0, + "content": "Evaluation Metrics: We employ three commonly used metrics for evaluating the estimated alpha matte: mean absolute error (MAE), mean squared error (MSE), and gradient error (GRAD). We use MAE unless otherwise specified." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.507, + 0.347, + 0.523 + ], + "angle": 0, + "content": "4.1. Results on Synthetic Dataset" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.53, + 0.483, + 0.756 + ], + "angle": 0, + "content": "Table 2 shows the quantitative comparisons on the noise-free and noisy synthetic datasets. The linear pCSM-DS on the noise-free dataset A and the non-linear pCSM-DS on the noise-free datasets A and B correctly estimate the true alpha matte with zero errors—verifying the validity of our theory. Although unregularized pCSM-DS is sensitive to noise, the polarized diffuse regularization improves robustness to noise. Consequently, the regularized non-linear pCSM-DS shows the best performance on noisy datasets A and B, in which the majority of real-world data is classified. On datasets C and D, since even the non-linear pCSM-DS is not solvable, all the methods produce a similar margin of error. PTM critically fails on dataset D because PTM only relies on polarization cues and the polarization condition is not satisfied." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Table 3 shows the processing time of BGMv2 and our prototype implementations of matting methods written in Python, where PTM and pCSM-S for GPU processing are implemented using PyTorch. Since our pCSM-S and PTM are only required to perform a simple calculation, they are more than 9 times faster than BGMv2. Since the (regularized) non-linear pCSM-DS needs to find the zeros of high" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 2. Results on the noise-free synthetic data (top) and noisy synthetic data (bottom). Numbers in this table are MAE. See Tab. 1 for the conditions that each data type satisfies." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.143, + 0.9, + 0.326 + ], + "angle": 0, + "content": "
Noise-free dataABCD
PTM83.6478.4616.07117.60
pCSM-S10.9210.7416.0717.87
Linear pCSM-DS0.00NANANA
Non-linear pCSM-DS0.000.00NANA
Regularized linear pCSM-DS6.0110.7416.0717.87
Regularized non-linear pCSM-DS3.945.7421.4817.87
Noisy dataABCD
PTM85.2579.8421.30118.60
pCSM-S20.3821.4323.6425.39
Linear pCSM-DS42.45NANANA
Non-linear pCSM-DS23.2837.48NANA
Regularized linear pCSM-DS19.1721.7324.1125.60
Regularized non-linear pCSM-DS17.7819.6524.1925.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.567, + 0.346, + 0.852, + 0.361 + ], + "angle": 0, + "content": "Table 3. Processing time on \\( 2\\mathrm{\\;K} \\) and \\( 4\\mathrm{\\;K} \\) images." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.372, + 0.907, + 0.521 + ], + "angle": 0, + "content": "
Resolution2K4K
BGMv2 (GPU)20.55 ms39.83 ms
PTM (GPU)1.80 ms6.61 ms
pCSM-S (GPU)2.28 ms8.39 ms
PTM (CPU)0.22 s1.15 s
pCSM-S (CPU)0.36 s1.83 s
Linear pCSM-DS (CPU)70 s259 s
Non-linear pCSM-DS (CPU)71 s276 s
Regularized linear pCSM-DS (CPU)63 s267 s
Regularized non-linear pCSM-DS (CPU)72 s277 s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.541, + 0.907, + 0.585 + ], + "angle": 0, + "content": "Table 4. Quantitative evaluations of Keylight, PTM [21], BGMv2 [18], PolarMatte [11], and our pCSM-S and pCSM-DS on the static scene test data with a green background [11]." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.594, + 0.907, + 0.699 + ], + "angle": 0, + "content": "
MAE ↓MSE ↓GRAD ↓
Keylight27.134.7412.18
BGMv285.2056.6212.51
PTM11.261.223.31
PolarMatte10.350.841.97
pCSM-S7.570.291.92
Regularized non-linear pCSM-DS7.480.261.68
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.907, + 0.822 + ], + "angle": 0, + "content": "order polynomials and solve an ordinary least squares problem for every critical point, it shows a less efficient result. We note that while BGMv2 or per-image methods often struggle with dealing with very high-resolution images, all our methods are operated in a per-pixel manner and applicable to any resolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The experimental results on the synthetic datasets show that the regularized non-linear pCSM-DS is the best choice for offline applications such as matting dataset labeling, and pCSM-S is the best choice for real-time applications such as video conferencing." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2The processing time was measured using Intel® Xeon® Platinum 8275CL CPU @ 3.00 GHz and NVIDIA ® Tesla® V100." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "397" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.907, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.542, + 0.907, + 0.585 + ], + "angle": 0, + "content": "Figure 5. Visual results of the reference alpha matte computed by triangulation matting (TM) [29], Keylight [31], BGMv2 [18], polarimetric triangulation matting (PTM) [21], PolarMatte [11], and our pCSM-S and Regularized non-linear pCSM-DS on the static scene test data. See more visual results in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.599, + 0.36, + 0.615 + ], + "angle": 0, + "content": "4.2. Results on Real-world Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.842 + ], + "angle": 0, + "content": "Given the results on the synthetic datasets, we only applied the regularized non-linear pCSM-DS to the real-world datasets as a representative of pCSM-DS. Table 4 shows quantitative comparisons on the static scene test data. These results suggest that our pCSM-S and pCSM-DS significantly improve the quality of matte extraction. While pCSM-DS shows marginal improvements from pCSM-S because the intensity of polarized diffuse reflections tend to be small in the real-world, visual improvements can be seen in Fig. 5. On the headphone scene, pCSM-S mitigates the errors in PTM and PolarMatte due to polarized reflections, and pCSM-DS further improves the alpha matte by considering polarized diffuse reflections. A similar trend can be seen in the zoomed-in figures of the model airplane scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We show qualitative comparisons on actual videos in the supplementary video, which demonstrate our accurate and temporally consistent alpha mattes and composite results." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.598, + 0.634, + 0.614 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.865 + ], + "angle": 0, + "content": "This paper introduced Polarized Color Screen Matting, a single-shot, per-pixel matting theory that fully leverages color and polarization cues. The theoretical analysis on the proposed diffuse-specular polarimetric compositing equation derived two matting methods, pCSM-S and pCSM-DS. pCSM-S offers inherently efficient matting by simplifying the compositing equation. pCSM-DS revealed the color and polarization conditions to solve the pCSM problem using the full expression of the compositing equation. By introducing polarized diffuse regularization, pCSM-DS gains robustness to noise and applicability to any scene. Experiments on synthetic and real-world datasets show the stable and accurate matting capability of our methods compared with solely color-based or polarization-based methods. Future work may include more general reflectance models for further generalization and applicability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements: We would like to thank Brian Maffitt for his contributions as the Keylight expert." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "398" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.155 + ], + "angle": 0, + "content": "[1] Adobe Inc. Adobe After Effects. https://www.adobe.com/products/aftereffects.html. Accessed: 2025-02-07. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.483, + 0.199 + ], + "angle": 0, + "content": "[2] Yagiz Aksoy, Tunç Ozan Aydin, Marc Pollefeys, and Aljosa Smolić. Interactive high-quality green-screen keying via color unmixing. ACM TOG, 36(4):1, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.201, + 0.482, + 0.253 + ], + "angle": 0, + "content": "[3] Yunhao Ba, Alex Gilbert, Franklin Wang, Jina Yang, Rui Chen, Yiqin Wang, Lei Yan, Boxin Shi, and Achuta Kadambi. Deep shape from polarization. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.482, + 0.297 + ], + "angle": 0, + "content": "[4] Seung-Hwan Baek, Daniel S Jeon, Xin Tong, and Min H Kim. Simultaneous acquisition of polarimetric svbrdf and normals. ACM TOG, 37(6):268, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[5] Seung-Hwan Baek, Tizian Zeltner, Hyunjin Ku, Inseung Hwang, Xin Tong, Wenzel Jakob, and Min H Kim. Image-based acquisition and modeling of polarimetric reflectance. ACM TOG, 39(4):139, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[6] Moshe Ben-Ezra. Segmentation with invisible keying signal. In CVPR, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[7] Walter Beyer. Traveling-matte photography and the bluescreen system: a tutorial paper. Journal of the SMPTE, 74 (3):217-239, 1965. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.482, + 0.465 + ], + "angle": 0, + "content": "[8] Calibrite LLC. ColorChecker Classic. https:// calibrite.com/us/product/colorchecker-classic/. Accessed: 2025-02-07. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[9] Zhaopeng Cui, Jinwei Gu, Boxin Shi, Ping Tan, and Jan Kautz. Polarimetric multi-view stereo. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.495, + 0.482, + 0.549 + ], + "angle": 0, + "content": "[10] Akshat Dave, Yannick Hold-Geoffroy, Milos Hasan, Kalyan Sunkavalli, and Ashok Veeraraghavan. Snapshot polarimetric diffuse-specular separation. Optics Express, 30(19): 34239-34255, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.551, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[11] Kenji Enomoto, TJ Rhodes, Brian Price, and Gavin Miller. PolarMatte: Fully computational ground-truth-quality alpha matte extraction for images and video using polarized screen matting. In CVPR, 2024. 2, 3, 4, 6, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[12] Mikhail Erofeev, Yury Gitman, Dmitriy Vatolin, Alexey Fedorov, and Jue Wang. Perceptually motivated benchmark for video matting. In BMVC, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[13] FLIR Systems Inc. BFS-U3-51S5PC-C USB 3.1. https://www.teledynevisionsolutions.com/ products/blackfly-s-usb3/?model=BFS-U3-51S5PC-C. Accessed: 2025-02-07. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[14] Hisense Group Co., Ltd. 55\" 4K UHD Hisense Android Smart TV (2021). https://www.hisense-usa.com/support/55a6g-55-4k-uhd-hisense android-smart-TV-2021. Accessed: 2025-02-07. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[15] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[16] Soma Kajiyama, Taihe Piao, Ryo Kawahara, and Takahiro Okabe. Separating partially-polarized diffuse and specular reflection components under unpolarized light sources. In WACV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[17] Chenhao Li, Taishi Ono, Takeshi Uemori, Hajime Mihara, Alexander Gatto, Hajime Nagahara, and Yuseke Moriuchi." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "NeISF: Neural incident stokes field for geometry and material estimation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[18] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In CVPR, 2021. 1, 2, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[19] Hao Lu, Yutong Dai, Chunhua Shen, and Songcen Xu. Indices matter: Learning to index for deep image matting. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[20] Youwei Lyu, Lingran Zhao, Si Li, and Boxin Shi. Shape from polarization with distant lighting estimation. IEEE TPAMI, 45(11):13991-14004, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[21] Morgan McGuire and Wojciech Matusik. Real-time triangulation matting using passive polarization. In SIGGRAPH Sketches. 2006. 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[22] Morgan McGuire, Wojciech Matusik, and William Yerazunis. Practical, Real-time Studio Matting using Dual Imagers. In Symposium on Rendering, 2006. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[23] GyuTae Park, SungJoon Son, JaeYoung Yoo, SeHo Kim, and Nojun Kwak. Matteformer: Transformer-based image matting via prior-tokens. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.904, + 0.42 + ], + "angle": 0, + "content": "[24] Vlahos Petro. Composite photography utilizing sodium vapor illumination, 1963. US Patent 3,095,304. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.422, + 0.904, + 0.448 + ], + "angle": 0, + "content": "[25] Thomas Porter and Tom Duff. Compositing digital images. SIGGRAPH, 1984. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.45, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[26] Steven A Shafer. Using color to separate reflection components. Color Research & Application, 10(4):210-218, 1985. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[27] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE TPAMI, 41(2):271-284, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.549, + 0.905, + 0.605 + ], + "angle": 0, + "content": "[28] Dmitriy Smirnov, Chloe LeGendre, Xueming Yu, and Paul Debevec. Magenta green screen: Spectrally multiplexed alpha matting with deep colorization. In The Digital Production Symposium (DigiPro), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.607, + 0.905, + 0.633 + ], + "angle": 0, + "content": "[29] Alvy Ray Smith and James F Blinn. Blue screen matting. In SIGGRAPH, 1996. 1, 2, 6, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.635, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[30] George Gabriel Stokes. On the composition and resolution of streams of polarized light from different sources. Transactions of the Cambridge Philosophical Society, 9:399, 1851. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.691, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[31] The Foundry Visionmongers Ltd. Keylight. https://learn.foundry.com/nuke/13.0/content/reference_guide/keyer_nodes/ keylight.html. Accessed: 2025-02-07. 1, 2, 7, 8, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[32] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.832 + ], + "angle": 0, + "content": "[33] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In ACM MM, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "399" + } + ] +] \ No newline at end of file diff --git a/2025/Polarized Color Screen Matting/4bc8f905-0a33-46ab-838e-6d058731820a_origin.pdf b/2025/Polarized Color Screen Matting/4bc8f905-0a33-46ab-838e-6d058731820a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a5badb7c6acce343f5c6d35774fe4a3d08e2a00a --- /dev/null +++ b/2025/Polarized Color Screen Matting/4bc8f905-0a33-46ab-838e-6d058731820a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddc00b7446c986ff3cf14bf47c9534a2cd42c348ea99c41ae73b12b62ca3532d +size 951196 diff --git a/2025/Polarized Color Screen Matting/full.md b/2025/Polarized Color Screen Matting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1dfdb3864d5b3f5eeb2417304f79ecbda92b3817 --- /dev/null +++ b/2025/Polarized Color Screen Matting/full.md @@ -0,0 +1,450 @@ +# Polarized Color Screen Matting + +Kenji Enomoto + +Scott Cohen + +Brian Price + +Adobe Research + +TJ Rhodes + +![](images/0676404ca766dac43b5b8a02829ede517d2550acb824f2073535671628529079.jpg) + +![](images/e95f43ac7e90b9ebac418a26311ad2a3d08a3b8dcba7b6d7937d071c82fdbcc0.jpg) +(a) Scene + +![](images/ce6f9c81fb53732e98eb389493d40b34a96ae4e2f62a6b5212b7c4e53a3759a0.jpg) + +![](images/c6438447227b25d72ca2a8e9bdb7ad59b4d825e488fd9b42f1cfcaa44919fa33.jpg) +Figure 1. (a) Scenes with fully polarized blue/green backgrounds under unpolarized known color illumination. (b) Commercial keying tools [31] use color information to extract an alpha matte, often struggling with hair details or motion blur due to the essential ill-posedness of the problem. (c) Polarimetric matting [22] uses polarization cues to extract an alpha matte without manual operations, often struggling with polarized foreground reflections as shown in the specular highlight. (d) Our Polarized Color Screen Matting leverages color and polarization cues, which enables more accurate alpha matte extraction for a wide range of real-world objects. + +![](images/69c96b4abb5b06b2071d3def511369b60a010a45cb65e915c17f764726b91d2d.jpg) + +![](images/5cad6778570a9389a0a3ac9d09ce863bb96634e94f38a8e7b9b6927a96ebd662.jpg) +(b) Chroma keying + +![](images/bed4f6d6d4e1a03af811c6f179af0054a70f98c21d97e9cf3b20b6141410b003.jpg) + +![](images/a5af54e8b8df1b94ca91f7ea6203238b69b6915162b021aa454df94f51b43b50.jpg) +(c) Polarimetric matting + +![](images/f880d8aef1391ebe6642591716bc5ed694a8cbaac7b0994b45b6dfdf7bdba2d0.jpg) + +![](images/50dd9ddca87d68425d7b2cc0222296af58854754c09d0bb75e17979a94c0db13.jpg) + +![](images/b527ec44273f4b9cdef9dcddba211b0c587f1db1d19a0b8f9cd2f5f5c62e7e0d.jpg) + +![](images/5cc1348bab708cab9fc053d68e931caa1241e4ee4ede2a96b16ab744692f5ec0.jpg) + +![](images/7673760209bd2dcc6c8eef3eb83ef57b2c52d9a53a3bf33d88034958b8ef9712.jpg) + +![](images/96572da303678e1e9df8f2f7b5de0272a3dd54b32048b9cd8a37ef830bd48393.jpg) +(d) Ours + +![](images/a57e20190af89c48e969dc6cfd5742f021175db4a29626edc9793cea21285089.jpg) + +![](images/42fd110c552e5a0f523e6182015f3774a4d6b3765acfa0c31b02da02d2368bad.jpg) + +# Abstract + +This paper considers the long-standing problem of extracting alpha mattes from video using a known background. While various color-based or polarization-based approaches have been studied in past decades, the problem remains ill-posed because the solutions solely rely on either color or polarization. We introduce Polarized Color Screen Matting, a single-shot, per-pixel matting theory for alpha matte and foreground color recovery using both color and polarization cues. Through a theoretical analysis of our diffuse-specular polarimetric compositing equation, we derive practical closed-form matting methods with their solvability conditions. Our theory concludes that an alpha matte can be extracted without manual corrections using off-the-shelf equipment such as an LCD monitor, polarization camera, and unpolarized lights with calibrated color. Experiments on synthetic and real-world datasets verify the validity of our theory and show the capability of our matting methods on real videos with quantitative and qualitative comparisons to color-based and polarization-based matting methods. + +# 1. Introduction + +Image and video editing rely on compositing [25] to blend images. The quality of the composite depends on how accurately the alpha matte depicts opacity. Modern matte extraction methods use inherent information such as color or polarization to extract the alpha matte. However, they are forced to compromise on quality of foreground color, alpha, or both. This paper combines color and polarimetric observations to constrain the problem. + +Color-based alpha matte extraction methods are most commonly used for compositing due to equipment availability. Chroma keying [2, 29] is the de facto standard in film making, but because it is underconstrained, it often requires manual corrections to extract high-quality alpha mattes. In addition to this, it often struggles with motion blur and transparent objects as depicted in Fig. 1. Recent learning-based matting methods [18, 19, 23, 32] can sometimes extract acceptable mattes but they still struggle with complex situations; potential improvements in modeling and dataset remain. Magenta green screen [28] can extract a more precise alpha matte by sacrificing the green color channel, but this later requires a restoration of the green color channel using machine-learning colorization. All these challenges are rooted in the fact that the use of + +color alone to recover an alpha matte is underconstrained. + +Alternatively, polarization-based matting methods [6, 11, 22] can extract the alpha matte using the polarization of light. For a scene backlit with polarized light, there is less ambiguity between phase angle than color, thus physically accurate alpha mattes can be obtained for the majority of pixels. However, in cases where the foreground subject reflects polarized light whose phase angle is comparable to the background, the alpha matte quality will be degraded, as shown in the headphone and glass scenes in Fig. 1. + +We introduce Polarized Color Screen Matting (pCSM), a single-shot, per-pixel matting theory that uses both color and polarization cues. Our theory is built upon our proposed diffuse-specular polarimetric compositing equation, where foreground appearance is explicitly modeled by a summation of diffuse and specular reflections. This new compositing equation steps forward a theoretical analysis of the matting problem, resulting in two matting methods: (1) $p\text{CSM-S}$ , a fast and robust matting method with an approximation of unpolarized diffuse reflections that is applicable to any scene, and (2) $p\text{CSM-DS}$ , an accurate and approximation-free matting method. Through our theoretical analysis, we found solvability conditions of pCSM-DS in terms of color and polarized reflection from the foreground subject. For the practical use of our theoretical matting method, we introduce a physics-inspired regularization, which grants our methods significant robustness to noisy real-world observations. + +The main contributions of this paper are threefold: + +- A diffuse-specular polarimetric compositing equation that unlocks further theoretical analysis of the polarized screen matting problem. +- A fast and simple matting method, pCSM-S, that is applicable to any scene. +- An accurate and approximation-free matting method, pCSM-DS, with solvability conditions. + +We show quantitatively and qualitatively that our methods produce more accurate results than state-of-the-art solely color-based or polarization-based matting methods. + +# 2. Related Work + +Early methods for computing alpha mattes from video have focused on capturing footage against a solid-colored background [7], often referred to as chroma keying. Indeed, chroma keying is an industry standard, implemented in commercial products [31] and used for generating video matting datasets [18, 33]. However, despite its popularity, chroma keying is underconstrained and often performs poorly unless an expert carefully sets up the capture environment and manually post-processes the matte. + +Due to the shortcomings of standard chroma keying, further work has focused on using color to facilitate matte extraction. The sodium vapor matting process [24] constrains + +the problem by narrowing the spectrum of backlight emitted from a low-pressure sodium vapor lamp. A beamsplitter and bandpass filter allow the background sodium vapor light to be captured on one strip of film, forming an alpha matte, and the foreground subject's color to be captured through a notch reject filter on a second strip of film, blocking the sodium vapor light. More recently, magenta green screen [28] looked to eliminate the need for a specialized camera system by placing a foreground subject strongly lit by magenta light in front of a green background to observe the green spectrum solely at the background. However, the foreground magenta light loses the true foreground color, requiring machine-learning colorization to recover the lost green color of the foreground. In theory, triangulation matting [29] can compute the true alpha matte by taking two (or more) images of a static subject in front of different colored backgrounds to triangulate the alpha matte. However, since the images must be taken at different times to allow the background color to change, triangulation matting is not suitable for dynamic scenes without sacrificing motion blur, such as with stop-motion photography [12]. + +Instead of using color to constrain the capture system, polarimetric triangulation matting (PTM) [22] employs polarization. PTM uses a beamsplitter with two polarization filters to simultaneously capture two polarization images of a target scene against a polarized background screen. These images can serve as input for triangulation matting if the foreground reflects only unpolarized light; however, in practice, polarized foreground reflections are observed, which cause perceivable errors in foreground subjects. + +Addressing the issue of polarized foreground reflections, PolarMatte [11] reformulates the polarized screen matting problem by leveraging four polarization images captured with a modern polarization camera (see Fig. 2). In this framework, an extended robust PTM is first applied to pixels with negligible polarized reflections, and the resulting high-confidence alpha mattes are subsequently interpolated to pixels with polarized foreground reflections using spatial context through a polarimetric matting Laplacian. However, two critical limitations persist. First, because PolarMatte is not formulated as a per-pixel method, establishing solvability conditions remains challenging, thereby complicating both its theoretical analysis and practical reliability. Second, its exclusive reliance on polarization cues (i.e., the use of white polarized background) renders it vulnerable when the polarized reflection aligns with the background polarity—a situation often encountered in the presence of specular highlights (e.g., from jewelry or glossy skin). In this paper, we aim to overcome these limitations by jointly integrating color and polarization cues, which more robustly constrains the matting problem without adding significant complexity to the capture process. + +![](images/371d023ffd82abc666dc055f83eaafda5f5050a9af943cffc5707a6aef7ee313.jpg) +Figure 2. Illustration of our capture setup. A recent polarization camera with a color polarization filter array sensor captures four polarization images in a single-shot. Our method is tailored for the polarization images of a foreground subject with a known polarized background under known color illuminations. + +# 3. Polarized Color Screen Matting + +This section delivers a proof that a single-shot and per-pixel matting problem has a closed-form solution given: (1) the background is known and polarized, (2) the foreground illumination color is known, and (3) a diffuse-specular reflectance model for the foreground subject. Our pCSM theory is tailored for polarization images at four angles $\left\{0, \frac{\pi}{4}, \frac{\pi}{2}, \frac{3\pi}{4}\right\}$ captured by a recent polarization camera shown in Fig. 2, similarly to PolarMatte [11]. The reader is encouraged to reference Table 5 for the notations used in the following sections. + +# 3.1. Diffuse-Specular Polarimetric Compositing Equation + +The matting problem is formulated as a decomposition of the color vector of a given pixel into foreground and background color vectors and the opacity of the foreground, commonly called an alpha matte. In a linear color space, the compositing process of a polarization image is described in the polarimetric compositing equation [11]: + +$$ +\mathbf {c} _ {\phi} = \alpha \mathbf {f} _ {\phi} + (1 - \alpha) \mathbf {b} _ {\phi}, \tag {1} +$$ + +where $\mathbf{c}_{\phi}, \mathbf{f}_{\phi}, \mathbf{b}_{\phi} \in \mathbb{R}_+^3$ are composite, foreground, and background color polarization vectors at an angle $\phi \in \{0, \frac{\pi}{4}, \frac{\pi}{2}, \frac{3\pi}{4}\}$ , and $\alpha \in [0,1]$ is an alpha matte. We employ a diffuse-specular reflectance model [16] on the foreground color, which is the sum of diffuse and specular terms: + +$$ +\mathbf {f} _ {\phi} = \underbrace {\left(u _ {d} + p _ {d} \cos^ {2} \left(\theta_ {d} - \phi\right)\right) \mathbf {f} _ {d}} _ {\text {d i f f u s e t e r m}} + \underbrace {\left(u _ {s} + p _ {s} \cos^ {2} \left(\theta_ {s} - \phi\right)\right) \mathbf {f} _ {s}} _ {\text {s p e c u l a r t e r m}}. \tag {2} +$$ + +Here, the diffuse and specular terms are decomposed into unpolarized and polarized terms, where $u_{d}, p_{d} \in \mathbb{R}_{+}$ + +are the intensities of the unpolarized and polarized diffuse reflections, respectively, and $u_{s}, p_{s} \in \mathbb{R}_{+}$ are the same for the specular reflection. $\mathbf{f}_d, \mathbf{f}_s \in S_+^2 \subset \mathbb{R}_+^3$ denote unit color vectors of diffuse and specular reflections, respectively. $\theta_d, \theta_s \in [0, \pi)$ are phase angles of the diffuse and specular reflections, and it is known that $\theta_d = \theta_s \pm \frac{\pi}{2}$ under unpolarized illumination [3, 9, 16]. Therefore, the diffuse term in Eq. (2) can be rewritten using $\theta_{s}$ as + +$$ +\left(u _ {d} + p _ {d} \sin^ {2} \left(\theta_ {s} - \phi\right)\right) \mathbf {f} _ {d}. \tag {3} +$$ + +In our problem setting, the background is fully polarized, which can be described in + +$$ +\mathbf {b} _ {\phi} = p _ {b} \cos^ {2} (\theta_ {b} - \phi) \mathbf {q}, \tag {4} +$$ + +where $p_b \in \mathbb{R}_+$ , $\theta_b \in [0, \pi)$ , and $\mathbf{q} \in S_+^2 \subset \mathbb{R}_+^3$ are the intensity, phase angle, and unit color vector of the backlight, respectively. Finally, the polarimetric compositing equation (1) can be written as + +$$ +\begin{array}{l} \mathbf {c} _ {\phi} = \alpha \left(\left(u _ {d} + p _ {d} \sin^ {2} \left(\theta_ {s} - \phi\right)\right) \mathbf {f} _ {d} + \left(u _ {s} + p _ {s} \cos^ {2} \left(\theta_ {s} - \phi\right)\right) \mathbf {f} _ {s}\right) \\ + (1 - \alpha) p _ {b} \cos^ {2} (\theta_ {b} - \phi) \mathbf {q}. \tag {5} \\ \end{array} +$$ + +We call this compositing process the diffuse-specular polarimetric compositing equation. The diffuse-specular reflectance model is widely used in polarimetric vision problems, such as diffuse-specular separation [10, 16], shape-from-polarization [17, 20], and polarimetric BRDF acquisition [4]. We extend it to the compositing equation to represent opacity. + +For ease of theoretical analysis, we transform observed polarization images into Stokes parameters [30]. This paper only considers linear polarization, and the Stokes parameters can also be described in the form of compositing equations: + +$$ +\mathbf {s} _ {0} = \frac {1}{2} \sum_ {\phi} \mathbf {c} _ {\phi} \stackrel {\mathrm {d e f}} {=} \mathbf {d} _ {0} + (1 - \alpha) \mathbf {q} _ {0}, \tag {6} +$$ + +$$ +\mathbf {s} _ {1} = \mathbf {c} _ {0} - \mathbf {c} _ {\frac {\pi}{2}} \stackrel {\mathrm {d e f}} {=} \mathbf {d} _ {1} + (1 - \alpha) \mathbf {q} _ {1}, \tag {7} +$$ + +$$ +\mathbf {s} _ {2} = \mathbf {c} _ {\frac {\pi}{4}} - \mathbf {c} _ {\frac {3 \pi}{4}} \stackrel {\text {d e f}} {=} \mathbf {d} _ {2} + (1 - \alpha) \mathbf {q} _ {2}, \tag {8} +$$ + +where + +$$ +\mathbf {d} _ {0} = \alpha \left(\left(2 u _ {d} + p _ {d}\right) \mathbf {f} _ {d} + \left(2 u _ {s} + p _ {s}\right) \mathbf {f} _ {s}\right), \quad \mathbf {q} _ {0} = p _ {b} \mathbf {q}, \tag {9} +$$ + +$$ +\mathbf {d} _ {1} = \alpha \cos 2 \theta_ {s} \left(p _ {s} \mathbf {f} _ {s} - p _ {d} \mathbf {f} _ {d}\right), \mathbf {q} _ {1} = p _ {b} \cos 2 \theta_ {b} \mathbf {q}, \tag {10} +$$ + +$$ +\mathbf {d} _ {2} = \alpha \sin 2 \theta_ {s} \left(p _ {s} \mathbf {f} _ {s} - p _ {d} \mathbf {f} _ {d}\right), \mathbf {q} _ {2} = p _ {b} \sin 2 \theta_ {b} \mathbf {q}. \tag {11} +$$ + +Here, $\mathbf{s}_0$ is equivalent to an ordinary color image, and $\mathbf{s}_1$ and $\mathbf{s}_2$ represent the polarization properties of an observation. + +This paper aims to recover the alpha matte $\alpha$ given the polarization images $\mathbf{c}_{\phi}$ and the background images $\mathbf{b}_{\phi}$ for all the angles $\phi \in \{0, \frac{\pi}{4}, \frac{\pi}{2}, \frac{3\pi}{4}\}$ . + +![](images/f44b2a1cc6a24d367f7a5c879f93c4118db12b0edb918cb8db7717f51ef82718.jpg) +Figure 3. (Left) A scene of a headphone in front of a blue polarized background and its polarized intensity computed as $\sqrt{\mathbf{s}_1^2 + \mathbf{s}_2^2}$ . (Right) The color distribution of the polarized intensities, showing that the majority of the polarized intensities derive from background and specular reflections. + +# 3.2. pCSM with Polarized Specular Reflection + +This section introduces a pCSM method, which we call $p\mathrm{CSM - S}$ that highlights the polarized specular reflection. Prior polarimetric matting methods [11, 22] show that polarized reflections from a foreground subject are a major source of error in the alpha matte estimation. In particular, polarized specular reflections have larger intensities than polarized diffuse reflections and a dominant effect on the estimated alpha matte (see Fig. 3). This observation motivates us to simplify the polarimetric compositing equation and Stokes parameters without the diffuse term (i.e., $p_d = 0$ ). With this simplification, Eqs. (7) and (8) can be rewritten as + +$$ +\mathbf {s} _ {1} = \alpha p _ {s} \cos 2 \theta_ {s} \mathbf {f} _ {s} + (1 - \alpha) \mathbf {q} _ {1}, \tag {12} +$$ + +$$ +\mathbf {s} _ {2} = \alpha p _ {s} \sin 2 \theta_ {s} \mathbf {f} _ {s} + (1 - \alpha) \mathbf {q} _ {2}. \tag {13} +$$ + +Assuming the dichromatic reflectance model [26], the specular reflection color $\mathbf{f}_s$ is independent of the object's albedo and is equal to the ambient light color. Therefore, we can assume the specular reflection color $\mathbf{f}_s$ to be known if the ambient light color is calibrated or ambient light is natural and white. In the following, we assume that the specular reflection color $\mathbf{f}_s$ is known. + +With Eqs. (12) and (13), the alpha matte can be estimated by solving the following linear system: + +$$ +\underbrace {\left[ \begin{array}{l} \mathbf {s} _ {1} - \mathbf {q} _ {1} \\ \mathbf {s} _ {2} - \mathbf {q} _ {2} \end{array} \right]} _ {\mathbf {y}} = \underbrace {\left[ \begin{array}{c c c} - \mathbf {q} _ {1} & \mathbf {f} _ {s} & \mathbf {0} \\ - \mathbf {q} _ {2} & \mathbf {0} & \mathbf {f} _ {s} \end{array} \right]} _ {\mathbf {A}} \underbrace {\left[ \begin{array}{c} \alpha \\ \alpha p _ {s} \cos 2 \theta_ {s} \\ \alpha p _ {s} \sin 2 \theta_ {s} \end{array} \right]} _ {\mathbf {x}}. \tag {14} +$$ + +Thus, this linear system has a closed-form solution, and in + +practice, the alpha matte can be computed as + +$$ +\alpha = 1 - \frac {\mathbf {q} _ {1} ^ {\top} \mathbf {s} _ {1} + \mathbf {q} _ {2} ^ {\top} \mathbf {s} _ {2} - \left(\mathbf {q} _ {1} ^ {\top} \mathbf {f} _ {s}\right) \left(\mathbf {s} _ {1} ^ {\top} \mathbf {f} _ {s}\right) - \left(\mathbf {q} _ {2} ^ {\top} \mathbf {f} _ {s}\right) \left(\mathbf {s} _ {2} ^ {\top} \mathbf {f} _ {s}\right)}{p _ {b} ^ {2} - \left(\mathbf {q} _ {1} ^ {\top} \mathbf {f} _ {s}\right) ^ {2} - \left(\mathbf {q} _ {2} ^ {\top} \mathbf {f} _ {s}\right) ^ {2}} \tag {15} +$$ + +without computing the pseudo inverse matrix of $\mathbf{A}$ for each pixel of a given image. + +Proposition 1. If $\mathbf{q} \neq \mathbf{f}_s$ , the $p$ CSM-S problem has a closed-form solution. + +Proof. The determinant of $\mathbf{A}^{\top}\mathbf{A}$ can be calculated as + +$$ +\det \left(\mathbf {A} ^ {\top} \mathbf {A}\right) = p _ {b} ^ {2} \left(1 - \left(\mathbf {q} ^ {\top} \mathbf {f} _ {s}\right) ^ {2}\right). \tag {16} +$$ + +Since $\mathbf{q},\mathbf{f}_s\geq \mathbf{0}$ and $\| \mathbf{q}\| _2 = \| \mathbf{f}_s\| _2 = 1$ ,det $(\mathbf{A}^{\top}\mathbf{A}) = 0$ if and only if $\mathbf{q} = \mathbf{f}_s$ + +Therefore, our pCSM-S can be always applied by using a backlight color sufficiently different from the ambient light color; for instance, a green background with natural white ambient light. + +# 3.3. pCSM with Polarized Diffuse-Specular Reflection + +This section introduces another pCSM method, which we call $p\mathrm{CSM - DS}$ , using the full expression of the diffuse-specular polarimetric compositing equation without any simplification. Research shows that more general modeling leads to more ambiguity [27], and our pCSM-DS problem is no exception. Here, we present the solvability conditions of our pCSM-DS and the solutions when each condition is satisfied. Since the pCSM-DS problem is equivalent to pCSM-S if $\mathbf{f}_d = \mathbf{f}_s$ or $p_d = 0$ , we suppose $\mathbf{f}_d \neq \mathbf{f}_s$ and $p_d > 0$ in the following. + +Proposition 2. If a strong color condition, span $\{\mathbf{q},\mathbf{f}_d,\mathbf{f}_s\} = \mathbb{R}^3$ is satisfied, the pCSM-DS problem has a closed-form solution. + +Proof. Since Eq. (6) can be written as + +$$ +\mathbf {s} _ {0} - p _ {b} \mathbf {q} = \left[ \begin{array}{l l l} - p _ {b} \mathbf {q} & \mathbf {f} _ {d} & \mathbf {f} _ {s} \end{array} \right] \left[ \begin{array}{c} \alpha \\ \alpha \left(2 u _ {d} + p _ {d}\right) \\ \alpha \left(2 u _ {s} + p _ {s}\right) \end{array} \right], \tag {17} +$$ + +if $\mathbf{q}$ , $\mathbf{f}_d$ , $\mathbf{f}_s$ are linearly independent, the matrix in Eq. (17) is invertible, and it has a closed-form solution. + +The missing variable to solve Eq. (17) is the diffuse color vector $\mathbf{f}_d$ , which is inherently difficult to be known. However, the use of polarization images with a fully polarized background allows us to compute a linear combination of $\mathbf{f}_d$ and $\mathbf{f}_s$ as + +$$ +\begin{array}{l} \mathbf {f} _ {d s} = \mathbf {c} _ {\theta_ {b} - \frac {\pi}{2}} \\ = \alpha \left(\left(u _ {d} + p _ {d} \cos^ {2} \left(\theta_ {s} - \theta_ {b}\right)\right) \mathbf {f} _ {d} + \left(u _ {s} + p _ {s} \sin^ {2} \left(\theta_ {s} - \theta_ {b}\right)\right) \mathbf {f} _ {s}\right). \tag {18} \\ \end{array} +$$ + +We call this vector $\mathbf{f}_{ds}$ the foreground basis. Since the unpolarized reflection is dominant in the diffuse reflection, the unpolarized diffuse intensity $u_{d}$ is almost always nonzero when the polarized diffuse intensity is positive $p_d > 0$ . Hence, the foreground basis and $\mathbf{f}_s$ are linearly independent unless $\mathbf{f}_d = \mathbf{f}_s$ . With this foreground basis, Eq. (17) can be rewritten as + +$$ +\mathbf {s} _ {0} - p _ {b} \mathbf {q} = \underbrace {\left[ - p _ {b} \mathbf {q} \quad \mathbf {f} _ {d s} \quad \mathbf {f} _ {s} \right]} _ {\mathbf {B}} \left[ \begin{array}{l} \alpha \\ v _ {d s} \\ v _ {s} \end{array} \right], \tag {19} +$$ + +where $v_{ds}, v_s \in \mathbb{R}$ . Therefore, if the matrix $\mathbf{B}$ is full-rank, i.e., the strong color condition is satisfied, this linear system has a closed-form solution, and the alpha matte is uniquely determined. We call this method linear $pCSM-DS$ . + +We then consider the case of $\mathbf{q} \in \operatorname{span}\{\mathbf{f}_d, \mathbf{f}_s\}$ . In this case, Eq. (6) can be satisfied with multiple alpha matte values, hence, we leverage Eqs. (7) and (8) to derive a unique alpha matte. + +Proposition 3. If a weak color condition, $\mathbf{q} \not\propto p_s\mathbf{f}_s - p_d\mathbf{f}_d$ and a polarization condition, $\theta_b \neq \theta_s$ nor $\theta_b \neq \theta_d$ , are satisfied, the $p$ CSM-DS problem has a unique solution, even if the strong color condition is not satisfied. + +Proof. See the supplementary material. + +Proposition 3 shows the well-posedness of the pCSM-DS problem when the weak color condition and polarization condition are satisfied; therefore, here we derive a solution to the pCSM-DS problem. The alpha matte can be found as a minimizer of the following objective function: + +$$ +\begin{array}{l} f (\alpha , \mathbf {w}, t) = \| \mathbf {s} _ {1} - \mathbf {d} _ {1} - (1 - \alpha) \mathbf {q} _ {1} \| _ {2} ^ {2} + \| \mathbf {s} _ {2} - \mathbf {d} _ {2} - (1 - \alpha) \mathbf {q} _ {2} \| _ {2} ^ {2} \\ = \| \mathbf {s} _ {1} - \mathbf {F} \mathbf {w} - (1 - \alpha) \mathbf {q} _ {1} \| _ {2} ^ {2} + \| \mathbf {s} _ {2} - t \mathbf {F} \mathbf {w} - (1 - \alpha) \mathbf {q} _ {2} \| _ {2} ^ {2} \tag {20} \\ \end{array} +$$ + +where $\mathbf{d}_1$ and $\mathbf{d}_2$ , defined in Eqs. (10) and (11), are represented by a basis matrix $\mathbf{F} = [\mathbf{f}_{ds}\mathbf{f}_s]\in \mathbb{R}^{3\times 2}$ , coefficient vector $\mathbf{w} = [w_{ds}w_s]^\top \in \mathbb{R}^2$ , and a scalar $t\in \mathbb{R}$ . This objective function is written in matrix form as + +$$ +\begin{array}{l} f (\alpha , \mathbf {w}, t) = \left\| \left[ \begin{array}{l} \mathbf {s} _ {1} - \mathbf {q} _ {1} \\ \mathbf {s} _ {2} - \mathbf {q} _ {2} \end{array} \right] - \left[ \begin{array}{l l} - \mathbf {q} _ {1} & \mathbf {F} \\ - \mathbf {q} _ {2} & t \mathbf {F} \end{array} \right] \left[ \begin{array}{l} \alpha \\ \mathbf {w} \end{array} \right] \right\| _ {2} ^ {2} \tag {21} \\ \stackrel {{\mathrm {d e f}}} {{=}} \left\| \mathbf {y} - \mathbf {D} (t) \left[ \begin{array}{c} \alpha \\ \mathbf {w} \end{array} \right] \right\| _ {2} ^ {2}. \\ \end{array} +$$ + +Since $t$ is unknown, the alpha matte cannot be derived in a closed-form manner. However, as shown in Eq. (22), $\alpha$ and $\mathbf{w}$ may be eliminated from Eq. (21), leaving a polynomial function only in $t$ . + +![](images/788694c48578e692902a03bb07909d3b4859fdd297e3a37d7c3bff0212762cd7.jpg) +Figure 4. An illustrative interpretation of pCSM-DS optimization. For a fixed $t$ , the objective function has a least-squares solution (blue points). Projecting the least-squares solutions to the $t - f(\alpha, \mathbf{w}, t)$ plane forms $g(t)$ , where the critical points (red and yellow points) can be derived as roots of the derivative $g'(t)$ . The final solution (red point) can be found by assessing the values of $f(\alpha, \mathbf{w}, t)$ at all the critical points. + +Proposition 4. Define $g(t)$ as + +$$ +g (t) = \min _ {\alpha , \mathbf {w}} f (\alpha , \mathbf {w}, t). \tag {22} +$$ + +Then, $g(t)$ is a ratio of polynomials in $t$ as + +$$ +g (t) = \frac {n (t)}{d (t)}, \tag {23} +$$ + +where $n(t)$ and $d(t)$ are fourth-degree polynomials in $t$ . + +Proof. See the supplementary material. + +![](images/7ed151b7b90a84b84bfc12aec7e1b2adc85f651d193cad847b5556aa60bc5186.jpg) + +With Proposition 4, we can find critical points of $g(t)$ as roots of the derivative $g'(t)$ and corresponding alpha matte values and coefficient vectors. Noting the range of the alpha matte value, we discard critical points if the corresponding alpha matte values are out of the range [0, 1], forming a set of valid critical points, say $\mathcal{T}$ . Therefore, the optimal variables can be found as + +$$ +t _ {m} = \underset {t \in \mathcal {T}} {\operatorname {a r g m i n}} g (t), \tag {24} +$$ + +$$ +\alpha_ {m}, \mathbf {w} _ {m} = \underset {\alpha , \mathbf {w}} {\operatorname {a r g m i n}} f (\alpha , \mathbf {w}, t = t _ {m}). \tag {25} +$$ + +An illustrative interpretation of this optimization are shown in Fig. 4. + +Note that we need to assess the boundaries of the alpha matte value separately. Given $\alpha = \alpha'$ , the objective function becomes + +$$ +g _ {\alpha^ {\prime}} (t) = \min _ {\mathbf {w}} f (\alpha = \alpha^ {\prime}, \mathbf {w}, t). \tag {26} +$$ + +Proposition 5. $g_{\alpha^{\prime}}(t)$ is a ratio of quadratic functions in $t$ + +Proof. See the supplementary material. + +![](images/91d07de54d10a712c2d7b3a586190e064552ba763ae80a1ea8fc8156592c22b6.jpg) + +Similarly with Eqs. (24) and (25), the optimal $t$ and coefficient vector for $g_0(t)$ and $g_{1}(t)$ can be found as $(t_0,\mathbf{w}_0)$ and $(t_{1},\mathbf{w}_{1})$ , respectively. Gathering all candidates of the solution in $\mathcal{V} = \left\{\left(\alpha_{m},\mathbf{w}_{m},t_{m}\right),(0,\mathbf{w}_{0},t_{0}),(1,\mathbf{w}_{1},t_{1})\right\}$ , the final solution can be written as + +$$ +\alpha^ {*}, \mathbf {w} ^ {*}, t ^ {*} = \underset {\alpha , \mathbf {w}, t \in \mathcal {V}} {\operatorname {a r g m i n}} f (\alpha , \mathbf {w}, t). \tag {27} +$$ + +This pCSM-DS optimizes a non-linear objective function; therefore, we call this method non-linear $p\mathrm{CSM - DS}$ + +# 3.4. Polarized Diffuse Regularization + +Our pCSM-DS is built upon a solid theory; however, we observed cases where the linear and non-linear pCSM-DS yield undesirable noisy alpha mattes. This behavior is due to the overfitting of the diffuse-specular polarimetric compositing equation. For instance, in the linear pCSM-DS, the matrix $\mathbf{B}$ in Eq. (19) spans the entire RGB space, and $\mathbf{s}_0$ is perfectly approximated even if $\mathbf{s}_0$ contains deviations from the model, e.g., noise. To avoid overfitting, we introduce a polarized diffuse regularization term, motivated by a physical observation that the polarized diffuse intensity $p_d$ tends to be small in most scenes. To introduce this regularization into the linear pCSM-DS, we reformulate the linear pCSM-DS with $\mathbf{s}_1$ and $\mathbf{s}_2$ as follows: + +$$ +\underbrace {\left[ \begin{array}{l} \mathbf {s} _ {1} - \mathbf {q} _ {1} \\ \mathbf {s} _ {2} - \mathbf {q} _ {2} \end{array} \right]} _ {\mathbf {y}} = \underbrace {\left[ \begin{array}{c c c c c} - \mathbf {q} _ {1} & \mathbf {f} _ {d s} & \mathbf {f} _ {s} & \mathbf {0} & \mathbf {0} \\ - \mathbf {q} _ {2} & \mathbf {0} & \mathbf {0} & \mathbf {f} _ {d s} & \mathbf {f} _ {s} \end{array} \right]} _ {\overline {{\mathbf {B}}}} \underbrace {\left[ \begin{array}{l} \alpha \\ m _ {d s} \\ m _ {s} \\ n _ {d s} \\ n _ {s} \end{array} \right]} _ {\overline {{\mathbf {x}}}}. \tag {28} +$$ + +Here, the polarized diffuse intensity $p_d$ is only included in $m_{ds}$ and $n_{ds}$ , hence, we regularize $m_{ds}$ and $n_{ds}$ to indirectly constrain the polarized diffuse intensity. We use the ridge regression to form the objective function as + +$$ +\bar {\mathbf {x}} ^ {*} = \underset {\bar {\mathbf {x}}} {\operatorname {a r g m i n}} \| \mathbf {y} - \bar {\mathbf {B}} \bar {\mathbf {x}} \| _ {2} ^ {2} + \lambda \| \mathbf {E} \bar {\mathbf {x}} \| _ {2} ^ {2}, \tag {29} +$$ + +$$ +\mathbf {E} = \left[ \begin{array}{l l l l l} 0 & 1 & 0 & 0 & 0 \\ 0 & 0 & 0 & 1 & 0 \end{array} \right], \tag {30} +$$ + +where $\lambda \in \mathbb{R}_+$ is a regularization parameter. This ridge regression problem can be solved in a closed-form manner. + +Next, we introduce a regularization in the non-linear pCSM-DS. In Eq. (21), the polarized diffuse intensity $p_d$ is only included in $w_{ds}$ , hence, we simply regularize $w_{ds}$ and obtain the objective function as + +$$ +\bar {f} (\alpha , \mathbf {w}, t) = \left\| \mathbf {y} - \mathbf {D} (t) \left[ \begin{array}{l} \alpha \\ \mathbf {w} \end{array} \right] \right\| _ {2} ^ {2} + \mu \| \mathbf {e} ^ {\top} \mathbf {w} \| _ {2} ^ {2}, \tag {31} +$$ + +where $\mathbf{e} = [10]^\top$ , $\mu \in \mathbb{R}_+$ is a regularization parameter. Since this objective function is a quadratic function with respect to $\alpha$ and $\mathbf{w}$ , we can still compute a closed-form expression for + +$$ +\bar {g} (t) = \min _ {\alpha , \mathbf {w}} \bar {f} (\alpha , \mathbf {w}, t), \tag {32} +$$ + +$$ +\bar {g} _ {\alpha^ {\prime}} (t) = \min _ {\mathbf {w}} \bar {f} (\alpha = \alpha^ {\prime}, \mathbf {w}, t). \tag {33} +$$ + +Therefore, the optimal alpha matte can be estimated in the same manner with the unregularized non-linear pCSM-DS. + +Proposition 6. $\overline{g}(t)$ is a ratio of fourth-degree polynomials in $t$ . + +Proof. See the supplementary material. + +![](images/0e396fb0ad517b49e7488443af63e9596d35b66d33d547fbbe03d5879cb57dda.jpg) + +Proposition 7. $\overline{g}_{\alpha'}(t)$ is a ratio of fourth-degree polynomials in $t$ . + +Proof. See the supplementary material. + +![](images/7c7ff336401856daeac0694080ab423bf7220aa574aa793c0ad19b2af236b303.jpg) + +Note that, similar to pCSM-S, the polarized diffuse regularization makes pCSM-DS applicable to any scene regardless of the color and polarization conditions. We assess the capability of each regularized pCSM-DS using synthetic datasets in Sec. 4.1 to identify the superior method. + +# 4. Experimental Evaluation + +Synthetic Dataset: We evaluated the capability of each matting method in isolation with four types of synthetic datasets, each of which satisfies different color or polarization conditions as shown in Tab. 1. We generated all the parameters and synthesized $\mathbf{c}_{\phi}$ and $\mathbf{b}_{\phi}$ following Eqs. (4) and (5). We generated 1000 samples (pixels) for each type of data. We also synthesized noisy datasets, to evaluate the robustness to noise, by adding the Gaussian noise $\mathbf{n}_{\phi} \sim \mathcal{N}(0, 0.01)$ as $\mathbf{c}_{\phi} + \mathbf{n}_{\phi}, \forall \phi$ . In total, we synthesized eight types of datasets: A, B, C, and D in Tab. 1 with and without noise. The details of how the parameters were generated can be found in the supplementary material. + +Real-world Dataset: For quantitative evaluations, we used the evaluation dataset used in PolarMatte [11], which contains 12 static scenes with a green polarized background. This dataset provides reference alpha mattes computed by triangulation matting [29] for quantitative evaluation. For further visual evaluations, we captured several videos containing various objects such as furry or transparent objects. + +Hardware Setup: All videos were captured with the Blackfly® S USB3 color polarization camera [13] and an LCD monitor [14] served as the polarized background. The camera's configuration was set to 30 frames per second, 1/30 second exposure time, and f/1.4. We used a color checker board [8] to linearize the camera response function and calibrate the color of illumination. + +Table 1. Color and polarization conditions for four types of synthetic datasets: A, B, C, and D. + +
ABCD
Strong color condition✓
Weak color condition✓✓✓
Polarization condition✓✓✓
+ +Hyperparameters: Throughout the experiments, we used $\lambda = 0.01$ and $\mu = 0.01$ for the polarized diffuse regularization. Since the PolarMatte static scene test data does not provide the color of illumination, we assume the ambient illumination is natural and white, and the specular reflection color is assumed to be $\mathbf{f}_s = (1 / \sqrt{3})\mathbf{1}$ . + +Comparative Methods: We compare our method with polarimetric triangulation matting (PTM) [21] using four polarization images, PolarMatte [11], background matting v2 (BGMv2) [18], and the Keylight plug-in [31] in Adobe After Effects [1]. The Keylight results shown in this paper were done by a Keylight expert in 5 minutes for each image to compare it with fully computational matting methods. + +Evaluation Metrics: We employ three commonly used metrics for evaluating the estimated alpha matte: mean absolute error (MAE), mean squared error (MSE), and gradient error (GRAD). We use MAE unless otherwise specified. + +# 4.1. Results on Synthetic Dataset + +Table 2 shows the quantitative comparisons on the noise-free and noisy synthetic datasets. The linear pCSM-DS on the noise-free dataset A and the non-linear pCSM-DS on the noise-free datasets A and B correctly estimate the true alpha matte with zero errors—verifying the validity of our theory. Although unregularized pCSM-DS is sensitive to noise, the polarized diffuse regularization improves robustness to noise. Consequently, the regularized non-linear pCSM-DS shows the best performance on noisy datasets A and B, in which the majority of real-world data is classified. On datasets C and D, since even the non-linear pCSM-DS is not solvable, all the methods produce a similar margin of error. PTM critically fails on dataset D because PTM only relies on polarization cues and the polarization condition is not satisfied. + +Table 3 shows the processing time of BGMv2 and our prototype implementations of matting methods written in Python, where PTM and pCSM-S for GPU processing are implemented using PyTorch. Since our pCSM-S and PTM are only required to perform a simple calculation, they are more than 9 times faster than BGMv2. Since the (regularized) non-linear pCSM-DS needs to find the zeros of high + +Table 2. Results on the noise-free synthetic data (top) and noisy synthetic data (bottom). Numbers in this table are MAE. See Tab. 1 for the conditions that each data type satisfies. + +
Noise-free dataABCD
PTM83.6478.4616.07117.60
pCSM-S10.9210.7416.0717.87
Linear pCSM-DS0.00NANANA
Non-linear pCSM-DS0.000.00NANA
Regularized linear pCSM-DS6.0110.7416.0717.87
Regularized non-linear pCSM-DS3.945.7421.4817.87
Noisy dataABCD
PTM85.2579.8421.30118.60
pCSM-S20.3821.4323.6425.39
Linear pCSM-DS42.45NANANA
Non-linear pCSM-DS23.2837.48NANA
Regularized linear pCSM-DS19.1721.7324.1125.60
Regularized non-linear pCSM-DS17.7819.6524.1925.82
+ +Table 3. Processing time on $2\mathrm{\;K}$ and $4\mathrm{\;K}$ images. + +
Resolution2K4K
BGMv2 (GPU)20.55 ms39.83 ms
PTM (GPU)1.80 ms6.61 ms
pCSM-S (GPU)2.28 ms8.39 ms
PTM (CPU)0.22 s1.15 s
pCSM-S (CPU)0.36 s1.83 s
Linear pCSM-DS (CPU)70 s259 s
Non-linear pCSM-DS (CPU)71 s276 s
Regularized linear pCSM-DS (CPU)63 s267 s
Regularized non-linear pCSM-DS (CPU)72 s277 s
+ +Table 4. Quantitative evaluations of Keylight, PTM [21], BGMv2 [18], PolarMatte [11], and our pCSM-S and pCSM-DS on the static scene test data with a green background [11]. + +
MAE ↓MSE ↓GRAD ↓
Keylight27.134.7412.18
BGMv285.2056.6212.51
PTM11.261.223.31
PolarMatte10.350.841.97
pCSM-S7.570.291.92
Regularized non-linear pCSM-DS7.480.261.68
+ +order polynomials and solve an ordinary least squares problem for every critical point, it shows a less efficient result. We note that while BGMv2 or per-image methods often struggle with dealing with very high-resolution images, all our methods are operated in a per-pixel manner and applicable to any resolution. + +The experimental results on the synthetic datasets show that the regularized non-linear pCSM-DS is the best choice for offline applications such as matting dataset labeling, and pCSM-S is the best choice for real-time applications such as video conferencing. + +![](images/ff1e952abb082207ac677aad896a560dfb76b97bded4dbc3198e47e43165c511.jpg) +Figure 5. Visual results of the reference alpha matte computed by triangulation matting (TM) [29], Keylight [31], BGMv2 [18], polarimetric triangulation matting (PTM) [21], PolarMatte [11], and our pCSM-S and Regularized non-linear pCSM-DS on the static scene test data. See more visual results in the supplementary materials. + +# 4.2. Results on Real-world Dataset + +Given the results on the synthetic datasets, we only applied the regularized non-linear pCSM-DS to the real-world datasets as a representative of pCSM-DS. Table 4 shows quantitative comparisons on the static scene test data. These results suggest that our pCSM-S and pCSM-DS significantly improve the quality of matte extraction. While pCSM-DS shows marginal improvements from pCSM-S because the intensity of polarized diffuse reflections tend to be small in the real-world, visual improvements can be seen in Fig. 5. On the headphone scene, pCSM-S mitigates the errors in PTM and PolarMatte due to polarized reflections, and pCSM-DS further improves the alpha matte by considering polarized diffuse reflections. A similar trend can be seen in the zoomed-in figures of the model airplane scene. + +We show qualitative comparisons on actual videos in the supplementary video, which demonstrate our accurate and temporally consistent alpha mattes and composite results. + +# 5. Conclusion + +This paper introduced Polarized Color Screen Matting, a single-shot, per-pixel matting theory that fully leverages color and polarization cues. The theoretical analysis on the proposed diffuse-specular polarimetric compositing equation derived two matting methods, pCSM-S and pCSM-DS. pCSM-S offers inherently efficient matting by simplifying the compositing equation. pCSM-DS revealed the color and polarization conditions to solve the pCSM problem using the full expression of the compositing equation. By introducing polarized diffuse regularization, pCSM-DS gains robustness to noise and applicability to any scene. Experiments on synthetic and real-world datasets show the stable and accurate matting capability of our methods compared with solely color-based or polarization-based methods. Future work may include more general reflectance models for further generalization and applicability. + +Acknowledgements: We would like to thank Brian Maffitt for his contributions as the Keylight expert. + +# References + +[1] Adobe Inc. Adobe After Effects. https://www.adobe.com/products/aftereffects.html. Accessed: 2025-02-07. 7 +[2] Yagiz Aksoy, Tunç Ozan Aydin, Marc Pollefeys, and Aljosa Smolić. Interactive high-quality green-screen keying via color unmixing. ACM TOG, 36(4):1, 2016. 1 +[3] Yunhao Ba, Alex Gilbert, Franklin Wang, Jina Yang, Rui Chen, Yiqin Wang, Lei Yan, Boxin Shi, and Achuta Kadambi. Deep shape from polarization. In ECCV, 2020. 3 +[4] Seung-Hwan Baek, Daniel S Jeon, Xin Tong, and Min H Kim. Simultaneous acquisition of polarimetric svbrdf and normals. ACM TOG, 37(6):268, 2018. 3 +[5] Seung-Hwan Baek, Tizian Zeltner, Hyunjin Ku, Inseung Hwang, Xin Tong, Wenzel Jakob, and Min H Kim. Image-based acquisition and modeling of polarimetric reflectance. ACM TOG, 39(4):139, 2020. 7, 8 +[6] Moshe Ben-Ezra. Segmentation with invisible keying signal. In CVPR, 2000. 2 +[7] Walter Beyer. Traveling-matte photography and the bluescreen system: a tutorial paper. Journal of the SMPTE, 74 (3):217-239, 1965. 2 +[8] Calibrite LLC. ColorChecker Classic. https:// calibrite.com/us/product/colorchecker-classic/. Accessed: 2025-02-07. 6 +[9] Zhaopeng Cui, Jinwei Gu, Boxin Shi, Ping Tan, and Jan Kautz. Polarimetric multi-view stereo. In CVPR, 2017. 3 +[10] Akshat Dave, Yannick Hold-Geoffroy, Milos Hasan, Kalyan Sunkavalli, and Ashok Veeraraghavan. Snapshot polarimetric diffuse-specular separation. Optics Express, 30(19): 34239-34255, 2022. 3 +[11] Kenji Enomoto, TJ Rhodes, Brian Price, and Gavin Miller. PolarMatte: Fully computational ground-truth-quality alpha matte extraction for images and video using polarized screen matting. In CVPR, 2024. 2, 3, 4, 6, 7, 8, 9, 10 +[12] Mikhail Erofeev, Yury Gitman, Dmitriy Vatolin, Alexey Fedorov, and Jue Wang. Perceptually motivated benchmark for video matting. In BMVC, 2015. 2 +[13] FLIR Systems Inc. BFS-U3-51S5PC-C USB 3.1. https://www.teledynevisionsolutions.com/ products/blackfly-s-usb3/?model=BFS-U3-51S5PC-C. Accessed: 2025-02-07. 6 +[14] Hisense Group Co., Ltd. 55" 4K UHD Hisense Android Smart TV (2021). https://www.hisense-usa.com/support/55a6g-55-4k-uhd-hisense android-smart-TV-2021. Accessed: 2025-02-07. 6 +[15] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 7 +[16] Soma Kajiyama, Taihe Piao, Ryo Kawahara, and Takahiro Okabe. Separating partially-polarized diffuse and specular reflection components under unpolarized light sources. In WACV, 2023. 3 +[17] Chenhao Li, Taishi Ono, Takeshi Uemori, Hajime Mihara, Alexander Gatto, Hajime Nagahara, and Yuseke Moriuchi. + +NeISF: Neural incident stokes field for geometry and material estimation. In CVPR, 2024. 3 +[18] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In CVPR, 2021. 1, 2, 7, 8, 9, 10 +[19] Hao Lu, Yutong Dai, Chunhua Shen, and Songcen Xu. Indices matter: Learning to index for deep image matting. In ICCV, 2019. 1 +[20] Youwei Lyu, Lingran Zhao, Si Li, and Boxin Shi. Shape from polarization with distant lighting estimation. IEEE TPAMI, 45(11):13991-14004, 2023. 3 +[21] Morgan McGuire and Wojciech Matusik. Real-time triangulation matting using passive polarization. In SIGGRAPH Sketches. 2006. 7, 8, 9, 10 +[22] Morgan McGuire, Wojciech Matusik, and William Yerazunis. Practical, Real-time Studio Matting using Dual Imagers. In Symposium on Rendering, 2006. 1, 2, 4 +[23] GyuTae Park, SungJoon Son, JaeYoung Yoo, SeHo Kim, and Nojun Kwak. Matteformer: Transformer-based image matting via prior-tokens. In CVPR, 2022. 1 +[24] Vlahos Petro. Composite photography utilizing sodium vapor illumination, 1963. US Patent 3,095,304. 2 +[25] Thomas Porter and Tom Duff. Compositing digital images. SIGGRAPH, 1984. 1 +[26] Steven A Shafer. Using color to separate reflection components. Color Research & Application, 10(4):210-218, 1985. 4 +[27] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE TPAMI, 41(2):271-284, 2019. 4 +[28] Dmitriy Smirnov, Chloe LeGendre, Xueming Yu, and Paul Debevec. Magenta green screen: Spectrally multiplexed alpha matting with deep colorization. In The Digital Production Symposium (DigiPro), 2023. 1, 2 +[29] Alvy Ray Smith and James F Blinn. Blue screen matting. In SIGGRAPH, 1996. 1, 2, 6, 8, 9, 10 +[30] George Gabriel Stokes. On the composition and resolution of streams of polarized light from different sources. Transactions of the Cambridge Philosophical Society, 9:399, 1851. 3 +[31] The Foundry Visionmongers Ltd. Keylight. https://learn.foundry.com/nuke/13.0/content/reference_guide/keyer_nodes/ keylight.html. Accessed: 2025-02-07. 1, 2, 7, 8, 9, 10 +[32] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In CVPR, 2017. 1 +[33] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In ACM MM, 2021. 2 \ No newline at end of file diff --git a/2025/Polarized Color Screen Matting/images.zip b/2025/Polarized Color Screen Matting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d0cd8449b065f3bec91e541f82d519c3f49fdfa4 --- /dev/null +++ b/2025/Polarized Color Screen Matting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ff5860e40f5b4362ee036b32f02a8571f6fbe21f8f7f5c670847124a773bc3d +size 605918 diff --git a/2025/Polarized Color Screen Matting/layout.json b/2025/Polarized Color Screen Matting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bd4dc72755b40ae3d3b829e68928cca21593c71c --- /dev/null +++ b/2025/Polarized Color Screen Matting/layout.json @@ -0,0 +1,10491 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 206, + 103, + 403, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 103, + 403, + 121 + ], + "spans": [ + { + "bbox": [ + 206, + 103, + 403, + 121 + ], + "type": "text", + "content": "Polarized Color Screen Matting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 143, + 220, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 143, + 220, + 158 + ], + "spans": [ + { + "bbox": [ + 145, + 143, + 220, + 158 + ], + "type": "text", + "content": "Kenji Enomoto" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 245, + 144, + 306, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 144, + 306, + 156 + ], + "spans": [ + { + "bbox": [ + 245, + 144, + 306, + 156 + ], + "type": "text", + "content": "Scott Cohen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 331, + 144, + 387, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 144, + 387, + 156 + ], + "spans": [ + { + "bbox": [ + 331, + 144, + 387, + 156 + ], + "type": "text", + "content": "Brian Price" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 265, + 158, + 345, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 158, + 345, + 170 + ], + "spans": [ + { + "bbox": [ + 265, + 158, + 345, + 170 + ], + "type": "text", + "content": "Adobe Research" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 412, + 144, + 465, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 144, + 465, + 156 + ], + "spans": [ + { + "bbox": [ + 412, + 144, + 465, + 156 + ], + "type": "text", + "content": "TJ Rhodes" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 193, + 119, + 252 + ], + "blocks": [ + { + "bbox": [ + 59, + 193, + 119, + 252 + ], + "lines": [ + { + "bbox": [ + 59, + 193, + 119, + 252 + ], + "spans": [ + { + "bbox": [ + 59, + 193, + 119, + 252 + ], + "type": "image", + "image_path": "0676404ca766dac43b5b8a02829ede517d2550acb824f2073535671628529079.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 254, + 119, + 314 + ], + "blocks": [ + { + "bbox": [ + 59, + 254, + 119, + 314 + ], + "lines": [ + { + "bbox": [ + 59, + 254, + 119, + 314 + ], + "spans": [ + { + "bbox": [ + 59, + 254, + 119, + 314 + ], + "type": "image", + "image_path": "e95f43ac7e90b9ebac418a26311ad2a3d08a3b8dcba7b6d7937d071c82fdbcc0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 322, + 139, + 333 + ], + "lines": [ + { + "bbox": [ + 99, + 322, + 139, + 333 + ], + "spans": [ + { + "bbox": [ + 99, + 322, + 139, + 333 + ], + "type": "text", + "content": "(a) Scene" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 120, + 193, + 180, + 253 + ], + "blocks": [ + { + "bbox": [ + 120, + 193, + 180, + 253 + ], + "lines": [ + { + "bbox": [ + 120, + 193, + 180, + 253 + ], + "spans": [ + { + "bbox": [ + 120, + 193, + 180, + 253 + ], + "type": "image", + "image_path": "ce6f9c81fb53732e98eb389493d40b34a96ae4e2f62a6b5212b7c4e53a3759a0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 120, + 254, + 180, + 314 + ], + "blocks": [ + { + "bbox": [ + 120, + 254, + 180, + 314 + ], + "lines": [ + { + "bbox": [ + 120, + 254, + 180, + 314 + ], + "spans": [ + { + "bbox": [ + 120, + 254, + 180, + 314 + ], + "type": "image", + "image_path": "c6438447227b25d72ca2a8e9bdb7ad59b4d825e488fd9b42f1cfcaa44919fa33.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 336, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 336, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 555, + 392 + ], + "type": "text", + "content": "Figure 1. (a) Scenes with fully polarized blue/green backgrounds under unpolarized known color illumination. (b) Commercial keying tools [31] use color information to extract an alpha matte, often struggling with hair details or motion blur due to the essential ill-posedness of the problem. (c) Polarimetric matting [22] uses polarization cues to extract an alpha matte without manual operations, often struggling with polarized foreground reflections as shown in the specular highlight. (d) Our Polarized Color Screen Matting leverages color and polarization cues, which enables more accurate alpha matte extraction for a wide range of real-world objects." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 181, + 193, + 242, + 253 + ], + "blocks": [ + { + "bbox": [ + 181, + 193, + 242, + 253 + ], + "lines": [ + { + "bbox": [ + 181, + 193, + 242, + 253 + ], + "spans": [ + { + "bbox": [ + 181, + 193, + 242, + 253 + ], + "type": "image", + "image_path": "69c96b4abb5b06b2071d3def511369b60a010a45cb65e915c17f764726b91d2d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 181, + 254, + 242, + 314 + ], + "blocks": [ + { + "bbox": [ + 181, + 254, + 242, + 314 + ], + "lines": [ + { + "bbox": [ + 181, + 254, + 242, + 314 + ], + "spans": [ + { + "bbox": [ + 181, + 254, + 242, + 314 + ], + "type": "image", + "image_path": "5cad6778570a9389a0a3ac9d09ce863bb96634e94f38a8e7b9b6927a96ebd662.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 322, + 282, + 333 + ], + "lines": [ + { + "bbox": [ + 205, + 322, + 282, + 333 + ], + "spans": [ + { + "bbox": [ + 205, + 322, + 282, + 333 + ], + "type": "text", + "content": "(b) Chroma keying" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 243, + 193, + 303, + 253 + ], + "blocks": [ + { + "bbox": [ + 243, + 193, + 303, + 253 + ], + "lines": [ + { + "bbox": [ + 243, + 193, + 303, + 253 + ], + "spans": [ + { + "bbox": [ + 243, + 193, + 303, + 253 + ], + "type": "image", + "image_path": "bed4f6d6d4e1a03af811c6f179af0054a70f98c21d97e9cf3b20b6141410b003.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 243, + 254, + 303, + 314 + ], + "blocks": [ + { + "bbox": [ + 243, + 254, + 303, + 314 + ], + "lines": [ + { + "bbox": [ + 243, + 254, + 303, + 314 + ], + "spans": [ + { + "bbox": [ + 243, + 254, + 303, + 314 + ], + "type": "image", + "image_path": "a5af54e8b8df1b94ca91f7ea6203238b69b6915162b021aa454df94f51b43b50.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 322, + 416, + 333 + ], + "lines": [ + { + "bbox": [ + 318, + 322, + 416, + 333 + ], + "spans": [ + { + "bbox": [ + 318, + 322, + 416, + 333 + ], + "type": "text", + "content": "(c) Polarimetric matting" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 304, + 193, + 365, + 253 + ], + "blocks": [ + { + "bbox": [ + 304, + 193, + 365, + 253 + ], + "lines": [ + { + "bbox": [ + 304, + 193, + 365, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 365, + 253 + ], + "type": "image", + "image_path": "f880d8aef1391ebe6642591716bc5ed694a8cbaac7b0994b45b6dfdf7bdba2d0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 304, + 254, + 365, + 314 + ], + "blocks": [ + { + "bbox": [ + 304, + 254, + 365, + 314 + ], + "lines": [ + { + "bbox": [ + 304, + 254, + 365, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 365, + 314 + ], + "type": "image", + "image_path": "50dd9ddca87d68425d7b2cc0222296af58854754c09d0bb75e17979a94c0db13.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 366, + 193, + 426, + 253 + ], + "blocks": [ + { + "bbox": [ + 366, + 193, + 426, + 253 + ], + "lines": [ + { + "bbox": [ + 366, + 193, + 426, + 253 + ], + "spans": [ + { + "bbox": [ + 366, + 193, + 426, + 253 + ], + "type": "image", + "image_path": "b527ec44273f4b9cdef9dcddba211b0c587f1db1d19a0b8f9cd2f5f5c62e7e0d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 366, + 254, + 426, + 314 + ], + "blocks": [ + { + "bbox": [ + 366, + 254, + 426, + 314 + ], + "lines": [ + { + "bbox": [ + 366, + 254, + 426, + 314 + ], + "spans": [ + { + "bbox": [ + 366, + 254, + 426, + 314 + ], + "type": "image", + "image_path": "5cc1348bab708cab9fc053d68e931caa1241e4ee4ede2a96b16ab744692f5ec0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 428, + 193, + 488, + 253 + ], + "blocks": [ + { + "bbox": [ + 428, + 193, + 488, + 253 + ], + "lines": [ + { + "bbox": [ + 428, + 193, + 488, + 253 + ], + "spans": [ + { + "bbox": [ + 428, + 193, + 488, + 253 + ], + "type": "image", + "image_path": "7673760209bd2dcc6c8eef3eb83ef57b2c52d9a53a3bf33d88034958b8ef9712.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 428, + 254, + 487, + 314 + ], + "blocks": [ + { + "bbox": [ + 428, + 254, + 487, + 314 + ], + "lines": [ + { + "bbox": [ + 428, + 254, + 487, + 314 + ], + "spans": [ + { + "bbox": [ + 428, + 254, + 487, + 314 + ], + "type": "image", + "image_path": "96572da303678e1e9df8f2f7b5de0272a3dd54b32048b9cd8a37ef830bd48393.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 322, + 508, + 332 + ], + "lines": [ + { + "bbox": [ + 473, + 322, + 508, + 332 + ], + "spans": [ + { + "bbox": [ + 473, + 322, + 508, + 332 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 490, + 193, + 549, + 253 + ], + "blocks": [ + { + "bbox": [ + 490, + 193, + 549, + 253 + ], + "lines": [ + { + "bbox": [ + 490, + 193, + 549, + 253 + ], + "spans": [ + { + "bbox": [ + 490, + 193, + 549, + 253 + ], + "type": "image", + "image_path": "a57e20190af89c48e969dc6cfd5742f021175db4a29626edc9793cea21285089.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 490, + 254, + 549, + 314 + ], + "blocks": [ + { + "bbox": [ + 490, + 254, + 549, + 314 + ], + "lines": [ + { + "bbox": [ + 490, + 254, + 549, + 314 + ], + "spans": [ + { + "bbox": [ + 490, + 254, + 549, + 314 + ], + "type": "image", + "image_path": "42fd110c552e5a0f523e6182015f3774a4d6b3765acfa0c31b02da02d2368bad.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 152, + 410, + 199, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 410, + 199, + 422 + ], + "spans": [ + { + "bbox": [ + 152, + 410, + 199, + 422 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 460, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 700 + ], + "type": "text", + "content": "This paper considers the long-standing problem of extracting alpha mattes from video using a known background. While various color-based or polarization-based approaches have been studied in past decades, the problem remains ill-posed because the solutions solely rely on either color or polarization. We introduce Polarized Color Screen Matting, a single-shot, per-pixel matting theory for alpha matte and foreground color recovery using both color and polarization cues. Through a theoretical analysis of our diffuse-specular polarimetric compositing equation, we derive practical closed-form matting methods with their solvability conditions. Our theory concludes that an alpha matte can be extracted without manual corrections using off-the-shelf equipment such as an LCD monitor, polarization camera, and unpolarized lights with calibrated color. Experiments on synthetic and real-world datasets verify the validity of our theory and show the capability of our matting methods on real videos with quantitative and qualitative comparisons to color-based and polarization-based matting methods." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "spans": [ + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 434, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 530 + ], + "type": "text", + "content": "Image and video editing rely on compositing [25] to blend images. The quality of the composite depends on how accurately the alpha matte depicts opacity. Modern matte extraction methods use inherent information such as color or polarization to extract the alpha matte. However, they are forced to compromise on quality of foreground color, alpha, or both. This paper combines color and polarimetric observations to constrain the problem." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Color-based alpha matte extraction methods are most commonly used for compositing due to equipment availability. Chroma keying [2, 29] is the de facto standard in film making, but because it is underconstrained, it often requires manual corrections to extract high-quality alpha mattes. In addition to this, it often struggles with motion blur and transparent objects as depicted in Fig. 1. Recent learning-based matting methods [18, 19, 23, 32] can sometimes extract acceptable mattes but they still struggle with complex situations; potential improvements in modeling and dataset remain. Magenta green screen [28] can extract a more precise alpha matte by sacrificing the green color channel, but this later requires a restoration of the green color channel using machine-learning colorization. All these challenges are rooted in the fact that the use of" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "391" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 288, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 288, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 288, + 83 + ], + "type": "text", + "content": "color alone to recover an alpha matte is underconstrained." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 84, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 84, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 84, + 294, + 191 + ], + "type": "text", + "content": "Alternatively, polarization-based matting methods [6, 11, 22] can extract the alpha matte using the polarization of light. For a scene backlit with polarized light, there is less ambiguity between phase angle than color, thus physically accurate alpha mattes can be obtained for the majority of pixels. However, in cases where the foreground subject reflects polarized light whose phase angle is comparable to the background, the alpha matte quality will be degraded, as shown in the headphone and glass scenes in Fig. 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 192, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 192, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 192, + 296, + 407 + ], + "type": "text", + "content": "We introduce Polarized Color Screen Matting (pCSM), a single-shot, per-pixel matting theory that uses both color and polarization cues. Our theory is built upon our proposed diffuse-specular polarimetric compositing equation, where foreground appearance is explicitly modeled by a summation of diffuse and specular reflections. This new compositing equation steps forward a theoretical analysis of the matting problem, resulting in two matting methods: (1) " + }, + { + "bbox": [ + 56, + 192, + 296, + 407 + ], + "type": "inline_equation", + "content": "p\\text{CSM-S}" + }, + { + "bbox": [ + 56, + 192, + 296, + 407 + ], + "type": "text", + "content": ", a fast and robust matting method with an approximation of unpolarized diffuse reflections that is applicable to any scene, and (2) " + }, + { + "bbox": [ + 56, + 192, + 296, + 407 + ], + "type": "inline_equation", + "content": "p\\text{CSM-DS}" + }, + { + "bbox": [ + 56, + 192, + 296, + 407 + ], + "type": "text", + "content": ", an accurate and approximation-free matting method. Through our theoretical analysis, we found solvability conditions of pCSM-DS in terms of color and polarized reflection from the foreground subject. For the practical use of our theoretical matting method, we introduce a physics-inspired regularization, which grants our methods significant robustness to noisy real-world observations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 407, + 272, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 407, + 272, + 418 + ], + "spans": [ + { + "bbox": [ + 67, + 407, + 272, + 418 + ], + "type": "text", + "content": "The main contributions of this paper are threefold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 419, + 295, + 503 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 419, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 295, + 455 + ], + "type": "text", + "content": "- A diffuse-specular polarimetric compositing equation that unlocks further theoretical analysis of the polarized screen matting problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 455, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 479 + ], + "type": "text", + "content": "- A fast and simple matting method, pCSM-S, that is applicable to any scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 503 + ], + "type": "text", + "content": "- An accurate and approximation-free matting method, pCSM-DS, with solvability conditions." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 503, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 539 + ], + "type": "text", + "content": "We show quantitatively and qualitatively that our methods produce more accurate results than state-of-the-art solely color-based or polarization-based matting methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 550, + 142, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 142, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 142, + 562 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": "Early methods for computing alpha mattes from video have focused on capturing footage against a solid-colored background [7], often referred to as chroma keying. Indeed, chroma keying is an industry standard, implemented in commercial products [31] and used for generating video matting datasets [18, 33]. However, despite its popularity, chroma keying is underconstrained and often performs poorly unless an expert carefully sets up the capture environment and manually post-processes the matte." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Due to the shortcomings of standard chroma keying, further work has focused on using color to facilitate matte extraction. The sodium vapor matting process [24] constrains" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 324 + ], + "type": "text", + "content": "the problem by narrowing the spectrum of backlight emitted from a low-pressure sodium vapor lamp. A beamsplitter and bandpass filter allow the background sodium vapor light to be captured on one strip of film, forming an alpha matte, and the foreground subject's color to be captured through a notch reject filter on a second strip of film, blocking the sodium vapor light. More recently, magenta green screen [28] looked to eliminate the need for a specialized camera system by placing a foreground subject strongly lit by magenta light in front of a green background to observe the green spectrum solely at the background. However, the foreground magenta light loses the true foreground color, requiring machine-learning colorization to recover the lost green color of the foreground. In theory, triangulation matting [29] can compute the true alpha matte by taking two (or more) images of a static subject in front of different colored backgrounds to triangulate the alpha matte. However, since the images must be taken at different times to allow the background color to change, triangulation matting is not suitable for dynamic scenes without sacrificing motion blur, such as with stop-motion photography [12]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 333, + 553, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 553, + 441 + ], + "type": "text", + "content": "Instead of using color to constrain the capture system, polarimetric triangulation matting (PTM) [22] employs polarization. PTM uses a beamsplitter with two polarization filters to simultaneously capture two polarization images of a target scene against a polarized background screen. These images can serve as input for triangulation matting if the foreground reflects only unpolarized light; however, in practice, polarized foreground reflections are observed, which cause perceivable errors in foreground subjects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 714 + ], + "type": "text", + "content": "Addressing the issue of polarized foreground reflections, PolarMatte [11] reformulates the polarized screen matting problem by leveraging four polarization images captured with a modern polarization camera (see Fig. 2). In this framework, an extended robust PTM is first applied to pixels with negligible polarized reflections, and the resulting high-confidence alpha mattes are subsequently interpolated to pixels with polarized foreground reflections using spatial context through a polarimetric matting Laplacian. However, two critical limitations persist. First, because PolarMatte is not formulated as a per-pixel method, establishing solvability conditions remains challenging, thereby complicating both its theoretical analysis and practical reliability. Second, its exclusive reliance on polarization cues (i.e., the use of white polarized background) renders it vulnerable when the polarized reflection aligns with the background polarity—a situation often encountered in the presence of specular highlights (e.g., from jewelry or glossy skin). In this paper, we aim to overcome these limitations by jointly integrating color and polarization cues, which more robustly constrains the matting problem without adding significant complexity to the capture process." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "392" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 295, + 218 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 295, + 218 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 295, + 218 + ], + "type": "image", + "image_path": "371d023ffd82abc666dc055f83eaafda5f5050a9af943cffc5707a6aef7ee313.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 226, + 296, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 281 + ], + "type": "text", + "content": "Figure 2. Illustration of our capture setup. A recent polarization camera with a color polarization filter array sensor captures four polarization images in a single-shot. Our method is tailored for the polarization images of a foreground subject with a known polarized background under known color illuminations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 234, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 234, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 234, + 314 + ], + "type": "text", + "content": "3. Polarized Color Screen Matting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 320, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 296, + 440 + ], + "type": "text", + "content": "This section delivers a proof that a single-shot and per-pixel matting problem has a closed-form solution given: (1) the background is known and polarized, (2) the foreground illumination color is known, and (3) a diffuse-specular reflectance model for the foreground subject. Our pCSM theory is tailored for polarization images at four angles " + }, + { + "bbox": [ + 55, + 320, + 296, + 440 + ], + "type": "inline_equation", + "content": "\\left\\{0, \\frac{\\pi}{4}, \\frac{\\pi}{2}, \\frac{3\\pi}{4}\\right\\}" + }, + { + "bbox": [ + 55, + 320, + 296, + 440 + ], + "type": "text", + "content": " captured by a recent polarization camera shown in Fig. 2, similarly to PolarMatte [11]. The reader is encouraged to reference Table 5 for the notations used in the following sections." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 446, + 296, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 472 + ], + "type": "text", + "content": "3.1. Diffuse-Specular Polarimetric Compositing Equation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 477, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 296, + 550 + ], + "type": "text", + "content": "The matting problem is formulated as a decomposition of the color vector of a given pixel into foreground and background color vectors and the opacity of the foreground, commonly called an alpha matte. In a linear color space, the compositing process of a polarization image is described in the polarimetric compositing equation [11]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 123, + 557, + 294, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 557, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 123, + 557, + 294, + 570 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {\\phi} = \\alpha \\mathbf {f} _ {\\phi} + (1 - \\alpha) \\mathbf {b} _ {\\phi}, \\tag {1}", + "image_path": "38a6d0c27040eca200f4d759cbd7ab6d04a5700d6d47537518cb368a0346f9ab.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 577, + 296, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 296, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 577, + 296, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\phi}, \\mathbf{f}_{\\phi}, \\mathbf{b}_{\\phi} \\in \\mathbb{R}_+^3" + }, + { + "bbox": [ + 55, + 577, + 296, + 648 + ], + "type": "text", + "content": " are composite, foreground, and background color polarization vectors at an angle " + }, + { + "bbox": [ + 55, + 577, + 296, + 648 + ], + "type": "inline_equation", + "content": "\\phi \\in \\{0, \\frac{\\pi}{4}, \\frac{\\pi}{2}, \\frac{3\\pi}{4}\\}" + }, + { + "bbox": [ + 55, + 577, + 296, + 648 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 577, + 296, + 648 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 55, + 577, + 296, + 648 + ], + "type": "text", + "content": " is an alpha matte. We employ a diffuse-specular reflectance model [16] on the foreground color, which is the sum of diffuse and specular terms:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 654, + 295, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 654, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 63, + 654, + 295, + 684 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {\\phi} = \\underbrace {\\left(u _ {d} + p _ {d} \\cos^ {2} \\left(\\theta_ {d} - \\phi\\right)\\right) \\mathbf {f} _ {d}} _ {\\text {d i f f u s e t e r m}} + \\underbrace {\\left(u _ {s} + p _ {s} \\cos^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {s}} _ {\\text {s p e c u l a r t e r m}}. \\tag {2}", + "image_path": "e7b60f1381152773384798d1c901eb02e2e7339b90730d3246fc986288d8ae6b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Here, the diffuse and specular terms are decomposed into unpolarized and polarized terms, where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "u_{d}, p_{d} \\in \\mathbb{R}_{+}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "are the intensities of the unpolarized and polarized diffuse reflections, respectively, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "u_{s}, p_{s} \\in \\mathbb{R}_{+}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " are the same for the specular reflection. " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d, \\mathbf{f}_s \\in S_+^2 \\subset \\mathbb{R}_+^3" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " denote unit color vectors of diffuse and specular reflections, respectively. " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\theta_d, \\theta_s \\in [0, \\pi)" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " are phase angles of the diffuse and specular reflections, and it is known that " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\theta_d = \\theta_s \\pm \\frac{\\pi}{2}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " under unpolarized illumination [3, 9, 16]. Therefore, the diffuse term in Eq. (2) can be rewritten using " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 177, + 555, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 177, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 378, + 177, + 555, + 192 + ], + "type": "interline_equation", + "content": "\\left(u _ {d} + p _ {d} \\sin^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {d}. \\tag {3}", + "image_path": "a4237f8c143636431e99ce09ae9bb06d9d1710e10ec6c003d81c7be13394a9d5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 201, + 554, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 554, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 554, + 224 + ], + "type": "text", + "content": "In our problem setting, the background is fully polarized, which can be described in" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 382, + 235, + 555, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 235, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 382, + 235, + 555, + 249 + ], + "type": "interline_equation", + "content": "\\mathbf {b} _ {\\phi} = p _ {b} \\cos^ {2} (\\theta_ {b} - \\phi) \\mathbf {q}, \\tag {4}", + "image_path": "6a5baaaa3992ce5967b715b2dd4566b62b0bd778e8fcc610d0595e2c9c74fd61.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 258, + 554, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 554, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 554, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 258, + 554, + 307 + ], + "type": "inline_equation", + "content": "p_b \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 313, + 258, + 554, + 307 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 258, + 554, + 307 + ], + "type": "inline_equation", + "content": "\\theta_b \\in [0, \\pi)" + }, + { + "bbox": [ + 313, + 258, + 554, + 307 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 258, + 554, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in S_+^2 \\subset \\mathbb{R}_+^3" + }, + { + "bbox": [ + 313, + 258, + 554, + 307 + ], + "type": "text", + "content": " are the intensity, phase angle, and unit color vector of the backlight, respectively. Finally, the polarimetric compositing equation (1) can be written as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 315, + 553, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 315, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 323, + 315, + 553, + 352 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {c} _ {\\phi} = \\alpha \\left(\\left(u _ {d} + p _ {d} \\sin^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {d} + \\left(u _ {s} + p _ {s} \\cos^ {2} \\left(\\theta_ {s} - \\phi\\right)\\right) \\mathbf {f} _ {s}\\right) \\\\ + (1 - \\alpha) p _ {b} \\cos^ {2} (\\theta_ {b} - \\phi) \\mathbf {q}. \\tag {5} \\\\ \\end{array}", + "image_path": "edc37f5459f03ed7bb1584be7389ffc57796460c76fe818e6d240769e30fe400.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 354, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 437 + ], + "type": "text", + "content": "We call this compositing process the diffuse-specular polarimetric compositing equation. The diffuse-specular reflectance model is widely used in polarimetric vision problems, such as diffuse-specular separation [10, 16], shape-from-polarization [17, 20], and polarimetric BRDF acquisition [4]. We extend it to the compositing equation to represent opacity." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "type": "text", + "content": "For ease of theoretical analysis, we transform observed polarization images into Stokes parameters [30]. This paper only considers linear polarization, and the Stokes parameters can also be described in the form of compositing equations:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 359, + 506, + 555, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 506, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 359, + 506, + 555, + 535 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {0} = \\frac {1}{2} \\sum_ {\\phi} \\mathbf {c} _ {\\phi} \\stackrel {\\mathrm {d e f}} {=} \\mathbf {d} _ {0} + (1 - \\alpha) \\mathbf {q} _ {0}, \\tag {6}", + "image_path": "842ca7e37d89227c9168b60a20051c75690eac119e92369a13d4e0222a4d3757.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 359, + 539, + 553, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 539, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 359, + 539, + 553, + 552 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {1} = \\mathbf {c} _ {0} - \\mathbf {c} _ {\\frac {\\pi}{2}} \\stackrel {\\mathrm {d e f}} {=} \\mathbf {d} _ {1} + (1 - \\alpha) \\mathbf {q} _ {1}, \\tag {7}", + "image_path": "2fe3c438cb9c2641ead4ea0595863b3d2bcb1b2de9638a50da922ce254953f2d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 359, + 555, + 553, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 555, + 553, + 571 + ], + "spans": [ + { + "bbox": [ + 359, + 555, + 553, + 571 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {2} = \\mathbf {c} _ {\\frac {\\pi}{4}} - \\mathbf {c} _ {\\frac {3 \\pi}{4}} \\stackrel {\\text {d e f}} {=} \\mathbf {d} _ {2} + (1 - \\alpha) \\mathbf {q} _ {2}, \\tag {8}", + "image_path": "cfd51558e1816dbd9bcbafaa5775edd6bdd1e3fde6623eb885a02fa361b30d86.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 578, + 342, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 578, + 342, + 588 + ], + "spans": [ + { + "bbox": [ + 314, + 578, + 342, + 588 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 600, + 553, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 600, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 320, + 600, + 553, + 613 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {0} = \\alpha \\left(\\left(2 u _ {d} + p _ {d}\\right) \\mathbf {f} _ {d} + \\left(2 u _ {s} + p _ {s}\\right) \\mathbf {f} _ {s}\\right), \\quad \\mathbf {q} _ {0} = p _ {b} \\mathbf {q}, \\tag {9}", + "image_path": "ddf096470994531bb0e90dcd2fec90de71b234759c647b3fa00e63fa70a963f3.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 325, + 616, + 553, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 616, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 325, + 616, + 553, + 628 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {1} = \\alpha \\cos 2 \\theta_ {s} \\left(p _ {s} \\mathbf {f} _ {s} - p _ {d} \\mathbf {f} _ {d}\\right), \\mathbf {q} _ {1} = p _ {b} \\cos 2 \\theta_ {b} \\mathbf {q}, \\tag {10}", + "image_path": "8ea5f02ce98112754990701e460b4ab5a1ef2396bef5afb77b52fd52f990b18c.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 326, + 631, + 553, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 631, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 326, + 631, + 553, + 643 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {2} = \\alpha \\sin 2 \\theta_ {s} \\left(p _ {s} \\mathbf {f} _ {s} - p _ {d} \\mathbf {f} _ {d}\\right), \\mathbf {q} _ {2} = p _ {b} \\sin 2 \\theta_ {b} \\mathbf {q}. \\tag {11}", + "image_path": "0811ae5995e75e8ee8f990d5869551563971c945d2f06390bb59657345e7a2c1.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "content": " is equivalent to an ordinary color image, and " + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_1" + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_2" + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "content": " represent the polarization properties of an observation." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "This paper aims to recover the alpha matte " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " given the polarization images " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\phi}" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " and the background images " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{\\phi}" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " for all the angles " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\phi \\in \\{0, \\frac{\\pi}{4}, \\frac{\\pi}{2}, \\frac{3\\pi}{4}\\}" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "393" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 293, + 236 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 293, + 236 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 293, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 293, + 236 + ], + "type": "image", + "image_path": "f44b2a1cc6a24d367f7a5c879f93c4118db12b0edb918cb8db7717f51ef82718.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 248, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 304 + ], + "type": "text", + "content": "Figure 3. (Left) A scene of a headphone in front of a blue polarized background and its polarized intensity computed as " + }, + { + "bbox": [ + 55, + 248, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\sqrt{\\mathbf{s}_1^2 + \\mathbf{s}_2^2}" + }, + { + "bbox": [ + 55, + 248, + 296, + 304 + ], + "type": "text", + "content": ". (Right) The color distribution of the polarized intensities, showing that the majority of the polarized intensities derive from background and specular reflections." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 324, + 274, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 274, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 274, + 338 + ], + "type": "text", + "content": "3.2. pCSM with Polarized Specular Reflection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 475 + ], + "type": "text", + "content": "This section introduces a pCSM method, which we call " + }, + { + "bbox": [ + 55, + 342, + 296, + 475 + ], + "type": "inline_equation", + "content": "p\\mathrm{CSM - S}" + }, + { + "bbox": [ + 55, + 342, + 296, + 475 + ], + "type": "text", + "content": " that highlights the polarized specular reflection. Prior polarimetric matting methods [11, 22] show that polarized reflections from a foreground subject are a major source of error in the alpha matte estimation. In particular, polarized specular reflections have larger intensities than polarized diffuse reflections and a dominant effect on the estimated alpha matte (see Fig. 3). This observation motivates us to simplify the polarimetric compositing equation and Stokes parameters without the diffuse term (i.e., " + }, + { + "bbox": [ + 55, + 342, + 296, + 475 + ], + "type": "inline_equation", + "content": "p_d = 0" + }, + { + "bbox": [ + 55, + 342, + 296, + 475 + ], + "type": "text", + "content": "). With this simplification, Eqs. (7) and (8) can be rewritten as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 486, + 295, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 486, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 295, + 498 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {1} = \\alpha p _ {s} \\cos 2 \\theta_ {s} \\mathbf {f} _ {s} + (1 - \\alpha) \\mathbf {q} _ {1}, \\tag {12}", + "image_path": "0495910b13baab0891a5fe39687d5df35da270d60504bbece8a3f43b8d28b0d2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 500, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 500, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 295, + 514 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {2} = \\alpha p _ {s} \\sin 2 \\theta_ {s} \\mathbf {f} _ {s} + (1 - \\alpha) \\mathbf {q} _ {2}. \\tag {13}", + "image_path": "75e7288e46c2bdd97b817d72615baef2cf0fb20e0e3c25e15d7ae9e5f8cb676a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 523, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 296, + 606 + ], + "type": "text", + "content": "Assuming the dichromatic reflectance model [26], the specular reflection color " + }, + { + "bbox": [ + 55, + 523, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 55, + 523, + 296, + 606 + ], + "type": "text", + "content": " is independent of the object's albedo and is equal to the ambient light color. Therefore, we can assume the specular reflection color " + }, + { + "bbox": [ + 55, + 523, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 55, + 523, + 296, + 606 + ], + "type": "text", + "content": " to be known if the ambient light color is calibrated or ambient light is natural and white. In the following, we assume that the specular reflection color " + }, + { + "bbox": [ + 55, + 523, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 55, + 523, + 296, + 606 + ], + "type": "text", + "content": " is known." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 608, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 295, + 632 + ], + "type": "text", + "content": "With Eqs. (12) and (13), the alpha matte can be estimated by solving the following linear system:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 72, + 639, + 295, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 639, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 72, + 639, + 295, + 693 + ], + "type": "interline_equation", + "content": "\\underbrace {\\left[ \\begin{array}{l} \\mathbf {s} _ {1} - \\mathbf {q} _ {1} \\\\ \\mathbf {s} _ {2} - \\mathbf {q} _ {2} \\end{array} \\right]} _ {\\mathbf {y}} = \\underbrace {\\left[ \\begin{array}{c c c} - \\mathbf {q} _ {1} & \\mathbf {f} _ {s} & \\mathbf {0} \\\\ - \\mathbf {q} _ {2} & \\mathbf {0} & \\mathbf {f} _ {s} \\end{array} \\right]} _ {\\mathbf {A}} \\underbrace {\\left[ \\begin{array}{c} \\alpha \\\\ \\alpha p _ {s} \\cos 2 \\theta_ {s} \\\\ \\alpha p _ {s} \\sin 2 \\theta_ {s} \\end{array} \\right]} _ {\\mathbf {x}}. \\tag {14}", + "image_path": "a2831f04f0ed89ae694ef67446e738459e1840a4b92d374046ed7c12f0865e33.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "Thus, this linear system has a closed-form solution, and in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 73, + 497, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 497, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 497, + 84 + ], + "type": "text", + "content": "practice, the alpha matte can be computed as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 87, + 555, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 87, + 555, + 114 + ], + "spans": [ + { + "bbox": [ + 321, + 87, + 555, + 114 + ], + "type": "interline_equation", + "content": "\\alpha = 1 - \\frac {\\mathbf {q} _ {1} ^ {\\top} \\mathbf {s} _ {1} + \\mathbf {q} _ {2} ^ {\\top} \\mathbf {s} _ {2} - \\left(\\mathbf {q} _ {1} ^ {\\top} \\mathbf {f} _ {s}\\right) \\left(\\mathbf {s} _ {1} ^ {\\top} \\mathbf {f} _ {s}\\right) - \\left(\\mathbf {q} _ {2} ^ {\\top} \\mathbf {f} _ {s}\\right) \\left(\\mathbf {s} _ {2} ^ {\\top} \\mathbf {f} _ {s}\\right)}{p _ {b} ^ {2} - \\left(\\mathbf {q} _ {1} ^ {\\top} \\mathbf {f} _ {s}\\right) ^ {2} - \\left(\\mathbf {q} _ {2} ^ {\\top} \\mathbf {f} _ {s}\\right) ^ {2}} \\tag {15}", + "image_path": "31030c918ecb13b10d16d6a7511edc2b21848f6707755f38ff3c9a814f898931.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 118, + 555, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 555, + 142 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 555, + 142 + ], + "type": "text", + "content": "without computing the pseudo inverse matrix of " + }, + { + "bbox": [ + 313, + 118, + 555, + 142 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 118, + 555, + 142 + ], + "type": "text", + "content": " for each pixel of a given image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 148, + 555, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 173 + ], + "type": "text", + "content": "Proposition 1. If " + }, + { + "bbox": [ + 313, + 148, + 555, + 173 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\neq \\mathbf{f}_s" + }, + { + "bbox": [ + 313, + 148, + 555, + 173 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 313, + 148, + 555, + 173 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 148, + 555, + 173 + ], + "type": "text", + "content": "CSM-S problem has a closed-form solution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 179, + 533, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 533, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 533, + 192 + ], + "type": "text", + "content": "Proof. The determinant of " + }, + { + "bbox": [ + 313, + 179, + 533, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{\\top}\\mathbf{A}" + }, + { + "bbox": [ + 313, + 179, + 533, + 192 + ], + "type": "text", + "content": " can be calculated as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 200, + 555, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 200, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 361, + 200, + 555, + 220 + ], + "type": "interline_equation", + "content": "\\det \\left(\\mathbf {A} ^ {\\top} \\mathbf {A}\\right) = p _ {b} ^ {2} \\left(1 - \\left(\\mathbf {q} ^ {\\top} \\mathbf {f} _ {s}\\right) ^ {2}\\right). \\tag {16}", + "image_path": "1aa579833efd3c3b12d6981cef36a93fe446086804c7182630c7c47352c82d89.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 228, + 555, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 254 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 313, + 228, + 555, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{q},\\mathbf{f}_s\\geq \\mathbf{0}" + }, + { + "bbox": [ + 313, + 228, + 555, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 228, + 555, + 254 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{q}\\| _2 = \\| \\mathbf{f}_s\\| _2 = 1" + }, + { + "bbox": [ + 313, + 228, + 555, + 254 + ], + "type": "text", + "content": " ,det " + }, + { + "bbox": [ + 313, + 228, + 555, + 254 + ], + "type": "inline_equation", + "content": "(\\mathbf{A}^{\\top}\\mathbf{A}) = 0" + }, + { + "bbox": [ + 313, + 228, + 555, + 254 + ], + "type": "text", + "content": " if and only if " + }, + { + "bbox": [ + 313, + 228, + 555, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = \\mathbf{f}_s" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 261, + 554, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 554, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 554, + 309 + ], + "type": "text", + "content": "Therefore, our pCSM-S can be always applied by using a backlight color sufficiently different from the ambient light color; for instance, a green background with natural white ambient light." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 316, + 554, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 340 + ], + "type": "text", + "content": "3.3. pCSM with Polarized Diffuse-Specular Reflection" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "content": "This section introduces another pCSM method, which we call " + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "inline_equation", + "content": "p\\mathrm{CSM - DS}" + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "content": ", using the full expression of the diffuse-specular polarimetric compositing equation without any simplification. Research shows that more general modeling leads to more ambiguity [27], and our pCSM-DS problem is no exception. Here, we present the solvability conditions of our pCSM-DS and the solutions when each condition is satisfied. Since the pCSM-DS problem is equivalent to pCSM-S if " + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d = \\mathbf{f}_s" + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "inline_equation", + "content": "p_d = 0" + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "content": ", we suppose " + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d \\neq \\mathbf{f}_s" + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "inline_equation", + "content": "p_d > 0" + }, + { + "bbox": [ + 313, + 346, + 555, + 467 + ], + "type": "text", + "content": " in the following." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "type": "text", + "content": "Proposition 2. If a strong color condition, span " + }, + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{q},\\mathbf{f}_d,\\mathbf{f}_s\\} = \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "type": "text", + "content": " is satisfied, the pCSM-DS problem has a closed-form solution." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 517, + 470, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 517, + 470, + 529 + ], + "spans": [ + { + "bbox": [ + 314, + 517, + 470, + 529 + ], + "type": "text", + "content": "Proof. Since Eq. (6) can be written as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 329, + 536, + 554, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 536, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 329, + 536, + 554, + 576 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {0} - p _ {b} \\mathbf {q} = \\left[ \\begin{array}{l l l} - p _ {b} \\mathbf {q} & \\mathbf {f} _ {d} & \\mathbf {f} _ {s} \\end{array} \\right] \\left[ \\begin{array}{c} \\alpha \\\\ \\alpha \\left(2 u _ {d} + p _ {d}\\right) \\\\ \\alpha \\left(2 u _ {s} + p _ {s}\\right) \\end{array} \\right], \\tag {17}", + "image_path": "81cec20c3e753ab23cb4e871d5dff14b231ada18abcd5b309bb8c17ab16b888e.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": "if " + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d" + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": " are linearly independent, the matrix in Eq. (17) is invertible, and it has a closed-form solution." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 613, + 555, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 555, + 669 + ], + "type": "text", + "content": "The missing variable to solve Eq. (17) is the diffuse color vector " + }, + { + "bbox": [ + 313, + 613, + 555, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d" + }, + { + "bbox": [ + 313, + 613, + 555, + 669 + ], + "type": "text", + "content": ", which is inherently difficult to be known. However, the use of polarization images with a fully polarized background allows us to compute a linear combination of " + }, + { + "bbox": [ + 313, + 613, + 555, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d" + }, + { + "bbox": [ + 313, + 613, + 555, + 669 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 613, + 555, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 313, + 613, + 555, + 669 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 675, + 553, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 675, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 675, + 553, + 713 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {f} _ {d s} = \\mathbf {c} _ {\\theta_ {b} - \\frac {\\pi}{2}} \\\\ = \\alpha \\left(\\left(u _ {d} + p _ {d} \\cos^ {2} \\left(\\theta_ {s} - \\theta_ {b}\\right)\\right) \\mathbf {f} _ {d} + \\left(u _ {s} + p _ {s} \\sin^ {2} \\left(\\theta_ {s} - \\theta_ {b}\\right)\\right) \\mathbf {f} _ {s}\\right). \\tag {18} \\\\ \\end{array}", + "image_path": "6d80988eb04a91acd0f0b95d242cb08690e07526bb2b4ddc99cd4a480b59cdc3.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "394" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "We call this vector " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ds}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " the foreground basis. Since the unpolarized reflection is dominant in the diffuse reflection, the unpolarized diffuse intensity " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "u_{d}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " is almost always nonzero when the polarized diffuse intensity is positive " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "p_d > 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". Hence, the foreground basis and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " are linearly independent unless " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_d = \\mathbf{f}_s" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". With this foreground basis, Eq. (17) can be rewritten as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 96, + 163, + 295, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 163, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 96, + 163, + 295, + 204 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {0} - p _ {b} \\mathbf {q} = \\underbrace {\\left[ - p _ {b} \\mathbf {q} \\quad \\mathbf {f} _ {d s} \\quad \\mathbf {f} _ {s} \\right]} _ {\\mathbf {B}} \\left[ \\begin{array}{l} \\alpha \\\\ v _ {d s} \\\\ v _ {s} \\end{array} \\right], \\tag {19}", + "image_path": "e15d9b117faa3701200323d39d8c7929fa40404e1aaedf2af3ca83940112de0e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 295, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 260 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 213, + 295, + 260 + ], + "type": "inline_equation", + "content": "v_{ds}, v_s \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 213, + 295, + 260 + ], + "type": "text", + "content": ". Therefore, if the matrix " + }, + { + "bbox": [ + 55, + 213, + 295, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 213, + 295, + 260 + ], + "type": "text", + "content": " is full-rank, i.e., the strong color condition is satisfied, this linear system has a closed-form solution, and the alpha matte is uniquely determined. We call this method linear " + }, + { + "bbox": [ + 55, + 213, + 295, + 260 + ], + "type": "inline_equation", + "content": "pCSM-DS" + }, + { + "bbox": [ + 55, + 213, + 295, + 260 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 261, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 296, + 308 + ], + "type": "text", + "content": "We then consider the case of " + }, + { + "bbox": [ + 55, + 261, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\operatorname{span}\\{\\mathbf{f}_d, \\mathbf{f}_s\\}" + }, + { + "bbox": [ + 55, + 261, + 296, + 308 + ], + "type": "text", + "content": ". In this case, Eq. (6) can be satisfied with multiple alpha matte values, hence, we leverage Eqs. (7) and (8) to derive a unique alpha matte." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "text", + "content": "Proposition 3. If a weak color condition, " + }, + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\not\\propto p_s\\mathbf{f}_s - p_d\\mathbf{f}_d" + }, + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "text", + "content": " and a polarization condition, " + }, + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\theta_b \\neq \\theta_s" + }, + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "text", + "content": " nor " + }, + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\theta_b \\neq \\theta_d" + }, + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "text", + "content": ", are satisfied, the " + }, + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "text", + "content": "CSM-DS problem has a unique solution, even if the strong color condition is not satisfied." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 373, + 217, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 217, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 217, + 385 + ], + "type": "text", + "content": "Proof. See the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 456 + ], + "type": "text", + "content": "Proposition 3 shows the well-posedness of the pCSM-DS problem when the weak color condition and polarization condition are satisfied; therefore, here we derive a solution to the pCSM-DS problem. The alpha matte can be found as a minimizer of the following objective function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 463, + 296, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 501 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f (\\alpha , \\mathbf {w}, t) = \\| \\mathbf {s} _ {1} - \\mathbf {d} _ {1} - (1 - \\alpha) \\mathbf {q} _ {1} \\| _ {2} ^ {2} + \\| \\mathbf {s} _ {2} - \\mathbf {d} _ {2} - (1 - \\alpha) \\mathbf {q} _ {2} \\| _ {2} ^ {2} \\\\ = \\| \\mathbf {s} _ {1} - \\mathbf {F} \\mathbf {w} - (1 - \\alpha) \\mathbf {q} _ {1} \\| _ {2} ^ {2} + \\| \\mathbf {s} _ {2} - t \\mathbf {F} \\mathbf {w} - (1 - \\alpha) \\mathbf {q} _ {2} \\| _ {2} ^ {2} \\tag {20} \\\\ \\end{array}", + "image_path": "1687e17fe7b1ec562be9a8d48ef9c7838f4e005a4b942d661086fbaf4c0f00c0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_1" + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_2" + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": ", defined in Eqs. (10) and (11), are represented by a basis matrix " + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = [\\mathbf{f}_{ds}\\mathbf{f}_s]\\in \\mathbb{R}^{3\\times 2}" + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": ", coefficient vector " + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = [w_{ds}w_s]^\\top \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": ", and a scalar " + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "inline_equation", + "content": "t\\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 502, + 296, + 552 + ], + "type": "text", + "content": ". This objective function is written in matrix form as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 559, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 559, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 67, + 559, + 295, + 620 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f (\\alpha , \\mathbf {w}, t) = \\left\\| \\left[ \\begin{array}{l} \\mathbf {s} _ {1} - \\mathbf {q} _ {1} \\\\ \\mathbf {s} _ {2} - \\mathbf {q} _ {2} \\end{array} \\right] - \\left[ \\begin{array}{l l} - \\mathbf {q} _ {1} & \\mathbf {F} \\\\ - \\mathbf {q} _ {2} & t \\mathbf {F} \\end{array} \\right] \\left[ \\begin{array}{l} \\alpha \\\\ \\mathbf {w} \\end{array} \\right] \\right\\| _ {2} ^ {2} \\tag {21} \\\\ \\stackrel {{\\mathrm {d e f}}} {{=}} \\left\\| \\mathbf {y} - \\mathbf {D} (t) \\left[ \\begin{array}{c} \\alpha \\\\ \\mathbf {w} \\end{array} \\right] \\right\\| _ {2} ^ {2}. \\\\ \\end{array}", + "image_path": "6d0454b3aa46bcb5fdcdcc2c935e07b830190bb2dcd45f1cf3169e6c62b246a4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "text", + "content": " is unknown, the alpha matte cannot be derived in a closed-form manner. However, as shown in Eq. (22), " + }, + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "text", + "content": " may be eliminated from Eq. (21), leaving a polynomial function only in " + }, + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 628, + 296, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 77, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 316, + 77, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 316, + 77, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 77, + 553, + 243 + ], + "type": "image", + "image_path": "788694c48578e692902a03bb07909d3b4859fdd297e3a37d7c3bff0212762cd7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "type": "text", + "content": "Figure 4. An illustrative interpretation of pCSM-DS optimization. For a fixed " + }, + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "type": "text", + "content": ", the objective function has a least-squares solution (blue points). Projecting the least-squares solutions to the " + }, + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "type": "inline_equation", + "content": "t - f(\\alpha, \\mathbf{w}, t)" + }, + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "type": "text", + "content": " plane forms " + }, + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "type": "inline_equation", + "content": "g(t)" + }, + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "type": "text", + "content": ", where the critical points (red and yellow points) can be derived as roots of the derivative " + }, + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "type": "inline_equation", + "content": "g'(t)" + }, + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "type": "text", + "content": ". The final solution (red point) can be found by assessing the values of " + }, + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "type": "inline_equation", + "content": "f(\\alpha, \\mathbf{w}, t)" + }, + { + "bbox": [ + 313, + 253, + 555, + 331 + ], + "type": "text", + "content": " at all the critical points." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 350, + 440, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 350, + 440, + 363 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 440, + 363 + ], + "type": "text", + "content": "Proposition 4. Define " + }, + { + "bbox": [ + 314, + 350, + 440, + 363 + ], + "type": "inline_equation", + "content": "g(t)" + }, + { + "bbox": [ + 314, + 350, + 440, + 363 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 386, + 371, + 553, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 371, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 386, + 371, + 553, + 388 + ], + "type": "interline_equation", + "content": "g (t) = \\min _ {\\alpha , \\mathbf {w}} f (\\alpha , \\mathbf {w}, t). \\tag {22}", + "image_path": "36dd6e50db19c18e4da45649a4a6bc976a68c471552193eb3e679b261aae10f6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 397, + 487, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 397, + 487, + 409 + ], + "spans": [ + { + "bbox": [ + 314, + 397, + 487, + 409 + ], + "type": "text", + "content": "Then, " + }, + { + "bbox": [ + 314, + 397, + 487, + 409 + ], + "type": "inline_equation", + "content": "g(t)" + }, + { + "bbox": [ + 314, + 397, + 487, + 409 + ], + "type": "text", + "content": " is a ratio of polynomials in " + }, + { + "bbox": [ + 314, + 397, + 487, + 409 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 314, + 397, + 487, + 409 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 406, + 416, + 553, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 416, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 406, + 416, + 553, + 443 + ], + "type": "interline_equation", + "content": "g (t) = \\frac {n (t)}{d (t)}, \\tag {23}", + "image_path": "9d2733892897db6178dcd8e4a32a9aad5580d413e083ab263de9fb91e57f507b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 450, + 539, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 539, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 539, + 463 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 450, + 539, + 463 + ], + "type": "inline_equation", + "content": "n(t)" + }, + { + "bbox": [ + 313, + 450, + 539, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 539, + 463 + ], + "type": "inline_equation", + "content": "d(t)" + }, + { + "bbox": [ + 313, + 450, + 539, + 463 + ], + "type": "text", + "content": " are fourth-degree polynomials in " + }, + { + "bbox": [ + 313, + 450, + 539, + 463 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 450, + 539, + 463 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 469, + 476, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 469, + 476, + 481 + ], + "spans": [ + { + "bbox": [ + 314, + 469, + 476, + 481 + ], + "type": "text", + "content": "Proof. See the supplementary material." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 544, + 470, + 554, + 480 + ], + "blocks": [ + { + "bbox": [ + 544, + 470, + 554, + 480 + ], + "lines": [ + { + "bbox": [ + 544, + 470, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 544, + 470, + 554, + 480 + ], + "type": "image", + "image_path": "7ed151b7b90a84b84bfc12aec7e1b2adc85f651d193cad847b5556aa60bc5186.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "type": "text", + "content": "With Proposition 4, we can find critical points of " + }, + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "type": "inline_equation", + "content": "g(t)" + }, + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "type": "text", + "content": " as roots of the derivative " + }, + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "type": "inline_equation", + "content": "g'(t)" + }, + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "type": "text", + "content": " and corresponding alpha matte values and coefficient vectors. Noting the range of the alpha matte value, we discard critical points if the corresponding alpha matte values are out of the range [0, 1], forming a set of valid critical points, say " + }, + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "type": "text", + "content": ". Therefore, the optimal variables can be found as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 394, + 582, + 553, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 582, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 394, + 582, + 553, + 600 + ], + "type": "interline_equation", + "content": "t _ {m} = \\underset {t \\in \\mathcal {T}} {\\operatorname {a r g m i n}} g (t), \\tag {24}", + "image_path": "36d56f026775a5cb4bf51b7479943c638dfccaca3e1a9db4d8e87c1e6323a701.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 358, + 604, + 553, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 604, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 358, + 604, + 553, + 622 + ], + "type": "interline_equation", + "content": "\\alpha_ {m}, \\mathbf {w} _ {m} = \\underset {\\alpha , \\mathbf {w}} {\\operatorname {a r g m i n}} f (\\alpha , \\mathbf {w}, t = t _ {m}). \\tag {25}", + "image_path": "66defa626976f2491033f2fce76870f668eb83ec509890e681e78432921f45af.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 631, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 553, + 654 + ], + "type": "text", + "content": "An illustrative interpretation of this optimization are shown in Fig. 4." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": "Note that we need to assess the boundaries of the alpha matte value separately. Given " + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\alpha = \\alpha'" + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": ", the objective function becomes" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 370, + 699, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 699, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 370, + 699, + 553, + 716 + ], + "type": "interline_equation", + "content": "g _ {\\alpha^ {\\prime}} (t) = \\min _ {\\mathbf {w}} f (\\alpha = \\alpha^ {\\prime}, \\mathbf {w}, t). \\tag {26}", + "image_path": "7e9e3a980efa308ba4fb4fb05a11a09facde75ebc89a4f407bf3fd55060b01d4.jpg" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "text", + "content": "1Note that the foreground basis is not necessary to be captured as one of four polarization images because the four polarization images allows us to compute the polarization image intensity at any angle." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "395" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "content": "Proposition 5. " + }, + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "inline_equation", + "content": "g_{\\alpha^{\\prime}}(t)" + }, + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "content": " is a ratio of quadratic functions in " + }, + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "inline_equation", + "content": "t" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 217, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 217, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 217, + 103 + ], + "type": "text", + "content": "Proof. See the supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 285, + 91, + 294, + 101 + ], + "blocks": [ + { + "bbox": [ + 285, + 91, + 294, + 101 + ], + "lines": [ + { + "bbox": [ + 285, + 91, + 294, + 101 + ], + "spans": [ + { + "bbox": [ + 285, + 91, + 294, + 101 + ], + "type": "image", + "image_path": "91d07de54d10a712c2d7b3a586190e064552ba763ae80a1ea8fc8156592c22b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "text", + "content": "Similarly with Eqs. (24) and (25), the optimal " + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "text", + "content": " and coefficient vector for " + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "inline_equation", + "content": "g_0(t)" + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "inline_equation", + "content": "g_{1}(t)" + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "text", + "content": " can be found as " + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "inline_equation", + "content": "(t_0,\\mathbf{w}_0)" + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "inline_equation", + "content": "(t_{1},\\mathbf{w}_{1})" + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "text", + "content": ", respectively. Gathering all candidates of the solution in " + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\left\\{\\left(\\alpha_{m},\\mathbf{w}_{m},t_{m}\\right),(0,\\mathbf{w}_{0},t_{0}),(1,\\mathbf{w}_{1},t_{1})\\right\\}" + }, + { + "bbox": [ + 55, + 110, + 296, + 170 + ], + "type": "text", + "content": ", the final solution can be written as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 178, + 295, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 178, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 108, + 178, + 295, + 198 + ], + "type": "interline_equation", + "content": "\\alpha^ {*}, \\mathbf {w} ^ {*}, t ^ {*} = \\underset {\\alpha , \\mathbf {w}, t \\in \\mathcal {V}} {\\operatorname {a r g m i n}} f (\\alpha , \\mathbf {w}, t). \\tag {27}", + "image_path": "586f470759bb11e917c0c1c652ad4fb6849a839793d99fad4840e9796ae2b4bb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 205, + 295, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 229 + ], + "type": "text", + "content": "This pCSM-DS optimizes a non-linear objective function; therefore, we call this method non-linear " + }, + { + "bbox": [ + 55, + 205, + 295, + 229 + ], + "type": "inline_equation", + "content": "p\\mathrm{CSM - DS}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 236, + 231, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 231, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 231, + 248 + ], + "type": "text", + "content": "3.4. Polarized Diffuse Regularization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "text", + "content": "Our pCSM-DS is built upon a solid theory; however, we observed cases where the linear and non-linear pCSM-DS yield undesirable noisy alpha mattes. This behavior is due to the overfitting of the diffuse-specular polarimetric compositing equation. For instance, in the linear pCSM-DS, the matrix " + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "text", + "content": " in Eq. (19) spans the entire RGB space, and " + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "text", + "content": " is perfectly approximated even if " + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "text", + "content": " contains deviations from the model, e.g., noise. To avoid overfitting, we introduce a polarized diffuse regularization term, motivated by a physical observation that the polarized diffuse intensity " + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "text", + "content": " tends to be small in most scenes. To introduce this regularization into the linear pCSM-DS, we reformulate the linear pCSM-DS with " + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_1" + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_2" + }, + { + "bbox": [ + 55, + 255, + 296, + 410 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 415, + 296, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 415, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 65, + 415, + 296, + 494 + ], + "type": "interline_equation", + "content": "\\underbrace {\\left[ \\begin{array}{l} \\mathbf {s} _ {1} - \\mathbf {q} _ {1} \\\\ \\mathbf {s} _ {2} - \\mathbf {q} _ {2} \\end{array} \\right]} _ {\\mathbf {y}} = \\underbrace {\\left[ \\begin{array}{c c c c c} - \\mathbf {q} _ {1} & \\mathbf {f} _ {d s} & \\mathbf {f} _ {s} & \\mathbf {0} & \\mathbf {0} \\\\ - \\mathbf {q} _ {2} & \\mathbf {0} & \\mathbf {0} & \\mathbf {f} _ {d s} & \\mathbf {f} _ {s} \\end{array} \\right]} _ {\\overline {{\\mathbf {B}}}} \\underbrace {\\left[ \\begin{array}{l} \\alpha \\\\ m _ {d s} \\\\ m _ {s} \\\\ n _ {d s} \\\\ n _ {s} \\end{array} \\right]} _ {\\overline {{\\mathbf {x}}}}. \\tag {28}", + "image_path": "57e3fdad72bc80603cb9b0074f045f30a3dc5ee502d1bb9f3a0f04a55dc92040.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "text", + "content": "Here, the polarized diffuse intensity " + }, + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "text", + "content": " is only included in " + }, + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "inline_equation", + "content": "m_{ds}" + }, + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "inline_equation", + "content": "n_{ds}" + }, + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "text", + "content": ", hence, we regularize " + }, + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "inline_equation", + "content": "m_{ds}" + }, + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "inline_equation", + "content": "n_{ds}" + }, + { + "bbox": [ + 55, + 499, + 296, + 548 + ], + "type": "text", + "content": " to indirectly constrain the polarized diffuse intensity. We use the ridge regression to form the objective function as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 97, + 554, + 294, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 554, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 97, + 554, + 294, + 574 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {x}} ^ {*} = \\underset {\\bar {\\mathbf {x}}} {\\operatorname {a r g m i n}} \\| \\mathbf {y} - \\bar {\\mathbf {B}} \\bar {\\mathbf {x}} \\| _ {2} ^ {2} + \\lambda \\| \\mathbf {E} \\bar {\\mathbf {x}} \\| _ {2} ^ {2}, \\tag {29}", + "image_path": "24b5902a81ad6572dd09617c2bea7932f306c60ac40306f933f728f528539194.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 123, + 575, + 295, + 603 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 575, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 123, + 575, + 295, + 603 + ], + "type": "interline_equation", + "content": "\\mathbf {E} = \\left[ \\begin{array}{l l l l l} 0 & 1 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 1 & 0 \\end{array} \\right], \\tag {30}", + "image_path": "75a62a51c9f43d4fc5b76688ca907fd341d6e6a1234dc33c228b2797bd645cef.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 55, + 609, + 295, + 632 + ], + "type": "text", + "content": " is a regularization parameter. This ridge regression problem can be solved in a closed-form manner." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "text", + "content": "Next, we introduce a regularization in the non-linear pCSM-DS. In Eq. (21), the polarized diffuse intensity " + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "text", + "content": " is only included in " + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "inline_equation", + "content": "w_{ds}" + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "text", + "content": ", hence, we simply regularize " + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "inline_equation", + "content": "w_{ds}" + }, + { + "bbox": [ + 55, + 633, + 296, + 681 + ], + "type": "text", + "content": " and obtain the objective function as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 72, + 687, + 296, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 687, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 72, + 687, + 296, + 717 + ], + "type": "interline_equation", + "content": "\\bar {f} (\\alpha , \\mathbf {w}, t) = \\left\\| \\mathbf {y} - \\mathbf {D} (t) \\left[ \\begin{array}{l} \\alpha \\\\ \\mathbf {w} \\end{array} \\right] \\right\\| _ {2} ^ {2} + \\mu \\| \\mathbf {e} ^ {\\top} \\mathbf {w} \\| _ {2} ^ {2}, \\tag {31}", + "image_path": "9a767677ef9e4f4edce994631bc10e84fd01f1060c9518056126dc49854eb9e9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{e} = [10]^\\top" + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "text", + "content": " is a regularization parameter. Since this objective function is a quadratic function with respect to " + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 71, + 554, + 120 + ], + "type": "text", + "content": ", we can still compute a closed-form expression for" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 386, + 128, + 555, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 128, + 555, + 146 + ], + "spans": [ + { + "bbox": [ + 386, + 128, + 555, + 146 + ], + "type": "interline_equation", + "content": "\\bar {g} (t) = \\min _ {\\alpha , \\mathbf {w}} \\bar {f} (\\alpha , \\mathbf {w}, t), \\tag {32}", + "image_path": "b976f54685921b8aa0e4ac0759bb5571f24b20bc9278999b316fd32cffa97879.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 371, + 149, + 555, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 149, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 371, + 149, + 555, + 167 + ], + "type": "interline_equation", + "content": "\\bar {g} _ {\\alpha^ {\\prime}} (t) = \\min _ {\\mathbf {w}} \\bar {f} (\\alpha = \\alpha^ {\\prime}, \\mathbf {w}, t). \\tag {33}", + "image_path": "3440fcf241c981a46080266cd0a8577e14d14ebc44ae98d9e00e8d653864aad1.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 174, + 554, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 554, + 198 + ], + "type": "text", + "content": "Therefore, the optimal alpha matte can be estimated in the same manner with the unregularized non-linear pCSM-DS." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 205, + 554, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 554, + 228 + ], + "type": "text", + "content": "Proposition 6. " + }, + { + "bbox": [ + 313, + 205, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\overline{g}(t)" + }, + { + "bbox": [ + 313, + 205, + 554, + 228 + ], + "type": "text", + "content": " is a ratio of fourth-degree polynomials in " + }, + { + "bbox": [ + 313, + 205, + 554, + 228 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 205, + 554, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 231, + 476, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 231, + 476, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 231, + 476, + 243 + ], + "type": "text", + "content": "Proof. See the supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 544, + 231, + 554, + 241 + ], + "blocks": [ + { + "bbox": [ + 544, + 231, + 554, + 241 + ], + "lines": [ + { + "bbox": [ + 544, + 231, + 554, + 241 + ], + "spans": [ + { + "bbox": [ + 544, + 231, + 554, + 241 + ], + "type": "image", + "image_path": "0e396fb0ad517b49e7488443af63e9596d35b66d33d547fbbe03d5879cb57dda.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 251, + 554, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 273 + ], + "type": "text", + "content": "Proposition 7. " + }, + { + "bbox": [ + 313, + 251, + 554, + 273 + ], + "type": "inline_equation", + "content": "\\overline{g}_{\\alpha'}(t)" + }, + { + "bbox": [ + 313, + 251, + 554, + 273 + ], + "type": "text", + "content": " is a ratio of fourth-degree polynomials in " + }, + { + "bbox": [ + 313, + 251, + 554, + 273 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 251, + 554, + 273 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 275, + 476, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 275, + 476, + 288 + ], + "spans": [ + { + "bbox": [ + 314, + 275, + 476, + 288 + ], + "type": "text", + "content": "Proof. See the supplementary material." + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 544, + 276, + 554, + 286 + ], + "blocks": [ + { + "bbox": [ + 544, + 276, + 554, + 286 + ], + "lines": [ + { + "bbox": [ + 544, + 276, + 554, + 286 + ], + "spans": [ + { + "bbox": [ + 544, + 276, + 554, + 286 + ], + "type": "image", + "image_path": "7c7ff336401856daeac0694080ab423bf7220aa574aa793c0ad19b2af236b303.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 295, + 554, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 554, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 554, + 356 + ], + "type": "text", + "content": "Note that, similar to pCSM-S, the polarized diffuse regularization makes pCSM-DS applicable to any scene regardless of the color and polarization conditions. We assess the capability of each regularized pCSM-DS using synthetic datasets in Sec. 4.1 to identify the superior method." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 365, + 459, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 459, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 459, + 379 + ], + "type": "text", + "content": "4. Experimental Evaluation" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 391, + 555, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 535 + ], + "type": "text", + "content": "Synthetic Dataset: We evaluated the capability of each matting method in isolation with four types of synthetic datasets, each of which satisfies different color or polarization conditions as shown in Tab. 1. We generated all the parameters and synthesized " + }, + { + "bbox": [ + 313, + 391, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\phi}" + }, + { + "bbox": [ + 313, + 391, + 555, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 391, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{\\phi}" + }, + { + "bbox": [ + 313, + 391, + 555, + 535 + ], + "type": "text", + "content": " following Eqs. (4) and (5). We generated 1000 samples (pixels) for each type of data. We also synthesized noisy datasets, to evaluate the robustness to noise, by adding the Gaussian noise " + }, + { + "bbox": [ + 313, + 391, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_{\\phi} \\sim \\mathcal{N}(0, 0.01)" + }, + { + "bbox": [ + 313, + 391, + 555, + 535 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 391, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\phi} + \\mathbf{n}_{\\phi}, \\forall \\phi" + }, + { + "bbox": [ + 313, + 391, + 555, + 535 + ], + "type": "text", + "content": ". In total, we synthesized eight types of datasets: A, B, C, and D in Tab. 1 with and without noise. The details of how the parameters were generated can be found in the supplementary material." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "type": "text", + "content": "Real-world Dataset: For quantitative evaluations, we used the evaluation dataset used in PolarMatte [11], which contains 12 static scenes with a green polarized background. This dataset provides reference alpha mattes computed by triangulation matting [29] for quantitative evaluation. For further visual evaluations, we captured several videos containing various objects such as furry or transparent objects." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "type": "text", + "content": "Hardware Setup: All videos were captured with the Blackfly® S USB3 color polarization camera [13] and an LCD monitor [14] served as the polarized background. The camera's configuration was set to 30 frames per second, 1/30 second exposure time, and f/1.4. We used a color checker board [8] to linearize the camera response function and calibrate the color of illumination." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "396" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 102, + 284, + 156 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "text", + "content": "Table 1. Color and polarization conditions for four types of synthetic datasets: A, B, C, and D." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 102, + 284, + 156 + ], + "lines": [ + { + "bbox": [ + 69, + 102, + 284, + 156 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 284, + 156 + ], + "type": "table", + "html": "
ABCD
Strong color condition✓
Weak color condition✓✓✓
Polarization condition✓✓✓
", + "image_path": "76c6ebcb9429e1c508bc68e48de91c02615e97df2deb4a8dc426139842afb2fc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 296, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 250 + ], + "type": "text", + "content": "Hyperparameters: Throughout the experiments, we used " + }, + { + "bbox": [ + 55, + 175, + 296, + 250 + ], + "type": "inline_equation", + "content": "\\lambda = 0.01" + }, + { + "bbox": [ + 55, + 175, + 296, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 175, + 296, + 250 + ], + "type": "inline_equation", + "content": "\\mu = 0.01" + }, + { + "bbox": [ + 55, + 175, + 296, + 250 + ], + "type": "text", + "content": " for the polarized diffuse regularization. Since the PolarMatte static scene test data does not provide the color of illumination, we assume the ambient illumination is natural and white, and the specular reflection color is assumed to be " + }, + { + "bbox": [ + 55, + 175, + 296, + 250 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s = (1 / \\sqrt{3})\\mathbf{1}" + }, + { + "bbox": [ + 55, + 175, + 296, + 250 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 254, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 339 + ], + "type": "text", + "content": "Comparative Methods: We compare our method with polarimetric triangulation matting (PTM) [21] using four polarization images, PolarMatte [11], background matting v2 (BGMv2) [18], and the Keylight plug-in [31] in Adobe After Effects [1]. The Keylight results shown in this paper were done by a Keylight expert in 5 minutes for each image to compare it with fully computational matting methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 392 + ], + "type": "text", + "content": "Evaluation Metrics: We employ three commonly used metrics for evaluating the estimated alpha matte: mean absolute error (MAE), mean squared error (MSE), and gradient error (GRAD). We use MAE unless otherwise specified." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 401, + 212, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 212, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 212, + 414 + ], + "type": "text", + "content": "4.1. Results on Synthetic Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 419, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 419, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 295, + 598 + ], + "type": "text", + "content": "Table 2 shows the quantitative comparisons on the noise-free and noisy synthetic datasets. The linear pCSM-DS on the noise-free dataset A and the non-linear pCSM-DS on the noise-free datasets A and B correctly estimate the true alpha matte with zero errors—verifying the validity of our theory. Although unregularized pCSM-DS is sensitive to noise, the polarized diffuse regularization improves robustness to noise. Consequently, the regularized non-linear pCSM-DS shows the best performance on noisy datasets A and B, in which the majority of real-world data is classified. On datasets C and D, since even the non-linear pCSM-DS is not solvable, all the methods produce a similar margin of error. PTM critically fails on dataset D because PTM only relies on polarization cues and the polarization condition is not satisfied." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 600, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 296, + 684 + ], + "type": "text", + "content": "Table 3 shows the processing time of BGMv2 and our prototype implementations of matting methods written in Python, where PTM and pCSM-S for GPU processing are implemented using PyTorch. Since our pCSM-S and PTM are only required to perform a simple calculation, they are more than 9 times faster than BGMv2. Since the (regularized) non-linear pCSM-DS needs to find the zeros of high" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 320, + 113, + 550, + 258 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. Results on the noise-free synthetic data (top) and noisy synthetic data (bottom). Numbers in this table are MAE. See Tab. 1 for the conditions that each data type satisfies." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 113, + 550, + 258 + ], + "lines": [ + { + "bbox": [ + 320, + 113, + 550, + 258 + ], + "spans": [ + { + "bbox": [ + 320, + 113, + 550, + 258 + ], + "type": "table", + "html": "
Noise-free dataABCD
PTM83.6478.4616.07117.60
pCSM-S10.9210.7416.0717.87
Linear pCSM-DS0.00NANANA
Non-linear pCSM-DS0.000.00NANA
Regularized linear pCSM-DS6.0110.7416.0717.87
Regularized non-linear pCSM-DS3.945.7421.4817.87
Noisy dataABCD
PTM85.2579.8421.30118.60
pCSM-S20.3821.4323.6425.39
Linear pCSM-DS42.45NANANA
Non-linear pCSM-DS23.2837.48NANA
Regularized linear pCSM-DS19.1721.7324.1125.60
Regularized non-linear pCSM-DS17.7819.6524.1925.82
", + "image_path": "712d1d555b734ef68bb0f1b86ff41e1d0b0577ef1cc77892bd5c21d91998620f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 294, + 555, + 412 + ], + "blocks": [ + { + "bbox": [ + 347, + 274, + 521, + 285 + ], + "lines": [ + { + "bbox": [ + 347, + 274, + 521, + 285 + ], + "spans": [ + { + "bbox": [ + 347, + 274, + 521, + 285 + ], + "type": "text", + "content": "Table 3. Processing time on " + }, + { + "bbox": [ + 347, + 274, + 521, + 285 + ], + "type": "inline_equation", + "content": "2\\mathrm{\\;K}" + }, + { + "bbox": [ + 347, + 274, + 521, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 347, + 274, + 521, + 285 + ], + "type": "inline_equation", + "content": "4\\mathrm{\\;K}" + }, + { + "bbox": [ + 347, + 274, + 521, + 285 + ], + "type": "text", + "content": " images." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 294, + 555, + 412 + ], + "lines": [ + { + "bbox": [ + 316, + 294, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 555, + 412 + ], + "type": "table", + "html": "
Resolution2K4K
BGMv2 (GPU)20.55 ms39.83 ms
PTM (GPU)1.80 ms6.61 ms
pCSM-S (GPU)2.28 ms8.39 ms
PTM (CPU)0.22 s1.15 s
pCSM-S (CPU)0.36 s1.83 s
Linear pCSM-DS (CPU)70 s259 s
Non-linear pCSM-DS (CPU)71 s276 s
Regularized linear pCSM-DS (CPU)63 s267 s
Regularized non-linear pCSM-DS (CPU)72 s277 s
", + "image_path": "3f292ff0e136aa0b3980d856d9edc27eb829c6d5002dfbf7437af43022105a12.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 470, + 555, + 553 + ], + "blocks": [ + { + "bbox": [ + 313, + 428, + 555, + 463 + ], + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 463 + ], + "type": "text", + "content": "Table 4. Quantitative evaluations of Keylight, PTM [21], BGMv2 [18], PolarMatte [11], and our pCSM-S and pCSM-DS on the static scene test data with a green background [11]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 470, + 555, + 553 + ], + "lines": [ + { + "bbox": [ + 315, + 470, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 315, + 470, + 555, + 553 + ], + "type": "table", + "html": "
MAE ↓MSE ↓GRAD ↓
Keylight27.134.7412.18
BGMv285.2056.6212.51
PTM11.261.223.31
PolarMatte10.350.841.97
pCSM-S7.570.291.92
Regularized non-linear pCSM-DS7.480.261.68
", + "image_path": "e395515516f4c9f46585189ef0cc1acd3989f8b1930cecd90dceee5ea68be6a1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 578, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 651 + ], + "type": "text", + "content": "order polynomials and solve an ordinary least squares problem for every critical point, it shows a less efficient result. We note that while BGMv2 or per-image methods often struggle with dealing with very high-resolution images, all our methods are operated in a per-pixel manner and applicable to any resolution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "The experimental results on the synthetic datasets show that the regularized non-linear pCSM-DS is the best choice for offline applications such as matting dataset labeling, and pCSM-S is the best choice for real-time applications such as video conferencing." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "2The processing time was measured using Intel® Xeon® Platinum 8275CL CPU @ 3.00 GHz and NVIDIA ® Tesla® V100." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "397" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 555, + 420 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 555, + 420 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 555, + 420 + ], + "type": "image", + "image_path": "ff1e952abb082207ac677aad896a560dfb76b97bded4dbc3198e47e43165c511.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 429, + 555, + 463 + ], + "lines": [ + { + "bbox": [ + 55, + 429, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 555, + 463 + ], + "type": "text", + "content": "Figure 5. Visual results of the reference alpha matte computed by triangulation matting (TM) [29], Keylight [31], BGMv2 [18], polarimetric triangulation matting (PTM) [21], PolarMatte [11], and our pCSM-S and Regularized non-linear pCSM-DS on the static scene test data. See more visual results in the supplementary materials." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 474, + 220, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 220, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 220, + 487 + ], + "type": "text", + "content": "4.2. Results on Real-world Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 666 + ], + "type": "text", + "content": "Given the results on the synthetic datasets, we only applied the regularized non-linear pCSM-DS to the real-world datasets as a representative of pCSM-DS. Table 4 shows quantitative comparisons on the static scene test data. These results suggest that our pCSM-S and pCSM-DS significantly improve the quality of matte extraction. While pCSM-DS shows marginal improvements from pCSM-S because the intensity of polarized diffuse reflections tend to be small in the real-world, visual improvements can be seen in Fig. 5. On the headphone scene, pCSM-S mitigates the errors in PTM and PolarMatte due to polarized reflections, and pCSM-DS further improves the alpha matte by considering polarized diffuse reflections. A similar trend can be seen in the zoomed-in figures of the model airplane scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "We show qualitative comparisons on actual videos in the supplementary video, which demonstrate our accurate and temporally consistent alpha mattes and composite results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 473, + 388, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 473, + 388, + 486 + ], + "spans": [ + { + "bbox": [ + 314, + 473, + 388, + 486 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 493, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 685 + ], + "type": "text", + "content": "This paper introduced Polarized Color Screen Matting, a single-shot, per-pixel matting theory that fully leverages color and polarization cues. The theoretical analysis on the proposed diffuse-specular polarimetric compositing equation derived two matting methods, pCSM-S and pCSM-DS. pCSM-S offers inherently efficient matting by simplifying the compositing equation. pCSM-DS revealed the color and polarization conditions to solve the pCSM problem using the full expression of the compositing equation. By introducing polarized diffuse regularization, pCSM-DS gains robustness to noise and applicability to any scene. Experiments on synthetic and real-world datasets show the stable and accurate matting capability of our methods compared with solely color-based or polarization-based methods. Future work may include more general reflectance models for further generalization and applicability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Acknowledgements: We would like to thank Brian Maffitt for his contributions as the Keylight expert." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "398" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "type": "text", + "content": "[1] Adobe Inc. Adobe After Effects. https://www.adobe.com/products/aftereffects.html. Accessed: 2025-02-07. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 295, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 295, + 157 + ], + "type": "text", + "content": "[2] Yagiz Aksoy, Tunç Ozan Aydin, Marc Pollefeys, and Aljosa Smolić. Interactive high-quality green-screen keying via color unmixing. ACM TOG, 36(4):1, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 294, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 294, + 200 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 294, + 200 + ], + "type": "text", + "content": "[3] Yunhao Ba, Alex Gilbert, Franklin Wang, Jina Yang, Rui Chen, Yiqin Wang, Lei Yan, Boxin Shi, and Achuta Kadambi. Deep shape from polarization. In ECCV, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 294, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 294, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 294, + 235 + ], + "type": "text", + "content": "[4] Seung-Hwan Baek, Daniel S Jeon, Xin Tong, and Min H Kim. Simultaneous acquisition of polarimetric svbrdf and normals. ACM TOG, 37(6):268, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 236, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 278 + ], + "type": "text", + "content": "[5] Seung-Hwan Baek, Tizian Zeltner, Hyunjin Ku, Inseung Hwang, Xin Tong, Wenzel Jakob, and Min H Kim. Image-based acquisition and modeling of polarimetric reflectance. ACM TOG, 39(4):139, 2020. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 280, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 294, + 301 + ], + "type": "text", + "content": "[6] Moshe Ben-Ezra. Segmentation with invisible keying signal. In CVPR, 2000. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 303, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 294, + 334 + ], + "type": "text", + "content": "[7] Walter Beyer. Traveling-matte photography and the bluescreen system: a tutorial paper. Journal of the SMPTE, 74 (3):217-239, 1965. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 335, + 294, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 294, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 294, + 368 + ], + "type": "text", + "content": "[8] Calibrite LLC. ColorChecker Classic. https:// calibrite.com/us/product/colorchecker-classic/. Accessed: 2025-02-07. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 369, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 390 + ], + "type": "text", + "content": "[9] Zhaopeng Cui, Jinwei Gu, Boxin Shi, Ping Tan, and Jan Kautz. Polarimetric multi-view stereo. In CVPR, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 392, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 392, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 57, + 392, + 294, + 434 + ], + "type": "text", + "content": "[10] Akshat Dave, Yannick Hold-Geoffroy, Milos Hasan, Kalyan Sunkavalli, and Ashok Veeraraghavan. Snapshot polarimetric diffuse-specular separation. Optics Express, 30(19): 34239-34255, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 436, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 436, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 57, + 436, + 294, + 479 + ], + "type": "text", + "content": "[11] Kenji Enomoto, TJ Rhodes, Brian Price, and Gavin Miller. PolarMatte: Fully computational ground-truth-quality alpha matte extraction for images and video using polarized screen matting. In CVPR, 2024. 2, 3, 4, 6, 7, 8, 9, 10" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 480, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 294, + 513 + ], + "type": "text", + "content": "[12] Mikhail Erofeev, Yury Gitman, Dmitriy Vatolin, Alexey Fedorov, and Jue Wang. Perceptually motivated benchmark for video matting. In BMVC, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 514, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 556 + ], + "type": "text", + "content": "[13] FLIR Systems Inc. BFS-U3-51S5PC-C USB 3.1. https://www.teledynevisionsolutions.com/ products/blackfly-s-usb3/?model=BFS-U3-51S5PC-C. Accessed: 2025-02-07. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 601 + ], + "type": "text", + "content": "[14] Hisense Group Co., Ltd. 55\" 4K UHD Hisense Android Smart TV (2021). https://www.hisense-usa.com/support/55a6g-55-4k-uhd-hisense android-smart-TV-2021. Accessed: 2025-02-07. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 602, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 646 + ], + "type": "text", + "content": "[15] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[16] Soma Kajiyama, Taihe Piao, Ryo Kawahara, and Takahiro Okabe. Separating partially-polarized diffuse and specular reflection components under unpolarized light sources. In WACV, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[17] Chenhao Li, Taishi Ono, Takeshi Uemori, Hajime Mihara, Alexander Gatto, Hajime Nagahara, and Yuseke Moriuchi." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 658 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "NeISF: Neural incident stokes field for geometry and material estimation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[18] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In CVPR, 2021. 1, 2, 7, 8, 9, 10" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "text", + "content": "[19] Hao Lu, Yutong Dai, Chunhua Shen, and Songcen Xu. Indices matter: Learning to index for deep image matting. In ICCV, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "text", + "content": "[20] Youwei Lyu, Lingran Zhao, Si Li, and Boxin Shi. Shape from polarization with distant lighting estimation. IEEE TPAMI, 45(11):13991-14004, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 209, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 240 + ], + "type": "text", + "content": "[21] Morgan McGuire and Wojciech Matusik. Real-time triangulation matting using passive polarization. In SIGGRAPH Sketches. 2006. 7, 8, 9, 10" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 243, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 275 + ], + "type": "text", + "content": "[22] Morgan McGuire, Wojciech Matusik, and William Yerazunis. Practical, Real-time Studio Matting using Dual Imagers. In Symposium on Rendering, 2006. 1, 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 277, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 308 + ], + "type": "text", + "content": "[23] GyuTae Park, SungJoon Son, JaeYoung Yoo, SeHo Kim, and Nojun Kwak. Matteformer: Transformer-based image matting via prior-tokens. In CVPR, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 310, + 553, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 332 + ], + "type": "text", + "content": "[24] Vlahos Petro. Composite photography utilizing sodium vapor illumination, 1963. US Patent 3,095,304. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 334, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 334, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 334, + 553, + 354 + ], + "type": "text", + "content": "[25] Thomas Porter and Tom Duff. Compositing digital images. SIGGRAPH, 1984. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 356, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 356, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 356, + 553, + 387 + ], + "type": "text", + "content": "[26] Steven A Shafer. Using color to separate reflection components. Color Research & Application, 10(4):210-218, 1985. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 389, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 433 + ], + "type": "text", + "content": "[27] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE TPAMI, 41(2):271-284, 2019. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 434, + 553, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 479 + ], + "type": "text", + "content": "[28] Dmitriy Smirnov, Chloe LeGendre, Xueming Yu, and Paul Debevec. Magenta green screen: Spectrally multiplexed alpha matting with deep colorization. In The Digital Production Symposium (DigiPro), 2023. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 480, + 553, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 480, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 316, + 480, + 553, + 501 + ], + "type": "text", + "content": "[29] Alvy Ray Smith and James F Blinn. Blue screen matting. In SIGGRAPH, 1996. 1, 2, 6, 8, 9, 10" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 502, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 545 + ], + "type": "text", + "content": "[30] George Gabriel Stokes. On the composition and resolution of streams of polarized light from different sources. Transactions of the Cambridge Philosophical Society, 9:399, 1851. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 547, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 547, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 547, + 553, + 591 + ], + "type": "text", + "content": "[31] The Foundry Visionmongers Ltd. Keylight. https://learn.foundry.com/nuke/13.0/content/reference_guide/keyer_nodes/ keylight.html. Accessed: 2025-02-07. 1, 2, 7, 8, 9, 10" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 593, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 613 + ], + "type": "text", + "content": "[32] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In CVPR, 2017. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "type": "text", + "content": "[33] Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. Attention-guided temporally coherent video object matting. In ACM MM, 2021. 2" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "399" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Poly-Autoregressive Prediction for Modeling Interactions/43d051a6-6821-4e3b-bd9a-a4c244857daa_content_list.json b/2025/Poly-Autoregressive Prediction for Modeling Interactions/43d051a6-6821-4e3b-bd9a-a4c244857daa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e157e112da17cebdfcace5b638b08af7575243f9 --- /dev/null +++ b/2025/Poly-Autoregressive Prediction for Modeling Interactions/43d051a6-6821-4e3b-bd9a-a4c244857daa_content_list.json @@ -0,0 +1,1628 @@ +[ + { + "type": "text", + "text": "Poly-Autoregressive Prediction for Modeling Interactions", + "text_level": 1, + "bbox": [ + 209, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neerja Thakkar $^{1}$ Tara Sadjadpour $^{1}$ Jathushan Rajasegeran $^{1}$ Shiry Ginosar $^{2,3}$ Jitendra Malik $^{1}$ $^{1}$ UC Berkeley, $^{2}$ Toyota Technical Institute at Chicago, $^{3}$ Google DeepMind", + "bbox": [ + 106, + 179, + 888, + 224 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/709f2fc30bfc60bc3d43308b1860f77821a18b7732f0f7e89851a55cdf7275f8.jpg", + "image_caption": [ + "(a) Autoregression (AR)" + ], + "image_footnote": [], + "bbox": [ + 98, + 277, + 460, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76fb006702cb31067a0dcf2db0afe7ff8004ce2ee609bee1bce0694edc3cb479.jpg", + "image_caption": [ + "(b) Poly-autoregression (PAR)", + "Figure 1. Inference for (a) autoregressive (AR) models and (b) our proposed poly-autoregressive (PAR) model. Solid indicates ground-truth tokens which represent a tracked data modality such as action or 6DOF pose; striped represents predicted output tokens. Color denotes agent identity. Compared to AR models, the PAR model takes other agents' tokens as inputs when making a prediction for the next timestep." + ], + "image_footnote": [], + "bbox": [ + 491, + 273, + 898, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 483, + 326, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a simple framework for predicting the behavior of an agent in multi-agent settings. In contrast to autoregressive (AR) tasks, such as language processing, our focus is on scenarios with multiple agents whose interactions are shaped by physical constraints and internal motivations. To this end, we propose Poly-Autoregressive (PAR) modeling, which forecasts an ego agent's future behavior by reasoning about the ego agent's state history and the past and current states of other interacting agents. At its core, PAR represents the behavior of all agents as a sequence of tokens, each representing an agent's state at a specific timestep. With minimal data pre-processing changes, we show that PAR can be applied to three different problems: human action forecasting in social situations, trajectory prediction for autonomous vehicles, and object pose forecasting during hand-object interaction. Using a small proof-of-concept transformer backbone, PAR outperforms AR across these three scenarios.", + "bbox": [ + 88, + 513, + 485, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 814, + 220, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) have been very successful with natural language processing (NLP) tasks, which require accurate reasoning over relationships words have within a body of text. A key component of LLMs is autoregressive", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(AR) modeling, where each word token is predicted based on a sequence of preceding word tokens. Building on the success of AR modeling in the NLP community, this work focuses on modeling the dynamic relationships between agents and entities in everyday interactions that occur in the physical world, such as social interactions, driving, and hand-object interactions.", + "bbox": [ + 511, + 484, + 906, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike language which is structured through grammar and semantics, interactions in the physical world are dictated by both the laws of physics (e.g. how a hand grasps an object) and the internal state of each agent (e.g. the trajectory an agent chooses to move a grasped object in), a latent variable that we know nothing about. Furthermore, as opposed to text where each word follows another unidimensionally, the states of multiple agents are changing simultaneously. For example, in social situations, the history of a single person's past states does not alone determine the dynamics of their future states; we also need to consider the states of other agents. We argue, therefore, that AR modeling alone is insufficient.", + "bbox": [ + 509, + 594, + 908, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce poly-autoregressive (PAR) modeling, a simple unifying approach to model the influence of other agents and entities on one's behavior. We model behavior as a temporal sequence of states and predict an ego agent's future behavior conditioned on the history of behavior of the ego agent as well as the rest of the agents. By considering other agents' behaviors, we demonstrate that our approach significantly improves upon the ill-posed problem", + "bbox": [ + 509, + 779, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12402", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of single-agent prediction in interactive settings.", + "bbox": [ + 89, + 90, + 410, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The PAR framework uses a transformer-based model for next-token prediction. Transformers have shown great success in language modeling and naturally lend themselves to predicting behavior over time. In an interaction scenario of $N$ agents, our model predicts the future behavior of an ego $(N^{\\mathrm{th}})$ agent conditioned on its past behavior and the behavior of the other $N - 1$ non-ego agents (see Figure 1b). Each prediction task may model a different behavior modality of interest, e.g. actions for social action prediction, or 6DoF pose for object forecasting in hand-object interactions. We apply PAR to three different case studies of common real-world interactions:", + "bbox": [ + 89, + 108, + 485, + 287 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Social action prediction. We test our method on the AVA benchmark [19] for action forecasting. By incorporating both the ego and another agent using PAR, we get a overall $+1.9$ absolute mAP gain over AR, which only models the ego agent to predict its future behavior, and an absolute $+3.5$ mAP gain on 2-person interaction classes.", + "2. Trajectory prediction for autonomous vehicles. When forecasting future $xy$ locations of an ego vehicle, incorporating the locations of neighboring vehicles with PAR outperforms AR, which only uses the ego vehicle's preceding trajectory as input. Specifically, on the nuScenes dataset [5], PAR outperforms AR with a relative improvement of $6.3\\%$ ADE and $6.4\\%$ FDE.", + "3. 6DoF object pose forecasting during hand-object interaction. We use the DexYCB dataset [8], where we treat the object as the ego agent and the hand as the interacting agent. While PAR integrates that hand's 3D location and object pose history, AR only uses the object's pose history to predict the object's 6DoF pose. PAR outperforms AR with relative improvements of $8.9\\%$ and $41.0\\%$ for the rotation and translation predictions, respectively." + ], + "bbox": [ + 89, + 292, + 483, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In all these settings, we find that incorporating the behavior of other agents in the scene improves predictions of the ego agent's behavior. All of these problems are modeled via the same simple PAR framework and implemented using the same proof-of-concept 4 million parameter transformer without any modifications to the base framework or architecture, only to data pre-processing and choice of tokenization. We also provide an example of a simple way to build on our architecture through a location positional encoding, Sec. 5.", + "bbox": [ + 89, + 611, + 485, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primary contribution of this work is a versatile framework that can be applied to a diverse range of settings, without modifications aside from domain-specific data processing. Our results suggest that PAR provides a simple formulation that, with a more complex transformer backbone and larger datasets, could enhance prediction of diverse multi-agent interactions across various problem domains. To facilitate further exploration and development, we have released our code, which contains the building blocks to use PAR for modeling other types of multi-agent interactions.", + "bbox": [ + 89, + 750, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 656, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoregressive models. Autoregressive modeling has a rich history in information theory and deep learning, tracing back to Shannon's paper on language prediction [45] and Attneave's study on visual perception [2]. These foundational works laid the groundwork for modern applications in deep learning, including [25], which revisits neural autoregressive models and [18, 50], which explore continuous-valued modeling. [53] developed PixelRNN and PixelCNN to autoregressively generate images one pixel at a time using RNN and CNN architectures, respectively.", + "bbox": [ + 511, + 114, + 906, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The development of the transformer model [54] spurred progress in computer vision with the image transformer [36] and the vision transformer [14], and autoregressive models [9, 41] and more notably in NLP, where the GPT family of models [4, 37, 38] has demonstrated the power of large-scale unsupervised autoregressive pre-training. Recent research has focused on multimodal learning, exemplified by the Flamingo [1] or LlaVa [28] models, which combine vision and language processing capabilities, illustrating the versatility of autoregressive models across various domains in artificial intelligence. While these approaches operate on image patches and word tokens, we operate on symbolic representations extracted from in-the-wild videos showing natural interactions. A recent approach [39] frames humanoid locomotion as an autoregressive next-token prediction task that operates on two types of continuous tokens: observations and actions. This approach projects continuous tokens to the hidden dimension and uses a shifted loss, similar to the next-timestep prediction proposed in our framework.", + "bbox": [ + 511, + 276, + 908, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-agent regressive models. Several prior works addressed modeling specific multi-agent problems via regressive models as one-off case studies. We introduce the PAR framework to unify these efforts into a single cohesive framework. Many behavior prediction works focus on two agents engaging in social interaction, whether it be dyadic communication [32-34] or social dance [30, 46]. These studies primarily tackle the challenge of predicting the state of an interacting partner (Person B) based on the input from Person A's state, sometimes extending predictions into the future [20, 30]. While earlier works used architectures such as variational RNNs [3], recent works have predominantly adopted transformer architectures for social interaction modeling [10, 20, 32, 33, 46], with some works exploring diffusion [27], or diffusion with attention [17]. Our PAR framework focuses on transformer models.", + "bbox": [ + 511, + 573, + 908, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Works encompassed by the PAR framework extend beyond human social interaction. Many multi-agent human or car trajectory prediction approaches use autoregressive prediction. For instance, MotionLM [44] utilizes a transformer decoder that processes multi-agent tokens, incorporating a", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12403", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learned agent ID embedding. This methodology informs our approach across all our case studies. Critically, in contrast to all prior multi-agent regressive works that designed solutions to address specific applications, we demonstrate that we can unify a diverse set of multi-agent regressive problems under a single PAR framework. See Appendix Sec. 9 for case-study-specific works.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Poly-Autoregressive Modeling", + "text_level": 1, + "bbox": [ + 89, + 214, + 366, + 232 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to model the behavior of an agent or entity while taking into account any other agents it interacts with, if any. To evaluate the performance of our model in capturing interaction dynamics, we predict the agent's future behavior and compare it against ground-truth data.", + "bbox": [ + 89, + 239, + 483, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define the following task: In an interaction comprised of $N$ agents, given the observed past states of the $N - 1$ interacting agents, and the observed or previously-predicted past states of the $N^{th}$ ego agent, predict the future states of the $N^{th}$ ego agent.", + "bbox": [ + 89, + 316, + 483, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define a transformer-based poly-autoregressive (PAR) predictor, $\\mathcal{P}$ , that learns to model temporally long-range interactions in the input sequence. The inputs to the predictor are the past states of the $N$ interacting agents, and its output is the predicted future state of the $N^{\\mathrm{th}}$ ego agent.", + "bbox": [ + 89, + 391, + 483, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 89, + 474, + 272, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{S} = \\{\\mathbf{s}_i\\}_{i=1}^T$ be a temporal sequence of agent states, $\\mathbf{s}_i$ . We use $\\mathbf{S}^N$ and $\\mathbf{S}^{1:N-1}$ to denote the temporal sequences of states of the $N_{th}$ agent and of the other $N-1$ agents, respectively. For each timestep $t \\in [t_{\\pi}, T]$ , where $t_{\\pi} \\in [1, T]$ is the time we start predicting, we take as input all other $N-1$ agents' past observed state sequences $\\mathbf{S}_{1:t-1}^{1:N-1}$ along with the $N_{th}$ agent's past observed states up to $t_{\\pi}$ , $\\mathbf{S}_{1:t_{\\pi}}^N$ , and any of its previously predicted past states $\\hat{\\mathbf{S}}_{t_{\\pi}+1:t-1}^N$ , if available (see Fig. 1). Our predictor, $\\mathcal{P}$ , then poly-autoregressively predicts the $N_{th}$ agent's future states one time-step at a time:", + "bbox": [ + 89, + 494, + 483, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {s}} _ {t} ^ {N} = \\mathcal {P} \\left(\\mathbf {S} _ {1: t - 1} ^ {1: N - 1}, \\mathbf {S} _ {1: t _ {\\pi}} ^ {N}, \\hat {\\mathbf {S}} _ {t _ {\\pi} + 1: t - 1} ^ {N}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 661, + 483, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathcal{P}$ learns to model the distribution over the next timestep of the $N_{th}$ agent's states, given all other agents' states:", + "bbox": [ + 89, + 685, + 483, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\hat {\\mathbf {s}} _ {t} ^ {N} \\mid \\mathbf {S} _ {1: t - 1} ^ {1: N - 1}, \\mathbf {S} _ {1: t - 1} ^ {N}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 720, + 482, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While we provide the observed ground truth states of other agents at inference, during training, we jointly maximize the likelihood of all $N$ agents by computing losses on their future state predictions.", + "bbox": [ + 89, + 747, + 483, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We train the predictor by maximizing the likelihood of the target state $y$ at time $t$ :", + "bbox": [ + 89, + 806, + 483, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {P}} = E _ {y \\sim p (y)} [ - \\log (p (\\mathbf {s} _ {t} ^ {N}) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 845, + 385, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the target state $y$ at $t$ is computed from the $N_{th}$ agent ground truth future state.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. The Poly-Autoregressive Framework", + "text_level": 1, + "bbox": [ + 511, + 90, + 831, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We address the problem of forecasting the future states of an agent (from time $t$ to $T$ ) in a data-driven way, given a temporal sequence of past states (from time 1 to $t - 1$ ). We assume that our agent has some feature, or a set of features, of interest in a video (e.g., 3D pose) that we can tokenize. We predict the future states of the agent in terms of this tokenized feature (or set of), where we use one token (or set of tokens) per time step. The predicted tokens can be discrete (i.e., an index into a feature codebook) or continuous (i.e., a vector of one or more continuous values). The loss $\\ell$ will depend on the problem's specifics and the type of token used. To train the model to predict the future, we rely on all the interaction dynamics of length $T$ in our training dataset as ground truth examples.", + "bbox": [ + 511, + 112, + 906, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a baseline, we consider the single-agent autoregressive (AR) paradigm, where a transformer is trained to perform next-token-prediction with teacher forcing. AR uses greedy sampling to generate sequences at inference time, predicting one next token at a time (Fig. 1(a)).", + "bbox": [ + 511, + 324, + 906, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, our multi-agent poly-autoregressive (PAR) framework considers the other $N - 1$ agents in the scene when predicting the future state of the $N$ th agent. In this setup, we utilize the features of interest of all $N$ agents, yielding $N$ tokens at each timestep for a total of $N * T$ tokens. In practice, we operate on a flattened sequence of $N * T$ tokens. Instead of using the AR training procedure in this multi-agent case (as in Fig. 3a), we jointly model the $N$ agents at each timestep by introducing the following features to our PAR framework.", + "bbox": [ + 511, + 398, + 905, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next-timestep prediction. A standard AR model predicts the next token. Given the flattened sequence of $N * T$ tokens our model operates on, next token prediction would take as input an agent $k$ at timestep $t$ and predict agent $k + 1$ 's state at the same timestep $t$ (as in Fig. 3a). However, our goal is to predict the input agent $k$ 's future state at time $t + 1$ . Therefore, we perform same-agent next-timestep prediction rather than next-token prediction (see Fig. 3b for an illustration of same-agent next-timestep at training).", + "bbox": [ + 511, + 556, + 906, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learned agent identity embedding. When giving a model information corresponding to multiple agents, the model can benefit from knowing which token corresponds to which agent. We give the model this information with a learned agent ID embedding.", + "bbox": [ + 511, + 699, + 905, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Joint training. We train the model to jointly predict the future of all agents by computing a loss on the predicted tokens of all agents (Fig. 3b). Please refer to Section 3.1 for our inference paradigm.", + "bbox": [ + 511, + 780, + 905, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Task-Specific Considerations", + "text_level": 1, + "bbox": [ + 511, + 848, + 769, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our simple PAR approach unifies diverse problems under a single framework and architecture without any modifications.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12404", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/82a17ddfb9150be20f986188f09277d60cac11c0377ae2ffa992a7c4207655f2.jpg", + "image_caption": [ + "Figure 2. The PAR Framework. We begin by collecting a video dataset, such as AVA (top) or DexYCB (bottom). Then, using dataset labels or computer vision techniques, a trajectory of a given modality for our prediction task is extracted for each agent, such as action class labels (top) or object pose and 3D hand translation (bottom). Data is then tokenized, either through discretization or directly using continuous values, with our framework supporting both formats. Based on the tokenization and prediction task, we choose the appropriate loss function for PAR training. After training with PAR, predicted tokens can be decoded back to data space and evaluated with relevant metrics." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 897, + 310 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b8c56e25ce69fce0b0be688f7fe04386f70a084587cfb72faf67a20e577e2e7.jpg", + "image_caption": [ + "(a) AR: multi-agent, next-token training." + ], + "image_footnote": [], + "bbox": [ + 135, + 405, + 450, + 512 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec2468950fe8402fa30610d7862e0da7ed3da055861f2cfd5821056d079c9eef.jpg", + "image_caption": [ + "(b) PAR: same-agent, next-timestep training.", + "Figure 3. Training with teacher forcing for (a) multi-agent next-token prediction in autoregressive models and (b) multi-agent poly-autoregressive models. Solid vs striped indicates a ground-truth vs predicted token, respectively. Color denotes agent identity. The AR model is trained for next-token prediction, while the PAR model is trained to predict the next timestep of the same agent. Three agents are shown for ease of visualization, but the PAR model supports an arbitrary number of agents." + ], + "image_footnote": [], + "bbox": [ + 135, + 527, + 450, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to formulate a problem as interaction-conditioned prediction, users must consider several task-specific details.", + "bbox": [ + 89, + 773, + 483, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 gives an overview of how the PAR framework disentangles multi-agent learning from problem-specific modeling.", + "bbox": [ + 89, + 804, + 483, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data. The input data source in our example tasks is always a collection of videos. From these videos, we extract various modalities relevant to the task at hand. These modalities can range from high-level features, such as action class labels, to", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "low-level ones, such as 3D pose (Fig. 2 first two columns). We assume that each agent in the dataset is detected at each frame and is associated with an agent ID.", + "bbox": [ + 511, + 407, + 906, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tokenization. Our framework supports both discrete, quantized tokens and continuous vector tokens. The choice between discrete and continuous depends on the nature of the task. In the case of discrete tokens, we use a standard embedding layer to project to the hidden dimension. For continuous tokens, we train a projection layer to project the token into the hidden dimension of the transformer. For instance, if our continuous token is a 3D vector with an $(x,y,z)$ 3D location coordinate and our hidden dimension is 128, our projection layer will project from 3 to 128 dimensions. We also train an un-projection layer that reverts the hidden dimension to the original token dimension.", + "bbox": [ + 511, + 460, + 908, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss. The type of token and task-specific considerations dictate the loss function $\\ell$ applied during model training. For discrete tokens, a classification loss is appropriate. For continuous tokens, we use a regression loss on the original token dimension.", + "bbox": [ + 511, + 650, + 906, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baselines. We compare to the following baselines, where applicable on a case-by-case basis:", + "bbox": [ + 511, + 733, + 906, + 763 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Random token: pick random tokens from the best available token space and use as the prediction.", + "- Random trajectory: pick at random a trajectory from the training dataset to use as the prediction.", + "- NN: Given an input agent $A$ 's trajectory history, find the closest trajectory to it in the training set, belonging to $A^T$ . Use $A^T$ 's future as the predicted future.", + "- Multiagent NN: In a dataset with two interacting partners $A$ and $B$ , where $B$ is the ego agent, given an input agent" + ], + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12405", + "bbox": [ + 478, + 916, + 517, + 926 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$A$ 's trajectory history, find the closest trajectory to it in the training set, belonging to $A^T$ . Use $A^T$ 's interaction partner's $B^T$ 's future as the prediction.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Mirror: In a dataset with two interacting partners $A$ and $B$ , use the ground truth future of agent $B$ as the predicted future for agent $A$ .", + "bbox": [ + 89, + 138, + 482, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Framework Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 191, + 401, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We keep the following implementation details constant for all case studies (see also Sec. 10).", + "bbox": [ + 89, + 214, + 482, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learned agent ID embedding. Our learned agent ID embedding consists of the integer agent ID mapped to a hidden dim-sized vector, and summed to the token embedding.", + "bbox": [ + 89, + 250, + 483, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture. For all case studies, we use the Llama [51] transformer decoder architecture with 8 layers, 8 attention heads, and a hidden and intermediate dimension of 128. The decoder has $\\sim 4.4\\mathrm{M}$ learned parameters, not including learned embedding layers which add a few thousand more parameters. A rotary positional encoding [47] is used in addition to other summed encodings (i.e. agent ID embedding, locational positional encoding in Sec. 5). We train using teacher forcing. The only hyperparameter that changes between case studies is the learning rate.", + "bbox": [ + 89, + 300, + 483, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Case Study 1: Social Action Forecasting", + "text_level": 1, + "bbox": [ + 89, + 463, + 447, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our first case study involves forecasting human actions. Human behaviors are fundamentally social; for instance, individuals frequently walk in groups and alternate between speaking and listening roles when conversing. Certain actions, like hugging or handshaking, are intrinsically multiperson. Therefore, modeling human interactions should help improve action forecasting performance, especially on multiperson actions, which we show in this case study.", + "bbox": [ + 89, + 488, + 483, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 617, + 281, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. The Atomic Visual Actions (AVA) dataset [19] comprises 235 training and 64 15-minute validation videos from movies. Annotations are provided at a 1Hz frequency, detailing bounding boxes and tracks for individuals within the frame, and each person's actions within a 1-second timeframe. Individuals may engage in multiple concurrent actions from a repertoire of 60 distinct action classes (e.g., sitting and talking simultaneously). For our analysis, we select clips featuring a continuous sequence of an agent's actions spanning at least 4s, splitting sequences exceeding 12s. We use the first half of each clip as history to predict the second half. For any ego agent trajectory, we pick a second agent by selecting the person present in the scene for the longest subset of the ego agent's trajectory.", + "bbox": [ + 89, + 638, + 482, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task-specific considerations. Each agent's token $\\mathcal{A}$ represents an 60-dimensional vector that corresponds to the actions performed at a specific timestep. Each element denotes", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/918d894c8eab9de6ce9581537fa3c4e1da73166755f84a2c911650fb8a7a4e47.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTimestep predAg-ID embdmAP ↑
1-agent ARN/AN/A40.7
2-agent ARXX38.0
2-agent PAR*X✓40.2
2-agent PAR*✓X40.0
2-agent PAR✓✓42.6
", + "bbox": [ + 517, + 89, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d4aa1b5f0fd67690160c6ee5ed3045f8b9bbfa25787737491f44344e0220aeb4.jpg", + "table_caption": [ + "Table 1. PAR action forecasting performance on AVA We evaluate 1 and 2-agent AR methods, two 2-agent PAR ablations (rows 3 and 4, $\\mathrm{PAR^{*}}$ ), and our PAR method. Without next-timestep prediction (see Fig. 3) or a learned agent ID embedding, our model struggles with multi-agent reasoning, performing worse than the AR baseline. With both components, the 2-agent PAR model achieves a $+1.9$ mAP gain over the AR method (see Fig. 11 and Fig. 5 for class breakdown)." + ], + "table_footnote": [], + "table_body": "
BaselineAgentsmAP ↑
Random Token13.46
Random Training Traj13.44
Nearest Neighbor113.17
Multiagent NN25.10
Mirror27.97
", + "bbox": [ + 568, + 321, + 849, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. AVA baselines While the nearest neighbor baseline performs best among baselines, it is still significantly worse than the AR model.", + "bbox": [ + 511, + 440, + 906, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the probability of a particular action class being enacted; ground-truth inputs are a binary vector. We implement an embedding layer that projects these tokens into the transformer's hidden dimension, as well as an un-projection layer that reverts them back to the original 60D token space for the purposes of loss calculation and output generation. We do not explicitly require the outputs to be values between 0 and 1. We use a MSE regression loss on the 60D action tokens: $\\mathcal{L} = \\frac{1}{n}\\sum_{i=1}^{n}(\\mathcal{A}_i - \\hat{\\mathcal{A}}_i)^2$ . Our evaluation metric is the mean average precision (mAP) on the 60 AVA classes.", + "bbox": [ + 511, + 491, + 905, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement all baselines described in 3.3, where Random Token corresponds to a random 60D vector sampled from 0 to 1. NN and Multiagent NN use Hamming distance as the distance metric.", + "bbox": [ + 511, + 643, + 906, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 511, + 712, + 607, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We report the performance of a single-agent AR model as a baseline, in the first line of Table 1. The AR model is significantly better than our baselines (see Table 2), the strongest baseline being the single-agent NN. We compare these baselines to our 2-agent PAR model (last line) and various ablations where we remove the agent ID embedding and perform next-token rather than same-agent next-timestep prediction. The second line of the table corresponds to multi-agent next-token prediction (Fig. 3a). We see that this approach confuses the model, and the performance is significantly worse than just training on and considering a single agent. However,", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12406", + "bbox": [ + 480, + 917, + 517, + 928 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f33dc690baa1e6313f48ddb1725c5177b489af2d575f1db9777a17af674f3a8e.jpg", + "image_caption": [ + "Figure 4. Action forecasting example. The distribution over ground truth actions are in white, and our predictions in red. A 6s action history (1Hz) is input, and 6s of future actions predicted. In the scene, the man and woman alternate between talking and listening. Initially, the man is talking. The AR model predicts he will continue talking, while the 2-agent PAR model recognizes the woman is talking and predicts more accurate turn-taking behavior." + ], + "image_footnote": [], + "bbox": [ + 140, + 92, + 864, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f4fb74698e8f8f190ab3c6ab86b66b3b909184799962d45dbcdabfa4e945f10.jpg", + "image_caption": [ + "Figure 5. Per-class mAP for AVA 2-person actions. We see performance improvement on almost all 2-person AVA action classes ((P) stands for \"a person\"). Some absolute mAP gains are particularly significant: listen to $+7.0$ , kiss $+8.3$ , fight/hit $+5.7$ , talk to $+4.4$ , hug $+5.7$ , and hand shake $+4.0$ ." + ], + "image_footnote": [], + "bbox": [ + 98, + 362, + 473, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as we add various components of our PAR approach, the performance improves, and with both the next timestep prediction and agent ID embedding, we get a $+1.9$ mAP gain. When only considering 2-person action classes (enumerated in Fig. 5), our mAP is 36.3 on the single agent PAR model and 39.8 on the 2-agent PAR model, a $+3.5$ mAP gain.", + "bbox": [ + 89, + 608, + 483, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 4 we see an example of action forecasting. In the input history, the man talks and the woman listens. In the future, the woman talks, and the man listens. Our 2-agent PAR model (bottom row) better understands that talking and listening actions are complementary actions, while the AR model doesn't learn this correlation. We see quantitative evidence of this in Fig. 5, with per-class mAPs for our AR vs 2-agent PAR model for 2-person action classes. Here, talk to gets a $+4.4$ mAP gain and listen to gets a $+7.0$ mAP gain when we train a multi-agent model. We see a significant boost on many other interaction-related action classes—for instance, kiss a person $+8.3$ and fight/hit a person $+5.7$ mAP—and on single-person actions, see Fig. 11.", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Case Study 2: Multiagent Car Trajectory Prediction", + "text_level": 1, + "bbox": [ + 511, + 358, + 906, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our second case study focuses on predicting car trajectories. Trajectory prediction requires a vehicle to be aware of other cars on the road to avoid collisions and promote cooperative behavior. This study demonstrates how our framework enables the joint modeling of multiple vehicles' movements.", + "bbox": [ + 511, + 402, + 908, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 486, + 705, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We use nuScenes [5], inputting 2 seconds of positions to forecast vehicle positions 6 seconds ahead. Specifically, our objective is to predict the $xy$ coordinates of each agent, exclusively considering vehicles as agents. We use the trajdata interface [23] to load and visualize the data.", + "bbox": [ + 511, + 508, + 906, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task-specific considerations. Instead of discretizing the $xy$ position space, we discretize the motion, resulting in discrete velocity or acceleration tokens. These integer tokens are projected to the transformer hidden dimension using the Llama token embedding layer. Inputting only these tokens results in our PAR model knowing what speed the other agents are going at, but not where they are. It is important the model has this awareness (it should know if two agents are going to collide), so our model needs to reason over this second modality of location. We implement this by passing locations relative to the agent we are predicting into a sincosine positional embedding (see details in Sec. 12.1), which we denote a location positional encoding (LPE). The LPE is summed to our token embeddings.", + "bbox": [ + 511, + 583, + 908, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use a cross-entropy classification loss on our discrete tokens: $\\mathcal{L} = E_{y\\sim p(y)}[-\\log (p(\\mathbf{s}_{t_\\pi}^T))]$ . We use the standard average displacement error (ADE) and final displacement error (FDE) to evaluate our predicted trajectories. For our baselines (Sec. 3.3), we use the closest agent at the current timestep for Multiagent NN and Mirror. For NN and Multiagent NN we use MSE as the distance metric.", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17", + "bbox": [ + 478, + 916, + 519, + 928 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/673e05d349b3b1c65bda8705ea689558450412e26a74c5e1045ed7631929ba5d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Token typeLPEMethodADE ↓FDE ↓
VelocityX1-agent AR1.503.64
VelocityX3-agent PAR1.453.51
AccelerationX1-agent AR1.443.57
AccelerationX3-agent PAR1.403.44
Accleration✓3-agent PAR1.353.34
", + "bbox": [ + 106, + 88, + 467, + 195 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0def927a57bd7e09893faf6f0d0db1bfcfb50997c4e09bbe28b32091a5d89d50.jpg", + "table_caption": [ + "Table 3. Car trajectory prediction performance. Using acceleration tokens and 3-agent PAR results in a stronger performance over velocity tokens and single-agent AR. Adding location via a positional encoding (LPE) further improves results." + ], + "table_footnote": [], + "table_body": "
BaselineAgentsADE ↓FDE ↓
Random Trajectory18.8916.51
NN11.804.13
Multiagent NNN6.4012.04
MirrorN11.5914.93
", + "bbox": [ + 124, + 276, + 450, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Car trajectory prediction baselines. Nearest neighbor performs best overall, but our learned single-agent AR models outperform all baselines.", + "bbox": [ + 89, + 378, + 482, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d13e6894e58ebababc4444c6e8055ffd8d0bcba1d06a06f5a548822a59ab381.jpg", + "image_caption": [ + "Figure 6. Example results from our single-agent AR model (top row) and three-agent PAR model with location positional encoding (bottom row) on nuScenes. The predicted agent's ground truth trajectory is in pink, and the predicted future in blue. For the PAR model, the other two agents' ground truth states are in green. Qualitatively, the PAR model handles situations where single-agent predictions might lead to collisions (A, B), uses other agents' behavior to better adhere to road areas (A, C) without environment data, and predicts based on the speed changes of other cars (D).", + "Figure 8. Translation forecasting qualitative result on test set. 3D predictions are projected onto the image, isolating translation results by showing the ground-truth rotation. Using the PAR framework (right) instead of AR (left) improves object pose prediction." + ], + "image_footnote": [], + "bbox": [ + 96, + 425, + 483, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 89, + 727, + 183, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train AR and 3-agent PAR models using velocity tokens, acceleration tokens, and acceleration tokens combined with our location positional encoding. The results can be seen in Table 3. Note that the 3-agent PAR model uses the agent ID embedding and next timestep prediction. Acceleration tokens consistently outperform velocity tokens both for agent AR and 3-agent PAR models. This could be because the vocabulary size for acceleration tokens is much smaller and therefore easier to optimize. Regardless, both ways of tokenizing result in models that outperform our baselines (see", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/def4b1307324993288b96bbab1b5936cf16c7eabc70bf2c40351dbca37af15e4.jpg", + "image_caption": [ + "1-Agent AR" + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 702, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfbf7d426bc29faf021bd3a45c9e36115c17a0a2b6199eff482202ed58baa139.jpg", + "image_caption": [ + "2-Agent PAR" + ], + "image_footnote": [], + "bbox": [ + 717, + 88, + 903, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42d64316489f0226b383bb7ee1615551eff970998cd7c5f20176090569f35980.jpg", + "image_caption": [ + "Figure 7. Rotation forecasting qualitative result on test set. 3D predictions are projected onto the image, isolating rotation results by showing the ground-truth translation. Incorporating the hand agent in the PAR framework (right) improves object pose prediction over object-only AR (left).", + "1-Agent AR" + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 702, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b427a6868d2dc4570d1314493ae331ddef2aa66f616355d44fb90648f2130b8b.jpg", + "image_caption": [ + "2-Agent PAR" + ], + "image_footnote": [], + "bbox": [ + 717, + 290, + 903, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 - NN has a relatively low error on this dataset), and highlight that our framework is flexible such that a user can experiment with different ways of representing entities. For both token types, the 3-agent PAR model that is blind to location outperforms the AR model. While location information should help the model, it is possible that simply knowing whether other agents are slowing down or accelerating can help the model make better predictions. When adding location information via the LPE to our 3-agent PAR model, we see another performance gain in ADE and FDE.", + "bbox": [ + 511, + 474, + 906, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative examples of the AR model (top row) and 3-agent location-aware PAR model (bottom row) can be seen in Figure 6. Our method uses no image or environment data (e.g., lanes) as input. However, by reasoning over multiple agents, its predictions lead to fewer collisions and better reasoning about speed changes and driveable areas based solely on other agents' behaviors.", + "bbox": [ + 511, + 628, + 906, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Case Study 3: Object Pose Forecasting During Hand-Object Interaction", + "text_level": 1, + "bbox": [ + 511, + 762, + 906, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our final case study explores how hand-object interaction can be leveraged for object pose estimation. We define the human hand and the interacting object as two agents, with tokens representing distinct state types. We show that our PAR framework can jointly model these agents, improving 3D translation and rotation predictions for the object.", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12408", + "bbox": [ + 480, + 917, + 517, + 928 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 90, + 279, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset. We use the DexYCB dataset, which includes 1000 videos of 10 subjects performing object manipulation tasks with 20 distinct objects from the YCB-Video dataset. The data is split into 800 training, 40 validation, and 160 testing videos. We use one of 8 provided camera views. In each trial, subjects pick up and lift objects in randomized conditions. Labels include the object's SO(3) rotation and 3D translation, and the hand's 3D translation. We focus on predicting the object's rotation or translation.", + "bbox": [ + 89, + 112, + 483, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task-specific considerations. We tokenize object information in object-only experiments and both object and hand information in hand-object experiments. The object is represented as a 4D token for rotation forecasting (quaternion from SO(3) rotation) or a 3D token for translation forecasting (Euclidean coordinates). In hand-object experiments, the hand token is included with a 3D translation vector, and agent ID embeddings distinguish between the hand and object. Normalization is applied to all 3D translation vectors in both AR and PAR experiments; quaternions are normalized by definition and require no additional processing. An embedding layer projects the tokens into the transformer's hidden dimension, and another layer projects them back for prediction.", + "bbox": [ + 89, + 258, + 483, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For rotation-only forecasting, the loss is $\\mathcal{L}_{rot} = 1 - |\\hat{q}\\cdot q|$ , where $\\hat{q}$ is the predicted quaternion and $q$ the ground-truth quaternion. For translation-only forecasting, the loss $\\mathcal{L}_t$ is the mean squared error (MSE) between predicted and ground-truth translations. For PAR we predict relative object-to-hand translations at each frame, using the current hand position as origin, while for AR, we predict absolute object translations without considering the interacting agent. For PAR models, we add the loss $\\mathcal{L}_h$ , a MSE on hand translation. The object-only AR rotation model is optimized with $\\mathcal{L}_{rot}$ , while the PAR rotation model combines $\\mathcal{L}_{rot} + \\mathcal{L}_h$ ; similarly, the object-only translation model is trained with $\\mathcal{L}_t$ , and the hand-object translation model uses $\\mathcal{L}_t + \\mathcal{L}_h$ . At inference, the first half of each video is provided, and object predictions are autoregressively generated for the second half. Translation is evaluated using MSE, while rotation is measured using geodesic distance (GEO) on SO(3).", + "bbox": [ + 89, + 472, + 483, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Results", + "text_level": 1, + "bbox": [ + 89, + 741, + 184, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the object-only AR models to the hand-object PAR models in Table 5 for the two prediction tasks. We also present the baselines described in Sec. 3.3 in Table 6. Figures 7 and 8 show qualitative results on the rotation and translation predictions, respectively. In both prediction tasks, we observe that incorporating the human hand's interaction with the object enhances accuracy: for rotation, PAR results in a relative improvement of $8.9\\%$ over AR, and for translation, $41\\%$ . See Section 11.2 for additional qualitative results", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/81446936dd837ebf540ac9ca60064800d69fc4e79940603f020fcb15f61cb769.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodMSE ↓GEO (rad) ↓
Translation1-agent AR3.68 × 10-3-
Translation2-agent PAR2.17 × 10-3-
Rotation1-agent AR-0.919
Rotation2-agent PAR-0.837
", + "bbox": [ + 514, + 88, + 908, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f6d17556517a6feaca8c2dde64dcd81c81391a9aac6cb2d88bb053810ce1d8f8.jpg", + "table_caption": [ + "Table 5. Test set results on DexYCB dataset. For both rotation and translation forecasting, the 2-agent PAR model, which treats the hand as an additional agent, improves results." + ], + "table_footnote": [], + "table_body": "
BaselineTranslation - MSE (m2) ↓Rotation - GEO (rad) ↓
Random0.2442.196
Random Trajectory1.60 × 10-22.146
NN1.69 × 10-22.179
Multiagent NN1.71 × 10-22.170
Mirror1.20 × 10-2-
", + "bbox": [ + 514, + 253, + 923, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Test set results for DexYCB baselines. We cannot provide rotation results for the Mirror baseline, because the ground-truth does not include hand rotation, only 3D translation.", + "bbox": [ + 511, + 353, + 906, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with more sampled frames.", + "bbox": [ + 511, + 422, + 694, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion", + "text_level": 1, + "bbox": [ + 511, + 450, + 625, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work introduced the Poly-Autoregressive (PAR) framework, a unifying approach to prediction for multi-agent interactions. By applying the same transformer architecture and hyperparameters across diverse tasks, including action forecasting in social settings, trajectory prediction for autonomous vehicles, and object pose forecasting during hand-object interaction, we have demonstrated the versatility and robustness of our framework.", + "bbox": [ + 511, + 477, + 906, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our findings underscore the crucial importance of considering the influence of multiple agents in a scene for prediction tasks. By modeling interactions, we significantly improved prediction accuracy over single-agent approaches on all three problems we considered. While we achieved promising results with a simple architecture, we have only provided a starting point that can be built upon extensively. For instance, incorporating environmental context or tokenizing pixel patches, especially as a way to relax our assumption on high-quality tracking, are avenues for further research using PAR. It would be interesting to experiment with scaling the data and model. We do not specifically consider the relative importance of neighboring agents, which is an interesting future research direction (ex. dynamic attention mechanism).", + "bbox": [ + 511, + 599, + 908, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our PAR framework provides a simple and generalizable foundation of universal building blocks, ready for extension or refinement in future tasks. The PAR framework holds potential for advancing AI systems, enhancing prediction capabilities and enabling more accurate navigation and operation in real-world, multi-agent interactions.", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12409", + "bbox": [ + 480, + 917, + 517, + 928 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 188, + 187, + 204 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2", + "[2] Fred Attneave. Some informational aspects of visual perception. Psychological review, 61(3):183, 1954. 2", + "[3] Murchana Baruah and Bonny Banerjee. A multimodal predictive agent model for human interaction generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 1022-1023, 2020. 2", + "[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2", + "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 6, 1", + "[6] Sergio Casas, Wenjie Luo, and Raquel Urtasun. Intentnet: Learning to predict intention from raw sensor data. In Conference on Robot Learning, pages 947-956. PMLR, 2018. 1", + "[7] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 1", + "[8] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 1", + "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2", + "[10] Baptiste Chopin, Hao Tang, Naima Otberdout, Mohamed Daoudi, and Nicu Sebe. Interaction transformer for human reaction generation. IEEE Transactions on Multimedia, pages 1-13, 2023. 2", + "[11] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, KueiDa Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1" + ], + "bbox": [ + 93, + 213, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "We thank Jane Wu, Himanshu Singh, Georgios Pavlakos, and João Carreira for useful discussions and feedback. This work was supported by ONR MURI N00014-21-1-2801 and NSF Graduate Fellowships to NT and TS.", + "[12] Henggang Cui, Vladan Radosavljevic, Fang-Chieh Chou, Tsung-Han Lin, Thi Nguyen, Tzu-Kuo Huang, Jeff Schneider, and Nemanja Djuric. Multimodal trajectory predictions for autonomous driving using deep convolutional networks. In 2019 international conference on robotics and automation (icra), pages 2090-2096. IEEE, 2019. 1", + "[13] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 1", + "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 1", + "[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 1", + "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1", + "[17] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. In European Conference on Computer Vision (ECCV), page 3, 2024. 2", + "[18] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In International Conference on Machine Learning, pages 1242-1250. PMLR, 2014. 2", + "[19] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6047-6056, 2018. 2, 5", + "[20] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13053-13064, 2022. 2", + "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1", + "[22] Yanjun Huang, Jiatong Du, Ziru Yang, Zewei Zhou, Lin Zhang, and Hong Chen. A survey on trajectory-prediction methods for autonomous driving. IEEE Transactions on Intelligent Vehicles, 7(3):652-674, 2022. 1", + "[23] Boris Ivanovic, Guanyu Song, Igor Gilitschenski, and Marco Pavone. trajdata: A unified interface to multiple human trajectory datasets. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, New Orleans, USA, 2023. 6", + "[24] Bolin Lai, Sam Toyer, Tushar Nagarajan, Rohit Girdhar, Shengxin Zha, James M Rehg, Kris Kitani, Kristen Grau" + ], + "bbox": [ + 93, + 92, + 908, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12410", + "bbox": [ + 478, + 917, + 519, + 929 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "man, Ruta Desai, and Miao Liu. Human action anticipation: A survey. arXiv preprint arXiv:2410.14045, 2024. 1", + "[25] Hugo Larochelle and Iain Murray. The neural autoregressive distribution estimator. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 29-37. JMLR Workshop and Conference Proceedings, 2011. 2", + "[26] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 1", + "[27] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 2", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2", + "[29] Siyuan Brandon Loh, Debaditya Roy, and Basura Fernando. Long-term action forecasting using multi-headed attention-based variational recurrent neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2419-2427, 2022. 1", + "[30] Vongani Maluleke, Lea Müller, Jathushan Rajasegaran, Georgios Pavlakos, Shiry Ginosar, Angjoo Kanazawa, and Jitendra Malik. Synergy and synchrony in couple dances. arXiv preprint arXiv:2409.04440, 2024. 2", + "[31] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 1", + "[32] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20395-20405, 2022. 2", + "[33] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2", + "[34] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In ArXiv, 2024. 2", + "[35] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 1", + "[36] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer, 2018. 2", + "[37] Alec Radford. Improving language understanding by generative pre-training. 2018. 2", + "[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Ilija Radosavovic, Bike Zhang, Baifeng Shi, Jathushan Rajasegaran, Sarthak Kamat, Trevor Darrell, Koushil Sreenath, and Jitendra Malik. Humanoid locomotion as next token prediction. arXiv preprint arXiv:2402.19469, 2024. 2", + "[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, Christoph Feichtenhofer, and Jitendra Malik. On the benefits of 3d pose and tracking for human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2023. 1", + "[41] Jathushan Rajasegaran, Ilija Radosavovic, Rahul Ravishankar, Yossi Gandelsman, Christoph Feichtenhofer, and Jitendra Malik. An empirical study of autoregressive pre-training from videos. arXiv preprint arXiv:2501.05453, 2025. 2", + "[42] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 1", + "[43] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 1", + "[44] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8579-8590, 2023. 2, 1", + "[45] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 30(1):50-64, 1951. 2", + "[46] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. arXiv preprint arXiv:2403.18811, 2024. 2", + "[47] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5", + "[48] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Rahul Sukthankar, Kevin Murphy, and Cordelia Schmid. Relational action forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 273-283, 2019. 1", + "[49] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 1", + "[50] Lucas Theis and Matthias Bethge. Generative image modeling using spatial lstms. Advances in neural information processing systems, 28, 2015. 2", + "[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12411", + "bbox": [ + 480, + 917, + 517, + 928 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 5", + "[52] Ameni Trabelsi, Mohamed Chaabane, Nathaniel Blanchard, and Ross Beveridge. A pose proposal and refinement network for better 6d object pose estimation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2382-2391, 2021. 1", + "[53] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 2", + "[54] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2", + "[55] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16611–16621, 2021. 1", + "[56] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373. IEEE, 2020. 1", + "[57] Bowen Wen, Jonathan Tremblay, Valts Blukis, Stephen Tyree, Thomas Müller, Alex Evans, Dieter Fox, Jan Kautz, and Stan Birchfield. Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 606-617, 2023.", + "[58] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868-17879, 2024. 1", + "[59] Jane Wu, Georgios Pavlakos, Georgia Gkioxari, and Jitendra Malik. Reconstructing hand-held objects in 3d. arXiv preprint arXiv:2404.06507, 2024. 1", + "[60] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 1", + "[61] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 1" + ], + "bbox": [ + 91, + 90, + 483, + 768 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12412", + "bbox": [ + 480, + 917, + 519, + 929 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Poly-Autoregressive Prediction for Modeling Interactions/43d051a6-6821-4e3b-bd9a-a4c244857daa_model.json b/2025/Poly-Autoregressive Prediction for Modeling Interactions/43d051a6-6821-4e3b-bd9a-a4c244857daa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a4a824ffd2b45b7aad9915a4115efd20f15a311b --- /dev/null +++ b/2025/Poly-Autoregressive Prediction for Modeling Interactions/43d051a6-6821-4e3b-bd9a-a4c244857daa_model.json @@ -0,0 +1,2455 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.21, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "Poly-Autoregressive Prediction for Modeling Interactions" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.18, + 0.89, + 0.226 + ], + "angle": 0, + "content": "Neerja Thakkar\\(^{1}\\) Tara Sadjadpour\\(^{1}\\) Jathushan Rajasegeran\\(^{1}\\) Shiry Ginosar\\(^{2,3}\\) Jitendra Malik\\(^{1}\\) \n\\(^{1}\\)UC Berkeley, \\(^{2}\\)Toyota Technical Institute at Chicago, \\(^{3}\\)Google DeepMind" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.279, + 0.462, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.394, + 0.33, + 0.408 + ], + "angle": 0, + "content": "(a) Autoregression (AR)" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.274, + 0.9, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.394, + 0.776, + 0.408 + ], + "angle": 0, + "content": "(b) Poly-autoregression (PAR)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.419, + 0.908, + 0.464 + ], + "angle": 0, + "content": "Figure 1. Inference for (a) autoregressive (AR) models and (b) our proposed poly-autoregressive (PAR) model. Solid indicates ground-truth tokens which represent a tracked data modality such as action or 6DOF pose; striped represents predicted output tokens. Color denotes agent identity. Compared to AR models, the PAR model takes other agents' tokens as inputs when making a prediction for the next timestep." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.484, + 0.327, + 0.5 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.515, + 0.486, + 0.787 + ], + "angle": 0, + "content": "We introduce a simple framework for predicting the behavior of an agent in multi-agent settings. In contrast to autoregressive (AR) tasks, such as language processing, our focus is on scenarios with multiple agents whose interactions are shaped by physical constraints and internal motivations. To this end, we propose Poly-Autoregressive (PAR) modeling, which forecasts an ego agent's future behavior by reasoning about the ego agent's state history and the past and current states of other interacting agents. At its core, PAR represents the behavior of all agents as a sequence of tokens, each representing an agent's state at a specific timestep. With minimal data pre-processing changes, we show that PAR can be applied to three different problems: human action forecasting in social situations, trajectory prediction for autonomous vehicles, and object pose forecasting during hand-object interaction. Using a small proof-of-concept transformer backbone, PAR outperforms AR across these three scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.222, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Large language models (LLMs) have been very successful with natural language processing (NLP) tasks, which require accurate reasoning over relationships words have within a body of text. A key component of LLMs is autoregressive" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.59 + ], + "angle": 0, + "content": "(AR) modeling, where each word token is predicted based on a sequence of preceding word tokens. Building on the success of AR modeling in the NLP community, this work focuses on modeling the dynamic relationships between agents and entities in everyday interactions that occur in the physical world, such as social interactions, driving, and hand-object interactions." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.595, + 0.909, + 0.777 + ], + "angle": 0, + "content": "Unlike language which is structured through grammar and semantics, interactions in the physical world are dictated by both the laws of physics (e.g. how a hand grasps an object) and the internal state of each agent (e.g. the trajectory an agent chooses to move a grasped object in), a latent variable that we know nothing about. Furthermore, as opposed to text where each word follows another unidimensionally, the states of multiple agents are changing simultaneously. For example, in social situations, the history of a single person's past states does not alone determine the dynamics of their future states; we also need to consider the states of other agents. We argue, therefore, that AR modeling alone is insufficient." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce poly-autoregressive (PAR) modeling, a simple unifying approach to model the influence of other agents and entities on one's behavior. We model behavior as a temporal sequence of states and predict an ego agent's future behavior conditioned on the history of behavior of the ego agent as well as the rest of the agents. By considering other agents' behaviors, we demonstrate that our approach significantly improves upon the ill-posed problem" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12402" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.411, + 0.106 + ], + "angle": 0, + "content": "of single-agent prediction in interactive settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.486, + 0.289 + ], + "angle": 0, + "content": "The PAR framework uses a transformer-based model for next-token prediction. Transformers have shown great success in language modeling and naturally lend themselves to predicting behavior over time. In an interaction scenario of \\( N \\) agents, our model predicts the future behavior of an ego \\( (N^{\\mathrm{th}}) \\) agent conditioned on its past behavior and the behavior of the other \\( N - 1 \\) non-ego agents (see Figure 1b). Each prediction task may model a different behavior modality of interest, e.g. actions for social action prediction, or 6DoF pose for object forecasting in hand-object interactions. We apply PAR to three different case studies of common real-world interactions:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.293, + 0.483, + 0.383 + ], + "angle": 0, + "content": "1. Social action prediction. We test our method on the AVA benchmark [19] for action forecasting. By incorporating both the ego and another agent using PAR, we get a overall \\(+1.9\\) absolute mAP gain over AR, which only models the ego agent to predict its future behavior, and an absolute \\(+3.5\\) mAP gain on 2-person interaction classes." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.384, + 0.484, + 0.488 + ], + "angle": 0, + "content": "2. Trajectory prediction for autonomous vehicles. When forecasting future \\(xy\\) locations of an ego vehicle, incorporating the locations of neighboring vehicles with PAR outperforms AR, which only uses the ego vehicle's preceding trajectory as input. Specifically, on the nuScenes dataset [5], PAR outperforms AR with a relative improvement of \\(6.3\\%\\) ADE and \\(6.4\\%\\) FDE." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.489, + 0.484, + 0.61 + ], + "angle": 0, + "content": "3. 6DoF object pose forecasting during hand-object interaction. We use the DexYCB dataset [8], where we treat the object as the ego agent and the hand as the interacting agent. While PAR integrates that hand's 3D location and object pose history, AR only uses the object's pose history to predict the object's 6DoF pose. PAR outperforms AR with relative improvements of \\(8.9\\%\\) and \\(41.0\\%\\) for the rotation and translation predictions, respectively." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.293, + 0.484, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.486, + 0.748 + ], + "angle": 0, + "content": "In all these settings, we find that incorporating the behavior of other agents in the scene improves predictions of the ego agent's behavior. All of these problems are modeled via the same simple PAR framework and implemented using the same proof-of-concept 4 million parameter transformer without any modifications to the base framework or architecture, only to data pre-processing and choice of tokenization. We also provide an example of a simple way to build on our architecture through a location positional encoding, Sec. 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The primary contribution of this work is a versatile framework that can be applied to a diverse range of settings, without modifications aside from domain-specific data processing. Our results suggest that PAR provides a simple formulation that, with a more complex transformer backbone and larger datasets, could enhance prediction of diverse multi-agent interactions across various problem domains. To facilitate further exploration and development, we have released our code, which contains the building blocks to use PAR for modeling other types of multi-agent interactions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.657, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.116, + 0.908, + 0.268 + ], + "angle": 0, + "content": "Autoregressive models. Autoregressive modeling has a rich history in information theory and deep learning, tracing back to Shannon's paper on language prediction [45] and Attneave's study on visual perception [2]. These foundational works laid the groundwork for modern applications in deep learning, including [25], which revisits neural autoregressive models and [18, 50], which explore continuous-valued modeling. [53] developed PixelRNN and PixelCNN to autoregressively generate images one pixel at a time using RNN and CNN architectures, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.909, + 0.565 + ], + "angle": 0, + "content": "The development of the transformer model [54] spurred progress in computer vision with the image transformer [36] and the vision transformer [14], and autoregressive models [9, 41] and more notably in NLP, where the GPT family of models [4, 37, 38] has demonstrated the power of large-scale unsupervised autoregressive pre-training. Recent research has focused on multimodal learning, exemplified by the Flamingo [1] or LlaVa [28] models, which combine vision and language processing capabilities, illustrating the versatility of autoregressive models across various domains in artificial intelligence. While these approaches operate on image patches and word tokens, we operate on symbolic representations extracted from in-the-wild videos showing natural interactions. A recent approach [39] frames humanoid locomotion as an autoregressive next-token prediction task that operates on two types of continuous tokens: observations and actions. This approach projects continuous tokens to the hidden dimension and uses a shifted loss, similar to the next-timestep prediction proposed in our framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.909, + 0.816 + ], + "angle": 0, + "content": "Multi-agent regressive models. Several prior works addressed modeling specific multi-agent problems via regressive models as one-off case studies. We introduce the PAR framework to unify these efforts into a single cohesive framework. Many behavior prediction works focus on two agents engaging in social interaction, whether it be dyadic communication [32-34] or social dance [30, 46]. These studies primarily tackle the challenge of predicting the state of an interacting partner (Person B) based on the input from Person A's state, sometimes extending predictions into the future [20, 30]. While earlier works used architectures such as variational RNNs [3], recent works have predominantly adopted transformer architectures for social interaction modeling [10, 20, 32, 33, 46], with some works exploring diffusion [27], or diffusion with attention [17]. Our PAR framework focuses on transformer models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Works encompassed by the PAR framework extend beyond human social interaction. Many multi-agent human or car trajectory prediction approaches use autoregressive prediction. For instance, MotionLM [44] utilizes a transformer decoder that processes multi-agent tokens, incorporating a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12403" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "learned agent ID embedding. This methodology informs our approach across all our case studies. Critically, in contrast to all prior multi-agent regressive works that designed solutions to address specific applications, we demonstrate that we can unify a diverse set of multi-agent regressive problems under a single PAR framework. See Appendix Sec. 9 for case-study-specific works." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.215, + 0.367, + 0.233 + ], + "angle": 0, + "content": "3. Poly-Autoregressive Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.485, + 0.316 + ], + "angle": 0, + "content": "Our goal is to model the behavior of an agent or entity while taking into account any other agents it interacts with, if any. To evaluate the performance of our model in capturing interaction dynamics, we predict the agent's future behavior and compare it against ground-truth data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.484, + 0.391 + ], + "angle": 0, + "content": "We define the following task: In an interaction comprised of \\( N \\) agents, given the observed past states of the \\( N - 1 \\) interacting agents, and the observed or previously-predicted past states of the \\( N^{th} \\) ego agent, predict the future states of the \\( N^{th} \\) ego agent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.468 + ], + "angle": 0, + "content": "We define a transformer-based poly-autoregressive (PAR) predictor, \\(\\mathcal{P}\\), that learns to model temporally long-range interactions in the input sequence. The inputs to the predictor are the past states of the \\(N\\) interacting agents, and its output is the predicted future state of the \\(N^{\\mathrm{th}}\\) ego agent." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.475, + 0.274, + 0.489 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.484, + 0.651 + ], + "angle": 0, + "content": "Let \\(\\mathbf{S} = \\{\\mathbf{s}_i\\}_{i=1}^T\\) be a temporal sequence of agent states, \\(\\mathbf{s}_i\\). We use \\(\\mathbf{S}^N\\) and \\(\\mathbf{S}^{1:N-1}\\) to denote the temporal sequences of states of the \\(N_{th}\\) agent and of the other \\(N-1\\) agents, respectively. For each timestep \\(t \\in [t_{\\pi}, T]\\), where \\(t_{\\pi} \\in [1, T]\\) is the time we start predicting, we take as input all other \\(N-1\\) agents' past observed state sequences \\(\\mathbf{S}_{1:t-1}^{1:N-1}\\) along with the \\(N_{th}\\) agent's past observed states up to \\(t_{\\pi}\\), \\(\\mathbf{S}_{1:t_{\\pi}}^N\\), and any of its previously predicted past states \\(\\hat{\\mathbf{S}}_{t_{\\pi}+1:t-1}^N\\), if available (see Fig. 1). Our predictor, \\(\\mathcal{P}\\), then poly-autoregressively predicts the \\(N_{th}\\) agent's future states one time-step at a time:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.662, + 0.484, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {s}} _ {t} ^ {N} = \\mathcal {P} \\left(\\mathbf {S} _ {1: t - 1} ^ {1: N - 1}, \\mathbf {S} _ {1: t _ {\\pi}} ^ {N}, \\hat {\\mathbf {S}} _ {t _ {\\pi} + 1: t - 1} ^ {N}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.716 + ], + "angle": 0, + "content": "\\(\\mathcal{P}\\) learns to model the distribution over the next timestep of the \\(N_{th}\\) agent's states, given all other agents' states:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.722, + 0.483, + 0.742 + ], + "angle": 0, + "content": "\\[\np \\left(\\hat {\\mathbf {s}} _ {t} ^ {N} \\mid \\mathbf {S} _ {1: t - 1} ^ {1: N - 1}, \\mathbf {S} _ {1: t - 1} ^ {N}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.484, + 0.808 + ], + "angle": 0, + "content": "While we provide the observed ground truth states of other agents at inference, during training, we jointly maximize the likelihood of all \\(N\\) agents by computing losses on their future state predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.808, + 0.484, + 0.839 + ], + "angle": 0, + "content": "We train the predictor by maximizing the likelihood of the target state \\(y\\) at time \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.846, + 0.387, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {P}} = E _ {y \\sim p (y)} [ - \\log (p (\\mathbf {s} _ {t} ^ {N}) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where the target state \\(y\\) at \\(t\\) is computed from the \\(N_{th}\\) agent ground truth future state." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.833, + 0.107 + ], + "angle": 0, + "content": "3.2. The Poly-Autoregressive Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.324 + ], + "angle": 0, + "content": "We address the problem of forecasting the future states of an agent (from time \\( t \\) to \\( T \\)) in a data-driven way, given a temporal sequence of past states (from time 1 to \\( t - 1 \\)). We assume that our agent has some feature, or a set of features, of interest in a video (e.g., 3D pose) that we can tokenize. We predict the future states of the agent in terms of this tokenized feature (or set of), where we use one token (or set of tokens) per time step. The predicted tokens can be discrete (i.e., an index into a feature codebook) or continuous (i.e., a vector of one or more continuous values). The loss \\( \\ell \\) will depend on the problem's specifics and the type of token used. To train the model to predict the future, we rely on all the interaction dynamics of length \\( T \\) in our training dataset as ground truth examples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.4 + ], + "angle": 0, + "content": "As a baseline, we consider the single-agent autoregressive (AR) paradigm, where a transformer is trained to perform next-token-prediction with teacher forcing. AR uses greedy sampling to generate sequences at inference time, predicting one next token at a time (Fig. 1(a))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.906, + 0.551 + ], + "angle": 0, + "content": "In contrast, our multi-agent poly-autoregressive (PAR) framework considers the other \\(N - 1\\) agents in the scene when predicting the future state of the \\(N\\)th agent. In this setup, we utilize the features of interest of all \\(N\\) agents, yielding \\(N\\) tokens at each timestep for a total of \\(N * T\\) tokens. In practice, we operate on a flattened sequence of \\(N * T\\) tokens. Instead of using the AR training procedure in this multi-agent case (as in Fig. 3a), we jointly model the \\(N\\) agents at each timestep by introducing the following features to our PAR framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.907, + 0.695 + ], + "angle": 0, + "content": "Next-timestep prediction. A standard AR model predicts the next token. Given the flattened sequence of \\( N * T \\) tokens our model operates on, next token prediction would take as input an agent \\( k \\) at timestep \\( t \\) and predict agent \\( k + 1 \\)'s state at the same timestep \\( t \\) (as in Fig. 3a). However, our goal is to predict the input agent \\( k \\)'s future state at time \\( t + 1 \\). Therefore, we perform same-agent next-timestep prediction rather than next-token prediction (see Fig. 3b for an illustration of same-agent next-timestep at training)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.906, + 0.775 + ], + "angle": 0, + "content": "Learned agent identity embedding. When giving a model information corresponding to multiple agents, the model can benefit from knowing which token corresponds to which agent. We give the model this information with a learned agent ID embedding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.842 + ], + "angle": 0, + "content": "Joint training. We train the model to jointly predict the future of all agents by computing a loss on the predicted tokens of all agents (Fig. 3b). Please refer to Section 3.1 for our inference paradigm." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.771, + 0.865 + ], + "angle": 0, + "content": "3.3. Task-Specific Considerations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Our simple PAR approach unifies diverse problems under a single framework and architecture without any modifications." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12404" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.898, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.324, + 0.908, + 0.396 + ], + "angle": 0, + "content": "Figure 2. The PAR Framework. We begin by collecting a video dataset, such as AVA (top) or DexYCB (bottom). Then, using dataset labels or computer vision techniques, a trajectory of a given modality for our prediction task is extracted for each agent, such as action class labels (top) or object pose and 3D hand translation (bottom). Data is then tokenized, either through discretization or directly using continuous values, with our framework supporting both formats. Based on the tokenization and prediction task, we choose the appropriate loss function for PAR training. After training with PAR, predicted tokens can be decoded back to data space and evaluated with relevant metrics." + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.406, + 0.451, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.515, + 0.4, + 0.527 + ], + "angle": 0, + "content": "(a) AR: multi-agent, next-token training." + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.528, + 0.451, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.639, + 0.411, + 0.652 + ], + "angle": 0, + "content": "(b) PAR: same-agent, next-timestep training." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.774 + ], + "angle": 0, + "content": "Figure 3. Training with teacher forcing for (a) multi-agent next-token prediction in autoregressive models and (b) multi-agent poly-autoregressive models. Solid vs striped indicates a ground-truth vs predicted token, respectively. Color denotes agent identity. The AR model is trained for next-token prediction, while the PAR model is trained to predict the next timestep of the same agent. Three agents are shown for ease of visualization, but the PAR model supports an arbitrary number of agents." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.484, + 0.804 + ], + "angle": 0, + "content": "In order to formulate a problem as interaction-conditioned prediction, users must consider several task-specific details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.484, + 0.835 + ], + "angle": 0, + "content": "Fig. 2 gives an overview of how the PAR framework disentangles multi-agent learning from problem-specific modeling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Data. The input data source in our example tasks is always a collection of videos. From these videos, we extract various modalities relevant to the task at hand. These modalities can range from high-level features, such as action class labels, to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.908, + 0.454 + ], + "angle": 0, + "content": "low-level ones, such as 3D pose (Fig. 2 first two columns). We assume that each agent in the dataset is detected at each frame and is associated with an agent ID." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.909, + 0.642 + ], + "angle": 0, + "content": "Tokenization. Our framework supports both discrete, quantized tokens and continuous vector tokens. The choice between discrete and continuous depends on the nature of the task. In the case of discrete tokens, we use a standard embedding layer to project to the hidden dimension. For continuous tokens, we train a projection layer to project the token into the hidden dimension of the transformer. For instance, if our continuous token is a 3D vector with an \\((x,y,z)\\) 3D location coordinate and our hidden dimension is 128, our projection layer will project from 3 to 128 dimensions. We also train an un-projection layer that reverts the hidden dimension to the original token dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.908, + 0.726 + ], + "angle": 0, + "content": "Loss. The type of token and task-specific considerations dictate the loss function \\(\\ell\\) applied during model training. For discrete tokens, a classification loss is appropriate. For continuous tokens, we use a regression loss on the original token dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Baselines. We compare to the following baselines, where applicable on a case-by-case basis:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.794 + ], + "angle": 0, + "content": "- Random token: pick random tokens from the best available token space and use as the prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.825 + ], + "angle": 0, + "content": "- Random trajectory: pick at random a trajectory from the training dataset to use as the prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.87 + ], + "angle": 0, + "content": "- NN: Given an input agent \\(A\\)'s trajectory history, find the closest trajectory to it in the training set, belonging to \\(A^T\\). Use \\(A^T\\)'s future as the predicted future." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Multiagent NN: In a dataset with two interacting partners \\( A \\) and \\( B \\), where \\( B \\) is the ego agent, given an input agent" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.917, + 0.519, + 0.928 + ], + "angle": 0, + "content": "12405" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "\\(A\\)'s trajectory history, find the closest trajectory to it in the training set, belonging to \\(A^T\\). Use \\(A^T\\)'s interaction partner's \\(B^T\\)'s future as the prediction." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.14, + 0.483, + 0.186 + ], + "angle": 0, + "content": "- Mirror: In a dataset with two interacting partners \\( A \\) and \\( B \\), use the ground truth future of agent \\( B \\) as the predicted future for agent \\( A \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.193, + 0.403, + 0.209 + ], + "angle": 0, + "content": "3.4. Framework Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.483, + 0.245 + ], + "angle": 0, + "content": "We keep the following implementation details constant for all case studies (see also Sec. 10)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.25, + 0.484, + 0.296 + ], + "angle": 0, + "content": "Learned agent ID embedding. Our learned agent ID embedding consists of the integer agent ID mapped to a hidden dim-sized vector, and summed to the token embedding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.484, + 0.452 + ], + "angle": 0, + "content": "Architecture. For all case studies, we use the Llama [51] transformer decoder architecture with 8 layers, 8 attention heads, and a hidden and intermediate dimension of 128. The decoder has \\(\\sim 4.4\\mathrm{M}\\) learned parameters, not including learned embedding layers which add a few thousand more parameters. A rotary positional encoding [47] is used in addition to other summed encodings (i.e. agent ID embedding, locational positional encoding in Sec. 5). We train using teacher forcing. The only hyperparameter that changes between case studies is the learning rate." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.464, + 0.449, + 0.481 + ], + "angle": 0, + "content": "4. Case Study 1: Social Action Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.61 + ], + "angle": 0, + "content": "Our first case study involves forecasting human actions. Human behaviors are fundamentally social; for instance, individuals frequently walk in groups and alternate between speaking and listening roles when conversing. Certain actions, like hugging or handshaking, are intrinsically multiperson. Therefore, modeling human interactions should help improve action forecasting performance, especially on multiperson actions, which we show in this case study." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.618, + 0.282, + 0.634 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Dataset. The Atomic Visual Actions (AVA) dataset [19] comprises 235 training and 64 15-minute validation videos from movies. Annotations are provided at a 1Hz frequency, detailing bounding boxes and tracks for individuals within the frame, and each person's actions within a 1-second timeframe. Individuals may engage in multiple concurrent actions from a repertoire of 60 distinct action classes (e.g., sitting and talking simultaneously). For our analysis, we select clips featuring a continuous sequence of an agent's actions spanning at least 4s, splitting sequences exceeding 12s. We use the first half of each clip as history to predict the second half. For any ego agent trajectory, we pick a second agent by selecting the person present in the scene for the longest subset of the ego agent's trajectory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Task-specific considerations. Each agent's token \\(\\mathcal{A}\\) represents an 60-dimensional vector that corresponds to the actions performed at a specific timestep. Each element denotes" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.197 + ], + "angle": 0, + "content": "
MethodTimestep predAg-ID embdmAP ↑
1-agent ARN/AN/A40.7
2-agent ARXX38.0
2-agent PAR*X✓40.2
2-agent PAR*✓X40.0
2-agent PAR✓✓42.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.207, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Table 1. PAR action forecasting performance on AVA We evaluate 1 and 2-agent AR methods, two 2-agent PAR ablations (rows 3 and 4, \\(\\mathrm{PAR^{*}}\\)), and our PAR method. Without next-timestep prediction (see Fig. 3) or a learned agent ID embedding, our model struggles with multi-agent reasoning, performing worse than the AR baseline. With both components, the 2-agent PAR model achieves a \\(+1.9\\) mAP gain over the AR method (see Fig. 11 and Fig. 5 for class breakdown)." + }, + { + "type": "table", + "bbox": [ + 0.569, + 0.323, + 0.851, + 0.43 + ], + "angle": 0, + "content": "
BaselineAgentsmAP ↑
Random Token13.46
Random Training Traj13.44
Nearest Neighbor113.17
Multiagent NN25.10
Mirror27.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Table 2. AVA baselines While the nearest neighbor baseline performs best among baselines, it is still significantly worse than the AR model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.643 + ], + "angle": 0, + "content": "the probability of a particular action class being enacted; ground-truth inputs are a binary vector. We implement an embedding layer that projects these tokens into the transformer's hidden dimension, as well as an un-projection layer that reverts them back to the original 60D token space for the purposes of loss calculation and output generation. We do not explicitly require the outputs to be values between 0 and 1. We use a MSE regression loss on the 60D action tokens: \\(\\mathcal{L} = \\frac{1}{n}\\sum_{i=1}^{n}(\\mathcal{A}_i - \\hat{\\mathcal{A}}_i)^2\\). Our evaluation metric is the mean average precision (mAP) on the 60 AVA classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.703 + ], + "angle": 0, + "content": "We implement all baselines described in 3.3, where Random Token corresponds to a random 60D vector sampled from 0 to 1. NN and Multiagent NN use Hamming distance as the distance metric." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.608, + 0.727 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We report the performance of a single-agent AR model as a baseline, in the first line of Table 1. The AR model is significantly better than our baselines (see Table 2), the strongest baseline being the single-agent NN. We compare these baselines to our 2-agent PAR model (last line) and various ablations where we remove the agent ID embedding and perform next-token rather than same-agent next-timestep prediction. The second line of the table corresponds to multi-agent next-token prediction (Fig. 3a). We see that this approach confuses the model, and the performance is significantly worse than just training on and considering a single agent. However," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.918, + 0.519, + 0.929 + ], + "angle": 0, + "content": "12406" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.093, + 0.865, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.3, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Figure 4. Action forecasting example. The distribution over ground truth actions are in white, and our predictions in red. A 6s action history (1Hz) is input, and 6s of future actions predicted. In the scene, the man and woman alternate between talking and listening. Initially, the man is talking. The AR model predicts he will continue talking, while the 2-agent PAR model recognizes the woman is talking and predicts more accurate turn-taking behavior." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.363, + 0.475, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.52, + 0.487, + 0.592 + ], + "angle": 0, + "content": "Figure 5. Per-class mAP for AVA 2-person actions. We see performance improvement on almost all 2-person AVA action classes ((P) stands for \"a person\"). Some absolute mAP gains are particularly significant: listen to \\(+7.0\\), kiss \\(+8.3\\), fight/hit \\(+5.7\\), talk to \\(+4.4\\), hug \\(+5.7\\), and hand shake \\(+4.0\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.484, + 0.701 + ], + "angle": 0, + "content": "as we add various components of our PAR approach, the performance improves, and with both the next timestep prediction and agent ID embedding, we get a \\(+1.9\\) mAP gain. When only considering 2-person action classes (enumerated in Fig. 5), our mAP is 36.3 on the single agent PAR model and 39.8 on the 2-agent PAR model, a \\(+3.5\\) mAP gain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In Fig. 4 we see an example of action forecasting. In the input history, the man talks and the woman listens. In the future, the woman talks, and the man listens. Our 2-agent PAR model (bottom row) better understands that talking and listening actions are complementary actions, while the AR model doesn't learn this correlation. We see quantitative evidence of this in Fig. 5, with per-class mAPs for our AR vs 2-agent PAR model for 2-person action classes. Here, talk to gets a \\(+4.4\\) mAP gain and listen to gets a \\(+7.0\\) mAP gain when we train a multi-agent model. We see a significant boost on many other interaction-related action classes—for instance, kiss a person \\(+8.3\\) and fight/hit a person \\(+5.7\\) mAP—and on single-person actions, see Fig. 11." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.907, + 0.392 + ], + "angle": 0, + "content": "5. Case Study 2: Multiagent Car Trajectory Prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.909, + 0.48 + ], + "angle": 0, + "content": "Our second case study focuses on predicting car trajectories. Trajectory prediction requires a vehicle to be aware of other cars on the road to avoid collisions and promote cooperative behavior. This study demonstrates how our framework enables the joint modeling of multiple vehicles' movements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.487, + 0.707, + 0.504 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.908, + 0.584 + ], + "angle": 0, + "content": "Dataset. We use nuScenes [5], inputting 2 seconds of positions to forecast vehicle positions 6 seconds ahead. Specifically, our objective is to predict the \\(xy\\) coordinates of each agent, exclusively considering vehicles as agents. We use the trajdata interface [23] to load and visualize the data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Task-specific considerations. Instead of discretizing the \\( xy \\) position space, we discretize the motion, resulting in discrete velocity or acceleration tokens. These integer tokens are projected to the transformer hidden dimension using the Llama token embedding layer. Inputting only these tokens results in our PAR model knowing what speed the other agents are going at, but not where they are. It is important the model has this awareness (it should know if two agents are going to collide), so our model needs to reason over this second modality of location. We implement this by passing locations relative to the agent we are predicting into a sincosine positional embedding (see details in Sec. 12.1), which we denote a location positional encoding (LPE). The LPE is summed to our token embeddings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We use a cross-entropy classification loss on our discrete tokens: \\(\\mathcal{L} = E_{y\\sim p(y)}[-\\log (p(\\mathbf{s}_{t_\\pi}^T))]\\). We use the standard average displacement error (ADE) and final displacement error (FDE) to evaluate our predicted trajectories. For our baselines (Sec. 3.3), we use the closest agent at the current timestep for Multiagent NN and Mirror. For NN and Multiagent NN we use MSE as the distance metric." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.917, + 0.521, + 0.929 + ], + "angle": 0, + "content": "17" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.468, + 0.196 + ], + "angle": 0, + "content": "
Token typeLPEMethodADE ↓FDE ↓
VelocityX1-agent AR1.503.64
VelocityX3-agent PAR1.453.51
AccelerationX1-agent AR1.443.57
AccelerationX3-agent PAR1.403.44
Accleration✓3-agent PAR1.353.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.206, + 0.483, + 0.263 + ], + "angle": 0, + "content": "Table 3. Car trajectory prediction performance. Using acceleration tokens and 3-agent PAR results in a stronger performance over velocity tokens and single-agent AR. Adding location via a positional encoding (LPE) further improves results." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.277, + 0.451, + 0.368 + ], + "angle": 0, + "content": "
BaselineAgentsADE ↓FDE ↓
Random Trajectory18.8916.51
NN11.804.13
Multiagent NNN6.4012.04
MirrorN11.5914.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.421 + ], + "angle": 0, + "content": "Table 4. Car trajectory prediction baselines. Nearest neighbor performs best overall, but our learned single-agent AR models outperform all baselines." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.426, + 0.485, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.723 + ], + "angle": 0, + "content": "Figure 6. Example results from our single-agent AR model (top row) and three-agent PAR model with location positional encoding (bottom row) on nuScenes. The predicted agent's ground truth trajectory is in pink, and the predicted future in blue. For the PAR model, the other two agents' ground truth states are in green. Qualitatively, the PAR model handles situations where single-agent predictions might lead to collisions (A, B), uses other agents' behavior to better adhere to road areas (A, C) without environment data, and predicts based on the speed changes of other cars (D)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.184, + 0.743 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We train AR and 3-agent PAR models using velocity tokens, acceleration tokens, and acceleration tokens combined with our location positional encoding. The results can be seen in Table 3. Note that the 3-agent PAR model uses the agent ID embedding and next timestep prediction. Acceleration tokens consistently outperform velocity tokens both for agent AR and 3-agent PAR models. This could be because the vocabulary size for acceleration tokens is much smaller and therefore easier to optimize. Regardless, both ways of tokenizing result in models that outperform our baselines (see" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.703, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.177, + 0.646, + 0.19 + ], + "angle": 0, + "content": "1-Agent AR" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.089, + 0.905, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.177, + 0.852, + 0.19 + ], + "angle": 0, + "content": "2-Agent PAR" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.269 + ], + "angle": 0, + "content": "Figure 7. Rotation forecasting qualitative result on test set. 3D predictions are projected onto the image, isolating rotation results by showing the ground-truth translation. Incorporating the hand agent in the PAR framework (right) improves object pose prediction over object-only AR (left)." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.703, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.38, + 0.646, + 0.393 + ], + "angle": 0, + "content": "1-Agent AR" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.291, + 0.905, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.38, + 0.852, + 0.393 + ], + "angle": 0, + "content": "2-Agent PAR" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.402, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 8. Translation forecasting qualitative result on test set. 3D predictions are projected onto the image, isolating translation results by showing the ground-truth rotation. Using the PAR framework (right) instead of AR (left) improves object pose prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.626 + ], + "angle": 0, + "content": "Table 4 - NN has a relatively low error on this dataset), and highlight that our framework is flexible such that a user can experiment with different ways of representing entities. For both token types, the 3-agent PAR model that is blind to location outperforms the AR model. While location information should help the model, it is possible that simply knowing whether other agents are slowing down or accelerating can help the model make better predictions. When adding location information via the LPE to our 3-agent PAR model, we see another performance gain in ADE and FDE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.907, + 0.736 + ], + "angle": 0, + "content": "Qualitative examples of the AR model (top row) and 3-agent location-aware PAR model (bottom row) can be seen in Figure 6. Our method uses no image or environment data (e.g., lanes) as input. However, by reasoning over multiple agents, its predictions lead to fewer collisions and better reasoning about speed changes and driveable areas based solely on other agents' behaviors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.763, + 0.908, + 0.799 + ], + "angle": 0, + "content": "6. Case Study 3: Object Pose Forecasting During Hand-Object Interaction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our final case study explores how hand-object interaction can be leveraged for object pose estimation. We define the human hand and the interacting object as two agents, with tokens representing distinct state types. We show that our PAR framework can jointly model these agents, improving 3D translation and rotation predictions for the object." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.918, + 0.519, + 0.929 + ], + "angle": 0, + "content": "12408" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.281, + 0.108 + ], + "angle": 0, + "content": "6.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.25 + ], + "angle": 0, + "content": "Dataset. We use the DexYCB dataset, which includes 1000 videos of 10 subjects performing object manipulation tasks with 20 distinct objects from the YCB-Video dataset. The data is split into 800 training, 40 validation, and 160 testing videos. We use one of 8 provided camera views. In each trial, subjects pick up and lift objects in randomized conditions. Labels include the object's SO(3) rotation and 3D translation, and the hand's 3D translation. We focus on predicting the object's rotation or translation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.484, + 0.472 + ], + "angle": 0, + "content": "Task-specific considerations. We tokenize object information in object-only experiments and both object and hand information in hand-object experiments. The object is represented as a 4D token for rotation forecasting (quaternion from SO(3) rotation) or a 3D token for translation forecasting (Euclidean coordinates). In hand-object experiments, the hand token is included with a 3D translation vector, and agent ID embeddings distinguish between the hand and object. Normalization is applied to all 3D translation vectors in both AR and PAR experiments; quaternions are normalized by definition and require no additional processing. An embedding layer projects the tokens into the transformer's hidden dimension, and another layer projects them back for prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.732 + ], + "angle": 0, + "content": "For rotation-only forecasting, the loss is \\(\\mathcal{L}_{rot} = 1 - |\\hat{q}\\cdot q|\\), where \\(\\hat{q}\\) is the predicted quaternion and \\(q\\) the ground-truth quaternion. For translation-only forecasting, the loss \\(\\mathcal{L}_t\\) is the mean squared error (MSE) between predicted and ground-truth translations. For PAR we predict relative object-to-hand translations at each frame, using the current hand position as origin, while for AR, we predict absolute object translations without considering the interacting agent. For PAR models, we add the loss \\(\\mathcal{L}_h\\), a MSE on hand translation. The object-only AR rotation model is optimized with \\(\\mathcal{L}_{rot}\\), while the PAR rotation model combines \\(\\mathcal{L}_{rot} + \\mathcal{L}_h\\); similarly, the object-only translation model is trained with \\(\\mathcal{L}_t\\), and the hand-object translation model uses \\(\\mathcal{L}_t + \\mathcal{L}_h\\). At inference, the first half of each video is provided, and object predictions are autoregressively generated for the second half. Translation is evaluated using MSE, while rotation is measured using geodesic distance (GEO) on SO(3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.185, + 0.756 + ], + "angle": 0, + "content": "6.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We compare the object-only AR models to the hand-object PAR models in Table 5 for the two prediction tasks. We also present the baselines described in Sec. 3.3 in Table 6. Figures 7 and 8 show qualitative results on the rotation and translation predictions, respectively. In both prediction tasks, we observe that incorporating the human hand's interaction with the object enhances accuracy: for rotation, PAR results in a relative improvement of \\(8.9\\%\\) over AR, and for translation, \\(41\\%\\). See Section 11.2 for additional qualitative results" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.909, + 0.188 + ], + "angle": 0, + "content": "
TypeMethodMSE ↓GEO (rad) ↓
Translation1-agent AR3.68 × 10-3-
Translation2-agent PAR2.17 × 10-3-
Rotation1-agent AR-0.919
Rotation2-agent PAR-0.837
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.198, + 0.907, + 0.24 + ], + "angle": 0, + "content": "Table 5. Test set results on DexYCB dataset. For both rotation and translation forecasting, the 2-agent PAR model, which treats the hand as an additional agent, improves results." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.254, + 0.924, + 0.343 + ], + "angle": 0, + "content": "
BaselineTranslation - MSE (m2) ↓Rotation - GEO (rad) ↓
Random0.2442.196
Random Trajectory1.60 × 10-22.146
NN1.69 × 10-22.179
Multiagent NN1.71 × 10-22.170
Mirror1.20 × 10-2-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.354, + 0.907, + 0.396 + ], + "angle": 0, + "content": "Table 6. Test set results for DexYCB baselines. We cannot provide rotation results for the Mirror baseline, because the ground-truth does not include hand rotation, only 3D translation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.423, + 0.695, + 0.437 + ], + "angle": 0, + "content": "with more sampled frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.452, + 0.627, + 0.468 + ], + "angle": 0, + "content": "7. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.598 + ], + "angle": 0, + "content": "This work introduced the Poly-Autoregressive (PAR) framework, a unifying approach to prediction for multi-agent interactions. By applying the same transformer architecture and hyperparameters across diverse tasks, including action forecasting in social settings, trajectory prediction for autonomous vehicles, and object pose forecasting during hand-object interaction, we have demonstrated the versatility and robustness of our framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Our findings underscore the crucial importance of considering the influence of multiple agents in a scene for prediction tasks. By modeling interactions, we significantly improved prediction accuracy over single-agent approaches on all three problems we considered. While we achieved promising results with a simple architecture, we have only provided a starting point that can be built upon extensively. For instance, incorporating environmental context or tokenizing pixel patches, especially as a way to relax our assumption on high-quality tracking, are avenues for further research using PAR. It would be interesting to experiment with scaling the data and model. We do not specifically consider the relative importance of neighboring agents, which is an interesting future research direction (ex. dynamic attention mechanism)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Our PAR framework provides a simple and generalizable foundation of universal building blocks, ready for extension or refinement in future tasks. The PAR framework holds potential for advancing AI systems, enhancing prediction capabilities and enabling more accurate navigation and operation in real-world, multi-agent interactions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.918, + 0.519, + 0.929 + ], + "angle": 0, + "content": "12409" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "8. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.177 + ], + "angle": 0, + "content": "We thank Jane Wu, Himanshu Singh, Georgios Pavlakos, and João Carreira for useful discussions and feedback. This work was supported by ONR MURI N00014-21-1-2801 and NSF Graduate Fellowships to NT and TS." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.189, + 0.188, + 0.205 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.486, + 0.296 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.297, + 0.485, + 0.327 + ], + "angle": 0, + "content": "[2] Fred Attneave. Some informational aspects of visual perception. Psychological review, 61(3):183, 1954. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.327, + 0.485, + 0.395 + ], + "angle": 0, + "content": "[3] Murchana Baruah and Bonny Banerjee. A multimodal predictive agent model for human interaction generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 1022-1023, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.485, + 0.481 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.482, + 0.485, + 0.535 + ], + "angle": 0, + "content": "[6] Sergio Casas, Wenjie Luo, and Raquel Urtasun. Intentnet: Learning to predict intention from raw sensor data. In Conference on Robot Learning, pages 947-956. PMLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.486, + 0.62 + ], + "angle": 0, + "content": "[7] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.486, + 0.704 + ], + "angle": 0, + "content": "[8] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.485, + 0.761 + ], + "angle": 0, + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[10] Baptiste Chopin, Hao Tang, Naima Otberdout, Mohamed Daoudi, and Nicu Sebe. Interaction transformer for human reaction generation. IEEE Transactions on Multimedia, pages 1-13, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.485, + 0.9 + ], + "angle": 0, + "content": "[11] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, KueiDa Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[12] Henggang Cui, Vladan Radosavljevic, Fang-Chieh Chou, Tsung-Han Lin, Thi Nguyen, Tzu-Kuo Huang, Jeff Schneider, and Nemanja Djuric. Multimodal trajectory predictions for autonomous driving using deep convolutional networks. In 2019 international conference on robotics and automation (icra), pages 2090-2096. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[13] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.907, + 0.273 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.909, + 0.344 + ], + "angle": 0, + "content": "[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.908, + 0.468 + ], + "angle": 0, + "content": "[17] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. In European Conference on Computer Vision (ECCV), page 3, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.908, + 0.524 + ], + "angle": 0, + "content": "[18] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In International Conference on Machine Learning, pages 1242-1250. PMLR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.908, + 0.621 + ], + "angle": 0, + "content": "[19] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6047-6056, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[20] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13053-13064, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.907, + 0.803 + ], + "angle": 0, + "content": "[22] Yanjun Huang, Jiatong Du, Ziru Yang, Zewei Zhou, Lin Zhang, and Hong Chen. A survey on trajectory-prediction methods for autonomous driving. IEEE Transactions on Intelligent Vehicles, 7(3):652-674, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.908, + 0.872 + ], + "angle": 0, + "content": "[23] Boris Ivanovic, Guanyu Song, Igor Gilitschenski, and Marco Pavone. trajdata: A unified interface to multiple human trajectory datasets. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, New Orleans, USA, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[24] Bolin Lai, Sam Toyer, Tushar Nagarajan, Rohit Girdhar, Shengxin Zha, James M Rehg, Kris Kitani, Kristen Grau" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.918, + 0.521, + 0.93 + ], + "angle": 0, + "content": "12410" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "man, Ruta Desai, and Miao Liu. Human action anticipation: A survey. arXiv preprint arXiv:2410.14045, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.188 + ], + "angle": 0, + "content": "[25] Hugo Larochelle and Iain Murray. The neural autoregressive distribution estimator. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 29-37. JMLR Workshop and Conference Proceedings, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[26] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.484, + 0.301 + ], + "angle": 0, + "content": "[27] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[29] Siyuan Brandon Loh, Debaditya Roy, and Basura Fernando. Long-term action forecasting using multi-headed attention-based variational recurrent neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2419-2427, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[30] Vongani Maluleke, Lea Müller, Jathushan Rajasegaran, Georgios Pavlakos, Shiry Ginosar, Angjoo Kanazawa, and Jitendra Malik. Synergy and synchrony in couple dances. arXiv preprint arXiv:2409.04440, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.484, + 0.524 + ], + "angle": 0, + "content": "[31] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[32] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20395-20405, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[33] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[34] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In ArXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[35] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[36] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.859 + ], + "angle": 0, + "content": "[37] Alec Radford. Improving language understanding by generative pre-training. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[39] Ilija Radosavovic, Bike Zhang, Baifeng Shi, Jathushan Rajasegaran, Sarthak Kamat, Trevor Darrell, Koushil Sreenath, and Jitendra Malik. Humanoid locomotion as next token prediction. arXiv preprint arXiv:2402.19469, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.907, + 0.231 + ], + "angle": 0, + "content": "[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, Christoph Feichtenhofer, and Jitendra Malik. On the benefits of 3d pose and tracking for human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[41] Jathushan Rajasegaran, Ilija Radosavovic, Rahul Ravishankar, Yossi Gandelsman, Christoph Feichtenhofer, and Jitendra Malik. An empirical study of autoregressive pre-training from videos. arXiv preprint arXiv:2501.05453, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[42] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.907, + 0.455 + ], + "angle": 0, + "content": "[43] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.907, + 0.537 + ], + "angle": 0, + "content": "[44] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8579-8590, 2023. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.538, + 0.907, + 0.566 + ], + "angle": 0, + "content": "[45] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 30(1):50-64, 1951. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.907, + 0.622 + ], + "angle": 0, + "content": "[46] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. arXiv preprint arXiv:2403.18811, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.907, + 0.676 + ], + "angle": 0, + "content": "[47] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[48] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Rahul Sukthankar, Kevin Murphy, and Cordelia Schmid. Relational action forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 273-283, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[49] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.873 + ], + "angle": 0, + "content": "[50] Lucas Theis and Matthias Bethge. Generative image modeling using spatial lstms. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.918, + 0.518, + 0.929 + ], + "angle": 0, + "content": "12411" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[52] Ameni Trabelsi, Mohamed Chaabane, Nathaniel Blanchard, and Ross Beveridge. A pose proposal and refinement network for better 6d object pose estimation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2382-2391, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[53] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.276, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[54] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[55] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16611–16621, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[56] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.484, + 0.529 + ], + "angle": 0, + "content": "[57] Bowen Wen, Jonathan Tremblay, Valts Blukis, Stephen Tyree, Thomas Müller, Alex Evans, Dieter Fox, Jan Kautz, and Stan Birchfield. Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 606-617, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.53, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[58] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868-17879, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.484, + 0.641 + ], + "angle": 0, + "content": "[59] Jane Wu, Georgios Pavlakos, Georgia Gkioxari, and Jitendra Malik. Reconstructing hand-held objects in 3d. arXiv preprint arXiv:2404.06507, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.484, + 0.698 + ], + "angle": 0, + "content": "[60] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.484, + 0.769 + ], + "angle": 0, + "content": "[61] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.918, + 0.52, + 0.93 + ], + "angle": 0, + "content": "12412" + } + ] +] \ No newline at end of file diff --git a/2025/Poly-Autoregressive Prediction for Modeling Interactions/43d051a6-6821-4e3b-bd9a-a4c244857daa_origin.pdf b/2025/Poly-Autoregressive Prediction for Modeling Interactions/43d051a6-6821-4e3b-bd9a-a4c244857daa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b5e51fb73eec9949453b354b5ed38fb3aed5e043 --- /dev/null +++ b/2025/Poly-Autoregressive Prediction for Modeling Interactions/43d051a6-6821-4e3b-bd9a-a4c244857daa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25fe1f4dc146a3372ae7d9c118f5a4ed366018af7b100e8b419e2220007706b5 +size 4129548 diff --git a/2025/Poly-Autoregressive Prediction for Modeling Interactions/full.md b/2025/Poly-Autoregressive Prediction for Modeling Interactions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2122d5e02e69efc43691d4dc9818ad7372c8d68b --- /dev/null +++ b/2025/Poly-Autoregressive Prediction for Modeling Interactions/full.md @@ -0,0 +1,327 @@ +# Poly-Autoregressive Prediction for Modeling Interactions + +Neerja Thakkar $^{1}$ Tara Sadjadpour $^{1}$ Jathushan Rajasegeran $^{1}$ Shiry Ginosar $^{2,3}$ Jitendra Malik $^{1}$ $^{1}$ UC Berkeley, $^{2}$ Toyota Technical Institute at Chicago, $^{3}$ Google DeepMind + +![](images/709f2fc30bfc60bc3d43308b1860f77821a18b7732f0f7e89851a55cdf7275f8.jpg) +(a) Autoregression (AR) + +![](images/76fb006702cb31067a0dcf2db0afe7ff8004ce2ee609bee1bce0694edc3cb479.jpg) +(b) Poly-autoregression (PAR) +Figure 1. Inference for (a) autoregressive (AR) models and (b) our proposed poly-autoregressive (PAR) model. Solid indicates ground-truth tokens which represent a tracked data modality such as action or 6DOF pose; striped represents predicted output tokens. Color denotes agent identity. Compared to AR models, the PAR model takes other agents' tokens as inputs when making a prediction for the next timestep. + +# Abstract + +We introduce a simple framework for predicting the behavior of an agent in multi-agent settings. In contrast to autoregressive (AR) tasks, such as language processing, our focus is on scenarios with multiple agents whose interactions are shaped by physical constraints and internal motivations. To this end, we propose Poly-Autoregressive (PAR) modeling, which forecasts an ego agent's future behavior by reasoning about the ego agent's state history and the past and current states of other interacting agents. At its core, PAR represents the behavior of all agents as a sequence of tokens, each representing an agent's state at a specific timestep. With minimal data pre-processing changes, we show that PAR can be applied to three different problems: human action forecasting in social situations, trajectory prediction for autonomous vehicles, and object pose forecasting during hand-object interaction. Using a small proof-of-concept transformer backbone, PAR outperforms AR across these three scenarios. + +# 1. Introduction + +Large language models (LLMs) have been very successful with natural language processing (NLP) tasks, which require accurate reasoning over relationships words have within a body of text. A key component of LLMs is autoregressive + +(AR) modeling, where each word token is predicted based on a sequence of preceding word tokens. Building on the success of AR modeling in the NLP community, this work focuses on modeling the dynamic relationships between agents and entities in everyday interactions that occur in the physical world, such as social interactions, driving, and hand-object interactions. + +Unlike language which is structured through grammar and semantics, interactions in the physical world are dictated by both the laws of physics (e.g. how a hand grasps an object) and the internal state of each agent (e.g. the trajectory an agent chooses to move a grasped object in), a latent variable that we know nothing about. Furthermore, as opposed to text where each word follows another unidimensionally, the states of multiple agents are changing simultaneously. For example, in social situations, the history of a single person's past states does not alone determine the dynamics of their future states; we also need to consider the states of other agents. We argue, therefore, that AR modeling alone is insufficient. + +In this paper, we introduce poly-autoregressive (PAR) modeling, a simple unifying approach to model the influence of other agents and entities on one's behavior. We model behavior as a temporal sequence of states and predict an ego agent's future behavior conditioned on the history of behavior of the ego agent as well as the rest of the agents. By considering other agents' behaviors, we demonstrate that our approach significantly improves upon the ill-posed problem + +of single-agent prediction in interactive settings. + +The PAR framework uses a transformer-based model for next-token prediction. Transformers have shown great success in language modeling and naturally lend themselves to predicting behavior over time. In an interaction scenario of $N$ agents, our model predicts the future behavior of an ego $(N^{\mathrm{th}})$ agent conditioned on its past behavior and the behavior of the other $N - 1$ non-ego agents (see Figure 1b). Each prediction task may model a different behavior modality of interest, e.g. actions for social action prediction, or 6DoF pose for object forecasting in hand-object interactions. We apply PAR to three different case studies of common real-world interactions: + +1. Social action prediction. We test our method on the AVA benchmark [19] for action forecasting. By incorporating both the ego and another agent using PAR, we get a overall $+1.9$ absolute mAP gain over AR, which only models the ego agent to predict its future behavior, and an absolute $+3.5$ mAP gain on 2-person interaction classes. +2. Trajectory prediction for autonomous vehicles. When forecasting future $xy$ locations of an ego vehicle, incorporating the locations of neighboring vehicles with PAR outperforms AR, which only uses the ego vehicle's preceding trajectory as input. Specifically, on the nuScenes dataset [5], PAR outperforms AR with a relative improvement of $6.3\%$ ADE and $6.4\%$ FDE. +3. 6DoF object pose forecasting during hand-object interaction. We use the DexYCB dataset [8], where we treat the object as the ego agent and the hand as the interacting agent. While PAR integrates that hand's 3D location and object pose history, AR only uses the object's pose history to predict the object's 6DoF pose. PAR outperforms AR with relative improvements of $8.9\%$ and $41.0\%$ for the rotation and translation predictions, respectively. + +In all these settings, we find that incorporating the behavior of other agents in the scene improves predictions of the ego agent's behavior. All of these problems are modeled via the same simple PAR framework and implemented using the same proof-of-concept 4 million parameter transformer without any modifications to the base framework or architecture, only to data pre-processing and choice of tokenization. We also provide an example of a simple way to build on our architecture through a location positional encoding, Sec. 5. + +The primary contribution of this work is a versatile framework that can be applied to a diverse range of settings, without modifications aside from domain-specific data processing. Our results suggest that PAR provides a simple formulation that, with a more complex transformer backbone and larger datasets, could enhance prediction of diverse multi-agent interactions across various problem domains. To facilitate further exploration and development, we have released our code, which contains the building blocks to use PAR for modeling other types of multi-agent interactions. + +# 2. Related Work + +Autoregressive models. Autoregressive modeling has a rich history in information theory and deep learning, tracing back to Shannon's paper on language prediction [45] and Attneave's study on visual perception [2]. These foundational works laid the groundwork for modern applications in deep learning, including [25], which revisits neural autoregressive models and [18, 50], which explore continuous-valued modeling. [53] developed PixelRNN and PixelCNN to autoregressively generate images one pixel at a time using RNN and CNN architectures, respectively. + +The development of the transformer model [54] spurred progress in computer vision with the image transformer [36] and the vision transformer [14], and autoregressive models [9, 41] and more notably in NLP, where the GPT family of models [4, 37, 38] has demonstrated the power of large-scale unsupervised autoregressive pre-training. Recent research has focused on multimodal learning, exemplified by the Flamingo [1] or LlaVa [28] models, which combine vision and language processing capabilities, illustrating the versatility of autoregressive models across various domains in artificial intelligence. While these approaches operate on image patches and word tokens, we operate on symbolic representations extracted from in-the-wild videos showing natural interactions. A recent approach [39] frames humanoid locomotion as an autoregressive next-token prediction task that operates on two types of continuous tokens: observations and actions. This approach projects continuous tokens to the hidden dimension and uses a shifted loss, similar to the next-timestep prediction proposed in our framework. + +Multi-agent regressive models. Several prior works addressed modeling specific multi-agent problems via regressive models as one-off case studies. We introduce the PAR framework to unify these efforts into a single cohesive framework. Many behavior prediction works focus on two agents engaging in social interaction, whether it be dyadic communication [32-34] or social dance [30, 46]. These studies primarily tackle the challenge of predicting the state of an interacting partner (Person B) based on the input from Person A's state, sometimes extending predictions into the future [20, 30]. While earlier works used architectures such as variational RNNs [3], recent works have predominantly adopted transformer architectures for social interaction modeling [10, 20, 32, 33, 46], with some works exploring diffusion [27], or diffusion with attention [17]. Our PAR framework focuses on transformer models. + +Works encompassed by the PAR framework extend beyond human social interaction. Many multi-agent human or car trajectory prediction approaches use autoregressive prediction. For instance, MotionLM [44] utilizes a transformer decoder that processes multi-agent tokens, incorporating a + +learned agent ID embedding. This methodology informs our approach across all our case studies. Critically, in contrast to all prior multi-agent regressive works that designed solutions to address specific applications, we demonstrate that we can unify a diverse set of multi-agent regressive problems under a single PAR framework. See Appendix Sec. 9 for case-study-specific works. + +# 3. Poly-Autoregressive Modeling + +Our goal is to model the behavior of an agent or entity while taking into account any other agents it interacts with, if any. To evaluate the performance of our model in capturing interaction dynamics, we predict the agent's future behavior and compare it against ground-truth data. + +We define the following task: In an interaction comprised of $N$ agents, given the observed past states of the $N - 1$ interacting agents, and the observed or previously-predicted past states of the $N^{th}$ ego agent, predict the future states of the $N^{th}$ ego agent. + +We define a transformer-based poly-autoregressive (PAR) predictor, $\mathcal{P}$ , that learns to model temporally long-range interactions in the input sequence. The inputs to the predictor are the past states of the $N$ interacting agents, and its output is the predicted future state of the $N^{\mathrm{th}}$ ego agent. + +# 3.1. Problem Definition + +Let $\mathbf{S} = \{\mathbf{s}_i\}_{i=1}^T$ be a temporal sequence of agent states, $\mathbf{s}_i$ . We use $\mathbf{S}^N$ and $\mathbf{S}^{1:N-1}$ to denote the temporal sequences of states of the $N_{th}$ agent and of the other $N-1$ agents, respectively. For each timestep $t \in [t_{\pi}, T]$ , where $t_{\pi} \in [1, T]$ is the time we start predicting, we take as input all other $N-1$ agents' past observed state sequences $\mathbf{S}_{1:t-1}^{1:N-1}$ along with the $N_{th}$ agent's past observed states up to $t_{\pi}$ , $\mathbf{S}_{1:t_{\pi}}^N$ , and any of its previously predicted past states $\hat{\mathbf{S}}_{t_{\pi}+1:t-1}^N$ , if available (see Fig. 1). Our predictor, $\mathcal{P}$ , then poly-autoregressively predicts the $N_{th}$ agent's future states one time-step at a time: + +$$ +\hat {\mathbf {s}} _ {t} ^ {N} = \mathcal {P} \left(\mathbf {S} _ {1: t - 1} ^ {1: N - 1}, \mathbf {S} _ {1: t _ {\pi}} ^ {N}, \hat {\mathbf {S}} _ {t _ {\pi} + 1: t - 1} ^ {N}\right). \tag {1} +$$ + +$\mathcal{P}$ learns to model the distribution over the next timestep of the $N_{th}$ agent's states, given all other agents' states: + +$$ +p \left(\hat {\mathbf {s}} _ {t} ^ {N} \mid \mathbf {S} _ {1: t - 1} ^ {1: N - 1}, \mathbf {S} _ {1: t - 1} ^ {N}\right). \tag {2} +$$ + +While we provide the observed ground truth states of other agents at inference, during training, we jointly maximize the likelihood of all $N$ agents by computing losses on their future state predictions. + +We train the predictor by maximizing the likelihood of the target state $y$ at time $t$ : + +$$ +\mathcal {L} _ {\mathcal {P}} = E _ {y \sim p (y)} [ - \log (p (\mathbf {s} _ {t} ^ {N}) ], +$$ + +where the target state $y$ at $t$ is computed from the $N_{th}$ agent ground truth future state. + +# 3.2. The Poly-Autoregressive Framework + +We address the problem of forecasting the future states of an agent (from time $t$ to $T$ ) in a data-driven way, given a temporal sequence of past states (from time 1 to $t - 1$ ). We assume that our agent has some feature, or a set of features, of interest in a video (e.g., 3D pose) that we can tokenize. We predict the future states of the agent in terms of this tokenized feature (or set of), where we use one token (or set of tokens) per time step. The predicted tokens can be discrete (i.e., an index into a feature codebook) or continuous (i.e., a vector of one or more continuous values). The loss $\ell$ will depend on the problem's specifics and the type of token used. To train the model to predict the future, we rely on all the interaction dynamics of length $T$ in our training dataset as ground truth examples. + +As a baseline, we consider the single-agent autoregressive (AR) paradigm, where a transformer is trained to perform next-token-prediction with teacher forcing. AR uses greedy sampling to generate sequences at inference time, predicting one next token at a time (Fig. 1(a)). + +In contrast, our multi-agent poly-autoregressive (PAR) framework considers the other $N - 1$ agents in the scene when predicting the future state of the $N$ th agent. In this setup, we utilize the features of interest of all $N$ agents, yielding $N$ tokens at each timestep for a total of $N * T$ tokens. In practice, we operate on a flattened sequence of $N * T$ tokens. Instead of using the AR training procedure in this multi-agent case (as in Fig. 3a), we jointly model the $N$ agents at each timestep by introducing the following features to our PAR framework. + +Next-timestep prediction. A standard AR model predicts the next token. Given the flattened sequence of $N * T$ tokens our model operates on, next token prediction would take as input an agent $k$ at timestep $t$ and predict agent $k + 1$ 's state at the same timestep $t$ (as in Fig. 3a). However, our goal is to predict the input agent $k$ 's future state at time $t + 1$ . Therefore, we perform same-agent next-timestep prediction rather than next-token prediction (see Fig. 3b for an illustration of same-agent next-timestep at training). + +Learned agent identity embedding. When giving a model information corresponding to multiple agents, the model can benefit from knowing which token corresponds to which agent. We give the model this information with a learned agent ID embedding. + +Joint training. We train the model to jointly predict the future of all agents by computing a loss on the predicted tokens of all agents (Fig. 3b). Please refer to Section 3.1 for our inference paradigm. + +# 3.3. Task-Specific Considerations + +Our simple PAR approach unifies diverse problems under a single framework and architecture without any modifications. + +![](images/82a17ddfb9150be20f986188f09277d60cac11c0377ae2ffa992a7c4207655f2.jpg) +Figure 2. The PAR Framework. We begin by collecting a video dataset, such as AVA (top) or DexYCB (bottom). Then, using dataset labels or computer vision techniques, a trajectory of a given modality for our prediction task is extracted for each agent, such as action class labels (top) or object pose and 3D hand translation (bottom). Data is then tokenized, either through discretization or directly using continuous values, with our framework supporting both formats. Based on the tokenization and prediction task, we choose the appropriate loss function for PAR training. After training with PAR, predicted tokens can be decoded back to data space and evaluated with relevant metrics. + +![](images/2b8c56e25ce69fce0b0be688f7fe04386f70a084587cfb72faf67a20e577e2e7.jpg) +(a) AR: multi-agent, next-token training. + +![](images/ec2468950fe8402fa30610d7862e0da7ed3da055861f2cfd5821056d079c9eef.jpg) +(b) PAR: same-agent, next-timestep training. +Figure 3. Training with teacher forcing for (a) multi-agent next-token prediction in autoregressive models and (b) multi-agent poly-autoregressive models. Solid vs striped indicates a ground-truth vs predicted token, respectively. Color denotes agent identity. The AR model is trained for next-token prediction, while the PAR model is trained to predict the next timestep of the same agent. Three agents are shown for ease of visualization, but the PAR model supports an arbitrary number of agents. + +In order to formulate a problem as interaction-conditioned prediction, users must consider several task-specific details. + +Fig. 2 gives an overview of how the PAR framework disentangles multi-agent learning from problem-specific modeling. + +Data. The input data source in our example tasks is always a collection of videos. From these videos, we extract various modalities relevant to the task at hand. These modalities can range from high-level features, such as action class labels, to + +low-level ones, such as 3D pose (Fig. 2 first two columns). We assume that each agent in the dataset is detected at each frame and is associated with an agent ID. + +Tokenization. Our framework supports both discrete, quantized tokens and continuous vector tokens. The choice between discrete and continuous depends on the nature of the task. In the case of discrete tokens, we use a standard embedding layer to project to the hidden dimension. For continuous tokens, we train a projection layer to project the token into the hidden dimension of the transformer. For instance, if our continuous token is a 3D vector with an $(x,y,z)$ 3D location coordinate and our hidden dimension is 128, our projection layer will project from 3 to 128 dimensions. We also train an un-projection layer that reverts the hidden dimension to the original token dimension. + +Loss. The type of token and task-specific considerations dictate the loss function $\ell$ applied during model training. For discrete tokens, a classification loss is appropriate. For continuous tokens, we use a regression loss on the original token dimension. + +Baselines. We compare to the following baselines, where applicable on a case-by-case basis: + +- Random token: pick random tokens from the best available token space and use as the prediction. +- Random trajectory: pick at random a trajectory from the training dataset to use as the prediction. +- NN: Given an input agent $A$ 's trajectory history, find the closest trajectory to it in the training set, belonging to $A^T$ . Use $A^T$ 's future as the predicted future. +- Multiagent NN: In a dataset with two interacting partners $A$ and $B$ , where $B$ is the ego agent, given an input agent + +$A$ 's trajectory history, find the closest trajectory to it in the training set, belonging to $A^T$ . Use $A^T$ 's interaction partner's $B^T$ 's future as the prediction. + +- Mirror: In a dataset with two interacting partners $A$ and $B$ , use the ground truth future of agent $B$ as the predicted future for agent $A$ . + +# 3.4. Framework Implementation Details + +We keep the following implementation details constant for all case studies (see also Sec. 10). + +Learned agent ID embedding. Our learned agent ID embedding consists of the integer agent ID mapped to a hidden dim-sized vector, and summed to the token embedding. + +Architecture. For all case studies, we use the Llama [51] transformer decoder architecture with 8 layers, 8 attention heads, and a hidden and intermediate dimension of 128. The decoder has $\sim 4.4\mathrm{M}$ learned parameters, not including learned embedding layers which add a few thousand more parameters. A rotary positional encoding [47] is used in addition to other summed encodings (i.e. agent ID embedding, locational positional encoding in Sec. 5). We train using teacher forcing. The only hyperparameter that changes between case studies is the learning rate. + +# 4. Case Study 1: Social Action Forecasting + +Our first case study involves forecasting human actions. Human behaviors are fundamentally social; for instance, individuals frequently walk in groups and alternate between speaking and listening roles when conversing. Certain actions, like hugging or handshaking, are intrinsically multiperson. Therefore, modeling human interactions should help improve action forecasting performance, especially on multiperson actions, which we show in this case study. + +# 4.1. Experimental Setup + +Dataset. The Atomic Visual Actions (AVA) dataset [19] comprises 235 training and 64 15-minute validation videos from movies. Annotations are provided at a 1Hz frequency, detailing bounding boxes and tracks for individuals within the frame, and each person's actions within a 1-second timeframe. Individuals may engage in multiple concurrent actions from a repertoire of 60 distinct action classes (e.g., sitting and talking simultaneously). For our analysis, we select clips featuring a continuous sequence of an agent's actions spanning at least 4s, splitting sequences exceeding 12s. We use the first half of each clip as history to predict the second half. For any ego agent trajectory, we pick a second agent by selecting the person present in the scene for the longest subset of the ego agent's trajectory. + +Task-specific considerations. Each agent's token $\mathcal{A}$ represents an 60-dimensional vector that corresponds to the actions performed at a specific timestep. Each element denotes + +
MethodTimestep predAg-ID embdmAP ↑
1-agent ARN/AN/A40.7
2-agent ARXX38.0
2-agent PAR*X✓40.2
2-agent PAR*✓X40.0
2-agent PAR✓✓42.6
+ +Table 1. PAR action forecasting performance on AVA We evaluate 1 and 2-agent AR methods, two 2-agent PAR ablations (rows 3 and 4, $\mathrm{PAR^{*}}$ ), and our PAR method. Without next-timestep prediction (see Fig. 3) or a learned agent ID embedding, our model struggles with multi-agent reasoning, performing worse than the AR baseline. With both components, the 2-agent PAR model achieves a $+1.9$ mAP gain over the AR method (see Fig. 11 and Fig. 5 for class breakdown). + +
BaselineAgentsmAP ↑
Random Token13.46
Random Training Traj13.44
Nearest Neighbor113.17
Multiagent NN25.10
Mirror27.97
+ +Table 2. AVA baselines While the nearest neighbor baseline performs best among baselines, it is still significantly worse than the AR model. + +the probability of a particular action class being enacted; ground-truth inputs are a binary vector. We implement an embedding layer that projects these tokens into the transformer's hidden dimension, as well as an un-projection layer that reverts them back to the original 60D token space for the purposes of loss calculation and output generation. We do not explicitly require the outputs to be values between 0 and 1. We use a MSE regression loss on the 60D action tokens: $\mathcal{L} = \frac{1}{n}\sum_{i=1}^{n}(\mathcal{A}_i - \hat{\mathcal{A}}_i)^2$ . Our evaluation metric is the mean average precision (mAP) on the 60 AVA classes. + +We implement all baselines described in 3.3, where Random Token corresponds to a random 60D vector sampled from 0 to 1. NN and Multiagent NN use Hamming distance as the distance metric. + +# 4.2. Results + +We report the performance of a single-agent AR model as a baseline, in the first line of Table 1. The AR model is significantly better than our baselines (see Table 2), the strongest baseline being the single-agent NN. We compare these baselines to our 2-agent PAR model (last line) and various ablations where we remove the agent ID embedding and perform next-token rather than same-agent next-timestep prediction. The second line of the table corresponds to multi-agent next-token prediction (Fig. 3a). We see that this approach confuses the model, and the performance is significantly worse than just training on and considering a single agent. However, + +![](images/f33dc690baa1e6313f48ddb1725c5177b489af2d575f1db9777a17af674f3a8e.jpg) +Figure 4. Action forecasting example. The distribution over ground truth actions are in white, and our predictions in red. A 6s action history (1Hz) is input, and 6s of future actions predicted. In the scene, the man and woman alternate between talking and listening. Initially, the man is talking. The AR model predicts he will continue talking, while the 2-agent PAR model recognizes the woman is talking and predicts more accurate turn-taking behavior. + +![](images/9f4fb74698e8f8f190ab3c6ab86b66b3b909184799962d45dbcdabfa4e945f10.jpg) +Figure 5. Per-class mAP for AVA 2-person actions. We see performance improvement on almost all 2-person AVA action classes ((P) stands for "a person"). Some absolute mAP gains are particularly significant: listen to $+7.0$ , kiss $+8.3$ , fight/hit $+5.7$ , talk to $+4.4$ , hug $+5.7$ , and hand shake $+4.0$ . + +as we add various components of our PAR approach, the performance improves, and with both the next timestep prediction and agent ID embedding, we get a $+1.9$ mAP gain. When only considering 2-person action classes (enumerated in Fig. 5), our mAP is 36.3 on the single agent PAR model and 39.8 on the 2-agent PAR model, a $+3.5$ mAP gain. + +In Fig. 4 we see an example of action forecasting. In the input history, the man talks and the woman listens. In the future, the woman talks, and the man listens. Our 2-agent PAR model (bottom row) better understands that talking and listening actions are complementary actions, while the AR model doesn't learn this correlation. We see quantitative evidence of this in Fig. 5, with per-class mAPs for our AR vs 2-agent PAR model for 2-person action classes. Here, talk to gets a $+4.4$ mAP gain and listen to gets a $+7.0$ mAP gain when we train a multi-agent model. We see a significant boost on many other interaction-related action classes—for instance, kiss a person $+8.3$ and fight/hit a person $+5.7$ mAP—and on single-person actions, see Fig. 11. + +# 5. Case Study 2: Multiagent Car Trajectory Prediction + +Our second case study focuses on predicting car trajectories. Trajectory prediction requires a vehicle to be aware of other cars on the road to avoid collisions and promote cooperative behavior. This study demonstrates how our framework enables the joint modeling of multiple vehicles' movements. + +# 5.1. Experimental Setup + +Dataset. We use nuScenes [5], inputting 2 seconds of positions to forecast vehicle positions 6 seconds ahead. Specifically, our objective is to predict the $xy$ coordinates of each agent, exclusively considering vehicles as agents. We use the trajdata interface [23] to load and visualize the data. + +Task-specific considerations. Instead of discretizing the $xy$ position space, we discretize the motion, resulting in discrete velocity or acceleration tokens. These integer tokens are projected to the transformer hidden dimension using the Llama token embedding layer. Inputting only these tokens results in our PAR model knowing what speed the other agents are going at, but not where they are. It is important the model has this awareness (it should know if two agents are going to collide), so our model needs to reason over this second modality of location. We implement this by passing locations relative to the agent we are predicting into a sincosine positional embedding (see details in Sec. 12.1), which we denote a location positional encoding (LPE). The LPE is summed to our token embeddings. + +We use a cross-entropy classification loss on our discrete tokens: $\mathcal{L} = E_{y\sim p(y)}[-\log (p(\mathbf{s}_{t_\pi}^T))]$ . We use the standard average displacement error (ADE) and final displacement error (FDE) to evaluate our predicted trajectories. For our baselines (Sec. 3.3), we use the closest agent at the current timestep for Multiagent NN and Mirror. For NN and Multiagent NN we use MSE as the distance metric. + +
Token typeLPEMethodADE ↓FDE ↓
VelocityX1-agent AR1.503.64
VelocityX3-agent PAR1.453.51
AccelerationX1-agent AR1.443.57
AccelerationX3-agent PAR1.403.44
Accleration✓3-agent PAR1.353.34
+ +Table 3. Car trajectory prediction performance. Using acceleration tokens and 3-agent PAR results in a stronger performance over velocity tokens and single-agent AR. Adding location via a positional encoding (LPE) further improves results. + +
BaselineAgentsADE ↓FDE ↓
Random Trajectory18.8916.51
NN11.804.13
Multiagent NNN6.4012.04
MirrorN11.5914.93
+ +Table 4. Car trajectory prediction baselines. Nearest neighbor performs best overall, but our learned single-agent AR models outperform all baselines. + +![](images/5d13e6894e58ebababc4444c6e8055ffd8d0bcba1d06a06f5a548822a59ab381.jpg) +Figure 6. Example results from our single-agent AR model (top row) and three-agent PAR model with location positional encoding (bottom row) on nuScenes. The predicted agent's ground truth trajectory is in pink, and the predicted future in blue. For the PAR model, the other two agents' ground truth states are in green. Qualitatively, the PAR model handles situations where single-agent predictions might lead to collisions (A, B), uses other agents' behavior to better adhere to road areas (A, C) without environment data, and predicts based on the speed changes of other cars (D). +Figure 8. Translation forecasting qualitative result on test set. 3D predictions are projected onto the image, isolating translation results by showing the ground-truth rotation. Using the PAR framework (right) instead of AR (left) improves object pose prediction. + +# 5.2. Results + +We train AR and 3-agent PAR models using velocity tokens, acceleration tokens, and acceleration tokens combined with our location positional encoding. The results can be seen in Table 3. Note that the 3-agent PAR model uses the agent ID embedding and next timestep prediction. Acceleration tokens consistently outperform velocity tokens both for agent AR and 3-agent PAR models. This could be because the vocabulary size for acceleration tokens is much smaller and therefore easier to optimize. Regardless, both ways of tokenizing result in models that outperform our baselines (see + +![](images/def4b1307324993288b96bbab1b5936cf16c7eabc70bf2c40351dbca37af15e4.jpg) +1-Agent AR + +![](images/cfbf7d426bc29faf021bd3a45c9e36115c17a0a2b6199eff482202ed58baa139.jpg) +2-Agent PAR + +![](images/42d64316489f0226b383bb7ee1615551eff970998cd7c5f20176090569f35980.jpg) +Figure 7. Rotation forecasting qualitative result on test set. 3D predictions are projected onto the image, isolating rotation results by showing the ground-truth translation. Incorporating the hand agent in the PAR framework (right) improves object pose prediction over object-only AR (left). +1-Agent AR + +![](images/b427a6868d2dc4570d1314493ae331ddef2aa66f616355d44fb90648f2130b8b.jpg) +2-Agent PAR + +Table 4 - NN has a relatively low error on this dataset), and highlight that our framework is flexible such that a user can experiment with different ways of representing entities. For both token types, the 3-agent PAR model that is blind to location outperforms the AR model. While location information should help the model, it is possible that simply knowing whether other agents are slowing down or accelerating can help the model make better predictions. When adding location information via the LPE to our 3-agent PAR model, we see another performance gain in ADE and FDE. + +Qualitative examples of the AR model (top row) and 3-agent location-aware PAR model (bottom row) can be seen in Figure 6. Our method uses no image or environment data (e.g., lanes) as input. However, by reasoning over multiple agents, its predictions lead to fewer collisions and better reasoning about speed changes and driveable areas based solely on other agents' behaviors. + +# 6. Case Study 3: Object Pose Forecasting During Hand-Object Interaction + +Our final case study explores how hand-object interaction can be leveraged for object pose estimation. We define the human hand and the interacting object as two agents, with tokens representing distinct state types. We show that our PAR framework can jointly model these agents, improving 3D translation and rotation predictions for the object. + +# 6.1. Experimental Setup + +Dataset. We use the DexYCB dataset, which includes 1000 videos of 10 subjects performing object manipulation tasks with 20 distinct objects from the YCB-Video dataset. The data is split into 800 training, 40 validation, and 160 testing videos. We use one of 8 provided camera views. In each trial, subjects pick up and lift objects in randomized conditions. Labels include the object's SO(3) rotation and 3D translation, and the hand's 3D translation. We focus on predicting the object's rotation or translation. + +Task-specific considerations. We tokenize object information in object-only experiments and both object and hand information in hand-object experiments. The object is represented as a 4D token for rotation forecasting (quaternion from SO(3) rotation) or a 3D token for translation forecasting (Euclidean coordinates). In hand-object experiments, the hand token is included with a 3D translation vector, and agent ID embeddings distinguish between the hand and object. Normalization is applied to all 3D translation vectors in both AR and PAR experiments; quaternions are normalized by definition and require no additional processing. An embedding layer projects the tokens into the transformer's hidden dimension, and another layer projects them back for prediction. + +For rotation-only forecasting, the loss is $\mathcal{L}_{rot} = 1 - |\hat{q}\cdot q|$ , where $\hat{q}$ is the predicted quaternion and $q$ the ground-truth quaternion. For translation-only forecasting, the loss $\mathcal{L}_t$ is the mean squared error (MSE) between predicted and ground-truth translations. For PAR we predict relative object-to-hand translations at each frame, using the current hand position as origin, while for AR, we predict absolute object translations without considering the interacting agent. For PAR models, we add the loss $\mathcal{L}_h$ , a MSE on hand translation. The object-only AR rotation model is optimized with $\mathcal{L}_{rot}$ , while the PAR rotation model combines $\mathcal{L}_{rot} + \mathcal{L}_h$ ; similarly, the object-only translation model is trained with $\mathcal{L}_t$ , and the hand-object translation model uses $\mathcal{L}_t + \mathcal{L}_h$ . At inference, the first half of each video is provided, and object predictions are autoregressively generated for the second half. Translation is evaluated using MSE, while rotation is measured using geodesic distance (GEO) on SO(3). + +# 6.2. Results + +We compare the object-only AR models to the hand-object PAR models in Table 5 for the two prediction tasks. We also present the baselines described in Sec. 3.3 in Table 6. Figures 7 and 8 show qualitative results on the rotation and translation predictions, respectively. In both prediction tasks, we observe that incorporating the human hand's interaction with the object enhances accuracy: for rotation, PAR results in a relative improvement of $8.9\%$ over AR, and for translation, $41\%$ . See Section 11.2 for additional qualitative results + +
TypeMethodMSE ↓GEO (rad) ↓
Translation1-agent AR3.68 × 10-3-
Translation2-agent PAR2.17 × 10-3-
Rotation1-agent AR-0.919
Rotation2-agent PAR-0.837
+ +Table 5. Test set results on DexYCB dataset. For both rotation and translation forecasting, the 2-agent PAR model, which treats the hand as an additional agent, improves results. + +
BaselineTranslation - MSE (m2) ↓Rotation - GEO (rad) ↓
Random0.2442.196
Random Trajectory1.60 × 10-22.146
NN1.69 × 10-22.179
Multiagent NN1.71 × 10-22.170
Mirror1.20 × 10-2-
+ +Table 6. Test set results for DexYCB baselines. We cannot provide rotation results for the Mirror baseline, because the ground-truth does not include hand rotation, only 3D translation. + +with more sampled frames. + +# 7. Discussion + +This work introduced the Poly-Autoregressive (PAR) framework, a unifying approach to prediction for multi-agent interactions. By applying the same transformer architecture and hyperparameters across diverse tasks, including action forecasting in social settings, trajectory prediction for autonomous vehicles, and object pose forecasting during hand-object interaction, we have demonstrated the versatility and robustness of our framework. + +Our findings underscore the crucial importance of considering the influence of multiple agents in a scene for prediction tasks. By modeling interactions, we significantly improved prediction accuracy over single-agent approaches on all three problems we considered. While we achieved promising results with a simple architecture, we have only provided a starting point that can be built upon extensively. For instance, incorporating environmental context or tokenizing pixel patches, especially as a way to relax our assumption on high-quality tracking, are avenues for further research using PAR. It would be interesting to experiment with scaling the data and model. We do not specifically consider the relative importance of neighboring agents, which is an interesting future research direction (ex. dynamic attention mechanism). + +Our PAR framework provides a simple and generalizable foundation of universal building blocks, ready for extension or refinement in future tasks. The PAR framework holds potential for advancing AI systems, enhancing prediction capabilities and enabling more accurate navigation and operation in real-world, multi-agent interactions. + +# 8. Acknowledgements + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2 +[2] Fred Attneave. Some informational aspects of visual perception. Psychological review, 61(3):183, 1954. 2 +[3] Murchana Baruah and Bonny Banerjee. A multimodal predictive agent model for human interaction generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 1022-1023, 2020. 2 +[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2 +[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 6, 1 +[6] Sergio Casas, Wenjie Luo, and Raquel Urtasun. Intentnet: Learning to predict intention from raw sensor data. In Conference on Robot Learning, pages 947-956. PMLR, 2018. 1 +[7] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 1 +[8] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 1 +[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2 +[10] Baptiste Chopin, Hao Tang, Naima Otberdout, Mohamed Daoudi, and Nicu Sebe. Interaction transformer for human reaction generation. IEEE Transactions on Multimedia, pages 1-13, 2023. 2 +[11] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, KueiDa Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1 + +We thank Jane Wu, Himanshu Singh, Georgios Pavlakos, and João Carreira for useful discussions and feedback. This work was supported by ONR MURI N00014-21-1-2801 and NSF Graduate Fellowships to NT and TS. +[12] Henggang Cui, Vladan Radosavljevic, Fang-Chieh Chou, Tsung-Han Lin, Thi Nguyen, Tzu-Kuo Huang, Jeff Schneider, and Nemanja Djuric. Multimodal trajectory predictions for autonomous driving using deep convolutional networks. In 2019 international conference on robotics and automation (icra), pages 2090-2096. IEEE, 2019. 1 +[13] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 1 +[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 1 +[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 1 +[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1 +[17] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. In European Conference on Computer Vision (ECCV), page 3, 2024. 2 +[18] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In International Conference on Machine Learning, pages 1242-1250. PMLR, 2014. 2 +[19] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6047-6056, 2018. 2, 5 +[20] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13053-13064, 2022. 2 +[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1 +[22] Yanjun Huang, Jiatong Du, Ziru Yang, Zewei Zhou, Lin Zhang, and Hong Chen. A survey on trajectory-prediction methods for autonomous driving. IEEE Transactions on Intelligent Vehicles, 7(3):652-674, 2022. 1 +[23] Boris Ivanovic, Guanyu Song, Igor Gilitschenski, and Marco Pavone. trajdata: A unified interface to multiple human trajectory datasets. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, New Orleans, USA, 2023. 6 +[24] Bolin Lai, Sam Toyer, Tushar Nagarajan, Rohit Girdhar, Shengxin Zha, James M Rehg, Kris Kitani, Kristen Grau + +man, Ruta Desai, and Miao Liu. Human action anticipation: A survey. arXiv preprint arXiv:2410.14045, 2024. 1 +[25] Hugo Larochelle and Iain Murray. The neural autoregressive distribution estimator. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 29-37. JMLR Workshop and Conference Proceedings, 2011. 2 +[26] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 1 +[27] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 2 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2 +[29] Siyuan Brandon Loh, Debaditya Roy, and Basura Fernando. Long-term action forecasting using multi-headed attention-based variational recurrent neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2419-2427, 2022. 1 +[30] Vongani Maluleke, Lea Müller, Jathushan Rajasegaran, Georgios Pavlakos, Shiry Ginosar, Angjoo Kanazawa, and Jitendra Malik. Synergy and synchrony in couple dances. arXiv preprint arXiv:2409.04440, 2024. 2 +[31] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 1 +[32] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20395-20405, 2022. 2 +[33] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2 +[34] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In ArXiv, 2024. 2 +[35] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 1 +[36] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer, 2018. 2 +[37] Alec Radford. Improving language understanding by generative pre-training. 2018. 2 +[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 + +[39] Ilija Radosavovic, Bike Zhang, Baifeng Shi, Jathushan Rajasegaran, Sarthak Kamat, Trevor Darrell, Koushil Sreenath, and Jitendra Malik. Humanoid locomotion as next token prediction. arXiv preprint arXiv:2402.19469, 2024. 2 +[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, Christoph Feichtenhofer, and Jitendra Malik. On the benefits of 3d pose and tracking for human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2023. 1 +[41] Jathushan Rajasegaran, Ilija Radosavovic, Rahul Ravishankar, Yossi Gandelsman, Christoph Feichtenhofer, and Jitendra Malik. An empirical study of autoregressive pre-training from videos. arXiv preprint arXiv:2501.05453, 2025. 2 +[42] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 1 +[43] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 1 +[44] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8579-8590, 2023. 2, 1 +[45] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 30(1):50-64, 1951. 2 +[46] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. arXiv preprint arXiv:2403.18811, 2024. 2 +[47] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5 +[48] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Rahul Sukthankar, Kevin Murphy, and Cordelia Schmid. Relational action forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 273-283, 2019. 1 +[49] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 1 +[50] Lucas Theis and Matthias Bethge. Generative image modeling using spatial lstms. Advances in neural information processing systems, 28, 2015. 2 +[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste + +Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 5 +[52] Ameni Trabelsi, Mohamed Chaabane, Nathaniel Blanchard, and Ross Beveridge. A pose proposal and refinement network for better 6d object pose estimation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2382-2391, 2021. 1 +[53] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 2 +[54] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2 +[55] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16611–16621, 2021. 1 +[56] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373. IEEE, 2020. 1 +[57] Bowen Wen, Jonathan Tremblay, Valts Blukis, Stephen Tyree, Thomas Müller, Alex Evans, Dieter Fox, Jan Kautz, and Stan Birchfield. Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 606-617, 2023. +[58] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868-17879, 2024. 1 +[59] Jane Wu, Georgios Pavlakos, Georgia Gkioxari, and Jitendra Malik. Reconstructing hand-held objects in 3d. arXiv preprint arXiv:2404.06507, 2024. 1 +[60] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 1 +[61] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 1 \ No newline at end of file diff --git a/2025/Poly-Autoregressive Prediction for Modeling Interactions/images.zip b/2025/Poly-Autoregressive Prediction for Modeling Interactions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..247fa7c0b62086310c1b0e94488ccbd5dd9d431b --- /dev/null +++ b/2025/Poly-Autoregressive Prediction for Modeling Interactions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2672ddb8a6af0b2d18dadb0d047fc91dbc0766a9552e2736bbed71249c3fb4a6 +size 435213 diff --git a/2025/Poly-Autoregressive Prediction for Modeling Interactions/layout.json b/2025/Poly-Autoregressive Prediction for Modeling Interactions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6268f4e16b6512d38669b9748786e92fed7f1257 --- /dev/null +++ b/2025/Poly-Autoregressive Prediction for Modeling Interactions/layout.json @@ -0,0 +1,9782 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 128, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 128, + 103, + 483, + 121 + ], + "type": "text", + "content": "Poly-Autoregressive Prediction for Modeling Interactions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "spans": [ + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "text", + "content": "Neerja Thakkar" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "text", + "content": " Tara Sadjadpour" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "text", + "content": " Jathushan Rajasegeran" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "text", + "content": " Shiry Ginosar" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "text", + "content": " Jitendra Malik" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "text", + "content": "UC Berkeley, " + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "text", + "content": "Toyota Technical Institute at Chicago, " + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 65, + 142, + 544, + 178 + ], + "type": "text", + "content": "Google DeepMind" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 220, + 282, + 291 + ], + "blocks": [ + { + "bbox": [ + 60, + 220, + 282, + 291 + ], + "lines": [ + { + "bbox": [ + 60, + 220, + 282, + 291 + ], + "spans": [ + { + "bbox": [ + 60, + 220, + 282, + 291 + ], + "type": "image", + "image_path": "709f2fc30bfc60bc3d43308b1860f77821a18b7732f0f7e89851a55cdf7275f8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 312, + 201, + 323 + ], + "lines": [ + { + "bbox": [ + 119, + 312, + 201, + 323 + ], + "spans": [ + { + "bbox": [ + 119, + 312, + 201, + 323 + ], + "type": "text", + "content": "(a) Autoregression (AR)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 301, + 217, + 550, + 308 + ], + "blocks": [ + { + "bbox": [ + 301, + 217, + 550, + 308 + ], + "lines": [ + { + "bbox": [ + 301, + 217, + 550, + 308 + ], + "spans": [ + { + "bbox": [ + 301, + 217, + 550, + 308 + ], + "type": "image", + "image_path": "76fb006702cb31067a0dcf2db0afe7ff8004ce2ee609bee1bce0694edc3cb479.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 312, + 474, + 323 + ], + "lines": [ + { + "bbox": [ + 371, + 312, + 474, + 323 + ], + "spans": [ + { + "bbox": [ + 371, + 312, + 474, + 323 + ], + "type": "text", + "content": "(b) Poly-autoregression (PAR)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 331, + 555, + 367 + ], + "lines": [ + { + "bbox": [ + 54, + 331, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 555, + 367 + ], + "type": "text", + "content": "Figure 1. Inference for (a) autoregressive (AR) models and (b) our proposed poly-autoregressive (PAR) model. Solid indicates ground-truth tokens which represent a tracked data modality such as action or 6DOF pose; striped represents predicted output tokens. Color denotes agent identity. Compared to AR models, the PAR model takes other agents' tokens as inputs when making a prediction for the next timestep." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 383, + 200, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 383, + 200, + 396 + ], + "spans": [ + { + "bbox": [ + 151, + 383, + 200, + 396 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 407, + 297, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 407, + 297, + 623 + ], + "spans": [ + { + "bbox": [ + 54, + 407, + 297, + 623 + ], + "type": "text", + "content": "We introduce a simple framework for predicting the behavior of an agent in multi-agent settings. In contrast to autoregressive (AR) tasks, such as language processing, our focus is on scenarios with multiple agents whose interactions are shaped by physical constraints and internal motivations. To this end, we propose Poly-Autoregressive (PAR) modeling, which forecasts an ego agent's future behavior by reasoning about the ego agent's state history and the past and current states of other interacting agents. At its core, PAR represents the behavior of all agents as a sequence of tokens, each representing an agent's state at a specific timestep. With minimal data pre-processing changes, we show that PAR can be applied to three different problems: human action forecasting in social situations, trajectory prediction for autonomous vehicles, and object pose forecasting during hand-object interaction. Using a small proof-of-concept transformer backbone, PAR outperforms AR across these three scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Large language models (LLMs) have been very successful with natural language processing (NLP) tasks, which require accurate reasoning over relationships words have within a body of text. A key component of LLMs is autoregressive" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 384, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 467 + ], + "type": "text", + "content": "(AR) modeling, where each word token is predicted based on a sequence of preceding word tokens. Building on the success of AR modeling in the NLP community, this work focuses on modeling the dynamic relationships between agents and entities in everyday interactions that occur in the physical world, such as social interactions, driving, and hand-object interactions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 471, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 471, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 312, + 471, + 556, + 615 + ], + "type": "text", + "content": "Unlike language which is structured through grammar and semantics, interactions in the physical world are dictated by both the laws of physics (e.g. how a hand grasps an object) and the internal state of each agent (e.g. the trajectory an agent chooses to move a grasped object in), a latent variable that we know nothing about. Furthermore, as opposed to text where each word follows another unidimensionally, the states of multiple agents are changing simultaneously. For example, in social situations, the history of a single person's past states does not alone determine the dynamics of their future states; we also need to consider the states of other agents. We argue, therefore, that AR modeling alone is insufficient." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 557, + 715 + ], + "type": "text", + "content": "In this paper, we introduce poly-autoregressive (PAR) modeling, a simple unifying approach to model the influence of other agents and entities on one's behavior. We model behavior as a temporal sequence of states and predict an ego agent's future behavior conditioned on the history of behavior of the ego agent as well as the rest of the agents. By considering other agents' behaviors, we demonstrate that our approach significantly improves upon the ill-posed problem" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12402" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "type": "text", + "content": "of single-agent prediction in interactive settings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 297, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 297, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 297, + 228 + ], + "type": "text", + "content": "The PAR framework uses a transformer-based model for next-token prediction. Transformers have shown great success in language modeling and naturally lend themselves to predicting behavior over time. In an interaction scenario of " + }, + { + "bbox": [ + 55, + 86, + 297, + 228 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 86, + 297, + 228 + ], + "type": "text", + "content": " agents, our model predicts the future behavior of an ego " + }, + { + "bbox": [ + 55, + 86, + 297, + 228 + ], + "type": "inline_equation", + "content": "(N^{\\mathrm{th}})" + }, + { + "bbox": [ + 55, + 86, + 297, + 228 + ], + "type": "text", + "content": " agent conditioned on its past behavior and the behavior of the other " + }, + { + "bbox": [ + 55, + 86, + 297, + 228 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 55, + 86, + 297, + 228 + ], + "type": "text", + "content": " non-ego agents (see Figure 1b). Each prediction task may model a different behavior modality of interest, e.g. actions for social action prediction, or 6DoF pose for object forecasting in hand-object interactions. We apply PAR to three different case studies of common real-world interactions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 296, + 483 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 232, + 295, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 295, + 303 + ], + "type": "text", + "content": "1. Social action prediction. We test our method on the AVA benchmark [19] for action forecasting. By incorporating both the ego and another agent using PAR, we get a overall " + }, + { + "bbox": [ + 56, + 232, + 295, + 303 + ], + "type": "inline_equation", + "content": "+1.9" + }, + { + "bbox": [ + 56, + 232, + 295, + 303 + ], + "type": "text", + "content": " absolute mAP gain over AR, which only models the ego agent to predict its future behavior, and an absolute " + }, + { + "bbox": [ + 56, + 232, + 295, + 303 + ], + "type": "inline_equation", + "content": "+3.5" + }, + { + "bbox": [ + 56, + 232, + 295, + 303 + ], + "type": "text", + "content": " mAP gain on 2-person interaction classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 386 + ], + "type": "text", + "content": "2. Trajectory prediction for autonomous vehicles. When forecasting future " + }, + { + "bbox": [ + 55, + 304, + 296, + 386 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 55, + 304, + 296, + 386 + ], + "type": "text", + "content": " locations of an ego vehicle, incorporating the locations of neighboring vehicles with PAR outperforms AR, which only uses the ego vehicle's preceding trajectory as input. Specifically, on the nuScenes dataset [5], PAR outperforms AR with a relative improvement of " + }, + { + "bbox": [ + 55, + 304, + 296, + 386 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 386 + ], + "type": "text", + "content": " ADE and " + }, + { + "bbox": [ + 55, + 304, + 296, + 386 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 386 + ], + "type": "text", + "content": " FDE." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": "3. 6DoF object pose forecasting during hand-object interaction. We use the DexYCB dataset [8], where we treat the object as the ego agent and the hand as the interacting agent. While PAR integrates that hand's 3D location and object pose history, AR only uses the object's pose history to predict the object's 6DoF pose. PAR outperforms AR with relative improvements of " + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "inline_equation", + "content": "41.0\\%" + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": " for the rotation and translation predictions, respectively." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 484, + 297, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 297, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 297, + 592 + ], + "type": "text", + "content": "In all these settings, we find that incorporating the behavior of other agents in the scene improves predictions of the ego agent's behavior. All of these problems are modeled via the same simple PAR framework and implemented using the same proof-of-concept 4 million parameter transformer without any modifications to the base framework or architecture, only to data pre-processing and choice of tokenization. We also provide an example of a simple way to build on our architecture through a location positional encoding, Sec. 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": "The primary contribution of this work is a versatile framework that can be applied to a diverse range of settings, without modifications aside from domain-specific data processing. Our results suggest that PAR provides a simple formulation that, with a more complex transformer backbone and larger datasets, could enhance prediction of diverse multi-agent interactions across various problem domains. To facilitate further exploration and development, we have released our code, which contains the building blocks to use PAR for modeling other types of multi-agent interactions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 71, + 402, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 402, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 402, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": "Autoregressive models. Autoregressive modeling has a rich history in information theory and deep learning, tracing back to Shannon's paper on language prediction [45] and Attneave's study on visual perception [2]. These foundational works laid the groundwork for modern applications in deep learning, including [25], which revisits neural autoregressive models and [18, 50], which explore continuous-valued modeling. [53] developed PixelRNN and PixelCNN to autoregressively generate images one pixel at a time using RNN and CNN architectures, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 219, + 556, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 556, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 556, + 447 + ], + "type": "text", + "content": "The development of the transformer model [54] spurred progress in computer vision with the image transformer [36] and the vision transformer [14], and autoregressive models [9, 41] and more notably in NLP, where the GPT family of models [4, 37, 38] has demonstrated the power of large-scale unsupervised autoregressive pre-training. Recent research has focused on multimodal learning, exemplified by the Flamingo [1] or LlaVa [28] models, which combine vision and language processing capabilities, illustrating the versatility of autoregressive models across various domains in artificial intelligence. While these approaches operate on image patches and word tokens, we operate on symbolic representations extracted from in-the-wild videos showing natural interactions. A recent approach [39] frames humanoid locomotion as an autoregressive next-token prediction task that operates on two types of continuous tokens: observations and actions. This approach projects continuous tokens to the hidden dimension and uses a shifted loss, similar to the next-timestep prediction proposed in our framework." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 454, + 556, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 556, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 556, + 646 + ], + "type": "text", + "content": "Multi-agent regressive models. Several prior works addressed modeling specific multi-agent problems via regressive models as one-off case studies. We introduce the PAR framework to unify these efforts into a single cohesive framework. Many behavior prediction works focus on two agents engaging in social interaction, whether it be dyadic communication [32-34] or social dance [30, 46]. These studies primarily tackle the challenge of predicting the state of an interacting partner (Person B) based on the input from Person A's state, sometimes extending predictions into the future [20, 30]. While earlier works used architectures such as variational RNNs [3], recent works have predominantly adopted transformer architectures for social interaction modeling [10, 20, 32, 33, 46], with some works exploring diffusion [27], or diffusion with attention [17]. Our PAR framework focuses on transformer models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Works encompassed by the PAR framework extend beyond human social interaction. Many multi-agent human or car trajectory prediction approaches use autoregressive prediction. For instance, MotionLM [44] utilizes a transformer decoder that processes multi-agent tokens, incorporating a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12403" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "learned agent ID embedding. This methodology informs our approach across all our case studies. Critically, in contrast to all prior multi-agent regressive works that designed solutions to address specific applications, we demonstrate that we can unify a diverse set of multi-agent regressive problems under a single PAR framework. See Appendix Sec. 9 for case-study-specific works." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 170, + 224, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 224, + 184 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 224, + 184 + ], + "type": "text", + "content": "3. Poly-Autoregressive Modeling" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 250 + ], + "type": "text", + "content": "Our goal is to model the behavior of an agent or entity while taking into account any other agents it interacts with, if any. To evaluate the performance of our model in capturing interaction dynamics, we predict the agent's future behavior and compare it against ground-truth data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 296, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 309 + ], + "type": "text", + "content": "We define the following task: In an interaction comprised of " + }, + { + "bbox": [ + 55, + 251, + 296, + 309 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 251, + 296, + 309 + ], + "type": "text", + "content": " agents, given the observed past states of the " + }, + { + "bbox": [ + 55, + 251, + 296, + 309 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 55, + 251, + 296, + 309 + ], + "type": "text", + "content": " interacting agents, and the observed or previously-predicted past states of the " + }, + { + "bbox": [ + 55, + 251, + 296, + 309 + ], + "type": "inline_equation", + "content": "N^{th}" + }, + { + "bbox": [ + 55, + 251, + 296, + 309 + ], + "type": "text", + "content": " ego agent, predict the future states of the " + }, + { + "bbox": [ + 55, + 251, + 296, + 309 + ], + "type": "inline_equation", + "content": "N^{th}" + }, + { + "bbox": [ + 55, + 251, + 296, + 309 + ], + "type": "text", + "content": " ego agent." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 310, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 370 + ], + "type": "text", + "content": "We define a transformer-based poly-autoregressive (PAR) predictor, " + }, + { + "bbox": [ + 55, + 310, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 310, + 296, + 370 + ], + "type": "text", + "content": ", that learns to model temporally long-range interactions in the input sequence. The inputs to the predictor are the past states of the " + }, + { + "bbox": [ + 55, + 310, + 296, + 370 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 310, + 296, + 370 + ], + "type": "text", + "content": " interacting agents, and its output is the predicted future state of the " + }, + { + "bbox": [ + 55, + 310, + 296, + 370 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 310, + 296, + 370 + ], + "type": "text", + "content": " ego agent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 376, + 167, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 167, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 167, + 387 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\{\\mathbf{s}_i\\}_{i=1}^T" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": " be a temporal sequence of agent states, " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^N" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{1:N-1}" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": " to denote the temporal sequences of states of the " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "N_{th}" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": " agent and of the other " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "N-1" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": " agents, respectively. For each timestep " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "t \\in [t_{\\pi}, T]" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "t_{\\pi} \\in [1, T]" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": " is the time we start predicting, we take as input all other " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "N-1" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": " agents' past observed state sequences " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{1:t-1}^{1:N-1}" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": " along with the " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "N_{th}" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": " agent's past observed states up to " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "t_{\\pi}" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{1:t_{\\pi}}^N" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": ", and any of its previously predicted past states " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}_{t_{\\pi}+1:t-1}^N" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": ", if available (see Fig. 1). Our predictor, " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": ", then poly-autoregressively predicts the " + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "inline_equation", + "content": "N_{th}" + }, + { + "bbox": [ + 55, + 392, + 296, + 515 + ], + "type": "text", + "content": " agent's future states one time-step at a time:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 101, + 524, + 296, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 524, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 101, + 524, + 296, + 540 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {s}} _ {t} ^ {N} = \\mathcal {P} \\left(\\mathbf {S} _ {1: t - 1} ^ {1: N - 1}, \\mathbf {S} _ {1: t _ {\\pi}} ^ {N}, \\hat {\\mathbf {S}} _ {t _ {\\pi} + 1: t - 1} ^ {N}\\right). \\tag {1}", + "image_path": "0a85f4cfd75678706522e11719802cfd368700a5200cba180902277515e94818.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 543, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 543, + 296, + 567 + ], + "type": "text", + "content": " learns to model the distribution over the next timestep of the " + }, + { + "bbox": [ + 55, + 543, + 296, + 567 + ], + "type": "inline_equation", + "content": "N_{th}" + }, + { + "bbox": [ + 55, + 543, + 296, + 567 + ], + "type": "text", + "content": " agent's states, given all other agents' states:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 571, + 295, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 571, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 127, + 571, + 295, + 587 + ], + "type": "interline_equation", + "content": "p \\left(\\hat {\\mathbf {s}} _ {t} ^ {N} \\mid \\mathbf {S} _ {1: t - 1} ^ {1: N - 1}, \\mathbf {S} _ {1: t - 1} ^ {N}\\right). \\tag {2}", + "image_path": "f1ef8fc1f570befd30f8a30d738c968dd113bdd4806c5ee7a1b2fc6139f03281.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 592, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 296, + 639 + ], + "type": "text", + "content": "While we provide the observed ground truth states of other agents at inference, during training, we jointly maximize the likelihood of all " + }, + { + "bbox": [ + 55, + 592, + 296, + 639 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 592, + 296, + 639 + ], + "type": "text", + "content": " agents by computing losses on their future state predictions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 639, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 664 + ], + "type": "text", + "content": "We train the predictor by maximizing the likelihood of the target state " + }, + { + "bbox": [ + 55, + 639, + 296, + 664 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 639, + 296, + 664 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 639, + 296, + 664 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 639, + 296, + 664 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 113, + 670, + 236, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 670, + 236, + 685 + ], + "spans": [ + { + "bbox": [ + 113, + 670, + 236, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {P}} = E _ {y \\sim p (y)} [ - \\log (p (\\mathbf {s} _ {t} ^ {N}) ],", + "image_path": "0f197b60ebf183851325a6f3aa9e28b703f05cc39c117055de29803e83786d61.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where the target state " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is computed from the " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{th}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " agent ground truth future state." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 509, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 509, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 509, + 84 + ], + "type": "text", + "content": "3.2. The Poly-Autoregressive Framework" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "text", + "content": "We address the problem of forecasting the future states of an agent (from time " + }, + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "text", + "content": ") in a data-driven way, given a temporal sequence of past states (from time 1 to " + }, + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "text", + "content": "). We assume that our agent has some feature, or a set of features, of interest in a video (e.g., 3D pose) that we can tokenize. We predict the future states of the agent in terms of this tokenized feature (or set of), where we use one token (or set of tokens) per time step. The predicted tokens can be discrete (i.e., an index into a feature codebook) or continuous (i.e., a vector of one or more continuous values). The loss " + }, + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "text", + "content": " will depend on the problem's specifics and the type of token used. To train the model to predict the future, we rely on all the interaction dynamics of length " + }, + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 89, + 555, + 256 + ], + "type": "text", + "content": " in our training dataset as ground truth examples." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 257, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 316 + ], + "type": "text", + "content": "As a baseline, we consider the single-agent autoregressive (AR) paradigm, where a transformer is trained to perform next-token-prediction with teacher forcing. AR uses greedy sampling to generate sequences at inference time, predicting one next token at a time (Fig. 1(a))." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "text", + "content": "In contrast, our multi-agent poly-autoregressive (PAR) framework considers the other " + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "text", + "content": " agents in the scene when predicting the future state of the " + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "text", + "content": "th agent. In this setup, we utilize the features of interest of all " + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "text", + "content": " agents, yielding " + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "text", + "content": " tokens at each timestep for a total of " + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "inline_equation", + "content": "N * T" + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "text", + "content": " tokens. In practice, we operate on a flattened sequence of " + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "inline_equation", + "content": "N * T" + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "text", + "content": " tokens. Instead of using the AR training procedure in this multi-agent case (as in Fig. 3a), we jointly model the " + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 316, + 554, + 436 + ], + "type": "text", + "content": " agents at each timestep by introducing the following features to our PAR framework." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "text", + "content": "Next-timestep prediction. A standard AR model predicts the next token. Given the flattened sequence of " + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "inline_equation", + "content": "N * T" + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "text", + "content": " tokens our model operates on, next token prediction would take as input an agent " + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "text", + "content": " and predict agent " + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "inline_equation", + "content": "k + 1" + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "text", + "content": "'s state at the same timestep " + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "text", + "content": " (as in Fig. 3a). However, our goal is to predict the input agent " + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "text", + "content": "'s future state at time " + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 313, + 441, + 555, + 550 + ], + "type": "text", + "content": ". Therefore, we perform same-agent next-timestep prediction rather than next-token prediction (see Fig. 3b for an illustration of same-agent next-timestep at training)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 554, + 554, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 554, + 613 + ], + "type": "text", + "content": "Learned agent identity embedding. When giving a model information corresponding to multiple agents, the model can benefit from knowing which token corresponds to which agent. We give the model this information with a learned agent ID embedding." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 666 + ], + "type": "text", + "content": "Joint training. We train the model to jointly predict the future of all agents by computing a loss on the predicted tokens of all agents (Fig. 3b). Please refer to Section 3.1 for our inference paradigm." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "type": "text", + "content": "3.3. Task-Specific Considerations" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Our simple PAR approach unifies diverse problems under a single framework and architecture without any modifications." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12404" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 549, + 246 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 549, + 246 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 549, + 246 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 549, + 246 + ], + "type": "image", + "image_path": "82a17ddfb9150be20f986188f09277d60cac11c0377ae2ffa992a7c4207655f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 256, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 54, + 256, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 555, + 313 + ], + "type": "text", + "content": "Figure 2. The PAR Framework. We begin by collecting a video dataset, such as AVA (top) or DexYCB (bottom). Then, using dataset labels or computer vision techniques, a trajectory of a given modality for our prediction task is extracted for each agent, such as action class labels (top) or object pose and 3D hand translation (bottom). Data is then tokenized, either through discretization or directly using continuous values, with our framework supporting both formats. Based on the tokenization and prediction task, we choose the appropriate loss function for PAR training. After training with PAR, predicted tokens can be decoded back to data space and evaluated with relevant metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 83, + 321, + 276, + 406 + ], + "blocks": [ + { + "bbox": [ + 83, + 321, + 276, + 406 + ], + "lines": [ + { + "bbox": [ + 83, + 321, + 276, + 406 + ], + "spans": [ + { + "bbox": [ + 83, + 321, + 276, + 406 + ], + "type": "image", + "image_path": "2b8c56e25ce69fce0b0be688f7fe04386f70a084587cfb72faf67a20e577e2e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 407, + 244, + 417 + ], + "lines": [ + { + "bbox": [ + 113, + 407, + 244, + 417 + ], + "spans": [ + { + "bbox": [ + 113, + 407, + 244, + 417 + ], + "type": "text", + "content": "(a) AR: multi-agent, next-token training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 83, + 418, + 276, + 503 + ], + "blocks": [ + { + "bbox": [ + 83, + 418, + 276, + 503 + ], + "lines": [ + { + "bbox": [ + 83, + 418, + 276, + 503 + ], + "spans": [ + { + "bbox": [ + 83, + 418, + 276, + 503 + ], + "type": "image", + "image_path": "ec2468950fe8402fa30610d7862e0da7ed3da055861f2cfd5821056d079c9eef.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 506, + 251, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 251, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 251, + 516 + ], + "type": "text", + "content": "(b) PAR: same-agent, next-timestep training." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 525, + 295, + 613 + ], + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 613 + ], + "type": "text", + "content": "Figure 3. Training with teacher forcing for (a) multi-agent next-token prediction in autoregressive models and (b) multi-agent poly-autoregressive models. Solid vs striped indicates a ground-truth vs predicted token, respectively. Color denotes agent identity. The AR model is trained for next-token prediction, while the PAR model is trained to predict the next timestep of the same agent. Three agents are shown for ease of visualization, but the PAR model supports an arbitrary number of agents." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 613, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 636 + ], + "type": "text", + "content": "In order to formulate a problem as interaction-conditioned prediction, users must consider several task-specific details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 637, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 296, + 661 + ], + "type": "text", + "content": "Fig. 2 gives an overview of how the PAR framework disentangles multi-agent learning from problem-specific modeling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "Data. The input data source in our example tasks is always a collection of videos. From these videos, we extract various modalities relevant to the task at hand. These modalities can range from high-level features, such as action class labels, to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 323, + 555, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 359 + ], + "type": "text", + "content": "low-level ones, such as 3D pose (Fig. 2 first two columns). We assume that each agent in the dataset is detected at each frame and is associated with an agent ID." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 556, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 508 + ], + "type": "text", + "content": "Tokenization. Our framework supports both discrete, quantized tokens and continuous vector tokens. The choice between discrete and continuous depends on the nature of the task. In the case of discrete tokens, we use a standard embedding layer to project to the hidden dimension. For continuous tokens, we train a projection layer to project the token into the hidden dimension of the transformer. For instance, if our continuous token is a 3D vector with an " + }, + { + "bbox": [ + 313, + 365, + 556, + 508 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 313, + 365, + 556, + 508 + ], + "type": "text", + "content": " 3D location coordinate and our hidden dimension is 128, our projection layer will project from 3 to 128 dimensions. We also train an un-projection layer that reverts the hidden dimension to the original token dimension." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 515, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 555, + 574 + ], + "type": "text", + "content": "Loss. The type of token and task-specific considerations dictate the loss function " + }, + { + "bbox": [ + 313, + 515, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 515, + 555, + 574 + ], + "type": "text", + "content": " applied during model training. For discrete tokens, a classification loss is appropriate. For continuous tokens, we use a regression loss on the original token dimension." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 581, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 605 + ], + "type": "text", + "content": "Baselines. We compare to the following baselines, where applicable on a case-by-case basis:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "text", + "content": "- Random token: pick random tokens from the best available token space and use as the prediction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 653 + ], + "type": "text", + "content": "- Random trajectory: pick at random a trajectory from the training dataset to use as the prediction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "type": "text", + "content": "- NN: Given an input agent " + }, + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "type": "text", + "content": "'s trajectory history, find the closest trajectory to it in the training set, belonging to " + }, + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "type": "inline_equation", + "content": "A^T" + }, + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "type": "text", + "content": ". Use " + }, + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "type": "inline_equation", + "content": "A^T" + }, + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "type": "text", + "content": "'s future as the predicted future." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "- Multiagent NN: In a dataset with two interacting partners " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " is the ego agent, given an input agent" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 726, + 317, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 726, + 317, + 734 + ], + "spans": [ + { + "bbox": [ + 293, + 726, + 317, + 734 + ], + "type": "text", + "content": "12405" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "'s trajectory history, find the closest trajectory to it in the training set, belonging to " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "A^T" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": ". Use " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "A^T" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "'s interaction partner's " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "B^T" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "'s future as the prediction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 295, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 147 + ], + "type": "text", + "content": "- Mirror: In a dataset with two interacting partners " + }, + { + "bbox": [ + 55, + 110, + 295, + 147 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 110, + 295, + 147 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 110, + 295, + 147 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 110, + 295, + 147 + ], + "type": "text", + "content": ", use the ground truth future of agent " + }, + { + "bbox": [ + 55, + 110, + 295, + 147 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 110, + 295, + 147 + ], + "type": "text", + "content": " as the predicted future for agent " + }, + { + "bbox": [ + 55, + 110, + 295, + 147 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 110, + 295, + 147 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 152, + 246, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 246, + 165 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 246, + 165 + ], + "type": "text", + "content": "3.4. Framework Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 170, + 295, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 295, + 194 + ], + "type": "text", + "content": "We keep the following implementation details constant for all case studies (see also Sec. 10)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 198, + 296, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 234 + ], + "type": "text", + "content": "Learned agent ID embedding. Our learned agent ID embedding consists of the integer agent ID mapped to a hidden dim-sized vector, and summed to the token embedding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 357 + ], + "type": "text", + "content": "Architecture. For all case studies, we use the Llama [51] transformer decoder architecture with 8 layers, 8 attention heads, and a hidden and intermediate dimension of 128. The decoder has " + }, + { + "bbox": [ + 55, + 238, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\sim 4.4\\mathrm{M}" + }, + { + "bbox": [ + 55, + 238, + 296, + 357 + ], + "type": "text", + "content": " learned parameters, not including learned embedding layers which add a few thousand more parameters. A rotary positional encoding [47] is used in addition to other summed encodings (i.e. agent ID embedding, locational positional encoding in Sec. 5). We train using teacher forcing. The only hyperparameter that changes between case studies is the learning rate." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 367, + 274, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 274, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 274, + 380 + ], + "type": "text", + "content": "4. Case Study 1: Social Action Forecasting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 483 + ], + "type": "text", + "content": "Our first case study involves forecasting human actions. Human behaviors are fundamentally social; for instance, individuals frequently walk in groups and alternate between speaking and listening roles when conversing. Certain actions, like hugging or handshaking, are intrinsically multiperson. Therefore, modeling human interactions should help improve action forecasting performance, especially on multiperson actions, which we show in this case study." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 489, + 172, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 172, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 172, + 502 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 506, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 673 + ], + "type": "text", + "content": "Dataset. The Atomic Visual Actions (AVA) dataset [19] comprises 235 training and 64 15-minute validation videos from movies. Annotations are provided at a 1Hz frequency, detailing bounding boxes and tracks for individuals within the frame, and each person's actions within a 1-second timeframe. Individuals may engage in multiple concurrent actions from a repertoire of 60 distinct action classes (e.g., sitting and talking simultaneously). For our analysis, we select clips featuring a continuous sequence of an agent's actions spanning at least 4s, splitting sequences exceeding 12s. We use the first half of each clip as history to predict the second half. For any ego agent trajectory, we pick a second agent by selecting the person present in the scene for the longest subset of the ego agent's trajectory." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Task-specific considerations. Each agent's token " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " represents an 60-dimensional vector that corresponds to the actions performed at a specific timestep. Each element denotes" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 156 + ], + "type": "table", + "html": "
MethodTimestep predAg-ID embdmAP ↑
1-agent ARN/AN/A40.7
2-agent ARXX38.0
2-agent PAR*X✓40.2
2-agent PAR*✓X40.0
2-agent PAR✓✓42.6
", + "image_path": "918d894c8eab9de6ce9581537fa3c4e1da73166755f84a2c911650fb8a7a4e47.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 348, + 255, + 520, + 340 + ], + "blocks": [ + { + "bbox": [ + 313, + 163, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 251 + ], + "type": "text", + "content": "Table 1. PAR action forecasting performance on AVA We evaluate 1 and 2-agent AR methods, two 2-agent PAR ablations (rows 3 and 4, " + }, + { + "bbox": [ + 313, + 163, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{PAR^{*}}" + }, + { + "bbox": [ + 313, + 163, + 555, + 251 + ], + "type": "text", + "content": "), and our PAR method. Without next-timestep prediction (see Fig. 3) or a learned agent ID embedding, our model struggles with multi-agent reasoning, performing worse than the AR baseline. With both components, the 2-agent PAR model achieves a " + }, + { + "bbox": [ + 313, + 163, + 555, + 251 + ], + "type": "inline_equation", + "content": "+1.9" + }, + { + "bbox": [ + 313, + 163, + 555, + 251 + ], + "type": "text", + "content": " mAP gain over the AR method (see Fig. 11 and Fig. 5 for class breakdown)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 348, + 255, + 520, + 340 + ], + "lines": [ + { + "bbox": [ + 348, + 255, + 520, + 340 + ], + "spans": [ + { + "bbox": [ + 348, + 255, + 520, + 340 + ], + "type": "table", + "html": "
BaselineAgentsmAP ↑
Random Token13.46
Random Training Traj13.44
Nearest Neighbor113.17
Multiagent NN25.10
Mirror27.97
", + "image_path": "d4aa1b5f0fd67690160c6ee5ed3045f8b9bbfa25787737491f44344e0220aeb4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 349, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 381 + ], + "type": "text", + "content": "Table 2. AVA baselines While the nearest neighbor baseline performs best among baselines, it is still significantly worse than the AR model." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 389, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 509 + ], + "type": "text", + "content": "the probability of a particular action class being enacted; ground-truth inputs are a binary vector. We implement an embedding layer that projects these tokens into the transformer's hidden dimension, as well as an un-projection layer that reverts them back to the original 60D token space for the purposes of loss calculation and output generation. We do not explicitly require the outputs to be values between 0 and 1. We use a MSE regression loss on the 60D action tokens: " + }, + { + "bbox": [ + 313, + 389, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\frac{1}{n}\\sum_{i=1}^{n}(\\mathcal{A}_i - \\hat{\\mathcal{A}}_i)^2" + }, + { + "bbox": [ + 313, + 389, + 554, + 509 + ], + "type": "text", + "content": ". Our evaluation metric is the mean average precision (mAP) on the 60 AVA classes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 556 + ], + "type": "text", + "content": "We implement all baselines described in 3.3, where Random Token corresponds to a random 60D vector sampled from 0 to 1. NN and Multiagent NN use Hamming distance as the distance metric." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 564, + 372, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 372, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 372, + 575 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "We report the performance of a single-agent AR model as a baseline, in the first line of Table 1. The AR model is significantly better than our baselines (see Table 2), the strongest baseline being the single-agent NN. We compare these baselines to our 2-agent PAR model (last line) and various ablations where we remove the agent ID embedding and perform next-token rather than same-agent next-timestep prediction. The second line of the table corresponds to multi-agent next-token prediction (Fig. 3a). We see that this approach confuses the model, and the performance is significantly worse than just training on and considering a single agent. However," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "spans": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "type": "text", + "content": "12406" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 73, + 529, + 228 + ], + "blocks": [ + { + "bbox": [ + 86, + 73, + 529, + 228 + ], + "lines": [ + { + "bbox": [ + 86, + 73, + 529, + 228 + ], + "spans": [ + { + "bbox": [ + 86, + 73, + 529, + 228 + ], + "type": "image", + "image_path": "f33dc690baa1e6313f48ddb1725c5177b489af2d575f1db9777a17af674f3a8e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 237, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 54, + 237, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 54, + 237, + 555, + 282 + ], + "type": "text", + "content": "Figure 4. Action forecasting example. The distribution over ground truth actions are in white, and our predictions in red. A 6s action history (1Hz) is input, and 6s of future actions predicted. In the scene, the man and woman alternate between talking and listening. Initially, the man is talking. The AR model predicts he will continue talking, while the 2-agent PAR model recognizes the woman is talking and predicts more accurate turn-taking behavior." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 287, + 290, + 399 + ], + "blocks": [ + { + "bbox": [ + 60, + 287, + 290, + 399 + ], + "lines": [ + { + "bbox": [ + 60, + 287, + 290, + 399 + ], + "spans": [ + { + "bbox": [ + 60, + 287, + 290, + 399 + ], + "type": "image", + "image_path": "9f4fb74698e8f8f190ab3c6ab86b66b3b909184799962d45dbcdabfa4e945f10.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "text", + "content": "Figure 5. Per-class mAP for AVA 2-person actions. We see performance improvement on almost all 2-person AVA action classes ((P) stands for \"a person\"). Some absolute mAP gains are particularly significant: listen to " + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "inline_equation", + "content": "+7.0" + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "text", + "content": ", kiss " + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "inline_equation", + "content": "+8.3" + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "text", + "content": ", fight/hit " + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "inline_equation", + "content": "+5.7" + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "text", + "content": ", talk to " + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "inline_equation", + "content": "+4.4" + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "text", + "content": ", hug " + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "inline_equation", + "content": "+5.7" + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "text", + "content": ", and hand shake " + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "inline_equation", + "content": "+4.0" + }, + { + "bbox": [ + 55, + 411, + 298, + 468 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 482, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 296, + 555 + ], + "type": "text", + "content": "as we add various components of our PAR approach, the performance improves, and with both the next timestep prediction and agent ID embedding, we get a " + }, + { + "bbox": [ + 55, + 482, + 296, + 555 + ], + "type": "inline_equation", + "content": "+1.9" + }, + { + "bbox": [ + 55, + 482, + 296, + 555 + ], + "type": "text", + "content": " mAP gain. When only considering 2-person action classes (enumerated in Fig. 5), our mAP is 36.3 on the single agent PAR model and 39.8 on the 2-agent PAR model, a " + }, + { + "bbox": [ + 55, + 482, + 296, + 555 + ], + "type": "inline_equation", + "content": "+3.5" + }, + { + "bbox": [ + 55, + 482, + 296, + 555 + ], + "type": "text", + "content": " mAP gain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "In Fig. 4 we see an example of action forecasting. In the input history, the man talks and the woman listens. In the future, the woman talks, and the man listens. Our 2-agent PAR model (bottom row) better understands that talking and listening actions are complementary actions, while the AR model doesn't learn this correlation. We see quantitative evidence of this in Fig. 5, with per-class mAPs for our AR vs 2-agent PAR model for 2-person action classes. Here, talk to gets a " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "+4.4" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " mAP gain and listen to gets a " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "+7.0" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " mAP gain when we train a multi-agent model. We see a significant boost on many other interaction-related action classes—for instance, kiss a person " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "+8.3" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " and fight/hit a person " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "+5.7" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " mAP—and on single-person actions, see Fig. 11." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 284, + 555, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 310 + ], + "type": "text", + "content": "5. Case Study 2: Multiagent Car Trajectory Prediction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 319, + 556, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 556, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 556, + 380 + ], + "type": "text", + "content": "Our second case study focuses on predicting car trajectories. Trajectory prediction requires a vehicle to be aware of other cars on the road to avoid collisions and promote cooperative behavior. This study demonstrates how our framework enables the joint modeling of multiple vehicles' movements." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 385, + 432, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 432, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 432, + 399 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 403, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 462 + ], + "type": "text", + "content": "Dataset. We use nuScenes [5], inputting 2 seconds of positions to forecast vehicle positions 6 seconds ahead. Specifically, our objective is to predict the " + }, + { + "bbox": [ + 313, + 403, + 555, + 462 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 313, + 403, + 555, + 462 + ], + "type": "text", + "content": " coordinates of each agent, exclusively considering vehicles as agents. We use the trajdata interface [23] to load and visualize the data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "text", + "content": "Task-specific considerations. Instead of discretizing the " + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 313, + 462, + 556, + 629 + ], + "type": "text", + "content": " position space, we discretize the motion, resulting in discrete velocity or acceleration tokens. These integer tokens are projected to the transformer hidden dimension using the Llama token embedding layer. Inputting only these tokens results in our PAR model knowing what speed the other agents are going at, but not where they are. It is important the model has this awareness (it should know if two agents are going to collide), so our model needs to reason over this second modality of location. We implement this by passing locations relative to the agent we are predicting into a sincosine positional embedding (see details in Sec. 12.1), which we denote a location positional encoding (LPE). The LPE is summed to our token embeddings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "We use a cross-entropy classification loss on our discrete tokens: " + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = E_{y\\sim p(y)}[-\\log (p(\\mathbf{s}_{t_\\pi}^T))]" + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": ". We use the standard average displacement error (ADE) and final displacement error (FDE) to evaluate our predicted trajectories. For our baselines (Sec. 3.3), we use the closest agent at the current timestep for Multiagent NN and Mirror. For NN and Multiagent NN we use MSE as the distance metric." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 726, + 318, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 726, + 318, + 735 + ], + "spans": [ + { + "bbox": [ + 293, + 726, + 318, + 735 + ], + "type": "text", + "content": "17" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 286, + 155 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 286, + 155 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 286, + 155 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 286, + 155 + ], + "type": "table", + "html": "
Token typeLPEMethodADE ↓FDE ↓
VelocityX1-agent AR1.503.64
VelocityX3-agent PAR1.453.51
AccelerationX1-agent AR1.443.57
AccelerationX3-agent PAR1.403.44
Accleration✓3-agent PAR1.353.34
", + "image_path": "673e05d349b3b1c65bda8705ea689558450412e26a74c5e1045ed7631929ba5d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 76, + 219, + 276, + 291 + ], + "blocks": [ + { + "bbox": [ + 55, + 163, + 295, + 208 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 295, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 295, + 208 + ], + "type": "text", + "content": "Table 3. Car trajectory prediction performance. Using acceleration tokens and 3-agent PAR results in a stronger performance over velocity tokens and single-agent AR. Adding location via a positional encoding (LPE) further improves results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 219, + 276, + 291 + ], + "lines": [ + { + "bbox": [ + 76, + 219, + 276, + 291 + ], + "spans": [ + { + "bbox": [ + 76, + 219, + 276, + 291 + ], + "type": "table", + "html": "
BaselineAgentsADE ↓FDE ↓
Random Trajectory18.8916.51
NN11.804.13
Multiagent NNN6.4012.04
MirrorN11.5914.93
", + "image_path": "0def927a57bd7e09893faf6f0d0db1bfcfb50997c4e09bbe28b32091a5d89d50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 295, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 333 + ], + "type": "text", + "content": "Table 4. Car trajectory prediction baselines. Nearest neighbor performs best overall, but our learned single-agent AR models outperform all baselines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 337, + 296, + 466 + ], + "blocks": [ + { + "bbox": [ + 59, + 337, + 296, + 466 + ], + "lines": [ + { + "bbox": [ + 59, + 337, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 59, + 337, + 296, + 466 + ], + "type": "image", + "image_path": "5d13e6894e58ebababc4444c6e8055ffd8d0bcba1d06a06f5a548822a59ab381.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 473, + 295, + 572 + ], + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 572 + ], + "type": "text", + "content": "Figure 6. Example results from our single-agent AR model (top row) and three-agent PAR model with location positional encoding (bottom row) on nuScenes. The predicted agent's ground truth trajectory is in pink, and the predicted future in blue. For the PAR model, the other two agents' ground truth states are in green. Qualitatively, the PAR model handles situations where single-agent predictions might lead to collisions (A, B), uses other agents' behavior to better adhere to road areas (A, C) without environment data, and predicts based on the speed changes of other cars (D)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 318, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 361 + ], + "type": "text", + "content": "Figure 8. Translation forecasting qualitative result on test set. 3D predictions are projected onto the image, isolating translation results by showing the ground-truth rotation. Using the PAR framework (right) instead of AR (left) improves object pose prediction." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 576, + 112, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 112, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 112, + 588 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "We train AR and 3-agent PAR models using velocity tokens, acceleration tokens, and acceleration tokens combined with our location positional encoding. The results can be seen in Table 3. Note that the 3-agent PAR model uses the agent ID embedding and next timestep prediction. Acceleration tokens consistently outperform velocity tokens both for agent AR and 3-agent PAR models. This could be because the vocabulary size for acceleration tokens is much smaller and therefore easier to optimize. Regardless, both ways of tokenizing result in models that outperform our baselines (see" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 430, + 137 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 430, + 137 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 430, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 430, + 137 + ], + "type": "image", + "image_path": "def4b1307324993288b96bbab1b5936cf16c7eabc70bf2c40351dbca37af15e4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 140, + 395, + 150 + ], + "lines": [ + { + "bbox": [ + 351, + 140, + 395, + 150 + ], + "spans": [ + { + "bbox": [ + 351, + 140, + 395, + 150 + ], + "type": "text", + "content": "1-Agent AR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 439, + 70, + 553, + 137 + ], + "blocks": [ + { + "bbox": [ + 439, + 70, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 439, + 70, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 439, + 70, + 553, + 137 + ], + "type": "image", + "image_path": "cfbf7d426bc29faf021bd3a45c9e36115c17a0a2b6199eff482202ed58baa139.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 140, + 521, + 150 + ], + "lines": [ + { + "bbox": [ + 471, + 140, + 521, + 150 + ], + "spans": [ + { + "bbox": [ + 471, + 140, + 521, + 150 + ], + "type": "text", + "content": "2-Agent PAR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 430, + 298 + ], + "blocks": [ + { + "bbox": [ + 313, + 157, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 213 + ], + "type": "text", + "content": "Figure 7. Rotation forecasting qualitative result on test set. 3D predictions are projected onto the image, isolating rotation results by showing the ground-truth translation. Incorporating the hand agent in the PAR framework (right) improves object pose prediction over object-only AR (left)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 230, + 430, + 298 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 430, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 430, + 298 + ], + "type": "image", + "image_path": "42d64316489f0226b383bb7ee1615551eff970998cd7c5f20176090569f35980.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 300, + 395, + 311 + ], + "lines": [ + { + "bbox": [ + 351, + 300, + 395, + 311 + ], + "spans": [ + { + "bbox": [ + 351, + 300, + 395, + 311 + ], + "type": "text", + "content": "1-Agent AR" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 439, + 230, + 553, + 298 + ], + "blocks": [ + { + "bbox": [ + 439, + 230, + 553, + 298 + ], + "lines": [ + { + "bbox": [ + 439, + 230, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 439, + 230, + 553, + 298 + ], + "type": "image", + "image_path": "b427a6868d2dc4570d1314493ae331ddef2aa66f616355d44fb90648f2130b8b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 300, + 521, + 311 + ], + "lines": [ + { + "bbox": [ + 471, + 300, + 521, + 311 + ], + "spans": [ + { + "bbox": [ + 471, + 300, + 521, + 311 + ], + "type": "text", + "content": "2-Agent PAR" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 376, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 495 + ], + "type": "text", + "content": "Table 4 - NN has a relatively low error on this dataset), and highlight that our framework is flexible such that a user can experiment with different ways of representing entities. For both token types, the 3-agent PAR model that is blind to location outperforms the AR model. While location information should help the model, it is possible that simply knowing whether other agents are slowing down or accelerating can help the model make better predictions. When adding location information via the LPE to our 3-agent PAR model, we see another performance gain in ADE and FDE." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "type": "text", + "content": "Qualitative examples of the AR model (top row) and 3-agent location-aware PAR model (bottom row) can be seen in Figure 6. Our method uses no image or environment data (e.g., lanes) as input. However, by reasoning over multiple agents, its predictions lead to fewer collisions and better reasoning about speed changes and driveable areas based solely on other agents' behaviors." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 604, + 555, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 632 + ], + "type": "text", + "content": "6. Case Study 3: Object Pose Forecasting During Hand-Object Interaction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Our final case study explores how hand-object interaction can be leveraged for object pose estimation. We define the human hand and the interacting object as two agents, with tokens representing distinct state types. We show that our PAR framework can jointly model these agents, improving 3D translation and rotation predictions for the object." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "spans": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "type": "text", + "content": "12408" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 171, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 171, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 171, + 85 + ], + "type": "text", + "content": "6.1. Experimental Setup" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 198 + ], + "type": "text", + "content": "Dataset. We use the DexYCB dataset, which includes 1000 videos of 10 subjects performing object manipulation tasks with 20 distinct objects from the YCB-Video dataset. The data is split into 800 training, 40 validation, and 160 testing videos. We use one of 8 provided camera views. In each trial, subjects pick up and lift objects in randomized conditions. Labels include the object's SO(3) rotation and 3D translation, and the hand's 3D translation. We focus on predicting the object's rotation or translation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 373 + ], + "type": "text", + "content": "Task-specific considerations. We tokenize object information in object-only experiments and both object and hand information in hand-object experiments. The object is represented as a 4D token for rotation forecasting (quaternion from SO(3) rotation) or a 3D token for translation forecasting (Euclidean coordinates). In hand-object experiments, the hand token is included with a 3D translation vector, and agent ID embeddings distinguish between the hand and object. Normalization is applied to all 3D translation vectors in both AR and PAR experiments; quaternions are normalized by definition and require no additional processing. An embedding layer projects the tokens into the transformer's hidden dimension, and another layer projects them back for prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "text", + "content": "For rotation-only forecasting, the loss is " + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rot} = 1 - |\\hat{q}\\cdot q|" + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "text", + "content": " is the predicted quaternion and " + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "text", + "content": " the ground-truth quaternion. For translation-only forecasting, the loss " + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "text", + "content": " is the mean squared error (MSE) between predicted and ground-truth translations. For PAR we predict relative object-to-hand translations at each frame, using the current hand position as origin, while for AR, we predict absolute object translations without considering the interacting agent. For PAR models, we add the loss " + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "text", + "content": ", a MSE on hand translation. The object-only AR rotation model is optimized with " + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rot}" + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "text", + "content": ", while the PAR rotation model combines " + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rot} + \\mathcal{L}_h" + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "text", + "content": "; similarly, the object-only translation model is trained with " + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "text", + "content": ", and the hand-object translation model uses " + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t + \\mathcal{L}_h" + }, + { + "bbox": [ + 55, + 374, + 296, + 579 + ], + "type": "text", + "content": ". At inference, the first half of each video is provided, and object predictions are autoregressively generated for the second half. Translation is evaluated using MSE, while rotation is measured using geodesic distance (GEO) on SO(3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 587, + 113, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 113, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 113, + 598 + ], + "type": "text", + "content": "6.2. Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "We compare the object-only AR models to the hand-object PAR models in Table 5 for the two prediction tasks. We also present the baselines described in Sec. 3.3 in Table 6. Figures 7 and 8 show qualitative results on the rotation and translation predictions, respectively. In both prediction tasks, we observe that incorporating the human hand's interaction with the object enhances accuracy: for rotation, PAR results in a relative improvement of " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " over AR, and for translation, " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "41\\%" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ". See Section 11.2 for additional qualitative results" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 556, + 148 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 556, + 148 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 556, + 148 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 556, + 148 + ], + "type": "table", + "html": "
TypeMethodMSE ↓GEO (rad) ↓
Translation1-agent AR3.68 × 10-3-
Translation2-agent PAR2.17 × 10-3-
Rotation1-agent AR-0.919
Rotation2-agent PAR-0.837
", + "image_path": "81446936dd837ebf540ac9ca60064800d69fc4e79940603f020fcb15f61cb769.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 201, + 565, + 271 + ], + "blocks": [ + { + "bbox": [ + 313, + 156, + 555, + 190 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 190 + ], + "type": "text", + "content": "Table 5. Test set results on DexYCB dataset. For both rotation and translation forecasting, the 2-agent PAR model, which treats the hand as an additional agent, improves results." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 201, + 565, + 271 + ], + "lines": [ + { + "bbox": [ + 315, + 201, + 565, + 271 + ], + "spans": [ + { + "bbox": [ + 315, + 201, + 565, + 271 + ], + "type": "table", + "html": "
BaselineTranslation - MSE (m2) ↓Rotation - GEO (rad) ↓
Random0.2442.196
Random Trajectory1.60 × 10-22.146
NN1.69 × 10-22.179
Multiagent NN1.71 × 10-22.170
Mirror1.20 × 10-2-
", + "image_path": "f6d17556517a6feaca8c2dde64dcd81c81391a9aac6cb2d88bb053810ce1d8f8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 280, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 313 + ], + "type": "text", + "content": "Table 6. Test set results for DexYCB baselines. We cannot provide rotation results for the Mirror baseline, because the ground-truth does not include hand rotation, only 3D translation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 335, + 425, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 425, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 425, + 346 + ], + "type": "text", + "content": "with more sampled frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 357, + 383, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 383, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 383, + 370 + ], + "type": "text", + "content": "7. Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 378, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 473 + ], + "type": "text", + "content": "This work introduced the Poly-Autoregressive (PAR) framework, a unifying approach to prediction for multi-agent interactions. By applying the same transformer architecture and hyperparameters across diverse tasks, including action forecasting in social settings, trajectory prediction for autonomous vehicles, and object pose forecasting during hand-object interaction, we have demonstrated the versatility and robustness of our framework." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 475, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 641 + ], + "type": "text", + "content": "Our findings underscore the crucial importance of considering the influence of multiple agents in a scene for prediction tasks. By modeling interactions, we significantly improved prediction accuracy over single-agent approaches on all three problems we considered. While we achieved promising results with a simple architecture, we have only provided a starting point that can be built upon extensively. For instance, incorporating environmental context or tokenizing pixel patches, especially as a way to relax our assumption on high-quality tracking, are avenues for further research using PAR. It would be interesting to experiment with scaling the data and model. We do not specifically consider the relative importance of neighboring agents, which is an interesting future research direction (ex. dynamic attention mechanism)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "Our PAR framework provides a simple and generalizable foundation of universal building blocks, ready for extension or refinement in future tasks. The PAR framework holds potential for advancing AI systems, enhancing prediction capabilities and enabling more accurate navigation and operation in real-world, multi-agent interactions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "spans": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "type": "text", + "content": "12409" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "8. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 169, + 297, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 297, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 297, + 234 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 235, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 235, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 235, + 296, + 258 + ], + "type": "text", + "content": "[2] Fred Attneave. Some informational aspects of visual perception. Psychological review, 61(3):183, 1954. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 258, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 258, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 296, + 312 + ], + "type": "text", + "content": "[3] Murchana Baruah and Bonny Banerjee. A multimodal predictive agent model for human interaction generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 1022-1023, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 314, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 296, + 335 + ], + "type": "text", + "content": "[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 336, + 296, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 296, + 380 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 6, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 381, + 296, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 381, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 381, + 296, + 423 + ], + "type": "text", + "content": "[6] Sergio Casas, Wenjie Luo, and Raquel Urtasun. Intentnet: Learning to predict intention from raw sensor data. In Conference on Robot Learning, pages 947-956. PMLR, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 426, + 297, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 297, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 297, + 491 + ], + "type": "text", + "content": "[7] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8748-8757, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 492, + 297, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 297, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 297, + 557 + ], + "type": "text", + "content": "[8] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 558, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 296, + 602 + ], + "type": "text", + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 603, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 296, + 646 + ], + "type": "text", + "content": "[10] Baptiste Chopin, Hao Tang, Naima Otberdout, Mohamed Daoudi, and Nicu Sebe. Interaction transformer for human reaction generation. IEEE Transactions on Multimedia, pages 1-13, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "text", + "content": "[11] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, KueiDa Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "content": "We thank Jane Wu, Himanshu Singh, Georgios Pavlakos, and João Carreira for useful discussions and feedback. This work was supported by ONR MURI N00014-21-1-2801 and NSF Graduate Fellowships to NT and TS." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "text", + "content": "[12] Henggang Cui, Vladan Radosavljevic, Fang-Chieh Chou, Tsung-Han Lin, Thi Nguyen, Tzu-Kuo Huang, Jeff Schneider, and Nemanja Djuric. Multimodal trajectory predictions for autonomous driving using deep convolutional networks. In 2019 international conference on robotics and automation (icra), pages 2090-2096. IEEE, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "text", + "content": "[13] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 184, + 555, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 555, + 216 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 556, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 556, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 556, + 272 + ], + "type": "text", + "content": "[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "text", + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 555, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 555, + 370 + ], + "type": "text", + "content": "[17] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. In European Conference on Computer Vision (ECCV), page 3, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "type": "text", + "content": "[18] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In International Conference on Machine Learning, pages 1242-1250. PMLR, 2014. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 491 + ], + "type": "text", + "content": "[19] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6047-6056, 2018. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 492, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 554, + 536 + ], + "type": "text", + "content": "[20] Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13053-13064, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 555, + 591 + ], + "type": "text", + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "type": "text", + "content": "[22] Yanjun Huang, Jiatong Du, Ziru Yang, Zewei Zhou, Lin Zhang, and Hong Chen. A survey on trajectory-prediction methods for autonomous driving. IEEE Transactions on Intelligent Vehicles, 7(3):652-674, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 555, + 690 + ], + "type": "text", + "content": "[23] Boris Ivanovic, Guanyu Song, Igor Gilitschenski, and Marco Pavone. trajdata: A unified interface to multiple human trajectory datasets. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, New Orleans, USA, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "type": "text", + "content": "[24] Bolin Lai, Sam Toyer, Tushar Nagarajan, Rohit Girdhar, Shengxin Zha, James M Rehg, Kris Kitani, Kristen Grau" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 727, + 318, + 736 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 727, + 318, + 736 + ], + "spans": [ + { + "bbox": [ + 293, + 727, + 318, + 736 + ], + "type": "text", + "content": "12410" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "type": "text", + "content": "man, Ruta Desai, and Miao Liu. Human action anticipation: A survey. arXiv preprint arXiv:2410.14045, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 148 + ], + "type": "text", + "content": "[25] Hugo Larochelle and Iain Murray. The neural autoregressive distribution estimator. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 29-37. JMLR Workshop and Conference Proceedings, 2011. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 194 + ], + "type": "text", + "content": "[26] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. Deepim: Deep iterative matching for 6d pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 683-698, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 296, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 238 + ], + "type": "text", + "content": "[27] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 261 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 316 + ], + "type": "text", + "content": "[29] Siyuan Brandon Loh, Debaditya Roy, and Basura Fernando. Long-term action forecasting using multi-headed attention-based variational recurrent neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2419-2427, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 296, + 360 + ], + "type": "text", + "content": "[30] Vongani Maluleke, Lea Müller, Jathushan Rajasegaran, Georgios Pavlakos, Shiry Ginosar, Angjoo Kanazawa, and Jitendra Malik. Synergy and synchrony in couple dances. arXiv preprint arXiv:2409.04440, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "text", + "content": "[31] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 470 + ], + "type": "text", + "content": "[32] Evonne Ng, Hanbyul Joo, Liwen Hu, Hao Li, Trevor Darrell, Angjoo Kanazawa, and Shiry Ginosar. Learning to listen: Modeling non-deterministic dyadic facial motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20395-20405, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "text", + "content": "[33] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 296, + 557 + ], + "type": "text", + "content": "[34] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In ArXiv, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 296, + 624 + ], + "type": "text", + "content": "[35] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 657 + ], + "type": "text", + "content": "[36] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 680 + ], + "type": "text", + "content": "[37] Alec Radford. Improving language understanding by generative pre-training. 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "text", + "content": "[38] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "text", + "content": "[39] Ilija Radosavovic, Bike Zhang, Baifeng Shi, Jathushan Rajasegaran, Sarthak Kamat, Trevor Darrell, Koushil Sreenath, and Jitendra Malik. Humanoid locomotion as next token prediction. arXiv preprint arXiv:2402.19469, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 555, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 182 + ], + "type": "text", + "content": "[40] Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, Christoph Feichtenhofer, and Jitendra Malik. On the benefits of 3d pose and tracking for human action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 640-649, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 183, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 555, + 227 + ], + "type": "text", + "content": "[41] Jathushan Rajasegaran, Ilija Radosavovic, Rahul Ravishankar, Yossi Gandelsman, Christoph Feichtenhofer, and Jitendra Malik. An empirical study of autoregressive pre-training from videos. arXiv preprint arXiv:2501.05453, 2025. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "type": "text", + "content": "[42] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 293, + 555, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 555, + 360 + ], + "type": "text", + "content": "[43] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 555, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 555, + 425 + ], + "type": "text", + "content": "[44] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8579-8590, 2023. 2, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 426, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 555, + 448 + ], + "type": "text", + "content": "[45] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 30(1):50-64, 1951. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 555, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 492 + ], + "type": "text", + "content": "[46] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. arXiv preprint arXiv:2403.18811, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 493, + 555, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 555, + 535 + ], + "type": "text", + "content": "[47] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 536, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 555, + 591 + ], + "type": "text", + "content": "[48] Chen Sun, Abhinav Shrivastava, Carl Vondrick, Rahul Sukthankar, Kevin Murphy, and Cordelia Schmid. Relational action forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 273-283, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 592, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 555, + 658 + ], + "type": "text", + "content": "[49] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 555, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 691 + ], + "type": "text", + "content": "[50] Lucas Theis and Matthias Bethge. Generative image modeling using spatial lstms. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "spans": [ + { + "bbox": [ + 294, + 727, + 317, + 735 + ], + "type": "text", + "content": "12411" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 609 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "text", + "content": "[52] Ameni Trabelsi, Mohamed Chaabane, Nathaniel Blanchard, and Ross Beveridge. A pose proposal and refinement network for better 6d object pose estimation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2382-2391, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 296, + 217 + ], + "type": "text", + "content": "[53] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 218, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 295, + 240 + ], + "type": "text", + "content": "[54] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 296, + 296 + ], + "type": "text", + "content": "[55] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16611–16621, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 352 + ], + "type": "text", + "content": "[56] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10367-10373. IEEE, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 296, + 418 + ], + "type": "text", + "content": "[57] Bowen Wen, Jonathan Tremblay, Valts Blukis, Stephen Tyree, Thomas Müller, Alex Evans, Dieter Fox, Jan Kautz, and Stan Birchfield. Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 606-617, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 419, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 419, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 419, + 296, + 474 + ], + "type": "text", + "content": "[58] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868-17879, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 296, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 296, + 507 + ], + "type": "text", + "content": "[59] Jane Wu, Georgios Pavlakos, Georgia Gkioxari, and Jitendra Malik. Reconstructing hand-held objects in 3d. arXiv preprint arXiv:2404.06507, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 509, + 296, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 296, + 552 + ], + "type": "text", + "content": "[60] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 554, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 296, + 609 + ], + "type": "text", + "content": "[61] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 727, + 318, + 736 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 727, + 318, + 736 + ], + "spans": [ + { + "bbox": [ + 294, + 727, + 318, + 736 + ], + "type": "text", + "content": "12412" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Population Normalization for Federated Learning/b6b8182e-2812-44b6-99cb-82e1b12982cb_content_list.json b/2025/Population Normalization for Federated Learning/b6b8182e-2812-44b6-99cb-82e1b12982cb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ef054668c87eb7fedc7d83d55064e92bdc5540b0 --- /dev/null +++ b/2025/Population Normalization for Federated Learning/b6b8182e-2812-44b6-99cb-82e1b12982cb_content_list.json @@ -0,0 +1,1796 @@ +[ + { + "type": "text", + "text": "Population Normalization for Federated Learning", + "text_level": 1, + "bbox": [ + 245, + 130, + 751, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhuoyao Wang $^{1*}$ , Fan Yi $^{1*}$ , Peizhu Gong $^{1}$ , Caitou He $^{2}$ , Cheng Jin $^{1,3}$ , Weizhong Zhang $^{1,4\\dagger}$", + "bbox": [ + 145, + 178, + 872, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Fudan University $^{2}$ Wenzhou University", + "bbox": [ + 346, + 196, + 669, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Innovation Center of Calligraphy and Painting Creation Technology, MCT, China", + "bbox": [ + 183, + 215, + 831, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Shanghai Key Laboratory of Intelligent Information Processing", + "bbox": [ + 254, + 233, + 761, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{18110240064,pzgong,jc,weizhongzhang}@fudan.edu.cn，ftyi21@m.fudan.edu.cn，hctou89@wzu.edu.cn", + "bbox": [ + 99, + 253, + 913, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Batch normalization (BN) is widely recognized as an essential method in training deep neural networks, facilitating convergence and enhancing model stability. However, in Federated Learning (FL) contexts, where training data are typically heterogeneous and clients often face resource constraints, the effectiveness of BN is considerably limited for two primary reasons. First, the population statistics, specifically the mean and variance of these heterogeneous datasets, vary substantially, resulting in inconsistent BN layers across client models, which ultimately drives these models to diverge further. Second, estimating statistics from a mini-batch is often imprecise since the batch size has to be small in resource-limited clients. This paper introduces Population Normalization, a novel technique for FL, in which the statistics are learned as trainable parameters rather than calculated from mini-batches as in BN. Thus, our normalization layers are homogeneous among the clients and the adverse impact of small batch size is eliminated as the model can be well-trained even when the batch size equals to one. To enhance the flexibility of our method in practical applications, we investigate the role of stochastic uncertainty in BN's statistical estimation. When larger batch sizes are available, we demonstrate that injecting simple artificial noise can effectively mimic this stochastic uncertainty and improve the model's generalization capability. Experimental results validate the efficacy of our approach across various FL tasks.", + "bbox": [ + 89, + 335, + 485, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 770, + 220, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Batch Normalization (BN) [17] has become a standard component in deep neural networks (DNNs). It is proposed to stabilize and accelerate the training of DNNs, and the original motivation was to reduce the so-called internal covariate shift. In each training step, it normalizes the layer input of a neural network using the mean and variance com", + "bbox": [ + 89, + 794, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "puted within a randomly sampled mini-batch. Specifically, consider a particular neuron in a network, and we denote $\\{x_{1},\\ldots ,x_{m}\\}$ to be its pre-activations on a mini-batch $\\mathcal{B}$ of size $m$ , BN first normalizes each $x_{i}$ as", + "bbox": [ + 511, + 305, + 906, + 361 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {i} = \\frac {x _ {i} - \\mu_ {\\mathcal {B}}}{\\sigma_ {\\mathcal {B}}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 369, + 906, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\mu_{\\mathcal{B}}$ and $\\sigma_{\\mathcal{B}}^{2}$ are the sample mean and variance of the pre-activation on $\\mathcal{B}$ . It then scales and shifts each $\\hat{x}_i$ as $y_i = \\alpha \\hat{x}_i + \\beta$ to compensate the effect of normalization, where $\\gamma$ and $\\beta$ are two parameters to be learned during the training process. BN finally takes $y_i$ as the new pre-activation of this neuron. It has been reported in literature [12, 16, 17, 37, 40] that BN can significantly improve the convergence of training and the generalization performance of the final learned models.", + "bbox": [ + 511, + 401, + 906, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, we notice that the effectiveness of BN is compromised in Federated Learning [14, 19, 42], which typically aims at collaboratively training models on resource limited clients (e.g., mobiles) with locally private and heterogeneous training data. The main reasons can be summarized as follows. One is that the statistics, i.e., the mean and variance, of heterogeneous datasets in the clients differ greatly. Eqn. (1) implies that this can lead to inconsistent BN layers among the client models and finally causes these models to drift further away from each other. The other is that in resource limited clients, the mini-batch size has to be small, which makes $\\mu_{\\mathcal{B}}$ and $\\sigma_{\\mathcal{B}}^{2}$ inaccurate estimations of the population mean and variance for normalization and the obtained BN would cause instability during training that slows down convergence.", + "bbox": [ + 511, + 537, + 908, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some other normalization techniques, such as batch renormalization [16], group normalization [40], and layer normalization [2], are proposed to address this heavy reliance on batch size. The key ideas are to estimate the population mean and variance as accurately as possible by using techniques such as moving average [16] or to normalize the activation along other dimensions such as the channel dimension [2]. Nevertheless, these approaches focus on training acceleration, which typically under-perform", + "bbox": [ + 511, + 765, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution, † Corresponding authors.", + "bbox": [ + 114, + 886, + 362, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10214", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "BN in terms of the trained model's accuracy [40]. Recently, FedBN [25] is proposed to extend BN to FL. It keeps the client BN layers updated locally, without communicating and aggregating them during training. This enables the clients to achieve better performances in their own respective environments. However, since the BN layers among the clients can be heterogeneous due to the non-IID datasets, FedBN cannot produce an off-the-shelf global model after training, which is often required in reality. Consequently, FedBN suffers from the cold start issue, i.e., the learned model cannot be directly deployed in the new clients due to the lack of trained local BN layers. What's more, FedBN still relies on estimation within mini-batches, which cannot address the issue of inaccurate estimation caused by small batch size in cross-device FL.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Population Normalization (PN), a simple yet effective normalization method for FL to eliminate the negative impacts of statistical heterogeneity and estimation inaccuracies on BN. The key idea is to learn population statistics as trainable parameters in the training process by introducing two constraints derived from the definitions of mean and variance (see Section 3), instead of estimating them from the mini-batches as most existing methods have done. As the population statistics take place of the sample mean and variance of the local datasets in BN, and are treated as normal trainable parameters in local training and aggregation, our normalization layers among the clients would be consistent although the datasets are heterogeneous. This enables our method to eventually produce a global model with the normalization layer among the clients. In addition, since we no longer need to calculate the sample mean and variance, the resulting training problem can be solved by stochastic algorithms of FL with arbitrary batch sizes (even equals to one) and thus the issue of inaccurate statistic estimation because of small batch size is eliminated.", + "bbox": [ + 91, + 321, + 483, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In practice, we observe that PN tends to under-perform BN given very large batch sizes. We then further explore this issue and find its cause to be that PN lacks the stochastic uncertainty in the estimation of mean and variance. Specifically, we show that the noise introduced by the mini-batch based mean and variance estimations in the BN is essential for its good generalization behavior. Our empirical results demonstrate that injecting a reasonable amount of data independent noise into BN with extremely large batch sizes can significantly improve the generalization performance to be comparable to that of BN with proper batch sizes. Inspired by this finding, we propose Noisy Population Normalization (NPN) by injecting simple data independent artificial noises into the learned population mean and variance in PN during training. The injected noise can serve the same purpose as the uncertainty in estimating population statistics in BN with small batch sizes. This makes the proposed", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method more flexible in the applications of FL where large batch size is available, e.g., collaboration among companies in cross-silo FL [9, 15, 21].", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we conduct a series of experiments to verify the effectiveness of our proposed PN and NPN in the context of FL with small and large batch sizes, respectively (Section 5). The results demonstrate that PN effectively reduces the reliance on training batch size and alleviates the negative impact of data heterogeneity on normalization. Furthermore, the artificial noise injected in our NPN method effectively improves the generalization performance given large training batch sizes.", + "bbox": [ + 511, + 137, + 906, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 529, + 273, + 872, + 287 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a normalization technique PN for federated learning, which demonstrates superior performance given small batch sizes and data heterogeneity.", + "- In training with large batch sizes, we propose NPN based on PN, which introduces stochastic uncertainty by injecting artificial noise to improve generalization ability.", + "- We conduct extensive experiments to verify the effectiveness of our method." + ], + "bbox": [ + 511, + 290, + 903, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notations: Given a positive integer $d$ , we define $[d]$ to be the set $\\{1,2,\\dots,d\\}$ . For a vector $x\\in \\mathbb{R}^d$ , we let $[x]_i$ be its $i$ -th component with $i\\in [d]$ . We denote by $w_{i,j}$ the element in the $i$ -th row and the $j$ -th column of a real-valued matrix $w$ . Let $A\\cup B$ be the union of two sets $A$ and $B$ . Moreover, we denote by $\\mathcal{N}(\\mu ,\\sigma^2)$ the Gaussian distribution with mean of $\\mu$ and variance of $\\sigma^2$ .", + "bbox": [ + 511, + 411, + 905, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 534, + 653, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Normalization Techniques", + "text_level": 1, + "bbox": [ + 511, + 559, + 751, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "BN [17] is a standard normalization technique, which enables the training of very deep neural networks, and the most representative result is ResNet[11], which cannot be easily trained to converge without BN. It can also achieve significant imporement in generalization. Therefore, the invention of BN is a major milestone in the development of deep learning.", + "bbox": [ + 511, + 582, + 905, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As aforementioned, small batch size or non-iid minibatch can diminish the effect of BN on the training acceleration or even result in training failure, due to the inaccurate mean and variance estimations [16]. Moreover, large batch size leads to huge memory cost. This issue hinders BN's usage in FL applications, where the clients often have limited computational resources and the datasets are usually heterogeneous. This in turn prevents the training of large networks in FL. Recently, some new normalization methods have been proposed as alternatives to BN, and they can be roughly divided into two categories, i.e., normalization along the batch dimension [16], and mini-batch independent normalization [2, 31, 37, 40]. The former alleviates the reliance on the batch size by using techniques such as", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10215", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "moving average. The latter ones achieve this by normalizing the activation of neurons in a network along other dimensions. For example, Group Normalization [40] divides the channels in each hidden layer into groups and normalizes the activations in each group with its mean and variance. Empirical studies show that all these methods achieve comparable training speedups with BN. However, the generalization performance they achieved is inferior to that of BN.", + "bbox": [ + 89, + 90, + 483, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides the above studies, there is another line of research [4, 26, 29, 32, 39] that focuses on understanding BN. One of the main conclusions in these investigations is that with BN, the effective objective function used in training is different from the original one. Specifically, [32] showed that by inserting BN layers into a network, the loss landscape becomes smoother, which naturally explains why BN enables one to use large learning rates. The studies [29, 39] focus on the regularization effect of BN. They showed that the objective function with BN can be divided into two parts: an empirical loss under population normalization and an explicit regularizer named gamma decay.", + "bbox": [ + 89, + 227, + 483, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Federated Learning", + "text_level": 1, + "bbox": [ + 89, + 417, + 281, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FL emerges as an effective paradigm to train models on distributed devices without centralizing local private data [3, 8, 13, 19, 42]. Specifically, it can be formulated as follows:", + "bbox": [ + 89, + 438, + 483, + 481 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathcal {L} (\\theta , \\mathcal {D}) \\triangleq \\sum_ {m = 1} ^ {M} \\rho_ {m} \\mathcal {L} _ {m} (\\theta , \\mathcal {D} _ {m}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 489, + 482, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M$ is the number of clients and $\\theta$ is the model to be learnt. The global dataset $\\mathcal{D} = \\cup_{m=1}^{M}\\mathcal{D}_{m}$ , where $\\mathcal{D}_{m} = \\{(\\pmb{x}_{i},y_{i})\\}_{i=1}^{|D_{m}|}$ is the locally kept dataset on client $m$ . $\\mathcal{L}_{m}(\\theta,\\mathcal{D}_{m}) = \\frac{1}{|\\mathcal{D}_{m}|}\\sum_{\\xi\\in\\mathcal{D}_{m}}\\ell(\\theta,\\xi)$ denotes the local empirical risk with the loss function $\\ell(\\cdot,\\cdot)$ and $\\rho_{m} = \\frac{|D_{m}|}{|\\mathcal{D}|}$ . Typically, the most popular FL algorithm FedAvg [30] solves the problem (2) based on local-SGD [34]. Specifically, in the $c^{th}$ round, it updates the global model on the server as follows:", + "bbox": [ + 89, + 536, + 483, + 680 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {c + 1} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\rho_ {m} \\theta_ {m} ^ {c}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 688, + 482, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{m}^{c}$ denotes the received local model from client $m$ after local training in this round. The updated global model $\\theta^{c + 1}$ is then distributed to the clients to initiate the $(c + 1)$ -th round of training.", + "bbox": [ + 89, + 734, + 482, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In FL, especially in mobile edge scenarios, both limited device capacity and non-IID distribution of clients' local data raise challenges to the training procedure [27]. Various works attempt to tackle the non-IID distribution of data for federated learning from the perspective of local training regularization [18, 22, 23], server aggregation [5, 28, 38, 41, 43], personalized federated learning [10, 20, 24, 35],", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Population Normalization", + "text_level": 1, + "bbox": [ + 514, + 90, + 777, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input: activation: $x$ ; mean and variance: $\\mu$ and $1 / \\gamma^2$ ; scale and shift parameters: $\\alpha$ and $\\beta$ .", + "bbox": [ + 527, + 109, + 906, + 141 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Normalize using the population statistics:", + "bbox": [ + 529, + 142, + 805, + 156 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = \\frac {x - \\mu}{\\sqrt {\\gamma^ {2} + \\epsilon}}\n$$\n", + "text_format": "latex", + "bbox": [ + 669, + 165, + 764, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scale and shift:", + "bbox": [ + 529, + 209, + 635, + 222 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = \\alpha \\hat {x} + \\beta \\equiv \\mathrm {P N} _ {\\alpha , \\beta} ^ {\\mu , \\gamma} (x)\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 233, + 795, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Return: y", + "text_level": 1, + "bbox": [ + 529, + 273, + 604, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "etc. However, they do not explore the reduced effectiveness of batch normalization in federated learning. FedBN [25] uses local BN parameters to tackle non-iid federated learning. However, FedBN is only applicable to local inferences where only clients participating in the training with local statistics can perform inference, without providing a global model for new clients or global data. Also, FedBN does not consider the small batches caused by limited local resources.", + "bbox": [ + 511, + 324, + 906, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Population Normalization", + "text_level": 1, + "bbox": [ + 511, + 479, + 754, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now present our method PN (see Alg.1). Its key idea is to eliminate the impacts of heterogeneous statistics and inaccurate statistic estimations by learning the population mean and variance during the training process, instead of estimating them within the mini-batch as BN has done. Precisely, given an activation $x$ , we normalize it as follows:", + "bbox": [ + 511, + 506, + 906, + 590 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = \\frac {x - \\mu}{\\sqrt {\\gamma^ {2} + \\epsilon}},\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 606, + 758, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu$ and $\\gamma^2$ are the population mean and variance, respectively, and they will be learned during the training process, of which the details are specified later in Eqn.(3). $\\epsilon > 0$ is a small constant. Then, to compensate the effect of normalization on the effective capacity of a neural network, we scale and shift the normalized activation $\\hat{x}$ as:", + "bbox": [ + 511, + 652, + 906, + 743 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = \\alpha \\hat {x} + \\beta \\equiv \\mathrm {P N} _ {\\alpha , \\beta} ^ {\\mu , \\gamma} (x), \\quad (\\mathrm {P N})\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 758, + 906, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the coefficients $\\alpha$ and $\\beta$ are learned in the training process together with $\\mu$ and $\\gamma$ .", + "bbox": [ + 511, + 791, + 905, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Below, we take a fully connected neural network $f(\\pmb{x}) \\in \\mathbb{R}$ as an example to show how to insert PN layers into DNNs. Specifically, given an input $\\pmb{x} \\in \\mathbb{R}^d$ , let the input layer be $f_j^{(0)}(\\pmb{x}) = [\\pmb{x}]_j$ with $j \\in [n^{(0)}] = [d]$ . The $\\ell$ -th hidden layer containing $n^{(\\ell)}$ hidden nodes with $\\ell \\in [L]$ takes", + "bbox": [ + 511, + 823, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10216", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the form of", + "bbox": [ + 89, + 90, + 171, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f _ {j} ^ {(\\ell)} (\\boldsymbol {x}) = h ^ {(\\ell)} \\left(z _ {j} ^ {(\\ell)} (\\boldsymbol {x})\\right) \\text {w i t h} \\tag {4} \\\\ z _ {j} ^ {(\\ell)} (\\boldsymbol {x}) = \\sum_ {k = 1} ^ {n ^ {(\\ell - 1)}} w _ {j, k} ^ {(\\ell)} \\underbrace {\\left(\\alpha_ {k} ^ {(\\ell - 1)} \\frac {f _ {k} ^ {(\\ell - 1)} (\\boldsymbol {x}) - \\mu_ {k} ^ {(\\ell - 1)}}{\\sqrt {\\left(\\gamma_ {k} ^ {(\\ell - 1)}\\right) ^ {2} + \\epsilon}} + \\beta_ {k} ^ {(\\ell - 1)}\\right)} _ {(\\ell - 1) - \\text {t h P N l a y e r}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 108, + 483, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $j\\in [n^{(\\ell)}],z_j^{(\\ell)}(\\pmb {x})$ is the pre-activation, $h^{(\\ell)}$ is the activation function and $w^{(\\ell)}\\in \\mathbb{R}^{n^{(\\ell)}\\times n^{(\\ell -1)}}$ is the weight matrix connecting layer $\\ell$ and $\\ell -1$ $\\mu_k^{(\\ell -1)}$ and $\\gamma_{k}^{(\\ell -1)}$ are the population mean and standard deviation of the hidden nodes $f_{k}^{(\\ell -1)}(\\pmb {x})$ $\\alpha_{k}^{(\\ell -1)}$ and $\\beta_{k}^{(\\ell -1)}$ are the parameters of the linear transformation. The final output layer is defined as", + "bbox": [ + 89, + 219, + 483, + 330 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (\\boldsymbol {x}) = \\sum_ {k = 1} ^ {n ^ {(L)}} u _ {k} \\underbrace {\\left(\\frac {f _ {k} ^ {(L)} (\\boldsymbol {x}) - \\mu_ {k} ^ {(L)}}{\\sqrt {\\left(\\gamma_ {k} ^ {(L)}\\right) ^ {2} + \\epsilon}}\\right)} _ {\\text {f i n a l P N l a y e r}} + c _ {0}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 335, + 482, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $u\\in \\mathbb{R}^{n^{(L)}}$ is the weight vector connecting the output to the $L$ -th hidden layer, $\\mu_k^{(L)}$ and $\\gamma_k^{(L)}$ are the population mean and standard deviation of the hidden nodes $f_{k}^{(L)}(\\pmb {x})$ The scaling and shifting parameters are absorbed into $u_{k}$ and $c_{0}$ , respectively.", + "bbox": [ + 89, + 420, + 482, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To eliminate the dependency of PN on batch size, we learn the statistics $\\mu_k^{(\\ell)}$ and $\\gamma_k^{(\\ell)}$ in training by introducing two natural constraints derived from the definition of mean and variance. The statistic estimation is optimized together with the model parameters during local training, and aggregated on the server by FL algorithms, such as FedAVG. To be precise, given the loss $\\phi(\\cdot, \\cdot): \\mathbb{R} \\times \\{0,1\\} \\to \\mathbb{R}$ and the global dataset $\\mathcal{D} = \\cup_{m=1}^{M} \\mathcal{D}_m$ , the training objective in Eqn. (2) of a FL problem becomes", + "bbox": [ + 89, + 503, + 483, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta} \\mathcal {L} (\\theta , \\mathcal {D}) s. t. \\left\\{ \\begin{array}{l} \\frac {1}{| \\mathcal {D} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\mathcal {D}} \\frac {\\left(f _ {k} ^ {(\\ell)} (\\boldsymbol {x}) - \\mu_ {k} ^ {(\\ell)}\\right) ^ {2}}{\\left(\\gamma_ {k} ^ {(\\ell)}\\right) ^ {2} + \\epsilon} = 1 \\\\ \\frac {1}{| \\mathcal {D} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\mathcal {D}} f _ {k} ^ {(\\ell)} (\\boldsymbol {x}) = \\mu_ {k} ^ {(\\ell)} \\end{array} \\right., \\\\ k \\in [ m ^ {(\\ell)} ], \\ell \\in \\{0 \\} \\cup [ L ] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 647, + 475, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ is the set of all the trainable parameters, i.e., $\\theta = \\{w^{(\\ell)}\\} \\cup \\{u\\} \\cup \\{\\mu^{(\\ell)}\\} \\cup \\{\\gamma^{(\\ell)}\\} \\cup \\{\\alpha^{(\\ell)}\\} \\cup \\{\\beta^{(\\ell)}\\} \\cup \\{c_0\\}$ and $\\mathcal{L}(\\theta ,\\mathcal{D}) = \\sum_{m = 1}^{M}\\rho_{m}\\sum_{(\\boldsymbol {x},y)\\in \\mathcal{D}_{m}}\\phi (f(\\boldsymbol {x}),y)$ . This training problem can be solved by various stochastic algorithms with arbitrary batch sizes (even equals to 1), such as moving average or the method of Lagrangian multipliers. In the experiments of this paper, we use stochastic Lagrangian multipliers and the details can be found in the appendix.", + "bbox": [ + 89, + 732, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. (1) Since the statistics $\\mu$ and $\\gamma$ are learned in the training process instead of being estimated from minibatches, the resulting training problem can be solved by", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "stochastic solvers of FL with arbitrary batch sizes (even equals to 1). Thus the reliance on batch size is eliminated. This enables resource limited clients to train larger deep neural networks. (2) Both $\\mu$ and $\\gamma$ are treated as normal trainable parameters in local training, communication and aggregation, our PN layers are consistent among the clients although the datasets on the clients are heterogeneous. This enables us to eventually obtain an off-the-shelf global model, which is often required in reality.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 3.1 We should point out that our PN layer can also be inserted before the activation functions.", + "bbox": [ + 511, + 234, + 906, + 266 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a58edbede5168160128fe734085f58f0965127749be6532017594e7fe1aa878f.jpg", + "image_caption": [ + "Figure 1. Training loss (left) and test loss (right) of BN, LBN and LBN+noise. Injecting artificial noise into estimated BN statistics successfully decreases the test loss and improves generalization." + ], + "image_footnote": [], + "bbox": [ + 519, + 287, + 694, + 388 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/81a9c9f0b96329886b774af40745de8efbfc5465b307444165d5496593bd56dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 287, + 885, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Noisy Population Normalization", + "text_level": 1, + "bbox": [ + 511, + 452, + 805, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that PN demonstrates inferior performance relative to BN when batch size becomes larger. We explore this issue and find its cause to be that PN lacks stochastic uncertainty in the mean and variance estimations. In practice, large batch sizes may often be used in cross-silo federated learning, where the participants are companies and organizations with huge amounts of data as well as abundant computational resources [9, 15, 21]. Therefore, to enable our method to be more flexibly used in such scenarios, we propose NPN, which injects artificial noise into the learned statistics during training to boost generalization. The details are presented in the following subsections.", + "bbox": [ + 511, + 477, + 906, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. The Role of Stochastic Uncertainty in BN", + "text_level": 1, + "bbox": [ + 511, + 667, + 864, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We provide an in-depth exploration into the role of stochastic uncertainty of the mini-batch statistics in BN. This provides a deep understanding of BN's batch size reliance.", + "bbox": [ + 511, + 688, + 905, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start with a simple experiment to show the effect of batch size on BN. Precisely, we compare the performances of BN with a reasonable batch size (m=128) and a very large size (m=2048) in training VGG-16 on CIFAR-10. For simplicity, we refer to BN with m=2048 as LBN. For fair comparison, in LBN, to eliminate the effect of large batch size on SGD, we only use large batches to obtain batch statistics and calculate the stochastic gradient with respect to the loss on mini-batches of size 128.", + "bbox": [ + 511, + 734, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present both the training and test losses of the learned models with BN and LBN over the training process in Fig", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10217", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1, from which we can see that LBN converges faster than BN, due to the more accurate population statistics estimation. However, there is a noticeable gap between their test losses, which becomes larger after the learning rate decays. Therefore, the effect of BN in improving generalization is reduced when the batch size becomes too large. This indicates that the stochastic uncertainty in estimates of the mean and variance is important for BN.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The next question we want to investigate is whether we can artificially inject a proper amount of data-independent noise into LBN to improve its generalization performance. Note that the noise of estimating $\\mu_{\\mathcal{B}}$ and $\\sigma_{\\mathcal{B}}$ in BN is data dependent. To answer this question, we should first calculate the uncertainty explicitly. The studies [29, 36] show that for sufficiently large batch size M, the estimated $\\mu_{\\mathcal{B}}$ and $\\sigma_{\\mathcal{B}}$ can be viewed as two random variables sampled from two Gaussian distributions, i.e.,", + "bbox": [ + 89, + 210, + 480, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {\\mathcal {B}} \\sim \\mathcal {N} \\left(\\mu , \\frac {\\sigma^ {2}}{m}\\right) \\text {a n d} \\sigma_ {\\mathcal {B}} \\sim \\mathcal {N} \\left(\\sigma , \\left(\\frac {\\mu_ {4} / \\sigma^ {4} - 1}{4 m}\\right) \\sigma^ {2}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 342, + 482, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu$ and $\\sigma$ are the mean and variance of the activation $x\\in \\mathbb{R}$ and $\\mu_{4}$ is its 4-th central moment, i.e., $\\mu_4 = \\mathbb{E}_x(x - \\mu)^4$ . The item $\\mu_4 / \\sigma^4$ in Eqn.(6) is well-known as the kurtosis [6]]. It holds that $\\mu_4 / \\sigma^4\\geq 1$ for any distribution. Although there is no upper limit to kurtosis of a general distribution, for most well-known distributions, e.g., Gaussian distribution, uniform distribution and logistic distribution, it is finite [6]. Therefore, the variance of $\\sigma_{\\mathcal{B}}$ can always be regarded as $\\mathcal{O}(\\frac{\\sigma^2}{m})$ .", + "bbox": [ + 89, + 381, + 482, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notice that the magnitudes of both $\\mu_{\\mathcal{B}}$ and $\\sigma_{\\mathcal{B}}$ of different neurons could be quite different. The analysis above indicates that to inject noise into $\\mu_{\\mathcal{B}}$ and $\\sigma_{\\mathcal{B}}$ , it is better to multiply them by a Gaussian noise centered at 1 with proper variance instead of add them by two zero-mean noises. One reason is that the standard deviation of $\\sigma_{\\mathcal{B}}$ is proportional to $\\sigma$ when $m$ is sufficiently large. More importantly, it is impossible to choose the right amount of noise for each $\\mu_{\\mathcal{B}}$ and $\\sigma_{\\mathcal{B}}$ if we add noises additively. Therefore, we multiplicatively inject noise into its mean and variance estimations of LBN as follows:", + "bbox": [ + 89, + 517, + 482, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {\\mathcal {B}} \\leftarrow \\mu_ {\\mathcal {B}} \\cdot n _ {\\mu} \\text {a n d} \\sigma_ {\\mathcal {B}} \\leftarrow \\sigma_ {\\mathcal {B}} \\cdot n _ {\\sigma},\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 695, + 401, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_{\\mu}$ and $n_{\\sigma}$ are two noises sampled from $\\mathcal{N}(1,0.1)$ and they are treated as untrainable parameters in each back propagation. The result is presented in Fig 1. It shows that with such noise, the performance of LBN can be boosted to be comparable with BN.", + "bbox": [ + 89, + 719, + 482, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The main conclusion we can reach in the experiment above is that data-independent noise can be used to improve the performance of LBN. Moreover, it also indicates that the effect of BN on improving generalization can be decomposed into two parts: one is population normalization (can be seen from the improvement of LBN over w/o bn) and the other is the stochastic uncertainty contained in the", + "bbox": [ + 89, + 795, + 482, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mini-batch statistics (can be seen from the improvement of LBN+noise over LBN). This phenomenon inspires us to inject stochastic noise into our PN to improve its generalization performance.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Noisy Population Normalization", + "text_level": 1, + "bbox": [ + 511, + 159, + 795, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Below, we develop our final method NPN (Alg.2) to achieve better generalization performance. It is inspired from our above understandings of the stochastic uncertainty in BN. In particular, we showed that data independent noise can be injected into LBN to boost the generalization performance.", + "bbox": [ + 511, + 180, + 903, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At first, for each activation $x$ , we insert data-independent noise into the corresponding $\\mu$ and $\\gamma$ to improve generalization. In order to reduce the effects of different magnitudes of $\\mu$ and $\\gamma$ corresponding to different neurons, similar to LBN+noise in Section 4.1, we multiply $\\mu$ and $\\gamma$ by two noises centered at 1 instead of adding two zero-mean ones as follows:", + "bbox": [ + 511, + 256, + 903, + 351 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mu} = \\mu \\left(1 + n _ {\\mu}\\right) \\text {a n d} \\hat {\\gamma} = \\max \\{0, \\gamma \\left(1 + n _ {\\gamma}\\right) \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 358, + 848, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_{\\mu}$ and $n_{\\gamma}$ are independently sampled noises from a zero-mean distribution, e.g., Gaussian, and max is used to avoid negative $\\hat{\\gamma}$ . We then normalize, scale and shift $x$ by:", + "bbox": [ + 511, + 378, + 903, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = \\frac {x - \\hat {\\mu}}{\\sqrt {\\hat {\\gamma} ^ {2} + \\epsilon}} \\text {a n d} y = \\alpha \\hat {x} + \\beta , \\tag {NPN}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 426, + 903, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu, \\gamma, \\alpha$ and $\\beta$ are trainable parameters. The training process with NPN is almost the same as PN, except that the injected noises $n_{\\mu}$ and $n_{\\gamma}$ are treated as constants in back propagation. The details of the training process can be found in the appendix.", + "bbox": [ + 511, + 460, + 903, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion. NPN has the following appealing features. As the injected noises $n_{\\mu}$ and $n_{\\gamma}$ are data independent, we are able to flexibly inject a proper amount of noise into the learnt estimation of statistics to improve generalization ability, which is particularly important when large batch sizes are used. This scenario is often encountered in cross silo federated learning. The idea of injecting data independent noise to improve generalization can also be combined with other normalization methods. The success of NPN shows that one can directly learn the population statistics in the training process instead of estimating them from the minibatches. This can inspire new normalization methods.", + "bbox": [ + 511, + 536, + 905, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 4.1 It is unclear that whether NPN has some other nice properties of BN, e.g., auto-tuning of learning rates due to scale invariance [1, 7], while our empirical results demonstrate the effectiveness of NPN. We argue that different from standard training, the performance improvements achieved by these properties could be negligible compared to normalization itself in FL as the training in FL is more challenging than standard training. Such theoretical analysis is left as future work. Thus, we would like to point out that NPN is a normalization technique for FL instead of a general approach.", + "bbox": [ + 511, + 724, + 906, + 891 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10218", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1d93c7b6b52f7d6062d5f0d8908078ee4c35ec458a60ed7fea0f52df49c9aa1f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
batch sizeNormalizationα (CIFAR-10)α (CIFAR-100)
0.040.080.161.000.040.080.161.00
8w/o69.8383.2285.0587.5945.1350.4354.2658.47
IN67.6878.5582.6284.0935.5140.0245.4845.45
GN71.0582.5483.3586.1145.4549.2252.2054.75
BN70.2282.2785.7789.4252.1355.9158.1664.88
FedBN71.4983.0584.4189.3052.0556.1257.9764.43
PN78.5985.4588.0589.6353.0657.4361.8765.00
4w/o70.6982.0285.1186.9346.9950.8355.0859.45
IN60.5374.0878.1182.3936.0442.9646.1749.89
GN68.6378.5382.0185.5244.2148.1851.8554.67
BN70.1882.4786.3289.5251.3155.7860.3564.32
FedBN72.2483.5786.0389.4651.4555.7759.9165.10
PN78.4085.0587.5289.5952.2856.8562.4265.82
2w/o67.0379.9883.4387.0143.9348.5654.7759.43
IN55.4268.9375.4378.7936.1543.9047.7247.30
GN66.0278.8281.7585.2344.4046.1750.6053.24
BN65.3279.6883.3087.2545.8750.6655.0658.74
FedBN65.5080.0282.1786.4046.0451.3055.5259.68
PN72.5382.5986.2388.9848.9754.9260.6465.46
1w/o60.4971.8378.6584.9338.2442.2649.4954.59
IN35.4754.9366.4573.7433.6039.0643.2948.82
GN60.5571.8375.2579.6846.0449.2653.9756.64
BNNANANANANANANANA
FedBNNANANANANANANANA
PN64.3878.8982.1386.6946.1349.3259.1962.97
", + "bbox": [ + 217, + 87, + 781, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. The results of different normalization methods on CIFAR-10/100 with different batch sizes and Non-IID degree $\\alpha$ . Or PN consistently outperforms other methods by a large margin across different $\\alpha$ and batch sizes.", + "bbox": [ + 89, + 414, + 906, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 2 Noisy Population Normalization", + "text_level": 1, + "bbox": [ + 91, + 469, + 397, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input: activation: $x$ ; population mean and variance: $\\mu$ and $1 / \\gamma^2$ ; scale and shift parameter: $\\alpha$ and $\\beta$ ; noise level: $\\sigma$ .", + "bbox": [ + 104, + 489, + 482, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inject noise into $\\mu$ and $\\gamma$ :", + "bbox": [ + 106, + 534, + 276, + 547 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mu} = \\mu \\left(1 + n _ {\\mu}\\right) \\text {a n d} \\hat {\\gamma} = \\max \\{0, \\gamma \\left(1 + n _ {\\gamma}\\right) \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 558, + 434, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $n_{\\mu}$ and $n_{\\gamma}$ are noises from $\\mathcal{N}(0,\\sigma^2)$ . Normalize using noisy population statistics:", + "bbox": [ + 107, + 582, + 400, + 613 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = \\frac {x - \\hat {\\mu}}{\\sqrt {\\hat {\\gamma} ^ {2} + \\epsilon}}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 622, + 339, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scale and shift:", + "bbox": [ + 107, + 666, + 212, + 680 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ny = \\alpha \\hat {x} + \\beta \\equiv \\mathrm {N P N} _ {\\alpha , \\beta} ^ {\\mu , \\gamma} (\\sigma ; x)\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 691, + 385, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Return: y", + "bbox": [ + 107, + 732, + 181, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 791, + 223, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct two sets of experiments. Firstly, we verify the effectiveness of of PN in the federated learning setting given small batch sizes and heterogeneous client data. Secondly, we demonstrate the superior generalization ability of our NPN given large training batch sizes.", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Settings and Baselines", + "text_level": 1, + "bbox": [ + 511, + 469, + 718, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To verify the advantage of our population normalization method, we compare it with other normalization counterparts in the FL setting. Specifically, we compare with batch normalization [16], group normalization [40], instance normalization [37] and without normalization. We compare with FedBN [25] by averaging the BN parameters in the clients. Extensive experiments are conducted on two image classification benchmarks (i.e., CIFAR-10 and CIFAR-100) with models VGG-16 [33] and ResNet [11] to evaluate the effectiveness of our method.", + "bbox": [ + 511, + 491, + 906, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For FL experiments, we adopt FedAvg [30] as the base algorithm. We use 80 clients in total with participation rate set to 0.4 throughout all experiments. For the clients' local update, we adopt SGD with momentum. For batch size of 8, the learning rates are set to be the best among $\\{0.005, 0.01, 0.15, 0.02, 0.25\\}$ for each normalization method, and linearly down-scaled for other smaller batch sizes. We run 200 global communication rounds with local epoch set to 1. We examine the superiority of our method against the baseline under various batch sizes and heterogeneity degrees. We follow prior work [5, 28] to use Dirichlet distribution for simulating the non-IID data distribution, where the degree of heterogeneity is defined by $\\alpha$ , smaller $\\alpha$ value corresponds to more severe heterogeneity. We report the global model's average performance in the last five rounds evaluated using the test split of the datasets. For other details, please refer to the Appendix.", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10219", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fccbce6add3e42effc06f4e77385fb8472154b3a7fabd6ca2bd7df33d985c30b.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 135, + 93, + 377, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e0cb30003cdbb0168fb67d854e4880702d5de8b46105a1e083faad685bccd29.jpg", + "image_caption": [ + "Figure 2. (a) Training (left) and test (right) loss of BN, PN, and NPN as the training processes; (b) Sensitivity of NPN on noise level." + ], + "image_footnote": [], + "bbox": [ + 377, + 93, + 620, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6eca362c319f2b1bf1acbcd8bde40903267ef46310e9ef53fc0bc7b79b29d37.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 620, + 90, + 864, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For experiments with large batch sizes, we adopt the distributed training setting, where the model is being trained 4 different participants. The batch size is set to 2048. Other hyper-parameters are set to be the same as in the setting. We conduct the experiments using VGG-16 train on CIFAR-10 and CIFAR-100 datasets.", + "bbox": [ + 89, + 292, + 457, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Federated Learning with Small Batch Size", + "text_level": 1, + "bbox": [ + 89, + 395, + 450, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We point out that with small batch sizes, SGD already introduces enough stochastic uncertainty in gradient estimation, which helps generalization. Therefore, we do not need to inject additional noise for PN and thus the experiments with NPN are omitted in this section.", + "bbox": [ + 89, + 417, + 483, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Various Normalization Methods We compare PN with other counterparts under different batch sizes and non-iid degrees in Table 1. Specifically, we conduct experiments with $\\alpha \\in \\{0.04, 0.08, 0.16, 1.00\\}$ and batch sizes $\\{1, 2, 4, 8\\}$ . We adopt VGG16 as the trained network. As demonstrated in the table, we observe that our method consistently outperforms BN and other baselines, while the advantage is more significant when the batch size is small and the heterogeneity degree is higher. Notably, our method works well even with batch size set to 1, while the vanilla BN becomes inapplicable under this scenario. This enables our method to be used when the computational resource is limited on the client side.", + "bbox": [ + 89, + 515, + 483, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiments with Various Networks We confirm the effectiveness of PN across different network architectures and capacities. As shown in Table 3, we conduct experiments with VGG16, ResNet18 and ResNet50. We fix the heterogeneity degree $\\alpha$ to be 0.08 and trial with batch sizes of $\\{1,2,4,8\\}$ . We observe consistent performance advantage over the baselines for different networks. Interestingly, one may notice that networks with higher capacities do not necessarily lead to better performances, which is opposite to standard training. This phenomenon is also reported in [5], which may be due to clients' local over-fitting or further", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "drifting from each other in the parameter space. However, PN still demonstrates superior performances compared with other normalization techniques.", + "bbox": [ + 511, + 292, + 906, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Federated Learning with Large Batch Size", + "text_level": 1, + "bbox": [ + 511, + 352, + 874, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now extend our PN to NPN to enable stronger generalization ability for training with large batch sizes. The results are demonstrated in Table 4. As can be seen, when training with large batch sizes, PN is outperformed by BN due to the lack of stochastic uncertainty in the normalization layer. On the other hand, NPN remedies this issue by injecting artificial noise into the statistics estimation, which in turn leads to better test performance. The training progress is also visualized in Figure 2(a). We can see that although the training loss of NPN is larger than that of BN, its testing loss is significantly smaller than BN. This demonstrates that benefiting from the injected noise, NPN can prevent the training process from over-fitting more effectively than BN.", + "bbox": [ + 511, + 376, + 906, + 571 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0ec3266f1977e5f452867db767163f128b4a420e40ef8c7d4fe45cf2024d7779.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLossTop-1Top-5Top-10
BRN0.4493.2197.5298.83
BRN + noise0.2993.6097.8399.02
", + "bbox": [ + 562, + 587, + 857, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Batch Renormalization + noise.", + "bbox": [ + 586, + 648, + 830, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we point out that there are only a small number of participants in this scenario, which makes the trained model less prone to the adverse impact of heterogeneity and thus shows better performance than the previous experiments with more participating clients (Table 1).", + "bbox": [ + 511, + 681, + 905, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Abative Experiments", + "text_level": 1, + "bbox": [ + 511, + 771, + 712, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noise Level Sensitivity. We evaluate the sensitivity of NPN to the injected noise level. We conduct experiments with VGG-16 on CIFAR-10. We set the batch size to 2048 and vary the variance of the noise injected into NPN in $\\{0.5, 0.1, 0.01, 0.001\\}$ . The results are presented in Figure 2(b). It shows that the test loss increases gradually and slowly when we vary the variance of the noise from 0.5 to", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10220", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d67f86aa6cd0eb6f40874594440a5518c00950b98298a628939a217ef17172e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Normalizationbatch size (VGG-16)batch size (ResNet-18)batch size (ResNet-50)
124812481248
w/o71.8379.9882.0283.2275.4879.9480.2181.4965.5972.3377.8079.94
IN35.4768.9374.0378.5563.5369.3670.6871.3560.3764.6568.2869.10
GN60.5578.8278.5382.5466.0370.4472.5975.1360.6368.7270.7371.25
BNNA78.6582.0982.27NA78.5579.6481.94NA72.3978.2081.83
PN78.8982.2285.1385.4580.2083.9784.5384.6767.7273.6379.5983.54
", + "bbox": [ + 96, + 88, + 906, + 191 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/403703db580c8bf8afeed380d45570d49a1483ac76719fa8ae88194306641eb2.jpg", + "table_caption": [ + "Table 3. Results of different normalization methods on CIFAR-10. PN consistently outperforms other methods for different nwork architectures and batch sizes." + ], + "table_footnote": [], + "table_body": "
NormalizationCIFAR-10CIFAR-100
LossTop-1Top-2Top-3LossTop-1Top-2Top-3
w/o0.5492.3196.9198.202.7770.3487.9692.11
IN0.4692.4597.0098.312.5470.6888.4992.33
GN0.4392.9297.1498.522.2371.3588.9393.10
BN0.3993.1397.6698.872.0572.2190.2893.60
PN0.4193.1097.7898.971.8572.1590.0793.89
NPN0.3293.6698.0599.121.2272.7491.4794.90
", + "bbox": [ + 254, + 243, + 743, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Results of different normalization methods on CIFAR-10 / CIFAR-100 with large batch size.", + "bbox": [ + 197, + 354, + 797, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "0.001. Therefore, it is easy to select a suitable noise variance that falls within a reasonable range.", + "bbox": [ + 89, + 395, + 480, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Artificial Noise for Other Normalization Methods. We demonstrate that the artificially injected noise is also able to improve existing normalization methods with limited stochastic uncertainty. We take batch renormalization (BRN) [16] as an example and conduct experiments on CIFAR-10 with VGG-16. For simplicity and excluding other factors, we conduct this experiment with regular training instead of FL. We inject noise sampled from $\\mathcal{N}(0,0.1)$ to the moving averages $\\mu$ and $\\sigma$ of BRN as we do for NPN. The results are given in Table 2. It shows that with the injected noise, the performance of BRN can be improved effectively. Therefore, the idea of data-independent noise injection is a very general regularization technique that can be used to improve existing normalization methods.", + "bbox": [ + 88, + 445, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Supervised Learning with PN", + "text_level": 1, + "bbox": [ + 89, + 666, + 354, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intuitively, PN can also be used in non-FL scenarios. We apply PN to supervised learning and conduct experiments on two image classification benchmarks (i.e., CIFAR-10, CIFAR-100) with VGG-16 and ResNet to evaluate the effectiveness of PN in supervised learning. The batch sizes of all the methods are set to 128. Table 5 presents the test loss and accuracy of the final learned models. From Table 5 we can see that our NPN can achieve much lower test loss than BN and comparable (if no better than) test accuracy with BN on all the three models. We can also observe that NPN can achieve a noticeable improvement over PN, which demonstrates that artificially injected data independent noise can be used to reinforce the performance of the normalization methods with limited stochastic under", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/58761bae11106903945ddf6a64386e7e1f37863696ba8470a113db263c3ebeba.jpg", + "table_caption": [ + "Table 5. Results of different normalization methods on image classification benchmarks" + ], + "table_footnote": [], + "table_body": "
NetworkNormalizationLossAccuracy
VGG-16BN0.3992.31
PN0.4193.63
NPN0.2893.66
ResNet-18BN0.4693.76
PN0.4593.97
NPN0.3894.26
ResNet-56BN0.3393.05
PN0.4292.56
NPN0.2593.99
", + "bbox": [ + 537, + 431, + 879, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tainty. The comparable performance of BN and NPN implies that the effect of BN can be decomposed into two parts: one is population normalization, and the other is the stochastic uncertainty containing in the mini-batch statistics.", + "bbox": [ + 511, + 616, + 906, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 693, + 638, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a learning-based normalization technique, Population Normalization (PN), which reduces the reliance on training batch size and improves robustness against data heterogeneity in federated learning (FL). Extensive experiments demonstrate that PN outperforms existing normalization techniques by a significant margin. We further develop Noisy Population Normalization (NPN), which injects artificial noise into the learnt normalization statistics to enhance generalization under large batch size conditions. We believe that our work will offer valuable insights for the design and development of more stable and adaptable normalization techniques for FL.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10221", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Nature Science Foundation of China (62472097) and AI for Science Foundation of Fudan University (FudanX24AI028). The computations in this research were performed on the CFFF platform of Fudan University.", + "bbox": [ + 89, + 113, + 483, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 227 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sanjeev Arora, Kaifeng Lyu, and Zhiyuan Li. Theoretical analysis of auto rate-tuning by batch normalization. In 7th International Conference on Learning Representations, ICLR 2019, 2019. 5", + "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 1, 2", + "[3] Tobias Bernecker, Annette Peters, Christopher L. Schlett, Fabian Bamberg, Fabian J. Theis, Daniel Rueckert, Jakob Weiss, and Shadi Albarqouni. Fednorm: Modality-based normalization in federated learning for multi-modal liver segmentation. ArXiv, abs/2205.11096, 2022. 3", + "[4] Yongqiang Cai, Qianxiao Li, and Zuowei Shen. A quantitative analysis of the effect of batch normalization on gradient descent. In International Conference on Machine Learning, pages 882-890, 2019. 3", + "[5] Hong-You Chen and Wei-Lun Chao. Fedbe: Making bayesian model ensemble applicable to federated learning. In ICLR, 2021. 3, 6, 7", + "[6] Wikipedia contributors. Kurtosis. Online; accessed 2-June-2020, 2020. Wikipedia, The Free Encyclopedia. 5", + "[7] Hadi Daneshmand, Amir Joudaki, and Francis Bach. Batch normalization orthogonalizes representations in deep random networks. In Advances in Neural Information Processing Systems, 2021. 5", + "[8] Jian-Hui Duan, Wenzhong Li, and Sanglu Lu. Feddna: Federated learning with decoupled normalization-layer aggregation for non-iid data. In Machine Learning and Knowledge Discovery in Databases. Research Track, pages 722-737, Cham, 2021. Springer International Publishing. 3", + "[9] Aiden Durrant, Milan Markovic, David Matthews, David May, Jessica Enright, and Georgios Leontidis. The role of cross-silo federated learning in facilitating data sharing in the agri-food sector. Computers and Electronics in Agriculture, 193:106648, 2022. 2, 4", + "[10] Filip Hanzely, Slavomír Hanzely, Samuel Horváth, and Peter Richtárik. Lower bounds and optimal algorithms for personalized federated learning. Advances in Neural Information Processing Systems, 33:2304-2315, 2020. 3", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 6", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + ], + "bbox": [ + 93, + 236, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip Gibbons. The non-IID data quagmire of decentralized machine learning. In Proceedings of the 37th International Conference on Machine Learning, pages 4387-4398. PMLR, 2020. 3", + "[14] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip B. Gibbons. The non-iid data quagmire of decentralized machine learning, 2020. 1", + "[15] Chao Huang, Jianwei Huang, and Xin Liu. Cross-silo federated learning: Challenges and opportunities. arXiv preprint arXiv:2206.12949, 2022. 2, 4", + "[16] Sergey Ioffe. Batch renormalization: Towards reducing minibatch dependence in batch-normalized models. In Advances in neural information processing systems, pages 1945-1953, 2017. 1, 2, 6, 8", + "[17] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning, pages 448-456, Lille, France, 2015. PMLR. 1, 2", + "[18] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for ondevice federated learning. In ICML, 2020. 3", + "[19] Jakub Konečny, H Brendan McMahan, Felix X Yu, Peter Richtárik, Ananda Theertha Suresh, and Dave Bacon. Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492, 2016. 1, 3", + "[20] Viraj Kulkarni, Milind Kulkarni, and Aniruddha Pant. Survey of personalization techniques for federated learning. In 2020 Fourth World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4), pages 794-797. IEEE, 2020. 3", + "[21] Qinbin Li, Bingsheng He, and Dawn Song. Practical oneshot federated learning for cross-silo setting. arXiv preprint arXiv:2010.01017, 2020. 2, 4", + "[22] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10713–10722, 2021. 3", + "[23] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In MLSys, 2020. 3", + "[24] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through personalization. In International Conference on Machine Learning, pages 6357-6368. PMLR, 2021. 3", + "[25] Xiaoxiao Li, Meirui Jiang, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fed{bn}: Federated learning on non- $\\{\\mathrm{i}\\mathrm{d}\\}$ features via local batch normalization. In International Conference on Learning Representations, 2021. 2, 3, 6", + "[26] Xiangru Lian and Ji Liu. Revisit batch normalization: New understanding and refinement via composition optimization. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 3254-3263, 2019. 3", + "[27] Wei Yang Bryan Lim, Nguyen Cong Luong, Dinh Thai Hoang, Yutao Jiao, Ying-Chang Liang, Qiang Yang, Dusit" + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10222", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Niyato, and Chunyan Miao. Federated learning in mobile edge networks: A comprehensive survey. IEEE Communications Surveys & Tutorials, 22(3):2031-2063, 2020. 3", + "[28] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. Advances in Neural Information Processing Systems, 33:2351-2363, 2020. 3, 6", + "[29] Ping Luo, Xinjiang Wang, Wenqi Shao, and Zhanglin Peng. Towards understanding regularization in batch normalization. arXiv preprint arXiv:1809.00846, 2018. 3, 5", + "[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 3, 6", + "[31] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. In Advances in neural information processing systems, pages 901-909, 2016. 2", + "[32] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? In Advances in Neural Information Processing Systems, pages 2483-2493, 2018. 3", + "[33] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 6", + "[34] Sebastian U Stich. Local sgd converges fast and communicates little. In International Conference on Learning Representations, 2018. 3", + "[35] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 3", + "[36] Mattias Teye, Hossein Azizpour, and Kevin Smith. Bayesian uncertainty estimation for batch normalized deep networks. In International Conference on Machine Learning, pages 4907-4916, 2018. 5", + "[37] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 1, 2, 6", + "[38] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 3", + "[39] Xiaoxia Wu, Edgar Dobriban, Tongzheng Ren, Shanshan Wu, Zhiyuan Li, Suriya Gunasekar, Rachel Ward, and Qiang Liu. Implicit regularization of normalization methods. arXiv preprint arXiv:1911.07956, 2019. 3", + "[40] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-19, 2018. 1, 2, 3, 6", + "[41] Jianhang Xiao, Chunhui Du, Zijing Duan, and Wei Guo. A novel server-side aggregation strategy for federated learning in non-id situations. In 2021 20th International Symposium on Parallel and Distributed Computing (ISPDC), pages 17-24. IEEE, 2021. 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1, 3", + "[43] Yousef Yeganeh, Azade Farshad, Nassir Navab, and Shadi Albarqouni. Inverse distance aggregation for federated learning with non-iid data. In Domain Adaptation and Representation Transfer, and Distributed and Collaborative Learning, pages 150-159. Springer, 2020. 3" + ], + "bbox": [ + 514, + 90, + 905, + 218 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10223", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Population Normalization for Federated Learning/b6b8182e-2812-44b6-99cb-82e1b12982cb_model.json b/2025/Population Normalization for Federated Learning/b6b8182e-2812-44b6-99cb-82e1b12982cb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..372fc7ab304e910a5ce9cda59bbdc6411e89864f --- /dev/null +++ b/2025/Population Normalization for Federated Learning/b6b8182e-2812-44b6-99cb-82e1b12982cb_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.246, + 0.131, + 0.753, + 0.154 + ], + "angle": 0, + "content": "Population Normalization for Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.179, + 0.874, + 0.2 + ], + "angle": 0, + "content": "Zhuoyao Wang\\(^{1*}\\), Fan Yi\\(^{1*}\\), Peizhu Gong\\(^{1}\\), Caitou He\\(^{2}\\), Cheng Jin\\(^{1,3}\\), Weizhong Zhang\\(^{1,4\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.198, + 0.671, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)Fudan University \\(^{2}\\) Wenzhou University" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.216, + 0.832, + 0.234 + ], + "angle": 0, + "content": "3Innovation Center of Calligraphy and Painting Creation Technology, MCT, China" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.234, + 0.762, + 0.253 + ], + "angle": 0, + "content": "\\(^{4}\\)Shanghai Key Laboratory of Intelligent Information Processing" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.255, + 0.914, + 0.27 + ], + "angle": 0, + "content": "{18110240064,pzgong,jc,weizhongzhang}@fudan.edu.cn，ftyi21@m.fudan.edu.cn，hctou89@wzu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.486, + 0.744 + ], + "angle": 0, + "content": "Batch normalization (BN) is widely recognized as an essential method in training deep neural networks, facilitating convergence and enhancing model stability. However, in Federated Learning (FL) contexts, where training data are typically heterogeneous and clients often face resource constraints, the effectiveness of BN is considerably limited for two primary reasons. First, the population statistics, specifically the mean and variance of these heterogeneous datasets, vary substantially, resulting in inconsistent BN layers across client models, which ultimately drives these models to diverge further. Second, estimating statistics from a mini-batch is often imprecise since the batch size has to be small in resource-limited clients. This paper introduces Population Normalization, a novel technique for FL, in which the statistics are learned as trainable parameters rather than calculated from mini-batches as in BN. Thus, our normalization layers are homogeneous among the clients and the adverse impact of small batch size is eliminated as the model can be well-trained even when the batch size equals to one. To enhance the flexibility of our method in practical applications, we investigate the role of stochastic uncertainty in BN's statistical estimation. When larger batch sizes are available, we demonstrate that injecting simple artificial noise can effectively mimic this stochastic uncertainty and improve the model's generalization capability. Experimental results validate the efficacy of our approach across various FL tasks." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.771, + 0.222, + 0.787 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Batch Normalization (BN) [17] has become a standard component in deep neural networks (DNNs). It is proposed to stabilize and accelerate the training of DNNs, and the original motivation was to reduce the so-called internal covariate shift. In each training step, it normalizes the layer input of a neural network using the mean and variance com" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.306, + 0.907, + 0.362 + ], + "angle": 0, + "content": "puted within a randomly sampled mini-batch. Specifically, consider a particular neuron in a network, and we denote \\(\\{x_{1},\\ldots ,x_{m}\\}\\) to be its pre-activations on a mini-batch \\(\\mathcal{B}\\) of size \\(m\\), BN first normalizes each \\(x_{i}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.37, + 0.907, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {i} = \\frac {x _ {i} - \\mu_ {\\mathcal {B}}}{\\sigma_ {\\mathcal {B}}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.538 + ], + "angle": 0, + "content": "where \\(\\mu_{\\mathcal{B}}\\) and \\(\\sigma_{\\mathcal{B}}^{2}\\) are the sample mean and variance of the pre-activation on \\(\\mathcal{B}\\). It then scales and shifts each \\(\\hat{x}_i\\) as \\(y_i = \\alpha \\hat{x}_i + \\beta\\) to compensate the effect of normalization, where \\(\\gamma\\) and \\(\\beta\\) are two parameters to be learned during the training process. BN finally takes \\(y_i\\) as the new pre-activation of this neuron. It has been reported in literature [12, 16, 17, 37, 40] that BN can significantly improve the convergence of training and the generalization performance of the final learned models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.765 + ], + "angle": 0, + "content": "However, we notice that the effectiveness of BN is compromised in Federated Learning [14, 19, 42], which typically aims at collaboratively training models on resource limited clients (e.g., mobiles) with locally private and heterogeneous training data. The main reasons can be summarized as follows. One is that the statistics, i.e., the mean and variance, of heterogeneous datasets in the clients differ greatly. Eqn. (1) implies that this can lead to inconsistent BN layers among the client models and finally causes these models to drift further away from each other. The other is that in resource limited clients, the mini-batch size has to be small, which makes \\(\\mu_{\\mathcal{B}}\\) and \\(\\sigma_{\\mathcal{B}}^{2}\\) inaccurate estimations of the population mean and variance for normalization and the obtained BN would cause instability during training that slows down convergence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Some other normalization techniques, such as batch renormalization [16], group normalization [40], and layer normalization [2], are proposed to address this heavy reliance on batch size. The key ideas are to estimate the population mean and variance as accurately as possible by using techniques such as moving average [16] or to normalize the activation along other dimensions such as the channel dimension [2]. Nevertheless, these approaches focus on training acceleration, which typically under-perform" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.363, + 0.901 + ], + "angle": 0, + "content": "*Equal Contribution, † Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10214" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "BN in terms of the trained model's accuracy [40]. Recently, FedBN [25] is proposed to extend BN to FL. It keeps the client BN layers updated locally, without communicating and aggregating them during training. This enables the clients to achieve better performances in their own respective environments. However, since the BN layers among the clients can be heterogeneous due to the non-IID datasets, FedBN cannot produce an off-the-shelf global model after training, which is often required in reality. Consequently, FedBN suffers from the cold start issue, i.e., the learned model cannot be directly deployed in the new clients due to the lack of trained local BN layers. What's more, FedBN still relies on estimation within mini-batches, which cannot address the issue of inaccurate estimation caused by small batch size in cross-device FL." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.323, + 0.485, + 0.639 + ], + "angle": 0, + "content": "In this paper, we propose Population Normalization (PN), a simple yet effective normalization method for FL to eliminate the negative impacts of statistical heterogeneity and estimation inaccuracies on BN. The key idea is to learn population statistics as trainable parameters in the training process by introducing two constraints derived from the definitions of mean and variance (see Section 3), instead of estimating them from the mini-batches as most existing methods have done. As the population statistics take place of the sample mean and variance of the local datasets in BN, and are treated as normal trainable parameters in local training and aggregation, our normalization layers among the clients would be consistent although the datasets are heterogeneous. This enables our method to eventually produce a global model with the normalization layer among the clients. In addition, since we no longer need to calculate the sample mean and variance, the resulting training problem can be solved by stochastic algorithms of FL with arbitrary batch sizes (even equals to one) and thus the issue of inaccurate statistic estimation because of small batch size is eliminated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In practice, we observe that PN tends to under-perform BN given very large batch sizes. We then further explore this issue and find its cause to be that PN lacks the stochastic uncertainty in the estimation of mean and variance. Specifically, we show that the noise introduced by the mini-batch based mean and variance estimations in the BN is essential for its good generalization behavior. Our empirical results demonstrate that injecting a reasonable amount of data independent noise into BN with extremely large batch sizes can significantly improve the generalization performance to be comparable to that of BN with proper batch sizes. Inspired by this finding, we propose Noisy Population Normalization (NPN) by injecting simple data independent artificial noises into the learned population mean and variance in PN during training. The injected noise can serve the same purpose as the uncertainty in estimating population statistics in BN with small batch sizes. This makes the proposed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "method more flexible in the applications of FL where large batch size is available, e.g., collaboration among companies in cross-silo FL [9, 15, 21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Finally, we conduct a series of experiments to verify the effectiveness of our proposed PN and NPN in the context of FL with small and large batch sizes, respectively (Section 5). The results demonstrate that PN effectively reduces the reliance on training batch size and alleviates the negative impact of data heterogeneity on normalization. Furthermore, the artificial noise injected in our NPN method effectively improves the generalization performance given large training batch sizes." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.275, + 0.874, + 0.289 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.291, + 0.905, + 0.336 + ], + "angle": 0, + "content": "- We propose a normalization technique PN for federated learning, which demonstrates superior performance given small batch sizes and data heterogeneity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.336, + 0.905, + 0.381 + ], + "angle": 0, + "content": "- In training with large batch sizes, we propose NPN based on PN, which introduces stochastic uncertainty by injecting artificial noise to improve generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.381, + 0.905, + 0.41 + ], + "angle": 0, + "content": "- We conduct extensive experiments to verify the effectiveness of our method." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.291, + 0.905, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.906, + 0.518 + ], + "angle": 0, + "content": "Notations: Given a positive integer \\(d\\), we define \\([d]\\) to be the set \\(\\{1,2,\\dots,d\\}\\). For a vector \\(x\\in \\mathbb{R}^d\\), we let \\([x]_i\\) be its \\(i\\)-th component with \\(i\\in [d]\\). We denote by \\(w_{i,j}\\) the element in the \\(i\\)-th row and the \\(j\\)-th column of a real-valued matrix \\(w\\). Let \\(A\\cup B\\) be the union of two sets \\(A\\) and \\(B\\). Moreover, we denote by \\(\\mathcal{N}(\\mu ,\\sigma^2)\\) the Gaussian distribution with mean of \\(\\mu\\) and variance of \\(\\sigma^2\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.535, + 0.655, + 0.55 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.753, + 0.577 + ], + "angle": 0, + "content": "2.1. Normalization Techniques" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.688 + ], + "angle": 0, + "content": "BN [17] is a standard normalization technique, which enables the training of very deep neural networks, and the most representative result is ResNet[11], which cannot be easily trained to converge without BN. It can also achieve significant imporement in generalization. Therefore, the invention of BN is a major milestone in the development of deep learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As aforementioned, small batch size or non-iid minibatch can diminish the effect of BN on the training acceleration or even result in training failure, due to the inaccurate mean and variance estimations [16]. Moreover, large batch size leads to huge memory cost. This issue hinders BN's usage in FL applications, where the clients often have limited computational resources and the datasets are usually heterogeneous. This in turn prevents the training of large networks in FL. Recently, some new normalization methods have been proposed as alternatives to BN, and they can be roughly divided into two categories, i.e., normalization along the batch dimension [16], and mini-batch independent normalization [2, 31, 37, 40]. The former alleviates the reliance on the batch size by using techniques such as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10215" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.226 + ], + "angle": 0, + "content": "moving average. The latter ones achieve this by normalizing the activation of neurons in a network along other dimensions. For example, Group Normalization [40] divides the channels in each hidden layer into groups and normalizes the activations in each group with its mean and variance. Empirical studies show that all these methods achieve comparable training speedups with BN. However, the generalization performance they achieved is inferior to that of BN." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.484, + 0.41 + ], + "angle": 0, + "content": "Besides the above studies, there is another line of research [4, 26, 29, 32, 39] that focuses on understanding BN. One of the main conclusions in these investigations is that with BN, the effective objective function used in training is different from the original one. Specifically, [32] showed that by inserting BN layers into a network, the loss landscape becomes smoother, which naturally explains why BN enables one to use large learning rates. The studies [29, 39] focus on the regularization effect of BN. They showed that the objective function with BN can be divided into two parts: an empirical loss under population normalization and an explicit regularizer named gamma decay." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.418, + 0.282, + 0.435 + ], + "angle": 0, + "content": "2.2. Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.484, + 0.482 + ], + "angle": 0, + "content": "FL emerges as an effective paradigm to train models on distributed devices without centralizing local private data [3, 8, 13, 19, 42]. Specifically, it can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.491, + 0.483, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathcal {L} (\\theta , \\mathcal {D}) \\triangleq \\sum_ {m = 1} ^ {M} \\rho_ {m} \\mathcal {L} _ {m} (\\theta , \\mathcal {D} _ {m}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.681 + ], + "angle": 0, + "content": "where \\(M\\) is the number of clients and \\(\\theta\\) is the model to be learnt. The global dataset \\(\\mathcal{D} = \\cup_{m=1}^{M}\\mathcal{D}_{m}\\), where \\(\\mathcal{D}_{m} = \\{(\\pmb{x}_{i},y_{i})\\}_{i=1}^{|D_{m}|}\\) is the locally kept dataset on client \\(m\\). \\(\\mathcal{L}_{m}(\\theta,\\mathcal{D}_{m}) = \\frac{1}{|\\mathcal{D}_{m}|}\\sum_{\\xi\\in\\mathcal{D}_{m}}\\ell(\\theta,\\xi)\\) denotes the local empirical risk with the loss function \\(\\ell(\\cdot,\\cdot)\\) and \\(\\rho_{m} = \\frac{|D_{m}|}{|\\mathcal{D}|}\\). Typically, the most popular FL algorithm FedAvg [30] solves the problem (2) based on local-SGD [34]. Specifically, in the \\(c^{th}\\) round, it updates the global model on the server as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.689, + 0.483, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\theta^ {c + 1} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\rho_ {m} \\theta_ {m} ^ {c}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.796 + ], + "angle": 0, + "content": "where \\(\\theta_{m}^{c}\\) denotes the received local model from client \\(m\\) after local training in this round. The updated global model \\(\\theta^{c + 1}\\) is then distributed to the clients to initiate the \\((c + 1)\\)-th round of training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In FL, especially in mobile edge scenarios, both limited device capacity and non-IID distribution of clients' local data raise challenges to the training procedure [27]. Various works attempt to tackle the non-IID distribution of data for federated learning from the perspective of local training regularization [18, 22, 23], server aggregation [5, 28, 38, 41, 43], personalized federated learning [10, 20, 24, 35]," + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.092, + 0.779, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Population Normalization" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.111, + 0.907, + 0.142 + ], + "angle": 0, + "content": "Input: activation: \\( x \\); mean and variance: \\( \\mu \\) and \\( 1 / \\gamma^2 \\); scale and shift parameters: \\( \\alpha \\) and \\( \\beta \\)." + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.143, + 0.806, + 0.157 + ], + "angle": 0, + "content": "Normalize using the population statistics:" + }, + { + "type": "equation", + "bbox": [ + 0.671, + 0.166, + 0.766, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = \\frac {x - \\mu}{\\sqrt {\\gamma^ {2} + \\epsilon}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.21, + 0.637, + 0.223 + ], + "angle": 0, + "content": "Scale and shift:" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.234, + 0.797, + 0.252 + ], + "angle": 0, + "content": "\\[\ny = \\alpha \\hat {x} + \\beta \\equiv \\mathrm {P N} _ {\\alpha , \\beta} ^ {\\mu , \\gamma} (x)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.531, + 0.275, + 0.605, + 0.291 + ], + "angle": 0, + "content": "Return: y" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.461 + ], + "angle": 0, + "content": "etc. However, they do not explore the reduced effectiveness of batch normalization in federated learning. FedBN [25] uses local BN parameters to tackle non-iid federated learning. However, FedBN is only applicable to local inferences where only clients participating in the training with local statistics can perform inference, without providing a global model for new clients or global data. Also, FedBN does not consider the small batches caused by limited local resources." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.756, + 0.499 + ], + "angle": 0, + "content": "3. Population Normalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.592 + ], + "angle": 0, + "content": "We now present our method PN (see Alg.1). Its key idea is to eliminate the impacts of heterogeneous statistics and inaccurate statistic estimations by learning the population mean and variance during the training process, instead of estimating them within the mini-batch as BN has done. Precisely, given an activation \\( x \\), we normalize it as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.607, + 0.759, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = \\frac {x - \\mu}{\\sqrt {\\gamma^ {2} + \\epsilon}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.653, + 0.907, + 0.744 + ], + "angle": 0, + "content": "where \\(\\mu\\) and \\(\\gamma^2\\) are the population mean and variance, respectively, and they will be learned during the training process, of which the details are specified later in Eqn.(3). \\(\\epsilon > 0\\) is a small constant. Then, to compensate the effect of normalization on the effective capacity of a neural network, we scale and shift the normalized activation \\(\\hat{x}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.759, + 0.907, + 0.78 + ], + "angle": 0, + "content": "\\[\ny = \\alpha \\hat {x} + \\beta \\equiv \\mathrm {P N} _ {\\alpha , \\beta} ^ {\\mu , \\gamma} (x), \\quad (\\mathrm {P N})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.906, + 0.823 + ], + "angle": 0, + "content": "where the coefficients \\(\\alpha\\) and \\(\\beta\\) are learned in the training process together with \\(\\mu\\) and \\(\\gamma\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Below, we take a fully connected neural network \\( f(\\pmb{x}) \\in \\mathbb{R} \\) as an example to show how to insert PN layers into DNNs. Specifically, given an input \\( \\pmb{x} \\in \\mathbb{R}^d \\), let the input layer be \\( f_j^{(0)}(\\pmb{x}) = [\\pmb{x}]_j \\) with \\( j \\in [n^{(0)}] = [d] \\). The \\( \\ell \\)-th hidden layer containing \\( n^{(\\ell)} \\) hidden nodes with \\( \\ell \\in [L] \\) takes" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10216" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.172, + 0.105 + ], + "angle": 0, + "content": "the form of" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.109, + 0.484, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f _ {j} ^ {(\\ell)} (\\boldsymbol {x}) = h ^ {(\\ell)} \\left(z _ {j} ^ {(\\ell)} (\\boldsymbol {x})\\right) \\text {w i t h} \\tag {4} \\\\ z _ {j} ^ {(\\ell)} (\\boldsymbol {x}) = \\sum_ {k = 1} ^ {n ^ {(\\ell - 1)}} w _ {j, k} ^ {(\\ell)} \\underbrace {\\left(\\alpha_ {k} ^ {(\\ell - 1)} \\frac {f _ {k} ^ {(\\ell - 1)} (\\boldsymbol {x}) - \\mu_ {k} ^ {(\\ell - 1)}}{\\sqrt {\\left(\\gamma_ {k} ^ {(\\ell - 1)}\\right) ^ {2} + \\epsilon}} + \\beta_ {k} ^ {(\\ell - 1)}\\right)} _ {(\\ell - 1) - \\text {t h P N l a y e r}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.484, + 0.332 + ], + "angle": 0, + "content": "where \\(j\\in [n^{(\\ell)}],z_j^{(\\ell)}(\\pmb {x})\\) is the pre-activation, \\(h^{(\\ell)}\\) is the activation function and \\(w^{(\\ell)}\\in \\mathbb{R}^{n^{(\\ell)}\\times n^{(\\ell -1)}}\\) is the weight matrix connecting layer \\(\\ell\\) and \\(\\ell -1\\) \\(\\mu_k^{(\\ell -1)}\\) and \\(\\gamma_{k}^{(\\ell -1)}\\) are the population mean and standard deviation of the hidden nodes \\(f_{k}^{(\\ell -1)}(\\pmb {x})\\) \\(\\alpha_{k}^{(\\ell -1)}\\) and \\(\\beta_{k}^{(\\ell -1)}\\) are the parameters of the linear transformation. The final output layer is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.336, + 0.483, + 0.414 + ], + "angle": 0, + "content": "\\[\nf (\\boldsymbol {x}) = \\sum_ {k = 1} ^ {n ^ {(L)}} u _ {k} \\underbrace {\\left(\\frac {f _ {k} ^ {(L)} (\\boldsymbol {x}) - \\mu_ {k} ^ {(L)}}{\\sqrt {\\left(\\gamma_ {k} ^ {(L)}\\right) ^ {2} + \\epsilon}}\\right)} _ {\\text {f i n a l P N l a y e r}} + c _ {0}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.483, + 0.503 + ], + "angle": 0, + "content": "where \\(u\\in \\mathbb{R}^{n^{(L)}}\\) is the weight vector connecting the output to the \\(L\\) -th hidden layer, \\(\\mu_k^{(L)}\\) and \\(\\gamma_k^{(L)}\\) are the population mean and standard deviation of the hidden nodes \\(f_{k}^{(L)}(\\pmb {x})\\) The scaling and shifting parameters are absorbed into \\(u_{k}\\) and \\(c_{0}\\) , respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.643 + ], + "angle": 0, + "content": "To eliminate the dependency of PN on batch size, we learn the statistics \\(\\mu_k^{(\\ell)}\\) and \\(\\gamma_k^{(\\ell)}\\) in training by introducing two natural constraints derived from the definition of mean and variance. The statistic estimation is optimized together with the model parameters during local training, and aggregated on the server by FL algorithms, such as FedAVG. To be precise, given the loss \\(\\phi(\\cdot, \\cdot): \\mathbb{R} \\times \\{0,1\\} \\to \\mathbb{R}\\) and the global dataset \\(\\mathcal{D} = \\cup_{m=1}^{M} \\mathcal{D}_m\\), the training objective in Eqn. (2) of a FL problem becomes" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.648, + 0.476, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta} \\mathcal {L} (\\theta , \\mathcal {D}) s. t. \\left\\{ \\begin{array}{l} \\frac {1}{| \\mathcal {D} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\mathcal {D}} \\frac {\\left(f _ {k} ^ {(\\ell)} (\\boldsymbol {x}) - \\mu_ {k} ^ {(\\ell)}\\right) ^ {2}}{\\left(\\gamma_ {k} ^ {(\\ell)}\\right) ^ {2} + \\epsilon} = 1 \\\\ \\frac {1}{| \\mathcal {D} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\mathcal {D}} f _ {k} ^ {(\\ell)} (\\boldsymbol {x}) = \\mu_ {k} ^ {(\\ell)} \\end{array} \\right., \\\\ k \\in [ m ^ {(\\ell)} ], \\ell \\in \\{0 \\} \\cup [ L ] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the set of all the trainable parameters, i.e., \\(\\theta = \\{w^{(\\ell)}\\} \\cup \\{u\\} \\cup \\{\\mu^{(\\ell)}\\} \\cup \\{\\gamma^{(\\ell)}\\} \\cup \\{\\alpha^{(\\ell)}\\} \\cup \\{\\beta^{(\\ell)}\\} \\cup \\{c_0\\}\\) and \\(\\mathcal{L}(\\theta ,\\mathcal{D}) = \\sum_{m = 1}^{M}\\rho_{m}\\sum_{(\\boldsymbol {x},y)\\in \\mathcal{D}_{m}}\\phi (f(\\boldsymbol {x}),y)\\). This training problem can be solved by various stochastic algorithms with arbitrary batch sizes (even equals to 1), such as moving average or the method of Lagrangian multipliers. In the experiments of this paper, we use stochastic Lagrangian multipliers and the details can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Discussion. (1) Since the statistics \\(\\mu\\) and \\(\\gamma\\) are learned in the training process instead of being estimated from minibatches, the resulting training problem can be solved by" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "stochastic solvers of FL with arbitrary batch sizes (even equals to 1). Thus the reliance on batch size is eliminated. This enables resource limited clients to train larger deep neural networks. (2) Both \\(\\mu\\) and \\(\\gamma\\) are treated as normal trainable parameters in local training, communication and aggregation, our PN layers are consistent among the clients although the datasets on the clients are heterogeneous. This enables us to eventually obtain an off-the-shelf global model, which is often required in reality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.235, + 0.907, + 0.267 + ], + "angle": 0, + "content": "Remark 3.1 We should point out that our PN layer can also be inserted before the activation functions." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.289, + 0.695, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.289, + 0.886, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.393, + 0.907, + 0.436 + ], + "angle": 0, + "content": "Figure 1. Training loss (left) and test loss (right) of BN, LBN and LBN+noise. Injecting artificial noise into estimated BN statistics successfully decreases the test loss and improves generalization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.807, + 0.47 + ], + "angle": 0, + "content": "4. Noisy Population Normalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.659 + ], + "angle": 0, + "content": "We observe that PN demonstrates inferior performance relative to BN when batch size becomes larger. We explore this issue and find its cause to be that PN lacks stochastic uncertainty in the mean and variance estimations. In practice, large batch sizes may often be used in cross-silo federated learning, where the participants are companies and organizations with huge amounts of data as well as abundant computational resources [9, 15, 21]. Therefore, to enable our method to be more flexibly used in such scenarios, we propose NPN, which injects artificial noise into the learned statistics during training to boost generalization. The details are presented in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.866, + 0.684 + ], + "angle": 0, + "content": "4.1. The Role of Stochastic Uncertainty in BN" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.735 + ], + "angle": 0, + "content": "We provide an in-depth exploration into the role of stochastic uncertainty of the mini-batch statistics in BN. This provides a deep understanding of BN's batch size reliance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.87 + ], + "angle": 0, + "content": "We start with a simple experiment to show the effect of batch size on BN. Precisely, we compare the performances of BN with a reasonable batch size (m=128) and a very large size (m=2048) in training VGG-16 on CIFAR-10. For simplicity, we refer to BN with m=2048 as LBN. For fair comparison, in LBN, to eliminate the effect of large batch size on SGD, we only use large batches to obtain batch statistics and calculate the stochastic gradient with respect to the loss on mini-batches of size 128." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We present both the training and test losses of the learned models with BN and LBN over the training process in Fig" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10217" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.211 + ], + "angle": 0, + "content": "1, from which we can see that LBN converges faster than BN, due to the more accurate population statistics estimation. However, there is a noticeable gap between their test losses, which becomes larger after the learning rate decays. Therefore, the effect of BN in improving generalization is reduced when the batch size becomes too large. This indicates that the stochastic uncertainty in estimates of the mean and variance is important for BN." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.482, + 0.336 + ], + "angle": 0, + "content": "The next question we want to investigate is whether we can artificially inject a proper amount of data-independent noise into LBN to improve its generalization performance. Note that the noise of estimating \\(\\mu_{\\mathcal{B}}\\) and \\(\\sigma_{\\mathcal{B}}\\) in BN is data dependent. To answer this question, we should first calculate the uncertainty explicitly. The studies [29, 36] show that for sufficiently large batch size M, the estimated \\(\\mu_{\\mathcal{B}}\\) and \\(\\sigma_{\\mathcal{B}}\\) can be viewed as two random variables sampled from two Gaussian distributions, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.343, + 0.483, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\mu_ {\\mathcal {B}} \\sim \\mathcal {N} \\left(\\mu , \\frac {\\sigma^ {2}}{m}\\right) \\text {a n d} \\sigma_ {\\mathcal {B}} \\sim \\mathcal {N} \\left(\\sigma , \\left(\\frac {\\mu_ {4} / \\sigma^ {4} - 1}{4 m}\\right) \\sigma^ {2}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.483, + 0.518 + ], + "angle": 0, + "content": "where \\(\\mu\\) and \\(\\sigma\\) are the mean and variance of the activation \\(x\\in \\mathbb{R}\\) and \\(\\mu_{4}\\) is its 4-th central moment, i.e., \\(\\mu_4 = \\mathbb{E}_x(x - \\mu)^4\\). The item \\(\\mu_4 / \\sigma^4\\) in Eqn.(6) is well-known as the kurtosis [6]]. It holds that \\(\\mu_4 / \\sigma^4\\geq 1\\) for any distribution. Although there is no upper limit to kurtosis of a general distribution, for most well-known distributions, e.g., Gaussian distribution, uniform distribution and logistic distribution, it is finite [6]. Therefore, the variance of \\(\\sigma_{\\mathcal{B}}\\) can always be regarded as \\(\\mathcal{O}(\\frac{\\sigma^2}{m})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.681 + ], + "angle": 0, + "content": "Notice that the magnitudes of both \\(\\mu_{\\mathcal{B}}\\) and \\(\\sigma_{\\mathcal{B}}\\) of different neurons could be quite different. The analysis above indicates that to inject noise into \\(\\mu_{\\mathcal{B}}\\) and \\(\\sigma_{\\mathcal{B}}\\), it is better to multiply them by a Gaussian noise centered at 1 with proper variance instead of add them by two zero-mean noises. One reason is that the standard deviation of \\(\\sigma_{\\mathcal{B}}\\) is proportional to \\(\\sigma\\) when \\(m\\) is sufficiently large. More importantly, it is impossible to choose the right amount of noise for each \\(\\mu_{\\mathcal{B}}\\) and \\(\\sigma_{\\mathcal{B}}\\) if we add noises additively. Therefore, we multiplicatively inject noise into its mean and variance estimations of LBN as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.696, + 0.402, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\mu_ {\\mathcal {B}} \\leftarrow \\mu_ {\\mathcal {B}} \\cdot n _ {\\mu} \\text {a n d} \\sigma_ {\\mathcal {B}} \\leftarrow \\sigma_ {\\mathcal {B}} \\cdot n _ {\\sigma},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.795 + ], + "angle": 0, + "content": "where \\( n_{\\mu} \\) and \\( n_{\\sigma} \\) are two noises sampled from \\( \\mathcal{N}(1,0.1) \\) and they are treated as untrainable parameters in each back propagation. The result is presented in Fig 1. It shows that with such noise, the performance of LBN can be boosted to be comparable with BN." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.9 + ], + "angle": 0, + "content": "The main conclusion we can reach in the experiment above is that data-independent noise can be used to improve the performance of LBN. Moreover, it also indicates that the effect of BN on improving generalization can be decomposed into two parts: one is population normalization (can be seen from the improvement of LBN over w/o bn) and the other is the stochastic uncertainty contained in the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "mini-batch statistics (can be seen from the improvement of LBN+noise over LBN). This phenomenon inspires us to inject stochastic noise into our PN to improve its generalization performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.16, + 0.796, + 0.175 + ], + "angle": 0, + "content": "4.2. Noisy Population Normalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.181, + 0.905, + 0.256 + ], + "angle": 0, + "content": "Below, we develop our final method NPN (Alg.2) to achieve better generalization performance. It is inspired from our above understandings of the stochastic uncertainty in BN. In particular, we showed that data independent noise can be injected into LBN to boost the generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.905, + 0.352 + ], + "angle": 0, + "content": "At first, for each activation \\( x \\), we insert data-independent noise into the corresponding \\( \\mu \\) and \\( \\gamma \\) to improve generalization. In order to reduce the effects of different magnitudes of \\( \\mu \\) and \\( \\gamma \\) corresponding to different neurons, similar to LBN+noise in Section 4.1, we multiply \\( \\mu \\) and \\( \\gamma \\) by two noises centered at 1 instead of adding two zero-mean ones as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.359, + 0.849, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mu} = \\mu \\left(1 + n _ {\\mu}\\right) \\text {a n d} \\hat {\\gamma} = \\max \\{0, \\gamma \\left(1 + n _ {\\gamma}\\right) \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.905, + 0.422 + ], + "angle": 0, + "content": "where \\(n_{\\mu}\\) and \\(n_{\\gamma}\\) are independently sampled noises from a zero-mean distribution, e.g., Gaussian, and max is used to avoid negative \\(\\hat{\\gamma}\\). We then normalize, scale and shift \\(x\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.427, + 0.905, + 0.458 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = \\frac {x - \\hat {\\mu}}{\\sqrt {\\hat {\\gamma} ^ {2} + \\epsilon}} \\text {a n d} y = \\alpha \\hat {x} + \\beta , \\tag {NPN}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.905, + 0.536 + ], + "angle": 0, + "content": "where \\(\\mu, \\gamma, \\alpha\\) and \\(\\beta\\) are trainable parameters. The training process with NPN is almost the same as PN, except that the injected noises \\(n_{\\mu}\\) and \\(n_{\\gamma}\\) are treated as constants in back propagation. The details of the training process can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.718 + ], + "angle": 0, + "content": "Discussion. NPN has the following appealing features. As the injected noises \\( n_{\\mu} \\) and \\( n_{\\gamma} \\) are data independent, we are able to flexibly inject a proper amount of noise into the learnt estimation of statistics to improve generalization ability, which is particularly important when large batch sizes are used. This scenario is often encountered in cross silo federated learning. The idea of injecting data independent noise to improve generalization can also be combined with other normalization methods. The success of NPN shows that one can directly learn the population statistics in the training process instead of estimating them from the minibatches. This can inspire new normalization methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.725, + 0.907, + 0.892 + ], + "angle": 0, + "content": "Remark 4.1 It is unclear that whether NPN has some other nice properties of BN, e.g., auto-tuning of learning rates due to scale invariance [1, 7], while our empirical results demonstrate the effectiveness of NPN. We argue that different from standard training, the performance improvements achieved by these properties could be negligible compared to normalization itself in FL as the training in FL is more challenging than standard training. Such theoretical analysis is left as future work. Thus, we would like to point out that NPN is a normalization technique for FL instead of a general approach." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10218" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.218, + 0.088, + 0.782, + 0.406 + ], + "angle": 0, + "content": "
batch sizeNormalizationα (CIFAR-10)α (CIFAR-100)
0.040.080.161.000.040.080.161.00
8w/o69.8383.2285.0587.5945.1350.4354.2658.47
IN67.6878.5582.6284.0935.5140.0245.4845.45
GN71.0582.5483.3586.1145.4549.2252.2054.75
BN70.2282.2785.7789.4252.1355.9158.1664.88
FedBN71.4983.0584.4189.3052.0556.1257.9764.43
PN78.5985.4588.0589.6353.0657.4361.8765.00
4w/o70.6982.0285.1186.9346.9950.8355.0859.45
IN60.5374.0878.1182.3936.0442.9646.1749.89
GN68.6378.5382.0185.5244.2148.1851.8554.67
BN70.1882.4786.3289.5251.3155.7860.3564.32
FedBN72.2483.5786.0389.4651.4555.7759.9165.10
PN78.4085.0587.5289.5952.2856.8562.4265.82
2w/o67.0379.9883.4387.0143.9348.5654.7759.43
IN55.4268.9375.4378.7936.1543.9047.7247.30
GN66.0278.8281.7585.2344.4046.1750.6053.24
BN65.3279.6883.3087.2545.8750.6655.0658.74
FedBN65.5080.0282.1786.4046.0451.3055.5259.68
PN72.5382.5986.2388.9848.9754.9260.6465.46
1w/o60.4971.8378.6584.9338.2442.2649.4954.59
IN35.4754.9366.4573.7433.6039.0643.2948.82
GN60.5571.8375.2579.6846.0449.2653.9756.64
BNNANANANANANANANA
FedBNNANANANANANANANA
PN64.3878.8982.1386.6946.1349.3259.1962.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.415, + 0.907, + 0.445 + ], + "angle": 0, + "content": "Table 1. The results of different normalization methods on CIFAR-10/100 with different batch sizes and Non-IID degree \\( \\alpha \\) . Or PN consistently outperforms other methods by a large margin across different \\( \\alpha \\) and batch sizes." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.47, + 0.398, + 0.485 + ], + "angle": 0, + "content": "Algorithm 2 Noisy Population Normalization" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.49, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Input: activation: \\( x \\); population mean and variance: \\( \\mu \\) and \\( 1 / \\gamma^2 \\); scale and shift parameter: \\( \\alpha \\) and \\( \\beta \\); noise level: \\( \\sigma \\)." + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.535, + 0.277, + 0.549 + ], + "angle": 0, + "content": "Inject noise into \\(\\mu\\) and \\(\\gamma\\):" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.559, + 0.435, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mu} = \\mu \\left(1 + n _ {\\mu}\\right) \\text {a n d} \\hat {\\gamma} = \\max \\{0, \\gamma \\left(1 + n _ {\\gamma}\\right) \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.583, + 0.401, + 0.614 + ], + "angle": 0, + "content": "where \\(n_{\\mu}\\) and \\(n_{\\gamma}\\) are noises from \\(\\mathcal{N}(0,\\sigma^2)\\). Normalize using noisy population statistics:" + }, + { + "type": "equation", + "bbox": [ + 0.248, + 0.623, + 0.341, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = \\frac {x - \\hat {\\mu}}{\\sqrt {\\hat {\\gamma} ^ {2} + \\epsilon}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.667, + 0.214, + 0.681 + ], + "angle": 0, + "content": "Scale and shift:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.692, + 0.387, + 0.71 + ], + "angle": 0, + "content": "\\[\ny = \\alpha \\hat {x} + \\beta \\equiv \\mathrm {N P N} _ {\\alpha , \\beta} ^ {\\mu , \\gamma} (\\sigma ; x)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.733, + 0.182, + 0.749 + ], + "angle": 0, + "content": "Return: y" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.224, + 0.81 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "In this section, we conduct two sets of experiments. Firstly, we verify the effectiveness of of PN in the federated learning setting given small batch sizes and heterogeneous client data. Secondly, we demonstrate the superior generalization ability of our NPN given large training batch sizes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.72, + 0.485 + ], + "angle": 0, + "content": "5.1. Settings and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.643 + ], + "angle": 0, + "content": "To verify the advantage of our population normalization method, we compare it with other normalization counterparts in the FL setting. Specifically, we compare with batch normalization [16], group normalization [40], instance normalization [37] and without normalization. We compare with FedBN [25] by averaging the BN parameters in the clients. Extensive experiments are conducted on two image classification benchmarks (i.e., CIFAR-10 and CIFAR-100) with models VGG-16 [33] and ResNet [11] to evaluate the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For FL experiments, we adopt FedAvg [30] as the base algorithm. We use 80 clients in total with participation rate set to 0.4 throughout all experiments. For the clients' local update, we adopt SGD with momentum. For batch size of 8, the learning rates are set to be the best among \\(\\{0.005, 0.01, 0.15, 0.02, 0.25\\}\\) for each normalization method, and linearly down-scaled for other smaller batch sizes. We run 200 global communication rounds with local epoch set to 1. We examine the superiority of our method against the baseline under various batch sizes and heterogeneity degrees. We follow prior work [5, 28] to use Dirichlet distribution for simulating the non-IID data distribution, where the degree of heterogeneity is defined by \\(\\alpha\\), smaller \\(\\alpha\\) value corresponds to more severe heterogeneity. We report the global model's average performance in the last five rounds evaluated using the test split of the datasets. For other details, please refer to the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10219" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.094, + 0.379, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.094, + 0.621, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.091, + 0.865, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.235, + 0.385, + 0.248 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.235, + 0.75, + 0.247 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.253, + 0.89, + 0.268 + ], + "angle": 0, + "content": "Figure 2. (a) Training (left) and test (right) loss of BN, PN, and NPN as the training processes; (b) Sensitivity of NPN on noise level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.458, + 0.385 + ], + "angle": 0, + "content": "For experiments with large batch sizes, we adopt the distributed training setting, where the model is being trained 4 different participants. The batch size is set to 2048. Other hyper-parameters are set to be the same as in the setting. We conduct the experiments using VGG-16 train on CIFAR-10 and CIFAR-100 datasets." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.396, + 0.452, + 0.412 + ], + "angle": 0, + "content": "5.2. Federated Learning with Small Batch Size" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.495 + ], + "angle": 0, + "content": "We point out that with small batch sizes, SGD already introduces enough stochastic uncertainty in gradient estimation, which helps generalization. Therefore, we do not need to inject additional noise for PN and thus the experiments with NPN are omitted in this section." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Comparison with Various Normalization Methods We compare PN with other counterparts under different batch sizes and non-iid degrees in Table 1. Specifically, we conduct experiments with \\(\\alpha \\in \\{0.04, 0.08, 0.16, 1.00\\}\\) and batch sizes \\(\\{1, 2, 4, 8\\}\\). We adopt VGG16 as the trained network. As demonstrated in the table, we observe that our method consistently outperforms BN and other baselines, while the advantage is more significant when the batch size is small and the heterogeneity degree is higher. Notably, our method works well even with batch size set to 1, while the vanilla BN becomes inapplicable under this scenario. This enables our method to be used when the computational resource is limited on the client side." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Experiments with Various Networks We confirm the effectiveness of PN across different network architectures and capacities. As shown in Table 3, we conduct experiments with VGG16, ResNet18 and ResNet50. We fix the heterogeneity degree \\(\\alpha\\) to be 0.08 and trial with batch sizes of \\(\\{1,2,4,8\\}\\). We observe consistent performance advantage over the baselines for different networks. Interestingly, one may notice that networks with higher capacities do not necessarily lead to better performances, which is opposite to standard training. This phenomenon is also reported in [5], which may be due to clients' local over-fitting or further" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.907, + 0.34 + ], + "angle": 0, + "content": "drifting from each other in the parameter space. However, PN still demonstrates superior performances compared with other normalization techniques." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.353, + 0.875, + 0.37 + ], + "angle": 0, + "content": "5.3. Federated Learning with Large Batch Size" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.573 + ], + "angle": 0, + "content": "We now extend our PN to NPN to enable stronger generalization ability for training with large batch sizes. The results are demonstrated in Table 4. As can be seen, when training with large batch sizes, PN is outperformed by BN due to the lack of stochastic uncertainty in the normalization layer. On the other hand, NPN remedies this issue by injecting artificial noise into the statistics estimation, which in turn leads to better test performance. The training progress is also visualized in Figure 2(a). We can see that although the training loss of NPN is larger than that of BN, its testing loss is significantly smaller than BN. This demonstrates that benefiting from the injected noise, NPN can prevent the training process from over-fitting more effectively than BN." + }, + { + "type": "table", + "bbox": [ + 0.563, + 0.588, + 0.859, + 0.638 + ], + "angle": 0, + "content": "
MethodLossTop-1Top-5Top-10
BRN0.4493.2197.5298.83
BRN + noise0.2993.6097.8399.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.588, + 0.649, + 0.831, + 0.662 + ], + "angle": 0, + "content": "Table 2. Batch Renormalization + noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.906, + 0.759 + ], + "angle": 0, + "content": "Finally, we point out that there are only a small number of participants in this scenario, which makes the trained model less prone to the adverse impact of heterogeneity and thus shows better performance than the previous experiments with more participating clients (Table 1)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.714, + 0.789 + ], + "angle": 0, + "content": "5.4. Abative Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Noise Level Sensitivity. We evaluate the sensitivity of NPN to the injected noise level. We conduct experiments with VGG-16 on CIFAR-10. We set the batch size to 2048 and vary the variance of the noise injected into NPN in \\(\\{0.5, 0.1, 0.01, 0.001\\}\\). The results are presented in Figure 2(b). It shows that the test loss increases gradually and slowly when we vary the variance of the noise from 0.5 to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10220" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.907, + 0.193 + ], + "angle": 0, + "content": "
Normalizationbatch size (VGG-16)batch size (ResNet-18)batch size (ResNet-50)
124812481248
w/o71.8379.9882.0283.2275.4879.9480.2181.4965.5972.3377.8079.94
IN35.4768.9374.0378.5563.5369.3670.6871.3560.3764.6568.2869.10
GN60.5578.8278.5382.5466.0370.4472.5975.1360.6368.7270.7371.25
BNNA78.6582.0982.27NA78.5579.6481.94NA72.3978.2081.83
PN78.8982.2285.1385.4580.2083.9784.5384.6767.7273.6379.5983.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.203, + 0.907, + 0.23 + ], + "angle": 0, + "content": "Table 3. Results of different normalization methods on CIFAR-10. PN consistently outperforms other methods for different nwork architectures and batch sizes." + }, + { + "type": "table", + "bbox": [ + 0.255, + 0.244, + 0.744, + 0.345 + ], + "angle": 0, + "content": "
NormalizationCIFAR-10CIFAR-100
LossTop-1Top-2Top-3LossTop-1Top-2Top-3
w/o0.5492.3196.9198.202.7770.3487.9692.11
IN0.4692.4597.0098.312.5470.6888.4992.33
GN0.4392.9297.1498.522.2371.3588.9393.10
BN0.3993.1397.6698.872.0572.2190.2893.60
PN0.4193.1097.7898.971.8572.1590.0793.89
NPN0.3293.6698.0599.121.2272.7491.4794.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.198, + 0.355, + 0.798, + 0.37 + ], + "angle": 0, + "content": "Table 4. Results of different normalization methods on CIFAR-10 / CIFAR-100 with large batch size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.482, + 0.427 + ], + "angle": 0, + "content": "0.001. Therefore, it is easy to select a suitable noise variance that falls within a reasonable range." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.446, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Artificial Noise for Other Normalization Methods. We demonstrate that the artificially injected noise is also able to improve existing normalization methods with limited stochastic uncertainty. We take batch renormalization (BRN) [16] as an example and conduct experiments on CIFAR-10 with VGG-16. For simplicity and excluding other factors, we conduct this experiment with regular training instead of FL. We inject noise sampled from \\(\\mathcal{N}(0,0.1)\\) to the moving averages \\(\\mu\\) and \\(\\sigma\\) of BRN as we do for NPN. The results are given in Table 2. It shows that with the injected noise, the performance of BRN can be improved effectively. Therefore, the idea of data-independent noise injection is a very general regularization technique that can be used to improve existing normalization methods." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.667, + 0.355, + 0.684 + ], + "angle": 0, + "content": "5.5. Supervised Learning with PN" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Intuitively, PN can also be used in non-FL scenarios. We apply PN to supervised learning and conduct experiments on two image classification benchmarks (i.e., CIFAR-10, CIFAR-100) with VGG-16 and ResNet to evaluate the effectiveness of PN in supervised learning. The batch sizes of all the methods are set to 128. Table 5 presents the test loss and accuracy of the final learned models. From Table 5 we can see that our NPN can achieve much lower test loss than BN and comparable (if no better than) test accuracy with BN on all the three models. We can also observe that NPN can achieve a noticeable improvement over PN, which demonstrates that artificially injected data independent noise can be used to reinforce the performance of the normalization methods with limited stochastic under" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.394, + 0.905, + 0.422 + ], + "angle": 0, + "content": "Table 5. Results of different normalization methods on image classification benchmarks" + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.433, + 0.88, + 0.591 + ], + "angle": 0, + "content": "
NetworkNormalizationLossAccuracy
VGG-16BN0.3992.31
PN0.4193.63
NPN0.2893.66
ResNet-18BN0.4693.76
PN0.4593.97
NPN0.3894.26
ResNet-56BN0.3393.05
PN0.4292.56
NPN0.2593.99
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.907, + 0.679 + ], + "angle": 0, + "content": "tainty. The comparable performance of BN and NPN implies that the effect of BN can be decomposed into two parts: one is population normalization, and the other is the stochastic uncertainty containing in the mini-batch statistics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.64, + 0.71 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose a learning-based normalization technique, Population Normalization (PN), which reduces the reliance on training batch size and improves robustness against data heterogeneity in federated learning (FL). Extensive experiments demonstrate that PN outperforms existing normalization techniques by a significant margin. We further develop Noisy Population Normalization (NPN), which injects artificial noise into the learnt normalization statistics to enhance generalization under large batch size conditions. We believe that our work will offer valuable insights for the design and development of more stable and adaptable normalization techniques for FL." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10221" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.201 + ], + "angle": 0, + "content": "This work was supported by the National Nature Science Foundation of China (62472097) and AI for Science Foundation of Fudan University (FudanX24AI028). The computations in this research were performed on the CFFF platform of Fudan University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.212, + 0.188, + 0.228 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.237, + 0.483, + 0.293 + ], + "angle": 0, + "content": "[1] Sanjeev Arora, Kaifeng Lyu, and Zhiyuan Li. Theoretical analysis of auto rate-tuning by batch normalization. In 7th International Conference on Learning Representations, ICLR 2019, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.293, + 0.484, + 0.335 + ], + "angle": 0, + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.336, + 0.484, + 0.406 + ], + "angle": 0, + "content": "[3] Tobias Bernecker, Annette Peters, Christopher L. Schlett, Fabian Bamberg, Fabian J. Theis, Daniel Rueckert, Jakob Weiss, and Shadi Albarqouni. Fednorm: Modality-based normalization in federated learning for multi-modal liver segmentation. ArXiv, abs/2205.11096, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.407, + 0.484, + 0.463 + ], + "angle": 0, + "content": "[4] Yongqiang Cai, Qianxiao Li, and Zuowei Shen. A quantitative analysis of the effect of batch normalization on gradient descent. In International Conference on Machine Learning, pages 882-890, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.464, + 0.484, + 0.505 + ], + "angle": 0, + "content": "[5] Hong-You Chen and Wei-Lun Chao. Fedbe: Making bayesian model ensemble applicable to federated learning. In ICLR, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.484, + 0.535 + ], + "angle": 0, + "content": "[6] Wikipedia contributors. Kurtosis. Online; accessed 2-June-2020, 2020. Wikipedia, The Free Encyclopedia. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.484, + 0.59 + ], + "angle": 0, + "content": "[7] Hadi Daneshmand, Amir Joudaki, and Francis Bach. Batch normalization orthogonalizes representations in deep random networks. In Advances in Neural Information Processing Systems, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.484, + 0.662 + ], + "angle": 0, + "content": "[8] Jian-Hui Duan, Wenzhong Li, and Sanglu Lu. Feddna: Federated learning with decoupled normalization-layer aggregation for non-iid data. In Machine Learning and Knowledge Discovery in Databases. Research Track, pages 722-737, Cham, 2021. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.662, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[9] Aiden Durrant, Milan Markovic, David Matthews, David May, Jessica Enright, and Georgios Leontidis. The role of cross-silo federated learning in facilitating data sharing in the agri-food sector. Computers and Electronics in Agriculture, 193:106648, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[10] Filip Hanzely, Slavomír Hanzely, Samuel Horváth, and Peter Richtárik. Lower bounds and optimal algorithms for personalized federated learning. Advances in Neural Information Processing Systems, 33:2304-2315, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.237, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[13] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip Gibbons. The non-IID data quagmire of decentralized machine learning. In Proceedings of the 37th International Conference on Machine Learning, pages 4387-4398. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[14] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip B. Gibbons. The non-iid data quagmire of decentralized machine learning, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[15] Chao Huang, Jianwei Huang, and Xin Liu. Cross-silo federated learning: Challenges and opportunities. arXiv preprint arXiv:2206.12949, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.305 + ], + "angle": 0, + "content": "[16] Sergey Ioffe. Batch renormalization: Towards reducing minibatch dependence in batch-normalized models. In Advances in neural information processing systems, pages 1945-1953, 2017. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[17] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning, pages 448-456, Lille, France, 2015. PMLR. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.906, + 0.433 + ], + "angle": 0, + "content": "[18] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for ondevice federated learning. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.49 + ], + "angle": 0, + "content": "[19] Jakub Konečny, H Brendan McMahan, Felix X Yu, Peter Richtárik, Ananda Theertha Suresh, and Dave Bacon. Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.559 + ], + "angle": 0, + "content": "[20] Viraj Kulkarni, Milind Kulkarni, and Aniruddha Pant. Survey of personalization techniques for federated learning. In 2020 Fourth World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4), pages 794-797. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.906, + 0.602 + ], + "angle": 0, + "content": "[21] Qinbin Li, Bingsheng He, and Dawn Song. Practical oneshot federated learning for cross-silo setting. arXiv preprint arXiv:2010.01017, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[22] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10713–10722, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[23] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In MLSys, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.906, + 0.759 + ], + "angle": 0, + "content": "[24] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through personalization. In International Conference on Machine Learning, pages 6357-6368. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[25] Xiaoxiao Li, Meirui Jiang, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fed{bn}: Federated learning on non- \\(\\{\\mathrm{i}\\mathrm{d}\\}\\) features via local batch normalization. In International Conference on Learning Representations, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[26] Xiangru Lian and Ji Liu. Revisit batch normalization: New understanding and refinement via composition optimization. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 3254-3263, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.903 + ], + "angle": 0, + "content": "[27] Wei Yang Bryan Lim, Nguyen Cong Luong, Dinh Thai Hoang, Yutao Jiao, Ying-Chang Liang, Qiang Yang, Dusit" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10222" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Niyato, and Chunyan Miao. Federated learning in mobile edge networks: A comprehensive survey. IEEE Communications Surveys & Tutorials, 22(3):2031-2063, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[28] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. Advances in Neural Information Processing Systems, 33:2351-2363, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[29] Ping Luo, Xinjiang Wang, Wenqi Shao, and Zhanglin Peng. Towards understanding regularization in batch normalization. arXiv preprint arXiv:1809.00846, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[31] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. In Advances in neural information processing systems, pages 901-909, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[32] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? In Advances in Neural Information Processing Systems, pages 2483-2493, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[33] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[34] Sebastian U Stich. Local sgd converges fast and communicates little. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[35] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[36] Mattias Teye, Hossein Azizpour, and Kevin Smith. Bayesian uncertainty estimation for batch normalized deep networks. In International Conference on Machine Learning, pages 4907-4916, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[37] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[38] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[39] Xiaoxia Wu, Edgar Dobriban, Tongzheng Ren, Shanshan Wu, Zhiyuan Li, Suriya Gunasekar, Rachel Ward, and Qiang Liu. Implicit regularization of normalization methods. arXiv preprint arXiv:1911.07956, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[40] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-19, 2018. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[41] Jianhang Xiao, Chunhui Du, Zijing Duan, and Wei Guo. A novel server-side aggregation strategy for federated learning in non-id situations. In 2021 20th International Symposium on Parallel and Distributed Computing (ISPDC), pages 17-24. IEEE, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[42] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[43] Yousef Yeganeh, Azade Farshad, Nassir Navab, and Shadi Albarqouni. Inverse distance aggregation for federated learning with non-iid data. In Domain Adaptation and Representation Transfer, and Distributed and Collaborative Learning, pages 150-159. Springer, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10223" + } + ] +] \ No newline at end of file diff --git a/2025/Population Normalization for Federated Learning/b6b8182e-2812-44b6-99cb-82e1b12982cb_origin.pdf b/2025/Population Normalization for Federated Learning/b6b8182e-2812-44b6-99cb-82e1b12982cb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2af602047ac8855d955a6a2b76f546c132cbe060 --- /dev/null +++ b/2025/Population Normalization for Federated Learning/b6b8182e-2812-44b6-99cb-82e1b12982cb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:378706588910c129e0e662e6d4d6f24ca5fec05e5fc21b5f8beec45937cbd9f0 +size 527937 diff --git a/2025/Population Normalization for Federated Learning/full.md b/2025/Population Normalization for Federated Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6a905804b3ce8bca26512fde35038d86576580f4 --- /dev/null +++ b/2025/Population Normalization for Federated Learning/full.md @@ -0,0 +1,359 @@ +# Population Normalization for Federated Learning + +Zhuoyao Wang $^{1*}$ , Fan Yi $^{1*}$ , Peizhu Gong $^{1}$ , Caitou He $^{2}$ , Cheng Jin $^{1,3}$ , Weizhong Zhang $^{1,4\dagger}$ + +$^{1}$ Fudan University $^{2}$ Wenzhou University + +3Innovation Center of Calligraphy and Painting Creation Technology, MCT, China + +$^{4}$ Shanghai Key Laboratory of Intelligent Information Processing + +{18110240064,pzgong,jc,weizhongzhang}@fudan.edu.cn，ftyi21@m.fudan.edu.cn，hctou89@wzu.edu.cn + +# Abstract + +Batch normalization (BN) is widely recognized as an essential method in training deep neural networks, facilitating convergence and enhancing model stability. However, in Federated Learning (FL) contexts, where training data are typically heterogeneous and clients often face resource constraints, the effectiveness of BN is considerably limited for two primary reasons. First, the population statistics, specifically the mean and variance of these heterogeneous datasets, vary substantially, resulting in inconsistent BN layers across client models, which ultimately drives these models to diverge further. Second, estimating statistics from a mini-batch is often imprecise since the batch size has to be small in resource-limited clients. This paper introduces Population Normalization, a novel technique for FL, in which the statistics are learned as trainable parameters rather than calculated from mini-batches as in BN. Thus, our normalization layers are homogeneous among the clients and the adverse impact of small batch size is eliminated as the model can be well-trained even when the batch size equals to one. To enhance the flexibility of our method in practical applications, we investigate the role of stochastic uncertainty in BN's statistical estimation. When larger batch sizes are available, we demonstrate that injecting simple artificial noise can effectively mimic this stochastic uncertainty and improve the model's generalization capability. Experimental results validate the efficacy of our approach across various FL tasks. + +# 1. Introduction + +Batch Normalization (BN) [17] has become a standard component in deep neural networks (DNNs). It is proposed to stabilize and accelerate the training of DNNs, and the original motivation was to reduce the so-called internal covariate shift. In each training step, it normalizes the layer input of a neural network using the mean and variance com + +puted within a randomly sampled mini-batch. Specifically, consider a particular neuron in a network, and we denote $\{x_{1},\ldots ,x_{m}\}$ to be its pre-activations on a mini-batch $\mathcal{B}$ of size $m$ , BN first normalizes each $x_{i}$ as + +$$ +\hat {x} _ {i} = \frac {x _ {i} - \mu_ {\mathcal {B}}}{\sigma_ {\mathcal {B}}}, \tag {1} +$$ + +where $\mu_{\mathcal{B}}$ and $\sigma_{\mathcal{B}}^{2}$ are the sample mean and variance of the pre-activation on $\mathcal{B}$ . It then scales and shifts each $\hat{x}_i$ as $y_i = \alpha \hat{x}_i + \beta$ to compensate the effect of normalization, where $\gamma$ and $\beta$ are two parameters to be learned during the training process. BN finally takes $y_i$ as the new pre-activation of this neuron. It has been reported in literature [12, 16, 17, 37, 40] that BN can significantly improve the convergence of training and the generalization performance of the final learned models. + +However, we notice that the effectiveness of BN is compromised in Federated Learning [14, 19, 42], which typically aims at collaboratively training models on resource limited clients (e.g., mobiles) with locally private and heterogeneous training data. The main reasons can be summarized as follows. One is that the statistics, i.e., the mean and variance, of heterogeneous datasets in the clients differ greatly. Eqn. (1) implies that this can lead to inconsistent BN layers among the client models and finally causes these models to drift further away from each other. The other is that in resource limited clients, the mini-batch size has to be small, which makes $\mu_{\mathcal{B}}$ and $\sigma_{\mathcal{B}}^{2}$ inaccurate estimations of the population mean and variance for normalization and the obtained BN would cause instability during training that slows down convergence. + +Some other normalization techniques, such as batch renormalization [16], group normalization [40], and layer normalization [2], are proposed to address this heavy reliance on batch size. The key ideas are to estimate the population mean and variance as accurately as possible by using techniques such as moving average [16] or to normalize the activation along other dimensions such as the channel dimension [2]. Nevertheless, these approaches focus on training acceleration, which typically under-perform + +BN in terms of the trained model's accuracy [40]. Recently, FedBN [25] is proposed to extend BN to FL. It keeps the client BN layers updated locally, without communicating and aggregating them during training. This enables the clients to achieve better performances in their own respective environments. However, since the BN layers among the clients can be heterogeneous due to the non-IID datasets, FedBN cannot produce an off-the-shelf global model after training, which is often required in reality. Consequently, FedBN suffers from the cold start issue, i.e., the learned model cannot be directly deployed in the new clients due to the lack of trained local BN layers. What's more, FedBN still relies on estimation within mini-batches, which cannot address the issue of inaccurate estimation caused by small batch size in cross-device FL. + +In this paper, we propose Population Normalization (PN), a simple yet effective normalization method for FL to eliminate the negative impacts of statistical heterogeneity and estimation inaccuracies on BN. The key idea is to learn population statistics as trainable parameters in the training process by introducing two constraints derived from the definitions of mean and variance (see Section 3), instead of estimating them from the mini-batches as most existing methods have done. As the population statistics take place of the sample mean and variance of the local datasets in BN, and are treated as normal trainable parameters in local training and aggregation, our normalization layers among the clients would be consistent although the datasets are heterogeneous. This enables our method to eventually produce a global model with the normalization layer among the clients. In addition, since we no longer need to calculate the sample mean and variance, the resulting training problem can be solved by stochastic algorithms of FL with arbitrary batch sizes (even equals to one) and thus the issue of inaccurate statistic estimation because of small batch size is eliminated. + +In practice, we observe that PN tends to under-perform BN given very large batch sizes. We then further explore this issue and find its cause to be that PN lacks the stochastic uncertainty in the estimation of mean and variance. Specifically, we show that the noise introduced by the mini-batch based mean and variance estimations in the BN is essential for its good generalization behavior. Our empirical results demonstrate that injecting a reasonable amount of data independent noise into BN with extremely large batch sizes can significantly improve the generalization performance to be comparable to that of BN with proper batch sizes. Inspired by this finding, we propose Noisy Population Normalization (NPN) by injecting simple data independent artificial noises into the learned population mean and variance in PN during training. The injected noise can serve the same purpose as the uncertainty in estimating population statistics in BN with small batch sizes. This makes the proposed + +method more flexible in the applications of FL where large batch size is available, e.g., collaboration among companies in cross-silo FL [9, 15, 21]. + +Finally, we conduct a series of experiments to verify the effectiveness of our proposed PN and NPN in the context of FL with small and large batch sizes, respectively (Section 5). The results demonstrate that PN effectively reduces the reliance on training batch size and alleviates the negative impact of data heterogeneity on normalization. Furthermore, the artificial noise injected in our NPN method effectively improves the generalization performance given large training batch sizes. + +Our main contributions are summarized as follows: + +- We propose a normalization technique PN for federated learning, which demonstrates superior performance given small batch sizes and data heterogeneity. +- In training with large batch sizes, we propose NPN based on PN, which introduces stochastic uncertainty by injecting artificial noise to improve generalization ability. +- We conduct extensive experiments to verify the effectiveness of our method. + +Notations: Given a positive integer $d$ , we define $[d]$ to be the set $\{1,2,\dots,d\}$ . For a vector $x\in \mathbb{R}^d$ , we let $[x]_i$ be its $i$ -th component with $i\in [d]$ . We denote by $w_{i,j}$ the element in the $i$ -th row and the $j$ -th column of a real-valued matrix $w$ . Let $A\cup B$ be the union of two sets $A$ and $B$ . Moreover, we denote by $\mathcal{N}(\mu ,\sigma^2)$ the Gaussian distribution with mean of $\mu$ and variance of $\sigma^2$ . + +# 2. Related Work + +# 2.1. Normalization Techniques + +BN [17] is a standard normalization technique, which enables the training of very deep neural networks, and the most representative result is ResNet[11], which cannot be easily trained to converge without BN. It can also achieve significant imporement in generalization. Therefore, the invention of BN is a major milestone in the development of deep learning. + +As aforementioned, small batch size or non-iid minibatch can diminish the effect of BN on the training acceleration or even result in training failure, due to the inaccurate mean and variance estimations [16]. Moreover, large batch size leads to huge memory cost. This issue hinders BN's usage in FL applications, where the clients often have limited computational resources and the datasets are usually heterogeneous. This in turn prevents the training of large networks in FL. Recently, some new normalization methods have been proposed as alternatives to BN, and they can be roughly divided into two categories, i.e., normalization along the batch dimension [16], and mini-batch independent normalization [2, 31, 37, 40]. The former alleviates the reliance on the batch size by using techniques such as + +moving average. The latter ones achieve this by normalizing the activation of neurons in a network along other dimensions. For example, Group Normalization [40] divides the channels in each hidden layer into groups and normalizes the activations in each group with its mean and variance. Empirical studies show that all these methods achieve comparable training speedups with BN. However, the generalization performance they achieved is inferior to that of BN. + +Besides the above studies, there is another line of research [4, 26, 29, 32, 39] that focuses on understanding BN. One of the main conclusions in these investigations is that with BN, the effective objective function used in training is different from the original one. Specifically, [32] showed that by inserting BN layers into a network, the loss landscape becomes smoother, which naturally explains why BN enables one to use large learning rates. The studies [29, 39] focus on the regularization effect of BN. They showed that the objective function with BN can be divided into two parts: an empirical loss under population normalization and an explicit regularizer named gamma decay. + +# 2.2. Federated Learning + +FL emerges as an effective paradigm to train models on distributed devices without centralizing local private data [3, 8, 13, 19, 42]. Specifically, it can be formulated as follows: + +$$ +\min _ {\theta} \mathcal {L} (\theta , \mathcal {D}) \triangleq \sum_ {m = 1} ^ {M} \rho_ {m} \mathcal {L} _ {m} (\theta , \mathcal {D} _ {m}), \tag {2} +$$ + +where $M$ is the number of clients and $\theta$ is the model to be learnt. The global dataset $\mathcal{D} = \cup_{m=1}^{M}\mathcal{D}_{m}$ , where $\mathcal{D}_{m} = \{(\pmb{x}_{i},y_{i})\}_{i=1}^{|D_{m}|}$ is the locally kept dataset on client $m$ . $\mathcal{L}_{m}(\theta,\mathcal{D}_{m}) = \frac{1}{|\mathcal{D}_{m}|}\sum_{\xi\in\mathcal{D}_{m}}\ell(\theta,\xi)$ denotes the local empirical risk with the loss function $\ell(\cdot,\cdot)$ and $\rho_{m} = \frac{|D_{m}|}{|\mathcal{D}|}$ . Typically, the most popular FL algorithm FedAvg [30] solves the problem (2) based on local-SGD [34]. Specifically, in the $c^{th}$ round, it updates the global model on the server as follows: + +$$ +\theta^ {c + 1} = \frac {1}{M} \sum_ {m = 1} ^ {M} \rho_ {m} \theta_ {m} ^ {c}, \tag {3} +$$ + +where $\theta_{m}^{c}$ denotes the received local model from client $m$ after local training in this round. The updated global model $\theta^{c + 1}$ is then distributed to the clients to initiate the $(c + 1)$ -th round of training. + +In FL, especially in mobile edge scenarios, both limited device capacity and non-IID distribution of clients' local data raise challenges to the training procedure [27]. Various works attempt to tackle the non-IID distribution of data for federated learning from the perspective of local training regularization [18, 22, 23], server aggregation [5, 28, 38, 41, 43], personalized federated learning [10, 20, 24, 35], + +# Algorithm 1 Population Normalization + +Input: activation: $x$ ; mean and variance: $\mu$ and $1 / \gamma^2$ ; scale and shift parameters: $\alpha$ and $\beta$ . + +Normalize using the population statistics: + +$$ +\hat {x} = \frac {x - \mu}{\sqrt {\gamma^ {2} + \epsilon}} +$$ + +Scale and shift: + +$$ +y = \alpha \hat {x} + \beta \equiv \mathrm {P N} _ {\alpha , \beta} ^ {\mu , \gamma} (x) +$$ + +# Return: y + +etc. However, they do not explore the reduced effectiveness of batch normalization in federated learning. FedBN [25] uses local BN parameters to tackle non-iid federated learning. However, FedBN is only applicable to local inferences where only clients participating in the training with local statistics can perform inference, without providing a global model for new clients or global data. Also, FedBN does not consider the small batches caused by limited local resources. + +# 3. Population Normalization + +We now present our method PN (see Alg.1). Its key idea is to eliminate the impacts of heterogeneous statistics and inaccurate statistic estimations by learning the population mean and variance during the training process, instead of estimating them within the mini-batch as BN has done. Precisely, given an activation $x$ , we normalize it as follows: + +$$ +\hat {x} = \frac {x - \mu}{\sqrt {\gamma^ {2} + \epsilon}}, +$$ + +where $\mu$ and $\gamma^2$ are the population mean and variance, respectively, and they will be learned during the training process, of which the details are specified later in Eqn.(3). $\epsilon > 0$ is a small constant. Then, to compensate the effect of normalization on the effective capacity of a neural network, we scale and shift the normalized activation $\hat{x}$ as: + +$$ +y = \alpha \hat {x} + \beta \equiv \mathrm {P N} _ {\alpha , \beta} ^ {\mu , \gamma} (x), \quad (\mathrm {P N}) +$$ + +where the coefficients $\alpha$ and $\beta$ are learned in the training process together with $\mu$ and $\gamma$ . + +Below, we take a fully connected neural network $f(\pmb{x}) \in \mathbb{R}$ as an example to show how to insert PN layers into DNNs. Specifically, given an input $\pmb{x} \in \mathbb{R}^d$ , let the input layer be $f_j^{(0)}(\pmb{x}) = [\pmb{x}]_j$ with $j \in [n^{(0)}] = [d]$ . The $\ell$ -th hidden layer containing $n^{(\ell)}$ hidden nodes with $\ell \in [L]$ takes + +the form of + +$$ +\begin{array}{l} f _ {j} ^ {(\ell)} (\boldsymbol {x}) = h ^ {(\ell)} \left(z _ {j} ^ {(\ell)} (\boldsymbol {x})\right) \text {w i t h} \tag {4} \\ z _ {j} ^ {(\ell)} (\boldsymbol {x}) = \sum_ {k = 1} ^ {n ^ {(\ell - 1)}} w _ {j, k} ^ {(\ell)} \underbrace {\left(\alpha_ {k} ^ {(\ell - 1)} \frac {f _ {k} ^ {(\ell - 1)} (\boldsymbol {x}) - \mu_ {k} ^ {(\ell - 1)}}{\sqrt {\left(\gamma_ {k} ^ {(\ell - 1)}\right) ^ {2} + \epsilon}} + \beta_ {k} ^ {(\ell - 1)}\right)} _ {(\ell - 1) - \text {t h P N l a y e r}}, \\ \end{array} +$$ + +where $j\in [n^{(\ell)}],z_j^{(\ell)}(\pmb {x})$ is the pre-activation, $h^{(\ell)}$ is the activation function and $w^{(\ell)}\in \mathbb{R}^{n^{(\ell)}\times n^{(\ell -1)}}$ is the weight matrix connecting layer $\ell$ and $\ell -1$ $\mu_k^{(\ell -1)}$ and $\gamma_{k}^{(\ell -1)}$ are the population mean and standard deviation of the hidden nodes $f_{k}^{(\ell -1)}(\pmb {x})$ $\alpha_{k}^{(\ell -1)}$ and $\beta_{k}^{(\ell -1)}$ are the parameters of the linear transformation. The final output layer is defined as + +$$ +f (\boldsymbol {x}) = \sum_ {k = 1} ^ {n ^ {(L)}} u _ {k} \underbrace {\left(\frac {f _ {k} ^ {(L)} (\boldsymbol {x}) - \mu_ {k} ^ {(L)}}{\sqrt {\left(\gamma_ {k} ^ {(L)}\right) ^ {2} + \epsilon}}\right)} _ {\text {f i n a l P N l a y e r}} + c _ {0}, \tag {5} +$$ + +where $u\in \mathbb{R}^{n^{(L)}}$ is the weight vector connecting the output to the $L$ -th hidden layer, $\mu_k^{(L)}$ and $\gamma_k^{(L)}$ are the population mean and standard deviation of the hidden nodes $f_{k}^{(L)}(\pmb {x})$ The scaling and shifting parameters are absorbed into $u_{k}$ and $c_{0}$ , respectively. + +To eliminate the dependency of PN on batch size, we learn the statistics $\mu_k^{(\ell)}$ and $\gamma_k^{(\ell)}$ in training by introducing two natural constraints derived from the definition of mean and variance. The statistic estimation is optimized together with the model parameters during local training, and aggregated on the server by FL algorithms, such as FedAVG. To be precise, given the loss $\phi(\cdot, \cdot): \mathbb{R} \times \{0,1\} \to \mathbb{R}$ and the global dataset $\mathcal{D} = \cup_{m=1}^{M} \mathcal{D}_m$ , the training objective in Eqn. (2) of a FL problem becomes + +$$ +\begin{array}{l} \min _ {\theta} \mathcal {L} (\theta , \mathcal {D}) s. t. \left\{ \begin{array}{l} \frac {1}{| \mathcal {D} |} \sum_ {(\boldsymbol {x}, y) \in \mathcal {D}} \frac {\left(f _ {k} ^ {(\ell)} (\boldsymbol {x}) - \mu_ {k} ^ {(\ell)}\right) ^ {2}}{\left(\gamma_ {k} ^ {(\ell)}\right) ^ {2} + \epsilon} = 1 \\ \frac {1}{| \mathcal {D} |} \sum_ {(\boldsymbol {x}, y) \in \mathcal {D}} f _ {k} ^ {(\ell)} (\boldsymbol {x}) = \mu_ {k} ^ {(\ell)} \end{array} \right., \\ k \in [ m ^ {(\ell)} ], \ell \in \{0 \} \cup [ L ] \\ \end{array} +$$ + +where $\theta$ is the set of all the trainable parameters, i.e., $\theta = \{w^{(\ell)}\} \cup \{u\} \cup \{\mu^{(\ell)}\} \cup \{\gamma^{(\ell)}\} \cup \{\alpha^{(\ell)}\} \cup \{\beta^{(\ell)}\} \cup \{c_0\}$ and $\mathcal{L}(\theta ,\mathcal{D}) = \sum_{m = 1}^{M}\rho_{m}\sum_{(\boldsymbol {x},y)\in \mathcal{D}_{m}}\phi (f(\boldsymbol {x}),y)$ . This training problem can be solved by various stochastic algorithms with arbitrary batch sizes (even equals to 1), such as moving average or the method of Lagrangian multipliers. In the experiments of this paper, we use stochastic Lagrangian multipliers and the details can be found in the appendix. + +Discussion. (1) Since the statistics $\mu$ and $\gamma$ are learned in the training process instead of being estimated from minibatches, the resulting training problem can be solved by + +stochastic solvers of FL with arbitrary batch sizes (even equals to 1). Thus the reliance on batch size is eliminated. This enables resource limited clients to train larger deep neural networks. (2) Both $\mu$ and $\gamma$ are treated as normal trainable parameters in local training, communication and aggregation, our PN layers are consistent among the clients although the datasets on the clients are heterogeneous. This enables us to eventually obtain an off-the-shelf global model, which is often required in reality. + +Remark 3.1 We should point out that our PN layer can also be inserted before the activation functions. + +![](images/a58edbede5168160128fe734085f58f0965127749be6532017594e7fe1aa878f.jpg) +Figure 1. Training loss (left) and test loss (right) of BN, LBN and LBN+noise. Injecting artificial noise into estimated BN statistics successfully decreases the test loss and improves generalization. + +![](images/81a9c9f0b96329886b774af40745de8efbfc5465b307444165d5496593bd56dd.jpg) + +# 4. Noisy Population Normalization + +We observe that PN demonstrates inferior performance relative to BN when batch size becomes larger. We explore this issue and find its cause to be that PN lacks stochastic uncertainty in the mean and variance estimations. In practice, large batch sizes may often be used in cross-silo federated learning, where the participants are companies and organizations with huge amounts of data as well as abundant computational resources [9, 15, 21]. Therefore, to enable our method to be more flexibly used in such scenarios, we propose NPN, which injects artificial noise into the learned statistics during training to boost generalization. The details are presented in the following subsections. + +# 4.1. The Role of Stochastic Uncertainty in BN + +We provide an in-depth exploration into the role of stochastic uncertainty of the mini-batch statistics in BN. This provides a deep understanding of BN's batch size reliance. + +We start with a simple experiment to show the effect of batch size on BN. Precisely, we compare the performances of BN with a reasonable batch size (m=128) and a very large size (m=2048) in training VGG-16 on CIFAR-10. For simplicity, we refer to BN with m=2048 as LBN. For fair comparison, in LBN, to eliminate the effect of large batch size on SGD, we only use large batches to obtain batch statistics and calculate the stochastic gradient with respect to the loss on mini-batches of size 128. + +We present both the training and test losses of the learned models with BN and LBN over the training process in Fig + +1, from which we can see that LBN converges faster than BN, due to the more accurate population statistics estimation. However, there is a noticeable gap between their test losses, which becomes larger after the learning rate decays. Therefore, the effect of BN in improving generalization is reduced when the batch size becomes too large. This indicates that the stochastic uncertainty in estimates of the mean and variance is important for BN. + +The next question we want to investigate is whether we can artificially inject a proper amount of data-independent noise into LBN to improve its generalization performance. Note that the noise of estimating $\mu_{\mathcal{B}}$ and $\sigma_{\mathcal{B}}$ in BN is data dependent. To answer this question, we should first calculate the uncertainty explicitly. The studies [29, 36] show that for sufficiently large batch size M, the estimated $\mu_{\mathcal{B}}$ and $\sigma_{\mathcal{B}}$ can be viewed as two random variables sampled from two Gaussian distributions, i.e., + +$$ +\mu_ {\mathcal {B}} \sim \mathcal {N} \left(\mu , \frac {\sigma^ {2}}{m}\right) \text {a n d} \sigma_ {\mathcal {B}} \sim \mathcal {N} \left(\sigma , \left(\frac {\mu_ {4} / \sigma^ {4} - 1}{4 m}\right) \sigma^ {2}\right), \tag {6} +$$ + +where $\mu$ and $\sigma$ are the mean and variance of the activation $x\in \mathbb{R}$ and $\mu_{4}$ is its 4-th central moment, i.e., $\mu_4 = \mathbb{E}_x(x - \mu)^4$ . The item $\mu_4 / \sigma^4$ in Eqn.(6) is well-known as the kurtosis [6]]. It holds that $\mu_4 / \sigma^4\geq 1$ for any distribution. Although there is no upper limit to kurtosis of a general distribution, for most well-known distributions, e.g., Gaussian distribution, uniform distribution and logistic distribution, it is finite [6]. Therefore, the variance of $\sigma_{\mathcal{B}}$ can always be regarded as $\mathcal{O}(\frac{\sigma^2}{m})$ . + +Notice that the magnitudes of both $\mu_{\mathcal{B}}$ and $\sigma_{\mathcal{B}}$ of different neurons could be quite different. The analysis above indicates that to inject noise into $\mu_{\mathcal{B}}$ and $\sigma_{\mathcal{B}}$ , it is better to multiply them by a Gaussian noise centered at 1 with proper variance instead of add them by two zero-mean noises. One reason is that the standard deviation of $\sigma_{\mathcal{B}}$ is proportional to $\sigma$ when $m$ is sufficiently large. More importantly, it is impossible to choose the right amount of noise for each $\mu_{\mathcal{B}}$ and $\sigma_{\mathcal{B}}$ if we add noises additively. Therefore, we multiplicatively inject noise into its mean and variance estimations of LBN as follows: + +$$ +\mu_ {\mathcal {B}} \leftarrow \mu_ {\mathcal {B}} \cdot n _ {\mu} \text {a n d} \sigma_ {\mathcal {B}} \leftarrow \sigma_ {\mathcal {B}} \cdot n _ {\sigma}, +$$ + +where $n_{\mu}$ and $n_{\sigma}$ are two noises sampled from $\mathcal{N}(1,0.1)$ and they are treated as untrainable parameters in each back propagation. The result is presented in Fig 1. It shows that with such noise, the performance of LBN can be boosted to be comparable with BN. + +The main conclusion we can reach in the experiment above is that data-independent noise can be used to improve the performance of LBN. Moreover, it also indicates that the effect of BN on improving generalization can be decomposed into two parts: one is population normalization (can be seen from the improvement of LBN over w/o bn) and the other is the stochastic uncertainty contained in the + +mini-batch statistics (can be seen from the improvement of LBN+noise over LBN). This phenomenon inspires us to inject stochastic noise into our PN to improve its generalization performance. + +# 4.2. Noisy Population Normalization + +Below, we develop our final method NPN (Alg.2) to achieve better generalization performance. It is inspired from our above understandings of the stochastic uncertainty in BN. In particular, we showed that data independent noise can be injected into LBN to boost the generalization performance. + +At first, for each activation $x$ , we insert data-independent noise into the corresponding $\mu$ and $\gamma$ to improve generalization. In order to reduce the effects of different magnitudes of $\mu$ and $\gamma$ corresponding to different neurons, similar to LBN+noise in Section 4.1, we multiply $\mu$ and $\gamma$ by two noises centered at 1 instead of adding two zero-mean ones as follows: + +$$ +\hat {\mu} = \mu \left(1 + n _ {\mu}\right) \text {a n d} \hat {\gamma} = \max \{0, \gamma \left(1 + n _ {\gamma}\right) \}, +$$ + +where $n_{\mu}$ and $n_{\gamma}$ are independently sampled noises from a zero-mean distribution, e.g., Gaussian, and max is used to avoid negative $\hat{\gamma}$ . We then normalize, scale and shift $x$ by: + +$$ +\hat {x} = \frac {x - \hat {\mu}}{\sqrt {\hat {\gamma} ^ {2} + \epsilon}} \text {a n d} y = \alpha \hat {x} + \beta , \tag {NPN} +$$ + +where $\mu, \gamma, \alpha$ and $\beta$ are trainable parameters. The training process with NPN is almost the same as PN, except that the injected noises $n_{\mu}$ and $n_{\gamma}$ are treated as constants in back propagation. The details of the training process can be found in the appendix. + +Discussion. NPN has the following appealing features. As the injected noises $n_{\mu}$ and $n_{\gamma}$ are data independent, we are able to flexibly inject a proper amount of noise into the learnt estimation of statistics to improve generalization ability, which is particularly important when large batch sizes are used. This scenario is often encountered in cross silo federated learning. The idea of injecting data independent noise to improve generalization can also be combined with other normalization methods. The success of NPN shows that one can directly learn the population statistics in the training process instead of estimating them from the minibatches. This can inspire new normalization methods. + +Remark 4.1 It is unclear that whether NPN has some other nice properties of BN, e.g., auto-tuning of learning rates due to scale invariance [1, 7], while our empirical results demonstrate the effectiveness of NPN. We argue that different from standard training, the performance improvements achieved by these properties could be negligible compared to normalization itself in FL as the training in FL is more challenging than standard training. Such theoretical analysis is left as future work. Thus, we would like to point out that NPN is a normalization technique for FL instead of a general approach. + +
batch sizeNormalizationα (CIFAR-10)α (CIFAR-100)
0.040.080.161.000.040.080.161.00
8w/o69.8383.2285.0587.5945.1350.4354.2658.47
IN67.6878.5582.6284.0935.5140.0245.4845.45
GN71.0582.5483.3586.1145.4549.2252.2054.75
BN70.2282.2785.7789.4252.1355.9158.1664.88
FedBN71.4983.0584.4189.3052.0556.1257.9764.43
PN78.5985.4588.0589.6353.0657.4361.8765.00
4w/o70.6982.0285.1186.9346.9950.8355.0859.45
IN60.5374.0878.1182.3936.0442.9646.1749.89
GN68.6378.5382.0185.5244.2148.1851.8554.67
BN70.1882.4786.3289.5251.3155.7860.3564.32
FedBN72.2483.5786.0389.4651.4555.7759.9165.10
PN78.4085.0587.5289.5952.2856.8562.4265.82
2w/o67.0379.9883.4387.0143.9348.5654.7759.43
IN55.4268.9375.4378.7936.1543.9047.7247.30
GN66.0278.8281.7585.2344.4046.1750.6053.24
BN65.3279.6883.3087.2545.8750.6655.0658.74
FedBN65.5080.0282.1786.4046.0451.3055.5259.68
PN72.5382.5986.2388.9848.9754.9260.6465.46
1w/o60.4971.8378.6584.9338.2442.2649.4954.59
IN35.4754.9366.4573.7433.6039.0643.2948.82
GN60.5571.8375.2579.6846.0449.2653.9756.64
BNNANANANANANANANA
FedBNNANANANANANANANA
PN64.3878.8982.1386.6946.1349.3259.1962.97
+ +Table 1. The results of different normalization methods on CIFAR-10/100 with different batch sizes and Non-IID degree $\alpha$ . Or PN consistently outperforms other methods by a large margin across different $\alpha$ and batch sizes. + +# Algorithm 2 Noisy Population Normalization + +Input: activation: $x$ ; population mean and variance: $\mu$ and $1 / \gamma^2$ ; scale and shift parameter: $\alpha$ and $\beta$ ; noise level: $\sigma$ . + +Inject noise into $\mu$ and $\gamma$ : + +$$ +\hat {\mu} = \mu \left(1 + n _ {\mu}\right) \text {a n d} \hat {\gamma} = \max \{0, \gamma \left(1 + n _ {\gamma}\right) \}, +$$ + +where $n_{\mu}$ and $n_{\gamma}$ are noises from $\mathcal{N}(0,\sigma^2)$ . Normalize using noisy population statistics: + +$$ +\hat {x} = \frac {x - \hat {\mu}}{\sqrt {\hat {\gamma} ^ {2} + \epsilon}} +$$ + +Scale and shift: + +$$ +y = \alpha \hat {x} + \beta \equiv \mathrm {N P N} _ {\alpha , \beta} ^ {\mu , \gamma} (\sigma ; x) +$$ + +Return: y + +# 5. Experiments + +In this section, we conduct two sets of experiments. Firstly, we verify the effectiveness of of PN in the federated learning setting given small batch sizes and heterogeneous client data. Secondly, we demonstrate the superior generalization ability of our NPN given large training batch sizes. + +# 5.1. Settings and Baselines + +To verify the advantage of our population normalization method, we compare it with other normalization counterparts in the FL setting. Specifically, we compare with batch normalization [16], group normalization [40], instance normalization [37] and without normalization. We compare with FedBN [25] by averaging the BN parameters in the clients. Extensive experiments are conducted on two image classification benchmarks (i.e., CIFAR-10 and CIFAR-100) with models VGG-16 [33] and ResNet [11] to evaluate the effectiveness of our method. + +For FL experiments, we adopt FedAvg [30] as the base algorithm. We use 80 clients in total with participation rate set to 0.4 throughout all experiments. For the clients' local update, we adopt SGD with momentum. For batch size of 8, the learning rates are set to be the best among $\{0.005, 0.01, 0.15, 0.02, 0.25\}$ for each normalization method, and linearly down-scaled for other smaller batch sizes. We run 200 global communication rounds with local epoch set to 1. We examine the superiority of our method against the baseline under various batch sizes and heterogeneity degrees. We follow prior work [5, 28] to use Dirichlet distribution for simulating the non-IID data distribution, where the degree of heterogeneity is defined by $\alpha$ , smaller $\alpha$ value corresponds to more severe heterogeneity. We report the global model's average performance in the last five rounds evaluated using the test split of the datasets. For other details, please refer to the Appendix. + +![](images/fccbce6add3e42effc06f4e77385fb8472154b3a7fabd6ca2bd7df33d985c30b.jpg) +(a) + +![](images/9e0cb30003cdbb0168fb67d854e4880702d5de8b46105a1e083faad685bccd29.jpg) +Figure 2. (a) Training (left) and test (right) loss of BN, PN, and NPN as the training processes; (b) Sensitivity of NPN on noise level. + +![](images/e6eca362c319f2b1bf1acbcd8bde40903267ef46310e9ef53fc0bc7b79b29d37.jpg) +(b) + +For experiments with large batch sizes, we adopt the distributed training setting, where the model is being trained 4 different participants. The batch size is set to 2048. Other hyper-parameters are set to be the same as in the setting. We conduct the experiments using VGG-16 train on CIFAR-10 and CIFAR-100 datasets. + +# 5.2. Federated Learning with Small Batch Size + +We point out that with small batch sizes, SGD already introduces enough stochastic uncertainty in gradient estimation, which helps generalization. Therefore, we do not need to inject additional noise for PN and thus the experiments with NPN are omitted in this section. + +Comparison with Various Normalization Methods We compare PN with other counterparts under different batch sizes and non-iid degrees in Table 1. Specifically, we conduct experiments with $\alpha \in \{0.04, 0.08, 0.16, 1.00\}$ and batch sizes $\{1, 2, 4, 8\}$ . We adopt VGG16 as the trained network. As demonstrated in the table, we observe that our method consistently outperforms BN and other baselines, while the advantage is more significant when the batch size is small and the heterogeneity degree is higher. Notably, our method works well even with batch size set to 1, while the vanilla BN becomes inapplicable under this scenario. This enables our method to be used when the computational resource is limited on the client side. + +Experiments with Various Networks We confirm the effectiveness of PN across different network architectures and capacities. As shown in Table 3, we conduct experiments with VGG16, ResNet18 and ResNet50. We fix the heterogeneity degree $\alpha$ to be 0.08 and trial with batch sizes of $\{1,2,4,8\}$ . We observe consistent performance advantage over the baselines for different networks. Interestingly, one may notice that networks with higher capacities do not necessarily lead to better performances, which is opposite to standard training. This phenomenon is also reported in [5], which may be due to clients' local over-fitting or further + +drifting from each other in the parameter space. However, PN still demonstrates superior performances compared with other normalization techniques. + +# 5.3. Federated Learning with Large Batch Size + +We now extend our PN to NPN to enable stronger generalization ability for training with large batch sizes. The results are demonstrated in Table 4. As can be seen, when training with large batch sizes, PN is outperformed by BN due to the lack of stochastic uncertainty in the normalization layer. On the other hand, NPN remedies this issue by injecting artificial noise into the statistics estimation, which in turn leads to better test performance. The training progress is also visualized in Figure 2(a). We can see that although the training loss of NPN is larger than that of BN, its testing loss is significantly smaller than BN. This demonstrates that benefiting from the injected noise, NPN can prevent the training process from over-fitting more effectively than BN. + +
MethodLossTop-1Top-5Top-10
BRN0.4493.2197.5298.83
BRN + noise0.2993.6097.8399.02
+ +Table 2. Batch Renormalization + noise. + +Finally, we point out that there are only a small number of participants in this scenario, which makes the trained model less prone to the adverse impact of heterogeneity and thus shows better performance than the previous experiments with more participating clients (Table 1). + +# 5.4. Abative Experiments + +Noise Level Sensitivity. We evaluate the sensitivity of NPN to the injected noise level. We conduct experiments with VGG-16 on CIFAR-10. We set the batch size to 2048 and vary the variance of the noise injected into NPN in $\{0.5, 0.1, 0.01, 0.001\}$ . The results are presented in Figure 2(b). It shows that the test loss increases gradually and slowly when we vary the variance of the noise from 0.5 to + +
Normalizationbatch size (VGG-16)batch size (ResNet-18)batch size (ResNet-50)
124812481248
w/o71.8379.9882.0283.2275.4879.9480.2181.4965.5972.3377.8079.94
IN35.4768.9374.0378.5563.5369.3670.6871.3560.3764.6568.2869.10
GN60.5578.8278.5382.5466.0370.4472.5975.1360.6368.7270.7371.25
BNNA78.6582.0982.27NA78.5579.6481.94NA72.3978.2081.83
PN78.8982.2285.1385.4580.2083.9784.5384.6767.7273.6379.5983.54
+ +Table 3. Results of different normalization methods on CIFAR-10. PN consistently outperforms other methods for different nwork architectures and batch sizes. + +
NormalizationCIFAR-10CIFAR-100
LossTop-1Top-2Top-3LossTop-1Top-2Top-3
w/o0.5492.3196.9198.202.7770.3487.9692.11
IN0.4692.4597.0098.312.5470.6888.4992.33
GN0.4392.9297.1498.522.2371.3588.9393.10
BN0.3993.1397.6698.872.0572.2190.2893.60
PN0.4193.1097.7898.971.8572.1590.0793.89
NPN0.3293.6698.0599.121.2272.7491.4794.90
+ +Table 4. Results of different normalization methods on CIFAR-10 / CIFAR-100 with large batch size. + +0.001. Therefore, it is easy to select a suitable noise variance that falls within a reasonable range. + +Artificial Noise for Other Normalization Methods. We demonstrate that the artificially injected noise is also able to improve existing normalization methods with limited stochastic uncertainty. We take batch renormalization (BRN) [16] as an example and conduct experiments on CIFAR-10 with VGG-16. For simplicity and excluding other factors, we conduct this experiment with regular training instead of FL. We inject noise sampled from $\mathcal{N}(0,0.1)$ to the moving averages $\mu$ and $\sigma$ of BRN as we do for NPN. The results are given in Table 2. It shows that with the injected noise, the performance of BRN can be improved effectively. Therefore, the idea of data-independent noise injection is a very general regularization technique that can be used to improve existing normalization methods. + +# 5.5. Supervised Learning with PN + +Intuitively, PN can also be used in non-FL scenarios. We apply PN to supervised learning and conduct experiments on two image classification benchmarks (i.e., CIFAR-10, CIFAR-100) with VGG-16 and ResNet to evaluate the effectiveness of PN in supervised learning. The batch sizes of all the methods are set to 128. Table 5 presents the test loss and accuracy of the final learned models. From Table 5 we can see that our NPN can achieve much lower test loss than BN and comparable (if no better than) test accuracy with BN on all the three models. We can also observe that NPN can achieve a noticeable improvement over PN, which demonstrates that artificially injected data independent noise can be used to reinforce the performance of the normalization methods with limited stochastic under + +Table 5. Results of different normalization methods on image classification benchmarks + +
NetworkNormalizationLossAccuracy
VGG-16BN0.3992.31
PN0.4193.63
NPN0.2893.66
ResNet-18BN0.4693.76
PN0.4593.97
NPN0.3894.26
ResNet-56BN0.3393.05
PN0.4292.56
NPN0.2593.99
+ +tainty. The comparable performance of BN and NPN implies that the effect of BN can be decomposed into two parts: one is population normalization, and the other is the stochastic uncertainty containing in the mini-batch statistics. + +# 6. Conclusions + +We propose a learning-based normalization technique, Population Normalization (PN), which reduces the reliance on training batch size and improves robustness against data heterogeneity in federated learning (FL). Extensive experiments demonstrate that PN outperforms existing normalization techniques by a significant margin. We further develop Noisy Population Normalization (NPN), which injects artificial noise into the learnt normalization statistics to enhance generalization under large batch size conditions. We believe that our work will offer valuable insights for the design and development of more stable and adaptable normalization techniques for FL. + +# 7. Acknowledgments + +This work was supported by the National Nature Science Foundation of China (62472097) and AI for Science Foundation of Fudan University (FudanX24AI028). The computations in this research were performed on the CFFF platform of Fudan University. + +# References + +[1] Sanjeev Arora, Kaifeng Lyu, and Zhiyuan Li. Theoretical analysis of auto rate-tuning by batch normalization. In 7th International Conference on Learning Representations, ICLR 2019, 2019. 5 +[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 1, 2 +[3] Tobias Bernecker, Annette Peters, Christopher L. Schlett, Fabian Bamberg, Fabian J. Theis, Daniel Rueckert, Jakob Weiss, and Shadi Albarqouni. Fednorm: Modality-based normalization in federated learning for multi-modal liver segmentation. ArXiv, abs/2205.11096, 2022. 3 +[4] Yongqiang Cai, Qianxiao Li, and Zuowei Shen. A quantitative analysis of the effect of batch normalization on gradient descent. In International Conference on Machine Learning, pages 882-890, 2019. 3 +[5] Hong-You Chen and Wei-Lun Chao. Fedbe: Making bayesian model ensemble applicable to federated learning. In ICLR, 2021. 3, 6, 7 +[6] Wikipedia contributors. Kurtosis. Online; accessed 2-June-2020, 2020. Wikipedia, The Free Encyclopedia. 5 +[7] Hadi Daneshmand, Amir Joudaki, and Francis Bach. Batch normalization orthogonalizes representations in deep random networks. In Advances in Neural Information Processing Systems, 2021. 5 +[8] Jian-Hui Duan, Wenzhong Li, and Sanglu Lu. Feddna: Federated learning with decoupled normalization-layer aggregation for non-iid data. In Machine Learning and Knowledge Discovery in Databases. Research Track, pages 722-737, Cham, 2021. Springer International Publishing. 3 +[9] Aiden Durrant, Milan Markovic, David Matthews, David May, Jessica Enright, and Georgios Leontidis. The role of cross-silo federated learning in facilitating data sharing in the agri-food sector. Computers and Electronics in Agriculture, 193:106648, 2022. 2, 4 +[10] Filip Hanzely, Slavomír Hanzely, Samuel Horváth, and Peter Richtárik. Lower bounds and optimal algorithms for personalized federated learning. Advances in Neural Information Processing Systems, 33:2304-2315, 2020. 3 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 6 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1 + +[13] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip Gibbons. The non-IID data quagmire of decentralized machine learning. In Proceedings of the 37th International Conference on Machine Learning, pages 4387-4398. PMLR, 2020. 3 +[14] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip B. Gibbons. The non-iid data quagmire of decentralized machine learning, 2020. 1 +[15] Chao Huang, Jianwei Huang, and Xin Liu. Cross-silo federated learning: Challenges and opportunities. arXiv preprint arXiv:2206.12949, 2022. 2, 4 +[16] Sergey Ioffe. Batch renormalization: Towards reducing minibatch dependence in batch-normalized models. In Advances in neural information processing systems, pages 1945-1953, 2017. 1, 2, 6, 8 +[17] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning, pages 448-456, Lille, France, 2015. PMLR. 1, 2 +[18] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for ondevice federated learning. In ICML, 2020. 3 +[19] Jakub Konečny, H Brendan McMahan, Felix X Yu, Peter Richtárik, Ananda Theertha Suresh, and Dave Bacon. Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492, 2016. 1, 3 +[20] Viraj Kulkarni, Milind Kulkarni, and Aniruddha Pant. Survey of personalization techniques for federated learning. In 2020 Fourth World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4), pages 794-797. IEEE, 2020. 3 +[21] Qinbin Li, Bingsheng He, and Dawn Song. Practical oneshot federated learning for cross-silo setting. arXiv preprint arXiv:2010.01017, 2020. 2, 4 +[22] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10713–10722, 2021. 3 +[23] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In MLSys, 2020. 3 +[24] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through personalization. In International Conference on Machine Learning, pages 6357-6368. PMLR, 2021. 3 +[25] Xiaoxiao Li, Meirui Jiang, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fed{bn}: Federated learning on non- $\{\mathrm{i}\mathrm{d}\}$ features via local batch normalization. In International Conference on Learning Representations, 2021. 2, 3, 6 +[26] Xiangru Lian and Ji Liu. Revisit batch normalization: New understanding and refinement via composition optimization. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 3254-3263, 2019. 3 +[27] Wei Yang Bryan Lim, Nguyen Cong Luong, Dinh Thai Hoang, Yutao Jiao, Ying-Chang Liang, Qiang Yang, Dusit + +Niyato, and Chunyan Miao. Federated learning in mobile edge networks: A comprehensive survey. IEEE Communications Surveys & Tutorials, 22(3):2031-2063, 2020. 3 +[28] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. Advances in Neural Information Processing Systems, 33:2351-2363, 2020. 3, 6 +[29] Ping Luo, Xinjiang Wang, Wenqi Shao, and Zhanglin Peng. Towards understanding regularization in batch normalization. arXiv preprint arXiv:1809.00846, 2018. 3, 5 +[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 3, 6 +[31] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. In Advances in neural information processing systems, pages 901-909, 2016. 2 +[32] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? In Advances in Neural Information Processing Systems, pages 2483-2493, 2018. 3 +[33] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 6 +[34] Sebastian U Stich. Local sgd converges fast and communicates little. In International Conference on Learning Representations, 2018. 3 +[35] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 3 +[36] Mattias Teye, Hossein Azizpour, and Kevin Smith. Bayesian uncertainty estimation for batch normalized deep networks. In International Conference on Machine Learning, pages 4907-4916, 2018. 5 +[37] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 1, 2, 6 +[38] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 3 +[39] Xiaoxia Wu, Edgar Dobriban, Tongzheng Ren, Shanshan Wu, Zhiyuan Li, Suriya Gunasekar, Rachel Ward, and Qiang Liu. Implicit regularization of normalization methods. arXiv preprint arXiv:1911.07956, 2019. 3 +[40] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-19, 2018. 1, 2, 3, 6 +[41] Jianhang Xiao, Chunhui Du, Zijing Duan, and Wei Guo. A novel server-side aggregation strategy for federated learning in non-id situations. In 2021 20th International Symposium on Parallel and Distributed Computing (ISPDC), pages 17-24. IEEE, 2021. 3 + +[42] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1, 3 +[43] Yousef Yeganeh, Azade Farshad, Nassir Navab, and Shadi Albarqouni. Inverse distance aggregation for federated learning with non-iid data. In Domain Adaptation and Representation Transfer, and Distributed and Collaborative Learning, pages 150-159. Springer, 2020. 3 \ No newline at end of file diff --git a/2025/Population Normalization for Federated Learning/images.zip b/2025/Population Normalization for Federated Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e2f18c354a0350ca21eb88141f0db561745dc8ae --- /dev/null +++ b/2025/Population Normalization for Federated Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54cb2b1edc33c5f442a3d5f1e220860b84b0a7f13a9e7a4884e1eb2b00ad42f8 +size 424530 diff --git a/2025/Population Normalization for Federated Learning/layout.json b/2025/Population Normalization for Federated Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..57864ad841af9e26f9e131c7d41cd2972037ad6d --- /dev/null +++ b/2025/Population Normalization for Federated Learning/layout.json @@ -0,0 +1,10225 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 150, + 103, + 460, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 103, + 460, + 121 + ], + "spans": [ + { + "bbox": [ + 150, + 103, + 460, + 121 + ], + "type": "text", + "content": "Population Normalization for Federated Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "spans": [ + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "text", + "content": "Zhuoyao Wang" + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "text", + "content": ", Fan Yi" + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "text", + "content": ", Peizhu Gong" + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "text", + "content": ", Caitou He" + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "text", + "content": ", Cheng Jin" + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "text", + "content": ", Weizhong Zhang" + }, + { + "bbox": [ + 89, + 141, + 534, + 158 + ], + "type": "inline_equation", + "content": "^{1,4\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 156, + 410, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 156, + 410, + 171 + ], + "spans": [ + { + "bbox": [ + 212, + 156, + 410, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 212, + 156, + 410, + 171 + ], + "type": "text", + "content": "Fudan University " + }, + { + "bbox": [ + 212, + 156, + 410, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 212, + 156, + 410, + 171 + ], + "type": "text", + "content": " Wenzhou University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 171, + 509, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 171, + 509, + 185 + ], + "spans": [ + { + "bbox": [ + 112, + 171, + 509, + 185 + ], + "type": "text", + "content": "3Innovation Center of Calligraphy and Painting Creation Technology, MCT, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 156, + 185, + 466, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 185, + 466, + 200 + ], + "spans": [ + { + "bbox": [ + 156, + 185, + 466, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 156, + 185, + 466, + 200 + ], + "type": "text", + "content": "Shanghai Key Laboratory of Intelligent Information Processing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 201, + 559, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 201, + 559, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 201, + 559, + 213 + ], + "type": "text", + "content": "{18110240064,pzgong,jc,weizhongzhang}@fudan.edu.cn，ftyi21@m.fudan.edu.cn，hctou89@wzu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 266, + 297, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 297, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 297, + 589 + ], + "type": "text", + "content": "Batch normalization (BN) is widely recognized as an essential method in training deep neural networks, facilitating convergence and enhancing model stability. However, in Federated Learning (FL) contexts, where training data are typically heterogeneous and clients often face resource constraints, the effectiveness of BN is considerably limited for two primary reasons. First, the population statistics, specifically the mean and variance of these heterogeneous datasets, vary substantially, resulting in inconsistent BN layers across client models, which ultimately drives these models to diverge further. Second, estimating statistics from a mini-batch is often imprecise since the batch size has to be small in resource-limited clients. This paper introduces Population Normalization, a novel technique for FL, in which the statistics are learned as trainable parameters rather than calculated from mini-batches as in BN. Thus, our normalization layers are homogeneous among the clients and the adverse impact of small batch size is eliminated as the model can be well-trained even when the batch size equals to one. To enhance the flexibility of our method in practical applications, we investigate the role of stochastic uncertainty in BN's statistical estimation. When larger batch sizes are available, we demonstrate that injecting simple artificial noise can effectively mimic this stochastic uncertainty and improve the model's generalization capability. Experimental results validate the efficacy of our approach across various FL tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 610, + 135, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 135, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 135, + 623 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 629, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 696 + ], + "type": "text", + "content": "Batch Normalization (BN) [17] has become a standard component in deep neural networks (DNNs). It is proposed to stabilize and accelerate the training of DNNs, and the original motivation was to reduce the so-called internal covariate shift. In each training step, it normalizes the layer input of a neural network using the mean and variance com" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "text", + "content": "puted within a randomly sampled mini-batch. Specifically, consider a particular neuron in a network, and we denote " + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\{x_{1},\\ldots ,x_{m}\\}" + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "text", + "content": " to be its pre-activations on a mini-batch " + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "text", + "content": ", BN first normalizes each " + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 403, + 293, + 555, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 293, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 403, + 293, + 555, + 312 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {i} = \\frac {x _ {i} - \\mu_ {\\mathcal {B}}}{\\sigma_ {\\mathcal {B}}}, \\tag {1}", + "image_path": "bd6d6494a9fb9e94a6254400e8a791173ba9406ecf41adf55cfef2dd71a4558a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{B}}^{2}" + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "content": " are the sample mean and variance of the pre-activation on " + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "content": ". It then scales and shifts each " + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\hat{x}_i" + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "inline_equation", + "content": "y_i = \\alpha \\hat{x}_i + \\beta" + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "content": " to compensate the effect of normalization, where " + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "content": " are two parameters to be learned during the training process. BN finally takes " + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 313, + 318, + 555, + 426 + ], + "type": "text", + "content": " as the new pre-activation of this neuron. It has been reported in literature [12, 16, 17, 37, 40] that BN can significantly improve the convergence of training and the generalization performance of the final learned models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 426, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 605 + ], + "type": "text", + "content": "However, we notice that the effectiveness of BN is compromised in Federated Learning [14, 19, 42], which typically aims at collaboratively training models on resource limited clients (e.g., mobiles) with locally private and heterogeneous training data. The main reasons can be summarized as follows. One is that the statistics, i.e., the mean and variance, of heterogeneous datasets in the clients differ greatly. Eqn. (1) implies that this can lead to inconsistent BN layers among the client models and finally causes these models to drift further away from each other. The other is that in resource limited clients, the mini-batch size has to be small, which makes " + }, + { + "bbox": [ + 313, + 426, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 426, + 556, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{B}}^{2}" + }, + { + "bbox": [ + 313, + 426, + 556, + 605 + ], + "type": "text", + "content": " inaccurate estimations of the population mean and variance for normalization and the obtained BN would cause instability during training that slows down convergence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": "Some other normalization techniques, such as batch renormalization [16], group normalization [40], and layer normalization [2], are proposed to address this heavy reliance on batch size. The key ideas are to estimate the population mean and variance as accurately as possible by using techniques such as moving average [16] or to normalize the activation along other dimensions such as the channel dimension [2]. Nevertheless, these approaches focus on training acceleration, which typically under-perform" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "type": "text", + "content": "*Equal Contribution, † Corresponding authors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10214" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "BN in terms of the trained model's accuracy [40]. Recently, FedBN [25] is proposed to extend BN to FL. It keeps the client BN layers updated locally, without communicating and aggregating them during training. This enables the clients to achieve better performances in their own respective environments. However, since the BN layers among the clients can be heterogeneous due to the non-IID datasets, FedBN cannot produce an off-the-shelf global model after training, which is often required in reality. Consequently, FedBN suffers from the cold start issue, i.e., the learned model cannot be directly deployed in the new clients due to the lack of trained local BN layers. What's more, FedBN still relies on estimation within mini-batches, which cannot address the issue of inaccurate estimation caused by small batch size in cross-device FL." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 255, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 296, + 506 + ], + "type": "text", + "content": "In this paper, we propose Population Normalization (PN), a simple yet effective normalization method for FL to eliminate the negative impacts of statistical heterogeneity and estimation inaccuracies on BN. The key idea is to learn population statistics as trainable parameters in the training process by introducing two constraints derived from the definitions of mean and variance (see Section 3), instead of estimating them from the mini-batches as most existing methods have done. As the population statistics take place of the sample mean and variance of the local datasets in BN, and are treated as normal trainable parameters in local training and aggregation, our normalization layers among the clients would be consistent although the datasets are heterogeneous. This enables our method to eventually produce a global model with the normalization layer among the clients. In addition, since we no longer need to calculate the sample mean and variance, the resulting training problem can be solved by stochastic algorithms of FL with arbitrary batch sizes (even equals to one) and thus the issue of inaccurate statistic estimation because of small batch size is eliminated." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "In practice, we observe that PN tends to under-perform BN given very large batch sizes. We then further explore this issue and find its cause to be that PN lacks the stochastic uncertainty in the estimation of mean and variance. Specifically, we show that the noise introduced by the mini-batch based mean and variance estimations in the BN is essential for its good generalization behavior. Our empirical results demonstrate that injecting a reasonable amount of data independent noise into BN with extremely large batch sizes can significantly improve the generalization performance to be comparable to that of BN with proper batch sizes. Inspired by this finding, we propose Noisy Population Normalization (NPN) by injecting simple data independent artificial noises into the learned population mean and variance in PN during training. The injected noise can serve the same purpose as the uncertainty in estimating population statistics in BN with small batch sizes. This makes the proposed" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "method more flexible in the applications of FL where large batch size is available, e.g., collaboration among companies in cross-silo FL [9, 15, 21]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 109, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 216 + ], + "type": "text", + "content": "Finally, we conduct a series of experiments to verify the effectiveness of our proposed PN and NPN in the context of FL with small and large batch sizes, respectively (Section 5). The results demonstrate that PN effectively reduces the reliance on training batch size and alleviates the negative impact of data heterogeneity on normalization. Furthermore, the artificial noise injected in our NPN method effectively improves the generalization performance given large training batch sizes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 324, + 217, + 534, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 217, + 534, + 228 + ], + "spans": [ + { + "bbox": [ + 324, + 217, + 534, + 228 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 230, + 553, + 324 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 230, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 230, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 230, + 553, + 266 + ], + "type": "text", + "content": "- We propose a normalization technique PN for federated learning, which demonstrates superior performance given small batch sizes and data heterogeneity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 266, + 553, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 301 + ], + "type": "text", + "content": "- In training with large batch sizes, we propose NPN based on PN, which introduces stochastic uncertainty by injecting artificial noise to improve generalization ability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 301, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 553, + 324 + ], + "type": "text", + "content": "- We conduct extensive experiments to verify the effectiveness of our method." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": "Notations: Given a positive integer " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "[d]" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": " to be the set " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\{1,2,\\dots,d\\}" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": ". For a vector " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "x\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": ", we let " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "[x]_i" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": " be its " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": "-th component with " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "i\\in [d]" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": ". We denote by " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "w_{i,j}" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": " the element in the " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": "-th row and the " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": "-th column of a real-valued matrix " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "A\\cup B" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": " be the union of two sets " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": ". Moreover, we denote by " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu ,\\sigma^2)" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": " the Gaussian distribution with mean of " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": " and variance of " + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 313, + 326, + 554, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 423, + 400, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 423, + 400, + 435 + ], + "spans": [ + { + "bbox": [ + 314, + 423, + 400, + 435 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 443, + 460, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 460, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 460, + 456 + ], + "type": "text", + "content": "2.1. Normalization Techniques" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 461, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 544 + ], + "type": "text", + "content": "BN [17] is a standard normalization technique, which enables the training of very deep neural networks, and the most representative result is ResNet[11], which cannot be easily trained to converge without BN. It can also achieve significant imporement in generalization. Therefore, the invention of BN is a major milestone in the development of deep learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "As aforementioned, small batch size or non-iid minibatch can diminish the effect of BN on the training acceleration or even result in training failure, due to the inaccurate mean and variance estimations [16]. Moreover, large batch size leads to huge memory cost. This issue hinders BN's usage in FL applications, where the clients often have limited computational resources and the datasets are usually heterogeneous. This in turn prevents the training of large networks in FL. Recently, some new normalization methods have been proposed as alternatives to BN, and they can be roughly divided into two categories, i.e., normalization along the batch dimension [16], and mini-batch independent normalization [2, 31, 37, 40]. The former alleviates the reliance on the batch size by using techniques such as" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10215" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 178 + ], + "type": "text", + "content": "moving average. The latter ones achieve this by normalizing the activation of neurons in a network along other dimensions. For example, Group Normalization [40] divides the channels in each hidden layer into groups and normalizes the activations in each group with its mean and variance. Empirical studies show that all these methods achieve comparable training speedups with BN. However, the generalization performance they achieved is inferior to that of BN." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 324 + ], + "type": "text", + "content": "Besides the above studies, there is another line of research [4, 26, 29, 32, 39] that focuses on understanding BN. One of the main conclusions in these investigations is that with BN, the effective objective function used in training is different from the original one. Specifically, [32] showed that by inserting BN layers into a network, the loss landscape becomes smoother, which naturally explains why BN enables one to use large learning rates. The studies [29, 39] focus on the regularization effect of BN. They showed that the objective function with BN can be divided into two parts: an empirical loss under population normalization and an explicit regularizer named gamma decay." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 331, + 172, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 172, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 172, + 344 + ], + "type": "text", + "content": "2.2. Federated Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 347, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 381 + ], + "type": "text", + "content": "FL emerges as an effective paradigm to train models on distributed devices without centralizing local private data [3, 8, 13, 19, 42]. Specifically, it can be formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 388, + 295, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 388, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 295, + 418 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathcal {L} (\\theta , \\mathcal {D}) \\triangleq \\sum_ {m = 1} ^ {M} \\rho_ {m} \\mathcal {L} _ {m} (\\theta , \\mathcal {D} _ {m}), \\tag {2}", + "image_path": "217c18686d40699016cd5d1e90fa3302d291c8abb5658142b5176618974b8527.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "text", + "content": " is the number of clients and " + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "text", + "content": " is the model to be learnt. The global dataset " + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\cup_{m=1}^{M}\\mathcal{D}_{m}" + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{m} = \\{(\\pmb{x}_{i},y_{i})\\}_{i=1}^{|D_{m}|}" + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "text", + "content": " is the locally kept dataset on client " + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{m}(\\theta,\\mathcal{D}_{m}) = \\frac{1}{|\\mathcal{D}_{m}|}\\sum_{\\xi\\in\\mathcal{D}_{m}}\\ell(\\theta,\\xi)" + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "text", + "content": " denotes the local empirical risk with the loss function " + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\ell(\\cdot,\\cdot)" + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\rho_{m} = \\frac{|D_{m}|}{|\\mathcal{D}|}" + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "text", + "content": ". Typically, the most popular FL algorithm FedAvg [30] solves the problem (2) based on local-SGD [34]. Specifically, in the " + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "inline_equation", + "content": "c^{th}" + }, + { + "bbox": [ + 55, + 425, + 296, + 539 + ], + "type": "text", + "content": " round, it updates the global model on the server as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 545, + 295, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 545, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 129, + 545, + 295, + 574 + ], + "type": "interline_equation", + "content": "\\theta^ {c + 1} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\rho_ {m} \\theta_ {m} ^ {c}, \\tag {3}", + "image_path": "95efc57974a19ec65f36d83ba38e9b7069d024c8a53bc3bdbb55d95cf61398fc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\theta_{m}^{c}" + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": " denotes the received local model from client " + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": " after local training in this round. The updated global model " + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\theta^{c + 1}" + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": " is then distributed to the clients to initiate the " + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "inline_equation", + "content": "(c + 1)" + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": "-th round of training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "In FL, especially in mobile edge scenarios, both limited device capacity and non-IID distribution of clients' local data raise challenges to the training procedure [27]. Various works attempt to tackle the non-IID distribution of data for federated learning from the perspective of local training regularization [18, 22, 23], server aggregation [5, 28, 38, 41, 43], personalized federated learning [10, 20, 24, 35]," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 72, + 476, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 476, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 476, + 84 + ], + "type": "text", + "content": "Algorithm 1 Population Normalization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "spans": [ + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "text", + "content": "Input: activation: " + }, + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "text", + "content": "; mean and variance: " + }, + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "inline_equation", + "content": "1 / \\gamma^2" + }, + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "text", + "content": "; scale and shift parameters: " + }, + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 323, + 87, + 555, + 112 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 113, + 493, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 113, + 493, + 124 + ], + "spans": [ + { + "bbox": [ + 324, + 113, + 493, + 124 + ], + "type": "text", + "content": "Normalize using the population statistics:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 410, + 131, + 468, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 131, + 468, + 155 + ], + "spans": [ + { + "bbox": [ + 410, + 131, + 468, + 155 + ], + "type": "interline_equation", + "content": "\\hat {x} = \\frac {x - \\mu}{\\sqrt {\\gamma^ {2} + \\epsilon}}", + "image_path": "d5c7193c1bfbe052477155b5811e992f0c666eb478f47d7cae59b3b37f5e4d1c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 166, + 389, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 166, + 389, + 176 + ], + "spans": [ + { + "bbox": [ + 324, + 166, + 389, + 176 + ], + "type": "text", + "content": "Scale and shift:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 391, + 185, + 487, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 185, + 487, + 199 + ], + "spans": [ + { + "bbox": [ + 391, + 185, + 487, + 199 + ], + "type": "interline_equation", + "content": "y = \\alpha \\hat {x} + \\beta \\equiv \\mathrm {P N} _ {\\alpha , \\beta} ^ {\\mu , \\gamma} (x)", + "image_path": "22b35a238a47cd8fcd2c1a7736ae24debc262cf9332d5735c2cf22b26bd98da8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 217, + 370, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 217, + 370, + 230 + ], + "spans": [ + { + "bbox": [ + 324, + 217, + 370, + 230 + ], + "type": "text", + "content": "Return: y" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 257, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 365 + ], + "type": "text", + "content": "etc. However, they do not explore the reduced effectiveness of batch normalization in federated learning. FedBN [25] uses local BN parameters to tackle non-iid federated learning. However, FedBN is only applicable to local inferences where only clients participating in the training with local statistics can perform inference, without providing a global model for new clients or global data. Also, FedBN does not consider the small batches caused by limited local resources." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 380, + 462, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 462, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 462, + 395 + ], + "type": "text", + "content": "3. Population Normalization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 401, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 468 + ], + "type": "text", + "content": "We now present our method PN (see Alg.1). Its key idea is to eliminate the impacts of heterogeneous statistics and inaccurate statistic estimations by learning the population mean and variance during the training process, instead of estimating them within the mini-batch as BN has done. Precisely, given an activation " + }, + { + "bbox": [ + 313, + 401, + 555, + 468 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 401, + 555, + 468 + ], + "type": "text", + "content": ", we normalize it as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 403, + 480, + 464, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 480, + 464, + 505 + ], + "spans": [ + { + "bbox": [ + 403, + 480, + 464, + 505 + ], + "type": "interline_equation", + "content": "\\hat {x} = \\frac {x - \\mu}{\\sqrt {\\gamma^ {2} + \\epsilon}},", + "image_path": "7c0de71172245463591006be459a701ab3bc1f393f1653f64c153ef247660b89.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\gamma^2" + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "content": " are the population mean and variance, respectively, and they will be learned during the training process, of which the details are specified later in Eqn.(3). " + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\epsilon > 0" + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "content": " is a small constant. Then, to compensate the effect of normalization on the effective capacity of a neural network, we scale and shift the normalized activation " + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 380, + 601, + 555, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 601, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 380, + 601, + 555, + 617 + ], + "type": "interline_equation", + "content": "y = \\alpha \\hat {x} + \\beta \\equiv \\mathrm {P N} _ {\\alpha , \\beta} ^ {\\mu , \\gamma} (x), \\quad (\\mathrm {P N})", + "image_path": "f2df8bac2d5f832996b1d6bf621d04231036f6b7d8270e4edec42a4c173735f6.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 627, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 554, + 651 + ], + "type": "text", + "content": "where the coefficients " + }, + { + "bbox": [ + 313, + 627, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 627, + 554, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 627, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 627, + 554, + 651 + ], + "type": "text", + "content": " are learned in the training process together with " + }, + { + "bbox": [ + 313, + 627, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 627, + 554, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 627, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 627, + 554, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "text", + "content": "Below, we take a fully connected neural network " + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}) \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "text", + "content": " as an example to show how to insert PN layers into DNNs. Specifically, given an input " + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "text", + "content": ", let the input layer be " + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_j^{(0)}(\\pmb{x}) = [\\pmb{x}]_j" + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "inline_equation", + "content": "j \\in [n^{(0)}] = [d]" + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "text", + "content": "-th hidden layer containing " + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "inline_equation", + "content": "n^{(\\ell)}" + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "text", + "content": " hidden nodes with " + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\ell \\in [L]" + }, + { + "bbox": [ + 313, + 652, + 555, + 714 + ], + "type": "text", + "content": " takes" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10216" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 105, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 105, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 105, + 83 + ], + "type": "text", + "content": "the form of" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 296, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 296, + 169 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f _ {j} ^ {(\\ell)} (\\boldsymbol {x}) = h ^ {(\\ell)} \\left(z _ {j} ^ {(\\ell)} (\\boldsymbol {x})\\right) \\text {w i t h} \\tag {4} \\\\ z _ {j} ^ {(\\ell)} (\\boldsymbol {x}) = \\sum_ {k = 1} ^ {n ^ {(\\ell - 1)}} w _ {j, k} ^ {(\\ell)} \\underbrace {\\left(\\alpha_ {k} ^ {(\\ell - 1)} \\frac {f _ {k} ^ {(\\ell - 1)} (\\boldsymbol {x}) - \\mu_ {k} ^ {(\\ell - 1)}}{\\sqrt {\\left(\\gamma_ {k} ^ {(\\ell - 1)}\\right) ^ {2} + \\epsilon}} + \\beta_ {k} ^ {(\\ell - 1)}\\right)} _ {(\\ell - 1) - \\text {t h P N l a y e r}}, \\\\ \\end{array}", + "image_path": "d5edc973e9c17b3f9c45a9b68d1beaff61f8bf1ed17d6f665ede142894971362.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "inline_equation", + "content": "j\\in [n^{(\\ell)}],z_j^{(\\ell)}(\\pmb {x})" + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "text", + "content": " is the pre-activation, " + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "inline_equation", + "content": "h^{(\\ell)}" + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "text", + "content": " is the activation function and " + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "inline_equation", + "content": "w^{(\\ell)}\\in \\mathbb{R}^{n^{(\\ell)}\\times n^{(\\ell -1)}}" + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "text", + "content": " is the weight matrix connecting layer " + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\ell -1" + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mu_k^{(\\ell -1)}" + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\gamma_{k}^{(\\ell -1)}" + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "text", + "content": " are the population mean and standard deviation of the hidden nodes " + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "inline_equation", + "content": "f_{k}^{(\\ell -1)}(\\pmb {x})" + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\alpha_{k}^{(\\ell -1)}" + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\beta_{k}^{(\\ell -1)}" + }, + { + "bbox": [ + 55, + 174, + 296, + 262 + ], + "type": "text", + "content": " are the parameters of the linear transformation. The final output layer is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 266, + 295, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 266, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 96, + 266, + 295, + 327 + ], + "type": "interline_equation", + "content": "f (\\boldsymbol {x}) = \\sum_ {k = 1} ^ {n ^ {(L)}} u _ {k} \\underbrace {\\left(\\frac {f _ {k} ^ {(L)} (\\boldsymbol {x}) - \\mu_ {k} ^ {(L)}}{\\sqrt {\\left(\\gamma_ {k} ^ {(L)}\\right) ^ {2} + \\epsilon}}\\right)} _ {\\text {f i n a l P N l a y e r}} + c _ {0}, \\tag {5}", + "image_path": "9e3ebfcf6b43f6f624a483ad99706a858e3911ed39e108b1eab0bcad2c76b9d8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "inline_equation", + "content": "u\\in \\mathbb{R}^{n^{(L)}}" + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "text", + "content": " is the weight vector connecting the output to the " + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "text", + "content": " -th hidden layer, " + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "inline_equation", + "content": "\\mu_k^{(L)}" + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "inline_equation", + "content": "\\gamma_k^{(L)}" + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "text", + "content": " are the population mean and standard deviation of the hidden nodes " + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "inline_equation", + "content": "f_{k}^{(L)}(\\pmb {x})" + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "text", + "content": " The scaling and shifting parameters are absorbed into " + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "inline_equation", + "content": "u_{k}" + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "inline_equation", + "content": "c_{0}" + }, + { + "bbox": [ + 55, + 333, + 295, + 398 + ], + "type": "text", + "content": " , respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 399, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 509 + ], + "type": "text", + "content": "To eliminate the dependency of PN on batch size, we learn the statistics " + }, + { + "bbox": [ + 55, + 399, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mu_k^{(\\ell)}" + }, + { + "bbox": [ + 55, + 399, + 296, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 399, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\gamma_k^{(\\ell)}" + }, + { + "bbox": [ + 55, + 399, + 296, + 509 + ], + "type": "text", + "content": " in training by introducing two natural constraints derived from the definition of mean and variance. The statistic estimation is optimized together with the model parameters during local training, and aggregated on the server by FL algorithms, such as FedAVG. To be precise, given the loss " + }, + { + "bbox": [ + 55, + 399, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot, \\cdot): \\mathbb{R} \\times \\{0,1\\} \\to \\mathbb{R}" + }, + { + "bbox": [ + 55, + 399, + 296, + 509 + ], + "type": "text", + "content": " and the global dataset " + }, + { + "bbox": [ + 55, + 399, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\cup_{m=1}^{M} \\mathcal{D}_m" + }, + { + "bbox": [ + 55, + 399, + 296, + 509 + ], + "type": "text", + "content": ", the training objective in Eqn. (2) of a FL problem becomes" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 513, + 291, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 513, + 291, + 575 + ], + "spans": [ + { + "bbox": [ + 59, + 513, + 291, + 575 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta} \\mathcal {L} (\\theta , \\mathcal {D}) s. t. \\left\\{ \\begin{array}{l} \\frac {1}{| \\mathcal {D} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\mathcal {D}} \\frac {\\left(f _ {k} ^ {(\\ell)} (\\boldsymbol {x}) - \\mu_ {k} ^ {(\\ell)}\\right) ^ {2}}{\\left(\\gamma_ {k} ^ {(\\ell)}\\right) ^ {2} + \\epsilon} = 1 \\\\ \\frac {1}{| \\mathcal {D} |} \\sum_ {(\\boldsymbol {x}, y) \\in \\mathcal {D}} f _ {k} ^ {(\\ell)} (\\boldsymbol {x}) = \\mu_ {k} ^ {(\\ell)} \\end{array} \\right., \\\\ k \\in [ m ^ {(\\ell)} ], \\ell \\in \\{0 \\} \\cup [ L ] \\\\ \\end{array}", + "image_path": "0c729d1e5b261e702a177863906b678e8cff14c22aea5a09b5879805249c943b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 580, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 580, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 580, + 295, + 677 + ], + "type": "text", + "content": " is the set of all the trainable parameters, i.e., " + }, + { + "bbox": [ + 55, + 580, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\theta = \\{w^{(\\ell)}\\} \\cup \\{u\\} \\cup \\{\\mu^{(\\ell)}\\} \\cup \\{\\gamma^{(\\ell)}\\} \\cup \\{\\alpha^{(\\ell)}\\} \\cup \\{\\beta^{(\\ell)}\\} \\cup \\{c_0\\}" + }, + { + "bbox": [ + 55, + 580, + 295, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 580, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta ,\\mathcal{D}) = \\sum_{m = 1}^{M}\\rho_{m}\\sum_{(\\boldsymbol {x},y)\\in \\mathcal{D}_{m}}\\phi (f(\\boldsymbol {x}),y)" + }, + { + "bbox": [ + 55, + 580, + 295, + 677 + ], + "type": "text", + "content": ". This training problem can be solved by various stochastic algorithms with arbitrary batch sizes (even equals to 1), such as moving average or the method of Lagrangian multipliers. In the experiments of this paper, we use stochastic Lagrangian multipliers and the details can be found in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Discussion. (1) Since the statistics " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " are learned in the training process instead of being estimated from minibatches, the resulting training problem can be solved by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "stochastic solvers of FL with arbitrary batch sizes (even equals to 1). Thus the reliance on batch size is eliminated. This enables resource limited clients to train larger deep neural networks. (2) Both " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " are treated as normal trainable parameters in local training, communication and aggregation, our PN layers are consistent among the clients although the datasets on the clients are heterogeneous. This enables us to eventually obtain an off-the-shelf global model, which is often required in reality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 186, + 555, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 211 + ], + "type": "text", + "content": "Remark 3.1 We should point out that our PN layer can also be inserted before the activation functions." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 228, + 425, + 308 + ], + "blocks": [ + { + "bbox": [ + 318, + 228, + 425, + 308 + ], + "lines": [ + { + "bbox": [ + 318, + 228, + 425, + 308 + ], + "spans": [ + { + "bbox": [ + 318, + 228, + 425, + 308 + ], + "type": "image", + "image_path": "a58edbede5168160128fe734085f58f0965127749be6532017594e7fe1aa878f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 311, + 555, + 345 + ], + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 345 + ], + "type": "text", + "content": "Figure 1. Training loss (left) and test loss (right) of BN, LBN and LBN+noise. Injecting artificial noise into estimated BN statistics successfully decreases the test loss and improves generalization." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 433, + 228, + 542, + 308 + ], + "blocks": [ + { + "bbox": [ + 433, + 228, + 542, + 308 + ], + "lines": [ + { + "bbox": [ + 433, + 228, + 542, + 308 + ], + "spans": [ + { + "bbox": [ + 433, + 228, + 542, + 308 + ], + "type": "image", + "image_path": "81a9c9f0b96329886b774af40745de8efbfc5465b307444165d5496593bd56dd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 358, + 493, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 493, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 493, + 372 + ], + "type": "text", + "content": "4. Noisy Population Normalization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 378, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 521 + ], + "type": "text", + "content": "We observe that PN demonstrates inferior performance relative to BN when batch size becomes larger. We explore this issue and find its cause to be that PN lacks stochastic uncertainty in the mean and variance estimations. In practice, large batch sizes may often be used in cross-silo federated learning, where the participants are companies and organizations with huge amounts of data as well as abundant computational resources [9, 15, 21]. Therefore, to enable our method to be more flexibly used in such scenarios, we propose NPN, which injects artificial noise into the learned statistics during training to boost generalization. The details are presented in the following subsections." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 529, + 529, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 529, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 529, + 541 + ], + "type": "text", + "content": "4.1. The Role of Stochastic Uncertainty in BN" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 545, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 582 + ], + "type": "text", + "content": "We provide an in-depth exploration into the role of stochastic uncertainty of the mini-batch statistics in BN. This provides a deep understanding of BN's batch size reliance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": "We start with a simple experiment to show the effect of batch size on BN. Precisely, we compare the performances of BN with a reasonable batch size (m=128) and a very large size (m=2048) in training VGG-16 on CIFAR-10. For simplicity, we refer to BN with m=2048 as LBN. For fair comparison, in LBN, to eliminate the effect of large batch size on SGD, we only use large batches to obtain batch statistics and calculate the stochastic gradient with respect to the loss on mini-batches of size 128." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "We present both the training and test losses of the learned models with BN and LBN over the training process in Fig" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10217" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "1, from which we can see that LBN converges faster than BN, due to the more accurate population statistics estimation. However, there is a noticeable gap between their test losses, which becomes larger after the learning rate decays. Therefore, the effect of BN in improving generalization is reduced when the batch size becomes too large. This indicates that the stochastic uncertainty in estimates of the mean and variance is important for BN." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 294, + 266 + ], + "type": "text", + "content": "The next question we want to investigate is whether we can artificially inject a proper amount of data-independent noise into LBN to improve its generalization performance. Note that the noise of estimating " + }, + { + "bbox": [ + 55, + 167, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 167, + 294, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 167, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 167, + 294, + 266 + ], + "type": "text", + "content": " in BN is data dependent. To answer this question, we should first calculate the uncertainty explicitly. The studies [29, 36] show that for sufficiently large batch size M, the estimated " + }, + { + "bbox": [ + 55, + 167, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 167, + 294, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 167, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 167, + 294, + 266 + ], + "type": "text", + "content": " can be viewed as two random variables sampled from two Gaussian distributions, i.e.," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 271, + 295, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 271, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 271, + 295, + 297 + ], + "type": "interline_equation", + "content": "\\mu_ {\\mathcal {B}} \\sim \\mathcal {N} \\left(\\mu , \\frac {\\sigma^ {2}}{m}\\right) \\text {a n d} \\sigma_ {\\mathcal {B}} \\sim \\mathcal {N} \\left(\\sigma , \\left(\\frac {\\mu_ {4} / \\sigma^ {4} - 1}{4 m}\\right) \\sigma^ {2}\\right), \\tag {6}", + "image_path": "a70e6cfd2f2ff3f047de5e1d58f265e36a66ab724188422804344bf43659ac9f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "text", + "content": " are the mean and variance of the activation " + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "inline_equation", + "content": "x\\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mu_{4}" + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "text", + "content": " is its 4-th central moment, i.e., " + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mu_4 = \\mathbb{E}_x(x - \\mu)^4" + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "text", + "content": ". The item " + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mu_4 / \\sigma^4" + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "text", + "content": " in Eqn.(6) is well-known as the kurtosis [6]]. It holds that " + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mu_4 / \\sigma^4\\geq 1" + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "text", + "content": " for any distribution. Although there is no upper limit to kurtosis of a general distribution, for most well-known distributions, e.g., Gaussian distribution, uniform distribution and logistic distribution, it is finite [6]. Therefore, the variance of " + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "text", + "content": " can always be regarded as " + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\frac{\\sigma^2}{m})" + }, + { + "bbox": [ + 55, + 302, + 295, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "text", + "content": "Notice that the magnitudes of both " + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "text", + "content": " of different neurons could be quite different. The analysis above indicates that to inject noise into " + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "text", + "content": ", it is better to multiply them by a Gaussian noise centered at 1 with proper variance instead of add them by two zero-mean noises. One reason is that the standard deviation of " + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "text", + "content": " is proportional to " + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "text", + "content": " is sufficiently large. More importantly, it is impossible to choose the right amount of noise for each " + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 410, + 295, + 539 + ], + "type": "text", + "content": " if we add noises additively. Therefore, we multiplicatively inject noise into its mean and variance estimations of LBN as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 551, + 246, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 551, + 246, + 562 + ], + "spans": [ + { + "bbox": [ + 104, + 551, + 246, + 562 + ], + "type": "interline_equation", + "content": "\\mu_ {\\mathcal {B}} \\leftarrow \\mu_ {\\mathcal {B}} \\cdot n _ {\\mu} \\text {a n d} \\sigma_ {\\mathcal {B}} \\leftarrow \\sigma_ {\\mathcal {B}} \\cdot n _ {\\sigma},", + "image_path": "4bc00f46a5dbfccd18d1b45fdf4ed255ad78dc385c50eadd0075ff7c1873f281.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "n_{\\mu}" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "n_{\\sigma}" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " are two noises sampled from " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(1,0.1)" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " and they are treated as untrainable parameters in each back propagation. The result is presented in Fig 1. It shows that with such noise, the performance of LBN can be boosted to be comparable with BN." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 712 + ], + "type": "text", + "content": "The main conclusion we can reach in the experiment above is that data-independent noise can be used to improve the performance of LBN. Moreover, it also indicates that the effect of BN on improving generalization can be decomposed into two parts: one is population normalization (can be seen from the improvement of LBN over w/o bn) and the other is the stochastic uncertainty contained in the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "mini-batch statistics (can be seen from the improvement of LBN+noise over LBN). This phenomenon inspires us to inject stochastic noise into our PN to improve its generalization performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 126, + 487, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 487, + 138 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 487, + 138 + ], + "type": "text", + "content": "4.2. Noisy Population Normalization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 143, + 553, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 143, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 553, + 202 + ], + "type": "text", + "content": "Below, we develop our final method NPN (Alg.2) to achieve better generalization performance. It is inspired from our above understandings of the stochastic uncertainty in BN. In particular, we showed that data independent noise can be injected into LBN to boost the generalization performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "text", + "content": "At first, for each activation " + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "text", + "content": ", we insert data-independent noise into the corresponding " + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "text", + "content": " to improve generalization. In order to reduce the effects of different magnitudes of " + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "text", + "content": " corresponding to different neurons, similar to LBN+noise in Section 4.1, we multiply " + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 203, + 553, + 278 + ], + "type": "text", + "content": " by two noises centered at 1 instead of adding two zero-mean ones as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 284, + 519, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 284, + 519, + 297 + ], + "spans": [ + { + "bbox": [ + 348, + 284, + 519, + 297 + ], + "type": "interline_equation", + "content": "\\hat {\\mu} = \\mu \\left(1 + n _ {\\mu}\\right) \\text {a n d} \\hat {\\gamma} = \\max \\{0, \\gamma \\left(1 + n _ {\\gamma}\\right) \\},", + "image_path": "2fe749e5bad6c595c85d21d177149bcee573af93678c2ee1a320a92b7003a2d8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "inline_equation", + "content": "n_{\\mu}" + }, + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "inline_equation", + "content": "n_{\\gamma}" + }, + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "text", + "content": " are independently sampled noises from a zero-mean distribution, e.g., Gaussian, and max is used to avoid negative " + }, + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "inline_equation", + "content": "\\hat{\\gamma}" + }, + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "text", + "content": ". We then normalize, scale and shift " + }, + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 373, + 338, + 553, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 338, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 373, + 338, + 553, + 362 + ], + "type": "interline_equation", + "content": "\\hat {x} = \\frac {x - \\hat {\\mu}}{\\sqrt {\\hat {\\gamma} ^ {2} + \\epsilon}} \\text {a n d} y = \\alpha \\hat {x} + \\beta , \\tag {NPN}", + "image_path": "9530f9698800aa1a70dd3999e36ba9a99b727fe44c534a5e17199d722711852f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 365, + 553, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 365, + 553, + 424 + ], + "type": "inline_equation", + "content": "\\mu, \\gamma, \\alpha" + }, + { + "bbox": [ + 313, + 365, + 553, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 553, + 424 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 365, + 553, + 424 + ], + "type": "text", + "content": " are trainable parameters. The training process with NPN is almost the same as PN, except that the injected noises " + }, + { + "bbox": [ + 313, + 365, + 553, + 424 + ], + "type": "inline_equation", + "content": "n_{\\mu}" + }, + { + "bbox": [ + 313, + 365, + 553, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 553, + 424 + ], + "type": "inline_equation", + "content": "n_{\\gamma}" + }, + { + "bbox": [ + 313, + 365, + 553, + 424 + ], + "type": "text", + "content": " are treated as constants in back propagation. The details of the training process can be found in the appendix." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 425, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 568 + ], + "type": "text", + "content": "Discussion. NPN has the following appealing features. As the injected noises " + }, + { + "bbox": [ + 313, + 425, + 554, + 568 + ], + "type": "inline_equation", + "content": "n_{\\mu}" + }, + { + "bbox": [ + 313, + 425, + 554, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 425, + 554, + 568 + ], + "type": "inline_equation", + "content": "n_{\\gamma}" + }, + { + "bbox": [ + 313, + 425, + 554, + 568 + ], + "type": "text", + "content": " are data independent, we are able to flexibly inject a proper amount of noise into the learnt estimation of statistics to improve generalization ability, which is particularly important when large batch sizes are used. This scenario is often encountered in cross silo federated learning. The idea of injecting data independent noise to improve generalization can also be combined with other normalization methods. The success of NPN shows that one can directly learn the population statistics in the training process instead of estimating them from the minibatches. This can inspire new normalization methods." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 574, + 555, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 706 + ], + "type": "text", + "content": "Remark 4.1 It is unclear that whether NPN has some other nice properties of BN, e.g., auto-tuning of learning rates due to scale invariance [1, 7], while our empirical results demonstrate the effectiveness of NPN. We argue that different from standard training, the performance improvements achieved by these properties could be negligible compared to normalization itself in FL as the training in FL is more challenging than standard training. Such theoretical analysis is left as future work. Thus, we would like to point out that NPN is a normalization technique for FL instead of a general approach." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10218" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 69, + 478, + 321 + ], + "blocks": [ + { + "bbox": [ + 133, + 69, + 478, + 321 + ], + "lines": [ + { + "bbox": [ + 133, + 69, + 478, + 321 + ], + "spans": [ + { + "bbox": [ + 133, + 69, + 478, + 321 + ], + "type": "table", + "html": "
batch sizeNormalizationα (CIFAR-10)α (CIFAR-100)
0.040.080.161.000.040.080.161.00
8w/o69.8383.2285.0587.5945.1350.4354.2658.47
IN67.6878.5582.6284.0935.5140.0245.4845.45
GN71.0582.5483.3586.1145.4549.2252.2054.75
BN70.2282.2785.7789.4252.1355.9158.1664.88
FedBN71.4983.0584.4189.3052.0556.1257.9764.43
PN78.5985.4588.0589.6353.0657.4361.8765.00
4w/o70.6982.0285.1186.9346.9950.8355.0859.45
IN60.5374.0878.1182.3936.0442.9646.1749.89
GN68.6378.5382.0185.5244.2148.1851.8554.67
BN70.1882.4786.3289.5251.3155.7860.3564.32
FedBN72.2483.5786.0389.4651.4555.7759.9165.10
PN78.4085.0587.5289.5952.2856.8562.4265.82
2w/o67.0379.9883.4387.0143.9348.5654.7759.43
IN55.4268.9375.4378.7936.1543.9047.7247.30
GN66.0278.8281.7585.2344.4046.1750.6053.24
BN65.3279.6883.3087.2545.8750.6655.0658.74
FedBN65.5080.0282.1786.4046.0451.3055.5259.68
PN72.5382.5986.2388.9848.9754.9260.6465.46
1w/o60.4971.8378.6584.9338.2442.2649.4954.59
IN35.4754.9366.4573.7433.6039.0643.2948.82
GN60.5571.8375.2579.6846.0449.2653.9756.64
BNNANANANANANANANA
FedBNNANANANANANANANA
PN64.3878.8982.1386.6946.1349.3259.1962.97
", + "image_path": "1d93c7b6b52f7d6062d5f0d8908078ee4c35ec458a60ed7fea0f52df49c9aa1f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "type": "text", + "content": "Table 1. The results of different normalization methods on CIFAR-10/100 with different batch sizes and Non-IID degree " + }, + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "type": "text", + "content": " . Or PN consistently outperforms other methods by a large margin across different " + }, + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "type": "text", + "content": " and batch sizes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 56, + 372, + 243, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 243, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 243, + 384 + ], + "type": "text", + "content": "Algorithm 2 Noisy Population Normalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "text", + "content": "Input: activation: " + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "text", + "content": "; population mean and variance: " + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "inline_equation", + "content": "1 / \\gamma^2" + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "text", + "content": "; scale and shift parameter: " + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "text", + "content": "; noise level: " + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 64, + 388, + 295, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 423, + 169, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 423, + 169, + 434 + ], + "spans": [ + { + "bbox": [ + 65, + 423, + 169, + 434 + ], + "type": "text", + "content": "Inject noise into " + }, + { + "bbox": [ + 65, + 423, + 169, + 434 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 65, + 423, + 169, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 65, + 423, + 169, + 434 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 65, + 423, + 169, + 434 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 442, + 266, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 442, + 266, + 455 + ], + "spans": [ + { + "bbox": [ + 94, + 442, + 266, + 455 + ], + "type": "interline_equation", + "content": "\\hat {\\mu} = \\mu \\left(1 + n _ {\\mu}\\right) \\text {a n d} \\hat {\\gamma} = \\max \\{0, \\gamma \\left(1 + n _ {\\gamma}\\right) \\},", + "image_path": "d36b374196f7b7d4695234c20e6b3539822bfa03be998a966516d446e0f3eeb4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 461, + 245, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 461, + 245, + 486 + ], + "spans": [ + { + "bbox": [ + 66, + 461, + 245, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 66, + 461, + 245, + 486 + ], + "type": "inline_equation", + "content": "n_{\\mu}" + }, + { + "bbox": [ + 66, + 461, + 245, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 66, + 461, + 245, + 486 + ], + "type": "inline_equation", + "content": "n_{\\gamma}" + }, + { + "bbox": [ + 66, + 461, + 245, + 486 + ], + "type": "text", + "content": " are noises from " + }, + { + "bbox": [ + 66, + 461, + 245, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\sigma^2)" + }, + { + "bbox": [ + 66, + 461, + 245, + 486 + ], + "type": "text", + "content": ". Normalize using noisy population statistics:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 493, + 208, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 493, + 208, + 517 + ], + "spans": [ + { + "bbox": [ + 151, + 493, + 208, + 517 + ], + "type": "interline_equation", + "content": "\\hat {x} = \\frac {x - \\hat {\\mu}}{\\sqrt {\\hat {\\gamma} ^ {2} + \\epsilon}}", + "image_path": "0a7fbec54264478ec792c842e4751b3594bb9badb4086cb02317ae4fdfa0cc2a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 528, + 130, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 528, + 130, + 539 + ], + "spans": [ + { + "bbox": [ + 66, + 528, + 130, + 539 + ], + "type": "text", + "content": "Scale and shift:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 123, + 548, + 236, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 548, + 236, + 562 + ], + "spans": [ + { + "bbox": [ + 123, + 548, + 236, + 562 + ], + "type": "interline_equation", + "content": "y = \\alpha \\hat {x} + \\beta \\equiv \\mathrm {N P N} _ {\\alpha , \\beta} ^ {\\mu , \\gamma} (\\sigma ; x)", + "image_path": "3ac94eee8790e6389b233081c222ab0edc42efc51ebd82201d1ffcf85556e009.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 580, + 111, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 580, + 111, + 593 + ], + "spans": [ + { + "bbox": [ + 66, + 580, + 111, + 593 + ], + "type": "text", + "content": "Return: y" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 627, + 137, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 137, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 137, + 641 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "In this section, we conduct two sets of experiments. Firstly, we verify the effectiveness of of PN in the federated learning setting given small batch sizes and heterogeneous client data. Secondly, we demonstrate the superior generalization ability of our NPN given large training batch sizes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 372, + 440, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 440, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 440, + 384 + ], + "type": "text", + "content": "5.1. Settings and Baselines" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 389, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 509 + ], + "type": "text", + "content": "To verify the advantage of our population normalization method, we compare it with other normalization counterparts in the FL setting. Specifically, we compare with batch normalization [16], group normalization [40], instance normalization [37] and without normalization. We compare with FedBN [25] by averaging the BN parameters in the clients. Extensive experiments are conducted on two image classification benchmarks (i.e., CIFAR-10 and CIFAR-100) with models VGG-16 [33] and ResNet [11] to evaluate the effectiveness of our method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "For FL experiments, we adopt FedAvg [30] as the base algorithm. We use 80 clients in total with participation rate set to 0.4 throughout all experiments. For the clients' local update, we adopt SGD with momentum. For batch size of 8, the learning rates are set to be the best among " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{0.005, 0.01, 0.15, 0.02, 0.25\\}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " for each normalization method, and linearly down-scaled for other smaller batch sizes. We run 200 global communication rounds with local epoch set to 1. We examine the superiority of our method against the baseline under various batch sizes and heterogeneity degrees. We follow prior work [5, 28] to use Dirichlet distribution for simulating the non-IID data distribution, where the degree of heterogeneity is defined by " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", smaller " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " value corresponds to more severe heterogeneity. We report the global model's average performance in the last five rounds evaluated using the test split of the datasets. For other details, please refer to the Appendix." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10219" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 74, + 231, + 182 + ], + "blocks": [ + { + "bbox": [ + 83, + 74, + 231, + 182 + ], + "lines": [ + { + "bbox": [ + 83, + 74, + 231, + 182 + ], + "spans": [ + { + "bbox": [ + 83, + 74, + 231, + 182 + ], + "type": "image", + "image_path": "fccbce6add3e42effc06f4e77385fb8472154b3a7fabd6ca2bd7df33d985c30b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 186, + 235, + 196 + ], + "lines": [ + { + "bbox": [ + 225, + 186, + 235, + 196 + ], + "spans": [ + { + "bbox": [ + 225, + 186, + 235, + 196 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 231, + 74, + 380, + 182 + ], + "blocks": [ + { + "bbox": [ + 231, + 74, + 380, + 182 + ], + "lines": [ + { + "bbox": [ + 231, + 74, + 380, + 182 + ], + "spans": [ + { + "bbox": [ + 231, + 74, + 380, + 182 + ], + "type": "image", + "image_path": "9e0cb30003cdbb0168fb67d854e4880702d5de8b46105a1e083faad685bccd29.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 200, + 544, + 212 + ], + "lines": [ + { + "bbox": [ + 64, + 200, + 544, + 212 + ], + "spans": [ + { + "bbox": [ + 64, + 200, + 544, + 212 + ], + "type": "text", + "content": "Figure 2. (a) Training (left) and test (right) loss of BN, PN, and NPN as the training processes; (b) Sensitivity of NPN on noise level." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 380, + 72, + 529, + 182 + ], + "blocks": [ + { + "bbox": [ + 380, + 72, + 529, + 182 + ], + "lines": [ + { + "bbox": [ + 380, + 72, + 529, + 182 + ], + "spans": [ + { + "bbox": [ + 380, + 72, + 529, + 182 + ], + "type": "image", + "image_path": "e6eca362c319f2b1bf1acbcd8bde40903267ef46310e9ef53fc0bc7b79b29d37.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 186, + 459, + 195 + ], + "lines": [ + { + "bbox": [ + 447, + 186, + 459, + 195 + ], + "spans": [ + { + "bbox": [ + 447, + 186, + 459, + 195 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 280, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 280, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 280, + 304 + ], + "type": "text", + "content": "For experiments with large batch sizes, we adopt the distributed training setting, where the model is being trained 4 different participants. The batch size is set to 2048. Other hyper-parameters are set to be the same as in the setting. We conduct the experiments using VGG-16 train on CIFAR-10 and CIFAR-100 datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 313, + 276, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 276, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 276, + 326 + ], + "type": "text", + "content": "5.2. Federated Learning with Small Batch Size" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "type": "text", + "content": "We point out that with small batch sizes, SGD already introduces enough stochastic uncertainty in gradient estimation, which helps generalization. Therefore, we do not need to inject additional noise for PN and thus the experiments with NPN are omitted in this section." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "content": "Comparison with Various Normalization Methods We compare PN with other counterparts under different batch sizes and non-iid degrees in Table 1. Specifically, we conduct experiments with " + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\{0.04, 0.08, 0.16, 1.00\\}" + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "content": " and batch sizes " + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\{1, 2, 4, 8\\}" + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "content": ". We adopt VGG16 as the trained network. As demonstrated in the table, we observe that our method consistently outperforms BN and other baselines, while the advantage is more significant when the batch size is small and the heterogeneity degree is higher. Notably, our method works well even with batch size set to 1, while the vanilla BN becomes inapplicable under this scenario. This enables our method to be used when the computational resource is limited on the client side." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Experiments with Various Networks We confirm the effectiveness of PN across different network architectures and capacities. As shown in Table 3, we conduct experiments with VGG16, ResNet18 and ResNet50. We fix the heterogeneity degree " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " to be 0.08 and trial with batch sizes of " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{1,2,4,8\\}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". We observe consistent performance advantage over the baselines for different networks. Interestingly, one may notice that networks with higher capacities do not necessarily lead to better performances, which is opposite to standard training. This phenomenon is also reported in [5], which may be due to clients' local over-fitting or further" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 232, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 269 + ], + "type": "text", + "content": "drifting from each other in the parameter space. However, PN still demonstrates superior performances compared with other normalization techniques." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 279, + 535, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 535, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 535, + 293 + ], + "type": "text", + "content": "5.3. Federated Learning with Large Batch Size" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 298, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 453 + ], + "type": "text", + "content": "We now extend our PN to NPN to enable stronger generalization ability for training with large batch sizes. The results are demonstrated in Table 4. As can be seen, when training with large batch sizes, PN is outperformed by BN due to the lack of stochastic uncertainty in the normalization layer. On the other hand, NPN remedies this issue by injecting artificial noise into the statistics estimation, which in turn leads to better test performance. The training progress is also visualized in Figure 2(a). We can see that although the training loss of NPN is larger than that of BN, its testing loss is significantly smaller than BN. This demonstrates that benefiting from the injected noise, NPN can prevent the training process from over-fitting more effectively than BN." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 344, + 465, + 525, + 505 + ], + "blocks": [ + { + "bbox": [ + 344, + 465, + 525, + 505 + ], + "lines": [ + { + "bbox": [ + 344, + 465, + 525, + 505 + ], + "spans": [ + { + "bbox": [ + 344, + 465, + 525, + 505 + ], + "type": "table", + "html": "
MethodLossTop-1Top-5Top-10
BRN0.4493.2197.5298.83
BRN + noise0.2993.6097.8399.02
", + "image_path": "0ec3266f1977e5f452867db767163f128b4a420e40ef8c7d4fe45cf2024d7779.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 514, + 508, + 524 + ], + "lines": [ + { + "bbox": [ + 359, + 514, + 508, + 524 + ], + "spans": [ + { + "bbox": [ + 359, + 514, + 508, + 524 + ], + "type": "text", + "content": "Table 2. Batch Renormalization + noise." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 540, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 601 + ], + "type": "text", + "content": "Finally, we point out that there are only a small number of participants in this scenario, which makes the trained model less prone to the adverse impact of heterogeneity and thus shows better performance than the previous experiments with more participating clients (Table 1)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 611, + 436, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 436, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 436, + 624 + ], + "type": "text", + "content": "5.4. Abative Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "Noise Level Sensitivity. We evaluate the sensitivity of NPN to the injected noise level. We conduct experiments with VGG-16 on CIFAR-10. We set the batch size to 2048 and vary the variance of the noise injected into NPN in " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{0.5, 0.1, 0.01, 0.001\\}" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": ". The results are presented in Figure 2(b). It shows that the test loss increases gradually and slowly when we vary the variance of the noise from 0.5 to" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10220" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 555, + 152 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 152 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 152 + ], + "type": "table", + "html": "
Normalizationbatch size (VGG-16)batch size (ResNet-18)batch size (ResNet-50)
124812481248
w/o71.8379.9882.0283.2275.4879.9480.2181.4965.5972.3377.8079.94
IN35.4768.9374.0378.5563.5369.3670.6871.3560.3764.6568.2869.10
GN60.5578.8278.5382.5466.0370.4472.5975.1360.6368.7270.7371.25
BNNA78.6582.0982.27NA78.5579.6481.94NA72.3978.2081.83
PN78.8982.2285.1385.4580.2083.9784.5384.6767.7273.6379.5983.54
", + "image_path": "8d67f86aa6cd0eb6f40874594440a5518c00950b98298a628939a217ef17172e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 156, + 193, + 455, + 273 + ], + "blocks": [ + { + "bbox": [ + 55, + 160, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 160, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 555, + 182 + ], + "type": "text", + "content": "Table 3. Results of different normalization methods on CIFAR-10. PN consistently outperforms other methods for different nwork architectures and batch sizes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 156, + 193, + 455, + 273 + ], + "lines": [ + { + "bbox": [ + 156, + 193, + 455, + 273 + ], + "spans": [ + { + "bbox": [ + 156, + 193, + 455, + 273 + ], + "type": "table", + "html": "
NormalizationCIFAR-10CIFAR-100
LossTop-1Top-2Top-3LossTop-1Top-2Top-3
w/o0.5492.3196.9198.202.7770.3487.9692.11
IN0.4692.4597.0098.312.5470.6888.4992.33
GN0.4392.9297.1498.522.2371.3588.9393.10
BN0.3993.1397.6698.872.0572.2190.2893.60
PN0.4193.1097.7898.971.8572.1590.0793.89
NPN0.3293.6698.0599.121.2272.7491.4794.90
", + "image_path": "403703db580c8bf8afeed380d45570d49a1483ac76719fa8ae88194306641eb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 281, + 488, + 293 + ], + "lines": [ + { + "bbox": [ + 121, + 281, + 488, + 293 + ], + "spans": [ + { + "bbox": [ + 121, + 281, + 488, + 293 + ], + "type": "text", + "content": "Table 4. Results of different normalization methods on CIFAR-10 / CIFAR-100 with large batch size." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 313, + 294, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 294, + 338 + ], + "type": "text", + "content": "0.001. Therefore, it is easy to select a suitable noise variance that falls within a reasonable range." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 353, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 353, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 295, + 521 + ], + "type": "text", + "content": "Artificial Noise for Other Normalization Methods. We demonstrate that the artificially injected noise is also able to improve existing normalization methods with limited stochastic uncertainty. We take batch renormalization (BRN) [16] as an example and conduct experiments on CIFAR-10 with VGG-16. For simplicity and excluding other factors, we conduct this experiment with regular training instead of FL. We inject noise sampled from " + }, + { + "bbox": [ + 54, + 353, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,0.1)" + }, + { + "bbox": [ + 54, + 353, + 295, + 521 + ], + "type": "text", + "content": " to the moving averages " + }, + { + "bbox": [ + 54, + 353, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 353, + 295, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 353, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 54, + 353, + 295, + 521 + ], + "type": "text", + "content": " of BRN as we do for NPN. The results are given in Table 2. It shows that with the injected noise, the performance of BRN can be improved effectively. Therefore, the idea of data-independent noise injection is a very general regularization technique that can be used to improve existing normalization methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 217, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 217, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 217, + 541 + ], + "type": "text", + "content": "5.5. Supervised Learning with PN" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Intuitively, PN can also be used in non-FL scenarios. We apply PN to supervised learning and conduct experiments on two image classification benchmarks (i.e., CIFAR-10, CIFAR-100) with VGG-16 and ResNet to evaluate the effectiveness of PN in supervised learning. The batch sizes of all the methods are set to 128. Table 5 presents the test loss and accuracy of the final learned models. From Table 5 we can see that our NPN can achieve much lower test loss than BN and comparable (if no better than) test accuracy with BN on all the three models. We can also observe that NPN can achieve a noticeable improvement over PN, which demonstrates that artificially injected data independent noise can be used to reinforce the performance of the normalization methods with limited stochastic under" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 329, + 342, + 538, + 468 + ], + "blocks": [ + { + "bbox": [ + 313, + 312, + 553, + 334 + ], + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 334 + ], + "type": "text", + "content": "Table 5. Results of different normalization methods on image classification benchmarks" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 342, + 538, + 468 + ], + "lines": [ + { + "bbox": [ + 329, + 342, + 538, + 468 + ], + "spans": [ + { + "bbox": [ + 329, + 342, + 538, + 468 + ], + "type": "table", + "html": "
NetworkNormalizationLossAccuracy
VGG-16BN0.3992.31
PN0.4193.63
NPN0.2893.66
ResNet-18BN0.4693.76
PN0.4593.97
NPN0.3894.26
ResNet-56BN0.3393.05
PN0.4292.56
NPN0.2593.99
", + "image_path": "58761bae11106903945ddf6a64386e7e1f37863696ba8470a113db263c3ebeba.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 488, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 555, + 537 + ], + "type": "text", + "content": "tainty. The comparable performance of BN and NPN implies that the effect of BN can be decomposed into two parts: one is population normalization, and the other is the stochastic uncertainty containing in the mini-batch statistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 549, + 391, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 391, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 391, + 562 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We propose a learning-based normalization technique, Population Normalization (PN), which reduces the reliance on training batch size and improves robustness against data heterogeneity in federated learning (FL). Extensive experiments demonstrate that PN outperforms existing normalization techniques by a significant margin. We further develop Noisy Population Normalization (NPN), which injects artificial noise into the learnt normalization statistics to enhance generalization under large batch size conditions. We believe that our work will offer valuable insights for the design and development of more stable and adaptable normalization techniques for FL." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10221" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 159 + ], + "type": "text", + "content": "This work was supported by the National Nature Science Foundation of China (62472097) and AI for Science Foundation of Fudan University (FudanX24AI028). The computations in this research were performed on the CFFF platform of Fudan University." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 187, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 187, + 295, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 187, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 187, + 295, + 232 + ], + "type": "text", + "content": "[1] Sanjeev Arora, Kaifeng Lyu, and Zhiyuan Li. Theoretical analysis of auto rate-tuning by batch normalization. In 7th International Conference on Learning Representations, ICLR 2019, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 232, + 296, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 232, + 296, + 265 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 296, + 265 + ], + "type": "text", + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 266, + 296, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 266, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 61, + 266, + 296, + 321 + ], + "type": "text", + "content": "[3] Tobias Bernecker, Annette Peters, Christopher L. Schlett, Fabian Bamberg, Fabian J. Theis, Daniel Rueckert, Jakob Weiss, and Shadi Albarqouni. Fednorm: Modality-based normalization in federated learning for multi-modal liver segmentation. ArXiv, abs/2205.11096, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 322, + 296, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 322, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 61, + 322, + 296, + 366 + ], + "type": "text", + "content": "[4] Yongqiang Cai, Qianxiao Li, and Zuowei Shen. A quantitative analysis of the effect of batch normalization on gradient descent. In International Conference on Machine Learning, pages 882-890, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 367, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 367, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 367, + 296, + 399 + ], + "type": "text", + "content": "[5] Hong-You Chen and Wei-Lun Chao. Fedbe: Making bayesian model ensemble applicable to federated learning. In ICLR, 2021. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 400, + 296, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 296, + 423 + ], + "type": "text", + "content": "[6] Wikipedia contributors. Kurtosis. Online; accessed 2-June-2020, 2020. Wikipedia, The Free Encyclopedia. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 424, + 296, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 296, + 467 + ], + "type": "text", + "content": "[7] Hadi Daneshmand, Amir Joudaki, and Francis Bach. Batch normalization orthogonalizes representations in deep random networks. In Advances in Neural Information Processing Systems, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 468, + 296, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 524 + ], + "type": "text", + "content": "[8] Jian-Hui Duan, Wenzhong Li, and Sanglu Lu. Feddna: Federated learning with decoupled normalization-layer aggregation for non-iid data. In Machine Learning and Knowledge Discovery in Databases. Research Track, pages 722-737, Cham, 2021. Springer International Publishing. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 524, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 524, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 524, + 296, + 579 + ], + "type": "text", + "content": "[9] Aiden Durrant, Milan Markovic, David Matthews, David May, Jessica Enright, and Georgios Leontidis. The role of cross-silo federated learning in facilitating data sharing in the agri-food sector. Computers and Electronics in Agriculture, 193:106648, 2022. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 580, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 296, + 624 + ], + "type": "text", + "content": "[10] Filip Hanzely, Slavomír Hanzely, Samuel Horváth, and Peter Richtárik. Lower bounds and optimal algorithms for personalized federated learning. Advances in Neural Information Processing Systems, 33:2304-2315, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 669 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[13] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip Gibbons. The non-IID data quagmire of decentralized machine learning. In Proceedings of the 37th International Conference on Machine Learning, pages 4387-4398. PMLR, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "type": "text", + "content": "[14] Kevin Hsieh, Amar Phanishayee, Onur Mutlu, and Phillip B. Gibbons. The non-iid data quagmire of decentralized machine learning, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 196 + ], + "type": "text", + "content": "[15] Chao Huang, Jianwei Huang, and Xin Liu. Cross-silo federated learning: Challenges and opportunities. arXiv preprint arXiv:2206.12949, 2022. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 554, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 241 + ], + "type": "text", + "content": "[16] Sergey Ioffe. Batch renormalization: Towards reducing minibatch dependence in batch-normalized models. In Advances in neural information processing systems, pages 1945-1953, 2017. 1, 2, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 242, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 554, + 297 + ], + "type": "text", + "content": "[17] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning, pages 448-456, Lille, France, 2015. PMLR. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "type": "text", + "content": "[18] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank J Reddi, Sebastian U Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for ondevice federated learning. In ICML, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 343, + 554, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 388 + ], + "type": "text", + "content": "[19] Jakub Konečny, H Brendan McMahan, Felix X Yu, Peter Richtárik, Ananda Theertha Suresh, and Dave Bacon. Federated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492, 2016. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 554, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 442 + ], + "type": "text", + "content": "[20] Viraj Kulkarni, Milind Kulkarni, and Aniruddha Pant. Survey of personalization techniques for federated learning. In 2020 Fourth World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4), pages 794-797. IEEE, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 443, + 554, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 554, + 476 + ], + "type": "text", + "content": "[21] Qinbin Li, Bingsheng He, and Dawn Song. Practical oneshot federated learning for cross-silo setting. arXiv preprint arXiv:2010.01017, 2020. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 477, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 554, + 521 + ], + "type": "text", + "content": "[22] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10713–10722, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 522, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 554, + 555 + ], + "type": "text", + "content": "[23] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. In MLSys, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 555, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 601 + ], + "type": "text", + "content": "[24] Tian Li, Shengyuan Hu, Ahmad Beirami, and Virginia Smith. Ditto: Fair and robust federated learning through personalization. In International Conference on Machine Learning, pages 6357-6368. PMLR, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 601, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 646 + ], + "type": "text", + "content": "[25] Xiaoxiao Li, Meirui Jiang, Xiaofei Zhang, Michael Kamp, and Qi Dou. Fed{bn}: Federated learning on non- " + }, + { + "bbox": [ + 316, + 601, + 554, + 646 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{i}\\mathrm{d}\\}" + }, + { + "bbox": [ + 316, + 601, + 554, + 646 + ], + "type": "text", + "content": " features via local batch normalization. In International Conference on Learning Representations, 2021. 2, 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "type": "text", + "content": "[26] Xiangru Lian and Ji Liu. Revisit batch normalization: New understanding and refinement via composition optimization. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 3254-3263, 2019. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "type": "text", + "content": "[27] Wei Yang Bryan Lim, Nguyen Cong Luong, Dinh Thai Hoang, Yutao Jiao, Ying-Chang Liang, Qiang Yang, Dusit" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10222" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Niyato, and Chunyan Miao. Federated learning in mobile edge networks: A comprehensive survey. IEEE Communications Surveys & Tutorials, 22(3):2031-2063, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "text", + "content": "[28] Tao Lin, Lingjing Kong, Sebastian U Stich, and Martin Jaggi. Ensemble distillation for robust model fusion in federated learning. Advances in Neural Information Processing Systems, 33:2351-2363, 2020. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[29] Ping Luo, Xinjiang Wang, Wenqi Shao, and Zhanglin Peng. Towards understanding regularization in batch normalization. arXiv preprint arXiv:1809.00846, 2018. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "text", + "content": "[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "type": "text", + "content": "[31] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. In Advances in neural information processing systems, pages 901-909, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 327 + ], + "type": "text", + "content": "[32] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? In Advances in Neural Information Processing Systems, pages 2483-2493, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "text", + "content": "[33] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "text", + "content": "[34] Sebastian U Stich. Local sgd converges fast and communicates little. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "type": "text", + "content": "[35] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with moreau envelopes. Advances in Neural Information Processing Systems, 33:21394-21405, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[36] Mattias Teye, Hossein Azizpour, and Kevin Smith. Bayesian uncertainty estimation for batch normalized deep networks. In International Conference on Machine Learning, pages 4907-4916, 2018. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 492, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 525 + ], + "type": "text", + "content": "[37] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "type": "text", + "content": "[38] Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H Vincent Poor. Tackling the objective inconsistency problem in heterogeneous federated optimization. Advances in neural information processing systems, 33:7611-7623, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "text", + "content": "[39] Xiaoxia Wu, Edgar Dobriban, Tongzheng Ren, Shanshan Wu, Zhiyuan Li, Suriya Gunasekar, Rachel Ward, and Qiang Liu. Implicit regularization of normalization methods. arXiv preprint arXiv:1911.07956, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "text", + "content": "[40] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 3-19, 2018. 1, 2, 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[41] Jianhang Xiao, Chunhui Du, Zijing Duan, and Wei Guo. A novel server-side aggregation strategy for federated learning in non-id situations. In 2021 20th International Symposium on Parallel and Distributed Computing (ISPDC), pages 17-24. IEEE, 2021. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 173 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "type": "text", + "content": "[42] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "text", + "content": "[43] Yousef Yeganeh, Azade Farshad, Nassir Navab, and Shadi Albarqouni. Inverse distance aggregation for federated learning with non-iid data. In Domain Adaptation and Representation Transfer, and Distributed and Collaborative Learning, pages 150-159. Springer, 2020. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10223" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/162f30eb-880f-42fb-95ca-56f10c93b7f6_content_list.json b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/162f30eb-880f-42fb-95ca-56f10c93b7f6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ef979b8281dc4c8e2a9a50133d679a03b2dc08e9 --- /dev/null +++ b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/162f30eb-880f-42fb-95ca-56f10c93b7f6_content_list.json @@ -0,0 +1,1563 @@ +[ + { + "type": "text", + "text": "Pos3R: 6D Pose Estimation for Unseen Objects Made Easy", + "text_level": 1, + "bbox": [ + 202, + 130, + 794, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weijian Deng $^{1}$ Dylan Campbell $^{1}$ Chunyi Sun $^{1}$ Jiahao Zhang $^{1}$", + "bbox": [ + 225, + 179, + 769, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shubham Kanitkar $^{2}$ Matthew E. Shaffer $^{2}$ Stephen Gould $^{1}$", + "bbox": [ + 250, + 198, + 745, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Australian National University $^{2}$ RIOS Intelligent Machines", + "bbox": [ + 248, + 215, + 748, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{\\text {firstname.lastname } \\} @ \\text { anu.edu.au } ^{2}\\{\\text {firstname.lastname } \\} @ \\text { rios.ai }$", + "bbox": [ + 133, + 233, + 849, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundation models have significantly reduced the need for task-specific training, while also enhancing generalizability. However, state-of-the-art 6D pose estimators either require further training with pose supervision or neglect advances obtainable from 3D foundation models. The latter is a missed opportunity, since these models are better equipped to predict 3D-consistent features, which are of significant utility for the pose estimation task. To address this gap, we propose Pos3R, a method for estimating the 6D pose of any object from a single RGB image, making extensive use of a 3D reconstruction foundation model and requiring no additional training. We identify template selection as a particular bottleneck for existing methods that is significantly alleviated by the use of a 3D model, which can more easily distinguish between template poses than a 2D model. Despite its simplicity, Pos3R achieves competitive performance on the Benchmark for 6D Object Pose Estimation (BOP), matching or surpassing existing refinement-free methods. Additionally, Pos3R integrates seamlessly with render-and-compare refinement techniques, demonstrating adaptability for high-precision applications.", + "bbox": [ + 88, + 316, + 485, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 662, + 223, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Six-dimensional (6D) object pose estimation—the task of determining the exact position and orientation of objects relative to a camera—is essential for applications in robotics, augmented reality, and autonomous systems. Reliable pose estimation enables critical tasks like object manipulation, grasping, and assembly, allowing these systems to interact effectively in complex and dynamic environments [6, 18, 29, 30, 49, 54]. Traditional approaches often rely on learning-based methods tailored to specific objects or categories, achieving high accuracy but struggling to generalize to new categories or unseen objects [4, 5, 7, 26, 33, 46, 47, 53-55, 59, 61]. This limitation is particularly problematic in dynamic, data-scarce environments where adaptability and flexibility are crucial.", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a51ba4ad4b7ceaff74356db999e6f95204ed479b6381b88b278f7a244ad1c10.jpg", + "image_caption": [ + "Figure 1. Illustration of In-Plane and Out-of-Plane Rotations with Correspondence Quality between DINOv2 and MASt3R." + ], + "image_footnote": [], + "bbox": [ + 517, + 287, + 906, + 401 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The left diagram illustrates in-plane rotations (within the image plane) and out-of-plane rotations (3D orientation changes) from a camera view. On the right, DINOv2 [40] (top row) and MASt3R [24] (bottom row) are compared in their handling of these rotations. DINOv2 exhibits sparse and inconsistent correspondences (shown by red lines), particularly under out-of-plane rotations, due to its 2D feature limitations. In contrast, MASt3R provides dense and stable correspondences across both types of rotations, reflecting its ability to produce 3D-consistent features.", + "bbox": [ + 511, + 441, + 906, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome these limitations, there has been a recent shift toward model-based approaches that aim to generalize pose estimation to unseen objects without object-specific training [2, 17, 41, 56]. These typically rely on a two-stage pipeline: detecting and localizing objects within a scene, followed by a render-and-compare process that matches detected object regions to a set of template models. Building on this concept, recent works have sought to improve generalization through large-scale model training with significant attention to object diversity. For example, Foundation-Pose [56] uses synthetic training data combined with a large language model and contrastive learning, enhancing feature alignment across varied domains for robust generalization.", + "bbox": [ + 509, + 595, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent emergence of training-free methods offers a promising alternative, enabling 6D pose estimation for unseen objects without object- or task-specific training. Foundation models like DINOv2 [40] have demonstrated strong zero-shot capabilities, capturing spatial and semantic details through learned features [1, 58]. FoundPose [41], for instance, leverages DINOv2 descriptors to bridge synthetic", + "bbox": [ + 509, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16818", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and real data, achieving competitive results on the BOP challenge [17]. These advancements underscore the potential of training-free frameworks in 6D pose estimation.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on this trend, we explore the strengths of 3D foundation models. Accurate pose estimation requires handling both in-plane and out-of-plane rotations, which are demonstrated in Fig. 1. In-plane rotations occur within the image plane, as when an object rotates around the camera's line of sight, while out-of-plane rotations involve changes in 3D orientation, such as tilting or turning, which significantly alters appearance due to perspective shifts. Addressing both types of rotation is essential for robust pose estimation, as objects often appear from varied angles. Although 2D foundation models like DINOv2 handle in-plane rotations well through training augmentations, they lack robustness to out-of-plane rotations due to their 2D limitations (shown in Fig. 1). In contrast, 3D foundation models, such as MASt3R [24], are specifically designed to produce 3D-consistent features across different viewpoints, enabling reliable feature alignment even with out-of-plane rotations.", + "bbox": [ + 89, + 138, + 482, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by this, we propose Pos3R, a training-free method for 6D pose estimation of unseen objects using only RGB inputs. Pos3R leverages the 3D foundation model MASt3R [24] for pose estimation without requiring additional training. The core of Pos3R lies in the image matching process between test crops and rendered templates from a CAD model. While common matchers like LoFTR [50] often struggle with synthetic-to-real matching due to domain gaps, MASt3R produces high-quality 2D correspondences in this case. These reliable matches enable Pos3R to establish the 2D-3D correspondences required for pose estimation using the PnP-RANSAC algorithm.", + "bbox": [ + 89, + 397, + 482, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By harnessing MASt3R's ability to produce dense correspondences robust to viewpoint and illumination changes, we simplify template generation and achieve computational efficiency. Unlike existing methods that render hundreds of templates, we place camera positions to cover both in-plane and out-of-plane rotations. Specifically, we use eight base templates per object, positioned at the vertices of a cube centered on the CAD model, to capture out-of-plane rotations. For each base template, we generate five in-plane rotational templates along the object's principal axis, resulting in forty templates per object. To ensure efficient and accurate pose estimation, Pos3R dynamically selects the optimal rotation based on matching quality. This simple yet effective strategy maintains computational efficiency while achieving reliable pose estimation. Tested on seven diverse datasets in the BOP challenge, Pos3R demonstrates strong performance as a robust, scalable option for training-free 6D pose estimation of unseen objects, adaptable to render-and-compare refinement techniques.", + "bbox": [ + 89, + 580, + 482, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 109, + 869, + 339, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- While existing methods rely on 2D foundation models", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for unseen object pose estimation, we present Pos3R, a training-free method that leverages the 3D foundation model MAST3R [24] to improve robustness.", + "bbox": [ + 526, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Leveraging MASt3R's robust dense correspondences, Pos3R uses only forty strategically placed templates per object to capture in-plane and out-of-plane rotations. A simple selection technique based on correspondence quality ensures accurate matching, achieving strong performance on the BOP challenge.", + "bbox": [ + 513, + 136, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 244, + 653, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Seen Object Pose Estimation. Instance-level object pose estimation refers to the task of estimating the poses of specific objects previously encountered during model training [5, 7, 26, 33, 46, 54, 61]. Among common approaches are correspondence-based methods, which learn to identify precise alignments between input data and CAD models of the objects [7, 27, 46]. Other methods use template-based strategies, where the model selects the most similar pose-labeled template from a predefined set of examples [25, 35, 52]. Additionally, regression-based approaches directly predict object poses from the learned object-specific features [12, 19, 26, 42].", + "bbox": [ + 511, + 272, + 903, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These instance-specific techniques offer high precision but typically require retraining for new object instances, limiting their generalizability. This limitation has led to the exploration of category-level methods, which generalize within specific object categories and allow estimation of unseen objects in known categories [4, 47, 53, 55, 59]. For example, SecondPose [4] enhances category-level 6D pose estimation by integrating object-specific geometric features with DINOv2 SE(3)-consistent semantic priors, effectively addressing intra-class shape variation. However, these methods struggle outside their target categories, while our approach enables pose estimation without relying on such constraints, allowing broader applications.", + "bbox": [ + 511, + 455, + 903, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unseen Object Pose Estimation. To enhance pose estimation flexibility, many methods aim to generalize across new object instances without object-specific training [22, 31, 37, 39, 43]. Approaches for this task can be broadly divided into manual reference view-based and CAD model-based methods. (i) Reference view-based methods use multi-view images with known poses as reference data for pose estimation [14, 31, 43, 51]. For example, OnePose [14, 51] reconstructs 3D object point clouds from posed RGB images, establishing 2D-3D correspondences to solve 6D poses. In a different approach, Nope [38], adopting the perspective of generating new views, trains models to predict discriminative embeddings for novel object perspectives. (ii) CAD model-based approaches leverage 3D object model to facilitate pose estimation. Some methods match features between the CAD model and the query image [20, 28, 60],", + "bbox": [ + 511, + 659, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16819", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f01f662c5b4ced6619d7971ccf11c13a4a43baacff99389e263fbea499a8d11e.jpg", + "image_caption": [ + "Figure 2. Overview of the 6D Pose Estimation Process in Pos3R. (1) Template Rendering (Sec. 3.2.1): A CAD model is used to generate a set of forty templates, each representing a unique orientation that covers both in-plane and out-of-plane rotations. This is achieved by positioning the camera at the vertices of a cube around the object. (2) Image Matching (Sec. 3.2.2): Given a test segment from CNOS [37], we establish 2D-2D correspondences between the test image and each rendered template. We leverage the 3D-consistent features generated by MASt3R [24] for accurate matching and select the best template based on matching quality. Each template also includes a 3D object coordinate map that records the corresponding 3D points. (3) Pose Fitting (Sec. 3.2.3): Using the selected correspondences, we apply the PnP-RANSAC algorithm [23] to obtain a final pose that aligns the object with its observed position in the scene." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 883, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while others employ template matching to find the closest initial pose, refining it further using specialized refiners for higher accuracy [22, 34, 39]. For example, GigaPose [39] introduces an efficient two-network system: one network for retrieving template views (out-of-plane rotation) and another for estimating the remaining degrees of freedom (in-plane rotation and 3D translation). This separation reduces computational costs, contrasting with methods like MegaPose [22], which applies a single network to every possible test crop-template pair. Our work builds on the CAD model-based approach, aligning with recent BOP challenge protocols [17]. We focus on a training-free pipeline for unseen object pose estimation, a direction distinct from methods that require extensive training of task-specific networks.", + "bbox": [ + 88, + 439, + 483, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training-Free Object Pose Estimation. Traditionally, 6D object pose estimation is achieved by establishing 3D-to-2D correspondences followed by a $\\mathrm{PnP}$ algorithm [11, 13, 44, 61]. Recently, leveraging features from foundation models has gained traction, especially with models like DINO, which offer spatial detail and semantic consistency for pose estimation [1, 58]. FoundPose [41] uses DINOv2 descriptors to bridge synthetic and real data domains, providing strong performance for symmetric objects with RGB-only inputs. FreeZe integrates 3D point descriptors from GeDi [45] with image features from DINOv2 [40] for RGB-D pose estimation. By leveraging the 3D consistent feature of MAST3R to produce dense correspondences, our work explores training-free, unseen object pose estimation with RGB-only inputs, enabling effective pose estimation across diverse object categories without task-specific training.", + "bbox": [ + 89, + 657, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.6D Pose Estimation with Pos3R", + "text_level": 1, + "bbox": [ + 511, + 438, + 800, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Definition", + "text_level": 1, + "bbox": [ + 511, + 463, + 666, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a 3D model of a query object $\\mathbf{Q}$ and an RGB image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ containing $\\mathbf{Q}$ , the task is to estimate the 6D pose of $\\mathbf{Q}$ relative to the camera's reference frame, with known intrinsics $\\mathbf{K}$ . Specifically, we aim to determine the 6DoF transformation $\\mathbf{T} = (\\mathbf{R}, \\mathbf{t})$ in 3D space, where $\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}$ represents the rotation matrix and $\\mathbf{t} \\in \\mathbb{R}^3$ denotes the translation vector. The segmented object region $\\mathbf{I}_m = \\mathbf{M} \\odot \\mathbf{I}$ is created by element-wise multiplication of the binary segmentation mask $\\mathbf{M}$ and $\\mathbf{I}$ , isolating the visible part of $\\mathbf{Q}$ . We note that $\\mathbf{Q}$ may be partially occluded.", + "bbox": [ + 511, + 486, + 906, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work leverages the 3D foundation model MASt3R [24] to tackle model-based 6D pose estimation for unseen objects. We introduce Pos3R, a training-free, RGB-only method that effectively estimates 6D poses without object- or task-specific training.", + "bbox": [ + 511, + 638, + 906, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Training-Free Pipeline", + "text_level": 1, + "bbox": [ + 511, + 724, + 723, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. Following the standard model-based pipeline for 6D pose estimation of unseen objects [2, 39, 41], Pos3R comprises two components: object detection and pose estimation. We keep each component frozen, avoiding any object- or task-specific training.", + "bbox": [ + 511, + 748, + 905, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For object detection, we use CNOS [37] to generate segmentation masks and object identities for each target instance, enabling localization of the target segment $\\mathbf{I}_m$ within the RGB image $\\mathbf{I}$ . As a default method in the BOP challenge for segmenting unseen objects, CNOS requires", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16820", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "only 3D models for onboarding and does not depend on additional data or object-specific training.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For 6D pose estimation, we illustrate three steps in Fig. 2. Specifically, given a set of templates rendered from a textured CAD model, Pos3R utilizes the 3D foundation model MAST3R [24] to extract features from both the target segment $\\mathbf{I}_m$ and each template, enabling estimation of the 6D transformation $\\mathbf{T} = (\\mathbf{R},\\mathbf{t})$ through 2D-3D correspondences using the PnP-RANSAC algorithm.", + "bbox": [ + 89, + 122, + 483, + 228 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Template Rendering", + "text_level": 1, + "bbox": [ + 89, + 238, + 279, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using a textured 3D CAD model, we render templates of the object from various orientations. The rendering process follows standard rasterization methods as described in [39], with a black background and fixed lighting. The rendering camera is configured with the same intrinsic parameters, $\\mathbf{K}$ , as the test camera, and the rendered templates match the size of the test image $\\mathbf{I}$ . The object remains centered in each template. Additionally, the 3D locations in the coordinate space of the 3D CAD model corresponding to each pixel in the rendered template are recorded, enabling the establishment of 2D-3D correspondences.", + "bbox": [ + 89, + 257, + 483, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A critical component of Pos3R is establishing correspondences between pixels in the target segment and the template. Accurate pose estimation requires addressing both in-plane and out-of-plane rotations, as objects are frequently viewed from diverse orientations. In-plane rotations occur within the image plane, such as when an object rotates around the camera's line of sight. Out-of-plane rotations, on the other hand, involve 3D orientation changes—such as tilting or turning—that significantly alter an object's appearance due to perspective shifts. Existing methods tackle this challenge through extensive template libraries and specialized selection mechanisms. For example, MegaPose [22] uses hundreds of templates to capture a broad range of object poses, but maintaining such a large template library necessitates a dedicated selection network.", + "bbox": [ + 89, + 425, + 483, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Template Configuration. To reduce reliance on extensive template libraries, Pos3R leverages the 3D foundation model MASt3R, which generates 3D-consistent features across viewpoints. This enables Pos3R to handle out-of-plane rotations effectively without requiring hundreds of templates or complex selection mechanisms. We use a set of eight base templates, denoted by $\\{\\mathbf{I}_i\\}_{i=1}^8$ , covering essential orientations. They are positioned at the vertices of a cube centered around the CAD model, effectively capturing out-of-plane rotations.", + "bbox": [ + 89, + 657, + 482, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further address ambiguities from in-plane (axial) rotations, which can impact correspondence quality, we apply controlled rotational variations to each base template. For each template $\\mathbf{I}_i$ , we generate $T$ rotations around the camera's principal axis, each rotation defined by an angle $\\theta_k = \\frac{2\\pi k}{T}$ , where $k = 0,\\dots,T - 1$ , covering the full", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$360^{\\circ}$ range. In our experiments, we set $T = 5$ to balance efficiency and accuracy, yielding a set of rotationally-augmented templates $\\{\\mathbf{I}_{i,k}\\}_{k=1}^{5}$ . In the following, we discuss the template selection process in detail.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Image Matching", + "text_level": 1, + "bbox": [ + 511, + 159, + 673, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MASt3R as an Image Matcher. Our approach builds upon MASt3R [24], a model for joint local 3D reconstruction and pixelwise matching between two input images $\\mathbf{I}_a$ and $\\mathbf{I}_b \\in \\mathbb{R}^{H \\times W \\times 3}$ . Conceptually, MASt3R operates as a mapping function $\\mathbf{f}(\\mathbf{I}_a, \\mathbf{I}_b) = \\mathrm{Dec}(\\mathrm{Enc}(\\mathbf{I}_a), \\mathrm{Enc}(\\mathbf{I}_b))$ , where the encoder $\\mathrm{Enc}(\\mathbf{I}) \\to \\mathbf{F}$ is a Siamese Vision Transformer (ViT) that processes image $\\mathbf{I}$ into token vectors of size $N \\times d$ , with $N = w \\times h$ , yielding $\\mathbf{F} \\in \\mathbb{R}^{N \\times d}$ . The decoder, $\\mathrm{Dec}(\\mathbf{F}_a, \\mathbf{F}_b)$ , employs twin ViT modules that produce pixelwise pointmaps $\\mathbf{X}$ , local feature maps $\\mathbf{D}$ for each image.", + "bbox": [ + 511, + 178, + 903, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using these local feature representations, correspondences between images are identified via the fastNN algorithm [24]. It efficiently establishes reciprocal matches between the feature maps $\\mathbf{D}_a$ and $\\mathbf{D}_b$ by initially seeding points across a uniform pixel grid and iteratively refining these seeds to form high-quality mutual correspondences. Through this process, fastNN accurately aligns key features across images. The resulting reciprocal pixel pairs between $\\mathbf{I}_a$ and $\\mathbf{I}_b$ are represented as $M_{a,b} = \\{(y_a^c,y_b^c)\\}_{c = 1}^{|M_{a,b}|}$ , where $\\mathbf{y}_a^c$ and $\\mathbf{y}_b^c\\in \\mathbb{N}^2$ denote the coordinates of matched pixels in each respective image.", + "bbox": [ + 511, + 330, + 905, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarity-Based Template Selection. Unlike methods that require hundreds of templates [22, 36, 41], Pos3R simplifies template selection by needing only forty templates, significantly reducing the complexity of the selection process. Rather than relying on a trained template selection network, we employ a straightforward, training-free approach based on the similarity of matched correspondences. Specifically, given a target segment $\\mathbf{I}_m$ and a set of eight base templates $\\{\\mathbf{I}_i\\}_{i=1}^8$ , each augmented with rotational variations $\\{\\mathbf{I}_{i,k}\\}_{k=1}^5$ , we identify the most similar template for pose estimation by calculating correspondences between the target segment and each template.", + "bbox": [ + 511, + 503, + 905, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each rotationally-augmented template $\\mathbf{I}_{i,k}$ , we obtain reciprocal pixel pairs between $\\mathbf{I}_m$ and $\\mathbf{I}_{i,k}$ , represented as: $M_{m,i,k} = \\{(y_m^p, y_{i,k}^p)\\}_{p=1}^{|M_{m,i,k}|}$ , where $M_{m,i,k}$ is the set of all reciprocal pixel pairs between $\\mathbf{I}_m$ and $\\mathbf{I}_{i,k}$ , and $\\mathbf{y}_m^p$ and $\\mathbf{y}_{i,k}^p \\in \\mathbb{N}^2$ denote the coordinates of matched pixels in the target segment and the template variant, respectively.", + "bbox": [ + 511, + 686, + 905, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each matched pair $(\\mathbf{y}_m^p, \\mathbf{y}_{i,k}^p)$ in $M_{m,i,k}$ , we retrieve the corresponding local features from the feature maps $\\mathbf{D}_m$ and $\\mathbf{D}_{i,k}$ generated by MASt3R. Specifically, we obtain the feature vector $\\mathbf{f}_m^p$ at coordinate $\\mathbf{y}_m^p$ in the target segment from $\\mathbf{D}_m$ , and the feature vector $\\mathbf{f}_{i,k}^p$ at coordinate $\\mathbf{y}_{i,k}^p$ in the template from $\\mathbf{D}_{i,k}$ . We compute the feature similarity for each matched pair as: $S(\\mathbf{f}_m^p, \\mathbf{f}_{i,k}^p) = \\mathbf{f}_m^p \\cdot \\mathbf{f}_{i,k}^p$ , where $\\cdot$ denotes the dot product.", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16821", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To calculate an overall similarity score between the target segment $\\mathbf{I}_m$ and each template variant $\\mathbf{I}_{i,k}$ , we aggregate the similarity scores across all matched pairs as follows:", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} \\left(\\mathbf {I} _ {m}, \\mathbf {I} _ {i, k}\\right) = \\sum_ {p = 1} ^ {\\left| M _ {m, i, k} \\right|} S \\left(\\mathbf {f} _ {m} ^ {p}, \\mathbf {f} _ {i, k} ^ {p}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 148, + 482, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{m,i,k}$ is the set of all matched pairs between $\\mathbf{I}_m$ and $\\mathbf{I}_{i,k}$ . After computing $\\mathrm{sim}(\\mathbf{I}_m,\\mathbf{I}_{i,k})$ for each pair, we select the template with the highest similarity score:", + "bbox": [ + 89, + 203, + 483, + 250 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(i _ {\\text {o p t}}, k _ {\\text {o p t}}\\right) = \\underset {i \\in \\{1, \\dots , 8 \\}, k \\in \\{1, \\dots , 5 \\}} {\\arg \\max } \\operatorname {s i m} \\left(\\mathbf {I} _ {m}, \\mathbf {I} _ {i, k}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 261, + 482, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The selected template $\\mathbf{I}_{i_{\\mathrm{opt}},k_{\\mathrm{opt}}}$ is then used as the closest match to the target segment for the pose estimation process.", + "bbox": [ + 89, + 299, + 483, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Pose Fitting", + "text_level": 1, + "bbox": [ + 89, + 337, + 220, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After selecting the suitable template $\\mathbf{I}_{i_{\\mathrm{opt}},k_{\\mathrm{opt}}}$ , we proceed to estimate pose $\\mathbf{T}_m = (\\mathbf{R}_m,\\mathbf{t}_m)$ , where $\\mathbf{R}_m$ is a 3D rotation matrix and $\\mathbf{t}_m$ is a 3D translation vector that transforms the object from model space to camera space. It relies on a set of 2D-3D correspondences $\\mathcal{C}_{t_{\\mathrm{final}}} = \\{(\\mathbf{y}_m^j,\\mathbf{P}^j)\\}_{j = 1}^{|M|}$ where $\\mathbf{y}_m^j\\in \\mathbb{R}^2$ represents the coordinates of matched pixels in the target segment $\\mathbf{I}_m$ , $\\mathbf{P}^j\\in \\mathbb{R}^3$ denotes the corresponding 3D points in model space from the selected template, and there are $|M|$ pairs. To determine $\\mathbf{T}_m$ , we solve the Perspective-n-Point (PnP) problem, which minimizes the reprojection error:", + "bbox": [ + 89, + 356, + 483, + 526 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\mathbf {R} _ {m}, \\mathbf {t} _ {m}} {\\arg \\min } \\sum_ {j = 1} ^ {| M |} \\left\\| \\mathbf {y} _ {m} ^ {j} - \\pi \\left(\\mathbf {R} _ {m} \\mathbf {P} ^ {j} + \\mathbf {t} _ {m}\\right) \\right\\| ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 537, + 482, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi$ is the projection function that maps 3D points to 2D image points according to the camera intrinsics.", + "bbox": [ + 89, + 593, + 482, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance robustness against outliers, we employ the Efficient PnP (EPnP) algorithm [23] combined with a RANSAC-based fitting strategy [11]. In this approach, PnP is applied iteratively to random subsets of four correspondences from $\\mathcal{C}_{t_{\\mathrm{final}}}$ , generating multiple pose hypotheses. For each hypothesis, we count the number of inliers, defined as correspondences where the reprojection error falls below a predefined threshold $\\epsilon$ :", + "bbox": [ + 89, + 623, + 483, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {i n l i e r s} = \\left| \\left\\{j: \\left\\| \\mathbf {y} _ {m} ^ {j} - \\pi \\left(\\mathbf {R} _ {m} \\mathbf {P} ^ {j} + \\mathbf {t} _ {m}\\right) \\right\\| < \\epsilon \\right\\} \\right|. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 755, + 482, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The hypothesis with the highest inlier count is selected as the final coarse pose estimate $\\mathbf{T}_m$ .", + "bbox": [ + 89, + 785, + 483, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 830, + 223, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first outline the experimental setup (Section 4.1). We then evaluate our method's performance in comparison with previous approaches on the seven core", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "datasets of the BOP challenge [17], examining accuracy, runtime efficiency, and effectiveness when using predicted 3D models (Section 4.2). This evaluation highlights the strengths and unique contributions of our approach. Lastly, we present an ablation study to investigate the impact of various configurations in our method (Section 4.3).", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 194, + 702, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Datasets. We evaluate our approach on the seven core datasets of the BOP challenge [17]: LineMod Occlusion (LM-O) [3], T-LESS [15], TUD-L [16], IC-BIN [8], ITODD [9], HomebrewedDB (HB)[21], and YCB-Video (YCB-V) [57]. Together, these datasets feature 132 distinct objects and 19,048 testing instances, each in complex, cluttered scenes with partial occlusions. Table 1 provides the instance count for each dataset. It is worth noting that the unseen object pose estimation task remains challenging, with significant room for improvement.", + "bbox": [ + 511, + 215, + 906, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use the BOP evaluation protocol [17] for 6D object localization, which assesses pose accuracy with three error metrics: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). VSD evaluates only the visible part of the object to handle ambiguities, MSSD measures 3D surface deviation with global symmetries, and MSPD assesses perceivable deviation using object symmetries in projection. A pose is considered correct for a metric $e$ if $e < \\theta_e$ , with $\\theta_e$ as the correctness threshold. The Average Recall (AR) for each metric $e$ , denoted $\\mathrm{AR}_e$ , is the mean Recall over various $\\theta_e$ thresholds, and for VSD, multiple misalignment tolerances $\\tau$ . The overall AR score is the average of the three metrics: $\\mathrm{AR} = (\\mathrm{AR}_{\\mathrm{VSD}} + \\mathrm{AR}_{\\mathrm{MSSD}} + \\mathrm{AR}_{\\mathrm{MSPD}}) / 3$ .", + "bbox": [ + 511, + 375, + 906, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Refinement. To show Pos3R can integrate with render-and-compare refinement techniques, we apply the refinement method from MegaPose [22] to our result. Given an input image and an estimated pose, the refiner predicts the relative transformation between the initial and ground-truth pose. The refiner is trained on a large-scale dataset with several well-designed techniques [22]. We use the similarity score Eq. 1 to select the top-5 templates and each test crop-template pair gives a pose with EPnP. Then, we use the refiner to update these five pose hypotheses.", + "bbox": [ + 511, + 608, + 905, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison With the State of the Art", + "text_level": 1, + "bbox": [ + 511, + 772, + 839, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare Pos3R with both training-free and training-based methods. The training-free methods include Found-Pose [41] and ZS6D [2], while the training-based methods include MegaPose [22], GigaPose [39], OSOP [48], and GenFlow [34]. For pose refinement, we apply the refiner from MegaPose [22]. We use the publicly available code from [39] with the default hyperparameters.", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16822", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/240f26f9858eeda1cbbd3b35e523c11fb1db5ea9ef6d2e0b53fcb0127cf4f8f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#MethodTraining-FreeRefinementDatasets (num. instances)MeanTime
LM-OT-LESSTUD-LIC-BINITODDHBYCB-V
144564236001786304116304123
Coarse Pose Estimation:
1GigaPose [39]X-29.927.330.223.118.834.829.027.60.9
2GenFlow [34]X-25.021.530.016.815.428.327.723.53.8
3MegaPose [22]X-22.917.725.815.210.825.128.120.815.5
4OSOP [48]X-31.2----49.233.2--
5ZS6D [2]✓-29.821.0----32.4--
6FoundPose [41]✓-39.633.846.723.920.450.845.237.21.7
7Baseline✓-30.523.643.221.415.442.549.332.30.7
7Pos3R (Ours)✓-32.331.547.333.125.153.753.939.51.4
With Pose Refinement (5 hypotheses):
8FoundPoseXMegaPose58.654.965.744.436.170.367.356.811.2
9GigaPoseXMegaPose59.957.064.546.739.772.266.357.97.3
10GenFlowXGenFlow56.352.368.445.339.573.963.357.120.9
11MegaPoseXMegaPose56.050.768.441.433.870.462.154.747.4
12Pos3R (ours)XMegaPose56.557.566.846.037.570.166.457.38.0
", + "bbox": [ + 89, + 88, + 906, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance Comparison on the Seven Datasets of BOP. This table reports the Average Recall (AR) scores per dataset, the mean AR score across all datasets, and the time required to estimate poses for all objects in an image (in seconds). The runtime data of other methods are sourced from FoundPose [41]. The upper section lists methods for coarse pose estimation without refinement, while the lower section presents methods that incorporate a refinement stage using multiple pose hypotheses, reporting the best refined pose. Methods without task-specific training are marked with a green check mark $(\\checkmark)$ . In the coarse estimation category, Pos3R achieves the highest AR scores on several datasets and demonstrates good overall generalization with a competitive mean AR score. While refinement-based methods yield slightly higher accuracy in certain cases, Pos3R remains a robust and efficient choice for coarse pose estimation and performs comparably to top refined methods when combined with a refinement stage. The best results in the coarse estimation section are highlighted in green, and the best results in the refinement section are highlighted in blue. The second-best results are underlined.", + "bbox": [ + 89, + 395, + 906, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 provides a detailed comparison of Pos3R (Pos3R) with other leading 6D pose estimation methods across the seven primary datasets in the BOP challenge. The upper portion of the table showcases results for coarse pose estimation without refinement, while the lower portion highlights performance after incorporating a pose refinement step with five hypotheses. It is important to note that pose refinement involves task-specific training on extensive pose datasets, making it non-training-free [22].", + "bbox": [ + 89, + 537, + 482, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the coarse pose estimation category, Pos3R demonstrates superior performance over other training-free methods, including FoundPose and ZS6D, achieving the highest mean accuracy and consistently strong results across most datasets. For instance, Pos3R attains the highest AR (Average Recall) on the TUD-L, HB, and YCB-V datasets, along with a mean AR of 39.5. This high performance across varied datasets reflects Pos3R's robust generalization capabilities for different object poses and scenarios. Additionally, with a runtime of only 1.4 seconds, Pos3R stands out as an efficient alternative to methods that either require substantial computational resources or are not training-free.", + "bbox": [ + 89, + 681, + 482, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pos3R remains competitive after pose refinement using MegaPose's refine. Although Pos3R is not inher", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ently designed for refinement, it achieves comparable accuracy to MegaPose and other refined methods across several datasets. These findings underscore the adaptability of Pos3R, as it provides accurate initial pose estimates and benefits from additional refinement techniques—demonstrating its potential for scalable, practical, and flexible applications in 6D pose estimation.", + "bbox": [ + 511, + 537, + 903, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, we observe that Pos3R faces challenges in achieving competitive accuracy on the LM-O dataset, where significant occlusion poses a major difficulty. Occlusion disrupts the initial image matching process, often resulting in suboptimal pose estimates that refinement methods struggle to correct. FoundPose [41] addresses this issue by employing a bag-of-words strategy for template selection, which has shown greater resilience to occlusion. To improve robustness in such scenarios, incorporating the local contrastive learning approach used by GigaPose [39] could be a promising direction. This method is specifically designed to enhance feature matching in occluded regions, and we consider this an area for future exploration.", + "bbox": [ + 511, + 643, + 906, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Component Analysis", + "text_level": 1, + "bbox": [ + 511, + 848, + 709, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose Estimation Using Predicted 3D Models. Model-based methods for unseen object pose estimation typically", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16823", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3f42a79633c2fa19bb7a16d6447db170c1fc6b55d9c77d494efffd2d142205c5.jpg", + "table_caption": [ + "Table 2. Comparison with Predicted 3D Models on LM-O [3]. Average Recall is reported. All methods use 3D models predicted by Wonder3D [32]. Results with pose refinement, applied to the generated 3D models, are also included for comparison." + ], + "table_footnote": [], + "table_body": "
#MethodSingle Image-to-3DGT 3D model w/o refinement
CoarseRefined
1MegaPose [22]15.425.222.7
2GigaPose [39]17.627.229.4
3Pos3R19.528.532.3
", + "bbox": [ + 91, + 155, + 483, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b1f91f54bc77e1fc3af0bbfd20e32af7bf2e860c1382c3e8bfe94551275a14c.jpg", + "image_caption": [ + "Figure 3. Rendered Images and 3D Coordinate Maps from Wonder3D and CAD Model. The top two rows show rendered images and 3D coordinate maps from Wonder3D [32], while the bottom two rows show those from the real CAD model." + ], + "image_footnote": [], + "bbox": [ + 112, + 234, + 467, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "require accurate, textured 3D models. To relax this requirement, we follow the approach in [39] and use 3D models predicted by Wonder3D [32] from a single reference image. The predicted 3D models use the same object coordinate system as the ground truth (GT) frame, ensuring aligned axes. We then assess the performance of our method using these reconstructed models in place of the high-fidelity CAD models provided by the dataset. For a fair comparison, we adopt the same settings as [39], including the reference image for Wonder3D and post-processing steps, and report results on the LM-O dataset [3]. Examples of rendered views from generated 3D model are shown in Fig. 3. As shown in Table 2, Pos3R achieves higher AR scores than both MegaPose [22] and GigaPose [39], both before and after pose refinement, showing the effectiveness of Pos3R.", + "bbox": [ + 88, + 501, + 482, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Template Selection Technique. In our experiments, we use the similarity of correspondence matches (Eq. 1) as the primary method for template selection. Additionally, we consider two alternative approaches: 1) Inliers: selecting templates based on the number of inliers (Eq. 4); 2) Confidence: leveraging the 3D pointmaps and confidence maps provided by MASt3R for each target segment-template pair [24], where the average confidence score is used for selection. As illustrated in Fig. 4, template selection based on inliers proves effective, while selection based on similarity score yields the best performance.", + "bbox": [ + 89, + 734, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison of Pose Estimation Method. We evaluate our proposed method (Pos3R), a variant without axial rotation augmentation ( $w/o$ axial rotation), a variant that use face centers as base templates, and a version that replaces the 3D foundation model MASt3R with the 2D foundation model DINOv2. We also replace MASt3R with another dense image matcher RoMa [10]. Pos3R achieves the highest accuracy across all datasets, highlighting the importance of axial rotation handling and the advantage of using 3D foundation models over DINOv2.", + "bbox": [ + 511, + 88, + 906, + 214 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/be0c224017e3467454bfb11da8ede080c9ca31e6a29e45c743397bf4fa2fd40b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#MethodLM-OTUD-LYCB-VT-LessIC-BINITODDHB
1Pos3R32.347.353.931.533.125.153.7
2w/o Ax. Rot.30.543.249.323.621.415.442.5
3Face Center31.545.553.328.029.422.452.5
4DINOv2 (L11)18.525.022.514.215.614.132.1
5DINOv2 (L9)20.526.024.015.616.614.832.9
6RoMa [10]20.825.032.016.217.015.525.1
", + "bbox": [ + 514, + 224, + 906, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8a399f8d81c565c2614b700e53b5e13b756594a42fcc14314d522266eed4fc2.jpg", + "image_caption": [ + "Figure 4. Comparison of Template Selection Techniques. We compare template selection methods based on confidence, inliers, and similarity score. The similarity score consistently achieves the highest average recall (AR) across all datasets." + ], + "image_footnote": [], + "bbox": [ + 514, + 335, + 903, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of In-Plane Rotation and 3D Consistency on Pose Estimation. Table 3 evaluates the impact of incorporating in-plane (axial) rotations and 3D-consistent features on pose estimation performance across seven datasets. Our method, Pos3R (row 1), achieves the highest accuracy on all datasets by combining controlled in-plane rotations with the 3D-consistent features provided by the MASt3R foundation model. In contrast, row 2 presents results without in-plane rotation, where only eight base templates are used to account for out-of-plane rotations. The performance drops considerably, highlighting the importance of addressing in-plane rotations for improving correspondence quality and overall pose accuracy. Rows 4 and 5 present a further variation in which we replace MASt3R with the 2D foundation model DINOv2. Despite following the same pipeline (e.g., using the same 40 templates), this substitution leads to a notable performance decline across all datasets. Additionally, in row 6, replacing MASt3R with the dense image matcher RoMa [10] also results in lower accuracy. We speculate that MASt3R's improved performance stems from its large-scale and 3D-aware training, which may enhance its ability to produce features that are well-suited for pose estimation. The above analysis highlights the importance of incorporating both in-plane rotations and 3D-consistent features to achieve robust and accurate 6D pose estimation.", + "bbox": [ + 511, + 523, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16824", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7709a89e402e4f20ffd30465c93613c0c1265eb2787a74991c4d2fabd9e091d0.jpg", + "image_caption": [ + "YCBV" + ], + "image_footnote": [], + "bbox": [ + 125, + 88, + 315, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/220a39336f50664e06af3f63b1e46c4395e6804d30f63768ebc87830171bab76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 88, + 506, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e242a91145d8e673b8c55d71fad0250ed17726fe09f2b109fd37c947f07f2af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 88, + 702, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4746bd19ad495afe185abe8da093ae570f45bf15cb0a5d3543a578782745e87a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 88, + 895, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5cf411d0aa5e04f1b209a24cd092e06d026ab70dd0b05339f0a520da4d2e447a.jpg", + "image_caption": [ + "LM-O" + ], + "image_footnote": [], + "bbox": [ + 127, + 203, + 313, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c342bd184f62cc991d6dedb7d2227046431709fdd4e5aa68b1d2fa38b0feb902.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 203, + 506, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7436092b2d84476e8a73dfcffd2a99e017986aa1e0c9f741e413f8320f63fbff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 203, + 702, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8265ead8de098df5edfe4ff46a2c706a113a1fdcb1c800be96bdd5107b7f4e1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 203, + 895, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c82b945bb82ddc732ddc98923c84d1a744160525024a5ee617d86bffa4534cc6.jpg", + "image_caption": [ + "IC-BIN" + ], + "image_footnote": [], + "bbox": [ + 125, + 316, + 313, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8e6f7634359499e72ef07d103faa2a6d91232dbf7c85e206cdaad2efb572a864.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 318, + 506, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ae717336c2440f9995024939bd955d98cffe0d138835cee9007bdb6299d3a45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 318, + 702, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/273a492fcb79ed9018c185b7246d278cf98e929f8940440728e063a12db6afe3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 318, + 895, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/793280ee2a1334859367abb45b377c5d6b5d2a33f3f4a28d98c86e4d362dfbbc.jpg", + "image_caption": [ + "T-LESS" + ], + "image_footnote": [], + "bbox": [ + 125, + 431, + 313, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/192e63f0b688d5785574f2e0044d73e42d63a37d3c30467c8a174be424d8eb34.jpg", + "image_caption": [ + "Figure 5. Qualitative Results of 6D Pose Estimation. Visualization of 6D object pose estimation of Pos3R across LM-O [3], T-LESS [15], IC-BIN [8], and YCB-V [57]. Pos3R effectively handles a variety of objects and maintains accuracy even in very crowded scenes (e.g., IC-BIN), where predictions remain reasonably robust. Furthermore, Pos3R adapts well to texture-less objects, as demonstrated in the T-LESS dataset. However, as marked by red crosses $(\\times)$ , heavy occlusion poses a challenge, with reduced accuracy in heavily occluded regions, illustrating both the strengths and limitations of our approach." + ], + "image_footnote": [], + "bbox": [ + 318, + 431, + 506, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4974464d2c96da5462c928acb4c403a46878563d75ef2be5403bdbcb2f495256.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 431, + 702, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c7339eb09b04f2cba68c29f2dfd067925e3c78f3f277ecd4cf39906a93f58019.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 431, + 895, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results of 6D Pose Estimation. Figure 5 shows pose estimation results of Pos3R across challenging datasets, including LM-O [3], T-LESS [15], IC-BIN [8], and YCB-V [57]. Pos3R demonstrates robust performance across diverse object types, sizes, and textures, handling complex environments effectively. In cluttered scenes like IC-BIN, Pos3R accurately localizes multiple, tightly packed objects. It also performs well on textureless objects in T-LESS, a challenge for vision-based methods due to minimal surface features. However, Figure 5 also reveals a limitation: in heavily occluded scenes, marked by red crosses, Pos3R cannot well handle. Reduced visibility of occluded objects results in deviations in the predicted bounding boxes. This limitation suggests future directions for improvement, such as integrating occlusion-aware methods (e.g., GigaPose [39]) or leveraging multi-view information to enhance robustness in these challenging cases.", + "bbox": [ + 89, + 640, + 482, + 896 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 638, + 633, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work introduces Pos3R, a training-free, RGB-only framework for 6D pose estimation of unseen objects. By leveraging the 3D foundation model MASt3R, Pos3R generates robust, 3D-consistent features that effectively handle both in-plane and out-of-plane rotations. Without relying on extensive datasets or object-specific training, Pos3R establishes a strong baseline for training-free research. Using a minimal set of templates captured from eight cube vertex viewpoints with controlled rotational variations, Pos3R achieves accurate and efficient pose estimation. Experiments on the BOP challenge demonstrate that Pos3R outperforms other training-free methods in coarse pose estimation and achieves competitive results with refined methods when combined with MegaPose refinement. Future work will focus on incorporating occlusion-aware techniques to enhance robustness in occluded settings.", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16825", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We thank all reviewers and ACs for their constructive comments. This work was generously supported through research collaboration with RIOS Intelligent Machines.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 151, + 187, + 167 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 1, 3", + "[2] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 463-469. IEEE, 2024. 1, 3, 5, 6", + "[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 536-551. Springer, 2014. 5, 7, 8", + "[4] Yamei Chen, Yan Di, Guangyao Zhai, Fabian Manhardt, Chenyangguang Zhang, Ruida Zhang, Federico Tombari, Nassir Navab, and Benjamin Busam. Secondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9959-9969, 2024. 1, 2", + "[5] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 1, 2", + "[6] Francesco Di Felice, Salvatore D'Avella, Alberto Remus, Paolo Tripicchio, and Carlo Alberto Avizzano. One-shot imitation learning with graph neural networks for pick-and-place manipulation tasks. IEEE Robotics and Automation Letters, 2023. 1", + "[7] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6608-6617, 2020. 1, 2", + "[8] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3583-3592, 2016. 5, 8", + "[9] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec itodd-a dataset for 3d object recognition in industry. In Proceedings of the IEEE international conference on computer vision workshops, pages 2200-2208, 2017. 5", + "[10] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 7", + "[11] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to" + ], + "bbox": [ + 93, + 176, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 5", + "[12] Ge Gao, Mikko Lauri, Yulong Wang, Xiaolin Hu, Jianwei Zhang, and Simone Frintrop. 6d object pose regression via supervised learning on point clouds. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3643-3649. IEEE, 2020. 2", + "[13] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1580-1589, 2020. 3", + "[14] Xingyi He, Jiaming Sun, Yang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without cad models. Advances in Neural Information Processing Systems, 35:35103-35115, 2022. 2", + "[15] Tomás Hodan, Pavel Haluza, Štepan Obdržálek, Jiri Matas, Manolis Lourakis, and Xenophon Zabulis. T-less: An rgb-d dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 880-888. IEEE, 2017. 5, 8", + "[16] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. Bop: Benchmark for 6d object pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 19-34, 2018. 5", + "[17] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. BOP challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 1, 2, 3, 5", + "[18] Sabera Hoque, Shuxiang Xu, Ananda Maiti, Yuchen Wei, and Md Yasir Arafat. Deep learning for 6d pose estimation of objects—a case study for autonomous driving. Expert Systems with Applications, 223:119838, 2023. 1", + "[19] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2930-2939, 2020. 2", + "[20] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10095-10105, 2024. 2", + "[21] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 5", + "[22] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapose: 6d pose estimation of novel objects via render & compare. arXiv preprint arXiv:2212.06870, 2022. 2, 3, 4, 5, 6, 7" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16826", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 3, 5", + "[24] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 1, 2, 3, 4, 7", + "[25] Hongyang Li, Jiehong Lin, and Kui Jia. Dcl-net: Deep correspondence learning network for 6d pose estimation. In European Conference on Computer Vision, pages 369-385. Springer, 2022. 2", + "[26] Yuelong Li, Yafei Mao, Raja Bala, and Sunil Hadap. Mrcnet: 6-dof pose estimation with multiscale residual correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10476-10486, 2024. 1, 2", + "[27] Ruyi Lian and Haibin Ling. Checkerpose: Progressive dense keypoint localization for object pose estimation with graph neural network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14022-14033, 2023. 2", + "[28] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27906-27916, 2024. 2", + "[29] Jian Liu, Wei Sun, Chongpei Liu, Xing Zhang, and Qiang Fu. Robotic continuous grasping system by shape transformer-guided multiobject category-level 6-d pose estimation. IEEE Transactions on Industrial Informatics, 19 (11):11171-11181, 2023. 1", + "[30] Jian Liu, Wei Sun, Hui Yang, Zhiwen Zeng, Chongpei Liu, Jin Zheng, Xingyu Liu, Hossein Rahmani, Nicu Sebe, and Ajmal Mian. Deep learning-based object pose estimation: A comprehensive survey. arXiv preprint arXiv:2405.07801, 2024. 1", + "[31] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In European Conference on Computer Vision, pages 298-315. Springer, 2022. 2", + "[32] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 7", + "[33] Ningkai Mo, Wanshui Gan, Naoto Yokoya, and Shifeng Chen. Es6d: A computation efficient and symmetry-aware 6d pose regression framework. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6718-6727, 2022. 1, 2", + "[34] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. Genflow: Generalizable recurrent flow for 6d pose refinement of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10039-10049, 2024. 3, 5, 6" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Fengjun Mu, Rui Huang, Ao Luo, Xin Li, Jing Qiu, and Hong Cheng. Temporalfusion: Temporal motion reasoning with multi-frame fusion for 6d object pose estimation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5930-5936. IEEE, 2021. 2", + "[36] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6771-6780, 2022. 4", + "[37] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2134-2140, 2023. 2, 3", + "[38] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Yinlin Hu, Renaud Marlet, Mathieu Salzmann, and Vincent Lepetit. Nope: Novel object pose estimation from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17923-17932, 2024. 2", + "[39] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9903–9913, 2024. 2, 3, 4, 5, 6, 7, 8", + "[40] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 3", + "[41] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In European Conference on Computer Vision, pages 163-182. Springer, 2025. 1, 3, 4, 5, 6", + "[42] Jaewoo Park and Nam Ik Cho. Dprost: Dynamic projective spatial transformer network for 6d pose estimation. In European Conference on Computer Vision, pages 363-379. Springer, 2022. 2", + "[43] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10710-10719, 2020. 2", + "[44] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4561-4570, 2019. 3", + "[45] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud registration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3979-3985, 2022. 3", + "[46] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16827", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "poses of challenging objects without using depth. In Proceedings of the IEEE international conference on computer vision, pages 3828-3836, 2017. 1, 2", + "[47] Caner Sahin and Tae-Kyun Kim. Category-level 6d object pose recovery in depth images. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 2", + "[48] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6835-6844, 2022. 5, 6", + "[49] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In 2019 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227. IEEE, 2019. 1", + "[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 2", + "[51] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6825-6834, 2022. 2", + "[52] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 2", + "[53] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 530-546. Springer, 2020. 1, 2", + "[54] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 108-125. Springer, 2020. 1, 2", + "[55] Jiaxin Wei, Xibin Song, Weizhe Liu, Laurent Kneip, Hongdong Li, and Pan Ji. Rgb-based category-level object pose estimation via decoupled metric scale recovery. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2036-2042. IEEE, 2024. 1, 2", + "[56] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868-17879, 2024. 1", + "[57] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 5, 8" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 1, 3", + "[59] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2025. 1, 2", + "[60] Heng Zhao, Shenxing Wei, Dahu Shi, Wenming Tan, Zheyang Li, Ye Ren, Xing Wei, Yi Yang, and Shiliang Pu. Learning symmetry-aware geometry correspondences for 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14045-14054, 2023. 2", + "[61] Wanqing Zhao, Shaobo Zhang, Ziyu Guan, Wei Zhao, Jinye Peng, and Jianping Fan. Learning deep network for detecting 3d object keypoints and 6d poses. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 14134-14142, 2020. 1, 2, 3" + ], + "bbox": [ + 516, + 90, + 906, + 386 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16828", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/162f30eb-880f-42fb-95ca-56f10c93b7f6_model.json b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/162f30eb-880f-42fb-95ca-56f10c93b7f6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6dddcac3d42c83bc24aa603c81d59bec617803e7 --- /dev/null +++ b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/162f30eb-880f-42fb-95ca-56f10c93b7f6_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.204, + 0.131, + 0.795, + 0.154 + ], + "angle": 0, + "content": "Pos3R: 6D Pose Estimation for Unseen Objects Made Easy" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.18, + 0.77, + 0.2 + ], + "angle": 0, + "content": "Weijian Deng\\(^{1}\\) Dylan Campbell\\(^{1}\\) Chunyi Sun\\(^{1}\\) Jiahao Zhang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.199, + 0.746, + 0.216 + ], + "angle": 0, + "content": "Shubham Kanitkar\\(^{2}\\) Matthew E. Shaffer\\(^{2}\\) Stephen Gould\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.216, + 0.749, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Australian National University \\(^{2}\\)RIOS Intelligent Machines" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.234, + 0.85, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\{\\text {firstname.lastname } \\} @ \\text { anu.edu.au } ^{2}\\{\\text {firstname.lastname } \\} @ \\text { rios.ai }\\)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.486, + 0.637 + ], + "angle": 0, + "content": "Foundation models have significantly reduced the need for task-specific training, while also enhancing generalizability. However, state-of-the-art 6D pose estimators either require further training with pose supervision or neglect advances obtainable from 3D foundation models. The latter is a missed opportunity, since these models are better equipped to predict 3D-consistent features, which are of significant utility for the pose estimation task. To address this gap, we propose Pos3R, a method for estimating the 6D pose of any object from a single RGB image, making extensive use of a 3D reconstruction foundation model and requiring no additional training. We identify template selection as a particular bottleneck for existing methods that is significantly alleviated by the use of a 3D model, which can more easily distinguish between template poses than a 2D model. Despite its simplicity, Pos3R achieves competitive performance on the Benchmark for 6D Object Pose Estimation (BOP), matching or surpassing existing refinement-free methods. Additionally, Pos3R integrates seamlessly with render-and-compare refinement techniques, demonstrating adaptability for high-precision applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.224, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Six-dimensional (6D) object pose estimation—the task of determining the exact position and orientation of objects relative to a camera—is essential for applications in robotics, augmented reality, and autonomous systems. Reliable pose estimation enables critical tasks like object manipulation, grasping, and assembly, allowing these systems to interact effectively in complex and dynamic environments [6, 18, 29, 30, 49, 54]. Traditional approaches often rely on learning-based methods tailored to specific objects or categories, achieving high accuracy but struggling to generalize to new categories or unseen objects [4, 5, 7, 26, 33, 46, 47, 53-55, 59, 61]. This limitation is particularly problematic in dynamic, data-scarce environments where adaptability and flexibility are crucial." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.288, + 0.907, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.441 + ], + "angle": 0, + "content": "Figure 1. Illustration of In-Plane and Out-of-Plane Rotations with Correspondence Quality between DINOv2 and MASt3R." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.907, + 0.567 + ], + "angle": 0, + "content": "The left diagram illustrates in-plane rotations (within the image plane) and out-of-plane rotations (3D orientation changes) from a camera view. On the right, DINOv2 [40] (top row) and MASt3R [24] (bottom row) are compared in their handling of these rotations. DINOv2 exhibits sparse and inconsistent correspondences (shown by red lines), particularly under out-of-plane rotations, due to its 2D feature limitations. In contrast, MASt3R provides dense and stable correspondences across both types of rotations, reflecting its ability to produce 3D-consistent features." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.597, + 0.907, + 0.793 + ], + "angle": 0, + "content": "To overcome these limitations, there has been a recent shift toward model-based approaches that aim to generalize pose estimation to unseen objects without object-specific training [2, 17, 41, 56]. These typically rely on a two-stage pipeline: detecting and localizing objects within a scene, followed by a render-and-compare process that matches detected object regions to a set of template models. Building on this concept, recent works have sought to improve generalization through large-scale model training with significant attention to object diversity. For example, Foundation-Pose [56] uses synthetic training data combined with a large language model and contrastive learning, enhancing feature alignment across varied domains for robust generalization." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The recent emergence of training-free methods offers a promising alternative, enabling 6D pose estimation for unseen objects without object- or task-specific training. Foundation models like DINOv2 [40] have demonstrated strong zero-shot capabilities, capturing spatial and semantic details through learned features [1, 58]. FoundPose [41], for instance, leverages DINOv2 descriptors to bridge synthetic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16818" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "and real data, achieving competitive results on the BOP challenge [17]. These advancements underscore the potential of training-free frameworks in 6D pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.483, + 0.395 + ], + "angle": 0, + "content": "Building on this trend, we explore the strengths of 3D foundation models. Accurate pose estimation requires handling both in-plane and out-of-plane rotations, which are demonstrated in Fig. 1. In-plane rotations occur within the image plane, as when an object rotates around the camera's line of sight, while out-of-plane rotations involve changes in 3D orientation, such as tilting or turning, which significantly alters appearance due to perspective shifts. Addressing both types of rotation is essential for robust pose estimation, as objects often appear from varied angles. Although 2D foundation models like DINOv2 handle in-plane rotations well through training augmentations, they lack robustness to out-of-plane rotations due to their 2D limitations (shown in Fig. 1). In contrast, 3D foundation models, such as MASt3R [24], are specifically designed to produce 3D-consistent features across different viewpoints, enabling reliable feature alignment even with out-of-plane rotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.579 + ], + "angle": 0, + "content": "Motivated by this, we propose Pos3R, a training-free method for 6D pose estimation of unseen objects using only RGB inputs. Pos3R leverages the 3D foundation model MASt3R [24] for pose estimation without requiring additional training. The core of Pos3R lies in the image matching process between test crops and rendered templates from a CAD model. While common matchers like LoFTR [50] often struggle with synthetic-to-real matching due to domain gaps, MASt3R produces high-quality 2D correspondences in this case. These reliable matches enable Pos3R to establish the 2D-3D correspondences required for pose estimation using the PnP-RANSAC algorithm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.867 + ], + "angle": 0, + "content": "By harnessing MASt3R's ability to produce dense correspondences robust to viewpoint and illumination changes, we simplify template generation and achieve computational efficiency. Unlike existing methods that render hundreds of templates, we place camera positions to cover both in-plane and out-of-plane rotations. Specifically, we use eight base templates per object, positioned at the vertices of a cube centered on the CAD model, to capture out-of-plane rotations. For each base template, we generate five in-plane rotational templates along the object's principal axis, resulting in forty templates per object. To ensure efficient and accurate pose estimation, Pos3R dynamically selects the optimal rotation based on matching quality. This simple yet effective strategy maintains computational efficiency while achieving reliable pose estimation. Tested on seven diverse datasets in the BOP challenge, Pos3R demonstrates strong performance as a robust, scalable option for training-free 6D pose estimation of unseen objects, adaptable to render-and-compare refinement techniques." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.87, + 0.341, + 0.884 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- While existing methods rely on 2D foundation models" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "for unseen object pose estimation, we present Pos3R, a training-free method that leverages the 3D foundation model MAST3R [24] to improve robustness." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.228 + ], + "angle": 0, + "content": "- Leveraging MASt3R's robust dense correspondences, Pos3R uses only forty strategically placed templates per object to capture in-plane and out-of-plane rotations. A simple selection technique based on correspondence quality ensures accurate matching, achieving strong performance on the BOP challenge." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.246, + 0.655, + 0.261 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.453 + ], + "angle": 0, + "content": "Seen Object Pose Estimation. Instance-level object pose estimation refers to the task of estimating the poses of specific objects previously encountered during model training [5, 7, 26, 33, 46, 54, 61]. Among common approaches are correspondence-based methods, which learn to identify precise alignments between input data and CAD models of the objects [7, 27, 46]. Other methods use template-based strategies, where the model selects the most similar pose-labeled template from a predefined set of examples [25, 35, 52]. Additionally, regression-based approaches directly predict object poses from the learned object-specific features [12, 19, 26, 42]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.456, + 0.905, + 0.651 + ], + "angle": 0, + "content": "These instance-specific techniques offer high precision but typically require retraining for new object instances, limiting their generalizability. This limitation has led to the exploration of category-level methods, which generalize within specific object categories and allow estimation of unseen objects in known categories [4, 47, 53, 55, 59]. For example, SecondPose [4] enhances category-level 6D pose estimation by integrating object-specific geometric features with DINOv2 SE(3)-consistent semantic priors, effectively addressing intra-class shape variation. However, these methods struggle outside their target categories, while our approach enables pose estimation without relying on such constraints, allowing broader applications." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Unseen Object Pose Estimation. To enhance pose estimation flexibility, many methods aim to generalize across new object instances without object-specific training [22, 31, 37, 39, 43]. Approaches for this task can be broadly divided into manual reference view-based and CAD model-based methods. (i) Reference view-based methods use multi-view images with known poses as reference data for pose estimation [14, 31, 43, 51]. For example, OnePose [14, 51] reconstructs 3D object point clouds from posed RGB images, establishing 2D-3D correspondences to solve 6D poses. In a different approach, Nope [38], adopting the perspective of generating new views, trains models to predict discriminative embeddings for novel object perspectives. (ii) CAD model-based approaches leverage 3D object model to facilitate pose estimation. Some methods match features between the CAD model and the query image [20, 28, 60]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16819" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.087, + 0.885, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.316, + 0.908, + 0.415 + ], + "angle": 0, + "content": "Figure 2. Overview of the 6D Pose Estimation Process in Pos3R. (1) Template Rendering (Sec. 3.2.1): A CAD model is used to generate a set of forty templates, each representing a unique orientation that covers both in-plane and out-of-plane rotations. This is achieved by positioning the camera at the vertices of a cube around the object. (2) Image Matching (Sec. 3.2.2): Given a test segment from CNOS [37], we establish 2D-2D correspondences between the test image and each rendered template. We leverage the 3D-consistent features generated by MASt3R [24] for accurate matching and select the best template based on matching quality. Each template also includes a 3D object coordinate map that records the corresponding 3D points. (3) Pose Fitting (Sec. 3.2.3): Using the selected correspondences, we apply the PnP-RANSAC algorithm [23] to obtain a final pose that aligns the object with its observed position in the scene." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.44, + 0.484, + 0.651 + ], + "angle": 0, + "content": "while others employ template matching to find the closest initial pose, refining it further using specialized refiners for higher accuracy [22, 34, 39]. For example, GigaPose [39] introduces an efficient two-network system: one network for retrieving template views (out-of-plane rotation) and another for estimating the remaining degrees of freedom (in-plane rotation and 3D translation). This separation reduces computational costs, contrasting with methods like MegaPose [22], which applies a single network to every possible test crop-template pair. Our work builds on the CAD model-based approach, aligning with recent BOP challenge protocols [17]. We focus on a training-free pipeline for unseen object pose estimation, a direction distinct from methods that require extensive training of task-specific networks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Training-Free Object Pose Estimation. Traditionally, 6D object pose estimation is achieved by establishing 3D-to-2D correspondences followed by a \\( \\mathrm{PnP} \\) algorithm [11, 13, 44, 61]. Recently, leveraging features from foundation models has gained traction, especially with models like DINO, which offer spatial detail and semantic consistency for pose estimation [1, 58]. FoundPose [41] uses DINOv2 descriptors to bridge synthetic and real data domains, providing strong performance for symmetric objects with RGB-only inputs. FreeZe integrates 3D point descriptors from GeDi [45] with image features from DINOv2 [40] for RGB-D pose estimation. By leveraging the 3D consistent feature of MAST3R to produce dense correspondences, our work explores training-free, unseen object pose estimation with RGB-only inputs, enabling effective pose estimation across diverse object categories without task-specific training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.802, + 0.455 + ], + "angle": 0, + "content": "3.6D Pose Estimation with Pos3R" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.464, + 0.668, + 0.479 + ], + "angle": 0, + "content": "3.1. Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.638 + ], + "angle": 0, + "content": "Given a 3D model of a query object \\(\\mathbf{Q}\\) and an RGB image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\) containing \\(\\mathbf{Q}\\), the task is to estimate the 6D pose of \\(\\mathbf{Q}\\) relative to the camera's reference frame, with known intrinsics \\(\\mathbf{K}\\). Specifically, we aim to determine the 6DoF transformation \\(\\mathbf{T} = (\\mathbf{R}, \\mathbf{t})\\) in 3D space, where \\(\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}\\) represents the rotation matrix and \\(\\mathbf{t} \\in \\mathbb{R}^3\\) denotes the translation vector. The segmented object region \\(\\mathbf{I}_m = \\mathbf{M} \\odot \\mathbf{I}\\) is created by element-wise multiplication of the binary segmentation mask \\(\\mathbf{M}\\) and \\(\\mathbf{I}\\), isolating the visible part of \\(\\mathbf{Q}\\). We note that \\(\\mathbf{Q}\\) may be partially occluded." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.715 + ], + "angle": 0, + "content": "This work leverages the 3D foundation model MASt3R [24] to tackle model-based 6D pose estimation for unseen objects. We introduce Pos3R, a training-free, RGB-only method that effectively estimates 6D poses without object- or task-specific training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.725, + 0.743 + ], + "angle": 0, + "content": "3.2. Training-Free Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Overview. Following the standard model-based pipeline for 6D pose estimation of unseen objects [2, 39, 41], Pos3R comprises two components: object detection and pose estimation. We keep each component frozen, avoiding any object- or task-specific training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "For object detection, we use CNOS [37] to generate segmentation masks and object identities for each target instance, enabling localization of the target segment \\(\\mathbf{I}_m\\) within the RGB image \\(\\mathbf{I}\\). As a default method in the BOP challenge for segmenting unseen objects, CNOS requires" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16820" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "only 3D models for onboarding and does not depend on additional data or object-specific training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.484, + 0.229 + ], + "angle": 0, + "content": "For 6D pose estimation, we illustrate three steps in Fig. 2. Specifically, given a set of templates rendered from a textured CAD model, Pos3R utilizes the 3D foundation model MAST3R [24] to extract features from both the target segment \\(\\mathbf{I}_m\\) and each template, enabling estimation of the 6D transformation \\(\\mathbf{T} = (\\mathbf{R},\\mathbf{t})\\) through 2D-3D correspondences using the PnP-RANSAC algorithm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.239, + 0.281, + 0.254 + ], + "angle": 0, + "content": "3.2.1. Template Rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.424 + ], + "angle": 0, + "content": "Using a textured 3D CAD model, we render templates of the object from various orientations. The rendering process follows standard rasterization methods as described in [39], with a black background and fixed lighting. The rendering camera is configured with the same intrinsic parameters, \\(\\mathbf{K}\\), as the test camera, and the rendered templates match the size of the test image \\(\\mathbf{I}\\). The object remains centered in each template. Additionally, the 3D locations in the coordinate space of the 3D CAD model corresponding to each pixel in the rendered template are recorded, enabling the establishment of 2D-3D correspondences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.651 + ], + "angle": 0, + "content": "A critical component of Pos3R is establishing correspondences between pixels in the target segment and the template. Accurate pose estimation requires addressing both in-plane and out-of-plane rotations, as objects are frequently viewed from diverse orientations. In-plane rotations occur within the image plane, such as when an object rotates around the camera's line of sight. Out-of-plane rotations, on the other hand, involve 3D orientation changes—such as tilting or turning—that significantly alter an object's appearance due to perspective shifts. Existing methods tackle this challenge through extensive template libraries and specialized selection mechanisms. For example, MegaPose [22] uses hundreds of templates to capture a broad range of object poses, but maintaining such a large template library necessitates a dedicated selection network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Template Configuration. To reduce reliance on extensive template libraries, Pos3R leverages the 3D foundation model MASt3R, which generates 3D-consistent features across viewpoints. This enables Pos3R to handle out-of-plane rotations effectively without requiring hundreds of templates or complex selection mechanisms. We use a set of eight base templates, denoted by \\(\\{\\mathbf{I}_i\\}_{i=1}^8\\), covering essential orientations. They are positioned at the vertices of a cube centered around the CAD model, effectively capturing out-of-plane rotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.904 + ], + "angle": 0, + "content": "To further address ambiguities from in-plane (axial) rotations, which can impact correspondence quality, we apply controlled rotational variations to each base template. For each template \\(\\mathbf{I}_i\\), we generate \\(T\\) rotations around the camera's principal axis, each rotation defined by an angle \\(\\theta_k = \\frac{2\\pi k}{T}\\), where \\(k = 0,\\dots,T - 1\\), covering the full" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "\\(360^{\\circ}\\) range. In our experiments, we set \\(T = 5\\) to balance efficiency and accuracy, yielding a set of rotationally-augmented templates \\(\\{\\mathbf{I}_{i,k}\\}_{k=1}^{5}\\). In the following, we discuss the template selection process in detail." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.16, + 0.674, + 0.175 + ], + "angle": 0, + "content": "3.2.2. Image Matching" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.179, + 0.905, + 0.33 + ], + "angle": 0, + "content": "MASt3R as an Image Matcher. Our approach builds upon MASt3R [24], a model for joint local 3D reconstruction and pixelwise matching between two input images \\(\\mathbf{I}_a\\) and \\(\\mathbf{I}_b \\in \\mathbb{R}^{H \\times W \\times 3}\\). Conceptually, MASt3R operates as a mapping function \\(\\mathbf{f}(\\mathbf{I}_a, \\mathbf{I}_b) = \\mathrm{Dec}(\\mathrm{Enc}(\\mathbf{I}_a), \\mathrm{Enc}(\\mathbf{I}_b))\\), where the encoder \\(\\mathrm{Enc}(\\mathbf{I}) \\to \\mathbf{F}\\) is a Siamese Vision Transformer (ViT) that processes image \\(\\mathbf{I}\\) into token vectors of size \\(N \\times d\\), with \\(N = w \\times h\\), yielding \\(\\mathbf{F} \\in \\mathbb{R}^{N \\times d}\\). The decoder, \\(\\mathrm{Dec}(\\mathbf{F}_a, \\mathbf{F}_b)\\), employs twin ViT modules that produce pixelwise pointmaps \\(\\mathbf{X}\\), local feature maps \\(\\mathbf{D}\\) for each image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.331, + 0.906, + 0.501 + ], + "angle": 0, + "content": "Using these local feature representations, correspondences between images are identified via the fastNN algorithm [24]. It efficiently establishes reciprocal matches between the feature maps \\(\\mathbf{D}_a\\) and \\(\\mathbf{D}_b\\) by initially seeding points across a uniform pixel grid and iteratively refining these seeds to form high-quality mutual correspondences. Through this process, fastNN accurately aligns key features across images. The resulting reciprocal pixel pairs between \\(\\mathbf{I}_a\\) and \\(\\mathbf{I}_b\\) are represented as \\(M_{a,b} = \\{(y_a^c,y_b^c)\\}_{c = 1}^{|M_{a,b}|}\\), where \\(\\mathbf{y}_a^c\\) and \\(\\mathbf{y}_b^c\\in \\mathbb{N}^2\\) denote the coordinates of matched pixels in each respective image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.505, + 0.906, + 0.686 + ], + "angle": 0, + "content": "Similarity-Based Template Selection. Unlike methods that require hundreds of templates [22, 36, 41], Pos3R simplifies template selection by needing only forty templates, significantly reducing the complexity of the selection process. Rather than relying on a trained template selection network, we employ a straightforward, training-free approach based on the similarity of matched correspondences. Specifically, given a target segment \\(\\mathbf{I}_m\\) and a set of eight base templates \\(\\{\\mathbf{I}_i\\}_{i=1}^8\\), each augmented with rotational variations \\(\\{\\mathbf{I}_{i,k}\\}_{k=1}^5\\), we identify the most similar template for pose estimation by calculating correspondences between the target segment and each template." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.687, + 0.906, + 0.78 + ], + "angle": 0, + "content": "For each rotationally-augmented template \\(\\mathbf{I}_{i,k}\\), we obtain reciprocal pixel pairs between \\(\\mathbf{I}_m\\) and \\(\\mathbf{I}_{i,k}\\), represented as: \\(M_{m,i,k} = \\{(y_m^p, y_{i,k}^p)\\}_{p=1}^{|M_{m,i,k}|}\\), where \\(M_{m,i,k}\\) is the set of all reciprocal pixel pairs between \\(\\mathbf{I}_m\\) and \\(\\mathbf{I}_{i,k}\\), and \\(\\mathbf{y}_m^p\\) and \\(\\mathbf{y}_{i,k}^p \\in \\mathbb{N}^2\\) denote the coordinates of matched pixels in the target segment and the template variant, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For each matched pair \\((\\mathbf{y}_m^p, \\mathbf{y}_{i,k}^p)\\) in \\(M_{m,i,k}\\), we retrieve the corresponding local features from the feature maps \\(\\mathbf{D}_m\\) and \\(\\mathbf{D}_{i,k}\\) generated by MASt3R. Specifically, we obtain the feature vector \\(\\mathbf{f}_m^p\\) at coordinate \\(\\mathbf{y}_m^p\\) in the target segment from \\(\\mathbf{D}_m\\), and the feature vector \\(\\mathbf{f}_{i,k}^p\\) at coordinate \\(\\mathbf{y}_{i,k}^p\\) in the template from \\(\\mathbf{D}_{i,k}\\). We compute the feature similarity for each matched pair as: \\(S(\\mathbf{f}_m^p, \\mathbf{f}_{i,k}^p) = \\mathbf{f}_m^p \\cdot \\mathbf{f}_{i,k}^p\\), where \\(\\cdot\\) denotes the dot product." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16821" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "To calculate an overall similarity score between the target segment \\(\\mathbf{I}_m\\) and each template variant \\(\\mathbf{I}_{i,k}\\), we aggregate the similarity scores across all matched pairs as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.149, + 0.483, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} \\left(\\mathbf {I} _ {m}, \\mathbf {I} _ {i, k}\\right) = \\sum_ {p = 1} ^ {\\left| M _ {m, i, k} \\right|} S \\left(\\mathbf {f} _ {m} ^ {p}, \\mathbf {f} _ {i, k} ^ {p}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.204, + 0.484, + 0.25 + ], + "angle": 0, + "content": "where \\(M_{m,i,k}\\) is the set of all matched pairs between \\(\\mathbf{I}_m\\) and \\(\\mathbf{I}_{i,k}\\). After computing \\(\\mathrm{sim}(\\mathbf{I}_m,\\mathbf{I}_{i,k})\\) for each pair, we select the template with the highest similarity score:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.262, + 0.483, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\left(i _ {\\text {o p t}}, k _ {\\text {o p t}}\\right) = \\underset {i \\in \\{1, \\dots , 8 \\}, k \\in \\{1, \\dots , 5 \\}} {\\arg \\max } \\operatorname {s i m} \\left(\\mathbf {I} _ {m}, \\mathbf {I} _ {i, k}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.332 + ], + "angle": 0, + "content": "The selected template \\(\\mathbf{I}_{i_{\\mathrm{opt}},k_{\\mathrm{opt}}}\\) is then used as the closest match to the target segment for the pose estimation process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.338, + 0.221, + 0.354 + ], + "angle": 0, + "content": "3.2.3. Pose Fitting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.527 + ], + "angle": 0, + "content": "After selecting the suitable template \\(\\mathbf{I}_{i_{\\mathrm{opt}},k_{\\mathrm{opt}}}\\), we proceed to estimate pose \\(\\mathbf{T}_m = (\\mathbf{R}_m,\\mathbf{t}_m)\\), where \\(\\mathbf{R}_m\\) is a 3D rotation matrix and \\(\\mathbf{t}_m\\) is a 3D translation vector that transforms the object from model space to camera space. It relies on a set of 2D-3D correspondences \\(\\mathcal{C}_{t_{\\mathrm{final}}} = \\{(\\mathbf{y}_m^j,\\mathbf{P}^j)\\}_{j = 1}^{|M|}\\) where \\(\\mathbf{y}_m^j\\in \\mathbb{R}^2\\) represents the coordinates of matched pixels in the target segment \\(\\mathbf{I}_m\\), \\(\\mathbf{P}^j\\in \\mathbb{R}^3\\) denotes the corresponding 3D points in model space from the selected template, and there are \\(|M|\\) pairs. To determine \\(\\mathbf{T}_m\\), we solve the Perspective-n-Point (PnP) problem, which minimizes the reprojection error:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.539, + 0.483, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\underset {\\mathbf {R} _ {m}, \\mathbf {t} _ {m}} {\\arg \\min } \\sum_ {j = 1} ^ {| M |} \\left\\| \\mathbf {y} _ {m} ^ {j} - \\pi \\left(\\mathbf {R} _ {m} \\mathbf {P} ^ {j} + \\mathbf {t} _ {m}\\right) \\right\\| ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.623 + ], + "angle": 0, + "content": "where \\(\\pi\\) is the projection function that maps 3D points to 2D image points according to the camera intrinsics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.746 + ], + "angle": 0, + "content": "To enhance robustness against outliers, we employ the Efficient PnP (EPnP) algorithm [23] combined with a RANSAC-based fitting strategy [11]. In this approach, PnP is applied iteratively to random subsets of four correspondences from \\( \\mathcal{C}_{t_{\\mathrm{final}}} \\), generating multiple pose hypotheses. For each hypothesis, we count the number of inliers, defined as correspondences where the reprojection error falls below a predefined threshold \\( \\epsilon \\):" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.756, + 0.483, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\text {i n l i e r s} = \\left| \\left\\{j: \\left\\| \\mathbf {y} _ {m} ^ {j} - \\pi \\left(\\mathbf {R} _ {m} \\mathbf {P} ^ {j} + \\mathbf {t} _ {m}\\right) \\right\\| < \\epsilon \\right\\} \\right|. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.484, + 0.817 + ], + "angle": 0, + "content": "The hypothesis with the highest inlier count is selected as the final coarse pose estimate \\(\\mathbf{T}_m\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.224, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we first outline the experimental setup (Section 4.1). We then evaluate our method's performance in comparison with previous approaches on the seven core" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "datasets of the BOP challenge [17], examining accuracy, runtime efficiency, and effectiveness when using predicted 3D models (Section 4.2). This evaluation highlights the strengths and unique contributions of our approach. Lastly, we present an ablation study to investigate the impact of various configurations in our method (Section 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.704, + 0.211 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.217, + 0.907, + 0.369 + ], + "angle": 0, + "content": "Evaluation Datasets. We evaluate our approach on the seven core datasets of the BOP challenge [17]: LineMod Occlusion (LM-O) [3], T-LESS [15], TUD-L [16], IC-BIN [8], ITODD [9], HomebrewedDB (HB)[21], and YCB-Video (YCB-V) [57]. Together, these datasets feature 132 distinct objects and 19,048 testing instances, each in complex, cluttered scenes with partial occlusions. Table 1 provides the instance count for each dataset. It is worth noting that the unseen object pose estimation task remains challenging, with significant room for improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Evaluation Metrics. We use the BOP evaluation protocol [17] for 6D object localization, which assesses pose accuracy with three error metrics: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). VSD evaluates only the visible part of the object to handle ambiguities, MSSD measures 3D surface deviation with global symmetries, and MSPD assesses perceivable deviation using object symmetries in projection. A pose is considered correct for a metric \\( e \\) if \\( e < \\theta_e \\), with \\( \\theta_e \\) as the correctness threshold. The Average Recall (AR) for each metric \\( e \\), denoted \\( \\mathrm{AR}_e \\), is the mean Recall over various \\( \\theta_e \\) thresholds, and for VSD, multiple misalignment tolerances \\( \\tau \\). The overall AR score is the average of the three metrics: \\( \\mathrm{AR} = (\\mathrm{AR}_{\\mathrm{VSD}} + \\mathrm{AR}_{\\mathrm{MSSD}} + \\mathrm{AR}_{\\mathrm{MSPD}}) / 3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.906, + 0.761 + ], + "angle": 0, + "content": "Pose Refinement. To show Pos3R can integrate with render-and-compare refinement techniques, we apply the refinement method from MegaPose [22] to our result. Given an input image and an estimated pose, the refiner predicts the relative transformation between the initial and ground-truth pose. The refiner is trained on a large-scale dataset with several well-designed techniques [22]. We use the similarity score Eq. 1 to select the top-5 templates and each test crop-template pair gives a pose with EPnP. Then, we use the refiner to update these five pose hypotheses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.84, + 0.789 + ], + "angle": 0, + "content": "4.2. Comparison With the State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We compare Pos3R with both training-free and training-based methods. The training-free methods include Found-Pose [41] and ZS6D [2], while the training-based methods include MegaPose [22], GigaPose [39], OSOP [48], and GenFlow [34]. For pose refinement, we apply the refiner from MegaPose [22]. We use the publicly available code from [39] with the default hyperparameters." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16822" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.907, + 0.386 + ], + "angle": 0, + "content": "
#MethodTraining-FreeRefinementDatasets (num. instances)MeanTime
LM-OT-LESSTUD-LIC-BINITODDHBYCB-V
144564236001786304116304123
Coarse Pose Estimation:
1GigaPose [39]X-29.927.330.223.118.834.829.027.60.9
2GenFlow [34]X-25.021.530.016.815.428.327.723.53.8
3MegaPose [22]X-22.917.725.815.210.825.128.120.815.5
4OSOP [48]X-31.2----49.233.2--
5ZS6D [2]✓-29.821.0----32.4--
6FoundPose [41]✓-39.633.846.723.920.450.845.237.21.7
7Baseline✓-30.523.643.221.415.442.549.332.30.7
7Pos3R (Ours)✓-32.331.547.333.125.153.753.939.51.4
With Pose Refinement (5 hypotheses):
8FoundPoseXMegaPose58.654.965.744.436.170.367.356.811.2
9GigaPoseXMegaPose59.957.064.546.739.772.266.357.97.3
10GenFlowXGenFlow56.352.368.445.339.573.963.357.120.9
11MegaPoseXMegaPose56.050.768.441.433.870.462.154.747.4
12Pos3R (ours)XMegaPose56.557.566.846.037.570.166.457.38.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.396, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Table 1. Performance Comparison on the Seven Datasets of BOP. This table reports the Average Recall (AR) scores per dataset, the mean AR score across all datasets, and the time required to estimate poses for all objects in an image (in seconds). The runtime data of other methods are sourced from FoundPose [41]. The upper section lists methods for coarse pose estimation without refinement, while the lower section presents methods that incorporate a refinement stage using multiple pose hypotheses, reporting the best refined pose. Methods without task-specific training are marked with a green check mark \\((\\checkmark)\\). In the coarse estimation category, Pos3R achieves the highest AR scores on several datasets and demonstrates good overall generalization with a competitive mean AR score. While refinement-based methods yield slightly higher accuracy in certain cases, Pos3R remains a robust and efficient choice for coarse pose estimation and performs comparably to top refined methods when combined with a refinement stage. The best results in the coarse estimation section are highlighted in green, and the best results in the refinement section are highlighted in blue. The second-best results are underlined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.676 + ], + "angle": 0, + "content": "Table 1 provides a detailed comparison of Pos3R (Pos3R) with other leading 6D pose estimation methods across the seven primary datasets in the BOP challenge. The upper portion of the table showcases results for coarse pose estimation without refinement, while the lower portion highlights performance after incorporating a pose refinement step with five hypotheses. It is important to note that pose refinement involves task-specific training on extensive pose datasets, making it non-training-free [22]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.483, + 0.864 + ], + "angle": 0, + "content": "In the coarse pose estimation category, Pos3R demonstrates superior performance over other training-free methods, including FoundPose and ZS6D, achieving the highest mean accuracy and consistently strong results across most datasets. For instance, Pos3R attains the highest AR (Average Recall) on the TUD-L, HB, and YCB-V datasets, along with a mean AR of 39.5. This high performance across varied datasets reflects Pos3R's robust generalization capabilities for different object poses and scenarios. Additionally, with a runtime of only 1.4 seconds, Pos3R stands out as an efficient alternative to methods that either require substantial computational resources or are not training-free." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Pos3R remains competitive after pose refinement using MegaPose's refine. Although Pos3R is not inher" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.905, + 0.645 + ], + "angle": 0, + "content": "ently designed for refinement, it achieves comparable accuracy to MegaPose and other refined methods across several datasets. These findings underscore the adaptability of Pos3R, as it provides accurate initial pose estimates and benefits from additional refinement techniques—demonstrating its potential for scalable, practical, and flexible applications in 6D pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.842 + ], + "angle": 0, + "content": "Moreover, we observe that Pos3R faces challenges in achieving competitive accuracy on the LM-O dataset, where significant occlusion poses a major difficulty. Occlusion disrupts the initial image matching process, often resulting in suboptimal pose estimates that refinement methods struggle to correct. FoundPose [41] addresses this issue by employing a bag-of-words strategy for template selection, which has shown greater resilience to occlusion. To improve robustness in such scenarios, incorporating the local contrastive learning approach used by GigaPose [39] could be a promising direction. This method is specifically designed to enhance feature matching in occluded regions, and we consider this an area for future exploration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.71, + 0.866 + ], + "angle": 0, + "content": "4.3. Component Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Pose Estimation Using Predicted 3D Models. Model-based methods for unseen object pose estimation typically" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16823" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 2. Comparison with Predicted 3D Models on LM-O [3]. Average Recall is reported. All methods use 3D models predicted by Wonder3D [32]. Results with pose refinement, applied to the generated 3D models, are also included for comparison." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.156, + 0.485, + 0.23 + ], + "angle": 0, + "content": "
#MethodSingle Image-to-3DGT 3D model w/o refinement
CoarseRefined
1MegaPose [22]15.425.222.7
2GigaPose [39]17.627.229.4
3Pos3R19.528.532.3
" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.236, + 0.468, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.426, + 0.486, + 0.482 + ], + "angle": 0, + "content": "Figure 3. Rendered Images and 3D Coordinate Maps from Wonder3D and CAD Model. The top two rows show rendered images and 3D coordinate maps from Wonder3D [32], while the bottom two rows show those from the real CAD model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.502, + 0.483, + 0.729 + ], + "angle": 0, + "content": "require accurate, textured 3D models. To relax this requirement, we follow the approach in [39] and use 3D models predicted by Wonder3D [32] from a single reference image. The predicted 3D models use the same object coordinate system as the ground truth (GT) frame, ensuring aligned axes. We then assess the performance of our method using these reconstructed models in place of the high-fidelity CAD models provided by the dataset. For a fair comparison, we adopt the same settings as [39], including the reference image for Wonder3D and post-processing steps, and report results on the LM-O dataset [3]. Examples of rendered views from generated 3D model are shown in Fig. 3. As shown in Table 2, Pos3R achieves higher AR scores than both MegaPose [22] and GigaPose [39], both before and after pose refinement, showing the effectiveness of Pos3R." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Template Selection Technique. In our experiments, we use the similarity of correspondence matches (Eq. 1) as the primary method for template selection. Additionally, we consider two alternative approaches: 1) Inliers: selecting templates based on the number of inliers (Eq. 4); 2) Confidence: leveraging the 3D pointmaps and confidence maps provided by MASt3R for each target segment-template pair [24], where the average confidence score is used for selection. As illustrated in Fig. 4, template selection based on inliers proves effective, while selection based on similarity score yields the best performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.089, + 0.908, + 0.215 + ], + "angle": 0, + "content": "Table 3. Comparison of Pose Estimation Method. We evaluate our proposed method (Pos3R), a variant without axial rotation augmentation (\\( w/o \\) axial rotation), a variant that use face centers as base templates, and a version that replaces the 3D foundation model MASt3R with the 2D foundation model DINOv2. We also replace MASt3R with another dense image matcher RoMa [10]. Pos3R achieves the highest accuracy across all datasets, highlighting the importance of axial rotation handling and the advantage of using 3D foundation models over DINOv2." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.226, + 0.907, + 0.321 + ], + "angle": 0, + "content": "
#MethodLM-OTUD-LYCB-VT-LessIC-BINITODDHB
1Pos3R32.347.353.931.533.125.153.7
2w/o Ax. Rot.30.543.249.323.621.415.442.5
3Face Center31.545.553.328.029.422.452.5
4DINOv2 (L11)18.525.022.514.215.614.132.1
5DINOv2 (L9)20.526.024.015.616.614.832.9
6RoMa [10]20.825.032.016.217.015.525.1
" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.336, + 0.905, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.439, + 0.908, + 0.496 + ], + "angle": 0, + "content": "Figure 4. Comparison of Template Selection Techniques. We compare template selection methods based on confidence, inliers, and similarity score. The similarity score consistently achieves the highest average recall (AR) across all datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Impact of In-Plane Rotation and 3D Consistency on Pose Estimation. Table 3 evaluates the impact of incorporating in-plane (axial) rotations and 3D-consistent features on pose estimation performance across seven datasets. Our method, Pos3R (row 1), achieves the highest accuracy on all datasets by combining controlled in-plane rotations with the 3D-consistent features provided by the MASt3R foundation model. In contrast, row 2 presents results without in-plane rotation, where only eight base templates are used to account for out-of-plane rotations. The performance drops considerably, highlighting the importance of addressing in-plane rotations for improving correspondence quality and overall pose accuracy. Rows 4 and 5 present a further variation in which we replace MASt3R with the 2D foundation model DINOv2. Despite following the same pipeline (e.g., using the same 40 templates), this substitution leads to a notable performance decline across all datasets. Additionally, in row 6, replacing MASt3R with the dense image matcher RoMa [10] also results in lower accuracy. We speculate that MASt3R's improved performance stems from its large-scale and 3D-aware training, which may enhance its ability to produce features that are well-suited for pose estimation. The above analysis highlights the importance of incorporating both in-plane rotations and 3D-consistent features to achieve robust and accurate 6D pose estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16824" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.129, + 0.119, + 0.159 + ], + "angle": 270, + "content": "YCBV" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.089, + 0.316, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.089, + 0.508, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.089, + 0.703, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.089, + 0.897, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.231, + 0.119, + 0.26 + ], + "angle": 270, + "content": "LM-O" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.204, + 0.315, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.204, + 0.508, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.204, + 0.703, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.204, + 0.897, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.357, + 0.119, + 0.39 + ], + "angle": 270, + "content": "IC-BIN" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.318, + 0.315, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.319, + 0.508, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.319, + 0.703, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.319, + 0.897, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.471, + 0.118, + 0.506 + ], + "angle": 270, + "content": "T-LESS" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.433, + 0.315, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.433, + 0.508, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.433, + 0.703, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.433, + 0.897, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.553, + 0.907, + 0.624 + ], + "angle": 0, + "content": "Figure 5. Qualitative Results of 6D Pose Estimation. Visualization of 6D object pose estimation of Pos3R across LM-O [3], T-LESS [15], IC-BIN [8], and YCB-V [57]. Pos3R effectively handles a variety of objects and maintains accuracy even in very crowded scenes (e.g., IC-BIN), where predictions remain reasonably robust. Furthermore, Pos3R adapts well to texture-less objects, as demonstrated in the T-LESS dataset. However, as marked by red crosses \\((\\times)\\), heavy occlusion poses a challenge, with reduced accuracy in heavily occluded regions, illustrating both the strengths and limitations of our approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.897 + ], + "angle": 0, + "content": "Qualitative Results of 6D Pose Estimation. Figure 5 shows pose estimation results of Pos3R across challenging datasets, including LM-O [3], T-LESS [15], IC-BIN [8], and YCB-V [57]. Pos3R demonstrates robust performance across diverse object types, sizes, and textures, handling complex environments effectively. In cluttered scenes like IC-BIN, Pos3R accurately localizes multiple, tightly packed objects. It also performs well on textureless objects in T-LESS, a challenge for vision-based methods due to minimal surface features. However, Figure 5 also reveals a limitation: in heavily occluded scenes, marked by red crosses, Pos3R cannot well handle. Reduced visibility of occluded objects results in deviations in the predicted bounding boxes. This limitation suggests future directions for improvement, such as integrating occlusion-aware methods (e.g., GigaPose [39]) or leveraging multi-view information to enhance robustness in these challenging cases." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.64, + 0.634, + 0.655 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work introduces Pos3R, a training-free, RGB-only framework for 6D pose estimation of unseen objects. By leveraging the 3D foundation model MASt3R, Pos3R generates robust, 3D-consistent features that effectively handle both in-plane and out-of-plane rotations. Without relying on extensive datasets or object-specific training, Pos3R establishes a strong baseline for training-free research. Using a minimal set of templates captured from eight cube vertex viewpoints with controlled rotational variations, Pos3R achieves accurate and efficient pose estimation. Experiments on the BOP challenge demonstrate that Pos3R outperforms other training-free methods in coarse pose estimation and achieves competitive results with refined methods when combined with MegaPose refinement. Future work will focus on incorporating occlusion-aware techniques to enhance robustness in occluded settings." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16825" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "Acknowledgments. We thank all reviewers and ACs for their constructive comments. This work was generously supported through research collaboration with RIOS Intelligent Machines." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.152, + 0.188, + 0.168 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.221, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[2] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 463-469. IEEE, 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.292, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 536-551. Springer, 2014. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.377, + 0.483, + 0.472 + ], + "angle": 0, + "content": "[4] Yamei Chen, Yan Di, Guangyao Zhai, Fabian Manhardt, Chenyangguang Zhang, Ruida Zhang, Federico Tombari, Nassir Navab, and Benjamin Busam. Secondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9959-9969, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[5] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[6] Francesco Di Felice, Salvatore D'Avella, Alberto Remus, Paolo Tripicchio, and Carlo Alberto Avizzano. One-shot imitation learning with graph neural networks for pick-and-place manipulation tasks. IEEE Robotics and Automation Letters, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[7] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6608-6617, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[8] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3583-3592, 2016. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[9] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec itodd-a dataset for 3d object recognition in industry. In Proceedings of the IEEE international conference on computer vision workshops, pages 2200-2208, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[10] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.177, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[12] Ge Gao, Mikko Lauri, Yulong Wang, Xiaolin Hu, Jianwei Zhang, and Simone Frintrop. 6d object pose regression via supervised learning on point clouds. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3643-3649. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[13] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1580-1589, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[14] Xingyi He, Jiaming Sun, Yang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without cad models. Advances in Neural Information Processing Systems, 35:35103-35115, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[15] Tomás Hodan, Pavel Haluza, Štepan Obdržálek, Jiri Matas, Manolis Lourakis, and Xenophon Zabulis. T-less: An rgb-d dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 880-888. IEEE, 2017. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[16] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. Bop: Benchmark for 6d object pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 19-34, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[17] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. BOP challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[18] Sabera Hoque, Shuxiang Xu, Ananda Maiti, Yuchen Wei, and Md Yasir Arafat. Deep learning for 6d pose estimation of objects—a case study for autonomous driving. Expert Systems with Applications, 223:119838, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.907, + 0.692 + ], + "angle": 0, + "content": "[19] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2930-2939, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[20] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10095-10105, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[21] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapose: 6d pose estimation of novel objects via render & compare. arXiv preprint arXiv:2212.06870, 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16826" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[23] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[24] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[25] Hongyang Li, Jiehong Lin, and Kui Jia. Dcl-net: Deep correspondence learning network for 6d pose estimation. In European Conference on Computer Vision, pages 369-385. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[26] Yuelong Li, Yafei Mao, Raja Bala, and Sunil Hadap. Mrcnet: 6-dof pose estimation with multiscale residual correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10476-10486, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[27] Ruyi Lian and Haibin Ling. Checkerpose: Progressive dense keypoint localization for object pose estimation with graph neural network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14022-14033, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[28] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27906-27916, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[29] Jian Liu, Wei Sun, Chongpei Liu, Xing Zhang, and Qiang Fu. Robotic continuous grasping system by shape transformer-guided multiobject category-level 6-d pose estimation. IEEE Transactions on Industrial Informatics, 19 (11):11171-11181, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[30] Jian Liu, Wei Sun, Hui Yang, Zhiwen Zeng, Chongpei Liu, Jin Zheng, Xingyu Liu, Hossein Rahmani, Nicu Sebe, and Ajmal Mian. Deep learning-based object pose estimation: A comprehensive survey. arXiv preprint arXiv:2405.07801, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[31] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In European Conference on Computer Vision, pages 298-315. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[32] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[33] Ningkai Mo, Wanshui Gan, Naoto Yokoya, and Shifeng Chen. Es6d: A computation efficient and symmetry-aware 6d pose regression framework. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6718-6727, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[34] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. Genflow: Generalizable recurrent flow for 6d pose refinement of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10039-10049, 2024. 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[35] Fengjun Mu, Rui Huang, Ao Luo, Xin Li, Jing Qiu, and Hong Cheng. Temporalfusion: Temporal motion reasoning with multi-frame fusion for 6d object pose estimation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5930-5936. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[36] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6771-6780, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[37] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2134-2140, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[38] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Yinlin Hu, Renaud Marlet, Mathieu Salzmann, and Vincent Lepetit. Nope: Novel object pose estimation from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17923-17932, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.907, + 0.482 + ], + "angle": 0, + "content": "[39] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9903–9913, 2024. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[40] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[41] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In European Conference on Computer Vision, pages 163-182. Springer, 2025. 1, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[42] Jaewoo Park and Nam Ik Cho. Dprost: Dynamic projective spatial transformer network for 6d pose estimation. In European Conference on Computer Vision, pages 363-379. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[43] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10710-10719, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[44] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4561-4570, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[45] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud registration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3979-3985, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16827" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "poses of challenging objects without using depth. In Proceedings of the IEEE international conference on computer vision, pages 3828-3836, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[47] Caner Sahin and Tae-Kyun Kim. Category-level 6d object pose recovery in depth images. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[48] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6835-6844, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[49] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In 2019 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[51] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6825-6834, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.484, + 0.547 + ], + "angle": 0, + "content": "[52] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.618 + ], + "angle": 0, + "content": "[53] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 530-546. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[54] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 108-125. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[55] Jiaxin Wei, Xibin Song, Weizhe Liu, Laurent Kneip, Hongdong Li, and Pan Ji. Rgb-based category-level object pose estimation via decoupled metric scale recovery. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2036-2042. IEEE, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[56] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868-17879, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[57] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[58] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[59] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.316 + ], + "angle": 0, + "content": "[60] Heng Zhao, Shenxing Wei, Dahu Shi, Wenming Tan, Zheyang Li, Ye Ren, Xing Wei, Yi Yang, and Shiliang Pu. Learning symmetry-aware geometry correspondences for 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14045-14054, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.908, + 0.387 + ], + "angle": 0, + "content": "[61] Wanqing Zhao, Shaobo Zhang, Ziyu Guan, Wei Zhao, Jinye Peng, and Jianping Fan. Learning deep network for detecting 3d object keypoints and 6d poses. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 14134-14142, 2020. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16828" + } + ] +] \ No newline at end of file diff --git a/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/162f30eb-880f-42fb-95ca-56f10c93b7f6_origin.pdf b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/162f30eb-880f-42fb-95ca-56f10c93b7f6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..81ba92e890cfb717fbaf11fc0ccea74adf1771f0 --- /dev/null +++ b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/162f30eb-880f-42fb-95ca-56f10c93b7f6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e94f73406b12301637bbd564afa09fe17af978d72efbefc165387626c970e6d +size 7107682 diff --git a/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/full.md b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ea67678ef734d2eaf391476ff300f8f1b4e21b4d --- /dev/null +++ b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/full.md @@ -0,0 +1,303 @@ +# Pos3R: 6D Pose Estimation for Unseen Objects Made Easy + +Weijian Deng $^{1}$ Dylan Campbell $^{1}$ Chunyi Sun $^{1}$ Jiahao Zhang $^{1}$ + +Shubham Kanitkar $^{2}$ Matthew E. Shaffer $^{2}$ Stephen Gould $^{1}$ + +$^{1}$ Australian National University $^{2}$ RIOS Intelligent Machines + +$^{1}\{\text {firstname.lastname } \} @ \text { anu.edu.au } ^{2}\{\text {firstname.lastname } \} @ \text { rios.ai }$ + +# Abstract + +Foundation models have significantly reduced the need for task-specific training, while also enhancing generalizability. However, state-of-the-art 6D pose estimators either require further training with pose supervision or neglect advances obtainable from 3D foundation models. The latter is a missed opportunity, since these models are better equipped to predict 3D-consistent features, which are of significant utility for the pose estimation task. To address this gap, we propose Pos3R, a method for estimating the 6D pose of any object from a single RGB image, making extensive use of a 3D reconstruction foundation model and requiring no additional training. We identify template selection as a particular bottleneck for existing methods that is significantly alleviated by the use of a 3D model, which can more easily distinguish between template poses than a 2D model. Despite its simplicity, Pos3R achieves competitive performance on the Benchmark for 6D Object Pose Estimation (BOP), matching or surpassing existing refinement-free methods. Additionally, Pos3R integrates seamlessly with render-and-compare refinement techniques, demonstrating adaptability for high-precision applications. + +# 1. Introduction + +Six-dimensional (6D) object pose estimation—the task of determining the exact position and orientation of objects relative to a camera—is essential for applications in robotics, augmented reality, and autonomous systems. Reliable pose estimation enables critical tasks like object manipulation, grasping, and assembly, allowing these systems to interact effectively in complex and dynamic environments [6, 18, 29, 30, 49, 54]. Traditional approaches often rely on learning-based methods tailored to specific objects or categories, achieving high accuracy but struggling to generalize to new categories or unseen objects [4, 5, 7, 26, 33, 46, 47, 53-55, 59, 61]. This limitation is particularly problematic in dynamic, data-scarce environments where adaptability and flexibility are crucial. + +![](images/4a51ba4ad4b7ceaff74356db999e6f95204ed479b6381b88b278f7a244ad1c10.jpg) +Figure 1. Illustration of In-Plane and Out-of-Plane Rotations with Correspondence Quality between DINOv2 and MASt3R. + +The left diagram illustrates in-plane rotations (within the image plane) and out-of-plane rotations (3D orientation changes) from a camera view. On the right, DINOv2 [40] (top row) and MASt3R [24] (bottom row) are compared in their handling of these rotations. DINOv2 exhibits sparse and inconsistent correspondences (shown by red lines), particularly under out-of-plane rotations, due to its 2D feature limitations. In contrast, MASt3R provides dense and stable correspondences across both types of rotations, reflecting its ability to produce 3D-consistent features. + +To overcome these limitations, there has been a recent shift toward model-based approaches that aim to generalize pose estimation to unseen objects without object-specific training [2, 17, 41, 56]. These typically rely on a two-stage pipeline: detecting and localizing objects within a scene, followed by a render-and-compare process that matches detected object regions to a set of template models. Building on this concept, recent works have sought to improve generalization through large-scale model training with significant attention to object diversity. For example, Foundation-Pose [56] uses synthetic training data combined with a large language model and contrastive learning, enhancing feature alignment across varied domains for robust generalization. + +The recent emergence of training-free methods offers a promising alternative, enabling 6D pose estimation for unseen objects without object- or task-specific training. Foundation models like DINOv2 [40] have demonstrated strong zero-shot capabilities, capturing spatial and semantic details through learned features [1, 58]. FoundPose [41], for instance, leverages DINOv2 descriptors to bridge synthetic + +and real data, achieving competitive results on the BOP challenge [17]. These advancements underscore the potential of training-free frameworks in 6D pose estimation. + +Building on this trend, we explore the strengths of 3D foundation models. Accurate pose estimation requires handling both in-plane and out-of-plane rotations, which are demonstrated in Fig. 1. In-plane rotations occur within the image plane, as when an object rotates around the camera's line of sight, while out-of-plane rotations involve changes in 3D orientation, such as tilting or turning, which significantly alters appearance due to perspective shifts. Addressing both types of rotation is essential for robust pose estimation, as objects often appear from varied angles. Although 2D foundation models like DINOv2 handle in-plane rotations well through training augmentations, they lack robustness to out-of-plane rotations due to their 2D limitations (shown in Fig. 1). In contrast, 3D foundation models, such as MASt3R [24], are specifically designed to produce 3D-consistent features across different viewpoints, enabling reliable feature alignment even with out-of-plane rotations. + +Motivated by this, we propose Pos3R, a training-free method for 6D pose estimation of unseen objects using only RGB inputs. Pos3R leverages the 3D foundation model MASt3R [24] for pose estimation without requiring additional training. The core of Pos3R lies in the image matching process between test crops and rendered templates from a CAD model. While common matchers like LoFTR [50] often struggle with synthetic-to-real matching due to domain gaps, MASt3R produces high-quality 2D correspondences in this case. These reliable matches enable Pos3R to establish the 2D-3D correspondences required for pose estimation using the PnP-RANSAC algorithm. + +By harnessing MASt3R's ability to produce dense correspondences robust to viewpoint and illumination changes, we simplify template generation and achieve computational efficiency. Unlike existing methods that render hundreds of templates, we place camera positions to cover both in-plane and out-of-plane rotations. Specifically, we use eight base templates per object, positioned at the vertices of a cube centered on the CAD model, to capture out-of-plane rotations. For each base template, we generate five in-plane rotational templates along the object's principal axis, resulting in forty templates per object. To ensure efficient and accurate pose estimation, Pos3R dynamically selects the optimal rotation based on matching quality. This simple yet effective strategy maintains computational efficiency while achieving reliable pose estimation. Tested on seven diverse datasets in the BOP challenge, Pos3R demonstrates strong performance as a robust, scalable option for training-free 6D pose estimation of unseen objects, adaptable to render-and-compare refinement techniques. + +In summary, our contributions are: + +- While existing methods rely on 2D foundation models + +for unseen object pose estimation, we present Pos3R, a training-free method that leverages the 3D foundation model MAST3R [24] to improve robustness. + +- Leveraging MASt3R's robust dense correspondences, Pos3R uses only forty strategically placed templates per object to capture in-plane and out-of-plane rotations. A simple selection technique based on correspondence quality ensures accurate matching, achieving strong performance on the BOP challenge. + +# 2. Related Work + +Seen Object Pose Estimation. Instance-level object pose estimation refers to the task of estimating the poses of specific objects previously encountered during model training [5, 7, 26, 33, 46, 54, 61]. Among common approaches are correspondence-based methods, which learn to identify precise alignments between input data and CAD models of the objects [7, 27, 46]. Other methods use template-based strategies, where the model selects the most similar pose-labeled template from a predefined set of examples [25, 35, 52]. Additionally, regression-based approaches directly predict object poses from the learned object-specific features [12, 19, 26, 42]. + +These instance-specific techniques offer high precision but typically require retraining for new object instances, limiting their generalizability. This limitation has led to the exploration of category-level methods, which generalize within specific object categories and allow estimation of unseen objects in known categories [4, 47, 53, 55, 59]. For example, SecondPose [4] enhances category-level 6D pose estimation by integrating object-specific geometric features with DINOv2 SE(3)-consistent semantic priors, effectively addressing intra-class shape variation. However, these methods struggle outside their target categories, while our approach enables pose estimation without relying on such constraints, allowing broader applications. + +Unseen Object Pose Estimation. To enhance pose estimation flexibility, many methods aim to generalize across new object instances without object-specific training [22, 31, 37, 39, 43]. Approaches for this task can be broadly divided into manual reference view-based and CAD model-based methods. (i) Reference view-based methods use multi-view images with known poses as reference data for pose estimation [14, 31, 43, 51]. For example, OnePose [14, 51] reconstructs 3D object point clouds from posed RGB images, establishing 2D-3D correspondences to solve 6D poses. In a different approach, Nope [38], adopting the perspective of generating new views, trains models to predict discriminative embeddings for novel object perspectives. (ii) CAD model-based approaches leverage 3D object model to facilitate pose estimation. Some methods match features between the CAD model and the query image [20, 28, 60], + +![](images/f01f662c5b4ced6619d7971ccf11c13a4a43baacff99389e263fbea499a8d11e.jpg) +Figure 2. Overview of the 6D Pose Estimation Process in Pos3R. (1) Template Rendering (Sec. 3.2.1): A CAD model is used to generate a set of forty templates, each representing a unique orientation that covers both in-plane and out-of-plane rotations. This is achieved by positioning the camera at the vertices of a cube around the object. (2) Image Matching (Sec. 3.2.2): Given a test segment from CNOS [37], we establish 2D-2D correspondences between the test image and each rendered template. We leverage the 3D-consistent features generated by MASt3R [24] for accurate matching and select the best template based on matching quality. Each template also includes a 3D object coordinate map that records the corresponding 3D points. (3) Pose Fitting (Sec. 3.2.3): Using the selected correspondences, we apply the PnP-RANSAC algorithm [23] to obtain a final pose that aligns the object with its observed position in the scene. + +while others employ template matching to find the closest initial pose, refining it further using specialized refiners for higher accuracy [22, 34, 39]. For example, GigaPose [39] introduces an efficient two-network system: one network for retrieving template views (out-of-plane rotation) and another for estimating the remaining degrees of freedom (in-plane rotation and 3D translation). This separation reduces computational costs, contrasting with methods like MegaPose [22], which applies a single network to every possible test crop-template pair. Our work builds on the CAD model-based approach, aligning with recent BOP challenge protocols [17]. We focus on a training-free pipeline for unseen object pose estimation, a direction distinct from methods that require extensive training of task-specific networks. + +Training-Free Object Pose Estimation. Traditionally, 6D object pose estimation is achieved by establishing 3D-to-2D correspondences followed by a $\mathrm{PnP}$ algorithm [11, 13, 44, 61]. Recently, leveraging features from foundation models has gained traction, especially with models like DINO, which offer spatial detail and semantic consistency for pose estimation [1, 58]. FoundPose [41] uses DINOv2 descriptors to bridge synthetic and real data domains, providing strong performance for symmetric objects with RGB-only inputs. FreeZe integrates 3D point descriptors from GeDi [45] with image features from DINOv2 [40] for RGB-D pose estimation. By leveraging the 3D consistent feature of MAST3R to produce dense correspondences, our work explores training-free, unseen object pose estimation with RGB-only inputs, enabling effective pose estimation across diverse object categories without task-specific training. + +# 3.6D Pose Estimation with Pos3R + +# 3.1. Task Definition + +Given a 3D model of a query object $\mathbf{Q}$ and an RGB image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ containing $\mathbf{Q}$ , the task is to estimate the 6D pose of $\mathbf{Q}$ relative to the camera's reference frame, with known intrinsics $\mathbf{K}$ . Specifically, we aim to determine the 6DoF transformation $\mathbf{T} = (\mathbf{R}, \mathbf{t})$ in 3D space, where $\mathbf{R} \in \mathbb{R}^{3 \times 3}$ represents the rotation matrix and $\mathbf{t} \in \mathbb{R}^3$ denotes the translation vector. The segmented object region $\mathbf{I}_m = \mathbf{M} \odot \mathbf{I}$ is created by element-wise multiplication of the binary segmentation mask $\mathbf{M}$ and $\mathbf{I}$ , isolating the visible part of $\mathbf{Q}$ . We note that $\mathbf{Q}$ may be partially occluded. + +This work leverages the 3D foundation model MASt3R [24] to tackle model-based 6D pose estimation for unseen objects. We introduce Pos3R, a training-free, RGB-only method that effectively estimates 6D poses without object- or task-specific training. + +# 3.2. Training-Free Pipeline + +Overview. Following the standard model-based pipeline for 6D pose estimation of unseen objects [2, 39, 41], Pos3R comprises two components: object detection and pose estimation. We keep each component frozen, avoiding any object- or task-specific training. + +For object detection, we use CNOS [37] to generate segmentation masks and object identities for each target instance, enabling localization of the target segment $\mathbf{I}_m$ within the RGB image $\mathbf{I}$ . As a default method in the BOP challenge for segmenting unseen objects, CNOS requires + +only 3D models for onboarding and does not depend on additional data or object-specific training. + +For 6D pose estimation, we illustrate three steps in Fig. 2. Specifically, given a set of templates rendered from a textured CAD model, Pos3R utilizes the 3D foundation model MAST3R [24] to extract features from both the target segment $\mathbf{I}_m$ and each template, enabling estimation of the 6D transformation $\mathbf{T} = (\mathbf{R},\mathbf{t})$ through 2D-3D correspondences using the PnP-RANSAC algorithm. + +# 3.2.1. Template Rendering + +Using a textured 3D CAD model, we render templates of the object from various orientations. The rendering process follows standard rasterization methods as described in [39], with a black background and fixed lighting. The rendering camera is configured with the same intrinsic parameters, $\mathbf{K}$ , as the test camera, and the rendered templates match the size of the test image $\mathbf{I}$ . The object remains centered in each template. Additionally, the 3D locations in the coordinate space of the 3D CAD model corresponding to each pixel in the rendered template are recorded, enabling the establishment of 2D-3D correspondences. + +A critical component of Pos3R is establishing correspondences between pixels in the target segment and the template. Accurate pose estimation requires addressing both in-plane and out-of-plane rotations, as objects are frequently viewed from diverse orientations. In-plane rotations occur within the image plane, such as when an object rotates around the camera's line of sight. Out-of-plane rotations, on the other hand, involve 3D orientation changes—such as tilting or turning—that significantly alter an object's appearance due to perspective shifts. Existing methods tackle this challenge through extensive template libraries and specialized selection mechanisms. For example, MegaPose [22] uses hundreds of templates to capture a broad range of object poses, but maintaining such a large template library necessitates a dedicated selection network. + +Template Configuration. To reduce reliance on extensive template libraries, Pos3R leverages the 3D foundation model MASt3R, which generates 3D-consistent features across viewpoints. This enables Pos3R to handle out-of-plane rotations effectively without requiring hundreds of templates or complex selection mechanisms. We use a set of eight base templates, denoted by $\{\mathbf{I}_i\}_{i=1}^8$ , covering essential orientations. They are positioned at the vertices of a cube centered around the CAD model, effectively capturing out-of-plane rotations. + +To further address ambiguities from in-plane (axial) rotations, which can impact correspondence quality, we apply controlled rotational variations to each base template. For each template $\mathbf{I}_i$ , we generate $T$ rotations around the camera's principal axis, each rotation defined by an angle $\theta_k = \frac{2\pi k}{T}$ , where $k = 0,\dots,T - 1$ , covering the full + +$360^{\circ}$ range. In our experiments, we set $T = 5$ to balance efficiency and accuracy, yielding a set of rotationally-augmented templates $\{\mathbf{I}_{i,k}\}_{k=1}^{5}$ . In the following, we discuss the template selection process in detail. + +# 3.2.2. Image Matching + +MASt3R as an Image Matcher. Our approach builds upon MASt3R [24], a model for joint local 3D reconstruction and pixelwise matching between two input images $\mathbf{I}_a$ and $\mathbf{I}_b \in \mathbb{R}^{H \times W \times 3}$ . Conceptually, MASt3R operates as a mapping function $\mathbf{f}(\mathbf{I}_a, \mathbf{I}_b) = \mathrm{Dec}(\mathrm{Enc}(\mathbf{I}_a), \mathrm{Enc}(\mathbf{I}_b))$ , where the encoder $\mathrm{Enc}(\mathbf{I}) \to \mathbf{F}$ is a Siamese Vision Transformer (ViT) that processes image $\mathbf{I}$ into token vectors of size $N \times d$ , with $N = w \times h$ , yielding $\mathbf{F} \in \mathbb{R}^{N \times d}$ . The decoder, $\mathrm{Dec}(\mathbf{F}_a, \mathbf{F}_b)$ , employs twin ViT modules that produce pixelwise pointmaps $\mathbf{X}$ , local feature maps $\mathbf{D}$ for each image. + +Using these local feature representations, correspondences between images are identified via the fastNN algorithm [24]. It efficiently establishes reciprocal matches between the feature maps $\mathbf{D}_a$ and $\mathbf{D}_b$ by initially seeding points across a uniform pixel grid and iteratively refining these seeds to form high-quality mutual correspondences. Through this process, fastNN accurately aligns key features across images. The resulting reciprocal pixel pairs between $\mathbf{I}_a$ and $\mathbf{I}_b$ are represented as $M_{a,b} = \{(y_a^c,y_b^c)\}_{c = 1}^{|M_{a,b}|}$ , where $\mathbf{y}_a^c$ and $\mathbf{y}_b^c\in \mathbb{N}^2$ denote the coordinates of matched pixels in each respective image. + +Similarity-Based Template Selection. Unlike methods that require hundreds of templates [22, 36, 41], Pos3R simplifies template selection by needing only forty templates, significantly reducing the complexity of the selection process. Rather than relying on a trained template selection network, we employ a straightforward, training-free approach based on the similarity of matched correspondences. Specifically, given a target segment $\mathbf{I}_m$ and a set of eight base templates $\{\mathbf{I}_i\}_{i=1}^8$ , each augmented with rotational variations $\{\mathbf{I}_{i,k}\}_{k=1}^5$ , we identify the most similar template for pose estimation by calculating correspondences between the target segment and each template. + +For each rotationally-augmented template $\mathbf{I}_{i,k}$ , we obtain reciprocal pixel pairs between $\mathbf{I}_m$ and $\mathbf{I}_{i,k}$ , represented as: $M_{m,i,k} = \{(y_m^p, y_{i,k}^p)\}_{p=1}^{|M_{m,i,k}|}$ , where $M_{m,i,k}$ is the set of all reciprocal pixel pairs between $\mathbf{I}_m$ and $\mathbf{I}_{i,k}$ , and $\mathbf{y}_m^p$ and $\mathbf{y}_{i,k}^p \in \mathbb{N}^2$ denote the coordinates of matched pixels in the target segment and the template variant, respectively. + +For each matched pair $(\mathbf{y}_m^p, \mathbf{y}_{i,k}^p)$ in $M_{m,i,k}$ , we retrieve the corresponding local features from the feature maps $\mathbf{D}_m$ and $\mathbf{D}_{i,k}$ generated by MASt3R. Specifically, we obtain the feature vector $\mathbf{f}_m^p$ at coordinate $\mathbf{y}_m^p$ in the target segment from $\mathbf{D}_m$ , and the feature vector $\mathbf{f}_{i,k}^p$ at coordinate $\mathbf{y}_{i,k}^p$ in the template from $\mathbf{D}_{i,k}$ . We compute the feature similarity for each matched pair as: $S(\mathbf{f}_m^p, \mathbf{f}_{i,k}^p) = \mathbf{f}_m^p \cdot \mathbf{f}_{i,k}^p$ , where $\cdot$ denotes the dot product. + +To calculate an overall similarity score between the target segment $\mathbf{I}_m$ and each template variant $\mathbf{I}_{i,k}$ , we aggregate the similarity scores across all matched pairs as follows: + +$$ +\operatorname {s i m} \left(\mathbf {I} _ {m}, \mathbf {I} _ {i, k}\right) = \sum_ {p = 1} ^ {\left| M _ {m, i, k} \right|} S \left(\mathbf {f} _ {m} ^ {p}, \mathbf {f} _ {i, k} ^ {p}\right), \tag {1} +$$ + +where $M_{m,i,k}$ is the set of all matched pairs between $\mathbf{I}_m$ and $\mathbf{I}_{i,k}$ . After computing $\mathrm{sim}(\mathbf{I}_m,\mathbf{I}_{i,k})$ for each pair, we select the template with the highest similarity score: + +$$ +\left(i _ {\text {o p t}}, k _ {\text {o p t}}\right) = \underset {i \in \{1, \dots , 8 \}, k \in \{1, \dots , 5 \}} {\arg \max } \operatorname {s i m} \left(\mathbf {I} _ {m}, \mathbf {I} _ {i, k}\right). \tag {2} +$$ + +The selected template $\mathbf{I}_{i_{\mathrm{opt}},k_{\mathrm{opt}}}$ is then used as the closest match to the target segment for the pose estimation process. + +# 3.2.3. Pose Fitting + +After selecting the suitable template $\mathbf{I}_{i_{\mathrm{opt}},k_{\mathrm{opt}}}$ , we proceed to estimate pose $\mathbf{T}_m = (\mathbf{R}_m,\mathbf{t}_m)$ , where $\mathbf{R}_m$ is a 3D rotation matrix and $\mathbf{t}_m$ is a 3D translation vector that transforms the object from model space to camera space. It relies on a set of 2D-3D correspondences $\mathcal{C}_{t_{\mathrm{final}}} = \{(\mathbf{y}_m^j,\mathbf{P}^j)\}_{j = 1}^{|M|}$ where $\mathbf{y}_m^j\in \mathbb{R}^2$ represents the coordinates of matched pixels in the target segment $\mathbf{I}_m$ , $\mathbf{P}^j\in \mathbb{R}^3$ denotes the corresponding 3D points in model space from the selected template, and there are $|M|$ pairs. To determine $\mathbf{T}_m$ , we solve the Perspective-n-Point (PnP) problem, which minimizes the reprojection error: + +$$ +\underset {\mathbf {R} _ {m}, \mathbf {t} _ {m}} {\arg \min } \sum_ {j = 1} ^ {| M |} \left\| \mathbf {y} _ {m} ^ {j} - \pi \left(\mathbf {R} _ {m} \mathbf {P} ^ {j} + \mathbf {t} _ {m}\right) \right\| ^ {2}, \tag {3} +$$ + +where $\pi$ is the projection function that maps 3D points to 2D image points according to the camera intrinsics. + +To enhance robustness against outliers, we employ the Efficient PnP (EPnP) algorithm [23] combined with a RANSAC-based fitting strategy [11]. In this approach, PnP is applied iteratively to random subsets of four correspondences from $\mathcal{C}_{t_{\mathrm{final}}}$ , generating multiple pose hypotheses. For each hypothesis, we count the number of inliers, defined as correspondences where the reprojection error falls below a predefined threshold $\epsilon$ : + +$$ +\text {i n l i e r s} = \left| \left\{j: \left\| \mathbf {y} _ {m} ^ {j} - \pi \left(\mathbf {R} _ {m} \mathbf {P} ^ {j} + \mathbf {t} _ {m}\right) \right\| < \epsilon \right\} \right|. \tag {4} +$$ + +The hypothesis with the highest inlier count is selected as the final coarse pose estimate $\mathbf{T}_m$ . + +# 4. Experiments + +In this section, we first outline the experimental setup (Section 4.1). We then evaluate our method's performance in comparison with previous approaches on the seven core + +datasets of the BOP challenge [17], examining accuracy, runtime efficiency, and effectiveness when using predicted 3D models (Section 4.2). This evaluation highlights the strengths and unique contributions of our approach. Lastly, we present an ablation study to investigate the impact of various configurations in our method (Section 4.3). + +# 4.1. Experimental Setup + +Evaluation Datasets. We evaluate our approach on the seven core datasets of the BOP challenge [17]: LineMod Occlusion (LM-O) [3], T-LESS [15], TUD-L [16], IC-BIN [8], ITODD [9], HomebrewedDB (HB)[21], and YCB-Video (YCB-V) [57]. Together, these datasets feature 132 distinct objects and 19,048 testing instances, each in complex, cluttered scenes with partial occlusions. Table 1 provides the instance count for each dataset. It is worth noting that the unseen object pose estimation task remains challenging, with significant room for improvement. + +Evaluation Metrics. We use the BOP evaluation protocol [17] for 6D object localization, which assesses pose accuracy with three error metrics: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). VSD evaluates only the visible part of the object to handle ambiguities, MSSD measures 3D surface deviation with global symmetries, and MSPD assesses perceivable deviation using object symmetries in projection. A pose is considered correct for a metric $e$ if $e < \theta_e$ , with $\theta_e$ as the correctness threshold. The Average Recall (AR) for each metric $e$ , denoted $\mathrm{AR}_e$ , is the mean Recall over various $\theta_e$ thresholds, and for VSD, multiple misalignment tolerances $\tau$ . The overall AR score is the average of the three metrics: $\mathrm{AR} = (\mathrm{AR}_{\mathrm{VSD}} + \mathrm{AR}_{\mathrm{MSSD}} + \mathrm{AR}_{\mathrm{MSPD}}) / 3$ . + +Pose Refinement. To show Pos3R can integrate with render-and-compare refinement techniques, we apply the refinement method from MegaPose [22] to our result. Given an input image and an estimated pose, the refiner predicts the relative transformation between the initial and ground-truth pose. The refiner is trained on a large-scale dataset with several well-designed techniques [22]. We use the similarity score Eq. 1 to select the top-5 templates and each test crop-template pair gives a pose with EPnP. Then, we use the refiner to update these five pose hypotheses. + +# 4.2. Comparison With the State of the Art + +We compare Pos3R with both training-free and training-based methods. The training-free methods include Found-Pose [41] and ZS6D [2], while the training-based methods include MegaPose [22], GigaPose [39], OSOP [48], and GenFlow [34]. For pose refinement, we apply the refiner from MegaPose [22]. We use the publicly available code from [39] with the default hyperparameters. + +
#MethodTraining-FreeRefinementDatasets (num. instances)MeanTime
LM-OT-LESSTUD-LIC-BINITODDHBYCB-V
144564236001786304116304123
Coarse Pose Estimation:
1GigaPose [39]X-29.927.330.223.118.834.829.027.60.9
2GenFlow [34]X-25.021.530.016.815.428.327.723.53.8
3MegaPose [22]X-22.917.725.815.210.825.128.120.815.5
4OSOP [48]X-31.2----49.233.2--
5ZS6D [2]✓-29.821.0----32.4--
6FoundPose [41]✓-39.633.846.723.920.450.845.237.21.7
7Baseline✓-30.523.643.221.415.442.549.332.30.7
7Pos3R (Ours)✓-32.331.547.333.125.153.753.939.51.4
With Pose Refinement (5 hypotheses):
8FoundPoseXMegaPose58.654.965.744.436.170.367.356.811.2
9GigaPoseXMegaPose59.957.064.546.739.772.266.357.97.3
10GenFlowXGenFlow56.352.368.445.339.573.963.357.120.9
11MegaPoseXMegaPose56.050.768.441.433.870.462.154.747.4
12Pos3R (ours)XMegaPose56.557.566.846.037.570.166.457.38.0
+ +Table 1. Performance Comparison on the Seven Datasets of BOP. This table reports the Average Recall (AR) scores per dataset, the mean AR score across all datasets, and the time required to estimate poses for all objects in an image (in seconds). The runtime data of other methods are sourced from FoundPose [41]. The upper section lists methods for coarse pose estimation without refinement, while the lower section presents methods that incorporate a refinement stage using multiple pose hypotheses, reporting the best refined pose. Methods without task-specific training are marked with a green check mark $(\checkmark)$ . In the coarse estimation category, Pos3R achieves the highest AR scores on several datasets and demonstrates good overall generalization with a competitive mean AR score. While refinement-based methods yield slightly higher accuracy in certain cases, Pos3R remains a robust and efficient choice for coarse pose estimation and performs comparably to top refined methods when combined with a refinement stage. The best results in the coarse estimation section are highlighted in green, and the best results in the refinement section are highlighted in blue. The second-best results are underlined. + +Table 1 provides a detailed comparison of Pos3R (Pos3R) with other leading 6D pose estimation methods across the seven primary datasets in the BOP challenge. The upper portion of the table showcases results for coarse pose estimation without refinement, while the lower portion highlights performance after incorporating a pose refinement step with five hypotheses. It is important to note that pose refinement involves task-specific training on extensive pose datasets, making it non-training-free [22]. + +In the coarse pose estimation category, Pos3R demonstrates superior performance over other training-free methods, including FoundPose and ZS6D, achieving the highest mean accuracy and consistently strong results across most datasets. For instance, Pos3R attains the highest AR (Average Recall) on the TUD-L, HB, and YCB-V datasets, along with a mean AR of 39.5. This high performance across varied datasets reflects Pos3R's robust generalization capabilities for different object poses and scenarios. Additionally, with a runtime of only 1.4 seconds, Pos3R stands out as an efficient alternative to methods that either require substantial computational resources or are not training-free. + +Pos3R remains competitive after pose refinement using MegaPose's refine. Although Pos3R is not inher + +ently designed for refinement, it achieves comparable accuracy to MegaPose and other refined methods across several datasets. These findings underscore the adaptability of Pos3R, as it provides accurate initial pose estimates and benefits from additional refinement techniques—demonstrating its potential for scalable, practical, and flexible applications in 6D pose estimation. + +Moreover, we observe that Pos3R faces challenges in achieving competitive accuracy on the LM-O dataset, where significant occlusion poses a major difficulty. Occlusion disrupts the initial image matching process, often resulting in suboptimal pose estimates that refinement methods struggle to correct. FoundPose [41] addresses this issue by employing a bag-of-words strategy for template selection, which has shown greater resilience to occlusion. To improve robustness in such scenarios, incorporating the local contrastive learning approach used by GigaPose [39] could be a promising direction. This method is specifically designed to enhance feature matching in occluded regions, and we consider this an area for future exploration. + +# 4.3. Component Analysis + +Pose Estimation Using Predicted 3D Models. Model-based methods for unseen object pose estimation typically + +Table 2. Comparison with Predicted 3D Models on LM-O [3]. Average Recall is reported. All methods use 3D models predicted by Wonder3D [32]. Results with pose refinement, applied to the generated 3D models, are also included for comparison. + +
#MethodSingle Image-to-3DGT 3D model w/o refinement
CoarseRefined
1MegaPose [22]15.425.222.7
2GigaPose [39]17.627.229.4
3Pos3R19.528.532.3
+ +![](images/4b1f91f54bc77e1fc3af0bbfd20e32af7bf2e860c1382c3e8bfe94551275a14c.jpg) +Figure 3. Rendered Images and 3D Coordinate Maps from Wonder3D and CAD Model. The top two rows show rendered images and 3D coordinate maps from Wonder3D [32], while the bottom two rows show those from the real CAD model. + +require accurate, textured 3D models. To relax this requirement, we follow the approach in [39] and use 3D models predicted by Wonder3D [32] from a single reference image. The predicted 3D models use the same object coordinate system as the ground truth (GT) frame, ensuring aligned axes. We then assess the performance of our method using these reconstructed models in place of the high-fidelity CAD models provided by the dataset. For a fair comparison, we adopt the same settings as [39], including the reference image for Wonder3D and post-processing steps, and report results on the LM-O dataset [3]. Examples of rendered views from generated 3D model are shown in Fig. 3. As shown in Table 2, Pos3R achieves higher AR scores than both MegaPose [22] and GigaPose [39], both before and after pose refinement, showing the effectiveness of Pos3R. + +Template Selection Technique. In our experiments, we use the similarity of correspondence matches (Eq. 1) as the primary method for template selection. Additionally, we consider two alternative approaches: 1) Inliers: selecting templates based on the number of inliers (Eq. 4); 2) Confidence: leveraging the 3D pointmaps and confidence maps provided by MASt3R for each target segment-template pair [24], where the average confidence score is used for selection. As illustrated in Fig. 4, template selection based on inliers proves effective, while selection based on similarity score yields the best performance. + +Table 3. Comparison of Pose Estimation Method. We evaluate our proposed method (Pos3R), a variant without axial rotation augmentation ( $w/o$ axial rotation), a variant that use face centers as base templates, and a version that replaces the 3D foundation model MASt3R with the 2D foundation model DINOv2. We also replace MASt3R with another dense image matcher RoMa [10]. Pos3R achieves the highest accuracy across all datasets, highlighting the importance of axial rotation handling and the advantage of using 3D foundation models over DINOv2. + +
#MethodLM-OTUD-LYCB-VT-LessIC-BINITODDHB
1Pos3R32.347.353.931.533.125.153.7
2w/o Ax. Rot.30.543.249.323.621.415.442.5
3Face Center31.545.553.328.029.422.452.5
4DINOv2 (L11)18.525.022.514.215.614.132.1
5DINOv2 (L9)20.526.024.015.616.614.832.9
6RoMa [10]20.825.032.016.217.015.525.1
+ +![](images/f8a399f8d81c565c2614b700e53b5e13b756594a42fcc14314d522266eed4fc2.jpg) +Figure 4. Comparison of Template Selection Techniques. We compare template selection methods based on confidence, inliers, and similarity score. The similarity score consistently achieves the highest average recall (AR) across all datasets. + +Impact of In-Plane Rotation and 3D Consistency on Pose Estimation. Table 3 evaluates the impact of incorporating in-plane (axial) rotations and 3D-consistent features on pose estimation performance across seven datasets. Our method, Pos3R (row 1), achieves the highest accuracy on all datasets by combining controlled in-plane rotations with the 3D-consistent features provided by the MASt3R foundation model. In contrast, row 2 presents results without in-plane rotation, where only eight base templates are used to account for out-of-plane rotations. The performance drops considerably, highlighting the importance of addressing in-plane rotations for improving correspondence quality and overall pose accuracy. Rows 4 and 5 present a further variation in which we replace MASt3R with the 2D foundation model DINOv2. Despite following the same pipeline (e.g., using the same 40 templates), this substitution leads to a notable performance decline across all datasets. Additionally, in row 6, replacing MASt3R with the dense image matcher RoMa [10] also results in lower accuracy. We speculate that MASt3R's improved performance stems from its large-scale and 3D-aware training, which may enhance its ability to produce features that are well-suited for pose estimation. The above analysis highlights the importance of incorporating both in-plane rotations and 3D-consistent features to achieve robust and accurate 6D pose estimation. + +![](images/7709a89e402e4f20ffd30465c93613c0c1265eb2787a74991c4d2fabd9e091d0.jpg) +YCBV + +![](images/220a39336f50664e06af3f63b1e46c4395e6804d30f63768ebc87830171bab76.jpg) + +![](images/6e242a91145d8e673b8c55d71fad0250ed17726fe09f2b109fd37c947f07f2af.jpg) + +![](images/4746bd19ad495afe185abe8da093ae570f45bf15cb0a5d3543a578782745e87a.jpg) + +![](images/5cf411d0aa5e04f1b209a24cd092e06d026ab70dd0b05339f0a520da4d2e447a.jpg) +LM-O + +![](images/c342bd184f62cc991d6dedb7d2227046431709fdd4e5aa68b1d2fa38b0feb902.jpg) + +![](images/7436092b2d84476e8a73dfcffd2a99e017986aa1e0c9f741e413f8320f63fbff.jpg) + +![](images/8265ead8de098df5edfe4ff46a2c706a113a1fdcb1c800be96bdd5107b7f4e1a.jpg) + +![](images/c82b945bb82ddc732ddc98923c84d1a744160525024a5ee617d86bffa4534cc6.jpg) +IC-BIN + +![](images/8e6f7634359499e72ef07d103faa2a6d91232dbf7c85e206cdaad2efb572a864.jpg) + +![](images/3ae717336c2440f9995024939bd955d98cffe0d138835cee9007bdb6299d3a45.jpg) + +![](images/273a492fcb79ed9018c185b7246d278cf98e929f8940440728e063a12db6afe3.jpg) + +![](images/793280ee2a1334859367abb45b377c5d6b5d2a33f3f4a28d98c86e4d362dfbbc.jpg) +T-LESS + +![](images/192e63f0b688d5785574f2e0044d73e42d63a37d3c30467c8a174be424d8eb34.jpg) +Figure 5. Qualitative Results of 6D Pose Estimation. Visualization of 6D object pose estimation of Pos3R across LM-O [3], T-LESS [15], IC-BIN [8], and YCB-V [57]. Pos3R effectively handles a variety of objects and maintains accuracy even in very crowded scenes (e.g., IC-BIN), where predictions remain reasonably robust. Furthermore, Pos3R adapts well to texture-less objects, as demonstrated in the T-LESS dataset. However, as marked by red crosses $(\times)$ , heavy occlusion poses a challenge, with reduced accuracy in heavily occluded regions, illustrating both the strengths and limitations of our approach. + +![](images/4974464d2c96da5462c928acb4c403a46878563d75ef2be5403bdbcb2f495256.jpg) + +![](images/c7339eb09b04f2cba68c29f2dfd067925e3c78f3f277ecd4cf39906a93f58019.jpg) + +Qualitative Results of 6D Pose Estimation. Figure 5 shows pose estimation results of Pos3R across challenging datasets, including LM-O [3], T-LESS [15], IC-BIN [8], and YCB-V [57]. Pos3R demonstrates robust performance across diverse object types, sizes, and textures, handling complex environments effectively. In cluttered scenes like IC-BIN, Pos3R accurately localizes multiple, tightly packed objects. It also performs well on textureless objects in T-LESS, a challenge for vision-based methods due to minimal surface features. However, Figure 5 also reveals a limitation: in heavily occluded scenes, marked by red crosses, Pos3R cannot well handle. Reduced visibility of occluded objects results in deviations in the predicted bounding boxes. This limitation suggests future directions for improvement, such as integrating occlusion-aware methods (e.g., GigaPose [39]) or leveraging multi-view information to enhance robustness in these challenging cases. + +# 5. Conclusion + +This work introduces Pos3R, a training-free, RGB-only framework for 6D pose estimation of unseen objects. By leveraging the 3D foundation model MASt3R, Pos3R generates robust, 3D-consistent features that effectively handle both in-plane and out-of-plane rotations. Without relying on extensive datasets or object-specific training, Pos3R establishes a strong baseline for training-free research. Using a minimal set of templates captured from eight cube vertex viewpoints with controlled rotational variations, Pos3R achieves accurate and efficient pose estimation. Experiments on the BOP challenge demonstrate that Pos3R outperforms other training-free methods in coarse pose estimation and achieves competitive results with refined methods when combined with MegaPose refinement. Future work will focus on incorporating occlusion-aware techniques to enhance robustness in occluded settings. + +Acknowledgments. We thank all reviewers and ACs for their constructive comments. This work was generously supported through research collaboration with RIOS Intelligent Machines. + +# References + +[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 1, 3 +[2] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 463-469. IEEE, 2024. 1, 3, 5, 6 +[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 536-551. Springer, 2014. 5, 7, 8 +[4] Yamei Chen, Yan Di, Guangyao Zhai, Fabian Manhardt, Chenyangguang Zhang, Ruida Zhang, Federico Tombari, Nassir Navab, and Benjamin Busam. Secondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9959-9969, 2024. 1, 2 +[5] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 1, 2 +[6] Francesco Di Felice, Salvatore D'Avella, Alberto Remus, Paolo Tripicchio, and Carlo Alberto Avizzano. One-shot imitation learning with graph neural networks for pick-and-place manipulation tasks. IEEE Robotics and Automation Letters, 2023. 1 +[7] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6608-6617, 2020. 1, 2 +[8] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3583-3592, 2016. 5, 8 +[9] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec itodd-a dataset for 3d object recognition in industry. In Proceedings of the IEEE international conference on computer vision workshops, pages 2200-2208, 2017. 5 +[10] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 7 +[11] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to + +image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 5 +[12] Ge Gao, Mikko Lauri, Yulong Wang, Xiaolin Hu, Jianwei Zhang, and Simone Frintrop. 6d object pose regression via supervised learning on point clouds. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3643-3649. IEEE, 2020. 2 +[13] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1580-1589, 2020. 3 +[14] Xingyi He, Jiaming Sun, Yang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without cad models. Advances in Neural Information Processing Systems, 35:35103-35115, 2022. 2 +[15] Tomás Hodan, Pavel Haluza, Štepan Obdržálek, Jiri Matas, Manolis Lourakis, and Xenophon Zabulis. T-less: An rgb-d dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 880-888. IEEE, 2017. 5, 8 +[16] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. Bop: Benchmark for 6d object pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 19-34, 2018. 5 +[17] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. BOP challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 1, 2, 3, 5 +[18] Sabera Hoque, Shuxiang Xu, Ananda Maiti, Yuchen Wei, and Md Yasir Arafat. Deep learning for 6d pose estimation of objects—a case study for autonomous driving. Expert Systems with Applications, 223:119838, 2023. 1 +[19] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2930-2939, 2020. 2 +[20] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10095-10105, 2024. 2 +[21] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 5 +[22] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapose: 6d pose estimation of novel objects via render & compare. arXiv preprint arXiv:2212.06870, 2022. 2, 3, 4, 5, 6, 7 + +[23] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 3, 5 +[24] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 1, 2, 3, 4, 7 +[25] Hongyang Li, Jiehong Lin, and Kui Jia. Dcl-net: Deep correspondence learning network for 6d pose estimation. In European Conference on Computer Vision, pages 369-385. Springer, 2022. 2 +[26] Yuelong Li, Yafei Mao, Raja Bala, and Sunil Hadap. Mrcnet: 6-dof pose estimation with multiscale residual correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10476-10486, 2024. 1, 2 +[27] Ruyi Lian and Haibin Ling. Checkerpose: Progressive dense keypoint localization for object pose estimation with graph neural network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14022-14033, 2023. 2 +[28] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27906-27916, 2024. 2 +[29] Jian Liu, Wei Sun, Chongpei Liu, Xing Zhang, and Qiang Fu. Robotic continuous grasping system by shape transformer-guided multiobject category-level 6-d pose estimation. IEEE Transactions on Industrial Informatics, 19 (11):11171-11181, 2023. 1 +[30] Jian Liu, Wei Sun, Hui Yang, Zhiwen Zeng, Chongpei Liu, Jin Zheng, Xingyu Liu, Hossein Rahmani, Nicu Sebe, and Ajmal Mian. Deep learning-based object pose estimation: A comprehensive survey. arXiv preprint arXiv:2405.07801, 2024. 1 +[31] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In European Conference on Computer Vision, pages 298-315. Springer, 2022. 2 +[32] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 7 +[33] Ningkai Mo, Wanshui Gan, Naoto Yokoya, and Shifeng Chen. Es6d: A computation efficient and symmetry-aware 6d pose regression framework. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6718-6727, 2022. 1, 2 +[34] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. Genflow: Generalizable recurrent flow for 6d pose refinement of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10039-10049, 2024. 3, 5, 6 + +[35] Fengjun Mu, Rui Huang, Ao Luo, Xin Li, Jing Qiu, and Hong Cheng. Temporalfusion: Temporal motion reasoning with multi-frame fusion for 6d object pose estimation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5930-5936. IEEE, 2021. 2 +[36] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6771-6780, 2022. 4 +[37] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2134-2140, 2023. 2, 3 +[38] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Yinlin Hu, Renaud Marlet, Mathieu Salzmann, and Vincent Lepetit. Nope: Novel object pose estimation from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17923-17932, 2024. 2 +[39] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9903–9913, 2024. 2, 3, 4, 5, 6, 7, 8 +[40] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 3 +[41] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In European Conference on Computer Vision, pages 163-182. Springer, 2025. 1, 3, 4, 5, 6 +[42] Jaewoo Park and Nam Ik Cho. Dprost: Dynamic projective spatial transformer network for 6d pose estimation. In European Conference on Computer Vision, pages 363-379. Springer, 2022. 2 +[43] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10710-10719, 2020. 2 +[44] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4561-4570, 2019. 3 +[45] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud registration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3979-3985, 2022. 3 +[46] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d + +poses of challenging objects without using depth. In Proceedings of the IEEE international conference on computer vision, pages 3828-3836, 2017. 1, 2 +[47] Caner Sahin and Tae-Kyun Kim. Category-level 6d object pose recovery in depth images. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 2 +[48] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6835-6844, 2022. 5, 6 +[49] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In 2019 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227. IEEE, 2019. 1 +[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 2 +[51] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6825-6834, 2022. 2 +[52] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 2 +[53] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 530-546. Springer, 2020. 1, 2 +[54] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 108-125. Springer, 2020. 1, 2 +[55] Jiaxin Wei, Xibin Song, Weizhe Liu, Laurent Kneip, Hongdong Li, and Pan Ji. Rgb-based category-level object pose estimation via decoupled metric scale recovery. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2036-2042. IEEE, 2024. 1, 2 +[56] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868-17879, 2024. 1 +[57] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 5, 8 + +[58] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 1, 3 +[59] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2025. 1, 2 +[60] Heng Zhao, Shenxing Wei, Dahu Shi, Wenming Tan, Zheyang Li, Ye Ren, Xing Wei, Yi Yang, and Shiliang Pu. Learning symmetry-aware geometry correspondences for 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14045-14054, 2023. 2 +[61] Wanqing Zhao, Shaobo Zhang, Ziyu Guan, Wei Zhao, Jinye Peng, and Jianping Fan. Learning deep network for detecting 3d object keypoints and 6d poses. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 14134-14142, 2020. 1, 2, 3 \ No newline at end of file diff --git a/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/images.zip b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f3a90f47da5e91be7f0b39c9d056ef98ccef7050 --- /dev/null +++ b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7a64de465f1ecf894a22c1481c302754385952e39722ffefaeb0c3d141bc9a9 +size 689374 diff --git a/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/layout.json b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3291971d0870bb9d0aadf926bd1b33b7c7ec2853 --- /dev/null +++ b/2025/Pos3R_ 6D Pose Estimation for Unseen Objects Made Easy/layout.json @@ -0,0 +1,9076 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 486, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 486, + 121 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 486, + 121 + ], + "type": "text", + "content": "Pos3R: 6D Pose Estimation for Unseen Objects Made Easy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 471, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 471, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 471, + 158 + ], + "type": "text", + "content": "Weijian Deng" + }, + { + "bbox": [ + 138, + 142, + 471, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 158 + ], + "type": "text", + "content": " Dylan Campbell" + }, + { + "bbox": [ + 138, + 142, + 471, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 158 + ], + "type": "text", + "content": " Chunyi Sun" + }, + { + "bbox": [ + 138, + 142, + 471, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 158 + ], + "type": "text", + "content": " Jiahao Zhang" + }, + { + "bbox": [ + 138, + 142, + 471, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 157, + 456, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 157, + 456, + 171 + ], + "spans": [ + { + "bbox": [ + 153, + 157, + 456, + 171 + ], + "type": "text", + "content": "Shubham Kanitkar" + }, + { + "bbox": [ + 153, + 157, + 456, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 157, + 456, + 171 + ], + "type": "text", + "content": " Matthew E. Shaffer" + }, + { + "bbox": [ + 153, + 157, + 456, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 157, + 456, + 171 + ], + "type": "text", + "content": " Stephen Gould" + }, + { + "bbox": [ + 153, + 157, + 456, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 171, + 458, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 171, + 458, + 185 + ], + "spans": [ + { + "bbox": [ + 152, + 171, + 458, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 152, + 171, + 458, + 185 + ], + "type": "text", + "content": "Australian National University " + }, + { + "bbox": [ + 152, + 171, + 458, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 152, + 171, + 458, + 185 + ], + "type": "text", + "content": "RIOS Intelligent Machines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 185, + 520, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 185, + 520, + 200 + ], + "spans": [ + { + "bbox": [ + 82, + 185, + 520, + 200 + ], + "type": "inline_equation", + "content": "^{1}\\{\\text {firstname.lastname } \\} @ \\text { anu.edu.au } ^{2}\\{\\text {firstname.lastname } \\} @ \\text { rios.ai }" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 251, + 297, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 504 + ], + "type": "text", + "content": "Foundation models have significantly reduced the need for task-specific training, while also enhancing generalizability. However, state-of-the-art 6D pose estimators either require further training with pose supervision or neglect advances obtainable from 3D foundation models. The latter is a missed opportunity, since these models are better equipped to predict 3D-consistent features, which are of significant utility for the pose estimation task. To address this gap, we propose Pos3R, a method for estimating the 6D pose of any object from a single RGB image, making extensive use of a 3D reconstruction foundation model and requiring no additional training. We identify template selection as a particular bottleneck for existing methods that is significantly alleviated by the use of a 3D model, which can more easily distinguish between template poses than a 2D model. Despite its simplicity, Pos3R achieves competitive performance on the Benchmark for 6D Object Pose Estimation (BOP), matching or surpassing existing refinement-free methods. Additionally, Pos3R integrates seamlessly with render-and-compare refinement techniques, demonstrating adaptability for high-precision applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 525, + 137, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 137, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 137, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Six-dimensional (6D) object pose estimation—the task of determining the exact position and orientation of objects relative to a camera—is essential for applications in robotics, augmented reality, and autonomous systems. Reliable pose estimation enables critical tasks like object manipulation, grasping, and assembly, allowing these systems to interact effectively in complex and dynamic environments [6, 18, 29, 30, 49, 54]. Traditional approaches often rely on learning-based methods tailored to specific objects or categories, achieving high accuracy but struggling to generalize to new categories or unseen objects [4, 5, 7, 26, 33, 46, 47, 53-55, 59, 61]. This limitation is particularly problematic in dynamic, data-scarce environments where adaptability and flexibility are crucial." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 228, + 555, + 318 + ], + "blocks": [ + { + "bbox": [ + 317, + 228, + 555, + 318 + ], + "lines": [ + { + "bbox": [ + 317, + 228, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 555, + 318 + ], + "type": "image", + "image_path": "4a51ba4ad4b7ceaff74356db999e6f95204ed479b6381b88b278f7a244ad1c10.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 327, + 555, + 349 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 349 + ], + "type": "text", + "content": "Figure 1. Illustration of In-Plane and Out-of-Plane Rotations with Correspondence Quality between DINOv2 and MASt3R." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 350, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 449 + ], + "type": "text", + "content": "The left diagram illustrates in-plane rotations (within the image plane) and out-of-plane rotations (3D orientation changes) from a camera view. On the right, DINOv2 [40] (top row) and MASt3R [24] (bottom row) are compared in their handling of these rotations. DINOv2 exhibits sparse and inconsistent correspondences (shown by red lines), particularly under out-of-plane rotations, due to its 2D feature limitations. In contrast, MASt3R provides dense and stable correspondences across both types of rotations, reflecting its ability to produce 3D-consistent features." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 472, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 555, + 628 + ], + "type": "text", + "content": "To overcome these limitations, there has been a recent shift toward model-based approaches that aim to generalize pose estimation to unseen objects without object-specific training [2, 17, 41, 56]. These typically rely on a two-stage pipeline: detecting and localizing objects within a scene, followed by a render-and-compare process that matches detected object regions to a set of template models. Building on this concept, recent works have sought to improve generalization through large-scale model training with significant attention to object diversity. For example, Foundation-Pose [56] uses synthetic training data combined with a large language model and contrastive learning, enhancing feature alignment across varied domains for robust generalization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "text", + "content": "The recent emergence of training-free methods offers a promising alternative, enabling 6D pose estimation for unseen objects without object- or task-specific training. Foundation models like DINOv2 [40] have demonstrated strong zero-shot capabilities, capturing spatial and semantic details through learned features [1, 58]. FoundPose [41], for instance, leverages DINOv2 descriptors to bridge synthetic" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16818" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "and real data, achieving competitive results on the BOP challenge [17]. These advancements underscore the potential of training-free frameworks in 6D pose estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 312 + ], + "type": "text", + "content": "Building on this trend, we explore the strengths of 3D foundation models. Accurate pose estimation requires handling both in-plane and out-of-plane rotations, which are demonstrated in Fig. 1. In-plane rotations occur within the image plane, as when an object rotates around the camera's line of sight, while out-of-plane rotations involve changes in 3D orientation, such as tilting or turning, which significantly alters appearance due to perspective shifts. Addressing both types of rotation is essential for robust pose estimation, as objects often appear from varied angles. Although 2D foundation models like DINOv2 handle in-plane rotations well through training augmentations, they lack robustness to out-of-plane rotations due to their 2D limitations (shown in Fig. 1). In contrast, 3D foundation models, such as MASt3R [24], are specifically designed to produce 3D-consistent features across different viewpoints, enabling reliable feature alignment even with out-of-plane rotations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 315, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 458 + ], + "type": "text", + "content": "Motivated by this, we propose Pos3R, a training-free method for 6D pose estimation of unseen objects using only RGB inputs. Pos3R leverages the 3D foundation model MASt3R [24] for pose estimation without requiring additional training. The core of Pos3R lies in the image matching process between test crops and rendered templates from a CAD model. While common matchers like LoFTR [50] often struggle with synthetic-to-real matching due to domain gaps, MASt3R produces high-quality 2D correspondences in this case. These reliable matches enable Pos3R to establish the 2D-3D correspondences required for pose estimation using the PnP-RANSAC algorithm." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 686 + ], + "type": "text", + "content": "By harnessing MASt3R's ability to produce dense correspondences robust to viewpoint and illumination changes, we simplify template generation and achieve computational efficiency. Unlike existing methods that render hundreds of templates, we place camera positions to cover both in-plane and out-of-plane rotations. Specifically, we use eight base templates per object, positioned at the vertices of a cube centered on the CAD model, to capture out-of-plane rotations. For each base template, we generate five in-plane rotational templates along the object's principal axis, resulting in forty templates per object. To ensure efficient and accurate pose estimation, Pos3R dynamically selects the optimal rotation based on matching quality. This simple yet effective strategy maintains computational efficiency while achieving reliable pose estimation. Tested on seven diverse datasets in the BOP challenge, Pos3R demonstrates strong performance as a robust, scalable option for training-free 6D pose estimation of unseen objects, adaptable to render-and-compare refinement techniques." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 689, + 208, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 689, + 208, + 700 + ], + "spans": [ + { + "bbox": [ + 67, + 689, + 208, + 700 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "- While existing methods rely on 2D foundation models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "content": "for unseen object pose estimation, we present Pos3R, a training-free method that leverages the 3D foundation model MAST3R [24] to improve robustness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 108, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 180 + ], + "type": "text", + "content": "- Leveraging MASt3R's robust dense correspondences, Pos3R uses only forty strategically placed templates per object to capture in-plane and out-of-plane rotations. A simple selection technique based on correspondence quality ensures accurate matching, achieving strong performance on the BOP challenge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 194, + 400, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 194, + 400, + 206 + ], + "spans": [ + { + "bbox": [ + 314, + 194, + 400, + 206 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 216, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 358 + ], + "type": "text", + "content": "Seen Object Pose Estimation. Instance-level object pose estimation refers to the task of estimating the poses of specific objects previously encountered during model training [5, 7, 26, 33, 46, 54, 61]. Among common approaches are correspondence-based methods, which learn to identify precise alignments between input data and CAD models of the objects [7, 27, 46]. Other methods use template-based strategies, where the model selects the most similar pose-labeled template from a predefined set of examples [25, 35, 52]. Additionally, regression-based approaches directly predict object poses from the learned object-specific features [12, 19, 26, 42]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 361, + 553, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 553, + 515 + ], + "type": "text", + "content": "These instance-specific techniques offer high precision but typically require retraining for new object instances, limiting their generalizability. This limitation has led to the exploration of category-level methods, which generalize within specific object categories and allow estimation of unseen objects in known categories [4, 47, 53, 55, 59]. For example, SecondPose [4] enhances category-level 6D pose estimation by integrating object-specific geometric features with DINOv2 SE(3)-consistent semantic priors, effectively addressing intra-class shape variation. However, these methods struggle outside their target categories, while our approach enables pose estimation without relying on such constraints, allowing broader applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "content": "Unseen Object Pose Estimation. To enhance pose estimation flexibility, many methods aim to generalize across new object instances without object-specific training [22, 31, 37, 39, 43]. Approaches for this task can be broadly divided into manual reference view-based and CAD model-based methods. (i) Reference view-based methods use multi-view images with known poses as reference data for pose estimation [14, 31, 43, 51]. For example, OnePose [14, 51] reconstructs 3D object point clouds from posed RGB images, establishing 2D-3D correspondences to solve 6D poses. In a different approach, Nope [38], adopting the perspective of generating new views, trains models to predict discriminative embeddings for novel object perspectives. (ii) CAD model-based approaches leverage 3D object model to facilitate pose estimation. Some methods match features between the CAD model and the query image [20, 28, 60]," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16819" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 541, + 241 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 541, + 241 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 541, + 241 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 541, + 241 + ], + "type": "image", + "image_path": "f01f662c5b4ced6619d7971ccf11c13a4a43baacff99389e263fbea499a8d11e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 250, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 54, + 250, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 54, + 250, + 555, + 328 + ], + "type": "text", + "content": "Figure 2. Overview of the 6D Pose Estimation Process in Pos3R. (1) Template Rendering (Sec. 3.2.1): A CAD model is used to generate a set of forty templates, each representing a unique orientation that covers both in-plane and out-of-plane rotations. This is achieved by positioning the camera at the vertices of a cube around the object. (2) Image Matching (Sec. 3.2.2): Given a test segment from CNOS [37], we establish 2D-2D correspondences between the test image and each rendered template. We leverage the 3D-consistent features generated by MASt3R [24] for accurate matching and select the best template based on matching quality. Each template also includes a 3D object coordinate map that records the corresponding 3D points. (3) Pose Fitting (Sec. 3.2.3): Using the selected correspondences, we apply the PnP-RANSAC algorithm [23] to obtain a final pose that aligns the object with its observed position in the scene." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 348, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 348, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 348, + 296, + 515 + ], + "type": "text", + "content": "while others employ template matching to find the closest initial pose, refining it further using specialized refiners for higher accuracy [22, 34, 39]. For example, GigaPose [39] introduces an efficient two-network system: one network for retrieving template views (out-of-plane rotation) and another for estimating the remaining degrees of freedom (in-plane rotation and 3D translation). This separation reduces computational costs, contrasting with methods like MegaPose [22], which applies a single network to every possible test crop-template pair. Our work builds on the CAD model-based approach, aligning with recent BOP challenge protocols [17]. We focus on a training-free pipeline for unseen object pose estimation, a direction distinct from methods that require extensive training of task-specific networks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "type": "text", + "content": "Training-Free Object Pose Estimation. Traditionally, 6D object pose estimation is achieved by establishing 3D-to-2D correspondences followed by a " + }, + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{PnP}" + }, + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "type": "text", + "content": " algorithm [11, 13, 44, 61]. Recently, leveraging features from foundation models has gained traction, especially with models like DINO, which offer spatial detail and semantic consistency for pose estimation [1, 58]. FoundPose [41] uses DINOv2 descriptors to bridge synthetic and real data domains, providing strong performance for symmetric objects with RGB-only inputs. FreeZe integrates 3D point descriptors from GeDi [45] with image features from DINOv2 [40] for RGB-D pose estimation. By leveraging the 3D consistent feature of MAST3R to produce dense correspondences, our work explores training-free, unseen object pose estimation with RGB-only inputs, enabling effective pose estimation across diverse object categories without task-specific training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 347, + 490, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 490, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 490, + 360 + ], + "type": "text", + "content": "3.6D Pose Estimation with Pos3R" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 367, + 408, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 408, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 408, + 379 + ], + "type": "text", + "content": "3.1. Task Definition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": "Given a 3D model of a query object " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " and an RGB image " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": ", the task is to estimate the 6D pose of " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " relative to the camera's reference frame, with known intrinsics " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": ". Specifically, we aim to determine the 6DoF transformation " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = (\\mathbf{R}, \\mathbf{t})" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " in 3D space, where " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " represents the rotation matrix and " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " denotes the translation vector. The segmented object region " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_m = \\mathbf{M} \\odot \\mathbf{I}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " is created by element-wise multiplication of the binary segmentation mask " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": ", isolating the visible part of " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": ". We note that " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " may be partially occluded." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 506, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 566 + ], + "type": "text", + "content": "This work leverages the 3D foundation model MASt3R [24] to tackle model-based 6D pose estimation for unseen objects. We introduce Pos3R, a training-free, RGB-only method that effectively estimates 6D poses without object- or task-specific training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 574, + 443, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 443, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 443, + 588 + ], + "type": "text", + "content": "3.2. Training-Free Pipeline" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "type": "text", + "content": "Overview. Following the standard model-based pipeline for 6D pose estimation of unseen objects [2, 39, 41], Pos3R comprises two components: object detection and pose estimation. We keep each component frozen, avoiding any object- or task-specific training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "For object detection, we use CNOS [37] to generate segmentation masks and object identities for each target instance, enabling localization of the target segment " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_m" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " within the RGB image " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": ". As a default method in the BOP challenge for segmenting unseen objects, CNOS requires" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16820" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "only 3D models for onboarding and does not depend on additional data or object-specific training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 296, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 296, + 181 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 296, + 181 + ], + "type": "text", + "content": "For 6D pose estimation, we illustrate three steps in Fig. 2. Specifically, given a set of templates rendered from a textured CAD model, Pos3R utilizes the 3D foundation model MAST3R [24] to extract features from both the target segment " + }, + { + "bbox": [ + 55, + 97, + 296, + 181 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_m" + }, + { + "bbox": [ + 55, + 97, + 296, + 181 + ], + "type": "text", + "content": " and each template, enabling estimation of the 6D transformation " + }, + { + "bbox": [ + 55, + 97, + 296, + 181 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = (\\mathbf{R},\\mathbf{t})" + }, + { + "bbox": [ + 55, + 97, + 296, + 181 + ], + "type": "text", + "content": " through 2D-3D correspondences using the PnP-RANSAC algorithm." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 189, + 171, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 171, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 171, + 201 + ], + "type": "text", + "content": "3.2.1. Template Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 204, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 335 + ], + "type": "text", + "content": "Using a textured 3D CAD model, we render templates of the object from various orientations. The rendering process follows standard rasterization methods as described in [39], with a black background and fixed lighting. The rendering camera is configured with the same intrinsic parameters, " + }, + { + "bbox": [ + 55, + 204, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 204, + 296, + 335 + ], + "type": "text", + "content": ", as the test camera, and the rendered templates match the size of the test image " + }, + { + "bbox": [ + 55, + 204, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 204, + 296, + 335 + ], + "type": "text", + "content": ". The object remains centered in each template. Additionally, the 3D locations in the coordinate space of the 3D CAD model corresponding to each pixel in the rendered template are recorded, enabling the establishment of 2D-3D correspondences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 337, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 515 + ], + "type": "text", + "content": "A critical component of Pos3R is establishing correspondences between pixels in the target segment and the template. Accurate pose estimation requires addressing both in-plane and out-of-plane rotations, as objects are frequently viewed from diverse orientations. In-plane rotations occur within the image plane, such as when an object rotates around the camera's line of sight. Out-of-plane rotations, on the other hand, involve 3D orientation changes—such as tilting or turning—that significantly alter an object's appearance due to perspective shifts. Existing methods tackle this challenge through extensive template libraries and specialized selection mechanisms. For example, MegaPose [22] uses hundreds of templates to capture a broad range of object poses, but maintaining such a large template library necessitates a dedicated selection network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": "Template Configuration. To reduce reliance on extensive template libraries, Pos3R leverages the 3D foundation model MASt3R, which generates 3D-consistent features across viewpoints. This enables Pos3R to handle out-of-plane rotations effectively without requiring hundreds of templates or complex selection mechanisms. We use a set of eight base templates, denoted by " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i\\}_{i=1}^8" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": ", covering essential orientations. They are positioned at the vertices of a cube centered around the CAD model, effectively capturing out-of-plane rotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "To further address ambiguities from in-plane (axial) rotations, which can impact correspondence quality, we apply controlled rotational variations to each base template. For each template " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ", we generate " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " rotations around the camera's principal axis, each rotation defined by an angle " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\theta_k = \\frac{2\\pi k}{T}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "k = 0,\\dots,T - 1" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ", covering the full" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " range. In our experiments, we set " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "T = 5" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " to balance efficiency and accuracy, yielding a set of rotationally-augmented templates " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{i,k}\\}_{k=1}^{5}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": ". In the following, we discuss the template selection process in detail." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 126, + 412, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 412, + 138 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 412, + 138 + ], + "type": "text", + "content": "3.2.2. Image Matching" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": "MASt3R as an Image Matcher. Our approach builds upon MASt3R [24], a model for joint local 3D reconstruction and pixelwise matching between two input images " + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_a" + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_b \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": ". Conceptually, MASt3R operates as a mapping function " + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(\\mathbf{I}_a, \\mathbf{I}_b) = \\mathrm{Dec}(\\mathrm{Enc}(\\mathbf{I}_a), \\mathrm{Enc}(\\mathbf{I}_b))" + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": ", where the encoder " + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\mathrm{Enc}(\\mathbf{I}) \\to \\mathbf{F}" + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": " is a Siamese Vision Transformer (ViT) that processes image " + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": " into token vectors of size " + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "inline_equation", + "content": "N \\times d" + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "inline_equation", + "content": "N = w \\times h" + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": ". The decoder, " + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\mathrm{Dec}(\\mathbf{F}_a, \\mathbf{F}_b)" + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": ", employs twin ViT modules that produce pixelwise pointmaps " + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": ", local feature maps " + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 141, + 553, + 261 + ], + "type": "text", + "content": " for each image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "text", + "content": "Using these local feature representations, correspondences between images are identified via the fastNN algorithm [24]. It efficiently establishes reciprocal matches between the feature maps " + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_a" + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_b" + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "text", + "content": " by initially seeding points across a uniform pixel grid and iteratively refining these seeds to form high-quality mutual correspondences. Through this process, fastNN accurately aligns key features across images. The resulting reciprocal pixel pairs between " + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_a" + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_b" + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "text", + "content": " are represented as " + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "inline_equation", + "content": "M_{a,b} = \\{(y_a^c,y_b^c)\\}_{c = 1}^{|M_{a,b}|}" + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_a^c" + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_b^c\\in \\mathbb{N}^2" + }, + { + "bbox": [ + 313, + 262, + 554, + 396 + ], + "type": "text", + "content": " denote the coordinates of matched pixels in each respective image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 399, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 543 + ], + "type": "text", + "content": "Similarity-Based Template Selection. Unlike methods that require hundreds of templates [22, 36, 41], Pos3R simplifies template selection by needing only forty templates, significantly reducing the complexity of the selection process. Rather than relying on a trained template selection network, we employ a straightforward, training-free approach based on the similarity of matched correspondences. Specifically, given a target segment " + }, + { + "bbox": [ + 313, + 399, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_m" + }, + { + "bbox": [ + 313, + 399, + 554, + 543 + ], + "type": "text", + "content": " and a set of eight base templates " + }, + { + "bbox": [ + 313, + 399, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i\\}_{i=1}^8" + }, + { + "bbox": [ + 313, + 399, + 554, + 543 + ], + "type": "text", + "content": ", each augmented with rotational variations " + }, + { + "bbox": [ + 313, + 399, + 554, + 543 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{i,k}\\}_{k=1}^5" + }, + { + "bbox": [ + 313, + 399, + 554, + 543 + ], + "type": "text", + "content": ", we identify the most similar template for pose estimation by calculating correspondences between the target segment and each template." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "text", + "content": "For each rotationally-augmented template " + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i,k}" + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "text", + "content": ", we obtain reciprocal pixel pairs between " + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_m" + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i,k}" + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "text", + "content": ", represented as: " + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "inline_equation", + "content": "M_{m,i,k} = \\{(y_m^p, y_{i,k}^p)\\}_{p=1}^{|M_{m,i,k}|}" + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "inline_equation", + "content": "M_{m,i,k}" + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "text", + "content": " is the set of all reciprocal pixel pairs between " + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_m" + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i,k}" + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_m^p" + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{i,k}^p \\in \\mathbb{N}^2" + }, + { + "bbox": [ + 313, + 544, + 554, + 617 + ], + "type": "text", + "content": " denote the coordinates of matched pixels in the target segment and the template variant, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "For each matched pair " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{y}_m^p, \\mathbf{y}_{i,k}^p)" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "M_{m,i,k}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": ", we retrieve the corresponding local features from the feature maps " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_m" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{i,k}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " generated by MASt3R. Specifically, we obtain the feature vector " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_m^p" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " at coordinate " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_m^p" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " in the target segment from " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_m" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": ", and the feature vector " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{i,k}^p" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " at coordinate " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{i,k}^p" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " in the template from " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{i,k}" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": ". We compute the feature similarity for each matched pair as: " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "S(\\mathbf{f}_m^p, \\mathbf{f}_{i,k}^p) = \\mathbf{f}_m^p \\cdot \\mathbf{f}_{i,k}^p" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " denotes the dot product." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16821" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "To calculate an overall similarity score between the target segment " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_m" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": " and each template variant " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i,k}" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": ", we aggregate the similarity scores across all matched pairs as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 101, + 118, + 295, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 118, + 295, + 153 + ], + "spans": [ + { + "bbox": [ + 101, + 118, + 295, + 153 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} \\left(\\mathbf {I} _ {m}, \\mathbf {I} _ {i, k}\\right) = \\sum_ {p = 1} ^ {\\left| M _ {m, i, k} \\right|} S \\left(\\mathbf {f} _ {m} ^ {p}, \\mathbf {f} _ {i, k} ^ {p}\\right), \\tag {1}", + "image_path": "6231add656687e89038cb392c136bfc780b1055b4e65435ee9d7ceb27c8a3488.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 161, + 296, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 296, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 161, + 296, + 198 + ], + "type": "inline_equation", + "content": "M_{m,i,k}" + }, + { + "bbox": [ + 55, + 161, + 296, + 198 + ], + "type": "text", + "content": " is the set of all matched pairs between " + }, + { + "bbox": [ + 55, + 161, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_m" + }, + { + "bbox": [ + 55, + 161, + 296, + 198 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 161, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i,k}" + }, + { + "bbox": [ + 55, + 161, + 296, + 198 + ], + "type": "text", + "content": ". After computing " + }, + { + "bbox": [ + 55, + 161, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\mathbf{I}_m,\\mathbf{I}_{i,k})" + }, + { + "bbox": [ + 55, + 161, + 296, + 198 + ], + "type": "text", + "content": " for each pair, we select the template with the highest similarity score:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 207, + 295, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 207, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 80, + 207, + 295, + 229 + ], + "type": "interline_equation", + "content": "\\left(i _ {\\text {o p t}}, k _ {\\text {o p t}}\\right) = \\underset {i \\in \\{1, \\dots , 8 \\}, k \\in \\{1, \\dots , 5 \\}} {\\arg \\max } \\operatorname {s i m} \\left(\\mathbf {I} _ {m}, \\mathbf {I} _ {i, k}\\right). \\tag {2}", + "image_path": "f0518fb02f8e92a01e0b5b4a7ce785615bb71ff29b43ae65a1ed1ad0028f9048.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 237, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 262 + ], + "type": "text", + "content": "The selected template " + }, + { + "bbox": [ + 55, + 237, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i_{\\mathrm{opt}},k_{\\mathrm{opt}}}" + }, + { + "bbox": [ + 55, + 237, + 296, + 262 + ], + "type": "text", + "content": " is then used as the closest match to the target segment for the pose estimation process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 267, + 135, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 135, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 135, + 280 + ], + "type": "text", + "content": "3.2.3. Pose Fitting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "content": "After selecting the suitable template " + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{i_{\\mathrm{opt}},k_{\\mathrm{opt}}}" + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "content": ", we proceed to estimate pose " + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_m = (\\mathbf{R}_m,\\mathbf{t}_m)" + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_m" + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "content": " is a 3D rotation matrix and " + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_m" + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "content": " is a 3D translation vector that transforms the object from model space to camera space. It relies on a set of 2D-3D correspondences " + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{t_{\\mathrm{final}}} = \\{(\\mathbf{y}_m^j,\\mathbf{P}^j)\\}_{j = 1}^{|M|}" + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_m^j\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "content": " represents the coordinates of matched pixels in the target segment " + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_m" + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^j\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "content": " denotes the corresponding 3D points in model space from the selected template, and there are " + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "inline_equation", + "content": "|M|" + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "content": " pairs. To determine " + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_m" + }, + { + "bbox": [ + 55, + 282, + 296, + 417 + ], + "type": "text", + "content": ", we solve the Perspective-n-Point (PnP) problem, which minimizes the reprojection error:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 92, + 426, + 295, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 426, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 92, + 426, + 295, + 462 + ], + "type": "interline_equation", + "content": "\\underset {\\mathbf {R} _ {m}, \\mathbf {t} _ {m}} {\\arg \\min } \\sum_ {j = 1} ^ {| M |} \\left\\| \\mathbf {y} _ {m} ^ {j} - \\pi \\left(\\mathbf {R} _ {m} \\mathbf {P} ^ {j} + \\mathbf {t} _ {m}\\right) \\right\\| ^ {2}, \\tag {3}", + "image_path": "dd5aa1f7442166bbd6915d781a71abff6fe8fedd4b84ae6ed6a23662c83aff52.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "text", + "content": " is the projection function that maps 3D points to 2D image points according to the camera intrinsics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 494, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 590 + ], + "type": "text", + "content": "To enhance robustness against outliers, we employ the Efficient PnP (EPnP) algorithm [23] combined with a RANSAC-based fitting strategy [11]. In this approach, PnP is applied iteratively to random subsets of four correspondences from " + }, + { + "bbox": [ + 55, + 494, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{t_{\\mathrm{final}}}" + }, + { + "bbox": [ + 55, + 494, + 296, + 590 + ], + "type": "text", + "content": ", generating multiple pose hypotheses. For each hypothesis, we count the number of inliers, defined as correspondences where the reprojection error falls below a predefined threshold " + }, + { + "bbox": [ + 55, + 494, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 494, + 296, + 590 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 598, + 295, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 598, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 70, + 598, + 295, + 613 + ], + "type": "interline_equation", + "content": "\\text {i n l i e r s} = \\left| \\left\\{j: \\left\\| \\mathbf {y} _ {m} ^ {j} - \\pi \\left(\\mathbf {R} _ {m} \\mathbf {P} ^ {j} + \\mathbf {t} _ {m}\\right) \\right\\| < \\epsilon \\right\\} \\right|. \\tag {4}", + "image_path": "fc120930350e0ea98f242ff4439e8fd324d77f2d666cdc3428da2cffa988bbd7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 622, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 647 + ], + "type": "text", + "content": "The hypothesis with the highest inlier count is selected as the final coarse pose estimate " + }, + { + "bbox": [ + 55, + 622, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_m" + }, + { + "bbox": [ + 55, + 622, + 296, + 647 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 658, + 137, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 137, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 137, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In this section, we first outline the experimental setup (Section 4.1). We then evaluate our method's performance in comparison with previous approaches on the seven core" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "datasets of the BOP challenge [17], examining accuracy, runtime efficiency, and effectiveness when using predicted 3D models (Section 4.2). This evaluation highlights the strengths and unique contributions of our approach. Lastly, we present an ablation study to investigate the impact of various configurations in our method (Section 4.3)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 154, + 430, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 430, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 430, + 167 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 171, + 555, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 292 + ], + "type": "text", + "content": "Evaluation Datasets. We evaluate our approach on the seven core datasets of the BOP challenge [17]: LineMod Occlusion (LM-O) [3], T-LESS [15], TUD-L [16], IC-BIN [8], ITODD [9], HomebrewedDB (HB)[21], and YCB-Video (YCB-V) [57]. Together, these datasets feature 132 distinct objects and 19,048 testing instances, each in complex, cluttered scenes with partial occlusions. Table 1 provides the instance count for each dataset. It is worth noting that the unseen object pose estimation task remains challenging, with significant room for improvement." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": "Evaluation Metrics. We use the BOP evaluation protocol [17] for 6D object localization, which assesses pose accuracy with three error metrics: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). VSD evaluates only the visible part of the object to handle ambiguities, MSSD measures 3D surface deviation with global symmetries, and MSPD assesses perceivable deviation using object symmetries in projection. A pose is considered correct for a metric " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "e < \\theta_e" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\theta_e" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": " as the correctness threshold. The Average Recall (AR) for each metric " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": ", denoted " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_e" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": ", is the mean Recall over various " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\theta_e" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": " thresholds, and for VSD, multiple misalignment tolerances " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": ". The overall AR score is the average of the three metrics: " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathrm{AR} = (\\mathrm{AR}_{\\mathrm{VSD}} + \\mathrm{AR}_{\\mathrm{MSSD}} + \\mathrm{AR}_{\\mathrm{MSPD}}) / 3" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 482, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 554, + 602 + ], + "type": "text", + "content": "Pose Refinement. To show Pos3R can integrate with render-and-compare refinement techniques, we apply the refinement method from MegaPose [22] to our result. Given an input image and an estimated pose, the refiner predicts the relative transformation between the initial and ground-truth pose. The refiner is trained on a large-scale dataset with several well-designed techniques [22]. We use the similarity score Eq. 1 to select the top-5 templates and each test crop-template pair gives a pose with EPnP. Then, we use the refiner to update these five pose hypotheses." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 612, + 514, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 514, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 514, + 624 + ], + "type": "text", + "content": "4.2. Comparison With the State of the Art" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "We compare Pos3R with both training-free and training-based methods. The training-free methods include Found-Pose [41] and ZS6D [2], while the training-based methods include MegaPose [22], GigaPose [39], OSOP [48], and GenFlow [34]. For pose refinement, we apply the refiner from MegaPose [22]. We use the publicly available code from [39] with the default hyperparameters." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16822" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 555, + 305 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 305 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 305 + ], + "type": "table", + "html": "
#MethodTraining-FreeRefinementDatasets (num. instances)MeanTime
LM-OT-LESSTUD-LIC-BINITODDHBYCB-V
144564236001786304116304123
Coarse Pose Estimation:
1GigaPose [39]X-29.927.330.223.118.834.829.027.60.9
2GenFlow [34]X-25.021.530.016.815.428.327.723.53.8
3MegaPose [22]X-22.917.725.815.210.825.128.120.815.5
4OSOP [48]X-31.2----49.233.2--
5ZS6D [2]✓-29.821.0----32.4--
6FoundPose [41]✓-39.633.846.723.920.450.845.237.21.7
7Baseline✓-30.523.643.221.415.442.549.332.30.7
7Pos3R (Ours)✓-32.331.547.333.125.153.753.939.51.4
With Pose Refinement (5 hypotheses):
8FoundPoseXMegaPose58.654.965.744.436.170.367.356.811.2
9GigaPoseXMegaPose59.957.064.546.739.772.266.357.97.3
10GenFlowXGenFlow56.352.368.445.339.573.963.357.120.9
11MegaPoseXMegaPose56.050.768.441.433.870.462.154.747.4
12Pos3R (ours)XMegaPose56.557.566.846.037.570.166.457.38.0
", + "image_path": "240f26f9858eeda1cbbd3b35e523c11fb1db5ea9ef6d2e0b53fcb0127cf4f8f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 313, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 555, + 413 + ], + "type": "text", + "content": "Table 1. Performance Comparison on the Seven Datasets of BOP. This table reports the Average Recall (AR) scores per dataset, the mean AR score across all datasets, and the time required to estimate poses for all objects in an image (in seconds). The runtime data of other methods are sourced from FoundPose [41]. The upper section lists methods for coarse pose estimation without refinement, while the lower section presents methods that incorporate a refinement stage using multiple pose hypotheses, reporting the best refined pose. Methods without task-specific training are marked with a green check mark " + }, + { + "bbox": [ + 55, + 313, + 555, + 413 + ], + "type": "inline_equation", + "content": "(\\checkmark)" + }, + { + "bbox": [ + 55, + 313, + 555, + 413 + ], + "type": "text", + "content": ". In the coarse estimation category, Pos3R achieves the highest AR scores on several datasets and demonstrates good overall generalization with a competitive mean AR score. While refinement-based methods yield slightly higher accuracy in certain cases, Pos3R remains a robust and efficient choice for coarse pose estimation and performs comparably to top refined methods when combined with a refinement stage. The best results in the coarse estimation section are highlighted in green, and the best results in the refinement section are highlighted in blue. The second-best results are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 426, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 535 + ], + "type": "text", + "content": "Table 1 provides a detailed comparison of Pos3R (Pos3R) with other leading 6D pose estimation methods across the seven primary datasets in the BOP challenge. The upper portion of the table showcases results for coarse pose estimation without refinement, while the lower portion highlights performance after incorporating a pose refinement step with five hypotheses. It is important to note that pose refinement involves task-specific training on extensive pose datasets, making it non-training-free [22]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 540, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 684 + ], + "type": "text", + "content": "In the coarse pose estimation category, Pos3R demonstrates superior performance over other training-free methods, including FoundPose and ZS6D, achieving the highest mean accuracy and consistently strong results across most datasets. For instance, Pos3R attains the highest AR (Average Recall) on the TUD-L, HB, and YCB-V datasets, along with a mean AR of 39.5. This high performance across varied datasets reflects Pos3R's robust generalization capabilities for different object poses and scenarios. Additionally, with a runtime of only 1.4 seconds, Pos3R stands out as an efficient alternative to methods that either require substantial computational resources or are not training-free." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Pos3R remains competitive after pose refinement using MegaPose's refine. Although Pos3R is not inher" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 426, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 510 + ], + "type": "text", + "content": "ently designed for refinement, it achieves comparable accuracy to MegaPose and other refined methods across several datasets. These findings underscore the adaptability of Pos3R, as it provides accurate initial pose estimates and benefits from additional refinement techniques—demonstrating its potential for scalable, practical, and flexible applications in 6D pose estimation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 510, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 666 + ], + "type": "text", + "content": "Moreover, we observe that Pos3R faces challenges in achieving competitive accuracy on the LM-O dataset, where significant occlusion poses a major difficulty. Occlusion disrupts the initial image matching process, often resulting in suboptimal pose estimates that refinement methods struggle to correct. FoundPose [41] addresses this issue by employing a bag-of-words strategy for template selection, which has shown greater resilience to occlusion. To improve robustness in such scenarios, incorporating the local contrastive learning approach used by GigaPose [39] could be a promising direction. This method is specifically designed to enhance feature matching in occluded regions, and we consider this an area for future exploration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 672, + 434, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 434, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 434, + 685 + ], + "type": "text", + "content": "4.3. Component Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Pose Estimation Using Predicted 3D Models. Model-based methods for unseen object pose estimation typically" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16823" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 123, + 296, + 182 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 2. Comparison with Predicted 3D Models on LM-O [3]. Average Recall is reported. All methods use 3D models predicted by Wonder3D [32]. Results with pose refinement, applied to the generated 3D models, are also included for comparison." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 123, + 296, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 123, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 123, + 296, + 182 + ], + "type": "table", + "html": "
#MethodSingle Image-to-3DGT 3D model w/o refinement
CoarseRefined
1MegaPose [22]15.425.222.7
2GigaPose [39]17.627.229.4
3Pos3R19.528.532.3
", + "image_path": "3f42a79633c2fa19bb7a16d6447db170c1fc6b55d9c77d494efffd2d142205c5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 69, + 186, + 286, + 329 + ], + "blocks": [ + { + "bbox": [ + 69, + 186, + 286, + 329 + ], + "lines": [ + { + "bbox": [ + 69, + 186, + 286, + 329 + ], + "spans": [ + { + "bbox": [ + 69, + 186, + 286, + 329 + ], + "type": "image", + "image_path": "4b1f91f54bc77e1fc3af0bbfd20e32af7bf2e860c1382c3e8bfe94551275a14c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 337, + 297, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 337, + 297, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 297, + 381 + ], + "type": "text", + "content": "Figure 3. Rendered Images and 3D Coordinate Maps from Wonder3D and CAD Model. The top two rows show rendered images and 3D coordinate maps from Wonder3D [32], while the bottom two rows show those from the real CAD model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 397, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 397, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 397, + 295, + 577 + ], + "type": "text", + "content": "require accurate, textured 3D models. To relax this requirement, we follow the approach in [39] and use 3D models predicted by Wonder3D [32] from a single reference image. The predicted 3D models use the same object coordinate system as the ground truth (GT) frame, ensuring aligned axes. We then assess the performance of our method using these reconstructed models in place of the high-fidelity CAD models provided by the dataset. For a fair comparison, we adopt the same settings as [39], including the reference image for Wonder3D and post-processing steps, and report results on the LM-O dataset [3]. Examples of rendered views from generated 3D model are shown in Fig. 3. As shown in Table 2, Pos3R achieves higher AR scores than both MegaPose [22] and GigaPose [39], both before and after pose refinement, showing the effectiveness of Pos3R." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": "Template Selection Technique. In our experiments, we use the similarity of correspondence matches (Eq. 1) as the primary method for template selection. Additionally, we consider two alternative approaches: 1) Inliers: selecting templates based on the number of inliers (Eq. 4); 2) Confidence: leveraging the 3D pointmaps and confidence maps provided by MASt3R for each target segment-template pair [24], where the average confidence score is used for selection. As illustrated in Fig. 4, template selection based on inliers proves effective, while selection based on similarity score yields the best performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 70, + 555, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 170 + ], + "type": "text", + "content": "Table 3. Comparison of Pose Estimation Method. We evaluate our proposed method (Pos3R), a variant without axial rotation augmentation (" + }, + { + "bbox": [ + 313, + 70, + 555, + 170 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 313, + 70, + 555, + 170 + ], + "type": "text", + "content": " axial rotation), a variant that use face centers as base templates, and a version that replaces the 3D foundation model MASt3R with the 2D foundation model DINOv2. We also replace MASt3R with another dense image matcher RoMa [10]. Pos3R achieves the highest accuracy across all datasets, highlighting the importance of axial rotation handling and the advantage of using 3D foundation models over DINOv2." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 178, + 555, + 254 + ], + "blocks": [ + { + "bbox": [ + 315, + 178, + 555, + 254 + ], + "lines": [ + { + "bbox": [ + 315, + 178, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 315, + 178, + 555, + 254 + ], + "type": "table", + "html": "
#MethodLM-OTUD-LYCB-VT-LessIC-BINITODDHB
1Pos3R32.347.353.931.533.125.153.7
2w/o Ax. Rot.30.543.249.323.621.415.442.5
3Face Center31.545.553.328.029.422.452.5
4DINOv2 (L11)18.525.022.514.215.614.132.1
5DINOv2 (L9)20.526.024.015.616.614.832.9
6RoMa [10]20.825.032.016.217.015.525.1
", + "image_path": "be0c224017e3467454bfb11da8ede080c9ca31e6a29e45c743397bf4fa2fd40b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 266, + 553, + 339 + ], + "blocks": [ + { + "bbox": [ + 315, + 266, + 553, + 339 + ], + "lines": [ + { + "bbox": [ + 315, + 266, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 315, + 266, + 553, + 339 + ], + "type": "image", + "image_path": "f8a399f8d81c565c2614b700e53b5e13b756594a42fcc14314d522266eed4fc2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 347, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 392 + ], + "type": "text", + "content": "Figure 4. Comparison of Template Selection Techniques. We compare template selection methods based on confidence, inliers, and similarity score. The similarity score consistently achieves the highest average recall (AR) across all datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 415, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 715 + ], + "type": "text", + "content": "Impact of In-Plane Rotation and 3D Consistency on Pose Estimation. Table 3 evaluates the impact of incorporating in-plane (axial) rotations and 3D-consistent features on pose estimation performance across seven datasets. Our method, Pos3R (row 1), achieves the highest accuracy on all datasets by combining controlled in-plane rotations with the 3D-consistent features provided by the MASt3R foundation model. In contrast, row 2 presents results without in-plane rotation, where only eight base templates are used to account for out-of-plane rotations. The performance drops considerably, highlighting the importance of addressing in-plane rotations for improving correspondence quality and overall pose accuracy. Rows 4 and 5 present a further variation in which we replace MASt3R with the 2D foundation model DINOv2. Despite following the same pipeline (e.g., using the same 40 templates), this substitution leads to a notable performance decline across all datasets. Additionally, in row 6, replacing MASt3R with the dense image matcher RoMa [10] also results in lower accuracy. We speculate that MASt3R's improved performance stems from its large-scale and 3D-aware training, which may enhance its ability to produce features that are well-suited for pose estimation. The above analysis highlights the importance of incorporating both in-plane rotations and 3D-consistent features to achieve robust and accurate 6D pose estimation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16824" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 70, + 193, + 157 + ], + "blocks": [ + { + "bbox": [ + 62, + 102, + 72, + 125 + ], + "lines": [ + { + "bbox": [ + 62, + 102, + 72, + 125 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 72, + 125 + ], + "type": "text", + "content": "YCBV" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 70, + 193, + 157 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 193, + 157 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 193, + 157 + ], + "type": "image", + "image_path": "7709a89e402e4f20ffd30465c93613c0c1265eb2787a74991c4d2fabd9e091d0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 195, + 70, + 310, + 157 + ], + "blocks": [ + { + "bbox": [ + 195, + 70, + 310, + 157 + ], + "lines": [ + { + "bbox": [ + 195, + 70, + 310, + 157 + ], + "spans": [ + { + "bbox": [ + 195, + 70, + 310, + 157 + ], + "type": "image", + "image_path": "220a39336f50664e06af3f63b1e46c4395e6804d30f63768ebc87830171bab76.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 314, + 70, + 430, + 157 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 430, + 157 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 430, + 157 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 430, + 157 + ], + "type": "image", + "image_path": "6e242a91145d8e673b8c55d71fad0250ed17726fe09f2b109fd37c947f07f2af.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 433, + 70, + 548, + 157 + ], + "blocks": [ + { + "bbox": [ + 433, + 70, + 548, + 157 + ], + "lines": [ + { + "bbox": [ + 433, + 70, + 548, + 157 + ], + "spans": [ + { + "bbox": [ + 433, + 70, + 548, + 157 + ], + "type": "image", + "image_path": "4746bd19ad495afe185abe8da093ae570f45bf15cb0a5d3543a578782745e87a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 78, + 161, + 192, + 248 + ], + "blocks": [ + { + "bbox": [ + 62, + 182, + 72, + 205 + ], + "lines": [ + { + "bbox": [ + 62, + 182, + 72, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 72, + 205 + ], + "type": "text", + "content": "LM-O" + } + ] + } + ], + "index": 5, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 161, + 192, + 248 + ], + "lines": [ + { + "bbox": [ + 78, + 161, + 192, + 248 + ], + "spans": [ + { + "bbox": [ + 78, + 161, + 192, + 248 + ], + "type": "image", + "image_path": "5cf411d0aa5e04f1b209a24cd092e06d026ab70dd0b05339f0a520da4d2e447a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 195, + 161, + 310, + 248 + ], + "blocks": [ + { + "bbox": [ + 195, + 161, + 310, + 248 + ], + "lines": [ + { + "bbox": [ + 195, + 161, + 310, + 248 + ], + "spans": [ + { + "bbox": [ + 195, + 161, + 310, + 248 + ], + "type": "image", + "image_path": "c342bd184f62cc991d6dedb7d2227046431709fdd4e5aa68b1d2fa38b0feb902.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 314, + 161, + 430, + 248 + ], + "blocks": [ + { + "bbox": [ + 314, + 161, + 430, + 248 + ], + "lines": [ + { + "bbox": [ + 314, + 161, + 430, + 248 + ], + "spans": [ + { + "bbox": [ + 314, + 161, + 430, + 248 + ], + "type": "image", + "image_path": "7436092b2d84476e8a73dfcffd2a99e017986aa1e0c9f741e413f8320f63fbff.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 433, + 161, + 548, + 248 + ], + "blocks": [ + { + "bbox": [ + 433, + 161, + 548, + 248 + ], + "lines": [ + { + "bbox": [ + 433, + 161, + 548, + 248 + ], + "spans": [ + { + "bbox": [ + 433, + 161, + 548, + 248 + ], + "type": "image", + "image_path": "8265ead8de098df5edfe4ff46a2c706a113a1fdcb1c800be96bdd5107b7f4e1a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 77, + 251, + 192, + 338 + ], + "blocks": [ + { + "bbox": [ + 62, + 282, + 72, + 308 + ], + "lines": [ + { + "bbox": [ + 62, + 282, + 72, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 72, + 308 + ], + "type": "text", + "content": "IC-BIN" + } + ] + } + ], + "index": 10, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 251, + 192, + 338 + ], + "lines": [ + { + "bbox": [ + 77, + 251, + 192, + 338 + ], + "spans": [ + { + "bbox": [ + 77, + 251, + 192, + 338 + ], + "type": "image", + "image_path": "c82b945bb82ddc732ddc98923c84d1a744160525024a5ee617d86bffa4534cc6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 195, + 252, + 310, + 339 + ], + "blocks": [ + { + "bbox": [ + 195, + 252, + 310, + 339 + ], + "lines": [ + { + "bbox": [ + 195, + 252, + 310, + 339 + ], + "spans": [ + { + "bbox": [ + 195, + 252, + 310, + 339 + ], + "type": "image", + "image_path": "8e6f7634359499e72ef07d103faa2a6d91232dbf7c85e206cdaad2efb572a864.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 314, + 252, + 430, + 339 + ], + "blocks": [ + { + "bbox": [ + 314, + 252, + 430, + 339 + ], + "lines": [ + { + "bbox": [ + 314, + 252, + 430, + 339 + ], + "spans": [ + { + "bbox": [ + 314, + 252, + 430, + 339 + ], + "type": "image", + "image_path": "3ae717336c2440f9995024939bd955d98cffe0d138835cee9007bdb6299d3a45.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 433, + 252, + 548, + 339 + ], + "blocks": [ + { + "bbox": [ + 433, + 252, + 548, + 339 + ], + "lines": [ + { + "bbox": [ + 433, + 252, + 548, + 339 + ], + "spans": [ + { + "bbox": [ + 433, + 252, + 548, + 339 + ], + "type": "image", + "image_path": "273a492fcb79ed9018c185b7246d278cf98e929f8940440728e063a12db6afe3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 77, + 342, + 192, + 430 + ], + "blocks": [ + { + "bbox": [ + 62, + 373, + 72, + 400 + ], + "lines": [ + { + "bbox": [ + 62, + 373, + 72, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 72, + 400 + ], + "type": "text", + "content": "T-LESS" + } + ] + } + ], + "index": 15, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 342, + 192, + 430 + ], + "lines": [ + { + "bbox": [ + 77, + 342, + 192, + 430 + ], + "spans": [ + { + "bbox": [ + 77, + 342, + 192, + 430 + ], + "type": "image", + "image_path": "793280ee2a1334859367abb45b377c5d6b5d2a33f3f4a28d98c86e4d362dfbbc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 195, + 342, + 310, + 430 + ], + "blocks": [ + { + "bbox": [ + 195, + 342, + 310, + 430 + ], + "lines": [ + { + "bbox": [ + 195, + 342, + 310, + 430 + ], + "spans": [ + { + "bbox": [ + 195, + 342, + 310, + 430 + ], + "type": "image", + "image_path": "192e63f0b688d5785574f2e0044d73e42d63a37d3c30467c8a174be424d8eb34.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 437, + 555, + 494 + ], + "lines": [ + { + "bbox": [ + 55, + 437, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 555, + 494 + ], + "type": "text", + "content": "Figure 5. Qualitative Results of 6D Pose Estimation. Visualization of 6D object pose estimation of Pos3R across LM-O [3], T-LESS [15], IC-BIN [8], and YCB-V [57]. Pos3R effectively handles a variety of objects and maintains accuracy even in very crowded scenes (e.g., IC-BIN), where predictions remain reasonably robust. Furthermore, Pos3R adapts well to texture-less objects, as demonstrated in the T-LESS dataset. However, as marked by red crosses " + }, + { + "bbox": [ + 55, + 437, + 555, + 494 + ], + "type": "inline_equation", + "content": "(\\times)" + }, + { + "bbox": [ + 55, + 437, + 555, + 494 + ], + "type": "text", + "content": ", heavy occlusion poses a challenge, with reduced accuracy in heavily occluded regions, illustrating both the strengths and limitations of our approach." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 314, + 342, + 430, + 430 + ], + "blocks": [ + { + "bbox": [ + 314, + 342, + 430, + 430 + ], + "lines": [ + { + "bbox": [ + 314, + 342, + 430, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 342, + 430, + 430 + ], + "type": "image", + "image_path": "4974464d2c96da5462c928acb4c403a46878563d75ef2be5403bdbcb2f495256.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 433, + 342, + 548, + 430 + ], + "blocks": [ + { + "bbox": [ + 433, + 342, + 548, + 430 + ], + "lines": [ + { + "bbox": [ + 433, + 342, + 548, + 430 + ], + "spans": [ + { + "bbox": [ + 433, + 342, + 548, + 430 + ], + "type": "image", + "image_path": "c7339eb09b04f2cba68c29f2dfd067925e3c78f3f277ecd4cf39906a93f58019.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 507, + 295, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 710 + ], + "type": "text", + "content": "Qualitative Results of 6D Pose Estimation. Figure 5 shows pose estimation results of Pos3R across challenging datasets, including LM-O [3], T-LESS [15], IC-BIN [8], and YCB-V [57]. Pos3R demonstrates robust performance across diverse object types, sizes, and textures, handling complex environments effectively. In cluttered scenes like IC-BIN, Pos3R accurately localizes multiple, tightly packed objects. It also performs well on textureless objects in T-LESS, a challenge for vision-based methods due to minimal surface features. However, Figure 5 also reveals a limitation: in heavily occluded scenes, marked by red crosses, Pos3R cannot well handle. Reduced visibility of occluded objects results in deviations in the predicted bounding boxes. This limitation suggests future directions for improvement, such as integrating occlusion-aware methods (e.g., GigaPose [39]) or leveraging multi-view information to enhance robustness in these challenging cases." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 506, + 388, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 506, + 388, + 518 + ], + "spans": [ + { + "bbox": [ + 314, + 506, + 388, + 518 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "This work introduces Pos3R, a training-free, RGB-only framework for 6D pose estimation of unseen objects. By leveraging the 3D foundation model MASt3R, Pos3R generates robust, 3D-consistent features that effectively handle both in-plane and out-of-plane rotations. Without relying on extensive datasets or object-specific training, Pos3R establishes a strong baseline for training-free research. Using a minimal set of templates captured from eight cube vertex viewpoints with controlled rotational variations, Pos3R achieves accurate and efficient pose estimation. Experiments on the BOP challenge demonstrate that Pos3R outperforms other training-free methods in coarse pose estimation and achieves competitive results with refined methods when combined with MegaPose refinement. Future work will focus on incorporating occlusion-aware techniques to enhance robustness in occluded settings." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16825" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "Acknowledgments. We thank all reviewers and ACs for their constructive comments. This work was generously supported through research collaboration with RIOS Intelligent Machines." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 120, + 115, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 120, + 115, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 120, + 115, + 133 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 140, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 140, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 295, + 173 + ], + "type": "text", + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 175, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 295, + 229 + ], + "type": "text", + "content": "[2] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 463-469. IEEE, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 231, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 231, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 231, + 294, + 297 + ], + "type": "text", + "content": "[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 536-551. Springer, 2014. 5, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 298, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 298, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 298, + 295, + 373 + ], + "type": "text", + "content": "[4] Yamei Chen, Yan Di, Guangyao Zhai, Fabian Manhardt, Chenyangguang Zhang, Ruida Zhang, Federico Tombari, Nassir Navab, and Benjamin Busam. Secondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9959-9969, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 376, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 295, + 419 + ], + "type": "text", + "content": "[5] Xinke Deng, Arsalan Mousavian, Yu Xiang, Fei Xia, Timothy Bretl, and Dieter Fox. Poserbpf: A rao-blackwellized particle filter for 6-d object pose tracking. IEEE Transactions on Robotics, 37(5):1328-1342, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 421, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 295, + 475 + ], + "type": "text", + "content": "[6] Francesco Di Felice, Salvatore D'Avella, Alberto Remus, Paolo Tripicchio, and Carlo Alberto Avizzano. One-shot imitation learning with graph neural networks for pick-and-place manipulation tasks. IEEE Robotics and Automation Letters, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "type": "text", + "content": "[7] Bardia Doosti, Shujon Naha, Majid Mirbagheri, and David J Crandall. Hope-net: A graph-based model for hand-object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6608-6617, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 534, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 295, + 588 + ], + "type": "text", + "content": "[8] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3583-3592, 2016. 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 590, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 645 + ], + "type": "text", + "content": "[9] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec itodd-a dataset for 3d object recognition in industry. In Proceedings of the IEEE international conference on computer vision workshops, pages 2200-2208, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[10] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[11] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "type": "text", + "content": "[12] Ge Gao, Mikko Lauri, Yulong Wang, Xiaolin Hu, Jianwei Zhang, and Simone Frintrop. 6d object pose regression via supervised learning on point clouds. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3643-3649. IEEE, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 151, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 204 + ], + "type": "text", + "content": "[13] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1580-1589, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "type": "text", + "content": "[14] Xingyi He, Jiaming Sun, Yang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without cad models. Advances in Neural Information Processing Systems, 35:35103-35115, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[15] Tomás Hodan, Pavel Haluza, Štepan Obdržálek, Jiri Matas, Manolis Lourakis, and Xenophon Zabulis. T-less: An rgb-d dataset for 6d pose estimation of texture-less objects. In 2017 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 880-888. IEEE, 2017. 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 317, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 553, + 380 + ], + "type": "text", + "content": "[16] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. Bop: Benchmark for 6d object pose estimation. In Proceedings of the European conference on computer vision (ECCV), pages 19-34, 2018. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 383, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 459 + ], + "type": "text", + "content": "[17] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. BOP challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "text", + "content": "[18] Sabera Hoque, Shuxiang Xu, Ananda Maiti, Yuchen Wei, and Md Yasir Arafat. Deep learning for 6d pose estimation of objects—a case study for autonomous driving. Expert Systems with Applications, 223:119838, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 504, + 555, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 555, + 548 + ], + "type": "text", + "content": "[19] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2930-2939, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "type": "text", + "content": "[20] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10095-10105, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "type": "text", + "content": "[21] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[22] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapose: 6d pose estimation of novel objects via render & compare. arXiv preprint arXiv:2212.06870, 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16826" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[23] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "text", + "content": "[24] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[25] Hongyang Li, Jiehong Lin, and Kui Jia. Dcl-net: Deep correspondence learning network for 6d pose estimation. In European Conference on Computer Vision, pages 369-385. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[26] Yuelong Li, Yafei Mao, Raja Bala, and Sunil Hadap. Mrcnet: 6-dof pose estimation with multiscale residual correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10476-10486, 2024. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[27] Ruyi Lian and Haibin Ling. Checkerpose: Progressive dense keypoint localization for object pose estimation with graph neural network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14022-14033, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 364 + ], + "type": "text", + "content": "[28] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27906-27916, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "text", + "content": "[29] Jian Liu, Wei Sun, Chongpei Liu, Xing Zhang, and Qiang Fu. Robotic continuous grasping system by shape transformer-guided multiobject category-level 6-d pose estimation. IEEE Transactions on Industrial Informatics, 19 (11):11171-11181, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "text", + "content": "[30] Jian Liu, Wei Sun, Hui Yang, Zhiwen Zeng, Chongpei Liu, Jin Zheng, Xingyu Liu, Hossein Rahmani, Nicu Sebe, and Ajmal Mian. Deep learning-based object pose estimation: A comprehensive survey. arXiv preprint arXiv:2405.07801, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "text", + "content": "[31] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In European Conference on Computer Vision, pages 298-315. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 535, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 601 + ], + "type": "text", + "content": "[32] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "text", + "content": "[33] Ningkai Mo, Wanshui Gan, Naoto Yokoya, and Shifeng Chen. Es6d: A computation efficient and symmetry-aware 6d pose regression framework. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6718-6727, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[34] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. Genflow: Generalizable recurrent flow for 6d pose refinement of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10039-10049, 2024. 3, 5, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "type": "text", + "content": "[35] Fengjun Mu, Rui Huang, Ao Luo, Xin Li, Jing Qiu, and Hong Cheng. Temporalfusion: Temporal motion reasoning with multi-frame fusion for 6d object pose estimation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5930-5936. IEEE, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 140, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 204 + ], + "type": "text", + "content": "[36] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6771-6780, 2022. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[37] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2134-2140, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 261, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 325 + ], + "type": "text", + "content": "[38] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Yinlin Hu, Renaud Marlet, Mathieu Salzmann, and Vincent Lepetit. Nope: Novel object pose estimation from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17923-17932, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 327, + 555, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 381 + ], + "type": "text", + "content": "[39] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9903–9913, 2024. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "type": "text", + "content": "[40] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "text", + "content": "[41] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In European Conference on Computer Vision, pages 163-182. Springer, 2025. 1, 3, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "text", + "content": "[42] Jaewoo Park and Nam Ik Cho. Dprost: Dynamic projective spatial transformer network for 6d pose estimation. In European Conference on Computer Vision, pages 363-379. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "text", + "content": "[43] Keunhong Park, Arsalan Mousavian, Yu Xiang, and Dieter Fox. Latentfusion: End-to-end differentiable reconstruction and rendering for unseen object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10710-10719, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "type": "text", + "content": "[44] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4561-4570, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[45] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud registration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3979-3985, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[46] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16827" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "poses of challenging objects without using depth. In Proceedings of the IEEE international conference on computer vision, pages 3828-3836, 2017. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "type": "text", + "content": "[47] Caner Sahin and Tae-Kyun Kim. Category-level 6d object pose recovery in depth images. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "text", + "content": "[48] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6835-6844, 2022. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 274 + ], + "type": "text", + "content": "[49] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In 2019 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227. IEEE, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 308 + ], + "type": "text", + "content": "[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. CVPR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 296, + 376 + ], + "type": "text", + "content": "[51] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6825-6834, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 296, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 296, + 433 + ], + "type": "text", + "content": "[52] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 699-715, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 296, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 489 + ], + "type": "text", + "content": "[53] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6d object pose and size estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 530-546. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 491, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 296, + 555 + ], + "type": "text", + "content": "[54] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 108-125. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 296, + 612 + ], + "type": "text", + "content": "[55] Jiaxin Wei, Xibin Song, Weizhe Liu, Laurent Kneip, Hongdong Li, and Pan Ji. Rgb-based category-level object pose estimation via decoupled metric scale recovery. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2036-2042. IEEE, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "type": "text", + "content": "[56] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17868-17879, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "type": "text", + "content": "[57] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 5, 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 306 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "text", + "content": "[58] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "text", + "content": "[59] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2025. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 554, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 250 + ], + "type": "text", + "content": "[60] Heng Zhao, Shenxing Wei, Dahu Shi, Wenming Tan, Zheyang Li, Ye Ren, Xing Wei, Yi Yang, and Shiliang Pu. Learning symmetry-aware geometry correspondences for 6d object pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14045-14054, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 251, + 555, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 306 + ], + "type": "text", + "content": "[61] Wanqing Zhao, Shaobo Zhang, Ziyu Guan, Wei Zhao, Jinye Peng, and Jianping Fan. Learning deep network for detecting 3d object keypoints and 6d poses. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 14134-14142, 2020. 1, 2, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16828" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Pose Priors from Language Models/9f701210-ac31-4714-b2b0-3e6ed2c2dfd9_content_list.json b/2025/Pose Priors from Language Models/9f701210-ac31-4714-b2b0-3e6ed2c2dfd9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..50d41487a50efc7da4c81cab01b029f686de7366 --- /dev/null +++ b/2025/Pose Priors from Language Models/9f701210-ac31-4714-b2b0-3e6ed2c2dfd9_content_list.json @@ -0,0 +1,1474 @@ +[ + { + "type": "text", + "text": "Pose Priors from Language Models", + "text_level": 1, + "bbox": [ + 320, + 130, + 678, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sanjay Subramanian1 Evonne Ng1 Lea Müller1 Dan Klein1 Shiry Ginosar2,3 Trevor Darrell1 \n1 University of California, Berkeley \n2 Google DeepMind \n3 Toyota Technological Institute at Chicago \n{sanjayss,evonne_ng,mueller,klein,trevordarrell}@berkeley.edu \nshiry@google.com", + "bbox": [ + 99, + 178, + 919, + 287 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f8f6ccd34f91ef92ce6ef19db6cdba97778ee3d527121d3bc54b2bdfa3ce926.jpg", + "image_caption": [ + "Figure 1. Optimizing human-to-human contacts in 3D pose. Our approach leverages the semantic priors of a Large Multimodal Model (LMM) to infer meaningful information about physical contact from images. Instead of relying on human annotations or motion capture data, we extract not only descriptive insights (\"... engaged in a dance or embrace ...\") but also structured constraints between body parts (underlined). By incorporating these LMM-derived constraints, we refine initial 3D human pose estimates, achieving realistic and semantically consistent reconstructions of contact. This scalable approach opens up new possibilities for contact-aware pose estimation without explicit contact annotations, making it a promising alternative to traditional methods." + ], + "image_footnote": [], + "bbox": [ + 102, + 306, + 903, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 636, + 326, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Language is often used to describe physical interaction, yet most 3D human pose estimation methods overlook this rich source of information. We bridge this gap by leveraging large multimodal models (LMMs) as priors for reconstructing contact poses, offering a scalable alternative to traditional methods that rely on human annotations or motion capture data. Our approach extracts contact-relevant descriptors from an LMM and translates them into tractable losses to constrain 3D human pose optimization. Despite its simplicity, our method produces compelling reconstructions for both two-person interactions and self-contact scenarios, accurately capturing the semantics of physical and social interactions. Our results demonstrate that LMMs can serve as powerful tools for contact prediction and pose estimation, offering an alternative to costly manual human annotations", + "bbox": [ + 88, + 675, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "or motion capture data. Our code is publicly available at https://prosepose.github.io.", + "bbox": [ + 513, + 638, + 906, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 695, + 645, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Language, as a human artifact, encodes a rich set of social and physical interactions. Over centuries, our vocabulary has evolved to describe the nuances of touch, with words and phrases capturing contexts as varied as hugs, handshakes, or postures in sports and yoga. Perceiving physical contact is essential for understanding human behavior: e.g. several forms of parent-child contact are associated with affection [3, 43], and some forms of self-contact signal stress [16].", + "bbox": [ + 509, + 720, + 906, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since written language discusses our physical interactions at great length, can large multimodal models (LMMs) trained on images and text correctly perceive physical contact in human pose? This question has practical significance", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7125", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "because scenes with contact are challenging for pose estimation methods as some body parts are frequently occluded. This particularly holds for methods solely relying on 2D keypoints which do not convey contact information. Previously proposed approaches address these issues by curating task-specific datasets via motion capture or human-annotated points of contact between body parts [11, 33, 34]. However, collecting these datasets is expensive, and existing publicly available datasets include only tens of thousands of images [10, 22, 33, 50]. If LMMs can accurately identify contact points, they could decrease the cost of curating such datasets", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we study the efficacy of LMMs as tools for contact prediction in pose estimation. Since LMMs output language rather than pose parameters, answering this question requires a way of eliciting the required information from an LMM and operationalizing it in pose estimation. We introduce a framework, called ProsePose, which prompts an LMM for formatted constraints about physical contact in the image, converts the constraints into a loss function, and optimizes this loss function (jointly with losses from other cues, such as 2D keypoints) to refine initial pose estimates.", + "bbox": [ + 89, + 257, + 480, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use ProsePose to evaluate LMMs on both 2-person interaction datasets and a dataset of complex yoga poses. Our framework improves pose estimates compared to strong baselines that do not use contact supervision. In our extensive analysis, we show that several components are important: mitigating LMM hallucinations by aggregating predictions from several LMM samples, careful construction of the prompt and loss functions, and integrating losses from other cues. Finally, we conduct an extensive analysis of LMM predictions and their role in the optimization results. With respect to LMM failures, we find that identifying the chirality of limbs is a particular challenge for LMMs. While existing supervised methods excel by training on large amounts of supervised data with contact labels, we show that we can extract useful priors for contact prediction from pretrained LMMs without fine tuning.", + "bbox": [ + 89, + 409, + 482, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are (1) we introduce a framework for applying LMMs as contact prediction tools in pose estimation, (2) we show that our framework can improve the quality of pose estimates in 2-person and 1-person settings, and (3) we provide an analysis of the components of our framework and LMM failure modes.", + "bbox": [ + 89, + 651, + 482, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 753, + 232, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D human pose reconstruction. Reconstructing 3D human poses from single images is an active area of research. Prior works have explored optimization-based approaches [13, 26, 35, 36, 39] or pure regression [2, 14, 20, 21, 25] to estimate the 3D body pose given a single image. HMR2 [12] is a recent state-of-the-art regression model in this line of work. Building on these approaches, some methods have looked into reconstructing multiple individuals jointly from a single", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image. These methods [19, 41, 51] use deep networks to reason about multiple people in a scene to directly output multi-person 3D pose predictions. BEV [42] accounts for the relative proximity of people explicitly using relative depth annotations to reason about proxemics when placing each individual in the scene (e.g. relative depth of people). However, approaches in both categories generally do not accurately capture physical contact between parts of a single person or between people [33, 34].", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contact inference in 3D pose reconstruction. 3D pose reconstruction is especially challenging when there is self-contact or inter-person contact. This has motivated a line of work on pose reconstruction approaches tailored for these settings. [33] focuses on predicting self contact regions for 3D pose estimation by leveraging a dataset with contact annotations to model complex poses such as crossed arms. [10] introduces the first dataset with hand-annotated ground-truth contact labels between two people. REMIPS [11] and BUDDI [34] train models on the person-to-person contact maps in this data in order to improve 3D pose estimation of multiple people from a single image. CloseInt [17] trains a physics-guided diffusion model on two-person motion capture data for this task. However, contact annotations, which are crucial for these approaches, are expensive to acquire. Our method does not require any training on such annotations. Instead, we leverage an LMM's implicit knowledge of pose to constrain pose optimization to capture both self- and person-to-person contact.", + "bbox": [ + 511, + 237, + 906, + 525 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language priors on human pose. There exists a plethora of text to 3D human pose and motion datasets [15, 37, 38], which have enabled work focused on generating 3D motion sequences of a single person performing a general action [18, 44, 52]. This line of work has been extended to generating the motion of two people conditioned on text [29, 40].", + "bbox": [ + 511, + 536, + 906, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PoseScript [7] is a method for generating a single person's pose from fine-grained descriptions, which uses training data from motion capture annotated with detailed text. PoseFix [8] introduces a labeled dataset for the task of modifying a pose given a fine-grained description of the desired change and trains a model on this data. PoseGPT [9] is a pose regressor that uses language as part of its training data. However, PoseGPT does not produce better pose estimates than previous state-of-the-art regressors (i.e. regressors that do not use language) and applies only to the one-person setting. [46] uses a text-only LM to improve action-conditioned human-object pose estimation. This method relies on a limited database of action-pose pairs to classify an input pose, and uses an LM to improve pose estimates based on the action retrieved from the database.", + "bbox": [ + 511, + 628, + 908, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work differs from previous work on language and pose in several ways. First, whereas all prior work relies on training data with pairs of language and pose, which", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7126", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bccbf40ac4ce4854f8466d257bdaeef8b3a8dd2789d155231527b821e73a939b.jpg", + "image_caption": [ + "Figure 2. LMM-guided Pose Estimation. (a) Method overview: ProsePose takes as input an image of one or two people in contact. We first obtain initial pose estimates for each person from a pose regressor. Then we use an LMM to generate contact constraints, each of which is a pair of body parts that should be touching. This list of contacts is converted into a loss function $\\mathcal{L}_{\\mathrm{LMM}}$ . We optimize the pose estimates using $\\mathcal{L}_{\\mathrm{LMM}}$ and other losses to produce a refined estimate of each person's pose that respects the predicted contacts. (b) Defining contact constraints: Given an image $\\pmb{I}$ , we can lift each individual into corresponding 3D meshes $\\pmb{V}$ . A contact constraint $\\pmb{c}$ is a pair of regions $(\\pmb{R}_a, \\pmb{R}_b)$ in contact. The loss is defined in terms of the distance between the vertices $(\\pmb{v}_a, \\pmb{v}_b)$ on the mesh." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 584, + 281 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/367f6409f63bade72da68af5b3fafc47aca26d5875d5a507d826286ab6845cd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 130, + 908, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is expensive to collect, our method leverages the existing knowledge in an LMM to reason about pose from a given image. Second, prior work in this area focuses on either the one-person or the two-person setting. In contrast, our work presents a single framework to reason about physical contacts within or between poses. Finally, in scenes with physical contact, we show that our method improves the pose estimates of state-of-the-art regressors.", + "bbox": [ + 88, + 398, + 486, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Guiding Pose Optimization with an LMM", + "text_level": 1, + "bbox": [ + 89, + 566, + 465, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an image, our goal is to estimate the 3D body pose of individuals in the image while capturing the self and cross-person contact points. While we cannot trivially use natural language responses (hug, kiss) to directly optimize 3D body poses, we leverage the key insight that LMMs understand how to articulate a given pose (arms around waist, lips touching). We propose a method to structure these articulations into constraints and convert them into loss functions.", + "bbox": [ + 88, + 602, + 486, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More concretely, our framework, illustrated by Figure 2a, takes as input the image $\\pmb{I}$ and the bounding boxes $B$ of the subjects of interest. In the first stage, a pose regressor takes the image and produces a rough estimate of the 3D pose $X^p$ for each individual $p$ in the image. In the second stage, an LMM takes the image and a set of instructions and generates a list of self- or inter-person contact constraints, which we then convert into a loss function (Sec. 3.2). Finally, in the third stage, we jointly optimize the generated loss function with several other pre-defined loss terms (Sec. 3.3). We refer to our framework as ProsePose.", + "bbox": [ + 88, + 734, + 486, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 398, + 653, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We focus our description on the two-person case to keep the exposition simple. We also demonstrate results on the one-person case, which is simply an extension of the two-person case. In particular, we apply our method to the one-person case by setting $X^0 = X^1$ . Please see Appendix § 7 for details on the differences between the two cases.", + "bbox": [ + 511, + 422, + 908, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Multimodal Models. An LMM is a model that takes as input an image and a text prompt and produces text output that answers the prompt based on the image. Our framework is agnostic to the architecture of the LMM. LMMs are typically trained to respond to a wide variety of instructions [6, 30]. However, LMMs are prone to hallucination [27, 28]. Handling cases of hallucination is a key challenge when using LMMs. We mitigate this issue by aggregating information across several samples from the LMM.", + "bbox": [ + 511, + 525, + 908, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose representation. We use a human body model [36] to represent each person $p \\in \\{0,1\\}$ . The body model is composed of a pose parameter that defines the joint rotations $\\theta \\in \\mathbb{R}^{d_{\\theta} \\times 3}$ , where $d_{\\theta}$ is the number of joints, and a shape parameter $\\beta \\in \\mathbb{R}^{d_{\\beta}}$ , where $d_{\\beta}$ is the dimensions of the shape parameter. We can apply a global rotation $\\Phi \\in \\mathbb{R}^3$ and translation $t \\in \\mathbb{R}^3$ to place each person in the world coordinate space. The full set of parameters for each person is denoted by $\\boldsymbol{X}^p = [\\theta^p, \\beta^p, \\Phi^p, t^p]$ . For simplicity, we refer to the parameter set $(\\boldsymbol{X}^0, \\boldsymbol{X}^1)$ as $\\boldsymbol{X}$ .", + "bbox": [ + 511, + 672, + 908, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These parameters can be plugged into a differentiable function that maps to a mesh consisting of $d_v$ vertices $\\mathbf{V} \\in \\mathbb{R}^{d_v \\times 3}$ . From the mesh, we can obtain the 3D locations of the body's joints $\\mathbf{J} \\in \\mathbb{R}^{d_j \\times 3}$ . From these joints, we can calculate the 2D keypoints $\\mathbf{K}_{proj}$ by projecting the 3D joints to 2D", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7127", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "using the camera intrinsics $\\Pi$ predicted from [36].", + "bbox": [ + 89, + 90, + 421, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {K} _ {p r o j} = \\Pi (\\boldsymbol {J}) \\in \\mathbb {R} ^ {d _ {j} \\times 2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 116, + 483, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vertex regions. In order to define contact constraints between body parts, we define a set of regions of vertices. Prior work on contact has partitioned the body into fine-grained regions [10]. However, since our constraints are specified by a LMM trained on natural language, the referenced body parts are often coarser in granularity. We therefore update the set of regions to reflect this language bias by combining these fine-grained regions into larger, more commonly referenced body parts such as arm, shoulder (front&back), back, and waist (front&back). Some regions, e.g. back and waist (back), overlap. Please see Appendix § 7.2 for a visualization of the coarse regions. Formally, we write $\\mathbf{R} \\in \\mathbb{R}^{d_r \\times 3}$ to denote a region with $d_r$ vertices, which is part of the full mesh $(\\mathbf{R} \\subset V)$ .", + "bbox": [ + 89, + 157, + 483, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Constraint definition. A contact constraint specifies which body parts from two meshes should touch. We define contact constraints as pairs of coarse regions $c = (R_a, R_b)$ between a region $R_a$ of one mesh and $R_b$ of the other mesh, as shown in Figure 2b. For instance, (\"hand\", \"arm\") indicates a hand should touch an arm.", + "bbox": [ + 89, + 377, + 483, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Constraint generation with a LMM", + "text_level": 1, + "bbox": [ + 89, + 477, + 398, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our key insight is to leverage a LMM to identify regions of contact between different body parts on the human body surface. As shown in Figure 2a, we prompt the LMM with an image and ask it to output a list of all region pairs that are in contact. However, we cannot simply use its output natural language descriptions to directly optimize a 3D mesh. As such, we convert these constraints into a loss function.", + "bbox": [ + 89, + 500, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LMM-based constraint generation. Given the image $I$ , we first use the bounding boxes $B$ to crop the part containing the subjects. We then use an image segmentation model to mask any extraneous individuals. While cropping and masking the image may remove information, we find the LMMs are relatively robust to missing context, and more importantly, this allows us to indicate which individuals to focus on. Given the segmented image, we ask the LMM to generate a set $C = \\{c_1, \\dots, c_m\\}$ of all pairs of body parts that are touching, where $m$ is the total number of constraints the LMM generates for the image.", + "bbox": [ + 89, + 613, + 483, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the prompt, we specify the full set of coarse regions to pick from. We find that LMMs fail to reliably reference the left and right limbs correctly or consistently, so the prompt instructs the LMM not to specify chirality for each limb (see Appendix 8.2 for prompt analysis). If the LMM uses \"left\" or \"right\" to reference a region, despite an instruction to not do so, we directly use the part of the region with the specified chirality rather than considering both possibilities.", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the chain-of-thought technique, which has been shown to improve language model performance on reasoning tasks [48], we ask the LMM to write its reasoning or describe the pose before listing the constraints. For the full prompt used in each setting, please refer to Appendix § 7.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We sample $N$ responses from the LMM. Below we describe (1) how we convert these natural language responses into $N$ sets of constraints $\\{C_1, C_2, \\dots, C_N\\}$ and (2) how we convert each constraint set $C_j$ into a loss.", + "bbox": [ + 511, + 167, + 906, + 228 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Canonicalizing Region Names and Assigning Chirality. Given the LMM's output, we must map the mentioned region names to our fixed set of coarse regions. Since the LMM may deviate from the names in the prompt, we check for some additional names (see Appendix 7.2). We then filter out contact pairs that occur fewer than $f$ times across constraint sets, where $f$ is a hyperparameter.", + "bbox": [ + 511, + 237, + 906, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we assign a chirality (left/right) to each hand/arm/foot/leg/shoulder region in cases when the LMM does not itself specify the chirality. We enumerate all possible assignments of left/right to these regions and take the one resulting in the minimum loss. In the two-person setting, we only consider assignments satisfying a condition designed for the case in which a region type occurs in multiple constraint pairs (see Appendix 7.3).", + "bbox": [ + 511, + 344, + 908, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss function generation. We compute a loss for each constraint by mapping the relevant regions to sets of vertices and calculating the minimum distance between vertices in the two sets. In particular, we first specify a mapping between each of the coarse region names and the fine-grained regions from [10]. We then use a mapping from fine-grained regions to SMPL-X vertices (provided by [34]) to obtain a set of vertices for each coarse region. Then for each contact pair of coarse regions $c = (R_a, R_b)$ in $C_j$ , we define $dist(c)$ as the minimum distance between the two regions:", + "bbox": [ + 511, + 474, + 908, + 626 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {d i s t} (\\boldsymbol {c}) = \\min \\| \\boldsymbol {v} _ {a} - \\boldsymbol {v} _ {b} \\| _ {2} \\quad \\forall \\boldsymbol {v} _ {a} \\in \\boldsymbol {R} _ {a}, \\forall \\boldsymbol {v} _ {b} \\in \\boldsymbol {R} _ {b} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 638, + 906, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{\\pmb{v}_a, \\pmb{v}_b\\} \\in \\mathbb{R}^3$ . In practice, the number of vertices in each region can be very large. To make this computation tractable, we first take a random sample of vertices from $R_a$ and from $R_b$ before computing distances between pairs of vertices in these samples.", + "bbox": [ + 511, + 669, + 906, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, since the ordering of the people in the LMM constraints is unknown (i.e. does $\\mathbf{R}_a$ come from the mesh defined by parameter $\\mathbf{X}^0$ or $\\mathbf{X}^1$ ), we compute the overall loss for both possibilities and take the minimum. We use $\\mathbf{c}^\\top = (\\mathbf{R}_b, \\mathbf{R}_a)$ to denote the flipped ordering. We then sum over all constraints in the list $C_j$ :", + "bbox": [ + 511, + 744, + 905, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {d i s t} _ {\\text {s u m}} \\left(\\boldsymbol {C} _ {j}\\right) = \\min \\left(\\sum_ {\\boldsymbol {c} \\in \\boldsymbol {C} _ {j}} \\operatorname {d i s t} (\\boldsymbol {c}), \\sum_ {\\boldsymbol {c} \\in \\boldsymbol {C} _ {j}} \\operatorname {d i s t} \\left(\\boldsymbol {c} ^ {\\top}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 849, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7128", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each constraint set sampled from the LMM is likely to contain noise or hallucination. To mitigate this issue, we average over all $N$ losses corresponding to each constraint set to obtain the overall LMM loss. This technique is similar to self-consistency [47], which is commonly used for code generation. Concretely, the overall LMM loss is defined as", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L M M}} = \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\operatorname {d i s t} _ {s u m} \\left(C _ {j}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 202, + 483, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If a constraint set $C_j$ is empty (i.e. the LMM does not suggest any contact pairs), then we set $dist_{\\mathrm{sum}}(C_j) = 0$ . If there are several such constraint sets, we infer that the LMM has low confidence about the contact points (if any) in the image. Consequently, we set a threshold $t$ and if the number of empty constraint sets is at least $t$ , we gracefully backoff to the appropriate baseline optimization procedure (described in Sections 4.1 and 4.2 for each setting). We also backoff to the baseline if the LMM-based optimization diverges.", + "bbox": [ + 89, + 253, + 483, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Constrained pose optimization", + "text_level": 1, + "bbox": [ + 89, + 398, + 362, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Drawing from previous optimization-based approaches [4, 34, 35], we employ several additional losses in the optimization. We then minimize the joint loss to obtain a refined subset of the body model parameters $\\mathbf{X}^{\\prime} = [\\pmb{\\theta}^{\\prime},\\pmb{\\beta}^{\\prime},\\pmb{t}^{\\prime}]$ :", + "bbox": [ + 89, + 421, + 483, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} [ \\boldsymbol {\\theta} ^ {\\prime}, \\boldsymbol {\\beta} ^ {\\prime}, \\boldsymbol {t} ^ {\\prime} ] = \\arg \\min (\\lambda_ {\\mathrm {L M M}} \\mathcal {L} _ {\\mathrm {L M M}} + \\lambda_ {\\mathrm {G M M}} \\mathcal {L} _ {\\mathrm {G M M}} + \\lambda_ {\\beta} \\mathcal {L} _ {\\beta} \\\\ + \\lambda_ {\\theta} \\mathcal {L} _ {\\theta} + \\lambda_ {2 D} \\mathcal {L} _ {2 D} + \\lambda_ {P} \\mathcal {L} _ {P}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 491, + 478, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [34], we divide the optimization into two stages. In the first stage, we optimize all three parameters. In the second stage, we optimize only $\\theta$ and $t$ , keeping the shape $\\beta$ fixed. We detail the remaining losses below.", + "bbox": [ + 89, + 537, + 483, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose and shape priors. We compute a loss $\\mathcal{L}_{\\mathrm{GMM}}$ based on the Gaussian Mixture pose prior of [4] and a shape loss $\\mathcal{L}_{\\beta} = \\| \\beta \\|_2^2$ , which penalizes extreme deviations from the body model's mean shape.", + "bbox": [ + 89, + 604, + 483, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initial pose loss. To ensure we do not stray too far from the initialization, we penalize large deviations from the initial pose $\\mathcal{L}_{\\theta} = ||\\pmb{\\theta}' - \\pmb{\\theta}||_2^2$", + "bbox": [ + 89, + 671, + 483, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2D keypoint loss. Similar to BUDDI [34], for each person in the image, we obtain pseudo ground truth 2D keypoints and their confidence from OpenPose [5] and ViTPose [49]. Given this pseudo ground truth, we merge all the keypoints into $\\pmb{K} \\in \\mathbb{R}^{d_j \\times 2}$ , and their corresponding confidence into $\\gamma \\in \\mathbb{R}^{d_j}$ . From the predicted $\\pmb{X}'$ , we can compute the 2D projection of each 3D joint location using Equation 3.1. Then, the 2D keypoint loss is defined as:", + "bbox": [ + 89, + 722, + 485, + 844 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2 D} = \\sum_ {j = 1} ^ {d _ {j}} \\boldsymbol {\\gamma} \\left(\\boldsymbol {K} _ {p r o j} - \\boldsymbol {K}\\right) ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 853, + 483, + 897 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Interpenetration loss. To prevent parts of one mesh from being in the interior of the other, we add an interpenetration loss. Generically, given two sets of vertices $V_{0}$ and $V_{1}$ , we use winding numbers to compute the subset of $V_{0}$ that intersects $V_{1}$ , which we denote as $V_{0,1}$ . Similarly, $V_{1,0}$ is the subset of $V_{1}$ that intersects $V_{0}$ . The interpenetration loss is then defined as", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P} = \\sum_ {x \\in \\boldsymbol {V} _ {0, 1}} \\min _ {v _ {1} \\in \\boldsymbol {V} _ {1}} \\| x - v _ {1} \\| _ {2} ^ {2} + \\sum_ {y \\in \\boldsymbol {V} _ {1, 0}} \\min _ {v _ {0} \\in \\boldsymbol {V} _ {0}} \\| y - v _ {0} \\| _ {2} ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 205, + 906, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For efficiency, this loss is computed on low-resolution versions of the two meshes (roughly 1000 vertices per mesh).", + "bbox": [ + 511, + 268, + 908, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 313, + 645, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Following prior work on two-person pose estimation [34], we use BEV [42] to initialize the poses since it was trained to predict both the body pose parameters and the placement of each person in the scene. However, on the single person yoga poses, we find that the pose parameter estimates of HMR2 [12] are much higher quality, so we initialize the body pose using HMR2.", + "bbox": [ + 511, + 347, + 906, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the SMPL-X [36] body model and (unless specified otherwise) GPT4-V [1] as the LMM with temperature $= 0.7$ when sampling from it. GPT4-V refers to the gpt-4-vision-preview model in the OpenAI API: platform.openai.com. In the OpenAI API, we use the \"high\" detail setting for image input. Appendix 8.2 provides results using other prompts and other LMMs (LLaVA [31], GPT-4o) and a running time analysis. Unless otherwise specified, we set $N = 20$ samples. For all of our 2-person experiments, $f = 1$ , while $f = 10$ in the 1-person setting. We set $t = 2$ for the experiment on the CHI3D dataset and $t = N$ for all other experiments. The hyperparameters and our prompts were chosen based on experiments on the validation sets. For other implementation details refer to Appendix § 7.", + "bbox": [ + 511, + 454, + 908, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. As is standard in the pose estimation literature, we report Procrustes-aligned Mean Per Joint Position Error (PA-MPJPE) in millimeters. This metric finds the best alignment between the estimated and ground-truth pose before computing the joint error. In the two-person setting, we focus on the joint PA-MPJPE, as this evaluation incorporates the relative translation and orientation of the two people. See Appendix § 8.2 for the per-person PA-MPJPE.", + "bbox": [ + 511, + 672, + 908, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also include the percentage of correct contact points (PCC) metric introduced by [34]. This metric captures the fraction of ground-truth contact pairs that are accurately predicted. For a given radius $r$ , a pair is classified as \"in contact\" if the two regions are both within the specified radius. We use the set of fine-grained regions defined in [10] to compute PCC. The metric is averaged over $r \\in$", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7129", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8235dbab5c70c75196a0bf8ba4f6bcbbdd56ace04a4cafbc8b9f4ac9448af16f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Two-person Results. Joint PA-MPJPE (abbreviated PM) (lower is better), Avg. PCC (higher is better), and F1 (higher is better). For FlickrCI3D, PA-MPJPE is computed using the pseudoground-truth fits. F1 measures the accuracy of coarse contacts predicted by the LMM, while the other two metrics evaluate the quality of the estimated 3D pose. Last line shows results using ground-truth contact pairs of coarse regions (heuristic is still used as the backoff method when needed). Bold indicates best method without/with contact supervision in each column." + ], + "table_body": "
Hi4DFlickrCI3DCHI3D
PM↓F1↑PM↓PCC↑F1↑PM↓PCC↑F1↑
w/o contact sup.
BEV [42]144-10664.8-9671.4-
Heuristic116-6777.8-10574.1-
ProsePose93245879.91310075.823
w. contact sup.
BUDDI [34]89-6681.9-6878.0-
BUDDI+ProsePose88-6583.2-6978.8-
Coarse GT contacts
Oracle8110043861008383.8100
", + "bbox": [ + 91, + 88, + 486, + 248 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "0, 5, 10, 15, ..., 95 mm. Since these regions are defined on the SMPL-X mesh topology, we convert the regression baselines-BEV and HMR2- from the SMPL mesh topology to SMPL-X to compute this metric. Please see the Appendix § 8.1 for more details on the regions and on the mesh conversion.", + "bbox": [ + 89, + 402, + 483, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we report the F1 of the LMM's predicted coarse region pairs. Specifically, we compute precision as the proportion of predicted pairs $(\\mathbf{R}_a,\\mathbf{R}_b)$ such that there is some ground-truth pair $(\\mathbf{R}_a^*,\\mathbf{R}_b^*)$ for which $\\mathbf{R}_a$ overlaps with $\\mathbf{R}_a^*$ and $\\mathbf{R}_b$ overlaps with $\\mathbf{R}_b^*$ . We compute recall similarly as a proportion of ground-truth pairs. For datasets that do not provide contact maps, we define the ground-truth pairs from the ground-truth meshes, using a threshold on the minimum distance between regions (0.01 and 0.02 for Hi4D and MOYO, respectively). We ignore chirality when computing these. For a pair of people we take the maximum F1 of the two possible orderings. This metric serves to evaluate the raw output of the LMMs without 3D optimization.", + "bbox": [ + 89, + 478, + 483, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Two-person Pose Refinement", + "text_level": 1, + "bbox": [ + 89, + 686, + 349, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate on three datasets, and our dataset processing largely follows [34]. Hi4D [50] is a motion capture dataset of pairs of people interacting. Each sequence has a subset of frames marked as contact frames, and we take every fifth contact frame. We use the images from a single camera, resulting in 241 images. Flickr Close Interactions 3D (FlickrCI3D) [10] is a collection of Flickr images of multiple people in close interaction. The dataset includes manual annotations of the contact maps between pairs of people. [34] used these contact maps to create pseudo-ground truth 3D meshes and curated a version of the test set to exclude noisy annotations, which has 1403 images. CHI3D [10] is", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ee0a71bf6a93ee1c93e47b90787cf393f73b950baaa912009388f7d6e0adc40c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Two-person PCC. Percent of correct contact points (PCC) for three different radii $r$ in mm. Bold indicates the best score without/with contact supervision in each column. At the ground-truth contact points, our method brings the meshes closer together than the baselines." + ], + "table_body": "
@ radius[mm]FlickrCI3D PCC↑CHI3D PCC↑
5101551015
W/o contact sup.
BEV [42]3.66.310.85.817.432.5
Heuristic14.633.949.311.128.045.3
ProsePose15.639.957.113.535.252.5
W/ contact sup.
BUDDI [34]18.544.261.815.539.056.6
BUDDI+ProsePose21.849.366.419.543.958.8
", + "bbox": [ + 517, + 88, + 903, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a motion capture dataset of pairs of people interacting. We present results on the validation set. There are 431 images, distributed across 4 cameras. The images come from 126 video sequences, each of which has a single \"contact frame.\"", + "bbox": [ + 511, + 316, + 906, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To develop our method, we experimented on the validation sets of FlickrCI3D and Hi4D, and a sample of the training set from CHI3D. For our experiments, we can compute the PCC on FlickrCI3D and CHI3D, which have annotated ground-truth contact maps. Following [34], we exclude from evaluation images where BEV or the keypoint detectors, which are used by the baselines as well, fail to detect one of the subjects in the interaction pair.", + "bbox": [ + 511, + 378, + 908, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We compare our estimated poses to the following:", + "bbox": [ + 514, + 510, + 906, + 526 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- BEV [42] Multi-person 3D pose estimation method. Uses relative depth to reason about spatial placement of individuals in the scene. ProsePose, Heuristic, and BUDDI use BEV to initialize pose estimates.", + "- Heuristic A contact heuristic which includes the auxiliary losses in Section 3.3 as well as a term that minimizes the minimum distance between the two meshes. Introduced by [34]. We use their hyperparameters for this heuristic. This baseline is also used as the backoff method for Prose-Pose when the number of empty constraint sets is at least the threshold $t$ or when the optimization diverges.", + "- BUDDI [34] This method uses a learned diffusion prior to constrain the optimization. We stress that BUDDI requires a large amount of annotated training data on pairs of interacting bodies, which is not used in our method." + ], + "bbox": [ + 514, + 527, + 906, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results Table 1 provides quantitative results on the three datasets. Across datasets, ProsePose consistently improves over the strongest baseline, Heuristic. On the Hi4D dataset, ProsePose reduces $85\\%$ of the gap in PA-MPJPE between Heuristic and the fully supervised BUDDI. On the FlickrCI3D and CHI3D datasets, ProsePose narrows the gap in the average PCC between Heuristic and BUDDI by more than one-third. (While ProsePose achieves a better PA-MPJPE than BUDDI on FlickrCI3D, for this dataset,", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7130", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6c2efec2ce40dd45c106171a96d1a9043fe73f7a0ee93222955a8926130539ad.jpg", + "image_caption": [ + "Figure 3. Two-person examples We show qualitative results from ProsePose, BUDDI [34], and the contact heuristic. For each example, we show GPT4-V's top 3 constraints and the number of times each constraint was predicted across all 20 samples. Our method correctly reconstructs people in a variety of interactions, and the predicted constraints generally align with each interaction type." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 467, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we rely primarily on PCC since PA-MPJPE is computed on pseudo-ground-truth fits.)", + "bbox": [ + 89, + 700, + 482, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On CHI3D, ProsePose outperforms Heuristic but underperforms BEV in terms of PA-MPJPE. On the subset of images where we do not default to the heuristic (i.e. on images where GPT4-V predicts enough non-empty constraint sets), the PA-MPJPE for ProsePose and BEV is 86 and 87, respectively. In other words, in the cases where our method is actually used, the joint error is slightly less than that of BEV. As a result, we can attribute the worse overall error to the poorer performance of the heuristic. The backoff method (which is the heuristic) is used in 13/241 Hi4D examples, 106/1403 Flickr examples, and 224/431 CHI3D examples.", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ed66116ef372112d24251ca8d321a11ce3dfcadfd07283e5984f0a8cf653466.jpg", + "image_caption": [ + "Figure 4. Single-person examples We show qualitative results from ProsePose, HMR2 [12], and HMR2-optim on complex yoga poses. Each example also shows the constraints that are predicted by the LMM at least $f = 10$ times (and are thus used to compute $\\mathcal{L}_{\\mathrm{LMM}}$ ) with their counts. ProsePose correctly identifies self-contact points and optimizes the poses to respect these contacts." + ], + "image_footnote": [], + "bbox": [ + 563, + 89, + 854, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall, our method improves over the other methods that do not use contact supervision in terms of both joint error and PCC. While not the focus of this work, Table 1 also shows that adding ProsePose, specifically $\\mathcal{L}_{\\mathrm{LMM}}$ , to BUDDI leads to improved PCC. The F1 scores show that LMMs' raw output is often flawed/incomplete, but our approach mitigates hallucinations in various ways (see the ablation study below). The last row in Table 1 shows the performance when the ground-truth coarse contacts (with correct left/right labels) are used in optimization. These results show the benefit of correct coarse contacts and the clear room for improvement from better LMM predictions.", + "bbox": [ + 511, + 580, + 908, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows the PCC for each method at various radii. The results show that ProsePose brings the meshes closer together at the correct contact points. On both the FlickrCI3D and CHI3D datasets, ProsePose outperforms the other baselines that do not use contact supervision. Next, we ablate important aspects of ProsePose. In Figure 5, we show that averaging the loss over several samples from the LMM improves performance, mitigating the effect of LMM hallucination. Finally, ablating the various losses in optimization indicates", + "bbox": [ + 511, + 763, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7131", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f7d56f72810ad803f3e384aa147ede287596822f2f8613606b608b9a13873c7.jpg", + "image_caption": [ + "Figure 5. More samples improve pose estimation. On the Flickr3D validation set, taking more samples from the LMM and averaging the resulting loss functions improves joint PA-MPJPE (left) and average PCC (right)." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 279, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e233f6ec7e030e3ae15bdffc97c715de01146a0c223a9a5e0832204716006c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 460, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/02065261a5d54f6145cf74809761f431644993a4e224aa8878bb9bef95abf6d2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PA-MPJPE↓PCC↑PCC↑@rF1↑
51015
HMR2 [12]8483.034.255.269.5-
HMR2+opt8185.247.765.574.6-
ProsePose8287.854.273.881.425
", + "bbox": [ + 91, + 286, + 480, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. One-person Results. PA-MPJPE (lower is better) and Avg. PCC and F1 (higher is better). ProsePose captures ground-truth contacts better than the baselines, as shown by the PCC.", + "bbox": [ + 89, + 378, + 483, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that our LMM-based loss and the 2D keypoint loss have the greatest impact on joint error: Using all losses results in a PA-MPJPE of 81. Removing $\\mathcal{L}_{\\mathrm{LMM}} / \\mathcal{L}_{\\mathrm{GMM}} / \\mathcal{L}_{\\beta} / \\mathcal{L}_{\\theta} / \\mathcal{L}_{2D} / \\mathcal{L}_P$ results in a PA-MPJPE of 138/85/91/84/130/78.", + "bbox": [ + 89, + 441, + 482, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results Figure 3 shows examples of reconstructions from ProsePose, Heuristic, and BUDDI. Below each of our predictions, we list the most common constraints predicted by GPT4-V for the image. The predicted constraints correctly capture the semantics of each interaction. For instance, in tango, one person's arm should touch the other's back. In a rugby tackle, a player's arms are usually wrapped around the other player. Using these constraints, ProsePose correctly reconstructs a variety of interactions, such as tackling, dancing, and holding hands. In contrast, Heuristic struggles to accurately position individuals and/or predict limb placements, often resulting in awkward distances.", + "bbox": [ + 89, + 510, + 485, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. One-person pose refinement", + "text_level": 1, + "bbox": [ + 89, + 717, + 344, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets Next, we evaluate ProsePose on a single-person setting. For this setting, we evaluate on MOYO [45], a motion capture dataset with videos of a single person performing various yoga poses. In total, our test set is composed of 76 examples from a single camera angle (side view). See Appendix 7.6 for further dataset details. Since this dataset does not have annotated region contact pairs, we compute the peso-ground-truth contact maps using the Euclidean and geodesic distance following [33] and report PCC for the subset of 67 examples where the ground-truth has self-contact.", + "bbox": [ + 89, + 750, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines We compare against the following baselines:", + "bbox": [ + 513, + 90, + 877, + 107 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- HMR2 [12] State-of-the-art pose regression method. We use HMR2 to initialize our pose estimates for optimization.", + "- HMR2+opt Optimization procedure that is identical to ours without $\\mathcal{L}_{\\mathrm{LMM}}$ . It is the default method when the number of empty constraint sets is at least the threshold $t$ ." + ], + "bbox": [ + 511, + 109, + 906, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Both the quantitative and qualitative results echo the trends discussed in the 2-person setting. Table 3 provides the quantitative results. The PCC metrics show that our LMM loss improves the predicted self-contact in complex yoga poses relative to the two baselines. The backoff method is used in 43/76 examples. Figure 4 provides a qualitative comparison of poses predicted by ProsePose versus the two baselines. Below each of our predictions, we list the corresponding constraints predicted by GPT4-V. In each case, the predicted constraint captures the correct self-contact, which is reflected in the final pose estimates. Using the semantically guided loss, ProsePose effectively refines the pose to ensure proper contact between hand-foot or hand-hand, an important detail consistently overlooked by the baselines.", + "bbox": [ + 511, + 189, + 906, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Limitations", + "text_level": 1, + "bbox": [ + 513, + 424, + 638, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While ProsePose consistently improves contact across settings and datasets, it has limitations which are related to failures of LMMs. First, as shown in Table 5 (in Appendix), prompting the LMM for left/right labels sometimes leads to worse results, suggesting that LMMs struggle with disambiguating chirality. Improving this approach depends in large part on correctly identifying limbs as left/right. Another limitation is the use of coarse regions. Future work could improve by eliciting more fine-grained constraints from an LMM. Finally, LMM accuracy varies moderately across camera angles (quantified in Appendix 8.2). In Appendix § 8.3, we provide examples of LMM failures.", + "bbox": [ + 511, + 450, + 908, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 659, + 633, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present ProsePose, a framework for refining 3D pose estimates to capture touch accurately using the implicit semantic knowledge of poses in LMMs. Our key novelty is that we generate structured pose descriptions from LMMs and convert them into loss functions used to optimize the pose. Our experiments show that in both one-person and two-person settings, ProsePose improves over previous baselines that do not use contact supervision. These results suggest that LMMs may be useful in creating larger datasets with contact annotations, which are otherwise expensive but are crucial for training state-of-the-art priors for pose estimation in situations with physical contact. More broadly, this work provides evidence that LMMs are promising tools for 3D pose estimation, which likely has implications beyond touch.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7132", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SS, EN, and TD were supported in part by the NSF, DoD, and/or the Berkeley Artificial Intelligence Research (BAIR) industrial alliance program.", + "bbox": [ + 89, + 116, + 485, + 162 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 183, + 187, + 199 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mo Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Benjamin Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Sim'on Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Raphael Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heating, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Hendrik Kirchner, Jamie Ryan Kiros, Matthew Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Adeola Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney Christine McLeavey Paul McMillan Jake McNeil David Medina Alok Mehta Jacob Menick Luke Metz Andrey Mishchenko Pamela Mishkin Vinnie Monaco Evan Morikawa Daniel P. Mossing Tong Mu Mira Murati Oleg Murk David M'ely Ashvin Nair Reichiro Nakano Rajeev Nayak Arvind Neelakantan Richard Ngo Hyeonwoo Noh Ouyang Long Cullen O'Keefe Jakub W. Pachocki Alex Paino Joe Palermo Ashley Pantuliano Giambattista Parascandolo Joel Parish Emy Parparita Alexandre Passos Mikhail Pavlov Andrew Peng Adam Perelman Filipe", + "bbox": [ + 98, + 205, + 485, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario D. Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin D. Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas A. Tezak, Madeleine Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cer'on Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao, Tao Xu, Sarah Yoo, Kevin Yu, Qiming Yuan, Wojciech Zaremba, Rowan Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. Gpt-4 technical report. 2023. 5", + "bbox": [ + 545, + 92, + 908, + 450 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "[2] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3395-3404, 2019. 2", + "[3] Ana Aznar and Harriet R Tenenbaum. Parent-child positive touch: Gender, age, and task differences. Journal of nonverbal behavior, 40:317-333, 2016. 1", + "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 561-578. Springer, 2016. 5", + "[5] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 5", + "[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 3", + "[7] Delmas, Ginger and Weinzaepfel, Philippe and Lucas, Thomas and Moreno-Noguer, Francesc and Rogez, Grégory. PoseScript: 3D Human Poses from Natural Language. In ECCV, 2022. 2", + "[8] Delmas, Ginger and Weinzaepfel, Philippe and Moreno-Noguer, Francesc and Rogez, Grégory. PoseFix: Correcting 3D Human Poses with Natural Language. In ICCV, 2023. 2" + ], + "bbox": [ + 521, + 454, + 908, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7133", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Posegpt: Chatting about 3d human pose. ArXiv, abs/2311.18836, 2023. 2", + "[10] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7214-7223, 2020. 2, 4, 5, 6", + "[11] Mihai Fieraru, Mihai Zanfir, Teodor Szente, Eduard Bazavan, Vlad Olaru, and Cristian Sminchisescu. Remips: Physically consistent 3d reconstruction of multiple interacting people under weak supervision. Advances in Neural Information Processing Systems, 34:19385-19397, 2021. 2", + "[12] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2, 5, 7, 8", + "[13] Peng Guan, Alexander Weiss, Alexandru O Balan, and Michael J Black. Estimating human shape and pose from a single image. In 2009 IEEE 12th International Conference on Computer Vision, pages 1381-1388. IEEE, 2009. 2", + "[14] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10884-10894, 2019. 2", + "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, 2022. 2", + "[16] Jinni A Harrigan. Self-touching as an indicator of underlying affect and language processes. Social science & medicine, 20 (11):1161-1168, 1985. 1", + "[17] Buzhen Huang, Chen Li, Chongyang Xu, Liang Pan, Yanggang Wang, and Gim Hee Lee. Closely interactive human reconstruction with proxemics and physics-guided adaption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1011-1021, 2024. 2", + "[18] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2", + "[19] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2020. 2", + "[20] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human model fitting towards inthe-wild 3d human pose estimation. In 2021 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2021. 2", + "[21] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Rawal Khirodkar, Jyun-Ting Song, Jinkun Cao, Zhengyi Luo, and Kris M. Kitani. Harmony4d: A video dataset for in-the-wild close human interactions. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2", + "[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 4", + "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3", + "[25] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2252-2261, 2019. 2", + "[26] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017. 2", + "[27] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv preprint arXiv:2311.16922, 2023. 3", + "[28] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3", + "[29] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023. 2", + "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3, 6", + "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.5,6", + "[32] Yujie Lu, Dongfu Jiang, Wenhu Chen, William Wang, Yejin Choi, and Bill Yuchen Lin. Wildvision arena: Benchmarking multimodal llms in the wild, 2024. 8", + "[33] Lea Muller, Ahmed AA Osman, Siyu Tang, Chun-Hao P Huang, and Michael J Black. On self-contact and human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9990-9999, 2021. 2, 8", + "[34] Lea Müller, Vickie Ye, Georgios Pavlakos, Michael Black, and Angjoo Kanazawa. Generative proxemics: A prior for 3d social interaction from images. arXiv preprint arXiv:2306.09337, 2023. 2, 4, 5, 6, 7, 9", + "[35] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2, 5", + "[36] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7134", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3, 4, 5", + "[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. *Big Data*, 4(4):236–252, 2016. 2", + "[38] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 722-731, 2021. 2", + "[39] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2", + "[40] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2", + "[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11179-11188, 2021. 2", + "[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13243-13252, 2022. 2, 5, 6", + "[43] Mika S Takeuchi, Hitoshi Miyaoka, Atsuko Tomoda, Masao Suzuki, Qingbo Liu, and Toshinori Kitamura. The effect of interpersonal touch during childhood on adult attachment and depression: A neglected area of family and developmental psychology? Journal of Child and Family Studies, 19:109-117, 2010. 1", + "[44] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[45] Shashank Tripathi, Lea Müller, Chun-Hao P. Huang, Taheri Omid, Michael J. Black, and Dimitrios Tzionas. 3D human pose estimation via intuitive physics. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 4713-4725, 2023. 8", + "[46] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 2022 International Conference on 3D Vision (3DV), pages 353-362. IEEE, 2022. 2, 5", + "[47] X Wang, J Wei, D Schuurmans, Q Le, E Chi, S Narang, A Chowdhery, and D Zhou. Self-consistency improves chain of thought reasoning in language models. arxiv. Preprint posted online March, 21:10-48550, 2022. 5", + "[48] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Huai hsin Chi, F. Xia, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. ArXiv, abs/2201.11903, 2022. 4" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In Advances in Neural Information Processing Systems, 2022. 5", + "[50] Yifei Yin, Chen Guo, Manuel Kaufmann, Juan José Zárate, Jie Song, and Otmar Hilliges. Hi4d: 4d instance segmentation of close human interaction. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17016-17027, 2023. 2, 6", + "[51] Andrei Zanfir, Elisabeta Marinhoiu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2148-2157, 2018. 2", + "[52] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + ], + "bbox": [ + 516, + 90, + 906, + 372 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7135", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Pose Priors from Language Models/9f701210-ac31-4714-b2b0-3e6ed2c2dfd9_model.json b/2025/Pose Priors from Language Models/9f701210-ac31-4714-b2b0-3e6ed2c2dfd9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..192fb99fcccc8dbbb9e9ee3a8bece2778cccf43c --- /dev/null +++ b/2025/Pose Priors from Language Models/9f701210-ac31-4714-b2b0-3e6ed2c2dfd9_model.json @@ -0,0 +1,2070 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.321, + 0.131, + 0.679, + 0.154 + ], + "angle": 0, + "content": "Pose Priors from Language Models" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.179, + 0.92, + 0.289 + ], + "angle": 0, + "content": "Sanjay Subramanian1 Evonne Ng1 Lea Müller1 Dan Klein1 Shiry Ginosar2,3 Trevor Darrell1 \n1 University of California, Berkeley \n2 Google DeepMind \n3 Toyota Technological Institute at Chicago \n{sanjayss,evonne_ng,mueller,klein,trevordarrell}@berkeley.edu \nshiry@google.com" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.308, + 0.905, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.532, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Figure 1. Optimizing human-to-human contacts in 3D pose. Our approach leverages the semantic priors of a Large Multimodal Model (LMM) to infer meaningful information about physical contact from images. Instead of relying on human annotations or motion capture data, we extract not only descriptive insights (\"... engaged in a dance or embrace ...\") but also structured constraints between body parts (underlined). By incorporating these LMM-derived constraints, we refine initial 3D human pose estimates, achieving realistic and semantically consistent reconstructions of contact. This scalable approach opens up new possibilities for contact-aware pose estimation without explicit contact annotations, making it a promising alternative to traditional methods." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.637, + 0.327, + 0.652 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.676, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Language is often used to describe physical interaction, yet most 3D human pose estimation methods overlook this rich source of information. We bridge this gap by leveraging large multimodal models (LMMs) as priors for reconstructing contact poses, offering a scalable alternative to traditional methods that rely on human annotations or motion capture data. Our approach extracts contact-relevant descriptors from an LMM and translates them into tractable losses to constrain 3D human pose optimization. Despite its simplicity, our method produces compelling reconstructions for both two-person interactions and self-contact scenarios, accurately capturing the semantics of physical and social interactions. Our results demonstrate that LMMs can serve as powerful tools for contact prediction and pose estimation, offering an alternative to costly manual human annotations" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.639, + 0.907, + 0.669 + ], + "angle": 0, + "content": "or motion capture data. Our code is publicly available at https://prosepose.github.io." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.696, + 0.646, + 0.711 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.721, + 0.907, + 0.842 + ], + "angle": 0, + "content": "Language, as a human artifact, encodes a rich set of social and physical interactions. Over centuries, our vocabulary has evolved to describe the nuances of touch, with words and phrases capturing contexts as varied as hugs, handshakes, or postures in sports and yoga. Perceiving physical contact is essential for understanding human behavior: e.g. several forms of parent-child contact are associated with affection [3, 43], and some forms of self-contact signal stress [16]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Since written language discusses our physical interactions at great length, can large multimodal models (LMMs) trained on images and text correctly perceive physical contact in human pose? This question has practical significance" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7125" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "because scenes with contact are challenging for pose estimation methods as some body parts are frequently occluded. This particularly holds for methods solely relying on 2D keypoints which do not convey contact information. Previously proposed approaches address these issues by curating task-specific datasets via motion capture or human-annotated points of contact between body parts [11, 33, 34]. However, collecting these datasets is expensive, and existing publicly available datasets include only tens of thousands of images [10, 22, 33, 50]. If LMMs can accurately identify contact points, they could decrease the cost of curating such datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.482, + 0.41 + ], + "angle": 0, + "content": "In this work, we study the efficacy of LMMs as tools for contact prediction in pose estimation. Since LMMs output language rather than pose parameters, answering this question requires a way of eliciting the required information from an LMM and operationalizing it in pose estimation. We introduce a framework, called ProsePose, which prompts an LMM for formatted constraints about physical contact in the image, converts the constraints into a loss function, and optimizes this loss function (jointly with losses from other cues, such as 2D keypoints) to refine initial pose estimates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.651 + ], + "angle": 0, + "content": "We use ProsePose to evaluate LMMs on both 2-person interaction datasets and a dataset of complex yoga poses. Our framework improves pose estimates compared to strong baselines that do not use contact supervision. In our extensive analysis, we show that several components are important: mitigating LMM hallucinations by aggregating predictions from several LMM samples, careful construction of the prompt and loss functions, and integrating losses from other cues. Finally, we conduct an extensive analysis of LMM predictions and their role in the optimization results. With respect to LMM failures, we find that identifying the chirality of limbs is a particular challenge for LMMs. While existing supervised methods excel by training on large amounts of supervised data with contact labels, we show that we can extract useful priors for contact prediction from pretrained LMMs without fine tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.742 + ], + "angle": 0, + "content": "In summary, our contributions are (1) we introduce a framework for applying LMMs as contact prediction tools in pose estimation, (2) we show that our framework can improve the quality of pose estimates in 2-person and 1-person settings, and (3) we provide an analysis of the components of our framework and LMM failure modes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.233, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "3D human pose reconstruction. Reconstructing 3D human poses from single images is an active area of research. Prior works have explored optimization-based approaches [13, 26, 35, 36, 39] or pure regression [2, 14, 20, 21, 25] to estimate the 3D body pose given a single image. HMR2 [12] is a recent state-of-the-art regression model in this line of work. Building on these approaches, some methods have looked into reconstructing multiple individuals jointly from a single" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "image. These methods [19, 41, 51] use deep networks to reason about multiple people in a scene to directly output multi-person 3D pose predictions. BEV [42] accounts for the relative proximity of people explicitly using relative depth annotations to reason about proxemics when placing each individual in the scene (e.g. relative depth of people). However, approaches in both categories generally do not accurately capture physical contact between parts of a single person or between people [33, 34]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Contact inference in 3D pose reconstruction. 3D pose reconstruction is especially challenging when there is self-contact or inter-person contact. This has motivated a line of work on pose reconstruction approaches tailored for these settings. [33] focuses on predicting self contact regions for 3D pose estimation by leveraging a dataset with contact annotations to model complex poses such as crossed arms. [10] introduces the first dataset with hand-annotated ground-truth contact labels between two people. REMIPS [11] and BUDDI [34] train models on the person-to-person contact maps in this data in order to improve 3D pose estimation of multiple people from a single image. CloseInt [17] trains a physics-guided diffusion model on two-person motion capture data for this task. However, contact annotations, which are crucial for these approaches, are expensive to acquire. Our method does not require any training on such annotations. Instead, we leverage an LMM's implicit knowledge of pose to constrain pose optimization to capture both self- and person-to-person contact." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Language priors on human pose. There exists a plethora of text to 3D human pose and motion datasets [15, 37, 38], which have enabled work focused on generating 3D motion sequences of a single person performing a general action [18, 44, 52]. This line of work has been extended to generating the motion of two people conditioned on text [29, 40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.853 + ], + "angle": 0, + "content": "PoseScript [7] is a method for generating a single person's pose from fine-grained descriptions, which uses training data from motion capture annotated with detailed text. PoseFix [8] introduces a labeled dataset for the task of modifying a pose given a fine-grained description of the desired change and trains a model on this data. PoseGPT [9] is a pose regressor that uses language as part of its training data. However, PoseGPT does not produce better pose estimates than previous state-of-the-art regressors (i.e. regressors that do not use language) and applies only to the one-person setting. [46] uses a text-only LM to improve action-conditioned human-object pose estimation. This method relies on a limited database of action-pose pairs to classify an input pose, and uses an LM to improve pose estimates based on the action retrieved from the database." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our work differs from previous work on language and pose in several ways. First, whereas all prior work relies on training data with pairs of language and pose, which" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7126" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.585, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.131, + 0.91, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.289, + 0.911, + 0.377 + ], + "angle": 0, + "content": "Figure 2. LMM-guided Pose Estimation. (a) Method overview: ProsePose takes as input an image of one or two people in contact. We first obtain initial pose estimates for each person from a pose regressor. Then we use an LMM to generate contact constraints, each of which is a pair of body parts that should be touching. This list of contacts is converted into a loss function \\(\\mathcal{L}_{\\mathrm{LMM}}\\). We optimize the pose estimates using \\(\\mathcal{L}_{\\mathrm{LMM}}\\) and other losses to produce a refined estimate of each person's pose that respects the predicted contacts. (b) Defining contact constraints: Given an image \\(\\pmb{I}\\), we can lift each individual into corresponding 3D meshes \\(\\pmb{V}\\). A contact constraint \\(\\pmb{c}\\) is a pair of regions \\((\\pmb{R}_a, \\pmb{R}_b)\\) in contact. The loss is defined in terms of the distance between the vertices \\((\\pmb{v}_a, \\pmb{v}_b)\\) on the mesh." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.487, + 0.522 + ], + "angle": 0, + "content": "is expensive to collect, our method leverages the existing knowledge in an LMM to reason about pose from a given image. Second, prior work in this area focuses on either the one-person or the two-person setting. In contrast, our work presents a single framework to reason about physical contacts within or between poses. Finally, in scenes with physical contact, we show that our method improves the pose estimates of state-of-the-art regressors." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.567, + 0.466, + 0.586 + ], + "angle": 0, + "content": "3. Guiding Pose Optimization with an LMM" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.603, + 0.487, + 0.725 + ], + "angle": 0, + "content": "Given an image, our goal is to estimate the 3D body pose of individuals in the image while capturing the self and cross-person contact points. While we cannot trivially use natural language responses (hug, kiss) to directly optimize 3D body poses, we leverage the key insight that LMMs understand how to articulate a given pose (arms around waist, lips touching). We propose a method to structure these articulations into constraints and convert them into loss functions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.487, + 0.903 + ], + "angle": 0, + "content": "More concretely, our framework, illustrated by Figure 2a, takes as input the image \\( \\pmb{I} \\) and the bounding boxes \\( B \\) of the subjects of interest. In the first stage, a pose regressor takes the image and produces a rough estimate of the 3D pose \\( X^p \\) for each individual \\( p \\) in the image. In the second stage, an LMM takes the image and a set of instructions and generates a list of self- or inter-person contact constraints, which we then convert into a loss function (Sec. 3.2). Finally, in the third stage, we jointly optimize the generated loss function with several other pre-defined loss terms (Sec. 3.3). We refer to our framework as ProsePose." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.399, + 0.655, + 0.415 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.909, + 0.515 + ], + "angle": 0, + "content": "We focus our description on the two-person case to keep the exposition simple. We also demonstrate results on the one-person case, which is simply an extension of the two-person case. In particular, we apply our method to the one-person case by setting \\( X^0 = X^1 \\). Please see Appendix § 7 for details on the differences between the two cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.909, + 0.663 + ], + "angle": 0, + "content": "Large Multimodal Models. An LMM is a model that takes as input an image and a text prompt and produces text output that answers the prompt based on the image. Our framework is agnostic to the architecture of the LMM. LMMs are typically trained to respond to a wide variety of instructions [6, 30]. However, LMMs are prone to hallucination [27, 28]. Handling cases of hallucination is a key challenge when using LMMs. We mitigate this issue by aggregating information across several samples from the LMM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Pose representation. We use a human body model [36] to represent each person \\(p \\in \\{0,1\\}\\). The body model is composed of a pose parameter that defines the joint rotations \\(\\theta \\in \\mathbb{R}^{d_{\\theta} \\times 3}\\), where \\(d_{\\theta}\\) is the number of joints, and a shape parameter \\(\\beta \\in \\mathbb{R}^{d_{\\beta}}\\), where \\(d_{\\beta}\\) is the dimensions of the shape parameter. We can apply a global rotation \\(\\Phi \\in \\mathbb{R}^3\\) and translation \\(t \\in \\mathbb{R}^3\\) to place each person in the world coordinate space. The full set of parameters for each person is denoted by \\(\\boldsymbol{X}^p = [\\theta^p, \\beta^p, \\Phi^p, t^p]\\). For simplicity, we refer to the parameter set \\((\\boldsymbol{X}^0, \\boldsymbol{X}^1)\\) as \\(\\boldsymbol{X}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.91, + 0.903 + ], + "angle": 0, + "content": "These parameters can be plugged into a differentiable function that maps to a mesh consisting of \\(d_v\\) vertices \\(\\mathbf{V} \\in \\mathbb{R}^{d_v \\times 3}\\). From the mesh, we can obtain the 3D locations of the body's joints \\(\\mathbf{J} \\in \\mathbb{R}^{d_j \\times 3}\\). From these joints, we can calculate the 2D keypoints \\(\\mathbf{K}_{proj}\\) by projecting the 3D joints to 2D" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7127" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.423, + 0.107 + ], + "angle": 0, + "content": "using the camera intrinsics \\(\\Pi\\) predicted from [36]." + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.117, + 0.484, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {K} _ {p r o j} = \\Pi (\\boldsymbol {J}) \\in \\mathbb {R} ^ {d _ {j} \\times 2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.159, + 0.485, + 0.37 + ], + "angle": 0, + "content": "Vertex regions. In order to define contact constraints between body parts, we define a set of regions of vertices. Prior work on contact has partitioned the body into fine-grained regions [10]. However, since our constraints are specified by a LMM trained on natural language, the referenced body parts are often coarser in granularity. We therefore update the set of regions to reflect this language bias by combining these fine-grained regions into larger, more commonly referenced body parts such as arm, shoulder (front&back), back, and waist (front&back). Some regions, e.g. back and waist (back), overlap. Please see Appendix § 7.2 for a visualization of the coarse regions. Formally, we write \\( \\mathbf{R} \\in \\mathbb{R}^{d_r \\times 3} \\) to denote a region with \\( d_r \\) vertices, which is part of the full mesh \\( (\\mathbf{R} \\subset V) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.468 + ], + "angle": 0, + "content": "Constraint definition. A contact constraint specifies which body parts from two meshes should touch. We define contact constraints as pairs of coarse regions \\( c = (R_a, R_b) \\) between a region \\( R_a \\) of one mesh and \\( R_b \\) of the other mesh, as shown in Figure 2b. For instance, (\"hand\", \"arm\") indicates a hand should touch an arm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.4, + 0.493 + ], + "angle": 0, + "content": "3.2. Constraint generation with a LMM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Our key insight is to leverage a LMM to identify regions of contact between different body parts on the human body surface. As shown in Figure 2a, we prompt the LMM with an image and ask it to output a list of all region pairs that are in contact. However, we cannot simply use its output natural language descriptions to directly optimize a 3D mesh. As such, we convert these constraints into a loss function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.78 + ], + "angle": 0, + "content": "LMM-based constraint generation. Given the image \\( I \\), we first use the bounding boxes \\( B \\) to crop the part containing the subjects. We then use an image segmentation model to mask any extraneous individuals. While cropping and masking the image may remove information, we find the LMMs are relatively robust to missing context, and more importantly, this allows us to indicate which individuals to focus on. Given the segmented image, we ask the LMM to generate a set \\( C = \\{c_1, \\dots, c_m\\} \\) of all pairs of body parts that are touching, where \\( m \\) is the total number of constraints the LMM generates for the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In the prompt, we specify the full set of coarse regions to pick from. We find that LMMs fail to reliably reference the left and right limbs correctly or consistently, so the prompt instructs the LMM not to specify chirality for each limb (see Appendix 8.2 for prompt analysis). If the LMM uses \"left\" or \"right\" to reference a region, despite an instruction to not do so, we directly use the part of the region with the specified chirality rather than considering both possibilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "Motivated by the chain-of-thought technique, which has been shown to improve language model performance on reasoning tasks [48], we ask the LMM to write its reasoning or describe the pose before listing the constraints. For the full prompt used in each setting, please refer to Appendix § 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.908, + 0.229 + ], + "angle": 0, + "content": "We sample \\(N\\) responses from the LMM. Below we describe (1) how we convert these natural language responses into \\(N\\) sets of constraints \\(\\{C_1, C_2, \\dots, C_N\\}\\) and (2) how we convert each constraint set \\(C_j\\) into a loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.908, + 0.344 + ], + "angle": 0, + "content": "Canonicalizing Region Names and Assigning Chirality. Given the LMM's output, we must map the mentioned region names to our fixed set of coarse regions. Since the LMM may deviate from the names in the prompt, we check for some additional names (see Appendix 7.2). We then filter out contact pairs that occur fewer than \\( f \\) times across constraint sets, where \\( f \\) is a hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.909, + 0.466 + ], + "angle": 0, + "content": "Next, we assign a chirality (left/right) to each hand/arm/foot/leg/shoulder region in cases when the LMM does not itself specify the chirality. We enumerate all possible assignments of left/right to these regions and take the one resulting in the minimum loss. In the two-person setting, we only consider assignments satisfying a condition designed for the case in which a region type occurs in multiple constraint pairs (see Appendix 7.3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.909, + 0.627 + ], + "angle": 0, + "content": "Loss function generation. We compute a loss for each constraint by mapping the relevant regions to sets of vertices and calculating the minimum distance between vertices in the two sets. In particular, we first specify a mapping between each of the coarse region names and the fine-grained regions from [10]. We then use a mapping from fine-grained regions to SMPL-X vertices (provided by [34]) to obtain a set of vertices for each coarse region. Then for each contact pair of coarse regions \\( c = (R_a, R_b) \\) in \\( C_j \\), we define \\( dist(c) \\) as the minimum distance between the two regions:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.64, + 0.907, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\operatorname {d i s t} (\\boldsymbol {c}) = \\min \\| \\boldsymbol {v} _ {a} - \\boldsymbol {v} _ {b} \\| _ {2} \\quad \\forall \\boldsymbol {v} _ {a} \\in \\boldsymbol {R} _ {a}, \\forall \\boldsymbol {v} _ {b} \\in \\boldsymbol {R} _ {b} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.907, + 0.745 + ], + "angle": 0, + "content": "where \\(\\{\\pmb{v}_a, \\pmb{v}_b\\} \\in \\mathbb{R}^3\\). In practice, the number of vertices in each region can be very large. To make this computation tractable, we first take a random sample of vertices from \\(R_a\\) and from \\(R_b\\) before computing distances between pairs of vertices in these samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Furthermore, since the ordering of the people in the LMM constraints is unknown (i.e. does \\( \\mathbf{R}_a \\) come from the mesh defined by parameter \\( \\mathbf{X}^0 \\) or \\( \\mathbf{X}^1 \\)), we compute the overall loss for both possibilities and take the minimum. We use \\( \\mathbf{c}^\\top = (\\mathbf{R}_b, \\mathbf{R}_a) \\) to denote the flipped ordering. We then sum over all constraints in the list \\( C_j \\):" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.85, + 0.907, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\operatorname {d i s t} _ {\\text {s u m}} \\left(\\boldsymbol {C} _ {j}\\right) = \\min \\left(\\sum_ {\\boldsymbol {c} \\in \\boldsymbol {C} _ {j}} \\operatorname {d i s t} (\\boldsymbol {c}), \\sum_ {\\boldsymbol {c} \\in \\boldsymbol {C} _ {j}} \\operatorname {d i s t} \\left(\\boldsymbol {c} ^ {\\top}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7128" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "Each constraint set sampled from the LMM is likely to contain noise or hallucination. To mitigate this issue, we average over all \\(N\\) losses corresponding to each constraint set to obtain the overall LMM loss. This technique is similar to self-consistency [47], which is commonly used for code generation. Concretely, the overall LMM loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.203, + 0.484, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L M M}} = \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\operatorname {d i s t} _ {s u m} \\left(C _ {j}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.255, + 0.485, + 0.392 + ], + "angle": 0, + "content": "If a constraint set \\( C_j \\) is empty (i.e. the LMM does not suggest any contact pairs), then we set \\( dist_{\\mathrm{sum}}(C_j) = 0 \\). If there are several such constraint sets, we infer that the LMM has low confidence about the contact points (if any) in the image. Consequently, we set a threshold \\( t \\) and if the number of empty constraint sets is at least \\( t \\), we gracefully backoff to the appropriate baseline optimization procedure (described in Sections 4.1 and 4.2 for each setting). We also backoff to the baseline if the LMM-based optimization diverges." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.399, + 0.363, + 0.417 + ], + "angle": 0, + "content": "3.3. Constrained pose optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.485, + 0.485 + ], + "angle": 0, + "content": "Drawing from previous optimization-based approaches [4, 34, 35], we employ several additional losses in the optimization. We then minimize the joint loss to obtain a refined subset of the body model parameters \\( \\mathbf{X}^{\\prime} = [\\pmb{\\theta}^{\\prime},\\pmb{\\beta}^{\\prime},\\pmb{t}^{\\prime}] \\):" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.492, + 0.479, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} [ \\boldsymbol {\\theta} ^ {\\prime}, \\boldsymbol {\\beta} ^ {\\prime}, \\boldsymbol {t} ^ {\\prime} ] = \\arg \\min (\\lambda_ {\\mathrm {L M M}} \\mathcal {L} _ {\\mathrm {L M M}} + \\lambda_ {\\mathrm {G M M}} \\mathcal {L} _ {\\mathrm {G M M}} + \\lambda_ {\\beta} \\mathcal {L} _ {\\beta} \\\\ + \\lambda_ {\\theta} \\mathcal {L} _ {\\theta} + \\lambda_ {2 D} \\mathcal {L} _ {2 D} + \\lambda_ {P} \\mathcal {L} _ {P}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.599 + ], + "angle": 0, + "content": "Following [34], we divide the optimization into two stages. In the first stage, we optimize all three parameters. In the second stage, we optimize only \\(\\theta\\) and \\(t\\), keeping the shape \\(\\beta\\) fixed. We detail the remaining losses below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.667 + ], + "angle": 0, + "content": "Pose and shape priors. We compute a loss \\(\\mathcal{L}_{\\mathrm{GMM}}\\) based on the Gaussian Mixture pose prior of [4] and a shape loss \\(\\mathcal{L}_{\\beta} = \\| \\beta \\|_2^2\\), which penalizes extreme deviations from the body model's mean shape." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.719 + ], + "angle": 0, + "content": "Initial pose loss. To ensure we do not stray too far from the initialization, we penalize large deviations from the initial pose \\(\\mathcal{L}_{\\theta} = ||\\pmb{\\theta}' - \\pmb{\\theta}||_2^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.486, + 0.845 + ], + "angle": 0, + "content": "2D keypoint loss. Similar to BUDDI [34], for each person in the image, we obtain pseudo ground truth 2D keypoints and their confidence from OpenPose [5] and ViTPose [49]. Given this pseudo ground truth, we merge all the keypoints into \\( \\pmb{K} \\in \\mathbb{R}^{d_j \\times 2} \\), and their corresponding confidence into \\( \\gamma \\in \\mathbb{R}^{d_j} \\). From the predicted \\( \\pmb{X}' \\), we can compute the 2D projection of each 3D joint location using Equation 3.1. Then, the 2D keypoint loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.854, + 0.484, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2 D} = \\sum_ {j = 1} ^ {d _ {j}} \\boldsymbol {\\gamma} \\left(\\boldsymbol {K} _ {p r o j} - \\boldsymbol {K}\\right) ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": "Interpenetration loss. To prevent parts of one mesh from being in the interior of the other, we add an interpenetration loss. Generically, given two sets of vertices \\( V_{0} \\) and \\( V_{1} \\), we use winding numbers to compute the subset of \\( V_{0} \\) that intersects \\( V_{1} \\), which we denote as \\( V_{0,1} \\). Similarly, \\( V_{1,0} \\) is the subset of \\( V_{1} \\) that intersects \\( V_{0} \\). The interpenetration loss is then defined as" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.207, + 0.907, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P} = \\sum_ {x \\in \\boldsymbol {V} _ {0, 1}} \\min _ {v _ {1} \\in \\boldsymbol {V} _ {1}} \\| x - v _ {1} \\| _ {2} ^ {2} + \\sum_ {y \\in \\boldsymbol {V} _ {1, 0}} \\min _ {v _ {0} \\in \\boldsymbol {V} _ {0}} \\| y - v _ {0} \\| _ {2} ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.269, + 0.909, + 0.3 + ], + "angle": 0, + "content": "For efficiency, this loss is computed on low-resolution versions of the two meshes (roughly 1000 vertices per mesh)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.314, + 0.646, + 0.332 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.908, + 0.454 + ], + "angle": 0, + "content": "Implementation details. Following prior work on two-person pose estimation [34], we use BEV [42] to initialize the poses since it was trained to predict both the body pose parameters and the placement of each person in the scene. However, on the single person yoga poses, we find that the pose parameter estimates of HMR2 [12] are much higher quality, so we initialize the body pose using HMR2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.909, + 0.668 + ], + "angle": 0, + "content": "We use the SMPL-X [36] body model and (unless specified otherwise) GPT4-V [1] as the LMM with temperature \\(= 0.7\\) when sampling from it. GPT4-V refers to the gpt-4-vision-preview model in the OpenAI API: platform.openai.com. In the OpenAI API, we use the \"high\" detail setting for image input. Appendix 8.2 provides results using other prompts and other LMMs (LLaVA [31], GPT-4o) and a running time analysis. Unless otherwise specified, we set \\(N = 20\\) samples. For all of our 2-person experiments, \\(f = 1\\), while \\(f = 10\\) in the 1-person setting. We set \\(t = 2\\) for the experiment on the CHI3D dataset and \\(t = N\\) for all other experiments. The hyperparameters and our prompts were chosen based on experiments on the validation sets. For other implementation details refer to Appendix § 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Metrics. As is standard in the pose estimation literature, we report Procrustes-aligned Mean Per Joint Position Error (PA-MPJPE) in millimeters. This metric finds the best alignment between the estimated and ground-truth pose before computing the joint error. In the two-person setting, we focus on the joint PA-MPJPE, as this evaluation incorporates the relative translation and orientation of the two people. See Appendix § 8.2 for the per-person PA-MPJPE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We also include the percentage of correct contact points (PCC) metric introduced by [34]. This metric captures the fraction of ground-truth contact pairs that are accurately predicted. For a given radius \\( r \\), a pair is classified as \"in contact\" if the two regions are both within the specified radius. We use the set of fine-grained regions defined in [10] to compute PCC. The metric is averaged over \\( r \\in \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7129" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.487, + 0.249 + ], + "angle": 0, + "content": "
Hi4DFlickrCI3DCHI3D
PM↓F1↑PM↓PCC↑F1↑PM↓PCC↑F1↑
w/o contact sup.
BEV [42]144-10664.8-9671.4-
Heuristic116-6777.8-10574.1-
ProsePose93245879.91310075.823
w. contact sup.
BUDDI [34]89-6681.9-6878.0-
BUDDI+ProsePose88-6583.2-6978.8-
Coarse GT contacts
Oracle8110043861008383.8100
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.26, + 0.485, + 0.386 + ], + "angle": 0, + "content": "Table 1. Two-person Results. Joint PA-MPJPE (abbreviated PM) (lower is better), Avg. PCC (higher is better), and F1 (higher is better). For FlickrCI3D, PA-MPJPE is computed using the pseudoground-truth fits. F1 measures the accuracy of coarse contacts predicted by the LMM, while the other two metrics evaluate the quality of the estimated 3D pose. Last line shows results using ground-truth contact pairs of coarse regions (heuristic is still used as the backoff method when needed). Bold indicates best method without/with contact supervision in each column." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.485, + 0.478 + ], + "angle": 0, + "content": "0, 5, 10, 15, ..., 95 mm. Since these regions are defined on the SMPL-X mesh topology, we convert the regression baselines-BEV and HMR2- from the SMPL mesh topology to SMPL-X to compute this metric. Please see the Appendix § 8.1 for more details on the regions and on the mesh conversion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.676 + ], + "angle": 0, + "content": "Finally, we report the F1 of the LMM's predicted coarse region pairs. Specifically, we compute precision as the proportion of predicted pairs \\((\\mathbf{R}_a,\\mathbf{R}_b)\\) such that there is some ground-truth pair \\((\\mathbf{R}_a^*,\\mathbf{R}_b^*)\\) for which \\(\\mathbf{R}_a\\) overlaps with \\(\\mathbf{R}_a^*\\) and \\(\\mathbf{R}_b\\) overlaps with \\(\\mathbf{R}_b^*\\). We compute recall similarly as a proportion of ground-truth pairs. For datasets that do not provide contact maps, we define the ground-truth pairs from the ground-truth meshes, using a threshold on the minimum distance between regions (0.01 and 0.02 for Hi4D and MOYO, respectively). We ignore chirality when computing these. For a pair of people we take the maximum F1 of the two possible orderings. This metric serves to evaluate the raw output of the LMMs without 3D optimization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.687, + 0.35, + 0.703 + ], + "angle": 0, + "content": "4.1. Two-person Pose Refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Datasets. We evaluate on three datasets, and our dataset processing largely follows [34]. Hi4D [50] is a motion capture dataset of pairs of people interacting. Each sequence has a subset of frames marked as contact frames, and we take every fifth contact frame. We use the images from a single camera, resulting in 241 images. Flickr Close Interactions 3D (FlickrCI3D) [10] is a collection of Flickr images of multiple people in close interaction. The dataset includes manual annotations of the contact maps between pairs of people. [34] used these contact maps to create pseudo-ground truth 3D meshes and curated a version of the test set to exclude noisy annotations, which has 1403 images. CHI3D [10] is" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.22 + ], + "angle": 0, + "content": "
@ radius[mm]FlickrCI3D PCC↑CHI3D PCC↑
5101551015
W/o contact sup.
BEV [42]3.66.310.85.817.432.5
Heuristic14.633.949.311.128.045.3
ProsePose15.639.957.113.535.252.5
W/ contact sup.
BUDDI [34]18.544.261.815.539.056.6
BUDDI+ProsePose21.849.366.419.543.958.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.512, + 0.23, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Table 2. Two-person PCC. Percent of correct contact points (PCC) for three different radii \\( r \\) in mm. Bold indicates the best score without/with contact supervision in each column. At the ground-truth contact points, our method brings the meshes closer together than the baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.908, + 0.378 + ], + "angle": 0, + "content": "a motion capture dataset of pairs of people interacting. We present results on the validation set. There are 431 images, distributed across 4 cameras. The images come from 126 video sequences, each of which has a single \"contact frame.\"" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.909, + 0.501 + ], + "angle": 0, + "content": "To develop our method, we experimented on the validation sets of FlickrCI3D and Hi4D, and a sample of the training set from CHI3D. For our experiments, we can compute the PCC on FlickrCI3D and CHI3D, which have annotated ground-truth contact maps. Following [34], we exclude from evaluation images where BEV or the keypoint detectors, which are used by the baselines as well, fail to detect one of the subjects in the interaction pair." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.511, + 0.908, + 0.527 + ], + "angle": 0, + "content": "Baselines We compare our estimated poses to the following:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.528, + 0.907, + 0.588 + ], + "angle": 0, + "content": "- BEV [42] Multi-person 3D pose estimation method. Uses relative depth to reason about spatial placement of individuals in the scene. ProsePose, Heuristic, and BUDDI use BEV to initialize pose estimates." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.589, + 0.908, + 0.693 + ], + "angle": 0, + "content": "- Heuristic A contact heuristic which includes the auxiliary losses in Section 3.3 as well as a term that minimizes the minimum distance between the two meshes. Introduced by [34]. We use their hyperparameters for this heuristic. This baseline is also used as the backoff method for Prose-Pose when the number of empty constraint sets is at least the threshold \\( t \\) or when the optimization diverges." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.694, + 0.908, + 0.755 + ], + "angle": 0, + "content": "- BUDDI [34] This method uses a learned diffusion prior to constrain the optimization. We stress that BUDDI requires a large amount of annotated training data on pairs of interacting bodies, which is not used in our method." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.528, + 0.908, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results Table 1 provides quantitative results on the three datasets. Across datasets, ProsePose consistently improves over the strongest baseline, Heuristic. On the Hi4D dataset, ProsePose reduces \\(85\\%\\) of the gap in PA-MPJPE between Heuristic and the fully supervised BUDDI. On the FlickrCI3D and CHI3D datasets, ProsePose narrows the gap in the average PCC between Heuristic and BUDDI by more than one-third. (While ProsePose achieves a better PA-MPJPE than BUDDI on FlickrCI3D, for this dataset," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7130" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.468, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.592, + 0.484, + 0.678 + ], + "angle": 0, + "content": "Figure 3. Two-person examples We show qualitative results from ProsePose, BUDDI [34], and the contact heuristic. For each example, we show GPT4-V's top 3 constraints and the number of times each constraint was predicted across all 20 samples. Our method correctly reconstructs people in a variety of interactions, and the predicted constraints generally align with each interaction type." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.731 + ], + "angle": 0, + "content": "we rely primarily on PCC since PA-MPJPE is computed on pseudo-ground-truth fits.)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "On CHI3D, ProsePose outperforms Heuristic but underperforms BEV in terms of PA-MPJPE. On the subset of images where we do not default to the heuristic (i.e. on images where GPT4-V predicts enough non-empty constraint sets), the PA-MPJPE for ProsePose and BEV is 86 and 87, respectively. In other words, in the cases where our method is actually used, the joint error is slightly less than that of BEV. As a result, we can attribute the worse overall error to the poorer performance of the heuristic. The backoff method (which is the heuristic) is used in 13/241 Hi4D examples, 106/1403 Flickr examples, and 224/431 CHI3D examples." + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.09, + 0.856, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.477, + 0.908, + 0.562 + ], + "angle": 0, + "content": "Figure 4. Single-person examples We show qualitative results from ProsePose, HMR2 [12], and HMR2-optim on complex yoga poses. Each example also shows the constraints that are predicted by the LMM at least \\( f = 10 \\) times (and are thus used to compute \\( \\mathcal{L}_{\\mathrm{LMM}} \\)) with their counts. ProsePose correctly identifies self-contact points and optimizes the poses to respect these contacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.909, + 0.763 + ], + "angle": 0, + "content": "Overall, our method improves over the other methods that do not use contact supervision in terms of both joint error and PCC. While not the focus of this work, Table 1 also shows that adding ProsePose, specifically \\(\\mathcal{L}_{\\mathrm{LMM}}\\), to BUDDI leads to improved PCC. The F1 scores show that LMMs' raw output is often flawed/incomplete, but our approach mitigates hallucinations in various ways (see the ablation study below). The last row in Table 1 shows the performance when the ground-truth coarse contacts (with correct left/right labels) are used in optimization. These results show the benefit of correct coarse contacts and the clear room for improvement from better LMM predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Table 2 shows the PCC for each method at various radii. The results show that ProsePose brings the meshes closer together at the correct contact points. On both the FlickrCI3D and CHI3D datasets, ProsePose outperforms the other baselines that do not use contact supervision. Next, we ablate important aspects of ProsePose. In Figure 5, we show that averaging the loss over several samples from the LMM improves performance, mitigating the effect of LMM hallucination. Finally, ablating the various losses in optimization indicates" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7131" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.28, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.462, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.214, + 0.486, + 0.271 + ], + "angle": 0, + "content": "Figure 5. More samples improve pose estimation. On the Flickr3D validation set, taking more samples from the LMM and averaging the resulting loss functions improves joint PA-MPJPE (left) and average PCC (right)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.287, + 0.482, + 0.369 + ], + "angle": 0, + "content": "
PA-MPJPE↓PCC↑PCC↑@rF1↑
51015
HMR2 [12]8483.034.255.269.5-
HMR2+opt8185.247.765.574.6-
ProsePose8287.854.273.881.425
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.38, + 0.485, + 0.422 + ], + "angle": 0, + "content": "Table 3. One-person Results. PA-MPJPE (lower is better) and Avg. PCC and F1 (higher is better). ProsePose captures ground-truth contacts better than the baselines, as shown by the PCC." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.483, + 0.503 + ], + "angle": 0, + "content": "that our LMM-based loss and the 2D keypoint loss have the greatest impact on joint error: Using all losses results in a PA-MPJPE of 81. Removing \\(\\mathcal{L}_{\\mathrm{LMM}} / \\mathcal{L}_{\\mathrm{GMM}} / \\mathcal{L}_{\\beta} / \\mathcal{L}_{\\theta} / \\mathcal{L}_{2D} / \\mathcal{L}_P\\) results in a PA-MPJPE of 138/85/91/84/130/78." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.486, + 0.707 + ], + "angle": 0, + "content": "Qualitative Results Figure 3 shows examples of reconstructions from ProsePose, Heuristic, and BUDDI. Below each of our predictions, we list the most common constraints predicted by GPT4-V for the image. The predicted constraints correctly capture the semantics of each interaction. For instance, in tango, one person's arm should touch the other's back. In a rugby tackle, a player's arms are usually wrapped around the other player. Using these constraints, ProsePose correctly reconstructs a variety of interactions, such as tackling, dancing, and holding hands. In contrast, Heuristic struggles to accurately position individuals and/or predict limb placements, often resulting in awkward distances." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.718, + 0.345, + 0.735 + ], + "angle": 0, + "content": "4.2. One-person pose refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Datasets Next, we evaluate ProsePose on a single-person setting. For this setting, we evaluate on MOYO [45], a motion capture dataset with videos of a single person performing various yoga poses. In total, our test set is composed of 76 examples from a single camera angle (side view). See Appendix 7.6 for further dataset details. Since this dataset does not have annotated region contact pairs, we compute the peso-ground-truth contact maps using the Euclidean and geodesic distance following [33] and report PCC for the subset of 67 examples where the ground-truth has self-contact." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.879, + 0.108 + ], + "angle": 0, + "content": "Baselines We compare against the following baselines:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.111, + 0.907, + 0.141 + ], + "angle": 0, + "content": "- HMR2 [12] State-of-the-art pose regression method. We use HMR2 to initialize our pose estimates for optimization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.142, + 0.907, + 0.187 + ], + "angle": 0, + "content": "- HMR2+opt Optimization procedure that is identical to ours without \\(\\mathcal{L}_{\\mathrm{LMM}}\\). It is the default method when the number of empty constraint sets is at least the threshold \\(t\\)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.111, + 0.907, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.908, + 0.403 + ], + "angle": 0, + "content": "Both the quantitative and qualitative results echo the trends discussed in the 2-person setting. Table 3 provides the quantitative results. The PCC metrics show that our LMM loss improves the predicted self-contact in complex yoga poses relative to the two baselines. The backoff method is used in 43/76 examples. Figure 4 provides a qualitative comparison of poses predicted by ProsePose versus the two baselines. Below each of our predictions, we list the corresponding constraints predicted by GPT4-V. In each case, the predicted constraint captures the correct self-contact, which is reflected in the final pose estimates. Using the semantically guided loss, ProsePose effectively refines the pose to ensure proper contact between hand-foot or hand-hand, an important detail consistently overlooked by the baselines." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.425, + 0.64, + 0.44 + ], + "angle": 0, + "content": "4.3. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.909, + 0.634 + ], + "angle": 0, + "content": "While ProsePose consistently improves contact across settings and datasets, it has limitations which are related to failures of LMMs. First, as shown in Table 5 (in Appendix), prompting the LMM for left/right labels sometimes leads to worse results, suggesting that LMMs struggle with disambiguating chirality. Improving this approach depends in large part on correctly identifying limbs as left/right. Another limitation is the use of coarse regions. Future work could improve by eliciting more fine-grained constraints from an LMM. Finally, LMM accuracy varies moderately across camera angles (quantified in Appendix 8.2). In Appendix § 8.3, we provide examples of LMM failures." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.66, + 0.634, + 0.675 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We present ProsePose, a framework for refining 3D pose estimates to capture touch accurately using the implicit semantic knowledge of poses in LMMs. Our key novelty is that we generate structured pose descriptions from LMMs and convert them into loss functions used to optimize the pose. Our experiments show that in both one-person and two-person settings, ProsePose improves over previous baselines that do not use contact supervision. These results suggest that LMMs may be useful in creating larger datasets with contact annotations, which are otherwise expensive but are crucial for training state-of-the-art priors for pose estimation in situations with physical contact. More broadly, this work provides evidence that LMMs are promising tools for 3D pose estimation, which likely has implications beyond touch." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7132" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.117, + 0.486, + 0.164 + ], + "angle": 0, + "content": "SS, EN, and TD were supported in part by the NSF, DoD, and/or the Berkeley Artificial Intelligence Research (BAIR) industrial alliance program." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.184, + 0.188, + 0.2 + ], + "angle": 0, + "content": "References" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.207, + 0.486, + 0.904 + ], + "angle": 0, + "content": "[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mo Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Benjamin Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Sim'on Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Raphael Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heating, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Hendrik Kirchner, Jamie Ryan Kiros, Matthew Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Adeola Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney Christine McLeavey Paul McMillan Jake McNeil David Medina Alok Mehta Jacob Menick Luke Metz Andrey Mishchenko Pamela Mishkin Vinnie Monaco Evan Morikawa Daniel P. Mossing Tong Mu Mira Murati Oleg Murk David M'ely Ashvin Nair Reichiro Nakano Rajeev Nayak Arvind Neelakantan Richard Ngo Hyeonwoo Noh Ouyang Long Cullen O'Keefe Jakub W. Pachocki Alex Paino Joe Palermo Ashley Pantuliano Giambattista Parascandolo Joel Parish Emy Parparita Alexandre Passos Mikhail Pavlov Andrew Peng Adam Perelman Filipe" + }, + { + "type": "text", + "bbox": [ + 0.546, + 0.093, + 0.909, + 0.452 + ], + "angle": 0, + "content": "de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario D. Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin D. Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas A. Tezak, Madeleine Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cer'on Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao, Tao Xu, Sarah Yoo, Kevin Yu, Qiming Yuan, Wojciech Zaremba, Rowan Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. Gpt-4 technical report. 2023. 5" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.455, + 0.909, + 0.524 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3395-3404, 2019. 2" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.527, + 0.907, + 0.569 + ], + "angle": 0, + "content": "[3] Ana Aznar and Harriet R Tenenbaum. Parent-child positive touch: Gender, age, and task differences. Journal of nonverbal behavior, 40:317-333, 2016. 1" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.571, + 0.909, + 0.668 + ], + "angle": 0, + "content": "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 561-578. Springer, 2016. 5" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.671, + 0.909, + 0.727 + ], + "angle": 0, + "content": "[5] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 5" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.729, + 0.909, + 0.799 + ], + "angle": 0, + "content": "[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 3" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.801, + 0.909, + 0.857 + ], + "angle": 0, + "content": "[7] Delmas, Ginger and Weinzaepfel, Philippe and Lucas, Thomas and Moreno-Noguer, Francesc and Rogez, Grégory. PoseScript: 3D Human Poses from Natural Language. In ECCV, 2022. 2" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.859, + 0.909, + 0.902 + ], + "angle": 0, + "content": "[8] Delmas, Ginger and Weinzaepfel, Philippe and Moreno-Noguer, Francesc and Rogez, Grégory. PoseFix: Correcting 3D Human Poses with Natural Language. In ICCV, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.455, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7133" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[9] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Posegpt: Chatting about 3d human pose. ArXiv, abs/2311.18836, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[10] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7214-7223, 2020. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.484, + 0.277 + ], + "angle": 0, + "content": "[11] Mihai Fieraru, Mihai Zanfir, Teodor Szente, Eduard Bazavan, Vlad Olaru, and Cristian Sminchisescu. Remips: Physically consistent 3d reconstruction of multiple interacting people under weak supervision. Advances in Neural Information Processing Systems, 34:19385-19397, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.281, + 0.484, + 0.349 + ], + "angle": 0, + "content": "[12] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.353, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[13] Peng Guan, Alexander Weiss, Alexandru O Balan, and Michael J Black. Estimating human shape and pose from a single image. In 2009 IEEE 12th International Conference on Computer Vision, pages 1381-1388. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.411, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[14] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10884-10894, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[16] Jinni A Harrigan. Self-touching as an indicator of underlying affect and language processes. Social science & medicine, 20 (11):1161-1168, 1985. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.482, + 0.667 + ], + "angle": 0, + "content": "[17] Buzhen Huang, Chen Li, Chongyang Xu, Liang Pan, Yanggang Wang, and Gim Hee Lee. Closely interactive human reconstruction with proxemics and physics-guided adaption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1011-1021, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.671, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[18] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.484, + 0.784 + ], + "angle": 0, + "content": "[19] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.787, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[20] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human model fitting towards inthe-wild 3d human pose estimation. In 2021 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[21] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.161 + ], + "angle": 0, + "content": "[22] Rawal Khirodkar, Jyun-Ting Song, Jinkun Cao, Zhengyi Luo, and Kris M. Kitani. Harmony4d: A video dataset for in-the-wild close human interactions. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.908, + 0.245 + ], + "angle": 0, + "content": "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.908, + 0.315 + ], + "angle": 0, + "content": "[25] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2252-2261, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.907, + 0.385 + ], + "angle": 0, + "content": "[26] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.907, + 0.453 + ], + "angle": 0, + "content": "[27] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv preprint arXiv:2311.16922, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.907, + 0.509 + ], + "angle": 0, + "content": "[28] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.907, + 0.565 + ], + "angle": 0, + "content": "[29] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.907, + 0.593 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.908, + 0.636 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[32] Yujie Lu, Dongfu Jiang, Wenhu Chen, William Wang, Yejin Choi, and Bill Yuchen Lin. Wildvision arena: Benchmarking multimodal llms in the wild, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.907, + 0.734 + ], + "angle": 0, + "content": "[33] Lea Muller, Ahmed AA Osman, Siyu Tang, Chun-Hao P Huang, and Michael J Black. On self-contact and human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9990-9999, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.907, + 0.789 + ], + "angle": 0, + "content": "[34] Lea Müller, Vickie Ye, Georgios Pavlakos, Michael Black, and Angjoo Kanazawa. Generative proxemics: A prior for 3d social interaction from images. arXiv preprint arXiv:2306.09337, 2023. 2, 4, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[35] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[36] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7134" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. *Big Data*, 4(4):236–252, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[38] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 722-731, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[39] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[40] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.484, + 0.433 + ], + "angle": 0, + "content": "[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11179-11188, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.484, + 0.504 + ], + "angle": 0, + "content": "[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13243-13252, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[43] Mika S Takeuchi, Hitoshi Miyaoka, Atsuko Tomoda, Masao Suzuki, Qingbo Liu, and Toshinori Kitamura. The effect of interpersonal touch during childhood on adult attachment and depression: A neglected area of family and developmental psychology? Journal of Child and Family Studies, 19:109-117, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[44] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[45] Shashank Tripathi, Lea Müller, Chun-Hao P. Huang, Taheri Omid, Michael J. Black, and Dimitrios Tzionas. 3D human pose estimation via intuitive physics. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 4713-4725, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[46] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 2022 International Conference on 3D Vision (3DV), pages 353-362. IEEE, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[47] X Wang, J Wei, D Schuurmans, Q Le, E Chi, S Narang, A Chowdhery, and D Zhou. Self-consistency improves chain of thought reasoning in language models. arxiv. Preprint posted online March, 21:10-48550, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[48] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Huai hsin Chi, F. Xia, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. ArXiv, abs/2201.11903, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[49] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In Advances in Neural Information Processing Systems, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[50] Yifei Yin, Chen Guo, Manuel Kaufmann, Juan José Zárate, Jie Song, and Otmar Hilliges. Hi4d: 4d instance segmentation of close human interaction. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17016-17027, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[51] Andrei Zanfir, Elisabeta Marinhoiu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2148-2157, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.907, + 0.373 + ], + "angle": 0, + "content": "[52] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7135" + } + ] +] \ No newline at end of file diff --git a/2025/Pose Priors from Language Models/9f701210-ac31-4714-b2b0-3e6ed2c2dfd9_origin.pdf b/2025/Pose Priors from Language Models/9f701210-ac31-4714-b2b0-3e6ed2c2dfd9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c6d01f39da787f0eaf64c13c15770892341c7c5f --- /dev/null +++ b/2025/Pose Priors from Language Models/9f701210-ac31-4714-b2b0-3e6ed2c2dfd9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de3eb80e05b86a407577f3c2a1925f039d77e0cf10f5d6c27749e4d71e3f9ff9 +size 2344039 diff --git a/2025/Pose Priors from Language Models/full.md b/2025/Pose Priors from Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c348f2989e8ec2da825edd3be24a11dd5647d22c --- /dev/null +++ b/2025/Pose Priors from Language Models/full.md @@ -0,0 +1,295 @@ +# Pose Priors from Language Models + +Sanjay Subramanian1 Evonne Ng1 Lea Müller1 Dan Klein1 Shiry Ginosar2,3 Trevor Darrell1 +1 University of California, Berkeley +2 Google DeepMind +3 Toyota Technological Institute at Chicago +{sanjayss,evonne_ng,mueller,klein,trevordarrell}@berkeley.edu +shiry@google.com + +![](images/7f8f6ccd34f91ef92ce6ef19db6cdba97778ee3d527121d3bc54b2bdfa3ce926.jpg) +Figure 1. Optimizing human-to-human contacts in 3D pose. Our approach leverages the semantic priors of a Large Multimodal Model (LMM) to infer meaningful information about physical contact from images. Instead of relying on human annotations or motion capture data, we extract not only descriptive insights ("... engaged in a dance or embrace ...") but also structured constraints between body parts (underlined). By incorporating these LMM-derived constraints, we refine initial 3D human pose estimates, achieving realistic and semantically consistent reconstructions of contact. This scalable approach opens up new possibilities for contact-aware pose estimation without explicit contact annotations, making it a promising alternative to traditional methods. + +# Abstract + +Language is often used to describe physical interaction, yet most 3D human pose estimation methods overlook this rich source of information. We bridge this gap by leveraging large multimodal models (LMMs) as priors for reconstructing contact poses, offering a scalable alternative to traditional methods that rely on human annotations or motion capture data. Our approach extracts contact-relevant descriptors from an LMM and translates them into tractable losses to constrain 3D human pose optimization. Despite its simplicity, our method produces compelling reconstructions for both two-person interactions and self-contact scenarios, accurately capturing the semantics of physical and social interactions. Our results demonstrate that LMMs can serve as powerful tools for contact prediction and pose estimation, offering an alternative to costly manual human annotations + +or motion capture data. Our code is publicly available at https://prosepose.github.io. + +# 1. Introduction + +Language, as a human artifact, encodes a rich set of social and physical interactions. Over centuries, our vocabulary has evolved to describe the nuances of touch, with words and phrases capturing contexts as varied as hugs, handshakes, or postures in sports and yoga. Perceiving physical contact is essential for understanding human behavior: e.g. several forms of parent-child contact are associated with affection [3, 43], and some forms of self-contact signal stress [16]. + +Since written language discusses our physical interactions at great length, can large multimodal models (LMMs) trained on images and text correctly perceive physical contact in human pose? This question has practical significance + +because scenes with contact are challenging for pose estimation methods as some body parts are frequently occluded. This particularly holds for methods solely relying on 2D keypoints which do not convey contact information. Previously proposed approaches address these issues by curating task-specific datasets via motion capture or human-annotated points of contact between body parts [11, 33, 34]. However, collecting these datasets is expensive, and existing publicly available datasets include only tens of thousands of images [10, 22, 33, 50]. If LMMs can accurately identify contact points, they could decrease the cost of curating such datasets + +In this work, we study the efficacy of LMMs as tools for contact prediction in pose estimation. Since LMMs output language rather than pose parameters, answering this question requires a way of eliciting the required information from an LMM and operationalizing it in pose estimation. We introduce a framework, called ProsePose, which prompts an LMM for formatted constraints about physical contact in the image, converts the constraints into a loss function, and optimizes this loss function (jointly with losses from other cues, such as 2D keypoints) to refine initial pose estimates. + +We use ProsePose to evaluate LMMs on both 2-person interaction datasets and a dataset of complex yoga poses. Our framework improves pose estimates compared to strong baselines that do not use contact supervision. In our extensive analysis, we show that several components are important: mitigating LMM hallucinations by aggregating predictions from several LMM samples, careful construction of the prompt and loss functions, and integrating losses from other cues. Finally, we conduct an extensive analysis of LMM predictions and their role in the optimization results. With respect to LMM failures, we find that identifying the chirality of limbs is a particular challenge for LMMs. While existing supervised methods excel by training on large amounts of supervised data with contact labels, we show that we can extract useful priors for contact prediction from pretrained LMMs without fine tuning. + +In summary, our contributions are (1) we introduce a framework for applying LMMs as contact prediction tools in pose estimation, (2) we show that our framework can improve the quality of pose estimates in 2-person and 1-person settings, and (3) we provide an analysis of the components of our framework and LMM failure modes. + +# 2. Related Work + +3D human pose reconstruction. Reconstructing 3D human poses from single images is an active area of research. Prior works have explored optimization-based approaches [13, 26, 35, 36, 39] or pure regression [2, 14, 20, 21, 25] to estimate the 3D body pose given a single image. HMR2 [12] is a recent state-of-the-art regression model in this line of work. Building on these approaches, some methods have looked into reconstructing multiple individuals jointly from a single + +image. These methods [19, 41, 51] use deep networks to reason about multiple people in a scene to directly output multi-person 3D pose predictions. BEV [42] accounts for the relative proximity of people explicitly using relative depth annotations to reason about proxemics when placing each individual in the scene (e.g. relative depth of people). However, approaches in both categories generally do not accurately capture physical contact between parts of a single person or between people [33, 34]. + +Contact inference in 3D pose reconstruction. 3D pose reconstruction is especially challenging when there is self-contact or inter-person contact. This has motivated a line of work on pose reconstruction approaches tailored for these settings. [33] focuses on predicting self contact regions for 3D pose estimation by leveraging a dataset with contact annotations to model complex poses such as crossed arms. [10] introduces the first dataset with hand-annotated ground-truth contact labels between two people. REMIPS [11] and BUDDI [34] train models on the person-to-person contact maps in this data in order to improve 3D pose estimation of multiple people from a single image. CloseInt [17] trains a physics-guided diffusion model on two-person motion capture data for this task. However, contact annotations, which are crucial for these approaches, are expensive to acquire. Our method does not require any training on such annotations. Instead, we leverage an LMM's implicit knowledge of pose to constrain pose optimization to capture both self- and person-to-person contact. + +Language priors on human pose. There exists a plethora of text to 3D human pose and motion datasets [15, 37, 38], which have enabled work focused on generating 3D motion sequences of a single person performing a general action [18, 44, 52]. This line of work has been extended to generating the motion of two people conditioned on text [29, 40]. + +PoseScript [7] is a method for generating a single person's pose from fine-grained descriptions, which uses training data from motion capture annotated with detailed text. PoseFix [8] introduces a labeled dataset for the task of modifying a pose given a fine-grained description of the desired change and trains a model on this data. PoseGPT [9] is a pose regressor that uses language as part of its training data. However, PoseGPT does not produce better pose estimates than previous state-of-the-art regressors (i.e. regressors that do not use language) and applies only to the one-person setting. [46] uses a text-only LM to improve action-conditioned human-object pose estimation. This method relies on a limited database of action-pose pairs to classify an input pose, and uses an LM to improve pose estimates based on the action retrieved from the database. + +Our work differs from previous work on language and pose in several ways. First, whereas all prior work relies on training data with pairs of language and pose, which + +![](images/bccbf40ac4ce4854f8466d257bdaeef8b3a8dd2789d155231527b821e73a939b.jpg) +Figure 2. LMM-guided Pose Estimation. (a) Method overview: ProsePose takes as input an image of one or two people in contact. We first obtain initial pose estimates for each person from a pose regressor. Then we use an LMM to generate contact constraints, each of which is a pair of body parts that should be touching. This list of contacts is converted into a loss function $\mathcal{L}_{\mathrm{LMM}}$ . We optimize the pose estimates using $\mathcal{L}_{\mathrm{LMM}}$ and other losses to produce a refined estimate of each person's pose that respects the predicted contacts. (b) Defining contact constraints: Given an image $\pmb{I}$ , we can lift each individual into corresponding 3D meshes $\pmb{V}$ . A contact constraint $\pmb{c}$ is a pair of regions $(\pmb{R}_a, \pmb{R}_b)$ in contact. The loss is defined in terms of the distance between the vertices $(\pmb{v}_a, \pmb{v}_b)$ on the mesh. + +![](images/367f6409f63bade72da68af5b3fafc47aca26d5875d5a507d826286ab6845cd3.jpg) + +is expensive to collect, our method leverages the existing knowledge in an LMM to reason about pose from a given image. Second, prior work in this area focuses on either the one-person or the two-person setting. In contrast, our work presents a single framework to reason about physical contacts within or between poses. Finally, in scenes with physical contact, we show that our method improves the pose estimates of state-of-the-art regressors. + +# 3. Guiding Pose Optimization with an LMM + +Given an image, our goal is to estimate the 3D body pose of individuals in the image while capturing the self and cross-person contact points. While we cannot trivially use natural language responses (hug, kiss) to directly optimize 3D body poses, we leverage the key insight that LMMs understand how to articulate a given pose (arms around waist, lips touching). We propose a method to structure these articulations into constraints and convert them into loss functions. + +More concretely, our framework, illustrated by Figure 2a, takes as input the image $\pmb{I}$ and the bounding boxes $B$ of the subjects of interest. In the first stage, a pose regressor takes the image and produces a rough estimate of the 3D pose $X^p$ for each individual $p$ in the image. In the second stage, an LMM takes the image and a set of instructions and generates a list of self- or inter-person contact constraints, which we then convert into a loss function (Sec. 3.2). Finally, in the third stage, we jointly optimize the generated loss function with several other pre-defined loss terms (Sec. 3.3). We refer to our framework as ProsePose. + +# 3.1. Preliminaries + +We focus our description on the two-person case to keep the exposition simple. We also demonstrate results on the one-person case, which is simply an extension of the two-person case. In particular, we apply our method to the one-person case by setting $X^0 = X^1$ . Please see Appendix § 7 for details on the differences between the two cases. + +Large Multimodal Models. An LMM is a model that takes as input an image and a text prompt and produces text output that answers the prompt based on the image. Our framework is agnostic to the architecture of the LMM. LMMs are typically trained to respond to a wide variety of instructions [6, 30]. However, LMMs are prone to hallucination [27, 28]. Handling cases of hallucination is a key challenge when using LMMs. We mitigate this issue by aggregating information across several samples from the LMM. + +Pose representation. We use a human body model [36] to represent each person $p \in \{0,1\}$ . The body model is composed of a pose parameter that defines the joint rotations $\theta \in \mathbb{R}^{d_{\theta} \times 3}$ , where $d_{\theta}$ is the number of joints, and a shape parameter $\beta \in \mathbb{R}^{d_{\beta}}$ , where $d_{\beta}$ is the dimensions of the shape parameter. We can apply a global rotation $\Phi \in \mathbb{R}^3$ and translation $t \in \mathbb{R}^3$ to place each person in the world coordinate space. The full set of parameters for each person is denoted by $\boldsymbol{X}^p = [\theta^p, \beta^p, \Phi^p, t^p]$ . For simplicity, we refer to the parameter set $(\boldsymbol{X}^0, \boldsymbol{X}^1)$ as $\boldsymbol{X}$ . + +These parameters can be plugged into a differentiable function that maps to a mesh consisting of $d_v$ vertices $\mathbf{V} \in \mathbb{R}^{d_v \times 3}$ . From the mesh, we can obtain the 3D locations of the body's joints $\mathbf{J} \in \mathbb{R}^{d_j \times 3}$ . From these joints, we can calculate the 2D keypoints $\mathbf{K}_{proj}$ by projecting the 3D joints to 2D + +using the camera intrinsics $\Pi$ predicted from [36]. + +$$ +\boldsymbol {K} _ {p r o j} = \Pi (\boldsymbol {J}) \in \mathbb {R} ^ {d _ {j} \times 2}. \tag {1} +$$ + +Vertex regions. In order to define contact constraints between body parts, we define a set of regions of vertices. Prior work on contact has partitioned the body into fine-grained regions [10]. However, since our constraints are specified by a LMM trained on natural language, the referenced body parts are often coarser in granularity. We therefore update the set of regions to reflect this language bias by combining these fine-grained regions into larger, more commonly referenced body parts such as arm, shoulder (front&back), back, and waist (front&back). Some regions, e.g. back and waist (back), overlap. Please see Appendix § 7.2 for a visualization of the coarse regions. Formally, we write $\mathbf{R} \in \mathbb{R}^{d_r \times 3}$ to denote a region with $d_r$ vertices, which is part of the full mesh $(\mathbf{R} \subset V)$ . + +Constraint definition. A contact constraint specifies which body parts from two meshes should touch. We define contact constraints as pairs of coarse regions $c = (R_a, R_b)$ between a region $R_a$ of one mesh and $R_b$ of the other mesh, as shown in Figure 2b. For instance, ("hand", "arm") indicates a hand should touch an arm. + +# 3.2. Constraint generation with a LMM + +Our key insight is to leverage a LMM to identify regions of contact between different body parts on the human body surface. As shown in Figure 2a, we prompt the LMM with an image and ask it to output a list of all region pairs that are in contact. However, we cannot simply use its output natural language descriptions to directly optimize a 3D mesh. As such, we convert these constraints into a loss function. + +LMM-based constraint generation. Given the image $I$ , we first use the bounding boxes $B$ to crop the part containing the subjects. We then use an image segmentation model to mask any extraneous individuals. While cropping and masking the image may remove information, we find the LMMs are relatively robust to missing context, and more importantly, this allows us to indicate which individuals to focus on. Given the segmented image, we ask the LMM to generate a set $C = \{c_1, \dots, c_m\}$ of all pairs of body parts that are touching, where $m$ is the total number of constraints the LMM generates for the image. + +In the prompt, we specify the full set of coarse regions to pick from. We find that LMMs fail to reliably reference the left and right limbs correctly or consistently, so the prompt instructs the LMM not to specify chirality for each limb (see Appendix 8.2 for prompt analysis). If the LMM uses "left" or "right" to reference a region, despite an instruction to not do so, we directly use the part of the region with the specified chirality rather than considering both possibilities. + +Motivated by the chain-of-thought technique, which has been shown to improve language model performance on reasoning tasks [48], we ask the LMM to write its reasoning or describe the pose before listing the constraints. For the full prompt used in each setting, please refer to Appendix § 7. + +We sample $N$ responses from the LMM. Below we describe (1) how we convert these natural language responses into $N$ sets of constraints $\{C_1, C_2, \dots, C_N\}$ and (2) how we convert each constraint set $C_j$ into a loss. + +Canonicalizing Region Names and Assigning Chirality. Given the LMM's output, we must map the mentioned region names to our fixed set of coarse regions. Since the LMM may deviate from the names in the prompt, we check for some additional names (see Appendix 7.2). We then filter out contact pairs that occur fewer than $f$ times across constraint sets, where $f$ is a hyperparameter. + +Next, we assign a chirality (left/right) to each hand/arm/foot/leg/shoulder region in cases when the LMM does not itself specify the chirality. We enumerate all possible assignments of left/right to these regions and take the one resulting in the minimum loss. In the two-person setting, we only consider assignments satisfying a condition designed for the case in which a region type occurs in multiple constraint pairs (see Appendix 7.3). + +Loss function generation. We compute a loss for each constraint by mapping the relevant regions to sets of vertices and calculating the minimum distance between vertices in the two sets. In particular, we first specify a mapping between each of the coarse region names and the fine-grained regions from [10]. We then use a mapping from fine-grained regions to SMPL-X vertices (provided by [34]) to obtain a set of vertices for each coarse region. Then for each contact pair of coarse regions $c = (R_a, R_b)$ in $C_j$ , we define $dist(c)$ as the minimum distance between the two regions: + +$$ +\operatorname {d i s t} (\boldsymbol {c}) = \min \| \boldsymbol {v} _ {a} - \boldsymbol {v} _ {b} \| _ {2} \quad \forall \boldsymbol {v} _ {a} \in \boldsymbol {R} _ {a}, \forall \boldsymbol {v} _ {b} \in \boldsymbol {R} _ {b} \tag {2} +$$ + +where $\{\pmb{v}_a, \pmb{v}_b\} \in \mathbb{R}^3$ . In practice, the number of vertices in each region can be very large. To make this computation tractable, we first take a random sample of vertices from $R_a$ and from $R_b$ before computing distances between pairs of vertices in these samples. + +Furthermore, since the ordering of the people in the LMM constraints is unknown (i.e. does $\mathbf{R}_a$ come from the mesh defined by parameter $\mathbf{X}^0$ or $\mathbf{X}^1$ ), we compute the overall loss for both possibilities and take the minimum. We use $\mathbf{c}^\top = (\mathbf{R}_b, \mathbf{R}_a)$ to denote the flipped ordering. We then sum over all constraints in the list $C_j$ : + +$$ +\operatorname {d i s t} _ {\text {s u m}} \left(\boldsymbol {C} _ {j}\right) = \min \left(\sum_ {\boldsymbol {c} \in \boldsymbol {C} _ {j}} \operatorname {d i s t} (\boldsymbol {c}), \sum_ {\boldsymbol {c} \in \boldsymbol {C} _ {j}} \operatorname {d i s t} \left(\boldsymbol {c} ^ {\top}\right)\right) \tag {3} +$$ + +Each constraint set sampled from the LMM is likely to contain noise or hallucination. To mitigate this issue, we average over all $N$ losses corresponding to each constraint set to obtain the overall LMM loss. This technique is similar to self-consistency [47], which is commonly used for code generation. Concretely, the overall LMM loss is defined as + +$$ +\mathcal {L} _ {\mathrm {L M M}} = \frac {1}{N} \sum_ {j = 1} ^ {N} \operatorname {d i s t} _ {s u m} \left(C _ {j}\right) \tag {4} +$$ + +If a constraint set $C_j$ is empty (i.e. the LMM does not suggest any contact pairs), then we set $dist_{\mathrm{sum}}(C_j) = 0$ . If there are several such constraint sets, we infer that the LMM has low confidence about the contact points (if any) in the image. Consequently, we set a threshold $t$ and if the number of empty constraint sets is at least $t$ , we gracefully backoff to the appropriate baseline optimization procedure (described in Sections 4.1 and 4.2 for each setting). We also backoff to the baseline if the LMM-based optimization diverges. + +# 3.3. Constrained pose optimization + +Drawing from previous optimization-based approaches [4, 34, 35], we employ several additional losses in the optimization. We then minimize the joint loss to obtain a refined subset of the body model parameters $\mathbf{X}^{\prime} = [\pmb{\theta}^{\prime},\pmb{\beta}^{\prime},\pmb{t}^{\prime}]$ : + +$$ +\begin{array}{l} [ \boldsymbol {\theta} ^ {\prime}, \boldsymbol {\beta} ^ {\prime}, \boldsymbol {t} ^ {\prime} ] = \arg \min (\lambda_ {\mathrm {L M M}} \mathcal {L} _ {\mathrm {L M M}} + \lambda_ {\mathrm {G M M}} \mathcal {L} _ {\mathrm {G M M}} + \lambda_ {\beta} \mathcal {L} _ {\beta} \\ + \lambda_ {\theta} \mathcal {L} _ {\theta} + \lambda_ {2 D} \mathcal {L} _ {2 D} + \lambda_ {P} \mathcal {L} _ {P}) \\ \end{array} +$$ + +Following [34], we divide the optimization into two stages. In the first stage, we optimize all three parameters. In the second stage, we optimize only $\theta$ and $t$ , keeping the shape $\beta$ fixed. We detail the remaining losses below. + +Pose and shape priors. We compute a loss $\mathcal{L}_{\mathrm{GMM}}$ based on the Gaussian Mixture pose prior of [4] and a shape loss $\mathcal{L}_{\beta} = \| \beta \|_2^2$ , which penalizes extreme deviations from the body model's mean shape. + +Initial pose loss. To ensure we do not stray too far from the initialization, we penalize large deviations from the initial pose $\mathcal{L}_{\theta} = ||\pmb{\theta}' - \pmb{\theta}||_2^2$ + +2D keypoint loss. Similar to BUDDI [34], for each person in the image, we obtain pseudo ground truth 2D keypoints and their confidence from OpenPose [5] and ViTPose [49]. Given this pseudo ground truth, we merge all the keypoints into $\pmb{K} \in \mathbb{R}^{d_j \times 2}$ , and their corresponding confidence into $\gamma \in \mathbb{R}^{d_j}$ . From the predicted $\pmb{X}'$ , we can compute the 2D projection of each 3D joint location using Equation 3.1. Then, the 2D keypoint loss is defined as: + +$$ +\mathcal {L} _ {2 D} = \sum_ {j = 1} ^ {d _ {j}} \boldsymbol {\gamma} \left(\boldsymbol {K} _ {p r o j} - \boldsymbol {K}\right) ^ {2} \tag {5} +$$ + +Interpenetration loss. To prevent parts of one mesh from being in the interior of the other, we add an interpenetration loss. Generically, given two sets of vertices $V_{0}$ and $V_{1}$ , we use winding numbers to compute the subset of $V_{0}$ that intersects $V_{1}$ , which we denote as $V_{0,1}$ . Similarly, $V_{1,0}$ is the subset of $V_{1}$ that intersects $V_{0}$ . The interpenetration loss is then defined as + +$$ +\mathcal {L} _ {P} = \sum_ {x \in \boldsymbol {V} _ {0, 1}} \min _ {v _ {1} \in \boldsymbol {V} _ {1}} \| x - v _ {1} \| _ {2} ^ {2} + \sum_ {y \in \boldsymbol {V} _ {1, 0}} \min _ {v _ {0} \in \boldsymbol {V} _ {0}} \| y - v _ {0} \| _ {2} ^ {2} \tag {6} +$$ + +For efficiency, this loss is computed on low-resolution versions of the two meshes (roughly 1000 vertices per mesh). + +# 4. Experiments + +Implementation details. Following prior work on two-person pose estimation [34], we use BEV [42] to initialize the poses since it was trained to predict both the body pose parameters and the placement of each person in the scene. However, on the single person yoga poses, we find that the pose parameter estimates of HMR2 [12] are much higher quality, so we initialize the body pose using HMR2. + +We use the SMPL-X [36] body model and (unless specified otherwise) GPT4-V [1] as the LMM with temperature $= 0.7$ when sampling from it. GPT4-V refers to the gpt-4-vision-preview model in the OpenAI API: platform.openai.com. In the OpenAI API, we use the "high" detail setting for image input. Appendix 8.2 provides results using other prompts and other LMMs (LLaVA [31], GPT-4o) and a running time analysis. Unless otherwise specified, we set $N = 20$ samples. For all of our 2-person experiments, $f = 1$ , while $f = 10$ in the 1-person setting. We set $t = 2$ for the experiment on the CHI3D dataset and $t = N$ for all other experiments. The hyperparameters and our prompts were chosen based on experiments on the validation sets. For other implementation details refer to Appendix § 7. + +Metrics. As is standard in the pose estimation literature, we report Procrustes-aligned Mean Per Joint Position Error (PA-MPJPE) in millimeters. This metric finds the best alignment between the estimated and ground-truth pose before computing the joint error. In the two-person setting, we focus on the joint PA-MPJPE, as this evaluation incorporates the relative translation and orientation of the two people. See Appendix § 8.2 for the per-person PA-MPJPE. + +We also include the percentage of correct contact points (PCC) metric introduced by [34]. This metric captures the fraction of ground-truth contact pairs that are accurately predicted. For a given radius $r$ , a pair is classified as "in contact" if the two regions are both within the specified radius. We use the set of fine-grained regions defined in [10] to compute PCC. The metric is averaged over $r \in$ + +
Hi4DFlickrCI3DCHI3D
PM↓F1↑PM↓PCC↑F1↑PM↓PCC↑F1↑
w/o contact sup.
BEV [42]144-10664.8-9671.4-
Heuristic116-6777.8-10574.1-
ProsePose93245879.91310075.823
w. contact sup.
BUDDI [34]89-6681.9-6878.0-
BUDDI+ProsePose88-6583.2-6978.8-
Coarse GT contacts
Oracle8110043861008383.8100
+ +Table 1. Two-person Results. Joint PA-MPJPE (abbreviated PM) (lower is better), Avg. PCC (higher is better), and F1 (higher is better). For FlickrCI3D, PA-MPJPE is computed using the pseudoground-truth fits. F1 measures the accuracy of coarse contacts predicted by the LMM, while the other two metrics evaluate the quality of the estimated 3D pose. Last line shows results using ground-truth contact pairs of coarse regions (heuristic is still used as the backoff method when needed). Bold indicates best method without/with contact supervision in each column. + +0, 5, 10, 15, ..., 95 mm. Since these regions are defined on the SMPL-X mesh topology, we convert the regression baselines-BEV and HMR2- from the SMPL mesh topology to SMPL-X to compute this metric. Please see the Appendix § 8.1 for more details on the regions and on the mesh conversion. + +Finally, we report the F1 of the LMM's predicted coarse region pairs. Specifically, we compute precision as the proportion of predicted pairs $(\mathbf{R}_a,\mathbf{R}_b)$ such that there is some ground-truth pair $(\mathbf{R}_a^*,\mathbf{R}_b^*)$ for which $\mathbf{R}_a$ overlaps with $\mathbf{R}_a^*$ and $\mathbf{R}_b$ overlaps with $\mathbf{R}_b^*$ . We compute recall similarly as a proportion of ground-truth pairs. For datasets that do not provide contact maps, we define the ground-truth pairs from the ground-truth meshes, using a threshold on the minimum distance between regions (0.01 and 0.02 for Hi4D and MOYO, respectively). We ignore chirality when computing these. For a pair of people we take the maximum F1 of the two possible orderings. This metric serves to evaluate the raw output of the LMMs without 3D optimization. + +# 4.1. Two-person Pose Refinement + +Datasets. We evaluate on three datasets, and our dataset processing largely follows [34]. Hi4D [50] is a motion capture dataset of pairs of people interacting. Each sequence has a subset of frames marked as contact frames, and we take every fifth contact frame. We use the images from a single camera, resulting in 241 images. Flickr Close Interactions 3D (FlickrCI3D) [10] is a collection of Flickr images of multiple people in close interaction. The dataset includes manual annotations of the contact maps between pairs of people. [34] used these contact maps to create pseudo-ground truth 3D meshes and curated a version of the test set to exclude noisy annotations, which has 1403 images. CHI3D [10] is + +
@ radius[mm]FlickrCI3D PCC↑CHI3D PCC↑
5101551015
W/o contact sup.
BEV [42]3.66.310.85.817.432.5
Heuristic14.633.949.311.128.045.3
ProsePose15.639.957.113.535.252.5
W/ contact sup.
BUDDI [34]18.544.261.815.539.056.6
BUDDI+ProsePose21.849.366.419.543.958.8
+ +Table 2. Two-person PCC. Percent of correct contact points (PCC) for three different radii $r$ in mm. Bold indicates the best score without/with contact supervision in each column. At the ground-truth contact points, our method brings the meshes closer together than the baselines. + +a motion capture dataset of pairs of people interacting. We present results on the validation set. There are 431 images, distributed across 4 cameras. The images come from 126 video sequences, each of which has a single "contact frame." + +To develop our method, we experimented on the validation sets of FlickrCI3D and Hi4D, and a sample of the training set from CHI3D. For our experiments, we can compute the PCC on FlickrCI3D and CHI3D, which have annotated ground-truth contact maps. Following [34], we exclude from evaluation images where BEV or the keypoint detectors, which are used by the baselines as well, fail to detect one of the subjects in the interaction pair. + +Baselines We compare our estimated poses to the following: + +- BEV [42] Multi-person 3D pose estimation method. Uses relative depth to reason about spatial placement of individuals in the scene. ProsePose, Heuristic, and BUDDI use BEV to initialize pose estimates. +- Heuristic A contact heuristic which includes the auxiliary losses in Section 3.3 as well as a term that minimizes the minimum distance between the two meshes. Introduced by [34]. We use their hyperparameters for this heuristic. This baseline is also used as the backoff method for Prose-Pose when the number of empty constraint sets is at least the threshold $t$ or when the optimization diverges. +- BUDDI [34] This method uses a learned diffusion prior to constrain the optimization. We stress that BUDDI requires a large amount of annotated training data on pairs of interacting bodies, which is not used in our method. + +Quantitative Results Table 1 provides quantitative results on the three datasets. Across datasets, ProsePose consistently improves over the strongest baseline, Heuristic. On the Hi4D dataset, ProsePose reduces $85\%$ of the gap in PA-MPJPE between Heuristic and the fully supervised BUDDI. On the FlickrCI3D and CHI3D datasets, ProsePose narrows the gap in the average PCC between Heuristic and BUDDI by more than one-third. (While ProsePose achieves a better PA-MPJPE than BUDDI on FlickrCI3D, for this dataset, + +![](images/6c2efec2ce40dd45c106171a96d1a9043fe73f7a0ee93222955a8926130539ad.jpg) +Figure 3. Two-person examples We show qualitative results from ProsePose, BUDDI [34], and the contact heuristic. For each example, we show GPT4-V's top 3 constraints and the number of times each constraint was predicted across all 20 samples. Our method correctly reconstructs people in a variety of interactions, and the predicted constraints generally align with each interaction type. + +we rely primarily on PCC since PA-MPJPE is computed on pseudo-ground-truth fits.) + +On CHI3D, ProsePose outperforms Heuristic but underperforms BEV in terms of PA-MPJPE. On the subset of images where we do not default to the heuristic (i.e. on images where GPT4-V predicts enough non-empty constraint sets), the PA-MPJPE for ProsePose and BEV is 86 and 87, respectively. In other words, in the cases where our method is actually used, the joint error is slightly less than that of BEV. As a result, we can attribute the worse overall error to the poorer performance of the heuristic. The backoff method (which is the heuristic) is used in 13/241 Hi4D examples, 106/1403 Flickr examples, and 224/431 CHI3D examples. + +![](images/7ed66116ef372112d24251ca8d321a11ce3dfcadfd07283e5984f0a8cf653466.jpg) +Figure 4. Single-person examples We show qualitative results from ProsePose, HMR2 [12], and HMR2-optim on complex yoga poses. Each example also shows the constraints that are predicted by the LMM at least $f = 10$ times (and are thus used to compute $\mathcal{L}_{\mathrm{LMM}}$ ) with their counts. ProsePose correctly identifies self-contact points and optimizes the poses to respect these contacts. + +Overall, our method improves over the other methods that do not use contact supervision in terms of both joint error and PCC. While not the focus of this work, Table 1 also shows that adding ProsePose, specifically $\mathcal{L}_{\mathrm{LMM}}$ , to BUDDI leads to improved PCC. The F1 scores show that LMMs' raw output is often flawed/incomplete, but our approach mitigates hallucinations in various ways (see the ablation study below). The last row in Table 1 shows the performance when the ground-truth coarse contacts (with correct left/right labels) are used in optimization. These results show the benefit of correct coarse contacts and the clear room for improvement from better LMM predictions. + +Table 2 shows the PCC for each method at various radii. The results show that ProsePose brings the meshes closer together at the correct contact points. On both the FlickrCI3D and CHI3D datasets, ProsePose outperforms the other baselines that do not use contact supervision. Next, we ablate important aspects of ProsePose. In Figure 5, we show that averaging the loss over several samples from the LMM improves performance, mitigating the effect of LMM hallucination. Finally, ablating the various losses in optimization indicates + +![](images/5f7d56f72810ad803f3e384aa147ede287596822f2f8613606b608b9a13873c7.jpg) +Figure 5. More samples improve pose estimation. On the Flickr3D validation set, taking more samples from the LMM and averaging the resulting loss functions improves joint PA-MPJPE (left) and average PCC (right). + +![](images/1e233f6ec7e030e3ae15bdffc97c715de01146a0c223a9a5e0832204716006c2.jpg) + +
PA-MPJPE↓PCC↑PCC↑@rF1↑
51015
HMR2 [12]8483.034.255.269.5-
HMR2+opt8185.247.765.574.6-
ProsePose8287.854.273.881.425
+ +Table 3. One-person Results. PA-MPJPE (lower is better) and Avg. PCC and F1 (higher is better). ProsePose captures ground-truth contacts better than the baselines, as shown by the PCC. + +that our LMM-based loss and the 2D keypoint loss have the greatest impact on joint error: Using all losses results in a PA-MPJPE of 81. Removing $\mathcal{L}_{\mathrm{LMM}} / \mathcal{L}_{\mathrm{GMM}} / \mathcal{L}_{\beta} / \mathcal{L}_{\theta} / \mathcal{L}_{2D} / \mathcal{L}_P$ results in a PA-MPJPE of 138/85/91/84/130/78. + +Qualitative Results Figure 3 shows examples of reconstructions from ProsePose, Heuristic, and BUDDI. Below each of our predictions, we list the most common constraints predicted by GPT4-V for the image. The predicted constraints correctly capture the semantics of each interaction. For instance, in tango, one person's arm should touch the other's back. In a rugby tackle, a player's arms are usually wrapped around the other player. Using these constraints, ProsePose correctly reconstructs a variety of interactions, such as tackling, dancing, and holding hands. In contrast, Heuristic struggles to accurately position individuals and/or predict limb placements, often resulting in awkward distances. + +# 4.2. One-person pose refinement + +Datasets Next, we evaluate ProsePose on a single-person setting. For this setting, we evaluate on MOYO [45], a motion capture dataset with videos of a single person performing various yoga poses. In total, our test set is composed of 76 examples from a single camera angle (side view). See Appendix 7.6 for further dataset details. Since this dataset does not have annotated region contact pairs, we compute the peso-ground-truth contact maps using the Euclidean and geodesic distance following [33] and report PCC for the subset of 67 examples where the ground-truth has self-contact. + +Baselines We compare against the following baselines: + +- HMR2 [12] State-of-the-art pose regression method. We use HMR2 to initialize our pose estimates for optimization. +- HMR2+opt Optimization procedure that is identical to ours without $\mathcal{L}_{\mathrm{LMM}}$ . It is the default method when the number of empty constraint sets is at least the threshold $t$ . + +Both the quantitative and qualitative results echo the trends discussed in the 2-person setting. Table 3 provides the quantitative results. The PCC metrics show that our LMM loss improves the predicted self-contact in complex yoga poses relative to the two baselines. The backoff method is used in 43/76 examples. Figure 4 provides a qualitative comparison of poses predicted by ProsePose versus the two baselines. Below each of our predictions, we list the corresponding constraints predicted by GPT4-V. In each case, the predicted constraint captures the correct self-contact, which is reflected in the final pose estimates. Using the semantically guided loss, ProsePose effectively refines the pose to ensure proper contact between hand-foot or hand-hand, an important detail consistently overlooked by the baselines. + +# 4.3. Limitations + +While ProsePose consistently improves contact across settings and datasets, it has limitations which are related to failures of LMMs. First, as shown in Table 5 (in Appendix), prompting the LMM for left/right labels sometimes leads to worse results, suggesting that LMMs struggle with disambiguating chirality. Improving this approach depends in large part on correctly identifying limbs as left/right. Another limitation is the use of coarse regions. Future work could improve by eliciting more fine-grained constraints from an LMM. Finally, LMM accuracy varies moderately across camera angles (quantified in Appendix 8.2). In Appendix § 8.3, we provide examples of LMM failures. + +# 5. Conclusion + +We present ProsePose, a framework for refining 3D pose estimates to capture touch accurately using the implicit semantic knowledge of poses in LMMs. Our key novelty is that we generate structured pose descriptions from LMMs and convert them into loss functions used to optimize the pose. Our experiments show that in both one-person and two-person settings, ProsePose improves over previous baselines that do not use contact supervision. These results suggest that LMMs may be useful in creating larger datasets with contact annotations, which are otherwise expensive but are crucial for training state-of-the-art priors for pose estimation in situations with physical contact. More broadly, this work provides evidence that LMMs are promising tools for 3D pose estimation, which likely has implications beyond touch. + +# 6. Acknowledgements + +SS, EN, and TD were supported in part by the NSF, DoD, and/or the Berkeley Artificial Intelligence Research (BAIR) industrial alliance program. + +# References + +[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mo Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Benjamin Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Sim'on Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Raphael Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heating, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Hendrik Kirchner, Jamie Ryan Kiros, Matthew Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Adeola Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney Christine McLeavey Paul McMillan Jake McNeil David Medina Alok Mehta Jacob Menick Luke Metz Andrey Mishchenko Pamela Mishkin Vinnie Monaco Evan Morikawa Daniel P. Mossing Tong Mu Mira Murati Oleg Murk David M'ely Ashvin Nair Reichiro Nakano Rajeev Nayak Arvind Neelakantan Richard Ngo Hyeonwoo Noh Ouyang Long Cullen O'Keefe Jakub W. Pachocki Alex Paino Joe Palermo Ashley Pantuliano Giambattista Parascandolo Joel Parish Emy Parparita Alexandre Passos Mikhail Pavlov Andrew Peng Adam Perelman Filipe + +de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario D. Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin D. Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas A. Tezak, Madeleine Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cer'on Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao, Tao Xu, Sarah Yoo, Kevin Yu, Qiming Yuan, Wojciech Zaremba, Rowan Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. Gpt-4 technical report. 2023. 5 + +[2] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3395-3404, 2019. 2 +[3] Ana Aznar and Harriet R Tenenbaum. Parent-child positive touch: Gender, age, and task differences. Journal of nonverbal behavior, 40:317-333, 2016. 1 +[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 561-578. Springer, 2016. 5 +[5] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 5 +[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 3 +[7] Delmas, Ginger and Weinzaepfel, Philippe and Lucas, Thomas and Moreno-Noguer, Francesc and Rogez, Grégory. PoseScript: 3D Human Poses from Natural Language. In ECCV, 2022. 2 +[8] Delmas, Ginger and Weinzaepfel, Philippe and Moreno-Noguer, Francesc and Rogez, Grégory. PoseFix: Correcting 3D Human Poses with Natural Language. In ICCV, 2023. 2 + +[9] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Posegpt: Chatting about 3d human pose. ArXiv, abs/2311.18836, 2023. 2 +[10] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7214-7223, 2020. 2, 4, 5, 6 +[11] Mihai Fieraru, Mihai Zanfir, Teodor Szente, Eduard Bazavan, Vlad Olaru, and Cristian Sminchisescu. Remips: Physically consistent 3d reconstruction of multiple interacting people under weak supervision. Advances in Neural Information Processing Systems, 34:19385-19397, 2021. 2 +[12] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2, 5, 7, 8 +[13] Peng Guan, Alexander Weiss, Alexandru O Balan, and Michael J Black. Estimating human shape and pose from a single image. In 2009 IEEE 12th International Conference on Computer Vision, pages 1381-1388. IEEE, 2009. 2 +[14] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10884-10894, 2019. 2 +[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, 2022. 2 +[16] Jinni A Harrigan. Self-touching as an indicator of underlying affect and language processes. Social science & medicine, 20 (11):1161-1168, 1985. 1 +[17] Buzhen Huang, Chen Li, Chongyang Xu, Liang Pan, Yanggang Wang, and Gim Hee Lee. Closely interactive human reconstruction with proxemics and physics-guided adaption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1011-1021, 2024. 2 +[18] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2 +[19] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2020. 2 +[20] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human model fitting towards inthe-wild 3d human pose estimation. In 2021 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2021. 2 +[21] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2 + +[22] Rawal Khirodkar, Jyun-Ting Song, Jinkun Cao, Zhengyi Luo, and Kris M. Kitani. Harmony4d: A video dataset for in-the-wild close human interactions. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2 +[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 4 +[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3 +[25] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2252-2261, 2019. 2 +[26] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017. 2 +[27] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv preprint arXiv:2311.16922, 2023. 3 +[28] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3 +[29] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023. 2 +[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3, 6 +[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.5,6 +[32] Yujie Lu, Dongfu Jiang, Wenhu Chen, William Wang, Yejin Choi, and Bill Yuchen Lin. Wildvision arena: Benchmarking multimodal llms in the wild, 2024. 8 +[33] Lea Muller, Ahmed AA Osman, Siyu Tang, Chun-Hao P Huang, and Michael J Black. On self-contact and human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9990-9999, 2021. 2, 8 +[34] Lea Müller, Vickie Ye, Georgios Pavlakos, Michael Black, and Angjoo Kanazawa. Generative proxemics: A prior for 3d social interaction from images. arXiv preprint arXiv:2306.09337, 2023. 2, 4, 5, 6, 7, 9 +[35] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2, 5 +[36] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and + +Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3, 4, 5 +[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. *Big Data*, 4(4):236–252, 2016. 2 +[38] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 722-731, 2021. 2 +[39] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2 +[40] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2 +[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11179-11188, 2021. 2 +[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13243-13252, 2022. 2, 5, 6 +[43] Mika S Takeuchi, Hitoshi Miyaoka, Atsuko Tomoda, Masao Suzuki, Qingbo Liu, and Toshinori Kitamura. The effect of interpersonal touch during childhood on adult attachment and depression: A neglected area of family and developmental psychology? Journal of Child and Family Studies, 19:109-117, 2010. 1 +[44] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2 +[45] Shashank Tripathi, Lea Müller, Chun-Hao P. Huang, Taheri Omid, Michael J. Black, and Dimitrios Tzionas. 3D human pose estimation via intuitive physics. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 4713-4725, 2023. 8 +[46] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 2022 International Conference on 3D Vision (3DV), pages 353-362. IEEE, 2022. 2, 5 +[47] X Wang, J Wei, D Schuurmans, Q Le, E Chi, S Narang, A Chowdhery, and D Zhou. Self-consistency improves chain of thought reasoning in language models. arxiv. Preprint posted online March, 21:10-48550, 2022. 5 +[48] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Huai hsin Chi, F. Xia, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. ArXiv, abs/2201.11903, 2022. 4 + +[49] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In Advances in Neural Information Processing Systems, 2022. 5 +[50] Yifei Yin, Chen Guo, Manuel Kaufmann, Juan José Zárate, Jie Song, and Otmar Hilliges. Hi4d: 4d instance segmentation of close human interaction. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17016-17027, 2023. 2, 6 +[51] Andrei Zanfir, Elisabeta Marinhoiu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2148-2157, 2018. 2 +[52] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 \ No newline at end of file diff --git a/2025/Pose Priors from Language Models/images.zip b/2025/Pose Priors from Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a6409503291a8b1f796cca5f50ea528e2eb851c9 --- /dev/null +++ b/2025/Pose Priors from Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9e1da4c9b8f6a45a92795dba6dde0b69c5c413ddd261ee148035d0c2e29dd69 +size 407009 diff --git a/2025/Pose Priors from Language Models/layout.json b/2025/Pose Priors from Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..434aa039c2437b6fe34cc225fb5f75609f44c81e --- /dev/null +++ b/2025/Pose Priors from Language Models/layout.json @@ -0,0 +1,8242 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 196, + 103, + 415, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 103, + 415, + 121 + ], + "spans": [ + { + "bbox": [ + 196, + 103, + 415, + 121 + ], + "type": "text", + "content": "Pose Priors from Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 563, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 563, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 563, + 228 + ], + "type": "text", + "content": "Sanjay Subramanian1 Evonne Ng1 Lea Müller1 Dan Klein1 Shiry Ginosar2,3 Trevor Darrell1 \n1 University of California, Berkeley \n2 Google DeepMind \n3 Toyota Technological Institute at Chicago \n{sanjayss,evonne_ng,mueller,klein,trevordarrell}@berkeley.edu \nshiry@google.com" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 243, + 553, + 407 + ], + "blocks": [ + { + "bbox": [ + 63, + 243, + 553, + 407 + ], + "lines": [ + { + "bbox": [ + 63, + 243, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 63, + 243, + 553, + 407 + ], + "type": "image", + "image_path": "7f8f6ccd34f91ef92ce6ef19db6cdba97778ee3d527121d3bc54b2bdfa3ce926.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "type": "text", + "content": "Figure 1. Optimizing human-to-human contacts in 3D pose. Our approach leverages the semantic priors of a Large Multimodal Model (LMM) to infer meaningful information about physical contact from images. Instead of relying on human annotations or motion capture data, we extract not only descriptive insights (\"... engaged in a dance or embrace ...\") but also structured constraints between body parts (underlined). By incorporating these LMM-derived constraints, we refine initial 3D human pose estimates, achieving realistic and semantically consistent reconstructions of contact. This scalable approach opens up new possibilities for contact-aware pose estimation without explicit contact annotations, making it a promising alternative to traditional methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 504, + 200, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 504, + 200, + 516 + ], + "spans": [ + { + "bbox": [ + 152, + 504, + 200, + 516 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 535, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 535, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 535, + 297, + 715 + ], + "type": "text", + "content": "Language is often used to describe physical interaction, yet most 3D human pose estimation methods overlook this rich source of information. We bridge this gap by leveraging large multimodal models (LMMs) as priors for reconstructing contact poses, offering a scalable alternative to traditional methods that rely on human annotations or motion capture data. Our approach extracts contact-relevant descriptors from an LMM and translates them into tractable losses to constrain 3D human pose optimization. Despite its simplicity, our method produces compelling reconstructions for both two-person interactions and self-contact scenarios, accurately capturing the semantics of physical and social interactions. Our results demonstrate that LMMs can serve as powerful tools for contact prediction and pose estimation, offering an alternative to costly manual human annotations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 506, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 506, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 314, + 506, + 555, + 529 + ], + "type": "text", + "content": "or motion capture data. Our code is publicly available at https://prosepose.github.io." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 551, + 395, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 551, + 395, + 563 + ], + "spans": [ + { + "bbox": [ + 314, + 551, + 395, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 571, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 571, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 312, + 571, + 555, + 666 + ], + "type": "text", + "content": "Language, as a human artifact, encodes a rich set of social and physical interactions. Over centuries, our vocabulary has evolved to describe the nuances of touch, with words and phrases capturing contexts as varied as hugs, handshakes, or postures in sports and yoga. Perceiving physical contact is essential for understanding human behavior: e.g. several forms of parent-child contact are associated with affection [3, 43], and some forms of self-contact signal stress [16]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Since written language discusses our physical interactions at great length, can large multimodal models (LMMs) trained on images and text correctly perceive physical contact in human pose? This question has practical significance" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7125" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "because scenes with contact are challenging for pose estimation methods as some body parts are frequently occluded. This particularly holds for methods solely relying on 2D keypoints which do not convey contact information. Previously proposed approaches address these issues by curating task-specific datasets via motion capture or human-annotated points of contact between body parts [11, 33, 34]. However, collecting these datasets is expensive, and existing publicly available datasets include only tens of thousands of images [10, 22, 33, 50]. If LMMs can accurately identify contact points, they could decrease the cost of curating such datasets" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 294, + 324 + ], + "type": "text", + "content": "In this work, we study the efficacy of LMMs as tools for contact prediction in pose estimation. Since LMMs output language rather than pose parameters, answering this question requires a way of eliciting the required information from an LMM and operationalizing it in pose estimation. We introduce a framework, called ProsePose, which prompts an LMM for formatted constraints about physical contact in the image, converts the constraints into a loss function, and optimizes this loss function (jointly with losses from other cues, such as 2D keypoints) to refine initial pose estimates." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 324, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 515 + ], + "type": "text", + "content": "We use ProsePose to evaluate LMMs on both 2-person interaction datasets and a dataset of complex yoga poses. Our framework improves pose estimates compared to strong baselines that do not use contact supervision. In our extensive analysis, we show that several components are important: mitigating LMM hallucinations by aggregating predictions from several LMM samples, careful construction of the prompt and loss functions, and integrating losses from other cues. Finally, we conduct an extensive analysis of LMM predictions and their role in the optimization results. With respect to LMM failures, we find that identifying the chirality of limbs is a particular challenge for LMMs. While existing supervised methods excel by training on large amounts of supervised data with contact labels, we show that we can extract useful priors for contact prediction from pretrained LMMs without fine tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 516, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 587 + ], + "type": "text", + "content": "In summary, our contributions are (1) we introduce a framework for applying LMMs as contact prediction tools in pose estimation, (2) we show that our framework can improve the quality of pose estimates in 2-person and 1-person settings, and (3) we provide an analysis of the components of our framework and LMM failure modes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "3D human pose reconstruction. Reconstructing 3D human poses from single images is an active area of research. Prior works have explored optimization-based approaches [13, 26, 35, 36, 39] or pure regression [2, 14, 20, 21, 25] to estimate the 3D body pose given a single image. HMR2 [12] is a recent state-of-the-art regression model in this line of work. Building on these approaches, some methods have looked into reconstructing multiple individuals jointly from a single" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "image. These methods [19, 41, 51] use deep networks to reason about multiple people in a scene to directly output multi-person 3D pose predictions. BEV [42] accounts for the relative proximity of people explicitly using relative depth annotations to reason about proxemics when placing each individual in the scene (e.g. relative depth of people). However, approaches in both categories generally do not accurately capture physical contact between parts of a single person or between people [33, 34]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 188, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 416 + ], + "type": "text", + "content": "Contact inference in 3D pose reconstruction. 3D pose reconstruction is especially challenging when there is self-contact or inter-person contact. This has motivated a line of work on pose reconstruction approaches tailored for these settings. [33] focuses on predicting self contact regions for 3D pose estimation by leveraging a dataset with contact annotations to model complex poses such as crossed arms. [10] introduces the first dataset with hand-annotated ground-truth contact labels between two people. REMIPS [11] and BUDDI [34] train models on the person-to-person contact maps in this data in order to improve 3D pose estimation of multiple people from a single image. CloseInt [17] trains a physics-guided diffusion model on two-person motion capture data for this task. However, contact annotations, which are crucial for these approaches, are expensive to acquire. Our method does not require any training on such annotations. Instead, we leverage an LMM's implicit knowledge of pose to constrain pose optimization to capture both self- and person-to-person contact." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 425, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 496 + ], + "type": "text", + "content": "Language priors on human pose. There exists a plethora of text to 3D human pose and motion datasets [15, 37, 38], which have enabled work focused on generating 3D motion sequences of a single person performing a general action [18, 44, 52]. This line of work has been extended to generating the motion of two people conditioned on text [29, 40]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 675 + ], + "type": "text", + "content": "PoseScript [7] is a method for generating a single person's pose from fine-grained descriptions, which uses training data from motion capture annotated with detailed text. PoseFix [8] introduces a labeled dataset for the task of modifying a pose given a fine-grained description of the desired change and trains a model on this data. PoseGPT [9] is a pose regressor that uses language as part of its training data. However, PoseGPT does not produce better pose estimates than previous state-of-the-art regressors (i.e. regressors that do not use language) and applies only to the one-person setting. [46] uses a text-only LM to improve action-conditioned human-object pose estimation. This method relies on a limited database of action-pose pairs to classify an input pose, and uses an LM to improve pose estimates based on the action retrieved from the database." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Our work differs from previous work on language and pose in several ways. First, whereas all prior work relies on training data with pairs of language and pose, which" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7126" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 358, + 223 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 358, + 223 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 358, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 358, + 223 + ], + "type": "image", + "image_path": "bccbf40ac4ce4854f8466d257bdaeef8b3a8dd2789d155231527b821e73a939b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "lines": [ + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "text", + "content": "Figure 2. LMM-guided Pose Estimation. (a) Method overview: ProsePose takes as input an image of one or two people in contact. We first obtain initial pose estimates for each person from a pose regressor. Then we use an LMM to generate contact constraints, each of which is a pair of body parts that should be touching. This list of contacts is converted into a loss function " + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LMM}}" + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "text", + "content": ". We optimize the pose estimates using " + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LMM}}" + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "text", + "content": " and other losses to produce a refined estimate of each person's pose that respects the predicted contacts. (b) Defining contact constraints: Given an image " + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "text", + "content": ", we can lift each individual into corresponding 3D meshes " + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "inline_equation", + "content": "\\pmb{V}" + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "text", + "content": ". A contact constraint " + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "inline_equation", + "content": "\\pmb{c}" + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "text", + "content": " is a pair of regions " + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "inline_equation", + "content": "(\\pmb{R}_a, \\pmb{R}_b)" + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "text", + "content": " in contact. The loss is defined in terms of the distance between the vertices " + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "inline_equation", + "content": "(\\pmb{v}_a, \\pmb{v}_b)" + }, + { + "bbox": [ + 54, + 228, + 557, + 298 + ], + "type": "text", + "content": " on the mesh." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 402, + 103, + 556, + 223 + ], + "blocks": [ + { + "bbox": [ + 402, + 103, + 556, + 223 + ], + "lines": [ + { + "bbox": [ + 402, + 103, + 556, + 223 + ], + "spans": [ + { + "bbox": [ + 402, + 103, + 556, + 223 + ], + "type": "image", + "image_path": "367f6409f63bade72da68af5b3fafc47aca26d5875d5a507d826286ab6845cd3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 316, + 298, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 298, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 298, + 413 + ], + "type": "text", + "content": "is expensive to collect, our method leverages the existing knowledge in an LMM to reason about pose from a given image. Second, prior work in this area focuses on either the one-person or the two-person setting. In contrast, our work presents a single framework to reason about physical contacts within or between poses. Finally, in scenes with physical contact, we show that our method improves the pose estimates of state-of-the-art regressors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 285, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 285, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 285, + 464 + ], + "type": "text", + "content": "3. Guiding Pose Optimization with an LMM" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 477, + 298, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 477, + 298, + 574 + ], + "spans": [ + { + "bbox": [ + 54, + 477, + 298, + 574 + ], + "type": "text", + "content": "Given an image, our goal is to estimate the 3D body pose of individuals in the image while capturing the self and cross-person contact points. While we cannot trivially use natural language responses (hug, kiss) to directly optimize 3D body poses, we leverage the key insight that LMMs understand how to articulate a given pose (arms around waist, lips touching). We propose a method to structure these articulations into constraints and convert them into loss functions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "content": "More concretely, our framework, illustrated by Figure 2a, takes as input the image " + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "content": " and the bounding boxes " + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "content": " of the subjects of interest. In the first stage, a pose regressor takes the image and produces a rough estimate of the 3D pose " + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "content": " for each individual " + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "content": " in the image. In the second stage, an LMM takes the image and a set of instructions and generates a list of self- or inter-person contact constraints, which we then convert into a loss function (Sec. 3.2). Finally, in the third stage, we jointly optimize the generated loss function with several other pre-defined loss terms (Sec. 3.3). We refer to our framework as ProsePose." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 400, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 400, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 400, + 328 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 335, + 556, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 556, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 556, + 407 + ], + "type": "text", + "content": "We focus our description on the two-person case to keep the exposition simple. We also demonstrate results on the one-person case, which is simply an extension of the two-person case. In particular, we apply our method to the one-person case by setting " + }, + { + "bbox": [ + 313, + 335, + 556, + 407 + ], + "type": "inline_equation", + "content": "X^0 = X^1" + }, + { + "bbox": [ + 313, + 335, + 556, + 407 + ], + "type": "text", + "content": ". Please see Appendix § 7 for details on the differences between the two cases." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 416, + 556, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 556, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 556, + 525 + ], + "type": "text", + "content": "Large Multimodal Models. An LMM is a model that takes as input an image and a text prompt and produces text output that answers the prompt based on the image. Our framework is agnostic to the architecture of the LMM. LMMs are typically trained to respond to a wide variety of instructions [6, 30]. However, LMMs are prone to hallucination [27, 28]. Handling cases of hallucination is a key challenge when using LMMs. We mitigate this issue by aggregating information across several samples from the LMM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": "Pose representation. We use a human body model [36] to represent each person " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "p \\in \\{0,1\\}" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": ". The body model is composed of a pose parameter that defines the joint rotations " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{d_{\\theta} \\times 3}" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "d_{\\theta}" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": " is the number of joints, and a shape parameter " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{d_{\\beta}}" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "d_{\\beta}" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": " is the dimensions of the shape parameter. We can apply a global rotation " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\Phi \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": " to place each person in the world coordinate space. The full set of parameters for each person is denoted by " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\boldsymbol{X}^p = [\\theta^p, \\beta^p, \\Phi^p, t^p]" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": ". For simplicity, we refer to the parameter set " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{X}^0, \\boldsymbol{X}^1)" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\boldsymbol{X}" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "These parameters can be plugged into a differentiable function that maps to a mesh consisting of " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " vertices " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{d_v \\times 3}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": ". From the mesh, we can obtain the 3D locations of the body's joints " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{J} \\in \\mathbb{R}^{d_j \\times 3}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": ". From these joints, we can calculate the 2D keypoints " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{proj}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " by projecting the 3D joints to 2D" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7127" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 258, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 258, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 258, + 84 + ], + "type": "text", + "content": "using the camera intrinsics " + }, + { + "bbox": [ + 55, + 72, + 258, + 84 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 55, + 72, + 258, + 84 + ], + "type": "text", + "content": " predicted from [36]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 122, + 92, + 296, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 92, + 296, + 107 + ], + "spans": [ + { + "bbox": [ + 122, + 92, + 296, + 107 + ], + "type": "interline_equation", + "content": "\\boldsymbol {K} _ {p r o j} = \\Pi (\\boldsymbol {J}) \\in \\mathbb {R} ^ {d _ {j} \\times 2}. \\tag {1}", + "image_path": "0c3e1bcbca3d88b71d6d1e45d616275353e68c1f3070465669574de004fe0336.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 125, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 125, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 125, + 296, + 293 + ], + "type": "text", + "content": "Vertex regions. In order to define contact constraints between body parts, we define a set of regions of vertices. Prior work on contact has partitioned the body into fine-grained regions [10]. However, since our constraints are specified by a LMM trained on natural language, the referenced body parts are often coarser in granularity. We therefore update the set of regions to reflect this language bias by combining these fine-grained regions into larger, more commonly referenced body parts such as arm, shoulder (front&back), back, and waist (front&back). Some regions, e.g. back and waist (back), overlap. Please see Appendix § 7.2 for a visualization of the coarse regions. Formally, we write " + }, + { + "bbox": [ + 55, + 125, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{d_r \\times 3}" + }, + { + "bbox": [ + 55, + 125, + 296, + 293 + ], + "type": "text", + "content": " to denote a region with " + }, + { + "bbox": [ + 55, + 125, + 296, + 293 + ], + "type": "inline_equation", + "content": "d_r" + }, + { + "bbox": [ + 55, + 125, + 296, + 293 + ], + "type": "text", + "content": " vertices, which is part of the full mesh " + }, + { + "bbox": [ + 55, + 125, + 296, + 293 + ], + "type": "inline_equation", + "content": "(\\mathbf{R} \\subset V)" + }, + { + "bbox": [ + 55, + 125, + 296, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 299, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 370 + ], + "type": "text", + "content": "Constraint definition. A contact constraint specifies which body parts from two meshes should touch. We define contact constraints as pairs of coarse regions " + }, + { + "bbox": [ + 55, + 299, + 296, + 370 + ], + "type": "inline_equation", + "content": "c = (R_a, R_b)" + }, + { + "bbox": [ + 55, + 299, + 296, + 370 + ], + "type": "text", + "content": " between a region " + }, + { + "bbox": [ + 55, + 299, + 296, + 370 + ], + "type": "inline_equation", + "content": "R_a" + }, + { + "bbox": [ + 55, + 299, + 296, + 370 + ], + "type": "text", + "content": " of one mesh and " + }, + { + "bbox": [ + 55, + 299, + 296, + 370 + ], + "type": "inline_equation", + "content": "R_b" + }, + { + "bbox": [ + 55, + 299, + 296, + 370 + ], + "type": "text", + "content": " of the other mesh, as shown in Figure 2b. For instance, (\"hand\", \"arm\") indicates a hand should touch an arm." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 244, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 244, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 244, + 390 + ], + "type": "text", + "content": "3.2. Constraint generation with a LMM" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 479 + ], + "type": "text", + "content": "Our key insight is to leverage a LMM to identify regions of contact between different body parts on the human body surface. As shown in Figure 2a, we prompt the LMM with an image and ask it to output a list of all region pairs that are in contact. However, we cannot simply use its output natural language descriptions to directly optimize a 3D mesh. As such, we convert these constraints into a loss function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "text", + "content": "LMM-based constraint generation. Given the image " + }, + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "text", + "content": ", we first use the bounding boxes " + }, + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "text", + "content": " to crop the part containing the subjects. We then use an image segmentation model to mask any extraneous individuals. While cropping and masking the image may remove information, we find the LMMs are relatively robust to missing context, and more importantly, this allows us to indicate which individuals to focus on. Given the segmented image, we ask the LMM to generate a set " + }, + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "inline_equation", + "content": "C = \\{c_1, \\dots, c_m\\}" + }, + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "text", + "content": " of all pairs of body parts that are touching, where " + }, + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 486, + 296, + 617 + ], + "type": "text", + "content": " is the total number of constraints the LMM generates for the image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "In the prompt, we specify the full set of coarse regions to pick from. We find that LMMs fail to reliably reference the left and right limbs correctly or consistently, so the prompt instructs the LMM not to specify chirality for each limb (see Appendix 8.2 for prompt analysis). If the LMM uses \"left\" or \"right\" to reference a region, despite an instruction to not do so, we directly use the part of the region with the specified chirality rather than considering both possibilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "Motivated by the chain-of-thought technique, which has been shown to improve language model performance on reasoning tasks [48], we ask the LMM to write its reasoning or describe the pose before listing the constraints. For the full prompt used in each setting, please refer to Appendix § 7." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": "We sample " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": " responses from the LMM. Below we describe (1) how we convert these natural language responses into " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": " sets of constraints " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "\\{C_1, C_2, \\dots, C_N\\}" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": " and (2) how we convert each constraint set " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": " into a loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 188, + 555, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 272 + ], + "type": "text", + "content": "Canonicalizing Region Names and Assigning Chirality. Given the LMM's output, we must map the mentioned region names to our fixed set of coarse regions. Since the LMM may deviate from the names in the prompt, we check for some additional names (see Appendix 7.2). We then filter out contact pairs that occur fewer than " + }, + { + "bbox": [ + 313, + 188, + 555, + 272 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 188, + 555, + 272 + ], + "type": "text", + "content": " times across constraint sets, where " + }, + { + "bbox": [ + 313, + 188, + 555, + 272 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 188, + 555, + 272 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 273, + 556, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 556, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 556, + 369 + ], + "type": "text", + "content": "Next, we assign a chirality (left/right) to each hand/arm/foot/leg/shoulder region in cases when the LMM does not itself specify the chirality. We enumerate all possible assignments of left/right to these regions and take the one resulting in the minimum loss. In the two-person setting, we only consider assignments satisfying a condition designed for the case in which a region type occurs in multiple constraint pairs (see Appendix 7.3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "text", + "content": "Loss function generation. We compute a loss for each constraint by mapping the relevant regions to sets of vertices and calculating the minimum distance between vertices in the two sets. In particular, we first specify a mapping between each of the coarse region names and the fine-grained regions from [10]. We then use a mapping from fine-grained regions to SMPL-X vertices (provided by [34]) to obtain a set of vertices for each coarse region. Then for each contact pair of coarse regions " + }, + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "inline_equation", + "content": "c = (R_a, R_b)" + }, + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "inline_equation", + "content": "dist(c)" + }, + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "text", + "content": " as the minimum distance between the two regions:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 506, + 555, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 506, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 324, + 506, + 555, + 520 + ], + "type": "interline_equation", + "content": "\\operatorname {d i s t} (\\boldsymbol {c}) = \\min \\| \\boldsymbol {v} _ {a} - \\boldsymbol {v} _ {b} \\| _ {2} \\quad \\forall \\boldsymbol {v} _ {a} \\in \\boldsymbol {R} _ {a}, \\forall \\boldsymbol {v} _ {b} \\in \\boldsymbol {R} _ {b} \\tag {2}", + "image_path": "381c30710143397d3654cc21ae6f80b72fdec2ce2fed255511d640e2bca2c6d0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\{\\pmb{v}_a, \\pmb{v}_b\\} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "type": "text", + "content": ". In practice, the number of vertices in each region can be very large. To make this computation tractable, we first take a random sample of vertices from " + }, + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "type": "inline_equation", + "content": "R_a" + }, + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "type": "text", + "content": " and from " + }, + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "type": "inline_equation", + "content": "R_b" + }, + { + "bbox": [ + 313, + 530, + 555, + 590 + ], + "type": "text", + "content": " before computing distances between pairs of vertices in these samples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "text", + "content": "Furthermore, since the ordering of the people in the LMM constraints is unknown (i.e. does " + }, + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_a" + }, + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "text", + "content": " come from the mesh defined by parameter " + }, + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^0" + }, + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^1" + }, + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "text", + "content": "), we compute the overall loss for both possibilities and take the minimum. We use " + }, + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^\\top = (\\mathbf{R}_b, \\mathbf{R}_a)" + }, + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "text", + "content": " to denote the flipped ordering. We then sum over all constraints in the list " + }, + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 313, + 590, + 554, + 664 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 673, + 555, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 673, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 673, + 555, + 712 + ], + "type": "interline_equation", + "content": "\\operatorname {d i s t} _ {\\text {s u m}} \\left(\\boldsymbol {C} _ {j}\\right) = \\min \\left(\\sum_ {\\boldsymbol {c} \\in \\boldsymbol {C} _ {j}} \\operatorname {d i s t} (\\boldsymbol {c}), \\sum_ {\\boldsymbol {c} \\in \\boldsymbol {C} _ {j}} \\operatorname {d i s t} \\left(\\boldsymbol {c} ^ {\\top}\\right)\\right) \\tag {3}", + "image_path": "bda1d7a10b9731621770f97d80a64ef4a15b16817ac38c6d03555491226e41ef.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7128" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "Each constraint set sampled from the LMM is likely to contain noise or hallucination. To mitigate this issue, we average over all " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " losses corresponding to each constraint set to obtain the overall LMM loss. This technique is similar to self-consistency [47], which is commonly used for code generation. Concretely, the overall LMM loss is defined as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 160, + 296, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 160, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 115, + 160, + 296, + 196 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L M M}} = \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\operatorname {d i s t} _ {s u m} \\left(C _ {j}\\right) \\tag {4}", + "image_path": "f5727d2a18ed9e2620f963dd74c6f6a1a55ebfa7d79181cd4edf486424381154.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 201, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 310 + ], + "type": "text", + "content": "If a constraint set " + }, + { + "bbox": [ + 55, + 201, + 296, + 310 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 55, + 201, + 296, + 310 + ], + "type": "text", + "content": " is empty (i.e. the LMM does not suggest any contact pairs), then we set " + }, + { + "bbox": [ + 55, + 201, + 296, + 310 + ], + "type": "inline_equation", + "content": "dist_{\\mathrm{sum}}(C_j) = 0" + }, + { + "bbox": [ + 55, + 201, + 296, + 310 + ], + "type": "text", + "content": ". If there are several such constraint sets, we infer that the LMM has low confidence about the contact points (if any) in the image. Consequently, we set a threshold " + }, + { + "bbox": [ + 55, + 201, + 296, + 310 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 201, + 296, + 310 + ], + "type": "text", + "content": " and if the number of empty constraint sets is at least " + }, + { + "bbox": [ + 55, + 201, + 296, + 310 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 201, + 296, + 310 + ], + "type": "text", + "content": ", we gracefully backoff to the appropriate baseline optimization procedure (described in Sections 4.1 and 4.2 for each setting). We also backoff to the baseline if the LMM-based optimization diverges." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 222, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 222, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 222, + 330 + ], + "type": "text", + "content": "3.3. Constrained pose optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 334, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 384 + ], + "type": "text", + "content": "Drawing from previous optimization-based approaches [4, 34, 35], we employ several additional losses in the optimization. We then minimize the joint loss to obtain a refined subset of the body model parameters " + }, + { + "bbox": [ + 55, + 334, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\prime} = [\\pmb{\\theta}^{\\prime},\\pmb{\\beta}^{\\prime},\\pmb{t}^{\\prime}]" + }, + { + "bbox": [ + 55, + 334, + 296, + 384 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 389, + 293, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 389, + 293, + 418 + ], + "spans": [ + { + "bbox": [ + 58, + 389, + 293, + 418 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} [ \\boldsymbol {\\theta} ^ {\\prime}, \\boldsymbol {\\beta} ^ {\\prime}, \\boldsymbol {t} ^ {\\prime} ] = \\arg \\min (\\lambda_ {\\mathrm {L M M}} \\mathcal {L} _ {\\mathrm {L M M}} + \\lambda_ {\\mathrm {G M M}} \\mathcal {L} _ {\\mathrm {G M M}} + \\lambda_ {\\beta} \\mathcal {L} _ {\\beta} \\\\ + \\lambda_ {\\theta} \\mathcal {L} _ {\\theta} + \\lambda_ {2 D} \\mathcal {L} _ {2 D} + \\lambda_ {P} \\mathcal {L} _ {P}) \\\\ \\end{array}", + "image_path": "6a4137ded0512022a732caa84dc6df854fbb205ffee8ab4848ccf8837e1a905f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "text", + "content": "Following [34], we divide the optimization into two stages. In the first stage, we optimize all three parameters. In the second stage, we optimize only " + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "text", + "content": ", keeping the shape " + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 426, + 296, + 474 + ], + "type": "text", + "content": " fixed. We detail the remaining losses below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "text", + "content": "Pose and shape priors. We compute a loss " + }, + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GMM}}" + }, + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "text", + "content": " based on the Gaussian Mixture pose prior of [4] and a shape loss " + }, + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\beta} = \\| \\beta \\|_2^2" + }, + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "text", + "content": ", which penalizes extreme deviations from the body model's mean shape." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 532, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 569 + ], + "type": "text", + "content": "Initial pose loss. To ensure we do not stray too far from the initialization, we penalize large deviations from the initial pose " + }, + { + "bbox": [ + 55, + 532, + 296, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta} = ||\\pmb{\\theta}' - \\pmb{\\theta}||_2^2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 572, + 297, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 297, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 297, + 669 + ], + "type": "text", + "content": "2D keypoint loss. Similar to BUDDI [34], for each person in the image, we obtain pseudo ground truth 2D keypoints and their confidence from OpenPose [5] and ViTPose [49]. Given this pseudo ground truth, we merge all the keypoints into " + }, + { + "bbox": [ + 55, + 572, + 297, + 669 + ], + "type": "inline_equation", + "content": "\\pmb{K} \\in \\mathbb{R}^{d_j \\times 2}" + }, + { + "bbox": [ + 55, + 572, + 297, + 669 + ], + "type": "text", + "content": ", and their corresponding confidence into " + }, + { + "bbox": [ + 55, + 572, + 297, + 669 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^{d_j}" + }, + { + "bbox": [ + 55, + 572, + 297, + 669 + ], + "type": "text", + "content": ". From the predicted " + }, + { + "bbox": [ + 55, + 572, + 297, + 669 + ], + "type": "inline_equation", + "content": "\\pmb{X}'" + }, + { + "bbox": [ + 55, + 572, + 297, + 669 + ], + "type": "text", + "content": ", we can compute the 2D projection of each 3D joint location using Equation 3.1. Then, the 2D keypoint loss is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 676, + 296, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 676, + 296, + 711 + ], + "spans": [ + { + "bbox": [ + 116, + 676, + 296, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2 D} = \\sum_ {j = 1} ^ {d _ {j}} \\boldsymbol {\\gamma} \\left(\\boldsymbol {K} _ {p r o j} - \\boldsymbol {K}\\right) ^ {2} \\tag {5}", + "image_path": "5d88f71ddd86ebb0cc976aa9cb91d7d446e2530b36ab4058bd9de5eecdffdefb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "Interpenetration loss. To prevent parts of one mesh from being in the interior of the other, we add an interpenetration loss. Generically, given two sets of vertices " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ", we use winding numbers to compute the subset of " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " that intersects " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ", which we denote as " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "V_{0,1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ". Similarly, " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "V_{1,0}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " is the subset of " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " that intersects " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "V_{0}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ". The interpenetration loss is then defined as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 163, + 555, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 163, + 555, + 202 + ], + "spans": [ + { + "bbox": [ + 320, + 163, + 555, + 202 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P} = \\sum_ {x \\in \\boldsymbol {V} _ {0, 1}} \\min _ {v _ {1} \\in \\boldsymbol {V} _ {1}} \\| x - v _ {1} \\| _ {2} ^ {2} + \\sum_ {y \\in \\boldsymbol {V} _ {1, 0}} \\min _ {v _ {0} \\in \\boldsymbol {V} _ {0}} \\| y - v _ {0} \\| _ {2} ^ {2} \\tag {6}", + "image_path": "c7cfa9af86469581f4e401e34c59aaa10544acb0dfec9f5696bb8132c86391db.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 213, + 556, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 556, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 556, + 237 + ], + "type": "text", + "content": "For efficiency, this loss is computed on low-resolution versions of the two meshes (roughly 1000 vertices per mesh)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 248, + 395, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 395, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 395, + 262 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 275, + 555, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 359 + ], + "type": "text", + "content": "Implementation details. Following prior work on two-person pose estimation [34], we use BEV [42] to initialize the poses since it was trained to predict both the body pose parameters and the placement of each person in the scene. However, on the single person yoga poses, we find that the pose parameter estimates of HMR2 [12] are much higher quality, so we initialize the body pose using HMR2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "text", + "content": "We use the SMPL-X [36] body model and (unless specified otherwise) GPT4-V [1] as the LMM with temperature " + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "inline_equation", + "content": "= 0.7" + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "text", + "content": " when sampling from it. GPT4-V refers to the gpt-4-vision-preview model in the OpenAI API: platform.openai.com. In the OpenAI API, we use the \"high\" detail setting for image input. Appendix 8.2 provides results using other prompts and other LMMs (LLaVA [31], GPT-4o) and a running time analysis. Unless otherwise specified, we set " + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "inline_equation", + "content": "N = 20" + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "text", + "content": " samples. For all of our 2-person experiments, " + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "inline_equation", + "content": "f = 1" + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "inline_equation", + "content": "f = 10" + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "text", + "content": " in the 1-person setting. We set " + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "inline_equation", + "content": "t = 2" + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "text", + "content": " for the experiment on the CHI3D dataset and " + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "inline_equation", + "content": "t = N" + }, + { + "bbox": [ + 313, + 360, + 556, + 529 + ], + "type": "text", + "content": " for all other experiments. The hyperparameters and our prompts were chosen based on experiments on the validation sets. For other implementation details refer to Appendix § 7." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 533, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 629 + ], + "type": "text", + "content": "Metrics. As is standard in the pose estimation literature, we report Procrustes-aligned Mean Per Joint Position Error (PA-MPJPE) in millimeters. This metric finds the best alignment between the estimated and ground-truth pose before computing the joint error. In the two-person setting, we focus on the joint PA-MPJPE, as this evaluation incorporates the relative translation and orientation of the two people. See Appendix § 8.2 for the per-person PA-MPJPE." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "We also include the percentage of correct contact points (PCC) metric introduced by [34]. This metric captures the fraction of ground-truth contact pairs that are accurately predicted. For a given radius " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", a pair is classified as \"in contact\" if the two regions are both within the specified radius. We use the set of fine-grained regions defined in [10] to compute PCC. The metric is averaged over " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "r \\in" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7129" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 298, + 197 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 197 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 197 + ], + "type": "table", + "html": "
Hi4DFlickrCI3DCHI3D
PM↓F1↑PM↓PCC↑F1↑PM↓PCC↑F1↑
w/o contact sup.
BEV [42]144-10664.8-9671.4-
Heuristic116-6777.8-10574.1-
ProsePose93245879.91310075.823
w. contact sup.
BUDDI [34]89-6681.9-6878.0-
BUDDI+ProsePose88-6583.2-6978.8-
Coarse GT contacts
Oracle8110043861008383.8100
", + "image_path": "8235dbab5c70c75196a0bf8ba4f6bcbbdd56ace04a4cafbc8b9f4ac9448af16f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 205, + 296, + 305 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 305 + ], + "type": "text", + "content": "Table 1. Two-person Results. Joint PA-MPJPE (abbreviated PM) (lower is better), Avg. PCC (higher is better), and F1 (higher is better). For FlickrCI3D, PA-MPJPE is computed using the pseudoground-truth fits. F1 measures the accuracy of coarse contacts predicted by the LMM, while the other two metrics evaluate the quality of the estimated 3D pose. Last line shows results using ground-truth contact pairs of coarse regions (heuristic is still used as the backoff method when needed). Bold indicates best method without/with contact supervision in each column." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": "0, 5, 10, 15, ..., 95 mm. Since these regions are defined on the SMPL-X mesh topology, we convert the regression baselines-BEV and HMR2- from the SMPL mesh topology to SMPL-X to compute this metric. Please see the Appendix § 8.1 for more details on the regions and on the mesh conversion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "text", + "content": "Finally, we report the F1 of the LMM's predicted coarse region pairs. Specifically, we compute precision as the proportion of predicted pairs " + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_a,\\mathbf{R}_b)" + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "text", + "content": " such that there is some ground-truth pair " + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_a^*,\\mathbf{R}_b^*)" + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "text", + "content": " for which " + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_a" + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "text", + "content": " overlaps with " + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_a^*" + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_b" + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "text", + "content": " overlaps with " + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_b^*" + }, + { + "bbox": [ + 55, + 379, + 296, + 535 + ], + "type": "text", + "content": ". We compute recall similarly as a proportion of ground-truth pairs. For datasets that do not provide contact maps, we define the ground-truth pairs from the ground-truth meshes, using a threshold on the minimum distance between regions (0.01 and 0.02 for Hi4D and MOYO, respectively). We ignore chirality when computing these. For a pair of people we take the maximum F1 of the two possible orderings. This metric serves to evaluate the raw output of the LMMs without 3D optimization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 214, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 214, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 214, + 556 + ], + "type": "text", + "content": "4.1. Two-person Pose Refinement" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "Datasets. We evaluate on three datasets, and our dataset processing largely follows [34]. Hi4D [50] is a motion capture dataset of pairs of people interacting. Each sequence has a subset of frames marked as contact frames, and we take every fifth contact frame. We use the images from a single camera, resulting in 241 images. Flickr Close Interactions 3D (FlickrCI3D) [10] is a collection of Flickr images of multiple people in close interaction. The dataset includes manual annotations of the contact maps between pairs of people. [34] used these contact maps to create pseudo-ground truth 3D meshes and curated a version of the test set to exclude noisy annotations, which has 1403 images. CHI3D [10] is" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 174 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 174 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 174 + ], + "type": "table", + "html": "
@ radius[mm]FlickrCI3D PCC↑CHI3D PCC↑
5101551015
W/o contact sup.
BEV [42]3.66.310.85.817.432.5
Heuristic14.633.949.311.128.045.3
ProsePose15.639.957.113.535.252.5
W/ contact sup.
BUDDI [34]18.544.261.815.539.056.6
BUDDI+ProsePose21.849.366.419.543.958.8
", + "image_path": "ee0a71bf6a93ee1c93e47b90787cf393f73b950baaa912009388f7d6e0adc40c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 182, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 237 + ], + "type": "text", + "content": "Table 2. Two-person PCC. Percent of correct contact points (PCC) for three different radii " + }, + { + "bbox": [ + 313, + 182, + 555, + 237 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 182, + 555, + 237 + ], + "type": "text", + "content": " in mm. Bold indicates the best score without/with contact supervision in each column. At the ground-truth contact points, our method brings the meshes closer together than the baselines." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 251, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 299 + ], + "type": "text", + "content": "a motion capture dataset of pairs of people interacting. We present results on the validation set. There are 431 images, distributed across 4 cameras. The images come from 126 video sequences, each of which has a single \"contact frame.\"" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 556, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 556, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 556, + 396 + ], + "type": "text", + "content": "To develop our method, we experimented on the validation sets of FlickrCI3D and Hi4D, and a sample of the training set from CHI3D. For our experiments, we can compute the PCC on FlickrCI3D and CHI3D, which have annotated ground-truth contact maps. Following [34], we exclude from evaluation images where BEV or the keypoint detectors, which are used by the baselines as well, fail to detect one of the subjects in the interaction pair." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 404, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 404, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 315, + 404, + 555, + 417 + ], + "type": "text", + "content": "Baselines We compare our estimated poses to the following:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 418, + 555, + 597 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 315, + 418, + 555, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 418, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 315, + 418, + 555, + 465 + ], + "type": "text", + "content": "- BEV [42] Multi-person 3D pose estimation method. Uses relative depth to reason about spatial placement of individuals in the scene. ProsePose, Heuristic, and BUDDI use BEV to initialize pose estimates." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 466, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 466, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 315, + 466, + 555, + 548 + ], + "type": "text", + "content": "- Heuristic A contact heuristic which includes the auxiliary losses in Section 3.3 as well as a term that minimizes the minimum distance between the two meshes. Introduced by [34]. We use their hyperparameters for this heuristic. This baseline is also used as the backoff method for Prose-Pose when the number of empty constraint sets is at least the threshold " + }, + { + "bbox": [ + 315, + 466, + 555, + 548 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 315, + 466, + 555, + 548 + ], + "type": "text", + "content": " or when the optimization diverges." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 549, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 549, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 315, + 549, + 555, + 597 + ], + "type": "text", + "content": "- BUDDI [34] This method uses a learned diffusion prior to constrain the optimization. We stress that BUDDI requires a large amount of annotated training data on pairs of interacting bodies, which is not used in our method." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Quantitative Results Table 1 provides quantitative results on the three datasets. Across datasets, ProsePose consistently improves over the strongest baseline, Heuristic. On the Hi4D dataset, ProsePose reduces " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " of the gap in PA-MPJPE between Heuristic and the fully supervised BUDDI. On the FlickrCI3D and CHI3D datasets, ProsePose narrows the gap in the average PCC between Heuristic and BUDDI by more than one-third. (While ProsePose achieves a better PA-MPJPE than BUDDI on FlickrCI3D, for this dataset," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7130" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 286, + 460 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 286, + 460 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 286, + 460 + ], + "type": "image", + "image_path": "6c2efec2ce40dd45c106171a96d1a9043fe73f7a0ee93222955a8926130539ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 468, + 296, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 536 + ], + "type": "text", + "content": "Figure 3. Two-person examples We show qualitative results from ProsePose, BUDDI [34], and the contact heuristic. For each example, we show GPT4-V's top 3 constraints and the number of times each constraint was predicted across all 20 samples. Our method correctly reconstructs people in a variety of interactions, and the predicted constraints generally align with each interaction type." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "type": "text", + "content": "we rely primarily on PCC since PA-MPJPE is computed on pseudo-ground-truth fits.)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "On CHI3D, ProsePose outperforms Heuristic but underperforms BEV in terms of PA-MPJPE. On the subset of images where we do not default to the heuristic (i.e. on images where GPT4-V predicts enough non-empty constraint sets), the PA-MPJPE for ProsePose and BEV is 86 and 87, respectively. In other words, in the cases where our method is actually used, the joint error is slightly less than that of BEV. As a result, we can attribute the worse overall error to the poorer performance of the heuristic. The backoff method (which is the heuristic) is used in 13/241 Hi4D examples, 106/1403 Flickr examples, and 224/431 CHI3D examples." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 345, + 71, + 523, + 373 + ], + "blocks": [ + { + "bbox": [ + 345, + 71, + 523, + 373 + ], + "lines": [ + { + "bbox": [ + 345, + 71, + 523, + 373 + ], + "spans": [ + { + "bbox": [ + 345, + 71, + 523, + 373 + ], + "type": "image", + "image_path": "7ed66116ef372112d24251ca8d321a11ce3dfcadfd07283e5984f0a8cf653466.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 377, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 445 + ], + "type": "text", + "content": "Figure 4. Single-person examples We show qualitative results from ProsePose, HMR2 [12], and HMR2-optim on complex yoga poses. Each example also shows the constraints that are predicted by the LMM at least " + }, + { + "bbox": [ + 313, + 377, + 555, + 445 + ], + "type": "inline_equation", + "content": "f = 10" + }, + { + "bbox": [ + 313, + 377, + 555, + 445 + ], + "type": "text", + "content": " times (and are thus used to compute " + }, + { + "bbox": [ + 313, + 377, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LMM}}" + }, + { + "bbox": [ + 313, + 377, + 555, + 445 + ], + "type": "text", + "content": ") with their counts. ProsePose correctly identifies self-contact points and optimizes the poses to respect these contacts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "content": "Overall, our method improves over the other methods that do not use contact supervision in terms of both joint error and PCC. While not the focus of this work, Table 1 also shows that adding ProsePose, specifically " + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LMM}}" + }, + { + "bbox": [ + 313, + 460, + 556, + 604 + ], + "type": "text", + "content": ", to BUDDI leads to improved PCC. The F1 scores show that LMMs' raw output is often flawed/incomplete, but our approach mitigates hallucinations in various ways (see the ablation study below). The last row in Table 1 shows the performance when the ground-truth coarse contacts (with correct left/right labels) are used in optimization. These results show the benefit of correct coarse contacts and the clear room for improvement from better LMM predictions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": "Table 2 shows the PCC for each method at various radii. The results show that ProsePose brings the meshes closer together at the correct contact points. On both the FlickrCI3D and CHI3D datasets, ProsePose outperforms the other baselines that do not use contact supervision. Next, we ablate important aspects of ProsePose. In Figure 5, we show that averaging the loss over several samples from the LMM improves performance, mitigating the effect of LMM hallucination. Finally, ablating the various losses in optimization indicates" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7131" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 171, + 159 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 171, + 159 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 171, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 171, + 159 + ], + "type": "image", + "image_path": "5f7d56f72810ad803f3e384aa147ede287596822f2f8613606b608b9a13873c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 169, + 297, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 169, + 297, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 297, + 214 + ], + "type": "text", + "content": "Figure 5. More samples improve pose estimation. On the Flickr3D validation set, taking more samples from the LMM and averaging the resulting loss functions improves joint PA-MPJPE (left) and average PCC (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 282, + 159 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 282, + 159 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 282, + 159 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 282, + 159 + ], + "type": "image", + "image_path": "1e233f6ec7e030e3ae15bdffc97c715de01146a0c223a9a5e0832204716006c2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 227, + 294, + 292 + ], + "blocks": [ + { + "bbox": [ + 56, + 227, + 294, + 292 + ], + "lines": [ + { + "bbox": [ + 56, + 227, + 294, + 292 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 294, + 292 + ], + "type": "table", + "html": "
PA-MPJPE↓PCC↑PCC↑@rF1↑
51015
HMR2 [12]8483.034.255.269.5-
HMR2+opt8185.247.765.574.6-
ProsePose8287.854.273.881.425
", + "image_path": "02065261a5d54f6145cf74809761f431644993a4e224aa8878bb9bef95abf6d2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 296, + 334 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 334 + ], + "type": "text", + "content": "Table 3. One-person Results. PA-MPJPE (lower is better) and Avg. PCC and F1 (higher is better). ProsePose captures ground-truth contacts better than the baselines, as shown by the PCC." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 350, + 295, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 398 + ], + "type": "text", + "content": "that our LMM-based loss and the 2D keypoint loss have the greatest impact on joint error: Using all losses results in a PA-MPJPE of 81. Removing " + }, + { + "bbox": [ + 55, + 350, + 295, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LMM}} / \\mathcal{L}_{\\mathrm{GMM}} / \\mathcal{L}_{\\beta} / \\mathcal{L}_{\\theta} / \\mathcal{L}_{2D} / \\mathcal{L}_P" + }, + { + "bbox": [ + 55, + 350, + 295, + 398 + ], + "type": "text", + "content": " results in a PA-MPJPE of 138/85/91/84/130/78." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 404, + 297, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 297, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 297, + 559 + ], + "type": "text", + "content": "Qualitative Results Figure 3 shows examples of reconstructions from ProsePose, Heuristic, and BUDDI. Below each of our predictions, we list the most common constraints predicted by GPT4-V for the image. The predicted constraints correctly capture the semantics of each interaction. For instance, in tango, one person's arm should touch the other's back. In a rugby tackle, a player's arms are usually wrapped around the other player. Using these constraints, ProsePose correctly reconstructs a variety of interactions, such as tackling, dancing, and holding hands. In contrast, Heuristic struggles to accurately position individuals and/or predict limb placements, often resulting in awkward distances." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 568, + 211, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 211, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 211, + 582 + ], + "type": "text", + "content": "4.2. One-person pose refinement" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 713 + ], + "type": "text", + "content": "Datasets Next, we evaluate ProsePose on a single-person setting. For this setting, we evaluate on MOYO [45], a motion capture dataset with videos of a single person performing various yoga poses. In total, our test set is composed of 76 examples from a single camera angle (side view). See Appendix 7.6 for further dataset details. Since this dataset does not have annotated region contact pairs, we compute the peso-ground-truth contact maps using the Euclidean and geodesic distance following [33] and report PCC for the subset of 67 examples where the ground-truth has self-contact." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 537, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 537, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 537, + 85 + ], + "type": "text", + "content": "Baselines We compare against the following baselines:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 87, + 555, + 148 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 87, + 555, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 87, + 555, + 111 + ], + "spans": [ + { + "bbox": [ + 314, + 87, + 555, + 111 + ], + "type": "text", + "content": "- HMR2 [12] State-of-the-art pose regression method. We use HMR2 to initialize our pose estimates for optimization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 112, + 555, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 112, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 112, + 555, + 148 + ], + "type": "text", + "content": "- HMR2+opt Optimization procedure that is identical to ours without " + }, + { + "bbox": [ + 313, + 112, + 555, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LMM}}" + }, + { + "bbox": [ + 313, + 112, + 555, + 148 + ], + "type": "text", + "content": ". It is the default method when the number of empty constraint sets is at least the threshold " + }, + { + "bbox": [ + 313, + 112, + 555, + 148 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 112, + 555, + 148 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 150, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 555, + 319 + ], + "type": "text", + "content": "Both the quantitative and qualitative results echo the trends discussed in the 2-person setting. Table 3 provides the quantitative results. The PCC metrics show that our LMM loss improves the predicted self-contact in complex yoga poses relative to the two baselines. The backoff method is used in 43/76 examples. Figure 4 provides a qualitative comparison of poses predicted by ProsePose versus the two baselines. Below each of our predictions, we list the corresponding constraints predicted by GPT4-V. In each case, the predicted constraint captures the correct self-contact, which is reflected in the final pose estimates. Using the semantically guided loss, ProsePose effectively refines the pose to ensure proper contact between hand-foot or hand-hand, an important detail consistently overlooked by the baselines." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 336, + 391, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 336, + 391, + 348 + ], + "spans": [ + { + "bbox": [ + 314, + 336, + 391, + 348 + ], + "type": "text", + "content": "4.3. Limitations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 357, + 556, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 556, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 556, + 502 + ], + "type": "text", + "content": "While ProsePose consistently improves contact across settings and datasets, it has limitations which are related to failures of LMMs. First, as shown in Table 5 (in Appendix), prompting the LMM for left/right labels sometimes leads to worse results, suggesting that LMMs struggle with disambiguating chirality. Improving this approach depends in large part on correctly identifying limbs as left/right. Another limitation is the use of coarse regions. Future work could improve by eliciting more fine-grained constraints from an LMM. Finally, LMM accuracy varies moderately across camera angles (quantified in Appendix 8.2). In Appendix § 8.3, we provide examples of LMM failures." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 522, + 388, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 522, + 388, + 534 + ], + "spans": [ + { + "bbox": [ + 314, + 522, + 388, + 534 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "We present ProsePose, a framework for refining 3D pose estimates to capture touch accurately using the implicit semantic knowledge of poses in LMMs. Our key novelty is that we generate structured pose descriptions from LMMs and convert them into loss functions used to optimize the pose. Our experiments show that in both one-person and two-person settings, ProsePose improves over previous baselines that do not use contact supervision. These results suggest that LMMs may be useful in creating larger datasets with contact annotations, which are otherwise expensive but are crucial for training state-of-the-art priors for pose estimation in situations with physical contact. More broadly, this work provides evidence that LMMs are promising tools for 3D pose estimation, which likely has implications beyond touch." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7132" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 297, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 297, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 297, + 129 + ], + "type": "text", + "content": "SS, EN, and TD were supported in part by the NSF, DoD, and/or the Berkeley Artificial Intelligence Research (BAIR) industrial alliance program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 145, + 115, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 145, + 115, + 158 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 115, + 158 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 163, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 163, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 60, + 163, + 297, + 715 + ], + "type": "text", + "content": "[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mo Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Benjamin Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Sim'on Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Raphael Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heating, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Hendrik Kirchner, Jamie Ryan Kiros, Matthew Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Adeola Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovski, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney Christine McLeavey Paul McMillan Jake McNeil David Medina Alok Mehta Jacob Menick Luke Metz Andrey Mishchenko Pamela Mishkin Vinnie Monaco Evan Morikawa Daniel P. Mossing Tong Mu Mira Murati Oleg Murk David M'ely Ashvin Nair Reichiro Nakano Rajeev Nayak Arvind Neelakantan Richard Ngo Hyeonwoo Noh Ouyang Long Cullen O'Keefe Jakub W. Pachocki Alex Paino Joe Palermo Ashley Pantuliano Giambattista Parascandolo Joel Parish Emy Parparita Alexandre Passos Mikhail Pavlov Andrew Peng Adam Perelman Filipe" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 334, + 73, + 556, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 73, + 556, + 357 + ], + "spans": [ + { + "bbox": [ + 334, + 73, + 556, + 357 + ], + "type": "text", + "content": "de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario D. Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin D. Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas A. Tezak, Madeleine Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cer'on Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao, Tao Xu, Sarah Yoo, Kevin Yu, Qiming Yuan, Wojciech Zaremba, Rowan Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. Gpt-4 technical report. 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 319, + 360, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 319, + 360, + 556, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 360, + 556, + 415 + ], + "spans": [ + { + "bbox": [ + 319, + 360, + 556, + 415 + ], + "type": "text", + "content": "[2] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3395-3404, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 319, + 417, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 417, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 319, + 417, + 555, + 450 + ], + "type": "text", + "content": "[3] Ana Aznar and Harriet R Tenenbaum. Parent-child positive touch: Gender, age, and task differences. Journal of nonverbal behavior, 40:317-333, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 319, + 452, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 452, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 319, + 452, + 556, + 529 + ], + "type": "text", + "content": "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pages 561-578. Springer, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 531, + 556, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 531, + 556, + 575 + ], + "spans": [ + { + "bbox": [ + 319, + 531, + 556, + 575 + ], + "type": "text", + "content": "[5] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 577, + 556, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 577, + 556, + 632 + ], + "spans": [ + { + "bbox": [ + 319, + 577, + 556, + 632 + ], + "type": "text", + "content": "[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Albert Li, Pascale Fung, and Steven C. H. Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. ArXiv, abs/2305.06500, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 634, + 556, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 634, + 556, + 678 + ], + "spans": [ + { + "bbox": [ + 319, + 634, + 556, + 678 + ], + "type": "text", + "content": "[7] Delmas, Ginger and Weinzaepfel, Philippe and Lucas, Thomas and Moreno-Noguer, Francesc and Rogez, Grégory. PoseScript: 3D Human Poses from Natural Language. In ECCV, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 680, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 680, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 319, + 680, + 556, + 714 + ], + "type": "text", + "content": "[8] Delmas, Ginger and Weinzaepfel, Philippe and Moreno-Noguer, Francesc and Rogez, Grégory. PoseFix: Correcting 3D Human Poses with Natural Language. In ICCV, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7133" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 294, + 106 + ], + "type": "text", + "content": "[9] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. Posegpt: Chatting about 3d human pose. ArXiv, abs/2311.18836, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "type": "text", + "content": "[10] Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Sminchisescu. Three-dimensional reconstruction of human interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7214-7223, 2020. 2, 4, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 296, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 296, + 219 + ], + "type": "text", + "content": "[11] Mihai Fieraru, Mihai Zanfir, Teodor Szente, Eduard Bazavan, Vlad Olaru, and Cristian Sminchisescu. Remips: Physically consistent 3d reconstruction of multiple interacting people under weak supervision. Advances in Neural Information Processing Systems, 34:19385-19397, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 222, + 296, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 296, + 276 + ], + "type": "text", + "content": "[12] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2, 5, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "type": "text", + "content": "[13] Peng Guan, Alexander Weiss, Alexandru O Balan, and Michael J Black. Estimating human shape and pose from a single image. In 2009 IEEE 12th International Conference on Computer Vision, pages 1381-1388. IEEE, 2009. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 325, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 296, + 369 + ], + "type": "text", + "content": "[14] Riza Alp Guler and Iasonas Kokkinos. Holopose: Holistic 3d human reconstruction in-the-wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10884-10894, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "text", + "content": "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152-5161, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 428, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 294, + 460 + ], + "type": "text", + "content": "[16] Jinni A Harrigan. Self-touching as an indicator of underlying affect and language processes. Social science & medicine, 20 (11):1161-1168, 1985. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 463, + 294, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 294, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 294, + 528 + ], + "type": "text", + "content": "[17] Buzhen Huang, Chen Li, Chongyang Xu, Liang Pan, Yanggang Wang, and Gim Hee Lee. Closely interactive human reconstruction with proxemics and physics-guided adaption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1011-1021, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 531, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 531, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 294, + 563 + ], + "type": "text", + "content": "[18] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 296, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 296, + 620 + ], + "type": "text", + "content": "[19] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 623, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 296, + 667 + ], + "type": "text", + "content": "[20] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3d human model fitting towards inthe-wild 3d human pose estimation. In 2021 International Conference on 3D Vision (3DV), pages 42-52. IEEE, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[21] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "type": "text", + "content": "[22] Rawal Khirodkar, Jyun-Ting Song, Jinkun Cao, Zhengyi Luo, and Kris M. Kitani. Harmony4d: A video dataset for in-the-wild close human interactions. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 149 + ], + "type": "text", + "content": "[23] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 555, + 194 + ], + "type": "text", + "content": "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 555, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 555, + 249 + ], + "type": "text", + "content": "[25] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2252-2261, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 555, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 555, + 304 + ], + "type": "text", + "content": "[26] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 304, + 555, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 555, + 358 + ], + "type": "text", + "content": "[27] Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. arXiv preprint arXiv:2311.16922, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 360, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 555, + 403 + ], + "type": "text", + "content": "[28] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 555, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 555, + 447 + ], + "type": "text", + "content": "[29] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. arXiv preprint arXiv:2304.05684, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 555, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 469 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 471, + 555, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 555, + 503 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.5,6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 535 + ], + "type": "text", + "content": "[32] Yujie Lu, Dongfu Jiang, Wenhu Chen, William Wang, Yejin Choi, and Bill Yuchen Lin. Wildvision arena: Benchmarking multimodal llms in the wild, 2024. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 555, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 555, + 581 + ], + "type": "text", + "content": "[33] Lea Muller, Ahmed AA Osman, Siyu Tang, Chun-Hao P Huang, and Michael J Black. On self-contact and human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9990-9999, 2021. 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "type": "text", + "content": "[34] Lea Müller, Vickie Ye, Georgios Pavlakos, Michael Black, and Angjoo Kanazawa. Generative proxemics: A prior for 3d social interaction from images. arXiv preprint arXiv:2306.09337, 2023. 2, 4, 5, 6, 7, 9" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 555, + 689 + ], + "type": "text", + "content": "[35] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 712 + ], + "type": "text", + "content": "[36] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7134" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2, 3, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 150 + ], + "type": "text", + "content": "[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. *Big Data*, 4(4):236–252, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "text", + "content": "[38] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 722-731, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "text", + "content": "[39] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "text", + "content": "[40] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 296, + 342 + ], + "type": "text", + "content": "[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11179-11188, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 296, + 399 + ], + "type": "text", + "content": "[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13243-13252, 2022. 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 400, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 296, + 465 + ], + "type": "text", + "content": "[43] Mika S Takeuchi, Hitoshi Miyaoka, Atsuko Tomoda, Masao Suzuki, Qingbo Liu, and Toshinori Kitamura. The effect of interpersonal touch during childhood on adult attachment and depression: A neglected area of family and developmental psychology? Journal of Child and Family Studies, 19:109-117, 2010. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "type": "text", + "content": "[44] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 513, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 296, + 566 + ], + "type": "text", + "content": "[45] Shashank Tripathi, Lea Müller, Chun-Hao P. Huang, Taheri Omid, Michael J. Black, and Dimitrios Tzionas. 3D human pose estimation via intuitive physics. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 4713-4725, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 622 + ], + "type": "text", + "content": "[46] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 2022 International Conference on 3D Vision (3DV), pages 353-362. IEEE, 2022. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "text", + "content": "[47] X Wang, J Wei, D Schuurmans, Q Le, E Chi, S Narang, A Chowdhery, and D Zhou. Self-consistency improves chain of thought reasoning in language models. arxiv. Preprint posted online March, 21:10-48550, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[48] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Huai hsin Chi, F. Xia, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. ArXiv, abs/2201.11903, 2022. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 295 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "type": "text", + "content": "[49] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In Advances in Neural Information Processing Systems, 2022. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 172 + ], + "type": "text", + "content": "[50] Yifei Yin, Chen Guo, Manuel Kaufmann, Juan José Zárate, Jie Song, and Otmar Hilliges. Hi4d: 4d instance segmentation of close human interaction. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17016-17027, 2023. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 555, + 228 + ], + "type": "text", + "content": "[51] Andrei Zanfir, Elisabeta Marinhoiu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2148-2157, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 295 + ], + "type": "text", + "content": "[52] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7135" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/61fa7d4b-03fc-4f0b-b84c-af353e8bd6fc_content_list.json b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/61fa7d4b-03fc-4f0b-b84c-af353e8bd6fc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d0d1c1a94dacd4cf6b67238b9ca344d2887c09a9 --- /dev/null +++ b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/61fa7d4b-03fc-4f0b-b84c-af353e8bd6fc_content_list.json @@ -0,0 +1,1957 @@ +[ + { + "type": "text", + "text": "Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction", + "text_level": 1, + "bbox": [ + 140, + 130, + 857, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kaixin Fan, Pengfei Ren\\*, Jingyu Wang, $^{\\dagger}$ Haifeng Sun, Qi Qi, Zirui Zhuang, Jianxin Liao \nState Key Laboratory of Networking and Switching Technology, \nBeijing University of Posts and Telecommunications", + "bbox": [ + 151, + 203, + 844, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fkxbupt, rpf, wangjingyu, hfsun, qiqi8266, zhuangzirui, liaojx}@bupt.edu.cn", + "bbox": [ + 158, + 258, + 831, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 309, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D hand reconstruction is essential in non-contact human-computer interaction applications, but existing methods struggle with low-resolution images, which occur in slightly distant interactive scenes. Leveraging temporal information can mitigate the limitations of individual low-resolution images that lack detailed appearance information, enhancing the robustness and accuracy of hand reconstruction. Existing temporal methods typically use joint features to represent temporal information, avoiding interference from redundant background information. However, joint features excessively disregard the spatial context of visual features, limiting hand reconstruction accuracy. We propose to integrate temporal joint features with visual features to construct a robust low-resolution visual representation. We adopt Triplane Features, a dense representation with 3D spatial awareness, to bridge the gap between the joint features and visual features that are misaligned in terms of representation form and semantics. Triplane Features are obtained by orthogonally projecting joint features, embedding hand structure information into the 3D spatial context. Furthermore, we compress the spatial information of the three planes into a 2D dense feature though Spatial-Aware Fusion to enhance the visual features. By using enhanced visual features enriched with temporal information for hand reconstruction, our method achieves competitive performance at much lower resolutions compared to state-of-the-art methods operating at high resolution on DexYCB, HanCo and H2O. Code is available at https://github.com/NewbieFan/Temp-LowRes-hand.", + "bbox": [ + 86, + 340, + 483, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 808, + 220, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D hand reconstruction, a crucial component in non-contact human-computer interaction applications, has made signifi-", + "bbox": [ + 89, + 834, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2618034399817112badc1206dd577a8dfd8ade133e6a17c642afce196b696cd4.jpg", + "image_caption": [ + "(a) No Interaction", + "(b) Cross Attention" + ], + "image_footnote": [], + "bbox": [ + 519, + 324, + 903, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f0bedb21c5bc037a15ca55228c1445547c7e5523607938d98c418d7af2f7305.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 392, + 903, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac48e0fe3f9bdf291ab41a7e40a1a357e219ef9dde255ddd99628040548fb0e0.jpg", + "image_caption": [ + "(c) Triplane Feature", + "Figure 1. Different way to utilizing joint features. a) Directly reconstruct hand from joint features, relying solely on sparse features without spatial context. b) Fusing joint and visual features through cross-attention, which avoids representation discrepancies but neglects the hand's structural priors. c) Using triplane features to bridge joint and visual features by decoding structural joint information into dense representations that enhance visual features." + ], + "image_footnote": [], + "bbox": [ + 521, + 501, + 903, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cant advancements in recent years with the advent of large-scale datasets and deep learning technologies. Many studies [6, 7, 30, 31, 34, 40] have focused on reconstructing 3D hands from a single RGB image. In typical interaction scenarios, these methods require the subject to stay close to the capture camera to obtain high-resolution (e.g., $256 \\times 256$ ) cropped hand images. However, as the interaction distance increases—such as in XR teaching activities within a classroom—the available cropped hand images are limited to lower resolutions (e.g., $64 \\times 64$ ). Current methods", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 109, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22627", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "perform poorly on such low-resolution images, as these images provide fewer cues for addressing joint self-similarity and depth ambiguity, significantly increasing the difficulty of 3D hand reconstruction.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the low-resolution problem, one intuitive approach is to lift the resolution through super-resolution techniques [14]. However, general super-resolution methods can lead to structural distortions and artifacts in the hand regions, which brings a domain gap between super-resolved image and real high-resolution image. Low-Res Pose [45] adopted confidence-aware learning to supervise continue joint offset fields, thereby alleviating the quantization errors. RSC-Net [48] adopted cross-resolution contrastive learning to align the estimation results of low-resolution images with high-resolution images. However, the performance of these methods is still limited due to the inherent ambiguity resulting from information loss in low resolution.", + "bbox": [ + 88, + 154, + 482, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A single low-resolution image lacks detailed hand appearance, leading to ambiguous visual representation. Continuous frames provide comprehensive hand appearance and motion context, offering strong disambiguation cues for inferring global hand structure. To reduce interference from redundant background information and computational costs, existing temporal methods typically compress single-frame information into joint features [51, 58], and then model the intra- and inter-frame correlation between joints through temporal interaction. However, joint features excessively discard spatial context of visual features, which limits the accuracy of pose estimation. Integrating the spatial context of visual features with the temporal context of joint features can construct robust visual representation with 3D spatial awareness. However, integrating joint features into visual features present two major challenges. Firstly, sparse joint features are not aligned with dense visual features in terms of representation form. Secondly, joint features contain prior geometric structural information in 3D space (e.g., hand scale, skeleton length), while visual features represent detailed appearance information in 2D space (e.g., texture, color), leading to a semantic misalignment.", + "bbox": [ + 91, + 354, + 483, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we integrate the temporal clue from historical frames into the current visual features to achieve robust and accurate 3D hand pose estimation under low-resolution scenarios. Specifically, we compress temporal information into joint representations, and model intra- and interframe relationships of joints through spatio-temporal interaction. Then, we adopt triplane features[4], a 3D spatial-aware dense visual representation, to bridge the sparse 3D joint representation and the dense 2D visual representation. We orthogonally project the joint features into three individual planes, with each plane independently embedding the sparse joint features into specific spatial contexts. The orthogonal planes ensure comprehensive coverage of 3D hand structure information in a low-redundancy manner. How", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ever, they still contain redundant spatial information, directly integrating these features into the visual plane leads to the overlap of spatial context from different planes. Thus, we compress the features from two planes orthogonal to the visual plane into 1D global features, guiding each pixel in visual features to integrate distinct spatial context from the other planes, reducing redundancy and clarifying the 3D spatial context. Finally, we reconstruct the hand based on the enhanced visual features.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results on several large-scale hand pose estimation video datasets, DexYCB [5], HanCo [60] and H2O [25], show the proposed method reach competitive performance at much lower resolution compared to both single-frame based and temporal based state-of-the-art (SOTA) methods at high resolution. In summary, our contribution is list as below:", + "bbox": [ + 511, + 227, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a temporal enhancement framework that integrates temporal joint features with visual features to construct a robust visual representation for low-resolution conditions, thereby improving hand reconstruction performance in low-resolution videos.", + "- We introduced triplane feature encoding to encode joint features into dense feature maps. Through triplane feature, visual features can perceive the prior hand structure information from the temporal joint features.", + "- Our method reaches competitive result at much lower resolution compared to SOTA methods at high resolution." + ], + "bbox": [ + 513, + 337, + 903, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Relate Work", + "text_level": 1, + "bbox": [ + 513, + 518, + 643, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Single Image based Hand Pose Estimation", + "text_level": 1, + "bbox": [ + 511, + 545, + 870, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early works [2, 17, 18, 36, 43, 59] focused on estimating hand pose solely from input images, which suffered from occlusion issues due to the lack of modeling the inherent structure of the hand. MANO [41] introduces a parametric hand model, which contains rich priors of human hands, and now most work focuses on reconstruct this pre-defined hand from image. Some works [1, 23, 32, 37-39, 52, 56] directly predict parameters of MANO, 3D joints and mesh vertices coordinates are then retrieved using the differentiable MANO model. However, predicting MANO parameters directly from a single RGB image is a challenging job, which is difficult for network optimization and is prone to mesh-image misalignment [44], especially in a low-res input condition. To alleviate these issues present in parametric predicting, some works [13, 24, 27] regress the 3D coordinates or occupancy directly for mesh vertices. For example, [53] adopts graph convolution neural network to regress 2.5D Heatmap for every mesh vertices. [26] proposed an End to end framework to recover mesh using transformer. These works show that using a regression-based approach can maintain flexibility while achieving comparable results to parametric methods. Thus, we adopt a regression-based", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22628", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method described in [53] to reconstruct the hand mesh from enhanced visual features.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Low Resolution Pose Estimation", + "text_level": 1, + "bbox": [ + 89, + 132, + 375, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, a few works have explored on low resolution pose estimation. A intuitive way is leveraging Super Resolution approach [14] to improve the image quality and pose estimation performance. However, the addition super-resolution network brings extra computation cost and may bring unpleasant artifacts to the input image. [45] find that offset-based methods can provide more reliable results in low resolution, and further proposed a confidence-aware learning to selectively weigh the learning of heatmap and offset. [48] controls the scaling of the CNN architecture model using a set of hyperparameters, enforcing it to output scale-invariant features at different resolutions through contrastive learning and self-supervised learning. Although specific learning schemas improve the poor results in low resolution, the features extracted by the network are still constrained by the lack of detailed hand appearance. Other work [33] explored on 3D pose estimation in low-resolution multi-view videos, showing that temporal and multi-view information could be a supplement to alleviate information loss. However, they only correct the prediction under certain adverse conditions, ignoring temporal information can reduce the inherent ambiguity in single low-resolution image and imporve the overall performance.", + "bbox": [ + 89, + 155, + 483, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Temporal Hand Pose Estimation", + "text_level": 1, + "bbox": [ + 89, + 513, + 375, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some recent methods [11, 15, 22, 29, 50, 55] have begun to explore leveraging temporal information in videos to overcome the limitations of single-view images. Hasson et al. [15] use photometric supervision to keep consistency between the re-projected 3D hand and the optical flow. VIBE [22] use a temporal generation network to predict model parameters for each frame, and employs a recurrent network as a discriminator to supervise motion sequence. However, they ignored the spatial and feature correlations in the video. More recently, [20] has harnessed temporal information by propagating frame-wise features with a sequence-to-sequence model based on LSTM [16], but it only provides sparse joint location predictions. TCMR [9] introduces a temporally consistent system that synthesizes implicit features from past and future frames to maintain temporal consistency in the mesh of the current frame. Deformer [12] deforms the hand poses from neighborhood frames using predicted motion to avoid unreliable queries generated with local feature when leveraging self-attention mechanism.", + "bbox": [ + 89, + 537, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In low resolutions, these methods are limited to the poor quality of single frame features. Additionally, they lack a focus center when processing temporal features, which may cause some estimation results to deviate from the actual image due to the influence of temporal context.", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of our method is shown in the Fig. 2. For a given frame $I_{T}$ in an image sequence $V = \\{I_{t}\\}_{t=1}^{T}$ , we extract temporal cues from the previous $T - 1$ frames to enhance the visual feature of the current frame and reconstruct the hand mesh from the enhanced visual feature. Specifically, we first compress the information of each historical frames into joint features and establish their inter- and intra-frame correlations through a spatio-temporal Transformer. Then, we encode the structural information from the enhanced joint features into 3D spatial-aware dense representations through triplane feature encoding to enhance the visual features of the current frame. Finally, we leverage the enhanced visual feature to reconstruct the hand mesh.", + "bbox": [ + 511, + 114, + 906, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Static Visual Features", + "text_level": 1, + "bbox": [ + 511, + 319, + 718, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 3, we adopt a Vision Transformer (ViT) [10] to extract visual features. During the extraction process, ambiguous local features would propagate from low-level features to subsequent high-level features, leading to ambiguities in both semantic content and fine-grained details in the final visual features. Thus, we incorporate temporal information into intermediate features at different levels to gradually reduce ambiguous local representations and ultimately obtain visual features with a clear perception of hand structure. However, static visual features lack the ability to perceive dynamic temporal information represented in joint features. Therefore, we introduce pose-aware tokens as anchors for fusing dynamic temporal information into a static representation, which are used for further interaction with visual features. These tokens capture hand pose and appearance information from the current visual feature through attention mechanism. We categorize the functionalities of these tokens based on finger semantic to ensure they incorporate both fine-grained visual details (e.g., texture) and high-level semantic content (e.g., skeletal structures).", + "bbox": [ + 511, + 340, + 908, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dynamic Temporal Joint Features", + "text_level": 1, + "bbox": [ + 511, + 652, + 812, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use compact joint features to represent dynamic temporal information from historical frames. Give a single frame, we obtain its visual feature map along with pose-aware tokens through the backbone. Each pose-aware token carries semantic information corresponding to different individual fingers. We further combine local visual features to expand each token into a complete joint feature, enabling fine-grained temporal interaction around each joints. We adopt a FC layer to predict 2D joint coordinates $p \\in \\mathbb{R}^{J \\times 2}$ from pose-aware tokens, and use it to sample joint-centric local features $J^{sample} \\in \\mathbb{R}^{J \\times C}$ from the visual feature map. The joint coordinates is encoded as pose feature $J^{pose}$ , and then add to the sampled feature $J^{sample}$ to obtain the initial joint feature $J^{init} = J^{pose} + J^{region}$ . We concatenate the joint features from each frame in the upstream sequence", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22629", + "bbox": [ + 478, + 945, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/faafee0f03e7d489097e453072fb8ebd15ec990db44bf1bb8932c1c5fc22d5ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 85, + 879, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2eb451d28a4708f478ee48e5c5e4b52db636d23418b8290c9329ecf0eaeecfa5.jpg", + "image_caption": [ + "Figure 2. Overview of Our Framework. We first construct joint features from the sampled past frame features and 2D poses using a Spatial-Temporal Transformer (a), capturing the temporal dynamics and spatial relationships between joints. We then use these joint features to enhance the visual features of the current frame (b) by converting them into spatial-aware dense features through Tripalne Feature Encoding (c). In (c), we fuse different joint features (from a) using pose-aware tokens (from b), and project them into three orthogonal planes to obtain the Triplane features. These features are then compressed into a single plane through space-aware fusion." + ], + "image_footnote": [], + "bbox": [ + 109, + 229, + 879, + 366 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e3485bd1c5523836e146b411c8b2d282169e5d2e5f5d8a8bd821da104457b043.jpg", + "image_caption": [ + "Figure 3. Backbone of our framework." + ], + "image_footnote": [], + "bbox": [ + 96, + 472, + 460, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to form the initial representation of temporal information $J^{temp} = \\text{Concat}(J_t^{init}) \\in \\mathbb{R}^{T \\times J \\times C}$ . Then, we adopt a spatio-temporal Transformer to establish inter- and intra-frame correlation between these joints. The spatio-temporal Transformer is a combination of alternately stacked temporal and spatial attention modules. In the spatial attention module, attention interactions occur among different joints within the same joint feature:", + "bbox": [ + 89, + 643, + 483, + 765 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS - M H S A (J ^ {t e m p}) = M H S A \\left(Q _ {j}, K _ {j}, V _ {j}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 773, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {j} = W _ {q} J _ {j} ^ {\\text {t e m p}}, K _ {j} = W _ {k} J _ {j} ^ {\\text {t e m p}}, V _ {j} = W _ {v} J _ {j} ^ {\\text {t e m p}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 797, + 482, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the temporal attention module, attention interactions occur among the same joint within different joint features:", + "bbox": [ + 89, + 820, + 482, + 851 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT - M H S A \\left(J ^ {t e m p}\\right) = M H S A \\left(Q _ {t}, K _ {t}, V _ {t}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 859, + 482, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {t} = W _ {q} J _ {t} ^ {t e m p}, K _ {j} = W _ {k} J _ {t} ^ {t e m p}, V _ {j} = W _ {v} J _ {t} ^ {t e m p} \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $W_{q}$ , $W_{k}$ , and $W_{v}$ are learnable weight matrices used to transform the input features into queries, keys, and values, respectively. $J_{j}^{temp}$ represents the joint features for the $j$ -th joint across all time steps, while $J_{t}^{temp}$ represents the joint features for the $t$ -th time step across all joints.", + "bbox": [ + 511, + 474, + 906, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Triplane Feature Encoding", + "text_level": 1, + "bbox": [ + 511, + 561, + 756, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Joint features contain temporal context of the current hand pose, which can be integrated into visual features to construct low-resolution robust visual representations. However, it is difficult to fuse joint features with visual features that are not aligned in semantic content and representation form. A straightforward approach is to treat joint features as global tokens and interact them with visual features through cross attention. However, visual features struggle to effectively perceive the 3D spatial properties of joint features through attention mechanisms. Thus, we propose to use triplane features as an intermediate representation to enhance the visual features. Unlike [4, 42, 46] that constructs dense Triplane features from images for spatial indexing, we leverage sparse joint features to generate triplane features, embedding 3D geometric attributes of the skeleton into specific spatial context.", + "bbox": [ + 511, + 583, + 906, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first copy each pose token to the corresponding finger joints as a Pose Template $J^{P} \\in \\mathbb{R}^{J \\times C}$ , and use it to fuse dynamic joint features into a static one. Then splat it onto three orthogonal planes as an intermediate representation for interaction with visual features. Fur", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22630", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "thermore, we integrate the information from different planes into a single plane through spatial-aware fusion, allowing for direct fusion with the visual features.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Triplane Projection of Joint Features Visual features represent specific spatial information in the form of dense representations. Transferring joint features into the 2D visual feature space through 3D-2D pose projection [40] can provide effective disambiguation cues for hand poses. However, a single projection plane cannot capture the complete 3D hand structure implied in joint features, resulting in the loss of crucial structural disambiguation cues. Therefore, we propose to project joint features onto three orthogonal planes to fully preserve 3D hand structure information. We first decode the 3D hand pose from the fused joint features, and then use it to generate distance-aware heatmaps corresponding to the x-y, y-z, and z-x planes. Then, we splat the fused joint feature across the three planes through these heatmaps, and collect the features from each plane as the Triplane Feature maps $F_{\\text{triplane}} = \\{F_{x,y}, F_{x,z}, F_{y,z}\\}$ .", + "bbox": [ + 89, + 148, + 483, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial-Aware Fusion Triplane encoding transforms the fused joint feature into dense 2D features, eliminating the misalignment in representation form with visual features. However, the initial plane features derived from projecting sparse joint information primarily encode geometric positions while overlooking the relative spatial relationships and interactions between neighboring joints, such as occlusions or depth variations, which are crucial for resolving structural ambiguities in visual features. Additionally, these plane features are spatially non-unified: The x-y plane shares the same spatial dimension with visual features, while the y-z and x-z planes are spatially orthogonal to visual features. To embed the complete 3D hand structure information represented by the Triplane features into the visual features in 2D space, it is essential to compress the y-z and x-z plane features as supplementary depth information relative to the x-y plane into the pixel-wise semantic content. However, directly integrating the y-z and x-z plane features into the x-y plane via cross-attention is inefficient and introduces redundant information. Therefore, we compress the y-z and x-z features along distinct directions into 1D global representations. This compression allows us to guide each pixel in the x-y plane to selectively integrate spatial context from the other two planes, reducing unnecessary overlap while preserving key spatial dependencies.", + "bbox": [ + 89, + 402, + 483, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first add 2D positional encoding to each plane feature and apply convolution layers to each plane feature to capture local spatial information (See Fig. 2 (c) project). Next, we perform summation along different directions for $F_{x,z}$ and $F_{y,z}$ to obtain the 1D features $I_x, V_x$ and $I_y, V_y$ (See Fig. 2 (c) compress), where $I$ denotes summation along the Z direction, and $V$ denotes summation along the X or Y direction. Finally, we integrate these", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "features into the pixel-wise content of $F_{x,y}$ through crossattention to obtain spatially compressed features.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Optimization", + "text_level": 1, + "bbox": [ + 511, + 128, + 653, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We supervise hand mesh regression through reconstruct loss, supervise temporal enhancement process through token pose loss and temporal pose loss.", + "bbox": [ + 511, + 150, + 905, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1. Reconstruct Loss", + "text_level": 1, + "bbox": [ + 511, + 200, + 679, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We used the same reconstruction loss as [53], including vertex/joint loss and mesh smoothing loss. Vertex/Joint loss supervises the prediction error of each mesh and joint point through L1 loss, i.e", + "bbox": [ + 511, + 219, + 905, + 280 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {V / J} = \\sum_ {i = 1} ^ {N} \\| \\nu_ {i} - \\nu_ {i} ^ {*} \\| _ {1} + \\sum_ {i = 1} ^ {K} \\| j _ {i} - j ^ {*} \\| _ {1} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 285, + 905, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nu_{i}$ is the $i^{th}$ vertex of the mesh, $j_{i}$ is the $i^{th}$ joint of the pose, and $*$ denotes the ground-truth.", + "bbox": [ + 511, + 332, + 905, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mesh smooth loss includes supervision of the surface normal vectors and edge lengths to obtain visually reasonable results. The loss functions for normal vector and edge length are defined as follows:", + "bbox": [ + 511, + 363, + 905, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {n o r m a l}} = \\sum_ {f} \\sum_ {\\{i, j \\} \\subset f} | \\left\\langle \\frac {\\nu_ {i} - \\nu_ {j}}{\\| \\nu_ {i} - \\nu_ {j} \\| _ {2}}, n _ {f} ^ {*} \\right\\rangle | \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 429, + 903, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e d g e} = \\sum_ {f} \\sum_ {\\{i, j \\} \\subset f} \\left| \\left\\| \\nu_ {i} - \\nu_ {j} \\right\\| _ {2} - \\left\\| \\nu_ {i} ^ {*} - \\nu_ {j} ^ {*} \\right\\| _ {2} \\right| \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 470, + 903, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f$ and $n_f$ indicate a mesh face and unit normal vector of face $f$ , respectively. $n_f^*$ is computed from ground-truth vertex sets $V^*$ , where * denotes ground-truth.", + "bbox": [ + 511, + 507, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final hand reconstruction task loss combines the vertex/joint loss and the mesh smoothness loss as $L_{hand}$ .", + "bbox": [ + 511, + 553, + 905, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. Pose Loss", + "text_level": 1, + "bbox": [ + 511, + 588, + 627, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For pose-aware tokens, we enable them to perceive pose changes by supervising their predicted 2D poses. During the temporal fusion process, we use the fused joint feature to predict 2.5D hand pose coordinates for triplane project. We supervise the predicted 2.5D coordinates at each temporal fusion stage to ensure accurate decoding of hand structure information. These two loss functions are represented as follows:", + "bbox": [ + 511, + 606, + 905, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t o k e n}} = \\sum_ {i} ^ {K} \\| j _ {i} ^ {2 d} - j ^ {2 d *} \\| _ {1} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 724, + 903, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t e m p o r a l}} = \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {K} \\| j _ {t, i} ^ {2. 5 d} - j ^ {2. 5 d *} \\| _ {1} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 768, + 903, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final training loss of our framework, which combines reconstruction loss and pose loss, can be expressed as follows:", + "bbox": [ + 511, + 811, + 905, + 856 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {h a n d}} + L _ {\\text {t o k e n}} + L _ {\\text {t e m p o r a l}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 864, + 903, + 880 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22631", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b782c227021313d6f0e6c1c8fe95d983ac9a1b83b56261a7f8a020e464af09f9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsInputResolutionPA-MPJPE(↓)PA-MPVPE(↓)AUC(↑)F@5(↑)F@15(↑)
METRO[26]Monocular2566.99----
Liu et. al.[28]Monocular2566.60----
MeshGraphormer[27]Monocular2566.41----
HandOccNet[34]Monocular2565.805.5088.478.099.0
H2ONet[47]Monocular1285.705.5089.180.199.0
SimpleHand[57]Monocular2565.505.50---
VIBE[22]Sequence2246.43-87.1--
TCMR[9]Sequence2246.28-87.5--
Deformer[12]Sequence5125.22-89.6--
HMR[19] + LR-ViTMonocular646.456.2487.373.098.5
VIBE[22] + LR-ViTSequence646.226.0987.574.198.7
OursSequence645.725.5288.578.498.7
OursSequence326.035.8288.076.598.6
", + "bbox": [ + 99, + 87, + 901, + 305 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9eaaa7df928d1037c1fb5d554de23bd881def84edc1d5f5b7b5a3af7e3315e3d.jpg", + "table_caption": [ + "Table 1. Comparison with the state-of-the-art on the DexYCB dataset." + ], + "table_footnote": [], + "table_body": "
MethodsInputMPJPEPA-MPJPE
Cho et al.[8]Sequence17.20-
HandOccNetMonocular16.196.13
HMR + LR-ViTMonocular30.727.49
VIBE + LR-ViTSequence27.897.05
OursSequence18.156.11
", + "bbox": [ + 109, + 353, + 464, + 449 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/47bbca064ee7cbe27132a949c17e1ea7d76a177f62fcfd27ba08772405350471.jpg", + "table_caption": [ + "Table 2. Comparison with the state-of-the-art on the H2O dataset." + ], + "table_footnote": [], + "table_body": "
MethodsInputJ-PAV-PAF@5F@15
HaMuCo[54]Monocular7.00---
EpipolarPose[21]Monocular6.10---
MobRecon[7]Monocular5.705.81--
HaMeR[35]Monocular5.995.6177.899.0
HMR + LR-ViTMonocular6.886.4473.498.7
VIBE + LR-ViTSequence6.796.2775.098.8
OursSequence5.775.7477.699.0
", + "bbox": [ + 94, + 488, + 480, + 614 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ed2bfbe8414d1ad597966cbe72144970b4cb4e901ff977d105f3379a64cec271.jpg", + "table_caption": [ + "Table 3. Comparison results on the HanCo dataset. The evaluation results for HaMuCo and EpipolarPose using monocular input are derived from [54]." + ], + "table_footnote": [], + "table_body": "
RetrainedHanOccNetSimpleHandHaMeROurs*
×6.63-8.70-
✓5.985.92-5.72
", + "bbox": [ + 96, + 680, + 480, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. SOTA performance (P-MPJPE) on interpolated low-resolution images (DexYCB). *Using raw low-resolution inputs.", + "bbox": [ + 89, + 737, + 482, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 792, + 223, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 816, + 192, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DexYCB contains multi-camera videos of humans grasping 20 YCB [3] objects. It consists of 1000 sequences with 582K frames captured by cameras from 8 different views. During the data collection, the hands are frequently oc", + "bbox": [ + 89, + 839, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cluded during hand-object interaction. We use the dataset split setup s0 from [5] for the training and test sets.", + "bbox": [ + 511, + 356, + 906, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "HanCo consists of 1,517 videos with multiple views and camera calibration. It has 860,304 frames in total, i.e. 107,538 time-step per view. Since HanCo does not have an official train/test split, we follow [54] use the first 1,200 sequences for training and validation and the last 317 sequences for testing in all experiments for fair comparisons. We conduct all ablation studies on this dataset.", + "bbox": [ + 511, + 388, + 906, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "H2O includes data from four participants performing 36 distinct actions involving interactions with eight different objects. It provides markerless 3D annotations for both hands. We follow the setup in [25] to divide the training and test sets for pose estimation.", + "bbox": [ + 511, + 496, + 906, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation", + "text_level": 1, + "bbox": [ + 511, + 582, + 671, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Handling Low Resolution Our experiment was conducted under two original resolution settings, $32 \\times 32$ and $64 \\times 64$ . Unless otherwise specified, all experiments use the original resolution of $64 \\times 64$ . To avoid additional noise introduced by excessive image interpolation, we set the size of the input image to $64 \\times 64$ . We have modified ViT pose-B [49] as the backbone. To adapt it to low resolution inputs, we replaced the patch embedding part with two convolution layers with zero padding, and reduce the number of tokens in ViT to 64. These 64 tokens were rearranged into an $8 \\times 8$ feature map during the fusion of temporal features and regression of hand mesh. We adopt a regression-based method in [53] to reconstruct hand from the final feature.", + "bbox": [ + 511, + 604, + 906, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Handling Temporal Input Our method utilizes the temporal information of historical frames to enhance the current hand reconstruction. We have chosen a sequence length of $T = 6$ with a 5-frames interval based on observations of each datasets. In training, we first extract the visual features and pose-aware tokens of the first 5 frames through the ViT", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22632", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "backbone without temporal fusion, then compressed them into joint features. During this phase, the parameters of the ViT backbone are not updated. These joint features are subsequently processed through spatio-temporal Transformer to enhance the feature extraction process of last frame. We train all network parameters in this phrase.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training We train and evaluate our method on a single server with an Nvidia GeForce RTX 4090 GPU. All experiments are implemented with PyTorch. For all datasets, the batch size is set to 64. Before formal training on each dataset, we fine-tune the backbone using the pre-trained weights of ViTPose-B. We conducted 30 epochs of fine-tuning for DexYCB and HanCo, and 20 epochs for H2O. For DexYCB ,H2O and HanCo, the model is trained for 50,30,50 epochs,respectively. The model is optimized with AdamW solver with parameters $(\\beta_{1}$ of $0.9,\\beta_{2}$ of 0.999,and weight decay of 0.1). We use an initial learning rate of $5e - 4$ and cosine schedule. We perform data augmentation including scaling $(\\pm 10\\%)$ , rotation $(\\pm 180^{\\circ})$ , translation $(\\pm 10$ pixels), and color jittering $(\\pm 20\\%)$ .", + "bbox": [ + 89, + 188, + 482, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 407, + 274, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employ the standard metric for hand pose estimation: Mean Per Joint Position Error (MPJPE, J-PE) and Mean Per Vertex Position Error (MPVPE, V-PE), measured after Procrustes alignment, to measure the accuracy of the pose and shape of the estimated hand meshes. In addition, we report the area under the curve of PCK (AUC) score and F-score at 5mm (F@5) and 15mm (F@15) threshold. All evaluation metrics are calculated in the root-aligned space.", + "bbox": [ + 89, + 429, + 482, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparisons with State-of-the-arts", + "text_level": 1, + "bbox": [ + 89, + 558, + 397, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compared the proposed low-resolution method with state-of-the-art video-based and single-view methods that use high-resolution input. Additionally, we further use our backbone network (LR-ViT) to establish a low-resolution benchmark for comparison. We trained HMR and VIBE with our backbone on raw low-resolution images to explore the effectiveness of commonly used methods. Comparison results are shown in Tab. 1, Tab. 2 and Tab. 3. In DexYCB, our method performs equally or better than other high-resolution methods at $32 \\times 32$ and $64 \\times 64$ resolution settings, except for Deformer that uses $512 \\times 512$ input resolution. In HanCo and H2O, our method also demonstrates competitive results. We also report the performance of SOTA methods on upsampled low-resolution images in Tab. 4, including cases where models were retrained with upsampled images and where they were not.", + "bbox": [ + 89, + 580, + 482, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 829, + 243, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.1. Backbone settings", + "text_level": 1, + "bbox": [ + 89, + 851, + 261, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we experiment with different backbone settings and report their performance with and without tem", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5d2bb487c5dfad6c6c2900818a0fbe4cc0ac7efb74ff1f7c90feed8d53e890ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Token SettingsMonocularTemporal
J-PAV-PAJ-PAV-PA
06.716.75--
16.586.646.036.09
6 (SP)6.196.195.775.74
21 (JP)6.076.055.805.82
", + "bbox": [ + 570, + 88, + 851, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation study on backbone and pose-aware tokens.", + "bbox": [ + 527, + 194, + 888, + 208 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "poral enhancement. Under monocular input, the backbone skips temporal fusion and uses only transformer blocks to process the input. Our experiment further explored the impact of Pose-Aware Tokens. We adopted different settings of introduced tokens to test their impact. Under monocular input, we still supervise these tokens to predict 2D hand poses. The experiment result is shown in Tab. 5. The introduction of temporal enhancement universally improves performance. Under monocular input, the original ViT (w/o pose-aware tokens) performs poorly at low resolutions, while other networks generally achieve performance gains. This demonstrates that supervising intermediate features through specific tokens is more effective than solely employing end-to-end training on the final regression results. Meanwhile, due to the ability of finger based pose-aware tokens (SP) to jointly carry high-level semantic information and low-level visual details, it is more effective in temporal enhancement than a single pose token and skeleton based pose tokens (JP).", + "bbox": [ + 511, + 236, + 903, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.2. Temporal Fusion Schemes", + "text_level": 1, + "bbox": [ + 511, + 534, + 741, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we explored the effects of different temporal fusion schemes. We consider three ways to fuse temporal features and visual features: cross attention, single plane encoding, and triplane encoding. The cross attention approach uses joint features as K and V, and visual features as Q to interact through attention mechanisms. The single plane encoding method uses 2D Heatmap to diffuse joint features into a single plane. The single plane feature is also embedded with 2D positional encoding and processed using convolutional layers. For the triplane encoding, we also tested the effect of directly integrating the full $F_{x,z}$ and $F_{y,z}$ with $F_{x,y}$ . The comparison results are presented in Table 6. Our proposed triplane encoding performs best. Compared to the other two methods, the triplane encoding effectively extracts crucial 3D structural information, providing cues for eliminating hand structure ambiguity in visual features. Additionally, using compressed 1D features for interaction is superior to using complete feature.", + "bbox": [ + 511, + 553, + 906, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.3. Position for inserting temporal fusion", + "text_level": 1, + "bbox": [ + 511, + 835, + 820, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we further investigated the impact of the insertion position of temporal fusion operations in the backbone. Specifically, we compared two strategies: one where", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22633", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea5865cdb67302b8e219683c12dbef5ea9e7459dba0c985dc77f2a85a85acd8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 87, + 187, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b611921dcfbc4b7f40c90bed1e52a22d44dd49b0bd2c45702a6947fbcdf25cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 88, + 289, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fda41ed7e1bb484c068938af830dd6f1205c512d6e973388c58557905312a097.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 88, + 390, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10dc607d8ed6d72e66823f71c67b908451e90c93b8ac578b9c438be5d4b1ab66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 88, + 493, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b10c8d38d58d159712230ae996fca18bf1622f2a9203ade9f2f1df88672b1798.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 176, + 187, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ce82f2843f4ec50aaefecb0297a08881aec626394e74a3b4a230dd1814a8c9c.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 192, + 176, + 289, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5b51293807e8934e5f16a1f81ba84ad11ec9f61667574ec59863089ccf0ed49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 176, + 390, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bac03c7ffd1ff9f485b6526cb5e46546604f5d057159966aea80065e4c4c7dfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 176, + 493, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d34a9177f37e4355a58ae2a92528356d5918089222e2e49fd46058463d674b10.jpg", + "image_caption": [ + "Figure 4. Qualitative Comparison Results on DexYCB. Regions with significant prediction errors are marked with red circles." + ], + "image_footnote": [], + "bbox": [ + 91, + 265, + 187, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3154c17d2689b5d91a116ce97aa424db4cf8406509489f2be28dc5bc99abf225.jpg", + "image_caption": [ + "(c) HMR", + "(e) Ours-Single" + ], + "image_footnote": [], + "bbox": [ + 192, + 265, + 289, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/13cdd829557a961a757a3b7c9ae2c16325e02da24cabccb059446782b1389dbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 265, + 390, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb37af80419fa58ed51b4650bcee52eefe2069c69b3ac4a77de1f9906bce3092.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 265, + 493, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80ba1a4c446113d6ec91fd3185a3af21229c628c9e1b8f7ca65b7fbcc9f10dd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 88, + 599, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f23552623d61819f8fe10e6c7ef1fbc96e963897dd1b9caf2fc1548cee18d11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 88, + 702, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a32d90e2a4d584ded64c1672f5350e9568b489dd708e305f14d298f35d50790.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 88, + 803, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd0968ebe67be0ff628f029e2c42a941dd31d9ada077a78002b77762206964e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 88, + 905, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1cae2ef4965d0ff7619dd25e629be4c82828fa627a4f7c54022d8401d9d3aff0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 176, + 599, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73625c413d28379546b685d57cbc5e83dcee2d15dcf74a382b1ef35054d85da8.jpg", + "image_caption": [ + "(b) GT" + ], + "image_footnote": [], + "bbox": [ + 604, + 176, + 702, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/188df50cd324cc163b8f058910df147172f1d483bc7b596def4fccaea5051702.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 176, + 803, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5c453b1d37eee1c358d0240da56a1583f3804d2a23ca0f52ea8a55cf40678bae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 176, + 905, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a87febfee9d113e4fe8ce382a2f9cbad5fe2b41ef6701e1e32b7bbdaabeae406.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 265, + 599, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d9a8dd3b049e5178f111bf925a3a9221f309470531c6c999cdfc6f72327409d3.jpg", + "image_caption": [ + "(d) VIBE", + "(f) Ours" + ], + "image_footnote": [], + "bbox": [ + 604, + 265, + 702, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da1e135f507fb77e750331426461cbcccf1be54cdb43735840e4738d968f5e22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 265, + 803, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f724e1dbb66a90f30e710ff9903a9db0c6ec885ceca571c6fab4df09c3ed9ed0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 265, + 905, + 339 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/68afa8b96721e0e144dc017dd15139798c77732685c3a0780c242b05b564e1aa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Fusion SchemeJ-PAV-PAF@5F@15
Cross-attention5.955.9975.898.9
Single plane Encoding5.855.9076.799.0
Triplane Encoding (w/o 1D compress)5.805.8576.999.0
Triplane Encoding5.775.7477.699.0
", + "bbox": [ + 119, + 402, + 455, + 500 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0f0158666546d943f98649c3cc9446a846409c7dec784c0a0c9a8f86d53ae0f6.jpg", + "table_caption": [ + "Table 6. Ablation study of different temporal fusion schemes." + ], + "table_footnote": [], + "table_body": "
Insert PositionJ-PAV-PAF@5F@15
Final layer6.016.0575.198.6
Middle layer5.775.7477.699.0
", + "bbox": [ + 145, + 537, + 429, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation Study on the Position of Temporal Information Fusion in the Backbone Network.", + "bbox": [ + 89, + 598, + 482, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "temporal features are fused at multiple intermediate levels throughout the network, and another where fusion occurs solely at the final level, as shown in Tab. 7. The experimental results in Tab. 7 demonstrate that fusing temporal features at multiple intermediate levels leads to superior performance compared to only performing fusion at the final level. This suggests that incorporating temporal information at various stages of the network helps to better capture ambiguous local features, contributing to more accurate and consistent hand pose reconstruction in low resolution.", + "bbox": [ + 89, + 654, + 483, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 816, + 272, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the qualitative comparison results on DexYCB in Fig. 4. Compared with low resolution benchmarks (HMR, VIBE) that use the same LR-ViT as the backbone, our method achieves better alignment of the reconstructed", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "hands with low-resolution images. Additionally, by leveraging historic temporal information to supplement missing details in visual features, our approach enhances the ability to cope with local occlusions compare to solely rely on a single frame (see columns 2 and 3 in Fig. 4 (e) and (f)).", + "bbox": [ + 511, + 406, + 906, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 494, + 633, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a temporal enhancement framework for low-resolution hand reconstruction. We represent temporal information through joint features to capture the spatiotemporal correlations between joints. Furthermore, we integrate the temporal context provided by joint features with the spatial context provided by visual features to construct a robust low-resolution representation with 3D spatial awareness. Specifically, we introduce a Triplane feature representation to address the misalignment in semantics and representation forms between joint and visual features, preserving 3D structural information during fusion. Quantitative experiments on multiple datasets have shown that our method achieves comparable performance to high-resolution SOTA methods at much lower resolutions. We also demonstrated the contribution of each design to the scheme performance through ablation study.", + "bbox": [ + 511, + 520, + 906, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "**Limitation and Future Work** Our work did not explicitly model long-term hand motion cues across video sequences, which may result in a lack of constraints on motion consistency. As a result, the model may struggle with maintaining consistent hand shapes over extended sequences. Future work should explore utilizing motion cues from long sequences to better constrain the reconstructed hand shape and improve temporal consistency.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22634", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported by the National Natural Science Foundation of China under Grants (62406039, 62101064, 62171057,62201072, U23B2001, 62001054, 62071067), the Ministry of Education and China Mobile Joint Fund (MCM20200202, MCM20180101), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center, BUPT Excellent Ph.D. Students Foundation (CX20241014), in part by the Project funded by China Postdoctoral Science Foundation (2023TQ0039, 2024M750257), the Postdoctoral Fellowship Program of CPSF (GZC20230320).", + "bbox": [ + 89, + 90, + 485, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 270, + 187, + 286 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10843-10852, 2019. 2", + "[2] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 666-682, 2018. 2", + "[3] Berk Calli, Arjun Singh, Aaron Walsman, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In 2015 international conference on advanced robotics (ICAR), pages 510-517. IEEE, 2015. 6", + "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 4", + "[5] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 6", + "[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12929–12938, 2021. 1", + "[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20544-20554, 2022. 1, 6", + "[8] Hoseong Cho, Chanwoo Kim, Jihyeon Kim, Seongyeong Lee, Elkhan Ismayilzada, and Seungryul Baek. Transformer-based unified recognition of two hands manipulating objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4769-4778, 2023. 6", + "[9] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consis" + ], + "bbox": [ + 99, + 295, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tent 3d human pose and shape from a video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1964-1973, 2021. 3, 6", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3", + "[11] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 3", + "[12] Qichen Fu, Xingyu Liu, Ran Xu, Juan Carlos Niebles, and Kris M Kitani. Deformer: Dynamic fusion transformer for robust hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23600-23611, 2023. 3, 6", + "[13] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10833-10842, 2019. 2", + "[14] Peter Hardy, Srinandan Dasmahapatra, and Hansung Kim. Can super resolution be used to improve human pose estimation in low resolution scenarios? arXiv preprint arXiv:2107.02108, 2021. 2, 3", + "[15] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 571-580, 2020. 3", + "[16] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3", + "[17] Weiting Huang, Pengfei Ren, Jingyu Wang, Qi Qi, and Haifeng Sun. Awr: Adaptive weighting regression for 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11061-11068, 2020. 2", + "[18] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In Proceedings of the European conference on computer vision (ECCV), pages 118-134, 2018. 2", + "[19] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018. 6", + "[20] Leyla Khaleghi, Alireza Sepas-Moghaddam, Josh Marshall, and Ali Etemad. Multiview video-based 3-d hand pose estimation. IEEE Transactions on Artificial Intelligence, 4:896-909, 2021. 3", + "[21] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF conference on" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22635", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "computer vision and pattern recognition, pages 1077-1086, 2019. 6", + "[22] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 3, 6", + "[23] Deying Kong, Linguang Zhang, Liangjian Chen, Haoyu Ma, Xiangyi Yan, Shanlin Sun, Xingwei Liu, Kun Han, and Xiaohui Xie. Identity-aware hand mesh estimation and personalization from rgb images. In European Conference on Computer Vision, pages 536-553. Springer, 2022. 2", + "[24] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4990-5000, 2020. 2", + "[25] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 6", + "[26] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 2, 6", + "[27] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12939-12948, 2021. 2, 6", + "[28] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 6", + "[29] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 3", + "[30] Xingyu Liu, Pengfei Ren, Yuanyuan Gao, Jingyu Wang, Haifeng Sun, Qi Qi, Zirui Zhuang, and Jianxin Liao. Keypoint fusion for rgb-d based 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3756-3764, 2024. 1", + "[31] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 752-768. Springer, 2020. 1", + "[32] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2308-2317, 2022. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Chihiro Nakatsuka and Satoshi Komorita. Stable 3d human pose estimation in low-resolution videos with a few views. In 2022 International Conference on Robotics and Automation (ICRA), pages 427-433. IEEE, 2022. 3", + "[34] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 6", + "[35] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 6", + "[36] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. Pose-guided hierarchical graph reasoning for 3-d hand pose estimation from a single depth image. IEEE Transactions on Cybernetics, 53(1):315-328, 2021. 2", + "[37] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. A dual-branch self-boosting framework for self-supervised 3d hand pose estimation. IEEE Transactions on Image Processing, 31:5052–5066, 2022. 2", + "[38] Pengfei Ren, Haifeng Sun, Jiachang Hao, Jingyu Wang, Qi Qi, and Jianxin Liao. Mining multi-view information: a strong self-supervised framework for depth-based 3d hand pose and mesh estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20555-20565, 2022.", + "[39] Pengfei Ren, Yuchen Chen, Jiachang Hao, Haifeng Sun, Qi Qi, Jingyu Wang, and Jianxin Liao. Two heads are better than one: Image-point cloud network for depth-based 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2163-2171, 2023. 2", + "[40] Pengfei Ren, Chao Wen, Xiaozheng Zheng, Zhou Xue, Haifeng Sun, Qi Qi, Jingyu Wang, and Jianxin Liao. Decoupled iterative refinement framework for interacting hands reconstruction from a single rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8014-8025, 2023. 1, 5", + "[41] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2", + "[42] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 4", + "[43] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 2", + "[44] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11698-11707, 2021. 2", + "[45] Chen Wang, Feng Zhang, Xiatian Zhu, and Shuzhi Sam Ge." + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22636", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Low-resolution human pose estimation. Pattern Recognition, 126:108579, 2022. 2, 3", + "[46] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4", + "[47] Hao Xu, Tianyu Wang, Xiao Tang, and Chi-Wing Fu. H2onet: Hand-occlusion-and-orientation-aware network for real-time 3d hand mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17048–17058, 2023. 6", + "[48] Xiangyu Xu, Hao Chen, Francesc Moreno-Noguer, Laszlo A Jeni, and Fernando De la Torre. 3d human pose, shape and texture from low-resolution images and videos. IEEE transactions on pattern analysis and machine intelligence, 44(9): 4490-4504, 2021. 2, 3", + "[49] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022. 6", + "[50] John Yang, Hyung Jin Chang, Seungeui Lee, and Nojun Kwak. Seqhand: Rgb-sequence-based 3d hand pose and shape estimation. In European Conference on Computer Vision, pages 122-139. Springer, 2020. 3", + "[51] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13232-13242, 2022. 2", + "[52] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 2", + "[53] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Sar: Spatial-aware regression for 3d hand pose and mesh reconstruction from a monocular rgb image. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 99-108. IEEE, 2021. 2, 3, 5, 6", + "[54] Xiaozheng Zheng, Chao Wen, Zhou Xue, Pengfei Ren, and Jingyu Wang. Hamuco: Hand pose estimation via multiview collaborative self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20763-20773, 2023. 6", + "[55] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3", + "[56] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 2", + "[57] Zhishan Zhou, Shihao Zhou, Zhi Lv, Minqiang Zou, Yao Tang, and Jiajun Liang. A simple baseline for efficient hand mesh reconstruction. In Proceedings of the IEEE/CVF Con-" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 1367-1376, 2024. 6", + "[58] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15085-15099, 2023. 2", + "[59] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 2", + "[60] Christian Zimmermann, Max Argus, and Thomas Brox. Contrastive representation learning for hand shape estimation. In Pattern Recognition, pages 250-264, Cham, 2021. Springer International Publishing. 2" + ], + "bbox": [ + 516, + 92, + 906, + 304 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22637", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/61fa7d4b-03fc-4f0b-b84c-af353e8bd6fc_model.json b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/61fa7d4b-03fc-4f0b-b84c-af353e8bd6fc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..749762f2fb0a997b467604a2052f05014ff5f972 --- /dev/null +++ b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/61fa7d4b-03fc-4f0b-b84c-af353e8bd6fc_model.json @@ -0,0 +1,2686 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.858, + 0.175 + ], + "angle": 0, + "content": "Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.204, + 0.846, + 0.257 + ], + "angle": 0, + "content": "Kaixin Fan, Pengfei Ren\\*, Jingyu Wang,\\(^{\\dagger}\\) Haifeng Sun, Qi Qi, Zirui Zhuang, Jianxin Liao \nState Key Laboratory of Networking and Switching Technology, \nBeijing University of Posts and Telecommunications" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.26, + 0.832, + 0.275 + ], + "angle": 0, + "content": "{fkxbupt, rpf, wangjingyu, hfsun, qiqi8266, zhuangzirui, liaojx}@bupt.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.31, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.342, + 0.485, + 0.78 + ], + "angle": 0, + "content": "3D hand reconstruction is essential in non-contact human-computer interaction applications, but existing methods struggle with low-resolution images, which occur in slightly distant interactive scenes. Leveraging temporal information can mitigate the limitations of individual low-resolution images that lack detailed appearance information, enhancing the robustness and accuracy of hand reconstruction. Existing temporal methods typically use joint features to represent temporal information, avoiding interference from redundant background information. However, joint features excessively disregard the spatial context of visual features, limiting hand reconstruction accuracy. We propose to integrate temporal joint features with visual features to construct a robust low-resolution visual representation. We adopt Triplane Features, a dense representation with 3D spatial awareness, to bridge the gap between the joint features and visual features that are misaligned in terms of representation form and semantics. Triplane Features are obtained by orthogonally projecting joint features, embedding hand structure information into the 3D spatial context. Furthermore, we compress the spatial information of the three planes into a 2D dense feature though Spatial-Aware Fusion to enhance the visual features. By using enhanced visual features enriched with temporal information for hand reconstruction, our method achieves competitive performance at much lower resolutions compared to state-of-the-art methods operating at high resolution on DexYCB, HanCo and H2O. Code is available at https://github.com/NewbieFan/Temp-LowRes-hand." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.809, + 0.222, + 0.824 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.835, + 0.484, + 0.866 + ], + "angle": 0, + "content": "3D hand reconstruction, a crucial component in non-contact human-computer interaction applications, has made signifi-" + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.313, + 0.612, + 0.323 + ], + "angle": 0, + "content": "(a) No Interaction" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.325, + 0.905, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.382, + 0.615, + 0.391 + ], + "angle": 0, + "content": "(b) Cross Attention" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.393, + 0.904, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.492, + 0.617, + 0.502 + ], + "angle": 0, + "content": "(c) Triplane Feature" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.502, + 0.904, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.619, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Figure 1. Different way to utilizing joint features. a) Directly reconstruct hand from joint features, relying solely on sparse features without spatial context. b) Fusing joint and visual features through cross-attention, which avoids representation discrepancies but neglects the hand's structural priors. c) Using triplane features to bridge joint and visual features by decoding structural joint information into dense representations that enhance visual features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "cant advancements in recent years with the advent of large-scale datasets and deep learning technologies. Many studies [6, 7, 30, 31, 34, 40] have focused on reconstructing 3D hands from a single RGB image. In typical interaction scenarios, these methods require the subject to stay close to the capture camera to obtain high-resolution (e.g., \\(256 \\times 256\\)) cropped hand images. However, as the interaction distance increases—such as in XR teaching activities within a classroom—the available cropped hand images are limited to lower resolutions (e.g., \\(64 \\times 64\\)). Current methods" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22627" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "perform poorly on such low-resolution images, as these images provide fewer cues for addressing joint self-similarity and depth ambiguity, significantly increasing the difficulty of 3D hand reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.155, + 0.483, + 0.352 + ], + "angle": 0, + "content": "To address the low-resolution problem, one intuitive approach is to lift the resolution through super-resolution techniques [14]. However, general super-resolution methods can lead to structural distortions and artifacts in the hand regions, which brings a domain gap between super-resolved image and real high-resolution image. Low-Res Pose [45] adopted confidence-aware learning to supervise continue joint offset fields, thereby alleviating the quantization errors. RSC-Net [48] adopted cross-resolution contrastive learning to align the estimation results of low-resolution images with high-resolution images. However, the performance of these methods is still limited due to the inherent ambiguity resulting from information loss in low resolution." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.355, + 0.485, + 0.687 + ], + "angle": 0, + "content": "A single low-resolution image lacks detailed hand appearance, leading to ambiguous visual representation. Continuous frames provide comprehensive hand appearance and motion context, offering strong disambiguation cues for inferring global hand structure. To reduce interference from redundant background information and computational costs, existing temporal methods typically compress single-frame information into joint features [51, 58], and then model the intra- and inter-frame correlation between joints through temporal interaction. However, joint features excessively discard spatial context of visual features, which limits the accuracy of pose estimation. Integrating the spatial context of visual features with the temporal context of joint features can construct robust visual representation with 3D spatial awareness. However, integrating joint features into visual features present two major challenges. Firstly, sparse joint features are not aligned with dense visual features in terms of representation form. Secondly, joint features contain prior geometric structural information in 3D space (e.g., hand scale, skeleton length), while visual features represent detailed appearance information in 2D space (e.g., texture, color), leading to a semantic misalignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we integrate the temporal clue from historical frames into the current visual features to achieve robust and accurate 3D hand pose estimation under low-resolution scenarios. Specifically, we compress temporal information into joint representations, and model intra- and interframe relationships of joints through spatio-temporal interaction. Then, we adopt triplane features[4], a 3D spatial-aware dense visual representation, to bridge the sparse 3D joint representation and the dense 2D visual representation. We orthogonally project the joint features into three individual planes, with each plane independently embedding the sparse joint features into specific spatial contexts. The orthogonal planes ensure comprehensive coverage of 3D hand structure information in a low-redundancy manner. How" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "ever, they still contain redundant spatial information, directly integrating these features into the visual plane leads to the overlap of spatial context from different planes. Thus, we compress the features from two planes orthogonal to the visual plane into 1D global features, guiding each pixel in visual features to integrate distinct spatial context from the other planes, reducing redundancy and clarifying the 3D spatial context. Finally, we reconstruct the hand based on the enhanced visual features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.333 + ], + "angle": 0, + "content": "Experimental results on several large-scale hand pose estimation video datasets, DexYCB [5], HanCo [60] and H2O [25], show the proposed method reach competitive performance at much lower resolution compared to both single-frame based and temporal based state-of-the-art (SOTA) methods at high resolution. In summary, our contribution is list as below:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.338, + 0.904, + 0.412 + ], + "angle": 0, + "content": "- We propose a temporal enhancement framework that integrates temporal joint features with visual features to construct a robust visual representation for low-resolution conditions, thereby improving hand reconstruction performance in low-resolution videos." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.414, + 0.904, + 0.473 + ], + "angle": 0, + "content": "- We introduced triplane feature encoding to encode joint features into dense feature maps. Through triplane feature, visual features can perceive the prior hand structure information from the temporal joint features." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.474, + 0.904, + 0.504 + ], + "angle": 0, + "content": "- Our method reaches competitive result at much lower resolution compared to SOTA methods at high resolution." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.338, + 0.904, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.52, + 0.644, + 0.536 + ], + "angle": 0, + "content": "2. Relate Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.546, + 0.871, + 0.562 + ], + "angle": 0, + "content": "2.1. Single Image based Hand Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Early works [2, 17, 18, 36, 43, 59] focused on estimating hand pose solely from input images, which suffered from occlusion issues due to the lack of modeling the inherent structure of the hand. MANO [41] introduces a parametric hand model, which contains rich priors of human hands, and now most work focuses on reconstruct this pre-defined hand from image. Some works [1, 23, 32, 37-39, 52, 56] directly predict parameters of MANO, 3D joints and mesh vertices coordinates are then retrieved using the differentiable MANO model. However, predicting MANO parameters directly from a single RGB image is a challenging job, which is difficult for network optimization and is prone to mesh-image misalignment [44], especially in a low-res input condition. To alleviate these issues present in parametric predicting, some works [13, 24, 27] regress the 3D coordinates or occupancy directly for mesh vertices. For example, [53] adopts graph convolution neural network to regress 2.5D Heatmap for every mesh vertices. [26] proposed an End to end framework to recover mesh using transformer. These works show that using a regression-based approach can maintain flexibility while achieving comparable results to parametric methods. Thus, we adopt a regression-based" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22628" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "method described in [53] to reconstruct the hand mesh from enhanced visual features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.133, + 0.376, + 0.149 + ], + "angle": 0, + "content": "2.2. Low Resolution Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.156, + 0.485, + 0.504 + ], + "angle": 0, + "content": "Recently, a few works have explored on low resolution pose estimation. A intuitive way is leveraging Super Resolution approach [14] to improve the image quality and pose estimation performance. However, the addition super-resolution network brings extra computation cost and may bring unpleasant artifacts to the input image. [45] find that offset-based methods can provide more reliable results in low resolution, and further proposed a confidence-aware learning to selectively weigh the learning of heatmap and offset. [48] controls the scaling of the CNN architecture model using a set of hyperparameters, enforcing it to output scale-invariant features at different resolutions through contrastive learning and self-supervised learning. Although specific learning schemas improve the poor results in low resolution, the features extracted by the network are still constrained by the lack of detailed hand appearance. Other work [33] explored on 3D pose estimation in low-resolution multi-view videos, showing that temporal and multi-view information could be a supplement to alleviate information loss. However, they only correct the prediction under certain adverse conditions, ignoring temporal information can reduce the inherent ambiguity in single low-resolution image and imporve the overall performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.515, + 0.377, + 0.531 + ], + "angle": 0, + "content": "2.3. Temporal Hand Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Some recent methods [11, 15, 22, 29, 50, 55] have begun to explore leveraging temporal information in videos to overcome the limitations of single-view images. Hasson et al. [15] use photometric supervision to keep consistency between the re-projected 3D hand and the optical flow. VIBE [22] use a temporal generation network to predict model parameters for each frame, and employs a recurrent network as a discriminator to supervise motion sequence. However, they ignored the spatial and feature correlations in the video. More recently, [20] has harnessed temporal information by propagating frame-wise features with a sequence-to-sequence model based on LSTM [16], but it only provides sparse joint location predictions. TCMR [9] introduces a temporally consistent system that synthesizes implicit features from past and future frames to maintain temporal consistency in the mesh of the current frame. Deformer [12] deforms the hand poses from neighborhood frames using predicted motion to avoid unreliable queries generated with local feature when leveraging self-attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.903 + ], + "angle": 0, + "content": "In low resolutions, these methods are limited to the poor quality of single frame features. Additionally, they lack a focus center when processing temporal features, which may cause some estimation results to deviate from the actual image due to the influence of temporal context." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.606, + 0.107 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.312 + ], + "angle": 0, + "content": "The overview of our method is shown in the Fig. 2. For a given frame \\( I_{T} \\) in an image sequence \\( V = \\{I_{t}\\}_{t=1}^{T} \\), we extract temporal cues from the previous \\( T - 1 \\) frames to enhance the visual feature of the current frame and reconstruct the hand mesh from the enhanced visual feature. Specifically, we first compress the information of each historical frames into joint features and establish their inter- and intra-frame correlations through a spatio-temporal Transformer. Then, we encode the structural information from the enhanced joint features into 3D spatial-aware dense representations through triplane feature encoding to enhance the visual features of the current frame. Finally, we leverage the enhanced visual feature to reconstruct the hand mesh." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.32, + 0.719, + 0.336 + ], + "angle": 0, + "content": "3.1. Static Visual Features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.909, + 0.644 + ], + "angle": 0, + "content": "As shown in Fig. 3, we adopt a Vision Transformer (ViT) [10] to extract visual features. During the extraction process, ambiguous local features would propagate from low-level features to subsequent high-level features, leading to ambiguities in both semantic content and fine-grained details in the final visual features. Thus, we incorporate temporal information into intermediate features at different levels to gradually reduce ambiguous local representations and ultimately obtain visual features with a clear perception of hand structure. However, static visual features lack the ability to perceive dynamic temporal information represented in joint features. Therefore, we introduce pose-aware tokens as anchors for fusing dynamic temporal information into a static representation, which are used for further interaction with visual features. These tokens capture hand pose and appearance information from the current visual feature through attention mechanism. We categorize the functionalities of these tokens based on finger semantic to ensure they incorporate both fine-grained visual details (e.g., texture) and high-level semantic content (e.g., skeletal structures)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.813, + 0.669 + ], + "angle": 0, + "content": "3.2. Dynamic Temporal Joint Features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We use compact joint features to represent dynamic temporal information from historical frames. Give a single frame, we obtain its visual feature map along with pose-aware tokens through the backbone. Each pose-aware token carries semantic information corresponding to different individual fingers. We further combine local visual features to expand each token into a complete joint feature, enabling fine-grained temporal interaction around each joints. We adopt a FC layer to predict 2D joint coordinates \\( p \\in \\mathbb{R}^{J \\times 2} \\) from pose-aware tokens, and use it to sample joint-centric local features \\( J^{sample} \\in \\mathbb{R}^{J \\times C} \\) from the visual feature map. The joint coordinates is encoded as pose feature \\( J^{pose} \\), and then add to the sampled feature \\( J^{sample} \\) to obtain the initial joint feature \\( J^{init} = J^{pose} + J^{region} \\). We concatenate the joint features from each frame in the upstream sequence" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22629" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.087, + 0.88, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.23, + 0.88, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.379, + 0.908, + 0.451 + ], + "angle": 0, + "content": "Figure 2. Overview of Our Framework. We first construct joint features from the sampled past frame features and 2D poses using a Spatial-Temporal Transformer (a), capturing the temporal dynamics and spatial relationships between joints. We then use these joint features to enhance the visual features of the current frame (b) by converting them into spatial-aware dense features through Tripalne Feature Encoding (c). In (c), we fuse different joint features (from a) using pose-aware tokens (from b), and project them into three orthogonal planes to obtain the Triplane features. These features are then compressed into a single plane through space-aware fusion." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.473, + 0.462, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.606, + 0.402, + 0.62 + ], + "angle": 0, + "content": "Figure 3. Backbone of our framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.766 + ], + "angle": 0, + "content": "to form the initial representation of temporal information \\(J^{temp} = \\text{Concat}(J_t^{init}) \\in \\mathbb{R}^{T \\times J \\times C}\\). Then, we adopt a spatio-temporal Transformer to establish inter- and intra-frame correlation between these joints. The spatio-temporal Transformer is a combination of alternately stacked temporal and spatial attention modules. In the spatial attention module, attention interactions occur among different joints within the same joint feature:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.774, + 0.483, + 0.793 + ], + "angle": 0, + "content": "\\[\nS - M H S A (J ^ {t e m p}) = M H S A \\left(Q _ {j}, K _ {j}, V _ {j}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.798, + 0.483, + 0.819 + ], + "angle": 0, + "content": "\\[\nQ _ {j} = W _ {q} J _ {j} ^ {\\text {t e m p}}, K _ {j} = W _ {k} J _ {j} ^ {\\text {t e m p}}, V _ {j} = W _ {v} J _ {j} ^ {\\text {t e m p}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.821, + 0.483, + 0.852 + ], + "angle": 0, + "content": "In the temporal attention module, attention interactions occur among the same joint within different joint features:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.86, + 0.483, + 0.878 + ], + "angle": 0, + "content": "\\[\nT - M H S A \\left(J ^ {t e m p}\\right) = M H S A \\left(Q _ {t}, K _ {t}, V _ {t}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nQ _ {t} = W _ {q} J _ {t} ^ {t e m p}, K _ {j} = W _ {k} J _ {t} ^ {t e m p}, V _ {j} = W _ {v} J _ {t} ^ {t e m p} \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Here, \\( W_{q} \\), \\( W_{k} \\), and \\( W_{v} \\) are learnable weight matrices used to transform the input features into queries, keys, and values, respectively. \\( J_{j}^{temp} \\) represents the joint features for the \\( j \\)-th joint across all time steps, while \\( J_{t}^{temp} \\) represents the joint features for the \\( t \\)-th time step across all joints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.757, + 0.579 + ], + "angle": 0, + "content": "3.3. Triplane Feature Encoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.827 + ], + "angle": 0, + "content": "Joint features contain temporal context of the current hand pose, which can be integrated into visual features to construct low-resolution robust visual representations. However, it is difficult to fuse joint features with visual features that are not aligned in semantic content and representation form. A straightforward approach is to treat joint features as global tokens and interact them with visual features through cross attention. However, visual features struggle to effectively perceive the 3D spatial properties of joint features through attention mechanisms. Thus, we propose to use triplane features as an intermediate representation to enhance the visual features. Unlike [4, 42, 46] that constructs dense Triplane features from images for spatial indexing, we leverage sparse joint features to generate triplane features, embedding 3D geometric attributes of the skeleton into specific spatial context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Specifically, we first copy each pose token to the corresponding finger joints as a Pose Template \\( J^{P} \\in \\mathbb{R}^{J \\times C} \\), and use it to fuse dynamic joint features into a static one. Then splat it onto three orthogonal planes as an intermediate representation for interaction with visual features. Fur" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22630" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "thermore, we integrate the information from different planes into a single plane through spatial-aware fusion, allowing for direct fusion with the visual features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.149, + 0.484, + 0.392 + ], + "angle": 0, + "content": "Triplane Projection of Joint Features Visual features represent specific spatial information in the form of dense representations. Transferring joint features into the 2D visual feature space through 3D-2D pose projection [40] can provide effective disambiguation cues for hand poses. However, a single projection plane cannot capture the complete 3D hand structure implied in joint features, resulting in the loss of crucial structural disambiguation cues. Therefore, we propose to project joint features onto three orthogonal planes to fully preserve 3D hand structure information. We first decode the 3D hand pose from the fused joint features, and then use it to generate distance-aware heatmaps corresponding to the x-y, y-z, and z-x planes. Then, we splat the fused joint feature across the three planes through these heatmaps, and collect the features from each plane as the Triplane Feature maps \\( F_{\\text{triplane}} = \\{F_{x,y}, F_{x,z}, F_{y,z}\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.78 + ], + "angle": 0, + "content": "Spatial-Aware Fusion Triplane encoding transforms the fused joint feature into dense 2D features, eliminating the misalignment in representation form with visual features. However, the initial plane features derived from projecting sparse joint information primarily encode geometric positions while overlooking the relative spatial relationships and interactions between neighboring joints, such as occlusions or depth variations, which are crucial for resolving structural ambiguities in visual features. Additionally, these plane features are spatially non-unified: The x-y plane shares the same spatial dimension with visual features, while the y-z and x-z planes are spatially orthogonal to visual features. To embed the complete 3D hand structure information represented by the Triplane features into the visual features in 2D space, it is essential to compress the y-z and x-z plane features as supplementary depth information relative to the x-y plane into the pixel-wise semantic content. However, directly integrating the y-z and x-z plane features into the x-y plane via cross-attention is inefficient and introduces redundant information. Therefore, we compress the y-z and x-z features along distinct directions into 1D global representations. This compression allows us to guide each pixel in the x-y plane to selectively integrate spatial context from the other two planes, reducing unnecessary overlap while preserving key spatial dependencies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Specifically, we first add 2D positional encoding to each plane feature and apply convolution layers to each plane feature to capture local spatial information (See Fig. 2 (c) project). Next, we perform summation along different directions for \\( F_{x,z} \\) and \\( F_{y,z} \\) to obtain the 1D features \\( I_x, V_x \\) and \\( I_y, V_y \\) (See Fig. 2 (c) compress), where \\( I \\) denotes summation along the Z direction, and \\( V \\) denotes summation along the X or Y direction. Finally, we integrate these" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "features into the pixel-wise content of \\( F_{x,y} \\) through crossattention to obtain spatially compressed features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.129, + 0.655, + 0.144 + ], + "angle": 0, + "content": "3.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.151, + 0.906, + 0.197 + ], + "angle": 0, + "content": "We supervise hand mesh regression through reconstruct loss, supervise temporal enhancement process through token pose loss and temporal pose loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.202, + 0.681, + 0.215 + ], + "angle": 0, + "content": "3.4.1. Reconstruct Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.906, + 0.281 + ], + "angle": 0, + "content": "We used the same reconstruction loss as [53], including vertex/joint loss and mesh smoothing loss. Vertex/Joint loss supervises the prediction error of each mesh and joint point through L1 loss, i.e" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.286, + 0.906, + 0.328 + ], + "angle": 0, + "content": "\\[\nL _ {V / J} = \\sum_ {i = 1} ^ {N} \\| \\nu_ {i} - \\nu_ {i} ^ {*} \\| _ {1} + \\sum_ {i = 1} ^ {K} \\| j _ {i} - j ^ {*} \\| _ {1} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.906, + 0.363 + ], + "angle": 0, + "content": "where \\(\\nu_{i}\\) is the \\(i^{th}\\) vertex of the mesh, \\(j_{i}\\) is the \\(i^{th}\\) joint of the pose, and \\(*\\) denotes the ground-truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.424 + ], + "angle": 0, + "content": "Mesh smooth loss includes supervision of the surface normal vectors and edge lengths to obtain visually reasonable results. The loss functions for normal vector and edge length are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.43, + 0.905, + 0.466 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {n o r m a l}} = \\sum_ {f} \\sum_ {\\{i, j \\} \\subset f} | \\left\\langle \\frac {\\nu_ {i} - \\nu_ {j}}{\\| \\nu_ {i} - \\nu_ {j} \\| _ {2}}, n _ {f} ^ {*} \\right\\rangle | \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.472, + 0.905, + 0.506 + ], + "angle": 0, + "content": "\\[\nL _ {e d g e} = \\sum_ {f} \\sum_ {\\{i, j \\} \\subset f} \\left| \\left\\| \\nu_ {i} - \\nu_ {j} \\right\\| _ {2} - \\left\\| \\nu_ {i} ^ {*} - \\nu_ {j} ^ {*} \\right\\| _ {2} \\right| \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.553 + ], + "angle": 0, + "content": "where \\( f \\) and \\( n_f \\) indicate a mesh face and unit normal vector of face \\( f \\), respectively. \\( n_f^* \\) is computed from ground-truth vertex sets \\( V^* \\), where * denotes ground-truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.584 + ], + "angle": 0, + "content": "The final hand reconstruction task loss combines the vertex/joint loss and the mesh smoothness loss as \\( L_{hand} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.628, + 0.602 + ], + "angle": 0, + "content": "3.4.2. Pose Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.726 + ], + "angle": 0, + "content": "For pose-aware tokens, we enable them to perceive pose changes by supervising their predicted 2D poses. During the temporal fusion process, we use the fused joint feature to predict 2.5D hand pose coordinates for triplane project. We supervise the predicted 2.5D coordinates at each temporal fusion stage to ensure accurate decoding of hand structure information. These two loss functions are represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.725, + 0.905, + 0.765 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t o k e n}} = \\sum_ {i} ^ {K} \\| j _ {i} ^ {2 d} - j ^ {2 d *} \\| _ {1} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.77, + 0.905, + 0.811 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t e m p o r a l}} = \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {K} \\| j _ {t, i} ^ {2. 5 d} - j ^ {2. 5 d *} \\| _ {1} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.813, + 0.906, + 0.857 + ], + "angle": 0, + "content": "The final training loss of our framework, which combines reconstruction loss and pose loss, can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.865, + 0.905, + 0.881 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {h a n d}} + L _ {\\text {t o k e n}} + L _ {\\text {t e m p o r a l}} \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22631" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.088, + 0.903, + 0.306 + ], + "angle": 0, + "content": "
MethodsInputResolutionPA-MPJPE(↓)PA-MPVPE(↓)AUC(↑)F@5(↑)F@15(↑)
METRO[26]Monocular2566.99----
Liu et. al.[28]Monocular2566.60----
MeshGraphormer[27]Monocular2566.41----
HandOccNet[34]Monocular2565.805.5088.478.099.0
H2ONet[47]Monocular1285.705.5089.180.199.0
SimpleHand[57]Monocular2565.505.50---
VIBE[22]Sequence2246.43-87.1--
TCMR[9]Sequence2246.28-87.5--
Deformer[12]Sequence5125.22-89.6--
HMR[19] + LR-ViTMonocular646.456.2487.373.098.5
VIBE[22] + LR-ViTSequence646.226.0987.574.198.7
OursSequence645.725.5288.578.498.7
OursSequence326.035.8288.076.598.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.29, + 0.316, + 0.707, + 0.33 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art on the DexYCB dataset." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.354, + 0.465, + 0.45 + ], + "angle": 0, + "content": "
MethodsInputMPJPEPA-MPJPE
Cho et al.[8]Sequence17.20-
HandOccNetMonocular16.196.13
HMR + LR-ViTMonocular30.727.49
VIBE + LR-ViTSequence27.897.05
OursSequence18.156.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.46, + 0.483, + 0.474 + ], + "angle": 0, + "content": "Table 2. Comparison with the state-of-the-art on the H2O dataset." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.489, + 0.481, + 0.616 + ], + "angle": 0, + "content": "
MethodsInputJ-PAV-PAF@5F@15
HaMuCo[54]Monocular7.00---
EpipolarPose[21]Monocular6.10---
MobRecon[7]Monocular5.705.81--
HaMeR[35]Monocular5.995.6177.899.0
HMR + LR-ViTMonocular6.886.4473.498.7
VIBE + LR-ViTSequence6.796.2775.098.8
OursSequence5.775.7477.699.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.668 + ], + "angle": 0, + "content": "Table 3. Comparison results on the HanCo dataset. The evaluation results for HaMuCo and EpipolarPose using monocular input are derived from [54]." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.681, + 0.481, + 0.728 + ], + "angle": 0, + "content": "
RetrainedHanOccNetSimpleHandHaMeROurs*
×6.63-8.70-
✓5.985.92-5.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.738, + 0.483, + 0.768 + ], + "angle": 0, + "content": "Table 4. SOTA performance (P-MPJPE) on interpolated low-resolution images (DexYCB). *Using raw low-resolution inputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.793, + 0.224, + 0.81 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.194, + 0.833 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.903 + ], + "angle": 0, + "content": "DexYCB contains multi-camera videos of humans grasping 20 YCB [3] objects. It consists of 1000 sequences with 582K frames captured by cameras from 8 different views. During the data collection, the hands are frequently oc" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.388 + ], + "angle": 0, + "content": "cluded during hand-object interaction. We use the dataset split setup s0 from [5] for the training and test sets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.907, + 0.495 + ], + "angle": 0, + "content": "HanCo consists of 1,517 videos with multiple views and camera calibration. It has 860,304 frames in total, i.e. 107,538 time-step per view. Since HanCo does not have an official train/test split, we follow [54] use the first 1,200 sequences for training and validation and the last 317 sequences for testing in all experiments for fair comparisons. We conduct all ablation studies on this dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.907, + 0.573 + ], + "angle": 0, + "content": "H2O includes data from four participants performing 36 distinct actions involving interactions with eight different objects. It provides markerless 3D annotations for both hands. We follow the setup in [25] to divide the training and test sets for pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.673, + 0.599 + ], + "angle": 0, + "content": "4.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.907, + 0.801 + ], + "angle": 0, + "content": "Handling Low Resolution Our experiment was conducted under two original resolution settings, \\(32 \\times 32\\) and \\(64 \\times 64\\). Unless otherwise specified, all experiments use the original resolution of \\(64 \\times 64\\). To avoid additional noise introduced by excessive image interpolation, we set the size of the input image to \\(64 \\times 64\\). We have modified ViT pose-B [49] as the backbone. To adapt it to low resolution inputs, we replaced the patch embedding part with two convolution layers with zero padding, and reduce the number of tokens in ViT to 64. These 64 tokens were rearranged into an \\(8 \\times 8\\) feature map during the fusion of temporal features and regression of hand mesh. We adopt a regression-based method in [53] to reconstruct hand from the final feature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Handling Temporal Input Our method utilizes the temporal information of historical frames to enhance the current hand reconstruction. We have chosen a sequence length of \\( T = 6 \\) with a 5-frames interval based on observations of each datasets. In training, we first extract the visual features and pose-aware tokens of the first 5 frames through the ViT" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22632" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "backbone without temporal fusion, then compressed them into joint features. During this phase, the parameters of the ViT backbone are not updated. These joint features are subsequently processed through spatio-temporal Transformer to enhance the feature extraction process of last frame. We train all network parameters in this phrase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.483, + 0.401 + ], + "angle": 0, + "content": "Training We train and evaluate our method on a single server with an Nvidia GeForce RTX 4090 GPU. All experiments are implemented with PyTorch. For all datasets, the batch size is set to 64. Before formal training on each dataset, we fine-tune the backbone using the pre-trained weights of ViTPose-B. We conducted 30 epochs of fine-tuning for DexYCB and HanCo, and 20 epochs for H2O. For DexYCB ,H2O and HanCo, the model is trained for 50,30,50 epochs,respectively. The model is optimized with AdamW solver with parameters \\((\\beta_{1}\\) of \\(0.9,\\beta_{2}\\) of 0.999,and weight decay of 0.1). We use an initial learning rate of \\(5e - 4\\) and cosine schedule. We perform data augmentation including scaling \\((\\pm 10\\%)\\), rotation \\((\\pm 180^{\\circ})\\), translation \\((\\pm 10\\) pixels), and color jittering \\((\\pm 20\\%)\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.408, + 0.275, + 0.423 + ], + "angle": 0, + "content": "4.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.552 + ], + "angle": 0, + "content": "We employ the standard metric for hand pose estimation: Mean Per Joint Position Error (MPJPE, J-PE) and Mean Per Vertex Position Error (MPVPE, V-PE), measured after Procrustes alignment, to measure the accuracy of the pose and shape of the estimated hand meshes. In addition, we report the area under the curve of PCK (AUC) score and F-score at 5mm (F@5) and 15mm (F@15) threshold. All evaluation metrics are calculated in the root-aligned space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.559, + 0.398, + 0.574 + ], + "angle": 0, + "content": "4.4. Comparisons with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.822 + ], + "angle": 0, + "content": "We compared the proposed low-resolution method with state-of-the-art video-based and single-view methods that use high-resolution input. Additionally, we further use our backbone network (LR-ViT) to establish a low-resolution benchmark for comparison. We trained HMR and VIBE with our backbone on raw low-resolution images to explore the effectiveness of commonly used methods. Comparison results are shown in Tab. 1, Tab. 2 and Tab. 3. In DexYCB, our method performs equally or better than other high-resolution methods at \\(32 \\times 32\\) and \\(64 \\times 64\\) resolution settings, except for Deformer that uses \\(512 \\times 512\\) input resolution. In HanCo and H2O, our method also demonstrates competitive results. We also report the performance of SOTA methods on upsampled low-resolution images in Tab. 4, including cases where models were retrained with upsampled images and where they were not." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.245, + 0.846 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.263, + 0.867 + ], + "angle": 0, + "content": "4.5.1. Backbone settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we experiment with different backbone settings and report their performance with and without tem" + }, + { + "type": "table", + "bbox": [ + 0.571, + 0.089, + 0.852, + 0.184 + ], + "angle": 0, + "content": "
Token SettingsMonocularTemporal
J-PAV-PAJ-PAV-PA
06.716.75--
16.586.646.036.09
6 (SP)6.196.195.775.74
21 (JP)6.076.055.805.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.195, + 0.89, + 0.209 + ], + "angle": 0, + "content": "Table 5. Ablation study on backbone and pose-aware tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.237, + 0.905, + 0.525 + ], + "angle": 0, + "content": "poral enhancement. Under monocular input, the backbone skips temporal fusion and uses only transformer blocks to process the input. Our experiment further explored the impact of Pose-Aware Tokens. We adopted different settings of introduced tokens to test their impact. Under monocular input, we still supervise these tokens to predict 2D hand poses. The experiment result is shown in Tab. 5. The introduction of temporal enhancement universally improves performance. Under monocular input, the original ViT (w/o pose-aware tokens) performs poorly at low resolutions, while other networks generally achieve performance gains. This demonstrates that supervising intermediate features through specific tokens is more effective than solely employing end-to-end training on the final regression results. Meanwhile, due to the ability of finger based pose-aware tokens (SP) to jointly carry high-level semantic information and low-level visual details, it is more effective in temporal enhancement than a single pose token and skeleton based pose tokens (JP)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.535, + 0.743, + 0.549 + ], + "angle": 0, + "content": "4.5.2. Temporal Fusion Schemes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.827 + ], + "angle": 0, + "content": "In this section, we explored the effects of different temporal fusion schemes. We consider three ways to fuse temporal features and visual features: cross attention, single plane encoding, and triplane encoding. The cross attention approach uses joint features as K and V, and visual features as Q to interact through attention mechanisms. The single plane encoding method uses 2D Heatmap to diffuse joint features into a single plane. The single plane feature is also embedded with 2D positional encoding and processed using convolutional layers. For the triplane encoding, we also tested the effect of directly integrating the full \\( F_{x,z} \\) and \\( F_{y,z} \\) with \\( F_{x,y} \\). The comparison results are presented in Table 6. Our proposed triplane encoding performs best. Compared to the other two methods, the triplane encoding effectively extracts crucial 3D structural information, providing cues for eliminating hand structure ambiguity in visual features. Additionally, using compressed 1D features for interaction is superior to using complete feature." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.836, + 0.821, + 0.851 + ], + "angle": 0, + "content": "4.5.3. Position for inserting temporal fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this section, we further investigated the impact of the insertion position of temporal fusion operations in the backbone. Specifically, we compared two strategies: one where" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22633" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.189, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.089, + 0.29, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.089, + 0.391, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.089, + 0.495, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.166, + 0.317, + 0.177 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.178, + 0.189, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.178, + 0.29, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.178, + 0.391, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.178, + 0.494, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.254, + 0.318, + 0.264 + ], + "angle": 0, + "content": "(c) HMR" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.266, + 0.189, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.266, + 0.29, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.266, + 0.391, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.266, + 0.494, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.342, + 0.335, + 0.354 + ], + "angle": 0, + "content": "(e) Ours-Single" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.601, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.089, + 0.703, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.089, + 0.805, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.089, + 0.906, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.166, + 0.724, + 0.176 + ], + "angle": 0, + "content": "(b) GT" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.178, + 0.601, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.178, + 0.703, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.178, + 0.805, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.178, + 0.906, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.254, + 0.731, + 0.265 + ], + "angle": 0, + "content": "(d) VIBE" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.266, + 0.601, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.266, + 0.703, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.266, + 0.805, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.266, + 0.906, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.342, + 0.728, + 0.353 + ], + "angle": 0, + "content": "(f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.366, + 0.869, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Qualitative Comparison Results on DexYCB. Regions with significant prediction errors are marked with red circles." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.404, + 0.457, + 0.5 + ], + "angle": 0, + "content": "
Fusion SchemeJ-PAV-PAF@5F@15
Cross-attention5.955.9975.898.9
Single plane Encoding5.855.9076.799.0
Triplane Encoding (w/o 1D compress)5.805.8576.999.0
Triplane Encoding5.775.7477.699.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.51, + 0.468, + 0.524 + ], + "angle": 0, + "content": "Table 6. Ablation study of different temporal fusion schemes." + }, + { + "type": "table", + "bbox": [ + 0.146, + 0.539, + 0.43, + 0.589 + ], + "angle": 0, + "content": "
Insert PositionJ-PAV-PAF@5F@15
Final layer6.016.0575.198.6
Middle layer5.775.7477.699.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.627 + ], + "angle": 0, + "content": "Table 7. Ablation Study on the Position of Temporal Information Fusion in the Backbone Network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.484, + 0.807 + ], + "angle": 0, + "content": "temporal features are fused at multiple intermediate levels throughout the network, and another where fusion occurs solely at the final level, as shown in Tab. 7. The experimental results in Tab. 7 demonstrate that fusing temporal features at multiple intermediate levels leads to superior performance compared to only performing fusion at the final level. This suggests that incorporating temporal information at various stages of the network helps to better capture ambiguous local features, contributing to more accurate and consistent hand pose reconstruction in low resolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.274, + 0.833 + ], + "angle": 0, + "content": "4.6. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We present the qualitative comparison results on DexYCB in Fig. 4. Compared with low resolution benchmarks (HMR, VIBE) that use the same LR-ViT as the backbone, our method achieves better alignment of the reconstructed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.907, + 0.483 + ], + "angle": 0, + "content": "hands with low-resolution images. Additionally, by leveraging historic temporal information to supplement missing details in visual features, our approach enhances the ability to cope with local occlusions compare to solely rely on a single frame (see columns 2 and 3 in Fig. 4 (e) and (f))." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.495, + 0.634, + 0.51 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.763 + ], + "angle": 0, + "content": "In this paper, we propose a temporal enhancement framework for low-resolution hand reconstruction. We represent temporal information through joint features to capture the spatiotemporal correlations between joints. Furthermore, we integrate the temporal context provided by joint features with the spatial context provided by visual features to construct a robust low-resolution representation with 3D spatial awareness. Specifically, we introduce a Triplane feature representation to address the misalignment in semantics and representation forms between joint and visual features, preserving 3D structural information during fusion. Quantitative experiments on multiple datasets have shown that our method achieves comparable performance to high-resolution SOTA methods at much lower resolutions. We also demonstrated the contribution of each design to the scheme performance through ablation study." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "**Limitation and Future Work** Our work did not explicitly model long-term hand motion cues across video sequences, which may result in a lack of constraints on motion consistency. As a result, the model may struggle with maintaining consistent hand shapes over extended sequences. Future work should explore utilizing motion cues from long sequences to better constrain the reconstructed hand shape and improve temporal consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22634" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.259 + ], + "angle": 0, + "content": "Acknowledgements This work was supported by the National Natural Science Foundation of China under Grants (62406039, 62101064, 62171057,62201072, U23B2001, 62001054, 62071067), the Ministry of Education and China Mobile Joint Fund (MCM20200202, MCM20180101), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center, BUPT Excellent Ph.D. Students Foundation (CX20241014), in part by the Project funded by China Postdoctoral Science Foundation (2023TQ0039, 2024M750257), the Postdoctoral Fellowship Program of CPSF (GZC20230320)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.271, + 0.188, + 0.287 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.296, + 0.482, + 0.351 + ], + "angle": 0, + "content": "[1] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10843-10852, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[2] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 666-682, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[3] Berk Calli, Arjun Singh, Aaron Walsman, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In 2015 international conference on advanced robotics (ICAR), pages 510-517. IEEE, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[5] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12929–12938, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20544-20554, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[8] Hoseong Cho, Chanwoo Kim, Jihyeon Kim, Seongyeong Lee, Elkhan Ismayilzada, and Seungryul Baek. Transformer-based unified recognition of two hands manipulating objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4769-4778, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[9] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consis" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.296, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "tent 3d human pose and shape from a video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1964-1973, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[11] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[12] Qichen Fu, Xingyu Liu, Ran Xu, Juan Carlos Niebles, and Kris M Kitani. Deformer: Dynamic fusion transformer for robust hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23600-23611, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.907, + 0.468 + ], + "angle": 0, + "content": "[13] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10833-10842, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[14] Peter Hardy, Srinandan Dasmahapatra, and Hansung Kim. Can super resolution be used to improve human pose estimation in low resolution scenarios? arXiv preprint arXiv:2107.02108, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.525, + 0.905, + 0.606 + ], + "angle": 0, + "content": "[15] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 571-580, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[16] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[17] Weiting Huang, Pengfei Ren, Jingyu Wang, Qi Qi, and Haifeng Sun. Awr: Adaptive weighting regression for 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11061-11068, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[18] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In Proceedings of the European conference on computer vision (ECCV), pages 118-134, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[19] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[20] Leyla Khaleghi, Alireza Sepas-Moghaddam, Josh Marshall, and Ali Etemad. Multiview video-based 3-d hand pose estimation. IEEE Transactions on Artificial Intelligence, 4:896-909, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF conference on" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22635" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "computer vision and pattern recognition, pages 1077-1086, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[22] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[23] Deying Kong, Linguang Zhang, Liangjian Chen, Haoyu Ma, Xiangyi Yan, Shanlin Sun, Xingwei Liu, Kun Han, and Xiaohui Xie. Identity-aware hand mesh estimation and personalization from rgb images. In European Conference on Computer Vision, pages 536-553. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[24] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4990-5000, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[25] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.476 + ], + "angle": 0, + "content": "[26] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[27] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12939-12948, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[28] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[29] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[30] Xingyu Liu, Pengfei Ren, Yuanyuan Gao, Jingyu Wang, Haifeng Sun, Qi Qi, Zirui Zhuang, and Jianxin Liao. Keypoint fusion for rgb-d based 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3756-3764, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[31] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 752-768. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[32] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2308-2317, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[33] Chihiro Nakatsuka and Satoshi Komorita. Stable 3d human pose estimation in low-resolution videos with a few views. In 2022 International Conference on Robotics and Automation (ICRA), pages 427-433. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[34] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.276 + ], + "angle": 0, + "content": "[35] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[36] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. Pose-guided hierarchical graph reasoning for 3-d hand pose estimation from a single depth image. IEEE Transactions on Cybernetics, 53(1):315-328, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[37] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. A dual-branch self-boosting framework for self-supervised 3d hand pose estimation. IEEE Transactions on Image Processing, 31:5052–5066, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[38] Pengfei Ren, Haifeng Sun, Jiachang Hao, Jingyu Wang, Qi Qi, and Jianxin Liao. Mining multi-view information: a strong self-supervised framework for depth-based 3d hand pose and mesh estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20555-20565, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[39] Pengfei Ren, Yuchen Chen, Jiachang Hao, Haifeng Sun, Qi Qi, Jingyu Wang, and Jianxin Liao. Two heads are better than one: Image-point cloud network for depth-based 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2163-2171, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[40] Pengfei Ren, Chao Wen, Xiaozheng Zheng, Zhou Xue, Haifeng Sun, Qi Qi, Jingyu Wang, and Jianxin Liao. Decoupled iterative refinement framework for interacting hands reconstruction from a single rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8014-8025, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[41] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[42] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[43] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.885 + ], + "angle": 0, + "content": "[44] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11698-11707, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.887, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[45] Chen Wang, Feng Zhang, Xiatian Zhu, and Shuzhi Sam Ge." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22636" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Low-resolution human pose estimation. Pattern Recognition, 126:108579, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[46] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[47] Hao Xu, Tianyu Wang, Xiao Tang, and Chi-Wing Fu. H2onet: Hand-occlusion-and-orientation-aware network for real-time 3d hand mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17048–17058, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[48] Xiangyu Xu, Hao Chen, Francesc Moreno-Noguer, Laszlo A Jeni, and Fernando De la Torre. 3d human pose, shape and texture from low-resolution images and videos. IEEE transactions on pattern analysis and machine intelligence, 44(9): 4490-4504, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[49] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[50] John Yang, Hyung Jin Chang, Seungeui Lee, and Nojun Kwak. Seqhand: Rgb-sequence-based 3d hand pose and shape estimation. In European Conference on Computer Vision, pages 122-139. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[51] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13232-13242, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[52] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[53] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Sar: Spatial-aware regression for 3d hand pose and mesh reconstruction from a monocular rgb image. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 99-108. IEEE, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[54] Xiaozheng Zheng, Chao Wen, Zhou Xue, Pengfei Ren, and Jingyu Wang. Hamuco: Hand pose estimation via multiview collaborative self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20763-20773, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[55] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[56] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[57] Zhishan Zhou, Shihao Zhou, Zhi Lv, Minqiang Zou, Yao Tang, and Jiajun Liang. A simple baseline for efficient hand mesh reconstruction. In Proceedings of the IEEE/CVF Con-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 1367-1376, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[58] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15085-15099, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[59] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[60] Christian Zimmermann, Max Argus, and Thomas Brox. Contrastive representation learning for hand shape estimation. In Pattern Recognition, pages 250-264, Cham, 2021. Springer International Publishing. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22637" + } + ] +] \ No newline at end of file diff --git a/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/61fa7d4b-03fc-4f0b-b84c-af353e8bd6fc_origin.pdf b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/61fa7d4b-03fc-4f0b-b84c-af353e8bd6fc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f035ee307aad44eb1a36751b00af8133165ea8d0 --- /dev/null +++ b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/61fa7d4b-03fc-4f0b-b84c-af353e8bd6fc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6366dca44a5ce6eff3865919b7e58fdb971da76c53e3014abf26d01e0bbace75 +size 5171903 diff --git a/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/full.md b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..522d67a50166746491846427eaa2fd2ff99218ea --- /dev/null +++ b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/full.md @@ -0,0 +1,382 @@ +# Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction + +Kaixin Fan, Pengfei Ren\*, Jingyu Wang, $^{\dagger}$ Haifeng Sun, Qi Qi, Zirui Zhuang, Jianxin Liao +State Key Laboratory of Networking and Switching Technology, +Beijing University of Posts and Telecommunications + +{fkxbupt, rpf, wangjingyu, hfsun, qiqi8266, zhuangzirui, liaojx}@bupt.edu.cn + +# Abstract + +3D hand reconstruction is essential in non-contact human-computer interaction applications, but existing methods struggle with low-resolution images, which occur in slightly distant interactive scenes. Leveraging temporal information can mitigate the limitations of individual low-resolution images that lack detailed appearance information, enhancing the robustness and accuracy of hand reconstruction. Existing temporal methods typically use joint features to represent temporal information, avoiding interference from redundant background information. However, joint features excessively disregard the spatial context of visual features, limiting hand reconstruction accuracy. We propose to integrate temporal joint features with visual features to construct a robust low-resolution visual representation. We adopt Triplane Features, a dense representation with 3D spatial awareness, to bridge the gap between the joint features and visual features that are misaligned in terms of representation form and semantics. Triplane Features are obtained by orthogonally projecting joint features, embedding hand structure information into the 3D spatial context. Furthermore, we compress the spatial information of the three planes into a 2D dense feature though Spatial-Aware Fusion to enhance the visual features. By using enhanced visual features enriched with temporal information for hand reconstruction, our method achieves competitive performance at much lower resolutions compared to state-of-the-art methods operating at high resolution on DexYCB, HanCo and H2O. Code is available at https://github.com/NewbieFan/Temp-LowRes-hand. + +# 1. Introduction + +3D hand reconstruction, a crucial component in non-contact human-computer interaction applications, has made signifi- + +![](images/2618034399817112badc1206dd577a8dfd8ade133e6a17c642afce196b696cd4.jpg) +(a) No Interaction +(b) Cross Attention + +![](images/8f0bedb21c5bc037a15ca55228c1445547c7e5523607938d98c418d7af2f7305.jpg) + +![](images/ac48e0fe3f9bdf291ab41a7e40a1a357e219ef9dde255ddd99628040548fb0e0.jpg) +(c) Triplane Feature +Figure 1. Different way to utilizing joint features. a) Directly reconstruct hand from joint features, relying solely on sparse features without spatial context. b) Fusing joint and visual features through cross-attention, which avoids representation discrepancies but neglects the hand's structural priors. c) Using triplane features to bridge joint and visual features by decoding structural joint information into dense representations that enhance visual features. + +cant advancements in recent years with the advent of large-scale datasets and deep learning technologies. Many studies [6, 7, 30, 31, 34, 40] have focused on reconstructing 3D hands from a single RGB image. In typical interaction scenarios, these methods require the subject to stay close to the capture camera to obtain high-resolution (e.g., $256 \times 256$ ) cropped hand images. However, as the interaction distance increases—such as in XR teaching activities within a classroom—the available cropped hand images are limited to lower resolutions (e.g., $64 \times 64$ ). Current methods + +perform poorly on such low-resolution images, as these images provide fewer cues for addressing joint self-similarity and depth ambiguity, significantly increasing the difficulty of 3D hand reconstruction. + +To address the low-resolution problem, one intuitive approach is to lift the resolution through super-resolution techniques [14]. However, general super-resolution methods can lead to structural distortions and artifacts in the hand regions, which brings a domain gap between super-resolved image and real high-resolution image. Low-Res Pose [45] adopted confidence-aware learning to supervise continue joint offset fields, thereby alleviating the quantization errors. RSC-Net [48] adopted cross-resolution contrastive learning to align the estimation results of low-resolution images with high-resolution images. However, the performance of these methods is still limited due to the inherent ambiguity resulting from information loss in low resolution. + +A single low-resolution image lacks detailed hand appearance, leading to ambiguous visual representation. Continuous frames provide comprehensive hand appearance and motion context, offering strong disambiguation cues for inferring global hand structure. To reduce interference from redundant background information and computational costs, existing temporal methods typically compress single-frame information into joint features [51, 58], and then model the intra- and inter-frame correlation between joints through temporal interaction. However, joint features excessively discard spatial context of visual features, which limits the accuracy of pose estimation. Integrating the spatial context of visual features with the temporal context of joint features can construct robust visual representation with 3D spatial awareness. However, integrating joint features into visual features present two major challenges. Firstly, sparse joint features are not aligned with dense visual features in terms of representation form. Secondly, joint features contain prior geometric structural information in 3D space (e.g., hand scale, skeleton length), while visual features represent detailed appearance information in 2D space (e.g., texture, color), leading to a semantic misalignment. + +In this paper, we integrate the temporal clue from historical frames into the current visual features to achieve robust and accurate 3D hand pose estimation under low-resolution scenarios. Specifically, we compress temporal information into joint representations, and model intra- and interframe relationships of joints through spatio-temporal interaction. Then, we adopt triplane features[4], a 3D spatial-aware dense visual representation, to bridge the sparse 3D joint representation and the dense 2D visual representation. We orthogonally project the joint features into three individual planes, with each plane independently embedding the sparse joint features into specific spatial contexts. The orthogonal planes ensure comprehensive coverage of 3D hand structure information in a low-redundancy manner. How + +ever, they still contain redundant spatial information, directly integrating these features into the visual plane leads to the overlap of spatial context from different planes. Thus, we compress the features from two planes orthogonal to the visual plane into 1D global features, guiding each pixel in visual features to integrate distinct spatial context from the other planes, reducing redundancy and clarifying the 3D spatial context. Finally, we reconstruct the hand based on the enhanced visual features. + +Experimental results on several large-scale hand pose estimation video datasets, DexYCB [5], HanCo [60] and H2O [25], show the proposed method reach competitive performance at much lower resolution compared to both single-frame based and temporal based state-of-the-art (SOTA) methods at high resolution. In summary, our contribution is list as below: + +- We propose a temporal enhancement framework that integrates temporal joint features with visual features to construct a robust visual representation for low-resolution conditions, thereby improving hand reconstruction performance in low-resolution videos. +- We introduced triplane feature encoding to encode joint features into dense feature maps. Through triplane feature, visual features can perceive the prior hand structure information from the temporal joint features. +- Our method reaches competitive result at much lower resolution compared to SOTA methods at high resolution. + +# 2. Relate Work + +# 2.1. Single Image based Hand Pose Estimation + +Early works [2, 17, 18, 36, 43, 59] focused on estimating hand pose solely from input images, which suffered from occlusion issues due to the lack of modeling the inherent structure of the hand. MANO [41] introduces a parametric hand model, which contains rich priors of human hands, and now most work focuses on reconstruct this pre-defined hand from image. Some works [1, 23, 32, 37-39, 52, 56] directly predict parameters of MANO, 3D joints and mesh vertices coordinates are then retrieved using the differentiable MANO model. However, predicting MANO parameters directly from a single RGB image is a challenging job, which is difficult for network optimization and is prone to mesh-image misalignment [44], especially in a low-res input condition. To alleviate these issues present in parametric predicting, some works [13, 24, 27] regress the 3D coordinates or occupancy directly for mesh vertices. For example, [53] adopts graph convolution neural network to regress 2.5D Heatmap for every mesh vertices. [26] proposed an End to end framework to recover mesh using transformer. These works show that using a regression-based approach can maintain flexibility while achieving comparable results to parametric methods. Thus, we adopt a regression-based + +method described in [53] to reconstruct the hand mesh from enhanced visual features. + +# 2.2. Low Resolution Pose Estimation + +Recently, a few works have explored on low resolution pose estimation. A intuitive way is leveraging Super Resolution approach [14] to improve the image quality and pose estimation performance. However, the addition super-resolution network brings extra computation cost and may bring unpleasant artifacts to the input image. [45] find that offset-based methods can provide more reliable results in low resolution, and further proposed a confidence-aware learning to selectively weigh the learning of heatmap and offset. [48] controls the scaling of the CNN architecture model using a set of hyperparameters, enforcing it to output scale-invariant features at different resolutions through contrastive learning and self-supervised learning. Although specific learning schemas improve the poor results in low resolution, the features extracted by the network are still constrained by the lack of detailed hand appearance. Other work [33] explored on 3D pose estimation in low-resolution multi-view videos, showing that temporal and multi-view information could be a supplement to alleviate information loss. However, they only correct the prediction under certain adverse conditions, ignoring temporal information can reduce the inherent ambiguity in single low-resolution image and imporve the overall performance. + +# 2.3. Temporal Hand Pose Estimation + +Some recent methods [11, 15, 22, 29, 50, 55] have begun to explore leveraging temporal information in videos to overcome the limitations of single-view images. Hasson et al. [15] use photometric supervision to keep consistency between the re-projected 3D hand and the optical flow. VIBE [22] use a temporal generation network to predict model parameters for each frame, and employs a recurrent network as a discriminator to supervise motion sequence. However, they ignored the spatial and feature correlations in the video. More recently, [20] has harnessed temporal information by propagating frame-wise features with a sequence-to-sequence model based on LSTM [16], but it only provides sparse joint location predictions. TCMR [9] introduces a temporally consistent system that synthesizes implicit features from past and future frames to maintain temporal consistency in the mesh of the current frame. Deformer [12] deforms the hand poses from neighborhood frames using predicted motion to avoid unreliable queries generated with local feature when leveraging self-attention mechanism. + +In low resolutions, these methods are limited to the poor quality of single frame features. Additionally, they lack a focus center when processing temporal features, which may cause some estimation results to deviate from the actual image due to the influence of temporal context. + +# 3. Method + +The overview of our method is shown in the Fig. 2. For a given frame $I_{T}$ in an image sequence $V = \{I_{t}\}_{t=1}^{T}$ , we extract temporal cues from the previous $T - 1$ frames to enhance the visual feature of the current frame and reconstruct the hand mesh from the enhanced visual feature. Specifically, we first compress the information of each historical frames into joint features and establish their inter- and intra-frame correlations through a spatio-temporal Transformer. Then, we encode the structural information from the enhanced joint features into 3D spatial-aware dense representations through triplane feature encoding to enhance the visual features of the current frame. Finally, we leverage the enhanced visual feature to reconstruct the hand mesh. + +# 3.1. Static Visual Features + +As shown in Fig. 3, we adopt a Vision Transformer (ViT) [10] to extract visual features. During the extraction process, ambiguous local features would propagate from low-level features to subsequent high-level features, leading to ambiguities in both semantic content and fine-grained details in the final visual features. Thus, we incorporate temporal information into intermediate features at different levels to gradually reduce ambiguous local representations and ultimately obtain visual features with a clear perception of hand structure. However, static visual features lack the ability to perceive dynamic temporal information represented in joint features. Therefore, we introduce pose-aware tokens as anchors for fusing dynamic temporal information into a static representation, which are used for further interaction with visual features. These tokens capture hand pose and appearance information from the current visual feature through attention mechanism. We categorize the functionalities of these tokens based on finger semantic to ensure they incorporate both fine-grained visual details (e.g., texture) and high-level semantic content (e.g., skeletal structures). + +# 3.2. Dynamic Temporal Joint Features + +We use compact joint features to represent dynamic temporal information from historical frames. Give a single frame, we obtain its visual feature map along with pose-aware tokens through the backbone. Each pose-aware token carries semantic information corresponding to different individual fingers. We further combine local visual features to expand each token into a complete joint feature, enabling fine-grained temporal interaction around each joints. We adopt a FC layer to predict 2D joint coordinates $p \in \mathbb{R}^{J \times 2}$ from pose-aware tokens, and use it to sample joint-centric local features $J^{sample} \in \mathbb{R}^{J \times C}$ from the visual feature map. The joint coordinates is encoded as pose feature $J^{pose}$ , and then add to the sampled feature $J^{sample}$ to obtain the initial joint feature $J^{init} = J^{pose} + J^{region}$ . We concatenate the joint features from each frame in the upstream sequence + +![](images/faafee0f03e7d489097e453072fb8ebd15ec990db44bf1bb8932c1c5fc22d5ca.jpg) + +![](images/2eb451d28a4708f478ee48e5c5e4b52db636d23418b8290c9329ecf0eaeecfa5.jpg) +Figure 2. Overview of Our Framework. We first construct joint features from the sampled past frame features and 2D poses using a Spatial-Temporal Transformer (a), capturing the temporal dynamics and spatial relationships between joints. We then use these joint features to enhance the visual features of the current frame (b) by converting them into spatial-aware dense features through Tripalne Feature Encoding (c). In (c), we fuse different joint features (from a) using pose-aware tokens (from b), and project them into three orthogonal planes to obtain the Triplane features. These features are then compressed into a single plane through space-aware fusion. + +![](images/e3485bd1c5523836e146b411c8b2d282169e5d2e5f5d8a8bd821da104457b043.jpg) +Figure 3. Backbone of our framework. + +to form the initial representation of temporal information $J^{temp} = \text{Concat}(J_t^{init}) \in \mathbb{R}^{T \times J \times C}$ . Then, we adopt a spatio-temporal Transformer to establish inter- and intra-frame correlation between these joints. The spatio-temporal Transformer is a combination of alternately stacked temporal and spatial attention modules. In the spatial attention module, attention interactions occur among different joints within the same joint feature: + +$$ +S - M H S A (J ^ {t e m p}) = M H S A \left(Q _ {j}, K _ {j}, V _ {j}\right) \tag {1} +$$ + +$$ +Q _ {j} = W _ {q} J _ {j} ^ {\text {t e m p}}, K _ {j} = W _ {k} J _ {j} ^ {\text {t e m p}}, V _ {j} = W _ {v} J _ {j} ^ {\text {t e m p}} \tag {2} +$$ + +In the temporal attention module, attention interactions occur among the same joint within different joint features: + +$$ +T - M H S A \left(J ^ {t e m p}\right) = M H S A \left(Q _ {t}, K _ {t}, V _ {t}\right) \tag {3} +$$ + +$$ +Q _ {t} = W _ {q} J _ {t} ^ {t e m p}, K _ {j} = W _ {k} J _ {t} ^ {t e m p}, V _ {j} = W _ {v} J _ {t} ^ {t e m p} \quad (4) +$$ + +Here, $W_{q}$ , $W_{k}$ , and $W_{v}$ are learnable weight matrices used to transform the input features into queries, keys, and values, respectively. $J_{j}^{temp}$ represents the joint features for the $j$ -th joint across all time steps, while $J_{t}^{temp}$ represents the joint features for the $t$ -th time step across all joints. + +# 3.3. Triplane Feature Encoding + +Joint features contain temporal context of the current hand pose, which can be integrated into visual features to construct low-resolution robust visual representations. However, it is difficult to fuse joint features with visual features that are not aligned in semantic content and representation form. A straightforward approach is to treat joint features as global tokens and interact them with visual features through cross attention. However, visual features struggle to effectively perceive the 3D spatial properties of joint features through attention mechanisms. Thus, we propose to use triplane features as an intermediate representation to enhance the visual features. Unlike [4, 42, 46] that constructs dense Triplane features from images for spatial indexing, we leverage sparse joint features to generate triplane features, embedding 3D geometric attributes of the skeleton into specific spatial context. + +Specifically, we first copy each pose token to the corresponding finger joints as a Pose Template $J^{P} \in \mathbb{R}^{J \times C}$ , and use it to fuse dynamic joint features into a static one. Then splat it onto three orthogonal planes as an intermediate representation for interaction with visual features. Fur + +thermore, we integrate the information from different planes into a single plane through spatial-aware fusion, allowing for direct fusion with the visual features. + +Triplane Projection of Joint Features Visual features represent specific spatial information in the form of dense representations. Transferring joint features into the 2D visual feature space through 3D-2D pose projection [40] can provide effective disambiguation cues for hand poses. However, a single projection plane cannot capture the complete 3D hand structure implied in joint features, resulting in the loss of crucial structural disambiguation cues. Therefore, we propose to project joint features onto three orthogonal planes to fully preserve 3D hand structure information. We first decode the 3D hand pose from the fused joint features, and then use it to generate distance-aware heatmaps corresponding to the x-y, y-z, and z-x planes. Then, we splat the fused joint feature across the three planes through these heatmaps, and collect the features from each plane as the Triplane Feature maps $F_{\text{triplane}} = \{F_{x,y}, F_{x,z}, F_{y,z}\}$ . + +Spatial-Aware Fusion Triplane encoding transforms the fused joint feature into dense 2D features, eliminating the misalignment in representation form with visual features. However, the initial plane features derived from projecting sparse joint information primarily encode geometric positions while overlooking the relative spatial relationships and interactions between neighboring joints, such as occlusions or depth variations, which are crucial for resolving structural ambiguities in visual features. Additionally, these plane features are spatially non-unified: The x-y plane shares the same spatial dimension with visual features, while the y-z and x-z planes are spatially orthogonal to visual features. To embed the complete 3D hand structure information represented by the Triplane features into the visual features in 2D space, it is essential to compress the y-z and x-z plane features as supplementary depth information relative to the x-y plane into the pixel-wise semantic content. However, directly integrating the y-z and x-z plane features into the x-y plane via cross-attention is inefficient and introduces redundant information. Therefore, we compress the y-z and x-z features along distinct directions into 1D global representations. This compression allows us to guide each pixel in the x-y plane to selectively integrate spatial context from the other two planes, reducing unnecessary overlap while preserving key spatial dependencies. + +Specifically, we first add 2D positional encoding to each plane feature and apply convolution layers to each plane feature to capture local spatial information (See Fig. 2 (c) project). Next, we perform summation along different directions for $F_{x,z}$ and $F_{y,z}$ to obtain the 1D features $I_x, V_x$ and $I_y, V_y$ (See Fig. 2 (c) compress), where $I$ denotes summation along the Z direction, and $V$ denotes summation along the X or Y direction. Finally, we integrate these + +features into the pixel-wise content of $F_{x,y}$ through crossattention to obtain spatially compressed features. + +# 3.4. Optimization + +We supervise hand mesh regression through reconstruct loss, supervise temporal enhancement process through token pose loss and temporal pose loss. + +# 3.4.1. Reconstruct Loss + +We used the same reconstruction loss as [53], including vertex/joint loss and mesh smoothing loss. Vertex/Joint loss supervises the prediction error of each mesh and joint point through L1 loss, i.e + +$$ +L _ {V / J} = \sum_ {i = 1} ^ {N} \| \nu_ {i} - \nu_ {i} ^ {*} \| _ {1} + \sum_ {i = 1} ^ {K} \| j _ {i} - j ^ {*} \| _ {1} \tag {5} +$$ + +where $\nu_{i}$ is the $i^{th}$ vertex of the mesh, $j_{i}$ is the $i^{th}$ joint of the pose, and $*$ denotes the ground-truth. + +Mesh smooth loss includes supervision of the surface normal vectors and edge lengths to obtain visually reasonable results. The loss functions for normal vector and edge length are defined as follows: + +$$ +L _ {\text {n o r m a l}} = \sum_ {f} \sum_ {\{i, j \} \subset f} | \left\langle \frac {\nu_ {i} - \nu_ {j}}{\| \nu_ {i} - \nu_ {j} \| _ {2}}, n _ {f} ^ {*} \right\rangle | \tag {6} +$$ + +$$ +L _ {e d g e} = \sum_ {f} \sum_ {\{i, j \} \subset f} \left| \left\| \nu_ {i} - \nu_ {j} \right\| _ {2} - \left\| \nu_ {i} ^ {*} - \nu_ {j} ^ {*} \right\| _ {2} \right| \tag {7} +$$ + +where $f$ and $n_f$ indicate a mesh face and unit normal vector of face $f$ , respectively. $n_f^*$ is computed from ground-truth vertex sets $V^*$ , where * denotes ground-truth. + +The final hand reconstruction task loss combines the vertex/joint loss and the mesh smoothness loss as $L_{hand}$ . + +# 3.4.2. Pose Loss + +For pose-aware tokens, we enable them to perceive pose changes by supervising their predicted 2D poses. During the temporal fusion process, we use the fused joint feature to predict 2.5D hand pose coordinates for triplane project. We supervise the predicted 2.5D coordinates at each temporal fusion stage to ensure accurate decoding of hand structure information. These two loss functions are represented as follows: + +$$ +L _ {\text {t o k e n}} = \sum_ {i} ^ {K} \| j _ {i} ^ {2 d} - j ^ {2 d *} \| _ {1} \tag {8} +$$ + +$$ +L _ {\text {t e m p o r a l}} = \sum_ {t = 1} ^ {T} \sum_ {i = 1} ^ {K} \| j _ {t, i} ^ {2. 5 d} - j ^ {2. 5 d *} \| _ {1} \tag {9} +$$ + +The final training loss of our framework, which combines reconstruction loss and pose loss, can be expressed as follows: + +$$ +L = L _ {\text {h a n d}} + L _ {\text {t o k e n}} + L _ {\text {t e m p o r a l}} \tag {10} +$$ + +
MethodsInputResolutionPA-MPJPE(↓)PA-MPVPE(↓)AUC(↑)F@5(↑)F@15(↑)
METRO[26]Monocular2566.99----
Liu et. al.[28]Monocular2566.60----
MeshGraphormer[27]Monocular2566.41----
HandOccNet[34]Monocular2565.805.5088.478.099.0
H2ONet[47]Monocular1285.705.5089.180.199.0
SimpleHand[57]Monocular2565.505.50---
VIBE[22]Sequence2246.43-87.1--
TCMR[9]Sequence2246.28-87.5--
Deformer[12]Sequence5125.22-89.6--
HMR[19] + LR-ViTMonocular646.456.2487.373.098.5
VIBE[22] + LR-ViTSequence646.226.0987.574.198.7
OursSequence645.725.5288.578.498.7
OursSequence326.035.8288.076.598.6
+ +Table 1. Comparison with the state-of-the-art on the DexYCB dataset. + +
MethodsInputMPJPEPA-MPJPE
Cho et al.[8]Sequence17.20-
HandOccNetMonocular16.196.13
HMR + LR-ViTMonocular30.727.49
VIBE + LR-ViTSequence27.897.05
OursSequence18.156.11
+ +Table 2. Comparison with the state-of-the-art on the H2O dataset. + +
MethodsInputJ-PAV-PAF@5F@15
HaMuCo[54]Monocular7.00---
EpipolarPose[21]Monocular6.10---
MobRecon[7]Monocular5.705.81--
HaMeR[35]Monocular5.995.6177.899.0
HMR + LR-ViTMonocular6.886.4473.498.7
VIBE + LR-ViTSequence6.796.2775.098.8
OursSequence5.775.7477.699.0
+ +Table 3. Comparison results on the HanCo dataset. The evaluation results for HaMuCo and EpipolarPose using monocular input are derived from [54]. + +
RetrainedHanOccNetSimpleHandHaMeROurs*
×6.63-8.70-
✓5.985.92-5.72
+ +Table 4. SOTA performance (P-MPJPE) on interpolated low-resolution images (DexYCB). *Using raw low-resolution inputs. + +# 4. Experiments + +# 4.1. Datasets + +DexYCB contains multi-camera videos of humans grasping 20 YCB [3] objects. It consists of 1000 sequences with 582K frames captured by cameras from 8 different views. During the data collection, the hands are frequently oc + +cluded during hand-object interaction. We use the dataset split setup s0 from [5] for the training and test sets. + +HanCo consists of 1,517 videos with multiple views and camera calibration. It has 860,304 frames in total, i.e. 107,538 time-step per view. Since HanCo does not have an official train/test split, we follow [54] use the first 1,200 sequences for training and validation and the last 317 sequences for testing in all experiments for fair comparisons. We conduct all ablation studies on this dataset. + +H2O includes data from four participants performing 36 distinct actions involving interactions with eight different objects. It provides markerless 3D annotations for both hands. We follow the setup in [25] to divide the training and test sets for pose estimation. + +# 4.2. Implementation + +Handling Low Resolution Our experiment was conducted under two original resolution settings, $32 \times 32$ and $64 \times 64$ . Unless otherwise specified, all experiments use the original resolution of $64 \times 64$ . To avoid additional noise introduced by excessive image interpolation, we set the size of the input image to $64 \times 64$ . We have modified ViT pose-B [49] as the backbone. To adapt it to low resolution inputs, we replaced the patch embedding part with two convolution layers with zero padding, and reduce the number of tokens in ViT to 64. These 64 tokens were rearranged into an $8 \times 8$ feature map during the fusion of temporal features and regression of hand mesh. We adopt a regression-based method in [53] to reconstruct hand from the final feature. + +Handling Temporal Input Our method utilizes the temporal information of historical frames to enhance the current hand reconstruction. We have chosen a sequence length of $T = 6$ with a 5-frames interval based on observations of each datasets. In training, we first extract the visual features and pose-aware tokens of the first 5 frames through the ViT + +backbone without temporal fusion, then compressed them into joint features. During this phase, the parameters of the ViT backbone are not updated. These joint features are subsequently processed through spatio-temporal Transformer to enhance the feature extraction process of last frame. We train all network parameters in this phrase. + +Training We train and evaluate our method on a single server with an Nvidia GeForce RTX 4090 GPU. All experiments are implemented with PyTorch. For all datasets, the batch size is set to 64. Before formal training on each dataset, we fine-tune the backbone using the pre-trained weights of ViTPose-B. We conducted 30 epochs of fine-tuning for DexYCB and HanCo, and 20 epochs for H2O. For DexYCB ,H2O and HanCo, the model is trained for 50,30,50 epochs,respectively. The model is optimized with AdamW solver with parameters $(\beta_{1}$ of $0.9,\beta_{2}$ of 0.999,and weight decay of 0.1). We use an initial learning rate of $5e - 4$ and cosine schedule. We perform data augmentation including scaling $(\pm 10\%)$ , rotation $(\pm 180^{\circ})$ , translation $(\pm 10$ pixels), and color jittering $(\pm 20\%)$ . + +# 4.3. Evaluation Metrics + +We employ the standard metric for hand pose estimation: Mean Per Joint Position Error (MPJPE, J-PE) and Mean Per Vertex Position Error (MPVPE, V-PE), measured after Procrustes alignment, to measure the accuracy of the pose and shape of the estimated hand meshes. In addition, we report the area under the curve of PCK (AUC) score and F-score at 5mm (F@5) and 15mm (F@15) threshold. All evaluation metrics are calculated in the root-aligned space. + +# 4.4. Comparisons with State-of-the-arts + +We compared the proposed low-resolution method with state-of-the-art video-based and single-view methods that use high-resolution input. Additionally, we further use our backbone network (LR-ViT) to establish a low-resolution benchmark for comparison. We trained HMR and VIBE with our backbone on raw low-resolution images to explore the effectiveness of commonly used methods. Comparison results are shown in Tab. 1, Tab. 2 and Tab. 3. In DexYCB, our method performs equally or better than other high-resolution methods at $32 \times 32$ and $64 \times 64$ resolution settings, except for Deformer that uses $512 \times 512$ input resolution. In HanCo and H2O, our method also demonstrates competitive results. We also report the performance of SOTA methods on upsampled low-resolution images in Tab. 4, including cases where models were retrained with upsampled images and where they were not. + +# 4.5. Ablation Study + +# 4.5.1. Backbone settings + +In this section, we experiment with different backbone settings and report their performance with and without tem + +
Token SettingsMonocularTemporal
J-PAV-PAJ-PAV-PA
06.716.75--
16.586.646.036.09
6 (SP)6.196.195.775.74
21 (JP)6.076.055.805.82
+ +Table 5. Ablation study on backbone and pose-aware tokens. + +poral enhancement. Under monocular input, the backbone skips temporal fusion and uses only transformer blocks to process the input. Our experiment further explored the impact of Pose-Aware Tokens. We adopted different settings of introduced tokens to test their impact. Under monocular input, we still supervise these tokens to predict 2D hand poses. The experiment result is shown in Tab. 5. The introduction of temporal enhancement universally improves performance. Under monocular input, the original ViT (w/o pose-aware tokens) performs poorly at low resolutions, while other networks generally achieve performance gains. This demonstrates that supervising intermediate features through specific tokens is more effective than solely employing end-to-end training on the final regression results. Meanwhile, due to the ability of finger based pose-aware tokens (SP) to jointly carry high-level semantic information and low-level visual details, it is more effective in temporal enhancement than a single pose token and skeleton based pose tokens (JP). + +# 4.5.2. Temporal Fusion Schemes + +In this section, we explored the effects of different temporal fusion schemes. We consider three ways to fuse temporal features and visual features: cross attention, single plane encoding, and triplane encoding. The cross attention approach uses joint features as K and V, and visual features as Q to interact through attention mechanisms. The single plane encoding method uses 2D Heatmap to diffuse joint features into a single plane. The single plane feature is also embedded with 2D positional encoding and processed using convolutional layers. For the triplane encoding, we also tested the effect of directly integrating the full $F_{x,z}$ and $F_{y,z}$ with $F_{x,y}$ . The comparison results are presented in Table 6. Our proposed triplane encoding performs best. Compared to the other two methods, the triplane encoding effectively extracts crucial 3D structural information, providing cues for eliminating hand structure ambiguity in visual features. Additionally, using compressed 1D features for interaction is superior to using complete feature. + +# 4.5.3. Position for inserting temporal fusion + +In this section, we further investigated the impact of the insertion position of temporal fusion operations in the backbone. Specifically, we compared two strategies: one where + +![](images/ea5865cdb67302b8e219683c12dbef5ea9e7459dba0c985dc77f2a85a85acd8c.jpg) + +![](images/3b611921dcfbc4b7f40c90bed1e52a22d44dd49b0bd2c45702a6947fbcdf25cf.jpg) + +![](images/fda41ed7e1bb484c068938af830dd6f1205c512d6e973388c58557905312a097.jpg) + +![](images/10dc607d8ed6d72e66823f71c67b908451e90c93b8ac578b9c438be5d4b1ab66.jpg) + +![](images/b10c8d38d58d159712230ae996fca18bf1622f2a9203ade9f2f1df88672b1798.jpg) + +![](images/0ce82f2843f4ec50aaefecb0297a08881aec626394e74a3b4a230dd1814a8c9c.jpg) +(a) Input + +![](images/d5b51293807e8934e5f16a1f81ba84ad11ec9f61667574ec59863089ccf0ed49.jpg) + +![](images/bac03c7ffd1ff9f485b6526cb5e46546604f5d057159966aea80065e4c4c7dfd.jpg) + +![](images/d34a9177f37e4355a58ae2a92528356d5918089222e2e49fd46058463d674b10.jpg) +Figure 4. Qualitative Comparison Results on DexYCB. Regions with significant prediction errors are marked with red circles. + +![](images/3154c17d2689b5d91a116ce97aa424db4cf8406509489f2be28dc5bc99abf225.jpg) +(c) HMR +(e) Ours-Single + +![](images/13cdd829557a961a757a3b7c9ae2c16325e02da24cabccb059446782b1389dbd.jpg) + +![](images/fb37af80419fa58ed51b4650bcee52eefe2069c69b3ac4a77de1f9906bce3092.jpg) + +![](images/80ba1a4c446113d6ec91fd3185a3af21229c628c9e1b8f7ca65b7fbcc9f10dd5.jpg) + +![](images/6f23552623d61819f8fe10e6c7ef1fbc96e963897dd1b9caf2fc1548cee18d11.jpg) + +![](images/0a32d90e2a4d584ded64c1672f5350e9568b489dd708e305f14d298f35d50790.jpg) + +![](images/fd0968ebe67be0ff628f029e2c42a941dd31d9ada077a78002b77762206964e7.jpg) + +![](images/1cae2ef4965d0ff7619dd25e629be4c82828fa627a4f7c54022d8401d9d3aff0.jpg) + +![](images/73625c413d28379546b685d57cbc5e83dcee2d15dcf74a382b1ef35054d85da8.jpg) +(b) GT + +![](images/188df50cd324cc163b8f058910df147172f1d483bc7b596def4fccaea5051702.jpg) + +![](images/5c453b1d37eee1c358d0240da56a1583f3804d2a23ca0f52ea8a55cf40678bae.jpg) + +![](images/a87febfee9d113e4fe8ce382a2f9cbad5fe2b41ef6701e1e32b7bbdaabeae406.jpg) + +![](images/d9a8dd3b049e5178f111bf925a3a9221f309470531c6c999cdfc6f72327409d3.jpg) +(d) VIBE +(f) Ours + +![](images/da1e135f507fb77e750331426461cbcccf1be54cdb43735840e4738d968f5e22.jpg) + +![](images/f724e1dbb66a90f30e710ff9903a9db0c6ec885ceca571c6fab4df09c3ed9ed0.jpg) + +
Fusion SchemeJ-PAV-PAF@5F@15
Cross-attention5.955.9975.898.9
Single plane Encoding5.855.9076.799.0
Triplane Encoding (w/o 1D compress)5.805.8576.999.0
Triplane Encoding5.775.7477.699.0
+ +Table 6. Ablation study of different temporal fusion schemes. + +
Insert PositionJ-PAV-PAF@5F@15
Final layer6.016.0575.198.6
Middle layer5.775.7477.699.0
+ +Table 7. Ablation Study on the Position of Temporal Information Fusion in the Backbone Network. + +temporal features are fused at multiple intermediate levels throughout the network, and another where fusion occurs solely at the final level, as shown in Tab. 7. The experimental results in Tab. 7 demonstrate that fusing temporal features at multiple intermediate levels leads to superior performance compared to only performing fusion at the final level. This suggests that incorporating temporal information at various stages of the network helps to better capture ambiguous local features, contributing to more accurate and consistent hand pose reconstruction in low resolution. + +# 4.6. Qualitative Results + +We present the qualitative comparison results on DexYCB in Fig. 4. Compared with low resolution benchmarks (HMR, VIBE) that use the same LR-ViT as the backbone, our method achieves better alignment of the reconstructed + +hands with low-resolution images. Additionally, by leveraging historic temporal information to supplement missing details in visual features, our approach enhances the ability to cope with local occlusions compare to solely rely on a single frame (see columns 2 and 3 in Fig. 4 (e) and (f)). + +# 5. Conclusion + +In this paper, we propose a temporal enhancement framework for low-resolution hand reconstruction. We represent temporal information through joint features to capture the spatiotemporal correlations between joints. Furthermore, we integrate the temporal context provided by joint features with the spatial context provided by visual features to construct a robust low-resolution representation with 3D spatial awareness. Specifically, we introduce a Triplane feature representation to address the misalignment in semantics and representation forms between joint and visual features, preserving 3D structural information during fusion. Quantitative experiments on multiple datasets have shown that our method achieves comparable performance to high-resolution SOTA methods at much lower resolutions. We also demonstrated the contribution of each design to the scheme performance through ablation study. + +**Limitation and Future Work** Our work did not explicitly model long-term hand motion cues across video sequences, which may result in a lack of constraints on motion consistency. As a result, the model may struggle with maintaining consistent hand shapes over extended sequences. Future work should explore utilizing motion cues from long sequences to better constrain the reconstructed hand shape and improve temporal consistency. + +Acknowledgements This work was supported by the National Natural Science Foundation of China under Grants (62406039, 62101064, 62171057,62201072, U23B2001, 62001054, 62071067), the Ministry of Education and China Mobile Joint Fund (MCM20200202, MCM20180101), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center, BUPT Excellent Ph.D. Students Foundation (CX20241014), in part by the Project funded by China Postdoctoral Science Foundation (2023TQ0039, 2024M750257), the Postdoctoral Fellowship Program of CPSF (GZC20230320). + +# References + +[1] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10843-10852, 2019. 2 +[2] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 666-682, 2018. 2 +[3] Berk Calli, Arjun Singh, Aaron Walsman, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In 2015 international conference on advanced robotics (ICAR), pages 510-517. IEEE, 2015. 6 +[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 4 +[5] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 6 +[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12929–12938, 2021. 1 +[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20544-20554, 2022. 1, 6 +[8] Hoseong Cho, Chanwoo Kim, Jihyeon Kim, Seongyeong Lee, Elkhan Ismayilzada, and Seungryul Baek. Transformer-based unified recognition of two hands manipulating objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4769-4778, 2023. 6 +[9] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consis + +tent 3d human pose and shape from a video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1964-1973, 2021. 3, 6 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3 +[11] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 3 +[12] Qichen Fu, Xingyu Liu, Ran Xu, Juan Carlos Niebles, and Kris M Kitani. Deformer: Dynamic fusion transformer for robust hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23600-23611, 2023. 3, 6 +[13] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10833-10842, 2019. 2 +[14] Peter Hardy, Srinandan Dasmahapatra, and Hansung Kim. Can super resolution be used to improve human pose estimation in low resolution scenarios? arXiv preprint arXiv:2107.02108, 2021. 2, 3 +[15] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 571-580, 2020. 3 +[16] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3 +[17] Weiting Huang, Pengfei Ren, Jingyu Wang, Qi Qi, and Haifeng Sun. Awr: Adaptive weighting regression for 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11061-11068, 2020. 2 +[18] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In Proceedings of the European conference on computer vision (ECCV), pages 118-134, 2018. 2 +[19] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018. 6 +[20] Leyla Khaleghi, Alireza Sepas-Moghaddam, Josh Marshall, and Ali Etemad. Multiview video-based 3-d hand pose estimation. IEEE Transactions on Artificial Intelligence, 4:896-909, 2021. 3 +[21] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF conference on + +computer vision and pattern recognition, pages 1077-1086, 2019. 6 +[22] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 3, 6 +[23] Deying Kong, Linguang Zhang, Liangjian Chen, Haoyu Ma, Xiangyi Yan, Shanlin Sun, Xingwei Liu, Kun Han, and Xiaohui Xie. Identity-aware hand mesh estimation and personalization from rgb images. In European Conference on Computer Vision, pages 536-553. Springer, 2022. 2 +[24] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4990-5000, 2020. 2 +[25] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 6 +[26] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 2, 6 +[27] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12939-12948, 2021. 2, 6 +[28] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 6 +[29] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 3 +[30] Xingyu Liu, Pengfei Ren, Yuanyuan Gao, Jingyu Wang, Haifeng Sun, Qi Qi, Zirui Zhuang, and Jianxin Liao. Keypoint fusion for rgb-d based 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3756-3764, 2024. 1 +[31] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 752-768. Springer, 2020. 1 +[32] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2308-2317, 2022. 2 + +[33] Chihiro Nakatsuka and Satoshi Komorita. Stable 3d human pose estimation in low-resolution videos with a few views. In 2022 International Conference on Robotics and Automation (ICRA), pages 427-433. IEEE, 2022. 3 +[34] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 6 +[35] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 6 +[36] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. Pose-guided hierarchical graph reasoning for 3-d hand pose estimation from a single depth image. IEEE Transactions on Cybernetics, 53(1):315-328, 2021. 2 +[37] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. A dual-branch self-boosting framework for self-supervised 3d hand pose estimation. IEEE Transactions on Image Processing, 31:5052–5066, 2022. 2 +[38] Pengfei Ren, Haifeng Sun, Jiachang Hao, Jingyu Wang, Qi Qi, and Jianxin Liao. Mining multi-view information: a strong self-supervised framework for depth-based 3d hand pose and mesh estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20555-20565, 2022. +[39] Pengfei Ren, Yuchen Chen, Jiachang Hao, Haifeng Sun, Qi Qi, Jingyu Wang, and Jianxin Liao. Two heads are better than one: Image-point cloud network for depth-based 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2163-2171, 2023. 2 +[40] Pengfei Ren, Chao Wen, Xiaozheng Zheng, Zhou Xue, Haifeng Sun, Qi Qi, Jingyu Wang, and Jianxin Liao. Decoupled iterative refinement framework for interacting hands reconstruction from a single rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8014-8025, 2023. 1, 5 +[41] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2 +[42] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 4 +[43] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 2 +[44] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11698-11707, 2021. 2 +[45] Chen Wang, Feng Zhang, Xiatian Zhu, and Shuzhi Sam Ge. + +Low-resolution human pose estimation. Pattern Recognition, 126:108579, 2022. 2, 3 +[46] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4 +[47] Hao Xu, Tianyu Wang, Xiao Tang, and Chi-Wing Fu. H2onet: Hand-occlusion-and-orientation-aware network for real-time 3d hand mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17048–17058, 2023. 6 +[48] Xiangyu Xu, Hao Chen, Francesc Moreno-Noguer, Laszlo A Jeni, and Fernando De la Torre. 3d human pose, shape and texture from low-resolution images and videos. IEEE transactions on pattern analysis and machine intelligence, 44(9): 4490-4504, 2021. 2, 3 +[49] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022. 6 +[50] John Yang, Hyung Jin Chang, Seungeui Lee, and Nojun Kwak. Seqhand: Rgb-sequence-based 3d hand pose and shape estimation. In European Conference on Computer Vision, pages 122-139. Springer, 2020. 3 +[51] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13232-13242, 2022. 2 +[52] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 2 +[53] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Sar: Spatial-aware regression for 3d hand pose and mesh reconstruction from a monocular rgb image. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 99-108. IEEE, 2021. 2, 3, 5, 6 +[54] Xiaozheng Zheng, Chao Wen, Zhou Xue, Pengfei Ren, and Jingyu Wang. Hamuco: Hand pose estimation via multiview collaborative self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20763-20773, 2023. 6 +[55] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3 +[56] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 2 +[57] Zhishan Zhou, Shihao Zhou, Zhi Lv, Minqiang Zou, Yao Tang, and Jiajun Liang. A simple baseline for efficient hand mesh reconstruction. In Proceedings of the IEEE/CVF Con- + +ference on Computer Vision and Pattern Recognition, pages 1367-1376, 2024. 6 +[58] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15085-15099, 2023. 2 +[59] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 2 +[60] Christian Zimmermann, Max Argus, and Thomas Brox. Contrastive representation learning for hand shape estimation. In Pattern Recognition, pages 250-264, Cham, 2021. Springer International Publishing. 2 \ No newline at end of file diff --git a/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/images.zip b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9b9c0070063e7f5605bd87676f40d7e25045395d --- /dev/null +++ b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbc52ef676c5f731ed7f45734fbf47cfd93d595f03e6723b2755f072b4cf7a59 +size 633525 diff --git a/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/layout.json b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..32218010ce0ca1d185fee6d9ed7d34bf6266be31 --- /dev/null +++ b/2025/Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction/layout.json @@ -0,0 +1,9568 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 525, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 525, + 138 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 525, + 138 + ], + "type": "text", + "content": "Pose-Guided Temporal Enhancement for Robust Low-Resolution Hand Reconstruction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 161, + 517, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 161, + 517, + 203 + ], + "spans": [ + { + "bbox": [ + 93, + 161, + 517, + 203 + ], + "type": "text", + "content": "Kaixin Fan, Pengfei Ren\\*, Jingyu Wang," + }, + { + "bbox": [ + 93, + 161, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 93, + 161, + 517, + 203 + ], + "type": "text", + "content": " Haifeng Sun, Qi Qi, Zirui Zhuang, Jianxin Liao \nState Key Laboratory of Networking and Switching Technology, \nBeijing University of Posts and Telecommunications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 205, + 509, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 205, + 509, + 217 + ], + "spans": [ + { + "bbox": [ + 97, + 205, + 509, + 217 + ], + "type": "text", + "content": "{fkxbupt, rpf, wangjingyu, hfsun, qiqi8266, zhuangzirui, liaojx}@bupt.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 270, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 296, + 617 + ], + "type": "text", + "content": "3D hand reconstruction is essential in non-contact human-computer interaction applications, but existing methods struggle with low-resolution images, which occur in slightly distant interactive scenes. Leveraging temporal information can mitigate the limitations of individual low-resolution images that lack detailed appearance information, enhancing the robustness and accuracy of hand reconstruction. Existing temporal methods typically use joint features to represent temporal information, avoiding interference from redundant background information. However, joint features excessively disregard the spatial context of visual features, limiting hand reconstruction accuracy. We propose to integrate temporal joint features with visual features to construct a robust low-resolution visual representation. We adopt Triplane Features, a dense representation with 3D spatial awareness, to bridge the gap between the joint features and visual features that are misaligned in terms of representation form and semantics. Triplane Features are obtained by orthogonally projecting joint features, embedding hand structure information into the 3D spatial context. Furthermore, we compress the spatial information of the three planes into a 2D dense feature though Spatial-Aware Fusion to enhance the visual features. By using enhanced visual features enriched with temporal information for hand reconstruction, our method achieves competitive performance at much lower resolutions compared to state-of-the-art methods operating at high resolution on DexYCB, HanCo and H2O. Code is available at https://github.com/NewbieFan/Temp-LowRes-hand." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 640, + 135, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 640, + 135, + 652 + ], + "spans": [ + { + "bbox": [ + 56, + 640, + 135, + 652 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 661, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 296, + 685 + ], + "type": "text", + "content": "3D hand reconstruction, a crucial component in non-contact human-computer interaction applications, has made signifi-" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 257, + 553, + 302 + ], + "blocks": [ + { + "bbox": [ + 321, + 247, + 374, + 255 + ], + "lines": [ + { + "bbox": [ + 321, + 247, + 374, + 255 + ], + "spans": [ + { + "bbox": [ + 321, + 247, + 374, + 255 + ], + "type": "text", + "content": "(a) No Interaction" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 257, + 553, + 302 + ], + "lines": [ + { + "bbox": [ + 318, + 257, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 318, + 257, + 553, + 302 + ], + "type": "image", + "image_path": "2618034399817112badc1206dd577a8dfd8ade133e6a17c642afce196b696cd4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 302, + 376, + 309 + ], + "lines": [ + { + "bbox": [ + 321, + 302, + 376, + 309 + ], + "spans": [ + { + "bbox": [ + 321, + 302, + 376, + 309 + ], + "type": "text", + "content": "(b) Cross Attention" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 311, + 553, + 388 + ], + "blocks": [ + { + "bbox": [ + 318, + 311, + 553, + 388 + ], + "lines": [ + { + "bbox": [ + 318, + 311, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 318, + 311, + 553, + 388 + ], + "type": "image", + "image_path": "8f0bedb21c5bc037a15ca55228c1445547c7e5523607938d98c418d7af2f7305.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 319, + 397, + 553, + 479 + ], + "blocks": [ + { + "bbox": [ + 320, + 389, + 377, + 397 + ], + "lines": [ + { + "bbox": [ + 320, + 389, + 377, + 397 + ], + "spans": [ + { + "bbox": [ + 320, + 389, + 377, + 397 + ], + "type": "text", + "content": "(c) Triplane Feature" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 397, + 553, + 479 + ], + "lines": [ + { + "bbox": [ + 319, + 397, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 319, + 397, + 553, + 479 + ], + "type": "image", + "image_path": "ac48e0fe3f9bdf291ab41a7e40a1a357e219ef9dde255ddd99628040548fb0e0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 490, + 555, + 567 + ], + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 567 + ], + "type": "text", + "content": "Figure 1. Different way to utilizing joint features. a) Directly reconstruct hand from joint features, relying solely on sparse features without spatial context. b) Fusing joint and visual features through cross-attention, which avoids representation discrepancies but neglects the hand's structural priors. c) Using triplane features to bridge joint and visual features by decoding structural joint information into dense representations that enhance visual features." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "cant advancements in recent years with the advent of large-scale datasets and deep learning technologies. Many studies [6, 7, 30, 31, 34, 40] have focused on reconstructing 3D hands from a single RGB image. In typical interaction scenarios, these methods require the subject to stay close to the capture camera to obtain high-resolution (e.g., " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ") cropped hand images. However, as the interaction distance increases—such as in XR teaching activities within a classroom—the available cropped hand images are limited to lower resolutions (e.g., " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "). Current methods" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22627" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "perform poorly on such low-resolution images, as these images provide fewer cues for addressing joint self-similarity and depth ambiguity, significantly increasing the difficulty of 3D hand reconstruction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 122, + 295, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 122, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 54, + 122, + 295, + 278 + ], + "type": "text", + "content": "To address the low-resolution problem, one intuitive approach is to lift the resolution through super-resolution techniques [14]. However, general super-resolution methods can lead to structural distortions and artifacts in the hand regions, which brings a domain gap between super-resolved image and real high-resolution image. Low-Res Pose [45] adopted confidence-aware learning to supervise continue joint offset fields, thereby alleviating the quantization errors. RSC-Net [48] adopted cross-resolution contrastive learning to align the estimation results of low-resolution images with high-resolution images. However, the performance of these methods is still limited due to the inherent ambiguity resulting from information loss in low resolution." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 281, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 281, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 296, + 544 + ], + "type": "text", + "content": "A single low-resolution image lacks detailed hand appearance, leading to ambiguous visual representation. Continuous frames provide comprehensive hand appearance and motion context, offering strong disambiguation cues for inferring global hand structure. To reduce interference from redundant background information and computational costs, existing temporal methods typically compress single-frame information into joint features [51, 58], and then model the intra- and inter-frame correlation between joints through temporal interaction. However, joint features excessively discard spatial context of visual features, which limits the accuracy of pose estimation. Integrating the spatial context of visual features with the temporal context of joint features can construct robust visual representation with 3D spatial awareness. However, integrating joint features into visual features present two major challenges. Firstly, sparse joint features are not aligned with dense visual features in terms of representation form. Secondly, joint features contain prior geometric structural information in 3D space (e.g., hand scale, skeleton length), while visual features represent detailed appearance information in 2D space (e.g., texture, color), leading to a semantic misalignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we integrate the temporal clue from historical frames into the current visual features to achieve robust and accurate 3D hand pose estimation under low-resolution scenarios. Specifically, we compress temporal information into joint representations, and model intra- and interframe relationships of joints through spatio-temporal interaction. Then, we adopt triplane features[4], a 3D spatial-aware dense visual representation, to bridge the sparse 3D joint representation and the dense 2D visual representation. We orthogonally project the joint features into three individual planes, with each plane independently embedding the sparse joint features into specific spatial contexts. The orthogonal planes ensure comprehensive coverage of 3D hand structure information in a low-redundancy manner. How" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "ever, they still contain redundant spatial information, directly integrating these features into the visual plane leads to the overlap of spatial context from different planes. Thus, we compress the features from two planes orthogonal to the visual plane into 1D global features, guiding each pixel in visual features to integrate distinct spatial context from the other planes, reducing redundancy and clarifying the 3D spatial context. Finally, we reconstruct the hand based on the enhanced visual features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 263 + ], + "type": "text", + "content": "Experimental results on several large-scale hand pose estimation video datasets, DexYCB [5], HanCo [60] and H2O [25], show the proposed method reach competitive performance at much lower resolution compared to both single-frame based and temporal based state-of-the-art (SOTA) methods at high resolution. In summary, our contribution is list as below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 267, + 553, + 399 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 267, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 267, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 314, + 267, + 553, + 326 + ], + "type": "text", + "content": "- We propose a temporal enhancement framework that integrates temporal joint features with visual features to construct a robust visual representation for low-resolution conditions, thereby improving hand reconstruction performance in low-resolution videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 327, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 327, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 314, + 327, + 553, + 374 + ], + "type": "text", + "content": "- We introduced triplane feature encoding to encode joint features into dense feature maps. Through triplane feature, visual features can perceive the prior hand structure information from the temporal joint features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 375, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 375, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 375, + 553, + 399 + ], + "type": "text", + "content": "- Our method reaches competitive result at much lower resolution compared to SOTA methods at high resolution." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 411, + 394, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 411, + 394, + 424 + ], + "spans": [ + { + "bbox": [ + 314, + 411, + 394, + 424 + ], + "type": "text", + "content": "2. Relate Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 432, + 533, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 533, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 533, + 445 + ], + "type": "text", + "content": "2.1. Single Image based Hand Pose Estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "Early works [2, 17, 18, 36, 43, 59] focused on estimating hand pose solely from input images, which suffered from occlusion issues due to the lack of modeling the inherent structure of the hand. MANO [41] introduces a parametric hand model, which contains rich priors of human hands, and now most work focuses on reconstruct this pre-defined hand from image. Some works [1, 23, 32, 37-39, 52, 56] directly predict parameters of MANO, 3D joints and mesh vertices coordinates are then retrieved using the differentiable MANO model. However, predicting MANO parameters directly from a single RGB image is a challenging job, which is difficult for network optimization and is prone to mesh-image misalignment [44], especially in a low-res input condition. To alleviate these issues present in parametric predicting, some works [13, 24, 27] regress the 3D coordinates or occupancy directly for mesh vertices. For example, [53] adopts graph convolution neural network to regress 2.5D Heatmap for every mesh vertices. [26] proposed an End to end framework to recover mesh using transformer. These works show that using a regression-based approach can maintain flexibility while achieving comparable results to parametric methods. Thus, we adopt a regression-based" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22628" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "method described in [53] to reconstruct the hand mesh from enhanced visual features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 105, + 230, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 105, + 230, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 105, + 230, + 118 + ], + "type": "text", + "content": "2.2. Low Resolution Pose Estimation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 123, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 123, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 296, + 399 + ], + "type": "text", + "content": "Recently, a few works have explored on low resolution pose estimation. A intuitive way is leveraging Super Resolution approach [14] to improve the image quality and pose estimation performance. However, the addition super-resolution network brings extra computation cost and may bring unpleasant artifacts to the input image. [45] find that offset-based methods can provide more reliable results in low resolution, and further proposed a confidence-aware learning to selectively weigh the learning of heatmap and offset. [48] controls the scaling of the CNN architecture model using a set of hyperparameters, enforcing it to output scale-invariant features at different resolutions through contrastive learning and self-supervised learning. Although specific learning schemas improve the poor results in low resolution, the features extracted by the network are still constrained by the lack of detailed hand appearance. Other work [33] explored on 3D pose estimation in low-resolution multi-view videos, showing that temporal and multi-view information could be a supplement to alleviate information loss. However, they only correct the prediction under certain adverse conditions, ignoring temporal information can reduce the inherent ambiguity in single low-resolution image and imporve the overall performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 230, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 230, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 230, + 420 + ], + "type": "text", + "content": "2.3. Temporal Hand Pose Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 653 + ], + "type": "text", + "content": "Some recent methods [11, 15, 22, 29, 50, 55] have begun to explore leveraging temporal information in videos to overcome the limitations of single-view images. Hasson et al. [15] use photometric supervision to keep consistency between the re-projected 3D hand and the optical flow. VIBE [22] use a temporal generation network to predict model parameters for each frame, and employs a recurrent network as a discriminator to supervise motion sequence. However, they ignored the spatial and feature correlations in the video. More recently, [20] has harnessed temporal information by propagating frame-wise features with a sequence-to-sequence model based on LSTM [16], but it only provides sparse joint location predictions. TCMR [9] introduces a temporally consistent system that synthesizes implicit features from past and future frames to maintain temporal consistency in the mesh of the current frame. Deformer [12] deforms the hand poses from neighborhood frames using predicted motion to avoid unreliable queries generated with local feature when leveraging self-attention mechanism." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "In low resolutions, these methods are limited to the poor quality of single frame features. Additionally, they lack a focus center when processing temporal features, which may cause some estimation results to deviate from the actual image due to the influence of temporal context." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": "The overview of our method is shown in the Fig. 2. For a given frame " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "I_{T}" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " in an image sequence " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "V = \\{I_{t}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ", we extract temporal cues from the previous " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " frames to enhance the visual feature of the current frame and reconstruct the hand mesh from the enhanced visual feature. Specifically, we first compress the information of each historical frames into joint features and establish their inter- and intra-frame correlations through a spatio-temporal Transformer. Then, we encode the structural information from the enhanced joint features into 3D spatial-aware dense representations through triplane feature encoding to enhance the visual features of the current frame. Finally, we leverage the enhanced visual feature to reconstruct the hand mesh." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 253, + 440, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 440, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 440, + 266 + ], + "type": "text", + "content": "3.1. Static Visual Features" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 270, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 556, + 510 + ], + "type": "text", + "content": "As shown in Fig. 3, we adopt a Vision Transformer (ViT) [10] to extract visual features. During the extraction process, ambiguous local features would propagate from low-level features to subsequent high-level features, leading to ambiguities in both semantic content and fine-grained details in the final visual features. Thus, we incorporate temporal information into intermediate features at different levels to gradually reduce ambiguous local representations and ultimately obtain visual features with a clear perception of hand structure. However, static visual features lack the ability to perceive dynamic temporal information represented in joint features. Therefore, we introduce pose-aware tokens as anchors for fusing dynamic temporal information into a static representation, which are used for further interaction with visual features. These tokens capture hand pose and appearance information from the current visual feature through attention mechanism. We categorize the functionalities of these tokens based on finger semantic to ensure they incorporate both fine-grained visual details (e.g., texture) and high-level semantic content (e.g., skeletal structures)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 517, + 497, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 497, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 497, + 529 + ], + "type": "text", + "content": "3.2. Dynamic Temporal Joint Features" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "We use compact joint features to represent dynamic temporal information from historical frames. Give a single frame, we obtain its visual feature map along with pose-aware tokens through the backbone. Each pose-aware token carries semantic information corresponding to different individual fingers. We further combine local visual features to expand each token into a complete joint feature, enabling fine-grained temporal interaction around each joints. We adopt a FC layer to predict 2D joint coordinates " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^{J \\times 2}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " from pose-aware tokens, and use it to sample joint-centric local features " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "J^{sample} \\in \\mathbb{R}^{J \\times C}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " from the visual feature map. The joint coordinates is encoded as pose feature " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "J^{pose}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": ", and then add to the sampled feature " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "J^{sample}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " to obtain the initial joint feature " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "J^{init} = J^{pose} + J^{region}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": ". We concatenate the joint features from each frame in the upstream sequence" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "type": "text", + "content": "22629" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 68, + 538, + 179 + ], + "blocks": [ + { + "bbox": [ + 67, + 68, + 538, + 179 + ], + "lines": [ + { + "bbox": [ + 67, + 68, + 538, + 179 + ], + "spans": [ + { + "bbox": [ + 67, + 68, + 538, + 179 + ], + "type": "image", + "image_path": "faafee0f03e7d489097e453072fb8ebd15ec990db44bf1bb8932c1c5fc22d5ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 182, + 538, + 290 + ], + "blocks": [ + { + "bbox": [ + 67, + 182, + 538, + 290 + ], + "lines": [ + { + "bbox": [ + 67, + 182, + 538, + 290 + ], + "spans": [ + { + "bbox": [ + 67, + 182, + 538, + 290 + ], + "type": "image", + "image_path": "2eb451d28a4708f478ee48e5c5e4b52db636d23418b8290c9329ecf0eaeecfa5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 555, + 357 + ], + "type": "text", + "content": "Figure 2. Overview of Our Framework. We first construct joint features from the sampled past frame features and 2D poses using a Spatial-Temporal Transformer (a), capturing the temporal dynamics and spatial relationships between joints. We then use these joint features to enhance the visual features of the current frame (b) by converting them into spatial-aware dense features through Tripalne Feature Encoding (c). In (c), we fuse different joint features (from a) using pose-aware tokens (from b), and project them into three orthogonal planes to obtain the Triplane features. These features are then compressed into a single plane through space-aware fusion." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 374, + 282, + 473 + ], + "blocks": [ + { + "bbox": [ + 59, + 374, + 282, + 473 + ], + "lines": [ + { + "bbox": [ + 59, + 374, + 282, + 473 + ], + "spans": [ + { + "bbox": [ + 59, + 374, + 282, + 473 + ], + "type": "image", + "image_path": "e3485bd1c5523836e146b411c8b2d282169e5d2e5f5d8a8bd821da104457b043.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 479, + 246, + 491 + ], + "lines": [ + { + "bbox": [ + 104, + 479, + 246, + 491 + ], + "spans": [ + { + "bbox": [ + 104, + 479, + 246, + 491 + ], + "type": "text", + "content": "Figure 3. Backbone of our framework." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "text", + "content": "to form the initial representation of temporal information " + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "inline_equation", + "content": "J^{temp} = \\text{Concat}(J_t^{init}) \\in \\mathbb{R}^{T \\times J \\times C}" + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "text", + "content": ". Then, we adopt a spatio-temporal Transformer to establish inter- and intra-frame correlation between these joints. The spatio-temporal Transformer is a combination of alternately stacked temporal and spatial attention modules. In the spatial attention module, attention interactions occur among different joints within the same joint feature:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 613, + 295, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 613, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 85, + 613, + 295, + 628 + ], + "type": "interline_equation", + "content": "S - M H S A (J ^ {t e m p}) = M H S A \\left(Q _ {j}, K _ {j}, V _ {j}\\right) \\tag {1}", + "image_path": "f502178a8493d048f082a44abf5dff6a36b6408a99dea1dacf14804f010ec315.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 632, + 295, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 632, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 67, + 632, + 295, + 648 + ], + "type": "interline_equation", + "content": "Q _ {j} = W _ {q} J _ {j} ^ {\\text {t e m p}}, K _ {j} = W _ {k} J _ {j} ^ {\\text {t e m p}}, V _ {j} = W _ {v} J _ {j} ^ {\\text {t e m p}} \\tag {2}", + "image_path": "2061ab139cbbba004debe427a93d3e53822919d95f520867ec6e9db78279a497.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 650, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 295, + 674 + ], + "type": "text", + "content": "In the temporal attention module, attention interactions occur among the same joint within different joint features:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 681, + 295, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 681, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 86, + 681, + 295, + 695 + ], + "type": "interline_equation", + "content": "T - M H S A \\left(J ^ {t e m p}\\right) = M H S A \\left(Q _ {t}, K _ {t}, V _ {t}\\right) \\tag {3}", + "image_path": "00d28c5d53e5715a9ec36d4355a43c61278d2a0422bff8bf37f5ee5869d2ab5a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 67, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "Q _ {t} = W _ {q} J _ {t} ^ {t e m p}, K _ {j} = W _ {k} J _ {t} ^ {t e m p}, V _ {j} = W _ {v} J _ {t} ^ {t e m p} \\quad (4)", + "image_path": "41242027e66f2c807f88c526cb9550ca54c4e14b3685d6e53825de80020449aa.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "text", + "content": " are learnable weight matrices used to transform the input features into queries, keys, and values, respectively. " + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "inline_equation", + "content": "J_{j}^{temp}" + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "text", + "content": " represents the joint features for the " + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "text", + "content": "-th joint across all time steps, while " + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "inline_equation", + "content": "J_{t}^{temp}" + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "text", + "content": " represents the joint features for the " + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 376, + 555, + 439 + ], + "type": "text", + "content": "-th time step across all joints." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 445, + 463, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 463, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 463, + 458 + ], + "type": "text", + "content": "3.3. Triplane Feature Encoding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 654 + ], + "type": "text", + "content": "Joint features contain temporal context of the current hand pose, which can be integrated into visual features to construct low-resolution robust visual representations. However, it is difficult to fuse joint features with visual features that are not aligned in semantic content and representation form. A straightforward approach is to treat joint features as global tokens and interact them with visual features through cross attention. However, visual features struggle to effectively perceive the 3D spatial properties of joint features through attention mechanisms. Thus, we propose to use triplane features as an intermediate representation to enhance the visual features. Unlike [4, 42, 46] that constructs dense Triplane features from images for spatial indexing, we leverage sparse joint features to generate triplane features, embedding 3D geometric attributes of the skeleton into specific spatial context." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Specifically, we first copy each pose token to the corresponding finger joints as a Pose Template " + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "inline_equation", + "content": "J^{P} \\in \\mathbb{R}^{J \\times C}" + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": ", and use it to fuse dynamic joint features into a static one. Then splat it onto three orthogonal planes as an intermediate representation for interaction with visual features. Fur" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22630" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "thermore, we integrate the information from different planes into a single plane through spatial-aware fusion, allowing for direct fusion with the visual features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 310 + ], + "type": "text", + "content": "Triplane Projection of Joint Features Visual features represent specific spatial information in the form of dense representations. Transferring joint features into the 2D visual feature space through 3D-2D pose projection [40] can provide effective disambiguation cues for hand poses. However, a single projection plane cannot capture the complete 3D hand structure implied in joint features, resulting in the loss of crucial structural disambiguation cues. Therefore, we propose to project joint features onto three orthogonal planes to fully preserve 3D hand structure information. We first decode the 3D hand pose from the fused joint features, and then use it to generate distance-aware heatmaps corresponding to the x-y, y-z, and z-x planes. Then, we splat the fused joint feature across the three planes through these heatmaps, and collect the features from each plane as the Triplane Feature maps " + }, + { + "bbox": [ + 55, + 118, + 296, + 310 + ], + "type": "inline_equation", + "content": "F_{\\text{triplane}} = \\{F_{x,y}, F_{x,z}, F_{y,z}\\}" + }, + { + "bbox": [ + 55, + 118, + 296, + 310 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 617 + ], + "type": "text", + "content": "Spatial-Aware Fusion Triplane encoding transforms the fused joint feature into dense 2D features, eliminating the misalignment in representation form with visual features. However, the initial plane features derived from projecting sparse joint information primarily encode geometric positions while overlooking the relative spatial relationships and interactions between neighboring joints, such as occlusions or depth variations, which are crucial for resolving structural ambiguities in visual features. Additionally, these plane features are spatially non-unified: The x-y plane shares the same spatial dimension with visual features, while the y-z and x-z planes are spatially orthogonal to visual features. To embed the complete 3D hand structure information represented by the Triplane features into the visual features in 2D space, it is essential to compress the y-z and x-z plane features as supplementary depth information relative to the x-y plane into the pixel-wise semantic content. However, directly integrating the y-z and x-z plane features into the x-y plane via cross-attention is inefficient and introduces redundant information. Therefore, we compress the y-z and x-z features along distinct directions into 1D global representations. This compression allows us to guide each pixel in the x-y plane to selectively integrate spatial context from the other two planes, reducing unnecessary overlap while preserving key spatial dependencies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "Specifically, we first add 2D positional encoding to each plane feature and apply convolution layers to each plane feature to capture local spatial information (See Fig. 2 (c) project). Next, we perform summation along different directions for " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{x,z}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{y,z}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " to obtain the 1D features " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_x, V_x" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_y, V_y" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " (See Fig. 2 (c) compress), where " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " denotes summation along the Z direction, and " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " denotes summation along the X or Y direction. Finally, we integrate these" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "features into the pixel-wise content of " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "F_{x,y}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " through crossattention to obtain spatially compressed features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 102, + 400, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 400, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 400, + 114 + ], + "type": "text", + "content": "3.4. Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 119, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 119, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 119, + 554, + 156 + ], + "type": "text", + "content": "We supervise hand mesh regression through reconstruct loss, supervise temporal enhancement process through token pose loss and temporal pose loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 159, + 416, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 416, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 416, + 170 + ], + "type": "text", + "content": "3.4.1. Reconstruct Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 174, + 554, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 554, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 554, + 222 + ], + "type": "text", + "content": "We used the same reconstruction loss as [53], including vertex/joint loss and mesh smoothing loss. Vertex/Joint loss supervises the prediction error of each mesh and joint point through L1 loss, i.e" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 226, + 554, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 226, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 348, + 226, + 554, + 259 + ], + "type": "interline_equation", + "content": "L _ {V / J} = \\sum_ {i = 1} ^ {N} \\| \\nu_ {i} - \\nu_ {i} ^ {*} \\| _ {1} + \\sum_ {i = 1} ^ {K} \\| j _ {i} - j ^ {*} \\| _ {1} \\tag {5}", + "image_path": "3fb95472952599e1881eda6912e6a00855a878d4dff1a3516f0ce770a1894ed3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "inline_equation", + "content": "\\nu_{i}" + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "text", + "content": " vertex of the mesh, " + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "inline_equation", + "content": "j_{i}" + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "text", + "content": " joint of the pose, and " + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "text", + "content": " denotes the ground-truth." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 554, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 335 + ], + "type": "text", + "content": "Mesh smooth loss includes supervision of the surface normal vectors and edge lengths to obtain visually reasonable results. The loss functions for normal vector and edge length are defined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 340, + 553, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 340, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 348, + 340, + 553, + 369 + ], + "type": "interline_equation", + "content": "L _ {\\text {n o r m a l}} = \\sum_ {f} \\sum_ {\\{i, j \\} \\subset f} | \\left\\langle \\frac {\\nu_ {i} - \\nu_ {j}}{\\| \\nu_ {i} - \\nu_ {j} \\| _ {2}}, n _ {f} ^ {*} \\right\\rangle | \\tag {6}", + "image_path": "a04ab3b69f342c006a490c251a19ee251a279b2619c25847e8d295ffb154b5dc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 331, + 373, + 553, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 373, + 553, + 400 + ], + "spans": [ + { + "bbox": [ + 331, + 373, + 553, + 400 + ], + "type": "interline_equation", + "content": "L _ {e d g e} = \\sum_ {f} \\sum_ {\\{i, j \\} \\subset f} \\left| \\left\\| \\nu_ {i} - \\nu_ {j} \\right\\| _ {2} - \\left\\| \\nu_ {i} ^ {*} - \\nu_ {j} ^ {*} \\right\\| _ {2} \\right| \\tag {7}", + "image_path": "d01ec45b0fd806fa540d536af2a0191fd4290adc030e4f156322125719dd8e7b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "inline_equation", + "content": "n_f" + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "text", + "content": " indicate a mesh face and unit normal vector of face " + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "inline_equation", + "content": "n_f^*" + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "text", + "content": " is computed from ground-truth vertex sets " + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "inline_equation", + "content": "V^*" + }, + { + "bbox": [ + 313, + 402, + 554, + 437 + ], + "type": "text", + "content": ", where * denotes ground-truth." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 438, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 462 + ], + "type": "text", + "content": "The final hand reconstruction task loss combines the vertex/joint loss and the mesh smoothness loss as " + }, + { + "bbox": [ + 313, + 438, + 554, + 462 + ], + "type": "inline_equation", + "content": "L_{hand}" + }, + { + "bbox": [ + 313, + 438, + 554, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 466, + 384, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 384, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 384, + 476 + ], + "type": "text", + "content": "3.4.2. Pose Loss" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 480, + 554, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 574 + ], + "type": "text", + "content": "For pose-aware tokens, we enable them to perceive pose changes by supervising their predicted 2D poses. During the temporal fusion process, we use the fused joint feature to predict 2.5D hand pose coordinates for triplane project. We supervise the predicted 2.5D coordinates at each temporal fusion stage to ensure accurate decoding of hand structure information. These two loss functions are represented as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 376, + 574, + 553, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 574, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 376, + 574, + 553, + 605 + ], + "type": "interline_equation", + "content": "L _ {\\text {t o k e n}} = \\sum_ {i} ^ {K} \\| j _ {i} ^ {2 d} - j ^ {2 d *} \\| _ {1} \\tag {8}", + "image_path": "ea5912e6011cd906754070ed4598d42415c140b5ad06ff99cc45d941c1cda60c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 354, + 609, + 553, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 609, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 354, + 609, + 553, + 642 + ], + "type": "interline_equation", + "content": "L _ {\\text {t e m p o r a l}} = \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {K} \\| j _ {t, i} ^ {2. 5 d} - j ^ {2. 5 d *} \\| _ {1} \\tag {9}", + "image_path": "3008a5128d5e5ad4088e76f7b8eaae1c05a7f2c6ee08ebc02beb2a60a7ce27b3.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 643, + 554, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 554, + 678 + ], + "type": "text", + "content": "The final training loss of our framework, which combines reconstruction loss and pose loss, can be expressed as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 364, + 685, + 553, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 685, + 553, + 697 + ], + "spans": [ + { + "bbox": [ + 364, + 685, + 553, + 697 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {h a n d}} + L _ {\\text {t o k e n}} + L _ {\\text {t e m p o r a l}} \\tag {10}", + "image_path": "7adeee1c680bbbfcf60973d2707f954226d521871879b1b8750b3f26ecef5ce0.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22631" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 69, + 552, + 242 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 552, + 242 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 552, + 242 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 552, + 242 + ], + "type": "table", + "html": "
MethodsInputResolutionPA-MPJPE(↓)PA-MPVPE(↓)AUC(↑)F@5(↑)F@15(↑)
METRO[26]Monocular2566.99----
Liu et. al.[28]Monocular2566.60----
MeshGraphormer[27]Monocular2566.41----
HandOccNet[34]Monocular2565.805.5088.478.099.0
H2ONet[47]Monocular1285.705.5089.180.199.0
SimpleHand[57]Monocular2565.505.50---
VIBE[22]Sequence2246.43-87.1--
TCMR[9]Sequence2246.28-87.5--
Deformer[12]Sequence5125.22-89.6--
HMR[19] + LR-ViTMonocular646.456.2487.373.098.5
VIBE[22] + LR-ViTSequence646.226.0987.574.198.7
OursSequence645.725.5288.578.498.7
OursSequence326.035.8288.076.598.6
", + "image_path": "b782c227021313d6f0e6c1c8fe95d983ac9a1b83b56261a7f8a020e464af09f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 280, + 284, + 356 + ], + "blocks": [ + { + "bbox": [ + 177, + 250, + 432, + 261 + ], + "lines": [ + { + "bbox": [ + 177, + 250, + 432, + 261 + ], + "spans": [ + { + "bbox": [ + 177, + 250, + 432, + 261 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art on the DexYCB dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 280, + 284, + 356 + ], + "lines": [ + { + "bbox": [ + 67, + 280, + 284, + 356 + ], + "spans": [ + { + "bbox": [ + 67, + 280, + 284, + 356 + ], + "type": "table", + "html": "
MethodsInputMPJPEPA-MPJPE
Cho et al.[8]Sequence17.20-
HandOccNetMonocular16.196.13
HMR + LR-ViTMonocular30.727.49
VIBE + LR-ViTSequence27.897.05
OursSequence18.156.11
", + "image_path": "9eaaa7df928d1037c1fb5d554de23bd881def84edc1d5f5b7b5a3af7e3315e3d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 387, + 294, + 487 + ], + "blocks": [ + { + "bbox": [ + 56, + 364, + 295, + 375 + ], + "lines": [ + { + "bbox": [ + 56, + 364, + 295, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 295, + 375 + ], + "type": "text", + "content": "Table 2. Comparison with the state-of-the-art on the H2O dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 387, + 294, + 487 + ], + "lines": [ + { + "bbox": [ + 58, + 387, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 58, + 387, + 294, + 487 + ], + "type": "table", + "html": "
MethodsInputJ-PAV-PAF@5F@15
HaMuCo[54]Monocular7.00---
EpipolarPose[21]Monocular6.10---
MobRecon[7]Monocular5.705.81--
HaMeR[35]Monocular5.995.6177.899.0
HMR + LR-ViTMonocular6.886.4473.498.7
VIBE + LR-ViTSequence6.796.2775.098.8
OursSequence5.775.7477.699.0
", + "image_path": "47bbca064ee7cbe27132a949c17e1ea7d76a177f62fcfd27ba08772405350471.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 59, + 539, + 294, + 576 + ], + "blocks": [ + { + "bbox": [ + 55, + 495, + 295, + 529 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 529 + ], + "type": "text", + "content": "Table 3. Comparison results on the HanCo dataset. The evaluation results for HaMuCo and EpipolarPose using monocular input are derived from [54]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 539, + 294, + 576 + ], + "lines": [ + { + "bbox": [ + 59, + 539, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 59, + 539, + 294, + 576 + ], + "type": "table", + "html": "
RetrainedHanOccNetSimpleHandHaMeROurs*
×6.63-8.70-
✓5.985.92-5.72
", + "image_path": "ed2bfbe8414d1ad597966cbe72144970b4cb4e901ff977d105f3379a64cec271.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 584, + 295, + 608 + ], + "lines": [ + { + "bbox": [ + 55, + 584, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 295, + 608 + ], + "type": "text", + "content": "Table 4. SOTA performance (P-MPJPE) on interpolated low-resolution images (DexYCB). *Using raw low-resolution inputs." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 118, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 118, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 118, + 659 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": "DexYCB contains multi-camera videos of humans grasping 20 YCB [3] objects. It consists of 1000 sequences with 582K frames captured by cameras from 8 different views. During the data collection, the hands are frequently oc" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 282, + 555, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 307 + ], + "type": "text", + "content": "cluded during hand-object interaction. We use the dataset split setup s0 from [5] for the training and test sets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 308, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 392 + ], + "type": "text", + "content": "HanCo consists of 1,517 videos with multiple views and camera calibration. It has 860,304 frames in total, i.e. 107,538 time-step per view. Since HanCo does not have an official train/test split, we follow [54] use the first 1,200 sequences for training and validation and the last 317 sequences for testing in all experiments for fair comparisons. We conduct all ablation studies on this dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 393, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 453 + ], + "type": "text", + "content": "H2O includes data from four participants performing 36 distinct actions involving interactions with eight different objects. It provides markerless 3D annotations for both hands. We follow the setup in [25] to divide the training and test sets for pose estimation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 461, + 411, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 411, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 411, + 474 + ], + "type": "text", + "content": "4.2. Implementation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "text", + "content": "Handling Low Resolution Our experiment was conducted under two original resolution settings, " + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "text", + "content": ". Unless otherwise specified, all experiments use the original resolution of " + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "text", + "content": ". To avoid additional noise introduced by excessive image interpolation, we set the size of the input image to " + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "text", + "content": ". We have modified ViT pose-B [49] as the backbone. To adapt it to low resolution inputs, we replaced the patch embedding part with two convolution layers with zero padding, and reduce the number of tokens in ViT to 64. These 64 tokens were rearranged into an " + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 479, + 555, + 634 + ], + "type": "text", + "content": " feature map during the fusion of temporal features and regression of hand mesh. We adopt a regression-based method in [53] to reconstruct hand from the final feature." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Handling Temporal Input Our method utilizes the temporal information of historical frames to enhance the current hand reconstruction. We have chosen a sequence length of " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "T = 6" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " with a 5-frames interval based on observations of each datasets. In training, we first extract the visual features and pose-aware tokens of the first 5 frames through the ViT" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22632" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "backbone without temporal fusion, then compressed them into joint features. During this phase, the parameters of the ViT backbone are not updated. These joint features are subsequently processed through spatio-temporal Transformer to enhance the feature extraction process of last frame. We train all network parameters in this phrase." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "text", + "content": "Training We train and evaluate our method on a single server with an Nvidia GeForce RTX 4090 GPU. All experiments are implemented with PyTorch. For all datasets, the batch size is set to 64. Before formal training on each dataset, we fine-tune the backbone using the pre-trained weights of ViTPose-B. We conducted 30 epochs of fine-tuning for DexYCB and HanCo, and 20 epochs for H2O. For DexYCB ,H2O and HanCo, the model is trained for 50,30,50 epochs,respectively. The model is optimized with AdamW solver with parameters " + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "inline_equation", + "content": "(\\beta_{1}" + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "inline_equation", + "content": "0.9,\\beta_{2}" + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "text", + "content": " of 0.999,and weight decay of 0.1). We use an initial learning rate of " + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "inline_equation", + "content": "5e - 4" + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "text", + "content": " and cosine schedule. We perform data augmentation including scaling " + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "inline_equation", + "content": "(\\pm 10\\%)" + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "inline_equation", + "content": "(\\pm 180^{\\circ})" + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "text", + "content": ", translation " + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "inline_equation", + "content": "(\\pm 10" + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "text", + "content": " pixels), and color jittering " + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "inline_equation", + "content": "(\\pm 20\\%)" + }, + { + "bbox": [ + 55, + 149, + 295, + 317 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 168, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 168, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 168, + 335 + ], + "type": "text", + "content": "4.3. Evaluation Metrics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 437 + ], + "type": "text", + "content": "We employ the standard metric for hand pose estimation: Mean Per Joint Position Error (MPJPE, J-PE) and Mean Per Vertex Position Error (MPVPE, V-PE), measured after Procrustes alignment, to measure the accuracy of the pose and shape of the estimated hand meshes. In addition, we report the area under the curve of PCK (AUC) score and F-score at 5mm (F@5) and 15mm (F@15) threshold. All evaluation metrics are calculated in the root-aligned space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 442, + 243, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 243, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 243, + 454 + ], + "type": "text", + "content": "4.4. Comparisons with State-of-the-arts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 651 + ], + "type": "text", + "content": "We compared the proposed low-resolution method with state-of-the-art video-based and single-view methods that use high-resolution input. Additionally, we further use our backbone network (LR-ViT) to establish a low-resolution benchmark for comparison. We trained HMR and VIBE with our backbone on raw low-resolution images to explore the effectiveness of commonly used methods. Comparison results are shown in Tab. 1, Tab. 2 and Tab. 3. In DexYCB, our method performs equally or better than other high-resolution methods at " + }, + { + "bbox": [ + 55, + 460, + 295, + 651 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 460, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 295, + 651 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 55, + 460, + 295, + 651 + ], + "type": "text", + "content": " resolution settings, except for Deformer that uses " + }, + { + "bbox": [ + 55, + 460, + 295, + 651 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 460, + 295, + 651 + ], + "type": "text", + "content": " input resolution. In HanCo and H2O, our method also demonstrates competitive results. We also report the performance of SOTA methods on upsampled low-resolution images in Tab. 4, including cases where models were retrained with upsampled images and where they were not." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 657, + 149, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 149, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 149, + 670 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 674, + 160, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 160, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 160, + 686 + ], + "type": "text", + "content": "4.5.1. Backbone settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "In this section, we experiment with different backbone settings and report their performance with and without tem" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 349, + 70, + 521, + 145 + ], + "blocks": [ + { + "bbox": [ + 349, + 70, + 521, + 145 + ], + "lines": [ + { + "bbox": [ + 349, + 70, + 521, + 145 + ], + "spans": [ + { + "bbox": [ + 349, + 70, + 521, + 145 + ], + "type": "table", + "html": "
Token SettingsMonocularTemporal
J-PAV-PAJ-PAV-PA
06.716.75--
16.586.646.036.09
6 (SP)6.196.195.775.74
21 (JP)6.076.055.805.82
", + "image_path": "5d2bb487c5dfad6c6c2900818a0fbe4cc0ac7efb74ff1f7c90feed8d53e890ff.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 323, + 154, + 544, + 165 + ], + "lines": [ + { + "bbox": [ + 323, + 154, + 544, + 165 + ], + "spans": [ + { + "bbox": [ + 323, + 154, + 544, + 165 + ], + "type": "text", + "content": "Table 5. Ablation study on backbone and pose-aware tokens." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 187, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 415 + ], + "type": "text", + "content": "poral enhancement. Under monocular input, the backbone skips temporal fusion and uses only transformer blocks to process the input. Our experiment further explored the impact of Pose-Aware Tokens. We adopted different settings of introduced tokens to test their impact. Under monocular input, we still supervise these tokens to predict 2D hand poses. The experiment result is shown in Tab. 5. The introduction of temporal enhancement universally improves performance. Under monocular input, the original ViT (w/o pose-aware tokens) performs poorly at low resolutions, while other networks generally achieve performance gains. This demonstrates that supervising intermediate features through specific tokens is more effective than solely employing end-to-end training on the final regression results. Meanwhile, due to the ability of finger based pose-aware tokens (SP) to jointly carry high-level semantic information and low-level visual details, it is more effective in temporal enhancement than a single pose token and skeleton based pose tokens (JP)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 423, + 454, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 454, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 454, + 434 + ], + "type": "text", + "content": "4.5.2. Temporal Fusion Schemes" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 438, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 654 + ], + "type": "text", + "content": "In this section, we explored the effects of different temporal fusion schemes. We consider three ways to fuse temporal features and visual features: cross attention, single plane encoding, and triplane encoding. The cross attention approach uses joint features as K and V, and visual features as Q to interact through attention mechanisms. The single plane encoding method uses 2D Heatmap to diffuse joint features into a single plane. The single plane feature is also embedded with 2D positional encoding and processed using convolutional layers. For the triplane encoding, we also tested the effect of directly integrating the full " + }, + { + "bbox": [ + 313, + 438, + 555, + 654 + ], + "type": "inline_equation", + "content": "F_{x,z}" + }, + { + "bbox": [ + 313, + 438, + 555, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 555, + 654 + ], + "type": "inline_equation", + "content": "F_{y,z}" + }, + { + "bbox": [ + 313, + 438, + 555, + 654 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 438, + 555, + 654 + ], + "type": "inline_equation", + "content": "F_{x,y}" + }, + { + "bbox": [ + 313, + 438, + 555, + 654 + ], + "type": "text", + "content": ". The comparison results are presented in Table 6. Our proposed triplane encoding performs best. Compared to the other two methods, the triplane encoding effectively extracts crucial 3D structural information, providing cues for eliminating hand structure ambiguity in visual features. Additionally, using compressed 1D features for interaction is superior to using complete feature." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 662, + 502, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 502, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 502, + 673 + ], + "type": "text", + "content": "4.5.3. Position for inserting temporal fusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "In this section, we further investigated the impact of the insertion position of temporal fusion operations in the backbone. Specifically, we compared two strategies: one where" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22633" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 115, + 129 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 115, + 129 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 115, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 115, + 129 + ], + "type": "image", + "image_path": "ea5865cdb67302b8e219683c12dbef5ea9e7459dba0c985dc77f2a85a85acd8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 118, + 70, + 177, + 129 + ], + "blocks": [ + { + "bbox": [ + 118, + 70, + 177, + 129 + ], + "lines": [ + { + "bbox": [ + 118, + 70, + 177, + 129 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 177, + 129 + ], + "type": "image", + "image_path": "3b611921dcfbc4b7f40c90bed1e52a22d44dd49b0bd2c45702a6947fbcdf25cf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 179, + 70, + 239, + 129 + ], + "blocks": [ + { + "bbox": [ + 179, + 70, + 239, + 129 + ], + "lines": [ + { + "bbox": [ + 179, + 70, + 239, + 129 + ], + "spans": [ + { + "bbox": [ + 179, + 70, + 239, + 129 + ], + "type": "image", + "image_path": "fda41ed7e1bb484c068938af830dd6f1205c512d6e973388c58557905312a097.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 70, + 302, + 129 + ], + "blocks": [ + { + "bbox": [ + 242, + 70, + 302, + 129 + ], + "lines": [ + { + "bbox": [ + 242, + 70, + 302, + 129 + ], + "spans": [ + { + "bbox": [ + 242, + 70, + 302, + 129 + ], + "type": "image", + "image_path": "10dc607d8ed6d72e66823f71c67b908451e90c93b8ac578b9c438be5d4b1ab66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 140, + 115, + 199 + ], + "blocks": [ + { + "bbox": [ + 56, + 140, + 115, + 199 + ], + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 199 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 199 + ], + "type": "image", + "image_path": "b10c8d38d58d159712230ae996fca18bf1622f2a9203ade9f2f1df88672b1798.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 118, + 140, + 177, + 199 + ], + "blocks": [ + { + "bbox": [ + 164, + 131, + 194, + 140 + ], + "lines": [ + { + "bbox": [ + 164, + 131, + 194, + 140 + ], + "spans": [ + { + "bbox": [ + 164, + 131, + 194, + 140 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 118, + 140, + 177, + 199 + ], + "lines": [ + { + "bbox": [ + 118, + 140, + 177, + 199 + ], + "spans": [ + { + "bbox": [ + 118, + 140, + 177, + 199 + ], + "type": "image", + "image_path": "0ce82f2843f4ec50aaefecb0297a08881aec626394e74a3b4a230dd1814a8c9c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 179, + 140, + 239, + 199 + ], + "blocks": [ + { + "bbox": [ + 179, + 140, + 239, + 199 + ], + "lines": [ + { + "bbox": [ + 179, + 140, + 239, + 199 + ], + "spans": [ + { + "bbox": [ + 179, + 140, + 239, + 199 + ], + "type": "image", + "image_path": "d5b51293807e8934e5f16a1f81ba84ad11ec9f61667574ec59863089ccf0ed49.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 242, + 140, + 302, + 199 + ], + "blocks": [ + { + "bbox": [ + 242, + 140, + 302, + 199 + ], + "lines": [ + { + "bbox": [ + 242, + 140, + 302, + 199 + ], + "spans": [ + { + "bbox": [ + 242, + 140, + 302, + 199 + ], + "type": "image", + "image_path": "bac03c7ffd1ff9f485b6526cb5e46546604f5d057159966aea80065e4c4c7dfd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 56, + 210, + 115, + 269 + ], + "blocks": [ + { + "bbox": [ + 56, + 210, + 115, + 269 + ], + "lines": [ + { + "bbox": [ + 56, + 210, + 115, + 269 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 115, + 269 + ], + "type": "image", + "image_path": "d34a9177f37e4355a58ae2a92528356d5918089222e2e49fd46058463d674b10.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 289, + 531, + 301 + ], + "lines": [ + { + "bbox": [ + 77, + 289, + 531, + 301 + ], + "spans": [ + { + "bbox": [ + 77, + 289, + 531, + 301 + ], + "type": "text", + "content": "Figure 4. Qualitative Comparison Results on DexYCB. Regions with significant prediction errors are marked with red circles." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 118, + 210, + 177, + 269 + ], + "blocks": [ + { + "bbox": [ + 164, + 201, + 194, + 209 + ], + "lines": [ + { + "bbox": [ + 164, + 201, + 194, + 209 + ], + "spans": [ + { + "bbox": [ + 164, + 201, + 194, + 209 + ], + "type": "text", + "content": "(c) HMR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 118, + 210, + 177, + 269 + ], + "lines": [ + { + "bbox": [ + 118, + 210, + 177, + 269 + ], + "spans": [ + { + "bbox": [ + 118, + 210, + 177, + 269 + ], + "type": "image", + "image_path": "3154c17d2689b5d91a116ce97aa424db4cf8406509489f2be28dc5bc99abf225.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 270, + 205, + 280 + ], + "lines": [ + { + "bbox": [ + 153, + 270, + 205, + 280 + ], + "spans": [ + { + "bbox": [ + 153, + 270, + 205, + 280 + ], + "type": "text", + "content": "(e) Ours-Single" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 179, + 210, + 239, + 269 + ], + "blocks": [ + { + "bbox": [ + 179, + 210, + 239, + 269 + ], + "lines": [ + { + "bbox": [ + 179, + 210, + 239, + 269 + ], + "spans": [ + { + "bbox": [ + 179, + 210, + 239, + 269 + ], + "type": "image", + "image_path": "13cdd829557a961a757a3b7c9ae2c16325e02da24cabccb059446782b1389dbd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 242, + 210, + 302, + 269 + ], + "blocks": [ + { + "bbox": [ + 242, + 210, + 302, + 269 + ], + "lines": [ + { + "bbox": [ + 242, + 210, + 302, + 269 + ], + "spans": [ + { + "bbox": [ + 242, + 210, + 302, + 269 + ], + "type": "image", + "image_path": "fb37af80419fa58ed51b4650bcee52eefe2069c69b3ac4a77de1f9906bce3092.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 367, + 129 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 367, + 129 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 367, + 129 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 367, + 129 + ], + "type": "image", + "image_path": "80ba1a4c446113d6ec91fd3185a3af21229c628c9e1b8f7ca65b7fbcc9f10dd5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 371, + 70, + 430, + 129 + ], + "blocks": [ + { + "bbox": [ + 371, + 70, + 430, + 129 + ], + "lines": [ + { + "bbox": [ + 371, + 70, + 430, + 129 + ], + "spans": [ + { + "bbox": [ + 371, + 70, + 430, + 129 + ], + "type": "image", + "image_path": "6f23552623d61819f8fe10e6c7ef1fbc96e963897dd1b9caf2fc1548cee18d11.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 433, + 70, + 492, + 129 + ], + "blocks": [ + { + "bbox": [ + 433, + 70, + 492, + 129 + ], + "lines": [ + { + "bbox": [ + 433, + 70, + 492, + 129 + ], + "spans": [ + { + "bbox": [ + 433, + 70, + 492, + 129 + ], + "type": "image", + "image_path": "0a32d90e2a4d584ded64c1672f5350e9568b489dd708e305f14d298f35d50790.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 495, + 70, + 554, + 129 + ], + "blocks": [ + { + "bbox": [ + 495, + 70, + 554, + 129 + ], + "lines": [ + { + "bbox": [ + 495, + 70, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 495, + 70, + 554, + 129 + ], + "type": "image", + "image_path": "fd0968ebe67be0ff628f029e2c42a941dd31d9ada077a78002b77762206964e7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 309, + 140, + 367, + 199 + ], + "blocks": [ + { + "bbox": [ + 309, + 140, + 367, + 199 + ], + "lines": [ + { + "bbox": [ + 309, + 140, + 367, + 199 + ], + "spans": [ + { + "bbox": [ + 309, + 140, + 367, + 199 + ], + "type": "image", + "image_path": "1cae2ef4965d0ff7619dd25e629be4c82828fa627a4f7c54022d8401d9d3aff0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 370, + 140, + 430, + 199 + ], + "blocks": [ + { + "bbox": [ + 420, + 131, + 443, + 139 + ], + "lines": [ + { + "bbox": [ + 420, + 131, + 443, + 139 + ], + "spans": [ + { + "bbox": [ + 420, + 131, + 443, + 139 + ], + "type": "text", + "content": "(b) GT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 140, + 430, + 199 + ], + "lines": [ + { + "bbox": [ + 370, + 140, + 430, + 199 + ], + "spans": [ + { + "bbox": [ + 370, + 140, + 430, + 199 + ], + "type": "image", + "image_path": "73625c413d28379546b685d57cbc5e83dcee2d15dcf74a382b1ef35054d85da8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 433, + 140, + 492, + 199 + ], + "blocks": [ + { + "bbox": [ + 433, + 140, + 492, + 199 + ], + "lines": [ + { + "bbox": [ + 433, + 140, + 492, + 199 + ], + "spans": [ + { + "bbox": [ + 433, + 140, + 492, + 199 + ], + "type": "image", + "image_path": "188df50cd324cc163b8f058910df147172f1d483bc7b596def4fccaea5051702.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 495, + 140, + 554, + 199 + ], + "blocks": [ + { + "bbox": [ + 495, + 140, + 554, + 199 + ], + "lines": [ + { + "bbox": [ + 495, + 140, + 554, + 199 + ], + "spans": [ + { + "bbox": [ + 495, + 140, + 554, + 199 + ], + "type": "image", + "image_path": "5c453b1d37eee1c358d0240da56a1583f3804d2a23ca0f52ea8a55cf40678bae.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 309, + 210, + 367, + 269 + ], + "blocks": [ + { + "bbox": [ + 309, + 210, + 367, + 269 + ], + "lines": [ + { + "bbox": [ + 309, + 210, + 367, + 269 + ], + "spans": [ + { + "bbox": [ + 309, + 210, + 367, + 269 + ], + "type": "image", + "image_path": "a87febfee9d113e4fe8ce382a2f9cbad5fe2b41ef6701e1e32b7bbdaabeae406.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 370, + 210, + 430, + 269 + ], + "blocks": [ + { + "bbox": [ + 416, + 201, + 447, + 209 + ], + "lines": [ + { + "bbox": [ + 416, + 201, + 447, + 209 + ], + "spans": [ + { + "bbox": [ + 416, + 201, + 447, + 209 + ], + "type": "text", + "content": "(d) VIBE" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 210, + 430, + 269 + ], + "lines": [ + { + "bbox": [ + 370, + 210, + 430, + 269 + ], + "spans": [ + { + "bbox": [ + 370, + 210, + 430, + 269 + ], + "type": "image", + "image_path": "d9a8dd3b049e5178f111bf925a3a9221f309470531c6c999cdfc6f72327409d3.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 270, + 445, + 279 + ], + "lines": [ + { + "bbox": [ + 417, + 270, + 445, + 279 + ], + "spans": [ + { + "bbox": [ + 417, + 270, + 445, + 279 + ], + "type": "text", + "content": "(f) Ours" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 433, + 210, + 492, + 269 + ], + "blocks": [ + { + "bbox": [ + 433, + 210, + 492, + 269 + ], + "lines": [ + { + "bbox": [ + 433, + 210, + 492, + 269 + ], + "spans": [ + { + "bbox": [ + 433, + 210, + 492, + 269 + ], + "type": "image", + "image_path": "da1e135f507fb77e750331426461cbcccf1be54cdb43735840e4738d968f5e22.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 495, + 210, + 554, + 269 + ], + "blocks": [ + { + "bbox": [ + 495, + 210, + 554, + 269 + ], + "lines": [ + { + "bbox": [ + 495, + 210, + 554, + 269 + ], + "spans": [ + { + "bbox": [ + 495, + 210, + 554, + 269 + ], + "type": "image", + "image_path": "f724e1dbb66a90f30e710ff9903a9db0c6ec885ceca571c6fab4df09c3ed9ed0.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 73, + 319, + 279, + 396 + ], + "blocks": [ + { + "bbox": [ + 73, + 319, + 279, + 396 + ], + "lines": [ + { + "bbox": [ + 73, + 319, + 279, + 396 + ], + "spans": [ + { + "bbox": [ + 73, + 319, + 279, + 396 + ], + "type": "table", + "html": "
Fusion SchemeJ-PAV-PAF@5F@15
Cross-attention5.955.9975.898.9
Single plane Encoding5.855.9076.799.0
Triplane Encoding (w/o 1D compress)5.805.8576.999.0
Triplane Encoding5.775.7477.699.0
", + "image_path": "68afa8b96721e0e144dc017dd15139798c77732685c3a0780c242b05b564e1aa.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 89, + 426, + 263, + 466 + ], + "blocks": [ + { + "bbox": [ + 64, + 403, + 286, + 415 + ], + "lines": [ + { + "bbox": [ + 64, + 403, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 64, + 403, + 286, + 415 + ], + "type": "text", + "content": "Table 6. Ablation study of different temporal fusion schemes." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 426, + 263, + 466 + ], + "lines": [ + { + "bbox": [ + 89, + 426, + 263, + 466 + ], + "spans": [ + { + "bbox": [ + 89, + 426, + 263, + 466 + ], + "type": "table", + "html": "
Insert PositionJ-PAV-PAF@5F@15
Final layer6.016.0575.198.6
Middle layer5.775.7477.699.0
", + "image_path": "0f0158666546d943f98649c3cc9446a846409c7dec784c0a0c9a8f86d53ae0f6.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 474, + 295, + 496 + ], + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 496 + ], + "type": "text", + "content": "Table 7. Ablation Study on the Position of Temporal Information Fusion in the Backbone Network." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 518, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 639 + ], + "type": "text", + "content": "temporal features are fused at multiple intermediate levels throughout the network, and another where fusion occurs solely at the final level, as shown in Tab. 7. The experimental results in Tab. 7 demonstrate that fusing temporal features at multiple intermediate levels leads to superior performance compared to only performing fusion at the final level. This suggests that incorporating temporal information at various stages of the network helps to better capture ambiguous local features, contributing to more accurate and consistent hand pose reconstruction in low resolution." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 647, + 167, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 167, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 167, + 659 + ], + "type": "text", + "content": "4.6. Qualitative Results" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "We present the qualitative comparison results on DexYCB in Fig. 4. Compared with low resolution benchmarks (HMR, VIBE) that use the same LR-ViT as the backbone, our method achieves better alignment of the reconstructed" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 322, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 382 + ], + "type": "text", + "content": "hands with low-resolution images. Additionally, by leveraging historic temporal information to supplement missing details in visual features, our approach enhances the ability to cope with local occlusions compare to solely rely on a single frame (see columns 2 and 3 in Fig. 4 (e) and (f))." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 392, + 388, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 388, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 388, + 403 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 412, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 604 + ], + "type": "text", + "content": "In this paper, we propose a temporal enhancement framework for low-resolution hand reconstruction. We represent temporal information through joint features to capture the spatiotemporal correlations between joints. Furthermore, we integrate the temporal context provided by joint features with the spatial context provided by visual features to construct a robust low-resolution representation with 3D spatial awareness. Specifically, we introduce a Triplane feature representation to address the misalignment in semantics and representation forms between joint and visual features, preserving 3D structural information during fusion. Quantitative experiments on multiple datasets have shown that our method achieves comparable performance to high-resolution SOTA methods at much lower resolutions. We also demonstrated the contribution of each design to the scheme performance through ablation study." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "**Limitation and Future Work** Our work did not explicitly model long-term hand motion cues across video sequences, which may result in a lack of constraints on motion consistency. As a result, the model may struggle with maintaining consistent hand shapes over extended sequences. Future work should explore utilizing motion cues from long sequences to better constrain the reconstructed hand shape and improve temporal consistency." + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22634" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "type": "text", + "content": "Acknowledgements This work was supported by the National Natural Science Foundation of China under Grants (62406039, 62101064, 62171057,62201072, U23B2001, 62001054, 62071067), the Ministry of Education and China Mobile Joint Fund (MCM20200202, MCM20180101), Beijing University of Posts and Telecommunications-China Mobile Research Institute Joint Innovation Center, BUPT Excellent Ph.D. Students Foundation (CX20241014), in part by the Project funded by China Postdoctoral Science Foundation (2023TQ0039, 2024M750257), the Postdoctoral Fellowship Program of CPSF (GZC20230320)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 214, + 115, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 214, + 115, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 214, + 115, + 227 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 234, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 234, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 234, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 234, + 294, + 277 + ], + "type": "text", + "content": "[1] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10843-10852, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 280, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 295, + 323 + ], + "type": "text", + "content": "[2] Yujun Cai, Liuhao Ge, Jianfei Cai, and Junsong Yuan. Weakly-supervised 3d hand pose estimation from monocular rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 666-682, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "text", + "content": "[3] Berk Calli, Arjun Singh, Aaron Walsman, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In 2015 international conference on advanced robotics (ICAR), pages 510-517. IEEE, 2015. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 380, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 434 + ], + "type": "text", + "content": "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 436, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 501 + ], + "type": "text", + "content": "[5] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 502, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 567 + ], + "type": "text", + "content": "[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12929–12938, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 569, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 294, + 635 + ], + "type": "text", + "content": "[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20544-20554, 2022. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "type": "text", + "content": "[8] Hoseong Cho, Chanwoo Kim, Jihyeon Kim, Seongyeong Lee, Elkhan Ismayilzada, and Seungryul Baek. Transformer-based unified recognition of two hands manipulating objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4769-4778, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[9] Hongsuk Choi, Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Beyond static features for temporally consis" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "tent 3d human pose and shape from a video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1964-1973, 2021. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 107, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 193 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 195, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 258 + ], + "type": "text", + "content": "[11] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 260, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 553, + 315 + ], + "type": "text", + "content": "[12] Qichen Fu, Xingyu Liu, Ran Xu, Juan Carlos Niebles, and Kris M Kitani. Deformer: Dynamic fusion transformer for robust hand pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23600-23611, 2023. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 316, + 555, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 555, + 370 + ], + "type": "text", + "content": "[13] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10833-10842, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "type": "text", + "content": "[14] Peter Hardy, Srinandan Dasmahapatra, and Hansung Kim. Can super resolution be used to improve human pose estimation in low resolution scenarios? arXiv preprint arXiv:2107.02108, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 415, + 553, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 553, + 479 + ], + "type": "text", + "content": "[15] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 571-580, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 482, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 503 + ], + "type": "text", + "content": "[16] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "type": "text", + "content": "[17] Weiting Huang, Pengfei Ren, Jingyu Wang, Qi Qi, and Haifeng Sun. Awr: Adaptive weighting regression for 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11061-11068, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "text", + "content": "[18] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In Proceedings of the European conference on computer vision (ECCV), pages 118-134, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "text", + "content": "[19] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 679 + ], + "type": "text", + "content": "[20] Leyla Khaleghi, Alireza Sepas-Moghaddam, Josh Marshall, and Ali Etemad. Multiview video-based 3-d hand pose estimation. IEEE Transactions on Artificial Intelligence, 4:896-909, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[21] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Self-supervised learning of 3d human pose using multi-view geometry. In Proceedings of the IEEE/CVF conference on" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22635" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "computer vision and pattern recognition, pages 1077-1086, 2019. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[22] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5253-5263, 2020. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[23] Deying Kong, Linguang Zhang, Liangjian Chen, Haoyu Ma, Xiangyi Yan, Shanlin Sun, Xingwei Liu, Kun Han, and Xiaohui Xie. Identity-aware hand mesh estimation and personalization from rgb images. In European Conference on Computer Vision, pages 536-553. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "text", + "content": "[24] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4990-5000, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 331 + ], + "type": "text", + "content": "[25] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[26] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 378, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 420 + ], + "type": "text", + "content": "[27] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12939-12948, 2021. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "text", + "content": "[28] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "text", + "content": "[29] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14687-14697, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 535, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 590 + ], + "type": "text", + "content": "[30] Xingyu Liu, Pengfei Ren, Yuanyuan Gao, Jingyu Wang, Haifeng Sun, Qi Qi, Zirui Zhuang, and Jianxin Liao. Keypoint fusion for rgb-d based 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3756-3764, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 591, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 657 + ], + "type": "text", + "content": "[31] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 752-768. Springer, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[32] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2308-2317, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "text", + "content": "[33] Chihiro Nakatsuka and Satoshi Komorita. Stable 3d human pose estimation in low-resolution videos with a few views. In 2022 International Conference on Robotics and Automation (ICRA), pages 427-433. IEEE, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[34] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 218 + ], + "type": "text", + "content": "[35] Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826-9836, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 273 + ], + "type": "text", + "content": "[36] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. Pose-guided hierarchical graph reasoning for 3-d hand pose estimation from a single depth image. IEEE Transactions on Cybernetics, 53(1):315-328, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "type": "text", + "content": "[37] Pengfei Ren, Haifeng Sun, Jiachang Hao, Qi Qi, Jingyu Wang, and Jianxin Liao. A dual-branch self-boosting framework for self-supervised 3d hand pose estimation. IEEE Transactions on Image Processing, 31:5052–5066, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 320, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 386 + ], + "type": "text", + "content": "[38] Pengfei Ren, Haifeng Sun, Jiachang Hao, Jingyu Wang, Qi Qi, and Jianxin Liao. Mining multi-view information: a strong self-supervised framework for depth-based 3d hand pose and mesh estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20555-20565, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[39] Pengfei Ren, Yuchen Chen, Jiachang Hao, Haifeng Sun, Qi Qi, Jingyu Wang, and Jianxin Liao. Two heads are better than one: Image-point cloud network for depth-based 3d hand pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2163-2171, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 444, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 509 + ], + "type": "text", + "content": "[40] Pengfei Ren, Chao Wen, Xiaozheng Zheng, Zhou Xue, Haifeng Sun, Qi Qi, Jingyu Wang, and Jianxin Liao. Decoupled iterative refinement framework for interacting hands reconstruction from a single rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8014-8025, 2023. 1, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "type": "text", + "content": "[41] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 555, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 610 + ], + "type": "text", + "content": "[42] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "text", + "content": "[43] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 89-98, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 700 + ], + "type": "text", + "content": "[44] Xiao Tang, Tianyu Wang, and Chi-Wing Fu. Towards accurate alignment in real-time 3d hand-mesh reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11698-11707, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "type": "text", + "content": "[45] Chen Wang, Feng Zhang, Xiatian Zhu, and Shuzhi Sam Ge." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22636" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "Low-resolution human pose estimation. Pattern Recognition, 126:108579, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[46] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "text", + "content": "[47] Hao Xu, Tianyu Wang, Xiao Tang, and Chi-Wing Fu. H2onet: Hand-occlusion-and-orientation-aware network for real-time 3d hand mesh reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17048–17058, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[48] Xiangyu Xu, Hao Chen, Francesc Moreno-Noguer, Laszlo A Jeni, and Fernando De la Torre. 3d human pose, shape and texture from low-resolution images and videos. IEEE transactions on pattern analysis and machine intelligence, 44(9): 4490-4504, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[49] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35:38571-38584, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 354 + ], + "type": "text", + "content": "[50] John Yang, Hyung Jin Chang, Seungeui Lee, and Nojun Kwak. Seqhand: Rgb-sequence-based 3d hand pose and shape estimation. In European Conference on Computer Vision, pages 122-139. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 296, + 409 + ], + "type": "text", + "content": "[51] Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Jun-song Yuan. Mixste: Seq2seq mixed spatio-temporal encoder for 3d human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13232-13242, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "text", + "content": "[52] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2354-2364, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "type": "text", + "content": "[53] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Sar: Spatial-aware regression for 3d hand pose and mesh reconstruction from a monocular rgb image. In 2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 99-108. IEEE, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[54] Xiaozheng Zheng, Chao Wen, Zhou Xue, Pengfei Ren, and Jingyu Wang. Hamuco: Hand pose estimation via multiview collaborative self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20763-20773, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 624 + ], + "type": "text", + "content": "[55] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. Toch: Spatio-temporal object-to-hand correspondence for motion refinement. In European Conference on Computer Vision, pages 1-19. Springer, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "text", + "content": "[56] Yuxiao Zhou, Marc Habermann, Weipeng Xu, Ikhsanul Habibie, Christian Theobalt, and Feng Xu. Monocular real-time hand shape and motion capture using multi-modal data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5346-5355, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[57] Zhishan Zhou, Shihao Zhou, Zhi Lv, Minqiang Zou, Yao Tang, and Jiajun Liang. A simple baseline for efficient hand mesh reconstruction. In Proceedings of the IEEE/CVF Con-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 241 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 1367-1376, 2024. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[58] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15085-15099, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "text", + "content": "[59] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In Proceedings of the IEEE international conference on computer vision, pages 4903-4911, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "text", + "content": "[60] Christian Zimmermann, Max Argus, and Thomas Brox. Contrastive representation learning for hand shape estimation. In Pattern Recognition, pages 250-264, Cham, 2021. Springer International Publishing. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22637" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/6abac681-ec51-468a-ba82-3e1de6acd774_content_list.json b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/6abac681-ec51-468a-ba82-3e1de6acd774_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f0caa08f492ac1ceea54b8def0a1253e4ec76f37 --- /dev/null +++ b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/6abac681-ec51-468a-ba82-3e1de6acd774_content_list.json @@ -0,0 +1,1731 @@ +[ + { + "type": "text", + "text": "PoseBH: Prototypical Multi-Dataset Training Beyond Human Pose Estimation", + "text_level": 1, + "bbox": [ + 104, + 130, + 893, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Uyoung Jeong $^{1}$ Jonathan Freer $^{2}$ Seungryul Baek $^{1}$ Hyung Jin Chang $^{2}$ Kwang In Kim $^{3}$ $^{1}$ UNIST $^{2}$ University of Birmingham $^{3}$ POSTECH", + "bbox": [ + 120, + 179, + 872, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 257, + 326, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We study multi-dataset training (MDT) for pose estimation, where skeletal heterogeneity presents a unique challenge that existing methods have yet to address. In traditional domains, e.g. regression and classification, MDT typically relies on dataset merging or multi-head supervision. However, the diversity of skeleton types and limited cross-dataset supervision complicate integration in pose estimation. To address these challenges, we introduce PoseBH, a new MDT framework that tackles keypoint heterogeneity and limited supervision through two key techniques. First, we propose nonparametric keypoint prototypes that learn within a unified embedding space, enabling seamless integration across skeleton types. Second, we develop a cross-type self-supervision mechanism that aligns keypoint predictions with keypoint embedding prototypes, providing supervision without relying on teacher-student models or additional augmentations. PoseBH substantially improves generalization across whole-body and animal pose datasets, including COCO-WholeBody, AP-10K, and APT-36K, while preserving performance on standard human pose benchmarks (COCO, MPII, and AIC). Furthermore, our learned keypoint embeddings transfer effectively to hand shape estimation (InterHand2.6M) and human body shape estimation (3DPW). The code for PoseBH is available at: https://github.com/uyoung-jeong/PoseBH.", + "bbox": [ + 89, + 289, + 483, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 694, + 220, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The demand for human pose estimators has grown substantially in recent years, driving their adoption across diverse applications, including 3D avatar generation [14, 25, 31], motion generation [15, 22, 33], human-robot interaction [20, 44, 59], safety monitoring [7, 18, 39], and VR/AR tracking [12, 43, 47]. These domains pose distinct operational challenges, often necessitating significant domain adaptation for robust pose estimation. To address this, we aim to train a human pose estimator that generalizes effectively across multiple datasets.", + "bbox": [ + 89, + 719, + 482, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, multi-dataset training (MDT) for pose estimation presents two key challenges. First, keypoints", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/43a33a2afd1a03df425af3a78dc42b1cae43b9c136b7c98400c354781c7235fa.jpg", + "image_caption": [ + "COCO" + ], + "image_footnote": [], + "bbox": [ + 517, + 256, + 612, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/19c165f3a24b569e93979b415d9428a0c35a93f90a4273c64793192e6c213cd4.jpg", + "image_caption": [ + "AIC" + ], + "image_footnote": [], + "bbox": [ + 614, + 256, + 709, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3242e21cae62c2ccf64299f92785151df94bdd3362da83fdac61e6e2eb32aba.jpg", + "image_caption": [ + "MPII" + ], + "image_footnote": [], + "bbox": [ + 709, + 256, + 803, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/008abc2c4d83cb96d3474a78c51c19a47b89dea96caf28779ad26e1c297a5778.jpg", + "image_caption": [ + "WholeBody" + ], + "image_footnote": [], + "bbox": [ + 805, + 256, + 901, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b711e282a34ca825b6b61038b9d3a57189d962ad717232e160e8d377614d375d.jpg", + "image_caption": [ + "AP-10K" + ], + "image_footnote": [], + "bbox": [ + 517, + 359, + 612, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8fa0aa361a62d037be1304f667b77e736bd9c3d47a2a882eee378c034a550a4a.jpg", + "image_caption": [ + "APT-36K" + ], + "image_footnote": [], + "bbox": [ + 614, + 359, + 707, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e13785908194b70429277d31d109d9a0e1f9beb14d1dca21d3e9a262edd6ff5.jpg", + "image_caption": [ + "3DPW" + ], + "image_footnote": [], + "bbox": [ + 709, + 359, + 805, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4f9a3fbc1c16f9aba82b1c44972e8feb9c32ce26e52c3bfee63f222f467cda4.jpg", + "image_caption": [ + "InterHand2.6M", + "Figure 1. PoseBH unifies diverse skeleton formats, including humans, hands, and animals. The displayed skeletons show pose estimation results from our method, with 3DPW and InterHand2.6M predictions from a transferred model." + ], + "image_footnote": [], + "bbox": [ + 807, + 359, + 901, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are heterogeneous across different datasets. For example, COCO [23] and MPII [1] define 17 and 16 keypoints, respectively, with 9 non-overlapping ones. Even when datasets share keypoints with identical names, their localization characteristics often differ due to domain gaps. For example, while COCO and animal datasets (e.g., APT-10K [58] and APT-36K [56]) use common keypoints names such as eyes, nose, and hips, significant anatomical differences exist (see Fig. 1). As a result, naively merging the identically named keypoints is ineffective.", + "bbox": [ + 511, + 561, + 906, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Second, pose estimation MDT suffers from a lack of adequate supervision. In MDT, multiple skeleton types from different datasets are trained for each input instance. However, only a single skeleton-type ground truth is provided per input, leaving others unsupervised. This limitation resembles a semi-supervised learning challenge, constraining performance on smaller datasets. Existing semi-supervised learning methods typically rely on input augmentation or teacher-student knowledge distillation, both of which require substantial extra computational overhead.", + "bbox": [ + 511, + 715, + 908, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing approaches fail to fully address these challenges, often relying on a shared backbone network [6] or", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12278", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a mixture of experts [53] to preserve downstream performance. However, these methods tend to focus on mainstream datasets like COCO or MPII, resulting in degraded performance on underrepresented datasets. We propose a method that enhances performance in discrepant domains while maintaining accuracy on standard human pose datasets. A straightforward fine-tuning approach on the target domain often disrupts performance on source datasets, as it fails to balance differences in skeleton structures and dataset distributions. Therefore, effective generalization requires jointly harmonizing skeletal representations and dataset diversity during training.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We achieve this through two primary techniques. First, we introduce prototypes to represent keypoints as embedding vectors, providing a unified representation across datasets. We regress pixel-wise keypoint embeddings from backbone features and match them to corresponding prototypes, automating keypoint mapping across datasets. This enables the integration of diverse keypoint types, including whole-body and animal keypoints, despite significant skeletal differences. Furthermore, our keypoint prototypes are non-learnable parameters, offering adaptability and computational efficiency for domain transfer.", + "bbox": [ + 89, + 272, + 480, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, we generate reliable self-supervision signals using the predicted keypoints and their embeddings, a process we term cross-type self-supervision. Since keypoint predictions can be obtained via a dot product between embeddings and prototypes, we effectively establish two distinct keypoint prediction modalities. By aligning these predictions, we filter out noisy keypoints and produce reliable labels for supervising unlabeled keypoints. This process enhances the utility of embeddings without requiring additional networks or input augmentation. As a result, we provide effective supervision for underrepresented and discrepant datasets, achieving a higher degree of generalizability.", + "bbox": [ + 89, + 439, + 480, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main technical contributions are as follows:", + "bbox": [ + 109, + 619, + 428, + 633 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Keypoint prototypes for learning arbitrary keypoints from multiple datasets ensuring high transferability.", + "- A cross-type self-supervision mechanism that aligns keypoint regression outputs with keypoint embeddings, enriching supervision for unlabeled keypoints.", + "- Strong generalization across estimation tasks, including whole-body, animal, and hand pose estimation, as well as human shape estimation." + ], + "bbox": [ + 91, + 635, + 480, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 89, + 768, + 235, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-dataset training has been widely explored in various computer vision tasks. Relabeling and pseudo-labeling are among the most straightforward yet resource-intensive approaches. For example, MSeg [19] employed relabeling for segmentation, while [63] used offline pseudo-labeling for object detection. However, both methods face scalability challenges, as adding a new dataset requires reprocessing", + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "annotations for all previously processed datasets to redefine the class set, leading to a quadratic computational cost.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several approaches have been proposed to improve scalability. In object detection, [4, 65] aim to unify object class labels by merging similar categories. However, they struggle to integrate localization features of the same object due to the absence of spatial information in class logits. UniTrack learns a unified identity embedding through self-supervised learning for multiple object-tracking tasks, including human pose tracking [49]. However, it prioritizes pose propagation over estimation and relies exclusively on the PoseTrack18 [2] skeleton, without incorporating multiple pose datasets.", + "bbox": [ + 511, + 121, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another class of methods [48, 51, 54, 62] employs network modules, such as graph convolutional networks (GCN), for cross-dataset adaptation. However, these dataset-specific modules limit domain transferability, as their input and output sizes are dictated by the label structure, making them incompatible with datasets that have different keypoint formats. In contrast, our approach achieves flexible transferability without relying on dataset-specific domain adaptation modules.", + "bbox": [ + 511, + 287, + 903, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human pose estimation with MDT. In 3D human shape estimation, several methods have used pseudo-ground truths (GTs) generated by the SMPL model for MDT [5, 17, 41, 42]. However, these pseudo-GTs are often inaccurate due to depth ambiguity in 2D images and the limited expressiveness of the parametric 3D model [29, 30]. Similar to [65], Sarandi et al.'s affine-combining autoencoder (ACAE) merges 3D joint coordinates via an autoencoder [36]. However, the autoencoder's learnable parameters are dataset-dependent, limiting its generalizability. Furthermore, ACAE does not preserve pixel-level features, making it unsuitable when spatial information is crucial.", + "bbox": [ + 511, + 446, + 903, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works on category-agnostic pose estimation integrate various objects, such as chairs and vehicles, by combining multiple datasets [13, 52]. These methods follow a relabeling approach, creating a superset of keypoint classes and detecting only the keypoints defined in their relabeled dataset. As a result, they do not address keypoint heterogeneity or label sparsity issues, as training and evaluation occur within a single labeled dataset. In contrast, our method uses multiple skeletons to improve generalization across diverse downstream domains.", + "bbox": [ + 511, + 628, + 903, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In 2D human pose estimation, UniHCP performs multiple human perception tasks using task-specific queries without incorporating task-specific layers inside the backbone [6]. This design limits its ability to generalize to distant domains such as animals and whole-body pose estimation. ViTPose++ [53] and Sapiens [16] employ multi-head architectures for MDT but do not address skeleton heterogeneity or the resulting label sparsity.", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12279", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa025e7daeaf31ca940cd2d70076f26054a16400641f4148c3a85d9ce74fa53d.jpg", + "image_caption": [ + "Figure 2. Overview of the PoseBH architecture. During training, the embedding head maps the backbone features into a unified keypoint embedding space. By matching these embeddings with prototypes, dataset-specific keypoint heatmaps $\\mathbf{k}_{\\mathrm{coco}}^{\\mathrm{P}}$ are generated. Prototypes are updated nonparametrically from the embeddings. During inference, the embedding head and the subsequent procedures are removed." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 624, + 318 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/54d4bb2b2991acf752d6e8ec4020400bdf2f0ce860d387059179f41c9b62041c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 89, + 905, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prototype learning can be interpreted as proxy-based metric learning [35], where the centers of a fixed set of clusters serve as prototypes. This approach has been studied in various domains, including semantic segmentation [8, 10, 64], person re-identification [32], meta-learning [24, 38, 57, 61], and visual interpretability [3, 9, 46].", + "bbox": [ + 88, + 398, + 480, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ProtoSeg [64] introduces a nonparametric approach to semantic segmentation by assigning multiple prototypes per class. ProMotion [27] employs prototypes to unify scene depth estimation and optical flow tasks. While our method draws inspiration from ProtoSeg and aligns with ProMotion in exploiting the benefits of prototypes for MDT, it differs in three key aspects. First, our approach applies keypoint-level supervision, which is inherently sparser than segmentation, depth, or optical flow. Second, we specifically address skeleton heterogeneity and cross-dataset relationships, aspects not considered by ProtoSeg and ProMotion. Third, while ProtoSeg and ProMotion apply fully supervised learning, our method tackles the challenge of unlabeled data, a problem unique to MDT pose estimation.", + "bbox": [ + 88, + 489, + 482, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 717, + 181, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We focus on multi-dataset training (MDT) for 2D human pose estimation. Given $D$ datasets, our model predicts a keypoint heatmap for a person, represented as an array $\\mathbf{k}_d \\in \\mathbb{R}^{J_d \\times H \\times W}$ , where $d$ is the dataset index, $J_{d}$ is the number of keypoints in the $d$ -th dataset, and $H$ and $W$ denote the height and width of the heatmap, respectively.", + "bbox": [ + 89, + 743, + 483, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-head baseline", + "text_level": 1, + "bbox": [ + 89, + 847, + 279, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The multi-head baseline model consists of a backbone network that extracts an output backbone feature map from", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an input RGB image. The backbone network is shared across all datasets, while individual keypoint heads independently regress dataset-wise keypoint heatmaps $\\{\\mathbf{k}_d\\}_{d=1}^D$ , each trained with its corresponding labels. The keypoint heads share the same architecture, except for the final layer, where the output channel size is the respective number of keypoints. This generic baseline does not share the keypoint output space, limiting the MDT effect up to the backbone.", + "bbox": [ + 511, + 398, + 906, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Keypoint prototype learning", + "text_level": 1, + "bbox": [ + 511, + 527, + 767, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. Our goal is to learn a unified keypoint space that standardizes keypoint output formats across datasets, scales with dataset size, and retains dataset-specific localization accuracy. To achieve this, we reformulate keypoint regression as a prototype-based distance metric learning problem (see Fig. 2). Our shared keypoint embedding module generates a normalized keypoint embedding map $\\mathbf{e} \\in \\mathbb{R}^{F \\times H \\times W}$ from an input feature map. The module consists of two deconvolution layers, followed by a residual block and two convolution layers. Additional details are provided in the supplementary document.", + "bbox": [ + 511, + 547, + 906, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using embeddings and prototypes, we perform keypoint classification by matching embeddings with prototypes $\\mathbf{P} \\in \\mathbb{R}^{J \\times M \\times F}$ , where $F$ is the embedding dimension, $J$ is the total number of keypoints across all datasets, $M$ is the number of in-class prototypes. For each heatmap pixel, we compute the cosine similarity between the prototypes and the embedding vector to obtain keypoint predictions:", + "bbox": [ + 511, + 715, + 905, + 821 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {k} _ {j, y, x} ^ {\\mathrm {P}} = \\max _ {m} \\frac {\\mathbf {P} _ {(j , m , :)} \\cdot \\mathbf {e} _ {(: , y , x)}}{\\left\\| \\mathbf {P} _ {(j , m , :)} \\right\\| \\left\\| \\mathbf {e} _ {(: , y , x)} \\right\\|}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 828, + 903, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x,y$ denote pixel indices in the heatmap, and $\\mathbf{P}_{(j,m,:)}$ is a row vector of size $F$ , extracted $\\mathbf{P}\\in \\mathbb{R}^{J\\times M\\times F}$ along its", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12280", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "first two dimensions. For each keypoint class $j$ , $\\mathbf{k}_{(j,:;:)}^{\\mathrm{P}} \\in \\mathbb{R}^{H \\times W}$ stores the matching score of the most probable prototype among $M$ prototypes. We refer to this as the prototype keypoint heatmap $\\mathbf{k}^{\\mathrm{P}}$ to distinguish it from the multihead outputs $\\mathbf{k}$ .", + "bbox": [ + 89, + 90, + 480, + 169 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the remainder of this paper, we use the following notation: $\\mathbf{A}_{(:,a)}$ denotes the 2D slice of a 3D array $\\mathbf{A}$ indexed at $a$ along the third dimension, while $\\mathbf{A}_{(:,b,:)}$ represents the 2D slice at index $b$ along the second dimension. Similarly, $\\mathbf{A}_{(d:e,:;)}$ represents the 3D sub-array spanning indices $d$ to $e$ along the first dimension. For a 2D array $\\mathbf{B}$ , $\\mathbf{B}_{(:,c)}$ indicates the $c$ -th column vector, whereas $\\mathbf{B}_{(d,c)}$ refers to the element at position $(d,c)$ .", + "bbox": [ + 89, + 169, + 483, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prototype learning. We adopt a nonparametric learning approach [64] to train $\\mathbf{P}$ . First, we obtain logits $\\{\\mathbf{l}_j\\}_{j=1}^J \\subset \\mathbb{R}^{M \\times N}$ and targets $\\{\\mathbf{t}_j \\in \\mathbb{R}^{N_j}\\}_{j=1}^J$ , where $N_j$ is the number of foreground samples (i.e., pixels in the ground-truth (GT) heatmap with nonzero values) for the $j$ -th joint, and $N$ is the total number of foreground samples across all joints. We define logits as sampled embedding vectors and targets as keypoint class labels, with each keypoint class selecting an in-class prototype from $M$ candidates.", + "bbox": [ + 89, + 306, + 483, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since no GT selection is explicitly provided, we assign an in-class prototype for each logit via online clustering using Sinkhorn-Knopp iteration with an equipartition constraint [37], and compute $\\mathbf{t}_j$ as follows:", + "bbox": [ + 89, + 445, + 483, + 506 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {l} _ {j} = \\mathbf {P} _ {(j,:;:)} \\bar {\\mathbf {e}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 512, + 480, + 529 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {j} = \\underset {m} {\\arg \\max } \\operatorname {d i a g} (\\mathbf {u}) \\exp \\left(\\frac {\\mathbf {l} _ {j}}{\\kappa}\\right) \\operatorname {d i a g} (\\mathbf {v}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 531, + 480, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\overline{\\mathbf{e}}$ is a 2D array of size $F\\times N_{j}$ , obtained by first flattening $\\mathbf{e}\\in \\mathbb{R}^{F\\times H\\times W}$ along the spatial dimensions into a 2D array of size $F\\times (HW)$ and then, selecting $N_{j}$ foreground objects. $\\mathrm{diag}(\\mathbf{u})$ is a diagonal matrix derived from the vector $\\mathbf{u}$ , and $\\exp (\\cdot)$ is applied element-wise.", + "bbox": [ + 89, + 569, + 483, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each prototype is then updated with momentum $\\lambda$ :", + "bbox": [ + 109, + 645, + 447, + 660 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {j, m} ^ {\\text {n e w}} = \\lambda \\mathbf {P} _ {j, m} ^ {\\text {o l d}} + (1 - \\lambda) \\frac {1}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} w _ {m n} \\overline {{\\mathbf {e}}} _ {(;, n)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 666, + 480, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{mn}$ is the clustering assignment weight, calculated similarly to Eq. (3), but using max instead of arg max.", + "bbox": [ + 89, + 713, + 480, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimize the learned embedding with respect to the prototype $\\mathbf{P}$ , we employ pixel-prototype contrastive learning, using the pixel-prototype contrastive loss (PPC) and pixel-prototype distance (PPD) from [64]:", + "bbox": [ + 89, + 744, + 483, + 804 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P P C}} \\left(\\mathbf {l} _ {j}, \\mathbf {t} _ {j}\\right) = \\frac {c _ {n}}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} \\operatorname {C E} \\left(\\mathbf {l} _ {j, n}, \\mathbf {t} _ {j, n}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 810, + 480, + 852 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {P P D}} \\left(\\mathbf {l} _ {j}, \\mathbf {t} _ {j}\\right) = \\frac {c _ {n}}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} \\left(1 - \\mathbf {l} _ {j, n} ^ {m +}\\right) ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 854, + 480, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4bda5829f2ed1d2ff71d0b311ee56f613716dce5b3382e919783c88b3fc0c80c.jpg", + "image_caption": [ + "Figure 3. An illustrative example of cross-type self-supervision. Given a COCO image, we jointly refine the AIC head and AIC prototype predictions to produce reliable AIC heatmaps." + ], + "image_footnote": [], + "bbox": [ + 519, + 87, + 906, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{n}$ is the confidence value from the GT heatmap, CE denotes the cross-entropy loss, and $m^{+}$ indexes the first dimension of $\\mathbf{l}_d$ . For each dataset $d$ , we collect logits $\\mathbf{l}_{j,n} \\in \\mathbb{R}^{M \\times N_d}$ and targets $\\mathbf{t}_d \\in \\mathbb{R}^{N_d}$ , where $N_{d}$ is the number of samples in dataset $d$ . The total prototype loss is then defined as $\\mathcal{L}_{\\mathrm{Proto}} = \\mathcal{L}_{\\mathrm{PPC}} + \\mathcal{L}_{\\mathrm{PPD}}$ .", + "bbox": [ + 511, + 272, + 903, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike standard contrastive learning, $\\mathcal{L}_{\\mathrm{Proto}}$ operates within a single dataset and lacks the push term for differing dataset skeletons. To introduce cross-dataset negatives, we apply K-means clustering to all prototypes, forming $K$ (=96) clusters. This clustering is performed only once during training, while online clustering for in-class prototypes is conducted at each iteration. These clusters serve as negative samples, where the embeddings closest to each centroid are used as logits, with their dataset-specific labels as targets. Finally, we obtain $\\mathbf{l}_c$ and $\\mathbf{t}_c$ for each cluster and apply Eqs. (5) and (6) to enable cross-dataset contrastive learning.", + "bbox": [ + 511, + 363, + 905, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Cross-type self-supervision", + "text_level": 1, + "bbox": [ + 511, + 537, + 756, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While our keypoint prototypes enable unified MDT, label sparsity remains a challenge, as each image is annotated for only one dataset, limiting supervision. To address this semi-supervised learning problem, we introduce a self-supervision strategy that bridges keypoint heatmaps and embeddings. By collaboratively filtering out noisy predictions from both the keypoint and embedding heads, we generate reliable keypoint heatmaps for training (see Fig. 3).", + "bbox": [ + 511, + 559, + 905, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To eliminate uncertain keypoint predictions, we apply two filtering conditions. First, for each keypoint class, the confidence scores of both the keypoint and embedding heads must exceed a threshold $(c_{thr} = 0.25)$ . Second, the root mean square distance between their predictions must be below a threshold $(d_{thr} = 2.1)$ . The filtered predictions are then combined using a weighted average:", + "bbox": [ + 511, + 680, + 905, + 786 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathbf {y}} _ {i} = s _ {i} \\widehat {\\mathbf {y}} _ {i} ^ {\\mathrm {k p t}} + (1 - s _ {i}) \\widehat {\\mathbf {y}} _ {i} ^ {\\mathrm {e m b}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 794, + 903, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_i = c_i^{\\mathrm{kpt}} / (c_i^{\\mathrm{kpt}} + c_i^{\\mathrm{emb}})$ , $i$ is the keypoint index, and $\\widehat{\\mathbf{y}}$ and $c$ represent the keypoint prediction and confidence score from each head, respectively.", + "bbox": [ + 511, + 821, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From $\\widehat{\\mathbf{y}}$ , we generate a reliable heatmap $\\mathbf{k}^{\\mathrm{CSS}}$ following the standard GT heatmap generation process. The loss for", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12281", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/5e762ad2a6b22a5999a7002facde101e1c9835ecbcf29bedef7cec39d9288ae7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGFLOPSCOCOAICMPIIAvg.
UniHCP18.976.832.690.966.8
ViTPose++18.577.031.693.167.2
Ours18.577.332.193.267.5
", + "bbox": [ + 114, + 88, + 457, + 160 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/482ad8f271afcae5b4af537b4e966614a6a877c6782a046750ea3070f7d5551c.jpg", + "table_caption": [ + "Table 1. Comparison of different multi-dataset training methods on general human pose benchmarks, with computational complexity measured in GFLOPS during the evaluation phase." + ], + "table_footnote": [], + "table_body": "
MethodAP-10KAPT-36KCOCO-WAvg.
UniHCP56.562.020.146.2
ViTPose++74.176.057.169.1
Ours75.087.257.973.4
", + "bbox": [ + 91, + 224, + 480, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Comparison of multi-dataset training methods on whole-body and animal domains. COCO-W refers to COCO-WholeBody. The reported metrics are AP(%)", + "bbox": [ + 89, + 319, + 482, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "unlabeled samples is then computed as:", + "bbox": [ + 89, + 387, + 352, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C S S}} = \\sum_ {d = 1} ^ {D} \\zeta \\left[ \\mathcal {L} _ {\\mathrm {h m}} \\left(\\mathbf {k} _ {d [ \\mathbf {u} ]}, \\mathbf {k} _ {d [ \\mathbf {u} ]} ^ {\\mathrm {C S S}}\\right) \\right. \\\\ + \\mathcal {L} _ {\\text {P r o t o}} \\left(\\mathbf {e} _ {[ \\mathbf {u} ]}, \\mathbf {k} _ {d [ \\mathbf {u} ]}, \\mathbf {k} _ {d [ \\mathbf {u} ]} ^ {\\mathrm {C S S}}\\right) ], \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 412, + 480, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{u}$ denotes the indices of the unlabeled samples. This cross-type self-supervision (CSS) loss enables self-distillation between the dataset-wise keypoint head and the prototype, mitigating supervision shortage without requiring a teacher model or input duplication.", + "bbox": [ + 89, + 486, + 482, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final training loss is defined as:", + "bbox": [ + 109, + 561, + 346, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M D T}} = \\mathcal {L} _ {\\mathrm {K P L}} + \\mathcal {L} _ {\\mathrm {C S S}}, \\text {w h e r e} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 588, + 482, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K P L}} = \\sum_ {d = 1} ^ {D} \\left[ \\mathcal {L} _ {\\mathrm {h m}} \\left(\\mathbf {k} _ {d}, \\mathbf {k} _ {d} ^ {\\mathrm {g t}}\\right) + \\mathcal {L} _ {\\mathrm {P r o t o}} \\left(\\mathbf {e}, \\mathbf {k} _ {d}, \\mathbf {k} _ {d} ^ {\\mathrm {g t}}\\right) \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 608, + 482, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\mathcal{L}_{\\mathrm{hm}}$ is the standard JointMSE loss [40].", + "bbox": [ + 89, + 659, + 382, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 686, + 223, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 712, + 277, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow the experimental setup of ViTPose++ [53], using six human pose datasets for training. (1) COCO consists of 57,000 training and 5,000 validation images, annotated with 17 keypoints [23]. (2) AIChallenger (AIC) contains 210,000 training and 30,000 validation images, annotated with 14 keypoints [50]. (3) MPII includes approximately 15,000 training, 2,729 validation, and 5,700 test images, with 16 keypoints [1]. (4) AP-10K is an animal pose dataset with approximately 7,000 training, 1,000 validation, and 2,000 test images [58]. (5) APT-36K, a video-based animal pose dataset, consists of about 25,000 training, 3,600 validation,", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and 7,000 test images [56]. Both AP-10K and APT-36K share the same skeletal format with 17 keypoints, largely following COCO. (6) COCO-WholeBody extends COCO with whole-body annotations covering 133 keypoints [50]. We follow standard evaluation protocols, using the average precision (AP) metric based on object keypoint similarity (OKS) for COCO, AIC, COCO-WholeBody, AP-10K, and APT-36K. For MPII, the percentage of correct keypoints (PCK) metric is used [1].", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the domain transfer experiments, we use InterHand2.6M [28] and 3DPW [45]. InterHand2.6M is a two-hand pose estimation dataset containing approximately 2.6M frames, annotated with 21 keypoints. We adopt the ViTPose++ configuration for frame sampling and dataset splits (training, validation, and test). For evaluation, we use PCK, AUC, and EPE, consistent with ViTPose++. 3DPW is an in-the-wild 3D human shape estimation dataset with 18,000 training and 26,000 test frames, annotated with 24 keypoints. We use COCO-style AP and AR as evaluation metrics. In both InterHand2.6M and 3DPW, the projected 2D keypoints are used as ground-truths.", + "bbox": [ + 511, + 227, + 905, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further details on the experimental setups, the impact of varying hyperparameters, and failure case analysis can be found in the supplemental document.", + "bbox": [ + 511, + 409, + 905, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 511, + 465, + 725, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our multi-head baseline model is ViTPose++ [53]. We use an input image size of $256 \\times 192$ and apply a flip test in all cases. The model is trained for 100 epochs using the AdamW optimizer with a weight decay factor of 0.1 and a step learning rate scheduler with an initial learning rate of 0.001. The learning rate is reduced by a factor of 0.1 at the 50th and 90th epochs. We use four NVIDIA A100 GPUs for training with the ViT-H backbone and four NVIDIA A6000 GPUs for training with the ViT-B backbone.", + "bbox": [ + 511, + 488, + 905, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ a scheduling strategy for both the losses and the learning rate. Initially, we freeze the multi-head model, updating only the embedding module and prototypes for 50 epochs. In the subsequent 40 epochs, we train the multi-heads and the embedding module while keeping the prototypes and the backbone frozen. Finally, during the last 10 epochs, we train the entire network except the prototypes. At this stage, we introduce cross-type self-supervision loss. This progressive optimization approach accelerates training compared to training the entire network from the outset.", + "bbox": [ + 511, + 625, + 905, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison with existing MDT methods", + "text_level": 1, + "bbox": [ + 511, + 787, + 862, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Tab. 1, we evaluate the competitiveness of our method against existing MDT approaches. For a fair comparison, all methods use the ViT-Base backbone and are trained under our dataset configuration. While UniHCP performs well on AIC, it achieves the lowest scores on COCO and MPII, resulting in an average of 66.8. We hypothe", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12282", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/74adebbf2682bec539674f578c1f8e5b09cac961cfe206adb259bad9a3813f39.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneDatasetsCOCOMPII
(AP↑)(AP50↑)(AP75↑)(AR↑)(PCKh↑)(PCKh@0.1↑)
HRNetHRNet-W48175.190.682.280.490.3†33.1†
HRFormerHRFormer-B175.690.882.880.8--
SimCCHRNet-W48176.190.682.981.290.036.8
PCT†Swin-B177.791.284.782.192.5-
PCT†Swin-H179.391.585.9---
UniHCPViT-B3376.1---93.2††-
ViTPose++-BViT-B677.073.484.082.692.839.1
ViTPose++-HViT-H679.491.985.784.894.241.6
OursViT-B677.390.884.282.493.239.3
OursViT-H679.591.985.884.594.242.0
", + "bbox": [ + 114, + 88, + 883, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "size that UniHCP's parameter-sharing strategy causes well-represented pose datasets to dominate, hindering its generalization across different skeleton structures. ViTPose++ serves as both a multi-head model and the baseline for our method. Our algorithm surpasses ViTPose++ by 0.3 in average score and outperforms it on individual datasets. Given that these benchmark scores are highly saturated and our method maintains the same computation complexity as ViTPose++ during inference, this improvement is significant.", + "bbox": [ + 88, + 364, + 480, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further assess generalization, we test our method on animal and whole-body pose estimation datasets. Table 2 shows that our method consistently improves performance, despite significant domain discrepancies. UniHCP lags behind with an average score of 27.2, likely due to its parameter sharing strategy, which does not address skeleton heterogeneity. Our method also surpasses ViTPose++ by an average score of 4.3, demonstrating superior generalization beyond human skeletons. Specifically, it achieves gains of 0.9 AP on AP-10K, 11.2 AP on APT-36K, and 0.8 AP on COCO-WholeBody. Unlike existing methods that compromise either mainstream performance or adaptation to distant domains, our approach enhances both synergistically.", + "bbox": [ + 88, + 502, + 482, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison to state-of-the-art methods", + "text_level": 1, + "bbox": [ + 89, + 710, + 431, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 3, we compare our method with state-of-the-art (SOTA) approaches on the COCO and MPII benchmarks, using scores reported in the respective papers. We also evaluate our method against SOTA approaches designed for single datasets, including HRNet [40], HRFformer [60], SimCC [21], and PCT [11]. On the COCO validation set, our method ranks second with the ViT-B backbone and achieves SOTA performance with the ViT-H backbone, reaching 79.5 AP. On the MPII validation set, our ViT-B model performs comparably to the fine-tuned UniHCP and surpasses the SOTA on the PCKh@0.1 metric by 0.4 PCKh.", + "bbox": [ + 88, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ad3ad1aeb996c57073b2cdcb7dbd4dcae69cfd84cabfd2f7928c5038058311bd.jpg", + "table_caption": [ + "Table 3. Comparison with state-of-the-art methods on COCO and MPII. Methods with ${}^{ \\dagger }$ uses ${256} \\times {256}$ input size. ${}^{\\dagger \\dagger }$ indicates additional fine-tuning on each dataset." + ], + "table_footnote": [], + "table_body": "
MethodInterHand2.6M3DPW
PCK↑AUC↑EPE↓AP↑AR↑
UniHCP98.587.03.7257.961.8
ViTPose++98.386.24.0281.785.2
Ours98.687.13.7083.687.1
", + "bbox": [ + 514, + 362, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Transfer results on hand (Interhand2.6M) and human shape (3DPW) domains.", + "bbox": [ + 511, + 470, + 906, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These results demonstrate that our method effectively balances adaptability to diverse domains while maintaining competitive performance on standard benchmarks, aligning with our primary objective of improving generalization beyond mainstream tasks.", + "bbox": [ + 511, + 529, + 906, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Transfer to hands and human shapes", + "text_level": 1, + "bbox": [ + 511, + 618, + 834, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PoseBH has the capability to generalize its learned representations beyond the training dataset. We validate this through domain transfer experiments by fine-tuning the model separately on the InterHand2.6M and 3DPW datasets. InterHand2.6M is a two-hand pose estimation dataset containing 21 MANO keypoints [34], while 3DPW is an in-the-wild human shape estimation dataset with 24 SMPL keypoints [26].", + "bbox": [ + 511, + 642, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During fine-tuning, we freeze the embedding module and nonparametrically learn the prototypes for the target datasets. After convergence, we fine-tune the backbone and head weights alongside the prototype loss. Freezing the embedding module ensures that the previously learned prototypes and those of the target dataset remain aligned within the same embedding space, preserving the learned representations during transfer. Since these experiments do not involve skeleton heterogeneity across datasets, no semi-", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12283", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53c6437367cbd68d4e582e4aabd6a24ea92f39629a1fea7044648f63e941409f.jpg", + "image_caption": [ + "Figure 4. Comparative pose estimation results on human (left) and animal (dog; right)." + ], + "image_footnote": [], + "bbox": [ + 109, + 85, + 883, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d7f0d136da98f5cbb22c318d7f45f00e14edbb52af7c57f0276cc828307751e.jpg", + "image_caption": [ + "Figure 5. Comparative results on InterHand2.6M (a-c) and 3DPW (e-f)." + ], + "image_footnote": [], + "bbox": [ + 112, + 348, + 883, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "supervised learning challenges arise, and CSS loss is not applied.", + "bbox": [ + 89, + 643, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 4, we report results on the InterHand2.6M and 3DPW test sets using the ViT-B backbone. Given its large dataset size ( $\\sim$ 2.6M instances), InterHand2.6M exhibits a smaller performance gap among different methods. Nonetheless, our method surpasses ViTPose++ by 0.3 PCK, 0.9 AUC, and 0.32 EPE, while also outperforming UniHCP by 0.1 PCK, 0.1 AUC, and 0.02 EPE. For the smaller 3DPW dataset, our method achieves a notable performance gain, surpassing the baseline by 1.9 AP and 1.9 AR, demonstrating strong generalization in keypoint embeddings.", + "bbox": [ + 89, + 674, + 482, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 833, + 251, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 5, we present the results of ablation experiments on the COCO (CO), AIC (AI), MPII (MP), AP-10K (AP), APT-36K (APT), and COCO-Wholebody (CW) validation", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sets using the ViT-Base backbone. The top row represents a multi-head baseline with a shared backbone and dataset-specific heads. This conventional multi-head approach exhibits relatively low performance, particularly on the AP-10K and APT-36K validation sets, due to substantial domain discrepancies. Our keypoint prototype method achieves an average score of 66.3, improving the baseline by 0.8. Additionally, incorporating cross-type self-supervision provides an additional average performance gain of 0.2. Collectively, these enhancements enable our final configuration to outperform the multi-head baseline by an average of 2.4 points across all six datasets.", + "bbox": [ + 511, + 643, + 906, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Visual analysis", + "text_level": 1, + "bbox": [ + 511, + 833, + 663, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 presents examples of human and animal pose estimation, demonstrating that our algorithm achieves more accurate keypoint localization under complex interactions and", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12284", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/351aed903ab13454ee08ad3ca52e8a221700c4e559c53379af1fe12d7b917069.jpg", + "image_caption": [ + "(a)", + "Figure 6. Examples of SMPL fitting using KITRO and with our 2D predictions." + ], + "image_footnote": [], + "bbox": [ + 96, + 99, + 478, + 416 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d222dc3e92cc7540ce20f6372649d98441204c481c739219c07d18dd7c318161.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCOAIMPAPAPTCWAvg.
Baseline77.031.693.175.374.857.168.2
+LProto77.031.893.076.486.457.570.4
+LCSS77.332.193.276.786.557.970.6
", + "bbox": [ + 117, + 473, + 455, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study of the proposed method. $\\mathcal{L}_{\\mathrm{Proto}}$ : keypoint prototype learning. $\\mathcal{L}_{\\mathrm{CSS}}$ : cross-type self-supervision.", + "bbox": [ + 89, + 566, + 482, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "occlusions, as seen in the right-hand keypoints of the standing person (left) and the right front toe of the dog (right).", + "bbox": [ + 89, + 626, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further compare the pose estimation results on InterHand2.6M and 3DPW in Fig. 5. In (a), ViTPose++ incorrectly assigns a little finger joint to the opposite hand, whereas our method provides more accurate localization. In (b), ViTPose++ struggles with occlusion from the other hand, producing noisy predictions, while our method correctly separates them. In (c), due to severe self-occlusion, ViTPose++ misestimates occluded keypoints and assigns them to the opposite hand, while our method generalizes well in challenging self-occlusion scenarios. In (d), ViTPose++ produces an implausible foot and hand pose under heavy occlusion, while our method remains robust. In (e), ViTPose++ swaps the left and right legs, while our method correctly distinguishes them. In (f), background clutter causes ViTPose++ to produce noisy predictions for the left hand, while our method correctly identifies the lo", + "bbox": [ + 88, + 659, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cation of the occluded left hand.", + "bbox": [ + 511, + 90, + 728, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further assess the effectiveness of our 2D SMPL keypoint predictions on the 3DPW test set, we employ a SMPL optimization technique. KITRO [55] optimizes SMPL parameters using 2D keypoint reprojection loss. For a comparative visual analysis, we replace KITRO's 2D keypoint predictions with ours. As shown in Figure 6(a) and (b), KITRO fails to align the mesh with the actual feet, whereas our method achieves more precise alignment. In (c), KITRO mislocalizes the right hand under partial occlusion, while our 2D predictions remain robust. In (d), KITRO exhibits misalignment in both the hands and feet, while replacing its 2D predictions with ours significantly improves the alignment of peripheral body parts.", + "bbox": [ + 511, + 109, + 906, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the supplemental material, we provide a visualization of the prototypes generated by our algorithm, illustrating how they effectively capture the diversity of representations within the embedding space.", + "bbox": [ + 511, + 310, + 906, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 396, + 633, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a new multi-dataset training (MDT) framework for pose estimation across diverse domains and datasets. Traditional approaches encounter challenges with label heterogeneity, dataset-specific parameters, and limited domain generalization. Our approach addresses the limitations with two main contributions. First, to resolve label heterogeneity and dataset-specific parameters, we employ nonparametric prototypes to unify diverse keypoint annotations across multiple datasets. Second, to address the limited multi-dataset supervision, we propose a cross-type self-supervision mechanism. The proposed method has demonstrated substantial improvements over existing MDT across a broad range of datasets, including human pose, animal pose, hand shape, and human shape estimation.", + "bbox": [ + 511, + 425, + 908, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. A limitation of our approach is that it requires at least a few-shot sample to learn prototypes, making zero-shot inference for unseen skeletons infeasible. Future work should explore extending our method for zero-shot human pose estimation. Future work should also investigate adapting the approach to 3D domains e.g. using axis-angle rotation representations, to facilitate representation learning on 3D manifolds.", + "bbox": [ + 511, + 647, + 908, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by the National Research Foundation of Korea (NRF) grant (No. 2021R1A2C2012195) and by the Institute of Information & communications Technology Planning & Evaluation (IITP) grants No. RS-2019-II191906 (AI Graduate School Program, POSTECH), No. RS-2020-II201336 (AI Graduate School Program, UNIST), and RS-2022-II220290, funded by the Korea government (MSIT).", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12285", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2D human pose estimation: New benchmark and state of the art analysis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1, 5", + "[2] Mykhaylo Andriluka, Umar Iqbal, Eldar Insafutdinov, Leonid Pishchulin, Anton Milan, Juergen Gall, and Bernt Schiele. PoseTrack: A benchmark for human pose estimation and tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5167-5176, 2018. 2", + "[3] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3", + "[4] Yanbei Chen, Manchen Wang, Abhay Mittal, Zhenlin Xu, Paolo Favaro, Joseph Tighe, and Davide Modolo. ScaleDet: A scalable multi-dataset object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7288-7297, 2023. 2", + "[5] Hongsuk Choi, Gyeongsik Moon, JoonKyu Park, and Kyoung Mu Lee. Learning to estimate robust 3D human mesh from in-the-wild crowded scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[6] Yuanzheng Ci, Yizhou Wang, Meilin Chen, Shixiang Tang, Lei Bai, Feng Zhu, Rui Zhao, Fengwei Yu, Donglian Qi, and Wanli Ouyang. UniHCP: A unified model for human-centric perceptions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17840-17852, 2023. 1, 2", + "[7] Mickael Cormier, Aris Clepe, Andreas Specker, and Jürgen Beyerer. Where are we with human pose estimation in real-world surveillance? In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 591-601, 2022. 1", + "[8] Anurag Das, Yongqin Xian, Dengxin Dai, and Bernt Schiele. Weakly-supervised domain adaptive semantic segmentation with prototypical contrastive learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15434-15443, 2023. 3", + "[9] Jon Donnelly, Alina Jade Barnett, and Chaofan Chen. Deformable protopnet: An interpretable image classifier using deformable prototypes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10265-10275, 2022. 3", + "[10] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4320–4329, 2022. 3", + "[11] Zigang Geng, Chunyu Wang, Yixuan Wei, Ze Liu, Houqiang Li, and Han Hu. Human pose as compositional tokens. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6", + "[12] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human POSEitioning System (HPS): 3D human pose estimation and self-localization in large scenes from" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "body-mounted sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4318-4329, 2021. 1", + "[13] Or Hirschorn and Shai Avidan. A graph-based approach for category-agnostic pose estimation, 2024. 2", + "[14] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. NeuMan: Neural human radiance field from a single video. In European Conference on Computer Vision (ECCV), 2022. 1", + "[15] Roy Kapon, Guy Tevet, Daniel Cohen-Or, and Amit H Bermano. MAS: Multi-view ancestral sampling for 3D motion generation using 2D diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1965–1974, 2024. 1", + "[16] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. arXiv preprint arXiv:2408.12569, 2024. 2", + "[17] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2", + "[18] Alberto Lamas, Siham Tabik, Antonio Cano Montes, Francisco Pérez-Hernández, Jorge García, Roberto Olmos, and Francisco Herrera. Human pose estimation for mitigating false negatives in weapon detection in video-surveillance. Neurocomputing, 489:488-503, 2022. 1", + "[19] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. MSeg: A composite dataset for multidomain semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[20] Duy Tho Le, Chenhui Gou, Stavya Datta, Hengcan Shi, Ian Reid, Jianfei Cai, and Hamid Rezatofighi. JRDB-PanoTrack: An open-world panoptic segmentation and tracking robotic dataset in crowded human environments. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22325-22334, 2024. 1", + "[21] Yanjie Li, Sen Yang, Peidong Liu, Shoukui Zhang, Yunxiao Wang, Zhicheng Wang, Wankou Yang, and Shu-Tao Xia. SimCC: A simple coordinate classification perspective for human pose estimation. In European Conference on Computer Vision (ECCV), pages 89-106. Springer, 2022. 6", + "[22] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-X: A large-scale 3D expressive whole-body human motion dataset. Advances in Neural Information Processing Systems (NeurIPS), 2023. 1", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 1, 5", + "[24] Sun-Ao Liu, Yiheng Zhang, Zhaofan Qiu, Hongtao Xie, Yongdong Zhang, and Ting Yao. Learning orthogonal prototypes for generalized few-shot semantic segmentation. In" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12286", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[25] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. HumanGaussian: Text-driven 3D human generation with gaussian splatt-ting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6646-6657, 2024. 1", + "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 6", + "[27] Yawen Lu, Dongfang Liu, Qifan Wang, Cheng Han, Yiming Cui, Zhiwen Cao, Xueling Zhang, Yingjie Victor Chen, and Heng Fan. ProMotion: Prototypes as motion learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28109-28119, 2024. 3", + "[28] Gyeongsik Moon, Shouou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. InterHand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 5", + "[29] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. NeuralAnnot: Neural annotator for 3D human mesh training sets. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2", + "[30] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three Recipes for Better 3D PseudogTs of 3D Human Mesh Estimation in the Wild. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 2", + "[31] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3D gaussian avatar. In European Conference on Computer Vision (ECCV), 2024. 1", + "[32] Haocong Rao and Chunyan Miao. TranSG: Transformer-based skeleton graph prototype contrastive learning with structure-trajectory prompted reconstruction for person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22118-22128, 2023. 3", + "[33] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D human motion model for robust pose estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1", + "[34] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 6", + "[35] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7420-7430, 2022. 3", + "[36] István Sárándi, Alexander Hermans, and Bastian Leibe. Learning 3D human pose estimation from dozens of datasets using a geometry-aware autoencoder to bridge between skeleton formats. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. 4", + "[38] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3", + "[39] Markus D Solbach and John K Tsotsos. Vision-based fallen person detection for the elderly. In IEEE/CVF international conference on computer vision workshops (ICCVW), pages 1433-1442, 2017. 1", + "[40] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5, 6", + "[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Black Michael J., and Tao Mei. Monocular, one-stage, regression of multiple 3D people. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3D people in depth. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[43] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xR-EgoPose: Egocentric 3D human pose from an hmd camera. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7728-7738, 2019. 1", + "[44] Edward Vendrow, Duy Tho Le, Jianfei Cai, and Hamid Rezatofighi. JRDB-Pose: A large-scale dataset for multiperson pose estimation and tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1", + "[45] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 5", + "[46] Chong Wang, Yuyuan Liu, Yuanhong Chen, Fengbei Liu, Yu Tian, Davis McCarthy, Helen Frazer, and Gustavo Carneiro. Learning support and trivial prototypes for interpretable image classification. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2062-2072, 2023. 3", + "[47] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3D human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13031-13040, 2023. 1", + "[48] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards universal object detection by domain attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7289-7298, 2019. 2", + "[49] Zhongdao Wang, Hengshuang Zhao, Ya-Li Li, Shengjin Wang, Philip Torr, and Luca Bertinetto. Do different tracking tasks require different appearance models? Advances in Neural Information Processing Systems (NeurIPS), 34:726-738, 2021. 2", + "[50] Jiahong Wu, He Zheng, Bo Zhao, Yixin Li, Baoming Yan, Rui Liang, Wenjia Wang, Shipei Zhou, Guosen Lin, Yanwei Fu, et al. AI Challenger: A large-scale dataset for" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12287", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "going deeper in image understanding. arXiv preprint arXiv:1711.06475, 2017. 5", + "[51] Hang Xu, Linpu Fang, Xiaodan Liang, Wenxiong Kang, and Zhenguo Li. Universal-RCNN: Universal object detector via transferable graph r-cnn. In AAAI Conference on Artificial Intelligence (AAAI), pages 12492–12499, 2020. 2", + "[52] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In European Conference on Computer Vision (ECCV), pages 398–416. Springer, 2022. 2", + "[53] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose++: Vision transformer for generic body pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 5", + "[54] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. UNIK: A unified framework for real-world skeleton-based action recognition. British Machine Visison Conference (BMVC), 2021. 2", + "[55] Fengyuan Yang, Kerui Gu, and Angela Yao. KITRO: Refining human mesh by 2D clues and kinematic-tree rotation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1052-1061, 2024. 8", + "[56] Yuxiang Yang, Junjie Yang, Yufei Xu, Jing Zhang, Long Lan, and Dacheng Tao. APT-36K: A large-scale benchmark for animal pose estimation and tracking. Advances in Neural Information Processing Systems (NeurIPS), 35:17301-17313, 2022. 1, 5", + "[57] Sung Whan Yoon, Jun Seo, and Jaekyun Moon. TapNet: Neural network augmented with task-adaptive projection for few-shot learning. In International conference on machine learning, pages 7115-7123. PMLR, 2019. 3", + "[58] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. AP-10K: A benchmark for animal pose estimation in the wild. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 1, 5", + "[59] Xinyi Yu, Xin Zhang, Chengjun Xu, and Linlin Ou. Humanrobot collaborative interaction with human perception and action recognition. Neurocomputing, 563:126827, 2024. 1", + "[60] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. HRFormer: High-resolution transformer for dense prediction. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 6", + "[61] Baoquan Zhang, Xutao Li, Yunming Ye, Zhichao Huang, and Lisai Zhang. Prototype completion with primitive knowledge for few-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3754-3762, 2021. 3", + "[62] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3D: A unified baseline for multi-dataset 3D object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9253-9262, 2023. 2", + "[63] Xiangyun Zhao, Samuel Schulter, Gaurav Sharma, Yi-Hsuan Tsai, Manmohan Chandraker, and Ying Wu. Object detection with a unified label space from multiple datasets. In" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference on Computer Vision (ECCV), pages 178-193. Springer, 2020. 2", + "[64] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking Semantic Segmentation: A Prototype View. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2582-2593, 2022. 3, 4", + "[65] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7571-7580, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 232 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12288", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/6abac681-ec51-468a-ba82-3e1de6acd774_model.json b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/6abac681-ec51-468a-ba82-3e1de6acd774_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1d1c5f53f93cb6d8a5cb81700716ac31c1b0df50 --- /dev/null +++ b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/6abac681-ec51-468a-ba82-3e1de6acd774_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.131, + 0.895, + 0.153 + ], + "angle": 0, + "content": "PoseBH: Prototypical Multi-Dataset Training Beyond Human Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.18, + 0.873, + 0.226 + ], + "angle": 0, + "content": "Uyoung Jeong\\(^{1}\\) Jonathan Freer\\(^{2}\\) Seungryul Baek\\(^{1}\\) Hyung Jin Chang\\(^{2}\\) Kwang In Kim\\(^{3}\\) \n\\(^{1}\\)UNIST \\(^{2}\\)University of Birmingham \\(^{3}\\)POSTECH" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.258, + 0.327, + 0.274 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.485, + 0.668 + ], + "angle": 0, + "content": "We study multi-dataset training (MDT) for pose estimation, where skeletal heterogeneity presents a unique challenge that existing methods have yet to address. In traditional domains, e.g. regression and classification, MDT typically relies on dataset merging or multi-head supervision. However, the diversity of skeleton types and limited cross-dataset supervision complicate integration in pose estimation. To address these challenges, we introduce PoseBH, a new MDT framework that tackles keypoint heterogeneity and limited supervision through two key techniques. First, we propose nonparametric keypoint prototypes that learn within a unified embedding space, enabling seamless integration across skeleton types. Second, we develop a cross-type self-supervision mechanism that aligns keypoint predictions with keypoint embedding prototypes, providing supervision without relying on teacher-student models or additional augmentations. PoseBH substantially improves generalization across whole-body and animal pose datasets, including COCO-WholeBody, AP-10K, and APT-36K, while preserving performance on standard human pose benchmarks (COCO, MPII, and AIC). Furthermore, our learned keypoint embeddings transfer effectively to hand shape estimation (InterHand2.6M) and human body shape estimation (3DPW). The code for PoseBH is available at: https://github.com/uyoung-jeong/PoseBH." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.222, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.87 + ], + "angle": 0, + "content": "The demand for human pose estimators has grown substantially in recent years, driving their adoption across diverse applications, including 3D avatar generation [14, 25, 31], motion generation [15, 22, 33], human-robot interaction [20, 44, 59], safety monitoring [7, 18, 39], and VR/AR tracking [12, 43, 47]. These domains pose distinct operational challenges, often necessitating significant domain adaptation for robust pose estimation. To address this, we aim to train a human pose estimator that generalizes effectively across multiple datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "However, multi-dataset training (MDT) for pose estimation presents two key challenges. First, keypoints" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.257, + 0.614, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.347, + 0.586, + 0.358 + ], + "angle": 0, + "content": "COCO" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.257, + 0.71, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.347, + 0.675, + 0.358 + ], + "angle": 0, + "content": "AIC" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.257, + 0.805, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.347, + 0.774, + 0.358 + ], + "angle": 0, + "content": "MPII" + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.257, + 0.902, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.347, + 0.887, + 0.359 + ], + "angle": 0, + "content": "WholeBody" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.36, + 0.614, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.451, + 0.588, + 0.461 + ], + "angle": 0, + "content": "AP-10K" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.36, + 0.709, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.451, + 0.689, + 0.461 + ], + "angle": 0, + "content": "APT-36K" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.36, + 0.806, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.451, + 0.777, + 0.461 + ], + "angle": 0, + "content": "3DPW" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.36, + 0.903, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.451, + 0.897, + 0.461 + ], + "angle": 0, + "content": "InterHand2.6M" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Figure 1. PoseBH unifies diverse skeleton formats, including humans, hands, and animals. The displayed skeletons show pose estimation results from our method, with 3DPW and InterHand2.6M predictions from a transferred model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.715 + ], + "angle": 0, + "content": "are heterogeneous across different datasets. For example, COCO [23] and MPII [1] define 17 and 16 keypoints, respectively, with 9 non-overlapping ones. Even when datasets share keypoints with identical names, their localization characteristics often differ due to domain gaps. For example, while COCO and animal datasets (e.g., APT-10K [58] and APT-36K [56]) use common keypoints names such as eyes, nose, and hips, significant anatomical differences exist (see Fig. 1). As a result, naively merging the identically named keypoints is ineffective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.909, + 0.868 + ], + "angle": 0, + "content": "Second, pose estimation MDT suffers from a lack of adequate supervision. In MDT, multiple skeleton types from different datasets are trained for each input instance. However, only a single skeleton-type ground truth is provided per input, leaving others unsupervised. This limitation resembles a semi-supervised learning challenge, constraining performance on smaller datasets. Existing semi-supervised learning methods typically rely on input augmentation or teacher-student knowledge distillation, both of which require substantial extra computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Existing approaches fail to fully address these challenges, often relying on a shared backbone network [6] or" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12278" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "a mixture of experts [53] to preserve downstream performance. However, these methods tend to focus on mainstream datasets like COCO or MPII, resulting in degraded performance on underrepresented datasets. We propose a method that enhances performance in discrepant domains while maintaining accuracy on standard human pose datasets. A straightforward fine-tuning approach on the target domain often disrupts performance on source datasets, as it fails to balance differences in skeleton structures and dataset distributions. Therefore, effective generalization requires jointly harmonizing skeletal representations and dataset diversity during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.437 + ], + "angle": 0, + "content": "We achieve this through two primary techniques. First, we introduce prototypes to represent keypoints as embedding vectors, providing a unified representation across datasets. We regress pixel-wise keypoint embeddings from backbone features and match them to corresponding prototypes, automating keypoint mapping across datasets. This enables the integration of diverse keypoint types, including whole-body and animal keypoints, despite significant skeletal differences. Furthermore, our keypoint prototypes are non-learnable parameters, offering adaptability and computational efficiency for domain transfer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.482, + 0.62 + ], + "angle": 0, + "content": "Second, we generate reliable self-supervision signals using the predicted keypoints and their embeddings, a process we term cross-type self-supervision. Since keypoint predictions can be obtained via a dot product between embeddings and prototypes, we effectively establish two distinct keypoint prediction modalities. By aligning these predictions, we filter out noisy keypoints and produce reliable labels for supervising unlabeled keypoints. This process enhances the utility of embeddings without requiring additional networks or input augmentation. As a result, we provide effective supervision for underrepresented and discrepant datasets, achieving a higher degree of generalizability." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.621, + 0.429, + 0.634 + ], + "angle": 0, + "content": "Our main technical contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.636, + 0.481, + 0.665 + ], + "angle": 0, + "content": "- Keypoint prototypes for learning arbitrary keypoints from multiple datasets ensuring high transferability." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.666, + 0.481, + 0.711 + ], + "angle": 0, + "content": "- A cross-type self-supervision mechanism that aligns keypoint regression outputs with keypoint embeddings, enriching supervision for unlabeled keypoints." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.712, + 0.481, + 0.756 + ], + "angle": 0, + "content": "- Strong generalization across estimation tasks, including whole-body, animal, and hand pose estimation, as well as human shape estimation." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.636, + 0.481, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.236, + 0.785 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Multi-dataset training has been widely explored in various computer vision tasks. Relabeling and pseudo-labeling are among the most straightforward yet resource-intensive approaches. For example, MSeg [19] employed relabeling for segmentation, while [63] used offline pseudo-labeling for object detection. However, both methods face scalability challenges, as adding a new dataset requires reprocessing" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "annotations for all previously processed datasets to redefine the class set, leading to a quadratic computational cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.288 + ], + "angle": 0, + "content": "Several approaches have been proposed to improve scalability. In object detection, [4, 65] aim to unify object class labels by merging similar categories. However, they struggle to integrate localization features of the same object due to the absence of spatial information in class logits. UniTrack learns a unified identity embedding through self-supervised learning for multiple object-tracking tasks, including human pose tracking [49]. However, it prioritizes pose propagation over estimation and relies exclusively on the PoseTrack18 [2] skeleton, without incorporating multiple pose datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.905, + 0.425 + ], + "angle": 0, + "content": "Another class of methods [48, 51, 54, 62] employs network modules, such as graph convolutional networks (GCN), for cross-dataset adaptation. However, these dataset-specific modules limit domain transferability, as their input and output sizes are dictated by the label structure, making them incompatible with datasets that have different keypoint formats. In contrast, our approach achieves flexible transferability without relying on dataset-specific domain adaptation modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.905, + 0.627 + ], + "angle": 0, + "content": "Human pose estimation with MDT. In 3D human shape estimation, several methods have used pseudo-ground truths (GTs) generated by the SMPL model for MDT [5, 17, 41, 42]. However, these pseudo-GTs are often inaccurate due to depth ambiguity in 2D images and the limited expressiveness of the parametric 3D model [29, 30]. Similar to [65], Sarandi et al.'s affine-combining autoencoder (ACAE) merges 3D joint coordinates via an autoencoder [36]. However, the autoencoder's learnable parameters are dataset-dependent, limiting its generalizability. Furthermore, ACAE does not preserve pixel-level features, making it unsuitable when spatial information is crucial." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.779 + ], + "angle": 0, + "content": "Recent works on category-agnostic pose estimation integrate various objects, such as chairs and vehicles, by combining multiple datasets [13, 52]. These methods follow a relabeling approach, creating a superset of keypoint classes and detecting only the keypoints defined in their relabeled dataset. As a result, they do not address keypoint heterogeneity or label sparsity issues, as training and evaluation occur within a single labeled dataset. In contrast, our method uses multiple skeletons to improve generalization across diverse downstream domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In 2D human pose estimation, UniHCP performs multiple human perception tasks using task-specific queries without incorporating task-specific layers inside the backbone [6]. This design limits its ability to generalize to distant domains such as animals and whole-body pose estimation. ViTPose++ [53] and Sapiens [16] employ multi-head architectures for MDT but do not address skeleton heterogeneity or the resulting label sparsity." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12279" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.625, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.09, + 0.906, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.33, + 0.908, + 0.375 + ], + "angle": 0, + "content": "Figure 2. Overview of the PoseBH architecture. During training, the embedding head maps the backbone features into a unified keypoint embedding space. By matching these embeddings with prototypes, dataset-specific keypoint heatmaps \\(\\mathbf{k}_{\\mathrm{coco}}^{\\mathrm{P}}\\) are generated. Prototypes are updated nonparametrically from the embeddings. During inference, the embedding head and the subsequent procedures are removed." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.399, + 0.482, + 0.489 + ], + "angle": 0, + "content": "Prototype learning can be interpreted as proxy-based metric learning [35], where the centers of a fixed set of clusters serve as prototypes. This approach has been studied in various domains, including semantic segmentation [8, 10, 64], person re-identification [32], meta-learning [24, 38, 57, 61], and visual interpretability [3, 9, 46]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.483, + 0.703 + ], + "angle": 0, + "content": "ProtoSeg [64] introduces a nonparametric approach to semantic segmentation by assigning multiple prototypes per class. ProMotion [27] employs prototypes to unify scene depth estimation and optical flow tasks. While our method draws inspiration from ProtoSeg and aligns with ProMotion in exploiting the benefits of prototypes for MDT, it differs in three key aspects. First, our approach applies keypoint-level supervision, which is inherently sparser than segmentation, depth, or optical flow. Second, we specifically address skeleton heterogeneity and cross-dataset relationships, aspects not considered by ProtoSeg and ProMotion. Third, while ProtoSeg and ProMotion apply fully supervised learning, our method tackles the challenge of unlabeled data, a problem unique to MDT pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.718, + 0.182, + 0.733 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.484, + 0.837 + ], + "angle": 0, + "content": "We focus on multi-dataset training (MDT) for 2D human pose estimation. Given \\( D \\) datasets, our model predicts a keypoint heatmap for a person, represented as an array \\( \\mathbf{k}_d \\in \\mathbb{R}^{J_d \\times H \\times W} \\), where \\( d \\) is the dataset index, \\( J_{d} \\) is the number of keypoints in the \\( d \\)-th dataset, and \\( H \\) and \\( W \\) denote the height and width of the heatmap, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.281, + 0.862 + ], + "angle": 0, + "content": "3.1. Multi-head baseline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The multi-head baseline model consists of a backbone network that extracts an output backbone feature map from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.52 + ], + "angle": 0, + "content": "an input RGB image. The backbone network is shared across all datasets, while individual keypoint heads independently regress dataset-wise keypoint heatmaps \\(\\{\\mathbf{k}_d\\}_{d=1}^D\\), each trained with its corresponding labels. The keypoint heads share the same architecture, except for the final layer, where the output channel size is the respective number of keypoints. This generic baseline does not share the keypoint output space, limiting the MDT effect up to the backbone." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.528, + 0.769, + 0.544 + ], + "angle": 0, + "content": "3.2. Keypoint prototype learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.715 + ], + "angle": 0, + "content": "Overview. Our goal is to learn a unified keypoint space that standardizes keypoint output formats across datasets, scales with dataset size, and retains dataset-specific localization accuracy. To achieve this, we reformulate keypoint regression as a prototype-based distance metric learning problem (see Fig. 2). Our shared keypoint embedding module generates a normalized keypoint embedding map \\(\\mathbf{e} \\in \\mathbb{R}^{F \\times H \\times W}\\) from an input feature map. The module consists of two deconvolution layers, followed by a residual block and two convolution layers. Additional details are provided in the supplementary document." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.906, + 0.822 + ], + "angle": 0, + "content": "Using embeddings and prototypes, we perform keypoint classification by matching embeddings with prototypes \\(\\mathbf{P} \\in \\mathbb{R}^{J \\times M \\times F}\\), where \\(F\\) is the embedding dimension, \\(J\\) is the total number of keypoints across all datasets, \\(M\\) is the number of in-class prototypes. For each heatmap pixel, we compute the cosine similarity between the prototypes and the embedding vector to obtain keypoint predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.829, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathbf {k} _ {j, y, x} ^ {\\mathrm {P}} = \\max _ {m} \\frac {\\mathbf {P} _ {(j , m , :)} \\cdot \\mathbf {e} _ {(: , y , x)}}{\\left\\| \\mathbf {P} _ {(j , m , :)} \\right\\| \\left\\| \\mathbf {e} _ {(: , y , x)} \\right\\|}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(x,y\\) denote pixel indices in the heatmap, and \\(\\mathbf{P}_{(j,m,:)}\\) is a row vector of size \\(F\\), extracted \\(\\mathbf{P}\\in \\mathbb{R}^{J\\times M\\times F}\\) along its" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12280" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.17 + ], + "angle": 0, + "content": "first two dimensions. For each keypoint class \\(j\\), \\(\\mathbf{k}_{(j,:;:)}^{\\mathrm{P}} \\in \\mathbb{R}^{H \\times W}\\) stores the matching score of the most probable prototype among \\(M\\) prototypes. We refer to this as the prototype keypoint heatmap \\(\\mathbf{k}^{\\mathrm{P}}\\) to distinguish it from the multihead outputs \\(\\mathbf{k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.484, + 0.292 + ], + "angle": 0, + "content": "For the remainder of this paper, we use the following notation: \\(\\mathbf{A}_{(:,a)}\\) denotes the 2D slice of a 3D array \\(\\mathbf{A}\\) indexed at \\(a\\) along the third dimension, while \\(\\mathbf{A}_{(:,b,:)}\\) represents the 2D slice at index \\(b\\) along the second dimension. Similarly, \\(\\mathbf{A}_{(d:e,:;)}\\) represents the 3D sub-array spanning indices \\(d\\) to \\(e\\) along the first dimension. For a 2D array \\(\\mathbf{B}\\), \\(\\mathbf{B}_{(:,c)}\\) indicates the \\(c\\)-th column vector, whereas \\(\\mathbf{B}_{(d,c)}\\) refers to the element at position \\((d,c)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.484, + 0.446 + ], + "angle": 0, + "content": "Prototype learning. We adopt a nonparametric learning approach [64] to train \\(\\mathbf{P}\\). First, we obtain logits \\(\\{\\mathbf{l}_j\\}_{j=1}^J \\subset \\mathbb{R}^{M \\times N}\\) and targets \\(\\{\\mathbf{t}_j \\in \\mathbb{R}^{N_j}\\}_{j=1}^J\\), where \\(N_j\\) is the number of foreground samples (i.e., pixels in the ground-truth (GT) heatmap with nonzero values) for the \\(j\\)-th joint, and \\(N\\) is the total number of foreground samples across all joints. We define logits as sampled embedding vectors and targets as keypoint class labels, with each keypoint class selecting an in-class prototype from \\(M\\) candidates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.507 + ], + "angle": 0, + "content": "Since no GT selection is explicitly provided, we assign an in-class prototype for each logit via online clustering using Sinkhorn-Knopp iteration with an equipartition constraint [37], and compute \\(\\mathbf{t}_j\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.513, + 0.482, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\mathbf {l} _ {j} = \\mathbf {P} _ {(j,:;:)} \\bar {\\mathbf {e}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.532, + 0.482, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {j} = \\underset {m} {\\arg \\max } \\operatorname {d i a g} (\\mathbf {u}) \\exp \\left(\\frac {\\mathbf {l} _ {j}}{\\kappa}\\right) \\operatorname {d i a g} (\\mathbf {v}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.646 + ], + "angle": 0, + "content": "where \\(\\overline{\\mathbf{e}}\\) is a 2D array of size \\(F\\times N_{j}\\), obtained by first flattening \\(\\mathbf{e}\\in \\mathbb{R}^{F\\times H\\times W}\\) along the spatial dimensions into a 2D array of size \\(F\\times (HW)\\) and then, selecting \\(N_{j}\\) foreground objects. \\(\\mathrm{diag}(\\mathbf{u})\\) is a diagonal matrix derived from the vector \\(\\mathbf{u}\\), and \\(\\exp (\\cdot)\\) is applied element-wise." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.646, + 0.448, + 0.661 + ], + "angle": 0, + "content": "Each prototype is then updated with momentum \\(\\lambda\\):" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.667, + 0.482, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {j, m} ^ {\\text {n e w}} = \\lambda \\mathbf {P} _ {j, m} ^ {\\text {o l d}} + (1 - \\lambda) \\frac {1}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} w _ {m n} \\overline {{\\mathbf {e}}} _ {(;, n)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.482, + 0.744 + ], + "angle": 0, + "content": "where \\(w_{mn}\\) is the clustering assignment weight, calculated similarly to Eq. (3), but using max instead of arg max." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.484, + 0.805 + ], + "angle": 0, + "content": "To optimize the learned embedding with respect to the prototype \\(\\mathbf{P}\\), we employ pixel-prototype contrastive learning, using the pixel-prototype contrastive loss (PPC) and pixel-prototype distance (PPD) from [64]:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.811, + 0.482, + 0.853 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P P C}} \\left(\\mathbf {l} _ {j}, \\mathbf {t} _ {j}\\right) = \\frac {c _ {n}}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} \\operatorname {C E} \\left(\\mathbf {l} _ {j, n}, \\mathbf {t} _ {j, n}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.856, + 0.482, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {P P D}} \\left(\\mathbf {l} _ {j}, \\mathbf {t} _ {j}\\right) = \\frac {c _ {n}}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} \\left(1 - \\mathbf {l} _ {j, n} ^ {m +}\\right) ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.088, + 0.907, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.206, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Figure 3. An illustrative example of cross-type self-supervision. Given a COCO image, we jointly refine the AIC head and AIC prototype predictions to produce reliable AIC heatmaps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.905, + 0.363 + ], + "angle": 0, + "content": "where \\( c_{n} \\) is the confidence value from the GT heatmap, CE denotes the cross-entropy loss, and \\( m^{+} \\) indexes the first dimension of \\( \\mathbf{l}_d \\). For each dataset \\( d \\), we collect logits \\( \\mathbf{l}_{j,n} \\in \\mathbb{R}^{M \\times N_d} \\) and targets \\( \\mathbf{t}_d \\in \\mathbb{R}^{N_d} \\), where \\( N_{d} \\) is the number of samples in dataset \\( d \\). The total prototype loss is then defined as \\( \\mathcal{L}_{\\mathrm{Proto}} = \\mathcal{L}_{\\mathrm{PPC}} + \\mathcal{L}_{\\mathrm{PPD}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.53 + ], + "angle": 0, + "content": "Unlike standard contrastive learning, \\(\\mathcal{L}_{\\mathrm{Proto}}\\) operates within a single dataset and lacks the push term for differing dataset skeletons. To introduce cross-dataset negatives, we apply K-means clustering to all prototypes, forming \\(K\\) (=96) clusters. This clustering is performed only once during training, while online clustering for in-class prototypes is conducted at each iteration. These clusters serve as negative samples, where the embeddings closest to each centroid are used as logits, with their dataset-specific labels as targets. Finally, we obtain \\(\\mathbf{l}_c\\) and \\(\\mathbf{t}_c\\) for each cluster and apply Eqs. (5) and (6) to enable cross-dataset contrastive learning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.538, + 0.757, + 0.554 + ], + "angle": 0, + "content": "3.3. Cross-type self-supervision" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.906, + 0.681 + ], + "angle": 0, + "content": "While our keypoint prototypes enable unified MDT, label sparsity remains a challenge, as each image is annotated for only one dataset, limiting supervision. To address this semi-supervised learning problem, we introduce a self-supervision strategy that bridges keypoint heatmaps and embeddings. By collaboratively filtering out noisy predictions from both the keypoint and embedding heads, we generate reliable keypoint heatmaps for training (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.906, + 0.787 + ], + "angle": 0, + "content": "To eliminate uncertain keypoint predictions, we apply two filtering conditions. First, for each keypoint class, the confidence scores of both the keypoint and embedding heads must exceed a threshold \\( (c_{thr} = 0.25) \\). Second, the root mean square distance between their predictions must be below a threshold \\( (d_{thr} = 2.1) \\). The filtered predictions are then combined using a weighted average:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.795, + 0.905, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathbf {y}} _ {i} = s _ {i} \\widehat {\\mathbf {y}} _ {i} ^ {\\mathrm {k p t}} + (1 - s _ {i}) \\widehat {\\mathbf {y}} _ {i} ^ {\\mathrm {e m b}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.823, + 0.905, + 0.871 + ], + "angle": 0, + "content": "where \\( s_i = c_i^{\\mathrm{kpt}} / (c_i^{\\mathrm{kpt}} + c_i^{\\mathrm{emb}}) \\), \\( i \\) is the keypoint index, and \\( \\widehat{\\mathbf{y}} \\) and \\( c \\) represent the keypoint prediction and confidence score from each head, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "From \\(\\widehat{\\mathbf{y}}\\), we generate a reliable heatmap \\(\\mathbf{k}^{\\mathrm{CSS}}\\) following the standard GT heatmap generation process. The loss for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12281" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.458, + 0.161 + ], + "angle": 0, + "content": "
MethodGFLOPSCOCOAICMPIIAvg.
UniHCP18.976.832.690.966.8
ViTPose++18.577.031.693.167.2
Ours18.577.332.193.267.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.17, + 0.483, + 0.214 + ], + "angle": 0, + "content": "Table 1. Comparison of different multi-dataset training methods on general human pose benchmarks, with computational complexity measured in GFLOPS during the evaluation phase." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.226, + 0.482, + 0.31 + ], + "angle": 0, + "content": "
MethodAP-10KAPT-36KCOCO-WAvg.
UniHCP56.562.020.146.2
ViTPose++74.176.057.169.1
Ours75.087.257.973.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.32, + 0.483, + 0.362 + ], + "angle": 0, + "content": "Table 2. Comparison of multi-dataset training methods on whole-body and animal domains. COCO-W refers to COCO-WholeBody. The reported metrics are AP(%)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.388, + 0.354, + 0.404 + ], + "angle": 0, + "content": "unlabeled samples is then computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.414, + 0.482, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C S S}} = \\sum_ {d = 1} ^ {D} \\zeta \\left[ \\mathcal {L} _ {\\mathrm {h m}} \\left(\\mathbf {k} _ {d [ \\mathbf {u} ]}, \\mathbf {k} _ {d [ \\mathbf {u} ]} ^ {\\mathrm {C S S}}\\right) \\right. \\\\ + \\mathcal {L} _ {\\text {P r o t o}} \\left(\\mathbf {e} _ {[ \\mathbf {u} ]}, \\mathbf {k} _ {d [ \\mathbf {u} ]}, \\mathbf {k} _ {d [ \\mathbf {u} ]} ^ {\\mathrm {C S S}}\\right) ], \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.483, + 0.562 + ], + "angle": 0, + "content": "where \\(\\mathbf{u}\\) denotes the indices of the unlabeled samples. This cross-type self-supervision (CSS) loss enables self-distillation between the dataset-wise keypoint head and the prototype, mitigating supervision shortage without requiring a teacher model or input duplication." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.563, + 0.348, + 0.577 + ], + "angle": 0, + "content": "The final training loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.589, + 0.483, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M D T}} = \\mathcal {L} _ {\\mathrm {K P L}} + \\mathcal {L} _ {\\mathrm {C S S}}, \\text {w h e r e} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.609, + 0.483, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K P L}} = \\sum_ {d = 1} ^ {D} \\left[ \\mathcal {L} _ {\\mathrm {h m}} \\left(\\mathbf {k} _ {d}, \\mathbf {k} _ {d} ^ {\\mathrm {g t}}\\right) + \\mathcal {L} _ {\\mathrm {P r o t o}} \\left(\\mathbf {e}, \\mathbf {k} _ {d}, \\mathbf {k} _ {d} ^ {\\mathrm {g t}}\\right) \\right], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.66, + 0.383, + 0.675 + ], + "angle": 0, + "content": "and \\(\\mathcal{L}_{\\mathrm{hm}}\\) is the standard JointMSE loss [40]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.688, + 0.224, + 0.705 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.279, + 0.73 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We follow the experimental setup of ViTPose++ [53], using six human pose datasets for training. (1) COCO consists of 57,000 training and 5,000 validation images, annotated with 17 keypoints [23]. (2) AIChallenger (AIC) contains 210,000 training and 30,000 validation images, annotated with 14 keypoints [50]. (3) MPII includes approximately 15,000 training, 2,729 validation, and 5,700 test images, with 16 keypoints [1]. (4) AP-10K is an animal pose dataset with approximately 7,000 training, 1,000 validation, and 2,000 test images [58]. (5) APT-36K, a video-based animal pose dataset, consists of about 25,000 training, 3,600 validation," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "and 7,000 test images [56]. Both AP-10K and APT-36K share the same skeletal format with 17 keypoints, largely following COCO. (6) COCO-WholeBody extends COCO with whole-body annotations covering 133 keypoints [50]. We follow standard evaluation protocols, using the average precision (AP) metric based on object keypoint similarity (OKS) for COCO, AIC, COCO-WholeBody, AP-10K, and APT-36K. For MPII, the percentage of correct keypoints (PCK) metric is used [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.906, + 0.409 + ], + "angle": 0, + "content": "For the domain transfer experiments, we use InterHand2.6M [28] and 3DPW [45]. InterHand2.6M is a two-hand pose estimation dataset containing approximately 2.6M frames, annotated with 21 keypoints. We adopt the ViTPose++ configuration for frame sampling and dataset splits (training, validation, and test). For evaluation, we use PCK, AUC, and EPE, consistent with ViTPose++. 3DPW is an in-the-wild 3D human shape estimation dataset with 18,000 training and 26,000 test frames, annotated with 24 keypoints. We use COCO-style AP and AR as evaluation metrics. In both InterHand2.6M and 3DPW, the projected 2D keypoints are used as ground-truths." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.456 + ], + "angle": 0, + "content": "Further details on the experimental setups, the impact of varying hyperparameters, and failure case analysis can be found in the supplemental document." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.727, + 0.483 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.906, + 0.624 + ], + "angle": 0, + "content": "Our multi-head baseline model is ViTPose++ [53]. We use an input image size of \\(256 \\times 192\\) and apply a flip test in all cases. The model is trained for 100 epochs using the AdamW optimizer with a weight decay factor of 0.1 and a step learning rate scheduler with an initial learning rate of 0.001. The learning rate is reduced by a factor of 0.1 at the 50th and 90th epochs. We use four NVIDIA A100 GPUs for training with the ViT-H backbone and four NVIDIA A6000 GPUs for training with the ViT-B backbone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.906, + 0.777 + ], + "angle": 0, + "content": "We employ a scheduling strategy for both the losses and the learning rate. Initially, we freeze the multi-head model, updating only the embedding module and prototypes for 50 epochs. In the subsequent 40 epochs, we train the multi-heads and the embedding module while keeping the prototypes and the backbone frozen. Finally, during the last 10 epochs, we train the entire network except the prototypes. At this stage, we introduce cross-type self-supervision loss. This progressive optimization approach accelerates training compared to training the entire network from the outset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.863, + 0.804 + ], + "angle": 0, + "content": "4.3. Comparison with existing MDT methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In Tab. 1, we evaluate the competitiveness of our method against existing MDT approaches. For a fair comparison, all methods use the ViT-Base backbone and are trained under our dataset configuration. While UniHCP performs well on AIC, it achieves the lowest scores on COCO and MPII, resulting in an average of 66.8. We hypothe" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12282" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.885, + 0.3 + ], + "angle": 0, + "content": "
MethodBackboneDatasetsCOCOMPII
(AP↑)(AP50↑)(AP75↑)(AR↑)(PCKh↑)(PCKh@0.1↑)
HRNetHRNet-W48175.190.682.280.490.3†33.1†
HRFormerHRFormer-B175.690.882.880.8--
SimCCHRNet-W48176.190.682.981.290.036.8
PCT†Swin-B177.791.284.782.192.5-
PCT†Swin-H179.391.585.9---
UniHCPViT-B3376.1---93.2††-
ViTPose++-BViT-B677.073.484.082.692.839.1
ViTPose++-HViT-H679.491.985.784.894.241.6
OursViT-B677.390.884.282.493.239.3
OursViT-H679.591.985.884.594.242.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.31, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Table 3. Comparison with state-of-the-art methods on COCO and MPII. Methods with \\( {}^{ \\dagger } \\) uses \\( {256} \\times {256} \\) input size. \\( {}^{\\dagger \\dagger } \\) indicates additional fine-tuning on each dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.482, + 0.501 + ], + "angle": 0, + "content": "size that UniHCP's parameter-sharing strategy causes well-represented pose datasets to dominate, hindering its generalization across different skeleton structures. ViTPose++ serves as both a multi-head model and the baseline for our method. Our algorithm surpasses ViTPose++ by 0.3 in average score and outperforms it on individual datasets. Given that these benchmark scores are highly saturated and our method maintains the same computation complexity as ViTPose++ during inference, this improvement is significant." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.503, + 0.483, + 0.7 + ], + "angle": 0, + "content": "To further assess generalization, we test our method on animal and whole-body pose estimation datasets. Table 2 shows that our method consistently improves performance, despite significant domain discrepancies. UniHCP lags behind with an average score of 27.2, likely due to its parameter sharing strategy, which does not address skeleton heterogeneity. Our method also surpasses ViTPose++ by an average score of 4.3, demonstrating superior generalization beyond human skeletons. Specifically, it achieves gains of 0.9 AP on AP-10K, 11.2 AP on APT-36K, and 0.8 AP on COCO-WholeBody. Unlike existing methods that compromise either mainstream performance or adaptation to distant domains, our approach enhances both synergistically." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.712, + 0.433, + 0.728 + ], + "angle": 0, + "content": "4.4. Comparison to state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In Tab. 3, we compare our method with state-of-the-art (SOTA) approaches on the COCO and MPII benchmarks, using scores reported in the respective papers. We also evaluate our method against SOTA approaches designed for single datasets, including HRNet [40], HRFformer [60], SimCC [21], and PCT [11]. On the COCO validation set, our method ranks second with the ViT-B backbone and achieves SOTA performance with the ViT-H backbone, reaching 79.5 AP. On the MPII validation set, our ViT-B model performs comparably to the fine-tuned UniHCP and surpasses the SOTA on the PCKh@0.1 metric by 0.4 PCKh." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.363, + 0.892, + 0.462 + ], + "angle": 0, + "content": "
MethodInterHand2.6M3DPW
PCK↑AUC↑EPE↓AP↑AR↑
UniHCP98.587.03.7257.961.8
ViTPose++98.386.24.0281.785.2
Ours98.687.13.7083.687.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.472, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Table 4. Transfer results on hand (Interhand2.6M) and human shape (3DPW) domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.606 + ], + "angle": 0, + "content": "These results demonstrate that our method effectively balances adaptability to diverse domains while maintaining competitive performance on standard benchmarks, aligning with our primary objective of improving generalization beyond mainstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.619, + 0.836, + 0.636 + ], + "angle": 0, + "content": "4.5. Transfer to hands and human shapes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.764 + ], + "angle": 0, + "content": "PoseBH has the capability to generalize its learned representations beyond the training dataset. We validate this through domain transfer experiments by fine-tuning the model separately on the InterHand2.6M and 3DPW datasets. InterHand2.6M is a two-hand pose estimation dataset containing 21 MANO keypoints [34], while 3DPW is an in-the-wild human shape estimation dataset with 24 SMPL keypoints [26]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "During fine-tuning, we freeze the embedding module and nonparametrically learn the prototypes for the target datasets. After convergence, we fine-tune the backbone and head weights alongside the prototype loss. Freezing the embedding module ensures that the previously learned prototypes and those of the target dataset remain aligned within the same embedding space, preserving the learned representations during transfer. Since these experiments do not involve skeleton heterogeneity across datasets, no semi-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12283" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.087, + 0.885, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.326, + 0.754, + 0.34 + ], + "angle": 0, + "content": "Figure 4. Comparative pose estimation results on human (left) and animal (dog; right)." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.349, + 0.885, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.603, + 0.713, + 0.617 + ], + "angle": 0, + "content": "Figure 5. Comparative results on InterHand2.6M (a-c) and 3DPW (e-f)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.674 + ], + "angle": 0, + "content": "supervised learning challenges arise, and CSS loss is not applied." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.827 + ], + "angle": 0, + "content": "In Tab. 4, we report results on the InterHand2.6M and 3DPW test sets using the ViT-B backbone. Given its large dataset size (\\(\\sim\\)2.6M instances), InterHand2.6M exhibits a smaller performance gap among different methods. Nonetheless, our method surpasses ViTPose++ by 0.3 PCK, 0.9 AUC, and 0.32 EPE, while also outperforming UniHCP by 0.1 PCK, 0.1 AUC, and 0.02 EPE. For the smaller 3DPW dataset, our method achieves a notable performance gain, surpassing the baseline by 1.9 AP and 1.9 AR, demonstrating strong generalization in keypoint embeddings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.253, + 0.849 + ], + "angle": 0, + "content": "4.6. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In Tab. 5, we present the results of ablation experiments on the COCO (CO), AIC (AI), MPII (MP), AP-10K (AP), APT-36K (APT), and COCO-Wholebody (CW) validation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.827 + ], + "angle": 0, + "content": "sets using the ViT-Base backbone. The top row represents a multi-head baseline with a shared backbone and dataset-specific heads. This conventional multi-head approach exhibits relatively low performance, particularly on the AP-10K and APT-36K validation sets, due to substantial domain discrepancies. Our keypoint prototype method achieves an average score of 66.3, improving the baseline by 0.8. Additionally, incorporating cross-type self-supervision provides an additional average performance gain of 0.2. Collectively, these enhancements enable our final configuration to outperform the multi-head baseline by an average of 2.4 points across all six datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.665, + 0.85 + ], + "angle": 0, + "content": "4.7. Visual analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Figure 4 presents examples of human and animal pose estimation, demonstrating that our algorithm achieves more accurate keypoint localization under complex interactions and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12284" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.09, + 0.191, + 0.101 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.101, + 0.48, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.427, + 0.483, + 0.456 + ], + "angle": 0, + "content": "Figure 6. Examples of SMPL fitting using KITRO and with our 2D predictions." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.474, + 0.457, + 0.558 + ], + "angle": 0, + "content": "
MethodCOAIMPAPAPTCWAvg.
Baseline77.031.693.175.374.857.168.2
+LProto77.031.893.076.486.457.570.4
+LCSS77.332.193.276.786.557.970.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.597 + ], + "angle": 0, + "content": "Table 5. Ablation study of the proposed method. \\( \\mathcal{L}_{\\mathrm{Proto}} \\): keypoint prototype learning. \\( \\mathcal{L}_{\\mathrm{CSS}} \\): cross-type self-supervision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.658 + ], + "angle": 0, + "content": "occlusions, as seen in the right-hand keypoints of the standing person (left) and the right front toe of the dog (right)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We further compare the pose estimation results on InterHand2.6M and 3DPW in Fig. 5. In (a), ViTPose++ incorrectly assigns a little finger joint to the opposite hand, whereas our method provides more accurate localization. In (b), ViTPose++ struggles with occlusion from the other hand, producing noisy predictions, while our method correctly separates them. In (c), due to severe self-occlusion, ViTPose++ misestimates occluded keypoints and assigns them to the opposite hand, while our method generalizes well in challenging self-occlusion scenarios. In (d), ViTPose++ produces an implausible foot and hand pose under heavy occlusion, while our method remains robust. In (e), ViTPose++ swaps the left and right legs, while our method correctly distinguishes them. In (f), background clutter causes ViTPose++ to produce noisy predictions for the left hand, while our method correctly identifies the lo" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.729, + 0.106 + ], + "angle": 0, + "content": "cation of the occluded left hand." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.111, + 0.907, + 0.308 + ], + "angle": 0, + "content": "To further assess the effectiveness of our 2D SMPL keypoint predictions on the 3DPW test set, we employ a SMPL optimization technique. KITRO [55] optimizes SMPL parameters using 2D keypoint reprojection loss. For a comparative visual analysis, we replace KITRO's 2D keypoint predictions with ours. As shown in Figure 6(a) and (b), KITRO fails to align the mesh with the actual feet, whereas our method achieves more precise alignment. In (c), KITRO mislocalizes the right hand under partial occlusion, while our 2D predictions remain robust. In (d), KITRO exhibits misalignment in both the hands and feet, while replacing its 2D predictions with ours significantly improves the alignment of peripheral body parts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.372 + ], + "angle": 0, + "content": "In the supplemental material, we provide a visualization of the prototypes generated by our algorithm, illustrating how they effectively capture the diversity of representations within the embedding space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.397, + 0.634, + 0.413 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.909, + 0.639 + ], + "angle": 0, + "content": "This paper presents a new multi-dataset training (MDT) framework for pose estimation across diverse domains and datasets. Traditional approaches encounter challenges with label heterogeneity, dataset-specific parameters, and limited domain generalization. Our approach addresses the limitations with two main contributions. First, to resolve label heterogeneity and dataset-specific parameters, we employ nonparametric prototypes to unify diverse keypoint annotations across multiple datasets. Second, to address the limited multi-dataset supervision, we propose a cross-type self-supervision mechanism. The proposed method has demonstrated substantial improvements over existing MDT across a broad range of datasets, including human pose, animal pose, hand shape, and human shape estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.909, + 0.77 + ], + "angle": 0, + "content": "Limitations and future work. A limitation of our approach is that it requires at least a few-shot sample to learn prototypes, making zero-shot inference for unseen skeletons infeasible. Future work should explore extending our method for zero-shot human pose estimation. Future work should also investigate adapting the approach to 3D domains e.g. using axis-angle rotation representations, to facilitate representation learning on 3D manifolds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by the National Research Foundation of Korea (NRF) grant (No. 2021R1A2C2012195) and by the Institute of Information & communications Technology Planning & Evaluation (IITP) grants No. RS-2019-II191906 (AI Graduate School Program, POSTECH), No. RS-2020-II201336 (AI Graduate School Program, UNIST), and RS-2022-II220290, funded by the Korea government (MSIT)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12285" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.183 + ], + "angle": 0, + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2D human pose estimation: New benchmark and state of the art analysis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Mykhaylo Andriluka, Umar Iqbal, Eldar Insafutdinov, Leonid Pishchulin, Anton Milan, Juergen Gall, and Bernt Schiele. PoseTrack: A benchmark for human pose estimation and tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5167-5176, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Yanbei Chen, Manchen Wang, Abhay Mittal, Zhenlin Xu, Paolo Favaro, Joseph Tighe, and Davide Modolo. ScaleDet: A scalable multi-dataset object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7288-7297, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[5] Hongsuk Choi, Gyeongsik Moon, JoonKyu Park, and Kyoung Mu Lee. Learning to estimate robust 3D human mesh from in-the-wild crowded scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.548 + ], + "angle": 0, + "content": "[6] Yuanzheng Ci, Yizhou Wang, Meilin Chen, Shixiang Tang, Lei Bai, Feng Zhu, Rui Zhao, Fengwei Yu, Donglian Qi, and Wanli Ouyang. UniHCP: A unified model for human-centric perceptions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17840-17852, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.606 + ], + "angle": 0, + "content": "[7] Mickael Cormier, Aris Clepe, Andreas Specker, and Jürgen Beyerer. Where are we with human pose estimation in real-world surveillance? In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 591-601, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[8] Anurag Das, Yongqin Xian, Dengxin Dai, and Bernt Schiele. Weakly-supervised domain adaptive semantic segmentation with prototypical contrastive learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15434-15443, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[9] Jon Donnelly, Alina Jade Barnett, and Chaofan Chen. Deformable protopnet: An interpretable image classifier using deformable prototypes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10265-10275, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[10] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4320–4329, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[11] Zigang Geng, Chunyu Wang, Yixuan Wei, Ze Liu, Houqiang Li, and Han Hu. Human pose as compositional tokens. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human POSEitioning System (HPS): 3D human pose estimation and self-localization in large scenes from" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "body-mounted sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4318-4329, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[13] Or Hirschorn and Shai Avidan. A graph-based approach for category-agnostic pose estimation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.166, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[14] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. NeuMan: Neural human radiance field from a single video. In European Conference on Computer Vision (ECCV), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.223, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[15] Roy Kapon, Guy Tevet, Daniel Cohen-Or, and Amit H Bermano. MAS: Multi-view ancestral sampling for 3D motion generation using 2D diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1965–1974, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[16] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. arXiv preprint arXiv:2408.12569, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[17] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[18] Alberto Lamas, Siham Tabik, Antonio Cano Montes, Francisco Pérez-Hernández, Jorge García, Roberto Olmos, and Francisco Herrera. Human pose estimation for mitigating false negatives in weapon detection in video-surveillance. Neurocomputing, 489:488-503, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[19] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. MSeg: A composite dataset for multidomain semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[20] Duy Tho Le, Chenhui Gou, Stavya Datta, Hengcan Shi, Ian Reid, Jianfei Cai, and Hamid Rezatofighi. JRDB-PanoTrack: An open-world panoptic segmentation and tracking robotic dataset in crowded human environments. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22325-22334, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[21] Yanjie Li, Sen Yang, Peidong Liu, Shoukui Zhang, Yunxiao Wang, Zhicheng Wang, Wankou Yang, and Shu-Tao Xia. SimCC: A simple coordinate classification perspective for human pose estimation. In European Conference on Computer Vision (ECCV), pages 89-106. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[22] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-X: A large-scale 3D expressive whole-body human motion dataset. Advances in Neural Information Processing Systems (NeurIPS), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Sun-Ao Liu, Yiheng Zhang, Zhaofan Qiu, Hongtao Xie, Yongdong Zhang, and Ting Yao. Learning orthogonal prototypes for generalized few-shot semantic segmentation. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12286" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[25] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. HumanGaussian: Text-driven 3D human generation with gaussian splatt-ting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6646-6657, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[27] Yawen Lu, Dongfang Liu, Qifan Wang, Cheng Han, Yiming Cui, Zhiwen Cao, Xueling Zhang, Yingjie Victor Chen, and Heng Fan. ProMotion: Prototypes as motion learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28109-28119, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[28] Gyeongsik Moon, Shouou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. InterHand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[29] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. NeuralAnnot: Neural annotator for 3D human mesh training sets. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[30] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three Recipes for Better 3D PseudogTs of 3D Human Mesh Estimation in the Wild. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[31] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3D gaussian avatar. In European Conference on Computer Vision (ECCV), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[32] Haocong Rao and Chunyan Miao. TranSG: Transformer-based skeleton graph prototype contrastive learning with structure-trajectory prompted reconstruction for person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22118-22128, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[33] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D human motion model for robust pose estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[34] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[35] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7420-7430, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] István Sárándi, Alexander Hermans, and Bastian Leibe. Learning 3D human pose estimation from dozens of datasets using a geometry-aware autoencoder to bridge between skeleton formats. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[37] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.177 + ], + "angle": 0, + "content": "[38] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[39] Markus D Solbach and John K Tsotsos. Vision-based fallen person detection for the elderly. In IEEE/CVF international conference on computer vision workshops (ICCVW), pages 1433-1442, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.906, + 0.291 + ], + "angle": 0, + "content": "[40] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Black Michael J., and Tao Mei. Monocular, one-stage, regression of multiple 3D people. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.404 + ], + "angle": 0, + "content": "[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3D people in depth. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.906, + 0.462 + ], + "angle": 0, + "content": "[43] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xR-EgoPose: Egocentric 3D human pose from an hmd camera. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7728-7738, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[44] Edward Vendrow, Duy Tho Le, Jianfei Cai, and Hamid Rezatofighi. JRDB-Pose: A large-scale dataset for multiperson pose estimation and tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[45] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[46] Chong Wang, Yuyuan Liu, Yuanhong Chen, Fengbei Liu, Yu Tian, Davis McCarthy, Helen Frazer, and Gustavo Carneiro. Learning support and trivial prototypes for interpretable image classification. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2062-2072, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.906, + 0.73 + ], + "angle": 0, + "content": "[47] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3D human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13031-13040, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[48] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards universal object detection by domain attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7289-7298, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.856 + ], + "angle": 0, + "content": "[49] Zhongdao Wang, Hengshuang Zhao, Ya-Li Li, Shengjin Wang, Philip Torr, and Luca Bertinetto. Do different tracking tasks require different appearance models? Advances in Neural Information Processing Systems (NeurIPS), 34:726-738, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[50] Jiahong Wu, He Zheng, Bo Zhao, Yixin Li, Baoming Yan, Rui Liang, Wenjia Wang, Shipei Zhou, Guosen Lin, Yanwei Fu, et al. AI Challenger: A large-scale dataset for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12287" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.119 + ], + "angle": 0, + "content": "going deeper in image understanding. arXiv preprint arXiv:1711.06475, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[51] Hang Xu, Linpu Fang, Xiaodan Liang, Wenxiong Kang, and Zhenguo Li. Universal-RCNN: Universal object detector via transferable graph r-cnn. In AAAI Conference on Artificial Intelligence (AAAI), pages 12492–12499, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.483, + 0.245 + ], + "angle": 0, + "content": "[52] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In European Conference on Computer Vision (ECCV), pages 398–416. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[53] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose++: Vision transformer for generic body pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[54] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. UNIK: A unified framework for real-world skeleton-based action recognition. British Machine Visison Conference (BMVC), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.427 + ], + "angle": 0, + "content": "[55] Fengyuan Yang, Kerui Gu, and Angela Yao. KITRO: Refining human mesh by 2D clues and kinematic-tree rotation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1052-1061, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[56] Yuxiang Yang, Junjie Yang, Yufei Xu, Jing Zhang, Long Lan, and Dacheng Tao. APT-36K: A large-scale benchmark for animal pose estimation and tracking. Advances in Neural Information Processing Systems (NeurIPS), 35:17301-17313, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.553 + ], + "angle": 0, + "content": "[57] Sung Whan Yoon, Jun Seo, and Jaekyun Moon. TapNet: Neural network augmented with task-adaptive projection for few-shot learning. In International conference on machine learning, pages 7115-7123. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[58] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. AP-10K: A benchmark for animal pose estimation in the wild. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.483, + 0.664 + ], + "angle": 0, + "content": "[59] Xinyi Yu, Xin Zhang, Chengjun Xu, and Linlin Ou. Humanrobot collaborative interaction with human perception and action recognition. Neurocomputing, 563:126827, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.72 + ], + "angle": 0, + "content": "[60] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. HRFormer: High-resolution transformer for dense prediction. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[61] Baoquan Zhang, Xutao Li, Yunming Ye, Zhichao Huang, and Lisai Zhang. Prototype completion with primitive knowledge for few-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3754-3762, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[62] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3D: A unified baseline for multi-dataset 3D object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9253-9262, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[63] Xiangyun Zhao, Samuel Schulter, Gaurav Sharma, Yi-Hsuan Tsai, Manmohan Chandraker, and Ying Wu. Object detection with a unified label space from multiple datasets. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "European Conference on Computer Vision (ECCV), pages 178-193. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[64] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking Semantic Segmentation: A Prototype View. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2582-2593, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[65] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7571-7580, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12288" + } + ] +] \ No newline at end of file diff --git a/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/6abac681-ec51-468a-ba82-3e1de6acd774_origin.pdf b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/6abac681-ec51-468a-ba82-3e1de6acd774_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..aeb3b319ee3c7f0e8d31e95d626fecff5dc8d8e4 --- /dev/null +++ b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/6abac681-ec51-468a-ba82-3e1de6acd774_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0fd9939948dc9c46a5490b2655e111c647b6cdf2703f57d47e2b587edc50c94 +size 9989793 diff --git a/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/full.md b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..81231b4c78ff21c5dfd4e633a885ec845d629342 --- /dev/null +++ b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/full.md @@ -0,0 +1,357 @@ +# PoseBH: Prototypical Multi-Dataset Training Beyond Human Pose Estimation + +Uyoung Jeong $^{1}$ Jonathan Freer $^{2}$ Seungryul Baek $^{1}$ Hyung Jin Chang $^{2}$ Kwang In Kim $^{3}$ $^{1}$ UNIST $^{2}$ University of Birmingham $^{3}$ POSTECH + +# Abstract + +We study multi-dataset training (MDT) for pose estimation, where skeletal heterogeneity presents a unique challenge that existing methods have yet to address. In traditional domains, e.g. regression and classification, MDT typically relies on dataset merging or multi-head supervision. However, the diversity of skeleton types and limited cross-dataset supervision complicate integration in pose estimation. To address these challenges, we introduce PoseBH, a new MDT framework that tackles keypoint heterogeneity and limited supervision through two key techniques. First, we propose nonparametric keypoint prototypes that learn within a unified embedding space, enabling seamless integration across skeleton types. Second, we develop a cross-type self-supervision mechanism that aligns keypoint predictions with keypoint embedding prototypes, providing supervision without relying on teacher-student models or additional augmentations. PoseBH substantially improves generalization across whole-body and animal pose datasets, including COCO-WholeBody, AP-10K, and APT-36K, while preserving performance on standard human pose benchmarks (COCO, MPII, and AIC). Furthermore, our learned keypoint embeddings transfer effectively to hand shape estimation (InterHand2.6M) and human body shape estimation (3DPW). The code for PoseBH is available at: https://github.com/uyoung-jeong/PoseBH. + +# 1. Introduction + +The demand for human pose estimators has grown substantially in recent years, driving their adoption across diverse applications, including 3D avatar generation [14, 25, 31], motion generation [15, 22, 33], human-robot interaction [20, 44, 59], safety monitoring [7, 18, 39], and VR/AR tracking [12, 43, 47]. These domains pose distinct operational challenges, often necessitating significant domain adaptation for robust pose estimation. To address this, we aim to train a human pose estimator that generalizes effectively across multiple datasets. + +However, multi-dataset training (MDT) for pose estimation presents two key challenges. First, keypoints + +![](images/43a33a2afd1a03df425af3a78dc42b1cae43b9c136b7c98400c354781c7235fa.jpg) +COCO + +![](images/19c165f3a24b569e93979b415d9428a0c35a93f90a4273c64793192e6c213cd4.jpg) +AIC + +![](images/f3242e21cae62c2ccf64299f92785151df94bdd3362da83fdac61e6e2eb32aba.jpg) +MPII + +![](images/008abc2c4d83cb96d3474a78c51c19a47b89dea96caf28779ad26e1c297a5778.jpg) +WholeBody + +![](images/b711e282a34ca825b6b61038b9d3a57189d962ad717232e160e8d377614d375d.jpg) +AP-10K + +![](images/8fa0aa361a62d037be1304f667b77e736bd9c3d47a2a882eee378c034a550a4a.jpg) +APT-36K + +![](images/9e13785908194b70429277d31d109d9a0e1f9beb14d1dca21d3e9a262edd6ff5.jpg) +3DPW + +![](images/d4f9a3fbc1c16f9aba82b1c44972e8feb9c32ce26e52c3bfee63f222f467cda4.jpg) +InterHand2.6M +Figure 1. PoseBH unifies diverse skeleton formats, including humans, hands, and animals. The displayed skeletons show pose estimation results from our method, with 3DPW and InterHand2.6M predictions from a transferred model. + +are heterogeneous across different datasets. For example, COCO [23] and MPII [1] define 17 and 16 keypoints, respectively, with 9 non-overlapping ones. Even when datasets share keypoints with identical names, their localization characteristics often differ due to domain gaps. For example, while COCO and animal datasets (e.g., APT-10K [58] and APT-36K [56]) use common keypoints names such as eyes, nose, and hips, significant anatomical differences exist (see Fig. 1). As a result, naively merging the identically named keypoints is ineffective. + +Second, pose estimation MDT suffers from a lack of adequate supervision. In MDT, multiple skeleton types from different datasets are trained for each input instance. However, only a single skeleton-type ground truth is provided per input, leaving others unsupervised. This limitation resembles a semi-supervised learning challenge, constraining performance on smaller datasets. Existing semi-supervised learning methods typically rely on input augmentation or teacher-student knowledge distillation, both of which require substantial extra computational overhead. + +Existing approaches fail to fully address these challenges, often relying on a shared backbone network [6] or + +a mixture of experts [53] to preserve downstream performance. However, these methods tend to focus on mainstream datasets like COCO or MPII, resulting in degraded performance on underrepresented datasets. We propose a method that enhances performance in discrepant domains while maintaining accuracy on standard human pose datasets. A straightforward fine-tuning approach on the target domain often disrupts performance on source datasets, as it fails to balance differences in skeleton structures and dataset distributions. Therefore, effective generalization requires jointly harmonizing skeletal representations and dataset diversity during training. + +We achieve this through two primary techniques. First, we introduce prototypes to represent keypoints as embedding vectors, providing a unified representation across datasets. We regress pixel-wise keypoint embeddings from backbone features and match them to corresponding prototypes, automating keypoint mapping across datasets. This enables the integration of diverse keypoint types, including whole-body and animal keypoints, despite significant skeletal differences. Furthermore, our keypoint prototypes are non-learnable parameters, offering adaptability and computational efficiency for domain transfer. + +Second, we generate reliable self-supervision signals using the predicted keypoints and their embeddings, a process we term cross-type self-supervision. Since keypoint predictions can be obtained via a dot product between embeddings and prototypes, we effectively establish two distinct keypoint prediction modalities. By aligning these predictions, we filter out noisy keypoints and produce reliable labels for supervising unlabeled keypoints. This process enhances the utility of embeddings without requiring additional networks or input augmentation. As a result, we provide effective supervision for underrepresented and discrepant datasets, achieving a higher degree of generalizability. + +Our main technical contributions are as follows: + +- Keypoint prototypes for learning arbitrary keypoints from multiple datasets ensuring high transferability. +- A cross-type self-supervision mechanism that aligns keypoint regression outputs with keypoint embeddings, enriching supervision for unlabeled keypoints. +- Strong generalization across estimation tasks, including whole-body, animal, and hand pose estimation, as well as human shape estimation. + +# 2. Related works + +Multi-dataset training has been widely explored in various computer vision tasks. Relabeling and pseudo-labeling are among the most straightforward yet resource-intensive approaches. For example, MSeg [19] employed relabeling for segmentation, while [63] used offline pseudo-labeling for object detection. However, both methods face scalability challenges, as adding a new dataset requires reprocessing + +annotations for all previously processed datasets to redefine the class set, leading to a quadratic computational cost. + +Several approaches have been proposed to improve scalability. In object detection, [4, 65] aim to unify object class labels by merging similar categories. However, they struggle to integrate localization features of the same object due to the absence of spatial information in class logits. UniTrack learns a unified identity embedding through self-supervised learning for multiple object-tracking tasks, including human pose tracking [49]. However, it prioritizes pose propagation over estimation and relies exclusively on the PoseTrack18 [2] skeleton, without incorporating multiple pose datasets. + +Another class of methods [48, 51, 54, 62] employs network modules, such as graph convolutional networks (GCN), for cross-dataset adaptation. However, these dataset-specific modules limit domain transferability, as their input and output sizes are dictated by the label structure, making them incompatible with datasets that have different keypoint formats. In contrast, our approach achieves flexible transferability without relying on dataset-specific domain adaptation modules. + +Human pose estimation with MDT. In 3D human shape estimation, several methods have used pseudo-ground truths (GTs) generated by the SMPL model for MDT [5, 17, 41, 42]. However, these pseudo-GTs are often inaccurate due to depth ambiguity in 2D images and the limited expressiveness of the parametric 3D model [29, 30]. Similar to [65], Sarandi et al.'s affine-combining autoencoder (ACAE) merges 3D joint coordinates via an autoencoder [36]. However, the autoencoder's learnable parameters are dataset-dependent, limiting its generalizability. Furthermore, ACAE does not preserve pixel-level features, making it unsuitable when spatial information is crucial. + +Recent works on category-agnostic pose estimation integrate various objects, such as chairs and vehicles, by combining multiple datasets [13, 52]. These methods follow a relabeling approach, creating a superset of keypoint classes and detecting only the keypoints defined in their relabeled dataset. As a result, they do not address keypoint heterogeneity or label sparsity issues, as training and evaluation occur within a single labeled dataset. In contrast, our method uses multiple skeletons to improve generalization across diverse downstream domains. + +In 2D human pose estimation, UniHCP performs multiple human perception tasks using task-specific queries without incorporating task-specific layers inside the backbone [6]. This design limits its ability to generalize to distant domains such as animals and whole-body pose estimation. ViTPose++ [53] and Sapiens [16] employ multi-head architectures for MDT but do not address skeleton heterogeneity or the resulting label sparsity. + +![](images/aa025e7daeaf31ca940cd2d70076f26054a16400641f4148c3a85d9ce74fa53d.jpg) +Figure 2. Overview of the PoseBH architecture. During training, the embedding head maps the backbone features into a unified keypoint embedding space. By matching these embeddings with prototypes, dataset-specific keypoint heatmaps $\mathbf{k}_{\mathrm{coco}}^{\mathrm{P}}$ are generated. Prototypes are updated nonparametrically from the embeddings. During inference, the embedding head and the subsequent procedures are removed. + +![](images/54d4bb2b2991acf752d6e8ec4020400bdf2f0ce860d387059179f41c9b62041c.jpg) + +Prototype learning can be interpreted as proxy-based metric learning [35], where the centers of a fixed set of clusters serve as prototypes. This approach has been studied in various domains, including semantic segmentation [8, 10, 64], person re-identification [32], meta-learning [24, 38, 57, 61], and visual interpretability [3, 9, 46]. + +ProtoSeg [64] introduces a nonparametric approach to semantic segmentation by assigning multiple prototypes per class. ProMotion [27] employs prototypes to unify scene depth estimation and optical flow tasks. While our method draws inspiration from ProtoSeg and aligns with ProMotion in exploiting the benefits of prototypes for MDT, it differs in three key aspects. First, our approach applies keypoint-level supervision, which is inherently sparser than segmentation, depth, or optical flow. Second, we specifically address skeleton heterogeneity and cross-dataset relationships, aspects not considered by ProtoSeg and ProMotion. Third, while ProtoSeg and ProMotion apply fully supervised learning, our method tackles the challenge of unlabeled data, a problem unique to MDT pose estimation. + +# 3. Method + +We focus on multi-dataset training (MDT) for 2D human pose estimation. Given $D$ datasets, our model predicts a keypoint heatmap for a person, represented as an array $\mathbf{k}_d \in \mathbb{R}^{J_d \times H \times W}$ , where $d$ is the dataset index, $J_{d}$ is the number of keypoints in the $d$ -th dataset, and $H$ and $W$ denote the height and width of the heatmap, respectively. + +# 3.1. Multi-head baseline + +The multi-head baseline model consists of a backbone network that extracts an output backbone feature map from + +an input RGB image. The backbone network is shared across all datasets, while individual keypoint heads independently regress dataset-wise keypoint heatmaps $\{\mathbf{k}_d\}_{d=1}^D$ , each trained with its corresponding labels. The keypoint heads share the same architecture, except for the final layer, where the output channel size is the respective number of keypoints. This generic baseline does not share the keypoint output space, limiting the MDT effect up to the backbone. + +# 3.2. Keypoint prototype learning + +Overview. Our goal is to learn a unified keypoint space that standardizes keypoint output formats across datasets, scales with dataset size, and retains dataset-specific localization accuracy. To achieve this, we reformulate keypoint regression as a prototype-based distance metric learning problem (see Fig. 2). Our shared keypoint embedding module generates a normalized keypoint embedding map $\mathbf{e} \in \mathbb{R}^{F \times H \times W}$ from an input feature map. The module consists of two deconvolution layers, followed by a residual block and two convolution layers. Additional details are provided in the supplementary document. + +Using embeddings and prototypes, we perform keypoint classification by matching embeddings with prototypes $\mathbf{P} \in \mathbb{R}^{J \times M \times F}$ , where $F$ is the embedding dimension, $J$ is the total number of keypoints across all datasets, $M$ is the number of in-class prototypes. For each heatmap pixel, we compute the cosine similarity between the prototypes and the embedding vector to obtain keypoint predictions: + +$$ +\mathbf {k} _ {j, y, x} ^ {\mathrm {P}} = \max _ {m} \frac {\mathbf {P} _ {(j , m , :)} \cdot \mathbf {e} _ {(: , y , x)}}{\left\| \mathbf {P} _ {(j , m , :)} \right\| \left\| \mathbf {e} _ {(: , y , x)} \right\|}, \tag {1} +$$ + +where $x,y$ denote pixel indices in the heatmap, and $\mathbf{P}_{(j,m,:)}$ is a row vector of size $F$ , extracted $\mathbf{P}\in \mathbb{R}^{J\times M\times F}$ along its + +first two dimensions. For each keypoint class $j$ , $\mathbf{k}_{(j,:;:)}^{\mathrm{P}} \in \mathbb{R}^{H \times W}$ stores the matching score of the most probable prototype among $M$ prototypes. We refer to this as the prototype keypoint heatmap $\mathbf{k}^{\mathrm{P}}$ to distinguish it from the multihead outputs $\mathbf{k}$ . + +For the remainder of this paper, we use the following notation: $\mathbf{A}_{(:,a)}$ denotes the 2D slice of a 3D array $\mathbf{A}$ indexed at $a$ along the third dimension, while $\mathbf{A}_{(:,b,:)}$ represents the 2D slice at index $b$ along the second dimension. Similarly, $\mathbf{A}_{(d:e,:;)}$ represents the 3D sub-array spanning indices $d$ to $e$ along the first dimension. For a 2D array $\mathbf{B}$ , $\mathbf{B}_{(:,c)}$ indicates the $c$ -th column vector, whereas $\mathbf{B}_{(d,c)}$ refers to the element at position $(d,c)$ . + +Prototype learning. We adopt a nonparametric learning approach [64] to train $\mathbf{P}$ . First, we obtain logits $\{\mathbf{l}_j\}_{j=1}^J \subset \mathbb{R}^{M \times N}$ and targets $\{\mathbf{t}_j \in \mathbb{R}^{N_j}\}_{j=1}^J$ , where $N_j$ is the number of foreground samples (i.e., pixels in the ground-truth (GT) heatmap with nonzero values) for the $j$ -th joint, and $N$ is the total number of foreground samples across all joints. We define logits as sampled embedding vectors and targets as keypoint class labels, with each keypoint class selecting an in-class prototype from $M$ candidates. + +Since no GT selection is explicitly provided, we assign an in-class prototype for each logit via online clustering using Sinkhorn-Knopp iteration with an equipartition constraint [37], and compute $\mathbf{t}_j$ as follows: + +$$ +\mathbf {l} _ {j} = \mathbf {P} _ {(j,:;:)} \bar {\mathbf {e}}, \tag {2} +$$ + +$$ +\mathbf {t} _ {j} = \underset {m} {\arg \max } \operatorname {d i a g} (\mathbf {u}) \exp \left(\frac {\mathbf {l} _ {j}}{\kappa}\right) \operatorname {d i a g} (\mathbf {v}), \tag {3} +$$ + +where $\overline{\mathbf{e}}$ is a 2D array of size $F\times N_{j}$ , obtained by first flattening $\mathbf{e}\in \mathbb{R}^{F\times H\times W}$ along the spatial dimensions into a 2D array of size $F\times (HW)$ and then, selecting $N_{j}$ foreground objects. $\mathrm{diag}(\mathbf{u})$ is a diagonal matrix derived from the vector $\mathbf{u}$ , and $\exp (\cdot)$ is applied element-wise. + +Each prototype is then updated with momentum $\lambda$ : + +$$ +\mathbf {P} _ {j, m} ^ {\text {n e w}} = \lambda \mathbf {P} _ {j, m} ^ {\text {o l d}} + (1 - \lambda) \frac {1}{N _ {j}} \sum_ {n = 1} ^ {N _ {j}} w _ {m n} \overline {{\mathbf {e}}} _ {(;, n)}, \tag {4} +$$ + +where $w_{mn}$ is the clustering assignment weight, calculated similarly to Eq. (3), but using max instead of arg max. + +To optimize the learned embedding with respect to the prototype $\mathbf{P}$ , we employ pixel-prototype contrastive learning, using the pixel-prototype contrastive loss (PPC) and pixel-prototype distance (PPD) from [64]: + +$$ +\mathcal {L} _ {\mathrm {P P C}} \left(\mathbf {l} _ {j}, \mathbf {t} _ {j}\right) = \frac {c _ {n}}{N _ {j}} \sum_ {n = 1} ^ {N _ {j}} \operatorname {C E} \left(\mathbf {l} _ {j, n}, \mathbf {t} _ {j, n}\right), \tag {5} +$$ + +$$ +\mathcal {L} _ {\mathrm {P P D}} \left(\mathbf {l} _ {j}, \mathbf {t} _ {j}\right) = \frac {c _ {n}}{N _ {j}} \sum_ {n = 1} ^ {N _ {j}} \left(1 - \mathbf {l} _ {j, n} ^ {m +}\right) ^ {2}, \tag {6} +$$ + +![](images/4bda5829f2ed1d2ff71d0b311ee56f613716dce5b3382e919783c88b3fc0c80c.jpg) +Figure 3. An illustrative example of cross-type self-supervision. Given a COCO image, we jointly refine the AIC head and AIC prototype predictions to produce reliable AIC heatmaps. + +where $c_{n}$ is the confidence value from the GT heatmap, CE denotes the cross-entropy loss, and $m^{+}$ indexes the first dimension of $\mathbf{l}_d$ . For each dataset $d$ , we collect logits $\mathbf{l}_{j,n} \in \mathbb{R}^{M \times N_d}$ and targets $\mathbf{t}_d \in \mathbb{R}^{N_d}$ , where $N_{d}$ is the number of samples in dataset $d$ . The total prototype loss is then defined as $\mathcal{L}_{\mathrm{Proto}} = \mathcal{L}_{\mathrm{PPC}} + \mathcal{L}_{\mathrm{PPD}}$ . + +Unlike standard contrastive learning, $\mathcal{L}_{\mathrm{Proto}}$ operates within a single dataset and lacks the push term for differing dataset skeletons. To introduce cross-dataset negatives, we apply K-means clustering to all prototypes, forming $K$ (=96) clusters. This clustering is performed only once during training, while online clustering for in-class prototypes is conducted at each iteration. These clusters serve as negative samples, where the embeddings closest to each centroid are used as logits, with their dataset-specific labels as targets. Finally, we obtain $\mathbf{l}_c$ and $\mathbf{t}_c$ for each cluster and apply Eqs. (5) and (6) to enable cross-dataset contrastive learning. + +# 3.3. Cross-type self-supervision + +While our keypoint prototypes enable unified MDT, label sparsity remains a challenge, as each image is annotated for only one dataset, limiting supervision. To address this semi-supervised learning problem, we introduce a self-supervision strategy that bridges keypoint heatmaps and embeddings. By collaboratively filtering out noisy predictions from both the keypoint and embedding heads, we generate reliable keypoint heatmaps for training (see Fig. 3). + +To eliminate uncertain keypoint predictions, we apply two filtering conditions. First, for each keypoint class, the confidence scores of both the keypoint and embedding heads must exceed a threshold $(c_{thr} = 0.25)$ . Second, the root mean square distance between their predictions must be below a threshold $(d_{thr} = 2.1)$ . The filtered predictions are then combined using a weighted average: + +$$ +\widehat {\mathbf {y}} _ {i} = s _ {i} \widehat {\mathbf {y}} _ {i} ^ {\mathrm {k p t}} + (1 - s _ {i}) \widehat {\mathbf {y}} _ {i} ^ {\mathrm {e m b}}, \tag {7} +$$ + +where $s_i = c_i^{\mathrm{kpt}} / (c_i^{\mathrm{kpt}} + c_i^{\mathrm{emb}})$ , $i$ is the keypoint index, and $\widehat{\mathbf{y}}$ and $c$ represent the keypoint prediction and confidence score from each head, respectively. + +From $\widehat{\mathbf{y}}$ , we generate a reliable heatmap $\mathbf{k}^{\mathrm{CSS}}$ following the standard GT heatmap generation process. The loss for + +
MethodGFLOPSCOCOAICMPIIAvg.
UniHCP18.976.832.690.966.8
ViTPose++18.577.031.693.167.2
Ours18.577.332.193.267.5
+ +Table 1. Comparison of different multi-dataset training methods on general human pose benchmarks, with computational complexity measured in GFLOPS during the evaluation phase. + +
MethodAP-10KAPT-36KCOCO-WAvg.
UniHCP56.562.020.146.2
ViTPose++74.176.057.169.1
Ours75.087.257.973.4
+ +Table 2. Comparison of multi-dataset training methods on whole-body and animal domains. COCO-W refers to COCO-WholeBody. The reported metrics are AP(%) + +unlabeled samples is then computed as: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {C S S}} = \sum_ {d = 1} ^ {D} \zeta \left[ \mathcal {L} _ {\mathrm {h m}} \left(\mathbf {k} _ {d [ \mathbf {u} ]}, \mathbf {k} _ {d [ \mathbf {u} ]} ^ {\mathrm {C S S}}\right) \right. \\ + \mathcal {L} _ {\text {P r o t o}} \left(\mathbf {e} _ {[ \mathbf {u} ]}, \mathbf {k} _ {d [ \mathbf {u} ]}, \mathbf {k} _ {d [ \mathbf {u} ]} ^ {\mathrm {C S S}}\right) ], \tag {8} \\ \end{array} +$$ + +where $\mathbf{u}$ denotes the indices of the unlabeled samples. This cross-type self-supervision (CSS) loss enables self-distillation between the dataset-wise keypoint head and the prototype, mitigating supervision shortage without requiring a teacher model or input duplication. + +The final training loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {M D T}} = \mathcal {L} _ {\mathrm {K P L}} + \mathcal {L} _ {\mathrm {C S S}}, \text {w h e r e} \tag {9} +$$ + +$$ +\mathcal {L} _ {\mathrm {K P L}} = \sum_ {d = 1} ^ {D} \left[ \mathcal {L} _ {\mathrm {h m}} \left(\mathbf {k} _ {d}, \mathbf {k} _ {d} ^ {\mathrm {g t}}\right) + \mathcal {L} _ {\mathrm {P r o t o}} \left(\mathbf {e}, \mathbf {k} _ {d}, \mathbf {k} _ {d} ^ {\mathrm {g t}}\right) \right], \tag {10} +$$ + +and $\mathcal{L}_{\mathrm{hm}}$ is the standard JointMSE loss [40]. + +# 4. Experiments + +# 4.1. Experimental setup + +We follow the experimental setup of ViTPose++ [53], using six human pose datasets for training. (1) COCO consists of 57,000 training and 5,000 validation images, annotated with 17 keypoints [23]. (2) AIChallenger (AIC) contains 210,000 training and 30,000 validation images, annotated with 14 keypoints [50]. (3) MPII includes approximately 15,000 training, 2,729 validation, and 5,700 test images, with 16 keypoints [1]. (4) AP-10K is an animal pose dataset with approximately 7,000 training, 1,000 validation, and 2,000 test images [58]. (5) APT-36K, a video-based animal pose dataset, consists of about 25,000 training, 3,600 validation, + +and 7,000 test images [56]. Both AP-10K and APT-36K share the same skeletal format with 17 keypoints, largely following COCO. (6) COCO-WholeBody extends COCO with whole-body annotations covering 133 keypoints [50]. We follow standard evaluation protocols, using the average precision (AP) metric based on object keypoint similarity (OKS) for COCO, AIC, COCO-WholeBody, AP-10K, and APT-36K. For MPII, the percentage of correct keypoints (PCK) metric is used [1]. + +For the domain transfer experiments, we use InterHand2.6M [28] and 3DPW [45]. InterHand2.6M is a two-hand pose estimation dataset containing approximately 2.6M frames, annotated with 21 keypoints. We adopt the ViTPose++ configuration for frame sampling and dataset splits (training, validation, and test). For evaluation, we use PCK, AUC, and EPE, consistent with ViTPose++. 3DPW is an in-the-wild 3D human shape estimation dataset with 18,000 training and 26,000 test frames, annotated with 24 keypoints. We use COCO-style AP and AR as evaluation metrics. In both InterHand2.6M and 3DPW, the projected 2D keypoints are used as ground-truths. + +Further details on the experimental setups, the impact of varying hyperparameters, and failure case analysis can be found in the supplemental document. + +# 4.2. Implementation details + +Our multi-head baseline model is ViTPose++ [53]. We use an input image size of $256 \times 192$ and apply a flip test in all cases. The model is trained for 100 epochs using the AdamW optimizer with a weight decay factor of 0.1 and a step learning rate scheduler with an initial learning rate of 0.001. The learning rate is reduced by a factor of 0.1 at the 50th and 90th epochs. We use four NVIDIA A100 GPUs for training with the ViT-H backbone and four NVIDIA A6000 GPUs for training with the ViT-B backbone. + +We employ a scheduling strategy for both the losses and the learning rate. Initially, we freeze the multi-head model, updating only the embedding module and prototypes for 50 epochs. In the subsequent 40 epochs, we train the multi-heads and the embedding module while keeping the prototypes and the backbone frozen. Finally, during the last 10 epochs, we train the entire network except the prototypes. At this stage, we introduce cross-type self-supervision loss. This progressive optimization approach accelerates training compared to training the entire network from the outset. + +# 4.3. Comparison with existing MDT methods + +In Tab. 1, we evaluate the competitiveness of our method against existing MDT approaches. For a fair comparison, all methods use the ViT-Base backbone and are trained under our dataset configuration. While UniHCP performs well on AIC, it achieves the lowest scores on COCO and MPII, resulting in an average of 66.8. We hypothe + +
MethodBackboneDatasetsCOCOMPII
(AP↑)(AP50↑)(AP75↑)(AR↑)(PCKh↑)(PCKh@0.1↑)
HRNetHRNet-W48175.190.682.280.490.3†33.1†
HRFormerHRFormer-B175.690.882.880.8--
SimCCHRNet-W48176.190.682.981.290.036.8
PCT†Swin-B177.791.284.782.192.5-
PCT†Swin-H179.391.585.9---
UniHCPViT-B3376.1---93.2††-
ViTPose++-BViT-B677.073.484.082.692.839.1
ViTPose++-HViT-H679.491.985.784.894.241.6
OursViT-B677.390.884.282.493.239.3
OursViT-H679.591.985.884.594.242.0
+ +size that UniHCP's parameter-sharing strategy causes well-represented pose datasets to dominate, hindering its generalization across different skeleton structures. ViTPose++ serves as both a multi-head model and the baseline for our method. Our algorithm surpasses ViTPose++ by 0.3 in average score and outperforms it on individual datasets. Given that these benchmark scores are highly saturated and our method maintains the same computation complexity as ViTPose++ during inference, this improvement is significant. + +To further assess generalization, we test our method on animal and whole-body pose estimation datasets. Table 2 shows that our method consistently improves performance, despite significant domain discrepancies. UniHCP lags behind with an average score of 27.2, likely due to its parameter sharing strategy, which does not address skeleton heterogeneity. Our method also surpasses ViTPose++ by an average score of 4.3, demonstrating superior generalization beyond human skeletons. Specifically, it achieves gains of 0.9 AP on AP-10K, 11.2 AP on APT-36K, and 0.8 AP on COCO-WholeBody. Unlike existing methods that compromise either mainstream performance or adaptation to distant domains, our approach enhances both synergistically. + +# 4.4. Comparison to state-of-the-art methods + +In Tab. 3, we compare our method with state-of-the-art (SOTA) approaches on the COCO and MPII benchmarks, using scores reported in the respective papers. We also evaluate our method against SOTA approaches designed for single datasets, including HRNet [40], HRFformer [60], SimCC [21], and PCT [11]. On the COCO validation set, our method ranks second with the ViT-B backbone and achieves SOTA performance with the ViT-H backbone, reaching 79.5 AP. On the MPII validation set, our ViT-B model performs comparably to the fine-tuned UniHCP and surpasses the SOTA on the PCKh@0.1 metric by 0.4 PCKh. + +Table 3. Comparison with state-of-the-art methods on COCO and MPII. Methods with ${}^{ \dagger }$ uses ${256} \times {256}$ input size. ${}^{\dagger \dagger }$ indicates additional fine-tuning on each dataset. + +
MethodInterHand2.6M3DPW
PCK↑AUC↑EPE↓AP↑AR↑
UniHCP98.587.03.7257.961.8
ViTPose++98.386.24.0281.785.2
Ours98.687.13.7083.687.1
+ +Table 4. Transfer results on hand (Interhand2.6M) and human shape (3DPW) domains. + +These results demonstrate that our method effectively balances adaptability to diverse domains while maintaining competitive performance on standard benchmarks, aligning with our primary objective of improving generalization beyond mainstream tasks. + +# 4.5. Transfer to hands and human shapes + +PoseBH has the capability to generalize its learned representations beyond the training dataset. We validate this through domain transfer experiments by fine-tuning the model separately on the InterHand2.6M and 3DPW datasets. InterHand2.6M is a two-hand pose estimation dataset containing 21 MANO keypoints [34], while 3DPW is an in-the-wild human shape estimation dataset with 24 SMPL keypoints [26]. + +During fine-tuning, we freeze the embedding module and nonparametrically learn the prototypes for the target datasets. After convergence, we fine-tune the backbone and head weights alongside the prototype loss. Freezing the embedding module ensures that the previously learned prototypes and those of the target dataset remain aligned within the same embedding space, preserving the learned representations during transfer. Since these experiments do not involve skeleton heterogeneity across datasets, no semi- + +![](images/53c6437367cbd68d4e582e4aabd6a24ea92f39629a1fea7044648f63e941409f.jpg) +Figure 4. Comparative pose estimation results on human (left) and animal (dog; right). + +![](images/3d7f0d136da98f5cbb22c318d7f45f00e14edbb52af7c57f0276cc828307751e.jpg) +Figure 5. Comparative results on InterHand2.6M (a-c) and 3DPW (e-f). + +supervised learning challenges arise, and CSS loss is not applied. + +In Tab. 4, we report results on the InterHand2.6M and 3DPW test sets using the ViT-B backbone. Given its large dataset size ( $\sim$ 2.6M instances), InterHand2.6M exhibits a smaller performance gap among different methods. Nonetheless, our method surpasses ViTPose++ by 0.3 PCK, 0.9 AUC, and 0.32 EPE, while also outperforming UniHCP by 0.1 PCK, 0.1 AUC, and 0.02 EPE. For the smaller 3DPW dataset, our method achieves a notable performance gain, surpassing the baseline by 1.9 AP and 1.9 AR, demonstrating strong generalization in keypoint embeddings. + +# 4.6. Ablation studies + +In Tab. 5, we present the results of ablation experiments on the COCO (CO), AIC (AI), MPII (MP), AP-10K (AP), APT-36K (APT), and COCO-Wholebody (CW) validation + +sets using the ViT-Base backbone. The top row represents a multi-head baseline with a shared backbone and dataset-specific heads. This conventional multi-head approach exhibits relatively low performance, particularly on the AP-10K and APT-36K validation sets, due to substantial domain discrepancies. Our keypoint prototype method achieves an average score of 66.3, improving the baseline by 0.8. Additionally, incorporating cross-type self-supervision provides an additional average performance gain of 0.2. Collectively, these enhancements enable our final configuration to outperform the multi-head baseline by an average of 2.4 points across all six datasets. + +# 4.7. Visual analysis + +Figure 4 presents examples of human and animal pose estimation, demonstrating that our algorithm achieves more accurate keypoint localization under complex interactions and + +![](images/351aed903ab13454ee08ad3ca52e8a221700c4e559c53379af1fe12d7b917069.jpg) +(a) +Figure 6. Examples of SMPL fitting using KITRO and with our 2D predictions. + +
MethodCOAIMPAPAPTCWAvg.
Baseline77.031.693.175.374.857.168.2
+LProto77.031.893.076.486.457.570.4
+LCSS77.332.193.276.786.557.970.6
+ +Table 5. Ablation study of the proposed method. $\mathcal{L}_{\mathrm{Proto}}$ : keypoint prototype learning. $\mathcal{L}_{\mathrm{CSS}}$ : cross-type self-supervision. + +occlusions, as seen in the right-hand keypoints of the standing person (left) and the right front toe of the dog (right). + +We further compare the pose estimation results on InterHand2.6M and 3DPW in Fig. 5. In (a), ViTPose++ incorrectly assigns a little finger joint to the opposite hand, whereas our method provides more accurate localization. In (b), ViTPose++ struggles with occlusion from the other hand, producing noisy predictions, while our method correctly separates them. In (c), due to severe self-occlusion, ViTPose++ misestimates occluded keypoints and assigns them to the opposite hand, while our method generalizes well in challenging self-occlusion scenarios. In (d), ViTPose++ produces an implausible foot and hand pose under heavy occlusion, while our method remains robust. In (e), ViTPose++ swaps the left and right legs, while our method correctly distinguishes them. In (f), background clutter causes ViTPose++ to produce noisy predictions for the left hand, while our method correctly identifies the lo + +cation of the occluded left hand. + +To further assess the effectiveness of our 2D SMPL keypoint predictions on the 3DPW test set, we employ a SMPL optimization technique. KITRO [55] optimizes SMPL parameters using 2D keypoint reprojection loss. For a comparative visual analysis, we replace KITRO's 2D keypoint predictions with ours. As shown in Figure 6(a) and (b), KITRO fails to align the mesh with the actual feet, whereas our method achieves more precise alignment. In (c), KITRO mislocalizes the right hand under partial occlusion, while our 2D predictions remain robust. In (d), KITRO exhibits misalignment in both the hands and feet, while replacing its 2D predictions with ours significantly improves the alignment of peripheral body parts. + +In the supplemental material, we provide a visualization of the prototypes generated by our algorithm, illustrating how they effectively capture the diversity of representations within the embedding space. + +# 5. Conclusion + +This paper presents a new multi-dataset training (MDT) framework for pose estimation across diverse domains and datasets. Traditional approaches encounter challenges with label heterogeneity, dataset-specific parameters, and limited domain generalization. Our approach addresses the limitations with two main contributions. First, to resolve label heterogeneity and dataset-specific parameters, we employ nonparametric prototypes to unify diverse keypoint annotations across multiple datasets. Second, to address the limited multi-dataset supervision, we propose a cross-type self-supervision mechanism. The proposed method has demonstrated substantial improvements over existing MDT across a broad range of datasets, including human pose, animal pose, hand shape, and human shape estimation. + +Limitations and future work. A limitation of our approach is that it requires at least a few-shot sample to learn prototypes, making zero-shot inference for unseen skeletons infeasible. Future work should explore extending our method for zero-shot human pose estimation. Future work should also investigate adapting the approach to 3D domains e.g. using axis-angle rotation representations, to facilitate representation learning on 3D manifolds. + +Acknowledgments. This work was supported by the National Research Foundation of Korea (NRF) grant (No. 2021R1A2C2012195) and by the Institute of Information & communications Technology Planning & Evaluation (IITP) grants No. RS-2019-II191906 (AI Graduate School Program, POSTECH), No. RS-2020-II201336 (AI Graduate School Program, UNIST), and RS-2022-II220290, funded by the Korea government (MSIT). + +# References + +[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2D human pose estimation: New benchmark and state of the art analysis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1, 5 +[2] Mykhaylo Andriluka, Umar Iqbal, Eldar Insafutdinov, Leonid Pishchulin, Anton Milan, Juergen Gall, and Bernt Schiele. PoseTrack: A benchmark for human pose estimation and tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5167-5176, 2018. 2 +[3] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3 +[4] Yanbei Chen, Manchen Wang, Abhay Mittal, Zhenlin Xu, Paolo Favaro, Joseph Tighe, and Davide Modolo. ScaleDet: A scalable multi-dataset object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7288-7297, 2023. 2 +[5] Hongsuk Choi, Gyeongsik Moon, JoonKyu Park, and Kyoung Mu Lee. Learning to estimate robust 3D human mesh from in-the-wild crowded scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[6] Yuanzheng Ci, Yizhou Wang, Meilin Chen, Shixiang Tang, Lei Bai, Feng Zhu, Rui Zhao, Fengwei Yu, Donglian Qi, and Wanli Ouyang. UniHCP: A unified model for human-centric perceptions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17840-17852, 2023. 1, 2 +[7] Mickael Cormier, Aris Clepe, Andreas Specker, and Jürgen Beyerer. Where are we with human pose estimation in real-world surveillance? In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 591-601, 2022. 1 +[8] Anurag Das, Yongqin Xian, Dengxin Dai, and Bernt Schiele. Weakly-supervised domain adaptive semantic segmentation with prototypical contrastive learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15434-15443, 2023. 3 +[9] Jon Donnelly, Alina Jade Barnett, and Chaofan Chen. Deformable protopnet: An interpretable image classifier using deformable prototypes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10265-10275, 2022. 3 +[10] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4320–4329, 2022. 3 +[11] Zigang Geng, Chunyu Wang, Yixuan Wei, Ze Liu, Houqiang Li, and Han Hu. Human pose as compositional tokens. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6 +[12] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human POSEitioning System (HPS): 3D human pose estimation and self-localization in large scenes from + +body-mounted sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4318-4329, 2021. 1 +[13] Or Hirschorn and Shai Avidan. A graph-based approach for category-agnostic pose estimation, 2024. 2 +[14] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. NeuMan: Neural human radiance field from a single video. In European Conference on Computer Vision (ECCV), 2022. 1 +[15] Roy Kapon, Guy Tevet, Daniel Cohen-Or, and Amit H Bermano. MAS: Multi-view ancestral sampling for 3D motion generation using 2D diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1965–1974, 2024. 1 +[16] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. arXiv preprint arXiv:2408.12569, 2024. 2 +[17] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2 +[18] Alberto Lamas, Siham Tabik, Antonio Cano Montes, Francisco Pérez-Hernández, Jorge García, Roberto Olmos, and Francisco Herrera. Human pose estimation for mitigating false negatives in weapon detection in video-surveillance. Neurocomputing, 489:488-503, 2022. 1 +[19] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. MSeg: A composite dataset for multidomain semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[20] Duy Tho Le, Chenhui Gou, Stavya Datta, Hengcan Shi, Ian Reid, Jianfei Cai, and Hamid Rezatofighi. JRDB-PanoTrack: An open-world panoptic segmentation and tracking robotic dataset in crowded human environments. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22325-22334, 2024. 1 +[21] Yanjie Li, Sen Yang, Peidong Liu, Shoukui Zhang, Yunxiao Wang, Zhicheng Wang, Wankou Yang, and Shu-Tao Xia. SimCC: A simple coordinate classification perspective for human pose estimation. In European Conference on Computer Vision (ECCV), pages 89-106. Springer, 2022. 6 +[22] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-X: A large-scale 3D expressive whole-body human motion dataset. Advances in Neural Information Processing Systems (NeurIPS), 2023. 1 +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 1, 5 +[24] Sun-Ao Liu, Yiheng Zhang, Zhaofan Qiu, Hongtao Xie, Yongdong Zhang, and Ting Yao. Learning orthogonal prototypes for generalized few-shot semantic segmentation. In + +IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[25] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. HumanGaussian: Text-driven 3D human generation with gaussian splatt-ting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6646-6657, 2024. 1 +[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 6 +[27] Yawen Lu, Dongfang Liu, Qifan Wang, Cheng Han, Yiming Cui, Zhiwen Cao, Xueling Zhang, Yingjie Victor Chen, and Heng Fan. ProMotion: Prototypes as motion learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28109-28119, 2024. 3 +[28] Gyeongsik Moon, Shouou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. InterHand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 5 +[29] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. NeuralAnnot: Neural annotator for 3D human mesh training sets. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2 +[30] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three Recipes for Better 3D PseudogTs of 3D Human Mesh Estimation in the Wild. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 2 +[31] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3D gaussian avatar. In European Conference on Computer Vision (ECCV), 2024. 1 +[32] Haocong Rao and Chunyan Miao. TranSG: Transformer-based skeleton graph prototype contrastive learning with structure-trajectory prompted reconstruction for person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22118-22128, 2023. 3 +[33] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D human motion model for robust pose estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1 +[34] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 6 +[35] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7420-7430, 2022. 3 +[36] István Sárándi, Alexander Hermans, and Bastian Leibe. Learning 3D human pose estimation from dozens of datasets using a geometry-aware autoencoder to bridge between skeleton formats. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023. 2 + +[37] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. 4 +[38] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3 +[39] Markus D Solbach and John K Tsotsos. Vision-based fallen person detection for the elderly. In IEEE/CVF international conference on computer vision workshops (ICCVW), pages 1433-1442, 2017. 1 +[40] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5, 6 +[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Black Michael J., and Tao Mei. Monocular, one-stage, regression of multiple 3D people. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3D people in depth. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[43] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xR-EgoPose: Egocentric 3D human pose from an hmd camera. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7728-7738, 2019. 1 +[44] Edward Vendrow, Duy Tho Le, Jianfei Cai, and Hamid Rezatofighi. JRDB-Pose: A large-scale dataset for multiperson pose estimation and tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1 +[45] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 5 +[46] Chong Wang, Yuyuan Liu, Yuanhong Chen, Fengbei Liu, Yu Tian, Davis McCarthy, Helen Frazer, and Gustavo Carneiro. Learning support and trivial prototypes for interpretable image classification. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2062-2072, 2023. 3 +[47] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3D human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13031-13040, 2023. 1 +[48] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards universal object detection by domain attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7289-7298, 2019. 2 +[49] Zhongdao Wang, Hengshuang Zhao, Ya-Li Li, Shengjin Wang, Philip Torr, and Luca Bertinetto. Do different tracking tasks require different appearance models? Advances in Neural Information Processing Systems (NeurIPS), 34:726-738, 2021. 2 +[50] Jiahong Wu, He Zheng, Bo Zhao, Yixin Li, Baoming Yan, Rui Liang, Wenjia Wang, Shipei Zhou, Guosen Lin, Yanwei Fu, et al. AI Challenger: A large-scale dataset for + +going deeper in image understanding. arXiv preprint arXiv:1711.06475, 2017. 5 +[51] Hang Xu, Linpu Fang, Xiaodan Liang, Wenxiong Kang, and Zhenguo Li. Universal-RCNN: Universal object detector via transferable graph r-cnn. In AAAI Conference on Artificial Intelligence (AAAI), pages 12492–12499, 2020. 2 +[52] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In European Conference on Computer Vision (ECCV), pages 398–416. Springer, 2022. 2 +[53] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose++: Vision transformer for generic body pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 5 +[54] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. UNIK: A unified framework for real-world skeleton-based action recognition. British Machine Visison Conference (BMVC), 2021. 2 +[55] Fengyuan Yang, Kerui Gu, and Angela Yao. KITRO: Refining human mesh by 2D clues and kinematic-tree rotation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1052-1061, 2024. 8 +[56] Yuxiang Yang, Junjie Yang, Yufei Xu, Jing Zhang, Long Lan, and Dacheng Tao. APT-36K: A large-scale benchmark for animal pose estimation and tracking. Advances in Neural Information Processing Systems (NeurIPS), 35:17301-17313, 2022. 1, 5 +[57] Sung Whan Yoon, Jun Seo, and Jaekyun Moon. TapNet: Neural network augmented with task-adaptive projection for few-shot learning. In International conference on machine learning, pages 7115-7123. PMLR, 2019. 3 +[58] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. AP-10K: A benchmark for animal pose estimation in the wild. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 1, 5 +[59] Xinyi Yu, Xin Zhang, Chengjun Xu, and Linlin Ou. Humanrobot collaborative interaction with human perception and action recognition. Neurocomputing, 563:126827, 2024. 1 +[60] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. HRFormer: High-resolution transformer for dense prediction. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 6 +[61] Baoquan Zhang, Xutao Li, Yunming Ye, Zhichao Huang, and Lisai Zhang. Prototype completion with primitive knowledge for few-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3754-3762, 2021. 3 +[62] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3D: A unified baseline for multi-dataset 3D object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9253-9262, 2023. 2 +[63] Xiangyun Zhao, Samuel Schulter, Gaurav Sharma, Yi-Hsuan Tsai, Manmohan Chandraker, and Ying Wu. Object detection with a unified label space from multiple datasets. In + +European Conference on Computer Vision (ECCV), pages 178-193. Springer, 2020. 2 +[64] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking Semantic Segmentation: A Prototype View. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2582-2593, 2022. 3, 4 +[65] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7571-7580, 2022. 2 \ No newline at end of file diff --git a/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/images.zip b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..29bd751e4b32e4e892460fdd490162a0accfdd9d --- /dev/null +++ b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:085c332d36f0ba6721d2d49d7e87e283312b108a221de70b3a33c0cc1e17fbe8 +size 698031 diff --git a/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/layout.json b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..48f889a7942f692287a1e88f648ae76a91488e9a --- /dev/null +++ b/2025/PoseBH_ Prototypical Multi-Dataset Training Beyond Human Pose Estimation/layout.json @@ -0,0 +1,9644 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "type": "text", + "content": "PoseBH: Prototypical Multi-Dataset Training Beyond Human Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "spans": [ + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "text", + "content": "Uyoung Jeong" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "text", + "content": " Jonathan Freer" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "text", + "content": " Seungryul Baek" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "text", + "content": " Hyung Jin Chang" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "text", + "content": " Kwang In Kim" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "text", + "content": "UNIST " + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "text", + "content": "University of Birmingham " + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 74, + 142, + 534, + 178 + ], + "type": "text", + "content": "POSTECH" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 204, + 200, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 204, + 200, + 217 + ], + "spans": [ + { + "bbox": [ + 151, + 204, + 200, + 217 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 229, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 529 + ], + "type": "text", + "content": "We study multi-dataset training (MDT) for pose estimation, where skeletal heterogeneity presents a unique challenge that existing methods have yet to address. In traditional domains, e.g. regression and classification, MDT typically relies on dataset merging or multi-head supervision. However, the diversity of skeleton types and limited cross-dataset supervision complicate integration in pose estimation. To address these challenges, we introduce PoseBH, a new MDT framework that tackles keypoint heterogeneity and limited supervision through two key techniques. First, we propose nonparametric keypoint prototypes that learn within a unified embedding space, enabling seamless integration across skeleton types. Second, we develop a cross-type self-supervision mechanism that aligns keypoint predictions with keypoint embedding prototypes, providing supervision without relying on teacher-student models or additional augmentations. PoseBH substantially improves generalization across whole-body and animal pose datasets, including COCO-WholeBody, AP-10K, and APT-36K, while preserving performance on standard human pose benchmarks (COCO, MPII, and AIC). Furthermore, our learned keypoint embeddings transfer effectively to hand shape estimation (InterHand2.6M) and human body shape estimation (3DPW). The code for PoseBH is available at: https://github.com/uyoung-jeong/PoseBH." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 550, + 135, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 135, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 135, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "content": "The demand for human pose estimators has grown substantially in recent years, driving their adoption across diverse applications, including 3D avatar generation [14, 25, 31], motion generation [15, 22, 33], human-robot interaction [20, 44, 59], safety monitoring [7, 18, 39], and VR/AR tracking [12, 43, 47]. These domains pose distinct operational challenges, often necessitating significant domain adaptation for robust pose estimation. To address this, we aim to train a human pose estimator that generalizes effectively across multiple datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "However, multi-dataset training (MDT) for pose estimation presents two key challenges. First, keypoints" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 203, + 375, + 273 + ], + "blocks": [ + { + "bbox": [ + 317, + 203, + 375, + 273 + ], + "lines": [ + { + "bbox": [ + 317, + 203, + 375, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 203, + 375, + 273 + ], + "type": "image", + "image_path": "43a33a2afd1a03df425af3a78dc42b1cae43b9c136b7c98400c354781c7235fa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 274, + 358, + 283 + ], + "lines": [ + { + "bbox": [ + 334, + 274, + 358, + 283 + ], + "spans": [ + { + "bbox": [ + 334, + 274, + 358, + 283 + ], + "type": "text", + "content": "COCO" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 376, + 203, + 434, + 273 + ], + "blocks": [ + { + "bbox": [ + 376, + 203, + 434, + 273 + ], + "lines": [ + { + "bbox": [ + 376, + 203, + 434, + 273 + ], + "spans": [ + { + "bbox": [ + 376, + 203, + 434, + 273 + ], + "type": "image", + "image_path": "19c165f3a24b569e93979b415d9428a0c35a93f90a4273c64793192e6c213cd4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 274, + 413, + 283 + ], + "lines": [ + { + "bbox": [ + 398, + 274, + 413, + 283 + ], + "spans": [ + { + "bbox": [ + 398, + 274, + 413, + 283 + ], + "type": "text", + "content": "AIC" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 434, + 203, + 492, + 273 + ], + "blocks": [ + { + "bbox": [ + 434, + 203, + 492, + 273 + ], + "lines": [ + { + "bbox": [ + 434, + 203, + 492, + 273 + ], + "spans": [ + { + "bbox": [ + 434, + 203, + 492, + 273 + ], + "type": "image", + "image_path": "f3242e21cae62c2ccf64299f92785151df94bdd3362da83fdac61e6e2eb32aba.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 274, + 473, + 283 + ], + "lines": [ + { + "bbox": [ + 455, + 274, + 473, + 283 + ], + "spans": [ + { + "bbox": [ + 455, + 274, + 473, + 283 + ], + "type": "text", + "content": "MPII" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 493, + 203, + 552, + 273 + ], + "blocks": [ + { + "bbox": [ + 493, + 203, + 552, + 273 + ], + "lines": [ + { + "bbox": [ + 493, + 203, + 552, + 273 + ], + "spans": [ + { + "bbox": [ + 493, + 203, + 552, + 273 + ], + "type": "image", + "image_path": "008abc2c4d83cb96d3474a78c51c19a47b89dea96caf28779ad26e1c297a5778.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 274, + 542, + 284 + ], + "lines": [ + { + "bbox": [ + 503, + 274, + 542, + 284 + ], + "spans": [ + { + "bbox": [ + 503, + 274, + 542, + 284 + ], + "type": "text", + "content": "WholeBody" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 285, + 375, + 354 + ], + "blocks": [ + { + "bbox": [ + 317, + 285, + 375, + 354 + ], + "lines": [ + { + "bbox": [ + 317, + 285, + 375, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 285, + 375, + 354 + ], + "type": "image", + "image_path": "b711e282a34ca825b6b61038b9d3a57189d962ad717232e160e8d377614d375d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 357, + 359, + 365 + ], + "lines": [ + { + "bbox": [ + 332, + 357, + 359, + 365 + ], + "spans": [ + { + "bbox": [ + 332, + 357, + 359, + 365 + ], + "type": "text", + "content": "AP-10K" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 376, + 285, + 433, + 354 + ], + "blocks": [ + { + "bbox": [ + 376, + 285, + 433, + 354 + ], + "lines": [ + { + "bbox": [ + 376, + 285, + 433, + 354 + ], + "spans": [ + { + "bbox": [ + 376, + 285, + 433, + 354 + ], + "type": "image", + "image_path": "8fa0aa361a62d037be1304f667b77e736bd9c3d47a2a882eee378c034a550a4a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 357, + 421, + 365 + ], + "lines": [ + { + "bbox": [ + 389, + 357, + 421, + 365 + ], + "spans": [ + { + "bbox": [ + 389, + 357, + 421, + 365 + ], + "type": "text", + "content": "APT-36K" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 434, + 285, + 493, + 354 + ], + "blocks": [ + { + "bbox": [ + 434, + 285, + 493, + 354 + ], + "lines": [ + { + "bbox": [ + 434, + 285, + 493, + 354 + ], + "spans": [ + { + "bbox": [ + 434, + 285, + 493, + 354 + ], + "type": "image", + "image_path": "9e13785908194b70429277d31d109d9a0e1f9beb14d1dca21d3e9a262edd6ff5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 357, + 475, + 365 + ], + "lines": [ + { + "bbox": [ + 452, + 357, + 475, + 365 + ], + "spans": [ + { + "bbox": [ + 452, + 357, + 475, + 365 + ], + "type": "text", + "content": "3DPW" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 494, + 285, + 552, + 354 + ], + "blocks": [ + { + "bbox": [ + 494, + 285, + 552, + 354 + ], + "lines": [ + { + "bbox": [ + 494, + 285, + 552, + 354 + ], + "spans": [ + { + "bbox": [ + 494, + 285, + 552, + 354 + ], + "type": "image", + "image_path": "d4f9a3fbc1c16f9aba82b1c44972e8feb9c32ce26e52c3bfee63f222f467cda4.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 497, + 357, + 548, + 365 + ], + "lines": [ + { + "bbox": [ + 497, + 357, + 548, + 365 + ], + "spans": [ + { + "bbox": [ + 497, + 357, + 548, + 365 + ], + "type": "text", + "content": "InterHand2.6M" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 376, + 555, + 419 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 419 + ], + "type": "text", + "content": "Figure 1. PoseBH unifies diverse skeleton formats, including humans, hands, and animals. The displayed skeletons show pose estimation results from our method, with 3DPW and InterHand2.6M predictions from a transferred model." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 445, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 566 + ], + "type": "text", + "content": "are heterogeneous across different datasets. For example, COCO [23] and MPII [1] define 17 and 16 keypoints, respectively, with 9 non-overlapping ones. Even when datasets share keypoints with identical names, their localization characteristics often differ due to domain gaps. For example, while COCO and animal datasets (e.g., APT-10K [58] and APT-36K [56]) use common keypoints names such as eyes, nose, and hips, significant anatomical differences exist (see Fig. 1). As a result, naively merging the identically named keypoints is ineffective." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 567, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 556, + 687 + ], + "type": "text", + "content": "Second, pose estimation MDT suffers from a lack of adequate supervision. In MDT, multiple skeleton types from different datasets are trained for each input instance. However, only a single skeleton-type ground truth is provided per input, leaving others unsupervised. This limitation resembles a semi-supervised learning challenge, constraining performance on smaller datasets. Existing semi-supervised learning methods typically rely on input augmentation or teacher-student knowledge distillation, both of which require substantial extra computational overhead." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Existing approaches fail to fully address these challenges, often relying on a shared backbone network [6] or" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12278" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "a mixture of experts [53] to preserve downstream performance. However, these methods tend to focus on mainstream datasets like COCO or MPII, resulting in degraded performance on underrepresented datasets. We propose a method that enhances performance in discrepant domains while maintaining accuracy on standard human pose datasets. A straightforward fine-tuning approach on the target domain often disrupts performance on source datasets, as it fails to balance differences in skeleton structures and dataset distributions. Therefore, effective generalization requires jointly harmonizing skeletal representations and dataset diversity during training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 294, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 346 + ], + "type": "text", + "content": "We achieve this through two primary techniques. First, we introduce prototypes to represent keypoints as embedding vectors, providing a unified representation across datasets. We regress pixel-wise keypoint embeddings from backbone features and match them to corresponding prototypes, automating keypoint mapping across datasets. This enables the integration of diverse keypoint types, including whole-body and animal keypoints, despite significant skeletal differences. Furthermore, our keypoint prototypes are non-learnable parameters, offering adaptability and computational efficiency for domain transfer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 348, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 294, + 491 + ], + "type": "text", + "content": "Second, we generate reliable self-supervision signals using the predicted keypoints and their embeddings, a process we term cross-type self-supervision. Since keypoint predictions can be obtained via a dot product between embeddings and prototypes, we effectively establish two distinct keypoint prediction modalities. By aligning these predictions, we filter out noisy keypoints and produce reliable labels for supervising unlabeled keypoints. This process enhances the utility of embeddings without requiring additional networks or input augmentation. As a result, we provide effective supervision for underrepresented and discrepant datasets, achieving a higher degree of generalizability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 491, + 262, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 491, + 262, + 502 + ], + "spans": [ + { + "bbox": [ + 67, + 491, + 262, + 502 + ], + "type": "text", + "content": "Our main technical contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 503, + 294, + 598 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 503, + 294, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 526 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 526 + ], + "type": "text", + "content": "- Keypoint prototypes for learning arbitrary keypoints from multiple datasets ensuring high transferability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 527, + 294, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 527, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 527, + 294, + 563 + ], + "type": "text", + "content": "- A cross-type self-supervision mechanism that aligns keypoint regression outputs with keypoint embeddings, enriching supervision for unlabeled keypoints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 563, + 294, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 563, + 294, + 598 + ], + "spans": [ + { + "bbox": [ + 56, + 563, + 294, + 598 + ], + "type": "text", + "content": "- Strong generalization across estimation tasks, including whole-body, animal, and hand pose estimation, as well as human shape estimation." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 609, + 144, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 144, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 144, + 621 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": "Multi-dataset training has been widely explored in various computer vision tasks. Relabeling and pseudo-labeling are among the most straightforward yet resource-intensive approaches. For example, MSeg [19] employed relabeling for segmentation, while [63] used offline pseudo-labeling for object detection. However, both methods face scalability challenges, as adding a new dataset requires reprocessing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "annotations for all previously processed datasets to redefine the class set, leading to a quadratic computational cost." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 96, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 228 + ], + "type": "text", + "content": "Several approaches have been proposed to improve scalability. In object detection, [4, 65] aim to unify object class labels by merging similar categories. However, they struggle to integrate localization features of the same object due to the absence of spatial information in class logits. UniTrack learns a unified identity embedding through self-supervised learning for multiple object-tracking tasks, including human pose tracking [49]. However, it prioritizes pose propagation over estimation and relies exclusively on the PoseTrack18 [2] skeleton, without incorporating multiple pose datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 228, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 336 + ], + "type": "text", + "content": "Another class of methods [48, 51, 54, 62] employs network modules, such as graph convolutional networks (GCN), for cross-dataset adaptation. However, these dataset-specific modules limit domain transferability, as their input and output sizes are dictated by the label structure, making them incompatible with datasets that have different keypoint formats. In contrast, our approach achieves flexible transferability without relying on dataset-specific domain adaptation modules." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 354, + 553, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 496 + ], + "type": "text", + "content": "Human pose estimation with MDT. In 3D human shape estimation, several methods have used pseudo-ground truths (GTs) generated by the SMPL model for MDT [5, 17, 41, 42]. However, these pseudo-GTs are often inaccurate due to depth ambiguity in 2D images and the limited expressiveness of the parametric 3D model [29, 30]. Similar to [65], Sarandi et al.'s affine-combining autoencoder (ACAE) merges 3D joint coordinates via an autoencoder [36]. However, the autoencoder's learnable parameters are dataset-dependent, limiting its generalizability. Furthermore, ACAE does not preserve pixel-level features, making it unsuitable when spatial information is crucial." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 616 + ], + "type": "text", + "content": "Recent works on category-agnostic pose estimation integrate various objects, such as chairs and vehicles, by combining multiple datasets [13, 52]. These methods follow a relabeling approach, creating a superset of keypoint classes and detecting only the keypoints defined in their relabeled dataset. As a result, they do not address keypoint heterogeneity or label sparsity issues, as training and evaluation occur within a single labeled dataset. In contrast, our method uses multiple skeletons to improve generalization across diverse downstream domains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "In 2D human pose estimation, UniHCP performs multiple human perception tasks using task-specific queries without incorporating task-specific layers inside the backbone [6]. This design limits its ability to generalize to distant domains such as animals and whole-body pose estimation. ViTPose++ [53] and Sapiens [16] employ multi-head architectures for MDT but do not address skeleton heterogeneity or the resulting label sparsity." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12279" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 382, + 252 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 382, + 252 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 382, + 252 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 382, + 252 + ], + "type": "image", + "image_path": "aa025e7daeaf31ca940cd2d70076f26054a16400641f4148c3a85d9ce74fa53d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 261, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 54, + 261, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 555, + 297 + ], + "type": "text", + "content": "Figure 2. Overview of the PoseBH architecture. During training, the embedding head maps the backbone features into a unified keypoint embedding space. By matching these embeddings with prototypes, dataset-specific keypoint heatmaps " + }, + { + "bbox": [ + 54, + 261, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{\\mathrm{coco}}^{\\mathrm{P}}" + }, + { + "bbox": [ + 54, + 261, + 555, + 297 + ], + "type": "text", + "content": " are generated. Prototypes are updated nonparametrically from the embeddings. During inference, the embedding head and the subsequent procedures are removed." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 384, + 71, + 554, + 253 + ], + "blocks": [ + { + "bbox": [ + 384, + 71, + 554, + 253 + ], + "lines": [ + { + "bbox": [ + 384, + 71, + 554, + 253 + ], + "spans": [ + { + "bbox": [ + 384, + 71, + 554, + 253 + ], + "type": "image", + "image_path": "54d4bb2b2991acf752d6e8ec4020400bdf2f0ce860d387059179f41c9b62041c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 316, + 294, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 294, + 387 + ], + "type": "text", + "content": "Prototype learning can be interpreted as proxy-based metric learning [35], where the centers of a fixed set of clusters serve as prototypes. This approach has been studied in various domains, including semantic segmentation [8, 10, 64], person re-identification [32], meta-learning [24, 38, 57, 61], and visual interpretability [3, 9, 46]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 388, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 556 + ], + "type": "text", + "content": "ProtoSeg [64] introduces a nonparametric approach to semantic segmentation by assigning multiple prototypes per class. ProMotion [27] employs prototypes to unify scene depth estimation and optical flow tasks. While our method draws inspiration from ProtoSeg and aligns with ProMotion in exploiting the benefits of prototypes for MDT, it differs in three key aspects. First, our approach applies keypoint-level supervision, which is inherently sparser than segmentation, depth, or optical flow. Second, we specifically address skeleton heterogeneity and cross-dataset relationships, aspects not considered by ProtoSeg and ProMotion. Third, while ProtoSeg and ProMotion apply fully supervised learning, our method tackles the challenge of unlabeled data, a problem unique to MDT pose estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 568, + 111, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 111, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 111, + 580 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "text", + "content": "We focus on multi-dataset training (MDT) for 2D human pose estimation. Given " + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "text", + "content": " datasets, our model predicts a keypoint heatmap for a person, represented as an array " + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_d \\in \\mathbb{R}^{J_d \\times H \\times W}" + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "text", + "content": " is the dataset index, " + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "inline_equation", + "content": "J_{d}" + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "text", + "content": " is the number of keypoints in the " + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "text", + "content": "-th dataset, and " + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 589, + 296, + 662 + ], + "type": "text", + "content": " denote the height and width of the heatmap, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 671, + 171, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 171, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 171, + 682 + ], + "type": "text", + "content": "3.1. Multi-head baseline" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "The multi-head baseline model consists of a backbone network that extracts an output backbone feature map from" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 316, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 411 + ], + "type": "text", + "content": "an input RGB image. The backbone network is shared across all datasets, while individual keypoint heads independently regress dataset-wise keypoint heatmaps " + }, + { + "bbox": [ + 313, + 316, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{k}_d\\}_{d=1}^D" + }, + { + "bbox": [ + 313, + 316, + 555, + 411 + ], + "type": "text", + "content": ", each trained with its corresponding labels. The keypoint heads share the same architecture, except for the final layer, where the output channel size is the respective number of keypoints. This generic baseline does not share the keypoint output space, limiting the MDT effect up to the backbone." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 418, + 470, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 470, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 470, + 430 + ], + "type": "text", + "content": "3.2. Keypoint prototype learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": "Overview. Our goal is to learn a unified keypoint space that standardizes keypoint output formats across datasets, scales with dataset size, and retains dataset-specific localization accuracy. To achieve this, we reformulate keypoint regression as a prototype-based distance metric learning problem (see Fig. 2). Our shared keypoint embedding module generates a normalized keypoint embedding map " + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{e} \\in \\mathbb{R}^{F \\times H \\times W}" + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": " from an input feature map. The module consists of two deconvolution layers, followed by a residual block and two convolution layers. Additional details are provided in the supplementary document." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "content": "Using embeddings and prototypes, we perform keypoint classification by matching embeddings with prototypes " + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{J \\times M \\times F}" + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "content": " is the embedding dimension, " + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "content": " is the total number of keypoints across all datasets, " + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 567, + 554, + 651 + ], + "type": "text", + "content": " is the number of in-class prototypes. For each heatmap pixel, we compute the cosine similarity between the prototypes and the embedding vector to obtain keypoint predictions:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 359, + 656, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 656, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 359, + 656, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\mathbf {k} _ {j, y, x} ^ {\\mathrm {P}} = \\max _ {m} \\frac {\\mathbf {P} _ {(j , m , :)} \\cdot \\mathbf {e} _ {(: , y , x)}}{\\left\\| \\mathbf {P} _ {(j , m , :)} \\right\\| \\left\\| \\mathbf {e} _ {(: , y , x)} \\right\\|}, \\tag {1}", + "image_path": "dac28de61b0a12836b38104f7222c4fc0d42a1ecb6dde4fc29c66d01d5fd6b50.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "x,y" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " denote pixel indices in the heatmap, and " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{(j,m,:)}" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " is a row vector of size " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": ", extracted " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\in \\mathbb{R}^{J\\times M\\times F}" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " along its" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12280" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "text", + "content": "first two dimensions. For each keypoint class " + }, + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_{(j,:;:)}^{\\mathrm{P}} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "text", + "content": " stores the matching score of the most probable prototype among " + }, + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "text", + "content": " prototypes. We refer to this as the prototype keypoint heatmap " + }, + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "inline_equation", + "content": "\\mathbf{k}^{\\mathrm{P}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "text", + "content": " to distinguish it from the multihead outputs " + }, + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 55, + 72, + 294, + 134 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": "For the remainder of this paper, we use the following notation: " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{(:,a)}" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": " denotes the 2D slice of a 3D array " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": " indexed at " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": " along the third dimension, while " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{(:,b,:)}" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": " represents the 2D slice at index " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": " along the second dimension. Similarly, " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{(d:e,:;)}" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": " represents the 3D sub-array spanning indices " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": " along the first dimension. For a 2D array " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{(:,c)}" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": " indicates the " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": "-th column vector, whereas " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{(d,c)}" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": " refers to the element at position " + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "inline_equation", + "content": "(d,c)" + }, + { + "bbox": [ + 55, + 134, + 296, + 231 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "text", + "content": "Prototype learning. We adopt a nonparametric learning approach [64] to train " + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "text", + "content": ". First, we obtain logits " + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{l}_j\\}_{j=1}^J \\subset \\mathbb{R}^{M \\times N}" + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "text", + "content": " and targets " + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{t}_j \\in \\mathbb{R}^{N_j}\\}_{j=1}^J" + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "inline_equation", + "content": "N_j" + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "text", + "content": " is the number of foreground samples (i.e., pixels in the ground-truth (GT) heatmap with nonzero values) for the " + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "text", + "content": "-th joint, and " + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "text", + "content": " is the total number of foreground samples across all joints. We define logits as sampled embedding vectors and targets as keypoint class labels, with each keypoint class selecting an in-class prototype from " + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 243, + 296, + 353 + ], + "type": "text", + "content": " candidates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "text", + "content": "Since no GT selection is explicitly provided, we assign an in-class prototype for each logit via online clustering using Sinkhorn-Knopp iteration with an equipartition constraint [37], and compute " + }, + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_j" + }, + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 406, + 294, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 406, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 89, + 406, + 294, + 419 + ], + "type": "interline_equation", + "content": "\\mathbf {l} _ {j} = \\mathbf {P} _ {(j,:;:)} \\bar {\\mathbf {e}}, \\tag {2}", + "image_path": "88dab09dd80954bdc0d85827bc91a6ba73457287de48b72ffa1c78e5f3852745.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 421, + 294, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 421, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 88, + 421, + 294, + 449 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {j} = \\underset {m} {\\arg \\max } \\operatorname {d i a g} (\\mathbf {u}) \\exp \\left(\\frac {\\mathbf {l} _ {j}}{\\kappa}\\right) \\operatorname {d i a g} (\\mathbf {v}), \\tag {3}", + "image_path": "76cf82d9e31eb75795a7399f8aadb3cb7694c9d051060761fe5737fac46bd70e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{e}}" + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "text", + "content": " is a 2D array of size " + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "inline_equation", + "content": "F\\times N_{j}" + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "text", + "content": ", obtained by first flattening " + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{e}\\in \\mathbb{R}^{F\\times H\\times W}" + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "text", + "content": " along the spatial dimensions into a 2D array of size " + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "inline_equation", + "content": "F\\times (HW)" + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "text", + "content": " and then, selecting " + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "inline_equation", + "content": "N_{j}" + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "text", + "content": " foreground objects. " + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(\\mathbf{u})" + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "text", + "content": " is a diagonal matrix derived from the vector " + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\exp (\\cdot)" + }, + { + "bbox": [ + 55, + 451, + 296, + 511 + ], + "type": "text", + "content": " is applied element-wise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 511, + 274, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 511, + 274, + 523 + ], + "spans": [ + { + "bbox": [ + 67, + 511, + 274, + 523 + ], + "type": "text", + "content": "Each prototype is then updated with momentum " + }, + { + "bbox": [ + 67, + 511, + 274, + 523 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 67, + 511, + 274, + 523 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 528, + 294, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 528, + 294, + 561 + ], + "spans": [ + { + "bbox": [ + 83, + 528, + 294, + 561 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {j, m} ^ {\\text {n e w}} = \\lambda \\mathbf {P} _ {j, m} ^ {\\text {o l d}} + (1 - \\lambda) \\frac {1}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} w _ {m n} \\overline {{\\mathbf {e}}} _ {(;, n)}, \\tag {4}", + "image_path": "3caa767ac20faa2611abe19b90b3e5ca02acedf06a4a3b2008d06f07b94c482b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 565, + 294, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 294, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 565, + 294, + 589 + ], + "type": "inline_equation", + "content": "w_{mn}" + }, + { + "bbox": [ + 55, + 565, + 294, + 589 + ], + "type": "text", + "content": " is the clustering assignment weight, calculated similarly to Eq. (3), but using max instead of arg max." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 590, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 637 + ], + "type": "text", + "content": "To optimize the learned embedding with respect to the prototype " + }, + { + "bbox": [ + 55, + 590, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 590, + 296, + 637 + ], + "type": "text", + "content": ", we employ pixel-prototype contrastive learning, using the pixel-prototype contrastive loss (PPC) and pixel-prototype distance (PPD) from [64]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 96, + 642, + 294, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 642, + 294, + 675 + ], + "spans": [ + { + "bbox": [ + 96, + 642, + 294, + 675 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P P C}} \\left(\\mathbf {l} _ {j}, \\mathbf {t} _ {j}\\right) = \\frac {c _ {n}}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} \\operatorname {C E} \\left(\\mathbf {l} _ {j, n}, \\mathbf {t} _ {j, n}\\right), \\tag {5}", + "image_path": "29284d3c74f1c4a05afc25e951916e0c4c6aa0a85b67ae532b51ccb612ca46bb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 96, + 677, + 294, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 677, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 96, + 677, + 294, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {P P D}} \\left(\\mathbf {l} _ {j}, \\mathbf {t} _ {j}\\right) = \\frac {c _ {n}}{N _ {j}} \\sum_ {n = 1} ^ {N _ {j}} \\left(1 - \\mathbf {l} _ {j, n} ^ {m +}\\right) ^ {2}, \\tag {6}", + "image_path": "7c61ac59837b0bede555951ceac30f035bf7969eed3172c204cd933d6a6dcbcc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 69, + 555, + 155 + ], + "blocks": [ + { + "bbox": [ + 318, + 69, + 555, + 155 + ], + "lines": [ + { + "bbox": [ + 318, + 69, + 555, + 155 + ], + "spans": [ + { + "bbox": [ + 318, + 69, + 555, + 155 + ], + "type": "image", + "image_path": "4bda5829f2ed1d2ff71d0b311ee56f613716dce5b3382e919783c88b3fc0c80c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 163, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 196 + ], + "type": "text", + "content": "Figure 3. An illustrative example of cross-type self-supervision. Given a COCO image, we jointly refine the AIC head and AIC prototype predictions to produce reliable AIC heatmaps." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "inline_equation", + "content": "c_{n}" + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "content": " is the confidence value from the GT heatmap, CE denotes the cross-entropy loss, and " + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "inline_equation", + "content": "m^{+}" + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "content": " indexes the first dimension of " + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_d" + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "content": ". For each dataset " + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "content": ", we collect logits " + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_{j,n} \\in \\mathbb{R}^{M \\times N_d}" + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "content": " and targets " + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_d \\in \\mathbb{R}^{N_d}" + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "inline_equation", + "content": "N_{d}" + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "content": " is the number of samples in dataset " + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "content": ". The total prototype loss is then defined as " + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Proto}} = \\mathcal{L}_{\\mathrm{PPC}} + \\mathcal{L}_{\\mathrm{PPD}}" + }, + { + "bbox": [ + 313, + 216, + 553, + 287 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "text", + "content": "Unlike standard contrastive learning, " + }, + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Proto}}" + }, + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "text", + "content": " operates within a single dataset and lacks the push term for differing dataset skeletons. To introduce cross-dataset negatives, we apply K-means clustering to all prototypes, forming " + }, + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "text", + "content": " (=96) clusters. This clustering is performed only once during training, while online clustering for in-class prototypes is conducted at each iteration. These clusters serve as negative samples, where the embeddings closest to each centroid are used as logits, with their dataset-specific labels as targets. Finally, we obtain " + }, + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_c" + }, + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "text", + "content": " for each cluster and apply Eqs. (5) and (6) to enable cross-dataset contrastive learning." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 426, + 463, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 463, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 463, + 438 + ], + "type": "text", + "content": "3.3. Cross-type self-supervision" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 443, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 554, + 539 + ], + "type": "text", + "content": "While our keypoint prototypes enable unified MDT, label sparsity remains a challenge, as each image is annotated for only one dataset, limiting supervision. To address this semi-supervised learning problem, we introduce a self-supervision strategy that bridges keypoint heatmaps and embeddings. By collaboratively filtering out noisy predictions from both the keypoint and embedding heads, we generate reliable keypoint heatmaps for training (see Fig. 3)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 539, + 554, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 554, + 623 + ], + "type": "text", + "content": "To eliminate uncertain keypoint predictions, we apply two filtering conditions. First, for each keypoint class, the confidence scores of both the keypoint and embedding heads must exceed a threshold " + }, + { + "bbox": [ + 313, + 539, + 554, + 623 + ], + "type": "inline_equation", + "content": "(c_{thr} = 0.25)" + }, + { + "bbox": [ + 313, + 539, + 554, + 623 + ], + "type": "text", + "content": ". Second, the root mean square distance between their predictions must be below a threshold " + }, + { + "bbox": [ + 313, + 539, + 554, + 623 + ], + "type": "inline_equation", + "content": "(d_{thr} = 2.1)" + }, + { + "bbox": [ + 313, + 539, + 554, + 623 + ], + "type": "text", + "content": ". The filtered predictions are then combined using a weighted average:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 375, + 629, + 553, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 629, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 375, + 629, + 553, + 644 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathbf {y}} _ {i} = s _ {i} \\widehat {\\mathbf {y}} _ {i} ^ {\\mathrm {k p t}} + (1 - s _ {i}) \\widehat {\\mathbf {y}} _ {i} ^ {\\mathrm {e m b}}, \\tag {7}", + "image_path": "e53089ad62146eed1a5e35abfa97b409087b214676472a71782b91e673679883.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 651, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 553, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 651, + 553, + 689 + ], + "type": "inline_equation", + "content": "s_i = c_i^{\\mathrm{kpt}} / (c_i^{\\mathrm{kpt}} + c_i^{\\mathrm{emb}})" + }, + { + "bbox": [ + 313, + 651, + 553, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 651, + 553, + 689 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 651, + 553, + 689 + ], + "type": "text", + "content": " is the keypoint index, and " + }, + { + "bbox": [ + 313, + 651, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{y}}" + }, + { + "bbox": [ + 313, + 651, + 553, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 651, + 553, + 689 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 651, + 553, + 689 + ], + "type": "text", + "content": " represent the keypoint prediction and confidence score from each head, respectively." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "From " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{y}}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": ", we generate a reliable heatmap " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{k}^{\\mathrm{CSS}}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " following the standard GT heatmap generation process. The loss for" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12281" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 280, + 127 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 280, + 127 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 280, + 127 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 280, + 127 + ], + "type": "table", + "html": "
MethodGFLOPSCOCOAICMPIIAvg.
UniHCP18.976.832.690.966.8
ViTPose++18.577.031.693.167.2
Ours18.577.332.193.267.5
", + "image_path": "5e762ad2a6b22a5999a7002facde101e1c9835ecbcf29bedef7cec39d9288ae7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 178, + 294, + 245 + ], + "blocks": [ + { + "bbox": [ + 55, + 134, + 295, + 169 + ], + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 169 + ], + "type": "text", + "content": "Table 1. Comparison of different multi-dataset training methods on general human pose benchmarks, with computational complexity measured in GFLOPS during the evaluation phase." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 178, + 294, + 245 + ], + "lines": [ + { + "bbox": [ + 56, + 178, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 294, + 245 + ], + "type": "table", + "html": "
MethodAP-10KAPT-36KCOCO-WAvg.
UniHCP56.562.020.146.2
ViTPose++74.176.057.169.1
Ours75.087.257.973.4
", + "image_path": "482ad8f271afcae5b4af537b4e966614a6a877c6782a046750ea3070f7d5551c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 253, + 295, + 286 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 286 + ], + "type": "text", + "content": "Table 2. Comparison of multi-dataset training methods on whole-body and animal domains. COCO-W refers to COCO-WholeBody. The reported metrics are AP(%)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 307, + 216, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 216, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 216, + 319 + ], + "type": "text", + "content": "unlabeled samples is then computed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 327, + 294, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 327, + 294, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 294, + 379 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C S S}} = \\sum_ {d = 1} ^ {D} \\zeta \\left[ \\mathcal {L} _ {\\mathrm {h m}} \\left(\\mathbf {k} _ {d [ \\mathbf {u} ]}, \\mathbf {k} _ {d [ \\mathbf {u} ]} ^ {\\mathrm {C S S}}\\right) \\right. \\\\ + \\mathcal {L} _ {\\text {P r o t o}} \\left(\\mathbf {e} _ {[ \\mathbf {u} ]}, \\mathbf {k} _ {d [ \\mathbf {u} ]}, \\mathbf {k} _ {d [ \\mathbf {u} ]} ^ {\\mathrm {C S S}}\\right) ], \\tag {8} \\\\ \\end{array}", + "image_path": "fc246edad6cd1ed52c37718232a7e421f484a6ead06e3612f0b053d9c0773867.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 385, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 295, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 385, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 385, + 295, + 445 + ], + "type": "text", + "content": " denotes the indices of the unlabeled samples. This cross-type self-supervision (CSS) loss enables self-distillation between the dataset-wise keypoint head and the prototype, mitigating supervision shortage without requiring a teacher model or input duplication." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 445, + 212, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 445, + 212, + 456 + ], + "spans": [ + { + "bbox": [ + 67, + 445, + 212, + 456 + ], + "type": "text", + "content": "The final training loss is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 466, + 295, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 466, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 67, + 466, + 295, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M D T}} = \\mathcal {L} _ {\\mathrm {K P L}} + \\mathcal {L} _ {\\mathrm {C S S}}, \\text {w h e r e} \\tag {9}", + "image_path": "f455be9b83799e78c6062ce4af43c0a6b78c49cfcd2a5b5d5b49b001c2f12f78.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 482, + 295, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 482, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 70, + 482, + 295, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K P L}} = \\sum_ {d = 1} ^ {D} \\left[ \\mathcal {L} _ {\\mathrm {h m}} \\left(\\mathbf {k} _ {d}, \\mathbf {k} _ {d} ^ {\\mathrm {g t}}\\right) + \\mathcal {L} _ {\\mathrm {P r o t o}} \\left(\\mathbf {e}, \\mathbf {k} _ {d}, \\mathbf {k} _ {d} ^ {\\mathrm {g t}}\\right) \\right], \\tag {10}", + "image_path": "ff5d693faed1a5e7b00e683b6f4a27b242e67a0c90645730af62d92364a3c7d2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 522, + 234, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 234, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 234, + 534 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 55, + 522, + 234, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{hm}}" + }, + { + "bbox": [ + 55, + 522, + 234, + 534 + ], + "type": "text", + "content": " is the standard JointMSE loss [40]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 544, + 137, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 137, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 137, + 558 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 564, + 170, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 170, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 170, + 578 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "We follow the experimental setup of ViTPose++ [53], using six human pose datasets for training. (1) COCO consists of 57,000 training and 5,000 validation images, annotated with 17 keypoints [23]. (2) AIChallenger (AIC) contains 210,000 training and 30,000 validation images, annotated with 14 keypoints [50]. (3) MPII includes approximately 15,000 training, 2,729 validation, and 5,700 test images, with 16 keypoints [1]. (4) AP-10K is an animal pose dataset with approximately 7,000 training, 1,000 validation, and 2,000 test images [58]. (5) APT-36K, a video-based animal pose dataset, consists of about 25,000 training, 3,600 validation," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "and 7,000 test images [56]. Both AP-10K and APT-36K share the same skeletal format with 17 keypoints, largely following COCO. (6) COCO-WholeBody extends COCO with whole-body annotations covering 133 keypoints [50]. We follow standard evaluation protocols, using the average precision (AP) metric based on object keypoint similarity (OKS) for COCO, AIC, COCO-WholeBody, AP-10K, and APT-36K. For MPII, the percentage of correct keypoints (PCK) metric is used [1]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 180, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 554, + 323 + ], + "type": "text", + "content": "For the domain transfer experiments, we use InterHand2.6M [28] and 3DPW [45]. InterHand2.6M is a two-hand pose estimation dataset containing approximately 2.6M frames, annotated with 21 keypoints. We adopt the ViTPose++ configuration for frame sampling and dataset splits (training, validation, and test). For evaluation, we use PCK, AUC, and EPE, consistent with ViTPose++. 3DPW is an in-the-wild 3D human shape estimation dataset with 18,000 training and 26,000 test frames, annotated with 24 keypoints. We use COCO-style AP and AR as evaluation metrics. In both InterHand2.6M and 3DPW, the projected 2D keypoints are used as ground-truths." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "type": "text", + "content": "Further details on the experimental setups, the impact of varying hyperparameters, and failure case analysis can be found in the supplemental document." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 369, + 444, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 444, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 444, + 382 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 387, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 494 + ], + "type": "text", + "content": "Our multi-head baseline model is ViTPose++ [53]. We use an input image size of " + }, + { + "bbox": [ + 313, + 387, + 554, + 494 + ], + "type": "inline_equation", + "content": "256 \\times 192" + }, + { + "bbox": [ + 313, + 387, + 554, + 494 + ], + "type": "text", + "content": " and apply a flip test in all cases. The model is trained for 100 epochs using the AdamW optimizer with a weight decay factor of 0.1 and a step learning rate scheduler with an initial learning rate of 0.001. The learning rate is reduced by a factor of 0.1 at the 50th and 90th epochs. We use four NVIDIA A100 GPUs for training with the ViT-H backbone and four NVIDIA A6000 GPUs for training with the ViT-B backbone." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 495, + 554, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 615 + ], + "type": "text", + "content": "We employ a scheduling strategy for both the losses and the learning rate. Initially, we freeze the multi-head model, updating only the embedding module and prototypes for 50 epochs. In the subsequent 40 epochs, we train the multi-heads and the embedding module while keeping the prototypes and the backbone frozen. Finally, during the last 10 epochs, we train the entire network except the prototypes. At this stage, we introduce cross-type self-supervision loss. This progressive optimization approach accelerates training compared to training the entire network from the outset." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 624, + 528, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 528, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 528, + 636 + ], + "type": "text", + "content": "4.3. Comparison with existing MDT methods" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "In Tab. 1, we evaluate the competitiveness of our method against existing MDT approaches. For a fair comparison, all methods use the ViT-Base backbone and are trained under our dataset configuration. While UniHCP performs well on AIC, it achieves the lowest scores on COCO and MPII, resulting in an average of 66.8. We hypothe" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12282" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 541, + 237 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 541, + 237 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 541, + 237 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 541, + 237 + ], + "type": "table", + "html": "
MethodBackboneDatasetsCOCOMPII
(AP↑)(AP50↑)(AP75↑)(AR↑)(PCKh↑)(PCKh@0.1↑)
HRNetHRNet-W48175.190.682.280.490.3†33.1†
HRFormerHRFormer-B175.690.882.880.8--
SimCCHRNet-W48176.190.682.981.290.036.8
PCT†Swin-B177.791.284.782.192.5-
PCT†Swin-H179.391.585.9---
UniHCPViT-B3376.1---93.2††-
ViTPose++-BViT-B677.073.484.082.692.839.1
ViTPose++-HViT-H679.491.985.784.894.241.6
OursViT-B677.390.884.282.493.239.3
OursViT-H679.591.985.884.594.242.0
", + "image_path": "74adebbf2682bec539674f578c1f8e5b09cac961cfe206adb259bad9a3813f39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 289, + 294, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 294, + 396 + ], + "type": "text", + "content": "size that UniHCP's parameter-sharing strategy causes well-represented pose datasets to dominate, hindering its generalization across different skeleton structures. ViTPose++ serves as both a multi-head model and the baseline for our method. Our algorithm surpasses ViTPose++ by 0.3 in average score and outperforms it on individual datasets. Given that these benchmark scores are highly saturated and our method maintains the same computation complexity as ViTPose++ during inference, this improvement is significant." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 398, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 398, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 54, + 398, + 295, + 554 + ], + "type": "text", + "content": "To further assess generalization, we test our method on animal and whole-body pose estimation datasets. Table 2 shows that our method consistently improves performance, despite significant domain discrepancies. UniHCP lags behind with an average score of 27.2, likely due to its parameter sharing strategy, which does not address skeleton heterogeneity. Our method also surpasses ViTPose++ by an average score of 4.3, demonstrating superior generalization beyond human skeletons. Specifically, it achieves gains of 0.9 AP on AP-10K, 11.2 AP on APT-36K, and 0.8 AP on COCO-WholeBody. Unlike existing methods that compromise either mainstream performance or adaptation to distant domains, our approach enhances both synergistically." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 563, + 264, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 264, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 264, + 576 + ], + "type": "text", + "content": "4.4. Comparison to state-of-the-art methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": "In Tab. 3, we compare our method with state-of-the-art (SOTA) approaches on the COCO and MPII benchmarks, using scores reported in the respective papers. We also evaluate our method against SOTA approaches designed for single datasets, including HRNet [40], HRFformer [60], SimCC [21], and PCT [11]. On the COCO validation set, our method ranks second with the ViT-B backbone and achieves SOTA performance with the ViT-H backbone, reaching 79.5 AP. On the MPII validation set, our ViT-B model performs comparably to the fine-tuned UniHCP and surpasses the SOTA on the PCKh@0.1 metric by 0.4 PCKh." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 287, + 545, + 365 + ], + "blocks": [ + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "type": "text", + "content": "Table 3. Comparison with state-of-the-art methods on COCO and MPII. Methods with " + }, + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "type": "text", + "content": " uses " + }, + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "type": "text", + "content": " input size. " + }, + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "type": "inline_equation", + "content": "{}^{\\dagger \\dagger }" + }, + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "type": "text", + "content": " indicates additional fine-tuning on each dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 287, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 315, + 287, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 315, + 287, + 545, + 365 + ], + "type": "table", + "html": "
MethodInterHand2.6M3DPW
PCK↑AUC↑EPE↓AP↑AR↑
UniHCP98.587.03.7257.961.8
ViTPose++98.386.24.0281.785.2
Ours98.687.13.7083.687.1
", + "image_path": "ad3ad1aeb996c57073b2cdcb7dbd4dcae69cfd84cabfd2f7928c5038058311bd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 373, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 396 + ], + "type": "text", + "content": "Table 4. Transfer results on hand (Interhand2.6M) and human shape (3DPW) domains." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 419, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 479 + ], + "type": "text", + "content": "These results demonstrate that our method effectively balances adaptability to diverse domains while maintaining competitive performance on standard benchmarks, aligning with our primary objective of improving generalization beyond mainstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 490, + 511, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 511, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 511, + 503 + ], + "type": "text", + "content": "4.5. Transfer to hands and human shapes" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": "PoseBH has the capability to generalize its learned representations beyond the training dataset. We validate this through domain transfer experiments by fine-tuning the model separately on the InterHand2.6M and 3DPW datasets. InterHand2.6M is a two-hand pose estimation dataset containing 21 MANO keypoints [34], while 3DPW is an in-the-wild human shape estimation dataset with 24 SMPL keypoints [26]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "During fine-tuning, we freeze the embedding module and nonparametrically learn the prototypes for the target datasets. After convergence, we fine-tune the backbone and head weights alongside the prototype loss. Freezing the embedding module ensures that the previously learned prototypes and those of the target dataset remain aligned within the same embedding space, preserving the learned representations during transfer. Since these experiments do not involve skeleton heterogeneity across datasets, no semi-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12283" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 68, + 541, + 256 + ], + "blocks": [ + { + "bbox": [ + 67, + 68, + 541, + 256 + ], + "lines": [ + { + "bbox": [ + 67, + 68, + 541, + 256 + ], + "spans": [ + { + "bbox": [ + 67, + 68, + 541, + 256 + ], + "type": "image", + "image_path": "53c6437367cbd68d4e582e4aabd6a24ea92f39629a1fea7044648f63e941409f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 258, + 461, + 269 + ], + "lines": [ + { + "bbox": [ + 147, + 258, + 461, + 269 + ], + "spans": [ + { + "bbox": [ + 147, + 258, + 461, + 269 + ], + "type": "text", + "content": "Figure 4. Comparative pose estimation results on human (left) and animal (dog; right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 276, + 541, + 473 + ], + "blocks": [ + { + "bbox": [ + 69, + 276, + 541, + 473 + ], + "lines": [ + { + "bbox": [ + 69, + 276, + 541, + 473 + ], + "spans": [ + { + "bbox": [ + 69, + 276, + 541, + 473 + ], + "type": "image", + "image_path": "3d7f0d136da98f5cbb22c318d7f45f00e14edbb52af7c57f0276cc828307751e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 477, + 436, + 488 + ], + "lines": [ + { + "bbox": [ + 173, + 477, + 436, + 488 + ], + "spans": [ + { + "bbox": [ + 173, + 477, + 436, + 488 + ], + "type": "text", + "content": "Figure 5. Comparative results on InterHand2.6M (a-c) and 3DPW (e-f)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "content": "supervised learning challenges arise, and CSS loss is not applied." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 654 + ], + "type": "text", + "content": "In Tab. 4, we report results on the InterHand2.6M and 3DPW test sets using the ViT-B backbone. Given its large dataset size (" + }, + { + "bbox": [ + 55, + 534, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 55, + 534, + 295, + 654 + ], + "type": "text", + "content": "2.6M instances), InterHand2.6M exhibits a smaller performance gap among different methods. Nonetheless, our method surpasses ViTPose++ by 0.3 PCK, 0.9 AUC, and 0.32 EPE, while also outperforming UniHCP by 0.1 PCK, 0.1 AUC, and 0.02 EPE. For the smaller 3DPW dataset, our method achieves a notable performance gain, surpassing the baseline by 1.9 AP and 1.9 AR, demonstrating strong generalization in keypoint embeddings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 660, + 154, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 154, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 154, + 672 + ], + "type": "text", + "content": "4.6. Ablation studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In Tab. 5, we present the results of ablation experiments on the COCO (CO), AIC (AI), MPII (MP), AP-10K (AP), APT-36K (APT), and COCO-Wholebody (CW) validation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 510, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 654 + ], + "type": "text", + "content": "sets using the ViT-Base backbone. The top row represents a multi-head baseline with a shared backbone and dataset-specific heads. This conventional multi-head approach exhibits relatively low performance, particularly on the AP-10K and APT-36K validation sets, due to substantial domain discrepancies. Our keypoint prototype method achieves an average score of 66.3, improving the baseline by 0.8. Additionally, incorporating cross-type self-supervision provides an additional average performance gain of 0.2. Collectively, these enhancements enable our final configuration to outperform the multi-head baseline by an average of 2.4 points across all six datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 660, + 406, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 406, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 406, + 673 + ], + "type": "text", + "content": "4.7. Visual analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Figure 4 presents examples of human and animal pose estimation, demonstrating that our algorithm achieves more accurate keypoint localization under complex interactions and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12284" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 79, + 293, + 330 + ], + "blocks": [ + { + "bbox": [ + 107, + 71, + 116, + 79 + ], + "lines": [ + { + "bbox": [ + 107, + 71, + 116, + 79 + ], + "spans": [ + { + "bbox": [ + 107, + 71, + 116, + 79 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 79, + 293, + 330 + ], + "lines": [ + { + "bbox": [ + 59, + 79, + 293, + 330 + ], + "spans": [ + { + "bbox": [ + 59, + 79, + 293, + 330 + ], + "type": "image", + "image_path": "351aed903ab13454ee08ad3ca52e8a221700c4e559c53379af1fe12d7b917069.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 338, + 295, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 361 + ], + "type": "text", + "content": "Figure 6. Examples of SMPL fitting using KITRO and with our 2D predictions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 72, + 375, + 279, + 441 + ], + "blocks": [ + { + "bbox": [ + 72, + 375, + 279, + 441 + ], + "lines": [ + { + "bbox": [ + 72, + 375, + 279, + 441 + ], + "spans": [ + { + "bbox": [ + 72, + 375, + 279, + 441 + ], + "type": "table", + "html": "
MethodCOAIMPAPAPTCWAvg.
Baseline77.031.693.175.374.857.168.2
+LProto77.031.893.076.486.457.570.4
+LCSS77.332.193.276.786.557.970.6
", + "image_path": "d222dc3e92cc7540ce20f6372649d98441204c481c739219c07d18dd7c318161.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 295, + 472 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 472 + ], + "type": "text", + "content": "Table 5. Ablation study of the proposed method. " + }, + { + "bbox": [ + 55, + 449, + 295, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Proto}}" + }, + { + "bbox": [ + 55, + 449, + 295, + 472 + ], + "type": "text", + "content": ": keypoint prototype learning. " + }, + { + "bbox": [ + 55, + 449, + 295, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CSS}}" + }, + { + "bbox": [ + 55, + 449, + 295, + 472 + ], + "type": "text", + "content": ": cross-type self-supervision." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 496, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 521 + ], + "type": "text", + "content": "occlusions, as seen in the right-hand keypoints of the standing person (left) and the right front toe of the dog (right)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 295, + 713 + ], + "type": "text", + "content": "We further compare the pose estimation results on InterHand2.6M and 3DPW in Fig. 5. In (a), ViTPose++ incorrectly assigns a little finger joint to the opposite hand, whereas our method provides more accurate localization. In (b), ViTPose++ struggles with occlusion from the other hand, producing noisy predictions, while our method correctly separates them. In (c), due to severe self-occlusion, ViTPose++ misestimates occluded keypoints and assigns them to the opposite hand, while our method generalizes well in challenging self-occlusion scenarios. In (d), ViTPose++ produces an implausible foot and hand pose under heavy occlusion, while our method remains robust. In (e), ViTPose++ swaps the left and right legs, while our method correctly distinguishes them. In (f), background clutter causes ViTPose++ to produce noisy predictions for the left hand, while our method correctly identifies the lo" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 446, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 446, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 446, + 83 + ], + "type": "text", + "content": "cation of the occluded left hand." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 87, + 555, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 243 + ], + "type": "text", + "content": "To further assess the effectiveness of our 2D SMPL keypoint predictions on the 3DPW test set, we employ a SMPL optimization technique. KITRO [55] optimizes SMPL parameters using 2D keypoint reprojection loss. For a comparative visual analysis, we replace KITRO's 2D keypoint predictions with ours. As shown in Figure 6(a) and (b), KITRO fails to align the mesh with the actual feet, whereas our method achieves more precise alignment. In (c), KITRO mislocalizes the right hand under partial occlusion, while our 2D predictions remain robust. In (d), KITRO exhibits misalignment in both the hands and feet, while replacing its 2D predictions with ours significantly improves the alignment of peripheral body parts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 246, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 294 + ], + "type": "text", + "content": "In the supplemental material, we provide a visualization of the prototypes generated by our algorithm, illustrating how they effectively capture the diversity of representations within the embedding space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 314, + 388, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 388, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 388, + 327 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 337, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 556, + 506 + ], + "type": "text", + "content": "This paper presents a new multi-dataset training (MDT) framework for pose estimation across diverse domains and datasets. Traditional approaches encounter challenges with label heterogeneity, dataset-specific parameters, and limited domain generalization. Our approach addresses the limitations with two main contributions. First, to resolve label heterogeneity and dataset-specific parameters, we employ nonparametric prototypes to unify diverse keypoint annotations across multiple datasets. Second, to address the limited multi-dataset supervision, we propose a cross-type self-supervision mechanism. The proposed method has demonstrated substantial improvements over existing MDT across a broad range of datasets, including human pose, animal pose, hand shape, and human shape estimation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 513, + 556, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 556, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 556, + 609 + ], + "type": "text", + "content": "Limitations and future work. A limitation of our approach is that it requires at least a few-shot sample to learn prototypes, making zero-shot inference for unseen skeletons infeasible. Future work should explore extending our method for zero-shot human pose estimation. Future work should also investigate adapting the approach to 3D domains e.g. using axis-angle rotation representations, to facilitate representation learning on 3D manifolds." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by the National Research Foundation of Korea (NRF) grant (No. 2021R1A2C2012195) and by the Institute of Information & communications Technology Planning & Evaluation (IITP) grants No. RS-2019-II191906 (AI Graduate School Program, POSTECH), No. RS-2020-II201336 (AI Graduate School Program, UNIST), and RS-2022-II220290, funded by the Korea government (MSIT)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12285" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "text", + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2D human pose estimation: New benchmark and state of the art analysis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "text", + "content": "[2] Mykhaylo Andriluka, Umar Iqbal, Eldar Insafutdinov, Leonid Pishchulin, Anton Milan, Juergen Gall, and Bernt Schiele. PoseTrack: A benchmark for human pose estimation and tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5167-5176, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "text", + "content": "[3] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: Deep learning for interpretable image recognition. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[4] Yanbei Chen, Manchen Wang, Abhay Mittal, Zhenlin Xu, Paolo Favaro, Joseph Tighe, and Davide Modolo. ScaleDet: A scalable multi-dataset object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7288-7297, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 314, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 295, + 367 + ], + "type": "text", + "content": "[5] Hongsuk Choi, Gyeongsik Moon, JoonKyu Park, and Kyoung Mu Lee. Learning to estimate robust 3D human mesh from in-the-wild crowded scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 434 + ], + "type": "text", + "content": "[6] Yuanzheng Ci, Yizhou Wang, Meilin Chen, Shixiang Tang, Lei Bai, Feng Zhu, Rui Zhao, Fengwei Yu, Donglian Qi, and Wanli Ouyang. UniHCP: A unified model for human-centric perceptions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17840-17852, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "text", + "content": "[7] Mickael Cormier, Aris Clepe, Andreas Specker, and Jürgen Beyerer. Where are we with human pose estimation in real-world surveillance? In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 591-601, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 481, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 534 + ], + "type": "text", + "content": "[8] Anurag Das, Yongqin Xian, Dengxin Dai, and Bernt Schiele. Weakly-supervised domain adaptive semantic segmentation with prototypical contrastive learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15434-15443, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "type": "text", + "content": "[9] Jon Donnelly, Alina Jade Barnett, and Chaofan Chen. Deformable protopnet: An interpretable image classifier using deformable prototypes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10265-10275, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[10] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4320–4329, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "type": "text", + "content": "[11] Zigang Geng, Chunyu Wang, Yixuan Wei, Ze Liu, Houqiang Li, and Han Hu. Human pose as compositional tokens. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[12] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human POSEitioning System (HPS): 3D human pose estimation and self-localization in large scenes from" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "body-mounted sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4318-4329, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "type": "text", + "content": "[13] Or Hirschorn and Shai Avidan. A graph-based approach for category-agnostic pose estimation, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 131, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 131, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 131, + 553, + 174 + ], + "type": "text", + "content": "[14] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. NeuMan: Neural human radiance field from a single video. In European Conference on Computer Vision (ECCV), 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 176, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 176, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 176, + 553, + 229 + ], + "type": "text", + "content": "[15] Roy Kapon, Guy Tevet, Daniel Cohen-Or, and Amit H Bermano. MAS: Multi-view ancestral sampling for 3D motion generation using 2D diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1965–1974, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "text", + "content": "[16] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. arXiv preprint arXiv:2408.12569, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 277, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 330 + ], + "type": "text", + "content": "[17] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "type": "text", + "content": "[18] Alberto Lamas, Siham Tabik, Antonio Cano Montes, Francisco Pérez-Hernández, Jorge García, Roberto Olmos, and Francisco Herrera. Human pose estimation for mitigating false negatives in weapon detection in video-surveillance. Neurocomputing, 489:488-503, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 389, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 442 + ], + "type": "text", + "content": "[19] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. MSeg: A composite dataset for multidomain semantic segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "type": "text", + "content": "[20] Duy Tho Le, Chenhui Gou, Stavya Datta, Hengcan Shi, Ian Reid, Jianfei Cai, and Hamid Rezatofighi. JRDB-PanoTrack: An open-world panoptic segmentation and tracking robotic dataset in crowded human environments. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22325-22334, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "text", + "content": "[21] Yanjie Li, Sen Yang, Peidong Liu, Shoukui Zhang, Yunxiao Wang, Zhicheng Wang, Wankou Yang, and Shu-Tao Xia. SimCC: A simple coordinate classification perspective for human pose estimation. In European Conference on Computer Vision (ECCV), pages 89-106. Springer, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 569, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 621 + ], + "type": "text", + "content": "[22] Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. Motion-X: A large-scale 3D expressive whole-body human motion dataset. Advances in Neural Information Processing Systems (NeurIPS), 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 679 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 1, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[24] Sun-Ao Liu, Yiheng Zhang, Zhaofan Qiu, Hongtao Xie, Yongdong Zhang, and Ting Yao. Learning orthogonal prototypes for generalized few-shot semantic segmentation. In" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12286" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[25] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. HumanGaussian: Text-driven 3D human generation with gaussian splatt-ting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6646-6657, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[26] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "text", + "content": "[27] Yawen Lu, Dongfang Liu, Qifan Wang, Cheng Han, Yiming Cui, Zhiwen Cao, Xueling Zhang, Yingjie Victor Chen, and Heng Fan. ProMotion: Prototypes as motion learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28109-28119, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[28] Gyeongsik Moon, Shouou-I Yu, He Wen, Takaaki Shiratori, and Kyoung Mu Lee. InterHand2.6M: A dataset and baseline for 3D interacting hand pose estimation from a single rgb image. In European Conference on Computer Vision (ECCV), 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[29] Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. NeuralAnnot: Neural annotator for 3D human mesh training sets. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "text", + "content": "[30] Gyeongsik Moon, Hongsuk Choi, Sanghyuk Chun, Jiyoung Lee, and Sangdoo Yun. Three Recipes for Better 3D PseudogTs of 3D Human Mesh Estimation in the Wild. In Computer Vision and Pattern Recognition Workshop (CVPRW), 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "text", + "content": "[31] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3D gaussian avatar. In European Conference on Computer Vision (ECCV), 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 509 + ], + "type": "text", + "content": "[32] Haocong Rao and Chunyan Miao. TranSG: Transformer-based skeleton graph prototype contrastive learning with structure-trajectory prompted reconstruction for person re-identification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22118-22128, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "text", + "content": "[33] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D human motion model for robust pose estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "text", + "content": "[34] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "text", + "content": "[35] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7420-7430, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[36] István Sárándi, Alexander Hermans, and Bastian Leibe. Learning 3D human pose estimation from dozens of datasets using a geometry-aware autoencoder to bridge between skeleton formats. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[37] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 140 + ], + "type": "text", + "content": "[38] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 184 + ], + "type": "text", + "content": "[39] Markus D Solbach and John K Tsotsos. Vision-based fallen person detection for the elderly. In IEEE/CVF international conference on computer vision workshops (ICCVW), pages 1433-1442, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 554, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 230 + ], + "type": "text", + "content": "[40] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 554, + 274 + ], + "type": "text", + "content": "[41] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Black Michael J., and Tao Mei. Monocular, one-stage, regression of multiple 3D people. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "text", + "content": "[42] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3D people in depth. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 322, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 554, + 365 + ], + "type": "text", + "content": "[43] Denis Tome, Patrick Peluse, Lourdes Agapito, and Hernan Badino. xR-EgoPose: Egocentric 3D human pose from an hmd camera. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7728-7738, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 367, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 554, + 420 + ], + "type": "text", + "content": "[44] Edward Vendrow, Duy Tho Le, Jianfei Cai, and Hamid Rezatofighi. JRDB-Pose: A large-scale dataset for multiperson pose estimation and tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 422, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 466 + ], + "type": "text", + "content": "[45] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 468, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 554, + 521 + ], + "type": "text", + "content": "[46] Chong Wang, Yuyuan Liu, Yuanhong Chen, Fengbei Liu, Yu Tian, Davis McCarthy, Helen Frazer, and Gustavo Carneiro. Learning support and trivial prototypes for interpretable image classification. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2062-2072, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "type": "text", + "content": "[47] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3D human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13031-13040, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 580, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 554, + 623 + ], + "type": "text", + "content": "[48] Xudong Wang, Zhaowei Cai, Dashan Gao, and Nuno Vasconcelos. Towards universal object detection by domain attention. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7289-7298, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 554, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 677 + ], + "type": "text", + "content": "[49] Zhongdao Wang, Hengshuang Zhao, Ya-Li Li, Shengjin Wang, Philip Torr, and Luca Bertinetto. Do different tracking tasks require different appearance models? Advances in Neural Information Processing Systems (NeurIPS), 34:726-738, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[50] Jiahong Wu, He Zheng, Bo Zhao, Yixin Li, Baoming Yan, Rui Liang, Wenjia Wang, Shipei Zhou, Guosen Lin, Yanwei Fu, et al. AI Challenger: A large-scale dataset for" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12287" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 94 + ], + "type": "text", + "content": "going deeper in image understanding. arXiv preprint arXiv:1711.06475, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 139 + ], + "type": "text", + "content": "[51] Hang Xu, Linpu Fang, Xiaodan Liang, Wenxiong Kang, and Zhenguo Li. Universal-RCNN: Universal object detector via transferable graph r-cnn. In AAAI Conference on Artificial Intelligence (AAAI), pages 12492–12499, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 194 + ], + "type": "text", + "content": "[52] Lumin Xu, Sheng Jin, Wang Zeng, Wentao Liu, Chen Qian, Wanli Ouyang, Ping Luo, and Xiaogang Wang. Pose for everything: Towards category-agnostic pose estimation. In European Conference on Computer Vision (ECCV), pages 398–416. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 295, + 239 + ], + "type": "text", + "content": "[53] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose++: Vision transformer for generic body pose estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "type": "text", + "content": "[54] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. UNIK: A unified framework for real-world skeleton-based action recognition. British Machine Visison Conference (BMVC), 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 338 + ], + "type": "text", + "content": "[55] Fengyuan Yang, Kerui Gu, and Angela Yao. KITRO: Refining human mesh by 2D clues and kinematic-tree rotation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1052-1061, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "text", + "content": "[56] Yuxiang Yang, Junjie Yang, Yufei Xu, Jing Zhang, Long Lan, and Dacheng Tao. APT-36K: A large-scale benchmark for animal pose estimation and tracking. Advances in Neural Information Processing Systems (NeurIPS), 35:17301-17313, 2022. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "type": "text", + "content": "[57] Sung Whan Yoon, Jun Seo, and Jaekyun Moon. TapNet: Neural network augmented with task-adaptive projection for few-shot learning. In International conference on machine learning, pages 7115-7123. PMLR, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 438, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 295, + 491 + ], + "type": "text", + "content": "[58] Hang Yu, Yufei Xu, Jing Zhang, Wei Zhao, Ziyu Guan, and Dacheng Tao. AP-10K: A benchmark for animal pose estimation in the wild. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 1, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 295, + 525 + ], + "type": "text", + "content": "[59] Xinyi Yu, Xin Zhang, Chengjun Xu, and Linlin Ou. Humanrobot collaborative interaction with human perception and action recognition. Neurocomputing, 563:126827, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 295, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 570 + ], + "type": "text", + "content": "[60] Yuhui Yuan, Rao Fu, Lang Huang, Weihong Lin, Chao Zhang, Xilin Chen, and Jingdong Wang. HRFormer: High-resolution transformer for dense prediction. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "type": "text", + "content": "[61] Baoquan Zhang, Xutao Li, Yunming Ye, Zhichao Huang, and Lisai Zhang. Prototype completion with primitive knowledge for few-shot learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3754-3762, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 678 + ], + "type": "text", + "content": "[62] Bo Zhang, Jiakang Yuan, Botian Shi, Tao Chen, Yikang Li, and Yu Qiao. Uni3D: A unified baseline for multi-dataset 3D object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9253-9262, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "text", + "content": "[63] Xiangyun Zhao, Samuel Schulter, Gaurav Sharma, Yi-Hsuan Tsai, Manmohan Chandraker, and Ying Wu. Object detection with a unified label space from multiple datasets. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 184 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "European Conference on Computer Vision (ECCV), pages 178-193. Springer, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[64] Tianfei Zhou, Wenguan Wang, Ender Konukoglu, and Luc Van Gool. Rethinking Semantic Segmentation: A Prototype View. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2582-2593, 2022. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[65] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7571-7580, 2022. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12288" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/8cdccae4-d7bc-4e3e-9398-db2b0c310587_content_list.json b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/8cdccae4-d7bc-4e3e-9398-db2b0c310587_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..201fd76f85fd5d74a8bbdc4390a5fb0e68b5ae70 --- /dev/null +++ b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/8cdccae4-d7bc-4e3e-9398-db2b0c310587_content_list.json @@ -0,0 +1,1679 @@ +[ + { + "type": "text", + "text": "PoseTraj: Pose-Aware Trajectory Control in Video Diffusion", + "text_level": 1, + "bbox": [ + 194, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Longbin Ji", + "bbox": [ + 256, + 179, + 354, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lei Zhong", + "bbox": [ + 375, + 180, + 468, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pengfei Wei2", + "bbox": [ + 490, + 180, + 596, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changjian Li", + "bbox": [ + 617, + 180, + 730, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ University of Edinburgh $^{2}$ Nanyang Technological University", + "bbox": [ + 243, + 204, + 748, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://robingg1.github.io/Pose-Traj/", + "bbox": [ + 351, + 227, + 640, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d8a515e4cadaeecb99fcec7f00238c427ecc970a822c08383060ce04c9ac172b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 257, + 225, + 329 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd1d1ee36163605c4cb8ca46b61ce448b39b3d0d5a8cf2c95c8c81f4edafc03d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 257, + 361, + 329 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dadc0f7b07c13d4558f45dfdf9adc360f5b0f26c6ee27e4b8a8a27b4b32ad7da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 257, + 496, + 329 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f11187bafe2f608908e87db374161a0148f3c8ecb831d71f8192232c6722750.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 257, + 633, + 329 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e82b3966821e1741d0120617a1b62af3b52db3e02f2465e2034f03c75768e36d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 257, + 769, + 329 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/973d4b95a84b9f9ccb2bf02f7c5696075313ccfbe7712952ca954f497c3d30bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 257, + 903, + 329 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d0d2b075072be0b2d0b75f5043287b03598af62cca53bc8efceba395699802f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 330, + 225, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/05eebfe5859eb591bd23f01c45b03730a7585106fd22295e0dd329d25dc28858.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 330, + 361, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df2f795b5570ca548141b719ef89082fa0845172a80edfffefb60946db015f5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 330, + 496, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f5f886ee2e3a6e5b113188950f74222c0e6fdad0ab58fa2cbe15372f4cf7d3b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 330, + 633, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d80524fb2a7e9420c2df122cc5cd19ea11749c7cf02ca2bf6e9147311cbff3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 330, + 769, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0867242021e791f963488ab35f4643b800fdb9fb5d0fbd15b1d45aa6e3c32e0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 330, + 903, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f83578d8b8e36c9cc2dd0f54a3d81c79d91d504fa94d5c7d694c91cc32e8361d.jpg", + "image_caption": [ + "Trajectory" + ], + "image_footnote": [], + "bbox": [ + 91, + 405, + 225, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dbd5ab5225201a46af04dae3e35fb9c2b615c0af10c2ae872bfca6b390dc24cd.jpg", + "image_caption": [ + "Frame 1" + ], + "image_footnote": [], + "bbox": [ + 227, + 405, + 361, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ccca32a98fa579b7343b0e064fcadf23b941f97927b0a8d92001713687650ff9.jpg", + "image_caption": [ + "Frame 4", + "Figure 1. PoseTraj produces plausible dragging videos, where objects follow a rotational trajectory with awareness of changing poses." + ], + "image_footnote": [], + "bbox": [ + 364, + 405, + 496, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f196f9ab28a7f0746416a91217697046f2404421da17784e619c3d8117e5593d.jpg", + "image_caption": [ + "Frame 7" + ], + "image_footnote": [], + "bbox": [ + 500, + 405, + 633, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e66b3ab2abc4b3d9992d10662124b1c16f4d1e21dd61253a98bde45e16c0178.jpg", + "image_caption": [ + "Frame 10" + ], + "image_footnote": [], + "bbox": [ + 635, + 405, + 769, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd5ea9fa9cc5d2a1e0d2ab18e6cb9f7f676d74f0447c2ed6e9cbc4d4ba2e17a0.jpg", + "image_caption": [ + "Frame 14" + ], + "image_footnote": [], + "bbox": [ + 771, + 405, + 903, + 476 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 526, + 326, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in trajectory-guided video generation have achieved notable progress. However, existing models still face challenges in generating object motions with potentially changing 6D poses under wide-range rotations, due to limited 3D understanding. To address this problem, we introduce PoseTraj, a pose-aware video dragging model for generating 3D-aligned motion from 2D trajectories. Our method adopts a novel two-stage pose-aware pretraining framework, improving 3D understanding across diverse trajectories. Specifically, we propose a large-scale synthetic dataset PoseTraj-10K, containing 10k videos of objects following rotational trajectories, and enhance the model perception of object pose changes by incorporating 3D bounding boxes as intermediate supervision signals. Following this, we fine-tune the trajectory-controlling module on real-world videos, applying an additional camera-disentanglement module to further refine motion accuracy. Experiments on various benchmark datasets demonstrate that our method not only excels in 3D pose-aligned dragging for rotational trajectories but also outperforms existing baselines in trajectory accuracy and video quality.", + "bbox": [ + 88, + 565, + 485, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 526, + 643, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address the problem of controlling video generation with trajectories involving potential object rotations, as illustrated in Fig. 1. Recently, significant advancements have been made in text-to-video generation [6, 10, 11]. Building on these foundational models, many works have explored the use of additional guidance signals, such as images [2, 19, 20], trajectories [36, 40, 43], audio [30, 37], and poses [14, 17, 33, 45] to control the content in the generated video. Among these, trajectory-guided motion control [36, 40] has been gaining increasing attention due to its interactive and user-friendly nature.", + "bbox": [ + 509, + 551, + 906, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Trajectory-guided motion control requires the target object to follow a specified path accurately, preserving temporal coherence and realism in the generated video. Earlier works IPoke [1] and MCDiff [7] leverage dragging animation only in the limited human domain. DragNUWA [40] is the first to attempt trajectory control in open-domain videos. DragAnything [36] further enhances object-level motion awareness by extracting object entities using corresponding masks and integrating them with trajectory guidance. However, these works constrain the object to follow the trajectory in a 2D image space, without considering potential changes of the object's 6D pose (position and ori", + "bbox": [ + 509, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22776", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0bd215226c197e72f4d12f3ff27145c8cad971495183c7a1dd054c5e770eecb.jpg", + "image_caption": [ + "Complex object entity maintenance under wide-range and steady trajectories." + ], + "image_footnote": [], + "bbox": [ + 116, + 108, + 883, + 273 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b9b776dc4da00fa3bd9cc8b5c13fa02356778a7006f0c92b37a8335c9dc4ada3.jpg", + "image_caption": [ + "Figure 2. The video generation performance of DragAnything vs. Ours under complex trajectories on static and dynamic objects. The yellow dashed line indicates the boat's orientation and helps perceive object rotation, while the orange circle highlights visual defects." + ], + "image_footnote": [], + "bbox": [ + 116, + 291, + 883, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "entation). Consequently, they can perform well in translational motion but fail to control objects when trajectories implicitly involve rotations (see Fig. 2). These limitations arise from two key factors: i) in the collected video datasets, object motions are more commonly translational, while rotations are rare and difficult to annotate automatically, ii) it is inherently ill-posed to infer potential rotations from 2D trajectories and objects in pixel space.", + "bbox": [ + 89, + 498, + 482, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a novel pose-aware, trajectory-guided motion control framework, dubbed PoseTraj, which can perceive potential rotations in the trajectory and perform corresponding control (as shown in Fig. 1). The ability to perceive rotations in the trajectory originates from a two-stage pose-aware pre-training pipeline. Specifically, we propose a specialized synthetic dataset PoseTraj-10K, containing over 10,000 videos with diverse trajectories for 2,000 distinct objects involving rotations, paired with the objects' 3D bounding boxes. This dataset circumvents the challenge of accurately estimating trajectories and 3D bounding boxes from open-world videos. With this synthetic dataset, in the first stage, we pre-train the model to generate video frames with accurate 3D bounding boxes following trajectories, without emphasizing appearance details. The introduction of 3D bounding box generation helps the model understand the object location along the driving trajectories and makes it aware of potential rota", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tional changes. In the second stage, we further pre-train the model to focus on refining object appearance with the learned knowledge of object entities and poses. After the two-stage pre-training, the model can perceive potential rotations and maintain stable object entities along the trajectory. Finally, we perform camera-disentangled fine-tuning on a real-world video dataset to enhance the model's generalization to real-world videos.", + "bbox": [ + 511, + 498, + 905, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments on the VIPSeg benchmark dataset demonstrate that PoseTraj achieves substantial improvements over DragNUWA and DragAnything in both quantitative metrics and qualitative visual performance. As illustrated in Fig. 2, our model provides more accurate motion tracking and superior object reconstruction quality, especially in scenarios involving complex rotations or dynamic objects.", + "bbox": [ + 511, + 622, + 905, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are:", + "bbox": [ + 531, + 731, + 777, + 744 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel pose-aware pre-training process that employs 3D bounding boxes as intermediate supervision to enhance the model's 6D pose understanding. This capability is further generalized to open-domain video through camera-disentangled fine-tuning.", + "- We introduce PoseTraj-10K, the first synthetic dataset augmented with both trajectories and 3D bounding boxes, featuring diverse animated objects undergoing wide-range rotational motions along complex paths.", + "- Experimental results show that PoseTraj outperforms" + ], + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "Accurate and 3D-aware motion generation under a sharp rotational trajectory.", + "bbox": [ + 256, + 93, + 728, + 107 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22777", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SoTA baselines in trajectory-following accuracy and video realism when dragging objects along a path, particularly in scenarios where rotations are included.", + "bbox": [ + 104, + 90, + 482, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 152, + 232, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Controllable Video Generation Model", + "text_level": 1, + "bbox": [ + 89, + 176, + 416, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video generation model. While image generation models [25, 25, 28, 42] have achieved impressive results, video generation remains a relatively nascent field. Early works such as NUWA [35] improved video quality by unifying representations across multimodal learning. CogVideo [13] was the first to fine-tune a text-to-video model using pretrained text-to-image weights with minimal additional parameters. Make-A-Video [29] further advanced efficiency by introducing pseudo-3D convolutional layers and temporal attention mechanisms. To mitigate the computational intensity of pixel-space diffusion, recent video diffusion models [3, 11, 14] have adopted latent diffusion methods [26]. With access to larger volumes of high-quality video data, state-of-the-art models such as VideoCrafter [6], Gen-2 [10], SVD [2], and SORA [4] have demonstrated remarkable realism and high-fidelity video generation.", + "bbox": [ + 89, + 200, + 483, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Trajectory-guided video generation. Recently, trajectory-guided video generation has gained significant attention due to its user-friendly and interactive nature. Early trajectory-based animation methods, such as IPoke [1] and MCDiff [7], were limited to human motion synthesis, relying on skeleton-based conditioning. DragNUWA [40] was the first to introduce dragging control for open-domain videos by fine-tuning the pre-trained image-to-video model SVD with trajectory guidance from large-scale real-world videos. However, maintaining temporal consistency during large motions remains a challenge, often leading to trajectory-following inaccuracies.", + "bbox": [ + 89, + 443, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this challenge, DragAnything [15] enhanced object identity awareness by leveraging a pre-trained diffusion-based image encoder for explicit entity extraction from object segmentation masks. Tora [43] further improved trajectory guidance by integrating fusion layers into each DiT basic block within a scalable architecture. Other works, such as Direct-A-Video [39] and MotionCtrl [34], aimed to disentangle object motion learning from passive camera movement, enabling more fine-grained control. However, existing models lack strong pose-awareness for handling 3D motion along rotational trajectories. Our work addresses this limitation by explicitly incorporating pose-aware pretraining to enhance the model's understanding of object pose variations.", + "bbox": [ + 89, + 625, + 483, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Training on Synthetic Dataset", + "text_level": 1, + "bbox": [ + 89, + 847, + 356, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the scarcity of comprehensive, large-scale real-world data in certain domains, many studies have leveraged syn", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The synthetic datasets to enhance model performance for specific tasks. Focusing on fine-grained 3D reconstruction with diverse materials, OpenMaterial [8] introduced a dataset of 1,001 3D objects, each manually rendered under varying materials and lighting conditions. For image decomposition, the intrinsic diffusion model was trained on the synthetic dataset InteriorVerse [44], which provides high-quality intrinsic rendering images. X-RGB [41] further developed a unified framework for image decomposition and synthesis, leveraging self-rendered synthetic image pairs conditioned on various intrinsic modalities. In controllable video generation, PuppetMaster [16] recently introduced a part-level animation model, rendering high-quality animatable objects sampled from Objaverse [9].", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For our pose-aware pretraining, we propose a synthetic dataset with complex trajectories and precisely annotated 3D bounding boxes, enabling the animation of diverse objects along various wide-range rotational trajectories.", + "bbox": [ + 511, + 303, + 903, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. PoseTraj-10K Dataset Construction", + "text_level": 1, + "bbox": [ + 511, + 380, + 831, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training a trajectory-guided video generation model requires paired data, i.e., high-quality video with clearly identifiable moving objects and precisely annotated trajectories.", + "bbox": [ + 511, + 407, + 903, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, obtaining such data from open-domain videos presents several significant challenges. First, complex motions, such as wide-range rotations, are typically rare in filtered video datasets and difficult to annotate accurately using point-tracking or flow-matching estimators. Moreover, large motions are often tightly coupled with camera movements, making it challenging to disentangle object motion from camera motion. As a result, obtaining precise 6D poses of moving objects from open-domain videos remains infeasible with current pose estimation methods [5, 22].", + "bbox": [ + 511, + 452, + 905, + 604 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4553f94bfbd32fddf816beb62b6bd869f841045c69d8bebf45bdea162e6f1e3c.jpg", + "image_caption": [ + "Figure 3. Data construction pipeline of our synthetic dataset PoseTraj-10K. The whole configuration, including environment setup, object sampling, and trajectory sampling, is displayed." + ], + "image_footnote": [], + "bbox": [ + 521, + 618, + 890, + 851 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22778", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7f5b9beeba915fd3d55f3ee0f89e7b20305d1f38fac15f773221bf3e328c925d.jpg", + "image_caption": [ + "Figure 4. Method overview. Our PoseTraj first utilizes two-stage pose-aware pre-training on our synthetic dataset to obtain 3D-enhanced awareness for rotational trajectory-following capacity and further exploits camera-disentangled finetuning to adapt this ability on open-domain videos. The dashed colorful arrows demonstrate the dedicated data flows during the three training stages, while the black arrows are shared in all stages. All yellow blocks are trainable, while other blocks are frozen." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 890, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We thus construct a synthetic dataset PoseTraj-10K, that offers three key advantages: i) it provides a diverse and comprehensive range of objects, enhancing robust motion understanding, particularly for complex geometric structures and wide-range rotations; ii) it enables the creation of a wide variety of precise trajectories, ensuring motion consistency without interference from camera movements; iii) through controlled rendering, we can effortlessly obtain precise 3D bounding box parameters, which serve as crucial intermediate supervision signals in our pretraining process.", + "bbox": [ + 88, + 395, + 480, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 illustrates our rendering pipeline implemented in Blender. First, we set up a realistic virtual scene and sampled 2,000 high-quality 3D models from Objaverse [9]. To ensure the quality of the selected models, we employ GPT-4v [21] to filter out overly complex or uncommon objects, followed by a manual selection, resulting in the top 2,000 models that best represent high-quality daily objects. Each model is normalized and assigned a unique random trajectory with randomized rotational angles, trajectory shape, and length. During the animation, the 3D model follows its designated trajectory while maintaining a rotation center, enabling rotation-focused motion for the final rendering. Fig. 3 bottom displays four frames of the rendered video, check out our dataset and the supplementary for more examples and further data construction details.", + "bbox": [ + 89, + 547, + 482, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 784, + 181, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ containing foreground objects and a driving trajectory $\\mathbf{tr} = \\{(x_i, y_i)\\}_{i=1}^L$ drawn as a set of line segment images $\\mathbf{I}_{\\mathbf{tr}} = \\{I_{tr}^i\\}_{i=1}^L$ (with last empty padding), our goal is to generate a sequence of animated video frames $\\{f_i\\}_{i=1}^L$ , where the relevant object accurately follows the trajectory, while maintaining a coherent", + "bbox": [ + 89, + 808, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and natural video appearance.", + "bbox": [ + 511, + 395, + 712, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 4 presents an overview of our approach, which is built upon a latent stable video diffusion model, SVD [2]. Our base framework mainly consists of three components: a latent diffusion model $\\epsilon_{\\theta}$ (3D U-Net [27]) to denoise the latent noise, and an encoder $(\\mathcal{E})$ - decoder $(\\mathcal{D})$ pair to compress video frames into the latent space and reconstruct the denoised latent features back into video frames. Inspired by the success of ControlNet [19] in guiding the pre-trained diffusion model, we design Traj-ControlNet, which consists of a trainable copy of the encoder blocks of SVD. Given a sequence of trajectory images $\\mathbf{I}_{\\mathrm{tr}}$ , we adopt a trajectory guider - a 3D ConvNet, to encode trajectory features, which is then fed into the Traj-ControlNet to predict the residual features for SVD, enabling the generated video to follow the trajectory. In the following subsections, we detail on the novel technical designs to achieve our goal.", + "bbox": [ + 511, + 411, + 906, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Two-stage Pose-aware Pretraining", + "text_level": 1, + "bbox": [ + 511, + 665, + 812, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance the model's ability to perceive potential 6D pose changes within the 2D-pixel space, we select the 3D bounding box (bbox) [32] as the primary 3D supervision signals, as they provide precise positional and pose information. We propose a novel two-stage pose-aware pretraining pipeline that incorporates 3Dbbox generation as an intermediate supervision signal. The two pre-training is on our PoseTraj-10K synthetic dataset.", + "bbox": [ + 511, + 688, + 905, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage One: 3D bbox-guided localization. Instead of separately regressing 3D bbox parameters and generating the object's appearance for each video frame, we simultaneously generate 3D bboxes in the image space along with the objects, as shown in Fig. 4 (red arrows). To this end, given one data example in our dataset, we first render the", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22779", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D bounding box into the pixel space on top of the image to obtain $\\mathbf{I}_{bbox}$ (the bottom-left figure in Fig. 4). We apply the same rendering process for other frames to obtain the bounding box augmented video for supervision. This joint generation process helps the model better understand the object's location along the driving trajectory and makes it aware of potential rotational changes.", + "bbox": [ + 88, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage Two: object-centric reconstruction. In practical applications, bounding boxes are not required in the generated video. Moreover, in the first stage, with explicit bounding box constraints, the model learns object entities and poses, but the generated object lacks appearance details. Therefore, in the second stage, we exclude bounding box generation and remove 3D bounding box supervision, finetuning the model to refine the object's appearance.", + "bbox": [ + 88, + 198, + 480, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are several ways to inject auxiliary 3D information, e.g., inputting depth maps or 3D bounding boxes as conditions. Our injection-by-reconstruction approach offers two key advantages. First, the bounding boxes explicitly enhance continuous 3D awareness by serving as a pixel-level supervision target (see evaluation and supplementary material). Second, these supervisions can be easily removed by simply changing the reconstruction target, thereby mitigating inference-stage mismatches caused by inaccurate estimation of these additional signals.", + "bbox": [ + 89, + 319, + 482, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Camera-disentangled Finetuning", + "text_level": 1, + "bbox": [ + 89, + 483, + 380, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the two-stage pre-training, the model learns to perceive potential rotations while preserving object shape and detailed appearance along a trajectory. However, adapting the pre-trained model to real-world videos remains a key challenge. Unlike the synthetic dataset, where the camera is static, real-world videos often exhibit unpredictable and irregular camera motions. Undistinguished camera motion and object motion can lead to errors in object tracking. Therefore, we introduce camera-disentangled finetuning, which incorporates additional camera motion information into the model to help distinguish active object movement from passive camera motion (orange arrows in Fig. 4).", + "bbox": [ + 89, + 507, + 482, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we annotate the real-world video dataset - VIPSeg [18], with motion trajectories and per-frame camera poses (i.e., extrinsic parameters, denoted as $\\{C a m^{i}\\}_{i = 1}^{L}$ ). We incorporate camera poses as additional inputs by passing them through an MLP layer and then concatenating them with the trajectory features. The concatenated feature is then fed into a zero-initialized MLP projection layer and further input to the Traj-ControlNet. At inference time, it is hard to provide the camera pose information, we then randomly (50%) drop camera poses in this fine-tuning stage, enabling the model to generate videos without camera information. This approach helps the model differentiate between active object movement and passive camera motion, thereby improving trajectory-following accuracy and en", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "hancing robustness across diverse video contexts.", + "bbox": [ + 511, + 90, + 841, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Training and Inference", + "text_level": 1, + "bbox": [ + 511, + 113, + 728, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The mean squared error (MSE) loss is commonly used in video generation training, and the objective is to minimize the discrepancy between the predicted noise from $\\epsilon_{\\theta}$ and the ground truth noise $\\epsilon$ for each latent noise $x_{t}$ at timestep $t$ . Both the two pre-training stages and fine-tuning stage leverage the same MSE loss, while the difference lies in the condition $C^i$ :", + "bbox": [ + 511, + 135, + 903, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M S E}} = \\mathbb {E} _ {x _ {t}, \\epsilon} \\left[ \\sum_ {i = 1} ^ {L} \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t}, t, C ^ {i}\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 253, + 903, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C^i$ is defined as:", + "bbox": [ + 511, + 297, + 671, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC ^ {i} = \\left\\{ \\begin{array}{l l} \\left\\{I _ {t r} ^ {i}, \\mathbf {I} _ {b b o x} \\right\\}, & \\text {s t a g e o n e ,} \\\\ \\left\\{I _ {t r} ^ {i}, \\mathbf {I} \\right\\}, & \\text {s t a g e t w o ,} \\\\ \\left\\{I _ {t r} ^ {i}, \\mathbf {I}, C a m ^ {i} \\right\\}, & \\text {f i n e t u n i n g .} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 318, + 903, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remeber that, in the first pre-training stage, we used the bbox-augmented video frames as the reconstruction target.", + "bbox": [ + 511, + 380, + 903, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial enhancement loss. Simply aggregating errors across all frames overlooks the spatial reconstruction accuracy of individual frames, potentially leading to object entity collapse under large rotational motions. Since each short per-frame trajectory segment provides precise localization information within individual frames, we introduce an additional spatial enhancement loss to improve spatial consistency by reconstructing the per-frame image. Specifically, we randomly sample the trajectory of the $j$ -th frame (i.e., $I_{tr}^{j}$ ) and use it as the Traj-ControlNet condition. From its corresponding latent noise $x_{t,j}$ and the initial frame $\\mathbf{I}$ ( $\\mathbf{I}_{bbox}$ for the first stage), the objective is to train the model to generate the target noise $\\epsilon_{j}$ for this specific frame (i.e., single frame image reconstruction), formulated as:", + "bbox": [ + 511, + 410, + 905, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S P A}} = \\left| \\right.\\left| \\right. \\epsilon_ {j} - \\epsilon_ {\\theta} \\left(x _ {t, j}, t, C ^ {j}\\right)\\left. \\right)\\left. \\right|\\left. \\right| _ {2} ^ {2}, j \\in (1 \\sim L). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 628, + 903, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During backpropagation, only the spatial layers are updated, ensuring the refinement focuses on per-frame spatial accuracy.", + "bbox": [ + 511, + 654, + 903, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, the training loss function of our framework for each stage is defined as:", + "bbox": [ + 511, + 699, + 903, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {M S E}} + \\lambda_ {\\text {S P A}} \\mathcal {L} _ {\\text {S P A}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 737, + 903, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{SPA}}$ is a hyperparameter.", + "bbox": [ + 511, + 761, + 728, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference time video generation. During inference, given an input image, users freely draw a custom trajectory, which can be either rotational or translational, our model will generate corresponding pose-aware videos. Note that our approach supports more than one trajectory, see Figs. 2 and 6 for examples. In addition, users can control the camera motion behavior by providing camera poses extracted from existing videos.", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22780", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/990f58f3dfd16340ba09be3ba509de14b9e222a27b148d0ae34554c66ab5773f.jpg", + "image_caption": [ + "Figure 5. Visual comparison. Given the same trajectory and the initial image, our method produces plausible video frames containing the moving object following the given trajectory. In contrast, DrafNUWA and DragAnything either introduce unexpected camera motions or fail to maintain the object entity, causing severe collapse." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 906, + 467 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/128e0de99688027651cc9293d1b7668ddfa5b008e28179c91a5a83a83f824fa1.jpg", + "table_caption": [ + "Table 1. Quantitative comparison between our approach and the SoTA methods." + ], + "table_footnote": [], + "table_body": "
MethodsVIPSeg 256 × 256VIPSeg 320 × 576DAVIS 320 × 576
ObjMC ↓FID ↓FVD ↓ObjMC ↓FID ↓FVD ↓ObjMC ↓FID ↓FVD ↓
DragNUWA 1.5 [40]173.6163.39449.59133.0541.88289.1574.0754.10952.87
DragAnything [36]100.2361.69410.7091.1239.29275.9347.0150.83771.78
Ours87.5646.60384.4177.4838.41267.3329.9251.48729.16
", + "bbox": [ + 174, + 542, + 820, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments and Evaluations", + "text_level": 1, + "bbox": [ + 89, + 654, + 362, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Our base model is SVD [2], trained to generate 14 frames given an initial frame at a resolution of $320 \\times 576$ pixels. All experiments are conducted using a single Tesla A100 GPU with a batch size of 1. More implementation details are in the supplementary.", + "bbox": [ + 89, + 684, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic dataset. For our synthetic dataset used in training, we generated 10,000 videos by sampling 5 random trajectories for each of 2,000 high-quality objects and rendering them in Blender through the Cycles engine at 5 fps. For evaluation, we create an additional trajectory for a subset of 200 sampled objects, which serves as our test set. During trajectory annotation, we augment the trajectory placement through a trajectory sampler (see the supplementary). Both training and evaluation on the synthetic dataset are", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "conducted in the resolution of $320 \\times 576$ pixels with 14 frames per sequence.", + "bbox": [ + 511, + 656, + 906, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-world dataset. For real-world finetuning, we adopted the VIPSeg dataset [18], following DragAnything. VIPSeg is a large-scale video segmentation dataset comprising over 3,000 high-quality videos with annotated object masks. We utilized its training set for finetuning and its validation set for evaluation, employing the same trajectory extraction pipeline as in DragAnything (see supplementary).", + "bbox": [ + 511, + 686, + 905, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further evaluate the generalization capability of our approach, we additionally select the DAVIS dataset [24], a high-quality video segmentation benchmark, for out-of-distribution (OOD) evaluation. Note that, we only annotate the trajectory for this dataset for OOD evaluation, and at inference, camera poses are not used.", + "bbox": [ + 511, + 794, + 905, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We follow established evaluation met-", + "bbox": [ + 511, + 885, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22781", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rics from previous trajectory-guided studies [23, 36, 40], assessing performance across two key aspects:", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Controlling accuracy. To quantify the precision of trajectory-guided dragging, we use the ObjMC metric, which computes the mean squared error between the temporal motion tendency of generated samples and that of the ground truth videos.", + "- Generation quality. We employ the Fréchet Inception Distance (FID) [12] to assess the single-frame spatial quality and the Fréchet Video Distance (FVD) [31] to measure temporal consistency." + ], + "bbox": [ + 89, + 122, + 482, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Comparison", + "text_level": 1, + "bbox": [ + 89, + 271, + 222, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Competitors. We primarily evaluate our model's performance against two recent trajectory-guided video generation models: DragNUWA 1.5 [40] and DragAnything [36]. DragNUWA is the first method designed for open-domain video dragging. It utilizes optical flow estimated by UniMatch [38] as trajectory guidance, combining these trajectories with text and image conditions to enable motion-controllable fine-tuning. Built upon DragNUWA, DragAnything enhances object-level animation by extracting entity feature maps from instance masks and integrating them into ControlNet for further fine-tuning. To ensure fair comparisons, the same trajectory is converted into each model's respective trajectory representation (e.g., the trajectory point set or trajectory segment images). Additionally, since DragAnything reports metric performance only at a resolution of $256 \\times 256$ , we evaluate our model at both the original $320 \\times 576$ training resolution and the downsampled $256 \\times 256$ resolution to ensure a fair comparison. Due to the unavailability of public model weights for Tora and MotionCtrl with SVD pretraining, we exclude these models from our comparison.", + "bbox": [ + 89, + 294, + 482, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. As shown in Tab. 1, PoseTraj outperforms competing methods across both resolutions in all trajectory-specific and video quality metrics, with particularly strong performance in lower-resolution synthesis. In terms of trajectory accuracy, our model achieves a significant reduction in error than DragAnything, with improvements of $13\\%$ , $15\\%$ , and $36\\%$ for lower and standard resolutions on the VIPSeg dataset and standard resolution on the DAVIS dataset, respectively. These results demonstrate the effectiveness of our pose-aware pretraining approach, which leverages high-precision synthetic data pairs with 3D bounding boxes and incorporates a camera-disentanglement module to enhance the model's understanding of object-oriented motion and passive camera movement.", + "bbox": [ + 89, + 612, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For single-frame reconstruction performance, measured by FID, our model outperforms DragAnything by a significant margin, achieving an improvement of 15.01 at the downsampled $256 \\times 256$ resolution and 0.88 at the standard $320 \\times 576$ resolution on the VIPSeg dataset. Fur", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/21147d34b8964b2b77f7074bcc144407ae68552462ee37f1f45a8c82dcf609f9.jpg", + "image_caption": [ + "Figure 6. Transitional samples on DAVIS dataset." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "thermore, our model demonstrates improved temporal consistency across frames, achieving substantial reductions in FVD scores: 26.29, 8.6, and 42.62 across the three evaluation settings. These results highlight our model's significant advancements in trajectory-guided motion control, achieving precise dragging-following and reliable motion. Notably, our approach shows strong performance in traditional in-distribution scenarios while exhibiting even greater superiority in out-of-distribution (OOD) settings (see the evaluation on DAVIS in Tab. 1).", + "bbox": [ + 511, + 275, + 903, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Visual comparisons are displayed in Fig. 5, where PoseTraj achieves superior performance in generating 3D and pose-aligned rotational motions that accurately follow 2D trajectories with precise pixel-level trajectory adherence. Furthermore, our model demonstrates enhanced capability in reliably reconstructing occluded object regions, even under changing 6D poses, such as rotational car doors. In contrast, DragNUWA primarily generates moving camera perspectives while keeping object poses static, resulting in poor rotational alignment. DragAnything produces unstable and mismatched rotational motions, where dragged pixels progressively drift across frames, potentially leading to entity collapse. This comparison highlights the effectiveness of our approach in generating more accurate, stable, and realistic motion sequences, particularly when handling complex trajectories.", + "bbox": [ + 511, + 426, + 905, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, in Fig. 6, we show visual results on OOD examples, where the trajectory is mostly translational instead of rotational. Our generated videos are superior to DragAnything.", + "bbox": [ + 511, + 667, + 905, + 728 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01a88f571a72a3edced5ee8d48117ae602eb08535215bad1eee8ecb975d51339.jpg", + "image_caption": [ + "Figure 7. The voting results of the user evaluation." + ], + "image_footnote": [], + "bbox": [ + 531, + 743, + 877, + 844 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Evaluation. Other than the visual and statistical evaluations, we conduct a subjective user evaluation, where par", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22782", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1d31600e54d0f6ed5e8664941d01231b783686921e088210ff1cc80238dc3186.jpg", + "table_caption": [ + "Table 2. Ablation results on the real-world video test set with several model variants." + ], + "table_footnote": [], + "table_body": "
Methods320 × 576
ObjMC ↓FID ↓FVD ↓
Full method77.4838.41267.33
Nobbox stage81.3641.90275.40
No pretrain145.7242.62486.84
No Cam-disen83.2239.71279.15
No SPA-loss137.2639.79436.56
", + "bbox": [ + 161, + 121, + 410, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e5783d82ebadafa89b74986b7a8629028c0122389f3acab8de26bdcb886f27d0.jpg", + "table_caption": [ + "Table 3. Comparison results on the synthetic validation set of our designed ablation studies on the pretraining stage." + ], + "table_footnote": [], + "table_body": "
MethodsSynthetic Dataset Val
ObjMC ↓FID ↓FVD ↓
Full method0.19647.18187.12
Nobbox stage0.26650.30223.74
No SPA-loss0.26546.72197.68
", + "bbox": [ + 151, + 252, + 419, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ticipants rate generated videos (20 sampled videos from the VIPSeg test set and manually annotated wild images) based on two criteria: overall motion accuracy and visual quality. We have invited seven experienced experts on the video generation topic and asked them to rate the videos.", + "bbox": [ + 89, + 340, + 483, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 7, our approach is favorably voted, receiving $43\\%$ more votes in trajectory-following accuracy and $39\\%$ more votes in video quality. These findings are consistent with the superior performance observed in quantitative metrics and qualitative visualizations.", + "bbox": [ + 89, + 416, + 483, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 500, + 243, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pose-aware pretraining. To evaluate the impact of our pretraining strategy, we propose two experimental variants: (1) training the model without 3D-aware pertaining (denoted as 'No pretrain'), and (2) training the model using only the second-stage pretraining while omitting 3D bounding box supervision (denoted as 'No bbox stage').", + "bbox": [ + 89, + 523, + 483, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tabs. 2 and 3, excluding 3D bounding box supervision during pretraining leads to a degradation in performance across all evaluation metrics on both the synthetic and open-domain datasets. Notably, in Tab. 2, compared with 'No pretraining', the absence of 3D bounding box pretraining has minimal impact on the metric. This is because 3D bounding boxes mainly affect object pose localization, and inaccurate poses do not significantly degrade motion accuracy or quality. However, the visual results in Fig. 8 clearly demonstrate the defects without bounding box pretraining, i.e., there is pronounced object collapse, exacerbating trajectory misalignment. More visual comparisons can be found in the supplementary material. Furthermore, omitting the two-stage pretraining approach results in even greater performance degradation, leading to a substantial drop in trajectory-following capability.", + "bbox": [ + 89, + 614, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spatial enhancement loss. Compared to standard video generation tasks, achieving both precise trajectory control and spatial visual consistency is significantly more chal", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af3c31a3e418b98a255d48c9332257dce4c6bdd518f848a70e3af9c1b1e5e903.jpg", + "image_caption": [ + "Figure 8. Visualization of ablation results in synthetic dataset." + ], + "image_footnote": [], + "bbox": [ + 519, + 89, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lenging, especially when the motion involves complex dynamics or large-angle rotations. Consequently, integrating spatial enhancement loss and formulating image-to-image prediction as a sub-task during training becomes essential. As demonstrated in Tabs. 2 and 3 and Fig. 8, models trained without spatial loss (denoted as 'No SPA-loss') show considerable degradation in trajectory alignment reliability and generation quality, with frequent blurring observed for fast-moving objects.", + "bbox": [ + 511, + 253, + 903, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Camera disentanglement. Considering the unexpected influence of camera motions, we incorporate camera disentangled fine-tuning to enhance the model's trajectory, controlling awareness of generating naturally activated object motion with passive camera movements. We evaluate the impact of this design by discarding camera poses (denoted as 'No Cam-disen'). Experimental results (Tab. 2) indicate that models without camera disentangling exhibit lower object motion-matching accuracy and a corresponding decrease in the FVD metric. This inaccuracy arises from the model's uncertainty regarding how to generate aligned passive camera motions and results in a reduced ability to handle trajectory variations involving background motion.", + "bbox": [ + 511, + 393, + 905, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show more visual results for ablation experiments in the supplementary.", + "bbox": [ + 511, + 590, + 905, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 643, + 633, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have presented PoseTraj, a novel approach for pose-aware trajectory control in video generation. We propose a two-stage pre-training strategy to enable the accurate trajectory-following ability, while explicitly exploiting 3D bounding boxes as a supervision signal to perceive the object's potential 6D pose changes along the driving trajectory. A new dataset, PoseTraj-10K, is constructed to achieve our goal. Our approach outperforms state-of-the-art methods quantitatively and qualitatively. We hope that the 3D bounding box signal and the injection-by-reconstruction design can inspire future research in controllable video generation or relevant topics.", + "bbox": [ + 511, + 671, + 905, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. LB completed part of this work during his Master's study at the University of Edinburgh. CJ was supported by a gift from Adobe.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22783", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14707-14717, 2021. 1, 3", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3, 4, 6", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3", + "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 3", + "[5] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. European Conference on Computer Vision (ECCV), 2024. 3", + "[6] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation, 2023. 1, 3", + "[7] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 1, 3", + "[8] Zheng Dang, Jialu Huang, Fei Wang, and Mathieu Salzmann. Openmaterial: A comprehensive dataset of complex materials for 3d reconstruction. arXiv preprint arXiv:2406.08894, 2024. 3", + "[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 3, 4", + "[10] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 3", + "[11] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 1, 3", + "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[13] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3", + "[14] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 1, 3", + "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3", + "[16] Ruining Li, Chuanxia Zheng, Christian Rupprecht, and Andrea Vedaldi. Puppet-master: Scaling interactive video generation as a motion prior for part-level dynamics. arXiv preprint arXiv:2408.04631, 2024. 3", + "[17] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024. 1", + "[18] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 5, 6", + "[19] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1, 4", + "[20] Haomiao Ni, Bernhard Egger, Suhas Lohit, Anoop Cherian, Ye Wang, Toshiaki Koike-Akino, Sharon X Huang, and Tim K Marks. Ti2v-zero: Zero-shot image conditioning for text-to-video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9015-9025, 2024. 1", + "[21] OpenAI. GPT (v4) [Large language model]. https://chat.openai.com, 2024.4", + "[22] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In European Conference on Computer Vision, pages 163-182. Springer, 2025. 3", + "[23] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 7", + "[24] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference" + ], + "bbox": [ + 517, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22784", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on computer vision and pattern recognition, pages 724-732, 2016. 6", + "[25] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4", + "[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3", + "[29] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3", + "[30] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive - generating expressive portrait videos with audio2video diffusion model under weak conditions, 2024. 1", + "[31] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7", + "[32] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis, 2024. 4", + "[33] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 1", + "[34] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. arXiv preprint arXiv:2312.03641, 2023. 3", + "[35] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In European conference on computer vision, pages 720-736. Springer, 2022. 3", + "[36] Wejia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control" + ], + "bbox": [ + 91, + 92, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for anything using entity representation. arXiv preprint arXiv:2403.07420, 2024.1,6,7", + "[37] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1", + "[38] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In CVPR, 2023. 7", + "[39] Shiuyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3", + "[40] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1, 3, 6, 7", + "[41] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgbx: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3", + "[42] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to $360^{\\circ}$ panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[43] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 1, 3", + "[44] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 3", + "[45] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 696 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22785", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/8cdccae4-d7bc-4e3e-9398-db2b0c310587_model.json b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/8cdccae4-d7bc-4e3e-9398-db2b0c310587_model.json new file mode 100644 index 0000000000000000000000000000000000000000..038a1a9fe75f13ff07570fdf86075c0378ff95e9 --- /dev/null +++ b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/8cdccae4-d7bc-4e3e-9398-db2b0c310587_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "PoseTraj: Pose-Aware Trajectory Control in Video Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.18, + 0.355, + 0.2 + ], + "angle": 0, + "content": "Longbin Ji" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.181, + 0.469, + 0.2 + ], + "angle": 0, + "content": "Lei Zhong" + }, + { + "type": "text", + "bbox": [ + 0.491, + 0.181, + 0.598, + 0.199 + ], + "angle": 0, + "content": "Pengfei Wei2" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.181, + 0.732, + 0.2 + ], + "angle": 0, + "content": "Changjian Li" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.205, + 0.749, + 0.225 + ], + "angle": 0, + "content": "\\(^{1}\\)University of Edinburgh \\(^{2}\\)Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.228, + 0.642, + 0.246 + ], + "angle": 0, + "content": "https://robingg1.github.io/Pose-Traj/" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.258, + 0.226, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.258, + 0.362, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.258, + 0.498, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.258, + 0.634, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.258, + 0.77, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.258, + 0.905, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.332, + 0.226, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.332, + 0.362, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.332, + 0.498, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.332, + 0.634, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.332, + 0.77, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.332, + 0.905, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.406, + 0.226, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.481, + 0.19, + 0.493 + ], + "angle": 0, + "content": "Trajectory" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.406, + 0.362, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.481, + 0.319, + 0.492 + ], + "angle": 0, + "content": "Frame 1" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.406, + 0.498, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.481, + 0.456, + 0.492 + ], + "angle": 0, + "content": "Frame 4" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.406, + 0.634, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.481, + 0.591, + 0.492 + ], + "angle": 0, + "content": "Frame 7" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.406, + 0.77, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.481, + 0.731, + 0.492 + ], + "angle": 0, + "content": "Frame 10" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.406, + 0.905, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.481, + 0.866, + 0.492 + ], + "angle": 0, + "content": "Frame 14" + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.5, + 0.893, + 0.516 + ], + "angle": 0, + "content": "Figure 1. PoseTraj produces plausible dragging videos, where objects follow a rotational trajectory with awareness of changing poses." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.527, + 0.327, + 0.542 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.566, + 0.486, + 0.884 + ], + "angle": 0, + "content": "Recent advancements in trajectory-guided video generation have achieved notable progress. However, existing models still face challenges in generating object motions with potentially changing 6D poses under wide-range rotations, due to limited 3D understanding. To address this problem, we introduce PoseTraj, a pose-aware video dragging model for generating 3D-aligned motion from 2D trajectories. Our method adopts a novel two-stage pose-aware pretraining framework, improving 3D understanding across diverse trajectories. Specifically, we propose a large-scale synthetic dataset PoseTraj-10K, containing 10k videos of objects following rotational trajectories, and enhance the model perception of object pose changes by incorporating 3D bounding boxes as intermediate supervision signals. Following this, we fine-tune the trajectory-controlling module on real-world videos, applying an additional camera-disentanglement module to further refine motion accuracy. Experiments on various benchmark datasets demonstrate that our method not only excels in 3D pose-aligned dragging for rotational trajectories but also outperforms existing baselines in trajectory accuracy and video quality." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.527, + 0.645, + 0.542 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.553, + 0.907, + 0.719 + ], + "angle": 0, + "content": "We address the problem of controlling video generation with trajectories involving potential object rotations, as illustrated in Fig. 1. Recently, significant advancements have been made in text-to-video generation [6, 10, 11]. Building on these foundational models, many works have explored the use of additional guidance signals, such as images [2, 19, 20], trajectories [36, 40, 43], audio [30, 37], and poses [14, 17, 33, 45] to control the content in the generated video. Among these, trajectory-guided motion control [36, 40] has been gaining increasing attention due to its interactive and user-friendly nature." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Trajectory-guided motion control requires the target object to follow a specified path accurately, preserving temporal coherence and realism in the generated video. Earlier works IPoke [1] and MCDiff [7] leverage dragging animation only in the limited human domain. DragNUWA [40] is the first to attempt trajectory control in open-domain videos. DragAnything [36] further enhances object-level motion awareness by extracting object entities using corresponding masks and integrating them with trajectory guidance. However, these works constrain the object to follow the trajectory in a 2D image space, without considering potential changes of the object's 6D pose (position and ori" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22776" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.257, + 0.094, + 0.729, + 0.108 + ], + "angle": 0, + "content": "Accurate and 3D-aware motion generation under a sharp rotational trajectory." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.109, + 0.884, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.276, + 0.735, + 0.29 + ], + "angle": 0, + "content": "Complex object entity maintenance under wide-range and steady trajectories." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.292, + 0.885, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.458, + 0.908, + 0.486 + ], + "angle": 0, + "content": "Figure 2. The video generation performance of DragAnything vs. Ours under complex trajectories on static and dynamic objects. The yellow dashed line indicates the boat's orientation and helps perceive object rotation, while the orange circle highlights visual defects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.483, + 0.62 + ], + "angle": 0, + "content": "entation). Consequently, they can perform well in translational motion but fail to control objects when trajectories implicitly involve rotations (see Fig. 2). These limitations arise from two key factors: i) in the collected video datasets, object motions are more commonly translational, while rotations are rare and difficult to annotate automatically, ii) it is inherently ill-posed to infer potential rotations from 2D trajectories and objects in pixel space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present a novel pose-aware, trajectory-guided motion control framework, dubbed PoseTraj, which can perceive potential rotations in the trajectory and perform corresponding control (as shown in Fig. 1). The ability to perceive rotations in the trajectory originates from a two-stage pose-aware pre-training pipeline. Specifically, we propose a specialized synthetic dataset PoseTraj-10K, containing over 10,000 videos with diverse trajectories for 2,000 distinct objects involving rotations, paired with the objects' 3D bounding boxes. This dataset circumvents the challenge of accurately estimating trajectories and 3D bounding boxes from open-world videos. With this synthetic dataset, in the first stage, we pre-train the model to generate video frames with accurate 3D bounding boxes following trajectories, without emphasizing appearance details. The introduction of 3D bounding box generation helps the model understand the object location along the driving trajectories and makes it aware of potential rota" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.906, + 0.619 + ], + "angle": 0, + "content": "tional changes. In the second stage, we further pre-train the model to focus on refining object appearance with the learned knowledge of object entities and poses. After the two-stage pre-training, the model can perceive potential rotations and maintain stable object entities along the trajectory. Finally, we perform camera-disentangled fine-tuning on a real-world video dataset to enhance the model's generalization to real-world videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.906, + 0.729 + ], + "angle": 0, + "content": "Experiments on the VIPSeg benchmark dataset demonstrate that PoseTraj achieves substantial improvements over DragNUWA and DragAnything in both quantitative metrics and qualitative visual performance. As illustrated in Fig. 2, our model provides more accurate motion tracking and superior object reconstruction quality, especially in scenarios involving complex rotations or dynamic objects." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.732, + 0.779, + 0.746 + ], + "angle": 0, + "content": "To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.825 + ], + "angle": 0, + "content": "- We propose a novel pose-aware pre-training process that employs 3D bounding boxes as intermediate supervision to enhance the model's 6D pose understanding. This capability is further generalized to open-domain video through camera-disentangled fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.886 + ], + "angle": 0, + "content": "- We introduce PoseTraj-10K, the first synthetic dataset augmented with both trajectories and 3D bounding boxes, featuring diverse animated objects undergoing wide-range rotational motions along complex paths." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Experimental results show that PoseTraj outperforms" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22777" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.105, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "SoTA baselines in trajectory-following accuracy and video realism when dragging objects along a path, particularly in scenarios where rotations are included." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.153, + 0.233, + 0.168 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.178, + 0.418, + 0.194 + ], + "angle": 0, + "content": "2.1. Controllable Video Generation Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.201, + 0.484, + 0.442 + ], + "angle": 0, + "content": "Video generation model. While image generation models [25, 25, 28, 42] have achieved impressive results, video generation remains a relatively nascent field. Early works such as NUWA [35] improved video quality by unifying representations across multimodal learning. CogVideo [13] was the first to fine-tune a text-to-video model using pretrained text-to-image weights with minimal additional parameters. Make-A-Video [29] further advanced efficiency by introducing pseudo-3D convolutional layers and temporal attention mechanisms. To mitigate the computational intensity of pixel-space diffusion, recent video diffusion models [3, 11, 14] have adopted latent diffusion methods [26]. With access to larger volumes of high-quality video data, state-of-the-art models such as VideoCrafter [6], Gen-2 [10], SVD [2], and SORA [4] have demonstrated remarkable realism and high-fidelity video generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.625 + ], + "angle": 0, + "content": "Trajectory-guided video generation. Recently, trajectory-guided video generation has gained significant attention due to its user-friendly and interactive nature. Early trajectory-based animation methods, such as IPoke [1] and MCDiff [7], were limited to human motion synthesis, relying on skeleton-based conditioning. DragNUWA [40] was the first to introduce dragging control for open-domain videos by fine-tuning the pre-trained image-to-video model SVD with trajectory guidance from large-scale real-world videos. However, maintaining temporal consistency during large motions remains a challenge, often leading to trajectory-following inaccuracies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.837 + ], + "angle": 0, + "content": "To address this challenge, DragAnything [15] enhanced object identity awareness by leveraging a pre-trained diffusion-based image encoder for explicit entity extraction from object segmentation masks. Tora [43] further improved trajectory guidance by integrating fusion layers into each DiT basic block within a scalable architecture. Other works, such as Direct-A-Video [39] and MotionCtrl [34], aimed to disentangle object motion learning from passive camera movement, enabling more fine-grained control. However, existing models lack strong pose-awareness for handling 3D motion along rotational trajectories. Our work addresses this limitation by explicitly incorporating pose-aware pretraining to enhance the model's understanding of object pose variations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.357, + 0.865 + ], + "angle": 0, + "content": "2.2. Training on Synthetic Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Due to the scarcity of comprehensive, large-scale real-world data in certain domains, many studies have leveraged syn" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "The synthetic datasets to enhance model performance for specific tasks. Focusing on fine-grained 3D reconstruction with diverse materials, OpenMaterial [8] introduced a dataset of 1,001 3D objects, each manually rendered under varying materials and lighting conditions. For image decomposition, the intrinsic diffusion model was trained on the synthetic dataset InteriorVerse [44], which provides high-quality intrinsic rendering images. X-RGB [41] further developed a unified framework for image decomposition and synthesis, leveraging self-rendered synthetic image pairs conditioned on various intrinsic modalities. In controllable video generation, PuppetMaster [16] recently introduced a part-level animation model, rendering high-quality animatable objects sampled from Objaverse [9]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.905, + 0.365 + ], + "angle": 0, + "content": "For our pose-aware pretraining, we propose a synthetic dataset with complex trajectories and precisely annotated 3D bounding boxes, enabling the animation of diverse objects along various wide-range rotational trajectories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.833, + 0.399 + ], + "angle": 0, + "content": "3. PoseTraj-10K Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.905, + 0.453 + ], + "angle": 0, + "content": "Training a trajectory-guided video generation model requires paired data, i.e., high-quality video with clearly identifiable moving objects and precisely annotated trajectories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.906, + 0.605 + ], + "angle": 0, + "content": "However, obtaining such data from open-domain videos presents several significant challenges. First, complex motions, such as wide-range rotations, are typically rare in filtered video datasets and difficult to annotate accurately using point-tracking or flow-matching estimators. Moreover, large motions are often tightly coupled with camera movements, making it challenging to disentangle object motion from camera motion. As a result, obtaining precise 6D poses of moving objects from open-domain videos remains infeasible with current pose estimation methods [5, 22]." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.619, + 0.892, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.855, + 0.905, + 0.898 + ], + "angle": 0, + "content": "Figure 3. Data construction pipeline of our synthetic dataset PoseTraj-10K. The whole configuration, including environment setup, object sampling, and trajectory sampling, is displayed." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22778" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.891, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.328, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 4. Method overview. Our PoseTraj first utilizes two-stage pose-aware pre-training on our synthetic dataset to obtain 3D-enhanced awareness for rotational trajectory-following capacity and further exploits camera-disentangled finetuning to adapt this ability on open-domain videos. The dashed colorful arrows demonstrate the dedicated data flows during the three training stages, while the black arrows are shared in all stages. All yellow blocks are trainable, while other blocks are frozen." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.396, + 0.482, + 0.546 + ], + "angle": 0, + "content": "We thus construct a synthetic dataset PoseTraj-10K, that offers three key advantages: i) it provides a diverse and comprehensive range of objects, enhancing robust motion understanding, particularly for complex geometric structures and wide-range rotations; ii) it enables the creation of a wide variety of precise trajectories, ensuring motion consistency without interference from camera movements; iii) through controlled rendering, we can effortlessly obtain precise 3D bounding box parameters, which serve as crucial intermediate supervision signals in our pretraining process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Fig. 3 illustrates our rendering pipeline implemented in Blender. First, we set up a realistic virtual scene and sampled 2,000 high-quality 3D models from Objaverse [9]. To ensure the quality of the selected models, we employ GPT-4v [21] to filter out overly complex or uncommon objects, followed by a manual selection, resulting in the top 2,000 models that best represent high-quality daily objects. Each model is normalized and assigned a unique random trajectory with randomized rotational angles, trajectory shape, and length. During the animation, the 3D model follows its designated trajectory while maintaining a rotation center, enabling rotation-focused motion for the final rendering. Fig. 3 bottom displays four frames of the rendered video, check out our dataset and the supplementary for more examples and further data construction details." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.785, + 0.182, + 0.8 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.809, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Given an input image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\) containing foreground objects and a driving trajectory \\(\\mathbf{tr} = \\{(x_i, y_i)\\}_{i=1}^L\\) drawn as a set of line segment images \\(\\mathbf{I}_{\\mathbf{tr}} = \\{I_{tr}^i\\}_{i=1}^L\\) (with last empty padding), our goal is to generate a sequence of animated video frames \\(\\{f_i\\}_{i=1}^L\\), where the relevant object accurately follows the trajectory, while maintaining a coherent" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.396, + 0.714, + 0.411 + ], + "angle": 0, + "content": "and natural video appearance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.655 + ], + "angle": 0, + "content": "Fig. 4 presents an overview of our approach, which is built upon a latent stable video diffusion model, SVD [2]. Our base framework mainly consists of three components: a latent diffusion model \\(\\epsilon_{\\theta}\\) (3D U-Net [27]) to denoise the latent noise, and an encoder \\((\\mathcal{E})\\) - decoder \\((\\mathcal{D})\\) pair to compress video frames into the latent space and reconstruct the denoised latent features back into video frames. Inspired by the success of ControlNet [19] in guiding the pre-trained diffusion model, we design Traj-ControlNet, which consists of a trainable copy of the encoder blocks of SVD. Given a sequence of trajectory images \\(\\mathbf{I}_{\\mathrm{tr}}\\), we adopt a trajectory guider - a 3D ConvNet, to encode trajectory features, which is then fed into the Traj-ControlNet to predict the residual features for SVD, enabling the generated video to follow the trajectory. In the following subsections, we detail on the novel technical designs to achieve our goal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.813, + 0.683 + ], + "angle": 0, + "content": "4.1. Two-stage Pose-aware Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.809 + ], + "angle": 0, + "content": "To enhance the model's ability to perceive potential 6D pose changes within the 2D-pixel space, we select the 3D bounding box (bbox) [32] as the primary 3D supervision signals, as they provide precise positional and pose information. We propose a novel two-stage pose-aware pretraining pipeline that incorporates 3Dbbox generation as an intermediate supervision signal. The two pre-training is on our PoseTraj-10K synthetic dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Stage One: 3D bbox-guided localization. Instead of separately regressing 3D bbox parameters and generating the object's appearance for each video frame, we simultaneously generate 3D bboxes in the image space along with the objects, as shown in Fig. 4 (red arrows). To this end, given one data example in our dataset, we first render the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22779" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "3D bounding box into the pixel space on top of the image to obtain \\(\\mathbf{I}_{bbox}\\) (the bottom-left figure in Fig. 4). We apply the same rendering process for other frames to obtain the bounding box augmented video for supervision. This joint generation process helps the model better understand the object's location along the driving trajectory and makes it aware of potential rotational changes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.199, + 0.482, + 0.319 + ], + "angle": 0, + "content": "Stage Two: object-centric reconstruction. In practical applications, bounding boxes are not required in the generated video. Moreover, in the first stage, with explicit bounding box constraints, the model learns object entities and poses, but the generated object lacks appearance details. Therefore, in the second stage, we exclude bounding box generation and remove 3D bounding box supervision, finetuning the model to refine the object's appearance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.483, + 0.472 + ], + "angle": 0, + "content": "There are several ways to inject auxiliary 3D information, e.g., inputting depth maps or 3D bounding boxes as conditions. Our injection-by-reconstruction approach offers two key advantages. First, the bounding boxes explicitly enhance continuous 3D awareness by serving as a pixel-level supervision target (see evaluation and supplementary material). Second, these supervisions can be easily removed by simply changing the reconstruction target, thereby mitigating inference-stage mismatches caused by inaccurate estimation of these additional signals." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.484, + 0.382, + 0.501 + ], + "angle": 0, + "content": "4.2. Camera-disentangled Finetuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.483, + 0.689 + ], + "angle": 0, + "content": "After the two-stage pre-training, the model learns to perceive potential rotations while preserving object shape and detailed appearance along a trajectory. However, adapting the pre-trained model to real-world videos remains a key challenge. Unlike the synthetic dataset, where the camera is static, real-world videos often exhibit unpredictable and irregular camera motions. Undistinguished camera motion and object motion can lead to errors in object tracking. Therefore, we introduce camera-disentangled finetuning, which incorporates additional camera motion information into the model to help distinguish active object movement from passive camera motion (orange arrows in Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Specifically, we annotate the real-world video dataset - VIPSeg [18], with motion trajectories and per-frame camera poses (i.e., extrinsic parameters, denoted as \\(\\{C a m^{i}\\}_{i = 1}^{L}\\)). We incorporate camera poses as additional inputs by passing them through an MLP layer and then concatenating them with the trajectory features. The concatenated feature is then fed into a zero-initialized MLP projection layer and further input to the Traj-ControlNet. At inference time, it is hard to provide the camera pose information, we then randomly (50%) drop camera poses in this fine-tuning stage, enabling the model to generate videos without camera information. This approach helps the model differentiate between active object movement and passive camera motion, thereby improving trajectory-following accuracy and en" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.842, + 0.107 + ], + "angle": 0, + "content": "hancing robustness across diverse video contexts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.114, + 0.729, + 0.13 + ], + "angle": 0, + "content": "4.3. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.136, + 0.905, + 0.241 + ], + "angle": 0, + "content": "The mean squared error (MSE) loss is commonly used in video generation training, and the objective is to minimize the discrepancy between the predicted noise from \\(\\epsilon_{\\theta}\\) and the ground truth noise \\(\\epsilon\\) for each latent noise \\(x_{t}\\) at timestep \\(t\\). Both the two pre-training stages and fine-tuning stage leverage the same MSE loss, while the difference lies in the condition \\(C^i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.254, + 0.905, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M S E}} = \\mathbb {E} _ {x _ {t}, \\epsilon} \\left[ \\sum_ {i = 1} ^ {L} \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t}, t, C ^ {i}\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.299, + 0.673, + 0.313 + ], + "angle": 0, + "content": "where \\(C^i\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.319, + 0.905, + 0.375 + ], + "angle": 0, + "content": "\\[\nC ^ {i} = \\left\\{ \\begin{array}{l l} \\left\\{I _ {t r} ^ {i}, \\mathbf {I} _ {b b o x} \\right\\}, & \\text {s t a g e o n e ,} \\\\ \\left\\{I _ {t r} ^ {i}, \\mathbf {I} \\right\\}, & \\text {s t a g e t w o ,} \\\\ \\left\\{I _ {t r} ^ {i}, \\mathbf {I}, C a m ^ {i} \\right\\}, & \\text {f i n e t u n i n g .} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.905, + 0.411 + ], + "angle": 0, + "content": "Remeber that, in the first pre-training stage, we used the bbox-augmented video frames as the reconstruction target." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.906, + 0.623 + ], + "angle": 0, + "content": "Spatial enhancement loss. Simply aggregating errors across all frames overlooks the spatial reconstruction accuracy of individual frames, potentially leading to object entity collapse under large rotational motions. Since each short per-frame trajectory segment provides precise localization information within individual frames, we introduce an additional spatial enhancement loss to improve spatial consistency by reconstructing the per-frame image. Specifically, we randomly sample the trajectory of the \\(j\\)-th frame (i.e., \\(I_{tr}^{j}\\)) and use it as the Traj-ControlNet condition. From its corresponding latent noise \\(x_{t,j}\\) and the initial frame \\(\\mathbf{I}\\) (\\(\\mathbf{I}_{bbox}\\) for the first stage), the objective is to train the model to generate the target noise \\(\\epsilon_{j}\\) for this specific frame (i.e., single frame image reconstruction), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.63, + 0.905, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S P A}} = \\left| \\right.\\left| \\right. \\epsilon_ {j} - \\epsilon_ {\\theta} \\left(x _ {t, j}, t, C ^ {j}\\right)\\left. \\right)\\left. \\right|\\left. \\right| _ {2} ^ {2}, j \\in (1 \\sim L). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.905, + 0.699 + ], + "angle": 0, + "content": "During backpropagation, only the spatial layers are updated, ensuring the refinement focuses on per-frame spatial accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.905, + 0.73 + ], + "angle": 0, + "content": "Overall, the training loss function of our framework for each stage is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.738, + 0.905, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {M S E}} + \\lambda_ {\\text {S P A}} \\mathcal {L} _ {\\text {S P A}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.762, + 0.729, + 0.777 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{SPA}}\\) is a hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Inference time video generation. During inference, given an input image, users freely draw a custom trajectory, which can be either rotational or translational, our model will generate corresponding pose-aware videos. Note that our approach supports more than one trajectory, see Figs. 2 and 6 for examples. In addition, users can control the camera motion behavior by providing camera poses extracted from existing videos." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22780" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.907, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.472, + 0.908, + 0.515 + ], + "angle": 0, + "content": "Figure 5. Visual comparison. Given the same trajectory and the initial image, our method produces plausible video frames containing the moving object following the given trajectory. In contrast, DrafNUWA and DragAnything either introduce unexpected camera motions or fail to maintain the object entity, causing severe collapse." + }, + { + "type": "table_caption", + "bbox": [ + 0.26, + 0.524, + 0.735, + 0.539 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison between our approach and the SoTA methods." + }, + { + "type": "table", + "bbox": [ + 0.175, + 0.543, + 0.821, + 0.637 + ], + "angle": 0, + "content": "
MethodsVIPSeg 256 × 256VIPSeg 320 × 576DAVIS 320 × 576
ObjMC ↓FID ↓FVD ↓ObjMC ↓FID ↓FVD ↓ObjMC ↓FID ↓FVD ↓
DragNUWA 1.5 [40]173.6163.39449.59133.0541.88289.1574.0754.10952.87
DragAnything [36]100.2361.69410.7091.1239.29275.9347.0150.83771.78
Ours87.5646.60384.4177.4838.41267.3329.9251.48729.16
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.364, + 0.673 + ], + "angle": 0, + "content": "5. Experiments and Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Implementation details. Our base model is SVD [2], trained to generate 14 frames given an initial frame at a resolution of \\(320 \\times 576\\) pixels. All experiments are conducted using a single Tesla A100 GPU with a batch size of 1. More implementation details are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Synthetic dataset. For our synthetic dataset used in training, we generated 10,000 videos by sampling 5 random trajectories for each of 2,000 high-quality objects and rendering them in Blender through the Cycles engine at 5 fps. For evaluation, we create an additional trajectory for a subset of 200 sampled objects, which serves as our test set. During trajectory annotation, we augment the trajectory placement through a trajectory sampler (see the supplementary). Both training and evaluation on the synthetic dataset are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.687 + ], + "angle": 0, + "content": "conducted in the resolution of \\(320 \\times 576\\) pixels with 14 frames per sequence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.794 + ], + "angle": 0, + "content": "Real-world dataset. For real-world finetuning, we adopted the VIPSeg dataset [18], following DragAnything. VIPSeg is a large-scale video segmentation dataset comprising over 3,000 high-quality videos with annotated object masks. We utilized its training set for finetuning and its validation set for evaluation, employing the same trajectory extraction pipeline as in DragAnything (see supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.884 + ], + "angle": 0, + "content": "To further evaluate the generalization capability of our approach, we additionally select the DAVIS dataset [24], a high-quality video segmentation benchmark, for out-of-distribution (OOD) evaluation. Note that, we only annotate the trajectory for this dataset for OOD evaluation, and at inference, camera poses are not used." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Evaluation metrics. We follow established evaluation met-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22781" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "rics from previous trajectory-guided studies [23, 36, 40], assessing performance across two key aspects:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.123, + 0.483, + 0.197 + ], + "angle": 0, + "content": "- Controlling accuracy. To quantify the precision of trajectory-guided dragging, we use the ObjMC metric, which computes the mean squared error between the temporal motion tendency of generated samples and that of the ground truth videos." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.199, + 0.483, + 0.259 + ], + "angle": 0, + "content": "- Generation quality. We employ the Fréchet Inception Distance (FID) [12] to assess the single-frame spatial quality and the Fréchet Video Distance (FVD) [31] to measure temporal consistency." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.123, + 0.483, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.272, + 0.223, + 0.287 + ], + "angle": 0, + "content": "5.1. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.295, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Competitors. We primarily evaluate our model's performance against two recent trajectory-guided video generation models: DragNUWA 1.5 [40] and DragAnything [36]. DragNUWA is the first method designed for open-domain video dragging. It utilizes optical flow estimated by UniMatch [38] as trajectory guidance, combining these trajectories with text and image conditions to enable motion-controllable fine-tuning. Built upon DragNUWA, DragAnything enhances object-level animation by extracting entity feature maps from instance masks and integrating them into ControlNet for further fine-tuning. To ensure fair comparisons, the same trajectory is converted into each model's respective trajectory representation (e.g., the trajectory point set or trajectory segment images). Additionally, since DragAnything reports metric performance only at a resolution of \\(256 \\times 256\\), we evaluate our model at both the original \\(320 \\times 576\\) training resolution and the downsampled \\(256 \\times 256\\) resolution to ensure a fair comparison. Due to the unavailability of public model weights for Tora and MotionCtrl with SVD pretraining, we exclude these models from our comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Quantitative Results. As shown in Tab. 1, PoseTraj outperforms competing methods across both resolutions in all trajectory-specific and video quality metrics, with particularly strong performance in lower-resolution synthesis. In terms of trajectory accuracy, our model achieves a significant reduction in error than DragAnything, with improvements of \\(13\\%\\), \\(15\\%\\), and \\(36\\%\\) for lower and standard resolutions on the VIPSeg dataset and standard resolution on the DAVIS dataset, respectively. These results demonstrate the effectiveness of our pose-aware pretraining approach, which leverages high-precision synthetic data pairs with 3D bounding boxes and incorporates a camera-disentanglement module to enhance the model's understanding of object-oriented motion and passive camera movement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For single-frame reconstruction performance, measured by FID, our model outperforms DragAnything by a significant margin, achieving an improvement of 15.01 at the downsampled \\(256 \\times 256\\) resolution and 0.88 at the standard \\(320 \\times 576\\) resolution on the VIPSeg dataset. Fur" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.25, + 0.856, + 0.263 + ], + "angle": 0, + "content": "Figure 6. Transitional samples on DAVIS dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.905, + 0.426 + ], + "angle": 0, + "content": "thermore, our model demonstrates improved temporal consistency across frames, achieving substantial reductions in FVD scores: 26.29, 8.6, and 42.62 across the three evaluation settings. These results highlight our model's significant advancements in trajectory-guided motion control, achieving precise dragging-following and reliable motion. Notably, our approach shows strong performance in traditional in-distribution scenarios while exhibiting even greater superiority in out-of-distribution (OOD) settings (see the evaluation on DAVIS in Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.906, + 0.669 + ], + "angle": 0, + "content": "Qualitative Results. Visual comparisons are displayed in Fig. 5, where PoseTraj achieves superior performance in generating 3D and pose-aligned rotational motions that accurately follow 2D trajectories with precise pixel-level trajectory adherence. Furthermore, our model demonstrates enhanced capability in reliably reconstructing occluded object regions, even under changing 6D poses, such as rotational car doors. In contrast, DragNUWA primarily generates moving camera perspectives while keeping object poses static, resulting in poor rotational alignment. DragAnything produces unstable and mismatched rotational motions, where dragged pixels progressively drift across frames, potentially leading to entity collapse. This comparison highlights the effectiveness of our approach in generating more accurate, stable, and realistic motion sequences, particularly when handling complex trajectories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.906, + 0.729 + ], + "angle": 0, + "content": "Additionally, in Fig. 6, we show visual results on OOD examples, where the trajectory is mostly translational instead of rotational. Our generated videos are superior to DragAnything." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.744, + 0.878, + 0.845 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.851, + 0.859, + 0.865 + ], + "angle": 0, + "content": "Figure 7. The voting results of the user evaluation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "User Evaluation. Other than the visual and statistical evaluations, we conduct a subjective user evaluation, where par" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22782" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 2. Ablation results on the real-world video test set with several model variants." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.122, + 0.411, + 0.216 + ], + "angle": 0, + "content": "
Methods320 × 576
ObjMC ↓FID ↓FVD ↓
Full method77.4838.41267.33
Nobbox stage81.3641.90275.40
No pretrain145.7242.62486.84
No Cam-disen83.2239.71279.15
No SPA-loss137.2639.79436.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.222, + 0.483, + 0.25 + ], + "angle": 0, + "content": "Table 3. Comparison results on the synthetic validation set of our designed ablation studies on the pretraining stage." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.253, + 0.42, + 0.328 + ], + "angle": 0, + "content": "
MethodsSynthetic Dataset Val
ObjMC ↓FID ↓FVD ↓
Full method0.19647.18187.12
Nobbox stage0.26650.30223.74
No SPA-loss0.26546.72197.68
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.416 + ], + "angle": 0, + "content": "ticipants rate generated videos (20 sampled videos from the VIPSeg test set and manually annotated wild images) based on two criteria: overall motion accuracy and visual quality. We have invited seven experienced experts on the video generation topic and asked them to rate the videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.493 + ], + "angle": 0, + "content": "As shown in Fig. 7, our approach is favorably voted, receiving \\(43\\%\\) more votes in trajectory-following accuracy and \\(39\\%\\) more votes in video quality. These findings are consistent with the superior performance observed in quantitative metrics and qualitative visualizations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.501, + 0.245, + 0.518 + ], + "angle": 0, + "content": "5.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Pose-aware pretraining. To evaluate the impact of our pretraining strategy, we propose two experimental variants: (1) training the model without 3D-aware pertaining (denoted as 'No pretrain'), and (2) training the model using only the second-stage pretraining while omitting 3D bounding box supervision (denoted as 'No bbox stage')." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.856 + ], + "angle": 0, + "content": "As shown in Tabs. 2 and 3, excluding 3D bounding box supervision during pretraining leads to a degradation in performance across all evaluation metrics on both the synthetic and open-domain datasets. Notably, in Tab. 2, compared with 'No pretraining', the absence of 3D bounding box pretraining has minimal impact on the metric. This is because 3D bounding boxes mainly affect object pose localization, and inaccurate poses do not significantly degrade motion accuracy or quality. However, the visual results in Fig. 8 clearly demonstrate the defects without bounding box pretraining, i.e., there is pronounced object collapse, exacerbating trajectory misalignment. More visual comparisons can be found in the supplementary material. Furthermore, omitting the two-stage pretraining approach results in even greater performance degradation, leading to a substantial drop in trajectory-following capability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Spatial enhancement loss. Compared to standard video generation tasks, achieving both precise trajectory control and spatial visual consistency is significantly more chal" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.09, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.231, + 0.893, + 0.245 + ], + "angle": 0, + "content": "Figure 8. Visualization of ablation results in synthetic dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.905, + 0.391 + ], + "angle": 0, + "content": "lenging, especially when the motion involves complex dynamics or large-angle rotations. Consequently, integrating spatial enhancement loss and formulating image-to-image prediction as a sub-task during training becomes essential. As demonstrated in Tabs. 2 and 3 and Fig. 8, models trained without spatial loss (denoted as 'No SPA-loss') show considerable degradation in trajectory alignment reliability and generation quality, with frequent blurring observed for fast-moving objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.906, + 0.59 + ], + "angle": 0, + "content": "Camera disentanglement. Considering the unexpected influence of camera motions, we incorporate camera disentangled fine-tuning to enhance the model's trajectory, controlling awareness of generating naturally activated object motion with passive camera movements. We evaluate the impact of this design by discarding camera poses (denoted as 'No Cam-disen'). Experimental results (Tab. 2) indicate that models without camera disentangling exhibit lower object motion-matching accuracy and a corresponding decrease in the FVD metric. This inaccuracy arises from the model's uncertainty regarding how to generate aligned passive camera motions and results in a reduced ability to handle trajectory variations involving background motion." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.592, + 0.906, + 0.623 + ], + "angle": 0, + "content": "We show more visual results for ablation experiments in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.644, + 0.634, + 0.66 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.853 + ], + "angle": 0, + "content": "In this paper, we have presented PoseTraj, a novel approach for pose-aware trajectory control in video generation. We propose a two-stage pre-training strategy to enable the accurate trajectory-following ability, while explicitly exploiting 3D bounding boxes as a supervision signal to perceive the object's potential 6D pose changes along the driving trajectory. A new dataset, PoseTraj-10K, is constructed to achieve our goal. Our approach outperforms state-of-the-art methods quantitatively and qualitatively. We hope that the 3D bounding box signal and the injection-by-reconstruction design can inspire future research in controllable video generation or relevant topics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. LB completed part of this work during his Master's study at the University of Edinburgh. CJ was supported by a gift from Adobe." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22783" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14707-14717, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.358, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.429, + 0.482, + 0.484 + ], + "angle": 0, + "content": "[5] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. European Conference on Computer Vision (ECCV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.482, + 0.555 + ], + "angle": 0, + "content": "[6] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.558, + 0.482, + 0.612 + ], + "angle": 0, + "content": "[7] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[8] Zheng Dang, Jialu Huang, Fei Wang, and Mathieu Salzmann. Openmaterial: A comprehensive dataset of complex materials for 3d reconstruction. arXiv preprint arXiv:2406.08894, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. Advances in Neural Information Processing Systems, 36, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[10] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[11] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.905, + 0.133 + ], + "angle": 0, + "content": "two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[14] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.251, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.322, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[16] Ruining Li, Chuanxia Zheng, Christian Rupprecht, and Andrea Vedaldi. Puppet-master: Scaling interactive video generation as a motion prior for part-level dynamics. arXiv preprint arXiv:2408.04631, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.379, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[17] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[18] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.907, + 0.588 + ], + "angle": 0, + "content": "[19] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[20] Haomiao Ni, Bernhard Egger, Suhas Lohit, Anoop Cherian, Ye Wang, Toshiaki Koike-Akino, Sharon X Huang, and Tim K Marks. Ti2v-zero: Zero-shot image conditioning for text-to-video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9015-9025, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[21] OpenAI. GPT (v4) [Large language model]. https://chat.openai.com, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[22] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In European Conference on Computer Vision, pages 163-182. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[23] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22784" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "on computer vision and pattern recognition, pages 724-732, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[25] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.333 + ], + "angle": 0, + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[29] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[30] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive - generating expressive portrait videos with audio2video diffusion model under weak conditions, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[31] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[32] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[33] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[34] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. arXiv preprint arXiv:2312.03641, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.855 + ], + "angle": 0, + "content": "[35] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In European conference on computer vision, pages 720-736. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[36] Wejia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "for anything using entity representation. arXiv preprint arXiv:2403.07420, 2024.1,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[37] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[38] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[39] Shiuyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[40] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[41] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgbx: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[42] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to \\(360^{\\circ}\\) panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[43] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[44] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[45] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22785" + } + ] +] \ No newline at end of file diff --git a/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/8cdccae4-d7bc-4e3e-9398-db2b0c310587_origin.pdf b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/8cdccae4-d7bc-4e3e-9398-db2b0c310587_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eabdcf5281df604a77e497dec17781fb859f02d1 --- /dev/null +++ b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/8cdccae4-d7bc-4e3e-9398-db2b0c310587_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13a17ebb2f52c04b329d069a8aafa7ba66310ac958e45ff6a75123d18744c60b +size 8589444 diff --git a/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/full.md b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a80acdb4e336b5b7c73cc0cabf11b7bc7ab423c1 --- /dev/null +++ b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/full.md @@ -0,0 +1,321 @@ +# PoseTraj: Pose-Aware Trajectory Control in Video Diffusion + +Longbin Ji + +Lei Zhong + +Pengfei Wei2 + +Changjian Li + +$^{1}$ University of Edinburgh $^{2}$ Nanyang Technological University + +https://robingg1.github.io/Pose-Traj/ + +![](images/d8a515e4cadaeecb99fcec7f00238c427ecc970a822c08383060ce04c9ac172b.jpg) + +![](images/fd1d1ee36163605c4cb8ca46b61ce448b39b3d0d5a8cf2c95c8c81f4edafc03d.jpg) + +![](images/dadc0f7b07c13d4558f45dfdf9adc360f5b0f26c6ee27e4b8a8a27b4b32ad7da.jpg) + +![](images/8f11187bafe2f608908e87db374161a0148f3c8ecb831d71f8192232c6722750.jpg) + +![](images/e82b3966821e1741d0120617a1b62af3b52db3e02f2465e2034f03c75768e36d.jpg) + +![](images/973d4b95a84b9f9ccb2bf02f7c5696075313ccfbe7712952ca954f497c3d30bc.jpg) + +![](images/2d0d2b075072be0b2d0b75f5043287b03598af62cca53bc8efceba395699802f.jpg) + +![](images/05eebfe5859eb591bd23f01c45b03730a7585106fd22295e0dd329d25dc28858.jpg) + +![](images/df2f795b5570ca548141b719ef89082fa0845172a80edfffefb60946db015f5f.jpg) + +![](images/f5f886ee2e3a6e5b113188950f74222c0e6fdad0ab58fa2cbe15372f4cf7d3b3.jpg) + +![](images/9d80524fb2a7e9420c2df122cc5cd19ea11749c7cf02ca2bf6e9147311cbff3b.jpg) + +![](images/0867242021e791f963488ab35f4643b800fdb9fb5d0fbd15b1d45aa6e3c32e0f.jpg) + +![](images/f83578d8b8e36c9cc2dd0f54a3d81c79d91d504fa94d5c7d694c91cc32e8361d.jpg) +Trajectory + +![](images/dbd5ab5225201a46af04dae3e35fb9c2b615c0af10c2ae872bfca6b390dc24cd.jpg) +Frame 1 + +![](images/ccca32a98fa579b7343b0e064fcadf23b941f97927b0a8d92001713687650ff9.jpg) +Frame 4 +Figure 1. PoseTraj produces plausible dragging videos, where objects follow a rotational trajectory with awareness of changing poses. + +![](images/f196f9ab28a7f0746416a91217697046f2404421da17784e619c3d8117e5593d.jpg) +Frame 7 + +![](images/9e66b3ab2abc4b3d9992d10662124b1c16f4d1e21dd61253a98bde45e16c0178.jpg) +Frame 10 + +![](images/cd5ea9fa9cc5d2a1e0d2ab18e6cb9f7f676d74f0447c2ed6e9cbc4d4ba2e17a0.jpg) +Frame 14 + +# Abstract + +Recent advancements in trajectory-guided video generation have achieved notable progress. However, existing models still face challenges in generating object motions with potentially changing 6D poses under wide-range rotations, due to limited 3D understanding. To address this problem, we introduce PoseTraj, a pose-aware video dragging model for generating 3D-aligned motion from 2D trajectories. Our method adopts a novel two-stage pose-aware pretraining framework, improving 3D understanding across diverse trajectories. Specifically, we propose a large-scale synthetic dataset PoseTraj-10K, containing 10k videos of objects following rotational trajectories, and enhance the model perception of object pose changes by incorporating 3D bounding boxes as intermediate supervision signals. Following this, we fine-tune the trajectory-controlling module on real-world videos, applying an additional camera-disentanglement module to further refine motion accuracy. Experiments on various benchmark datasets demonstrate that our method not only excels in 3D pose-aligned dragging for rotational trajectories but also outperforms existing baselines in trajectory accuracy and video quality. + +# 1. Introduction + +We address the problem of controlling video generation with trajectories involving potential object rotations, as illustrated in Fig. 1. Recently, significant advancements have been made in text-to-video generation [6, 10, 11]. Building on these foundational models, many works have explored the use of additional guidance signals, such as images [2, 19, 20], trajectories [36, 40, 43], audio [30, 37], and poses [14, 17, 33, 45] to control the content in the generated video. Among these, trajectory-guided motion control [36, 40] has been gaining increasing attention due to its interactive and user-friendly nature. + +Trajectory-guided motion control requires the target object to follow a specified path accurately, preserving temporal coherence and realism in the generated video. Earlier works IPoke [1] and MCDiff [7] leverage dragging animation only in the limited human domain. DragNUWA [40] is the first to attempt trajectory control in open-domain videos. DragAnything [36] further enhances object-level motion awareness by extracting object entities using corresponding masks and integrating them with trajectory guidance. However, these works constrain the object to follow the trajectory in a 2D image space, without considering potential changes of the object's 6D pose (position and ori + +![](images/c0bd215226c197e72f4d12f3ff27145c8cad971495183c7a1dd054c5e770eecb.jpg) +Complex object entity maintenance under wide-range and steady trajectories. + +![](images/b9b776dc4da00fa3bd9cc8b5c13fa02356778a7006f0c92b37a8335c9dc4ada3.jpg) +Figure 2. The video generation performance of DragAnything vs. Ours under complex trajectories on static and dynamic objects. The yellow dashed line indicates the boat's orientation and helps perceive object rotation, while the orange circle highlights visual defects. + +entation). Consequently, they can perform well in translational motion but fail to control objects when trajectories implicitly involve rotations (see Fig. 2). These limitations arise from two key factors: i) in the collected video datasets, object motions are more commonly translational, while rotations are rare and difficult to annotate automatically, ii) it is inherently ill-posed to infer potential rotations from 2D trajectories and objects in pixel space. + +In this paper, we present a novel pose-aware, trajectory-guided motion control framework, dubbed PoseTraj, which can perceive potential rotations in the trajectory and perform corresponding control (as shown in Fig. 1). The ability to perceive rotations in the trajectory originates from a two-stage pose-aware pre-training pipeline. Specifically, we propose a specialized synthetic dataset PoseTraj-10K, containing over 10,000 videos with diverse trajectories for 2,000 distinct objects involving rotations, paired with the objects' 3D bounding boxes. This dataset circumvents the challenge of accurately estimating trajectories and 3D bounding boxes from open-world videos. With this synthetic dataset, in the first stage, we pre-train the model to generate video frames with accurate 3D bounding boxes following trajectories, without emphasizing appearance details. The introduction of 3D bounding box generation helps the model understand the object location along the driving trajectories and makes it aware of potential rota + +tional changes. In the second stage, we further pre-train the model to focus on refining object appearance with the learned knowledge of object entities and poses. After the two-stage pre-training, the model can perceive potential rotations and maintain stable object entities along the trajectory. Finally, we perform camera-disentangled fine-tuning on a real-world video dataset to enhance the model's generalization to real-world videos. + +Experiments on the VIPSeg benchmark dataset demonstrate that PoseTraj achieves substantial improvements over DragNUWA and DragAnything in both quantitative metrics and qualitative visual performance. As illustrated in Fig. 2, our model provides more accurate motion tracking and superior object reconstruction quality, especially in scenarios involving complex rotations or dynamic objects. + +To summarize, our contributions are: + +- We propose a novel pose-aware pre-training process that employs 3D bounding boxes as intermediate supervision to enhance the model's 6D pose understanding. This capability is further generalized to open-domain video through camera-disentangled fine-tuning. +- We introduce PoseTraj-10K, the first synthetic dataset augmented with both trajectories and 3D bounding boxes, featuring diverse animated objects undergoing wide-range rotational motions along complex paths. +- Experimental results show that PoseTraj outperforms + +SoTA baselines in trajectory-following accuracy and video realism when dragging objects along a path, particularly in scenarios where rotations are included. + +# 2. Related Work + +# 2.1. Controllable Video Generation Model + +Video generation model. While image generation models [25, 25, 28, 42] have achieved impressive results, video generation remains a relatively nascent field. Early works such as NUWA [35] improved video quality by unifying representations across multimodal learning. CogVideo [13] was the first to fine-tune a text-to-video model using pretrained text-to-image weights with minimal additional parameters. Make-A-Video [29] further advanced efficiency by introducing pseudo-3D convolutional layers and temporal attention mechanisms. To mitigate the computational intensity of pixel-space diffusion, recent video diffusion models [3, 11, 14] have adopted latent diffusion methods [26]. With access to larger volumes of high-quality video data, state-of-the-art models such as VideoCrafter [6], Gen-2 [10], SVD [2], and SORA [4] have demonstrated remarkable realism and high-fidelity video generation. + +Trajectory-guided video generation. Recently, trajectory-guided video generation has gained significant attention due to its user-friendly and interactive nature. Early trajectory-based animation methods, such as IPoke [1] and MCDiff [7], were limited to human motion synthesis, relying on skeleton-based conditioning. DragNUWA [40] was the first to introduce dragging control for open-domain videos by fine-tuning the pre-trained image-to-video model SVD with trajectory guidance from large-scale real-world videos. However, maintaining temporal consistency during large motions remains a challenge, often leading to trajectory-following inaccuracies. + +To address this challenge, DragAnything [15] enhanced object identity awareness by leveraging a pre-trained diffusion-based image encoder for explicit entity extraction from object segmentation masks. Tora [43] further improved trajectory guidance by integrating fusion layers into each DiT basic block within a scalable architecture. Other works, such as Direct-A-Video [39] and MotionCtrl [34], aimed to disentangle object motion learning from passive camera movement, enabling more fine-grained control. However, existing models lack strong pose-awareness for handling 3D motion along rotational trajectories. Our work addresses this limitation by explicitly incorporating pose-aware pretraining to enhance the model's understanding of object pose variations. + +# 2.2. Training on Synthetic Dataset + +Due to the scarcity of comprehensive, large-scale real-world data in certain domains, many studies have leveraged syn + +The synthetic datasets to enhance model performance for specific tasks. Focusing on fine-grained 3D reconstruction with diverse materials, OpenMaterial [8] introduced a dataset of 1,001 3D objects, each manually rendered under varying materials and lighting conditions. For image decomposition, the intrinsic diffusion model was trained on the synthetic dataset InteriorVerse [44], which provides high-quality intrinsic rendering images. X-RGB [41] further developed a unified framework for image decomposition and synthesis, leveraging self-rendered synthetic image pairs conditioned on various intrinsic modalities. In controllable video generation, PuppetMaster [16] recently introduced a part-level animation model, rendering high-quality animatable objects sampled from Objaverse [9]. + +For our pose-aware pretraining, we propose a synthetic dataset with complex trajectories and precisely annotated 3D bounding boxes, enabling the animation of diverse objects along various wide-range rotational trajectories. + +# 3. PoseTraj-10K Dataset Construction + +Training a trajectory-guided video generation model requires paired data, i.e., high-quality video with clearly identifiable moving objects and precisely annotated trajectories. + +However, obtaining such data from open-domain videos presents several significant challenges. First, complex motions, such as wide-range rotations, are typically rare in filtered video datasets and difficult to annotate accurately using point-tracking or flow-matching estimators. Moreover, large motions are often tightly coupled with camera movements, making it challenging to disentangle object motion from camera motion. As a result, obtaining precise 6D poses of moving objects from open-domain videos remains infeasible with current pose estimation methods [5, 22]. + +![](images/4553f94bfbd32fddf816beb62b6bd869f841045c69d8bebf45bdea162e6f1e3c.jpg) +Figure 3. Data construction pipeline of our synthetic dataset PoseTraj-10K. The whole configuration, including environment setup, object sampling, and trajectory sampling, is displayed. + +![](images/7f5b9beeba915fd3d55f3ee0f89e7b20305d1f38fac15f773221bf3e328c925d.jpg) +Figure 4. Method overview. Our PoseTraj first utilizes two-stage pose-aware pre-training on our synthetic dataset to obtain 3D-enhanced awareness for rotational trajectory-following capacity and further exploits camera-disentangled finetuning to adapt this ability on open-domain videos. The dashed colorful arrows demonstrate the dedicated data flows during the three training stages, while the black arrows are shared in all stages. All yellow blocks are trainable, while other blocks are frozen. + +We thus construct a synthetic dataset PoseTraj-10K, that offers three key advantages: i) it provides a diverse and comprehensive range of objects, enhancing robust motion understanding, particularly for complex geometric structures and wide-range rotations; ii) it enables the creation of a wide variety of precise trajectories, ensuring motion consistency without interference from camera movements; iii) through controlled rendering, we can effortlessly obtain precise 3D bounding box parameters, which serve as crucial intermediate supervision signals in our pretraining process. + +Fig. 3 illustrates our rendering pipeline implemented in Blender. First, we set up a realistic virtual scene and sampled 2,000 high-quality 3D models from Objaverse [9]. To ensure the quality of the selected models, we employ GPT-4v [21] to filter out overly complex or uncommon objects, followed by a manual selection, resulting in the top 2,000 models that best represent high-quality daily objects. Each model is normalized and assigned a unique random trajectory with randomized rotational angles, trajectory shape, and length. During the animation, the 3D model follows its designated trajectory while maintaining a rotation center, enabling rotation-focused motion for the final rendering. Fig. 3 bottom displays four frames of the rendered video, check out our dataset and the supplementary for more examples and further data construction details. + +# 4. Method + +Given an input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ containing foreground objects and a driving trajectory $\mathbf{tr} = \{(x_i, y_i)\}_{i=1}^L$ drawn as a set of line segment images $\mathbf{I}_{\mathbf{tr}} = \{I_{tr}^i\}_{i=1}^L$ (with last empty padding), our goal is to generate a sequence of animated video frames $\{f_i\}_{i=1}^L$ , where the relevant object accurately follows the trajectory, while maintaining a coherent + +and natural video appearance. + +Fig. 4 presents an overview of our approach, which is built upon a latent stable video diffusion model, SVD [2]. Our base framework mainly consists of three components: a latent diffusion model $\epsilon_{\theta}$ (3D U-Net [27]) to denoise the latent noise, and an encoder $(\mathcal{E})$ - decoder $(\mathcal{D})$ pair to compress video frames into the latent space and reconstruct the denoised latent features back into video frames. Inspired by the success of ControlNet [19] in guiding the pre-trained diffusion model, we design Traj-ControlNet, which consists of a trainable copy of the encoder blocks of SVD. Given a sequence of trajectory images $\mathbf{I}_{\mathrm{tr}}$ , we adopt a trajectory guider - a 3D ConvNet, to encode trajectory features, which is then fed into the Traj-ControlNet to predict the residual features for SVD, enabling the generated video to follow the trajectory. In the following subsections, we detail on the novel technical designs to achieve our goal. + +# 4.1. Two-stage Pose-aware Pretraining + +To enhance the model's ability to perceive potential 6D pose changes within the 2D-pixel space, we select the 3D bounding box (bbox) [32] as the primary 3D supervision signals, as they provide precise positional and pose information. We propose a novel two-stage pose-aware pretraining pipeline that incorporates 3Dbbox generation as an intermediate supervision signal. The two pre-training is on our PoseTraj-10K synthetic dataset. + +Stage One: 3D bbox-guided localization. Instead of separately regressing 3D bbox parameters and generating the object's appearance for each video frame, we simultaneously generate 3D bboxes in the image space along with the objects, as shown in Fig. 4 (red arrows). To this end, given one data example in our dataset, we first render the + +3D bounding box into the pixel space on top of the image to obtain $\mathbf{I}_{bbox}$ (the bottom-left figure in Fig. 4). We apply the same rendering process for other frames to obtain the bounding box augmented video for supervision. This joint generation process helps the model better understand the object's location along the driving trajectory and makes it aware of potential rotational changes. + +Stage Two: object-centric reconstruction. In practical applications, bounding boxes are not required in the generated video. Moreover, in the first stage, with explicit bounding box constraints, the model learns object entities and poses, but the generated object lacks appearance details. Therefore, in the second stage, we exclude bounding box generation and remove 3D bounding box supervision, finetuning the model to refine the object's appearance. + +There are several ways to inject auxiliary 3D information, e.g., inputting depth maps or 3D bounding boxes as conditions. Our injection-by-reconstruction approach offers two key advantages. First, the bounding boxes explicitly enhance continuous 3D awareness by serving as a pixel-level supervision target (see evaluation and supplementary material). Second, these supervisions can be easily removed by simply changing the reconstruction target, thereby mitigating inference-stage mismatches caused by inaccurate estimation of these additional signals. + +# 4.2. Camera-disentangled Finetuning + +After the two-stage pre-training, the model learns to perceive potential rotations while preserving object shape and detailed appearance along a trajectory. However, adapting the pre-trained model to real-world videos remains a key challenge. Unlike the synthetic dataset, where the camera is static, real-world videos often exhibit unpredictable and irregular camera motions. Undistinguished camera motion and object motion can lead to errors in object tracking. Therefore, we introduce camera-disentangled finetuning, which incorporates additional camera motion information into the model to help distinguish active object movement from passive camera motion (orange arrows in Fig. 4). + +Specifically, we annotate the real-world video dataset - VIPSeg [18], with motion trajectories and per-frame camera poses (i.e., extrinsic parameters, denoted as $\{C a m^{i}\}_{i = 1}^{L}$ ). We incorporate camera poses as additional inputs by passing them through an MLP layer and then concatenating them with the trajectory features. The concatenated feature is then fed into a zero-initialized MLP projection layer and further input to the Traj-ControlNet. At inference time, it is hard to provide the camera pose information, we then randomly (50%) drop camera poses in this fine-tuning stage, enabling the model to generate videos without camera information. This approach helps the model differentiate between active object movement and passive camera motion, thereby improving trajectory-following accuracy and en + +hancing robustness across diverse video contexts. + +# 4.3. Training and Inference + +The mean squared error (MSE) loss is commonly used in video generation training, and the objective is to minimize the discrepancy between the predicted noise from $\epsilon_{\theta}$ and the ground truth noise $\epsilon$ for each latent noise $x_{t}$ at timestep $t$ . Both the two pre-training stages and fine-tuning stage leverage the same MSE loss, while the difference lies in the condition $C^i$ : + +$$ +\mathcal {L} _ {\mathrm {M S E}} = \mathbb {E} _ {x _ {t}, \epsilon} \left[ \sum_ {i = 1} ^ {L} \left\| \epsilon - \epsilon_ {\theta} \left(x _ {t}, t, C ^ {i}\right) \right\| _ {2} ^ {2} \right], \tag {1} +$$ + +where $C^i$ is defined as: + +$$ +C ^ {i} = \left\{ \begin{array}{l l} \left\{I _ {t r} ^ {i}, \mathbf {I} _ {b b o x} \right\}, & \text {s t a g e o n e ,} \\ \left\{I _ {t r} ^ {i}, \mathbf {I} \right\}, & \text {s t a g e t w o ,} \\ \left\{I _ {t r} ^ {i}, \mathbf {I}, C a m ^ {i} \right\}, & \text {f i n e t u n i n g .} \end{array} \right. \tag {2} +$$ + +Remeber that, in the first pre-training stage, we used the bbox-augmented video frames as the reconstruction target. + +Spatial enhancement loss. Simply aggregating errors across all frames overlooks the spatial reconstruction accuracy of individual frames, potentially leading to object entity collapse under large rotational motions. Since each short per-frame trajectory segment provides precise localization information within individual frames, we introduce an additional spatial enhancement loss to improve spatial consistency by reconstructing the per-frame image. Specifically, we randomly sample the trajectory of the $j$ -th frame (i.e., $I_{tr}^{j}$ ) and use it as the Traj-ControlNet condition. From its corresponding latent noise $x_{t,j}$ and the initial frame $\mathbf{I}$ ( $\mathbf{I}_{bbox}$ for the first stage), the objective is to train the model to generate the target noise $\epsilon_{j}$ for this specific frame (i.e., single frame image reconstruction), formulated as: + +$$ +\mathcal {L} _ {\mathrm {S P A}} = \left| \right.\left| \right. \epsilon_ {j} - \epsilon_ {\theta} \left(x _ {t, j}, t, C ^ {j}\right)\left. \right)\left. \right|\left. \right| _ {2} ^ {2}, j \in (1 \sim L). \tag {3} +$$ + +During backpropagation, only the spatial layers are updated, ensuring the refinement focuses on per-frame spatial accuracy. + +Overall, the training loss function of our framework for each stage is defined as: + +$$ +\mathcal {L} _ {\text {a l l}} = \mathcal {L} _ {\text {M S E}} + \lambda_ {\text {S P A}} \mathcal {L} _ {\text {S P A}}, \tag {4} +$$ + +where $\lambda_{\mathrm{SPA}}$ is a hyperparameter. + +Inference time video generation. During inference, given an input image, users freely draw a custom trajectory, which can be either rotational or translational, our model will generate corresponding pose-aware videos. Note that our approach supports more than one trajectory, see Figs. 2 and 6 for examples. In addition, users can control the camera motion behavior by providing camera poses extracted from existing videos. + +![](images/990f58f3dfd16340ba09be3ba509de14b9e222a27b148d0ae34554c66ab5773f.jpg) +Figure 5. Visual comparison. Given the same trajectory and the initial image, our method produces plausible video frames containing the moving object following the given trajectory. In contrast, DrafNUWA and DragAnything either introduce unexpected camera motions or fail to maintain the object entity, causing severe collapse. + +Table 1. Quantitative comparison between our approach and the SoTA methods. + +
MethodsVIPSeg 256 × 256VIPSeg 320 × 576DAVIS 320 × 576
ObjMC ↓FID ↓FVD ↓ObjMC ↓FID ↓FVD ↓ObjMC ↓FID ↓FVD ↓
DragNUWA 1.5 [40]173.6163.39449.59133.0541.88289.1574.0754.10952.87
DragAnything [36]100.2361.69410.7091.1239.29275.9347.0150.83771.78
Ours87.5646.60384.4177.4838.41267.3329.9251.48729.16
+ +# 5. Experiments and Evaluations + +Implementation details. Our base model is SVD [2], trained to generate 14 frames given an initial frame at a resolution of $320 \times 576$ pixels. All experiments are conducted using a single Tesla A100 GPU with a batch size of 1. More implementation details are in the supplementary. + +Synthetic dataset. For our synthetic dataset used in training, we generated 10,000 videos by sampling 5 random trajectories for each of 2,000 high-quality objects and rendering them in Blender through the Cycles engine at 5 fps. For evaluation, we create an additional trajectory for a subset of 200 sampled objects, which serves as our test set. During trajectory annotation, we augment the trajectory placement through a trajectory sampler (see the supplementary). Both training and evaluation on the synthetic dataset are + +conducted in the resolution of $320 \times 576$ pixels with 14 frames per sequence. + +Real-world dataset. For real-world finetuning, we adopted the VIPSeg dataset [18], following DragAnything. VIPSeg is a large-scale video segmentation dataset comprising over 3,000 high-quality videos with annotated object masks. We utilized its training set for finetuning and its validation set for evaluation, employing the same trajectory extraction pipeline as in DragAnything (see supplementary). + +To further evaluate the generalization capability of our approach, we additionally select the DAVIS dataset [24], a high-quality video segmentation benchmark, for out-of-distribution (OOD) evaluation. Note that, we only annotate the trajectory for this dataset for OOD evaluation, and at inference, camera poses are not used. + +Evaluation metrics. We follow established evaluation met- + +rics from previous trajectory-guided studies [23, 36, 40], assessing performance across two key aspects: + +- Controlling accuracy. To quantify the precision of trajectory-guided dragging, we use the ObjMC metric, which computes the mean squared error between the temporal motion tendency of generated samples and that of the ground truth videos. +- Generation quality. We employ the Fréchet Inception Distance (FID) [12] to assess the single-frame spatial quality and the Fréchet Video Distance (FVD) [31] to measure temporal consistency. + +# 5.1. Comparison + +Competitors. We primarily evaluate our model's performance against two recent trajectory-guided video generation models: DragNUWA 1.5 [40] and DragAnything [36]. DragNUWA is the first method designed for open-domain video dragging. It utilizes optical flow estimated by UniMatch [38] as trajectory guidance, combining these trajectories with text and image conditions to enable motion-controllable fine-tuning. Built upon DragNUWA, DragAnything enhances object-level animation by extracting entity feature maps from instance masks and integrating them into ControlNet for further fine-tuning. To ensure fair comparisons, the same trajectory is converted into each model's respective trajectory representation (e.g., the trajectory point set or trajectory segment images). Additionally, since DragAnything reports metric performance only at a resolution of $256 \times 256$ , we evaluate our model at both the original $320 \times 576$ training resolution and the downsampled $256 \times 256$ resolution to ensure a fair comparison. Due to the unavailability of public model weights for Tora and MotionCtrl with SVD pretraining, we exclude these models from our comparison. + +Quantitative Results. As shown in Tab. 1, PoseTraj outperforms competing methods across both resolutions in all trajectory-specific and video quality metrics, with particularly strong performance in lower-resolution synthesis. In terms of trajectory accuracy, our model achieves a significant reduction in error than DragAnything, with improvements of $13\%$ , $15\%$ , and $36\%$ for lower and standard resolutions on the VIPSeg dataset and standard resolution on the DAVIS dataset, respectively. These results demonstrate the effectiveness of our pose-aware pretraining approach, which leverages high-precision synthetic data pairs with 3D bounding boxes and incorporates a camera-disentanglement module to enhance the model's understanding of object-oriented motion and passive camera movement. + +For single-frame reconstruction performance, measured by FID, our model outperforms DragAnything by a significant margin, achieving an improvement of 15.01 at the downsampled $256 \times 256$ resolution and 0.88 at the standard $320 \times 576$ resolution on the VIPSeg dataset. Fur + +![](images/21147d34b8964b2b77f7074bcc144407ae68552462ee37f1f45a8c82dcf609f9.jpg) +Figure 6. Transitional samples on DAVIS dataset. + +thermore, our model demonstrates improved temporal consistency across frames, achieving substantial reductions in FVD scores: 26.29, 8.6, and 42.62 across the three evaluation settings. These results highlight our model's significant advancements in trajectory-guided motion control, achieving precise dragging-following and reliable motion. Notably, our approach shows strong performance in traditional in-distribution scenarios while exhibiting even greater superiority in out-of-distribution (OOD) settings (see the evaluation on DAVIS in Tab. 1). + +Qualitative Results. Visual comparisons are displayed in Fig. 5, where PoseTraj achieves superior performance in generating 3D and pose-aligned rotational motions that accurately follow 2D trajectories with precise pixel-level trajectory adherence. Furthermore, our model demonstrates enhanced capability in reliably reconstructing occluded object regions, even under changing 6D poses, such as rotational car doors. In contrast, DragNUWA primarily generates moving camera perspectives while keeping object poses static, resulting in poor rotational alignment. DragAnything produces unstable and mismatched rotational motions, where dragged pixels progressively drift across frames, potentially leading to entity collapse. This comparison highlights the effectiveness of our approach in generating more accurate, stable, and realistic motion sequences, particularly when handling complex trajectories. + +Additionally, in Fig. 6, we show visual results on OOD examples, where the trajectory is mostly translational instead of rotational. Our generated videos are superior to DragAnything. + +![](images/01a88f571a72a3edced5ee8d48117ae602eb08535215bad1eee8ecb975d51339.jpg) +Figure 7. The voting results of the user evaluation. + +User Evaluation. Other than the visual and statistical evaluations, we conduct a subjective user evaluation, where par + +Table 2. Ablation results on the real-world video test set with several model variants. + +
Methods320 × 576
ObjMC ↓FID ↓FVD ↓
Full method77.4838.41267.33
Nobbox stage81.3641.90275.40
No pretrain145.7242.62486.84
No Cam-disen83.2239.71279.15
No SPA-loss137.2639.79436.56
+ +Table 3. Comparison results on the synthetic validation set of our designed ablation studies on the pretraining stage. + +
MethodsSynthetic Dataset Val
ObjMC ↓FID ↓FVD ↓
Full method0.19647.18187.12
Nobbox stage0.26650.30223.74
No SPA-loss0.26546.72197.68
+ +ticipants rate generated videos (20 sampled videos from the VIPSeg test set and manually annotated wild images) based on two criteria: overall motion accuracy and visual quality. We have invited seven experienced experts on the video generation topic and asked them to rate the videos. + +As shown in Fig. 7, our approach is favorably voted, receiving $43\%$ more votes in trajectory-following accuracy and $39\%$ more votes in video quality. These findings are consistent with the superior performance observed in quantitative metrics and qualitative visualizations. + +# 5.2. Ablation Study + +Pose-aware pretraining. To evaluate the impact of our pretraining strategy, we propose two experimental variants: (1) training the model without 3D-aware pertaining (denoted as 'No pretrain'), and (2) training the model using only the second-stage pretraining while omitting 3D bounding box supervision (denoted as 'No bbox stage'). + +As shown in Tabs. 2 and 3, excluding 3D bounding box supervision during pretraining leads to a degradation in performance across all evaluation metrics on both the synthetic and open-domain datasets. Notably, in Tab. 2, compared with 'No pretraining', the absence of 3D bounding box pretraining has minimal impact on the metric. This is because 3D bounding boxes mainly affect object pose localization, and inaccurate poses do not significantly degrade motion accuracy or quality. However, the visual results in Fig. 8 clearly demonstrate the defects without bounding box pretraining, i.e., there is pronounced object collapse, exacerbating trajectory misalignment. More visual comparisons can be found in the supplementary material. Furthermore, omitting the two-stage pretraining approach results in even greater performance degradation, leading to a substantial drop in trajectory-following capability. + +Spatial enhancement loss. Compared to standard video generation tasks, achieving both precise trajectory control and spatial visual consistency is significantly more chal + +![](images/af3c31a3e418b98a255d48c9332257dce4c6bdd518f848a70e3af9c1b1e5e903.jpg) +Figure 8. Visualization of ablation results in synthetic dataset. + +lenging, especially when the motion involves complex dynamics or large-angle rotations. Consequently, integrating spatial enhancement loss and formulating image-to-image prediction as a sub-task during training becomes essential. As demonstrated in Tabs. 2 and 3 and Fig. 8, models trained without spatial loss (denoted as 'No SPA-loss') show considerable degradation in trajectory alignment reliability and generation quality, with frequent blurring observed for fast-moving objects. + +Camera disentanglement. Considering the unexpected influence of camera motions, we incorporate camera disentangled fine-tuning to enhance the model's trajectory, controlling awareness of generating naturally activated object motion with passive camera movements. We evaluate the impact of this design by discarding camera poses (denoted as 'No Cam-disen'). Experimental results (Tab. 2) indicate that models without camera disentangling exhibit lower object motion-matching accuracy and a corresponding decrease in the FVD metric. This inaccuracy arises from the model's uncertainty regarding how to generate aligned passive camera motions and results in a reduced ability to handle trajectory variations involving background motion. + +We show more visual results for ablation experiments in the supplementary. + +# 6. Conclusion + +In this paper, we have presented PoseTraj, a novel approach for pose-aware trajectory control in video generation. We propose a two-stage pre-training strategy to enable the accurate trajectory-following ability, while explicitly exploiting 3D bounding boxes as a supervision signal to perceive the object's potential 6D pose changes along the driving trajectory. A new dataset, PoseTraj-10K, is constructed to achieve our goal. Our approach outperforms state-of-the-art methods quantitatively and qualitatively. We hope that the 3D bounding box signal and the injection-by-reconstruction design can inspire future research in controllable video generation or relevant topics. + +Acknowledgment. LB completed part of this work during his Master's study at the University of Edinburgh. CJ was supported by a gift from Adobe. + +# References + +[1] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14707-14717, 2021. 1, 3 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3, 4, 6 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3 +[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 3 +[5] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. European Conference on Computer Vision (ECCV), 2024. 3 +[6] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation, 2023. 1, 3 +[7] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 1, 3 +[8] Zheng Dang, Jialu Huang, Fei Wang, and Mathieu Salzmann. Openmaterial: A comprehensive dataset of complex materials for 3d reconstruction. arXiv preprint arXiv:2406.08894, 2024. 3 +[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 3, 4 +[10] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 3 +[11] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 1, 3 +[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a + +two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[13] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3 +[14] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 1, 3 +[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3 +[16] Ruining Li, Chuanxia Zheng, Christian Rupprecht, and Andrea Vedaldi. Puppet-master: Scaling interactive video generation as a motion prior for part-level dynamics. arXiv preprint arXiv:2408.04631, 2024. 3 +[17] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024. 1 +[18] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 5, 6 +[19] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1, 4 +[20] Haomiao Ni, Bernhard Egger, Suhas Lohit, Anoop Cherian, Ye Wang, Toshiaki Koike-Akino, Sharon X Huang, and Tim K Marks. Ti2v-zero: Zero-shot image conditioning for text-to-video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9015-9025, 2024. 1 +[21] OpenAI. GPT (v4) [Large language model]. https://chat.openai.com, 2024.4 +[22] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In European Conference on Computer Vision, pages 163-182. Springer, 2025. 3 +[23] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 7 +[24] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference + +on computer vision and pattern recognition, pages 724-732, 2016. 6 +[25] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4 +[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3 +[29] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3 +[30] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive - generating expressive portrait videos with audio2video diffusion model under weak conditions, 2024. 1 +[31] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7 +[32] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis, 2024. 4 +[33] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 1 +[34] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. arXiv preprint arXiv:2312.03641, 2023. 3 +[35] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In European conference on computer vision, pages 720-736. Springer, 2022. 3 +[36] Wejia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control + +for anything using entity representation. arXiv preprint arXiv:2403.07420, 2024.1,6,7 +[37] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1 +[38] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In CVPR, 2023. 7 +[39] Shiuyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3 +[40] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1, 3, 6, 7 +[41] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgbx: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[42] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to $360^{\circ}$ panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[43] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 1, 3 +[44] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 3 +[45] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 1 \ No newline at end of file diff --git a/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/images.zip b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1519d40296933f6059a426ff58dec9dd329875f2 --- /dev/null +++ b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:279d93465ea9f2e015d5ed7b0444a18fafcc175b5733990af39dc07ae0cb4c4a +size 763015 diff --git a/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/layout.json b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1843d5879ec799a6f4b42ae4b126545063cd81cd --- /dev/null +++ b/2025/PoseTraj_ Pose-Aware Trajectory Control in Video Diffusion/layout.json @@ -0,0 +1,7958 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "text", + "content": "PoseTraj: Pose-Aware Trajectory Control in Video Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 142, + 217, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 142, + 217, + 158 + ], + "spans": [ + { + "bbox": [ + 157, + 142, + 217, + 158 + ], + "type": "text", + "content": "Longbin Ji" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 230, + 143, + 287, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 143, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 230, + 143, + 287, + 158 + ], + "type": "text", + "content": "Lei Zhong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 300, + 143, + 365, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 143, + 365, + 157 + ], + "spans": [ + { + "bbox": [ + 300, + 143, + 365, + 157 + ], + "type": "text", + "content": "Pengfei Wei2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 378, + 143, + 447, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 143, + 447, + 158 + ], + "spans": [ + { + "bbox": [ + 378, + 143, + 447, + 158 + ], + "type": "text", + "content": "Changjian Li" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 149, + 162, + 458, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 162, + 458, + 178 + ], + "spans": [ + { + "bbox": [ + 149, + 162, + 458, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 162, + 458, + 178 + ], + "type": "text", + "content": "University of Edinburgh " + }, + { + "bbox": [ + 149, + 162, + 458, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 162, + 458, + 178 + ], + "type": "text", + "content": "Nanyang Technological University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 215, + 180, + 392, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 180, + 392, + 194 + ], + "spans": [ + { + "bbox": [ + 215, + 180, + 392, + 194 + ], + "type": "text", + "content": "https://robingg1.github.io/Pose-Traj/" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 56, + 204, + 138, + 261 + ], + "blocks": [ + { + "bbox": [ + 56, + 204, + 138, + 261 + ], + "lines": [ + { + "bbox": [ + 56, + 204, + 138, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 138, + 261 + ], + "type": "image", + "image_path": "d8a515e4cadaeecb99fcec7f00238c427ecc970a822c08383060ce04c9ac172b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 139, + 204, + 221, + 261 + ], + "blocks": [ + { + "bbox": [ + 139, + 204, + 221, + 261 + ], + "lines": [ + { + "bbox": [ + 139, + 204, + 221, + 261 + ], + "spans": [ + { + "bbox": [ + 139, + 204, + 221, + 261 + ], + "type": "image", + "image_path": "fd1d1ee36163605c4cb8ca46b61ce448b39b3d0d5a8cf2c95c8c81f4edafc03d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 223, + 204, + 304, + 261 + ], + "blocks": [ + { + "bbox": [ + 223, + 204, + 304, + 261 + ], + "lines": [ + { + "bbox": [ + 223, + 204, + 304, + 261 + ], + "spans": [ + { + "bbox": [ + 223, + 204, + 304, + 261 + ], + "type": "image", + "image_path": "dadc0f7b07c13d4558f45dfdf9adc360f5b0f26c6ee27e4b8a8a27b4b32ad7da.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 306, + 204, + 388, + 261 + ], + "blocks": [ + { + "bbox": [ + 306, + 204, + 388, + 261 + ], + "lines": [ + { + "bbox": [ + 306, + 204, + 388, + 261 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 388, + 261 + ], + "type": "image", + "image_path": "8f11187bafe2f608908e87db374161a0148f3c8ecb831d71f8192232c6722750.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 389, + 204, + 471, + 261 + ], + "blocks": [ + { + "bbox": [ + 389, + 204, + 471, + 261 + ], + "lines": [ + { + "bbox": [ + 389, + 204, + 471, + 261 + ], + "spans": [ + { + "bbox": [ + 389, + 204, + 471, + 261 + ], + "type": "image", + "image_path": "e82b3966821e1741d0120617a1b62af3b52db3e02f2465e2034f03c75768e36d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 472, + 204, + 553, + 261 + ], + "blocks": [ + { + "bbox": [ + 472, + 204, + 553, + 261 + ], + "lines": [ + { + "bbox": [ + 472, + 204, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 472, + 204, + 553, + 261 + ], + "type": "image", + "image_path": "973d4b95a84b9f9ccb2bf02f7c5696075313ccfbe7712952ca954f497c3d30bc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 56, + 262, + 138, + 319 + ], + "blocks": [ + { + "bbox": [ + 56, + 262, + 138, + 319 + ], + "lines": [ + { + "bbox": [ + 56, + 262, + 138, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 138, + 319 + ], + "type": "image", + "image_path": "2d0d2b075072be0b2d0b75f5043287b03598af62cca53bc8efceba395699802f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 139, + 262, + 221, + 319 + ], + "blocks": [ + { + "bbox": [ + 139, + 262, + 221, + 319 + ], + "lines": [ + { + "bbox": [ + 139, + 262, + 221, + 319 + ], + "spans": [ + { + "bbox": [ + 139, + 262, + 221, + 319 + ], + "type": "image", + "image_path": "05eebfe5859eb591bd23f01c45b03730a7585106fd22295e0dd329d25dc28858.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 223, + 262, + 304, + 319 + ], + "blocks": [ + { + "bbox": [ + 223, + 262, + 304, + 319 + ], + "lines": [ + { + "bbox": [ + 223, + 262, + 304, + 319 + ], + "spans": [ + { + "bbox": [ + 223, + 262, + 304, + 319 + ], + "type": "image", + "image_path": "df2f795b5570ca548141b719ef89082fa0845172a80edfffefb60946db015f5f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 306, + 262, + 388, + 319 + ], + "blocks": [ + { + "bbox": [ + 306, + 262, + 388, + 319 + ], + "lines": [ + { + "bbox": [ + 306, + 262, + 388, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 388, + 319 + ], + "type": "image", + "image_path": "f5f886ee2e3a6e5b113188950f74222c0e6fdad0ab58fa2cbe15372f4cf7d3b3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 389, + 262, + 471, + 319 + ], + "blocks": [ + { + "bbox": [ + 389, + 262, + 471, + 319 + ], + "lines": [ + { + "bbox": [ + 389, + 262, + 471, + 319 + ], + "spans": [ + { + "bbox": [ + 389, + 262, + 471, + 319 + ], + "type": "image", + "image_path": "9d80524fb2a7e9420c2df122cc5cd19ea11749c7cf02ca2bf6e9147311cbff3b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 472, + 262, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 472, + 262, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 472, + 262, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 472, + 262, + 553, + 319 + ], + "type": "image", + "image_path": "0867242021e791f963488ab35f4643b800fdb9fb5d0fbd15b1d45aa6e3c32e0f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 56, + 321, + 138, + 378 + ], + "blocks": [ + { + "bbox": [ + 56, + 321, + 138, + 378 + ], + "lines": [ + { + "bbox": [ + 56, + 321, + 138, + 378 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 138, + 378 + ], + "type": "image", + "image_path": "f83578d8b8e36c9cc2dd0f54a3d81c79d91d504fa94d5c7d694c91cc32e8361d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 380, + 116, + 390 + ], + "lines": [ + { + "bbox": [ + 80, + 380, + 116, + 390 + ], + "spans": [ + { + "bbox": [ + 80, + 380, + 116, + 390 + ], + "type": "text", + "content": "Trajectory" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 139, + 321, + 221, + 377 + ], + "blocks": [ + { + "bbox": [ + 139, + 321, + 221, + 377 + ], + "lines": [ + { + "bbox": [ + 139, + 321, + 221, + 377 + ], + "spans": [ + { + "bbox": [ + 139, + 321, + 221, + 377 + ], + "type": "image", + "image_path": "dbd5ab5225201a46af04dae3e35fb9c2b615c0af10c2ae872bfca6b390dc24cd.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 380, + 195, + 389 + ], + "lines": [ + { + "bbox": [ + 166, + 380, + 195, + 389 + ], + "spans": [ + { + "bbox": [ + 166, + 380, + 195, + 389 + ], + "type": "text", + "content": "Frame 1" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 223, + 321, + 304, + 377 + ], + "blocks": [ + { + "bbox": [ + 223, + 321, + 304, + 377 + ], + "lines": [ + { + "bbox": [ + 223, + 321, + 304, + 377 + ], + "spans": [ + { + "bbox": [ + 223, + 321, + 304, + 377 + ], + "type": "image", + "image_path": "ccca32a98fa579b7343b0e064fcadf23b941f97927b0a8d92001713687650ff9.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 380, + 279, + 389 + ], + "lines": [ + { + "bbox": [ + 249, + 380, + 279, + 389 + ], + "spans": [ + { + "bbox": [ + 249, + 380, + 279, + 389 + ], + "type": "text", + "content": "Frame 4" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 396, + 546, + 408 + ], + "lines": [ + { + "bbox": [ + 62, + 396, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 546, + 408 + ], + "type": "text", + "content": "Figure 1. PoseTraj produces plausible dragging videos, where objects follow a rotational trajectory with awareness of changing poses." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 306, + 321, + 388, + 377 + ], + "blocks": [ + { + "bbox": [ + 306, + 321, + 388, + 377 + ], + "lines": [ + { + "bbox": [ + 306, + 321, + 388, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 321, + 388, + 377 + ], + "type": "image", + "image_path": "f196f9ab28a7f0746416a91217697046f2404421da17784e619c3d8117e5593d.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 380, + 361, + 389 + ], + "lines": [ + { + "bbox": [ + 332, + 380, + 361, + 389 + ], + "spans": [ + { + "bbox": [ + 332, + 380, + 361, + 389 + ], + "type": "text", + "content": "Frame 7" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 389, + 321, + 471, + 377 + ], + "blocks": [ + { + "bbox": [ + 389, + 321, + 471, + 377 + ], + "lines": [ + { + "bbox": [ + 389, + 321, + 471, + 377 + ], + "spans": [ + { + "bbox": [ + 389, + 321, + 471, + 377 + ], + "type": "image", + "image_path": "9e66b3ab2abc4b3d9992d10662124b1c16f4d1e21dd61253a98bde45e16c0178.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 380, + 447, + 389 + ], + "lines": [ + { + "bbox": [ + 413, + 380, + 447, + 389 + ], + "spans": [ + { + "bbox": [ + 413, + 380, + 447, + 389 + ], + "type": "text", + "content": "Frame 10" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 472, + 321, + 553, + 377 + ], + "blocks": [ + { + "bbox": [ + 472, + 321, + 553, + 377 + ], + "lines": [ + { + "bbox": [ + 472, + 321, + 553, + 377 + ], + "spans": [ + { + "bbox": [ + 472, + 321, + 553, + 377 + ], + "type": "image", + "image_path": "cd5ea9fa9cc5d2a1e0d2ab18e6cb9f7f676d74f0447c2ed6e9cbc4d4ba2e17a0.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 380, + 529, + 389 + ], + "lines": [ + { + "bbox": [ + 496, + 380, + 529, + 389 + ], + "spans": [ + { + "bbox": [ + 496, + 380, + 529, + 389 + ], + "type": "text", + "content": "Frame 14" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "bbox": [ + 152, + 417, + 200, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 417, + 200, + 429 + ], + "spans": [ + { + "bbox": [ + 152, + 417, + 200, + 429 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 54, + 448, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 448, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 297, + 700 + ], + "type": "text", + "content": "Recent advancements in trajectory-guided video generation have achieved notable progress. However, existing models still face challenges in generating object motions with potentially changing 6D poses under wide-range rotations, due to limited 3D understanding. To address this problem, we introduce PoseTraj, a pose-aware video dragging model for generating 3D-aligned motion from 2D trajectories. Our method adopts a novel two-stage pose-aware pretraining framework, improving 3D understanding across diverse trajectories. Specifically, we propose a large-scale synthetic dataset PoseTraj-10K, containing 10k videos of objects following rotational trajectories, and enhance the model perception of object pose changes by incorporating 3D bounding boxes as intermediate supervision signals. Following this, we fine-tune the trajectory-controlling module on real-world videos, applying an additional camera-disentanglement module to further refine motion accuracy. Experiments on various benchmark datasets demonstrate that our method not only excels in 3D pose-aligned dragging for rotational trajectories but also outperforms existing baselines in trajectory accuracy and video quality." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 315, + 417, + 394, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 417, + 394, + 429 + ], + "spans": [ + { + "bbox": [ + 315, + 417, + 394, + 429 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 312, + 437, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 555, + 569 + ], + "type": "text", + "content": "We address the problem of controlling video generation with trajectories involving potential object rotations, as illustrated in Fig. 1. Recently, significant advancements have been made in text-to-video generation [6, 10, 11]. Building on these foundational models, many works have explored the use of additional guidance signals, such as images [2, 19, 20], trajectories [36, 40, 43], audio [30, 37], and poses [14, 17, 33, 45] to control the content in the generated video. Among these, trajectory-guided motion control [36, 40] has been gaining increasing attention due to its interactive and user-friendly nature." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "type": "text", + "content": "Trajectory-guided motion control requires the target object to follow a specified path accurately, preserving temporal coherence and realism in the generated video. Earlier works IPoke [1] and MCDiff [7] leverage dragging animation only in the limited human domain. DragNUWA [40] is the first to attempt trajectory control in open-domain videos. DragAnything [36] further enhances object-level motion awareness by extracting object entities using corresponding masks and integrating them with trajectory guidance. However, these works constrain the object to follow the trajectory in a 2D image space, without considering potential changes of the object's 6D pose (position and ori" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22776" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 86, + 541, + 217 + ], + "blocks": [ + { + "bbox": [ + 71, + 86, + 541, + 217 + ], + "lines": [ + { + "bbox": [ + 71, + 86, + 541, + 217 + ], + "spans": [ + { + "bbox": [ + 71, + 86, + 541, + 217 + ], + "type": "image", + "image_path": "c0bd215226c197e72f4d12f3ff27145c8cad971495183c7a1dd054c5e770eecb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 218, + 449, + 229 + ], + "lines": [ + { + "bbox": [ + 161, + 218, + 449, + 229 + ], + "spans": [ + { + "bbox": [ + 161, + 218, + 449, + 229 + ], + "type": "text", + "content": "Complex object entity maintenance under wide-range and steady trajectories." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 71, + 231, + 541, + 360 + ], + "blocks": [ + { + "bbox": [ + 71, + 231, + 541, + 360 + ], + "lines": [ + { + "bbox": [ + 71, + 231, + 541, + 360 + ], + "spans": [ + { + "bbox": [ + 71, + 231, + 541, + 360 + ], + "type": "image", + "image_path": "b9b776dc4da00fa3bd9cc8b5c13fa02356778a7006f0c92b37a8335c9dc4ada3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 362, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 362, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 555, + 384 + ], + "type": "text", + "content": "Figure 2. The video generation performance of DragAnything vs. Ours under complex trajectories on static and dynamic objects. The yellow dashed line indicates the boat's orientation and helps perceive object rotation, while the orange circle highlights visual defects." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 395, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 295, + 491 + ], + "type": "text", + "content": "entation). Consequently, they can perform well in translational motion but fail to control objects when trajectories implicitly involve rotations (see Fig. 2). These limitations arise from two key factors: i) in the collected video datasets, object motions are more commonly translational, while rotations are rare and difficult to annotate automatically, ii) it is inherently ill-posed to infer potential rotations from 2D trajectories and objects in pixel space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we present a novel pose-aware, trajectory-guided motion control framework, dubbed PoseTraj, which can perceive potential rotations in the trajectory and perform corresponding control (as shown in Fig. 1). The ability to perceive rotations in the trajectory originates from a two-stage pose-aware pre-training pipeline. Specifically, we propose a specialized synthetic dataset PoseTraj-10K, containing over 10,000 videos with diverse trajectories for 2,000 distinct objects involving rotations, paired with the objects' 3D bounding boxes. This dataset circumvents the challenge of accurately estimating trajectories and 3D bounding boxes from open-world videos. With this synthetic dataset, in the first stage, we pre-train the model to generate video frames with accurate 3D bounding boxes following trajectories, without emphasizing appearance details. The introduction of 3D bounding box generation helps the model understand the object location along the driving trajectories and makes it aware of potential rota" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 395, + 554, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 554, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 554, + 490 + ], + "type": "text", + "content": "tional changes. In the second stage, we further pre-train the model to focus on refining object appearance with the learned knowledge of object entities and poses. After the two-stage pre-training, the model can perceive potential rotations and maintain stable object entities along the trajectory. Finally, we perform camera-disentangled fine-tuning on a real-world video dataset to enhance the model's generalization to real-world videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 493, + 554, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 554, + 577 + ], + "type": "text", + "content": "Experiments on the VIPSeg benchmark dataset demonstrate that PoseTraj achieves substantial improvements over DragNUWA and DragAnything in both quantitative metrics and qualitative visual performance. As illustrated in Fig. 2, our model provides more accurate motion tracking and superior object reconstruction quality, especially in scenarios involving complex rotations or dynamic objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 579, + 476, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 579, + 476, + 590 + ], + "spans": [ + { + "bbox": [ + 325, + 579, + 476, + 590 + ], + "type": "text", + "content": "To summarize, our contributions are:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 594, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 653 + ], + "type": "text", + "content": "- We propose a novel pose-aware pre-training process that employs 3D bounding boxes as intermediate supervision to enhance the model's 6D pose understanding. This capability is further generalized to open-domain video through camera-disentangled fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 553, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 701 + ], + "type": "text", + "content": "- We introduce PoseTraj-10K, the first synthetic dataset augmented with both trajectories and 3D bounding boxes, featuring diverse animated objects undergoing wide-range rotational motions along complex paths." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "- Experimental results show that PoseTraj outperforms" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 157, + 74, + 446, + 85 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 74, + 446, + 85 + ], + "spans": [ + { + "bbox": [ + 157, + 74, + 446, + 85 + ], + "type": "text", + "content": "Accurate and 3D-aware motion generation under a sharp rotational trajectory." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22777" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 64, + 72, + 295, + 109 + ], + "type": "text", + "content": "SoTA baselines in trajectory-following accuracy and video realism when dragging objects along a path, particularly in scenarios where rotations are included." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 142, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 142, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 142, + 133 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 255, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 255, + 153 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 255, + 153 + ], + "type": "text", + "content": "2.1. Controllable Video Generation Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 159, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 350 + ], + "type": "text", + "content": "Video generation model. While image generation models [25, 25, 28, 42] have achieved impressive results, video generation remains a relatively nascent field. Early works such as NUWA [35] improved video quality by unifying representations across multimodal learning. CogVideo [13] was the first to fine-tune a text-to-video model using pretrained text-to-image weights with minimal additional parameters. Make-A-Video [29] further advanced efficiency by introducing pseudo-3D convolutional layers and temporal attention mechanisms. To mitigate the computational intensity of pixel-space diffusion, recent video diffusion models [3, 11, 14] have adopted latent diffusion methods [26]. With access to larger volumes of high-quality video data, state-of-the-art models such as VideoCrafter [6], Gen-2 [10], SVD [2], and SORA [4] have demonstrated remarkable realism and high-fidelity video generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 351, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 495 + ], + "type": "text", + "content": "Trajectory-guided video generation. Recently, trajectory-guided video generation has gained significant attention due to its user-friendly and interactive nature. Early trajectory-based animation methods, such as IPoke [1] and MCDiff [7], were limited to human motion synthesis, relying on skeleton-based conditioning. DragNUWA [40] was the first to introduce dragging control for open-domain videos by fine-tuning the pre-trained image-to-video model SVD with trajectory guidance from large-scale real-world videos. However, maintaining temporal consistency during large motions remains a challenge, often leading to trajectory-following inaccuracies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 662 + ], + "type": "text", + "content": "To address this challenge, DragAnything [15] enhanced object identity awareness by leveraging a pre-trained diffusion-based image encoder for explicit entity extraction from object segmentation masks. Tora [43] further improved trajectory guidance by integrating fusion layers into each DiT basic block within a scalable architecture. Other works, such as Direct-A-Video [39] and MotionCtrl [34], aimed to disentangle object motion learning from passive camera movement, enabling more fine-grained control. However, existing models lack strong pose-awareness for handling 3D motion along rotational trajectories. Our work addresses this limitation by explicitly incorporating pose-aware pretraining to enhance the model's understanding of object pose variations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 218, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 218, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 218, + 685 + ], + "type": "text", + "content": "2.2. Training on Synthetic Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Due to the scarcity of comprehensive, large-scale real-world data in certain domains, many studies have leveraged syn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "The synthetic datasets to enhance model performance for specific tasks. Focusing on fine-grained 3D reconstruction with diverse materials, OpenMaterial [8] introduced a dataset of 1,001 3D objects, each manually rendered under varying materials and lighting conditions. For image decomposition, the intrinsic diffusion model was trained on the synthetic dataset InteriorVerse [44], which provides high-quality intrinsic rendering images. X-RGB [41] further developed a unified framework for image decomposition and synthesis, leveraging self-rendered synthetic image pairs conditioned on various intrinsic modalities. In controllable video generation, PuppetMaster [16] recently introduced a part-level animation model, rendering high-quality animatable objects sampled from Objaverse [9]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 240, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 289 + ], + "type": "text", + "content": "For our pose-aware pretraining, we propose a synthetic dataset with complex trajectories and precisely annotated 3D bounding boxes, enabling the animation of diverse objects along various wide-range rotational trajectories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 301, + 509, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 509, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 509, + 316 + ], + "type": "text", + "content": "3. PoseTraj-10K Dataset Construction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 323, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 358 + ], + "type": "text", + "content": "Training a trajectory-guided video generation model requires paired data, i.e., high-quality video with clearly identifiable moving objects and precisely annotated trajectories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 358, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 479 + ], + "type": "text", + "content": "However, obtaining such data from open-domain videos presents several significant challenges. First, complex motions, such as wide-range rotations, are typically rare in filtered video datasets and difficult to annotate accurately using point-tracking or flow-matching estimators. Moreover, large motions are often tightly coupled with camera movements, making it challenging to disentangle object motion from camera motion. As a result, obtaining precise 6D poses of moving objects from open-domain videos remains infeasible with current pose estimation methods [5, 22]." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 319, + 490, + 545, + 674 + ], + "blocks": [ + { + "bbox": [ + 319, + 490, + 545, + 674 + ], + "lines": [ + { + "bbox": [ + 319, + 490, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 319, + 490, + 545, + 674 + ], + "type": "image", + "image_path": "4553f94bfbd32fddf816beb62b6bd869f841045c69d8bebf45bdea162e6f1e3c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 677, + 553, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 711 + ], + "type": "text", + "content": "Figure 3. Data construction pipeline of our synthetic dataset PoseTraj-10K. The whole configuration, including environment setup, object sampling, and trajectory sampling, is displayed." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22778" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 545, + 255 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 545, + 255 + ], + "type": "image", + "image_path": "7f5b9beeba915fd3d55f3ee0f89e7b20305d1f38fac15f773221bf3e328c925d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 259, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 54, + 259, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 54, + 259, + 555, + 304 + ], + "type": "text", + "content": "Figure 4. Method overview. Our PoseTraj first utilizes two-stage pose-aware pre-training on our synthetic dataset to obtain 3D-enhanced awareness for rotational trajectory-following capacity and further exploits camera-disentangled finetuning to adapt this ability on open-domain videos. The dashed colorful arrows demonstrate the dedicated data flows during the three training stages, while the black arrows are shared in all stages. All yellow blocks are trainable, while other blocks are frozen." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 313, + 294, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 313, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 54, + 313, + 294, + 432 + ], + "type": "text", + "content": "We thus construct a synthetic dataset PoseTraj-10K, that offers three key advantages: i) it provides a diverse and comprehensive range of objects, enhancing robust motion understanding, particularly for complex geometric structures and wide-range rotations; ii) it enables the creation of a wide variety of precise trajectories, ensuring motion consistency without interference from camera movements; iii) through controlled rendering, we can effortlessly obtain precise 3D bounding box parameters, which serve as crucial intermediate supervision signals in our pretraining process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 612 + ], + "type": "text", + "content": "Fig. 3 illustrates our rendering pipeline implemented in Blender. First, we set up a realistic virtual scene and sampled 2,000 high-quality 3D models from Objaverse [9]. To ensure the quality of the selected models, we employ GPT-4v [21] to filter out overly complex or uncommon objects, followed by a manual selection, resulting in the top 2,000 models that best represent high-quality daily objects. Each model is normalized and assigned a unique random trajectory with randomized rotational angles, trajectory shape, and length. During the animation, the 3D model follows its designated trajectory while maintaining a rotation center, enabling rotation-focused motion for the final rendering. Fig. 3 bottom displays four frames of the rendered video, check out our dataset and the supplementary for more examples and further data construction details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 621, + 111, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 111, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 111, + 633 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " containing foreground objects and a driving trajectory " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{tr} = \\{(x_i, y_i)\\}_{i=1}^L" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " drawn as a set of line segment images " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathbf{tr}} = \\{I_{tr}^i\\}_{i=1}^L" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " (with last empty padding), our goal is to generate a sequence of animated video frames " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{f_i\\}_{i=1}^L" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": ", where the relevant object accurately follows the trajectory, while maintaining a coherent" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 313, + 436, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 436, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 436, + 325 + ], + "type": "text", + "content": "and natural video appearance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "text", + "content": "Fig. 4 presents an overview of our approach, which is built upon a latent stable video diffusion model, SVD [2]. Our base framework mainly consists of three components: a latent diffusion model " + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "text", + "content": " (3D U-Net [27]) to denoise the latent noise, and an encoder " + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "inline_equation", + "content": "(\\mathcal{E})" + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "text", + "content": " - decoder " + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "inline_equation", + "content": "(\\mathcal{D})" + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "text", + "content": " pair to compress video frames into the latent space and reconstruct the denoised latent features back into video frames. Inspired by the success of ControlNet [19] in guiding the pre-trained diffusion model, we design Traj-ControlNet, which consists of a trainable copy of the encoder blocks of SVD. Given a sequence of trajectory images " + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{tr}}" + }, + { + "bbox": [ + 313, + 326, + 555, + 518 + ], + "type": "text", + "content": ", we adopt a trajectory guider - a 3D ConvNet, to encode trajectory features, which is then fed into the Traj-ControlNet to predict the residual features for SVD, enabling the generated video to follow the trajectory. In the following subsections, we detail on the novel technical designs to achieve our goal." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 527, + 497, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 497, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 497, + 540 + ], + "type": "text", + "content": "4.1. Two-stage Pose-aware Pretraining" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 640 + ], + "type": "text", + "content": "To enhance the model's ability to perceive potential 6D pose changes within the 2D-pixel space, we select the 3D bounding box (bbox) [32] as the primary 3D supervision signals, as they provide precise positional and pose information. We propose a novel two-stage pose-aware pretraining pipeline that incorporates 3Dbbox generation as an intermediate supervision signal. The two pre-training is on our PoseTraj-10K synthetic dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Stage One: 3D bbox-guided localization. Instead of separately regressing 3D bbox parameters and generating the object's appearance for each video frame, we simultaneously generate 3D bboxes in the image space along with the objects, as shown in Fig. 4 (red arrows). To this end, given one data example in our dataset, we first render the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22779" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "content": "3D bounding box into the pixel space on top of the image to obtain " + }, + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{bbox}" + }, + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "content": " (the bottom-left figure in Fig. 4). We apply the same rendering process for other frames to obtain the bounding box augmented video for supervision. This joint generation process helps the model better understand the object's location along the driving trajectory and makes it aware of potential rotational changes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 157, + 294, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 157, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 54, + 157, + 294, + 252 + ], + "type": "text", + "content": "Stage Two: object-centric reconstruction. In practical applications, bounding boxes are not required in the generated video. Moreover, in the first stage, with explicit bounding box constraints, the model learns object entities and poses, but the generated object lacks appearance details. Therefore, in the second stage, we exclude bounding box generation and remove 3D bounding box supervision, finetuning the model to refine the object's appearance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 253, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 373 + ], + "type": "text", + "content": "There are several ways to inject auxiliary 3D information, e.g., inputting depth maps or 3D bounding boxes as conditions. Our injection-by-reconstruction approach offers two key advantages. First, the bounding boxes explicitly enhance continuous 3D awareness by serving as a pixel-level supervision target (see evaluation and supplementary material). Second, these supervisions can be easily removed by simply changing the reconstruction target, thereby mitigating inference-stage mismatches caused by inaccurate estimation of these additional signals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 383, + 233, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 233, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 233, + 396 + ], + "type": "text", + "content": "4.2. Camera-disentangled Finetuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 402, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 545 + ], + "type": "text", + "content": "After the two-stage pre-training, the model learns to perceive potential rotations while preserving object shape and detailed appearance along a trajectory. However, adapting the pre-trained model to real-world videos remains a key challenge. Unlike the synthetic dataset, where the camera is static, real-world videos often exhibit unpredictable and irregular camera motions. Undistinguished camera motion and object motion can lead to errors in object tracking. Therefore, we introduce camera-disentangled finetuning, which incorporates additional camera motion information into the model to help distinguish active object movement from passive camera motion (orange arrows in Fig. 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Specifically, we annotate the real-world video dataset - VIPSeg [18], with motion trajectories and per-frame camera poses (i.e., extrinsic parameters, denoted as " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\{C a m^{i}\\}_{i = 1}^{L}" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "). We incorporate camera poses as additional inputs by passing them through an MLP layer and then concatenating them with the trajectory features. The concatenated feature is then fed into a zero-initialized MLP projection layer and further input to the Traj-ControlNet. At inference time, it is hard to provide the camera pose information, we then randomly (50%) drop camera poses in this fine-tuning stage, enabling the model to generate videos without camera information. This approach helps the model differentiate between active object movement and passive camera motion, thereby improving trajectory-following accuracy and en" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 515, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 515, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 515, + 84 + ], + "type": "text", + "content": "hancing robustness across diverse video contexts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 90, + 446, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 446, + 102 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 446, + 102 + ], + "type": "text", + "content": "4.3. Training and Inference" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "text", + "content": "The mean squared error (MSE) loss is commonly used in video generation training, and the objective is to minimize the discrepancy between the predicted noise from " + }, + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "text", + "content": " and the ground truth noise " + }, + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "text", + "content": " for each latent noise " + }, + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "text", + "content": ". Both the two pre-training stages and fine-tuning stage leverage the same MSE loss, while the difference lies in the condition " + }, + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 313, + 107, + 553, + 190 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 347, + 201, + 553, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 201, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 347, + 201, + 553, + 233 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M S E}} = \\mathbb {E} _ {x _ {t}, \\epsilon} \\left[ \\sum_ {i = 1} ^ {L} \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(x _ {t}, t, C ^ {i}\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "4a9cad8494afd9c06414bb1064d6be97475ec7ca4a1cde235d557ad78122517a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 236, + 411, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 411, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 411, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 236, + 411, + 247 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 313, + 236, + 411, + 247 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 356, + 252, + 553, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 252, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 356, + 252, + 553, + 297 + ], + "type": "interline_equation", + "content": "C ^ {i} = \\left\\{ \\begin{array}{l l} \\left\\{I _ {t r} ^ {i}, \\mathbf {I} _ {b b o x} \\right\\}, & \\text {s t a g e o n e ,} \\\\ \\left\\{I _ {t r} ^ {i}, \\mathbf {I} \\right\\}, & \\text {s t a g e t w o ,} \\\\ \\left\\{I _ {t r} ^ {i}, \\mathbf {I}, C a m ^ {i} \\right\\}, & \\text {f i n e t u n i n g .} \\end{array} \\right. \\tag {2}", + "image_path": "96b50e4dced29b63f001590354c78f624e7a42e339abf155e2f0f96707d5ae5e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 301, + 553, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 553, + 325 + ], + "type": "text", + "content": "Remeber that, in the first pre-training stage, we used the bbox-augmented video frames as the reconstruction target." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "text", + "content": "Spatial enhancement loss. Simply aggregating errors across all frames overlooks the spatial reconstruction accuracy of individual frames, potentially leading to object entity collapse under large rotational motions. Since each short per-frame trajectory segment provides precise localization information within individual frames, we introduce an additional spatial enhancement loss to improve spatial consistency by reconstructing the per-frame image. Specifically, we randomly sample the trajectory of the " + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "text", + "content": "-th frame (i.e., " + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "inline_equation", + "content": "I_{tr}^{j}" + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "text", + "content": ") and use it as the Traj-ControlNet condition. From its corresponding latent noise " + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "inline_equation", + "content": "x_{t,j}" + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "text", + "content": " and the initial frame " + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{bbox}" + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "text", + "content": " for the first stage), the objective is to train the model to generate the target noise " + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\epsilon_{j}" + }, + { + "bbox": [ + 313, + 325, + 554, + 493 + ], + "type": "text", + "content": " for this specific frame (i.e., single frame image reconstruction), formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 498, + 553, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 498, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 339, + 498, + 553, + 513 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S P A}} = \\left| \\right.\\left| \\right. \\epsilon_ {j} - \\epsilon_ {\\theta} \\left(x _ {t, j}, t, C ^ {j}\\right)\\left. \\right)\\left. \\right|\\left. \\right| _ {2} ^ {2}, j \\in (1 \\sim L). \\tag {3}", + "image_path": "3139d5a75b7ba53d23aa4151a8e9de738af58a25ecc211d57d21db23cf2cc48f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 518, + 553, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 553, + 553 + ], + "type": "text", + "content": "During backpropagation, only the spatial layers are updated, ensuring the refinement focuses on per-frame spatial accuracy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 554, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 553, + 578 + ], + "type": "text", + "content": "Overall, the training loss function of our framework for each stage is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 381, + 584, + 553, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 584, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 381, + 584, + 553, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {M S E}} + \\lambda_ {\\text {S P A}} \\mathcal {L} _ {\\text {S P A}}, \\tag {4}", + "image_path": "0d71c91fe114d178085cb9d832283de243903a80d229a91cd535f7896cdf6606.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 603, + 446, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 446, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 446, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 603, + 446, + 615 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SPA}}" + }, + { + "bbox": [ + 313, + 603, + 446, + 615 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Inference time video generation. During inference, given an input image, users freely draw a custom trajectory, which can be either rotational or translational, our model will generate corresponding pose-aware videos. Note that our approach supports more than one trajectory, see Figs. 2 and 6 for examples. In addition, users can control the camera motion behavior by providing camera poses extracted from existing videos." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22780" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 555, + 370 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 370 + ], + "type": "image", + "image_path": "990f58f3dfd16340ba09be3ba509de14b9e222a27b148d0ae34554c66ab5773f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 373, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 555, + 407 + ], + "type": "text", + "content": "Figure 5. Visual comparison. Given the same trajectory and the initial image, our method produces plausible video frames containing the moving object following the given trajectory. In contrast, DrafNUWA and DragAnything either introduce unexpected camera motions or fail to maintain the object entity, causing severe collapse." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 107, + 430, + 502, + 504 + ], + "blocks": [ + { + "bbox": [ + 159, + 415, + 449, + 426 + ], + "lines": [ + { + "bbox": [ + 159, + 415, + 449, + 426 + ], + "spans": [ + { + "bbox": [ + 159, + 415, + 449, + 426 + ], + "type": "text", + "content": "Table 1. Quantitative comparison between our approach and the SoTA methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 107, + 430, + 502, + 504 + ], + "lines": [ + { + "bbox": [ + 107, + 430, + 502, + 504 + ], + "spans": [ + { + "bbox": [ + 107, + 430, + 502, + 504 + ], + "type": "table", + "html": "
MethodsVIPSeg 256 × 256VIPSeg 320 × 576DAVIS 320 × 576
ObjMC ↓FID ↓FVD ↓ObjMC ↓FID ↓FVD ↓ObjMC ↓FID ↓FVD ↓
DragNUWA 1.5 [40]173.6163.39449.59133.0541.88289.1574.0754.10952.87
DragAnything [36]100.2361.69410.7091.1239.29275.9347.0150.83771.78
Ours87.5646.60384.4177.4838.41267.3329.9251.48729.16
", + "image_path": "128e0de99688027651cc9293d1b7668ddfa5b008e28179c91a5a83a83f824fa1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 518, + 222, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 222, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 222, + 533 + ], + "type": "text", + "content": "5. Experiments and Evaluations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "text", + "content": "Implementation details. Our base model is SVD [2], trained to generate 14 frames given an initial frame at a resolution of " + }, + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "inline_equation", + "content": "320 \\times 576" + }, + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "text", + "content": " pixels. All experiments are conducted using a single Tesla A100 GPU with a batch size of 1. More implementation details are in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Synthetic dataset. For our synthetic dataset used in training, we generated 10,000 videos by sampling 5 random trajectories for each of 2,000 high-quality objects and rendering them in Blender through the Cycles engine at 5 fps. For evaluation, we create an additional trajectory for a subset of 200 sampled objects, which serves as our test set. During trajectory annotation, we augment the trajectory placement through a trajectory sampler (see the supplementary). Both training and evaluation on the synthetic dataset are" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 520, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 544 + ], + "type": "text", + "content": "conducted in the resolution of " + }, + { + "bbox": [ + 313, + 520, + 555, + 544 + ], + "type": "inline_equation", + "content": "320 \\times 576" + }, + { + "bbox": [ + 313, + 520, + 555, + 544 + ], + "type": "text", + "content": " pixels with 14 frames per sequence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 628 + ], + "type": "text", + "content": "Real-world dataset. For real-world finetuning, we adopted the VIPSeg dataset [18], following DragAnything. VIPSeg is a large-scale video segmentation dataset comprising over 3,000 high-quality videos with annotated object masks. We utilized its training set for finetuning and its validation set for evaluation, employing the same trajectory extraction pipeline as in DragAnything (see supplementary)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 700 + ], + "type": "text", + "content": "To further evaluate the generalization capability of our approach, we additionally select the DAVIS dataset [24], a high-quality video segmentation benchmark, for out-of-distribution (OOD) evaluation. Note that, we only annotate the trajectory for this dataset for OOD evaluation, and at inference, camera poses are not used." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 701, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 712 + ], + "type": "text", + "content": "Evaluation metrics. We follow established evaluation met-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22781" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "rics from previous trajectory-guided studies [23, 36, 40], assessing performance across two key aspects:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 295, + 205 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 97, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 295, + 156 + ], + "type": "text", + "content": "- Controlling accuracy. To quantify the precision of trajectory-guided dragging, we use the ObjMC metric, which computes the mean squared error between the temporal motion tendency of generated samples and that of the ground truth videos." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 157, + 295, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 205 + ], + "type": "text", + "content": "- Generation quality. We employ the Fréchet Inception Distance (FID) [12] to assess the single-frame spatial quality and the Fréchet Video Distance (FVD) [31] to measure temporal consistency." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 215, + 136, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 136, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 136, + 227 + ], + "type": "text", + "content": "5.1. Comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 233, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 295, + 483 + ], + "type": "text", + "content": "Competitors. We primarily evaluate our model's performance against two recent trajectory-guided video generation models: DragNUWA 1.5 [40] and DragAnything [36]. DragNUWA is the first method designed for open-domain video dragging. It utilizes optical flow estimated by UniMatch [38] as trajectory guidance, combining these trajectories with text and image conditions to enable motion-controllable fine-tuning. Built upon DragNUWA, DragAnything enhances object-level animation by extracting entity feature maps from instance masks and integrating them into ControlNet for further fine-tuning. To ensure fair comparisons, the same trajectory is converted into each model's respective trajectory representation (e.g., the trajectory point set or trajectory segment images). Additionally, since DragAnything reports metric performance only at a resolution of " + }, + { + "bbox": [ + 55, + 233, + 295, + 483 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 233, + 295, + 483 + ], + "type": "text", + "content": ", we evaluate our model at both the original " + }, + { + "bbox": [ + 55, + 233, + 295, + 483 + ], + "type": "inline_equation", + "content": "320 \\times 576" + }, + { + "bbox": [ + 55, + 233, + 295, + 483 + ], + "type": "text", + "content": " training resolution and the downsampled " + }, + { + "bbox": [ + 55, + 233, + 295, + 483 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 233, + 295, + 483 + ], + "type": "text", + "content": " resolution to ensure a fair comparison. Due to the unavailability of public model weights for Tora and MotionCtrl with SVD pretraining, we exclude these models from our comparison." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": "Quantitative Results. As shown in Tab. 1, PoseTraj outperforms competing methods across both resolutions in all trajectory-specific and video quality metrics, with particularly strong performance in lower-resolution synthesis. In terms of trajectory accuracy, our model achieves a significant reduction in error than DragAnything, with improvements of " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": " for lower and standard resolutions on the VIPSeg dataset and standard resolution on the DAVIS dataset, respectively. These results demonstrate the effectiveness of our pose-aware pretraining approach, which leverages high-precision synthetic data pairs with 3D bounding boxes and incorporates a camera-disentanglement module to enhance the model's understanding of object-oriented motion and passive camera movement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "For single-frame reconstruction performance, measured by FID, our model outperforms DragAnything by a significant margin, achieving an improvement of 15.01 at the downsampled " + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": " resolution and 0.88 at the standard " + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "inline_equation", + "content": "320 \\times 576" + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": " resolution on the VIPSeg dataset. Fur" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 194 + ], + "type": "image", + "image_path": "21147d34b8964b2b77f7074bcc144407ae68552462ee37f1f45a8c82dcf609f9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 198, + 523, + 208 + ], + "lines": [ + { + "bbox": [ + 343, + 198, + 523, + 208 + ], + "spans": [ + { + "bbox": [ + 343, + 198, + 523, + 208 + ], + "type": "text", + "content": "Figure 6. Transitional samples on DAVIS dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 218, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 337 + ], + "type": "text", + "content": "thermore, our model demonstrates improved temporal consistency across frames, achieving substantial reductions in FVD scores: 26.29, 8.6, and 42.62 across the three evaluation settings. These results highlight our model's significant advancements in trajectory-guided motion control, achieving precise dragging-following and reliable motion. Notably, our approach shows strong performance in traditional in-distribution scenarios while exhibiting even greater superiority in out-of-distribution (OOD) settings (see the evaluation on DAVIS in Tab. 1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 338, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 529 + ], + "type": "text", + "content": "Qualitative Results. Visual comparisons are displayed in Fig. 5, where PoseTraj achieves superior performance in generating 3D and pose-aligned rotational motions that accurately follow 2D trajectories with precise pixel-level trajectory adherence. Furthermore, our model demonstrates enhanced capability in reliably reconstructing occluded object regions, even under changing 6D poses, such as rotational car doors. In contrast, DragNUWA primarily generates moving camera perspectives while keeping object poses static, resulting in poor rotational alignment. DragAnything produces unstable and mismatched rotational motions, where dragged pixels progressively drift across frames, potentially leading to entity collapse. This comparison highlights the effectiveness of our approach in generating more accurate, stable, and realistic motion sequences, particularly when handling complex trajectories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 529, + 554, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 577 + ], + "type": "text", + "content": "Additionally, in Fig. 6, we show visual results on OOD examples, where the trajectory is mostly translational instead of rotational. Our generated videos are superior to DragAnything." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 325, + 589, + 537, + 669 + ], + "blocks": [ + { + "bbox": [ + 325, + 589, + 537, + 669 + ], + "lines": [ + { + "bbox": [ + 325, + 589, + 537, + 669 + ], + "spans": [ + { + "bbox": [ + 325, + 589, + 537, + 669 + ], + "type": "image", + "image_path": "01a88f571a72a3edced5ee8d48117ae602eb08535215bad1eee8ecb975d51339.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 673, + 525, + 685 + ], + "lines": [ + { + "bbox": [ + 341, + 673, + 525, + 685 + ], + "spans": [ + { + "bbox": [ + 341, + 673, + 525, + 685 + ], + "type": "text", + "content": "Figure 7. The voting results of the user evaluation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "User Evaluation. Other than the visual and statistical evaluations, we conduct a subjective user evaluation, where par" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22782" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 96, + 251, + 171 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 2. Ablation results on the real-world video test set with several model variants." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 96, + 251, + 171 + ], + "lines": [ + { + "bbox": [ + 99, + 96, + 251, + 171 + ], + "spans": [ + { + "bbox": [ + 99, + 96, + 251, + 171 + ], + "type": "table", + "html": "
Methods320 × 576
ObjMC ↓FID ↓FVD ↓
Full method77.4838.41267.33
Nobbox stage81.3641.90275.40
No pretrain145.7242.62486.84
No Cam-disen83.2239.71279.15
No SPA-loss137.2639.79436.56
", + "image_path": "1d31600e54d0f6ed5e8664941d01231b783686921e088210ff1cc80238dc3186.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 93, + 200, + 257, + 259 + ], + "blocks": [ + { + "bbox": [ + 55, + 175, + 295, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 295, + 198 + ], + "type": "text", + "content": "Table 3. Comparison results on the synthetic validation set of our designed ablation studies on the pretraining stage." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 200, + 257, + 259 + ], + "lines": [ + { + "bbox": [ + 93, + 200, + 257, + 259 + ], + "spans": [ + { + "bbox": [ + 93, + 200, + 257, + 259 + ], + "type": "table", + "html": "
MethodsSynthetic Dataset Val
ObjMC ↓FID ↓FVD ↓
Full method0.19647.18187.12
Nobbox stage0.26650.30223.74
No SPA-loss0.26546.72197.68
", + "image_path": "e5783d82ebadafa89b74986b7a8629028c0122389f3acab8de26bdcb886f27d0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 270, + 296, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 329 + ], + "type": "text", + "content": "ticipants rate generated videos (20 sampled videos from the VIPSeg test set and manually annotated wild images) based on two criteria: overall motion accuracy and visual quality. We have invited seven experienced experts on the video generation topic and asked them to rate the videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 330, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 390 + ], + "type": "text", + "content": "As shown in Fig. 7, our approach is favorably voted, receiving " + }, + { + "bbox": [ + 55, + 330, + 296, + 390 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 55, + 330, + 296, + 390 + ], + "type": "text", + "content": " more votes in trajectory-following accuracy and " + }, + { + "bbox": [ + 55, + 330, + 296, + 390 + ], + "type": "inline_equation", + "content": "39\\%" + }, + { + "bbox": [ + 55, + 330, + 296, + 390 + ], + "type": "text", + "content": " more votes in video quality. These findings are consistent with the superior performance observed in quantitative metrics and qualitative visualizations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 396, + 149, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 149, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 149, + 410 + ], + "type": "text", + "content": "5.2. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "text", + "content": "Pose-aware pretraining. To evaluate the impact of our pretraining strategy, we propose two experimental variants: (1) training the model without 3D-aware pertaining (denoted as 'No pretrain'), and (2) training the model using only the second-stage pretraining while omitting 3D bounding box supervision (denoted as 'No bbox stage')." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 487, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 677 + ], + "type": "text", + "content": "As shown in Tabs. 2 and 3, excluding 3D bounding box supervision during pretraining leads to a degradation in performance across all evaluation metrics on both the synthetic and open-domain datasets. Notably, in Tab. 2, compared with 'No pretraining', the absence of 3D bounding box pretraining has minimal impact on the metric. This is because 3D bounding boxes mainly affect object pose localization, and inaccurate poses do not significantly degrade motion accuracy or quality. However, the visual results in Fig. 8 clearly demonstrate the defects without bounding box pretraining, i.e., there is pronounced object collapse, exacerbating trajectory misalignment. More visual comparisons can be found in the supplementary material. Furthermore, omitting the two-stage pretraining approach results in even greater performance degradation, leading to a substantial drop in trajectory-following capability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Spatial enhancement loss. Compared to standard video generation tasks, achieving both precise trajectory control and spatial visual consistency is significantly more chal" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 553, + 178 + ], + "type": "image", + "image_path": "af3c31a3e418b98a255d48c9332257dce4c6bdd518f848a70e3af9c1b1e5e903.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 182, + 546, + 194 + ], + "lines": [ + { + "bbox": [ + 321, + 182, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 321, + 182, + 546, + 194 + ], + "type": "text", + "content": "Figure 8. Visualization of ablation results in synthetic dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 201, + 553, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 553, + 309 + ], + "type": "text", + "content": "lenging, especially when the motion involves complex dynamics or large-angle rotations. Consequently, integrating spatial enhancement loss and formulating image-to-image prediction as a sub-task during training becomes essential. As demonstrated in Tabs. 2 and 3 and Fig. 8, models trained without spatial loss (denoted as 'No SPA-loss') show considerable degradation in trajectory alignment reliability and generation quality, with frequent blurring observed for fast-moving objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 312, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 467 + ], + "type": "text", + "content": "Camera disentanglement. Considering the unexpected influence of camera motions, we incorporate camera disentangled fine-tuning to enhance the model's trajectory, controlling awareness of generating naturally activated object motion with passive camera movements. We evaluate the impact of this design by discarding camera poses (denoted as 'No Cam-disen'). Experimental results (Tab. 2) indicate that models without camera disentangling exhibit lower object motion-matching accuracy and a corresponding decrease in the FVD metric. This inaccuracy arises from the model's uncertainty regarding how to generate aligned passive camera motions and results in a reduced ability to handle trajectory variations involving background motion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 468, + 554, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 493 + ], + "type": "text", + "content": "We show more visual results for ablation experiments in the supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 510, + 388, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 388, + 522 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 388, + 522 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 532, + 554, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 675 + ], + "type": "text", + "content": "In this paper, we have presented PoseTraj, a novel approach for pose-aware trajectory control in video generation. We propose a two-stage pre-training strategy to enable the accurate trajectory-following ability, while explicitly exploiting 3D bounding boxes as a supervision signal to perceive the object's potential 6D pose changes along the driving trajectory. A new dataset, PoseTraj-10K, is constructed to achieve our goal. Our approach outperforms state-of-the-art methods quantitatively and qualitatively. We hope that the 3D bounding box signal and the injection-by-reconstruction design can inspire future research in controllable video generation or relevant topics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Acknowledgment. LB completed part of this work during his Master's study at the University of Edinburgh. CJ was supported by a gift from Adobe." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22783" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14707-14717, 2021. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 216, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 294, + 280 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 283, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 283, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 283, + 294, + 337 + ], + "type": "text", + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 339, + 294, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 339, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 62, + 339, + 294, + 383 + ], + "type": "text", + "content": "[5] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. European Conference on Computer Vision (ECCV), 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 385, + 294, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 294, + 439 + ], + "type": "text", + "content": "[6] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 441, + 294, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 294, + 484 + ], + "type": "text", + "content": "[7] Tsai-Shien Chen, Chieh Hubert Lin, Hung-Yu Tseng, Tsung-Yi Lin, and Ming-Hsuan Yang. Motion-conditioned diffusion model for controllable video synthesis. arXiv preprint arXiv:2304.14404, 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 487, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 294, + 529 + ], + "type": "text", + "content": "[8] Zheng Dang, Jialu Huang, Fei Wang, and Mathieu Salzmann. Openmaterial: A comprehensive dataset of complex materials for 3d reconstruction. arXiv preprint arXiv:2406.08894, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "type": "text", + "content": "[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 62, + 533, + 294, + 586 + ], + "type": "text", + "content": " objects. Advances in Neural Information Processing Systems, 36, 2024. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 644 + ], + "type": "text", + "content": "[10] Patrick Esser, Johnathan Chiu, Parmida Atighechian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[11] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 692, + 294, + 713 + ], + "type": "text", + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 74, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 553, + 105 + ], + "type": "text", + "content": "two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "text", + "content": "[13] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 153, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 555, + 196 + ], + "type": "text", + "content": "[14] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 198, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 252 + ], + "type": "text", + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 255, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 553, + 297 + ], + "type": "text", + "content": "[16] Ruining Li, Chuanxia Zheng, Christian Rupprecht, and Andrea Vedaldi. Puppet-master: Scaling interactive video generation as a motion prior for part-level dynamics. arXiv preprint arXiv:2408.04631, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 300, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 553, + 353 + ], + "type": "text", + "content": "[17] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Xiu Li, and Qifeng Chen. Follow your pose: Pose-guided text-to-video generation using pose-free videos. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4117-4125, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "type": "text", + "content": "[18] Jiaxu Miao, Xiaohan Wang, Yu Wu, Wei Li, Xu Zhang, Yunchao Wei, and Yi Yang. Large-scale video panoptic segmentation in the wild: A benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2022. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 411, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 555, + 465 + ], + "type": "text", + "content": "[19] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 468, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 555, + 533 + ], + "type": "text", + "content": "[20] Haomiao Ni, Bernhard Egger, Suhas Lohit, Anoop Cherian, Ye Wang, Toshiaki Koike-Akino, Sharon X Huang, and Tim K Marks. Ti2v-zero: Zero-shot image conditioning for text-to-video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9015-9025, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 535, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 555 + ], + "type": "text", + "content": "[21] OpenAI. GPT (v4) [Large language model]. https://chat.openai.com, 2024.4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 558, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 553, + 612 + ], + "type": "text", + "content": "[22] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In European Conference on Computer Vision, pages 163-182. Springer, 2025. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 614, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 668 + ], + "type": "text", + "content": "[23] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[24] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22784" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 94 + ], + "type": "text", + "content": "on computer vision and pattern recognition, pages 724-732, 2016. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[25] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 263 + ], + "type": "text", + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 295, + 330 + ], + "type": "text", + "content": "[28] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 295, + 385 + ], + "type": "text", + "content": "[29] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 430 + ], + "type": "text", + "content": "[30] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive - generating expressive portrait videos with audio2video diffusion model under weak conditions, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 295, + 476 + ], + "type": "text", + "content": "[31] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 520 + ], + "type": "text", + "content": "[32] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis, 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 577 + ], + "type": "text", + "content": "[33] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "text", + "content": "[34] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. arXiv preprint arXiv:2312.03641, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 677 + ], + "type": "text", + "content": "[35] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Nüwa: Visual synthesis pretraining for neural visual world creation. In European conference on computer vision, pages 720-736. Springer, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "text", + "content": "[36] Wejia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 552 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "for anything using entity representation. arXiv preprint arXiv:2403.07420, 2024.1,6,7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[37] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": "[38] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In CVPR, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "text", + "content": "[39] Shiuyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[40] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "text", + "content": "[41] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgbx: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "text", + "content": "[42] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to " + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "text", + "content": " panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "type": "text", + "content": "[43] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 432, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 496 + ], + "type": "text", + "content": "[44] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "text", + "content": "[45] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 1" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22785" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/867e9811-fe1d-4217-8c41-8e621b6da43b_content_list.json b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/867e9811-fe1d-4217-8c41-8e621b6da43b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eb5383977aea39056c3e00a5fa3e549b53a7cb58 --- /dev/null +++ b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/867e9811-fe1d-4217-8c41-8e621b6da43b_content_list.json @@ -0,0 +1,1920 @@ +[ + { + "type": "text", + "text": "Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising", + "text_level": 1, + "bbox": [ + 197, + 130, + 799, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tong Li $^{1}$ Lizhi Wang $^{2,3*}$ Zhiyuan Xu $^{1}$ Lin Zhu $^{1}$ Wanxuan Lu $^{4}$ Hua Huang $^{2,3}$", + "bbox": [ + 133, + 202, + 926, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Technology, Beijing Institute of Technology $^{2}$ School of Artificial Intelligence, Beijing Normal University", + "bbox": [ + 217, + 220, + 844, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Engineering Research Center of Intelligent Technology and Educational Application, Ministry of Education", + "bbox": [ + 99, + 256, + 960, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Aerospace Information Research Institute, Chinese Academy of Sciences", + "bbox": [ + 238, + 273, + 823, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image denoising enhances image quality, serving as a foundational technique across various computational photography applications. The obstacle to clean image acquisition in real scenarios necessitates the development of self-supervised image denoising methods only depending on noisy images, especially a single noisy image. Existing self-supervised image denoising paradigms (Noise2Noise and Noise2Void) rely heavily on information-lossy operations, such as downsampling and masking, culminating in low-quality denoising performance. In this paper, we propose a novel self-supervised single image denoising paradigm, Positive2Negative, to break the information-lossy barrier. Our paradigm involves two key steps: Renoised Data Construction (RDC) and Denoised Consistency Supervision (DCS). RDC renoes the predicted denoised image by the predicted noise to construct multiple noisy images, preserving all the information of the original image. DCS ensures consistency across the multiple denoised images, supervising the network to learn robust denoising. Our Positive2Negative paradigm achieves state-of-the-art performance in self-supervised single image denoising with significant speed improvements. The code is released to the public at https://github.com/Li-Tong-621/P2N.", + "bbox": [ + 89, + 358, + 483, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 747, + 220, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image denoising plays a vital role in image processing to enhance visual quality [13], serving as a foundational technique across various computational photography applications [3, 39]. The widespread popularity of paired clean-noisy image datasets [1, 35] has catalyzed advancements in supervised methods, yielding remarkable performance [7, 51]. However, supervised methods fall short in", + "bbox": [ + 89, + 773, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6030aee1e7cbeaca920ad71a0402be5f35f7867661fa67f2fd60255d45328d20.jpg", + "image_caption": [ + "Noise2Noise-Based Paradigms" + ], + "image_footnote": [], + "bbox": [ + 519, + 327, + 903, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a003c6634aeed0e0b9ed91b36d2b0cf0004cb1e5683948feefe0ed9452d53f89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 445, + 901, + 547 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50164d06ded5d99c14301ce13480195e748ee747b9db641fe115f1da6cecc3b6.jpg", + "image_caption": [ + "Noise2Void-Based Paradigms", + "Positive2Negative Paradigm", + "Figure 1. Information-lossy operations in current paradigms. The noise addition operation reduces the signal-to-noise ratio, the downsampling operation diminishes the sampling density, and the masking operation disregards crucial central pixels. Each of these operations introduces an information-lossy barrier [8, 31, 37], culminating in imprecise denoising." + ], + "image_footnote": [], + "bbox": [ + 519, + 561, + 901, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dynamic scenes [18, 22], where long exposure results in misalignment and blur [22, 42], hindering the acquisition of clean images. Therefore, advancing self-supervised image denoising methods using only noisy images [25, 43], especially a single noisy image [9, 11, 24, 27], holds significant importance.", + "bbox": [ + 511, + 791, + 906, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The crux of self-supervised image denoising methods", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author: Lizhi Wang (wanglizhi@bnu.edu.cn)", + "bbox": [ + 107, + 886, + 437, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17924", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4df928a1a33cb13fde08a59c6a9b34639ebf6b25cc8dd8aacfb75a2aeb52132.jpg", + "image_caption": [ + "Figure 2. The information-lossy barrier significantly compromises the image quality, resulting in residual noise, aliasing effects or texture loss." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 478, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lies in the construction of the training data and the formulation of the supervision strategy. Current methods are divided into Noise2Noise-based paradigms and Noise2Void-based paradigms, as illustrated in Figure 1. Noise2Noise-based paradigms add additional noise to the noisy image [28, 33] or down sample the noisy image [14, 27] to approximately construct two independently noisy observations as training data. These paradigms learn denoising by supervising one observation with the denoised result of other observation. Noise2Void-based paradigms mask either the noisy image [18, 40] or the neural network [8, 21, 41] to construct masked training data. These paradigms learn denoising by supervising the denoised result of the masked central pixel with the original noisy pixel, where the prediction only relies on the surrounding pixels.", + "bbox": [ + 88, + 321, + 482, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing methods rely heavily on information-lossy operations, resulting in low-quality denoising performance [21, 32, 40, 41]. Specifically, the noise addition operation reduces the signal-to-noise ratio, the downsampling operation diminishes the sampling density and the masking operation disregards crucial central pixels. Each of these operations loses the information from the original noisy image [8, 31, 37], inevitably compromising the denoising ability [41, 44]. Consequently, denoised images always suffer from remaining noise, aliasing effects or texture loss [45, 46], as shown in Figure 2. As these information-lossy operations constitute the indispensable foundation of the Noise2Noise-based and Noise2Void-based paradigms, it is imperative to establish a new paradigm that transcends the Noise2Noise-based and Noise2Void-based paradigms.", + "bbox": [ + 89, + 551, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose the concise Positive2Negative paradigm to break the information-lossy barrier in self-supervised single image denoising. Our inspiration stems from the observation that noise distributions are approximately symmetrical and centered around zero, implying that the opposite noise also follows the distribution of the original noise. Leveraging this insight, we can construct multiple noisy images corresponding to the same clean image, which", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "serves as the basis of Positive2Negative. Concretely, Positive2Negative relies on two steps: Renoised Data Construction (RDC) and Denoised Consistency Supervision (DCS).", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The basic idea of RDC is to construct training data. Specifically, we design a zero-mean sampling strategy to generate multi-scale Positive noise and Negative noise based on the predicted noise, which are then added back to construct Positive noisy images and Negative noisy images. The constructed data comprehensively covers the information of the given single image; thus the Positive2Negative paradigm breaks the information-lossy barrier.", + "bbox": [ + 511, + 136, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The basic idea of DCS is to formulate a supervision strategy. Specifically, given the constructed multi-scale noisy images, DCS supervises the network to generate consistent denoised outputs during the training process. Furthermore, we provide a theoretical analysis showing that the Positive2Negative paradigm can learn robust denoising, through supervising the multiple denoised images.", + "bbox": [ + 511, + 257, + 903, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 532, + 363, + 834, + 377 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new self-supervised single image denoising paradigm, Positive2Negative, to break the information-lossy barrier.", + "- We propose a data construction method, which constructs multi-scale similar noisy images for training.", + "- We propose a denoising supervision method, which is theoretically guaranteed to learn robust denoising.", + "- Positive2Negative achieves SOTA performance over self-supervised single image denoising." + ], + "bbox": [ + 535, + 380, + 903, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 527, + 653, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From the training data perspective, the self-supervised image denoising methods can be classified into noisy dataset based methods and single image based methods.", + "bbox": [ + 511, + 553, + 903, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Noisy Dataset based Methods", + "text_level": 1, + "bbox": [ + 511, + 606, + 774, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Noisy dataset based methods represent the most active research direction in image denoising. These methods learn to denoise from a given noisy dataset. Generally, a larger volume of the noisy dataset improves the denoising performance. As a result, in scenarios where only a single noisy image is available, the denoising performance of whole noisy dataset-based methods decreases significantly.", + "bbox": [ + 511, + 628, + 903, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Noise2Noise [22] learns denoising by mapping one noisy observation to another independent observation of the same scene. Subsequent studies focus on generating observations of the same scene from a single observation, typically achieved through downsampling [14, 27] or by introducing additional noise [28, 33, 50, 54, 58]. After that, another elegant method Noise2Void [18] is proposed. Noise2Void [18] assumes that the noise distribution is zero-mean, the noise is spatially independent, and the signal is spatially correlated. Noise2Void [18] predicts the clean signal by leveraging surrounding noisy signals, with mask", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17925", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing the central pixel. Various other works propose alternative masking schemes, either mask the input [11, 36, 40] or mask the network [2, 19, 20, 34, 48, 52]. Noise2Void struggles with real-world noise, which often exhibits spatial correlation. Therefore, subsequent research endeavors to break the noise spatial correlation [8, 21, 23, 41, 47, 57]. There are also some methods that amalgamate frameworks established by both Noise2Noise and Noise2Void [15, 32] or get rid of frameworks established by Noise2Noise and Noise2Void [5, 16, 17, 30, 56], but these methods are usually less developed due to the complexity.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Single Image based Methods", + "text_level": 1, + "bbox": [ + 89, + 266, + 346, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single image based methods represent the most challenging situation for denoising and the least demanding requirements for data. These methods learn to denoise from a given single noisy image without any dependence on paired noisy-clean or noisy datasets. As a result, single image based methods are less developed due to the inherent difficulty, only gaining attention in the past two years.", + "bbox": [ + 89, + 287, + 483, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DIP [38] is the earliest self-supervised image denoising method. DIP [38] learns to map random noise to a given noisy image and employs early stopping to prevent overfitting. Self2Self [36] uses dropout to create two Bernoulli-sampled observations of a noisy image. R2R [33] constructs noisy images by adding new Gaussian noise and applying reciprocal coefficients. ScoreDVI [9] performs real denoising using a pre-trained Gaussian denoiser and an estimated noise model. ZS-N2N [27] extend Noise2Noise [22] to a single image version by downsampling. MASH [11] employs masking and local pixel shuffling to break the noise spatial correlation. Recently, DMID stimulates the diffusion model for self-supervised single image denoising [24].", + "bbox": [ + 89, + 393, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose Positive2Negative, a new self-supervised single image denoising paradigm, to break the information-lossy barrier.", + "bbox": [ + 89, + 590, + 483, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Positive2Negative", + "text_level": 1, + "bbox": [ + 89, + 648, + 263, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first demonstrate the motivation of Positive2Negative in Section 3.1. After that, we present the two critical steps of Positive2Negative: RDC and DCS, respectively in Section 3.2 and Section 3.3. Finally, we provide an analysis and proof of Positive2Negative in Section 3.4.", + "bbox": [ + 89, + 672, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 89, + 757, + 210, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our inspiration stems from the widely recognized assumptions that the noise distribution is zero-mean and approximately symmetrical [18, 22], which is consistent with our observation. In our observation experiment, we calculate the noise distribution of a fixed pixel across multiple observations of the same scene. Specifically, following SIDD [1], ELD [42] and PMN [12], we collect noisy image $y$ and clean image $c$ . Then we calculate the noise value $n$ at each", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/886d228308623bd181382553d5106bdc587a173134a3eb334629454421e61d37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/697cfef3d7501d8e3ace79726fdad6473499e5855cae7834a6c43d12583cc41d.jpg", + "image_caption": [ + "(a). noisy image", + "(b). green circle" + ], + "image_footnote": [], + "bbox": [ + 516, + 272, + 705, + 354 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c35e222e7404ee6411235128c20f70c831b3088ee5aa5a0c6998326b95f7539b.jpg", + "image_caption": [ + "(c). red circle" + ], + "image_footnote": [], + "bbox": [ + 712, + 272, + 901, + 354 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b516d66181a60f87abc87956a2bb38162b2c6694552d10e4b25316008d5f73ef.jpg", + "image_caption": [ + "Figure 3. Noise distribution is zero-mean and approximately symmetrical. (a) shows a noisy image. (b) and (c) show the noise distributions, which are calculated at the center pixels marked by the green circle and the red circle in the noisy image, respectively. It is evident that the noise distribution is zero-mean and approximately symmetrical.", + "Original noise", + "n" + ], + "image_footnote": [], + "bbox": [ + 521, + 477, + 614, + 549 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/699519e8f7265c12a2d3c7d1429fb3a80cade584cea652d197f3d4d286490148.jpg", + "image_caption": [ + "Original noisy", + "$y = c + n$", + "Figure 4. The opposite noise $-n$ is similar to the original noise $n$ . The opposite noise $-n$ approximately follows the original noise's distribution, which is zero-mean and approximately symmetrical." + ], + "image_footnote": [], + "bbox": [ + 617, + 477, + 710, + 550 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7f331d157e6699eb6a5d689453d34f19844d1e76cbd046d7bd9f69f4851a252a.jpg", + "image_caption": [ + "Opposite noise", + "$n_{n} = -n$" + ], + "image_footnote": [], + "bbox": [ + 712, + 477, + 803, + 549 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/16ae37fb525c992e55db198498e7f9ce967b66372b1017ed8b05a504e3caee60.jpg", + "image_caption": [ + "Opposite noisy", + "$y_{n} = c + (-n)$" + ], + "image_footnote": [], + "bbox": [ + 807, + 477, + 900, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "spatial location $(i,j)$ . Additionally, we compute the frequency of $n_{i,j}$ cross noisy images of the same scene. The results are shown in Figure 3, which demonstrates the noise distribution is zero-mean and approximately symmetrical.", + "bbox": [ + 511, + 659, + 905, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this observation, an intuitive insight is that the opposite noise $-n$ also follows the same distribution as that of the original noise $n$ . This intuitive idea inspires us to construct the opposite noise $n_{n}$ and corresponding opposite noisy image $y_{n}$ :", + "bbox": [ + 511, + 720, + 905, + 796 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nn _ {n} = - n, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 810, + 903, + 824 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {n} = x + n _ {n} = x - n. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 828, + 903, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The opposite noise $n_n$ and opposite noisy image $y_n$ are similar to the original noise $n$ and original noisy image $y$ , respectively, as shown in Figure 4.", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17926", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0b829f6c936d354a94499a951522460a36a8e9749d5a496bdbaedc8629236c4a.jpg", + "image_caption": [ + "Figure 5. Overview of the proposed self-supervised single image denoising paradigm. For training, the Positive2Negative paradigm consists of 2 steps. The first step is Renoised Data Construction (RDC): multiply the predicted noise by positive and negative coefficients to construct multi-scale Positive noise and Negative noise, respectively. The second step is Denoised Consistency Supervision (DCS): train the network through supervising the consistency of the denoised images. For inference, denoising can be achieved in just one pass through the network. (The networks shown in the figure are the same one. In addition, for better visualization, the images within the red boxes on the bottom left have been zoomed in and are displayed in the red boxes on the bottom right.)" + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above observations and insights, we propose the Positive2Negative paradigm, as shown in Figure 5. Firstly, we extend our intuitive idea into a flexible and upgraded version: Renoised Data Construction (RDC), which constructs multiple positive and negative noises. Secondly, we propose the Denoised Consistency Supervision (DCS), which supervises the consistency of the predicted denoised images to learn a denoising neural network.", + "bbox": [ + 88, + 359, + 480, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is important to note that we do not require the noise distribution to be strictly zero-mean or symmetric. Although real noise generally has a zero-mean, certain cases where the noise is not strictly symmetric can be easily addressed by adjusting the norm in the loss function [22].", + "bbox": [ + 89, + 479, + 482, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Renoised Data Construction (RDC)", + "text_level": 1, + "bbox": [ + 89, + 563, + 398, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The basic idea of RDC is to denoise the image to predict a denoised image and renoise the predicted denoised image by the predicted noise to construct multiple noisy images.", + "bbox": [ + 89, + 585, + 482, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a given noisy image $y = x + n$ , where $x$ presents the clean image and $n$ presents the noise. RDC firstly denoises the noisy image $y$ once through the neural network $\\mathcal{F}$ parametered by $\\theta$ :", + "bbox": [ + 89, + 632, + 483, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = \\mathcal {F} _ {\\theta} (y). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 702, + 480, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\hat{x}$ is the predicted denoised image of the noisy observation $y$ . Then we can calculate the predicted noise $\\hat{n}$ within the noisy observation $y$ :", + "bbox": [ + 89, + 728, + 482, + 772 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {n} = y - \\hat {x}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 784, + 480, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, RDC constructs noisy images as training data based on the predicted noise $\\hat{n}$ and the predicted denoised image $\\hat{x}$ . As noise is zero-mean, we further design a zero-mean sampling strategy to construct new noises. Specifically, RDC multiplies the predicted noise $\\hat{n}$ by both positive scale parameters $\\sigma_{p}$ and negative scale parameters $-\\sigma_{n}$", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sampled from Gaussian distribution $\\mathcal{N}$ , constructing multiscale Positive noise $n_p$ and Negative noise $n_n$ , respectively:", + "bbox": [ + 511, + 359, + 906, + 390 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {n}, \\sigma_ {p} \\sim \\mathcal {N} (1, \\sigma), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 417, + 906, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nn _ {p} = \\sigma_ {n} \\hat {n}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 436, + 903, + 453 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {p} = \\hat {x} + n _ {p} = \\hat {x} + \\sigma_ {n} \\hat {n}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 455, + 903, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nn _ {n} = - \\sigma_ {p} \\hat {n}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 476, + 903, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {n} = \\hat {x} + n _ {n} = \\hat {x} - \\sigma_ {p} \\hat {n}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 494, + 903, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Denoised Consistency Supervision (DCS)", + "text_level": 1, + "bbox": [ + 511, + 534, + 862, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The basic idea of DCS is that the noisy observations $y_{p}$ and $y_{n}$ should correspond to the same clean image. Therefore, the loss supervises the consistency of the predicted denoised images:", + "bbox": [ + 511, + 556, + 906, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p 2 n} = \\left\\| \\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) - \\mathcal {F} _ {\\theta} \\left(y _ {n}\\right) \\right\\|, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 631, + 903, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\|\\cdot\\|$ represents the norm. Considering the general nature of noise [22], we employ a gradually varying norm function $\\|\\cdot\\|_{2 \\to 1.5}$ to cover a broader range of unknown distributions:", + "bbox": [ + 511, + 660, + 906, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\| x \\| _ {2 \\rightarrow 1. 5} = (| x | + \\epsilon) ^ {\\gamma}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 734, + 903, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon = 10^{-8}$ and $\\gamma$ varies linearly from 2 to 1.5 during training. The advantage of the varying $\\| \\cdot \\|_{2\\to 1.5}$ norm is that $\\| \\cdot \\|_{2\\to 1.5}$ does not require fine-tuning of parameters, yet $\\| \\cdot \\|_{2\\to 1.5}$ consistently delivers robust performance.", + "bbox": [ + 511, + 763, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The noisy observations $y_{p}$ and $y_{n}$ contain multi-scale noise. Through DCS, the neural network $\\mathcal{F}_{\\theta}$ gradually learns to become insensitive to input noise and consistently outputs the clean signal, achieving the desired denoising performance.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17927", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/ba49c0365d71233d462fb23fff8e5d65d7c52dd6b5eb17f9d6fcac0d0cc00c32.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on SIDD, CC, PolyU and FMDD datasets. The best and second-best results (PSNR↑ / SSIM↑) are marked in red and blue in self-supervised single image denoising methods. The results of noisy dataset based self-supervised methods and supervised methods are also provided, only as a reference comparison. SIDD val and SIDD ben represent the SIDD validation dataset and the SIDD online benchmark, respectively." + ], + "table_footnote": [], + "table_body": "
CategoryMethodSIDD val [1]SIDD ben [1]CC [29]PolyU [49]FMDD [53]
SupervisedUnprocessing [4]28.12 / 0.52931.64 / 0.53135.18 / 0.90836.99 / 0.94128.45 / 0.574
MaskedDenoising [6]28.67 / 0.60431.99 / 0.60133.87 / 0.93034.56 / 0.93629.86 / 0.643
CLIPDenoising [10]34.79 / 0.86635.82 / 0.85936.30 / 0.94137.54 / 0.96030.56 / 0.698
Self-Supervised (Noisy Dataset Based Methods)Noise2Void [18]29.35 / 0.65127.68 / 0.66832.27 / 0.86233.83 / 0.873-
CVF-SID [30]34.81 / 0.87235.05 / 0.85629.10 /0.85233.05 / 0.91131.64 / 0.837
LUD-VAE [56]34.91 / 0.89235.49 / 0.88331.25 / 0.92033.68 / 0.94232.00 / 0.822
Self-Supervised (Single Image Based Methods)DIP [38]32.11 / 0.740-35.61 / 0.91237.17 / 0.91232.90 / 0.854
PD-denoising [57]33.97 / 0.820-35.85 / 0.92337.04 / 0.94033.01 / 0.856
NN+denoiser [55]--36.52 / 0.94337.66 / 0.95632.21 / 0.831
Self2Self [36]29.46 / 0.59529.51 / 0.65137.44 / 0.94837.52 / 0.95130.76 / 0.695
R2R [33]24.59 / 0.387-33.43 / 0.86536.23 / 0.93127.17 / 0.526
APBSN-single [21]30.90 / 0.818-27.72 / 0.89129.61 / 0.89728.43 / 0.804
ScoreDVI [9]34.75 / 0.85635.39 / 0.85937.09 / 0.94537.77 / 0.95933.10 / 0.865
ZS-N2N [27]25.59 / 0.43330.19 / 0.42833.51 / 0.85135.99 / 0.91431.65 / 0.767
MASH [11]35.06 / 0.85134.80 / 0.81431.17 / 0.88937.62 / 0.93233.71 / 0.882
ATBSN-single [8]32.39 / 0.79627.86 / 0.31231.87 / 0.89733.21 / 0.92531.23 / 0.829
Positive2Negative35.24 / 0.88935.80 / 0.87537.92 / 0.95438.20 / 0.96134.00 / 0.887
", + "bbox": [ + 91, + 155, + 906, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Analysis", + "text_level": 1, + "bbox": [ + 89, + 508, + 192, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we provide an analysis and proof demonstrating how the Positive2Negative paradigm achieves denoising.", + "bbox": [ + 89, + 531, + 482, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Taking the 2-norm as an example, the loss function can be expressed as:", + "bbox": [ + 89, + 575, + 482, + 606 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p 2 n} = \\mathbb {E} \\left[ \\| \\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) - \\mathcal {F} _ {\\theta} \\left(y _ {n}\\right) \\| ^ {2} \\right]. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 614, + 480, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through Taylor expansion, we can approximate the predicted denoised images $\\mathcal{F}_{\\theta}(y_p)$ and $\\mathcal{F}_{\\theta}(y_n)$ :", + "bbox": [ + 89, + 643, + 482, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) = \\mathcal {F} _ {\\theta} \\left(\\mathcal {F} _ {\\theta} (y)\\right) + \\sigma_ {p} \\cdot \\frac {\\partial \\mathcal {F} _ {\\theta} (y)}{\\partial y} \\cdot n + \\mathcal {O} \\left(n ^ {2}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 683, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\theta} (y _ {n}) = \\mathcal {F} _ {\\theta} \\left(\\mathcal {F} _ {\\theta} (y)\\right) - \\sigma_ {n} \\cdot \\frac {\\partial \\mathcal {F} _ {\\theta} (y)}{\\partial y} \\cdot n + \\mathcal {O} \\left(n ^ {2}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 717, + 482, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Substituting the approximations, we get:", + "bbox": [ + 109, + 758, + 380, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) - \\mathcal {F} _ {\\theta} \\left(y _ {n}\\right) \\approx \\left(\\sigma_ {p} + \\sigma_ {n}\\right) \\cdot \\frac {\\partial \\mathcal {F} _ {\\theta} (y)}{\\partial y} \\cdot n. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 781, + 482, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With further Taylor expansion, the loss function can be simplified as:", + "bbox": [ + 89, + 821, + 482, + 853 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p 2 n} \\approx (4 + 2 \\sigma^ {2}) \\left\\| \\frac {\\partial \\mathcal {F} _ {\\theta} (x)}{\\partial x} \\cdot n \\right\\| ^ {2}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 861, + 482, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through consistency loss $L_{p2n}$ , we optimize the network $\\mathcal{F}_{\\theta}$ such that $\\frac{\\partial\\mathcal{F}_{\\theta}(x)}{\\partial x}\\cdot n$ is optimized to drive towards zero, which means that the output $\\mathcal{F}_{\\theta}(y)$ tends to:", + "bbox": [ + 511, + 510, + 906, + 558 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {F} _ {\\theta} (y) = \\mathcal {F} _ {\\theta} (x + n), (17) \\\\ = \\mathcal {F} _ {\\theta} (x) + \\frac {\\partial \\mathcal {F} _ {\\theta} (x)}{\\partial x} \\cdot n + \\mathcal {O} \\left(n ^ {2}\\right), (18) \\\\ \\approx \\mathcal {F} _ {\\theta} (x). (19) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 510, + 906, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ideally, the neural network $\\mathcal{F}_{\\theta}$ has learned image priors, the model output $\\mathcal{F}_{\\theta}(x)$ matches the clean signal $x$ and we have $\\mathcal{F}_{\\theta}(x) = x$ . Thus, the ultimate goal achieves:", + "bbox": [ + 511, + 638, + 905, + 684 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\theta} (y) \\approx x. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 668, + 691, + 903, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the neural network $\\mathcal{F}_{\\theta}$ gradually learns to denoise the noisy image $y$ effectively.", + "bbox": [ + 511, + 713, + 906, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 755, + 645, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the experimental settings in Section 4.1. After that, we demonstrate the performance and comparisons with other methods in Section 4.2. Finally, we conduct a detailed ablation study in Section 4.3.", + "bbox": [ + 511, + 780, + 906, + 842 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 511, + 848, + 612, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we briefly introduce the experimental settings and details.", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17928", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c96a765d330fce118da6ea577a15bc3e09b17a8a64bbc57d17a9ea9203bf0cb6.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons of P2N against other self-supervised single image denoising methods in SIDD validation dataset. The denoised results of ScoreDVI are always flaky, with discontinuous lines and loss of texture. ZS-N2N does not completely denoise the images. MASH and ATBSN-single tend to either meet aliasing effect or blur the details. In summary, other methods meet aliasing effects and loss of texture, while Positive2Negative achieves both detailed texture preservation and comprehensive noise removal. In the figure, P2N represents Positive2Negative." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 893, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets details. Following previous methods [9, 11], we evaluate the proposed paradigm on the widely-used datasets in self-supervised single image denoising: the SIDD validation dataset [1], the SIDD benchmark dataset [1], CC [29], PolyU [49] and FMDD [53]. The SIDD, PolyU and CC datasets consist of natural sRGB images, while the FMDD dataset consists of fluorescence microscopy images.", + "bbox": [ + 89, + 512, + 483, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Following the default settings of self-supervised single image denoising methods [9, 11, 22, 40], we employ the same neural network architecture as Noise2Noise [22] and MASH [11] and initialize the neural network parameters with those pre-trained on DIV2K dataset with Gaussian noise [9, 55]. For training details, we employ AdamW [26] optimizer with a learning rate of 0.0001 and set scale parameter $\\sigma = 0.75$ , for all the experiments. No image enhancement techniques are applied.", + "bbox": [ + 89, + 638, + 483, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. Following previous methods [9, 11, 24], we mainly compare with self-supervised single image denoising methods. In addition, we also present the results of noisy dataset based self-supervised and supervised image denoising methods only as a reference comparison. All the results of the comparison methods are sourced from previous work [9, 11, 24] or evaluated by ourselves.", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 512, + 653, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we demonstrate the performance of Positive2Negative and compare Positive2Negative with other methods both quantitatively and qualitatively.", + "bbox": [ + 511, + 534, + 906, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparisons. The quantitative comparisons are described in Table 1. Previous self-supervised single image denoising methods have shown varying performance across different datasets, but Positive2Negative consistently achieves better results. In fact, Positive2Negative even outperforms some dataset-based and supervised methods. Notably, in the online evaluation of the SIDD benchmark dataset, we achieve an improvement of over $0.4\\mathrm{dB}$ .", + "bbox": [ + 511, + 580, + 905, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparisons. The qualitative comparisons are described in Figure 6. Clearly, the Positive2Negative paradigm achieves significant qualitative improvements by either enhancing details or effectively reducing noise. In contrast, other methods face limitations in both detail enhancement and noise removal due to the difficulty in learning denoising, resulting from the information-lossy barrier.", + "bbox": [ + 511, + 700, + 905, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 511, + 816, + 624, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first present the training process in Section 4.3.1. Then, we conduct ablation studies on the components and hyperparameters in Section 4.3.2 and Section 4.3.3, respectively.", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17929", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f39650f028abd0cdfee01828197ff698054cc50ec2be2a5690178a7b25e7184e.jpg", + "image_caption": [ + "Figure 7. Visual results of the training process. The top row shows the denoised results, while the bottom row displays the reconstructed negative noisy images in the training process. On the far right of each set is the reference image. The denoised image and reconstructed noisy image gradually become more like the reference images." + ], + "image_footnote": [], + "bbox": [ + 101, + 87, + 901, + 289 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1ad394bf256415ba94def0fe56a911aeb7ecc12a04955e4a4dd590880c3c7f81.jpg", + "table_caption": [ + "Table 2. Ablation of components. Each step is crucial for Positive2Negative. N2N and N2V represent Noise2Noise and Noise2Void, respectively." + ], + "table_footnote": [], + "table_body": "
MethodCC [29]SIDD [1]
PSNRSSIMPSNRSSIM
Baseline35.400.92733.500.849
RDC+N2N34.080.86825.420.414
RDC+N2V33.790.86124.510.368
N2N+DCS17.550.69824.200.815
N2V+DCS14.140.49422.300.743
RDC+DCS37.920.95435.240.889
", + "bbox": [ + 94, + 407, + 480, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1. Ablation of training", + "text_level": 1, + "bbox": [ + 89, + 571, + 276, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present visual results of the training process to illustrate the effectiveness of our Positive2Negative paradigm. As shown in Figure 7, throughout the training process, the denoised images and the reconstructed negative noisy images gradually converge towards their respective references.", + "bbox": [ + 89, + 590, + 483, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2. Ablation of components", + "text_level": 1, + "bbox": [ + 89, + 674, + 303, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of each component of the Positive2Negative paradigm on the CC and SIDD datasets. RDC constructs precise training data and DCS trains robust denoising neural networks. Both RDC and DCS are crucial for Positive2Negative. Replacing either component would result in a decline in performance, as shown in Table 2.", + "bbox": [ + 89, + 693, + 483, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.3. Ablation of hyperparameters", + "text_level": 1, + "bbox": [ + 89, + 791, + 339, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Throughout the entire training process, the only adjustable hyperparameter is the scale parameter $\\sigma$ , which is set to 0.75 for all experiments. As shown in Table 3, the performance remains similar for $\\sigma$ values ranging from 0.25 to 0.75. The scale parameter $\\sigma$ introduces a dynamic and wide range of noise levels, while the \"Fixed\" setting corresponds to $\\sigma = 0$ .", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c56ecc646ee4af664bf995b4ee2059e1505f4fa791710672117d01848c09bcdc.jpg", + "table_caption": [ + "Table 3. Ablation of hyperparameters. The scale parameter is the only hyperparameter and is very robust. We fix the scale parameter $\\sigma = 0.75$ for all the experiments." + ], + "table_footnote": [], + "table_body": "
MethodCC [29]SIDD [1]
PSNRSSIMPSNRSSIM
Fixed37.330.95235.240.892
σ = 0.2537.640.94535.290.893
σ = 0.5037.840.95235.300.891
σ = 0.7537.920.95435.240.889
σ = 1.0037.860.95435.110.884
", + "bbox": [ + 519, + 407, + 903, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussions", + "text_level": 1, + "bbox": [ + 513, + 546, + 633, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Distribution", + "text_level": 1, + "bbox": [ + 513, + 571, + 643, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different norms are suitable for different noise distributions [22]. For simplicity, we have uniformly adopted the $\\| \\cdot \\|_{2\\to 1.5}$ norm, as real-world noise is typically centered around a zero mean with slight fluctuations. For noise that remains nearly zero-mean, such as in PolyU [49], the $\\| \\cdot \\|_{2\\to 1.5}$ norm and the $\\| \\cdot \\|_2$ norm yield similar performance. For noise that exhibits slight deviations from zero mean, as in SIDD [1], the $\\| \\cdot \\|_{2\\to 1.5}$ norm can outperform the $\\| \\cdot \\|_2$ norm, as shown in Table 4. The slight increase in PSNR further suggests that the deviation remains minimal.", + "bbox": [ + 511, + 595, + 906, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Convergence", + "text_level": 1, + "bbox": [ + 511, + 757, + 648, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present the PSNR variation over iterations on SIDD dataset [1] in Figure 8. Around the 100th iteration, the Positive2Negative paradigm converges and shows minimal fluctuations thereafter. Additionally, it is worth noting that the constructed data by RDC is unsuitable for Noise2Noise and Noise2Void. This is mainly because the noise in constructed images by RDC is correlated, which would quickly collapse the training processes of the other two methods.", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17930", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2936e11417bd1b2ee4f56634df75f54ad813229a1460fb7d7f7e20de18fad3d9.jpg", + "table_caption": [ + "Table 4. Discussion of distributions. Different norms are suitable for different noise distributions [22]. For simplicity, we have uniformly adopted the $\\| \\cdot \\|_{2\\to 1.5}$ norm." + ], + "table_footnote": [], + "table_body": "
Dataset||·||0||·||1||·||2||·||2→1.5
PolyU [49]37.8838.0238.2038.20
SIDD [1]35.1235.2635.2035.24
", + "bbox": [ + 91, + 142, + 483, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fee973ec3e3634645272cf323da138f5bf041cb7f4e5b2000e11c5bb1133e3e7.jpg", + "image_caption": [ + "Figure 8. Convergence comparisons. P2N converges quickly and remains stable after convergence on SIDD. The P2N-Fixed version (where $\\sigma = 0.00$ ), in particular, even achieves rapid convergence in approximately 10 iterations. P2N represents Positive2Negative." + ], + "image_footnote": [], + "bbox": [ + 93, + 215, + 285, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a52c55de9b69ce102160154975397945498e416d8127e52e2b4c8069d26260d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 215, + 480, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Efficiency", + "text_level": 1, + "bbox": [ + 89, + 395, + 205, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For self-supervised single image denoising methods, computational efficiency is crucial, as each noisy image must be trained individually. Here, we demonstrate the number of training iterations, parameter, and computational cost, as detailed in Table 5. It is obvious that our paradigm achieves significant efficiency improvements.", + "bbox": [ + 89, + 417, + 483, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Comparison with Noise2Noise and Noise2Void", + "text_level": 1, + "bbox": [ + 89, + 522, + 480, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compared with Noise2Noise [22] and Noise2Void [18], Positive2Negative also presents an elegant paradigm The key differences are present in the following aspects.", + "bbox": [ + 89, + 545, + 482, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Firstly, for the constructed data, the synthetic data constructed by Positive2Negative breaks the information-lossy barrier of Noise2Noise and Noise2Void. Secondly, for the supervision strategy, Positive2Negative enforces consistency between denoised results, whereas Noise2Noise and Noise2Void supervise the consistency between a denoised result and a noisy image. Thirdly, from the assumption perspective, Noise2Noise, Noise2Void, and Positive2Negative all have theoretical guarantees, while Noise2Noise and Noise2Void operate under more restrictive assumptions. Noise2Noise requires two independent noisy images, while Noise2Void assumes that the noise is spatially correlated.", + "bbox": [ + 89, + 592, + 482, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Differences with R2R and CVF-SID", + "text_level": 1, + "bbox": [ + 89, + 786, + 403, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the constructed data, R2R employs new Gaussian noise and applies reciprocal coefficients to the noisy image, while CVF-SID employs two networks to predict three components and applies fixed manual rules. In contrast, our method employs a single network to predict noise and applies random multiscale symmetrical parameters.", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f9b10a56752869f8379e809419f1c3ac0eb7f031cce9a329b3f40605ab823c11.jpg", + "table_caption": [ + "Table 5. Efficiency comparisons under input images with dimensions of $256 \\times 256 \\times 3$ . With similar parameters and FLOPs, P2N requires the fewest training iterations. P2N converges rapidly, and the P2N-Fixed version (where $\\sigma = 0.00$ ) converges even just within 20 iterations. P2N represents Positive2Negative." + ], + "table_footnote": [], + "table_body": "
MethodIterationsParams (M)FLOPs (G)
N2N [22]-0.7018.7
B2U [40]-1.1018.7
Self2Self [36]4500001.009.6
R2R [33]80000.5636.7
ScoreDVI [9]40013.537.9
MASH [11]800 × 30.9911.4
P2N-Fixed<200.9911.4
P2N<1000.9911.4
", + "bbox": [ + 517, + 169, + 903, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the supervision strategy, R2R maps the denoised image to the noisy input, while CVF-SID primarily maps the first cyclic result to the noisy input and maps the second cyclic result to the first cyclic result. In contrast, our method maps the denoised image to another denoised image.", + "bbox": [ + 511, + 354, + 906, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Limitation and Future Work", + "text_level": 1, + "bbox": [ + 511, + 446, + 769, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The primary limitation is that Positive2Negative has only been validated on the single image based self-supervised image denoising task. To prevent convergence to trivial solutions such as zero mapping or identity mapping, Positive2Negative requires training with a pre-trained Gaussian denoising model. While the single image based self-supervised denoising task allows training from a pre-trained model on Gaussian noise, the noisy dataset based self-supervised image denoising task trains from scratch. Therefore, extending the Positive2Negative paradigm to the noisy dataset based self-supervised image denoising task could be a promising direction for future work.", + "bbox": [ + 511, + 470, + 908, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 672, + 633, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a self-supervised single image denoising paradigm, Positive2Negative, to break information-lossy barrier. Specifically, Positive2Negative includes two steps: RDC and DCS. Positive2Negative achieves SOTA performance compared to existing methods across various public benchmarks.", + "bbox": [ + 511, + 700, + 906, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 811, + 671, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China under Grants 62322204, 62131003, and the Key Laboratory of Target Cognition and Application Technology.", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17931", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S. Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1700, 2018. 1, 3, 5, 6, 7, 8", + "[2] Joshua Batson and Loic Royer. Noise2Self: Blind denoising by self-supervision. In International Conference on Machine Learning (ICML), pages 524-533, 2019. 3", + "[3] Chinmay Belthangady and Loic A Royer. Applications, promises, and pitfalls of deep learning for fluorescence image reconstruction. In Nature Methods, pages 1215-1225, 2019. 1", + "[4] Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing images for learned raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11036-11045, 2019. 5", + "[5] Jaeseok Byun, Sungmin Cha, and Taesup Moon. Fbidenoiser: Fast blind image denoiser for poisson-gaussian noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5768-5777, 2021. 3", + "[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1703, 2023. 5", + "[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision (ECCV), pages 17-33, 2022. 1", + "[8] Shiyan Chen, Jiyuan Zhang, Zhaofei Yu, and Tiejun Huang. Exploring efficient asymmetric blind-spots for self-supervised denoising in real-world scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2814-2823, 2024. 1, 2, 3, 5, 6", + "[9] Jun Cheng, Tao Liu, and Shan Tan. Score priors guided deep variational inference for unsupervised real-world single image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12937-12948, 2023. 1, 3, 5, 6, 8", + "[10] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25974-25984, 2024. 5", + "[11] Hamadi Chihaoui and Paolo Favaro. Masked and shuffled blind spot denoising for real-world images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3025-3034, 2024. 1, 2, 3, 5, 6, 8", + "[12] Hansen Feng, Lizhi Wang, Yuzhi Wang, Haoqiang Fan, and Hua Huang. Learnability enhancement for low-light raw image denoising: A data perspective. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 370-387, 2024. 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. In ACM Transactions on Graphics (ToG), pages 1-12, 2016. 1", + "[14] Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. Neighbor2Neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14781-14790, 2021. 2", + "[15] Yeong Il Jang, Keuntek Lee, Gu Yong Park, Seyun Kim, and Nam Ik Cho. Self-supervised image denoising with downsampled invariance loss and conditional blind-spot network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12196-12205, 2023. 3", + "[16] Kwanyoung Kim and Jong Chul Ye. Noise2score: tweedie's approach to self-supervised image denoising without clean images. In Advances in Neural Information Processing Systems, pages 864-874, 2021. 3", + "[17] Kwanyoung Kim, Taesung Kwon, and Jong Chul Ye. Noise distribution adaptive self-supervised image denoising using tweedie distribution and score matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1998-2006, 2022. 3", + "[18] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void: Learning denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2129-2137, 2019. 1, 2, 3, 5, 8", + "[19] Alexander Krull, Tomáš Vičar, Mangal Prakash, Manan Lalit, and Florian Jug. Probabilistic noise2void: Unsupervised content-aware denoising. In Frontiers in Computer Science, pages 2-5, 2020. 3", + "[20] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. In Proceedings of International Conference on Neural Information Processing Systems (NeurIPS), pages 6970-6980, 2019. 3", + "[21] Wooseok Lee, Sanghyun Son, and Kyoung Mu Lee. APBSN: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17725-17734, 2022. 2, 3, 5", + "[22] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2Noise: Learning image restoration without clean data. In International Conference on Machine Learning (ICML), pages 4620-4631, 2018. 1, 2, 3, 4, 6, 7, 8", + "[23] Junyi Li, Zhilu Zhang, Xiaoyu Liu, Chaoyu Feng, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Spatially adaptive self-supervised learning for real-world image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914-9924, 2023. 3", + "[24] Tong Li, Hansen Feng, Lizhi Wang, Zhiwei Xiong, and Hua Huang. Stimulating the diffusion model for image denoising via adaptive embedding and ensembling. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 8240-8257, 2024. 1, 3, 6" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17932", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Xinyang Li, Yixin Li, Yiliang Zhou, Jiamin Wu, Zhifeng Zhao, Jiaqi Fan, Fei Deng, Zhaofa Wu, Guihua Xiao, Jing He, Yuanlong Zhang, Guoxun Zhang, Xiaowan Hu, Zhang Yi Chen, Xingye and, Hui Qiao, Hao Xie, Yulong Li, Haoqian Wang, Lu Fang, and Qionghai Dai. Real-time denoising enables high-sensitivity fluorescence time-lapse imaging beyond the shot-noise limit. In Nature Biotechnology, page 282-292, 2023. 1", + "[26] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. In arXiv preprint arXiv:1711.05101, pages 1-10, 2017. 6", + "[27] Youssef Mansour and Reinhard Heckel. Zero-Shot Noise2Noise: Efficient image denoising without any data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14018-14027, 2023. 1, 2, 3, 5, 6", + "[28] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2Noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12064-12072, 2020. 2", + "[29] Seonghyeon Nam, Youngbae Hwang, Yasuyuki Matsushita, and Seon Joo Kim. A holistic approach to cross-channel image noise modeling and its application to image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1683-1691, 2016. 5, 6, 7", + "[30] Reyhaneh Neshatavar, Mohsen Yavartanoo, Sanghyun Son, and Kyoung Mu Lee. CVF-SID: Cyclic multi-variate function for self-supervised image denoising by disentangling noise from image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17583-17591, 2022. 3, 5", + "[31] Harry Nyquist. Certain topics in telegraph transmission theory. In Transactions of the American Institute of Electrical Engineers, pages 617-644, 1928. 1, 2", + "[32] Yizhong Pan, Xiao Liu, Xiangyu Liao, Yuanzhouhan Cao, and Chao Ren. Random sub-samples generation for self-supervised real image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12150-12159, 2023. 2, 3", + "[33] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-Recorrupted: Unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2043–2052, 2021. 2, 3, 5, 8", + "[34] Mikhail Papkov, Pavel Chizhov, and Leopold Parts. Swinia: Self-supervised blind-spot image denoising without convolutions. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 7071-7080, 2025. 3", + "[35] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1586-1595, 2017. 1", + "[36] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2Self with dropout: Learning self-supervised denoising" + ], + "bbox": [ + 91, + 90, + 480, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "from single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1890-1898, 2020. 3, 5, 8", + "[37] Claude Elwood Shannon. A mathematical theory of communication. In The Bell System Technical Journal, pages 379-423, 1948. 1, 2", + "[38] Andrea Vedaldi Ulyanov Dmitry and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018. 3, 5", + "[39] Yuzhi Wang, Haibin Huang, Qin Xu, Jiaming Liu, Yiqun Liu, and Jue Wang. Practical deep raw image denoising on mobile devices. In European Conference on Computer Vision (ECCV), pages 1-16, 2020. 1", + "[40] Zejin Wang, Jiazheng Liu, Guoqing Li, and Hua Han. Blind2Unblind: Self-supervised image denoising with visible blind spots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2027-2036, 2022. 2, 3, 6, 8", + "[41] Zichun Wang, Ying Fu, Ji Liu, and Yulun Zhang. LG-BPN: Local and global blind-patch network for self-supervised real-world denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18156-18165, 2023. 2, 3", + "[42] Kaixuan Wei, Ying Fu, Yinqiang Zheng, and Jiaolong Yang. Physics-based noise modeling for extreme low-light photography. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 8520-8537, 2021. 1, 3", + "[43] Martin Weigert, Uwe Schmidt, Tobias Boothe, Andreas Müller, Alexandr Dibrov, Akanksha Jain, Benjamin Wilhelm, Deborah Schmidt, Coleman Broaddus, Sián Culley, et al. Content-aware image restoration: pushing the limits of fluorescence microscopy. In Nature Methods, pages 1090–1097, 2018. 1", + "[44] Halbert White. Connectionist nonparametric regression: Multilayer feedforward networks can learn arbitrary mappings. In Neural networks, pages 535-549, 1990. 2", + "[45] David Wu, Damian Tan, Marilyn Baird, John DeCampo, Chris White, and Hong Ren Wu. Perceptually lossless medical image coding. In IEEE Transactions on Medical Imaging, pages 335-344, 2006. 2", + "[46] Hong Ren Wu, Weisi Lin, and Lina J. Karam. An overview of perceptual processing for digital pictures. In IEEE International Conference on Multimedia and Expo Workshops, pages 113-120, 2012. 2", + "[47] Xiaohe Wu, Ming Liu, Yue Cao, Dongwei Ren, and Wangmeng Zuo. Unpaired learning of deep image denoising. In European Conference on Computer Vision (ECCV), pages 352-368, 2020. 3", + "[48] Yaochen Xie, Zhengyang Wang, and Shuiwang Ji. Noise2same: Optimizing a self-supervised bound for image denoising. In Advances in Neural Information Processing Systems, pages 20320-20330, 2020. 3", + "[49] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. In arXiv preprint arXiv:1804.02603, pages 1-13, 2018. 5, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17933", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-Clean: Learning self-supervised denoising from corrupted image. In IEEE Transactions on Image Processing (TIP), pages 9316-9329, 2020. 2", + "[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728-5739, 2022. 1", + "[52] Dan Zhang, Fangfang Zhou, Yuwen Jiang, and Zhengming Fu. Mm-bsn: Self-supervised image denoising for real-world with multi-mask based on blind-spot network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 4189–4198, 2023. 3", + "[53] Yide Zhang, Yinhao Zhu, Evan Nichols, Qingfei Wang, Siyuan Zhang, Cody Smith, and Scott Howard. Aoisson-gaussian denoising dataset with real fluorescence microscopy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11710-11718, 2019. 5, 6", + "[54] Yi Zhang, Dasong Li, Ka Lung Law, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. IDR: Self-supervised image denoising via iterative data refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2098-2107, 2022. 2", + "[55] Dihan Zheng, Sia Huat Tan, Xiaowen Zhang, Zuoqiang Shi, Kaisheng Ma, and Chenglong Bao. An unsupervised deep learning approach for real-world image denoising. In International Conference on Learning Representations (ICLR), pages 1-9, 2020. 5, 6", + "[56] Dihan Zheng, Xiaowen Zhang, Kaisheng Ma, and Chenglong Bao. Learn from unpaired data for image restoration: A variational bayes approach. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 5889-5903, 2022. 3, 5", + "[57] Yuqian Zhou, Jianbo Jiao, Haibin Huang, Yang Wang, Jue Wang, Honghui Shi, and Thomas Huang. When awgn-based denoiser meets real noises. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 13074-13081, 2020. 3, 5", + "[58] Yunhao Zou, Chenggang Yan, and Ying Fu. Iterative denoiser and noise estimator for self-supervised image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13265-13274, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 752 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17934", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/867e9811-fe1d-4217-8c41-8e621b6da43b_model.json b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/867e9811-fe1d-4217-8c41-8e621b6da43b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c88c3f36d14258b5e84a8494d96eba51aed6ae8a --- /dev/null +++ b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/867e9811-fe1d-4217-8c41-8e621b6da43b_model.json @@ -0,0 +1,2719 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.131, + 0.8, + 0.177 + ], + "angle": 0, + "content": "Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.203, + 0.928, + 0.223 + ], + "angle": 0, + "content": "Tong Li\\(^{1}\\) Lizhi Wang\\(^{2,3*}\\) Zhiyuan Xu\\(^{1}\\) Lin Zhu\\(^{1}\\) Wanxuan Lu\\(^{4}\\) Hua Huang\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.222, + 0.845, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science and Technology, Beijing Institute of Technology \\(^{2}\\) School of Artificial Intelligence, Beijing Normal University" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.257, + 0.962, + 0.275 + ], + "angle": 0, + "content": "3 Engineering Research Center of Intelligent Technology and Educational Application, Ministry of Education" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.275, + 0.825, + 0.293 + ], + "angle": 0, + "content": "\\(^{4}\\)Aerospace Information Research Institute, Chinese Academy of Sciences" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.485, + 0.707 + ], + "angle": 0, + "content": "Image denoising enhances image quality, serving as a foundational technique across various computational photography applications. The obstacle to clean image acquisition in real scenarios necessitates the development of self-supervised image denoising methods only depending on noisy images, especially a single noisy image. Existing self-supervised image denoising paradigms (Noise2Noise and Noise2Void) rely heavily on information-lossy operations, such as downsampling and masking, culminating in low-quality denoising performance. In this paper, we propose a novel self-supervised single image denoising paradigm, Positive2Negative, to break the information-lossy barrier. Our paradigm involves two key steps: Renoised Data Construction (RDC) and Denoised Consistency Supervision (DCS). RDC renoes the predicted denoised image by the predicted noise to construct multiple noisy images, preserving all the information of the original image. DCS ensures consistency across the multiple denoised images, supervising the network to learn robust denoising. Our Positive2Negative paradigm achieves state-of-the-art performance in self-supervised single image denoising with significant speed improvements. The code is released to the public at https://github.com/Li-Tong-621/P2N." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.222, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Image denoising plays a vital role in image processing to enhance visual quality [13], serving as a foundational technique across various computational photography applications [3, 39]. The widespread popularity of paired clean-noisy image datasets [1, 35] has catalyzed advancements in supervised methods, yielding remarkable performance [7, 51]. However, supervised methods fall short in" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.328, + 0.905, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.433, + 0.783, + 0.444 + ], + "angle": 0, + "content": "Noise2Noise-Based Paradigms" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.446, + 0.903, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.551, + 0.777, + 0.562 + ], + "angle": 0, + "content": "Noise2Void-Based Paradigms" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.563, + 0.902, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.666, + 0.774, + 0.677 + ], + "angle": 0, + "content": "Positive2Negative Paradigm" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.688, + 0.907, + 0.773 + ], + "angle": 0, + "content": "Figure 1. Information-lossy operations in current paradigms. The noise addition operation reduces the signal-to-noise ratio, the downsampling operation diminishes the sampling density, and the masking operation disregards crucial central pixels. Each of these operations introduces an information-lossy barrier [8, 31, 37], culminating in imprecise denoising." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.907, + 0.882 + ], + "angle": 0, + "content": "dynamic scenes [18, 22], where long exposure results in misalignment and blur [22, 42], hindering the acquisition of clean images. Therefore, advancing self-supervised image denoising methods using only noisy images [25, 43], especially a single noisy image [9, 11, 24, 27], holds significant importance." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The crux of self-supervised image denoising methods" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.438, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author: Lizhi Wang (wanglizhi@bnu.edu.cn)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17924" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.48, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.306 + ], + "angle": 0, + "content": "Figure 2. The information-lossy barrier significantly compromises the image quality, resulting in residual noise, aliasing effects or texture loss." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.483, + 0.549 + ], + "angle": 0, + "content": "lies in the construction of the training data and the formulation of the supervision strategy. Current methods are divided into Noise2Noise-based paradigms and Noise2Void-based paradigms, as illustrated in Figure 1. Noise2Noise-based paradigms add additional noise to the noisy image [28, 33] or down sample the noisy image [14, 27] to approximately construct two independently noisy observations as training data. These paradigms learn denoising by supervising one observation with the denoised result of other observation. Noise2Void-based paradigms mask either the noisy image [18, 40] or the neural network [8, 21, 41] to construct masked training data. These paradigms learn denoising by supervising the denoised result of the masked central pixel with the original noisy pixel, where the prediction only relies on the surrounding pixels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.779 + ], + "angle": 0, + "content": "However, existing methods rely heavily on information-lossy operations, resulting in low-quality denoising performance [21, 32, 40, 41]. Specifically, the noise addition operation reduces the signal-to-noise ratio, the downsampling operation diminishes the sampling density and the masking operation disregards crucial central pixels. Each of these operations loses the information from the original noisy image [8, 31, 37], inevitably compromising the denoising ability [41, 44]. Consequently, denoised images always suffer from remaining noise, aliasing effects or texture loss [45, 46], as shown in Figure 2. As these information-lossy operations constitute the indispensable foundation of the Noise2Noise-based and Noise2Void-based paradigms, it is imperative to establish a new paradigm that transcends the Noise2Noise-based and Noise2Void-based paradigms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose the concise Positive2Negative paradigm to break the information-lossy barrier in self-supervised single image denoising. Our inspiration stems from the observation that noise distributions are approximately symmetrical and centered around zero, implying that the opposite noise also follows the distribution of the original noise. Leveraging this insight, we can construct multiple noisy images corresponding to the same clean image, which" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "serves as the basis of Positive2Negative. Concretely, Positive2Negative relies on two steps: Renoised Data Construction (RDC) and Denoised Consistency Supervision (DCS)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.905, + 0.258 + ], + "angle": 0, + "content": "The basic idea of RDC is to construct training data. Specifically, we design a zero-mean sampling strategy to generate multi-scale Positive noise and Negative noise based on the predicted noise, which are then added back to construct Positive noisy images and Negative noisy images. The constructed data comprehensively covers the information of the given single image; thus the Positive2Negative paradigm breaks the information-lossy barrier." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.363 + ], + "angle": 0, + "content": "The basic idea of DCS is to formulate a supervision strategy. Specifically, given the constructed multi-scale noisy images, DCS supervises the network to generate consistent denoised outputs during the training process. Furthermore, we provide a theoretical analysis showing that the Positive2Negative paradigm can learn robust denoising, through supervising the multiple denoised images." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.364, + 0.835, + 0.378 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.381, + 0.905, + 0.425 + ], + "angle": 0, + "content": "- We propose a new self-supervised single image denoising paradigm, Positive2Negative, to break the information-lossy barrier." + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.426, + 0.905, + 0.455 + ], + "angle": 0, + "content": "- We propose a data construction method, which constructs multi-scale similar noisy images for training." + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.456, + 0.905, + 0.485 + ], + "angle": 0, + "content": "- We propose a denoising supervision method, which is theoretically guaranteed to learn robust denoising." + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.486, + 0.905, + 0.516 + ], + "angle": 0, + "content": "- Positive2Negative achieves SOTA performance over self-supervised single image denoising." + }, + { + "type": "list", + "bbox": [ + 0.537, + 0.381, + 0.905, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.655, + 0.544 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.905, + 0.599 + ], + "angle": 0, + "content": "From the training data perspective, the self-supervised image denoising methods can be classified into noisy dataset based methods and single image based methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.776, + 0.623 + ], + "angle": 0, + "content": "2.1. Noisy Dataset based Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.735 + ], + "angle": 0, + "content": "Noisy dataset based methods represent the most active research direction in image denoising. These methods learn to denoise from a given noisy dataset. Generally, a larger volume of the noisy dataset improves the denoising performance. As a result, in scenarios where only a single noisy image is available, the denoising performance of whole noisy dataset-based methods decreases significantly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Noise2Noise [22] learns denoising by mapping one noisy observation to another independent observation of the same scene. Subsequent studies focus on generating observations of the same scene from a single observation, typically achieved through downsampling [14, 27] or by introducing additional noise [28, 33, 50, 54, 58]. After that, another elegant method Noise2Void [18] is proposed. Noise2Void [18] assumes that the noise distribution is zero-mean, the noise is spatially independent, and the signal is spatially correlated. Noise2Void [18] predicts the clean signal by leveraging surrounding noisy signals, with mask" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17925" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "ing the central pixel. Various other works propose alternative masking schemes, either mask the input [11, 36, 40] or mask the network [2, 19, 20, 34, 48, 52]. Noise2Void struggles with real-world noise, which often exhibits spatial correlation. Therefore, subsequent research endeavors to break the noise spatial correlation [8, 21, 23, 41, 47, 57]. There are also some methods that amalgamate frameworks established by both Noise2Noise and Noise2Void [15, 32] or get rid of frameworks established by Noise2Noise and Noise2Void [5, 16, 17, 30, 56], but these methods are usually less developed due to the complexity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.267, + 0.347, + 0.283 + ], + "angle": 0, + "content": "2.2. Single Image based Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.394 + ], + "angle": 0, + "content": "Single image based methods represent the most challenging situation for denoising and the least demanding requirements for data. These methods learn to denoise from a given single noisy image without any dependence on paired noisy-clean or noisy datasets. As a result, single image based methods are less developed due to the inherent difficulty, only gaining attention in the past two years." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.591 + ], + "angle": 0, + "content": "DIP [38] is the earliest self-supervised image denoising method. DIP [38] learns to map random noise to a given noisy image and employs early stopping to prevent overfitting. Self2Self [36] uses dropout to create two Bernoulli-sampled observations of a noisy image. R2R [33] constructs noisy images by adding new Gaussian noise and applying reciprocal coefficients. ScoreDVI [9] performs real denoising using a pre-trained Gaussian denoiser and an estimated noise model. ZS-N2N [27] extend Noise2Noise [22] to a single image version by downsampling. MASH [11] employs masking and local pixel shuffling to break the noise spatial correlation. Recently, DMID stimulates the diffusion model for self-supervised single image denoising [24]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.636 + ], + "angle": 0, + "content": "In this paper, we propose Positive2Negative, a new self-supervised single image denoising paradigm, to break the information-lossy barrier." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.264, + 0.666 + ], + "angle": 0, + "content": "3. Positive2Negative" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.75 + ], + "angle": 0, + "content": "In this section, we first demonstrate the motivation of Positive2Negative in Section 3.1. After that, we present the two critical steps of Positive2Negative: RDC and DCS, respectively in Section 3.2 and Section 3.3. Finally, we provide an analysis and proof of Positive2Negative in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.212, + 0.773 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our inspiration stems from the widely recognized assumptions that the noise distribution is zero-mean and approximately symmetrical [18, 22], which is consistent with our observation. In our observation experiment, we calculate the noise distribution of a fixed pixel across multiple observations of the same scene. Specifically, following SIDD [1], ELD [42] and PMN [12], we collect noisy image \\(y\\) and clean image \\(c\\). Then we calculate the noise value \\(n\\) at each" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.262, + 0.765, + 0.273 + ], + "angle": 0, + "content": "(a). noisy image" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.273, + 0.707, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.357, + 0.667, + 0.369 + ], + "angle": 0, + "content": "(b). green circle" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.273, + 0.903, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.357, + 0.86, + 0.369 + ], + "angle": 0, + "content": "(c). red circle" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.38, + 0.906, + 0.464 + ], + "angle": 0, + "content": "Figure 3. Noise distribution is zero-mean and approximately symmetrical. (a) shows a noisy image. (b) and (c) show the noise distributions, which are calculated at the center pixels marked by the green circle and the red circle in the noisy image, respectively. It is evident that the noise distribution is zero-mean and approximately symmetrical." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.478, + 0.616, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.552, + 0.612, + 0.565 + ], + "angle": 0, + "content": "Original noise" + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.478, + 0.712, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.553, + 0.706, + 0.565 + ], + "angle": 0, + "content": "Original noisy" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.478, + 0.805, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.553, + 0.803, + 0.565 + ], + "angle": 0, + "content": "Opposite noise" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.478, + 0.901, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.808, + 0.553, + 0.9, + 0.565 + ], + "angle": 0, + "content": "Opposite noisy" + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.567, + 0.575, + 0.577 + ], + "angle": 0, + "content": "n" + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.567, + 0.696, + 0.579 + ], + "angle": 0, + "content": "\\(y = c + n\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.567, + 0.789, + 0.579 + ], + "angle": 0, + "content": "\\(n_{n} = -n\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.567, + 0.9, + 0.58 + ], + "angle": 0, + "content": "\\(y_{n} = c + (-n)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.592, + 0.906, + 0.634 + ], + "angle": 0, + "content": "Figure 4. The opposite noise \\(-n\\) is similar to the original noise \\(n\\). The opposite noise \\(-n\\) approximately follows the original noise's distribution, which is zero-mean and approximately symmetrical." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.72 + ], + "angle": 0, + "content": "spatial location \\((i,j)\\). Additionally, we compute the frequency of \\(n_{i,j}\\) cross noisy images of the same scene. The results are shown in Figure 3, which demonstrates the noise distribution is zero-mean and approximately symmetrical." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.906, + 0.797 + ], + "angle": 0, + "content": "Based on this observation, an intuitive insight is that the opposite noise \\(-n\\) also follows the same distribution as that of the original noise \\(n\\). This intuitive idea inspires us to construct the opposite noise \\(n_{n}\\) and corresponding opposite noisy image \\(y_{n}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.811, + 0.905, + 0.825 + ], + "angle": 0, + "content": "\\[\nn _ {n} = - n, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.829, + 0.905, + 0.844 + ], + "angle": 0, + "content": "\\[\ny _ {n} = x + n _ {n} = x - n. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The opposite noise \\( n_n \\) and opposite noisy image \\( y_n \\) are similar to the original noise \\( n \\) and original noisy image \\( y \\), respectively, as shown in Figure 4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17926" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.09, + 0.904, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.249, + 0.907, + 0.336 + ], + "angle": 0, + "content": "Figure 5. Overview of the proposed self-supervised single image denoising paradigm. For training, the Positive2Negative paradigm consists of 2 steps. The first step is Renoised Data Construction (RDC): multiply the predicted noise by positive and negative coefficients to construct multi-scale Positive noise and Negative noise, respectively. The second step is Denoised Consistency Supervision (DCS): train the network through supervising the consistency of the denoised images. For inference, denoising can be achieved in just one pass through the network. (The networks shown in the figure are the same one. In addition, for better visualization, the images within the red boxes on the bottom left have been zoomed in and are displayed in the red boxes on the bottom right.)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.36, + 0.482, + 0.48 + ], + "angle": 0, + "content": "Based on the above observations and insights, we propose the Positive2Negative paradigm, as shown in Figure 5. Firstly, we extend our intuitive idea into a flexible and upgraded version: Renoised Data Construction (RDC), which constructs multiple positive and negative noises. Secondly, we propose the Denoised Consistency Supervision (DCS), which supervises the consistency of the predicted denoised images to learn a denoising neural network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.483, + 0.556 + ], + "angle": 0, + "content": "It is important to note that we do not require the noise distribution to be strictly zero-mean or symmetric. Although real noise generally has a zero-mean, certain cases where the noise is not strictly symmetric can be easily addressed by adjusting the norm in the loss function [22]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.564, + 0.4, + 0.58 + ], + "angle": 0, + "content": "3.2. Renoised Data Construction (RDC)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.483, + 0.632 + ], + "angle": 0, + "content": "The basic idea of RDC is to denoise the image to predict a denoised image and renoise the predicted denoised image by the predicted noise to construct multiple noisy images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.484, + 0.692 + ], + "angle": 0, + "content": "For a given noisy image \\( y = x + n \\), where \\( x \\) presents the clean image and \\( n \\) presents the noise. RDC firstly denoises the noisy image \\( y \\) once through the neural network \\( \\mathcal{F} \\) parametered by \\( \\theta \\):" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.703, + 0.482, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = \\mathcal {F} _ {\\theta} (y). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Here \\(\\hat{x}\\) is the predicted denoised image of the noisy observation \\(y\\). Then we can calculate the predicted noise \\(\\hat{n}\\) within the noisy observation \\(y\\):" + }, + { + "type": "equation", + "bbox": [ + 0.247, + 0.785, + 0.482, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\hat {n} = y - \\hat {x}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Next, RDC constructs noisy images as training data based on the predicted noise \\(\\hat{n}\\) and the predicted denoised image \\(\\hat{x}\\). As noise is zero-mean, we further design a zero-mean sampling strategy to construct new noises. Specifically, RDC multiplies the predicted noise \\(\\hat{n}\\) by both positive scale parameters \\(\\sigma_{p}\\) and negative scale parameters \\(-\\sigma_{n}\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.36, + 0.907, + 0.391 + ], + "angle": 0, + "content": "sampled from Gaussian distribution \\(\\mathcal{N}\\), constructing multiscale Positive noise \\(n_p\\) and Negative noise \\(n_n\\), respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.418, + 0.907, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {n}, \\sigma_ {p} \\sim \\mathcal {N} (1, \\sigma), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.438, + 0.905, + 0.454 + ], + "angle": 0, + "content": "\\[\nn _ {p} = \\sigma_ {n} \\hat {n}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.457, + 0.905, + 0.472 + ], + "angle": 0, + "content": "\\[\ny _ {p} = \\hat {x} + n _ {p} = \\hat {x} + \\sigma_ {n} \\hat {n}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.477, + 0.905, + 0.492 + ], + "angle": 0, + "content": "\\[\nn _ {n} = - \\sigma_ {p} \\hat {n}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.495, + 0.905, + 0.51 + ], + "angle": 0, + "content": "\\[\ny _ {n} = \\hat {x} + n _ {n} = \\hat {x} - \\sigma_ {p} \\hat {n}. \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.535, + 0.864, + 0.551 + ], + "angle": 0, + "content": "3.3. Denoised Consistency Supervision (DCS)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.558, + 0.907, + 0.618 + ], + "angle": 0, + "content": "The basic idea of DCS is that the noisy observations \\( y_{p} \\) and \\( y_{n} \\) should correspond to the same clean image. Therefore, the loss supervises the consistency of the predicted denoised images:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.632, + 0.905, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p 2 n} = \\left\\| \\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) - \\mathcal {F} _ {\\theta} \\left(y _ {n}\\right) \\right\\|, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.907, + 0.72 + ], + "angle": 0, + "content": "where \\( \\|\\cdot\\| \\) represents the norm. Considering the general nature of noise [22], we employ a gradually varying norm function \\( \\|\\cdot\\|_{2 \\to 1.5} \\) to cover a broader range of unknown distributions:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.735, + 0.905, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\| x \\| _ {2 \\rightarrow 1. 5} = (| x | + \\epsilon) ^ {\\gamma}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.825 + ], + "angle": 0, + "content": "where \\(\\epsilon = 10^{-8}\\) and \\(\\gamma\\) varies linearly from 2 to 1.5 during training. The advantage of the varying \\(\\| \\cdot \\|_{2\\to 1.5}\\) norm is that \\(\\| \\cdot \\|_{2\\to 1.5}\\) does not require fine-tuning of parameters, yet \\(\\| \\cdot \\|_{2\\to 1.5}\\) consistently delivers robust performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The noisy observations \\( y_{p} \\) and \\( y_{n} \\) contain multi-scale noise. Through DCS, the neural network \\( \\mathcal{F}_{\\theta} \\) gradually learns to become insensitive to input noise and consistently outputs the clean signal, achieving the desired denoising performance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17927" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on SIDD, CC, PolyU and FMDD datasets. The best and second-best results (PSNR↑ / SSIM↑) are marked in red and blue in self-supervised single image denoising methods. The results of noisy dataset based self-supervised methods and supervised methods are also provided, only as a reference comparison. SIDD val and SIDD ben represent the SIDD validation dataset and the SIDD online benchmark, respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.156, + 0.907, + 0.485 + ], + "angle": 0, + "content": "
CategoryMethodSIDD val [1]SIDD ben [1]CC [29]PolyU [49]FMDD [53]
SupervisedUnprocessing [4]28.12 / 0.52931.64 / 0.53135.18 / 0.90836.99 / 0.94128.45 / 0.574
MaskedDenoising [6]28.67 / 0.60431.99 / 0.60133.87 / 0.93034.56 / 0.93629.86 / 0.643
CLIPDenoising [10]34.79 / 0.86635.82 / 0.85936.30 / 0.94137.54 / 0.96030.56 / 0.698
Self-Supervised (Noisy Dataset Based Methods)Noise2Void [18]29.35 / 0.65127.68 / 0.66832.27 / 0.86233.83 / 0.873-
CVF-SID [30]34.81 / 0.87235.05 / 0.85629.10 /0.85233.05 / 0.91131.64 / 0.837
LUD-VAE [56]34.91 / 0.89235.49 / 0.88331.25 / 0.92033.68 / 0.94232.00 / 0.822
Self-Supervised (Single Image Based Methods)DIP [38]32.11 / 0.740-35.61 / 0.91237.17 / 0.91232.90 / 0.854
PD-denoising [57]33.97 / 0.820-35.85 / 0.92337.04 / 0.94033.01 / 0.856
NN+denoiser [55]--36.52 / 0.94337.66 / 0.95632.21 / 0.831
Self2Self [36]29.46 / 0.59529.51 / 0.65137.44 / 0.94837.52 / 0.95130.76 / 0.695
R2R [33]24.59 / 0.387-33.43 / 0.86536.23 / 0.93127.17 / 0.526
APBSN-single [21]30.90 / 0.818-27.72 / 0.89129.61 / 0.89728.43 / 0.804
ScoreDVI [9]34.75 / 0.85635.39 / 0.85937.09 / 0.94537.77 / 0.95933.10 / 0.865
ZS-N2N [27]25.59 / 0.43330.19 / 0.42833.51 / 0.85135.99 / 0.91431.65 / 0.767
MASH [11]35.06 / 0.85134.80 / 0.81431.17 / 0.88937.62 / 0.93233.71 / 0.882
ATBSN-single [8]32.39 / 0.79627.86 / 0.31231.87 / 0.89733.21 / 0.92531.23 / 0.829
Positive2Negative35.24 / 0.88935.80 / 0.87537.92 / 0.95438.20 / 0.96134.00 / 0.887
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.51, + 0.194, + 0.526 + ], + "angle": 0, + "content": "3.4. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.483, + 0.576 + ], + "angle": 0, + "content": "In this section, we provide an analysis and proof demonstrating how the Positive2Negative paradigm achieves denoising." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.607 + ], + "angle": 0, + "content": "Taking the 2-norm as an example, the loss function can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.616, + 0.482, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p 2 n} = \\mathbb {E} \\left[ \\| \\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) - \\mathcal {F} _ {\\theta} \\left(y _ {n}\\right) \\| ^ {2} \\right]. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.676 + ], + "angle": 0, + "content": "Through Taylor expansion, we can approximate the predicted denoised images \\(\\mathcal{F}_{\\theta}(y_p)\\) and \\(\\mathcal{F}_{\\theta}(y_n)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.684, + 0.483, + 0.717 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) = \\mathcal {F} _ {\\theta} \\left(\\mathcal {F} _ {\\theta} (y)\\right) + \\sigma_ {p} \\cdot \\frac {\\partial \\mathcal {F} _ {\\theta} (y)}{\\partial y} \\cdot n + \\mathcal {O} \\left(n ^ {2}\\right), \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.718, + 0.483, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\theta} (y _ {n}) = \\mathcal {F} _ {\\theta} \\left(\\mathcal {F} _ {\\theta} (y)\\right) - \\sigma_ {n} \\cdot \\frac {\\partial \\mathcal {F} _ {\\theta} (y)}{\\partial y} \\cdot n + \\mathcal {O} \\left(n ^ {2}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.759, + 0.381, + 0.775 + ], + "angle": 0, + "content": "Substituting the approximations, we get:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.782, + 0.483, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) - \\mathcal {F} _ {\\theta} \\left(y _ {n}\\right) \\approx \\left(\\sigma_ {p} + \\sigma_ {n}\\right) \\cdot \\frac {\\partial \\mathcal {F} _ {\\theta} (y)}{\\partial y} \\cdot n. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.823, + 0.483, + 0.854 + ], + "angle": 0, + "content": "With further Taylor expansion, the loss function can be simplified as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.862, + 0.483, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p 2 n} \\approx (4 + 2 \\sigma^ {2}) \\left\\| \\frac {\\partial \\mathcal {F} _ {\\theta} (x)}{\\partial x} \\cdot n \\right\\| ^ {2}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.511, + 0.907, + 0.559 + ], + "angle": 0, + "content": "Through consistency loss \\(L_{p2n}\\), we optimize the network \\(\\mathcal{F}_{\\theta}\\) such that \\(\\frac{\\partial\\mathcal{F}_{\\theta}(x)}{\\partial x}\\cdot n\\) is optimized to drive towards zero, which means that the output \\(\\mathcal{F}_{\\theta}(y)\\) tends to:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.511, + 0.907, + 0.633 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {F} _ {\\theta} (y) = \\mathcal {F} _ {\\theta} (x + n), (17) \\\\ = \\mathcal {F} _ {\\theta} (x) + \\frac {\\partial \\mathcal {F} _ {\\theta} (x)}{\\partial x} \\cdot n + \\mathcal {O} \\left(n ^ {2}\\right), (18) \\\\ \\approx \\mathcal {F} _ {\\theta} (x). (19) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.906, + 0.685 + ], + "angle": 0, + "content": "Ideally, the neural network \\(\\mathcal{F}_{\\theta}\\) has learned image priors, the model output \\(\\mathcal{F}_{\\theta}(x)\\) matches the clean signal \\(x\\) and we have \\(\\mathcal{F}_{\\theta}(x) = x\\). Thus, the ultimate goal achieves:" + }, + { + "type": "equation", + "bbox": [ + 0.669, + 0.692, + 0.905, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\theta} (y) \\approx x. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.714, + 0.907, + 0.745 + ], + "angle": 0, + "content": "In summary, the neural network \\(\\mathcal{F}_{\\theta}\\) gradually learns to denoise the noisy image \\(y\\) effectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.646, + 0.773 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.843 + ], + "angle": 0, + "content": "In this section, we first introduce the experimental settings in Section 4.1. After that, we demonstrate the performance and comparisons with other methods in Section 4.2. Finally, we conduct a detailed ablation study in Section 4.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.613, + 0.865 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this section, we briefly introduce the experimental settings and details." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17928" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.895, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.43, + 0.907, + 0.5 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons of P2N against other self-supervised single image denoising methods in SIDD validation dataset. The denoised results of ScoreDVI are always flaky, with discontinuous lines and loss of texture. ZS-N2N does not completely denoise the images. MASH and ATBSN-single tend to either meet aliasing effect or blur the details. In summary, other methods meet aliasing effects and loss of texture, while Positive2Negative achieves both detailed texture preservation and comprehensive noise removal. In the figure, P2N represents Positive2Negative." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.484, + 0.62 + ], + "angle": 0, + "content": "Datasets details. Following previous methods [9, 11], we evaluate the proposed paradigm on the widely-used datasets in self-supervised single image denoising: the SIDD validation dataset [1], the SIDD benchmark dataset [1], CC [29], PolyU [49] and FMDD [53]. The SIDD, PolyU and CC datasets consist of natural sRGB images, while the FMDD dataset consists of fluorescence microscopy images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.484, + 0.777 + ], + "angle": 0, + "content": "Implementation details. Following the default settings of self-supervised single image denoising methods [9, 11, 22, 40], we employ the same neural network architecture as Noise2Noise [22] and MASH [11] and initialize the neural network parameters with those pre-trained on DIV2K dataset with Gaussian noise [9, 55]. For training details, we employ AdamW [26] optimizer with a learning rate of 0.0001 and set scale parameter \\(\\sigma = 0.75\\), for all the experiments. No image enhancement techniques are applied." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Compared methods. Following previous methods [9, 11, 24], we mainly compare with self-supervised single image denoising methods. In addition, we also present the results of noisy dataset based self-supervised and supervised image denoising methods only as a reference comparison. All the results of the comparison methods are sourced from previous work [9, 11, 24] or evaluated by ourselves." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.513, + 0.655, + 0.53 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.581 + ], + "angle": 0, + "content": "In this section, we demonstrate the performance of Positive2Negative and compare Positive2Negative with other methods both quantitatively and qualitatively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.701 + ], + "angle": 0, + "content": "Quantitative comparisons. The quantitative comparisons are described in Table 1. Previous self-supervised single image denoising methods have shown varying performance across different datasets, but Positive2Negative consistently achieves better results. In fact, Positive2Negative even outperforms some dataset-based and supervised methods. Notably, in the online evaluation of the SIDD benchmark dataset, we achieve an improvement of over \\(0.4\\mathrm{dB}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Qualitative comparisons. The qualitative comparisons are described in Figure 6. Clearly, the Positive2Negative paradigm achieves significant qualitative improvements by either enhancing details or effectively reducing noise. In contrast, other methods face limitations in both detail enhancement and noise removal due to the difficulty in learning denoising, resulting from the information-lossy barrier." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.625, + 0.833 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "In this section, we first present the training process in Section 4.3.1. Then, we conduct ablation studies on the components and hyperparameters in Section 4.3.2 and Section 4.3.3, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17929" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.088, + 0.903, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.908, + 0.343 + ], + "angle": 0, + "content": "Figure 7. Visual results of the training process. The top row shows the denoised results, while the bottom row displays the reconstructed negative noisy images in the training process. On the far right of each set is the reference image. The denoised image and reconstructed noisy image gradually become more like the reference images." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.355, + 0.486, + 0.398 + ], + "angle": 0, + "content": "Table 2. Ablation of components. Each step is crucial for Positive2Negative. N2N and N2V represent Noise2Noise and Noise2Void, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.409, + 0.482, + 0.55 + ], + "angle": 0, + "content": "
MethodCC [29]SIDD [1]
PSNRSSIMPSNRSSIM
Baseline35.400.92733.500.849
RDC+N2N34.080.86825.420.414
RDC+N2V33.790.86124.510.368
N2N+DCS17.550.69824.200.815
N2V+DCS14.140.49422.300.743
RDC+DCS37.920.95435.240.889
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.573, + 0.277, + 0.589 + ], + "angle": 0, + "content": "4.3.1. Ablation of training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.484, + 0.669 + ], + "angle": 0, + "content": "We present visual results of the training process to illustrate the effectiveness of our Positive2Negative paradigm. As shown in Figure 7, throughout the training process, the denoised images and the reconstructed negative noisy images gradually converge towards their respective references." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.675, + 0.304, + 0.691 + ], + "angle": 0, + "content": "4.3.2. Ablation of components" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.484, + 0.785 + ], + "angle": 0, + "content": "We evaluate the effectiveness of each component of the Positive2Negative paradigm on the CC and SIDD datasets. RDC constructs precise training data and DCS trains robust denoising neural networks. Both RDC and DCS are crucial for Positive2Negative. Replacing either component would result in a decline in performance, as shown in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.34, + 0.807 + ], + "angle": 0, + "content": "4.3.3. Ablation of hyperparameters" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Throughout the entire training process, the only adjustable hyperparameter is the scale parameter \\(\\sigma\\), which is set to 0.75 for all experiments. As shown in Table 3, the performance remains similar for \\(\\sigma\\) values ranging from 0.25 to 0.75. The scale parameter \\(\\sigma\\) introduces a dynamic and wide range of noise levels, while the \"Fixed\" setting corresponds to \\(\\sigma = 0\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.355, + 0.908, + 0.398 + ], + "angle": 0, + "content": "Table 3. Ablation of hyperparameters. The scale parameter is the only hyperparameter and is very robust. We fix the scale parameter \\(\\sigma = 0.75\\) for all the experiments." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.409, + 0.905, + 0.523 + ], + "angle": 0, + "content": "
MethodCC [29]SIDD [1]
PSNRSSIMPSNRSSIM
Fixed37.330.95235.240.892
σ = 0.2537.640.94535.290.893
σ = 0.5037.840.95235.300.891
σ = 0.7537.920.95435.240.889
σ = 1.0037.860.95435.110.884
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.547, + 0.635, + 0.563 + ], + "angle": 0, + "content": "5. Discussions" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.573, + 0.644, + 0.587 + ], + "angle": 0, + "content": "5.1. Distribution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.747 + ], + "angle": 0, + "content": "Different norms are suitable for different noise distributions [22]. For simplicity, we have uniformly adopted the \\(\\| \\cdot \\|_{2\\to 1.5}\\) norm, as real-world noise is typically centered around a zero mean with slight fluctuations. For noise that remains nearly zero-mean, such as in PolyU [49], the \\(\\| \\cdot \\|_{2\\to 1.5}\\) norm and the \\(\\| \\cdot \\|_2\\) norm yield similar performance. For noise that exhibits slight deviations from zero mean, as in SIDD [1], the \\(\\| \\cdot \\|_{2\\to 1.5}\\) norm can outperform the \\(\\| \\cdot \\|_2\\) norm, as shown in Table 4. The slight increase in PSNR further suggests that the deviation remains minimal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.65, + 0.774 + ], + "angle": 0, + "content": "5.2. Convergence" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present the PSNR variation over iterations on SIDD dataset [1] in Figure 8. Around the 100th iteration, the Positive2Negative paradigm converges and shows minimal fluctuations thereafter. Additionally, it is worth noting that the constructed data by RDC is unsuitable for Noise2Noise and Noise2Void. This is mainly because the noise in constructed images by RDC is correlated, which would quickly collapse the training processes of the other two methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17930" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 4. Discussion of distributions. Different norms are suitable for different noise distributions [22]. For simplicity, we have uniformly adopted the \\( \\| \\cdot \\|_{2\\to 1.5} \\) norm." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.143, + 0.485, + 0.202 + ], + "angle": 0, + "content": "
Dataset||·||0||·||1||·||2||·||2→1.5
PolyU [49]37.8838.0238.2038.20
SIDD [1]35.1235.2635.2035.24
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.216, + 0.287, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.216, + 0.482, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.312, + 0.485, + 0.37 + ], + "angle": 0, + "content": "Figure 8. Convergence comparisons. P2N converges quickly and remains stable after convergence on SIDD. The P2N-Fixed version (where \\(\\sigma = 0.00\\)), in particular, even achieves rapid convergence in approximately 10 iterations. P2N represents Positive2Negative." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.396, + 0.206, + 0.412 + ], + "angle": 0, + "content": "5.3. Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.511 + ], + "angle": 0, + "content": "For self-supervised single image denoising methods, computational efficiency is crucial, as each noisy image must be trained individually. Here, we demonstrate the number of training iterations, parameter, and computational cost, as detailed in Table 5. It is obvious that our paradigm achieves significant efficiency improvements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.523, + 0.482, + 0.54 + ], + "angle": 0, + "content": "5.4. Comparison with Noise2Noise and Noise2Void" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.592 + ], + "angle": 0, + "content": "Compared with Noise2Noise [22] and Noise2Void [18], Positive2Negative also presents an elegant paradigm The key differences are present in the following aspects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.775 + ], + "angle": 0, + "content": "Firstly, for the constructed data, the synthetic data constructed by Positive2Negative breaks the information-lossy barrier of Noise2Noise and Noise2Void. Secondly, for the supervision strategy, Positive2Negative enforces consistency between denoised results, whereas Noise2Noise and Noise2Void supervise the consistency between a denoised result and a noisy image. Thirdly, from the assumption perspective, Noise2Noise, Noise2Void, and Positive2Negative all have theoretical guarantees, while Noise2Noise and Noise2Void operate under more restrictive assumptions. Noise2Noise requires two independent noisy images, while Noise2Void assumes that the noise is spatially correlated." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.405, + 0.803 + ], + "angle": 0, + "content": "5.5. Differences with R2R and CVF-SID" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.903 + ], + "angle": 0, + "content": "For the constructed data, R2R employs new Gaussian noise and applies reciprocal coefficients to the noisy image, while CVF-SID employs two networks to predict three components and applies fixed manual rules. In contrast, our method employs a single network to predict noise and applies random multiscale symmetrical parameters." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.161 + ], + "angle": 0, + "content": "Table 5. Efficiency comparisons under input images with dimensions of \\(256 \\times 256 \\times 3\\). With similar parameters and FLOPs, P2N requires the fewest training iterations. P2N converges rapidly, and the P2N-Fixed version (where \\(\\sigma = 0.00\\)) converges even just within 20 iterations. P2N represents Positive2Negative." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.17, + 0.905, + 0.326 + ], + "angle": 0, + "content": "
MethodIterationsParams (M)FLOPs (G)
N2N [22]-0.7018.7
B2U [40]-1.1018.7
Self2Self [36]4500001.009.6
R2R [33]80000.5636.7
ScoreDVI [9]40013.537.9
MASH [11]800 × 30.9911.4
P2N-Fixed<200.9911.4
P2N<1000.9911.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.432 + ], + "angle": 0, + "content": "For the supervision strategy, R2R maps the denoised image to the noisy input, while CVF-SID primarily maps the first cyclic result to the noisy input and maps the second cyclic result to the first cyclic result. In contrast, our method maps the denoised image to another denoised image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.447, + 0.771, + 0.463 + ], + "angle": 0, + "content": "5.6. Limitation and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.909, + 0.654 + ], + "angle": 0, + "content": "The primary limitation is that Positive2Negative has only been validated on the single image based self-supervised image denoising task. To prevent convergence to trivial solutions such as zero mapping or identity mapping, Positive2Negative requires training with a pre-trained Gaussian denoising model. While the single image based self-supervised denoising task allows training from a pre-trained model on Gaussian noise, the noisy dataset based self-supervised image denoising task trains from scratch. Therefore, extending the Positive2Negative paradigm to the noisy dataset based self-supervised image denoising task could be a promising direction for future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.674, + 0.634, + 0.69 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.907, + 0.793 + ], + "angle": 0, + "content": "In this paper, we propose a self-supervised single image denoising paradigm, Positive2Negative, to break information-lossy barrier. Specifically, Positive2Negative includes two steps: RDC and DCS. Positive2Negative achieves SOTA performance compared to existing methods across various public benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.813, + 0.673, + 0.83 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China under Grants 62322204, 62131003, and the Key Laboratory of Target Cognition and Application Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17931" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S. Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1700, 2018. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.185, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Joshua Batson and Loic Royer. Noise2Self: Blind denoising by self-supervision. In International Conference on Machine Learning (ICML), pages 524-533, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[3] Chinmay Belthangady and Loic A Royer. Applications, promises, and pitfalls of deep learning for fluorescence image reconstruction. In Nature Methods, pages 1215-1225, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.284, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[4] Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing images for learned raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11036-11045, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.355, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[5] Jaeseok Byun, Sungmin Cha, and Taesup Moon. Fbidenoiser: Fast blind image denoiser for poisson-gaussian noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5768-5777, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.424, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1703, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision (ECCV), pages 17-33, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[8] Shiyan Chen, Jiyuan Zhang, Zhaofei Yu, and Tiejun Huang. Exploring efficient asymmetric blind-spots for self-supervised denoising in real-world scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2814-2823, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[9] Jun Cheng, Tao Liu, and Shan Tan. Score priors guided deep variational inference for unsupervised real-world single image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12937-12948, 2023. 1, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[10] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25974-25984, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[11] Hamadi Chihaoui and Paolo Favaro. Masked and shuffled blind spot denoising for real-world images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3025-3034, 2024. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Hansen Feng, Lizhi Wang, Yuzhi Wang, Haoqiang Fan, and Hua Huang. Learnability enhancement for low-light raw image denoising: A data perspective. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 370-387, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. In ACM Transactions on Graphics (ToG), pages 1-12, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[14] Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. Neighbor2Neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14781-14790, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[15] Yeong Il Jang, Keuntek Lee, Gu Yong Park, Seyun Kim, and Nam Ik Cho. Self-supervised image denoising with downsampled invariance loss and conditional blind-spot network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12196-12205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[16] Kwanyoung Kim and Jong Chul Ye. Noise2score: tweedie's approach to self-supervised image denoising without clean images. In Advances in Neural Information Processing Systems, pages 864-874, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[17] Kwanyoung Kim, Taesung Kwon, and Jong Chul Ye. Noise distribution adaptive self-supervised image denoising using tweedie distribution and score matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1998-2006, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[18] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void: Learning denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2129-2137, 2019. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[19] Alexander Krull, Tomáš Vičar, Mangal Prakash, Manan Lalit, and Florian Jug. Probabilistic noise2void: Unsupervised content-aware denoising. In Frontiers in Computer Science, pages 2-5, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[20] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. In Proceedings of International Conference on Neural Information Processing Systems (NeurIPS), pages 6970-6980, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.907, + 0.688 + ], + "angle": 0, + "content": "[21] Wooseok Lee, Sanghyun Son, and Kyoung Mu Lee. APBSN: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17725-17734, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[22] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2Noise: Learning image restoration without clean data. In International Conference on Machine Learning (ICML), pages 4620-4631, 2018. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[23] Junyi Li, Zhilu Zhang, Xiaoyu Liu, Chaoyu Feng, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Spatially adaptive self-supervised learning for real-world image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914-9924, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] Tong Li, Hansen Feng, Lizhi Wang, Zhiwei Xiong, and Hua Huang. Stimulating the diffusion model for image denoising via adaptive embedding and ensembling. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 8240-8257, 2024. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17932" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[25] Xinyang Li, Yixin Li, Yiliang Zhou, Jiamin Wu, Zhifeng Zhao, Jiaqi Fan, Fei Deng, Zhaofa Wu, Guihua Xiao, Jing He, Yuanlong Zhang, Guoxun Zhang, Xiaowan Hu, Zhang Yi Chen, Xingye and, Hui Qiao, Hao Xie, Yulong Li, Haoqian Wang, Lu Fang, and Qionghai Dai. Real-time denoising enables high-sensitivity fluorescence time-lapse imaging beyond the shot-noise limit. In Nature Biotechnology, page 282-292, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. In arXiv preprint arXiv:1711.05101, pages 1-10, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[27] Youssef Mansour and Reinhard Heckel. Zero-Shot Noise2Noise: Efficient image denoising without any data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14018-14027, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[28] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2Noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12064-12072, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[29] Seonghyeon Nam, Youngbae Hwang, Yasuyuki Matsushita, and Seon Joo Kim. A holistic approach to cross-channel image noise modeling and its application to image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1683-1691, 2016. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[30] Reyhaneh Neshatavar, Mohsen Yavartanoo, Sanghyun Son, and Kyoung Mu Lee. CVF-SID: Cyclic multi-variate function for self-supervised image denoising by disentangling noise from image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17583-17591, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[31] Harry Nyquist. Certain topics in telegraph transmission theory. In Transactions of the American Institute of Electrical Engineers, pages 617-644, 1928. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[32] Yizhong Pan, Xiao Liu, Xiangyu Liao, Yuanzhouhan Cao, and Chao Ren. Random sub-samples generation for self-supervised real image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12150-12159, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[33] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-Recorrupted: Unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2043–2052, 2021. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[34] Mikhail Papkov, Pavel Chizhov, and Leopold Parts. Swinia: Self-supervised blind-spot image denoising without convolutions. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 7071-7080, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[35] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1586-1595, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.903 + ], + "angle": 0, + "content": "[36] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2Self with dropout: Learning self-supervised denoising" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "from single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1890-1898, 2020. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[37] Claude Elwood Shannon. A mathematical theory of communication. In The Bell System Technical Journal, pages 379-423, 1948. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[38] Andrea Vedaldi Ulyanov Dmitry and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[39] Yuzhi Wang, Haibin Huang, Qin Xu, Jiaming Liu, Yiqun Liu, and Jue Wang. Practical deep raw image denoising on mobile devices. In European Conference on Computer Vision (ECCV), pages 1-16, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[40] Zejin Wang, Jiazheng Liu, Guoqing Li, and Hua Han. Blind2Unblind: Self-supervised image denoising with visible blind spots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2027-2036, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[41] Zichun Wang, Ying Fu, Ji Liu, and Yulun Zhang. LG-BPN: Local and global blind-patch network for self-supervised real-world denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18156-18165, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[42] Kaixuan Wei, Ying Fu, Yinqiang Zheng, and Jiaolong Yang. Physics-based noise modeling for extreme low-light photography. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 8520-8537, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[43] Martin Weigert, Uwe Schmidt, Tobias Boothe, Andreas Müller, Alexandr Dibrov, Akanksha Jain, Benjamin Wilhelm, Deborah Schmidt, Coleman Broaddus, Sián Culley, et al. Content-aware image restoration: pushing the limits of fluorescence microscopy. In Nature Methods, pages 1090–1097, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[44] Halbert White. Connectionist nonparametric regression: Multilayer feedforward networks can learn arbitrary mappings. In Neural networks, pages 535-549, 1990. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[45] David Wu, Damian Tan, Marilyn Baird, John DeCampo, Chris White, and Hong Ren Wu. Perceptually lossless medical image coding. In IEEE Transactions on Medical Imaging, pages 335-344, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[46] Hong Ren Wu, Weisi Lin, and Lina J. Karam. An overview of perceptual processing for digital pictures. In IEEE International Conference on Multimedia and Expo Workshops, pages 113-120, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[47] Xiaohe Wu, Ming Liu, Yue Cao, Dongwei Ren, and Wangmeng Zuo. Unpaired learning of deep image denoising. In European Conference on Computer Vision (ECCV), pages 352-368, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[48] Yaochen Xie, Zhengyang Wang, and Shuiwang Ji. Noise2same: Optimizing a self-supervised bound for image denoising. In Advances in Neural Information Processing Systems, pages 20320-20330, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. In arXiv preprint arXiv:1804.02603, pages 1-13, 2018. 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17933" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[50] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-Clean: Learning self-supervised denoising from corrupted image. In IEEE Transactions on Image Processing (TIP), pages 9316-9329, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728-5739, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[52] Dan Zhang, Fangfang Zhou, Yuwen Jiang, and Zhengming Fu. Mm-bsn: Self-supervised image denoising for real-world with multi-mask based on blind-spot network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 4189–4198, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[53] Yide Zhang, Yinhao Zhu, Evan Nichols, Qingfei Wang, Siyuan Zhang, Cody Smith, and Scott Howard. Aoisson-gaussian denoising dataset with real fluorescence microscopy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11710-11718, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[54] Yi Zhang, Dasong Li, Ka Lung Law, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. IDR: Self-supervised image denoising via iterative data refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2098-2107, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.473, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[55] Dihan Zheng, Sia Huat Tan, Xiaowen Zhang, Zuoqiang Shi, Kaisheng Ma, and Chenglong Bao. An unsupervised deep learning approach for real-world image denoising. In International Conference on Learning Representations (ICLR), pages 1-9, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.544, + 0.482, + 0.611 + ], + "angle": 0, + "content": "[56] Dihan Zheng, Xiaowen Zhang, Kaisheng Ma, and Chenglong Bao. Learn from unpaired data for image restoration: A variational bayes approach. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 5889-5903, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.614, + 0.482, + 0.682 + ], + "angle": 0, + "content": "[57] Yuqian Zhou, Jianbo Jiao, Haibin Huang, Yang Wang, Jue Wang, Honghui Shi, and Thomas Huang. When awgn-based denoiser meets real noises. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 13074-13081, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.684, + 0.482, + 0.753 + ], + "angle": 0, + "content": "[58] Yunhao Zou, Chenggang Yan, and Ying Fu. Iterative denoiser and noise estimator for self-supervised image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13265-13274, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17934" + } + ] +] \ No newline at end of file diff --git a/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/867e9811-fe1d-4217-8c41-8e621b6da43b_origin.pdf b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/867e9811-fe1d-4217-8c41-8e621b6da43b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d456b898664b4527bc1de86feeeea222108f1c62 --- /dev/null +++ b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/867e9811-fe1d-4217-8c41-8e621b6da43b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c004d8ebdea079d18d3a5a2ca1470eb2e1c650b896b66612f7c9716ec2d41e3 +size 7499844 diff --git a/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/full.md b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9d0480e11c391ff980fd4d21ae3b02fa83799bf2 --- /dev/null +++ b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/full.md @@ -0,0 +1,406 @@ +# Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising + +Tong Li $^{1}$ Lizhi Wang $^{2,3*}$ Zhiyuan Xu $^{1}$ Lin Zhu $^{1}$ Wanxuan Lu $^{4}$ Hua Huang $^{2,3}$ + +$^{1}$ School of Computer Science and Technology, Beijing Institute of Technology $^{2}$ School of Artificial Intelligence, Beijing Normal University + +3 Engineering Research Center of Intelligent Technology and Educational Application, Ministry of Education + +$^{4}$ Aerospace Information Research Institute, Chinese Academy of Sciences + +# Abstract + +Image denoising enhances image quality, serving as a foundational technique across various computational photography applications. The obstacle to clean image acquisition in real scenarios necessitates the development of self-supervised image denoising methods only depending on noisy images, especially a single noisy image. Existing self-supervised image denoising paradigms (Noise2Noise and Noise2Void) rely heavily on information-lossy operations, such as downsampling and masking, culminating in low-quality denoising performance. In this paper, we propose a novel self-supervised single image denoising paradigm, Positive2Negative, to break the information-lossy barrier. Our paradigm involves two key steps: Renoised Data Construction (RDC) and Denoised Consistency Supervision (DCS). RDC renoes the predicted denoised image by the predicted noise to construct multiple noisy images, preserving all the information of the original image. DCS ensures consistency across the multiple denoised images, supervising the network to learn robust denoising. Our Positive2Negative paradigm achieves state-of-the-art performance in self-supervised single image denoising with significant speed improvements. The code is released to the public at https://github.com/Li-Tong-621/P2N. + +# 1. Introduction + +Image denoising plays a vital role in image processing to enhance visual quality [13], serving as a foundational technique across various computational photography applications [3, 39]. The widespread popularity of paired clean-noisy image datasets [1, 35] has catalyzed advancements in supervised methods, yielding remarkable performance [7, 51]. However, supervised methods fall short in + +![](images/6030aee1e7cbeaca920ad71a0402be5f35f7867661fa67f2fd60255d45328d20.jpg) +Noise2Noise-Based Paradigms + +![](images/a003c6634aeed0e0b9ed91b36d2b0cf0004cb1e5683948feefe0ed9452d53f89.jpg) + +![](images/50164d06ded5d99c14301ce13480195e748ee747b9db641fe115f1da6cecc3b6.jpg) +Noise2Void-Based Paradigms +Positive2Negative Paradigm +Figure 1. Information-lossy operations in current paradigms. The noise addition operation reduces the signal-to-noise ratio, the downsampling operation diminishes the sampling density, and the masking operation disregards crucial central pixels. Each of these operations introduces an information-lossy barrier [8, 31, 37], culminating in imprecise denoising. + +dynamic scenes [18, 22], where long exposure results in misalignment and blur [22, 42], hindering the acquisition of clean images. Therefore, advancing self-supervised image denoising methods using only noisy images [25, 43], especially a single noisy image [9, 11, 24, 27], holds significant importance. + +The crux of self-supervised image denoising methods + +![](images/e4df928a1a33cb13fde08a59c6a9b34639ebf6b25cc8dd8aacfb75a2aeb52132.jpg) +Figure 2. The information-lossy barrier significantly compromises the image quality, resulting in residual noise, aliasing effects or texture loss. + +lies in the construction of the training data and the formulation of the supervision strategy. Current methods are divided into Noise2Noise-based paradigms and Noise2Void-based paradigms, as illustrated in Figure 1. Noise2Noise-based paradigms add additional noise to the noisy image [28, 33] or down sample the noisy image [14, 27] to approximately construct two independently noisy observations as training data. These paradigms learn denoising by supervising one observation with the denoised result of other observation. Noise2Void-based paradigms mask either the noisy image [18, 40] or the neural network [8, 21, 41] to construct masked training data. These paradigms learn denoising by supervising the denoised result of the masked central pixel with the original noisy pixel, where the prediction only relies on the surrounding pixels. + +However, existing methods rely heavily on information-lossy operations, resulting in low-quality denoising performance [21, 32, 40, 41]. Specifically, the noise addition operation reduces the signal-to-noise ratio, the downsampling operation diminishes the sampling density and the masking operation disregards crucial central pixels. Each of these operations loses the information from the original noisy image [8, 31, 37], inevitably compromising the denoising ability [41, 44]. Consequently, denoised images always suffer from remaining noise, aliasing effects or texture loss [45, 46], as shown in Figure 2. As these information-lossy operations constitute the indispensable foundation of the Noise2Noise-based and Noise2Void-based paradigms, it is imperative to establish a new paradigm that transcends the Noise2Noise-based and Noise2Void-based paradigms. + +In this paper, we propose the concise Positive2Negative paradigm to break the information-lossy barrier in self-supervised single image denoising. Our inspiration stems from the observation that noise distributions are approximately symmetrical and centered around zero, implying that the opposite noise also follows the distribution of the original noise. Leveraging this insight, we can construct multiple noisy images corresponding to the same clean image, which + +serves as the basis of Positive2Negative. Concretely, Positive2Negative relies on two steps: Renoised Data Construction (RDC) and Denoised Consistency Supervision (DCS). + +The basic idea of RDC is to construct training data. Specifically, we design a zero-mean sampling strategy to generate multi-scale Positive noise and Negative noise based on the predicted noise, which are then added back to construct Positive noisy images and Negative noisy images. The constructed data comprehensively covers the information of the given single image; thus the Positive2Negative paradigm breaks the information-lossy barrier. + +The basic idea of DCS is to formulate a supervision strategy. Specifically, given the constructed multi-scale noisy images, DCS supervises the network to generate consistent denoised outputs during the training process. Furthermore, we provide a theoretical analysis showing that the Positive2Negative paradigm can learn robust denoising, through supervising the multiple denoised images. + +In summary, our contributions are as follows: + +- We propose a new self-supervised single image denoising paradigm, Positive2Negative, to break the information-lossy barrier. +- We propose a data construction method, which constructs multi-scale similar noisy images for training. +- We propose a denoising supervision method, which is theoretically guaranteed to learn robust denoising. +- Positive2Negative achieves SOTA performance over self-supervised single image denoising. + +# 2. Related Work + +From the training data perspective, the self-supervised image denoising methods can be classified into noisy dataset based methods and single image based methods. + +# 2.1. Noisy Dataset based Methods + +Noisy dataset based methods represent the most active research direction in image denoising. These methods learn to denoise from a given noisy dataset. Generally, a larger volume of the noisy dataset improves the denoising performance. As a result, in scenarios where only a single noisy image is available, the denoising performance of whole noisy dataset-based methods decreases significantly. + +Noise2Noise [22] learns denoising by mapping one noisy observation to another independent observation of the same scene. Subsequent studies focus on generating observations of the same scene from a single observation, typically achieved through downsampling [14, 27] or by introducing additional noise [28, 33, 50, 54, 58]. After that, another elegant method Noise2Void [18] is proposed. Noise2Void [18] assumes that the noise distribution is zero-mean, the noise is spatially independent, and the signal is spatially correlated. Noise2Void [18] predicts the clean signal by leveraging surrounding noisy signals, with mask + +ing the central pixel. Various other works propose alternative masking schemes, either mask the input [11, 36, 40] or mask the network [2, 19, 20, 34, 48, 52]. Noise2Void struggles with real-world noise, which often exhibits spatial correlation. Therefore, subsequent research endeavors to break the noise spatial correlation [8, 21, 23, 41, 47, 57]. There are also some methods that amalgamate frameworks established by both Noise2Noise and Noise2Void [15, 32] or get rid of frameworks established by Noise2Noise and Noise2Void [5, 16, 17, 30, 56], but these methods are usually less developed due to the complexity. + +# 2.2. Single Image based Methods + +Single image based methods represent the most challenging situation for denoising and the least demanding requirements for data. These methods learn to denoise from a given single noisy image without any dependence on paired noisy-clean or noisy datasets. As a result, single image based methods are less developed due to the inherent difficulty, only gaining attention in the past two years. + +DIP [38] is the earliest self-supervised image denoising method. DIP [38] learns to map random noise to a given noisy image and employs early stopping to prevent overfitting. Self2Self [36] uses dropout to create two Bernoulli-sampled observations of a noisy image. R2R [33] constructs noisy images by adding new Gaussian noise and applying reciprocal coefficients. ScoreDVI [9] performs real denoising using a pre-trained Gaussian denoiser and an estimated noise model. ZS-N2N [27] extend Noise2Noise [22] to a single image version by downsampling. MASH [11] employs masking and local pixel shuffling to break the noise spatial correlation. Recently, DMID stimulates the diffusion model for self-supervised single image denoising [24]. + +In this paper, we propose Positive2Negative, a new self-supervised single image denoising paradigm, to break the information-lossy barrier. + +# 3. Positive2Negative + +In this section, we first demonstrate the motivation of Positive2Negative in Section 3.1. After that, we present the two critical steps of Positive2Negative: RDC and DCS, respectively in Section 3.2 and Section 3.3. Finally, we provide an analysis and proof of Positive2Negative in Section 3.4. + +# 3.1. Motivation + +Our inspiration stems from the widely recognized assumptions that the noise distribution is zero-mean and approximately symmetrical [18, 22], which is consistent with our observation. In our observation experiment, we calculate the noise distribution of a fixed pixel across multiple observations of the same scene. Specifically, following SIDD [1], ELD [42] and PMN [12], we collect noisy image $y$ and clean image $c$ . Then we calculate the noise value $n$ at each + +![](images/886d228308623bd181382553d5106bdc587a173134a3eb334629454421e61d37.jpg) + +![](images/697cfef3d7501d8e3ace79726fdad6473499e5855cae7834a6c43d12583cc41d.jpg) +(a). noisy image +(b). green circle + +![](images/c35e222e7404ee6411235128c20f70c831b3088ee5aa5a0c6998326b95f7539b.jpg) +(c). red circle + +![](images/b516d66181a60f87abc87956a2bb38162b2c6694552d10e4b25316008d5f73ef.jpg) +Figure 3. Noise distribution is zero-mean and approximately symmetrical. (a) shows a noisy image. (b) and (c) show the noise distributions, which are calculated at the center pixels marked by the green circle and the red circle in the noisy image, respectively. It is evident that the noise distribution is zero-mean and approximately symmetrical. +Original noise +n + +![](images/699519e8f7265c12a2d3c7d1429fb3a80cade584cea652d197f3d4d286490148.jpg) +Original noisy +$y = c + n$ +Figure 4. The opposite noise $-n$ is similar to the original noise $n$ . The opposite noise $-n$ approximately follows the original noise's distribution, which is zero-mean and approximately symmetrical. + +![](images/7f331d157e6699eb6a5d689453d34f19844d1e76cbd046d7bd9f69f4851a252a.jpg) +Opposite noise +$n_{n} = -n$ + +![](images/16ae37fb525c992e55db198498e7f9ce967b66372b1017ed8b05a504e3caee60.jpg) +Opposite noisy +$y_{n} = c + (-n)$ + +spatial location $(i,j)$ . Additionally, we compute the frequency of $n_{i,j}$ cross noisy images of the same scene. The results are shown in Figure 3, which demonstrates the noise distribution is zero-mean and approximately symmetrical. + +Based on this observation, an intuitive insight is that the opposite noise $-n$ also follows the same distribution as that of the original noise $n$ . This intuitive idea inspires us to construct the opposite noise $n_{n}$ and corresponding opposite noisy image $y_{n}$ : + +$$ +n _ {n} = - n, \tag {1} +$$ + +$$ +y _ {n} = x + n _ {n} = x - n. \tag {2} +$$ + +The opposite noise $n_n$ and opposite noisy image $y_n$ are similar to the original noise $n$ and original noisy image $y$ , respectively, as shown in Figure 4. + +![](images/0b829f6c936d354a94499a951522460a36a8e9749d5a496bdbaedc8629236c4a.jpg) +Figure 5. Overview of the proposed self-supervised single image denoising paradigm. For training, the Positive2Negative paradigm consists of 2 steps. The first step is Renoised Data Construction (RDC): multiply the predicted noise by positive and negative coefficients to construct multi-scale Positive noise and Negative noise, respectively. The second step is Denoised Consistency Supervision (DCS): train the network through supervising the consistency of the denoised images. For inference, denoising can be achieved in just one pass through the network. (The networks shown in the figure are the same one. In addition, for better visualization, the images within the red boxes on the bottom left have been zoomed in and are displayed in the red boxes on the bottom right.) + +Based on the above observations and insights, we propose the Positive2Negative paradigm, as shown in Figure 5. Firstly, we extend our intuitive idea into a flexible and upgraded version: Renoised Data Construction (RDC), which constructs multiple positive and negative noises. Secondly, we propose the Denoised Consistency Supervision (DCS), which supervises the consistency of the predicted denoised images to learn a denoising neural network. + +It is important to note that we do not require the noise distribution to be strictly zero-mean or symmetric. Although real noise generally has a zero-mean, certain cases where the noise is not strictly symmetric can be easily addressed by adjusting the norm in the loss function [22]. + +# 3.2. Renoised Data Construction (RDC) + +The basic idea of RDC is to denoise the image to predict a denoised image and renoise the predicted denoised image by the predicted noise to construct multiple noisy images. + +For a given noisy image $y = x + n$ , where $x$ presents the clean image and $n$ presents the noise. RDC firstly denoises the noisy image $y$ once through the neural network $\mathcal{F}$ parametered by $\theta$ : + +$$ +\hat {x} = \mathcal {F} _ {\theta} (y). \tag {3} +$$ + +Here $\hat{x}$ is the predicted denoised image of the noisy observation $y$ . Then we can calculate the predicted noise $\hat{n}$ within the noisy observation $y$ : + +$$ +\hat {n} = y - \hat {x}. \tag {4} +$$ + +Next, RDC constructs noisy images as training data based on the predicted noise $\hat{n}$ and the predicted denoised image $\hat{x}$ . As noise is zero-mean, we further design a zero-mean sampling strategy to construct new noises. Specifically, RDC multiplies the predicted noise $\hat{n}$ by both positive scale parameters $\sigma_{p}$ and negative scale parameters $-\sigma_{n}$ + +sampled from Gaussian distribution $\mathcal{N}$ , constructing multiscale Positive noise $n_p$ and Negative noise $n_n$ , respectively: + +$$ +\sigma_ {n}, \sigma_ {p} \sim \mathcal {N} (1, \sigma), \tag {5} +$$ + +$$ +n _ {p} = \sigma_ {n} \hat {n}, \tag {6} +$$ + +$$ +y _ {p} = \hat {x} + n _ {p} = \hat {x} + \sigma_ {n} \hat {n}, \tag {7} +$$ + +$$ +n _ {n} = - \sigma_ {p} \hat {n}, \tag {8} +$$ + +$$ +y _ {n} = \hat {x} + n _ {n} = \hat {x} - \sigma_ {p} \hat {n}. \tag {9} +$$ + +# 3.3. Denoised Consistency Supervision (DCS) + +The basic idea of DCS is that the noisy observations $y_{p}$ and $y_{n}$ should correspond to the same clean image. Therefore, the loss supervises the consistency of the predicted denoised images: + +$$ +\mathcal {L} _ {p 2 n} = \left\| \mathcal {F} _ {\theta} \left(y _ {p}\right) - \mathcal {F} _ {\theta} \left(y _ {n}\right) \right\|, \tag {10} +$$ + +where $\|\cdot\|$ represents the norm. Considering the general nature of noise [22], we employ a gradually varying norm function $\|\cdot\|_{2 \to 1.5}$ to cover a broader range of unknown distributions: + +$$ +\| x \| _ {2 \rightarrow 1. 5} = (| x | + \epsilon) ^ {\gamma}, \tag {11} +$$ + +where $\epsilon = 10^{-8}$ and $\gamma$ varies linearly from 2 to 1.5 during training. The advantage of the varying $\| \cdot \|_{2\to 1.5}$ norm is that $\| \cdot \|_{2\to 1.5}$ does not require fine-tuning of parameters, yet $\| \cdot \|_{2\to 1.5}$ consistently delivers robust performance. + +The noisy observations $y_{p}$ and $y_{n}$ contain multi-scale noise. Through DCS, the neural network $\mathcal{F}_{\theta}$ gradually learns to become insensitive to input noise and consistently outputs the clean signal, achieving the desired denoising performance. + +Table 1. Quantitative comparisons on SIDD, CC, PolyU and FMDD datasets. The best and second-best results (PSNR↑ / SSIM↑) are marked in red and blue in self-supervised single image denoising methods. The results of noisy dataset based self-supervised methods and supervised methods are also provided, only as a reference comparison. SIDD val and SIDD ben represent the SIDD validation dataset and the SIDD online benchmark, respectively. + +
CategoryMethodSIDD val [1]SIDD ben [1]CC [29]PolyU [49]FMDD [53]
SupervisedUnprocessing [4]28.12 / 0.52931.64 / 0.53135.18 / 0.90836.99 / 0.94128.45 / 0.574
MaskedDenoising [6]28.67 / 0.60431.99 / 0.60133.87 / 0.93034.56 / 0.93629.86 / 0.643
CLIPDenoising [10]34.79 / 0.86635.82 / 0.85936.30 / 0.94137.54 / 0.96030.56 / 0.698
Self-Supervised (Noisy Dataset Based Methods)Noise2Void [18]29.35 / 0.65127.68 / 0.66832.27 / 0.86233.83 / 0.873-
CVF-SID [30]34.81 / 0.87235.05 / 0.85629.10 /0.85233.05 / 0.91131.64 / 0.837
LUD-VAE [56]34.91 / 0.89235.49 / 0.88331.25 / 0.92033.68 / 0.94232.00 / 0.822
Self-Supervised (Single Image Based Methods)DIP [38]32.11 / 0.740-35.61 / 0.91237.17 / 0.91232.90 / 0.854
PD-denoising [57]33.97 / 0.820-35.85 / 0.92337.04 / 0.94033.01 / 0.856
NN+denoiser [55]--36.52 / 0.94337.66 / 0.95632.21 / 0.831
Self2Self [36]29.46 / 0.59529.51 / 0.65137.44 / 0.94837.52 / 0.95130.76 / 0.695
R2R [33]24.59 / 0.387-33.43 / 0.86536.23 / 0.93127.17 / 0.526
APBSN-single [21]30.90 / 0.818-27.72 / 0.89129.61 / 0.89728.43 / 0.804
ScoreDVI [9]34.75 / 0.85635.39 / 0.85937.09 / 0.94537.77 / 0.95933.10 / 0.865
ZS-N2N [27]25.59 / 0.43330.19 / 0.42833.51 / 0.85135.99 / 0.91431.65 / 0.767
MASH [11]35.06 / 0.85134.80 / 0.81431.17 / 0.88937.62 / 0.93233.71 / 0.882
ATBSN-single [8]32.39 / 0.79627.86 / 0.31231.87 / 0.89733.21 / 0.92531.23 / 0.829
Positive2Negative35.24 / 0.88935.80 / 0.87537.92 / 0.95438.20 / 0.96134.00 / 0.887
+ +# 3.4. Analysis + +In this section, we provide an analysis and proof demonstrating how the Positive2Negative paradigm achieves denoising. + +Taking the 2-norm as an example, the loss function can be expressed as: + +$$ +\mathcal {L} _ {p 2 n} = \mathbb {E} \left[ \| \mathcal {F} _ {\theta} \left(y _ {p}\right) - \mathcal {F} _ {\theta} \left(y _ {n}\right) \| ^ {2} \right]. \tag {12} +$$ + +Through Taylor expansion, we can approximate the predicted denoised images $\mathcal{F}_{\theta}(y_p)$ and $\mathcal{F}_{\theta}(y_n)$ : + +$$ +\mathcal {F} _ {\theta} \left(y _ {p}\right) = \mathcal {F} _ {\theta} \left(\mathcal {F} _ {\theta} (y)\right) + \sigma_ {p} \cdot \frac {\partial \mathcal {F} _ {\theta} (y)}{\partial y} \cdot n + \mathcal {O} \left(n ^ {2}\right), \tag {13} +$$ + +$$ +\mathcal {F} _ {\theta} (y _ {n}) = \mathcal {F} _ {\theta} \left(\mathcal {F} _ {\theta} (y)\right) - \sigma_ {n} \cdot \frac {\partial \mathcal {F} _ {\theta} (y)}{\partial y} \cdot n + \mathcal {O} \left(n ^ {2}\right). \tag {14} +$$ + +Substituting the approximations, we get: + +$$ +\mathcal {F} _ {\theta} \left(y _ {p}\right) - \mathcal {F} _ {\theta} \left(y _ {n}\right) \approx \left(\sigma_ {p} + \sigma_ {n}\right) \cdot \frac {\partial \mathcal {F} _ {\theta} (y)}{\partial y} \cdot n. \tag {15} +$$ + +With further Taylor expansion, the loss function can be simplified as: + +$$ +\mathcal {L} _ {p 2 n} \approx (4 + 2 \sigma^ {2}) \left\| \frac {\partial \mathcal {F} _ {\theta} (x)}{\partial x} \cdot n \right\| ^ {2}. \tag {16} +$$ + +Through consistency loss $L_{p2n}$ , we optimize the network $\mathcal{F}_{\theta}$ such that $\frac{\partial\mathcal{F}_{\theta}(x)}{\partial x}\cdot n$ is optimized to drive towards zero, which means that the output $\mathcal{F}_{\theta}(y)$ tends to: + +$$ +\begin{array}{l} \mathcal {F} _ {\theta} (y) = \mathcal {F} _ {\theta} (x + n), (17) \\ = \mathcal {F} _ {\theta} (x) + \frac {\partial \mathcal {F} _ {\theta} (x)}{\partial x} \cdot n + \mathcal {O} \left(n ^ {2}\right), (18) \\ \approx \mathcal {F} _ {\theta} (x). (19) \\ \end{array} +$$ + +Ideally, the neural network $\mathcal{F}_{\theta}$ has learned image priors, the model output $\mathcal{F}_{\theta}(x)$ matches the clean signal $x$ and we have $\mathcal{F}_{\theta}(x) = x$ . Thus, the ultimate goal achieves: + +$$ +\mathcal {F} _ {\theta} (y) \approx x. \tag {20} +$$ + +In summary, the neural network $\mathcal{F}_{\theta}$ gradually learns to denoise the noisy image $y$ effectively. + +# 4. Experiments + +In this section, we first introduce the experimental settings in Section 4.1. After that, we demonstrate the performance and comparisons with other methods in Section 4.2. Finally, we conduct a detailed ablation study in Section 4.3. + +# 4.1. Settings + +In this section, we briefly introduce the experimental settings and details. + +![](images/c96a765d330fce118da6ea577a15bc3e09b17a8a64bbc57d17a9ea9203bf0cb6.jpg) +Figure 6. Qualitative comparisons of P2N against other self-supervised single image denoising methods in SIDD validation dataset. The denoised results of ScoreDVI are always flaky, with discontinuous lines and loss of texture. ZS-N2N does not completely denoise the images. MASH and ATBSN-single tend to either meet aliasing effect or blur the details. In summary, other methods meet aliasing effects and loss of texture, while Positive2Negative achieves both detailed texture preservation and comprehensive noise removal. In the figure, P2N represents Positive2Negative. + +Datasets details. Following previous methods [9, 11], we evaluate the proposed paradigm on the widely-used datasets in self-supervised single image denoising: the SIDD validation dataset [1], the SIDD benchmark dataset [1], CC [29], PolyU [49] and FMDD [53]. The SIDD, PolyU and CC datasets consist of natural sRGB images, while the FMDD dataset consists of fluorescence microscopy images. + +Implementation details. Following the default settings of self-supervised single image denoising methods [9, 11, 22, 40], we employ the same neural network architecture as Noise2Noise [22] and MASH [11] and initialize the neural network parameters with those pre-trained on DIV2K dataset with Gaussian noise [9, 55]. For training details, we employ AdamW [26] optimizer with a learning rate of 0.0001 and set scale parameter $\sigma = 0.75$ , for all the experiments. No image enhancement techniques are applied. + +Compared methods. Following previous methods [9, 11, 24], we mainly compare with self-supervised single image denoising methods. In addition, we also present the results of noisy dataset based self-supervised and supervised image denoising methods only as a reference comparison. All the results of the comparison methods are sourced from previous work [9, 11, 24] or evaluated by ourselves. + +# 4.2. Comparisons + +In this section, we demonstrate the performance of Positive2Negative and compare Positive2Negative with other methods both quantitatively and qualitatively. + +Quantitative comparisons. The quantitative comparisons are described in Table 1. Previous self-supervised single image denoising methods have shown varying performance across different datasets, but Positive2Negative consistently achieves better results. In fact, Positive2Negative even outperforms some dataset-based and supervised methods. Notably, in the online evaluation of the SIDD benchmark dataset, we achieve an improvement of over $0.4\mathrm{dB}$ . + +Qualitative comparisons. The qualitative comparisons are described in Figure 6. Clearly, the Positive2Negative paradigm achieves significant qualitative improvements by either enhancing details or effectively reducing noise. In contrast, other methods face limitations in both detail enhancement and noise removal due to the difficulty in learning denoising, resulting from the information-lossy barrier. + +# 4.3. Ablations + +In this section, we first present the training process in Section 4.3.1. Then, we conduct ablation studies on the components and hyperparameters in Section 4.3.2 and Section 4.3.3, respectively. + +![](images/f39650f028abd0cdfee01828197ff698054cc50ec2be2a5690178a7b25e7184e.jpg) +Figure 7. Visual results of the training process. The top row shows the denoised results, while the bottom row displays the reconstructed negative noisy images in the training process. On the far right of each set is the reference image. The denoised image and reconstructed noisy image gradually become more like the reference images. + +Table 2. Ablation of components. Each step is crucial for Positive2Negative. N2N and N2V represent Noise2Noise and Noise2Void, respectively. + +
MethodCC [29]SIDD [1]
PSNRSSIMPSNRSSIM
Baseline35.400.92733.500.849
RDC+N2N34.080.86825.420.414
RDC+N2V33.790.86124.510.368
N2N+DCS17.550.69824.200.815
N2V+DCS14.140.49422.300.743
RDC+DCS37.920.95435.240.889
+ +# 4.3.1. Ablation of training + +We present visual results of the training process to illustrate the effectiveness of our Positive2Negative paradigm. As shown in Figure 7, throughout the training process, the denoised images and the reconstructed negative noisy images gradually converge towards their respective references. + +# 4.3.2. Ablation of components + +We evaluate the effectiveness of each component of the Positive2Negative paradigm on the CC and SIDD datasets. RDC constructs precise training data and DCS trains robust denoising neural networks. Both RDC and DCS are crucial for Positive2Negative. Replacing either component would result in a decline in performance, as shown in Table 2. + +# 4.3.3. Ablation of hyperparameters + +Throughout the entire training process, the only adjustable hyperparameter is the scale parameter $\sigma$ , which is set to 0.75 for all experiments. As shown in Table 3, the performance remains similar for $\sigma$ values ranging from 0.25 to 0.75. The scale parameter $\sigma$ introduces a dynamic and wide range of noise levels, while the "Fixed" setting corresponds to $\sigma = 0$ . + +Table 3. Ablation of hyperparameters. The scale parameter is the only hyperparameter and is very robust. We fix the scale parameter $\sigma = 0.75$ for all the experiments. + +
MethodCC [29]SIDD [1]
PSNRSSIMPSNRSSIM
Fixed37.330.95235.240.892
σ = 0.2537.640.94535.290.893
σ = 0.5037.840.95235.300.891
σ = 0.7537.920.95435.240.889
σ = 1.0037.860.95435.110.884
+ +# 5. Discussions + +# 5.1. Distribution + +Different norms are suitable for different noise distributions [22]. For simplicity, we have uniformly adopted the $\| \cdot \|_{2\to 1.5}$ norm, as real-world noise is typically centered around a zero mean with slight fluctuations. For noise that remains nearly zero-mean, such as in PolyU [49], the $\| \cdot \|_{2\to 1.5}$ norm and the $\| \cdot \|_2$ norm yield similar performance. For noise that exhibits slight deviations from zero mean, as in SIDD [1], the $\| \cdot \|_{2\to 1.5}$ norm can outperform the $\| \cdot \|_2$ norm, as shown in Table 4. The slight increase in PSNR further suggests that the deviation remains minimal. + +# 5.2. Convergence + +We present the PSNR variation over iterations on SIDD dataset [1] in Figure 8. Around the 100th iteration, the Positive2Negative paradigm converges and shows minimal fluctuations thereafter. Additionally, it is worth noting that the constructed data by RDC is unsuitable for Noise2Noise and Noise2Void. This is mainly because the noise in constructed images by RDC is correlated, which would quickly collapse the training processes of the other two methods. + +Table 4. Discussion of distributions. Different norms are suitable for different noise distributions [22]. For simplicity, we have uniformly adopted the $\| \cdot \|_{2\to 1.5}$ norm. + +
Dataset||·||0||·||1||·||2||·||2→1.5
PolyU [49]37.8838.0238.2038.20
SIDD [1]35.1235.2635.2035.24
+ +![](images/fee973ec3e3634645272cf323da138f5bf041cb7f4e5b2000e11c5bb1133e3e7.jpg) +Figure 8. Convergence comparisons. P2N converges quickly and remains stable after convergence on SIDD. The P2N-Fixed version (where $\sigma = 0.00$ ), in particular, even achieves rapid convergence in approximately 10 iterations. P2N represents Positive2Negative. + +![](images/a52c55de9b69ce102160154975397945498e416d8127e52e2b4c8069d26260d5.jpg) + +# 5.3. Efficiency + +For self-supervised single image denoising methods, computational efficiency is crucial, as each noisy image must be trained individually. Here, we demonstrate the number of training iterations, parameter, and computational cost, as detailed in Table 5. It is obvious that our paradigm achieves significant efficiency improvements. + +# 5.4. Comparison with Noise2Noise and Noise2Void + +Compared with Noise2Noise [22] and Noise2Void [18], Positive2Negative also presents an elegant paradigm The key differences are present in the following aspects. + +Firstly, for the constructed data, the synthetic data constructed by Positive2Negative breaks the information-lossy barrier of Noise2Noise and Noise2Void. Secondly, for the supervision strategy, Positive2Negative enforces consistency between denoised results, whereas Noise2Noise and Noise2Void supervise the consistency between a denoised result and a noisy image. Thirdly, from the assumption perspective, Noise2Noise, Noise2Void, and Positive2Negative all have theoretical guarantees, while Noise2Noise and Noise2Void operate under more restrictive assumptions. Noise2Noise requires two independent noisy images, while Noise2Void assumes that the noise is spatially correlated. + +# 5.5. Differences with R2R and CVF-SID + +For the constructed data, R2R employs new Gaussian noise and applies reciprocal coefficients to the noisy image, while CVF-SID employs two networks to predict three components and applies fixed manual rules. In contrast, our method employs a single network to predict noise and applies random multiscale symmetrical parameters. + +Table 5. Efficiency comparisons under input images with dimensions of $256 \times 256 \times 3$ . With similar parameters and FLOPs, P2N requires the fewest training iterations. P2N converges rapidly, and the P2N-Fixed version (where $\sigma = 0.00$ ) converges even just within 20 iterations. P2N represents Positive2Negative. + +
MethodIterationsParams (M)FLOPs (G)
N2N [22]-0.7018.7
B2U [40]-1.1018.7
Self2Self [36]4500001.009.6
R2R [33]80000.5636.7
ScoreDVI [9]40013.537.9
MASH [11]800 × 30.9911.4
P2N-Fixed<200.9911.4
P2N<1000.9911.4
+ +For the supervision strategy, R2R maps the denoised image to the noisy input, while CVF-SID primarily maps the first cyclic result to the noisy input and maps the second cyclic result to the first cyclic result. In contrast, our method maps the denoised image to another denoised image. + +# 5.6. Limitation and Future Work + +The primary limitation is that Positive2Negative has only been validated on the single image based self-supervised image denoising task. To prevent convergence to trivial solutions such as zero mapping or identity mapping, Positive2Negative requires training with a pre-trained Gaussian denoising model. While the single image based self-supervised denoising task allows training from a pre-trained model on Gaussian noise, the noisy dataset based self-supervised image denoising task trains from scratch. Therefore, extending the Positive2Negative paradigm to the noisy dataset based self-supervised image denoising task could be a promising direction for future work. + +# 6. Conclusion + +In this paper, we propose a self-supervised single image denoising paradigm, Positive2Negative, to break information-lossy barrier. Specifically, Positive2Negative includes two steps: RDC and DCS. Positive2Negative achieves SOTA performance compared to existing methods across various public benchmarks. + +# Acknowledgments + +This work is supported by the National Natural Science Foundation of China under Grants 62322204, 62131003, and the Key Laboratory of Target Cognition and Application Technology. + +# References + +[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S. Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1700, 2018. 1, 3, 5, 6, 7, 8 +[2] Joshua Batson and Loic Royer. Noise2Self: Blind denoising by self-supervision. In International Conference on Machine Learning (ICML), pages 524-533, 2019. 3 +[3] Chinmay Belthangady and Loic A Royer. Applications, promises, and pitfalls of deep learning for fluorescence image reconstruction. In Nature Methods, pages 1215-1225, 2019. 1 +[4] Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing images for learned raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11036-11045, 2019. 5 +[5] Jaeseok Byun, Sungmin Cha, and Taesup Moon. Fbidenoiser: Fast blind image denoiser for poisson-gaussian noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5768-5777, 2021. 3 +[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1703, 2023. 5 +[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision (ECCV), pages 17-33, 2022. 1 +[8] Shiyan Chen, Jiyuan Zhang, Zhaofei Yu, and Tiejun Huang. Exploring efficient asymmetric blind-spots for self-supervised denoising in real-world scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2814-2823, 2024. 1, 2, 3, 5, 6 +[9] Jun Cheng, Tao Liu, and Shan Tan. Score priors guided deep variational inference for unsupervised real-world single image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12937-12948, 2023. 1, 3, 5, 6, 8 +[10] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25974-25984, 2024. 5 +[11] Hamadi Chihaoui and Paolo Favaro. Masked and shuffled blind spot denoising for real-world images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3025-3034, 2024. 1, 2, 3, 5, 6, 8 +[12] Hansen Feng, Lizhi Wang, Yuzhi Wang, Haoqiang Fan, and Hua Huang. Learnability enhancement for low-light raw image denoising: A data perspective. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 370-387, 2024. 3 + +[13] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. In ACM Transactions on Graphics (ToG), pages 1-12, 2016. 1 +[14] Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. Neighbor2Neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14781-14790, 2021. 2 +[15] Yeong Il Jang, Keuntek Lee, Gu Yong Park, Seyun Kim, and Nam Ik Cho. Self-supervised image denoising with downsampled invariance loss and conditional blind-spot network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12196-12205, 2023. 3 +[16] Kwanyoung Kim and Jong Chul Ye. Noise2score: tweedie's approach to self-supervised image denoising without clean images. In Advances in Neural Information Processing Systems, pages 864-874, 2021. 3 +[17] Kwanyoung Kim, Taesung Kwon, and Jong Chul Ye. Noise distribution adaptive self-supervised image denoising using tweedie distribution and score matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1998-2006, 2022. 3 +[18] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void: Learning denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2129-2137, 2019. 1, 2, 3, 5, 8 +[19] Alexander Krull, Tomáš Vičar, Mangal Prakash, Manan Lalit, and Florian Jug. Probabilistic noise2void: Unsupervised content-aware denoising. In Frontiers in Computer Science, pages 2-5, 2020. 3 +[20] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. In Proceedings of International Conference on Neural Information Processing Systems (NeurIPS), pages 6970-6980, 2019. 3 +[21] Wooseok Lee, Sanghyun Son, and Kyoung Mu Lee. APBSN: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17725-17734, 2022. 2, 3, 5 +[22] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2Noise: Learning image restoration without clean data. In International Conference on Machine Learning (ICML), pages 4620-4631, 2018. 1, 2, 3, 4, 6, 7, 8 +[23] Junyi Li, Zhilu Zhang, Xiaoyu Liu, Chaoyu Feng, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Spatially adaptive self-supervised learning for real-world image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914-9924, 2023. 3 +[24] Tong Li, Hansen Feng, Lizhi Wang, Zhiwei Xiong, and Hua Huang. Stimulating the diffusion model for image denoising via adaptive embedding and ensembling. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 8240-8257, 2024. 1, 3, 6 + +[25] Xinyang Li, Yixin Li, Yiliang Zhou, Jiamin Wu, Zhifeng Zhao, Jiaqi Fan, Fei Deng, Zhaofa Wu, Guihua Xiao, Jing He, Yuanlong Zhang, Guoxun Zhang, Xiaowan Hu, Zhang Yi Chen, Xingye and, Hui Qiao, Hao Xie, Yulong Li, Haoqian Wang, Lu Fang, and Qionghai Dai. Real-time denoising enables high-sensitivity fluorescence time-lapse imaging beyond the shot-noise limit. In Nature Biotechnology, page 282-292, 2023. 1 +[26] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. In arXiv preprint arXiv:1711.05101, pages 1-10, 2017. 6 +[27] Youssef Mansour and Reinhard Heckel. Zero-Shot Noise2Noise: Efficient image denoising without any data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14018-14027, 2023. 1, 2, 3, 5, 6 +[28] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2Noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12064-12072, 2020. 2 +[29] Seonghyeon Nam, Youngbae Hwang, Yasuyuki Matsushita, and Seon Joo Kim. A holistic approach to cross-channel image noise modeling and its application to image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1683-1691, 2016. 5, 6, 7 +[30] Reyhaneh Neshatavar, Mohsen Yavartanoo, Sanghyun Son, and Kyoung Mu Lee. CVF-SID: Cyclic multi-variate function for self-supervised image denoising by disentangling noise from image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17583-17591, 2022. 3, 5 +[31] Harry Nyquist. Certain topics in telegraph transmission theory. In Transactions of the American Institute of Electrical Engineers, pages 617-644, 1928. 1, 2 +[32] Yizhong Pan, Xiao Liu, Xiangyu Liao, Yuanzhouhan Cao, and Chao Ren. Random sub-samples generation for self-supervised real image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12150-12159, 2023. 2, 3 +[33] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-Recorrupted: Unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2043–2052, 2021. 2, 3, 5, 8 +[34] Mikhail Papkov, Pavel Chizhov, and Leopold Parts. Swinia: Self-supervised blind-spot image denoising without convolutions. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 7071-7080, 2025. 3 +[35] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1586-1595, 2017. 1 +[36] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2Self with dropout: Learning self-supervised denoising + +from single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1890-1898, 2020. 3, 5, 8 +[37] Claude Elwood Shannon. A mathematical theory of communication. In The Bell System Technical Journal, pages 379-423, 1948. 1, 2 +[38] Andrea Vedaldi Ulyanov Dmitry and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018. 3, 5 +[39] Yuzhi Wang, Haibin Huang, Qin Xu, Jiaming Liu, Yiqun Liu, and Jue Wang. Practical deep raw image denoising on mobile devices. In European Conference on Computer Vision (ECCV), pages 1-16, 2020. 1 +[40] Zejin Wang, Jiazheng Liu, Guoqing Li, and Hua Han. Blind2Unblind: Self-supervised image denoising with visible blind spots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2027-2036, 2022. 2, 3, 6, 8 +[41] Zichun Wang, Ying Fu, Ji Liu, and Yulun Zhang. LG-BPN: Local and global blind-patch network for self-supervised real-world denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18156-18165, 2023. 2, 3 +[42] Kaixuan Wei, Ying Fu, Yinqiang Zheng, and Jiaolong Yang. Physics-based noise modeling for extreme low-light photography. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 8520-8537, 2021. 1, 3 +[43] Martin Weigert, Uwe Schmidt, Tobias Boothe, Andreas Müller, Alexandr Dibrov, Akanksha Jain, Benjamin Wilhelm, Deborah Schmidt, Coleman Broaddus, Sián Culley, et al. Content-aware image restoration: pushing the limits of fluorescence microscopy. In Nature Methods, pages 1090–1097, 2018. 1 +[44] Halbert White. Connectionist nonparametric regression: Multilayer feedforward networks can learn arbitrary mappings. In Neural networks, pages 535-549, 1990. 2 +[45] David Wu, Damian Tan, Marilyn Baird, John DeCampo, Chris White, and Hong Ren Wu. Perceptually lossless medical image coding. In IEEE Transactions on Medical Imaging, pages 335-344, 2006. 2 +[46] Hong Ren Wu, Weisi Lin, and Lina J. Karam. An overview of perceptual processing for digital pictures. In IEEE International Conference on Multimedia and Expo Workshops, pages 113-120, 2012. 2 +[47] Xiaohe Wu, Ming Liu, Yue Cao, Dongwei Ren, and Wangmeng Zuo. Unpaired learning of deep image denoising. In European Conference on Computer Vision (ECCV), pages 352-368, 2020. 3 +[48] Yaochen Xie, Zhengyang Wang, and Shuiwang Ji. Noise2same: Optimizing a self-supervised bound for image denoising. In Advances in Neural Information Processing Systems, pages 20320-20330, 2020. 3 +[49] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. In arXiv preprint arXiv:1804.02603, pages 1-13, 2018. 5, 6, 7, 8 + +[50] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-Clean: Learning self-supervised denoising from corrupted image. In IEEE Transactions on Image Processing (TIP), pages 9316-9329, 2020. 2 +[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728-5739, 2022. 1 +[52] Dan Zhang, Fangfang Zhou, Yuwen Jiang, and Zhengming Fu. Mm-bsn: Self-supervised image denoising for real-world with multi-mask based on blind-spot network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 4189–4198, 2023. 3 +[53] Yide Zhang, Yinhao Zhu, Evan Nichols, Qingfei Wang, Siyuan Zhang, Cody Smith, and Scott Howard. Aoisson-gaussian denoising dataset with real fluorescence microscopy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11710-11718, 2019. 5, 6 +[54] Yi Zhang, Dasong Li, Ka Lung Law, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. IDR: Self-supervised image denoising via iterative data refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2098-2107, 2022. 2 +[55] Dihan Zheng, Sia Huat Tan, Xiaowen Zhang, Zuoqiang Shi, Kaisheng Ma, and Chenglong Bao. An unsupervised deep learning approach for real-world image denoising. In International Conference on Learning Representations (ICLR), pages 1-9, 2020. 5, 6 +[56] Dihan Zheng, Xiaowen Zhang, Kaisheng Ma, and Chenglong Bao. Learn from unpaired data for image restoration: A variational bayes approach. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 5889-5903, 2022. 3, 5 +[57] Yuqian Zhou, Jianbo Jiao, Haibin Huang, Yang Wang, Jue Wang, Honghui Shi, and Thomas Huang. When awgn-based denoiser meets real noises. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 13074-13081, 2020. 3, 5 +[58] Yunhao Zou, Chenggang Yan, and Ying Fu. Iterative denoiser and noise estimator for self-supervised image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13265-13274, 2023. 2 \ No newline at end of file diff --git a/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/images.zip b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..06abb10208251cb46fcfcd1112598afb45e46a61 --- /dev/null +++ b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc70da8e8cde7ff903b945cf995b5c967e82d44e2fc4dcff9708cabeb6a5f059 +size 866399 diff --git a/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/layout.json b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae0eb297c3a1a2e33b804671c24554b9e780d70f --- /dev/null +++ b/2025/Positive2Negative_ Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising/layout.json @@ -0,0 +1,9993 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 489, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 489, + 140 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 489, + 140 + ], + "type": "text", + "content": "Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "spans": [ + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "text", + "content": "Tong Li" + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "text", + "content": " Lizhi Wang" + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "text", + "content": " Zhiyuan Xu" + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "text", + "content": " Lin Zhu" + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "text", + "content": " Wanxuan Lu" + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "text", + "content": " Hua Huang" + }, + { + "bbox": [ + 82, + 160, + 567, + 176 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 175, + 517, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 175, + 517, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 175, + 517, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 175, + 517, + 204 + ], + "type": "text", + "content": " School of Computer Science and Technology, Beijing Institute of Technology " + }, + { + "bbox": [ + 133, + 175, + 517, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 175, + 517, + 204 + ], + "type": "text", + "content": " School of Artificial Intelligence, Beijing Normal University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 203, + 588, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 203, + 588, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 203, + 588, + 217 + ], + "type": "text", + "content": "3 Engineering Research Center of Intelligent Technology and Educational Application, Ministry of Education" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 217, + 504, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 217, + 504, + 232 + ], + "spans": [ + { + "bbox": [ + 146, + 217, + 504, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 146, + 217, + 504, + 232 + ], + "type": "text", + "content": "Aerospace Information Research Institute, Chinese Academy of Sciences" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 284, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 559 + ], + "type": "text", + "content": "Image denoising enhances image quality, serving as a foundational technique across various computational photography applications. The obstacle to clean image acquisition in real scenarios necessitates the development of self-supervised image denoising methods only depending on noisy images, especially a single noisy image. Existing self-supervised image denoising paradigms (Noise2Noise and Noise2Void) rely heavily on information-lossy operations, such as downsampling and masking, culminating in low-quality denoising performance. In this paper, we propose a novel self-supervised single image denoising paradigm, Positive2Negative, to break the information-lossy barrier. Our paradigm involves two key steps: Renoised Data Construction (RDC) and Denoised Consistency Supervision (DCS). RDC renoes the predicted denoised image by the predicted noise to construct multiple noisy images, preserving all the information of the original image. DCS ensures consistency across the multiple denoised images, supervising the network to learn robust denoising. Our Positive2Negative paradigm achieves state-of-the-art performance in self-supervised single image denoising with significant speed improvements. The code is released to the public at https://github.com/Li-Tong-621/P2N." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 592, + 135, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 135, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 135, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 613, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 696 + ], + "type": "text", + "content": "Image denoising plays a vital role in image processing to enhance visual quality [13], serving as a foundational technique across various computational photography applications [3, 39]. The widespread popularity of paired clean-noisy image datasets [1, 35] has catalyzed advancements in supervised methods, yielding remarkable performance [7, 51]. However, supervised methods fall short in" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 259, + 553, + 342 + ], + "blocks": [ + { + "bbox": [ + 318, + 259, + 553, + 342 + ], + "lines": [ + { + "bbox": [ + 318, + 259, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 318, + 259, + 553, + 342 + ], + "type": "image", + "image_path": "6030aee1e7cbeaca920ad71a0402be5f35f7867661fa67f2fd60255d45328d20.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 342, + 479, + 351 + ], + "lines": [ + { + "bbox": [ + 390, + 342, + 479, + 351 + ], + "spans": [ + { + "bbox": [ + 390, + 342, + 479, + 351 + ], + "type": "text", + "content": "Noise2Noise-Based Paradigms" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 353, + 552, + 434 + ], + "blocks": [ + { + "bbox": [ + 317, + 353, + 552, + 434 + ], + "lines": [ + { + "bbox": [ + 317, + 353, + 552, + 434 + ], + "spans": [ + { + "bbox": [ + 317, + 353, + 552, + 434 + ], + "type": "image", + "image_path": "a003c6634aeed0e0b9ed91b36d2b0cf0004cb1e5683948feefe0ed9452d53f89.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 445, + 552, + 526 + ], + "blocks": [ + { + "bbox": [ + 390, + 436, + 475, + 445 + ], + "lines": [ + { + "bbox": [ + 390, + 436, + 475, + 445 + ], + "spans": [ + { + "bbox": [ + 390, + 436, + 475, + 445 + ], + "type": "text", + "content": "Noise2Void-Based Paradigms" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 445, + 552, + 526 + ], + "lines": [ + { + "bbox": [ + 318, + 445, + 552, + 526 + ], + "spans": [ + { + "bbox": [ + 318, + 445, + 552, + 526 + ], + "type": "image", + "image_path": "50164d06ded5d99c14301ce13480195e748ee747b9db641fe115f1da6cecc3b6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 527, + 473, + 536 + ], + "lines": [ + { + "bbox": [ + 391, + 527, + 473, + 536 + ], + "spans": [ + { + "bbox": [ + 391, + 527, + 473, + 536 + ], + "type": "text", + "content": "Positive2Negative Paradigm" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 544, + 555, + 612 + ], + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 612 + ], + "type": "text", + "content": "Figure 1. Information-lossy operations in current paradigms. The noise addition operation reduces the signal-to-noise ratio, the downsampling operation diminishes the sampling density, and the masking operation disregards crucial central pixels. Each of these operations introduces an information-lossy barrier [8, 31, 37], culminating in imprecise denoising." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 627, + 555, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 555, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 555, + 698 + ], + "type": "text", + "content": "dynamic scenes [18, 22], where long exposure results in misalignment and blur [22, 42], hindering the acquisition of clean images. Therefore, advancing self-supervised image denoising methods using only noisy images [25, 43], especially a single noisy image [9, 11, 24, 27], holds significant importance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "The crux of self-supervised image denoising methods" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "type": "text", + "content": "*Corresponding Author: Lizhi Wang (wanglizhi@bnu.edu.cn)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17924" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 293, + 199 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 293, + 199 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 293, + 199 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 293, + 199 + ], + "type": "image", + "image_path": "e4df928a1a33cb13fde08a59c6a9b34639ebf6b25cc8dd8aacfb75a2aeb52132.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 208, + 295, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 242 + ], + "type": "text", + "content": "Figure 2. The information-lossy barrier significantly compromises the image quality, resulting in residual noise, aliasing effects or texture loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 255, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 295, + 434 + ], + "type": "text", + "content": "lies in the construction of the training data and the formulation of the supervision strategy. Current methods are divided into Noise2Noise-based paradigms and Noise2Void-based paradigms, as illustrated in Figure 1. Noise2Noise-based paradigms add additional noise to the noisy image [28, 33] or down sample the noisy image [14, 27] to approximately construct two independently noisy observations as training data. These paradigms learn denoising by supervising one observation with the denoised result of other observation. Noise2Void-based paradigms mask either the noisy image [18, 40] or the neural network [8, 21, 41] to construct masked training data. These paradigms learn denoising by supervising the denoised result of the masked central pixel with the original noisy pixel, where the prediction only relies on the surrounding pixels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 616 + ], + "type": "text", + "content": "However, existing methods rely heavily on information-lossy operations, resulting in low-quality denoising performance [21, 32, 40, 41]. Specifically, the noise addition operation reduces the signal-to-noise ratio, the downsampling operation diminishes the sampling density and the masking operation disregards crucial central pixels. Each of these operations loses the information from the original noisy image [8, 31, 37], inevitably compromising the denoising ability [41, 44]. Consequently, denoised images always suffer from remaining noise, aliasing effects or texture loss [45, 46], as shown in Figure 2. As these information-lossy operations constitute the indispensable foundation of the Noise2Noise-based and Noise2Void-based paradigms, it is imperative to establish a new paradigm that transcends the Noise2Noise-based and Noise2Void-based paradigms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we propose the concise Positive2Negative paradigm to break the information-lossy barrier in self-supervised single image denoising. Our inspiration stems from the observation that noise distributions are approximately symmetrical and centered around zero, implying that the opposite noise also follows the distribution of the original noise. Leveraging this insight, we can construct multiple noisy images corresponding to the same clean image, which" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "serves as the basis of Positive2Negative. Concretely, Positive2Negative relies on two steps: Renoised Data Construction (RDC) and Denoised Consistency Supervision (DCS)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 108, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 204 + ], + "type": "text", + "content": "The basic idea of RDC is to construct training data. Specifically, we design a zero-mean sampling strategy to generate multi-scale Positive noise and Negative noise based on the predicted noise, which are then added back to construct Positive noisy images and Negative noisy images. The constructed data comprehensively covers the information of the given single image; thus the Positive2Negative paradigm breaks the information-lossy barrier." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 287 + ], + "type": "text", + "content": "The basic idea of DCS is to formulate a supervision strategy. Specifically, given the constructed multi-scale noisy images, DCS supervises the network to generate consistent denoised outputs during the training process. Furthermore, we provide a theoretical analysis showing that the Positive2Negative paradigm can learn robust denoising, through supervising the multiple denoised images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 288, + 511, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 288, + 511, + 299 + ], + "spans": [ + { + "bbox": [ + 326, + 288, + 511, + 299 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 301, + 553, + 408 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 328, + 301, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 301, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 328, + 301, + 553, + 336 + ], + "type": "text", + "content": "- We propose a new self-supervised single image denoising paradigm, Positive2Negative, to break the information-lossy barrier." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 328, + 337, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 337, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 328, + 337, + 553, + 360 + ], + "type": "text", + "content": "- We propose a data construction method, which constructs multi-scale similar noisy images for training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 361, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 361, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 328, + 361, + 553, + 384 + ], + "type": "text", + "content": "- We propose a denoising supervision method, which is theoretically guaranteed to learn robust denoising." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 328, + 384, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 384, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 328, + 384, + 553, + 408 + ], + "type": "text", + "content": "- Positive2Negative achieves SOTA performance over self-supervised single image denoising." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 418, + 400, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 400, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 400, + 430 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 438, + 553, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 553, + 474 + ], + "type": "text", + "content": "From the training data perspective, the self-supervised image denoising methods can be classified into noisy dataset based methods and single image based methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 474, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 474, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 474, + 493 + ], + "type": "text", + "content": "2.1. Noisy Dataset based Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 582 + ], + "type": "text", + "content": "Noisy dataset based methods represent the most active research direction in image denoising. These methods learn to denoise from a given noisy dataset. Generally, a larger volume of the noisy dataset improves the denoising performance. As a result, in scenarios where only a single noisy image is available, the denoising performance of whole noisy dataset-based methods decreases significantly." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Noise2Noise [22] learns denoising by mapping one noisy observation to another independent observation of the same scene. Subsequent studies focus on generating observations of the same scene from a single observation, typically achieved through downsampling [14, 27] or by introducing additional noise [28, 33, 50, 54, 58]. After that, another elegant method Noise2Void [18] is proposed. Noise2Void [18] assumes that the noise distribution is zero-mean, the noise is spatially independent, and the signal is spatially correlated. Noise2Void [18] predicts the clean signal by leveraging surrounding noisy signals, with mask" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17925" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "ing the central pixel. Various other works propose alternative masking schemes, either mask the input [11, 36, 40] or mask the network [2, 19, 20, 34, 48, 52]. Noise2Void struggles with real-world noise, which often exhibits spatial correlation. Therefore, subsequent research endeavors to break the noise spatial correlation [8, 21, 23, 41, 47, 57]. There are also some methods that amalgamate frameworks established by both Noise2Noise and Noise2Void [15, 32] or get rid of frameworks established by Noise2Noise and Noise2Void [5, 16, 17, 30, 56], but these methods are usually less developed due to the complexity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 211, + 212, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 212, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 212, + 224 + ], + "type": "text", + "content": "2.2. Single Image based Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 312 + ], + "type": "text", + "content": "Single image based methods represent the most challenging situation for denoising and the least demanding requirements for data. These methods learn to denoise from a given single noisy image without any dependence on paired noisy-clean or noisy datasets. As a result, single image based methods are less developed due to the inherent difficulty, only gaining attention in the past two years." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 468 + ], + "type": "text", + "content": "DIP [38] is the earliest self-supervised image denoising method. DIP [38] learns to map random noise to a given noisy image and employs early stopping to prevent overfitting. Self2Self [36] uses dropout to create two Bernoulli-sampled observations of a noisy image. R2R [33] constructs noisy images by adding new Gaussian noise and applying reciprocal coefficients. ScoreDVI [9] performs real denoising using a pre-trained Gaussian denoiser and an estimated noise model. ZS-N2N [27] extend Noise2Noise [22] to a single image version by downsampling. MASH [11] employs masking and local pixel shuffling to break the noise spatial correlation. Recently, DMID stimulates the diffusion model for self-supervised single image denoising [24]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 296, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 503 + ], + "type": "text", + "content": "In this paper, we propose Positive2Negative, a new self-supervised single image denoising paradigm, to break the information-lossy barrier." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 161, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 161, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 161, + 527 + ], + "type": "text", + "content": "3. Positive2Negative" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 594 + ], + "type": "text", + "content": "In this section, we first demonstrate the motivation of Positive2Negative in Section 3.1. After that, we present the two critical steps of Positive2Negative: RDC and DCS, respectively in Section 3.2 and Section 3.3. Finally, we provide an analysis and proof of Positive2Negative in Section 3.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 600, + 129, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 129, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 129, + 612 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "Our inspiration stems from the widely recognized assumptions that the noise distribution is zero-mean and approximately symmetrical [18, 22], which is consistent with our observation. In our observation experiment, we calculate the noise distribution of a fixed pixel across multiple observations of the same scene. Specifically, following SIDD [1], ELD [42] and PMN [12], we collect noisy image " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " and clean image " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": ". Then we calculate the noise value " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " at each" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 205 + ], + "type": "image", + "image_path": "886d228308623bd181382553d5106bdc587a173134a3eb334629454421e61d37.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 216, + 432, + 281 + ], + "blocks": [ + { + "bbox": [ + 413, + 207, + 468, + 216 + ], + "lines": [ + { + "bbox": [ + 413, + 207, + 468, + 216 + ], + "spans": [ + { + "bbox": [ + 413, + 207, + 468, + 216 + ], + "type": "text", + "content": "(a). noisy image" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 216, + 432, + 281 + ], + "lines": [ + { + "bbox": [ + 316, + 216, + 432, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 216, + 432, + 281 + ], + "type": "image", + "image_path": "697cfef3d7501d8e3ace79726fdad6473499e5855cae7834a6c43d12583cc41d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 282, + 408, + 292 + ], + "lines": [ + { + "bbox": [ + 355, + 282, + 408, + 292 + ], + "spans": [ + { + "bbox": [ + 355, + 282, + 408, + 292 + ], + "type": "text", + "content": "(b). green circle" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 436, + 216, + 552, + 281 + ], + "blocks": [ + { + "bbox": [ + 436, + 216, + 552, + 281 + ], + "lines": [ + { + "bbox": [ + 436, + 216, + 552, + 281 + ], + "spans": [ + { + "bbox": [ + 436, + 216, + 552, + 281 + ], + "type": "image", + "image_path": "c35e222e7404ee6411235128c20f70c831b3088ee5aa5a0c6998326b95f7539b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 282, + 526, + 292 + ], + "lines": [ + { + "bbox": [ + 481, + 282, + 526, + 292 + ], + "spans": [ + { + "bbox": [ + 481, + 282, + 526, + 292 + ], + "type": "text", + "content": "(c). red circle" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 319, + 378, + 376, + 435 + ], + "blocks": [ + { + "bbox": [ + 313, + 300, + 554, + 367 + ], + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 367 + ], + "type": "text", + "content": "Figure 3. Noise distribution is zero-mean and approximately symmetrical. (a) shows a noisy image. (b) and (c) show the noise distributions, which are calculated at the center pixels marked by the green circle and the red circle in the noisy image, respectively. It is evident that the noise distribution is zero-mean and approximately symmetrical." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 378, + 376, + 435 + ], + "lines": [ + { + "bbox": [ + 319, + 378, + 376, + 435 + ], + "spans": [ + { + "bbox": [ + 319, + 378, + 376, + 435 + ], + "type": "image", + "image_path": "b516d66181a60f87abc87956a2bb38162b2c6694552d10e4b25316008d5f73ef.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 437, + 374, + 447 + ], + "lines": [ + { + "bbox": [ + 322, + 437, + 374, + 447 + ], + "spans": [ + { + "bbox": [ + 322, + 437, + 374, + 447 + ], + "type": "text", + "content": "Original noise" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 449, + 351, + 456 + ], + "lines": [ + { + "bbox": [ + 343, + 449, + 351, + 456 + ], + "spans": [ + { + "bbox": [ + 343, + 449, + 351, + 456 + ], + "type": "text", + "content": "n" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 378, + 378, + 435, + 436 + ], + "blocks": [ + { + "bbox": [ + 378, + 378, + 435, + 436 + ], + "lines": [ + { + "bbox": [ + 378, + 378, + 435, + 436 + ], + "spans": [ + { + "bbox": [ + 378, + 378, + 435, + 436 + ], + "type": "image", + "image_path": "699519e8f7265c12a2d3c7d1429fb3a80cade584cea652d197f3d4d286490148.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 437, + 432, + 447 + ], + "lines": [ + { + "bbox": [ + 380, + 437, + 432, + 447 + ], + "spans": [ + { + "bbox": [ + 380, + 437, + 432, + 447 + ], + "type": "text", + "content": "Original noisy" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 449, + 425, + 458 + ], + "lines": [ + { + "bbox": [ + 386, + 449, + 425, + 458 + ], + "spans": [ + { + "bbox": [ + 386, + 449, + 425, + 458 + ], + "type": "inline_equation", + "content": "y = c + n" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 468, + 554, + 502 + ], + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 502 + ], + "type": "text", + "content": "Figure 4. The opposite noise " + }, + { + "bbox": [ + 313, + 468, + 554, + 502 + ], + "type": "inline_equation", + "content": "-n" + }, + { + "bbox": [ + 313, + 468, + 554, + 502 + ], + "type": "text", + "content": " is similar to the original noise " + }, + { + "bbox": [ + 313, + 468, + 554, + 502 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 468, + 554, + 502 + ], + "type": "text", + "content": ". The opposite noise " + }, + { + "bbox": [ + 313, + 468, + 554, + 502 + ], + "type": "inline_equation", + "content": "-n" + }, + { + "bbox": [ + 313, + 468, + 554, + 502 + ], + "type": "text", + "content": " approximately follows the original noise's distribution, which is zero-mean and approximately symmetrical." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 436, + 378, + 492, + 435 + ], + "blocks": [ + { + "bbox": [ + 436, + 378, + 492, + 435 + ], + "lines": [ + { + "bbox": [ + 436, + 378, + 492, + 435 + ], + "spans": [ + { + "bbox": [ + 436, + 378, + 492, + 435 + ], + "type": "image", + "image_path": "7f331d157e6699eb6a5d689453d34f19844d1e76cbd046d7bd9f69f4851a252a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 437, + 491, + 447 + ], + "lines": [ + { + "bbox": [ + 437, + 437, + 491, + 447 + ], + "spans": [ + { + "bbox": [ + 437, + 437, + 491, + 447 + ], + "type": "text", + "content": "Opposite noise" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 449, + 482, + 458 + ], + "lines": [ + { + "bbox": [ + 446, + 449, + 482, + 458 + ], + "spans": [ + { + "bbox": [ + 446, + 449, + 482, + 458 + ], + "type": "inline_equation", + "content": "n_{n} = -n" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 494, + 378, + 551, + 436 + ], + "blocks": [ + { + "bbox": [ + 494, + 378, + 551, + 436 + ], + "lines": [ + { + "bbox": [ + 494, + 378, + 551, + 436 + ], + "spans": [ + { + "bbox": [ + 494, + 378, + 551, + 436 + ], + "type": "image", + "image_path": "16ae37fb525c992e55db198498e7f9ce967b66372b1017ed8b05a504e3caee60.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 494, + 437, + 550, + 447 + ], + "lines": [ + { + "bbox": [ + 494, + 437, + 550, + 447 + ], + "spans": [ + { + "bbox": [ + 494, + 437, + 550, + 447 + ], + "type": "text", + "content": "Opposite noisy" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 493, + 449, + 550, + 459 + ], + "lines": [ + { + "bbox": [ + 493, + 449, + 550, + 459 + ], + "spans": [ + { + "bbox": [ + 493, + 449, + 550, + 459 + ], + "type": "inline_equation", + "content": "y_{n} = c + (-n)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "content": "spatial location " + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "content": ". Additionally, we compute the frequency of " + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "inline_equation", + "content": "n_{i,j}" + }, + { + "bbox": [ + 313, + 522, + 554, + 570 + ], + "type": "text", + "content": " cross noisy images of the same scene. The results are shown in Figure 3, which demonstrates the noise distribution is zero-mean and approximately symmetrical." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "text", + "content": "Based on this observation, an intuitive insight is that the opposite noise " + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "inline_equation", + "content": "-n" + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "text", + "content": " also follows the same distribution as that of the original noise " + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "text", + "content": ". This intuitive idea inspires us to construct the opposite noise " + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "inline_equation", + "content": "n_{n}" + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "text", + "content": " and corresponding opposite noisy image " + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "inline_equation", + "content": "y_{n}" + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 386, + 642, + 553, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 642, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 386, + 642, + 553, + 653 + ], + "type": "interline_equation", + "content": "n _ {n} = - n, \\tag {1}", + "image_path": "c7c69048f73f67976b64e616409ec59b7dedb1668d509ae022a009556d955ba1.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 386, + 656, + 553, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 656, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 386, + 656, + 553, + 668 + ], + "type": "interline_equation", + "content": "y _ {n} = x + n _ {n} = x - n. \\tag {2}", + "image_path": "d6bb9698b83b0714dc4d8cefc02047fbd18762bec0f86c7cdfad26d791449b85.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "The opposite noise " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "n_n" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " and opposite noisy image " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "y_n" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " are similar to the original noise " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " and original noisy image " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ", respectively, as shown in Figure 4." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17926" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 186 + ], + "type": "image", + "image_path": "0b829f6c936d354a94499a951522460a36a8e9749d5a496bdbaedc8629236c4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 197, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 197, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 197, + 555, + 266 + ], + "type": "text", + "content": "Figure 5. Overview of the proposed self-supervised single image denoising paradigm. For training, the Positive2Negative paradigm consists of 2 steps. The first step is Renoised Data Construction (RDC): multiply the predicted noise by positive and negative coefficients to construct multi-scale Positive noise and Negative noise, respectively. The second step is Denoised Consistency Supervision (DCS): train the network through supervising the consistency of the denoised images. For inference, denoising can be achieved in just one pass through the network. (The networks shown in the figure are the same one. In addition, for better visualization, the images within the red boxes on the bottom left have been zoomed in and are displayed in the red boxes on the bottom right.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 285, + 294, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 294, + 380 + ], + "type": "text", + "content": "Based on the above observations and insights, we propose the Positive2Negative paradigm, as shown in Figure 5. Firstly, we extend our intuitive idea into a flexible and upgraded version: Renoised Data Construction (RDC), which constructs multiple positive and negative noises. Secondly, we propose the Denoised Consistency Supervision (DCS), which supervises the consistency of the predicted denoised images to learn a denoising neural network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 380, + 295, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 440 + ], + "type": "text", + "content": "It is important to note that we do not require the noise distribution to be strictly zero-mean or symmetric. Although real noise generally has a zero-mean, certain cases where the noise is not strictly symmetric can be easily addressed by adjusting the norm in the loss function [22]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 446, + 244, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 244, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 244, + 459 + ], + "type": "text", + "content": "3.2. Renoised Data Construction (RDC)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "content": "The basic idea of RDC is to denoise the image to predict a denoised image and renoise the predicted denoised image by the predicted noise to construct multiple noisy images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "content": "For a given noisy image " + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "inline_equation", + "content": "y = x + n" + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "content": " presents the clean image and " + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "content": " presents the noise. RDC firstly denoises the noisy image " + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "content": " once through the neural network " + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "content": " parametered by " + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 501, + 296, + 548 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 150, + 556, + 294, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 556, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 150, + 556, + 294, + 569 + ], + "type": "interline_equation", + "content": "\\hat {x} = \\mathcal {F} _ {\\theta} (y). \\tag {3}", + "image_path": "15f2d774bd5863b90e8a8ccbe4f45427d5b1a198c8ddb31605c7f870ba15da2b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 577, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 295, + 612 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 577, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 55, + 577, + 295, + 612 + ], + "type": "text", + "content": " is the predicted denoised image of the noisy observation " + }, + { + "bbox": [ + 55, + 577, + 295, + 612 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 577, + 295, + 612 + ], + "type": "text", + "content": ". Then we can calculate the predicted noise " + }, + { + "bbox": [ + 55, + 577, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\hat{n}" + }, + { + "bbox": [ + 55, + 577, + 295, + 612 + ], + "type": "text", + "content": " within the noisy observation " + }, + { + "bbox": [ + 55, + 577, + 295, + 612 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 577, + 295, + 612 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 621, + 294, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 621, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 151, + 621, + 294, + 633 + ], + "type": "interline_equation", + "content": "\\hat {n} = y - \\hat {x}. \\tag {4}", + "image_path": "da816c55d3cc6225bd81a946011fc79b10fe9c2c8afd13aa61121cb058ab389f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Next, RDC constructs noisy images as training data based on the predicted noise " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{n}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and the predicted denoised image " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". As noise is zero-mean, we further design a zero-mean sampling strategy to construct new noises. Specifically, RDC multiplies the predicted noise " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{n}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " by both positive scale parameters " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{p}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and negative scale parameters " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "-\\sigma_{n}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 285, + 555, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 309 + ], + "type": "text", + "content": "sampled from Gaussian distribution " + }, + { + "bbox": [ + 313, + 285, + 555, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 313, + 285, + 555, + 309 + ], + "type": "text", + "content": ", constructing multiscale Positive noise " + }, + { + "bbox": [ + 313, + 285, + 555, + 309 + ], + "type": "inline_equation", + "content": "n_p" + }, + { + "bbox": [ + 313, + 285, + 555, + 309 + ], + "type": "text", + "content": " and Negative noise " + }, + { + "bbox": [ + 313, + 285, + 555, + 309 + ], + "type": "inline_equation", + "content": "n_n" + }, + { + "bbox": [ + 313, + 285, + 555, + 309 + ], + "type": "text", + "content": ", respectively:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 380, + 331, + 555, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 331, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 380, + 331, + 555, + 344 + ], + "type": "interline_equation", + "content": "\\sigma_ {n}, \\sigma_ {p} \\sim \\mathcal {N} (1, \\sigma), \\tag {5}", + "image_path": "074f199f1235c9428262a2bdd54a8fae6e13bc767eda2c6104048c7be394e0f1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 382, + 346, + 553, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 346, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 382, + 346, + 553, + 359 + ], + "type": "interline_equation", + "content": "n _ {p} = \\sigma_ {n} \\hat {n}, \\tag {6}", + "image_path": "bb6089a97757f6475b4d6b2860c279bf9aabe137f42eae4cf4a9453d2918b799.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 382, + 361, + 553, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 361, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 382, + 361, + 553, + 373 + ], + "type": "interline_equation", + "content": "y _ {p} = \\hat {x} + n _ {p} = \\hat {x} + \\sigma_ {n} \\hat {n}, \\tag {7}", + "image_path": "c9e473f1946c2ce97c9caf8e3f0b394d7add0a913dd80b7e33bc5eacf9287c22.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 381, + 377, + 553, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 377, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 381, + 377, + 553, + 389 + ], + "type": "interline_equation", + "content": "n _ {n} = - \\sigma_ {p} \\hat {n}, \\tag {8}", + "image_path": "1ccc3e7b5adfb0058ae527c411bd2ec3dd5e9a4454361c18ddba53bb240e1fc2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 382, + 392, + 553, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 392, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 382, + 392, + 553, + 403 + ], + "type": "interline_equation", + "content": "y _ {n} = \\hat {x} + n _ {n} = \\hat {x} - \\sigma_ {p} \\hat {n}. \\tag {9}", + "image_path": "3536c4ef7d0c136cf2a7e49200021336319b85be4f7d7f9cd536439f78a5f1bf.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 423, + 528, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 528, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 528, + 436 + ], + "type": "text", + "content": "3.3. Denoised Consistency Supervision (DCS)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "text", + "content": "The basic idea of DCS is that the noisy observations " + }, + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "inline_equation", + "content": "y_{p}" + }, + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "inline_equation", + "content": "y_{n}" + }, + { + "bbox": [ + 313, + 441, + 555, + 489 + ], + "type": "text", + "content": " should correspond to the same clean image. Therefore, the loss supervises the consistency of the predicted denoised images:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 373, + 500, + 553, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 500, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 373, + 500, + 553, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p 2 n} = \\left\\| \\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) - \\mathcal {F} _ {\\theta} \\left(y _ {n}\\right) \\right\\|, \\tag {10}", + "image_path": "a18f7adc3c2a5620721b1d2483d78085fae4471869f92aa93d5d07419e14d8e1.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 523, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 555, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 523, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 313, + 523, + 555, + 570 + ], + "type": "text", + "content": " represents the norm. Considering the general nature of noise [22], we employ a gradually varying norm function " + }, + { + "bbox": [ + 313, + 523, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|_{2 \\to 1.5}" + }, + { + "bbox": [ + 313, + 523, + 555, + 570 + ], + "type": "text", + "content": " to cover a broader range of unknown distributions:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 385, + 582, + 553, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 582, + 553, + 595 + ], + "spans": [ + { + "bbox": [ + 385, + 582, + 553, + 595 + ], + "type": "interline_equation", + "content": "\\| x \\| _ {2 \\rightarrow 1. 5} = (| x | + \\epsilon) ^ {\\gamma}, \\tag {11}", + "image_path": "0db13221fb18ac1cb33c16bbd3043a164f769c349e12ff775f6f75676901f799.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\epsilon = 10^{-8}" + }, + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "text", + "content": " varies linearly from 2 to 1.5 during training. The advantage of the varying " + }, + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{2\\to 1.5}" + }, + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "text", + "content": " norm is that " + }, + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{2\\to 1.5}" + }, + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "text", + "content": " does not require fine-tuning of parameters, yet " + }, + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{2\\to 1.5}" + }, + { + "bbox": [ + 313, + 605, + 555, + 653 + ], + "type": "text", + "content": " consistently delivers robust performance." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "The noisy observations " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "y_{p}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "y_{n}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " contain multi-scale noise. Through DCS, the neural network " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " gradually learns to become insensitive to input noise and consistently outputs the clean signal, achieving the desired denoising performance." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17927" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 123, + 555, + 384 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on SIDD, CC, PolyU and FMDD datasets. The best and second-best results (PSNR↑ / SSIM↑) are marked in red and blue in self-supervised single image denoising methods. The results of noisy dataset based self-supervised methods and supervised methods are also provided, only as a reference comparison. SIDD val and SIDD ben represent the SIDD validation dataset and the SIDD online benchmark, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 123, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 56, + 123, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 123, + 555, + 384 + ], + "type": "table", + "html": "
CategoryMethodSIDD val [1]SIDD ben [1]CC [29]PolyU [49]FMDD [53]
SupervisedUnprocessing [4]28.12 / 0.52931.64 / 0.53135.18 / 0.90836.99 / 0.94128.45 / 0.574
MaskedDenoising [6]28.67 / 0.60431.99 / 0.60133.87 / 0.93034.56 / 0.93629.86 / 0.643
CLIPDenoising [10]34.79 / 0.86635.82 / 0.85936.30 / 0.94137.54 / 0.96030.56 / 0.698
Self-Supervised (Noisy Dataset Based Methods)Noise2Void [18]29.35 / 0.65127.68 / 0.66832.27 / 0.86233.83 / 0.873-
CVF-SID [30]34.81 / 0.87235.05 / 0.85629.10 /0.85233.05 / 0.91131.64 / 0.837
LUD-VAE [56]34.91 / 0.89235.49 / 0.88331.25 / 0.92033.68 / 0.94232.00 / 0.822
Self-Supervised (Single Image Based Methods)DIP [38]32.11 / 0.740-35.61 / 0.91237.17 / 0.91232.90 / 0.854
PD-denoising [57]33.97 / 0.820-35.85 / 0.92337.04 / 0.94033.01 / 0.856
NN+denoiser [55]--36.52 / 0.94337.66 / 0.95632.21 / 0.831
Self2Self [36]29.46 / 0.59529.51 / 0.65137.44 / 0.94837.52 / 0.95130.76 / 0.695
R2R [33]24.59 / 0.387-33.43 / 0.86536.23 / 0.93127.17 / 0.526
APBSN-single [21]30.90 / 0.818-27.72 / 0.89129.61 / 0.89728.43 / 0.804
ScoreDVI [9]34.75 / 0.85635.39 / 0.85937.09 / 0.94537.77 / 0.95933.10 / 0.865
ZS-N2N [27]25.59 / 0.43330.19 / 0.42833.51 / 0.85135.99 / 0.91431.65 / 0.767
MASH [11]35.06 / 0.85134.80 / 0.81431.17 / 0.88937.62 / 0.93233.71 / 0.882
ATBSN-single [8]32.39 / 0.79627.86 / 0.31231.87 / 0.89733.21 / 0.92531.23 / 0.829
Positive2Negative35.24 / 0.88935.80 / 0.87537.92 / 0.95438.20 / 0.96134.00 / 0.887
", + "image_path": "ba49c0365d71233d462fb23fff8e5d65d7c52dd6b5eb17f9d6fcac0d0cc00c32.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 403, + 118, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 118, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 118, + 416 + ], + "type": "text", + "content": "3.4. Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 421, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 295, + 456 + ], + "type": "text", + "content": "In this section, we provide an analysis and proof demonstrating how the Positive2Negative paradigm achieves denoising." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 480 + ], + "type": "text", + "content": "Taking the 2-norm as an example, the loss function can be expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 487, + 294, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 487, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 104, + 487, + 294, + 504 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p 2 n} = \\mathbb {E} \\left[ \\| \\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) - \\mathcal {F} _ {\\theta} \\left(y _ {n}\\right) \\| ^ {2} \\right]. \\tag {12}", + "image_path": "571d47f3a991d06f4bad60ae1f667e558a44a091a0af55c9361908df8967ef57.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "text", + "content": "Through Taylor expansion, we can approximate the predicted denoised images " + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}(y_p)" + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}(y_n)" + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 541, + 295, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 541, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 63, + 541, + 295, + 567 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) = \\mathcal {F} _ {\\theta} \\left(\\mathcal {F} _ {\\theta} (y)\\right) + \\sigma_ {p} \\cdot \\frac {\\partial \\mathcal {F} _ {\\theta} (y)}{\\partial y} \\cdot n + \\mathcal {O} \\left(n ^ {2}\\right), \\tag {13}", + "image_path": "b3b7655fc9f27f4efe966e7a51098dff3103d2f1fe511032f29d18782fd989ea.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 568, + 295, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 568, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 63, + 568, + 295, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\theta} (y _ {n}) = \\mathcal {F} _ {\\theta} \\left(\\mathcal {F} _ {\\theta} (y)\\right) - \\sigma_ {n} \\cdot \\frac {\\partial \\mathcal {F} _ {\\theta} (y)}{\\partial y} \\cdot n + \\mathcal {O} \\left(n ^ {2}\\right). \\tag {14}", + "image_path": "8c7fae7d3dfd994bdf39dcc483e85f334f765277fae2ad9ce50d652334667c41.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 601, + 233, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 601, + 233, + 613 + ], + "spans": [ + { + "bbox": [ + 67, + 601, + 233, + 613 + ], + "type": "text", + "content": "Substituting the approximations, we get:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 74, + 619, + 295, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 619, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 74, + 619, + 295, + 645 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\theta} \\left(y _ {p}\\right) - \\mathcal {F} _ {\\theta} \\left(y _ {n}\\right) \\approx \\left(\\sigma_ {p} + \\sigma_ {n}\\right) \\cdot \\frac {\\partial \\mathcal {F} _ {\\theta} (y)}{\\partial y} \\cdot n. \\tag {15}", + "image_path": "58c41aa3c87475a4d55fe097e69dce4a3f9e0a4574b9c1802c497c97df123e42.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 651, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 676 + ], + "type": "text", + "content": "With further Taylor expansion, the loss function can be simplified as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 102, + 682, + 295, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 682, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 102, + 682, + 295, + 712 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p 2 n} \\approx (4 + 2 \\sigma^ {2}) \\left\\| \\frac {\\partial \\mathcal {F} _ {\\theta} (x)}{\\partial x} \\cdot n \\right\\| ^ {2}. \\tag {16}", + "image_path": "d32ff7e5a9e6d5e319f75ada034217467cbe2fd32c0d932a2543c98b429a9cf5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 404, + 555, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 555, + 442 + ], + "type": "text", + "content": "Through consistency loss " + }, + { + "bbox": [ + 313, + 404, + 555, + 442 + ], + "type": "inline_equation", + "content": "L_{p2n}" + }, + { + "bbox": [ + 313, + 404, + 555, + 442 + ], + "type": "text", + "content": ", we optimize the network " + }, + { + "bbox": [ + 313, + 404, + 555, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 313, + 404, + 555, + 442 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 404, + 555, + 442 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{F}_{\\theta}(x)}{\\partial x}\\cdot n" + }, + { + "bbox": [ + 313, + 404, + 555, + 442 + ], + "type": "text", + "content": " is optimized to drive towards zero, which means that the output " + }, + { + "bbox": [ + 313, + 404, + 555, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}(y)" + }, + { + "bbox": [ + 313, + 404, + 555, + 442 + ], + "type": "text", + "content": " tends to:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 404, + 555, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 555, + 501 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {F} _ {\\theta} (y) = \\mathcal {F} _ {\\theta} (x + n), (17) \\\\ = \\mathcal {F} _ {\\theta} (x) + \\frac {\\partial \\mathcal {F} _ {\\theta} (x)}{\\partial x} \\cdot n + \\mathcal {O} \\left(n ^ {2}\\right), (18) \\\\ \\approx \\mathcal {F} _ {\\theta} (x). (19) \\\\ \\end{array}", + "image_path": "c2ca70b5e289af79641701c08cafe9e0da244154139abe64ad942272223ad3a4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 506, + 554, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 542 + ], + "type": "text", + "content": "Ideally, the neural network " + }, + { + "bbox": [ + 313, + 506, + 554, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 313, + 506, + 554, + 542 + ], + "type": "text", + "content": " has learned image priors, the model output " + }, + { + "bbox": [ + 313, + 506, + 554, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}(x)" + }, + { + "bbox": [ + 313, + 506, + 554, + 542 + ], + "type": "text", + "content": " matches the clean signal " + }, + { + "bbox": [ + 313, + 506, + 554, + 542 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 506, + 554, + 542 + ], + "type": "text", + "content": " and we have " + }, + { + "bbox": [ + 313, + 506, + 554, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}(x) = x" + }, + { + "bbox": [ + 313, + 506, + 554, + 542 + ], + "type": "text", + "content": ". Thus, the ultimate goal achieves:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 409, + 548, + 553, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 548, + 553, + 560 + ], + "spans": [ + { + "bbox": [ + 409, + 548, + 553, + 560 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\theta} (y) \\approx x. \\tag {20}", + "image_path": "346c0f78c4a72d208a9db68858aba2201f5b0b199db27d76a302e4bdd25140dd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 565, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 555, + 590 + ], + "type": "text", + "content": "In summary, the neural network " + }, + { + "bbox": [ + 313, + 565, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 313, + 565, + 555, + 590 + ], + "type": "text", + "content": " gradually learns to denoise the noisy image " + }, + { + "bbox": [ + 313, + 565, + 555, + 590 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 565, + 555, + 590 + ], + "type": "text", + "content": " effectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 598, + 395, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 395, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 395, + 612 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "text", + "content": "In this section, we first introduce the experimental settings in Section 4.1. After that, we demonstrate the performance and comparisons with other methods in Section 4.2. Finally, we conduct a detailed ablation study in Section 4.3." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 672, + 375, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 375, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 375, + 685 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "In this section, we briefly introduce the experimental settings and details." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17928" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 547, + 331 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 547, + 331 + ], + "type": "image", + "image_path": "c96a765d330fce118da6ea577a15bc3e09b17a8a64bbc57d17a9ea9203bf0cb6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 340, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 555, + 396 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons of P2N against other self-supervised single image denoising methods in SIDD validation dataset. The denoised results of ScoreDVI are always flaky, with discontinuous lines and loss of texture. ZS-N2N does not completely denoise the images. MASH and ATBSN-single tend to either meet aliasing effect or blur the details. In summary, other methods meet aliasing effects and loss of texture, while Positive2Negative achieves both detailed texture preservation and comprehensive noise removal. In the figure, P2N represents Positive2Negative." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 406, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 296, + 491 + ], + "type": "text", + "content": "Datasets details. Following previous methods [9, 11], we evaluate the proposed paradigm on the widely-used datasets in self-supervised single image denoising: the SIDD validation dataset [1], the SIDD benchmark dataset [1], CC [29], PolyU [49] and FMDD [53]. The SIDD, PolyU and CC datasets consist of natural sRGB images, while the FMDD dataset consists of fluorescence microscopy images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 506, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 615 + ], + "type": "text", + "content": "Implementation details. Following the default settings of self-supervised single image denoising methods [9, 11, 22, 40], we employ the same neural network architecture as Noise2Noise [22] and MASH [11] and initialize the neural network parameters with those pre-trained on DIV2K dataset with Gaussian noise [9, 55]. For training details, we employ AdamW [26] optimizer with a learning rate of 0.0001 and set scale parameter " + }, + { + "bbox": [ + 55, + 506, + 296, + 615 + ], + "type": "inline_equation", + "content": "\\sigma = 0.75" + }, + { + "bbox": [ + 55, + 506, + 296, + 615 + ], + "type": "text", + "content": ", for all the experiments. No image enhancement techniques are applied." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Compared methods. Following previous methods [9, 11, 24], we mainly compare with self-supervised single image denoising methods. In addition, we also present the results of noisy dataset based self-supervised and supervised image denoising methods only as a reference comparison. All the results of the comparison methods are sourced from previous work [9, 11, 24] or evaluated by ourselves." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 406, + 400, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 400, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 400, + 419 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 423, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 460 + ], + "type": "text", + "content": "In this section, we demonstrate the performance of Positive2Negative and compare Positive2Negative with other methods both quantitatively and qualitatively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "type": "text", + "content": "Quantitative comparisons. The quantitative comparisons are described in Table 1. Previous self-supervised single image denoising methods have shown varying performance across different datasets, but Positive2Negative consistently achieves better results. In fact, Positive2Negative even outperforms some dataset-based and supervised methods. Notably, in the online evaluation of the SIDD benchmark dataset, we achieve an improvement of over " + }, + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 555, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 640 + ], + "type": "text", + "content": "Qualitative comparisons. The qualitative comparisons are described in Figure 6. Clearly, the Positive2Negative paradigm achieves significant qualitative improvements by either enhancing details or effectively reducing noise. In contrast, other methods face limitations in both detail enhancement and noise removal due to the difficulty in learning denoising, resulting from the information-lossy barrier." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 647, + 382, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 382, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 382, + 659 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "In this section, we first present the training process in Section 4.3.1. Then, we conduct ablation studies on the components and hyperparameters in Section 4.3.2 and Section 4.3.3, respectively." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17929" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 69, + 552, + 229 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 552, + 229 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 552, + 229 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 552, + 229 + ], + "type": "image", + "image_path": "f39650f028abd0cdfee01828197ff698054cc50ec2be2a5690178a7b25e7184e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 271 + ], + "type": "text", + "content": "Figure 7. Visual results of the training process. The top row shows the denoised results, while the bottom row displays the reconstructed negative noisy images in the training process. On the far right of each set is the reference image. The denoised image and reconstructed noisy image gradually become more like the reference images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 323, + 294, + 435 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 297, + 315 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 297, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 297, + 315 + ], + "type": "text", + "content": "Table 2. Ablation of components. Each step is crucial for Positive2Negative. N2N and N2V represent Noise2Noise and Noise2Void, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 323, + 294, + 435 + ], + "lines": [ + { + "bbox": [ + 58, + 323, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 58, + 323, + 294, + 435 + ], + "type": "table", + "html": "
MethodCC [29]SIDD [1]
PSNRSSIMPSNRSSIM
Baseline35.400.92733.500.849
RDC+N2N34.080.86825.420.414
RDC+N2V33.790.86124.510.368
N2N+DCS17.550.69824.200.815
N2V+DCS14.140.49422.300.743
RDC+DCS37.920.95435.240.889
", + "image_path": "1ad394bf256415ba94def0fe56a911aeb7ecc12a04955e4a4dd590880c3c7f81.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 169, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 169, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 169, + 466 + ], + "type": "text", + "content": "4.3.1. Ablation of training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 529 + ], + "type": "text", + "content": "We present visual results of the training process to illustrate the effectiveness of our Positive2Negative paradigm. As shown in Figure 7, throughout the training process, the denoised images and the reconstructed negative noisy images gradually converge towards their respective references." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 186, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 186, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 186, + 547 + ], + "type": "text", + "content": "4.3.2. Ablation of components" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 549, + 296, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 296, + 621 + ], + "type": "text", + "content": "We evaluate the effectiveness of each component of the Positive2Negative paradigm on the CC and SIDD datasets. RDC constructs precise training data and DCS trains robust denoising neural networks. Both RDC and DCS are crucial for Positive2Negative. Replacing either component would result in a decline in performance, as shown in Table 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 627, + 208, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 208, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 208, + 639 + ], + "type": "text", + "content": "4.3.3. Ablation of hyperparameters" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "Throughout the entire training process, the only adjustable hyperparameter is the scale parameter " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": ", which is set to 0.75 for all experiments. As shown in Table 3, the performance remains similar for " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " values ranging from 0.25 to 0.75. The scale parameter " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " introduces a dynamic and wide range of noise levels, while the \"Fixed\" setting corresponds to " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sigma = 0" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 318, + 323, + 553, + 414 + ], + "blocks": [ + { + "bbox": [ + 313, + 281, + 555, + 315 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 315 + ], + "type": "text", + "content": "Table 3. Ablation of hyperparameters. The scale parameter is the only hyperparameter and is very robust. We fix the scale parameter " + }, + { + "bbox": [ + 313, + 281, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\sigma = 0.75" + }, + { + "bbox": [ + 313, + 281, + 555, + 315 + ], + "type": "text", + "content": " for all the experiments." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 323, + 553, + 414 + ], + "lines": [ + { + "bbox": [ + 318, + 323, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 318, + 323, + 553, + 414 + ], + "type": "table", + "html": "
MethodCC [29]SIDD [1]
PSNRSSIMPSNRSSIM
Fixed37.330.95235.240.892
σ = 0.2537.640.94535.290.893
σ = 0.5037.840.95235.300.891
σ = 0.7537.920.95435.240.889
σ = 1.0037.860.95435.110.884
", + "image_path": "c56ecc646ee4af664bf995b4ee2059e1505f4fa791710672117d01848c09bcdc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 433, + 388, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 433, + 388, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 433, + 388, + 445 + ], + "type": "text", + "content": "5. Discussions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 453, + 394, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 453, + 394, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 394, + 464 + ], + "type": "text", + "content": "5.1. Distribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "text", + "content": "Different norms are suitable for different noise distributions [22]. For simplicity, we have uniformly adopted the " + }, + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{2\\to 1.5}" + }, + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "text", + "content": " norm, as real-world noise is typically centered around a zero mean with slight fluctuations. For noise that remains nearly zero-mean, such as in PolyU [49], the " + }, + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{2\\to 1.5}" + }, + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "text", + "content": " norm and the " + }, + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "text", + "content": " norm yield similar performance. For noise that exhibits slight deviations from zero mean, as in SIDD [1], the " + }, + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{2\\to 1.5}" + }, + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "text", + "content": " norm can outperform the " + }, + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 313, + 472, + 555, + 591 + ], + "type": "text", + "content": " norm, as shown in Table 4. The slight increase in PSNR further suggests that the deviation remains minimal." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 600, + 397, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 397, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 397, + 613 + ], + "type": "text", + "content": "5.2. Convergence" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "We present the PSNR variation over iterations on SIDD dataset [1] in Figure 8. Around the 100th iteration, the Positive2Negative paradigm converges and shows minimal fluctuations thereafter. Additionally, it is worth noting that the constructed data by RDC is unsuitable for Noise2Noise and Noise2Void. This is mainly because the noise in constructed images by RDC is correlated, which would quickly collapse the training processes of the other two methods." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17930" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 113, + 296, + 159 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 4. Discussion of distributions. Different norms are suitable for different noise distributions [22]. For simplicity, we have uniformly adopted the " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{2\\to 1.5}" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " norm." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 113, + 296, + 159 + ], + "lines": [ + { + "bbox": [ + 56, + 113, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 113, + 296, + 159 + ], + "type": "table", + "html": "
Dataset||·||0||·||1||·||2||·||2→1.5
PolyU [49]37.8838.0238.2038.20
SIDD [1]35.1235.2635.2035.24
", + "image_path": "2936e11417bd1b2ee4f56634df75f54ad813229a1460fb7d7f7e20de18fad3d9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 171, + 175, + 239 + ], + "blocks": [ + { + "bbox": [ + 57, + 171, + 175, + 239 + ], + "lines": [ + { + "bbox": [ + 57, + 171, + 175, + 239 + ], + "spans": [ + { + "bbox": [ + 57, + 171, + 175, + 239 + ], + "type": "image", + "image_path": "fee973ec3e3634645272cf323da138f5bf041cb7f4e5b2000e11c5bb1133e3e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 296, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 293 + ], + "type": "text", + "content": "Figure 8. Convergence comparisons. P2N converges quickly and remains stable after convergence on SIDD. The P2N-Fixed version (where " + }, + { + "bbox": [ + 55, + 247, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\sigma = 0.00" + }, + { + "bbox": [ + 55, + 247, + 296, + 293 + ], + "type": "text", + "content": "), in particular, even achieves rapid convergence in approximately 10 iterations. P2N represents Positive2Negative." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 178, + 171, + 294, + 239 + ], + "blocks": [ + { + "bbox": [ + 178, + 171, + 294, + 239 + ], + "lines": [ + { + "bbox": [ + 178, + 171, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 178, + 171, + 294, + 239 + ], + "type": "image", + "image_path": "a52c55de9b69ce102160154975397945498e416d8127e52e2b4c8069d26260d5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 313, + 126, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 126, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 126, + 326 + ], + "type": "text", + "content": "5.3. Efficiency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 331, + 296, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 404 + ], + "type": "text", + "content": "For self-supervised single image denoising methods, computational efficiency is crucial, as each noisy image must be trained individually. Here, we demonstrate the number of training iterations, parameter, and computational cost, as detailed in Table 5. It is obvious that our paradigm achieves significant efficiency improvements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 414, + 294, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 294, + 427 + ], + "type": "text", + "content": "5.4. Comparison with Noise2Noise and Noise2Void" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 432, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 468 + ], + "type": "text", + "content": "Compared with Noise2Noise [22] and Noise2Void [18], Positive2Negative also presents an elegant paradigm The key differences are present in the following aspects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 469, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 613 + ], + "type": "text", + "content": "Firstly, for the constructed data, the synthetic data constructed by Positive2Negative breaks the information-lossy barrier of Noise2Noise and Noise2Void. Secondly, for the supervision strategy, Positive2Negative enforces consistency between denoised results, whereas Noise2Noise and Noise2Void supervise the consistency between a denoised result and a noisy image. Thirdly, from the assumption perspective, Noise2Noise, Noise2Void, and Positive2Negative all have theoretical guarantees, while Noise2Noise and Noise2Void operate under more restrictive assumptions. Noise2Noise requires two independent noisy images, while Noise2Void assumes that the noise is spatially correlated." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 623, + 247, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 247, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 247, + 635 + ], + "type": "text", + "content": "5.5. Differences with R2R and CVF-SID" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "For the constructed data, R2R employs new Gaussian noise and applies reciprocal coefficients to the noisy image, while CVF-SID employs two networks to predict three components and applies fixed manual rules. In contrast, our method employs a single network to predict noise and applies random multiscale symmetrical parameters." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 134, + 553, + 258 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "text", + "content": "Table 5. Efficiency comparisons under input images with dimensions of " + }, + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 3" + }, + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "text", + "content": ". With similar parameters and FLOPs, P2N requires the fewest training iterations. P2N converges rapidly, and the P2N-Fixed version (where " + }, + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "inline_equation", + "content": "\\sigma = 0.00" + }, + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "text", + "content": ") converges even just within 20 iterations. P2N represents Positive2Negative." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 134, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 317, + 134, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 134, + 553, + 258 + ], + "type": "table", + "html": "
MethodIterationsParams (M)FLOPs (G)
N2N [22]-0.7018.7
B2U [40]-1.1018.7
Self2Self [36]4500001.009.6
R2R [33]80000.5636.7
ScoreDVI [9]40013.537.9
MASH [11]800 × 30.9911.4
P2N-Fixed<200.9911.4
P2N<1000.9911.4
", + "image_path": "f9b10a56752869f8379e809419f1c3ac0eb7f031cce9a329b3f40605ab823c11.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "text", + "content": "For the supervision strategy, R2R maps the denoised image to the noisy input, while CVF-SID primarily maps the first cyclic result to the noisy input and maps the second cyclic result to the first cyclic result. In contrast, our method maps the denoised image to another denoised image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 354, + 471, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 471, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 471, + 366 + ], + "type": "text", + "content": "5.6. Limitation and Future Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 373, + 556, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 517 + ], + "type": "text", + "content": "The primary limitation is that Positive2Negative has only been validated on the single image based self-supervised image denoising task. To prevent convergence to trivial solutions such as zero mapping or identity mapping, Positive2Negative requires training with a pre-trained Gaussian denoising model. While the single image based self-supervised denoising task allows training from a pre-trained model on Gaussian noise, the noisy dataset based self-supervised image denoising task trains from scratch. Therefore, extending the Positive2Negative paradigm to the noisy dataset based self-supervised image denoising task could be a promising direction for future work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 533, + 388, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 388, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 388, + 546 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 555, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 628 + ], + "type": "text", + "content": "In this paper, we propose a self-supervised single image denoising paradigm, Positive2Negative, to break information-lossy barrier. Specifically, Positive2Negative includes two steps: RDC and DCS. Positive2Negative achieves SOTA performance compared to existing methods across various public benchmarks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 643, + 411, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 411, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 411, + 657 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China under Grants 62322204, 62131003, and the Key Laboratory of Target Cognition and Application Technology." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17931" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Abdelrahman Abdelhamed, Stephen Lin, and Michael S. Brown. A high-quality denoising dataset for smartphone cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1700, 2018. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "type": "text", + "content": "[2] Joshua Batson and Loic Royer. Noise2Self: Blind denoising by self-supervision. In International Conference on Machine Learning (ICML), pages 524-533, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 180, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 295, + 223 + ], + "type": "text", + "content": "[3] Chinmay Belthangady and Loic A Royer. Applications, promises, and pitfalls of deep learning for fluorescence image reconstruction. In Nature Methods, pages 1215-1225, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 224, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 295, + 280 + ], + "type": "text", + "content": "[4] Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing images for learned raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11036-11045, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 281, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 295, + 334 + ], + "type": "text", + "content": "[5] Jaeseok Byun, Sungmin Cha, and Taesup Moon. Fbidenoiser: Fast blind image denoiser for poisson-gaussian noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5768-5777, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 335, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 335, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 61, + 335, + 295, + 401 + ], + "type": "text", + "content": "[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1692-1703, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 435 + ], + "type": "text", + "content": "[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision (ECCV), pages 17-33, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 501 + ], + "type": "text", + "content": "[8] Shiyan Chen, Jiyuan Zhang, Zhaofei Yu, and Tiejun Huang. Exploring efficient asymmetric blind-spots for self-supervised denoising in real-world scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2814-2823, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "type": "text", + "content": "[9] Jun Cheng, Tao Liu, and Shan Tan. Score priors guided deep variational inference for unsupervised real-world single image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12937-12948, 2023. 1, 3, 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 602 + ], + "type": "text", + "content": "[10] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25974-25984, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 657 + ], + "type": "text", + "content": "[11] Hamadi Chihaoui and Paolo Favaro. Masked and shuffled blind spot denoising for real-world images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3025-3034, 2024. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[12] Hansen Feng, Lizhi Wang, Yuzhi Wang, Haoqiang Fan, and Hua Huang. Learnability enhancement for low-light raw image denoising: A data perspective. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 370-387, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[13] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. In ACM Transactions on Graphics (ToG), pages 1-12, 2016. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 185 + ], + "type": "text", + "content": "[14] Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. Neighbor2Neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14781-14790, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "text", + "content": "[15] Yeong Il Jang, Keuntek Lee, Gu Yong Park, Seyun Kim, and Nam Ik Cho. Self-supervised image denoising with downsampled invariance loss and conditional blind-spot network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12196-12205, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[16] Kwanyoung Kim and Jong Chul Ye. Noise2score: tweedie's approach to self-supervised image denoising without clean images. In Advances in Neural Information Processing Systems, pages 864-874, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 342 + ], + "type": "text", + "content": "[17] Kwanyoung Kim, Taesung Kwon, and Jong Chul Ye. Noise distribution adaptive self-supervised image denoising using tweedie distribution and score matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1998-2006, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "text", + "content": "[18] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void: Learning denoising from single noisy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2129-2137, 2019. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[19] Alexander Krull, Tomáš Vičar, Mangal Prakash, Manan Lalit, and Florian Jug. Probabilistic noise2void: Unsupervised content-aware denoising. In Frontiers in Computer Science, pages 2-5, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 444, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 488 + ], + "type": "text", + "content": "[20] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. In Proceedings of International Conference on Neural Information Processing Systems (NeurIPS), pages 6970-6980, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 489, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 555, + 544 + ], + "type": "text", + "content": "[21] Wooseok Lee, Sanghyun Son, and Kyoung Mu Lee. APBSN: Self-supervised denoising for real-world images via asymmetric pd and blind-spot network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17725-17734, 2022. 2, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "type": "text", + "content": "[22] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2Noise: Learning image restoration without clean data. In International Conference on Machine Learning (ICML), pages 4620-4631, 2018. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 657 + ], + "type": "text", + "content": "[23] Junyi Li, Zhilu Zhang, Xiaoyu Liu, Chaoyu Feng, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Spatially adaptive self-supervised learning for real-world image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914-9924, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[24] Tong Li, Hansen Feng, Lizhi Wang, Zhiwei Xiong, and Hua Huang. Stimulating the diffusion model for image denoising via adaptive embedding and ensembling. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 8240-8257, 2024. 1, 3, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17932" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 715 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 161 + ], + "type": "text", + "content": "[25] Xinyang Li, Yixin Li, Yiliang Zhou, Jiamin Wu, Zhifeng Zhao, Jiaqi Fan, Fei Deng, Zhaofa Wu, Guihua Xiao, Jing He, Yuanlong Zhang, Guoxun Zhang, Xiaowan Hu, Zhang Yi Chen, Xingye and, Hui Qiao, Hao Xie, Yulong Li, Haoqian Wang, Lu Fang, and Qionghai Dai. Real-time denoising enables high-sensitivity fluorescence time-lapse imaging beyond the shot-noise limit. In Nature Biotechnology, page 282-292, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "text", + "content": "[26] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. In arXiv preprint arXiv:1711.05101, pages 1-10, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[27] Youssef Mansour and Reinhard Heckel. Zero-Shot Noise2Noise: Efficient image denoising without any data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14018-14027, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "text", + "content": "[28] Nick Moran, Dan Schmidt, Yu Zhong, and Patrick Coady. Noisier2Noise: Learning to denoise from unpaired noisy data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12064-12072, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 374 + ], + "type": "text", + "content": "[29] Seonghyeon Nam, Youngbae Hwang, Yasuyuki Matsushita, and Seon Joo Kim. A holistic approach to cross-channel image noise modeling and its application to image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1683-1691, 2016. 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 441 + ], + "type": "text", + "content": "[30] Reyhaneh Neshatavar, Mohsen Yavartanoo, Sanghyun Son, and Kyoung Mu Lee. CVF-SID: Cyclic multi-variate function for self-supervised image denoising by disentangling noise from image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17583-17591, 2022. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 443, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 476 + ], + "type": "text", + "content": "[31] Harry Nyquist. Certain topics in telegraph transmission theory. In Transactions of the American Institute of Electrical Engineers, pages 617-644, 1928. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "text", + "content": "[32] Yizhong Pan, Xiao Liu, Xiangyu Liao, Yuanzhouhan Cao, and Chao Ren. Random sub-samples generation for self-supervised real image denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12150-12159, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "type": "text", + "content": "[33] Tongyao Pang, Huan Zheng, Yuhui Quan, and Hui Ji. Recorrupted-to-Recorrupted: Unsupervised deep learning for image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2043–2052, 2021. 2, 3, 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "text", + "content": "[34] Mikhail Papkov, Pavel Chizhov, and Leopold Parts. Swinia: Self-supervised blind-spot image denoising without convolutions. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 7071-7080, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[35] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1586-1595, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "text", + "content": "[36] Yuhui Quan, Mingqin Chen, Tongyao Pang, and Hui Ji. Self2Self with dropout: Learning self-supervised denoising" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "from single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1890-1898, 2020. 3, 5, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[37] Claude Elwood Shannon. A mathematical theory of communication. In The Bell System Technical Journal, pages 379-423, 1948. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[38] Andrea Vedaldi Ulyanov Dmitry and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9446-9454, 2018. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[39] Yuzhi Wang, Haibin Huang, Qin Xu, Jiaming Liu, Yiqun Liu, and Jue Wang. Practical deep raw image denoising on mobile devices. In European Conference on Computer Vision (ECCV), pages 1-16, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "text", + "content": "[40] Zejin Wang, Jiazheng Liu, Guoqing Li, and Hua Han. Blind2Unblind: Self-supervised image denoising with visible blind spots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2027-2036, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "type": "text", + "content": "[41] Zichun Wang, Ying Fu, Ji Liu, and Yulun Zhang. LG-BPN: Local and global blind-patch network for self-supervised real-world denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18156-18165, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 344, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 388 + ], + "type": "text", + "content": "[42] Kaixuan Wei, Ying Fu, Yinqiang Zheng, and Jiaolong Yang. Physics-based noise modeling for extreme low-light photography. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 8520-8537, 2021. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 453 + ], + "type": "text", + "content": "[43] Martin Weigert, Uwe Schmidt, Tobias Boothe, Andreas Müller, Alexandr Dibrov, Akanksha Jain, Benjamin Wilhelm, Deborah Schmidt, Coleman Broaddus, Sián Culley, et al. Content-aware image restoration: pushing the limits of fluorescence microscopy. In Nature Methods, pages 1090–1097, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "text", + "content": "[44] Halbert White. Connectionist nonparametric regression: Multilayer feedforward networks can learn arbitrary mappings. In Neural networks, pages 535-549, 1990. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "text", + "content": "[45] David Wu, Damian Tan, Marilyn Baird, John DeCampo, Chris White, and Hong Ren Wu. Perceptually lossless medical image coding. In IEEE Transactions on Medical Imaging, pages 335-344, 2006. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[46] Hong Ren Wu, Weisi Lin, and Lina J. Karam. An overview of perceptual processing for digital pictures. In IEEE International Conference on Multimedia and Expo Workshops, pages 113-120, 2012. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 580, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 623 + ], + "type": "text", + "content": "[47] Xiaohe Wu, Ming Liu, Yue Cao, Dongwei Ren, and Wangmeng Zuo. Unpaired learning of deep image denoising. In European Conference on Computer Vision (ECCV), pages 352-368, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "text", + "content": "[48] Yaochen Xie, Zhengyang Wang, and Shuiwang Ji. Noise2same: Optimizing a self-supervised bound for image denoising. In Advances in Neural Information Processing Systems, pages 20320-20330, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[49] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. In arXiv preprint arXiv:1804.02603, pages 1-13, 2018. 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17933" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 596 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[50] Jun Xu, Yuan Huang, Ming-Ming Cheng, Li Liu, Fan Zhu, Zhou Xu, and Ling Shao. Noisy-as-Clean: Learning self-supervised denoising from corrupted image. In IEEE Transactions on Image Processing (TIP), pages 9316-9329, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "text", + "content": "[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728-5739, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 294, + 251 + ], + "type": "text", + "content": "[52] Dan Zhang, Fangfang Zhou, Yuwen Jiang, and Zhengming Fu. Mm-bsn: Self-supervised image denoising for real-world with multi-mask based on blind-spot network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 4189–4198, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 317 + ], + "type": "text", + "content": "[53] Yide Zhang, Yinhao Zhu, Evan Nichols, Qingfei Wang, Siyuan Zhang, Cody Smith, and Scott Howard. Aoisson-gaussian denoising dataset with real fluorescence microscopy images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11710-11718, 2019. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 319, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 373 + ], + "type": "text", + "content": "[54] Yi Zhang, Dasong Li, Ka Lung Law, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. IDR: Self-supervised image denoising via iterative data refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2098-2107, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 374, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 374, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 374, + 294, + 430 + ], + "type": "text", + "content": "[55] Dihan Zheng, Sia Huat Tan, Xiaowen Zhang, Zuoqiang Shi, Kaisheng Ma, and Chenglong Bao. An unsupervised deep learning approach for real-world image denoising. In International Conference on Learning Representations (ICLR), pages 1-9, 2020. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 430, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 294, + 483 + ], + "type": "text", + "content": "[56] Dihan Zheng, Xiaowen Zhang, Kaisheng Ma, and Chenglong Bao. Learn from unpaired data for image restoration: A variational bayes approach. In IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 5889-5903, 2022. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 486, + 294, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 294, + 540 + ], + "type": "text", + "content": "[57] Yuqian Zhou, Jianbo Jiao, Haibin Huang, Yang Wang, Jue Wang, Honghui Shi, and Thomas Huang. When awgn-based denoiser meets real noises. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 13074-13081, 2020. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 541, + 294, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 541, + 294, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 541, + 294, + 596 + ], + "type": "text", + "content": "[58] Yunhao Zou, Chenggang Yan, and Ying Fu. Iterative denoiser and noise estimator for self-supervised image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13265-13274, 2023. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17934" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/1196e091-a513-40f9-85ef-51604d8681b7_content_list.json b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/1196e091-a513-40f9-85ef-51604d8681b7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7126d67365c3d2ab41bed699c8f3e6db4244edb8 --- /dev/null +++ b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/1196e091-a513-40f9-85ef-51604d8681b7_content_list.json @@ -0,0 +1,1717 @@ +[ + { + "type": "text", + "text": "Post-pre-training for Modality Alignment in Vision-Language Foundation Models", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shin'ya Yamaguchi* NTT, Kyoto University", + "bbox": [ + 102, + 180, + 290, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dewei Feng \nMIT", + "bbox": [ + 323, + 181, + 423, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sekitoshi Kanai NTT", + "bbox": [ + 457, + 181, + 586, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kazuki Adachi NTT", + "bbox": [ + 622, + 181, + 743, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daiki Chijiwa NTT", + "bbox": [ + 779, + 181, + 893, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive language image pre-training (CLIP) is an essential component of building modern vision-language foundation models. While CLIP demonstrates remarkable zero-shot performance on downstream tasks, the multi-modal feature spaces still suffer from a modality gap, which is a gap between image and text feature clusters and limits downstream task performance. Although existing works attempt to address the modality gap by modifying pre-training or fine-tuning, they struggle with heavy training costs with large datasets or degradations of zero-shot performance. This paper presents CLIP-Refine, a post-pre-training method for CLIP models at a phase between pre-training and fine-tuning. CLIP-Refine aims to align the feature space with 1 epoch training on small image-text datasets without zero-shot performance degradations. To this end, we introduce two techniques: random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). RaFA aligns the image and text features to follow a shared prior distribution by minimizing the distance to random reference vectors sampled from the prior. HyCD updates the model with hybrid soft labels generated by combining ground-truth image-text pair labels and outputs from the pre-trained CLIP model. This contributes to achieving both maintaining the past knowledge and learning new knowledge to align features. Our extensive experiments with multiple classification and retrieval tasks show that CLIP-Refine succeeds in mitigating the modality gap and improving the zero-shot performance1.", + "bbox": [ + 86, + 300, + 485, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 734, + 220, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive language image pre-training (CLIP, [21, 40]) is a standard method to build modern vision-language foundation models. CLIP enables models to learn multimodal representations to map images and texts into a shared feature space with contrastive loss on large-scale image-text pair datasets. Since the pre-trained CLIP models provide a cross-modal understanding of input image/text data in various domains,", + "bbox": [ + 89, + 761, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d73534414c3d8d7094939027bbfc8aae79a0a8f314d80db69639ba5512a8d68.jpg", + "image_caption": [ + "Figure 1. Scope of our work: post-pre-training to align the modality gap in pre-trained CLIP models. We aim to address the modality gap and enhance the generalization performance of pre-trained CLIP models through lightweight training." + ], + "image_footnote": [], + "bbox": [ + 522, + 270, + 903, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "they are widely used as a foundation of many applications, including zero-shot classification [14, 40, 55], cross-modal retrieval [16, 20], text-to-image generation [41, 44], and visual question answering [30, 31].", + "bbox": [ + 511, + 410, + 906, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While CLIP achieves remarkable performance in broad domains and tasks, its image and text alignment is still not perfect. For example, CLIP models tend to encode images and texts into different clusters for each modality, and thus, there is a modality gap between images and text features even after sufficiently training with the massive datasets [28, 39, 66]. This modality gap suggests that CLIP has difficulty in precisely mapping images and text. In fact, Liang et al. [28] have shown that the modality gap largely affects downstream task performance, especially in fine-grained classification tasks, and Ray et al. [43] have demonstrated that CLIP models often fail to retrieve localized objects and attributes in images from corresponding captions.", + "bbox": [ + 509, + 474, + 908, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the modality gap in CLIP, existing literature mainly focuses on refining the CLIP feature spaces in the pre-training or fine-tuning phases. For pre-training, one approach is multi-task learning of the contrastive objective and auxiliary losses such as self-supervised loss with data augmentations [26, 27]. Another direction for pre-training is to reduce the modality gap by modifying the encoder architectures to explicitly share the weights or feature maps between the image and text encoders [6, 64]. For fine-tuning, previous works have demonstrated that the target downstream task performance can be improved significantly by prompt tuning that optimizes trainable visual/textual tokens on target tasks to match image and text in the feature space (i.e., reducing the modality gap) [23, 33, 68]. More recently, Yang et al. [62] have introduced adapter parameters to share the infor", + "bbox": [ + 511, + 674, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. shinya.yamaguchi@ntt.com \n1Code: https://github.com/yshinya6/clip-refine", + "bbox": [ + 107, + 875, + 460, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4256", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mation across modalities. This cross-modal shared adapter is fine-tuned to solve the target tasks.", + "bbox": [ + 93, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although these pre-training/fine-tuning methods succeed in improving the cross-modal alignment capability, they face difficulties in terms of computational cost and zero-shot transfer performance. On the one hand, pre-training methods require training from scratch with million-scale image-text datasets like LAION-400M [46]. Since the previous works often do not publish pre-trained models using such large datasets, naive CLIP models published by OpenAI [40] or OpenCLIP [8] are still a practical choice under the limited budget. On the other hand, fine-tuning methods do not require huge computation costs because their datasets are much smaller than pre-training ones. Nevertheless, fine-tuning pretrained CLIP models degrades the general zero-shot transfer performance due to focusing on target tasks [25]. This is inevitable as the primary goal of fine-tuning is to improve the target task performance. Therefore, we seek a new approach that is lighter than pre-training and avoids degrading zero-shot transfer performance in fine-tuning.", + "bbox": [ + 93, + 125, + 482, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we tackle aligning the image and text features of CLIP models in a training phase called post-pretraining between pre-training and fine-tuning (Fig. 1). The goal of post-pre-training is to mitigate the modality gap and enhance zero-shot transfer performance of off-the-shelf pretrained CLIP models by only using reasonable computation resources and datasets (e.g., a single GPU and COCO Captions [29]). This is challenging because naive alignment methods, such as directly minimizing the gap between image and text features, corrupt the feature spaces in terms of the uniformity on the hypersphere [53]; the uniformity is an important property of contrastive learning representation that indicates the amount of input data information. Furthermore, post-pre-training with contrastive loss causes catastrophic forgetting of the general knowledge in pre-trained CLIP models due to the overfitting by the restricted mini-batch sizes.", + "bbox": [ + 93, + 398, + 482, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we introduce a method for post-pre-training called CLIP-Refine (Figure 2), which is composed of random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). Instead of directly minimizing the gap between features, our idea is to refine the image and text feature distributions to follow a shared prior (e.g., standard Gaussian). To achieve this, RaFA minimizes the gap between the image/text feature vectors and the reference vectors, which are randomly generated from the prior. In RaFA, the reference vectors are shared for the pairs of input image and text. Thus, RaFA penalizes image and text features to explicitly follow the same distribution. By matching the image and text feature distributions, the features are expected to avoid excessive concentration and preserve uniformity among samples, resulting in a good balance of multi-modal alignment and uniformity in the feature space. Conversely, HyCD is designed to avoid catastrophic forgetting by modified self-distillation", + "bbox": [ + 91, + 643, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "loss with the supervision of image-text pairs. Specifically, we apply the knowledge distillation where the teacher is the pre-trained model, i.e., minimizing KL-divergence between the outputs of the teacher and (student) post-pre-training model. Furthermore, we encourage learning new knowledge by blending the teacher's similarity matrix and the identity matrix, which represents the supervision of matched image-text pairs. Combining RaFA and HyCD, the CLIP models can reduce the modality gap while maintaining the uniformity of features without forgetting past knowledge.", + "bbox": [ + 517, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our extensive experiments on zero-shot classification (12 datasets) and cross-modal retrieval (2 datasets) demonstrate that our method significantly improves zero-shot performance over post-pre-training baselines with contrastive loss. Through the quantitative and qualitative analysis of feature spaces, we found that CLIP-Refine not only reduces the modality gap but also improves the uniformity of the hypersphere. This suggests that post-pre-training facilitates effective cross-modal alignment and well-clustered features.", + "bbox": [ + 517, + 242, + 903, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 516, + 392, + 653, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CLIP and Modality Gap. CLIP [21, 40] is a multi-modal representation learning technique that trains a model to embed inputs from image and text modalities into shared feature space. This is done by making the correct pairs of image and text closer in the feature space while the other pairs in mini-batch repelled via the InfoNCE-based contrastive loss [37]. While this simple multi-modal pre-training has remarkably advanced multi-modal research in various fields [4, 5, 7, 51, 52, 54, 58], Liang et al. [28] have revealed that pre-trained CLIP models encode image and text into different clusters for each modality, i.e., modality gap, and the small temperature parameter in the CLIP's contrastive loss can cause the modality gap empirically. The successor work by Qian et al. [39] has theoretically shown that contrastive loss can not reduce the modality gap perfectly.", + "bbox": [ + 517, + 416, + 903, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pre-training Modifications. To mitigate the modality gap and improve the cross-modal alignment, several works have modified the contrastive loss of CLIP by adding auxiliary losses, including geometric cyclic consistency between image and text features [15], the fine-grained similarity between output tokens of both modalities [63], self-supervised loss with data augmentation [26, 27], supervised contrastive learning [61], textual augmented contrastive loss using LLM-generated positive/negative texts [10]. Another direction is to explicitly share the information of image and text modalities by introducing shared encoder weights [64] or shared feature spaces of finite discrete tokens [6] to the pre-training. These methods, however, require large-scale image-text pair datasets (e.g., CC12M [3] and LAION-400M [46]) to achieve practical performance, incurring high computational costs on multiple GPUs. In contrast, our post-pre-training method improves the generalization performance of off-the", + "bbox": [ + 517, + 643, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4257", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c37fd29f21c5043056c83372df2d3daf26f072ac24d98476e4033eef5e809084.jpg", + "image_caption": [ + "Random Feature Alignment" + ], + "image_footnote": [], + "bbox": [ + 156, + 116, + 472, + 266 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1d936d63ae7949b54b4104ce354ec93c3e115be645ac5ca6b846ef9544848e72.jpg", + "image_caption": [ + "Hybrid Contrastive-Distillation", + "Figure 2. Intuition of CLIP-Refine (proposed method). CLIP-Refine modifies a pre-trained CLIP model by Random Feature Alignment (RaFA, left) and Hybrid Contrastive-Distillation (HyCD, right). RaFA penalizes image and text features to follow a shared prior distribution by minimizing the gap between the features and a random reference vector sampled from the prior. HyCD trains the model with the KL-divergence-based distillation loss using hybrid soft labels composed by alpha blending the ground truth label of image-text pairs with the output of the pre-trained model (teacher)." + ], + "image_footnote": [], + "bbox": [ + 545, + 114, + 823, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "shelf pre-trained models by only using small datasets like Flickr8/30K [42] and COCO Caption [29] with a single GPU. Besides, our method is cooperative with these pre-training methods because it can be used for any pre-trained model.", + "bbox": [ + 91, + 359, + 483, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning Methods2. The original paper of CLIP [40] has reported lower zero-shot performance in several fine-grained classification tasks, such as Aircraft [34], demonstrating the imperfectness of the cross-modal alignment of CLIP. Motivated by this, Zhou et al. [68, 69] have proposed to optimize text prompts to mitigate the modality gap by updating additional trainable vectors while the pre-trained models are fixed. This simple approach helps match image and text in the feature space (i.e., reducing the modality gap) and significantly improves target performance. Similarly, The method of Jia et al. [22] learns the additional trainable parameters for image input (visual prompt), and the successor works [23, 48] have unified the visual and textual prompt tuning by the multi-task learning objectives. Another approach is to directly share the low-rank adaptation parameters across the modalities and update them in fine-tuning [62]. In the perspective of the refinement of CLIP feature spaces, Oh et al. [36] have theoretically and empirically shown that fine-tuned CLIP models retain the suboptimal feature space in terms of alignment and uniformity, which are important properties to assess the feature quality in contrastive learning [53]. To refine the feature space, they have provided a mixup-based fine-tuning method, which generates hard negative samples on the hypersphere in a cross-modal manner ( $m^2$ -mix). These fine-tuning techniques succeed in reducing the gap in target datasets, but they do not intend to improve the zero-shot transfer performance. Our post-pre-training method differs from the fine-tuning methods in the objective: our method aims to improve the zero-shot transfer performance by reducing the modality gap. Obviously, we can", + "bbox": [ + 91, + 421, + 483, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 CLIP-Refine for Post-pre-training", + "text_level": 1, + "bbox": [ + 514, + 359, + 826, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Require: Training dataset $\\mathcal{D}$ , vision encoder $f_{\\mathrm{V}}$ , text encoder $f_{\\mathrm{T}}$ , pre-trained parameters $\\{\\theta_{\\mathrm{V}}^{\\mathrm{P}}, \\theta_{\\mathrm{T}}^{\\mathrm{P}}\\}$ , training batchsize $B$ , step size $\\eta$ , temperature parameter $\\tau$ , hyper-parameters $\\alpha$", + "bbox": [ + 514, + 378, + 906, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ensure: Post-pre-trained parameters $\\{\\theta_{\\mathrm{V}},\\theta_{\\mathrm{T}}\\}$", + "bbox": [ + 514, + 421, + 795, + 435 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: # Initialize parameters", + "2: $\\theta_{\\mathrm{V}} \\gets \\theta_{\\mathrm{V}}^{\\mathrm{P}}$", + "3: $\\theta_{\\mathrm{T}} \\gets \\theta_{\\mathrm{T}}^{\\mathrm{P}}$", + "4: while not converged do", + "5: $\\{(x^i,t^i)\\}_{i = 1}^B\\sim \\mathcal{D}$", + "6: $\\{z_{\\mathrm{img}}^i,z_{\\mathrm{txt}}^i\\}_{i = 1}^B\\gets \\{f_{\\mathrm{V}}(x^i;\\theta_{\\mathrm{V}}),f_{\\mathrm{T}}(t^i;\\theta_{\\mathrm{T}})\\}_{i = 1}^B$", + "7: $\\{z_{\\mathrm{img}}^{i,\\mathrm{p}}, z_{\\mathrm{xt}}^{i,\\mathrm{p}}\\}_{i=1}^{B} \\gets \\{f_{\\mathrm{V}}(x^i; \\theta_{\\mathrm{V}}^{\\mathrm{p}}), f_{\\mathrm{T}}(t^i; \\theta_{\\mathrm{T}}^{\\mathrm{p}})\\}_{i=1}^{B}$", + "8: # Compute random feature alignment loss", + "9: $\\{z_{\\mathrm{ref}}^{\\mathrm{i}}|z_{\\mathrm{ref}}^{\\mathrm{i}}\\sim p(z)\\}_{i = 1}^{B}$", + "10: $\\mathcal{L}_{\\mathrm{RaFA}}\\gets \\frac{1}{2B}\\sum_{i = 1}^{B}\\| z_{\\mathrm{img}}^i -z_{\\mathrm{ref}}^i\\| _2^2 +\\| z_{\\mathrm{txt}}^i -z_{\\mathrm{ref}}^i\\| _2^2$", + "11: # Compute hybrid contrastive-distillation loss", + "12: Compute student outputs $p_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}}$ and $p_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}$ for all sample combinations in a batch by Eq. (6) with $\\tau$ .", + "13: Compute teacher outputs $\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}}$ and $\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}$ for all sample combinations in a batch by Eq. (8) with $\\tau$ and $\\alpha$ .", + "14: $\\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{I}\\rightarrow \\mathrm{T}}\\gets \\frac{1}{B}\\sum_{i = 1}^{B}\\sum_{j = 1}^{B}\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}}\\log \\frac{\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow\\mathrm{T}}}{p_{i,j}^{\\mathrm{I}\\rightarrow\\mathrm{T}}}$", + "15: $\\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{T}\\rightarrow \\mathrm{I}}\\gets \\frac{1}{B}\\sum_{i = 1}^{B}\\sum_{j = 1}^{B}\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}\\log \\frac{\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}}{p_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}}$", + "16: $\\mathcal{L}_{\\mathrm{HyCD}}\\gets \\frac{1}{2} (\\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{I}\\to \\mathrm{T}} + \\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{T}\\to \\mathrm{I}})$", + "17: $\\theta_{\\mathrm{V}}\\gets \\theta_{\\mathrm{V}} - \\frac{\\eta}{2}\\nabla_{\\theta_{\\mathrm{V}}}\\left(\\mathcal{L}_{\\mathrm{RaFA}} + \\mathcal{L}_{\\mathrm{HyCD}}\\right)$", + "18: $\\theta_{\\mathrm{T}}\\gets \\theta_{\\mathrm{T}} - \\frac{\\eta}{2}\\nabla_{\\theta_{\\mathrm{T}}}(\\mathcal{L}_{\\mathrm{RaFA}} + \\mathcal{L}_{\\mathrm{HyCD}})$", + "19: end while" + ], + "bbox": [ + 519, + 436, + 903, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "leverage any fine-tuning techniques after post-pre-training by our method, and it helps to boost the baseline performance.", + "bbox": [ + 511, + 781, + 906, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 828, + 604, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose post-pre-training of CLIP in between pretraining and fine-tuning to mitigate the modality gap (Figure 1). To this end, we present CLIP-Refine (Figure 2),", + "bbox": [ + 511, + 854, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2We refer to training models on specific target tasks as fine-tuning.", + "bbox": [ + 107, + 886, + 460, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4258", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which is composed of random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). RaFA penalizes the model to induce both modal features into a single shared distribution by minimizing the gap between the paired features and a random reference vector. To reconcile learning new knowledge with retaining past knowledge, HyCD updates the models with knowledge distillation loss using fixed pretrained models, where the teacher outputs are mixed with ground-truth labels of image-text pairs. Algorithm 1 shows the overall procedure of CLIP-Refine.", + "bbox": [ + 89, + 90, + 483, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Problem Setting: Post-pre-training", + "text_level": 1, + "bbox": [ + 89, + 251, + 397, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider a problem setting called post-pre-training, where we aim to improve the cross-modal alignment and generalization performance of pre-trained vision-language models. This setting allows to access a vision encoder $f_{\\mathrm{V}}: \\mathcal{X} \\to \\mathbb{R}^{d}$ and a text encoder $f_{\\mathrm{T}}: \\mathcal{T} \\to \\mathbb{R}^{d}$ with the parameter $\\theta^{\\mathrm{CLIP}} = \\{\\theta_{\\mathrm{V}}^{\\mathrm{CLIP}}, \\theta_{\\mathrm{T}}^{\\mathrm{CLIP}}\\}$ pre-trained by CLIP, where $\\mathcal{X}$ and $\\mathcal{T}$ are the image and text space. We optimize the parameters $\\theta_{\\mathrm{V}}$ of $f_{\\mathrm{V}}$ and $\\theta_{\\mathrm{T}}$ of $f_{\\mathrm{T}}$ on a post-pre-training image-text pair dataset $\\mathcal{D} = \\{(x_i, t_i)\\}_{i=1}^N$ ; we basically assume that $\\mathcal{D}$ is much smaller than the pre-training dataset and contains images and text captions in general domain.", + "bbox": [ + 89, + 273, + 483, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Objective Function", + "text_level": 1, + "bbox": [ + 89, + 448, + 274, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In CLIP-Refine, we optimize $\\theta_{\\mathrm{V}}$ and $\\theta_{\\mathrm{T}}$ by the following objective function:", + "bbox": [ + 89, + 469, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {\\mathrm {V}}, \\theta_ {\\mathrm {T}}} \\mathcal {L} _ {\\mathrm {R a F A}} \\left(\\theta_ {\\mathrm {V}}, \\theta_ {\\mathrm {T}}\\right) + \\mathcal {L} _ {\\mathrm {H y C D}} \\left(\\theta_ {\\mathrm {V}}, \\theta_ {\\mathrm {T}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 510, + 483, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{RaFA}}$ is the random feature alignment loss and $\\mathcal{L}_{\\mathrm{HyCD}}$ is the hybrid contrastive-distillation loss. One can consider balancing $\\mathcal{L}_{\\mathrm{RaFA}}$ and $\\mathcal{L}_{\\mathrm{HyCD}}$ by hyperparameters, but we found that their equal contribution achieves the best performance (see Appendix). We describe the details of $\\mathcal{L}_{\\mathrm{RaFA}}$ and $\\mathcal{L}_{\\mathrm{HyCD}}$ in the following sections.", + "bbox": [ + 89, + 542, + 483, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Random Feature Alignment", + "text_level": 1, + "bbox": [ + 89, + 643, + 344, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The primary goal in post-pre-training is to minimize the modality gap and encourage cross-modal alignment. To this end, the most straightforward approach is just minimizing the feature gap between the modalities. Let $z_{\\mathrm{img}}^i = f_{\\mathrm{V}}(x^i;\\theta_{\\mathrm{V}})$ and $z_{\\mathrm{xt}}^i = f_{\\mathrm{T}}(t^i;\\theta_{\\mathrm{T}})$ are the normalized feature vectors of an image-text pair $(x^{i},t^{i})$ . A naive loss for minimizing the modality gap can be defined by an $L_{2}$ distance form as", + "bbox": [ + 89, + 664, + 483, + 772 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\mathbb {E} _ {(x ^ {i}, t ^ {i}) \\in \\mathcal {D}} \\| z _ {\\text {i m g}} ^ {i} - z _ {\\text {t x t}} ^ {i} \\| _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 782, + 483, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, minimizing this loss function degrades the generalization performance. This can be explained by the balance of alignment and uniformity [53], which are desirable properties of transferability in contrastive representation learning. Alignment is defined by the gap between positive (i.e., imagetext) pairs in the feature space, and uniformity is defined by", + "bbox": [ + 89, + 810, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "equal separation among all features on the hypersphere. That is, minimizing Eq. (2) enhances the alignment of the positive pair but destroys the uniformity on the hypersphere by forcing changes to the feature distribution. In fact, Wang and Isola [53] have shown that using only the alignment loss in post-pre-training degrades the validation performance due to corrupt uniformity.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this challenge, we introduce the idea of matching feature distributions instead of directly matching the paired features. In other words, we refine the multi-modal features of pre-trained CLIP models to follow a prior distribution $p(z)$ (e.g., standard Gaussian $\\mathcal{N}(0, I)$ ) shared across the image and text modalities. Inspired by random feature regularization in single modal fine-tuning [59, 67], we formulate this idea as minimization between image/text feature vectors $(z_{\\mathrm{img}}^i, z_{\\mathrm{txt}}^i)$ and a random reference vector $z_{\\mathrm{ref}}^i$ sampled from the shared prior $p(z)$ :", + "bbox": [ + 511, + 198, + 908, + 348 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R a F A}} = \\mathbb {E} _ {(x ^ {i}, t ^ {i}) \\in \\mathcal {D}} \\frac {1}{2} (\\| z _ {\\mathrm {i m g}} ^ {i} - z _ {\\mathrm {r e f}} ^ {i} \\| _ {2} ^ {2} + \\| z _ {\\mathrm {t x t}} ^ {i} - z _ {\\mathrm {r e f}} ^ {i} \\| _ {2} ^ {2}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 359, + 906, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through this loss function, we can expect that (i) the modality gap is indirectly minimized via $z_{\\mathrm{ref}}^i$ shared with each image-text pair, (ii) the feature vectors follow the modality shared prior $p(z)$ gradually [59], (iii) the randomness of $z_{\\mathrm{ref}}^i$ helps to avoid the models from overfitting to $\\mathcal{D}$ [67] and reform the models' capacity [49]. Furthermore, we found that $\\mathcal{L}_{\\mathrm{RaFA}}$ can improve the uniformity from baselines (Table 3), meaning that the learned representations transfer well to downstream tasks. Throughout this paper, we use standard Gaussian $\\mathcal{N}(0,I)$ as $p(z)$ by default; we discuss the effect of prior choice in Sec. 4.4.1.", + "bbox": [ + 511, + 398, + 908, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Hybrid Contrastive-Distillation", + "text_level": 1, + "bbox": [ + 511, + 574, + 790, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another important goal of post-pre-training is to retain the past knowledge in pre-trained CLIP models. To this end, our basic strategy is to penalize the post-pre-training models with knowledge distillation loss [19] where the frozen pre-trained CLIP model performs as the teacher. We define the minibatch-wise knowledge distillation loss with KL-divergence according to the formulation of CLIP-KD [60] as", + "bbox": [ + 511, + 597, + 908, + 703 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K D}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {K D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} + \\mathcal {L} _ {\\mathrm {K D}} ^ {\\mathrm {T} \\rightarrow \\mathrm {I}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 713, + 906, + 732 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {B} q _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} \\log \\frac {q _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}{p _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 734, + 906, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {\\exp \\left(z _ {\\mathrm {i m g}} ^ {i} \\cdot z _ {\\mathrm {t x t}} ^ {j} / \\tau\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(z _ {\\mathrm {i m g}} ^ {i} \\cdot z _ {\\mathrm {t x t}} ^ {k} / \\tau\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 763, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {\\exp \\left(z _ {\\mathrm {i m g}} ^ {i , \\mathrm {p}} \\cdot z _ {\\mathrm {t x t}} ^ {j , \\mathrm {p}} / \\tau\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(z _ {\\mathrm {i m g}} ^ {i , \\mathrm {p}} \\cdot z _ {\\mathrm {t x t}} ^ {k , \\mathrm {p}} / \\tau\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 795, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature parameter, $(z_{\\mathrm{img}}^{i,\\mathrm{p}},z_{\\mathrm{xt}}^{i,\\mathrm{p}}) = (f_{\\mathrm{V}}(x^{i};\\theta_{\\mathrm{V}}^{\\mathrm{CLIP}}),f_{\\mathrm{T}}(t^{i};\\theta_{\\mathrm{T}}^{\\mathrm{CLIP}}))$ are the feature vectors generated by the pre-trained image/text encoders, and $\\mathcal{L}_{\\mathrm{KD}}^{\\mathrm{T}\\rightarrow \\mathrm{I}}$ is defined by swapping the order of image features and text", + "bbox": [ + 511, + 837, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4259", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/ae6722b5eda3466910d11b4522fc85326e5d2d72c5ba33a9181788de97c9d4cb.jpg", + "table_caption": [ + "Table 1. Zero-shot classification performance on 12 datasets with CLIP ViT-B/32." + ], + "table_footnote": [], + "table_body": "
MethodAirBirdCalCarDTDEuroFloFoodINPetSUNUCFAvg.
Pre-trained18.8149.3723.9860.0740.9038.2059.7580.2159.0481.7960.8559.9052.74
Contrastive13.9840.0723.7747.6336.2234.6846.2870.3152.9670.1354.7858.2645.75
m2-mix [36]14.6440.7623.4949.8636.4934.6547.4271.2753.5870.3256.5158.7646.48
Self-KD18.9651.5223.9460.0441.1738.3659.5580.0259.0681.8560.6860.1652.94
HyCD18.7250.5023.9859.8340.4840.9058.0680.1959.7181.4562.0461.7053.13
HyCD + L_align13.2934.2023.7343.7035.0039.2746.2773.3152.5772.4255.0158.5545.61
CLIP-Refine (Ours)20.7752.7224.7960.6341.4941.4462.6481.8760.9383.5763.6361.8554.69
", + "bbox": [ + 96, + 103, + 906, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "features in Eq. (6) and (7). Although minimizing Eq. (4) retains the past knowledge, we found that it interferes with RaFA because it strongly enchains the models' parameters to the pre-trained ones.", + "bbox": [ + 89, + 251, + 482, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reconcile RaFA and the retention of past knowledge, we present a technique that blends the ground-truth label of an image-text pair with the teacher output. Concretely, we modify Eq. (7) by alpha-blending of an indicator function $\\mathbb{I}_{i = j}$ , which returns one if $i = j$ otherwise zero, and the teacher output signal:", + "bbox": [ + 89, + 311, + 483, + 404 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {q} _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\alpha \\mathbb {I} _ {i = j} + (1 - \\alpha) q _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 409, + 483, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyperparameter to balance acquiring new knowledge and retaining past knowledge; we use $\\alpha = 0.5$ by default and discuss the effect of $\\alpha$ in the Appendix. With this modification, the post-pre-training models can further enhance the cross-modal alignment by learning the correct image-text pair (i.e., $i = j$ ) with reference to the relative knowledge contained in the $i \\neq j$ coordinates of the teacher's outputs. By using $\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}}$ and $\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}$ , we compute the HyCD loss function as follows:", + "bbox": [ + 89, + 436, + 483, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {H y C D}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {H y C D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} + \\mathcal {L} _ {\\mathrm {H y C D}} ^ {\\mathrm {T} \\rightarrow \\mathrm {I}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 580, + 482, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {H y C D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {B} \\hat {q} _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} \\log \\frac {\\hat {q} _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}{p _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 601, + 482, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While $\\mathcal{L}_{\\mathrm{HyCD}}$ solely improves the generalization performance of the pre-trained CLIP models, combining it with $\\mathcal{L}_{\\mathrm{RaFA}}$ achieves larger improvements by minimizing the modality gaps without losing the past knowledge.", + "bbox": [ + 89, + 636, + 483, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 709, + 223, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate CLIP-Refine on 12 classification datasets and 2 cross-modal retrieval datasets by using multiple pre-trained models. We also conduct qualitative and quantitative analyses to assess the feature space through modality gap, alignment, uniformity metrics, and PCA visualization.", + "bbox": [ + 89, + 734, + 485, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 89, + 818, + 189, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our CLIP-Refine with the following baselines of post-pre-training. Pre-trained: predicting with pre-trained weights without additional training. Contrastive: post-pre-training with the contrastive", + "bbox": [ + 89, + 839, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loss [21, 40]. $m^2$ -mix [36]: post-pre-training by the modified contrastive loss with multi-modal mixup; we use this fine-tuning method as a baseline because it can be used for refining the feature spaces as suggested in the original paper [36]. Self-KD: using Eq. (4) for post-pre-training. HyCD: using only Eq. (9) for training. HyCD + L align: combining Eq. (9) and Eq. (2).", + "bbox": [ + 511, + 251, + 906, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Post-pre-training Datasets. We used COCO Caption [29] as the default post-pre-training dataset. We also used Flickr8K/30K [42], CC3M [47], and CC12M [3] to confirm the dataset size effect in post-pre-training.", + "bbox": [ + 511, + 364, + 906, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test Datasets. We used 12 image classification datasets containing various image domains: Aircraft (Air) [34], Bird [56], Caltech-101 (Cal) [12] Car [24], DTD [9], EuroSAT (Euro) [18], Flower (Flo) [35], Food [2], ImageNet (IN) [45], Pet [38], SUN397 [57], and UCF-101 [50]. We also evaluated our method on COCO2017-Val and Flickr8K/30K for the image and text retrieval tasks; we denote text-to-image retrieval as $\\mathrm{T} \\rightarrow \\mathrm{I}$ and image-to-text retrieval as $\\mathrm{I} \\rightarrow \\mathrm{T}$ . We adopt these datasets since they are often used to evaluate the zero-shot transfer performance of CLIP models [68]. We evaluated models on the test sets except for ImageNet, and the official validation set for ImageNet.", + "bbox": [ + 511, + 431, + 908, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-trained Models. By default, we used CLIP-ViT-B/32 [40] with the pre-trained weights downloaded from the OpenAI's official repositories. We also tested other pretrained models, including OpenCLIP-ViT-H/14, OpenCLIP-ViT-bigG/14, SigLIP [65] and DFN [11] with the pre-trained weights on the OpenCLIP [8] repository.", + "bbox": [ + 511, + 619, + 908, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Post-pre-training. In all settings, we trained models by the AdamW [32] optimizer with the learning rate of $1.0 \\times 10^{-6}$ for one epoch; we determined the learning rate by the zero-shot classification validation using the subset of the ImageNet training set constructed by uniform sampling. We used mini-batch sizes of 512 as the default. For CLIP-Refine, we used $\\alpha = 0.5$ . The input samples were preprocessed using the default image transformations provided by each pre-trained model. We implemented the training and evaluation with PyTorch-1.13. We ran the experiments on a 24-core Intel Xeon CPU with a single NVIDIA A100 GPU with 80GB VRAM.", + "bbox": [ + 511, + 719, + 908, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4260", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/da4a9eb69a642e8e6d23e870f88f2a379da89442c803dad14c67fd17995f2c38.jpg", + "table_caption": [ + "Table 2. Zero-shot retrieval performance on COCO2017-Val and Flickr8K/30K with CLIP ViT-B/32." + ], + "table_footnote": [], + "table_body": "
MethodCOCO2017-ValFlickr8KFlickr30K
T→II→TR@1T→IR@1I→TR@1T→IR@1I→T
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@5R@10R@1R@5R@10R@1R@5R@10
Pre-trained30.5654.9265.2633.2659.1068.7830.0653.8363.2932.4255.5765.2621.0640.9650.5324.5244.5754.14
Contrastive34.8861.5072.1031.8656.8067.4834.9459.4169.3628.8351.5561.0725.6947.3256.9318.4236.2945.37
m2-mix [36]36.2862.1873.0832.9257.9667.5635.0459.6669.6628.6450.6560.7725.9947.1857.1919.8837.6346.51
Self-KD31.0455.5865.9035.3659.0069.7229.9853.8363.5532.4755.1165.1021.9241.5150.7524.9444.8254.27
HyCD36.0462.2873.1437.8862.5472.8435.7260.3470.1434.6957.1967.2227.1149.1258.7827.1747.6656.94
HyCD + L align33.9261.1872.0635.1461.7472.0631.7355.6365.9032.5856.9466.9923.7844.7354.4723.5845.0755.23
CLIP-Refine (Ours)37.6463.5474.4238.7865.0475.1236.1161.2970.5138.2861.5570.9426.6549.7059.0330.9951.5461.31
", + "bbox": [ + 93, + 103, + 906, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For zero-shot transfer performance, we report top-1 accuracy in classification tasks and recall@ $k$ scores in cross-modal retrieval tasks; recall@ $k$ indicates the proportion of correct answers in the retrieved top- $k$ candidates for all test input samples. We denote recall@ $k$ as $R@k$ . To quantitatively evaluate the post-pre-trained features, we measure modality gap, alignment, and uniformity scores. Modality gap [28] assesses a cluster-wise gap between image and text features:", + "bbox": [ + 89, + 237, + 485, + 375 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {M o d a l i t y} \\mathrm {G a p} := \\| \\bar {f} _ {\\mathrm {V}} - \\bar {f} _ {\\mathrm {T}} \\| _ {2} ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 383, + 482, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\bar{f}$ is a mean feature vector for all test samples of each modality; lower is better. Alignment [53] is defined by", + "bbox": [ + 89, + 412, + 483, + 445 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {A l i g n m e n t} := \\frac {1}{N _ {\\text {t e s t}}} \\sum_ {i = 1} ^ {N _ {\\text {t e s t}}} \\| f _ {\\mathrm {V}} \\left(x ^ {i}\\right) - f _ {\\mathrm {T}} \\left(t ^ {i}\\right) \\| _ {2} ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 455, + 482, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "representing how well the positive pair is aligned in the feature space; lower is better. Uniformity [53] evaluates how uniformly distributed the feature vectors are on the hypersphere, which is defined by radial basis function (RBF) kernel as", + "bbox": [ + 89, + 507, + 483, + 582 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {U n i f o r m i t y} := \\frac {1}{2 N _ {\\text {t e s t}}} \\sum_ {f _ {1}, f _ {2} \\in F} \\exp (- 2 \\| f _ {1} - f _ {2} \\| _ {2} ^ {2}), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 590, + 482, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F = \\{f_{\\mathrm{V}}(x^i)\\}_{i=1}^{N_{\\mathrm{test}}} \\cup \\{f_{\\mathrm{T}}(t^i)\\}_{i=1}^{N_{\\mathrm{test}}}$ ; lower is better. We report the average score for each metric by running experiments three times.", + "bbox": [ + 89, + 640, + 483, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Zero-shot Transfer Evaluation", + "text_level": 1, + "bbox": [ + 89, + 696, + 362, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first demonstrate the zero-shot performance of our CLIP-Refine. Tables 1 and 2 list the zero-shot test performance on classification and cross-modal retrieval tasks after post-pretraining. For zero-shot classification tasks, post-pre-training by the contrastive loss (i.e., Contrastive and $m^2$ -mix) largely degraded the top-1 accuracy performance for all datasets. This can be caused by catastrophic forgetting since Self-KD did not show performance degradation. One cause of catastrophic forgetting is a smaller batchsize than at pre-training; the OpenAI's CLIP uses a batchsize of 32,768 with hundreds of GPUs [40], whereas our setting uses ones of 512 with a single GPU. Since a smaller batch contains a limited number", + "bbox": [ + 89, + 719, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of negative samples, contrastive learning tends to overfit and catastrophically forget past knowledge. We discuss more details about the effect of batchsize in Sec. 4.4.4. In contrast, our CLIP-Refine impressively improved the zero-shot performance for all datasets. The ablation studies of HyCD show that the performance gain comes from the combination of HyCD with RaFA. Notably, $\\mathrm{HyCD} + \\mathcal{L}_{\\mathrm{align}}$ resulted in the worst performance among the baselines, suggesting that naively minimizing the gap between image and text features leads to the degradations of generalization performance.", + "bbox": [ + 511, + 237, + 906, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For zero-shot retrieval, our CLIP-Refine achieved the best performance with a larger margin than that in classification. Interestingly, in contrast to classification tasks, we found that Contrastive and $m^2$ -mix improved the pre-trained models in the $\\mathrm{T} \\rightarrow \\mathrm{I}$ retrieval tasks while the $\\mathrm{I} \\rightarrow \\mathrm{T}$ performance was degraded. This indicates that the image encoders are more prone to overfitting than the text encoders. This can be explained by the difference between image and text in the data spaces; images are represented by continuous-valued pixels and thus have a high degree of freedom, while texts are represented by discrete tokens in vocabulary and thus have a limited degree of freedom. Therefore, the image encoder tends to overfit to a limited number of negative samples in small batches, and the text encoder can learn new knowledge from them while keeping the past knowledge. Since the zero-shot classification is a $\\mathrm{I} \\rightarrow \\mathrm{T}$ retrieval task, this can also be a reason for the performance degradation in classification. In this perspective, our CLIP-Refine successfully prevents the overfitting in $\\mathrm{I} \\rightarrow \\mathrm{T}$ and greatly improves the performance in both $\\mathrm{I} \\rightarrow \\mathrm{T}$ and $\\mathrm{T} \\rightarrow \\mathrm{I}$ .", + "bbox": [ + 511, + 390, + 908, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Analysis of Feature Space", + "text_level": 1, + "bbox": [ + 511, + 709, + 746, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we aim to answer our fundamental research question: how much can we alleviate the modality gap by post-pre-training with CLIP-Refine? For quantitative evaluations, we summarize the modality gap, alignment, and uniformity scores in Table 3. HyCD $+\\mathcal{L}_{\\mathrm{align}}$ achieved the best modality gap and alignment scores. However, it significantly degraded the uniformity score, indicating that directly minimizing $\\mathcal{L}_{\\mathrm{align}}$ can destroy feature space well constructed by the CLIP training. In contrast, our CLIP-Refine consistently improved all of the scores. This suggests that it is important to minimize not only the gap between image and text features", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4261", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/175c129fbc52c2a0c63eed8aa993b0a054828eaf1477152aaa892f5bd87f894a.jpg", + "image_caption": [ + "(a) Pre-trained" + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 300, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00546a37b2fd7af3859b1c13b57b7b7d1f25749cfa11b66e0d0dadd2f512e183.jpg", + "image_caption": [ + "(b) Contrastive", + "Figure 3. PCA visualization of multi-modal feature spaces of trained models (Flickr8K, ViT-B/32). CLIP-Refine reduces the modality gap while keeping the form of feature distributions for each modality." + ], + "image_footnote": [], + "bbox": [ + 313, + 87, + 514, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ae6438814ad3ee5a2154ec0f2b995be71ee0d5f396709cf53a1af220f07b791.jpg", + "image_caption": [ + "(c) $\\mathrm{HyCD} + \\mathcal{L}_{\\mathrm{align}}$" + ], + "image_footnote": [], + "bbox": [ + 527, + 87, + 728, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0a2a744612d987be37aa739e461e7625d8b0f996e00a866734476cbca987f85.jpg", + "image_caption": [ + "(d) CLIP-Refine (Ours)" + ], + "image_footnote": [], + "bbox": [ + 741, + 87, + 942, + 200 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8497bdc23c48c62be6f6e251c50e1e044d7ba7db1c2b2213c9e8366e07212fa3.jpg", + "table_caption": [ + "Table 3. Quantitative evaluation of features (Flickr8K, ViT-B/32)." + ], + "table_footnote": [], + "table_body": "
MethodModality Gap ×10-3(↓)Alignment (↓)Uniformity (↓)
Pre-trained1.33451.37240.0895
Contrastive2.17141.43000.2136
m2-mix [36]2.09111.41300.2019
Self-KD1.27541.37740.0901
HyCD1.38231.35780.0971
HyCD+L align0.15900.74130.1172
CLIP-Refine (Ours)0.79341.28490.0495
", + "bbox": [ + 93, + 294, + 480, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "but also the uniformity for improving the generalization performance of pre-trained CLIP models, which is consistent with the findings in [53]. In such a sense, since CLIP-Refine penalizes the image and text features to follow the shared reference distribution $p(z) = \\mathcal{N}(0,I)$ , where the samples tend to be uniformly distributed on the hypersphere [1], it can naturally achieve a good balance of alignment and uniformity, and reasonably helps to reduce the modality gap.", + "bbox": [ + 88, + 404, + 483, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As a qualitative evaluation, Figure 3 illustrates the PCA visualization of post-pre-trained features for each method. We can see that the contrastive model makes the modality gap rather large. $\\mathrm{HyCD} + \\mathcal{L}_{\\mathrm{align}}$ succeeds in gathering the image and text features into the same cluster but distorts the feature distribution for each modality. Meanwhile, CLIP-Refine certainly reduces the modality gap while keeping the shape of the pre-trained features. Although the modality gap is not completely zero, it is natural because there remains a constant value in the lower bound of the modality gap [39].", + "bbox": [ + 88, + 526, + 483, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Detailed Analysis", + "text_level": 1, + "bbox": [ + 89, + 689, + 261, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1. Prior Distributions", + "text_level": 1, + "bbox": [ + 89, + 712, + 269, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we analyze CLIP-Refine by varying the prior distribution $p(z)$ . According to the prior works [59, 67], we tried standard Gaussian $\\mathcal{N}(0,I)$ , uniform distribution $U(0,1)$ , Gaussian distributions using pre-trained text feature statistics $\\mathcal{N}(\\mu_{\\mathrm{txt}}^{\\mathrm{P}},(\\sigma_{\\mathrm{txt}}^{\\mathrm{P}})^2)$ , image feature statistics $\\mathcal{N}(\\mu_{\\mathrm{img}}^{\\mathrm{P}},(\\sigma_{\\mathrm{img}}^{\\mathrm{P}})^2)$ , statistics of all image and text features $\\mathcal{N}(\\mu_{\\mathrm{all}}^{\\mathrm{P}},(\\sigma_{\\mathrm{all}}^{\\mathrm{P}})^2)$ , where $\\mu^{\\mathrm{P}},\\sigma^{\\mathrm{P}}$ is mean and variance of image or/and text feature vectors computed on a pre-trained CLIP model. We also tested variations of $\\mathcal{N}(0,\\beta I)$ by varying $\\beta$ in $\\{0,0.01,0.1,10,100\\}$ ; $\\beta = 0$ means RaFA without randomness, i.e., $\\frac{1}{2} (\\| z_{\\mathrm{img}}^{i}\\|_{2}^{2} + \\| z_{\\mathrm{txt}}^{i}\\|_{2}^{2})$ .", + "bbox": [ + 89, + 732, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7562a72399eecbc15fd57a1839dfb4939170c3c69899213bfa632d12432d3105.jpg", + "table_caption": [ + "Table 4. Evaluation of CLIP-Refine varying prior distributions in $\\mathcal{L}_{\\mathrm{RaFA}}$ . ZS Cls. is the zero-shot classification accuracy averaged on 12 datasets." + ], + "table_footnote": [], + "table_body": "
PriorZS Cls.AlignmentUniformity
Pretrained52.741.37240.0895
U(0,1)52.911.30810.0661
N(μpimg, (σpimg)2)52.931.31420.0916
N(μptxt, (σptxt)2)53.401.33040.0873
N(μpall, (σpall)2)53.291.33610.0888
N(0,0I)53.791.30520.0554
N(0,0.01I)54.191.30790.0544
N(0,0.1I)54.321.30210.0515
N(0,I)54.691.28490.0495
N(0,10I)54.231.30850.0537
N(0,100I)53.591.34410.0663
", + "bbox": [ + 535, + 321, + 880, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 shows the zero-shot classification performance, alignment, and uniformity scores when varying $p(z)$ . Using the uniform distribution $U(0,1)$ did not show large improvements on the zero-shot performance and uniformity score. This may be because random reference vectors from $U(0,1)$ do not necessarily concentrate on the hypersphere in contrast to $\\mathcal{N}(0,I)$ , possibly causing the collapse of uniformity by $\\mathcal{L}_{\\mathrm{RaFA}}$ . In this perspective, we recommend using $\\mathcal{N}(0,I)$ or families of Gaussian distributions. Interestingly, among the Gaussian distributions with the pre-trained statistics, the prior with the text feature, i.e., $\\mathcal{N}(\\mu_{\\mathrm{txt}}^{\\mathrm{p}}(\\sigma_{\\mathrm{txt}}^{\\mathrm{p}})^2)$ largely improved the zero-shot performance. This indicates that aligning image features with text features is effective in zero-shot classification tasks and also evidences the performance improvements in the $\\mathrm{I} \\rightarrow \\mathrm{T}$ tasks observed in Sec. 4.2. More importantly, the ablation studies of the variance magnitude $\\beta$ in the zero-mean Gaussian distributions emphasize the importance of randomness in $\\mathcal{L}_{\\mathrm{RaFA}}$ .", + "bbox": [ + 511, + 518, + 908, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2. Post-pre-training Datasets", + "text_level": 1, + "bbox": [ + 511, + 804, + 743, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the effects of the dataset choice in post-pretraining. Table 5 shows the zero-shot classification/retrieval performance of CLIP-Refine with Flickr8K/30K, COCO Caption, CC3M, and CC12M as post-pre-training datasets. We confirm that most of the CLIP-Refine models outper", + "bbox": [ + 511, + 824, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4262", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/45c18597a686633e8bf55b8e7293b010c40be81348217b2bd1d7d1ae4899b6b8.jpg", + "table_caption": [ + "Table 5. Zero-shot classification (12 datasets) and retrieval (Flicker30K) performance of CLIP-Refine with various post-pre-training datasets with ViT-B/32." + ], + "table_footnote": [], + "table_body": "
Post-pre-training Dataset# of SamplesZero-shot Cls. Acc.Zero-shot Ret. R@5
Avg. of 12 datasetsT→II→T
None (Pre-trained)052.7440.9644.57
Flickr8K8K53.0445.5746.80
Flickr30K30K53.7448.6649.25
COCO Caption118K54.6949.7051.54
CC3M3M53.8049.5650.24
CC12M12M52.4645.3547.79
Filtered CC3M1M55.5052.4553.00
Filtered CC12M4M60.1254.9757.88
", + "bbox": [ + 91, + 142, + 480, + 260 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b1ba50ac5d464808ded963c825f02f8e0c37d48f9e4fc939a23bd86abda53968.jpg", + "table_caption": [ + "Table 6. Zero-shot classification (12 datasets) and retrieval (Flickr30K) performance with various pre-trained CLIP models." + ], + "table_footnote": [], + "table_body": "
ModelZero-shot Cls. Acc.Zero-shot Ret. R@5 (T→I/I→T)
Pre-trainedContrastiveCLIP-RefinePre-trainedContrastiveCLIP-Refine
ViT-B/3252.7445.7654.6940.96/44.5747.32/36.2949.70/51.54
ViT-L/1462.9661.9063.8849.13/52.7751.07/51.0058.38/58.41
OpenCLIP ViT-H/1468.5257.6169.0565.84/66.0257.47/43.4367.82/68.48
OpenCLIP ViT-G/1471.2857.4472.5968.73/68.3866.35/42.3670.94/69.76
SigLip ViT-SO400M/1472.5765.6373.6172.77/72.8364.63/62.9776.54/76.49
DFN ViT-H/1477.3076.3978.2971.93/70.8972.81/70.0976.39/75.44
", + "bbox": [ + 91, + 313, + 480, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formed the pre-trained baselines. Notably, larger datasets do not always achieve high performance; CC3M and CC12M are inferior to COCO Caption. This suggests that the quality of image-text pairs in a dataset is an important factor in post-pre-training. Since CC3M and CC12M contain mismatched and noisy image-text pairs [17], COCO Caption that has higher-quality text captions can be superior in terms of the modality alignment of pre-trained CLIP models. In fact, when we filtered out low-quality image-text pairs by CLIP-Score-based filtering of DataComp [13], the models achieved significant performance improvements. This indicates that the caption quality is crucial and our method is scalable if we have high-quality image-text pairs.", + "bbox": [ + 88, + 392, + 485, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.3. Pre-trained Models", + "text_level": 1, + "bbox": [ + 89, + 597, + 272, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the generalizability of CLIP-Refine to pretrained CLIP models. Table 6 shows the zero-shot performance with 6 pre-trained models. We observed that our CLIP-Refine stably improved the pre-trained baseline. This suggests that the bigger and stronger models still have the modality gap, and modifying the gap by CLIP-Refine helps improve the generalization performance for various pretrained models.", + "bbox": [ + 88, + 614, + 485, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.4. Mini-Batch Sizes", + "text_level": 1, + "bbox": [ + 89, + 744, + 256, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide the analysis of varying mini-batch sizes in post-pre-training. We tried the mini-batch sizes in $\\{32, 64, 128, 256, 512, 1024, 2048\\}$ . We used gradient accumulation by following the OpenCLIP [8] implementation for the mini-batch sizes of 1024 and 2048. Figure 4 shows the zero-shot performance versus mini-batch sizes. The contrastive loss gradually improved the accuracy according to the mini-batch size. However, the accuracy curve implies that it is difficult to exceed the pre-trained baseline with", + "bbox": [ + 88, + 763, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/065f34356cbd634a4a0697963f5ad2f05b36a7e0a5d810dc0b7196fe5c3925bb.jpg", + "image_caption": [ + "Figure 4. Zero-shot classification accuracy averaged on 12 datasets when varying mini-batch sizes in post-pre-training." + ], + "image_footnote": [], + "bbox": [ + 571, + 87, + 851, + 204 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3df8148a21e082750f7ef6e5e2195b204f6969026675727d1bc1a0387e442e80.jpg", + "table_caption": [ + "Table 7. Fine-tuning top-1 accuracy on the ImageNet validation set." + ], + "table_footnote": [], + "table_body": "
ModelPre-trainedContrastiveCLIP-Refine
ViT-B/3273.4069.4674.14
", + "bbox": [ + 553, + 272, + 861, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "feasible mini-batch sizes. This is probably due to the large discrepancy from the mini-batch size in the pre-training (i.e., 32,768), which causes the model to forget knowledge in the predictions for negative pairs in a mini-batch. In contrast, CLIP-Refine always outperformed the contrastive loss and succeeded in improving the pre-trained baseline with feasible mini-batch sizes. This suggests that CLIP-Refine's HyCD prevents knowledge forgetting even for small mini-batch sizes with few negative pairs and is able to modify the feature space with new knowledge.", + "bbox": [ + 511, + 318, + 908, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.5. Effects on Fine-tuning", + "text_level": 1, + "bbox": [ + 511, + 477, + 718, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show the transferability of CLIP-Refine to fine-tuning. We fine-tuned the linear classification head on ImageNet while freezing post-pre-trained models. Tabel 7 shows the results. CLIP-Refine outperformed the pre-trained and contrastive loss baselines. This emphasizes that our method can generate useful representation even in fine-tuning.", + "bbox": [ + 511, + 496, + 908, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 599, + 633, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presented CLIP-Refine, a post-pre-training method for pre-trained CLIP models to align the modality gap between the image and text features. CLIP-Refine addresses the modality gap by penalizing the multi-modal features to follow a shared prior distribution by minimizing the distance to the random reference vectors sampled from the prior. To maintain the past knowledge in the CLIP models and promote the feature alignment simultaneously, CLIP-Refine also trains the model with knowledge distillation loss using hybrid soft labels composed of ground-truth image-text pair labels and outputs from the pre-trained CLIP model. Through extensive experiments, we show that CLIP-Refine can improve the zero-shot performance of the pre-trained CLIP by addressing the modality gap and enhancing uniformity. We believe that our work not only provides a practical method but also opens up a new research field where we refine the pre-trained CLIP models by post-pre-training with much smaller computational costs than that of pre-training.", + "bbox": [ + 511, + 626, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4263", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Christopher M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag, Berlin, Heidelberg, 2006. 7", + "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5", + "[3] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 2, 5", + "[4] Lianggangxu Chen, Xuejiao Wang, Jiale Lu, Shaohui Lin, Changbo Wang, and Gaoqi He. Clip-driven open-vocabulary 3d scene graph generation via cross-modality contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[5] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yu Hou, Yuenan and Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[6] Yuxiao Chen, Jianbo Yuan, Yu Tian, Shijie Geng, Xinyu Li, Ding Zhou, Dimitris N Metaxas, and Hongxia Yang. Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15095-15104, 2023. 1, 2", + "[7] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 2, 5, 8", + "[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2014. 5", + "[10] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2", + "[11] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. In International Conference on Learning Representations, 2024. 5" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In Conference on CVPR Workshop, 2004. 5", + "[13] Gadre et al. Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 2023. 8", + "[14] Yunhao Ge, Jie Ren, Andrew Gallagher, Yuxiao Wang, Ming-Hsuan Yang, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11093-11101, 2023. 1", + "[15] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. In Advances in Neural Information Processing Systems, 2022. 2", + "[16] Haochen Han, Qinghua Zheng, Guang Dai, Minnan Luo, and Jingdong Wang. Learning to rematch mismatched pairs for robust cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1", + "[17] Haochen Han, Qinghua Zheng, Guang Dai, Minnan Luo, and Jingdong Wang. Learning to rematch mismatched pairs for robust cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26679-26688, 2024. 8", + "[18] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5", + "[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 4", + "[20] Ahmet Iscen, Mathilde Caron, Alireza Fathi, and Cordelia Schmid. Retrieval-enhanced contrastive vision-text models. In International Conference on Learning Representations, 2024. 1", + "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, 2021. 1, 2, 5", + "[22] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision. Springer, 2022. 3", + "[23] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multimodal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 1, 3", + "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition, Sydney, Australia, 2013. 5" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4264", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Zhengfeng Lai, Noranart Vesdapunt, Ning Zhou, Jun Wu, Cong Phuoc Huynh, Xuelu Li, Kah Kuen Fu, and Chen-nee Chuah. Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2", + "[26] Janghyeon Lee, Jongsuk Kim, Hyounguk Shon, Bumsoo Kim, Seung Hwan Kim, Honglak Lee, and Junmo Kim. Uniclip: Unified framework for contrastive language-image pre-training. Advances in Neural Information Processing Systems, 2022. 1, 2", + "[27] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In International Conference on Learning Representations, 2021. 1, 2", + "[28] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 2022. 1, 2, 6", + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision. Springer, 2014. 2, 3, 5", + "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1", + "[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1", + "[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5", + "[33] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1", + "[34] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. arXiv, 2013. 3, 5", + "[35] M-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5", + "[36] Changdae Oh, Junhyuk So, Hoyoon Byun, YongTaek Lim, Minchul Shin, Jong-June Jeon, and Kyungwoo Song. Geodesic multi-modal mixup for robust fine-tuning. In Advances in Neural Information Processing Systems, 2023. 3, 5, 6, 7", + "[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2", + "[38] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 5" + ], + "bbox": [ + 91, + 92, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In Advances in Neural Information Processing Systems, 2024. 1, 2, 7", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 1, 2, 3, 5, 6", + "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[42] Cyrus Rashtchian, Peter Young, Micah Hodosh, and Julia Hockenmaier. Collecting image annotations using amazon's mechanical turk. In Proceedings of the NAACL HLT 2010 workshop on creating speech and language data with Amazon's Mechanical Turk, pages 139-147, 2010. 3, 5", + "[43] Arijit Ray, Filip Radenovic, Abhimanyu Dubey, Bryan Plummer, Ranjay Krishna, and Kate Saenko. Cola: A benchmark for compositional text-to-image retrieval. In Advances in Neural Information Processing Systems, 2023. 1", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1", + "[45] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3), 2015. 5", + "[46] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2", + "[47] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, 2018. 5", + "[48] Sheng Shen, Shijia Yang, Tianjun Zhang, Bohan Zhai, Joseph E Gonzalez, Kurt Keutzer, and Trevor Darrell. Multitask vision-language prompt tuning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 3", + "[49] Sicong Shen, Yang Zhou, Bingzheng Wei, Eric I Chang, Yan Xu, et al. Tuning stable rank shrinkage: Aiming at the overlooked structural risk in fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4", + "[50] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5", + "[51] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4265", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3835-3844, 2022. 2", + "[52] Jingyun Wang and Guoliang Kang. Learn to rectify the bias of clip for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4102-4112, 2024. 2", + "[53] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 2, 3, 4, 6, 7", + "[54] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2022. 2", + "[55] Zhengbo Wang, Jian Liang, Ran He, Nan Xu, Zilei Wang, and Tieniu Tan. Improving zero-shot generalization for clip with synthesized prompts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1", + "[56] P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD Birds 200. Technical report, California Institute of Technology, 2010. 5", + "[57] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5", + "[58] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2021. 2", + "[59] Shin'ya Yamaguchi, Sekitoshi Kanai, Kazuki Adachi, and Daiki Chijiwa. Adaptive random feature regularization on fine-tuning deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4, 7", + "[60] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. Clipkd: An empirical study of clip model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15952-15962, 2024. 4", + "[61] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19163-19173, 2022. 2", + "[62] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 1, 3", + "[63] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and" + ], + "bbox": [ + 91, + 90, + 483, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chunjing Xu. FILIP: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022. 2", + "[64] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. In European Conference on Computer Vision. Springer, 2022. 1, 2", + "[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 5", + "[66] Yuhui Zhang, Jeff Z. HaoChen, Shih-Cheng Huang, Kuan-Chieh Wang, James Zou, and Serena Yeung. Diagnosing and rectifying vision models using language. In The Eleventh International Conference on Learning Representations, 2023. 1", + "[67] Yang Zhong and Atsuto Maki. Regularizing cnn transfer learning with randomised regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13637-13646, 2020. 4, 7", + "[68] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 1, 3, 5", + "[69] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 3" + ], + "bbox": [ + 516, + 90, + 906, + 488 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4266", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/1196e091-a513-40f9-85ef-51604d8681b7_model.json b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/1196e091-a513-40f9-85ef-51604d8681b7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..81a0e2a0e52e93a7558af877b253116c93af5aeb --- /dev/null +++ b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/1196e091-a513-40f9-85ef-51604d8681b7_model.json @@ -0,0 +1,2708 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.153 + ], + "angle": 0, + "content": "Post-pre-training for Modality Alignment in Vision-Language Foundation Models" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.181, + 0.291, + 0.218 + ], + "angle": 0, + "content": "Shin'ya Yamaguchi* NTT, Kyoto University" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.182, + 0.424, + 0.216 + ], + "angle": 0, + "content": "Dewei Feng \nMIT" + }, + { + "type": "text", + "bbox": [ + 0.459, + 0.182, + 0.588, + 0.216 + ], + "angle": 0, + "content": "Sekitoshi Kanai NTT" + }, + { + "type": "text", + "bbox": [ + 0.624, + 0.182, + 0.744, + 0.216 + ], + "angle": 0, + "content": "Kazuki Adachi NTT" + }, + { + "type": "text", + "bbox": [ + 0.78, + 0.182, + 0.895, + 0.216 + ], + "angle": 0, + "content": "Daiki Chijiwa NTT" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.301, + 0.486, + 0.709 + ], + "angle": 0, + "content": "Contrastive language image pre-training (CLIP) is an essential component of building modern vision-language foundation models. While CLIP demonstrates remarkable zero-shot performance on downstream tasks, the multi-modal feature spaces still suffer from a modality gap, which is a gap between image and text feature clusters and limits downstream task performance. Although existing works attempt to address the modality gap by modifying pre-training or fine-tuning, they struggle with heavy training costs with large datasets or degradations of zero-shot performance. This paper presents CLIP-Refine, a post-pre-training method for CLIP models at a phase between pre-training and fine-tuning. CLIP-Refine aims to align the feature space with 1 epoch training on small image-text datasets without zero-shot performance degradations. To this end, we introduce two techniques: random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). RaFA aligns the image and text features to follow a shared prior distribution by minimizing the distance to random reference vectors sampled from the prior. HyCD updates the model with hybrid soft labels generated by combining ground-truth image-text pair labels and outputs from the pre-trained CLIP model. This contributes to achieving both maintaining the past knowledge and learning new knowledge to align features. Our extensive experiments with multiple classification and retrieval tasks show that CLIP-Refine succeeds in mitigating the modality gap and improving the zero-shot performance1." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.736, + 0.222, + 0.751 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.485, + 0.868 + ], + "angle": 0, + "content": "Contrastive language image pre-training (CLIP, [21, 40]) is a standard method to build modern vision-language foundation models. CLIP enables models to learn multimodal representations to map images and texts into a shared feature space with contrastive loss on large-scale image-text pair datasets. Since the pre-trained CLIP models provide a cross-modal understanding of input image/text data in various domains," + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.271, + 0.905, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.339, + 0.908, + 0.396 + ], + "angle": 0, + "content": "Figure 1. Scope of our work: post-pre-training to align the modality gap in pre-trained CLIP models. We aim to address the modality gap and enhance the generalization performance of pre-trained CLIP models through lightweight training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.472 + ], + "angle": 0, + "content": "they are widely used as a foundation of many applications, including zero-shot classification [14, 40, 55], cross-modal retrieval [16, 20], text-to-image generation [41, 44], and visual question answering [30, 31]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.475, + 0.909, + 0.672 + ], + "angle": 0, + "content": "While CLIP achieves remarkable performance in broad domains and tasks, its image and text alignment is still not perfect. For example, CLIP models tend to encode images and texts into different clusters for each modality, and thus, there is a modality gap between images and text features even after sufficiently training with the massive datasets [28, 39, 66]. This modality gap suggests that CLIP has difficulty in precisely mapping images and text. In fact, Liang et al. [28] have shown that the modality gap largely affects downstream task performance, especially in fine-grained classification tasks, and Ray et al. [43] have demonstrated that CLIP models often fail to retrieve localized objects and attributes in images from corresponding captions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.902 + ], + "angle": 0, + "content": "To address the modality gap in CLIP, existing literature mainly focuses on refining the CLIP feature spaces in the pre-training or fine-tuning phases. For pre-training, one approach is multi-task learning of the contrastive objective and auxiliary losses such as self-supervised loss with data augmentations [26, 27]. Another direction for pre-training is to reduce the modality gap by modifying the encoder architectures to explicitly share the weights or feature maps between the image and text encoders [6, 64]. For fine-tuning, previous works have demonstrated that the target downstream task performance can be improved significantly by prompt tuning that optimizes trainable visual/textual tokens on target tasks to match image and text in the feature space (i.e., reducing the modality gap) [23, 33, 68]. More recently, Yang et al. [62] have introduced adapter parameters to share the infor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.876, + 0.461, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author. shinya.yamaguchi@ntt.com \n1Code: https://github.com/yshinya6/clip-refine" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4256" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "mation across modalities. This cross-modal shared adapter is fine-tuned to solve the target tasks." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.125, + 0.483, + 0.396 + ], + "angle": 0, + "content": "Although these pre-training/fine-tuning methods succeed in improving the cross-modal alignment capability, they face difficulties in terms of computational cost and zero-shot transfer performance. On the one hand, pre-training methods require training from scratch with million-scale image-text datasets like LAION-400M [46]. Since the previous works often do not publish pre-trained models using such large datasets, naive CLIP models published by OpenAI [40] or OpenCLIP [8] are still a practical choice under the limited budget. On the other hand, fine-tuning methods do not require huge computation costs because their datasets are much smaller than pre-training ones. Nevertheless, fine-tuning pretrained CLIP models degrades the general zero-shot transfer performance due to focusing on target tasks [25]. This is inevitable as the primary goal of fine-tuning is to improve the target task performance. Therefore, we seek a new approach that is lighter than pre-training and avoids degrading zero-shot transfer performance in fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.4, + 0.483, + 0.641 + ], + "angle": 0, + "content": "In this paper, we tackle aligning the image and text features of CLIP models in a training phase called post-pretraining between pre-training and fine-tuning (Fig. 1). The goal of post-pre-training is to mitigate the modality gap and enhance zero-shot transfer performance of off-the-shelf pretrained CLIP models by only using reasonable computation resources and datasets (e.g., a single GPU and COCO Captions [29]). This is challenging because naive alignment methods, such as directly minimizing the gap between image and text features, corrupt the feature spaces in terms of the uniformity on the hypersphere [53]; the uniformity is an important property of contrastive learning representation that indicates the amount of input data information. Furthermore, post-pre-training with contrastive loss causes catastrophic forgetting of the general knowledge in pre-trained CLIP models due to the overfitting by the restricted mini-batch sizes." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.645, + 0.483, + 0.9 + ], + "angle": 0, + "content": "To this end, we introduce a method for post-pre-training called CLIP-Refine (Figure 2), which is composed of random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). Instead of directly minimizing the gap between features, our idea is to refine the image and text feature distributions to follow a shared prior (e.g., standard Gaussian). To achieve this, RaFA minimizes the gap between the image/text feature vectors and the reference vectors, which are randomly generated from the prior. In RaFA, the reference vectors are shared for the pairs of input image and text. Thus, RaFA penalizes image and text features to explicitly follow the same distribution. By matching the image and text feature distributions, the features are expected to avoid excessive concentration and preserve uniformity among samples, resulting in a good balance of multi-modal alignment and uniformity in the feature space. Conversely, HyCD is designed to avoid catastrophic forgetting by modified self-distillation" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "loss with the supervision of image-text pairs. Specifically, we apply the knowledge distillation where the teacher is the pre-trained model, i.e., minimizing KL-divergence between the outputs of the teacher and (student) post-pre-training model. Furthermore, we encourage learning new knowledge by blending the teacher's similarity matrix and the identity matrix, which represents the supervision of matched image-text pairs. Combining RaFA and HyCD, the CLIP models can reduce the modality gap while maintaining the uniformity of features without forgetting past knowledge." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.243, + 0.905, + 0.378 + ], + "angle": 0, + "content": "Our extensive experiments on zero-shot classification (12 datasets) and cross-modal retrieval (2 datasets) demonstrate that our method significantly improves zero-shot performance over post-pre-training baselines with contrastive loss. Through the quantitative and qualitative analysis of feature spaces, we found that CLIP-Refine not only reduces the modality gap but also improves the uniformity of the hypersphere. This suggests that post-pre-training facilitates effective cross-modal alignment and well-clustered features." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.393, + 0.654, + 0.408 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.417, + 0.905, + 0.644 + ], + "angle": 0, + "content": "CLIP and Modality Gap. CLIP [21, 40] is a multi-modal representation learning technique that trains a model to embed inputs from image and text modalities into shared feature space. This is done by making the correct pairs of image and text closer in the feature space while the other pairs in mini-batch repelled via the InfoNCE-based contrastive loss [37]. While this simple multi-modal pre-training has remarkably advanced multi-modal research in various fields [4, 5, 7, 51, 52, 54, 58], Liang et al. [28] have revealed that pre-trained CLIP models encode image and text into different clusters for each modality, i.e., modality gap, and the small temperature parameter in the CLIP's contrastive loss can cause the modality gap empirically. The successor work by Qian et al. [39] has theoretically shown that contrastive loss can not reduce the modality gap perfectly." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.645, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Pre-training Modifications. To mitigate the modality gap and improve the cross-modal alignment, several works have modified the contrastive loss of CLIP by adding auxiliary losses, including geometric cyclic consistency between image and text features [15], the fine-grained similarity between output tokens of both modalities [63], self-supervised loss with data augmentation [26, 27], supervised contrastive learning [61], textual augmented contrastive loss using LLM-generated positive/negative texts [10]. Another direction is to explicitly share the information of image and text modalities by introducing shared encoder weights [64] or shared feature spaces of finite discrete tokens [6] to the pre-training. These methods, however, require large-scale image-text pair datasets (e.g., CC12M [3] and LAION-400M [46]) to achieve practical performance, incurring high computational costs on multiple GPUs. In contrast, our post-pre-training method improves the generalization performance of off-the" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4257" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.095, + 0.419, + 0.11 + ], + "angle": 0, + "content": "Random Feature Alignment" + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.117, + 0.473, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.095, + 0.792, + 0.11 + ], + "angle": 0, + "content": "Hybrid Contrastive-Distillation" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.116, + 0.825, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.282, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Figure 2. Intuition of CLIP-Refine (proposed method). CLIP-Refine modifies a pre-trained CLIP model by Random Feature Alignment (RaFA, left) and Hybrid Contrastive-Distillation (HyCD, right). RaFA penalizes image and text features to follow a shared prior distribution by minimizing the gap between the features and a random reference vector sampled from the prior. HyCD trains the model with the KL-divergence-based distillation loss using hybrid soft labels composed by alpha blending the ground truth label of image-text pairs with the output of the pre-trained model (teacher)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.361, + 0.485, + 0.422 + ], + "angle": 0, + "content": "shelf pre-trained models by only using small datasets like Flickr8/30K [42] and COCO Caption [29] with a single GPU. Besides, our method is cooperative with these pre-training methods because it can be used for any pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.422, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Fine-tuning Methods2. The original paper of CLIP [40] has reported lower zero-shot performance in several fine-grained classification tasks, such as Aircraft [34], demonstrating the imperfectness of the cross-modal alignment of CLIP. Motivated by this, Zhou et al. [68, 69] have proposed to optimize text prompts to mitigate the modality gap by updating additional trainable vectors while the pre-trained models are fixed. This simple approach helps match image and text in the feature space (i.e., reducing the modality gap) and significantly improves target performance. Similarly, The method of Jia et al. [22] learns the additional trainable parameters for image input (visual prompt), and the successor works [23, 48] have unified the visual and textual prompt tuning by the multi-task learning objectives. Another approach is to directly share the low-rank adaptation parameters across the modalities and update them in fine-tuning [62]. In the perspective of the refinement of CLIP feature spaces, Oh et al. [36] have theoretically and empirically shown that fine-tuned CLIP models retain the suboptimal feature space in terms of alignment and uniformity, which are important properties to assess the feature quality in contrastive learning [53]. To refine the feature space, they have provided a mixup-based fine-tuning method, which generates hard negative samples on the hypersphere in a cross-modal manner (\\(m^2\\)-mix). These fine-tuning techniques succeed in reducing the gap in target datasets, but they do not intend to improve the zero-shot transfer performance. Our post-pre-training method differs from the fine-tuning methods in the objective: our method aims to improve the zero-shot transfer performance by reducing the modality gap. Obviously, we can" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.361, + 0.828, + 0.376 + ], + "angle": 0, + "content": "Algorithm 1 CLIP-Refine for Post-pre-training" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.38, + 0.907, + 0.422 + ], + "angle": 0, + "content": "Require: Training dataset \\(\\mathcal{D}\\), vision encoder \\(f_{\\mathrm{V}}\\), text encoder \\(f_{\\mathrm{T}}\\), pre-trained parameters \\(\\{\\theta_{\\mathrm{V}}^{\\mathrm{P}}, \\theta_{\\mathrm{T}}^{\\mathrm{P}}\\}\\), training batchsize \\(B\\), step size \\(\\eta\\), temperature parameter \\(\\tau\\), hyper-parameters \\(\\alpha\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.422, + 0.797, + 0.436 + ], + "angle": 0, + "content": "Ensure: Post-pre-trained parameters \\(\\{\\theta_{\\mathrm{V}},\\theta_{\\mathrm{T}}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.437, + 0.678, + 0.449 + ], + "angle": 0, + "content": "1: # Initialize parameters" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.45, + 0.602, + 0.463 + ], + "angle": 0, + "content": "2: \\(\\theta_{\\mathrm{V}} \\gets \\theta_{\\mathrm{V}}^{\\mathrm{P}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.464, + 0.602, + 0.477 + ], + "angle": 0, + "content": "3: \\(\\theta_{\\mathrm{T}} \\gets \\theta_{\\mathrm{T}}^{\\mathrm{P}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.478, + 0.686, + 0.49 + ], + "angle": 0, + "content": "4: while not converged do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.491, + 0.674, + 0.505 + ], + "angle": 0, + "content": "5: \\(\\{(x^i,t^i)\\}_{i = 1}^B\\sim \\mathcal{D}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.505, + 0.847, + 0.52 + ], + "angle": 0, + "content": "6: \\(\\{z_{\\mathrm{img}}^i,z_{\\mathrm{txt}}^i\\}_{i = 1}^B\\gets \\{f_{\\mathrm{V}}(x^i;\\theta_{\\mathrm{V}}),f_{\\mathrm{T}}(t^i;\\theta_{\\mathrm{T}})\\}_{i = 1}^B\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.52, + 0.847, + 0.535 + ], + "angle": 0, + "content": "7: \\(\\{z_{\\mathrm{img}}^{i,\\mathrm{p}}, z_{\\mathrm{xt}}^{i,\\mathrm{p}}\\}_{i=1}^{B} \\gets \\{f_{\\mathrm{V}}(x^i; \\theta_{\\mathrm{V}}^{\\mathrm{p}}), f_{\\mathrm{T}}(t^i; \\theta_{\\mathrm{T}}^{\\mathrm{p}})\\}_{i=1}^{B}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.535, + 0.808, + 0.548 + ], + "angle": 0, + "content": "8: # Compute random feature alignment loss" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.548, + 0.692, + 0.563 + ], + "angle": 0, + "content": "9: \\(\\{z_{\\mathrm{ref}}^{\\mathrm{i}}|z_{\\mathrm{ref}}^{\\mathrm{i}}\\sim p(z)\\}_{i = 1}^{B}\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.562, + 0.882, + 0.577 + ], + "angle": 0, + "content": "10: \\(\\mathcal{L}_{\\mathrm{RaFA}}\\gets \\frac{1}{2B}\\sum_{i = 1}^{B}\\| z_{\\mathrm{img}}^i -z_{\\mathrm{ref}}^i\\| _2^2 +\\| z_{\\mathrm{txt}}^i -z_{\\mathrm{ref}}^i\\| _2^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.577, + 0.831, + 0.589 + ], + "angle": 0, + "content": "11: # Compute hybrid contrastive-distillation loss" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.589, + 0.905, + 0.617 + ], + "angle": 0, + "content": "12: Compute student outputs \\( p_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}} \\) and \\( p_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}} \\) for all sample combinations in a batch by Eq. (6) with \\( \\tau \\)." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.617, + 0.905, + 0.645 + ], + "angle": 0, + "content": "13: Compute teacher outputs \\(\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}}\\) and \\(\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}\\) for all sample combinations in a batch by Eq. (8) with \\(\\tau\\) and \\(\\alpha\\)." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.645, + 0.815, + 0.669 + ], + "angle": 0, + "content": "14: \\(\\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{I}\\rightarrow \\mathrm{T}}\\gets \\frac{1}{B}\\sum_{i = 1}^{B}\\sum_{j = 1}^{B}\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}}\\log \\frac{\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow\\mathrm{T}}}{p_{i,j}^{\\mathrm{I}\\rightarrow\\mathrm{T}}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.669, + 0.815, + 0.695 + ], + "angle": 0, + "content": "15: \\(\\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{T}\\rightarrow \\mathrm{I}}\\gets \\frac{1}{B}\\sum_{i = 1}^{B}\\sum_{j = 1}^{B}\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}\\log \\frac{\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}}{p_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.694, + 0.756, + 0.709 + ], + "angle": 0, + "content": "16: \\(\\mathcal{L}_{\\mathrm{HyCD}}\\gets \\frac{1}{2} (\\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{I}\\to \\mathrm{T}} + \\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{T}\\to \\mathrm{I}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.709, + 0.791, + 0.723 + ], + "angle": 0, + "content": "17: \\(\\theta_{\\mathrm{V}}\\gets \\theta_{\\mathrm{V}} - \\frac{\\eta}{2}\\nabla_{\\theta_{\\mathrm{V}}}\\left(\\mathcal{L}_{\\mathrm{RaFA}} + \\mathcal{L}_{\\mathrm{HyCD}}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.723, + 0.789, + 0.738 + ], + "angle": 0, + "content": "18: \\(\\theta_{\\mathrm{T}}\\gets \\theta_{\\mathrm{T}} - \\frac{\\eta}{2}\\nabla_{\\theta_{\\mathrm{T}}}(\\mathcal{L}_{\\mathrm{RaFA}} + \\mathcal{L}_{\\mathrm{HyCD}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.738, + 0.607, + 0.75 + ], + "angle": 0, + "content": "19: end while" + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.437, + 0.905, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.782, + 0.907, + 0.813 + ], + "angle": 0, + "content": "leverage any fine-tuning techniques after post-pre-training by our method, and it helps to boost the baseline performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.829, + 0.605, + 0.844 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We propose post-pre-training of CLIP in between pretraining and fine-tuning to mitigate the modality gap (Figure 1). To this end, we present CLIP-Refine (Figure 2)," + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.461, + 0.901 + ], + "angle": 0, + "content": "2We refer to training models on specific target tasks as fine-tuning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4258" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.243 + ], + "angle": 0, + "content": "which is composed of random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). RaFA penalizes the model to induce both modal features into a single shared distribution by minimizing the gap between the paired features and a random reference vector. To reconcile learning new knowledge with retaining past knowledge, HyCD updates the models with knowledge distillation loss using fixed pretrained models, where the teacher outputs are mixed with ground-truth labels of image-text pairs. Algorithm 1 shows the overall procedure of CLIP-Refine." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.252, + 0.398, + 0.269 + ], + "angle": 0, + "content": "3.1. Problem Setting: Post-pre-training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.485, + 0.44 + ], + "angle": 0, + "content": "We consider a problem setting called post-pre-training, where we aim to improve the cross-modal alignment and generalization performance of pre-trained vision-language models. This setting allows to access a vision encoder \\( f_{\\mathrm{V}}: \\mathcal{X} \\to \\mathbb{R}^{d} \\) and a text encoder \\( f_{\\mathrm{T}}: \\mathcal{T} \\to \\mathbb{R}^{d} \\) with the parameter \\( \\theta^{\\mathrm{CLIP}} = \\{\\theta_{\\mathrm{V}}^{\\mathrm{CLIP}}, \\theta_{\\mathrm{T}}^{\\mathrm{CLIP}}\\} \\) pre-trained by CLIP, where \\( \\mathcal{X} \\) and \\( \\mathcal{T} \\) are the image and text space. We optimize the parameters \\( \\theta_{\\mathrm{V}} \\) of \\( f_{\\mathrm{V}} \\) and \\( \\theta_{\\mathrm{T}} \\) of \\( f_{\\mathrm{T}} \\) on a post-pre-training image-text pair dataset \\( \\mathcal{D} = \\{(x_i, t_i)\\}_{i=1}^N \\); we basically assume that \\( \\mathcal{D} \\) is much smaller than the pre-training dataset and contains images and text captions in general domain." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.449, + 0.276, + 0.465 + ], + "angle": 0, + "content": "3.2. Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.501 + ], + "angle": 0, + "content": "In CLIP-Refine, we optimize \\(\\theta_{\\mathrm{V}}\\) and \\(\\theta_{\\mathrm{T}}\\) by the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.511, + 0.484, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {\\mathrm {V}}, \\theta_ {\\mathrm {T}}} \\mathcal {L} _ {\\mathrm {R a F A}} \\left(\\theta_ {\\mathrm {V}}, \\theta_ {\\mathrm {T}}\\right) + \\mathcal {L} _ {\\mathrm {H y C D}} \\left(\\theta_ {\\mathrm {V}}, \\theta_ {\\mathrm {T}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.485, + 0.636 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{RaFA}}\\) is the random feature alignment loss and \\(\\mathcal{L}_{\\mathrm{HyCD}}\\) is the hybrid contrastive-distillation loss. One can consider balancing \\(\\mathcal{L}_{\\mathrm{RaFA}}\\) and \\(\\mathcal{L}_{\\mathrm{HyCD}}\\) by hyperparameters, but we found that their equal contribution achieves the best performance (see Appendix). We describe the details of \\(\\mathcal{L}_{\\mathrm{RaFA}}\\) and \\(\\mathcal{L}_{\\mathrm{HyCD}}\\) in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.644, + 0.345, + 0.66 + ], + "angle": 0, + "content": "3.3. Random Feature Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.484, + 0.773 + ], + "angle": 0, + "content": "The primary goal in post-pre-training is to minimize the modality gap and encourage cross-modal alignment. To this end, the most straightforward approach is just minimizing the feature gap between the modalities. Let \\( z_{\\mathrm{img}}^i = f_{\\mathrm{V}}(x^i;\\theta_{\\mathrm{V}}) \\) and \\( z_{\\mathrm{xt}}^i = f_{\\mathrm{T}}(t^i;\\theta_{\\mathrm{T}}) \\) are the normalized feature vectors of an image-text pair \\( (x^{i},t^{i}) \\). A naive loss for minimizing the modality gap can be defined by an \\( L_{2} \\) distance form as" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.783, + 0.484, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\mathbb {E} _ {(x ^ {i}, t ^ {i}) \\in \\mathcal {D}} \\| z _ {\\text {i m g}} ^ {i} - z _ {\\text {t x t}} ^ {i} \\| _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.902 + ], + "angle": 0, + "content": "However, minimizing this loss function degrades the generalization performance. This can be explained by the balance of alignment and uniformity [53], which are desirable properties of transferability in contrastive representation learning. Alignment is defined by the gap between positive (i.e., imagetext) pairs in the feature space, and uniformity is defined by" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": "equal separation among all features on the hypersphere. That is, minimizing Eq. (2) enhances the alignment of the positive pair but destroys the uniformity on the hypersphere by forcing changes to the feature distribution. In fact, Wang and Isola [53] have shown that using only the alignment loss in post-pre-training degrades the validation performance due to corrupt uniformity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.909, + 0.349 + ], + "angle": 0, + "content": "To address this challenge, we introduce the idea of matching feature distributions instead of directly matching the paired features. In other words, we refine the multi-modal features of pre-trained CLIP models to follow a prior distribution \\( p(z) \\) (e.g., standard Gaussian \\( \\mathcal{N}(0, I) \\)) shared across the image and text modalities. Inspired by random feature regularization in single modal fine-tuning [59, 67], we formulate this idea as minimization between image/text feature vectors \\( (z_{\\mathrm{img}}^i, z_{\\mathrm{txt}}^i) \\) and a random reference vector \\( z_{\\mathrm{ref}}^i \\) sampled from the shared prior \\( p(z) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.36, + 0.907, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R a F A}} = \\mathbb {E} _ {(x ^ {i}, t ^ {i}) \\in \\mathcal {D}} \\frac {1}{2} (\\| z _ {\\mathrm {i m g}} ^ {i} - z _ {\\mathrm {r e f}} ^ {i} \\| _ {2} ^ {2} + \\| z _ {\\mathrm {t x t}} ^ {i} - z _ {\\mathrm {r e f}} ^ {i} \\| _ {2} ^ {2}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.909, + 0.565 + ], + "angle": 0, + "content": "Through this loss function, we can expect that (i) the modality gap is indirectly minimized via \\( z_{\\mathrm{ref}}^i \\) shared with each image-text pair, (ii) the feature vectors follow the modality shared prior \\( p(z) \\) gradually [59], (iii) the randomness of \\( z_{\\mathrm{ref}}^i \\) helps to avoid the models from overfitting to \\( \\mathcal{D} \\) [67] and reform the models' capacity [49]. Furthermore, we found that \\( \\mathcal{L}_{\\mathrm{RaFA}} \\) can improve the uniformity from baselines (Table 3), meaning that the learned representations transfer well to downstream tasks. Throughout this paper, we use standard Gaussian \\( \\mathcal{N}(0,I) \\) as \\( p(z) \\) by default; we discuss the effect of prior choice in Sec. 4.4.1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.792, + 0.59 + ], + "angle": 0, + "content": "3.4. Hybrid Contrastive-Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.704 + ], + "angle": 0, + "content": "Another important goal of post-pre-training is to retain the past knowledge in pre-trained CLIP models. To this end, our basic strategy is to penalize the post-pre-training models with knowledge distillation loss [19] where the frozen pre-trained CLIP model performs as the teacher. We define the minibatch-wise knowledge distillation loss with KL-divergence according to the formulation of CLIP-KD [60] as" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.714, + 0.907, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K D}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {K D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} + \\mathcal {L} _ {\\mathrm {K D}} ^ {\\mathrm {T} \\rightarrow \\mathrm {I}}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.735, + 0.907, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {B} q _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} \\log \\frac {q _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}{p _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.765, + 0.907, + 0.793 + ], + "angle": 0, + "content": "\\[\np _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {\\exp \\left(z _ {\\mathrm {i m g}} ^ {i} \\cdot z _ {\\mathrm {t x t}} ^ {j} / \\tau\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(z _ {\\mathrm {i m g}} ^ {i} \\cdot z _ {\\mathrm {t x t}} ^ {k} / \\tau\\right)}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.796, + 0.907, + 0.825 + ], + "angle": 0, + "content": "\\[\nq _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {\\exp \\left(z _ {\\mathrm {i m g}} ^ {i , \\mathrm {p}} \\cdot z _ {\\mathrm {t x t}} ^ {j , \\mathrm {p}} / \\tau\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(z _ {\\mathrm {i m g}} ^ {i , \\mathrm {p}} \\cdot z _ {\\mathrm {t x t}} ^ {k , \\mathrm {p}} / \\tau\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.838, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature parameter, \\((z_{\\mathrm{img}}^{i,\\mathrm{p}},z_{\\mathrm{xt}}^{i,\\mathrm{p}}) = (f_{\\mathrm{V}}(x^{i};\\theta_{\\mathrm{V}}^{\\mathrm{CLIP}}),f_{\\mathrm{T}}(t^{i};\\theta_{\\mathrm{T}}^{\\mathrm{CLIP}}))\\) are the feature vectors generated by the pre-trained image/text encoders, and \\(\\mathcal{L}_{\\mathrm{KD}}^{\\mathrm{T}\\rightarrow \\mathrm{I}}\\) is defined by swapping the order of image features and text" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4259" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.255, + 0.089, + 0.743, + 0.103 + ], + "angle": 0, + "content": "Table 1. Zero-shot classification performance on 12 datasets with CLIP ViT-B/32." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.104, + 0.907, + 0.235 + ], + "angle": 0, + "content": "
MethodAirBirdCalCarDTDEuroFloFoodINPetSUNUCFAvg.
Pre-trained18.8149.3723.9860.0740.9038.2059.7580.2159.0481.7960.8559.9052.74
Contrastive13.9840.0723.7747.6336.2234.6846.2870.3152.9670.1354.7858.2645.75
m2-mix [36]14.6440.7623.4949.8636.4934.6547.4271.2753.5870.3256.5158.7646.48
Self-KD18.9651.5223.9460.0441.1738.3659.5580.0259.0681.8560.6860.1652.94
HyCD18.7250.5023.9859.8340.4840.9058.0680.1959.7181.4562.0461.7053.13
HyCD + L_align13.2934.2023.7343.7035.0039.2746.2773.3152.5772.4255.0158.5545.61
CLIP-Refine (Ours)20.7752.7224.7960.6341.4941.4462.6481.8760.9383.5763.6361.8554.69
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.312 + ], + "angle": 0, + "content": "features in Eq. (6) and (7). Although minimizing Eq. (4) retains the past knowledge, we found that it interferes with RaFA because it strongly enchains the models' parameters to the pre-trained ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.484, + 0.405 + ], + "angle": 0, + "content": "To reconcile RaFA and the retention of past knowledge, we present a technique that blends the ground-truth label of an image-text pair with the teacher output. Concretely, we modify Eq. (7) by alpha-blending of an indicator function \\(\\mathbb{I}_{i = j}\\), which returns one if \\(i = j\\) otherwise zero, and the teacher output signal:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.41, + 0.484, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\hat {q} _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\alpha \\mathbb {I} _ {i = j} + (1 - \\alpha) q _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.485, + 0.573 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyperparameter to balance acquiring new knowledge and retaining past knowledge; we use \\(\\alpha = 0.5\\) by default and discuss the effect of \\(\\alpha\\) in the Appendix. With this modification, the post-pre-training models can further enhance the cross-modal alignment by learning the correct image-text pair (i.e., \\(i = j\\)) with reference to the relative knowledge contained in the \\(i \\neq j\\) coordinates of the teacher's outputs. By using \\(\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}}\\) and \\(\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}\\), we compute the HyCD loss function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.581, + 0.483, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {H y C D}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {H y C D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} + \\mathcal {L} _ {\\mathrm {H y C D}} ^ {\\mathrm {T} \\rightarrow \\mathrm {I}}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.602, + 0.483, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {H y C D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {B} \\hat {q} _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} \\log \\frac {\\hat {q} _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}{p _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.485, + 0.699 + ], + "angle": 0, + "content": "While \\(\\mathcal{L}_{\\mathrm{HyCD}}\\) solely improves the generalization performance of the pre-trained CLIP models, combining it with \\(\\mathcal{L}_{\\mathrm{RaFA}}\\) achieves larger improvements by minimizing the modality gaps without losing the past knowledge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.224, + 0.727 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.811 + ], + "angle": 0, + "content": "We evaluate CLIP-Refine on 12 classification datasets and 2 cross-modal retrieval datasets by using multiple pre-trained models. We also conduct qualitative and quantitative analyses to assess the feature space through modality gap, alignment, uniformity metrics, and PCA visualization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.19, + 0.835 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Baselines. We compare our CLIP-Refine with the following baselines of post-pre-training. Pre-trained: predicting with pre-trained weights without additional training. Contrastive: post-pre-training with the contrastive" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.908, + 0.359 + ], + "angle": 0, + "content": "loss [21, 40]. \\( m^2 \\)-mix [36]: post-pre-training by the modified contrastive loss with multi-modal mixup; we use this fine-tuning method as a baseline because it can be used for refining the feature spaces as suggested in the original paper [36]. Self-KD: using Eq. (4) for post-pre-training. HyCD: using only Eq. (9) for training. HyCD + L align: combining Eq. (9) and Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.907, + 0.427 + ], + "angle": 0, + "content": "Post-pre-training Datasets. We used COCO Caption [29] as the default post-pre-training dataset. We also used Flickr8K/30K [42], CC3M [47], and CC12M [3] to confirm the dataset size effect in post-pre-training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.909, + 0.614 + ], + "angle": 0, + "content": "Test Datasets. We used 12 image classification datasets containing various image domains: Aircraft (Air) [34], Bird [56], Caltech-101 (Cal) [12] Car [24], DTD [9], EuroSAT (Euro) [18], Flower (Flo) [35], Food [2], ImageNet (IN) [45], Pet [38], SUN397 [57], and UCF-101 [50]. We also evaluated our method on COCO2017-Val and Flickr8K/30K for the image and text retrieval tasks; we denote text-to-image retrieval as \\(\\mathrm{T} \\rightarrow \\mathrm{I}\\) and image-to-text retrieval as \\(\\mathrm{I} \\rightarrow \\mathrm{T}\\). We adopt these datasets since they are often used to evaluate the zero-shot transfer performance of CLIP models [68]. We evaluated models on the test sets except for ImageNet, and the official validation set for ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.909, + 0.713 + ], + "angle": 0, + "content": "Pre-trained Models. By default, we used CLIP-ViT-B/32 [40] with the pre-trained weights downloaded from the OpenAI's official repositories. We also tested other pretrained models, including OpenCLIP-ViT-H/14, OpenCLIP-ViT-bigG/14, SigLIP [65] and DFN [11] with the pre-trained weights on the OpenCLIP [8] repository." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Post-pre-training. In all settings, we trained models by the AdamW [32] optimizer with the learning rate of \\(1.0 \\times 10^{-6}\\) for one epoch; we determined the learning rate by the zero-shot classification validation using the subset of the ImageNet training set constructed by uniform sampling. We used mini-batch sizes of 512 as the default. For CLIP-Refine, we used \\(\\alpha = 0.5\\). The input samples were preprocessed using the default image transformations provided by each pre-trained model. We implemented the training and evaluation with PyTorch-1.13. We ran the experiments on a 24-core Intel Xeon CPU with a single NVIDIA A100 GPU with 80GB VRAM." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4260" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.199, + 0.089, + 0.796, + 0.103 + ], + "angle": 0, + "content": "Table 2. Zero-shot retrieval performance on COCO2017-Val and Flickr8K/30K with CLIP ViT-B/32." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.104, + 0.907, + 0.221 + ], + "angle": 0, + "content": "
MethodCOCO2017-ValFlickr8KFlickr30K
T→II→TR@1T→IR@1I→TR@1T→IR@1I→T
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@5R@10R@1R@5R@10R@1R@5R@10
Pre-trained30.5654.9265.2633.2659.1068.7830.0653.8363.2932.4255.5765.2621.0640.9650.5324.5244.5754.14
Contrastive34.8861.5072.1031.8656.8067.4834.9459.4169.3628.8351.5561.0725.6947.3256.9318.4236.2945.37
m2-mix [36]36.2862.1873.0832.9257.9667.5635.0459.6669.6628.6450.6560.7725.9947.1857.1919.8837.6346.51
Self-KD31.0455.5865.9035.3659.0069.7229.9853.8363.5532.4755.1165.1021.9241.5150.7524.9444.8254.27
HyCD36.0462.2873.1437.8862.5472.8435.7260.3470.1434.6957.1967.2227.1149.1258.7827.1747.6656.94
HyCD + L align33.9261.1872.0635.1461.7472.0631.7355.6365.9032.5856.9466.9923.7844.7354.4723.5845.0755.23
CLIP-Refine (Ours)37.6463.5474.4238.7865.0475.1236.1161.2970.5138.2861.5570.9426.6549.7059.0330.9951.5461.31
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.486, + 0.375 + ], + "angle": 0, + "content": "Evaluation Metrics. For zero-shot transfer performance, we report top-1 accuracy in classification tasks and recall@ \\( k \\) scores in cross-modal retrieval tasks; recall@ \\( k \\) indicates the proportion of correct answers in the retrieved top- \\( k \\) candidates for all test input samples. We denote recall@ \\( k \\) as \\( R@k \\). To quantitatively evaluate the post-pre-trained features, we measure modality gap, alignment, and uniformity scores. Modality gap [28] assesses a cluster-wise gap between image and text features:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.385, + 0.483, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\text {M o d a l i t y} \\mathrm {G a p} := \\| \\bar {f} _ {\\mathrm {V}} - \\bar {f} _ {\\mathrm {T}} \\| _ {2} ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.446 + ], + "angle": 0, + "content": "where \\(\\bar{f}\\) is a mean feature vector for all test samples of each modality; lower is better. Alignment [53] is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.457, + 0.483, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\text {A l i g n m e n t} := \\frac {1}{N _ {\\text {t e s t}}} \\sum_ {i = 1} ^ {N _ {\\text {t e s t}}} \\| f _ {\\mathrm {V}} \\left(x ^ {i}\\right) - f _ {\\mathrm {T}} \\left(t ^ {i}\\right) \\| _ {2} ^ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.583 + ], + "angle": 0, + "content": "representing how well the positive pair is aligned in the feature space; lower is better. Uniformity [53] evaluates how uniformly distributed the feature vectors are on the hypersphere, which is defined by radial basis function (RBF) kernel as" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.592, + 0.483, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\text {U n i f o r m i t y} := \\frac {1}{2 N _ {\\text {t e s t}}} \\sum_ {f _ {1}, f _ {2} \\in F} \\exp (- 2 \\| f _ {1} - f _ {2} \\| _ {2} ^ {2}), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.484, + 0.688 + ], + "angle": 0, + "content": "where \\( F = \\{f_{\\mathrm{V}}(x^i)\\}_{i=1}^{N_{\\mathrm{test}}} \\cup \\{f_{\\mathrm{T}}(t^i)\\}_{i=1}^{N_{\\mathrm{test}}} \\); lower is better. We report the average score for each metric by running experiments three times." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.363, + 0.713 + ], + "angle": 0, + "content": "4.2. Zero-shot Transfer Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We first demonstrate the zero-shot performance of our CLIP-Refine. Tables 1 and 2 list the zero-shot test performance on classification and cross-modal retrieval tasks after post-pretraining. For zero-shot classification tasks, post-pre-training by the contrastive loss (i.e., Contrastive and \\( m^2 \\)-mix) largely degraded the top-1 accuracy performance for all datasets. This can be caused by catastrophic forgetting since Self-KD did not show performance degradation. One cause of catastrophic forgetting is a smaller batchsize than at pre-training; the OpenAI's CLIP uses a batchsize of 32,768 with hundreds of GPUs [40], whereas our setting uses ones of 512 with a single GPU. Since a smaller batch contains a limited number" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.907, + 0.389 + ], + "angle": 0, + "content": "of negative samples, contrastive learning tends to overfit and catastrophically forget past knowledge. We discuss more details about the effect of batchsize in Sec. 4.4.4. In contrast, our CLIP-Refine impressively improved the zero-shot performance for all datasets. The ablation studies of HyCD show that the performance gain comes from the combination of HyCD with RaFA. Notably, \\(\\mathrm{HyCD} + \\mathcal{L}_{\\mathrm{align}}\\) resulted in the worst performance among the baselines, suggesting that naively minimizing the gap between image and text features leads to the degradations of generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.909, + 0.694 + ], + "angle": 0, + "content": "For zero-shot retrieval, our CLIP-Refine achieved the best performance with a larger margin than that in classification. Interestingly, in contrast to classification tasks, we found that Contrastive and \\( m^2 \\)-mix improved the pre-trained models in the \\( \\mathrm{T} \\rightarrow \\mathrm{I} \\) retrieval tasks while the \\( \\mathrm{I} \\rightarrow \\mathrm{T} \\) performance was degraded. This indicates that the image encoders are more prone to overfitting than the text encoders. This can be explained by the difference between image and text in the data spaces; images are represented by continuous-valued pixels and thus have a high degree of freedom, while texts are represented by discrete tokens in vocabulary and thus have a limited degree of freedom. Therefore, the image encoder tends to overfit to a limited number of negative samples in small batches, and the text encoder can learn new knowledge from them while keeping the past knowledge. Since the zero-shot classification is a \\( \\mathrm{I} \\rightarrow \\mathrm{T} \\) retrieval task, this can also be a reason for the performance degradation in classification. In this perspective, our CLIP-Refine successfully prevents the overfitting in \\( \\mathrm{I} \\rightarrow \\mathrm{T} \\) and greatly improves the performance in both \\( \\mathrm{I} \\rightarrow \\mathrm{T} \\) and \\( \\mathrm{T} \\rightarrow \\mathrm{I} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.748, + 0.727 + ], + "angle": 0, + "content": "4.3. Analysis of Feature Space" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this section, we aim to answer our fundamental research question: how much can we alleviate the modality gap by post-pre-training with CLIP-Refine? For quantitative evaluations, we summarize the modality gap, alignment, and uniformity scores in Table 3. HyCD \\(+\\mathcal{L}_{\\mathrm{align}}\\) achieved the best modality gap and alignment scores. However, it significantly degraded the uniformity score, indicating that directly minimizing \\(\\mathcal{L}_{\\mathrm{align}}\\) can destroy feature space well constructed by the CLIP training. In contrast, our CLIP-Refine consistently improved all of the scores. This suggests that it is important to minimize not only the gap between image and text features" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4261" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.301, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.204, + 0.241, + 0.216 + ], + "angle": 0, + "content": "(a) Pre-trained" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.088, + 0.515, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.204, + 0.456, + 0.216 + ], + "angle": 0, + "content": "(b) Contrastive" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.088, + 0.729, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.204, + 0.676, + 0.217 + ], + "angle": 0, + "content": "(c) \\(\\mathrm{HyCD} + \\mathcal{L}_{\\mathrm{align}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.088, + 0.943, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.204, + 0.904, + 0.216 + ], + "angle": 0, + "content": "(d) CLIP-Refine (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.228, + 0.908, + 0.258 + ], + "angle": 0, + "content": "Figure 3. PCA visualization of multi-modal feature spaces of trained models (Flickr8K, ViT-B/32). CLIP-Refine reduces the modality gap while keeping the form of feature distributions for each modality." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.271, + 0.482, + 0.285 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation of features (Flickr8K, ViT-B/32)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.295, + 0.481, + 0.391 + ], + "angle": 0, + "content": "
MethodModality Gap ×10-3(↓)Alignment (↓)Uniformity (↓)
Pre-trained1.33451.37240.0895
Contrastive2.17141.43000.2136
m2-mix [36]2.09111.41300.2019
Self-KD1.27541.37740.0901
HyCD1.38231.35780.0971
HyCD+L align0.15900.74130.1172
CLIP-Refine (Ours)0.79341.28490.0495
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.405, + 0.484, + 0.526 + ], + "angle": 0, + "content": "but also the uniformity for improving the generalization performance of pre-trained CLIP models, which is consistent with the findings in [53]. In such a sense, since CLIP-Refine penalizes the image and text features to follow the shared reference distribution \\( p(z) = \\mathcal{N}(0,I) \\), where the samples tend to be uniformly distributed on the hypersphere [1], it can naturally achieve a good balance of alignment and uniformity, and reasonably helps to reduce the modality gap." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.527, + 0.484, + 0.68 + ], + "angle": 0, + "content": "As a qualitative evaluation, Figure 3 illustrates the PCA visualization of post-pre-trained features for each method. We can see that the contrastive model makes the modality gap rather large. \\(\\mathrm{HyCD} + \\mathcal{L}_{\\mathrm{align}}\\) succeeds in gathering the image and text features into the same cluster but distorts the feature distribution for each modality. Meanwhile, CLIP-Refine certainly reduces the modality gap while keeping the shape of the pre-trained features. Although the modality gap is not completely zero, it is natural because there remains a constant value in the lower bound of the modality gap [39]." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.69, + 0.262, + 0.707 + ], + "angle": 0, + "content": "4.4. Detailed Analysis" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.27, + 0.727 + ], + "angle": 0, + "content": "4.4.1. Prior Distributions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.486, + 0.904 + ], + "angle": 0, + "content": "Here, we analyze CLIP-Refine by varying the prior distribution \\(p(z)\\). According to the prior works [59, 67], we tried standard Gaussian \\(\\mathcal{N}(0,I)\\), uniform distribution \\(U(0,1)\\), Gaussian distributions using pre-trained text feature statistics \\(\\mathcal{N}(\\mu_{\\mathrm{txt}}^{\\mathrm{P}},(\\sigma_{\\mathrm{txt}}^{\\mathrm{P}})^2)\\), image feature statistics \\(\\mathcal{N}(\\mu_{\\mathrm{img}}^{\\mathrm{P}},(\\sigma_{\\mathrm{img}}^{\\mathrm{P}})^2)\\), statistics of all image and text features \\(\\mathcal{N}(\\mu_{\\mathrm{all}}^{\\mathrm{P}},(\\sigma_{\\mathrm{all}}^{\\mathrm{P}})^2)\\), where \\(\\mu^{\\mathrm{P}},\\sigma^{\\mathrm{P}}\\) is mean and variance of image or/and text feature vectors computed on a pre-trained CLIP model. We also tested variations of \\(\\mathcal{N}(0,\\beta I)\\) by varying \\(\\beta\\) in \\(\\{0,0.01,0.1,10,100\\}\\); \\(\\beta = 0\\) means RaFA without randomness, i.e., \\(\\frac{1}{2} (\\| z_{\\mathrm{img}}^{i}\\|_{2}^{2} + \\| z_{\\mathrm{txt}}^{i}\\|_{2}^{2})\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.271, + 0.907, + 0.312 + ], + "angle": 0, + "content": "Table 4. Evaluation of CLIP-Refine varying prior distributions in \\(\\mathcal{L}_{\\mathrm{RaFA}}\\). ZS Cls. is the zero-shot classification accuracy averaged on 12 datasets." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.323, + 0.882, + 0.505 + ], + "angle": 0, + "content": "
PriorZS Cls.AlignmentUniformity
Pretrained52.741.37240.0895
U(0,1)52.911.30810.0661
N(μpimg, (σpimg)2)52.931.31420.0916
N(μptxt, (σptxt)2)53.401.33040.0873
N(μpall, (σpall)2)53.291.33610.0888
N(0,0I)53.791.30520.0554
N(0,0.01I)54.191.30790.0544
N(0,0.1I)54.321.30210.0515
N(0,I)54.691.28490.0495
N(0,10I)54.231.30850.0537
N(0,100I)53.591.34410.0663
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.909, + 0.791 + ], + "angle": 0, + "content": "Table 4 shows the zero-shot classification performance, alignment, and uniformity scores when varying \\( p(z) \\). Using the uniform distribution \\( U(0,1) \\) did not show large improvements on the zero-shot performance and uniformity score. This may be because random reference vectors from \\( U(0,1) \\) do not necessarily concentrate on the hypersphere in contrast to \\( \\mathcal{N}(0,I) \\), possibly causing the collapse of uniformity by \\( \\mathcal{L}_{\\mathrm{RaFA}} \\). In this perspective, we recommend using \\( \\mathcal{N}(0,I) \\) or families of Gaussian distributions. Interestingly, among the Gaussian distributions with the pre-trained statistics, the prior with the text feature, i.e., \\( \\mathcal{N}(\\mu_{\\mathrm{txt}}^{\\mathrm{p}}(\\sigma_{\\mathrm{txt}}^{\\mathrm{p}})^2) \\) largely improved the zero-shot performance. This indicates that aligning image features with text features is effective in zero-shot classification tasks and also evidences the performance improvements in the \\( \\mathrm{I} \\rightarrow \\mathrm{T} \\) tasks observed in Sec. 4.2. More importantly, the ablation studies of the variance magnitude \\( \\beta \\) in the zero-mean Gaussian distributions emphasize the importance of randomness in \\( \\mathcal{L}_{\\mathrm{RaFA}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.805, + 0.744, + 0.82 + ], + "angle": 0, + "content": "4.4.2. Post-pre-training Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We evaluate the effects of the dataset choice in post-pretraining. Table 5 shows the zero-shot classification/retrieval performance of CLIP-Refine with Flickr8K/30K, COCO Caption, CC3M, and CC12M as post-pre-training datasets. We confirm that most of the CLIP-Refine models outper" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4262" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 5. Zero-shot classification (12 datasets) and retrieval (Flicker30K) performance of CLIP-Refine with various post-pre-training datasets with ViT-B/32." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.143, + 0.482, + 0.261 + ], + "angle": 0, + "content": "
Post-pre-training Dataset# of SamplesZero-shot Cls. Acc.Zero-shot Ret. R@5
Avg. of 12 datasetsT→II→T
None (Pre-trained)052.7440.9644.57
Flickr8K8K53.0445.5746.80
Flickr30K30K53.7448.6649.25
COCO Caption118K54.6949.7051.54
CC3M3M53.8049.5650.24
CC12M12M52.4645.3547.79
Filtered CC3M1M55.5052.4553.00
Filtered CC12M4M60.1254.9757.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Table 6. Zero-shot classification (12 datasets) and retrieval (Flickr30K) performance with various pre-trained CLIP models." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.314, + 0.482, + 0.384 + ], + "angle": 0, + "content": "
ModelZero-shot Cls. Acc.Zero-shot Ret. R@5 (T→I/I→T)
Pre-trainedContrastiveCLIP-RefinePre-trainedContrastiveCLIP-Refine
ViT-B/3252.7445.7654.6940.96/44.5747.32/36.2949.70/51.54
ViT-L/1462.9661.9063.8849.13/52.7751.07/51.0058.38/58.41
OpenCLIP ViT-H/1468.5257.6169.0565.84/66.0257.47/43.4367.82/68.48
OpenCLIP ViT-G/1471.2857.4472.5968.73/68.3866.35/42.3670.94/69.76
SigLip ViT-SO400M/1472.5765.6373.6172.77/72.8364.63/62.9776.54/76.49
DFN ViT-H/1477.3076.3978.2971.93/70.8972.81/70.0976.39/75.44
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.393, + 0.486, + 0.592 + ], + "angle": 0, + "content": "formed the pre-trained baselines. Notably, larger datasets do not always achieve high performance; CC3M and CC12M are inferior to COCO Caption. This suggests that the quality of image-text pairs in a dataset is an important factor in post-pre-training. Since CC3M and CC12M contain mismatched and noisy image-text pairs [17], COCO Caption that has higher-quality text captions can be superior in terms of the modality alignment of pre-trained CLIP models. In fact, when we filtered out low-quality image-text pairs by CLIP-Score-based filtering of DataComp [13], the models achieved significant performance improvements. This indicates that the caption quality is crucial and our method is scalable if we have high-quality image-text pairs." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.598, + 0.273, + 0.612 + ], + "angle": 0, + "content": "4.4.3. Pre-trained Models" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.616, + 0.486, + 0.738 + ], + "angle": 0, + "content": "We investigate the generalizability of CLIP-Refine to pretrained CLIP models. Table 6 shows the zero-shot performance with 6 pre-trained models. We observed that our CLIP-Refine stably improved the pre-trained baseline. This suggests that the bigger and stronger models still have the modality gap, and modifying the gap by CLIP-Refine helps improve the generalization performance for various pretrained models." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.746, + 0.258, + 0.76 + ], + "angle": 0, + "content": "4.4.4. Mini-Batch Sizes" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We provide the analysis of varying mini-batch sizes in post-pre-training. We tried the mini-batch sizes in \\(\\{32, 64, 128, 256, 512, 1024, 2048\\}\\). We used gradient accumulation by following the OpenCLIP [8] implementation for the mini-batch sizes of 1024 and 2048. Figure 4 shows the zero-shot performance versus mini-batch sizes. The contrastive loss gradually improved the accuracy according to the mini-batch size. However, the accuracy curve implies that it is difficult to exceed the pre-trained baseline with" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.088, + 0.852, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.216, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Figure 4. Zero-shot classification accuracy averaged on 12 datasets when varying mini-batch sizes in post-pre-training." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.248, + 0.907, + 0.263 + ], + "angle": 0, + "content": "Table 7. Fine-tuning top-1 accuracy on the ImageNet validation set." + }, + { + "type": "table", + "bbox": [ + 0.555, + 0.273, + 0.862, + 0.313 + ], + "angle": 0, + "content": "
ModelPre-trainedContrastiveCLIP-Refine
ViT-B/3273.4069.4674.14
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.909, + 0.472 + ], + "angle": 0, + "content": "feasible mini-batch sizes. This is probably due to the large discrepancy from the mini-batch size in the pre-training (i.e., 32,768), which causes the model to forget knowledge in the predictions for negative pairs in a mini-batch. In contrast, CLIP-Refine always outperformed the contrastive loss and succeeded in improving the pre-trained baseline with feasible mini-batch sizes. This suggests that CLIP-Refine's HyCD prevents knowledge forgetting even for small mini-batch sizes with few negative pairs and is able to modify the feature space with new knowledge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.478, + 0.719, + 0.494 + ], + "angle": 0, + "content": "4.4.5. Effects on Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.909, + 0.589 + ], + "angle": 0, + "content": "We show the transferability of CLIP-Refine to fine-tuning. We fine-tuned the linear classification head on ImageNet while freezing post-pre-trained models. Tabel 7 shows the results. CLIP-Refine outperformed the pre-trained and contrastive loss baselines. This emphasizes that our method can generate useful representation even in fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.634, + 0.617 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.909, + 0.9 + ], + "angle": 0, + "content": "This paper presented CLIP-Refine, a post-pre-training method for pre-trained CLIP models to align the modality gap between the image and text features. CLIP-Refine addresses the modality gap by penalizing the multi-modal features to follow a shared prior distribution by minimizing the distance to the random reference vectors sampled from the prior. To maintain the past knowledge in the CLIP models and promote the feature alignment simultaneously, CLIP-Refine also trains the model with knowledge distillation loss using hybrid soft labels composed of ground-truth image-text pair labels and outputs from the pre-trained CLIP model. Through extensive experiments, we show that CLIP-Refine can improve the zero-shot performance of the pre-trained CLIP by addressing the modality gap and enhancing uniformity. We believe that our work not only provides a practical method but also opens up a new research field where we refine the pre-trained CLIP models by post-pre-training with much smaller computational costs than that of pre-training." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4263" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Christopher M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag, Berlin, Heidelberg, 2006. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[3] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.289, + 0.483, + 0.358 + ], + "angle": 0, + "content": "[4] Lianggangxu Chen, Xuejiao Wang, Jiale Lu, Shaohui Lin, Changbo Wang, and Gaoqi He. Clip-driven open-vocabulary 3d scene graph generation via cross-modality contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[5] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yu Hou, Yuenan and Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.484, + 0.541 + ], + "angle": 0, + "content": "[6] Yuxiao Chen, Jianbo Yuan, Yu Tian, Shijie Geng, Xinyu Li, Ding Zhou, Dimitris N Metaxas, and Hongxia Yang. Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15095-15104, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[7] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.484, + 0.699 + ], + "angle": 0, + "content": "[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[10] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. In International Conference on Learning Representations, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[12] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In Conference on CVPR Workshop, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[13] Gadre et al. Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[14] Yunhao Ge, Jie Ren, Andrew Gallagher, Yuxiao Wang, Ming-Hsuan Yang, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11093-11101, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[15] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.335, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[16] Haochen Han, Qinghua Zheng, Guang Dai, Minnan Luo, and Jingdong Wang. Learning to rematch mismatched pairs for robust cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.907, + 0.474 + ], + "angle": 0, + "content": "[17] Haochen Han, Qinghua Zheng, Guang Dai, Minnan Luo, and Jingdong Wang. Learning to rematch mismatched pairs for robust cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26679-26688, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[18] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.907, + 0.588 + ], + "angle": 0, + "content": "[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.907, + 0.644 + ], + "angle": 0, + "content": "[20] Ahmet Iscen, Mathilde Caron, Alireza Fathi, and Cordelia Schmid. Retrieval-enhanced contrastive vision-text models. In International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[22] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[23] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multimodal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition, Sydney, Australia, 2013. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4264" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[25] Zhengfeng Lai, Noranart Vesdapunt, Ning Zhou, Jun Wu, Cong Phuoc Huynh, Xuelu Li, Kah Kuen Fu, and Chen-nee Chuah. Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[26] Janghyeon Lee, Jongsuk Kim, Hyounguk Shon, Bumsoo Kim, Seung Hwan Kim, Honglak Lee, and Junmo Kim. Uniclip: Unified framework for contrastive language-image pre-training. Advances in Neural Information Processing Systems, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[27] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In International Conference on Learning Representations, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[28] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.389, + 0.483, + 0.457 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision. Springer, 2014. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[33] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[34] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. arXiv, 2013. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[35] M-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[36] Changdae Oh, Junhyuk So, Hoyoon Byun, YongTaek Lim, Minchul Shin, Jong-June Jeon, and Kyungwoo Song. Geodesic multi-modal mixup for robust fine-tuning. In Advances in Neural Information Processing Systems, 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[38] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[39] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In Advances in Neural Information Processing Systems, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.908, + 0.231 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.908, + 0.286 + ], + "angle": 0, + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.908, + 0.357 + ], + "angle": 0, + "content": "[42] Cyrus Rashtchian, Peter Young, Micah Hodosh, and Julia Hockenmaier. Collecting image annotations using amazon's mechanical turk. In Proceedings of the NAACL HLT 2010 workshop on creating speech and language data with Amazon's Mechanical Turk, pages 139-147, 2010. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.908, + 0.413 + ], + "angle": 0, + "content": "[43] Arijit Ray, Filip Radenovic, Abhimanyu Dubey, Bryan Plummer, Ranjay Krishna, and Kate Saenko. Cola: A benchmark for compositional text-to-image retrieval. In Advances in Neural Information Processing Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.908, + 0.482 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.908, + 0.552 + ], + "angle": 0, + "content": "[45] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.908, + 0.621 + ], + "angle": 0, + "content": "[46] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.908, + 0.691 + ], + "angle": 0, + "content": "[47] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.908, + 0.761 + ], + "angle": 0, + "content": "[48] Sheng Shen, Shijia Yang, Tianjun Zhang, Bohan Zhai, Joseph E Gonzalez, Kurt Keutzer, and Trevor Darrell. Multitask vision-language prompt tuning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.908, + 0.831 + ], + "angle": 0, + "content": "[49] Sicong Shen, Yang Zhou, Bingzheng Wei, Eric I Chang, Yan Xu, et al. Tuning stable rank shrinkage: Aiming at the overlooked structural risk in fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.908, + 0.871 + ], + "angle": 0, + "content": "[50] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[51] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4265" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3835-3844, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[52] Jingyun Wang and Guoliang Kang. Learn to rectify the bias of clip for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4102-4112, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.263 + ], + "angle": 0, + "content": "[53] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.32 + ], + "angle": 0, + "content": "[54] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.377 + ], + "angle": 0, + "content": "[55] Zhengbo Wang, Jian Liang, Ran He, Nan Xu, Zilei Wang, and Tieniu Tan. Improving zero-shot generalization for clip with synthesized prompts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.484, + 0.42 + ], + "angle": 0, + "content": "[56] P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD Birds 200. Technical report, California Institute of Technology, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[57] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.484, + 0.575 + ], + "angle": 0, + "content": "[58] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[59] Shin'ya Yamaguchi, Sekitoshi Kanai, Kazuki Adachi, and Daiki Chijiwa. Adaptive random feature regularization on fine-tuning deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[60] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. Clipkd: An empirical study of clip model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15952-15962, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[61] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19163-19173, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[62] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[63] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.908, + 0.135 + ], + "angle": 0, + "content": "Chunjing Xu. FILIP: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.205 + ], + "angle": 0, + "content": "[64] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. In European Conference on Computer Vision. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.907, + 0.33 + ], + "angle": 0, + "content": "[66] Yuhui Zhang, Jeff Z. HaoChen, Shih-Cheng Huang, Kuan-Chieh Wang, James Zou, and Serena Yeung. Diagnosing and rectifying vision models using language. In The Eleventh International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[67] Yang Zhong and Atsuto Maki. Regularizing cnn transfer learning with randomised regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13637-13646, 2020. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[68] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.907, + 0.489 + ], + "angle": 0, + "content": "[69] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4266" + } + ] +] \ No newline at end of file diff --git a/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/1196e091-a513-40f9-85ef-51604d8681b7_origin.pdf b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/1196e091-a513-40f9-85ef-51604d8681b7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3ceb85f222158fb17af68805885b9ce3248220e6 --- /dev/null +++ b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/1196e091-a513-40f9-85ef-51604d8681b7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b96aa0127dbc3ea0e3b6a9d5149d7a0889d8e77c158e69401797e308b6ab117 +size 1020197 diff --git a/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/full.md b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d42130a711601d34b63df6b14339984e0b4d5151 --- /dev/null +++ b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/full.md @@ -0,0 +1,377 @@ +# Post-pre-training for Modality Alignment in Vision-Language Foundation Models + +Shin'ya Yamaguchi* NTT, Kyoto University + +Dewei Feng +MIT + +Sekitoshi Kanai NTT + +Kazuki Adachi NTT + +Daiki Chijiwa NTT + +# Abstract + +Contrastive language image pre-training (CLIP) is an essential component of building modern vision-language foundation models. While CLIP demonstrates remarkable zero-shot performance on downstream tasks, the multi-modal feature spaces still suffer from a modality gap, which is a gap between image and text feature clusters and limits downstream task performance. Although existing works attempt to address the modality gap by modifying pre-training or fine-tuning, they struggle with heavy training costs with large datasets or degradations of zero-shot performance. This paper presents CLIP-Refine, a post-pre-training method for CLIP models at a phase between pre-training and fine-tuning. CLIP-Refine aims to align the feature space with 1 epoch training on small image-text datasets without zero-shot performance degradations. To this end, we introduce two techniques: random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). RaFA aligns the image and text features to follow a shared prior distribution by minimizing the distance to random reference vectors sampled from the prior. HyCD updates the model with hybrid soft labels generated by combining ground-truth image-text pair labels and outputs from the pre-trained CLIP model. This contributes to achieving both maintaining the past knowledge and learning new knowledge to align features. Our extensive experiments with multiple classification and retrieval tasks show that CLIP-Refine succeeds in mitigating the modality gap and improving the zero-shot performance1. + +# 1. Introduction + +Contrastive language image pre-training (CLIP, [21, 40]) is a standard method to build modern vision-language foundation models. CLIP enables models to learn multimodal representations to map images and texts into a shared feature space with contrastive loss on large-scale image-text pair datasets. Since the pre-trained CLIP models provide a cross-modal understanding of input image/text data in various domains, + +![](images/9d73534414c3d8d7094939027bbfc8aae79a0a8f314d80db69639ba5512a8d68.jpg) +Figure 1. Scope of our work: post-pre-training to align the modality gap in pre-trained CLIP models. We aim to address the modality gap and enhance the generalization performance of pre-trained CLIP models through lightweight training. + +they are widely used as a foundation of many applications, including zero-shot classification [14, 40, 55], cross-modal retrieval [16, 20], text-to-image generation [41, 44], and visual question answering [30, 31]. + +While CLIP achieves remarkable performance in broad domains and tasks, its image and text alignment is still not perfect. For example, CLIP models tend to encode images and texts into different clusters for each modality, and thus, there is a modality gap between images and text features even after sufficiently training with the massive datasets [28, 39, 66]. This modality gap suggests that CLIP has difficulty in precisely mapping images and text. In fact, Liang et al. [28] have shown that the modality gap largely affects downstream task performance, especially in fine-grained classification tasks, and Ray et al. [43] have demonstrated that CLIP models often fail to retrieve localized objects and attributes in images from corresponding captions. + +To address the modality gap in CLIP, existing literature mainly focuses on refining the CLIP feature spaces in the pre-training or fine-tuning phases. For pre-training, one approach is multi-task learning of the contrastive objective and auxiliary losses such as self-supervised loss with data augmentations [26, 27]. Another direction for pre-training is to reduce the modality gap by modifying the encoder architectures to explicitly share the weights or feature maps between the image and text encoders [6, 64]. For fine-tuning, previous works have demonstrated that the target downstream task performance can be improved significantly by prompt tuning that optimizes trainable visual/textual tokens on target tasks to match image and text in the feature space (i.e., reducing the modality gap) [23, 33, 68]. More recently, Yang et al. [62] have introduced adapter parameters to share the infor + +mation across modalities. This cross-modal shared adapter is fine-tuned to solve the target tasks. + +Although these pre-training/fine-tuning methods succeed in improving the cross-modal alignment capability, they face difficulties in terms of computational cost and zero-shot transfer performance. On the one hand, pre-training methods require training from scratch with million-scale image-text datasets like LAION-400M [46]. Since the previous works often do not publish pre-trained models using such large datasets, naive CLIP models published by OpenAI [40] or OpenCLIP [8] are still a practical choice under the limited budget. On the other hand, fine-tuning methods do not require huge computation costs because their datasets are much smaller than pre-training ones. Nevertheless, fine-tuning pretrained CLIP models degrades the general zero-shot transfer performance due to focusing on target tasks [25]. This is inevitable as the primary goal of fine-tuning is to improve the target task performance. Therefore, we seek a new approach that is lighter than pre-training and avoids degrading zero-shot transfer performance in fine-tuning. + +In this paper, we tackle aligning the image and text features of CLIP models in a training phase called post-pretraining between pre-training and fine-tuning (Fig. 1). The goal of post-pre-training is to mitigate the modality gap and enhance zero-shot transfer performance of off-the-shelf pretrained CLIP models by only using reasonable computation resources and datasets (e.g., a single GPU and COCO Captions [29]). This is challenging because naive alignment methods, such as directly minimizing the gap between image and text features, corrupt the feature spaces in terms of the uniformity on the hypersphere [53]; the uniformity is an important property of contrastive learning representation that indicates the amount of input data information. Furthermore, post-pre-training with contrastive loss causes catastrophic forgetting of the general knowledge in pre-trained CLIP models due to the overfitting by the restricted mini-batch sizes. + +To this end, we introduce a method for post-pre-training called CLIP-Refine (Figure 2), which is composed of random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). Instead of directly minimizing the gap between features, our idea is to refine the image and text feature distributions to follow a shared prior (e.g., standard Gaussian). To achieve this, RaFA minimizes the gap between the image/text feature vectors and the reference vectors, which are randomly generated from the prior. In RaFA, the reference vectors are shared for the pairs of input image and text. Thus, RaFA penalizes image and text features to explicitly follow the same distribution. By matching the image and text feature distributions, the features are expected to avoid excessive concentration and preserve uniformity among samples, resulting in a good balance of multi-modal alignment and uniformity in the feature space. Conversely, HyCD is designed to avoid catastrophic forgetting by modified self-distillation + +loss with the supervision of image-text pairs. Specifically, we apply the knowledge distillation where the teacher is the pre-trained model, i.e., minimizing KL-divergence between the outputs of the teacher and (student) post-pre-training model. Furthermore, we encourage learning new knowledge by blending the teacher's similarity matrix and the identity matrix, which represents the supervision of matched image-text pairs. Combining RaFA and HyCD, the CLIP models can reduce the modality gap while maintaining the uniformity of features without forgetting past knowledge. + +Our extensive experiments on zero-shot classification (12 datasets) and cross-modal retrieval (2 datasets) demonstrate that our method significantly improves zero-shot performance over post-pre-training baselines with contrastive loss. Through the quantitative and qualitative analysis of feature spaces, we found that CLIP-Refine not only reduces the modality gap but also improves the uniformity of the hypersphere. This suggests that post-pre-training facilitates effective cross-modal alignment and well-clustered features. + +# 2. Related Work + +CLIP and Modality Gap. CLIP [21, 40] is a multi-modal representation learning technique that trains a model to embed inputs from image and text modalities into shared feature space. This is done by making the correct pairs of image and text closer in the feature space while the other pairs in mini-batch repelled via the InfoNCE-based contrastive loss [37]. While this simple multi-modal pre-training has remarkably advanced multi-modal research in various fields [4, 5, 7, 51, 52, 54, 58], Liang et al. [28] have revealed that pre-trained CLIP models encode image and text into different clusters for each modality, i.e., modality gap, and the small temperature parameter in the CLIP's contrastive loss can cause the modality gap empirically. The successor work by Qian et al. [39] has theoretically shown that contrastive loss can not reduce the modality gap perfectly. + +Pre-training Modifications. To mitigate the modality gap and improve the cross-modal alignment, several works have modified the contrastive loss of CLIP by adding auxiliary losses, including geometric cyclic consistency between image and text features [15], the fine-grained similarity between output tokens of both modalities [63], self-supervised loss with data augmentation [26, 27], supervised contrastive learning [61], textual augmented contrastive loss using LLM-generated positive/negative texts [10]. Another direction is to explicitly share the information of image and text modalities by introducing shared encoder weights [64] or shared feature spaces of finite discrete tokens [6] to the pre-training. These methods, however, require large-scale image-text pair datasets (e.g., CC12M [3] and LAION-400M [46]) to achieve practical performance, incurring high computational costs on multiple GPUs. In contrast, our post-pre-training method improves the generalization performance of off-the + +![](images/c37fd29f21c5043056c83372df2d3daf26f072ac24d98476e4033eef5e809084.jpg) +Random Feature Alignment + +![](images/1d936d63ae7949b54b4104ce354ec93c3e115be645ac5ca6b846ef9544848e72.jpg) +Hybrid Contrastive-Distillation +Figure 2. Intuition of CLIP-Refine (proposed method). CLIP-Refine modifies a pre-trained CLIP model by Random Feature Alignment (RaFA, left) and Hybrid Contrastive-Distillation (HyCD, right). RaFA penalizes image and text features to follow a shared prior distribution by minimizing the gap between the features and a random reference vector sampled from the prior. HyCD trains the model with the KL-divergence-based distillation loss using hybrid soft labels composed by alpha blending the ground truth label of image-text pairs with the output of the pre-trained model (teacher). + +shelf pre-trained models by only using small datasets like Flickr8/30K [42] and COCO Caption [29] with a single GPU. Besides, our method is cooperative with these pre-training methods because it can be used for any pre-trained model. + +Fine-tuning Methods2. The original paper of CLIP [40] has reported lower zero-shot performance in several fine-grained classification tasks, such as Aircraft [34], demonstrating the imperfectness of the cross-modal alignment of CLIP. Motivated by this, Zhou et al. [68, 69] have proposed to optimize text prompts to mitigate the modality gap by updating additional trainable vectors while the pre-trained models are fixed. This simple approach helps match image and text in the feature space (i.e., reducing the modality gap) and significantly improves target performance. Similarly, The method of Jia et al. [22] learns the additional trainable parameters for image input (visual prompt), and the successor works [23, 48] have unified the visual and textual prompt tuning by the multi-task learning objectives. Another approach is to directly share the low-rank adaptation parameters across the modalities and update them in fine-tuning [62]. In the perspective of the refinement of CLIP feature spaces, Oh et al. [36] have theoretically and empirically shown that fine-tuned CLIP models retain the suboptimal feature space in terms of alignment and uniformity, which are important properties to assess the feature quality in contrastive learning [53]. To refine the feature space, they have provided a mixup-based fine-tuning method, which generates hard negative samples on the hypersphere in a cross-modal manner ( $m^2$ -mix). These fine-tuning techniques succeed in reducing the gap in target datasets, but they do not intend to improve the zero-shot transfer performance. Our post-pre-training method differs from the fine-tuning methods in the objective: our method aims to improve the zero-shot transfer performance by reducing the modality gap. Obviously, we can + +# Algorithm 1 CLIP-Refine for Post-pre-training + +Require: Training dataset $\mathcal{D}$ , vision encoder $f_{\mathrm{V}}$ , text encoder $f_{\mathrm{T}}$ , pre-trained parameters $\{\theta_{\mathrm{V}}^{\mathrm{P}}, \theta_{\mathrm{T}}^{\mathrm{P}}\}$ , training batchsize $B$ , step size $\eta$ , temperature parameter $\tau$ , hyper-parameters $\alpha$ + +Ensure: Post-pre-trained parameters $\{\theta_{\mathrm{V}},\theta_{\mathrm{T}}\}$ + +1: # Initialize parameters +2: $\theta_{\mathrm{V}} \gets \theta_{\mathrm{V}}^{\mathrm{P}}$ +3: $\theta_{\mathrm{T}} \gets \theta_{\mathrm{T}}^{\mathrm{P}}$ +4: while not converged do +5: $\{(x^i,t^i)\}_{i = 1}^B\sim \mathcal{D}$ +6: $\{z_{\mathrm{img}}^i,z_{\mathrm{txt}}^i\}_{i = 1}^B\gets \{f_{\mathrm{V}}(x^i;\theta_{\mathrm{V}}),f_{\mathrm{T}}(t^i;\theta_{\mathrm{T}})\}_{i = 1}^B$ +7: $\{z_{\mathrm{img}}^{i,\mathrm{p}}, z_{\mathrm{xt}}^{i,\mathrm{p}}\}_{i=1}^{B} \gets \{f_{\mathrm{V}}(x^i; \theta_{\mathrm{V}}^{\mathrm{p}}), f_{\mathrm{T}}(t^i; \theta_{\mathrm{T}}^{\mathrm{p}})\}_{i=1}^{B}$ +8: # Compute random feature alignment loss +9: $\{z_{\mathrm{ref}}^{\mathrm{i}}|z_{\mathrm{ref}}^{\mathrm{i}}\sim p(z)\}_{i = 1}^{B}$ +10: $\mathcal{L}_{\mathrm{RaFA}}\gets \frac{1}{2B}\sum_{i = 1}^{B}\| z_{\mathrm{img}}^i -z_{\mathrm{ref}}^i\| _2^2 +\| z_{\mathrm{txt}}^i -z_{\mathrm{ref}}^i\| _2^2$ +11: # Compute hybrid contrastive-distillation loss +12: Compute student outputs $p_{i,j}^{\mathrm{I}\rightarrow \mathrm{T}}$ and $p_{i,j}^{\mathrm{T}\rightarrow \mathrm{I}}$ for all sample combinations in a batch by Eq. (6) with $\tau$ . +13: Compute teacher outputs $\hat{q}_{i,j}^{\mathrm{I}\rightarrow \mathrm{T}}$ and $\hat{q}_{i,j}^{\mathrm{T}\rightarrow \mathrm{I}}$ for all sample combinations in a batch by Eq. (8) with $\tau$ and $\alpha$ . +14: $\mathcal{L}_{\mathrm{HyCD}}^{\mathrm{I}\rightarrow \mathrm{T}}\gets \frac{1}{B}\sum_{i = 1}^{B}\sum_{j = 1}^{B}\hat{q}_{i,j}^{\mathrm{I}\rightarrow \mathrm{T}}\log \frac{\hat{q}_{i,j}^{\mathrm{I}\rightarrow\mathrm{T}}}{p_{i,j}^{\mathrm{I}\rightarrow\mathrm{T}}}$ +15: $\mathcal{L}_{\mathrm{HyCD}}^{\mathrm{T}\rightarrow \mathrm{I}}\gets \frac{1}{B}\sum_{i = 1}^{B}\sum_{j = 1}^{B}\hat{q}_{i,j}^{\mathrm{T}\rightarrow \mathrm{I}}\log \frac{\hat{q}_{i,j}^{\mathrm{T}\rightarrow \mathrm{I}}}{p_{i,j}^{\mathrm{T}\rightarrow \mathrm{I}}}$ +16: $\mathcal{L}_{\mathrm{HyCD}}\gets \frac{1}{2} (\mathcal{L}_{\mathrm{HyCD}}^{\mathrm{I}\to \mathrm{T}} + \mathcal{L}_{\mathrm{HyCD}}^{\mathrm{T}\to \mathrm{I}})$ +17: $\theta_{\mathrm{V}}\gets \theta_{\mathrm{V}} - \frac{\eta}{2}\nabla_{\theta_{\mathrm{V}}}\left(\mathcal{L}_{\mathrm{RaFA}} + \mathcal{L}_{\mathrm{HyCD}}\right)$ +18: $\theta_{\mathrm{T}}\gets \theta_{\mathrm{T}} - \frac{\eta}{2}\nabla_{\theta_{\mathrm{T}}}(\mathcal{L}_{\mathrm{RaFA}} + \mathcal{L}_{\mathrm{HyCD}})$ +19: end while + +leverage any fine-tuning techniques after post-pre-training by our method, and it helps to boost the baseline performance. + +# 3. Method + +We propose post-pre-training of CLIP in between pretraining and fine-tuning to mitigate the modality gap (Figure 1). To this end, we present CLIP-Refine (Figure 2), + +which is composed of random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). RaFA penalizes the model to induce both modal features into a single shared distribution by minimizing the gap between the paired features and a random reference vector. To reconcile learning new knowledge with retaining past knowledge, HyCD updates the models with knowledge distillation loss using fixed pretrained models, where the teacher outputs are mixed with ground-truth labels of image-text pairs. Algorithm 1 shows the overall procedure of CLIP-Refine. + +# 3.1. Problem Setting: Post-pre-training + +We consider a problem setting called post-pre-training, where we aim to improve the cross-modal alignment and generalization performance of pre-trained vision-language models. This setting allows to access a vision encoder $f_{\mathrm{V}}: \mathcal{X} \to \mathbb{R}^{d}$ and a text encoder $f_{\mathrm{T}}: \mathcal{T} \to \mathbb{R}^{d}$ with the parameter $\theta^{\mathrm{CLIP}} = \{\theta_{\mathrm{V}}^{\mathrm{CLIP}}, \theta_{\mathrm{T}}^{\mathrm{CLIP}}\}$ pre-trained by CLIP, where $\mathcal{X}$ and $\mathcal{T}$ are the image and text space. We optimize the parameters $\theta_{\mathrm{V}}$ of $f_{\mathrm{V}}$ and $\theta_{\mathrm{T}}$ of $f_{\mathrm{T}}$ on a post-pre-training image-text pair dataset $\mathcal{D} = \{(x_i, t_i)\}_{i=1}^N$ ; we basically assume that $\mathcal{D}$ is much smaller than the pre-training dataset and contains images and text captions in general domain. + +# 3.2. Objective Function + +In CLIP-Refine, we optimize $\theta_{\mathrm{V}}$ and $\theta_{\mathrm{T}}$ by the following objective function: + +$$ +\min _ {\theta_ {\mathrm {V}}, \theta_ {\mathrm {T}}} \mathcal {L} _ {\mathrm {R a F A}} \left(\theta_ {\mathrm {V}}, \theta_ {\mathrm {T}}\right) + \mathcal {L} _ {\mathrm {H y C D}} \left(\theta_ {\mathrm {V}}, \theta_ {\mathrm {T}}\right), \tag {1} +$$ + +where $\mathcal{L}_{\mathrm{RaFA}}$ is the random feature alignment loss and $\mathcal{L}_{\mathrm{HyCD}}$ is the hybrid contrastive-distillation loss. One can consider balancing $\mathcal{L}_{\mathrm{RaFA}}$ and $\mathcal{L}_{\mathrm{HyCD}}$ by hyperparameters, but we found that their equal contribution achieves the best performance (see Appendix). We describe the details of $\mathcal{L}_{\mathrm{RaFA}}$ and $\mathcal{L}_{\mathrm{HyCD}}$ in the following sections. + +# 3.3. Random Feature Alignment + +The primary goal in post-pre-training is to minimize the modality gap and encourage cross-modal alignment. To this end, the most straightforward approach is just minimizing the feature gap between the modalities. Let $z_{\mathrm{img}}^i = f_{\mathrm{V}}(x^i;\theta_{\mathrm{V}})$ and $z_{\mathrm{xt}}^i = f_{\mathrm{T}}(t^i;\theta_{\mathrm{T}})$ are the normalized feature vectors of an image-text pair $(x^{i},t^{i})$ . A naive loss for minimizing the modality gap can be defined by an $L_{2}$ distance form as + +$$ +\mathcal {L} _ {\text {a l i g n}} = \mathbb {E} _ {(x ^ {i}, t ^ {i}) \in \mathcal {D}} \| z _ {\text {i m g}} ^ {i} - z _ {\text {t x t}} ^ {i} \| _ {2} ^ {2}. \tag {2} +$$ + +However, minimizing this loss function degrades the generalization performance. This can be explained by the balance of alignment and uniformity [53], which are desirable properties of transferability in contrastive representation learning. Alignment is defined by the gap between positive (i.e., imagetext) pairs in the feature space, and uniformity is defined by + +equal separation among all features on the hypersphere. That is, minimizing Eq. (2) enhances the alignment of the positive pair but destroys the uniformity on the hypersphere by forcing changes to the feature distribution. In fact, Wang and Isola [53] have shown that using only the alignment loss in post-pre-training degrades the validation performance due to corrupt uniformity. + +To address this challenge, we introduce the idea of matching feature distributions instead of directly matching the paired features. In other words, we refine the multi-modal features of pre-trained CLIP models to follow a prior distribution $p(z)$ (e.g., standard Gaussian $\mathcal{N}(0, I)$ ) shared across the image and text modalities. Inspired by random feature regularization in single modal fine-tuning [59, 67], we formulate this idea as minimization between image/text feature vectors $(z_{\mathrm{img}}^i, z_{\mathrm{txt}}^i)$ and a random reference vector $z_{\mathrm{ref}}^i$ sampled from the shared prior $p(z)$ : + +$$ +\mathcal {L} _ {\mathrm {R a F A}} = \mathbb {E} _ {(x ^ {i}, t ^ {i}) \in \mathcal {D}} \frac {1}{2} (\| z _ {\mathrm {i m g}} ^ {i} - z _ {\mathrm {r e f}} ^ {i} \| _ {2} ^ {2} + \| z _ {\mathrm {t x t}} ^ {i} - z _ {\mathrm {r e f}} ^ {i} \| _ {2} ^ {2}). \tag {3} +$$ + +Through this loss function, we can expect that (i) the modality gap is indirectly minimized via $z_{\mathrm{ref}}^i$ shared with each image-text pair, (ii) the feature vectors follow the modality shared prior $p(z)$ gradually [59], (iii) the randomness of $z_{\mathrm{ref}}^i$ helps to avoid the models from overfitting to $\mathcal{D}$ [67] and reform the models' capacity [49]. Furthermore, we found that $\mathcal{L}_{\mathrm{RaFA}}$ can improve the uniformity from baselines (Table 3), meaning that the learned representations transfer well to downstream tasks. Throughout this paper, we use standard Gaussian $\mathcal{N}(0,I)$ as $p(z)$ by default; we discuss the effect of prior choice in Sec. 4.4.1. + +# 3.4. Hybrid Contrastive-Distillation + +Another important goal of post-pre-training is to retain the past knowledge in pre-trained CLIP models. To this end, our basic strategy is to penalize the post-pre-training models with knowledge distillation loss [19] where the frozen pre-trained CLIP model performs as the teacher. We define the minibatch-wise knowledge distillation loss with KL-divergence according to the formulation of CLIP-KD [60] as + +$$ +\mathcal {L} _ {\mathrm {K D}} = \frac {1}{2} \left(\mathcal {L} _ {\mathrm {K D}} ^ {\mathrm {I} \rightarrow \mathrm {T}} + \mathcal {L} _ {\mathrm {K D}} ^ {\mathrm {T} \rightarrow \mathrm {I}}\right), \tag {4} +$$ + +$$ +\mathcal {L} _ {\mathrm {K D}} ^ {\mathrm {I} \rightarrow \mathrm {T}} = \frac {1}{B} \sum_ {i = 1} ^ {B} \sum_ {j = 1} ^ {B} q _ {i, j} ^ {\mathrm {I} \rightarrow \mathrm {T}} \log \frac {q _ {i , j} ^ {\mathrm {I} \rightarrow \mathrm {T}}}{p _ {i , j} ^ {\mathrm {I} \rightarrow \mathrm {T}}}, \tag {5} +$$ + +$$ +p _ {i, j} ^ {\mathrm {I} \rightarrow \mathrm {T}} = \frac {\exp \left(z _ {\mathrm {i m g}} ^ {i} \cdot z _ {\mathrm {t x t}} ^ {j} / \tau\right)}{\sum_ {k = 1} ^ {B} \exp \left(z _ {\mathrm {i m g}} ^ {i} \cdot z _ {\mathrm {t x t}} ^ {k} / \tau\right)}, \tag {6} +$$ + +$$ +q _ {i, j} ^ {\mathrm {I} \rightarrow \mathrm {T}} = \frac {\exp \left(z _ {\mathrm {i m g}} ^ {i , \mathrm {p}} \cdot z _ {\mathrm {t x t}} ^ {j , \mathrm {p}} / \tau\right)}{\sum_ {k = 1} ^ {B} \exp \left(z _ {\mathrm {i m g}} ^ {i , \mathrm {p}} \cdot z _ {\mathrm {t x t}} ^ {k , \mathrm {p}} / \tau\right)}, \tag {7} +$$ + +where $\tau$ is the temperature parameter, $(z_{\mathrm{img}}^{i,\mathrm{p}},z_{\mathrm{xt}}^{i,\mathrm{p}}) = (f_{\mathrm{V}}(x^{i};\theta_{\mathrm{V}}^{\mathrm{CLIP}}),f_{\mathrm{T}}(t^{i};\theta_{\mathrm{T}}^{\mathrm{CLIP}}))$ are the feature vectors generated by the pre-trained image/text encoders, and $\mathcal{L}_{\mathrm{KD}}^{\mathrm{T}\rightarrow \mathrm{I}}$ is defined by swapping the order of image features and text + +Table 1. Zero-shot classification performance on 12 datasets with CLIP ViT-B/32. + +
MethodAirBirdCalCarDTDEuroFloFoodINPetSUNUCFAvg.
Pre-trained18.8149.3723.9860.0740.9038.2059.7580.2159.0481.7960.8559.9052.74
Contrastive13.9840.0723.7747.6336.2234.6846.2870.3152.9670.1354.7858.2645.75
m2-mix [36]14.6440.7623.4949.8636.4934.6547.4271.2753.5870.3256.5158.7646.48
Self-KD18.9651.5223.9460.0441.1738.3659.5580.0259.0681.8560.6860.1652.94
HyCD18.7250.5023.9859.8340.4840.9058.0680.1959.7181.4562.0461.7053.13
HyCD + L_align13.2934.2023.7343.7035.0039.2746.2773.3152.5772.4255.0158.5545.61
CLIP-Refine (Ours)20.7752.7224.7960.6341.4941.4462.6481.8760.9383.5763.6361.8554.69
+ +features in Eq. (6) and (7). Although minimizing Eq. (4) retains the past knowledge, we found that it interferes with RaFA because it strongly enchains the models' parameters to the pre-trained ones. + +To reconcile RaFA and the retention of past knowledge, we present a technique that blends the ground-truth label of an image-text pair with the teacher output. Concretely, we modify Eq. (7) by alpha-blending of an indicator function $\mathbb{I}_{i = j}$ , which returns one if $i = j$ otherwise zero, and the teacher output signal: + +$$ +\hat {q} _ {i, j} ^ {\mathrm {I} \rightarrow \mathrm {T}} = \alpha \mathbb {I} _ {i = j} + (1 - \alpha) q _ {i, j} ^ {\mathrm {I} \rightarrow \mathrm {T}}, \tag {8} +$$ + +where $\alpha$ is a hyperparameter to balance acquiring new knowledge and retaining past knowledge; we use $\alpha = 0.5$ by default and discuss the effect of $\alpha$ in the Appendix. With this modification, the post-pre-training models can further enhance the cross-modal alignment by learning the correct image-text pair (i.e., $i = j$ ) with reference to the relative knowledge contained in the $i \neq j$ coordinates of the teacher's outputs. By using $\hat{q}_{i,j}^{\mathrm{I}\rightarrow \mathrm{T}}$ and $\hat{q}_{i,j}^{\mathrm{T}\rightarrow \mathrm{I}}$ , we compute the HyCD loss function as follows: + +$$ +\mathcal {L} _ {\mathrm {H y C D}} = \frac {1}{2} \left(\mathcal {L} _ {\mathrm {H y C D}} ^ {\mathrm {I} \rightarrow \mathrm {T}} + \mathcal {L} _ {\mathrm {H y C D}} ^ {\mathrm {T} \rightarrow \mathrm {I}}\right), \tag {9} +$$ + +$$ +\mathcal {L} _ {\mathrm {H y C D}} ^ {\mathrm {I} \rightarrow \mathrm {T}} = \frac {1}{B} \sum_ {i = 1} ^ {B} \sum_ {j = 1} ^ {B} \hat {q} _ {i, j} ^ {\mathrm {I} \rightarrow \mathrm {T}} \log \frac {\hat {q} _ {i , j} ^ {\mathrm {I} \rightarrow \mathrm {T}}}{p _ {i , j} ^ {\mathrm {I} \rightarrow \mathrm {T}}}. \tag {10} +$$ + +While $\mathcal{L}_{\mathrm{HyCD}}$ solely improves the generalization performance of the pre-trained CLIP models, combining it with $\mathcal{L}_{\mathrm{RaFA}}$ achieves larger improvements by minimizing the modality gaps without losing the past knowledge. + +# 4. Experiments + +We evaluate CLIP-Refine on 12 classification datasets and 2 cross-modal retrieval datasets by using multiple pre-trained models. We also conduct qualitative and quantitative analyses to assess the feature space through modality gap, alignment, uniformity metrics, and PCA visualization. + +# 4.1. Settings + +Baselines. We compare our CLIP-Refine with the following baselines of post-pre-training. Pre-trained: predicting with pre-trained weights without additional training. Contrastive: post-pre-training with the contrastive + +loss [21, 40]. $m^2$ -mix [36]: post-pre-training by the modified contrastive loss with multi-modal mixup; we use this fine-tuning method as a baseline because it can be used for refining the feature spaces as suggested in the original paper [36]. Self-KD: using Eq. (4) for post-pre-training. HyCD: using only Eq. (9) for training. HyCD + L align: combining Eq. (9) and Eq. (2). + +Post-pre-training Datasets. We used COCO Caption [29] as the default post-pre-training dataset. We also used Flickr8K/30K [42], CC3M [47], and CC12M [3] to confirm the dataset size effect in post-pre-training. + +Test Datasets. We used 12 image classification datasets containing various image domains: Aircraft (Air) [34], Bird [56], Caltech-101 (Cal) [12] Car [24], DTD [9], EuroSAT (Euro) [18], Flower (Flo) [35], Food [2], ImageNet (IN) [45], Pet [38], SUN397 [57], and UCF-101 [50]. We also evaluated our method on COCO2017-Val and Flickr8K/30K for the image and text retrieval tasks; we denote text-to-image retrieval as $\mathrm{T} \rightarrow \mathrm{I}$ and image-to-text retrieval as $\mathrm{I} \rightarrow \mathrm{T}$ . We adopt these datasets since they are often used to evaluate the zero-shot transfer performance of CLIP models [68]. We evaluated models on the test sets except for ImageNet, and the official validation set for ImageNet. + +Pre-trained Models. By default, we used CLIP-ViT-B/32 [40] with the pre-trained weights downloaded from the OpenAI's official repositories. We also tested other pretrained models, including OpenCLIP-ViT-H/14, OpenCLIP-ViT-bigG/14, SigLIP [65] and DFN [11] with the pre-trained weights on the OpenCLIP [8] repository. + +Post-pre-training. In all settings, we trained models by the AdamW [32] optimizer with the learning rate of $1.0 \times 10^{-6}$ for one epoch; we determined the learning rate by the zero-shot classification validation using the subset of the ImageNet training set constructed by uniform sampling. We used mini-batch sizes of 512 as the default. For CLIP-Refine, we used $\alpha = 0.5$ . The input samples were preprocessed using the default image transformations provided by each pre-trained model. We implemented the training and evaluation with PyTorch-1.13. We ran the experiments on a 24-core Intel Xeon CPU with a single NVIDIA A100 GPU with 80GB VRAM. + +Table 2. Zero-shot retrieval performance on COCO2017-Val and Flickr8K/30K with CLIP ViT-B/32. + +
MethodCOCO2017-ValFlickr8KFlickr30K
T→II→TR@1T→IR@1I→TR@1T→IR@1I→T
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@5R@10R@1R@5R@10R@1R@5R@10
Pre-trained30.5654.9265.2633.2659.1068.7830.0653.8363.2932.4255.5765.2621.0640.9650.5324.5244.5754.14
Contrastive34.8861.5072.1031.8656.8067.4834.9459.4169.3628.8351.5561.0725.6947.3256.9318.4236.2945.37
m2-mix [36]36.2862.1873.0832.9257.9667.5635.0459.6669.6628.6450.6560.7725.9947.1857.1919.8837.6346.51
Self-KD31.0455.5865.9035.3659.0069.7229.9853.8363.5532.4755.1165.1021.9241.5150.7524.9444.8254.27
HyCD36.0462.2873.1437.8862.5472.8435.7260.3470.1434.6957.1967.2227.1149.1258.7827.1747.6656.94
HyCD + L align33.9261.1872.0635.1461.7472.0631.7355.6365.9032.5856.9466.9923.7844.7354.4723.5845.0755.23
CLIP-Refine (Ours)37.6463.5474.4238.7865.0475.1236.1161.2970.5138.2861.5570.9426.6549.7059.0330.9951.5461.31
+ +Evaluation Metrics. For zero-shot transfer performance, we report top-1 accuracy in classification tasks and recall@ $k$ scores in cross-modal retrieval tasks; recall@ $k$ indicates the proportion of correct answers in the retrieved top- $k$ candidates for all test input samples. We denote recall@ $k$ as $R@k$ . To quantitatively evaluate the post-pre-trained features, we measure modality gap, alignment, and uniformity scores. Modality gap [28] assesses a cluster-wise gap between image and text features: + +$$ +\text {M o d a l i t y} \mathrm {G a p} := \| \bar {f} _ {\mathrm {V}} - \bar {f} _ {\mathrm {T}} \| _ {2} ^ {2}, \tag {11} +$$ + +where $\bar{f}$ is a mean feature vector for all test samples of each modality; lower is better. Alignment [53] is defined by + +$$ +\text {A l i g n m e n t} := \frac {1}{N _ {\text {t e s t}}} \sum_ {i = 1} ^ {N _ {\text {t e s t}}} \| f _ {\mathrm {V}} \left(x ^ {i}\right) - f _ {\mathrm {T}} \left(t ^ {i}\right) \| _ {2} ^ {2}, \tag {12} +$$ + +representing how well the positive pair is aligned in the feature space; lower is better. Uniformity [53] evaluates how uniformly distributed the feature vectors are on the hypersphere, which is defined by radial basis function (RBF) kernel as + +$$ +\text {U n i f o r m i t y} := \frac {1}{2 N _ {\text {t e s t}}} \sum_ {f _ {1}, f _ {2} \in F} \exp (- 2 \| f _ {1} - f _ {2} \| _ {2} ^ {2}), \tag {13} +$$ + +where $F = \{f_{\mathrm{V}}(x^i)\}_{i=1}^{N_{\mathrm{test}}} \cup \{f_{\mathrm{T}}(t^i)\}_{i=1}^{N_{\mathrm{test}}}$ ; lower is better. We report the average score for each metric by running experiments three times. + +# 4.2. Zero-shot Transfer Evaluation + +We first demonstrate the zero-shot performance of our CLIP-Refine. Tables 1 and 2 list the zero-shot test performance on classification and cross-modal retrieval tasks after post-pretraining. For zero-shot classification tasks, post-pre-training by the contrastive loss (i.e., Contrastive and $m^2$ -mix) largely degraded the top-1 accuracy performance for all datasets. This can be caused by catastrophic forgetting since Self-KD did not show performance degradation. One cause of catastrophic forgetting is a smaller batchsize than at pre-training; the OpenAI's CLIP uses a batchsize of 32,768 with hundreds of GPUs [40], whereas our setting uses ones of 512 with a single GPU. Since a smaller batch contains a limited number + +of negative samples, contrastive learning tends to overfit and catastrophically forget past knowledge. We discuss more details about the effect of batchsize in Sec. 4.4.4. In contrast, our CLIP-Refine impressively improved the zero-shot performance for all datasets. The ablation studies of HyCD show that the performance gain comes from the combination of HyCD with RaFA. Notably, $\mathrm{HyCD} + \mathcal{L}_{\mathrm{align}}$ resulted in the worst performance among the baselines, suggesting that naively minimizing the gap between image and text features leads to the degradations of generalization performance. + +For zero-shot retrieval, our CLIP-Refine achieved the best performance with a larger margin than that in classification. Interestingly, in contrast to classification tasks, we found that Contrastive and $m^2$ -mix improved the pre-trained models in the $\mathrm{T} \rightarrow \mathrm{I}$ retrieval tasks while the $\mathrm{I} \rightarrow \mathrm{T}$ performance was degraded. This indicates that the image encoders are more prone to overfitting than the text encoders. This can be explained by the difference between image and text in the data spaces; images are represented by continuous-valued pixels and thus have a high degree of freedom, while texts are represented by discrete tokens in vocabulary and thus have a limited degree of freedom. Therefore, the image encoder tends to overfit to a limited number of negative samples in small batches, and the text encoder can learn new knowledge from them while keeping the past knowledge. Since the zero-shot classification is a $\mathrm{I} \rightarrow \mathrm{T}$ retrieval task, this can also be a reason for the performance degradation in classification. In this perspective, our CLIP-Refine successfully prevents the overfitting in $\mathrm{I} \rightarrow \mathrm{T}$ and greatly improves the performance in both $\mathrm{I} \rightarrow \mathrm{T}$ and $\mathrm{T} \rightarrow \mathrm{I}$ . + +# 4.3. Analysis of Feature Space + +In this section, we aim to answer our fundamental research question: how much can we alleviate the modality gap by post-pre-training with CLIP-Refine? For quantitative evaluations, we summarize the modality gap, alignment, and uniformity scores in Table 3. HyCD $+\mathcal{L}_{\mathrm{align}}$ achieved the best modality gap and alignment scores. However, it significantly degraded the uniformity score, indicating that directly minimizing $\mathcal{L}_{\mathrm{align}}$ can destroy feature space well constructed by the CLIP training. In contrast, our CLIP-Refine consistently improved all of the scores. This suggests that it is important to minimize not only the gap between image and text features + +![](images/175c129fbc52c2a0c63eed8aa993b0a054828eaf1477152aaa892f5bd87f894a.jpg) +(a) Pre-trained + +![](images/00546a37b2fd7af3859b1c13b57b7b7d1f25749cfa11b66e0d0dadd2f512e183.jpg) +(b) Contrastive +Figure 3. PCA visualization of multi-modal feature spaces of trained models (Flickr8K, ViT-B/32). CLIP-Refine reduces the modality gap while keeping the form of feature distributions for each modality. + +![](images/2ae6438814ad3ee5a2154ec0f2b995be71ee0d5f396709cf53a1af220f07b791.jpg) +(c) $\mathrm{HyCD} + \mathcal{L}_{\mathrm{align}}$ + +![](images/d0a2a744612d987be37aa739e461e7625d8b0f996e00a866734476cbca987f85.jpg) +(d) CLIP-Refine (Ours) + +Table 3. Quantitative evaluation of features (Flickr8K, ViT-B/32). + +
MethodModality Gap ×10-3(↓)Alignment (↓)Uniformity (↓)
Pre-trained1.33451.37240.0895
Contrastive2.17141.43000.2136
m2-mix [36]2.09111.41300.2019
Self-KD1.27541.37740.0901
HyCD1.38231.35780.0971
HyCD+L align0.15900.74130.1172
CLIP-Refine (Ours)0.79341.28490.0495
+ +but also the uniformity for improving the generalization performance of pre-trained CLIP models, which is consistent with the findings in [53]. In such a sense, since CLIP-Refine penalizes the image and text features to follow the shared reference distribution $p(z) = \mathcal{N}(0,I)$ , where the samples tend to be uniformly distributed on the hypersphere [1], it can naturally achieve a good balance of alignment and uniformity, and reasonably helps to reduce the modality gap. + +As a qualitative evaluation, Figure 3 illustrates the PCA visualization of post-pre-trained features for each method. We can see that the contrastive model makes the modality gap rather large. $\mathrm{HyCD} + \mathcal{L}_{\mathrm{align}}$ succeeds in gathering the image and text features into the same cluster but distorts the feature distribution for each modality. Meanwhile, CLIP-Refine certainly reduces the modality gap while keeping the shape of the pre-trained features. Although the modality gap is not completely zero, it is natural because there remains a constant value in the lower bound of the modality gap [39]. + +# 4.4. Detailed Analysis + +# 4.4.1. Prior Distributions + +Here, we analyze CLIP-Refine by varying the prior distribution $p(z)$ . According to the prior works [59, 67], we tried standard Gaussian $\mathcal{N}(0,I)$ , uniform distribution $U(0,1)$ , Gaussian distributions using pre-trained text feature statistics $\mathcal{N}(\mu_{\mathrm{txt}}^{\mathrm{P}},(\sigma_{\mathrm{txt}}^{\mathrm{P}})^2)$ , image feature statistics $\mathcal{N}(\mu_{\mathrm{img}}^{\mathrm{P}},(\sigma_{\mathrm{img}}^{\mathrm{P}})^2)$ , statistics of all image and text features $\mathcal{N}(\mu_{\mathrm{all}}^{\mathrm{P}},(\sigma_{\mathrm{all}}^{\mathrm{P}})^2)$ , where $\mu^{\mathrm{P}},\sigma^{\mathrm{P}}$ is mean and variance of image or/and text feature vectors computed on a pre-trained CLIP model. We also tested variations of $\mathcal{N}(0,\beta I)$ by varying $\beta$ in $\{0,0.01,0.1,10,100\}$ ; $\beta = 0$ means RaFA without randomness, i.e., $\frac{1}{2} (\| z_{\mathrm{img}}^{i}\|_{2}^{2} + \| z_{\mathrm{txt}}^{i}\|_{2}^{2})$ . + +Table 4. Evaluation of CLIP-Refine varying prior distributions in $\mathcal{L}_{\mathrm{RaFA}}$ . ZS Cls. is the zero-shot classification accuracy averaged on 12 datasets. + +
PriorZS Cls.AlignmentUniformity
Pretrained52.741.37240.0895
U(0,1)52.911.30810.0661
N(μpimg, (σpimg)2)52.931.31420.0916
N(μptxt, (σptxt)2)53.401.33040.0873
N(μpall, (σpall)2)53.291.33610.0888
N(0,0I)53.791.30520.0554
N(0,0.01I)54.191.30790.0544
N(0,0.1I)54.321.30210.0515
N(0,I)54.691.28490.0495
N(0,10I)54.231.30850.0537
N(0,100I)53.591.34410.0663
+ +Table 4 shows the zero-shot classification performance, alignment, and uniformity scores when varying $p(z)$ . Using the uniform distribution $U(0,1)$ did not show large improvements on the zero-shot performance and uniformity score. This may be because random reference vectors from $U(0,1)$ do not necessarily concentrate on the hypersphere in contrast to $\mathcal{N}(0,I)$ , possibly causing the collapse of uniformity by $\mathcal{L}_{\mathrm{RaFA}}$ . In this perspective, we recommend using $\mathcal{N}(0,I)$ or families of Gaussian distributions. Interestingly, among the Gaussian distributions with the pre-trained statistics, the prior with the text feature, i.e., $\mathcal{N}(\mu_{\mathrm{txt}}^{\mathrm{p}}(\sigma_{\mathrm{txt}}^{\mathrm{p}})^2)$ largely improved the zero-shot performance. This indicates that aligning image features with text features is effective in zero-shot classification tasks and also evidences the performance improvements in the $\mathrm{I} \rightarrow \mathrm{T}$ tasks observed in Sec. 4.2. More importantly, the ablation studies of the variance magnitude $\beta$ in the zero-mean Gaussian distributions emphasize the importance of randomness in $\mathcal{L}_{\mathrm{RaFA}}$ . + +# 4.4.2. Post-pre-training Datasets + +We evaluate the effects of the dataset choice in post-pretraining. Table 5 shows the zero-shot classification/retrieval performance of CLIP-Refine with Flickr8K/30K, COCO Caption, CC3M, and CC12M as post-pre-training datasets. We confirm that most of the CLIP-Refine models outper + +Table 5. Zero-shot classification (12 datasets) and retrieval (Flicker30K) performance of CLIP-Refine with various post-pre-training datasets with ViT-B/32. + +
Post-pre-training Dataset# of SamplesZero-shot Cls. Acc.Zero-shot Ret. R@5
Avg. of 12 datasetsT→II→T
None (Pre-trained)052.7440.9644.57
Flickr8K8K53.0445.5746.80
Flickr30K30K53.7448.6649.25
COCO Caption118K54.6949.7051.54
CC3M3M53.8049.5650.24
CC12M12M52.4645.3547.79
Filtered CC3M1M55.5052.4553.00
Filtered CC12M4M60.1254.9757.88
+ +Table 6. Zero-shot classification (12 datasets) and retrieval (Flickr30K) performance with various pre-trained CLIP models. + +
ModelZero-shot Cls. Acc.Zero-shot Ret. R@5 (T→I/I→T)
Pre-trainedContrastiveCLIP-RefinePre-trainedContrastiveCLIP-Refine
ViT-B/3252.7445.7654.6940.96/44.5747.32/36.2949.70/51.54
ViT-L/1462.9661.9063.8849.13/52.7751.07/51.0058.38/58.41
OpenCLIP ViT-H/1468.5257.6169.0565.84/66.0257.47/43.4367.82/68.48
OpenCLIP ViT-G/1471.2857.4472.5968.73/68.3866.35/42.3670.94/69.76
SigLip ViT-SO400M/1472.5765.6373.6172.77/72.8364.63/62.9776.54/76.49
DFN ViT-H/1477.3076.3978.2971.93/70.8972.81/70.0976.39/75.44
+ +formed the pre-trained baselines. Notably, larger datasets do not always achieve high performance; CC3M and CC12M are inferior to COCO Caption. This suggests that the quality of image-text pairs in a dataset is an important factor in post-pre-training. Since CC3M and CC12M contain mismatched and noisy image-text pairs [17], COCO Caption that has higher-quality text captions can be superior in terms of the modality alignment of pre-trained CLIP models. In fact, when we filtered out low-quality image-text pairs by CLIP-Score-based filtering of DataComp [13], the models achieved significant performance improvements. This indicates that the caption quality is crucial and our method is scalable if we have high-quality image-text pairs. + +# 4.4.3. Pre-trained Models + +We investigate the generalizability of CLIP-Refine to pretrained CLIP models. Table 6 shows the zero-shot performance with 6 pre-trained models. We observed that our CLIP-Refine stably improved the pre-trained baseline. This suggests that the bigger and stronger models still have the modality gap, and modifying the gap by CLIP-Refine helps improve the generalization performance for various pretrained models. + +# 4.4.4. Mini-Batch Sizes + +We provide the analysis of varying mini-batch sizes in post-pre-training. We tried the mini-batch sizes in $\{32, 64, 128, 256, 512, 1024, 2048\}$ . We used gradient accumulation by following the OpenCLIP [8] implementation for the mini-batch sizes of 1024 and 2048. Figure 4 shows the zero-shot performance versus mini-batch sizes. The contrastive loss gradually improved the accuracy according to the mini-batch size. However, the accuracy curve implies that it is difficult to exceed the pre-trained baseline with + +![](images/065f34356cbd634a4a0697963f5ad2f05b36a7e0a5d810dc0b7196fe5c3925bb.jpg) +Figure 4. Zero-shot classification accuracy averaged on 12 datasets when varying mini-batch sizes in post-pre-training. + +Table 7. Fine-tuning top-1 accuracy on the ImageNet validation set. + +
ModelPre-trainedContrastiveCLIP-Refine
ViT-B/3273.4069.4674.14
+ +feasible mini-batch sizes. This is probably due to the large discrepancy from the mini-batch size in the pre-training (i.e., 32,768), which causes the model to forget knowledge in the predictions for negative pairs in a mini-batch. In contrast, CLIP-Refine always outperformed the contrastive loss and succeeded in improving the pre-trained baseline with feasible mini-batch sizes. This suggests that CLIP-Refine's HyCD prevents knowledge forgetting even for small mini-batch sizes with few negative pairs and is able to modify the feature space with new knowledge. + +# 4.4.5. Effects on Fine-tuning + +We show the transferability of CLIP-Refine to fine-tuning. We fine-tuned the linear classification head on ImageNet while freezing post-pre-trained models. Tabel 7 shows the results. CLIP-Refine outperformed the pre-trained and contrastive loss baselines. This emphasizes that our method can generate useful representation even in fine-tuning. + +# 5. Conclusion + +This paper presented CLIP-Refine, a post-pre-training method for pre-trained CLIP models to align the modality gap between the image and text features. CLIP-Refine addresses the modality gap by penalizing the multi-modal features to follow a shared prior distribution by minimizing the distance to the random reference vectors sampled from the prior. To maintain the past knowledge in the CLIP models and promote the feature alignment simultaneously, CLIP-Refine also trains the model with knowledge distillation loss using hybrid soft labels composed of ground-truth image-text pair labels and outputs from the pre-trained CLIP model. Through extensive experiments, we show that CLIP-Refine can improve the zero-shot performance of the pre-trained CLIP by addressing the modality gap and enhancing uniformity. We believe that our work not only provides a practical method but also opens up a new research field where we refine the pre-trained CLIP models by post-pre-training with much smaller computational costs than that of pre-training. + +# References + +[1] Christopher M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag, Berlin, Heidelberg, 2006. 7 +[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5 +[3] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 2, 5 +[4] Lianggangxu Chen, Xuejiao Wang, Jiale Lu, Shaohui Lin, Changbo Wang, and Gaoqi He. Clip-driven open-vocabulary 3d scene graph generation via cross-modality contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[5] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yu Hou, Yuenan and Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[6] Yuxiao Chen, Jianbo Yuan, Yu Tian, Shijie Geng, Xinyu Li, Ding Zhou, Dimitris N Metaxas, and Hongxia Yang. Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15095-15104, 2023. 1, 2 +[7] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 2, 5, 8 +[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2014. 5 +[10] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2 +[11] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. In International Conference on Learning Representations, 2024. 5 + +[12] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In Conference on CVPR Workshop, 2004. 5 +[13] Gadre et al. Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 2023. 8 +[14] Yunhao Ge, Jie Ren, Andrew Gallagher, Yuxiao Wang, Ming-Hsuan Yang, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11093-11101, 2023. 1 +[15] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. In Advances in Neural Information Processing Systems, 2022. 2 +[16] Haochen Han, Qinghua Zheng, Guang Dai, Minnan Luo, and Jingdong Wang. Learning to rematch mismatched pairs for robust cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1 +[17] Haochen Han, Qinghua Zheng, Guang Dai, Minnan Luo, and Jingdong Wang. Learning to rematch mismatched pairs for robust cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26679-26688, 2024. 8 +[18] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5 +[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 4 +[20] Ahmet Iscen, Mathilde Caron, Alireza Fathi, and Cordelia Schmid. Retrieval-enhanced contrastive vision-text models. In International Conference on Learning Representations, 2024. 1 +[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, 2021. 1, 2, 5 +[22] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision. Springer, 2022. 3 +[23] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multimodal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 1, 3 +[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition, Sydney, Australia, 2013. 5 + +[25] Zhengfeng Lai, Noranart Vesdapunt, Ning Zhou, Jun Wu, Cong Phuoc Huynh, Xuelu Li, Kah Kuen Fu, and Chen-nee Chuah. Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2 +[26] Janghyeon Lee, Jongsuk Kim, Hyounguk Shon, Bumsoo Kim, Seung Hwan Kim, Honglak Lee, and Junmo Kim. Uniclip: Unified framework for contrastive language-image pre-training. Advances in Neural Information Processing Systems, 2022. 1, 2 +[27] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In International Conference on Learning Representations, 2021. 1, 2 +[28] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 2022. 1, 2, 6 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision. Springer, 2014. 2, 3, 5 +[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1 +[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1 +[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5 +[33] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1 +[34] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. arXiv, 2013. 3, 5 +[35] M-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5 +[36] Changdae Oh, Junhyuk So, Hoyoon Byun, YongTaek Lim, Minchul Shin, Jong-June Jeon, and Kyungwoo Song. Geodesic multi-modal mixup for robust fine-tuning. In Advances in Neural Information Processing Systems, 2023. 3, 5, 6, 7 +[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2 +[38] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 5 + +[39] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In Advances in Neural Information Processing Systems, 2024. 1, 2, 7 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 1, 2, 3, 5, 6 +[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[42] Cyrus Rashtchian, Peter Young, Micah Hodosh, and Julia Hockenmaier. Collecting image annotations using amazon's mechanical turk. In Proceedings of the NAACL HLT 2010 workshop on creating speech and language data with Amazon's Mechanical Turk, pages 139-147, 2010. 3, 5 +[43] Arijit Ray, Filip Radenovic, Abhimanyu Dubey, Bryan Plummer, Ranjay Krishna, and Kate Saenko. Cola: A benchmark for compositional text-to-image retrieval. In Advances in Neural Information Processing Systems, 2023. 1 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1 +[45] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3), 2015. 5 +[46] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2 +[47] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, 2018. 5 +[48] Sheng Shen, Shijia Yang, Tianjun Zhang, Bohan Zhai, Joseph E Gonzalez, Kurt Keutzer, and Trevor Darrell. Multitask vision-language prompt tuning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 3 +[49] Sicong Shen, Yang Zhou, Bingzheng Wei, Eric I Chang, Yan Xu, et al. Tuning stable rank shrinkage: Aiming at the overlooked structural risk in fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4 +[50] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5 +[51] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation + +of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3835-3844, 2022. 2 +[52] Jingyun Wang and Guoliang Kang. Learn to rectify the bias of clip for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4102-4112, 2024. 2 +[53] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 2, 3, 4, 6, 7 +[54] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2022. 2 +[55] Zhengbo Wang, Jian Liang, Ran He, Nan Xu, Zilei Wang, and Tieniu Tan. Improving zero-shot generalization for clip with synthesized prompts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1 +[56] P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD Birds 200. Technical report, California Institute of Technology, 2010. 5 +[57] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5 +[58] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2021. 2 +[59] Shin'ya Yamaguchi, Sekitoshi Kanai, Kazuki Adachi, and Daiki Chijiwa. Adaptive random feature regularization on fine-tuning deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4, 7 +[60] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. Clipkd: An empirical study of clip model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15952-15962, 2024. 4 +[61] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19163-19173, 2022. 2 +[62] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 1, 3 +[63] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and + +Chunjing Xu. FILIP: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022. 2 +[64] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. In European Conference on Computer Vision. Springer, 2022. 1, 2 +[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 5 +[66] Yuhui Zhang, Jeff Z. HaoChen, Shih-Cheng Huang, Kuan-Chieh Wang, James Zou, and Serena Yeung. Diagnosing and rectifying vision models using language. In The Eleventh International Conference on Learning Representations, 2023. 1 +[67] Yang Zhong and Atsuto Maki. Regularizing cnn transfer learning with randomised regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13637-13646, 2020. 4, 7 +[68] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 1, 3, 5 +[69] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 3 \ No newline at end of file diff --git a/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/images.zip b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..695e795b023bb0b20c6a3b4b0a93322becb24d7c --- /dev/null +++ b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:281b7e5154d0f462b0e3123044736c1bc3bd41c8d4edf9181ac40879207c8329 +size 458868 diff --git a/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/layout.json b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d07ee784ccf90b698d1cc37bb61caf90a8536fd7 --- /dev/null +++ b/2025/Post-pre-training for Modality Alignment in Vision-Language Foundation Models/layout.json @@ -0,0 +1,10656 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "text", + "content": "Post-pre-training for Modality Alignment in Vision-Language Foundation Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 143, + 178, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 143, + 178, + 172 + ], + "spans": [ + { + "bbox": [ + 63, + 143, + 178, + 172 + ], + "type": "text", + "content": "Shin'ya Yamaguchi* NTT, Kyoto University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 144, + 259, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 144, + 259, + 171 + ], + "spans": [ + { + "bbox": [ + 198, + 144, + 259, + 171 + ], + "type": "text", + "content": "Dewei Feng \nMIT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 280, + 144, + 359, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 144, + 359, + 171 + ], + "spans": [ + { + "bbox": [ + 280, + 144, + 359, + 171 + ], + "type": "text", + "content": "Sekitoshi Kanai NTT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 381, + 144, + 455, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 144, + 455, + 171 + ], + "spans": [ + { + "bbox": [ + 381, + 144, + 455, + 171 + ], + "type": "text", + "content": "Kazuki Adachi NTT" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 477, + 144, + 547, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 477, + 144, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 477, + 144, + 547, + 171 + ], + "type": "text", + "content": "Daiki Chijiwa NTT" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 238, + 297, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 297, + 561 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 297, + 561 + ], + "type": "text", + "content": "Contrastive language image pre-training (CLIP) is an essential component of building modern vision-language foundation models. While CLIP demonstrates remarkable zero-shot performance on downstream tasks, the multi-modal feature spaces still suffer from a modality gap, which is a gap between image and text feature clusters and limits downstream task performance. Although existing works attempt to address the modality gap by modifying pre-training or fine-tuning, they struggle with heavy training costs with large datasets or degradations of zero-shot performance. This paper presents CLIP-Refine, a post-pre-training method for CLIP models at a phase between pre-training and fine-tuning. CLIP-Refine aims to align the feature space with 1 epoch training on small image-text datasets without zero-shot performance degradations. To this end, we introduce two techniques: random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). RaFA aligns the image and text features to follow a shared prior distribution by minimizing the distance to random reference vectors sampled from the prior. HyCD updates the model with hybrid soft labels generated by combining ground-truth image-text pair labels and outputs from the pre-trained CLIP model. This contributes to achieving both maintaining the past knowledge and learning new knowledge to align features. Our extensive experiments with multiple classification and retrieval tasks show that CLIP-Refine succeeds in mitigating the modality gap and improving the zero-shot performance1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 135, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 135, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 135, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 603, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 296, + 687 + ], + "type": "text", + "content": "Contrastive language image pre-training (CLIP, [21, 40]) is a standard method to build modern vision-language foundation models. CLIP enables models to learn multimodal representations to map images and texts into a shared feature space with contrastive loss on large-scale image-text pair datasets. Since the pre-trained CLIP models provide a cross-modal understanding of input image/text data in various domains," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 214, + 553, + 259 + ], + "blocks": [ + { + "bbox": [ + 320, + 214, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 320, + 214, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 320, + 214, + 553, + 259 + ], + "type": "image", + "image_path": "9d73534414c3d8d7094939027bbfc8aae79a0a8f314d80db69639ba5512a8d68.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 268, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 313 + ], + "type": "text", + "content": "Figure 1. Scope of our work: post-pre-training to align the modality gap in pre-trained CLIP models. We aim to address the modality gap and enhance the generalization performance of pre-trained CLIP models through lightweight training." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 325, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 373 + ], + "type": "text", + "content": "they are widely used as a foundation of many applications, including zero-shot classification [14, 40, 55], cross-modal retrieval [16, 20], text-to-image generation [41, 44], and visual question answering [30, 31]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 376, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 376, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 312, + 376, + 556, + 532 + ], + "type": "text", + "content": "While CLIP achieves remarkable performance in broad domains and tasks, its image and text alignment is still not perfect. For example, CLIP models tend to encode images and texts into different clusters for each modality, and thus, there is a modality gap between images and text features even after sufficiently training with the massive datasets [28, 39, 66]. This modality gap suggests that CLIP has difficulty in precisely mapping images and text. In fact, Liang et al. [28] have shown that the modality gap largely affects downstream task performance, especially in fine-grained classification tasks, and Ray et al. [43] have demonstrated that CLIP models often fail to retrieve localized objects and attributes in images from corresponding captions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "text", + "content": "To address the modality gap in CLIP, existing literature mainly focuses on refining the CLIP feature spaces in the pre-training or fine-tuning phases. For pre-training, one approach is multi-task learning of the contrastive objective and auxiliary losses such as self-supervised loss with data augmentations [26, 27]. Another direction for pre-training is to reduce the modality gap by modifying the encoder architectures to explicitly share the weights or feature maps between the image and text encoders [6, 64]. For fine-tuning, previous works have demonstrated that the target downstream task performance can be improved significantly by prompt tuning that optimizes trainable visual/textual tokens on target tasks to match image and text in the feature space (i.e., reducing the modality gap) [23, 33, 68]. More recently, Yang et al. [62] have introduced adapter parameters to share the infor" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 282, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 282, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 282, + 713 + ], + "type": "text", + "content": "*Corresponding author. shinya.yamaguchi@ntt.com \n1Code: https://github.com/yshinya6/clip-refine" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4256" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 95 + ], + "type": "text", + "content": "mation across modalities. This cross-modal shared adapter is fine-tuned to solve the target tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 99, + 295, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 99, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 57, + 99, + 295, + 313 + ], + "type": "text", + "content": "Although these pre-training/fine-tuning methods succeed in improving the cross-modal alignment capability, they face difficulties in terms of computational cost and zero-shot transfer performance. On the one hand, pre-training methods require training from scratch with million-scale image-text datasets like LAION-400M [46]. Since the previous works often do not publish pre-trained models using such large datasets, naive CLIP models published by OpenAI [40] or OpenCLIP [8] are still a practical choice under the limited budget. On the other hand, fine-tuning methods do not require huge computation costs because their datasets are much smaller than pre-training ones. Nevertheless, fine-tuning pretrained CLIP models degrades the general zero-shot transfer performance due to focusing on target tasks [25]. This is inevitable as the primary goal of fine-tuning is to improve the target task performance. Therefore, we seek a new approach that is lighter than pre-training and avoids degrading zero-shot transfer performance in fine-tuning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 316, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 316, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 57, + 316, + 295, + 507 + ], + "type": "text", + "content": "In this paper, we tackle aligning the image and text features of CLIP models in a training phase called post-pretraining between pre-training and fine-tuning (Fig. 1). The goal of post-pre-training is to mitigate the modality gap and enhance zero-shot transfer performance of off-the-shelf pretrained CLIP models by only using reasonable computation resources and datasets (e.g., a single GPU and COCO Captions [29]). This is challenging because naive alignment methods, such as directly minimizing the gap between image and text features, corrupt the feature spaces in terms of the uniformity on the hypersphere [53]; the uniformity is an important property of contrastive learning representation that indicates the amount of input data information. Furthermore, post-pre-training with contrastive loss causes catastrophic forgetting of the general knowledge in pre-trained CLIP models due to the overfitting by the restricted mini-batch sizes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "text", + "content": "To this end, we introduce a method for post-pre-training called CLIP-Refine (Figure 2), which is composed of random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). Instead of directly minimizing the gap between features, our idea is to refine the image and text feature distributions to follow a shared prior (e.g., standard Gaussian). To achieve this, RaFA minimizes the gap between the image/text feature vectors and the reference vectors, which are randomly generated from the prior. In RaFA, the reference vectors are shared for the pairs of input image and text. Thus, RaFA penalizes image and text features to explicitly follow the same distribution. By matching the image and text feature distributions, the features are expected to avoid excessive concentration and preserve uniformity among samples, resulting in a good balance of multi-modal alignment and uniformity in the feature space. Conversely, HyCD is designed to avoid catastrophic forgetting by modified self-distillation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 191 + ], + "type": "text", + "content": "loss with the supervision of image-text pairs. Specifically, we apply the knowledge distillation where the teacher is the pre-trained model, i.e., minimizing KL-divergence between the outputs of the teacher and (student) post-pre-training model. Furthermore, we encourage learning new knowledge by blending the teacher's similarity matrix and the identity matrix, which represents the supervision of matched image-text pairs. Combining RaFA and HyCD, the CLIP models can reduce the modality gap while maintaining the uniformity of features without forgetting past knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 192, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 192, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 192, + 553, + 299 + ], + "type": "text", + "content": "Our extensive experiments on zero-shot classification (12 datasets) and cross-modal retrieval (2 datasets) demonstrate that our method significantly improves zero-shot performance over post-pre-training baselines with contrastive loss. Through the quantitative and qualitative analysis of feature spaces, we found that CLIP-Refine not only reduces the modality gap but also improves the uniformity of the hypersphere. This suggests that post-pre-training facilitates effective cross-modal alignment and well-clustered features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 311, + 400, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 311, + 400, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 311, + 400, + 323 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 330, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 330, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 330, + 553, + 510 + ], + "type": "text", + "content": "CLIP and Modality Gap. CLIP [21, 40] is a multi-modal representation learning technique that trains a model to embed inputs from image and text modalities into shared feature space. This is done by making the correct pairs of image and text closer in the feature space while the other pairs in mini-batch repelled via the InfoNCE-based contrastive loss [37]. While this simple multi-modal pre-training has remarkably advanced multi-modal research in various fields [4, 5, 7, 51, 52, 54, 58], Liang et al. [28] have revealed that pre-trained CLIP models encode image and text into different clusters for each modality, i.e., modality gap, and the small temperature parameter in the CLIP's contrastive loss can cause the modality gap empirically. The successor work by Qian et al. [39] has theoretically shown that contrastive loss can not reduce the modality gap perfectly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 510, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 510, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 510, + 553, + 712 + ], + "type": "text", + "content": "Pre-training Modifications. To mitigate the modality gap and improve the cross-modal alignment, several works have modified the contrastive loss of CLIP by adding auxiliary losses, including geometric cyclic consistency between image and text features [15], the fine-grained similarity between output tokens of both modalities [63], self-supervised loss with data augmentation [26, 27], supervised contrastive learning [61], textual augmented contrastive loss using LLM-generated positive/negative texts [10]. Another direction is to explicitly share the information of image and text modalities by introducing shared encoder weights [64] or shared feature spaces of finite discrete tokens [6] to the pre-training. These methods, however, require large-scale image-text pair datasets (e.g., CC12M [3] and LAION-400M [46]) to achieve practical performance, incurring high computational costs on multiple GPUs. In contrast, our post-pre-training method improves the generalization performance of off-the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "4257" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 92, + 289, + 211 + ], + "blocks": [ + { + "bbox": [ + 132, + 75, + 256, + 87 + ], + "lines": [ + { + "bbox": [ + 132, + 75, + 256, + 87 + ], + "spans": [ + { + "bbox": [ + 132, + 75, + 256, + 87 + ], + "type": "text", + "content": "Random Feature Alignment" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 96, + 92, + 289, + 211 + ], + "lines": [ + { + "bbox": [ + 96, + 92, + 289, + 211 + ], + "spans": [ + { + "bbox": [ + 96, + 92, + 289, + 211 + ], + "type": "image", + "image_path": "c37fd29f21c5043056c83372df2d3daf26f072ac24d98476e4033eef5e809084.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 334, + 91, + 504, + 211 + ], + "blocks": [ + { + "bbox": [ + 348, + 75, + 484, + 87 + ], + "lines": [ + { + "bbox": [ + 348, + 75, + 484, + 87 + ], + "spans": [ + { + "bbox": [ + 348, + 75, + 484, + 87 + ], + "type": "text", + "content": "Hybrid Contrastive-Distillation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 91, + 504, + 211 + ], + "lines": [ + { + "bbox": [ + 334, + 91, + 504, + 211 + ], + "spans": [ + { + "bbox": [ + 334, + 91, + 504, + 211 + ], + "type": "image", + "image_path": "1d936d63ae7949b54b4104ce354ec93c3e115be645ac5ca6b846ef9544848e72.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 223, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 54, + 223, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 54, + 223, + 555, + 278 + ], + "type": "text", + "content": "Figure 2. Intuition of CLIP-Refine (proposed method). CLIP-Refine modifies a pre-trained CLIP model by Random Feature Alignment (RaFA, left) and Hybrid Contrastive-Distillation (HyCD, right). RaFA penalizes image and text features to follow a shared prior distribution by minimizing the gap between the features and a random reference vector sampled from the prior. HyCD trains the model with the KL-divergence-based distillation loss using hybrid soft labels composed by alpha blending the ground truth label of image-text pairs with the output of the pre-trained model (teacher)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 296, + 334 + ], + "type": "text", + "content": "shelf pre-trained models by only using small datasets like Flickr8/30K [42] and COCO Caption [29] with a single GPU. Besides, our method is cooperative with these pre-training methods because it can be used for any pre-trained model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 334, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 334, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 56, + 334, + 296, + 693 + ], + "type": "text", + "content": "Fine-tuning Methods2. The original paper of CLIP [40] has reported lower zero-shot performance in several fine-grained classification tasks, such as Aircraft [34], demonstrating the imperfectness of the cross-modal alignment of CLIP. Motivated by this, Zhou et al. [68, 69] have proposed to optimize text prompts to mitigate the modality gap by updating additional trainable vectors while the pre-trained models are fixed. This simple approach helps match image and text in the feature space (i.e., reducing the modality gap) and significantly improves target performance. Similarly, The method of Jia et al. [22] learns the additional trainable parameters for image input (visual prompt), and the successor works [23, 48] have unified the visual and textual prompt tuning by the multi-task learning objectives. Another approach is to directly share the low-rank adaptation parameters across the modalities and update them in fine-tuning [62]. In the perspective of the refinement of CLIP feature spaces, Oh et al. [36] have theoretically and empirically shown that fine-tuned CLIP models retain the suboptimal feature space in terms of alignment and uniformity, which are important properties to assess the feature quality in contrastive learning [53]. To refine the feature space, they have provided a mixup-based fine-tuning method, which generates hard negative samples on the hypersphere in a cross-modal manner (" + }, + { + "bbox": [ + 56, + 334, + 296, + 693 + ], + "type": "inline_equation", + "content": "m^2" + }, + { + "bbox": [ + 56, + 334, + 296, + 693 + ], + "type": "text", + "content": "-mix). These fine-tuning techniques succeed in reducing the gap in target datasets, but they do not intend to improve the zero-shot transfer performance. Our post-pre-training method differs from the fine-tuning methods in the objective: our method aims to improve the zero-shot transfer performance by reducing the modality gap. Obviously, we can" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 285, + 506, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 285, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 315, + 285, + 506, + 297 + ], + "type": "text", + "content": "Algorithm 1 CLIP-Refine for Post-pre-training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "text", + "content": "Require: Training dataset " + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "text", + "content": ", vision encoder " + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{V}}" + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "text", + "content": ", text encoder " + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{T}}" + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "text", + "content": ", pre-trained parameters " + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "inline_equation", + "content": "\\{\\theta_{\\mathrm{V}}^{\\mathrm{P}}, \\theta_{\\mathrm{T}}^{\\mathrm{P}}\\}" + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "text", + "content": ", training batchsize " + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "text", + "content": ", step size " + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "text", + "content": ", temperature parameter " + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "text", + "content": ", hyper-parameters " + }, + { + "bbox": [ + 315, + 300, + 555, + 334 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 334, + 487, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 334, + 487, + 345 + ], + "spans": [ + { + "bbox": [ + 315, + 334, + 487, + 345 + ], + "type": "text", + "content": "Ensure: Post-pre-trained parameters " + }, + { + "bbox": [ + 315, + 334, + 487, + 345 + ], + "type": "inline_equation", + "content": "\\{\\theta_{\\mathrm{V}},\\theta_{\\mathrm{T}}\\}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 346, + 553, + 594 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 321, + 346, + 414, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 346, + 414, + 355 + ], + "spans": [ + { + "bbox": [ + 321, + 346, + 414, + 355 + ], + "type": "text", + "content": "1: # Initialize parameters" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 356, + 368, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 356, + 368, + 366 + ], + "spans": [ + { + "bbox": [ + 321, + 356, + 368, + 366 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 321, + 356, + 368, + 366 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{V}} \\gets \\theta_{\\mathrm{V}}^{\\mathrm{P}}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 367, + 368, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 367, + 368, + 377 + ], + "spans": [ + { + "bbox": [ + 321, + 367, + 368, + 377 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 321, + 367, + 368, + 377 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{T}} \\gets \\theta_{\\mathrm{T}}^{\\mathrm{P}}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 378, + 419, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 378, + 419, + 388 + ], + "spans": [ + { + "bbox": [ + 321, + 378, + 419, + 388 + ], + "type": "text", + "content": "4: while not converged do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 388, + 412, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 388, + 412, + 399 + ], + "spans": [ + { + "bbox": [ + 321, + 388, + 412, + 399 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 321, + 388, + 412, + 399 + ], + "type": "inline_equation", + "content": "\\{(x^i,t^i)\\}_{i = 1}^B\\sim \\mathcal{D}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 399, + 518, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 399, + 518, + 411 + ], + "spans": [ + { + "bbox": [ + 321, + 399, + 518, + 411 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 321, + 399, + 518, + 411 + ], + "type": "inline_equation", + "content": "\\{z_{\\mathrm{img}}^i,z_{\\mathrm{txt}}^i\\}_{i = 1}^B\\gets \\{f_{\\mathrm{V}}(x^i;\\theta_{\\mathrm{V}}),f_{\\mathrm{T}}(t^i;\\theta_{\\mathrm{T}})\\}_{i = 1}^B" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 411, + 518, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 411, + 518, + 423 + ], + "spans": [ + { + "bbox": [ + 321, + 411, + 518, + 423 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 321, + 411, + 518, + 423 + ], + "type": "inline_equation", + "content": "\\{z_{\\mathrm{img}}^{i,\\mathrm{p}}, z_{\\mathrm{xt}}^{i,\\mathrm{p}}\\}_{i=1}^{B} \\gets \\{f_{\\mathrm{V}}(x^i; \\theta_{\\mathrm{V}}^{\\mathrm{p}}), f_{\\mathrm{T}}(t^i; \\theta_{\\mathrm{T}}^{\\mathrm{p}})\\}_{i=1}^{B}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 423, + 494, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 423, + 494, + 434 + ], + "spans": [ + { + "bbox": [ + 321, + 423, + 494, + 434 + ], + "type": "text", + "content": "8: # Compute random feature alignment loss" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 434, + 423, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 434, + 423, + 445 + ], + "spans": [ + { + "bbox": [ + 321, + 434, + 423, + 445 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 321, + 434, + 423, + 445 + ], + "type": "inline_equation", + "content": "\\{z_{\\mathrm{ref}}^{\\mathrm{i}}|z_{\\mathrm{ref}}^{\\mathrm{i}}\\sim p(z)\\}_{i = 1}^{B}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 445, + 539, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 445, + 539, + 456 + ], + "spans": [ + { + "bbox": [ + 318, + 445, + 539, + 456 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 318, + 445, + 539, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RaFA}}\\gets \\frac{1}{2B}\\sum_{i = 1}^{B}\\| z_{\\mathrm{img}}^i -z_{\\mathrm{ref}}^i\\| _2^2 +\\| z_{\\mathrm{txt}}^i -z_{\\mathrm{ref}}^i\\| _2^2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 456, + 508, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 456, + 508, + 466 + ], + "spans": [ + { + "bbox": [ + 318, + 456, + 508, + 466 + ], + "type": "text", + "content": "11: # Compute hybrid contrastive-distillation loss" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 466, + 553, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 466, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 318, + 466, + 553, + 488 + ], + "type": "text", + "content": "12: Compute student outputs " + }, + { + "bbox": [ + 318, + 466, + 553, + 488 + ], + "type": "inline_equation", + "content": "p_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}}" + }, + { + "bbox": [ + 318, + 466, + 553, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 318, + 466, + 553, + 488 + ], + "type": "inline_equation", + "content": "p_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}" + }, + { + "bbox": [ + 318, + 466, + 553, + 488 + ], + "type": "text", + "content": " for all sample combinations in a batch by Eq. (6) with " + }, + { + "bbox": [ + 318, + 466, + 553, + 488 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 318, + 466, + 553, + 488 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 488, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 488, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 318, + 488, + 553, + 510 + ], + "type": "text", + "content": "13: Compute teacher outputs " + }, + { + "bbox": [ + 318, + 488, + 553, + 510 + ], + "type": "inline_equation", + "content": "\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}}" + }, + { + "bbox": [ + 318, + 488, + 553, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 318, + 488, + 553, + 510 + ], + "type": "inline_equation", + "content": "\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}" + }, + { + "bbox": [ + 318, + 488, + 553, + 510 + ], + "type": "text", + "content": " for all sample combinations in a batch by Eq. (8) with " + }, + { + "bbox": [ + 318, + 488, + 553, + 510 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 318, + 488, + 553, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 318, + 488, + 553, + 510 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 318, + 488, + 553, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 510, + 498, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 510, + 498, + 529 + ], + "spans": [ + { + "bbox": [ + 318, + 510, + 498, + 529 + ], + "type": "text", + "content": "14: " + }, + { + "bbox": [ + 318, + 510, + 498, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{I}\\rightarrow \\mathrm{T}}\\gets \\frac{1}{B}\\sum_{i = 1}^{B}\\sum_{j = 1}^{B}\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}}\\log \\frac{\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow\\mathrm{T}}}{p_{i,j}^{\\mathrm{I}\\rightarrow\\mathrm{T}}}" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 529, + 498, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 529, + 498, + 550 + ], + "spans": [ + { + "bbox": [ + 318, + 529, + 498, + 550 + ], + "type": "text", + "content": "15: " + }, + { + "bbox": [ + 318, + 529, + 498, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{T}\\rightarrow \\mathrm{I}}\\gets \\frac{1}{B}\\sum_{i = 1}^{B}\\sum_{j = 1}^{B}\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}\\log \\frac{\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}}{p_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 549, + 462, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 549, + 462, + 561 + ], + "spans": [ + { + "bbox": [ + 318, + 549, + 462, + 561 + ], + "type": "text", + "content": "16: " + }, + { + "bbox": [ + 318, + 549, + 462, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HyCD}}\\gets \\frac{1}{2} (\\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{I}\\to \\mathrm{T}} + \\mathcal{L}_{\\mathrm{HyCD}}^{\\mathrm{T}\\to \\mathrm{I}})" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 561, + 484, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 561, + 484, + 572 + ], + "spans": [ + { + "bbox": [ + 318, + 561, + 484, + 572 + ], + "type": "text", + "content": "17: " + }, + { + "bbox": [ + 318, + 561, + 484, + 572 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{V}}\\gets \\theta_{\\mathrm{V}} - \\frac{\\eta}{2}\\nabla_{\\theta_{\\mathrm{V}}}\\left(\\mathcal{L}_{\\mathrm{RaFA}} + \\mathcal{L}_{\\mathrm{HyCD}}\\right)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 572, + 482, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 572, + 482, + 584 + ], + "spans": [ + { + "bbox": [ + 318, + 572, + 482, + 584 + ], + "type": "text", + "content": "18: " + }, + { + "bbox": [ + 318, + 572, + 482, + 584 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{T}}\\gets \\theta_{\\mathrm{T}} - \\frac{\\eta}{2}\\nabla_{\\theta_{\\mathrm{T}}}(\\mathcal{L}_{\\mathrm{RaFA}} + \\mathcal{L}_{\\mathrm{HyCD}})" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 584, + 371, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 584, + 371, + 594 + ], + "spans": [ + { + "bbox": [ + 318, + 584, + 371, + 594 + ], + "type": "text", + "content": "19: end while" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 619, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 555, + 643 + ], + "type": "text", + "content": "leverage any fine-tuning techniques after post-pre-training by our method, and it helps to boost the baseline performance." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 656, + 370, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 370, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 370, + 668 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "content": "We propose post-pre-training of CLIP in between pretraining and fine-tuning to mitigate the modality gap (Figure 1). To this end, we present CLIP-Refine (Figure 2)," + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 282, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 282, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 282, + 713 + ], + "type": "text", + "content": "2We refer to training models on specific target tasks as fine-tuning." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4258" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": "which is composed of random feature alignment (RaFA) and hybrid contrastive-distillation (HyCD). RaFA penalizes the model to induce both modal features into a single shared distribution by minimizing the gap between the paired features and a random reference vector. To reconcile learning new knowledge with retaining past knowledge, HyCD updates the models with knowledge distillation loss using fixed pretrained models, where the teacher outputs are mixed with ground-truth labels of image-text pairs. Algorithm 1 shows the overall procedure of CLIP-Refine." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 199, + 243, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 243, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 243, + 213 + ], + "type": "text", + "content": "3.1. Problem Setting: Post-pre-training" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": "We consider a problem setting called post-pre-training, where we aim to improve the cross-modal alignment and generalization performance of pre-trained vision-language models. This setting allows to access a vision encoder " + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{V}}: \\mathcal{X} \\to \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{T}}: \\mathcal{T} \\to \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": " with the parameter " + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{CLIP}} = \\{\\theta_{\\mathrm{V}}^{\\mathrm{CLIP}}, \\theta_{\\mathrm{T}}^{\\mathrm{CLIP}}\\}" + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": " pre-trained by CLIP, where " + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": " are the image and text space. We optimize the parameters " + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{V}}" + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{V}}" + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": " on a post-pre-training image-text pair dataset " + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_i, t_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": "; we basically assume that " + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": " is much smaller than the pre-training dataset and contains images and text captions in general domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 355, + 168, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 168, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 168, + 368 + ], + "type": "text", + "content": "3.2. Objective Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 396 + ], + "type": "text", + "content": "In CLIP-Refine, we optimize " + }, + { + "bbox": [ + 55, + 372, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{V}}" + }, + { + "bbox": [ + 55, + 372, + 296, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 372, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 372, + 296, + 396 + ], + "type": "text", + "content": " by the following objective function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 404, + 296, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 404, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 92, + 404, + 296, + 423 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {\\mathrm {V}}, \\theta_ {\\mathrm {T}}} \\mathcal {L} _ {\\mathrm {R a F A}} \\left(\\theta_ {\\mathrm {V}}, \\theta_ {\\mathrm {T}}\\right) + \\mathcal {L} _ {\\mathrm {H y C D}} \\left(\\theta_ {\\mathrm {V}}, \\theta_ {\\mathrm {T}}\\right), \\tag {1}", + "image_path": "09e45b9271efa7e52cf42e9ffa3d2c9ca6d5d7b198c7444d39eea521595ac2d0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RaFA}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "text", + "content": " is the random feature alignment loss and " + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HyCD}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "text", + "content": " is the hybrid contrastive-distillation loss. One can consider balancing " + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RaFA}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HyCD}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "text", + "content": " by hyperparameters, but we found that their equal contribution achieves the best performance (see Appendix). We describe the details of " + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RaFA}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HyCD}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 503 + ], + "type": "text", + "content": " in the following sections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 211, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 211, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 211, + 522 + ], + "type": "text", + "content": "3.3. Random Feature Alignment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 526, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 296, + 612 + ], + "type": "text", + "content": "The primary goal in post-pre-training is to minimize the modality gap and encourage cross-modal alignment. To this end, the most straightforward approach is just minimizing the feature gap between the modalities. Let " + }, + { + "bbox": [ + 55, + 526, + 296, + 612 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{img}}^i = f_{\\mathrm{V}}(x^i;\\theta_{\\mathrm{V}})" + }, + { + "bbox": [ + 55, + 526, + 296, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 526, + 296, + 612 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{xt}}^i = f_{\\mathrm{T}}(t^i;\\theta_{\\mathrm{T}})" + }, + { + "bbox": [ + 55, + 526, + 296, + 612 + ], + "type": "text", + "content": " are the normalized feature vectors of an image-text pair " + }, + { + "bbox": [ + 55, + 526, + 296, + 612 + ], + "type": "inline_equation", + "content": "(x^{i},t^{i})" + }, + { + "bbox": [ + 55, + 526, + 296, + 612 + ], + "type": "text", + "content": ". A naive loss for minimizing the modality gap can be defined by an " + }, + { + "bbox": [ + 55, + 526, + 296, + 612 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 526, + 296, + 612 + ], + "type": "text", + "content": " distance form as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 620, + 296, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 620, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 102, + 620, + 296, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\mathbb {E} _ {(x ^ {i}, t ^ {i}) \\in \\mathcal {D}} \\| z _ {\\text {i m g}} ^ {i} - z _ {\\text {t x t}} ^ {i} \\| _ {2} ^ {2}. \\tag {2}", + "image_path": "80fa80f855babbfca9bb3f61c9119583ea8756c41a938ae5af032d563b766e6c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "content": "However, minimizing this loss function degrades the generalization performance. This can be explained by the balance of alignment and uniformity [53], which are desirable properties of transferability in contrastive representation learning. Alignment is defined by the gap between positive (i.e., imagetext) pairs in the feature space, and uniformity is defined by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "equal separation among all features on the hypersphere. That is, minimizing Eq. (2) enhances the alignment of the positive pair but destroys the uniformity on the hypersphere by forcing changes to the feature distribution. In fact, Wang and Isola [53] have shown that using only the alignment loss in post-pre-training degrades the validation performance due to corrupt uniformity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "text", + "content": "To address this challenge, we introduce the idea of matching feature distributions instead of directly matching the paired features. In other words, we refine the multi-modal features of pre-trained CLIP models to follow a prior distribution " + }, + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "text", + "content": " (e.g., standard Gaussian " + }, + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "text", + "content": ") shared across the image and text modalities. Inspired by random feature regularization in single modal fine-tuning [59, 67], we formulate this idea as minimization between image/text feature vectors " + }, + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "inline_equation", + "content": "(z_{\\mathrm{img}}^i, z_{\\mathrm{txt}}^i)" + }, + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "text", + "content": " and a random reference vector " + }, + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{ref}}^i" + }, + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "text", + "content": " sampled from the shared prior " + }, + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 313, + 157, + 556, + 276 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 285, + 555, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 285, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 319, + 285, + 555, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R a F A}} = \\mathbb {E} _ {(x ^ {i}, t ^ {i}) \\in \\mathcal {D}} \\frac {1}{2} (\\| z _ {\\mathrm {i m g}} ^ {i} - z _ {\\mathrm {r e f}} ^ {i} \\| _ {2} ^ {2} + \\| z _ {\\mathrm {t x t}} ^ {i} - z _ {\\mathrm {r e f}} ^ {i} \\| _ {2} ^ {2}). \\tag {3}", + "image_path": "ddf001e418e776fad508ad3d3169830ae79f0f3956ba620f5b830ed08a21850a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "text", + "content": "Through this loss function, we can expect that (i) the modality gap is indirectly minimized via " + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{ref}}^i" + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "text", + "content": " shared with each image-text pair, (ii) the feature vectors follow the modality shared prior " + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "text", + "content": " gradually [59], (iii) the randomness of " + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{ref}}^i" + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "text", + "content": " helps to avoid the models from overfitting to " + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "text", + "content": " [67] and reform the models' capacity [49]. Furthermore, we found that " + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RaFA}}" + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "text", + "content": " can improve the uniformity from baselines (Table 3), meaning that the learned representations transfer well to downstream tasks. Throughout this paper, we use standard Gaussian " + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 313, + 316, + 556, + 447 + ], + "type": "text", + "content": " by default; we discuss the effect of prior choice in Sec. 4.4.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 455, + 484, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 484, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 484, + 467 + ], + "type": "text", + "content": "3.4. Hybrid Contrastive-Distillation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 473, + 556, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 557 + ], + "type": "text", + "content": "Another important goal of post-pre-training is to retain the past knowledge in pre-trained CLIP models. To this end, our basic strategy is to penalize the post-pre-training models with knowledge distillation loss [19] where the frozen pre-trained CLIP model performs as the teacher. We define the minibatch-wise knowledge distillation loss with KL-divergence according to the formulation of CLIP-KD [60] as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 377, + 565, + 555, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 565, + 555, + 580 + ], + "spans": [ + { + "bbox": [ + 377, + 565, + 555, + 580 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K D}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {K D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} + \\mathcal {L} _ {\\mathrm {K D}} ^ {\\mathrm {T} \\rightarrow \\mathrm {I}}\\right), \\tag {4}", + "image_path": "492e51867e0eb0c0f9cfc96b91347be193252c93b3dadce83ccac3e4787fad9b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 351, + 582, + 555, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 582, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 351, + 582, + 555, + 604 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {B} q _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} \\log \\frac {q _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}{p _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}, \\tag {5}", + "image_path": "b734d8e2dcd5cc4e4d30b23d3be9c6028ef1baf9b24107c71da3849ed98d0583.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 370, + 605, + 555, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 605, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 370, + 605, + 555, + 628 + ], + "type": "interline_equation", + "content": "p _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {\\exp \\left(z _ {\\mathrm {i m g}} ^ {i} \\cdot z _ {\\mathrm {t x t}} ^ {j} / \\tau\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(z _ {\\mathrm {i m g}} ^ {i} \\cdot z _ {\\mathrm {t x t}} ^ {k} / \\tau\\right)}, \\tag {6}", + "image_path": "a1c6bffd1d75ddb01acd7a8dac59ef97d924066b31c18ae3da060c89789010cd.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 372, + 630, + 555, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 630, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 372, + 630, + 555, + 653 + ], + "type": "interline_equation", + "content": "q _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {\\exp \\left(z _ {\\mathrm {i m g}} ^ {i , \\mathrm {p}} \\cdot z _ {\\mathrm {t x t}} ^ {j , \\mathrm {p}} / \\tau\\right)}{\\sum_ {k = 1} ^ {B} \\exp \\left(z _ {\\mathrm {i m g}} ^ {i , \\mathrm {p}} \\cdot z _ {\\mathrm {t x t}} ^ {k , \\mathrm {p}} / \\tau\\right)}, \\tag {7}", + "image_path": "cdca1f3717eac2fca6daeb58f19a38a49c802db22366f4e6968a7060740f2b7b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 663, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 556, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 663, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 663, + 556, + 714 + ], + "type": "text", + "content": " is the temperature parameter, " + }, + { + "bbox": [ + 313, + 663, + 556, + 714 + ], + "type": "inline_equation", + "content": "(z_{\\mathrm{img}}^{i,\\mathrm{p}},z_{\\mathrm{xt}}^{i,\\mathrm{p}}) = (f_{\\mathrm{V}}(x^{i};\\theta_{\\mathrm{V}}^{\\mathrm{CLIP}}),f_{\\mathrm{T}}(t^{i};\\theta_{\\mathrm{T}}^{\\mathrm{CLIP}}))" + }, + { + "bbox": [ + 313, + 663, + 556, + 714 + ], + "type": "text", + "content": " are the feature vectors generated by the pre-trained image/text encoders, and " + }, + { + "bbox": [ + 313, + 663, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KD}}^{\\mathrm{T}\\rightarrow \\mathrm{I}}" + }, + { + "bbox": [ + 313, + 663, + 556, + 714 + ], + "type": "text", + "content": " is defined by swapping the order of image features and text" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4259" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 82, + 555, + 186 + ], + "blocks": [ + { + "bbox": [ + 156, + 70, + 454, + 81 + ], + "lines": [ + { + "bbox": [ + 156, + 70, + 454, + 81 + ], + "spans": [ + { + "bbox": [ + 156, + 70, + 454, + 81 + ], + "type": "text", + "content": "Table 1. Zero-shot classification performance on 12 datasets with CLIP ViT-B/32." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 82, + 555, + 186 + ], + "lines": [ + { + "bbox": [ + 59, + 82, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 59, + 82, + 555, + 186 + ], + "type": "table", + "html": "
MethodAirBirdCalCarDTDEuroFloFoodINPetSUNUCFAvg.
Pre-trained18.8149.3723.9860.0740.9038.2059.7580.2159.0481.7960.8559.9052.74
Contrastive13.9840.0723.7747.6336.2234.6846.2870.3152.9670.1354.7858.2645.75
m2-mix [36]14.6440.7623.4949.8636.4934.6547.4271.2753.5870.3256.5158.7646.48
Self-KD18.9651.5223.9460.0441.1738.3659.5580.0259.0681.8560.6860.1652.94
HyCD18.7250.5023.9859.8340.4840.9058.0680.1959.7181.4562.0461.7053.13
HyCD + L_align13.2934.2023.7343.7035.0039.2746.2773.3152.5772.4255.0158.5545.61
CLIP-Refine (Ours)20.7752.7224.7960.6341.4941.4462.6481.8760.9383.5763.6361.8554.69
", + "image_path": "ae6722b5eda3466910d11b4522fc85326e5d2d72c5ba33a9181788de97c9d4cb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 199, + 295, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 247 + ], + "type": "text", + "content": "features in Eq. (6) and (7). Although minimizing Eq. (4) retains the past knowledge, we found that it interferes with RaFA because it strongly enchains the models' parameters to the pre-trained ones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 247, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 320 + ], + "type": "text", + "content": "To reconcile RaFA and the retention of past knowledge, we present a technique that blends the ground-truth label of an image-text pair with the teacher output. Concretely, we modify Eq. (7) by alpha-blending of an indicator function " + }, + { + "bbox": [ + 55, + 247, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{i = j}" + }, + { + "bbox": [ + 55, + 247, + 296, + 320 + ], + "type": "text", + "content": ", which returns one if " + }, + { + "bbox": [ + 55, + 247, + 296, + 320 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 55, + 247, + 296, + 320 + ], + "type": "text", + "content": " otherwise zero, and the teacher output signal:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 324, + 296, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 324, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 110, + 324, + 296, + 341 + ], + "type": "interline_equation", + "content": "\\hat {q} _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\alpha \\mathbb {I} _ {i = j} + (1 - \\alpha) q _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}, \\tag {8}", + "image_path": "093383427e8645bb8858b80f5ae6004684dd66667ccab734c712bc3759e8387f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "text", + "content": " is a hyperparameter to balance acquiring new knowledge and retaining past knowledge; we use " + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "text", + "content": " by default and discuss the effect of " + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "text", + "content": " in the Appendix. With this modification, the post-pre-training models can further enhance the cross-modal alignment by learning the correct image-text pair (i.e., " + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "text", + "content": ") with reference to the relative knowledge contained in the " + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "text", + "content": " coordinates of the teacher's outputs. By using " + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\hat{q}_{i,j}^{\\mathrm{I}\\rightarrow \\mathrm{T}}" + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\hat{q}_{i,j}^{\\mathrm{T}\\rightarrow \\mathrm{I}}" + }, + { + "bbox": [ + 55, + 346, + 296, + 453 + ], + "type": "text", + "content": ", we compute the HyCD loss function as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 460, + 295, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 460, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 107, + 460, + 295, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {H y C D}} = \\frac {1}{2} \\left(\\mathcal {L} _ {\\mathrm {H y C D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} + \\mathcal {L} _ {\\mathrm {H y C D}} ^ {\\mathrm {T} \\rightarrow \\mathrm {I}}\\right), \\tag {9}", + "image_path": "a51067387eafc3da22a7bdfc30c67b5166cdb37688dbdf5736cc1dcaccfd4897.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 476, + 295, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 476, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 91, + 476, + 295, + 499 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {H y C D}} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} = \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\sum_ {j = 1} ^ {B} \\hat {q} _ {i, j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}} \\log \\frac {\\hat {q} _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}{p _ {i , j} ^ {\\mathrm {I} \\rightarrow \\mathrm {T}}}. \\tag {10}", + "image_path": "1d959ee566756937879abdf2e23b2fe14e344bcdb75a9f4b630adddead37e51a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 504, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 553 + ], + "type": "text", + "content": "While " + }, + { + "bbox": [ + 55, + 504, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HyCD}}" + }, + { + "bbox": [ + 55, + 504, + 296, + 553 + ], + "type": "text", + "content": " solely improves the generalization performance of the pre-trained CLIP models, combining it with " + }, + { + "bbox": [ + 55, + 504, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RaFA}}" + }, + { + "bbox": [ + 55, + 504, + 296, + 553 + ], + "type": "text", + "content": " achieves larger improvements by minimizing the modality gaps without losing the past knowledge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 562, + 137, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 137, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 137, + 575 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 297, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 642 + ], + "type": "text", + "content": "We evaluate CLIP-Refine on 12 classification datasets and 2 cross-modal retrieval datasets by using multiple pre-trained models. We also conduct qualitative and quantitative analyses to assess the feature space through modality gap, alignment, uniformity metrics, and PCA visualization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 648, + 116, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 116, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 116, + 661 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": "Baselines. We compare our CLIP-Refine with the following baselines of post-pre-training. Pre-trained: predicting with pre-trained weights without additional training. Contrastive: post-pre-training with the contrastive" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 199, + 555, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 284 + ], + "type": "text", + "content": "loss [21, 40]. " + }, + { + "bbox": [ + 313, + 199, + 555, + 284 + ], + "type": "inline_equation", + "content": "m^2" + }, + { + "bbox": [ + 313, + 199, + 555, + 284 + ], + "type": "text", + "content": "-mix [36]: post-pre-training by the modified contrastive loss with multi-modal mixup; we use this fine-tuning method as a baseline because it can be used for refining the feature spaces as suggested in the original paper [36]. Self-KD: using Eq. (4) for post-pre-training. HyCD: using only Eq. (9) for training. HyCD + L align: combining Eq. (9) and Eq. (2)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 289, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 338 + ], + "type": "text", + "content": "Post-pre-training Datasets. We used COCO Caption [29] as the default post-pre-training dataset. We also used Flickr8K/30K [42], CC3M [47], and CC12M [3] to confirm the dataset size effect in post-pre-training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 342, + 556, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 486 + ], + "type": "text", + "content": "Test Datasets. We used 12 image classification datasets containing various image domains: Aircraft (Air) [34], Bird [56], Caltech-101 (Cal) [12] Car [24], DTD [9], EuroSAT (Euro) [18], Flower (Flo) [35], Food [2], ImageNet (IN) [45], Pet [38], SUN397 [57], and UCF-101 [50]. We also evaluated our method on COCO2017-Val and Flickr8K/30K for the image and text retrieval tasks; we denote text-to-image retrieval as " + }, + { + "bbox": [ + 313, + 342, + 556, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{T} \\rightarrow \\mathrm{I}" + }, + { + "bbox": [ + 313, + 342, + 556, + 486 + ], + "type": "text", + "content": " and image-to-text retrieval as " + }, + { + "bbox": [ + 313, + 342, + 556, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{I} \\rightarrow \\mathrm{T}" + }, + { + "bbox": [ + 313, + 342, + 556, + 486 + ], + "type": "text", + "content": ". We adopt these datasets since they are often used to evaluate the zero-shot transfer performance of CLIP models [68]. We evaluated models on the test sets except for ImageNet, and the official validation set for ImageNet." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 491, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 556, + 564 + ], + "type": "text", + "content": "Pre-trained Models. By default, we used CLIP-ViT-B/32 [40] with the pre-trained weights downloaded from the OpenAI's official repositories. We also tested other pretrained models, including OpenCLIP-ViT-H/14, OpenCLIP-ViT-bigG/14, SigLIP [65] and DFN [11] with the pre-trained weights on the OpenCLIP [8] repository." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "type": "text", + "content": "Post-pre-training. In all settings, we trained models by the AdamW [32] optimizer with the learning rate of " + }, + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "type": "inline_equation", + "content": "1.0 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "type": "text", + "content": " for one epoch; we determined the learning rate by the zero-shot classification validation using the subset of the ImageNet training set constructed by uniform sampling. We used mini-batch sizes of 512 as the default. For CLIP-Refine, we used " + }, + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "type": "text", + "content": ". The input samples were preprocessed using the default image transformations provided by each pre-trained model. We implemented the training and evaluation with PyTorch-1.13. We ran the experiments on a 24-core Intel Xeon CPU with a single NVIDIA A100 GPU with 80GB VRAM." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4260" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 82, + 555, + 175 + ], + "blocks": [ + { + "bbox": [ + 121, + 70, + 487, + 81 + ], + "lines": [ + { + "bbox": [ + 121, + 70, + 487, + 81 + ], + "spans": [ + { + "bbox": [ + 121, + 70, + 487, + 81 + ], + "type": "text", + "content": "Table 2. Zero-shot retrieval performance on COCO2017-Val and Flickr8K/30K with CLIP ViT-B/32." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 82, + 555, + 175 + ], + "lines": [ + { + "bbox": [ + 57, + 82, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 57, + 82, + 555, + 175 + ], + "type": "table", + "html": "
MethodCOCO2017-ValFlickr8KFlickr30K
T→II→TR@1T→IR@1I→TR@1T→IR@1I→T
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@5R@10R@1R@5R@10R@1R@5R@10
Pre-trained30.5654.9265.2633.2659.1068.7830.0653.8363.2932.4255.5765.2621.0640.9650.5324.5244.5754.14
Contrastive34.8861.5072.1031.8656.8067.4834.9459.4169.3628.8351.5561.0725.6947.3256.9318.4236.2945.37
m2-mix [36]36.2862.1873.0832.9257.9667.5635.0459.6669.6628.6450.6560.7725.9947.1857.1919.8837.6346.51
Self-KD31.0455.5865.9035.3659.0069.7229.9853.8363.5532.4755.1165.1021.9241.5150.7524.9444.8254.27
HyCD36.0462.2873.1437.8862.5472.8435.7260.3470.1434.6957.1967.2227.1149.1258.7827.1747.6656.94
HyCD + L align33.9261.1872.0635.1461.7472.0631.7355.6365.9032.5856.9466.9923.7844.7354.4723.5845.0755.23
CLIP-Refine (Ours)37.6463.5474.4238.7865.0475.1236.1161.2970.5138.2861.5570.9426.6549.7059.0330.9951.5461.31
", + "image_path": "da4a9eb69a642e8e6d23e870f88f2a379da89442c803dad14c67fd17995f2c38.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "text", + "content": "Evaluation Metrics. For zero-shot transfer performance, we report top-1 accuracy in classification tasks and recall@ " + }, + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "text", + "content": " scores in cross-modal retrieval tasks; recall@ " + }, + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "text", + "content": " indicates the proportion of correct answers in the retrieved top- " + }, + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "text", + "content": " candidates for all test input samples. We denote recall@ " + }, + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "inline_equation", + "content": "R@k" + }, + { + "bbox": [ + 55, + 188, + 297, + 297 + ], + "type": "text", + "content": ". To quantitatively evaluate the post-pre-trained features, we measure modality gap, alignment, and uniformity scores. Modality gap [28] assesses a cluster-wise gap between image and text features:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 304, + 295, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 304, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 111, + 304, + 295, + 319 + ], + "type": "interline_equation", + "content": "\\text {M o d a l i t y} \\mathrm {G a p} := \\| \\bar {f} _ {\\mathrm {V}} - \\bar {f} _ {\\mathrm {T}} \\| _ {2} ^ {2}, \\tag {11}", + "image_path": "55d3ee082e46899dfd3879db49b875a603f135ab247926b9eea2ed9395d21980.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 327, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 353 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 327, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\bar{f}" + }, + { + "bbox": [ + 55, + 327, + 296, + 353 + ], + "type": "text", + "content": " is a mean feature vector for all test samples of each modality; lower is better. Alignment [53] is defined by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 361, + 295, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 361, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 72, + 361, + 295, + 393 + ], + "type": "interline_equation", + "content": "\\text {A l i g n m e n t} := \\frac {1}{N _ {\\text {t e s t}}} \\sum_ {i = 1} ^ {N _ {\\text {t e s t}}} \\| f _ {\\mathrm {V}} \\left(x ^ {i}\\right) - f _ {\\mathrm {T}} \\left(t ^ {i}\\right) \\| _ {2} ^ {2}, \\tag {12}", + "image_path": "e607e3b7284f25d63cb0505980bd10924be039a60cecf44835355df5cb3ec5c6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 402, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 461 + ], + "type": "text", + "content": "representing how well the positive pair is aligned in the feature space; lower is better. Uniformity [53] evaluates how uniformly distributed the feature vectors are on the hypersphere, which is defined by radial basis function (RBF) kernel as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 468, + 295, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 468, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 63, + 468, + 295, + 498 + ], + "type": "interline_equation", + "content": "\\text {U n i f o r m i t y} := \\frac {1}{2 N _ {\\text {t e s t}}} \\sum_ {f _ {1}, f _ {2} \\in F} \\exp (- 2 \\| f _ {1} - f _ {2} \\| _ {2} ^ {2}), \\tag {13}", + "image_path": "eb056450f3971fd04d3a7e017c5969c0a1ba83fde006fdef496ad09759e9b2bc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 507, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 296, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 507, + 296, + 544 + ], + "type": "inline_equation", + "content": "F = \\{f_{\\mathrm{V}}(x^i)\\}_{i=1}^{N_{\\mathrm{test}}} \\cup \\{f_{\\mathrm{T}}(t^i)\\}_{i=1}^{N_{\\mathrm{test}}}" + }, + { + "bbox": [ + 55, + 507, + 296, + 544 + ], + "type": "text", + "content": "; lower is better. We report the average score for each metric by running experiments three times." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 552, + 222, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 222, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 222, + 564 + ], + "type": "text", + "content": "4.2. Zero-shot Transfer Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "content": "We first demonstrate the zero-shot performance of our CLIP-Refine. Tables 1 and 2 list the zero-shot test performance on classification and cross-modal retrieval tasks after post-pretraining. For zero-shot classification tasks, post-pre-training by the contrastive loss (i.e., Contrastive and " + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "m^2" + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "content": "-mix) largely degraded the top-1 accuracy performance for all datasets. This can be caused by catastrophic forgetting since Self-KD did not show performance degradation. One cause of catastrophic forgetting is a smaller batchsize than at pre-training; the OpenAI's CLIP uses a batchsize of 32,768 with hundreds of GPUs [40], whereas our setting uses ones of 512 with a single GPU. Since a smaller batch contains a limited number" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 188, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 308 + ], + "type": "text", + "content": "of negative samples, contrastive learning tends to overfit and catastrophically forget past knowledge. We discuss more details about the effect of batchsize in Sec. 4.4.4. In contrast, our CLIP-Refine impressively improved the zero-shot performance for all datasets. The ablation studies of HyCD show that the performance gain comes from the combination of HyCD with RaFA. Notably, " + }, + { + "bbox": [ + 313, + 188, + 555, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{HyCD} + \\mathcal{L}_{\\mathrm{align}}" + }, + { + "bbox": [ + 313, + 188, + 555, + 308 + ], + "type": "text", + "content": " resulted in the worst performance among the baselines, suggesting that naively minimizing the gap between image and text features leads to the degradations of generalization performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "text", + "content": "For zero-shot retrieval, our CLIP-Refine achieved the best performance with a larger margin than that in classification. Interestingly, in contrast to classification tasks, we found that Contrastive and " + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "inline_equation", + "content": "m^2" + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "text", + "content": "-mix improved the pre-trained models in the " + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathrm{T} \\rightarrow \\mathrm{I}" + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "text", + "content": " retrieval tasks while the " + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathrm{I} \\rightarrow \\mathrm{T}" + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "text", + "content": " performance was degraded. This indicates that the image encoders are more prone to overfitting than the text encoders. This can be explained by the difference between image and text in the data spaces; images are represented by continuous-valued pixels and thus have a high degree of freedom, while texts are represented by discrete tokens in vocabulary and thus have a limited degree of freedom. Therefore, the image encoder tends to overfit to a limited number of negative samples in small batches, and the text encoder can learn new knowledge from them while keeping the past knowledge. Since the zero-shot classification is a " + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathrm{I} \\rightarrow \\mathrm{T}" + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "text", + "content": " retrieval task, this can also be a reason for the performance degradation in classification. In this perspective, our CLIP-Refine successfully prevents the overfitting in " + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathrm{I} \\rightarrow \\mathrm{T}" + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "text", + "content": " and greatly improves the performance in both " + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathrm{I} \\rightarrow \\mathrm{T}" + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathrm{T} \\rightarrow \\mathrm{I}" + }, + { + "bbox": [ + 313, + 309, + 556, + 549 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 562, + 457, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 457, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 457, + 575 + ], + "type": "text", + "content": "4.3. Analysis of Feature Space" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "In this section, we aim to answer our fundamental research question: how much can we alleviate the modality gap by post-pre-training with CLIP-Refine? For quantitative evaluations, we summarize the modality gap, alignment, and uniformity scores in Table 3. HyCD " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "+\\mathcal{L}_{\\mathrm{align}}" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": " achieved the best modality gap and alignment scores. However, it significantly degraded the uniformity score, indicating that directly minimizing " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{align}}" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": " can destroy feature space well constructed by the CLIP training. In contrast, our CLIP-Refine consistently improved all of the scores. This suggests that it is important to minimize not only the gap between image and text features" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4261" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 184, + 159 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 184, + 159 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 184, + 159 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 184, + 159 + ], + "type": "image", + "image_path": "175c129fbc52c2a0c63eed8aa993b0a054828eaf1477152aaa892f5bd87f894a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 161, + 147, + 171 + ], + "lines": [ + { + "bbox": [ + 99, + 161, + 147, + 171 + ], + "spans": [ + { + "bbox": [ + 99, + 161, + 147, + 171 + ], + "type": "text", + "content": "(a) Pre-trained" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 192, + 69, + 315, + 159 + ], + "blocks": [ + { + "bbox": [ + 192, + 69, + 315, + 159 + ], + "lines": [ + { + "bbox": [ + 192, + 69, + 315, + 159 + ], + "spans": [ + { + "bbox": [ + 192, + 69, + 315, + 159 + ], + "type": "image", + "image_path": "00546a37b2fd7af3859b1c13b57b7b7d1f25749cfa11b66e0d0dadd2f512e183.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 161, + 279, + 171 + ], + "lines": [ + { + "bbox": [ + 229, + 161, + 279, + 171 + ], + "spans": [ + { + "bbox": [ + 229, + 161, + 279, + 171 + ], + "type": "text", + "content": "(b) Contrastive" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 180, + 555, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 555, + 204 + ], + "type": "text", + "content": "Figure 3. PCA visualization of multi-modal feature spaces of trained models (Flickr8K, ViT-B/32). CLIP-Refine reduces the modality gap while keeping the form of feature distributions for each modality." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 323, + 69, + 446, + 159 + ], + "blocks": [ + { + "bbox": [ + 323, + 69, + 446, + 159 + ], + "lines": [ + { + "bbox": [ + 323, + 69, + 446, + 159 + ], + "spans": [ + { + "bbox": [ + 323, + 69, + 446, + 159 + ], + "type": "image", + "image_path": "2ae6438814ad3ee5a2154ec0f2b995be71ee0d5f396709cf53a1af220f07b791.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 161, + 413, + 171 + ], + "lines": [ + { + "bbox": [ + 354, + 161, + 413, + 171 + ], + "spans": [ + { + "bbox": [ + 354, + 161, + 413, + 171 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 354, + 161, + 413, + 171 + ], + "type": "inline_equation", + "content": "\\mathrm{HyCD} + \\mathcal{L}_{\\mathrm{align}}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 454, + 69, + 577, + 159 + ], + "blocks": [ + { + "bbox": [ + 454, + 69, + 577, + 159 + ], + "lines": [ + { + "bbox": [ + 454, + 69, + 577, + 159 + ], + "spans": [ + { + "bbox": [ + 454, + 69, + 577, + 159 + ], + "type": "image", + "image_path": "d0a2a744612d987be37aa739e461e7625d8b0f996e00a866734476cbca987f85.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 161, + 553, + 171 + ], + "lines": [ + { + "bbox": [ + 477, + 161, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 477, + 161, + 553, + 171 + ], + "type": "text", + "content": "(d) CLIP-Refine (Ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 57, + 233, + 294, + 309 + ], + "blocks": [ + { + "bbox": [ + 55, + 214, + 294, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 294, + 225 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation of features (Flickr8K, ViT-B/32)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 233, + 294, + 309 + ], + "lines": [ + { + "bbox": [ + 57, + 233, + 294, + 309 + ], + "spans": [ + { + "bbox": [ + 57, + 233, + 294, + 309 + ], + "type": "table", + "html": "
MethodModality Gap ×10-3(↓)Alignment (↓)Uniformity (↓)
Pre-trained1.33451.37240.0895
Contrastive2.17141.43000.2136
m2-mix [36]2.09111.41300.2019
Self-KD1.27541.37740.0901
HyCD1.38231.35780.0971
HyCD+L align0.15900.74130.1172
CLIP-Refine (Ours)0.79341.28490.0495
", + "image_path": "8497bdc23c48c62be6f6e251c50e1e044d7ba7db1c2b2213c9e8366e07212fa3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 320, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 320, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 54, + 320, + 296, + 416 + ], + "type": "text", + "content": "but also the uniformity for improving the generalization performance of pre-trained CLIP models, which is consistent with the findings in [53]. In such a sense, since CLIP-Refine penalizes the image and text features to follow the shared reference distribution " + }, + { + "bbox": [ + 54, + 320, + 296, + 416 + ], + "type": "inline_equation", + "content": "p(z) = \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 54, + 320, + 296, + 416 + ], + "type": "text", + "content": ", where the samples tend to be uniformly distributed on the hypersphere [1], it can naturally achieve a good balance of alignment and uniformity, and reasonably helps to reduce the modality gap." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 417, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 417, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 54, + 417, + 296, + 538 + ], + "type": "text", + "content": "As a qualitative evaluation, Figure 3 illustrates the PCA visualization of post-pre-trained features for each method. We can see that the contrastive model makes the modality gap rather large. " + }, + { + "bbox": [ + 54, + 417, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{HyCD} + \\mathcal{L}_{\\mathrm{align}}" + }, + { + "bbox": [ + 54, + 417, + 296, + 538 + ], + "type": "text", + "content": " succeeds in gathering the image and text features into the same cluster but distorts the feature distribution for each modality. Meanwhile, CLIP-Refine certainly reduces the modality gap while keeping the shape of the pre-trained features. Although the modality gap is not completely zero, it is natural because there remains a constant value in the lower bound of the modality gap [39]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 546, + 160, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 160, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 160, + 559 + ], + "type": "text", + "content": "4.4. Detailed Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 564, + 165, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 165, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 165, + 575 + ], + "type": "text", + "content": "4.4.1. Prior Distributions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": "Here, we analyze CLIP-Refine by varying the prior distribution " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": ". According to the prior works [59, 67], we tried standard Gaussian " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": ", uniform distribution " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "U(0,1)" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": ", Gaussian distributions using pre-trained text feature statistics " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_{\\mathrm{txt}}^{\\mathrm{P}},(\\sigma_{\\mathrm{txt}}^{\\mathrm{P}})^2)" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": ", image feature statistics " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_{\\mathrm{img}}^{\\mathrm{P}},(\\sigma_{\\mathrm{img}}^{\\mathrm{P}})^2)" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": ", statistics of all image and text features " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_{\\mathrm{all}}^{\\mathrm{P}},(\\sigma_{\\mathrm{all}}^{\\mathrm{P}})^2)" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mu^{\\mathrm{P}},\\sigma^{\\mathrm{P}}" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": " is mean and variance of image or/and text feature vectors computed on a pre-trained CLIP model. We also tested variations of " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\beta I)" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": " by varying " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\{0,0.01,0.1,10,100\\}" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": " means RaFA without randomness, i.e., " + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\frac{1}{2} (\\| z_{\\mathrm{img}}^{i}\\|_{2}^{2} + \\| z_{\\mathrm{txt}}^{i}\\|_{2}^{2})" + }, + { + "bbox": [ + 55, + 580, + 297, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 328, + 255, + 539, + 399 + ], + "blocks": [ + { + "bbox": [ + 313, + 214, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 247 + ], + "type": "text", + "content": "Table 4. Evaluation of CLIP-Refine varying prior distributions in " + }, + { + "bbox": [ + 313, + 214, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RaFA}}" + }, + { + "bbox": [ + 313, + 214, + 555, + 247 + ], + "type": "text", + "content": ". ZS Cls. is the zero-shot classification accuracy averaged on 12 datasets." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 255, + 539, + 399 + ], + "lines": [ + { + "bbox": [ + 328, + 255, + 539, + 399 + ], + "spans": [ + { + "bbox": [ + 328, + 255, + 539, + 399 + ], + "type": "table", + "html": "
PriorZS Cls.AlignmentUniformity
Pretrained52.741.37240.0895
U(0,1)52.911.30810.0661
N(μpimg, (σpimg)2)52.931.31420.0916
N(μptxt, (σptxt)2)53.401.33040.0873
N(μpall, (σpall)2)53.291.33610.0888
N(0,0I)53.791.30520.0554
N(0,0.01I)54.191.30790.0544
N(0,0.1I)54.321.30210.0515
N(0,I)54.691.28490.0495
N(0,10I)54.231.30850.0537
N(0,100I)53.591.34410.0663
", + "image_path": "7562a72399eecbc15fd57a1839dfb4939170c3c69899213bfa632d12432d3105.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "content": "Table 4 shows the zero-shot classification performance, alignment, and uniformity scores when varying " + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "content": ". Using the uniform distribution " + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "inline_equation", + "content": "U(0,1)" + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "content": " did not show large improvements on the zero-shot performance and uniformity score. This may be because random reference vectors from " + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "inline_equation", + "content": "U(0,1)" + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "content": " do not necessarily concentrate on the hypersphere in contrast to " + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "content": ", possibly causing the collapse of uniformity by " + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RaFA}}" + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "content": ". In this perspective, we recommend using " + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "content": " or families of Gaussian distributions. Interestingly, among the Gaussian distributions with the pre-trained statistics, the prior with the text feature, i.e., " + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_{\\mathrm{txt}}^{\\mathrm{p}}(\\sigma_{\\mathrm{txt}}^{\\mathrm{p}})^2)" + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "content": " largely improved the zero-shot performance. This indicates that aligning image features with text features is effective in zero-shot classification tasks and also evidences the performance improvements in the " + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{I} \\rightarrow \\mathrm{T}" + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "content": " tasks observed in Sec. 4.2. More importantly, the ablation studies of the variance magnitude " + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "content": " in the zero-mean Gaussian distributions emphasize the importance of randomness in " + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RaFA}}" + }, + { + "bbox": [ + 313, + 411, + 556, + 626 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 637, + 455, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 455, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 455, + 649 + ], + "type": "text", + "content": "4.4.2. Post-pre-training Datasets" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 653, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 714 + ], + "type": "text", + "content": "We evaluate the effects of the dataset choice in post-pretraining. Table 5 shows the zero-shot classification/retrieval performance of CLIP-Refine with Flickr8K/30K, COCO Caption, CC3M, and CC12M as post-pre-training datasets. We confirm that most of the CLIP-Refine models outper" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4262" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 113, + 294, + 206 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 5. Zero-shot classification (12 datasets) and retrieval (Flicker30K) performance of CLIP-Refine with various post-pre-training datasets with ViT-B/32." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 113, + 294, + 206 + ], + "lines": [ + { + "bbox": [ + 56, + 113, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 113, + 294, + 206 + ], + "type": "table", + "html": "
Post-pre-training Dataset# of SamplesZero-shot Cls. Acc.Zero-shot Ret. R@5
Avg. of 12 datasetsT→II→T
None (Pre-trained)052.7440.9644.57
Flickr8K8K53.0445.5746.80
Flickr30K30K53.7448.6649.25
COCO Caption118K54.6949.7051.54
CC3M3M53.8049.5650.24
CC12M12M52.4645.3547.79
Filtered CC3M1M55.5052.4553.00
Filtered CC12M4M60.1254.9757.88
", + "image_path": "45c18597a686633e8bf55b8e7293b010c40be81348217b2bd1d7d1ae4899b6b8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 248, + 294, + 304 + ], + "blocks": [ + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "type": "text", + "content": "Table 6. Zero-shot classification (12 datasets) and retrieval (Flickr30K) performance with various pre-trained CLIP models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 248, + 294, + 304 + ], + "lines": [ + { + "bbox": [ + 56, + 248, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 248, + 294, + 304 + ], + "type": "table", + "html": "
ModelZero-shot Cls. Acc.Zero-shot Ret. R@5 (T→I/I→T)
Pre-trainedContrastiveCLIP-RefinePre-trainedContrastiveCLIP-Refine
ViT-B/3252.7445.7654.6940.96/44.5747.32/36.2949.70/51.54
ViT-L/1462.9661.9063.8849.13/52.7751.07/51.0058.38/58.41
OpenCLIP ViT-H/1468.5257.6169.0565.84/66.0257.47/43.4367.82/68.48
OpenCLIP ViT-G/1471.2857.4472.5968.73/68.3866.35/42.3670.94/69.76
SigLip ViT-SO400M/1472.5765.6373.6172.77/72.8364.63/62.9776.54/76.49
DFN ViT-H/1477.3076.3978.2971.93/70.8972.81/70.0976.39/75.44
", + "image_path": "b1ba50ac5d464808ded963c825f02f8e0c37d48f9e4fc939a23bd86abda53968.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 311, + 297, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 311, + 297, + 468 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 297, + 468 + ], + "type": "text", + "content": "formed the pre-trained baselines. Notably, larger datasets do not always achieve high performance; CC3M and CC12M are inferior to COCO Caption. This suggests that the quality of image-text pairs in a dataset is an important factor in post-pre-training. Since CC3M and CC12M contain mismatched and noisy image-text pairs [17], COCO Caption that has higher-quality text captions can be superior in terms of the modality alignment of pre-trained CLIP models. In fact, when we filtered out low-quality image-text pairs by CLIP-Score-based filtering of DataComp [13], the models achieved significant performance improvements. This indicates that the caption quality is crucial and our method is scalable if we have high-quality image-text pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 167, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 167, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 167, + 484 + ], + "type": "text", + "content": "4.4.3. Pre-trained Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 487, + 297, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 297, + 584 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 297, + 584 + ], + "type": "text", + "content": "We investigate the generalizability of CLIP-Refine to pretrained CLIP models. Table 6 shows the zero-shot performance with 6 pre-trained models. We observed that our CLIP-Refine stably improved the pre-trained baseline. This suggests that the bigger and stronger models still have the modality gap, and modifying the gap by CLIP-Refine helps improve the generalization performance for various pretrained models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 590, + 157, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 157, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 157, + 601 + ], + "type": "text", + "content": "4.4.4. Mini-Batch Sizes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "content": "We provide the analysis of varying mini-batch sizes in post-pre-training. We tried the mini-batch sizes in " + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\{32, 64, 128, 256, 512, 1024, 2048\\}" + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "content": ". We used gradient accumulation by following the OpenCLIP [8] implementation for the mini-batch sizes of 1024 and 2048. Figure 4 shows the zero-shot performance versus mini-batch sizes. The contrastive loss gradually improved the accuracy according to the mini-batch size. However, the accuracy curve implies that it is difficult to exceed the pre-trained baseline with" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 350, + 69, + 521, + 162 + ], + "blocks": [ + { + "bbox": [ + 350, + 69, + 521, + 162 + ], + "lines": [ + { + "bbox": [ + 350, + 69, + 521, + 162 + ], + "spans": [ + { + "bbox": [ + 350, + 69, + 521, + 162 + ], + "type": "image", + "image_path": "065f34356cbd634a4a0697963f5ad2f05b36a7e0a5d810dc0b7196fe5c3925bb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 171, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 194 + ], + "type": "text", + "content": "Figure 4. Zero-shot classification accuracy averaged on 12 datasets when varying mini-batch sizes in post-pre-training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 339, + 216, + 527, + 247 + ], + "blocks": [ + { + "bbox": [ + 313, + 196, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 196, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 555, + 208 + ], + "type": "text", + "content": "Table 7. Fine-tuning top-1 accuracy on the ImageNet validation set." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 216, + 527, + 247 + ], + "lines": [ + { + "bbox": [ + 339, + 216, + 527, + 247 + ], + "spans": [ + { + "bbox": [ + 339, + 216, + 527, + 247 + ], + "type": "table", + "html": "
ModelPre-trainedContrastiveCLIP-Refine
ViT-B/3273.4069.4674.14
", + "image_path": "3df8148a21e082750f7ef6e5e2195b204f6969026675727d1bc1a0387e442e80.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 252, + 556, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 556, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 556, + 373 + ], + "type": "text", + "content": "feasible mini-batch sizes. This is probably due to the large discrepancy from the mini-batch size in the pre-training (i.e., 32,768), which causes the model to forget knowledge in the predictions for negative pairs in a mini-batch. In contrast, CLIP-Refine always outperformed the contrastive loss and succeeded in improving the pre-trained baseline with feasible mini-batch sizes. This suggests that CLIP-Refine's HyCD prevents knowledge forgetting even for small mini-batch sizes with few negative pairs and is able to modify the feature space with new knowledge." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 378, + 440, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 440, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 440, + 391 + ], + "type": "text", + "content": "4.4.5. Effects on Fine-tuning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 393, + 556, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 556, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 556, + 466 + ], + "type": "text", + "content": "We show the transferability of CLIP-Refine to fine-tuning. We fine-tuned the linear classification head on ImageNet while freezing post-pre-trained models. Tabel 7 shows the results. CLIP-Refine outperformed the pre-trained and contrastive loss baselines. This emphasizes that our method can generate useful representation even in fine-tuning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 496, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 556, + 712 + ], + "type": "text", + "content": "This paper presented CLIP-Refine, a post-pre-training method for pre-trained CLIP models to align the modality gap between the image and text features. CLIP-Refine addresses the modality gap by penalizing the multi-modal features to follow a shared prior distribution by minimizing the distance to the random reference vectors sampled from the prior. To maintain the past knowledge in the CLIP models and promote the feature alignment simultaneously, CLIP-Refine also trains the model with knowledge distillation loss using hybrid soft labels composed of ground-truth image-text pair labels and outputs from the pre-trained CLIP model. Through extensive experiments, we show that CLIP-Refine can improve the zero-shot performance of the pre-trained CLIP by addressing the modality gap and enhancing uniformity. We believe that our work not only provides a practical method but also opens up a new research field where we refine the pre-trained CLIP models by post-pre-training with much smaller computational costs than that of pre-training." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4263" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Christopher M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag, Berlin, Heidelberg, 2006. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "type": "text", + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 296, + 227 + ], + "type": "text", + "content": "[3] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 295, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 295, + 283 + ], + "type": "text", + "content": "[4] Lianggangxu Chen, Xuejiao Wang, Jiale Lu, Shaohui Lin, Changbo Wang, and Gaoqi He. Clip-driven open-vocabulary 3d scene graph generation via cross-modality contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 350 + ], + "type": "text", + "content": "[5] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yu Hou, Yuenan and Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 296, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 296, + 428 + ], + "type": "text", + "content": "[6] Yuxiao Chen, Jianbo Yuan, Yu Tian, Shijie Geng, Xinyu Li, Ding Zhou, Dimitris N Metaxas, and Hongxia Yang. Revisiting multimodal representation in contrastive learning: from patch and token embeddings to finite discrete tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15095-15104, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "type": "text", + "content": "[7] Jun Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 477, + 296, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 296, + 553 + ], + "type": "text", + "content": "[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 2, 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 556, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 296, + 600 + ], + "type": "text", + "content": "[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2014. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "type": "text", + "content": "[10] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. In International Conference on Learning Representations, 2024. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[12] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In Conference on CVPR Workshop, 2004. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "text", + "content": "[13] Gadre et al. Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 153, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 555, + 217 + ], + "type": "text", + "content": "[14] Yunhao Ge, Jie Ren, Andrew Gallagher, Yuxiao Wang, Ming-Hsuan Yang, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11093-11101, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 262 + ], + "type": "text", + "content": "[15] Shashank Goel, Hritik Bansal, Sumit Bhatia, Ryan Rossi, Vishwa Vinay, and Aditya Grover. Cyclip: Cyclic contrastive language-image pretraining. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 265, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 265, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 265, + 555, + 319 + ], + "type": "text", + "content": "[16] Haochen Han, Qinghua Zheng, Guang Dai, Minnan Luo, and Jingdong Wang. Learning to rematch mismatched pairs for robust cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 321, + 555, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 555, + 375 + ], + "type": "text", + "content": "[17] Haochen Han, Qinghua Zheng, Guang Dai, Minnan Luo, and Jingdong Wang. Learning to rematch mismatched pairs for robust cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26679-26688, 2024. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 377, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 554, + 431 + ], + "type": "text", + "content": "[18] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 434, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 555, + 465 + ], + "type": "text", + "content": "[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 468, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 555, + 510 + ], + "type": "text", + "content": "[20] Ahmet Iscen, Mathilde Caron, Alireza Fathi, and Cordelia Schmid. Retrieval-enhanced contrastive vision-text models. In International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 512, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 555, + 567 + ], + "type": "text", + "content": "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, 2021. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 569, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 555, + 612 + ], + "type": "text", + "content": "[22] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision. Springer, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 614, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 555, + 668 + ], + "type": "text", + "content": "[23] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multimodal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition, Sydney, Australia, 2013. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4264" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 139 + ], + "type": "text", + "content": "[25] Zhengfeng Lai, Noranart Vesdapunt, Ning Zhou, Jun Wu, Cong Phuoc Huynh, Xuelu Li, Kah Kuen Fu, and Chen-nee Chuah. Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "type": "text", + "content": "[26] Janghyeon Lee, Jongsuk Kim, Hyounguk Shon, Bumsoo Kim, Seung Hwan Kim, Honglak Lee, and Junmo Kim. Uniclip: Unified framework for contrastive language-image pre-training. Advances in Neural Information Processing Systems, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 251 + ], + "type": "text", + "content": "[27] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In International Conference on Learning Representations, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 295, + 306 + ], + "type": "text", + "content": "[28] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 2022. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 295, + 361 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision. Springer, 2014. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 364, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 295, + 397 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 442 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 295, + 464 + ], + "type": "text", + "content": "[32] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "type": "text", + "content": "[33] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 511, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 295, + 543 + ], + "type": "text", + "content": "[34] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. arXiv, 2013. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "type": "text", + "content": "[35] M-E. Nilsback and A. Zisserman. Automated flower classification over a large number of classes. In Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "type": "text", + "content": "[36] Changdae Oh, Junhyuk So, Hoyoon Byun, YongTaek Lim, Minchul Shin, Jong-June Jeon, and Kyungwoo Song. Geodesic multi-modal mixup for robust fine-tuning. In Advances in Neural Information Processing Systems, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "type": "text", + "content": "[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "type": "text", + "content": "[38] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[39] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In Advances in Neural Information Processing Systems, 2024. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 555, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 182 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 555, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 555, + 226 + ], + "type": "text", + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 282 + ], + "type": "text", + "content": "[42] Cyrus Rashtchian, Peter Young, Micah Hodosh, and Julia Hockenmaier. Collecting image annotations using amazon's mechanical turk. In Proceedings of the NAACL HLT 2010 workshop on creating speech and language data with Amazon's Mechanical Turk, pages 139-147, 2010. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 284, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 555, + 327 + ], + "type": "text", + "content": "[43] Arijit Ray, Filip Radenovic, Abhimanyu Dubey, Bryan Plummer, Ranjay Krishna, and Kate Saenko. Cola: A benchmark for compositional text-to-image retrieval. In Advances in Neural Information Processing Systems, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 328, + 555, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 555, + 381 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 555, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 555, + 437 + ], + "type": "text", + "content": "[45] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3), 2015. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 438, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 555, + 491 + ], + "type": "text", + "content": "[46] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 555, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 555, + 547 + ], + "type": "text", + "content": "[47] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, 2018. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 602 + ], + "type": "text", + "content": "[48] Sheng Shen, Shijia Yang, Tianjun Zhang, Bohan Zhai, Joseph E Gonzalez, Kurt Keutzer, and Trevor Darrell. Multitask vision-language prompt tuning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 604, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 555, + 658 + ], + "type": "text", + "content": "[49] Sicong Shen, Yang Zhou, Bingzheng Wei, Eric I Chang, Yan Xu, et al. Tuning stable rank shrinkage: Aiming at the overlooked structural risk in fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 659, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 555, + 689 + ], + "type": "text", + "content": "[50] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[51] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4265" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3835-3844, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[52] Jingyun Wang and Guoliang Kang. Learn to rectify the bias of clip for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4102-4112, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 208 + ], + "type": "text", + "content": "[53] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 253 + ], + "type": "text", + "content": "[54] Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 298 + ], + "type": "text", + "content": "[55] Zhengbo Wang, Jian Liang, Ran He, Nan Xu, Zilei Wang, and Tieniu Tan. Improving zero-shot generalization for clip with synthesized prompts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 296, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 296, + 332 + ], + "type": "text", + "content": "[56] P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD Birds 200. Technical report, California Institute of Technology, 2010. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 296, + 388 + ], + "type": "text", + "content": "[57] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 296, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 455 + ], + "type": "text", + "content": "[58] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 511 + ], + "type": "text", + "content": "[59] Shin'ya Yamaguchi, Sekitoshi Kanai, Kazuki Adachi, and Daiki Chijiwa. Adaptive random feature regularization on fine-tuning deep neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 577 + ], + "type": "text", + "content": "[60] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. Clipkd: An empirical study of clip model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15952-15962, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 296, + 635 + ], + "type": "text", + "content": "[61] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19163-19173, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "type": "text", + "content": "[62] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[63] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 387 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "type": "text", + "content": "Chunjing Xu. FILIP: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "text", + "content": "[64] Haoxuan You, Luowei Zhou, Bin Xiao, Noel Codella, Yu Cheng, Ruochen Xu, Shih-Fu Chang, and Lu Yuan. Learning visual representation from modality-shared contrastive language-image pre-training. In European Conference on Computer Vision. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 207 + ], + "type": "text", + "content": "[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 555, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 555, + 261 + ], + "type": "text", + "content": "[66] Yuhui Zhang, Jeff Z. HaoChen, Shih-Cheng Huang, Kuan-Chieh Wang, James Zou, and Serena Yeung. Diagnosing and rectifying vision models using language. In The Eleventh International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 555, + 308 + ], + "type": "text", + "content": "[67] Yang Zhong and Atsuto Maki. Regularizing cnn transfer learning with randomised regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13637-13646, 2020. 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 342 + ], + "type": "text", + "content": "[68] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 2022. 1, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 555, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 555, + 387 + ], + "type": "text", + "content": "[69] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4266" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/b07ff10a-9a3d-44d4-80c1-c5cc5cfdc9e1_content_list.json b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/b07ff10a-9a3d-44d4-80c1-c5cc5cfdc9e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..88cf6692bfe33bb3bcbbe21050e567bde65df9ec --- /dev/null +++ b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/b07ff10a-9a3d-44d4-80c1-c5cc5cfdc9e1_content_list.json @@ -0,0 +1,1817 @@ +[ + { + "type": "text", + "text": "PosterMaker: Towards High-Quality Product Poster Generation with Accurate Text Rendering", + "text_level": 1, + "bbox": [ + 173, + 130, + 823, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifan $\\mathrm{Gao}^{1,2*^{\\dagger}}$ , Zihang Lin $^{2*}$ , Chuanbin Liu $^{1\\ddagger}$ , Min Zhou $^{2}$", + "bbox": [ + 261, + 202, + 733, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tiezheng Ge², Bo Zheng², Hongtao Xie¹", + "bbox": [ + 328, + 220, + 668, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China 2Taubao & Tmall Group of Alibaba", + "bbox": [ + 161, + 238, + 834, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eafn@mail.ustc.edu.cn {liucb92, hxtie}@ustc.edu.cn", + "bbox": [ + 272, + 258, + 723, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{linzihang.lzh, yunqi.zm, tiezheng.gtz, bozheng}@alibaba-inc.com", + "bbox": [ + 215, + 277, + 781, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page: https://poster-maker.github.io", + "bbox": [ + 297, + 295, + 699, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt", + "text_level": 1, + "bbox": [ + 163, + 311, + 218, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The box of fish oil supplements is placed on a wooden table, with a background of a serene ocean and clear sky, symbolizing purity and the natural source of the product", + "bbox": [ + 116, + 335, + 256, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The subject rests on a smooth, dark wooden table, surrounded by a few scattered leaves and delicate flowers, with a serene garden scene complete with blooming flowers and lush greenery in the background.", + "bbox": [ + 116, + 428, + 256, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Subject", + "text_level": 1, + "bbox": [ + 272, + 311, + 331, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c39803913d39788b94ec04fef3ac80e74421a6198196bc02a5bd02b718fd765a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 344, + 331, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75947218702e92302820cb9baa7de9463820ef706e74c7e044188115851dd1f4.jpg", + "image_caption": [ + "Subject", + "Text" + ], + "image_footnote": [], + "bbox": [ + 349, + 325, + 460, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f724ce88fd72832683cb32c80146975374055324d5acc373f59b6e617fbf87a7.jpg", + "image_caption": [ + "Poster" + ], + "image_footnote": [], + "bbox": [ + 462, + 325, + 575, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36d5b12e25f4ec3efd638bd460c8de49074587021651dd875e78639bc38e7983.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 352, + 658, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/42180260bd03b80fb04721970272421b10653b361ab2ab1290b9295303dae9ae.jpg", + "image_caption": [ + "", + "Sence Generation" + ], + "image_footnote": [], + "bbox": [ + 661, + 327, + 750, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3cf9e98e53396d396a582892f6dc825377942a392dcafad67beabd74181c5a54.jpg", + "image_caption": [ + "Previous: two stage", + "Text Rendering" + ], + "image_footnote": [], + "bbox": [ + 774, + 325, + 867, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71e69ab1def6963bfd89c44458898882e15fca0e09fcc0c0e58bdb2d3b2b242c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 414, + 449, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(a)", + "bbox": [ + 323, + 515, + 338, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db85dda03cb016ad13a61f099800a9ec0edbaf73e499f6d6e661672628c7ab82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 412, + 573, + 526 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f7e1363ad398d6c802d00a1eb780e597a51affa9a5c1649ab970300e87fb5b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 455, + 658, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a42a935afba24f668b33dcbf3435955c67ba4285a34a9f47c87e7b64a90ba578.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 455, + 754, + 497 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee9f4c06ad5b211cb99a39bc982efb594ad0fb928f3dac715280a44621d82256.jpg", + "image_caption": [ + "Ours: end to end", + "End2End Generation" + ], + "image_footnote": [], + "bbox": [ + 750, + 438, + 867, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/55d0a972d63f570075dc671b97f11b22dddc3ebfdf8f32121bef7a2edcc13f63.jpg", + "image_caption": [ + "Figure 1. (a) Definition of the advertising product poster generation task. The input includes the prompt, subject image, and the texts to be rendered with their layouts. The output is the poster image. (b) The comparison of our method with the previous method. PosterMaker generates posters end-to-end, while previous methods first generate poster backgrounds and then render texts. (c) Visualization results demonstrate that PosterMaker can generate harmonious and aesthetically pleasing posters with accurate texts and maintain subject fidelity." + ], + "image_footnote": [], + "bbox": [ + 114, + 527, + 292, + 734 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbfd7bc0afe4f41a64333017094c1a0270611b5dad8340758c94680f87e94d1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 446, + 348, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(a)", + "bbox": [ + 323, + 515, + 338, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f098cc62326e39aa571b819f3e0889cf597a922e2ee525f1fd2454e54daef7f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 530, + 428, + 736 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/678f30e21ce9dfd0a82a8949952406ab66063d64f8085b3dce4c7fc308546fa8.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 429, + 530, + 560, + 734 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed7095cd8c8bd5b29e57ad87f9a250a42ebc2c7348274ff0742b615a0c323821.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 530, + 696, + 736 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/84c47e8a1266912a9a1442a5a20ad3930e2be592fcf338d890e6dc43a6150735.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 530, + 875, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 806, + 326, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Product posters, which integrate subject, scene, and text, are crucial promotional tools for attracting customers. Cre", + "bbox": [ + 89, + 833, + 482, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ating such posters using modern image generation methods is valuable, while the main challenge lies in accurately rendering text, especially for complex writing systems like Chinese, which contains over 10,000 individual characters.", + "bbox": [ + 511, + 811, + 906, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we identify the key to precise text rendering", + "bbox": [ + 514, + 871, + 903, + 887 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 31 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. ‡ Corresponding author.", + "bbox": [ + 112, + 875, + 362, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Work done during the internship at Alibaba Group.", + "bbox": [ + 112, + 888, + 393, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8083", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as constructing a character-discriminative visual feature as a control signal. Based on this insight, we propose a robust character-wise representation as control and we develop TextRenderNet, which achieves a high text rendering accuracy of over $90\\%$ . Another challenge in poster generation is maintaining the fidelity of user-specific products. We address this by introducing SceneGenNet, an inpainting-based model, and propose subject fidelity feedback learning to further enhance fidelity. Based on TextRenderNet and SceneGenNet, we present PosterMaker, an end-to-end generation framework. To optimize PosterMaker efficiently, we implement a two-stage training strategy that decouples text rendering and background generation learning. Experimental results show that PosterMaker outperforms existing baselines by a remarkable margin, which demonstrates its effectiveness.", + "bbox": [ + 89, + 90, + 480, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 345, + 222, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Product posters, which showcase items for sale within well-chosen background scenes and include descriptive text, play a vital role in e-commerce advertising by capturing customers' attention and boosting sales. Creating such posters necessitates photographing the product in carefully selected environments that highlight its features, as well as thoughtfully choosing text colors and fonts to ensure that the text is appealing, legible, and harmonious with the background. This process can be quite expensive. With the significant advancements in large-scale text-to-image (T2I) models [12, 31, 35], synthesizing such product posters with image generation models attracts increasing attention. In this paper, we focus on the product poster generation task. Specifically, given a prompt describing the background scene, the foreground image of the user-specified subject and some texts together with their layouts, we aim to develop a model to generate the subject into the desired scene background and accurately render the text in an end-to-end manner (as shown in Fig. 1 (a)).", + "bbox": [ + 89, + 369, + 482, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A straightforward solution for this task is to first generate the subject into the desired scene [2, 10, 36], and then predict the text attributes (such as color and font) [13, 22] and render them on the image. However, such two-stage approach suffers from disharmony between the text and the poster background(as shown in Fig. 2 (b)). And collecting training data is also challenging since the text attributes, especially the text font, are difficult to extract from the poster. Another solution is learning to generate the poster using a per-pixel synthesis approach, which can benefit from directly learning the distribution of professionally designed posters. We focus on such one-stage solution. The main challenge is how to ensure the text rendering accuracy.", + "bbox": [ + 89, + 659, + 482, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many recent works [12, 23, 38, 45] have been proposed to improve the text rendering accuracy for large diffusion models. Great progress has been made and some", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "recent work can achieve high rendering accuracy for English. However, for non-Latin languages like Chinese, one of the most widely spoken languages, achieving high rendering accuracy remains challenging. This difficulty stems from the existence of over 10,000 characters, with Chinese characters characterized by complex and diverse stroke patterns, making it extremely difficult to train a model to memorize the rendering of each individual character. Recent studies [4, 26, 38] have focused on extracting visual features of text as control signals. Typically, these approaches render text lines into glyph images and extract line-level text visual features to guide generation. Nevertheless, line-level visual features often lack the discriminative power to capture character-level visual nuances. To address this limitation, GlyphByT5 [23, 24] introduced a box-level contrastive loss with sophisticated glyph augmentation strategies to enhance character-level discriminativeness, achieving promising results. In this paper, we point out that the key to high-accuracy text rendering lies in constructing character-discriminative visual features as control signals. Specifically, we render each character as a glyph image and extract visual features via a visual encoder. These features are then concatenated with positional embeddings to form a character-level representation. Then we propose TextRenderNet, an SD3 [12] controlnet-like [49] architecture that takes the character-level representation as the control signal to render visual text. Our experiments demonstrate that the proposed character-level representation is effectively capable of achieving accurate text rendering.", + "bbox": [ + 511, + 90, + 906, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the task of poster generation, another important thing is to generate the user-specific subject into a desired scene while keeping high subject fidelity. Recent subject-driven controllable generation methods [36, 40, 47] can synthesize such images, but they still cannot ensure that the user-specified subject is completely consistent in the generated details (e.g., the logo on the product may be inaccurately generated), which could potentially mislead customers. Therefore, we follow poster generation methods [5, 10, 21] to address this task via introducing an inpainting-based module named SceneGenNet. However, we found that even using inpainting methods, subject consistency is not always achieved as the inpainting model sometimes extends the subject shape (as shown in Fig. 2 (a)). Similar phenomenon is also observed in [10, 11]. To address this issue, we elaboratively develop a detector to detect the foreground extension cases. Then we employ the detector as a reward model to train the SceneGenNet via feedback learning for further improving subject fidelity.", + "bbox": [ + 511, + 532, + 908, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Combining the proposed TextRenderNet and SceneGenNet, we develop a framework named PosterMaker that can synthesize the product poster in an end-to-end manner. To efficiently optimize PosterMaker, we introduce a two-stage training strategy to separately train TextRenderNet and Sce", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8084", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3bca4c2fe91c83af5f8bc57bb9f87d5a99b10eb79ad5aef18ca44b9109f43476.jpg", + "image_caption": [ + "User-Specified Texts", + "Figure 2. The illustration of the three challenges faced by poster generation, which seriously hinder the practical application." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 251, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c2bbb6e96fa79a8c0ed43fdbb36306144d76c90eaddc9dc2374e6cab60a705f.jpg", + "image_caption": [ + "User-Specified Subject" + ], + "image_footnote": [], + "bbox": [ + 254, + 90, + 372, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eed9665624db8c382f7b9579f934d5325a2c8fea2f84f3571e373f1646036d4b.jpg", + "image_caption": [ + "(a) Foreground Extension" + ], + "image_footnote": [], + "bbox": [ + 375, + 90, + 496, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bc694b1aafee0ae3795f6d58a09278daa381b8dc389e97f872d6d9b944c81747.jpg", + "image_caption": [ + "(b)Text-Scene Disharmony" + ], + "image_footnote": [], + "bbox": [ + 496, + 90, + 617, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/32f43f9a3f3bbb936aa874f05bf6711eb69c6e2951d13756d160e81588271b22.jpg", + "image_caption": [ + "(c) Poor Text Rendering" + ], + "image_footnote": [], + "bbox": [ + 619, + 90, + 738, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d0cadd493438ccd8585a7d1de28306f44365d9f6578bb23d96ef2a7fdc69f606.jpg", + "image_caption": [ + "Our Result" + ], + "image_footnote": [], + "bbox": [ + 741, + 90, + 862, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "neGenNet. This training strategy decouples the learning of text rendering and background image generation, thus TextRenderNet and SceneGenNet can focus on their specific tasks. Qualitative results (as shown in Fig. 1 (c)) demonstrate our training strategy is effective for training PosterMaker and it achieves promising poster generation results.", + "bbox": [ + 89, + 215, + 483, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows:", + "bbox": [ + 109, + 306, + 426, + 321 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We proposed a novel framework named PosterMaker, which mainly consists of a TextRenderNet and a SceneGenNet. With a two-stage training strategy, PosterMaker can synthesis aesthetically product posters with texts accurately and harmoniously rendered on it.", + "- We reveal the core of achieving accurate Chinese text rendering is to construct a robust character-level text representation as the control condition. These findings can inspire future research on improving the text rendering abilities of T2I models.", + "- We improve the subject fidelity via subject fidelity feedback learning, which is shown effective in addressing the subject inconsistency issue." + ], + "bbox": [ + 89, + 324, + 483, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 532, + 232, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Poster Generation", + "text_level": 1, + "bbox": [ + 89, + 558, + 266, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generating posters involves combining various elements like a subject image, a background scene image, and text to ensure the subject and text are prominently and accurately displayed while maintaining an appealing look. Automating this process is quite complex and challenging. Methods like AutoPoster [22], Prompt2Poster [41], and COLE [15] break it down into stages: creating images and layout, predicting the visual properties of text, and rendering the poster. These approaches have several steps and often struggle to precisely obtain all the necessary visual attributes like font and color gradients. With the emergence of more advanced generative models [31], methods like JoyType [18], Glyphbyt5 [23], and GlyphDraw2 [26] can directly generate the image and text together at the pixel level based on the poster prompt, text content, and layout. This more streamlined approach can leverage more readily available poster pixel data for training, but there is still room for improvement in terms of the overall poster cohesion and text accuracy. Our method is also a one-stage, direct pixel-level generation approach that simultaneously creates the image and text. However, our focus is on generating posters for a given", + "bbox": [ + 89, + 583, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "product subject, where the input includes the subject image, prompt, text content, and layout. In addition to considering text rendering accuracy and overall poster harmony, we also need to maintain the fidelity of the product.", + "bbox": [ + 511, + 215, + 906, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Visual Text Rendering", + "text_level": 1, + "bbox": [ + 511, + 285, + 720, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, text-to-image (T2I) models [1, 12, 37] have made significant strides in enhancing English text rendering by introducing stronger text encoders, such as T5 [34]. However, multilingual text image generation still faces significant challenges due to the large number of non-Latin characters and complex stroke structures. Early work [45] has explored the ControlNet-based method [49], using low-level visual images such as glyph images as the control signal for text image generation. However, glyph images are easily affected by text size and shape, especially complex stroke details. Besides, some recent works [4, 25, 26, 38, 48, 50] utilize more robust visual features, such as line-level OCR features as control conditions to further improve the text accuracy. But the line-level visual features still perform poorly in representing stroke details for each character. To address this issue, GlyphByT5 [23, 24] proposes a method with box-level contrastive learning to align the text features extracted from the language model with the features extracted from the visual encoder. To effectively learn such alignment, GlyphByT5 relies on collecting massive amounts of data and developing complex data augmentation strategies for the alignment pre-training, which lacks flexibility. In contrast, in this paper, we reveal that the key to high-accuracy text rendering lies in constructing discriminative character-level visual features. Thus we propose a plug-and-play and robust character-level text representation derived from off-the-shelf OCR encoders, which can accurately represent the visual structure of the text without additional training and enable precise text rendering.", + "bbox": [ + 511, + 308, + 906, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Subject-Preserved Scene Generation", + "text_level": 1, + "bbox": [ + 511, + 756, + 828, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To create a scene image with a product subject while ensuring subject fidelity, two main methods are commonly used. One is the subject-driven method [3, 6, 19, 32, 36], which adjusts the position, angle and lighting of the subject based on the prompt to create a harmonious image. However, it often struggles to preserve the significant features of the subject. The other utilizes inpainting-based background completion techniques [2, 10, 39]. It only generates the non-", + "bbox": [ + 511, + 780, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8085", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f4b117659f00639c8b4bd3674bba491facc16adbac19530cd80757917def6165.jpg", + "image_caption": [ + "Figure 3. The framework of the PosterMaker, which is based on the SD3. To precisely generate multilingual texts and create aesthetically pleasing poster scenes, TextRenderNet and SenceGenNet are introduced, whose outputs are used as control conditions added to the SD3." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 898, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "subject areas of an image and naturally keeps consistency in the original subject area. But it sometimes extends the foreground subject [10, 11], such as adding an extra handle to a cup, which also reduces subject fidelity. To maximize subject fidelity, our method uses background completion and a reward model to determine whether the foreground extension occurred, thereby enhancing subject fidelity.", + "bbox": [ + 89, + 301, + 483, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 414, + 181, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 439, + 294, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This paper focuses on the creation of product posters, which typically consist of multiple elements such as text, subjects, and scenes, as illustrated in Fig. 1 (a). The central challenge of this task is to generate these elements accurately and harmoniously, offering both research and practical applications. The task is defined as:", + "bbox": [ + 89, + 463, + 483, + 553 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {g} = f \\left(I _ {s}, M _ {s}, T, P\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 565, + 482, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{g}$ denotes the generated poster image, $I_{s}$ represents the subject image, and $M_{s}$ is the subject mask. The variable $T$ signifies the content and the position of text and $P$ is the prompt describing the background scene. Subsequent sections will detail the design of PosterMaker, and our proposed solution to this task.", + "bbox": [ + 89, + 594, + 483, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Framework", + "text_level": 1, + "bbox": [ + 89, + 695, + 218, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, PosterMaker is developed based on Stable Diffusion 3 (SD3) [12], which contains a strong VAE for reconstructing the image details like text stroke. And we propose two modules, i.e., TextRenderNet and SceneGenNet, to address the poster generation task. TextRenderNet is specifically designed to learn visual text rendering, taking character-level visual text representations as input to achieve precise and controllable text rendering. SceneGenNet, on the other hand, accepts a masked image (indicating which content should remain unchanged) and a prompt, learning to generate the foreground subject within the desired scene described by the prompt. Both TextRenderNet", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/042150ef1df01697695703c3c6f7189b7317020c8da26186af947fd013b306a5.jpg", + "image_caption": [ + "Figure 4. The details of TextRenderNet and SceneGenNet, showcasing their model architectures and their interactions with SD3." + ], + "image_footnote": [], + "bbox": [ + 531, + 297, + 893, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and SceneGenNet are grounded in a ControlNet-like [48] architecture derived from SD3 and their architectures are detailed in Fig. 4. They share the same internal structure, comprising several cascaded MM-DiT blocks [12], with weights copied from the base model for initialization. The output of each MM-DiT block is added to the corresponding block of the base model after passing through a zero convolution layer [49]. The key distinction between the two modules lies in their input configurations. SceneGenNet takes the prompt as input to the text condition branch, and for the visual branch, the input is derived by the latent feature at timestep $t$ , the subject mask, and the masked latent to preserve the foreground area. In contrast, TextRenderNet receives text representations (detailed in the next section) in the text condition branch for text rendering. An adapter, consisting of a linear layer and layer normalization, adjusts the feature dimensions of these text representations before they are input to TextRenderNet. The outputs of each block in TextRenderNet and SceneGenNet are directly added to the corresponding block outputs of the SD3 base model.", + "bbox": [ + 511, + 494, + 906, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Character-level Visual Representation for Precise Text Rendering", + "text_level": 1, + "bbox": [ + 511, + 804, + 903, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, some works have explored multilingual visual text generation. Among them, a promising approach is based on ControlNet-like methods [38], which utilize both glyph images and line-level OCR features as conditions.", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8086", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a2487c0173392901a1993e9f7b58903ed376fa0038493606e2b51bf152014ec4.jpg", + "image_caption": [ + "Figure 5. The distinction between the multilingual character-level text representation we proposed and the line-level methods of previous works like AnyText [38] and GlyphDraw2 [26]." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 482, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, this control information cannot accurately represent characters: 1) glyph images are easily affected by text size and shape, making them less robust. 2) line-level visual features lack fine-grained stroke features and are limited by the OCR model's poor capability to recognize long texts. To address these challenges, this paper proposes a plug-and-play and robust character-level text representation, where each character is precisely represented by one token.", + "bbox": [ + 89, + 203, + 483, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the text $C$ has $n$ characters. For each character $c_{i}$ , its feature is separately extracted by a pre-trained OCR encoder $f_{v}$ and then averaged and pooled to obtain a compact character representation vector $r_{c_i} \\in \\mathbb{R}^c$ . Thus, the character-level text representation is defined as follows:", + "bbox": [ + 89, + 324, + 483, + 400 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {c i} = \\operatorname {a v g p o o l} \\left(f _ {v} \\left(I _ {c i}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 405, + 480, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {c} = \\left[ r _ {c _ {1}}, r _ {c _ {2}}, \\dots , r _ {c _ {n}} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 422, + 480, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{c_i}$ is the $i$ -th character image rendered in a fixed font, and $R_{c} \\in \\mathbb{R}^{n \\times c}$ is the char-level text representation.", + "bbox": [ + 89, + 452, + 482, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 5, compared to previous methods, our key difference is extracting representations from character glyph images. This enables the model to perceive character stroke structures and achieve high text accuracy. Additionally, since the number of characters is fixed, we can pre-extract the representations of each character and store them in a dictionary, eliminating the need for online rendering and feature extraction. This significantly simplifies the training and inference pipeline.", + "bbox": [ + 89, + 483, + 483, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, this text representation lacks order and positional information. Thus, the character order encoding $P_{rank}$ is introduced to represent the order of characters in the text, which is implemented through a sinusoidal position encoding of the char order. Besides, inspired by GLIGEN [20], the text position coordinates are mapped to sinusoidal position encoding $P_{bbox}$ to control the position of the text. Then we concatenate $P_{rank}$ , $P_{bbox}$ and $R_c$ along the feature dimension to construct the final text representation.", + "bbox": [ + 89, + 619, + 483, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Improving Subject Fidelity", + "text_level": 1, + "bbox": [ + 89, + 760, + 334, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the task of generating product posters, it is crucial to maintain subject fidelity, i.e., ensuring that the subject in the generated poster remains consistent with the user-specified subject. To achieve this goal, we employ SceneGenNet to perform background inpainting, which is trained to precisely preserve the foreground subject and only inpaint the background according to the prompt. However, inpainting-based models sometimes extend the foreground subject into", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/710505a03bc6f8844bdcd454e6243181700a8f0265f7bd6e2dee32382cd07913.jpg", + "image_caption": [ + "Figure 6. The model details of the foreground extension detector." + ], + "image_footnote": [], + "bbox": [ + 524, + 87, + 885, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ce3b15cc56ae8e48077fda94b7c57af0ec7f4b3f143c543b9d3235facd61e7f8.jpg", + "image_caption": [ + "Figure 7. The illustration of our two-stage training strategy for efficiently optimizing PosterMaker." + ], + "image_footnote": [], + "bbox": [ + 516, + 234, + 687, + 329 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/144c221f07e4d286416725b244293eebfd4d20a6634e8dc9ab9801c828d020a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 234, + 898, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "another subject (as shown in Fig. 2 (b)) thereby compromising subject fidelity. We refer to this as \"foreground extension\". To mitigate this issue, we develop a model to detect foreground extension and employ it as a reward model to fine-tune PosterMaker to improve subject fidelity.", + "bbox": [ + 511, + 372, + 905, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Foreground Extension Detector. We develop the foreground extension detector $S_{\\theta}$ based on HQ-SAM [16]. As shown in Fig. 6, we input the generated image $I_{g}$ to SAM [17] image encoder. The subject mask $M_{s}$ and box $B_{s}$ are provided as mask prompt and box prompt, respectively, to the HQ-SAM decoder to obtain an intermediate mask $M_{i}$ . Next, we concatenate the image features extracted from SAM encoder with $M_{s}$ , $M_{i}$ and $M_{s} - M_{i}$ at the channel dimension. The concatenated features are processed through convolutional layers and MLP layers to predict whether the foreground has been extended in the generated image. We collected 20k manually annotated images to train the foreground extension detector $S_{\\theta}$ .", + "bbox": [ + 511, + 449, + 905, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subject Fidelity Feedback Learning. The foreground extension detector $S_{\\theta}$ , after the offline training, is used as a reward model to supervise PosterMaker to improve subject fidelity. Specifically, assuming the reverse process has a total of $T'$ steps, we follow ReFL [43] to first sample $z_{T'} \\sim \\mathcal{N}(0,1)$ and after $T' - t'$ steps of inference $(z_{T'} \\rightarrow z_{T'-1} \\rightarrow \\dots \\rightarrow z_{t'})$ , we obtain $z_{t'}$ , where $t' \\sim [1, t_1]$ . Then, we directly perform a one-step inference $z_{t'} \\rightarrow z_0$ to accelerate the reverse process. Furthermore, $z_0$ is decoded to the generated image $x_0$ . The detector $S_{\\theta}$ predicts the foreground extension score for $x_0$ , and this score is used as the reward loss to optimize the generator $G_{\\phi}$ (i.e., PostMaker). The reward loss is defined as follows:", + "bbox": [ + 511, + 647, + 906, + 843 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e w a r d}} (\\phi) = - \\mathbb {E} _ {(x, c, m) \\sim \\mathcal {D} _ {\\text {t r a i n}}, t ^ {\\prime} \\sim [ 1, t _ {1} ], z _ {T ^ {\\prime}} \\sim \\mathcal {N} (0, 1)} \\\\ \\log \\sigma \\left(1 - S _ {\\theta} \\left(G _ {\\phi} \\left(z _ {T ^ {\\prime}}, x, c, m, t ^ {\\prime}\\right), m\\right)\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 849, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8087", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0070060a364d9b3a711de82fba80cfa82d9491fd8c946ccda917b1fe47648d5c.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison with different methods. Best viewed on Screen. To aid comprehension, Chinese text lines in the image are translated into English and annotated using corresponding colors." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 903, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $x, c, m$ sampled from the train data $\\mathcal{D}_{\\mathrm{train}}$ , represent the subject image, control conditions, and subject mask respectively. To avoid overfitting, we don't calculate reward loss for the cases where the foreground extension score is below 0.3. Our total training loss is defined as:", + "bbox": [ + 89, + 407, + 483, + 484 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {d e n o i s e}} + \\lambda \\mathcal {L} _ {\\text {r e w a r d}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 501, + 482, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is the hyperparameter to adjust the weight of reward loss and the denoise loss.", + "bbox": [ + 89, + 535, + 482, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Training Strategy", + "text_level": 1, + "bbox": [ + 89, + 590, + 264, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To efficiently train PosterMaker, this paper introduces a two-stage training strategy, as shown in Fig. 7, aimed at decoupling the learning for text rendering and background image generation. Specifically, in the first stage, the training task is local text editing. We freeze SceneGenNet and only the TextRenderNet and adapter are optimized. Since we initialize SceneGenNet with pre-trained weights of inpainting-controlnet [7], it can fill the local background well thus TextRenderNet can focus on learning text generation. In the second stage, the training task is subject-based text-to-image generation. Here we froze TextRenderNet and only train the SceneGenNet. In this stage, SceneGenNet focuses on learning poster scenes and creative design from the train data. Notably, Stage 1 learns local text editing/inpainting and Stage 2 learns background inpainting, thus the input images indicating the area to inpaint are different (See Fig. 7). With such a two-stage training strategy, TextRenderNet and SceneGenNet can be efficiently optimized since they can focus on their specific tasks.", + "bbox": [ + 89, + 617, + 483, + 905 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 405, + 645, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 426, + 702, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We crawl product posters from online e-commerce platforms to construct our training set. Our training data mainly consists of Chinese posters, we first employ PPOCRv4 model [30] to extract the text content and their bounding boxes from the images as a coarse annotation. And we ask some annotators to further refine the bounding boxes and correct the text content to improve the annotation quality. Resulting in a dataset containing 160k images. We generate image captions with GPT4-o [29] and extract foreground subject masks with $\\mathrm{U}^2$ -Net [33] and VitMatte [46]. We randomly select 302 images for evaluation and leave the rest for training. To better evaluate the performance of our method, we use LLM [9] to generate some background prompts and text layouts as evaluation samples, after manually checking and removing those irrational ones, we obtain another 198 evaluation samples to form a final evaluation set named PosterBenchmark containing 500 samples.", + "bbox": [ + 511, + 446, + 906, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We follow Anytext [38] to evaluate text rendering accuracy using two metrics: sentence accuracy (Sen. Acc) and normalized edit distance (NED). Specifically, we crop the text line from the generated image according to the provided bounding box and utilize the OCR model [28] to predict the content $s_{\\mathrm{pred}}$ of the generated text line. We denote the ground truth text content as $s_{\\mathrm{gt}}$ . A text line is considered to be correctly generated if $s_{\\mathrm{pred}} = s_{\\mathrm{gt}}$ ; this condition is used to calculate Sen. Acc. Additionally, we compute the normalized edit distance (NED) between $s_{\\mathrm{pred}}$ and $s_{\\mathrm{gt}}$ to measure their similarity. We further calculate FID [14] to measure the visual quality and CLIP-T [36] metric for evaluating text-image alignment.", + "bbox": [ + 511, + 704, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8088", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e5da996d5adf47c36398dbed7f22968c7167086e4f13cb775bcc235b7d5f337.jpg", + "image_caption": [ + "Figure 9. Qualitative comparison using various text features. It is obvious that the character-level OCR features we used (PPOCR Char) are the most effective at maintaining character accuracy." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 903, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. Our SceneGenNet is initialized from pre-trained SD3 Inpainting-Controlnet [7] and TextRenderNet is initialized from SD3 [12] weight with the same configuration as in [8]. For Subject Fidelity Feedback Learning, we follow existing work [43] to uniformly sample $t'$ between [1, 10]. Within this range, the one-step inference result of image $x_0$ from $t'$ is close to the full inference result. The weight coefficient of $\\lambda$ is set to 0.0005. The learning rate is set to 1e-4 and the batch size is set to 192. We train our framework for 26k and 29.5k steps for training stage1 and stage2, respectively. Finally, PosterMaker was trained on 32 A100 GPUs for 3 days. During the sampling process, based on the statistical information, a maximum of 7 lines of text and 16 characters per line of text are selected from each image to render onto the image, as this setting can cover most situations in the dataset.", + "bbox": [ + 88, + 407, + 485, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with Prior Works", + "text_level": 1, + "bbox": [ + 89, + 657, + 362, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline methods. We carefully designed the following baseline approaches based on existing open-sourced techniques for comparative analysis. SD3_inpaint_byt5: We encode the text content into prompt embeddings using ByT5 [44] and employ an adapter to map these embeddings to the original prompt embedding space of SD3 before feeding them into the controlnet, which enables the controlnet to render multilingual text. SD3_canny&inpaint: First render the text into a white-background image and extract the canny edge from it as control. Then finetune a pre-trained SD3 canny controlnet together with an inpainting controlnet to achieve multilingual text rendering. Anytext: It is the SOTA open-sourced T2I method that supports multilin", + "bbox": [ + 89, + 676, + 483, + 875 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c5a57855734fcf807b46be2eb5b85b9a9826b9ab1fdafffb7b589e4271f68b52.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSen. ACC ↑NED ↑FID ↓CLIP-T ↑FG Ext. Ratio ↓
SD3 inpaint_AnyText52.78%75.27%100.8726.9014.82%
SD3 inpaint_byt552.28%86.57%65.4526.7114.60%
AnyText63.90%82.81%71.2726.6919.25%
Glyph-ByT5-v269.54%87.65%79.2326.6018.91%
SD3_canny&inpaint80.75%92.75%67.1927.0314.38%
GlyphDraw286.14%96.78%72.4926.7216.52%
GT (w/ SD1.5 Rec.)76.95%89.91%---
GT (w/ SD3 Rec.)98.09%99.36%---
GT98.53%99.59%---
Ours (SD1.5)72.12%88.01%68.1726.93-
Ours93.36%98.39%65.3527.0411.57%
", + "bbox": [ + 517, + 404, + 898, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Comparison with baseline methods.", + "bbox": [ + 575, + 535, + 841, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gual text rendering and its text editing mode supports text inpainting [38]. So we directly finetune it on our data using its text editing training pipeline. SD3_inpaint_Anytext: First generate the background with SD3 inpainting control-net, then render the text on the corresponding region using Anytext. Glyph-ByT5-v2 and GlyphDraw2: They are both the SOTA T2I methods that support multilingual text rendering [24, 26]. However, they don't have open-sourced pre-trained weights, so we reproduced them on our dataset. And we added an inpainting controlnet for them to support subject-preserved generation.", + "bbox": [ + 511, + 551, + 906, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. We trained all baseline models on the same dataset, and then quantitatively compared all methods on the PosterBenchmark, as shown in Tab. 1. It is worth noting that SD3 is used as the base model by default, but since we observed that the SD1.5 VAE leads to significant error in reconstruction, to enable a more equitable comparison between our method and AnyText (SD1.5 architecture), we also implemented an SD1.5 version of PosterMaker with the same experimental setup as AnyText. As the VAEs, especially SD1.5, introduce some reconstruction error and the OCR model may incorrectly recognize some characters, we also report the metrics on ground truth", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1Details can be found in the Appendix.", + "bbox": [ + 107, + 886, + 316, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8089", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f3f05887672be46327d87fc359aea44d5457c5e334261d53af6cc072989a02d5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Text FeatureTypeSen. ACCNED
ByT5textual feat.33.48%54.50%
Cannyimg81.50%92.72%
TrOCR Linevisual feat.26.58%49.46%
TrOCR Charvisual feat.94.27%98.54%
PPOCR Linevisual feat.38.91%53.86%
PPOCR Char (Ours)visual feat.95.15%98.75%
GT (w/o Rec.)-98.53%99.59%
GT (w/ SD3 Rec.)-98.09%99.36%
", + "bbox": [ + 130, + 88, + 441, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e65d1a650bc7ff1db543d002925588ec2a67d84bcdcabf36c7fd8d7604d9bdd6.jpg", + "table_caption": [ + "Table 2. Quantitative comparison using various text features." + ], + "table_footnote": [], + "table_body": "
MethodFG Ext. Ratio↓Sen. ACC ↑NED↑FID↓CLIP-T↑
Ours11.57%93.36%98.39%65.3527.04
Ours w/o Lreward15.05%93.11%98.21%65.1027.04
", + "bbox": [ + 96, + 218, + 475, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Evaluation on the subject fidelity feedback learning.", + "bbox": [ + 104, + 253, + 467, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "images as an upper bound. As shown in Tab. 1, our method achieves the best performance on all metrics. Notably, on text rendering metrics Sen. ACC and NED, our model outperforms the baselines by an impressive margin and is already close to the upper bound. The promising results demonstrate the effectiveness of the proposed PosterMaker. Qualitative Comparison. The results are shown in Fig. 8. Compared to the baselines, our PosterMaker generates more readable and accurate poster images with texts, particularly for smaller texts. Notably, as an end-to-end generation method, PosterMaker automatically creates underlays to enhance the contrast between text and background, effectively highlighting the text. This feature is crucial in product poster design for capturing viewers' attention. These findings demonstrate that our PosterMaker successfully learns the distribution of posters created by human designers.", + "bbox": [ + 89, + 271, + 482, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study and Analysis", + "text_level": 1, + "bbox": [ + 89, + 518, + 346, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How to achieve high text rendering accuracy? We conduct experiments to explore the effectiveness of different control conditions for visual text rendering. Due to the fact that text rendering accuracy is primarily determined by the first training stage, we discard the second training stage in this experiment to save computational resources. The results are summarized in Tab. 2. We observed several valuable experimental results: 1) The use of char-level features significantly outperforms previous line-level features, benefiting from finer-grained representation. This explains why previous methods [4, 26, 38], achieve inferior performance (PPOCR Line is used in [26, 38], TrOCR Line is used in [4]). Recent concurrent works [27, 42] have also found similar experimental findings as ours. 2) Char-level feature representation is superior to low-level image features such as Canny. 3) PPOCR outperforms TrOCR, which is attributed to PPOCR being a multi-language OCR model, while TrOCR is an English version model. 4) Even though TrOCR has not been trained on multi-language text data, it still achieves decent results, likely because it extracts universal visual structural features. 5) ByT5 extracts char-level features but the performance is inferior to OCR features, because it extracts semantic features rather than character structural features, while T2I models' text rendering", + "bbox": [ + 89, + 537, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e867053f7adcdd70aee505f1f17453b337873f1f67286a27fd6d271f02ade6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 88, + 630, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b8d73b6df260a6783137dee1b957ecef10b8e5eed6ebbc01fb5548e27544670f.jpg", + "image_caption": [ + "Figure 10. Visual examples showing the effect of $\\mathcal{L}_{reward}$ ." + ], + "image_footnote": [], + "bbox": [ + 633, + 88, + 901, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "capability relies more on character structural features. We present visualization results in Fig. 9. We observe that when using line-level features as a control, the generated text occasionally becomes completely unrecognizable. This suggests that line-level features are insufficient for achieving precise text rendering. Additionally, it is evident that using canny control always introduces stroke artifacts, particularly in smaller texts (as seen in row 3 of Fig. 9). This further demonstrates that canny control is also not an ideal condition for text rendering. In summary, the char-level feature extracted by PPOCR performs optimally and the accuracy is already close to the upper bound, indicating the discriminative char-level visual feature is the key to achieve high text rendering accuracy.", + "bbox": [ + 511, + 263, + 906, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of subject fidelity feedback learning. We calculate the foreground extension ratio (termed as FG Ext. Ratio) by asking human annotators to manually check each generated image whether the foreground subject is incorrectly extended. As demonstrated in Tab. 3, training our model with $\\mathcal{L}_{reward}$ effectively reduces FG Ext. Ratio by $3.4\\%$ , while maintaining subtle variations in other performance metrics. Representative visual examples are presented in Fig. 10. Besides, our model outperforms baseline methods in FG Ext. Ratio (see Tab. 1). These results show the efficacy of our proposed subject fidelity feedback learning approach in mitigating foreground extension artifacts.", + "bbox": [ + 511, + 474, + 908, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 666, + 633, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The application of image generation in poster creation is often impeded by subpar text rendering and inconsistent subjects. To address these challenges, this paper introduces a novel framework, PosterMaker, which synthesizes aesthetically pleasing product posters with accurate and harmonious texts and contents. Moreover, we reveal that the key underlying successful multilingual text rendering is the construction of robust character-level visual text representations. Additionally, we propose subject fidelity feedback learning to mitigate inconsistencies in subjects. Through extensive experiments, our method demonstrates a significant improvement in both high-precision text generation and subject fidelity. These findings not only advance poster generation but also inspire future research on T2I models.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8090", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Nature Science Foundation of China (62425114, 62121002, U23B2028, 62232006, 62272436) and Alibaba Group (Alibaba Research Intern Program).", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 193, + 187, + 209 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3", + "[2] Tingfeng Cao, Junsheng Kong, Xue Zhao, Wenqing Yao, Junwei Ding, Jinhui Zhu, and Jiandong Zhang. Product2img: Prompt-free e-commerce product background generation with diffusion model and self-improved LMM. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pages 10774-10783. ACM, 2024. 2, 3", + "[3] Kelvin C. K. Chan, Yang Zhao, Xuhui Jia, Ming-Hsuan Yang, and Huisheng Wang. Improving subject-driven image synthesis with subject-agnostic guidance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 6733-6742. IEEE, 2024. 3", + "[4] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Jun Lan, Xing Zheng, Yaohui Li, Changhua Meng, Huijia Zhu, and Weiqiang Wang. Diffuse: Universal text editing diffusion model. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2, 3, 8", + "[5] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8724-8733, 2024. 2", + "[6] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W. Cohen. Subject-driven text-to-image generation via apprenticeship learning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3", + "[7] Alimama Creative. Sd3-controlnet-inpainting. https://huggingface.co/alimama-creative/SD3-Controlnet-Inpainting, 2024.6, 7", + "[8] Alimama Creative. Sd3-controlnet-softedge. https://huggingface.co/alimama-creative/SD3-Controlnet-Softedge, 2024.7", + "[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen," + ], + "bbox": [ + 99, + 217, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Ji-aqi Wang. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6", + "[10] Zhenbang Du, Wei Feng, Haohan Wang, Yaoyu Li, Jingsen Wang, Jian Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junsheng Jin, et al. Towards reliable advertising image generation using human feedback. In European Conference on Computer Vision, pages 399-415. Springer, 2024. 2, 3, 4", + "[11] Amir Erfan Eshratifar, Joao V.B. Soares, Kapil Thadani, Shaunak Mishra, Mikhail Kuznetsov, Yueh-Ning Ku, and Paloma De Juan. Salient object-aware background generation using text-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 7489-7499, 2024. 2, 4", + "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2, 3, 4, 7", + "[13] Yifan Gao, Jinpeng Lin, Min Zhou, Chuanbin Liu, Hongtao Xie, Tiezheng Ge, and Yuning Jiang. Textpainter: Multimodal text image generation with visual-harmony and text-comprehension for poster design. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7236-7246, 2023. 2", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 6626-6637, 2017. 6", + "[15] Peidong Jia, Chenxuan Li, Yuhui Yuan, Zeyu Liu, Yichao Shen, Bohan Chen, Xingru Chen, Yinglin Zheng, Dong Chen, Ji Li, Xiaodong Xie, Shanghang Zhang, and Baining Guo. Cole: A hierarchical generation framework for multilayered and editable graphic design, 2024. 3", + "[16] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In Advances in Neural Information Processing Systems, pages 29914–29934. Curran Associates, Inc., 2023. 5", + "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 5", + "[18] Chao Li, Chen Jiang, Xiaolong Liu, Jun Zhao, and Guoxin Wang. Joytype: A robust design for multilingual visual text creation. arXiv preprint arXiv:2409.17524, 2024. 3", + "[19] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In Advances in Neural Infor" + ], + "bbox": [ + 514, + 92, + 908, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8091", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mation Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3", + "[20] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 5", + "[21] Zhaochen Li, Fengheng Li, Wei Feng, Honghe Zhu, An Liu, Yaoyu Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junjie Shen, et al. Planning and rendering: Towards end-to-end product poster generation. arXiv preprint arXiv:2312.08822, 2023. 2", + "[22] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1250-1260, 2023. 2, 3", + "[23] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. In European Conference on Computer Vision, pages 361-377. Springer, 2024. 2, 3", + "[24] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 2, 3, 7", + "[25] Zhiying Lu, Chuanbin Liu, Xiaojun Chang, Yongdong Zhang, and Hongtao Xie. Dhvt: Dynamic hybrid vision transformer for small dataset recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3", + "[26] Jian Ma, Yonglin Deng, Chen Chen, Haonan Lu, and Zhenyu Yang. Glyphdraw2: Automatic generation of complex glyph posters with diffusion models and large language models. arXiv preprint arXiv:2407.02252, 2024. 2, 3, 5, 7, 8", + "[27] Lichen Ma, Tiezhu Yue, Pei Fu, Yujie Zhong, Kai Zhou, Xiaoming Wei, and Jie Hu. Chargen: High accurate character-level visual text generation model with multimodal encoder. arXiv preprint arXiv:2412.17225, 2024. 8", + "[28] ModelScope. https://modelscope.cn/models/damo/cv_convnextTinyOCR-recognition-general_damo/summary, 2023.6", + "[29] OpenAI. https://openai.com/index/hello-gpt-40/, 2024.6", + "[30] PaddlePaddle. https://github.com/PaddlePaddle/PaddleOCR, 2023.6", + "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In *The Twelfth International Conference on Learning Representations*, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2, 3", + "[32] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentan" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8693-8702, 2024. 3", + "[33] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. page 107404, 2020. 6", + "[34] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and PeterJ. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv: Learning, arXiv: Learning, 2019. 3", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2", + "[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3, 6", + "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3", + "[38] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2, 3, 4, 5, 6, 7, 8", + "[39] Haohan Wang, Wei Feng, Yaoyu Li, Zheng Zhang, Jingjing Lv, Junjie Shen, Zhangang Lin, and Jingping Shao. Generate e-commerce product background by integrating category commonality and personalized style. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2025. 3", + "[40] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2", + "[41] Shaodong Wang, Yunyang Ge, Liuhan Chen, Haiyang Zhou, Qian Wang, Xinhua Cheng, and Li Yuan. Prompt2poster: Automatically artistic chinese poster creation from prompt only. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pages 10716-10724. ACM, 2024. 3", + "[42] Tong Wang, Xiaochao Qu, and Ting Liu. Textmastero: Mastering high-quality scene text editing in diverse languages and styles. arXiv preprint arXiv:2408.10623, 2024. 8", + "[43] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere-" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8092", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, pages 15903-15935. Curran Associates, Inc., 2023. 5, 7", + "[44] Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. ByT5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics, 10:291-306, 2022. 7", + "[45] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2, 3", + "[46] Jingfeng Yao, Xinggang Wang, Shusheng Yang, and Baoyuan Wang. Vitmatte: Boosting image matting with pretrained plain vision transformers. Information Fusion, 103: 102091, 2024. 6", + "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2", + "[48] Boqiang Zhang, Zuan Gao, Yadong Qu, and Hongtao Xie. How control information influences multilingual text image generation and editing? arXiv preprint arXiv:2407.11502, 2024. 3, 4", + "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4", + "[50] Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, and Zhibo Yang. Visual text generation in the wild. In European Conference on Computer Vision, pages 89-106. Springer, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 599 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8093", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/b07ff10a-9a3d-44d4-80c1-c5cc5cfdc9e1_model.json b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/b07ff10a-9a3d-44d4-80c1-c5cc5cfdc9e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c9a09ce3986074ee3c7a84941673cf116866509b --- /dev/null +++ b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/b07ff10a-9a3d-44d4-80c1-c5cc5cfdc9e1_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.032 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.174, + 0.131, + 0.825, + 0.177 + ], + "angle": 0, + "content": "PosterMaker: Towards High-Quality Product Poster Generation with Accurate Text Rendering" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.203, + 0.735, + 0.221 + ], + "angle": 0, + "content": "Yifan \\(\\mathrm{Gao}^{1,2*^{\\dagger}}\\), Zihang Lin\\(^{2*}\\), Chuanbin Liu\\(^{1\\ddagger}\\), Min Zhou\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.222, + 0.669, + 0.24 + ], + "angle": 0, + "content": "Tiezheng Ge², Bo Zheng², Hongtao Xie¹" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.239, + 0.836, + 0.258 + ], + "angle": 0, + "content": "1University of Science and Technology of China 2Taubao & Tmall Group of Alibaba" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.26, + 0.725, + 0.275 + ], + "angle": 0, + "content": "eafn@mail.ustc.edu.cn {liucb92, hxtie}@ustc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.278, + 0.782, + 0.293 + ], + "angle": 0, + "content": "{linzihang.lzh, yunqi.zm, tiezheng.gtz, bozheng}@alibaba-inc.com" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.296, + 0.7, + 0.31 + ], + "angle": 0, + "content": "Project page: https://poster-maker.github.io" + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.312, + 0.22, + 0.327 + ], + "angle": 0, + "content": "Prompt" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.336, + 0.258, + 0.395 + ], + "angle": 0, + "content": "The box of fish oil supplements is placed on a wooden table, with a background of a serene ocean and clear sky, symbolizing purity and the natural source of the product" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.429, + 0.258, + 0.499 + ], + "angle": 0, + "content": "The subject rests on a smooth, dark wooden table, surrounded by a few scattered leaves and delicate flowers, with a serene garden scene complete with blooming flowers and lush greenery in the background." + }, + { + "type": "title", + "bbox": [ + 0.274, + 0.312, + 0.333, + 0.327 + ], + "angle": 0, + "content": "Subject" + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.319, + 0.333, + 0.33 + ], + "angle": 0, + "content": "Subject" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.345, + 0.333, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.313, + 0.424, + 0.325 + ], + "angle": 0, + "content": "Text" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.326, + 0.462, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.313, + 0.537, + 0.325 + ], + "angle": 0, + "content": "Poster" + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.326, + 0.576, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.312, + 0.796, + 0.326 + ], + "angle": 0, + "content": "Previous: two stage" + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.353, + 0.659, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.328, + 0.751, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.353, + 0.772, + 0.369 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.327, + 0.868, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.398, + 0.755, + 0.408 + ], + "angle": 0, + "content": "Sence Generation" + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.398, + 0.858, + 0.408 + ], + "angle": 0, + "content": "Text Rendering" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.415, + 0.45, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.516, + 0.339, + 0.528 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.414, + 0.575, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.421, + 0.78, + 0.433 + ], + "angle": 0, + "content": "Ours: end to end" + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.456, + 0.659, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.456, + 0.755, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.439, + 0.868, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.511, + 0.874, + 0.52 + ], + "angle": 0, + "content": "End2End Generation" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.528, + 0.294, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.447, + 0.349, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.516, + 0.339, + 0.528 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.531, + 0.429, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.531, + 0.562, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.531, + 0.697, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.531, + 0.877, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.736, + 0.504, + 0.745 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.742, + 0.907, + 0.798 + ], + "angle": 0, + "content": "Figure 1. (a) Definition of the advertising product poster generation task. The input includes the prompt, subject image, and the texts to be rendered with their layouts. The output is the poster image. (b) The comparison of our method with the previous method. PosterMaker generates posters end-to-end, while previous methods first generate poster backgrounds and then render texts. (c) Visualization results demonstrate that PosterMaker can generate harmonious and aesthetically pleasing posters with accurate texts and maintain subject fidelity." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.807, + 0.327, + 0.823 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.834, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Product posters, which integrate subject, scene, and text, are crucial promotional tools for attracting customers. Cre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.907, + 0.871 + ], + "angle": 0, + "content": "ating such posters using modern image generation methods is valuable, while the main challenge lies in accurately rendering text, especially for complex writing systems like Chinese, which contains over 10,000 individual characters." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.872, + 0.905, + 0.888 + ], + "angle": 0, + "content": "In this work, we identify the key to precise text rendering" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.876, + 0.364, + 0.889 + ], + "angle": 0, + "content": "* Equal contribution. ‡ Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.889, + 0.395, + 0.901 + ], + "angle": 0, + "content": "† Work done during the internship at Alibaba Group." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.876, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8083" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.334 + ], + "angle": 0, + "content": "as constructing a character-discriminative visual feature as a control signal. Based on this insight, we propose a robust character-wise representation as control and we develop TextRenderNet, which achieves a high text rendering accuracy of over \\(90\\%\\). Another challenge in poster generation is maintaining the fidelity of user-specific products. We address this by introducing SceneGenNet, an inpainting-based model, and propose subject fidelity feedback learning to further enhance fidelity. Based on TextRenderNet and SceneGenNet, we present PosterMaker, an end-to-end generation framework. To optimize PosterMaker efficiently, we implement a two-stage training strategy that decouples text rendering and background generation learning. Experimental results show that PosterMaker outperforms existing baselines by a remarkable margin, which demonstrates its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.346, + 0.223, + 0.361 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Product posters, which showcase items for sale within well-chosen background scenes and include descriptive text, play a vital role in e-commerce advertising by capturing customers' attention and boosting sales. Creating such posters necessitates photographing the product in carefully selected environments that highlight its features, as well as thoughtfully choosing text colors and fonts to ensure that the text is appealing, legible, and harmonious with the background. This process can be quite expensive. With the significant advancements in large-scale text-to-image (T2I) models [12, 31, 35], synthesizing such product posters with image generation models attracts increasing attention. In this paper, we focus on the product poster generation task. Specifically, given a prompt describing the background scene, the foreground image of the user-specified subject and some texts together with their layouts, we aim to develop a model to generate the subject into the desired scene background and accurately render the text in an end-to-end manner (as shown in Fig. 1 (a))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.856 + ], + "angle": 0, + "content": "A straightforward solution for this task is to first generate the subject into the desired scene [2, 10, 36], and then predict the text attributes (such as color and font) [13, 22] and render them on the image. However, such two-stage approach suffers from disharmony between the text and the poster background(as shown in Fig. 2 (b)). And collecting training data is also challenging since the text attributes, especially the text font, are difficult to extract from the poster. Another solution is learning to generate the poster using a per-pixel synthesis approach, which can benefit from directly learning the distribution of professionally designed posters. We focus on such one-stage solution. The main challenge is how to ensure the text rendering accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Many recent works [12, 23, 38, 45] have been proposed to improve the text rendering accuracy for large diffusion models. Great progress has been made and some" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.53 + ], + "angle": 0, + "content": "recent work can achieve high rendering accuracy for English. However, for non-Latin languages like Chinese, one of the most widely spoken languages, achieving high rendering accuracy remains challenging. This difficulty stems from the existence of over 10,000 characters, with Chinese characters characterized by complex and diverse stroke patterns, making it extremely difficult to train a model to memorize the rendering of each individual character. Recent studies [4, 26, 38] have focused on extracting visual features of text as control signals. Typically, these approaches render text lines into glyph images and extract line-level text visual features to guide generation. Nevertheless, line-level visual features often lack the discriminative power to capture character-level visual nuances. To address this limitation, GlyphByT5 [23, 24] introduced a box-level contrastive loss with sophisticated glyph augmentation strategies to enhance character-level discriminativeness, achieving promising results. In this paper, we point out that the key to high-accuracy text rendering lies in constructing character-discriminative visual features as control signals. Specifically, we render each character as a glyph image and extract visual features via a visual encoder. These features are then concatenated with positional embeddings to form a character-level representation. Then we propose TextRenderNet, an SD3 [12] controlnet-like [49] architecture that takes the character-level representation as the control signal to render visual text. Our experiments demonstrate that the proposed character-level representation is effectively capable of achieving accurate text rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.909, + 0.822 + ], + "angle": 0, + "content": "In the task of poster generation, another important thing is to generate the user-specific subject into a desired scene while keeping high subject fidelity. Recent subject-driven controllable generation methods [36, 40, 47] can synthesize such images, but they still cannot ensure that the user-specified subject is completely consistent in the generated details (e.g., the logo on the product may be inaccurately generated), which could potentially mislead customers. Therefore, we follow poster generation methods [5, 10, 21] to address this task via introducing an inpainting-based module named SceneGenNet. However, we found that even using inpainting methods, subject consistency is not always achieved as the inpainting model sometimes extends the subject shape (as shown in Fig. 2 (a)). Similar phenomenon is also observed in [10, 11]. To address this issue, we elaboratively develop a detector to detect the foreground extension cases. Then we employ the detector as a reward model to train the SceneGenNet via feedback learning for further improving subject fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Combining the proposed TextRenderNet and SceneGenNet, we develop a framework named PosterMaker that can synthesize the product poster in an end-to-end manner. To efficiently optimize PosterMaker, we introduce a two-stage training strategy to separately train TextRenderNet and Sce" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8084" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.252, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.183, + 0.232, + 0.19 + ], + "angle": 0, + "content": "User-Specified Texts" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.091, + 0.374, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.183, + 0.351, + 0.191 + ], + "angle": 0, + "content": "User-Specified Subject" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.091, + 0.497, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.183, + 0.488, + 0.191 + ], + "angle": 0, + "content": "(a) Foreground Extension" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.091, + 0.618, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.183, + 0.61, + 0.191 + ], + "angle": 0, + "content": "(b)Text-Scene Disharmony" + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.091, + 0.74, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.183, + 0.724, + 0.191 + ], + "angle": 0, + "content": "(c) Poor Text Rendering" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.091, + 0.864, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.183, + 0.822, + 0.19 + ], + "angle": 0, + "content": "Our Result" + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.193, + 0.865, + 0.207 + ], + "angle": 0, + "content": "Figure 2. The illustration of the three challenges faced by poster generation, which seriously hinder the practical application." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.484, + 0.306 + ], + "angle": 0, + "content": "neGenNet. This training strategy decouples the learning of text rendering and background image generation, thus TextRenderNet and SceneGenNet can focus on their specific tasks. Qualitative results (as shown in Fig. 1 (c)) demonstrate our training strategy is effective for training PosterMaker and it achieves promising poster generation results." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.308, + 0.427, + 0.323 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.325, + 0.483, + 0.4 + ], + "angle": 0, + "content": "- We proposed a novel framework named PosterMaker, which mainly consists of a TextRenderNet and a SceneGenNet. With a two-stage training strategy, PosterMaker can synthesis aesthetically product posters with texts accurately and harmoniously rendered on it." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.401, + 0.483, + 0.475 + ], + "angle": 0, + "content": "- We reveal the core of achieving accurate Chinese text rendering is to construct a robust character-level text representation as the control condition. These findings can inspire future research on improving the text rendering abilities of T2I models." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.476, + 0.484, + 0.522 + ], + "angle": 0, + "content": "- We improve the subject fidelity via subject fidelity feedback learning, which is shown effective in addressing the subject inconsistency issue." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.325, + 0.484, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.533, + 0.233, + 0.547 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.559, + 0.267, + 0.573 + ], + "angle": 0, + "content": "2.1. Poster Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Generating posters involves combining various elements like a subject image, a background scene image, and text to ensure the subject and text are prominently and accurately displayed while maintaining an appealing look. Automating this process is quite complex and challenging. Methods like AutoPoster [22], Prompt2Poster [41], and COLE [15] break it down into stages: creating images and layout, predicting the visual properties of text, and rendering the poster. These approaches have several steps and often struggle to precisely obtain all the necessary visual attributes like font and color gradients. With the emergence of more advanced generative models [31], methods like JoyType [18], Glyphbyt5 [23], and GlyphDraw2 [26] can directly generate the image and text together at the pixel level based on the poster prompt, text content, and layout. This more streamlined approach can leverage more readily available poster pixel data for training, but there is still room for improvement in terms of the overall poster cohesion and text accuracy. Our method is also a one-stage, direct pixel-level generation approach that simultaneously creates the image and text. However, our focus is on generating posters for a given" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.216, + 0.907, + 0.277 + ], + "angle": 0, + "content": "product subject, where the input includes the subject image, prompt, text content, and layout. In addition to considering text rendering accuracy and overall poster harmony, we also need to maintain the fidelity of the product." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.286, + 0.722, + 0.303 + ], + "angle": 0, + "content": "2.2. Visual Text Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.908, + 0.747 + ], + "angle": 0, + "content": "Recently, text-to-image (T2I) models [1, 12, 37] have made significant strides in enhancing English text rendering by introducing stronger text encoders, such as T5 [34]. However, multilingual text image generation still faces significant challenges due to the large number of non-Latin characters and complex stroke structures. Early work [45] has explored the ControlNet-based method [49], using low-level visual images such as glyph images as the control signal for text image generation. However, glyph images are easily affected by text size and shape, especially complex stroke details. Besides, some recent works [4, 25, 26, 38, 48, 50] utilize more robust visual features, such as line-level OCR features as control conditions to further improve the text accuracy. But the line-level visual features still perform poorly in representing stroke details for each character. To address this issue, GlyphByT5 [23, 24] proposes a method with box-level contrastive learning to align the text features extracted from the language model with the features extracted from the visual encoder. To effectively learn such alignment, GlyphByT5 relies on collecting massive amounts of data and developing complex data augmentation strategies for the alignment pre-training, which lacks flexibility. In contrast, in this paper, we reveal that the key to high-accuracy text rendering lies in constructing discriminative character-level visual features. Thus we propose a plug-and-play and robust character-level text representation derived from off-the-shelf OCR encoders, which can accurately represent the visual structure of the text without additional training and enable precise text rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.83, + 0.773 + ], + "angle": 0, + "content": "2.3. Subject-Preserved Scene Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To create a scene image with a product subject while ensuring subject fidelity, two main methods are commonly used. One is the subject-driven method [3, 6, 19, 32, 36], which adjusts the position, angle and lighting of the subject based on the prompt to create a harmonious image. However, it often struggles to preserve the significant features of the subject. The other utilizes inpainting-based background completion techniques [2, 10, 39]. It only generates the non-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8085" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.093, + 0.899, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.265, + 0.905, + 0.294 + ], + "angle": 0, + "content": "Figure 3. The framework of the PosterMaker, which is based on the SD3. To precisely generate multilingual texts and create aesthetically pleasing poster scenes, TextRenderNet and SenceGenNet are introduced, whose outputs are used as control conditions added to the SD3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.484, + 0.409 + ], + "angle": 0, + "content": "subject areas of an image and naturally keeps consistency in the original subject area. But it sometimes extends the foreground subject [10, 11], such as adding an extra handle to a cup, which also reduces subject fidelity. To maximize subject fidelity, our method uses background completion and a reward model to determine whether the foreground extension occurred, thereby enhancing subject fidelity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.415, + 0.182, + 0.43 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.295, + 0.456 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.484, + 0.554 + ], + "angle": 0, + "content": "This paper focuses on the creation of product posters, which typically consist of multiple elements such as text, subjects, and scenes, as illustrated in Fig. 1 (a). The central challenge of this task is to generate these elements accurately and harmoniously, offering both research and practical applications. The task is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.566, + 0.483, + 0.584 + ], + "angle": 0, + "content": "\\[\nI _ {g} = f \\left(I _ {s}, M _ {s}, T, P\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.685 + ], + "angle": 0, + "content": "where \\( I_{g} \\) denotes the generated poster image, \\( I_{s} \\) represents the subject image, and \\( M_{s} \\) is the subject mask. The variable \\( T \\) signifies the content and the position of text and \\( P \\) is the prompt describing the background scene. Subsequent sections will detail the design of PosterMaker, and our proposed solution to this task." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.219, + 0.71 + ], + "angle": 0, + "content": "3.2. Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 3, PosterMaker is developed based on Stable Diffusion 3 (SD3) [12], which contains a strong VAE for reconstructing the image details like text stroke. And we propose two modules, i.e., TextRenderNet and SceneGenNet, to address the poster generation task. TextRenderNet is specifically designed to learn visual text rendering, taking character-level visual text representations as input to achieve precise and controllable text rendering. SceneGenNet, on the other hand, accepts a masked image (indicating which content should remain unchanged) and a prompt, learning to generate the foreground subject within the desired scene described by the prompt. Both TextRenderNet" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.299, + 0.895, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.459, + 0.905, + 0.487 + ], + "angle": 0, + "content": "Figure 4. The details of TextRenderNet and SceneGenNet, showcasing their model architectures and their interactions with SD3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.907, + 0.799 + ], + "angle": 0, + "content": "and SceneGenNet are grounded in a ControlNet-like [48] architecture derived from SD3 and their architectures are detailed in Fig. 4. They share the same internal structure, comprising several cascaded MM-DiT blocks [12], with weights copied from the base model for initialization. The output of each MM-DiT block is added to the corresponding block of the base model after passing through a zero convolution layer [49]. The key distinction between the two modules lies in their input configurations. SceneGenNet takes the prompt as input to the text condition branch, and for the visual branch, the input is derived by the latent feature at timestep \\( t \\), the subject mask, and the masked latent to preserve the foreground area. In contrast, TextRenderNet receives text representations (detailed in the next section) in the text condition branch for text rendering. An adapter, consisting of a linear layer and layer normalization, adjusts the feature dimensions of these text representations before they are input to TextRenderNet. The outputs of each block in TextRenderNet and SceneGenNet are directly added to the corresponding block outputs of the SD3 base model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.805, + 0.905, + 0.838 + ], + "angle": 0, + "content": "3.3. Character-level Visual Representation for Precise Text Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Recently, some works have explored multilingual visual text generation. Among them, a promising approach is based on ControlNet-like methods [38], which utilize both glyph images and line-level OCR features as conditions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8086" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.483, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.197 + ], + "angle": 0, + "content": "Figure 5. The distinction between the multilingual character-level text representation we proposed and the line-level methods of previous works like AnyText [38] and GlyphDraw2 [26]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.204, + 0.484, + 0.325 + ], + "angle": 0, + "content": "However, this control information cannot accurately represent characters: 1) glyph images are easily affected by text size and shape, making them less robust. 2) line-level visual features lack fine-grained stroke features and are limited by the OCR model's poor capability to recognize long texts. To address these challenges, this paper proposes a plug-and-play and robust character-level text representation, where each character is precisely represented by one token." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.401 + ], + "angle": 0, + "content": "Specifically, the text \\(C\\) has \\(n\\) characters. For each character \\(c_{i}\\), its feature is separately extracted by a pre-trained OCR encoder \\(f_{v}\\) and then averaged and pooled to obtain a compact character representation vector \\(r_{c_i} \\in \\mathbb{R}^c\\). Thus, the character-level text representation is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.406, + 0.482, + 0.423 + ], + "angle": 0, + "content": "\\[\nr _ {c i} = \\operatorname {a v g p o o l} \\left(f _ {v} \\left(I _ {c i}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.424, + 0.482, + 0.442 + ], + "angle": 0, + "content": "\\[\nR _ {c} = \\left[ r _ {c _ {1}}, r _ {c _ {2}}, \\dots , r _ {c _ {n}} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.483, + 0.483 + ], + "angle": 0, + "content": "where \\(I_{c_i}\\) is the \\(i\\)-th character image rendered in a fixed font, and \\(R_{c} \\in \\mathbb{R}^{n \\times c}\\) is the char-level text representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.619 + ], + "angle": 0, + "content": "As shown in Fig. 5, compared to previous methods, our key difference is extracting representations from character glyph images. This enables the model to perceive character stroke structures and achieve high text accuracy. Additionally, since the number of characters is fixed, we can pre-extract the representations of each character and store them in a dictionary, eliminating the need for online rendering and feature extraction. This significantly simplifies the training and inference pipeline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.755 + ], + "angle": 0, + "content": "Finally, this text representation lacks order and positional information. Thus, the character order encoding \\( P_{rank} \\) is introduced to represent the order of characters in the text, which is implemented through a sinusoidal position encoding of the char order. Besides, inspired by GLIGEN [20], the text position coordinates are mapped to sinusoidal position encoding \\( P_{bbox} \\) to control the position of the text. Then we concatenate \\( P_{rank} \\), \\( P_{bbox} \\) and \\( R_c \\) along the feature dimension to construct the final text representation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.761, + 0.336, + 0.777 + ], + "angle": 0, + "content": "3.4. Improving Subject Fidelity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In the task of generating product posters, it is crucial to maintain subject fidelity, i.e., ensuring that the subject in the generated poster remains consistent with the user-specified subject. To achieve this goal, we employ SceneGenNet to perform background inpainting, which is trained to precisely preserve the foreground subject and only inpaint the background according to the prompt. However, inpainting-based models sometimes extend the foreground subject into" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.088, + 0.887, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.215, + 0.903, + 0.228 + ], + "angle": 0, + "content": "Figure 6. The model details of the foreground extension detector." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.235, + 0.688, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.236, + 0.899, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.335, + 0.905, + 0.362 + ], + "angle": 0, + "content": "Figure 7. The illustration of our two-stage training strategy for efficiently optimizing PosterMaker." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.906, + 0.449 + ], + "angle": 0, + "content": "another subject (as shown in Fig. 2 (b)) thereby compromising subject fidelity. We refer to this as \"foreground extension\". To mitigate this issue, we develop a model to detect foreground extension and employ it as a reward model to fine-tune PosterMaker to improve subject fidelity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.45, + 0.906, + 0.646 + ], + "angle": 0, + "content": "Foreground Extension Detector. We develop the foreground extension detector \\( S_{\\theta} \\) based on HQ-SAM [16]. As shown in Fig. 6, we input the generated image \\( I_{g} \\) to SAM [17] image encoder. The subject mask \\( M_{s} \\) and box \\( B_{s} \\) are provided as mask prompt and box prompt, respectively, to the HQ-SAM decoder to obtain an intermediate mask \\( M_{i} \\). Next, we concatenate the image features extracted from SAM encoder with \\( M_{s} \\), \\( M_{i} \\) and \\( M_{s} - M_{i} \\) at the channel dimension. The concatenated features are processed through convolutional layers and MLP layers to predict whether the foreground has been extended in the generated image. We collected 20k manually annotated images to train the foreground extension detector \\( S_{\\theta} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.648, + 0.907, + 0.844 + ], + "angle": 0, + "content": "Subject Fidelity Feedback Learning. The foreground extension detector \\( S_{\\theta} \\), after the offline training, is used as a reward model to supervise PosterMaker to improve subject fidelity. Specifically, assuming the reverse process has a total of \\( T' \\) steps, we follow ReFL [43] to first sample \\( z_{T'} \\sim \\mathcal{N}(0,1) \\) and after \\( T' - t' \\) steps of inference \\( (z_{T'} \\rightarrow z_{T'-1} \\rightarrow \\dots \\rightarrow z_{t'}) \\), we obtain \\( z_{t'} \\), where \\( t' \\sim [1, t_1] \\). Then, we directly perform a one-step inference \\( z_{t'} \\rightarrow z_0 \\) to accelerate the reverse process. Furthermore, \\( z_0 \\) is decoded to the generated image \\( x_0 \\). The detector \\( S_{\\theta} \\) predicts the foreground extension score for \\( x_0 \\), and this score is used as the reward loss to optimize the generator \\( G_{\\phi} \\) (i.e., PostMaker). The reward loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.85, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e w a r d}} (\\phi) = - \\mathbb {E} _ {(x, c, m) \\sim \\mathcal {D} _ {\\text {t r a i n}}, t ^ {\\prime} \\sim [ 1, t _ {1} ], z _ {T ^ {\\prime}} \\sim \\mathcal {N} (0, 1)} \\\\ \\log \\sigma \\left(1 - S _ {\\theta} \\left(G _ {\\phi} \\left(z _ {T ^ {\\prime}}, x, c, m, t ^ {\\prime}\\right), m\\right)\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8087" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.905, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.375, + 0.907, + 0.403 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison with different methods. Best viewed on Screen. To aid comprehension, Chinese text lines in the image are translated into English and annotated using corresponding colors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.485 + ], + "angle": 0, + "content": "where \\( x, c, m \\) sampled from the train data \\( \\mathcal{D}_{\\mathrm{train}} \\), represent the subject image, control conditions, and subject mask respectively. To avoid overfitting, we don't calculate reward loss for the cases where the foreground extension score is below 0.3. Our total training loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.502, + 0.483, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {d e n o i s e}} + \\lambda \\mathcal {L} _ {\\text {r e w a r d}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.483, + 0.566 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the hyperparameter to adjust the weight of reward loss and the denoise loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.265, + 0.609 + ], + "angle": 0, + "content": "3.5. Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.484, + 0.906 + ], + "angle": 0, + "content": "To efficiently train PosterMaker, this paper introduces a two-stage training strategy, as shown in Fig. 7, aimed at decoupling the learning for text rendering and background image generation. Specifically, in the first stage, the training task is local text editing. We freeze SceneGenNet and only the TextRenderNet and adapter are optimized. Since we initialize SceneGenNet with pre-trained weights of inpainting-controlnet [7], it can fill the local background well thus TextRenderNet can focus on learning text generation. In the second stage, the training task is subject-based text-to-image generation. Here we froze TextRenderNet and only train the SceneGenNet. In this stage, SceneGenNet focuses on learning poster scenes and creative design from the train data. Notably, Stage 1 learns local text editing/inpainting and Stage 2 learns background inpainting, thus the input images indicating the area to inpaint are different (See Fig. 7). With such a two-stage training strategy, TextRenderNet and SceneGenNet can be efficiently optimized since they can focus on their specific tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.406, + 0.646, + 0.423 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.428, + 0.704, + 0.444 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.704 + ], + "angle": 0, + "content": "Dataset. We crawl product posters from online e-commerce platforms to construct our training set. Our training data mainly consists of Chinese posters, we first employ PPOCRv4 model [30] to extract the text content and their bounding boxes from the images as a coarse annotation. And we ask some annotators to further refine the bounding boxes and correct the text content to improve the annotation quality. Resulting in a dataset containing 160k images. We generate image captions with GPT4-o [29] and extract foreground subject masks with \\(\\mathrm{U}^2\\)-Net [33] and VitMatte [46]. We randomly select 302 images for evaluation and leave the rest for training. To better evaluate the performance of our method, we use LLM [9] to generate some background prompts and text layouts as evaluation samples, after manually checking and removing those irrational ones, we obtain another 198 evaluation samples to form a final evaluation set named PosterBenchmark containing 500 samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. We follow Anytext [38] to evaluate text rendering accuracy using two metrics: sentence accuracy (Sen. Acc) and normalized edit distance (NED). Specifically, we crop the text line from the generated image according to the provided bounding box and utilize the OCR model [28] to predict the content \\( s_{\\mathrm{pred}} \\) of the generated text line. We denote the ground truth text content as \\( s_{\\mathrm{gt}} \\). A text line is considered to be correctly generated if \\( s_{\\mathrm{pred}} = s_{\\mathrm{gt}} \\); this condition is used to calculate Sen. Acc. Additionally, we compute the normalized edit distance (NED) between \\( s_{\\mathrm{pred}} \\) and \\( s_{\\mathrm{gt}} \\) to measure their similarity. We further calculate FID [14] to measure the visual quality and CLIP-T [36] metric for evaluating text-image alignment." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8088" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.905, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.372, + 0.907, + 0.399 + ], + "angle": 0, + "content": "Figure 9. Qualitative comparison using various text features. It is obvious that the character-level OCR features we used (PPOCR Char) are the most effective at maintaining character accuracy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.408, + 0.486, + 0.651 + ], + "angle": 0, + "content": "Implementation Details. Our SceneGenNet is initialized from pre-trained SD3 Inpainting-Controlnet [7] and TextRenderNet is initialized from SD3 [12] weight with the same configuration as in [8]. For Subject Fidelity Feedback Learning, we follow existing work [43] to uniformly sample \\( t' \\) between [1, 10]. Within this range, the one-step inference result of image \\( x_0 \\) from \\( t' \\) is close to the full inference result. The weight coefficient of \\( \\lambda \\) is set to 0.0005. The learning rate is set to 1e-4 and the batch size is set to 192. We train our framework for 26k and 29.5k steps for training stage1 and stage2, respectively. Finally, PosterMaker was trained on 32 A100 GPUs for 3 days. During the sampling process, based on the statistical information, a maximum of 7 lines of text and 16 characters per line of text are selected from each image to render onto the image, as this setting can cover most situations in the dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.363, + 0.673 + ], + "angle": 0, + "content": "4.2. Comparison with Prior Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Baseline methods. We carefully designed the following baseline approaches based on existing open-sourced techniques for comparative analysis. SD3_inpaint_byt5: We encode the text content into prompt embeddings using ByT5 [44] and employ an adapter to map these embeddings to the original prompt embedding space of SD3 before feeding them into the controlnet, which enables the controlnet to render multilingual text. SD3_canny&inpaint: First render the text into a white-background image and extract the canny edge from it as control. Then finetune a pre-trained SD3 canny controlnet together with an inpainting controlnet to achieve multilingual text rendering. Anytext: It is the SOTA open-sourced T2I method that supports multilin" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.405, + 0.9, + 0.535 + ], + "angle": 0, + "content": "
ModelSen. ACC ↑NED ↑FID ↓CLIP-T ↑FG Ext. Ratio ↓
SD3 inpaint_AnyText52.78%75.27%100.8726.9014.82%
SD3 inpaint_byt552.28%86.57%65.4526.7114.60%
AnyText63.90%82.81%71.2726.6919.25%
Glyph-ByT5-v269.54%87.65%79.2326.6018.91%
SD3_canny&inpaint80.75%92.75%67.1927.0314.38%
GlyphDraw286.14%96.78%72.4926.7216.52%
GT (w/ SD1.5 Rec.)76.95%89.91%---
GT (w/ SD3 Rec.)98.09%99.36%---
GT98.53%99.59%---
Ours (SD1.5)72.12%88.01%68.1726.93-
Ours93.36%98.39%65.3527.0411.57%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.576, + 0.536, + 0.843, + 0.549 + ], + "angle": 0, + "content": "Table 1. Comparison with baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.718 + ], + "angle": 0, + "content": "gual text rendering and its text editing mode supports text inpainting [38]. So we directly finetune it on our data using its text editing training pipeline. SD3_inpaint_Anytext: First generate the background with SD3 inpainting control-net, then render the text on the corresponding region using Anytext. Glyph-ByT5-v2 and GlyphDraw2: They are both the SOTA T2I methods that support multilingual text rendering [24, 26]. However, they don't have open-sourced pre-trained weights, so we reproduced them on our dataset. And we added an inpainting controlnet for them to support subject-preserved generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Quantitative Comparison. We trained all baseline models on the same dataset, and then quantitatively compared all methods on the PosterBenchmark, as shown in Tab. 1. It is worth noting that SD3 is used as the base model by default, but since we observed that the SD1.5 VAE leads to significant error in reconstruction, to enable a more equitable comparison between our method and AnyText (SD1.5 architecture), we also implemented an SD1.5 version of PosterMaker with the same experimental setup as AnyText. As the VAEs, especially SD1.5, introduce some reconstruction error and the OCR model may incorrectly recognize some characters, we also report the metrics on ground truth" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.318, + 0.901 + ], + "angle": 0, + "content": "1Details can be found in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8089" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.089, + 0.442, + 0.201 + ], + "angle": 0, + "content": "
Text FeatureTypeSen. ACCNED
ByT5textual feat.33.48%54.50%
Cannyimg81.50%92.72%
TrOCR Linevisual feat.26.58%49.46%
TrOCR Charvisual feat.94.27%98.54%
PPOCR Linevisual feat.38.91%53.86%
PPOCR Char (Ours)visual feat.95.15%98.75%
GT (w/o Rec.)-98.53%99.59%
GT (w/ SD3 Rec.)-98.09%99.36%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.201, + 0.468, + 0.215 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison using various text features." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.219, + 0.476, + 0.255 + ], + "angle": 0, + "content": "
MethodFG Ext. Ratio↓Sen. ACC ↑NED↑FID↓CLIP-T↑
Ours11.57%93.36%98.39%65.3527.04
Ours w/o Lreward15.05%93.11%98.21%65.1027.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.255, + 0.468, + 0.269 + ], + "angle": 0, + "content": "Table 3. Evaluation on the subject fidelity feedback learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.483, + 0.514 + ], + "angle": 0, + "content": "images as an upper bound. As shown in Tab. 1, our method achieves the best performance on all metrics. Notably, on text rendering metrics Sen. ACC and NED, our model outperforms the baselines by an impressive margin and is already close to the upper bound. The promising results demonstrate the effectiveness of the proposed PosterMaker. Qualitative Comparison. The results are shown in Fig. 8. Compared to the baselines, our PosterMaker generates more readable and accurate poster images with texts, particularly for smaller texts. Notably, as an end-to-end generation method, PosterMaker automatically creates underlays to enhance the contrast between text and background, effectively highlighting the text. This feature is crucial in product poster design for capturing viewers' attention. These findings demonstrate that our PosterMaker successfully learns the distribution of posters created by human designers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.519, + 0.347, + 0.535 + ], + "angle": 0, + "content": "4.3. Ablation Study and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.902 + ], + "angle": 0, + "content": "How to achieve high text rendering accuracy? We conduct experiments to explore the effectiveness of different control conditions for visual text rendering. Due to the fact that text rendering accuracy is primarily determined by the first training stage, we discard the second training stage in this experiment to save computational resources. The results are summarized in Tab. 2. We observed several valuable experimental results: 1) The use of char-level features significantly outperforms previous line-level features, benefiting from finer-grained representation. This explains why previous methods [4, 26, 38], achieve inferior performance (PPOCR Line is used in [26, 38], TrOCR Line is used in [4]). Recent concurrent works [27, 42] have also found similar experimental findings as ours. 2) Char-level feature representation is superior to low-level image features such as Canny. 3) PPOCR outperforms TrOCR, which is attributed to PPOCR being a multi-language OCR model, while TrOCR is an English version model. 4) Even though TrOCR has not been trained on multi-language text data, it still achieves decent results, likely because it extracts universal visual structural features. 5) ByT5 extracts char-level features but the performance is inferior to OCR features, because it extracts semantic features rather than character structural features, while T2I models' text rendering" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.089, + 0.632, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.089, + 0.903, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.238, + 0.885, + 0.253 + ], + "angle": 0, + "content": "Figure 10. Visual examples showing the effect of \\(\\mathcal{L}_{reward}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.264, + 0.907, + 0.475 + ], + "angle": 0, + "content": "capability relies more on character structural features. We present visualization results in Fig. 9. We observe that when using line-level features as a control, the generated text occasionally becomes completely unrecognizable. This suggests that line-level features are insufficient for achieving precise text rendering. Additionally, it is evident that using canny control always introduces stroke artifacts, particularly in smaller texts (as seen in row 3 of Fig. 9). This further demonstrates that canny control is also not an ideal condition for text rendering. In summary, the char-level feature extracted by PPOCR performs optimally and the accuracy is already close to the upper bound, indicating the discriminative char-level visual feature is the key to achieve high text rendering accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.909, + 0.658 + ], + "angle": 0, + "content": "Effectiveness of subject fidelity feedback learning. We calculate the foreground extension ratio (termed as FG Ext. Ratio) by asking human annotators to manually check each generated image whether the foreground subject is incorrectly extended. As demonstrated in Tab. 3, training our model with \\(\\mathcal{L}_{reward}\\) effectively reduces FG Ext. Ratio by \\(3.4\\%\\), while maintaining subtle variations in other performance metrics. Representative visual examples are presented in Fig. 10. Besides, our model outperforms baseline methods in FG Ext. Ratio (see Tab. 1). These results show the efficacy of our proposed subject fidelity feedback learning approach in mitigating foreground extension artifacts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.634, + 0.683 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The application of image generation in poster creation is often impeded by subpar text rendering and inconsistent subjects. To address these challenges, this paper introduces a novel framework, PosterMaker, which synthesizes aesthetically pleasing product posters with accurate and harmonious texts and contents. Moreover, we reveal that the key underlying successful multilingual text rendering is the construction of robust character-level visual text representations. Additionally, we propose subject fidelity feedback learning to mitigate inconsistencies in subjects. Through extensive experiments, our method demonstrates a significant improvement in both high-precision text generation and subject fidelity. These findings not only advance poster generation but also inspire future research on T2I models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8090" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.178 + ], + "angle": 0, + "content": "This work was supported by the National Nature Science Foundation of China (62425114, 62121002, U23B2028, 62232006, 62272436) and Alibaba Group (Alibaba Research Intern Program)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.194, + 0.188, + 0.21 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.218, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.29, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[2] Tingfeng Cao, Junsheng Kong, Xue Zhao, Wenqing Yao, Junwei Ding, Jinhui Zhu, and Jiandong Zhang. Product2img: Prompt-free e-commerce product background generation with diffusion model and self-improved LMM. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pages 10774-10783. ACM, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.403, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[3] Kelvin C. K. Chan, Yang Zhao, Xuhui Jia, Ming-Hsuan Yang, and Huisheng Wang. Improving subject-driven image synthesis with subject-agnostic guidance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 6733-6742. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.488, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[4] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Jun Lan, Xing Zheng, Yaohui Li, Changhua Meng, Huijia Zhu, and Weiqiang Wang. Diffuse: Universal text editing diffusion model. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[5] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8724-8733, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.484, + 0.756 + ], + "angle": 0, + "content": "[6] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W. Cohen. Subject-driven text-to-image generation via apprenticeship learning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.758, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[7] Alimama Creative. Sd3-controlnet-inpainting. https://huggingface.co/alimama-creative/SD3-Controlnet-Inpainting, 2024.6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[8] Alimama Creative. Sd3-controlnet-softedge. https://huggingface.co/alimama-creative/SD3-Controlnet-Softedge, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen," + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.218, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Ji-aqi Wang. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[10] Zhenbang Du, Wei Feng, Haohan Wang, Yaoyu Li, Jingsen Wang, Jian Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junsheng Jin, et al. Towards reliable advertising image generation using human feedback. In European Conference on Computer Vision, pages 399-415. Springer, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.909, + 0.316 + ], + "angle": 0, + "content": "[11] Amir Erfan Eshratifar, Joao V.B. Soares, Kapil Thadani, Shaunak Mishra, Mikhail Kuznetsov, Yueh-Ning Ku, and Paloma De Juan. Salient object-aware background generation using text-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 7489-7499, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.907, + 0.487 + ], + "angle": 0, + "content": "[13] Yifan Gao, Jinpeng Lin, Min Zhou, Chuanbin Liu, Hongtao Xie, Tiezheng Ge, and Yuning Jiang. Textpainter: Multimodal text image generation with visual-harmony and text-comprehension for poster design. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7236-7246, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.587 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 6626-6637, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[15] Peidong Jia, Chenxuan Li, Yuhui Yuan, Zeyu Liu, Yichao Shen, Bohan Chen, Xingru Chen, Yinglin Zheng, Dong Chen, Ji Li, Xiaodong Xie, Shanghang Zhang, and Baining Guo. Cole: A hierarchical generation framework for multilayered and editable graphic design, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.906, + 0.728 + ], + "angle": 0, + "content": "[16] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In Advances in Neural Information Processing Systems, pages 29914–29934. Curran Associates, Inc., 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.817, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[18] Chao Li, Chen Jiang, Xiaolong Liu, Jun Zhao, and Guoxin Wang. Joytype: A robust design for multilingual visual text creation. arXiv preprint arXiv:2409.17524, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[19] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In Advances in Neural Infor" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8091" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "mation Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[20] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[21] Zhaochen Li, Fengheng Li, Wei Feng, Honghe Zhu, An Liu, Yaoyu Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junjie Shen, et al. Planning and rendering: Towards end-to-end product poster generation. arXiv preprint arXiv:2312.08822, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[22] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1250-1260, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[23] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. In European Conference on Computer Vision, pages 361-377. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[24] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[25] Zhiying Lu, Chuanbin Liu, Xiaojun Chang, Yongdong Zhang, and Hongtao Xie. Dhvt: Dynamic hybrid vision transformer for small dataset recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[26] Jian Ma, Yonglin Deng, Chen Chen, Haonan Lu, and Zhenyu Yang. Glyphdraw2: Automatic generation of complex glyph posters with diffusion models and large language models. arXiv preprint arXiv:2407.02252, 2024. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[27] Lichen Ma, Tiezhu Yue, Pei Fu, Yujie Zhong, Kai Zhou, Xiaoming Wei, and Jie Hu. Chargen: High accurate character-level visual text generation model with multimodal encoder. arXiv preprint arXiv:2412.17225, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[28] ModelScope. https://modelscope.cn/models/damo/cv_convnextTinyOCR-recognition-general_damo/summary, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[29] OpenAI. https://openai.com/index/hello-gpt-40/, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[30] PaddlePaddle. https://github.com/PaddlePaddle/PaddleOCR, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In *The Twelfth International Conference on Learning Representations*, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[32] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentan" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "gled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8693-8702, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[33] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. page 107404, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[34] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and PeterJ. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv: Learning, arXiv: Learning, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.265, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[38] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[39] Haohan Wang, Wei Feng, Yaoyu Li, Zheng Zhang, Jingjing Lv, Junjie Shen, Zhangang Lin, and Jingping Shao. Generate e-commerce product background by integrating category commonality and personalized style. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[40] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.827 + ], + "angle": 0, + "content": "[41] Shaodong Wang, Yunyang Ge, Liuhan Chen, Haiyang Zhou, Qian Wang, Xinhua Cheng, and Li Yuan. Prompt2poster: Automatically artistic chinese poster creation from prompt only. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pages 10716-10724. ACM, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[42] Tong Wang, Xiaochao Qu, and Ting Liu. Textmastero: Mastering high-quality scene text editing in diverse languages and styles. arXiv preprint arXiv:2408.10623, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[43] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8092" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "ward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, pages 15903-15935. Curran Associates, Inc., 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[44] Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. ByT5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics, 10:291-306, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[45] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[46] Jingfeng Yao, Xinggang Wang, Shusheng Yang, and Baoyuan Wang. Vitmatte: Boosting image matting with pretrained plain vision transformers. Information Fusion, 103: 102091, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.431, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[48] Boqiang Zhang, Zuan Gao, Yadong Qu, and Hongtao Xie. How control information influences multilingual text image generation and editing? arXiv preprint arXiv:2407.11502, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[50] Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, and Zhibo Yang. Visual text generation in the wild. In European Conference on Computer Vision, pages 89-106. Springer, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8093" + } + ] +] \ No newline at end of file diff --git a/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/b07ff10a-9a3d-44d4-80c1-c5cc5cfdc9e1_origin.pdf b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/b07ff10a-9a3d-44d4-80c1-c5cc5cfdc9e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b4d2344ae718db99315b47ff783327d39b65673b --- /dev/null +++ b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/b07ff10a-9a3d-44d4-80c1-c5cc5cfdc9e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d19e5bd6e7b35f316070737a0b5e9cf1e1d5fae09e3e3a53170a8ea1cf71c245 +size 9433696 diff --git a/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/full.md b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..481b28ea13144f1c35dec0090e67f86f33b771c4 --- /dev/null +++ b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/full.md @@ -0,0 +1,342 @@ +# PosterMaker: Towards High-Quality Product Poster Generation with Accurate Text Rendering + +Yifan $\mathrm{Gao}^{1,2*^{\dagger}}$ , Zihang Lin $^{2*}$ , Chuanbin Liu $^{1\ddagger}$ , Min Zhou $^{2}$ + +Tiezheng Ge², Bo Zheng², Hongtao Xie¹ + +1University of Science and Technology of China 2Taubao & Tmall Group of Alibaba + +eafn@mail.ustc.edu.cn {liucb92, hxtie}@ustc.edu.cn + +{linzihang.lzh, yunqi.zm, tiezheng.gtz, bozheng}@alibaba-inc.com + +Project page: https://poster-maker.github.io + +# Prompt + +The box of fish oil supplements is placed on a wooden table, with a background of a serene ocean and clear sky, symbolizing purity and the natural source of the product + +The subject rests on a smooth, dark wooden table, surrounded by a few scattered leaves and delicate flowers, with a serene garden scene complete with blooming flowers and lush greenery in the background. + +# Subject + +![](images/c39803913d39788b94ec04fef3ac80e74421a6198196bc02a5bd02b718fd765a.jpg) + +![](images/75947218702e92302820cb9baa7de9463820ef706e74c7e044188115851dd1f4.jpg) +Subject +Text + +![](images/f724ce88fd72832683cb32c80146975374055324d5acc373f59b6e617fbf87a7.jpg) +Poster + +![](images/36d5b12e25f4ec3efd638bd460c8de49074587021651dd875e78639bc38e7983.jpg) + +![](images/42180260bd03b80fb04721970272421b10653b361ab2ab1290b9295303dae9ae.jpg) + +Sence Generation + +![](images/3cf9e98e53396d396a582892f6dc825377942a392dcafad67beabd74181c5a54.jpg) +Previous: two stage +Text Rendering + +![](images/71e69ab1def6963bfd89c44458898882e15fca0e09fcc0c0e58bdb2d3b2b242c.jpg) + +(a) + +![](images/db85dda03cb016ad13a61f099800a9ec0edbaf73e499f6d6e661672628c7ab82.jpg) + +![](images/8f7e1363ad398d6c802d00a1eb780e597a51affa9a5c1649ab970300e87fb5b3.jpg) + +![](images/a42a935afba24f668b33dcbf3435955c67ba4285a34a9f47c87e7b64a90ba578.jpg) + +![](images/ee9f4c06ad5b211cb99a39bc982efb594ad0fb928f3dac715280a44621d82256.jpg) +Ours: end to end +End2End Generation + +![](images/55d0a972d63f570075dc671b97f11b22dddc3ebfdf8f32121bef7a2edcc13f63.jpg) +Figure 1. (a) Definition of the advertising product poster generation task. The input includes the prompt, subject image, and the texts to be rendered with their layouts. The output is the poster image. (b) The comparison of our method with the previous method. PosterMaker generates posters end-to-end, while previous methods first generate poster backgrounds and then render texts. (c) Visualization results demonstrate that PosterMaker can generate harmonious and aesthetically pleasing posters with accurate texts and maintain subject fidelity. + +![](images/bbfd7bc0afe4f41a64333017094c1a0270611b5dad8340758c94680f87e94d1d.jpg) + +(a) + +![](images/f098cc62326e39aa571b819f3e0889cf597a922e2ee525f1fd2454e54daef7f5.jpg) + +![](images/678f30e21ce9dfd0a82a8949952406ab66063d64f8085b3dce4c7fc308546fa8.jpg) +(c) + +![](images/ed7095cd8c8bd5b29e57ad87f9a250a42ebc2c7348274ff0742b615a0c323821.jpg) + +![](images/84c47e8a1266912a9a1442a5a20ad3930e2be592fcf338d890e6dc43a6150735.jpg) + +# Abstract + +Product posters, which integrate subject, scene, and text, are crucial promotional tools for attracting customers. Cre + +ating such posters using modern image generation methods is valuable, while the main challenge lies in accurately rendering text, especially for complex writing systems like Chinese, which contains over 10,000 individual characters. + +In this work, we identify the key to precise text rendering + +as constructing a character-discriminative visual feature as a control signal. Based on this insight, we propose a robust character-wise representation as control and we develop TextRenderNet, which achieves a high text rendering accuracy of over $90\%$ . Another challenge in poster generation is maintaining the fidelity of user-specific products. We address this by introducing SceneGenNet, an inpainting-based model, and propose subject fidelity feedback learning to further enhance fidelity. Based on TextRenderNet and SceneGenNet, we present PosterMaker, an end-to-end generation framework. To optimize PosterMaker efficiently, we implement a two-stage training strategy that decouples text rendering and background generation learning. Experimental results show that PosterMaker outperforms existing baselines by a remarkable margin, which demonstrates its effectiveness. + +# 1. Introduction + +Product posters, which showcase items for sale within well-chosen background scenes and include descriptive text, play a vital role in e-commerce advertising by capturing customers' attention and boosting sales. Creating such posters necessitates photographing the product in carefully selected environments that highlight its features, as well as thoughtfully choosing text colors and fonts to ensure that the text is appealing, legible, and harmonious with the background. This process can be quite expensive. With the significant advancements in large-scale text-to-image (T2I) models [12, 31, 35], synthesizing such product posters with image generation models attracts increasing attention. In this paper, we focus on the product poster generation task. Specifically, given a prompt describing the background scene, the foreground image of the user-specified subject and some texts together with their layouts, we aim to develop a model to generate the subject into the desired scene background and accurately render the text in an end-to-end manner (as shown in Fig. 1 (a)). + +A straightforward solution for this task is to first generate the subject into the desired scene [2, 10, 36], and then predict the text attributes (such as color and font) [13, 22] and render them on the image. However, such two-stage approach suffers from disharmony between the text and the poster background(as shown in Fig. 2 (b)). And collecting training data is also challenging since the text attributes, especially the text font, are difficult to extract from the poster. Another solution is learning to generate the poster using a per-pixel synthesis approach, which can benefit from directly learning the distribution of professionally designed posters. We focus on such one-stage solution. The main challenge is how to ensure the text rendering accuracy. + +Many recent works [12, 23, 38, 45] have been proposed to improve the text rendering accuracy for large diffusion models. Great progress has been made and some + +recent work can achieve high rendering accuracy for English. However, for non-Latin languages like Chinese, one of the most widely spoken languages, achieving high rendering accuracy remains challenging. This difficulty stems from the existence of over 10,000 characters, with Chinese characters characterized by complex and diverse stroke patterns, making it extremely difficult to train a model to memorize the rendering of each individual character. Recent studies [4, 26, 38] have focused on extracting visual features of text as control signals. Typically, these approaches render text lines into glyph images and extract line-level text visual features to guide generation. Nevertheless, line-level visual features often lack the discriminative power to capture character-level visual nuances. To address this limitation, GlyphByT5 [23, 24] introduced a box-level contrastive loss with sophisticated glyph augmentation strategies to enhance character-level discriminativeness, achieving promising results. In this paper, we point out that the key to high-accuracy text rendering lies in constructing character-discriminative visual features as control signals. Specifically, we render each character as a glyph image and extract visual features via a visual encoder. These features are then concatenated with positional embeddings to form a character-level representation. Then we propose TextRenderNet, an SD3 [12] controlnet-like [49] architecture that takes the character-level representation as the control signal to render visual text. Our experiments demonstrate that the proposed character-level representation is effectively capable of achieving accurate text rendering. + +In the task of poster generation, another important thing is to generate the user-specific subject into a desired scene while keeping high subject fidelity. Recent subject-driven controllable generation methods [36, 40, 47] can synthesize such images, but they still cannot ensure that the user-specified subject is completely consistent in the generated details (e.g., the logo on the product may be inaccurately generated), which could potentially mislead customers. Therefore, we follow poster generation methods [5, 10, 21] to address this task via introducing an inpainting-based module named SceneGenNet. However, we found that even using inpainting methods, subject consistency is not always achieved as the inpainting model sometimes extends the subject shape (as shown in Fig. 2 (a)). Similar phenomenon is also observed in [10, 11]. To address this issue, we elaboratively develop a detector to detect the foreground extension cases. Then we employ the detector as a reward model to train the SceneGenNet via feedback learning for further improving subject fidelity. + +Combining the proposed TextRenderNet and SceneGenNet, we develop a framework named PosterMaker that can synthesize the product poster in an end-to-end manner. To efficiently optimize PosterMaker, we introduce a two-stage training strategy to separately train TextRenderNet and Sce + +![](images/3bca4c2fe91c83af5f8bc57bb9f87d5a99b10eb79ad5aef18ca44b9109f43476.jpg) +User-Specified Texts +Figure 2. The illustration of the three challenges faced by poster generation, which seriously hinder the practical application. + +![](images/2c2bbb6e96fa79a8c0ed43fdbb36306144d76c90eaddc9dc2374e6cab60a705f.jpg) +User-Specified Subject + +![](images/eed9665624db8c382f7b9579f934d5325a2c8fea2f84f3571e373f1646036d4b.jpg) +(a) Foreground Extension + +![](images/bc694b1aafee0ae3795f6d58a09278daa381b8dc389e97f872d6d9b944c81747.jpg) +(b)Text-Scene Disharmony + +![](images/32f43f9a3f3bbb936aa874f05bf6711eb69c6e2951d13756d160e81588271b22.jpg) +(c) Poor Text Rendering + +![](images/d0cadd493438ccd8585a7d1de28306f44365d9f6578bb23d96ef2a7fdc69f606.jpg) +Our Result + +neGenNet. This training strategy decouples the learning of text rendering and background image generation, thus TextRenderNet and SceneGenNet can focus on their specific tasks. Qualitative results (as shown in Fig. 1 (c)) demonstrate our training strategy is effective for training PosterMaker and it achieves promising poster generation results. + +To summarize, our contributions are as follows: + +- We proposed a novel framework named PosterMaker, which mainly consists of a TextRenderNet and a SceneGenNet. With a two-stage training strategy, PosterMaker can synthesis aesthetically product posters with texts accurately and harmoniously rendered on it. +- We reveal the core of achieving accurate Chinese text rendering is to construct a robust character-level text representation as the control condition. These findings can inspire future research on improving the text rendering abilities of T2I models. +- We improve the subject fidelity via subject fidelity feedback learning, which is shown effective in addressing the subject inconsistency issue. + +# 2. Related Work + +# 2.1. Poster Generation + +Generating posters involves combining various elements like a subject image, a background scene image, and text to ensure the subject and text are prominently and accurately displayed while maintaining an appealing look. Automating this process is quite complex and challenging. Methods like AutoPoster [22], Prompt2Poster [41], and COLE [15] break it down into stages: creating images and layout, predicting the visual properties of text, and rendering the poster. These approaches have several steps and often struggle to precisely obtain all the necessary visual attributes like font and color gradients. With the emergence of more advanced generative models [31], methods like JoyType [18], Glyphbyt5 [23], and GlyphDraw2 [26] can directly generate the image and text together at the pixel level based on the poster prompt, text content, and layout. This more streamlined approach can leverage more readily available poster pixel data for training, but there is still room for improvement in terms of the overall poster cohesion and text accuracy. Our method is also a one-stage, direct pixel-level generation approach that simultaneously creates the image and text. However, our focus is on generating posters for a given + +product subject, where the input includes the subject image, prompt, text content, and layout. In addition to considering text rendering accuracy and overall poster harmony, we also need to maintain the fidelity of the product. + +# 2.2. Visual Text Rendering + +Recently, text-to-image (T2I) models [1, 12, 37] have made significant strides in enhancing English text rendering by introducing stronger text encoders, such as T5 [34]. However, multilingual text image generation still faces significant challenges due to the large number of non-Latin characters and complex stroke structures. Early work [45] has explored the ControlNet-based method [49], using low-level visual images such as glyph images as the control signal for text image generation. However, glyph images are easily affected by text size and shape, especially complex stroke details. Besides, some recent works [4, 25, 26, 38, 48, 50] utilize more robust visual features, such as line-level OCR features as control conditions to further improve the text accuracy. But the line-level visual features still perform poorly in representing stroke details for each character. To address this issue, GlyphByT5 [23, 24] proposes a method with box-level contrastive learning to align the text features extracted from the language model with the features extracted from the visual encoder. To effectively learn such alignment, GlyphByT5 relies on collecting massive amounts of data and developing complex data augmentation strategies for the alignment pre-training, which lacks flexibility. In contrast, in this paper, we reveal that the key to high-accuracy text rendering lies in constructing discriminative character-level visual features. Thus we propose a plug-and-play and robust character-level text representation derived from off-the-shelf OCR encoders, which can accurately represent the visual structure of the text without additional training and enable precise text rendering. + +# 2.3. Subject-Preserved Scene Generation + +To create a scene image with a product subject while ensuring subject fidelity, two main methods are commonly used. One is the subject-driven method [3, 6, 19, 32, 36], which adjusts the position, angle and lighting of the subject based on the prompt to create a harmonious image. However, it often struggles to preserve the significant features of the subject. The other utilizes inpainting-based background completion techniques [2, 10, 39]. It only generates the non- + +![](images/f4b117659f00639c8b4bd3674bba491facc16adbac19530cd80757917def6165.jpg) +Figure 3. The framework of the PosterMaker, which is based on the SD3. To precisely generate multilingual texts and create aesthetically pleasing poster scenes, TextRenderNet and SenceGenNet are introduced, whose outputs are used as control conditions added to the SD3. + +subject areas of an image and naturally keeps consistency in the original subject area. But it sometimes extends the foreground subject [10, 11], such as adding an extra handle to a cup, which also reduces subject fidelity. To maximize subject fidelity, our method uses background completion and a reward model to determine whether the foreground extension occurred, thereby enhancing subject fidelity. + +# 3. Method + +# 3.1. Problem Formulation + +This paper focuses on the creation of product posters, which typically consist of multiple elements such as text, subjects, and scenes, as illustrated in Fig. 1 (a). The central challenge of this task is to generate these elements accurately and harmoniously, offering both research and practical applications. The task is defined as: + +$$ +I _ {g} = f \left(I _ {s}, M _ {s}, T, P\right), \tag {1} +$$ + +where $I_{g}$ denotes the generated poster image, $I_{s}$ represents the subject image, and $M_{s}$ is the subject mask. The variable $T$ signifies the content and the position of text and $P$ is the prompt describing the background scene. Subsequent sections will detail the design of PosterMaker, and our proposed solution to this task. + +# 3.2. Framework + +As shown in Fig. 3, PosterMaker is developed based on Stable Diffusion 3 (SD3) [12], which contains a strong VAE for reconstructing the image details like text stroke. And we propose two modules, i.e., TextRenderNet and SceneGenNet, to address the poster generation task. TextRenderNet is specifically designed to learn visual text rendering, taking character-level visual text representations as input to achieve precise and controllable text rendering. SceneGenNet, on the other hand, accepts a masked image (indicating which content should remain unchanged) and a prompt, learning to generate the foreground subject within the desired scene described by the prompt. Both TextRenderNet + +![](images/042150ef1df01697695703c3c6f7189b7317020c8da26186af947fd013b306a5.jpg) +Figure 4. The details of TextRenderNet and SceneGenNet, showcasing their model architectures and their interactions with SD3. + +and SceneGenNet are grounded in a ControlNet-like [48] architecture derived from SD3 and their architectures are detailed in Fig. 4. They share the same internal structure, comprising several cascaded MM-DiT blocks [12], with weights copied from the base model for initialization. The output of each MM-DiT block is added to the corresponding block of the base model after passing through a zero convolution layer [49]. The key distinction between the two modules lies in their input configurations. SceneGenNet takes the prompt as input to the text condition branch, and for the visual branch, the input is derived by the latent feature at timestep $t$ , the subject mask, and the masked latent to preserve the foreground area. In contrast, TextRenderNet receives text representations (detailed in the next section) in the text condition branch for text rendering. An adapter, consisting of a linear layer and layer normalization, adjusts the feature dimensions of these text representations before they are input to TextRenderNet. The outputs of each block in TextRenderNet and SceneGenNet are directly added to the corresponding block outputs of the SD3 base model. + +# 3.3. Character-level Visual Representation for Precise Text Rendering + +Recently, some works have explored multilingual visual text generation. Among them, a promising approach is based on ControlNet-like methods [38], which utilize both glyph images and line-level OCR features as conditions. + +![](images/a2487c0173392901a1993e9f7b58903ed376fa0038493606e2b51bf152014ec4.jpg) +Figure 5. The distinction between the multilingual character-level text representation we proposed and the line-level methods of previous works like AnyText [38] and GlyphDraw2 [26]. + +However, this control information cannot accurately represent characters: 1) glyph images are easily affected by text size and shape, making them less robust. 2) line-level visual features lack fine-grained stroke features and are limited by the OCR model's poor capability to recognize long texts. To address these challenges, this paper proposes a plug-and-play and robust character-level text representation, where each character is precisely represented by one token. + +Specifically, the text $C$ has $n$ characters. For each character $c_{i}$ , its feature is separately extracted by a pre-trained OCR encoder $f_{v}$ and then averaged and pooled to obtain a compact character representation vector $r_{c_i} \in \mathbb{R}^c$ . Thus, the character-level text representation is defined as follows: + +$$ +r _ {c i} = \operatorname {a v g p o o l} \left(f _ {v} \left(I _ {c i}\right)\right), \tag {2} +$$ + +$$ +R _ {c} = \left[ r _ {c _ {1}}, r _ {c _ {2}}, \dots , r _ {c _ {n}} \right], \tag {3} +$$ + +where $I_{c_i}$ is the $i$ -th character image rendered in a fixed font, and $R_{c} \in \mathbb{R}^{n \times c}$ is the char-level text representation. + +As shown in Fig. 5, compared to previous methods, our key difference is extracting representations from character glyph images. This enables the model to perceive character stroke structures and achieve high text accuracy. Additionally, since the number of characters is fixed, we can pre-extract the representations of each character and store them in a dictionary, eliminating the need for online rendering and feature extraction. This significantly simplifies the training and inference pipeline. + +Finally, this text representation lacks order and positional information. Thus, the character order encoding $P_{rank}$ is introduced to represent the order of characters in the text, which is implemented through a sinusoidal position encoding of the char order. Besides, inspired by GLIGEN [20], the text position coordinates are mapped to sinusoidal position encoding $P_{bbox}$ to control the position of the text. Then we concatenate $P_{rank}$ , $P_{bbox}$ and $R_c$ along the feature dimension to construct the final text representation. + +# 3.4. Improving Subject Fidelity + +In the task of generating product posters, it is crucial to maintain subject fidelity, i.e., ensuring that the subject in the generated poster remains consistent with the user-specified subject. To achieve this goal, we employ SceneGenNet to perform background inpainting, which is trained to precisely preserve the foreground subject and only inpaint the background according to the prompt. However, inpainting-based models sometimes extend the foreground subject into + +![](images/710505a03bc6f8844bdcd454e6243181700a8f0265f7bd6e2dee32382cd07913.jpg) +Figure 6. The model details of the foreground extension detector. + +![](images/ce3b15cc56ae8e48077fda94b7c57af0ec7f4b3f143c543b9d3235facd61e7f8.jpg) +Figure 7. The illustration of our two-stage training strategy for efficiently optimizing PosterMaker. + +![](images/144c221f07e4d286416725b244293eebfd4d20a6634e8dc9ab9801c828d020a6.jpg) + +another subject (as shown in Fig. 2 (b)) thereby compromising subject fidelity. We refer to this as "foreground extension". To mitigate this issue, we develop a model to detect foreground extension and employ it as a reward model to fine-tune PosterMaker to improve subject fidelity. + +Foreground Extension Detector. We develop the foreground extension detector $S_{\theta}$ based on HQ-SAM [16]. As shown in Fig. 6, we input the generated image $I_{g}$ to SAM [17] image encoder. The subject mask $M_{s}$ and box $B_{s}$ are provided as mask prompt and box prompt, respectively, to the HQ-SAM decoder to obtain an intermediate mask $M_{i}$ . Next, we concatenate the image features extracted from SAM encoder with $M_{s}$ , $M_{i}$ and $M_{s} - M_{i}$ at the channel dimension. The concatenated features are processed through convolutional layers and MLP layers to predict whether the foreground has been extended in the generated image. We collected 20k manually annotated images to train the foreground extension detector $S_{\theta}$ . + +Subject Fidelity Feedback Learning. The foreground extension detector $S_{\theta}$ , after the offline training, is used as a reward model to supervise PosterMaker to improve subject fidelity. Specifically, assuming the reverse process has a total of $T'$ steps, we follow ReFL [43] to first sample $z_{T'} \sim \mathcal{N}(0,1)$ and after $T' - t'$ steps of inference $(z_{T'} \rightarrow z_{T'-1} \rightarrow \dots \rightarrow z_{t'})$ , we obtain $z_{t'}$ , where $t' \sim [1, t_1]$ . Then, we directly perform a one-step inference $z_{t'} \rightarrow z_0$ to accelerate the reverse process. Furthermore, $z_0$ is decoded to the generated image $x_0$ . The detector $S_{\theta}$ predicts the foreground extension score for $x_0$ , and this score is used as the reward loss to optimize the generator $G_{\phi}$ (i.e., PostMaker). The reward loss is defined as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {r e w a r d}} (\phi) = - \mathbb {E} _ {(x, c, m) \sim \mathcal {D} _ {\text {t r a i n}}, t ^ {\prime} \sim [ 1, t _ {1} ], z _ {T ^ {\prime}} \sim \mathcal {N} (0, 1)} \\ \log \sigma \left(1 - S _ {\theta} \left(G _ {\phi} \left(z _ {T ^ {\prime}}, x, c, m, t ^ {\prime}\right), m\right)\right), \tag {4} \\ \end{array} +$$ + +![](images/0070060a364d9b3a711de82fba80cfa82d9491fd8c946ccda917b1fe47648d5c.jpg) +Figure 8. Qualitative comparison with different methods. Best viewed on Screen. To aid comprehension, Chinese text lines in the image are translated into English and annotated using corresponding colors. + +where $x, c, m$ sampled from the train data $\mathcal{D}_{\mathrm{train}}$ , represent the subject image, control conditions, and subject mask respectively. To avoid overfitting, we don't calculate reward loss for the cases where the foreground extension score is below 0.3. Our total training loss is defined as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {d e n o i s e}} + \lambda \mathcal {L} _ {\text {r e w a r d}}, \tag {5} +$$ + +where $\lambda$ is the hyperparameter to adjust the weight of reward loss and the denoise loss. + +# 3.5. Training Strategy + +To efficiently train PosterMaker, this paper introduces a two-stage training strategy, as shown in Fig. 7, aimed at decoupling the learning for text rendering and background image generation. Specifically, in the first stage, the training task is local text editing. We freeze SceneGenNet and only the TextRenderNet and adapter are optimized. Since we initialize SceneGenNet with pre-trained weights of inpainting-controlnet [7], it can fill the local background well thus TextRenderNet can focus on learning text generation. In the second stage, the training task is subject-based text-to-image generation. Here we froze TextRenderNet and only train the SceneGenNet. In this stage, SceneGenNet focuses on learning poster scenes and creative design from the train data. Notably, Stage 1 learns local text editing/inpainting and Stage 2 learns background inpainting, thus the input images indicating the area to inpaint are different (See Fig. 7). With such a two-stage training strategy, TextRenderNet and SceneGenNet can be efficiently optimized since they can focus on their specific tasks. + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset. We crawl product posters from online e-commerce platforms to construct our training set. Our training data mainly consists of Chinese posters, we first employ PPOCRv4 model [30] to extract the text content and their bounding boxes from the images as a coarse annotation. And we ask some annotators to further refine the bounding boxes and correct the text content to improve the annotation quality. Resulting in a dataset containing 160k images. We generate image captions with GPT4-o [29] and extract foreground subject masks with $\mathrm{U}^2$ -Net [33] and VitMatte [46]. We randomly select 302 images for evaluation and leave the rest for training. To better evaluate the performance of our method, we use LLM [9] to generate some background prompts and text layouts as evaluation samples, after manually checking and removing those irrational ones, we obtain another 198 evaluation samples to form a final evaluation set named PosterBenchmark containing 500 samples. + +Evaluation Metrics. We follow Anytext [38] to evaluate text rendering accuracy using two metrics: sentence accuracy (Sen. Acc) and normalized edit distance (NED). Specifically, we crop the text line from the generated image according to the provided bounding box and utilize the OCR model [28] to predict the content $s_{\mathrm{pred}}$ of the generated text line. We denote the ground truth text content as $s_{\mathrm{gt}}$ . A text line is considered to be correctly generated if $s_{\mathrm{pred}} = s_{\mathrm{gt}}$ ; this condition is used to calculate Sen. Acc. Additionally, we compute the normalized edit distance (NED) between $s_{\mathrm{pred}}$ and $s_{\mathrm{gt}}$ to measure their similarity. We further calculate FID [14] to measure the visual quality and CLIP-T [36] metric for evaluating text-image alignment. + +![](images/1e5da996d5adf47c36398dbed7f22968c7167086e4f13cb775bcc235b7d5f337.jpg) +Figure 9. Qualitative comparison using various text features. It is obvious that the character-level OCR features we used (PPOCR Char) are the most effective at maintaining character accuracy. + +Implementation Details. Our SceneGenNet is initialized from pre-trained SD3 Inpainting-Controlnet [7] and TextRenderNet is initialized from SD3 [12] weight with the same configuration as in [8]. For Subject Fidelity Feedback Learning, we follow existing work [43] to uniformly sample $t'$ between [1, 10]. Within this range, the one-step inference result of image $x_0$ from $t'$ is close to the full inference result. The weight coefficient of $\lambda$ is set to 0.0005. The learning rate is set to 1e-4 and the batch size is set to 192. We train our framework for 26k and 29.5k steps for training stage1 and stage2, respectively. Finally, PosterMaker was trained on 32 A100 GPUs for 3 days. During the sampling process, based on the statistical information, a maximum of 7 lines of text and 16 characters per line of text are selected from each image to render onto the image, as this setting can cover most situations in the dataset. + +# 4.2. Comparison with Prior Works + +Baseline methods. We carefully designed the following baseline approaches based on existing open-sourced techniques for comparative analysis. SD3_inpaint_byt5: We encode the text content into prompt embeddings using ByT5 [44] and employ an adapter to map these embeddings to the original prompt embedding space of SD3 before feeding them into the controlnet, which enables the controlnet to render multilingual text. SD3_canny&inpaint: First render the text into a white-background image and extract the canny edge from it as control. Then finetune a pre-trained SD3 canny controlnet together with an inpainting controlnet to achieve multilingual text rendering. Anytext: It is the SOTA open-sourced T2I method that supports multilin + +
ModelSen. ACC ↑NED ↑FID ↓CLIP-T ↑FG Ext. Ratio ↓
SD3 inpaint_AnyText52.78%75.27%100.8726.9014.82%
SD3 inpaint_byt552.28%86.57%65.4526.7114.60%
AnyText63.90%82.81%71.2726.6919.25%
Glyph-ByT5-v269.54%87.65%79.2326.6018.91%
SD3_canny&inpaint80.75%92.75%67.1927.0314.38%
GlyphDraw286.14%96.78%72.4926.7216.52%
GT (w/ SD1.5 Rec.)76.95%89.91%---
GT (w/ SD3 Rec.)98.09%99.36%---
GT98.53%99.59%---
Ours (SD1.5)72.12%88.01%68.1726.93-
Ours93.36%98.39%65.3527.0411.57%
+ +Table 1. Comparison with baseline methods. + +gual text rendering and its text editing mode supports text inpainting [38]. So we directly finetune it on our data using its text editing training pipeline. SD3_inpaint_Anytext: First generate the background with SD3 inpainting control-net, then render the text on the corresponding region using Anytext. Glyph-ByT5-v2 and GlyphDraw2: They are both the SOTA T2I methods that support multilingual text rendering [24, 26]. However, they don't have open-sourced pre-trained weights, so we reproduced them on our dataset. And we added an inpainting controlnet for them to support subject-preserved generation. + +Quantitative Comparison. We trained all baseline models on the same dataset, and then quantitatively compared all methods on the PosterBenchmark, as shown in Tab. 1. It is worth noting that SD3 is used as the base model by default, but since we observed that the SD1.5 VAE leads to significant error in reconstruction, to enable a more equitable comparison between our method and AnyText (SD1.5 architecture), we also implemented an SD1.5 version of PosterMaker with the same experimental setup as AnyText. As the VAEs, especially SD1.5, introduce some reconstruction error and the OCR model may incorrectly recognize some characters, we also report the metrics on ground truth + +
Text FeatureTypeSen. ACCNED
ByT5textual feat.33.48%54.50%
Cannyimg81.50%92.72%
TrOCR Linevisual feat.26.58%49.46%
TrOCR Charvisual feat.94.27%98.54%
PPOCR Linevisual feat.38.91%53.86%
PPOCR Char (Ours)visual feat.95.15%98.75%
GT (w/o Rec.)-98.53%99.59%
GT (w/ SD3 Rec.)-98.09%99.36%
+ +Table 2. Quantitative comparison using various text features. + +
MethodFG Ext. Ratio↓Sen. ACC ↑NED↑FID↓CLIP-T↑
Ours11.57%93.36%98.39%65.3527.04
Ours w/o Lreward15.05%93.11%98.21%65.1027.04
+ +Table 3. Evaluation on the subject fidelity feedback learning. + +images as an upper bound. As shown in Tab. 1, our method achieves the best performance on all metrics. Notably, on text rendering metrics Sen. ACC and NED, our model outperforms the baselines by an impressive margin and is already close to the upper bound. The promising results demonstrate the effectiveness of the proposed PosterMaker. Qualitative Comparison. The results are shown in Fig. 8. Compared to the baselines, our PosterMaker generates more readable and accurate poster images with texts, particularly for smaller texts. Notably, as an end-to-end generation method, PosterMaker automatically creates underlays to enhance the contrast between text and background, effectively highlighting the text. This feature is crucial in product poster design for capturing viewers' attention. These findings demonstrate that our PosterMaker successfully learns the distribution of posters created by human designers. + +# 4.3. Ablation Study and Analysis + +How to achieve high text rendering accuracy? We conduct experiments to explore the effectiveness of different control conditions for visual text rendering. Due to the fact that text rendering accuracy is primarily determined by the first training stage, we discard the second training stage in this experiment to save computational resources. The results are summarized in Tab. 2. We observed several valuable experimental results: 1) The use of char-level features significantly outperforms previous line-level features, benefiting from finer-grained representation. This explains why previous methods [4, 26, 38], achieve inferior performance (PPOCR Line is used in [26, 38], TrOCR Line is used in [4]). Recent concurrent works [27, 42] have also found similar experimental findings as ours. 2) Char-level feature representation is superior to low-level image features such as Canny. 3) PPOCR outperforms TrOCR, which is attributed to PPOCR being a multi-language OCR model, while TrOCR is an English version model. 4) Even though TrOCR has not been trained on multi-language text data, it still achieves decent results, likely because it extracts universal visual structural features. 5) ByT5 extracts char-level features but the performance is inferior to OCR features, because it extracts semantic features rather than character structural features, while T2I models' text rendering + +![](images/2e867053f7adcdd70aee505f1f17453b337873f1f67286a27fd6d271f02ade6e.jpg) + +![](images/b8d73b6df260a6783137dee1b957ecef10b8e5eed6ebbc01fb5548e27544670f.jpg) +Figure 10. Visual examples showing the effect of $\mathcal{L}_{reward}$ . + +capability relies more on character structural features. We present visualization results in Fig. 9. We observe that when using line-level features as a control, the generated text occasionally becomes completely unrecognizable. This suggests that line-level features are insufficient for achieving precise text rendering. Additionally, it is evident that using canny control always introduces stroke artifacts, particularly in smaller texts (as seen in row 3 of Fig. 9). This further demonstrates that canny control is also not an ideal condition for text rendering. In summary, the char-level feature extracted by PPOCR performs optimally and the accuracy is already close to the upper bound, indicating the discriminative char-level visual feature is the key to achieve high text rendering accuracy. + +Effectiveness of subject fidelity feedback learning. We calculate the foreground extension ratio (termed as FG Ext. Ratio) by asking human annotators to manually check each generated image whether the foreground subject is incorrectly extended. As demonstrated in Tab. 3, training our model with $\mathcal{L}_{reward}$ effectively reduces FG Ext. Ratio by $3.4\%$ , while maintaining subtle variations in other performance metrics. Representative visual examples are presented in Fig. 10. Besides, our model outperforms baseline methods in FG Ext. Ratio (see Tab. 1). These results show the efficacy of our proposed subject fidelity feedback learning approach in mitigating foreground extension artifacts. + +# 5. Conclusion + +The application of image generation in poster creation is often impeded by subpar text rendering and inconsistent subjects. To address these challenges, this paper introduces a novel framework, PosterMaker, which synthesizes aesthetically pleasing product posters with accurate and harmonious texts and contents. Moreover, we reveal that the key underlying successful multilingual text rendering is the construction of robust character-level visual text representations. Additionally, we propose subject fidelity feedback learning to mitigate inconsistencies in subjects. Through extensive experiments, our method demonstrates a significant improvement in both high-precision text generation and subject fidelity. These findings not only advance poster generation but also inspire future research on T2I models. + +# Acknowledgments + +This work was supported by the National Nature Science Foundation of China (62425114, 62121002, U23B2028, 62232006, 62272436) and Alibaba Group (Alibaba Research Intern Program). + +# References + +[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3 +[2] Tingfeng Cao, Junsheng Kong, Xue Zhao, Wenqing Yao, Junwei Ding, Jinhui Zhu, and Jiandong Zhang. Product2img: Prompt-free e-commerce product background generation with diffusion model and self-improved LMM. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pages 10774-10783. ACM, 2024. 2, 3 +[3] Kelvin C. K. Chan, Yang Zhao, Xuhui Jia, Ming-Hsuan Yang, and Huisheng Wang. Improving subject-driven image synthesis with subject-agnostic guidance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 6733-6742. IEEE, 2024. 3 +[4] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Jun Lan, Xing Zheng, Yaohui Li, Changhua Meng, Huijia Zhu, and Weiqiang Wang. Diffuse: Universal text editing diffusion model. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2, 3, 8 +[5] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8724-8733, 2024. 2 +[6] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W. Cohen. Subject-driven text-to-image generation via apprenticeship learning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3 +[7] Alimama Creative. Sd3-controlnet-inpainting. https://huggingface.co/alimama-creative/SD3-Controlnet-Inpainting, 2024.6, 7 +[8] Alimama Creative. Sd3-controlnet-softedge. https://huggingface.co/alimama-creative/SD3-Controlnet-Softedge, 2024.7 +[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, + +Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Ji-aqi Wang. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6 +[10] Zhenbang Du, Wei Feng, Haohan Wang, Yaoyu Li, Jingsen Wang, Jian Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junsheng Jin, et al. Towards reliable advertising image generation using human feedback. In European Conference on Computer Vision, pages 399-415. Springer, 2024. 2, 3, 4 +[11] Amir Erfan Eshratifar, Joao V.B. Soares, Kapil Thadani, Shaunak Mishra, Mikhail Kuznetsov, Yueh-Ning Ku, and Paloma De Juan. Salient object-aware background generation using text-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 7489-7499, 2024. 2, 4 +[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2, 3, 4, 7 +[13] Yifan Gao, Jinpeng Lin, Min Zhou, Chuanbin Liu, Hongtao Xie, Tiezheng Ge, and Yuning Jiang. Textpainter: Multimodal text image generation with visual-harmony and text-comprehension for poster design. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7236-7246, 2023. 2 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 6626-6637, 2017. 6 +[15] Peidong Jia, Chenxuan Li, Yuhui Yuan, Zeyu Liu, Yichao Shen, Bohan Chen, Xingru Chen, Yinglin Zheng, Dong Chen, Ji Li, Xiaodong Xie, Shanghang Zhang, and Baining Guo. Cole: A hierarchical generation framework for multilayered and editable graphic design, 2024. 3 +[16] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In Advances in Neural Information Processing Systems, pages 29914–29934. Curran Associates, Inc., 2023. 5 +[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 5 +[18] Chao Li, Chen Jiang, Xiaolong Liu, Jun Zhao, and Guoxin Wang. Joytype: A robust design for multilingual visual text creation. arXiv preprint arXiv:2409.17524, 2024. 3 +[19] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In Advances in Neural Infor + +mation Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3 +[20] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 5 +[21] Zhaochen Li, Fengheng Li, Wei Feng, Honghe Zhu, An Liu, Yaoyu Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junjie Shen, et al. Planning and rendering: Towards end-to-end product poster generation. arXiv preprint arXiv:2312.08822, 2023. 2 +[22] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1250-1260, 2023. 2, 3 +[23] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. In European Conference on Computer Vision, pages 361-377. Springer, 2024. 2, 3 +[24] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 2, 3, 7 +[25] Zhiying Lu, Chuanbin Liu, Xiaojun Chang, Yongdong Zhang, and Hongtao Xie. Dhvt: Dynamic hybrid vision transformer for small dataset recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3 +[26] Jian Ma, Yonglin Deng, Chen Chen, Haonan Lu, and Zhenyu Yang. Glyphdraw2: Automatic generation of complex glyph posters with diffusion models and large language models. arXiv preprint arXiv:2407.02252, 2024. 2, 3, 5, 7, 8 +[27] Lichen Ma, Tiezhu Yue, Pei Fu, Yujie Zhong, Kai Zhou, Xiaoming Wei, and Jie Hu. Chargen: High accurate character-level visual text generation model with multimodal encoder. arXiv preprint arXiv:2412.17225, 2024. 8 +[28] ModelScope. https://modelscope.cn/models/damo/cv_convnextTinyOCR-recognition-general_damo/summary, 2023.6 +[29] OpenAI. https://openai.com/index/hello-gpt-40/, 2024.6 +[30] PaddlePaddle. https://github.com/PaddlePaddle/PaddleOCR, 2023.6 +[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In *The Twelfth International Conference on Learning Representations*, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2, 3 +[32] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentan + +gled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8693-8702, 2024. 3 +[33] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. page 107404, 2020. 6 +[34] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and PeterJ. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv: Learning, arXiv: Learning, 2019. 3 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2 +[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3, 6 +[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3 +[38] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2, 3, 4, 5, 6, 7, 8 +[39] Haohan Wang, Wei Feng, Yaoyu Li, Zheng Zhang, Jingjing Lv, Junjie Shen, Zhangang Lin, and Jingping Shao. Generate e-commerce product background by integrating category commonality and personalized style. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2025. 3 +[40] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2 +[41] Shaodong Wang, Yunyang Ge, Liuhan Chen, Haiyang Zhou, Qian Wang, Xinhua Cheng, and Li Yuan. Prompt2poster: Automatically artistic chinese poster creation from prompt only. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pages 10716-10724. ACM, 2024. 3 +[42] Tong Wang, Xiaochao Qu, and Ting Liu. Textmastero: Mastering high-quality scene text editing in diverse languages and styles. arXiv preprint arXiv:2408.10623, 2024. 8 +[43] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- + +ward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, pages 15903-15935. Curran Associates, Inc., 2023. 5, 7 +[44] Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. ByT5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics, 10:291-306, 2022. 7 +[45] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2, 3 +[46] Jingfeng Yao, Xinggang Wang, Shusheng Yang, and Baoyuan Wang. Vitmatte: Boosting image matting with pretrained plain vision transformers. Information Fusion, 103: 102091, 2024. 6 +[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2 +[48] Boqiang Zhang, Zuan Gao, Yadong Qu, and Hongtao Xie. How control information influences multilingual text image generation and editing? arXiv preprint arXiv:2407.11502, 2024. 3, 4 +[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4 +[50] Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, and Zhibo Yang. Visual text generation in the wild. In European Conference on Computer Vision, pages 89-106. Springer, 2024. 3 \ No newline at end of file diff --git a/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/images.zip b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7b44a0285622625b1a8dedfde40df5654db80c6d --- /dev/null +++ b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b26d9ae372d795f1ac0982138e7eb1fecb65af95e00067086f11675dc06bbf60 +size 991409 diff --git a/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/layout.json b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..640a4d866840809ba2daf758e947fe341a5a738c --- /dev/null +++ b/2025/PosterMaker_ Towards High-Quality Product Poster Generation with Accurate Text Rendering/layout.json @@ -0,0 +1,9240 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 106, + 103, + 504, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 103, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 504, + 140 + ], + "type": "text", + "content": "PosterMaker: Towards High-Quality Product Poster Generation with Accurate Text Rendering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 160, + 160, + 449, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 160, + 449, + 175 + ], + "spans": [ + { + "bbox": [ + 160, + 160, + 449, + 175 + ], + "type": "text", + "content": "Yifan " + }, + { + "bbox": [ + 160, + 160, + 449, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Gao}^{1,2*^{\\dagger}}" + }, + { + "bbox": [ + 160, + 160, + 449, + 175 + ], + "type": "text", + "content": ", Zihang Lin" + }, + { + "bbox": [ + 160, + 160, + 449, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 160, + 160, + 449, + 175 + ], + "type": "text", + "content": ", Chuanbin Liu" + }, + { + "bbox": [ + 160, + 160, + 449, + 175 + ], + "type": "inline_equation", + "content": "^{1\\ddagger}" + }, + { + "bbox": [ + 160, + 160, + 449, + 175 + ], + "type": "text", + "content": ", Min Zhou" + }, + { + "bbox": [ + 160, + 160, + 449, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 201, + 175, + 409, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 175, + 409, + 190 + ], + "spans": [ + { + "bbox": [ + 201, + 175, + 409, + 190 + ], + "type": "text", + "content": "Tiezheng Ge², Bo Zheng², Hongtao Xie¹" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 189, + 511, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 189, + 511, + 204 + ], + "spans": [ + { + "bbox": [ + 99, + 189, + 511, + 204 + ], + "type": "text", + "content": "1University of Science and Technology of China 2Taubao & Tmall Group of Alibaba" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 167, + 205, + 443, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 205, + 443, + 217 + ], + "spans": [ + { + "bbox": [ + 167, + 205, + 443, + 217 + ], + "type": "text", + "content": "eafn@mail.ustc.edu.cn {liucb92, hxtie}@ustc.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 132, + 220, + 478, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 220, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 478, + 232 + ], + "type": "text", + "content": "{linzihang.lzh, yunqi.zm, tiezheng.gtz, bozheng}@alibaba-inc.com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 182, + 234, + 428, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 234, + 428, + 245 + ], + "spans": [ + { + "bbox": [ + 182, + 234, + 428, + 245 + ], + "type": "text", + "content": "Project page: https://poster-maker.github.io" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 100, + 247, + 134, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 247, + 134, + 258 + ], + "spans": [ + { + "bbox": [ + 100, + 247, + 134, + 258 + ], + "type": "text", + "content": "Prompt" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 71, + 266, + 157, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 266, + 157, + 312 + ], + "spans": [ + { + "bbox": [ + 71, + 266, + 157, + 312 + ], + "type": "text", + "content": "The box of fish oil supplements is placed on a wooden table, with a background of a serene ocean and clear sky, symbolizing purity and the natural source of the product" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 71, + 339, + 157, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 339, + 157, + 395 + ], + "spans": [ + { + "bbox": [ + 71, + 339, + 157, + 395 + ], + "type": "text", + "content": "The subject rests on a smooth, dark wooden table, surrounded by a few scattered leaves and delicate flowers, with a serene garden scene complete with blooming flowers and lush greenery in the background." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 167, + 247, + 203, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 247, + 203, + 258 + ], + "spans": [ + { + "bbox": [ + 167, + 247, + 203, + 258 + ], + "type": "text", + "content": "Subject" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 160, + 273, + 203, + 315 + ], + "blocks": [ + { + "bbox": [ + 160, + 273, + 203, + 315 + ], + "lines": [ + { + "bbox": [ + 160, + 273, + 203, + 315 + ], + "spans": [ + { + "bbox": [ + 160, + 273, + 203, + 315 + ], + "type": "image", + "image_path": "c39803913d39788b94ec04fef3ac80e74421a6198196bc02a5bd02b718fd765a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 214, + 258, + 282, + 326 + ], + "blocks": [ + { + "bbox": [ + 160, + 252, + 203, + 261 + ], + "lines": [ + { + "bbox": [ + 160, + 252, + 203, + 261 + ], + "spans": [ + { + "bbox": [ + 160, + 252, + 203, + 261 + ], + "type": "text", + "content": "Subject" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 236, + 247, + 259, + 257 + ], + "lines": [ + { + "bbox": [ + 236, + 247, + 259, + 257 + ], + "spans": [ + { + "bbox": [ + 236, + 247, + 259, + 257 + ], + "type": "text", + "content": "Text" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 214, + 258, + 282, + 326 + ], + "lines": [ + { + "bbox": [ + 214, + 258, + 282, + 326 + ], + "spans": [ + { + "bbox": [ + 214, + 258, + 282, + 326 + ], + "type": "image", + "image_path": "75947218702e92302820cb9baa7de9463820ef706e74c7e044188115851dd1f4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 283, + 258, + 352, + 326 + ], + "blocks": [ + { + "bbox": [ + 299, + 247, + 328, + 257 + ], + "lines": [ + { + "bbox": [ + 299, + 247, + 328, + 257 + ], + "spans": [ + { + "bbox": [ + 299, + 247, + 328, + 257 + ], + "type": "text", + "content": "Poster" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 283, + 258, + 352, + 326 + ], + "lines": [ + { + "bbox": [ + 283, + 258, + 352, + 326 + ], + "spans": [ + { + "bbox": [ + 283, + 258, + 352, + 326 + ], + "type": "image", + "image_path": "f724ce88fd72832683cb32c80146975374055324d5acc373f59b6e617fbf87a7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 356, + 279, + 403, + 312 + ], + "blocks": [ + { + "bbox": [ + 356, + 279, + 403, + 312 + ], + "lines": [ + { + "bbox": [ + 356, + 279, + 403, + 312 + ], + "spans": [ + { + "bbox": [ + 356, + 279, + 403, + 312 + ], + "type": "image", + "image_path": "36d5b12e25f4ec3efd638bd460c8de49074587021651dd875e78639bc38e7983.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 405, + 259, + 459, + 313 + ], + "blocks": [ + { + "bbox": [ + 405, + 259, + 459, + 313 + ], + "lines": [ + { + "bbox": [ + 405, + 259, + 459, + 313 + ], + "spans": [ + { + "bbox": [ + 405, + 259, + 459, + 313 + ], + "type": "image", + "image_path": "42180260bd03b80fb04721970272421b10653b361ab2ab1290b9295303dae9ae.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 279, + 472, + 292 + ], + "lines": [ + { + "bbox": [ + 460, + 279, + 472, + 292 + ], + "spans": [ + { + "bbox": [ + 460, + 279, + 472, + 292 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 405, + 315, + 462, + 323 + ], + "lines": [ + { + "bbox": [ + 405, + 315, + 462, + 323 + ], + "spans": [ + { + "bbox": [ + 405, + 315, + 462, + 323 + ], + "type": "text", + "content": "Sence Generation" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 474, + 258, + 531, + 313 + ], + "blocks": [ + { + "bbox": [ + 398, + 247, + 487, + 258 + ], + "lines": [ + { + "bbox": [ + 398, + 247, + 487, + 258 + ], + "spans": [ + { + "bbox": [ + 398, + 247, + 487, + 258 + ], + "type": "text", + "content": "Previous: two stage" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 258, + 531, + 313 + ], + "lines": [ + { + "bbox": [ + 474, + 258, + 531, + 313 + ], + "spans": [ + { + "bbox": [ + 474, + 258, + 531, + 313 + ], + "type": "image", + "image_path": "3cf9e98e53396d396a582892f6dc825377942a392dcafad67beabd74181c5a54.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 315, + 525, + 323 + ], + "lines": [ + { + "bbox": [ + 476, + 315, + 525, + 323 + ], + "spans": [ + { + "bbox": [ + 476, + 315, + 525, + 323 + ], + "type": "text", + "content": "Text Rendering" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 214, + 328, + 275, + 417 + ], + "blocks": [ + { + "bbox": [ + 214, + 328, + 275, + 417 + ], + "lines": [ + { + "bbox": [ + 214, + 328, + 275, + 417 + ], + "spans": [ + { + "bbox": [ + 214, + 328, + 275, + 417 + ], + "type": "image", + "image_path": "71e69ab1def6963bfd89c44458898882e15fca0e09fcc0c0e58bdb2d3b2b242c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 198, + 408, + 207, + 418 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 408, + 207, + 418 + ], + "spans": [ + { + "bbox": [ + 198, + 408, + 207, + 418 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 29, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 283, + 327, + 351, + 417 + ], + "blocks": [ + { + "bbox": [ + 283, + 327, + 351, + 417 + ], + "lines": [ + { + "bbox": [ + 283, + 327, + 351, + 417 + ], + "spans": [ + { + "bbox": [ + 283, + 327, + 351, + 417 + ], + "type": "image", + "image_path": "db85dda03cb016ad13a61f099800a9ec0edbaf73e499f6d6e661672628c7ab82.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 356, + 361, + 403, + 411 + ], + "blocks": [ + { + "bbox": [ + 356, + 361, + 403, + 411 + ], + "lines": [ + { + "bbox": [ + 356, + 361, + 403, + 411 + ], + "spans": [ + { + "bbox": [ + 356, + 361, + 403, + 411 + ], + "type": "image", + "image_path": "8f7e1363ad398d6c802d00a1eb780e597a51affa9a5c1649ab970300e87fb5b3.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 405, + 361, + 462, + 394 + ], + "blocks": [ + { + "bbox": [ + 405, + 361, + 462, + 394 + ], + "lines": [ + { + "bbox": [ + 405, + 361, + 462, + 394 + ], + "spans": [ + { + "bbox": [ + 405, + 361, + 462, + 394 + ], + "type": "image", + "image_path": "a42a935afba24f668b33dcbf3435955c67ba4285a34a9f47c87e7b64a90ba578.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 459, + 347, + 531, + 403 + ], + "blocks": [ + { + "bbox": [ + 403, + 333, + 477, + 342 + ], + "lines": [ + { + "bbox": [ + 403, + 333, + 477, + 342 + ], + "spans": [ + { + "bbox": [ + 403, + 333, + 477, + 342 + ], + "type": "text", + "content": "Ours: end to end" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 347, + 531, + 403 + ], + "lines": [ + { + "bbox": [ + 459, + 347, + 531, + 403 + ], + "spans": [ + { + "bbox": [ + 459, + 347, + 531, + 403 + ], + "type": "image", + "image_path": "ee9f4c06ad5b211cb99a39bc982efb594ad0fb928f3dac715280a44621d82256.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 404, + 534, + 411 + ], + "lines": [ + { + "bbox": [ + 469, + 404, + 534, + 411 + ], + "spans": [ + { + "bbox": [ + 469, + 404, + 534, + 411 + ], + "type": "text", + "content": "End2End Generation" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 70, + 418, + 179, + 582 + ], + "blocks": [ + { + "bbox": [ + 70, + 418, + 179, + 582 + ], + "lines": [ + { + "bbox": [ + 70, + 418, + 179, + 582 + ], + "spans": [ + { + "bbox": [ + 70, + 418, + 179, + 582 + ], + "type": "image", + "image_path": "55d0a972d63f570075dc671b97f11b22dddc3ebfdf8f32121bef7a2edcc13f63.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 587, + 555, + 632 + ], + "lines": [ + { + "bbox": [ + 55, + 587, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 555, + 632 + ], + "type": "text", + "content": "Figure 1. (a) Definition of the advertising product poster generation task. The input includes the prompt, subject image, and the texts to be rendered with their layouts. The output is the poster image. (b) The comparison of our method with the previous method. PosterMaker generates posters end-to-end, while previous methods first generate poster backgrounds and then render texts. (c) Visualization results demonstrate that PosterMaker can generate harmonious and aesthetically pleasing posters with accurate texts and maintain subject fidelity." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 179, + 354, + 213, + 388 + ], + "blocks": [ + { + "bbox": [ + 179, + 354, + 213, + 388 + ], + "lines": [ + { + "bbox": [ + 179, + 354, + 213, + 388 + ], + "spans": [ + { + "bbox": [ + 179, + 354, + 213, + 388 + ], + "type": "image", + "image_path": "bbfd7bc0afe4f41a64333017094c1a0270611b5dad8340758c94680f87e94d1d.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 198, + 408, + 207, + 418 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 408, + 207, + 418 + ], + "spans": [ + { + "bbox": [ + 198, + 408, + 207, + 418 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 38, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 181, + 420, + 262, + 583 + ], + "blocks": [ + { + "bbox": [ + 181, + 420, + 262, + 583 + ], + "lines": [ + { + "bbox": [ + 181, + 420, + 262, + 583 + ], + "spans": [ + { + "bbox": [ + 181, + 420, + 262, + 583 + ], + "type": "image", + "image_path": "f098cc62326e39aa571b819f3e0889cf597a922e2ee525f1fd2454e54daef7f5.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 263, + 420, + 343, + 582 + ], + "blocks": [ + { + "bbox": [ + 263, + 420, + 343, + 582 + ], + "lines": [ + { + "bbox": [ + 263, + 420, + 343, + 582 + ], + "spans": [ + { + "bbox": [ + 263, + 420, + 343, + 582 + ], + "type": "image", + "image_path": "678f30e21ce9dfd0a82a8949952406ab66063d64f8085b3dce4c7fc308546fa8.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 582, + 308, + 590 + ], + "lines": [ + { + "bbox": [ + 299, + 582, + 308, + 590 + ], + "spans": [ + { + "bbox": [ + 299, + 582, + 308, + 590 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 343, + 420, + 426, + 583 + ], + "blocks": [ + { + "bbox": [ + 343, + 420, + 426, + 583 + ], + "lines": [ + { + "bbox": [ + 343, + 420, + 426, + 583 + ], + "spans": [ + { + "bbox": [ + 343, + 420, + 426, + 583 + ], + "type": "image", + "image_path": "ed7095cd8c8bd5b29e57ad87f9a250a42ebc2c7348274ff0742b615a0c323821.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 426, + 420, + 536, + 582 + ], + "blocks": [ + { + "bbox": [ + 426, + 420, + 536, + 582 + ], + "lines": [ + { + "bbox": [ + 426, + 420, + 536, + 582 + ], + "spans": [ + { + "bbox": [ + 426, + 420, + 536, + 582 + ], + "type": "image", + "image_path": "84c47e8a1266912a9a1442a5a20ad3930e2be592fcf338d890e6dc43a6150735.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 152, + 639, + 200, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 639, + 200, + 651 + ], + "spans": [ + { + "bbox": [ + 152, + 639, + 200, + 651 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 55, + 660, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 295, + 684 + ], + "type": "text", + "content": "Product posters, which integrate subject, scene, and text, are crucial promotional tools for attracting customers. Cre" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 643, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 555, + 689 + ], + "type": "text", + "content": "ating such posters using modern image generation methods is valuable, while the main challenge lies in accurately rendering text, especially for complex writing systems like Chinese, which contains over 10,000 individual characters." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 315, + 690, + 553, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 690, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 315, + 690, + 553, + 703 + ], + "type": "text", + "content": "In this work, we identify the key to precise text rendering" + } + ] + } + ], + "index": 48 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 693, + 222, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 222, + 704 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 222, + 704 + ], + "type": "text", + "content": "* Equal contribution. ‡ Corresponding author." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 69, + 704, + 241, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 704, + 241, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 704, + 241, + 713 + ], + "type": "text", + "content": "† Work done during the internship at Alibaba Group." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8083" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "content": "as constructing a character-discriminative visual feature as a control signal. Based on this insight, we propose a robust character-wise representation as control and we develop TextRenderNet, which achieves a high text rendering accuracy of over " + }, + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "content": ". Another challenge in poster generation is maintaining the fidelity of user-specific products. We address this by introducing SceneGenNet, an inpainting-based model, and propose subject fidelity feedback learning to further enhance fidelity. Based on TextRenderNet and SceneGenNet, we present PosterMaker, an end-to-end generation framework. To optimize PosterMaker efficiently, we implement a two-stage training strategy that decouples text rendering and background generation learning. Experimental results show that PosterMaker outperforms existing baselines by a remarkable margin, which demonstrates its effectiveness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 274, + 136, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 136, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 136, + 285 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 293, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 521 + ], + "type": "text", + "content": "Product posters, which showcase items for sale within well-chosen background scenes and include descriptive text, play a vital role in e-commerce advertising by capturing customers' attention and boosting sales. Creating such posters necessitates photographing the product in carefully selected environments that highlight its features, as well as thoughtfully choosing text colors and fonts to ensure that the text is appealing, legible, and harmonious with the background. This process can be quite expensive. With the significant advancements in large-scale text-to-image (T2I) models [12, 31, 35], synthesizing such product posters with image generation models attracts increasing attention. In this paper, we focus on the product poster generation task. Specifically, given a prompt describing the background scene, the foreground image of the user-specified subject and some texts together with their layouts, we aim to develop a model to generate the subject into the desired scene background and accurately render the text in an end-to-end manner (as shown in Fig. 1 (a))." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 677 + ], + "type": "text", + "content": "A straightforward solution for this task is to first generate the subject into the desired scene [2, 10, 36], and then predict the text attributes (such as color and font) [13, 22] and render them on the image. However, such two-stage approach suffers from disharmony between the text and the poster background(as shown in Fig. 2 (b)). And collecting training data is also challenging since the text attributes, especially the text font, are difficult to extract from the poster. Another solution is learning to generate the poster using a per-pixel synthesis approach, which can benefit from directly learning the distribution of professionally designed posters. We focus on such one-stage solution. The main challenge is how to ensure the text rendering accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Many recent works [12, 23, 38, 45] have been proposed to improve the text rendering accuracy for large diffusion models. Great progress has been made and some" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 419 + ], + "type": "text", + "content": "recent work can achieve high rendering accuracy for English. However, for non-Latin languages like Chinese, one of the most widely spoken languages, achieving high rendering accuracy remains challenging. This difficulty stems from the existence of over 10,000 characters, with Chinese characters characterized by complex and diverse stroke patterns, making it extremely difficult to train a model to memorize the rendering of each individual character. Recent studies [4, 26, 38] have focused on extracting visual features of text as control signals. Typically, these approaches render text lines into glyph images and extract line-level text visual features to guide generation. Nevertheless, line-level visual features often lack the discriminative power to capture character-level visual nuances. To address this limitation, GlyphByT5 [23, 24] introduced a box-level contrastive loss with sophisticated glyph augmentation strategies to enhance character-level discriminativeness, achieving promising results. In this paper, we point out that the key to high-accuracy text rendering lies in constructing character-discriminative visual features as control signals. Specifically, we render each character as a glyph image and extract visual features via a visual encoder. These features are then concatenated with positional embeddings to form a character-level representation. Then we propose TextRenderNet, an SD3 [12] controlnet-like [49] architecture that takes the character-level representation as the control signal to render visual text. Our experiments demonstrate that the proposed character-level representation is effectively capable of achieving accurate text rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 422, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 556, + 651 + ], + "type": "text", + "content": "In the task of poster generation, another important thing is to generate the user-specific subject into a desired scene while keeping high subject fidelity. Recent subject-driven controllable generation methods [36, 40, 47] can synthesize such images, but they still cannot ensure that the user-specified subject is completely consistent in the generated details (e.g., the logo on the product may be inaccurately generated), which could potentially mislead customers. Therefore, we follow poster generation methods [5, 10, 21] to address this task via introducing an inpainting-based module named SceneGenNet. However, we found that even using inpainting methods, subject consistency is not always achieved as the inpainting model sometimes extends the subject shape (as shown in Fig. 2 (a)). Similar phenomenon is also observed in [10, 11]. To address this issue, we elaboratively develop a detector to detect the foreground extension cases. Then we employ the detector as a reward model to train the SceneGenNet via feedback learning for further improving subject fidelity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Combining the proposed TextRenderNet and SceneGenNet, we develop a framework named PosterMaker that can synthesize the product poster in an end-to-end manner. To efficiently optimize PosterMaker, we introduce a two-stage training strategy to separately train TextRenderNet and Sce" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8084" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 154, + 144 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 154, + 144 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 154, + 144 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 154, + 144 + ], + "type": "image", + "image_path": "3bca4c2fe91c83af5f8bc57bb9f87d5a99b10eb79ad5aef18ca44b9109f43476.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 144, + 141, + 150 + ], + "lines": [ + { + "bbox": [ + 92, + 144, + 141, + 150 + ], + "spans": [ + { + "bbox": [ + 92, + 144, + 141, + 150 + ], + "type": "text", + "content": "User-Specified Texts" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 79, + 152, + 529, + 163 + ], + "lines": [ + { + "bbox": [ + 79, + 152, + 529, + 163 + ], + "spans": [ + { + "bbox": [ + 79, + 152, + 529, + 163 + ], + "type": "text", + "content": "Figure 2. The illustration of the three challenges faced by poster generation, which seriously hinder the practical application." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 72, + 228, + 144 + ], + "blocks": [ + { + "bbox": [ + 156, + 72, + 228, + 144 + ], + "lines": [ + { + "bbox": [ + 156, + 72, + 228, + 144 + ], + "spans": [ + { + "bbox": [ + 156, + 72, + 228, + 144 + ], + "type": "image", + "image_path": "2c2bbb6e96fa79a8c0ed43fdbb36306144d76c90eaddc9dc2374e6cab60a705f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 144, + 214, + 151 + ], + "lines": [ + { + "bbox": [ + 162, + 144, + 214, + 151 + ], + "spans": [ + { + "bbox": [ + 162, + 144, + 214, + 151 + ], + "type": "text", + "content": "User-Specified Subject" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 230, + 72, + 304, + 144 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 304, + 144 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 304, + 144 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 304, + 144 + ], + "type": "image", + "image_path": "eed9665624db8c382f7b9579f934d5325a2c8fea2f84f3571e373f1646036d4b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 144, + 298, + 151 + ], + "lines": [ + { + "bbox": [ + 238, + 144, + 298, + 151 + ], + "spans": [ + { + "bbox": [ + 238, + 144, + 298, + 151 + ], + "type": "text", + "content": "(a) Foreground Extension" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 378, + 144 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 378, + 144 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 378, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 378, + 144 + ], + "type": "image", + "image_path": "bc694b1aafee0ae3795f6d58a09278daa381b8dc389e97f872d6d9b944c81747.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 144, + 373, + 151 + ], + "lines": [ + { + "bbox": [ + 310, + 144, + 373, + 151 + ], + "spans": [ + { + "bbox": [ + 310, + 144, + 373, + 151 + ], + "type": "text", + "content": "(b)Text-Scene Disharmony" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 379, + 72, + 452, + 144 + ], + "blocks": [ + { + "bbox": [ + 379, + 72, + 452, + 144 + ], + "lines": [ + { + "bbox": [ + 379, + 72, + 452, + 144 + ], + "spans": [ + { + "bbox": [ + 379, + 72, + 452, + 144 + ], + "type": "image", + "image_path": "32f43f9a3f3bbb936aa874f05bf6711eb69c6e2951d13756d160e81588271b22.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 144, + 443, + 151 + ], + "lines": [ + { + "bbox": [ + 387, + 144, + 443, + 151 + ], + "spans": [ + { + "bbox": [ + 387, + 144, + 443, + 151 + ], + "type": "text", + "content": "(c) Poor Text Rendering" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 454, + 72, + 528, + 144 + ], + "blocks": [ + { + "bbox": [ + 454, + 72, + 528, + 144 + ], + "lines": [ + { + "bbox": [ + 454, + 72, + 528, + 144 + ], + "spans": [ + { + "bbox": [ + 454, + 72, + 528, + 144 + ], + "type": "image", + "image_path": "d0cadd493438ccd8585a7d1de28306f44365d9f6578bb23d96ef2a7fdc69f606.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 144, + 503, + 150 + ], + "lines": [ + { + "bbox": [ + 477, + 144, + 503, + 150 + ], + "spans": [ + { + "bbox": [ + 477, + 144, + 503, + 150 + ], + "type": "text", + "content": "Our Result" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 171, + 296, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 242 + ], + "type": "text", + "content": "neGenNet. This training strategy decouples the learning of text rendering and background image generation, thus TextRenderNet and SceneGenNet can focus on their specific tasks. Qualitative results (as shown in Fig. 1 (c)) demonstrate our training strategy is effective for training PosterMaker and it achieves promising poster generation results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 243, + 261, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 243, + 261, + 255 + ], + "spans": [ + { + "bbox": [ + 67, + 243, + 261, + 255 + ], + "type": "text", + "content": "To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 257, + 296, + 413 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 55, + 257, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 295, + 316 + ], + "type": "text", + "content": "- We proposed a novel framework named PosterMaker, which mainly consists of a TextRenderNet and a SceneGenNet. With a two-stage training strategy, PosterMaker can synthesis aesthetically product posters with texts accurately and harmoniously rendered on it." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 317, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 295, + 376 + ], + "type": "text", + "content": "- We reveal the core of achieving accurate Chinese text rendering is to construct a robust character-level text representation as the control condition. These findings can inspire future research on improving the text rendering abilities of T2I models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 376, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 413 + ], + "type": "text", + "content": "- We improve the subject fidelity via subject fidelity feedback learning, which is shown effective in addressing the subject inconsistency issue." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 422, + 142, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 142, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 142, + 433 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 442, + 163, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 163, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 163, + 453 + ], + "type": "text", + "content": "2.1. Poster Generation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 462, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 714 + ], + "type": "text", + "content": "Generating posters involves combining various elements like a subject image, a background scene image, and text to ensure the subject and text are prominently and accurately displayed while maintaining an appealing look. Automating this process is quite complex and challenging. Methods like AutoPoster [22], Prompt2Poster [41], and COLE [15] break it down into stages: creating images and layout, predicting the visual properties of text, and rendering the poster. These approaches have several steps and often struggle to precisely obtain all the necessary visual attributes like font and color gradients. With the emergence of more advanced generative models [31], methods like JoyType [18], Glyphbyt5 [23], and GlyphDraw2 [26] can directly generate the image and text together at the pixel level based on the poster prompt, text content, and layout. This more streamlined approach can leverage more readily available poster pixel data for training, but there is still room for improvement in terms of the overall poster cohesion and text accuracy. Our method is also a one-stage, direct pixel-level generation approach that simultaneously creates the image and text. However, our focus is on generating posters for a given" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 171, + 555, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 219 + ], + "type": "text", + "content": "product subject, where the input includes the subject image, prompt, text content, and layout. In addition to considering text rendering accuracy and overall poster harmony, we also need to maintain the fidelity of the product." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 226, + 441, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 441, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 441, + 239 + ], + "type": "text", + "content": "2.2. Visual Text Rendering" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 244, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 555, + 591 + ], + "type": "text", + "content": "Recently, text-to-image (T2I) models [1, 12, 37] have made significant strides in enhancing English text rendering by introducing stronger text encoders, such as T5 [34]. However, multilingual text image generation still faces significant challenges due to the large number of non-Latin characters and complex stroke structures. Early work [45] has explored the ControlNet-based method [49], using low-level visual images such as glyph images as the control signal for text image generation. However, glyph images are easily affected by text size and shape, especially complex stroke details. Besides, some recent works [4, 25, 26, 38, 48, 50] utilize more robust visual features, such as line-level OCR features as control conditions to further improve the text accuracy. But the line-level visual features still perform poorly in representing stroke details for each character. To address this issue, GlyphByT5 [23, 24] proposes a method with box-level contrastive learning to align the text features extracted from the language model with the features extracted from the visual encoder. To effectively learn such alignment, GlyphByT5 relies on collecting massive amounts of data and developing complex data augmentation strategies for the alignment pre-training, which lacks flexibility. In contrast, in this paper, we reveal that the key to high-accuracy text rendering lies in constructing discriminative character-level visual features. Thus we propose a plug-and-play and robust character-level text representation derived from off-the-shelf OCR encoders, which can accurately represent the visual structure of the text without additional training and enable precise text rendering." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 599, + 507, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 507, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 507, + 612 + ], + "type": "text", + "content": "2.3. Subject-Preserved Scene Generation" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": "To create a scene image with a product subject while ensuring subject fidelity, two main methods are commonly used. One is the subject-driven method [3, 6, 19, 32, 36], which adjusts the position, angle and lighting of the subject based on the prompt to create a harmonious image. However, it often struggles to preserve the significant features of the subject. The other utilizes inpainting-based background completion techniques [2, 10, 39]. It only generates the non-" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8085" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 550, + 207 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 550, + 207 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 550, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 550, + 207 + ], + "type": "image", + "image_path": "f4b117659f00639c8b4bd3674bba491facc16adbac19530cd80757917def6165.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 553, + 232 + ], + "type": "text", + "content": "Figure 3. The framework of the PosterMaker, which is based on the SD3. To precisely generate multilingual texts and create aesthetically pleasing poster scenes, TextRenderNet and SenceGenNet are introduced, whose outputs are used as control conditions added to the SD3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 323 + ], + "type": "text", + "content": "subject areas of an image and naturally keeps consistency in the original subject area. But it sometimes extends the foreground subject [10, 11], such as adding an extra handle to a cup, which also reduces subject fidelity. To maximize subject fidelity, our method uses background completion and a reward model to determine whether the foreground extension occurred, thereby enhancing subject fidelity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 328, + 111, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 111, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 111, + 340 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 180, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 180, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 180, + 361 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 367, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 296, + 438 + ], + "type": "text", + "content": "This paper focuses on the creation of product posters, which typically consist of multiple elements such as text, subjects, and scenes, as illustrated in Fig. 1 (a). The central challenge of this task is to generate these elements accurately and harmoniously, offering both research and practical applications. The task is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 448, + 295, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 448, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 129, + 448, + 295, + 462 + ], + "type": "interline_equation", + "content": "I _ {g} = f \\left(I _ {s}, M _ {s}, T, P\\right), \\tag {1}", + "image_path": "e9123f9df8db6d099ecd30090fbb250e35e84693dcafa9f867f85d88ef17d42c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "inline_equation", + "content": "I_{g}" + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "text", + "content": " denotes the generated poster image, " + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "text", + "content": " represents the subject image, and " + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "text", + "content": " is the subject mask. The variable " + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "text", + "content": " signifies the content and the position of text and " + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "text", + "content": " is the prompt describing the background scene. Subsequent sections will detail the design of PosterMaker, and our proposed solution to this task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 551, + 134, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 134, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 134, + 562 + ], + "type": "text", + "content": "3.2. Framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "As shown in Fig. 3, PosterMaker is developed based on Stable Diffusion 3 (SD3) [12], which contains a strong VAE for reconstructing the image details like text stroke. And we propose two modules, i.e., TextRenderNet and SceneGenNet, to address the poster generation task. TextRenderNet is specifically designed to learn visual text rendering, taking character-level visual text representations as input to achieve precise and controllable text rendering. SceneGenNet, on the other hand, accepts a masked image (indicating which content should remain unchanged) and a prompt, learning to generate the foreground subject within the desired scene described by the prompt. Both TextRenderNet" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 325, + 236, + 547, + 361 + ], + "blocks": [ + { + "bbox": [ + 325, + 236, + 547, + 361 + ], + "lines": [ + { + "bbox": [ + 325, + 236, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 325, + 236, + 547, + 361 + ], + "type": "image", + "image_path": "042150ef1df01697695703c3c6f7189b7317020c8da26186af947fd013b306a5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 363, + 553, + 385 + ], + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 385 + ], + "type": "text", + "content": "Figure 4. The details of TextRenderNet and SceneGenNet, showcasing their model architectures and their interactions with SD3." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 392, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 632 + ], + "type": "text", + "content": "and SceneGenNet are grounded in a ControlNet-like [48] architecture derived from SD3 and their architectures are detailed in Fig. 4. They share the same internal structure, comprising several cascaded MM-DiT blocks [12], with weights copied from the base model for initialization. The output of each MM-DiT block is added to the corresponding block of the base model after passing through a zero convolution layer [49]. The key distinction between the two modules lies in their input configurations. SceneGenNet takes the prompt as input to the text condition branch, and for the visual branch, the input is derived by the latent feature at timestep " + }, + { + "bbox": [ + 313, + 392, + 555, + 632 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 392, + 555, + 632 + ], + "type": "text", + "content": ", the subject mask, and the masked latent to preserve the foreground area. In contrast, TextRenderNet receives text representations (detailed in the next section) in the text condition branch for text rendering. An adapter, consisting of a linear layer and layer normalization, adjusts the feature dimensions of these text representations before they are input to TextRenderNet. The outputs of each block in TextRenderNet and SceneGenNet are directly added to the corresponding block outputs of the SD3 base model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 637, + 553, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 553, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 553, + 663 + ], + "type": "text", + "content": "3.3. Character-level Visual Representation for Precise Text Rendering" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Recently, some works have explored multilingual visual text generation. Among them, a promising approach is based on ControlNet-like methods [38], which utilize both glyph images and line-level OCR features as conditions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8086" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 295, + 121 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 295, + 121 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 295, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 295, + 121 + ], + "type": "image", + "image_path": "a2487c0173392901a1993e9f7b58903ed376fa0038493606e2b51bf152014ec4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 122, + 295, + 156 + ], + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 156 + ], + "type": "text", + "content": "Figure 5. The distinction between the multilingual character-level text representation we proposed and the line-level methods of previous works like AnyText [38] and GlyphDraw2 [26]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 161, + 296, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 296, + 257 + ], + "type": "text", + "content": "However, this control information cannot accurately represent characters: 1) glyph images are easily affected by text size and shape, making them less robust. 2) line-level visual features lack fine-grained stroke features and are limited by the OCR model's poor capability to recognize long texts. To address these challenges, this paper proposes a plug-and-play and robust character-level text representation, where each character is precisely represented by one token." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "text", + "content": "Specifically, the text " + }, + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "text", + "content": " characters. For each character " + }, + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "text", + "content": ", its feature is separately extracted by a pre-trained OCR encoder " + }, + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "text", + "content": " and then averaged and pooled to obtain a compact character representation vector " + }, + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "inline_equation", + "content": "r_{c_i} \\in \\mathbb{R}^c" + }, + { + "bbox": [ + 55, + 257, + 296, + 317 + ], + "type": "text", + "content": ". Thus, the character-level text representation is defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 124, + 321, + 294, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 321, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 124, + 321, + 294, + 335 + ], + "type": "interline_equation", + "content": "r _ {c i} = \\operatorname {a v g p o o l} \\left(f _ {v} \\left(I _ {c i}\\right)\\right), \\tag {2}", + "image_path": "806a349677a8601725d6a519604607bc72f697ce6a3850363d3f8cf320b86147.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 335, + 294, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 335, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 130, + 335, + 294, + 350 + ], + "type": "interline_equation", + "content": "R _ {c} = \\left[ r _ {c _ {1}}, r _ {c _ {2}}, \\dots , r _ {c _ {n}} \\right], \\tag {3}", + "image_path": "7ffd89c62d0e1925795ad91a8529a95dca878d95b47cd8e957e31ca20e9599fe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 358, + 295, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 382 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 358, + 295, + 382 + ], + "type": "inline_equation", + "content": "I_{c_i}" + }, + { + "bbox": [ + 55, + 358, + 295, + 382 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 358, + 295, + 382 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 358, + 295, + 382 + ], + "type": "text", + "content": "-th character image rendered in a fixed font, and " + }, + { + "bbox": [ + 55, + 358, + 295, + 382 + ], + "type": "inline_equation", + "content": "R_{c} \\in \\mathbb{R}^{n \\times c}" + }, + { + "bbox": [ + 55, + 358, + 295, + 382 + ], + "type": "text", + "content": " is the char-level text representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 383, + 296, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 490 + ], + "type": "text", + "content": "As shown in Fig. 5, compared to previous methods, our key difference is extracting representations from character glyph images. This enables the model to perceive character stroke structures and achieve high text accuracy. Additionally, since the number of characters is fixed, we can pre-extract the representations of each character and store them in a dictionary, eliminating the need for online rendering and feature extraction. This significantly simplifies the training and inference pipeline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "text", + "content": "Finally, this text representation lacks order and positional information. Thus, the character order encoding " + }, + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "inline_equation", + "content": "P_{rank}" + }, + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "text", + "content": " is introduced to represent the order of characters in the text, which is implemented through a sinusoidal position encoding of the char order. Besides, inspired by GLIGEN [20], the text position coordinates are mapped to sinusoidal position encoding " + }, + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "inline_equation", + "content": "P_{bbox}" + }, + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "text", + "content": " to control the position of the text. Then we concatenate " + }, + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "inline_equation", + "content": "P_{rank}" + }, + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "inline_equation", + "content": "P_{bbox}" + }, + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "inline_equation", + "content": "R_c" + }, + { + "bbox": [ + 55, + 491, + 296, + 597 + ], + "type": "text", + "content": " along the feature dimension to construct the final text representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 602, + 205, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 205, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 205, + 615 + ], + "type": "text", + "content": "3.4. Improving Subject Fidelity" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "In the task of generating product posters, it is crucial to maintain subject fidelity, i.e., ensuring that the subject in the generated poster remains consistent with the user-specified subject. To achieve this goal, we employ SceneGenNet to perform background inpainting, which is trained to precisely preserve the foreground subject and only inpaint the background according to the prompt. However, inpainting-based models sometimes extend the foreground subject into" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 69, + 542, + 167 + ], + "blocks": [ + { + "bbox": [ + 321, + 69, + 542, + 167 + ], + "lines": [ + { + "bbox": [ + 321, + 69, + 542, + 167 + ], + "spans": [ + { + "bbox": [ + 321, + 69, + 542, + 167 + ], + "type": "image", + "image_path": "710505a03bc6f8844bdcd454e6243181700a8f0265f7bd6e2dee32382cd07913.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 170, + 552, + 180 + ], + "lines": [ + { + "bbox": [ + 315, + 170, + 552, + 180 + ], + "spans": [ + { + "bbox": [ + 315, + 170, + 552, + 180 + ], + "type": "text", + "content": "Figure 6. The model details of the foreground extension detector." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 186, + 421, + 261 + ], + "blocks": [ + { + "bbox": [ + 316, + 186, + 421, + 261 + ], + "lines": [ + { + "bbox": [ + 316, + 186, + 421, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 421, + 261 + ], + "type": "image", + "image_path": "ce3b15cc56ae8e48077fda94b7c57af0ec7f4b3f143c543b9d3235facd61e7f8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 265, + 553, + 286 + ], + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 286 + ], + "type": "text", + "content": "Figure 7. The illustration of our two-stage training strategy for efficiently optimizing PosterMaker." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 429, + 186, + 550, + 263 + ], + "blocks": [ + { + "bbox": [ + 429, + 186, + 550, + 263 + ], + "lines": [ + { + "bbox": [ + 429, + 186, + 550, + 263 + ], + "spans": [ + { + "bbox": [ + 429, + 186, + 550, + 263 + ], + "type": "image", + "image_path": "144c221f07e4d286416725b244293eebfd4d20a6634e8dc9ab9801c828d020a6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 295, + 554, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 554, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 554, + 355 + ], + "type": "text", + "content": "another subject (as shown in Fig. 2 (b)) thereby compromising subject fidelity. We refer to this as \"foreground extension\". To mitigate this issue, we develop a model to detect foreground extension and employ it as a reward model to fine-tune PosterMaker to improve subject fidelity." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "text", + "content": "Foreground Extension Detector. We develop the foreground extension detector " + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "text", + "content": " based on HQ-SAM [16]. As shown in Fig. 6, we input the generated image " + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "inline_equation", + "content": "I_{g}" + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "text", + "content": " to SAM [17] image encoder. The subject mask " + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "text", + "content": " and box " + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "inline_equation", + "content": "B_{s}" + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "text", + "content": " are provided as mask prompt and box prompt, respectively, to the HQ-SAM decoder to obtain an intermediate mask " + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "text", + "content": ". Next, we concatenate the image features extracted from SAM encoder with " + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "inline_equation", + "content": "M_{s} - M_{i}" + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "text", + "content": " at the channel dimension. The concatenated features are processed through convolutional layers and MLP layers to predict whether the foreground has been extended in the generated image. We collected 20k manually annotated images to train the foreground extension detector " + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 313, + 356, + 554, + 511 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": "Subject Fidelity Feedback Learning. The foreground extension detector " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": ", after the offline training, is used as a reward model to supervise PosterMaker to improve subject fidelity. Specifically, assuming the reverse process has a total of " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": " steps, we follow ReFL [43] to first sample " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "z_{T'} \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": " and after " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "T' - t'" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": " steps of inference " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "(z_{T'} \\rightarrow z_{T'-1} \\rightarrow \\dots \\rightarrow z_{t'})" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": ", we obtain " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "z_{t'}" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "t' \\sim [1, t_1]" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": ". Then, we directly perform a one-step inference " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "z_{t'} \\rightarrow z_0" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": " to accelerate the reverse process. Furthermore, " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": " is decoded to the generated image " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": ". The detector " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": " predicts the foreground extension score for " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": ", and this score is used as the reward loss to optimize the generator " + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "inline_equation", + "content": "G_{\\phi}" + }, + { + "bbox": [ + 313, + 513, + 555, + 668 + ], + "type": "text", + "content": " (i.e., PostMaker). The reward loss is defined as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 673, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 673, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 673, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {r e w a r d}} (\\phi) = - \\mathbb {E} _ {(x, c, m) \\sim \\mathcal {D} _ {\\text {t r a i n}}, t ^ {\\prime} \\sim [ 1, t _ {1} ], z _ {T ^ {\\prime}} \\sim \\mathcal {N} (0, 1)} \\\\ \\log \\sigma \\left(1 - S _ {\\theta} \\left(G _ {\\phi} \\left(z _ {T ^ {\\prime}}, x, c, m, t ^ {\\prime}\\right), m\\right)\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "d2a12c03aaad424737c750f680f77f2d7c43f9a133f0e7e4cce1d482302f72b2.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8087" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 553, + 296 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 553, + 296 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 553, + 296 + ], + "type": "image", + "image_path": "0070060a364d9b3a711de82fba80cfa82d9491fd8c946ccda917b1fe47648d5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison with different methods. Best viewed on Screen. To aid comprehension, Chinese text lines in the image are translated into English and annotated using corresponding colors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "inline_equation", + "content": "x, c, m" + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "text", + "content": " sampled from the train data " + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{train}}" + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "text", + "content": ", represent the subject image, control conditions, and subject mask respectively. To avoid overfitting, we don't calculate reward loss for the cases where the foreground extension score is below 0.3. Our total training loss is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 397, + 295, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 397, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 119, + 397, + 295, + 410 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {d e n o i s e}} + \\lambda \\mathcal {L} _ {\\text {r e w a r d}}, \\tag {5}", + "image_path": "fe5a50ace377dd78b0cdc95153dff52855788196b32ce40af46dbd2a289d55b9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 424, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 295, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 424, + 295, + 448 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 424, + 295, + 448 + ], + "type": "text", + "content": " is the hyperparameter to adjust the weight of reward loss and the denoise loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 162, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 162, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 162, + 482 + ], + "type": "text", + "content": "3.5. Training Strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 489, + 296, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 296, + 717 + ], + "type": "text", + "content": "To efficiently train PosterMaker, this paper introduces a two-stage training strategy, as shown in Fig. 7, aimed at decoupling the learning for text rendering and background image generation. Specifically, in the first stage, the training task is local text editing. We freeze SceneGenNet and only the TextRenderNet and adapter are optimized. Since we initialize SceneGenNet with pre-trained weights of inpainting-controlnet [7], it can fill the local background well thus TextRenderNet can focus on learning text generation. In the second stage, the training task is subject-based text-to-image generation. Here we froze TextRenderNet and only train the SceneGenNet. In this stage, SceneGenNet focuses on learning poster scenes and creative design from the train data. Notably, Stage 1 learns local text editing/inpainting and Stage 2 learns background inpainting, thus the input images indicating the area to inpaint are different (See Fig. 7). With such a two-stage training strategy, TextRenderNet and SceneGenNet can be efficiently optimized since they can focus on their specific tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 321, + 395, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 395, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 395, + 335 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 338, + 430, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 430, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 430, + 351 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 354, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 557 + ], + "type": "text", + "content": "Dataset. We crawl product posters from online e-commerce platforms to construct our training set. Our training data mainly consists of Chinese posters, we first employ PPOCRv4 model [30] to extract the text content and their bounding boxes from the images as a coarse annotation. And we ask some annotators to further refine the bounding boxes and correct the text content to improve the annotation quality. Resulting in a dataset containing 160k images. We generate image captions with GPT4-o [29] and extract foreground subject masks with " + }, + { + "bbox": [ + 313, + 354, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{U}^2" + }, + { + "bbox": [ + 313, + 354, + 555, + 557 + ], + "type": "text", + "content": "-Net [33] and VitMatte [46]. We randomly select 302 images for evaluation and leave the rest for training. To better evaluate the performance of our method, we use LLM [9] to generate some background prompts and text layouts as evaluation samples, after manually checking and removing those irrational ones, we obtain another 198 evaluation samples to form a final evaluation set named PosterBenchmark containing 500 samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. We follow Anytext [38] to evaluate text rendering accuracy using two metrics: sentence accuracy (Sen. Acc) and normalized edit distance (NED). Specifically, we crop the text line from the generated image according to the provided bounding box and utilize the OCR model [28] to predict the content " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{pred}}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": " of the generated text line. We denote the ground truth text content as " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": ". A text line is considered to be correctly generated if " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{pred}} = s_{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": "; this condition is used to calculate Sen. Acc. Additionally, we compute the normalized edit distance (NED) between " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{pred}}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": " to measure their similarity. We further calculate FID [14] to measure the visual quality and CLIP-T [36] metric for evaluating text-image alignment." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8088" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 553, + 294 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 294 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 294 + ], + "type": "image", + "image_path": "1e5da996d5adf47c36398dbed7f22968c7167086e4f13cb775bcc235b7d5f337.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 294, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 555, + 316 + ], + "type": "text", + "content": "Figure 9. Qualitative comparison using various text features. It is obvious that the character-level OCR features we used (PPOCR Char) are the most effective at maintaining character accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 323, + 297, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 323, + 297, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 297, + 515 + ], + "type": "text", + "content": "Implementation Details. Our SceneGenNet is initialized from pre-trained SD3 Inpainting-Controlnet [7] and TextRenderNet is initialized from SD3 [12] weight with the same configuration as in [8]. For Subject Fidelity Feedback Learning, we follow existing work [43] to uniformly sample " + }, + { + "bbox": [ + 54, + 323, + 297, + 515 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 54, + 323, + 297, + 515 + ], + "type": "text", + "content": " between [1, 10]. Within this range, the one-step inference result of image " + }, + { + "bbox": [ + 54, + 323, + 297, + 515 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 54, + 323, + 297, + 515 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 323, + 297, + 515 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 54, + 323, + 297, + 515 + ], + "type": "text", + "content": " is close to the full inference result. The weight coefficient of " + }, + { + "bbox": [ + 54, + 323, + 297, + 515 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 323, + 297, + 515 + ], + "type": "text", + "content": " is set to 0.0005. The learning rate is set to 1e-4 and the batch size is set to 192. We train our framework for 26k and 29.5k steps for training stage1 and stage2, respectively. Finally, PosterMaker was trained on 32 A100 GPUs for 3 days. During the sampling process, based on the statistical information, a maximum of 7 lines of text and 16 characters per line of text are selected from each image to render onto the image, as this setting can cover most situations in the dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 521, + 222, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 222, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 222, + 533 + ], + "type": "text", + "content": "4.2. Comparison with Prior Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 536, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 693 + ], + "type": "text", + "content": "Baseline methods. We carefully designed the following baseline approaches based on existing open-sourced techniques for comparative analysis. SD3_inpaint_byt5: We encode the text content into prompt embeddings using ByT5 [44] and employ an adapter to map these embeddings to the original prompt embedding space of SD3 before feeding them into the controlnet, which enables the controlnet to render multilingual text. SD3_canny&inpaint: First render the text into a white-background image and extract the canny edge from it as control. Then finetune a pre-trained SD3 canny controlnet together with an inpainting controlnet to achieve multilingual text rendering. Anytext: It is the SOTA open-sourced T2I method that supports multilin" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 320, + 550, + 423 + ], + "blocks": [ + { + "bbox": [ + 317, + 320, + 550, + 423 + ], + "lines": [ + { + "bbox": [ + 317, + 320, + 550, + 423 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 550, + 423 + ], + "type": "table", + "html": "
ModelSen. ACC ↑NED ↑FID ↓CLIP-T ↑FG Ext. Ratio ↓
SD3 inpaint_AnyText52.78%75.27%100.8726.9014.82%
SD3 inpaint_byt552.28%86.57%65.4526.7114.60%
AnyText63.90%82.81%71.2726.6919.25%
Glyph-ByT5-v269.54%87.65%79.2326.6018.91%
SD3_canny&inpaint80.75%92.75%67.1927.0314.38%
GlyphDraw286.14%96.78%72.4926.7216.52%
GT (w/ SD1.5 Rec.)76.95%89.91%---
GT (w/ SD3 Rec.)98.09%99.36%---
GT98.53%99.59%---
Ours (SD1.5)72.12%88.01%68.1726.93-
Ours93.36%98.39%65.3527.0411.57%
", + "image_path": "c5a57855734fcf807b46be2eb5b85b9a9826b9ab1fdafffb7b589e4271f68b52.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 352, + 424, + 515, + 434 + ], + "lines": [ + { + "bbox": [ + 352, + 424, + 515, + 434 + ], + "spans": [ + { + "bbox": [ + 352, + 424, + 515, + 434 + ], + "type": "text", + "content": "Table 1. Comparison with baseline methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 437, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 568 + ], + "type": "text", + "content": "gual text rendering and its text editing mode supports text inpainting [38]. So we directly finetune it on our data using its text editing training pipeline. SD3_inpaint_Anytext: First generate the background with SD3 inpainting control-net, then render the text on the corresponding region using Anytext. Glyph-ByT5-v2 and GlyphDraw2: They are both the SOTA T2I methods that support multilingual text rendering [24, 26]. However, they don't have open-sourced pre-trained weights, so we reproduced them on our dataset. And we added an inpainting controlnet for them to support subject-preserved generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Quantitative Comparison. We trained all baseline models on the same dataset, and then quantitatively compared all methods on the PosterBenchmark, as shown in Tab. 1. It is worth noting that SD3 is used as the base model by default, but since we observed that the SD1.5 VAE leads to significant error in reconstruction, to enable a more equitable comparison between our method and AnyText (SD1.5 architecture), we also implemented an SD1.5 version of PosterMaker with the same experimental setup as AnyText. As the VAEs, especially SD1.5, introduce some reconstruction error and the OCR model may incorrectly recognize some characters, we also report the metrics on ground truth" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 194, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 194, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 194, + 713 + ], + "type": "text", + "content": "1Details can be found in the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8089" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 270, + 159 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 270, + 159 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 270, + 159 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 270, + 159 + ], + "type": "table", + "html": "
Text FeatureTypeSen. ACCNED
ByT5textual feat.33.48%54.50%
Cannyimg81.50%92.72%
TrOCR Linevisual feat.26.58%49.46%
TrOCR Charvisual feat.94.27%98.54%
PPOCR Linevisual feat.38.91%53.86%
PPOCR Char (Ours)visual feat.95.15%98.75%
GT (w/o Rec.)-98.53%99.59%
GT (w/ SD3 Rec.)-98.09%99.36%
", + "image_path": "f3f05887672be46327d87fc359aea44d5457c5e334261d53af6cc072989a02d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 173, + 291, + 201 + ], + "blocks": [ + { + "bbox": [ + 64, + 159, + 286, + 170 + ], + "lines": [ + { + "bbox": [ + 64, + 159, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 64, + 159, + 286, + 170 + ], + "type": "text", + "content": "Table 2. Quantitative comparison using various text features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 173, + 291, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 173, + 291, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 173, + 291, + 201 + ], + "type": "table", + "html": "
MethodFG Ext. Ratio↓Sen. ACC ↑NED↑FID↓CLIP-T↑
Ours11.57%93.36%98.39%65.3527.04
Ours w/o Lreward15.05%93.11%98.21%65.1027.04
", + "image_path": "e65d1a650bc7ff1db543d002925588ec2a67d84bcdcabf36c7fd8d7604d9bdd6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 201, + 286, + 213 + ], + "lines": [ + { + "bbox": [ + 64, + 201, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 64, + 201, + 286, + 213 + ], + "type": "text", + "content": "Table 3. Evaluation on the subject fidelity feedback learning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 215, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 295, + 407 + ], + "type": "text", + "content": "images as an upper bound. As shown in Tab. 1, our method achieves the best performance on all metrics. Notably, on text rendering metrics Sen. ACC and NED, our model outperforms the baselines by an impressive margin and is already close to the upper bound. The promising results demonstrate the effectiveness of the proposed PosterMaker. Qualitative Comparison. The results are shown in Fig. 8. Compared to the baselines, our PosterMaker generates more readable and accurate poster images with texts, particularly for smaller texts. Notably, as an end-to-end generation method, PosterMaker automatically creates underlays to enhance the contrast between text and background, effectively highlighting the text. This feature is crucial in product poster design for capturing viewers' attention. These findings demonstrate that our PosterMaker successfully learns the distribution of posters created by human designers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 411, + 212, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 212, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 212, + 423 + ], + "type": "text", + "content": "4.3. Ablation Study and Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "type": "text", + "content": "How to achieve high text rendering accuracy? We conduct experiments to explore the effectiveness of different control conditions for visual text rendering. Due to the fact that text rendering accuracy is primarily determined by the first training stage, we discard the second training stage in this experiment to save computational resources. The results are summarized in Tab. 2. We observed several valuable experimental results: 1) The use of char-level features significantly outperforms previous line-level features, benefiting from finer-grained representation. This explains why previous methods [4, 26, 38], achieve inferior performance (PPOCR Line is used in [26, 38], TrOCR Line is used in [4]). Recent concurrent works [27, 42] have also found similar experimental findings as ours. 2) Char-level feature representation is superior to low-level image features such as Canny. 3) PPOCR outperforms TrOCR, which is attributed to PPOCR being a multi-language OCR model, while TrOCR is an English version model. 4) Even though TrOCR has not been trained on multi-language text data, it still achieves decent results, likely because it extracts universal visual structural features. 5) ByT5 extracts char-level features but the performance is inferior to OCR features, because it extracts semantic features rather than character structural features, while T2I models' text rendering" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 320, + 70, + 386, + 121 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 386, + 121 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 386, + 121 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 386, + 121 + ], + "type": "image", + "image_path": "2e867053f7adcdd70aee505f1f17453b337873f1f67286a27fd6d271f02ade6e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 388, + 70, + 552, + 187 + ], + "blocks": [ + { + "bbox": [ + 388, + 70, + 552, + 187 + ], + "lines": [ + { + "bbox": [ + 388, + 70, + 552, + 187 + ], + "spans": [ + { + "bbox": [ + 388, + 70, + 552, + 187 + ], + "type": "image", + "image_path": "b8d73b6df260a6783137dee1b957ecef10b8e5eed6ebbc01fb5548e27544670f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 188, + 541, + 200 + ], + "lines": [ + { + "bbox": [ + 326, + 188, + 541, + 200 + ], + "spans": [ + { + "bbox": [ + 326, + 188, + 541, + 200 + ], + "type": "text", + "content": "Figure 10. Visual examples showing the effect of " + }, + { + "bbox": [ + 326, + 188, + 541, + 200 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reward}" + }, + { + "bbox": [ + 326, + 188, + 541, + 200 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 209, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 376 + ], + "type": "text", + "content": "capability relies more on character structural features. We present visualization results in Fig. 9. We observe that when using line-level features as a control, the generated text occasionally becomes completely unrecognizable. This suggests that line-level features are insufficient for achieving precise text rendering. Additionally, it is evident that using canny control always introduces stroke artifacts, particularly in smaller texts (as seen in row 3 of Fig. 9). This further demonstrates that canny control is also not an ideal condition for text rendering. In summary, the char-level feature extracted by PPOCR performs optimally and the accuracy is already close to the upper bound, indicating the discriminative char-level visual feature is the key to achieve high text rendering accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 556, + 521 + ], + "type": "text", + "content": "Effectiveness of subject fidelity feedback learning. We calculate the foreground extension ratio (termed as FG Ext. Ratio) by asking human annotators to manually check each generated image whether the foreground subject is incorrectly extended. As demonstrated in Tab. 3, training our model with " + }, + { + "bbox": [ + 313, + 376, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reward}" + }, + { + "bbox": [ + 313, + 376, + 556, + 521 + ], + "type": "text", + "content": " effectively reduces FG Ext. Ratio by " + }, + { + "bbox": [ + 313, + 376, + 556, + 521 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 376, + 556, + 521 + ], + "type": "text", + "content": ", while maintaining subtle variations in other performance metrics. Representative visual examples are presented in Fig. 10. Besides, our model outperforms baseline methods in FG Ext. Ratio (see Tab. 1). These results show the efficacy of our proposed subject fidelity feedback learning approach in mitigating foreground extension artifacts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 528, + 388, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 388, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 388, + 540 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "The application of image generation in poster creation is often impeded by subpar text rendering and inconsistent subjects. To address these challenges, this paper introduces a novel framework, PosterMaker, which synthesizes aesthetically pleasing product posters with accurate and harmonious texts and contents. Moreover, we reveal that the key underlying successful multilingual text rendering is the construction of robust character-level visual text representations. Additionally, we propose subject fidelity feedback learning to mitigate inconsistencies in subjects. Through extensive experiments, our method demonstrates a significant improvement in both high-precision text generation and subject fidelity. These findings not only advance poster generation but also inspire future research on T2I models." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8090" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work was supported by the National Nature Science Foundation of China (62425114, 62121002, U23B2028, 62232006, 62272436) and Alibaba Group (Alibaba Research Intern Program)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 172, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 172, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 172, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 172, + 296, + 228 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 229, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 229, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 229, + 296, + 316 + ], + "type": "text", + "content": "[2] Tingfeng Cao, Junsheng Kong, Xue Zhao, Wenqing Yao, Junwei Ding, Jinhui Zhu, and Jiandong Zhang. Product2img: Prompt-free e-commerce product background generation with diffusion model and self-improved LMM. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pages 10774-10783. ACM, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 319, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 319, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 61, + 319, + 296, + 384 + ], + "type": "text", + "content": "[3] Kelvin C. K. Chan, Yang Zhao, Xuhui Jia, Ming-Hsuan Yang, and Huisheng Wang. Improving subject-driven image synthesis with subject-agnostic guidance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 6733-6742. IEEE, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 386, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 386, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 61, + 386, + 296, + 463 + ], + "type": "text", + "content": "[4] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Jun Lan, Xing Zheng, Yaohui Li, Changhua Meng, Huijia Zhu, and Weiqiang Wang. Diffuse: Universal text editing diffusion model. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2, 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 465, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 296, + 520 + ], + "type": "text", + "content": "[5] Ruidong Chen, Lanjun Wang, Weizhi Nie, Yongdong Zhang, and An-An Liu. Anyscene: Customized image synthesis with composited foreground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8724-8733, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 521, + 296, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 296, + 598 + ], + "type": "text", + "content": "[6] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W. Cohen. Subject-driven text-to-image generation via apprenticeship learning. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 600, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 600, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 600, + 295, + 633 + ], + "type": "text", + "content": "[7] Alimama Creative. Sd3-controlnet-inpainting. https://huggingface.co/alimama-creative/SD3-Controlnet-Inpainting, 2024.6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 635, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 295, + 667 + ], + "type": "text", + "content": "[8] Alimama Creative. Sd3-controlnet-softedge. https://huggingface.co/alimama-creative/SD3-Controlnet-Softedge, 2024.7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 714 + ], + "type": "text", + "content": "[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "text", + "content": "Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Ji-aqi Wang. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "text", + "content": "[10] Zhenbang Du, Wei Feng, Haohan Wang, Yaoyu Li, Jingsen Wang, Jian Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junsheng Jin, et al. Towards reliable advertising image generation using human feedback. In European Conference on Computer Vision, pages 399-415. Springer, 2024. 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 556, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 556, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 556, + 250 + ], + "type": "text", + "content": "[11] Amir Erfan Eshratifar, Joao V.B. Soares, Kapil Thadani, Shaunak Mishra, Mikhail Kuznetsov, Yueh-Ning Ku, and Paloma De Juan. Salient object-aware background generation using text-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 7489-7499, 2024. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 253, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 319 + ], + "type": "text", + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 2, 3, 4, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 320, + 555, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 555, + 385 + ], + "type": "text", + "content": "[13] Yifan Gao, Jinpeng Lin, Min Zhou, Chuanbin Liu, Hongtao Xie, Tiezheng Ge, and Yuning Jiang. Textpainter: Multimodal text image generation with visual-harmony and text-comprehension for poster design. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7236-7246, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 388, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 464 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 6626-6637, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 466, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 554, + 520 + ], + "type": "text", + "content": "[15] Peidong Jia, Chenxuan Li, Yuhui Yuan, Zeyu Liu, Yichao Shen, Bohan Chen, Xingru Chen, Yinglin Zheng, Dong Chen, Ji Li, Xiaodong Xie, Shanghang Zhang, and Baining Guo. Cole: A hierarchical generation framework for multilayered and editable graphic design, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 522, + 554, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 554, + 576 + ], + "type": "text", + "content": "[16] Lei Ke, Mingqiao Ye, Martin Danelljan, Yifan liu, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In Advances in Neural Information Processing Systems, pages 29914–29934. Curran Associates, Inc., 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 578, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 554, + 644 + ], + "type": "text", + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 647, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 554, + 678 + ], + "type": "text", + "content": "[18] Chao Li, Chen Jiang, Xiaolong Liu, Jun Zhao, and Guoxin Wang. Joytype: A robust design for multilingual visual text creation. arXiv preprint arXiv:2409.17524, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 680, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 714 + ], + "type": "text", + "content": "[19] Dongxu Li, Junnan Li, and Steven C. H. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. In Advances in Neural Infor" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8091" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "type": "text", + "content": "mation Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[20] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[21] Zhaochen Li, Fengheng Li, Wei Feng, Honghe Zhu, An Liu, Yaoyu Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junjie Shen, et al. Planning and rendering: Towards end-to-end product poster generation. arXiv preprint arXiv:2312.08822, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 283 + ], + "type": "text", + "content": "[22] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1250-1260, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 340 + ], + "type": "text", + "content": "[23] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. In European Conference on Computer Vision, pages 361-377. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "type": "text", + "content": "[24] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "text", + "content": "[25] Zhiying Lu, Chuanbin Liu, Xiaojun Chang, Yongdong Zhang, and Hongtao Xie. Dhvt: Dynamic hybrid vision transformer for small dataset recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 475 + ], + "type": "text", + "content": "[26] Jian Ma, Yonglin Deng, Chen Chen, Haonan Lu, and Zhenyu Yang. Glyphdraw2: Automatic generation of complex glyph posters with diffusion models and large language models. arXiv preprint arXiv:2407.02252, 2024. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "type": "text", + "content": "[27] Lichen Ma, Tiezhu Yue, Pei Fu, Yujie Zhong, Kai Zhou, Xiaoming Wei, and Jie Hu. Chargen: High accurate character-level visual text generation model with multimodal encoder. arXiv preprint arXiv:2412.17225, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "text", + "content": "[28] ModelScope. https://modelscope.cn/models/damo/cv_convnextTinyOCR-recognition-general_damo/summary, 2023.6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 578 + ], + "type": "text", + "content": "[29] OpenAI. https://openai.com/index/hello-gpt-40/, 2024.6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 601 + ], + "type": "text", + "content": "[30] PaddlePaddle. https://github.com/PaddlePaddle/PaddleOCR, 2023.6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 677 + ], + "type": "text", + "content": "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. In *The Twelfth International Conference on Learning Representations*, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[32] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentan" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "gled representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8693-8702, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "type": "text", + "content": "[33] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. page 107404, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "type": "text", + "content": "[34] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and PeterJ. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv: Learning, arXiv: Learning, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 274 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 341 + ], + "type": "text", + "content": "[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 409 + ], + "type": "text", + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[38] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 467, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 533 + ], + "type": "text", + "content": "[39] Haohan Wang, Wei Feng, Yaoyu Li, Zheng Zhang, Jingjing Lv, Junjie Shen, Zhangang Lin, and Jingping Shao. Generate e-commerce product background by integrating category commonality and personalized style. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2025. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 534, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 576 + ], + "type": "text", + "content": "[40] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 654 + ], + "type": "text", + "content": "[41] Shaodong Wang, Yunyang Ge, Liuhan Chen, Haiyang Zhou, Qian Wang, Xinhua Cheng, and Li Yuan. Prompt2poster: Automatically artistic chinese poster creation from prompt only. In Proceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024, pages 10716-10724. ACM, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[42] Tong Wang, Xiaochao Qu, and Ting Liu. Textmastero: Mastering high-quality scene text editing in diverse languages and styles. arXiv preprint arXiv:2408.10623, 2024. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[43] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8092" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 475 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 116 + ], + "type": "text", + "content": "ward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, pages 15903-15935. Curran Associates, Inc., 2023. 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[44] Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. ByT5: Towards a token-free future with pre-trained byte-to-byte models. Transactions of the Association for Computational Linguistics, 10:291-306, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 250 + ], + "type": "text", + "content": "[45] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 295 + ], + "type": "text", + "content": "[46] Jingfeng Yao, Xinggang Wang, Shusheng Yang, and Baoyuan Wang. Vitmatte: Boosting image matting with pretrained plain vision transformers. Information Fusion, 103: 102091, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 339 + ], + "type": "text", + "content": "[47] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 341, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 294, + 384 + ], + "type": "text", + "content": "[48] Boqiang Zhang, Zuan Gao, Yadong Qu, and Hongtao Xie. How control information influences multilingual text image generation and editing? arXiv preprint arXiv:2407.11502, 2024. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 430 + ], + "type": "text", + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 475 + ], + "type": "text", + "content": "[50] Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, and Zhibo Yang. Visual text generation in the wild. In European Conference on Computer Vision, pages 89-106. Springer, 2024. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8093" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/28953876-24ed-4c02-8fbd-02d889f7bc8b_content_list.json b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/28953876-24ed-4c02-8fbd-02d889f7bc8b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c7dfa9a59ba47615663bb66f05a50a1b01d67410 --- /dev/null +++ b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/28953876-24ed-4c02-8fbd-02d889f7bc8b_content_list.json @@ -0,0 +1,1635 @@ +[ + { + "type": "text", + "text": "PosterO: Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation", + "text_level": 1, + "bbox": [ + 179, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "HsiaoYuan Hsu and Yuxin Peng* \nWangxuan Institute of Computer Technology, Peking University", + "bbox": [ + 243, + 203, + 751, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kslh99@stu.pku.edu.cn, pengyuxin@pku.edu.cn", + "bbox": [ + 305, + 242, + 691, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/238b9e451c661a20780a71bf7fd062a0ca0425da339f0ba93cd0285c85253502.jpg", + "image_caption": [ + "Figure 1. Illustration of the generalized settings in content-aware layout generation from both (a) data and (b) approach perspectives." + ], + "image_footnote": [], + "bbox": [ + 102, + 260, + 478, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b35deafe8e070892b506f38ac775a00bfe3d22452e3fb9c41cb329ce9c675e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 261, + 890, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 481, + 326, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In poster design, content-aware layout generation is crucial for automatically arranging visual-textual elements on the given image. With limited training data, existing work focused on image-centric enhancement. However, this neglects the diversity of layouts and fails to cope with shape-variant elements or diverse design intents in generalized settings. To this end, we proposed a layout-centric approach that leverages layout knowledge implicit in large language models (LLMs) to create posters for omnifarious purposes, hence the name PosterO. Specifically, it structures layouts from datasets as trees in SVG language by universal shape, design intent vectorization, and hierarchical node representation. Then, it applies LLMs during inference to predict new layout trees by in-context learning with intent-aligned example selection. After layout trees are generated, we can seamlessly realize them into poster designs by editing the chat with LLMs. Extensive experimental results have demonstrated that PosterO can generate visually appealing layouts for given images, achieving new state-of-the-art performance across various benchmarks. To further explore PosterO's abilities under the generalized settings, we built PStylish7, the first dataset with multi-purpose posters and various-shaped elements, further offering a challenging test for advanced research. Code and dataset will be publicly available at https://thekinsley.github.io/PosterO/.", + "bbox": [ + 86, + 503, + 483, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 481, + 643, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Content-aware layout generation makes the poster design process intelligent by automatically arranging visual-textual elements on given images. As the demand for effective design grows rapidly, a mainstream platform surpasses 185 million monthly users [3], and $90\\%$ of survey participants take AI-empowered features as key considerations [2]. This technology is gaining more attention for its practical value in various applications [32, 49, 50, 53], as shown in Fig. 1.", + "bbox": [ + 511, + 502, + 906, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike other generation tasks [22, 43], collecting image-Layout pairs for the target task requires costly annotation and inpainting processes [19, 55]. Therefore, existing work has incorporated object hints [4, 55], prior information [7, 19], retrieved references [17, 31], and augmented data [44] to confront the challenge of training data scarcity. However, because these are mainly image-centric enhancements without increasing the adaptability and diversity of layouts, they are prone to get trapped in the local solution space [18]. On the other hand, while leveraging related knowledge implicit in large language models (LLMs) [11, 37, 42, 47] showed the potential for layout-centric enhancements [12, 46], existing work [31, 44] has suffered from the semantically impoverished layout representations. The two primary issues are the monotonous rectangular shape of elements and the isolated visual constraints, which led to failures at capturing the complexity of real-world layouts and the intricate visual relationship between images and elements.", + "bbox": [ + 509, + 628, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 114, + 887, + 241, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8117", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To pursue data efficiency and semantic richness required for advances in content-aware layout generation, we proposed PosterO to enable in-context learning (ICL) of LLMs by structuring layout tree representations. Specifically, it consists of three procedures, including (a) layout tree construction that jointly represents layout elements and design intents (i.e., available areas on images) as hierarchical nodes in SVG trees [9], (b) layout tree generation that dynamically selects intent-aligned learning examples for the test input to facilitate ICL, and (c) poster design realization that integrates design materials into the generated layout trees to create posters in the subsequent chats with LLMs.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conducted extensive experiments on public benchmarks [19, 55] and demonstrated that PosterO has achieved new state-of-the-art performance. It particularly shows stability in managing domain adaption problems [52] and spatial distribution shifts [17] that make the current approaches stumped. In additional efforts, we enhanced the evaluation process through newly added intent-aware content metrics that are more credible than currently used saliency-aware ones. We also demonstrated the adaptability of PosterO to different design intents and small-scale LLMs [35], highlighting its applicability. To further explore its potential under the generalized settings of diverse poster purposes (e.g., commercial or education) and shape-variant elements (e.g., circles or curves), we built the first dataset, PStylish7, consisting of 152 few-shot learning samples and 100 test images, covering seven representative poster purposes and eight element types. Serving as a challenging test, it is believed to facilitate advanced research in this field.", + "bbox": [ + 89, + 273, + 482, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the contribution of this paper as follows:", + "bbox": [ + 109, + 546, + 478, + 560 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A language-based content-aware layout representation, layout tree, uses hierarchical nodes to model both layout elements and design intents with arbitrary shapes.", + "- A layout-centric approach, PosterO, leverages implicit knowledge of LLMs by applying a few intent-aligned layout trees as examples to perform in-context learning.", + "- Extensive evaluations and newly added metrics verify the superior effectiveness of PosterO in confronting various adaptation problems that plague current approaches.", + "- A dataset for generalized content-aware layout generation, PStylish7, incorporates seven representative poster purposes and eight shape-variant element types." + ], + "bbox": [ + 89, + 561, + 482, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As our best knowledge, this work is the first to consider the generalized settings of content-aware layout generation.", + "bbox": [ + 89, + 744, + 482, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 791, + 232, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Content-aware Layout Generation", + "text_level": 1, + "bbox": [ + 89, + 816, + 390, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Apart from general layout tasks [1, 15, 23, 25, 28], content-aware layout generation is a cross-modal task [10, 16] integrating visual conditions, i.e., image canvases, which are inherent in various real-world applications [32, 49, 50, 53],", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "highlighting its practical utility. With advances in deep generative models, data-centric approaches have become dominant in this field, categorized into GAN-based [18, 19, 52, 55], auto-regression-based [4, 5, 17, 30], and diffusion model-based [7, 27] approaches.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Entering into data construction of image layout pairs, CGL-GAN [55] leveraged an inpainting model [45] to remove design elements from posters and a saliency object detection model [48] to give generative networks additional hints about image compositions. Similar saliency-enhanced paradigm [26, 39, 40] has been adopted by most subsequent approaches. Dealing with limited training data, DS-GAN [19] introduced prior knowledge [14, 29] to arrange layout elements in a meticulously designed order, making the patterns existing in the data more efficiently mined. RALF introduced retrieval augmentation, leveraging a composition-aware similarity evaluation model [13] to find the nearest neighbors for a given image, and incorporating their layout features [23] as additional inputs for the generative network. Moreover, Chai et al. [7] adapted a general layout diffusion model [6] for content-aware tasks by incorporating aesthetic constraints and a saliency-aware layout plausibility ranker.", + "bbox": [ + 511, + 167, + 905, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As current approaches are engaged in image-centric enhancement yet seldom expand the diversity of layouts, the accessible solution space is narrowed, impeding their ability to address spatially varying design intents across different applications. In light of this, we propose a layout-centric, intent-aware approach in this work.", + "bbox": [ + 511, + 426, + 905, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Language Model-based Layout Generation", + "text_level": 1, + "bbox": [ + 511, + 527, + 877, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While numerical layout representations risk losing the semantic information of attributes [46], current advances in large language models (LLMs) [11, 42, 47] open up new possibilities with HTML-based [31, 38], CSS-based [8, 12], and SVG-based [44, 46] language representations as these approaches enable leverage of layout knowledge implicit in LLMs. However, due to the challenge of guiding LLMs to perceive visual content, only a few approaches [31, 44] have supported content-aware layouts.", + "bbox": [ + 511, + 551, + 905, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LayoutPrompter [31] represented each layout element as a
in HTML. While adopting a saliency paradigm, it only extracted the minimum bounding rectangle from the binary map as coarse content constraints and then retrieved layout examples for in-context learning. In contrast, PosterLlama [44] adopted a SVG-based representation [46] using shapes. It then fine-tuned DINOv2 [36, 56] with CodeLlama [20, 42] enhanced by data augmentation [54].", + "bbox": [ + 511, + 688, + 905, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As current approaches employed monotype element representations and isolated visual representations, they failed to exploit the rich semantics of languages nor explicitly capture the relationship between elements and images. In light of this, we propose layout trees to represent various-shaped elements and visual hints jointly.", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8118", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4074032f7c66cdc63cb79a0ec59d251fa818eeb5e0b4a7f022598b817e7e2e78.jpg", + "image_caption": [ + "Figure 2. An overview of PosterO. First, (a) takes image layout pairs $(I,L)$ from datasets as input for data preparation, jointly modeling various-shaped elements $E$ and design intents $D$ towards layout trees $T$ and building up the latent space of $D$ . Then, (b) takes test images $I_{t}$ as input and searches examples based on the predicted design intents $f_{t}$ to apply LLM $\\mathcal{M}$ through in-context learning. After obtaining generated layout trees $\\hat{T}$ , (c) can continue the conversation with $\\mathcal{M}$ to create poster designs seamlessly." + ], + "image_footnote": [], + "bbox": [ + 91, + 92, + 459, + 299 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ac9e0efde70344e1d54552de7f293726700faa26298f1236c30c677955e5e48d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 92, + 764, + 300 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa93599ed41f955c2539228a3dcb175333db26475495f61243acb43b2e9672e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 93, + 901, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Approach: PosterO", + "text_level": 1, + "bbox": [ + 89, + 375, + 398, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the limitations of image-centric and language model-based approaches, we propose PosterO. It is layout-centric and enables large language models (LLMs) to generate visually appealing layouts for given images by structurally representing the content-aware layouts as trees $T$ in SVG language. An overview of PosterO is shown in Fig. 2. In brief, it is composed of three procedures, namely, (a) layout tree construction for data preparation, (b) layout tree generation applying an LLM $\\mathcal{M}$ through $k$ -shot in-context learning, and (c) poster design realization through the subsequent chat with $\\mathcal{M}$ .", + "bbox": [ + 89, + 398, + 483, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Layout Tree Construction", + "text_level": 1, + "bbox": [ + 89, + 573, + 326, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Universal Shape Vectorization", + "text_level": 1, + "bbox": [ + 89, + 595, + 346, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In content-aware layout generation, each data pair contains an image $I$ and a layout $L$ . Usually, $L$ is represented as a set of numerical tuples $\\{e_i\\}_{i}^n = \\{(c_i, b_i)\\}_{i}^n$ , where $c_i$ and $b_i = (x, y, w, h)$ indicate the category and bounding box of the $i$ -th element $e_i$ out of all $n$ elements. As this is adequate only for rectangular elements, a more generalized representation becomes essential for the diverse shapes encountered in real-world posters, such as circles and curves. In light of this, SVG (Scalable Vector Graphics) emerges as an ideal choice because of its ability to represent arbitrary shapes and, more importantly, its nature as a language to facilitate the layout knowledge implicit in LLMs.", + "bbox": [ + 89, + 613, + 483, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the published SVG standards [9], we define five shapes to encompass a wide range of elements typically found in posters, as shown in Fig. 3. First, (a) regular rectangle is vectorized using $\\langle \\text{rect} \\rangle$ . Also, its two variants, including (b) vertical rectangle sensitive to the aspect ratio and (c) rotated rectangle using the rotate function in the transform attribute to indicate the rotation degree.", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8f213634ecc780f05c86c798374b0ab2428589bc8dd2eb5cc551cee6a7b356cc.jpg", + "image_caption": [ + "(a) General rectangle" + ], + "image_footnote": [], + "bbox": [ + 516, + 371, + 637, + 460 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c413ee1a7ab9f381a74b21df64c809230c402e0b05fda8f0b891e7e0b433867f.jpg", + "image_caption": [ + "(b) Vertical rectangle" + ], + "image_footnote": [], + "bbox": [ + 648, + 371, + 771, + 460 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/33abfb8ceac90a7d23ab9b60c19ed372a1343ea3e2b9cdf8d8b64a75566a21f0.jpg", + "image_caption": [ + "(c) Rotated rectangle" + ], + "image_footnote": [], + "bbox": [ + 784, + 371, + 903, + 460 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4c0796c747d085794b6628493d16236b254d117e40b0ca532761ebb144d50ea9.jpg", + "image_caption": [ + "(d) Ellipse" + ], + "image_footnote": [], + "bbox": [ + 517, + 474, + 629, + 558 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a9cf75fab83f5a938aed6f09c77c20083ae463ec281d4b15b949e705faf7574a.jpg", + "image_caption": [ + "(e) Complex curve" + ], + "image_footnote": [], + "bbox": [ + 648, + 474, + 769, + 558 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8929c0e5cea4c1dc80a98fe81e10dd56c53a870ae6a2042e82e622f4e52adc76.jpg", + "image_caption": [ + "(f) Polygon", + "Figure 3. Various shapes of elements in PStylish7 dataset." + ], + "image_footnote": [], + "bbox": [ + 782, + 474, + 903, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When delving into more intricate shapes, (d) ellipse is vectorized using , which can also handle the circle as a special case, and (e) complex curve is vectorized using approximated by multiple cubic Bezier curves.", + "bbox": [ + 511, + 599, + 906, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. Design Intent Vectorization", + "text_level": 1, + "bbox": [ + 511, + 670, + 750, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to visual-textual elements, the design intents associated with the layout $L$ must also be vectorized. Specifically, design intents $D$ are embodied by a set of areas within the image $I$ , suitable or intended for placing the elements. By incorporating these areas' representations $N_{D}$ , the layout tree becomes content-aware and capable of conveying visual relationships between elements and images to LLMs, which cannot 'see' the images. As $D$ can be learned from existing data or customized case by case, this also provides a flexible interface for indicating various purposes.", + "bbox": [ + 511, + 688, + 906, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By leveraging image layout pairs from posters with a similar purpose, a design intent detection model $S$ based on U-Net [41] is trained in a semi-supervised manner. Referring to [18], $S$ takes the bitmap of layout elements as the", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8119", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "supervision signal during the optimization, while the consistency loss is canceled to clearly segment out the areas. With the pre-trained $S$ , PosterO can obtain design intents $D$ aligning the target purpose for any image $I$ . Specifically, after thresholding and morphological operations, we approximate the contour of the predicted intent maps $m$ and vectorize the obtained irregular areas using in SVG language. An example is shown in Fig. 3(f). Besides, the intermediate embeddings $f = S_{\\mathfrak{e}}(I)$ , where $S_{\\mathfrak{e}}$ denotes the last block of the encoder, are also collected to serve as their representations in the latent space of design intents $\\mathbf{Z}_D$ .", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3. Hierarchical Node Representation", + "text_level": 1, + "bbox": [ + 89, + 263, + 375, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the nodes of layout elements $N_{E}$ and design intents $N_{D}$ are in place, the layout tree $T$ is constructed and encapsulated within an SVG container aware of the image resolution $(w_{I}, h_{I})$ , as: $<\\mathrm{svg~width}=\\{\"w_{I}\\}\" \\mathrm{height}=\\ {\"h_{I}\\}\" \\mathrm{xmlns}=\\text{\"http://www.w3.org/2000/sv\"}> \\{N_{D}\\} \\{N_{E}\\}<\\mathrm{svg}>$ . $N_{D}$ is put ahead to ensure that $N_{E}$ is arranged considering their probabilistic dependencies in the autoregression, thus generating suitable content-aware layouts. In addition, as the special elements that wrap around others, such as underlays, are usually used in poster design, a novel hierarchical representation explicitly modeling these enveloping properties is proposed. Specifically, $N_{E}$ is sorted by prioritizing underlays with larger areas. Then, given a node $N_{a}$ and the corresponding nodes $N_{B} = \\{N_{b} \\mid N_{b} \\in \\{N_{E}\\}_{a + 1}^{n} \\land C(N_{a}, N_{b})\\}$ , where the condition function $C(\\cdot)$ is defined as:", + "bbox": [ + 89, + 282, + 486, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} C (\\cdot) = \\left(| N _ {a}. x - N _ {b}. x | \\leq \\epsilon\\right) \\wedge \\left(| N _ {a}. y - N _ {b}. y | \\leq \\epsilon\\right) \\\\ \\wedge \\left(\\left| N _ {a}. x + N _ {a}. w - N _ {b}. x - N _ {b}. w \\right| \\leq \\epsilon\\right) \\tag {1} \\\\ \\wedge \\left(\\left| N _ {a}. y + N _ {a}. h - N _ {b}. y - N _ {b}. h \\right| \\leq \\epsilon\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 515, + 480, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a subtree is constructed as $<\\text{svg} \\times=\" \\{N_{a}.x\\}\" \\text{y=\"} \\{N_{a}.y\\}\">$ $\\{N_{a}'\\} \\{N_{B}'\\} < / \\text{svg}>$ and inserted at the position of $N_{a}$ , where the processed $N_{i}'$ has the relative geometry attributes $(N_{i}'.x, N_{i}'.y) = (N_{i}.x - N_{a}.x, N_{i}.y - N_{a}.y)$ . All leaf nodes of $T$ are assigned unique identifiers $\\mathfrak{id} = \"\\{c_{i} - i\\}\"$ to facilitate precise manipulation and querying.", + "bbox": [ + 89, + 577, + 483, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Layout Tree Generation", + "text_level": 1, + "bbox": [ + 89, + 676, + 312, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Intent-aligned Example Selection", + "text_level": 1, + "bbox": [ + 89, + 698, + 367, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Upon findings from [34], LLMs are capable of in-context learning (ICL), requiring only a few examples to perform downstream tasks. This significantly reduces the need for task-specific data and avoids the extensive costs or catastrophic forgetting [24, 33] associated with fine-tuning. To make the best use of layout knowledge implicit in LLMs during ICL, selecting examples that align with the design intent is crucial. Therefore, given the test image $I_{t}$ , its embedding $f_{t}$ is first extracted by the detection model $\\mathcal{S}$ ; subsequently, $k$ layout trees $\\{T_j\\}_j^k$ are selected as in-context examples, where their embeddings $\\{f_j\\}_j^k$ are the $k$ closest to $f_{t}$ in the latent space $\\mathbf{Z}_{D}$ .", + "bbox": [ + 89, + 717, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. In-context Learning for Layout Generation", + "text_level": 1, + "bbox": [ + 511, + 90, + 861, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To provide clear structure and objectives in the prompt $P$ for the LLM $\\mathcal{M}$ , each intent-aligned example $T_{j}$ is preceded by a short sentence that informs its number $j$ , the resolution of its corresponding image, design intents $D$ , and identifiers of its nodes, as shown in Fig. 2(b). The example field is then encapsulated within a preface and postscript, with the postscript delivering the request message of the test image $I_{t}$ through a short sentence mirroring that of the example. Complete prompts $P$ are presented in supplementary material. Finally, by inputting $P$ into $\\mathcal{M}$ and interpreting the responses, PosterO completes the layout generation process.", + "bbox": [ + 511, + 109, + 906, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Poster Design Realization", + "text_level": 1, + "bbox": [ + 511, + 286, + 746, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Benefiting from the essence of the generated layout tree $\\hat{T}$ as graphics and the rigid id assignment for elements, guiding $\\mathcal{M}$ to transform $\\hat{T}$ into a practical poster design is feasible and straightforward. Specifically, the transformation contains mockup creation, where shape elements are converted to or regarding their categories, and material synthesis, where elements are filled with actual design contents, e.g., slogans and hyperlinks to images. Notably, this transformation is achieved with zero-shot learning, yet $\\mathcal{M}$ brings impressive results (in supplementary material) and even predicts suitable font sizes for texts, highlighting the value of design knowledge implicit in LLMs.", + "bbox": [ + 511, + 306, + 906, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 503, + 645, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 529, + 614, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To evaluate the proposed PosterO, two public e-commerce poster datasets are used, while their train+valid/annotated test/unannotated test splits are allocated as in [17] to ensure a fair comparison with existing work. Specifically, PKU PosterLayout [19] comprises 8,734/1,000/905 images with three layout element types, including text, logo, and underlay; CGL [55] comprises 54,546/6,002/1,000 images with four layout element types, including text, logo, underlay, and embellishment. Furthermore, to explore PosterO's abilities under the generalized settings, we built the first multipurpose poster dataset, PStylish7, containing layout elements of various shapes. More details about PStylish7 are depicted in Sec. 4.6 and supplementary material.", + "bbox": [ + 511, + 551, + 906, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 757, + 697, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following previous work [17, 19, 55], we use graphic and content metrics. Graphic metrics include overlay $Ove\\downarrow$ alignment $Ali\\downarrow$ , and loose, strict underlay effectiveness $Und_{l}\\uparrow ,Und_{s}\\uparrow ;$ Content metrics include non-salient space utilization $Uti\\uparrow$ , salient space occlusion $Occ\\downarrow$ , and readability $Rea\\downarrow$ . Moreover, we newly proposed two content metrics, design intent coverage $Cov\\uparrow$ and conflict $Con\\downarrow$ based on a design intent detection model $S_{pc}$ that is compre", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8120", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d38091159da8a2c3b94d72ed52200cc5f0bd33325a0be2440f4a850da0e4951d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOve ↓Ali ↓Undl ↑Unds ↑Int ↓Sal ↓Rea ↓Avg ↓
Annotated Test Split
Real data0.00100.00380.99420.99030.02890.03680.01090.0138
CGL-GAN0.09660.00350.78540.35700.12040.27740.01910.1964
DS-GAN0.02610.00380.83500.58040.06770.31000.01990.1446
ICVT0.25720.04050.53840.39320.78311.08360.02590.4655
LayoutDM†0.15620.00180.64260.38730.50660.24290.01850.2709
RALF0.00840.00280.98080.92010.11620.04790.01280.0410
Tuning-required
PosterLlama‡0.00080.00060.99990.99820.34070.26190.01770.0891
LLM-based
Tuning-free
LayoutPrompterL20.00830.00300.37160.16330.32000.94330.03190.3959
LayoutPrompterCL0.00400.00300.44150.19370.36620.90350.03010.3817
LayoutPrompterL30.00170.00280.40850.16130.51990.85670.03090.4060
PosterOL2 (Ours)0.00040.00270.75970.56050.36790.14910.01880.1741
PosterOLCL (Ours)0.00040.00280.96210.86060.23970.22950.01890.0955
PosterOL3 (Ours)0.00060.00250.98880.92110.10800.11100.01710.0470
Unannoted Test Split
Non-LLMTrain data0.00120.00270.99690.99000.00000.00000.0108
CGL-GAN0.10100.00480.73260.27430.22750.77550.0327
DS-GAN0.02480.00460.78590.46760.16590.58680.0320
ICVT0.27860.04800.49390.35490.79471.32890.0347
LayoutDM†0.16380.00290.59870.36950.53570.31720.0264
RALF0.01750.00690.95480.86530.59860.17790.0231
Tuning-required
PosterLlama‡0.00060.00060.99860.99170.26670.38230.02850.0983
LLM-based
Tuning-free
LayoutPrompterL20.00950.00320.34010.11610.53491.20500.04080.4767
LayoutPrompterCL0.00320.00280.40520.14640.51481.11060.03950.4456
LayoutPrompterL30.00100.00260.40540.16210.70051.08380.04120.4660
PosterOL2 (Ours)0.00040.00310.72840.51880.40300.33510.02650.2173
PosterOLCL (Ours)0.00050.00280.95950.84510.27960.39450.02730.1286
PosterOL3 (Ours)0.00040.00250.98560.92410.14270.21310.02480.0677
", + "bbox": [ + 91, + 88, + 648, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "hensively pre-trained on [19] and [55]. Calculating $Cov\\uparrow$ and $Con\\downarrow$ is similar to $Uti\\uparrow$ and $Occ\\downarrow$ but using the predicted intent maps. As reported in Tab. 2, intent maps have doubled the matched portion with real layout data, reduced the unmatched portion by $34.7\\%$ , and required only $1.3\\%$ computational costs. These findings have demonstrated that intent-aware metrics are more credible and efficient. For a holistic understanding, we standardize intent-aware and saliency-aware metrics to $Int\\downarrow$ and $Sal\\downarrow$ , as follows:", + "bbox": [ + 89, + 547, + 482, + 683 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI n t = \\Sigma (\\frac {\\left| C o v _ {\\hat {L}} - C o v _ {L} \\right|}{1 - C o v _ {L}}, \\frac {\\left| C o n _ {\\hat {L}} - C o n _ {L} \\right|}{C o n _ {L}}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 694, + 482, + 734 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS a l = \\Sigma (\\frac {| U t i _ {\\hat {L}} - U t i _ {L} |}{1 - U t i _ {L}}, \\frac {| O c c _ {\\hat {L}} - O c c _ {L} |}{O c c _ {L}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 729, + 428, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L$ and $\\hat{L}$ are the layouts from the train split and generation. Finally, we introduce an average metric $Avg\\downarrow = \\Sigma (Ove,Ali,1 - Und_{l},1 - Und_{s},Int,Sal,Rea)$ .", + "bbox": [ + 89, + 775, + 482, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 832, + 307, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure independent evaluations on [19] and [55], we correspondingly train two detection models, $S_{p}$ and $S_{c}$ , with mean squared error loss and a learning rate of $1e^{-6}$ . Each", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/da94c947185bbe62879a25a49fd895c465362ebb3ecdadd71315ebb59ea20ca9.jpg", + "image_caption": [ + "Figure 4. Comparisons of visualized results on the unannotated test split of PKU PosterLayout dataset." + ], + "image_footnote": [], + "bbox": [ + 666, + 89, + 898, + 492 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3959275293406a32bd9b97ec4cef394fda7b05b4ebb2908a477ed13a8054e097.jpg", + "table_caption": [ + "Table 1. Comparisons of quantitative results on PKU PosterLayout dataset." + ], + "table_footnote": [], + "table_body": "
PredictionPKUCGL
Match ↑Unmatch ↓Match ↑Unmatch ↓
Saliency map0.22380.11930.19840.1352
Design intent map0.44140.09960.40360.0883
Inference costNetwork architecture#FLOPs(G)#Params(M)
Saliency mapISNet + BASNet287.83131.11
Design intent mapUNet (MiT-B1)3.7416.43
", + "bbox": [ + 514, + 544, + 903, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Comparisons of the predicted saliency and design intent maps on the annotated test splits of PKU PosterLayout and CGL datasets, showing that intent-aware metrics are more credible.", + "bbox": [ + 511, + 669, + 906, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "detection model is with a Transformer encoder tailored for semantic segmentation, MiT-B1 [51]. For in-context learning (ICL), we implement $\\mathcal{M}$ with three open-source LLMs, including Llama 2-7B (L2) [47], CodeLlama-7B (CL) [42], and Llama 3.1-8B (L3) [11], with an unified sampling temperature of 0.7. Considering the context lengths and remaining consistent with LayoutPrompter [31], the example size $k$ is 5, 10, and 10, respectively. Still adhering to [31], a layout ranker is employed, while mIoU in the quality function measures the similarity between generated layouts and predicted design intents rather than real data, minimizing the risk of data leakage.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8121", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e0e104a76d75c46ca0a1c05b37baf03e2e1c45520053b371fa49d5be90543dc5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOve ↓Ali ↓Undl ↑Unds ↑Int ↓Sal ↓Rea ↓Avg ↓
Annotated Test Split
Non-LLMReal data0.00030.00240.99630.98800.00880.00970.01190.0070
CGL-GAN0.22910.01230.64660.22810.70880.46270.02130.3656
DS-GAN0.04600.00220.90810.63080.39500.15390.01810.1537
ICVT0.24530.01790.51500.33260.36320.51390.02110.3305
LayoutDM†0.01840.00210.92160.81590.04110.02860.01370.0523
RALF0.00420.00240.99120.97560.03660.07240.01800.0238
Tuning-required PosterLlama‡0.00060.00060.99870.98900.13370.25590.01840.0602
LLM-basedTuning-free
LayoutPrompterL20.00870.00330.33690.14550.33620.82340.03270.3888
LayoutPrompterCL0.00370.00310.39650.16160.37180.77990.03170.3760
LayoutPrompterL30.00170.00300.38300.17400.55930.79570.03270.4051
PosterOL2 (Ours)0.00060.00230.78010.61500.25420.04670.01760.1323
PosterOLCL (Ours)0.00050.00230.96820.88680.12460.07000.01680.0513
PosterOL3 (Ours)0.00060.00220.98420.93400.03880.00260.01610.0203
Unannotated Test Split
Non-LLMTrain data0.00030.00270.99450.98580.00000.00000.01180.0049
CGL-GAN0.26680.03160.67740.16560.97072.39730.05120.6964
DS-GAN0.09910.01380.75660.28100.46072.27320.05260.5517
ICVT0.20450.10100.43570.25991.28132.65120.03970.7974
LayoutDM†0.07930.18220.63040.38531.50023.29060.06120.8711
RALF0.02730.01890.97560.93151.02371.67980.03970.4118
Tuning-required PosterLlama‡0.00140.00070.99710.97710.35102.61700.05550.4359
LLM-basedTuning-free
LayoutPrompterL20.00690.01030.28470.11261.02572.37020.06280.7255
LayoutPrompterCL0.00450.00240.32390.13411.58002.53800.06500.8188
LayoutPrompterL30.00260.00160.26930.11421.62492.41280.06440.8175
PosterOL2 (Ours)0.00050.00330.66630.47700.35221.23650.03820.3553
PosterOLCL (Ours)0.00030.00370.93690.80710.27711.32230.03790.2710
PosterOL3 (Ours)0.00040.00320.98160.90340.43281.24070.03650.2612
", + "bbox": [ + 91, + 88, + 648, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparisons of quantitative results on CGL dataset.", + "bbox": [ + 194, + 530, + 558, + 544 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0843d030f2c5bb261d80b85a93409ecdfc2aefd61ddae899d084d6fe9c13bf93.jpg", + "image_caption": [ + "Figure 5. Comparisons of visualized results on the unannotated test split on CGL dataset." + ], + "image_footnote": [], + "bbox": [ + 666, + 89, + 901, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 89, + 555, + 390, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We select approaches with open-sourced implementation as baselines, including GAN-based CGL-GAN [55], DS-GAN [19], auto-regression-based ICVT [4], RALF [17], diffusion model-based LayoutDM† [21], and also LLM-based PosterLlama‡ [44], LayoutPrompter [31].", + "bbox": [ + 89, + 577, + 482, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline comparison. Tab. 1 reports the quantitative results on PKU PosterLayout. The proposed PosterO shows the best overall performance across the test splits. Particularly, when confronting the challenging unannotated split that poses domain adaptation problems [52], it still achieves new state-of-the-art (SOTA) results with outstanding stability. Compared to the current non-LLM and LLM-based SOTA approaches, i.e., RALF and PosterLlama, ours significantly improves the average metric $Avg\\downarrow$ by $52.8\\%$ and $31.1\\%$ . Consistent results are obtained on CGL, as reported in Tab. 3, ours improves $Avg\\downarrow$ by $36.6\\%$ and $40.1\\%$ .", + "bbox": [ + 89, + 655, + 482, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "More precisely, PosterO shows impressive graphic metrics (2nd column), closely resembling the distribution of real data, with only a slight deficit in the Ali $\\downarrow$ metric on the", + "bbox": [ + 89, + 821, + 482, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "unannotated split of CGL compared to LayoutPrompter. As for content metrics (3rdcolumn), although PosterO uses only a small design intent detection model $S$ , it achieves comparable or even superior results to the baselines that utilize larger saliency detection models [26, 40] or costly fine-tuned DINOv2 [36]. Especially, PosterO is best at dealing with the spatial shifts in subject distributions [17] present in the unannotated split of CGL, maintaining an overwhelming lead in content metrics.", + "bbox": [ + 511, + 556, + 906, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Backbone LLMs $\\mathcal{M}$ . Comparing the performance of tuning-free approaches across various LLMs is crucial for realizing how effectively the layout knowledge implicit in $\\mathcal{M}$ has been leveraged. As reported in Tab. 1 and Tab. 3, PosterO demonstrates overwhelming superiority over LayoutPrompter in all counterparts with the same $\\mathcal{M}$ . More importantly, from Poster $_{L2,CL}$ , to Poster $_{L3}$ , it is evident that the proposed approach has experienced improvements as the capabilities of its based $\\mathcal{M}$ increased. This finding indicates that new knowledge, such as programming, introduced during model updates, has indeed been leveraged to generate layouts. With the continuous development of LLMs, PosterO is anticipated to keep its rising potential.", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "The extended version presented in [17].", + "bbox": [ + 112, + 875, + 331, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ With the released weights tuned on [19], [55], and synthesized data.", + "bbox": [ + 112, + 887, + 478, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8122", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff6f2d96fedffd543d6edde4a787b275c821615ed260be0315d68a94c9558f65.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method \\ k1257810
LayoutPrompter0.77110.66900.53530.48830.49190.4660
PosterO (Ours)0.06780.07460.07320.06630.06510.0677
", + "bbox": [ + 94, + 88, + 478, + 135 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8c86786cb1843b018447bc051224241845c9ffa14fb8b599b5f6591da0101fa.jpg", + "image_caption": [ + "Figure 6. Content metrics over in-context example size $k$ ." + ], + "image_footnote": [], + "bbox": [ + 98, + 162, + 464, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualized results. Fig. 4 and Fig. 5 show the layouts generated by different approaches. The results illustrate that PosterO generates high-quality layouts that the distribution of elements is generous and coordinated. In contrast, the approaches based on training or tuning tend to place elements centrally and thus fail to confront images with compositions that are less common in the training data. More visualized results are presented in supplementary material.", + "bbox": [ + 89, + 363, + 482, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In-context example size $k$ . To explore the impact of example selection during ICL, we test a set of candidates $\\{k\\} = \\{1,2,5,7,8,10\\}$ on the unannotated split of [19], as reported in Tab. 4. Amazingly, PosterO consistently outperforms LayoutPrompter and all other baselines in Tab. 1 across all settings of $k$ . As the optimal performance occurs at $k = 7$ and 8, it is proved that $\\mathcal{M}$ has truly benefited from learning with the selected intent-aligned examples. To further observe how well these examples help $\\mathcal{M}$ understand visual contents, which is the focal point of content-aware layout generation, we plot the resultant content metrics over $k$ in Fig. 6. It is noteworthy that directly adopting the most intent-aligned example (Top-1 Retrieval) surpasses all baselines, indicating a successful alignment between intent and layout in the latent space $\\mathbf{Z}_D$ . Nevertheless, we still observe improvements when involving in-context examples.", + "bbox": [ + 89, + 487, + 482, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 739, + 243, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate our design choice in layout tree construction (Sec. 3.1) and learning example selection (Sec. 3.2), as well as to explore PosterO's adaptability to customized design intents and small-scale LLMs, we conducted the following ablation studies, mainly on the unannotated split of [19].", + "bbox": [ + 89, + 761, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Design choices for layout tree construction. We investigate the effectiveness of the design intent vectorization $(d)$ and hierarchical node representation $(h)$ during constructing layout trees, as reported in Tab. 5. Due to space constraints,", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ebb3cf575df61de4bc99366e30c831c42c8207109590fe6ef04cc4967ef48f14.jpg", + "table_caption": [ + "Table 4. Comparisons of $Avg\\downarrow$ over in-context example size $k$ ." + ], + "table_footnote": [], + "table_body": "
dhUndl↑Unds↑Int↓Sal↓Rea↓Avg↓
✓✓0.98560.92410.14270.21310.02480.0677
✓0.86090.66500.09810.15600.02400.1078
✓0.94970.88360.98580.39910.02740.2260
0.76280.64130.81620.27290.02460.2447
", + "bbox": [ + 514, + 88, + 903, + 170 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/eabbab6e171e75a09bac78f8ddf5f45bb2edf1bca547d4260f427c4b5dc6356b.jpg", + "table_caption": [ + "Table 5. Ablation study on the $(d)$ design intent vectorization and $(h)$ hierarchical node representation in layout tree construction." + ], + "table_footnote": [], + "table_body": "
SelectionUndl↑Unds↑Int↓Sal↓Rea↓Avg↓
f-aligned0.98560.92410.14270.21310.02480.0677
D-aligned0.98660.92970.22210.25310.02360.0836
E-aligned0.97430.90260.34310.41770.02640.1304
Random0.98850.93020.32250.46880.02950.1293
", + "bbox": [ + 514, + 212, + 901, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation study on different references for in-context example selection. (f: Embeddings of design intents, $D$ : Bounding boxes of design intents, $E$ : Sets of element types.)", + "bbox": [ + 511, + 296, + 903, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$Ove\\downarrow$ and $Ali\\downarrow$ are omitted, considering their slight fluctuations. As observed, the combination of $d$ and $h$ yields the best results. Although removing $h$ appears to improve content metrics, this is primarily because the invalid empty underlays artificially inflate the utilization rate, as analyzed in supplementary material. The significant decreases witnessed in underlay effectiveness, $Und_{l}$ , $Und_{s}$ , by up to $28\\%$ , further underscore this issue. In contrast, removing $d$ from both the examples and postscript makes LLMs blindly mimic the provided instances. Here, however, shows a $Sal\\downarrow$ surpasses some baselines, such as CGL-GAN, DS-GAN, and LayoutPrompter, while $Int\\downarrow$ fails far behind, reflecting the robustness of the newly proposed metric $Int\\downarrow$ . If both $d$ and $h$ are removed, the most severe degradation is observed, up to $261\\%$ . In this way, we demonstrate the necessity of these design choices.", + "bbox": [ + 511, + 347, + 906, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Design choices for in-context example selection. We have seen the remarkable performance of design intent embeddings $f$ -aligned example selection, which relies on the learned latent space $\\mathbf{Z}_D$ . To investigate alternative references, we have considered bounding boxes of design intents ( $D$ -aligned), sets of element types ( $E$ -aligned), and even with no specific alignment (Random). As reported in Tab. 6, although all $D$ -aligned, $E$ -aligned, and random selections make declines in average metrics, the outcomes are still superior to most baseline approaches. In particular, using $D$ -aligned selection also achieves a new SOTA result, which supports its eligibility as an alternative enabling PosterO to deliver acceptable content-aware layouts when faced with roughly sketched $D$ from users.", + "bbox": [ + 511, + 594, + 908, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaption to customized design intent. Upon the discussions from the previous paragraph, we take a further step to verify PosterO's adaptability to customized design intents. Fig. 7 demonstrates the results of using the original $D$ predicted by the detection model $S$ and three customized $D$ with different levels of complexity. As the input $D$ varies,", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8123", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d451a8f3afa3c1033b231830c533ecafbe61454190ae7c2cd9faf8656f722e51.jpg", + "image_caption": [ + "Figure 7. PosterO's adaptability to different design intents $D$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 477, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da9cf6671ae11f0d46753805b19dbb4605b4cfa06335b3229ca63045e6131088.jpg", + "image_caption": [ + "Table 7. Ablation study on $\\mathcal{M}$ accommodating small-scale LLMs." + ], + "image_footnote": [], + "bbox": [ + 93, + 253, + 480, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PosterO dynamically adjusts the spatial distribution of elements in the generated layouts while maintaining remarkable graphic metrics. Notably, this capability is unique to PosterO, attributed to its intent-based paradigm.", + "bbox": [ + 89, + 513, + 482, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaption to small-scale backbone LLM $\\mathcal{M}$ . To address resource-limited environments and to accelerate the layout generation process, we investigate PosterO's adaptability to small-scale LLMs [35], including Llama3.2-1B (L3.2-1B) and Llama3.2-3B (L3.2-3B). As reported in Tab. 7, PosterO maintains very competitive and or even better results with the smaller $\\mathcal{M}$ , showcasing its potential for more flexible and efficient applications.", + "bbox": [ + 89, + 579, + 483, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Generalized Content-aware Layout Generation", + "text_level": 1, + "bbox": [ + 89, + 710, + 482, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify PosterO's potential in generalized content-aware layout generation, which is presented for the first time in this work, we built a new dataset, PStylish7. It comprises 152 few-shot learning samples and 100 test images, covering seven poster categories. Each serves as a representative prototype for a segment in the spectrum of poster design purposes, encompassing artwork exhibition (Paint), cultural education (Poem), public safety (Metro), entertainment marketing (Movie), merchandising display (Menu), public advocacy (Animal), and social-media interaction (Instagram). Besides, it covers eight element types, including", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3db045aaa7fa609e61be95283a46d273c2bae66fb24d90f812290adffc3be709.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OursCLPaintPoemMetroMovieMenuAnimalInsta.
Ove ↓0.00470.00340.06180.01060.00830.00320.0010
Int ↓0.76900.21980.69870.12160.23590.30200.2674
Sal ↓3.88380.84841.36040.32750.71912.62800.6652
Avg ↓1.55250.35720.70700.15320.32110.97770.3112
OursL3PaintPoemMetroMovieMenuAnimalInsta.
Ove ↓0.00320.00280.07770.00040.00560.14950.0004
Int ↓0.62190.62740.74780.33750.16200.53831.1431
Sal ↓2.45120.16001.11110.13470.57902.92841.0810
Avg ↓1.02540.26340.64550.15750.24891.20540.7415
", + "bbox": [ + 516, + 88, + 903, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Quantitative results on PStylish7 dataset.", + "bbox": [ + 560, + 239, + 856, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d3404882d1ea2e5be60dd528fbd27c8d741a0bf8c203d39b15cfd1d0b60e398.jpg", + "image_caption": [ + "Figure 8. Visualized results on PStylish7 dataset." + ], + "image_footnote": [], + "bbox": [ + 516, + 261, + 903, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "all those in the existing datasets and four text variants depicted in Sec. 3.1. Considering feasibility, we employ five metrics, including $Ove\\downarrow$ , $(Cov\\uparrow, Con\\downarrow)$ , standardized as $Int\\downarrow$ , and $(Uti\\uparrow, Occ\\downarrow)$ , standardized as $Sal\\downarrow$ . More details about PStylish7 are given in supplementary material.", + "bbox": [ + 511, + 347, + 903, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The quantitative results are reported in Tab. 8, and the visualized results are presented in Fig. 8. Here, the detection model $S_{pc}$ is employed to obtain design intents $D$ and their features $f$ . While PosterO $_{L3}$ outperforms PosterO $_{CL}$ in most categories, as the first attempt in this field, there is still plenty of room for improvement. Overall, the proposed PStylish7 is significantly more challenging than the existing datasets, attributed to its broader and sparser data distribution. This requires approaches with good generalization capabilities, mirroring real-world scenarios, and can stimulate more advanced work in the future.", + "bbox": [ + 511, + 422, + 905, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 603, + 761, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work presented a tuning-free language model-based approach for content-aware layout generation, named PosterO. Through extensive experiments, we demonstrated that PosterO achieves new state-of-the-art results on various benchmarks, mainly attributed to the structural representation, layout trees. Moreover, we delved into the task under generalized settings and built the first available dataset, PStylish7, which covers seven poster purposes and eight element types. Employing PStylish7, we verified PosterO's potential to solve complicated tasks in real-world scenarios, underscoring its practicality. The future work mainly lies in two aspects. First, enhancing the generalized capabilities of layout generation approaches is highly valuable, as PosterO aims. With the challenging test offered by the proposed PStylish7 dataset, it is increasingly attainable. Second, integrating feedback mechanisms is envisioned to facilitate self-optimization and allow the incorporation of user suggestions, forming a more interactive, user-centric process.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8124", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (61925201, 62132001, 62432001) and Beijing Natural Science Foundation (L247006).", + "bbox": [ + 91, + 114, + 483, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Diego Martin Arroyo, Janis Postels, and Federico Tombari. Variational transformer networks for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13642-13652, 2021. 2", + "[2] Canva. Visual economy report 2024 edition: How visual communication fuels new business opportunities. https://www.canva.com/visual-economy-report/, 2024.1", + "[3] Canva. 10 highlights from canva create 2024: Redesigned for work. https://www.canva.com/newsroom/news/what-happened-at-canva-create-2024/, 2024. 1", + "[4] Yunning Cao, Ye Ma, Min Zhou, Chuanbin Liu, Hongtao Xie, Tiezheng Ge, and Yuning Jiang. Geometry aligned variational transformer for image-conditioned layout generation. In Proceedings of the ACM International Conference on Multimedia, pages 1561-1571, 2022. 1, 2, 6", + "[5] Yunning Cao, Chuanbin Liu, Ye Ma, Min Zhou, Tiezheng Ge, Yuning Jiang, and Hongtao Xie. Self-refined variational transformer for image-conditioned layout generation. International Journal of Machine Learning and Cybernetics, pages 1-18, 2024. 2", + "[6] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18349-18358, 2023. 2", + "[7] Shang Chai, Liansheng Zhuang, Fengying Yan, and Zihan Zhou. Two-stage content-aware layout generation for poster designs. In Proceedings of the ACM International Conference on Multimedia, pages 8415-8423, 2023. 1, 2", + "[8] Jian Chen, Ruiyi Zhang, Yufan Zhou, Jennifer Healey, Jiuxiang Gu, Zhiqiang Xu, and Changyou Chen. Textlap: Customizing language models for text-to layout planning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing: Findings, 2024. 2", + "[9] Erik Dahlstrom, Anthony Grasso, Niklas Hagelroth, Chris Lilley, Cameron McCormack, and Doug Schepers. Scalable vector graphics (svg) 2.0 specification. SVG working group, w3c. https://dev.w3.org/SVG/profiles/2.0/publish/, 2002. 2, 3", + "[10] Leichao Dai, Lin Feng, Xinglin Shang, and Han Su. Cross modal adaptive few-shot learning based on task dependence. Chinese Journal of Electronics, 32(1):85-96, 2023. 2", + "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 2, 5", + "[12] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and" + ], + "bbox": [ + 93, + 200, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Proceedings of the International Conference on Neural Information Processing Systems, 36, 2024. 1, 2", + "[13] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. In Proceedings of the International Conference on Neural Information Processing Systems, pages 50742-50768, 2024. 2", + "[14] Shunan Guo, Zhuochen Jin, Fuling Sun, Jingwen Li, Zhaorui Li, Yang Shi, and Nan Cao. Vinci: an intelligent graphic design system for generating advertising posters. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1-17, 2021. 2", + "[15] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004–1014, 2021. 2", + "[16] Sun Haoran, Wang Yang, Liu Haipeng, and Qian Biao. Fine-grained cross-modal fusion based refinement for text-to-image synthesis. Chinese Journal of Electronics, 32(6): 1329-1340, 2023. 2", + "[17] Daichi Horita, Naoto Inoue, Kotaro Kikuchi, Kota Yamaguchi, and Kiyoharu Aizawa. Retrieval-augmented layout transformer for content-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 67-76, 2024. 1, 2, 4, 6", + "[18] HsiaoYuan Hsu, Xiangteng He, and Yuxin Peng. Densitylayout: Density-conditioned layout gan for visual-textual presentation designs. In Proceedings of the International Conference on Image and Graphics, pages 187-199, 2023. 1, 2, 3", + "[19] HsiaoYuan Hsu, Xiangteng He, Yuxin Peng, Hao Kong, and Qing Zhang. Posterlayout: A new benchmark and approach for content-aware visual-textual presentation layout. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6018-6026, 2023. 1, 2, 4, 5, 6, 7", + "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2", + "[21] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Layoutdm: Discrete diffusion model for controllable layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10167-10176, 2023. 6, 2", + "[22] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the International Conference on Machine Learning, pages 4904-4916, 2021. 1", + "[23] Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Constrained graphic layout generation via la" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8125", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tent optimization. In Proceedings of the ACM International Conference on Multimedia, pages 88-96, 2021. 2", + "[24] Suhas Kotha, Jacob Mitchell Springer, and Aditi Raghunathan. Understanding catastrophic forgetting in language models via implicit inference. In The Twelfth International Conference on Learning Representations, 2024. 4", + "[25] Hsin-Ying Lee, Lu Jiang, Irfan Essa, Phuong B Le, Haifeng Gong, Ming-Hsuan Yang, and Weilong Yang. Neural design network: Graphic layout generation with constraints. In Proceedings of the European Conference on Computer Vision, pages 491-506, 2020. 2", + "[26] Chenhui Li, Peiying Zhang, and Changbo Wang. HarmOnious textual layout generation over natural images via deep aesthetics learning. IEEE Transactions on Multimedia, 2021. 2, 6", + "[27] Fengheng Li, An Liu, Wei Feng, Honghe Zhu, Yaoyu Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junjie Shen, Zhangang Lin, et al. Relation-aware diffusion model for controllable poster layout generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 1249-1258, 2023. 2", + "[28] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. LayoutGAN: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7): 2388-2399, 2020. 2", + "[29] Jianan Li, Jimei Yang, Jianming Zhang, Chang Liu, Christina Wang, and Tingfa Xu. Attribute-conditioned layout GAN for automatic graphic design. IEEE Transactions on Visualization and Computer Graphics, 27(10):4039-4048, 2020. 2", + "[30] Yinan Li, Jia Chen, Yin Bai, Jia Cheng, and Jun Lei. Design element aware poster layout generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 1296-1305, 2024. 2", + "[31] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. Layoutprompt: Awaken the design ability of large language models. In Proceedings of the International Conference on Neural Information Processing Systems, pages 43852-43879, 2023. 1, 2, 5, 6", + "[32] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the ACM International Conference on Multimedia, pages 1250-1260, 2023. 1, 2", + "[33] Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. arXiv preprint arXiv:2308.08747, 2023. 4", + "[34] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. In Proceedings of the International Conference on Neural Information Processing Systems, pages 1877-1901, 2020. 4", + "[35] Meta. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/, 2024. 2, 8" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 6", + "[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In Proceedings of the International Conference on Neural Information Processing Systems, pages 27730-27744, 2022. 1", + "[38] Yi-Hao Peng, Faria Huq, Yue Jiang, Jason Wu, Amanda Xin Yue Li, Jeffrey Bigham, and Amy Pavel. Dreamstruct: Understanding slides and user interfaces via synthetic data generation. In Proceedings of the European Conference on Computer Vision, 2024. 2", + "[39] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7479-7489, 2019. 2", + "[40] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In Proceedings of the European Conference on Computer Vision, pages 38-56, 2022. 2, 6", + "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Intervention, pages 234-241, 2015. 3", + "[42] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1, 2, 5", + "[43] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In Proceedings of the International Conference on Neural Information Processing Systems, pages 25278–25294, 2022. 1", + "[44] Jaejung Seol, Seojun Kim, and Jaejun Yoo. Posterllama: Bridging design ability of language model to content-aware layout generation. In Proceedings of the European Conference on Computer Vision, pages 451-468, 2024. 1, 2, 6", + "[45] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF WInter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 2", + "[46] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. Lay-outnuwa: Revealing the hidden layout expertise of large language models. Proceedings of International Conference on Learning Representations, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8126", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 5", + "[48] Bo Wang, Quan Chen, Min Zhou, Zhiqiang Zhang, Xiaogang Jin, and Kun Gai. Progressive feature polishing network for salient object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12128-12135, 2020. 2", + "[49] Shaodong Wang, Yunyang Ge, Liuhan Chen, Haiyang Zhou, Qian Wang, Xinhua Cheng, and Li Yuan. Prompt2poster: Automatically artistic chinese poster creation from prompt only. In Proceedings of the ACM International Conference on Multimedia, pages 10716-10724, 2024. 1, 2", + "[50] Haohan Weng, Danqing Huang, Yu Qiao, Zheng Hu, ChinYew Lin, Tong Zhang, and CL Chen. Desigen: A pipeline for controllable design template generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12721-12732, 2024. 1, 2", + "[51] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Proceedings of the International Conference on Neural Information Processing Systems, pages 12077-12090, 2021. 5", + "[52] Chenchen Xu, Min Zhou, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10114-10123, 2023. 2, 6", + "[53] Tao Yang, Fan Wang, Junfan Lin, Zhongang Qi, Yang Wu, Jing Xu, Ying Shan, and Changwen Chen. Toward human perception-centric video thumbnail generation. In Proceedings of the ACM International Conference on Multimedia, pages 6653-6664, 2023. 1, 2", + "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2", + "[55] Min Zhou, Chenchen Xu, Ye Ma, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Composition-aware graphic layout GAN for visual-textual presentation designs. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 4995-5001, 2022. 1, 2, 4, 5, 6", + "[56] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 781 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8127", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/28953876-24ed-4c02-8fbd-02d889f7bc8b_model.json b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/28953876-24ed-4c02-8fbd-02d889f7bc8b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2dc16eb5f824b4300fbf8767e2d870a4c99ba00e --- /dev/null +++ b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/28953876-24ed-4c02-8fbd-02d889f7bc8b_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.82, + 0.177 + ], + "angle": 0, + "content": "PosterO: Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.204, + 0.753, + 0.241 + ], + "angle": 0, + "content": "HsiaoYuan Hsu and Yuxin Peng* \nWangxuan Institute of Computer Technology, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.243, + 0.692, + 0.257 + ], + "angle": 0, + "content": "kslh99@stu.pku.edu.cn, pengyuxin@pku.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.261, + 0.48, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.262, + 0.891, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.462, + 0.89, + 0.476 + ], + "angle": 0, + "content": "Figure 1. Illustration of the generalized settings in content-aware layout generation from both (a) data and (b) approach perspectives." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.482, + 0.327, + 0.497 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.504, + 0.485, + 0.881 + ], + "angle": 0, + "content": "In poster design, content-aware layout generation is crucial for automatically arranging visual-textual elements on the given image. With limited training data, existing work focused on image-centric enhancement. However, this neglects the diversity of layouts and fails to cope with shape-variant elements or diverse design intents in generalized settings. To this end, we proposed a layout-centric approach that leverages layout knowledge implicit in large language models (LLMs) to create posters for omnifarious purposes, hence the name PosterO. Specifically, it structures layouts from datasets as trees in SVG language by universal shape, design intent vectorization, and hierarchical node representation. Then, it applies LLMs during inference to predict new layout trees by in-context learning with intent-aligned example selection. After layout trees are generated, we can seamlessly realize them into poster designs by editing the chat with LLMs. Extensive experimental results have demonstrated that PosterO can generate visually appealing layouts for given images, achieving new state-of-the-art performance across various benchmarks. To further explore PosterO's abilities under the generalized settings, we built PStylish7, the first dataset with multi-purpose posters and various-shaped elements, further offering a challenging test for advanced research. Code and dataset will be publicly available at https://thekinsley.github.io/PosterO/." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.482, + 0.645, + 0.497 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.625 + ], + "angle": 0, + "content": "Content-aware layout generation makes the poster design process intelligent by automatically arranging visual-textual elements on given images. As the demand for effective design grows rapidly, a mainstream platform surpasses 185 million monthly users [3], and \\(90\\%\\) of survey participants take AI-empowered features as key considerations [2]. This technology is gaining more attention for its practical value in various applications [32, 49, 50, 53], as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Unlike other generation tasks [22, 43], collecting image-Layout pairs for the target task requires costly annotation and inpainting processes [19, 55]. Therefore, existing work has incorporated object hints [4, 55], prior information [7, 19], retrieved references [17, 31], and augmented data [44] to confront the challenge of training data scarcity. However, because these are mainly image-centric enhancements without increasing the adaptability and diversity of layouts, they are prone to get trapped in the local solution space [18]. On the other hand, while leveraging related knowledge implicit in large language models (LLMs) [11, 37, 42, 47] showed the potential for layout-centric enhancements [12, 46], existing work [31, 44] has suffered from the semantically impoverished layout representations. The two primary issues are the monotonous rectangular shape of elements and the isolated visual constraints, which led to failures at capturing the complexity of real-world layouts and the intricate visual relationship between images and elements." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.242, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8117" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "To pursue data efficiency and semantic richness required for advances in content-aware layout generation, we proposed PosterO to enable in-context learning (ICL) of LLMs by structuring layout tree representations. Specifically, it consists of three procedures, including (a) layout tree construction that jointly represents layout elements and design intents (i.e., available areas on images) as hierarchical nodes in SVG trees [9], (b) layout tree generation that dynamically selects intent-aligned learning examples for the test input to facilitate ICL, and (c) poster design realization that integrates design materials into the generated layout trees to create posters in the subsequent chats with LLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.545 + ], + "angle": 0, + "content": "We conducted extensive experiments on public benchmarks [19, 55] and demonstrated that PosterO has achieved new state-of-the-art performance. It particularly shows stability in managing domain adaption problems [52] and spatial distribution shifts [17] that make the current approaches stumped. In additional efforts, we enhanced the evaluation process through newly added intent-aware content metrics that are more credible than currently used saliency-aware ones. We also demonstrated the adaptability of PosterO to different design intents and small-scale LLMs [35], highlighting its applicability. To further explore its potential under the generalized settings of diverse poster purposes (e.g., commercial or education) and shape-variant elements (e.g., circles or curves), we built the first dataset, PStylish7, consisting of 152 few-shot learning samples and 100 test images, covering seven representative poster purposes and eight element types. Serving as a challenging test, it is believed to facilitate advanced research in this field." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.547, + 0.48, + 0.561 + ], + "angle": 0, + "content": "We summarize the contribution of this paper as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.563, + 0.483, + 0.608 + ], + "angle": 0, + "content": "- A language-based content-aware layout representation, layout tree, uses hierarchical nodes to model both layout elements and design intents with arbitrary shapes." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.609, + 0.483, + 0.653 + ], + "angle": 0, + "content": "- A layout-centric approach, PosterO, leverages implicit knowledge of LLMs by applying a few intent-aligned layout trees as examples to perform in-context learning." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.654, + 0.483, + 0.699 + ], + "angle": 0, + "content": "- Extensive evaluations and newly added metrics verify the superior effectiveness of PosterO in confronting various adaptation problems that plague current approaches." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.7, + 0.483, + 0.744 + ], + "angle": 0, + "content": "- A dataset for generalized content-aware layout generation, PStylish7, incorporates seven representative poster purposes and eight shape-variant element types." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.563, + 0.483, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.483, + 0.775 + ], + "angle": 0, + "content": "As our best knowledge, this work is the first to consider the generalized settings of content-aware layout generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.233, + 0.807 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.392, + 0.833 + ], + "angle": 0, + "content": "2.1. Content-aware Layout Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Apart from general layout tasks [1, 15, 23, 25, 28], content-aware layout generation is a cross-modal task [10, 16] integrating visual conditions, i.e., image canvases, which are inherent in various real-world applications [32, 49, 50, 53]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "highlighting its practical utility. With advances in deep generative models, data-centric approaches have become dominant in this field, categorized into GAN-based [18, 19, 52, 55], auto-regression-based [4, 5, 17, 30], and diffusion model-based [7, 27] approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.906, + 0.426 + ], + "angle": 0, + "content": "Entering into data construction of image layout pairs, CGL-GAN [55] leveraged an inpainting model [45] to remove design elements from posters and a saliency object detection model [48] to give generative networks additional hints about image compositions. Similar saliency-enhanced paradigm [26, 39, 40] has been adopted by most subsequent approaches. Dealing with limited training data, DS-GAN [19] introduced prior knowledge [14, 29] to arrange layout elements in a meticulously designed order, making the patterns existing in the data more efficiently mined. RALF introduced retrieval augmentation, leveraging a composition-aware similarity evaluation model [13] to find the nearest neighbors for a given image, and incorporating their layout features [23] as additional inputs for the generative network. Moreover, Chai et al. [7] adapted a general layout diffusion model [6] for content-aware tasks by incorporating aesthetic constraints and a saliency-aware layout plausibility ranker." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.906, + 0.516 + ], + "angle": 0, + "content": "As current approaches are engaged in image-centric enhancement yet seldom expand the diversity of layouts, the accessible solution space is narrowed, impeding their ability to address spatially varying design intents across different applications. In light of this, we propose a layout-centric, intent-aware approach in this work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.879, + 0.545 + ], + "angle": 0, + "content": "2.2. Language Model-based Layout Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.688 + ], + "angle": 0, + "content": "While numerical layout representations risk losing the semantic information of attributes [46], current advances in large language models (LLMs) [11, 42, 47] open up new possibilities with HTML-based [31, 38], CSS-based [8, 12], and SVG-based [44, 46] language representations as these approaches enable leverage of layout knowledge implicit in LLMs. However, due to the challenge of guiding LLMs to perceive visual content, only a few approaches [31, 44] have supported content-aware layouts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.81 + ], + "angle": 0, + "content": "LayoutPrompter [31] represented each layout element as a
in HTML. While adopting a saliency paradigm, it only extracted the minimum bounding rectangle from the binary map as coarse content constraints and then retrieved layout examples for in-context learning. In contrast, PosterLlama [44] adopted a SVG-based representation [46] using shapes. It then fine-tuned DINOv2 [36, 56] with CodeLlama [20, 42] enhanced by data augmentation [54]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As current approaches employed monotype element representations and isolated visual representations, they failed to exploit the rich semantics of languages nor explicitly capture the relationship between elements and images. In light of this, we propose layout trees to represent various-shaped elements and visual hints jointly." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8118" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.093, + 0.46, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.093, + 0.765, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.094, + 0.903, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.308, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Figure 2. An overview of PosterO. First, (a) takes image layout pairs \\((I,L)\\) from datasets as input for data preparation, jointly modeling various-shaped elements \\(E\\) and design intents \\(D\\) towards layout trees \\(T\\) and building up the latent space of \\(D\\). Then, (b) takes test images \\(I_{t}\\) as input and searches examples based on the predicted design intents \\(f_{t}\\) to apply LLM \\(\\mathcal{M}\\) through in-context learning. After obtaining generated layout trees \\(\\hat{T}\\), (c) can continue the conversation with \\(\\mathcal{M}\\) to create poster designs seamlessly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.375, + 0.4, + 0.391 + ], + "angle": 0, + "content": "3. The Proposed Approach: PosterO" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Considering the limitations of image-centric and language model-based approaches, we propose PosterO. It is layout-centric and enables large language models (LLMs) to generate visually appealing layouts for given images by structurally representing the content-aware layouts as trees \\( T \\) in SVG language. An overview of PosterO is shown in Fig. 2. In brief, it is composed of three procedures, namely, (a) layout tree construction for data preparation, (b) layout tree generation applying an LLM \\( \\mathcal{M} \\) through \\( k \\)-shot in-context learning, and (c) poster design realization through the subsequent chat with \\( \\mathcal{M} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.327, + 0.589 + ], + "angle": 0, + "content": "3.1. Layout Tree Construction" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.596, + 0.347, + 0.61 + ], + "angle": 0, + "content": "3.1.1. Universal Shape Vectorization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.795 + ], + "angle": 0, + "content": "In content-aware layout generation, each data pair contains an image \\( I \\) and a layout \\( L \\). Usually, \\( L \\) is represented as a set of numerical tuples \\( \\{e_i\\}_{i}^n = \\{(c_i, b_i)\\}_{i}^n \\), where \\( c_i \\) and \\( b_i = (x, y, w, h) \\) indicate the category and bounding box of the \\( i \\)-th element \\( e_i \\) out of all \\( n \\) elements. As this is adequate only for rectangular elements, a more generalized representation becomes essential for the diverse shapes encountered in real-world posters, such as circles and curves. In light of this, SVG (Scalable Vector Graphics) emerges as an ideal choice because of its ability to represent arbitrary shapes and, more importantly, its nature as a language to facilitate the layout knowledge implicit in LLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Following the published SVG standards [9], we define five shapes to encompass a wide range of elements typically found in posters, as shown in Fig. 3. First, (a) regular rectangle is vectorized using \\( \\langle \\text{rect} \\rangle \\). Also, its two variants, including (b) vertical rectangle sensitive to the aspect ratio and (c) rotated rectangle using the rotate function in the transform attribute to indicate the rotation degree." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.372, + 0.638, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.463, + 0.63, + 0.474 + ], + "angle": 0, + "content": "(a) General rectangle" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.372, + 0.772, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.463, + 0.764, + 0.474 + ], + "angle": 0, + "content": "(b) Vertical rectangle" + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.372, + 0.905, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.463, + 0.898, + 0.474 + ], + "angle": 0, + "content": "(c) Rotated rectangle" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.476, + 0.63, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.56, + 0.604, + 0.572 + ], + "angle": 0, + "content": "(d) Ellipse" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.475, + 0.771, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.56, + 0.758, + 0.572 + ], + "angle": 0, + "content": "(e) Complex curve" + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.475, + 0.905, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.561, + 0.874, + 0.572 + ], + "angle": 0, + "content": "(f) Polygon" + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.575, + 0.882, + 0.59 + ], + "angle": 0, + "content": "Figure 3. Various shapes of elements in PStylish7 dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.907, + 0.662 + ], + "angle": 0, + "content": "When delving into more intricate shapes, (d) ellipse is vectorized using , which can also handle the circle as a special case, and (e) complex curve is vectorized using approximated by multiple cubic Bezier curves." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.671, + 0.75, + 0.685 + ], + "angle": 0, + "content": "3.1.2. Design Intent Vectorization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.841 + ], + "angle": 0, + "content": "In addition to visual-textual elements, the design intents associated with the layout \\( L \\) must also be vectorized. Specifically, design intents \\( D \\) are embodied by a set of areas within the image \\( I \\), suitable or intended for placing the elements. By incorporating these areas' representations \\( N_{D} \\), the layout tree becomes content-aware and capable of conveying visual relationships between elements and images to LLMs, which cannot 'see' the images. As \\( D \\) can be learned from existing data or customized case by case, this also provides a flexible interface for indicating various purposes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.901 + ], + "angle": 0, + "content": "By leveraging image layout pairs from posters with a similar purpose, a design intent detection model \\(S\\) based on U-Net [41] is trained in a semi-supervised manner. Referring to [18], \\(S\\) takes the bitmap of layout elements as the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8119" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "supervision signal during the optimization, while the consistency loss is canceled to clearly segment out the areas. With the pre-trained \\( S \\), PosterO can obtain design intents \\( D \\) aligning the target purpose for any image \\( I \\). Specifically, after thresholding and morphological operations, we approximate the contour of the predicted intent maps \\( m \\) and vectorize the obtained irregular areas using in SVG language. An example is shown in Fig. 3(f). Besides, the intermediate embeddings \\( f = S_{\\mathfrak{e}}(I) \\), where \\( S_{\\mathfrak{e}} \\) denotes the last block of the encoder, are also collected to serve as their representations in the latent space of design intents \\( \\mathbf{Z}_D \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.264, + 0.376, + 0.279 + ], + "angle": 0, + "content": "3.1.3. Hierarchical Node Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.487, + 0.51 + ], + "angle": 0, + "content": "As the nodes of layout elements \\( N_{E} \\) and design intents \\( N_{D} \\) are in place, the layout tree \\( T \\) is constructed and encapsulated within an SVG container aware of the image resolution \\( (w_{I}, h_{I}) \\), as: \\( <\\mathrm{svg~width}=\\{\"w_{I}\\}\" \\mathrm{height}=\\ {\"h_{I}\\}\" \\mathrm{xmlns}=\\text{\"http://www.w3.org/2000/sv\"}> \\{N_{D}\\} \\{N_{E}\\}<\\mathrm{svg}> \\). \\( N_{D} \\) is put ahead to ensure that \\( N_{E} \\) is arranged considering their probabilistic dependencies in the autoregression, thus generating suitable content-aware layouts. In addition, as the special elements that wrap around others, such as underlays, are usually used in poster design, a novel hierarchical representation explicitly modeling these enveloping properties is proposed. Specifically, \\( N_{E} \\) is sorted by prioritizing underlays with larger areas. Then, given a node \\( N_{a} \\) and the corresponding nodes \\( N_{B} = \\{N_{b} \\mid N_{b} \\in \\{N_{E}\\}_{a + 1}^{n} \\land C(N_{a}, N_{b})\\} \\), where the condition function \\( C(\\cdot) \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.516, + 0.482, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} C (\\cdot) = \\left(| N _ {a}. x - N _ {b}. x | \\leq \\epsilon\\right) \\wedge \\left(| N _ {a}. y - N _ {b}. y | \\leq \\epsilon\\right) \\\\ \\wedge \\left(\\left| N _ {a}. x + N _ {a}. w - N _ {b}. x - N _ {b}. w \\right| \\leq \\epsilon\\right) \\tag {1} \\\\ \\wedge \\left(\\left| N _ {a}. y + N _ {a}. h - N _ {b}. y - N _ {b}. h \\right| \\leq \\epsilon\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.485, + 0.671 + ], + "angle": 0, + "content": "a subtree is constructed as \\(<\\text{svg} \\times=\" \\{N_{a}.x\\}\" \\text{y=\"} \\{N_{a}.y\\}\">\\) \\(\\{N_{a}'\\} \\{N_{B}'\\} < / \\text{svg}>\\) and inserted at the position of \\(N_{a}\\), where the processed \\(N_{i}'\\) has the relative geometry attributes \\((N_{i}'.x, N_{i}'.y) = (N_{i}.x - N_{a}.x, N_{i}.y - N_{a}.y)\\). All leaf nodes of \\(T\\) are assigned unique identifiers \\(\\mathfrak{id} = \"\\{c_{i} - i\\}\"\\) to facilitate precise manipulation and querying." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.677, + 0.313, + 0.693 + ], + "angle": 0, + "content": "3.2. Layout Tree Generation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.699, + 0.368, + 0.715 + ], + "angle": 0, + "content": "3.2.1. Intent-aligned Example Selection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Upon findings from [34], LLMs are capable of in-context learning (ICL), requiring only a few examples to perform downstream tasks. This significantly reduces the need for task-specific data and avoids the extensive costs or catastrophic forgetting [24, 33] associated with fine-tuning. To make the best use of layout knowledge implicit in LLMs during ICL, selecting examples that align with the design intent is crucial. Therefore, given the test image \\(I_{t}\\), its embedding \\(f_{t}\\) is first extracted by the detection model \\(\\mathcal{S}\\); subsequently, \\(k\\) layout trees \\(\\{T_j\\}_j^k\\) are selected as in-context examples, where their embeddings \\(\\{f_j\\}_j^k\\) are the \\(k\\) closest to \\(f_{t}\\) in the latent space \\(\\mathbf{Z}_{D}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.092, + 0.862, + 0.107 + ], + "angle": 0, + "content": "3.2.2. In-context Learning for Layout Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.111, + 0.907, + 0.278 + ], + "angle": 0, + "content": "To provide clear structure and objectives in the prompt \\(P\\) for the LLM \\(\\mathcal{M}\\), each intent-aligned example \\(T_{j}\\) is preceded by a short sentence that informs its number \\(j\\), the resolution of its corresponding image, design intents \\(D\\), and identifiers of its nodes, as shown in Fig. 2(b). The example field is then encapsulated within a preface and postscript, with the postscript delivering the request message of the test image \\(I_{t}\\) through a short sentence mirroring that of the example. Complete prompts \\(P\\) are presented in supplementary material. Finally, by inputting \\(P\\) into \\(\\mathcal{M}\\) and interpreting the responses, PosterO completes the layout generation process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.287, + 0.747, + 0.303 + ], + "angle": 0, + "content": "3.3. Poster Design Realization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Benefiting from the essence of the generated layout tree \\(\\hat{T}\\) as graphics and the rigid id assignment for elements, guiding \\(\\mathcal{M}\\) to transform \\(\\hat{T}\\) into a practical poster design is feasible and straightforward. Specifically, the transformation contains mockup creation, where shape elements are converted to or regarding their categories, and material synthesis, where elements are filled with actual design contents, e.g., slogans and hyperlinks to images. Notably, this transformation is achieved with zero-shot learning, yet \\(\\mathcal{M}\\) brings impressive results (in supplementary material) and even predicts suitable font sizes for texts, highlighting the value of design knowledge implicit in LLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.505, + 0.646, + 0.521 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.616, + 0.544 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.75 + ], + "angle": 0, + "content": "To evaluate the proposed PosterO, two public e-commerce poster datasets are used, while their train+valid/annotated test/unannotated test splits are allocated as in [17] to ensure a fair comparison with existing work. Specifically, PKU PosterLayout [19] comprises 8,734/1,000/905 images with three layout element types, including text, logo, and underlay; CGL [55] comprises 54,546/6,002/1,000 images with four layout element types, including text, logo, underlay, and embellishment. Furthermore, to explore PosterO's abilities under the generalized settings, we built the first multipurpose poster dataset, PStylish7, containing layout elements of various shapes. More details about PStylish7 are depicted in Sec. 4.6 and supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.698, + 0.773 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Following previous work [17, 19, 55], we use graphic and content metrics. Graphic metrics include overlay \\(Ove\\downarrow\\) alignment \\(Ali\\downarrow\\) , and loose, strict underlay effectiveness \\(Und_{l}\\uparrow ,Und_{s}\\uparrow ;\\) Content metrics include non-salient space utilization \\(Uti\\uparrow\\) , salient space occlusion \\(Occ\\downarrow\\) , and readability \\(Rea\\downarrow\\) . Moreover, we newly proposed two content metrics, design intent coverage \\(Cov\\uparrow\\) and conflict \\(Con\\downarrow\\) based on a design intent detection model \\(S_{pc}\\) that is compre" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8120" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.649, + 0.52 + ], + "angle": 0, + "content": "
MethodOve ↓Ali ↓Undl ↑Unds ↑Int ↓Sal ↓Rea ↓Avg ↓
Annotated Test Split
Real data0.00100.00380.99420.99030.02890.03680.01090.0138
CGL-GAN0.09660.00350.78540.35700.12040.27740.01910.1964
DS-GAN0.02610.00380.83500.58040.06770.31000.01990.1446
ICVT0.25720.04050.53840.39320.78311.08360.02590.4655
LayoutDM†0.15620.00180.64260.38730.50660.24290.01850.2709
RALF0.00840.00280.98080.92010.11620.04790.01280.0410
Tuning-required
PosterLlama‡0.00080.00060.99990.99820.34070.26190.01770.0891
LLM-based
Tuning-free
LayoutPrompterL20.00830.00300.37160.16330.32000.94330.03190.3959
LayoutPrompterCL0.00400.00300.44150.19370.36620.90350.03010.3817
LayoutPrompterL30.00170.00280.40850.16130.51990.85670.03090.4060
PosterOL2 (Ours)0.00040.00270.75970.56050.36790.14910.01880.1741
PosterOLCL (Ours)0.00040.00280.96210.86060.23970.22950.01890.0955
PosterOL3 (Ours)0.00060.00250.98880.92110.10800.11100.01710.0470
Unannoted Test Split
Non-LLMTrain data0.00120.00270.99690.99000.00000.00000.0108
CGL-GAN0.10100.00480.73260.27430.22750.77550.0327
DS-GAN0.02480.00460.78590.46760.16590.58680.0320
ICVT0.27860.04800.49390.35490.79471.32890.0347
LayoutDM†0.16380.00290.59870.36950.53570.31720.0264
RALF0.01750.00690.95480.86530.59860.17790.0231
Tuning-required
PosterLlama‡0.00060.00060.99860.99170.26670.38230.02850.0983
LLM-based
Tuning-free
LayoutPrompterL20.00950.00320.34010.11610.53491.20500.04080.4767
LayoutPrompterCL0.00320.00280.40520.14640.51481.11060.03950.4456
LayoutPrompterL30.00100.00260.40540.16210.70051.08380.04120.4660
PosterOL2 (Ours)0.00040.00310.72840.51880.40300.33510.02650.2173
PosterOLCL (Ours)0.00050.00280.95950.84510.27960.39450.02730.1286
PosterOL3 (Ours)0.00040.00250.98560.92410.14270.21310.02480.0677
" + }, + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.524, + 0.599, + 0.538 + ], + "angle": 0, + "content": "Table 1. Comparisons of quantitative results on PKU PosterLayout dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.684 + ], + "angle": 0, + "content": "hensively pre-trained on [19] and [55]. Calculating \\( Cov\\uparrow \\) and \\( Con\\downarrow \\) is similar to \\( Uti\\uparrow \\) and \\( Occ\\downarrow \\) but using the predicted intent maps. As reported in Tab. 2, intent maps have doubled the matched portion with real layout data, reduced the unmatched portion by \\( 34.7\\% \\), and required only \\( 1.3\\% \\) computational costs. These findings have demonstrated that intent-aware metrics are more credible and efficient. For a holistic understanding, we standardize intent-aware and saliency-aware metrics to \\( Int\\downarrow \\) and \\( Sal\\downarrow \\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.695, + 0.483, + 0.735 + ], + "angle": 0, + "content": "\\[\nI n t = \\Sigma (\\frac {\\left| C o v _ {\\hat {L}} - C o v _ {L} \\right|}{1 - C o v _ {L}}, \\frac {\\left| C o n _ {\\hat {L}} - C o n _ {L} \\right|}{C o n _ {L}}), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.731, + 0.429, + 0.763 + ], + "angle": 0, + "content": "\\[\nS a l = \\Sigma (\\frac {| U t i _ {\\hat {L}} - U t i _ {L} |}{1 - U t i _ {L}}, \\frac {| O c c _ {\\hat {L}} - O c c _ {L} |}{O c c _ {L}}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.483, + 0.822 + ], + "angle": 0, + "content": "where \\(L\\) and \\(\\hat{L}\\) are the layouts from the train split and generation. Finally, we introduce an average metric \\(Avg\\downarrow = \\Sigma (Ove,Ali,1 - Und_{l},1 - Und_{s},Int,Sal,Rea)\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.308, + 0.849 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To ensure independent evaluations on [19] and [55], we correspondingly train two detection models, \\(S_{p}\\) and \\(S_{c}\\), with mean squared error loss and a learning rate of \\(1e^{-6}\\). Each" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.09, + 0.9, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.496, + 0.902, + 0.537 + ], + "angle": 0, + "content": "Figure 4. Comparisons of visualized results on the unannotated test split of PKU PosterLayout dataset." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.545, + 0.905, + 0.661 + ], + "angle": 0, + "content": "
PredictionPKUCGL
Match ↑Unmatch ↓Match ↑Unmatch ↓
Saliency map0.22380.11930.19840.1352
Design intent map0.44140.09960.40360.0883
Inference costNetwork architecture#FLOPs(G)#Params(M)
Saliency mapISNet + BASNet287.83131.11
Design intent mapUNet (MiT-B1)3.7416.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.67, + 0.907, + 0.713 + ], + "angle": 0, + "content": "Table 2. Comparisons of the predicted saliency and design intent maps on the annotated test splits of PKU PosterLayout and CGL datasets, showing that intent-aware metrics are more credible." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "detection model is with a Transformer encoder tailored for semantic segmentation, MiT-B1 [51]. For in-context learning (ICL), we implement \\(\\mathcal{M}\\) with three open-source LLMs, including Llama 2-7B (L2) [47], CodeLlama-7B (CL) [42], and Llama 3.1-8B (L3) [11], with an unified sampling temperature of 0.7. Considering the context lengths and remaining consistent with LayoutPrompter [31], the example size \\(k\\) is 5, 10, and 10, respectively. Still adhering to [31], a layout ranker is employed, while mIoU in the quality function measures the similarity between generated layouts and predicted design intents rather than real data, minimizing the risk of data leakage." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8121" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.649, + 0.52 + ], + "angle": 0, + "content": "
MethodOve ↓Ali ↓Undl ↑Unds ↑Int ↓Sal ↓Rea ↓Avg ↓
Annotated Test Split
Non-LLMReal data0.00030.00240.99630.98800.00880.00970.01190.0070
CGL-GAN0.22910.01230.64660.22810.70880.46270.02130.3656
DS-GAN0.04600.00220.90810.63080.39500.15390.01810.1537
ICVT0.24530.01790.51500.33260.36320.51390.02110.3305
LayoutDM†0.01840.00210.92160.81590.04110.02860.01370.0523
RALF0.00420.00240.99120.97560.03660.07240.01800.0238
Tuning-required PosterLlama‡0.00060.00060.99870.98900.13370.25590.01840.0602
LLM-basedTuning-free
LayoutPrompterL20.00870.00330.33690.14550.33620.82340.03270.3888
LayoutPrompterCL0.00370.00310.39650.16160.37180.77990.03170.3760
LayoutPrompterL30.00170.00300.38300.17400.55930.79570.03270.4051
PosterOL2 (Ours)0.00060.00230.78010.61500.25420.04670.01760.1323
PosterOLCL (Ours)0.00050.00230.96820.88680.12460.07000.01680.0513
PosterOL3 (Ours)0.00060.00220.98420.93400.03880.00260.01610.0203
Unannotated Test Split
Non-LLMTrain data0.00030.00270.99450.98580.00000.00000.01180.0049
CGL-GAN0.26680.03160.67740.16560.97072.39730.05120.6964
DS-GAN0.09910.01380.75660.28100.46072.27320.05260.5517
ICVT0.20450.10100.43570.25991.28132.65120.03970.7974
LayoutDM†0.07930.18220.63040.38531.50023.29060.06120.8711
RALF0.02730.01890.97560.93151.02371.67980.03970.4118
Tuning-required PosterLlama‡0.00140.00070.99710.97710.35102.61700.05550.4359
LLM-basedTuning-free
LayoutPrompterL20.00690.01030.28470.11261.02572.37020.06280.7255
LayoutPrompterCL0.00450.00240.32390.13411.58002.53800.06500.8188
LayoutPrompterL30.00260.00160.26930.11421.62492.41280.06440.8175
PosterOL2 (Ours)0.00050.00330.66630.47700.35221.23650.03820.3553
PosterOLCL (Ours)0.00030.00370.93690.80710.27711.32230.03790.2710
PosterOL3 (Ours)0.00040.00320.98160.90340.43281.24070.03650.2612
" + }, + { + "type": "table_caption", + "bbox": [ + 0.195, + 0.531, + 0.559, + 0.545 + ], + "angle": 0, + "content": "Table 3. Comparisons of quantitative results on CGL dataset." + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.09, + 0.902, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.502, + 0.902, + 0.543 + ], + "angle": 0, + "content": "Figure 5. Comparisons of visualized results on the unannotated test split on CGL dataset." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.556, + 0.391, + 0.572 + ], + "angle": 0, + "content": "4.4. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.483, + 0.655 + ], + "angle": 0, + "content": "We select approaches with open-sourced implementation as baselines, including GAN-based CGL-GAN [55], DS-GAN [19], auto-regression-based ICVT [4], RALF [17], diffusion model-based LayoutDM† [21], and also LLM-based PosterLlama‡ [44], LayoutPrompter [31]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.821 + ], + "angle": 0, + "content": "Baseline comparison. Tab. 1 reports the quantitative results on PKU PosterLayout. The proposed PosterO shows the best overall performance across the test splits. Particularly, when confronting the challenging unannotated split that poses domain adaptation problems [52], it still achieves new state-of-the-art (SOTA) results with outstanding stability. Compared to the current non-LLM and LLM-based SOTA approaches, i.e., RALF and PosterLlama, ours significantly improves the average metric \\(Avg\\downarrow\\) by \\(52.8\\%\\) and \\(31.1\\%\\). Consistent results are obtained on CGL, as reported in Tab. 3, ours improves \\(Avg\\downarrow\\) by \\(36.6\\%\\) and \\(40.1\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.483, + 0.867 + ], + "angle": 0, + "content": "More precisely, PosterO shows impressive graphic metrics (2nd column), closely resembling the distribution of real data, with only a slight deficit in the Ali \\(\\downarrow\\) metric on the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.907, + 0.693 + ], + "angle": 0, + "content": "unannotated split of CGL compared to LayoutPrompter. As for content metrics (3rdcolumn), although PosterO uses only a small design intent detection model \\(S\\), it achieves comparable or even superior results to the baselines that utilize larger saliency detection models [26, 40] or costly fine-tuned DINOv2 [36]. Especially, PosterO is best at dealing with the spatial shifts in subject distributions [17] present in the unannotated split of CGL, maintaining an overwhelming lead in content metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Backbone LLMs \\(\\mathcal{M}\\). Comparing the performance of tuning-free approaches across various LLMs is crucial for realizing how effectively the layout knowledge implicit in \\(\\mathcal{M}\\) has been leveraged. As reported in Tab. 1 and Tab. 3, PosterO demonstrates overwhelming superiority over LayoutPrompter in all counterparts with the same \\(\\mathcal{M}\\). More importantly, from Poster\\(_{L2,CL}\\), to Poster\\(_{L3}\\), it is evident that the proposed approach has experienced improvements as the capabilities of its based \\(\\mathcal{M}\\) increased. This finding indicates that new knowledge, such as programming, introduced during model updates, has indeed been leveraged to generate layouts. With the continuous development of LLMs, PosterO is anticipated to keep its rising potential." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.333, + 0.888 + ], + "angle": 0, + "content": "The extended version presented in [17]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.479, + 0.901 + ], + "angle": 0, + "content": "\\(\\ddagger\\)With the released weights tuned on [19], [55], and synthesized data." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.479, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8122" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.136 + ], + "angle": 0, + "content": "
Method \\ k1257810
LayoutPrompter0.77110.66900.53530.48830.49190.4660
PosterO (Ours)0.06780.07460.07320.06630.06510.0677
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.141, + 0.476, + 0.155 + ], + "angle": 0, + "content": "Table 4. Comparisons of \\(Avg\\downarrow\\) over in-context example size \\(k\\)." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.164, + 0.465, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.338, + 0.458, + 0.352 + ], + "angle": 0, + "content": "Figure 6. Content metrics over in-context example size \\( k \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.485 + ], + "angle": 0, + "content": "Visualized results. Fig. 4 and Fig. 5 show the layouts generated by different approaches. The results illustrate that PosterO generates high-quality layouts that the distribution of elements is generous and coordinated. In contrast, the approaches based on training or tuning tend to place elements centrally and thus fail to confront images with compositions that are less common in the training data. More visualized results are presented in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.483, + 0.73 + ], + "angle": 0, + "content": "In-context example size \\( k \\). To explore the impact of example selection during ICL, we test a set of candidates \\( \\{k\\} = \\{1,2,5,7,8,10\\} \\) on the unannotated split of [19], as reported in Tab. 4. Amazingly, PosterO consistently outperforms LayoutPrompter and all other baselines in Tab. 1 across all settings of \\( k \\). As the optimal performance occurs at \\( k = 7 \\) and 8, it is proved that \\( \\mathcal{M} \\) has truly benefited from learning with the selected intent-aligned examples. To further observe how well these examples help \\( \\mathcal{M} \\) understand visual contents, which is the focal point of content-aware layout generation, we plot the resultant content metrics over \\( k \\) in Fig. 6. It is noteworthy that directly adopting the most intent-aligned example (Top-1 Retrieval) surpasses all baselines, indicating a successful alignment between intent and layout in the latent space \\( \\mathbf{Z}_D \\). Nevertheless, we still observe improvements when involving in-context examples." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.245, + 0.756 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.839 + ], + "angle": 0, + "content": "To investigate our design choice in layout tree construction (Sec. 3.1) and learning example selection (Sec. 3.2), as well as to explore PosterO's adaptability to customized design intents and small-scale LLMs, we conducted the following ablation studies, mainly on the unannotated split of [19]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Design choices for layout tree construction. We investigate the effectiveness of the design intent vectorization \\((d)\\) and hierarchical node representation \\((h)\\) during constructing layout trees, as reported in Tab. 5. Due to space constraints," + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.905, + 0.171 + ], + "angle": 0, + "content": "
dhUndl↑Unds↑Int↓Sal↓Rea↓Avg↓
✓✓0.98560.92410.14270.21310.02480.0677
✓0.86090.66500.09810.15600.02400.1078
✓0.94970.88360.98580.39910.02740.2260
0.76280.64130.81620.27290.02460.2447
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.176, + 0.905, + 0.204 + ], + "angle": 0, + "content": "Table 5. Ablation study on the \\((d)\\) design intent vectorization and \\((h)\\) hierarchical node representation in layout tree construction." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.213, + 0.903, + 0.292 + ], + "angle": 0, + "content": "
SelectionUndl↑Unds↑Int↓Sal↓Rea↓Avg↓
f-aligned0.98560.92410.14270.21310.02480.0677
D-aligned0.98660.92970.22210.25310.02360.0836
E-aligned0.97430.90260.34310.41770.02640.1304
Random0.98850.93020.32250.46880.02950.1293
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.297, + 0.905, + 0.34 + ], + "angle": 0, + "content": "Table 6. Ablation study on different references for in-context example selection. (f: Embeddings of design intents, \\(D\\): Bounding boxes of design intents, \\(E\\): Sets of element types.)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.59 + ], + "angle": 0, + "content": "\\(Ove\\downarrow\\) and \\(Ali\\downarrow\\) are omitted, considering their slight fluctuations. As observed, the combination of \\(d\\) and \\(h\\) yields the best results. Although removing \\(h\\) appears to improve content metrics, this is primarily because the invalid empty underlays artificially inflate the utilization rate, as analyzed in supplementary material. The significant decreases witnessed in underlay effectiveness, \\(Und_{l}\\), \\(Und_{s}\\), by up to \\(28\\%\\), further underscore this issue. In contrast, removing \\(d\\) from both the examples and postscript makes LLMs blindly mimic the provided instances. Here, however, shows a \\(Sal\\downarrow\\) surpasses some baselines, such as CGL-GAN, DS-GAN, and LayoutPrompter, while \\(Int\\downarrow\\) fails far behind, reflecting the robustness of the newly proposed metric \\(Int\\downarrow\\). If both \\(d\\) and \\(h\\) are removed, the most severe degradation is observed, up to \\(261\\%\\). In this way, we demonstrate the necessity of these design choices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.909, + 0.806 + ], + "angle": 0, + "content": "Design choices for in-context example selection. We have seen the remarkable performance of design intent embeddings \\( f \\)-aligned example selection, which relies on the learned latent space \\( \\mathbf{Z}_D \\). To investigate alternative references, we have considered bounding boxes of design intents (\\( D \\)-aligned), sets of element types (\\( E \\)-aligned), and even with no specific alignment (Random). As reported in Tab. 6, although all \\( D \\)-aligned, \\( E \\)-aligned, and random selections make declines in average metrics, the outcomes are still superior to most baseline approaches. In particular, using \\( D \\)-aligned selection also achieves a new SOTA result, which supports its eligibility as an alternative enabling PosterO to deliver acceptable content-aware layouts when faced with roughly sketched \\( D \\) from users." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Adaption to customized design intent. Upon the discussions from the previous paragraph, we take a further step to verify PosterO's adaptability to customized design intents. Fig. 7 demonstrates the results of using the original \\(D\\) predicted by the detection model \\(S\\) and three customized \\(D\\) with different levels of complexity. As the input \\(D\\) varies," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8123" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.478, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.232, + 0.47, + 0.246 + ], + "angle": 0, + "content": "Figure 7. PosterO's adaptability to different design intents \\(D\\)." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.254, + 0.482, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.491, + 0.482, + 0.504 + ], + "angle": 0, + "content": "Table 7. Ablation study on \\( \\mathcal{M} \\) accommodating small-scale LLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.483, + 0.575 + ], + "angle": 0, + "content": "PosterO dynamically adjusts the spatial distribution of elements in the generated layouts while maintaining remarkable graphic metrics. Notably, this capability is unique to PosterO, attributed to its intent-based paradigm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.701 + ], + "angle": 0, + "content": "Adaption to small-scale backbone LLM \\(\\mathcal{M}\\). To address resource-limited environments and to accelerate the layout generation process, we investigate PosterO's adaptability to small-scale LLMs [35], including Llama3.2-1B (L3.2-1B) and Llama3.2-3B (L3.2-3B). As reported in Tab. 7, PosterO maintains very competitive and or even better results with the smaller \\(\\mathcal{M}\\), showcasing its potential for more flexible and efficient applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.483, + 0.729 + ], + "angle": 0, + "content": "4.6. Generalized Content-aware Layout Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To verify PosterO's potential in generalized content-aware layout generation, which is presented for the first time in this work, we built a new dataset, PStylish7. It comprises 152 few-shot learning samples and 100 test images, covering seven poster categories. Each serves as a representative prototype for a segment in the spectrum of poster design purposes, encompassing artwork exhibition (Paint), cultural education (Poem), public safety (Metro), entertainment marketing (Movie), merchandising display (Menu), public advocacy (Animal), and social-media interaction (Instagram). Besides, it covers eight element types, including" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.235 + ], + "angle": 0, + "content": "
OursCLPaintPoemMetroMovieMenuAnimalInsta.
Ove ↓0.00470.00340.06180.01060.00830.00320.0010
Int ↓0.76900.21980.69870.12160.23590.30200.2674
Sal ↓3.88380.84841.36040.32750.71912.62800.6652
Avg ↓1.55250.35720.70700.15320.32110.97770.3112
OursL3PaintPoemMetroMovieMenuAnimalInsta.
Ove ↓0.00320.00280.07770.00040.00560.14950.0004
Int ↓0.62190.62740.74780.33750.16200.53831.1431
Sal ↓2.45120.16001.11110.13470.57902.92841.0810
Avg ↓1.02540.26340.64550.15750.24891.20540.7415
" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.24, + 0.857, + 0.254 + ], + "angle": 0, + "content": "Table 8. Quantitative results on PStylish7 dataset." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.326, + 0.856, + 0.34 + ], + "angle": 0, + "content": "Figure 8. Visualized results on PStylish7 dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.905, + 0.424 + ], + "angle": 0, + "content": "all those in the existing datasets and four text variants depicted in Sec. 3.1. Considering feasibility, we employ five metrics, including \\(Ove\\downarrow\\), \\((Cov\\uparrow, Con\\downarrow)\\), standardized as \\(Int\\downarrow\\), and \\((Uti\\uparrow, Occ\\downarrow)\\), standardized as \\(Sal\\downarrow\\). More details about PStylish7 are given in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.59 + ], + "angle": 0, + "content": "The quantitative results are reported in Tab. 8, and the visualized results are presented in Fig. 8. Here, the detection model \\( S_{pc} \\) is employed to obtain design intents \\( D \\) and their features \\( f \\). While PosterO\\(_{L3}\\) outperforms PosterO\\(_{CL}\\) in most categories, as the first attempt in this field, there is still plenty of room for improvement. Overall, the proposed PStylish7 is significantly more challenging than the existing datasets, attributed to its broader and sparser data distribution. This requires approaches with good generalization capabilities, mirroring real-world scenarios, and can stimulate more advanced work in the future." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.763, + 0.619 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work presented a tuning-free language model-based approach for content-aware layout generation, named PosterO. Through extensive experiments, we demonstrated that PosterO achieves new state-of-the-art results on various benchmarks, mainly attributed to the structural representation, layout trees. Moreover, we delved into the task under generalized settings and built the first available dataset, PStylish7, which covers seven poster purposes and eight element types. Employing PStylish7, we verified PosterO's potential to solve complicated tasks in real-world scenarios, underscoring its practicality. The future work mainly lies in two aspects. First, enhancing the generalized capabilities of layout generation approaches is highly valuable, as PosterO aims. With the challenging test offered by the proposed PStylish7 dataset, it is increasingly attainable. Second, integrating feedback mechanisms is envisioned to facilitate self-optimization and allow the incorporation of user suggestions, forming a more interactive, user-centric process." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8124" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.116, + 0.484, + 0.162 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (61925201, 62132001, 62432001) and Beijing Natural Science Foundation (L247006)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.177, + 0.188, + 0.193 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.202, + 0.484, + 0.258 + ], + "angle": 0, + "content": "[1] Diego Martin Arroyo, Janis Postels, and Federico Tombari. Variational transformer networks for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13642-13652, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[2] Canva. Visual economy report 2024 edition: How visual communication fuels new business opportunities. https://www.canva.com/visual-economy-report/, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[3] Canva. 10 highlights from canva create 2024: Redesigned for work. https://www.canva.com/newsroom/news/what-happened-at-canva-create-2024/, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[4] Yunning Cao, Ye Ma, Min Zhou, Chuanbin Liu, Hongtao Xie, Tiezheng Ge, and Yuning Jiang. Geometry aligned variational transformer for image-conditioned layout generation. In Proceedings of the ACM International Conference on Multimedia, pages 1561-1571, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[5] Yunning Cao, Chuanbin Liu, Ye Ma, Min Zhou, Tiezheng Ge, Yuning Jiang, and Hongtao Xie. Self-refined variational transformer for image-conditioned layout generation. International Journal of Machine Learning and Cybernetics, pages 1-18, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[6] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18349-18358, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[7] Shang Chai, Liansheng Zhuang, Fengying Yan, and Zihan Zhou. Two-stage content-aware layout generation for poster designs. In Proceedings of the ACM International Conference on Multimedia, pages 8415-8423, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[8] Jian Chen, Ruiyi Zhang, Yufan Zhou, Jennifer Healey, Jiuxiang Gu, Zhiqiang Xu, and Changyou Chen. Textlap: Customizing language models for text-to layout planning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing: Findings, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[9] Erik Dahlstrom, Anthony Grasso, Niklas Hagelroth, Chris Lilley, Cameron McCormack, and Doug Schepers. Scalable vector graphics (svg) 2.0 specification. SVG working group, w3c. https://dev.w3.org/SVG/profiles/2.0/publish/, 2002. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[10] Leichao Dai, Lin Feng, Xinglin Shang, and Han Su. Cross modal adaptive few-shot learning based on task dependence. Chinese Journal of Electronics, 32(1):85-96, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[12] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.202, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Proceedings of the International Conference on Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[13] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. In Proceedings of the International Conference on Neural Information Processing Systems, pages 50742-50768, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[14] Shunan Guo, Zhuochen Jin, Fuling Sun, Jingwen Li, Zhaorui Li, Yang Shi, and Nan Cao. Vinci: an intelligent graphic design system for generating advertising posters. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1-17, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[15] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004–1014, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[16] Sun Haoran, Wang Yang, Liu Haipeng, and Qian Biao. Fine-grained cross-modal fusion based refinement for text-to-image synthesis. Chinese Journal of Electronics, 32(6): 1329-1340, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[17] Daichi Horita, Naoto Inoue, Kotaro Kikuchi, Kota Yamaguchi, and Kiyoharu Aizawa. Retrieval-augmented layout transformer for content-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 67-76, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[18] HsiaoYuan Hsu, Xiangteng He, and Yuxin Peng. Densitylayout: Density-conditioned layout gan for visual-textual presentation designs. In Proceedings of the International Conference on Image and Graphics, pages 187-199, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[19] HsiaoYuan Hsu, Xiangteng He, Yuxin Peng, Hao Kong, and Qing Zhang. Posterlayout: A new benchmark and approach for content-aware visual-textual presentation layout. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6018-6026, 2023. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[21] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Layoutdm: Discrete diffusion model for controllable layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10167-10176, 2023. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.907, + 0.87 + ], + "angle": 0, + "content": "[22] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the International Conference on Machine Learning, pages 4904-4916, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[23] Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Constrained graphic layout generation via la" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8125" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "tent optimization. In Proceedings of the ACM International Conference on Multimedia, pages 88-96, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[24] Suhas Kotha, Jacob Mitchell Springer, and Aditi Raghunathan. Understanding catastrophic forgetting in language models via implicit inference. In The Twelfth International Conference on Learning Representations, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[25] Hsin-Ying Lee, Lu Jiang, Irfan Essa, Phuong B Le, Haifeng Gong, Ming-Hsuan Yang, and Weilong Yang. Neural design network: Graphic layout generation with constraints. In Proceedings of the European Conference on Computer Vision, pages 491-506, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[26] Chenhui Li, Peiying Zhang, and Changbo Wang. HarmOnious textual layout generation over natural images via deep aesthetics learning. IEEE Transactions on Multimedia, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[27] Fengheng Li, An Liu, Wei Feng, Honghe Zhu, Yaoyu Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junjie Shen, Zhangang Lin, et al. Relation-aware diffusion model for controllable poster layout generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 1249-1258, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[28] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. LayoutGAN: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7): 2388-2399, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[29] Jianan Li, Jimei Yang, Jianming Zhang, Chang Liu, Christina Wang, and Tingfa Xu. Attribute-conditioned layout GAN for automatic graphic design. IEEE Transactions on Visualization and Computer Graphics, 27(10):4039-4048, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[30] Yinan Li, Jia Chen, Yin Bai, Jia Cheng, and Jun Lei. Design element aware poster layout generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 1296-1305, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[31] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. Layoutprompt: Awaken the design ability of large language models. In Proceedings of the International Conference on Neural Information Processing Systems, pages 43852-43879, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[32] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the ACM International Conference on Multimedia, pages 1250-1260, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[33] Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. arXiv preprint arXiv:2308.08747, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[34] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. In Proceedings of the International Conference on Neural Information Processing Systems, pages 1877-1901, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Meta. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/, 2024. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[36] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In Proceedings of the International Conference on Neural Information Processing Systems, pages 27730-27744, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[38] Yi-Hao Peng, Faria Huq, Yue Jiang, Jason Wu, Amanda Xin Yue Li, Jeffrey Bigham, and Amy Pavel. Dreamstruct: Understanding slides and user interfaces via synthetic data generation. In Proceedings of the European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[39] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7479-7489, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[40] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In Proceedings of the European Conference on Computer Vision, pages 38-56, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Intervention, pages 234-241, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[42] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[43] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In Proceedings of the International Conference on Neural Information Processing Systems, pages 25278–25294, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[44] Jaejung Seol, Seojun Kim, and Jaejun Yoo. Posterllama: Bridging design ability of language model to content-aware layout generation. In Proceedings of the European Conference on Computer Vision, pages 451-468, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[45] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF WInter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. Lay-outnuwa: Revealing the hidden layout expertise of large language models. Proceedings of International Conference on Learning Representations, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8126" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[48] Bo Wang, Quan Chen, Min Zhou, Zhiqiang Zhang, Xiaogang Jin, and Kun Gai. Progressive feature polishing network for salient object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12128-12135, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[49] Shaodong Wang, Yunyang Ge, Liuhan Chen, Haiyang Zhou, Qian Wang, Xinhua Cheng, and Li Yuan. Prompt2poster: Automatically artistic chinese poster creation from prompt only. In Proceedings of the ACM International Conference on Multimedia, pages 10716-10724, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[50] Haohan Weng, Danqing Huang, Yu Qiao, Zheng Hu, ChinYew Lin, Tong Zhang, and CL Chen. Desigen: A pipeline for controllable design template generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12721-12732, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.456 + ], + "angle": 0, + "content": "[51] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Proceedings of the International Conference on Neural Information Processing Systems, pages 12077-12090, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.459, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[52] Chenchen Xu, Min Zhou, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10114-10123, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.53, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[53] Tao Yang, Fan Wang, Junfan Lin, Zhongang Qi, Yang Wu, Jing Xu, Ying Shan, and Changwen Chen. Toward human perception-centric video thumbnail generation. In Proceedings of the ACM International Conference on Multimedia, pages 6653-6664, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.725 + ], + "angle": 0, + "content": "[55] Min Zhou, Chenchen Xu, Ye Ma, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Composition-aware graphic layout GAN for visual-textual presentation designs. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 4995-5001, 2022. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.727, + 0.482, + 0.782 + ], + "angle": 0, + "content": "[56] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8127" + } + ] +] \ No newline at end of file diff --git a/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/28953876-24ed-4c02-8fbd-02d889f7bc8b_origin.pdf b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/28953876-24ed-4c02-8fbd-02d889f7bc8b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..39c006e8870c2d3593e3aee7c6c87b10e1aac8ce --- /dev/null +++ b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/28953876-24ed-4c02-8fbd-02d889f7bc8b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c346d3e9807af8d1d62d8caf18ba74543bc2a1d3d804fec2f02b59f081e75694 +size 4185112 diff --git a/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/full.md b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fdbbc2d6ee63f748657483aa3a5decf597646aa2 --- /dev/null +++ b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/full.md @@ -0,0 +1,313 @@ +# PosterO: Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation + +HsiaoYuan Hsu and Yuxin Peng* +Wangxuan Institute of Computer Technology, Peking University + +kslh99@stu.pku.edu.cn, pengyuxin@pku.edu.cn + +![](images/238b9e451c661a20780a71bf7fd062a0ca0425da339f0ba93cd0285c85253502.jpg) +Figure 1. Illustration of the generalized settings in content-aware layout generation from both (a) data and (b) approach perspectives. + +![](images/8b35deafe8e070892b506f38ac775a00bfe3d22452e3fb9c41cb329ce9c675e1.jpg) + +# Abstract + +In poster design, content-aware layout generation is crucial for automatically arranging visual-textual elements on the given image. With limited training data, existing work focused on image-centric enhancement. However, this neglects the diversity of layouts and fails to cope with shape-variant elements or diverse design intents in generalized settings. To this end, we proposed a layout-centric approach that leverages layout knowledge implicit in large language models (LLMs) to create posters for omnifarious purposes, hence the name PosterO. Specifically, it structures layouts from datasets as trees in SVG language by universal shape, design intent vectorization, and hierarchical node representation. Then, it applies LLMs during inference to predict new layout trees by in-context learning with intent-aligned example selection. After layout trees are generated, we can seamlessly realize them into poster designs by editing the chat with LLMs. Extensive experimental results have demonstrated that PosterO can generate visually appealing layouts for given images, achieving new state-of-the-art performance across various benchmarks. To further explore PosterO's abilities under the generalized settings, we built PStylish7, the first dataset with multi-purpose posters and various-shaped elements, further offering a challenging test for advanced research. Code and dataset will be publicly available at https://thekinsley.github.io/PosterO/. + +# 1. Introduction + +Content-aware layout generation makes the poster design process intelligent by automatically arranging visual-textual elements on given images. As the demand for effective design grows rapidly, a mainstream platform surpasses 185 million monthly users [3], and $90\%$ of survey participants take AI-empowered features as key considerations [2]. This technology is gaining more attention for its practical value in various applications [32, 49, 50, 53], as shown in Fig. 1. + +Unlike other generation tasks [22, 43], collecting image-Layout pairs for the target task requires costly annotation and inpainting processes [19, 55]. Therefore, existing work has incorporated object hints [4, 55], prior information [7, 19], retrieved references [17, 31], and augmented data [44] to confront the challenge of training data scarcity. However, because these are mainly image-centric enhancements without increasing the adaptability and diversity of layouts, they are prone to get trapped in the local solution space [18]. On the other hand, while leveraging related knowledge implicit in large language models (LLMs) [11, 37, 42, 47] showed the potential for layout-centric enhancements [12, 46], existing work [31, 44] has suffered from the semantically impoverished layout representations. The two primary issues are the monotonous rectangular shape of elements and the isolated visual constraints, which led to failures at capturing the complexity of real-world layouts and the intricate visual relationship between images and elements. + +To pursue data efficiency and semantic richness required for advances in content-aware layout generation, we proposed PosterO to enable in-context learning (ICL) of LLMs by structuring layout tree representations. Specifically, it consists of three procedures, including (a) layout tree construction that jointly represents layout elements and design intents (i.e., available areas on images) as hierarchical nodes in SVG trees [9], (b) layout tree generation that dynamically selects intent-aligned learning examples for the test input to facilitate ICL, and (c) poster design realization that integrates design materials into the generated layout trees to create posters in the subsequent chats with LLMs. + +We conducted extensive experiments on public benchmarks [19, 55] and demonstrated that PosterO has achieved new state-of-the-art performance. It particularly shows stability in managing domain adaption problems [52] and spatial distribution shifts [17] that make the current approaches stumped. In additional efforts, we enhanced the evaluation process through newly added intent-aware content metrics that are more credible than currently used saliency-aware ones. We also demonstrated the adaptability of PosterO to different design intents and small-scale LLMs [35], highlighting its applicability. To further explore its potential under the generalized settings of diverse poster purposes (e.g., commercial or education) and shape-variant elements (e.g., circles or curves), we built the first dataset, PStylish7, consisting of 152 few-shot learning samples and 100 test images, covering seven representative poster purposes and eight element types. Serving as a challenging test, it is believed to facilitate advanced research in this field. + +We summarize the contribution of this paper as follows: + +- A language-based content-aware layout representation, layout tree, uses hierarchical nodes to model both layout elements and design intents with arbitrary shapes. +- A layout-centric approach, PosterO, leverages implicit knowledge of LLMs by applying a few intent-aligned layout trees as examples to perform in-context learning. +- Extensive evaluations and newly added metrics verify the superior effectiveness of PosterO in confronting various adaptation problems that plague current approaches. +- A dataset for generalized content-aware layout generation, PStylish7, incorporates seven representative poster purposes and eight shape-variant element types. + +As our best knowledge, this work is the first to consider the generalized settings of content-aware layout generation. + +# 2. Related Work + +# 2.1. Content-aware Layout Generation + +Apart from general layout tasks [1, 15, 23, 25, 28], content-aware layout generation is a cross-modal task [10, 16] integrating visual conditions, i.e., image canvases, which are inherent in various real-world applications [32, 49, 50, 53], + +highlighting its practical utility. With advances in deep generative models, data-centric approaches have become dominant in this field, categorized into GAN-based [18, 19, 52, 55], auto-regression-based [4, 5, 17, 30], and diffusion model-based [7, 27] approaches. + +Entering into data construction of image layout pairs, CGL-GAN [55] leveraged an inpainting model [45] to remove design elements from posters and a saliency object detection model [48] to give generative networks additional hints about image compositions. Similar saliency-enhanced paradigm [26, 39, 40] has been adopted by most subsequent approaches. Dealing with limited training data, DS-GAN [19] introduced prior knowledge [14, 29] to arrange layout elements in a meticulously designed order, making the patterns existing in the data more efficiently mined. RALF introduced retrieval augmentation, leveraging a composition-aware similarity evaluation model [13] to find the nearest neighbors for a given image, and incorporating their layout features [23] as additional inputs for the generative network. Moreover, Chai et al. [7] adapted a general layout diffusion model [6] for content-aware tasks by incorporating aesthetic constraints and a saliency-aware layout plausibility ranker. + +As current approaches are engaged in image-centric enhancement yet seldom expand the diversity of layouts, the accessible solution space is narrowed, impeding their ability to address spatially varying design intents across different applications. In light of this, we propose a layout-centric, intent-aware approach in this work. + +# 2.2. Language Model-based Layout Generation + +While numerical layout representations risk losing the semantic information of attributes [46], current advances in large language models (LLMs) [11, 42, 47] open up new possibilities with HTML-based [31, 38], CSS-based [8, 12], and SVG-based [44, 46] language representations as these approaches enable leverage of layout knowledge implicit in LLMs. However, due to the challenge of guiding LLMs to perceive visual content, only a few approaches [31, 44] have supported content-aware layouts. + +LayoutPrompter [31] represented each layout element as a
in HTML. While adopting a saliency paradigm, it only extracted the minimum bounding rectangle from the binary map as coarse content constraints and then retrieved layout examples for in-context learning. In contrast, PosterLlama [44] adopted a SVG-based representation [46] using shapes. It then fine-tuned DINOv2 [36, 56] with CodeLlama [20, 42] enhanced by data augmentation [54]. + +As current approaches employed monotype element representations and isolated visual representations, they failed to exploit the rich semantics of languages nor explicitly capture the relationship between elements and images. In light of this, we propose layout trees to represent various-shaped elements and visual hints jointly. + +![](images/4074032f7c66cdc63cb79a0ec59d251fa818eeb5e0b4a7f022598b817e7e2e78.jpg) +Figure 2. An overview of PosterO. First, (a) takes image layout pairs $(I,L)$ from datasets as input for data preparation, jointly modeling various-shaped elements $E$ and design intents $D$ towards layout trees $T$ and building up the latent space of $D$ . Then, (b) takes test images $I_{t}$ as input and searches examples based on the predicted design intents $f_{t}$ to apply LLM $\mathcal{M}$ through in-context learning. After obtaining generated layout trees $\hat{T}$ , (c) can continue the conversation with $\mathcal{M}$ to create poster designs seamlessly. + +![](images/ac9e0efde70344e1d54552de7f293726700faa26298f1236c30c677955e5e48d.jpg) + +![](images/fa93599ed41f955c2539228a3dcb175333db26475495f61243acb43b2e9672e2.jpg) + +# 3. The Proposed Approach: PosterO + +Considering the limitations of image-centric and language model-based approaches, we propose PosterO. It is layout-centric and enables large language models (LLMs) to generate visually appealing layouts for given images by structurally representing the content-aware layouts as trees $T$ in SVG language. An overview of PosterO is shown in Fig. 2. In brief, it is composed of three procedures, namely, (a) layout tree construction for data preparation, (b) layout tree generation applying an LLM $\mathcal{M}$ through $k$ -shot in-context learning, and (c) poster design realization through the subsequent chat with $\mathcal{M}$ . + +# 3.1. Layout Tree Construction + +# 3.1.1. Universal Shape Vectorization + +In content-aware layout generation, each data pair contains an image $I$ and a layout $L$ . Usually, $L$ is represented as a set of numerical tuples $\{e_i\}_{i}^n = \{(c_i, b_i)\}_{i}^n$ , where $c_i$ and $b_i = (x, y, w, h)$ indicate the category and bounding box of the $i$ -th element $e_i$ out of all $n$ elements. As this is adequate only for rectangular elements, a more generalized representation becomes essential for the diverse shapes encountered in real-world posters, such as circles and curves. In light of this, SVG (Scalable Vector Graphics) emerges as an ideal choice because of its ability to represent arbitrary shapes and, more importantly, its nature as a language to facilitate the layout knowledge implicit in LLMs. + +Following the published SVG standards [9], we define five shapes to encompass a wide range of elements typically found in posters, as shown in Fig. 3. First, (a) regular rectangle is vectorized using $\langle \text{rect} \rangle$ . Also, its two variants, including (b) vertical rectangle sensitive to the aspect ratio and (c) rotated rectangle using the rotate function in the transform attribute to indicate the rotation degree. + +![](images/8f213634ecc780f05c86c798374b0ab2428589bc8dd2eb5cc551cee6a7b356cc.jpg) +(a) General rectangle + +![](images/c413ee1a7ab9f381a74b21df64c809230c402e0b05fda8f0b891e7e0b433867f.jpg) +(b) Vertical rectangle + +![](images/33abfb8ceac90a7d23ab9b60c19ed372a1343ea3e2b9cdf8d8b64a75566a21f0.jpg) +(c) Rotated rectangle + +![](images/4c0796c747d085794b6628493d16236b254d117e40b0ca532761ebb144d50ea9.jpg) +(d) Ellipse + +![](images/a9cf75fab83f5a938aed6f09c77c20083ae463ec281d4b15b949e705faf7574a.jpg) +(e) Complex curve + +![](images/8929c0e5cea4c1dc80a98fe81e10dd56c53a870ae6a2042e82e622f4e52adc76.jpg) +(f) Polygon +Figure 3. Various shapes of elements in PStylish7 dataset. + +When delving into more intricate shapes, (d) ellipse is vectorized using , which can also handle the circle as a special case, and (e) complex curve is vectorized using approximated by multiple cubic Bezier curves. + +# 3.1.2. Design Intent Vectorization + +In addition to visual-textual elements, the design intents associated with the layout $L$ must also be vectorized. Specifically, design intents $D$ are embodied by a set of areas within the image $I$ , suitable or intended for placing the elements. By incorporating these areas' representations $N_{D}$ , the layout tree becomes content-aware and capable of conveying visual relationships between elements and images to LLMs, which cannot 'see' the images. As $D$ can be learned from existing data or customized case by case, this also provides a flexible interface for indicating various purposes. + +By leveraging image layout pairs from posters with a similar purpose, a design intent detection model $S$ based on U-Net [41] is trained in a semi-supervised manner. Referring to [18], $S$ takes the bitmap of layout elements as the + +supervision signal during the optimization, while the consistency loss is canceled to clearly segment out the areas. With the pre-trained $S$ , PosterO can obtain design intents $D$ aligning the target purpose for any image $I$ . Specifically, after thresholding and morphological operations, we approximate the contour of the predicted intent maps $m$ and vectorize the obtained irregular areas using in SVG language. An example is shown in Fig. 3(f). Besides, the intermediate embeddings $f = S_{\mathfrak{e}}(I)$ , where $S_{\mathfrak{e}}$ denotes the last block of the encoder, are also collected to serve as their representations in the latent space of design intents $\mathbf{Z}_D$ . + +# 3.1.3. Hierarchical Node Representation + +As the nodes of layout elements $N_{E}$ and design intents $N_{D}$ are in place, the layout tree $T$ is constructed and encapsulated within an SVG container aware of the image resolution $(w_{I}, h_{I})$ , as: $<\mathrm{svg~width}=\{"w_{I}\}" \mathrm{height}=\ {"h_{I}\}" \mathrm{xmlns}=\text{"http://www.w3.org/2000/sv"}> \{N_{D}\} \{N_{E}\}<\mathrm{svg}>$ . $N_{D}$ is put ahead to ensure that $N_{E}$ is arranged considering their probabilistic dependencies in the autoregression, thus generating suitable content-aware layouts. In addition, as the special elements that wrap around others, such as underlays, are usually used in poster design, a novel hierarchical representation explicitly modeling these enveloping properties is proposed. Specifically, $N_{E}$ is sorted by prioritizing underlays with larger areas. Then, given a node $N_{a}$ and the corresponding nodes $N_{B} = \{N_{b} \mid N_{b} \in \{N_{E}\}_{a + 1}^{n} \land C(N_{a}, N_{b})\}$ , where the condition function $C(\cdot)$ is defined as: + +$$ +\begin{array}{l} C (\cdot) = \left(| N _ {a}. x - N _ {b}. x | \leq \epsilon\right) \wedge \left(| N _ {a}. y - N _ {b}. y | \leq \epsilon\right) \\ \wedge \left(\left| N _ {a}. x + N _ {a}. w - N _ {b}. x - N _ {b}. w \right| \leq \epsilon\right) \tag {1} \\ \wedge \left(\left| N _ {a}. y + N _ {a}. h - N _ {b}. y - N _ {b}. h \right| \leq \epsilon\right), \\ \end{array} +$$ + +a subtree is constructed as $<\text{svg} \times=" \{N_{a}.x\}" \text{y="} \{N_{a}.y\}">$ $\{N_{a}'\} \{N_{B}'\} < / \text{svg}>$ and inserted at the position of $N_{a}$ , where the processed $N_{i}'$ has the relative geometry attributes $(N_{i}'.x, N_{i}'.y) = (N_{i}.x - N_{a}.x, N_{i}.y - N_{a}.y)$ . All leaf nodes of $T$ are assigned unique identifiers $\mathfrak{id} = "\{c_{i} - i\}"$ to facilitate precise manipulation and querying. + +# 3.2. Layout Tree Generation + +# 3.2.1. Intent-aligned Example Selection + +Upon findings from [34], LLMs are capable of in-context learning (ICL), requiring only a few examples to perform downstream tasks. This significantly reduces the need for task-specific data and avoids the extensive costs or catastrophic forgetting [24, 33] associated with fine-tuning. To make the best use of layout knowledge implicit in LLMs during ICL, selecting examples that align with the design intent is crucial. Therefore, given the test image $I_{t}$ , its embedding $f_{t}$ is first extracted by the detection model $\mathcal{S}$ ; subsequently, $k$ layout trees $\{T_j\}_j^k$ are selected as in-context examples, where their embeddings $\{f_j\}_j^k$ are the $k$ closest to $f_{t}$ in the latent space $\mathbf{Z}_{D}$ . + +# 3.2.2. In-context Learning for Layout Generation + +To provide clear structure and objectives in the prompt $P$ for the LLM $\mathcal{M}$ , each intent-aligned example $T_{j}$ is preceded by a short sentence that informs its number $j$ , the resolution of its corresponding image, design intents $D$ , and identifiers of its nodes, as shown in Fig. 2(b). The example field is then encapsulated within a preface and postscript, with the postscript delivering the request message of the test image $I_{t}$ through a short sentence mirroring that of the example. Complete prompts $P$ are presented in supplementary material. Finally, by inputting $P$ into $\mathcal{M}$ and interpreting the responses, PosterO completes the layout generation process. + +# 3.3. Poster Design Realization + +Benefiting from the essence of the generated layout tree $\hat{T}$ as graphics and the rigid id assignment for elements, guiding $\mathcal{M}$ to transform $\hat{T}$ into a practical poster design is feasible and straightforward. Specifically, the transformation contains mockup creation, where shape elements are converted to or regarding their categories, and material synthesis, where elements are filled with actual design contents, e.g., slogans and hyperlinks to images. Notably, this transformation is achieved with zero-shot learning, yet $\mathcal{M}$ brings impressive results (in supplementary material) and even predicts suitable font sizes for texts, highlighting the value of design knowledge implicit in LLMs. + +# 4. Experiments + +# 4.1. Datasets + +To evaluate the proposed PosterO, two public e-commerce poster datasets are used, while their train+valid/annotated test/unannotated test splits are allocated as in [17] to ensure a fair comparison with existing work. Specifically, PKU PosterLayout [19] comprises 8,734/1,000/905 images with three layout element types, including text, logo, and underlay; CGL [55] comprises 54,546/6,002/1,000 images with four layout element types, including text, logo, underlay, and embellishment. Furthermore, to explore PosterO's abilities under the generalized settings, we built the first multipurpose poster dataset, PStylish7, containing layout elements of various shapes. More details about PStylish7 are depicted in Sec. 4.6 and supplementary material. + +# 4.2. Evaluation Metrics + +Following previous work [17, 19, 55], we use graphic and content metrics. Graphic metrics include overlay $Ove\downarrow$ alignment $Ali\downarrow$ , and loose, strict underlay effectiveness $Und_{l}\uparrow ,Und_{s}\uparrow ;$ Content metrics include non-salient space utilization $Uti\uparrow$ , salient space occlusion $Occ\downarrow$ , and readability $Rea\downarrow$ . Moreover, we newly proposed two content metrics, design intent coverage $Cov\uparrow$ and conflict $Con\downarrow$ based on a design intent detection model $S_{pc}$ that is compre + +
MethodOve ↓Ali ↓Undl ↑Unds ↑Int ↓Sal ↓Rea ↓Avg ↓
Annotated Test Split
Real data0.00100.00380.99420.99030.02890.03680.01090.0138
CGL-GAN0.09660.00350.78540.35700.12040.27740.01910.1964
DS-GAN0.02610.00380.83500.58040.06770.31000.01990.1446
ICVT0.25720.04050.53840.39320.78311.08360.02590.4655
LayoutDM†0.15620.00180.64260.38730.50660.24290.01850.2709
RALF0.00840.00280.98080.92010.11620.04790.01280.0410
Tuning-required
PosterLlama‡0.00080.00060.99990.99820.34070.26190.01770.0891
LLM-based
Tuning-free
LayoutPrompterL20.00830.00300.37160.16330.32000.94330.03190.3959
LayoutPrompterCL0.00400.00300.44150.19370.36620.90350.03010.3817
LayoutPrompterL30.00170.00280.40850.16130.51990.85670.03090.4060
PosterOL2 (Ours)0.00040.00270.75970.56050.36790.14910.01880.1741
PosterOLCL (Ours)0.00040.00280.96210.86060.23970.22950.01890.0955
PosterOL3 (Ours)0.00060.00250.98880.92110.10800.11100.01710.0470
Unannoted Test Split
Non-LLMTrain data0.00120.00270.99690.99000.00000.00000.0108
CGL-GAN0.10100.00480.73260.27430.22750.77550.0327
DS-GAN0.02480.00460.78590.46760.16590.58680.0320
ICVT0.27860.04800.49390.35490.79471.32890.0347
LayoutDM†0.16380.00290.59870.36950.53570.31720.0264
RALF0.01750.00690.95480.86530.59860.17790.0231
Tuning-required
PosterLlama‡0.00060.00060.99860.99170.26670.38230.02850.0983
LLM-based
Tuning-free
LayoutPrompterL20.00950.00320.34010.11610.53491.20500.04080.4767
LayoutPrompterCL0.00320.00280.40520.14640.51481.11060.03950.4456
LayoutPrompterL30.00100.00260.40540.16210.70051.08380.04120.4660
PosterOL2 (Ours)0.00040.00310.72840.51880.40300.33510.02650.2173
PosterOLCL (Ours)0.00050.00280.95950.84510.27960.39450.02730.1286
PosterOL3 (Ours)0.00040.00250.98560.92410.14270.21310.02480.0677
+ +hensively pre-trained on [19] and [55]. Calculating $Cov\uparrow$ and $Con\downarrow$ is similar to $Uti\uparrow$ and $Occ\downarrow$ but using the predicted intent maps. As reported in Tab. 2, intent maps have doubled the matched portion with real layout data, reduced the unmatched portion by $34.7\%$ , and required only $1.3\%$ computational costs. These findings have demonstrated that intent-aware metrics are more credible and efficient. For a holistic understanding, we standardize intent-aware and saliency-aware metrics to $Int\downarrow$ and $Sal\downarrow$ , as follows: + +$$ +I n t = \Sigma (\frac {\left| C o v _ {\hat {L}} - C o v _ {L} \right|}{1 - C o v _ {L}}, \frac {\left| C o n _ {\hat {L}} - C o n _ {L} \right|}{C o n _ {L}}), \tag {2} +$$ + +$$ +S a l = \Sigma (\frac {| U t i _ {\hat {L}} - U t i _ {L} |}{1 - U t i _ {L}}, \frac {| O c c _ {\hat {L}} - O c c _ {L} |}{O c c _ {L}}), +$$ + +where $L$ and $\hat{L}$ are the layouts from the train split and generation. Finally, we introduce an average metric $Avg\downarrow = \Sigma (Ove,Ali,1 - Und_{l},1 - Und_{s},Int,Sal,Rea)$ . + +# 4.3. Implementation Details + +To ensure independent evaluations on [19] and [55], we correspondingly train two detection models, $S_{p}$ and $S_{c}$ , with mean squared error loss and a learning rate of $1e^{-6}$ . Each + +![](images/da94c947185bbe62879a25a49fd895c465362ebb3ecdadd71315ebb59ea20ca9.jpg) +Figure 4. Comparisons of visualized results on the unannotated test split of PKU PosterLayout dataset. + +Table 1. Comparisons of quantitative results on PKU PosterLayout dataset. + +
PredictionPKUCGL
Match ↑Unmatch ↓Match ↑Unmatch ↓
Saliency map0.22380.11930.19840.1352
Design intent map0.44140.09960.40360.0883
Inference costNetwork architecture#FLOPs(G)#Params(M)
Saliency mapISNet + BASNet287.83131.11
Design intent mapUNet (MiT-B1)3.7416.43
+ +Table 2. Comparisons of the predicted saliency and design intent maps on the annotated test splits of PKU PosterLayout and CGL datasets, showing that intent-aware metrics are more credible. + +detection model is with a Transformer encoder tailored for semantic segmentation, MiT-B1 [51]. For in-context learning (ICL), we implement $\mathcal{M}$ with three open-source LLMs, including Llama 2-7B (L2) [47], CodeLlama-7B (CL) [42], and Llama 3.1-8B (L3) [11], with an unified sampling temperature of 0.7. Considering the context lengths and remaining consistent with LayoutPrompter [31], the example size $k$ is 5, 10, and 10, respectively. Still adhering to [31], a layout ranker is employed, while mIoU in the quality function measures the similarity between generated layouts and predicted design intents rather than real data, minimizing the risk of data leakage. + +
MethodOve ↓Ali ↓Undl ↑Unds ↑Int ↓Sal ↓Rea ↓Avg ↓
Annotated Test Split
Non-LLMReal data0.00030.00240.99630.98800.00880.00970.01190.0070
CGL-GAN0.22910.01230.64660.22810.70880.46270.02130.3656
DS-GAN0.04600.00220.90810.63080.39500.15390.01810.1537
ICVT0.24530.01790.51500.33260.36320.51390.02110.3305
LayoutDM†0.01840.00210.92160.81590.04110.02860.01370.0523
RALF0.00420.00240.99120.97560.03660.07240.01800.0238
Tuning-required PosterLlama‡0.00060.00060.99870.98900.13370.25590.01840.0602
LLM-basedTuning-free
LayoutPrompterL20.00870.00330.33690.14550.33620.82340.03270.3888
LayoutPrompterCL0.00370.00310.39650.16160.37180.77990.03170.3760
LayoutPrompterL30.00170.00300.38300.17400.55930.79570.03270.4051
PosterOL2 (Ours)0.00060.00230.78010.61500.25420.04670.01760.1323
PosterOLCL (Ours)0.00050.00230.96820.88680.12460.07000.01680.0513
PosterOL3 (Ours)0.00060.00220.98420.93400.03880.00260.01610.0203
Unannotated Test Split
Non-LLMTrain data0.00030.00270.99450.98580.00000.00000.01180.0049
CGL-GAN0.26680.03160.67740.16560.97072.39730.05120.6964
DS-GAN0.09910.01380.75660.28100.46072.27320.05260.5517
ICVT0.20450.10100.43570.25991.28132.65120.03970.7974
LayoutDM†0.07930.18220.63040.38531.50023.29060.06120.8711
RALF0.02730.01890.97560.93151.02371.67980.03970.4118
Tuning-required PosterLlama‡0.00140.00070.99710.97710.35102.61700.05550.4359
LLM-basedTuning-free
LayoutPrompterL20.00690.01030.28470.11261.02572.37020.06280.7255
LayoutPrompterCL0.00450.00240.32390.13411.58002.53800.06500.8188
LayoutPrompterL30.00260.00160.26930.11421.62492.41280.06440.8175
PosterOL2 (Ours)0.00050.00330.66630.47700.35221.23650.03820.3553
PosterOLCL (Ours)0.00030.00370.93690.80710.27711.32230.03790.2710
PosterOL3 (Ours)0.00040.00320.98160.90340.43281.24070.03650.2612
+ +Table 3. Comparisons of quantitative results on CGL dataset. + +![](images/0843d030f2c5bb261d80b85a93409ecdfc2aefd61ddae899d084d6fe9c13bf93.jpg) +Figure 5. Comparisons of visualized results on the unannotated test split on CGL dataset. + +# 4.4. Comparison with State-of-the-arts + +We select approaches with open-sourced implementation as baselines, including GAN-based CGL-GAN [55], DS-GAN [19], auto-regression-based ICVT [4], RALF [17], diffusion model-based LayoutDM† [21], and also LLM-based PosterLlama‡ [44], LayoutPrompter [31]. + +Baseline comparison. Tab. 1 reports the quantitative results on PKU PosterLayout. The proposed PosterO shows the best overall performance across the test splits. Particularly, when confronting the challenging unannotated split that poses domain adaptation problems [52], it still achieves new state-of-the-art (SOTA) results with outstanding stability. Compared to the current non-LLM and LLM-based SOTA approaches, i.e., RALF and PosterLlama, ours significantly improves the average metric $Avg\downarrow$ by $52.8\%$ and $31.1\%$ . Consistent results are obtained on CGL, as reported in Tab. 3, ours improves $Avg\downarrow$ by $36.6\%$ and $40.1\%$ . + +More precisely, PosterO shows impressive graphic metrics (2nd column), closely resembling the distribution of real data, with only a slight deficit in the Ali $\downarrow$ metric on the + +unannotated split of CGL compared to LayoutPrompter. As for content metrics (3rdcolumn), although PosterO uses only a small design intent detection model $S$ , it achieves comparable or even superior results to the baselines that utilize larger saliency detection models [26, 40] or costly fine-tuned DINOv2 [36]. Especially, PosterO is best at dealing with the spatial shifts in subject distributions [17] present in the unannotated split of CGL, maintaining an overwhelming lead in content metrics. + +Backbone LLMs $\mathcal{M}$ . Comparing the performance of tuning-free approaches across various LLMs is crucial for realizing how effectively the layout knowledge implicit in $\mathcal{M}$ has been leveraged. As reported in Tab. 1 and Tab. 3, PosterO demonstrates overwhelming superiority over LayoutPrompter in all counterparts with the same $\mathcal{M}$ . More importantly, from Poster $_{L2,CL}$ , to Poster $_{L3}$ , it is evident that the proposed approach has experienced improvements as the capabilities of its based $\mathcal{M}$ increased. This finding indicates that new knowledge, such as programming, introduced during model updates, has indeed been leveraged to generate layouts. With the continuous development of LLMs, PosterO is anticipated to keep its rising potential. + +
Method \ k1257810
LayoutPrompter0.77110.66900.53530.48830.49190.4660
PosterO (Ours)0.06780.07460.07320.06630.06510.0677
+ +![](images/d8c86786cb1843b018447bc051224241845c9ffa14fb8b599b5f6591da0101fa.jpg) +Figure 6. Content metrics over in-context example size $k$ . + +Visualized results. Fig. 4 and Fig. 5 show the layouts generated by different approaches. The results illustrate that PosterO generates high-quality layouts that the distribution of elements is generous and coordinated. In contrast, the approaches based on training or tuning tend to place elements centrally and thus fail to confront images with compositions that are less common in the training data. More visualized results are presented in supplementary material. + +In-context example size $k$ . To explore the impact of example selection during ICL, we test a set of candidates $\{k\} = \{1,2,5,7,8,10\}$ on the unannotated split of [19], as reported in Tab. 4. Amazingly, PosterO consistently outperforms LayoutPrompter and all other baselines in Tab. 1 across all settings of $k$ . As the optimal performance occurs at $k = 7$ and 8, it is proved that $\mathcal{M}$ has truly benefited from learning with the selected intent-aligned examples. To further observe how well these examples help $\mathcal{M}$ understand visual contents, which is the focal point of content-aware layout generation, we plot the resultant content metrics over $k$ in Fig. 6. It is noteworthy that directly adopting the most intent-aligned example (Top-1 Retrieval) surpasses all baselines, indicating a successful alignment between intent and layout in the latent space $\mathbf{Z}_D$ . Nevertheless, we still observe improvements when involving in-context examples. + +# 4.5. Ablation Study + +To investigate our design choice in layout tree construction (Sec. 3.1) and learning example selection (Sec. 3.2), as well as to explore PosterO's adaptability to customized design intents and small-scale LLMs, we conducted the following ablation studies, mainly on the unannotated split of [19]. + +Design choices for layout tree construction. We investigate the effectiveness of the design intent vectorization $(d)$ and hierarchical node representation $(h)$ during constructing layout trees, as reported in Tab. 5. Due to space constraints, + +Table 4. Comparisons of $Avg\downarrow$ over in-context example size $k$ . + +
dhUndl↑Unds↑Int↓Sal↓Rea↓Avg↓
✓✓0.98560.92410.14270.21310.02480.0677
✓0.86090.66500.09810.15600.02400.1078
✓0.94970.88360.98580.39910.02740.2260
0.76280.64130.81620.27290.02460.2447
+ +Table 5. Ablation study on the $(d)$ design intent vectorization and $(h)$ hierarchical node representation in layout tree construction. + +
SelectionUndl↑Unds↑Int↓Sal↓Rea↓Avg↓
f-aligned0.98560.92410.14270.21310.02480.0677
D-aligned0.98660.92970.22210.25310.02360.0836
E-aligned0.97430.90260.34310.41770.02640.1304
Random0.98850.93020.32250.46880.02950.1293
+ +Table 6. Ablation study on different references for in-context example selection. (f: Embeddings of design intents, $D$ : Bounding boxes of design intents, $E$ : Sets of element types.) + +$Ove\downarrow$ and $Ali\downarrow$ are omitted, considering their slight fluctuations. As observed, the combination of $d$ and $h$ yields the best results. Although removing $h$ appears to improve content metrics, this is primarily because the invalid empty underlays artificially inflate the utilization rate, as analyzed in supplementary material. The significant decreases witnessed in underlay effectiveness, $Und_{l}$ , $Und_{s}$ , by up to $28\%$ , further underscore this issue. In contrast, removing $d$ from both the examples and postscript makes LLMs blindly mimic the provided instances. Here, however, shows a $Sal\downarrow$ surpasses some baselines, such as CGL-GAN, DS-GAN, and LayoutPrompter, while $Int\downarrow$ fails far behind, reflecting the robustness of the newly proposed metric $Int\downarrow$ . If both $d$ and $h$ are removed, the most severe degradation is observed, up to $261\%$ . In this way, we demonstrate the necessity of these design choices. + +Design choices for in-context example selection. We have seen the remarkable performance of design intent embeddings $f$ -aligned example selection, which relies on the learned latent space $\mathbf{Z}_D$ . To investigate alternative references, we have considered bounding boxes of design intents ( $D$ -aligned), sets of element types ( $E$ -aligned), and even with no specific alignment (Random). As reported in Tab. 6, although all $D$ -aligned, $E$ -aligned, and random selections make declines in average metrics, the outcomes are still superior to most baseline approaches. In particular, using $D$ -aligned selection also achieves a new SOTA result, which supports its eligibility as an alternative enabling PosterO to deliver acceptable content-aware layouts when faced with roughly sketched $D$ from users. + +Adaption to customized design intent. Upon the discussions from the previous paragraph, we take a further step to verify PosterO's adaptability to customized design intents. Fig. 7 demonstrates the results of using the original $D$ predicted by the detection model $S$ and three customized $D$ with different levels of complexity. As the input $D$ varies, + +![](images/d451a8f3afa3c1033b231830c533ecafbe61454190ae7c2cd9faf8656f722e51.jpg) +Figure 7. PosterO's adaptability to different design intents $D$ . + +![](images/da9cf6671ae11f0d46753805b19dbb4605b4cfa06335b3229ca63045e6131088.jpg) +Table 7. Ablation study on $\mathcal{M}$ accommodating small-scale LLMs. + +PosterO dynamically adjusts the spatial distribution of elements in the generated layouts while maintaining remarkable graphic metrics. Notably, this capability is unique to PosterO, attributed to its intent-based paradigm. + +Adaption to small-scale backbone LLM $\mathcal{M}$ . To address resource-limited environments and to accelerate the layout generation process, we investigate PosterO's adaptability to small-scale LLMs [35], including Llama3.2-1B (L3.2-1B) and Llama3.2-3B (L3.2-3B). As reported in Tab. 7, PosterO maintains very competitive and or even better results with the smaller $\mathcal{M}$ , showcasing its potential for more flexible and efficient applications. + +# 4.6. Generalized Content-aware Layout Generation + +To verify PosterO's potential in generalized content-aware layout generation, which is presented for the first time in this work, we built a new dataset, PStylish7. It comprises 152 few-shot learning samples and 100 test images, covering seven poster categories. Each serves as a representative prototype for a segment in the spectrum of poster design purposes, encompassing artwork exhibition (Paint), cultural education (Poem), public safety (Metro), entertainment marketing (Movie), merchandising display (Menu), public advocacy (Animal), and social-media interaction (Instagram). Besides, it covers eight element types, including + +
OursCLPaintPoemMetroMovieMenuAnimalInsta.
Ove ↓0.00470.00340.06180.01060.00830.00320.0010
Int ↓0.76900.21980.69870.12160.23590.30200.2674
Sal ↓3.88380.84841.36040.32750.71912.62800.6652
Avg ↓1.55250.35720.70700.15320.32110.97770.3112
OursL3PaintPoemMetroMovieMenuAnimalInsta.
Ove ↓0.00320.00280.07770.00040.00560.14950.0004
Int ↓0.62190.62740.74780.33750.16200.53831.1431
Sal ↓2.45120.16001.11110.13470.57902.92841.0810
Avg ↓1.02540.26340.64550.15750.24891.20540.7415
+ +Table 8. Quantitative results on PStylish7 dataset. + +![](images/5d3404882d1ea2e5be60dd528fbd27c8d741a0bf8c203d39b15cfd1d0b60e398.jpg) +Figure 8. Visualized results on PStylish7 dataset. + +all those in the existing datasets and four text variants depicted in Sec. 3.1. Considering feasibility, we employ five metrics, including $Ove\downarrow$ , $(Cov\uparrow, Con\downarrow)$ , standardized as $Int\downarrow$ , and $(Uti\uparrow, Occ\downarrow)$ , standardized as $Sal\downarrow$ . More details about PStylish7 are given in supplementary material. + +The quantitative results are reported in Tab. 8, and the visualized results are presented in Fig. 8. Here, the detection model $S_{pc}$ is employed to obtain design intents $D$ and their features $f$ . While PosterO $_{L3}$ outperforms PosterO $_{CL}$ in most categories, as the first attempt in this field, there is still plenty of room for improvement. Overall, the proposed PStylish7 is significantly more challenging than the existing datasets, attributed to its broader and sparser data distribution. This requires approaches with good generalization capabilities, mirroring real-world scenarios, and can stimulate more advanced work in the future. + +# 5. Conclusion and Discussion + +This work presented a tuning-free language model-based approach for content-aware layout generation, named PosterO. Through extensive experiments, we demonstrated that PosterO achieves new state-of-the-art results on various benchmarks, mainly attributed to the structural representation, layout trees. Moreover, we delved into the task under generalized settings and built the first available dataset, PStylish7, which covers seven poster purposes and eight element types. Employing PStylish7, we verified PosterO's potential to solve complicated tasks in real-world scenarios, underscoring its practicality. The future work mainly lies in two aspects. First, enhancing the generalized capabilities of layout generation approaches is highly valuable, as PosterO aims. With the challenging test offered by the proposed PStylish7 dataset, it is increasingly attainable. Second, integrating feedback mechanisms is envisioned to facilitate self-optimization and allow the incorporation of user suggestions, forming a more interactive, user-centric process. + +# Acknowledgments + +This work was supported by the National Natural Science Foundation of China (61925201, 62132001, 62432001) and Beijing Natural Science Foundation (L247006). + +# References + +[1] Diego Martin Arroyo, Janis Postels, and Federico Tombari. Variational transformer networks for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13642-13652, 2021. 2 +[2] Canva. Visual economy report 2024 edition: How visual communication fuels new business opportunities. https://www.canva.com/visual-economy-report/, 2024.1 +[3] Canva. 10 highlights from canva create 2024: Redesigned for work. https://www.canva.com/newsroom/news/what-happened-at-canva-create-2024/, 2024. 1 +[4] Yunning Cao, Ye Ma, Min Zhou, Chuanbin Liu, Hongtao Xie, Tiezheng Ge, and Yuning Jiang. Geometry aligned variational transformer for image-conditioned layout generation. In Proceedings of the ACM International Conference on Multimedia, pages 1561-1571, 2022. 1, 2, 6 +[5] Yunning Cao, Chuanbin Liu, Ye Ma, Min Zhou, Tiezheng Ge, Yuning Jiang, and Hongtao Xie. Self-refined variational transformer for image-conditioned layout generation. International Journal of Machine Learning and Cybernetics, pages 1-18, 2024. 2 +[6] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18349-18358, 2023. 2 +[7] Shang Chai, Liansheng Zhuang, Fengying Yan, and Zihan Zhou. Two-stage content-aware layout generation for poster designs. In Proceedings of the ACM International Conference on Multimedia, pages 8415-8423, 2023. 1, 2 +[8] Jian Chen, Ruiyi Zhang, Yufan Zhou, Jennifer Healey, Jiuxiang Gu, Zhiqiang Xu, and Changyou Chen. Textlap: Customizing language models for text-to layout planning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing: Findings, 2024. 2 +[9] Erik Dahlstrom, Anthony Grasso, Niklas Hagelroth, Chris Lilley, Cameron McCormack, and Doug Schepers. Scalable vector graphics (svg) 2.0 specification. SVG working group, w3c. https://dev.w3.org/SVG/profiles/2.0/publish/, 2002. 2, 3 +[10] Leichao Dai, Lin Feng, Xinglin Shang, and Han Su. Cross modal adaptive few-shot learning based on task dependence. Chinese Journal of Electronics, 32(1):85-96, 2023. 2 +[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 2, 5 +[12] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and + +William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Proceedings of the International Conference on Neural Information Processing Systems, 36, 2024. 1, 2 +[13] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. In Proceedings of the International Conference on Neural Information Processing Systems, pages 50742-50768, 2024. 2 +[14] Shunan Guo, Zhuochen Jin, Fuling Sun, Jingwen Li, Zhaorui Li, Yang Shi, and Nan Cao. Vinci: an intelligent graphic design system for generating advertising posters. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1-17, 2021. 2 +[15] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004–1014, 2021. 2 +[16] Sun Haoran, Wang Yang, Liu Haipeng, and Qian Biao. Fine-grained cross-modal fusion based refinement for text-to-image synthesis. Chinese Journal of Electronics, 32(6): 1329-1340, 2023. 2 +[17] Daichi Horita, Naoto Inoue, Kotaro Kikuchi, Kota Yamaguchi, and Kiyoharu Aizawa. Retrieval-augmented layout transformer for content-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 67-76, 2024. 1, 2, 4, 6 +[18] HsiaoYuan Hsu, Xiangteng He, and Yuxin Peng. Densitylayout: Density-conditioned layout gan for visual-textual presentation designs. In Proceedings of the International Conference on Image and Graphics, pages 187-199, 2023. 1, 2, 3 +[19] HsiaoYuan Hsu, Xiangteng He, Yuxin Peng, Hao Kong, and Qing Zhang. Posterlayout: A new benchmark and approach for content-aware visual-textual presentation layout. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6018-6026, 2023. 1, 2, 4, 5, 6, 7 +[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2 +[21] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Layoutdm: Discrete diffusion model for controllable layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10167-10176, 2023. 6, 2 +[22] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the International Conference on Machine Learning, pages 4904-4916, 2021. 1 +[23] Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Constrained graphic layout generation via la + +tent optimization. In Proceedings of the ACM International Conference on Multimedia, pages 88-96, 2021. 2 +[24] Suhas Kotha, Jacob Mitchell Springer, and Aditi Raghunathan. Understanding catastrophic forgetting in language models via implicit inference. In The Twelfth International Conference on Learning Representations, 2024. 4 +[25] Hsin-Ying Lee, Lu Jiang, Irfan Essa, Phuong B Le, Haifeng Gong, Ming-Hsuan Yang, and Weilong Yang. Neural design network: Graphic layout generation with constraints. In Proceedings of the European Conference on Computer Vision, pages 491-506, 2020. 2 +[26] Chenhui Li, Peiying Zhang, and Changbo Wang. HarmOnious textual layout generation over natural images via deep aesthetics learning. IEEE Transactions on Multimedia, 2021. 2, 6 +[27] Fengheng Li, An Liu, Wei Feng, Honghe Zhu, Yaoyu Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junjie Shen, Zhangang Lin, et al. Relation-aware diffusion model for controllable poster layout generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 1249-1258, 2023. 2 +[28] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. LayoutGAN: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7): 2388-2399, 2020. 2 +[29] Jianan Li, Jimei Yang, Jianming Zhang, Chang Liu, Christina Wang, and Tingfa Xu. Attribute-conditioned layout GAN for automatic graphic design. IEEE Transactions on Visualization and Computer Graphics, 27(10):4039-4048, 2020. 2 +[30] Yinan Li, Jia Chen, Yin Bai, Jia Cheng, and Jun Lei. Design element aware poster layout generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 1296-1305, 2024. 2 +[31] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. Layoutprompt: Awaken the design ability of large language models. In Proceedings of the International Conference on Neural Information Processing Systems, pages 43852-43879, 2023. 1, 2, 5, 6 +[32] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the ACM International Conference on Multimedia, pages 1250-1260, 2023. 1, 2 +[33] Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. arXiv preprint arXiv:2308.08747, 2023. 4 +[34] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. In Proceedings of the International Conference on Neural Information Processing Systems, pages 1877-1901, 2020. 4 +[35] Meta. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/, 2024. 2, 8 + +[36] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 6 +[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In Proceedings of the International Conference on Neural Information Processing Systems, pages 27730-27744, 2022. 1 +[38] Yi-Hao Peng, Faria Huq, Yue Jiang, Jason Wu, Amanda Xin Yue Li, Jeffrey Bigham, and Amy Pavel. Dreamstruct: Understanding slides and user interfaces via synthetic data generation. In Proceedings of the European Conference on Computer Vision, 2024. 2 +[39] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7479-7489, 2019. 2 +[40] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In Proceedings of the European Conference on Computer Vision, pages 38-56, 2022. 2, 6 +[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Intervention, pages 234-241, 2015. 3 +[42] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1, 2, 5 +[43] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In Proceedings of the International Conference on Neural Information Processing Systems, pages 25278–25294, 2022. 1 +[44] Jaejung Seol, Seojun Kim, and Jaejun Yoo. Posterllama: Bridging design ability of language model to content-aware layout generation. In Proceedings of the European Conference on Computer Vision, pages 451-468, 2024. 1, 2, 6 +[45] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF WInter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 2 +[46] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. Lay-outnuwa: Revealing the hidden layout expertise of large language models. Proceedings of International Conference on Learning Representations, 2024. 1, 2 + +[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 5 +[48] Bo Wang, Quan Chen, Min Zhou, Zhiqiang Zhang, Xiaogang Jin, and Kun Gai. Progressive feature polishing network for salient object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12128-12135, 2020. 2 +[49] Shaodong Wang, Yunyang Ge, Liuhan Chen, Haiyang Zhou, Qian Wang, Xinhua Cheng, and Li Yuan. Prompt2poster: Automatically artistic chinese poster creation from prompt only. In Proceedings of the ACM International Conference on Multimedia, pages 10716-10724, 2024. 1, 2 +[50] Haohan Weng, Danqing Huang, Yu Qiao, Zheng Hu, ChinYew Lin, Tong Zhang, and CL Chen. Desigen: A pipeline for controllable design template generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12721-12732, 2024. 1, 2 +[51] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Proceedings of the International Conference on Neural Information Processing Systems, pages 12077-12090, 2021. 5 +[52] Chenchen Xu, Min Zhou, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10114-10123, 2023. 2, 6 +[53] Tao Yang, Fan Wang, Junfan Lin, Zhongang Qi, Yang Wu, Jing Xu, Ying Shan, and Changwen Chen. Toward human perception-centric video thumbnail generation. In Proceedings of the ACM International Conference on Multimedia, pages 6653-6664, 2023. 1, 2 +[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[55] Min Zhou, Chenchen Xu, Ye Ma, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Composition-aware graphic layout GAN for visual-textual presentation designs. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 4995-5001, 2022. 1, 2, 4, 5, 6 +[56] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 \ No newline at end of file diff --git a/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/images.zip b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..01ba64adf973105013924bcba71b391bbbe445e8 --- /dev/null +++ b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4146d7e5d49a0b9455895050b49ae675fca32ae6935784cf1c728cde01cc2fe6 +size 1112721 diff --git a/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/layout.json b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6201496886344223c6648f3e4335de9ca7f79ac4 --- /dev/null +++ b/2025/PosterO_ Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation/layout.json @@ -0,0 +1,10784 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 501, + 140 + ], + "type": "text", + "content": "PosterO: Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 161, + 460, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 161, + 460, + 190 + ], + "spans": [ + { + "bbox": [ + 149, + 161, + 460, + 190 + ], + "type": "text", + "content": "HsiaoYuan Hsu and Yuxin Peng* \nWangxuan Institute of Computer Technology, Peking University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 192, + 423, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 192, + 423, + 203 + ], + "spans": [ + { + "bbox": [ + 187, + 192, + 423, + 203 + ], + "type": "text", + "content": "kslh99@stu.pku.edu.cn, pengyuxin@pku.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 206, + 293, + 363 + ], + "blocks": [ + { + "bbox": [ + 63, + 206, + 293, + 363 + ], + "lines": [ + { + "bbox": [ + 63, + 206, + 293, + 363 + ], + "spans": [ + { + "bbox": [ + 63, + 206, + 293, + 363 + ], + "type": "image", + "image_path": "238b9e451c661a20780a71bf7fd062a0ca0425da339f0ba93cd0285c85253502.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 365, + 544, + 376 + ], + "lines": [ + { + "bbox": [ + 64, + 365, + 544, + 376 + ], + "spans": [ + { + "bbox": [ + 64, + 365, + 544, + 376 + ], + "type": "text", + "content": "Figure 1. Illustration of the generalized settings in content-aware layout generation from both (a) data and (b) approach perspectives." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 295, + 207, + 545, + 363 + ], + "blocks": [ + { + "bbox": [ + 295, + 207, + 545, + 363 + ], + "lines": [ + { + "bbox": [ + 295, + 207, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 295, + 207, + 545, + 363 + ], + "type": "image", + "image_path": "8b35deafe8e070892b506f38ac775a00bfe3d22452e3fb9c41cb329ce9c675e1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 381, + 200, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 381, + 200, + 393 + ], + "spans": [ + { + "bbox": [ + 151, + 381, + 200, + 393 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 399, + 296, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 296, + 697 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 296, + 697 + ], + "type": "text", + "content": "In poster design, content-aware layout generation is crucial for automatically arranging visual-textual elements on the given image. With limited training data, existing work focused on image-centric enhancement. However, this neglects the diversity of layouts and fails to cope with shape-variant elements or diverse design intents in generalized settings. To this end, we proposed a layout-centric approach that leverages layout knowledge implicit in large language models (LLMs) to create posters for omnifarious purposes, hence the name PosterO. Specifically, it structures layouts from datasets as trees in SVG language by universal shape, design intent vectorization, and hierarchical node representation. Then, it applies LLMs during inference to predict new layout trees by in-context learning with intent-aligned example selection. After layout trees are generated, we can seamlessly realize them into poster designs by editing the chat with LLMs. Extensive experimental results have demonstrated that PosterO can generate visually appealing layouts for given images, achieving new state-of-the-art performance across various benchmarks. To further explore PosterO's abilities under the generalized settings, we built PStylish7, the first dataset with multi-purpose posters and various-shaped elements, further offering a challenging test for advanced research. Code and dataset will be publicly available at https://thekinsley.github.io/PosterO/." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 381, + 394, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 381, + 394, + 393 + ], + "spans": [ + { + "bbox": [ + 315, + 381, + 394, + 393 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 398, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 495 + ], + "type": "text", + "content": "Content-aware layout generation makes the poster design process intelligent by automatically arranging visual-textual elements on given images. As the demand for effective design grows rapidly, a mainstream platform surpasses 185 million monthly users [3], and " + }, + { + "bbox": [ + 313, + 398, + 555, + 495 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 398, + 555, + 495 + ], + "type": "text", + "content": " of survey participants take AI-empowered features as key considerations [2]. This technology is gaining more attention for its practical value in various applications [32, 49, 50, 53], as shown in Fig. 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 556, + 714 + ], + "type": "text", + "content": "Unlike other generation tasks [22, 43], collecting image-Layout pairs for the target task requires costly annotation and inpainting processes [19, 55]. Therefore, existing work has incorporated object hints [4, 55], prior information [7, 19], retrieved references [17, 31], and augmented data [44] to confront the challenge of training data scarcity. However, because these are mainly image-centric enhancements without increasing the adaptability and diversity of layouts, they are prone to get trapped in the local solution space [18]. On the other hand, while leveraging related knowledge implicit in large language models (LLMs) [11, 37, 42, 47] showed the potential for layout-centric enhancements [12, 46], existing work [31, 44] has suffered from the semantically impoverished layout representations. The two primary issues are the monotonous rectangular shape of elements and the isolated visual constraints, which led to failures at capturing the complexity of real-world layouts and the intricate visual relationship between images and elements." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8117" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "To pursue data efficiency and semantic richness required for advances in content-aware layout generation, we proposed PosterO to enable in-context learning (ICL) of LLMs by structuring layout tree representations. Specifically, it consists of three procedures, including (a) layout tree construction that jointly represents layout elements and design intents (i.e., available areas on images) as hierarchical nodes in SVG trees [9], (b) layout tree generation that dynamically selects intent-aligned learning examples for the test input to facilitate ICL, and (c) poster design realization that integrates design materials into the generated layout trees to create posters in the subsequent chats with LLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 431 + ], + "type": "text", + "content": "We conducted extensive experiments on public benchmarks [19, 55] and demonstrated that PosterO has achieved new state-of-the-art performance. It particularly shows stability in managing domain adaption problems [52] and spatial distribution shifts [17] that make the current approaches stumped. In additional efforts, we enhanced the evaluation process through newly added intent-aware content metrics that are more credible than currently used saliency-aware ones. We also demonstrated the adaptability of PosterO to different design intents and small-scale LLMs [35], highlighting its applicability. To further explore its potential under the generalized settings of diverse poster purposes (e.g., commercial or education) and shape-variant elements (e.g., circles or curves), we built the first dataset, PStylish7, consisting of 152 few-shot learning samples and 100 test images, covering seven representative poster purposes and eight element types. Serving as a challenging test, it is believed to facilitate advanced research in this field." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 433, + 293, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 433, + 293, + 444 + ], + "spans": [ + { + "bbox": [ + 67, + 433, + 293, + 444 + ], + "type": "text", + "content": "We summarize the contribution of this paper as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 295, + 589 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 445, + 295, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 481 + ], + "type": "text", + "content": "- A language-based content-aware layout representation, layout tree, uses hierarchical nodes to model both layout elements and design intents with arbitrary shapes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 482, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 295, + 517 + ], + "type": "text", + "content": "- A layout-centric approach, PosterO, leverages implicit knowledge of LLMs by applying a few intent-aligned layout trees as examples to perform in-context learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 517, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 517, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 517, + 295, + 553 + ], + "type": "text", + "content": "- Extensive evaluations and newly added metrics verify the superior effectiveness of PosterO in confronting various adaptation problems that plague current approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 554, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 295, + 589 + ], + "type": "text", + "content": "- A dataset for generalized content-aware layout generation, PStylish7, incorporates seven representative poster purposes and eight shape-variant element types." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 590, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 613 + ], + "type": "text", + "content": "As our best knowledge, this work is the first to consider the generalized settings of content-aware layout generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 627, + 142, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 142, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 142, + 639 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 239, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 239, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 239, + 659 + ], + "type": "text", + "content": "2.1. Content-aware Layout Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Apart from general layout tasks [1, 15, 23, 25, 28], content-aware layout generation is a cross-modal task [10, 16] integrating visual conditions, i.e., image canvases, which are inherent in various real-world applications [32, 49, 50, 53]," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "highlighting its practical utility. With advances in deep generative models, data-centric approaches have become dominant in this field, categorized into GAN-based [18, 19, 52, 55], auto-regression-based [4, 5, 17, 30], and diffusion model-based [7, 27] approaches." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 133, + 554, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 554, + 337 + ], + "type": "text", + "content": "Entering into data construction of image layout pairs, CGL-GAN [55] leveraged an inpainting model [45] to remove design elements from posters and a saliency object detection model [48] to give generative networks additional hints about image compositions. Similar saliency-enhanced paradigm [26, 39, 40] has been adopted by most subsequent approaches. Dealing with limited training data, DS-GAN [19] introduced prior knowledge [14, 29] to arrange layout elements in a meticulously designed order, making the patterns existing in the data more efficiently mined. RALF introduced retrieval augmentation, leveraging a composition-aware similarity evaluation model [13] to find the nearest neighbors for a given image, and incorporating their layout features [23] as additional inputs for the generative network. Moreover, Chai et al. [7] adapted a general layout diffusion model [6] for content-aware tasks by incorporating aesthetic constraints and a saliency-aware layout plausibility ranker." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 338, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 408 + ], + "type": "text", + "content": "As current approaches are engaged in image-centric enhancement yet seldom expand the diversity of layouts, the accessible solution space is narrowed, impeding their ability to address spatially varying design intents across different applications. In light of this, we propose a layout-centric, intent-aware approach in this work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 418, + 537, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 537, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 537, + 431 + ], + "type": "text", + "content": "2.2. Language Model-based Layout Generation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "type": "text", + "content": "While numerical layout representations risk losing the semantic information of attributes [46], current advances in large language models (LLMs) [11, 42, 47] open up new possibilities with HTML-based [31, 38], CSS-based [8, 12], and SVG-based [44, 46] language representations as these approaches enable leverage of layout knowledge implicit in LLMs. However, due to the challenge of guiding LLMs to perceive visual content, only a few approaches [31, 44] have supported content-aware layouts." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 545, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 641 + ], + "type": "text", + "content": "LayoutPrompter [31] represented each layout element as a
in HTML. While adopting a saliency paradigm, it only extracted the minimum bounding rectangle from the binary map as coarse content constraints and then retrieved layout examples for in-context learning. In contrast, PosterLlama [44] adopted a SVG-based representation [46] using shapes. It then fine-tuned DINOv2 [36, 56] with CodeLlama [20, 42] enhanced by data augmentation [54]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "As current approaches employed monotype element representations and isolated visual representations, they failed to exploit the rich semantics of languages nor explicitly capture the relationship between elements and images. In light of this, we propose layout trees to represent various-shaped elements and visual hints jointly." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8118" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 73, + 281, + 237 + ], + "blocks": [ + { + "bbox": [ + 56, + 73, + 281, + 237 + ], + "lines": [ + { + "bbox": [ + 56, + 73, + 281, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 281, + 237 + ], + "type": "image", + "image_path": "4074032f7c66cdc63cb79a0ec59d251fa818eeb5e0b4a7f022598b817e7e2e78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "text", + "content": "Figure 2. An overview of PosterO. First, (a) takes image layout pairs " + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "inline_equation", + "content": "(I,L)" + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "text", + "content": " from datasets as input for data preparation, jointly modeling various-shaped elements " + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "text", + "content": " and design intents " + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "text", + "content": " towards layout trees " + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "text", + "content": " and building up the latent space of " + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "text", + "content": ". Then, (b) takes test images " + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "text", + "content": " as input and searches examples based on the predicted design intents " + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "text", + "content": " to apply LLM " + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "text", + "content": " through in-context learning. After obtaining generated layout trees " + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\hat{T}" + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "text", + "content": ", (c) can continue the conversation with " + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 243, + 555, + 289 + ], + "type": "text", + "content": " to create poster designs seamlessly." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 285, + 73, + 468, + 238 + ], + "blocks": [ + { + "bbox": [ + 285, + 73, + 468, + 238 + ], + "lines": [ + { + "bbox": [ + 285, + 73, + 468, + 238 + ], + "spans": [ + { + "bbox": [ + 285, + 73, + 468, + 238 + ], + "type": "image", + "image_path": "ac9e0efde70344e1d54552de7f293726700faa26298f1236c30c677955e5e48d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 470, + 74, + 552, + 237 + ], + "blocks": [ + { + "bbox": [ + 470, + 74, + 552, + 237 + ], + "lines": [ + { + "bbox": [ + 470, + 74, + 552, + 237 + ], + "spans": [ + { + "bbox": [ + 470, + 74, + 552, + 237 + ], + "type": "image", + "image_path": "fa93599ed41f955c2539228a3dcb175333db26475495f61243acb43b2e9672e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 244, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 244, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 244, + 309 + ], + "type": "text", + "content": "3. The Proposed Approach: PosterO" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 316, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 448 + ], + "type": "text", + "content": "Considering the limitations of image-centric and language model-based approaches, we propose PosterO. It is layout-centric and enables large language models (LLMs) to generate visually appealing layouts for given images by structurally representing the content-aware layouts as trees " + }, + { + "bbox": [ + 55, + 316, + 296, + 448 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 316, + 296, + 448 + ], + "type": "text", + "content": " in SVG language. An overview of PosterO is shown in Fig. 2. In brief, it is composed of three procedures, namely, (a) layout tree construction for data preparation, (b) layout tree generation applying an LLM " + }, + { + "bbox": [ + 55, + 316, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 316, + 296, + 448 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 55, + 316, + 296, + 448 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 316, + 296, + 448 + ], + "type": "text", + "content": "-shot in-context learning, and (c) poster design realization through the subsequent chat with " + }, + { + "bbox": [ + 55, + 316, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 316, + 296, + 448 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 454, + 200, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 200, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 200, + 466 + ], + "type": "text", + "content": "3.1. Layout Tree Construction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 212, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 212, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 212, + 483 + ], + "type": "text", + "content": "3.1.1. Universal Shape Vectorization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": "In content-aware layout generation, each data pair contains an image " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": " and a layout " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": ". Usually, " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": " is represented as a set of numerical tuples " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\{e_i\\}_{i}^n = \\{(c_i, b_i)\\}_{i}^n" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "b_i = (x, y, w, h)" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": " indicate the category and bounding box of the " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": "-th element " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": " out of all " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": " elements. As this is adequate only for rectangular elements, a more generalized representation becomes essential for the diverse shapes encountered in real-world posters, such as circles and curves. In light of this, SVG (Scalable Vector Graphics) emerges as an ideal choice because of its ability to represent arbitrary shapes and, more importantly, its nature as a language to facilitate the layout knowledge implicit in LLMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Following the published SVG standards [9], we define five shapes to encompass a wide range of elements typically found in posters, as shown in Fig. 3. First, (a) regular rectangle is vectorized using " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\langle \\text{rect} \\rangle" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ". Also, its two variants, including (b) vertical rectangle sensitive to the aspect ratio and (c) rotated rectangle using the rotate function in the transform attribute to indicate the rotation degree." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 294, + 390, + 365 + ], + "blocks": [ + { + "bbox": [ + 316, + 294, + 390, + 365 + ], + "lines": [ + { + "bbox": [ + 316, + 294, + 390, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 390, + 365 + ], + "type": "image", + "image_path": "8f213634ecc780f05c86c798374b0ab2428589bc8dd2eb5cc551cee6a7b356cc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 366, + 385, + 375 + ], + "lines": [ + { + "bbox": [ + 320, + 366, + 385, + 375 + ], + "spans": [ + { + "bbox": [ + 320, + 366, + 385, + 375 + ], + "type": "text", + "content": "(a) General rectangle" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 397, + 294, + 472, + 365 + ], + "blocks": [ + { + "bbox": [ + 397, + 294, + 472, + 365 + ], + "lines": [ + { + "bbox": [ + 397, + 294, + 472, + 365 + ], + "spans": [ + { + "bbox": [ + 397, + 294, + 472, + 365 + ], + "type": "image", + "image_path": "c413ee1a7ab9f381a74b21df64c809230c402e0b05fda8f0b891e7e0b433867f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 366, + 467, + 375 + ], + "lines": [ + { + "bbox": [ + 403, + 366, + 467, + 375 + ], + "spans": [ + { + "bbox": [ + 403, + 366, + 467, + 375 + ], + "type": "text", + "content": "(b) Vertical rectangle" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 480, + 294, + 553, + 365 + ], + "blocks": [ + { + "bbox": [ + 480, + 294, + 553, + 365 + ], + "lines": [ + { + "bbox": [ + 480, + 294, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 480, + 294, + 553, + 365 + ], + "type": "image", + "image_path": "33abfb8ceac90a7d23ab9b60c19ed372a1343ea3e2b9cdf8d8b64a75566a21f0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 366, + 549, + 375 + ], + "lines": [ + { + "bbox": [ + 485, + 366, + 549, + 375 + ], + "spans": [ + { + "bbox": [ + 485, + 366, + 549, + 375 + ], + "type": "text", + "content": "(c) Rotated rectangle" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 376, + 385, + 442 + ], + "blocks": [ + { + "bbox": [ + 317, + 376, + 385, + 442 + ], + "lines": [ + { + "bbox": [ + 317, + 376, + 385, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 385, + 442 + ], + "type": "image", + "image_path": "4c0796c747d085794b6628493d16236b254d117e40b0ca532761ebb144d50ea9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 443, + 369, + 453 + ], + "lines": [ + { + "bbox": [ + 336, + 443, + 369, + 453 + ], + "spans": [ + { + "bbox": [ + 336, + 443, + 369, + 453 + ], + "type": "text", + "content": "(d) Ellipse" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 397, + 376, + 471, + 442 + ], + "blocks": [ + { + "bbox": [ + 397, + 376, + 471, + 442 + ], + "lines": [ + { + "bbox": [ + 397, + 376, + 471, + 442 + ], + "spans": [ + { + "bbox": [ + 397, + 376, + 471, + 442 + ], + "type": "image", + "image_path": "a9cf75fab83f5a938aed6f09c77c20083ae463ec281d4b15b949e705faf7574a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 443, + 463, + 453 + ], + "lines": [ + { + "bbox": [ + 406, + 443, + 463, + 453 + ], + "spans": [ + { + "bbox": [ + 406, + 443, + 463, + 453 + ], + "type": "text", + "content": "(e) Complex curve" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 479, + 376, + 553, + 442 + ], + "blocks": [ + { + "bbox": [ + 479, + 376, + 553, + 442 + ], + "lines": [ + { + "bbox": [ + 479, + 376, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 479, + 376, + 553, + 442 + ], + "type": "image", + "image_path": "8929c0e5cea4c1dc80a98fe81e10dd56c53a870ae6a2042e82e622f4e52adc76.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 444, + 534, + 453 + ], + "lines": [ + { + "bbox": [ + 499, + 444, + 534, + 453 + ], + "spans": [ + { + "bbox": [ + 499, + 444, + 534, + 453 + ], + "type": "text", + "content": "(f) Polygon" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 455, + 539, + 467 + ], + "lines": [ + { + "bbox": [ + 328, + 455, + 539, + 467 + ], + "spans": [ + { + "bbox": [ + 328, + 455, + 539, + 467 + ], + "type": "text", + "content": "Figure 3. Various shapes of elements in PStylish7 dataset." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 475, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 524 + ], + "type": "text", + "content": "When delving into more intricate shapes, (d) ellipse is vectorized using , which can also handle the circle as a special case, and (e) complex curve is vectorized using approximated by multiple cubic Bezier curves." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 531, + 459, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 459, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 459, + 542 + ], + "type": "text", + "content": "3.1.2. Design Intent Vectorization" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "text", + "content": "In addition to visual-textual elements, the design intents associated with the layout " + }, + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "text", + "content": " must also be vectorized. Specifically, design intents " + }, + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "text", + "content": " are embodied by a set of areas within the image " + }, + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "text", + "content": ", suitable or intended for placing the elements. By incorporating these areas' representations " + }, + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "inline_equation", + "content": "N_{D}" + }, + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "text", + "content": ", the layout tree becomes content-aware and capable of conveying visual relationships between elements and images to LLMs, which cannot 'see' the images. As " + }, + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 545, + 555, + 666 + ], + "type": "text", + "content": " can be learned from existing data or customized case by case, this also provides a flexible interface for indicating various purposes." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "By leveraging image layout pairs from posters with a similar purpose, a design intent detection model " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": " based on U-Net [41] is trained in a semi-supervised manner. Referring to [18], " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": " takes the bitmap of layout elements as the" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8119" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "supervision signal during the optimization, while the consistency loss is canceled to clearly segment out the areas. With the pre-trained " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ", PosterO can obtain design intents " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " aligning the target purpose for any image " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ". Specifically, after thresholding and morphological operations, we approximate the contour of the predicted intent maps " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " and vectorize the obtained irregular areas using in SVG language. An example is shown in Fig. 3(f). Besides, the intermediate embeddings " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "f = S_{\\mathfrak{e}}(I)" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "S_{\\mathfrak{e}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " denotes the last block of the encoder, are also collected to serve as their representations in the latent space of design intents " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_D" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 230, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 230, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 230, + 220 + ], + "type": "text", + "content": "3.1.3. Hierarchical Node Representation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": "As the nodes of layout elements " + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "inline_equation", + "content": "N_{E}" + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": " and design intents " + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "inline_equation", + "content": "N_{D}" + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": " are in place, the layout tree " + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": " is constructed and encapsulated within an SVG container aware of the image resolution " + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "inline_equation", + "content": "(w_{I}, h_{I})" + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": ", as: " + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "inline_equation", + "content": "<\\mathrm{svg~width}=\\{\"w_{I}\\}\" \\mathrm{height}=\\ {\"h_{I}\\}\" \\mathrm{xmlns}=\\text{\"http://www.w3.org/2000/sv\"}> \\{N_{D}\\} \\{N_{E}\\}<\\mathrm{svg}>" + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "inline_equation", + "content": "N_{D}" + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": " is put ahead to ensure that " + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "inline_equation", + "content": "N_{E}" + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": " is arranged considering their probabilistic dependencies in the autoregression, thus generating suitable content-aware layouts. In addition, as the special elements that wrap around others, such as underlays, are usually used in poster design, a novel hierarchical representation explicitly modeling these enveloping properties is proposed. Specifically, " + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "inline_equation", + "content": "N_{E}" + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": " is sorted by prioritizing underlays with larger areas. Then, given a node " + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": " and the corresponding nodes " + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "inline_equation", + "content": "N_{B} = \\{N_{b} \\mid N_{b} \\in \\{N_{E}\\}_{a + 1}^{n} \\land C(N_{a}, N_{b})\\}" + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": ", where the condition function " + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "inline_equation", + "content": "C(\\cdot)" + }, + { + "bbox": [ + 55, + 224, + 298, + 403 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 408, + 294, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 294, + 452 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} C (\\cdot) = \\left(| N _ {a}. x - N _ {b}. x | \\leq \\epsilon\\right) \\wedge \\left(| N _ {a}. y - N _ {b}. y | \\leq \\epsilon\\right) \\\\ \\wedge \\left(\\left| N _ {a}. x + N _ {a}. w - N _ {b}. x - N _ {b}. w \\right| \\leq \\epsilon\\right) \\tag {1} \\\\ \\wedge \\left(\\left| N _ {a}. y + N _ {a}. h - N _ {b}. y - N _ {b}. h \\right| \\leq \\epsilon\\right), \\\\ \\end{array}", + "image_path": "c099750b622ffda0d84b5e7b41999ec3cbf59c1aad23716021071c336eacff71.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "text", + "content": "a subtree is constructed as " + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "inline_equation", + "content": "<\\text{svg} \\times=\" \\{N_{a}.x\\}\" \\text{y=\"} \\{N_{a}.y\\}\">" + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\{N_{a}'\\} \\{N_{B}'\\} < / \\text{svg}>" + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "text", + "content": " and inserted at the position of " + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "text", + "content": ", where the processed " + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "inline_equation", + "content": "N_{i}'" + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "text", + "content": " has the relative geometry attributes " + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "inline_equation", + "content": "(N_{i}'.x, N_{i}'.y) = (N_{i}.x - N_{a}.x, N_{i}.y - N_{a}.y)" + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "text", + "content": ". All leaf nodes of " + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "text", + "content": " are assigned unique identifiers " + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\mathfrak{id} = \"\\{c_{i} - i\\}\"" + }, + { + "bbox": [ + 55, + 457, + 296, + 531 + ], + "type": "text", + "content": " to facilitate precise manipulation and querying." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 536, + 191, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 191, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 191, + 548 + ], + "type": "text", + "content": "3.2. Layout Tree Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 553, + 225, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 225, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 225, + 566 + ], + "type": "text", + "content": "3.2.1. Intent-aligned Example Selection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": "Upon findings from [34], LLMs are capable of in-context learning (ICL), requiring only a few examples to perform downstream tasks. This significantly reduces the need for task-specific data and avoids the extensive costs or catastrophic forgetting [24, 33] associated with fine-tuning. To make the best use of layout knowledge implicit in LLMs during ICL, selecting examples that align with the design intent is crucial. Therefore, given the test image " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": ", its embedding " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " is first extracted by the detection model " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": "; subsequently, " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " layout trees " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{T_j\\}_j^k" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " are selected as in-context examples, where their embeddings " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{f_j\\}_j^k" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " closest to " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": " in the latent space " + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{D}" + }, + { + "bbox": [ + 55, + 568, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 527, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 527, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 527, + 84 + ], + "type": "text", + "content": "3.2.2. In-context Learning for Layout Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "text", + "content": "To provide clear structure and objectives in the prompt " + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "text", + "content": " for the LLM " + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "text", + "content": ", each intent-aligned example " + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "text", + "content": " is preceded by a short sentence that informs its number " + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "text", + "content": ", the resolution of its corresponding image, design intents " + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "text", + "content": ", and identifiers of its nodes, as shown in Fig. 2(b). The example field is then encapsulated within a preface and postscript, with the postscript delivering the request message of the test image " + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "text", + "content": " through a short sentence mirroring that of the example. Complete prompts " + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "text", + "content": " are presented in supplementary material. Finally, by inputting " + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 87, + 555, + 220 + ], + "type": "text", + "content": " and interpreting the responses, PosterO completes the layout generation process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 227, + 457, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 457, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 457, + 239 + ], + "type": "text", + "content": "3.3. Poster Design Realization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 243, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 389 + ], + "type": "text", + "content": "Benefiting from the essence of the generated layout tree " + }, + { + "bbox": [ + 313, + 243, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\hat{T}" + }, + { + "bbox": [ + 313, + 243, + 555, + 389 + ], + "type": "text", + "content": " as graphics and the rigid id assignment for elements, guiding " + }, + { + "bbox": [ + 313, + 243, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 243, + 555, + 389 + ], + "type": "text", + "content": " to transform " + }, + { + "bbox": [ + 313, + 243, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\hat{T}" + }, + { + "bbox": [ + 313, + 243, + 555, + 389 + ], + "type": "text", + "content": " into a practical poster design is feasible and straightforward. Specifically, the transformation contains mockup creation, where shape elements are converted to or regarding their categories, and material synthesis, where elements are filled with actual design contents, e.g., slogans and hyperlinks to images. Notably, this transformation is achieved with zero-shot learning, yet " + }, + { + "bbox": [ + 313, + 243, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 243, + 555, + 389 + ], + "type": "text", + "content": " brings impressive results (in supplementary material) and even predicts suitable font sizes for texts, highlighting the value of design knowledge implicit in LLMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 399, + 395, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 395, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 395, + 412 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 419, + 376, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 376, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 376, + 430 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 594 + ], + "type": "text", + "content": "To evaluate the proposed PosterO, two public e-commerce poster datasets are used, while their train+valid/annotated test/unannotated test splits are allocated as in [17] to ensure a fair comparison with existing work. Specifically, PKU PosterLayout [19] comprises 8,734/1,000/905 images with three layout element types, including text, logo, and underlay; CGL [55] comprises 54,546/6,002/1,000 images with four layout element types, including text, logo, underlay, and embellishment. Furthermore, to explore PosterO's abilities under the generalized settings, we built the first multipurpose poster dataset, PStylish7, containing layout elements of various shapes. More details about PStylish7 are depicted in Sec. 4.6 and supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 600, + 427, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 427, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 427, + 612 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Following previous work [17, 19, 55], we use graphic and content metrics. Graphic metrics include overlay " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "Ove\\downarrow" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " alignment " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "Ali\\downarrow" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " , and loose, strict underlay effectiveness " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "Und_{l}\\uparrow ,Und_{s}\\uparrow ;" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " Content metrics include non-salient space utilization " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "Uti\\uparrow" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " , salient space occlusion " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "Occ\\downarrow" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " , and readability " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "Rea\\downarrow" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " . Moreover, we newly proposed two content metrics, design intent coverage " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "Cov\\uparrow" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " and conflict " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "Con\\downarrow" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " based on a design intent detection model " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "S_{pc}" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " that is compre" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8120" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 397, + 411 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 397, + 411 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 397, + 411 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 397, + 411 + ], + "type": "table", + "html": "
MethodOve ↓Ali ↓Undl ↑Unds ↑Int ↓Sal ↓Rea ↓Avg ↓
Annotated Test Split
Real data0.00100.00380.99420.99030.02890.03680.01090.0138
CGL-GAN0.09660.00350.78540.35700.12040.27740.01910.1964
DS-GAN0.02610.00380.83500.58040.06770.31000.01990.1446
ICVT0.25720.04050.53840.39320.78311.08360.02590.4655
LayoutDM†0.15620.00180.64260.38730.50660.24290.01850.2709
RALF0.00840.00280.98080.92010.11620.04790.01280.0410
Tuning-required
PosterLlama‡0.00080.00060.99990.99820.34070.26190.01770.0891
LLM-based
Tuning-free
LayoutPrompterL20.00830.00300.37160.16330.32000.94330.03190.3959
LayoutPrompterCL0.00400.00300.44150.19370.36620.90350.03010.3817
LayoutPrompterL30.00170.00280.40850.16130.51990.85670.03090.4060
PosterOL2 (Ours)0.00040.00270.75970.56050.36790.14910.01880.1741
PosterOLCL (Ours)0.00040.00280.96210.86060.23970.22950.01890.0955
PosterOL3 (Ours)0.00060.00250.98880.92110.10800.11100.01710.0470
Unannoted Test Split
Non-LLMTrain data0.00120.00270.99690.99000.00000.00000.0108
CGL-GAN0.10100.00480.73260.27430.22750.77550.0327
DS-GAN0.02480.00460.78590.46760.16590.58680.0320
ICVT0.27860.04800.49390.35490.79471.32890.0347
LayoutDM†0.16380.00290.59870.36950.53570.31720.0264
RALF0.01750.00690.95480.86530.59860.17790.0231
Tuning-required
PosterLlama‡0.00060.00060.99860.99170.26670.38230.02850.0983
LLM-based
Tuning-free
LayoutPrompterL20.00950.00320.34010.11610.53491.20500.04080.4767
LayoutPrompterCL0.00320.00280.40520.14640.51481.11060.03950.4456
LayoutPrompterL30.00100.00260.40540.16210.70051.08380.04120.4660
PosterOL2 (Ours)0.00040.00310.72840.51880.40300.33510.02650.2173
PosterOLCL (Ours)0.00050.00280.95950.84510.27960.39450.02730.1286
PosterOL3 (Ours)0.00040.00250.98560.92410.14270.21310.02480.0677
", + "image_path": "d38091159da8a2c3b94d72ed52200cc5f0bd33325a0be2440f4a850da0e4951d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "text", + "content": "hensively pre-trained on [19] and [55]. Calculating " + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "inline_equation", + "content": "Cov\\uparrow" + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "inline_equation", + "content": "Con\\downarrow" + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "text", + "content": " is similar to " + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "inline_equation", + "content": "Uti\\uparrow" + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "inline_equation", + "content": "Occ\\downarrow" + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "text", + "content": " but using the predicted intent maps. As reported in Tab. 2, intent maps have doubled the matched portion with real layout data, reduced the unmatched portion by " + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "inline_equation", + "content": "34.7\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "text", + "content": ", and required only " + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "text", + "content": " computational costs. These findings have demonstrated that intent-aware metrics are more credible and efficient. For a holistic understanding, we standardize intent-aware and saliency-aware metrics to " + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "inline_equation", + "content": "Int\\downarrow" + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "inline_equation", + "content": "Sal\\downarrow" + }, + { + "bbox": [ + 55, + 434, + 295, + 541 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 550, + 295, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 550, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 80, + 550, + 295, + 582 + ], + "type": "interline_equation", + "content": "I n t = \\Sigma (\\frac {\\left| C o v _ {\\hat {L}} - C o v _ {L} \\right|}{1 - C o v _ {L}}, \\frac {\\left| C o n _ {\\hat {L}} - C o n _ {L} \\right|}{C o n _ {L}}), \\tag {2}", + "image_path": "b871455a59651da4d2db902dfa54b9fda871d645b1c8a8e2fb16af684d47270f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 578, + 262, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 578, + 262, + 604 + ], + "spans": [ + { + "bbox": [ + 82, + 578, + 262, + 604 + ], + "type": "interline_equation", + "content": "S a l = \\Sigma (\\frac {| U t i _ {\\hat {L}} - U t i _ {L} |}{1 - U t i _ {L}}, \\frac {| O c c _ {\\hat {L}} - O c c _ {L} |}{O c c _ {L}}),", + "image_path": "53c4daf779cb988b79928ec6f58a27c731a28b6c0d8b60b79301465cbb8a12f4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\hat{L}" + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "text", + "content": " are the layouts from the train split and generation. Finally, we introduce an average metric " + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "inline_equation", + "content": "Avg\\downarrow = \\Sigma (Ove,Ali,1 - Und_{l},1 - Und_{s},Int,Sal,Rea)" + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 188, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 188, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 188, + 672 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "To ensure independent evaluations on [19] and [55], we correspondingly train two detection models, " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "S_{p}" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": ", with mean squared error loss and a learning rate of " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "1e^{-6}" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": ". Each" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 408, + 71, + 550, + 390 + ], + "blocks": [ + { + "bbox": [ + 408, + 71, + 550, + 390 + ], + "lines": [ + { + "bbox": [ + 408, + 71, + 550, + 390 + ], + "spans": [ + { + "bbox": [ + 408, + 71, + 550, + 390 + ], + "type": "image", + "image_path": "da94c947185bbe62879a25a49fd895c465362ebb3ecdadd71315ebb59ea20ca9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 392, + 552, + 425 + ], + "lines": [ + { + "bbox": [ + 403, + 392, + 552, + 425 + ], + "spans": [ + { + "bbox": [ + 403, + 392, + 552, + 425 + ], + "type": "text", + "content": "Figure 4. Comparisons of visualized results on the unannotated test split of PKU PosterLayout dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 431, + 553, + 523 + ], + "blocks": [ + { + "bbox": [ + 94, + 415, + 366, + 426 + ], + "lines": [ + { + "bbox": [ + 94, + 415, + 366, + 426 + ], + "spans": [ + { + "bbox": [ + 94, + 415, + 366, + 426 + ], + "type": "text", + "content": "Table 1. Comparisons of quantitative results on PKU PosterLayout dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 431, + 553, + 523 + ], + "lines": [ + { + "bbox": [ + 315, + 431, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 315, + 431, + 553, + 523 + ], + "type": "table", + "html": "
PredictionPKUCGL
Match ↑Unmatch ↓Match ↑Unmatch ↓
Saliency map0.22380.11930.19840.1352
Design intent map0.44140.09960.40360.0883
Inference costNetwork architecture#FLOPs(G)#Params(M)
Saliency mapISNet + BASNet287.83131.11
Design intent mapUNet (MiT-B1)3.7416.43
", + "image_path": "3959275293406a32bd9b97ec4cef394fda7b05b4ebb2908a477ed13a8054e097.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 530, + 555, + 564 + ], + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 564 + ], + "type": "text", + "content": "Table 2. Comparisons of the predicted saliency and design intent maps on the annotated test splits of PKU PosterLayout and CGL datasets, showing that intent-aware metrics are more credible." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "detection model is with a Transformer encoder tailored for semantic segmentation, MiT-B1 [51]. For in-context learning (ICL), we implement " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " with three open-source LLMs, including Llama 2-7B (L2) [47], CodeLlama-7B (CL) [42], and Llama 3.1-8B (L3) [11], with an unified sampling temperature of 0.7. Considering the context lengths and remaining consistent with LayoutPrompter [31], the example size " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " is 5, 10, and 10, respectively. Still adhering to [31], a layout ranker is employed, while mIoU in the quality function measures the similarity between generated layouts and predicted design intents rather than real data, minimizing the risk of data leakage." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8121" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 397, + 411 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 397, + 411 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 397, + 411 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 397, + 411 + ], + "type": "table", + "html": "
MethodOve ↓Ali ↓Undl ↑Unds ↑Int ↓Sal ↓Rea ↓Avg ↓
Annotated Test Split
Non-LLMReal data0.00030.00240.99630.98800.00880.00970.01190.0070
CGL-GAN0.22910.01230.64660.22810.70880.46270.02130.3656
DS-GAN0.04600.00220.90810.63080.39500.15390.01810.1537
ICVT0.24530.01790.51500.33260.36320.51390.02110.3305
LayoutDM†0.01840.00210.92160.81590.04110.02860.01370.0523
RALF0.00420.00240.99120.97560.03660.07240.01800.0238
Tuning-required PosterLlama‡0.00060.00060.99870.98900.13370.25590.01840.0602
LLM-basedTuning-free
LayoutPrompterL20.00870.00330.33690.14550.33620.82340.03270.3888
LayoutPrompterCL0.00370.00310.39650.16160.37180.77990.03170.3760
LayoutPrompterL30.00170.00300.38300.17400.55930.79570.03270.4051
PosterOL2 (Ours)0.00060.00230.78010.61500.25420.04670.01760.1323
PosterOLCL (Ours)0.00050.00230.96820.88680.12460.07000.01680.0513
PosterOL3 (Ours)0.00060.00220.98420.93400.03880.00260.01610.0203
Unannotated Test Split
Non-LLMTrain data0.00030.00270.99450.98580.00000.00000.01180.0049
CGL-GAN0.26680.03160.67740.16560.97072.39730.05120.6964
DS-GAN0.09910.01380.75660.28100.46072.27320.05260.5517
ICVT0.20450.10100.43570.25991.28132.65120.03970.7974
LayoutDM†0.07930.18220.63040.38531.50023.29060.06120.8711
RALF0.02730.01890.97560.93151.02371.67980.03970.4118
Tuning-required PosterLlama‡0.00140.00070.99710.97710.35102.61700.05550.4359
LLM-basedTuning-free
LayoutPrompterL20.00690.01030.28470.11261.02572.37020.06280.7255
LayoutPrompterCL0.00450.00240.32390.13411.58002.53800.06500.8188
LayoutPrompterL30.00260.00160.26930.11421.62492.41280.06440.8175
PosterOL2 (Ours)0.00050.00330.66630.47700.35221.23650.03820.3553
PosterOLCL (Ours)0.00030.00370.93690.80710.27711.32230.03790.2710
PosterOL3 (Ours)0.00040.00320.98160.90340.43281.24070.03650.2612
", + "image_path": "e0e104a76d75c46ca0a1c05b37baf03e2e1c45520053b371fa49d5be90543dc5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 119, + 420, + 342, + 431 + ], + "lines": [ + { + "bbox": [ + 119, + 420, + 342, + 431 + ], + "spans": [ + { + "bbox": [ + 119, + 420, + 342, + 431 + ], + "type": "text", + "content": "Table 3. Comparisons of quantitative results on CGL dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 408, + 71, + 552, + 392 + ], + "blocks": [ + { + "bbox": [ + 408, + 71, + 552, + 392 + ], + "lines": [ + { + "bbox": [ + 408, + 71, + 552, + 392 + ], + "spans": [ + { + "bbox": [ + 408, + 71, + 552, + 392 + ], + "type": "image", + "image_path": "0843d030f2c5bb261d80b85a93409ecdfc2aefd61ddae899d084d6fe9c13bf93.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 397, + 552, + 430 + ], + "lines": [ + { + "bbox": [ + 403, + 397, + 552, + 430 + ], + "spans": [ + { + "bbox": [ + 403, + 397, + 552, + 430 + ], + "type": "text", + "content": "Figure 5. Comparisons of visualized results on the unannotated test split on CGL dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 440, + 239, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 239, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 239, + 453 + ], + "type": "text", + "content": "4.4. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 457, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 295, + 518 + ], + "type": "text", + "content": "We select approaches with open-sourced implementation as baselines, including GAN-based CGL-GAN [55], DS-GAN [19], auto-regression-based ICVT [4], RALF [17], diffusion model-based LayoutDM† [21], and also LLM-based PosterLlama‡ [44], LayoutPrompter [31]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "text", + "content": "Baseline comparison. Tab. 1 reports the quantitative results on PKU PosterLayout. The proposed PosterO shows the best overall performance across the test splits. Particularly, when confronting the challenging unannotated split that poses domain adaptation problems [52], it still achieves new state-of-the-art (SOTA) results with outstanding stability. Compared to the current non-LLM and LLM-based SOTA approaches, i.e., RALF and PosterLlama, ours significantly improves the average metric " + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "inline_equation", + "content": "Avg\\downarrow" + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "inline_equation", + "content": "52.8\\%" + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "inline_equation", + "content": "31.1\\%" + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "text", + "content": ". Consistent results are obtained on CGL, as reported in Tab. 3, ours improves " + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "inline_equation", + "content": "Avg\\downarrow" + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "inline_equation", + "content": "36.6\\%" + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "inline_equation", + "content": "40.1\\%" + }, + { + "bbox": [ + 55, + 519, + 295, + 650 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "text", + "content": "More precisely, PosterO shows impressive graphic metrics (2nd column), closely resembling the distribution of real data, with only a slight deficit in the Ali " + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 651, + 295, + 686 + ], + "type": "text", + "content": " metric on the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 441, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 548 + ], + "type": "text", + "content": "unannotated split of CGL compared to LayoutPrompter. As for content metrics (3rdcolumn), although PosterO uses only a small design intent detection model " + }, + { + "bbox": [ + 313, + 441, + 555, + 548 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 441, + 555, + 548 + ], + "type": "text", + "content": ", it achieves comparable or even superior results to the baselines that utilize larger saliency detection models [26, 40] or costly fine-tuned DINOv2 [36]. Especially, PosterO is best at dealing with the spatial shifts in subject distributions [17] present in the unannotated split of CGL, maintaining an overwhelming lead in content metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "Backbone LLMs " + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": ". Comparing the performance of tuning-free approaches across various LLMs is crucial for realizing how effectively the layout knowledge implicit in " + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": " has been leveraged. As reported in Tab. 1 and Tab. 3, PosterO demonstrates overwhelming superiority over LayoutPrompter in all counterparts with the same " + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": ". More importantly, from Poster" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "inline_equation", + "content": "_{L2,CL}" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": ", to Poster" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "inline_equation", + "content": "_{L3}" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": ", it is evident that the proposed approach has experienced improvements as the capabilities of its based " + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": " increased. This finding indicates that new knowledge, such as programming, introduced during model updates, has indeed been leveraged to generate layouts. With the continuous development of LLMs, PosterO is anticipated to keep its rising potential." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 69, + 693, + 203, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 203, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 203, + 703 + ], + "type": "text", + "content": "The extended version presented in [17]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 703, + 293, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 293, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 293, + 713 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 69, + 703, + 293, + 713 + ], + "type": "text", + "content": "With the released weights tuned on [19], [55], and synthesized data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8122" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 107 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 107 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 107 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 107 + ], + "type": "table", + "html": "
Method \\ k1257810
LayoutPrompter0.77110.66900.53530.48830.49190.4660
PosterO (Ours)0.06780.07460.07320.06630.06510.0677
", + "image_path": "ff6f2d96fedffd543d6edde4a787b275c821615ed260be0315d68a94c9558f65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 129, + 284, + 265 + ], + "blocks": [ + { + "bbox": [ + 60, + 129, + 284, + 265 + ], + "lines": [ + { + "bbox": [ + 60, + 129, + 284, + 265 + ], + "spans": [ + { + "bbox": [ + 60, + 129, + 284, + 265 + ], + "type": "image", + "image_path": "d8c86786cb1843b018447bc051224241845c9ffa14fb8b599b5f6591da0101fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 267, + 280, + 278 + ], + "lines": [ + { + "bbox": [ + 70, + 267, + 280, + 278 + ], + "spans": [ + { + "bbox": [ + 70, + 267, + 280, + 278 + ], + "type": "text", + "content": "Figure 6. Content metrics over in-context example size " + }, + { + "bbox": [ + 70, + 267, + 280, + 278 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 70, + 267, + 280, + 278 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 384 + ], + "type": "text", + "content": "Visualized results. Fig. 4 and Fig. 5 show the layouts generated by different approaches. The results illustrate that PosterO generates high-quality layouts that the distribution of elements is generous and coordinated. In contrast, the approaches based on training or tuning tend to place elements centrally and thus fail to confront images with compositions that are less common in the training data. More visualized results are presented in supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "text", + "content": "In-context example size " + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "text", + "content": ". To explore the impact of example selection during ICL, we test a set of candidates " + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "inline_equation", + "content": "\\{k\\} = \\{1,2,5,7,8,10\\}" + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "text", + "content": " on the unannotated split of [19], as reported in Tab. 4. Amazingly, PosterO consistently outperforms LayoutPrompter and all other baselines in Tab. 1 across all settings of " + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "text", + "content": ". As the optimal performance occurs at " + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "inline_equation", + "content": "k = 7" + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "text", + "content": " and 8, it is proved that " + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "text", + "content": " has truly benefited from learning with the selected intent-aligned examples. To further observe how well these examples help " + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "text", + "content": " understand visual contents, which is the focal point of content-aware layout generation, we plot the resultant content metrics over " + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "text", + "content": " in Fig. 6. It is noteworthy that directly adopting the most intent-aligned example (Top-1 Retrieval) surpasses all baselines, indicating a successful alignment between intent and layout in the latent space " + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_D" + }, + { + "bbox": [ + 55, + 386, + 295, + 578 + ], + "type": "text", + "content": ". Nevertheless, we still observe improvements when involving in-context examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 586, + 149, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 149, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 149, + 598 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 603, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 664 + ], + "type": "text", + "content": "To investigate our design choice in layout tree construction (Sec. 3.1) and learning example selection (Sec. 3.2), as well as to explore PosterO's adaptability to customized design intents and small-scale LLMs, we conducted the following ablation studies, mainly on the unannotated split of [19]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Design choices for layout tree construction. We investigate the effectiveness of the design intent vectorization " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "(d)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and hierarchical node representation " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "(h)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " during constructing layout trees, as reported in Tab. 5. Due to space constraints," + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 553, + 135 + ], + "blocks": [ + { + "bbox": [ + 60, + 111, + 291, + 122 + ], + "lines": [ + { + "bbox": [ + 60, + 111, + 291, + 122 + ], + "spans": [ + { + "bbox": [ + 60, + 111, + 291, + 122 + ], + "type": "text", + "content": "Table 4. Comparisons of " + }, + { + "bbox": [ + 60, + 111, + 291, + 122 + ], + "type": "inline_equation", + "content": "Avg\\downarrow" + }, + { + "bbox": [ + 60, + 111, + 291, + 122 + ], + "type": "text", + "content": " over in-context example size " + }, + { + "bbox": [ + 60, + 111, + 291, + 122 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 60, + 111, + 291, + 122 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 553, + 135 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 135 + ], + "type": "table", + "html": "
dhUndl↑Unds↑Int↓Sal↓Rea↓Avg↓
✓✓0.98560.92410.14270.21310.02480.0677
✓0.86090.66500.09810.15600.02400.1078
✓0.94970.88360.98580.39910.02740.2260
0.76280.64130.81620.27290.02460.2447
", + "image_path": "ebb3cf575df61de4bc99366e30c831c42c8207109590fe6ef04cc4967ef48f14.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 168, + 552, + 231 + ], + "blocks": [ + { + "bbox": [ + 314, + 139, + 553, + 161 + ], + "lines": [ + { + "bbox": [ + 314, + 139, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 314, + 139, + 553, + 161 + ], + "type": "text", + "content": "Table 5. Ablation study on the " + }, + { + "bbox": [ + 314, + 139, + 553, + 161 + ], + "type": "inline_equation", + "content": "(d)" + }, + { + "bbox": [ + 314, + 139, + 553, + 161 + ], + "type": "text", + "content": " design intent vectorization and " + }, + { + "bbox": [ + 314, + 139, + 553, + 161 + ], + "type": "inline_equation", + "content": "(h)" + }, + { + "bbox": [ + 314, + 139, + 553, + 161 + ], + "type": "text", + "content": " hierarchical node representation in layout tree construction." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 168, + 552, + 231 + ], + "lines": [ + { + "bbox": [ + 315, + 168, + 552, + 231 + ], + "spans": [ + { + "bbox": [ + 315, + 168, + 552, + 231 + ], + "type": "table", + "html": "
SelectionUndl↑Unds↑Int↓Sal↓Rea↓Avg↓
f-aligned0.98560.92410.14270.21310.02480.0677
D-aligned0.98660.92970.22210.25310.02360.0836
E-aligned0.97430.90260.34310.41770.02640.1304
Random0.98850.93020.32250.46880.02950.1293
", + "image_path": "eabbab6e171e75a09bac78f8ddf5f45bb2edf1bca547d4260f427c4b5dc6356b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 235, + 553, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 553, + 269 + ], + "type": "text", + "content": "Table 6. Ablation study on different references for in-context example selection. (f: Embeddings of design intents, " + }, + { + "bbox": [ + 313, + 235, + 553, + 269 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 235, + 553, + 269 + ], + "type": "text", + "content": ": Bounding boxes of design intents, " + }, + { + "bbox": [ + 313, + 235, + 553, + 269 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 235, + 553, + 269 + ], + "type": "text", + "content": ": Sets of element types.)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "Ove\\downarrow" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "Ali\\downarrow" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": " are omitted, considering their slight fluctuations. As observed, the combination of " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": " yields the best results. Although removing " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": " appears to improve content metrics, this is primarily because the invalid empty underlays artificially inflate the utilization rate, as analyzed in supplementary material. The significant decreases witnessed in underlay effectiveness, " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "Und_{l}" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "Und_{s}" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": ", by up to " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": ", further underscore this issue. In contrast, removing " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": " from both the examples and postscript makes LLMs blindly mimic the provided instances. Here, however, shows a " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "Sal\\downarrow" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": " surpasses some baselines, such as CGL-GAN, DS-GAN, and LayoutPrompter, while " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "Int\\downarrow" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": " fails far behind, reflecting the robustness of the newly proposed metric " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "Int\\downarrow" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": ". If both " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": " are removed, the most severe degradation is observed, up to " + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "inline_equation", + "content": "261\\%" + }, + { + "bbox": [ + 313, + 275, + 555, + 467 + ], + "type": "text", + "content": ". In this way, we demonstrate the necessity of these design choices." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "text", + "content": "Design choices for in-context example selection. We have seen the remarkable performance of design intent embeddings " + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "text", + "content": "-aligned example selection, which relies on the learned latent space " + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_D" + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "text", + "content": ". To investigate alternative references, we have considered bounding boxes of design intents (" + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "text", + "content": "-aligned), sets of element types (" + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "text", + "content": "-aligned), and even with no specific alignment (Random). As reported in Tab. 6, although all " + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "text", + "content": "-aligned, " + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "text", + "content": "-aligned, and random selections make declines in average metrics, the outcomes are still superior to most baseline approaches. In particular, using " + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "text", + "content": "-aligned selection also achieves a new SOTA result, which supports its eligibility as an alternative enabling PosterO to deliver acceptable content-aware layouts when faced with roughly sketched " + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 471, + 556, + 638 + ], + "type": "text", + "content": " from users." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Adaption to customized design intent. Upon the discussions from the previous paragraph, we take a further step to verify PosterO's adaptability to customized design intents. Fig. 7 demonstrates the results of using the original " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " predicted by the detection model " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " and three customized " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " with different levels of complexity. As the input " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " varies," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8123" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 292, + 179 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 292, + 179 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 292, + 179 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 292, + 179 + ], + "type": "image", + "image_path": "d451a8f3afa3c1033b231830c533ecafbe61454190ae7c2cd9faf8656f722e51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 183, + 287, + 194 + ], + "lines": [ + { + "bbox": [ + 64, + 183, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 64, + 183, + 287, + 194 + ], + "type": "text", + "content": "Figure 7. PosterO's adaptability to different design intents " + }, + { + "bbox": [ + 64, + 183, + 287, + 194 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 64, + 183, + 287, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 201, + 294, + 381 + ], + "blocks": [ + { + "bbox": [ + 57, + 201, + 294, + 381 + ], + "lines": [ + { + "bbox": [ + 57, + 201, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 57, + 201, + 294, + 381 + ], + "type": "image", + "image_path": "da9cf6671ae11f0d46753805b19dbb4605b4cfa06335b3229ca63045e6131088.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "type": "text", + "content": "Table 7. Ablation study on " + }, + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 388, + 294, + 399 + ], + "type": "text", + "content": " accommodating small-scale LLMs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 455 + ], + "type": "text", + "content": "PosterO dynamically adjusts the spatial distribution of elements in the generated layouts while maintaining remarkable graphic metrics. Notably, this capability is unique to PosterO, attributed to its intent-based paradigm." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "text", + "content": "Adaption to small-scale backbone LLM " + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "text", + "content": ". To address resource-limited environments and to accelerate the layout generation process, we investigate PosterO's adaptability to small-scale LLMs [35], including Llama3.2-1B (L3.2-1B) and Llama3.2-3B (L3.2-3B). As reported in Tab. 7, PosterO maintains very competitive and or even better results with the smaller " + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "text", + "content": ", showcasing its potential for more flexible and efficient applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 295, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 577 + ], + "type": "text", + "content": "4.6. Generalized Content-aware Layout Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "To verify PosterO's potential in generalized content-aware layout generation, which is presented for the first time in this work, we built a new dataset, PStylish7. It comprises 152 few-shot learning samples and 100 test images, covering seven poster categories. Each serves as a representative prototype for a segment in the spectrum of poster design purposes, encompassing artwork exhibition (Paint), cultural education (Poem), public safety (Metro), entertainment marketing (Movie), merchandising display (Menu), public advocacy (Animal), and social-media interaction (Instagram). Besides, it covers eight element types, including" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 186 + ], + "type": "table", + "html": "
OursCLPaintPoemMetroMovieMenuAnimalInsta.
Ove ↓0.00470.00340.06180.01060.00830.00320.0010
Int ↓0.76900.21980.69870.12160.23590.30200.2674
Sal ↓3.88380.84841.36040.32750.71912.62800.6652
Avg ↓1.55250.35720.70700.15320.32110.97770.3112
OursL3PaintPoemMetroMovieMenuAnimalInsta.
Ove ↓0.00320.00280.07770.00040.00560.14950.0004
Int ↓0.62190.62740.74780.33750.16200.53831.1431
Sal ↓2.45120.16001.11110.13470.57902.92841.0810
Avg ↓1.02540.26340.64550.15750.24891.20540.7415
", + "image_path": "3db045aaa7fa609e61be95283a46d273c2bae66fb24d90f812290adffc3be709.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 343, + 190, + 524, + 201 + ], + "lines": [ + { + "bbox": [ + 343, + 190, + 524, + 201 + ], + "spans": [ + { + "bbox": [ + 343, + 190, + 524, + 201 + ], + "type": "text", + "content": "Table 8. Quantitative results on PStylish7 dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 207, + 553, + 257 + ], + "blocks": [ + { + "bbox": [ + 316, + 207, + 553, + 257 + ], + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 257 + ], + "type": "image", + "image_path": "5d3404882d1ea2e5be60dd528fbd27c8d741a0bf8c203d39b15cfd1d0b60e398.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 258, + 523, + 269 + ], + "lines": [ + { + "bbox": [ + 344, + 258, + 523, + 269 + ], + "spans": [ + { + "bbox": [ + 344, + 258, + 523, + 269 + ], + "type": "text", + "content": "Figure 8. Visualized results on PStylish7 dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "text", + "content": "all those in the existing datasets and four text variants depicted in Sec. 3.1. Considering feasibility, we employ five metrics, including " + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "inline_equation", + "content": "Ove\\downarrow" + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "inline_equation", + "content": "(Cov\\uparrow, Con\\downarrow)" + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "text", + "content": ", standardized as " + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "inline_equation", + "content": "Int\\downarrow" + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "inline_equation", + "content": "(Uti\\uparrow, Occ\\downarrow)" + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "text", + "content": ", standardized as " + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "inline_equation", + "content": "Sal\\downarrow" + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "text", + "content": ". More details about PStylish7 are given in supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "text", + "content": "The quantitative results are reported in Tab. 8, and the visualized results are presented in Fig. 8. Here, the detection model " + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "inline_equation", + "content": "S_{pc}" + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "text", + "content": " is employed to obtain design intents " + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "text", + "content": " and their features " + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "text", + "content": ". While PosterO" + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "inline_equation", + "content": "_{L3}" + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "text", + "content": " outperforms PosterO" + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "inline_equation", + "content": "_{CL}" + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "text", + "content": " in most categories, as the first attempt in this field, there is still plenty of room for improvement. Overall, the proposed PStylish7 is significantly more challenging than the existing datasets, attributed to its broader and sparser data distribution. This requires approaches with good generalization capabilities, mirroring real-world scenarios, and can stimulate more advanced work in the future." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 478, + 466, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 466, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 466, + 490 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "This work presented a tuning-free language model-based approach for content-aware layout generation, named PosterO. Through extensive experiments, we demonstrated that PosterO achieves new state-of-the-art results on various benchmarks, mainly attributed to the structural representation, layout trees. Moreover, we delved into the task under generalized settings and built the first available dataset, PStylish7, which covers seven poster purposes and eight element types. Employing PStylish7, we verified PosterO's potential to solve complicated tasks in real-world scenarios, underscoring its practicality. The future work mainly lies in two aspects. First, enhancing the generalized capabilities of layout generation approaches is highly valuable, as PosterO aims. With the challenging test offered by the proposed PStylish7 dataset, it is increasingly attainable. Second, integrating feedback mechanisms is envisioned to facilitate self-optimization and allow the incorporation of user suggestions, forming a more interactive, user-centric process." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8124" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 296, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 296, + 128 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (61925201, 62132001, 62432001) and Beijing Natural Science Foundation (L247006)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 159, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 204 + ], + "type": "text", + "content": "[1] Diego Martin Arroyo, Janis Postels, and Federico Tombari. Variational transformer networks for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13642-13652, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 294, + 239 + ], + "type": "text", + "content": "[2] Canva. Visual economy report 2024 edition: How visual communication fuels new business opportunities. https://www.canva.com/visual-economy-report/, 2024.1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 240, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 294, + 273 + ], + "type": "text", + "content": "[3] Canva. 10 highlights from canva create 2024: Redesigned for work. https://www.canva.com/newsroom/news/what-happened-at-canva-create-2024/, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "type": "text", + "content": "[4] Yunning Cao, Ye Ma, Min Zhou, Chuanbin Liu, Hongtao Xie, Tiezheng Ge, and Yuning Jiang. Geometry aligned variational transformer for image-conditioned layout generation. In Proceedings of the ACM International Conference on Multimedia, pages 1561-1571, 2022. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 294, + 385 + ], + "type": "text", + "content": "[5] Yunning Cao, Chuanbin Liu, Ye Ma, Min Zhou, Tiezheng Ge, Yuning Jiang, and Hongtao Xie. Self-refined variational transformer for image-conditioned layout generation. International Journal of Machine Learning and Cybernetics, pages 1-18, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 387, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 294, + 441 + ], + "type": "text", + "content": "[6] Shang Chai, Liansheng Zhuang, and Fengying Yan. Layout: Transformer-based diffusion model for layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18349-18358, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 294, + 487 + ], + "type": "text", + "content": "[7] Shang Chai, Liansheng Zhuang, Fengying Yan, and Zihan Zhou. Two-stage content-aware layout generation for poster designs. In Proceedings of the ACM International Conference on Multimedia, pages 8415-8423, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 488, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 294, + 543 + ], + "type": "text", + "content": "[8] Jian Chen, Ruiyi Zhang, Yufan Zhou, Jennifer Healey, Jiuxiang Gu, Zhiqiang Xu, and Changyou Chen. Textlap: Customizing language models for text-to layout planning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing: Findings, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 544, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 294, + 597 + ], + "type": "text", + "content": "[9] Erik Dahlstrom, Anthony Grasso, Niklas Hagelroth, Chris Lilley, Cameron McCormack, and Doug Schepers. Scalable vector graphics (svg) 2.0 specification. SVG working group, w3c. https://dev.w3.org/SVG/profiles/2.0/publish/, 2002. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 633 + ], + "type": "text", + "content": "[10] Leichao Dai, Lin Feng, Xinglin Shang, and Han Su. Cross modal adaptive few-shot learning based on task dependence. Chinese Journal of Electronics, 32(1):85-96, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "text", + "content": "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "text", + "content": "[12] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "text", + "content": "William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Proceedings of the International Conference on Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 119, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 185 + ], + "type": "text", + "content": "[13] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. In Proceedings of the International Conference on Neural Information Processing Systems, pages 50742-50768, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 240 + ], + "type": "text", + "content": "[14] Shunan Guo, Zhuochen Jin, Fuling Sun, Jingwen Li, Zhaorui Li, Yang Shi, and Nan Cao. Vinci: an intelligent graphic design system for generating advertising posters. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1-17, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 243, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 297 + ], + "type": "text", + "content": "[15] Kamal Gupta, Justin Lazarow, Alessandro Achille, Larry S Davis, Vijay Mahadevan, and Abhinav Shrivastava. Layout-transformer: Layout generation and completion with self-attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1004–1014, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "type": "text", + "content": "[16] Sun Haoran, Wang Yang, Liu Haipeng, and Qian Biao. Fine-grained cross-modal fusion based refinement for text-to-image synthesis. Chinese Journal of Electronics, 32(6): 1329-1340, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 343, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 398 + ], + "type": "text", + "content": "[17] Daichi Horita, Naoto Inoue, Kotaro Kikuchi, Kota Yamaguchi, and Kiyoharu Aizawa. Retrieval-augmented layout transformer for content-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 67-76, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "text", + "content": "[18] HsiaoYuan Hsu, Xiangteng He, and Yuxin Peng. Densitylayout: Density-conditioned layout gan for visual-textual presentation designs. In Proceedings of the International Conference on Image and Graphics, pages 187-199, 2023. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 520 + ], + "type": "text", + "content": "[19] HsiaoYuan Hsu, Xiangteng He, Yuxin Peng, Hao Kong, and Qing Zhang. Posterlayout: A new benchmark and approach for content-aware visual-textual presentation layout. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6018-6026, 2023. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "text", + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 568, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 555, + 623 + ], + "type": "text", + "content": "[21] Naoto Inoue, Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Layoutdm: Discrete diffusion model for controllable layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10167-10176, 2023. 6, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 624, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 555, + 689 + ], + "type": "text", + "content": "[22] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the International Conference on Machine Learning, pages 4904-4916, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[23] Kotaro Kikuchi, Edgar Simo-Serra, Mayu Otani, and Kota Yamaguchi. Constrained graphic layout generation via la" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8125" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "text", + "content": "tent optimization. In Proceedings of the ACM International Conference on Multimedia, pages 88-96, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "text", + "content": "[24] Suhas Kotha, Jacob Mitchell Springer, and Aditi Raghunathan. Understanding catastrophic forgetting in language models via implicit inference. In The Twelfth International Conference on Learning Representations, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[25] Hsin-Ying Lee, Lu Jiang, Irfan Essa, Phuong B Le, Haifeng Gong, Ming-Hsuan Yang, and Weilong Yang. Neural design network: Graphic layout generation with constraints. In Proceedings of the European Conference on Computer Vision, pages 491-506, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "type": "text", + "content": "[26] Chenhui Li, Peiying Zhang, and Changbo Wang. HarmOnious textual layout generation over natural images via deep aesthetics learning. IEEE Transactions on Multimedia, 2021. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 304 + ], + "type": "text", + "content": "[27] Fengheng Li, An Liu, Wei Feng, Honghe Zhu, Yaoyu Li, Zheng Zhang, Jingjing Lv, Xin Zhu, Junjie Shen, Zhangang Lin, et al. Relation-aware diffusion model for controllable poster layout generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 1249-1258, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 358 + ], + "type": "text", + "content": "[28] Jianan Li, Jimei Yang, Aaron Hertzmann, Jianming Zhang, and Tingfa Xu. LayoutGAN: Synthesizing graphic layouts with vector-wireframe adversarial networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(7): 2388-2399, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "type": "text", + "content": "[29] Jianan Li, Jimei Yang, Jianming Zhang, Chang Liu, Christina Wang, and Tingfa Xu. Attribute-conditioned layout GAN for automatic graphic design. IEEE Transactions on Visualization and Computer Graphics, 27(10):4039-4048, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "text", + "content": "[30] Yinan Li, Jia Chen, Yin Bai, Jia Cheng, and Jun Lei. Design element aware poster layout generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 1296-1305, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[31] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. Layoutprompt: Awaken the design ability of large language models. In Proceedings of the International Conference on Neural Information Processing Systems, pages 43852-43879, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 579 + ], + "type": "text", + "content": "[32] Jinpeng Lin, Min Zhou, Ye Ma, Yifan Gao, Chenxi Fei, Yangjian Chen, Zhang Yu, and Tiezheng Ge. Autoposter: A highly automatic and content-aware design system for advertising poster generation. In Proceedings of the ACM International Conference on Multimedia, pages 1250-1260, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "text", + "content": "[33] Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. arXiv preprint arXiv:2308.08747, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "text", + "content": "[34] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. In Proceedings of the International Conference on Neural Information Processing Systems, pages 1877-1901, 2020. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[35] Meta. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/, 2024. 2, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[36] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 204 + ], + "type": "text", + "content": "[37] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In Proceedings of the International Conference on Neural Information Processing Systems, pages 27730-27744, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 205, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 259 + ], + "type": "text", + "content": "[38] Yi-Hao Peng, Faria Huq, Yue Jiang, Jason Wu, Amanda Xin Yue Li, Jeffrey Bigham, and Amy Pavel. Dreamstruct: Understanding slides and user interfaces via synthetic data generation. In Proceedings of the European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[39] Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7479-7489, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "text", + "content": "[40] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In Proceedings of the European Conference on Computer Vision, pages 38-56, 2022. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "text", + "content": "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Intervention, pages 234-241, 2015. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "text", + "content": "[42] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1, 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 471, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 547 + ], + "type": "text", + "content": "[43] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In Proceedings of the International Conference on Neural Information Processing Systems, pages 25278–25294, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "text", + "content": "[44] Jaejung Seol, Seojun Kim, and Jaejun Yoo. Posterllama: Bridging design ability of language model to content-aware layout generation. In Proceedings of the European Conference on Computer Vision, pages 451-468, 2024. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 593, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 668 + ], + "type": "text", + "content": "[45] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF WInter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[46] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. Lay-outnuwa: Revealing the hidden layout expertise of large language models. Proceedings of International Conference on Learning Representations, 2024. 1, 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8126" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 619 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[48] Bo Wang, Quan Chen, Min Zhou, Zhiqiang Zhang, Xiaogang Jin, and Kun Gai. Progressive feature polishing network for salient object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12128-12135, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[49] Shaodong Wang, Yunyang Ge, Liuhan Chen, Haiyang Zhou, Qian Wang, Xinhua Cheng, and Li Yuan. Prompt2poster: Automatically artistic chinese poster creation from prompt only. In Proceedings of the ACM International Conference on Multimedia, pages 10716-10724, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 295 + ], + "type": "text", + "content": "[50] Haohan Weng, Danqing Huang, Yu Qiao, Zheng Hu, ChinYew Lin, Tong Zhang, and CL Chen. Desigen: A pipeline for controllable design template generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12721-12732, 2024. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 361 + ], + "type": "text", + "content": "[51] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Proceedings of the International Conference on Neural Information Processing Systems, pages 12077-12090, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 363, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 294, + 418 + ], + "type": "text", + "content": "[52] Chenchen Xu, Min Zhou, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Unsupervised domain adaption with pixel-level discriminator for image-aware layout generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10114-10123, 2023. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 419, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 419, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 419, + 294, + 474 + ], + "type": "text", + "content": "[53] Tao Yang, Fan Wang, Junfan Lin, Zhongang Qi, Yang Wu, Jing Xu, Ying Shan, and Changwen Chen. Toward human perception-centric video thumbnail generation. In Proceedings of the ACM International Conference on Multimedia, pages 6653-6664, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "text", + "content": "[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "type": "text", + "content": "[55] Min Zhou, Chenchen Xu, Ye Ma, Tiezheng Ge, Yuning Jiang, and Weiwei Xu. Composition-aware graphic layout GAN for visual-textual presentation designs. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 4995-5001, 2022. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "type": "text", + "content": "[56] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8127" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Potential Field Based Deep Metric Learning/c5f58da8-3e8b-4fa3-accf-6f5c33627040_content_list.json b/2025/Potential Field Based Deep Metric Learning/c5f58da8-3e8b-4fa3-accf-6f5c33627040_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..075ec34c394b94248ebea4f47d1a9d5d1758fd8e --- /dev/null +++ b/2025/Potential Field Based Deep Metric Learning/c5f58da8-3e8b-4fa3-accf-6f5c33627040_content_list.json @@ -0,0 +1,1541 @@ +[ + { + "type": "text", + "text": "Potential Field Based Deep Metric Learning", + "text_level": 1, + "bbox": [ + 276, + 130, + 722, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shubhang Bhatnagar Narendra Ahuja University of Illinois Urbana-Champaign", + "bbox": [ + 328, + 181, + 666, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sb56，n-ahuja}@illinois.edu", + "bbox": [ + 372, + 219, + 620, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep metric learning (DML) involves training a network to learn a semantically meaningful representation space. Many current approaches mine $n$ -tuples of examples and model interactions within each tuplets. We present a novel, compositional DML model that instead of in tuples, represents the influence of each example (embedding) by a continuous potential field, and superposes the fields to obtain their combined global potential field. We use attractive/repulsive potential fields to represent interactions among embeddings from images of the same/different classes. Contrary to typical learning methods, where mutual influence of samples is proportional to their distance, we enforce reduction in such influence with distance, leading to a decaying field. We show that such decay helps improve performance on real world datasets with large intra-class variations and label noise. Like other proxy-based methods, we also use proxies to succinctly represent sub-populations of examples. We evaluate our method on three standard DML benchmarks-Cars-196,CUB-200-2011,and SOP datasets where it outperforms state-of-the-art baselines.", + "bbox": [ + 89, + 300, + 485, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 632, + 220, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The goal of visual metric learning is to learn a representation function where common distance metrics like the $\\ell_2$ distance capture similarity between images. Semantic similarity is useful for applications such as image/video retrieval and search [3, 4, 21, 34], open set classification and segmentation [6, 14, 32, 37], few-shot learning [41, 45], person re-identification[7, 54] and face verification [10, 17, 31]. Deep neural networks have helped make significant advances in visual metric learning [40], as they are trained to learn a non-linear mapping of images onto a semantic space.", + "bbox": [ + 89, + 657, + 482, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing Methods: Popular loss functions to train such networks help force tuples of examples/samples/embeddings of the same class to be closer than those in different classes. Examples include the triplet loss [52], the contrastive loss [8] and their extensions [42, 51].", + "bbox": [ + 89, + 809, + 482, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The utility of supervision provided by samples in a tu", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plet depends on samples' relative locations and labels. Most tuplet-based methods use sample mining strategies [15, 40] to choose a subset of informative tuplets to train the network, without which they suffer from lower performance and slower convergence [53]. Such strategies not only require the computationally demanding calculation of the relative hardness of all tuples formed $(\\mathrm{O}(n^{3}))$ for triplets among $n$ samples), but also incur a loss of valuable supervision by discarding interactions. Specifically, limiting interactions to only a subset of all possible tuplets leads to greater susceptibility to noise [25], overfitting on the training set [26], and reduced learning of intra-class variations and generalizable features [38].", + "bbox": [ + 511, + 268, + 906, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Proxy-based methods, first proposed in [34] attempt to mitigate some of these shortcomings by instead using class representative 'proxies' as one or more hypothesized centers for each class. Computation of the loss is then in terms of sample-proxy distances instead of sample-sample distances, helping circumvent the problem of tuplet sampling. However, these methods have been found to learn less generalizable intra-class features due to the exclusive use of proxy-sample distances while ignoring sample-sample distances[39]. This effect is compounded when proxies, whose locations are themselves learned, end up far from sample embedding locations they are intended to represent.", + "bbox": [ + 511, + 470, + 908, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Proposed Representation: To address these challenges, we propose Potential Field based Deep Metric learning (PFML), a metric learning framework that uses a continuous, potential field representation of all data samples motivated by potential fields used to model influence of charges in physics. Besides its analogical relevance, our motivation for using this representation comes from its previous, very successful uses, where it also helps overcome complexity in modeling interactions. These applications include planning paths of a robot avoiding obstacles [19] and planning an efficient strategy for a robot manipulator to grasp an object [20].", + "bbox": [ + 511, + 654, + 908, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In PFML, we view each data sample as a charge, creating a field. The fields due to the individual samples are added (superposed) to obtain the global potential field. This global field thus constitutes a compositional representation, cap-", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25549", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "turing the combined influence at a given location (or sample) as a weighted addition of the influences of the individual samples. Each sample generates two types of fields: (1) an attraction field to drive similar examples closer to it, and (2) a repulsion field that drives dissimilar examples away from it, each weakening with distance.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We also make use of proxies like previous work, but only to augment the potential field generated by the current batch of samples. This is in contrast to past work that completely replaces sample-sample interactions with sample-proxy interactions. Preserving sample-sample relations through the potential field helps learn better intra-class features.", + "bbox": [ + 89, + 181, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advantages of Potential Field: (1) Using the potential field representation enables us to model interactions among all sample embeddings, as opposed to modeling those among smaller number of samples (or proxies) as done in methods using point-tuplet based (e.g., triplet) loss. (2) Addition of potentials due to all points increases robustness to noise since the effect of noise on interactions among a smaller number of samples will have a larger variance while also (3) Considering all interactions improves the quality of features learned, unlike tuplet based methods which lose performance when all interactions are considered[53]. (4) This is because of a major feature that differentiates our potential field from past methods: The variation of strength of interaction between two points as the distance between them increases: instead of remaining constant or even becoming stronger, as is the case with most existing methods, e.g. [8, 21], in our model it becomes weaker with distance. The decay property is helpful in several ways: (4a) It ensures the intuitive expectation that two distant positive samples are too different to be considered as variants of each other, helping treat them as different varieties (e.g., associated with different proxies). (4b) The decay property also significantly improves PFML performance for the specific type of noise affecting labels, e.g., due to annotation errors common in real-world datasets (Sec. 4.3). (4c) As a result of the decay, the learned proxies remain closer to (at smaller Wasserstein distance $W_{2}$ from) the sample embeddings they represent than for (e.g., current proxy-based) methods where interactions strengthen with distance [21, 34, 36]; Sec 3.4), thereby enhancing their desired role.", + "bbox": [ + 91, + 273, + 483, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions: The main contributions of the proposed approach are:", + "bbox": [ + 89, + 743, + 482, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present PFML, a novel framework for metric learning with a compositional representation using a continuous, potential field to model all sample interactions instead of only interactions between a subset of all possible n-tuplets. This enables better feature learning, especially", + "bbox": [ + 89, + 773, + 482, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in the presence of annotation noise.", + "bbox": [ + 526, + 90, + 763, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- PFML reverses the almost universally used model of inter-sample interactions strengthening with distance; interaction strength instead weakens with distance. This (1) helps improve robustness in the presence of label noise, and (2) we prove that it helps better align the distribution of proxies with the sample embeddings they represent.", + "- We evaluate PFML on three zero-shot image retrieval benchmarks: Cars-196, CUB-200-2011, and the SOP dataset where it outperforms current, state-of-the-art techniques in both the standard no-label-noise scenario, and with even greater margins ( $>7\\%$ R@1) in the more realistic cases of label noise." + ], + "bbox": [ + 511, + 106, + 906, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 306, + 653, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Broadly, DML may be divided into two types, ranking/pair-based and proxy-based [21, 51].", + "bbox": [ + 511, + 335, + 903, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ranking/Pair based DML: These methods utilize losses that derive information from relations between intersample distances in pairs/ tuples of sample embeddings. Examples include the classical contrastive loss [8] or the triplet loss[40, 52, 53], which extends the concept to take into account relative distances in a triplet. [42] and [43] generalize these constraints to include multiple dissimilar samples. Mining informative pairs/tuples from a mini-batch plays an important role in the effectiveness of pair-based methods. Several strategies [15, 40, 57] have been proposed to pick samples based on their relative difficulty, with hard and semi-hard negative (picking pair of inter-class examples close together) and easy positive (same class examples close together) [25, 55] being popular. PFML naturally emphasizes such local (dis) similarities by virtue of its decaying formulation without the complexity of mining tuples. Also note that unlike PFML, none of these strategies optimize all interactions among a batch of samples.", + "bbox": [ + 511, + 367, + 906, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Proxy-based DML: ProxyNCA [34] and its variants [23, 48] attempt to minimize the distance of a sample from the proxy of its class while maximizing the distance from proxies of other classes. Proxy Anchor [21] builds on these by proposing a modified loss that takes into account sample hardness. [10, 36] introduced the use of multiple proxies per class to better model sub-clusters.", + "bbox": [ + 511, + 641, + 905, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Potential Field: Potential fields have been used in a variety of applications to model interactions of objects. Of interest here is their use in robot path planning [9, 19], shape representation [2, 18], image segmentation [1, 47] and object manipulation [20]. Here, they have been used to direct a robot towards its goal using an attraction field while avoiding obstacles using repulsion fields. Training supervision for DML shares a similar goal of moving sample embeddings towards other nearby samples of the same class, while avoiding dissimilar examples belonging to other classes.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1It remains constant over a small radius and decays outside; however, for brevity, we will disregard this detail in the rest of this paper, and refer to the influence as decaying with distance, or as the decay property.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25550", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/813a457e2650901ca8de8d420e0881ffa96b784bbe17ed61fbed1b82b643c9ef.jpg", + "image_caption": [ + "Figure 1. Overview of our Potential-field based DML pipeline. The process includes (1) Computing attraction and repulsion fields generated by each embedding and proxy, (2) Computing the class potential fields by superposition of individual fields (3) Evaluating total potential energy by summing up the potentials of embeddings and proxies under the class potential field and (4) Updating locations of sample embeddings (through network parameters) and proxies to minimize total potential energy through backprop." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 480, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 421, + 181, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setup", + "text_level": 1, + "bbox": [ + 89, + 446, + 241, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of deep metric learning is to define a semantic distance metric $d(\\mathbf{x}_1, \\mathbf{x}_2)$ for any pair of samples $\\mathbf{x}_1, \\mathbf{x}_2 \\in \\mathcal{D}$ , where $\\mathcal{D} = \\{\\mathbf{x}_i, y_i\\}$ , $i \\in \\{1, \\dots, |\\mathcal{D}|\\}$ and labels $y_i \\in \\{1, \\dots, N\\}$ represent $N$ classes. The learned distance metric for any two points $d(\\mathbf{x}_1, \\mathbf{x}_2)$ should represent their semantic dissimilarity. Instead of directly learning the distance metric, a projector function $f_\\theta$ is often represented by a neural network, parameterized by $\\theta$ learned from data and distance $d(\\mathbf{x}_1, \\mathbf{x}_2) = \\|f_\\theta(\\mathbf{x}_1), f_\\theta(\\mathbf{x}_2)\\|_2$ is given by the Euclidean distance between the projections/embeddings. For simplicity, we will refer to the embeddings $f_\\theta(\\mathbf{x}_i)$ belonging to a batch $\\mathcal{B}$ sampled from $\\mathcal{D}$ as $\\mathbf{z}_i$ .", + "bbox": [ + 89, + 470, + 483, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. The Potential Field", + "text_level": 1, + "bbox": [ + 89, + 664, + 272, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With our potential field representation, embeddings need to be driven towards other nearby embeddings belonging to the same class, while also being driven away from embeddings of other classes. This is reminiscent of the behavior of an isolated system of electric charges, where dissimilar charges are drawn together while similar ones are repelled. Analogously, we define attractive and repulsive potentials created by a single embedding, with the net potential generated by a set of samples given by superposing individual potentials exerted by each embedding. The gradient of the net potential at a point yields the net force (on network update) on a sample embedding located at that point.", + "bbox": [ + 89, + 686, + 482, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As samples of each class have a different set of similar and dissimilar points, we define a separate potential field", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\Psi_{j}$ , for each class $j \\in \\{1, \\dots, N\\}$ . The class potential field $\\Psi_{j}$ models the force applied on embeddings belonging to class $j$ due to all embeddings. It is formed from a superposition of potentials belonging to sample embeddings from all classes, with embeddings from the same class exerting an attractive potential field (to pull similar points closer) while those from other classes exert a repulsive potential field (to push apart dissimilar points). We assume that each sample embedding $z_{i}$ has a unit charge, so the net force on it is the gradient of the potential field $\\Psi_{y_i}$ of its class. We design the attractive and repulsive potential functions by taking motivation from both, electrostatic fields and metric learning literature, which we discuss in detail in the following subsections. Figure 1 provides an overview of our complete pipeline of generation and use of the potential field for metric learning.", + "bbox": [ + 511, + 90, + 906, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Potential field exerted by a single point", + "text_level": 1, + "bbox": [ + 511, + 340, + 828, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By using forces on the embeddings, we wish the training process to move the embeddings to a state of equilibrium at the end of training in which all embeddings from a class are nearby. To have a potential function exhibiting this behavior, we ensure that individual potential field has the following two properties:", + "bbox": [ + 511, + 364, + 905, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Property 1 - Extremum Property: Potential generated by an individual embedding achieves an extremum at the embedding (minimum for attraction field, maximum for repulsion). This ensures that similar points are pulled together and dissimilar ones are pushed apart on minimizing potential.", + "bbox": [ + 511, + 455, + 905, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Property 2 - Decay Property: The magnitude of the gradient of the potential (force) due to an embedding decays with distance. Such a decay ensures two desired characteristics. (1) An embedding is not pulled towards a distant sample embedding belonging to the same class, which is likely to be too different to be considered as a variant of it; instead, the decayed influence helps treat the distant sample as significantly different variety of the class, associated with other nearby embeddings of the class. (2) Like for a single charge in the Extremum property, a superposition of potentials of different embeddings still yields local minima at or near sample embeddings. As shown in Sec 3.4, this helps proxies come close to such minima and thus help better model the data distribution.", + "bbox": [ + 511, + 545, + 906, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given these properties motivated by electrostatic fields, we select the attraction and repulsion potentials $\\psi_{att},\\psi_{rep}$ generated by an embedding $\\mathbf{z}_i$ , with their strength at location $\\mathbf{r}$ defined as:", + "bbox": [ + 511, + 756, + 905, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {a t t} (\\mathbf {r}, \\mathbf {z} _ {i}): = \\left\\{ \\begin{array}{l l} - \\frac {1}{\\delta^ {\\alpha}} & \\text {i f} \\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} < \\delta \\\\ - \\frac {1}{\\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} ^ {\\alpha}} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 834, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25551", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2e7644140292bd2441b2028d58a4b5c4b51a0f0ea687f50f5cc7a25ec850fc17.jpg", + "image_caption": [ + "(a) Class Potential field $\\Psi_{1}$" + ], + "image_footnote": [], + "bbox": [ + 148, + 114, + 442, + 339 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2268a3e65bdeb8101e6be370584568f4ed0ab166f27ed7ba22e3b0ce06f93d4b.jpg", + "image_caption": [ + "(b) Class Potential field $\\Psi_{2}$", + "Figure 2. An example of the class potential fields $\\Psi_{1}$ and $\\Psi_{2}$ (Sec. 3.2.3), created by superposing the fields of individual embeddings (Sec 3.2.1) belonging to Classes 1 and 2. Arrows denote the gradient of the respective potentials representing the net force on them. (a) $\\Psi_{1}$ draws samples/proxies of class 1 towards nearby samples/ proxies of class 1 while keeping them at least $\\delta$ distance away from embeddings of class 2. Proxies at starred locations are drawn towards the nearest Class 1 embeddings which are potential minima, helping them better model the data distribution (Sec 3.4). (b) $\\Psi_{2}$ draws embeddings of class 2 towards other nearby embeddings of class 2 instead of distant class 2 embeddings, which might be a significantly different variant of the class (or potentially a mislabeled data point), helping better feature learning (or improving noise robustness if mislabeled). $\\Psi_{2}$ also keeps class 2 embeddings $\\delta$ distance them away from embeddings of class 1." + ], + "image_footnote": [], + "bbox": [ + 457, + 113, + 856, + 338 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {r e p} (\\mathbf {r}, \\mathbf {z} _ {i}): = \\left\\{ \\begin{array}{l l} \\frac {1}{\\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} ^ {\\alpha}} & \\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} < \\delta \\\\ \\frac {1}{\\delta^ {\\alpha}} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 474, + 482, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\alpha$ is a hyperparameter that determines the rate of decay of the field (we evaluate it's effect in Sec. 4.5 instead of fixing $\\alpha = 1$ as done in electrostatic fields). Note that the attraction potential is negative and increases with distance, implying that the force between the embedding and the point is attractive. The repulsion potential is positive and decreasing with distance, indicating repulsion.", + "bbox": [ + 88, + 551, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A constant attraction potential inside the $\\delta$ sphere (i.e., force $= 0$ ) ensures that sample embeddings within $\\delta$ distance of one another do not continue to attract each other, thus preventing the network from embeddings from collapsing to a single point which harms generalizability of the features learned (analogous to the effect margins used in tuplet-based losses have).", + "bbox": [ + 88, + 657, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Limiting the repulsion potential to a sphere of radius $\\delta$ ensures that embeddings not belonging to the same class are driven to a distance $> \\delta$ . Since the aforementioned constant attraction potential ensures that $\\delta$ is the maximum distance between samples $z_{i}$ from the same class, this leads to connected components of $\\delta$ spheres formed by nearby embeddings of the same class. Extending repulsion to distances $> \\delta$ does not provide any useful additional supervisory information while forcing further unmotivated inter-class sep", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "aration; as we empirically show in Sec. 4.5, such excessive separation does indeed harm performance.", + "bbox": [ + 511, + 478, + 906, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Proxies to Represent Sample Populations", + "text_level": 1, + "bbox": [ + 511, + 515, + 844, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To represent the image embeddings not sampled in the current batch $\\mathcal{B}$ , we propose to use a set of $M$ proxies per class for each of the N classes, $\\mathbf{p}_{j,k}, k \\in \\{1, \\dots, M\\}, j \\in \\{1, \\dots, N\\}$ , and augment the potential field generated by the embeddings in the batch by those generated by the proxies. The proxies serve as stand-ins for the larger numbers of out-of-the-batch embeddings for calculating the attraction and repulsion fields, while also being affected by the field themselves like all embeddings. The proxies, being learnable parameters with high learning rates, tend to gravitate toward the nearest class modes of the potential field (and hence the class data distribution). The number of proxies M chosen should be large enough to capture the class-modes.", + "bbox": [ + 511, + 534, + 906, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. Class Potential Field", + "text_level": 1, + "bbox": [ + 511, + 737, + 702, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a batch $\\mathcal{B}$ of samples and all proxies, we calculate the class potential field $\\Psi_{j}(\\mathbf{r})$ at a location $\\mathbf{r}$ due to a class $j$ as a superposition of their attraction and repulsion potentials:", + "bbox": [ + 511, + 757, + 905, + 803 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi_ {j} (\\mathbf {r}) = \\Psi_ {j, a t t} (\\mathbf {r}) + \\Psi_ {j, r e p} (\\mathbf {r}), \\text {w h e r e} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 814, + 906, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi_ {j, a t t} (\\mathbf {r}) = \\sum_ {i = 1} ^ {\\| \\mathcal {B} \\|, y _ {i} = j} \\psi_ {a t t} (\\mathbf {r}, \\mathbf {z} _ {i}) + \\sum_ {k = 1} ^ {M} \\psi_ {a t t} (\\mathbf {r}, \\mathbf {p} _ {j, k}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 854, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25552", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi_ {j, r e p} (\\mathbf {r}) = \\sum_ {i = 1} ^ {\\| \\mathcal {B} \\|, y _ {i} \\neq j} \\psi_ {r e p} (\\mathbf {r}, \\mathbf {z} _ {i}) + \\sum_ {\\gamma = 1, \\gamma \\neq j} ^ {N} \\sum_ {k = 1} ^ {M} \\psi_ {r e p} (\\mathbf {r}, \\mathbf {p} _ {\\gamma , k}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 88, + 493, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 visualizes the class potential fields $\\Psi_{1}$ and $\\Psi_{2}$ for a given set of embeddings belonging to two classes.", + "bbox": [ + 89, + 157, + 485, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 89, + 196, + 197, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using gradient descent, the network is trained to minimize the total potential energy $\\mathcal{U}$ of all proxies and embeddings in the batch $\\mathbf{z}_i\\in \\mathcal{B}$ by applying a force on embeddings and proxies in the direction of $-\\nabla \\Psi_{y_i}(\\mathbf{z}_i)$ . $\\mathcal{U}$ given by:", + "bbox": [ + 89, + 219, + 483, + 281 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {U} = \\sum_ {i = 1} ^ {\\| \\mathcal {B} \\|} \\Psi_ {y _ {i}} (\\mathbf {z} _ {i}) + \\sum_ {j = 1} ^ {N} \\sum_ {k = 1} ^ {M} \\Psi_ {j} (\\mathbf {p} _ {j, k}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 290, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Why decay interaction strength with distance?", + "text_level": 1, + "bbox": [ + 89, + 343, + 482, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PFML uses a potential field where the gradient (interaction strength) decays with distance, which is unlike previous approaches. This leads to the two benefits described in Sec 3.2.1, which we discuss in more detail here.", + "bbox": [ + 89, + 364, + 483, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1. Robustness when learning with datasets having large intra-class variations and noisy labels", + "text_level": 1, + "bbox": [ + 89, + 433, + 483, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The decay property helps PFML model all interactions even when using datasets with large intra-class variations. As seen in Figure 2, PFML draws embeddings towards other nearby embeddings from the same class rather than distant class embeddings that might be too different to be considered a variation of it. It also helps PFML become significantly more robust to the effect of incorrectly labeled examples as compared to existing methods. Recent work [26, 30] shows that such examples are common in metric learning benchmarks. The decay de-emphasizes the contribution of such unlabeled examples, that are typically located away from correctly labeled samples. We empirically confirm this in Sec. 4.3, demonstrating that methods not adhering to the decay property (i.e., all previous tuplet-based and proxy-based methods) experience significantly greater performance degradation due to label noise compared to PFML.", + "bbox": [ + 88, + 467, + 483, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. Better alignment of proxies with data distribution", + "text_level": 1, + "bbox": [ + 89, + 731, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The decaying interaction strength helps PFML better align the distribution of proxies with the underlying sample data distribution (of the complete dataset) that they represent. To analyze this, we compare PFML with the closely related contrastive loss [8] that also models pairwise interactions like PFML (defined in Eq. 1 & 2), and hence can be analyzed similarly by using a potential field. Specifically, we introduce Contrastive Potential based Metric Learning (CPML) by simply extending the contrastive loss by using multiple proxies per class, and subsequently analyzing the", + "bbox": [ + 88, + 750, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "resulting loss function through a potential field framework. In CPML, strength of interaction is directly proportional to the square of the distance from it $(\\|\\mathbf{r} - \\mathbf{r}_0\\|_2^2)$ like in the original contrastive loss. For fair comparison, we keep all other details (use of $\\delta$ , proxies) as the same, with its associated potential fields denoted by $\\Psi_{att}^*$ , $\\Psi_{rep}^*$ (formally defined in Supplement Sec 1). We analyze this theoretically using the following proposition and its corollary:", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 1: Let $Z = \\{z_1 \\ldots z_n\\}$ be a set of sample embeddings belonging to a class, then there exists a $0 < \\delta < \\frac{\\min_{i,j} \\|z_i - z_j\\|_2}{2\\left(1 + \\frac{1}{n}\\right)}$ and points $z_{min,i}$ within $\\delta$ distance from each embedding $z_i$ , $i \\in \\{1 \\ldots n\\}$ such that the attractive potential field $\\Psi_{att}$ defined using $Z$ , $\\delta$ has a minimum at each $z_{min,i}$ . The field $\\Psi_{att}^*$ defined by them does not achieve a minimum at points within $\\delta$ distance from all $z_i$ .", + "bbox": [ + 511, + 215, + 908, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 1 is a consequence of the potential field of an embedding dominating in its vicinity over potential fields of other distant embeddings. Its proof relies on the Decay property, which is not satisfied by $\\Psi^{*}$ and can be found in Supp. Sec 1. A consequence of this is:", + "bbox": [ + 511, + 358, + 906, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Corollary 1: Let $Z = \\{z_{1}\\dots z_{n}\\}$ be a set of sample embeddings from a class exerting an attraction field on a set of proxies $P = \\{\\pmb {p}_1\\dots \\pmb {p}_m\\}$ . Consider the equilibrium distribution $P_{eq}$ of proxies minimizing the potential energy. If the potential field is defined by $\\Psi_{att}$ , then the Wasserstein distance $W_{2}$ between $P_{eq}$ and the subset of data they represent is lower than when the potential field is defined by $\\Psi_{att}^*$ .", + "bbox": [ + 511, + 436, + 908, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proof for Corollary 1 relies on Proposition 1 and can be found in the Supplement (Sec. 2).", + "bbox": [ + 511, + 547, + 906, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion: As proved in Corollary 1, CPML leads to proxies having a larger distance from the data distribution they represent (measured by $W_{2}$ ) than our proposed PFML, making the use of proxies in such a framework less effective. This leads to the proxies forming poorer stand-ins for the sample embeddings they represent, reducing the quality of features learned (verified in Sec. 4). This is caused by CPML not following the decay property (which is its only difference in design wrt PFML). Other proxy-based losses like the Proxy NCA [34], Soft Triplet [36], and Proxy Anchor [21] which perform similarly as CPML (see Tab. 1), also share with it the property of force increasing with distance (violating the decay property). Due to this, we expect them to similarly have a poorer alignment of proxies with underlying data distribution as compared to PFML (Applying an analysis similar to the one in Proposition 1 to them is considerably more challenging due to the absence of a potential field structure.). We verify this empirically in Sec. 4.4 by measuring the average distance between proxies and the closest data points ( $W_{2}$ distance) for these methods.", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25553", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Advantages in Learning Fine-grained Features", + "text_level": 1, + "bbox": [ + 89, + 90, + 480, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast with all previous proxy-based methods, PFML (1) Models sample-sample interactions directly using the potential field, gaining fine-grained supervision, and (2) Makes use of an explicit margin-like $\\delta$ for proxies, which has been shown to help zero-shot generalization and fine-grained feature learning in tuplet-based methods[53]. Turning off the interaction within the $\\delta$ radius also prevents a bias towards learning isotropic distributions as is present in other proxy-based methods [39], helping us obtain better intra-class resolution.", + "bbox": [ + 89, + 117, + 482, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 89, + 297, + 323, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 329, + 171, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: We empirically compare PFML with current state-of-the-art baselines on three public benchmark datasets (1) The Cars-196 dataset [24] which includes 16,185 images from 196 car model categories, (2) CUB200-2011 dataset [49] consisting of 11,788 images from 200 bird species, and (3) Stanford Online Products (SOP) dataset [43] containing 120,053 images of 22,634 different products sold online. The training and testing splits for the three datasets follow standard zero-shot retrieval-based evaluation protocol [21, 51, 61] of using the first half of the classes for training and the second half for testing.", + "bbox": [ + 89, + 354, + 482, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Backbone:** To enable standardized comparison with a wide variety of methods, we evaluate our method using the Inception with batch normalization (IBN) [46], ResNet-50 [16], ViT [11] and DINO [5] backbones commonly used by DML methods. ImageNet pre-training is used for initialization, and only the last fully connected layer is changed to fit the dimension of the embedding needed. We follow past work [21, 36, 48] and $\\ell_2$ normalize the final output. An input image size of $224 \\times 224$ is used for all experiments.", + "bbox": [ + 89, + 527, + 482, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training parameters: We train the backbone for 200 epochs on each dataset. We use the Adam optimizer with the learning rate chosen as $5e^{-4}$ . We scale up the learning rate of our proxies by 100 times following [21, 34, 36] for better convergence. We choose values of $\\delta$ between $[0.1, 0.3]$ , $\\alpha$ between $\\{0, 6\\}$ via cross-validation. We use $M = 15$ for the CUB-200 and Cars-196 datasets, while $M = 2$ is used for the SOP dataset. All experiments are performed on a single NVIDIA A4000 GPU.", + "bbox": [ + 89, + 667, + 482, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Settings: We follow [21, 22, 29, 51] in using $224 \\times 224$ sized center crops from $256 \\times 256$ images for evaluation. We measure image retrieval performance using Recall@K, which computes the percentage of samples which have a valid similar neighbor (belonging to same class) among its K nearest neighbors.", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Image Retrieval Performance", + "text_level": 1, + "bbox": [ + 511, + 90, + 774, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As seen in Table 1, our method outperforms all other methods on Image Retrieval on the CUB-200-2011, Cars-196 and SOP datasets as measured by the Recall@K metric. It consistently outperforms other methods irrespective of the backbone used, significantly outperforming the best-performing proxy-based methods (which similar to us, only change the loss function) like ProxyAnchor[21], ProxyGML[61] and SoftTriplet [36] in terms of Recall@1 (R@1), by more than $5\\%$ on Cars-196, $3.7\\%$ on CUB-200-2011 and $1.5\\%$ on the SOP dataset. It also outperforms the significantly more complex, current state-of-the-art HIST[28] by $3.1\\%$ , $2\\%$ and $1.5\\%$ in terms of R@1 when using the Resnet-50 backbone on the Cars-196, CUB-200-2011 and SOP datasets respectively. It also significantly outperforms the best tuplet-based loss: the Multi-Similarity loss by $6\\%$ , $5.8\\%$ and $2.4\\%$ in terms of R@1 when using the BN Inception backbone on the Cars-196, CUB-200 and SOP datasets respectively. Additionally, our method outperforms CPML (Sec. 3.4.2, a simple combination of the classical contrastive loss with proxies) by $7.5\\%$ , $5.1\\%$ , and $3.5\\%$ on CUB200, Cars and SOP datasets. Note that our performance gains are larger than current SOTA compared to improvements reported by previous efforts on the same benchmarks, despite our intuitive and simpler method.", + "bbox": [ + 511, + 112, + 903, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluate our method 1) Using the protocol and metrics proposed in [35] and (2) When used to learn smaller 128-dim embeddings for more efficient retrieval. Our method still outperforms all SOTA methods in both cases. Details of these experiments are in Supp. Sec. 3. We also show qualitative retrieval results of our method in Supp. Sec 6.", + "bbox": [ + 511, + 474, + 903, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Image Retrieval in Presence of Label Noise", + "text_level": 1, + "bbox": [ + 511, + 590, + 877, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the robustness of our method vs. state-of-the-art baselines in the realistic scenario of label noise (e.g., due to mislabeling, which is very common, particularly in fine-grained datasets such as those used for DML). Specifically, we randomly corrupt labels of $20\\%$ of samples in the training set. As seen in the results in Table 2, this leads to a significant decline in the performance of all methods. But our method is the least affected by the noise. Other methods like the triplet [15] or ProxyAnchor[21] suffer much greater declines because they are affected more by distant points represented by the mislabeled examples. Our method significantly outperforms Proxy Anchor [21] (the second-best method) by $6\\%$ and $7.6\\%$ in terms of R@1 on the CUB-200-2011 and Cars-196 datasets.", + "bbox": [ + 511, + 612, + 903, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Distance Between Proxies and Sample Embeddings", + "text_level": 1, + "bbox": [ + 511, + 832, + 903, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We measure the average (over a complete epoch) Wasserstein distance $W_{2}$ between proxies and the subset of training", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25554", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e69f2d3045918d87ec0a983973cb7a967cf110879d03559b827bfc587f4d5ffb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Benchmarks →CUB-200-2011Cars-196SOP
Methods ↓ (Chronological)R@1R@2R@4R@1R@2R@4R@1R@10R@100
ResNet50 (512 dim)
ESPHN [55]64.975.383.582.789.393.078.390.796.3
N.Softmax [58]61.373.983.584.290.494.478.290.696.2
DiVA [33]69.279.3-87.692.9-79.691.2-
Proxy NCA++ [48]64.7---85.1--79.6--
Proxy Anchor [21]69.780.087.087.792.995.8---
DCML-MDW [59]68.477.986.185.291.896.079.890.895.8
MS+DAS [29]69.279.287.187.893.196.080.691.896.7
HIST [28]71.481.188.189.693.996.481.492.096.7
HIER[22]70.179.486.988.293.095.680.291.596.6
HSE-PA [56]70.680.187.189.693.896.080.091.496.3
CPML (Sec. 3.4)68.378.786.285.291.595.279.490.796.1
Potential Field (Ours)73.4 ± 0.382.4± 0.188.8± 0.192.7± 0.395.5 ± 0.197.6± 0.182.9 ± 0.292.5 ± 0.296.8 ± 0.1
BN Inception (512 dim)
HTL [13]57.168.878.781.488.092.774.888.394.8
MultiSimilarity [51]65.777.086.384.190.494.078.290.596.0
SoftTriple [36]65.476.484.584.590.794.578.390.395.9
CircleLoss [44]66.777.486.283.489.894.178.390.596.1
DiVA [33]66.877.7-84.190.7-78.190.6-
ProxyGML [61]66.677.686.485.591.895.378.090.696.2
Proxy Anchor [21]68.479.286.886.191.795.079.190.896.2
DRML-PA [60]68.778.686.386.992.195.271.585.293.0
MS+DAS [29]67.178.1186.485.791.695.378.290.396.0
HIST [28]69.780.087.387.492.595.479.691.096.2
DFML-PA [50]69.3--88.4-----
HSE-M [56]67.678.085.882.088.993.3---
PA+niV [23]69.580.0-86.492.0-79.290.4-
Potential Field (Ours)71.5± 0.381.2± 0.288.3± 0.290.1± 0.293.9± 0.196.3± 0.180.6 ± 0.391.8 ± 0.196.4 ± 0.1
DINO (384 dim)
DINO [5]70.881.188.842.953.964.263.478.188.3
Hyp [12]80.987.692.489.294.196.785.194.497.8
HIER [22]81.188.293.391.395.297.185.794.697.8
Potential Field (Ours)83.1 ± 0.389.3 ± 0.294.2 ± 0.194.7 ± 0.196.5 ± 0.197.8 ± 0.186.5 ± 0.395.1 ± 0.398.0 ± 0.2
ViT (384 dim)
ViT-S [11]83.190.494.447.860.272.262.177.789.0
Hyp [12]85.691.494.886.592.195.385.994.998.1
HIER [22]85.791.394.488.393.296.186.195.098.0
Potential Field (Ours)87.8 ± 0.292.6 ± 0.295.7 ± 0.191.5 ± 0.395.2 ± 0.297.4 ± 0.188.2 ± 0.195.7 ± 0.198.6 ± 0.1
", + "bbox": [ + 132, + 87, + 866, + 614 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/594d26ff7f09f0afe06ece9a60487d720994c6fca599f47a7950dab3fe07995c.jpg", + "table_caption": [ + "Table 1. Comparison of the Recall@ $K$ (%) achieved by our method on the CUB-200-2011, Cars-196 and SOP datasets with state-of-the-art baselines under standard settings. The table reports the average performance and standard deviations of our method over 5 runs." + ], + "table_footnote": [], + "table_body": "
CUB-200-2011Cars-196
MethodsR@1R@2R@1R@2
Triplet[52]55.168.767.577.9
MS [51]58.971.870.479.8
PNCA[34]60.174.774.382.4
PA [21]60.775.176.983.1
HIST[27]59.774.672.981.8
Ours66.7 ± 0.676.9± 0.384.5± 0.588.6± 0.3
", + "bbox": [ + 86, + 659, + 594, + 787 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/49e59f42447e35ed1546c5349f91c1f7355e3a2ab6bae4bded77d6a48f669e6b.jpg", + "table_caption": [ + "Table 2. A comparison of Recall@ $K$ (%) achieved by our method with SOTA baselines on CUB-200-2011 and Cars-196 datasets with label noise when using a Resnet-50 backbone (512-dim). Performance and std. deviations reported are over 5 runs." + ], + "table_footnote": [], + "table_body": "
MethodAverage W2
S.Triplet [36]0.24 ± 0.04
PNCA [21]0.36 ± 0.02
PA [21]0.38 ± 0.03
CPML (Sec. 3.4)0.26 ± 0.03
Ours0.16 ± 0.03
", + "bbox": [ + 638, + 674, + 883, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Average Wasserstein distance between proxies and the subset of data closest to them for different methods on CUB-200-2011 dataset. Mean and std. deviations are over 5 runs.", + "bbox": [ + 611, + 782, + 910, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "data represented by/closest to them ${}^{2}$ for different proxy-", + "bbox": [ + 89, + 854, + 482, + 872 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "based methods. We train networks on the CUB-200-2011 using the settings in Section 4.1 and use the same number", + "bbox": [ + 511, + 854, + 906, + 886 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2For methods using a single proxy per class, this equals the distance between a proxy and the closest embedding belonging to its class", + "bbox": [ + 89, + 877, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25555", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83c48e49831ecbab92c541f90b6f3a53a6191a03366d83260c4ad7d758c87a36.jpg", + "image_caption": [ + "Figure 3. Variation in Recall@1 with $M$ , $\\delta$ , $\\alpha$ and $\\delta_{rep} - \\delta_{att}$ on the Cars-196 and CUB-200-2011 datasets. Error bars represent std. deviations over 5 runs." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 297, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f78862198ec348277fa0eb0ca5d637f108e1e078ac781392dfd98b544255fa50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 88, + 496, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2189e2657748765a4f2dffad8fcc2bc9d0323c667997ac10b20e4315d8bfb76b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 691, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/92e2a474b7da5dcf709eb8a568784c588174eced910efb75b936f870f69f40fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 88, + 890, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of proxies for all methods utilizing multiple proxies. As seen in Table 3, our method shows $33.3\\%$ lower $W_{2}$ distance between the proxies and the subset of data than the second best SoftTriplet [36] loss, empirically validating our discussion in Sec. 3.4.2.", + "bbox": [ + 89, + 237, + 483, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 327, + 251, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In all experiments, we use the parameters mentioned in Section 4.1 with a ResNet-50 backbone on both the Cars-196 and CUB-200-2011 datasets.", + "bbox": [ + 89, + 351, + 483, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of proxies and their number: Proxies augment the potential field, and the number of proxies per class $M$ decides the granularity at which the proxies model the distribution of sample embeddings. We study their effect on performance by varying $M$ between [0, 30]. From the results plotted in Figure 3(a) we observe that (1) Using proxies to augment the field boosts performance (vs using no-proxies) and (2) Network performance remains stable for a wide range of M, decreasing slightly for small values of M (still outperforming most other proxy-based methods). This behavior is expected as using potential field created using a very small M would not be able to effectively model underlying sample distribution, leading to worse supervision. We also observe that using a larger number of proxies per class does not significantly affect training times, as bulk of computation is dominated by the network. For further details, please refer to Supp. Sec. 5.", + "bbox": [ + 89, + 398, + 483, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of varying $\\delta$ : We conduct an empirical study to determine the effect of varying the radius parameter $\\delta$ on performance by varying it over [0.1, 0.35]. As seen in Figure 3(b), we observe that performance remains relatively stable for a wide range of $\\delta$ values, demonstrating our method's robustness to it.", + "bbox": [ + 89, + 657, + 482, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of potential decay $\\alpha$ : The decay parameter $\\alpha$ determines the rate at which the potential field of an embedding decays with distance. We study its effect by varying it over [0, 6]. As seen in Figure 3(c), performance is worse when the potential field does not decay ( $\\alpha = 0$ ). A field that decays strongly $\\alpha \\in [3, 6]$ seems to perform better than one with a much milder decay ( $\\alpha \\in [1, 2]$ ). The likely reason for this trend is that when the potential field has no decay /decays mildly, the force on an embedding would still be dominated by a large number of distant sample embeddings", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rather than those in its vicinity.", + "bbox": [ + 511, + 238, + 718, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Restricting repulsion to a distance of $\\delta$ : Unlike most proxy-based methods, our design of potential field $\\Psi_{rep}$ limits the range of repulsion to within $\\delta$ distance from embeddings or proxies. We examine extending this range of repulsion beyond the $\\delta$ radius by having $\\delta_{rep} > \\delta_{att}$ where $\\delta_{att}$ and $\\delta_{rep}$ represent the values of $\\delta$ used for attraction (Eq.1) and repulsion (Eq.2) fields. Fixing $\\delta_{att} = 0.2$ , we increase $\\delta_{rep}$ to study the effect. As shown in Figure 3(d), his reduces performance, validating our design. This is due to individual potentials repelling dissimilar points to a distance of at least $\\delta_{rep}$ , greater than the distance $\\delta_{att}$ within which all points of the same class are attracted to. Using $\\delta_{rep} > \\delta_{att}$ enforces excessive separation between dissimilar examples, compressing intra-class features more than needed.", + "bbox": [ + 511, + 253, + 906, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 511, + 489, + 635, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PFML relies on proxies to augment potential fields in each batch, hence sharing the main limitation of proxy-based methods: with a large number of classes, the computational cost during training increases due to linear scaling of proxies with class count. Despite this, proxy-based methods have been widely used in applications with a large number of classes like face recognition (e.g. [10] on 500,000 classes), likely because (1) proxies only affect training, not inference complexity and (2) The parameters/compute complexity added by the proxies is relatively low for the performance gains they offer, especially when such compute complexity is compared to the forward and backward passes of the neural network that consume bulk of compute in DML.", + "bbox": [ + 511, + 508, + 906, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 715, + 633, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PFML introduces a novel paradigm to capture interactions between all samples through a continuous potential field, where strength of sample influence decreases with distance. This reversal of the common model of increasing influence with distance improves robustness to label noise and better aligns proxy distributions with sample embeddings. The overall design of our method is validated by its performance on standard image retrieval benchmarks, where it outperforms state-of-the-art methods both in the standard noise-free setting and in the presence of label noise unavoidable in real-world datasets.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25556", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Narendra Ahuja. A transform for multiscale image segmentation by integrated edge and region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 18 (12):1211-1235, 1996. 2", + "[2] N. Ahuja and Jen-Hui Chuang. Shape representation using a generalized potential field model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(2):169-176, 1997. 2", + "[3] Shubhang Bhatnagar and Narendra Ahuja. Piecewise-linear manifolds for deep metric learning. In Conference on Parsimony and Learning (Proceedings Track), 2023. 1", + "[4] Biagio Brattoli, Joseph Tighe, Fedor Zhdanov, Pietro Perona, and Krzysztof Chalupka. Rethinking zero-shot video classification: End-to-end training for realistic applications. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 4612-4622. Computer Vision Foundation / IEEE, 2020. 1", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6, 7", + "[6] Jun Cen, Peng Yun, Junhao Cai, Michael Yu Wang, and Ming Liu. Deep metric learning for open world semantic segmentation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15313-15322, 2021. 1", + "[7] Weihua Chen, Xiaotang Chen, Jianguo Zhang, and Kaiqi Huang. Beyond triplet loss: A deep quadruplet network for person re-identification. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1320-1329. IEEE Computer Society, 2017. 1", + "[8] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 539-546. IEEE, 2005. 1, 2, 5", + "[9] Jen-Hui Chuang and N. Ahuja. An analytically tractable potential field model of free space and its application in obstacle avoidance. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 28(5):729-736, 1998. 2", + "[10] J. Deng, J. Guo, N. Xue, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4685-4694, 2019. 1, 2, 8", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6, 7", + "[12] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7409-7419, 2022. 7", + "[13] Weifeng Ge, Weilin Huang, Dengke Dong, and Matthew R. Scott. Deep metric learning with hierarchical triplet loss. In Computer Vision - ECCV 2018, pages 272-288, Cham, 2018. Springer International Publishing. 7", + "[14] Matheus Gutoski, André Eugénio Lazzaretti, and Heitor Silvério Lopes. Deep metric learning for open-set human action recognition in videos. Neural Comput. Appl., 33(4): 1207-1220, 2021. 1", + "[15] Ben Harwood, Vijay Kumar BG, Gustavo Carneiro, Ian Reid, and Tom Drummond. Smart mining for deep metric learning. In Proceedings of the IEEE international conference on computer vision, pages 2821-2829, 2017. 1, 2, 6", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[17] J. Hu, J. Lu, and Y. Tan. Discriminative deep metric learning for face verification in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014. 1", + "[18] Y.K. Hwang and N. Ahuja. Path planning using a potential field representation. In Proceedings. 1988 IEEE International Conference on Robotics and Automation, pages 648-649 vol.1, 1988. 2", + "[19] Y.K. Hwang and N. Ahuja. A potential field approach to path planning. IEEE Transactions on Robotics and Automation, 8 (1):23-32, 1992. 1, 2", + "[20] O. Khatib. Real-time obstacle avoidance for manipulators and mobile robots. In Proceedings. 1985 IEEE International Conference on Robotics and Automation, pages 500-505, 1985. 1, 2", + "[21] Sungyeon Kim, Dongwon Kim, Minsu Cho, and Suha Kwak. Proxy anchor loss for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 5, 6, 7", + "[22] Sungyeon Kim, Boseung Jeong, and Suha Kwak. Hier: Metric learning beyond class labels via hierarchical regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19903-19912, 2023. 6, 7", + "[23] M. Kirchhof, K. Roth, Z. Akata, and E. Kasneci. A non-isotropic probabilistic take on proxy-based deep metric learning. In Computer Vision - ECCV 2022 - 17th European Conference, Proceedings, Part XXVI, pages 435-454. Springer, 2022. 2, 7", + "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 6", + "[25] Elad Levi, Tete Xiao, Xiaolong Wang, and Trevor Darrell. Rethinking preventing class-collapsing in metric learning with margin-based losses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10316-10325, 2021. 1, 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25557", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Christopher Liao, Theodoros Tsiligkaridis, and Brian Kulis. Supervised metric learning to rank for retrieval via contextual similarity optimization. In International Conference on Machine Learning, pages 20906-20938. PMLR, 2023. 1, 5", + "[27] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning code. https://github.com/ljin0429/HIST, 2022. 7", + "[28] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 212-222, 2022. 6, 7", + "[29] Lizhao Liu, Shangxin Huang, Zhuangwei Zhuang, Ran Yang, Mingkui Tan, and Yaowei Wang. Das: Densely-anchored sampling for deep metric learning. In European Conference on Computer Vision, pages 399–417. Springer, 2022. 6, 7", + "[30] Ruofan Liu, Yun Lin, XIANGLIN YANG, and Jin Song Dong. Debugging and explaining metric learning approaches: An influence function based perspective. In Advances in Neural Information Processing Systems, 2022. 5", + "[31] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1", + "[32] Benjamin J. Meyer and Tom Drummond. The importance of metric learning for robotic vision: Open set recognition and active learning. In 2019 International Conference on Robotics and Automation (ICRA), pages 2924-2931, 2019. 1", + "[33] Timo Milbich, Karsten Roth, Homanga Bharadhwaj, Samarth Sinha, Yoshua Bengio, Björn Ommer, and Joseph Paul Cohen. Diva: Diverse visual feature aggregation for deep metric learning. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VIII, page 590–607, Berlin, Heidelberg, 2020. Springer-Verlag. 7", + "[34] Yair Movshovitz-Attias, Alexander Toshev, Thomas K Leung, Sergey Ioffe, and Saurabh Singh. No fuss distance metric learning using proxies. In Proceedings of the IEEE international conference on computer vision, pages 360–368, 2017. 1, 2, 5, 6, 7", + "[35] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 6", + "[36] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 5, 6, 7, 8", + "[37] Samyak Rawlekar, Shubhang Bhatnagar, Vishnuvardhan Pogunulu Srinivasulu, and Narendra Ahuja. Improving multi-label recognition using class co-occurrence probabilities. In International Conference on Pattern Recognition, pages 424–439. Springer, 2025. 1" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Karsten Roth, Timo Milbich, Samarth Sinha, Prateek Gupta, Björn Ommer, and Joseph Paul Cohen. Revisiting training strategies and generalization performance in deep metric learning. In Proceedings of the 37th International Conference on Machine Learning. JMLR.org, 2020. 1", + "[39] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7420-7430, 2022. 1, 6", + "[40] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1, 2", + "[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1", + "[42] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1, 2", + "[43] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6", + "[44] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6397-6406, 2020. 7", + "[45] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H. S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1199-1208, 2017. 1", + "[46] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6", + "[47] Mark Tabb and Narendra Ahuja. Multiscale image segmentation by integrated edge and region detection. IEEE transactions on image processing, 6(5):642-655, 1997. 2", + "[48] Eu Wern Teh, Terrance DeVries, and Graham W. Taylor. Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV, pages 448-464. Springer, 2020. 2, 6, 7", + "[49] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6", + "[50] Chengkun Wang, Wenzhao Zheng, Junlong Li, Jie Zhou, and Jiwen Lu. Deep factorized metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2023. 7" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25558", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6, 7", + "[52] Kilian Q Weinberger, John Blitzer, and Lawrence Saul. Distance metric learning for large margin nearest neighbor classification. In Advances in Neural Information Processing Systems. MIT Press, 2005. 1, 2, 7", + "[53] Chao-Yuan Wu, R Manmatha, Alexander J Smola, and Philipp Krahenbuhl. Sampling matters in deep embedding learning. In Proceedings of the IEEE international conference on computer vision, pages 2840-2848, 2017. 1, 2, 6", + "[54] Tong Xiao, Shuang Li, Bochao Wang, Liang Lin, and Xiaogang Wang. Joint detection and identification feature learning for person search. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3415-3424, 2017. 1", + "[55] Hong Xuan, Abby Stylianou, and Robert Pless. Improved embeddings with easy positive triplet mining. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2463-2471, 2019. 2, 7", + "[56] Bailin Yang, Haoqiang Sun, Frederick WB Li, Zheng Chen, Jianlu Cai, and Chao Song. Hse: Hybrid species embedding for deep metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11047-11057, 2023. 7", + "[57] Yuhui Yuan, Kuiyuan Yang, and Chao Zhang. Hard-aware deeply cascaded embedding. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2", + "[58] Andrew Zhai and Hao-Yu Wu. Classification is a strong baseline for deep metric learning. In *British Machine Vision Conference*, 2018. 7", + "[59] Wenzhao Zheng, Chengkun Wang, Jiwen Lu, and Jie Zhou. Deep compositional metric learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9316-9325, 2021. 7", + "[60] Wenzhao Zheng, Borui Zhang, Jiwen Lu, and Jie Zhou. Deep relational metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12065-12074, 2021. 7", + "[61] Yuehua Zhu, Muli Yang, Cheng Deng, and Wei Liu. Fewer is more: A deep graph metric learning perspective using fewer proxies. In Advances in Neural Information Processing Systems, pages 17792-17803. Curran Associates, Inc., 2020. 6, 7" + ], + "bbox": [ + 91, + 92, + 485, + 752 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25559", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Potential Field Based Deep Metric Learning/c5f58da8-3e8b-4fa3-accf-6f5c33627040_model.json b/2025/Potential Field Based Deep Metric Learning/c5f58da8-3e8b-4fa3-accf-6f5c33627040_model.json new file mode 100644 index 0000000000000000000000000000000000000000..58e7c0cdd36ac57e1aa135e45a2fa06301432f33 --- /dev/null +++ b/2025/Potential Field Based Deep Metric Learning/c5f58da8-3e8b-4fa3-accf-6f5c33627040_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.277, + 0.131, + 0.723, + 0.154 + ], + "angle": 0, + "content": "Potential Field Based Deep Metric Learning" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.182, + 0.668, + 0.218 + ], + "angle": 0, + "content": "Shubhang Bhatnagar Narendra Ahuja University of Illinois Urbana-Champaign" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.22, + 0.621, + 0.233 + ], + "angle": 0, + "content": "{sb56，n-ahuja}@illinois.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.604 + ], + "angle": 0, + "content": "Deep metric learning (DML) involves training a network to learn a semantically meaningful representation space. Many current approaches mine \\( n \\)-tuples of examples and model interactions within each tuplets. We present a novel, compositional DML model that instead of in tuples, represents the influence of each example (embedding) by a continuous potential field, and superposes the fields to obtain their combined global potential field. We use attractive/repulsive potential fields to represent interactions among embeddings from images of the same/different classes. Contrary to typical learning methods, where mutual influence of samples is proportional to their distance, we enforce reduction in such influence with distance, leading to a decaying field. We show that such decay helps improve performance on real world datasets with large intra-class variations and label noise. Like other proxy-based methods, we also use proxies to succinctly represent sub-populations of examples. We evaluate our method on three standard DML benchmarks-Cars-196,CUB-200-2011,and SOP datasets where it outperforms state-of-the-art baselines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.222, + 0.649 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.81 + ], + "angle": 0, + "content": "The goal of visual metric learning is to learn a representation function where common distance metrics like the \\(\\ell_2\\) distance capture similarity between images. Semantic similarity is useful for applications such as image/video retrieval and search [3, 4, 21, 34], open set classification and segmentation [6, 14, 32, 37], few-shot learning [41, 45], person re-identification[7, 54] and face verification [10, 17, 31]. Deep neural networks have helped make significant advances in visual metric learning [40], as they are trained to learn a non-linear mapping of images onto a semantic space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.886 + ], + "angle": 0, + "content": "Existing Methods: Popular loss functions to train such networks help force tuples of examples/samples/embeddings of the same class to be closer than those in different classes. Examples include the triplet loss [52], the contrastive loss [8] and their extensions [42, 51]." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The utility of supervision provided by samples in a tu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.908, + 0.468 + ], + "angle": 0, + "content": "plet depends on samples' relative locations and labels. Most tuplet-based methods use sample mining strategies [15, 40] to choose a subset of informative tuplets to train the network, without which they suffer from lower performance and slower convergence [53]. Such strategies not only require the computationally demanding calculation of the relative hardness of all tuples formed \\((\\mathrm{O}(n^{3}))\\) for triplets among \\(n\\) samples), but also incur a loss of valuable supervision by discarding interactions. Specifically, limiting interactions to only a subset of all possible tuplets leads to greater susceptibility to noise [25], overfitting on the training set [26], and reduced learning of intra-class variations and generalizable features [38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.909, + 0.653 + ], + "angle": 0, + "content": "Proxy-based methods, first proposed in [34] attempt to mitigate some of these shortcomings by instead using class representative 'proxies' as one or more hypothesized centers for each class. Computation of the loss is then in terms of sample-proxy distances instead of sample-sample distances, helping circumvent the problem of tuplet sampling. However, these methods have been found to learn less generalizable intra-class features due to the exclusive use of proxy-sample distances while ignoring sample-sample distances[39]. This effect is compounded when proxies, whose locations are themselves learned, end up far from sample embedding locations they are intended to represent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Proposed Representation: To address these challenges, we propose Potential Field based Deep Metric learning (PFML), a metric learning framework that uses a continuous, potential field representation of all data samples motivated by potential fields used to model influence of charges in physics. Besides its analogical relevance, our motivation for using this representation comes from its previous, very successful uses, where it also helps overcome complexity in modeling interactions. These applications include planning paths of a robot avoiding obstacles [19] and planning an efficient strategy for a robot manipulator to grasp an object [20]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In PFML, we view each data sample as a charge, creating a field. The fields due to the individual samples are added (superposed) to obtain the global potential field. This global field thus constitutes a compositional representation, cap-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25549" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "turing the combined influence at a given location (or sample) as a weighted addition of the influences of the individual samples. Each sample generates two types of fields: (1) an attraction field to drive similar examples closer to it, and (2) a repulsion field that drives dissimilar examples away from it, each weakening with distance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.482, + 0.273 + ], + "angle": 0, + "content": "We also make use of proxies like previous work, but only to augment the potential field generated by the current batch of samples. This is in contrast to past work that completely replaces sample-sample interactions with sample-proxy interactions. Preserving sample-sample relations through the potential field helps learn better intra-class features." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.275, + 0.485, + 0.742 + ], + "angle": 0, + "content": "Advantages of Potential Field: (1) Using the potential field representation enables us to model interactions among all sample embeddings, as opposed to modeling those among smaller number of samples (or proxies) as done in methods using point-tuplet based (e.g., triplet) loss. (2) Addition of potentials due to all points increases robustness to noise since the effect of noise on interactions among a smaller number of samples will have a larger variance while also (3) Considering all interactions improves the quality of features learned, unlike tuplet based methods which lose performance when all interactions are considered[53]. (4) This is because of a major feature that differentiates our potential field from past methods: The variation of strength of interaction between two points as the distance between them increases: instead of remaining constant or even becoming stronger, as is the case with most existing methods, e.g. [8, 21], in our model it becomes weaker with distance. The decay property is helpful in several ways: (4a) It ensures the intuitive expectation that two distant positive samples are too different to be considered as variants of each other, helping treat them as different varieties (e.g., associated with different proxies). (4b) The decay property also significantly improves PFML performance for the specific type of noise affecting labels, e.g., due to annotation errors common in real-world datasets (Sec. 4.3). (4c) As a result of the decay, the learned proxies remain closer to (at smaller Wasserstein distance \\( W_{2} \\) from) the sample embeddings they represent than for (e.g., current proxy-based) methods where interactions strengthen with distance [21, 34, 36]; Sec 3.4), thereby enhancing their desired role." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Contributions: The main contributions of the proposed approach are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.775, + 0.483, + 0.851 + ], + "angle": 0, + "content": "- We present PFML, a novel framework for metric learning with a compositional representation using a continuous, potential field to model all sample interactions instead of only interactions between a subset of all possible n-tuplets. This enables better feature learning, especially" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.764, + 0.106 + ], + "angle": 0, + "content": "in the presence of annotation noise." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.907, + 0.197 + ], + "angle": 0, + "content": "- PFML reverses the almost universally used model of inter-sample interactions strengthening with distance; interaction strength instead weakens with distance. This (1) helps improve robustness in the presence of label noise, and (2) we prove that it helps better align the distribution of proxies with the sample embeddings they represent." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.198, + 0.907, + 0.288 + ], + "angle": 0, + "content": "- We evaluate PFML on three zero-shot image retrieval benchmarks: Cars-196, CUB-200-2011, and the SOP dataset where it outperforms current, state-of-the-art techniques in both the standard no-label-noise scenario, and with even greater margins (\\(>7\\%\\) R@1) in the more realistic cases of label noise." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.107, + 0.907, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.308, + 0.655, + 0.323 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.336, + 0.905, + 0.366 + ], + "angle": 0, + "content": "Broadly, DML may be divided into two types, ranking/pair-based and proxy-based [21, 51]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.368, + 0.907, + 0.64 + ], + "angle": 0, + "content": "Ranking/Pair based DML: These methods utilize losses that derive information from relations between intersample distances in pairs/ tuples of sample embeddings. Examples include the classical contrastive loss [8] or the triplet loss[40, 52, 53], which extends the concept to take into account relative distances in a triplet. [42] and [43] generalize these constraints to include multiple dissimilar samples. Mining informative pairs/tuples from a mini-batch plays an important role in the effectiveness of pair-based methods. Several strategies [15, 40, 57] have been proposed to pick samples based on their relative difficulty, with hard and semi-hard negative (picking pair of inter-class examples close together) and easy positive (same class examples close together) [25, 55] being popular. PFML naturally emphasizes such local (dis) similarities by virtue of its decaying formulation without the complexity of mining tuples. Also note that unlike PFML, none of these strategies optimize all interactions among a batch of samples." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.642, + 0.906, + 0.747 + ], + "angle": 0, + "content": "Proxy-based DML: ProxyNCA [34] and its variants [23, 48] attempt to minimize the distance of a sample from the proxy of its class while maximizing the distance from proxies of other classes. Proxy Anchor [21] builds on these by proposing a modified loss that takes into account sample hardness. [10, 36] introduced the use of multiple proxies per class to better model sub-clusters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Potential Field: Potential fields have been used in a variety of applications to model interactions of objects. Of interest here is their use in robot path planning [9, 19], shape representation [2, 18], image segmentation [1, 47] and object manipulation [20]. Here, they have been used to direct a robot towards its goal using an attraction field while avoiding obstacles using repulsion fields. Training supervision for DML shares a similar goal of moving sample embeddings towards other nearby samples of the same class, while avoiding dissimilar examples belonging to other classes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1It remains constant over a small radius and decays outside; however, for brevity, we will disregard this detail in the rest of this paper, and refer to the influence as decaying with distance, or as the decay property." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25550" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.482, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.414 + ], + "angle": 0, + "content": "Figure 1. Overview of our Potential-field based DML pipeline. The process includes (1) Computing attraction and repulsion fields generated by each embedding and proxy, (2) Computing the class potential fields by superposition of individual fields (3) Evaluating total potential energy by summing up the potentials of embeddings and proxies under the class potential field and (4) Updating locations of sample embeddings (through network parameters) and proxies to minimize total potential energy through backprop." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.422, + 0.182, + 0.437 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.448, + 0.242, + 0.464 + ], + "angle": 0, + "content": "3.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.653 + ], + "angle": 0, + "content": "The goal of deep metric learning is to define a semantic distance metric \\( d(\\mathbf{x}_1, \\mathbf{x}_2) \\) for any pair of samples \\( \\mathbf{x}_1, \\mathbf{x}_2 \\in \\mathcal{D} \\), where \\( \\mathcal{D} = \\{\\mathbf{x}_i, y_i\\} \\), \\( i \\in \\{1, \\dots, |\\mathcal{D}|\\} \\) and labels \\( y_i \\in \\{1, \\dots, N\\} \\) represent \\( N \\) classes. The learned distance metric for any two points \\( d(\\mathbf{x}_1, \\mathbf{x}_2) \\) should represent their semantic dissimilarity. Instead of directly learning the distance metric, a projector function \\( f_\\theta \\) is often represented by a neural network, parameterized by \\( \\theta \\) learned from data and distance \\( d(\\mathbf{x}_1, \\mathbf{x}_2) = \\|f_\\theta(\\mathbf{x}_1), f_\\theta(\\mathbf{x}_2)\\|_2 \\) is given by the Euclidean distance between the projections/embeddings. For simplicity, we will refer to the embeddings \\( f_\\theta(\\mathbf{x}_i) \\) belonging to a batch \\( \\mathcal{B} \\) sampled from \\( \\mathcal{D} \\) as \\( \\mathbf{z}_i \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.274, + 0.68 + ], + "angle": 0, + "content": "3.2. The Potential Field" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.869 + ], + "angle": 0, + "content": "With our potential field representation, embeddings need to be driven towards other nearby embeddings belonging to the same class, while also being driven away from embeddings of other classes. This is reminiscent of the behavior of an isolated system of electric charges, where dissimilar charges are drawn together while similar ones are repelled. Analogously, we define attractive and repulsive potentials created by a single embedding, with the net potential generated by a set of samples given by superposing individual potentials exerted by each embedding. The gradient of the net potential at a point yields the net force (on network update) on a sample embedding located at that point." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As samples of each class have a different set of similar and dissimilar points, we define a separate potential field" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.333 + ], + "angle": 0, + "content": "\\(\\Psi_{j}\\), for each class \\(j \\in \\{1, \\dots, N\\}\\). The class potential field \\(\\Psi_{j}\\) models the force applied on embeddings belonging to class \\(j\\) due to all embeddings. It is formed from a superposition of potentials belonging to sample embeddings from all classes, with embeddings from the same class exerting an attractive potential field (to pull similar points closer) while those from other classes exert a repulsive potential field (to push apart dissimilar points). We assume that each sample embedding \\(z_{i}\\) has a unit charge, so the net force on it is the gradient of the potential field \\(\\Psi_{y_i}\\) of its class. We design the attractive and repulsive potential functions by taking motivation from both, electrostatic fields and metric learning literature, which we discuss in detail in the following subsections. Figure 1 provides an overview of our complete pipeline of generation and use of the potential field for metric learning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.341, + 0.83, + 0.357 + ], + "angle": 0, + "content": "3.2.1. Potential field exerted by a single point" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.906, + 0.456 + ], + "angle": 0, + "content": "By using forces on the embeddings, we wish the training process to move the embeddings to a state of equilibrium at the end of training in which all embeddings from a class are nearby. To have a potential function exhibiting this behavior, we ensure that individual potential field has the following two properties:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.906, + 0.544 + ], + "angle": 0, + "content": "Property 1 - Extremum Property: Potential generated by an individual embedding achieves an extremum at the embedding (minimum for attraction field, maximum for repulsion). This ensures that similar points are pulled together and dissimilar ones are pushed apart on minimizing potential." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Property 2 - Decay Property: The magnitude of the gradient of the potential (force) due to an embedding decays with distance. Such a decay ensures two desired characteristics. (1) An embedding is not pulled towards a distant sample embedding belonging to the same class, which is likely to be too different to be considered as a variant of it; instead, the decayed influence helps treat the distant sample as significantly different variety of the class, associated with other nearby embeddings of the class. (2) Like for a single charge in the Extremum property, a superposition of potentials of different embeddings still yields local minima at or near sample embeddings. As shown in Sec 3.4, this helps proxies come close to such minima and thus help better model the data distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.757, + 0.906, + 0.817 + ], + "angle": 0, + "content": "Given these properties motivated by electrostatic fields, we select the attraction and repulsion potentials \\(\\psi_{att},\\psi_{rep}\\) generated by an embedding \\(\\mathbf{z}_i\\) , with their strength at location \\(\\mathbf{r}\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.835, + 0.907, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\psi_ {a t t} (\\mathbf {r}, \\mathbf {z} _ {i}): = \\left\\{ \\begin{array}{l l} - \\frac {1}{\\delta^ {\\alpha}} & \\text {i f} \\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} < \\delta \\\\ - \\frac {1}{\\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} ^ {\\alpha}} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25551" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.098, + 0.387, + 0.113 + ], + "angle": 0, + "content": "(a) Class Potential field \\(\\Psi_{1}\\)" + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.115, + 0.443, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.098, + 0.744, + 0.113 + ], + "angle": 0, + "content": "(b) Class Potential field \\(\\Psi_{2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.114, + 0.857, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.359, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Figure 2. An example of the class potential fields \\(\\Psi_{1}\\) and \\(\\Psi_{2}\\) (Sec. 3.2.3), created by superposing the fields of individual embeddings (Sec 3.2.1) belonging to Classes 1 and 2. Arrows denote the gradient of the respective potentials representing the net force on them. (a) \\(\\Psi_{1}\\) draws samples/proxies of class 1 towards nearby samples/ proxies of class 1 while keeping them at least \\(\\delta\\) distance away from embeddings of class 2. Proxies at starred locations are drawn towards the nearest Class 1 embeddings which are potential minima, helping them better model the data distribution (Sec 3.4). (b) \\(\\Psi_{2}\\) draws embeddings of class 2 towards other nearby embeddings of class 2 instead of distant class 2 embeddings, which might be a significantly different variant of the class (or potentially a mislabeled data point), helping better feature learning (or improving noise robustness if mislabeled). \\(\\Psi_{2}\\) also keeps class 2 embeddings \\(\\delta\\) distance them away from embeddings of class 1." + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.475, + 0.483, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\psi_ {r e p} (\\mathbf {r}, \\mathbf {z} _ {i}): = \\left\\{ \\begin{array}{l l} \\frac {1}{\\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} ^ {\\alpha}} & \\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} < \\delta \\\\ \\frac {1}{\\delta^ {\\alpha}} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.552, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Here, \\(\\alpha\\) is a hyperparameter that determines the rate of decay of the field (we evaluate it's effect in Sec. 4.5 instead of fixing \\(\\alpha = 1\\) as done in electrostatic fields). Note that the attraction potential is negative and increases with distance, implying that the force between the embedding and the point is attractive. The repulsion potential is positive and decreasing with distance, indicating repulsion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.764 + ], + "angle": 0, + "content": "A constant attraction potential inside the \\(\\delta\\) sphere (i.e., force \\(= 0\\)) ensures that sample embeddings within \\(\\delta\\) distance of one another do not continue to attract each other, thus preventing the network from embeddings from collapsing to a single point which harms generalizability of the features learned (analogous to the effect margins used in tuplet-based losses have)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Limiting the repulsion potential to a sphere of radius \\(\\delta\\) ensures that embeddings not belonging to the same class are driven to a distance \\(> \\delta\\). Since the aforementioned constant attraction potential ensures that \\(\\delta\\) is the maximum distance between samples \\(z_{i}\\) from the same class, this leads to connected components of \\(\\delta\\) spheres formed by nearby embeddings of the same class. Extending repulsion to distances \\(> \\delta\\) does not provide any useful additional supervisory information while forcing further unmotivated inter-class sep" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.479, + 0.907, + 0.51 + ], + "angle": 0, + "content": "aration; as we empirically show in Sec. 4.5, such excessive separation does indeed harm performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.845, + 0.532 + ], + "angle": 0, + "content": "3.2.2. Proxies to Represent Sample Populations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.732 + ], + "angle": 0, + "content": "To represent the image embeddings not sampled in the current batch \\(\\mathcal{B}\\), we propose to use a set of \\(M\\) proxies per class for each of the N classes, \\(\\mathbf{p}_{j,k}, k \\in \\{1, \\dots, M\\}, j \\in \\{1, \\dots, N\\}\\), and augment the potential field generated by the embeddings in the batch by those generated by the proxies. The proxies serve as stand-ins for the larger numbers of out-of-the-batch embeddings for calculating the attraction and repulsion fields, while also being affected by the field themselves like all embeddings. The proxies, being learnable parameters with high learning rates, tend to gravitate toward the nearest class modes of the potential field (and hence the class data distribution). The number of proxies M chosen should be large enough to capture the class-modes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.738, + 0.704, + 0.753 + ], + "angle": 0, + "content": "3.2.3. Class Potential Field" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.906, + 0.804 + ], + "angle": 0, + "content": "Given a batch \\(\\mathcal{B}\\) of samples and all proxies, we calculate the class potential field \\(\\Psi_{j}(\\mathbf{r})\\) at a location \\(\\mathbf{r}\\) due to a class \\(j\\) as a superposition of their attraction and repulsion potentials:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.815, + 0.907, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\Psi_ {j} (\\mathbf {r}) = \\Psi_ {j, a t t} (\\mathbf {r}) + \\Psi_ {j, r e p} (\\mathbf {r}), \\text {w h e r e} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.855, + 0.907, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\Psi_ {j, a t t} (\\mathbf {r}) = \\sum_ {i = 1} ^ {\\| \\mathcal {B} \\|, y _ {i} = j} \\psi_ {a t t} (\\mathbf {r}, \\mathbf {z} _ {i}) + \\sum_ {k = 1} ^ {M} \\psi_ {a t t} (\\mathbf {r}, \\mathbf {p} _ {j, k}) \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25552" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.092, + 0.089, + 0.495, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\Psi_ {j, r e p} (\\mathbf {r}) = \\sum_ {i = 1} ^ {\\| \\mathcal {B} \\|, y _ {i} \\neq j} \\psi_ {r e p} (\\mathbf {r}, \\mathbf {z} _ {i}) + \\sum_ {\\gamma = 1, \\gamma \\neq j} ^ {N} \\sum_ {k = 1} ^ {M} \\psi_ {r e p} (\\mathbf {r}, \\mathbf {p} _ {\\gamma , k}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.159, + 0.486, + 0.19 + ], + "angle": 0, + "content": "Figure 2 visualizes the class potential fields \\(\\Psi_{1}\\) and \\(\\Psi_{2}\\) for a given set of embeddings belonging to two classes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.198, + 0.198, + 0.215 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.485, + 0.282 + ], + "angle": 0, + "content": "Using gradient descent, the network is trained to minimize the total potential energy \\(\\mathcal{U}\\) of all proxies and embeddings in the batch \\(\\mathbf{z}_i\\in \\mathcal{B}\\) by applying a force on embeddings and proxies in the direction of \\(-\\nabla \\Psi_{y_i}(\\mathbf{z}_i)\\). \\(\\mathcal{U}\\) given by:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.291, + 0.484, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathcal {U} = \\sum_ {i = 1} ^ {\\| \\mathcal {B} \\|} \\Psi_ {y _ {i}} (\\mathbf {z} _ {i}) + \\sum_ {j = 1} ^ {N} \\sum_ {k = 1} ^ {M} \\Psi_ {j} (\\mathbf {p} _ {j, k}) \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.344, + 0.483, + 0.361 + ], + "angle": 0, + "content": "3.4. Why decay interaction strength with distance?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.427 + ], + "angle": 0, + "content": "PFML uses a potential field where the gradient (interaction strength) decays with distance, which is unlike previous approaches. This leads to the two benefits described in Sec 3.2.1, which we discuss in more detail here." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.464 + ], + "angle": 0, + "content": "3.4.1. Robustness when learning with datasets having large intra-class variations and noisy labels" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.468, + 0.485, + 0.724 + ], + "angle": 0, + "content": "The decay property helps PFML model all interactions even when using datasets with large intra-class variations. As seen in Figure 2, PFML draws embeddings towards other nearby embeddings from the same class rather than distant class embeddings that might be too different to be considered a variation of it. It also helps PFML become significantly more robust to the effect of incorrectly labeled examples as compared to existing methods. Recent work [26, 30] shows that such examples are common in metric learning benchmarks. The decay de-emphasizes the contribution of such unlabeled examples, that are typically located away from correctly labeled samples. We empirically confirm this in Sec. 4.3, demonstrating that methods not adhering to the decay property (i.e., all previous tuplet-based and proxy-based methods) experience significantly greater performance degradation due to label noise compared to PFML." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.732, + 0.483, + 0.746 + ], + "angle": 0, + "content": "3.4.2. Better alignment of proxies with data distribution" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The decaying interaction strength helps PFML better align the distribution of proxies with the underlying sample data distribution (of the complete dataset) that they represent. To analyze this, we compare PFML with the closely related contrastive loss [8] that also models pairwise interactions like PFML (defined in Eq. 1 & 2), and hence can be analyzed similarly by using a potential field. Specifically, we introduce Contrastive Potential based Metric Learning (CPML) by simply extending the contrastive loss by using multiple proxies per class, and subsequently analyzing the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "resulting loss function through a potential field framework. In CPML, strength of interaction is directly proportional to the square of the distance from it \\((\\|\\mathbf{r} - \\mathbf{r}_0\\|_2^2)\\) like in the original contrastive loss. For fair comparison, we keep all other details (use of \\(\\delta\\), proxies) as the same, with its associated potential fields denoted by \\(\\Psi_{att}^*\\), \\(\\Psi_{rep}^*\\) (formally defined in Supplement Sec 1). We analyze this theoretically using the following proposition and its corollary:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.217, + 0.909, + 0.355 + ], + "angle": 0, + "content": "Proposition 1: Let \\( Z = \\{z_1 \\ldots z_n\\} \\) be a set of sample embeddings belonging to a class, then there exists a \\( 0 < \\delta < \\frac{\\min_{i,j} \\|z_i - z_j\\|_2}{2\\left(1 + \\frac{1}{n}\\right)} \\) and points \\( z_{min,i} \\) within \\( \\delta \\) distance from each embedding \\( z_i \\), \\( i \\in \\{1 \\ldots n\\} \\) such that the attractive potential field \\( \\Psi_{att} \\) defined using \\( Z \\), \\( \\delta \\) has a minimum at each \\( z_{min,i} \\). The field \\( \\Psi_{att}^* \\) defined by them does not achieve a minimum at points within \\( \\delta \\) distance from all \\( z_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Proposition 1 is a consequence of the potential field of an embedding dominating in its vicinity over potential fields of other distant embeddings. Its proof relies on the Decay property, which is not satisfied by \\(\\Psi^{*}\\) and can be found in Supp. Sec 1. A consequence of this is:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.909, + 0.547 + ], + "angle": 0, + "content": "Corollary 1: Let \\( Z = \\{z_{1}\\dots z_{n}\\} \\) be a set of sample embeddings from a class exerting an attraction field on a set of proxies \\( P = \\{\\pmb {p}_1\\dots \\pmb {p}_m\\} \\). Consider the equilibrium distribution \\( P_{eq} \\) of proxies minimizing the potential energy. If the potential field is defined by \\( \\Psi_{att} \\), then the Wasserstein distance \\( W_{2} \\) between \\( P_{eq} \\) and the subset of data they represent is lower than when the potential field is defined by \\( \\Psi_{att}^* \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.908, + 0.58 + ], + "angle": 0, + "content": "The proof for Corollary 1 relies on Proposition 1 and can be found in the Supplement (Sec. 2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Discussion: As proved in Corollary 1, CPML leads to proxies having a larger distance from the data distribution they represent (measured by \\( W_{2} \\)) than our proposed PFML, making the use of proxies in such a framework less effective. This leads to the proxies forming poorer stand-ins for the sample embeddings they represent, reducing the quality of features learned (verified in Sec. 4). This is caused by CPML not following the decay property (which is its only difference in design wrt PFML). Other proxy-based losses like the Proxy NCA [34], Soft Triplet [36], and Proxy Anchor [21] which perform similarly as CPML (see Tab. 1), also share with it the property of force increasing with distance (violating the decay property). Due to this, we expect them to similarly have a poorer alignment of proxies with underlying data distribution as compared to PFML (Applying an analysis similar to the one in Proposition 1 to them is considerably more challenging due to the absence of a potential field structure.). We verify this empirically in Sec. 4.4 by measuring the average distance between proxies and the closest data points (\\( W_{2} \\) distance) for these methods." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "25553" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.482, + 0.108 + ], + "angle": 0, + "content": "3.5. Advantages in Learning Fine-grained Features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.118, + 0.483, + 0.269 + ], + "angle": 0, + "content": "In contrast with all previous proxy-based methods, PFML (1) Models sample-sample interactions directly using the potential field, gaining fine-grained supervision, and (2) Makes use of an explicit margin-like \\(\\delta\\) for proxies, which has been shown to help zero-shot generalization and fine-grained feature learning in tuplet-based methods[53]. Turning off the interaction within the \\(\\delta\\) radius also prevents a bias towards learning isotropic distributions as is present in other proxy-based methods [39], helping us obtain better intra-class resolution." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.299, + 0.325, + 0.316 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.33, + 0.173, + 0.345 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.483, + 0.522 + ], + "angle": 0, + "content": "Datasets: We empirically compare PFML with current state-of-the-art baselines on three public benchmark datasets (1) The Cars-196 dataset [24] which includes 16,185 images from 196 car model categories, (2) CUB200-2011 dataset [49] consisting of 11,788 images from 200 bird species, and (3) Stanford Online Products (SOP) dataset [43] containing 120,053 images of 22,634 different products sold online. The training and testing splits for the three datasets follow standard zero-shot retrieval-based evaluation protocol [21, 51, 61] of using the first half of the classes for training and the second half for testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.483, + 0.664 + ], + "angle": 0, + "content": "**Backbone:** To enable standardized comparison with a wide variety of methods, we evaluate our method using the Inception with batch normalization (IBN) [46], ResNet-50 [16], ViT [11] and DINO [5] backbones commonly used by DML methods. ImageNet pre-training is used for initialization, and only the last fully connected layer is changed to fit the dimension of the embedding needed. We follow past work [21, 36, 48] and \\(\\ell_2\\) normalize the final output. An input image size of \\(224 \\times 224\\) is used for all experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.669, + 0.483, + 0.805 + ], + "angle": 0, + "content": "Training parameters: We train the backbone for 200 epochs on each dataset. We use the Adam optimizer with the learning rate chosen as \\(5e^{-4}\\). We scale up the learning rate of our proxies by 100 times following [21, 34, 36] for better convergence. We choose values of \\(\\delta\\) between \\([0.1, 0.3]\\), \\(\\alpha\\) between \\(\\{0, 6\\}\\) via cross-validation. We use \\(M = 15\\) for the CUB-200 and Cars-196 datasets, while \\(M = 2\\) is used for the SOP dataset. All experiments are performed on a single NVIDIA A4000 GPU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation Settings: We follow [21, 22, 29, 51] in using \\(224 \\times 224\\) sized center crops from \\(256 \\times 256\\) images for evaluation. We measure image retrieval performance using Recall@K, which computes the percentage of samples which have a valid similar neighbor (belonging to same class) among its K nearest neighbors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.776, + 0.107 + ], + "angle": 0, + "content": "4.2. Image Retrieval Performance" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.476 + ], + "angle": 0, + "content": "As seen in Table 1, our method outperforms all other methods on Image Retrieval on the CUB-200-2011, Cars-196 and SOP datasets as measured by the Recall@K metric. It consistently outperforms other methods irrespective of the backbone used, significantly outperforming the best-performing proxy-based methods (which similar to us, only change the loss function) like ProxyAnchor[21], ProxyGML[61] and SoftTriplet [36] in terms of Recall@1 (R@1), by more than \\(5\\%\\) on Cars-196, \\(3.7\\%\\) on CUB-200-2011 and \\(1.5\\%\\) on the SOP dataset. It also outperforms the significantly more complex, current state-of-the-art HIST[28] by \\(3.1\\%\\), \\(2\\%\\) and \\(1.5\\%\\) in terms of R@1 when using the Resnet-50 backbone on the Cars-196, CUB-200-2011 and SOP datasets respectively. It also significantly outperforms the best tuplet-based loss: the Multi-Similarity loss by \\(6\\%\\), \\(5.8\\%\\) and \\(2.4\\%\\) in terms of R@1 when using the BN Inception backbone on the Cars-196, CUB-200 and SOP datasets respectively. Additionally, our method outperforms CPML (Sec. 3.4.2, a simple combination of the classical contrastive loss with proxies) by \\(7.5\\%\\), \\(5.1\\%\\), and \\(3.5\\%\\) on CUB200, Cars and SOP datasets. Note that our performance gains are larger than current SOTA compared to improvements reported by previous efforts on the same benchmarks, despite our intuitive and simpler method." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.476, + 0.905, + 0.581 + ], + "angle": 0, + "content": "We also evaluate our method 1) Using the protocol and metrics proposed in [35] and (2) When used to learn smaller 128-dim embeddings for more efficient retrieval. Our method still outperforms all SOTA methods in both cases. Details of these experiments are in Supp. Sec. 3. We also show qualitative retrieval results of our method in Supp. Sec 6." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.878, + 0.607 + ], + "angle": 0, + "content": "4.3. Image Retrieval in Presence of Label Noise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.905, + 0.823 + ], + "angle": 0, + "content": "We evaluate the robustness of our method vs. state-of-the-art baselines in the realistic scenario of label noise (e.g., due to mislabeling, which is very common, particularly in fine-grained datasets such as those used for DML). Specifically, we randomly corrupt labels of \\(20\\%\\) of samples in the training set. As seen in the results in Table 2, this leads to a significant decline in the performance of all methods. But our method is the least affected by the noise. Other methods like the triplet [15] or ProxyAnchor[21] suffer much greater declines because they are affected more by distant points represented by the mislabeled examples. Our method significantly outperforms Proxy Anchor [21] (the second-best method) by \\(6\\%\\) and \\(7.6\\%\\) in terms of R@1 on the CUB-200-2011 and Cars-196 datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.905, + 0.865 + ], + "angle": 0, + "content": "4.4. Distance Between Proxies and Sample Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "We measure the average (over a complete epoch) Wasserstein distance \\( W_{2} \\) between proxies and the subset of training" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25554" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.133, + 0.088, + 0.867, + 0.616 + ], + "angle": 0, + "content": "
Benchmarks →CUB-200-2011Cars-196SOP
Methods ↓ (Chronological)R@1R@2R@4R@1R@2R@4R@1R@10R@100
ResNet50 (512 dim)
ESPHN [55]64.975.383.582.789.393.078.390.796.3
N.Softmax [58]61.373.983.584.290.494.478.290.696.2
DiVA [33]69.279.3-87.692.9-79.691.2-
Proxy NCA++ [48]64.7---85.1--79.6--
Proxy Anchor [21]69.780.087.087.792.995.8---
DCML-MDW [59]68.477.986.185.291.896.079.890.895.8
MS+DAS [29]69.279.287.187.893.196.080.691.896.7
HIST [28]71.481.188.189.693.996.481.492.096.7
HIER[22]70.179.486.988.293.095.680.291.596.6
HSE-PA [56]70.680.187.189.693.896.080.091.496.3
CPML (Sec. 3.4)68.378.786.285.291.595.279.490.796.1
Potential Field (Ours)73.4 ± 0.382.4± 0.188.8± 0.192.7± 0.395.5 ± 0.197.6± 0.182.9 ± 0.292.5 ± 0.296.8 ± 0.1
BN Inception (512 dim)
HTL [13]57.168.878.781.488.092.774.888.394.8
MultiSimilarity [51]65.777.086.384.190.494.078.290.596.0
SoftTriple [36]65.476.484.584.590.794.578.390.395.9
CircleLoss [44]66.777.486.283.489.894.178.390.596.1
DiVA [33]66.877.7-84.190.7-78.190.6-
ProxyGML [61]66.677.686.485.591.895.378.090.696.2
Proxy Anchor [21]68.479.286.886.191.795.079.190.896.2
DRML-PA [60]68.778.686.386.992.195.271.585.293.0
MS+DAS [29]67.178.1186.485.791.695.378.290.396.0
HIST [28]69.780.087.387.492.595.479.691.096.2
DFML-PA [50]69.3--88.4-----
HSE-M [56]67.678.085.882.088.993.3---
PA+niV [23]69.580.0-86.492.0-79.290.4-
Potential Field (Ours)71.5± 0.381.2± 0.288.3± 0.290.1± 0.293.9± 0.196.3± 0.180.6 ± 0.391.8 ± 0.196.4 ± 0.1
DINO (384 dim)
DINO [5]70.881.188.842.953.964.263.478.188.3
Hyp [12]80.987.692.489.294.196.785.194.497.8
HIER [22]81.188.293.391.395.297.185.794.697.8
Potential Field (Ours)83.1 ± 0.389.3 ± 0.294.2 ± 0.194.7 ± 0.196.5 ± 0.197.8 ± 0.186.5 ± 0.395.1 ± 0.398.0 ± 0.2
ViT (384 dim)
ViT-S [11]83.190.494.447.860.272.262.177.789.0
Hyp [12]85.691.494.886.592.195.385.994.998.1
HIER [22]85.791.394.488.393.296.186.195.098.0
Potential Field (Ours)87.8 ± 0.292.6 ± 0.295.7 ± 0.191.5 ± 0.395.2 ± 0.297.4 ± 0.188.2 ± 0.195.7 ± 0.198.6 ± 0.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.617, + 0.908, + 0.647 + ], + "angle": 0, + "content": "Table 1. Comparison of the Recall@ \\( K \\) (%) achieved by our method on the CUB-200-2011, Cars-196 and SOP datasets with state-of-the-art baselines under standard settings. The table reports the average performance and standard deviations of our method over 5 runs." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.66, + 0.596, + 0.789 + ], + "angle": 0, + "content": "
CUB-200-2011Cars-196
MethodsR@1R@2R@1R@2
Triplet[52]55.168.767.577.9
MS [51]58.971.870.479.8
PNCA[34]60.174.774.382.4
PA [21]60.775.176.983.1
HIST[27]59.774.672.981.8
Ours66.7 ± 0.676.9± 0.384.5± 0.588.6± 0.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.799, + 0.595, + 0.842 + ], + "angle": 0, + "content": "Table 2. A comparison of Recall@ \\( K \\) (%) achieved by our method with SOTA baselines on CUB-200-2011 and Cars-196 datasets with label noise when using a Resnet-50 backbone (512-dim). Performance and std. deviations reported are over 5 runs." + }, + { + "type": "table", + "bbox": [ + 0.64, + 0.675, + 0.885, + 0.773 + ], + "angle": 0, + "content": "
MethodAverage W2
S.Triplet [36]0.24 ± 0.04
PNCA [21]0.36 ± 0.02
PA [21]0.38 ± 0.03
CPML (Sec. 3.4)0.26 ± 0.03
Ours0.16 ± 0.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.612, + 0.784, + 0.911, + 0.84 + ], + "angle": 0, + "content": "Table 3. Average Wasserstein distance between proxies and the subset of data closest to them for different methods on CUB-200-2011 dataset. Mean and std. deviations are over 5 runs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.873 + ], + "angle": 0, + "content": "data represented by/closest to them \\( {}^{2} \\) for different proxy-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.878, + 0.483, + 0.904 + ], + "angle": 0, + "content": "2For methods using a single proxy per class, this equals the distance between a proxy and the closest embedding belonging to its class" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.887 + ], + "angle": 0, + "content": "based methods. We train networks on the CUB-200-2011 using the settings in Section 4.1 and use the same number" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25555" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.298, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.089, + 0.498, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.692, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.089, + 0.892, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.208, + 0.905, + 0.236 + ], + "angle": 0, + "content": "Figure 3. Variation in Recall@1 with \\( M \\), \\( \\delta \\), \\( \\alpha \\) and \\( \\delta_{rep} - \\delta_{att} \\) on the Cars-196 and CUB-200-2011 datasets. Error bars represent std. deviations over 5 runs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.484, + 0.314 + ], + "angle": 0, + "content": "of proxies for all methods utilizing multiple proxies. As seen in Table 3, our method shows \\(33.3\\%\\) lower \\(W_{2}\\) distance between the proxies and the subset of data than the second best SoftTriplet [36] loss, empirically validating our discussion in Sec. 3.4.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.328, + 0.253, + 0.344 + ], + "angle": 0, + "content": "4.5. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.398 + ], + "angle": 0, + "content": "In all experiments, we use the parameters mentioned in Section 4.1 with a ResNet-50 backbone on both the Cars-196 and CUB-200-2011 datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Effect of proxies and their number: Proxies augment the potential field, and the number of proxies per class \\(M\\) decides the granularity at which the proxies model the distribution of sample embeddings. We study their effect on performance by varying \\(M\\) between [0, 30]. From the results plotted in Figure 3(a) we observe that (1) Using proxies to augment the field boosts performance (vs using no-proxies) and (2) Network performance remains stable for a wide range of M, decreasing slightly for small values of M (still outperforming most other proxy-based methods). This behavior is expected as using potential field created using a very small M would not be able to effectively model underlying sample distribution, leading to worse supervision. We also observe that using a larger number of proxies per class does not significantly affect training times, as bulk of computation is dominated by the network. For further details, please refer to Supp. Sec. 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Effect of varying \\(\\delta\\): We conduct an empirical study to determine the effect of varying the radius parameter \\(\\delta\\) on performance by varying it over [0.1, 0.35]. As seen in Figure 3(b), we observe that performance remains relatively stable for a wide range of \\(\\delta\\) values, demonstrating our method's robustness to it." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Effect of potential decay \\(\\alpha\\): The decay parameter \\(\\alpha\\) determines the rate at which the potential field of an embedding decays with distance. We study its effect by varying it over [0, 6]. As seen in Figure 3(c), performance is worse when the potential field does not decay (\\(\\alpha = 0\\)). A field that decays strongly \\(\\alpha \\in [3, 6]\\) seems to perform better than one with a much milder decay (\\(\\alpha \\in [1, 2]\\)). The likely reason for this trend is that when the potential field has no decay /decays mildly, the force on an embedding would still be dominated by a large number of distant sample embeddings" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.239, + 0.72, + 0.254 + ], + "angle": 0, + "content": "rather than those in its vicinity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.907, + 0.48 + ], + "angle": 0, + "content": "Restricting repulsion to a distance of \\(\\delta\\): Unlike most proxy-based methods, our design of potential field \\(\\Psi_{rep}\\) limits the range of repulsion to within \\(\\delta\\) distance from embeddings or proxies. We examine extending this range of repulsion beyond the \\(\\delta\\) radius by having \\(\\delta_{rep} > \\delta_{att}\\) where \\(\\delta_{att}\\) and \\(\\delta_{rep}\\) represent the values of \\(\\delta\\) used for attraction (Eq.1) and repulsion (Eq.2) fields. Fixing \\(\\delta_{att} = 0.2\\), we increase \\(\\delta_{rep}\\) to study the effect. As shown in Figure 3(d), his reduces performance, validating our design. This is due to individual potentials repelling dissimilar points to a distance of at least \\(\\delta_{rep}\\), greater than the distance \\(\\delta_{att}\\) within which all points of the same class are attracted to. Using \\(\\delta_{rep} > \\delta_{att}\\) enforces excessive separation between dissimilar examples, compressing intra-class features more than needed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.49, + 0.637, + 0.505 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.706 + ], + "angle": 0, + "content": "PFML relies on proxies to augment potential fields in each batch, hence sharing the main limitation of proxy-based methods: with a large number of classes, the computational cost during training increases due to linear scaling of proxies with class count. Despite this, proxy-based methods have been widely used in applications with a large number of classes like face recognition (e.g. [10] on 500,000 classes), likely because (1) proxies only affect training, not inference complexity and (2) The parameters/compute complexity added by the proxies is relatively low for the performance gains they offer, especially when such compute complexity is compared to the forward and backward passes of the neural network that consume bulk of compute in DML." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.716, + 0.634, + 0.731 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "PFML introduces a novel paradigm to capture interactions between all samples through a continuous potential field, where strength of sample influence decreases with distance. This reversal of the common model of increasing influence with distance improves robustness to label noise and better aligns proxy distributions with sample embeddings. The overall design of our method is validated by its performance on standard image retrieval benchmarks, where it outperforms state-of-the-art methods both in the standard noise-free setting and in the presence of label noise unavoidable in real-world datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25556" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Narendra Ahuja. A transform for multiscale image segmentation by integrated edge and region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 18 (12):1211-1235, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] N. Ahuja and Jen-Hui Chuang. Shape representation using a generalized potential field model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(2):169-176, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.269 + ], + "angle": 0, + "content": "[3] Shubhang Bhatnagar and Narendra Ahuja. Piecewise-linear manifolds for deep metric learning. In Conference on Parsimony and Learning (Proceedings Track), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.366 + ], + "angle": 0, + "content": "[4] Biagio Brattoli, Joseph Tighe, Fedor Zhdanov, Pietro Perona, and Krzysztof Chalupka. Rethinking zero-shot video classification: End-to-end training for realistic applications. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 4612-4622. Computer Vision Foundation / IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.494 + ], + "angle": 0, + "content": "[6] Jun Cen, Peng Yun, Junhao Cai, Michael Yu Wang, and Ming Liu. Deep metric learning for open world semantic segmentation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15313-15322, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.483, + 0.578 + ], + "angle": 0, + "content": "[7] Weihua Chen, Xiaotang Chen, Jianguo Zhang, and Kaiqi Huang. Beyond triplet loss: A deep quadruplet network for person re-identification. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1320-1329. IEEE Computer Society, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[8] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 539-546. IEEE, 2005. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[9] Jen-Hui Chuang and N. Ahuja. An analytically tractable potential field model of free space and its application in obstacle avoidance. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 28(5):729-736, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[10] J. Deng, J. Guo, N. Xue, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4685-4694, 2019. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7409-7419, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.193 + ], + "angle": 0, + "content": "[13] Weifeng Ge, Weilin Huang, Dengke Dong, and Matthew R. Scott. Deep metric learning with hierarchical triplet loss. In Computer Vision - ECCV 2018, pages 272-288, Cham, 2018. Springer International Publishing. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[14] Matheus Gutoski, André Eugénio Lazzaretti, and Heitor Silvério Lopes. Deep metric learning for open-set human action recognition in videos. Neural Comput. Appl., 33(4): 1207-1220, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.252, + 0.905, + 0.308 + ], + "angle": 0, + "content": "[15] Ben Harwood, Vijay Kumar BG, Gustavo Carneiro, Ian Reid, and Tom Drummond. Smart mining for deep metric learning. In Proceedings of the IEEE international conference on computer vision, pages 2821-2829, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.905, + 0.366 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.905, + 0.41 + ], + "angle": 0, + "content": "[17] J. Hu, J. Lu, and Y. Tan. Discriminative deep metric learning for face verification in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[18] Y.K. Hwang and N. Ahuja. Path planning using a potential field representation. In Proceedings. 1988 IEEE International Conference on Robotics and Automation, pages 648-649 vol.1, 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.512 + ], + "angle": 0, + "content": "[19] Y.K. Hwang and N. Ahuja. A potential field approach to path planning. IEEE Transactions on Robotics and Automation, 8 (1):23-32, 1992. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.514, + 0.905, + 0.569 + ], + "angle": 0, + "content": "[20] O. Khatib. Real-time obstacle avoidance for manipulators and mobile robots. In Proceedings. 1985 IEEE International Conference on Robotics and Automation, pages 500-505, 1985. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.907, + 0.627 + ], + "angle": 0, + "content": "[21] Sungyeon Kim, Dongwon Kim, Minsu Cho, and Suha Kwak. Proxy anchor loss for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[22] Sungyeon Kim, Boseung Jeong, and Suha Kwak. Hier: Metric learning beyond class labels via hierarchical regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19903-19912, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[23] M. Kirchhof, K. Roth, Z. Akata, and E. Kasneci. A non-isotropic probabilistic take on proxy-based deep metric learning. In Computer Vision - ECCV 2022 - 17th European Conference, Proceedings, Part XXVI, pages 435-454. Springer, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Elad Levi, Tete Xiao, Xiaolong Wang, and Trevor Darrell. Rethinking preventing class-collapsing in metric learning with margin-based losses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10316-10325, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25557" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[26] Christopher Liao, Theodoros Tsiligkaridis, and Brian Kulis. Supervised metric learning to rank for retrieval via contextual similarity optimization. In International Conference on Machine Learning, pages 20906-20938. PMLR, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[27] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning code. https://github.com/ljin0429/HIST, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[28] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 212-222, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[29] Lizhao Liu, Shangxin Huang, Zhuangwei Zhuang, Ran Yang, Mingkui Tan, and Yaowei Wang. Das: Densely-anchored sampling for deep metric learning. In European Conference on Computer Vision, pages 399–417. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[30] Ruofan Liu, Yun Lin, XIANGLIN YANG, and Jin Song Dong. Debugging and explaining metric learning approaches: An influence function based perspective. In Advances in Neural Information Processing Systems, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[31] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[32] Benjamin J. Meyer and Tom Drummond. The importance of metric learning for robotic vision: Open set recognition and active learning. In 2019 International Conference on Robotics and Automation (ICRA), pages 2924-2931, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[33] Timo Milbich, Karsten Roth, Homanga Bharadhwaj, Samarth Sinha, Yoshua Bengio, Björn Ommer, and Joseph Paul Cohen. Diva: Diverse visual feature aggregation for deep metric learning. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VIII, page 590–607, Berlin, Heidelberg, 2020. Springer-Verlag. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[34] Yair Movshovitz-Attias, Alexander Toshev, Thomas K Leung, Sergey Ioffe, and Saurabh Singh. No fuss distance metric learning using proxies. In Proceedings of the IEEE international conference on computer vision, pages 360–368, 2017. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[35] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[36] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Samyak Rawlekar, Shubhang Bhatnagar, Vishnuvardhan Pogunulu Srinivasulu, and Narendra Ahuja. Improving multi-label recognition using class co-occurrence probabilities. In International Conference on Pattern Recognition, pages 424–439. Springer, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[38] Karsten Roth, Timo Milbich, Samarth Sinha, Prateek Gupta, Björn Ommer, and Joseph Paul Cohen. Revisiting training strategies and generalization performance in deep metric learning. In Proceedings of the 37th International Conference on Machine Learning. JMLR.org, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[39] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7420-7430, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[40] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[42] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[43] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[44] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6397-6406, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[45] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H. S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1199-1208, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[46] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[47] Mark Tabb and Narendra Ahuja. Multiscale image segmentation by integrated edge and region detection. IEEE transactions on image processing, 6(5):642-655, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[48] Eu Wern Teh, Terrance DeVries, and Graham W. Taylor. Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV, pages 448-464. Springer, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[49] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Chengkun Wang, Wenzhao Zheng, Junlong Li, Jie Zhou, and Jiwen Lu. Deep factorized metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25558" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[51] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[52] Kilian Q Weinberger, John Blitzer, and Lawrence Saul. Distance metric learning for large margin nearest neighbor classification. In Advances in Neural Information Processing Systems. MIT Press, 2005. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[53] Chao-Yuan Wu, R Manmatha, Alexander J Smola, and Philipp Krahenbuhl. Sampling matters in deep embedding learning. In Proceedings of the IEEE international conference on computer vision, pages 2840-2848, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[54] Tong Xiao, Shuang Li, Bochao Wang, Liang Lin, and Xiaogang Wang. Joint detection and identification feature learning for person search. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3415-3424, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.483, + 0.402 + ], + "angle": 0, + "content": "[55] Hong Xuan, Abby Stylianou, and Robert Pless. Improved embeddings with easy positive triplet mining. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2463-2471, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[56] Bailin Yang, Haoqiang Sun, Frederick WB Li, Zheng Chen, Jianlu Cai, and Chao Song. Hse: Hybrid species embedding for deep metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11047-11057, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[57] Yuhui Yuan, Kuiyuan Yang, and Chao Zhang. Hard-aware deeply cascaded embedding. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[58] Andrew Zhai and Hao-Yu Wu. Classification is a strong baseline for deep metric learning. In *British Machine Vision Conference*, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[59] Wenzhao Zheng, Chengkun Wang, Jiwen Lu, and Jie Zhou. Deep compositional metric learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9316-9325, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[60] Wenzhao Zheng, Borui Zhang, Jiwen Lu, and Jie Zhou. Deep relational metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12065-12074, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.483, + 0.753 + ], + "angle": 0, + "content": "[61] Yuehua Zhu, Muli Yang, Cheng Deng, and Wei Liu. Fewer is more: A deep graph metric learning perspective using fewer proxies. In Advances in Neural Information Processing Systems, pages 17792-17803. Curran Associates, Inc., 2020. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25559" + } + ] +] \ No newline at end of file diff --git a/2025/Potential Field Based Deep Metric Learning/c5f58da8-3e8b-4fa3-accf-6f5c33627040_origin.pdf b/2025/Potential Field Based Deep Metric Learning/c5f58da8-3e8b-4fa3-accf-6f5c33627040_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ff14491410849671a841224f606411e77e9ddebe --- /dev/null +++ b/2025/Potential Field Based Deep Metric Learning/c5f58da8-3e8b-4fa3-accf-6f5c33627040_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd17ad4f5af4b11fa431f60fc5d4672389712a6c3be977e23093f536f99c7ee7 +size 819335 diff --git a/2025/Potential Field Based Deep Metric Learning/full.md b/2025/Potential Field Based Deep Metric Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..212f77bbebbe61d6165989b33a98da8574a2d616 --- /dev/null +++ b/2025/Potential Field Based Deep Metric Learning/full.md @@ -0,0 +1,302 @@ +# Potential Field Based Deep Metric Learning + +Shubhang Bhatnagar Narendra Ahuja University of Illinois Urbana-Champaign + +{sb56，n-ahuja}@illinois.edu + +# Abstract + +Deep metric learning (DML) involves training a network to learn a semantically meaningful representation space. Many current approaches mine $n$ -tuples of examples and model interactions within each tuplets. We present a novel, compositional DML model that instead of in tuples, represents the influence of each example (embedding) by a continuous potential field, and superposes the fields to obtain their combined global potential field. We use attractive/repulsive potential fields to represent interactions among embeddings from images of the same/different classes. Contrary to typical learning methods, where mutual influence of samples is proportional to their distance, we enforce reduction in such influence with distance, leading to a decaying field. We show that such decay helps improve performance on real world datasets with large intra-class variations and label noise. Like other proxy-based methods, we also use proxies to succinctly represent sub-populations of examples. We evaluate our method on three standard DML benchmarks-Cars-196,CUB-200-2011,and SOP datasets where it outperforms state-of-the-art baselines. + +# 1. Introduction + +The goal of visual metric learning is to learn a representation function where common distance metrics like the $\ell_2$ distance capture similarity between images. Semantic similarity is useful for applications such as image/video retrieval and search [3, 4, 21, 34], open set classification and segmentation [6, 14, 32, 37], few-shot learning [41, 45], person re-identification[7, 54] and face verification [10, 17, 31]. Deep neural networks have helped make significant advances in visual metric learning [40], as they are trained to learn a non-linear mapping of images onto a semantic space. + +Existing Methods: Popular loss functions to train such networks help force tuples of examples/samples/embeddings of the same class to be closer than those in different classes. Examples include the triplet loss [52], the contrastive loss [8] and their extensions [42, 51]. + +The utility of supervision provided by samples in a tu + +plet depends on samples' relative locations and labels. Most tuplet-based methods use sample mining strategies [15, 40] to choose a subset of informative tuplets to train the network, without which they suffer from lower performance and slower convergence [53]. Such strategies not only require the computationally demanding calculation of the relative hardness of all tuples formed $(\mathrm{O}(n^{3}))$ for triplets among $n$ samples), but also incur a loss of valuable supervision by discarding interactions. Specifically, limiting interactions to only a subset of all possible tuplets leads to greater susceptibility to noise [25], overfitting on the training set [26], and reduced learning of intra-class variations and generalizable features [38]. + +Proxy-based methods, first proposed in [34] attempt to mitigate some of these shortcomings by instead using class representative 'proxies' as one or more hypothesized centers for each class. Computation of the loss is then in terms of sample-proxy distances instead of sample-sample distances, helping circumvent the problem of tuplet sampling. However, these methods have been found to learn less generalizable intra-class features due to the exclusive use of proxy-sample distances while ignoring sample-sample distances[39]. This effect is compounded when proxies, whose locations are themselves learned, end up far from sample embedding locations they are intended to represent. + +Proposed Representation: To address these challenges, we propose Potential Field based Deep Metric learning (PFML), a metric learning framework that uses a continuous, potential field representation of all data samples motivated by potential fields used to model influence of charges in physics. Besides its analogical relevance, our motivation for using this representation comes from its previous, very successful uses, where it also helps overcome complexity in modeling interactions. These applications include planning paths of a robot avoiding obstacles [19] and planning an efficient strategy for a robot manipulator to grasp an object [20]. + +In PFML, we view each data sample as a charge, creating a field. The fields due to the individual samples are added (superposed) to obtain the global potential field. This global field thus constitutes a compositional representation, cap- + +turing the combined influence at a given location (or sample) as a weighted addition of the influences of the individual samples. Each sample generates two types of fields: (1) an attraction field to drive similar examples closer to it, and (2) a repulsion field that drives dissimilar examples away from it, each weakening with distance. + +We also make use of proxies like previous work, but only to augment the potential field generated by the current batch of samples. This is in contrast to past work that completely replaces sample-sample interactions with sample-proxy interactions. Preserving sample-sample relations through the potential field helps learn better intra-class features. + +Advantages of Potential Field: (1) Using the potential field representation enables us to model interactions among all sample embeddings, as opposed to modeling those among smaller number of samples (or proxies) as done in methods using point-tuplet based (e.g., triplet) loss. (2) Addition of potentials due to all points increases robustness to noise since the effect of noise on interactions among a smaller number of samples will have a larger variance while also (3) Considering all interactions improves the quality of features learned, unlike tuplet based methods which lose performance when all interactions are considered[53]. (4) This is because of a major feature that differentiates our potential field from past methods: The variation of strength of interaction between two points as the distance between them increases: instead of remaining constant or even becoming stronger, as is the case with most existing methods, e.g. [8, 21], in our model it becomes weaker with distance. The decay property is helpful in several ways: (4a) It ensures the intuitive expectation that two distant positive samples are too different to be considered as variants of each other, helping treat them as different varieties (e.g., associated with different proxies). (4b) The decay property also significantly improves PFML performance for the specific type of noise affecting labels, e.g., due to annotation errors common in real-world datasets (Sec. 4.3). (4c) As a result of the decay, the learned proxies remain closer to (at smaller Wasserstein distance $W_{2}$ from) the sample embeddings they represent than for (e.g., current proxy-based) methods where interactions strengthen with distance [21, 34, 36]; Sec 3.4), thereby enhancing their desired role. + +Contributions: The main contributions of the proposed approach are: + +- We present PFML, a novel framework for metric learning with a compositional representation using a continuous, potential field to model all sample interactions instead of only interactions between a subset of all possible n-tuplets. This enables better feature learning, especially + +in the presence of annotation noise. + +- PFML reverses the almost universally used model of inter-sample interactions strengthening with distance; interaction strength instead weakens with distance. This (1) helps improve robustness in the presence of label noise, and (2) we prove that it helps better align the distribution of proxies with the sample embeddings they represent. +- We evaluate PFML on three zero-shot image retrieval benchmarks: Cars-196, CUB-200-2011, and the SOP dataset where it outperforms current, state-of-the-art techniques in both the standard no-label-noise scenario, and with even greater margins ( $>7\%$ R@1) in the more realistic cases of label noise. + +# 2. Related Work + +Broadly, DML may be divided into two types, ranking/pair-based and proxy-based [21, 51]. + +Ranking/Pair based DML: These methods utilize losses that derive information from relations between intersample distances in pairs/ tuples of sample embeddings. Examples include the classical contrastive loss [8] or the triplet loss[40, 52, 53], which extends the concept to take into account relative distances in a triplet. [42] and [43] generalize these constraints to include multiple dissimilar samples. Mining informative pairs/tuples from a mini-batch plays an important role in the effectiveness of pair-based methods. Several strategies [15, 40, 57] have been proposed to pick samples based on their relative difficulty, with hard and semi-hard negative (picking pair of inter-class examples close together) and easy positive (same class examples close together) [25, 55] being popular. PFML naturally emphasizes such local (dis) similarities by virtue of its decaying formulation without the complexity of mining tuples. Also note that unlike PFML, none of these strategies optimize all interactions among a batch of samples. + +Proxy-based DML: ProxyNCA [34] and its variants [23, 48] attempt to minimize the distance of a sample from the proxy of its class while maximizing the distance from proxies of other classes. Proxy Anchor [21] builds on these by proposing a modified loss that takes into account sample hardness. [10, 36] introduced the use of multiple proxies per class to better model sub-clusters. + +Potential Field: Potential fields have been used in a variety of applications to model interactions of objects. Of interest here is their use in robot path planning [9, 19], shape representation [2, 18], image segmentation [1, 47] and object manipulation [20]. Here, they have been used to direct a robot towards its goal using an attraction field while avoiding obstacles using repulsion fields. Training supervision for DML shares a similar goal of moving sample embeddings towards other nearby samples of the same class, while avoiding dissimilar examples belonging to other classes. + +![](images/813a457e2650901ca8de8d420e0881ffa96b784bbe17ed61fbed1b82b643c9ef.jpg) +Figure 1. Overview of our Potential-field based DML pipeline. The process includes (1) Computing attraction and repulsion fields generated by each embedding and proxy, (2) Computing the class potential fields by superposition of individual fields (3) Evaluating total potential energy by summing up the potentials of embeddings and proxies under the class potential field and (4) Updating locations of sample embeddings (through network parameters) and proxies to minimize total potential energy through backprop. + +# 3. Method + +# 3.1. Problem Setup + +The goal of deep metric learning is to define a semantic distance metric $d(\mathbf{x}_1, \mathbf{x}_2)$ for any pair of samples $\mathbf{x}_1, \mathbf{x}_2 \in \mathcal{D}$ , where $\mathcal{D} = \{\mathbf{x}_i, y_i\}$ , $i \in \{1, \dots, |\mathcal{D}|\}$ and labels $y_i \in \{1, \dots, N\}$ represent $N$ classes. The learned distance metric for any two points $d(\mathbf{x}_1, \mathbf{x}_2)$ should represent their semantic dissimilarity. Instead of directly learning the distance metric, a projector function $f_\theta$ is often represented by a neural network, parameterized by $\theta$ learned from data and distance $d(\mathbf{x}_1, \mathbf{x}_2) = \|f_\theta(\mathbf{x}_1), f_\theta(\mathbf{x}_2)\|_2$ is given by the Euclidean distance between the projections/embeddings. For simplicity, we will refer to the embeddings $f_\theta(\mathbf{x}_i)$ belonging to a batch $\mathcal{B}$ sampled from $\mathcal{D}$ as $\mathbf{z}_i$ . + +# 3.2. The Potential Field + +With our potential field representation, embeddings need to be driven towards other nearby embeddings belonging to the same class, while also being driven away from embeddings of other classes. This is reminiscent of the behavior of an isolated system of electric charges, where dissimilar charges are drawn together while similar ones are repelled. Analogously, we define attractive and repulsive potentials created by a single embedding, with the net potential generated by a set of samples given by superposing individual potentials exerted by each embedding. The gradient of the net potential at a point yields the net force (on network update) on a sample embedding located at that point. + +As samples of each class have a different set of similar and dissimilar points, we define a separate potential field + +$\Psi_{j}$ , for each class $j \in \{1, \dots, N\}$ . The class potential field $\Psi_{j}$ models the force applied on embeddings belonging to class $j$ due to all embeddings. It is formed from a superposition of potentials belonging to sample embeddings from all classes, with embeddings from the same class exerting an attractive potential field (to pull similar points closer) while those from other classes exert a repulsive potential field (to push apart dissimilar points). We assume that each sample embedding $z_{i}$ has a unit charge, so the net force on it is the gradient of the potential field $\Psi_{y_i}$ of its class. We design the attractive and repulsive potential functions by taking motivation from both, electrostatic fields and metric learning literature, which we discuss in detail in the following subsections. Figure 1 provides an overview of our complete pipeline of generation and use of the potential field for metric learning. + +# 3.2.1. Potential field exerted by a single point + +By using forces on the embeddings, we wish the training process to move the embeddings to a state of equilibrium at the end of training in which all embeddings from a class are nearby. To have a potential function exhibiting this behavior, we ensure that individual potential field has the following two properties: + +Property 1 - Extremum Property: Potential generated by an individual embedding achieves an extremum at the embedding (minimum for attraction field, maximum for repulsion). This ensures that similar points are pulled together and dissimilar ones are pushed apart on minimizing potential. + +Property 2 - Decay Property: The magnitude of the gradient of the potential (force) due to an embedding decays with distance. Such a decay ensures two desired characteristics. (1) An embedding is not pulled towards a distant sample embedding belonging to the same class, which is likely to be too different to be considered as a variant of it; instead, the decayed influence helps treat the distant sample as significantly different variety of the class, associated with other nearby embeddings of the class. (2) Like for a single charge in the Extremum property, a superposition of potentials of different embeddings still yields local minima at or near sample embeddings. As shown in Sec 3.4, this helps proxies come close to such minima and thus help better model the data distribution. + +Given these properties motivated by electrostatic fields, we select the attraction and repulsion potentials $\psi_{att},\psi_{rep}$ generated by an embedding $\mathbf{z}_i$ , with their strength at location $\mathbf{r}$ defined as: + +$$ +\psi_ {a t t} (\mathbf {r}, \mathbf {z} _ {i}): = \left\{ \begin{array}{l l} - \frac {1}{\delta^ {\alpha}} & \text {i f} \| \mathbf {r} - \mathbf {z} _ {i} \| _ {2} < \delta \\ - \frac {1}{\| \mathbf {r} - \mathbf {z} _ {i} \| _ {2} ^ {\alpha}} & \text {o t h e r w i s e .} \end{array} \right. \tag {1} +$$ + +![](images/2e7644140292bd2441b2028d58a4b5c4b51a0f0ea687f50f5cc7a25ec850fc17.jpg) +(a) Class Potential field $\Psi_{1}$ + +![](images/2268a3e65bdeb8101e6be370584568f4ed0ab166f27ed7ba22e3b0ce06f93d4b.jpg) +(b) Class Potential field $\Psi_{2}$ +Figure 2. An example of the class potential fields $\Psi_{1}$ and $\Psi_{2}$ (Sec. 3.2.3), created by superposing the fields of individual embeddings (Sec 3.2.1) belonging to Classes 1 and 2. Arrows denote the gradient of the respective potentials representing the net force on them. (a) $\Psi_{1}$ draws samples/proxies of class 1 towards nearby samples/ proxies of class 1 while keeping them at least $\delta$ distance away from embeddings of class 2. Proxies at starred locations are drawn towards the nearest Class 1 embeddings which are potential minima, helping them better model the data distribution (Sec 3.4). (b) $\Psi_{2}$ draws embeddings of class 2 towards other nearby embeddings of class 2 instead of distant class 2 embeddings, which might be a significantly different variant of the class (or potentially a mislabeled data point), helping better feature learning (or improving noise robustness if mislabeled). $\Psi_{2}$ also keeps class 2 embeddings $\delta$ distance them away from embeddings of class 1. + +$$ +\psi_ {r e p} (\mathbf {r}, \mathbf {z} _ {i}): = \left\{ \begin{array}{l l} \frac {1}{\| \mathbf {r} - \mathbf {z} _ {i} \| _ {2} ^ {\alpha}} & \| \mathbf {r} - \mathbf {z} _ {i} \| _ {2} < \delta \\ \frac {1}{\delta^ {\alpha}} & \text {o t h e r w i s e .} \end{array} \right. \tag {2} +$$ + +Here, $\alpha$ is a hyperparameter that determines the rate of decay of the field (we evaluate it's effect in Sec. 4.5 instead of fixing $\alpha = 1$ as done in electrostatic fields). Note that the attraction potential is negative and increases with distance, implying that the force between the embedding and the point is attractive. The repulsion potential is positive and decreasing with distance, indicating repulsion. + +A constant attraction potential inside the $\delta$ sphere (i.e., force $= 0$ ) ensures that sample embeddings within $\delta$ distance of one another do not continue to attract each other, thus preventing the network from embeddings from collapsing to a single point which harms generalizability of the features learned (analogous to the effect margins used in tuplet-based losses have). + +Limiting the repulsion potential to a sphere of radius $\delta$ ensures that embeddings not belonging to the same class are driven to a distance $> \delta$ . Since the aforementioned constant attraction potential ensures that $\delta$ is the maximum distance between samples $z_{i}$ from the same class, this leads to connected components of $\delta$ spheres formed by nearby embeddings of the same class. Extending repulsion to distances $> \delta$ does not provide any useful additional supervisory information while forcing further unmotivated inter-class sep + +aration; as we empirically show in Sec. 4.5, such excessive separation does indeed harm performance. + +# 3.2.2. Proxies to Represent Sample Populations + +To represent the image embeddings not sampled in the current batch $\mathcal{B}$ , we propose to use a set of $M$ proxies per class for each of the N classes, $\mathbf{p}_{j,k}, k \in \{1, \dots, M\}, j \in \{1, \dots, N\}$ , and augment the potential field generated by the embeddings in the batch by those generated by the proxies. The proxies serve as stand-ins for the larger numbers of out-of-the-batch embeddings for calculating the attraction and repulsion fields, while also being affected by the field themselves like all embeddings. The proxies, being learnable parameters with high learning rates, tend to gravitate toward the nearest class modes of the potential field (and hence the class data distribution). The number of proxies M chosen should be large enough to capture the class-modes. + +# 3.2.3. Class Potential Field + +Given a batch $\mathcal{B}$ of samples and all proxies, we calculate the class potential field $\Psi_{j}(\mathbf{r})$ at a location $\mathbf{r}$ due to a class $j$ as a superposition of their attraction and repulsion potentials: + +$$ +\Psi_ {j} (\mathbf {r}) = \Psi_ {j, a t t} (\mathbf {r}) + \Psi_ {j, r e p} (\mathbf {r}), \text {w h e r e} \tag {3} +$$ + +$$ +\Psi_ {j, a t t} (\mathbf {r}) = \sum_ {i = 1} ^ {\| \mathcal {B} \|, y _ {i} = j} \psi_ {a t t} (\mathbf {r}, \mathbf {z} _ {i}) + \sum_ {k = 1} ^ {M} \psi_ {a t t} (\mathbf {r}, \mathbf {p} _ {j, k}) \tag {4} +$$ + +$$ +\Psi_ {j, r e p} (\mathbf {r}) = \sum_ {i = 1} ^ {\| \mathcal {B} \|, y _ {i} \neq j} \psi_ {r e p} (\mathbf {r}, \mathbf {z} _ {i}) + \sum_ {\gamma = 1, \gamma \neq j} ^ {N} \sum_ {k = 1} ^ {M} \psi_ {r e p} (\mathbf {r}, \mathbf {p} _ {\gamma , k}) \tag {5} +$$ + +Figure 2 visualizes the class potential fields $\Psi_{1}$ and $\Psi_{2}$ for a given set of embeddings belonging to two classes. + +# 3.3. Training + +Using gradient descent, the network is trained to minimize the total potential energy $\mathcal{U}$ of all proxies and embeddings in the batch $\mathbf{z}_i\in \mathcal{B}$ by applying a force on embeddings and proxies in the direction of $-\nabla \Psi_{y_i}(\mathbf{z}_i)$ . $\mathcal{U}$ given by: + +$$ +\mathcal {U} = \sum_ {i = 1} ^ {\| \mathcal {B} \|} \Psi_ {y _ {i}} (\mathbf {z} _ {i}) + \sum_ {j = 1} ^ {N} \sum_ {k = 1} ^ {M} \Psi_ {j} (\mathbf {p} _ {j, k}) \tag {6} +$$ + +# 3.4. Why decay interaction strength with distance? + +PFML uses a potential field where the gradient (interaction strength) decays with distance, which is unlike previous approaches. This leads to the two benefits described in Sec 3.2.1, which we discuss in more detail here. + +# 3.4.1. Robustness when learning with datasets having large intra-class variations and noisy labels + +The decay property helps PFML model all interactions even when using datasets with large intra-class variations. As seen in Figure 2, PFML draws embeddings towards other nearby embeddings from the same class rather than distant class embeddings that might be too different to be considered a variation of it. It also helps PFML become significantly more robust to the effect of incorrectly labeled examples as compared to existing methods. Recent work [26, 30] shows that such examples are common in metric learning benchmarks. The decay de-emphasizes the contribution of such unlabeled examples, that are typically located away from correctly labeled samples. We empirically confirm this in Sec. 4.3, demonstrating that methods not adhering to the decay property (i.e., all previous tuplet-based and proxy-based methods) experience significantly greater performance degradation due to label noise compared to PFML. + +# 3.4.2. Better alignment of proxies with data distribution + +The decaying interaction strength helps PFML better align the distribution of proxies with the underlying sample data distribution (of the complete dataset) that they represent. To analyze this, we compare PFML with the closely related contrastive loss [8] that also models pairwise interactions like PFML (defined in Eq. 1 & 2), and hence can be analyzed similarly by using a potential field. Specifically, we introduce Contrastive Potential based Metric Learning (CPML) by simply extending the contrastive loss by using multiple proxies per class, and subsequently analyzing the + +resulting loss function through a potential field framework. In CPML, strength of interaction is directly proportional to the square of the distance from it $(\|\mathbf{r} - \mathbf{r}_0\|_2^2)$ like in the original contrastive loss. For fair comparison, we keep all other details (use of $\delta$ , proxies) as the same, with its associated potential fields denoted by $\Psi_{att}^*$ , $\Psi_{rep}^*$ (formally defined in Supplement Sec 1). We analyze this theoretically using the following proposition and its corollary: + +Proposition 1: Let $Z = \{z_1 \ldots z_n\}$ be a set of sample embeddings belonging to a class, then there exists a $0 < \delta < \frac{\min_{i,j} \|z_i - z_j\|_2}{2\left(1 + \frac{1}{n}\right)}$ and points $z_{min,i}$ within $\delta$ distance from each embedding $z_i$ , $i \in \{1 \ldots n\}$ such that the attractive potential field $\Psi_{att}$ defined using $Z$ , $\delta$ has a minimum at each $z_{min,i}$ . The field $\Psi_{att}^*$ defined by them does not achieve a minimum at points within $\delta$ distance from all $z_i$ . + +Proposition 1 is a consequence of the potential field of an embedding dominating in its vicinity over potential fields of other distant embeddings. Its proof relies on the Decay property, which is not satisfied by $\Psi^{*}$ and can be found in Supp. Sec 1. A consequence of this is: + +Corollary 1: Let $Z = \{z_{1}\dots z_{n}\}$ be a set of sample embeddings from a class exerting an attraction field on a set of proxies $P = \{\pmb {p}_1\dots \pmb {p}_m\}$ . Consider the equilibrium distribution $P_{eq}$ of proxies minimizing the potential energy. If the potential field is defined by $\Psi_{att}$ , then the Wasserstein distance $W_{2}$ between $P_{eq}$ and the subset of data they represent is lower than when the potential field is defined by $\Psi_{att}^*$ . + +The proof for Corollary 1 relies on Proposition 1 and can be found in the Supplement (Sec. 2). + +Discussion: As proved in Corollary 1, CPML leads to proxies having a larger distance from the data distribution they represent (measured by $W_{2}$ ) than our proposed PFML, making the use of proxies in such a framework less effective. This leads to the proxies forming poorer stand-ins for the sample embeddings they represent, reducing the quality of features learned (verified in Sec. 4). This is caused by CPML not following the decay property (which is its only difference in design wrt PFML). Other proxy-based losses like the Proxy NCA [34], Soft Triplet [36], and Proxy Anchor [21] which perform similarly as CPML (see Tab. 1), also share with it the property of force increasing with distance (violating the decay property). Due to this, we expect them to similarly have a poorer alignment of proxies with underlying data distribution as compared to PFML (Applying an analysis similar to the one in Proposition 1 to them is considerably more challenging due to the absence of a potential field structure.). We verify this empirically in Sec. 4.4 by measuring the average distance between proxies and the closest data points ( $W_{2}$ distance) for these methods. + +# 3.5. Advantages in Learning Fine-grained Features + +In contrast with all previous proxy-based methods, PFML (1) Models sample-sample interactions directly using the potential field, gaining fine-grained supervision, and (2) Makes use of an explicit margin-like $\delta$ for proxies, which has been shown to help zero-shot generalization and fine-grained feature learning in tuplet-based methods[53]. Turning off the interaction within the $\delta$ radius also prevents a bias towards learning isotropic distributions as is present in other proxy-based methods [39], helping us obtain better intra-class resolution. + +# 4. Experiments and Results + +# 4.1. Setup + +Datasets: We empirically compare PFML with current state-of-the-art baselines on three public benchmark datasets (1) The Cars-196 dataset [24] which includes 16,185 images from 196 car model categories, (2) CUB200-2011 dataset [49] consisting of 11,788 images from 200 bird species, and (3) Stanford Online Products (SOP) dataset [43] containing 120,053 images of 22,634 different products sold online. The training and testing splits for the three datasets follow standard zero-shot retrieval-based evaluation protocol [21, 51, 61] of using the first half of the classes for training and the second half for testing. + +**Backbone:** To enable standardized comparison with a wide variety of methods, we evaluate our method using the Inception with batch normalization (IBN) [46], ResNet-50 [16], ViT [11] and DINO [5] backbones commonly used by DML methods. ImageNet pre-training is used for initialization, and only the last fully connected layer is changed to fit the dimension of the embedding needed. We follow past work [21, 36, 48] and $\ell_2$ normalize the final output. An input image size of $224 \times 224$ is used for all experiments. + +Training parameters: We train the backbone for 200 epochs on each dataset. We use the Adam optimizer with the learning rate chosen as $5e^{-4}$ . We scale up the learning rate of our proxies by 100 times following [21, 34, 36] for better convergence. We choose values of $\delta$ between $[0.1, 0.3]$ , $\alpha$ between $\{0, 6\}$ via cross-validation. We use $M = 15$ for the CUB-200 and Cars-196 datasets, while $M = 2$ is used for the SOP dataset. All experiments are performed on a single NVIDIA A4000 GPU. + +Evaluation Settings: We follow [21, 22, 29, 51] in using $224 \times 224$ sized center crops from $256 \times 256$ images for evaluation. We measure image retrieval performance using Recall@K, which computes the percentage of samples which have a valid similar neighbor (belonging to same class) among its K nearest neighbors. + +# 4.2. Image Retrieval Performance + +As seen in Table 1, our method outperforms all other methods on Image Retrieval on the CUB-200-2011, Cars-196 and SOP datasets as measured by the Recall@K metric. It consistently outperforms other methods irrespective of the backbone used, significantly outperforming the best-performing proxy-based methods (which similar to us, only change the loss function) like ProxyAnchor[21], ProxyGML[61] and SoftTriplet [36] in terms of Recall@1 (R@1), by more than $5\%$ on Cars-196, $3.7\%$ on CUB-200-2011 and $1.5\%$ on the SOP dataset. It also outperforms the significantly more complex, current state-of-the-art HIST[28] by $3.1\%$ , $2\%$ and $1.5\%$ in terms of R@1 when using the Resnet-50 backbone on the Cars-196, CUB-200-2011 and SOP datasets respectively. It also significantly outperforms the best tuplet-based loss: the Multi-Similarity loss by $6\%$ , $5.8\%$ and $2.4\%$ in terms of R@1 when using the BN Inception backbone on the Cars-196, CUB-200 and SOP datasets respectively. Additionally, our method outperforms CPML (Sec. 3.4.2, a simple combination of the classical contrastive loss with proxies) by $7.5\%$ , $5.1\%$ , and $3.5\%$ on CUB200, Cars and SOP datasets. Note that our performance gains are larger than current SOTA compared to improvements reported by previous efforts on the same benchmarks, despite our intuitive and simpler method. + +We also evaluate our method 1) Using the protocol and metrics proposed in [35] and (2) When used to learn smaller 128-dim embeddings for more efficient retrieval. Our method still outperforms all SOTA methods in both cases. Details of these experiments are in Supp. Sec. 3. We also show qualitative retrieval results of our method in Supp. Sec 6. + +# 4.3. Image Retrieval in Presence of Label Noise + +We evaluate the robustness of our method vs. state-of-the-art baselines in the realistic scenario of label noise (e.g., due to mislabeling, which is very common, particularly in fine-grained datasets such as those used for DML). Specifically, we randomly corrupt labels of $20\%$ of samples in the training set. As seen in the results in Table 2, this leads to a significant decline in the performance of all methods. But our method is the least affected by the noise. Other methods like the triplet [15] or ProxyAnchor[21] suffer much greater declines because they are affected more by distant points represented by the mislabeled examples. Our method significantly outperforms Proxy Anchor [21] (the second-best method) by $6\%$ and $7.6\%$ in terms of R@1 on the CUB-200-2011 and Cars-196 datasets. + +# 4.4. Distance Between Proxies and Sample Embeddings + +We measure the average (over a complete epoch) Wasserstein distance $W_{2}$ between proxies and the subset of training + +
Benchmarks →CUB-200-2011Cars-196SOP
Methods ↓ (Chronological)R@1R@2R@4R@1R@2R@4R@1R@10R@100
ResNet50 (512 dim)
ESPHN [55]64.975.383.582.789.393.078.390.796.3
N.Softmax [58]61.373.983.584.290.494.478.290.696.2
DiVA [33]69.279.3-87.692.9-79.691.2-
Proxy NCA++ [48]64.7---85.1--79.6--
Proxy Anchor [21]69.780.087.087.792.995.8---
DCML-MDW [59]68.477.986.185.291.896.079.890.895.8
MS+DAS [29]69.279.287.187.893.196.080.691.896.7
HIST [28]71.481.188.189.693.996.481.492.096.7
HIER[22]70.179.486.988.293.095.680.291.596.6
HSE-PA [56]70.680.187.189.693.896.080.091.496.3
CPML (Sec. 3.4)68.378.786.285.291.595.279.490.796.1
Potential Field (Ours)73.4 ± 0.382.4± 0.188.8± 0.192.7± 0.395.5 ± 0.197.6± 0.182.9 ± 0.292.5 ± 0.296.8 ± 0.1
BN Inception (512 dim)
HTL [13]57.168.878.781.488.092.774.888.394.8
MultiSimilarity [51]65.777.086.384.190.494.078.290.596.0
SoftTriple [36]65.476.484.584.590.794.578.390.395.9
CircleLoss [44]66.777.486.283.489.894.178.390.596.1
DiVA [33]66.877.7-84.190.7-78.190.6-
ProxyGML [61]66.677.686.485.591.895.378.090.696.2
Proxy Anchor [21]68.479.286.886.191.795.079.190.896.2
DRML-PA [60]68.778.686.386.992.195.271.585.293.0
MS+DAS [29]67.178.1186.485.791.695.378.290.396.0
HIST [28]69.780.087.387.492.595.479.691.096.2
DFML-PA [50]69.3--88.4-----
HSE-M [56]67.678.085.882.088.993.3---
PA+niV [23]69.580.0-86.492.0-79.290.4-
Potential Field (Ours)71.5± 0.381.2± 0.288.3± 0.290.1± 0.293.9± 0.196.3± 0.180.6 ± 0.391.8 ± 0.196.4 ± 0.1
DINO (384 dim)
DINO [5]70.881.188.842.953.964.263.478.188.3
Hyp [12]80.987.692.489.294.196.785.194.497.8
HIER [22]81.188.293.391.395.297.185.794.697.8
Potential Field (Ours)83.1 ± 0.389.3 ± 0.294.2 ± 0.194.7 ± 0.196.5 ± 0.197.8 ± 0.186.5 ± 0.395.1 ± 0.398.0 ± 0.2
ViT (384 dim)
ViT-S [11]83.190.494.447.860.272.262.177.789.0
Hyp [12]85.691.494.886.592.195.385.994.998.1
HIER [22]85.791.394.488.393.296.186.195.098.0
Potential Field (Ours)87.8 ± 0.292.6 ± 0.295.7 ± 0.191.5 ± 0.395.2 ± 0.297.4 ± 0.188.2 ± 0.195.7 ± 0.198.6 ± 0.1
+ +Table 1. Comparison of the Recall@ $K$ (%) achieved by our method on the CUB-200-2011, Cars-196 and SOP datasets with state-of-the-art baselines under standard settings. The table reports the average performance and standard deviations of our method over 5 runs. + +
CUB-200-2011Cars-196
MethodsR@1R@2R@1R@2
Triplet[52]55.168.767.577.9
MS [51]58.971.870.479.8
PNCA[34]60.174.774.382.4
PA [21]60.775.176.983.1
HIST[27]59.774.672.981.8
Ours66.7 ± 0.676.9± 0.384.5± 0.588.6± 0.3
+ +Table 2. A comparison of Recall@ $K$ (%) achieved by our method with SOTA baselines on CUB-200-2011 and Cars-196 datasets with label noise when using a Resnet-50 backbone (512-dim). Performance and std. deviations reported are over 5 runs. + +
MethodAverage W2
S.Triplet [36]0.24 ± 0.04
PNCA [21]0.36 ± 0.02
PA [21]0.38 ± 0.03
CPML (Sec. 3.4)0.26 ± 0.03
Ours0.16 ± 0.03
+ +Table 3. Average Wasserstein distance between proxies and the subset of data closest to them for different methods on CUB-200-2011 dataset. Mean and std. deviations are over 5 runs. + +data represented by/closest to them ${}^{2}$ for different proxy- + +based methods. We train networks on the CUB-200-2011 using the settings in Section 4.1 and use the same number + +![](images/83c48e49831ecbab92c541f90b6f3a53a6191a03366d83260c4ad7d758c87a36.jpg) +Figure 3. Variation in Recall@1 with $M$ , $\delta$ , $\alpha$ and $\delta_{rep} - \delta_{att}$ on the Cars-196 and CUB-200-2011 datasets. Error bars represent std. deviations over 5 runs. + +![](images/f78862198ec348277fa0eb0ca5d637f108e1e078ac781392dfd98b544255fa50.jpg) + +![](images/2189e2657748765a4f2dffad8fcc2bc9d0323c667997ac10b20e4315d8bfb76b.jpg) + +![](images/92e2a474b7da5dcf709eb8a568784c588174eced910efb75b936f870f69f40fa.jpg) + +of proxies for all methods utilizing multiple proxies. As seen in Table 3, our method shows $33.3\%$ lower $W_{2}$ distance between the proxies and the subset of data than the second best SoftTriplet [36] loss, empirically validating our discussion in Sec. 3.4.2. + +# 4.5. Ablation studies + +In all experiments, we use the parameters mentioned in Section 4.1 with a ResNet-50 backbone on both the Cars-196 and CUB-200-2011 datasets. + +Effect of proxies and their number: Proxies augment the potential field, and the number of proxies per class $M$ decides the granularity at which the proxies model the distribution of sample embeddings. We study their effect on performance by varying $M$ between [0, 30]. From the results plotted in Figure 3(a) we observe that (1) Using proxies to augment the field boosts performance (vs using no-proxies) and (2) Network performance remains stable for a wide range of M, decreasing slightly for small values of M (still outperforming most other proxy-based methods). This behavior is expected as using potential field created using a very small M would not be able to effectively model underlying sample distribution, leading to worse supervision. We also observe that using a larger number of proxies per class does not significantly affect training times, as bulk of computation is dominated by the network. For further details, please refer to Supp. Sec. 5. + +Effect of varying $\delta$ : We conduct an empirical study to determine the effect of varying the radius parameter $\delta$ on performance by varying it over [0.1, 0.35]. As seen in Figure 3(b), we observe that performance remains relatively stable for a wide range of $\delta$ values, demonstrating our method's robustness to it. + +Effect of potential decay $\alpha$ : The decay parameter $\alpha$ determines the rate at which the potential field of an embedding decays with distance. We study its effect by varying it over [0, 6]. As seen in Figure 3(c), performance is worse when the potential field does not decay ( $\alpha = 0$ ). A field that decays strongly $\alpha \in [3, 6]$ seems to perform better than one with a much milder decay ( $\alpha \in [1, 2]$ ). The likely reason for this trend is that when the potential field has no decay /decays mildly, the force on an embedding would still be dominated by a large number of distant sample embeddings + +rather than those in its vicinity. + +Restricting repulsion to a distance of $\delta$ : Unlike most proxy-based methods, our design of potential field $\Psi_{rep}$ limits the range of repulsion to within $\delta$ distance from embeddings or proxies. We examine extending this range of repulsion beyond the $\delta$ radius by having $\delta_{rep} > \delta_{att}$ where $\delta_{att}$ and $\delta_{rep}$ represent the values of $\delta$ used for attraction (Eq.1) and repulsion (Eq.2) fields. Fixing $\delta_{att} = 0.2$ , we increase $\delta_{rep}$ to study the effect. As shown in Figure 3(d), his reduces performance, validating our design. This is due to individual potentials repelling dissimilar points to a distance of at least $\delta_{rep}$ , greater than the distance $\delta_{att}$ within which all points of the same class are attracted to. Using $\delta_{rep} > \delta_{att}$ enforces excessive separation between dissimilar examples, compressing intra-class features more than needed. + +# 5. Limitations + +PFML relies on proxies to augment potential fields in each batch, hence sharing the main limitation of proxy-based methods: with a large number of classes, the computational cost during training increases due to linear scaling of proxies with class count. Despite this, proxy-based methods have been widely used in applications with a large number of classes like face recognition (e.g. [10] on 500,000 classes), likely because (1) proxies only affect training, not inference complexity and (2) The parameters/compute complexity added by the proxies is relatively low for the performance gains they offer, especially when such compute complexity is compared to the forward and backward passes of the neural network that consume bulk of compute in DML. + +# 6. Conclusion + +PFML introduces a novel paradigm to capture interactions between all samples through a continuous potential field, where strength of sample influence decreases with distance. This reversal of the common model of increasing influence with distance improves robustness to label noise and better aligns proxy distributions with sample embeddings. The overall design of our method is validated by its performance on standard image retrieval benchmarks, where it outperforms state-of-the-art methods both in the standard noise-free setting and in the presence of label noise unavoidable in real-world datasets. + +# References + +[1] Narendra Ahuja. A transform for multiscale image segmentation by integrated edge and region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 18 (12):1211-1235, 1996. 2 +[2] N. Ahuja and Jen-Hui Chuang. Shape representation using a generalized potential field model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(2):169-176, 1997. 2 +[3] Shubhang Bhatnagar and Narendra Ahuja. Piecewise-linear manifolds for deep metric learning. In Conference on Parsimony and Learning (Proceedings Track), 2023. 1 +[4] Biagio Brattoli, Joseph Tighe, Fedor Zhdanov, Pietro Perona, and Krzysztof Chalupka. Rethinking zero-shot video classification: End-to-end training for realistic applications. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 4612-4622. Computer Vision Foundation / IEEE, 2020. 1 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6, 7 +[6] Jun Cen, Peng Yun, Junhao Cai, Michael Yu Wang, and Ming Liu. Deep metric learning for open world semantic segmentation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15313-15322, 2021. 1 +[7] Weihua Chen, Xiaotang Chen, Jianguo Zhang, and Kaiqi Huang. Beyond triplet loss: A deep quadruplet network for person re-identification. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1320-1329. IEEE Computer Society, 2017. 1 +[8] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 539-546. IEEE, 2005. 1, 2, 5 +[9] Jen-Hui Chuang and N. Ahuja. An analytically tractable potential field model of free space and its application in obstacle avoidance. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 28(5):729-736, 1998. 2 +[10] J. Deng, J. Guo, N. Xue, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4685-4694, 2019. 1, 2, 8 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6, 7 +[12] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric + +learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7409-7419, 2022. 7 +[13] Weifeng Ge, Weilin Huang, Dengke Dong, and Matthew R. Scott. Deep metric learning with hierarchical triplet loss. In Computer Vision - ECCV 2018, pages 272-288, Cham, 2018. Springer International Publishing. 7 +[14] Matheus Gutoski, André Eugénio Lazzaretti, and Heitor Silvério Lopes. Deep metric learning for open-set human action recognition in videos. Neural Comput. Appl., 33(4): 1207-1220, 2021. 1 +[15] Ben Harwood, Vijay Kumar BG, Gustavo Carneiro, Ian Reid, and Tom Drummond. Smart mining for deep metric learning. In Proceedings of the IEEE international conference on computer vision, pages 2821-2829, 2017. 1, 2, 6 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[17] J. Hu, J. Lu, and Y. Tan. Discriminative deep metric learning for face verification in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014. 1 +[18] Y.K. Hwang and N. Ahuja. Path planning using a potential field representation. In Proceedings. 1988 IEEE International Conference on Robotics and Automation, pages 648-649 vol.1, 1988. 2 +[19] Y.K. Hwang and N. Ahuja. A potential field approach to path planning. IEEE Transactions on Robotics and Automation, 8 (1):23-32, 1992. 1, 2 +[20] O. Khatib. Real-time obstacle avoidance for manipulators and mobile robots. In Proceedings. 1985 IEEE International Conference on Robotics and Automation, pages 500-505, 1985. 1, 2 +[21] Sungyeon Kim, Dongwon Kim, Minsu Cho, and Suha Kwak. Proxy anchor loss for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 5, 6, 7 +[22] Sungyeon Kim, Boseung Jeong, and Suha Kwak. Hier: Metric learning beyond class labels via hierarchical regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19903-19912, 2023. 6, 7 +[23] M. Kirchhof, K. Roth, Z. Akata, and E. Kasneci. A non-isotropic probabilistic take on proxy-based deep metric learning. In Computer Vision - ECCV 2022 - 17th European Conference, Proceedings, Part XXVI, pages 435-454. Springer, 2022. 2, 7 +[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 6 +[25] Elad Levi, Tete Xiao, Xiaolong Wang, and Trevor Darrell. Rethinking preventing class-collapsing in metric learning with margin-based losses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10316-10325, 2021. 1, 2 + +[26] Christopher Liao, Theodoros Tsiligkaridis, and Brian Kulis. Supervised metric learning to rank for retrieval via contextual similarity optimization. In International Conference on Machine Learning, pages 20906-20938. PMLR, 2023. 1, 5 +[27] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning code. https://github.com/ljin0429/HIST, 2022. 7 +[28] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 212-222, 2022. 6, 7 +[29] Lizhao Liu, Shangxin Huang, Zhuangwei Zhuang, Ran Yang, Mingkui Tan, and Yaowei Wang. Das: Densely-anchored sampling for deep metric learning. In European Conference on Computer Vision, pages 399–417. Springer, 2022. 6, 7 +[30] Ruofan Liu, Yun Lin, XIANGLIN YANG, and Jin Song Dong. Debugging and explaining metric learning approaches: An influence function based perspective. In Advances in Neural Information Processing Systems, 2022. 5 +[31] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1 +[32] Benjamin J. Meyer and Tom Drummond. The importance of metric learning for robotic vision: Open set recognition and active learning. In 2019 International Conference on Robotics and Automation (ICRA), pages 2924-2931, 2019. 1 +[33] Timo Milbich, Karsten Roth, Homanga Bharadhwaj, Samarth Sinha, Yoshua Bengio, Björn Ommer, and Joseph Paul Cohen. Diva: Diverse visual feature aggregation for deep metric learning. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VIII, page 590–607, Berlin, Heidelberg, 2020. Springer-Verlag. 7 +[34] Yair Movshovitz-Attias, Alexander Toshev, Thomas K Leung, Sergey Ioffe, and Saurabh Singh. No fuss distance metric learning using proxies. In Proceedings of the IEEE international conference on computer vision, pages 360–368, 2017. 1, 2, 5, 6, 7 +[35] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 6 +[36] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 5, 6, 7, 8 +[37] Samyak Rawlekar, Shubhang Bhatnagar, Vishnuvardhan Pogunulu Srinivasulu, and Narendra Ahuja. Improving multi-label recognition using class co-occurrence probabilities. In International Conference on Pattern Recognition, pages 424–439. Springer, 2025. 1 + +[38] Karsten Roth, Timo Milbich, Samarth Sinha, Prateek Gupta, Björn Ommer, and Joseph Paul Cohen. Revisiting training strategies and generalization performance in deep metric learning. In Proceedings of the 37th International Conference on Machine Learning. JMLR.org, 2020. 1 +[39] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7420-7430, 2022. 1, 6 +[40] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1, 2 +[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1 +[42] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1, 2 +[43] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6 +[44] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6397-6406, 2020. 7 +[45] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H. S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1199-1208, 2017. 1 +[46] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6 +[47] Mark Tabb and Narendra Ahuja. Multiscale image segmentation by integrated edge and region detection. IEEE transactions on image processing, 6(5):642-655, 1997. 2 +[48] Eu Wern Teh, Terrance DeVries, and Graham W. Taylor. Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV, pages 448-464. Springer, 2020. 2, 6, 7 +[49] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6 +[50] Chengkun Wang, Wenzhao Zheng, Junlong Li, Jie Zhou, and Jiwen Lu. Deep factorized metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2023. 7 + +[51] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6, 7 +[52] Kilian Q Weinberger, John Blitzer, and Lawrence Saul. Distance metric learning for large margin nearest neighbor classification. In Advances in Neural Information Processing Systems. MIT Press, 2005. 1, 2, 7 +[53] Chao-Yuan Wu, R Manmatha, Alexander J Smola, and Philipp Krahenbuhl. Sampling matters in deep embedding learning. In Proceedings of the IEEE international conference on computer vision, pages 2840-2848, 2017. 1, 2, 6 +[54] Tong Xiao, Shuang Li, Bochao Wang, Liang Lin, and Xiaogang Wang. Joint detection and identification feature learning for person search. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3415-3424, 2017. 1 +[55] Hong Xuan, Abby Stylianou, and Robert Pless. Improved embeddings with easy positive triplet mining. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2463-2471, 2019. 2, 7 +[56] Bailin Yang, Haoqiang Sun, Frederick WB Li, Zheng Chen, Jianlu Cai, and Chao Song. Hse: Hybrid species embedding for deep metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11047-11057, 2023. 7 +[57] Yuhui Yuan, Kuiyuan Yang, and Chao Zhang. Hard-aware deeply cascaded embedding. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2 +[58] Andrew Zhai and Hao-Yu Wu. Classification is a strong baseline for deep metric learning. In *British Machine Vision Conference*, 2018. 7 +[59] Wenzhao Zheng, Chengkun Wang, Jiwen Lu, and Jie Zhou. Deep compositional metric learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9316-9325, 2021. 7 +[60] Wenzhao Zheng, Borui Zhang, Jiwen Lu, and Jie Zhou. Deep relational metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12065-12074, 2021. 7 +[61] Yuehua Zhu, Muli Yang, Cheng Deng, and Wei Liu. Fewer is more: A deep graph metric learning perspective using fewer proxies. In Advances in Neural Information Processing Systems, pages 17792-17803. Curran Associates, Inc., 2020. 6, 7 \ No newline at end of file diff --git a/2025/Potential Field Based Deep Metric Learning/images.zip b/2025/Potential Field Based Deep Metric Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e8b698662ff7f987edcb79fa1571ca84f2be14b9 --- /dev/null +++ b/2025/Potential Field Based Deep Metric Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de92be3674fd8f14925b3d71ae9f69b1cc388be55ee0a70e73e862f381328aea +size 441133 diff --git a/2025/Potential Field Based Deep Metric Learning/layout.json b/2025/Potential Field Based Deep Metric Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..66d1272fd6b96c673ce53306f14813e44eb1062f --- /dev/null +++ b/2025/Potential Field Based Deep Metric Learning/layout.json @@ -0,0 +1,9646 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 169, + 103, + 442, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 103, + 442, + 121 + ], + "spans": [ + { + "bbox": [ + 169, + 103, + 442, + 121 + ], + "type": "text", + "content": "Potential Field Based Deep Metric Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 201, + 144, + 408, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 144, + 408, + 172 + ], + "spans": [ + { + "bbox": [ + 201, + 144, + 408, + 172 + ], + "type": "text", + "content": "Shubhang Bhatnagar Narendra Ahuja University of Illinois Urbana-Champaign" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 174, + 380, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 174, + 380, + 184 + ], + "spans": [ + { + "bbox": [ + 228, + 174, + 380, + 184 + ], + "type": "text", + "content": "{sb56，n-ahuja}@illinois.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "type": "text", + "content": "Deep metric learning (DML) involves training a network to learn a semantically meaningful representation space. Many current approaches mine " + }, + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "type": "text", + "content": "-tuples of examples and model interactions within each tuplets. We present a novel, compositional DML model that instead of in tuples, represents the influence of each example (embedding) by a continuous potential field, and superposes the fields to obtain their combined global potential field. We use attractive/repulsive potential fields to represent interactions among embeddings from images of the same/different classes. Contrary to typical learning methods, where mutual influence of samples is proportional to their distance, we enforce reduction in such influence with distance, leading to a decaying field. We show that such decay helps improve performance on real world datasets with large intra-class variations and label noise. Like other proxy-based methods, we also use proxies to succinctly represent sub-populations of examples. We evaluate our method on three standard DML benchmarks-Cars-196,CUB-200-2011,and SOP datasets where it outperforms state-of-the-art baselines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 501, + 135, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 135, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 135, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 641 + ], + "type": "text", + "content": "The goal of visual metric learning is to learn a representation function where common distance metrics like the " + }, + { + "bbox": [ + 55, + 521, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 55, + 521, + 295, + 641 + ], + "type": "text", + "content": " distance capture similarity between images. Semantic similarity is useful for applications such as image/video retrieval and search [3, 4, 21, 34], open set classification and segmentation [6, 14, 32, 37], few-shot learning [41, 45], person re-identification[7, 54] and face verification [10, 17, 31]. Deep neural networks have helped make significant advances in visual metric learning [40], as they are trained to learn a non-linear mapping of images onto a semantic space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 701 + ], + "type": "text", + "content": "Existing Methods: Popular loss functions to train such networks help force tuples of examples/samples/embeddings of the same class to be closer than those in different classes. Examples include the triplet loss [52], the contrastive loss [8] and their extensions [42, 51]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "The utility of supervision provided by samples in a tu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "type": "text", + "content": "plet depends on samples' relative locations and labels. Most tuplet-based methods use sample mining strategies [15, 40] to choose a subset of informative tuplets to train the network, without which they suffer from lower performance and slower convergence [53]. Such strategies not only require the computationally demanding calculation of the relative hardness of all tuples formed " + }, + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "type": "inline_equation", + "content": "(\\mathrm{O}(n^{3}))" + }, + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "type": "text", + "content": " for triplets among " + }, + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "type": "text", + "content": " samples), but also incur a loss of valuable supervision by discarding interactions. Specifically, limiting interactions to only a subset of all possible tuplets leads to greater susceptibility to noise [25], overfitting on the training set [26], and reduced learning of intra-class variations and generalizable features [38]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 373, + 556, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 517 + ], + "type": "text", + "content": "Proxy-based methods, first proposed in [34] attempt to mitigate some of these shortcomings by instead using class representative 'proxies' as one or more hypothesized centers for each class. Computation of the loss is then in terms of sample-proxy distances instead of sample-sample distances, helping circumvent the problem of tuplet sampling. However, these methods have been found to learn less generalizable intra-class features due to the exclusive use of proxy-sample distances while ignoring sample-sample distances[39]. This effect is compounded when proxies, whose locations are themselves learned, end up far from sample embedding locations they are intended to represent." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 518, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 556, + 663 + ], + "type": "text", + "content": "Proposed Representation: To address these challenges, we propose Potential Field based Deep Metric learning (PFML), a metric learning framework that uses a continuous, potential field representation of all data samples motivated by potential fields used to model influence of charges in physics. Besides its analogical relevance, our motivation for using this representation comes from its previous, very successful uses, where it also helps overcome complexity in modeling interactions. These applications include planning paths of a robot avoiding obstacles [19] and planning an efficient strategy for a robot manipulator to grasp an object [20]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "In PFML, we view each data sample as a charge, creating a field. The fields due to the individual samples are added (superposed) to obtain the global potential field. This global field thus constitutes a compositional representation, cap-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25549" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "turing the combined influence at a given location (or sample) as a weighted addition of the influences of the individual samples. Each sample generates two types of fields: (1) an attraction field to drive similar examples closer to it, and (2) a repulsion field that drives dissimilar examples away from it, each weakening with distance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 294, + 216 + ], + "type": "text", + "content": "We also make use of proxies like previous work, but only to augment the potential field generated by the current batch of samples. This is in contrast to past work that completely replaces sample-sample interactions with sample-proxy interactions. Preserving sample-sample relations through the potential field helps learn better intra-class features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 217, + 296, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 587 + ], + "type": "text", + "content": "Advantages of Potential Field: (1) Using the potential field representation enables us to model interactions among all sample embeddings, as opposed to modeling those among smaller number of samples (or proxies) as done in methods using point-tuplet based (e.g., triplet) loss. (2) Addition of potentials due to all points increases robustness to noise since the effect of noise on interactions among a smaller number of samples will have a larger variance while also (3) Considering all interactions improves the quality of features learned, unlike tuplet based methods which lose performance when all interactions are considered[53]. (4) This is because of a major feature that differentiates our potential field from past methods: The variation of strength of interaction between two points as the distance between them increases: instead of remaining constant or even becoming stronger, as is the case with most existing methods, e.g. [8, 21], in our model it becomes weaker with distance. The decay property is helpful in several ways: (4a) It ensures the intuitive expectation that two distant positive samples are too different to be considered as variants of each other, helping treat them as different varieties (e.g., associated with different proxies). (4b) The decay property also significantly improves PFML performance for the specific type of noise affecting labels, e.g., due to annotation errors common in real-world datasets (Sec. 4.3). (4c) As a result of the decay, the learned proxies remain closer to (at smaller Wasserstein distance " + }, + { + "bbox": [ + 56, + 217, + 296, + 587 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 56, + 217, + 296, + 587 + ], + "type": "text", + "content": " from) the sample embeddings they represent than for (e.g., current proxy-based) methods where interactions strengthen with distance [21, 34, 36]; Sec 3.4), thereby enhancing their desired role." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "text", + "content": "Contributions: The main contributions of the proposed approach are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 673 + ], + "type": "text", + "content": "- We present PFML, a novel framework for metric learning with a compositional representation using a continuous, potential field to model all sample interactions instead of only interactions between a subset of all possible n-tuplets. This enables better feature learning, especially" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 322, + 72, + 467, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 467, + 83 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 467, + 83 + ], + "type": "text", + "content": "in the presence of annotation noise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 84, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 555, + 156 + ], + "type": "text", + "content": "- PFML reverses the almost universally used model of inter-sample interactions strengthening with distance; interaction strength instead weakens with distance. This (1) helps improve robustness in the presence of label noise, and (2) we prove that it helps better align the distribution of proxies with the sample embeddings they represent." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 156, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 228 + ], + "type": "text", + "content": "- We evaluate PFML on three zero-shot image retrieval benchmarks: Cars-196, CUB-200-2011, and the SOP dataset where it outperforms current, state-of-the-art techniques in both the standard no-label-noise scenario, and with even greater margins (" + }, + { + "bbox": [ + 313, + 156, + 555, + 228 + ], + "type": "inline_equation", + "content": ">7\\%" + }, + { + "bbox": [ + 313, + 156, + 555, + 228 + ], + "type": "text", + "content": " R@1) in the more realistic cases of label noise." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 243, + 400, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 243, + 400, + 255 + ], + "spans": [ + { + "bbox": [ + 314, + 243, + 400, + 255 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "type": "text", + "content": "Broadly, DML may be divided into two types, ranking/pair-based and proxy-based [21, 51]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 291, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 506 + ], + "type": "text", + "content": "Ranking/Pair based DML: These methods utilize losses that derive information from relations between intersample distances in pairs/ tuples of sample embeddings. Examples include the classical contrastive loss [8] or the triplet loss[40, 52, 53], which extends the concept to take into account relative distances in a triplet. [42] and [43] generalize these constraints to include multiple dissimilar samples. Mining informative pairs/tuples from a mini-batch plays an important role in the effectiveness of pair-based methods. Several strategies [15, 40, 57] have been proposed to pick samples based on their relative difficulty, with hard and semi-hard negative (picking pair of inter-class examples close together) and easy positive (same class examples close together) [25, 55] being popular. PFML naturally emphasizes such local (dis) similarities by virtue of its decaying formulation without the complexity of mining tuples. Also note that unlike PFML, none of these strategies optimize all interactions among a batch of samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": "Proxy-based DML: ProxyNCA [34] and its variants [23, 48] attempt to minimize the distance of a sample from the proxy of its class while maximizing the distance from proxies of other classes. Proxy Anchor [21] builds on these by proposing a modified loss that takes into account sample hardness. [10, 36] introduced the use of multiple proxies per class to better model sub-clusters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Potential Field: Potential fields have been used in a variety of applications to model interactions of objects. Of interest here is their use in robot path planning [9, 19], shape representation [2, 18], image segmentation [1, 47] and object manipulation [20]. Here, they have been used to direct a robot towards its goal using an attraction field while avoiding obstacles using repulsion fields. Training supervision for DML shares a similar goal of moving sample embeddings towards other nearby samples of the same class, while avoiding dissimilar examples belonging to other classes." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "1It remains constant over a small radius and decays outside; however, for brevity, we will disregard this detail in the rest of this paper, and refer to the influence as decaying with distance, or as the decay property." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25550" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 294, + 229 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 294, + 229 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 294, + 229 + ], + "type": "image", + "image_path": "813a457e2650901ca8de8d420e0881ffa96b784bbe17ed61fbed1b82b643c9ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 239, + 296, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 327 + ], + "type": "text", + "content": "Figure 1. Overview of our Potential-field based DML pipeline. The process includes (1) Computing attraction and repulsion fields generated by each embedding and proxy, (2) Computing the class potential fields by superposition of individual fields (3) Evaluating total potential energy by summing up the potentials of embeddings and proxies under the class potential field and (4) Updating locations of sample embeddings (through network parameters) and proxies to minimize total potential energy through backprop." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 334, + 111, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 111, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 111, + 346 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 148, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 148, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 148, + 367 + ], + "type": "text", + "content": "3.1. Problem Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": "The goal of deep metric learning is to define a semantic distance metric " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}_1, \\mathbf{x}_2)" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " for any pair of samples " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1, \\mathbf{x}_2 \\in \\mathcal{D}" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{x}_i, y_i\\}" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, |\\mathcal{D}|\\}" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " and labels " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "y_i \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " represent " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " classes. The learned distance metric for any two points " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}_1, \\mathbf{x}_2)" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " should represent their semantic dissimilarity. Instead of directly learning the distance metric, a projector function " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "f_\\theta" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " is often represented by a neural network, parameterized by " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " learned from data and distance " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}_1, \\mathbf{x}_2) = \\|f_\\theta(\\mathbf{x}_1), f_\\theta(\\mathbf{x}_2)\\|_2" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " is given by the Euclidean distance between the projections/embeddings. For simplicity, we will refer to the embeddings " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "f_\\theta(\\mathbf{x}_i)" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " belonging to a batch " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " sampled from " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 55, + 373, + 296, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 526, + 167, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 167, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 167, + 538 + ], + "type": "text", + "content": "3.2. The Potential Field" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 688 + ], + "type": "text", + "content": "With our potential field representation, embeddings need to be driven towards other nearby embeddings belonging to the same class, while also being driven away from embeddings of other classes. This is reminiscent of the behavior of an isolated system of electric charges, where dissimilar charges are drawn together while similar ones are repelled. Analogously, we define attractive and repulsive potentials created by a single embedding, with the net potential generated by a set of samples given by superposing individual potentials exerted by each embedding. The gradient of the net potential at a point yields the net force (on network update) on a sample embedding located at that point." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "As samples of each class have a different set of similar and dissimilar points, we define a separate potential field" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\Psi_{j}" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": ", for each class " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "j \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": ". The class potential field " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\Psi_{j}" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": " models the force applied on embeddings belonging to class " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": " due to all embeddings. It is formed from a superposition of potentials belonging to sample embeddings from all classes, with embeddings from the same class exerting an attractive potential field (to pull similar points closer) while those from other classes exert a repulsive potential field (to push apart dissimilar points). We assume that each sample embedding " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": " has a unit charge, so the net force on it is the gradient of the potential field " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\Psi_{y_i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": " of its class. We design the attractive and repulsive potential functions by taking motivation from both, electrostatic fields and metric learning literature, which we discuss in detail in the following subsections. Figure 1 provides an overview of our complete pipeline of generation and use of the potential field for metric learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 270, + 507, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 507, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 507, + 282 + ], + "type": "text", + "content": "3.2.1. Potential field exerted by a single point" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 289, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 361 + ], + "type": "text", + "content": "By using forces on the embeddings, we wish the training process to move the embeddings to a state of equilibrium at the end of training in which all embeddings from a class are nearby. To have a potential function exhibiting this behavior, we ensure that individual potential field has the following two properties:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 430 + ], + "type": "text", + "content": "Property 1 - Extremum Property: Potential generated by an individual embedding achieves an extremum at the embedding (minimum for attraction field, maximum for repulsion). This ensures that similar points are pulled together and dissimilar ones are pushed apart on minimizing potential." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 432, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 598 + ], + "type": "text", + "content": "Property 2 - Decay Property: The magnitude of the gradient of the potential (force) due to an embedding decays with distance. Such a decay ensures two desired characteristics. (1) An embedding is not pulled towards a distant sample embedding belonging to the same class, which is likely to be too different to be considered as a variant of it; instead, the decayed influence helps treat the distant sample as significantly different variety of the class, associated with other nearby embeddings of the class. (2) Like for a single charge in the Extremum property, a superposition of potentials of different embeddings still yields local minima at or near sample embeddings. As shown in Sec 3.4, this helps proxies come close to such minima and thus help better model the data distribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 599, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 554, + 647 + ], + "type": "text", + "content": "Given these properties motivated by electrostatic fields, we select the attraction and repulsion potentials " + }, + { + "bbox": [ + 313, + 599, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\psi_{att},\\psi_{rep}" + }, + { + "bbox": [ + 313, + 599, + 554, + 647 + ], + "type": "text", + "content": " generated by an embedding " + }, + { + "bbox": [ + 313, + 599, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 313, + 599, + 554, + 647 + ], + "type": "text", + "content": " , with their strength at location " + }, + { + "bbox": [ + 313, + 599, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 599, + 554, + 647 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 328, + 661, + 555, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 661, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 328, + 661, + 555, + 712 + ], + "type": "interline_equation", + "content": "\\psi_ {a t t} (\\mathbf {r}, \\mathbf {z} _ {i}): = \\left\\{ \\begin{array}{l l} - \\frac {1}{\\delta^ {\\alpha}} & \\text {i f} \\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} < \\delta \\\\ - \\frac {1}{\\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} ^ {\\alpha}} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {1}", + "image_path": "2a0a5f35f02feea4e9898ba7cbc86ec3c876aaf000449e420ecfd1b52119f80a.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25551" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 91, + 271, + 269 + ], + "blocks": [ + { + "bbox": [ + 111, + 77, + 236, + 89 + ], + "lines": [ + { + "bbox": [ + 111, + 77, + 236, + 89 + ], + "spans": [ + { + "bbox": [ + 111, + 77, + 236, + 89 + ], + "type": "text", + "content": "(a) Class Potential field " + }, + { + "bbox": [ + 111, + 77, + 236, + 89 + ], + "type": "inline_equation", + "content": "\\Psi_{1}" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 91, + 271, + 269 + ], + "lines": [ + { + "bbox": [ + 91, + 91, + 271, + 269 + ], + "spans": [ + { + "bbox": [ + 91, + 91, + 271, + 269 + ], + "type": "image", + "image_path": "2e7644140292bd2441b2028d58a4b5c4b51a0f0ea687f50f5cc7a25ec850fc17.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 280, + 90, + 524, + 268 + ], + "blocks": [ + { + "bbox": [ + 328, + 77, + 455, + 89 + ], + "lines": [ + { + "bbox": [ + 328, + 77, + 455, + 89 + ], + "spans": [ + { + "bbox": [ + 328, + 77, + 455, + 89 + ], + "type": "text", + "content": "(b) Class Potential field " + }, + { + "bbox": [ + 328, + 77, + 455, + 89 + ], + "type": "inline_equation", + "content": "\\Psi_{2}" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 280, + 90, + 524, + 268 + ], + "lines": [ + { + "bbox": [ + 280, + 90, + 524, + 268 + ], + "spans": [ + { + "bbox": [ + 280, + 90, + 524, + 268 + ], + "type": "image", + "image_path": "2268a3e65bdeb8101e6be370584568f4ed0ab166f27ed7ba22e3b0ce06f93d4b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "text", + "content": "Figure 2. An example of the class potential fields " + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\Psi_{1}" + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\Psi_{2}" + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "text", + "content": " (Sec. 3.2.3), created by superposing the fields of individual embeddings (Sec 3.2.1) belonging to Classes 1 and 2. Arrows denote the gradient of the respective potentials representing the net force on them. (a) " + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\Psi_{1}" + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "text", + "content": " draws samples/proxies of class 1 towards nearby samples/ proxies of class 1 while keeping them at least " + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "text", + "content": " distance away from embeddings of class 2. Proxies at starred locations are drawn towards the nearest Class 1 embeddings which are potential minima, helping them better model the data distribution (Sec 3.4). (b) " + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\Psi_{2}" + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "text", + "content": " draws embeddings of class 2 towards other nearby embeddings of class 2 instead of distant class 2 embeddings, which might be a significantly different variant of the class (or potentially a mislabeled data point), helping better feature learning (or improving noise robustness if mislabeled). " + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\Psi_{2}" + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "text", + "content": " also keeps class 2 embeddings " + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 284, + 555, + 373 + ], + "type": "text", + "content": " distance them away from embeddings of class 1." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 376, + 295, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 376, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 82, + 376, + 295, + 427 + ], + "type": "interline_equation", + "content": "\\psi_ {r e p} (\\mathbf {r}, \\mathbf {z} _ {i}): = \\left\\{ \\begin{array}{l l} \\frac {1}{\\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} ^ {\\alpha}} & \\| \\mathbf {r} - \\mathbf {z} _ {i} \\| _ {2} < \\delta \\\\ \\frac {1}{\\delta^ {\\alpha}} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {2}", + "image_path": "4d5830129814b645dd3d74805f2f504bdf7003d23e4893b908dc2f127f5b1fe6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 437, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 521 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 54, + 437, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 437, + 295, + 521 + ], + "type": "text", + "content": " is a hyperparameter that determines the rate of decay of the field (we evaluate it's effect in Sec. 4.5 instead of fixing " + }, + { + "bbox": [ + 54, + 437, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 54, + 437, + 295, + 521 + ], + "type": "text", + "content": " as done in electrostatic fields). Note that the attraction potential is negative and increases with distance, implying that the force between the embedding and the point is attractive. The repulsion potential is positive and decreasing with distance, indicating repulsion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "content": "A constant attraction potential inside the " + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "content": " sphere (i.e., force " + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "content": ") ensures that sample embeddings within " + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "content": " distance of one another do not continue to attract each other, thus preventing the network from embeddings from collapsing to a single point which harms generalizability of the features learned (analogous to the effect margins used in tuplet-based losses have)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Limiting the repulsion potential to a sphere of radius " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " ensures that embeddings not belonging to the same class are driven to a distance " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "> \\delta" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". Since the aforementioned constant attraction potential ensures that " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " is the maximum distance between samples " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " from the same class, this leads to connected components of " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " spheres formed by nearby embeddings of the same class. Extending repulsion to distances " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "> \\delta" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " does not provide any useful additional supervisory information while forcing further unmotivated inter-class sep" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 379, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 403 + ], + "type": "text", + "content": "aration; as we empirically show in Sec. 4.5, such excessive separation does indeed harm performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 408, + 517, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 517, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 517, + 421 + ], + "type": "text", + "content": "3.2.2. Proxies to Represent Sample Populations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 423, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 579 + ], + "type": "text", + "content": "To represent the image embeddings not sampled in the current batch " + }, + { + "bbox": [ + 313, + 423, + 555, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 423, + 555, + 579 + ], + "type": "text", + "content": ", we propose to use a set of " + }, + { + "bbox": [ + 313, + 423, + 555, + 579 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 423, + 555, + 579 + ], + "type": "text", + "content": " proxies per class for each of the N classes, " + }, + { + "bbox": [ + 313, + 423, + 555, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{j,k}, k \\in \\{1, \\dots, M\\}, j \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 313, + 423, + 555, + 579 + ], + "type": "text", + "content": ", and augment the potential field generated by the embeddings in the batch by those generated by the proxies. The proxies serve as stand-ins for the larger numbers of out-of-the-batch embeddings for calculating the attraction and repulsion fields, while also being affected by the field themselves like all embeddings. The proxies, being learnable parameters with high learning rates, tend to gravitate toward the nearest class modes of the potential field (and hence the class data distribution). The number of proxies M chosen should be large enough to capture the class-modes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 584, + 430, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 430, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 430, + 596 + ], + "type": "text", + "content": "3.2.3. Class Potential Field" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 600, + 554, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 554, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 554, + 636 + ], + "type": "text", + "content": "Given a batch " + }, + { + "bbox": [ + 313, + 600, + 554, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 600, + 554, + 636 + ], + "type": "text", + "content": " of samples and all proxies, we calculate the class potential field " + }, + { + "bbox": [ + 313, + 600, + 554, + 636 + ], + "type": "inline_equation", + "content": "\\Psi_{j}(\\mathbf{r})" + }, + { + "bbox": [ + 313, + 600, + 554, + 636 + ], + "type": "text", + "content": " at a location " + }, + { + "bbox": [ + 313, + 600, + 554, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 600, + 554, + 636 + ], + "type": "text", + "content": " due to a class " + }, + { + "bbox": [ + 313, + 600, + 554, + 636 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 600, + 554, + 636 + ], + "type": "text", + "content": " as a superposition of their attraction and repulsion potentials:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 354, + 645, + 555, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 645, + 555, + 659 + ], + "spans": [ + { + "bbox": [ + 354, + 645, + 555, + 659 + ], + "type": "interline_equation", + "content": "\\Psi_ {j} (\\mathbf {r}) = \\Psi_ {j, a t t} (\\mathbf {r}) + \\Psi_ {j, r e p} (\\mathbf {r}), \\text {w h e r e} \\tag {3}", + "image_path": "221fbd86997e4dcc3fa6ed98d8f906345486242241c32ba8fa578f4076ae5cd4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 677, + 555, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 677, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 677, + 555, + 712 + ], + "type": "interline_equation", + "content": "\\Psi_ {j, a t t} (\\mathbf {r}) = \\sum_ {i = 1} ^ {\\| \\mathcal {B} \\|, y _ {i} = j} \\psi_ {a t t} (\\mathbf {r}, \\mathbf {z} _ {i}) + \\sum_ {k = 1} ^ {M} \\psi_ {a t t} (\\mathbf {r}, \\mathbf {p} _ {j, k}) \\tag {4}", + "image_path": "323db1120e4b844ce2c41cc41f24870175717f2c72da8c920ceff58577585ba6.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25552" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 70, + 302, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 70, + 302, + 119 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 302, + 119 + ], + "type": "interline_equation", + "content": "\\Psi_ {j, r e p} (\\mathbf {r}) = \\sum_ {i = 1} ^ {\\| \\mathcal {B} \\|, y _ {i} \\neq j} \\psi_ {r e p} (\\mathbf {r}, \\mathbf {z} _ {i}) + \\sum_ {\\gamma = 1, \\gamma \\neq j} ^ {N} \\sum_ {k = 1} ^ {M} \\psi_ {r e p} (\\mathbf {r}, \\mathbf {p} _ {\\gamma , k}) \\tag {5}", + "image_path": "47385394944ae331ee93aaf8db1fa209d102bb78426f17d663f907487fb3e6dc.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 125, + 297, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 125, + 297, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 125, + 297, + 150 + ], + "type": "text", + "content": "Figure 2 visualizes the class potential fields " + }, + { + "bbox": [ + 55, + 125, + 297, + 150 + ], + "type": "inline_equation", + "content": "\\Psi_{1}" + }, + { + "bbox": [ + 55, + 125, + 297, + 150 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 125, + 297, + 150 + ], + "type": "inline_equation", + "content": "\\Psi_{2}" + }, + { + "bbox": [ + 55, + 125, + 297, + 150 + ], + "type": "text", + "content": " for a given set of embeddings belonging to two classes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 121, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 121, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 121, + 170 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 174, + 296, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 296, + 223 + ], + "type": "text", + "content": "Using gradient descent, the network is trained to minimize the total potential energy " + }, + { + "bbox": [ + 55, + 174, + 296, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 55, + 174, + 296, + 223 + ], + "type": "text", + "content": " of all proxies and embeddings in the batch " + }, + { + "bbox": [ + 55, + 174, + 296, + 223 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i\\in \\mathcal{B}" + }, + { + "bbox": [ + 55, + 174, + 296, + 223 + ], + "type": "text", + "content": " by applying a force on embeddings and proxies in the direction of " + }, + { + "bbox": [ + 55, + 174, + 296, + 223 + ], + "type": "inline_equation", + "content": "-\\nabla \\Psi_{y_i}(\\mathbf{z}_i)" + }, + { + "bbox": [ + 55, + 174, + 296, + 223 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 174, + 296, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 55, + 174, + 296, + 223 + ], + "type": "text", + "content": " given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 98, + 230, + 296, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 230, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 98, + 230, + 296, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {U} = \\sum_ {i = 1} ^ {\\| \\mathcal {B} \\|} \\Psi_ {y _ {i}} (\\mathbf {z} _ {i}) + \\sum_ {j = 1} ^ {N} \\sum_ {k = 1} ^ {M} \\Psi_ {j} (\\mathbf {p} _ {j, k}) \\tag {6}", + "image_path": "e63521e0a759a7e7966effdcf05e63c2aea160e6eceb3639686eb56572cd6271.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 272, + 295, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 295, + 285 + ], + "type": "text", + "content": "3.4. Why decay interaction strength with distance?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "text", + "content": "PFML uses a potential field where the gradient (interaction strength) decays with distance, which is unlike previous approaches. This leads to the two benefits described in Sec 3.2.1, which we discuss in more detail here." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 343, + 296, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 367 + ], + "type": "text", + "content": "3.4.1. Robustness when learning with datasets having large intra-class variations and noisy labels" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 370, + 296, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 370, + 296, + 573 + ], + "spans": [ + { + "bbox": [ + 54, + 370, + 296, + 573 + ], + "type": "text", + "content": "The decay property helps PFML model all interactions even when using datasets with large intra-class variations. As seen in Figure 2, PFML draws embeddings towards other nearby embeddings from the same class rather than distant class embeddings that might be too different to be considered a variation of it. It also helps PFML become significantly more robust to the effect of incorrectly labeled examples as compared to existing methods. Recent work [26, 30] shows that such examples are common in metric learning benchmarks. The decay de-emphasizes the contribution of such unlabeled examples, that are typically located away from correctly labeled samples. We empirically confirm this in Sec. 4.3, demonstrating that methods not adhering to the decay property (i.e., all previous tuplet-based and proxy-based methods) experience significantly greater performance degradation due to label noise compared to PFML." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 579, + 295, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 295, + 590 + ], + "type": "text", + "content": "3.4.2. Better alignment of proxies with data distribution" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 297, + 714 + ], + "type": "text", + "content": "The decaying interaction strength helps PFML better align the distribution of proxies with the underlying sample data distribution (of the complete dataset) that they represent. To analyze this, we compare PFML with the closely related contrastive loss [8] that also models pairwise interactions like PFML (defined in Eq. 1 & 2), and hence can be analyzed similarly by using a potential field. Specifically, we introduce Contrastive Potential based Metric Learning (CPML) by simply extending the contrastive loss by using multiple proxies per class, and subsequently analyzing the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "resulting loss function through a potential field framework. In CPML, strength of interaction is directly proportional to the square of the distance from it " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "(\\|\\mathbf{r} - \\mathbf{r}_0\\|_2^2)" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " like in the original contrastive loss. For fair comparison, we keep all other details (use of " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ", proxies) as the same, with its associated potential fields denoted by " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\Psi_{att}^*" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\Psi_{rep}^*" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " (formally defined in Supplement Sec 1). We analyze this theoretically using the following proposition and its corollary:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": "Proposition 1: Let " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "Z = \\{z_1 \\ldots z_n\\}" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": " be a set of sample embeddings belonging to a class, then there exists a " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "0 < \\delta < \\frac{\\min_{i,j} \\|z_i - z_j\\|_2}{2\\left(1 + \\frac{1}{n}\\right)}" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": " and points " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "z_{min,i}" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": " distance from each embedding " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "i \\in \\{1 \\ldots n\\}" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": " such that the attractive potential field " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "\\Psi_{att}" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": " defined using " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": " has a minimum at each " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "z_{min,i}" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": ". The field " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "\\Psi_{att}^*" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": " defined by them does not achieve a minimum at points within " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": " distance from all " + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "inline_equation", + "content": "z_i" + }, + { + "bbox": [ + 313, + 171, + 556, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 284, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 344 + ], + "type": "text", + "content": "Proposition 1 is a consequence of the potential field of an embedding dominating in its vicinity over potential fields of other distant embeddings. Its proof relies on the Decay property, which is not satisfied by " + }, + { + "bbox": [ + 313, + 284, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\Psi^{*}" + }, + { + "bbox": [ + 313, + 284, + 555, + 344 + ], + "type": "text", + "content": " and can be found in Supp. Sec 1. A consequence of this is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "text", + "content": "Corollary 1: Let " + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "inline_equation", + "content": "Z = \\{z_{1}\\dots z_{n}\\}" + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "text", + "content": " be a set of sample embeddings from a class exerting an attraction field on a set of proxies " + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "inline_equation", + "content": "P = \\{\\pmb {p}_1\\dots \\pmb {p}_m\\}" + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "text", + "content": ". Consider the equilibrium distribution " + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "inline_equation", + "content": "P_{eq}" + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "text", + "content": " of proxies minimizing the potential energy. If the potential field is defined by " + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "inline_equation", + "content": "\\Psi_{att}" + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "text", + "content": ", then the Wasserstein distance " + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "inline_equation", + "content": "P_{eq}" + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "text", + "content": " and the subset of data they represent is lower than when the potential field is defined by " + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "inline_equation", + "content": "\\Psi_{att}^*" + }, + { + "bbox": [ + 313, + 346, + 556, + 433 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 434, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 459 + ], + "type": "text", + "content": "The proof for Corollary 1 relies on Proposition 1 and can be found in the Supplement (Sec. 2)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "Discussion: As proved in Corollary 1, CPML leads to proxies having a larger distance from the data distribution they represent (measured by " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ") than our proposed PFML, making the use of proxies in such a framework less effective. This leads to the proxies forming poorer stand-ins for the sample embeddings they represent, reducing the quality of features learned (verified in Sec. 4). This is caused by CPML not following the decay property (which is its only difference in design wrt PFML). Other proxy-based losses like the Proxy NCA [34], Soft Triplet [36], and Proxy Anchor [21] which perform similarly as CPML (see Tab. 1), also share with it the property of force increasing with distance (violating the decay property). Due to this, we expect them to similarly have a poorer alignment of proxies with underlying data distribution as compared to PFML (Applying an analysis similar to the one in Proposition 1 to them is considerably more challenging due to the absence of a potential field structure.). We verify this empirically in Sec. 4.4 by measuring the average distance between proxies and the closest data points (" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " distance) for these methods." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25553" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "content": "3.5. Advantages in Learning Fine-grained Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 93, + 295, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 93, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 93, + 295, + 213 + ], + "type": "text", + "content": "In contrast with all previous proxy-based methods, PFML (1) Models sample-sample interactions directly using the potential field, gaining fine-grained supervision, and (2) Makes use of an explicit margin-like " + }, + { + "bbox": [ + 55, + 93, + 295, + 213 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 93, + 295, + 213 + ], + "type": "text", + "content": " for proxies, which has been shown to help zero-shot generalization and fine-grained feature learning in tuplet-based methods[53]. Turning off the interaction within the " + }, + { + "bbox": [ + 55, + 93, + 295, + 213 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 93, + 295, + 213 + ], + "type": "text", + "content": " radius also prevents a bias towards learning isotropic distributions as is present in other proxy-based methods [39], helping us obtain better intra-class resolution." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 236, + 198, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 198, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 198, + 250 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 261, + 105, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 105, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 105, + 273 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 281, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 413 + ], + "type": "text", + "content": "Datasets: We empirically compare PFML with current state-of-the-art baselines on three public benchmark datasets (1) The Cars-196 dataset [24] which includes 16,185 images from 196 car model categories, (2) CUB200-2011 dataset [49] consisting of 11,788 images from 200 bird species, and (3) Stanford Online Products (SOP) dataset [43] containing 120,053 images of 22,634 different products sold online. The training and testing splits for the three datasets follow standard zero-shot retrieval-based evaluation protocol [21, 51, 61] of using the first half of the classes for training and the second half for testing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 525 + ], + "type": "text", + "content": "**Backbone:** To enable standardized comparison with a wide variety of methods, we evaluate our method using the Inception with batch normalization (IBN) [46], ResNet-50 [16], ViT [11] and DINO [5] backbones commonly used by DML methods. ImageNet pre-training is used for initialization, and only the last fully connected layer is changed to fit the dimension of the embedding needed. We follow past work [21, 36, 48] and " + }, + { + "bbox": [ + 55, + 418, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 55, + 418, + 295, + 525 + ], + "type": "text", + "content": " normalize the final output. An input image size of " + }, + { + "bbox": [ + 55, + 418, + 295, + 525 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 418, + 295, + 525 + ], + "type": "text", + "content": " is used for all experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "content": "Training parameters: We train the backbone for 200 epochs on each dataset. We use the Adam optimizer with the learning rate chosen as " + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "inline_equation", + "content": "5e^{-4}" + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "content": ". We scale up the learning rate of our proxies by 100 times following [21, 34, 36] for better convergence. We choose values of " + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "inline_equation", + "content": "[0.1, 0.3]" + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "inline_equation", + "content": "\\{0, 6\\}" + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "content": " via cross-validation. We use " + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "inline_equation", + "content": "M = 15" + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "content": " for the CUB-200 and Cars-196 datasets, while " + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "inline_equation", + "content": "M = 2" + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "content": " is used for the SOP dataset. All experiments are performed on a single NVIDIA A4000 GPU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Evaluation Settings: We follow [21, 22, 29, 51] in using " + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": " sized center crops from " + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": " images for evaluation. We measure image retrieval performance using Recall@K, which computes the percentage of samples which have a valid similar neighbor (belonging to same class) among its K nearest neighbors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 474, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 474, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 474, + 84 + ], + "type": "text", + "content": "4.2. Image Retrieval Performance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": "As seen in Table 1, our method outperforms all other methods on Image Retrieval on the CUB-200-2011, Cars-196 and SOP datasets as measured by the Recall@K metric. It consistently outperforms other methods irrespective of the backbone used, significantly outperforming the best-performing proxy-based methods (which similar to us, only change the loss function) like ProxyAnchor[21], ProxyGML[61] and SoftTriplet [36] in terms of Recall@1 (R@1), by more than " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": " on Cars-196, " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": " on CUB-200-2011 and " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": " on the SOP dataset. It also outperforms the significantly more complex, current state-of-the-art HIST[28] by " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": " in terms of R@1 when using the Resnet-50 backbone on the Cars-196, CUB-200-2011 and SOP datasets respectively. It also significantly outperforms the best tuplet-based loss: the Multi-Similarity loss by " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": " in terms of R@1 when using the BN Inception backbone on the Cars-196, CUB-200 and SOP datasets respectively. Additionally, our method outperforms CPML (Sec. 3.4.2, a simple combination of the classical contrastive loss with proxies) by " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 376 + ], + "type": "text", + "content": " on CUB200, Cars and SOP datasets. Note that our performance gains are larger than current SOTA compared to improvements reported by previous efforts on the same benchmarks, despite our intuitive and simpler method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "text", + "content": "We also evaluate our method 1) Using the protocol and metrics proposed in [35] and (2) When used to learn smaller 128-dim embeddings for more efficient retrieval. Our method still outperforms all SOTA methods in both cases. Details of these experiments are in Supp. Sec. 3. We also show qualitative retrieval results of our method in Supp. Sec 6." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 468, + 537, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 537, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 537, + 480 + ], + "type": "text", + "content": "4.3. Image Retrieval in Presence of Label Noise" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 485, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 553, + 651 + ], + "type": "text", + "content": "We evaluate the robustness of our method vs. state-of-the-art baselines in the realistic scenario of label noise (e.g., due to mislabeling, which is very common, particularly in fine-grained datasets such as those used for DML). Specifically, we randomly corrupt labels of " + }, + { + "bbox": [ + 313, + 485, + 553, + 651 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 485, + 553, + 651 + ], + "type": "text", + "content": " of samples in the training set. As seen in the results in Table 2, this leads to a significant decline in the performance of all methods. But our method is the least affected by the noise. Other methods like the triplet [15] or ProxyAnchor[21] suffer much greater declines because they are affected more by distant points represented by the mislabeled examples. Our method significantly outperforms Proxy Anchor [21] (the second-best method) by " + }, + { + "bbox": [ + 313, + 485, + 553, + 651 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 485, + 553, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 485, + 553, + 651 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 313, + 485, + 553, + 651 + ], + "type": "text", + "content": " in terms of R@1 on the CUB-200-2011 and Cars-196 datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 659, + 553, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 553, + 685 + ], + "type": "text", + "content": "4.4. Distance Between Proxies and Sample Embeddings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "We measure the average (over a complete epoch) Wasserstein distance " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " between proxies and the subset of training" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25554" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 81, + 69, + 530, + 487 + ], + "blocks": [ + { + "bbox": [ + 81, + 69, + 530, + 487 + ], + "lines": [ + { + "bbox": [ + 81, + 69, + 530, + 487 + ], + "spans": [ + { + "bbox": [ + 81, + 69, + 530, + 487 + ], + "type": "table", + "html": "
Benchmarks →CUB-200-2011Cars-196SOP
Methods ↓ (Chronological)R@1R@2R@4R@1R@2R@4R@1R@10R@100
ResNet50 (512 dim)
ESPHN [55]64.975.383.582.789.393.078.390.796.3
N.Softmax [58]61.373.983.584.290.494.478.290.696.2
DiVA [33]69.279.3-87.692.9-79.691.2-
Proxy NCA++ [48]64.7---85.1--79.6--
Proxy Anchor [21]69.780.087.087.792.995.8---
DCML-MDW [59]68.477.986.185.291.896.079.890.895.8
MS+DAS [29]69.279.287.187.893.196.080.691.896.7
HIST [28]71.481.188.189.693.996.481.492.096.7
HIER[22]70.179.486.988.293.095.680.291.596.6
HSE-PA [56]70.680.187.189.693.896.080.091.496.3
CPML (Sec. 3.4)68.378.786.285.291.595.279.490.796.1
Potential Field (Ours)73.4 ± 0.382.4± 0.188.8± 0.192.7± 0.395.5 ± 0.197.6± 0.182.9 ± 0.292.5 ± 0.296.8 ± 0.1
BN Inception (512 dim)
HTL [13]57.168.878.781.488.092.774.888.394.8
MultiSimilarity [51]65.777.086.384.190.494.078.290.596.0
SoftTriple [36]65.476.484.584.590.794.578.390.395.9
CircleLoss [44]66.777.486.283.489.894.178.390.596.1
DiVA [33]66.877.7-84.190.7-78.190.6-
ProxyGML [61]66.677.686.485.591.895.378.090.696.2
Proxy Anchor [21]68.479.286.886.191.795.079.190.896.2
DRML-PA [60]68.778.686.386.992.195.271.585.293.0
MS+DAS [29]67.178.1186.485.791.695.378.290.396.0
HIST [28]69.780.087.387.492.595.479.691.096.2
DFML-PA [50]69.3--88.4-----
HSE-M [56]67.678.085.882.088.993.3---
PA+niV [23]69.580.0-86.492.0-79.290.4-
Potential Field (Ours)71.5± 0.381.2± 0.288.3± 0.290.1± 0.293.9± 0.196.3± 0.180.6 ± 0.391.8 ± 0.196.4 ± 0.1
DINO (384 dim)
DINO [5]70.881.188.842.953.964.263.478.188.3
Hyp [12]80.987.692.489.294.196.785.194.497.8
HIER [22]81.188.293.391.395.297.185.794.697.8
Potential Field (Ours)83.1 ± 0.389.3 ± 0.294.2 ± 0.194.7 ± 0.196.5 ± 0.197.8 ± 0.186.5 ± 0.395.1 ± 0.398.0 ± 0.2
ViT (384 dim)
ViT-S [11]83.190.494.447.860.272.262.177.789.0
Hyp [12]85.691.494.886.592.195.385.994.998.1
HIER [22]85.791.394.488.393.296.186.195.098.0
Potential Field (Ours)87.8 ± 0.292.6 ± 0.295.7 ± 0.191.5 ± 0.395.2 ± 0.297.4 ± 0.188.2 ± 0.195.7 ± 0.198.6 ± 0.1
", + "image_path": "e69f2d3045918d87ec0a983973cb7a967cf110879d03559b827bfc587f4d5ffb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 522, + 364, + 624 + ], + "blocks": [ + { + "bbox": [ + 55, + 488, + 555, + 512 + ], + "lines": [ + { + "bbox": [ + 55, + 488, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 555, + 512 + ], + "type": "text", + "content": "Table 1. Comparison of the Recall@ " + }, + { + "bbox": [ + 55, + 488, + 555, + 512 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 488, + 555, + 512 + ], + "type": "text", + "content": " (%) achieved by our method on the CUB-200-2011, Cars-196 and SOP datasets with state-of-the-art baselines under standard settings. The table reports the average performance and standard deviations of our method over 5 runs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 522, + 364, + 624 + ], + "lines": [ + { + "bbox": [ + 53, + 522, + 364, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 522, + 364, + 624 + ], + "type": "table", + "html": "
CUB-200-2011Cars-196
MethodsR@1R@2R@1R@2
Triplet[52]55.168.767.577.9
MS [51]58.971.870.479.8
PNCA[34]60.174.774.382.4
PA [21]60.775.176.983.1
HIST[27]59.774.672.981.8
Ours66.7 ± 0.676.9± 0.384.5± 0.588.6± 0.3
", + "image_path": "594d26ff7f09f0afe06ece9a60487d720994c6fca599f47a7950dab3fe07995c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 391, + 534, + 541, + 612 + ], + "blocks": [ + { + "bbox": [ + 52, + 632, + 364, + 666 + ], + "lines": [ + { + "bbox": [ + 52, + 632, + 364, + 666 + ], + "spans": [ + { + "bbox": [ + 52, + 632, + 364, + 666 + ], + "type": "text", + "content": "Table 2. A comparison of Recall@ " + }, + { + "bbox": [ + 52, + 632, + 364, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 52, + 632, + 364, + 666 + ], + "type": "text", + "content": " (%) achieved by our method with SOTA baselines on CUB-200-2011 and Cars-196 datasets with label noise when using a Resnet-50 backbone (512-dim). Performance and std. deviations reported are over 5 runs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 391, + 534, + 541, + 612 + ], + "lines": [ + { + "bbox": [ + 391, + 534, + 541, + 612 + ], + "spans": [ + { + "bbox": [ + 391, + 534, + 541, + 612 + ], + "type": "table", + "html": "
MethodAverage W2
S.Triplet [36]0.24 ± 0.04
PNCA [21]0.36 ± 0.02
PA [21]0.38 ± 0.03
CPML (Sec. 3.4)0.26 ± 0.03
Ours0.16 ± 0.03
", + "image_path": "49e59f42447e35ed1546c5349f91c1f7355e3a2ab6bae4bded77d6a48f669e6b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 374, + 620, + 557, + 665 + ], + "lines": [ + { + "bbox": [ + 374, + 620, + 557, + 665 + ], + "spans": [ + { + "bbox": [ + 374, + 620, + 557, + 665 + ], + "type": "text", + "content": "Table 3. Average Wasserstein distance between proxies and the subset of data closest to them for different methods on CUB-200-2011 dataset. Mean and std. deviations are over 5 runs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 677, + 295, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 691 + ], + "type": "text", + "content": "data represented by/closest to them " + }, + { + "bbox": [ + 55, + 677, + 295, + 691 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 55, + 677, + 295, + 691 + ], + "type": "text", + "content": " for different proxy-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 677, + 555, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 702 + ], + "type": "text", + "content": "based methods. We train networks on the CUB-200-2011 using the settings in Section 4.1 and use the same number" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 695, + 295, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 695, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 695, + 295, + 715 + ], + "type": "text", + "content": "2For methods using a single proxy per class, this equals the distance between a proxy and the closest embedding belonging to its class" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25555" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 182, + 155 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 182, + 155 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 182, + 155 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 182, + 155 + ], + "type": "image", + "image_path": "83c48e49831ecbab92c541f90b6f3a53a6191a03366d83260c4ad7d758c87a36.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 164, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 55, + 164, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 553, + 186 + ], + "type": "text", + "content": "Figure 3. Variation in Recall@1 with " + }, + { + "bbox": [ + 55, + 164, + 553, + 186 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 164, + 553, + 186 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 164, + 553, + 186 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 164, + 553, + 186 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 164, + 553, + 186 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 164, + 553, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 164, + 553, + 186 + ], + "type": "inline_equation", + "content": "\\delta_{rep} - \\delta_{att}" + }, + { + "bbox": [ + 55, + 164, + 553, + 186 + ], + "type": "text", + "content": " on the Cars-196 and CUB-200-2011 datasets. Error bars represent std. deviations over 5 runs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 185, + 70, + 304, + 155 + ], + "blocks": [ + { + "bbox": [ + 185, + 70, + 304, + 155 + ], + "lines": [ + { + "bbox": [ + 185, + 70, + 304, + 155 + ], + "spans": [ + { + "bbox": [ + 185, + 70, + 304, + 155 + ], + "type": "image", + "image_path": "f78862198ec348277fa0eb0ca5d637f108e1e078ac781392dfd98b544255fa50.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 423, + 156 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 423, + 156 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 423, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 423, + 156 + ], + "type": "image", + "image_path": "2189e2657748765a4f2dffad8fcc2bc9d0323c667997ac10b20e4315d8bfb76b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 426, + 70, + 545, + 156 + ], + "blocks": [ + { + "bbox": [ + 426, + 70, + 545, + 156 + ], + "lines": [ + { + "bbox": [ + 426, + 70, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 426, + 70, + 545, + 156 + ], + "type": "image", + "image_path": "92e2a474b7da5dcf709eb8a568784c588174eced910efb75b936f870f69f40fa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": "of proxies for all methods utilizing multiple proxies. As seen in Table 3, our method shows " + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "inline_equation", + "content": "33.3\\%" + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": " lower " + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": " distance between the proxies and the subset of data than the second best SoftTriplet [36] loss, empirically validating our discussion in Sec. 3.4.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 259, + 154, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 154, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 154, + 272 + ], + "type": "text", + "content": "4.5. Ablation studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 278, + 296, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 315 + ], + "type": "text", + "content": "In all experiments, we use the parameters mentioned in Section 4.1 with a ResNet-50 backbone on both the Cars-196 and CUB-200-2011 datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 316, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 519 + ], + "type": "text", + "content": "Effect of proxies and their number: Proxies augment the potential field, and the number of proxies per class " + }, + { + "bbox": [ + 55, + 316, + 296, + 519 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 316, + 296, + 519 + ], + "type": "text", + "content": " decides the granularity at which the proxies model the distribution of sample embeddings. We study their effect on performance by varying " + }, + { + "bbox": [ + 55, + 316, + 296, + 519 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 316, + 296, + 519 + ], + "type": "text", + "content": " between [0, 30]. From the results plotted in Figure 3(a) we observe that (1) Using proxies to augment the field boosts performance (vs using no-proxies) and (2) Network performance remains stable for a wide range of M, decreasing slightly for small values of M (still outperforming most other proxy-based methods). This behavior is expected as using potential field created using a very small M would not be able to effectively model underlying sample distribution, leading to worse supervision. We also observe that using a larger number of proxies per class does not significantly affect training times, as bulk of computation is dominated by the network. For further details, please refer to Supp. Sec. 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "type": "text", + "content": "Effect of varying " + }, + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "type": "text", + "content": ": We conduct an empirical study to determine the effect of varying the radius parameter " + }, + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "type": "text", + "content": " on performance by varying it over [0.1, 0.35]. As seen in Figure 3(b), we observe that performance remains relatively stable for a wide range of " + }, + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 521, + 295, + 592 + ], + "type": "text", + "content": " values, demonstrating our method's robustness to it." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "Effect of potential decay " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": ": The decay parameter " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " determines the rate at which the potential field of an embedding decays with distance. We study its effect by varying it over [0, 6]. As seen in Figure 3(c), performance is worse when the potential field does not decay (" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "). A field that decays strongly " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha \\in [3, 6]" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " seems to perform better than one with a much milder decay (" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha \\in [1, 2]" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "). The likely reason for this trend is that when the potential field has no decay /decays mildly, the force on an embedding would still be dominated by a large number of distant sample embeddings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 189, + 440, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 440, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 440, + 201 + ], + "type": "text", + "content": "rather than those in its vicinity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": "Restricting repulsion to a distance of " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": ": Unlike most proxy-based methods, our design of potential field " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\Psi_{rep}" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": " limits the range of repulsion to within " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": " distance from embeddings or proxies. We examine extending this range of repulsion beyond the " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": " radius by having " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta_{rep} > \\delta_{att}" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta_{att}" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta_{rep}" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": " represent the values of " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": " used for attraction (Eq.1) and repulsion (Eq.2) fields. Fixing " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta_{att} = 0.2" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": ", we increase " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta_{rep}" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": " to study the effect. As shown in Figure 3(d), his reduces performance, validating our design. This is due to individual potentials repelling dissimilar points to a distance of at least " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta_{rep}" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": ", greater than the distance " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta_{att}" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": " within which all points of the same class are attracted to. Using " + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\delta_{rep} > \\delta_{att}" + }, + { + "bbox": [ + 313, + 201, + 555, + 380 + ], + "type": "text", + "content": " enforces excessive separation between dissimilar examples, compressing intra-class features more than needed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 388, + 389, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 389, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 389, + 399 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 403, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 559 + ], + "type": "text", + "content": "PFML relies on proxies to augment potential fields in each batch, hence sharing the main limitation of proxy-based methods: with a large number of classes, the computational cost during training increases due to linear scaling of proxies with class count. Despite this, proxy-based methods have been widely used in applications with a large number of classes like face recognition (e.g. [10] on 500,000 classes), likely because (1) proxies only affect training, not inference complexity and (2) The parameters/compute complexity added by the proxies is relatively low for the performance gains they offer, especially when such compute complexity is compared to the forward and backward passes of the neural network that consume bulk of compute in DML." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 567, + 388, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 388, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 388, + 578 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "PFML introduces a novel paradigm to capture interactions between all samples through a continuous potential field, where strength of sample influence decreases with distance. This reversal of the common model of increasing influence with distance improves robustness to label noise and better aligns proxy distributions with sample embeddings. The overall design of our method is validated by its performance on standard image retrieval benchmarks, where it outperforms state-of-the-art methods both in the standard noise-free setting and in the presence of label noise unavoidable in real-world datasets." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25556" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Narendra Ahuja. A transform for multiscale image segmentation by integrated edge and region detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 18 (12):1211-1235, 1996. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] N. Ahuja and Jen-Hui Chuang. Shape representation using a generalized potential field model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(2):169-176, 1997. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 213 + ], + "type": "text", + "content": "[3] Shubhang Bhatnagar and Narendra Ahuja. Piecewise-linear manifolds for deep metric learning. In Conference on Parsimony and Learning (Proceedings Track), 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 295, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 289 + ], + "type": "text", + "content": "[4] Biagio Brattoli, Joseph Tighe, Fedor Zhdanov, Pietro Perona, and Krzysztof Chalupka. Rethinking zero-shot video classification: End-to-end training for realistic applications. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 4612-4622. Computer Vision Foundation / IEEE, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 346 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 295, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 391 + ], + "type": "text", + "content": "[6] Jun Cen, Peng Yun, Junhao Cai, Michael Yu Wang, and Ming Liu. Deep metric learning for open world semantic segmentation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15313-15322, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 295, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 457 + ], + "type": "text", + "content": "[7] Weihua Chen, Xiaotang Chen, Jianguo Zhang, and Kaiqi Huang. Beyond triplet loss: A deep quadruplet network for person re-identification. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1320-1329. IEEE Computer Society, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 295, + 512 + ], + "type": "text", + "content": "[8] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 539-546. IEEE, 2005. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "text", + "content": "[9] Jen-Hui Chuang and N. Ahuja. An analytically tractable potential field model of free space and its application in obstacle avoidance. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 28(5):729-736, 1998. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 602 + ], + "type": "text", + "content": "[10] J. Deng, J. Guo, N. Xue, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4685-4694, 2019. 1, 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 603, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 678 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[12] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7409-7419, 2022. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "type": "text", + "content": "[13] Weifeng Ge, Weilin Huang, Dengke Dong, and Matthew R. Scott. Deep metric learning with hierarchical triplet loss. In Computer Vision - ECCV 2018, pages 272-288, Cham, 2018. Springer International Publishing. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 154, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 196 + ], + "type": "text", + "content": "[14] Matheus Gutoski, André Eugénio Lazzaretti, and Heitor Silvério Lopes. Deep metric learning for open-set human action recognition in videos. Neural Comput. Appl., 33(4): 1207-1220, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 199, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 199, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 199, + 553, + 243 + ], + "type": "text", + "content": "[15] Ben Harwood, Vijay Kumar BG, Gustavo Carneiro, Ian Reid, and Tom Drummond. Smart mining for deep metric learning. In Proceedings of the IEEE international conference on computer vision, pages 2821-2829, 2017. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 246, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 553, + 289 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 291, + 553, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 324 + ], + "type": "text", + "content": "[17] J. Hu, J. Lu, and Y. Tan. Discriminative deep metric learning for face verification in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "text", + "content": "[18] Y.K. Hwang and N. Ahuja. Path planning using a potential field representation. In Proceedings. 1988 IEEE International Conference on Robotics and Automation, pages 648-649 vol.1, 1988. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 553, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 405 + ], + "type": "text", + "content": "[19] Y.K. Hwang and N. Ahuja. A potential field approach to path planning. IEEE Transactions on Robotics and Automation, 8 (1):23-32, 1992. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 407, + 553, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 553, + 450 + ], + "type": "text", + "content": "[20] O. Khatib. Real-time obstacle avoidance for manipulators and mobile robots. In Proceedings. 1985 IEEE International Conference on Robotics and Automation, pages 500-505, 1985. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 453, + 555, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 555, + 496 + ], + "type": "text", + "content": "[21] Sungyeon Kim, Dongwon Kim, Minsu Cho, and Suha Kwak. Proxy anchor loss for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "text", + "content": "[22] Sungyeon Kim, Boseung Jeong, and Suha Kwak. Hier: Metric learning beyond class labels via hierarchical regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19903-19912, 2023. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 555, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 610 + ], + "type": "text", + "content": "[23] M. Kirchhof, K. Roth, Z. Akata, and E. Kasneci. A non-isotropic probabilistic take on proxy-based deep metric learning. In Computer Vision - ECCV 2022 - 17th European Conference, Proceedings, Part XXVI, pages 435-454. Springer, 2022. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "text", + "content": "[24] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D object representations for fine-grained categorization. In 4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[25] Elad Levi, Tete Xiao, Xiaolong Wang, and Trevor Darrell. Rethinking preventing class-collapsing in metric learning with margin-based losses. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10316-10325, 2021. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25557" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[26] Christopher Liao, Theodoros Tsiligkaridis, and Brian Kulis. Supervised metric learning to rank for retrieval via contextual similarity optimization. In International Conference on Machine Learning, pages 20906-20938. PMLR, 2023. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[27] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning code. https://github.com/ljin0429/HIST, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[28] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 212-222, 2022. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[29] Lizhao Liu, Shangxin Huang, Zhuangwei Zhuang, Ran Yang, Mingkui Tan, and Yaowei Wang. Das: Densely-anchored sampling for deep metric learning. In European Conference on Computer Vision, pages 399–417. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[30] Ruofan Liu, Yun Lin, XIANGLIN YANG, and Jin Song Dong. Debugging and explaining metric learning approaches: An influence function based perspective. In Advances in Neural Information Processing Systems, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "text", + "content": "[31] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "text", + "content": "[32] Benjamin J. Meyer and Tom Drummond. The importance of metric learning for robotic vision: Open set recognition and active learning. In 2019 International Conference on Robotics and Automation (ICRA), pages 2924-2931, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 499 + ], + "type": "text", + "content": "[33] Timo Milbich, Karsten Roth, Homanga Bharadhwaj, Samarth Sinha, Yoshua Bengio, Björn Ommer, and Joseph Paul Cohen. Diva: Diverse visual feature aggregation for deep metric learning. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VIII, page 590–607, Berlin, Heidelberg, 2020. Springer-Verlag. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[34] Yair Movshovitz-Attias, Alexander Toshev, Thomas K Leung, Sergey Ioffe, and Saurabh Singh. No fuss distance metric learning using proxies. In Proceedings of the IEEE international conference on computer vision, pages 360–368, 2017. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 556, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 610 + ], + "type": "text", + "content": "[35] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "text", + "content": "[36] Qi Qian, Lei Shang, Baigui Sun, Juhua Hu, Hao Li, and Rong Jin. Softtriple loss: Deep metric learning without triplet sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[37] Samyak Rawlekar, Shubhang Bhatnagar, Vishnuvardhan Pogunulu Srinivasulu, and Narendra Ahuja. Improving multi-label recognition using class co-occurrence probabilities. In International Conference on Pattern Recognition, pages 424–439. Springer, 2025. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[38] Karsten Roth, Timo Milbich, Samarth Sinha, Prateek Gupta, Björn Ommer, and Joseph Paul Cohen. Revisiting training strategies and generalization performance in deep metric learning. In Proceedings of the 37th International Conference on Machine Learning. JMLR.org, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[39] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7420-7430, 2022. 1, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 229 + ], + "type": "text", + "content": "[40] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "text", + "content": "[41] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "type": "text", + "content": "[42] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "type": "text", + "content": "[43] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 355, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 410 + ], + "type": "text", + "content": "[44] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6397-6406, 2020. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "text", + "content": "[45] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H. S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1199-1208, 2017. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 467, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 531 + ], + "type": "text", + "content": "[46] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 567 + ], + "type": "text", + "content": "[47] Mark Tabb and Narendra Ahuja. Multiscale image segmentation by integrated edge and region detection. IEEE transactions on image processing, 6(5):642-655, 1997. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 633 + ], + "type": "text", + "content": "[48] Eu Wern Teh, Terrance DeVries, and Graham W. Taylor. Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV, pages 448-464. Springer, 2020. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[49] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[50] Chengkun Wang, Wenzhao Zheng, Junlong Li, Jie Zhou, and Jiwen Lu. Deep factorized metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7672-7682, 2023. 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25558" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 596 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[51] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "text", + "content": "[52] Kilian Q Weinberger, John Blitzer, and Lawrence Saul. Distance metric learning for large margin nearest neighbor classification. In Advances in Neural Information Processing Systems. MIT Press, 2005. 1, 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 217 + ], + "type": "text", + "content": "[53] Chao-Yuan Wu, R Manmatha, Alexander J Smola, and Philipp Krahenbuhl. Sampling matters in deep embedding learning. In Proceedings of the IEEE international conference on computer vision, pages 2840-2848, 2017. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 295, + 272 + ], + "type": "text", + "content": "[54] Tong Xiao, Shuang Li, Bochao Wang, Liang Lin, and Xiaogang Wang. Joint detection and identification feature learning for person search. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3415-3424, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 295, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 295, + 318 + ], + "type": "text", + "content": "[55] Hong Xuan, Abby Stylianou, and Robert Pless. Improved embeddings with easy positive triplet mining. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2463-2471, 2019. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 295, + 373 + ], + "type": "text", + "content": "[56] Bailin Yang, Haoqiang Sun, Frederick WB Li, Zheng Chen, Jianlu Cai, and Chao Song. Hse: Hybrid species embedding for deep metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11047-11057, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 295, + 418 + ], + "type": "text", + "content": "[57] Yuhui Yuan, Kuiyuan Yang, and Chao Zhang. Hard-aware deeply cascaded embedding. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 295, + 453 + ], + "type": "text", + "content": "[58] Andrew Zhai and Hao-Yu Wu. Classification is a strong baseline for deep metric learning. In *British Machine Vision Conference*, 2018. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 295, + 497 + ], + "type": "text", + "content": "[59] Wenzhao Zheng, Chengkun Wang, Jiwen Lu, and Jie Zhou. Deep compositional metric learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9316-9325, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 295, + 542 + ], + "type": "text", + "content": "[60] Wenzhao Zheng, Borui Zhang, Jiwen Lu, and Jie Zhou. Deep relational metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12065-12074, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 295, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 596 + ], + "type": "text", + "content": "[61] Yuehua Zhu, Muli Yang, Cheng Deng, and Wei Liu. Fewer is more: A deep graph metric learning perspective using fewer proxies. In Advances in Neural Information Processing Systems, pages 17792-17803. Curran Associates, Inc., 2020. 6, 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25559" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/fb9435c5-f90b-40a7-b204-e0d114acb0c9_content_list.json b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/fb9435c5-f90b-40a7-b204-e0d114acb0c9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..301f68ada06e5a3df00af8ed294c08a65311fb77 --- /dev/null +++ b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/fb9435c5-f90b-40a7-b204-e0d114acb0c9_content_list.json @@ -0,0 +1,1673 @@ +[ + { + "type": "text", + "text": "Pow3R: Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors", + "text_level": 1, + "bbox": [ + 264, + 130, + 733, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wonbong Jang* Philippe Weinzaepfel† Vincent Leroy† \nLourdes Agapito* Jerome Revaud† \n*UCL †Naver Labs Europe", + "bbox": [ + 261, + 202, + 754, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ucabwja,1.agapito}@ucl.ac.uk firstname.lastname@naverlabs.com", + "bbox": [ + 215, + 260, + 777, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd468f966acef602dc94daffba7df92ffce29633c66132b6a368cb37e4b2ad19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 292, + 276, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7996d15fe1224731089515b787255ed643c7e38300242e17396ae4643c12fe8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 391, + 274, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/908ee6513e718a96c7c64ef3920fb3a7b74c29bfbb673f221e9a03e007190970.jpg", + "image_caption": [ + "Figure 1. Qualitative comparison of raw 3D reconstructions between DUSt3R and Pow3R for the image pair on the left. DUSt3R cannot leverage auxiliary information potentially available at test time, resulting in noisy and low-resolution estimates, whereas Pow3R allows to inject such priors at test time in a versatile manner, enabling new capabilities such as high-resolution processing as a direct by-product." + ], + "image_footnote": [], + "bbox": [ + 277, + 292, + 578, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/657a18594bd983f7c7c17f58b7385fd5ba70716fe8c587b43ae61d0c62f47d96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 292, + 905, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 537, + 326, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Pow3R, a novel large 3D vision regression model that is highly versatile in the input modalities it accepts. Unlike previous feed-forward models that lack any mechanism to exploit known camera or scene priors at test time, Pow3R incorporates any combination of auxiliary information such as intrinsics, relative pose, dense or sparse depth, alongside input images, within a single network. Building upon the recent DUSt3R paradigm, a transformer-based architecture that leverages powerful pre-training, our lightweight and versatile conditioning acts as additional guidance for the network to predict more accurate estimates when auxiliary information is available. During training we feed the model with random subsets of modalities at each iteration, which enables the model to operate under different levels of known priors at test time. This in turn opens up new capabilities, such as performing inference in native image resolution, or point-cloud completion. Our experiments on 3D reconstruction, depth completion, multi-view depth prediction, multi-view stereo, and multi-view pose estimation tasks yield state-of-the-art results and confirm the effectiveness of Pow3R at exploiting all available information. The project webpage is https://europe.naverlabs.com/pow3r.", + "bbox": [ + 86, + 568, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 537, + 643, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building general models for 3D perception that can unify different 3D vision tasks such as depth estimation, keypoint matching, dense reconstruction or camera pose prediction, remains a complex and exciting open challenge. Traditional Structure from Motion (SfM) approaches such as COLMAP [59] do not apply pre-learned priors, so each scene must be optimized independently. On the other hand, most learning-based approaches focus on specific 3D vision tasks such as depth completion [26, 52], monocular and stereo depth estimation [46], pose estimation [73, 89] or novel view synthesis [27, 42]. While these dedicated approaches have been successful to some extent, this trend contrasts with other domains such as natural language processing where a single model like ChatGPT [45], trained with a vast collection of data, solves a range of different tasks without requiring re-training. While some models like CroCo-v2, DINOv2 or Stable Diffusion [15, 54, 77], trained on large image datasets, have served as learned priors for other tasks, including 3D vision ones, they are still far from multi-purpose foundation models like ChatGPT is for NLP.", + "bbox": [ + 511, + 566, + 906, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DUSt3R [75] marked a significant step forward in regression-based 3D foundation models. DUSt3R performs", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1071", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D dense reconstruction from uncalibrated and unposed images, solving many downstream vision tasks at once without the need for fine-tuning. Its success can be attributed to two main factors: (i) its use of a pointmap representation, which provides flexibility over its output space and enables to regress camera parameters, poses and depth easily and (ii) its use of a transformer architecture that exploits powerful pretraining combined with a vast amount of 3D training data. Despite its huge success, DUSt3R [75] and its successor MASt3R [31] are extremely limited in terms of their input space, i.e. they only take RGB images. Yet, in practice, many real-world applications do provide extra modalities such that calibrated camera intrinsics, sparse or dense depth from RGB-D sensors or LIDAR, etc. This raises the question of how best to exploit additional input modalities, and whether this not just boosts performance but also enables new capabilities, such as obtaining the full depth from sparse point clouds of SfM or LiDAR.", + "bbox": [ + 89, + 90, + 480, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce Pow3R, a new 3D feedforward model that empowers DUSt3R with any subset of priors available at test time such as camera intrinsic, sparse or dense depth, or relative camera poses. Each modality is injected into the network in a lightweight fashion. To allow the model to operate under different conditions at test time, random subsets of input modalities are fed at each training iteration. As a result, we obtain a single model that performs on par with DUSt3R when no prior information is available but outperforms it when it exists. Our model also gains new capabilities as a by-product: for instance, the camera intrinsic input allows to process images whose principal point is far from the center, thus allowing to perform extreme cropping, e.g. for performing sliding window inference. As a side contribution, our model directly outputs the pointmaps of the second image in its coordinate system, allowing faster relative pose estimation. To sum up, our contributions are:", + "bbox": [ + 91, + 363, + 482, + 633 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pow3R is a holistic 3D geometric vision model capable of taking any subset (including none) of camera intrinsics, pose and depthmaps with corresponding input images.", + "- Extensive experiments demonstrate an important boost in performance over DUSt3R which cannot exploit priors, yielding state-of-the-art results in three benchmarks.", + "- By predicting the same pointmaps in two different camera coordinate systems, we can achieve more accurate relative pose, orders of magnitude faster." + ], + "bbox": [ + 89, + 633, + 482, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 782, + 227, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Structure-from-Motion and MVS. SfM [12, 13, 22, 25, 59] builds 3D sparse maps by optimizing camera parameters using only RGB images. Traditional approaches consist of handcrafted pipelines to match keypoints [5, 7, 21, 36, 55], form pixel correspondences across multiple images, determine geometric relations, and perform bundle", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adjustment to optimize camera locations. Recently, SfM has incorporated learning-based approaches leveraging neural networks as components [4, 10, 35, 57, 66, 70, 74, 85], but it remains brittle to outliers and challenging conditions. Likewise, multi-view stereo (MVS) aims to reconstruct dense 3D surface through handcrafted pipelines based on triangulation from multiple viewpoints [18, 19, 60]. Learning-based approaches have also been incorporated for MVS to reconstruct 3D from few images [24, 30, 44, 64, 78, 83], and more recently by adopting the RAFT architecture [40] or using generic Transformers [9]. In contrast, the recent DUSt3R framework [31, 75], upon which we build our approach, departs from these methods altogether by defining a simple and unified neural framework to 3D reconstruction based on the paradigm of pointmap regression. The idea is that, ideally, learning-based MVS should be capable of reconstructing 3D from a few images without strictly requiring the availability of camera parameters.", + "bbox": [ + 511, + 90, + 903, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Guiding 3D. In this paper, we refer to guidance as the notion that a method can optionally take in 3D priors (typically camera parameters or sparse depth) to help with some prediction task. This can take different forms in the vast literature of 3D vision works. COLMAP [60], for instance, can take camera parameters for rectifying camera views; however, it is an optimization-based method, meaning that each scene has to be processed independently. Recently, UniDepth [46] is a monocular depth and intrinsic estimation method capable of optionally providing camera intrinsics as additional guide to generate depthmap, with an architecture specifically handcrafted for that purpose. Also, depth completion models [68, 69, 76, 80] such as CompletionFormer [26] take sparse depth input as guidance in order to inpaint the full depthmap, but are also typically capable of inferring depth from an image alone. Overall, however, there is currently no learning-based model to the best of our knowledge that guides its output by taking any subset of sparse depth, camera intrinsics and camera extrinsics.", + "bbox": [ + 511, + 364, + 903, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth Completion. Given sparse depths and a corresponding RGB image, the task is to densify depth by propagating sparse measurements. CNN-based approaches were adopted by [38, 39], and [11] proposed convolutional spatial propagation network (CSPN). Multi-branch networks, which process RGB and sparse depth separately, were used in [23, 43, 48, 52, 68, 71, 79, 90]. CompletionFormer [26] leverages both CNN and ViT architectures to capture both global and local features. Most methods focus only on the specific depth completion task from a single image and partial depth as the input, often without incorporating camera intrinsics. Using the pointmap representation, Pow3R, based on a ViT architecture, can complete the depth map from both RGB images and sparse point clouds, while taking relative pose or camera intrinsics as auxiliary input.", + "bbox": [ + 511, + 655, + 903, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pose Estimation from Sparse views. Learning-based ap", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1072", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6b7cba457c5a6465b3b0f24a09a752630003b667442bbfacaf8b712112300c24.jpg", + "image_caption": [ + "Figure 2. Overview of the model architecture. Following DUSt3R [75], images are encoded then decoded with a ViT backbone into pointmaps, from which focals, depthmaps and relative pose can be extracted. Pow3R introduces optional inputs to guide the regression with prior knowledge about the camera intrinsics and depth (fed into the encoder) and the pose (into the decoder)." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 482, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "proaches involve regressing relative pose, applying probabilistic methods, using energy-based or diffusion models [3, 33, 53, 73, 88]. RayDiffusion [89] models camera rays instead of camera parameters, applying diffusion on rays, then recovers the camera pose using DLT [2]. DUSt3R discovers camera pose from pointmaps using PnP, while Pow3R predicts pointmaps in both camera coordinate frames, allowing the use of Procrustes alignment.", + "bbox": [ + 88, + 313, + 482, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DUSt3R and its developments. DUSt3R [75] does not require calibrated images for 3D reconstruction, can handle extreme viewpoint changes, and deals with both single and a few images within the same framework. Its pointmap representation eases the training by not having to deal with camera parameters directly. MAST3R [31] improves DUSt3R by also predicting dense local features, trained with an additional matching loss. Splatt3R [65] marries MAST3R with GS. Spann3R [72] predicts pointmaps in global coordinate and adopts an external spatial memory that tracks relevant 3D information. MonST3R [87] further trains DUSt3R on dynamic scenes and generates time-varying dynamic point clouds, as well as camera intrinsics and poses from a dynamic video. In this work, we take a different and complementary direction and instead improve the core capabilities of DUSt3R, which could benefit all aforementioned works.", + "bbox": [ + 89, + 436, + 483, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 693, + 225, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to train a network $\\mathcal{F}$ that can take two input images $I^1, I^2 \\in \\mathbb{R}^{W \\times H \\times 3}$ of a given static scene and any subset of auxiliary information $\\Omega \\subseteq \\{K_1, K_2, P_{1,2}, D^1, D^2\\}$ , in order to regress a 3D reconstruction of the scene. Here, $K_1, K_2 \\in \\mathbb{R}^{3 \\times 3}$ are camera intrinsics, $P_{1,2} \\in \\mathbb{R}^{4 \\times 4}$ denote the relative pose between the two cameras, and $D^1, D^2 \\in \\mathbb{R}^{W \\times H}$ are depthmaps with associated masks $M^1, M^2 \\in \\{0,1\\}^{W \\times H}$ specifying pixels with valid depth data (i.e. masks are possibly sparse). The network $\\mathcal{F}$ is tasked to regress several pointmaps from which the camera intrinsic and extrinsic parameters as well as the dense depthmaps can be straightforwardly extracted for both im", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ages as described below.", + "bbox": [ + 511, + 92, + 679, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pointmaps. For each pixel $(i,j)$ in an image $I$ , we assume there exists a corresponding single 3D point $X_{i,j}$ where $X \\in \\mathbb{R}^{W \\times H \\times 3}$ is a pointmap. Given camera intrinsics $K$ and a depthmap $D$ , we can compute $X_{i,j} = K^{-1}[iD_{i,j},jD_{i,j},D_{i,j}]$ in the camera coordinate system. In the following, we denote as $X^{n,m}$ the pointmap of image $I^n$ expressed in the coordinate system of camera $I^m$ . To swap the coordinate system from camera $I^k$ to camera $I^m$ , we have $X^{n,k} = P_{m,k}X^{n,m}$ where $P_{m,k} = P_kP_m^{-1}$ .", + "bbox": [ + 511, + 108, + 906, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall architecture", + "text_level": 1, + "bbox": [ + 511, + 255, + 705, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow the recently proposed DUSt3R framework [75]. DUSt3R is a breakthrough foundation model based on vision transformers (ViT) that can regress two 3D pointmaps $X^{1,1}$ , $X^{2,1}$ given solely two unposed and uncalibrated input images. Nevertheless, DUSt3R's inability to leverage potentially available information about cameras or depth limits its practical use. To remedy these drawbacks, we enhance the original DUSt3R network $\\mathcal{F}$ in two ways.", + "bbox": [ + 511, + 277, + 906, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we introduce specific modules to incorporate any subset of extra information such as camera intrinsics, camera poses and depthmaps seamlessly into DUSt3R. Second, we predict an additional pointmap $X^{2,2}$ , which represents the pointmap of image $I^2$ in its own coordinate system. Predicting three pointmaps offers further capabilities, namely the possibility of recovering all information about both cameras in a single forward pass, as explained in Sec. 3.3. We first describe the overall architecture, see Fig. 2, and then detail the dedicated modules to inject auxiliary information.", + "bbox": [ + 511, + 398, + 908, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoder. We use a shared ViT encoder to encode both images independently. In addition to $I$ , the encoder can receive auxiliary information about intrinsics $K$ and depth $D$ for each image. For the two input images $I^1, I^2$ and their respective auxiliary information $\\Omega_1 \\in \\sigma(\\{K_1, D_1\\}), \\Omega_2 \\in \\sigma(\\{K_2, D_2\\})$ , where $\\sigma$ denotes the set of all subsets, the encoder processes the information in a Siamese manner:", + "bbox": [ + 511, + 551, + 906, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {1} = \\operatorname {E n c o d e r} \\left(I ^ {1}, \\Omega_ {1}\\right), F ^ {2} = \\operatorname {E n c o d e r} \\left(I ^ {2}, \\Omega_ {2}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 669, + 906, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decoder. Likewise, the network has two decoders, each with its corresponding head, one predicting $X^{1,1}$ and the other one estimating $X^{2,1}$ and $X^{2,2}$ . Both decoders communicate via cross-attention between their own tokens and the outputs of the previous block of the other decoder. Each decoder may receive the relative pose $P_{12}$ as additional input or not. Assume we provide the auxiliary information $\\Omega_D \\in \\sigma(\\{P_{12}\\})$ at the $i$ -th block of both decoders:", + "bbox": [ + 511, + 702, + 906, + 824 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} G _ {i} ^ {1} = \\operatorname {D e c o d e r B l o c k} _ {i} ^ {1} \\left(G _ {i - 1} ^ {1}, G _ {i - 1} ^ {2}, \\Omega_ {D}\\right), \\\\ G _ {i} ^ {2} = \\operatorname {D e c o d e r B l o c k} _ {i} ^ {2} \\left(G _ {i - 1} ^ {2}, G _ {i - 1} ^ {1}, \\Omega_ {D}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 834, + 903, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After $B$ decoder blocks in each branch, the head regresses", + "bbox": [ + 511, + 885, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1073", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "several pointmaps and their associated confidence map:", + "bbox": [ + 89, + 90, + 457, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {1, 1}, C ^ {1, 1} = \\operatorname {H e a d} ^ {1} \\left(G _ {B} ^ {1}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 116, + 398, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {2, 1}, X ^ {2, 2}, C ^ {2, 1}, C ^ {2, 2} = \\operatorname {H e a d} ^ {2} \\left(G _ {B} ^ {2}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 137, + 482, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike DUSt3R which uses a DPT head [50], here we employ a linear head for both $\\mathrm{Head}^1$ and $\\mathrm{Head}^2$ as it is more efficient and performs similarly.", + "bbox": [ + 89, + 166, + 482, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Regression Loss. Following DUSt3R, we supervise the model to minimize the distance between ground-truth and predicted pointmaps in a scale-invariant manner, allowing the model to train on multiple datasets with various scales. The regression loss between predicted and ground-truth pointmaps (resp. $X^{n,m}$ and $\\hat{X}^{n,m}$ ) at pixel $(i,j)$ is defined as $\\mathcal{L}_{i,j}^{\\mathrm{gr}}(n,m) = \\| X_{i,j}^{n,m} / z_m - \\hat{X}_{i,j}^{n,m} / \\hat{z}_m\\|$ , where $z_{m},\\hat{z}_{m}$ serve as scale normalizer. That is, $z_{m}$ is the average norm of all valid 3D points expressed in coordinate system of image $I^{m}$ , i.e. $z_{1} = \\mathrm{norm}(X^{1,1}\\cup X^{2,1})$ , $z_{2} = \\mathrm{norm}(X^{2,2})$ and likewise for $\\hat{z}_1,\\hat{z}_2$ , with $\\mathrm{norm}(X) = \\mathrm{mean}(\\| X_{i,j}\\| |i,j\\in \\mathcal{D}_X)$ ) and $\\mathcal{D}_X$ the set of valid pixels.", + "bbox": [ + 89, + 214, + 483, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Confidence-aware loss. The model jointly learns to predict a confidence level $C_{i,j}^{n,m}$ of each pixel $(i,j)$ . The confidence-aware regression loss for a given pointmap $X^{n,m}$ can be expressed as the 3D regression loss $\\mathcal{L}^{\\mathrm{reg}}$ weighted by the confidence map:", + "bbox": [ + 89, + 397, + 483, + 472 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {c o n f}} (n, m) = \\sum_ {i, j \\in \\mathcal {D}} C _ {i, j} ^ {n, m} \\mathcal {L} _ {i, j} ^ {\\text {r e g r}} (n, m) - \\alpha \\log C _ {i, j} ^ {n, m}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 483, + 482, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\alpha = 0.2$ . This loss penalizes the model less when the prediction is not accurate on harder areas, encouraging the model to extrapolate. The final loss is expressed as", + "bbox": [ + 89, + 527, + 483, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} ^ {\\text {c o n f}} (1, 1) + \\mathcal {L} ^ {\\text {c o n f}} (2, 1) + \\beta \\mathcal {L} ^ {\\text {c o n f}} (2, 2), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 583, + 482, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ is a hyper-parameter typically set to $\\beta = 1$ .", + "bbox": [ + 89, + 612, + 434, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Adding Versatile Conditioning", + "text_level": 1, + "bbox": [ + 89, + 637, + 362, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The knowledge of auxiliary information can significantly enhance 3D predictions at test time. Our model leverages up to five different modalities, which include two intrinsics $K_{1}, K_{2}$ , two depthmaps $D^{1}, D^{2}$ for each image and the relative pose $P_{12}$ . To condition the output on it, we first embed the auxiliary information using dedicated MLPs and then inject these embeddings at different points in the pipeline.", + "bbox": [ + 89, + 659, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We discuss different options in this regard. In the first strategy, denoted as 'embed', we simply add the auxiliary embeddings to the token embeddings before the first transformer block, as would be done for standard positional embeddings [16]. In the second strategy, denoted as 'inject- $n$ ' we instead insert dedicated MLPs for each modality in a subset of $n$ transformer blocks, as shown in Fig. 3. Our experiments show that 'inject-1' performs slightly better than 'embed' and similarly with 'inject- $n$ ' where $n > 1$ , see", + "bbox": [ + 89, + 765, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ab31b72c696d7f4c4a54a16400083834c04f94c8bb7b56e7c245739b0289e580.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 90, + 875, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c77f3013c264b67a5524128d9598d745fffda9b53e92dc598ef6661508aa0cfd.jpg", + "image_caption": [ + "Figure 3. Architecture for injecting auxiliary modalities. Top: injection of optional intrinsics and depth into the encoder. Intrinsic s are encoded into ray patches, sparse depth is patchified. Each of these modalities goes into a block-specific MLP and are tokenwise added in the middle of the transformer block. Bottom: injection of optional relative pose into the decoder. The relative pose is fed to a first embedding layer followed by an MLP. This token is added to the CLS token of the decoder after the self-attention (SA) and cross-attention (CA), but before the MLP. Experiments show that injection in the first block only suffices." + ], + "image_footnote": [], + "bbox": [ + 532, + 191, + 877, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sec. 4.5. We now describe how to compute the initial embeddings for each specific modality.", + "bbox": [ + 511, + 431, + 905, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intrinsics. Following [89], we generate camera rays from the intrinsic matrix $K \\in \\mathbb{R}^{3 \\times 3}$ , thereby establishing a direct correspondence between RGB pixels and rays. The ray at pixel location $(i,j)$ is computed as $K^{-1}[i,j,1]$ and encodes the viewing direction of that pixel w.r.t. the current camera pose. This allows us to potentially process non-centered crops and hence to perform inference in higher image resolutions, see Sec. 3.3. Similarly to RGB inputs, we patchify and embed dense rays and provide them to the encoder.", + "bbox": [ + 511, + 465, + 905, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depthmaps / Point Clouds. Given a depthmap $D$ and its sparsity mask $M$ , we first normalize $D' = D / \\mathrm{norm}(D)$ to handle any depth ranges at train and test time. As for RGB and rays, we patchify the stacked maps $[D', M] \\in \\mathbb{R}^{W \\times H \\times 2}$ and compute patch embeddings, which are then fed to the encoder. By jointly patchifying the depth and its valid mask we can handle any level of sparsity.", + "bbox": [ + 511, + 603, + 905, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera Pose. Given the relative pose $P_{12} = [R_{12}|t_{12}]$ , we normalize the translation scale as $t_{12}' = t_{12} / \\|t_{12}\\|$ since our output is unscaled. Unlike depthmaps or camera intrinsics, the camera pose cannot be expressed as a dense pixel map. Rather, camera pose affects the whole pixels between two images, so the embedding is instead added to the global CLS token of both decoders.", + "bbox": [ + 511, + 710, + 905, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Downstream Tasks", + "text_level": 1, + "bbox": [ + 513, + 830, + 694, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depthmaps. In the pointmap representation, the $z$ -axis of $X^{1,1}, X^{2,2}$ directly corresponds to the depthmaps of the first and second image, respectively. Note that DUSt3R", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1074", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/17d580084d03bb1334327108baa4f1b493068a4a757a8cba3f3467c63d8eaee0.jpg", + "image_caption": [ + "Figure 4. Native resolution example. Methods like DUSt3R require centered principal point and would thus downscale the image to its training resolution, leading to blurry prediction (top). With camera intrinsics input, we can process any crop in the image leading to native resolution prediction (bottom row), for example by dividing the image into 4 crops, that are independently process as the first input image, while the second input image of the network is set to the downscaled image." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 385, + 232 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/babbe8245315e7913c89d97a8f11bc2104888d68a413b018c1eddb82ae7f6cdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 88, + 766, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/01f03d6f78e2f1123369db61f5134fd05c3d7451b82eb8111f8cfe6681fda966.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 88, + 903, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "only outputs $X^{1,1}$ , hence it requires two forward passes of the pairs $(I^1, I^2)$ and $(I^2, I^1)$ to extract both depthmaps.", + "bbox": [ + 89, + 308, + 482, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "High-resolution processing. A limitation of DUSt3R is that it is trained for a specific resolution and does not generalize to larger resolutions. A simple fix would be to include higher resolutions during training, but (i) many datasets do not provide high-resolution ground-truths, and (ii) the training cost could become prohibitive. Pow3R, in contrast, can handle crops natively given camera intrinsic's of the crop, as these provide the crop position information (i.e. via focal length and principal point). We can thus perform prediction in a sliding window fashion, yielding predictions matching any target resolution by simple stitching. Comparison between standard and higher resolution processing are shown in Fig. 4. Note that prediction for each crop may have a different scale, by design, and cannot be stitched directly. We find that simply computing the median scale factor in overlapping areas is sufficient, and we blend the overlapping regions based on confidence without further post-processing.", + "bbox": [ + 89, + 340, + 482, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Focal Estimation. Similar to DUSt3R, we can recover fccals for both input images from pointmap $X^{1,1}$ and $X^{2,2}$ with the robust Weiszfeld fast iterative solver [47]. Similar to depthmap prediction, DUSt3R needs to infer $(I_2, I_1)$ in order to compute the focal of the second image, while Pow3R predicts them in a single pass.", + "bbox": [ + 89, + 599, + 482, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Relative Pose Estimation. Pow3R predicts the relative pose directly by Procrustes alignment [8, 37] to get the scaled relative pose $P^{*} = \\left[R^{*}|t^{*}\\right]$ between $X^{2,2}$ and $X^{2,1}$ , as it predicts the pointmaps of the second image in two different camera coordinates.", + "bbox": [ + 89, + 690, + 482, + 761 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {*}, t ^ {*} = \\arg \\min _ {\\sigma , R, t} \\sum_ {i, j} \\sqrt {C _ {i , j} ^ {2 , 2} C _ {i , j} ^ {2 , 1}} \\left\\| \\sigma \\left(R X _ {i, j} ^ {2, 2} + t\\right) - X _ {i, j} ^ {2, 1} \\right\\| ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 768, + 482, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Procrustes alignment is known to be sensitive to noise and outliers; however, in our case, it performs similar to RANSAC [17] with PnP [22, 29], while being an order of magnitude faster (see Sec. 4.4).", + "bbox": [ + 89, + 808, + 483, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Global alignment. The network $\\mathcal{F}$ predicts pointmaps for image pairs. To align all predictions in the same world coor", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dinate system, we resort to the global alignment algorithm from DUSt3R [75] that minimizes a global energy function to find per-camera intrinsics, depthmaps and poses that are consistent with all the pairwise predictions. Results of the optimization are global scene point-clouds, which can for instance serve for multi-view stereo estimation.", + "bbox": [ + 511, + 309, + 906, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Procedure", + "text_level": 1, + "bbox": [ + 511, + 410, + 700, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we feed the network with annotated image associated with random subsets of auxiliary information, the goal being for it to learn to handle any situation at test time. For each pair, we first chose a random number $m$ of modalities with uniform probability, and then randomly select the $m$ modalities likewise. Depthmaps are randomly sparsified. When giving intrinsics, we also perform aggressive non-centered cropping with $50\\%$ probability, so that the network learns to perform high-resolution inference.", + "bbox": [ + 511, + 431, + 906, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training data and recipe. We follow the training recipe from DUSt3R [75] in terms of training data and protocol. Namely, we use the provided pairs for 8 datasets: Habitat [58], MegaDepth [32], ARKitScenes [6], Static Scenes 3D [41], BlendedMVS [82], ScanNet++ [84], Co3Dv2 [51] and Waymo [67], resulting in a total of 8.5M image pairs. These datasets include indoor and outdoor scenes, as well as real and synthetic ones. To train the model, we follow [75] and first train the network with a resolution of 224px for 3 days on 8 A100 GPUs, and then finetune it in 512px with variable aspect-ratio for 2 more days.", + "bbox": [ + 511, + 570, + 908, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 511, + 751, + 750, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first study the impact of auxiliary information guidance (Sec. 4.1). We then experiment across many 3D vision tasks (Secs. 4.2 to 4.4) without retraining, i.e. mostly zero-shot settings. We finally present some ablations in Sec. 4.5.", + "bbox": [ + 511, + 777, + 906, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DUSt3R baseline. We point out that our overall training data, protocols and architecture are exactly the same as DUSt3R, except for the additional head that predict $X^{2,2}$ ( $+0.1\\%$ parameters) and for the modules used for injecting", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1075", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ec9bc7837c5d6c28fbacbabe7fc824f95cc113bb8b1dbc2d04b4585431661af0.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Impact of guiding at test time for models trained at $224 \\times 224$ resolution. We report performances on Habitat for DUSt3R, which cannot handle auxiliary modalities, and our single model with different sets of modalities; we show in green the absolute improvement w.r.t. the results without auxiliary modality." + ], + "table_body": "
aux. modalitiesfocal acc@1.015depth τ@1.03rel. pose
K1K2D1D2RTRRA@2°RTA@2°
DUSt3R×××××36.677.669.249.7
Pow3R×××××39.479.471.953.8
✓××××75.4(+36.0)80.6 (+1.2)77.1 (+5.3)62.8 (+8.9)
×✓×××67.3(+27.9)80.3 (+0.9)74.5 (+2.6)59.1 (+5.2)
✓✓×××98.0(+58.6)81.4 (+2.0)80.3 (+8.5)74.2(+20.3)
××✓××48.9 (+9.5)89.1 (+9.7)82.5(+10.6)64.9(+11.0)
×××✓×49.5(+10.1)91.0(+11.7)83.4(+11.5)64.8(+10.9)
××✓✓×58.2(+18.9)95.4(+16.0)89.6(+17.7)77.6(+23.8)
××××✓48.6 (+9.2)81.6 (+2.2)92.3(+20.5)77.0(+23.1)
✓✓✓✓×99.2(+59.8)95.4(+16.0)95.7(+23.9)94.3(+40.5)
✓✓××✓98.1(+58.7)82.9 (+3.6)95.1(+23.2)87.3(+33.5)
××✓✓✓68.6(+29.2)95.4(+16.0)98.1(+26.2)91.3(+37.5)
✓✓✓✓✓99.3(+59.9)95.4(+16.0)99.0(+27.2)98.1(+44.3)
", + "bbox": [ + 93, + 88, + 480, + 248 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the auxiliary information $(+4\\%$ parameters for the inject-1 variant that we use unless stated otherwise).", + "bbox": [ + 89, + 333, + 482, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. When evaluating focal lengths, we report the accuracy for a tight relative error threshold of $1.5\\%$ . For evaluating depthmap, we report the Absolute Relative Error (rel) and the Inlier Ratio $(\\tau)$ at a threshold of $3\\%$ . When evaluating poses, we opt for scale-invariant metrics and measure the relative rotation and translation accuracies (RRA and RTA resp.) of the relative pose at task-specific thresholds, and the mean Average Accuracy (mAA), defined as the area under the curve accuracy of the angular differences at min(RRA, RTA).", + "bbox": [ + 88, + 364, + 482, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Guiding the output prediction", + "text_level": 1, + "bbox": [ + 89, + 527, + 357, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of guidance. We first study the impact of providing different subsets of auxiliary information in terms of downstream 3D vision performance. Namely, we report accuracies in Tab. 1 for focal, depth, and rotation prediction on the Habitat validation dataset. In the supplementary material, we also provide results on the Infinigen dataset [49], which is not included in our training set, and observe similar outcomes. We first note that in the absence of modalities, Pow3R performs slightly better than a DUSt3R baseline, even though Pow3R is trained with just RGB images during a fraction of its training. This shows that Pow3R does not lose its native strength for unconstrained regression despite having gained more capabilities.", + "bbox": [ + 88, + 551, + 482, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interestingly, we observe in Tab. 1 the impact of each modality for each prediction task, and the synergetic effects of interacting modalities. For instance, the model performs better for focal length estimation with depthmaps and/or relative pose, and Pow3R's depth prediction improves with intrinsics and/or relative pose. Moreover, relative pose prediction becomes more accurate with intrinsics and/or depthmap. Overall, the model's performance consistently improves when provided with additional camera or scene priors, demonstrating the effectiveness of Pow3Rin", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/826dd2e89c2688b8204dd337cb9dc4237dd5a3a1d8bc67d6b14bb35cdffc7c4f.jpg", + "image_caption": [ + "Figure 5. Depth completion performance on NYUd for various input depth sparsity." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 903, + 205 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a20042584d1526d1423e3699a3a4d22b569cfd97994cdbf5f1648b32d0d9f91d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Multi-view depth estimation results in high-resolution using a sliding-window scheme and providing the crop intrinsics to Pow3R. (n): naively inputting high-resolution images without downsampling to Pow3R." + ], + "table_body": "
aux. mod.High-Res.KITTIT&TTime (sec)↓
KsRTrel↓τ↑rel↓τ↑
DUSt3R512×××5.449.53.375.10.13
Pow3R512✓✓×5.348.73.278.20.13
Pow3R512✓✓(n)7.534.43.968.00.48
Pow3R512✓✓✓4.653.52.582.30.40
", + "bbox": [ + 517, + 243, + 903, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "handling any combination of priors in a unified fashion. We are not aware of any other approach that can seamlessly integrate all these priors in a learning-based method.", + "bbox": [ + 511, + 407, + 903, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Depth completion. We also evaluate Pow3R on the depth completion task given RGB images with sparse depthmaps. We vary the sparsity (the ratio of pixels with known depth) and demonstrate its impact on the NYUv2 dataset [63] in Figure 5. Although Pow3R is not trained on the NYUv2 dataset, it still achieves state-of-the-art results for sparse depth completion. Our analysis indicates that the true performance of Pow3R on NYUv2 is likely underestimated due to numerous errors in the ground-truth depth annotations; further details are available in the supplementary material. On the KITTI [20] dataset, we obtain a mean absolute error below $30\\mathrm{cm}$ , which is competitive with the state of the art, considering that KITTI is not included in the training set.", + "bbox": [ + 511, + 455, + 906, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "High-Resolution processing. With the possibility of specifying camera intrinsics, new capabilities arise. Namely, Pow3R can perform high-resolution pointmap regression with a simple sliding-window approach. We process fixed-resolution crops of the original images under the guidance of both (i) the crop information, provided in term of camera intrinsics, and (ii) a second image, itself possibly down-scaled or cropped as well. More details are provided in Sec. 3.3 and Fig. 4. When applied, results significantly improve compared to processing the images at 512-pixel resolution, as shown in Tab. 2. We also compare with a naive baseline where we simply feed high-resolution images directly to Pow3R, but it dramatically fails.", + "bbox": [ + 511, + 655, + 905, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Camera controllability. We study how controllable are the regression outputs on 10K Habitat test pairs given arbitrary camera parameters. For each test pair with ground-truth fo", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1076", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f5fa1fa4f400cd95e2de9546aeab43e83115042a21d3c66b1bba5662cec59172.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 90, + 285, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b1c173f2a5a1298768c924cd3cf424f9cf3cd32a8bf924bf5780380fdc36fa5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 90, + 478, + 202 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b2d66dbb9e23093374d1bf4125e289508593aa82cbf9c68e7b0aaa3576c439e0.jpg", + "image_caption": [ + "Figure 6. Controllability study. We study how the model reacts when presented with false focal (top) and false relative pose (bottom). We plot the input vs. predicted camera parameters in the left column and the average confidence in the right column. Interestingly, the model seems to not take the input information at face value and instead makes an informed decision whether to use it or not, in which latter case the confidence becomes low." + ], + "image_footnote": [], + "bbox": [ + 94, + 205, + 284, + 318 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd045e1f9063569778ba0a83867a01cfd78d48c4a4366ef39388058716586544.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 205, + 478, + 318 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cal $f$ and relative pose $P_{12}$ , we purposefully input a wrong focal (resp. relative pose) and measure the output focal $\\hat{f}$ (resp. pose $\\hat{P}_{12}$ ) as predicted from the raw pointmaps. Results are presented in Fig. 6 in terms of focal ratio $\\hat{f} / f$ and angle $d_{\\text{geodesic}}(P_{12}, \\hat{P}_{12})$ . Interestingly, the model accepts to follow the guidance whenever it is close to the ground truth value (i.e. staying close to the identity transformation) but then refuses to follow it when it gets too different, showing that the model is able to somehow 'judge' of the guidance's worth. In all cases, the output confidence clearly correlates with the agreement between the guidance and the ground truth. Overall, the model is thus not taking the auxiliary information at face value, and instead combines it with the other modalities (e.g. RGB) it in a more subtle and informed way than what could be achieved with handcrafted methods. This suggests that Pow3R would be robust to noise in the auxiliary information.", + "bbox": [ + 88, + 436, + 482, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Multi-View Depth Estimation", + "text_level": 1, + "bbox": [ + 89, + 707, + 354, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following RobustMVD [62], performances are measured on the KITTI [20], ScanNet [14], ETH3D [61], DTU [1], and Tanks and Temples [28] datasets. We report in Tab. 3 the Absolute Relative Error (rel) and Inlier Ratio $(\\tau)$ for each dataset when using (1) images alone, (2) images with relative poses, (3) images with camera intrinsics, and (4) images with both modalities, and provide results for more state-of-the-art methods in the supplementary material.", + "bbox": [ + 89, + 729, + 482, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Pow3R again performs roughly on par with DUSt3R in the absence of auxiliary information (e.g. 3.73 average error across all datasets for DUSt3R-512 vs. 3.64", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6d0b83e7a49f5d04069a94b1b91281bf4b53f2a6d81a8816b94fd4f470864a3a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGT intrinsicsCo3Dv2RealEstate10K mAA(30)Speed (fps)
RRA@15RTA@15mAA(30)
Colmap+SG [56]✓36.127.325.345.2-
PixSfM [34]✓33.732.930.149.4-
RelPose [88]×57.1----
(a) PoseDiff [73]×80.579.866.548.0-
RelPose++ [33]×85.5----
RayDiff [89]×93.3----
DUSt3R (PnP) [75]×94.388.477.261.73.2
(b) Pow3R (PnP)×94.889.978.562.53.2
Pow3R (Pro)×94.690.378.166.330.9
Pow3R w/ K (Pro)✓95.092.182.272.530.1
", + "bbox": [ + 514, + 88, + 903, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Multiview pose estimation on the CO3Dv2 and RealEstate10K. (a) denotes multiview methods, while (b) denotes pairwise methods. Pow3R with Procrustes alignment (Pro) performs better than DUSt3R (PnP) and other multi-view based pose estimation approaches, and even more so with known intrinsics, while being much faster. RealEstate10K is not in the training set.", + "bbox": [ + 511, + 223, + 903, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for Pow3R-512). However, it is clear that providing auxiliary information consistently improves the depth estimates. In fact, Pow3R-512 with both pose and intrinsics achieves state-of-the-art performance on the ETH3D, DTU, and Tanks and Temples datasets, as well as on average across datasets. Interestingly, intrinsics tend to improve performance more than the GT pose, which alone seems to have a negligible impact. This might be due to the fact that camera pose primarily helps the prediction of $X^{2,1}$ , while the depthmap of this task is obtained from the $z$ axis of $X^{1,1}$ alone. Still, we observe a clear synergy in providing jointly intrinsics and relative pose.", + "bbox": [ + 511, + 325, + 906, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Multi-View Stereo", + "text_level": 1, + "bbox": [ + 511, + 532, + 691, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the quality of global 3D reconstructions on the DTU [1] benchmark in the presence and absence of auxiliary information. For each subset of available information from $\\{(K_1, K_2), P_{12}\\}$ , we extract pairwise predictions from image pairs and align them in a global coordinate system, as described in Sec 3.3. Performance is evaluated in terms of accuracy, which is the smallest Euclidean distance to the ground-truth, and completeness as the smallest distance to the reconstructed shape, with the overall average.", + "bbox": [ + 511, + 561, + 906, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results are presented in Tab. 4. We first attempt at reproducing the original DUSt3R results using the official codebase and obtain slightly worse results than reported in [75]. Using the same codebase, Pow3R with RGB only is able to improve over these and match the original results. Interestingly, adding ground-truth intrinsics results in a significant boost (19% rel. improvement), but adding the camera poses instead slightly degrades the results (-4%). However, using both intrinsics and pose results in a major improvement (36% rel. improvement), demonstrating a synergetic effect between both modalities. We hypothesize that the pose is of little value for easy cases like DTU, and that it does not help the model to properly resolve the focals.", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1077", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/abf11af29b525001d32d11e8e4a069972d86ef1e92d19f080e9269fb9e9e163d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Multi-view depth evaluation: Pow3R with both pose and intrinsics performs better than DUSt3R. More comparisons with the state of the art can be found in the supplementary material. (Parentheses) denote training on data from the same domain." + ], + "table_body": "
MethodGT rangeKITTIScanNetETH3DDTUT&TAverage
rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑
COLMAP [59, 60] (K+RT)×12.058.214.634.216.455.10.796.52.795.09.367.8
COLMAP Dense [59, 60] (K+RT)×26.952.738.022.589.823.220.869.325.776.440.248.8
MVSNet [81] (K+RT)✓18.630.722.720.921.635.6(1.8)(86.7)6.574.614.249.7
Vis-MVSNet [86] (K+RT)✓9.555.48.933.510.843.3(1.8)(87.4)4.187.27.061.4
MVS-Former++ [9] (K+RT)✓29.215.215.221.921.432.5(1.2)(91.9)7.671.514.946.6
CER-MVS [40] (K+RT)×14.332.221.124.311.747.54.171.36.482.111.551.5
DUSt3R [75]×5.449.5(3.1)(71.8)3.076.03.968.63.375.13.768.2
Pow3R×5.745.7(3.2)(68.8)3.074.73.074.33.376.63.668.0
Pow3R w/ RT×5.745.8(3.2)(69.7)2.975.63.371.63.277.93.768.1
Pow3R w/K×5.348.3(3.1)(70.8)2.976.01.689.93.277.33.272.5
Pow3R w/K+RT×5.348.7(3.1)(71.4)2.877.11.591.13.278.23.273.3
", + "bbox": [ + 93, + 88, + 620, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/86d00f68540f876391859338965f036f8644d4d865cecfccc47ccc2b1ba8f7a2.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. MVS results with the accuracy, completeness and their average (in mm) on the DTU dataset. $\\mathrm{DUSt3R}^{\\dagger}$ means results from the original paper, and 'repr.' means reproduced with the official public codebase and checkpoint." + ], + "table_body": "
MethodAcc.↓Comp.↓Overall↓
DUSt3R†[75]2.6770.8051.741
DUSt3R (repr.)2.1911.5981.894
Pow3R2.1161.3701.743
Pow3Rw/ K1.7221.1191.420
Pow3Rw/ Rt2.2051.4291.817
Pow3Rw/ K+RT1.3840.8461.115
", + "bbox": [ + 632, + 90, + 898, + 184 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Multi-View Pose estimation", + "text_level": 1, + "bbox": [ + 89, + 287, + 338, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then evaluate Pow3R for the task of multi-view pose prediction on the Co3Dv2 [51] and RealEstate10K [91] datasets. Co3Dv2 comprises over 37,000 videos across MS-COCO 51 categories with ground-truth camera derived from running COLMAP on 200 frames per video. RealEstate10k contains 80,000 YouTube video clips, containing indoor/outdoor scenes, with camera poses obtained via SLAM followed by Bundle Adjustment. For Co3Dv2, we follow the protocol of [73] and for RealEstate10K, we use its test set of 1,800 videos. For each sequence, we randomly select 10 frames and feed 45 pairs to Pow3R. We report the results in terms of RRA, RTA and mAA in Tab. 5.", + "bbox": [ + 89, + 309, + 483, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. Pow3R predicts $X^{2,2}$ , allowing to estimate the relative pose between 2 images in a single forward using Procrustes Alignment (see Sec. 3.3). In contrast, DUSt3R does not output $X^{2,2}$ and relies on PnP to obtain camera poses. Pow3R with PnP achieves slightly better results than DUSt3R with PnP. Using Procrustes alignment yields similar results, while being orders of magnitude faster than RANSAC, with a roughly $10 \\times$ speed-up overall when combining inference and pose estimation times. Finally, performance improves even further with known intrinsic, significantly outperforming all the state of the art.", + "bbox": [ + 89, + 492, + 483, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Architecture ablation", + "text_level": 1, + "bbox": [ + 89, + 667, + 290, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the impact of adding the $X^{2,2}$ prediction in Tab. 6. For these ablations, models are trained on a subset of Habitat and tested on focal, depth, and relative pose prediction, with results averaged over all combinations of auxiliary modalities. DUSt3R serves as a baseline, and DUSt3R-ft is finetuned with an additional $X^{2,2}$ head. This output can be used at test time to directly predict focals or relative poses, within a single forward pass, instead of having to forward both the original pair and its swapped version. We also report the performance, with and without predicting $X^{2,2}$ and/or using it. In all cases, predicting $X^{2,2}$ leads to slightly better performance while evaluations can be performed in a single forward pass.", + "bbox": [ + 89, + 688, + 482, + 885 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also explore different ways of conditioning auxil", + "bbox": [ + 109, + 885, + 480, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6f021b2cbb55be9ca7699c5aa3c75b7ccad0cc8cf78c13dd6b1efd3dd45e2042.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Impact of regressing $X^{2,2}$ . Metrics (in %) on Habitat are averaged over all combinations of input modalities. The 'with $X^{2,2}$ ' column denotes methods trained to predict it. 'Single forward' denotes evaluations of models based on $X^{2,2}$ prediction, i.e., in a single forward pass, instead of forwarding image pairs and their swapped versions. The top part shows the original DUSt3R, or finetune (-ft) with $X^{2,2}$ prediction." + ], + "table_body": "
Methodwith X2,2Single ForwardFocal acc@1.015Depth τ@1.03RRA@2 (Pro)RTA@2 (PnP)Avg.
Pro)PnP)
DUSt3R××36.885.469.272.449.857.3
DUSt3R-ft✓×38.979.171.774.452.959.2
✓✓38.378.573.174.456.359.2
Pow3R××66.085.781.983.868.673.3
✓×67.286.182.984.770.074.6
✓✓64.885.084.284.773.074.6
", + "bbox": [ + 516, + 286, + 903, + 377 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5ee9dcad4cb46b7e9155ffa7dd41da9a26a51d0be14e25791146f2225d811f8b.jpg", + "table_caption": [], + "table_footnote": [ + "Table 7. Architecture ablation. Results (in %), on Habitat, are averaged over all combinations of input modalities. We compare the injection architecture for different number of blocks with the 'embed' architecture in which the extra signals are added before the transformer blocks. 'Inject1' offers a good compromise between performance and number of extra parameters." + ], + "table_body": "
Arch.Extra ParamsFocal acc@1.015Depth τ@1.03RRA@2 (Pro) (PnP)RTA@2 (Pro) (PnP)Avg.
Embed16.0M64.784.784.184.572.873.877.4
Inject122.6M64.885.084.284.773.074.677.7
Inject442.3M65.985.184.885.273.775.478.3
", + "bbox": [ + 517, + 489, + 903, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ary information in Tab. 7. To incorporate additional priors, Pow3R injects those embeddings into the first block of the encoder or decoder, see Fig. 3. We compare this approach to another design choice, where the model embeds them before going through attention blocks. We have also tried to perform injection in more than one block, without significant improvement while the computational cost increases.", + "bbox": [ + 511, + 659, + 905, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 782, + 633, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented Pow3R, a novel approach to guide large 3D vision models, e.g. DUSt3R, with camera and scene priors in an implicit manner. Not only does it results in consistent gains whenever auxiliary information is available, but it also unlocks new capabilities such as high-resolution processing and sparse-to-dense depth completion.", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1078", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. IJCV, 2016. 7", + "[2] Youset I Abdel-Aziz, Hauck Michael Karara, and Michael Hauck. Direct linear transformation from comparator coordinates into object space coordinates in close range photogrammetry. Photogrammetric engineering & remote sensing, 2015. 3", + "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR, 2017. 3", + "[4] Daniel Barath, Dmytro Mishkin, Luca Cavalli, Paul-Edouard Sarlin, Petr Hruby, and Marc Pollefeys. Affineglue: Joint matching and robust estimation. arXiv preprint arXiv:2307.15381, 2023. 2", + "[5] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In ICCV, 2019. 2", + "[6] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In NeurIPS, 2021. 5", + "[7] Herbert Bay. Surf: Speeded up robust features. ECCV, 2006. 2", + "[8] Romain Brégier. Deep regression on manifolds: a 3D rotation case study. In 3DV, 2021. 5", + "[9] Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer++: Revealing the devil in transformer's details for multi-view stereo. In ICLR, 2024. 2, 8", + "[10] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, 2022. 2", + "[11] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In ECCV, 2018. 2", + "[12] David J Crandall, Andrew Owens, Noah Snavely, and Daniel P Huttenlocher. Sfm with mrfs: Discrete-continuous optimization for large-scale structure from motion. IEEE trans. PAMI, 2012. 2", + "[13] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In CVPR, 2017. 2", + "[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 7", + "[15] Timothee Darcet, Mathilde Caron, Piotr Bojanowski, Ishan Misra, Armand Joulin, Julien Mairal, Hervé Jégou, and Hugo Touvron. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 1", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4", + "[17] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 5", + "[18] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 2015. 2", + "[19] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In ICCV, 2015. 2", + "[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. *IJRR*, 2013. 6, 7", + "[21] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, 1988. 2", + "[22] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2, 5", + "[23] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In ICRA, 2021. 2", + "[24] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In ICCV, 2021. 2", + "[25] Nianjuan Jiang, Zhaopeng Cui, and Ping Tan. A global linear method for camera pose registration. In ICCV, 2013. 2", + "[26] Shengyu Jiang, Chen Wang, Yuhong Zhang, and Zhibin Wang. Completionformer: Efficient and accurate depth completion via hybrid attention with transformers. In CVPR, 2023. 1, 2", + "[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM trans. Graphics, 2023. 1", + "[28] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM trans. Graphics, 2017. 7", + "[29] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. IJCV, 2009. 5", + "[30] Vincent Leroy, Jean-Sébastien Franco, and Edmond Boyer. Volume sweeping: Learning photoconsistency for multi-view shape reconstruction. IJCV, 2021. 2", + "[31] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with MAST3R. In ECCV, 2024. 2, 3", + "[32] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 5", + "[33] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 3, 7", + "[34] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 7" + ], + "bbox": [ + 516, + 93, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1079", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In ICCV, 2023. 2", + "[36] David G Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 2004. 2", + "[37] Bin Luo and Edwin R. Hancock. Procrustes alignment with the em algorithm. In CAIP, 1999. 5", + "[38] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In ICRA, 2018. 2", + "[39] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In ICRA, 2019. 2", + "[40] Zeyu Ma, Zachary Teed, and Jia Deng. Multiview stereo with cascaded epipolar raft. In ECCV, 2022. 2, 8", + "[41] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5", + "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 1", + "[43] Danish Nazir, Marcus Liwicki, Didier Stricker, and Muhammad Zeshan Afzal. Semattnet: Towards attention-based semantic aware guided depth completion. arXiv preprint arXiv:2204.13635, 2022. 2", + "[44] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 2", + "[45] OpenAI. Chatgpt-4, 2024. 1", + "[46] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 1, 2", + "[47] Frank Plastria. The weiszfeld algorithm: Proof, amendments, and extensions. Foundations of Location Analysis, 2011. 5", + "[48] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deep lidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In CVPR, 2019. 2", + "[49] Alexander Raistrick, Lahav Lipson, Zeyu Ma, Lingjie Mei, Mingzhe Wang, Yiming Zuo, Karhan Kayan, Hongyu Wen, Beining Han, Yihan Wang, Alejandro Newell, Hei Law, Ankit Goyal, Kaiyu Yang, and Jia Deng. Infinite photorealistic worlds using procedural generation. In CVPR, 2023. 6", + "[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 4", + "[51] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 5, 8" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In CVPR, 2022. 1, 2", + "[53] Chris Rockwell, Justin Johnson, and David F Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 3DV, 2022. 3", + "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1", + "[55] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In ECCV, 2006. 2", + "[56] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In CVPR, 2019. 7", + "[57] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In CVPR, 2020. 2", + "[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 5", + "[59] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2, 8", + "[60] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 8", + "[61] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In CVPR, 2017.", + "[62] Philipp Schröppel, Jan Bechtold, Artemij Amiranashvili, and Thomas Brox. A benchmark and a baseline for robust multiview depth estimation. In 3DV, 2022. 7", + "[63] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6", + "[64] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2", + "[65] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 3", + "[66] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 2", + "[67] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5", + "[68] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE trans. Image Processing, 2020. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1080", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, 2024. 2", + "[70] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 2", + "[71] Wouter Van Gansbeke, Davy Neven, Bert De Brabandere, and Luc Van Gool. Sparse and noisy lidar completion with rgb guidance and uncertainty. In MVA, 2019. 2", + "[72] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. In 3DV, 2025. 3", + "[73] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In ICCV, 2023. 1, 3, 7, 8", + "[74] Shuzhe Wang, Juho Kannala, Marc Pollefeys, and Daniel Barath. Guiding local feature matching with surface curvature. In ICCV, 2023. 2", + "[75] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 3, 5, 7, 8", + "[76] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In ICCV, 2023. 2", + "[77] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Bregier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In ICCV, 2023. 1", + "[78] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 2", + "[79] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet++: Semantic assisted repetitive image guided network for depth completion. In ECCV, 2022. 2", + "[80] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, 2024. 2", + "[81] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In ECCV, 2018. 8", + "[82] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In CVPR, 2020. 5", + "[83] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 2", + "[84] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 5", + "[85] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In ECCV, 2016. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[86] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. In BMVC, 2020. 8", + "[87] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. In ICLR, 2025. 3", + "[88] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Rel-pose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 3, 7", + "[89] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In ICLR, 2024. 1, 3, 4, 7", + "[90] Yinda Zhang and Thomas Funkhouser. Deep depth completion of a single rgb-d image. In CVPR, 2018. 2", + "[91] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. SIGGRAPH, 2018. 8" + ], + "bbox": [ + 516, + 90, + 903, + 361 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1081", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/fb9435c5-f90b-40a7-b204-e0d114acb0c9_model.json b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/fb9435c5-f90b-40a7-b204-e0d114acb0c9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f22c7d6fbafac48fd737b3838f8a7df880410d93 --- /dev/null +++ b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/fb9435c5-f90b-40a7-b204-e0d114acb0c9_model.json @@ -0,0 +1,2664 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.265, + 0.131, + 0.734, + 0.175 + ], + "angle": 0, + "content": "Pow3R: Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.203, + 0.756, + 0.26 + ], + "angle": 0, + "content": "Wonbong Jang* Philippe Weinzaepfel† Vincent Leroy† \nLourdes Agapito* Jerome Revaud† \n*UCL †Naver Labs Europe" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.261, + 0.778, + 0.275 + ], + "angle": 0, + "content": "{ucabwja,1.agapito}@ucl.ac.uk firstname.lastname@naverlabs.com" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.293, + 0.277, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.392, + 0.276, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.293, + 0.58, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.293, + 0.906, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.486, + 0.908, + 0.527 + ], + "angle": 0, + "content": "Figure 1. Qualitative comparison of raw 3D reconstructions between DUSt3R and Pow3R for the image pair on the left. DUSt3R cannot leverage auxiliary information potentially available at test time, resulting in noisy and low-resolution estimates, whereas Pow3R allows to inject such priors at test time in a versatile manner, enabling new capabilities such as high-resolution processing as a direct by-product." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.539, + 0.327, + 0.555 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.569, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We present Pow3R, a novel large 3D vision regression model that is highly versatile in the input modalities it accepts. Unlike previous feed-forward models that lack any mechanism to exploit known camera or scene priors at test time, Pow3R incorporates any combination of auxiliary information such as intrinsics, relative pose, dense or sparse depth, alongside input images, within a single network. Building upon the recent DUSt3R paradigm, a transformer-based architecture that leverages powerful pre-training, our lightweight and versatile conditioning acts as additional guidance for the network to predict more accurate estimates when auxiliary information is available. During training we feed the model with random subsets of modalities at each iteration, which enables the model to operate under different levels of known priors at test time. This in turn opens up new capabilities, such as performing inference in native image resolution, or point-cloud completion. Our experiments on 3D reconstruction, depth completion, multi-view depth prediction, multi-view stereo, and multi-view pose estimation tasks yield state-of-the-art results and confirm the effectiveness of Pow3R at exploiting all available information. The project webpage is https://europe.naverlabs.com/pow3r." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.539, + 0.645, + 0.555 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Building general models for 3D perception that can unify different 3D vision tasks such as depth estimation, keypoint matching, dense reconstruction or camera pose prediction, remains a complex and exciting open challenge. Traditional Structure from Motion (SfM) approaches such as COLMAP [59] do not apply pre-learned priors, so each scene must be optimized independently. On the other hand, most learning-based approaches focus on specific 3D vision tasks such as depth completion [26, 52], monocular and stereo depth estimation [46], pose estimation [73, 89] or novel view synthesis [27, 42]. While these dedicated approaches have been successful to some extent, this trend contrasts with other domains such as natural language processing where a single model like ChatGPT [45], trained with a vast collection of data, solves a range of different tasks without requiring re-training. While some models like CroCo-v2, DINOv2 or Stable Diffusion [15, 54, 77], trained on large image datasets, have served as learned priors for other tasks, including 3D vision ones, they are still far from multi-purpose foundation models like ChatGPT is for NLP." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "DUSt3R [75] marked a significant step forward in regression-based 3D foundation models. DUSt3R performs" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1071" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.362 + ], + "angle": 0, + "content": "3D dense reconstruction from uncalibrated and unposed images, solving many downstream vision tasks at once without the need for fine-tuning. Its success can be attributed to two main factors: (i) its use of a pointmap representation, which provides flexibility over its output space and enables to regress camera parameters, poses and depth easily and (ii) its use of a transformer architecture that exploits powerful pretraining combined with a vast amount of 3D training data. Despite its huge success, DUSt3R [75] and its successor MASt3R [31] are extremely limited in terms of their input space, i.e. they only take RGB images. Yet, in practice, many real-world applications do provide extra modalities such that calibrated camera intrinsics, sparse or dense depth from RGB-D sensors or LIDAR, etc. This raises the question of how best to exploit additional input modalities, and whether this not just boosts performance but also enables new capabilities, such as obtaining the full depth from sparse point clouds of SfM or LiDAR." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.364, + 0.483, + 0.635 + ], + "angle": 0, + "content": "In this paper, we introduce Pow3R, a new 3D feedforward model that empowers DUSt3R with any subset of priors available at test time such as camera intrinsic, sparse or dense depth, or relative camera poses. Each modality is injected into the network in a lightweight fashion. To allow the model to operate under different conditions at test time, random subsets of input modalities are fed at each training iteration. As a result, we obtain a single model that performs on par with DUSt3R when no prior information is available but outperforms it when it exists. Our model also gains new capabilities as a by-product: for instance, the camera intrinsic input allows to process images whose principal point is far from the center, thus allowing to perform extreme cropping, e.g. for performing sliding window inference. As a side contribution, our model directly outputs the pointmaps of the second image in its coordinate system, allowing faster relative pose estimation. To sum up, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.635, + 0.483, + 0.68 + ], + "angle": 0, + "content": "- Pow3R is a holistic 3D geometric vision model capable of taking any subset (including none) of camera intrinsics, pose and depthmaps with corresponding input images." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.681, + 0.483, + 0.725 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate an important boost in performance over DUSt3R which cannot exploit priors, yielding state-of-the-art results in three benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.726, + 0.483, + 0.771 + ], + "angle": 0, + "content": "- By predicting the same pointmaps in two different camera coordinate systems, we can achieve more accurate relative pose, orders of magnitude faster." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.635, + 0.483, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.783, + 0.228, + 0.799 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Structure-from-Motion and MVS. SfM [12, 13, 22, 25, 59] builds 3D sparse maps by optimizing camera parameters using only RGB images. Traditional approaches consist of handcrafted pipelines to match keypoints [5, 7, 21, 36, 55], form pixel correspondences across multiple images, determine geometric relations, and perform bundle" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.364 + ], + "angle": 0, + "content": "adjustment to optimize camera locations. Recently, SfM has incorporated learning-based approaches leveraging neural networks as components [4, 10, 35, 57, 66, 70, 74, 85], but it remains brittle to outliers and challenging conditions. Likewise, multi-view stereo (MVS) aims to reconstruct dense 3D surface through handcrafted pipelines based on triangulation from multiple viewpoints [18, 19, 60]. Learning-based approaches have also been incorporated for MVS to reconstruct 3D from few images [24, 30, 44, 64, 78, 83], and more recently by adopting the RAFT architecture [40] or using generic Transformers [9]. In contrast, the recent DUSt3R framework [31, 75], upon which we build our approach, departs from these methods altogether by defining a simple and unified neural framework to 3D reconstruction based on the paradigm of pointmap regression. The idea is that, ideally, learning-based MVS should be capable of reconstructing 3D from a few images without strictly requiring the availability of camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.905, + 0.653 + ], + "angle": 0, + "content": "Guiding 3D. In this paper, we refer to guidance as the notion that a method can optionally take in 3D priors (typically camera parameters or sparse depth) to help with some prediction task. This can take different forms in the vast literature of 3D vision works. COLMAP [60], for instance, can take camera parameters for rectifying camera views; however, it is an optimization-based method, meaning that each scene has to be processed independently. Recently, UniDepth [46] is a monocular depth and intrinsic estimation method capable of optionally providing camera intrinsics as additional guide to generate depthmap, with an architecture specifically handcrafted for that purpose. Also, depth completion models [68, 69, 76, 80] such as CompletionFormer [26] take sparse depth input as guidance in order to inpaint the full depthmap, but are also typically capable of inferring depth from an image alone. Overall, however, there is currently no learning-based model to the best of our knowledge that guides its output by taking any subset of sparse depth, camera intrinsics and camera extrinsics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.905, + 0.883 + ], + "angle": 0, + "content": "Depth Completion. Given sparse depths and a corresponding RGB image, the task is to densify depth by propagating sparse measurements. CNN-based approaches were adopted by [38, 39], and [11] proposed convolutional spatial propagation network (CSPN). Multi-branch networks, which process RGB and sparse depth separately, were used in [23, 43, 48, 52, 68, 71, 79, 90]. CompletionFormer [26] leverages both CNN and ViT architectures to capture both global and local features. Most methods focus only on the specific depth completion task from a single image and partial depth as the input, often without incorporating camera intrinsics. Using the pointmap representation, Pow3R, based on a ViT architecture, can complete the depth map from both RGB images and sparse point clouds, while taking relative pose or camera intrinsics as auxiliary input." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Pose Estimation from Sparse views. Learning-based ap" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1072" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.483, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.21, + 0.484, + 0.295 + ], + "angle": 0, + "content": "Figure 2. Overview of the model architecture. Following DUSt3R [75], images are encoded then decoded with a ViT backbone into pointmaps, from which focals, depthmaps and relative pose can be extracted. Pow3R introduces optional inputs to guide the regression with prior knowledge about the camera intrinsics and depth (fed into the encoder) and the pose (into the decoder)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.314, + 0.483, + 0.435 + ], + "angle": 0, + "content": "proaches involve regressing relative pose, applying probabilistic methods, using energy-based or diffusion models [3, 33, 53, 73, 88]. RayDiffusion [89] models camera rays instead of camera parameters, applying diffusion on rays, then recovers the camera pose using DLT [2]. DUSt3R discovers camera pose from pointmaps using PnP, while Pow3R predicts pointmaps in both camera coordinate frames, allowing the use of Procrustes alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.484, + 0.679 + ], + "angle": 0, + "content": "DUSt3R and its developments. DUSt3R [75] does not require calibrated images for 3D reconstruction, can handle extreme viewpoint changes, and deals with both single and a few images within the same framework. Its pointmap representation eases the training by not having to deal with camera parameters directly. MAST3R [31] improves DUSt3R by also predicting dense local features, trained with an additional matching loss. Splatt3R [65] marries MAST3R with GS. Spann3R [72] predicts pointmaps in global coordinate and adopts an external spatial memory that tracks relevant 3D information. MonST3R [87] further trains DUSt3R on dynamic scenes and generates time-varying dynamic point clouds, as well as camera intrinsics and poses from a dynamic video. In this work, we take a different and complementary direction and instead improve the core capabilities of DUSt3R, which could benefit all aforementioned works." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.227, + 0.712 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We aim to train a network \\(\\mathcal{F}\\) that can take two input images \\(I^1, I^2 \\in \\mathbb{R}^{W \\times H \\times 3}\\) of a given static scene and any subset of auxiliary information \\(\\Omega \\subseteq \\{K_1, K_2, P_{1,2}, D^1, D^2\\}\\), in order to regress a 3D reconstruction of the scene. Here, \\(K_1, K_2 \\in \\mathbb{R}^{3 \\times 3}\\) are camera intrinsics, \\(P_{1,2} \\in \\mathbb{R}^{4 \\times 4}\\) denote the relative pose between the two cameras, and \\(D^1, D^2 \\in \\mathbb{R}^{W \\times H}\\) are depthmaps with associated masks \\(M^1, M^2 \\in \\{0,1\\}^{W \\times H}\\) specifying pixels with valid depth data (i.e. masks are possibly sparse). The network \\(\\mathcal{F}\\) is tasked to regress several pointmaps from which the camera intrinsic and extrinsic parameters as well as the dense depthmaps can be straightforwardly extracted for both im" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.68, + 0.107 + ], + "angle": 0, + "content": "ages as described below." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.109, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Pointmaps. For each pixel \\((i,j)\\) in an image \\(I\\), we assume there exists a corresponding single 3D point \\(X_{i,j}\\) where \\(X \\in \\mathbb{R}^{W \\times H \\times 3}\\) is a pointmap. Given camera intrinsics \\(K\\) and a depthmap \\(D\\), we can compute \\(X_{i,j} = K^{-1}[iD_{i,j},jD_{i,j},D_{i,j}]\\) in the camera coordinate system. In the following, we denote as \\(X^{n,m}\\) the pointmap of image \\(I^n\\) expressed in the coordinate system of camera \\(I^m\\). To swap the coordinate system from camera \\(I^k\\) to camera \\(I^m\\), we have \\(X^{n,k} = P_{m,k}X^{n,m}\\) where \\(P_{m,k} = P_kP_m^{-1}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.256, + 0.707, + 0.27 + ], + "angle": 0, + "content": "3.1. Overall architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.907, + 0.399 + ], + "angle": 0, + "content": "We follow the recently proposed DUSt3R framework [75]. DUSt3R is a breakthrough foundation model based on vision transformers (ViT) that can regress two 3D pointmaps \\( X^{1,1} \\), \\( X^{2,1} \\) given solely two unposed and uncalibrated input images. Nevertheless, DUSt3R's inability to leverage potentially available information about cameras or depth limits its practical use. To remedy these drawbacks, we enhance the original DUSt3R network \\( \\mathcal{F} \\) in two ways." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.909, + 0.551 + ], + "angle": 0, + "content": "First, we introduce specific modules to incorporate any subset of extra information such as camera intrinsics, camera poses and depthmaps seamlessly into DUSt3R. Second, we predict an additional pointmap \\( X^{2,2} \\), which represents the pointmap of image \\( I^2 \\) in its own coordinate system. Predicting three pointmaps offers further capabilities, namely the possibility of recovering all information about both cameras in a single forward pass, as explained in Sec. 3.3. We first describe the overall architecture, see Fig. 2, and then detail the dedicated modules to inject auxiliary information." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.553, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Encoder. We use a shared ViT encoder to encode both images independently. In addition to \\(I\\), the encoder can receive auxiliary information about intrinsics \\(K\\) and depth \\(D\\) for each image. For the two input images \\(I^1, I^2\\) and their respective auxiliary information \\(\\Omega_1 \\in \\sigma(\\{K_1, D_1\\}), \\Omega_2 \\in \\sigma(\\{K_2, D_2\\})\\), where \\(\\sigma\\) denotes the set of all subsets, the encoder processes the information in a Siamese manner:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.67, + 0.907, + 0.689 + ], + "angle": 0, + "content": "\\[\nF ^ {1} = \\operatorname {E n c o d e r} \\left(I ^ {1}, \\Omega_ {1}\\right), F ^ {2} = \\operatorname {E n c o d e r} \\left(I ^ {2}, \\Omega_ {2}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Decoder. Likewise, the network has two decoders, each with its corresponding head, one predicting \\( X^{1,1} \\) and the other one estimating \\( X^{2,1} \\) and \\( X^{2,2} \\). Both decoders communicate via cross-attention between their own tokens and the outputs of the previous block of the other decoder. Each decoder may receive the relative pose \\( P_{12} \\) as additional input or not. Assume we provide the auxiliary information \\( \\Omega_D \\in \\sigma(\\{P_{12}\\}) \\) at the \\( i \\)-th block of both decoders:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.835, + 0.905, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} G _ {i} ^ {1} = \\operatorname {D e c o d e r B l o c k} _ {i} ^ {1} \\left(G _ {i - 1} ^ {1}, G _ {i - 1} ^ {2}, \\Omega_ {D}\\right), \\\\ G _ {i} ^ {2} = \\operatorname {D e c o d e r B l o c k} _ {i} ^ {2} \\left(G _ {i - 1} ^ {2}, G _ {i - 1} ^ {1}, \\Omega_ {D}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "After \\(B\\) decoder blocks in each branch, the head regresses" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1073" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.459, + 0.108 + ], + "angle": 0, + "content": "several pointmaps and their associated confidence map:" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.117, + 0.4, + 0.136 + ], + "angle": 0, + "content": "\\[\nX ^ {1, 1}, C ^ {1, 1} = \\operatorname {H e a d} ^ {1} \\left(G _ {B} ^ {1}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.138, + 0.483, + 0.158 + ], + "angle": 0, + "content": "\\[\nX ^ {2, 1}, X ^ {2, 2}, C ^ {2, 1}, C ^ {2, 2} = \\operatorname {H e a d} ^ {2} \\left(G _ {B} ^ {2}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.167, + 0.483, + 0.213 + ], + "angle": 0, + "content": "Unlike DUSt3R which uses a DPT head [50], here we employ a linear head for both \\(\\mathrm{Head}^1\\) and \\(\\mathrm{Head}^2\\) as it is more efficient and performs similarly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.484, + 0.397 + ], + "angle": 0, + "content": "3D Regression Loss. Following DUSt3R, we supervise the model to minimize the distance between ground-truth and predicted pointmaps in a scale-invariant manner, allowing the model to train on multiple datasets with various scales. The regression loss between predicted and ground-truth pointmaps (resp. \\(X^{n,m}\\) and \\(\\hat{X}^{n,m}\\)) at pixel \\((i,j)\\) is defined as \\(\\mathcal{L}_{i,j}^{\\mathrm{gr}}(n,m) = \\| X_{i,j}^{n,m} / z_m - \\hat{X}_{i,j}^{n,m} / \\hat{z}_m\\|\\), where \\(z_{m},\\hat{z}_{m}\\) serve as scale normalizer. That is, \\(z_{m}\\) is the average norm of all valid 3D points expressed in coordinate system of image \\(I^{m}\\), i.e. \\(z_{1} = \\mathrm{norm}(X^{1,1}\\cup X^{2,1})\\), \\(z_{2} = \\mathrm{norm}(X^{2,2})\\) and likewise for \\(\\hat{z}_1,\\hat{z}_2\\), with \\(\\mathrm{norm}(X) = \\mathrm{mean}(\\| X_{i,j}\\| |i,j\\in \\mathcal{D}_X)\\)) and \\(\\mathcal{D}_X\\) the set of valid pixels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.473 + ], + "angle": 0, + "content": "Confidence-aware loss. The model jointly learns to predict a confidence level \\( C_{i,j}^{n,m} \\) of each pixel \\( (i,j) \\). The confidence-aware regression loss for a given pointmap \\( X^{n,m} \\) can be expressed as the 3D regression loss \\( \\mathcal{L}^{\\mathrm{reg}} \\) weighted by the confidence map:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.484, + 0.483, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {c o n f}} (n, m) = \\sum_ {i, j \\in \\mathcal {D}} C _ {i, j} ^ {n, m} \\mathcal {L} _ {i, j} ^ {\\text {r e g r}} (n, m) - \\alpha \\log C _ {i, j} ^ {n, m}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.573 + ], + "angle": 0, + "content": "with \\(\\alpha = 0.2\\). This loss penalizes the model less when the prediction is not accurate on harder areas, encouraging the model to extrapolate. The final loss is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.584, + 0.483, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} ^ {\\text {c o n f}} (1, 1) + \\mathcal {L} ^ {\\text {c o n f}} (2, 1) + \\beta \\mathcal {L} ^ {\\text {c o n f}} (2, 2), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.613, + 0.436, + 0.629 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a hyper-parameter typically set to \\(\\beta = 1\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.363, + 0.654 + ], + "angle": 0, + "content": "3.2. Adding Versatile Conditioning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.765 + ], + "angle": 0, + "content": "The knowledge of auxiliary information can significantly enhance 3D predictions at test time. Our model leverages up to five different modalities, which include two intrinsics \\( K_{1}, K_{2} \\), two depthmaps \\( D^{1}, D^{2} \\) for each image and the relative pose \\( P_{12} \\). To condition the output on it, we first embed the auxiliary information using dedicated MLPs and then inject these embeddings at different points in the pipeline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We discuss different options in this regard. In the first strategy, denoted as 'embed', we simply add the auxiliary embeddings to the token embeddings before the first transformer block, as would be done for standard positional embeddings [16]. In the second strategy, denoted as 'inject- \\( n \\) ' we instead insert dedicated MLPs for each modality in a subset of \\( n \\) transformer blocks, as shown in Fig. 3. Our experiments show that 'inject-1' performs slightly better than 'embed' and similarly with 'inject- \\( n \\) ' where \\( n > 1 \\), see" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.091, + 0.877, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.192, + 0.878, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.279, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Figure 3. Architecture for injecting auxiliary modalities. Top: injection of optional intrinsics and depth into the encoder. Intrinsic s are encoded into ray patches, sparse depth is patchified. Each of these modalities goes into a block-specific MLP and are tokenwise added in the middle of the transformer block. Bottom: injection of optional relative pose into the decoder. The relative pose is fed to a first embedding layer followed by an MLP. This token is added to the CLS token of the decoder after the self-attention (SA) and cross-attention (CA), but before the MLP. Experiments show that injection in the first block only suffices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.906, + 0.463 + ], + "angle": 0, + "content": "Sec. 4.5. We now describe how to compute the initial embeddings for each specific modality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.906, + 0.601 + ], + "angle": 0, + "content": "Intrinsics. Following [89], we generate camera rays from the intrinsic matrix \\( K \\in \\mathbb{R}^{3 \\times 3} \\), thereby establishing a direct correspondence between RGB pixels and rays. The ray at pixel location \\( (i,j) \\) is computed as \\( K^{-1}[i,j,1] \\) and encodes the viewing direction of that pixel w.r.t. the current camera pose. This allows us to potentially process non-centered crops and hence to perform inference in higher image resolutions, see Sec. 3.3. Similarly to RGB inputs, we patchify and embed dense rays and provide them to the encoder." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.604, + 0.906, + 0.71 + ], + "angle": 0, + "content": "Depthmaps / Point Clouds. Given a depthmap \\( D \\) and its sparsity mask \\( M \\), we first normalize \\( D' = D / \\mathrm{norm}(D) \\) to handle any depth ranges at train and test time. As for RGB and rays, we patchify the stacked maps \\( [D', M] \\in \\mathbb{R}^{W \\times H \\times 2} \\) and compute patch embeddings, which are then fed to the encoder. By jointly patchifying the depth and its valid mask we can handle any level of sparsity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.906, + 0.817 + ], + "angle": 0, + "content": "Camera Pose. Given the relative pose \\( P_{12} = [R_{12}|t_{12}] \\), we normalize the translation scale as \\( t_{12}' = t_{12} / \\|t_{12}\\| \\) since our output is unscaled. Unlike depthmaps or camera intrinsics, the camera pose cannot be expressed as a dense pixel map. Rather, camera pose affects the whole pixels between two images, so the embedding is instead added to the global CLS token of both decoders." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.831, + 0.696, + 0.845 + ], + "angle": 0, + "content": "3.3. Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Depthmaps. In the pointmap representation, the \\(z\\)-axis of \\(X^{1,1}, X^{2,2}\\) directly corresponds to the depthmaps of the first and second image, respectively. Note that DUSt3R" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1074" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.387, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.089, + 0.767, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.089, + 0.905, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.908, + 0.292 + ], + "angle": 0, + "content": "Figure 4. Native resolution example. Methods like DUSt3R require centered principal point and would thus downscale the image to its training resolution, leading to blurry prediction (top). With camera intrinsics input, we can process any crop in the image leading to native resolution prediction (bottom row), for example by dividing the image into 4 crops, that are independently process as the first input image, while the second input image of the network is set to the downscaled image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.483, + 0.34 + ], + "angle": 0, + "content": "only outputs \\( X^{1,1} \\), hence it requires two forward passes of the pairs \\( (I^1, I^2) \\) and \\( (I^2, I^1) \\) to extract both depthmaps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.483, + 0.599 + ], + "angle": 0, + "content": "High-resolution processing. A limitation of DUSt3R is that it is trained for a specific resolution and does not generalize to larger resolutions. A simple fix would be to include higher resolutions during training, but (i) many datasets do not provide high-resolution ground-truths, and (ii) the training cost could become prohibitive. Pow3R, in contrast, can handle crops natively given camera intrinsic's of the crop, as these provide the crop position information (i.e. via focal length and principal point). We can thus perform prediction in a sliding window fashion, yielding predictions matching any target resolution by simple stitching. Comparison between standard and higher resolution processing are shown in Fig. 4. Note that prediction for each crop may have a different scale, by design, and cannot be stitched directly. We find that simply computing the median scale factor in overlapping areas is sufficient, and we blend the overlapping regions based on confidence without further post-processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.691 + ], + "angle": 0, + "content": "Focal Estimation. Similar to DUSt3R, we can recover fccals for both input images from pointmap \\( X^{1,1} \\) and \\( X^{2,2} \\) with the robust Weiszfeld fast iterative solver [47]. Similar to depthmap prediction, DUSt3R needs to infer \\( (I_2, I_1) \\) in order to compute the focal of the second image, while Pow3R predicts them in a single pass." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.483, + 0.762 + ], + "angle": 0, + "content": "Relative Pose Estimation. Pow3R predicts the relative pose directly by Procrustes alignment [8, 37] to get the scaled relative pose \\( P^{*} = \\left[R^{*}|t^{*}\\right] \\) between \\( X^{2,2} \\) and \\( X^{2,1} \\), as it predicts the pointmaps of the second image in two different camera coordinates." + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.769, + 0.483, + 0.801 + ], + "angle": 0, + "content": "\\[\nR ^ {*}, t ^ {*} = \\arg \\min _ {\\sigma , R, t} \\sum_ {i, j} \\sqrt {C _ {i , j} ^ {2 , 2} C _ {i , j} ^ {2 , 1}} \\left\\| \\sigma \\left(R X _ {i, j} ^ {2, 2} + t\\right) - X _ {i, j} ^ {2, 1} \\right\\| ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.809, + 0.484, + 0.869 + ], + "angle": 0, + "content": "Procrustes alignment is known to be sensitive to noise and outliers; however, in our case, it performs similar to RANSAC [17] with PnP [22, 29], while being an order of magnitude faster (see Sec. 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Global alignment. The network \\(\\mathcal{F}\\) predicts pointmaps for image pairs. To align all predictions in the same world coor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.4 + ], + "angle": 0, + "content": "dinate system, we resort to the global alignment algorithm from DUSt3R [75] that minimizes a global energy function to find per-camera intrinsics, depthmaps and poses that are consistent with all the pairwise predictions. Results of the optimization are global scene point-clouds, which can for instance serve for multi-view stereo estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.411, + 0.702, + 0.427 + ], + "angle": 0, + "content": "3.4. Training Procedure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.569 + ], + "angle": 0, + "content": "During training, we feed the network with annotated image associated with random subsets of auxiliary information, the goal being for it to learn to handle any situation at test time. For each pair, we first chose a random number \\( m \\) of modalities with uniform probability, and then randomly select the \\( m \\) modalities likewise. Depthmaps are randomly sparsified. When giving intrinsics, we also perform aggressive non-centered cropping with \\( 50\\% \\) probability, so that the network learns to perform high-resolution inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.909, + 0.738 + ], + "angle": 0, + "content": "Training data and recipe. We follow the training recipe from DUSt3R [75] in terms of training data and protocol. Namely, we use the provided pairs for 8 datasets: Habitat [58], MegaDepth [32], ARKitScenes [6], Static Scenes 3D [41], BlendedMVS [82], ScanNet++ [84], Co3Dv2 [51] and Waymo [67], resulting in a total of 8.5M image pairs. These datasets include indoor and outdoor scenes, as well as real and synthetic ones. To train the model, we follow [75] and first train the network with a resolution of 224px for 3 days on 8 A100 GPUs, and then finetune it in 512px with variable aspect-ratio for 2 more days." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.752, + 0.75, + 0.77 + ], + "angle": 0, + "content": "4. Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.778, + 0.907, + 0.839 + ], + "angle": 0, + "content": "We first study the impact of auxiliary information guidance (Sec. 4.1). We then experiment across many 3D vision tasks (Secs. 4.2 to 4.4) without retraining, i.e. mostly zero-shot settings. We finally present some ablations in Sec. 4.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "DUSt3R baseline. We point out that our overall training data, protocols and architecture are exactly the same as DUSt3R, except for the additional head that predict \\( X^{2,2} \\) (\\(+0.1\\%\\) parameters) and for the modules used for injecting" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1075" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.481, + 0.249 + ], + "angle": 0, + "content": "
aux. modalitiesfocal acc@1.015depth τ@1.03rel. pose
K1K2D1D2RTRRA@2°RTA@2°
DUSt3R×××××36.677.669.249.7
Pow3R×××××39.479.471.953.8
✓××××75.4(+36.0)80.6 (+1.2)77.1 (+5.3)62.8 (+8.9)
×✓×××67.3(+27.9)80.3 (+0.9)74.5 (+2.6)59.1 (+5.2)
✓✓×××98.0(+58.6)81.4 (+2.0)80.3 (+8.5)74.2(+20.3)
××✓××48.9 (+9.5)89.1 (+9.7)82.5(+10.6)64.9(+11.0)
×××✓×49.5(+10.1)91.0(+11.7)83.4(+11.5)64.8(+10.9)
××✓✓×58.2(+18.9)95.4(+16.0)89.6(+17.7)77.6(+23.8)
××××✓48.6 (+9.2)81.6 (+2.2)92.3(+20.5)77.0(+23.1)
✓✓✓✓×99.2(+59.8)95.4(+16.0)95.7(+23.9)94.3(+40.5)
✓✓××✓98.1(+58.7)82.9 (+3.6)95.1(+23.2)87.3(+33.5)
××✓✓✓68.6(+29.2)95.4(+16.0)98.1(+26.2)91.3(+37.5)
✓✓✓✓✓99.3(+59.9)95.4(+16.0)99.0(+27.2)98.1(+44.3)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.251, + 0.483, + 0.322 + ], + "angle": 0, + "content": "Table 1. Impact of guiding at test time for models trained at \\(224 \\times 224\\) resolution. We report performances on Habitat for DUSt3R, which cannot handle auxiliary modalities, and our single model with different sets of modalities; we show in green the absolute improvement w.r.t. the results without auxiliary modality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.483, + 0.365 + ], + "angle": 0, + "content": "the auxiliary information \\((+4\\%\\) parameters for the inject-1 variant that we use unless stated otherwise)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.483, + 0.518 + ], + "angle": 0, + "content": "Metrics. When evaluating focal lengths, we report the accuracy for a tight relative error threshold of \\(1.5\\%\\). For evaluating depthmap, we report the Absolute Relative Error (rel) and the Inlier Ratio \\((\\tau)\\) at a threshold of \\(3\\%\\). When evaluating poses, we opt for scale-invariant metrics and measure the relative rotation and translation accuracies (RRA and RTA resp.) of the relative pose at task-specific thresholds, and the mean Average Accuracy (mAA), defined as the area under the curve accuracy of the angular differences at min(RRA, RTA)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.529, + 0.358, + 0.545 + ], + "angle": 0, + "content": "4.1. Guiding the output prediction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Impact of guidance. We first study the impact of providing different subsets of auxiliary information in terms of downstream 3D vision performance. Namely, we report accuracies in Tab. 1 for focal, depth, and rotation prediction on the Habitat validation dataset. In the supplementary material, we also provide results on the Infinigen dataset [49], which is not included in our training set, and observe similar outcomes. We first note that in the absence of modalities, Pow3R performs slightly better than a DUSt3R baseline, even though Pow3R is trained with just RGB images during a fraction of its training. This shows that Pow3R does not lose its native strength for unconstrained regression despite having gained more capabilities." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Interestingly, we observe in Tab. 1 the impact of each modality for each prediction task, and the synergetic effects of interacting modalities. For instance, the model performs better for focal length estimation with depthmaps and/or relative pose, and Pow3R's depth prediction improves with intrinsics and/or relative pose. Moreover, relative pose prediction becomes more accurate with intrinsics and/or depthmap. Overall, the model's performance consistently improves when provided with additional camera or scene priors, demonstrating the effectiveness of Pow3Rin" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.905, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.207, + 0.905, + 0.235 + ], + "angle": 0, + "content": "Figure 5. Depth completion performance on NYUd for various input depth sparsity." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.244, + 0.905, + 0.336 + ], + "angle": 0, + "content": "
aux. mod.High-Res.KITTIT&TTime (sec)↓
KsRTrel↓τ↑rel↓τ↑
DUSt3R512×××5.449.53.375.10.13
Pow3R512✓✓×5.348.73.278.20.13
Pow3R512✓✓(n)7.534.43.968.00.48
Pow3R512✓✓✓4.653.52.582.30.40
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.338, + 0.905, + 0.393 + ], + "angle": 0, + "content": "Table 2. Multi-view depth estimation results in high-resolution using a sliding-window scheme and providing the crop intrinsics to Pow3R. (n): naively inputting high-resolution images without downsampling to Pow3R." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.905, + 0.455 + ], + "angle": 0, + "content": "handling any combination of priors in a unified fashion. We are not aware of any other approach that can seamlessly integrate all these priors in a learning-based method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.907, + 0.653 + ], + "angle": 0, + "content": "Depth completion. We also evaluate Pow3R on the depth completion task given RGB images with sparse depthmaps. We vary the sparsity (the ratio of pixels with known depth) and demonstrate its impact on the NYUv2 dataset [63] in Figure 5. Although Pow3R is not trained on the NYUv2 dataset, it still achieves state-of-the-art results for sparse depth completion. Our analysis indicates that the true performance of Pow3R on NYUv2 is likely underestimated due to numerous errors in the ground-truth depth annotations; further details are available in the supplementary material. On the KITTI [20] dataset, we obtain a mean absolute error below \\(30\\mathrm{cm}\\), which is competitive with the state of the art, considering that KITTI is not included in the training set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.906, + 0.853 + ], + "angle": 0, + "content": "High-Resolution processing. With the possibility of specifying camera intrinsics, new capabilities arise. Namely, Pow3R can perform high-resolution pointmap regression with a simple sliding-window approach. We process fixed-resolution crops of the original images under the guidance of both (i) the crop information, provided in term of camera intrinsics, and (ii) a second image, itself possibly down-scaled or cropped as well. More details are provided in Sec. 3.3 and Fig. 4. When applied, results significantly improve compared to processing the images at 512-pixel resolution, as shown in Tab. 2. We also compare with a naive baseline where we simply feed high-resolution images directly to Pow3R, but it dramatically fails." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Camera controllability. We study how controllable are the regression outputs on 10K Habitat test pairs given arbitrary camera parameters. For each test pair with ground-truth fo" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1076" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.287, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.091, + 0.48, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.206, + 0.285, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.207, + 0.48, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.32, + 0.483, + 0.417 + ], + "angle": 0, + "content": "Figure 6. Controllability study. We study how the model reacts when presented with false focal (top) and false relative pose (bottom). We plot the input vs. predicted camera parameters in the left column and the average confidence in the right column. Interestingly, the model seems to not take the input information at face value and instead makes an informed decision whether to use it or not, in which latter case the confidence becomes low." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.438, + 0.483, + 0.695 + ], + "angle": 0, + "content": "cal \\( f \\) and relative pose \\( P_{12} \\), we purposefully input a wrong focal (resp. relative pose) and measure the output focal \\( \\hat{f} \\) (resp. pose \\( \\hat{P}_{12} \\)) as predicted from the raw pointmaps. Results are presented in Fig. 6 in terms of focal ratio \\( \\hat{f} / f \\) and angle \\( d_{\\text{geodesic}}(P_{12}, \\hat{P}_{12}) \\). Interestingly, the model accepts to follow the guidance whenever it is close to the ground truth value (i.e. staying close to the identity transformation) but then refuses to follow it when it gets too different, showing that the model is able to somehow 'judge' of the guidance's worth. In all cases, the output confidence clearly correlates with the agreement between the guidance and the ground truth. Overall, the model is thus not taking the auxiliary information at face value, and instead combines it with the other modalities (e.g. RGB) it in a more subtle and informed way than what could be achieved with handcrafted methods. This suggests that Pow3R would be robust to noise in the auxiliary information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.355, + 0.723 + ], + "angle": 0, + "content": "4.2. Multi-View Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Following RobustMVD [62], performances are measured on the KITTI [20], ScanNet [14], ETH3D [61], DTU [1], and Tanks and Temples [28] datasets. We report in Tab. 3 the Absolute Relative Error (rel) and Inlier Ratio \\((\\tau)\\) for each dataset when using (1) images alone, (2) images with relative poses, (3) images with camera intrinsics, and (4) images with both modalities, and provide results for more state-of-the-art methods in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Results. Pow3R again performs roughly on par with DUSt3R in the absence of auxiliary information (e.g. 3.73 average error across all datasets for DUSt3R-512 vs. 3.64" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.905, + 0.223 + ], + "angle": 0, + "content": "
MethodGT intrinsicsCo3Dv2RealEstate10K mAA(30)Speed (fps)
RRA@15RTA@15mAA(30)
Colmap+SG [56]✓36.127.325.345.2-
PixSfM [34]✓33.732.930.149.4-
RelPose [88]×57.1----
(a) PoseDiff [73]×80.579.866.548.0-
RelPose++ [33]×85.5----
RayDiff [89]×93.3----
DUSt3R (PnP) [75]×94.388.477.261.73.2
(b) Pow3R (PnP)×94.889.978.562.53.2
Pow3R (Pro)×94.690.378.166.330.9
Pow3R w/ K (Pro)✓95.092.182.272.530.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.224, + 0.905, + 0.307 + ], + "angle": 0, + "content": "Table 5. Multiview pose estimation on the CO3Dv2 and RealEstate10K. (a) denotes multiview methods, while (b) denotes pairwise methods. Pow3R with Procrustes alignment (Pro) performs better than DUSt3R (PnP) and other multi-view based pose estimation approaches, and even more so with known intrinsics, while being much faster. RealEstate10K is not in the training set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.509 + ], + "angle": 0, + "content": "for Pow3R-512). However, it is clear that providing auxiliary information consistently improves the depth estimates. In fact, Pow3R-512 with both pose and intrinsics achieves state-of-the-art performance on the ETH3D, DTU, and Tanks and Temples datasets, as well as on average across datasets. Interestingly, intrinsics tend to improve performance more than the GT pose, which alone seems to have a negligible impact. This might be due to the fact that camera pose primarily helps the prediction of \\( X^{2,1} \\), while the depthmap of this task is obtained from the \\( z \\) axis of \\( X^{1,1} \\) alone. Still, we observe a clear synergy in providing jointly intrinsics and relative pose." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.534, + 0.692, + 0.548 + ], + "angle": 0, + "content": "4.3. Multi-View Stereo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.698 + ], + "angle": 0, + "content": "We evaluate the quality of global 3D reconstructions on the DTU [1] benchmark in the presence and absence of auxiliary information. For each subset of available information from \\(\\{(K_1, K_2), P_{12}\\}\\), we extract pairwise predictions from image pairs and align them in a global coordinate system, as described in Sec 3.3. Performance is evaluated in terms of accuracy, which is the smallest Euclidean distance to the ground-truth, and completeness as the smallest distance to the reconstructed shape, with the overall average." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Results are presented in Tab. 4. We first attempt at reproducing the original DUSt3R results using the official codebase and obtain slightly worse results than reported in [75]. Using the same codebase, Pow3R with RGB only is able to improve over these and match the original results. Interestingly, adding ground-truth intrinsics results in a significant boost (19% rel. improvement), but adding the camera poses instead slightly degrades the results (-4%). However, using both intrinsics and pose results in a major improvement (36% rel. improvement), demonstrating a synergetic effect between both modalities. We hypothesize that the pose is of little value for easy cases like DTU, and that it does not help the model to properly resolve the focals." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1077" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.621, + 0.228 + ], + "angle": 0, + "content": "
MethodGT rangeKITTIScanNetETH3DDTUT&TAverage
rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑
COLMAP [59, 60] (K+RT)×12.058.214.634.216.455.10.796.52.795.09.367.8
COLMAP Dense [59, 60] (K+RT)×26.952.738.022.589.823.220.869.325.776.440.248.8
MVSNet [81] (K+RT)✓18.630.722.720.921.635.6(1.8)(86.7)6.574.614.249.7
Vis-MVSNet [86] (K+RT)✓9.555.48.933.510.843.3(1.8)(87.4)4.187.27.061.4
MVS-Former++ [9] (K+RT)✓29.215.215.221.921.432.5(1.2)(91.9)7.671.514.946.6
CER-MVS [40] (K+RT)×14.332.221.124.311.747.54.171.36.482.111.551.5
DUSt3R [75]×5.449.5(3.1)(71.8)3.076.03.968.63.375.13.768.2
Pow3R×5.745.7(3.2)(68.8)3.074.73.074.33.376.63.668.0
Pow3R w/ RT×5.745.8(3.2)(69.7)2.975.63.371.63.277.93.768.1
Pow3R w/K×5.348.3(3.1)(70.8)2.976.01.689.93.277.33.272.5
Pow3R w/K+RT×5.348.7(3.1)(71.4)2.877.11.591.13.278.23.273.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.231, + 0.622, + 0.273 + ], + "angle": 0, + "content": "Table 3. Multi-view depth evaluation: Pow3R with both pose and intrinsics performs better than DUSt3R. More comparisons with the state of the art can be found in the supplementary material. (Parentheses) denote training on data from the same domain." + }, + { + "type": "table", + "bbox": [ + 0.633, + 0.092, + 0.899, + 0.185 + ], + "angle": 0, + "content": "
MethodAcc.↓Comp.↓Overall↓
DUSt3R†[75]2.6770.8051.741
DUSt3R (repr.)2.1911.5981.894
Pow3R2.1161.3701.743
Pow3Rw/ K1.7221.1191.420
Pow3Rw/ Rt2.2051.4291.817
Pow3Rw/ K+RT1.3840.8461.115
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.628, + 0.187, + 0.902, + 0.27 + ], + "angle": 0, + "content": "Table 4. MVS results with the accuracy, completeness and their average (in mm) on the DTU dataset. \\(\\mathrm{DUSt3R}^{\\dagger}\\) means results from the original paper, and 'repr.' means reproduced with the official public codebase and checkpoint." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.289, + 0.339, + 0.303 + ], + "angle": 0, + "content": "4.4. Multi-View Pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.492 + ], + "angle": 0, + "content": "We then evaluate Pow3R for the task of multi-view pose prediction on the Co3Dv2 [51] and RealEstate10K [91] datasets. Co3Dv2 comprises over 37,000 videos across MS-COCO 51 categories with ground-truth camera derived from running COLMAP on 200 frames per video. RealEstate10k contains 80,000 YouTube video clips, containing indoor/outdoor scenes, with camera poses obtained via SLAM followed by Bundle Adjustment. For Co3Dv2, we follow the protocol of [73] and for RealEstate10K, we use its test set of 1,800 videos. For each sequence, we randomly select 10 frames and feed 45 pairs to Pow3R. We report the results in terms of RRA, RTA and mAA in Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.66 + ], + "angle": 0, + "content": "Results. Pow3R predicts \\( X^{2,2} \\), allowing to estimate the relative pose between 2 images in a single forward using Procrustes Alignment (see Sec. 3.3). In contrast, DUSt3R does not output \\( X^{2,2} \\) and relies on PnP to obtain camera poses. Pow3R with PnP achieves slightly better results than DUSt3R with PnP. Using Procrustes alignment yields similar results, while being orders of magnitude faster than RANSAC, with a roughly \\( 10 \\times \\) speed-up overall when combining inference and pose estimation times. Finally, performance improves even further with known intrinsic, significantly outperforming all the state of the art." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.292, + 0.682 + ], + "angle": 0, + "content": "4.5. Architecture ablation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.886 + ], + "angle": 0, + "content": "We evaluate the impact of adding the \\(X^{2,2}\\) prediction in Tab. 6. For these ablations, models are trained on a subset of Habitat and tested on focal, depth, and relative pose prediction, with results averaged over all combinations of auxiliary modalities. DUSt3R serves as a baseline, and DUSt3R-ft is finetuned with an additional \\(X^{2,2}\\) head. This output can be used at test time to directly predict focals or relative poses, within a single forward pass, instead of having to forward both the original pair and its swapped version. We also report the performance, with and without predicting \\(X^{2,2}\\) and/or using it. In all cases, predicting \\(X^{2,2}\\) leads to slightly better performance while evaluations can be performed in a single forward pass." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "We also explore different ways of conditioning auxil" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.287, + 0.905, + 0.378 + ], + "angle": 0, + "content": "
Methodwith X2,2Single ForwardFocal acc@1.015Depth τ@1.03RRA@2 (Pro)RTA@2 (PnP)Avg.
Pro)PnP)
DUSt3R××36.885.469.272.449.857.3
DUSt3R-ft✓×38.979.171.774.452.959.2
✓✓38.378.573.174.456.359.2
Pow3R××66.085.781.983.868.673.3
✓×67.286.182.984.770.074.6
✓✓64.885.084.284.773.074.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.38, + 0.905, + 0.477 + ], + "angle": 0, + "content": "Table 6. Impact of regressing \\( X^{2,2} \\). Metrics (in %) on Habitat are averaged over all combinations of input modalities. The 'with \\( X^{2,2} \\)' column denotes methods trained to predict it. 'Single forward' denotes evaluations of models based on \\( X^{2,2} \\) prediction, i.e., in a single forward pass, instead of forwarding image pairs and their swapped versions. The top part shows the original DUSt3R, or finetune (-ft) with \\( X^{2,2} \\) prediction." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.556 + ], + "angle": 0, + "content": "
Arch.Extra ParamsFocal acc@1.015Depth τ@1.03RRA@2 (Pro) (PnP)RTA@2 (Pro) (PnP)Avg.
Embed16.0M64.784.784.184.572.873.877.4
Inject122.6M64.885.084.284.773.074.677.7
Inject442.3M65.985.184.885.273.775.478.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.558, + 0.905, + 0.641 + ], + "angle": 0, + "content": "Table 7. Architecture ablation. Results (in %), on Habitat, are averaged over all combinations of input modalities. We compare the injection architecture for different number of blocks with the 'embed' architecture in which the extra signals are added before the transformer blocks. 'Inject1' offers a good compromise between performance and number of extra parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.766 + ], + "angle": 0, + "content": "ary information in Tab. 7. To incorporate additional priors, Pow3R injects those embeddings into the first block of the encoder or decoder, see Fig. 3. We compare this approach to another design choice, where the model embeds them before going through attention blocks. We have also tried to perform injection in more than one block, without significant improvement while the computational cost increases." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.784, + 0.634, + 0.799 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We have presented Pow3R, a novel approach to guide large 3D vision models, e.g. DUSt3R, with camera and scene priors in an implicit manner. Not only does it results in consistent gains whenever auxiliary information is available, but it also unlocks new capabilities such as high-resolution processing and sparse-to-dense depth completion." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1078" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. IJCV, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Youset I Abdel-Aziz, Hauck Michael Karara, and Michael Hauck. Direct linear transformation from comparator coordinates into object space coordinates in close range photogrammetry. Photogrammetric engineering & remote sensing, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.484, + 0.271 + ], + "angle": 0, + "content": "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[4] Daniel Barath, Dmytro Mishkin, Luca Cavalli, Paul-Edouard Sarlin, Petr Hruby, and Marc Pollefeys. Affineglue: Joint matching and robust estimation. arXiv preprint arXiv:2307.15381, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[5] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[6] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In NeurIPS, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[7] Herbert Bay. Surf: Speeded up robust features. ECCV, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[8] Romain Brégier. Deep regression on manifolds: a 3D rotation case study. In 3DV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[9] Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer++: Revealing the devil in transformer's details for multi-view stereo. In ICLR, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.546, + 0.483, + 0.6 + ], + "angle": 0, + "content": "[10] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.602, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[11] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[12] David J Crandall, Andrew Owens, Noah Snavely, and Daniel P Huttenlocher. Sfm with mrfs: Discrete-continuous optimization for large-scale structure from motion. IEEE trans. PAMI, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[13] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[15] Timothee Darcet, Mathilde Caron, Piotr Bojanowski, Ishan Misra, Armand Joulin, Julien Mairal, Hervé Jégou, and Hugo Touvron. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.906, + 0.12 + ], + "angle": 0, + "content": "worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.906, + 0.178 + ], + "angle": 0, + "content": "[17] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.18, + 0.906, + 0.221 + ], + "angle": 0, + "content": "[18] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.224, + 0.906, + 0.263 + ], + "angle": 0, + "content": "[19] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.267, + 0.906, + 0.307 + ], + "angle": 0, + "content": "[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. *IJRR*, 2013. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.311, + 0.906, + 0.338 + ], + "angle": 0, + "content": "[21] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.34, + 0.906, + 0.38 + ], + "angle": 0, + "content": "[22] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.383, + 0.906, + 0.424 + ], + "angle": 0, + "content": "[23] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In ICRA, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.426, + 0.906, + 0.466 + ], + "angle": 0, + "content": "[24] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.497 + ], + "angle": 0, + "content": "[25] Nianjuan Jiang, Zhaopeng Cui, and Ping Tan. A global linear method for camera pose registration. In ICCV, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.5, + 0.906, + 0.553 + ], + "angle": 0, + "content": "[26] Shengyu Jiang, Chen Wang, Yuhong Zhang, and Zhibin Wang. Completionformer: Efficient and accurate depth completion via hybrid attention with transformers. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.556, + 0.906, + 0.598 + ], + "angle": 0, + "content": "[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM trans. Graphics, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.6, + 0.906, + 0.641 + ], + "angle": 0, + "content": "[28] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM trans. Graphics, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.643, + 0.906, + 0.683 + ], + "angle": 0, + "content": "[29] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. IJCV, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.687, + 0.906, + 0.728 + ], + "angle": 0, + "content": "[30] Vincent Leroy, Jean-Sébastien Franco, and Edmond Boyer. Volume sweeping: Learning photoconsistency for multi-view shape reconstruction. IJCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.73, + 0.906, + 0.77 + ], + "angle": 0, + "content": "[31] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with MAST3R. In ECCV, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.773, + 0.906, + 0.813 + ], + "angle": 0, + "content": "[32] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[33] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[34] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1079" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[35] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[36] David G Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.165, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[37] Bin Luo and Edwin R. Hancock. Procrustes alignment with the em algorithm. In CAIP, 1999. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.194, + 0.483, + 0.235 + ], + "angle": 0, + "content": "[38] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In ICRA, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[39] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In ICRA, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.293, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[40] Zeyu Ma, Zachary Teed, and Jia Deng. Multiview stereo with cascaded epipolar raft. In ECCV, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.483, + 0.377 + ], + "angle": 0, + "content": "[41] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[43] Danish Nazir, Marcus Liwicki, Didier Stricker, and Muhammad Zeshan Afzal. Semattnet: Towards attention-based semantic aware guided depth completion. arXiv preprint arXiv:2204.13635, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[44] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.548, + 0.295, + 0.562 + ], + "angle": 0, + "content": "[45] OpenAI. Chatgpt-4, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.564, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[46] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[47] Frank Plastria. The weiszfeld algorithm: Proof, amendments, and extensions. Foundations of Location Analysis, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[48] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deep lidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[49] Alexander Raistrick, Lahav Lipson, Zeyu Ma, Lingjie Mei, Mingzhe Wang, Yiming Zuo, Karhan Kayan, Hongyu Wen, Beining Han, Yihan Wang, Alejandro Newell, Hei Law, Ankit Goyal, Kaiyu Yang, and Jia Deng. Infinite photorealistic worlds using procedural generation. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[51] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[52] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[53] Chris Rockwell, Justin Johnson, and David F Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 3DV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.18, + 0.905, + 0.221 + ], + "angle": 0, + "content": "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.224, + 0.905, + 0.251 + ], + "angle": 0, + "content": "[55] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In ECCV, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.254, + 0.905, + 0.295 + ], + "angle": 0, + "content": "[56] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In CVPR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.297, + 0.905, + 0.339 + ], + "angle": 0, + "content": "[57] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.341, + 0.905, + 0.395 + ], + "angle": 0, + "content": "[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.398, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[59] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[60] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.472, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[61] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.905, + 0.584 + ], + "angle": 0, + "content": "[62] Philipp Schröppel, Jan Bechtold, Artemij Amiranashvili, and Thomas Brox. A benchmark and a baseline for robust multiview depth estimation. In 3DV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[63] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[64] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[65] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[66] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[67] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[68] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE trans. Image Processing, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1080" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[69] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.164 + ], + "angle": 0, + "content": "[70] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.166, + 0.482, + 0.208 + ], + "angle": 0, + "content": "[71] Wouter Van Gansbeke, Davy Neven, Bert De Brabandere, and Luc Van Gool. Sparse and noisy lidar completion with rgb guidance and uncertainty. In MVA, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.21, + 0.482, + 0.238 + ], + "angle": 0, + "content": "[72] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. In 3DV, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.482, + 0.281 + ], + "angle": 0, + "content": "[73] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In ICCV, 2023. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.283, + 0.482, + 0.324 + ], + "angle": 0, + "content": "[74] Shuzhe Wang, Juho Kannala, Marc Pollefeys, and Daniel Barath. Guiding local feature matching with surface curvature. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.327, + 0.482, + 0.368 + ], + "angle": 0, + "content": "[75] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.37, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[76] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[77] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Bregier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.499, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[78] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.543, + 0.482, + 0.597 + ], + "angle": 0, + "content": "[79] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet++: Semantic assisted repetitive image guided network for depth completion. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[80] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.698 + ], + "angle": 0, + "content": "[81] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In ECCV, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[82] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[83] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[84] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[85] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In ECCV, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[86] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. In BMVC, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[87] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. In ICLR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[88] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Rel-pose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[89] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In ICLR, 2024. 1, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[90] Yinda Zhang and Thomas Funkhouser. Deep depth completion of a single rgb-d image. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[91] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. SIGGRAPH, 2018. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1081" + } + ] +] \ No newline at end of file diff --git a/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/fb9435c5-f90b-40a7-b204-e0d114acb0c9_origin.pdf b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/fb9435c5-f90b-40a7-b204-e0d114acb0c9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45f45178b879c961ee8b52c0769490ff377e4706 --- /dev/null +++ b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/fb9435c5-f90b-40a7-b204-e0d114acb0c9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ac8b156d9644691974daadd3866688485a842d00d651fc2f3ae4e171e3de20a +size 2416232 diff --git a/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/full.md b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..79ccfa5bb922d7193b07bf2cba5b46a89d753a0f --- /dev/null +++ b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/full.md @@ -0,0 +1,361 @@ +# Pow3R: Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors + +Wonbong Jang* Philippe Weinzaepfel† Vincent Leroy† +Lourdes Agapito* Jerome Revaud† +*UCL †Naver Labs Europe + +{ucabwja,1.agapito}@ucl.ac.uk firstname.lastname@naverlabs.com + +![](images/cd468f966acef602dc94daffba7df92ffce29633c66132b6a368cb37e4b2ad19.jpg) + +![](images/b7996d15fe1224731089515b787255ed643c7e38300242e17396ae4643c12fe8.jpg) + +![](images/908ee6513e718a96c7c64ef3920fb3a7b74c29bfbb673f221e9a03e007190970.jpg) +Figure 1. Qualitative comparison of raw 3D reconstructions between DUSt3R and Pow3R for the image pair on the left. DUSt3R cannot leverage auxiliary information potentially available at test time, resulting in noisy and low-resolution estimates, whereas Pow3R allows to inject such priors at test time in a versatile manner, enabling new capabilities such as high-resolution processing as a direct by-product. + +![](images/657a18594bd983f7c7c17f58b7385fd5ba70716fe8c587b43ae61d0c62f47d96.jpg) + +# Abstract + +We present Pow3R, a novel large 3D vision regression model that is highly versatile in the input modalities it accepts. Unlike previous feed-forward models that lack any mechanism to exploit known camera or scene priors at test time, Pow3R incorporates any combination of auxiliary information such as intrinsics, relative pose, dense or sparse depth, alongside input images, within a single network. Building upon the recent DUSt3R paradigm, a transformer-based architecture that leverages powerful pre-training, our lightweight and versatile conditioning acts as additional guidance for the network to predict more accurate estimates when auxiliary information is available. During training we feed the model with random subsets of modalities at each iteration, which enables the model to operate under different levels of known priors at test time. This in turn opens up new capabilities, such as performing inference in native image resolution, or point-cloud completion. Our experiments on 3D reconstruction, depth completion, multi-view depth prediction, multi-view stereo, and multi-view pose estimation tasks yield state-of-the-art results and confirm the effectiveness of Pow3R at exploiting all available information. The project webpage is https://europe.naverlabs.com/pow3r. + +# 1. Introduction + +Building general models for 3D perception that can unify different 3D vision tasks such as depth estimation, keypoint matching, dense reconstruction or camera pose prediction, remains a complex and exciting open challenge. Traditional Structure from Motion (SfM) approaches such as COLMAP [59] do not apply pre-learned priors, so each scene must be optimized independently. On the other hand, most learning-based approaches focus on specific 3D vision tasks such as depth completion [26, 52], monocular and stereo depth estimation [46], pose estimation [73, 89] or novel view synthesis [27, 42]. While these dedicated approaches have been successful to some extent, this trend contrasts with other domains such as natural language processing where a single model like ChatGPT [45], trained with a vast collection of data, solves a range of different tasks without requiring re-training. While some models like CroCo-v2, DINOv2 or Stable Diffusion [15, 54, 77], trained on large image datasets, have served as learned priors for other tasks, including 3D vision ones, they are still far from multi-purpose foundation models like ChatGPT is for NLP. + +DUSt3R [75] marked a significant step forward in regression-based 3D foundation models. DUSt3R performs + +3D dense reconstruction from uncalibrated and unposed images, solving many downstream vision tasks at once without the need for fine-tuning. Its success can be attributed to two main factors: (i) its use of a pointmap representation, which provides flexibility over its output space and enables to regress camera parameters, poses and depth easily and (ii) its use of a transformer architecture that exploits powerful pretraining combined with a vast amount of 3D training data. Despite its huge success, DUSt3R [75] and its successor MASt3R [31] are extremely limited in terms of their input space, i.e. they only take RGB images. Yet, in practice, many real-world applications do provide extra modalities such that calibrated camera intrinsics, sparse or dense depth from RGB-D sensors or LIDAR, etc. This raises the question of how best to exploit additional input modalities, and whether this not just boosts performance but also enables new capabilities, such as obtaining the full depth from sparse point clouds of SfM or LiDAR. + +In this paper, we introduce Pow3R, a new 3D feedforward model that empowers DUSt3R with any subset of priors available at test time such as camera intrinsic, sparse or dense depth, or relative camera poses. Each modality is injected into the network in a lightweight fashion. To allow the model to operate under different conditions at test time, random subsets of input modalities are fed at each training iteration. As a result, we obtain a single model that performs on par with DUSt3R when no prior information is available but outperforms it when it exists. Our model also gains new capabilities as a by-product: for instance, the camera intrinsic input allows to process images whose principal point is far from the center, thus allowing to perform extreme cropping, e.g. for performing sliding window inference. As a side contribution, our model directly outputs the pointmaps of the second image in its coordinate system, allowing faster relative pose estimation. To sum up, our contributions are: + +- Pow3R is a holistic 3D geometric vision model capable of taking any subset (including none) of camera intrinsics, pose and depthmaps with corresponding input images. +- Extensive experiments demonstrate an important boost in performance over DUSt3R which cannot exploit priors, yielding state-of-the-art results in three benchmarks. +- By predicting the same pointmaps in two different camera coordinate systems, we can achieve more accurate relative pose, orders of magnitude faster. + +# 2. Related work + +Structure-from-Motion and MVS. SfM [12, 13, 22, 25, 59] builds 3D sparse maps by optimizing camera parameters using only RGB images. Traditional approaches consist of handcrafted pipelines to match keypoints [5, 7, 21, 36, 55], form pixel correspondences across multiple images, determine geometric relations, and perform bundle + +adjustment to optimize camera locations. Recently, SfM has incorporated learning-based approaches leveraging neural networks as components [4, 10, 35, 57, 66, 70, 74, 85], but it remains brittle to outliers and challenging conditions. Likewise, multi-view stereo (MVS) aims to reconstruct dense 3D surface through handcrafted pipelines based on triangulation from multiple viewpoints [18, 19, 60]. Learning-based approaches have also been incorporated for MVS to reconstruct 3D from few images [24, 30, 44, 64, 78, 83], and more recently by adopting the RAFT architecture [40] or using generic Transformers [9]. In contrast, the recent DUSt3R framework [31, 75], upon which we build our approach, departs from these methods altogether by defining a simple and unified neural framework to 3D reconstruction based on the paradigm of pointmap regression. The idea is that, ideally, learning-based MVS should be capable of reconstructing 3D from a few images without strictly requiring the availability of camera parameters. + +Guiding 3D. In this paper, we refer to guidance as the notion that a method can optionally take in 3D priors (typically camera parameters or sparse depth) to help with some prediction task. This can take different forms in the vast literature of 3D vision works. COLMAP [60], for instance, can take camera parameters for rectifying camera views; however, it is an optimization-based method, meaning that each scene has to be processed independently. Recently, UniDepth [46] is a monocular depth and intrinsic estimation method capable of optionally providing camera intrinsics as additional guide to generate depthmap, with an architecture specifically handcrafted for that purpose. Also, depth completion models [68, 69, 76, 80] such as CompletionFormer [26] take sparse depth input as guidance in order to inpaint the full depthmap, but are also typically capable of inferring depth from an image alone. Overall, however, there is currently no learning-based model to the best of our knowledge that guides its output by taking any subset of sparse depth, camera intrinsics and camera extrinsics. + +Depth Completion. Given sparse depths and a corresponding RGB image, the task is to densify depth by propagating sparse measurements. CNN-based approaches were adopted by [38, 39], and [11] proposed convolutional spatial propagation network (CSPN). Multi-branch networks, which process RGB and sparse depth separately, were used in [23, 43, 48, 52, 68, 71, 79, 90]. CompletionFormer [26] leverages both CNN and ViT architectures to capture both global and local features. Most methods focus only on the specific depth completion task from a single image and partial depth as the input, often without incorporating camera intrinsics. Using the pointmap representation, Pow3R, based on a ViT architecture, can complete the depth map from both RGB images and sparse point clouds, while taking relative pose or camera intrinsics as auxiliary input. + +Pose Estimation from Sparse views. Learning-based ap + +![](images/6b7cba457c5a6465b3b0f24a09a752630003b667442bbfacaf8b712112300c24.jpg) +Figure 2. Overview of the model architecture. Following DUSt3R [75], images are encoded then decoded with a ViT backbone into pointmaps, from which focals, depthmaps and relative pose can be extracted. Pow3R introduces optional inputs to guide the regression with prior knowledge about the camera intrinsics and depth (fed into the encoder) and the pose (into the decoder). + +proaches involve regressing relative pose, applying probabilistic methods, using energy-based or diffusion models [3, 33, 53, 73, 88]. RayDiffusion [89] models camera rays instead of camera parameters, applying diffusion on rays, then recovers the camera pose using DLT [2]. DUSt3R discovers camera pose from pointmaps using PnP, while Pow3R predicts pointmaps in both camera coordinate frames, allowing the use of Procrustes alignment. + +DUSt3R and its developments. DUSt3R [75] does not require calibrated images for 3D reconstruction, can handle extreme viewpoint changes, and deals with both single and a few images within the same framework. Its pointmap representation eases the training by not having to deal with camera parameters directly. MAST3R [31] improves DUSt3R by also predicting dense local features, trained with an additional matching loss. Splatt3R [65] marries MAST3R with GS. Spann3R [72] predicts pointmaps in global coordinate and adopts an external spatial memory that tracks relevant 3D information. MonST3R [87] further trains DUSt3R on dynamic scenes and generates time-varying dynamic point clouds, as well as camera intrinsics and poses from a dynamic video. In this work, we take a different and complementary direction and instead improve the core capabilities of DUSt3R, which could benefit all aforementioned works. + +# 3. Methodology + +We aim to train a network $\mathcal{F}$ that can take two input images $I^1, I^2 \in \mathbb{R}^{W \times H \times 3}$ of a given static scene and any subset of auxiliary information $\Omega \subseteq \{K_1, K_2, P_{1,2}, D^1, D^2\}$ , in order to regress a 3D reconstruction of the scene. Here, $K_1, K_2 \in \mathbb{R}^{3 \times 3}$ are camera intrinsics, $P_{1,2} \in \mathbb{R}^{4 \times 4}$ denote the relative pose between the two cameras, and $D^1, D^2 \in \mathbb{R}^{W \times H}$ are depthmaps with associated masks $M^1, M^2 \in \{0,1\}^{W \times H}$ specifying pixels with valid depth data (i.e. masks are possibly sparse). The network $\mathcal{F}$ is tasked to regress several pointmaps from which the camera intrinsic and extrinsic parameters as well as the dense depthmaps can be straightforwardly extracted for both im + +ages as described below. + +Pointmaps. For each pixel $(i,j)$ in an image $I$ , we assume there exists a corresponding single 3D point $X_{i,j}$ where $X \in \mathbb{R}^{W \times H \times 3}$ is a pointmap. Given camera intrinsics $K$ and a depthmap $D$ , we can compute $X_{i,j} = K^{-1}[iD_{i,j},jD_{i,j},D_{i,j}]$ in the camera coordinate system. In the following, we denote as $X^{n,m}$ the pointmap of image $I^n$ expressed in the coordinate system of camera $I^m$ . To swap the coordinate system from camera $I^k$ to camera $I^m$ , we have $X^{n,k} = P_{m,k}X^{n,m}$ where $P_{m,k} = P_kP_m^{-1}$ . + +# 3.1. Overall architecture + +We follow the recently proposed DUSt3R framework [75]. DUSt3R is a breakthrough foundation model based on vision transformers (ViT) that can regress two 3D pointmaps $X^{1,1}$ , $X^{2,1}$ given solely two unposed and uncalibrated input images. Nevertheless, DUSt3R's inability to leverage potentially available information about cameras or depth limits its practical use. To remedy these drawbacks, we enhance the original DUSt3R network $\mathcal{F}$ in two ways. + +First, we introduce specific modules to incorporate any subset of extra information such as camera intrinsics, camera poses and depthmaps seamlessly into DUSt3R. Second, we predict an additional pointmap $X^{2,2}$ , which represents the pointmap of image $I^2$ in its own coordinate system. Predicting three pointmaps offers further capabilities, namely the possibility of recovering all information about both cameras in a single forward pass, as explained in Sec. 3.3. We first describe the overall architecture, see Fig. 2, and then detail the dedicated modules to inject auxiliary information. + +Encoder. We use a shared ViT encoder to encode both images independently. In addition to $I$ , the encoder can receive auxiliary information about intrinsics $K$ and depth $D$ for each image. For the two input images $I^1, I^2$ and their respective auxiliary information $\Omega_1 \in \sigma(\{K_1, D_1\}), \Omega_2 \in \sigma(\{K_2, D_2\})$ , where $\sigma$ denotes the set of all subsets, the encoder processes the information in a Siamese manner: + +$$ +F ^ {1} = \operatorname {E n c o d e r} \left(I ^ {1}, \Omega_ {1}\right), F ^ {2} = \operatorname {E n c o d e r} \left(I ^ {2}, \Omega_ {2}\right). \tag {1} +$$ + +Decoder. Likewise, the network has two decoders, each with its corresponding head, one predicting $X^{1,1}$ and the other one estimating $X^{2,1}$ and $X^{2,2}$ . Both decoders communicate via cross-attention between their own tokens and the outputs of the previous block of the other decoder. Each decoder may receive the relative pose $P_{12}$ as additional input or not. Assume we provide the auxiliary information $\Omega_D \in \sigma(\{P_{12}\})$ at the $i$ -th block of both decoders: + +$$ +\begin{array}{l} G _ {i} ^ {1} = \operatorname {D e c o d e r B l o c k} _ {i} ^ {1} \left(G _ {i - 1} ^ {1}, G _ {i - 1} ^ {2}, \Omega_ {D}\right), \\ G _ {i} ^ {2} = \operatorname {D e c o d e r B l o c k} _ {i} ^ {2} \left(G _ {i - 1} ^ {2}, G _ {i - 1} ^ {1}, \Omega_ {D}\right), \tag {2} \\ \end{array} +$$ + +After $B$ decoder blocks in each branch, the head regresses + +several pointmaps and their associated confidence map: + +$$ +X ^ {1, 1}, C ^ {1, 1} = \operatorname {H e a d} ^ {1} \left(G _ {B} ^ {1}\right), +$$ + +$$ +X ^ {2, 1}, X ^ {2, 2}, C ^ {2, 1}, C ^ {2, 2} = \operatorname {H e a d} ^ {2} \left(G _ {B} ^ {2}\right). \tag {3} +$$ + +Unlike DUSt3R which uses a DPT head [50], here we employ a linear head for both $\mathrm{Head}^1$ and $\mathrm{Head}^2$ as it is more efficient and performs similarly. + +3D Regression Loss. Following DUSt3R, we supervise the model to minimize the distance between ground-truth and predicted pointmaps in a scale-invariant manner, allowing the model to train on multiple datasets with various scales. The regression loss between predicted and ground-truth pointmaps (resp. $X^{n,m}$ and $\hat{X}^{n,m}$ ) at pixel $(i,j)$ is defined as $\mathcal{L}_{i,j}^{\mathrm{gr}}(n,m) = \| X_{i,j}^{n,m} / z_m - \hat{X}_{i,j}^{n,m} / \hat{z}_m\|$ , where $z_{m},\hat{z}_{m}$ serve as scale normalizer. That is, $z_{m}$ is the average norm of all valid 3D points expressed in coordinate system of image $I^{m}$ , i.e. $z_{1} = \mathrm{norm}(X^{1,1}\cup X^{2,1})$ , $z_{2} = \mathrm{norm}(X^{2,2})$ and likewise for $\hat{z}_1,\hat{z}_2$ , with $\mathrm{norm}(X) = \mathrm{mean}(\| X_{i,j}\| |i,j\in \mathcal{D}_X)$ ) and $\mathcal{D}_X$ the set of valid pixels. + +Confidence-aware loss. The model jointly learns to predict a confidence level $C_{i,j}^{n,m}$ of each pixel $(i,j)$ . The confidence-aware regression loss for a given pointmap $X^{n,m}$ can be expressed as the 3D regression loss $\mathcal{L}^{\mathrm{reg}}$ weighted by the confidence map: + +$$ +\mathcal {L} ^ {\text {c o n f}} (n, m) = \sum_ {i, j \in \mathcal {D}} C _ {i, j} ^ {n, m} \mathcal {L} _ {i, j} ^ {\text {r e g r}} (n, m) - \alpha \log C _ {i, j} ^ {n, m}, \tag {4} +$$ + +with $\alpha = 0.2$ . This loss penalizes the model less when the prediction is not accurate on harder areas, encouraging the model to extrapolate. The final loss is expressed as + +$$ +\mathcal {L} = \mathcal {L} ^ {\text {c o n f}} (1, 1) + \mathcal {L} ^ {\text {c o n f}} (2, 1) + \beta \mathcal {L} ^ {\text {c o n f}} (2, 2), \tag {5} +$$ + +where $\beta$ is a hyper-parameter typically set to $\beta = 1$ . + +# 3.2. Adding Versatile Conditioning + +The knowledge of auxiliary information can significantly enhance 3D predictions at test time. Our model leverages up to five different modalities, which include two intrinsics $K_{1}, K_{2}$ , two depthmaps $D^{1}, D^{2}$ for each image and the relative pose $P_{12}$ . To condition the output on it, we first embed the auxiliary information using dedicated MLPs and then inject these embeddings at different points in the pipeline. + +We discuss different options in this regard. In the first strategy, denoted as 'embed', we simply add the auxiliary embeddings to the token embeddings before the first transformer block, as would be done for standard positional embeddings [16]. In the second strategy, denoted as 'inject- $n$ ' we instead insert dedicated MLPs for each modality in a subset of $n$ transformer blocks, as shown in Fig. 3. Our experiments show that 'inject-1' performs slightly better than 'embed' and similarly with 'inject- $n$ ' where $n > 1$ , see + +![](images/ab31b72c696d7f4c4a54a16400083834c04f94c8bb7b56e7c245739b0289e580.jpg) + +![](images/c77f3013c264b67a5524128d9598d745fffda9b53e92dc598ef6661508aa0cfd.jpg) +Figure 3. Architecture for injecting auxiliary modalities. Top: injection of optional intrinsics and depth into the encoder. Intrinsic s are encoded into ray patches, sparse depth is patchified. Each of these modalities goes into a block-specific MLP and are tokenwise added in the middle of the transformer block. Bottom: injection of optional relative pose into the decoder. The relative pose is fed to a first embedding layer followed by an MLP. This token is added to the CLS token of the decoder after the self-attention (SA) and cross-attention (CA), but before the MLP. Experiments show that injection in the first block only suffices. + +Sec. 4.5. We now describe how to compute the initial embeddings for each specific modality. + +Intrinsics. Following [89], we generate camera rays from the intrinsic matrix $K \in \mathbb{R}^{3 \times 3}$ , thereby establishing a direct correspondence between RGB pixels and rays. The ray at pixel location $(i,j)$ is computed as $K^{-1}[i,j,1]$ and encodes the viewing direction of that pixel w.r.t. the current camera pose. This allows us to potentially process non-centered crops and hence to perform inference in higher image resolutions, see Sec. 3.3. Similarly to RGB inputs, we patchify and embed dense rays and provide them to the encoder. + +Depthmaps / Point Clouds. Given a depthmap $D$ and its sparsity mask $M$ , we first normalize $D' = D / \mathrm{norm}(D)$ to handle any depth ranges at train and test time. As for RGB and rays, we patchify the stacked maps $[D', M] \in \mathbb{R}^{W \times H \times 2}$ and compute patch embeddings, which are then fed to the encoder. By jointly patchifying the depth and its valid mask we can handle any level of sparsity. + +Camera Pose. Given the relative pose $P_{12} = [R_{12}|t_{12}]$ , we normalize the translation scale as $t_{12}' = t_{12} / \|t_{12}\|$ since our output is unscaled. Unlike depthmaps or camera intrinsics, the camera pose cannot be expressed as a dense pixel map. Rather, camera pose affects the whole pixels between two images, so the embedding is instead added to the global CLS token of both decoders. + +# 3.3. Downstream Tasks + +Depthmaps. In the pointmap representation, the $z$ -axis of $X^{1,1}, X^{2,2}$ directly corresponds to the depthmaps of the first and second image, respectively. Note that DUSt3R + +![](images/17d580084d03bb1334327108baa4f1b493068a4a757a8cba3f3467c63d8eaee0.jpg) +Figure 4. Native resolution example. Methods like DUSt3R require centered principal point and would thus downscale the image to its training resolution, leading to blurry prediction (top). With camera intrinsics input, we can process any crop in the image leading to native resolution prediction (bottom row), for example by dividing the image into 4 crops, that are independently process as the first input image, while the second input image of the network is set to the downscaled image. + +![](images/babbe8245315e7913c89d97a8f11bc2104888d68a413b018c1eddb82ae7f6cdf.jpg) + +![](images/01f03d6f78e2f1123369db61f5134fd05c3d7451b82eb8111f8cfe6681fda966.jpg) + +only outputs $X^{1,1}$ , hence it requires two forward passes of the pairs $(I^1, I^2)$ and $(I^2, I^1)$ to extract both depthmaps. + +High-resolution processing. A limitation of DUSt3R is that it is trained for a specific resolution and does not generalize to larger resolutions. A simple fix would be to include higher resolutions during training, but (i) many datasets do not provide high-resolution ground-truths, and (ii) the training cost could become prohibitive. Pow3R, in contrast, can handle crops natively given camera intrinsic's of the crop, as these provide the crop position information (i.e. via focal length and principal point). We can thus perform prediction in a sliding window fashion, yielding predictions matching any target resolution by simple stitching. Comparison between standard and higher resolution processing are shown in Fig. 4. Note that prediction for each crop may have a different scale, by design, and cannot be stitched directly. We find that simply computing the median scale factor in overlapping areas is sufficient, and we blend the overlapping regions based on confidence without further post-processing. + +Focal Estimation. Similar to DUSt3R, we can recover fccals for both input images from pointmap $X^{1,1}$ and $X^{2,2}$ with the robust Weiszfeld fast iterative solver [47]. Similar to depthmap prediction, DUSt3R needs to infer $(I_2, I_1)$ in order to compute the focal of the second image, while Pow3R predicts them in a single pass. + +Relative Pose Estimation. Pow3R predicts the relative pose directly by Procrustes alignment [8, 37] to get the scaled relative pose $P^{*} = \left[R^{*}|t^{*}\right]$ between $X^{2,2}$ and $X^{2,1}$ , as it predicts the pointmaps of the second image in two different camera coordinates. + +$$ +R ^ {*}, t ^ {*} = \arg \min _ {\sigma , R, t} \sum_ {i, j} \sqrt {C _ {i , j} ^ {2 , 2} C _ {i , j} ^ {2 , 1}} \left\| \sigma \left(R X _ {i, j} ^ {2, 2} + t\right) - X _ {i, j} ^ {2, 1} \right\| ^ {2}. \tag {6} +$$ + +Procrustes alignment is known to be sensitive to noise and outliers; however, in our case, it performs similar to RANSAC [17] with PnP [22, 29], while being an order of magnitude faster (see Sec. 4.4). + +Global alignment. The network $\mathcal{F}$ predicts pointmaps for image pairs. To align all predictions in the same world coor + +dinate system, we resort to the global alignment algorithm from DUSt3R [75] that minimizes a global energy function to find per-camera intrinsics, depthmaps and poses that are consistent with all the pairwise predictions. Results of the optimization are global scene point-clouds, which can for instance serve for multi-view stereo estimation. + +# 3.4. Training Procedure + +During training, we feed the network with annotated image associated with random subsets of auxiliary information, the goal being for it to learn to handle any situation at test time. For each pair, we first chose a random number $m$ of modalities with uniform probability, and then randomly select the $m$ modalities likewise. Depthmaps are randomly sparsified. When giving intrinsics, we also perform aggressive non-centered cropping with $50\%$ probability, so that the network learns to perform high-resolution inference. + +Training data and recipe. We follow the training recipe from DUSt3R [75] in terms of training data and protocol. Namely, we use the provided pairs for 8 datasets: Habitat [58], MegaDepth [32], ARKitScenes [6], Static Scenes 3D [41], BlendedMVS [82], ScanNet++ [84], Co3Dv2 [51] and Waymo [67], resulting in a total of 8.5M image pairs. These datasets include indoor and outdoor scenes, as well as real and synthetic ones. To train the model, we follow [75] and first train the network with a resolution of 224px for 3 days on 8 A100 GPUs, and then finetune it in 512px with variable aspect-ratio for 2 more days. + +# 4. Experimental Evaluation + +We first study the impact of auxiliary information guidance (Sec. 4.1). We then experiment across many 3D vision tasks (Secs. 4.2 to 4.4) without retraining, i.e. mostly zero-shot settings. We finally present some ablations in Sec. 4.5. + +DUSt3R baseline. We point out that our overall training data, protocols and architecture are exactly the same as DUSt3R, except for the additional head that predict $X^{2,2}$ ( $+0.1\%$ parameters) and for the modules used for injecting + +
aux. modalitiesfocal acc@1.015depth τ@1.03rel. pose
K1K2D1D2RTRRA@2°RTA@2°
DUSt3R×××××36.677.669.249.7
Pow3R×××××39.479.471.953.8
✓××××75.4(+36.0)80.6 (+1.2)77.1 (+5.3)62.8 (+8.9)
×✓×××67.3(+27.9)80.3 (+0.9)74.5 (+2.6)59.1 (+5.2)
✓✓×××98.0(+58.6)81.4 (+2.0)80.3 (+8.5)74.2(+20.3)
××✓××48.9 (+9.5)89.1 (+9.7)82.5(+10.6)64.9(+11.0)
×××✓×49.5(+10.1)91.0(+11.7)83.4(+11.5)64.8(+10.9)
××✓✓×58.2(+18.9)95.4(+16.0)89.6(+17.7)77.6(+23.8)
××××✓48.6 (+9.2)81.6 (+2.2)92.3(+20.5)77.0(+23.1)
✓✓✓✓×99.2(+59.8)95.4(+16.0)95.7(+23.9)94.3(+40.5)
✓✓××✓98.1(+58.7)82.9 (+3.6)95.1(+23.2)87.3(+33.5)
××✓✓✓68.6(+29.2)95.4(+16.0)98.1(+26.2)91.3(+37.5)
✓✓✓✓✓99.3(+59.9)95.4(+16.0)99.0(+27.2)98.1(+44.3)
+ +Table 1. Impact of guiding at test time for models trained at $224 \times 224$ resolution. We report performances on Habitat for DUSt3R, which cannot handle auxiliary modalities, and our single model with different sets of modalities; we show in green the absolute improvement w.r.t. the results without auxiliary modality. + +the auxiliary information $(+4\%$ parameters for the inject-1 variant that we use unless stated otherwise). + +Metrics. When evaluating focal lengths, we report the accuracy for a tight relative error threshold of $1.5\%$ . For evaluating depthmap, we report the Absolute Relative Error (rel) and the Inlier Ratio $(\tau)$ at a threshold of $3\%$ . When evaluating poses, we opt for scale-invariant metrics and measure the relative rotation and translation accuracies (RRA and RTA resp.) of the relative pose at task-specific thresholds, and the mean Average Accuracy (mAA), defined as the area under the curve accuracy of the angular differences at min(RRA, RTA). + +# 4.1. Guiding the output prediction + +Impact of guidance. We first study the impact of providing different subsets of auxiliary information in terms of downstream 3D vision performance. Namely, we report accuracies in Tab. 1 for focal, depth, and rotation prediction on the Habitat validation dataset. In the supplementary material, we also provide results on the Infinigen dataset [49], which is not included in our training set, and observe similar outcomes. We first note that in the absence of modalities, Pow3R performs slightly better than a DUSt3R baseline, even though Pow3R is trained with just RGB images during a fraction of its training. This shows that Pow3R does not lose its native strength for unconstrained regression despite having gained more capabilities. + +Interestingly, we observe in Tab. 1 the impact of each modality for each prediction task, and the synergetic effects of interacting modalities. For instance, the model performs better for focal length estimation with depthmaps and/or relative pose, and Pow3R's depth prediction improves with intrinsics and/or relative pose. Moreover, relative pose prediction becomes more accurate with intrinsics and/or depthmap. Overall, the model's performance consistently improves when provided with additional camera or scene priors, demonstrating the effectiveness of Pow3Rin + +![](images/826dd2e89c2688b8204dd337cb9dc4237dd5a3a1d8bc67d6b14bb35cdffc7c4f.jpg) +Figure 5. Depth completion performance on NYUd for various input depth sparsity. + +
aux. mod.High-Res.KITTIT&TTime (sec)↓
KsRTrel↓τ↑rel↓τ↑
DUSt3R512×××5.449.53.375.10.13
Pow3R512✓✓×5.348.73.278.20.13
Pow3R512✓✓(n)7.534.43.968.00.48
Pow3R512✓✓✓4.653.52.582.30.40
+ +Table 2. Multi-view depth estimation results in high-resolution using a sliding-window scheme and providing the crop intrinsics to Pow3R. (n): naively inputting high-resolution images without downsampling to Pow3R. + +handling any combination of priors in a unified fashion. We are not aware of any other approach that can seamlessly integrate all these priors in a learning-based method. + +Depth completion. We also evaluate Pow3R on the depth completion task given RGB images with sparse depthmaps. We vary the sparsity (the ratio of pixels with known depth) and demonstrate its impact on the NYUv2 dataset [63] in Figure 5. Although Pow3R is not trained on the NYUv2 dataset, it still achieves state-of-the-art results for sparse depth completion. Our analysis indicates that the true performance of Pow3R on NYUv2 is likely underestimated due to numerous errors in the ground-truth depth annotations; further details are available in the supplementary material. On the KITTI [20] dataset, we obtain a mean absolute error below $30\mathrm{cm}$ , which is competitive with the state of the art, considering that KITTI is not included in the training set. + +High-Resolution processing. With the possibility of specifying camera intrinsics, new capabilities arise. Namely, Pow3R can perform high-resolution pointmap regression with a simple sliding-window approach. We process fixed-resolution crops of the original images under the guidance of both (i) the crop information, provided in term of camera intrinsics, and (ii) a second image, itself possibly down-scaled or cropped as well. More details are provided in Sec. 3.3 and Fig. 4. When applied, results significantly improve compared to processing the images at 512-pixel resolution, as shown in Tab. 2. We also compare with a naive baseline where we simply feed high-resolution images directly to Pow3R, but it dramatically fails. + +Camera controllability. We study how controllable are the regression outputs on 10K Habitat test pairs given arbitrary camera parameters. For each test pair with ground-truth fo + +![](images/f5fa1fa4f400cd95e2de9546aeab43e83115042a21d3c66b1bba5662cec59172.jpg) + +![](images/3b1c173f2a5a1298768c924cd3cf424f9cf3cd32a8bf924bf5780380fdc36fa5.jpg) + +![](images/b2d66dbb9e23093374d1bf4125e289508593aa82cbf9c68e7b0aaa3576c439e0.jpg) +Figure 6. Controllability study. We study how the model reacts when presented with false focal (top) and false relative pose (bottom). We plot the input vs. predicted camera parameters in the left column and the average confidence in the right column. Interestingly, the model seems to not take the input information at face value and instead makes an informed decision whether to use it or not, in which latter case the confidence becomes low. + +![](images/dd045e1f9063569778ba0a83867a01cfd78d48c4a4366ef39388058716586544.jpg) + +cal $f$ and relative pose $P_{12}$ , we purposefully input a wrong focal (resp. relative pose) and measure the output focal $\hat{f}$ (resp. pose $\hat{P}_{12}$ ) as predicted from the raw pointmaps. Results are presented in Fig. 6 in terms of focal ratio $\hat{f} / f$ and angle $d_{\text{geodesic}}(P_{12}, \hat{P}_{12})$ . Interestingly, the model accepts to follow the guidance whenever it is close to the ground truth value (i.e. staying close to the identity transformation) but then refuses to follow it when it gets too different, showing that the model is able to somehow 'judge' of the guidance's worth. In all cases, the output confidence clearly correlates with the agreement between the guidance and the ground truth. Overall, the model is thus not taking the auxiliary information at face value, and instead combines it with the other modalities (e.g. RGB) it in a more subtle and informed way than what could be achieved with handcrafted methods. This suggests that Pow3R would be robust to noise in the auxiliary information. + +# 4.2. Multi-View Depth Estimation + +Following RobustMVD [62], performances are measured on the KITTI [20], ScanNet [14], ETH3D [61], DTU [1], and Tanks and Temples [28] datasets. We report in Tab. 3 the Absolute Relative Error (rel) and Inlier Ratio $(\tau)$ for each dataset when using (1) images alone, (2) images with relative poses, (3) images with camera intrinsics, and (4) images with both modalities, and provide results for more state-of-the-art methods in the supplementary material. + +Results. Pow3R again performs roughly on par with DUSt3R in the absence of auxiliary information (e.g. 3.73 average error across all datasets for DUSt3R-512 vs. 3.64 + +
MethodGT intrinsicsCo3Dv2RealEstate10K mAA(30)Speed (fps)
RRA@15RTA@15mAA(30)
Colmap+SG [56]✓36.127.325.345.2-
PixSfM [34]✓33.732.930.149.4-
RelPose [88]×57.1----
(a) PoseDiff [73]×80.579.866.548.0-
RelPose++ [33]×85.5----
RayDiff [89]×93.3----
DUSt3R (PnP) [75]×94.388.477.261.73.2
(b) Pow3R (PnP)×94.889.978.562.53.2
Pow3R (Pro)×94.690.378.166.330.9
Pow3R w/ K (Pro)✓95.092.182.272.530.1
+ +Table 5. Multiview pose estimation on the CO3Dv2 and RealEstate10K. (a) denotes multiview methods, while (b) denotes pairwise methods. Pow3R with Procrustes alignment (Pro) performs better than DUSt3R (PnP) and other multi-view based pose estimation approaches, and even more so with known intrinsics, while being much faster. RealEstate10K is not in the training set. + +for Pow3R-512). However, it is clear that providing auxiliary information consistently improves the depth estimates. In fact, Pow3R-512 with both pose and intrinsics achieves state-of-the-art performance on the ETH3D, DTU, and Tanks and Temples datasets, as well as on average across datasets. Interestingly, intrinsics tend to improve performance more than the GT pose, which alone seems to have a negligible impact. This might be due to the fact that camera pose primarily helps the prediction of $X^{2,1}$ , while the depthmap of this task is obtained from the $z$ axis of $X^{1,1}$ alone. Still, we observe a clear synergy in providing jointly intrinsics and relative pose. + +# 4.3. Multi-View Stereo + +We evaluate the quality of global 3D reconstructions on the DTU [1] benchmark in the presence and absence of auxiliary information. For each subset of available information from $\{(K_1, K_2), P_{12}\}$ , we extract pairwise predictions from image pairs and align them in a global coordinate system, as described in Sec 3.3. Performance is evaluated in terms of accuracy, which is the smallest Euclidean distance to the ground-truth, and completeness as the smallest distance to the reconstructed shape, with the overall average. + +Results are presented in Tab. 4. We first attempt at reproducing the original DUSt3R results using the official codebase and obtain slightly worse results than reported in [75]. Using the same codebase, Pow3R with RGB only is able to improve over these and match the original results. Interestingly, adding ground-truth intrinsics results in a significant boost (19% rel. improvement), but adding the camera poses instead slightly degrades the results (-4%). However, using both intrinsics and pose results in a major improvement (36% rel. improvement), demonstrating a synergetic effect between both modalities. We hypothesize that the pose is of little value for easy cases like DTU, and that it does not help the model to properly resolve the focals. + +
MethodGT rangeKITTIScanNetETH3DDTUT&TAverage
rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑
COLMAP [59, 60] (K+RT)×12.058.214.634.216.455.10.796.52.795.09.367.8
COLMAP Dense [59, 60] (K+RT)×26.952.738.022.589.823.220.869.325.776.440.248.8
MVSNet [81] (K+RT)✓18.630.722.720.921.635.6(1.8)(86.7)6.574.614.249.7
Vis-MVSNet [86] (K+RT)✓9.555.48.933.510.843.3(1.8)(87.4)4.187.27.061.4
MVS-Former++ [9] (K+RT)✓29.215.215.221.921.432.5(1.2)(91.9)7.671.514.946.6
CER-MVS [40] (K+RT)×14.332.221.124.311.747.54.171.36.482.111.551.5
DUSt3R [75]×5.449.5(3.1)(71.8)3.076.03.968.63.375.13.768.2
Pow3R×5.745.7(3.2)(68.8)3.074.73.074.33.376.63.668.0
Pow3R w/ RT×5.745.8(3.2)(69.7)2.975.63.371.63.277.93.768.1
Pow3R w/K×5.348.3(3.1)(70.8)2.976.01.689.93.277.33.272.5
Pow3R w/K+RT×5.348.7(3.1)(71.4)2.877.11.591.13.278.23.273.3
+ +Table 3. Multi-view depth evaluation: Pow3R with both pose and intrinsics performs better than DUSt3R. More comparisons with the state of the art can be found in the supplementary material. (Parentheses) denote training on data from the same domain. + +
MethodAcc.↓Comp.↓Overall↓
DUSt3R†[75]2.6770.8051.741
DUSt3R (repr.)2.1911.5981.894
Pow3R2.1161.3701.743
Pow3Rw/ K1.7221.1191.420
Pow3Rw/ Rt2.2051.4291.817
Pow3Rw/ K+RT1.3840.8461.115
+ +Table 4. MVS results with the accuracy, completeness and their average (in mm) on the DTU dataset. $\mathrm{DUSt3R}^{\dagger}$ means results from the original paper, and 'repr.' means reproduced with the official public codebase and checkpoint. + +# 4.4. Multi-View Pose estimation + +We then evaluate Pow3R for the task of multi-view pose prediction on the Co3Dv2 [51] and RealEstate10K [91] datasets. Co3Dv2 comprises over 37,000 videos across MS-COCO 51 categories with ground-truth camera derived from running COLMAP on 200 frames per video. RealEstate10k contains 80,000 YouTube video clips, containing indoor/outdoor scenes, with camera poses obtained via SLAM followed by Bundle Adjustment. For Co3Dv2, we follow the protocol of [73] and for RealEstate10K, we use its test set of 1,800 videos. For each sequence, we randomly select 10 frames and feed 45 pairs to Pow3R. We report the results in terms of RRA, RTA and mAA in Tab. 5. + +Results. Pow3R predicts $X^{2,2}$ , allowing to estimate the relative pose between 2 images in a single forward using Procrustes Alignment (see Sec. 3.3). In contrast, DUSt3R does not output $X^{2,2}$ and relies on PnP to obtain camera poses. Pow3R with PnP achieves slightly better results than DUSt3R with PnP. Using Procrustes alignment yields similar results, while being orders of magnitude faster than RANSAC, with a roughly $10 \times$ speed-up overall when combining inference and pose estimation times. Finally, performance improves even further with known intrinsic, significantly outperforming all the state of the art. + +# 4.5. Architecture ablation + +We evaluate the impact of adding the $X^{2,2}$ prediction in Tab. 6. For these ablations, models are trained on a subset of Habitat and tested on focal, depth, and relative pose prediction, with results averaged over all combinations of auxiliary modalities. DUSt3R serves as a baseline, and DUSt3R-ft is finetuned with an additional $X^{2,2}$ head. This output can be used at test time to directly predict focals or relative poses, within a single forward pass, instead of having to forward both the original pair and its swapped version. We also report the performance, with and without predicting $X^{2,2}$ and/or using it. In all cases, predicting $X^{2,2}$ leads to slightly better performance while evaluations can be performed in a single forward pass. + +We also explore different ways of conditioning auxil + +
Methodwith X2,2Single ForwardFocal acc@1.015Depth τ@1.03RRA@2 (Pro)RTA@2 (PnP)Avg.
Pro)PnP)
DUSt3R××36.885.469.272.449.857.3
DUSt3R-ft✓×38.979.171.774.452.959.2
✓✓38.378.573.174.456.359.2
Pow3R××66.085.781.983.868.673.3
✓×67.286.182.984.770.074.6
✓✓64.885.084.284.773.074.6
+ +Table 6. Impact of regressing $X^{2,2}$ . Metrics (in %) on Habitat are averaged over all combinations of input modalities. The 'with $X^{2,2}$ ' column denotes methods trained to predict it. 'Single forward' denotes evaluations of models based on $X^{2,2}$ prediction, i.e., in a single forward pass, instead of forwarding image pairs and their swapped versions. The top part shows the original DUSt3R, or finetune (-ft) with $X^{2,2}$ prediction. + +
Arch.Extra ParamsFocal acc@1.015Depth τ@1.03RRA@2 (Pro) (PnP)RTA@2 (Pro) (PnP)Avg.
Embed16.0M64.784.784.184.572.873.877.4
Inject122.6M64.885.084.284.773.074.677.7
Inject442.3M65.985.184.885.273.775.478.3
+ +Table 7. Architecture ablation. Results (in %), on Habitat, are averaged over all combinations of input modalities. We compare the injection architecture for different number of blocks with the 'embed' architecture in which the extra signals are added before the transformer blocks. 'Inject1' offers a good compromise between performance and number of extra parameters. + +ary information in Tab. 7. To incorporate additional priors, Pow3R injects those embeddings into the first block of the encoder or decoder, see Fig. 3. We compare this approach to another design choice, where the model embeds them before going through attention blocks. We have also tried to perform injection in more than one block, without significant improvement while the computational cost increases. + +# 5. Conclusion + +We have presented Pow3R, a novel approach to guide large 3D vision models, e.g. DUSt3R, with camera and scene priors in an implicit manner. Not only does it results in consistent gains whenever auxiliary information is available, but it also unlocks new capabilities such as high-resolution processing and sparse-to-dense depth completion. + +# References + +[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. IJCV, 2016. 7 +[2] Youset I Abdel-Aziz, Hauck Michael Karara, and Michael Hauck. Direct linear transformation from comparator coordinates into object space coordinates in close range photogrammetry. Photogrammetric engineering & remote sensing, 2015. 3 +[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR, 2017. 3 +[4] Daniel Barath, Dmytro Mishkin, Luca Cavalli, Paul-Edouard Sarlin, Petr Hruby, and Marc Pollefeys. Affineglue: Joint matching and robust estimation. arXiv preprint arXiv:2307.15381, 2023. 2 +[5] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In ICCV, 2019. 2 +[6] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In NeurIPS, 2021. 5 +[7] Herbert Bay. Surf: Speeded up robust features. ECCV, 2006. 2 +[8] Romain Brégier. Deep regression on manifolds: a 3D rotation case study. In 3DV, 2021. 5 +[9] Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer++: Revealing the devil in transformer's details for multi-view stereo. In ICLR, 2024. 2, 8 +[10] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, 2022. 2 +[11] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In ECCV, 2018. 2 +[12] David J Crandall, Andrew Owens, Noah Snavely, and Daniel P Huttenlocher. Sfm with mrfs: Discrete-continuous optimization for large-scale structure from motion. IEEE trans. PAMI, 2012. 2 +[13] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In CVPR, 2017. 2 +[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 7 +[15] Timothee Darcet, Mathilde Caron, Piotr Bojanowski, Ishan Misra, Armand Joulin, Julien Mairal, Hervé Jégou, and Hugo Touvron. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 1 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is + +worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4 +[17] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 5 +[18] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 2015. 2 +[19] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In ICCV, 2015. 2 +[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. *IJRR*, 2013. 6, 7 +[21] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, 1988. 2 +[22] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2, 5 +[23] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In ICRA, 2021. 2 +[24] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In ICCV, 2021. 2 +[25] Nianjuan Jiang, Zhaopeng Cui, and Ping Tan. A global linear method for camera pose registration. In ICCV, 2013. 2 +[26] Shengyu Jiang, Chen Wang, Yuhong Zhang, and Zhibin Wang. Completionformer: Efficient and accurate depth completion via hybrid attention with transformers. In CVPR, 2023. 1, 2 +[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM trans. Graphics, 2023. 1 +[28] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM trans. Graphics, 2017. 7 +[29] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. IJCV, 2009. 5 +[30] Vincent Leroy, Jean-Sébastien Franco, and Edmond Boyer. Volume sweeping: Learning photoconsistency for multi-view shape reconstruction. IJCV, 2021. 2 +[31] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with MAST3R. In ECCV, 2024. 2, 3 +[32] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 5 +[33] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 3, 7 +[34] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 7 + +[35] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In ICCV, 2023. 2 +[36] David G Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 2004. 2 +[37] Bin Luo and Edwin R. Hancock. Procrustes alignment with the em algorithm. In CAIP, 1999. 5 +[38] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In ICRA, 2018. 2 +[39] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In ICRA, 2019. 2 +[40] Zeyu Ma, Zachary Teed, and Jia Deng. Multiview stereo with cascaded epipolar raft. In ECCV, 2022. 2, 8 +[41] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5 +[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 1 +[43] Danish Nazir, Marcus Liwicki, Didier Stricker, and Muhammad Zeshan Afzal. Semattnet: Towards attention-based semantic aware guided depth completion. arXiv preprint arXiv:2204.13635, 2022. 2 +[44] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 2 +[45] OpenAI. Chatgpt-4, 2024. 1 +[46] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 1, 2 +[47] Frank Plastria. The weiszfeld algorithm: Proof, amendments, and extensions. Foundations of Location Analysis, 2011. 5 +[48] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deep lidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In CVPR, 2019. 2 +[49] Alexander Raistrick, Lahav Lipson, Zeyu Ma, Lingjie Mei, Mingzhe Wang, Yiming Zuo, Karhan Kayan, Hongyu Wen, Beining Han, Yihan Wang, Alejandro Newell, Hei Law, Ankit Goyal, Kaiyu Yang, and Jia Deng. Infinite photorealistic worlds using procedural generation. In CVPR, 2023. 6 +[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 4 +[51] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 5, 8 + +[52] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In CVPR, 2022. 1, 2 +[53] Chris Rockwell, Justin Johnson, and David F Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 3DV, 2022. 3 +[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1 +[55] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In ECCV, 2006. 2 +[56] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In CVPR, 2019. 7 +[57] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In CVPR, 2020. 2 +[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 5 +[59] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2, 8 +[60] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 8 +[61] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In CVPR, 2017. +[62] Philipp Schröppel, Jan Bechtold, Artemij Amiranashvili, and Thomas Brox. A benchmark and a baseline for robust multiview depth estimation. In 3DV, 2022. 7 +[63] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6 +[64] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2 +[65] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 3 +[66] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 2 +[67] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5 +[68] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE trans. Image Processing, 2020. 2 + +[69] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, 2024. 2 +[70] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 2 +[71] Wouter Van Gansbeke, Davy Neven, Bert De Brabandere, and Luc Van Gool. Sparse and noisy lidar completion with rgb guidance and uncertainty. In MVA, 2019. 2 +[72] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. In 3DV, 2025. 3 +[73] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In ICCV, 2023. 1, 3, 7, 8 +[74] Shuzhe Wang, Juho Kannala, Marc Pollefeys, and Daniel Barath. Guiding local feature matching with surface curvature. In ICCV, 2023. 2 +[75] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 3, 5, 7, 8 +[76] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In ICCV, 2023. 2 +[77] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Bregier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In ICCV, 2023. 1 +[78] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 2 +[79] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet++: Semantic assisted repetitive image guided network for depth completion. In ECCV, 2022. 2 +[80] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, 2024. 2 +[81] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In ECCV, 2018. 8 +[82] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In CVPR, 2020. 5 +[83] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 2 +[84] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 5 +[85] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In ECCV, 2016. 2 + +[86] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. In BMVC, 2020. 8 +[87] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. In ICLR, 2025. 3 +[88] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Rel-pose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 3, 7 +[89] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In ICLR, 2024. 1, 3, 4, 7 +[90] Yinda Zhang and Thomas Funkhouser. Deep depth completion of a single rgb-d image. In CVPR, 2018. 2 +[91] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. SIGGRAPH, 2018. 8 \ No newline at end of file diff --git a/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/images.zip b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2e6053c3826827caf06b0289a32e9f99325f52cc --- /dev/null +++ b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:656f713eb629b8ae8411eb818c7e14464ce3e1cb189821948e1d3ba0662cf63b +size 609769 diff --git a/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/layout.json b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0608ce52adcf5578e9b642e64830c1c209e94c08 --- /dev/null +++ b/2025/Pow3R_ Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors/layout.json @@ -0,0 +1,10652 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 103, + 449, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 103, + 449, + 138 + ], + "spans": [ + { + "bbox": [ + 162, + 103, + 449, + 138 + ], + "type": "text", + "content": "Pow3R: Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 160, + 462, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 160, + 462, + 205 + ], + "spans": [ + { + "bbox": [ + 160, + 160, + 462, + 205 + ], + "type": "text", + "content": "Wonbong Jang* Philippe Weinzaepfel† Vincent Leroy† \nLourdes Agapito* Jerome Revaud† \n*UCL †Naver Labs Europe" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 206, + 476, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 206, + 476, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 206, + 476, + 217 + ], + "type": "text", + "content": "{ucabwja,1.agapito}@ucl.ac.uk firstname.lastname@naverlabs.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 232, + 169, + 308 + ], + "blocks": [ + { + "bbox": [ + 56, + 232, + 169, + 308 + ], + "lines": [ + { + "bbox": [ + 56, + 232, + 169, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 169, + 308 + ], + "type": "image", + "image_path": "cd468f966acef602dc94daffba7df92ffce29633c66132b6a368cb37e4b2ad19.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 310, + 168, + 384 + ], + "blocks": [ + { + "bbox": [ + 56, + 310, + 168, + 384 + ], + "lines": [ + { + "bbox": [ + 56, + 310, + 168, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 168, + 384 + ], + "type": "image", + "image_path": "b7996d15fe1224731089515b787255ed643c7e38300242e17396ae4643c12fe8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 170, + 232, + 354, + 384 + ], + "blocks": [ + { + "bbox": [ + 170, + 232, + 354, + 384 + ], + "lines": [ + { + "bbox": [ + 170, + 232, + 354, + 384 + ], + "spans": [ + { + "bbox": [ + 170, + 232, + 354, + 384 + ], + "type": "image", + "image_path": "908ee6513e718a96c7c64ef3920fb3a7b74c29bfbb673f221e9a03e007190970.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 384, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 555, + 417 + ], + "type": "text", + "content": "Figure 1. Qualitative comparison of raw 3D reconstructions between DUSt3R and Pow3R for the image pair on the left. DUSt3R cannot leverage auxiliary information potentially available at test time, resulting in noisy and low-resolution estimates, whereas Pow3R allows to inject such priors at test time in a versatile manner, enabling new capabilities such as high-resolution processing as a direct by-product." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 354, + 232, + 554, + 384 + ], + "blocks": [ + { + "bbox": [ + 354, + 232, + 554, + 384 + ], + "lines": [ + { + "bbox": [ + 354, + 232, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 354, + 232, + 554, + 384 + ], + "type": "image", + "image_path": "657a18594bd983f7c7c17f58b7385fd5ba70716fe8c587b43ae61d0c62f47d96.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 426, + 200, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 426, + 200, + 439 + ], + "spans": [ + { + "bbox": [ + 152, + 426, + 200, + 439 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 450, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 450, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 53, + 450, + 297, + 715 + ], + "type": "text", + "content": "We present Pow3R, a novel large 3D vision regression model that is highly versatile in the input modalities it accepts. Unlike previous feed-forward models that lack any mechanism to exploit known camera or scene priors at test time, Pow3R incorporates any combination of auxiliary information such as intrinsics, relative pose, dense or sparse depth, alongside input images, within a single network. Building upon the recent DUSt3R paradigm, a transformer-based architecture that leverages powerful pre-training, our lightweight and versatile conditioning acts as additional guidance for the network to predict more accurate estimates when auxiliary information is available. During training we feed the model with random subsets of modalities at each iteration, which enables the model to operate under different levels of known priors at test time. This in turn opens up new capabilities, such as performing inference in native image resolution, or point-cloud completion. Our experiments on 3D reconstruction, depth completion, multi-view depth prediction, multi-view stereo, and multi-view pose estimation tasks yield state-of-the-art results and confirm the effectiveness of Pow3R at exploiting all available information. The project webpage is https://europe.naverlabs.com/pow3r." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 426, + 394, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 426, + 394, + 439 + ], + "spans": [ + { + "bbox": [ + 315, + 426, + 394, + 439 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 688 + ], + "type": "text", + "content": "Building general models for 3D perception that can unify different 3D vision tasks such as depth estimation, keypoint matching, dense reconstruction or camera pose prediction, remains a complex and exciting open challenge. Traditional Structure from Motion (SfM) approaches such as COLMAP [59] do not apply pre-learned priors, so each scene must be optimized independently. On the other hand, most learning-based approaches focus on specific 3D vision tasks such as depth completion [26, 52], monocular and stereo depth estimation [46], pose estimation [73, 89] or novel view synthesis [27, 42]. While these dedicated approaches have been successful to some extent, this trend contrasts with other domains such as natural language processing where a single model like ChatGPT [45], trained with a vast collection of data, solves a range of different tasks without requiring re-training. While some models like CroCo-v2, DINOv2 or Stable Diffusion [15, 54, 77], trained on large image datasets, have served as learned priors for other tasks, including 3D vision ones, they are still far from multi-purpose foundation models like ChatGPT is for NLP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "DUSt3R [75] marked a significant step forward in regression-based 3D foundation models. DUSt3R performs" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1071" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "type": "text", + "content": "3D dense reconstruction from uncalibrated and unposed images, solving many downstream vision tasks at once without the need for fine-tuning. Its success can be attributed to two main factors: (i) its use of a pointmap representation, which provides flexibility over its output space and enables to regress camera parameters, poses and depth easily and (ii) its use of a transformer architecture that exploits powerful pretraining combined with a vast amount of 3D training data. Despite its huge success, DUSt3R [75] and its successor MASt3R [31] are extremely limited in terms of their input space, i.e. they only take RGB images. Yet, in practice, many real-world applications do provide extra modalities such that calibrated camera intrinsics, sparse or dense depth from RGB-D sensors or LIDAR, etc. This raises the question of how best to exploit additional input modalities, and whether this not just boosts performance but also enables new capabilities, such as obtaining the full depth from sparse point clouds of SfM or LiDAR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 288, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 295, + 502 + ], + "type": "text", + "content": "In this paper, we introduce Pow3R, a new 3D feedforward model that empowers DUSt3R with any subset of priors available at test time such as camera intrinsic, sparse or dense depth, or relative camera poses. Each modality is injected into the network in a lightweight fashion. To allow the model to operate under different conditions at test time, random subsets of input modalities are fed at each training iteration. As a result, we obtain a single model that performs on par with DUSt3R when no prior information is available but outperforms it when it exists. Our model also gains new capabilities as a by-product: for instance, the camera intrinsic input allows to process images whose principal point is far from the center, thus allowing to perform extreme cropping, e.g. for performing sliding window inference. As a side contribution, our model directly outputs the pointmaps of the second image in its coordinate system, allowing faster relative pose estimation. To sum up, our contributions are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 502, + 295, + 610 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 502, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 538 + ], + "type": "text", + "content": "- Pow3R is a holistic 3D geometric vision model capable of taking any subset (including none) of camera intrinsics, pose and depthmaps with corresponding input images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 539, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 539, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 539, + 295, + 574 + ], + "type": "text", + "content": "- Extensive experiments demonstrate an important boost in performance over DUSt3R which cannot exploit priors, yielding state-of-the-art results in three benchmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 574, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 574, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 574, + 295, + 610 + ], + "type": "text", + "content": "- By predicting the same pointmaps in two different camera coordinate systems, we can achieve more accurate relative pose, orders of magnitude faster." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 620, + 139, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 139, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 139, + 632 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Structure-from-Motion and MVS. SfM [12, 13, 22, 25, 59] builds 3D sparse maps by optimizing camera parameters using only RGB images. Traditional approaches consist of handcrafted pipelines to match keypoints [5, 7, 21, 36, 55], form pixel correspondences across multiple images, determine geometric relations, and perform bundle" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 288 + ], + "type": "text", + "content": "adjustment to optimize camera locations. Recently, SfM has incorporated learning-based approaches leveraging neural networks as components [4, 10, 35, 57, 66, 70, 74, 85], but it remains brittle to outliers and challenging conditions. Likewise, multi-view stereo (MVS) aims to reconstruct dense 3D surface through handcrafted pipelines based on triangulation from multiple viewpoints [18, 19, 60]. Learning-based approaches have also been incorporated for MVS to reconstruct 3D from few images [24, 30, 44, 64, 78, 83], and more recently by adopting the RAFT architecture [40] or using generic Transformers [9]. In contrast, the recent DUSt3R framework [31, 75], upon which we build our approach, departs from these methods altogether by defining a simple and unified neural framework to 3D reconstruction based on the paradigm of pointmap regression. The idea is that, ideally, learning-based MVS should be capable of reconstructing 3D from a few images without strictly requiring the availability of camera parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 289, + 553, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 517 + ], + "type": "text", + "content": "Guiding 3D. In this paper, we refer to guidance as the notion that a method can optionally take in 3D priors (typically camera parameters or sparse depth) to help with some prediction task. This can take different forms in the vast literature of 3D vision works. COLMAP [60], for instance, can take camera parameters for rectifying camera views; however, it is an optimization-based method, meaning that each scene has to be processed independently. Recently, UniDepth [46] is a monocular depth and intrinsic estimation method capable of optionally providing camera intrinsics as additional guide to generate depthmap, with an architecture specifically handcrafted for that purpose. Also, depth completion models [68, 69, 76, 80] such as CompletionFormer [26] take sparse depth input as guidance in order to inpaint the full depthmap, but are also typically capable of inferring depth from an image alone. Overall, however, there is currently no learning-based model to the best of our knowledge that guides its output by taking any subset of sparse depth, camera intrinsics and camera extrinsics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 519, + 553, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 553, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 553, + 699 + ], + "type": "text", + "content": "Depth Completion. Given sparse depths and a corresponding RGB image, the task is to densify depth by propagating sparse measurements. CNN-based approaches were adopted by [38, 39], and [11] proposed convolutional spatial propagation network (CSPN). Multi-branch networks, which process RGB and sparse depth separately, were used in [23, 43, 48, 52, 68, 71, 79, 90]. CompletionFormer [26] leverages both CNN and ViT architectures to capture both global and local features. Most methods focus only on the specific depth completion task from a single image and partial depth as the input, often without incorporating camera intrinsics. Using the pointmap representation, Pow3R, based on a ViT architecture, can complete the depth map from both RGB images and sparse point clouds, while taking relative pose or camera intrinsics as auxiliary input." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Pose Estimation from Sparse views. Learning-based ap" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1072" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 295, + 165 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 295, + 165 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 295, + 165 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 295, + 165 + ], + "type": "image", + "image_path": "6b7cba457c5a6465b3b0f24a09a752630003b667442bbfacaf8b712112300c24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 166, + 296, + 233 + ], + "lines": [ + { + "bbox": [ + 55, + 166, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 296, + 233 + ], + "type": "text", + "content": "Figure 2. Overview of the model architecture. Following DUSt3R [75], images are encoded then decoded with a ViT backbone into pointmaps, from which focals, depthmaps and relative pose can be extracted. Pow3R introduces optional inputs to guide the regression with prior knowledge about the camera intrinsics and depth (fed into the encoder) and the pose (into the decoder)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 248, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 248, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 248, + 295, + 344 + ], + "type": "text", + "content": "proaches involve regressing relative pose, applying probabilistic methods, using energy-based or diffusion models [3, 33, 53, 73, 88]. RayDiffusion [89] models camera rays instead of camera parameters, applying diffusion on rays, then recovers the camera pose using DLT [2]. DUSt3R discovers camera pose from pointmaps using PnP, while Pow3R predicts pointmaps in both camera coordinate frames, allowing the use of Procrustes alignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 296, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 537 + ], + "type": "text", + "content": "DUSt3R and its developments. DUSt3R [75] does not require calibrated images for 3D reconstruction, can handle extreme viewpoint changes, and deals with both single and a few images within the same framework. Its pointmap representation eases the training by not having to deal with camera parameters directly. MAST3R [31] improves DUSt3R by also predicting dense local features, trained with an additional matching loss. Splatt3R [65] marries MAST3R with GS. Spann3R [72] predicts pointmaps in global coordinate and adopts an external spatial memory that tracks relevant 3D information. MonST3R [87] further trains DUSt3R on dynamic scenes and generates time-varying dynamic point clouds, as well as camera intrinsics and poses from a dynamic video. In this work, we take a different and complementary direction and instead improve the core capabilities of DUSt3R, which could benefit all aforementioned works." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 549, + 138, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 138, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 138, + 563 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "We aim to train a network " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " that can take two input images " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "I^1, I^2 \\in \\mathbb{R}^{W \\times H \\times 3}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " of a given static scene and any subset of auxiliary information " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Omega \\subseteq \\{K_1, K_2, P_{1,2}, D^1, D^2\\}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", in order to regress a 3D reconstruction of the scene. Here, " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "K_1, K_2 \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " are camera intrinsics, " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "P_{1,2} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " denote the relative pose between the two cameras, and " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "D^1, D^2 \\in \\mathbb{R}^{W \\times H}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " are depthmaps with associated masks " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "M^1, M^2 \\in \\{0,1\\}^{W \\times H}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " specifying pixels with valid depth data (i.e. masks are possibly sparse). The network " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " is tasked to regress several pointmaps from which the camera intrinsic and extrinsic parameters as well as the dense depthmaps can be straightforwardly extracted for both im" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 73, + 416, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 416, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 416, + 84 + ], + "type": "text", + "content": "ages as described below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": "Pointmaps. For each pixel " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": " in an image " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": ", we assume there exists a corresponding single 3D point " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "X_{i,j}" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{W \\times H \\times 3}" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": " is a pointmap. Given camera intrinsics " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": " and a depthmap " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": ", we can compute " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "X_{i,j} = K^{-1}[iD_{i,j},jD_{i,j},D_{i,j}]" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": " in the camera coordinate system. In the following, we denote as " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "X^{n,m}" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": " the pointmap of image " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "I^n" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": " expressed in the coordinate system of camera " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "I^m" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": ". To swap the coordinate system from camera " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "I^k" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": " to camera " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "I^m" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "X^{n,k} = P_{m,k}X^{n,m}" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "inline_equation", + "content": "P_{m,k} = P_kP_m^{-1}" + }, + { + "bbox": [ + 313, + 86, + 555, + 196 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 202, + 432, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 432, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 432, + 213 + ], + "type": "text", + "content": "3.1. Overall architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "text", + "content": "We follow the recently proposed DUSt3R framework [75]. DUSt3R is a breakthrough foundation model based on vision transformers (ViT) that can regress two 3D pointmaps " + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "inline_equation", + "content": "X^{1,1}" + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "inline_equation", + "content": "X^{2,1}" + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "text", + "content": " given solely two unposed and uncalibrated input images. Nevertheless, DUSt3R's inability to leverage potentially available information about cameras or depth limits its practical use. To remedy these drawbacks, we enhance the original DUSt3R network " + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "text", + "content": " in two ways." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 556, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 436 + ], + "type": "text", + "content": "First, we introduce specific modules to incorporate any subset of extra information such as camera intrinsics, camera poses and depthmaps seamlessly into DUSt3R. Second, we predict an additional pointmap " + }, + { + "bbox": [ + 313, + 316, + 556, + 436 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 313, + 316, + 556, + 436 + ], + "type": "text", + "content": ", which represents the pointmap of image " + }, + { + "bbox": [ + 313, + 316, + 556, + 436 + ], + "type": "inline_equation", + "content": "I^2" + }, + { + "bbox": [ + 313, + 316, + 556, + 436 + ], + "type": "text", + "content": " in its own coordinate system. Predicting three pointmaps offers further capabilities, namely the possibility of recovering all information about both cameras in a single forward pass, as explained in Sec. 3.3. We first describe the overall architecture, see Fig. 2, and then detail the dedicated modules to inject auxiliary information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "text", + "content": "Encoder. We use a shared ViT encoder to encode both images independently. In addition to " + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "text", + "content": ", the encoder can receive auxiliary information about intrinsics " + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "text", + "content": " for each image. For the two input images " + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "inline_equation", + "content": "I^1, I^2" + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "text", + "content": " and their respective auxiliary information " + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\Omega_1 \\in \\sigma(\\{K_1, D_1\\}), \\Omega_2 \\in \\sigma(\\{K_2, D_2\\})" + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "text", + "content": " denotes the set of all subsets, the encoder processes the information in a Siamese manner:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 328, + 530, + 555, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 530, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 328, + 530, + 555, + 545 + ], + "type": "interline_equation", + "content": "F ^ {1} = \\operatorname {E n c o d e r} \\left(I ^ {1}, \\Omega_ {1}\\right), F ^ {2} = \\operatorname {E n c o d e r} \\left(I ^ {2}, \\Omega_ {2}\\right). \\tag {1}", + "image_path": "1a946f33f5aa91f35cfe3ce0bf21f6ba31da90a05994f648b3a557b640b82764.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "text", + "content": "Decoder. Likewise, the network has two decoders, each with its corresponding head, one predicting " + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "inline_equation", + "content": "X^{1,1}" + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "text", + "content": " and the other one estimating " + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "inline_equation", + "content": "X^{2,1}" + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "text", + "content": ". Both decoders communicate via cross-attention between their own tokens and the outputs of the previous block of the other decoder. Each decoder may receive the relative pose " + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "inline_equation", + "content": "P_{12}" + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "text", + "content": " as additional input or not. Assume we provide the auxiliary information " + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\Omega_D \\in \\sigma(\\{P_{12}\\})" + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 556, + 555, + 653 + ], + "type": "text", + "content": "-th block of both decoders:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 336, + 661, + 553, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 661, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 336, + 661, + 553, + 693 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} G _ {i} ^ {1} = \\operatorname {D e c o d e r B l o c k} _ {i} ^ {1} \\left(G _ {i - 1} ^ {1}, G _ {i - 1} ^ {2}, \\Omega_ {D}\\right), \\\\ G _ {i} ^ {2} = \\operatorname {D e c o d e r B l o c k} _ {i} ^ {2} \\left(G _ {i - 1} ^ {2}, G _ {i - 1} ^ {1}, \\Omega_ {D}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "45d9aecae18fa5330239cfbc93fb008195c295bee019e45a954a0ab9ccc81514.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": "After " + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": " decoder blocks in each branch, the head regresses" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1073" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 280, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 280, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 280, + 85 + ], + "type": "text", + "content": "several pointmaps and their associated confidence map:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 130, + 92, + 244, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 92, + 244, + 107 + ], + "spans": [ + { + "bbox": [ + 130, + 92, + 244, + 107 + ], + "type": "interline_equation", + "content": "X ^ {1, 1}, C ^ {1, 1} = \\operatorname {H e a d} ^ {1} \\left(G _ {B} ^ {1}\\right),", + "image_path": "d840626d45c7aaa8e90fe1e829082180863ab5d4c76808a4c884c27dc49410b7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 85, + 109, + 295, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 109, + 295, + 125 + ], + "spans": [ + { + "bbox": [ + 85, + 109, + 295, + 125 + ], + "type": "interline_equation", + "content": "X ^ {2, 1}, X ^ {2, 2}, C ^ {2, 1}, C ^ {2, 2} = \\operatorname {H e a d} ^ {2} \\left(G _ {B} ^ {2}\\right). \\tag {3}", + "image_path": "0fc52a61e2b6936e2a470bbf4f0005191b54d5e1d21d5cbd824754b59fcc96e4.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 132, + 295, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 132, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 132, + 295, + 168 + ], + "type": "text", + "content": "Unlike DUSt3R which uses a DPT head [50], here we employ a linear head for both " + }, + { + "bbox": [ + 55, + 132, + 295, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{Head}^1" + }, + { + "bbox": [ + 55, + 132, + 295, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 132, + 295, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{Head}^2" + }, + { + "bbox": [ + 55, + 132, + 295, + 168 + ], + "type": "text", + "content": " as it is more efficient and performs similarly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": "3D Regression Loss. Following DUSt3R, we supervise the model to minimize the distance between ground-truth and predicted pointmaps in a scale-invariant manner, allowing the model to train on multiple datasets with various scales. The regression loss between predicted and ground-truth pointmaps (resp. " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "X^{n,m}" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "\\hat{X}^{n,m}" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": ") at pixel " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{i,j}^{\\mathrm{gr}}(n,m) = \\| X_{i,j}^{n,m} / z_m - \\hat{X}_{i,j}^{n,m} / \\hat{z}_m\\|" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "z_{m},\\hat{z}_{m}" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": " serve as scale normalizer. That is, " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "z_{m}" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": " is the average norm of all valid 3D points expressed in coordinate system of image " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "I^{m}" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "z_{1} = \\mathrm{norm}(X^{1,1}\\cup X^{2,1})" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "z_{2} = \\mathrm{norm}(X^{2,2})" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": " and likewise for " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "\\hat{z}_1,\\hat{z}_2" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "\\mathrm{norm}(X) = \\mathrm{mean}(\\| X_{i,j}\\| |i,j\\in \\mathcal{D}_X)" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_X" + }, + { + "bbox": [ + 55, + 170, + 296, + 314 + ], + "type": "text", + "content": " the set of valid pixels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "text", + "content": "Confidence-aware loss. The model jointly learns to predict a confidence level " + }, + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "inline_equation", + "content": "C_{i,j}^{n,m}" + }, + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "text", + "content": " of each pixel " + }, + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "text", + "content": ". The confidence-aware regression loss for a given pointmap " + }, + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "inline_equation", + "content": "X^{n,m}" + }, + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "text", + "content": " can be expressed as the 3D regression loss " + }, + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{reg}}" + }, + { + "bbox": [ + 55, + 315, + 296, + 374 + ], + "type": "text", + "content": " weighted by the confidence map:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 383, + 295, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 295, + 409 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {c o n f}} (n, m) = \\sum_ {i, j \\in \\mathcal {D}} C _ {i, j} ^ {n, m} \\mathcal {L} _ {i, j} ^ {\\text {r e g r}} (n, m) - \\alpha \\log C _ {i, j} ^ {n, m}, \\tag {4}", + "image_path": "d42d2a625455826f245abb5b65a96147c1b1d9e6f72ec5fab07bfd1eca0518cd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 418, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 453 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 418, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\alpha = 0.2" + }, + { + "bbox": [ + 55, + 418, + 296, + 453 + ], + "type": "text", + "content": ". This loss penalizes the model less when the prediction is not accurate on harder areas, encouraging the model to extrapolate. The final loss is expressed as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 462, + 295, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 462, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 84, + 462, + 295, + 476 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} ^ {\\text {c o n f}} (1, 1) + \\mathcal {L} ^ {\\text {c o n f}} (2, 1) + \\beta \\mathcal {L} ^ {\\text {c o n f}} (2, 2), \\tag {5}", + "image_path": "2983c95a2da4a39904a3164ae5c2d4c5c8ac3109d075873bf7c2a7f80ee67ce5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 485, + 266, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 266, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 266, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 485, + 266, + 498 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 485, + 266, + 498 + ], + "type": "text", + "content": " is a hyper-parameter typically set to " + }, + { + "bbox": [ + 55, + 485, + 266, + 498 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 55, + 485, + 266, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 505, + 222, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 222, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 222, + 517 + ], + "type": "text", + "content": "3.2. Adding Versatile Conditioning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": "The knowledge of auxiliary information can significantly enhance 3D predictions at test time. Our model leverages up to five different modalities, which include two intrinsics " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "K_{1}, K_{2}" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": ", two depthmaps " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "D^{1}, D^{2}" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": " for each image and the relative pose " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "P_{12}" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": ". To condition the output on it, we first embed the auxiliary information using dedicated MLPs and then inject these embeddings at different points in the pipeline." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": "We discuss different options in this regard. In the first strategy, denoted as 'embed', we simply add the auxiliary embeddings to the token embeddings before the first transformer block, as would be done for standard positional embeddings [16]. In the second strategy, denoted as 'inject- " + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": " ' we instead insert dedicated MLPs for each modality in a subset of " + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": " transformer blocks, as shown in Fig. 3. Our experiments show that 'inject-1' performs slightly better than 'embed' and similarly with 'inject- " + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": " ' where " + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "n > 1" + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": ", see" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 326, + 72, + 536, + 148 + ], + "blocks": [ + { + "bbox": [ + 326, + 72, + 536, + 148 + ], + "lines": [ + { + "bbox": [ + 326, + 72, + 536, + 148 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 536, + 148 + ], + "type": "image", + "image_path": "ab31b72c696d7f4c4a54a16400083834c04f94c8bb7b56e7c245739b0289e580.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 326, + 152, + 537, + 219 + ], + "blocks": [ + { + "bbox": [ + 326, + 152, + 537, + 219 + ], + "lines": [ + { + "bbox": [ + 326, + 152, + 537, + 219 + ], + "spans": [ + { + "bbox": [ + 326, + 152, + 537, + 219 + ], + "type": "image", + "image_path": "c77f3013c264b67a5524128d9598d745fffda9b53e92dc598ef6661508aa0cfd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 220, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 331 + ], + "type": "text", + "content": "Figure 3. Architecture for injecting auxiliary modalities. Top: injection of optional intrinsics and depth into the encoder. Intrinsic s are encoded into ray patches, sparse depth is patchified. Each of these modalities goes into a block-specific MLP and are tokenwise added in the middle of the transformer block. Bottom: injection of optional relative pose into the decoder. The relative pose is fed to a first embedding layer followed by an MLP. This token is added to the CLS token of the decoder after the self-attention (SA) and cross-attention (CA), but before the MLP. Experiments show that injection in the first block only suffices." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 342, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 366 + ], + "type": "text", + "content": "Sec. 4.5. We now describe how to compute the initial embeddings for each specific modality." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 369, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 475 + ], + "type": "text", + "content": "Intrinsics. Following [89], we generate camera rays from the intrinsic matrix " + }, + { + "bbox": [ + 313, + 369, + 554, + 475 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 313, + 369, + 554, + 475 + ], + "type": "text", + "content": ", thereby establishing a direct correspondence between RGB pixels and rays. The ray at pixel location " + }, + { + "bbox": [ + 313, + 369, + 554, + 475 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 369, + 554, + 475 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 313, + 369, + 554, + 475 + ], + "type": "inline_equation", + "content": "K^{-1}[i,j,1]" + }, + { + "bbox": [ + 313, + 369, + 554, + 475 + ], + "type": "text", + "content": " and encodes the viewing direction of that pixel w.r.t. the current camera pose. This allows us to potentially process non-centered crops and hence to perform inference in higher image resolutions, see Sec. 3.3. Similarly to RGB inputs, we patchify and embed dense rays and provide them to the encoder." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "text", + "content": "Depthmaps / Point Clouds. Given a depthmap " + }, + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "text", + "content": " and its sparsity mask " + }, + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "text", + "content": ", we first normalize " + }, + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "inline_equation", + "content": "D' = D / \\mathrm{norm}(D)" + }, + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "text", + "content": " to handle any depth ranges at train and test time. As for RGB and rays, we patchify the stacked maps " + }, + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "inline_equation", + "content": "[D', M] \\in \\mathbb{R}^{W \\times H \\times 2}" + }, + { + "bbox": [ + 313, + 478, + 554, + 562 + ], + "type": "text", + "content": " and compute patch embeddings, which are then fed to the encoder. By jointly patchifying the depth and its valid mask we can handle any level of sparsity." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "content": "Camera Pose. Given the relative pose " + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "inline_equation", + "content": "P_{12} = [R_{12}|t_{12}]" + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "content": ", we normalize the translation scale as " + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "inline_equation", + "content": "t_{12}' = t_{12} / \\|t_{12}\\|" + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "content": " since our output is unscaled. Unlike depthmaps or camera intrinsics, the camera pose cannot be expressed as a dense pixel map. Rather, camera pose affects the whole pixels between two images, so the embedding is instead added to the global CLS token of both decoders." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 658, + 425, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 658, + 425, + 669 + ], + "spans": [ + { + "bbox": [ + 314, + 658, + 425, + 669 + ], + "type": "text", + "content": "3.3. Downstream Tasks" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Depthmaps. In the pointmap representation, the " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "-axis of " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "X^{1,1}, X^{2,2}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " directly corresponds to the depthmaps of the first and second image, respectively. Note that DUSt3R" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1074" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 236, + 184 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 236, + 184 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 236, + 184 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 236, + 184 + ], + "type": "image", + "image_path": "17d580084d03bb1334327108baa4f1b493068a4a757a8cba3f3467c63d8eaee0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 555, + 231 + ], + "type": "text", + "content": "Figure 4. Native resolution example. Methods like DUSt3R require centered principal point and would thus downscale the image to its training resolution, leading to blurry prediction (top). With camera intrinsics input, we can process any crop in the image leading to native resolution prediction (bottom row), for example by dividing the image into 4 crops, that are independently process as the first input image, while the second input image of the network is set to the downscaled image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 70, + 469, + 185 + ], + "blocks": [ + { + "bbox": [ + 238, + 70, + 469, + 185 + ], + "lines": [ + { + "bbox": [ + 238, + 70, + 469, + 185 + ], + "spans": [ + { + "bbox": [ + 238, + 70, + 469, + 185 + ], + "type": "image", + "image_path": "babbe8245315e7913c89d97a8f11bc2104888d68a413b018c1eddb82ae7f6cdf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 470, + 70, + 553, + 185 + ], + "blocks": [ + { + "bbox": [ + 470, + 70, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 470, + 70, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 470, + 70, + 553, + 185 + ], + "type": "image", + "image_path": "01f03d6f78e2f1123369db61f5134fd05c3d7451b82eb8111f8cfe6681fda966.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "type": "text", + "content": "only outputs " + }, + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "type": "inline_equation", + "content": "X^{1,1}" + }, + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "type": "text", + "content": ", hence it requires two forward passes of the pairs " + }, + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "type": "inline_equation", + "content": "(I^1, I^2)" + }, + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "type": "inline_equation", + "content": "(I^2, I^1)" + }, + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "type": "text", + "content": " to extract both depthmaps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 270, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 474 + ], + "type": "text", + "content": "High-resolution processing. A limitation of DUSt3R is that it is trained for a specific resolution and does not generalize to larger resolutions. A simple fix would be to include higher resolutions during training, but (i) many datasets do not provide high-resolution ground-truths, and (ii) the training cost could become prohibitive. Pow3R, in contrast, can handle crops natively given camera intrinsic's of the crop, as these provide the crop position information (i.e. via focal length and principal point). We can thus perform prediction in a sliding window fashion, yielding predictions matching any target resolution by simple stitching. Comparison between standard and higher resolution processing are shown in Fig. 4. Note that prediction for each crop may have a different scale, by design, and cannot be stitched directly. We find that simply computing the median scale factor in overlapping areas is sufficient, and we blend the overlapping regions based on confidence without further post-processing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 475, + 295, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 547 + ], + "type": "text", + "content": "Focal Estimation. Similar to DUSt3R, we can recover fccals for both input images from pointmap " + }, + { + "bbox": [ + 55, + 475, + 295, + 547 + ], + "type": "inline_equation", + "content": "X^{1,1}" + }, + { + "bbox": [ + 55, + 475, + 295, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 475, + 295, + 547 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 55, + 475, + 295, + 547 + ], + "type": "text", + "content": " with the robust Weiszfeld fast iterative solver [47]. Similar to depthmap prediction, DUSt3R needs to infer " + }, + { + "bbox": [ + 55, + 475, + 295, + 547 + ], + "type": "inline_equation", + "content": "(I_2, I_1)" + }, + { + "bbox": [ + 55, + 475, + 295, + 547 + ], + "type": "text", + "content": " in order to compute the focal of the second image, while Pow3R predicts them in a single pass." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 547, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 295, + 603 + ], + "type": "text", + "content": "Relative Pose Estimation. Pow3R predicts the relative pose directly by Procrustes alignment [8, 37] to get the scaled relative pose " + }, + { + "bbox": [ + 55, + 547, + 295, + 603 + ], + "type": "inline_equation", + "content": "P^{*} = \\left[R^{*}|t^{*}\\right]" + }, + { + "bbox": [ + 55, + 547, + 295, + 603 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 55, + 547, + 295, + 603 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 55, + 547, + 295, + 603 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 547, + 295, + 603 + ], + "type": "inline_equation", + "content": "X^{2,1}" + }, + { + "bbox": [ + 55, + 547, + 295, + 603 + ], + "type": "text", + "content": ", as it predicts the pointmaps of the second image in two different camera coordinates." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 609, + 295, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 609, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 61, + 609, + 295, + 634 + ], + "type": "interline_equation", + "content": "R ^ {*}, t ^ {*} = \\arg \\min _ {\\sigma , R, t} \\sum_ {i, j} \\sqrt {C _ {i , j} ^ {2 , 2} C _ {i , j} ^ {2 , 1}} \\left\\| \\sigma \\left(R X _ {i, j} ^ {2, 2} + t\\right) - X _ {i, j} ^ {2, 1} \\right\\| ^ {2}. \\tag {6}", + "image_path": "a128ceca8fa1b31df383b8b1fe134984266b5faf28b7ad4e2289dd123729a9be.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 640, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 296, + 688 + ], + "type": "text", + "content": "Procrustes alignment is known to be sensitive to noise and outliers; however, in our case, it performs similar to RANSAC [17] with PnP [22, 29], while being an order of magnitude faster (see Sec. 4.4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Global alignment. The network " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " predicts pointmaps for image pairs. To align all predictions in the same world coor" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 245, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 316 + ], + "type": "text", + "content": "dinate system, we resort to the global alignment algorithm from DUSt3R [75] that minimizes a global energy function to find per-camera intrinsics, depthmaps and poses that are consistent with all the pairwise predictions. Results of the optimization are global scene point-clouds, which can for instance serve for multi-view stereo estimation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 325, + 429, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 429, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 429, + 338 + ], + "type": "text", + "content": "3.4. Training Procedure" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "text", + "content": "During training, we feed the network with annotated image associated with random subsets of auxiliary information, the goal being for it to learn to handle any situation at test time. For each pair, we first chose a random number " + }, + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "text", + "content": " of modalities with uniform probability, and then randomly select the " + }, + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "text", + "content": " modalities likewise. Depthmaps are randomly sparsified. When giving intrinsics, we also perform aggressive non-centered cropping with " + }, + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "text", + "content": " probability, so that the network learns to perform high-resolution inference." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 452, + 556, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 556, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 556, + 584 + ], + "type": "text", + "content": "Training data and recipe. We follow the training recipe from DUSt3R [75] in terms of training data and protocol. Namely, we use the provided pairs for 8 datasets: Habitat [58], MegaDepth [32], ARKitScenes [6], Static Scenes 3D [41], BlendedMVS [82], ScanNet++ [84], Co3Dv2 [51] and Waymo [67], resulting in a total of 8.5M image pairs. These datasets include indoor and outdoor scenes, as well as real and synthetic ones. To train the model, we follow [75] and first train the network with a resolution of 224px for 3 days on 8 A100 GPUs, and then finetune it in 512px with variable aspect-ratio for 2 more days." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 595, + 459, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 595, + 459, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 459, + 609 + ], + "type": "text", + "content": "4. Experimental Evaluation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 616, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 664 + ], + "type": "text", + "content": "We first study the impact of auxiliary information guidance (Sec. 4.1). We then experiment across many 3D vision tasks (Secs. 4.2 to 4.4) without retraining, i.e. mostly zero-shot settings. We finally present some ablations in Sec. 4.5." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "DUSt3R baseline. We point out that our overall training data, protocols and architecture are exactly the same as DUSt3R, except for the additional head that predict " + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "inline_equation", + "content": "+0.1\\%" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": " parameters) and for the modules used for injecting" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1075" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 294, + 197 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 197 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 197 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 197 + ], + "type": "table", + "html": "
aux. modalitiesfocal acc@1.015depth τ@1.03rel. pose
K1K2D1D2RTRRA@2°RTA@2°
DUSt3R×××××36.677.669.249.7
Pow3R×××××39.479.471.953.8
✓××××75.4(+36.0)80.6 (+1.2)77.1 (+5.3)62.8 (+8.9)
×✓×××67.3(+27.9)80.3 (+0.9)74.5 (+2.6)59.1 (+5.2)
✓✓×××98.0(+58.6)81.4 (+2.0)80.3 (+8.5)74.2(+20.3)
××✓××48.9 (+9.5)89.1 (+9.7)82.5(+10.6)64.9(+11.0)
×××✓×49.5(+10.1)91.0(+11.7)83.4(+11.5)64.8(+10.9)
××✓✓×58.2(+18.9)95.4(+16.0)89.6(+17.7)77.6(+23.8)
××××✓48.6 (+9.2)81.6 (+2.2)92.3(+20.5)77.0(+23.1)
✓✓✓✓×99.2(+59.8)95.4(+16.0)95.7(+23.9)94.3(+40.5)
✓✓××✓98.1(+58.7)82.9 (+3.6)95.1(+23.2)87.3(+33.5)
××✓✓✓68.6(+29.2)95.4(+16.0)98.1(+26.2)91.3(+37.5)
✓✓✓✓✓99.3(+59.9)95.4(+16.0)99.0(+27.2)98.1(+44.3)
", + "image_path": "ec9bc7837c5d6c28fbacbabe7fc824f95cc113bb8b1dbc2d04b4585431661af0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 198, + 295, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 255 + ], + "type": "text", + "content": "Table 1. Impact of guiding at test time for models trained at " + }, + { + "bbox": [ + 55, + 198, + 295, + 255 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 198, + 295, + 255 + ], + "type": "text", + "content": " resolution. We report performances on Habitat for DUSt3R, which cannot handle auxiliary modalities, and our single model with different sets of modalities; we show in green the absolute improvement w.r.t. the results without auxiliary modality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 264, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 295, + 289 + ], + "type": "text", + "content": "the auxiliary information " + }, + { + "bbox": [ + 55, + 264, + 295, + 289 + ], + "type": "inline_equation", + "content": "(+4\\%" + }, + { + "bbox": [ + 55, + 264, + 295, + 289 + ], + "type": "text", + "content": " parameters for the inject-1 variant that we use unless stated otherwise)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 289, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 295, + 410 + ], + "type": "text", + "content": "Metrics. When evaluating focal lengths, we report the accuracy for a tight relative error threshold of " + }, + { + "bbox": [ + 54, + 289, + 295, + 410 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 54, + 289, + 295, + 410 + ], + "type": "text", + "content": ". For evaluating depthmap, we report the Absolute Relative Error (rel) and the Inlier Ratio " + }, + { + "bbox": [ + 54, + 289, + 295, + 410 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 54, + 289, + 295, + 410 + ], + "type": "text", + "content": " at a threshold of " + }, + { + "bbox": [ + 54, + 289, + 295, + 410 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 54, + 289, + 295, + 410 + ], + "type": "text", + "content": ". When evaluating poses, we opt for scale-invariant metrics and measure the relative rotation and translation accuracies (RRA and RTA resp.) of the relative pose at task-specific thresholds, and the mean Average Accuracy (mAA), defined as the area under the curve accuracy of the angular differences at min(RRA, RTA)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 418, + 219, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 219, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 219, + 431 + ], + "type": "text", + "content": "4.1. Guiding the output prediction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 437, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 593 + ], + "type": "text", + "content": "Impact of guidance. We first study the impact of providing different subsets of auxiliary information in terms of downstream 3D vision performance. Namely, we report accuracies in Tab. 1 for focal, depth, and rotation prediction on the Habitat validation dataset. In the supplementary material, we also provide results on the Infinigen dataset [49], which is not included in our training set, and observe similar outcomes. We first note that in the absence of modalities, Pow3R performs slightly better than a DUSt3R baseline, even though Pow3R is trained with just RGB images during a fraction of its training. This shows that Pow3R does not lose its native strength for unconstrained regression despite having gained more capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Interestingly, we observe in Tab. 1 the impact of each modality for each prediction task, and the synergetic effects of interacting modalities. For instance, the model performs better for focal length estimation with depthmaps and/or relative pose, and Pow3R's depth prediction improves with intrinsics and/or relative pose. Moreover, relative pose prediction becomes more accurate with intrinsics and/or depthmap. Overall, the model's performance consistently improves when provided with additional camera or scene priors, demonstrating the effectiveness of Pow3Rin" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 553, + 163 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 163 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 163 + ], + "type": "image", + "image_path": "826dd2e89c2688b8204dd337cb9dc4237dd5a3a1d8bc67d6b14bb35cdffc7c4f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 163, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 314, + 163, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 314, + 163, + 553, + 186 + ], + "type": "text", + "content": "Figure 5. Depth completion performance on NYUd for various input depth sparsity." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 193, + 553, + 266 + ], + "blocks": [ + { + "bbox": [ + 317, + 193, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 317, + 193, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 317, + 193, + 553, + 266 + ], + "type": "table", + "html": "
aux. mod.High-Res.KITTIT&TTime (sec)↓
KsRTrel↓τ↑rel↓τ↑
DUSt3R512×××5.449.53.375.10.13
Pow3R512✓✓×5.348.73.278.20.13
Pow3R512✓✓(n)7.534.43.968.00.48
Pow3R512✓✓✓4.653.52.582.30.40
", + "image_path": "a20042584d1526d1423e3699a3a4d22b569cfd97994cdbf5f1648b32d0d9f91d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 267, + 553, + 311 + ], + "lines": [ + { + "bbox": [ + 313, + 267, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 553, + 311 + ], + "type": "text", + "content": "Table 2. Multi-view depth estimation results in high-resolution using a sliding-window scheme and providing the crop intrinsics to Pow3R. (n): naively inputting high-resolution images without downsampling to Pow3R." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 323, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 360 + ], + "type": "text", + "content": "handling any combination of priors in a unified fashion. We are not aware of any other approach that can seamlessly integrate all these priors in a learning-based method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 361, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 517 + ], + "type": "text", + "content": "Depth completion. We also evaluate Pow3R on the depth completion task given RGB images with sparse depthmaps. We vary the sparsity (the ratio of pixels with known depth) and demonstrate its impact on the NYUv2 dataset [63] in Figure 5. Although Pow3R is not trained on the NYUv2 dataset, it still achieves state-of-the-art results for sparse depth completion. Our analysis indicates that the true performance of Pow3R on NYUv2 is likely underestimated due to numerous errors in the ground-truth depth annotations; further details are available in the supplementary material. On the KITTI [20] dataset, we obtain a mean absolute error below " + }, + { + "bbox": [ + 313, + 361, + 555, + 517 + ], + "type": "inline_equation", + "content": "30\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 361, + 555, + 517 + ], + "type": "text", + "content": ", which is competitive with the state of the art, considering that KITTI is not included in the training set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 519, + 554, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 554, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 554, + 675 + ], + "type": "text", + "content": "High-Resolution processing. With the possibility of specifying camera intrinsics, new capabilities arise. Namely, Pow3R can perform high-resolution pointmap regression with a simple sliding-window approach. We process fixed-resolution crops of the original images under the guidance of both (i) the crop information, provided in term of camera intrinsics, and (ii) a second image, itself possibly down-scaled or cropped as well. More details are provided in Sec. 3.3 and Fig. 4. When applied, results significantly improve compared to processing the images at 512-pixel resolution, as shown in Tab. 2. We also compare with a naive baseline where we simply feed high-resolution images directly to Pow3R, but it dramatically fails." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Camera controllability. We study how controllable are the regression outputs on 10K Habitat test pairs given arbitrary camera parameters. For each test pair with ground-truth fo" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1076" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 175, + 160 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 175, + 160 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 175, + 160 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 175, + 160 + ], + "type": "image", + "image_path": "f5fa1fa4f400cd95e2de9546aeab43e83115042a21d3c66b1bba5662cec59172.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 72, + 293, + 160 + ], + "blocks": [ + { + "bbox": [ + 178, + 72, + 293, + 160 + ], + "lines": [ + { + "bbox": [ + 178, + 72, + 293, + 160 + ], + "spans": [ + { + "bbox": [ + 178, + 72, + 293, + 160 + ], + "type": "image", + "image_path": "3b1c173f2a5a1298768c924cd3cf424f9cf3cd32a8bf924bf5780380fdc36fa5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 163, + 174, + 252 + ], + "blocks": [ + { + "bbox": [ + 58, + 163, + 174, + 252 + ], + "lines": [ + { + "bbox": [ + 58, + 163, + 174, + 252 + ], + "spans": [ + { + "bbox": [ + 58, + 163, + 174, + 252 + ], + "type": "image", + "image_path": "b2d66dbb9e23093374d1bf4125e289508593aa82cbf9c68e7b0aaa3576c439e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 253, + 295, + 330 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 330 + ], + "type": "text", + "content": "Figure 6. Controllability study. We study how the model reacts when presented with false focal (top) and false relative pose (bottom). We plot the input vs. predicted camera parameters in the left column and the average confidence in the right column. Interestingly, the model seems to not take the input information at face value and instead makes an informed decision whether to use it or not, in which latter case the confidence becomes low." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 178, + 163, + 293, + 252 + ], + "blocks": [ + { + "bbox": [ + 178, + 163, + 293, + 252 + ], + "lines": [ + { + "bbox": [ + 178, + 163, + 293, + 252 + ], + "spans": [ + { + "bbox": [ + 178, + 163, + 293, + 252 + ], + "type": "image", + "image_path": "dd045e1f9063569778ba0a83867a01cfd78d48c4a4366ef39388058716586544.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "text", + "content": "cal " + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "text", + "content": " and relative pose " + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "inline_equation", + "content": "P_{12}" + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "text", + "content": ", we purposefully input a wrong focal (resp. relative pose) and measure the output focal " + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "text", + "content": " (resp. pose " + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "inline_equation", + "content": "\\hat{P}_{12}" + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "text", + "content": ") as predicted from the raw pointmaps. Results are presented in Fig. 6 in terms of focal ratio " + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "inline_equation", + "content": "\\hat{f} / f" + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "text", + "content": " and angle " + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "inline_equation", + "content": "d_{\\text{geodesic}}(P_{12}, \\hat{P}_{12})" + }, + { + "bbox": [ + 54, + 346, + 295, + 550 + ], + "type": "text", + "content": ". Interestingly, the model accepts to follow the guidance whenever it is close to the ground truth value (i.e. staying close to the identity transformation) but then refuses to follow it when it gets too different, showing that the model is able to somehow 'judge' of the guidance's worth. In all cases, the output confidence clearly correlates with the agreement between the guidance and the ground truth. Overall, the model is thus not taking the auxiliary information at face value, and instead combines it with the other modalities (e.g. RGB) it in a more subtle and informed way than what could be achieved with handcrafted methods. This suggests that Pow3R would be robust to noise in the auxiliary information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 560, + 217, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 217, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 217, + 572 + ], + "type": "text", + "content": "4.2. Multi-View Depth Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 675 + ], + "type": "text", + "content": "Following RobustMVD [62], performances are measured on the KITTI [20], ScanNet [14], ETH3D [61], DTU [1], and Tanks and Temples [28] datasets. We report in Tab. 3 the Absolute Relative Error (rel) and Inlier Ratio " + }, + { + "bbox": [ + 55, + 578, + 295, + 675 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 55, + 578, + 295, + 675 + ], + "type": "text", + "content": " for each dataset when using (1) images alone, (2) images with relative poses, (3) images with camera intrinsics, and (4) images with both modalities, and provide results for more state-of-the-art methods in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Results. Pow3R again performs roughly on par with DUSt3R in the absence of auxiliary information (e.g. 3.73 average error across all datasets for DUSt3R-512 vs. 3.64" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 553, + 176 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 553, + 176 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 176 + ], + "type": "table", + "html": "
MethodGT intrinsicsCo3Dv2RealEstate10K mAA(30)Speed (fps)
RRA@15RTA@15mAA(30)
Colmap+SG [56]✓36.127.325.345.2-
PixSfM [34]✓33.732.930.149.4-
RelPose [88]×57.1----
(a) PoseDiff [73]×80.579.866.548.0-
RelPose++ [33]×85.5----
RayDiff [89]×93.3----
DUSt3R (PnP) [75]×94.388.477.261.73.2
(b) Pow3R (PnP)×94.889.978.562.53.2
Pow3R (Pro)×94.690.378.166.330.9
Pow3R w/ K (Pro)✓95.092.182.272.530.1
", + "image_path": "6d0b83e7a49f5d04069a94b1b91281bf4b53f2a6d81a8816b94fd4f470864a3a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 177, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 243 + ], + "type": "text", + "content": "Table 5. Multiview pose estimation on the CO3Dv2 and RealEstate10K. (a) denotes multiview methods, while (b) denotes pairwise methods. Pow3R with Procrustes alignment (Pro) performs better than DUSt3R (PnP) and other multi-view based pose estimation approaches, and even more so with known intrinsics, while being much faster. RealEstate10K is not in the training set." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "type": "text", + "content": "for Pow3R-512). However, it is clear that providing auxiliary information consistently improves the depth estimates. In fact, Pow3R-512 with both pose and intrinsics achieves state-of-the-art performance on the ETH3D, DTU, and Tanks and Temples datasets, as well as on average across datasets. Interestingly, intrinsics tend to improve performance more than the GT pose, which alone seems to have a negligible impact. This might be due to the fact that camera pose primarily helps the prediction of " + }, + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "type": "inline_equation", + "content": "X^{2,1}" + }, + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "type": "text", + "content": ", while the depthmap of this task is obtained from the " + }, + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "type": "text", + "content": " axis of " + }, + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "type": "inline_equation", + "content": "X^{1,1}" + }, + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "type": "text", + "content": " alone. Still, we observe a clear synergy in providing jointly intrinsics and relative pose." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 422, + 423, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 423, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 423, + 434 + ], + "type": "text", + "content": "4.3. Multi-View Stereo" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 445, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 552 + ], + "type": "text", + "content": "We evaluate the quality of global 3D reconstructions on the DTU [1] benchmark in the presence and absence of auxiliary information. For each subset of available information from " + }, + { + "bbox": [ + 313, + 445, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\{(K_1, K_2), P_{12}\\}" + }, + { + "bbox": [ + 313, + 445, + 555, + 552 + ], + "type": "text", + "content": ", we extract pairwise predictions from image pairs and align them in a global coordinate system, as described in Sec 3.3. Performance is evaluated in terms of accuracy, which is the smallest Euclidean distance to the ground-truth, and completeness as the smallest distance to the reconstructed shape, with the overall average." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Results are presented in Tab. 4. We first attempt at reproducing the original DUSt3R results using the official codebase and obtain slightly worse results than reported in [75]. Using the same codebase, Pow3R with RGB only is able to improve over these and match the original results. Interestingly, adding ground-truth intrinsics results in a significant boost (19% rel. improvement), but adding the camera poses instead slightly degrades the results (-4%). However, using both intrinsics and pose results in a major improvement (36% rel. improvement), demonstrating a synergetic effect between both modalities. We hypothesize that the pose is of little value for easy cases like DTU, and that it does not help the model to properly resolve the focals." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1077" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 380, + 180 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 380, + 180 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 380, + 180 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 380, + 180 + ], + "type": "table", + "html": "
MethodGT rangeKITTIScanNetETH3DDTUT&TAverage
rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑rel↓τ↑
COLMAP [59, 60] (K+RT)×12.058.214.634.216.455.10.796.52.795.09.367.8
COLMAP Dense [59, 60] (K+RT)×26.952.738.022.589.823.220.869.325.776.440.248.8
MVSNet [81] (K+RT)✓18.630.722.720.921.635.6(1.8)(86.7)6.574.614.249.7
Vis-MVSNet [86] (K+RT)✓9.555.48.933.510.843.3(1.8)(87.4)4.187.27.061.4
MVS-Former++ [9] (K+RT)✓29.215.215.221.921.432.5(1.2)(91.9)7.671.514.946.6
CER-MVS [40] (K+RT)×14.332.221.124.311.747.54.171.36.482.111.551.5
DUSt3R [75]×5.449.5(3.1)(71.8)3.076.03.968.63.375.13.768.2
Pow3R×5.745.7(3.2)(68.8)3.074.73.074.33.376.63.668.0
Pow3R w/ RT×5.745.8(3.2)(69.7)2.975.63.371.63.277.93.768.1
Pow3R w/K×5.348.3(3.1)(70.8)2.976.01.689.93.277.33.272.5
Pow3R w/K+RT×5.348.7(3.1)(71.4)2.877.11.591.13.278.23.273.3
", + "image_path": "abf11af29b525001d32d11e8e4a069972d86ef1e92d19f080e9269fb9e9e163d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 182, + 380, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 380, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 380, + 216 + ], + "type": "text", + "content": "Table 3. Multi-view depth evaluation: Pow3R with both pose and intrinsics performs better than DUSt3R. More comparisons with the state of the art can be found in the supplementary material. (Parentheses) denote training on data from the same domain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 387, + 72, + 550, + 146 + ], + "blocks": [ + { + "bbox": [ + 387, + 72, + 550, + 146 + ], + "lines": [ + { + "bbox": [ + 387, + 72, + 550, + 146 + ], + "spans": [ + { + "bbox": [ + 387, + 72, + 550, + 146 + ], + "type": "table", + "html": "
MethodAcc.↓Comp.↓Overall↓
DUSt3R†[75]2.6770.8051.741
DUSt3R (repr.)2.1911.5981.894
Pow3R2.1161.3701.743
Pow3Rw/ K1.7221.1191.420
Pow3Rw/ Rt2.2051.4291.817
Pow3Rw/ K+RT1.3840.8461.115
", + "image_path": "86d00f68540f876391859338965f036f8644d4d865cecfccc47ccc2b1ba8f7a2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 384, + 148, + 552, + 213 + ], + "lines": [ + { + "bbox": [ + 384, + 148, + 552, + 213 + ], + "spans": [ + { + "bbox": [ + 384, + 148, + 552, + 213 + ], + "type": "text", + "content": "Table 4. MVS results with the accuracy, completeness and their average (in mm) on the DTU dataset. " + }, + { + "bbox": [ + 384, + 148, + 552, + 213 + ], + "type": "inline_equation", + "content": "\\mathrm{DUSt3R}^{\\dagger}" + }, + { + "bbox": [ + 384, + 148, + 552, + 213 + ], + "type": "text", + "content": " means results from the original paper, and 'repr.' means reproduced with the official public codebase and checkpoint." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 228, + 207, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 207, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 207, + 239 + ], + "type": "text", + "content": "4.4. Multi-View Pose estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 245, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 389 + ], + "type": "text", + "content": "We then evaluate Pow3R for the task of multi-view pose prediction on the Co3Dv2 [51] and RealEstate10K [91] datasets. Co3Dv2 comprises over 37,000 videos across MS-COCO 51 categories with ground-truth camera derived from running COLMAP on 200 frames per video. RealEstate10k contains 80,000 YouTube video clips, containing indoor/outdoor scenes, with camera poses obtained via SLAM followed by Bundle Adjustment. For Co3Dv2, we follow the protocol of [73] and for RealEstate10K, we use its test set of 1,800 videos. For each sequence, we randomly select 10 frames and feed 45 pairs to Pow3R. We report the results in terms of RRA, RTA and mAA in Tab. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 390, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 522 + ], + "type": "text", + "content": "Results. Pow3R predicts " + }, + { + "bbox": [ + 55, + 390, + 296, + 522 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 55, + 390, + 296, + 522 + ], + "type": "text", + "content": ", allowing to estimate the relative pose between 2 images in a single forward using Procrustes Alignment (see Sec. 3.3). In contrast, DUSt3R does not output " + }, + { + "bbox": [ + 55, + 390, + 296, + 522 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 55, + 390, + 296, + 522 + ], + "type": "text", + "content": " and relies on PnP to obtain camera poses. Pow3R with PnP achieves slightly better results than DUSt3R with PnP. Using Procrustes alignment yields similar results, while being orders of magnitude faster than RANSAC, with a roughly " + }, + { + "bbox": [ + 55, + 390, + 296, + 522 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 55, + 390, + 296, + 522 + ], + "type": "text", + "content": " speed-up overall when combining inference and pose estimation times. Finally, performance improves even further with known intrinsic, significantly outperforming all the state of the art." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 529, + 178, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 178, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 178, + 540 + ], + "type": "text", + "content": "4.5. Architecture ablation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "text", + "content": "We evaluate the impact of adding the " + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "text", + "content": " prediction in Tab. 6. For these ablations, models are trained on a subset of Habitat and tested on focal, depth, and relative pose prediction, with results averaged over all combinations of auxiliary modalities. DUSt3R serves as a baseline, and DUSt3R-ft is finetuned with an additional " + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "text", + "content": " head. This output can be used at test time to directly predict focals or relative poses, within a single forward pass, instead of having to forward both the original pair and its swapped version. We also report the performance, with and without predicting " + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "text", + "content": " and/or using it. In all cases, predicting " + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 55, + 545, + 295, + 701 + ], + "type": "text", + "content": " leads to slightly better performance while evaluations can be performed in a single forward pass." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "content": "We also explore different ways of conditioning auxil" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 227, + 553, + 299 + ], + "blocks": [ + { + "bbox": [ + 316, + 227, + 553, + 299 + ], + "lines": [ + { + "bbox": [ + 316, + 227, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 553, + 299 + ], + "type": "table", + "html": "
Methodwith X2,2Single ForwardFocal acc@1.015Depth τ@1.03RRA@2 (Pro)RTA@2 (PnP)Avg.
Pro)PnP)
DUSt3R××36.885.469.272.449.857.3
DUSt3R-ft✓×38.979.171.774.452.959.2
✓✓38.378.573.174.456.359.2
Pow3R××66.085.781.983.868.673.3
✓×67.286.182.984.770.074.6
✓✓64.885.084.284.773.074.6
", + "image_path": "6f021b2cbb55be9ca7699c5aa3c75b7ccad0cc8cf78c13dd6b1efd3dd45e2042.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 300, + 553, + 377 + ], + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 377 + ], + "type": "text", + "content": "Table 6. Impact of regressing " + }, + { + "bbox": [ + 313, + 300, + 553, + 377 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 313, + 300, + 553, + 377 + ], + "type": "text", + "content": ". Metrics (in %) on Habitat are averaged over all combinations of input modalities. The 'with " + }, + { + "bbox": [ + 313, + 300, + 553, + 377 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 313, + 300, + 553, + 377 + ], + "type": "text", + "content": "' column denotes methods trained to predict it. 'Single forward' denotes evaluations of models based on " + }, + { + "bbox": [ + 313, + 300, + 553, + 377 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 313, + 300, + 553, + 377 + ], + "type": "text", + "content": " prediction, i.e., in a single forward pass, instead of forwarding image pairs and their swapped versions. The top part shows the original DUSt3R, or finetune (-ft) with " + }, + { + "bbox": [ + 313, + 300, + 553, + 377 + ], + "type": "inline_equation", + "content": "X^{2,2}" + }, + { + "bbox": [ + 313, + 300, + 553, + 377 + ], + "type": "text", + "content": " prediction." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 388, + 553, + 440 + ], + "blocks": [ + { + "bbox": [ + 317, + 388, + 553, + 440 + ], + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 440 + ], + "type": "table", + "html": "
Arch.Extra ParamsFocal acc@1.015Depth τ@1.03RRA@2 (Pro) (PnP)RTA@2 (Pro) (PnP)Avg.
Embed16.0M64.784.784.184.572.873.877.4
Inject122.6M64.885.084.284.773.074.677.7
Inject442.3M65.985.184.885.273.775.478.3
", + "image_path": "5ee9dcad4cb46b7e9155ffa7dd41da9a26a51d0be14e25791146f2225d811f8b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 441, + 553, + 507 + ], + "lines": [ + { + "bbox": [ + 313, + 441, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 553, + 507 + ], + "type": "text", + "content": "Table 7. Architecture ablation. Results (in %), on Habitat, are averaged over all combinations of input modalities. We compare the injection architecture for different number of blocks with the 'embed' architecture in which the extra signals are added before the transformer blocks. 'Inject1' offers a good compromise between performance and number of extra parameters." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "content": "ary information in Tab. 7. To incorporate additional priors, Pow3R injects those embeddings into the first block of the encoder or decoder, see Fig. 3. We compare this approach to another design choice, where the model embeds them before going through attention blocks. We have also tried to perform injection in more than one block, without significant improvement while the computational cost increases." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 620, + 388, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 620, + 388, + 632 + ], + "spans": [ + { + "bbox": [ + 314, + 620, + 388, + 632 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "We have presented Pow3R, a novel approach to guide large 3D vision models, e.g. DUSt3R, with camera and scene priors in an implicit manner. Not only does it results in consistent gains whenever auxiliary information is available, but it also unlocks new capabilities such as high-resolution processing and sparse-to-dense depth completion." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1078" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. IJCV, 2016. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 179 + ], + "type": "text", + "content": "[2] Youset I Abdel-Aziz, Hauck Michael Karara, and Michael Hauck. Direct linear transformation from comparator coordinates into object space coordinates in close range photogrammetry. Photogrammetric engineering & remote sensing, 2015. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 296, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 296, + 214 + ], + "type": "text", + "content": "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. Hpatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 216, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 295, + 258 + ], + "type": "text", + "content": "[4] Daniel Barath, Dmytro Mishkin, Luca Cavalli, Paul-Edouard Sarlin, Petr Hruby, and Marc Pollefeys. Affineglue: Joint matching and robust estimation. arXiv preprint arXiv:2307.15381, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "text", + "content": "[5] Axel Barroso-Laguna, Edgar Riba, Daniel Ponsa, and Krystian Mikolajczyk. Key. net: Keypoint detection by handcrafted and learned cnn filters. In ICCV, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 295, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 295, + 350 + ], + "type": "text", + "content": "[6] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. In NeurIPS, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 351, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 294, + 372 + ], + "type": "text", + "content": "[7] Herbert Bay. Surf: Speeded up robust features. ECCV, 2006. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 374, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 295, + 396 + ], + "type": "text", + "content": "[8] Romain Brégier. Deep regression on manifolds: a 3D rotation case study. In 3DV, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 397, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 430 + ], + "type": "text", + "content": "[9] Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer++: Revealing the devil in transformer's details for multi-view stereo. In ICLR, 2024. 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 432, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 432, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 57, + 432, + 295, + 475 + ], + "type": "text", + "content": "[10] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Ming-min Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 476, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 476, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 57, + 476, + 295, + 509 + ], + "type": "text", + "content": "[11] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In ECCV, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 511, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 295, + 554 + ], + "type": "text", + "content": "[12] David J Crandall, Andrew Owens, Noah Snavely, and Daniel P Huttenlocher. Sfm with mrfs: Discrete-continuous optimization for large-scale structure from motion. IEEE trans. PAMI, 2012. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 556, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 295, + 578 + ], + "type": "text", + "content": "[13] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In CVPR, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "text", + "content": "[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "text", + "content": "[15] Timothee Darcet, Mathilde Caron, Piotr Bojanowski, Ishan Misra, Armand Joulin, Julien Mairal, Hervé Jégou, and Hugo Touvron. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "text", + "content": "worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 97, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 554, + 140 + ], + "type": "text", + "content": "[17] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 142, + 554, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 142, + 554, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 142, + 554, + 175 + ], + "type": "text", + "content": "[18] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and Trends® in Computer Graphics and Vision, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 177, + 554, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 177, + 554, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 177, + 554, + 208 + ], + "type": "text", + "content": "[19] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In ICCV, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 211, + 554, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 211, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 317, + 211, + 554, + 243 + ], + "type": "text", + "content": "[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. *IJRR*, 2013. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 246, + 554, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 246, + 554, + 267 + ], + "spans": [ + { + "bbox": [ + 317, + 246, + 554, + 267 + ], + "type": "text", + "content": "[21] Chris Harris, Mike Stephens, et al. A combined corner and edge detector. In Alvey vision conference, 1988. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 269, + 554, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 269, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 317, + 269, + 554, + 300 + ], + "type": "text", + "content": "[22] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 303, + 554, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 303, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 317, + 303, + 554, + 335 + ], + "type": "text", + "content": "[23] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In ICRA, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 337, + 554, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 337, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 337, + 554, + 369 + ], + "type": "text", + "content": "[24] Wonbong Jang and Lourdes Agapito. Codenerf: Disentangled neural radiance fields for object categories. In ICCV, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 372, + 554, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 393 + ], + "type": "text", + "content": "[25] Nianjuan Jiang, Zhaopeng Cui, and Ping Tan. A global linear method for camera pose registration. In ICCV, 2013. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 396, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 396, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 396, + 554, + 437 + ], + "type": "text", + "content": "[26] Shengyu Jiang, Chen Wang, Yuhong Zhang, and Zhibin Wang. Completionformer: Efficient and accurate depth completion via hybrid attention with transformers. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 440, + 554, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 440, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 317, + 440, + 554, + 473 + ], + "type": "text", + "content": "[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM trans. Graphics, 2023. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 475, + 554, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 554, + 507 + ], + "type": "text", + "content": "[28] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM trans. Graphics, 2017. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 509, + 554, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 509, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 317, + 509, + 554, + 540 + ], + "type": "text", + "content": "[29] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. IJCV, 2009. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 544, + 554, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 544, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 554, + 576 + ], + "type": "text", + "content": "[30] Vincent Leroy, Jean-Sébastien Franco, and Edmond Boyer. Volume sweeping: Learning photoconsistency for multi-view shape reconstruction. IJCV, 2021. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 578, + 554, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 554, + 609 + ], + "type": "text", + "content": "[31] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with MAST3R. In ECCV, 2024. 2, 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 612, + 554, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 612, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 612, + 554, + 643 + ], + "type": "text", + "content": "[32] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 5" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 317, + 647, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 554, + 678 + ], + "type": "text", + "content": "[33] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 3, 7" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 317, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 554, + 713 + ], + "type": "text", + "content": "[34] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 7" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1079" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[35] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In ICCV, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "text", + "content": "[36] David G Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 2004. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 130, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 130, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 57, + 130, + 294, + 152 + ], + "type": "text", + "content": "[37] Bin Luo and Edwin R. Hancock. Procrustes alignment with the em algorithm. In CAIP, 1999. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 153, + 295, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 153, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 153, + 295, + 186 + ], + "type": "text", + "content": "[38] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In ICRA, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "text", + "content": "[39] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In ICRA, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 232, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 232, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 294, + 253 + ], + "type": "text", + "content": "[40] Zeyu Ma, Zachary Teed, and Jia Deng. Multiview stereo with cascaded epipolar raft. In ECCV, 2022. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 255, + 295, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 295, + 298 + ], + "type": "text", + "content": "[41] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "text", + "content": "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 344, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 387 + ], + "type": "text", + "content": "[43] Danish Nazir, Marcus Liwicki, Didier Stricker, and Muhammad Zeshan Afzal. Semattnet: Towards attention-based semantic aware guided depth completion. arXiv preprint arXiv:2204.13635, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 388, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 432 + ], + "type": "text", + "content": "[44] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In CVPR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 434, + 180, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 434, + 180, + 445 + ], + "spans": [ + { + "bbox": [ + 57, + 434, + 180, + 445 + ], + "type": "text", + "content": "[45] OpenAI. Chatgpt-4, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 446, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 446, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 57, + 446, + 295, + 488 + ], + "type": "text", + "content": "[46] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 490, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 295, + 521 + ], + "type": "text", + "content": "[47] Frank Plastria. The weiszfeld algorithm: Proof, amendments, and extensions. Foundations of Location Analysis, 2011. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 524, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 295, + 578 + ], + "type": "text", + "content": "[48] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deep lidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In CVPR, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 580, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 644 + ], + "type": "text", + "content": "[49] Alexander Raistrick, Lahav Lipson, Zeyu Ma, Lingjie Mei, Mingzhe Wang, Yiming Zuo, Karhan Kayan, Hongyu Wen, Beining Han, Yihan Wang, Alejandro Newell, Hei Law, Ankit Goyal, Kaiyu Yang, and Jia Deng. Infinite photorealistic worlds using procedural generation. In CVPR, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 669 + ], + "type": "text", + "content": "[50] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[51] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV, 2021. 5, 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[52] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "text", + "content": "[53] Chris Rockwell, Justin Johnson, and David F Fouhey. The 8-point algorithm as an inductive bias for relative pose prediction by vits. In 3DV, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 142, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 142, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 142, + 553, + 175 + ], + "type": "text", + "content": "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 177, + 553, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 177, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 317, + 177, + 553, + 198 + ], + "type": "text", + "content": "[55] Edward Rosten and Tom Drummond. Machine learning for high-speed corner detection. In ECCV, 2006. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "type": "text", + "content": "[56] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In CVPR, 2019. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 235, + 553, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 235, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 316, + 235, + 553, + 268 + ], + "type": "text", + "content": "[57] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In CVPR, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 270, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 553, + 312 + ], + "type": "text", + "content": "[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 315, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 553, + 336 + ], + "type": "text", + "content": "[59] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 338, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 371 + ], + "type": "text", + "content": "[60] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "type": "text", + "content": "[61] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In CVPR, 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 430, + 553, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 462 + ], + "type": "text", + "content": "[62] Philipp Schröppel, Jan Bechtold, Artemij Amiranashvili, and Thomas Brox. A benchmark and a baseline for robust multiview depth estimation. In 3DV, 2022. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 464, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 497 + ], + "type": "text", + "content": "[63] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "text", + "content": "[64] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 544, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 586 + ], + "type": "text", + "content": "[65] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "text", + "content": "[66] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "text", + "content": "[67] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[68] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE trans. Image Processing, 2020. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1080" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "text", + "content": "[69] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 129 + ], + "type": "text", + "content": "[70] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 131, + 294, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 131, + 294, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 131, + 294, + 164 + ], + "type": "text", + "content": "[71] Wouter Van Gansbeke, Davy Neven, Bert De Brabandere, and Luc Van Gool. Sparse and noisy lidar completion with rgb guidance and uncertainty. In MVA, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 166, + 294, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 166, + 294, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 166, + 294, + 188 + ], + "type": "text", + "content": "[72] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. In 3DV, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 189, + 294, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 294, + 222 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 294, + 222 + ], + "type": "text", + "content": "[73] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In ICCV, 2023. 1, 3, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 224, + 294, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 224, + 294, + 256 + ], + "spans": [ + { + "bbox": [ + 56, + 224, + 294, + 256 + ], + "type": "text", + "content": "[74] Shuzhe Wang, Juho Kannala, Marc Pollefeys, and Daniel Barath. Guiding local feature matching with surface curvature. In ICCV, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 258, + 294, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 294, + 291 + ], + "type": "text", + "content": "[75] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "text", + "content": "[76] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In ICCV, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "text", + "content": "[77] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Bregier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In ICCV, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 395, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 395, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 56, + 395, + 294, + 427 + ], + "type": "text", + "content": "[78] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 430, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 294, + 472 + ], + "type": "text", + "content": "[79] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet++: Semantic assisted repetitive image guided network for depth completion. In ECCV, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "text", + "content": "[80] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 520, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 552 + ], + "type": "text", + "content": "[81] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In ECCV, 2018. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "type": "text", + "content": "[82] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In CVPR, 2020. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 643 + ], + "type": "text", + "content": "[83] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "type": "text", + "content": "[84] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[85] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In ECCV, 2016. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 286 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "text", + "content": "[86] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. In BMVC, 2020. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[87] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. In ICLR, 2025. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "type": "text", + "content": "[88] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Rel-pose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, 2022. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "text", + "content": "[89] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In ICLR, 2024. 1, 3, 4, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "type": "text", + "content": "[90] Yinda Zhang and Thomas Funkhouser. Deep depth completion of a single rgb-d image. In CVPR, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 254, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 286 + ], + "type": "text", + "content": "[91] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. SIGGRAPH, 2018. 8" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1081" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/0f1a120c-8d95-4d35-a553-6cd4183aad0f_content_list.json b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/0f1a120c-8d95-4d35-a553-6cd4183aad0f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9e357dc0038d83f7197cc3dcf616fdcbf3eb656c --- /dev/null +++ b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/0f1a120c-8d95-4d35-a553-6cd4183aad0f_content_list.json @@ -0,0 +1,1812 @@ +[ + { + "type": "text", + "text": "PrEditor3D: Fast and Precise 3D Shape Editing", + "text_level": 1, + "bbox": [ + 256, + 130, + 741, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziya Erkoc1", + "bbox": [ + 122, + 179, + 220, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can Gümeli", + "bbox": [ + 259, + 179, + 366, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peter Wonka2,3", + "bbox": [ + 254, + 204, + 374, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chaoyang Wang²", + "bbox": [ + 406, + 179, + 547, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hsin-Ying Lee", + "bbox": [ + 434, + 204, + 562, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthias Nießner1", + "bbox": [ + 584, + 180, + 730, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peiye Zhuang", + "bbox": [ + 622, + 205, + 741, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Technical University of Munich $^{2}$ Snap Inc $^{3}$ KAUST", + "bbox": [ + 281, + 229, + 715, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://ziyaerkoc.com/preditor3d", + "bbox": [ + 305, + 267, + 686, + 284 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c48c9fdd0cf5631c221117521fc1b9b1344c40a5bc7f89944adb8367fd7393f.jpg", + "image_caption": [ + "\"chicken in a racing car\"" + ], + "image_footnote": [], + "bbox": [ + 124, + 335, + 225, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/23b1c8cd2d7cb0bc1f34e5aa25c4deccd09603d79b4764c929d9fafc8d399d92.jpg", + "image_caption": [ + "cannon" + ], + "image_footnote": [], + "bbox": [ + 112, + 425, + 236, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c81bd58f932972b1c74cc61ced5ca46c670d94d066d3653ab6b0cc544bd913b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 396, + 267, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8cc7d880f57287871f29f335a086755527b4ce1063a9a502313a2e31a52423a8.jpg", + "image_caption": [ + "\"chicken cat ...\"" + ], + "image_footnote": [], + "bbox": [ + 282, + 337, + 398, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62692890255cfc1898b373239421ee2c28b5017bfc6ec5b0913767a15e98fa34.jpg", + "image_caption": [ + "cannon" + ], + "image_footnote": [], + "bbox": [ + 272, + 425, + 401, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ed015752295ae7fb404eb6679c783c6a3b6d80264dcc3c4a299031e7a8dab01.jpg", + "image_caption": [ + "chicken dog ..." + ], + "image_footnote": [], + "bbox": [ + 421, + 337, + 555, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a74bf6ce14f352c08e3a4df63f544caf575a117ecb090213c45a7592a14d9fe4.jpg", + "image_caption": [ + "cannon" + ], + "image_footnote": [], + "bbox": [ + 421, + 425, + 555, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/16d93914e8b02aeb8cce88ee3ff194f52947c3c1761d2a8996a9341f63b48ba9.jpg", + "image_caption": [ + "\"... raeing-car jeep\"" + ], + "image_footnote": [], + "bbox": [ + 576, + 337, + 666, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f468347a2e0645996089876c58fcf8febc3b6aea3d6910bf2622c55c710342a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 422, + 687, + 476 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Angela Dai", + "bbox": [ + 772, + 180, + 870, + 198 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/533aa91a06d59cef9f34e86ec606b1fe42a9400ade9b166c34905d771ce16dcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 310, + 893, + 488 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c93d13830cebbdc3fd20e421c2d0047868e94200efb6bad0a586edd0658c52ca.jpg", + "image_caption": [ + "in a wooden cart\"" + ], + "image_footnote": [], + "bbox": [ + 109, + 529, + 200, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd4fa33ecb83d425c3c2b706ba40c06eaf88ee091aa0d2522c8e93e8cf235f1c.jpg", + "image_caption": [ + "viking with a" + ], + "image_footnote": [], + "bbox": [ + 207, + 539, + 230, + 569 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e70736e11c74e377b631b1205aeeb3a93298b6a794b715578fd9fabe176811fc.jpg", + "image_caption": [ + "flamethrower ...\"" + ], + "image_footnote": [], + "bbox": [ + 235, + 529, + 326, + 587 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/119220910980ff47d723ef27312b3b92c8980c71bc90980fc36c8a42d7640f0a.jpg", + "image_caption": [ + "trebuchet ..." + ], + "image_footnote": [], + "bbox": [ + 334, + 530, + 426, + 587 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff9645074f11352b99d8ea555fd6396ae70b8abe9949202ea70ee6f0da360177.jpg", + "image_caption": [ + "\"house in a forest\"", + "\"chicken in a" + ], + "image_footnote": [], + "bbox": [ + 441, + 522, + 534, + 587 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/875172e5833e4a5e78739a0eab9283c8d2f51a49479b1727569ecc004d9ead00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 537, + 566, + 566 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ae10615da9277a41397c1aeaf094502e7942c78d403c55322ad5e8bd99226f0.jpg", + "image_caption": [ + "house treasure cave ...", + "\"chicken" + ], + "image_footnote": [], + "bbox": [ + 571, + 521, + 655, + 587 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8096860005b6cb0122e73a16c5a9a1c5a04fafb01844457f358e956358d1390a.jpg", + "image_caption": [ + "\"truck carrying clay\"", + "\"android\"" + ], + "image_footnote": [], + "bbox": [ + 684, + 525, + 764, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 776, + 537, + 800, + 569 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf2f124969d224eeb3a8ed43c086bec87e9a83d8b021634703c3deb304d76a1e.jpg", + "image_caption": [ + "... elay pizza", + "\"android chicken" + ], + "image_footnote": [], + "bbox": [ + 776, + 527, + 882, + 594 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0421c88025359e3c64200ad95c50bf5c163fd5d4a231a5f4eee04bc92a7f5738.jpg", + "image_caption": [ + "mustache, helmet", + "... helmet party hat" + ], + "image_footnote": [], + "bbox": [ + 246, + 627, + 315, + 698 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c32bcbd3cff098474abb37d3b96be16ae7a075be91fdbe18cc8942e510e83875.jpg", + "image_caption": [ + "...red pepper mustache ..." + ], + "image_footnote": [], + "bbox": [ + 364, + 628, + 437, + 699 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/406f732f82a18a122c29eb4202639c5106e88fb59cdbe8529e2921515f6efec5.jpg", + "image_caption": [ + "racing car\"" + ], + "image_footnote": [], + "bbox": [ + 462, + 628, + 532, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 462, + 641, + 532, + 700 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/073e30d0b0f8862b586d53260b056453641089686f0676451908ab775ad0519b.jpg", + "image_caption": [ + "cat with a tail ...", + "Figure 1. PrEditor3D is a (top) fast and high-quality editing method that can perform precise and consistent editing only in the intended regions, keeping the rest identical. (mid) It can handle diverse editing prompts with any given 3D object. (bottom) Furthermore, it can support iterative editing, facilitating artistic workflow, and can also support editing multiple regions in a single run." + ], + "image_footnote": [], + "bbox": [ + 578, + 632, + 653, + 703 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69e72e7cebb6abbe44a9aaa6cb9ed6c067af4dbca1aed5e99fae4134d47cba3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 633, + 771, + 696 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eeb47fe2344c3616e6b6291f1645cee84b9d9f33d3d38968afc03dc693e117d3.jpg", + "image_caption": [ + "wearing a tie\"" + ], + "image_footnote": [], + "bbox": [ + 803, + 632, + 882, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 760, + 326, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a training-free approach to 3D editing that enables the editing of a single shape within a few minutes. The edited 3D mesh aligns well with the prompts, and remains identical for regions that are not intended to be altered. To this end, we first project the 3D object onto 4-view images and perform synchronized multi-view image editing along with user-guided text prompts and user-provided", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rough masks. However, the targeted regions to be edited are ambiguous due to projection from 3D to 2D. To ensure precise editing only in intended regions, we develop a 3D segmentation pipeline that detects edited areas in 3D space, followed by a merging algorithm to seamlessly integrate edited 3D regions with the original input. Extensive experiments demonstrate the superiority of our method over previous approaches, enabling fast, high-quality editing while preserving unintended regions.", + "bbox": [ + 511, + 761, + 906, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "640", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent 3D diffusion models can generate high-quality assets that closely align with the text prompts in the form of neural fields [23, 38], meshes [2, 39], or Gaussian point clouds [48]. Although these methods generate impressive results, they lack the essential capability for precise and controllable editing of the generated outputs, a critical requirement for iterative artistic workflows. Effective 3D editing demands: (1) it should be fast enough to provide quick feedback, ideally comparable to fast 3D generation algorithms, and (2) it must allow for precise local control, enabling users to keep specific parts of the model unchanged.", + "bbox": [ + 89, + 114, + 483, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Enabling precise and controllable editing is still an open challenge. Several initial approaches have been proposed to tackle the challenge of 3D editing [8, 9, 14, 17, 32, 37], providing promising results but suffering from slow runtime, lack of precise control, and/or lack of 3D consistency and quality. Optimization-based techniques like SDS [9, 14, 37] or multiview training dataset updates [17] are computationally expensive, making interactive editing out of reach.", + "bbox": [ + 89, + 281, + 483, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, they offer limited control over specific parts of the shape, as text prompts alone cannot precisely localize regions to be edited [8, 9, 17, 32, 37]. While Vox-E [37] and Shap-Editor [9] propose a mechanism to prevent original parts of the shape from being altered during editing, they do not enable precise editing due to having only text as input. Finally, one can observe various visual quality problems, such as the Janus problem, blurring, over-saturation, and overemphasizing texture changes while leaving the geometry intact or degrading.", + "bbox": [ + 89, + 402, + 483, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a novel editing pipeline for 3D assets that is faster, more precise, and delivers high-quality results (See Fig. 1). As our primary goal is faster editing, we propose an editing framework leveraging a pipeline that consists of two components: a multiview diffusion algorithm and a feed-forward mesh reconstruction. Multi-view diffusion models can leverage superior 2D editing techniques, and the feed-forward mesh reconstruction bridges the gap between 2D and 3D. For better controllability, we extend multi-view image generation to multi-view image editing using 2D masks to constrain the edits to user-specified regions. The 2D masks can take various forms, including manually selected regions, handbrushed areas, or automatically generated segmentations. Automatically generated masks can constrain the edit to a limited region and make the editing more restrictive, we further elaborate on this discussion in the supplementary material. We adopt DDPM inversion [21] to extract initial noise vectors from input multi-view images and execute Prompt-to-Prompt [18] on a multi-view diffusion model [38]. We use 2D user-provided masks to blend edited and original views during the denoising. Due to the occlusion issues that may naturally occur in 2D renderings, we cannot ob", + "bbox": [ + 89, + 553, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ed053e6f3f5efb47dd28d94b8cee1b43fd32c5310ab876809527fce6cfbba889.jpg", + "image_caption": [ + "Figure 2. Ambiguous intended regions. The intended region to be edited is clear in 3D (e.g. the cat tail). However, after projecting to 2D, regardless of the granularity of the user-provided masks, the editing will either alter some unintended regions (e.g. the robot cat) or be too limited for reasonable editing, especially because of occlusion caused by the tail." + ], + "image_footnote": [], + "bbox": [ + 517, + 93, + 870, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tain ideal intended regions in 2D regardless of the granularity of the masks, as shown in Fig. 2. The masks are either too coarse so the unintended regions will be changed, or too fine-grained to allow reasonable editing. Without additional spatial information in 3D, multi-view editing approaches cannot fully address this challenge. Therefore, simply adopting a feed-forward reconstruction method [51] to convert edited multi-views into a 3D mesh often leads to undesirable results.", + "bbox": [ + 511, + 292, + 906, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this issue, we propose using the original 3D input and 3D segmentation. We first detect the intended editing region using Grounding DINO [24] and SAM 2 [35] with the user mask and prompts. This gives an initial 2D segmentation that we subsequently lift to 3D. For this purpose, we use color coding in a multi-view to 3D reconstruction pipeline, named GTR [51], to end up with a 3D segmentation that we can use during merging. Specifically, we paint 2D segmentations in a specific color, (e.g., green) and after reconstruction, we can detect which regions are edited by querying the color in the 3D field.", + "bbox": [ + 511, + 430, + 908, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Then, we perform merging to maintain the original parts of the shape. We use 3D masks to detect edited/replaced parts in GTR's [51] voxel-feature space. We extract the edited part from the new shape and replace it with the old part from the original shape. That way, we guarantee that the remaining shape will remain identical. We apply a final average blending operation so that new parts and the original shape blend smoothly.", + "bbox": [ + 511, + 597, + 908, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we make the following contributions:", + "bbox": [ + 531, + 719, + 867, + 733 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel method for diffusion-based 3D object editing that is faster than previous work and enables precise, interactive editing.", + "- The proposed method consists of user-guided and multiview-synchronized editing and a feed-forward 3D reconstruction, enabling fast editing in a feed-forward manner.", + "- To enable precise editing only to intended regions, we propose a voxel-based 3D segmentation method that utilizes multi-view segmentation information and propagates it to 3D, followed by an average blending operation to merge the edited and original objects." + ], + "bbox": [ + 511, + 734, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "641", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d674394a442d9900b55227a5829b8f095cb9e0a3c2630d948156ec9be7838c96.jpg", + "image_caption": [ + "Figure 3. Overview of PrEditor3D. Given an input 3D object, we first render its multi-view images from 4 orthogonal views. We then obtain editing input from the user, describing in text as well as rough 2D masks the desired edits. We perform synchronized multi-view editing based on the text prompts as well as the user-provided masks (Sec.3.1. Due to the rough masks and the unclear intended regions caused by ambiguous 3D-2D projection, we detect the intended regions with Grounding Dino and SAM 2 (Sec.3.2, where the segmentation results are lifted to 3D for the final merging operation (Sec.3.3)." + ], + "image_footnote": [], + "bbox": [ + 102, + 92, + 903, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Our editing method has superior quality to previous work. We show significant improvements in GPTEval3D [45], directional CLIP metrics, and extensive user studies.", + "bbox": [ + 89, + 378, + 480, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 440, + 240, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D Editing applies global or local modifications to an image based on user instructions. It has gained significant attention for enabling a more interactive user experience in content creation. To achieve this, existing methods either fine-tune text-to-image models with specialized instructional editing datasets [46, 50], or use training-free approaches with inpainting [3, 30] or cross-attention mechanisms [6, 7, 18]. To edit user-provided images, various diffusion inversion techniques have been developed. For example, SDEdit [26] introduces noise to images to capture intermediate steps in the diffusion process, null-text inversion [28] inverts the deterministic DDIM inversion [40], and recent works [4, 21] use DDPM inversion [19] to enhance editing capabilities.", + "bbox": [ + 88, + 470, + 482, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, one step to achieve 3D editing is to perform 2D multi-view image editing. We apply the DDPM inversion [19] and Prompt2Prompt [18] to operate within a diffusion-based sparse multi-view generation model.", + "bbox": [ + 89, + 683, + 483, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Reconstruction from Sparse Multi-Views refers to the task of reconstructing a 3D instance from a limited number of multi-view images. To achieve this, Score Distillation Sampling (SDS) [31] and its variants [10, 23, 33, 42, 43] optimize a 3D scene representation by reconstructing the given sparse views and generating novel views through gradients from large-scale pre-trained text-to-image diffusion models [36]. However, these approaches are often computationally demanding. Beyond per-scene optimization, PixelNeRF [49] generates a NeRF representation [27] of", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a scene from sparse-view images by training the model across multiple scenes. With recent advancements in largescale 3D datasets such as Objaverse [13], follow-up works [20, 22, 25, 29, 41, 44, 47, 51] have improved this feedforward 3D reconstruction approach for object-centric 3D assets. In this work, we utilize one recent method, GTR [51], to quickly generate 3D meshes from 4 views.", + "bbox": [ + 511, + 378, + 903, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Editing presents additional challenges compared to 2D editing due to the need to maintain spatial consistency in 3D. To address this, some methods train 3D editing models using paired 3D datasets [1, 46], however, these approaches are limited by a lack of diverse and complex datasets. Other methods adapt 2D editing models, such as InstructPix2Pix [5], for the 3D domain [17, 32], by iteratively updating multi-view images of a scene. However, without synchronized multi-view updates, this approach often results in flickering or inconsistent views. Alternatively, some methods propose a generation-reconstruction loop to modify 3D representations using intermediate denoised images [8] or SDS gradients [9, 11, 12, 14, 37] in the diffusion process. While these methods can achieve 3D consistency, they often struggle with quality or suffer from high computational costs. In our work, we perform multi-view editing and reconstruct the edited object in 3D. Beyond that, to preserve unchanged regions, we carefully design an approach by detecting the edited 3D regions and integrating the intended 3D edited regions into the original shapes.", + "bbox": [ + 511, + 486, + 906, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 799, + 604, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to achieve fast 3D asset editing in a training-free manner, allowing for precise and user-guided edits. Our approach achieves this through multi-view image editing in 2D, followed by lifting the 2D edits into 3D. This process can be summarized into 3 main steps: (1) synchronized", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "642", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sparse multi-view editing in 2D (Sec.3.1), (2) detecting intended editing regions across 2D views through the Grounding Dino [24] and SAM 2 [35] approach (Sec.3.2), and (3) lifting the intended editing regions to 3D and merging the edited shape into the original (Sec.3.3). Our approach is illustrated in Fig. 3.", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Synchronized Sparse Multi-View Editing", + "text_level": 1, + "bbox": [ + 89, + 190, + 441, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To edit a given 3D object $\\mathcal{O}$ , we leverage the power of 2D editing through multi-view diffusion models. We first perform synchronized sparse multi-view edits using a pretrained multi-view diffusion model. In practice, we use MVDream [38], which generates 4 orthogonal views based on a text prompt. Note that our approach remains agnostic to the specific diffusion-based multi-view generation model used.", + "bbox": [ + 89, + 212, + 483, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first render multi-view images from $\\mathcal{O}$ , then apply the DDPM diffusion inversion mechanism [21] to revert the multi-view images to their initial noise vectors, denoted as $\\pmb{x}^T$ , where $T$ is the number of diffusion timesteps during the denoising process. We will use these vectors $\\pmb{x}^T$ as the initial latent vectors for the editing process in diffusion. We denote the text prompt for the input shape and the edited shape as $y_i$ and $y_e$ , respectively. Also, to enable better editing control and interaction, we further take user-provided masks in 4 views as input. These masks indicate target regions for the edit, denoted as $M_{\\mathrm{user}} \\in \\mathbb{R}^{4 \\times H \\times W}$ where $H$ and $W$ are the height and weight of the images. Note that our method does not require precise and accurate masks.", + "bbox": [ + 89, + 319, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To edit multi-view images, we apply the Prompt-to-Prompt [18] approach on the multi-view diffusion model. For simplicity, we present the basic operation at each diffusion timestep and within each attention block. To be specific, Prompt-to-Prompt [18] generates an edited latent vector $\\boldsymbol{x}_e'$ by replacing the self- and cross-attention weights of the original latent vector, $\\boldsymbol{x}_i$ , with the edited prompt $y_e$ . To confine modifications to the desired regions, we blend the latent vectors $\\boldsymbol{x}_e'$ and $\\boldsymbol{x}_i$ using user-provided masks $M_{\\mathrm{user}}$ . We denote the final edited latent vector at each step as $\\boldsymbol{x}_e$ :", + "bbox": [ + 89, + 515, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {e} \\leftarrow M _ {\\text {u s e r}} \\cdot \\boldsymbol {x} _ {e} ^ {\\prime} + \\left(1 - M _ {\\text {u s e r}}\\right) \\cdot \\boldsymbol {x} _ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 676, + 482, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, we downsample the user-provided masks $M_{\\mathrm{user}}$ to match the feature resolution at each model layer, and the edited latent vector, $x_{e}$ , serves as the input to the next layer.", + "bbox": [ + 89, + 704, + 482, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the inverted noise vector $\\pmb{x}^T$ ensures that the edited results align with the original texture style. However, the masks $M_{\\mathrm{user}}$ are often imprecise or misaligned with the target regions, which can still affect regions that are not intended to be altered. Furthermore, occlusions along the depth dimension introduce additional challenges in accurately and semantically localizing the intended editing regions in 3D. We illustrate the issue in Fig. 2. To address this, we apply an automated grounding approach that detects the intended editing region in both 2D and 3D.", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Merging Voxel Features" + ], + "code_body": "Input: $V_{i}$ and $V_{e} \\in \\mathbb{R}^{A \\times A \\times F}, M_{i}$ and $M_{e} \\in \\mathbb{R}^{A \\times A \\times A}, d \\in \\mathbb{N}$ , and $\\theta \\in [0,1]$ \nOutput: $V_{\\text{blend}} \\in \\mathbb{R}^{A \\times A \\times A \\times F}$ \n1 $V_{i}[M_{i}] \\gets \\emptyset$ \n2 $V_{i}[M_{e}] \\gets V_{e}[M_{e}]$ \n3 $N \\gets \\text{Dilation}(M_{e}, d)$ \n4 $K \\gets M_{e} \\oplus N // \\oplus$ means XOR \n5 $V_{\\text{blend}} \\gets V_{i}$ \n6 $V_{\\text{blend}}[K] \\gets \\theta \\odot V_{i}[K] + (1 - \\theta) \\odot V_{e}[K]$ \n7 return $V_{\\text{blend}}$", + "bbox": [ + 516, + 111, + 903, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Detection of Intended Editing Regions in 2D", + "text_level": 1, + "bbox": [ + 511, + 297, + 887, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The intended editing regions refer to the specific semantic areas that correspond to the editing prompt. For example, in Fig. 1, changing \"chicken\" to \"cat\" implies that the desired editing region pertains only to the areas representing the \"chicken\" and \"cat\" concepts. Regions (both 2D and 3D) that are not semantically related to these concepts should remain unchanged.", + "bbox": [ + 511, + 320, + 906, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure that only intended editing regions are edited while allowing rough user-provided masks, we propose to detect the intended editing regions by applying Grounding Dino [24] and SAM 2 [35] to both the original and the edited multi-view images. To begin with, we identify the changing concept by comparing the original prompt $y_{i}$ and the editing prompt $y_{e}$ . We then localize the changing concept within the user-provided mask regions $M_{\\mathrm{user}}$ and obtain corresponding bounding boxes in multiple views. Formally, we write the procedure as", + "bbox": [ + 511, + 426, + 908, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {b b o x} _ {x} \\leftarrow \\operatorname {G r o u n d i n g} \\left(x, M _ {\\text {u s e r}}, y _ {i}, y _ {e}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 590, + 903, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathsf{bbox}_x$ are the bounding boxes for the changing concept in the multi-view images $x\\in \\{\\pmb {x}_i,\\pmb {x}_e\\}$", + "bbox": [ + 511, + 618, + 903, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterward, we segment and track the changing elements across views using the grounding bounding boxes $\\mathsf{bbox}_x$ . Formally, this procedure can be described as $\\mathrm{SAM}(\\pmb{x},\\mathsf{bbox}_x)$ . This process yields the intended editing regions in segmentation format for both the original and edited multi-views.", + "bbox": [ + 511, + 648, + 906, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Lifting and Merging Edits in 3D", + "text_level": 1, + "bbox": [ + 511, + 748, + 797, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we lift the 2D edits to 3D and merge the 3D-edited regions into the original shape.", + "bbox": [ + 511, + 770, + 903, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Segmentation by Lifting 2D Segmentations. We mark the intended editing regions on multi-view images using a green color. The color-coded multi-view images are then reconstructed in 3D using an offline 3D reconstruction", + "bbox": [ + 511, + 804, + 906, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We use $\\pmb{x}$ to represent both latent vectors and images, without distinguishing between the two.", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "643", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fc2fa62a782513e34a7ad643267821a407c09f1c29730176f900bc37f2e9e090.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison. Our method can perform diverse editing samples and only edit the intended regions." + ], + "image_footnote": [], + "bbox": [ + 91, + 93, + 880, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model [51]. This model takes multiple views as input and represents the shape as a triplane feature. Two separate decoders—one for Signed Distance Function (SDF) and one for color—generate a geometry field and a color field, respectively. Through this process, we create a 3D segmentation field from the color-coded multi-view images. For each 3D position within the 3D space represented by the triplane, we determine whether the position lies within the 3D intended editing regions based on its color value. That is, we apply a decision threshold for the distance between each color value and the preset green color to identify targeted regions. This produces two 3D masks, indicating the intended editing regions in 3D for both the original shape and the edited shape, denoted as $M_{i}$ and $M_{e}$ , respectively. Our selection of green color is arbitrary. We can flexibly change the color, if the shape already has the specific green color we use.", + "bbox": [ + 88, + 614, + 485, + 872 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Merging Edits in 3D. As aforementioned, coarse user", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "provided masks and occlusion issues in 2D projections often result in unwanted alterations, compromising the preservation of unaffected 3D regions. Therefore, the directly reconstructed shapes from the edited multi-view images using the reconstruction model [51] cannot be the final editing output, as illustrated in Fig. 7.", + "bbox": [ + 511, + 614, + 906, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the 3D reconstruction model [51], we extract voxel features for both original and edited shapes by interpolating their triplane features, denoted as $\\mathbf{V}_i$ and $\\mathbf{V}_e$ for the original and edited shapes, respectively. Both $\\mathbf{V}_i, \\mathbf{V}_e \\in \\mathbb{R}^{A \\times A \\times A \\times F}$ , where $A$ is the voxel resolution, and $F$ is the feature dimension, $A = 256$ and $F = 40$ in practice.", + "bbox": [ + 511, + 705, + 908, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To merge 3D features, we first nullify the original specific regions $M_{i}$ from the original voxel feature $V_{i}$ , and then replace the target edited regions $M_{e}$ with edited feature $V_{e}[M_{e}]$ . We write the above operations as follows:", + "bbox": [ + 511, + 797, + 906, + 858 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {i} \\left[ M _ {i} \\right] \\leftarrow \\emptyset , \\text {a n d} V _ {i} \\left[ M _ {e} \\right] \\leftarrow V _ {e} \\left[ M _ {e} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 863, + 903, + 880 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We refer to this approach as a naive copy-paste method.", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "644", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/86854e91fad33408573a2435b059a587351325bb154a437faa1cb460899ed5ce.jpg", + "image_caption": [ + "Figure 5. More editing results from PrEditor3D. Our method can perform a wide range of editing on various shapes." + ], + "image_footnote": [], + "bbox": [ + 125, + 93, + 885, + 436 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6584f56e1b45b43ab56a660ef11d1e744a567b51568bf0006b0e4367201f8b96.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPrompt Align.3D PlausibilityTextureOverall
Tailor3D [32]98%99%99%99%
MVEdit [8]57%55%55%57%
Vox-E [37]53%68%50%55%
", + "bbox": [ + 94, + 469, + 491, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While theoretically plausible, we observe that this straightforward approach typically introduces discontinuities at the 3D editing boundaries, as shown in Fig. 7. To address this, we propose an averaged merging approach that provides a more robust blend of 3D features. In the improved method, we dilate the 3D mask $M_{e}$ by a dilation $d$ and then use an exclusive or (a.k.a. XOR) operation to select the boundary mask regions for smooth blending. Next, we linearly interpolate the two voxel features $V_{i}$ and $V_{e}$ within the boundary regions using a coefficient $\\theta$ , in practice $\\theta = 0.5$ . We illustrate the merging process in Alg. 1. After merging, we generate a textured mesh from the blended voxel feature, using the decoders in the 3D reconstruction model [51].", + "bbox": [ + 89, + 614, + 483, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 823, + 223, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 848, + 210, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our evaluation dataset contains 18 unique shapes and 40 editing prompts. We use shapes from GSO [15] and Obja", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/29966c18ba2f5d551d837f1b0e63e598c808fca77ea1ba84599ea77cc878bc66.jpg", + "table_caption": [ + "Table 1. Comparison using GPTEval3D [45]. Scores indicate the percentage of our method being selected over baselines. We ask GPT-4V to compare our method against the baselines and choose one based on pre-defined categories. Each number here shows the percentage of our method being chosen." + ], + "table_footnote": [], + "table_body": "
MethodPrompt Align.Visual QualityPreserving Shape
Tailor3D [32]96%97%99%
MVEdit [8]68%68%97%
Vox-E [37]78%88%89%
", + "bbox": [ + 517, + 469, + 915, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. User study results comparing our method against baselines. The percentage shows the preference for our method. We show users binary comparisons of our method against the baselines and collect their response on which one they prefer.", + "bbox": [ + 511, + 527, + 906, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "verse [13]. We evaluate our method based on the quality of the editing and consistency with the input shapes.", + "bbox": [ + 511, + 608, + 906, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the GPTEval3D [45] metric to evaluate the quality of edited shapes and their alignment with the text prompts. GPT-4V is provided with multi-view renderings of two methods at a time, and instructed to pick one based on text-prompt alignment, 3D plausibility, and texture details. There were 120 total questions, each answered 3 times.", + "bbox": [ + 511, + 641, + 906, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the GPTEval3D metric does not consider the input shape, it cannot measure whether the shape remained intact, and whether the edited shape is consistent with the high-level style of the input shape. Therefore, we adopt the directional CLIP [34] score metric, $\\mathrm{CLIP}_{\\mathrm{dir}}$ from previous works [16, 37]. $\\mathrm{CLIP}_{\\mathrm{dir}}$ evaluates the average difference between text feature change direction and image feature change direction, where images are multi-view renderings of input and edited shapes. To ensure our evaluation is not affected by a particular implementation of this metric, we introduce three variants. $\\mathrm{CLIP}_{\\mathrm{dir - cos}}$ replaces the", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "645", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/048113bc701dc0c58c381255420ce7ebbd80c5f26fb78a12fdee87605de64b98.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCLIPdir↑CLIPdir-cos↑CLIPdir-avg↑CLIPdir-avg-cos↑CLIPdiff-edit↓CLIPdiff-noedit↓
Tailor3D [32]1.4163.7101.4174.59312.0505.619
MVEdit [8]0.7822.5780.7833.16410.0144.118
Vox-E [37]1.6226.1781.6218.15310.5283.432
PrEditor3D (Ours)1.7827.6791.78211.5548.8122.636
", + "bbox": [ + 114, + 87, + 883, + 167 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/50860f0b52e770b119c8958ac0fc3009931d272ac19d8ac2cb8e31ff5b425522.jpg", + "table_caption": [ + "Table 3. Directional CLIP score metrics [37] for evaluating editing fidelity and prompt consistency. Our method outperforms baselines across all directional CLIP metrics. Metrics are scaled by 100 to ease reading and allow for more precision." + ], + "table_footnote": [], + "table_body": "
MethodEditingMergingTotal
Tailor3D [32]26 sec-26 sec
MVEdit [8]6 min-6 min
Vox-E[37]60 min15 min75 min
PrEditor3D (Ours)24 sec50 sec74 sec
", + "bbox": [ + 109, + 210, + 464, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Runtime comparison. We measure the runtime of our baseline methods.", + "bbox": [ + 89, + 292, + 482, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "text-image direction vector difference with cosine distance, while $\\mathrm{CLIP}_{\\mathrm{dir - avg}}$ and $\\mathrm{CLIP}_{\\mathrm{dir - avg - cos}}$ compute the same metrics by averaging image vectors first rather than scores.", + "bbox": [ + 89, + 335, + 482, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We finally introduce two additional directional metrics. $\\mathrm{CLIP}_{\\mathrm{diff - edit}}$ is the CLIP score difference between input and output image-text pairs concerning only the edited part of the input and output text prompts. $\\mathrm{CLIP}_{\\mathrm{diff - noedit}}$ is the CLIP score difference between input and output using a fixed text prompt where the edited part of the input text is replaced with a generic word, i.e., \"object.\" These metrics enforce that the CLIP text matching scores are preserved between the input and edited shapes, both for edited and unedited parts of the text.", + "bbox": [ + 89, + 381, + 482, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report all directional metrics multiplied by 100 for higher precision. We refer readers to our supplementary material for further details about the evaluation metrics.", + "bbox": [ + 89, + 532, + 483, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 590, + 184, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method can generate various impressive edited shapes from complex input shapes and prompts. We illustrate a variety of our results in Fig. 5. Our approach flexibly edits various different elements of the 3D objects, for instance replacing a \"sword\" of a skull warrior with a \"viking axe,\" resulting in coherent, seamless edits in both texture and geometry. Our edits also follow the structure of the input shape when applicable; for instance, when replacing a curvy fork with a worm, the worm maintains the same curved structure as the initial fork. We can also insert new objects, such as a \"pillow\" or \"sunglasses.\" Our method even enables both replacement and addition at the same time as in \"cat with a tail\" example, replacing the chicken with a cat and simultaneously placing a tail at the back of the car.", + "bbox": [ + 88, + 612, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the same prompt, our method can generate different results with different seeds. In Fig. 6, we show generations of \"cat\" and \"dog\" samples with various seeds. The resulting shapes vary in their head, eye, and ear structures with different colors and sizes. We can further con", + "bbox": [ + 88, + 824, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1febbf828f36e56d7bf2d1ffac47820f98213d68676758ce42b26068966e7fe.jpg", + "image_caption": [ + "Figure 6. Multiple generations and detailed control through prompt. Our method can generate different results for the same prompt using different seeds. Moreover, Our method can handle detailed prompts that can modify various aspects of the shape such as appearance and mood. For instance, here we can define the type of the cat (e.g. ginger cat) and the mood (e.g., happy)." + ], + "image_footnote": [], + "bbox": [ + 540, + 210, + 872, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trol the various aspects of the generated shape through user prompts. This is also shown in Fig. 6, where we adjust the mood and appearance of the generated shape.", + "bbox": [ + 511, + 636, + 906, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Baselines. Fig. 4 shows a qualitative comparison of our method against several state-of-the-art methods: Tailor3D [32], MVEdit [8] and Vox-E [37]. Our approach shows significant improvements, in both editing quality as well as consistency with the original shape.", + "bbox": [ + 511, + 686, + 905, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similar to our method, Vox-E allows controllable editing through merging but at the expense of an expensive SDS-based optimization that can tend towards more global changes than local ones. Since Tailor3D accepts edited front and back views as input, we ran their method using our multi-view editing results. Tab. 1 shows a comparison using GPTEval3D. While our method is consistently preferred, improvements are not as large since this metric does not measure consistency with the input. A method could", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "646", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d3e6ac84cfa0f993766bdb6f902025fafb4de616ce8915e80e633b29aa2eac5.jpg", + "image_caption": [ + "Figure 7. Qualitative ablation of our merging algorithm. We can keep the original parts of the input fixed. Here when we insert a cat, the editing breaks the neighboring regions. Thanks to our merging algorithm, we can recover the original parts of the shape." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 473, + 246 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e7a1f1310ab2516d2b82a86b37c27cc213a3d34d5d4bca17b0a4d33fdc5dade6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodChamfer Distance ↓
w/o Merging2.95
w/o Average Merging2.29
Ours2.28
", + "bbox": [ + 127, + 311, + 444, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Quantitative ablation study of our merging algorithm. We calculate the chamfer distance to the input shape for each ablation. Chamfer Distance value is multiplied by $10^{3}$ . Our algorithm is effective in keeping the edited shape consistent with the input.", + "bbox": [ + 89, + 392, + 482, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "globally change the shape and still achieve better results. This is because this metric does not take input shape into account and only considers edited output and the prompt. To complement this metric, we calculated the directional CLIP score and its few variants in Tab. 3; this considers consistency with the input, and demonstrates that our approach achieves significant improvements over the baselines.", + "bbox": [ + 89, + 465, + 483, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Perceptual Study. We prepare a perceptual study to compare our method with three baselines asking users three questions: \"Select the one that follows the following prompt more closely\", \"Select the one with better visual quality\", \"Which example better preserves the parts that were not instructed to be edited with the prompt?\". There are 360 questions in total, each answered by 10 different participants, totaling 3600 responses. Results are presented in Tab. 2. In all questions, our method is preferred over the baselines.", + "bbox": [ + 89, + 577, + 483, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Runtime Analysis. Our method enables fast iteration, taking around 24 seconds to obtain initial multi-view editing results. Merging then takes another 50 seconds to produce a final refined shape. Tab. 4 shows a comparison with baselines, using a single RTX 3090 for measurements, except for MVDream, which we run on RTX A6000. Tailor3D [32], concurrent to our work, also operates fast, taking 2 seconds for a forward pass using our multi-view editing results as input (26 seconds in total). MVEdit [8] does not employ any merging, performing editing in around 6 minutes. Since Vox-E [37] involves a long SDS optimization process, its overall inference can take around an hour.", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 513, + 90, + 624, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 5 and Fig. 7 ablate our merging approach, measuring the chamfer distance between the edited and input shapes.", + "bbox": [ + 511, + 112, + 906, + 142 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Shape Preservation through Merging. Our merging algorithm ensures that only the regions described by the user through a mask and prompt changes. In this ablation study, we only do multi-view editing, and leave out the merging operation. As shown in Fig. 7, without any merging, regions that are not intended by the user can change. In the \"chicken in a racing car\" example, when the user replaces the chicken with a cat, some part of the car is also altered since the user mask covers that area. In our merging step, we detect the changed region (\"cat\") and erased region (\"chicken\") so that we keep the rest of the shape (\"car\") fixed.", + "bbox": [ + 511, + 146, + 906, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Average Merging. After edited regions are detected, we merge the voxel grids of the input and edited reconstructions to preserve consistency with the input. In contrast, VoxE [37] uses copy-pasting for merging. That is, they copy the detected part from the edited shape and paste it into the original shape. However, a simple copy-paste approach can create boundary artifacts such as gaps between the edited region and the original shape, as shown in Fig. 7. To fix these boundary problems, we dilate the masks. Within the dilated region, we take the average of the edited shape and the original shape, which provides a smoother transition.", + "bbox": [ + 511, + 316, + 908, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although our method can generate high-quality editing results, we are limited by the 256x256 resolution of the multi-view diffusion model, MVDream [38]. In addition, our method currently focuses on 3D assets that can be rendered from four inward-facing views. However, this assumption cannot effectively capture large-scale scenes, such as indoor rooms where more views within the scene are needed.", + "bbox": [ + 511, + 486, + 908, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 619, + 633, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a fast and controllable 3D editing method that can handle a wide variety of 3D shapes and editing prompts. We employ the editing strength of powerful multi-view models, lift edits to 3D, and merge edits in 3D in order to ensure unedited regions remain consistent with the input shape. Hence, our method produces high-quality editing results with fast runtime speeds. We believe this shows a significant potential for high-quality, controllable, seamless, and fast 3D editing.", + "bbox": [ + 511, + 645, + 908, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work is partially done during Ziya's and Can's internships at Snap. Matthias Nießner was supported by the ERC Starting Grant Scan2CAD (804724) and Angela Dai was supported by the ERC Starting Grant SpatialSem (101076253).", + "bbox": [ + 511, + 811, + 908, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "647", + "bbox": [ + 486, + 944, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. Shapetalk: A language dataset and framework for 3d shape edits and deformations. In CVPR, 2023. 3", + "[2] Antonio Alliegro, Yawar Siddiqui, Tatiana Tommasi, and Matthias Nießner. Polydiff: Generating 3d polygonal meshes with diffusion models. arXiv preprint arXiv:2312.11417, 2023. 2", + "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18208-18218, 2022. 3", + "[4] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In CVPR, 2024. 3", + "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 3", + "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3", + "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG, 2023. 3", + "[8] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 2, 3, 5, 6, 7, 8", + "[9] Minghao Chen, Junyu Xie, Iro Laina, and Andrea Vedaldi. Shap-editor: Instruction-guided latent 3d editing in seconds. In CVPR, 2024. 2, 3", + "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 3", + "[11] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023. 3", + "[12] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. In CVPR, 2024. 3", + "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 3, 6", + "[14] Shaocong Dong, Lihe Ding, Zhanpeng Huang, Zibin Wang, Tianfan Xue, and Dan Xu. Interactive3d: Create what you want by interactive 3d generation. In CVPR, 2024. 2, 3", + "[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "quality dataset of 3d scanned household items. In ICRA, 2022.6", + "[16] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG, 2022. 6", + "[17] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In ICCV, 2023. 2, 3", + "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In ICLR, 2022. 2, 3, 4", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3", + "[20] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In ICLR, 2024. 3", + "[21] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In CVPR, 2024. 2, 3, 4", + "[22] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2024. 3", + "[23] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 2, 3", + "[24] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 4", + "[25] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In CVPR, 2023. 3", + "[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 3", + "[27] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3", + "[28] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 3", + "[29] Bharath Raj Nagoor Kani, Hsin-Ying Lee, Sergey Tulyakov, and Shubham Tulsiani. Upfusion: Novel view diffusion from unposed sparse view observations. In ECCV, 2025. 3", + "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "648", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.3", + "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3", + "[32] Zhangyang Qi, Yunhan Yang, Mengchen Zhang, Long Xing, Xiaoyang Wu, Tong Wu, Dahua Lin, Xihui Liu, Jiaqi Wang, and Hengshuang Zhao. Tailor3d: Customized 3d assets editing and generation with dual-side images. arXiv preprint arXiv:2407.06191, 2024. 2, 3, 5, 6, 7, 8", + "[33] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In ICLR, 2024. 3", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6", + "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 4", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3", + "[37] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. In ICCV, 2023. 2, 3, 5, 6, 7, 8", + "[38] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2024. 2, 4, 8", + "[39] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In CVPR, 2024. 2", + "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 3", + "[41] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3", + "[42] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 3", + "[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 3", + "[44] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 3" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. In CVPR, 2024. 3, 6", + "[46] Jiale Xu, Xintao Wang, Yan-Pei Cao, Weihao Cheng, Ying Shan, and Shenghua Gao. Instructp2p: Learning to edit 3d point clouds with text instructions. arXiv preprint arXiv:2306.07154, 2023. 3", + "[47] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al. Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. In ICLR, 2024. 3", + "[48] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, 2024. 2", + "[49] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 3", + "[50] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In NeurIPS, 2024. 3", + "[51] Peiye Zhuang, Songfang Han, Chaoyang Wang, Aliaksandr Siarohin, Jiaxu Zou, Michael Vasilkovsky, Vladislav Shakhrai, Sergey Korolev, Sergey Tulyakov, and HsinYing Lee. Gtr: Improving large 3d reconstruction models through geometry and texture refinement. arXiv preprint arXiv:2406.05649, 2024. 2, 3, 5, 6" + ], + "bbox": [ + 516, + 90, + 903, + 500 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "649", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/0f1a120c-8d95-4d35-a553-6cd4183aad0f_model.json b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/0f1a120c-8d95-4d35-a553-6cd4183aad0f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..76372781fcdd14da2ea8d72dbd4f9e185f8c96cb --- /dev/null +++ b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/0f1a120c-8d95-4d35-a553-6cd4183aad0f_model.json @@ -0,0 +1,2607 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.258, + 0.131, + 0.743, + 0.154 + ], + "angle": 0, + "content": "PrEditor3D: Fast and Precise 3D Shape Editing" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.18, + 0.222, + 0.201 + ], + "angle": 0, + "content": "Ziya Erkoc1" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.18, + 0.367, + 0.199 + ], + "angle": 0, + "content": "Can Gümeli" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.205, + 0.375, + 0.223 + ], + "angle": 0, + "content": "Peter Wonka2,3" + }, + { + "type": "text", + "bbox": [ + 0.407, + 0.18, + 0.548, + 0.2 + ], + "angle": 0, + "content": "Chaoyang Wang²" + }, + { + "type": "text", + "bbox": [ + 0.436, + 0.205, + 0.563, + 0.224 + ], + "angle": 0, + "content": "Hsin-Ying Lee" + }, + { + "type": "text", + "bbox": [ + 0.586, + 0.181, + 0.732, + 0.198 + ], + "angle": 0, + "content": "Matthias Nießner1" + }, + { + "type": "text", + "bbox": [ + 0.624, + 0.206, + 0.743, + 0.225 + ], + "angle": 0, + "content": "Peiye Zhuang" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.23, + 0.717, + 0.251 + ], + "angle": 0, + "content": "\\(^{1}\\)Technical University of Munich \\(^{2}\\)Snap Inc \\(^{3}\\)KAUST" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.268, + 0.687, + 0.285 + ], + "angle": 0, + "content": "https://ziyaerkoc.com/preditor3d" + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.313, + 0.254, + 0.327 + ], + "angle": 0, + "content": "\"chicken in a racing car\"" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.337, + 0.226, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.426, + 0.237, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.397, + 0.268, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.313, + 0.381, + 0.326 + ], + "angle": 0, + "content": "\"chicken cat ...\"" + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.338, + 0.399, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.426, + 0.402, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.314, + 0.532, + 0.327 + ], + "angle": 0, + "content": "chicken dog ..." + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.338, + 0.556, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.426, + 0.556, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.314, + 0.672, + 0.327 + ], + "angle": 0, + "content": "\"... raeing-car jeep\"" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.338, + 0.668, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.424, + 0.688, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.774, + 0.181, + 0.871, + 0.199 + ], + "angle": 0, + "content": "Angela Dai" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.311, + 0.895, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.498, + 0.181, + 0.508 + ], + "angle": 0, + "content": "cannon" + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.51, + 0.205, + 0.52 + ], + "angle": 0, + "content": "in a wooden cart\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.499, + 0.305, + 0.508 + ], + "angle": 0, + "content": "cannon" + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.51, + 0.329, + 0.52 + ], + "angle": 0, + "content": "flamethrower ...\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.499, + 0.407, + 0.508 + ], + "angle": 0, + "content": "cannon" + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.51, + 0.42, + 0.52 + ], + "angle": 0, + "content": "trebuchet ..." + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.503, + 0.539, + 0.514 + ], + "angle": 0, + "content": "\"house in a forest\"" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.53, + 0.202, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.54, + 0.231, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.53, + 0.328, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.531, + 0.428, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.523, + 0.535, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.539, + 0.568, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.499, + 0.656, + 0.52 + ], + "angle": 0, + "content": "house treasure cave ..." + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.522, + 0.656, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.505, + 0.781, + 0.517 + ], + "angle": 0, + "content": "\"truck carrying clay\"" + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.526, + 0.766, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.538, + 0.801, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.808, + 0.505, + 0.886, + 0.517 + ], + "angle": 0, + "content": "... elay pizza" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.528, + 0.883, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.604, + 0.184, + 0.614 + ], + "angle": 0, + "content": "viking with a" + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.615, + 0.199, + 0.625 + ], + "angle": 0, + "content": "mustache, helmet" + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.609, + 0.335, + 0.621 + ], + "angle": 0, + "content": "... helmet party hat" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.628, + 0.316, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.605, + 0.44, + 0.627 + ], + "angle": 0, + "content": "...red pepper mustache ..." + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.629, + 0.438, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.603, + 0.532, + 0.613 + ], + "angle": 0, + "content": "\"chicken in a" + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.616, + 0.528, + 0.628 + ], + "angle": 0, + "content": "racing car\"" + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.63, + 0.534, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.602, + 0.641, + 0.613 + ], + "angle": 0, + "content": "\"chicken" + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.616, + 0.662, + 0.628 + ], + "angle": 0, + "content": "cat with a tail ..." + }, + { + "type": "image", + "bbox": [ + 0.463, + 0.642, + 0.534, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.633, + 0.654, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.611, + 0.759, + 0.621 + ], + "angle": 0, + "content": "\"android\"" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.634, + 0.772, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.606, + 0.892, + 0.616 + ], + "angle": 0, + "content": "\"android chicken" + }, + { + "type": "image_caption", + "bbox": [ + 0.809, + 0.617, + 0.884, + 0.631 + ], + "angle": 0, + "content": "wearing a tie\"" + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.633, + 0.883, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.705, + 0.907, + 0.749 + ], + "angle": 0, + "content": "Figure 1. PrEditor3D is a (top) fast and high-quality editing method that can perform precise and consistent editing only in the intended regions, keeping the rest identical. (mid) It can handle diverse editing prompts with any given 3D object. (bottom) Furthermore, it can support iterative editing, facilitating artistic workflow, and can also support editing multiple regions in a single run." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.761, + 0.327, + 0.776 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We propose a training-free approach to 3D editing that enables the editing of a single shape within a few minutes. The edited 3D mesh aligns well with the prompts, and remains identical for regions that are not intended to be altered. To this end, we first project the 3D object onto 4-view images and perform synchronized multi-view image editing along with user-guided text prompts and user-provided" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.907, + 0.899 + ], + "angle": 0, + "content": "rough masks. However, the targeted regions to be edited are ambiguous due to projection from 3D to 2D. To ensure precise editing only in intended regions, we develop a 3D segmentation pipeline that detects edited areas in 3D space, followed by a merging algorithm to seamlessly integrate edited 3D regions with the original input. Extensive experiments demonstrate the superiority of our method over previous approaches, enabling fast, high-quality editing while preserving unintended regions." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "640" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.282 + ], + "angle": 0, + "content": "Recent 3D diffusion models can generate high-quality assets that closely align with the text prompts in the form of neural fields [23, 38], meshes [2, 39], or Gaussian point clouds [48]. Although these methods generate impressive results, they lack the essential capability for precise and controllable editing of the generated outputs, a critical requirement for iterative artistic workflows. Effective 3D editing demands: (1) it should be fast enough to provide quick feedback, ideally comparable to fast 3D generation algorithms, and (2) it must allow for precise local control, enabling users to keep specific parts of the model unchanged." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.403 + ], + "angle": 0, + "content": "Enabling precise and controllable editing is still an open challenge. Several initial approaches have been proposed to tackle the challenge of 3D editing [8, 9, 14, 17, 32, 37], providing promising results but suffering from slow runtime, lack of precise control, and/or lack of 3D consistency and quality. Optimization-based techniques like SDS [9, 14, 37] or multiview training dataset updates [17] are computationally expensive, making interactive editing out of reach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.553 + ], + "angle": 0, + "content": "Additionally, they offer limited control over specific parts of the shape, as text prompts alone cannot precisely localize regions to be edited [8, 9, 17, 32, 37]. While Vox-E [37] and Shap-Editor [9] propose a mechanism to prevent original parts of the shape from being altered during editing, they do not enable precise editing due to having only text as input. Finally, one can observe various visual quality problems, such as the Janus problem, blurring, over-saturation, and overemphasizing texture changes while leaving the geometry intact or degrading." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, we propose a novel editing pipeline for 3D assets that is faster, more precise, and delivers high-quality results (See Fig. 1). As our primary goal is faster editing, we propose an editing framework leveraging a pipeline that consists of two components: a multiview diffusion algorithm and a feed-forward mesh reconstruction. Multi-view diffusion models can leverage superior 2D editing techniques, and the feed-forward mesh reconstruction bridges the gap between 2D and 3D. For better controllability, we extend multi-view image generation to multi-view image editing using 2D masks to constrain the edits to user-specified regions. The 2D masks can take various forms, including manually selected regions, handbrushed areas, or automatically generated segmentations. Automatically generated masks can constrain the edit to a limited region and make the editing more restrictive, we further elaborate on this discussion in the supplementary material. We adopt DDPM inversion [21] to extract initial noise vectors from input multi-view images and execute Prompt-to-Prompt [18] on a multi-view diffusion model [38]. We use 2D user-provided masks to blend edited and original views during the denoising. Due to the occlusion issues that may naturally occur in 2D renderings, we cannot ob" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.094, + 0.872, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.2, + 0.907, + 0.284 + ], + "angle": 0, + "content": "Figure 2. Ambiguous intended regions. The intended region to be edited is clear in 3D (e.g. the cat tail). However, after projecting to 2D, regardless of the granularity of the user-provided masks, the editing will either alter some unintended regions (e.g. the robot cat) or be too limited for reasonable editing, especially because of occlusion caused by the tail." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.907, + 0.43 + ], + "angle": 0, + "content": "tain ideal intended regions in 2D regardless of the granularity of the masks, as shown in Fig. 2. The masks are either too coarse so the unintended regions will be changed, or too fine-grained to allow reasonable editing. Without additional spatial information in 3D, multi-view editing approaches cannot fully address this challenge. Therefore, simply adopting a feed-forward reconstruction method [51] to convert edited multi-views into a 3D mesh often leads to undesirable results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.909, + 0.597 + ], + "angle": 0, + "content": "To tackle this issue, we propose using the original 3D input and 3D segmentation. We first detect the intended editing region using Grounding DINO [24] and SAM 2 [35] with the user mask and prompts. This gives an initial 2D segmentation that we subsequently lift to 3D. For this purpose, we use color coding in a multi-view to 3D reconstruction pipeline, named GTR [51], to end up with a 3D segmentation that we can use during merging. Specifically, we paint 2D segmentations in a specific color, (e.g., green) and after reconstruction, we can detect which regions are edited by querying the color in the 3D field." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Then, we perform merging to maintain the original parts of the shape. We use 3D masks to detect edited/replaced parts in GTR's [51] voxel-feature space. We extract the edited part from the new shape and replace it with the old part from the original shape. That way, we guarantee that the remaining shape will remain identical. We apply a final average blending operation so that new parts and the original shape blend smoothly." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.72, + 0.868, + 0.734 + ], + "angle": 0, + "content": "In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.78 + ], + "angle": 0, + "content": "- We propose a novel method for diffusion-based 3D object editing that is faster than previous work and enables precise, interactive editing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.905, + 0.825 + ], + "angle": 0, + "content": "- The proposed method consists of user-guided and multiview-synchronized editing and a feed-forward 3D reconstruction, enabling fast editing in a feed-forward manner." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.9 + ], + "angle": 0, + "content": "- To enable precise editing only to intended regions, we propose a voxel-based 3D segmentation method that utilizes multi-view segmentation information and propagates it to 3D, followed by an average blending operation to merge the edited and original objects." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "641" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.093, + 0.905, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.289, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Overview of PrEditor3D. Given an input 3D object, we first render its multi-view images from 4 orthogonal views. We then obtain editing input from the user, describing in text as well as rough 2D masks the desired edits. We perform synchronized multi-view editing based on the text prompts as well as the user-provided masks (Sec.3.1. Due to the rough masks and the unclear intended regions caused by ambiguous 3D-2D projection, we detect the intended regions with Grounding Dino and SAM 2 (Sec.3.2, where the segmentation results are lifted to 3D for the final merging operation (Sec.3.3)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.379, + 0.482, + 0.424 + ], + "angle": 0, + "content": "- Our editing method has superior quality to previous work. We show significant improvements in GPTEval3D [45], directional CLIP metrics, and extensive user studies." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.441, + 0.241, + 0.456 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.472, + 0.483, + 0.681 + ], + "angle": 0, + "content": "2D Editing applies global or local modifications to an image based on user instructions. It has gained significant attention for enabling a more interactive user experience in content creation. To achieve this, existing methods either fine-tune text-to-image models with specialized instructional editing datasets [46, 50], or use training-free approaches with inpainting [3, 30] or cross-attention mechanisms [6, 7, 18]. To edit user-provided images, various diffusion inversion techniques have been developed. For example, SDEdit [26] introduces noise to images to capture intermediate steps in the diffusion process, null-text inversion [28] inverts the deterministic DDIM inversion [40], and recent works [4, 21] use DDPM inversion [19] to enhance editing capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.484, + 0.745 + ], + "angle": 0, + "content": "In this work, one step to achieve 3D editing is to perform 2D multi-view image editing. We apply the DDPM inversion [19] and Prompt2Prompt [18] to operate within a diffusion-based sparse multi-view generation model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "3D Reconstruction from Sparse Multi-Views refers to the task of reconstructing a 3D instance from a limited number of multi-view images. To achieve this, Score Distillation Sampling (SDS) [31] and its variants [10, 23, 33, 42, 43] optimize a 3D scene representation by reconstructing the given sparse views and generating novel views through gradients from large-scale pre-trained text-to-image diffusion models [36]. However, these approaches are often computationally demanding. Beyond per-scene optimization, PixelNeRF [49] generates a NeRF representation [27] of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.905, + 0.485 + ], + "angle": 0, + "content": "a scene from sparse-view images by training the model across multiple scenes. With recent advancements in largescale 3D datasets such as Objaverse [13], follow-up works [20, 22, 25, 29, 41, 44, 47, 51] have improved this feedforward 3D reconstruction approach for object-centric 3D assets. In this work, we utilize one recent method, GTR [51], to quickly generate 3D meshes from 4 views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.791 + ], + "angle": 0, + "content": "3D Editing presents additional challenges compared to 2D editing due to the need to maintain spatial consistency in 3D. To address this, some methods train 3D editing models using paired 3D datasets [1, 46], however, these approaches are limited by a lack of diverse and complex datasets. Other methods adapt 2D editing models, such as InstructPix2Pix [5], for the 3D domain [17, 32], by iteratively updating multi-view images of a scene. However, without synchronized multi-view updates, this approach often results in flickering or inconsistent views. Alternatively, some methods propose a generation-reconstruction loop to modify 3D representations using intermediate denoised images [8] or SDS gradients [9, 11, 12, 14, 37] in the diffusion process. While these methods can achieve 3D consistency, they often struggle with quality or suffer from high computational costs. In our work, we perform multi-view editing and reconstruct the edited object in 3D. Beyond that, to preserve unchanged regions, we carefully design an approach by detecting the edited 3D regions and integrating the intended 3D edited regions into the original shapes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.605, + 0.815 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We aim to achieve fast 3D asset editing in a training-free manner, allowing for precise and user-guided edits. Our approach achieves this through multi-view image editing in 2D, followed by lifting the 2D edits into 3D. This process can be summarized into 3 main steps: (1) synchronized" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "642" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "sparse multi-view editing in 2D (Sec.3.1), (2) detecting intended editing regions across 2D views through the Grounding Dino [24] and SAM 2 [35] approach (Sec.3.2), and (3) lifting the intended editing regions to 3D and merging the edited shape into the original (Sec.3.3). Our approach is illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.191, + 0.442, + 0.208 + ], + "angle": 0, + "content": "3.1. Synchronized Sparse Multi-View Editing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.484, + 0.319 + ], + "angle": 0, + "content": "To edit a given 3D object \\(\\mathcal{O}\\), we leverage the power of 2D editing through multi-view diffusion models. We first perform synchronized sparse multi-view edits using a pretrained multi-view diffusion model. In practice, we use MVDream [38], which generates 4 orthogonal views based on a text prompt. Note that our approach remains agnostic to the specific diffusion-based multi-view generation model used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.515 + ], + "angle": 0, + "content": "We first render multi-view images from \\(\\mathcal{O}\\), then apply the DDPM diffusion inversion mechanism [21] to revert the multi-view images to their initial noise vectors, denoted as \\(\\pmb{x}^T\\), where \\(T\\) is the number of diffusion timesteps during the denoising process. We will use these vectors \\(\\pmb{x}^T\\) as the initial latent vectors for the editing process in diffusion. We denote the text prompt for the input shape and the edited shape as \\(y_i\\) and \\(y_e\\), respectively. Also, to enable better editing control and interaction, we further take user-provided masks in 4 views as input. These masks indicate target regions for the edit, denoted as \\(M_{\\mathrm{user}} \\in \\mathbb{R}^{4 \\times H \\times W}\\) where \\(H\\) and \\(W\\) are the height and weight of the images. Note that our method does not require precise and accurate masks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.668 + ], + "angle": 0, + "content": "To edit multi-view images, we apply the Prompt-to-Prompt [18] approach on the multi-view diffusion model. For simplicity, we present the basic operation at each diffusion timestep and within each attention block. To be specific, Prompt-to-Prompt [18] generates an edited latent vector \\( \\boldsymbol{x}_e' \\) by replacing the self- and cross-attention weights of the original latent vector, \\( \\boldsymbol{x}_i \\), with the edited prompt \\( y_e \\). To confine modifications to the desired regions, we blend the latent vectors \\( \\boldsymbol{x}_e' \\) and \\( \\boldsymbol{x}_i \\) using user-provided masks \\( M_{\\mathrm{user}} \\). We denote the final edited latent vector at each step as \\( \\boldsymbol{x}_e \\):" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.677, + 0.483, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {e} \\leftarrow M _ {\\text {u s e r}} \\cdot \\boldsymbol {x} _ {e} ^ {\\prime} + \\left(1 - M _ {\\text {u s e r}}\\right) \\cdot \\boldsymbol {x} _ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.749 + ], + "angle": 0, + "content": "In practice, we downsample the user-provided masks \\( M_{\\mathrm{user}} \\) to match the feature resolution at each model layer, and the edited latent vector, \\( x_{e} \\), serves as the input to the next layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Using the inverted noise vector \\(\\pmb{x}^T\\) ensures that the edited results align with the original texture style. However, the masks \\(M_{\\mathrm{user}}\\) are often imprecise or misaligned with the target regions, which can still affect regions that are not intended to be altered. Furthermore, occlusions along the depth dimension introduce additional challenges in accurately and semantically localizing the intended editing regions in 3D. We illustrate the issue in Fig. 2. To address this, we apply an automated grounding approach that detects the intended editing region in both 2D and 3D." + }, + { + "type": "code_caption", + "bbox": [ + 0.523, + 0.095, + 0.78, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Merging Voxel Features" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.112, + 0.905, + 0.271 + ], + "angle": 0, + "content": "Input: \\(V_{i}\\) and \\(V_{e} \\in \\mathbb{R}^{A \\times A \\times F}, M_{i}\\) and \\(M_{e} \\in \\mathbb{R}^{A \\times A \\times A}, d \\in \\mathbb{N}\\), and \\(\\theta \\in [0,1]\\) \nOutput: \\(V_{\\text{blend}} \\in \\mathbb{R}^{A \\times A \\times A \\times F}\\) \n1 \\(V_{i}[M_{i}] \\gets \\emptyset\\) \n2 \\(V_{i}[M_{e}] \\gets V_{e}[M_{e}]\\) \n3 \\(N \\gets \\text{Dilation}(M_{e}, d)\\) \n4 \\(K \\gets M_{e} \\oplus N // \\oplus\\) means XOR \n5 \\(V_{\\text{blend}} \\gets V_{i}\\) \n6 \\(V_{\\text{blend}}[K] \\gets \\theta \\odot V_{i}[K] + (1 - \\theta) \\odot V_{e}[K]\\) \n7 return \\(V_{\\text{blend}}\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.299, + 0.888, + 0.316 + ], + "angle": 0, + "content": "3.2. Detection of Intended Editing Regions in 2D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.907, + 0.427 + ], + "angle": 0, + "content": "The intended editing regions refer to the specific semantic areas that correspond to the editing prompt. For example, in Fig. 1, changing \"chicken\" to \"cat\" implies that the desired editing region pertains only to the areas representing the \"chicken\" and \"cat\" concepts. Regions (both 2D and 3D) that are not semantically related to these concepts should remain unchanged." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.909, + 0.58 + ], + "angle": 0, + "content": "To ensure that only intended editing regions are edited while allowing rough user-provided masks, we propose to detect the intended editing regions by applying Grounding Dino [24] and SAM 2 [35] to both the original and the edited multi-view images. To begin with, we identify the changing concept by comparing the original prompt \\( y_{i} \\) and the editing prompt \\( y_{e} \\). We then localize the changing concept within the user-provided mask regions \\( M_{\\mathrm{user}} \\) and obtain corresponding bounding boxes in multiple views. Formally, we write the procedure as" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.591, + 0.905, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\operatorname {b b o x} _ {x} \\leftarrow \\operatorname {G r o u n d i n g} \\left(x, M _ {\\text {u s e r}}, y _ {i}, y _ {e}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.905, + 0.649 + ], + "angle": 0, + "content": "where \\(\\mathsf{bbox}_x\\) are the bounding boxes for the changing concept in the multi-view images \\(x\\in \\{\\pmb {x}_i,\\pmb {x}_e\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.739 + ], + "angle": 0, + "content": "Afterward, we segment and track the changing elements across views using the grounding bounding boxes \\(\\mathsf{bbox}_x\\). Formally, this procedure can be described as \\(\\mathrm{SAM}(\\pmb{x},\\mathsf{bbox}_x)\\). This process yields the intended editing regions in segmentation format for both the original and edited multi-views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.749, + 0.798, + 0.766 + ], + "angle": 0, + "content": "3.3. Lifting and Merging Edits in 3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.771, + 0.905, + 0.802 + ], + "angle": 0, + "content": "Finally, we lift the 2D edits to 3D and merge the 3D-edited regions into the original shape." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.805, + 0.907, + 0.867 + ], + "angle": 0, + "content": "3D Segmentation by Lifting 2D Segmentations. We mark the intended editing regions on multi-view images using a green color. The color-coded multi-view images are then reconstructed in 3D using an offline 3D reconstruction" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1We use \\( \\pmb{x} \\) to represent both latent vectors and images, without distinguishing between the two." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "643" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.094, + 0.881, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.575, + 0.852, + 0.59 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison. Our method can perform diverse editing samples and only edit the intended regions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.616, + 0.486, + 0.873 + ], + "angle": 0, + "content": "model [51]. This model takes multiple views as input and represents the shape as a triplane feature. Two separate decoders—one for Signed Distance Function (SDF) and one for color—generate a geometry field and a color field, respectively. Through this process, we create a 3D segmentation field from the color-coded multi-view images. For each 3D position within the 3D space represented by the triplane, we determine whether the position lies within the 3D intended editing regions based on its color value. That is, we apply a decision threshold for the distance between each color value and the preset green color to identify targeted regions. This produces two 3D masks, indicating the intended editing regions in 3D for both the original shape and the edited shape, denoted as \\( M_{i} \\) and \\( M_{e} \\), respectively. Our selection of green color is arbitrary. We can flexibly change the color, if the shape already has the specific green color we use." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Merging Edits in 3D. As aforementioned, coarse user" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.907, + 0.707 + ], + "angle": 0, + "content": "provided masks and occlusion issues in 2D projections often result in unwanted alterations, compromising the preservation of unaffected 3D regions. Therefore, the directly reconstructed shapes from the edited multi-view images using the reconstruction model [51] cannot be the final editing output, as illustrated in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.909, + 0.798 + ], + "angle": 0, + "content": "Using the 3D reconstruction model [51], we extract voxel features for both original and edited shapes by interpolating their triplane features, denoted as \\( \\mathbf{V}_i \\) and \\( \\mathbf{V}_e \\) for the original and edited shapes, respectively. Both \\( \\mathbf{V}_i, \\mathbf{V}_e \\in \\mathbb{R}^{A \\times A \\times A \\times F} \\), where \\( A \\) is the voxel resolution, and \\( F \\) is the feature dimension, \\( A = 256 \\) and \\( F = 40 \\) in practice." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.798, + 0.908, + 0.859 + ], + "angle": 0, + "content": "To merge 3D features, we first nullify the original specific regions \\( M_{i} \\) from the original voxel feature \\( V_{i} \\), and then replace the target edited regions \\( M_{e} \\) with edited feature \\( V_{e}[M_{e}] \\). We write the above operations as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.864, + 0.905, + 0.881 + ], + "angle": 0, + "content": "\\[\nV _ {i} \\left[ M _ {i} \\right] \\leftarrow \\emptyset , \\text {a n d} V _ {i} \\left[ M _ {e} \\right] \\leftarrow V _ {e} \\left[ M _ {e} \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "We refer to this approach as a naive copy-paste method." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "644" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.094, + 0.887, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.44, + 0.851, + 0.457 + ], + "angle": 0, + "content": "Figure 5. More editing results from PrEditor3D. Our method can perform a wide range of editing on various shapes." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.47, + 0.493, + 0.525 + ], + "angle": 0, + "content": "
MethodPrompt Align.3D PlausibilityTextureOverall
Tailor3D [32]98%99%99%99%
MVEdit [8]57%55%55%57%
Vox-E [37]53%68%50%55%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.598 + ], + "angle": 0, + "content": "Table 1. Comparison using GPTEval3D [45]. Scores indicate the percentage of our method being selected over baselines. We ask GPT-4V to compare our method against the baselines and choose one based on pre-defined categories. Each number here shows the percentage of our method being chosen." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.812 + ], + "angle": 0, + "content": "While theoretically plausible, we observe that this straightforward approach typically introduces discontinuities at the 3D editing boundaries, as shown in Fig. 7. To address this, we propose an averaged merging approach that provides a more robust blend of 3D features. In the improved method, we dilate the 3D mask \\( M_{e} \\) by a dilation \\( d \\) and then use an exclusive or (a.k.a. XOR) operation to select the boundary mask regions for smooth blending. Next, we linearly interpolate the two voxel features \\( V_{i} \\) and \\( V_{e} \\) within the boundary regions using a coefficient \\( \\theta \\), in practice \\( \\theta = 0.5 \\). We illustrate the merging process in Alg. 1. After merging, we generate a textured mesh from the blended voxel feature, using the decoders in the 3D reconstruction model [51]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.224, + 0.841 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.212, + 0.863 + ], + "angle": 0, + "content": "4.1. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Our evaluation dataset contains 18 unique shapes and 40 editing prompts. We use shapes from GSO [15] and Obja" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.47, + 0.916, + 0.526 + ], + "angle": 0, + "content": "
MethodPrompt Align.Visual QualityPreserving Shape
Tailor3D [32]96%97%99%
MVEdit [8]68%68%97%
Vox-E [37]78%88%89%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.528, + 0.907, + 0.585 + ], + "angle": 0, + "content": "Table 2. User study results comparing our method against baselines. The percentage shows the preference for our method. We show users binary comparisons of our method against the baselines and collect their response on which one they prefer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.64 + ], + "angle": 0, + "content": "verse [13]. We evaluate our method based on the quality of the editing and consistency with the input shapes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.732 + ], + "angle": 0, + "content": "We use the GPTEval3D [45] metric to evaluate the quality of edited shapes and their alignment with the text prompts. GPT-4V is provided with multi-view renderings of two methods at a time, and instructed to pick one based on text-prompt alignment, 3D plausibility, and texture details. There were 120 total questions, each answered 3 times." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Since the GPTEval3D metric does not consider the input shape, it cannot measure whether the shape remained intact, and whether the edited shape is consistent with the high-level style of the input shape. Therefore, we adopt the directional CLIP [34] score metric, \\(\\mathrm{CLIP}_{\\mathrm{dir}}\\) from previous works [16, 37]. \\(\\mathrm{CLIP}_{\\mathrm{dir}}\\) evaluates the average difference between text feature change direction and image feature change direction, where images are multi-view renderings of input and edited shapes. To ensure our evaluation is not affected by a particular implementation of this metric, we introduce three variants. \\(\\mathrm{CLIP}_{\\mathrm{dir - cos}}\\) replaces the" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "645" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.088, + 0.885, + 0.169 + ], + "angle": 0, + "content": "
MethodCLIPdir↑CLIPdir-cos↑CLIPdir-avg↑CLIPdir-avg-cos↑CLIPdiff-edit↓CLIPdiff-noedit↓
Tailor3D [32]1.4163.7101.4174.59312.0505.619
MVEdit [8]0.7822.5780.7833.16410.0144.118
Vox-E [37]1.6226.1781.6218.15310.5283.432
PrEditor3D (Ours)1.7827.6791.78211.5548.8122.636
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.172, + 0.908, + 0.201 + ], + "angle": 0, + "content": "Table 3. Directional CLIP score metrics [37] for evaluating editing fidelity and prompt consistency. Our method outperforms baselines across all directional CLIP metrics. Metrics are scaled by 100 to ease reading and allow for more precision." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.212, + 0.465, + 0.294 + ], + "angle": 0, + "content": "
MethodEditingMergingTotal
Tailor3D [32]26 sec-26 sec
MVEdit [8]6 min-6 min
Vox-E[37]60 min15 min75 min
PrEditor3D (Ours)24 sec50 sec74 sec
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.294, + 0.483, + 0.322 + ], + "angle": 0, + "content": "Table 4. Runtime comparison. We measure the runtime of our baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.483, + 0.381 + ], + "angle": 0, + "content": "text-image direction vector difference with cosine distance, while \\(\\mathrm{CLIP}_{\\mathrm{dir - avg}}\\) and \\(\\mathrm{CLIP}_{\\mathrm{dir - avg - cos}}\\) compute the same metrics by averaging image vectors first rather than scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.483, + 0.532 + ], + "angle": 0, + "content": "We finally introduce two additional directional metrics. \\(\\mathrm{CLIP}_{\\mathrm{diff - edit}}\\) is the CLIP score difference between input and output image-text pairs concerning only the edited part of the input and output text prompts. \\(\\mathrm{CLIP}_{\\mathrm{diff - noedit}}\\) is the CLIP score difference between input and output using a fixed text prompt where the edited part of the input text is replaced with a generic word, i.e., \"object.\" These metrics enforce that the CLIP text matching scores are preserved between the input and edited shapes, both for edited and unedited parts of the text." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.579 + ], + "angle": 0, + "content": "We report all directional metrics multiplied by 100 for higher precision. We refer readers to our supplementary material for further details about the evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.591, + 0.185, + 0.605 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.613, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Our method can generate various impressive edited shapes from complex input shapes and prompts. We illustrate a variety of our results in Fig. 5. Our approach flexibly edits various different elements of the 3D objects, for instance replacing a \"sword\" of a skull warrior with a \"viking axe,\" resulting in coherent, seamless edits in both texture and geometry. Our edits also follow the structure of the input shape when applicable; for instance, when replacing a curvy fork with a worm, the worm maintains the same curved structure as the initial fork. We can also insert new objects, such as a \"pillow\" or \"sunglasses.\" Our method even enables both replacement and addition at the same time as in \"cat with a tail\" example, replacing the chicken with a cat and simultaneously placing a tail at the back of the car." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Given the same prompt, our method can generate different results with different seeds. In Fig. 6, we show generations of \"cat\" and \"dog\" samples with various seeds. The resulting shapes vary in their head, eye, and ear structures with different colors and sizes. We can further con" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.212, + 0.874, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.532, + 0.907, + 0.616 + ], + "angle": 0, + "content": "Figure 6. Multiple generations and detailed control through prompt. Our method can generate different results for the same prompt using different seeds. Moreover, Our method can handle detailed prompts that can modify various aspects of the shape such as appearance and mood. For instance, here we can define the type of the cat (e.g. ginger cat) and the mood (e.g., happy)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.907, + 0.682 + ], + "angle": 0, + "content": "trol the various aspects of the generated shape through user prompts. This is also shown in Fig. 6, where we adjust the mood and appearance of the generated shape." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Comparison with Baselines. Fig. 4 shows a qualitative comparison of our method against several state-of-the-art methods: Tailor3D [32], MVEdit [8] and Vox-E [37]. Our approach shows significant improvements, in both editing quality as well as consistency with the original shape." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Similar to our method, Vox-E allows controllable editing through merging but at the expense of an expensive SDS-based optimization that can tend towards more global changes than local ones. Since Tailor3D accepts edited front and back views as input, we ran their method using our multi-view editing results. Tab. 1 shows a comparison using GPTEval3D. While our method is consistently preferred, improvements are not as large since this metric does not measure consistency with the input. A method could" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "646" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.475, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.248, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Figure 7. Qualitative ablation of our merging algorithm. We can keep the original parts of the input fixed. Here when we insert a cat, the editing breaks the neighboring regions. Thanks to our merging algorithm, we can recover the original parts of the shape." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.312, + 0.445, + 0.389 + ], + "angle": 0, + "content": "
MethodChamfer Distance ↓
w/o Merging2.95
w/o Average Merging2.29
Ours2.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.393, + 0.483, + 0.45 + ], + "angle": 0, + "content": "Table 5. Quantitative ablation study of our merging algorithm. We calculate the chamfer distance to the input shape for each ablation. Chamfer Distance value is multiplied by \\(10^{3}\\). Our algorithm is effective in keeping the edited shape consistent with the input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.572 + ], + "angle": 0, + "content": "globally change the shape and still achieve better results. This is because this metric does not take input shape into account and only considers edited output and the prompt. To complement this metric, we calculated the directional CLIP score and its few variants in Tab. 3; this considers consistency with the input, and demonstrates that our approach achieves significant improvements over the baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Perceptual Study. We prepare a perceptual study to compare our method with three baselines asking users three questions: \"Select the one that follows the following prompt more closely\", \"Select the one with better visual quality\", \"Which example better preserves the parts that were not instructed to be edited with the prompt?\". There are 360 questions in total, each answered by 10 different participants, totaling 3600 responses. Results are presented in Tab. 2. In all questions, our method is preferred over the baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Runtime Analysis. Our method enables fast iteration, taking around 24 seconds to obtain initial multi-view editing results. Merging then takes another 50 seconds to produce a final refined shape. Tab. 4 shows a comparison with baselines, using a single RTX 3090 for measurements, except for MVDream, which we run on RTX A6000. Tailor3D [32], concurrent to our work, also operates fast, taking 2 seconds for a forward pass using our multi-view editing results as input (26 seconds in total). MVEdit [8] does not employ any merging, performing editing in around 6 minutes. Since Vox-E [37] involves a long SDS optimization process, its overall inference can take around an hour." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.625, + 0.105 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.143 + ], + "angle": 0, + "content": "Tab. 5 and Fig. 7 ablate our merging approach, measuring the chamfer distance between the edited and input shapes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.147, + 0.907, + 0.314 + ], + "angle": 0, + "content": "Shape Preservation through Merging. Our merging algorithm ensures that only the regions described by the user through a mask and prompt changes. In this ablation study, we only do multi-view editing, and leave out the merging operation. As shown in Fig. 7, without any merging, regions that are not intended by the user can change. In the \"chicken in a racing car\" example, when the user replaces the chicken with a cat, some part of the car is also altered since the user mask covers that area. In our merging step, we detect the changed region (\"cat\") and erased region (\"chicken\") so that we keep the rest of the shape (\"car\") fixed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.317, + 0.909, + 0.483 + ], + "angle": 0, + "content": "Average Merging. After edited regions are detected, we merge the voxel grids of the input and edited reconstructions to preserve consistency with the input. In contrast, VoxE [37] uses copy-pasting for merging. That is, they copy the detected part from the edited shape and paste it into the original shape. However, a simple copy-paste approach can create boundary artifacts such as gaps between the edited region and the original shape, as shown in Fig. 7. To fix these boundary problems, we dilate the masks. Within the dilated region, we take the average of the edited shape and the original shape, which provides a smoother transition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.909, + 0.609 + ], + "angle": 0, + "content": "Limitations. Although our method can generate high-quality editing results, we are limited by the 256x256 resolution of the multi-view diffusion model, MVDream [38]. In addition, our method currently focuses on 3D assets that can be rendered from four inward-facing views. However, this assumption cannot effectively capture large-scale scenes, such as indoor rooms where more views within the scene are needed." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.634, + 0.637 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.909, + 0.783 + ], + "angle": 0, + "content": "We propose a fast and controllable 3D editing method that can handle a wide variety of 3D shapes and editing prompts. We employ the editing strength of powerful multi-view models, lift edits to 3D, and merge edits in 3D in order to ensure unedited regions remain consistent with the input shape. Hence, our method produces high-quality editing results with fast runtime speeds. We believe this shows a significant potential for high-quality, controllable, seamless, and fast 3D editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.909, + 0.889 + ], + "angle": 0, + "content": "Acknowledgments This work is partially done during Ziya's and Can's internships at Snap. Matthias Nießner was supported by the ERC Starting Grant Scan2CAD (804724) and Angela Dai was supported by the ERC Starting Grant SpatialSem (101076253)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "647" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.17 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. Shapetalk: A language dataset and framework for 3d shape edits and deformations. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.482, + 0.226 + ], + "angle": 0, + "content": "[2] Antonio Alliegro, Yawar Siddiqui, Tatiana Tommasi, and Matthias Nießner. Polydiff: Generating 3d polygonal meshes with diffusion models. arXiv preprint arXiv:2312.11417, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18208-18218, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[4] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.556 + ], + "angle": 0, + "content": "[8] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.481, + 0.599 + ], + "angle": 0, + "content": "[9] Minghao Chen, Junyu Xie, Iro Laina, and Andrea Vedaldi. Shap-editor: Instruction-guided latent 3d editing in seconds. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.602, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[11] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[12] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[14] Shaocong Dong, Lihe Ding, Zhanpeng Huang, Zibin Wang, Tianfan Xue, and Dan Xu. Interactive3d: Create what you want by interactive 3d generation. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "quality dataset of 3d scanned household items. In ICRA, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[16] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[17] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In ICLR, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.904, + 0.291 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.294, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[20] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.391 + ], + "angle": 0, + "content": "[21] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In CVPR, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.393, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[22] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[23] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[24] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[25] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[27] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[28] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[29] Bharath Raj Nagoor Kani, Hsin-Ying Lee, Sergey Tulyakov, and Shubham Tulsiani. Upfusion: Novel view diffusion from unposed sparse view observations. In ECCV, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "648" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.162 + ], + "angle": 0, + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.484, + 0.234 + ], + "angle": 0, + "content": "[32] Zhangyang Qi, Yunhan Yang, Mengchen Zhang, Long Xing, Xiaoyang Wu, Tong Wu, Dahua Lin, Xihui Liu, Jiaqi Wang, and Hengshuang Zhao. Tailor3d: Customized 3d assets editing and generation with dual-side images. arXiv preprint arXiv:2407.06191, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[33] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.374 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.483, + 0.445 + ], + "angle": 0, + "content": "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.531 + ], + "angle": 0, + "content": "[37] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. In ICCV, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[38] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2024. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[39] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.66 + ], + "angle": 0, + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[41] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[42] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[44] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[45] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. In CVPR, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[46] Jiale Xu, Xintao Wang, Yan-Pei Cao, Weihao Cheng, Ying Shan, and Shenghua Gao. Instructp2p: Learning to edit 3d point clouds with text instructions. arXiv preprint arXiv:2306.07154, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[47] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al. Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[48] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.904, + 0.373 + ], + "angle": 0, + "content": "[49] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[50] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[51] Peiye Zhuang, Songfang Han, Chaoyang Wang, Aliaksandr Siarohin, Jiaxu Zou, Michael Vasilkovsky, Vladislav Shakhrai, Sergey Korolev, Sergey Tulyakov, and HsinYing Lee. Gtr: Improving large 3d reconstruction models through geometry and texture refinement. arXiv preprint arXiv:2406.05649, 2024. 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "649" + } + ] +] \ No newline at end of file diff --git a/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/0f1a120c-8d95-4d35-a553-6cd4183aad0f_origin.pdf b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/0f1a120c-8d95-4d35-a553-6cd4183aad0f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a1053cef0e3abcb0733c8634d2a24ab0386d561c --- /dev/null +++ b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/0f1a120c-8d95-4d35-a553-6cd4183aad0f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf69f4830efd349555490e611c994e0572a513e9e57b1f001b5c7e1b81c83cb5 +size 3502913 diff --git a/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/full.md b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e48702b7f65744955907ab316a33c3adeca306f0 --- /dev/null +++ b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/full.md @@ -0,0 +1,359 @@ +# PrEditor3D: Fast and Precise 3D Shape Editing + +Ziya Erkoc1 + +Can Gümeli + +Peter Wonka2,3 + +Chaoyang Wang² + +Hsin-Ying Lee + +Matthias Nießner1 + +Peiye Zhuang + +$^{1}$ Technical University of Munich $^{2}$ Snap Inc $^{3}$ KAUST + +https://ziyaerkoc.com/preditor3d + +![](images/9c48c9fdd0cf5631c221117521fc1b9b1344c40a5bc7f89944adb8367fd7393f.jpg) +"chicken in a racing car" + +![](images/23b1c8cd2d7cb0bc1f34e5aa25c4deccd09603d79b4764c929d9fafc8d399d92.jpg) +cannon + +![](images/2c81bd58f932972b1c74cc61ced5ca46c670d94d066d3653ab6b0cc544bd913b.jpg) + +![](images/8cc7d880f57287871f29f335a086755527b4ce1063a9a502313a2e31a52423a8.jpg) +"chicken cat ..." + +![](images/62692890255cfc1898b373239421ee2c28b5017bfc6ec5b0913767a15e98fa34.jpg) +cannon + +![](images/0ed015752295ae7fb404eb6679c783c6a3b6d80264dcc3c4a299031e7a8dab01.jpg) +chicken dog ... + +![](images/a74bf6ce14f352c08e3a4df63f544caf575a117ecb090213c45a7592a14d9fe4.jpg) +cannon + +![](images/16d93914e8b02aeb8cce88ee3ff194f52947c3c1761d2a8996a9341f63b48ba9.jpg) +"... raeing-car jeep" + +![](images/2f468347a2e0645996089876c58fcf8febc3b6aea3d6910bf2622c55c710342a.jpg) + +Angela Dai + +![](images/533aa91a06d59cef9f34e86ec606b1fe42a9400ade9b166c34905d771ce16dcc.jpg) + +![](images/c93d13830cebbdc3fd20e421c2d0047868e94200efb6bad0a586edd0658c52ca.jpg) +in a wooden cart" + +![](images/fd4fa33ecb83d425c3c2b706ba40c06eaf88ee091aa0d2522c8e93e8cf235f1c.jpg) +viking with a + +![](images/e70736e11c74e377b631b1205aeeb3a93298b6a794b715578fd9fabe176811fc.jpg) +flamethrower ..." + +![](images/119220910980ff47d723ef27312b3b92c8980c71bc90980fc36c8a42d7640f0a.jpg) +trebuchet ... + +![](images/ff9645074f11352b99d8ea555fd6396ae70b8abe9949202ea70ee6f0da360177.jpg) +"house in a forest" +"chicken in a + +![](images/875172e5833e4a5e78739a0eab9283c8d2f51a49479b1727569ecc004d9ead00.jpg) + +![](images/7ae10615da9277a41397c1aeaf094502e7942c78d403c55322ad5e8bd99226f0.jpg) +house treasure cave ... +"chicken + +![](images/8096860005b6cb0122e73a16c5a9a1c5a04fafb01844457f358e956358d1390a.jpg) +"truck carrying clay" +"android" + +![](images/bf2f124969d224eeb3a8ed43c086bec87e9a83d8b021634703c3deb304d76a1e.jpg) +... elay pizza +"android chicken + +![](images/0421c88025359e3c64200ad95c50bf5c163fd5d4a231a5f4eee04bc92a7f5738.jpg) +mustache, helmet +... helmet party hat + +![](images/c32bcbd3cff098474abb37d3b96be16ae7a075be91fdbe18cc8942e510e83875.jpg) +...red pepper mustache ... + +![](images/406f732f82a18a122c29eb4202639c5106e88fb59cdbe8529e2921515f6efec5.jpg) +racing car" + +![](images/073e30d0b0f8862b586d53260b056453641089686f0676451908ab775ad0519b.jpg) +cat with a tail ... +Figure 1. PrEditor3D is a (top) fast and high-quality editing method that can perform precise and consistent editing only in the intended regions, keeping the rest identical. (mid) It can handle diverse editing prompts with any given 3D object. (bottom) Furthermore, it can support iterative editing, facilitating artistic workflow, and can also support editing multiple regions in a single run. + +![](images/69e72e7cebb6abbe44a9aaa6cb9ed6c067af4dbca1aed5e99fae4134d47cba3f.jpg) + +![](images/eeb47fe2344c3616e6b6291f1645cee84b9d9f33d3d38968afc03dc693e117d3.jpg) +wearing a tie" + +# Abstract + +We propose a training-free approach to 3D editing that enables the editing of a single shape within a few minutes. The edited 3D mesh aligns well with the prompts, and remains identical for regions that are not intended to be altered. To this end, we first project the 3D object onto 4-view images and perform synchronized multi-view image editing along with user-guided text prompts and user-provided + +rough masks. However, the targeted regions to be edited are ambiguous due to projection from 3D to 2D. To ensure precise editing only in intended regions, we develop a 3D segmentation pipeline that detects edited areas in 3D space, followed by a merging algorithm to seamlessly integrate edited 3D regions with the original input. Extensive experiments demonstrate the superiority of our method over previous approaches, enabling fast, high-quality editing while preserving unintended regions. + +# 1. Introduction + +Recent 3D diffusion models can generate high-quality assets that closely align with the text prompts in the form of neural fields [23, 38], meshes [2, 39], or Gaussian point clouds [48]. Although these methods generate impressive results, they lack the essential capability for precise and controllable editing of the generated outputs, a critical requirement for iterative artistic workflows. Effective 3D editing demands: (1) it should be fast enough to provide quick feedback, ideally comparable to fast 3D generation algorithms, and (2) it must allow for precise local control, enabling users to keep specific parts of the model unchanged. + +Enabling precise and controllable editing is still an open challenge. Several initial approaches have been proposed to tackle the challenge of 3D editing [8, 9, 14, 17, 32, 37], providing promising results but suffering from slow runtime, lack of precise control, and/or lack of 3D consistency and quality. Optimization-based techniques like SDS [9, 14, 37] or multiview training dataset updates [17] are computationally expensive, making interactive editing out of reach. + +Additionally, they offer limited control over specific parts of the shape, as text prompts alone cannot precisely localize regions to be edited [8, 9, 17, 32, 37]. While Vox-E [37] and Shap-Editor [9] propose a mechanism to prevent original parts of the shape from being altered during editing, they do not enable precise editing due to having only text as input. Finally, one can observe various visual quality problems, such as the Janus problem, blurring, over-saturation, and overemphasizing texture changes while leaving the geometry intact or degrading. + +To address these challenges, we propose a novel editing pipeline for 3D assets that is faster, more precise, and delivers high-quality results (See Fig. 1). As our primary goal is faster editing, we propose an editing framework leveraging a pipeline that consists of two components: a multiview diffusion algorithm and a feed-forward mesh reconstruction. Multi-view diffusion models can leverage superior 2D editing techniques, and the feed-forward mesh reconstruction bridges the gap between 2D and 3D. For better controllability, we extend multi-view image generation to multi-view image editing using 2D masks to constrain the edits to user-specified regions. The 2D masks can take various forms, including manually selected regions, handbrushed areas, or automatically generated segmentations. Automatically generated masks can constrain the edit to a limited region and make the editing more restrictive, we further elaborate on this discussion in the supplementary material. We adopt DDPM inversion [21] to extract initial noise vectors from input multi-view images and execute Prompt-to-Prompt [18] on a multi-view diffusion model [38]. We use 2D user-provided masks to blend edited and original views during the denoising. Due to the occlusion issues that may naturally occur in 2D renderings, we cannot ob + +![](images/ed053e6f3f5efb47dd28d94b8cee1b43fd32c5310ab876809527fce6cfbba889.jpg) +Figure 2. Ambiguous intended regions. The intended region to be edited is clear in 3D (e.g. the cat tail). However, after projecting to 2D, regardless of the granularity of the user-provided masks, the editing will either alter some unintended regions (e.g. the robot cat) or be too limited for reasonable editing, especially because of occlusion caused by the tail. + +tain ideal intended regions in 2D regardless of the granularity of the masks, as shown in Fig. 2. The masks are either too coarse so the unintended regions will be changed, or too fine-grained to allow reasonable editing. Without additional spatial information in 3D, multi-view editing approaches cannot fully address this challenge. Therefore, simply adopting a feed-forward reconstruction method [51] to convert edited multi-views into a 3D mesh often leads to undesirable results. + +To tackle this issue, we propose using the original 3D input and 3D segmentation. We first detect the intended editing region using Grounding DINO [24] and SAM 2 [35] with the user mask and prompts. This gives an initial 2D segmentation that we subsequently lift to 3D. For this purpose, we use color coding in a multi-view to 3D reconstruction pipeline, named GTR [51], to end up with a 3D segmentation that we can use during merging. Specifically, we paint 2D segmentations in a specific color, (e.g., green) and after reconstruction, we can detect which regions are edited by querying the color in the 3D field. + +Then, we perform merging to maintain the original parts of the shape. We use 3D masks to detect edited/replaced parts in GTR's [51] voxel-feature space. We extract the edited part from the new shape and replace it with the old part from the original shape. That way, we guarantee that the remaining shape will remain identical. We apply a final average blending operation so that new parts and the original shape blend smoothly. + +In summary, we make the following contributions: + +- We propose a novel method for diffusion-based 3D object editing that is faster than previous work and enables precise, interactive editing. +- The proposed method consists of user-guided and multiview-synchronized editing and a feed-forward 3D reconstruction, enabling fast editing in a feed-forward manner. +- To enable precise editing only to intended regions, we propose a voxel-based 3D segmentation method that utilizes multi-view segmentation information and propagates it to 3D, followed by an average blending operation to merge the edited and original objects. + +![](images/d674394a442d9900b55227a5829b8f095cb9e0a3c2630d948156ec9be7838c96.jpg) +Figure 3. Overview of PrEditor3D. Given an input 3D object, we first render its multi-view images from 4 orthogonal views. We then obtain editing input from the user, describing in text as well as rough 2D masks the desired edits. We perform synchronized multi-view editing based on the text prompts as well as the user-provided masks (Sec.3.1. Due to the rough masks and the unclear intended regions caused by ambiguous 3D-2D projection, we detect the intended regions with Grounding Dino and SAM 2 (Sec.3.2, where the segmentation results are lifted to 3D for the final merging operation (Sec.3.3). + +- Our editing method has superior quality to previous work. We show significant improvements in GPTEval3D [45], directional CLIP metrics, and extensive user studies. + +# 2. Related Works + +2D Editing applies global or local modifications to an image based on user instructions. It has gained significant attention for enabling a more interactive user experience in content creation. To achieve this, existing methods either fine-tune text-to-image models with specialized instructional editing datasets [46, 50], or use training-free approaches with inpainting [3, 30] or cross-attention mechanisms [6, 7, 18]. To edit user-provided images, various diffusion inversion techniques have been developed. For example, SDEdit [26] introduces noise to images to capture intermediate steps in the diffusion process, null-text inversion [28] inverts the deterministic DDIM inversion [40], and recent works [4, 21] use DDPM inversion [19] to enhance editing capabilities. + +In this work, one step to achieve 3D editing is to perform 2D multi-view image editing. We apply the DDPM inversion [19] and Prompt2Prompt [18] to operate within a diffusion-based sparse multi-view generation model. + +3D Reconstruction from Sparse Multi-Views refers to the task of reconstructing a 3D instance from a limited number of multi-view images. To achieve this, Score Distillation Sampling (SDS) [31] and its variants [10, 23, 33, 42, 43] optimize a 3D scene representation by reconstructing the given sparse views and generating novel views through gradients from large-scale pre-trained text-to-image diffusion models [36]. However, these approaches are often computationally demanding. Beyond per-scene optimization, PixelNeRF [49] generates a NeRF representation [27] of + +a scene from sparse-view images by training the model across multiple scenes. With recent advancements in largescale 3D datasets such as Objaverse [13], follow-up works [20, 22, 25, 29, 41, 44, 47, 51] have improved this feedforward 3D reconstruction approach for object-centric 3D assets. In this work, we utilize one recent method, GTR [51], to quickly generate 3D meshes from 4 views. + +3D Editing presents additional challenges compared to 2D editing due to the need to maintain spatial consistency in 3D. To address this, some methods train 3D editing models using paired 3D datasets [1, 46], however, these approaches are limited by a lack of diverse and complex datasets. Other methods adapt 2D editing models, such as InstructPix2Pix [5], for the 3D domain [17, 32], by iteratively updating multi-view images of a scene. However, without synchronized multi-view updates, this approach often results in flickering or inconsistent views. Alternatively, some methods propose a generation-reconstruction loop to modify 3D representations using intermediate denoised images [8] or SDS gradients [9, 11, 12, 14, 37] in the diffusion process. While these methods can achieve 3D consistency, they often struggle with quality or suffer from high computational costs. In our work, we perform multi-view editing and reconstruct the edited object in 3D. Beyond that, to preserve unchanged regions, we carefully design an approach by detecting the edited 3D regions and integrating the intended 3D edited regions into the original shapes. + +# 3. Method + +We aim to achieve fast 3D asset editing in a training-free manner, allowing for precise and user-guided edits. Our approach achieves this through multi-view image editing in 2D, followed by lifting the 2D edits into 3D. This process can be summarized into 3 main steps: (1) synchronized + +sparse multi-view editing in 2D (Sec.3.1), (2) detecting intended editing regions across 2D views through the Grounding Dino [24] and SAM 2 [35] approach (Sec.3.2), and (3) lifting the intended editing regions to 3D and merging the edited shape into the original (Sec.3.3). Our approach is illustrated in Fig. 3. + +# 3.1. Synchronized Sparse Multi-View Editing + +To edit a given 3D object $\mathcal{O}$ , we leverage the power of 2D editing through multi-view diffusion models. We first perform synchronized sparse multi-view edits using a pretrained multi-view diffusion model. In practice, we use MVDream [38], which generates 4 orthogonal views based on a text prompt. Note that our approach remains agnostic to the specific diffusion-based multi-view generation model used. + +We first render multi-view images from $\mathcal{O}$ , then apply the DDPM diffusion inversion mechanism [21] to revert the multi-view images to their initial noise vectors, denoted as $\pmb{x}^T$ , where $T$ is the number of diffusion timesteps during the denoising process. We will use these vectors $\pmb{x}^T$ as the initial latent vectors for the editing process in diffusion. We denote the text prompt for the input shape and the edited shape as $y_i$ and $y_e$ , respectively. Also, to enable better editing control and interaction, we further take user-provided masks in 4 views as input. These masks indicate target regions for the edit, denoted as $M_{\mathrm{user}} \in \mathbb{R}^{4 \times H \times W}$ where $H$ and $W$ are the height and weight of the images. Note that our method does not require precise and accurate masks. + +To edit multi-view images, we apply the Prompt-to-Prompt [18] approach on the multi-view diffusion model. For simplicity, we present the basic operation at each diffusion timestep and within each attention block. To be specific, Prompt-to-Prompt [18] generates an edited latent vector $\boldsymbol{x}_e'$ by replacing the self- and cross-attention weights of the original latent vector, $\boldsymbol{x}_i$ , with the edited prompt $y_e$ . To confine modifications to the desired regions, we blend the latent vectors $\boldsymbol{x}_e'$ and $\boldsymbol{x}_i$ using user-provided masks $M_{\mathrm{user}}$ . We denote the final edited latent vector at each step as $\boldsymbol{x}_e$ : + +$$ +\boldsymbol {x} _ {e} \leftarrow M _ {\text {u s e r}} \cdot \boldsymbol {x} _ {e} ^ {\prime} + \left(1 - M _ {\text {u s e r}}\right) \cdot \boldsymbol {x} _ {i}. \tag {1} +$$ + +In practice, we downsample the user-provided masks $M_{\mathrm{user}}$ to match the feature resolution at each model layer, and the edited latent vector, $x_{e}$ , serves as the input to the next layer. + +Using the inverted noise vector $\pmb{x}^T$ ensures that the edited results align with the original texture style. However, the masks $M_{\mathrm{user}}$ are often imprecise or misaligned with the target regions, which can still affect regions that are not intended to be altered. Furthermore, occlusions along the depth dimension introduce additional challenges in accurately and semantically localizing the intended editing regions in 3D. We illustrate the issue in Fig. 2. To address this, we apply an automated grounding approach that detects the intended editing region in both 2D and 3D. + +Algorithm 1: Merging Voxel Features +Input: $V_{i}$ and $V_{e} \in \mathbb{R}^{A \times A \times F}, M_{i}$ and $M_{e} \in \mathbb{R}^{A \times A \times A}, d \in \mathbb{N}$ , and $\theta \in [0,1]$ +Output: $V_{\text{blend}} \in \mathbb{R}^{A \times A \times A \times F}$ +1 $V_{i}[M_{i}] \gets \emptyset$ +2 $V_{i}[M_{e}] \gets V_{e}[M_{e}]$ +3 $N \gets \text{Dilation}(M_{e}, d)$ +4 $K \gets M_{e} \oplus N // \oplus$ means XOR +5 $V_{\text{blend}} \gets V_{i}$ +6 $V_{\text{blend}}[K] \gets \theta \odot V_{i}[K] + (1 - \theta) \odot V_{e}[K]$ +7 return $V_{\text{blend}}$ + +# 3.2. Detection of Intended Editing Regions in 2D + +The intended editing regions refer to the specific semantic areas that correspond to the editing prompt. For example, in Fig. 1, changing "chicken" to "cat" implies that the desired editing region pertains only to the areas representing the "chicken" and "cat" concepts. Regions (both 2D and 3D) that are not semantically related to these concepts should remain unchanged. + +To ensure that only intended editing regions are edited while allowing rough user-provided masks, we propose to detect the intended editing regions by applying Grounding Dino [24] and SAM 2 [35] to both the original and the edited multi-view images. To begin with, we identify the changing concept by comparing the original prompt $y_{i}$ and the editing prompt $y_{e}$ . We then localize the changing concept within the user-provided mask regions $M_{\mathrm{user}}$ and obtain corresponding bounding boxes in multiple views. Formally, we write the procedure as + +$$ +\operatorname {b b o x} _ {x} \leftarrow \operatorname {G r o u n d i n g} \left(x, M _ {\text {u s e r}}, y _ {i}, y _ {e}\right), \tag {2} +$$ + +where $\mathsf{bbox}_x$ are the bounding boxes for the changing concept in the multi-view images $x\in \{\pmb {x}_i,\pmb {x}_e\}$ + +Afterward, we segment and track the changing elements across views using the grounding bounding boxes $\mathsf{bbox}_x$ . Formally, this procedure can be described as $\mathrm{SAM}(\pmb{x},\mathsf{bbox}_x)$ . This process yields the intended editing regions in segmentation format for both the original and edited multi-views. + +# 3.3. Lifting and Merging Edits in 3D + +Finally, we lift the 2D edits to 3D and merge the 3D-edited regions into the original shape. + +3D Segmentation by Lifting 2D Segmentations. We mark the intended editing regions on multi-view images using a green color. The color-coded multi-view images are then reconstructed in 3D using an offline 3D reconstruction + +![](images/fc2fa62a782513e34a7ad643267821a407c09f1c29730176f900bc37f2e9e090.jpg) +Figure 4. Qualitative comparison. Our method can perform diverse editing samples and only edit the intended regions. + +model [51]. This model takes multiple views as input and represents the shape as a triplane feature. Two separate decoders—one for Signed Distance Function (SDF) and one for color—generate a geometry field and a color field, respectively. Through this process, we create a 3D segmentation field from the color-coded multi-view images. For each 3D position within the 3D space represented by the triplane, we determine whether the position lies within the 3D intended editing regions based on its color value. That is, we apply a decision threshold for the distance between each color value and the preset green color to identify targeted regions. This produces two 3D masks, indicating the intended editing regions in 3D for both the original shape and the edited shape, denoted as $M_{i}$ and $M_{e}$ , respectively. Our selection of green color is arbitrary. We can flexibly change the color, if the shape already has the specific green color we use. + +Merging Edits in 3D. As aforementioned, coarse user + +provided masks and occlusion issues in 2D projections often result in unwanted alterations, compromising the preservation of unaffected 3D regions. Therefore, the directly reconstructed shapes from the edited multi-view images using the reconstruction model [51] cannot be the final editing output, as illustrated in Fig. 7. + +Using the 3D reconstruction model [51], we extract voxel features for both original and edited shapes by interpolating their triplane features, denoted as $\mathbf{V}_i$ and $\mathbf{V}_e$ for the original and edited shapes, respectively. Both $\mathbf{V}_i, \mathbf{V}_e \in \mathbb{R}^{A \times A \times A \times F}$ , where $A$ is the voxel resolution, and $F$ is the feature dimension, $A = 256$ and $F = 40$ in practice. + +To merge 3D features, we first nullify the original specific regions $M_{i}$ from the original voxel feature $V_{i}$ , and then replace the target edited regions $M_{e}$ with edited feature $V_{e}[M_{e}]$ . We write the above operations as follows: + +$$ +V _ {i} \left[ M _ {i} \right] \leftarrow \emptyset , \text {a n d} V _ {i} \left[ M _ {e} \right] \leftarrow V _ {e} \left[ M _ {e} \right]. \tag {3} +$$ + +We refer to this approach as a naive copy-paste method. + +![](images/86854e91fad33408573a2435b059a587351325bb154a437faa1cb460899ed5ce.jpg) +Figure 5. More editing results from PrEditor3D. Our method can perform a wide range of editing on various shapes. + +
MethodPrompt Align.3D PlausibilityTextureOverall
Tailor3D [32]98%99%99%99%
MVEdit [8]57%55%55%57%
Vox-E [37]53%68%50%55%
+ +While theoretically plausible, we observe that this straightforward approach typically introduces discontinuities at the 3D editing boundaries, as shown in Fig. 7. To address this, we propose an averaged merging approach that provides a more robust blend of 3D features. In the improved method, we dilate the 3D mask $M_{e}$ by a dilation $d$ and then use an exclusive or (a.k.a. XOR) operation to select the boundary mask regions for smooth blending. Next, we linearly interpolate the two voxel features $V_{i}$ and $V_{e}$ within the boundary regions using a coefficient $\theta$ , in practice $\theta = 0.5$ . We illustrate the merging process in Alg. 1. After merging, we generate a textured mesh from the blended voxel feature, using the decoders in the 3D reconstruction model [51]. + +# 4. Experiments + +# 4.1. Evaluation + +Our evaluation dataset contains 18 unique shapes and 40 editing prompts. We use shapes from GSO [15] and Obja + +Table 1. Comparison using GPTEval3D [45]. Scores indicate the percentage of our method being selected over baselines. We ask GPT-4V to compare our method against the baselines and choose one based on pre-defined categories. Each number here shows the percentage of our method being chosen. + +
MethodPrompt Align.Visual QualityPreserving Shape
Tailor3D [32]96%97%99%
MVEdit [8]68%68%97%
Vox-E [37]78%88%89%
+ +Table 2. User study results comparing our method against baselines. The percentage shows the preference for our method. We show users binary comparisons of our method against the baselines and collect their response on which one they prefer. + +verse [13]. We evaluate our method based on the quality of the editing and consistency with the input shapes. + +We use the GPTEval3D [45] metric to evaluate the quality of edited shapes and their alignment with the text prompts. GPT-4V is provided with multi-view renderings of two methods at a time, and instructed to pick one based on text-prompt alignment, 3D plausibility, and texture details. There were 120 total questions, each answered 3 times. + +Since the GPTEval3D metric does not consider the input shape, it cannot measure whether the shape remained intact, and whether the edited shape is consistent with the high-level style of the input shape. Therefore, we adopt the directional CLIP [34] score metric, $\mathrm{CLIP}_{\mathrm{dir}}$ from previous works [16, 37]. $\mathrm{CLIP}_{\mathrm{dir}}$ evaluates the average difference between text feature change direction and image feature change direction, where images are multi-view renderings of input and edited shapes. To ensure our evaluation is not affected by a particular implementation of this metric, we introduce three variants. $\mathrm{CLIP}_{\mathrm{dir - cos}}$ replaces the + +
MethodCLIPdir↑CLIPdir-cos↑CLIPdir-avg↑CLIPdir-avg-cos↑CLIPdiff-edit↓CLIPdiff-noedit↓
Tailor3D [32]1.4163.7101.4174.59312.0505.619
MVEdit [8]0.7822.5780.7833.16410.0144.118
Vox-E [37]1.6226.1781.6218.15310.5283.432
PrEditor3D (Ours)1.7827.6791.78211.5548.8122.636
+ +Table 3. Directional CLIP score metrics [37] for evaluating editing fidelity and prompt consistency. Our method outperforms baselines across all directional CLIP metrics. Metrics are scaled by 100 to ease reading and allow for more precision. + +
MethodEditingMergingTotal
Tailor3D [32]26 sec-26 sec
MVEdit [8]6 min-6 min
Vox-E[37]60 min15 min75 min
PrEditor3D (Ours)24 sec50 sec74 sec
+ +Table 4. Runtime comparison. We measure the runtime of our baseline methods. + +text-image direction vector difference with cosine distance, while $\mathrm{CLIP}_{\mathrm{dir - avg}}$ and $\mathrm{CLIP}_{\mathrm{dir - avg - cos}}$ compute the same metrics by averaging image vectors first rather than scores. + +We finally introduce two additional directional metrics. $\mathrm{CLIP}_{\mathrm{diff - edit}}$ is the CLIP score difference between input and output image-text pairs concerning only the edited part of the input and output text prompts. $\mathrm{CLIP}_{\mathrm{diff - noedit}}$ is the CLIP score difference between input and output using a fixed text prompt where the edited part of the input text is replaced with a generic word, i.e., "object." These metrics enforce that the CLIP text matching scores are preserved between the input and edited shapes, both for edited and unedited parts of the text. + +We report all directional metrics multiplied by 100 for higher precision. We refer readers to our supplementary material for further details about the evaluation metrics. + +# 4.2. Results + +Our method can generate various impressive edited shapes from complex input shapes and prompts. We illustrate a variety of our results in Fig. 5. Our approach flexibly edits various different elements of the 3D objects, for instance replacing a "sword" of a skull warrior with a "viking axe," resulting in coherent, seamless edits in both texture and geometry. Our edits also follow the structure of the input shape when applicable; for instance, when replacing a curvy fork with a worm, the worm maintains the same curved structure as the initial fork. We can also insert new objects, such as a "pillow" or "sunglasses." Our method even enables both replacement and addition at the same time as in "cat with a tail" example, replacing the chicken with a cat and simultaneously placing a tail at the back of the car. + +Given the same prompt, our method can generate different results with different seeds. In Fig. 6, we show generations of "cat" and "dog" samples with various seeds. The resulting shapes vary in their head, eye, and ear structures with different colors and sizes. We can further con + +![](images/b1febbf828f36e56d7bf2d1ffac47820f98213d68676758ce42b26068966e7fe.jpg) +Figure 6. Multiple generations and detailed control through prompt. Our method can generate different results for the same prompt using different seeds. Moreover, Our method can handle detailed prompts that can modify various aspects of the shape such as appearance and mood. For instance, here we can define the type of the cat (e.g. ginger cat) and the mood (e.g., happy). + +trol the various aspects of the generated shape through user prompts. This is also shown in Fig. 6, where we adjust the mood and appearance of the generated shape. + +Comparison with Baselines. Fig. 4 shows a qualitative comparison of our method against several state-of-the-art methods: Tailor3D [32], MVEdit [8] and Vox-E [37]. Our approach shows significant improvements, in both editing quality as well as consistency with the original shape. + +Similar to our method, Vox-E allows controllable editing through merging but at the expense of an expensive SDS-based optimization that can tend towards more global changes than local ones. Since Tailor3D accepts edited front and back views as input, we ran their method using our multi-view editing results. Tab. 1 shows a comparison using GPTEval3D. While our method is consistently preferred, improvements are not as large since this metric does not measure consistency with the input. A method could + +![](images/1d3e6ac84cfa0f993766bdb6f902025fafb4de616ce8915e80e633b29aa2eac5.jpg) +Figure 7. Qualitative ablation of our merging algorithm. We can keep the original parts of the input fixed. Here when we insert a cat, the editing breaks the neighboring regions. Thanks to our merging algorithm, we can recover the original parts of the shape. + +
MethodChamfer Distance ↓
w/o Merging2.95
w/o Average Merging2.29
Ours2.28
+ +Table 5. Quantitative ablation study of our merging algorithm. We calculate the chamfer distance to the input shape for each ablation. Chamfer Distance value is multiplied by $10^{3}$ . Our algorithm is effective in keeping the edited shape consistent with the input. + +globally change the shape and still achieve better results. This is because this metric does not take input shape into account and only considers edited output and the prompt. To complement this metric, we calculated the directional CLIP score and its few variants in Tab. 3; this considers consistency with the input, and demonstrates that our approach achieves significant improvements over the baselines. + +Perceptual Study. We prepare a perceptual study to compare our method with three baselines asking users three questions: "Select the one that follows the following prompt more closely", "Select the one with better visual quality", "Which example better preserves the parts that were not instructed to be edited with the prompt?". There are 360 questions in total, each answered by 10 different participants, totaling 3600 responses. Results are presented in Tab. 2. In all questions, our method is preferred over the baselines. + +Runtime Analysis. Our method enables fast iteration, taking around 24 seconds to obtain initial multi-view editing results. Merging then takes another 50 seconds to produce a final refined shape. Tab. 4 shows a comparison with baselines, using a single RTX 3090 for measurements, except for MVDream, which we run on RTX A6000. Tailor3D [32], concurrent to our work, also operates fast, taking 2 seconds for a forward pass using our multi-view editing results as input (26 seconds in total). MVEdit [8] does not employ any merging, performing editing in around 6 minutes. Since Vox-E [37] involves a long SDS optimization process, its overall inference can take around an hour. + +# 4.3. Ablations + +Tab. 5 and Fig. 7 ablate our merging approach, measuring the chamfer distance between the edited and input shapes. + +Shape Preservation through Merging. Our merging algorithm ensures that only the regions described by the user through a mask and prompt changes. In this ablation study, we only do multi-view editing, and leave out the merging operation. As shown in Fig. 7, without any merging, regions that are not intended by the user can change. In the "chicken in a racing car" example, when the user replaces the chicken with a cat, some part of the car is also altered since the user mask covers that area. In our merging step, we detect the changed region ("cat") and erased region ("chicken") so that we keep the rest of the shape ("car") fixed. + +Average Merging. After edited regions are detected, we merge the voxel grids of the input and edited reconstructions to preserve consistency with the input. In contrast, VoxE [37] uses copy-pasting for merging. That is, they copy the detected part from the edited shape and paste it into the original shape. However, a simple copy-paste approach can create boundary artifacts such as gaps between the edited region and the original shape, as shown in Fig. 7. To fix these boundary problems, we dilate the masks. Within the dilated region, we take the average of the edited shape and the original shape, which provides a smoother transition. + +Limitations. Although our method can generate high-quality editing results, we are limited by the 256x256 resolution of the multi-view diffusion model, MVDream [38]. In addition, our method currently focuses on 3D assets that can be rendered from four inward-facing views. However, this assumption cannot effectively capture large-scale scenes, such as indoor rooms where more views within the scene are needed. + +# 5. Conclusion + +We propose a fast and controllable 3D editing method that can handle a wide variety of 3D shapes and editing prompts. We employ the editing strength of powerful multi-view models, lift edits to 3D, and merge edits in 3D in order to ensure unedited regions remain consistent with the input shape. Hence, our method produces high-quality editing results with fast runtime speeds. We believe this shows a significant potential for high-quality, controllable, seamless, and fast 3D editing. + +Acknowledgments This work is partially done during Ziya's and Can's internships at Snap. Matthias Nießner was supported by the ERC Starting Grant Scan2CAD (804724) and Angela Dai was supported by the ERC Starting Grant SpatialSem (101076253). + +# References + +[1] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. Shapetalk: A language dataset and framework for 3d shape edits and deformations. In CVPR, 2023. 3 +[2] Antonio Alliegro, Yawar Siddiqui, Tatiana Tommasi, and Matthias Nießner. Polydiff: Generating 3d polygonal meshes with diffusion models. arXiv preprint arXiv:2312.11417, 2023. 2 +[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18208-18218, 2022. 3 +[4] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In CVPR, 2024. 3 +[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 3 +[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3 +[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG, 2023. 3 +[8] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 2, 3, 5, 6, 7, 8 +[9] Minghao Chen, Junyu Xie, Iro Laina, and Andrea Vedaldi. Shap-editor: Instruction-guided latent 3d editing in seconds. In CVPR, 2024. 2, 3 +[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 3 +[11] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023. 3 +[12] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. In CVPR, 2024. 3 +[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 3, 6 +[14] Shaocong Dong, Lihe Ding, Zhanpeng Huang, Zibin Wang, Tianfan Xue, and Dan Xu. Interactive3d: Create what you want by interactive 3d generation. In CVPR, 2024. 2, 3 +[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high + +quality dataset of 3d scanned household items. In ICRA, 2022.6 +[16] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG, 2022. 6 +[17] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In ICCV, 2023. 2, 3 +[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In ICLR, 2022. 2, 3, 4 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3 +[20] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In ICLR, 2024. 3 +[21] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In CVPR, 2024. 2, 3, 4 +[22] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2024. 3 +[23] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 2, 3 +[24] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 4 +[25] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In CVPR, 2023. 3 +[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 3 +[27] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3 +[28] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 3 +[29] Bharath Raj Nagoor Kani, Hsin-Ying Lee, Sergey Tulyakov, and Shubham Tulsiani. Upfusion: Novel view diffusion from unposed sparse view observations. In ECCV, 2025. 3 +[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation + +and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.3 +[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3 +[32] Zhangyang Qi, Yunhan Yang, Mengchen Zhang, Long Xing, Xiaoyang Wu, Tong Wu, Dahua Lin, Xihui Liu, Jiaqi Wang, and Hengshuang Zhao. Tailor3d: Customized 3d assets editing and generation with dual-side images. arXiv preprint arXiv:2407.06191, 2024. 2, 3, 5, 6, 7, 8 +[33] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In ICLR, 2024. 3 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6 +[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 4 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3 +[37] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. In ICCV, 2023. 2, 3, 5, 6, 7, 8 +[38] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2024. 2, 4, 8 +[39] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In CVPR, 2024. 2 +[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 3 +[41] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3 +[42] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 3 +[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 3 +[44] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 3 + +[45] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. In CVPR, 2024. 3, 6 +[46] Jiale Xu, Xintao Wang, Yan-Pei Cao, Weihao Cheng, Ying Shan, and Shenghua Gao. Instructp2p: Learning to edit 3d point clouds with text instructions. arXiv preprint arXiv:2306.07154, 2023. 3 +[47] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al. Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. In ICLR, 2024. 3 +[48] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, 2024. 2 +[49] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 3 +[50] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In NeurIPS, 2024. 3 +[51] Peiye Zhuang, Songfang Han, Chaoyang Wang, Aliaksandr Siarohin, Jiaxu Zou, Michael Vasilkovsky, Vladislav Shakhrai, Sergey Korolev, Sergey Tulyakov, and HsinYing Lee. Gtr: Improving large 3d reconstruction models through geometry and texture refinement. arXiv preprint arXiv:2406.05649, 2024. 2, 3, 5, 6 \ No newline at end of file diff --git a/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/images.zip b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e45ded5b3b3e8e9eceafe915727fae7b8a9bcdd1 --- /dev/null +++ b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:987c67fc98de4c42f4cfd111a69ac049cf82246bf290d887f87ad75a209d9777 +size 635263 diff --git a/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/layout.json b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..86988cadc046d9c7e693804deaa1de3b5313b47b --- /dev/null +++ b/2025/PrEditor3D_ Fast and Precise 3D Shape Editing/layout.json @@ -0,0 +1,9619 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 103, + 454, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 103, + 454, + 121 + ], + "spans": [ + { + "bbox": [ + 157, + 103, + 454, + 121 + ], + "type": "text", + "content": "PrEditor3D: Fast and Precise 3D Shape Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 142, + 135, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 142, + 135, + 159 + ], + "spans": [ + { + "bbox": [ + 75, + 142, + 135, + 159 + ], + "type": "text", + "content": "Ziya Erkoc1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 142, + 224, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 142, + 224, + 157 + ], + "spans": [ + { + "bbox": [ + 159, + 142, + 224, + 157 + ], + "type": "text", + "content": "Can Gümeli" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 162, + 229, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 162, + 229, + 176 + ], + "spans": [ + { + "bbox": [ + 156, + 162, + 229, + 176 + ], + "type": "text", + "content": "Peter Wonka2,3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 249, + 142, + 335, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 142, + 335, + 158 + ], + "spans": [ + { + "bbox": [ + 249, + 142, + 335, + 158 + ], + "type": "text", + "content": "Chaoyang Wang²" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 266, + 162, + 344, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 162, + 344, + 177 + ], + "spans": [ + { + "bbox": [ + 266, + 162, + 344, + 177 + ], + "type": "text", + "content": "Hsin-Ying Lee" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 143, + 447, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 143, + 447, + 156 + ], + "spans": [ + { + "bbox": [ + 358, + 143, + 447, + 156 + ], + "type": "text", + "content": "Matthias Nießner1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 381, + 163, + 454, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 163, + 454, + 178 + ], + "spans": [ + { + "bbox": [ + 381, + 163, + 454, + 178 + ], + "type": "text", + "content": "Peiye Zhuang" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 172, + 182, + 438, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 182, + 438, + 198 + ], + "spans": [ + { + "bbox": [ + 172, + 182, + 438, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 172, + 182, + 438, + 198 + ], + "type": "text", + "content": "Technical University of Munich " + }, + { + "bbox": [ + 172, + 182, + 438, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 172, + 182, + 438, + 198 + ], + "type": "text", + "content": "Snap Inc " + }, + { + "bbox": [ + 172, + 182, + 438, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 172, + 182, + 438, + 198 + ], + "type": "text", + "content": "KAUST" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 187, + 212, + 420, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 212, + 420, + 225 + ], + "spans": [ + { + "bbox": [ + 187, + 212, + 420, + 225 + ], + "type": "text", + "content": "https://ziyaerkoc.com/preditor3d" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 76, + 266, + 138, + 334 + ], + "blocks": [ + { + "bbox": [ + 67, + 247, + 155, + 258 + ], + "lines": [ + { + "bbox": [ + 67, + 247, + 155, + 258 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 155, + 258 + ], + "type": "text", + "content": "\"chicken in a racing car\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 266, + 138, + 334 + ], + "lines": [ + { + "bbox": [ + 76, + 266, + 138, + 334 + ], + "spans": [ + { + "bbox": [ + 76, + 266, + 138, + 334 + ], + "type": "image", + "image_path": "9c48c9fdd0cf5631c221117521fc1b9b1344c40a5bc7f89944adb8367fd7393f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 69, + 337, + 145, + 380 + ], + "blocks": [ + { + "bbox": [ + 69, + 337, + 145, + 380 + ], + "lines": [ + { + "bbox": [ + 69, + 337, + 145, + 380 + ], + "spans": [ + { + "bbox": [ + 69, + 337, + 145, + 380 + ], + "type": "image", + "image_path": "23b1c8cd2d7cb0bc1f34e5aa25c4deccd09603d79b4764c929d9fafc8d399d92.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 394, + 110, + 402 + ], + "lines": [ + { + "bbox": [ + 82, + 394, + 110, + 402 + ], + "spans": [ + { + "bbox": [ + 82, + 394, + 110, + 402 + ], + "type": "text", + "content": "cannon" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 149, + 314, + 164, + 342 + ], + "blocks": [ + { + "bbox": [ + 149, + 314, + 164, + 342 + ], + "lines": [ + { + "bbox": [ + 149, + 314, + 164, + 342 + ], + "spans": [ + { + "bbox": [ + 149, + 314, + 164, + 342 + ], + "type": "image", + "image_path": "2c81bd58f932972b1c74cc61ced5ca46c670d94d066d3653ab6b0cc544bd913b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 173, + 267, + 244, + 334 + ], + "blocks": [ + { + "bbox": [ + 175, + 247, + 233, + 258 + ], + "lines": [ + { + "bbox": [ + 175, + 247, + 233, + 258 + ], + "spans": [ + { + "bbox": [ + 175, + 247, + 233, + 258 + ], + "type": "text", + "content": "\"chicken cat ...\"" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 267, + 244, + 334 + ], + "lines": [ + { + "bbox": [ + 173, + 267, + 244, + 334 + ], + "spans": [ + { + "bbox": [ + 173, + 267, + 244, + 334 + ], + "type": "image", + "image_path": "8cc7d880f57287871f29f335a086755527b4ce1063a9a502313a2e31a52423a8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 167, + 337, + 246, + 380 + ], + "blocks": [ + { + "bbox": [ + 167, + 337, + 246, + 380 + ], + "lines": [ + { + "bbox": [ + 167, + 337, + 246, + 380 + ], + "spans": [ + { + "bbox": [ + 167, + 337, + 246, + 380 + ], + "type": "image", + "image_path": "62692890255cfc1898b373239421ee2c28b5017bfc6ec5b0913767a15e98fa34.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 395, + 186, + 402 + ], + "lines": [ + { + "bbox": [ + 159, + 395, + 186, + 402 + ], + "spans": [ + { + "bbox": [ + 159, + 395, + 186, + 402 + ], + "type": "text", + "content": "cannon" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 258, + 267, + 340, + 334 + ], + "blocks": [ + { + "bbox": [ + 271, + 248, + 325, + 258 + ], + "lines": [ + { + "bbox": [ + 271, + 248, + 325, + 258 + ], + "spans": [ + { + "bbox": [ + 271, + 248, + 325, + 258 + ], + "type": "text", + "content": "chicken dog ..." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 267, + 340, + 334 + ], + "lines": [ + { + "bbox": [ + 258, + 267, + 340, + 334 + ], + "spans": [ + { + "bbox": [ + 258, + 267, + 340, + 334 + ], + "type": "image", + "image_path": "0ed015752295ae7fb404eb6679c783c6a3b6d80264dcc3c4a299031e7a8dab01.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 258, + 337, + 340, + 380 + ], + "blocks": [ + { + "bbox": [ + 258, + 337, + 340, + 380 + ], + "lines": [ + { + "bbox": [ + 258, + 337, + 340, + 380 + ], + "spans": [ + { + "bbox": [ + 258, + 337, + 340, + 380 + ], + "type": "image", + "image_path": "a74bf6ce14f352c08e3a4df63f544caf575a117ecb090213c45a7592a14d9fe4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 395, + 249, + 402 + ], + "lines": [ + { + "bbox": [ + 222, + 395, + 249, + 402 + ], + "spans": [ + { + "bbox": [ + 222, + 395, + 249, + 402 + ], + "type": "text", + "content": "cannon" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 353, + 267, + 408, + 334 + ], + "blocks": [ + { + "bbox": [ + 342, + 248, + 411, + 258 + ], + "lines": [ + { + "bbox": [ + 342, + 248, + 411, + 258 + ], + "spans": [ + { + "bbox": [ + 342, + 248, + 411, + 258 + ], + "type": "text", + "content": "\"... raeing-car jeep\"" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 353, + 267, + 408, + 334 + ], + "lines": [ + { + "bbox": [ + 353, + 267, + 408, + 334 + ], + "spans": [ + { + "bbox": [ + 353, + 267, + 408, + 334 + ], + "type": "image", + "image_path": "16d93914e8b02aeb8cce88ee3ff194f52947c3c1761d2a8996a9341f63b48ba9.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 353, + 335, + 421, + 377 + ], + "blocks": [ + { + "bbox": [ + 353, + 335, + 421, + 377 + ], + "lines": [ + { + "bbox": [ + 353, + 335, + 421, + 377 + ], + "spans": [ + { + "bbox": [ + 353, + 335, + 421, + 377 + ], + "type": "image", + "image_path": "2f468347a2e0645996089876c58fcf8febc3b6aea3d6910bf2622c55c710342a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 473, + 143, + 533, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 473, + 143, + 533, + 157 + ], + "spans": [ + { + "bbox": [ + 473, + 143, + 533, + 157 + ], + "type": "text", + "content": "Angela Dai" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 427, + 246, + 547, + 387 + ], + "blocks": [ + { + "bbox": [ + 427, + 246, + 547, + 387 + ], + "lines": [ + { + "bbox": [ + 427, + 246, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 427, + 246, + 547, + 387 + ], + "type": "image", + "image_path": "533aa91a06d59cef9f34e86ec606b1fe42a9400ade9b166c34905d771ce16dcc.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 67, + 419, + 123, + 464 + ], + "blocks": [ + { + "bbox": [ + 67, + 403, + 125, + 411 + ], + "lines": [ + { + "bbox": [ + 67, + 403, + 125, + 411 + ], + "spans": [ + { + "bbox": [ + 67, + 403, + 125, + 411 + ], + "type": "text", + "content": "in a wooden cart\"" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 419, + 123, + 464 + ], + "lines": [ + { + "bbox": [ + 67, + 419, + 123, + 464 + ], + "spans": [ + { + "bbox": [ + 67, + 419, + 123, + 464 + ], + "type": "image", + "image_path": "c93d13830cebbdc3fd20e421c2d0047868e94200efb6bad0a586edd0658c52ca.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 127, + 427, + 141, + 451 + ], + "blocks": [ + { + "bbox": [ + 127, + 427, + 141, + 451 + ], + "lines": [ + { + "bbox": [ + 127, + 427, + 141, + 451 + ], + "spans": [ + { + "bbox": [ + 127, + 427, + 141, + 451 + ], + "type": "image", + "image_path": "fd4fa33ecb83d425c3c2b706ba40c06eaf88ee091aa0d2522c8e93e8cf235f1c.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 478, + 112, + 486 + ], + "lines": [ + { + "bbox": [ + 72, + 478, + 112, + 486 + ], + "spans": [ + { + "bbox": [ + 72, + 478, + 112, + 486 + ], + "type": "text", + "content": "viking with a" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 144, + 419, + 200, + 465 + ], + "blocks": [ + { + "bbox": [ + 146, + 403, + 201, + 411 + ], + "lines": [ + { + "bbox": [ + 146, + 403, + 201, + 411 + ], + "spans": [ + { + "bbox": [ + 146, + 403, + 201, + 411 + ], + "type": "text", + "content": "flamethrower ...\"" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 144, + 419, + 200, + 465 + ], + "lines": [ + { + "bbox": [ + 144, + 419, + 200, + 465 + ], + "spans": [ + { + "bbox": [ + 144, + 419, + 200, + 465 + ], + "type": "image", + "image_path": "e70736e11c74e377b631b1205aeeb3a93298b6a794b715578fd9fabe176811fc.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 205, + 420, + 261, + 465 + ], + "blocks": [ + { + "bbox": [ + 213, + 403, + 257, + 411 + ], + "lines": [ + { + "bbox": [ + 213, + 403, + 257, + 411 + ], + "spans": [ + { + "bbox": [ + 213, + 403, + 257, + 411 + ], + "type": "text", + "content": "trebuchet ..." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 420, + 261, + 465 + ], + "lines": [ + { + "bbox": [ + 205, + 420, + 261, + 465 + ], + "spans": [ + { + "bbox": [ + 205, + 420, + 261, + 465 + ], + "type": "image", + "image_path": "119220910980ff47d723ef27312b3b92c8980c71bc90980fc36c8a42d7640f0a.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 270, + 414, + 327, + 465 + ], + "blocks": [ + { + "bbox": [ + 269, + 398, + 329, + 407 + ], + "lines": [ + { + "bbox": [ + 269, + 398, + 329, + 407 + ], + "spans": [ + { + "bbox": [ + 269, + 398, + 329, + 407 + ], + "type": "text", + "content": "\"house in a forest\"" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 270, + 414, + 327, + 465 + ], + "lines": [ + { + "bbox": [ + 270, + 414, + 327, + 465 + ], + "spans": [ + { + "bbox": [ + 270, + 414, + 327, + 465 + ], + "type": "image", + "image_path": "ff9645074f11352b99d8ea555fd6396ae70b8abe9949202ea70ee6f0da360177.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 477, + 325, + 485 + ], + "lines": [ + { + "bbox": [ + 284, + 477, + 325, + 485 + ], + "spans": [ + { + "bbox": [ + 284, + 477, + 325, + 485 + ], + "type": "text", + "content": "\"chicken in a" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 333, + 426, + 347, + 449 + ], + "blocks": [ + { + "bbox": [ + 333, + 426, + 347, + 449 + ], + "lines": [ + { + "bbox": [ + 333, + 426, + 347, + 449 + ], + "spans": [ + { + "bbox": [ + 333, + 426, + 347, + 449 + ], + "type": "image", + "image_path": "875172e5833e4a5e78739a0eab9283c8d2f51a49479b1727569ecc004d9ead00.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 350, + 413, + 401, + 465 + ], + "blocks": [ + { + "bbox": [ + 349, + 395, + 401, + 411 + ], + "lines": [ + { + "bbox": [ + 349, + 395, + 401, + 411 + ], + "spans": [ + { + "bbox": [ + 349, + 395, + 401, + 411 + ], + "type": "text", + "content": "house treasure cave ..." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 350, + 413, + 401, + 465 + ], + "lines": [ + { + "bbox": [ + 350, + 413, + 401, + 465 + ], + "spans": [ + { + "bbox": [ + 350, + 413, + 401, + 465 + ], + "type": "image", + "image_path": "7ae10615da9277a41397c1aeaf094502e7942c78d403c55322ad5e8bd99226f0.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 476, + 392, + 485 + ], + "lines": [ + { + "bbox": [ + 364, + 476, + 392, + 485 + ], + "spans": [ + { + "bbox": [ + 364, + 476, + 392, + 485 + ], + "type": "text", + "content": "\"chicken" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 419, + 416, + 468, + 472 + ], + "blocks": [ + { + "bbox": [ + 413, + 399, + 477, + 409 + ], + "lines": [ + { + "bbox": [ + 413, + 399, + 477, + 409 + ], + "spans": [ + { + "bbox": [ + 413, + 399, + 477, + 409 + ], + "type": "text", + "content": "\"truck carrying clay\"" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 419, + 416, + 468, + 472 + ], + "lines": [ + { + "bbox": [ + 419, + 416, + 468, + 472 + ], + "spans": [ + { + "bbox": [ + 419, + 416, + 468, + 472 + ], + "type": "image", + "image_path": "8096860005b6cb0122e73a16c5a9a1c5a04fafb01844457f358e956358d1390a.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 483, + 464, + 491 + ], + "lines": [ + { + "bbox": [ + 433, + 483, + 464, + 491 + ], + "spans": [ + { + "bbox": [ + 433, + 483, + 464, + 491 + ], + "type": "text", + "content": "\"android\"" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "bbox": [ + 475, + 426, + 490, + 451 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 426, + 490, + 451 + ], + "spans": [ + { + "bbox": [ + 475, + 426, + 490, + 451 + ], + "type": "image", + "image_path": "ce3db1734d521bfc9e3c08688351d95fc74e480eb6cb42914cdcfad76b0a0057.jpg" + } + ] + } + ], + "index": 44, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 475, + 418, + 540, + 471 + ], + "blocks": [ + { + "bbox": [ + 494, + 399, + 542, + 409 + ], + "lines": [ + { + "bbox": [ + 494, + 399, + 542, + 409 + ], + "spans": [ + { + "bbox": [ + 494, + 399, + 542, + 409 + ], + "type": "text", + "content": "... elay pizza" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 475, + 418, + 540, + 471 + ], + "lines": [ + { + "bbox": [ + 475, + 418, + 540, + 471 + ], + "spans": [ + { + "bbox": [ + 475, + 418, + 540, + 471 + ], + "type": "image", + "image_path": "bf2f124969d224eeb3a8ed43c086bec87e9a83d8b021634703c3deb304d76a1e.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 479, + 545, + 487 + ], + "lines": [ + { + "bbox": [ + 490, + 479, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 490, + 479, + 545, + 487 + ], + "type": "text", + "content": "\"android chicken" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 151, + 497, + 193, + 553 + ], + "blocks": [ + { + "bbox": [ + 61, + 487, + 121, + 495 + ], + "lines": [ + { + "bbox": [ + 61, + 487, + 121, + 495 + ], + "spans": [ + { + "bbox": [ + 61, + 487, + 121, + 495 + ], + "type": "text", + "content": "mustache, helmet" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 482, + 205, + 491 + ], + "lines": [ + { + "bbox": [ + 138, + 482, + 205, + 491 + ], + "spans": [ + { + "bbox": [ + 138, + 482, + 205, + 491 + ], + "type": "text", + "content": "... helmet party hat" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 151, + 497, + 193, + 553 + ], + "lines": [ + { + "bbox": [ + 151, + 497, + 193, + 553 + ], + "spans": [ + { + "bbox": [ + 151, + 497, + 193, + 553 + ], + "type": "image", + "image_path": "0421c88025359e3c64200ad95c50bf5c163fd5d4a231a5f4eee04bc92a7f5738.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 223, + 498, + 268, + 554 + ], + "blocks": [ + { + "bbox": [ + 221, + 479, + 269, + 496 + ], + "lines": [ + { + "bbox": [ + 221, + 479, + 269, + 496 + ], + "spans": [ + { + "bbox": [ + 221, + 479, + 269, + 496 + ], + "type": "text", + "content": "...red pepper mustache ..." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 498, + 268, + 554 + ], + "lines": [ + { + "bbox": [ + 223, + 498, + 268, + 554 + ], + "spans": [ + { + "bbox": [ + 223, + 498, + 268, + 554 + ], + "type": "image", + "image_path": "c32bcbd3cff098474abb37d3b96be16ae7a075be91fdbe18cc8942e510e83875.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 283, + 498, + 326, + 555 + ], + "blocks": [ + { + "bbox": [ + 286, + 487, + 323, + 497 + ], + "lines": [ + { + "bbox": [ + 286, + 487, + 323, + 497 + ], + "spans": [ + { + "bbox": [ + 286, + 487, + 323, + 497 + ], + "type": "text", + "content": "racing car\"" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 283, + 498, + 326, + 555 + ], + "lines": [ + { + "bbox": [ + 283, + 498, + 326, + 555 + ], + "spans": [ + { + "bbox": [ + 283, + 498, + 326, + 555 + ], + "type": "image", + "image_path": "406f732f82a18a122c29eb4202639c5106e88fb59cdbe8529e2921515f6efec5.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "bbox": [ + 283, + 508, + 326, + 555 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 508, + 326, + 555 + ], + "spans": [ + { + "bbox": [ + 283, + 508, + 326, + 555 + ], + "type": "image", + "image_path": "e78138e3a04a4b109655ba4f8e8b39d355f4ca32ad28bfa2ea81c8959af6f05f.jpg" + } + ] + } + ], + "index": 58, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 354, + 501, + 400, + 557 + ], + "blocks": [ + { + "bbox": [ + 350, + 487, + 405, + 497 + ], + "lines": [ + { + "bbox": [ + 350, + 487, + 405, + 497 + ], + "spans": [ + { + "bbox": [ + 350, + 487, + 405, + 497 + ], + "type": "text", + "content": "cat with a tail ..." + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 501, + 400, + 557 + ], + "lines": [ + { + "bbox": [ + 354, + 501, + 400, + 557 + ], + "spans": [ + { + "bbox": [ + 354, + 501, + 400, + 557 + ], + "type": "image", + "image_path": "073e30d0b0f8862b586d53260b056453641089686f0676451908ab775ad0519b.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 558, + 555, + 593 + ], + "lines": [ + { + "bbox": [ + 55, + 558, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 555, + 593 + ], + "type": "text", + "content": "Figure 1. PrEditor3D is a (top) fast and high-quality editing method that can perform precise and consistent editing only in the intended regions, keeping the rest identical. (mid) It can handle diverse editing prompts with any given 3D object. (bottom) Furthermore, it can support iterative editing, facilitating artistic workflow, and can also support editing multiple regions in a single run." + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 425, + 502, + 472, + 552 + ], + "blocks": [ + { + "bbox": [ + 425, + 502, + 472, + 552 + ], + "lines": [ + { + "bbox": [ + 425, + 502, + 472, + 552 + ], + "spans": [ + { + "bbox": [ + 425, + 502, + 472, + 552 + ], + "type": "image", + "image_path": "69e72e7cebb6abbe44a9aaa6cb9ed6c067af4dbca1aed5e99fae4134d47cba3f.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 492, + 501, + 540, + 554 + ], + "blocks": [ + { + "bbox": [ + 495, + 488, + 541, + 499 + ], + "lines": [ + { + "bbox": [ + 495, + 488, + 541, + 499 + ], + "spans": [ + { + "bbox": [ + 495, + 488, + 541, + 499 + ], + "type": "text", + "content": "wearing a tie\"" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 492, + 501, + 540, + 554 + ], + "lines": [ + { + "bbox": [ + 492, + 501, + 540, + 554 + ], + "spans": [ + { + "bbox": [ + 492, + 501, + 540, + 554 + ], + "type": "image", + "image_path": "eeb47fe2344c3616e6b6291f1645cee84b9d9f33d3d38968afc03dc693e117d3.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + } + ], + "index": 64 + }, + { + "bbox": [ + 152, + 602, + 200, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 602, + 200, + 614 + ], + "spans": [ + { + "bbox": [ + 152, + 602, + 200, + 614 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "We propose a training-free approach to 3D editing that enables the editing of a single shape within a few minutes. The edited 3D mesh aligns well with the prompts, and remains identical for regions that are not intended to be altered. To this end, we first project the 3D object onto 4-view images and perform synchronized multi-view image editing along with user-guided text prompts and user-provided" + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 313, + 603, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 555, + 712 + ], + "type": "text", + "content": "rough masks. However, the targeted regions to be edited are ambiguous due to projection from 3D to 2D. To ensure precise editing only in intended regions, we develop a 3D segmentation pipeline that detects edited areas in 3D space, followed by a merging algorithm to seamlessly integrate edited 3D regions with the original input. Extensive experiments demonstrate the superiority of our method over previous approaches, enabling fast, high-quality editing while preserving unintended regions." + } + ] + } + ], + "index": 68 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "640" + } + ] + } + ], + "index": 69 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 223 + ], + "type": "text", + "content": "Recent 3D diffusion models can generate high-quality assets that closely align with the text prompts in the form of neural fields [23, 38], meshes [2, 39], or Gaussian point clouds [48]. Although these methods generate impressive results, they lack the essential capability for precise and controllable editing of the generated outputs, a critical requirement for iterative artistic workflows. Effective 3D editing demands: (1) it should be fast enough to provide quick feedback, ideally comparable to fast 3D generation algorithms, and (2) it must allow for precise local control, enabling users to keep specific parts of the model unchanged." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 223, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 319 + ], + "type": "text", + "content": "Enabling precise and controllable editing is still an open challenge. Several initial approaches have been proposed to tackle the challenge of 3D editing [8, 9, 14, 17, 32, 37], providing promising results but suffering from slow runtime, lack of precise control, and/or lack of 3D consistency and quality. Optimization-based techniques like SDS [9, 14, 37] or multiview training dataset updates [17] are computationally expensive, making interactive editing out of reach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "type": "text", + "content": "Additionally, they offer limited control over specific parts of the shape, as text prompts alone cannot precisely localize regions to be edited [8, 9, 17, 32, 37]. While Vox-E [37] and Shap-Editor [9] propose a mechanism to prevent original parts of the shape from being altered during editing, they do not enable precise editing due to having only text as input. Finally, one can observe various visual quality problems, such as the Janus problem, blurring, over-saturation, and overemphasizing texture changes while leaving the geometry intact or degrading." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "type": "text", + "content": "To address these challenges, we propose a novel editing pipeline for 3D assets that is faster, more precise, and delivers high-quality results (See Fig. 1). As our primary goal is faster editing, we propose an editing framework leveraging a pipeline that consists of two components: a multiview diffusion algorithm and a feed-forward mesh reconstruction. Multi-view diffusion models can leverage superior 2D editing techniques, and the feed-forward mesh reconstruction bridges the gap between 2D and 3D. For better controllability, we extend multi-view image generation to multi-view image editing using 2D masks to constrain the edits to user-specified regions. The 2D masks can take various forms, including manually selected regions, handbrushed areas, or automatically generated segmentations. Automatically generated masks can constrain the edit to a limited region and make the editing more restrictive, we further elaborate on this discussion in the supplementary material. We adopt DDPM inversion [21] to extract initial noise vectors from input multi-view images and execute Prompt-to-Prompt [18] on a multi-view diffusion model [38]. We use 2D user-provided masks to blend edited and original views during the denoising. Due to the occlusion issues that may naturally occur in 2D renderings, we cannot ob" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 74, + 533, + 156 + ], + "blocks": [ + { + "bbox": [ + 317, + 74, + 533, + 156 + ], + "lines": [ + { + "bbox": [ + 317, + 74, + 533, + 156 + ], + "spans": [ + { + "bbox": [ + 317, + 74, + 533, + 156 + ], + "type": "image", + "image_path": "ed053e6f3f5efb47dd28d94b8cee1b43fd32c5310ab876809527fce6cfbba889.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 158, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 224 + ], + "type": "text", + "content": "Figure 2. Ambiguous intended regions. The intended region to be edited is clear in 3D (e.g. the cat tail). However, after projecting to 2D, regardless of the granularity of the user-provided masks, the editing will either alter some unintended regions (e.g. the robot cat) or be too limited for reasonable editing, especially because of occlusion caused by the tail." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 232, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 340 + ], + "type": "text", + "content": "tain ideal intended regions in 2D regardless of the granularity of the masks, as shown in Fig. 2. The masks are either too coarse so the unintended regions will be changed, or too fine-grained to allow reasonable editing. Without additional spatial information in 3D, multi-view editing approaches cannot fully address this challenge. Therefore, simply adopting a feed-forward reconstruction method [51] to convert edited multi-views into a 3D mesh often leads to undesirable results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 341, + 556, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 556, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 556, + 472 + ], + "type": "text", + "content": "To tackle this issue, we propose using the original 3D input and 3D segmentation. We first detect the intended editing region using Grounding DINO [24] and SAM 2 [35] with the user mask and prompts. This gives an initial 2D segmentation that we subsequently lift to 3D. For this purpose, we use color coding in a multi-view to 3D reconstruction pipeline, named GTR [51], to end up with a 3D segmentation that we can use during merging. Specifically, we paint 2D segmentations in a specific color, (e.g., green) and after reconstruction, we can detect which regions are edited by querying the color in the 3D field." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 473, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 568 + ], + "type": "text", + "content": "Then, we perform merging to maintain the original parts of the shape. We use 3D masks to detect edited/replaced parts in GTR's [51] voxel-feature space. We extract the edited part from the new shape and replace it with the old part from the original shape. That way, we guarantee that the remaining shape will remain identical. We apply a final average blending operation so that new parts and the original shape blend smoothly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 570, + 531, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 570, + 531, + 581 + ], + "spans": [ + { + "bbox": [ + 325, + 570, + 531, + 581 + ], + "type": "text", + "content": "In summary, we make the following contributions:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "content": "- We propose a novel method for diffusion-based 3D object editing that is faster than previous work and enables precise, interactive editing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 618, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 653 + ], + "type": "text", + "content": "- The proposed method consists of user-guided and multiview-synchronized editing and a feed-forward 3D reconstruction, enabling fast editing in a feed-forward manner." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "content": "- To enable precise editing only to intended regions, we propose a voxel-based 3D segmentation method that utilizes multi-view segmentation information and propagates it to 3D, followed by an average blending operation to merge the edited and original objects." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "641" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 73, + 553, + 225 + ], + "blocks": [ + { + "bbox": [ + 63, + 73, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 63, + 73, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 63, + 73, + 553, + 225 + ], + "type": "image", + "image_path": "d674394a442d9900b55227a5829b8f095cb9e0a3c2630d948156ec9be7838c96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 555, + 285 + ], + "type": "text", + "content": "Figure 3. Overview of PrEditor3D. Given an input 3D object, we first render its multi-view images from 4 orthogonal views. We then obtain editing input from the user, describing in text as well as rough 2D masks the desired edits. We perform synchronized multi-view editing based on the text prompts as well as the user-provided masks (Sec.3.1. Due to the rough masks and the unclear intended regions caused by ambiguous 3D-2D projection, we detect the intended regions with Grounding Dino and SAM 2 (Sec.3.2, where the segmentation results are lifted to 3D for the final merging operation (Sec.3.3)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 294, + 335 + ], + "type": "text", + "content": "- Our editing method has superior quality to previous work. We show significant improvements in GPTEval3D [45], directional CLIP metrics, and extensive user studies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 349, + 147, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 147, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 147, + 361 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 373, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 539 + ], + "type": "text", + "content": "2D Editing applies global or local modifications to an image based on user instructions. It has gained significant attention for enabling a more interactive user experience in content creation. To achieve this, existing methods either fine-tune text-to-image models with specialized instructional editing datasets [46, 50], or use training-free approaches with inpainting [3, 30] or cross-attention mechanisms [6, 7, 18]. To edit user-provided images, various diffusion inversion techniques have been developed. For example, SDEdit [26] introduces noise to images to capture intermediate steps in the diffusion process, null-text inversion [28] inverts the deterministic DDIM inversion [40], and recent works [4, 21] use DDPM inversion [19] to enhance editing capabilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "text", + "content": "In this work, one step to achieve 3D editing is to perform 2D multi-view image editing. We apply the DDPM inversion [19] and Prompt2Prompt [18] to operate within a diffusion-based sparse multi-view generation model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "3D Reconstruction from Sparse Multi-Views refers to the task of reconstructing a 3D instance from a limited number of multi-view images. To achieve this, Score Distillation Sampling (SDS) [31] and its variants [10, 23, 33, 42, 43] optimize a 3D scene representation by reconstructing the given sparse views and generating novel views through gradients from large-scale pre-trained text-to-image diffusion models [36]. However, these approaches are often computationally demanding. Beyond per-scene optimization, PixelNeRF [49] generates a NeRF representation [27] of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 300, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 384 + ], + "type": "text", + "content": "a scene from sparse-view images by training the model across multiple scenes. With recent advancements in largescale 3D datasets such as Objaverse [13], follow-up works [20, 22, 25, 29, 41, 44, 47, 51] have improved this feedforward 3D reconstruction approach for object-centric 3D assets. In this work, we utilize one recent method, GTR [51], to quickly generate 3D meshes from 4 views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 385, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 626 + ], + "type": "text", + "content": "3D Editing presents additional challenges compared to 2D editing due to the need to maintain spatial consistency in 3D. To address this, some methods train 3D editing models using paired 3D datasets [1, 46], however, these approaches are limited by a lack of diverse and complex datasets. Other methods adapt 2D editing models, such as InstructPix2Pix [5], for the 3D domain [17, 32], by iteratively updating multi-view images of a scene. However, without synchronized multi-view updates, this approach often results in flickering or inconsistent views. Alternatively, some methods propose a generation-reconstruction loop to modify 3D representations using intermediate denoised images [8] or SDS gradients [9, 11, 12, 14, 37] in the diffusion process. While these methods can achieve 3D consistency, they often struggle with quality or suffer from high computational costs. In our work, we perform multi-view editing and reconstruct the edited object in 3D. Beyond that, to preserve unchanged regions, we carefully design an approach by detecting the edited 3D regions and integrating the intended 3D edited regions into the original shapes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 633, + 370, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 370, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 370, + 645 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "We aim to achieve fast 3D asset editing in a training-free manner, allowing for precise and user-guided edits. Our approach achieves this through multi-view image editing in 2D, followed by lifting the 2D edits into 3D. This process can be summarized into 3 main steps: (1) synchronized" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "642" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "sparse multi-view editing in 2D (Sec.3.1), (2) detecting intended editing regions across 2D views through the Grounding Dino [24] and SAM 2 [35] approach (Sec.3.2), and (3) lifting the intended editing regions to 3D and merging the edited shape into the original (Sec.3.3). Our approach is illustrated in Fig. 3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 151, + 270, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 270, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 270, + 164 + ], + "type": "text", + "content": "3.1. Synchronized Sparse Multi-View Editing" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 168, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 252 + ], + "type": "text", + "content": "To edit a given 3D object " + }, + { + "bbox": [ + 55, + 168, + 296, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 168, + 296, + 252 + ], + "type": "text", + "content": ", we leverage the power of 2D editing through multi-view diffusion models. We first perform synchronized sparse multi-view edits using a pretrained multi-view diffusion model. In practice, we use MVDream [38], which generates 4 orthogonal views based on a text prompt. Note that our approach remains agnostic to the specific diffusion-based multi-view generation model used." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "text", + "content": "We first render multi-view images from " + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "text", + "content": ", then apply the DDPM diffusion inversion mechanism [21] to revert the multi-view images to their initial noise vectors, denoted as " + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\pmb{x}^T" + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "text", + "content": " is the number of diffusion timesteps during the denoising process. We will use these vectors " + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\pmb{x}^T" + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "text", + "content": " as the initial latent vectors for the editing process in diffusion. We denote the text prompt for the input shape and the edited shape as " + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "inline_equation", + "content": "y_e" + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "text", + "content": ", respectively. Also, to enable better editing control and interaction, we further take user-provided masks in 4 views as input. These masks indicate target regions for the edit, denoted as " + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{user}} \\in \\mathbb{R}^{4 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 253, + 296, + 407 + ], + "type": "text", + "content": " are the height and weight of the images. Note that our method does not require precise and accurate masks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "text", + "content": "To edit multi-view images, we apply the Prompt-to-Prompt [18] approach on the multi-view diffusion model. For simplicity, we present the basic operation at each diffusion timestep and within each attention block. To be specific, Prompt-to-Prompt [18] generates an edited latent vector " + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_e'" + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "text", + "content": " by replacing the self- and cross-attention weights of the original latent vector, " + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i" + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "text", + "content": ", with the edited prompt " + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "inline_equation", + "content": "y_e" + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "text", + "content": ". To confine modifications to the desired regions, we blend the latent vectors " + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_e'" + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i" + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "text", + "content": " using user-provided masks " + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{user}}" + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "text", + "content": ". We denote the final edited latent vector at each step as " + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_e" + }, + { + "bbox": [ + 55, + 408, + 296, + 529 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 98, + 536, + 295, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 536, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 98, + 536, + 295, + 550 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {e} \\leftarrow M _ {\\text {u s e r}} \\cdot \\boldsymbol {x} _ {e} ^ {\\prime} + \\left(1 - M _ {\\text {u s e r}}\\right) \\cdot \\boldsymbol {x} _ {i}. \\tag {1}", + "image_path": "f8c53b42f7daeaa80e6563a16b8e61d355166e50ae122cf63887c0efdb6fdc90.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "text", + "content": "In practice, we downsample the user-provided masks " + }, + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{user}}" + }, + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "text", + "content": " to match the feature resolution at each model layer, and the edited latent vector, " + }, + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "inline_equation", + "content": "x_{e}" + }, + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "text", + "content": ", serves as the input to the next layer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Using the inverted noise vector " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}^T" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " ensures that the edited results align with the original texture style. However, the masks " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{user}}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " are often imprecise or misaligned with the target regions, which can still affect regions that are not intended to be altered. Furthermore, occlusions along the depth dimension introduce additional challenges in accurately and semantically localizing the intended editing regions in 3D. We illustrate the issue in Fig. 2. To address this, we apply an automated grounding approach that detects the intended editing region in both 2D and 3D." + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 316, + 88, + 553, + 214 + ], + "blocks": [ + { + "bbox": [ + 320, + 75, + 477, + 87 + ], + "lines": [ + { + "bbox": [ + 320, + 75, + 477, + 87 + ], + "spans": [ + { + "bbox": [ + 320, + 75, + 477, + 87 + ], + "type": "text", + "content": "Algorithm 1: Merging Voxel Features" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "lines": [ + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "V_{e} \\in \\mathbb{R}^{A \\times A \\times F}, M_{i}" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "M_{e} \\in \\mathbb{R}^{A \\times A \\times A}, d \\in \\mathbb{N}" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "\\theta \\in [0,1]" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": " \nOutput: " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "V_{\\text{blend}} \\in \\mathbb{R}^{A \\times A \\times A \\times F}" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": " \n1 " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "V_{i}[M_{i}] \\gets \\emptyset" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": " \n2 " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "V_{i}[M_{e}] \\gets V_{e}[M_{e}]" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": " \n3 " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "N \\gets \\text{Dilation}(M_{e}, d)" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": " \n4 " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "K \\gets M_{e} \\oplus N // \\oplus" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": " means XOR \n5 " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "V_{\\text{blend}} \\gets V_{i}" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": " \n6 " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "V_{\\text{blend}}[K] \\gets \\theta \\odot V_{i}[K] + (1 - \\theta) \\odot V_{e}[K]" + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "text", + "content": " \n7 return " + }, + { + "bbox": [ + 316, + 88, + 553, + 214 + ], + "type": "inline_equation", + "content": "V_{\\text{blend}}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 236, + 543, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 543, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 543, + 250 + ], + "type": "text", + "content": "3.2. Detection of Intended Editing Regions in 2D" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 254, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 338 + ], + "type": "text", + "content": "The intended editing regions refer to the specific semantic areas that correspond to the editing prompt. For example, in Fig. 1, changing \"chicken\" to \"cat\" implies that the desired editing region pertains only to the areas representing the \"chicken\" and \"cat\" concepts. Regions (both 2D and 3D) that are not semantically related to these concepts should remain unchanged." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 338, + 556, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 556, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 556, + 459 + ], + "type": "text", + "content": "To ensure that only intended editing regions are edited while allowing rough user-provided masks, we propose to detect the intended editing regions by applying Grounding Dino [24] and SAM 2 [35] to both the original and the edited multi-view images. To begin with, we identify the changing concept by comparing the original prompt " + }, + { + "bbox": [ + 313, + 338, + 556, + 459 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 338, + 556, + 459 + ], + "type": "text", + "content": " and the editing prompt " + }, + { + "bbox": [ + 313, + 338, + 556, + 459 + ], + "type": "inline_equation", + "content": "y_{e}" + }, + { + "bbox": [ + 313, + 338, + 556, + 459 + ], + "type": "text", + "content": ". We then localize the changing concept within the user-provided mask regions " + }, + { + "bbox": [ + 313, + 338, + 556, + 459 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{user}}" + }, + { + "bbox": [ + 313, + 338, + 556, + 459 + ], + "type": "text", + "content": " and obtain corresponding bounding boxes in multiple views. Formally, we write the procedure as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 468, + 553, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 468, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 354, + 468, + 553, + 480 + ], + "type": "interline_equation", + "content": "\\operatorname {b b o x} _ {x} \\leftarrow \\operatorname {G r o u n d i n g} \\left(x, M _ {\\text {u s e r}}, y _ {i}, y _ {e}\\right), \\tag {2}", + "image_path": "8e65a49b22ef7940845be0b86392054e53360aabf6dee37f8a1c0c0a8633cd3c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 490, + 553, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 553, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 490, + 553, + 514 + ], + "type": "inline_equation", + "content": "\\mathsf{bbox}_x" + }, + { + "bbox": [ + 313, + 490, + 553, + 514 + ], + "type": "text", + "content": " are the bounding boxes for the changing concept in the multi-view images " + }, + { + "bbox": [ + 313, + 490, + 553, + 514 + ], + "type": "inline_equation", + "content": "x\\in \\{\\pmb {x}_i,\\pmb {x}_e\\}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 514, + 555, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 585 + ], + "type": "text", + "content": "Afterward, we segment and track the changing elements across views using the grounding bounding boxes " + }, + { + "bbox": [ + 313, + 514, + 555, + 585 + ], + "type": "inline_equation", + "content": "\\mathsf{bbox}_x" + }, + { + "bbox": [ + 313, + 514, + 555, + 585 + ], + "type": "text", + "content": ". Formally, this procedure can be described as " + }, + { + "bbox": [ + 313, + 514, + 555, + 585 + ], + "type": "inline_equation", + "content": "\\mathrm{SAM}(\\pmb{x},\\mathsf{bbox}_x)" + }, + { + "bbox": [ + 313, + 514, + 555, + 585 + ], + "type": "text", + "content": ". This process yields the intended editing regions in segmentation format for both the original and edited multi-views." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 593, + 488, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 488, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 488, + 606 + ], + "type": "text", + "content": "3.3. Lifting and Merging Edits in 3D" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 610, + 553, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 553, + 635 + ], + "type": "text", + "content": "Finally, we lift the 2D edits to 3D and merge the 3D-edited regions into the original shape." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 637, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 555, + 686 + ], + "type": "text", + "content": "3D Segmentation by Lifting 2D Segmentations. We mark the intended editing regions on multi-view images using a green color. The color-coded multi-view images are then reconstructed in 3D using an offline 3D reconstruction" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "1We use " + }, + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": " to represent both latent vectors and images, without distinguishing between the two." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "643" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 74, + 539, + 454 + ], + "blocks": [ + { + "bbox": [ + 56, + 74, + 539, + 454 + ], + "lines": [ + { + "bbox": [ + 56, + 74, + 539, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 74, + 539, + 454 + ], + "type": "image", + "image_path": "fc2fa62a782513e34a7ad643267821a407c09f1c29730176f900bc37f2e9e090.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 455, + 521, + 467 + ], + "lines": [ + { + "bbox": [ + 87, + 455, + 521, + 467 + ], + "spans": [ + { + "bbox": [ + 87, + 455, + 521, + 467 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison. Our method can perform diverse editing samples and only edit the intended regions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 487, + 297, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 297, + 691 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 297, + 691 + ], + "type": "text", + "content": "model [51]. This model takes multiple views as input and represents the shape as a triplane feature. Two separate decoders—one for Signed Distance Function (SDF) and one for color—generate a geometry field and a color field, respectively. Through this process, we create a 3D segmentation field from the color-coded multi-view images. For each 3D position within the 3D space represented by the triplane, we determine whether the position lies within the 3D intended editing regions based on its color value. That is, we apply a decision threshold for the distance between each color value and the preset green color to identify targeted regions. This produces two 3D masks, indicating the intended editing regions in 3D for both the original shape and the edited shape, denoted as " + }, + { + "bbox": [ + 54, + 487, + 297, + 691 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 54, + 487, + 297, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 487, + 297, + 691 + ], + "type": "inline_equation", + "content": "M_{e}" + }, + { + "bbox": [ + 54, + 487, + 297, + 691 + ], + "type": "text", + "content": ", respectively. Our selection of green color is arbitrary. We can flexibly change the color, if the shape already has the specific green color we use." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Merging Edits in 3D. As aforementioned, coarse user" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 487, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 559 + ], + "type": "text", + "content": "provided masks and occlusion issues in 2D projections often result in unwanted alterations, compromising the preservation of unaffected 3D regions. Therefore, the directly reconstructed shapes from the edited multi-view images using the reconstruction model [51] cannot be the final editing output, as illustrated in Fig. 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "text", + "content": "Using the 3D reconstruction model [51], we extract voxel features for both original and edited shapes by interpolating their triplane features, denoted as " + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i" + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_e" + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "text", + "content": " for the original and edited shapes, respectively. Both " + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i, \\mathbf{V}_e \\in \\mathbb{R}^{A \\times A \\times A \\times F}" + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "text", + "content": " is the voxel resolution, and " + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "text", + "content": " is the feature dimension, " + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "inline_equation", + "content": "A = 256" + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "inline_equation", + "content": "F = 40" + }, + { + "bbox": [ + 313, + 559, + 556, + 632 + ], + "type": "text", + "content": " in practice." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 632, + 555, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 555, + 680 + ], + "type": "text", + "content": "To merge 3D features, we first nullify the original specific regions " + }, + { + "bbox": [ + 313, + 632, + 555, + 680 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 313, + 632, + 555, + 680 + ], + "type": "text", + "content": " from the original voxel feature " + }, + { + "bbox": [ + 313, + 632, + 555, + 680 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 313, + 632, + 555, + 680 + ], + "type": "text", + "content": ", and then replace the target edited regions " + }, + { + "bbox": [ + 313, + 632, + 555, + 680 + ], + "type": "inline_equation", + "content": "M_{e}" + }, + { + "bbox": [ + 313, + 632, + 555, + 680 + ], + "type": "text", + "content": " with edited feature " + }, + { + "bbox": [ + 313, + 632, + 555, + 680 + ], + "type": "inline_equation", + "content": "V_{e}[M_{e}]" + }, + { + "bbox": [ + 313, + 632, + 555, + 680 + ], + "type": "text", + "content": ". We write the above operations as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 353, + 684, + 553, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 684, + 553, + 697 + ], + "spans": [ + { + "bbox": [ + 353, + 684, + 553, + 697 + ], + "type": "interline_equation", + "content": "V _ {i} \\left[ M _ {i} \\right] \\leftarrow \\emptyset , \\text {a n d} V _ {i} \\left[ M _ {e} \\right] \\leftarrow V _ {e} \\left[ M _ {e} \\right]. \\tag {3}", + "image_path": "ed2ec83aaa5299905d083b6d31b44922bb745dbfecf026128a2be474418269c2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "We refer to this approach as a naive copy-paste method." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "644" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 74, + 542, + 346 + ], + "blocks": [ + { + "bbox": [ + 77, + 74, + 542, + 346 + ], + "lines": [ + { + "bbox": [ + 77, + 74, + 542, + 346 + ], + "spans": [ + { + "bbox": [ + 77, + 74, + 542, + 346 + ], + "type": "image", + "image_path": "86854e91fad33408573a2435b059a587351325bb154a437faa1cb460899ed5ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 348, + 520, + 361 + ], + "lines": [ + { + "bbox": [ + 88, + 348, + 520, + 361 + ], + "spans": [ + { + "bbox": [ + 88, + 348, + 520, + 361 + ], + "type": "text", + "content": "Figure 5. More editing results from PrEditor3D. Our method can perform a wide range of editing on various shapes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 372, + 301, + 415 + ], + "blocks": [ + { + "bbox": [ + 58, + 372, + 301, + 415 + ], + "lines": [ + { + "bbox": [ + 58, + 372, + 301, + 415 + ], + "spans": [ + { + "bbox": [ + 58, + 372, + 301, + 415 + ], + "type": "table", + "html": "
MethodPrompt Align.3D PlausibilityTextureOverall
Tailor3D [32]98%99%99%99%
MVEdit [8]57%55%55%57%
Vox-E [37]53%68%50%55%
", + "image_path": "6584f56e1b45b43ab56a660ef11d1e744a567b51568bf0006b0e4367201f8b96.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "text", + "content": "While theoretically plausible, we observe that this straightforward approach typically introduces discontinuities at the 3D editing boundaries, as shown in Fig. 7. To address this, we propose an averaged merging approach that provides a more robust blend of 3D features. In the improved method, we dilate the 3D mask " + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "inline_equation", + "content": "M_{e}" + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "text", + "content": " by a dilation " + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "text", + "content": " and then use an exclusive or (a.k.a. XOR) operation to select the boundary mask regions for smooth blending. Next, we linearly interpolate the two voxel features " + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "inline_equation", + "content": "V_{e}" + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "text", + "content": " within the boundary regions using a coefficient " + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "text", + "content": ", in practice " + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\theta = 0.5" + }, + { + "bbox": [ + 55, + 487, + 296, + 643 + ], + "type": "text", + "content": ". We illustrate the merging process in Alg. 1. After merging, we generate a textured mesh from the blended voxel feature, using the decoders in the 3D reconstruction model [51]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 129, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 129, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 129, + 683 + ], + "type": "text", + "content": "4.1. Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Our evaluation dataset contains 18 unique shapes and 40 editing prompts. We use shapes from GSO [15] and Obja" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 372, + 560, + 416 + ], + "blocks": [ + { + "bbox": [ + 55, + 417, + 296, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 473 + ], + "type": "text", + "content": "Table 1. Comparison using GPTEval3D [45]. Scores indicate the percentage of our method being selected over baselines. We ask GPT-4V to compare our method against the baselines and choose one based on pre-defined categories. Each number here shows the percentage of our method being chosen." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 372, + 560, + 416 + ], + "lines": [ + { + "bbox": [ + 317, + 372, + 560, + 416 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 560, + 416 + ], + "type": "table", + "html": "
MethodPrompt Align.Visual QualityPreserving Shape
Tailor3D [32]96%97%99%
MVEdit [8]68%68%97%
Vox-E [37]78%88%89%
", + "image_path": "29966c18ba2f5d551d837f1b0e63e598c808fca77ea1ba84599ea77cc878bc66.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 418, + 555, + 463 + ], + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 463 + ], + "type": "text", + "content": "Table 2. User study results comparing our method against baselines. The percentage shows the preference for our method. We show users binary comparisons of our method against the baselines and collect their response on which one they prefer." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 482, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 506 + ], + "type": "text", + "content": "verse [13]. We evaluate our method based on the quality of the editing and consistency with the input shapes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 508, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 579 + ], + "type": "text", + "content": "We use the GPTEval3D [45] metric to evaluate the quality of edited shapes and their alignment with the text prompts. GPT-4V is provided with multi-view renderings of two methods at a time, and instructed to pick one based on text-prompt alignment, 3D plausibility, and texture details. There were 120 total questions, each answered 3 times." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Since the GPTEval3D metric does not consider the input shape, it cannot measure whether the shape remained intact, and whether the edited shape is consistent with the high-level style of the input shape. Therefore, we adopt the directional CLIP [34] score metric, " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " from previous works [16, 37]. " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{dir}}" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " evaluates the average difference between text feature change direction and image feature change direction, where images are multi-view renderings of input and edited shapes. To ensure our evaluation is not affected by a particular implementation of this metric, we introduce three variants. " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{dir - cos}}" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " replaces the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "645" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 69, + 541, + 133 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 541, + 133 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 541, + 133 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 541, + 133 + ], + "type": "table", + "html": "
MethodCLIPdir↑CLIPdir-cos↑CLIPdir-avg↑CLIPdir-avg-cos↑CLIPdiff-edit↓CLIPdiff-noedit↓
Tailor3D [32]1.4163.7101.4174.59312.0505.619
MVEdit [8]0.7822.5780.7833.16410.0144.118
Vox-E [37]1.6226.1781.6218.15310.5283.432
PrEditor3D (Ours)1.7827.6791.78211.5548.8122.636
", + "image_path": "048113bc701dc0c58c381255420ce7ebbd80c5f26fb78a12fdee87605de64b98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 167, + 284, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 136, + 555, + 159 + ], + "lines": [ + { + "bbox": [ + 55, + 136, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 555, + 159 + ], + "type": "text", + "content": "Table 3. Directional CLIP score metrics [37] for evaluating editing fidelity and prompt consistency. Our method outperforms baselines across all directional CLIP metrics. Metrics are scaled by 100 to ease reading and allow for more precision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 167, + 284, + 232 + ], + "lines": [ + { + "bbox": [ + 67, + 167, + 284, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 167, + 284, + 232 + ], + "type": "table", + "html": "
MethodEditingMergingTotal
Tailor3D [32]26 sec-26 sec
MVEdit [8]6 min-6 min
Vox-E[37]60 min15 min75 min
PrEditor3D (Ours)24 sec50 sec74 sec
", + "image_path": "50860f0b52e770b119c8958ac0fc3009931d272ac19d8ac2cb8e31ff5b425522.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 255 + ], + "type": "text", + "content": "Table 4. Runtime comparison. We measure the runtime of our baseline methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 266, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 301 + ], + "type": "text", + "content": "text-image direction vector difference with cosine distance, while " + }, + { + "bbox": [ + 55, + 266, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{dir - avg}}" + }, + { + "bbox": [ + 55, + 266, + 295, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 266, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{dir - avg - cos}}" + }, + { + "bbox": [ + 55, + 266, + 295, + 301 + ], + "type": "text", + "content": " compute the same metrics by averaging image vectors first rather than scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 302, + 295, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 295, + 421 + ], + "type": "text", + "content": "We finally introduce two additional directional metrics. " + }, + { + "bbox": [ + 55, + 302, + 295, + 421 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{diff - edit}}" + }, + { + "bbox": [ + 55, + 302, + 295, + 421 + ], + "type": "text", + "content": " is the CLIP score difference between input and output image-text pairs concerning only the edited part of the input and output text prompts. " + }, + { + "bbox": [ + 55, + 302, + 295, + 421 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_{\\mathrm{diff - noedit}}" + }, + { + "bbox": [ + 55, + 302, + 295, + 421 + ], + "type": "text", + "content": " is the CLIP score difference between input and output using a fixed text prompt where the edited part of the input text is replaced with a generic word, i.e., \"object.\" These metrics enforce that the CLIP text matching scores are preserved between the input and edited shapes, both for edited and unedited parts of the text." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 422, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 458 + ], + "type": "text", + "content": "We report all directional metrics multiplied by 100 for higher precision. We refer readers to our supplementary material for further details about the evaluation metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 468, + 113, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 113, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 113, + 479 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 485, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 485, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 295, + 652 + ], + "type": "text", + "content": "Our method can generate various impressive edited shapes from complex input shapes and prompts. We illustrate a variety of our results in Fig. 5. Our approach flexibly edits various different elements of the 3D objects, for instance replacing a \"sword\" of a skull warrior with a \"viking axe,\" resulting in coherent, seamless edits in both texture and geometry. Our edits also follow the structure of the input shape when applicable; for instance, when replacing a curvy fork with a worm, the worm maintains the same curved structure as the initial fork. We can also insert new objects, such as a \"pillow\" or \"sunglasses.\" Our method even enables both replacement and addition at the same time as in \"cat with a tail\" example, replacing the chicken with a cat and simultaneously placing a tail at the back of the car." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 653, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 712 + ], + "type": "text", + "content": "Given the same prompt, our method can generate different results with different seeds. In Fig. 6, we show generations of \"cat\" and \"dog\" samples with various seeds. The resulting shapes vary in their head, eye, and ear structures with different colors and sizes. We can further con" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 167, + 534, + 422 + ], + "blocks": [ + { + "bbox": [ + 331, + 167, + 534, + 422 + ], + "lines": [ + { + "bbox": [ + 331, + 167, + 534, + 422 + ], + "spans": [ + { + "bbox": [ + 331, + 167, + 534, + 422 + ], + "type": "image", + "image_path": "b1febbf828f36e56d7bf2d1ffac47820f98213d68676758ce42b26068966e7fe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 421, + 555, + 487 + ], + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 487 + ], + "type": "text", + "content": "Figure 6. Multiple generations and detailed control through prompt. Our method can generate different results for the same prompt using different seeds. Moreover, Our method can handle detailed prompts that can modify various aspects of the shape such as appearance and mood. For instance, here we can define the type of the cat (e.g. ginger cat) and the mood (e.g., happy)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 504, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 540 + ], + "type": "text", + "content": "trol the various aspects of the generated shape through user prompts. This is also shown in Fig. 6, where we adjust the mood and appearance of the generated shape." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": "Comparison with Baselines. Fig. 4 shows a qualitative comparison of our method against several state-of-the-art methods: Tailor3D [32], MVEdit [8] and Vox-E [37]. Our approach shows significant improvements, in both editing quality as well as consistency with the original shape." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Similar to our method, Vox-E allows controllable editing through merging but at the expense of an expensive SDS-based optimization that can tend towards more global changes than local ones. Since Tailor3D accepts edited front and back views as input, we ran their method using our multi-view editing results. Tab. 1 shows a comparison using GPTEval3D. While our method is consistently preferred, improvements are not as large since this metric does not measure consistency with the input. A method could" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "646" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 290, + 195 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 290, + 195 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 290, + 195 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 290, + 195 + ], + "type": "image", + "image_path": "1d3e6ac84cfa0f993766bdb6f902025fafb4de616ce8915e80e633b29aa2eac5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 196, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 196, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 295, + 240 + ], + "type": "text", + "content": "Figure 7. Qualitative ablation of our merging algorithm. We can keep the original parts of the input fixed. Here when we insert a cat, the editing breaks the neighboring regions. Thanks to our merging algorithm, we can recover the original parts of the shape." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 78, + 247, + 272, + 308 + ], + "blocks": [ + { + "bbox": [ + 78, + 247, + 272, + 308 + ], + "lines": [ + { + "bbox": [ + 78, + 247, + 272, + 308 + ], + "spans": [ + { + "bbox": [ + 78, + 247, + 272, + 308 + ], + "type": "table", + "html": "
MethodChamfer Distance ↓
w/o Merging2.95
w/o Average Merging2.29
Ours2.28
", + "image_path": "e7a1f1310ab2516d2b82a86b37c27cc213a3d34d5d4bca17b0a4d33fdc5dade6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 311, + 295, + 356 + ], + "lines": [ + { + "bbox": [ + 55, + 311, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 295, + 356 + ], + "type": "text", + "content": "Table 5. Quantitative ablation study of our merging algorithm. We calculate the chamfer distance to the input shape for each ablation. Chamfer Distance value is multiplied by " + }, + { + "bbox": [ + 55, + 311, + 295, + 356 + ], + "type": "inline_equation", + "content": "10^{3}" + }, + { + "bbox": [ + 55, + 311, + 295, + 356 + ], + "type": "text", + "content": ". Our algorithm is effective in keeping the edited shape consistent with the input." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": "globally change the shape and still achieve better results. This is because this metric does not take input shape into account and only considers edited output and the prompt. To complement this metric, we calculated the directional CLIP score and its few variants in Tab. 3; this considers consistency with the input, and demonstrates that our approach achieves significant improvements over the baselines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 457, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 296, + 567 + ], + "type": "text", + "content": "Perceptual Study. We prepare a perceptual study to compare our method with three baselines asking users three questions: \"Select the one that follows the following prompt more closely\", \"Select the one with better visual quality\", \"Which example better preserves the parts that were not instructed to be edited with the prompt?\". There are 360 questions in total, each answered by 10 different participants, totaling 3600 responses. Results are presented in Tab. 2. In all questions, our method is preferred over the baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Runtime Analysis. Our method enables fast iteration, taking around 24 seconds to obtain initial multi-view editing results. Merging then takes another 50 seconds to produce a final refined shape. Tab. 4 shows a comparison with baselines, using a single RTX 3090 for measurements, except for MVDream, which we run on RTX A6000. Tailor3D [32], concurrent to our work, also operates fast, taking 2 seconds for a forward pass using our multi-view editing results as input (26 seconds in total). MVEdit [8] does not employ any merging, performing editing in around 6 minutes. Since Vox-E [37] involves a long SDS optimization process, its overall inference can take around an hour." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 382, + 83 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 555, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 113 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 113 + ], + "type": "text", + "content": "Tab. 5 and Fig. 7 ablate our merging approach, measuring the chamfer distance between the edited and input shapes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 116, + 555, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 116, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 555, + 248 + ], + "type": "text", + "content": "Shape Preservation through Merging. Our merging algorithm ensures that only the regions described by the user through a mask and prompt changes. In this ablation study, we only do multi-view editing, and leave out the merging operation. As shown in Fig. 7, without any merging, regions that are not intended by the user can change. In the \"chicken in a racing car\" example, when the user replaces the chicken with a cat, some part of the car is also altered since the user mask covers that area. In our merging step, we detect the changed region (\"cat\") and erased region (\"chicken\") so that we keep the rest of the shape (\"car\") fixed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 251, + 556, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 556, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 556, + 382 + ], + "type": "text", + "content": "Average Merging. After edited regions are detected, we merge the voxel grids of the input and edited reconstructions to preserve consistency with the input. In contrast, VoxE [37] uses copy-pasting for merging. That is, they copy the detected part from the edited shape and paste it into the original shape. However, a simple copy-paste approach can create boundary artifacts such as gaps between the edited region and the original shape, as shown in Fig. 7. To fix these boundary problems, we dilate the masks. Within the dilated region, we take the average of the edited shape and the original shape, which provides a smoother transition." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 385, + 556, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 556, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 556, + 482 + ], + "type": "text", + "content": "Limitations. Although our method can generate high-quality editing results, we are limited by the 256x256 resolution of the multi-view diffusion model, MVDream [38]. In addition, our method currently focuses on 3D assets that can be rendered from four inward-facing views. However, this assumption cannot effectively capture large-scale scenes, such as indoor rooms where more views within the scene are needed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 491, + 388, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 388, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 388, + 504 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 511, + 556, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 556, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 556, + 620 + ], + "type": "text", + "content": "We propose a fast and controllable 3D editing method that can handle a wide variety of 3D shapes and editing prompts. We employ the editing strength of powerful multi-view models, lift edits to 3D, and merge edits in 3D in order to ensure unedited regions remain consistent with the input shape. Hence, our method produces high-quality editing results with fast runtime speeds. We believe this shows a significant potential for high-quality, controllable, seamless, and fast 3D editing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 643, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 556, + 704 + ], + "type": "text", + "content": "Acknowledgments This work is partially done during Ziya's and Can's internships at Snap. Matthias Nießner was supported by the ERC Starting Grant Scan2CAD (804724) and Angela Dai was supported by the ERC Starting Grant SpatialSem (101076253)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "text", + "content": "647" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 134 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. Shapetalk: A language dataset and framework for 3d shape edits and deformations. In CVPR, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 294, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 294, + 178 + ], + "type": "text", + "content": "[2] Antonio Alliegro, Yawar Siddiqui, Tatiana Tommasi, and Matthias Nießner. Polydiff: Generating 3d polygonal meshes with diffusion models. arXiv preprint arXiv:2312.11417, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 215 + ], + "type": "text", + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18208-18218, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "type": "text", + "content": "[4] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In CVPR, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "type": "text", + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 338 + ], + "type": "text", + "content": "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In ICCV, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "type": "text", + "content": "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM TOG, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 294, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 440 + ], + "type": "text", + "content": "[8] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 442, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 294, + 474 + ], + "type": "text", + "content": "[9] Minghao Chen, Junyu Xie, Iro Laina, and Andrea Vedaldi. Shap-editor: Instruction-guided latent 3d editing in seconds. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 476, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 476, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 57, + 476, + 294, + 509 + ], + "type": "text", + "content": "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 510, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 294, + 564 + ], + "type": "text", + "content": "[11] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "text", + "content": "[12] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 644 + ], + "type": "text", + "content": "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "text", + "content": "[14] Shaocong Dong, Lihe Ding, Zhanpeng Huang, Zibin Wang, Tianfan Xue, and Dan Xu. Interactive3d: Create what you want by interactive 3d generation. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "text", + "content": "quality dataset of 3d scanned household items. In ICRA, 2022.6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[16] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "text", + "content": "[17] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "text", + "content": "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In ICLR, 2022. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 209, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 230 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[20] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In ICLR, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 277, + 553, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 309 + ], + "type": "text", + "content": "[21] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In CVPR, 2024. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 311, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 311, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 311, + 553, + 364 + ], + "type": "text", + "content": "[22] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 366, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 553, + 409 + ], + "type": "text", + "content": "[23] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 411, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 465 + ], + "type": "text", + "content": "[24] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "type": "text", + "content": "[25] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In CVPR, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 523, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 566 + ], + "type": "text", + "content": "[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 568, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 610 + ], + "type": "text", + "content": "[27] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "type": "text", + "content": "[28] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 678 + ], + "type": "text", + "content": "[29] Bharath Raj Nagoor Kani, Hsin-Ying Lee, Sergey Tulyakov, and Shubham Tulsiani. Upfusion: Novel view diffusion from unposed sparse view observations. In ECCV, 2025. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "648" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 128 + ], + "type": "text", + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 296, + 185 + ], + "type": "text", + "content": "[32] Zhangyang Qi, Yunhan Yang, Mengchen Zhang, Long Xing, Xiaoyang Wu, Tong Wu, Dahua Lin, Xihui Liu, Jiaqi Wang, and Hengshuang Zhao. Tailor3d: Customized 3d assets editing and generation with dual-side images. arXiv preprint arXiv:2407.06191, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "text", + "content": "[33] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. In ICLR, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 295, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 296 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 295, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 295, + 352 + ], + "type": "text", + "content": "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 386 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 295, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 420 + ], + "type": "text", + "content": "[37] Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. Vox-e: Text-guided voxel editing of 3d objects. In ICCV, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 453 + ], + "type": "text", + "content": "[38] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2024. 2, 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "type": "text", + "content": "[39] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 501, + 295, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 522 + ], + "type": "text", + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 524, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 295, + 567 + ], + "type": "text", + "content": "[41] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 568, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 611 + ], + "type": "text", + "content": "[42] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "type": "text", + "content": "[43] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[44] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. arXiv preprint arXiv:2403.05034, 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 396 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[45] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v (ision) is a human-aligned evaluator for text-to-3d generation. In CVPR, 2024. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[46] Jiale Xu, Xintao Wang, Yan-Pei Cao, Weihao Cheng, Ying Shan, and Shenghua Gao. Instructp2p: Learning to edit 3d point clouds with text instructions. arXiv preprint arXiv:2306.07154, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[47] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al. Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. In ICLR, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "text", + "content": "[48] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 263, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 295 + ], + "type": "text", + "content": "[49] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "text", + "content": "[50] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 396 + ], + "type": "text", + "content": "[51] Peiye Zhuang, Songfang Han, Chaoyang Wang, Aliaksandr Siarohin, Jiaxu Zou, Michael Vasilkovsky, Vladislav Shakhrai, Sergey Korolev, Sergey Tulyakov, and HsinYing Lee. Gtr: Improving large 3d reconstruction models through geometry and texture refinement. arXiv preprint arXiv:2406.05649, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "649" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/3fec933e-2dde-4be1-8ad1-28710156467f_content_list.json b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/3fec933e-2dde-4be1-8ad1-28710156467f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0fa582f9358dacc19b9e4ce258290ee3629eed62 --- /dev/null +++ b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/3fec933e-2dde-4be1-8ad1-28710156467f_content_list.json @@ -0,0 +1,1266 @@ +[ + { + "type": "text", + "text": "Practical solutions to the relative pose of three calibrated cameras", + "text_level": 1, + "bbox": [ + 166, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Charalambos Tzamos $^{1,*}$ Viktor Kocur $^{2,*}$ Yaqing Ding $^{1}$ Daniel Barath $^{3}$ Zuzana Berger Haladová $^{2}$ Torsten Sattler $^{4}$ Zuzana Kukelova $^{1}$", + "bbox": [ + 99, + 179, + 937, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague $^{2}$ Faculty of Mathematics, Physics and Informatics, Comenius University in Bratislava $^{3}$ ETH Zürich (Zurich), HUN-REN SZTAKI (Budapest)", + "bbox": [ + 122, + 227, + 915, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague", + "bbox": [ + 135, + 281, + 898, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 334, + 326, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We study the challenging problem of estimating the relative pose of three calibrated cameras from four point correspondences. We propose novel efficient solutions to this problem that are based on the simple idea of using four correspondences to estimate an approximate geometry of the first two views. We model this geometry either as an affine or a fully perspective geometry estimated using one additional approximate correspondence. We generate such an approximate correspondence using a very simple and efficient strategy, where the new point is the mean point of three corresponding input points. The new solvers are efficient and easy to implement, since they are based on existing efficient minimal solvers, i.e., the 4-point affine fundamental matrix, the well-known 5-point relative pose solver, and the P3P solver. Extensive experiments on real data show that the proposed solvers, when properly coupled with local optimization, achieve state-of-the-art results, with the novel solver based on approximate mean-point correspondences being more robust and accurate than the affine-based solver.", + "bbox": [ + 88, + 364, + 485, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 672, + 220, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camera geometry estimation is crucial in many computer vision applications, e.g., visual navigation [48], Structure-from-Motion [51], augmented reality [4], self-driving cars [16], and visual localization [47]. Due to noise and outliers in input correspondences, the predominant way for camera geometry estimation is to use a hypothesis-and-test framework, e.g., RANSAC [2, 6, 14, 45]. For RANSAC-like methods, using as few (ideally the minimal number of) correspondences as possible for estimation is important since the number of RANSAC iterations (and thus its run-time) grows exponentially with the number of correspondences required for the model estimation.", + "bbox": [ + 89, + 699, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4ceef2ab7f660eb7eb7af3c4704fe4d64a2870881275d38e82103d2b3d8579e7.jpg", + "image_caption": [ + "Figure 1. Visualization of the four-points-in-three-views (4p3v) problem and our solution based on using four correspondences to efficiently estimate an approximate geometry of the first two views and then register the third view using a P3P solver [40]." + ], + "image_footnote": [], + "bbox": [ + 601, + 332, + 820, + 439 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minimal camera geometry problems often result in complex systems of polynomial equations. Efficient algebraic methods helped solve many previously unsolved problems [3, 25, 26, 29, 53, 54]. Still, they fail to generate efficient and/or numerically stable solutions for some configurations. In this paper, we study one such challenging problem: Estimating the relative pose of three calibrated cameras. This problem has received attention for a long time [1, 19, 31, 34, 44]. However, due to its complexity, it is still not considered fully solved. There are no efficient and practical solutions for most of the minimal configurations of point and/or line correspondences [24]. One such configuration that is particularly interesting is the notoriously difficult configuration of four points in three calibrated views [35, 44], known as the $4\\mathrm{p}3\\mathrm{v}$ problem.", + "bbox": [ + 509, + 522, + 906, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State-of-the-art algebraic and numerical methods are known to fail in generating efficient and numerically stable solutions to the 4p3v problem. The existing methods for solving this problem are only approximate [21, 36]. By solving only for one (or a few) solutions from the 272 solutions of the 4p3v problem [21], and by discretely sampling the space of potential solutions [36], the existing 4p3v methods can often fail, i.e., the returned solution can be, in general, arbitrarily far from the geometrically correct solution. To decrease the failure rate, both methods [21, 36]", + "bbox": [ + 509, + 750, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 114, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21913", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "require a lot of tuning and are challenging to re-implement.1", + "bbox": [ + 89, + 90, + 480, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to complex solutions to the configuration of four points in three views, there is a simple and efficient solution for the configuration where five correspondences are detected in the first two views and three of these points are also visible in the third view. In this case, the $5\\mathrm{pt} + \\mathrm{P3P}$ solver first estimates the relative pose of two cameras from five correspondences using an efficient 5pt solver [35], and then registers the third camera using a P3P solver [40].", + "bbox": [ + 88, + 107, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the idea of first estimating the relative pose of two cameras, we propose two novel approaches for solving the 4p3v problem. Our solutions are based on the simple idea of using four correspondences to estimate an approximate geometry of the first two views and then registering the third camera using a P3P solver [40].", + "bbox": [ + 88, + 227, + 480, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the first approach, we approximate the geometry of the first two views using affine cameras. The approximation using affine cameras was shown to provide good accuracy for relative pose estimation if coupled with geometry refitting and local optimization inside RANSAC [41]. We denote the proposed affine-based 4p3v solver as 4p3v (A).", + "bbox": [ + 88, + 319, + 480, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the second approach, we use a less restrictive approximation. In this case, we estimate the full perspective geometry, i.e., the full 5DoF essential matrix, using four input correspondences and one additional approximate correspondence in two views. We generate the approximate correspondence using the locations of three of the four input correspondences. Under the assumption of a paraperspective projection, i.e., of affine geometry, the mean point of three 3D points is projected to the mean points in both images [57]. Thus, the new correspondence is generated as the correspondence between the mean points of three corresponding points detected in two views. This approximate mean-point correspondence can also be seen as a correspondence under the $1^{st}$ -order approximation of the homography defined by the plane passing through the three corresponding 3D points. We denote the proposed mean point-based 4p3v solver as 4p3v (M).", + "bbox": [ + 91, + 411, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the proposed $4\\mathrm{p}3\\mathrm{v}$ (M) solver returns significantly more accurate poses than the $4\\mathrm{p}3\\mathrm{v}$ (A) solver, both solvers may not be accurate enough if not properly treated inside RANSAC. Thus, in this paper, we propose several ways to improve the accuracy of the proposed approximate solvers. (i) To compensate for noise in the mean point correspondences, we introduce the $4\\mathrm{p}3\\mathrm{v}$ $(\\mathbb{M}\\pm \\delta)$ solver that generates two additional correspondences in the vicinity of the mean point, and in the first step calls the 5pt solver [35] three times. (ii) To improve the approximate geometry estimated between the first two views, we refit this geometry using a non-minimal relative pose solver and inliers obtained from the approximate geometry (ENM). (iii) We use", + "bbox": [ + 89, + 667, + 482, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the fourth correspondence in the third view to filter out geometrically infeasible solutions $(+F)$ and to refine the solutions on four input correspondences in three views using just a few iterations of Levenberg-Marquardt (LM) refinement $(+R)$ . While conceptually simple and efficient, the novel solvers achieve state-of-the-art results on real data. The contributions of the paper are as follows:", + "bbox": [ + 511, + 90, + 903, + 195 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose two groups of novel solutions for the well-known and challenging 4p3v problem: $4\\mathrm{p}3\\mathrm{v}(\\mathrm{A})$ - and $4\\mathrm{p}3\\mathrm{v}(\\mathrm{M})$ -based solvers. These solutions are based on the simple idea of using four correspondences to estimate an approximate geometry in the first two views. Compared to state-of-the-art 4p3v solvers [21, 36], which are non-trivial and difficult to re-implement such that they are numerically stable and fast, our new solutions can be easily implemented using existing efficient implementations of the linear 4pt affine fundamental matrix solver [17], the 5pt solver [35] and the P3P solver [40]. The source code is available at https://github.com/kocurvik/threeview.", + "- We present several ways of improving the accuracy and the speed of the proposed solvers, as well as a strategy for efficiently using these approximate solvers inside a RANSAC-style paradigm. We show that the new solvers achieve state-of-the-art results in terms of accuracy on real data. While both $4\\mathrm{p}3\\mathrm{v}$ (A)- and $4\\mathrm{p}3\\mathrm{v}$ (M)-based solvers achieve comparable results when coupled with the suggested non-minimal geometry refitting and local optimization, our $4\\mathrm{p}3\\mathrm{v}$ (M)-based solvers are more robust to the scene geometry and RANSAC inlier thresholds.", + "- To our knowledge, we are the first to extensively evaluate solutions to the 4p3v problem on a large variety of real-world scenes and within state-of-the-art RANSAC frameworks, and to compare them to the baseline 5pt+P3P solver. We report results on 3 datasets, consisting of 18 different scenes and altogether 90,000 camera triplets." + ], + "bbox": [ + 511, + 196, + 903, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 511, + 648, + 651, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Estimating the relative pose of three cameras from a minimal number of point and line correspondences is known as an extremely challenging problem. For three uncalibrated cameras, 6 point correspondences are necessary to estimate the trifocal tensor, with a solution known for a long time [43, 55]. Solutions to three minimal combinations of points and lines are presented in [38]. The configuration of 9 lines is more challenging and was solved only recently [28]. Yet, the solver is far from practical and runs 17.8s.", + "bbox": [ + 511, + 674, + 903, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For calibrated cameras, the configuration that attracts most of the attention is the configuration of four points in three views (the 4p3v problem). Note that this is not a minimal configuration since it generates 12 constraints for 11 degrees-of-freedom (DoF). The 4p3v problem is known to be extremely difficult to solve. Several papers present", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "There is no publicly available implementation for [36]. Implementation of [21] is quite complex and requires a non-negligible effort to run.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21914", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mostly theoretical results [1, 31, 34]. For four triplets of exact points without noise, it is shown that the 4p3v problem has, in general, a unique solution [19, 44].", + "bbox": [ + 89, + 90, + 480, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To the best of our knowledge, there are only two reasonably efficient solutions to the $4\\mathrm{p}3\\mathrm{v}$ problem reported in the literature. The first solver [36] is based on a similar idea as our solvers, i.e., to first estimate the relative pose of two cameras and then register the third camera using a $\\mathsf{P}3\\mathsf{P}$ solver [40]. To compute the pose of the first two cameras using information only from four point correspondences, the solver needs additional information about the position of one epipole. The paper shows that the four point correspondences between two calibrated views constrain the epipole in each image to lie on a curve of degree ten. Thus, the solver performs one-dimensional exhaustive search and sweeps a $10^{th}$ -degree curve of possible epipoles. For each potential epipole, it computes the relative pose of two cameras, registers the third camera using three triangulated points, and finally extracts the solution minimizing the reprojection error of the fourth point in the third view. Evaluation of the solver on one potential epipole is fast. Yet, in contrast to our proposed $4\\mathrm{p}3\\mathrm{v}$ (M)-based solvers, the error of the sampled epipole for one fixed point on the curve is not bounded since the true epipole can lie anywhere on the curve, and in many scenarios it is very far from the image center (e.g., outside the image). Thus, to obtain reasonably accurate and stable results, usually 1,000 candidates need to be evaluated. Even then, refinement at multiple local minima is required to improve the accuracy. The runtimes reported for this solver were $1 - 12ms$ depending on the number of points searched. On a small number of synthetic experiments, the paper shows that the translation error returned by the proposed solver is usually $1 - 10\\mathrm{deg}$ higher than the error returned by the $5\\mathrm{pt} + \\mathsf{P}3\\mathsf{P}$ solver. The literature does not compare against [36] as there is no publicly available implementation and it is hard to re-implement.", + "bbox": [ + 91, + 141, + 485, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second efficient solver to the 4p3v problem was published only recently [21]. In this paper, the authors first transform the 4p3v problem into a minimal problem by considering a line passing through the last correspondence in the third view. The resulting system of equations is solved using an efficient Homotopy continuation (HC) method [13, 52]. To avoid computing large numbers of spurious solutions, an MLP-based classifier is trained. For a given problem $p$ , it selects one or several starting problem-solution pairs (so-called anchors), such that the geometrically meaningful/correct solution of $p$ can be obtained by HC starting from this anchor. This strategy is fast, running $16.3\\mu s$ on average per solution. However, it has a high failure rate. The success rate of the 4p3v solver reported in [21] on two test datasets and data without noise is $26.3\\%$ . [21] do not show results for a real scenario, i.e., a RANSAC-like framework with noisy data. Providing such an evaluation,", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we show that our much simpler solvers outperform [21].", + "bbox": [ + 511, + 90, + 883, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Solutions to the 4p3v problem for orthographic and par-perspective views were presented in [33, 56]. In [33], the author suggested an iterative approach for updating to perspective views, but reported results only on a few synthetic instances. According to our extensive experiments, the update does not work on real data with general perspective cameras. This solver is returning large errors even after incorporating it into RANSAC with local optimization. We see two main reasons: (i) The set of inliers that satisfy an approximate para-perspective camera model in all three images is usually quite small, and (ii) the estimated model is often very far from the perspective optimum, and thus local optimization does not converge to a good solution.", + "bbox": [ + 511, + 107, + 903, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [11, 12, 24], the authors aim to classify and derive the number of solutions for different minimal configurations of points and lines in three calibrated views. Solutions to two minimal configurations combining points and lines were proposed in [13] and solved using a HC method [52]. Due to their complexity, the solvers are not practical. A GPU HC method was also used to solve minimal problems of four points/six lines in three views for a generalized camera in [10] and of four points in three cameras with an unknown shared focal length (4p3vf) in [5, 7]. Efficient GPU implementations of the 4p3vf solvers run $16.7ms$ to $154ms$ . These times are still too slow for practical applications.", + "bbox": [ + 511, + 303, + 903, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An approximation of perspective cameras using affine ones was used in several papers on camera geometry estimation [8]. [37] and [20] propose iterative approaches to the absolute camera pose estimation problem, i.e., the PnP problem. Both methods first compute the pose for an affine camera (weak- or para-perspective). Then the error induced by the affine camera approximation is used to adjust the constraints on the pose and recompute it. These methods do not guarantee convergence to the perspective model solution. In [41], affine cameras are used to efficiently solve the relative pose problem of two uncalibrated cameras from two affine correspondences. These affine correspondences are transformed to six point or two ellipse correspondences and then used to estimate the affine fundamental matrix. The paper also proposes a RANSAC framework that uses local optimization [30] to estimate the full fundamental matrix using inliers from the approximate affine model.", + "bbox": [ + 511, + 484, + 903, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Points that are sampled based on feature geometry to generate point correspondences from affine or scale-and orientation-invariant feature correspondences were also used in [15, 39, 42]. In contrast, our M-based solvers use only point correspondences, without associated feature geometry, to generate an additional point correspondence.", + "bbox": [ + 511, + 741, + 903, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Estimating the relative pose of three cameras", + "text_level": 1, + "bbox": [ + 511, + 845, + 903, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe different solutions for estimating the relative pose of three calibrated cameras. We start", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21915", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with a baseline solution for the minimal configuration of three points visible in all three cameras and two additional points visible in two of the three cameras (the $(5,5,3)$ configuration). Next, we present our novel solutions for the configuration of four points visible in all three cameras (the $(4,4,4)$ configuration). This configuration generates an over-constrained problem. In this case, we have one more constraint than DoF. A minimal solution would need to drop one constraint, e.g., by considering only a line passing through one of the points in the third view [21] or by considering a \"half\" point correspondence. Since, in practice, we always have full correspondences and sampling one less point in one view leads to an under-constrained problem, the $(4,4,4)$ configuration, is usually considered \"minimal\".", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5pt+P3P solver: The 5pt+P3P solver first estimates the relative pose of two cameras from 5 image point correspondences using the efficient 5pt solver [35]. Next, the three points visible in all three views are triangulated. Finally, the third camera is registered using the three 2D-3D point correspondences and the well-known efficient P3P solver [40].", + "bbox": [ + 89, + 306, + 480, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This straightforward solver, which is based on existing efficient solvers [35, 40], was discussed in several works [12, 35, 36, 46]. [36] showed that the $5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P}$ solver performs better than their dedicated $4\\mathrm{p}3\\mathrm{v}$ solver on synthetic data. Yet, the most recent works [13, 21] that study the three view relative pose problem do not discuss the $5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P}$ solver and do not use it as a baseline for comparison. To the best of our knowledge, the performance of this solver on real data and within state-of-the-art RANSAC frameworks in the context of the $4\\mathrm{p}3\\mathrm{v}$ problem has not been extensively studied. Our paper fills this gap in the literature.", + "bbox": [ + 89, + 398, + 480, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Approximate solutions to the 4p3v problem", + "text_level": 1, + "bbox": [ + 89, + 571, + 459, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to the (5,5,3) configuration, the (4,4,4) configuration, i.e., the configuration of four points in three views, leads to significantly more complex equations. State-of-the-art algebraic and numerical methods are known to fail in generating efficient and numerically stable solutions to these equations. Thus, solutions to the $4\\mathrm{p}3\\mathrm{v}$ problem require some approximations to be practical in real-world applications. This section introduces several practical approximate solutions to the $4\\mathrm{p}3\\mathrm{v}$ problem. Similarly to the $5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P}$ solver, we decompose the problem into the problem of first estimating the relative pose of two cameras and then registering the third camera with an efficient P3P solver [40]. The idea is to assume an approximate geometry only in the first two views and use four input point correspondences to efficiently estimate this geometry. Next we describe two groups of such solvers.", + "bbox": [ + 89, + 593, + 482, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4p3v(A) solver: The first group of solvers approximates the geometry in the first two views using affine cameras. Although the approximation can be quite far from the correct geometry, [41] showed that an approximate affine camera", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model provides good accuracy for relative pose estimation if coupled with geometry refitting (using the full 7DoF fundamental matrix) and local optimization inside RANSAC. The proposed $4\\mathrm{p}3\\mathrm{v}$ (A) solver first uses four point correspondences in two views to efficiently estimate the affine fundamental matrix $\\mathbf{F}_{\\mathrm{A}}$ [17] and then registers the third camera using three triangulated points and a P3P solver [40].", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4p3v(M) solver: The affine camera model used in the 4p3v (A)-based solvers can be quite imprecise. This motivates us to introduce solvers that are based on a less restrictive approximation. In this case, we approximate only one point correspondence in two views. Under the assumption of a para-perspective projection, i.e., of affine geometry, it is known that the mean point of three 3D points is projected to the mean points of their projections in both images [57]. Thus, we generate one new approximate correspondence in two views as the correspondence between the mean points of three corresponding points detected in these views. Let $\\mathbf{m}^l$ be the mean point of three points $\\{\\mathbf{x}_i^l,\\mathbf{x}_j^l,\\mathbf{x}_k^l\\}$ , $i,j,k\\in \\{1,\\ldots ,4\\}$ , in the view $l\\in \\{1,2\\}$ , then $\\mathbf{m}^{1}\\leftrightarrow \\mathbf{m}^{2}$ is our new approximate correspondence. The new 4p3v (M) solver solves the 4p3v problem by first estimating the full 5DoF essential matrix from four original correspondences $\\mathbf{x}_i^1\\leftrightarrow \\mathbf{x}_i^2$ , $i = 1,\\dots ,4$ and the $5^{th}$ correspondence $\\mathbf{m}^{1}\\leftrightarrow \\mathbf{m}^{2}$ . This is done using the efficient 5pt relative pose solver [35]. As with previous solvers, the third camera is registered using a P3P solver [40]. In short, the 4p3v (M) solver solves the 4p3v problem using the 5pt+P3P solver and one approximate correspondence.", + "bbox": [ + 511, + 202, + 903, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we show on large amounts of synthetic and real data (see Sec. 4 and Supp. mat. (SM)), the $4\\mathrm{p}3\\mathrm{v}$ (M) solver provides much more accurate estimates than the $4\\mathrm{p}3\\mathrm{v}$ (A) solver. Moreover, even in the basic form (without the modifications presented in the next section), on many scenes it performs comparably to the state-of-the-art HC solver [21]. This can be attributed to several facts and observations: (1) The $\\mathbf{m}^1 \\leftrightarrow \\mathbf{m}^2$ correspondence does not need to be seen as a correspondence of points that are projections of the mean point of three 3D points $\\mathbf{X}_i, \\mathbf{X}_j,$ and $\\mathbf{X}_k$ (in which case both $\\mathbf{m}^1$ and $\\mathbf{m}^2$ would have some error). We can look at this correspondence as a correspondence in which we fix a point in one view, e.g., $\\mathbf{m}^1$ , and generate a corresponding point in the second view. In this case, to generate a good correspondence, we only require the point in the second view to be reasonably close to the epipolar line defined by the mean point $\\mathbf{m}^1$ in the first view, i.e., the 2D point does not need to correspond to one particular 3D point with a given depth. (2) The ray from the center of the first camera through the mean point $\\mathbf{m}^1$ intersects the plane of $\\mathbf{X}_i, \\mathbf{X}_j,$ and $\\mathbf{X}_k$ at point M, which lies inside the triangle formed by these 3D points. The projection of M into the second camera lies inside the triangle formed by $\\{\\mathbf{x}_i^2, \\mathbf{x}_j^2, \\mathbf{x}_k^2\\}$ . By construction, this projection lies on the epipolar line $\\mathbf{E}\\mathbf{m}^1$ .", + "bbox": [ + 511, + 537, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21916", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus the epipolar line defined by $\\mathbf{m}^1$ passes through the triangle defined by $\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}$ . Consequently, the maximum distance of $\\mathbf{m}^2$ in the second image from the epipolar line is bounded by the maximum distance of $\\mathbf{m}^2$ from $\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}$ . For a formal lemma and proof, see SM. (3) For practical applications, when used in RANSAC, it is not necessary that each triplet $\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}$ generates a good correspondence $\\mathbf{m}^1\\leftrightarrow \\mathbf{m}^2$ . Samples with a high level of noise in the mean-point correspondence are filtered inside RANSAC. On a large number of different scenes, we observed that even if some image pairs have triplets of points that generate very noisy mean-point correspondences, there are usually enough triplets for which the noise in $\\mathbf{m}^2$ is reasonably small to lead to good estimates. (4) Four point correspondences in two views usually fix the space of possible poses such that the $5^{th}$ correspondence, even if noisy, often generates a pose that is not very far from the ground truth pose. Such a pose is usually sufficient for filtering out outliers and a good initialization for non-linear optimization on the original four points in three views and subsequent local optimization on detected inliers. We support our observations by experiments on a large amount of data.", + "bbox": [ + 89, + 90, + 483, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Making approximate solvers practical", + "text_level": 1, + "bbox": [ + 89, + 436, + 419, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $4\\mathrm{p}3\\mathrm{v}$ (A)-based solvers, when used without any modifications, generally provide imprecise results even when used inside RANSAC with local optimization on three views. While the accuracy of the pure $4\\mathrm{p}3\\mathrm{v}$ (M) solver inside LO-RANSAC [30] (RANSAC with Local Optimization) is much better, there is still room for improvement. Here we present several simple modifications of these solvers that significantly boost their performance.", + "bbox": [ + 89, + 458, + 483, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4p3v(M±δ) solver: The mean point correspondence used in the 4p3v (M) solver can provide a good approximation of a correct correspondence. Yet, as mentioned above, it can also be noisy. In the 4p3v (M±δ) solver, we thus, in addition to the mean point $\\mathbf{m}^2 = [x,y]$ of three points $\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}$ in the second image, generate two additional points. These points are (1) $\\mathbf{m}_{\\pm \\delta}^2 = [x\\pm \\delta ,y]$ if the longest dimension of the triangle $\\mathcal{T}^2 = \\Delta \\left\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\right\\}$ is in the x-direction or (2) $\\mathbf{m}_{\\pm \\delta}^2 = [x,y\\pm \\delta ]$ if it is in the y-direction. All three points, i.e., $\\mathbf{m}^2,\\mathbf{m}_{-\\delta}^2$ and $\\mathbf{m}_{+\\delta}^{2}$ are placed in correspondence with the mean point $\\mathbf{m}^1$ . The 4p3v (M±δ) solver in the first step calls the 5pt solver [35] three times, with the $5^{th}$ correspondence being either $\\mathbf{m}^1\\leftrightarrow \\mathbf{m}^2$ $\\mathbf{m}^1\\leftrightarrow \\mathbf{m}_{-\\delta}^2$ or $\\mathbf{m}^1\\leftrightarrow \\mathbf{m}_{+\\delta}^2$ . The results of these three 5pt solvers are collected to create hypotheses for the relative pose of the first two cameras inside RANSAC. The shift $\\delta$ is selected relative to the size of the triangle $\\mathcal{T}^2$ .", + "bbox": [ + 89, + 584, + 483, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Early non-minimal refitting (ENM): The geometry es", + "bbox": [ + 89, + 849, + 482, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "timated between the first two cameras using the proposed $4\\mathrm{p}3\\mathrm{v}$ (A) and $4\\mathrm{p}3\\mathrm{v}$ (M)- based solvers is only approximate. In standard LO-RANSAC, such a geometry is optimized in local optimization (LO) after registering the third view. However, this can lead to propagation of errors into triangulated points and subsequently into errors in the pose of the third camera. Inliers in three views w.r.t. such cameras together with imprecise pose initializations may not be sufficient for LO to converge to a good solution. We observed this especially for $4\\mathrm{p}3\\mathrm{v}$ (A)-based solvers. Fortunately, the design of our solvers allows us to optimize the geometry already after estimating the approximate geometry between the first two cameras. Here, even a very imprecise geometry returned by the $4\\mathrm{p}3\\mathrm{v}$ (A) solvers is usually sufficient to filter out outliers. Thus, after running the 4pt $\\mathrm{F_A / 5pt}$ solver in the first step of the proposed solvers, we use the estimated approximate models to detect inliers in two views. We then refit the estimated geometry using the non-minimal version of the 5pt solver [35], which instead of a 4-dim null space of a $5\\times 9$ matrix uses the last four vectors from the SVD/QR decomposition of a $n\\times 9$ matrix.", + "bbox": [ + 511, + 90, + 906, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$4^{\\mathrm{th}}$ point in the third view: $4\\mathrm{p}3\\mathrm{v}$ (A) and $4\\mathrm{p}3\\mathrm{v}$ (M) solvers actually solve the configuration $(4,4,3)$ , i.e., they do not use the information from the point $\\mathbf{x}_4^3$ in the third view. The information from $\\mathbf{x}_4^3$ can be used in two different ways:", + "bbox": [ + 511, + 422, + 906, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Filtering $(+F)$ : $\\mathbf{x}_4^3$ can be used to filter out geometrically infeasible solutions returned by the proposed solvers. Note that the P3P and 5pt solvers used inside the proposed methods and the $\\delta$ -based strategy return multiple solutions that can be evaluated w.r.t. $\\mathbf{x}_4^3$ to filter out spurious solutions. Since the returned solutions can be affected by the proposed approximation, we do not simply select the solution with the smallest error on $\\mathbf{x}_4^3$ , but we keep all solutions that have an epipolar error on $\\mathbf{x}_4^3$ smaller than twice the threshold used inside RANSAC. Our experiments show that this filtering can improve the speed of the proposed solvers. However, as a trade-off, there is sometimes a small drop in the accuracy of the solvers since, in some cases, geometrically correct solutions are filtered out.", + "bbox": [ + 511, + 496, + 908, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Refinement $(+\\mathbb{R})$ : $\\mathbf{x}_4^3$ can be used to refine the solutions returned by the proposed solvers. We want solutions that minimize the epipolar error on the original 4 points in all 3 views, i.e., that solve the original (4,4,4) configuration. (4,4,4) is an overconstrained configuration; thus, for noisy data, there is, in general, no solution with zero error on all 4 points in 3 views. We refine the poses by minimizing the epipolar error of the original four points in three views using LM optimization, initialized using the solutions from the $4p3v(A)$ - and $4p3v(M)$ -based solvers. Experiments with different numbers of iterations show that two iterations are usually sufficient to obtain an improvement (see SM).", + "bbox": [ + 511, + 719, + 910, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2This property was also used in the HC solver [21], which completely fails for many samples. These samples are filtered within RANSAC.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21917", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/73c81f1cb63439e0eecb94e0a58fc9c6154737b6782db60d828d2448b4cb0da6.jpg", + "image_caption": [ + "Figure 2. Results of a synthetic experiment measuring the accuracy of two-view variants of our solvers depending on the angle between the principal axes of the cameras." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 475, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 300, + 222, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We extensively evaluated the proposed solvers on a large variety of synthetic and real data to test their robustness to noise, outliers, and scene properties, and to assess their performance inside state-of-the-art RANSAC-frameworks [2, 27]. We compare our novel solvers with the homotopy continuation $4\\mathrm{p}3\\mathrm{v}$ (HC) solver [21] and the $5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P}$ baseline minimal solver for the $(5,5,3)$ configuration.", + "bbox": [ + 89, + 325, + 483, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental setup. To obtain feature correspondences, we use SuperPoint [9] features with the Light-Flue [32] matcher. We extract at most 2048 features per image. We perform matching for all three image pairs and keep only those matches that were consistently matched across all three views. We perform evaluation within two RANSAC frameworks: PoseLib [27] and GC-RANSAC [2]. For the 5pt solver, we use [35] and for the P3P solver, we use [40]. In PoseLib, we perform LO [6] using LM optimization. In GC-RANSAC, we perform LO using non-minimal solvers [18, 35] for fitting models to larger-than-minimal samples. We evaluated different shifts for our $\\delta$ -based solvers using a validation scene (for the ablation study, see SM) and selected $\\delta = 0.08 * (\\text{longest triangle dim.})$ . The choice of $\\delta$ is not critical, i.e., different values perform similarly.", + "bbox": [ + 89, + 436, + 483, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation measures. Inspired by [22], we define the pose error as max $\\left(0.5\\left(\\mathrm{R}_{err}^{12} + \\mathrm{R}_{err}^{13}\\right), 0.5\\left(\\mathrm{t}_{err}^{12} + \\mathrm{t}_{err}^{13}\\right)\\right)$ , where $\\mathbf{R}_{err}^{ij}$ and $\\mathbf{t}_{err}^{ij}$ are the angular errors of rotation and translation for pair $ij$ in degrees [22]. We also report AUC values [22] at different thresholds for the pose error. We include results for an alternative pose error definition which includes $\\mathbf{R}_{err}^{23}$ and $\\mathbf{t}_{err}^{23}$ in SM.", + "bbox": [ + 89, + 684, + 482, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Approximate camera geometry. The first experiments aim to support our idea of estimating approximate geometry in the first two views. The accuracy of the used approximations depends on a large number of variables, including the depths of the points w.r.t. the cameras, the angle under which the points are observed, the type of motion, etc. A detailed analysis of all these factors, e.g., through synthetic", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/866947a811cca479d5cd44a944d4ec17d3e87a71a804da5a42ff398c48a3eb69.jpg", + "image_caption": [ + "Figure 3. Distribution of the (left) rotation error (0.3373, 0.3349); and (right) percentage of inliers gathered (0.3266, 0.3434), as a function of the barycentric coordinates of the triangle in the second image w.r.t. the mean point of the corresponding triangle in the first image on 465k four-tuples of correspondences from St. Peter's Square scene from the PhotoTourism dataset [22]. We fit a 2D Gaussian distribution to the results and report the mean in brackets." + ], + "image_footnote": [], + "bbox": [ + 550, + 90, + 875, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "experiments, is beyond the scope of this paper.", + "bbox": [ + 511, + 319, + 821, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An approximation error introduced by the paraperspective projection, i.e., the affine geometry, and its effect on absolute camera pose determination is studied in the literature [20, 57]. Here, we thus only study the effect of the proposed approximation on relative pose estimation for real-world data and interesting synthetic scenarios. In the following experiments, all solvers are tested on two views without RANSAC. The two-view variants of our solvers are denoted as $4p(\\mathsf{A})$ , $4p(\\mathsf{M})$ , and $4p(\\mathsf{M} \\pm \\delta)$ .", + "bbox": [ + 511, + 334, + 905, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the first set of experiments, we consider several interesting synthetic setups. Fig. 2 presents the results for increasing the angle between the principal axes of the two cameras. We generate 3D points in a $2000 \\times 2000 \\times 100$ cube uniformly at random. The distances of the camera centers are random between 1000 and 1200. The cameras are looking towards the scene and their principal axes form a specified angle. To simulate realistic scenarios, we add 1px noise to the image correspondences and $20\\%$ outliers. It can be seen that the results for both approximate solvers deteriorate with increasing angle, with $4p(\\mathbb{A})$ being significantly less accurate than $4p(\\mathbb{M})$ and $4p(\\mathbb{M} \\pm \\delta)$ . However, ENM can significantly decrease errors, with the $4p(\\mathbb{M} \\pm \\delta) + ENM$ solver returning almost identical results to $5pt + ENM$ . Results for varying distances of the cameras from the scene, depth of the scene, and image noise are in SM.", + "bbox": [ + 511, + 470, + 905, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar observations as on synthetic data can be made on real-world data. Tab. 1 shows results on different scenes from the PhotoTourism dataset [22]. As can be seen, the approximate solvers are not as accurate as the $5\\mathrm{pt}$ solver when used without ENM. The gap between the $4\\mathrm{p}(\\mathrm{M}) / 4\\mathrm{p}(\\mathrm{M}\\pm \\delta)$ solvers and the $5\\mathrm{pt}$ solver is for some scenes larger (Sacre Coeur, Trevi Fountain), and for some noticeably smaller (Reichstag, St. Peter's Square), showing that the performance of our solvers is scene-dependent. The $4\\mathrm{p}(\\mathrm{A})$ solver returns large errors. The accuracy of", + "bbox": [ + 511, + 712, + 905, + 864 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "${}^{3}$ Outliers are only used in the ENM part to simulate that for approximate geometry outliers might be contained in the non-minimal samples.", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21918", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/27157558dd71e82ebcd20ef3afe6158ae9fcffe7cba93221b9505acdddb7d799.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MED (°)\\( {20}^{\\text{th }} \\) perc. (°)
w/o ENM
Scene5p (E)4p (M)4p (M±δ)4p (A)5p (E)4p (M)4p (M±δ)4p (A)
Brandenburg Gate16.1419.0215.6975.327.229.837.4645.99
Buckingham Palace18.8721.1217.7466.198.0410.018.2340.59
Coloseum Exterior17.4023.5018.9366.676.4411.368.3440.82
Grand Place Brussels18.4020.2816.7869.318.0410.098.0744.45
Notre Dame Front Facade15.8223.9719.6171.606.4012.279.0148.39
P Palace of Westminster16.0417.6214.7269.384.997.986.4046.49
Puntheon Exterior22.6525.7121.3161.3310.6514.0910.9438.12
Reichstag12.1512.9910.2782.855.156.164.5760.12
Sacre Coeur11.8117.6114.3174.213.697.955.8743.72
St. Peter's Square17.8518.7515.4872.758.519.717.6644.60
Taj Mahal9.9511.208.5682.743.774.903.6364.27
Temple Nara Japan18.5721.5317.0164.977.6010.638.1632.21
Trevi Fountain20.9327.7722.8950.128.1013.8910.7831.32
w/ ENM
Scene5p (E)4p (M)4p (M±δ)4p (A)5p (E)4p (M)4p (M±δ)4p (A)
Brandenburg Gate14.5317.0514.3828.406.748.866.9316.96
Buckingham Palace16.7118.7616.0735.077.358.967.4619.72
Colosseum Exterior16.2121.4617.7034.906.3510.718.0421.98
Grand Place Brussels16.2117.9615.2331.537.278.857.2418.40
Notre Dume Front Facade15.0522.0818.4634.496.3011.658.7122.08
Palace of Westminster13.8314.9512.7429.064.687.135.8416.41
Puntheon Exterior21.1924.0320.3037.2110.3213.2810.5524.42
Reichstag9.3210.038.3617.694.405.063.979.39
Sacre Coeur10.7115.1912.6424.923.617.235.4913.41
St. Peter's Square16.0116.9414.3528.607.758.637.0315.68
Taj Mahal8.329.327.4315.853.344.103.187.72
Temple Nara Japan16.2419.3815.7429.976.839.197.3216.65
Trevi Fountain20.2226.4222.0939.678.0213.4310.5326.62
", + "bbox": [ + 91, + 88, + 460, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Accuracy of two-view solvers on PhotoTourism scenes.", + "bbox": [ + 94, + 339, + 475, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the $4\\mathrm{p}(\\mathrm{M}) / 4\\mathrm{p}(\\mathrm{M}\\pm \\delta)$ solvers is not very far from the 5pt solver, with the $4\\mathrm{p}(\\mathrm{M}\\pm \\delta)$ solver even slightly outperforming the 5pt solver on some scenes (Reichstag, Taj Mahal). ENM helps to increase the precision of all solvers, especially $4\\mathrm{p}(\\mathrm{A})$ . Still, the $4\\mathrm{p}(\\mathrm{A}) + \\mathrm{ENM}$ solver returns quite high errors. However, in practice, all solvers are used inside RANSAC, where the current best pose is locally optimized on inliers. LO compensates for less accurate pose estimates and, as we show on real experiments for three as well as two views (see SM), it can suppress the errors of the $4\\mathrm{p}(\\mathrm{A}) / 4\\mathrm{p}3\\mathrm{v}(\\mathrm{A})$ -based solvers, making them comparable to the $4\\mathrm{p}(\\mathrm{M}) / 4\\mathrm{p}3\\mathrm{v}(\\mathrm{M})$ -based solvers. For the three view scenario, the performance of all of our solvers can be further improved by pose refinement (+R) and filtering (+F).", + "bbox": [ + 88, + 383, + 482, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the last set of experiments, we study the accuracy of the mean point correspondence. We sample 100 four-tuples of point correspondences consistent with the ground truth relative pose, i.e., inliers, for each image pair in scenes from the PhotoTourism dataset [22]. We use the first three correspondences to define the triangles in both images. Then, we establish correspondences between the mean of the triangle in one image and various points in the triangle in the second image. We express points in the second triangle via their barycentric coordinates and uniformly sample $19 \\times 19$ barycentric coordinates $(a, b) \\in [0, 1]^2$ , such that $a + b \\leq 1$ .", + "bbox": [ + 89, + 597, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 3 shows the results for the rotation error and the percentage of inliers consistent with the pose obtained with the $4\\mathrm{p}$ (M) solver on the St. Peter's Square scene. The optimum of the metrics is reached around the mean point of the triangles (the mean values of 2D Gaussians fitted to the results are very close to the mean point $(0.3,0.3)$ of the triangles). A detailed description, the translation and the symmetric epipolar errors, and the results for more scenes are in SM. For all tested scenes, we observed a similar behavior.", + "bbox": [ + 89, + 765, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiments on real data. We test the solvers on all scenes from the PhotoTourism dataset [22, 50] which provide ground truth poses and intrinsics via a COLMAP [49] reconstruction. In the results, we do not include the St. Peter's Square scene that we used for the validation of $\\delta$ and the number of refinement iterations (see SM). We also include results for the Cambridge Landmarks dataset [23] (except the Street scene, which is commonly not used due to issues with its ground truth) and Aachen Day-Night v1.1 [58]. For PhotoTourism and Aachen, we use the images in their original resolution. For Cambridge Landmarks, we resize them so that the larger side is $800\\mathrm{px}$ . For each scene, we sample 5,000 random image triplets with at least 10 matches obtained with [9, 32] with at least $10\\%$ overlap [22].", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows the results for PoseLib RANSAC with early termination and 5px epipolar threshold. Results for GC-RANSAC are in SM. With the suggested modifications (see Sec. 3.2) all proposed solvers outperform the state-of-the-art $4\\mathrm{p}3\\mathrm{v}$ (HC) solver [21] in terms of pose accuracy with many of the variants also achieving faster runtimes.", + "bbox": [ + 511, + 304, + 903, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using filtering $(+F)$ improves the run-time of RANSAC at the cost of a decrease in pose accuracy. Still, our $4p3v(M) + R + F$ solver outperforms $4p3v(HC)$ in terms of both accuracy and run-time. The $4p3v(M \\pm \\delta)$ and $4p3v(M \\pm \\delta) + R$ solvers clearly improve upon the $4p3v(M)$ solvers, albeit at an increased run-time. Still the $4p3v(M \\pm \\delta) + R + F$ solver provides, in general, the best speed-accuracy trade-off. The results also show that early non-minimal refinement $(+ENM)$ is necessary for the $4p3v(A)$ solver to perform well. For the other approximate solvers, this form of refinement also leads to improved accuracy at some cost to the runtime. Assuming that the $5pt + P3P$ solver samples five points from which at least four have projections in all three cameras, the suggested pose refinement $(+R)$ and filtering $(+F)$ can also be applied to the baseline $5pt + P3P$ by using the fourth (potentially the fifth) point in the third view. The suggested modifications improve the performance of the $5pt + P3P$ solver, with the $5pt + P3P + R + F + ENM$ solver having an accuracy similar to that of the $4p3v(M) + R + F + ENM$ solver, however at slower run-times. Moreover, it is less accurate than the proposed $4p3v(M \\pm \\delta) + R + F + ENM$ solver.", + "bbox": [ + 511, + 397, + 903, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also investigate the speed-accuracy trade-off of the solvers by running PoseLib RANSAC [27] for a set of fixed numbers of iterations. Runtimes are reported for 1 core of a $2\\mathrm{GHz}$ Intel Xeon Gold 6338 CPU. As shown in Fig. 4, the proposed $4\\mathrm{p}3\\mathrm{v}$ $(\\mathbb{M}\\pm \\delta) + \\mathbb{R} + \\mathbb{F}$ solver consistently provides the best speed-accuracy trade-off both with and without ENM. $4\\mathrm{p}3\\mathrm{v}$ (A) $+\\mathbb{R} + \\mathbb{F} + \\mathbb{ENM}$ provides a similar performance, typically beating $4\\mathrm{p}3\\mathrm{v}$ (HC) [21]. However, as shown in Fig. 4d for the St. Mary's Church scene [23], it may perform worse for some specific scenes.", + "bbox": [ + 511, + 732, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The SM provides detailed ablation studies on the signif", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21919", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/63e5c894136f5d0be36096f78de823ad5f38a80e8109eeb53ab1c4f30aa0d01b.jpg", + "image_caption": [ + "(a) Phototourism [22]", + "(b) Cambridge Landmarks [23]", + "(c) Aachen Day-Night v1.1 [58]", + "(d) St. Mary's Church [23]", + "Figure 4. Speed-accuracy trade-off for (a) all scenes from PhotoTourism [22] except St. Peter's Square, (b) 5 scenes from Cambridge Landmarks [23], (c) Aachen Day-Night v1.1 [58], and (d) St. Mary's Church scene from the Cambridge Landmarks dataset [23]. We report the AUC@10° of the pose error and vary the number of PoseLib RANSAC iterations (100, 200, 500, 1000, 2000, 5000, 10,000) with a 5 px epipolar threshold. Runtimes are averaged over all image triplets." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 903, + 220 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/442b33bf3f54da311f50ba99ac9fc2ee26501849307ece6ab58ac2408e40dd8c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EstimatorPhotoTourism [22]Cambridge Landmarks [23]Aachen Day-Night v1.1 [58]
AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓
4p3v (HC) [21]6.371.6262.4173.5381.9464.108.732.6747.9964.5876.6560.1112.433.5542.2855.2066.5567.26
5pt+P3P5.401.6062.9374.3083.0033.777.462.6548.3665.3377.7524.0410.523.3843.4356.7368.2853.34
5pt+P3P+ENM5.231.5663.7775.0383.5048.797.202.6348.7465.6878.0434.8210.593.4143.3256.6468.0671.61
5pt+P3P+R+F5.101.5064.9075.9084.1138.127.152.5849.4266.3578.5531.8210.373.3643.7357.1568.7154.34
5pt+P3P+R+F+ENM4.991.5065.1876.2084.3056.526.982.5849.3866.3178.5747.0710.203.3443.7557.1168.6376.86
4p3v (A)37.4825.1622.9128.5935.1616.5841.2321.7322.7531.4738.9813.3336.4019.5324.4131.7339.2432.04
4p3v (A)+ENM4.971.4865.4976.4784.5340.457.022.5949.3566.3278.5328.3510.353.3843.6256.9968.5562.44
4p3v (A)+R33.4519.9525.8632.2239.1616.3236.9811.5927.0336.7044.4812.4834.8418.2225.2732.7740.2929.56
4p3v (A)+R+F35.6523.4723.9929.9736.6411.1038.3414.4725.8935.1942.809.3536.7720.7623.9431.2338.5819.75
4p3v (A)+R+F+ENM4.981.4865.6276.5684.5532.376.992.5749.6266.5578.7223.4310.433.3843.5756.9568.5642.36
4p3v (M)6.021.7160.8672.5681.7335.188.362.7447.1564.0676.5925.1512.093.6241.7855.0666.6754.89
4p3v (M)+ENM5.281.5863.6174.9683.4648.837.212.6348.6965.5777.9534.9310.673.4343.3156.6168.0870.87
4p3v (M)+R5.491.5663.6674.8883.3741.527.692.6348.6965.6777.9930.7611.173.4643.2856.3967.8860.40
4p3v (M)+R+F5.481.5463.9975.1483.5130.887.752.6248.7765.7177.9822.7211.163.4843.0456.2867.7342.38
4p3v (M)+R+F+ENM5.011.5065.1876.1684.2844.516.992.5849.4366.3178.5632.4810.253.3943.7457.2168.8058.08
4p3v (M±δ)5.551.6861.8673.7382.7883.707.652.6748.1665.2577.7659.2311.133.4742.7456.1167.74125.02
4p3v (M±δ)+ENM4.991.5664.0275.5184.01125.666.882.6049.1166.1378.5189.0510.263.3543.6656.9968.55175.54
4p3v (M±δ)+R4.861.4765.9777.0085.01100.617.102.5649.7966.8979.0873.9410.373.3443.9357.2668.82138.53
4p3v (M±δ)+R+F4.921.4765.8576.8784.9071.737.192.5649.7166.7578.9552.8410.523.3643.7957.1768.7392.89
4p3v (M±δ)+R+F+ENM4.661.4666.1277.1485.15112.606.652.5549.8766.9579.2081.989.963.3244.0257.4268.97139.19
", + "bbox": [ + 91, + 319, + 906, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Results for different solvers implemented in the PoseLib framework [27] on 12 scenes from PhotoTourism [22], 5 scenes from Cambridge Landmarks [23] and Aachen Day-Night v1.1 [58]. We mark the best and second best results. Runtimes are reported in ms for the whole RANSAC with early termination (0.9999 confidence, minimum 100 iterations) and epipolar threshold set to 5 px.", + "bbox": [ + 89, + 517, + 906, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "icance of each individual modification $(+R / + C / + ENM)$", + "bbox": [ + 89, + 585, + 450, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. As discussed, the accuracy of the proposed approximate solvers is scene dependent. However, after using the proposed ENM refitting, the scene-dependency is negligible. Especially for the $4\\mathrm{p}3\\mathrm{v}$ $(\\mathbb{M}\\pm \\delta)$ +ENM solver, we have not noticed performance drops for some specific scene geometries or camera configurations. While the two-view variants of our solvers $4\\mathrm{p}$ $(\\mathbb{M}) / 4\\mathrm{p}$ $(\\mathbb{M}\\pm \\delta)$ provide very similar results to the 5pt solver in two views, in this case they do not outperform the 5pt solver. In the three-view scenario, the better performance is achieved thanks to the proposed pose refinement $(+\\mathbf{R})$ and filtering $(+\\mathbf{F})$", + "bbox": [ + 88, + 606, + 482, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed ENM refitting cannot be applied to the $4\\mathrm{p}3\\mathrm{v}$ (HC) solver [21], since this solver estimates the pose of all three cameras together. It is theoretically applicable to the solver proposed in [36]. However, here ENM would have needed to be run on hundreds of candidate poses corresponding to tens-to-hundreds of sampled epipoles. This", + "bbox": [ + 89, + 773, + 482, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is because for one fixed point on the curve of epipoles, the error of the sampled epipole is not bounded since the true epipole can lie anywhere on the curve, even outside the image. Thus, ENM might avoid local minima present in the second step of the original solver, but would significantly slow down the solver in the first step, making it much slower than our $4\\mathrm{p}3\\mathrm{v}$ $(\\mathbf{M}) + \\mathsf{ENM} / 4\\mathrm{p}3\\mathrm{v}$ $(\\mathbf{M}\\pm \\delta)$ +ENM solvers.", + "bbox": [ + 511, + 585, + 906, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 708, + 633, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We consider the highly challenging problem of relative pose estimation of three calibrated cameras from four correspondences. We propose novel approaches that solve the problem by utilizing solvers based on approximate geometry. The best performing solver uses a simple, yet novel strategy by using mean coordinates of three input points and points in their vicinity as an approximate $5^{th}$ correspondences. Extensive experiments show that our solvers achieve state-of-the-art performance on a large variety of real scenes. At the same time, our solvers are simple to implement, especially compared to the current state-of-the-art [21].", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "This weakness also applies to [21], since the scene needs to be similar enough to the training scenes for the MLP-based classifier to work well.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21920", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was funded by the Czech Science Foundation (GACR) JUNIOR STAR Grant No. 22-23183M (supporting C.T., Y.D., and Z.K.), the Grant Agency of the Czech Technical University in Prague grant no. SGS23/173/OHK3/3T/13 (supporting C.T.), the EU NextGenerationEU through the Recovery and Resilience Plan for Slovakia under the project \"InnovAite Slovakia, Illuminating Pathways for AI-Driven Breakthroughs\" No. 09I02-03-V01-00029, the TERAIS project - a Horizon-Widera-2021 program of the European Union under the Grant agreement number 101079338 (supporting V.K. and Z.B.H.), and the EU Horizon 2020 project RICAIP grant agreement No. 857306 (supporting T.S.). Part of the research results was obtained using the computational resources procured in the national project National competence centre for high performance computing project code: 311070AKF2, funded by European Regional Development Fund, EU Structural Funds Informatization of society, Operational Program Integrated Infrastructure.", + "bbox": [ + 89, + 90, + 485, + 380 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 93, + 392, + 187, + 407 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chris Aholt and Luke Oeding. The ideal of the trifocal variety. Mathematics of Computation, 83(289):2553-2574, 2014. 1, 3", + "[2] D. Barath and J. Matas. Graph-Cut RANSAC. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6733-6741, 2018. 1, 6", + "[3] Martin Bujnak, Zuzana Kukelova, and Tomas Pajdla. A general solution to the p4p problem for camera with unknown focal length. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1", + "[4] Robert Castle, Georg Klein, and David W. Murray. Videorelate localization in multiple maps for wearable augmented reality. In ISWC, 2008. 1", + "[5] Chiang-Heng Chien, Hongyi Fan, Ahmad Abdelfattah, Elias Tsigaridas, Stanimire Tomov, and Benjamin Kimia. Gpubased homotopy continuation for minimal problems in computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15765-15776, 2022. 3", + "[6] Ondrej Chum, Jiří Matas, and Josef Kittler. Locally optimized ransac. In Pattern Recognition, pages 236-243. Springer Berlin Heidelberg, 2003. 1, 6", + "[7] Andrea Porfiri Dal Cin, Timothy Duff, Luca Magri, and Tomas Pajdla. Minimal perspective autocalibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5064-5073, 2024. 3", + "[8] Toby Collins and Adrien Bartoli. Infinitesimal plane-based pose estimation. International Journal of Computer Vision, 109:252-286, 2014. 3", + "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 224-236, 2018. 6, 7" + ], + "bbox": [ + 99, + 417, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Yaqing Ding, Chiang-Heng Chien, Viktor Larsson, Karl Åström, and Benjamin Kimia. Minimal solutions to generalized three-view relative pose problem. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8156-8164, 2023. 3", + "[11] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. Plmp-point-line minimal problems in complete multiview visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1675-1684, 2019. 3", + "[12] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomás Pajdla. $\\mathrm{Pl}_1\\mathfrak{p}$ - point-line minimal problems under partial visibility in three views. In ECCV (26), pages 175-192. Springer, 2020. 3, 4", + "[13] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H. Regan, David da Costa de Pinho, Elias P. Tsigaridas, Charles W. Wampler, Jonathan D. Hauenstein, Peter J. Giblin, Benjamin B. Kimia, Anton Leykin, and Tomás Pajdla. TRPLP - trifocal relative pose from lines at points. In CVPR, pages 12070-12080. Computer Vision Foundation / IEEE, 2020. 3, 4", + "[14] M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6): 381-395, 1981. 1", + "[15] Liran Goshen and Ilan Shimshoni. Balanced exploration and exploitation model search for efficient epipolar geometry estimation. In Computer Vision - ECCV 2006, pages 151-164, Berlin, Heidelberg, 2006. Springer Berlin Heidelberg. 3", + "[16] Christian Häne, Lionel Heng, Gim Hee Lee, Friedrich Fraun-dorfer, Paul Furgale, Torsten Sattler, and Marc Pollefeys. 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image and Vision Computing, 68:14-27, 2017. 1", + "[17] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision, 2006. 2, 4", + "[18] Joel A. Hesch and Stergios I. Roumeliotis. A direct least-squares (dls) method for pnp. In 2011 International Conference on Computer Vision, pages 383-390, 2011. 6", + "[19] R.J. Holt and A.N. Netravali. Uniqueness of solutions to three perspective views of four points. IEEE Transactions on Pattern Analysis and Machine Intelligence, 17(3):303-307, 1995. 1, 3", + "[20] Radu Horaud, Fadi Dornaika, and Bart Lamiroy. Object pose: The link between weak perspective, parapspective, and full perspective. Int. J. Comput. Vision, 22(2):173-189, 1997. 3, 6", + "[21] Petr Hruby, Timothy Duff, Anton Leykin, and Tomas Pajdla. Learning to solve hard minimal problems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5532-5542, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[22] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21921", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Proceedings of the IEEE international conference on computer vision, pages 2938-2946, 2015. 7, 8", + "[24] J. Kileel. Minimal problems for the calibrated trifocal variety. SIAM Journal on Applied Algebra and Geometry, 1(1): 575-598, 2017. 1, 3", + "[25] Z. Kukelova and T. Pajdla. A minimal solution to the autocalibration of radial distortion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2007), 2007. 1", + "[26] Zuzana Kukelova, Martin Bujnak, and Tomas Pajdla. Real-time solution to the absolute pose problem with unknown radial distortion and focal length. In Proceedings of the IEEE International Conference on Computer Vision, pages 2816-2823, 2013. 1", + "[27] Viktor Larsson. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 6, 7, 8", + "[28] Viktor Larsson, Kalle Åström, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Computer Vision and Pattern Recognition (CVPR), pages 820-829, 2017. 2", + "[29] Viktor Larsson, Torsten Sattler, Zuzana Kukelova, and Marc Pollefeys. Revisiting radial distortion absolute pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1062-1071, 2019. 1", + "[30] Karel Lebeda, Jiri Matas, and Ondrej Chum. Fixing the locally optimized ransac. In Proceedings of the British Machine Vision Conference 2012, pages 4961-4974. British Machine Vision Association, 2012. 3, 5", + "[31] S. Leonardos, R. Tron, and K. Daniilidis. A metric parametrization for trifocal tensors with non-colinear pinholes. In Computer Vision and Pattern Recognition (CVPR), pages 259-267. IEEE, 2015. 1, 3", + "[32] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In International Conference on Computer Vision (ICCV), pages 17627–17638, 2023. 6, 7", + "[33] H. Christopher Longuet-Higgins. A method of obtaining the relative positions of 4 points from 3 perspective projections. In BMVC91, pages 86-94, London, 1991. Springer London. 3", + "[34] Evgeniy V. Martyushev. On some properties of calibrated trifocal tensors. Journal of Mathematical Imaging and Vision, 58(2):321-332, 2017. 1, 3", + "[35] D. Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 1, 2, 4, 5, 6", + "[36] D. Nister and F. Schaffalitzky. Four points in two or three calibrated views: Theory and practice. International Journal of Computer Vision, 67(2):211-231, 2006. 1, 2, 3, 4, 8", + "[37] Denis Oberkampf, Daniel F DeMenton, and Larry S Davis. Iterative pose estimation using coplanar feature points. Computer Vision and Image Understanding, 63(3):495-511, 1996. 3" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Magnus Oskarsson, Andrew Zisserman, and Kalle Åström. Minimal projective reconstruction for combinations of points and lines in three views. Image and Vision Computing, 22 (10):777-785, 2004. British Machine Vision Computing 2002. 2", + "[39] Michal Perdoch, Jiri Matas, and Ondrej Chum. Epipolar geometry from two correspondences. In 18th International Conference on Pattern Recognition (ICPR'06), pages 215-219. IEEE, 2006. 3", + "[40] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (p3p) solver. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318-332, 2018. 1, 2, 3, 4, 6", + "[41] James Pritts, Ondrej Chum, and Jiri Matas. Approximate models for fast and accurate epipolar geometry estimation. In 28th International Conference on Image and Vision Computing New Zealand, IVCNZ 2013, Wellington, New Zealand, November 27-29, 2013, pages 106-111. IEEE, 2013. 2, 3, 4", + "[42] James Pritts, Zuzana Kukelova, Viktor Larsson, and Ondrej Chum. Radially-distorted conjugate translations. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 1993-2001. IEEE Computer Society, 2018. 3", + "[43] Long Quan. Invariants of six points and projective reconstruction from three uncalibrated images. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 17(1): 34-46, 1995. 2", + "[44] L. Quan, B. Triggs, and B. Mourrain. Some results on minimal euclidean reconstruction from four points. Journal of Mathematical Imaging and Vision, 24(3):341-348, 2006. 1, 3", + "[45] R. Raguram, O. Chum, M. Pollefeys, J. Matas, and J.-M. Frahm. USAC: A universal framework for random sample consensus. IEEE Transactions on Pattern Recognition and Machine Intelligence, 35(8):2022-2038, 2013. 1", + "[46] Volker Rodehorst. Evaluation of the metric trifocal tensor for relative three-view orientation. In Digital Proceedings, International Conference on the Applications of Computer Science and Mathematics in Architecture and Civil Engineering: July 20 - 22 2015, Bauhaus-University Weimar, 2017. 4", + "[47] T. Sattler, B. Leibe, and L. Kobbelt. Efficient & Effective Prioritized Matching for Large-Scale Image-Based Localization. IEEE Transactions on Pattern Recognition and Machine Intelligence, 39(9):1744-1756, 2016. 1", + "[48] D. Scaramuzza and F. Fraundorfer. Visual odometry [tutorial]. IEEE Robot. Automat. Mag., 18(4):80-92, 2011. 1", + "[49] Johannes L. Schonberger and Jan-Michael Frahm. Structure-From-Motion Revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104–4113, 2016. 7", + "[50] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 7", + "[51] N. Snavely, S. M. Seitz, and R. Szeliski. Modeling the world from internet photo collections. International Journal Computer Vision, 80(2):189-210, 2008. 1" + ], + "bbox": [ + 516, + 90, + 906, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21922", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Andrew J Sommese, Andrew J Sommese, and Charles W Wampler. The numerical solution of systems of polynomials arising in engineering and science. World Scientific Publishing Co. Pte. Ltd, Singapore, 2005. 3", + "[53] H. Stewenius, D. Nister, F. Kahl, and F. Schaffalitzky. A minimal solution for relative pose with unknown focal length. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2005), 2005. 1", + "[54] H. Stewenius, C. Engels, and D. Nister. Recent developments on direct relative orientation. ISPRS J. of Photogrammetry and Remote Sensing, 60:284-294, 2006. 1", + "[55] P. H. S. Torr and A. Zisserman. Robust parameterization and computation of the trifocal tensor. Image and Vision Computing, 15:591-605, 1997. 2", + "[56] G. Xu and N. Sugimoto. A linear algorithm for motion from three weak perspective images using euler angles. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(1):54-57, 1999. 3", + "[57] Zhengyou Zhang. Weak Perspective Projection, pages 877-883. Springer US, Boston, MA, 2014. 2, 4, 6", + "[58] Zichao Zhang, Torsten Sattler, and Davide Scaramuzza. Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision, 129(4):821-844, 2021. 7, 8" + ], + "bbox": [ + 91, + 90, + 480, + 431 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21923", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/3fec933e-2dde-4be1-8ad1-28710156467f_model.json b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/3fec933e-2dde-4be1-8ad1-28710156467f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..381451e6082cbf7e210db9deaacdecc54d498173 --- /dev/null +++ b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/3fec933e-2dde-4be1-8ad1-28710156467f_model.json @@ -0,0 +1,1938 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "Practical solutions to the relative pose of three calibrated cameras" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.938, + 0.217 + ], + "angle": 0, + "content": "Charalambos Tzamos\\(^{1,*}\\) Viktor Kocur\\(^{2,*}\\) Yaqing Ding\\(^{1}\\) Daniel Barath\\(^{3}\\) Zuzana Berger Haladová\\(^{2}\\) Torsten Sattler\\(^{4}\\) Zuzana Kukelova\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.228, + 0.916, + 0.281 + ], + "angle": 0, + "content": "\\(^{1}\\)Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague \\(^{2}\\)Faculty of Mathematics, Physics and Informatics, Comenius University in Bratislava \\(^{3}\\)ETH Zürich (Zurich), HUN-REN SZTAKI (Budapest)" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.282, + 0.9, + 0.3 + ], + "angle": 0, + "content": "\\(^{4}\\)Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.335, + 0.327, + 0.351 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.486, + 0.653 + ], + "angle": 0, + "content": "We study the challenging problem of estimating the relative pose of three calibrated cameras from four point correspondences. We propose novel efficient solutions to this problem that are based on the simple idea of using four correspondences to estimate an approximate geometry of the first two views. We model this geometry either as an affine or a fully perspective geometry estimated using one additional approximate correspondence. We generate such an approximate correspondence using a very simple and efficient strategy, where the new point is the mean point of three corresponding input points. The new solvers are efficient and easy to implement, since they are based on existing efficient minimal solvers, i.e., the 4-point affine fundamental matrix, the well-known 5-point relative pose solver, and the P3P solver. Extensive experiments on real data show that the proposed solvers, when properly coupled with local optimization, achieve state-of-the-art results, with the novel solver based on approximate mean-point correspondences being more robust and accurate than the affine-based solver." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.674, + 0.222, + 0.689 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.485, + 0.88 + ], + "angle": 0, + "content": "Camera geometry estimation is crucial in many computer vision applications, e.g., visual navigation [48], Structure-from-Motion [51], augmented reality [4], self-driving cars [16], and visual localization [47]. Due to noise and outliers in input correspondences, the predominant way for camera geometry estimation is to use a hypothesis-and-test framework, e.g., RANSAC [2, 6, 14, 45]. For RANSAC-like methods, using as few (ideally the minimal number of) correspondences as possible for estimation is important since the number of RANSAC iterations (and thus its run-time) grows exponentially with the number of correspondences required for the model estimation." + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.333, + 0.821, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.51 + ], + "angle": 0, + "content": "Figure 1. Visualization of the four-points-in-three-views (4p3v) problem and our solution based on using four correspondences to efficiently estimate an approximate geometry of the first two views and then register the third view using a P3P solver [40]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.523, + 0.907, + 0.749 + ], + "angle": 0, + "content": "Minimal camera geometry problems often result in complex systems of polynomial equations. Efficient algebraic methods helped solve many previously unsolved problems [3, 25, 26, 29, 53, 54]. Still, they fail to generate efficient and/or numerically stable solutions for some configurations. In this paper, we study one such challenging problem: Estimating the relative pose of three calibrated cameras. This problem has received attention for a long time [1, 19, 31, 34, 44]. However, due to its complexity, it is still not considered fully solved. There are no efficient and practical solutions for most of the minimal configurations of point and/or line correspondences [24]. One such configuration that is particularly interesting is the notoriously difficult configuration of four points in three calibrated views [35, 44], known as the \\(4\\mathrm{p}3\\mathrm{v}\\) problem." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "State-of-the-art algebraic and numerical methods are known to fail in generating efficient and numerically stable solutions to the 4p3v problem. The existing methods for solving this problem are only approximate [21, 36]. By solving only for one (or a few) solutions from the 272 solutions of the 4p3v problem [21], and by discretely sampling the space of potential solutions [36], the existing 4p3v methods can often fail, i.e., the returned solution can be, in general, arbitrarily far from the geometrically correct solution. To decrease the failure rate, both methods [21, 36]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21913" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.107 + ], + "angle": 0, + "content": "require a lot of tuning and are challenging to re-implement.1" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.108, + 0.482, + 0.228 + ], + "angle": 0, + "content": "In contrast to complex solutions to the configuration of four points in three views, there is a simple and efficient solution for the configuration where five correspondences are detected in the first two views and three of these points are also visible in the third view. In this case, the \\(5\\mathrm{pt} + \\mathrm{P3P}\\) solver first estimates the relative pose of two cameras from five correspondences using an efficient 5pt solver [35], and then registers the third camera using a P3P solver [40]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.228, + 0.482, + 0.319 + ], + "angle": 0, + "content": "Following the idea of first estimating the relative pose of two cameras, we propose two novel approaches for solving the 4p3v problem. Our solutions are based on the simple idea of using four correspondences to estimate an approximate geometry of the first two views and then registering the third camera using a P3P solver [40]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.32, + 0.482, + 0.41 + ], + "angle": 0, + "content": "In the first approach, we approximate the geometry of the first two views using affine cameras. The approximation using affine cameras was shown to provide good accuracy for relative pose estimation if coupled with geometry refitting and local optimization inside RANSAC [41]. We denote the proposed affine-based 4p3v solver as 4p3v (A)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.412, + 0.483, + 0.668 + ], + "angle": 0, + "content": "In the second approach, we use a less restrictive approximation. In this case, we estimate the full perspective geometry, i.e., the full 5DoF essential matrix, using four input correspondences and one additional approximate correspondence in two views. We generate the approximate correspondence using the locations of three of the four input correspondences. Under the assumption of a paraperspective projection, i.e., of affine geometry, the mean point of three 3D points is projected to the mean points in both images [57]. Thus, the new correspondence is generated as the correspondence between the mean points of three corresponding points detected in two views. This approximate mean-point correspondence can also be seen as a correspondence under the \\(1^{st}\\)-order approximation of the homography defined by the plane passing through the three corresponding 3D points. We denote the proposed mean point-based 4p3v solver as 4p3v (M)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.483, + 0.865 + ], + "angle": 0, + "content": "While the proposed \\(4\\mathrm{p}3\\mathrm{v}\\) (M) solver returns significantly more accurate poses than the \\(4\\mathrm{p}3\\mathrm{v}\\) (A) solver, both solvers may not be accurate enough if not properly treated inside RANSAC. Thus, in this paper, we propose several ways to improve the accuracy of the proposed approximate solvers. (i) To compensate for noise in the mean point correspondences, we introduce the \\(4\\mathrm{p}3\\mathrm{v}\\) \\((\\mathbb{M}\\pm \\delta)\\) solver that generates two additional correspondences in the vicinity of the mean point, and in the first step calls the 5pt solver [35] three times. (ii) To improve the approximate geometry estimated between the first two views, we refit this geometry using a non-minimal relative pose solver and inliers obtained from the approximate geometry (ENM). (iii) We use" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.196 + ], + "angle": 0, + "content": "the fourth correspondence in the third view to filter out geometrically infeasible solutions \\((+F)\\) and to refine the solutions on four input correspondences in three views using just a few iterations of Levenberg-Marquardt (LM) refinement \\((+R)\\). While conceptually simple and efficient, the novel solvers achieve state-of-the-art results on real data. The contributions of the paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.393 + ], + "angle": 0, + "content": "- We propose two groups of novel solutions for the well-known and challenging 4p3v problem: \\(4\\mathrm{p}3\\mathrm{v}(\\mathrm{A})\\) - and \\(4\\mathrm{p}3\\mathrm{v}(\\mathrm{M})\\) -based solvers. These solutions are based on the simple idea of using four correspondences to estimate an approximate geometry in the first two views. Compared to state-of-the-art 4p3v solvers [21, 36], which are non-trivial and difficult to re-implement such that they are numerically stable and fast, our new solutions can be easily implemented using existing efficient implementations of the linear 4pt affine fundamental matrix solver [17], the 5pt solver [35] and the P3P solver [40]. The source code is available at https://github.com/kocurvik/threeview." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.905, + 0.544 + ], + "angle": 0, + "content": "- We present several ways of improving the accuracy and the speed of the proposed solvers, as well as a strategy for efficiently using these approximate solvers inside a RANSAC-style paradigm. We show that the new solvers achieve state-of-the-art results in terms of accuracy on real data. While both \\(4\\mathrm{p}3\\mathrm{v}\\) (A)- and \\(4\\mathrm{p}3\\mathrm{v}\\) (M)-based solvers achieve comparable results when coupled with the suggested non-minimal geometry refitting and local optimization, our \\(4\\mathrm{p}3\\mathrm{v}\\) (M)-based solvers are more robust to the scene geometry and RANSAC inlier thresholds." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.545, + 0.905, + 0.635 + ], + "angle": 0, + "content": "- To our knowledge, we are the first to extensively evaluate solutions to the 4p3v problem on a large variety of real-world scenes and within state-of-the-art RANSAC frameworks, and to compare them to the baseline 5pt+P3P solver. We report results on 3 datasets, consisting of 18 different scenes and altogether 90,000 camera triplets." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.652, + 0.665 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.81 + ], + "angle": 0, + "content": "Estimating the relative pose of three cameras from a minimal number of point and line correspondences is known as an extremely challenging problem. For three uncalibrated cameras, 6 point correspondences are necessary to estimate the trifocal tensor, with a solution known for a long time [43, 55]. Solutions to three minimal combinations of points and lines are presented in [38]. The configuration of 9 lines is more challenging and was solved only recently [28]. Yet, the solver is far from practical and runs 17.8s." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "For calibrated cameras, the configuration that attracts most of the attention is the configuration of four points in three views (the 4p3v problem). Note that this is not a minimal configuration since it generates 12 constraints for 11 degrees-of-freedom (DoF). The 4p3v problem is known to be extremely difficult to solve. Several papers present" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "There is no publicly available implementation for [36]. Implementation of [21] is quite complex and requires a non-negligible effort to run." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21914" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.138 + ], + "angle": 0, + "content": "mostly theoretical results [1, 31, 34]. For four triplets of exact points without noise, it is shown that the 4p3v problem has, in general, a unique solution [19, 44]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.142, + 0.486, + 0.641 + ], + "angle": 0, + "content": "To the best of our knowledge, there are only two reasonably efficient solutions to the \\(4\\mathrm{p}3\\mathrm{v}\\) problem reported in the literature. The first solver [36] is based on a similar idea as our solvers, i.e., to first estimate the relative pose of two cameras and then register the third camera using a \\(\\mathsf{P}3\\mathsf{P}\\) solver [40]. To compute the pose of the first two cameras using information only from four point correspondences, the solver needs additional information about the position of one epipole. The paper shows that the four point correspondences between two calibrated views constrain the epipole in each image to lie on a curve of degree ten. Thus, the solver performs one-dimensional exhaustive search and sweeps a \\(10^{th}\\)-degree curve of possible epipoles. For each potential epipole, it computes the relative pose of two cameras, registers the third camera using three triangulated points, and finally extracts the solution minimizing the reprojection error of the fourth point in the third view. Evaluation of the solver on one potential epipole is fast. Yet, in contrast to our proposed \\(4\\mathrm{p}3\\mathrm{v}\\) (M)-based solvers, the error of the sampled epipole for one fixed point on the curve is not bounded since the true epipole can lie anywhere on the curve, and in many scenarios it is very far from the image center (e.g., outside the image). Thus, to obtain reasonably accurate and stable results, usually 1,000 candidates need to be evaluated. Even then, refinement at multiple local minima is required to improve the accuracy. The runtimes reported for this solver were \\(1 - 12ms\\) depending on the number of points searched. On a small number of synthetic experiments, the paper shows that the translation error returned by the proposed solver is usually \\(1 - 10\\mathrm{deg}\\) higher than the error returned by the \\(5\\mathrm{pt} + \\mathsf{P}3\\mathsf{P}\\) solver. The literature does not compare against [36] as there is no publicly available implementation and it is hard to re-implement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.485, + 0.902 + ], + "angle": 0, + "content": "The second efficient solver to the 4p3v problem was published only recently [21]. In this paper, the authors first transform the 4p3v problem into a minimal problem by considering a line passing through the last correspondence in the third view. The resulting system of equations is solved using an efficient Homotopy continuation (HC) method [13, 52]. To avoid computing large numbers of spurious solutions, an MLP-based classifier is trained. For a given problem \\( p \\), it selects one or several starting problem-solution pairs (so-called anchors), such that the geometrically meaningful/correct solution of \\( p \\) can be obtained by HC starting from this anchor. This strategy is fast, running \\( 16.3\\mu s \\) on average per solution. However, it has a high failure rate. The success rate of the 4p3v solver reported in [21] on two test datasets and data without noise is \\( 26.3\\% \\). [21] do not show results for a real scenario, i.e., a RANSAC-like framework with noisy data. Providing such an evaluation," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.885, + 0.107 + ], + "angle": 0, + "content": "we show that our much simpler solvers outperform [21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.905, + 0.303 + ], + "angle": 0, + "content": "Solutions to the 4p3v problem for orthographic and par-perspective views were presented in [33, 56]. In [33], the author suggested an iterative approach for updating to perspective views, but reported results only on a few synthetic instances. According to our extensive experiments, the update does not work on real data with general perspective cameras. This solver is returning large errors even after incorporating it into RANSAC with local optimization. We see two main reasons: (i) The set of inliers that satisfy an approximate para-perspective camera model in all three images is usually quite small, and (ii) the estimated model is often very far from the perspective optimum, and thus local optimization does not converge to a good solution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.905, + 0.485 + ], + "angle": 0, + "content": "In [11, 12, 24], the authors aim to classify and derive the number of solutions for different minimal configurations of points and lines in three calibrated views. Solutions to two minimal configurations combining points and lines were proposed in [13] and solved using a HC method [52]. Due to their complexity, the solvers are not practical. A GPU HC method was also used to solve minimal problems of four points/six lines in three views for a generalized camera in [10] and of four points in three cameras with an unknown shared focal length (4p3vf) in [5, 7]. Efficient GPU implementations of the 4p3vf solvers run \\(16.7ms\\) to \\(154ms\\). These times are still too slow for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.904, + 0.741 + ], + "angle": 0, + "content": "An approximation of perspective cameras using affine ones was used in several papers on camera geometry estimation [8]. [37] and [20] propose iterative approaches to the absolute camera pose estimation problem, i.e., the PnP problem. Both methods first compute the pose for an affine camera (weak- or para-perspective). Then the error induced by the affine camera approximation is used to adjust the constraints on the pose and recompute it. These methods do not guarantee convergence to the perspective model solution. In [41], affine cameras are used to efficiently solve the relative pose problem of two uncalibrated cameras from two affine correspondences. These affine correspondences are transformed to six point or two ellipse correspondences and then used to estimate the affine fundamental matrix. The paper also proposes a RANSAC framework that uses local optimization [30] to estimate the full fundamental matrix using inliers from the approximate affine model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.742, + 0.905, + 0.832 + ], + "angle": 0, + "content": "Points that are sampled based on feature geometry to generate point correspondences from affine or scale-and orientation-invariant feature correspondences were also used in [15, 39, 42]. In contrast, our M-based solvers use only point correspondences, without associated feature geometry, to generate an additional point correspondence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.846, + 0.905, + 0.863 + ], + "angle": 0, + "content": "3. Estimating the relative pose of three cameras" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this section, we describe different solutions for estimating the relative pose of three calibrated cameras. We start" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21915" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "with a baseline solution for the minimal configuration of three points visible in all three cameras and two additional points visible in two of the three cameras (the \\((5,5,3)\\) configuration). Next, we present our novel solutions for the configuration of four points visible in all three cameras (the \\((4,4,4)\\) configuration). This configuration generates an over-constrained problem. In this case, we have one more constraint than DoF. A minimal solution would need to drop one constraint, e.g., by considering only a line passing through one of the points in the third view [21] or by considering a \"half\" point correspondence. Since, in practice, we always have full correspondences and sampling one less point in one view leads to an under-constrained problem, the \\((4,4,4)\\) configuration, is usually considered \"minimal\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.482, + 0.399 + ], + "angle": 0, + "content": "5pt+P3P solver: The 5pt+P3P solver first estimates the relative pose of two cameras from 5 image point correspondences using the efficient 5pt solver [35]. Next, the three points visible in all three views are triangulated. Finally, the third camera is registered using the three 2D-3D point correspondences and the well-known efficient P3P solver [40]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.482, + 0.565 + ], + "angle": 0, + "content": "This straightforward solver, which is based on existing efficient solvers [35, 40], was discussed in several works [12, 35, 36, 46]. [36] showed that the \\(5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P}\\) solver performs better than their dedicated \\(4\\mathrm{p}3\\mathrm{v}\\) solver on synthetic data. Yet, the most recent works [13, 21] that study the three view relative pose problem do not discuss the \\(5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P}\\) solver and do not use it as a baseline for comparison. To the best of our knowledge, the performance of this solver on real data and within state-of-the-art RANSAC frameworks in the context of the \\(4\\mathrm{p}3\\mathrm{v}\\) problem has not been extensively studied. Our paper fills this gap in the literature." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.572, + 0.46, + 0.589 + ], + "angle": 0, + "content": "3.1. Approximate solutions to the 4p3v problem" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.836 + ], + "angle": 0, + "content": "In contrast to the (5,5,3) configuration, the (4,4,4) configuration, i.e., the configuration of four points in three views, leads to significantly more complex equations. State-of-the-art algebraic and numerical methods are known to fail in generating efficient and numerically stable solutions to these equations. Thus, solutions to the \\( 4\\mathrm{p}3\\mathrm{v} \\) problem require some approximations to be practical in real-world applications. This section introduces several practical approximate solutions to the \\( 4\\mathrm{p}3\\mathrm{v} \\) problem. Similarly to the \\( 5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P} \\) solver, we decompose the problem into the problem of first estimating the relative pose of two cameras and then registering the third camera with an efficient P3P solver [40]. The idea is to assume an approximate geometry only in the first two views and use four input point correspondences to efficiently estimate this geometry. Next we describe two groups of such solvers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "4p3v(A) solver: The first group of solvers approximates the geometry in the first two views using affine cameras. Although the approximation can be quite far from the correct geometry, [41] showed that an approximate affine camera" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "model provides good accuracy for relative pose estimation if coupled with geometry refitting (using the full 7DoF fundamental matrix) and local optimization inside RANSAC. The proposed \\(4\\mathrm{p}3\\mathrm{v}\\) (A) solver first uses four point correspondences in two views to efficiently estimate the affine fundamental matrix \\(\\mathbf{F}_{\\mathrm{A}}\\) [17] and then registers the third camera using three triangulated points and a P3P solver [40]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.203, + 0.905, + 0.536 + ], + "angle": 0, + "content": "4p3v(M) solver: The affine camera model used in the 4p3v (A)-based solvers can be quite imprecise. This motivates us to introduce solvers that are based on a less restrictive approximation. In this case, we approximate only one point correspondence in two views. Under the assumption of a para-perspective projection, i.e., of affine geometry, it is known that the mean point of three 3D points is projected to the mean points of their projections in both images [57]. Thus, we generate one new approximate correspondence in two views as the correspondence between the mean points of three corresponding points detected in these views. Let \\(\\mathbf{m}^l\\) be the mean point of three points \\(\\{\\mathbf{x}_i^l,\\mathbf{x}_j^l,\\mathbf{x}_k^l\\}\\), \\(i,j,k\\in \\{1,\\ldots ,4\\}\\), in the view \\(l\\in \\{1,2\\}\\), then \\(\\mathbf{m}^{1}\\leftrightarrow \\mathbf{m}^{2}\\) is our new approximate correspondence. The new 4p3v (M) solver solves the 4p3v problem by first estimating the full 5DoF essential matrix from four original correspondences \\(\\mathbf{x}_i^1\\leftrightarrow \\mathbf{x}_i^2\\), \\(i = 1,\\dots ,4\\) and the \\(5^{th}\\) correspondence \\(\\mathbf{m}^{1}\\leftrightarrow \\mathbf{m}^{2}\\). This is done using the efficient 5pt relative pose solver [35]. As with previous solvers, the third camera is registered using a P3P solver [40]. In short, the 4p3v (M) solver solves the 4p3v problem using the 5pt+P3P solver and one approximate correspondence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As we show on large amounts of synthetic and real data (see Sec. 4 and Supp. mat. (SM)), the \\(4\\mathrm{p}3\\mathrm{v}\\) (M) solver provides much more accurate estimates than the \\(4\\mathrm{p}3\\mathrm{v}\\) (A) solver. Moreover, even in the basic form (without the modifications presented in the next section), on many scenes it performs comparably to the state-of-the-art HC solver [21]. This can be attributed to several facts and observations: (1) The \\(\\mathbf{m}^1 \\leftrightarrow \\mathbf{m}^2\\) correspondence does not need to be seen as a correspondence of points that are projections of the mean point of three 3D points \\(\\mathbf{X}_i, \\mathbf{X}_j,\\) and \\(\\mathbf{X}_k\\) (in which case both \\(\\mathbf{m}^1\\) and \\(\\mathbf{m}^2\\) would have some error). We can look at this correspondence as a correspondence in which we fix a point in one view, e.g., \\(\\mathbf{m}^1\\), and generate a corresponding point in the second view. In this case, to generate a good correspondence, we only require the point in the second view to be reasonably close to the epipolar line defined by the mean point \\(\\mathbf{m}^1\\) in the first view, i.e., the 2D point does not need to correspond to one particular 3D point with a given depth. (2) The ray from the center of the first camera through the mean point \\(\\mathbf{m}^1\\) intersects the plane of \\(\\mathbf{X}_i, \\mathbf{X}_j,\\) and \\(\\mathbf{X}_k\\) at point M, which lies inside the triangle formed by these 3D points. The projection of M into the second camera lies inside the triangle formed by \\(\\{\\mathbf{x}_i^2, \\mathbf{x}_j^2, \\mathbf{x}_k^2\\}\\). By construction, this projection lies on the epipolar line \\(\\mathbf{E}\\mathbf{m}^1\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21916" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.429 + ], + "angle": 0, + "content": "Thus the epipolar line defined by \\(\\mathbf{m}^1\\) passes through the triangle defined by \\(\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}\\). Consequently, the maximum distance of \\(\\mathbf{m}^2\\) in the second image from the epipolar line is bounded by the maximum distance of \\(\\mathbf{m}^2\\) from \\(\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}\\). For a formal lemma and proof, see SM. (3) For practical applications, when used in RANSAC, it is not necessary that each triplet \\(\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}\\) generates a good correspondence \\(\\mathbf{m}^1\\leftrightarrow \\mathbf{m}^2\\). Samples with a high level of noise in the mean-point correspondence are filtered inside RANSAC. On a large number of different scenes, we observed that even if some image pairs have triplets of points that generate very noisy mean-point correspondences, there are usually enough triplets for which the noise in \\(\\mathbf{m}^2\\) is reasonably small to lead to good estimates. (4) Four point correspondences in two views usually fix the space of possible poses such that the \\(5^{th}\\) correspondence, even if noisy, often generates a pose that is not very far from the ground truth pose. Such a pose is usually sufficient for filtering out outliers and a good initialization for non-linear optimization on the original four points in three views and subsequent local optimization on detected inliers. We support our observations by experiments on a large amount of data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.437, + 0.421, + 0.453 + ], + "angle": 0, + "content": "3.2. Making approximate solvers practical" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.581 + ], + "angle": 0, + "content": "The \\(4\\mathrm{p}3\\mathrm{v}\\) (A)-based solvers, when used without any modifications, generally provide imprecise results even when used inside RANSAC with local optimization on three views. While the accuracy of the pure \\(4\\mathrm{p}3\\mathrm{v}\\) (M) solver inside LO-RANSAC [30] (RANSAC with Local Optimization) is much better, there is still room for improvement. Here we present several simple modifications of these solvers that significantly boost their performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.846 + ], + "angle": 0, + "content": "4p3v(M±δ) solver: The mean point correspondence used in the 4p3v (M) solver can provide a good approximation of a correct correspondence. Yet, as mentioned above, it can also be noisy. In the 4p3v (M±δ) solver, we thus, in addition to the mean point \\(\\mathbf{m}^2 = [x,y]\\) of three points \\(\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}\\) in the second image, generate two additional points. These points are (1) \\(\\mathbf{m}_{\\pm \\delta}^2 = [x\\pm \\delta ,y]\\) if the longest dimension of the triangle \\(\\mathcal{T}^2 = \\Delta \\left\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\right\\}\\) is in the x-direction or (2) \\(\\mathbf{m}_{\\pm \\delta}^2 = [x,y\\pm \\delta ]\\) if it is in the y-direction. All three points, i.e., \\(\\mathbf{m}^2,\\mathbf{m}_{-\\delta}^2\\) and \\(\\mathbf{m}_{+\\delta}^{2}\\) are placed in correspondence with the mean point \\(\\mathbf{m}^1\\). The 4p3v (M±δ) solver in the first step calls the 5pt solver [35] three times, with the \\(5^{th}\\) correspondence being either \\(\\mathbf{m}^1\\leftrightarrow \\mathbf{m}^2\\) \\(\\mathbf{m}^1\\leftrightarrow \\mathbf{m}_{-\\delta}^2\\) or \\(\\mathbf{m}^1\\leftrightarrow \\mathbf{m}_{+\\delta}^2\\). The results of these three 5pt solvers are collected to create hypotheses for the relative pose of the first two cameras inside RANSAC. The shift \\(\\delta\\) is selected relative to the size of the triangle \\(\\mathcal{T}^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.85, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Early non-minimal refitting (ENM): The geometry es" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.41 + ], + "angle": 0, + "content": "timated between the first two cameras using the proposed \\(4\\mathrm{p}3\\mathrm{v}\\) (A) and \\(4\\mathrm{p}3\\mathrm{v}\\) (M)- based solvers is only approximate. In standard LO-RANSAC, such a geometry is optimized in local optimization (LO) after registering the third view. However, this can lead to propagation of errors into triangulated points and subsequently into errors in the pose of the third camera. Inliers in three views w.r.t. such cameras together with imprecise pose initializations may not be sufficient for LO to converge to a good solution. We observed this especially for \\(4\\mathrm{p}3\\mathrm{v}\\) (A)-based solvers. Fortunately, the design of our solvers allows us to optimize the geometry already after estimating the approximate geometry between the first two cameras. Here, even a very imprecise geometry returned by the \\(4\\mathrm{p}3\\mathrm{v}\\) (A) solvers is usually sufficient to filter out outliers. Thus, after running the 4pt \\(\\mathrm{F_A / 5pt}\\) solver in the first step of the proposed solvers, we use the estimated approximate models to detect inliers in two views. We then refit the estimated geometry using the non-minimal version of the 5pt solver [35], which instead of a 4-dim null space of a \\(5\\times 9\\) matrix uses the last four vectors from the SVD/QR decomposition of a \\(n\\times 9\\) matrix." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.424, + 0.907, + 0.488 + ], + "angle": 0, + "content": "\\(4^{\\mathrm{th}}\\) point in the third view: \\(4\\mathrm{p}3\\mathrm{v}\\) (A) and \\(4\\mathrm{p}3\\mathrm{v}\\) (M) solvers actually solve the configuration \\((4,4,3)\\), i.e., they do not use the information from the point \\(\\mathbf{x}_4^3\\) in the third view. The information from \\(\\mathbf{x}_4^3\\) can be used in two different ways:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.909, + 0.709 + ], + "angle": 0, + "content": "Filtering \\((+F)\\): \\(\\mathbf{x}_4^3\\) can be used to filter out geometrically infeasible solutions returned by the proposed solvers. Note that the P3P and 5pt solvers used inside the proposed methods and the \\(\\delta\\)-based strategy return multiple solutions that can be evaluated w.r.t. \\(\\mathbf{x}_4^3\\) to filter out spurious solutions. Since the returned solutions can be affected by the proposed approximation, we do not simply select the solution with the smallest error on \\(\\mathbf{x}_4^3\\), but we keep all solutions that have an epipolar error on \\(\\mathbf{x}_4^3\\) smaller than twice the threshold used inside RANSAC. Our experiments show that this filtering can improve the speed of the proposed solvers. However, as a trade-off, there is sometimes a small drop in the accuracy of the solvers since, in some cases, geometrically correct solutions are filtered out." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Refinement \\((+\\mathbb{R})\\): \\(\\mathbf{x}_4^3\\) can be used to refine the solutions returned by the proposed solvers. We want solutions that minimize the epipolar error on the original 4 points in all 3 views, i.e., that solve the original (4,4,4) configuration. (4,4,4) is an overconstrained configuration; thus, for noisy data, there is, in general, no solution with zero error on all 4 points in 3 views. We refine the poses by minimizing the epipolar error of the original four points in three views using LM optimization, initialized using the solutions from the \\(4p3v(A)\\) - and \\(4p3v(M)\\) -based solvers. Experiments with different numbers of iterations show that two iterations are usually sufficient to obtain an improvement (see SM)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2This property was also used in the HC solver [21], which completely fails for many samples. These samples are filtered within RANSAC." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21917" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.477, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.234, + 0.483, + 0.276 + ], + "angle": 0, + "content": "Figure 2. Results of a synthetic experiment measuring the accuracy of two-view variants of our solvers depending on the angle between the principal axes of the cameras." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.301, + 0.223, + 0.318 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.432 + ], + "angle": 0, + "content": "We extensively evaluated the proposed solvers on a large variety of synthetic and real data to test their robustness to noise, outliers, and scene properties, and to assess their performance inside state-of-the-art RANSAC-frameworks [2, 27]. We compare our novel solvers with the homotopy continuation \\(4\\mathrm{p}3\\mathrm{v}\\) (HC) solver [21] and the \\(5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P}\\) baseline minimal solver for the \\((5,5,3)\\) configuration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.68 + ], + "angle": 0, + "content": "Experimental setup. To obtain feature correspondences, we use SuperPoint [9] features with the Light-Flue [32] matcher. We extract at most 2048 features per image. We perform matching for all three image pairs and keep only those matches that were consistently matched across all three views. We perform evaluation within two RANSAC frameworks: PoseLib [27] and GC-RANSAC [2]. For the 5pt solver, we use [35] and for the P3P solver, we use [40]. In PoseLib, we perform LO [6] using LM optimization. In GC-RANSAC, we perform LO using non-minimal solvers [18, 35] for fitting models to larger-than-minimal samples. We evaluated different shifts for our \\(\\delta\\)-based solvers using a validation scene (for the ablation study, see SM) and selected \\(\\delta = 0.08 * (\\text{longest triangle dim.})\\). The choice of \\(\\delta\\) is not critical, i.e., different values perform similarly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Evaluation measures. Inspired by [22], we define the pose error as max \\(\\left(0.5\\left(\\mathrm{R}_{err}^{12} + \\mathrm{R}_{err}^{13}\\right), 0.5\\left(\\mathrm{t}_{err}^{12} + \\mathrm{t}_{err}^{13}\\right)\\right)\\), where \\(\\mathbf{R}_{err}^{ij}\\) and \\(\\mathbf{t}_{err}^{ij}\\) are the angular errors of rotation and translation for pair \\(ij\\) in degrees [22]. We also report AUC values [22] at different thresholds for the pose error. We include results for an alternative pose error definition which includes \\(\\mathbf{R}_{err}^{23}\\) and \\(\\mathbf{t}_{err}^{23}\\) in SM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Approximate camera geometry. The first experiments aim to support our idea of estimating approximate geometry in the first two views. The accuracy of the used approximations depends on a large number of variables, including the depths of the points w.r.t. the cameras, the angle under which the points are observed, the type of motion, etc. A detailed analysis of all these factors, e.g., through synthetic" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.092, + 0.877, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.194, + 0.907, + 0.291 + ], + "angle": 0, + "content": "Figure 3. Distribution of the (left) rotation error (0.3373, 0.3349); and (right) percentage of inliers gathered (0.3266, 0.3434), as a function of the barycentric coordinates of the triangle in the second image w.r.t. the mean point of the corresponding triangle in the first image on 465k four-tuples of correspondences from St. Peter's Square scene from the PhotoTourism dataset [22]. We fit a 2D Gaussian distribution to the results and report the mean in brackets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.823, + 0.334 + ], + "angle": 0, + "content": "experiments, is beyond the scope of this paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.906, + 0.47 + ], + "angle": 0, + "content": "An approximation error introduced by the paraperspective projection, i.e., the affine geometry, and its effect on absolute camera pose determination is studied in the literature [20, 57]. Here, we thus only study the effect of the proposed approximation on relative pose estimation for real-world data and interesting synthetic scenarios. In the following experiments, all solvers are tested on two views without RANSAC. The two-view variants of our solvers are denoted as \\(4p(\\mathsf{A})\\), \\(4p(\\mathsf{M})\\), and \\(4p(\\mathsf{M} \\pm \\delta)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.906, + 0.713 + ], + "angle": 0, + "content": "In the first set of experiments, we consider several interesting synthetic setups. Fig. 2 presents the results for increasing the angle between the principal axes of the two cameras. We generate 3D points in a \\(2000 \\times 2000 \\times 100\\) cube uniformly at random. The distances of the camera centers are random between 1000 and 1200. The cameras are looking towards the scene and their principal axes form a specified angle. To simulate realistic scenarios, we add 1px noise to the image correspondences and \\(20\\%\\) outliers. It can be seen that the results for both approximate solvers deteriorate with increasing angle, with \\(4p(\\mathbb{A})\\) being significantly less accurate than \\(4p(\\mathbb{M})\\) and \\(4p(\\mathbb{M} \\pm \\delta)\\). However, ENM can significantly decrease errors, with the \\(4p(\\mathbb{M} \\pm \\delta) + ENM\\) solver returning almost identical results to \\(5pt + ENM\\). Results for varying distances of the cameras from the scene, depth of the scene, and image noise are in SM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.906, + 0.865 + ], + "angle": 0, + "content": "Similar observations as on synthetic data can be made on real-world data. Tab. 1 shows results on different scenes from the PhotoTourism dataset [22]. As can be seen, the approximate solvers are not as accurate as the \\(5\\mathrm{pt}\\) solver when used without ENM. The gap between the \\(4\\mathrm{p}(\\mathrm{M}) / 4\\mathrm{p}(\\mathrm{M}\\pm \\delta)\\) solvers and the \\(5\\mathrm{pt}\\) solver is for some scenes larger (Sacre Coeur, Trevi Fountain), and for some noticeably smaller (Reichstag, St. Peter's Square), showing that the performance of our solvers is scene-dependent. The \\(4\\mathrm{p}(\\mathrm{A})\\) solver returns large errors. The accuracy of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{3} \\) Outliers are only used in the ENM part to simulate that for approximate geometry outliers might be contained in the non-minimal samples." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21918" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.462, + 0.329 + ], + "angle": 0, + "content": "
MED (°)\\( {20}^{\\text{th }} \\) perc. (°)
w/o ENM
Scene5p (E)4p (M)4p (M±δ)4p (A)5p (E)4p (M)4p (M±δ)4p (A)
Brandenburg Gate16.1419.0215.6975.327.229.837.4645.99
Buckingham Palace18.8721.1217.7466.198.0410.018.2340.59
Coloseum Exterior17.4023.5018.9366.676.4411.368.3440.82
Grand Place Brussels18.4020.2816.7869.318.0410.098.0744.45
Notre Dame Front Facade15.8223.9719.6171.606.4012.279.0148.39
P Palace of Westminster16.0417.6214.7269.384.997.986.4046.49
Puntheon Exterior22.6525.7121.3161.3310.6514.0910.9438.12
Reichstag12.1512.9910.2782.855.156.164.5760.12
Sacre Coeur11.8117.6114.3174.213.697.955.8743.72
St. Peter's Square17.8518.7515.4872.758.519.717.6644.60
Taj Mahal9.9511.208.5682.743.774.903.6364.27
Temple Nara Japan18.5721.5317.0164.977.6010.638.1632.21
Trevi Fountain20.9327.7722.8950.128.1013.8910.7831.32
w/ ENM
Scene5p (E)4p (M)4p (M±δ)4p (A)5p (E)4p (M)4p (M±δ)4p (A)
Brandenburg Gate14.5317.0514.3828.406.748.866.9316.96
Buckingham Palace16.7118.7616.0735.077.358.967.4619.72
Colosseum Exterior16.2121.4617.7034.906.3510.718.0421.98
Grand Place Brussels16.2117.9615.2331.537.278.857.2418.40
Notre Dume Front Facade15.0522.0818.4634.496.3011.658.7122.08
Palace of Westminster13.8314.9512.7429.064.687.135.8416.41
Puntheon Exterior21.1924.0320.3037.2110.3213.2810.5524.42
Reichstag9.3210.038.3617.694.405.063.979.39
Sacre Coeur10.7115.1912.6424.923.617.235.4913.41
St. Peter's Square16.0116.9414.3528.607.758.637.0315.68
Taj Mahal8.329.327.4315.853.344.103.187.72
Temple Nara Japan16.2419.3815.7429.976.839.197.3216.65
Trevi Fountain20.2226.4222.0939.678.0213.4310.5326.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.34, + 0.477, + 0.354 + ], + "angle": 0, + "content": "Table 1. Accuracy of two-view solvers on PhotoTourism scenes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.384, + 0.483, + 0.597 + ], + "angle": 0, + "content": "the \\(4\\mathrm{p}(\\mathrm{M}) / 4\\mathrm{p}(\\mathrm{M}\\pm \\delta)\\) solvers is not very far from the 5pt solver, with the \\(4\\mathrm{p}(\\mathrm{M}\\pm \\delta)\\) solver even slightly outperforming the 5pt solver on some scenes (Reichstag, Taj Mahal). ENM helps to increase the precision of all solvers, especially \\(4\\mathrm{p}(\\mathrm{A})\\). Still, the \\(4\\mathrm{p}(\\mathrm{A}) + \\mathrm{ENM}\\) solver returns quite high errors. However, in practice, all solvers are used inside RANSAC, where the current best pose is locally optimized on inliers. LO compensates for less accurate pose estimates and, as we show on real experiments for three as well as two views (see SM), it can suppress the errors of the \\(4\\mathrm{p}(\\mathrm{A}) / 4\\mathrm{p}3\\mathrm{v}(\\mathrm{A})\\) -based solvers, making them comparable to the \\(4\\mathrm{p}(\\mathrm{M}) / 4\\mathrm{p}3\\mathrm{v}(\\mathrm{M})\\) -based solvers. For the three view scenario, the performance of all of our solvers can be further improved by pose refinement (+R) and filtering (+F)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.765 + ], + "angle": 0, + "content": "In the last set of experiments, we study the accuracy of the mean point correspondence. We sample 100 four-tuples of point correspondences consistent with the ground truth relative pose, i.e., inliers, for each image pair in scenes from the PhotoTourism dataset [22]. We use the first three correspondences to define the triangles in both images. Then, we establish correspondences between the mean of the triangle in one image and various points in the triangle in the second image. We express points in the second triangle via their barycentric coordinates and uniformly sample \\(19 \\times 19\\) barycentric coordinates \\((a, b) \\in [0, 1]^2\\), such that \\(a + b \\leq 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Fig. 3 shows the results for the rotation error and the percentage of inliers consistent with the pose obtained with the \\(4\\mathrm{p}\\) (M) solver on the St. Peter's Square scene. The optimum of the metrics is reached around the mean point of the triangles (the mean values of 2D Gaussians fitted to the results are very close to the mean point \\((0.3,0.3)\\) of the triangles). A detailed description, the translation and the symmetric epipolar errors, and the results for more scenes are in SM. For all tested scenes, we observed a similar behavior." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "Experiments on real data. We test the solvers on all scenes from the PhotoTourism dataset [22, 50] which provide ground truth poses and intrinsics via a COLMAP [49] reconstruction. In the results, we do not include the St. Peter's Square scene that we used for the validation of \\(\\delta\\) and the number of refinement iterations (see SM). We also include results for the Cambridge Landmarks dataset [23] (except the Street scene, which is commonly not used due to issues with its ground truth) and Aachen Day-Night v1.1 [58]. For PhotoTourism and Aachen, we use the images in their original resolution. For Cambridge Landmarks, we resize them so that the larger side is \\(800\\mathrm{px}\\). For each scene, we sample 5,000 random image triplets with at least 10 matches obtained with [9, 32] with at least \\(10\\%\\) overlap [22]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.905, + 0.397 + ], + "angle": 0, + "content": "Tab. 2 shows the results for PoseLib RANSAC with early termination and 5px epipolar threshold. Results for GC-RANSAC are in SM. With the suggested modifications (see Sec. 3.2) all proposed solvers outperform the state-of-the-art \\(4\\mathrm{p}3\\mathrm{v}\\) (HC) solver [21] in terms of pose accuracy with many of the variants also achieving faster runtimes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.905, + 0.73 + ], + "angle": 0, + "content": "Using filtering \\((+F)\\) improves the run-time of RANSAC at the cost of a decrease in pose accuracy. Still, our \\(4p3v(M) + R + F\\) solver outperforms \\(4p3v(HC)\\) in terms of both accuracy and run-time. The \\(4p3v(M \\pm \\delta)\\) and \\(4p3v(M \\pm \\delta) + R\\) solvers clearly improve upon the \\(4p3v(M)\\) solvers, albeit at an increased run-time. Still the \\(4p3v(M \\pm \\delta) + R + F\\) solver provides, in general, the best speed-accuracy trade-off. The results also show that early non-minimal refinement \\((+ENM)\\) is necessary for the \\(4p3v(A)\\) solver to perform well. For the other approximate solvers, this form of refinement also leads to improved accuracy at some cost to the runtime. Assuming that the \\(5pt + P3P\\) solver samples five points from which at least four have projections in all three cameras, the suggested pose refinement \\((+R)\\) and filtering \\((+F)\\) can also be applied to the baseline \\(5pt + P3P\\) by using the fourth (potentially the fifth) point in the third view. The suggested modifications improve the performance of the \\(5pt + P3P\\) solver, with the \\(5pt + P3P + R + F + ENM\\) solver having an accuracy similar to that of the \\(4p3v(M) + R + F + ENM\\) solver, however at slower run-times. Moreover, it is less accurate than the proposed \\(4p3v(M \\pm \\delta) + R + F + ENM\\) solver." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.884 + ], + "angle": 0, + "content": "We also investigate the speed-accuracy trade-off of the solvers by running PoseLib RANSAC [27] for a set of fixed numbers of iterations. Runtimes are reported for 1 core of a \\(2\\mathrm{GHz}\\) Intel Xeon Gold 6338 CPU. As shown in Fig. 4, the proposed \\(4\\mathrm{p}3\\mathrm{v}\\) \\((\\mathbb{M}\\pm \\delta) + \\mathbb{R} + \\mathbb{F}\\) solver consistently provides the best speed-accuracy trade-off both with and without ENM. \\(4\\mathrm{p}3\\mathrm{v}\\) (A) \\(+\\mathbb{R} + \\mathbb{F} + \\mathbb{ENM}\\) provides a similar performance, typically beating \\(4\\mathrm{p}3\\mathrm{v}\\) (HC) [21]. However, as shown in Fig. 4d for the St. Mary's Church scene [23], it may perform worse for some specific scenes." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "The SM provides detailed ablation studies on the signif" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21919" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.904, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.226, + 0.25, + 0.239 + ], + "angle": 0, + "content": "(a) Phototourism [22]" + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.226, + 0.48, + 0.239 + ], + "angle": 0, + "content": "(b) Cambridge Landmarks [23]" + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.226, + 0.686, + 0.239 + ], + "angle": 0, + "content": "(c) Aachen Day-Night v1.1 [58]" + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.226, + 0.877, + 0.239 + ], + "angle": 0, + "content": "(d) St. Mary's Church [23]" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.251, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Figure 4. Speed-accuracy trade-off for (a) all scenes from PhotoTourism [22] except St. Peter's Square, (b) 5 scenes from Cambridge Landmarks [23], (c) Aachen Day-Night v1.1 [58], and (d) St. Mary's Church scene from the Cambridge Landmarks dataset [23]. We report the AUC@10° of the pose error and vary the number of PoseLib RANSAC iterations (100, 200, 500, 1000, 2000, 5000, 10,000) with a 5 px epipolar threshold. Runtimes are averaged over all image triplets." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.32, + 0.907, + 0.507 + ], + "angle": 0, + "content": "
EstimatorPhotoTourism [22]Cambridge Landmarks [23]Aachen Day-Night v1.1 [58]
AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓
4p3v (HC) [21]6.371.6262.4173.5381.9464.108.732.6747.9964.5876.6560.1112.433.5542.2855.2066.5567.26
5pt+P3P5.401.6062.9374.3083.0033.777.462.6548.3665.3377.7524.0410.523.3843.4356.7368.2853.34
5pt+P3P+ENM5.231.5663.7775.0383.5048.797.202.6348.7465.6878.0434.8210.593.4143.3256.6468.0671.61
5pt+P3P+R+F5.101.5064.9075.9084.1138.127.152.5849.4266.3578.5531.8210.373.3643.7357.1568.7154.34
5pt+P3P+R+F+ENM4.991.5065.1876.2084.3056.526.982.5849.3866.3178.5747.0710.203.3443.7557.1168.6376.86
4p3v (A)37.4825.1622.9128.5935.1616.5841.2321.7322.7531.4738.9813.3336.4019.5324.4131.7339.2432.04
4p3v (A)+ENM4.971.4865.4976.4784.5340.457.022.5949.3566.3278.5328.3510.353.3843.6256.9968.5562.44
4p3v (A)+R33.4519.9525.8632.2239.1616.3236.9811.5927.0336.7044.4812.4834.8418.2225.2732.7740.2929.56
4p3v (A)+R+F35.6523.4723.9929.9736.6411.1038.3414.4725.8935.1942.809.3536.7720.7623.9431.2338.5819.75
4p3v (A)+R+F+ENM4.981.4865.6276.5684.5532.376.992.5749.6266.5578.7223.4310.433.3843.5756.9568.5642.36
4p3v (M)6.021.7160.8672.5681.7335.188.362.7447.1564.0676.5925.1512.093.6241.7855.0666.6754.89
4p3v (M)+ENM5.281.5863.6174.9683.4648.837.212.6348.6965.5777.9534.9310.673.4343.3156.6168.0870.87
4p3v (M)+R5.491.5663.6674.8883.3741.527.692.6348.6965.6777.9930.7611.173.4643.2856.3967.8860.40
4p3v (M)+R+F5.481.5463.9975.1483.5130.887.752.6248.7765.7177.9822.7211.163.4843.0456.2867.7342.38
4p3v (M)+R+F+ENM5.011.5065.1876.1684.2844.516.992.5849.4366.3178.5632.4810.253.3943.7457.2168.8058.08
4p3v (M±δ)5.551.6861.8673.7382.7883.707.652.6748.1665.2577.7659.2311.133.4742.7456.1167.74125.02
4p3v (M±δ)+ENM4.991.5664.0275.5184.01125.666.882.6049.1166.1378.5189.0510.263.3543.6656.9968.55175.54
4p3v (M±δ)+R4.861.4765.9777.0085.01100.617.102.5649.7966.8979.0873.9410.373.3443.9357.2668.82138.53
4p3v (M±δ)+R+F4.921.4765.8576.8784.9071.737.192.5649.7166.7578.9552.8410.523.3643.7957.1768.7392.89
4p3v (M±δ)+R+F+ENM4.661.4666.1277.1485.15112.606.652.5549.8766.9579.2081.989.963.3244.0257.4268.97139.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.518, + 0.908, + 0.561 + ], + "angle": 0, + "content": "Table 2. Results for different solvers implemented in the PoseLib framework [27] on 12 scenes from PhotoTourism [22], 5 scenes from Cambridge Landmarks [23] and Aachen Day-Night v1.1 [58]. We mark the best and second best results. Runtimes are reported in ms for the whole RANSAC with early termination (0.9999 confidence, minimum 100 iterations) and epipolar threshold set to 5 px." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.452, + 0.602 + ], + "angle": 0, + "content": "icance of each individual modification \\((+R / + C / + ENM)\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.607, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Limitations. As discussed, the accuracy of the proposed approximate solvers is scene dependent. However, after using the proposed ENM refitting, the scene-dependency is negligible. Especially for the \\(4\\mathrm{p}3\\mathrm{v}\\) \\((\\mathbb{M}\\pm \\delta)\\) +ENM solver, we have not noticed performance drops for some specific scene geometries or camera configurations. While the two-view variants of our solvers \\(4\\mathrm{p}\\) \\((\\mathbb{M}) / 4\\mathrm{p}\\) \\((\\mathbb{M}\\pm \\delta)\\) provide very similar results to the 5pt solver in two views, in this case they do not outperform the 5pt solver. In the three-view scenario, the better performance is achieved thanks to the proposed pose refinement \\((+\\mathbf{R})\\) and filtering \\((+\\mathbf{F})\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.483, + 0.866 + ], + "angle": 0, + "content": "The proposed ENM refitting cannot be applied to the \\(4\\mathrm{p}3\\mathrm{v}\\) (HC) solver [21], since this solver estimates the pose of all three cameras together. It is theoretically applicable to the solver proposed in [36]. However, here ENM would have needed to be run on hundreds of candidate poses corresponding to tens-to-hundreds of sampled epipoles. This" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.907, + 0.693 + ], + "angle": 0, + "content": "is because for one fixed point on the curve of epipoles, the error of the sampled epipole is not bounded since the true epipole can lie anywhere on the curve, even outside the image. Thus, ENM might avoid local minima present in the second step of the original solver, but would significantly slow down the solver in the first step, making it much slower than our \\(4\\mathrm{p}3\\mathrm{v}\\) \\((\\mathbf{M}) + \\mathsf{ENM} / 4\\mathrm{p}3\\mathrm{v}\\) \\((\\mathbf{M}\\pm \\delta)\\) +ENM solvers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.634, + 0.724 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We consider the highly challenging problem of relative pose estimation of three calibrated cameras from four correspondences. We propose novel approaches that solve the problem by utilizing solvers based on approximate geometry. The best performing solver uses a simple, yet novel strategy by using mean coordinates of three input points and points in their vicinity as an approximate \\(5^{th}\\) correspondences. Extensive experiments show that our solvers achieve state-of-the-art performance on a large variety of real scenes. At the same time, our solvers are simple to implement, especially compared to the current state-of-the-art [21]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "This weakness also applies to [21], since the scene needs to be similar enough to the training scenes for the MLP-based classifier to work well." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21920" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.381 + ], + "angle": 0, + "content": "Acknowledgements. This work was funded by the Czech Science Foundation (GACR) JUNIOR STAR Grant No. 22-23183M (supporting C.T., Y.D., and Z.K.), the Grant Agency of the Czech Technical University in Prague grant no. SGS23/173/OHK3/3T/13 (supporting C.T.), the EU NextGenerationEU through the Recovery and Resilience Plan for Slovakia under the project \"InnovAite Slovakia, Illuminating Pathways for AI-Driven Breakthroughs\" No. 09I02-03-V01-00029, the TERAIS project - a Horizon-Widera-2021 program of the European Union under the Grant agreement number 101079338 (supporting V.K. and Z.B.H.), and the EU Horizon 2020 project RICAIP grant agreement No. 857306 (supporting T.S.). Part of the research results was obtained using the computational resources procured in the national project National competence centre for high performance computing project code: 311070AKF2, funded by European Regional Development Fund, EU Structural Funds Informatization of society, Operational Program Integrated Infrastructure." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.393, + 0.188, + 0.409 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.418, + 0.482, + 0.458 + ], + "angle": 0, + "content": "[1] Chris Aholt and Luke Oeding. The ideal of the trifocal variety. Mathematics of Computation, 83(289):2553-2574, 2014. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.461, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[2] D. Barath and J. Matas. Graph-Cut RANSAC. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6733-6741, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.504, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[3] Martin Bujnak, Zuzana Kukelova, and Tomas Pajdla. A general solution to the p4p problem for camera with unknown focal length. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[4] Robert Castle, Georg Klein, and David W. Murray. Videorelate localization in multiple maps for wearable augmented reality. In ISWC, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[5] Chiang-Heng Chien, Hongyi Fan, Ahmad Abdelfattah, Elias Tsigaridas, Stanimire Tomov, and Benjamin Kimia. Gpubased homotopy continuation for minimal problems in computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15765-15776, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[6] Ondrej Chum, Jiří Matas, and Josef Kittler. Locally optimized ransac. In Pattern Recognition, pages 236-243. Springer Berlin Heidelberg, 2003. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[7] Andrea Porfiri Dal Cin, Timothy Duff, Luca Magri, and Tomas Pajdla. Minimal perspective autocalibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5064-5073, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[8] Toby Collins and Adrien Bartoli. Infinitesimal plane-based pose estimation. International Journal of Computer Vision, 109:252-286, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 224-236, 2018. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.418, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[10] Yaqing Ding, Chiang-Heng Chien, Viktor Larsson, Karl Åström, and Benjamin Kimia. Minimal solutions to generalized three-view relative pose problem. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8156-8164, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[11] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. Plmp-point-line minimal problems in complete multiview visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1675-1684, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[12] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomás Pajdla. \\(\\mathrm{Pl}_1\\mathfrak{p}\\) - point-line minimal problems under partial visibility in three views. In ECCV (26), pages 175-192. Springer, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.906, + 0.386 + ], + "angle": 0, + "content": "[13] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H. Regan, David da Costa de Pinho, Elias P. Tsigaridas, Charles W. Wampler, Jonathan D. Hauenstein, Peter J. Giblin, Benjamin B. Kimia, Anton Leykin, and Tomás Pajdla. TRPLP - trifocal relative pose from lines at points. In CVPR, pages 12070-12080. Computer Vision Foundation / IEEE, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.906, + 0.444 + ], + "angle": 0, + "content": "[14] M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6): 381-395, 1981. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[15] Liran Goshen and Ilan Shimshoni. Balanced exploration and exploitation model search for efficient epipolar geometry estimation. In Computer Vision - ECCV 2006, pages 151-164, Berlin, Heidelberg, 2006. Springer Berlin Heidelberg. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.906, + 0.574 + ], + "angle": 0, + "content": "[16] Christian Häne, Lionel Heng, Gim Hee Lee, Friedrich Fraun-dorfer, Paul Furgale, Torsten Sattler, and Marc Pollefeys. 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image and Vision Computing, 68:14-27, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[17] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision, 2006. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[18] Joel A. Hesch and Stergios I. Roumeliotis. A direct least-squares (dls) method for pnp. In 2011 International Conference on Computer Vision, pages 383-390, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[19] R.J. Holt and A.N. Netravali. Uniqueness of solutions to three perspective views of four points. IEEE Transactions on Pattern Analysis and Machine Intelligence, 17(3):303-307, 1995. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[20] Radu Horaud, Fadi Dornaika, and Bart Lamiroy. Object pose: The link between weak perspective, parapspective, and full perspective. Int. J. Comput. Vision, 22(2):173-189, 1997. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.908, + 0.829 + ], + "angle": 0, + "content": "[21] Petr Hruby, Timothy Duff, Anton Leykin, and Tomas Pajdla. Learning to solve hard minimal problems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5532-5542, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[22] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21921" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[23] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Proceedings of the IEEE international conference on computer vision, pages 2938-2946, 2015. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.207 + ], + "angle": 0, + "content": "[24] J. Kileel. Minimal problems for the calibrated trifocal variety. SIAM Journal on Applied Algebra and Geometry, 1(1): 575-598, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.484, + 0.263 + ], + "angle": 0, + "content": "[25] Z. Kukelova and T. Pajdla. A minimal solution to the autocalibration of radial distortion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2007), 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.266, + 0.484, + 0.335 + ], + "angle": 0, + "content": "[26] Zuzana Kukelova, Martin Bujnak, and Tomas Pajdla. Real-time solution to the absolute pose problem with unknown radial distortion and focal length. In Proceedings of the IEEE International Conference on Computer Vision, pages 2816-2823, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.337, + 0.484, + 0.365 + ], + "angle": 0, + "content": "[27] Viktor Larsson. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.368, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[28] Viktor Larsson, Kalle Åström, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Computer Vision and Pattern Recognition (CVPR), pages 820-829, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.484, + 0.482 + ], + "angle": 0, + "content": "[29] Viktor Larsson, Torsten Sattler, Zuzana Kukelova, and Marc Pollefeys. Revisiting radial distortion absolute pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1062-1071, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.484, + 0.539 + ], + "angle": 0, + "content": "[30] Karel Lebeda, Jiri Matas, and Ondrej Chum. Fixing the locally optimized ransac. In Proceedings of the British Machine Vision Conference 2012, pages 4961-4974. British Machine Vision Association, 2012. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.484, + 0.597 + ], + "angle": 0, + "content": "[31] S. Leonardos, R. Tron, and K. Daniilidis. A metric parametrization for trifocal tensors with non-colinear pinholes. In Computer Vision and Pattern Recognition (CVPR), pages 259-267. IEEE, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.484, + 0.654 + ], + "angle": 0, + "content": "[32] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In International Conference on Computer Vision (ICCV), pages 17627–17638, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.656, + 0.484, + 0.711 + ], + "angle": 0, + "content": "[33] H. Christopher Longuet-Higgins. A method of obtaining the relative positions of 4 points from 3 perspective projections. In BMVC91, pages 86-94, London, 1991. Springer London. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.714, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[34] Evgeniy V. Martyushev. On some properties of calibrated trifocal tensors. Journal of Mathematical Imaging and Vision, 58(2):321-332, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[35] D. Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[36] D. Nister and F. Schaffalitzky. Four points in two or three calibrated views: Theory and practice. International Journal of Computer Vision, 67(2):211-231, 2006. 1, 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[37] Denis Oberkampf, Daniel F DeMenton, and Larry S Davis. Iterative pose estimation using coplanar feature points. Computer Vision and Image Understanding, 63(3):495-511, 1996. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[38] Magnus Oskarsson, Andrew Zisserman, and Kalle Åström. Minimal projective reconstruction for combinations of points and lines in three views. Image and Vision Computing, 22 (10):777-785, 2004. British Machine Vision Computing 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.908, + 0.219 + ], + "angle": 0, + "content": "[39] Michal Perdoch, Jiri Matas, and Ondrej Chum. Epipolar geometry from two correspondences. In 18th International Conference on Pattern Recognition (ICPR'06), pages 215-219. IEEE, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.277 + ], + "angle": 0, + "content": "[40] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (p3p) solver. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318-332, 2018. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[41] James Pritts, Ondrej Chum, and Jiri Matas. Approximate models for fast and accurate epipolar geometry estimation. In 28th International Conference on Image and Vision Computing New Zealand, IVCNZ 2013, Wellington, New Zealand, November 27-29, 2013, pages 106-111. IEEE, 2013. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[42] James Pritts, Zuzana Kukelova, Viktor Larsson, and Ondrej Chum. Radially-distorted conjugate translations. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 1993-2001. IEEE Computer Society, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[43] Long Quan. Invariants of six points and projective reconstruction from three uncalibrated images. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 17(1): 34-46, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.908, + 0.544 + ], + "angle": 0, + "content": "[44] L. Quan, B. Triggs, and B. Mourrain. Some results on minimal euclidean reconstruction from four points. Journal of Mathematical Imaging and Vision, 24(3):341-348, 2006. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.907, + 0.602 + ], + "angle": 0, + "content": "[45] R. Raguram, O. Chum, M. Pollefeys, J. Matas, and J.-M. Frahm. USAC: A universal framework for random sample consensus. IEEE Transactions on Pattern Recognition and Machine Intelligence, 35(8):2022-2038, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[46] Volker Rodehorst. Evaluation of the metric trifocal tensor for relative three-view orientation. In Digital Proceedings, International Conference on the Applications of Computer Science and Mathematics in Architecture and Civil Engineering: July 20 - 22 2015, Bauhaus-University Weimar, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[47] T. Sattler, B. Leibe, and L. Kobbelt. Efficient & Effective Prioritized Matching for Large-Scale Image-Based Localization. IEEE Transactions on Pattern Recognition and Machine Intelligence, 39(9):1744-1756, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.907, + 0.759 + ], + "angle": 0, + "content": "[48] D. Scaramuzza and F. Fraundorfer. Visual odometry [tutorial]. IEEE Robot. Automat. Mag., 18(4):80-92, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[49] Johannes L. Schonberger and Jan-Michael Frahm. Structure-From-Motion Revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104–4113, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.907, + 0.859 + ], + "angle": 0, + "content": "[50] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.903 + ], + "angle": 0, + "content": "[51] N. Snavely, S. M. Seitz, and R. Szeliski. Modeling the world from internet photo collections. International Journal Computer Vision, 80(2):189-210, 2008. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21922" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[52] Andrew J Sommese, Andrew J Sommese, and Charles W Wampler. The numerical solution of systems of polynomials arising in engineering and science. World Scientific Publishing Co. Pte. Ltd, Singapore, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[53] H. Stewenius, D. Nister, F. Kahl, and F. Schaffalitzky. A minimal solution for relative pose with unknown focal length. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2005), 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[54] H. Stewenius, C. Engels, and D. Nister. Recent developments on direct relative orientation. ISPRS J. of Photogrammetry and Remote Sensing, 60:284-294, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[55] P. H. S. Torr and A. Zisserman. Robust parameterization and computation of the trifocal tensor. Image and Vision Computing, 15:591-605, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[56] G. Xu and N. Sugimoto. A linear algorithm for motion from three weak perspective images using euler angles. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(1):54-57, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[57] Zhengyou Zhang. Weak Perspective Projection, pages 877-883. Springer US, Boston, MA, 2014. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[58] Zichao Zhang, Torsten Sattler, and Davide Scaramuzza. Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision, 129(4):821-844, 2021. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "21923" + } + ] +] \ No newline at end of file diff --git a/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/3fec933e-2dde-4be1-8ad1-28710156467f_origin.pdf b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/3fec933e-2dde-4be1-8ad1-28710156467f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bf1fb7a8b6375ea023d4375e4ea154878fa45071 --- /dev/null +++ b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/3fec933e-2dde-4be1-8ad1-28710156467f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a566be019f3b962f119f8ee8a3a4d2b54e987fd6b5fdee0f44c1296a280e6b8 +size 2177243 diff --git a/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/full.md b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/full.md new file mode 100644 index 0000000000000000000000000000000000000000..45ae91f06e3739795888fe2ff2450d94c494ec9c --- /dev/null +++ b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/full.md @@ -0,0 +1,236 @@ +# Practical solutions to the relative pose of three calibrated cameras + +Charalambos Tzamos $^{1,*}$ Viktor Kocur $^{2,*}$ Yaqing Ding $^{1}$ Daniel Barath $^{3}$ Zuzana Berger Haladová $^{2}$ Torsten Sattler $^{4}$ Zuzana Kukelova $^{1}$ + +$^{1}$ Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague $^{2}$ Faculty of Mathematics, Physics and Informatics, Comenius University in Bratislava $^{3}$ ETH Zürich (Zurich), HUN-REN SZTAKI (Budapest) + +$^{4}$ Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague + +# Abstract + +We study the challenging problem of estimating the relative pose of three calibrated cameras from four point correspondences. We propose novel efficient solutions to this problem that are based on the simple idea of using four correspondences to estimate an approximate geometry of the first two views. We model this geometry either as an affine or a fully perspective geometry estimated using one additional approximate correspondence. We generate such an approximate correspondence using a very simple and efficient strategy, where the new point is the mean point of three corresponding input points. The new solvers are efficient and easy to implement, since they are based on existing efficient minimal solvers, i.e., the 4-point affine fundamental matrix, the well-known 5-point relative pose solver, and the P3P solver. Extensive experiments on real data show that the proposed solvers, when properly coupled with local optimization, achieve state-of-the-art results, with the novel solver based on approximate mean-point correspondences being more robust and accurate than the affine-based solver. + +# 1. Introduction + +Camera geometry estimation is crucial in many computer vision applications, e.g., visual navigation [48], Structure-from-Motion [51], augmented reality [4], self-driving cars [16], and visual localization [47]. Due to noise and outliers in input correspondences, the predominant way for camera geometry estimation is to use a hypothesis-and-test framework, e.g., RANSAC [2, 6, 14, 45]. For RANSAC-like methods, using as few (ideally the minimal number of) correspondences as possible for estimation is important since the number of RANSAC iterations (and thus its run-time) grows exponentially with the number of correspondences required for the model estimation. + +![](images/4ceef2ab7f660eb7eb7af3c4704fe4d64a2870881275d38e82103d2b3d8579e7.jpg) +Figure 1. Visualization of the four-points-in-three-views (4p3v) problem and our solution based on using four correspondences to efficiently estimate an approximate geometry of the first two views and then register the third view using a P3P solver [40]. + +Minimal camera geometry problems often result in complex systems of polynomial equations. Efficient algebraic methods helped solve many previously unsolved problems [3, 25, 26, 29, 53, 54]. Still, they fail to generate efficient and/or numerically stable solutions for some configurations. In this paper, we study one such challenging problem: Estimating the relative pose of three calibrated cameras. This problem has received attention for a long time [1, 19, 31, 34, 44]. However, due to its complexity, it is still not considered fully solved. There are no efficient and practical solutions for most of the minimal configurations of point and/or line correspondences [24]. One such configuration that is particularly interesting is the notoriously difficult configuration of four points in three calibrated views [35, 44], known as the $4\mathrm{p}3\mathrm{v}$ problem. + +State-of-the-art algebraic and numerical methods are known to fail in generating efficient and numerically stable solutions to the 4p3v problem. The existing methods for solving this problem are only approximate [21, 36]. By solving only for one (or a few) solutions from the 272 solutions of the 4p3v problem [21], and by discretely sampling the space of potential solutions [36], the existing 4p3v methods can often fail, i.e., the returned solution can be, in general, arbitrarily far from the geometrically correct solution. To decrease the failure rate, both methods [21, 36] + +require a lot of tuning and are challenging to re-implement.1 + +In contrast to complex solutions to the configuration of four points in three views, there is a simple and efficient solution for the configuration where five correspondences are detected in the first two views and three of these points are also visible in the third view. In this case, the $5\mathrm{pt} + \mathrm{P3P}$ solver first estimates the relative pose of two cameras from five correspondences using an efficient 5pt solver [35], and then registers the third camera using a P3P solver [40]. + +Following the idea of first estimating the relative pose of two cameras, we propose two novel approaches for solving the 4p3v problem. Our solutions are based on the simple idea of using four correspondences to estimate an approximate geometry of the first two views and then registering the third camera using a P3P solver [40]. + +In the first approach, we approximate the geometry of the first two views using affine cameras. The approximation using affine cameras was shown to provide good accuracy for relative pose estimation if coupled with geometry refitting and local optimization inside RANSAC [41]. We denote the proposed affine-based 4p3v solver as 4p3v (A). + +In the second approach, we use a less restrictive approximation. In this case, we estimate the full perspective geometry, i.e., the full 5DoF essential matrix, using four input correspondences and one additional approximate correspondence in two views. We generate the approximate correspondence using the locations of three of the four input correspondences. Under the assumption of a paraperspective projection, i.e., of affine geometry, the mean point of three 3D points is projected to the mean points in both images [57]. Thus, the new correspondence is generated as the correspondence between the mean points of three corresponding points detected in two views. This approximate mean-point correspondence can also be seen as a correspondence under the $1^{st}$ -order approximation of the homography defined by the plane passing through the three corresponding 3D points. We denote the proposed mean point-based 4p3v solver as 4p3v (M). + +While the proposed $4\mathrm{p}3\mathrm{v}$ (M) solver returns significantly more accurate poses than the $4\mathrm{p}3\mathrm{v}$ (A) solver, both solvers may not be accurate enough if not properly treated inside RANSAC. Thus, in this paper, we propose several ways to improve the accuracy of the proposed approximate solvers. (i) To compensate for noise in the mean point correspondences, we introduce the $4\mathrm{p}3\mathrm{v}$ $(\mathbb{M}\pm \delta)$ solver that generates two additional correspondences in the vicinity of the mean point, and in the first step calls the 5pt solver [35] three times. (ii) To improve the approximate geometry estimated between the first two views, we refit this geometry using a non-minimal relative pose solver and inliers obtained from the approximate geometry (ENM). (iii) We use + +the fourth correspondence in the third view to filter out geometrically infeasible solutions $(+F)$ and to refine the solutions on four input correspondences in three views using just a few iterations of Levenberg-Marquardt (LM) refinement $(+R)$ . While conceptually simple and efficient, the novel solvers achieve state-of-the-art results on real data. The contributions of the paper are as follows: + +- We propose two groups of novel solutions for the well-known and challenging 4p3v problem: $4\mathrm{p}3\mathrm{v}(\mathrm{A})$ - and $4\mathrm{p}3\mathrm{v}(\mathrm{M})$ -based solvers. These solutions are based on the simple idea of using four correspondences to estimate an approximate geometry in the first two views. Compared to state-of-the-art 4p3v solvers [21, 36], which are non-trivial and difficult to re-implement such that they are numerically stable and fast, our new solutions can be easily implemented using existing efficient implementations of the linear 4pt affine fundamental matrix solver [17], the 5pt solver [35] and the P3P solver [40]. The source code is available at https://github.com/kocurvik/threeview. +- We present several ways of improving the accuracy and the speed of the proposed solvers, as well as a strategy for efficiently using these approximate solvers inside a RANSAC-style paradigm. We show that the new solvers achieve state-of-the-art results in terms of accuracy on real data. While both $4\mathrm{p}3\mathrm{v}$ (A)- and $4\mathrm{p}3\mathrm{v}$ (M)-based solvers achieve comparable results when coupled with the suggested non-minimal geometry refitting and local optimization, our $4\mathrm{p}3\mathrm{v}$ (M)-based solvers are more robust to the scene geometry and RANSAC inlier thresholds. +- To our knowledge, we are the first to extensively evaluate solutions to the 4p3v problem on a large variety of real-world scenes and within state-of-the-art RANSAC frameworks, and to compare them to the baseline 5pt+P3P solver. We report results on 3 datasets, consisting of 18 different scenes and altogether 90,000 camera triplets. + +# 2. Related work + +Estimating the relative pose of three cameras from a minimal number of point and line correspondences is known as an extremely challenging problem. For three uncalibrated cameras, 6 point correspondences are necessary to estimate the trifocal tensor, with a solution known for a long time [43, 55]. Solutions to three minimal combinations of points and lines are presented in [38]. The configuration of 9 lines is more challenging and was solved only recently [28]. Yet, the solver is far from practical and runs 17.8s. + +For calibrated cameras, the configuration that attracts most of the attention is the configuration of four points in three views (the 4p3v problem). Note that this is not a minimal configuration since it generates 12 constraints for 11 degrees-of-freedom (DoF). The 4p3v problem is known to be extremely difficult to solve. Several papers present + +mostly theoretical results [1, 31, 34]. For four triplets of exact points without noise, it is shown that the 4p3v problem has, in general, a unique solution [19, 44]. + +To the best of our knowledge, there are only two reasonably efficient solutions to the $4\mathrm{p}3\mathrm{v}$ problem reported in the literature. The first solver [36] is based on a similar idea as our solvers, i.e., to first estimate the relative pose of two cameras and then register the third camera using a $\mathsf{P}3\mathsf{P}$ solver [40]. To compute the pose of the first two cameras using information only from four point correspondences, the solver needs additional information about the position of one epipole. The paper shows that the four point correspondences between two calibrated views constrain the epipole in each image to lie on a curve of degree ten. Thus, the solver performs one-dimensional exhaustive search and sweeps a $10^{th}$ -degree curve of possible epipoles. For each potential epipole, it computes the relative pose of two cameras, registers the third camera using three triangulated points, and finally extracts the solution minimizing the reprojection error of the fourth point in the third view. Evaluation of the solver on one potential epipole is fast. Yet, in contrast to our proposed $4\mathrm{p}3\mathrm{v}$ (M)-based solvers, the error of the sampled epipole for one fixed point on the curve is not bounded since the true epipole can lie anywhere on the curve, and in many scenarios it is very far from the image center (e.g., outside the image). Thus, to obtain reasonably accurate and stable results, usually 1,000 candidates need to be evaluated. Even then, refinement at multiple local minima is required to improve the accuracy. The runtimes reported for this solver were $1 - 12ms$ depending on the number of points searched. On a small number of synthetic experiments, the paper shows that the translation error returned by the proposed solver is usually $1 - 10\mathrm{deg}$ higher than the error returned by the $5\mathrm{pt} + \mathsf{P}3\mathsf{P}$ solver. The literature does not compare against [36] as there is no publicly available implementation and it is hard to re-implement. + +The second efficient solver to the 4p3v problem was published only recently [21]. In this paper, the authors first transform the 4p3v problem into a minimal problem by considering a line passing through the last correspondence in the third view. The resulting system of equations is solved using an efficient Homotopy continuation (HC) method [13, 52]. To avoid computing large numbers of spurious solutions, an MLP-based classifier is trained. For a given problem $p$ , it selects one or several starting problem-solution pairs (so-called anchors), such that the geometrically meaningful/correct solution of $p$ can be obtained by HC starting from this anchor. This strategy is fast, running $16.3\mu s$ on average per solution. However, it has a high failure rate. The success rate of the 4p3v solver reported in [21] on two test datasets and data without noise is $26.3\%$ . [21] do not show results for a real scenario, i.e., a RANSAC-like framework with noisy data. Providing such an evaluation, + +we show that our much simpler solvers outperform [21]. + +Solutions to the 4p3v problem for orthographic and par-perspective views were presented in [33, 56]. In [33], the author suggested an iterative approach for updating to perspective views, but reported results only on a few synthetic instances. According to our extensive experiments, the update does not work on real data with general perspective cameras. This solver is returning large errors even after incorporating it into RANSAC with local optimization. We see two main reasons: (i) The set of inliers that satisfy an approximate para-perspective camera model in all three images is usually quite small, and (ii) the estimated model is often very far from the perspective optimum, and thus local optimization does not converge to a good solution. + +In [11, 12, 24], the authors aim to classify and derive the number of solutions for different minimal configurations of points and lines in three calibrated views. Solutions to two minimal configurations combining points and lines were proposed in [13] and solved using a HC method [52]. Due to their complexity, the solvers are not practical. A GPU HC method was also used to solve minimal problems of four points/six lines in three views for a generalized camera in [10] and of four points in three cameras with an unknown shared focal length (4p3vf) in [5, 7]. Efficient GPU implementations of the 4p3vf solvers run $16.7ms$ to $154ms$ . These times are still too slow for practical applications. + +An approximation of perspective cameras using affine ones was used in several papers on camera geometry estimation [8]. [37] and [20] propose iterative approaches to the absolute camera pose estimation problem, i.e., the PnP problem. Both methods first compute the pose for an affine camera (weak- or para-perspective). Then the error induced by the affine camera approximation is used to adjust the constraints on the pose and recompute it. These methods do not guarantee convergence to the perspective model solution. In [41], affine cameras are used to efficiently solve the relative pose problem of two uncalibrated cameras from two affine correspondences. These affine correspondences are transformed to six point or two ellipse correspondences and then used to estimate the affine fundamental matrix. The paper also proposes a RANSAC framework that uses local optimization [30] to estimate the full fundamental matrix using inliers from the approximate affine model. + +Points that are sampled based on feature geometry to generate point correspondences from affine or scale-and orientation-invariant feature correspondences were also used in [15, 39, 42]. In contrast, our M-based solvers use only point correspondences, without associated feature geometry, to generate an additional point correspondence. + +# 3. Estimating the relative pose of three cameras + +In this section, we describe different solutions for estimating the relative pose of three calibrated cameras. We start + +with a baseline solution for the minimal configuration of three points visible in all three cameras and two additional points visible in two of the three cameras (the $(5,5,3)$ configuration). Next, we present our novel solutions for the configuration of four points visible in all three cameras (the $(4,4,4)$ configuration). This configuration generates an over-constrained problem. In this case, we have one more constraint than DoF. A minimal solution would need to drop one constraint, e.g., by considering only a line passing through one of the points in the third view [21] or by considering a "half" point correspondence. Since, in practice, we always have full correspondences and sampling one less point in one view leads to an under-constrained problem, the $(4,4,4)$ configuration, is usually considered "minimal". + +5pt+P3P solver: The 5pt+P3P solver first estimates the relative pose of two cameras from 5 image point correspondences using the efficient 5pt solver [35]. Next, the three points visible in all three views are triangulated. Finally, the third camera is registered using the three 2D-3D point correspondences and the well-known efficient P3P solver [40]. + +This straightforward solver, which is based on existing efficient solvers [35, 40], was discussed in several works [12, 35, 36, 46]. [36] showed that the $5\mathrm{pt} + \mathrm{P}3\mathrm{P}$ solver performs better than their dedicated $4\mathrm{p}3\mathrm{v}$ solver on synthetic data. Yet, the most recent works [13, 21] that study the three view relative pose problem do not discuss the $5\mathrm{pt} + \mathrm{P}3\mathrm{P}$ solver and do not use it as a baseline for comparison. To the best of our knowledge, the performance of this solver on real data and within state-of-the-art RANSAC frameworks in the context of the $4\mathrm{p}3\mathrm{v}$ problem has not been extensively studied. Our paper fills this gap in the literature. + +# 3.1. Approximate solutions to the 4p3v problem + +In contrast to the (5,5,3) configuration, the (4,4,4) configuration, i.e., the configuration of four points in three views, leads to significantly more complex equations. State-of-the-art algebraic and numerical methods are known to fail in generating efficient and numerically stable solutions to these equations. Thus, solutions to the $4\mathrm{p}3\mathrm{v}$ problem require some approximations to be practical in real-world applications. This section introduces several practical approximate solutions to the $4\mathrm{p}3\mathrm{v}$ problem. Similarly to the $5\mathrm{pt} + \mathrm{P}3\mathrm{P}$ solver, we decompose the problem into the problem of first estimating the relative pose of two cameras and then registering the third camera with an efficient P3P solver [40]. The idea is to assume an approximate geometry only in the first two views and use four input point correspondences to efficiently estimate this geometry. Next we describe two groups of such solvers. + +4p3v(A) solver: The first group of solvers approximates the geometry in the first two views using affine cameras. Although the approximation can be quite far from the correct geometry, [41] showed that an approximate affine camera + +model provides good accuracy for relative pose estimation if coupled with geometry refitting (using the full 7DoF fundamental matrix) and local optimization inside RANSAC. The proposed $4\mathrm{p}3\mathrm{v}$ (A) solver first uses four point correspondences in two views to efficiently estimate the affine fundamental matrix $\mathbf{F}_{\mathrm{A}}$ [17] and then registers the third camera using three triangulated points and a P3P solver [40]. + +4p3v(M) solver: The affine camera model used in the 4p3v (A)-based solvers can be quite imprecise. This motivates us to introduce solvers that are based on a less restrictive approximation. In this case, we approximate only one point correspondence in two views. Under the assumption of a para-perspective projection, i.e., of affine geometry, it is known that the mean point of three 3D points is projected to the mean points of their projections in both images [57]. Thus, we generate one new approximate correspondence in two views as the correspondence between the mean points of three corresponding points detected in these views. Let $\mathbf{m}^l$ be the mean point of three points $\{\mathbf{x}_i^l,\mathbf{x}_j^l,\mathbf{x}_k^l\}$ , $i,j,k\in \{1,\ldots ,4\}$ , in the view $l\in \{1,2\}$ , then $\mathbf{m}^{1}\leftrightarrow \mathbf{m}^{2}$ is our new approximate correspondence. The new 4p3v (M) solver solves the 4p3v problem by first estimating the full 5DoF essential matrix from four original correspondences $\mathbf{x}_i^1\leftrightarrow \mathbf{x}_i^2$ , $i = 1,\dots ,4$ and the $5^{th}$ correspondence $\mathbf{m}^{1}\leftrightarrow \mathbf{m}^{2}$ . This is done using the efficient 5pt relative pose solver [35]. As with previous solvers, the third camera is registered using a P3P solver [40]. In short, the 4p3v (M) solver solves the 4p3v problem using the 5pt+P3P solver and one approximate correspondence. + +As we show on large amounts of synthetic and real data (see Sec. 4 and Supp. mat. (SM)), the $4\mathrm{p}3\mathrm{v}$ (M) solver provides much more accurate estimates than the $4\mathrm{p}3\mathrm{v}$ (A) solver. Moreover, even in the basic form (without the modifications presented in the next section), on many scenes it performs comparably to the state-of-the-art HC solver [21]. This can be attributed to several facts and observations: (1) The $\mathbf{m}^1 \leftrightarrow \mathbf{m}^2$ correspondence does not need to be seen as a correspondence of points that are projections of the mean point of three 3D points $\mathbf{X}_i, \mathbf{X}_j,$ and $\mathbf{X}_k$ (in which case both $\mathbf{m}^1$ and $\mathbf{m}^2$ would have some error). We can look at this correspondence as a correspondence in which we fix a point in one view, e.g., $\mathbf{m}^1$ , and generate a corresponding point in the second view. In this case, to generate a good correspondence, we only require the point in the second view to be reasonably close to the epipolar line defined by the mean point $\mathbf{m}^1$ in the first view, i.e., the 2D point does not need to correspond to one particular 3D point with a given depth. (2) The ray from the center of the first camera through the mean point $\mathbf{m}^1$ intersects the plane of $\mathbf{X}_i, \mathbf{X}_j,$ and $\mathbf{X}_k$ at point M, which lies inside the triangle formed by these 3D points. The projection of M into the second camera lies inside the triangle formed by $\{\mathbf{x}_i^2, \mathbf{x}_j^2, \mathbf{x}_k^2\}$ . By construction, this projection lies on the epipolar line $\mathbf{E}\mathbf{m}^1$ . + +Thus the epipolar line defined by $\mathbf{m}^1$ passes through the triangle defined by $\{\mathbf{x}_i^2,\mathbf{x}_j^2,\mathbf{x}_k^2\}$ . Consequently, the maximum distance of $\mathbf{m}^2$ in the second image from the epipolar line is bounded by the maximum distance of $\mathbf{m}^2$ from $\{\mathbf{x}_i^2,\mathbf{x}_j^2,\mathbf{x}_k^2\}$ . For a formal lemma and proof, see SM. (3) For practical applications, when used in RANSAC, it is not necessary that each triplet $\{\mathbf{x}_i^2,\mathbf{x}_j^2,\mathbf{x}_k^2\}$ generates a good correspondence $\mathbf{m}^1\leftrightarrow \mathbf{m}^2$ . Samples with a high level of noise in the mean-point correspondence are filtered inside RANSAC. On a large number of different scenes, we observed that even if some image pairs have triplets of points that generate very noisy mean-point correspondences, there are usually enough triplets for which the noise in $\mathbf{m}^2$ is reasonably small to lead to good estimates. (4) Four point correspondences in two views usually fix the space of possible poses such that the $5^{th}$ correspondence, even if noisy, often generates a pose that is not very far from the ground truth pose. Such a pose is usually sufficient for filtering out outliers and a good initialization for non-linear optimization on the original four points in three views and subsequent local optimization on detected inliers. We support our observations by experiments on a large amount of data. + +# 3.2. Making approximate solvers practical + +The $4\mathrm{p}3\mathrm{v}$ (A)-based solvers, when used without any modifications, generally provide imprecise results even when used inside RANSAC with local optimization on three views. While the accuracy of the pure $4\mathrm{p}3\mathrm{v}$ (M) solver inside LO-RANSAC [30] (RANSAC with Local Optimization) is much better, there is still room for improvement. Here we present several simple modifications of these solvers that significantly boost their performance. + +4p3v(M±δ) solver: The mean point correspondence used in the 4p3v (M) solver can provide a good approximation of a correct correspondence. Yet, as mentioned above, it can also be noisy. In the 4p3v (M±δ) solver, we thus, in addition to the mean point $\mathbf{m}^2 = [x,y]$ of three points $\{\mathbf{x}_i^2,\mathbf{x}_j^2,\mathbf{x}_k^2\}$ in the second image, generate two additional points. These points are (1) $\mathbf{m}_{\pm \delta}^2 = [x\pm \delta ,y]$ if the longest dimension of the triangle $\mathcal{T}^2 = \Delta \left\{\mathbf{x}_i^2,\mathbf{x}_j^2,\mathbf{x}_k^2\right\}$ is in the x-direction or (2) $\mathbf{m}_{\pm \delta}^2 = [x,y\pm \delta ]$ if it is in the y-direction. All three points, i.e., $\mathbf{m}^2,\mathbf{m}_{-\delta}^2$ and $\mathbf{m}_{+\delta}^{2}$ are placed in correspondence with the mean point $\mathbf{m}^1$ . The 4p3v (M±δ) solver in the first step calls the 5pt solver [35] three times, with the $5^{th}$ correspondence being either $\mathbf{m}^1\leftrightarrow \mathbf{m}^2$ $\mathbf{m}^1\leftrightarrow \mathbf{m}_{-\delta}^2$ or $\mathbf{m}^1\leftrightarrow \mathbf{m}_{+\delta}^2$ . The results of these three 5pt solvers are collected to create hypotheses for the relative pose of the first two cameras inside RANSAC. The shift $\delta$ is selected relative to the size of the triangle $\mathcal{T}^2$ . + +Early non-minimal refitting (ENM): The geometry es + +timated between the first two cameras using the proposed $4\mathrm{p}3\mathrm{v}$ (A) and $4\mathrm{p}3\mathrm{v}$ (M)- based solvers is only approximate. In standard LO-RANSAC, such a geometry is optimized in local optimization (LO) after registering the third view. However, this can lead to propagation of errors into triangulated points and subsequently into errors in the pose of the third camera. Inliers in three views w.r.t. such cameras together with imprecise pose initializations may not be sufficient for LO to converge to a good solution. We observed this especially for $4\mathrm{p}3\mathrm{v}$ (A)-based solvers. Fortunately, the design of our solvers allows us to optimize the geometry already after estimating the approximate geometry between the first two cameras. Here, even a very imprecise geometry returned by the $4\mathrm{p}3\mathrm{v}$ (A) solvers is usually sufficient to filter out outliers. Thus, after running the 4pt $\mathrm{F_A / 5pt}$ solver in the first step of the proposed solvers, we use the estimated approximate models to detect inliers in two views. We then refit the estimated geometry using the non-minimal version of the 5pt solver [35], which instead of a 4-dim null space of a $5\times 9$ matrix uses the last four vectors from the SVD/QR decomposition of a $n\times 9$ matrix. + +$4^{\mathrm{th}}$ point in the third view: $4\mathrm{p}3\mathrm{v}$ (A) and $4\mathrm{p}3\mathrm{v}$ (M) solvers actually solve the configuration $(4,4,3)$ , i.e., they do not use the information from the point $\mathbf{x}_4^3$ in the third view. The information from $\mathbf{x}_4^3$ can be used in two different ways: + +Filtering $(+F)$ : $\mathbf{x}_4^3$ can be used to filter out geometrically infeasible solutions returned by the proposed solvers. Note that the P3P and 5pt solvers used inside the proposed methods and the $\delta$ -based strategy return multiple solutions that can be evaluated w.r.t. $\mathbf{x}_4^3$ to filter out spurious solutions. Since the returned solutions can be affected by the proposed approximation, we do not simply select the solution with the smallest error on $\mathbf{x}_4^3$ , but we keep all solutions that have an epipolar error on $\mathbf{x}_4^3$ smaller than twice the threshold used inside RANSAC. Our experiments show that this filtering can improve the speed of the proposed solvers. However, as a trade-off, there is sometimes a small drop in the accuracy of the solvers since, in some cases, geometrically correct solutions are filtered out. + +Refinement $(+\mathbb{R})$ : $\mathbf{x}_4^3$ can be used to refine the solutions returned by the proposed solvers. We want solutions that minimize the epipolar error on the original 4 points in all 3 views, i.e., that solve the original (4,4,4) configuration. (4,4,4) is an overconstrained configuration; thus, for noisy data, there is, in general, no solution with zero error on all 4 points in 3 views. We refine the poses by minimizing the epipolar error of the original four points in three views using LM optimization, initialized using the solutions from the $4p3v(A)$ - and $4p3v(M)$ -based solvers. Experiments with different numbers of iterations show that two iterations are usually sufficient to obtain an improvement (see SM). + +![](images/73c81f1cb63439e0eecb94e0a58fc9c6154737b6782db60d828d2448b4cb0da6.jpg) +Figure 2. Results of a synthetic experiment measuring the accuracy of two-view variants of our solvers depending on the angle between the principal axes of the cameras. + +# 4. Experiments + +We extensively evaluated the proposed solvers on a large variety of synthetic and real data to test their robustness to noise, outliers, and scene properties, and to assess their performance inside state-of-the-art RANSAC-frameworks [2, 27]. We compare our novel solvers with the homotopy continuation $4\mathrm{p}3\mathrm{v}$ (HC) solver [21] and the $5\mathrm{pt} + \mathrm{P}3\mathrm{P}$ baseline minimal solver for the $(5,5,3)$ configuration. + +Experimental setup. To obtain feature correspondences, we use SuperPoint [9] features with the Light-Flue [32] matcher. We extract at most 2048 features per image. We perform matching for all three image pairs and keep only those matches that were consistently matched across all three views. We perform evaluation within two RANSAC frameworks: PoseLib [27] and GC-RANSAC [2]. For the 5pt solver, we use [35] and for the P3P solver, we use [40]. In PoseLib, we perform LO [6] using LM optimization. In GC-RANSAC, we perform LO using non-minimal solvers [18, 35] for fitting models to larger-than-minimal samples. We evaluated different shifts for our $\delta$ -based solvers using a validation scene (for the ablation study, see SM) and selected $\delta = 0.08 * (\text{longest triangle dim.})$ . The choice of $\delta$ is not critical, i.e., different values perform similarly. + +Evaluation measures. Inspired by [22], we define the pose error as max $\left(0.5\left(\mathrm{R}_{err}^{12} + \mathrm{R}_{err}^{13}\right), 0.5\left(\mathrm{t}_{err}^{12} + \mathrm{t}_{err}^{13}\right)\right)$ , where $\mathbf{R}_{err}^{ij}$ and $\mathbf{t}_{err}^{ij}$ are the angular errors of rotation and translation for pair $ij$ in degrees [22]. We also report AUC values [22] at different thresholds for the pose error. We include results for an alternative pose error definition which includes $\mathbf{R}_{err}^{23}$ and $\mathbf{t}_{err}^{23}$ in SM. + +Approximate camera geometry. The first experiments aim to support our idea of estimating approximate geometry in the first two views. The accuracy of the used approximations depends on a large number of variables, including the depths of the points w.r.t. the cameras, the angle under which the points are observed, the type of motion, etc. A detailed analysis of all these factors, e.g., through synthetic + +![](images/866947a811cca479d5cd44a944d4ec17d3e87a71a804da5a42ff398c48a3eb69.jpg) +Figure 3. Distribution of the (left) rotation error (0.3373, 0.3349); and (right) percentage of inliers gathered (0.3266, 0.3434), as a function of the barycentric coordinates of the triangle in the second image w.r.t. the mean point of the corresponding triangle in the first image on 465k four-tuples of correspondences from St. Peter's Square scene from the PhotoTourism dataset [22]. We fit a 2D Gaussian distribution to the results and report the mean in brackets. + +experiments, is beyond the scope of this paper. + +An approximation error introduced by the paraperspective projection, i.e., the affine geometry, and its effect on absolute camera pose determination is studied in the literature [20, 57]. Here, we thus only study the effect of the proposed approximation on relative pose estimation for real-world data and interesting synthetic scenarios. In the following experiments, all solvers are tested on two views without RANSAC. The two-view variants of our solvers are denoted as $4p(\mathsf{A})$ , $4p(\mathsf{M})$ , and $4p(\mathsf{M} \pm \delta)$ . + +In the first set of experiments, we consider several interesting synthetic setups. Fig. 2 presents the results for increasing the angle between the principal axes of the two cameras. We generate 3D points in a $2000 \times 2000 \times 100$ cube uniformly at random. The distances of the camera centers are random between 1000 and 1200. The cameras are looking towards the scene and their principal axes form a specified angle. To simulate realistic scenarios, we add 1px noise to the image correspondences and $20\%$ outliers. It can be seen that the results for both approximate solvers deteriorate with increasing angle, with $4p(\mathbb{A})$ being significantly less accurate than $4p(\mathbb{M})$ and $4p(\mathbb{M} \pm \delta)$ . However, ENM can significantly decrease errors, with the $4p(\mathbb{M} \pm \delta) + ENM$ solver returning almost identical results to $5pt + ENM$ . Results for varying distances of the cameras from the scene, depth of the scene, and image noise are in SM. + +Similar observations as on synthetic data can be made on real-world data. Tab. 1 shows results on different scenes from the PhotoTourism dataset [22]. As can be seen, the approximate solvers are not as accurate as the $5\mathrm{pt}$ solver when used without ENM. The gap between the $4\mathrm{p}(\mathrm{M}) / 4\mathrm{p}(\mathrm{M}\pm \delta)$ solvers and the $5\mathrm{pt}$ solver is for some scenes larger (Sacre Coeur, Trevi Fountain), and for some noticeably smaller (Reichstag, St. Peter's Square), showing that the performance of our solvers is scene-dependent. The $4\mathrm{p}(\mathrm{A})$ solver returns large errors. The accuracy of + +
MED (°)\( {20}^{\text{th }} \) perc. (°)
w/o ENM
Scene5p (E)4p (M)4p (M±δ)4p (A)5p (E)4p (M)4p (M±δ)4p (A)
Brandenburg Gate16.1419.0215.6975.327.229.837.4645.99
Buckingham Palace18.8721.1217.7466.198.0410.018.2340.59
Coloseum Exterior17.4023.5018.9366.676.4411.368.3440.82
Grand Place Brussels18.4020.2816.7869.318.0410.098.0744.45
Notre Dame Front Facade15.8223.9719.6171.606.4012.279.0148.39
P Palace of Westminster16.0417.6214.7269.384.997.986.4046.49
Puntheon Exterior22.6525.7121.3161.3310.6514.0910.9438.12
Reichstag12.1512.9910.2782.855.156.164.5760.12
Sacre Coeur11.8117.6114.3174.213.697.955.8743.72
St. Peter's Square17.8518.7515.4872.758.519.717.6644.60
Taj Mahal9.9511.208.5682.743.774.903.6364.27
Temple Nara Japan18.5721.5317.0164.977.6010.638.1632.21
Trevi Fountain20.9327.7722.8950.128.1013.8910.7831.32
w/ ENM
Scene5p (E)4p (M)4p (M±δ)4p (A)5p (E)4p (M)4p (M±δ)4p (A)
Brandenburg Gate14.5317.0514.3828.406.748.866.9316.96
Buckingham Palace16.7118.7616.0735.077.358.967.4619.72
Colosseum Exterior16.2121.4617.7034.906.3510.718.0421.98
Grand Place Brussels16.2117.9615.2331.537.278.857.2418.40
Notre Dume Front Facade15.0522.0818.4634.496.3011.658.7122.08
Palace of Westminster13.8314.9512.7429.064.687.135.8416.41
Puntheon Exterior21.1924.0320.3037.2110.3213.2810.5524.42
Reichstag9.3210.038.3617.694.405.063.979.39
Sacre Coeur10.7115.1912.6424.923.617.235.4913.41
St. Peter's Square16.0116.9414.3528.607.758.637.0315.68
Taj Mahal8.329.327.4315.853.344.103.187.72
Temple Nara Japan16.2419.3815.7429.976.839.197.3216.65
Trevi Fountain20.2226.4222.0939.678.0213.4310.5326.62
+ +Table 1. Accuracy of two-view solvers on PhotoTourism scenes. + +the $4\mathrm{p}(\mathrm{M}) / 4\mathrm{p}(\mathrm{M}\pm \delta)$ solvers is not very far from the 5pt solver, with the $4\mathrm{p}(\mathrm{M}\pm \delta)$ solver even slightly outperforming the 5pt solver on some scenes (Reichstag, Taj Mahal). ENM helps to increase the precision of all solvers, especially $4\mathrm{p}(\mathrm{A})$ . Still, the $4\mathrm{p}(\mathrm{A}) + \mathrm{ENM}$ solver returns quite high errors. However, in practice, all solvers are used inside RANSAC, where the current best pose is locally optimized on inliers. LO compensates for less accurate pose estimates and, as we show on real experiments for three as well as two views (see SM), it can suppress the errors of the $4\mathrm{p}(\mathrm{A}) / 4\mathrm{p}3\mathrm{v}(\mathrm{A})$ -based solvers, making them comparable to the $4\mathrm{p}(\mathrm{M}) / 4\mathrm{p}3\mathrm{v}(\mathrm{M})$ -based solvers. For the three view scenario, the performance of all of our solvers can be further improved by pose refinement (+R) and filtering (+F). + +In the last set of experiments, we study the accuracy of the mean point correspondence. We sample 100 four-tuples of point correspondences consistent with the ground truth relative pose, i.e., inliers, for each image pair in scenes from the PhotoTourism dataset [22]. We use the first three correspondences to define the triangles in both images. Then, we establish correspondences between the mean of the triangle in one image and various points in the triangle in the second image. We express points in the second triangle via their barycentric coordinates and uniformly sample $19 \times 19$ barycentric coordinates $(a, b) \in [0, 1]^2$ , such that $a + b \leq 1$ . + +Fig. 3 shows the results for the rotation error and the percentage of inliers consistent with the pose obtained with the $4\mathrm{p}$ (M) solver on the St. Peter's Square scene. The optimum of the metrics is reached around the mean point of the triangles (the mean values of 2D Gaussians fitted to the results are very close to the mean point $(0.3,0.3)$ of the triangles). A detailed description, the translation and the symmetric epipolar errors, and the results for more scenes are in SM. For all tested scenes, we observed a similar behavior. + +Experiments on real data. We test the solvers on all scenes from the PhotoTourism dataset [22, 50] which provide ground truth poses and intrinsics via a COLMAP [49] reconstruction. In the results, we do not include the St. Peter's Square scene that we used for the validation of $\delta$ and the number of refinement iterations (see SM). We also include results for the Cambridge Landmarks dataset [23] (except the Street scene, which is commonly not used due to issues with its ground truth) and Aachen Day-Night v1.1 [58]. For PhotoTourism and Aachen, we use the images in their original resolution. For Cambridge Landmarks, we resize them so that the larger side is $800\mathrm{px}$ . For each scene, we sample 5,000 random image triplets with at least 10 matches obtained with [9, 32] with at least $10\%$ overlap [22]. + +Tab. 2 shows the results for PoseLib RANSAC with early termination and 5px epipolar threshold. Results for GC-RANSAC are in SM. With the suggested modifications (see Sec. 3.2) all proposed solvers outperform the state-of-the-art $4\mathrm{p}3\mathrm{v}$ (HC) solver [21] in terms of pose accuracy with many of the variants also achieving faster runtimes. + +Using filtering $(+F)$ improves the run-time of RANSAC at the cost of a decrease in pose accuracy. Still, our $4p3v(M) + R + F$ solver outperforms $4p3v(HC)$ in terms of both accuracy and run-time. The $4p3v(M \pm \delta)$ and $4p3v(M \pm \delta) + R$ solvers clearly improve upon the $4p3v(M)$ solvers, albeit at an increased run-time. Still the $4p3v(M \pm \delta) + R + F$ solver provides, in general, the best speed-accuracy trade-off. The results also show that early non-minimal refinement $(+ENM)$ is necessary for the $4p3v(A)$ solver to perform well. For the other approximate solvers, this form of refinement also leads to improved accuracy at some cost to the runtime. Assuming that the $5pt + P3P$ solver samples five points from which at least four have projections in all three cameras, the suggested pose refinement $(+R)$ and filtering $(+F)$ can also be applied to the baseline $5pt + P3P$ by using the fourth (potentially the fifth) point in the third view. The suggested modifications improve the performance of the $5pt + P3P$ solver, with the $5pt + P3P + R + F + ENM$ solver having an accuracy similar to that of the $4p3v(M) + R + F + ENM$ solver, however at slower run-times. Moreover, it is less accurate than the proposed $4p3v(M \pm \delta) + R + F + ENM$ solver. + +We also investigate the speed-accuracy trade-off of the solvers by running PoseLib RANSAC [27] for a set of fixed numbers of iterations. Runtimes are reported for 1 core of a $2\mathrm{GHz}$ Intel Xeon Gold 6338 CPU. As shown in Fig. 4, the proposed $4\mathrm{p}3\mathrm{v}$ $(\mathbb{M}\pm \delta) + \mathbb{R} + \mathbb{F}$ solver consistently provides the best speed-accuracy trade-off both with and without ENM. $4\mathrm{p}3\mathrm{v}$ (A) $+\mathbb{R} + \mathbb{F} + \mathbb{ENM}$ provides a similar performance, typically beating $4\mathrm{p}3\mathrm{v}$ (HC) [21]. However, as shown in Fig. 4d for the St. Mary's Church scene [23], it may perform worse for some specific scenes. + +The SM provides detailed ablation studies on the signif + +![](images/63e5c894136f5d0be36096f78de823ad5f38a80e8109eeb53ab1c4f30aa0d01b.jpg) +(a) Phototourism [22] +(b) Cambridge Landmarks [23] +(c) Aachen Day-Night v1.1 [58] +(d) St. Mary's Church [23] +Figure 4. Speed-accuracy trade-off for (a) all scenes from PhotoTourism [22] except St. Peter's Square, (b) 5 scenes from Cambridge Landmarks [23], (c) Aachen Day-Night v1.1 [58], and (d) St. Mary's Church scene from the Cambridge Landmarks dataset [23]. We report the AUC@10° of the pose error and vary the number of PoseLib RANSAC iterations (100, 200, 500, 1000, 2000, 5000, 10,000) with a 5 px epipolar threshold. Runtimes are averaged over all image triplets. + +
EstimatorPhotoTourism [22]Cambridge Landmarks [23]Aachen Day-Night v1.1 [58]
AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓
4p3v (HC) [21]6.371.6262.4173.5381.9464.108.732.6747.9964.5876.6560.1112.433.5542.2855.2066.5567.26
5pt+P3P5.401.6062.9374.3083.0033.777.462.6548.3665.3377.7524.0410.523.3843.4356.7368.2853.34
5pt+P3P+ENM5.231.5663.7775.0383.5048.797.202.6348.7465.6878.0434.8210.593.4143.3256.6468.0671.61
5pt+P3P+R+F5.101.5064.9075.9084.1138.127.152.5849.4266.3578.5531.8210.373.3643.7357.1568.7154.34
5pt+P3P+R+F+ENM4.991.5065.1876.2084.3056.526.982.5849.3866.3178.5747.0710.203.3443.7557.1168.6376.86
4p3v (A)37.4825.1622.9128.5935.1616.5841.2321.7322.7531.4738.9813.3336.4019.5324.4131.7339.2432.04
4p3v (A)+ENM4.971.4865.4976.4784.5340.457.022.5949.3566.3278.5328.3510.353.3843.6256.9968.5562.44
4p3v (A)+R33.4519.9525.8632.2239.1616.3236.9811.5927.0336.7044.4812.4834.8418.2225.2732.7740.2929.56
4p3v (A)+R+F35.6523.4723.9929.9736.6411.1038.3414.4725.8935.1942.809.3536.7720.7623.9431.2338.5819.75
4p3v (A)+R+F+ENM4.981.4865.6276.5684.5532.376.992.5749.6266.5578.7223.4310.433.3843.5756.9568.5642.36
4p3v (M)6.021.7160.8672.5681.7335.188.362.7447.1564.0676.5925.1512.093.6241.7855.0666.6754.89
4p3v (M)+ENM5.281.5863.6174.9683.4648.837.212.6348.6965.5777.9534.9310.673.4343.3156.6168.0870.87
4p3v (M)+R5.491.5663.6674.8883.3741.527.692.6348.6965.6777.9930.7611.173.4643.2856.3967.8860.40
4p3v (M)+R+F5.481.5463.9975.1483.5130.887.752.6248.7765.7177.9822.7211.163.4843.0456.2867.7342.38
4p3v (M)+R+F+ENM5.011.5065.1876.1684.2844.516.992.5849.4366.3178.5632.4810.253.3943.7457.2168.8058.08
4p3v (M±δ)5.551.6861.8673.7382.7883.707.652.6748.1665.2577.7659.2311.133.4742.7456.1167.74125.02
4p3v (M±δ)+ENM4.991.5664.0275.5184.01125.666.882.6049.1166.1378.5189.0510.263.3543.6656.9968.55175.54
4p3v (M±δ)+R4.861.4765.9777.0085.01100.617.102.5649.7966.8979.0873.9410.373.3443.9357.2668.82138.53
4p3v (M±δ)+R+F4.921.4765.8576.8784.9071.737.192.5649.7166.7578.9552.8410.523.3643.7957.1768.7392.89
4p3v (M±δ)+R+F+ENM4.661.4666.1277.1485.15112.606.652.5549.8766.9579.2081.989.963.3244.0257.4268.97139.19
+ +Table 2. Results for different solvers implemented in the PoseLib framework [27] on 12 scenes from PhotoTourism [22], 5 scenes from Cambridge Landmarks [23] and Aachen Day-Night v1.1 [58]. We mark the best and second best results. Runtimes are reported in ms for the whole RANSAC with early termination (0.9999 confidence, minimum 100 iterations) and epipolar threshold set to 5 px. + +icance of each individual modification $(+R / + C / + ENM)$ + +Limitations. As discussed, the accuracy of the proposed approximate solvers is scene dependent. However, after using the proposed ENM refitting, the scene-dependency is negligible. Especially for the $4\mathrm{p}3\mathrm{v}$ $(\mathbb{M}\pm \delta)$ +ENM solver, we have not noticed performance drops for some specific scene geometries or camera configurations. While the two-view variants of our solvers $4\mathrm{p}$ $(\mathbb{M}) / 4\mathrm{p}$ $(\mathbb{M}\pm \delta)$ provide very similar results to the 5pt solver in two views, in this case they do not outperform the 5pt solver. In the three-view scenario, the better performance is achieved thanks to the proposed pose refinement $(+\mathbf{R})$ and filtering $(+\mathbf{F})$ + +The proposed ENM refitting cannot be applied to the $4\mathrm{p}3\mathrm{v}$ (HC) solver [21], since this solver estimates the pose of all three cameras together. It is theoretically applicable to the solver proposed in [36]. However, here ENM would have needed to be run on hundreds of candidate poses corresponding to tens-to-hundreds of sampled epipoles. This + +is because for one fixed point on the curve of epipoles, the error of the sampled epipole is not bounded since the true epipole can lie anywhere on the curve, even outside the image. Thus, ENM might avoid local minima present in the second step of the original solver, but would significantly slow down the solver in the first step, making it much slower than our $4\mathrm{p}3\mathrm{v}$ $(\mathbf{M}) + \mathsf{ENM} / 4\mathrm{p}3\mathrm{v}$ $(\mathbf{M}\pm \delta)$ +ENM solvers. + +# 5. Conclusion + +We consider the highly challenging problem of relative pose estimation of three calibrated cameras from four correspondences. We propose novel approaches that solve the problem by utilizing solvers based on approximate geometry. The best performing solver uses a simple, yet novel strategy by using mean coordinates of three input points and points in their vicinity as an approximate $5^{th}$ correspondences. Extensive experiments show that our solvers achieve state-of-the-art performance on a large variety of real scenes. At the same time, our solvers are simple to implement, especially compared to the current state-of-the-art [21]. + +Acknowledgements. This work was funded by the Czech Science Foundation (GACR) JUNIOR STAR Grant No. 22-23183M (supporting C.T., Y.D., and Z.K.), the Grant Agency of the Czech Technical University in Prague grant no. SGS23/173/OHK3/3T/13 (supporting C.T.), the EU NextGenerationEU through the Recovery and Resilience Plan for Slovakia under the project "InnovAite Slovakia, Illuminating Pathways for AI-Driven Breakthroughs" No. 09I02-03-V01-00029, the TERAIS project - a Horizon-Widera-2021 program of the European Union under the Grant agreement number 101079338 (supporting V.K. and Z.B.H.), and the EU Horizon 2020 project RICAIP grant agreement No. 857306 (supporting T.S.). Part of the research results was obtained using the computational resources procured in the national project National competence centre for high performance computing project code: 311070AKF2, funded by European Regional Development Fund, EU Structural Funds Informatization of society, Operational Program Integrated Infrastructure. + +# References + +[1] Chris Aholt and Luke Oeding. The ideal of the trifocal variety. Mathematics of Computation, 83(289):2553-2574, 2014. 1, 3 +[2] D. Barath and J. Matas. Graph-Cut RANSAC. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6733-6741, 2018. 1, 6 +[3] Martin Bujnak, Zuzana Kukelova, and Tomas Pajdla. A general solution to the p4p problem for camera with unknown focal length. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1 +[4] Robert Castle, Georg Klein, and David W. Murray. Videorelate localization in multiple maps for wearable augmented reality. In ISWC, 2008. 1 +[5] Chiang-Heng Chien, Hongyi Fan, Ahmad Abdelfattah, Elias Tsigaridas, Stanimire Tomov, and Benjamin Kimia. Gpubased homotopy continuation for minimal problems in computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15765-15776, 2022. 3 +[6] Ondrej Chum, Jiří Matas, and Josef Kittler. Locally optimized ransac. In Pattern Recognition, pages 236-243. Springer Berlin Heidelberg, 2003. 1, 6 +[7] Andrea Porfiri Dal Cin, Timothy Duff, Luca Magri, and Tomas Pajdla. Minimal perspective autocalibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5064-5073, 2024. 3 +[8] Toby Collins and Adrien Bartoli. Infinitesimal plane-based pose estimation. International Journal of Computer Vision, 109:252-286, 2014. 3 +[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 224-236, 2018. 6, 7 + +[10] Yaqing Ding, Chiang-Heng Chien, Viktor Larsson, Karl Åström, and Benjamin Kimia. Minimal solutions to generalized three-view relative pose problem. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8156-8164, 2023. 3 +[11] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. Plmp-point-line minimal problems in complete multiview visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1675-1684, 2019. 3 +[12] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomás Pajdla. $\mathrm{Pl}_1\mathfrak{p}$ - point-line minimal problems under partial visibility in three views. In ECCV (26), pages 175-192. Springer, 2020. 3, 4 +[13] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H. Regan, David da Costa de Pinho, Elias P. Tsigaridas, Charles W. Wampler, Jonathan D. Hauenstein, Peter J. Giblin, Benjamin B. Kimia, Anton Leykin, and Tomás Pajdla. TRPLP - trifocal relative pose from lines at points. In CVPR, pages 12070-12080. Computer Vision Foundation / IEEE, 2020. 3, 4 +[14] M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6): 381-395, 1981. 1 +[15] Liran Goshen and Ilan Shimshoni. Balanced exploration and exploitation model search for efficient epipolar geometry estimation. In Computer Vision - ECCV 2006, pages 151-164, Berlin, Heidelberg, 2006. Springer Berlin Heidelberg. 3 +[16] Christian Häne, Lionel Heng, Gim Hee Lee, Friedrich Fraun-dorfer, Paul Furgale, Torsten Sattler, and Marc Pollefeys. 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image and Vision Computing, 68:14-27, 2017. 1 +[17] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision, 2006. 2, 4 +[18] Joel A. Hesch and Stergios I. Roumeliotis. A direct least-squares (dls) method for pnp. In 2011 International Conference on Computer Vision, pages 383-390, 2011. 6 +[19] R.J. Holt and A.N. Netravali. Uniqueness of solutions to three perspective views of four points. IEEE Transactions on Pattern Analysis and Machine Intelligence, 17(3):303-307, 1995. 1, 3 +[20] Radu Horaud, Fadi Dornaika, and Bart Lamiroy. Object pose: The link between weak perspective, parapspective, and full perspective. Int. J. Comput. Vision, 22(2):173-189, 1997. 3, 6 +[21] Petr Hruby, Timothy Duff, Anton Leykin, and Tomas Pajdla. Learning to solve hard minimal problems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5532-5542, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[22] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 6, 7, 8 + +[23] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Proceedings of the IEEE international conference on computer vision, pages 2938-2946, 2015. 7, 8 +[24] J. Kileel. Minimal problems for the calibrated trifocal variety. SIAM Journal on Applied Algebra and Geometry, 1(1): 575-598, 2017. 1, 3 +[25] Z. Kukelova and T. Pajdla. A minimal solution to the autocalibration of radial distortion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2007), 2007. 1 +[26] Zuzana Kukelova, Martin Bujnak, and Tomas Pajdla. Real-time solution to the absolute pose problem with unknown radial distortion and focal length. In Proceedings of the IEEE International Conference on Computer Vision, pages 2816-2823, 2013. 1 +[27] Viktor Larsson. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 6, 7, 8 +[28] Viktor Larsson, Kalle Åström, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Computer Vision and Pattern Recognition (CVPR), pages 820-829, 2017. 2 +[29] Viktor Larsson, Torsten Sattler, Zuzana Kukelova, and Marc Pollefeys. Revisiting radial distortion absolute pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1062-1071, 2019. 1 +[30] Karel Lebeda, Jiri Matas, and Ondrej Chum. Fixing the locally optimized ransac. In Proceedings of the British Machine Vision Conference 2012, pages 4961-4974. British Machine Vision Association, 2012. 3, 5 +[31] S. Leonardos, R. Tron, and K. Daniilidis. A metric parametrization for trifocal tensors with non-colinear pinholes. In Computer Vision and Pattern Recognition (CVPR), pages 259-267. IEEE, 2015. 1, 3 +[32] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In International Conference on Computer Vision (ICCV), pages 17627–17638, 2023. 6, 7 +[33] H. Christopher Longuet-Higgins. A method of obtaining the relative positions of 4 points from 3 perspective projections. In BMVC91, pages 86-94, London, 1991. Springer London. 3 +[34] Evgeniy V. Martyushev. On some properties of calibrated trifocal tensors. Journal of Mathematical Imaging and Vision, 58(2):321-332, 2017. 1, 3 +[35] D. Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 1, 2, 4, 5, 6 +[36] D. Nister and F. Schaffalitzky. Four points in two or three calibrated views: Theory and practice. International Journal of Computer Vision, 67(2):211-231, 2006. 1, 2, 3, 4, 8 +[37] Denis Oberkampf, Daniel F DeMenton, and Larry S Davis. Iterative pose estimation using coplanar feature points. Computer Vision and Image Understanding, 63(3):495-511, 1996. 3 + +[38] Magnus Oskarsson, Andrew Zisserman, and Kalle Åström. Minimal projective reconstruction for combinations of points and lines in three views. Image and Vision Computing, 22 (10):777-785, 2004. British Machine Vision Computing 2002. 2 +[39] Michal Perdoch, Jiri Matas, and Ondrej Chum. Epipolar geometry from two correspondences. In 18th International Conference on Pattern Recognition (ICPR'06), pages 215-219. IEEE, 2006. 3 +[40] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (p3p) solver. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318-332, 2018. 1, 2, 3, 4, 6 +[41] James Pritts, Ondrej Chum, and Jiri Matas. Approximate models for fast and accurate epipolar geometry estimation. In 28th International Conference on Image and Vision Computing New Zealand, IVCNZ 2013, Wellington, New Zealand, November 27-29, 2013, pages 106-111. IEEE, 2013. 2, 3, 4 +[42] James Pritts, Zuzana Kukelova, Viktor Larsson, and Ondrej Chum. Radially-distorted conjugate translations. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 1993-2001. IEEE Computer Society, 2018. 3 +[43] Long Quan. Invariants of six points and projective reconstruction from three uncalibrated images. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 17(1): 34-46, 1995. 2 +[44] L. Quan, B. Triggs, and B. Mourrain. Some results on minimal euclidean reconstruction from four points. Journal of Mathematical Imaging and Vision, 24(3):341-348, 2006. 1, 3 +[45] R. Raguram, O. Chum, M. Pollefeys, J. Matas, and J.-M. Frahm. USAC: A universal framework for random sample consensus. IEEE Transactions on Pattern Recognition and Machine Intelligence, 35(8):2022-2038, 2013. 1 +[46] Volker Rodehorst. Evaluation of the metric trifocal tensor for relative three-view orientation. In Digital Proceedings, International Conference on the Applications of Computer Science and Mathematics in Architecture and Civil Engineering: July 20 - 22 2015, Bauhaus-University Weimar, 2017. 4 +[47] T. Sattler, B. Leibe, and L. Kobbelt. Efficient & Effective Prioritized Matching for Large-Scale Image-Based Localization. IEEE Transactions on Pattern Recognition and Machine Intelligence, 39(9):1744-1756, 2016. 1 +[48] D. Scaramuzza and F. Fraundorfer. Visual odometry [tutorial]. IEEE Robot. Automat. Mag., 18(4):80-92, 2011. 1 +[49] Johannes L. Schonberger and Jan-Michael Frahm. Structure-From-Motion Revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104–4113, 2016. 7 +[50] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 7 +[51] N. Snavely, S. M. Seitz, and R. Szeliski. Modeling the world from internet photo collections. International Journal Computer Vision, 80(2):189-210, 2008. 1 + +[52] Andrew J Sommese, Andrew J Sommese, and Charles W Wampler. The numerical solution of systems of polynomials arising in engineering and science. World Scientific Publishing Co. Pte. Ltd, Singapore, 2005. 3 +[53] H. Stewenius, D. Nister, F. Kahl, and F. Schaffalitzky. A minimal solution for relative pose with unknown focal length. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2005), 2005. 1 +[54] H. Stewenius, C. Engels, and D. Nister. Recent developments on direct relative orientation. ISPRS J. of Photogrammetry and Remote Sensing, 60:284-294, 2006. 1 +[55] P. H. S. Torr and A. Zisserman. Robust parameterization and computation of the trifocal tensor. Image and Vision Computing, 15:591-605, 1997. 2 +[56] G. Xu and N. Sugimoto. A linear algorithm for motion from three weak perspective images using euler angles. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(1):54-57, 1999. 3 +[57] Zhengyou Zhang. Weak Perspective Projection, pages 877-883. Springer US, Boston, MA, 2014. 2, 4, 6 +[58] Zichao Zhang, Torsten Sattler, and Davide Scaramuzza. Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision, 129(4):821-844, 2021. 7, 8 \ No newline at end of file diff --git a/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/images.zip b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..41adc13db4270f0eaa89a6cc2fa3d64e50a3c614 --- /dev/null +++ b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbe33a7cd06bb4bcc2efb51a2310cb161d980a6f72428c3dcfd7a72e8fb093dc +size 309430 diff --git a/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/layout.json b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..54b5967811af68af1690458c222ce18b9e2e3ca7 --- /dev/null +++ b/2025/Practical Solutions to the Relative Pose of Three Calibrated Cameras/layout.json @@ -0,0 +1,9679 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "type": "text", + "content": "Practical solutions to the relative pose of three calibrated cameras" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "text", + "content": "Charalambos Tzamos" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "text", + "content": " Viktor Kocur" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "text", + "content": " Yaqing Ding" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "text", + "content": " Daniel Barath" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "text", + "content": " Zuzana Berger Haladová" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "text", + "content": " Torsten Sattler" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "text", + "content": " Zuzana Kukelova" + }, + { + "bbox": [ + 61, + 142, + 574, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 180, + 560, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 180, + 560, + 222 + ], + "spans": [ + { + "bbox": [ + 75, + 180, + 560, + 222 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 180, + 560, + 222 + ], + "type": "text", + "content": "Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague " + }, + { + "bbox": [ + 75, + 180, + 560, + 222 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 180, + 560, + 222 + ], + "type": "text", + "content": "Faculty of Mathematics, Physics and Informatics, Comenius University in Bratislava " + }, + { + "bbox": [ + 75, + 180, + 560, + 222 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 180, + 560, + 222 + ], + "type": "text", + "content": "ETH Zürich (Zurich), HUN-REN SZTAKI (Budapest)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 223, + 550, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 223, + 550, + 237 + ], + "spans": [ + { + "bbox": [ + 83, + 223, + 550, + 237 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 223, + 550, + 237 + ], + "type": "text", + "content": "Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 265, + 200, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 265, + 200, + 277 + ], + "spans": [ + { + "bbox": [ + 151, + 265, + 200, + 277 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 289, + 297, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 297, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 297, + 517 + ], + "type": "text", + "content": "We study the challenging problem of estimating the relative pose of three calibrated cameras from four point correspondences. We propose novel efficient solutions to this problem that are based on the simple idea of using four correspondences to estimate an approximate geometry of the first two views. We model this geometry either as an affine or a fully perspective geometry estimated using one additional approximate correspondence. We generate such an approximate correspondence using a very simple and efficient strategy, where the new point is the mean point of three corresponding input points. The new solvers are efficient and easy to implement, since they are based on existing efficient minimal solvers, i.e., the 4-point affine fundamental matrix, the well-known 5-point relative pose solver, and the P3P solver. Extensive experiments on real data show that the proposed solvers, when properly coupled with local optimization, achieve state-of-the-art results, with the novel solver based on approximate mean-point correspondences being more robust and accurate than the affine-based solver." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 533, + 135, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 135, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 135, + 545 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 554, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 696 + ], + "type": "text", + "content": "Camera geometry estimation is crucial in many computer vision applications, e.g., visual navigation [48], Structure-from-Motion [51], augmented reality [4], self-driving cars [16], and visual localization [47]. Due to noise and outliers in input correspondences, the predominant way for camera geometry estimation is to use a hypothesis-and-test framework, e.g., RANSAC [2, 6, 14, 45]. For RANSAC-like methods, using as few (ideally the minimal number of) correspondences as possible for estimation is important since the number of RANSAC iterations (and thus its run-time) grows exponentially with the number of correspondences required for the model estimation." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 368, + 263, + 502, + 348 + ], + "blocks": [ + { + "bbox": [ + 368, + 263, + 502, + 348 + ], + "lines": [ + { + "bbox": [ + 368, + 263, + 502, + 348 + ], + "spans": [ + { + "bbox": [ + 368, + 263, + 502, + 348 + ], + "type": "image", + "image_path": "4ceef2ab7f660eb7eb7af3c4704fe4d64a2870881275d38e82103d2b3d8579e7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 358, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. Visualization of the four-points-in-three-views (4p3v) problem and our solution based on using four correspondences to efficiently estimate an approximate geometry of the first two views and then register the third view using a P3P solver [40]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 414, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 414, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 312, + 414, + 555, + 593 + ], + "type": "text", + "content": "Minimal camera geometry problems often result in complex systems of polynomial equations. Efficient algebraic methods helped solve many previously unsolved problems [3, 25, 26, 29, 53, 54]. Still, they fail to generate efficient and/or numerically stable solutions for some configurations. In this paper, we study one such challenging problem: Estimating the relative pose of three calibrated cameras. This problem has received attention for a long time [1, 19, 31, 34, 44]. However, due to its complexity, it is still not considered fully solved. There are no efficient and practical solutions for most of the minimal configurations of point and/or line correspondences [24]. One such configuration that is particularly interesting is the notoriously difficult configuration of four points in three calibrated views [35, 44], known as the " + }, + { + "bbox": [ + 312, + 414, + 555, + 593 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 312, + 414, + 555, + 593 + ], + "type": "text", + "content": " problem." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 556, + 714 + ], + "type": "text", + "content": "State-of-the-art algebraic and numerical methods are known to fail in generating efficient and numerically stable solutions to the 4p3v problem. The existing methods for solving this problem are only approximate [21, 36]. By solving only for one (or a few) solutions from the 272 solutions of the 4p3v problem [21], and by discretely sampling the space of potential solutions [36], the existing 4p3v methods can often fail, i.e., the returned solution can be, in general, arbitrarily far from the geometrically correct solution. To decrease the failure rate, both methods [21, 36]" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 138, + 712 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21913" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "type": "text", + "content": "require a lot of tuning and are challenging to re-implement.1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 85, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 85, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 54, + 85, + 294, + 180 + ], + "type": "text", + "content": "In contrast to complex solutions to the configuration of four points in three views, there is a simple and efficient solution for the configuration where five correspondences are detected in the first two views and three of these points are also visible in the third view. In this case, the " + }, + { + "bbox": [ + 54, + 85, + 294, + 180 + ], + "type": "inline_equation", + "content": "5\\mathrm{pt} + \\mathrm{P3P}" + }, + { + "bbox": [ + 54, + 85, + 294, + 180 + ], + "type": "text", + "content": " solver first estimates the relative pose of two cameras from five correspondences using an efficient 5pt solver [35], and then registers the third camera using a P3P solver [40]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 180, + 294, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 180, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 54, + 180, + 294, + 252 + ], + "type": "text", + "content": "Following the idea of first estimating the relative pose of two cameras, we propose two novel approaches for solving the 4p3v problem. Our solutions are based on the simple idea of using four correspondences to estimate an approximate geometry of the first two views and then registering the third camera using a P3P solver [40]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 253, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 253, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 294, + 324 + ], + "type": "text", + "content": "In the first approach, we approximate the geometry of the first two views using affine cameras. The approximation using affine cameras was shown to provide good accuracy for relative pose estimation if coupled with geometry refitting and local optimization inside RANSAC [41]. We denote the proposed affine-based 4p3v solver as 4p3v (A)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 326, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 326, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 326, + 295, + 529 + ], + "type": "text", + "content": "In the second approach, we use a less restrictive approximation. In this case, we estimate the full perspective geometry, i.e., the full 5DoF essential matrix, using four input correspondences and one additional approximate correspondence in two views. We generate the approximate correspondence using the locations of three of the four input correspondences. Under the assumption of a paraperspective projection, i.e., of affine geometry, the mean point of three 3D points is projected to the mean points in both images [57]. Thus, the new correspondence is generated as the correspondence between the mean points of three corresponding points detected in two views. This approximate mean-point correspondence can also be seen as a correspondence under the " + }, + { + "bbox": [ + 56, + 326, + 295, + 529 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 56, + 326, + 295, + 529 + ], + "type": "text", + "content": "-order approximation of the homography defined by the plane passing through the three corresponding 3D points. We denote the proposed mean point-based 4p3v solver as 4p3v (M)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 529, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 685 + ], + "type": "text", + "content": "While the proposed " + }, + { + "bbox": [ + 55, + 529, + 295, + 685 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 55, + 529, + 295, + 685 + ], + "type": "text", + "content": " (M) solver returns significantly more accurate poses than the " + }, + { + "bbox": [ + 55, + 529, + 295, + 685 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 55, + 529, + 295, + 685 + ], + "type": "text", + "content": " (A) solver, both solvers may not be accurate enough if not properly treated inside RANSAC. Thus, in this paper, we propose several ways to improve the accuracy of the proposed approximate solvers. (i) To compensate for noise in the mean point correspondences, we introduce the " + }, + { + "bbox": [ + 55, + 529, + 295, + 685 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 55, + 529, + 295, + 685 + ], + "type": "inline_equation", + "content": "(\\mathbb{M}\\pm \\delta)" + }, + { + "bbox": [ + 55, + 529, + 295, + 685 + ], + "type": "text", + "content": " solver that generates two additional correspondences in the vicinity of the mean point, and in the first step calls the 5pt solver [35] three times. (ii) To improve the approximate geometry estimated between the first two views, we refit this geometry using a non-minimal relative pose solver and inliers obtained from the approximate geometry (ENM). (iii) We use" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "content": "the fourth correspondence in the third view to filter out geometrically infeasible solutions " + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "inline_equation", + "content": "(+F)" + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "content": " and to refine the solutions on four input correspondences in three views using just a few iterations of Levenberg-Marquardt (LM) refinement " + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "inline_equation", + "content": "(+R)" + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "content": ". While conceptually simple and efficient, the novel solvers achieve state-of-the-art results on real data. The contributions of the paper are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 156, + 553, + 502 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 156, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 311 + ], + "type": "text", + "content": "- We propose two groups of novel solutions for the well-known and challenging 4p3v problem: " + }, + { + "bbox": [ + 313, + 156, + 553, + 311 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}(\\mathrm{A})" + }, + { + "bbox": [ + 313, + 156, + 553, + 311 + ], + "type": "text", + "content": " - and " + }, + { + "bbox": [ + 313, + 156, + 553, + 311 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}(\\mathrm{M})" + }, + { + "bbox": [ + 313, + 156, + 553, + 311 + ], + "type": "text", + "content": " -based solvers. These solutions are based on the simple idea of using four correspondences to estimate an approximate geometry in the first two views. Compared to state-of-the-art 4p3v solvers [21, 36], which are non-trivial and difficult to re-implement such that they are numerically stable and fast, our new solutions can be easily implemented using existing efficient implementations of the linear 4pt affine fundamental matrix solver [17], the 5pt solver [35] and the P3P solver [40]. The source code is available at https://github.com/kocurvik/threeview." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 430 + ], + "type": "text", + "content": "- We present several ways of improving the accuracy and the speed of the proposed solvers, as well as a strategy for efficiently using these approximate solvers inside a RANSAC-style paradigm. We show that the new solvers achieve state-of-the-art results in terms of accuracy on real data. While both " + }, + { + "bbox": [ + 313, + 312, + 553, + 430 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 312, + 553, + 430 + ], + "type": "text", + "content": " (A)- and " + }, + { + "bbox": [ + 313, + 312, + 553, + 430 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 312, + 553, + 430 + ], + "type": "text", + "content": " (M)-based solvers achieve comparable results when coupled with the suggested non-minimal geometry refitting and local optimization, our " + }, + { + "bbox": [ + 313, + 312, + 553, + 430 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 312, + 553, + 430 + ], + "type": "text", + "content": " (M)-based solvers are more robust to the scene geometry and RANSAC inlier thresholds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 431, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 553, + 502 + ], + "type": "text", + "content": "- To our knowledge, we are the first to extensively evaluate solutions to the 4p3v problem on a large variety of real-world scenes and within state-of-the-art RANSAC frameworks, and to compare them to the baseline 5pt+P3P solver. We report results on 3 datasets, consisting of 18 different scenes and altogether 90,000 camera triplets." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 514, + 399, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 399, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 399, + 526 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 641 + ], + "type": "text", + "content": "Estimating the relative pose of three cameras from a minimal number of point and line correspondences is known as an extremely challenging problem. For three uncalibrated cameras, 6 point correspondences are necessary to estimate the trifocal tensor, with a solution known for a long time [43, 55]. Solutions to three minimal combinations of points and lines are presented in [38]. The configuration of 9 lines is more challenging and was solved only recently [28]. Yet, the solver is far from practical and runs 17.8s." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "For calibrated cameras, the configuration that attracts most of the attention is the configuration of four points in three views (the 4p3v problem). Note that this is not a minimal configuration since it generates 12 constraints for 11 degrees-of-freedom (DoF). The 4p3v problem is known to be extremely difficult to solve. Several papers present" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "There is no publicly available implementation for [36]. Implementation of [21] is quite complex and requires a non-negligible effort to run." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21914" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "content": "mostly theoretical results [1, 31, 34]. For four triplets of exact points without noise, it is shown that the 4p3v problem has, in general, a unique solution [19, 44]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "text", + "content": "To the best of our knowledge, there are only two reasonably efficient solutions to the " + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "text", + "content": " problem reported in the literature. The first solver [36] is based on a similar idea as our solvers, i.e., to first estimate the relative pose of two cameras and then register the third camera using a " + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "inline_equation", + "content": "\\mathsf{P}3\\mathsf{P}" + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "text", + "content": " solver [40]. To compute the pose of the first two cameras using information only from four point correspondences, the solver needs additional information about the position of one epipole. The paper shows that the four point correspondences between two calibrated views constrain the epipole in each image to lie on a curve of degree ten. Thus, the solver performs one-dimensional exhaustive search and sweeps a " + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "inline_equation", + "content": "10^{th}" + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "text", + "content": "-degree curve of possible epipoles. For each potential epipole, it computes the relative pose of two cameras, registers the third camera using three triangulated points, and finally extracts the solution minimizing the reprojection error of the fourth point in the third view. Evaluation of the solver on one potential epipole is fast. Yet, in contrast to our proposed " + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "text", + "content": " (M)-based solvers, the error of the sampled epipole for one fixed point on the curve is not bounded since the true epipole can lie anywhere on the curve, and in many scenarios it is very far from the image center (e.g., outside the image). Thus, to obtain reasonably accurate and stable results, usually 1,000 candidates need to be evaluated. Even then, refinement at multiple local minima is required to improve the accuracy. The runtimes reported for this solver were " + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "inline_equation", + "content": "1 - 12ms" + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "text", + "content": " depending on the number of points searched. On a small number of synthetic experiments, the paper shows that the translation error returned by the proposed solver is usually " + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "inline_equation", + "content": "1 - 10\\mathrm{deg}" + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "text", + "content": " higher than the error returned by the " + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "inline_equation", + "content": "5\\mathrm{pt} + \\mathsf{P}3\\mathsf{P}" + }, + { + "bbox": [ + 56, + 112, + 297, + 507 + ], + "type": "text", + "content": " solver. The literature does not compare against [36] as there is no publicly available implementation and it is hard to re-implement." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "The second efficient solver to the 4p3v problem was published only recently [21]. In this paper, the authors first transform the 4p3v problem into a minimal problem by considering a line passing through the last correspondence in the third view. The resulting system of equations is solved using an efficient Homotopy continuation (HC) method [13, 52]. To avoid computing large numbers of spurious solutions, an MLP-based classifier is trained. For a given problem " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ", it selects one or several starting problem-solution pairs (so-called anchors), such that the geometrically meaningful/correct solution of " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " can be obtained by HC starting from this anchor. This strategy is fast, running " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "16.3\\mu s" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " on average per solution. However, it has a high failure rate. The success rate of the 4p3v solver reported in [21] on two test datasets and data without noise is " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "26.3\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ". [21] do not show results for a real scenario, i.e., a RANSAC-like framework with noisy data. Providing such an evaluation," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "type": "text", + "content": "we show that our much simpler solvers outperform [21]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 85, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 553, + 239 + ], + "type": "text", + "content": "Solutions to the 4p3v problem for orthographic and par-perspective views were presented in [33, 56]. In [33], the author suggested an iterative approach for updating to perspective views, but reported results only on a few synthetic instances. According to our extensive experiments, the update does not work on real data with general perspective cameras. This solver is returning large errors even after incorporating it into RANSAC with local optimization. We see two main reasons: (i) The set of inliers that satisfy an approximate para-perspective camera model in all three images is usually quite small, and (ii) the estimated model is often very far from the perspective optimum, and thus local optimization does not converge to a good solution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 240, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 384 + ], + "type": "text", + "content": "In [11, 12, 24], the authors aim to classify and derive the number of solutions for different minimal configurations of points and lines in three calibrated views. Solutions to two minimal configurations combining points and lines were proposed in [13] and solved using a HC method [52]. Due to their complexity, the solvers are not practical. A GPU HC method was also used to solve minimal problems of four points/six lines in three views for a generalized camera in [10] and of four points in three cameras with an unknown shared focal length (4p3vf) in [5, 7]. Efficient GPU implementations of the 4p3vf solvers run " + }, + { + "bbox": [ + 313, + 240, + 553, + 384 + ], + "type": "inline_equation", + "content": "16.7ms" + }, + { + "bbox": [ + 313, + 240, + 553, + 384 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 240, + 553, + 384 + ], + "type": "inline_equation", + "content": "154ms" + }, + { + "bbox": [ + 313, + 240, + 553, + 384 + ], + "type": "text", + "content": ". These times are still too slow for practical applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 384, + 553, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 553, + 586 + ], + "type": "text", + "content": "An approximation of perspective cameras using affine ones was used in several papers on camera geometry estimation [8]. [37] and [20] propose iterative approaches to the absolute camera pose estimation problem, i.e., the PnP problem. Both methods first compute the pose for an affine camera (weak- or para-perspective). Then the error induced by the affine camera approximation is used to adjust the constraints on the pose and recompute it. These methods do not guarantee convergence to the perspective model solution. In [41], affine cameras are used to efficiently solve the relative pose problem of two uncalibrated cameras from two affine correspondences. These affine correspondences are transformed to six point or two ellipse correspondences and then used to estimate the affine fundamental matrix. The paper also proposes a RANSAC framework that uses local optimization [30] to estimate the full fundamental matrix using inliers from the approximate affine model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 587, + 553, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 553, + 658 + ], + "type": "text", + "content": "Points that are sampled based on feature geometry to generate point correspondences from affine or scale-and orientation-invariant feature correspondences were also used in [15, 39, 42]. In contrast, our M-based solvers use only point correspondences, without associated feature geometry, to generate an additional point correspondence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 670, + 553, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 553, + 683 + ], + "type": "text", + "content": "3. Estimating the relative pose of three cameras" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "In this section, we describe different solutions for estimating the relative pose of three calibrated cameras. We start" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21915" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "with a baseline solution for the minimal configuration of three points visible in all three cameras and two additional points visible in two of the three cameras (the " + }, + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "inline_equation", + "content": "(5,5,3)" + }, + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": " configuration). Next, we present our novel solutions for the configuration of four points visible in all three cameras (the " + }, + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "inline_equation", + "content": "(4,4,4)" + }, + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": " configuration). This configuration generates an over-constrained problem. In this case, we have one more constraint than DoF. A minimal solution would need to drop one constraint, e.g., by considering only a line passing through one of the points in the third view [21] or by considering a \"half\" point correspondence. Since, in practice, we always have full correspondences and sampling one less point in one view leads to an under-constrained problem, the " + }, + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "inline_equation", + "content": "(4,4,4)" + }, + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": " configuration, is usually considered \"minimal\"." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 243, + 294, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 294, + 316 + ], + "type": "text", + "content": "5pt+P3P solver: The 5pt+P3P solver first estimates the relative pose of two cameras from 5 image point correspondences using the efficient 5pt solver [35]. Next, the three points visible in all three views are triangulated. Finally, the third camera is registered using the three 2D-3D point correspondences and the well-known efficient P3P solver [40]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 294, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 294, + 447 + ], + "type": "text", + "content": "This straightforward solver, which is based on existing efficient solvers [35, 40], was discussed in several works [12, 35, 36, 46]. [36] showed that the " + }, + { + "bbox": [ + 55, + 316, + 294, + 447 + ], + "type": "inline_equation", + "content": "5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P}" + }, + { + "bbox": [ + 55, + 316, + 294, + 447 + ], + "type": "text", + "content": " solver performs better than their dedicated " + }, + { + "bbox": [ + 55, + 316, + 294, + 447 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 55, + 316, + 294, + 447 + ], + "type": "text", + "content": " solver on synthetic data. Yet, the most recent works [13, 21] that study the three view relative pose problem do not discuss the " + }, + { + "bbox": [ + 55, + 316, + 294, + 447 + ], + "type": "inline_equation", + "content": "5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P}" + }, + { + "bbox": [ + 55, + 316, + 294, + 447 + ], + "type": "text", + "content": " solver and do not use it as a baseline for comparison. To the best of our knowledge, the performance of this solver on real data and within state-of-the-art RANSAC frameworks in the context of the " + }, + { + "bbox": [ + 55, + 316, + 294, + 447 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 55, + 316, + 294, + 447 + ], + "type": "text", + "content": " problem has not been extensively studied. Our paper fills this gap in the literature." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 281, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 281, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 281, + 466 + ], + "type": "text", + "content": "3.1. Approximate solutions to the 4p3v problem" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 470, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 662 + ], + "type": "text", + "content": "In contrast to the (5,5,3) configuration, the (4,4,4) configuration, i.e., the configuration of four points in three views, leads to significantly more complex equations. State-of-the-art algebraic and numerical methods are known to fail in generating efficient and numerically stable solutions to these equations. Thus, solutions to the " + }, + { + "bbox": [ + 55, + 470, + 295, + 662 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 55, + 470, + 295, + 662 + ], + "type": "text", + "content": " problem require some approximations to be practical in real-world applications. This section introduces several practical approximate solutions to the " + }, + { + "bbox": [ + 55, + 470, + 295, + 662 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 55, + 470, + 295, + 662 + ], + "type": "text", + "content": " problem. Similarly to the " + }, + { + "bbox": [ + 55, + 470, + 295, + 662 + ], + "type": "inline_equation", + "content": "5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P}" + }, + { + "bbox": [ + 55, + 470, + 295, + 662 + ], + "type": "text", + "content": " solver, we decompose the problem into the problem of first estimating the relative pose of two cameras and then registering the third camera with an efficient P3P solver [40]. The idea is to assume an approximate geometry only in the first two views and use four input point correspondences to efficiently estimate this geometry. Next we describe two groups of such solvers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "4p3v(A) solver: The first group of solvers approximates the geometry in the first two views using affine cameras. Although the approximation can be quite far from the correct geometry, [41] showed that an approximate affine camera" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "model provides good accuracy for relative pose estimation if coupled with geometry refitting (using the full 7DoF fundamental matrix) and local optimization inside RANSAC. The proposed " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " (A) solver first uses four point correspondences in two views to efficiently estimate the affine fundamental matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{A}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " [17] and then registers the third camera using three triangulated points and a P3P solver [40]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "text", + "content": "4p3v(M) solver: The affine camera model used in the 4p3v (A)-based solvers can be quite imprecise. This motivates us to introduce solvers that are based on a less restrictive approximation. In this case, we approximate only one point correspondence in two views. Under the assumption of a para-perspective projection, i.e., of affine geometry, it is known that the mean point of three 3D points is projected to the mean points of their projections in both images [57]. Thus, we generate one new approximate correspondence in two views as the correspondence between the mean points of three corresponding points detected in these views. Let " + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^l" + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "text", + "content": " be the mean point of three points " + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^l,\\mathbf{x}_j^l,\\mathbf{x}_k^l\\}" + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "inline_equation", + "content": "i,j,k\\in \\{1,\\ldots ,4\\}" + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "text", + "content": ", in the view " + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "inline_equation", + "content": "l\\in \\{1,2\\}" + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{1}\\leftrightarrow \\mathbf{m}^{2}" + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "text", + "content": " is our new approximate correspondence. The new 4p3v (M) solver solves the 4p3v problem by first estimating the full 5DoF essential matrix from four original correspondences " + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^1\\leftrightarrow \\mathbf{x}_i^2" + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,4" + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "text", + "content": " correspondence " + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{1}\\leftrightarrow \\mathbf{m}^{2}" + }, + { + "bbox": [ + 313, + 160, + 553, + 424 + ], + "type": "text", + "content": ". This is done using the efficient 5pt relative pose solver [35]. As with previous solvers, the third camera is registered using a P3P solver [40]. In short, the 4p3v (M) solver solves the 4p3v problem using the 5pt+P3P solver and one approximate correspondence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": "As we show on large amounts of synthetic and real data (see Sec. 4 and Supp. mat. (SM)), the " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": " (M) solver provides much more accurate estimates than the " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": " (A) solver. Moreover, even in the basic form (without the modifications presented in the next section), on many scenes it performs comparably to the state-of-the-art HC solver [21]. This can be attributed to several facts and observations: (1) The " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^1 \\leftrightarrow \\mathbf{m}^2" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": " correspondence does not need to be seen as a correspondence of points that are projections of the mean point of three 3D points " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i, \\mathbf{X}_j," + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_k" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": " (in which case both " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^1" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^2" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": " would have some error). We can look at this correspondence as a correspondence in which we fix a point in one view, e.g., " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^1" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": ", and generate a corresponding point in the second view. In this case, to generate a good correspondence, we only require the point in the second view to be reasonably close to the epipolar line defined by the mean point " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^1" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": " in the first view, i.e., the 2D point does not need to correspond to one particular 3D point with a given depth. (2) The ray from the center of the first camera through the mean point " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^1" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": " intersects the plane of " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i, \\mathbf{X}_j," + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_k" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": " at point M, which lies inside the triangle formed by these 3D points. The projection of M into the second camera lies inside the triangle formed by " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^2, \\mathbf{x}_j^2, \\mathbf{x}_k^2\\}" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": ". By construction, this projection lies on the epipolar line " + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}\\mathbf{m}^1" + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21916" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "text", + "content": "Thus the epipolar line defined by " + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^1" + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "text", + "content": " passes through the triangle defined by " + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "text", + "content": ". Consequently, the maximum distance of " + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^2" + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "text", + "content": " in the second image from the epipolar line is bounded by the maximum distance of " + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^2" + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "text", + "content": ". For a formal lemma and proof, see SM. (3) For practical applications, when used in RANSAC, it is not necessary that each triplet " + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "text", + "content": " generates a good correspondence " + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^1\\leftrightarrow \\mathbf{m}^2" + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "text", + "content": ". Samples with a high level of noise in the mean-point correspondence are filtered inside RANSAC. On a large number of different scenes, we observed that even if some image pairs have triplets of points that generate very noisy mean-point correspondences, there are usually enough triplets for which the noise in " + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^2" + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "text", + "content": " is reasonably small to lead to good estimates. (4) Four point correspondences in two views usually fix the space of possible poses such that the " + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 55, + 72, + 296, + 339 + ], + "type": "text", + "content": " correspondence, even if noisy, often generates a pose that is not very far from the ground truth pose. Such a pose is usually sufficient for filtering out outliers and a good initialization for non-linear optimization on the original four points in three views and subsequent local optimization on detected inliers. We support our observations by experiments on a large amount of data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 346, + 257, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 257, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 257, + 358 + ], + "type": "text", + "content": "3.2. Making approximate solvers practical" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "text", + "content": " (A)-based solvers, when used without any modifications, generally provide imprecise results even when used inside RANSAC with local optimization on three views. While the accuracy of the pure " + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 55, + 363, + 296, + 460 + ], + "type": "text", + "content": " (M) solver inside LO-RANSAC [30] (RANSAC with Local Optimization) is much better, there is still room for improvement. Here we present several simple modifications of these solvers that significantly boost their performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": "4p3v(M±δ) solver: The mean point correspondence used in the 4p3v (M) solver can provide a good approximation of a correct correspondence. Yet, as mentioned above, it can also be noisy. In the 4p3v (M±δ) solver, we thus, in addition to the mean point " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^2 = [x,y]" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": " of three points " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\}" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": " in the second image, generate two additional points. These points are (1) " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\pm \\delta}^2 = [x\\pm \\delta ,y]" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": " if the longest dimension of the triangle " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^2 = \\Delta \\left\\{\\mathbf{x}_i^2,\\mathbf{x}_j^2,\\mathbf{x}_k^2\\right\\}" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": " is in the x-direction or (2) " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\pm \\delta}^2 = [x,y\\pm \\delta ]" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": " if it is in the y-direction. All three points, i.e., " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^2,\\mathbf{m}_{-\\delta}^2" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{+\\delta}^{2}" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": " are placed in correspondence with the mean point " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^1" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": ". The 4p3v (M±δ) solver in the first step calls the 5pt solver [35] three times, with the " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": " correspondence being either " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^1\\leftrightarrow \\mathbf{m}^2" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^1\\leftrightarrow \\mathbf{m}_{-\\delta}^2" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^1\\leftrightarrow \\mathbf{m}_{+\\delta}^2" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": ". The results of these three 5pt solvers are collected to create hypotheses for the relative pose of the first two cameras inside RANSAC. The shift " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": " is selected relative to the size of the triangle " + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^2" + }, + { + "bbox": [ + 55, + 463, + 296, + 670 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 673, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 685 + ], + "type": "text", + "content": "Early non-minimal refitting (ENM): The geometry es" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": "timated between the first two cameras using the proposed " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " (A) and " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " (M)- based solvers is only approximate. In standard LO-RANSAC, such a geometry is optimized in local optimization (LO) after registering the third view. However, this can lead to propagation of errors into triangulated points and subsequently into errors in the pose of the third camera. Inliers in three views w.r.t. such cameras together with imprecise pose initializations may not be sufficient for LO to converge to a good solution. We observed this especially for " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " (A)-based solvers. Fortunately, the design of our solvers allows us to optimize the geometry already after estimating the approximate geometry between the first two cameras. Here, even a very imprecise geometry returned by the " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " (A) solvers is usually sufficient to filter out outliers. Thus, after running the 4pt " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathrm{F_A / 5pt}" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " solver in the first step of the proposed solvers, we use the estimated approximate models to detect inliers in two views. We then refit the estimated geometry using the non-minimal version of the 5pt solver [35], which instead of a 4-dim null space of a " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "5\\times 9" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " matrix uses the last four vectors from the SVD/QR decomposition of a " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "n\\times 9" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " matrix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "text", + "content": " point in the third view: " + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "text", + "content": " (A) and " + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "text", + "content": " (M) solvers actually solve the configuration " + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "inline_equation", + "content": "(4,4,3)" + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "text", + "content": ", i.e., they do not use the information from the point " + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_4^3" + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "text", + "content": " in the third view. The information from " + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_4^3" + }, + { + "bbox": [ + 313, + 335, + 555, + 386 + ], + "type": "text", + "content": " can be used in two different ways:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "text", + "content": "Filtering " + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "inline_equation", + "content": "(+F)" + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_4^3" + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "text", + "content": " can be used to filter out geometrically infeasible solutions returned by the proposed solvers. Note that the P3P and 5pt solvers used inside the proposed methods and the " + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "text", + "content": "-based strategy return multiple solutions that can be evaluated w.r.t. " + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_4^3" + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "text", + "content": " to filter out spurious solutions. Since the returned solutions can be affected by the proposed approximation, we do not simply select the solution with the smallest error on " + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_4^3" + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "text", + "content": ", but we keep all solutions that have an epipolar error on " + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_4^3" + }, + { + "bbox": [ + 313, + 393, + 556, + 561 + ], + "type": "text", + "content": " smaller than twice the threshold used inside RANSAC. Our experiments show that this filtering can improve the speed of the proposed solvers. However, as a trade-off, there is sometimes a small drop in the accuracy of the solvers since, in some cases, geometrically correct solutions are filtered out." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": "Refinement " + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "inline_equation", + "content": "(+\\mathbb{R})" + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_4^3" + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": " can be used to refine the solutions returned by the proposed solvers. We want solutions that minimize the epipolar error on the original 4 points in all 3 views, i.e., that solve the original (4,4,4) configuration. (4,4,4) is an overconstrained configuration; thus, for noisy data, there is, in general, no solution with zero error on all 4 points in 3 views. We refine the poses by minimizing the epipolar error of the original four points in three views using LM optimization, initialized using the solutions from the " + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "inline_equation", + "content": "4p3v(A)" + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": " - and " + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "inline_equation", + "content": "4p3v(M)" + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": " -based solvers. Experiments with different numbers of iterations show that two iterations are usually sufficient to obtain an improvement (see SM)." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "2This property was also used in the HC solver [21], which completely fails for many samples. These samples are filtered within RANSAC." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21917" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 291, + 177 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 291, + 177 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 291, + 177 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 291, + 177 + ], + "type": "image", + "image_path": "73c81f1cb63439e0eecb94e0a58fc9c6154737b6782db60d828d2448b4cb0da6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 295, + 218 + ], + "type": "text", + "content": "Figure 2. Results of a synthetic experiment measuring the accuracy of two-view variants of our solvers depending on the angle between the principal axes of the cameras." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 238, + 136, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 136, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 136, + 251 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 258, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 342 + ], + "type": "text", + "content": "We extensively evaluated the proposed solvers on a large variety of synthetic and real data to test their robustness to noise, outliers, and scene properties, and to assess their performance inside state-of-the-art RANSAC-frameworks [2, 27]. We compare our novel solvers with the homotopy continuation " + }, + { + "bbox": [ + 55, + 258, + 296, + 342 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 55, + 258, + 296, + 342 + ], + "type": "text", + "content": " (HC) solver [21] and the " + }, + { + "bbox": [ + 55, + 258, + 296, + 342 + ], + "type": "inline_equation", + "content": "5\\mathrm{pt} + \\mathrm{P}3\\mathrm{P}" + }, + { + "bbox": [ + 55, + 258, + 296, + 342 + ], + "type": "text", + "content": " baseline minimal solver for the " + }, + { + "bbox": [ + 55, + 258, + 296, + 342 + ], + "type": "inline_equation", + "content": "(5,5,3)" + }, + { + "bbox": [ + 55, + 258, + 296, + 342 + ], + "type": "text", + "content": " configuration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 346, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 538 + ], + "type": "text", + "content": "Experimental setup. To obtain feature correspondences, we use SuperPoint [9] features with the Light-Flue [32] matcher. We extract at most 2048 features per image. We perform matching for all three image pairs and keep only those matches that were consistently matched across all three views. We perform evaluation within two RANSAC frameworks: PoseLib [27] and GC-RANSAC [2]. For the 5pt solver, we use [35] and for the P3P solver, we use [40]. In PoseLib, we perform LO [6] using LM optimization. In GC-RANSAC, we perform LO using non-minimal solvers [18, 35] for fitting models to larger-than-minimal samples. We evaluated different shifts for our " + }, + { + "bbox": [ + 55, + 346, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 346, + 296, + 538 + ], + "type": "text", + "content": "-based solvers using a validation scene (for the ablation study, see SM) and selected " + }, + { + "bbox": [ + 55, + 346, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\delta = 0.08 * (\\text{longest triangle dim.})" + }, + { + "bbox": [ + 55, + 346, + 296, + 538 + ], + "type": "text", + "content": ". The choice of " + }, + { + "bbox": [ + 55, + 346, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 346, + 296, + 538 + ], + "type": "text", + "content": " is not critical, i.e., different values perform similarly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "content": "Evaluation measures. Inspired by [22], we define the pose error as max " + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\left(0.5\\left(\\mathrm{R}_{err}^{12} + \\mathrm{R}_{err}^{13}\\right), 0.5\\left(\\mathrm{t}_{err}^{12} + \\mathrm{t}_{err}^{13}\\right)\\right)" + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{err}^{ij}" + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{err}^{ij}" + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "content": " are the angular errors of rotation and translation for pair " + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "inline_equation", + "content": "ij" + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "content": " in degrees [22]. We also report AUC values [22] at different thresholds for the pose error. We include results for an alternative pose error definition which includes " + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{err}^{23}" + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{err}^{23}" + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "content": " in SM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Approximate camera geometry. The first experiments aim to support our idea of estimating approximate geometry in the first two views. The accuracy of the used approximations depends on a large number of variables, including the depths of the points w.r.t. the cameras, the angle under which the points are observed, the type of motion, etc. A detailed analysis of all these factors, e.g., through synthetic" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 337, + 72, + 536, + 144 + ], + "blocks": [ + { + "bbox": [ + 337, + 72, + 536, + 144 + ], + "lines": [ + { + "bbox": [ + 337, + 72, + 536, + 144 + ], + "spans": [ + { + "bbox": [ + 337, + 72, + 536, + 144 + ], + "type": "image", + "image_path": "866947a811cca479d5cd44a944d4ec17d3e87a71a804da5a42ff398c48a3eb69.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 153, + 555, + 230 + ], + "lines": [ + { + "bbox": [ + 313, + 153, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 555, + 230 + ], + "type": "text", + "content": "Figure 3. Distribution of the (left) rotation error (0.3373, 0.3349); and (right) percentage of inliers gathered (0.3266, 0.3434), as a function of the barycentric coordinates of the triangle in the second image w.r.t. the mean point of the corresponding triangle in the first image on 465k four-tuples of correspondences from St. Peter's Square scene from the PhotoTourism dataset [22]. We fit a 2D Gaussian distribution to the results and report the mean in brackets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 253, + 503, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 503, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 503, + 264 + ], + "type": "text", + "content": "experiments, is beyond the scope of this paper." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 265, + 554, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 554, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 554, + 372 + ], + "type": "text", + "content": "An approximation error introduced by the paraperspective projection, i.e., the affine geometry, and its effect on absolute camera pose determination is studied in the literature [20, 57]. Here, we thus only study the effect of the proposed approximation on relative pose estimation for real-world data and interesting synthetic scenarios. In the following experiments, all solvers are tested on two views without RANSAC. The two-view variants of our solvers are denoted as " + }, + { + "bbox": [ + 313, + 265, + 554, + 372 + ], + "type": "inline_equation", + "content": "4p(\\mathsf{A})" + }, + { + "bbox": [ + 313, + 265, + 554, + 372 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 265, + 554, + 372 + ], + "type": "inline_equation", + "content": "4p(\\mathsf{M})" + }, + { + "bbox": [ + 313, + 265, + 554, + 372 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 265, + 554, + 372 + ], + "type": "inline_equation", + "content": "4p(\\mathsf{M} \\pm \\delta)" + }, + { + "bbox": [ + 313, + 265, + 554, + 372 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "text", + "content": "In the first set of experiments, we consider several interesting synthetic setups. Fig. 2 presents the results for increasing the angle between the principal axes of the two cameras. We generate 3D points in a " + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "inline_equation", + "content": "2000 \\times 2000 \\times 100" + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "text", + "content": " cube uniformly at random. The distances of the camera centers are random between 1000 and 1200. The cameras are looking towards the scene and their principal axes form a specified angle. To simulate realistic scenarios, we add 1px noise to the image correspondences and " + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "text", + "content": " outliers. It can be seen that the results for both approximate solvers deteriorate with increasing angle, with " + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "inline_equation", + "content": "4p(\\mathbb{A})" + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "text", + "content": " being significantly less accurate than " + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "inline_equation", + "content": "4p(\\mathbb{M})" + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "inline_equation", + "content": "4p(\\mathbb{M} \\pm \\delta)" + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "text", + "content": ". However, ENM can significantly decrease errors, with the " + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "inline_equation", + "content": "4p(\\mathbb{M} \\pm \\delta) + ENM" + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "text", + "content": " solver returning almost identical results to " + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "inline_equation", + "content": "5pt + ENM" + }, + { + "bbox": [ + 313, + 373, + 554, + 564 + ], + "type": "text", + "content": ". Results for varying distances of the cameras from the scene, depth of the scene, and image noise are in SM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 564, + 554, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 554, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 554, + 685 + ], + "type": "text", + "content": "Similar observations as on synthetic data can be made on real-world data. Tab. 1 shows results on different scenes from the PhotoTourism dataset [22]. As can be seen, the approximate solvers are not as accurate as the " + }, + { + "bbox": [ + 313, + 564, + 554, + 685 + ], + "type": "inline_equation", + "content": "5\\mathrm{pt}" + }, + { + "bbox": [ + 313, + 564, + 554, + 685 + ], + "type": "text", + "content": " solver when used without ENM. The gap between the " + }, + { + "bbox": [ + 313, + 564, + 554, + 685 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}(\\mathrm{M}) / 4\\mathrm{p}(\\mathrm{M}\\pm \\delta)" + }, + { + "bbox": [ + 313, + 564, + 554, + 685 + ], + "type": "text", + "content": " solvers and the " + }, + { + "bbox": [ + 313, + 564, + 554, + 685 + ], + "type": "inline_equation", + "content": "5\\mathrm{pt}" + }, + { + "bbox": [ + 313, + 564, + 554, + 685 + ], + "type": "text", + "content": " solver is for some scenes larger (Sacre Coeur, Trevi Fountain), and for some noticeably smaller (Reichstag, St. Peter's Square), showing that the performance of our solvers is scene-dependent. The " + }, + { + "bbox": [ + 313, + 564, + 554, + 685 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}(\\mathrm{A})" + }, + { + "bbox": [ + 313, + 564, + 554, + 685 + ], + "type": "text", + "content": " solver returns large errors. The accuracy of" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": " Outliers are only used in the ENM part to simulate that for approximate geometry outliers might be contained in the non-minimal samples." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21918" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 282, + 260 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 282, + 260 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 282, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 282, + 260 + ], + "type": "table", + "html": "
MED (°)\\( {20}^{\\text{th }} \\) perc. (°)
w/o ENM
Scene5p (E)4p (M)4p (M±δ)4p (A)5p (E)4p (M)4p (M±δ)4p (A)
Brandenburg Gate16.1419.0215.6975.327.229.837.4645.99
Buckingham Palace18.8721.1217.7466.198.0410.018.2340.59
Coloseum Exterior17.4023.5018.9366.676.4411.368.3440.82
Grand Place Brussels18.4020.2816.7869.318.0410.098.0744.45
Notre Dame Front Facade15.8223.9719.6171.606.4012.279.0148.39
P Palace of Westminster16.0417.6214.7269.384.997.986.4046.49
Puntheon Exterior22.6525.7121.3161.3310.6514.0910.9438.12
Reichstag12.1512.9910.2782.855.156.164.5760.12
Sacre Coeur11.8117.6114.3174.213.697.955.8743.72
St. Peter's Square17.8518.7515.4872.758.519.717.6644.60
Taj Mahal9.9511.208.5682.743.774.903.6364.27
Temple Nara Japan18.5721.5317.0164.977.6010.638.1632.21
Trevi Fountain20.9327.7722.8950.128.1013.8910.7831.32
w/ ENM
Scene5p (E)4p (M)4p (M±δ)4p (A)5p (E)4p (M)4p (M±δ)4p (A)
Brandenburg Gate14.5317.0514.3828.406.748.866.9316.96
Buckingham Palace16.7118.7616.0735.077.358.967.4619.72
Colosseum Exterior16.2121.4617.7034.906.3510.718.0421.98
Grand Place Brussels16.2117.9615.2331.537.278.857.2418.40
Notre Dume Front Facade15.0522.0818.4634.496.3011.658.7122.08
Palace of Westminster13.8314.9512.7429.064.687.135.8416.41
Puntheon Exterior21.1924.0320.3037.2110.3213.2810.5524.42
Reichstag9.3210.038.3617.694.405.063.979.39
Sacre Coeur10.7115.1912.6424.923.617.235.4913.41
St. Peter's Square16.0116.9414.3528.607.758.637.0315.68
Taj Mahal8.329.327.4315.853.344.103.187.72
Temple Nara Japan16.2419.3815.7429.976.839.197.3216.65
Trevi Fountain20.2226.4222.0939.678.0213.4310.5326.62
", + "image_path": "27157558dd71e82ebcd20ef3afe6158ae9fcffe7cba93221b9505acdddb7d799.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 269, + 291, + 280 + ], + "lines": [ + { + "bbox": [ + 58, + 269, + 291, + 280 + ], + "spans": [ + { + "bbox": [ + 58, + 269, + 291, + 280 + ], + "type": "text", + "content": "Table 1. Accuracy of two-view solvers on PhotoTourism scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "text", + "content": "the " + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}(\\mathrm{M}) / 4\\mathrm{p}(\\mathrm{M}\\pm \\delta)" + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "text", + "content": " solvers is not very far from the 5pt solver, with the " + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}(\\mathrm{M}\\pm \\delta)" + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "text", + "content": " solver even slightly outperforming the 5pt solver on some scenes (Reichstag, Taj Mahal). ENM helps to increase the precision of all solvers, especially " + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}(\\mathrm{A})" + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "text", + "content": ". Still, the " + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}(\\mathrm{A}) + \\mathrm{ENM}" + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "text", + "content": " solver returns quite high errors. However, in practice, all solvers are used inside RANSAC, where the current best pose is locally optimized on inliers. LO compensates for less accurate pose estimates and, as we show on real experiments for three as well as two views (see SM), it can suppress the errors of the " + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}(\\mathrm{A}) / 4\\mathrm{p}3\\mathrm{v}(\\mathrm{A})" + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "text", + "content": " -based solvers, making them comparable to the " + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}(\\mathrm{M}) / 4\\mathrm{p}3\\mathrm{v}(\\mathrm{M})" + }, + { + "bbox": [ + 54, + 304, + 295, + 472 + ], + "type": "text", + "content": " -based solvers. For the three view scenario, the performance of all of our solvers can be further improved by pose refinement (+R) and filtering (+F)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 473, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 605 + ], + "type": "text", + "content": "In the last set of experiments, we study the accuracy of the mean point correspondence. We sample 100 four-tuples of point correspondences consistent with the ground truth relative pose, i.e., inliers, for each image pair in scenes from the PhotoTourism dataset [22]. We use the first three correspondences to define the triangles in both images. Then, we establish correspondences between the mean of the triangle in one image and various points in the triangle in the second image. We express points in the second triangle via their barycentric coordinates and uniformly sample " + }, + { + "bbox": [ + 55, + 473, + 295, + 605 + ], + "type": "inline_equation", + "content": "19 \\times 19" + }, + { + "bbox": [ + 55, + 473, + 295, + 605 + ], + "type": "text", + "content": " barycentric coordinates " + }, + { + "bbox": [ + 55, + 473, + 295, + 605 + ], + "type": "inline_equation", + "content": "(a, b) \\in [0, 1]^2" + }, + { + "bbox": [ + 55, + 473, + 295, + 605 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 55, + 473, + 295, + 605 + ], + "type": "inline_equation", + "content": "a + b \\leq 1" + }, + { + "bbox": [ + 55, + 473, + 295, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "text", + "content": "Fig. 3 shows the results for the rotation error and the percentage of inliers consistent with the pose obtained with the " + }, + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}" + }, + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "text", + "content": " (M) solver on the St. Peter's Square scene. The optimum of the metrics is reached around the mean point of the triangles (the mean values of 2D Gaussians fitted to the results are very close to the mean point " + }, + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "inline_equation", + "content": "(0.3,0.3)" + }, + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "text", + "content": " of the triangles). A detailed description, the translation and the symmetric epipolar errors, and the results for more scenes are in SM. For all tested scenes, we observed a similar behavior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "Experiments on real data. We test the solvers on all scenes from the PhotoTourism dataset [22, 50] which provide ground truth poses and intrinsics via a COLMAP [49] reconstruction. In the results, we do not include the St. Peter's Square scene that we used for the validation of " + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": " and the number of refinement iterations (see SM). We also include results for the Cambridge Landmarks dataset [23] (except the Street scene, which is commonly not used due to issues with its ground truth) and Aachen Day-Night v1.1 [58]. For PhotoTourism and Aachen, we use the images in their original resolution. For Cambridge Landmarks, we resize them so that the larger side is " + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "inline_equation", + "content": "800\\mathrm{px}" + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": ". For each scene, we sample 5,000 random image triplets with at least 10 matches obtained with [9, 32] with at least " + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": " overlap [22]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 241, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 553, + 314 + ], + "type": "text", + "content": "Tab. 2 shows the results for PoseLib RANSAC with early termination and 5px epipolar threshold. Results for GC-RANSAC are in SM. With the suggested modifications (see Sec. 3.2) all proposed solvers outperform the state-of-the-art " + }, + { + "bbox": [ + 313, + 241, + 553, + 314 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 241, + 553, + 314 + ], + "type": "text", + "content": " (HC) solver [21] in terms of pose accuracy with many of the variants also achieving faster runtimes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": "Using filtering " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "(+F)" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " improves the run-time of RANSAC at the cost of a decrease in pose accuracy. Still, our " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "4p3v(M) + R + F" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " solver outperforms " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "4p3v(HC)" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " in terms of both accuracy and run-time. The " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "4p3v(M \\pm \\delta)" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "4p3v(M \\pm \\delta) + R" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " solvers clearly improve upon the " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "4p3v(M)" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " solvers, albeit at an increased run-time. Still the " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "4p3v(M \\pm \\delta) + R + F" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " solver provides, in general, the best speed-accuracy trade-off. The results also show that early non-minimal refinement " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "(+ENM)" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " is necessary for the " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "4p3v(A)" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " solver to perform well. For the other approximate solvers, this form of refinement also leads to improved accuracy at some cost to the runtime. Assuming that the " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "5pt + P3P" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " solver samples five points from which at least four have projections in all three cameras, the suggested pose refinement " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "(+R)" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " and filtering " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "(+F)" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " can also be applied to the baseline " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "5pt + P3P" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " by using the fourth (potentially the fifth) point in the third view. The suggested modifications improve the performance of the " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "5pt + P3P" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " solver, with the " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "5pt + P3P + R + F + ENM" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " solver having an accuracy similar to that of the " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "4p3v(M) + R + F + ENM" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " solver, however at slower run-times. Moreover, it is less accurate than the proposed " + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "inline_equation", + "content": "4p3v(M \\pm \\delta) + R + F + ENM" + }, + { + "bbox": [ + 313, + 315, + 553, + 578 + ], + "type": "text", + "content": " solver." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "text", + "content": "We also investigate the speed-accuracy trade-off of the solvers by running PoseLib RANSAC [27] for a set of fixed numbers of iterations. Runtimes are reported for 1 core of a " + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "inline_equation", + "content": "2\\mathrm{GHz}" + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "text", + "content": " Intel Xeon Gold 6338 CPU. As shown in Fig. 4, the proposed " + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "inline_equation", + "content": "(\\mathbb{M}\\pm \\delta) + \\mathbb{R} + \\mathbb{F}" + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "text", + "content": " solver consistently provides the best speed-accuracy trade-off both with and without ENM. " + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "text", + "content": " (A) " + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "inline_equation", + "content": "+\\mathbb{R} + \\mathbb{F} + \\mathbb{ENM}" + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "text", + "content": " provides a similar performance, typically beating " + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 580, + 555, + 700 + ], + "type": "text", + "content": " (HC) [21]. However, as shown in Fig. 4d for the St. Mary's Church scene [23], it may perform worse for some specific scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "The SM provides detailed ablation studies on the signif" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21919" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 553, + 175 + ], + "type": "image", + "image_path": "63e5c894136f5d0be36096f78de823ad5f38a80e8109eeb53ab1c4f30aa0d01b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 178, + 153, + 189 + ], + "lines": [ + { + "bbox": [ + 82, + 178, + 153, + 189 + ], + "spans": [ + { + "bbox": [ + 82, + 178, + 153, + 189 + ], + "type": "text", + "content": "(a) Phototourism [22]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 192, + 178, + 293, + 189 + ], + "lines": [ + { + "bbox": [ + 192, + 178, + 293, + 189 + ], + "spans": [ + { + "bbox": [ + 192, + 178, + 293, + 189 + ], + "type": "text", + "content": "(b) Cambridge Landmarks [23]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 178, + 419, + 189 + ], + "lines": [ + { + "bbox": [ + 315, + 178, + 419, + 189 + ], + "spans": [ + { + "bbox": [ + 315, + 178, + 419, + 189 + ], + "type": "text", + "content": "(c) Aachen Day-Night v1.1 [58]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 450, + 178, + 536, + 189 + ], + "lines": [ + { + "bbox": [ + 450, + 178, + 536, + 189 + ], + "spans": [ + { + "bbox": [ + 450, + 178, + 536, + 189 + ], + "type": "text", + "content": "(d) St. Mary's Church [23]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 243 + ], + "type": "text", + "content": "Figure 4. Speed-accuracy trade-off for (a) all scenes from PhotoTourism [22] except St. Peter's Square, (b) 5 scenes from Cambridge Landmarks [23], (c) Aachen Day-Night v1.1 [58], and (d) St. Mary's Church scene from the Cambridge Landmarks dataset [23]. We report the AUC@10° of the pose error and vary the number of PoseLib RANSAC iterations (100, 200, 500, 1000, 2000, 5000, 10,000) with a 5 px epipolar threshold. Runtimes are averaged over all image triplets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 253, + 555, + 401 + ], + "blocks": [ + { + "bbox": [ + 56, + 253, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 56, + 253, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 555, + 401 + ], + "type": "table", + "html": "
EstimatorPhotoTourism [22]Cambridge Landmarks [23]Aachen Day-Night v1.1 [58]
AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓AVG (°) ↓MED (°) ↓AUC@5↑@10↑@20↑Runtime ↓
4p3v (HC) [21]6.371.6262.4173.5381.9464.108.732.6747.9964.5876.6560.1112.433.5542.2855.2066.5567.26
5pt+P3P5.401.6062.9374.3083.0033.777.462.6548.3665.3377.7524.0410.523.3843.4356.7368.2853.34
5pt+P3P+ENM5.231.5663.7775.0383.5048.797.202.6348.7465.6878.0434.8210.593.4143.3256.6468.0671.61
5pt+P3P+R+F5.101.5064.9075.9084.1138.127.152.5849.4266.3578.5531.8210.373.3643.7357.1568.7154.34
5pt+P3P+R+F+ENM4.991.5065.1876.2084.3056.526.982.5849.3866.3178.5747.0710.203.3443.7557.1168.6376.86
4p3v (A)37.4825.1622.9128.5935.1616.5841.2321.7322.7531.4738.9813.3336.4019.5324.4131.7339.2432.04
4p3v (A)+ENM4.971.4865.4976.4784.5340.457.022.5949.3566.3278.5328.3510.353.3843.6256.9968.5562.44
4p3v (A)+R33.4519.9525.8632.2239.1616.3236.9811.5927.0336.7044.4812.4834.8418.2225.2732.7740.2929.56
4p3v (A)+R+F35.6523.4723.9929.9736.6411.1038.3414.4725.8935.1942.809.3536.7720.7623.9431.2338.5819.75
4p3v (A)+R+F+ENM4.981.4865.6276.5684.5532.376.992.5749.6266.5578.7223.4310.433.3843.5756.9568.5642.36
4p3v (M)6.021.7160.8672.5681.7335.188.362.7447.1564.0676.5925.1512.093.6241.7855.0666.6754.89
4p3v (M)+ENM5.281.5863.6174.9683.4648.837.212.6348.6965.5777.9534.9310.673.4343.3156.6168.0870.87
4p3v (M)+R5.491.5663.6674.8883.3741.527.692.6348.6965.6777.9930.7611.173.4643.2856.3967.8860.40
4p3v (M)+R+F5.481.5463.9975.1483.5130.887.752.6248.7765.7177.9822.7211.163.4843.0456.2867.7342.38
4p3v (M)+R+F+ENM5.011.5065.1876.1684.2844.516.992.5849.4366.3178.5632.4810.253.3943.7457.2168.8058.08
4p3v (M±δ)5.551.6861.8673.7382.7883.707.652.6748.1665.2577.7659.2311.133.4742.7456.1167.74125.02
4p3v (M±δ)+ENM4.991.5664.0275.5184.01125.666.882.6049.1166.1378.5189.0510.263.3543.6656.9968.55175.54
4p3v (M±δ)+R4.861.4765.9777.0085.01100.617.102.5649.7966.8979.0873.9410.373.3443.9357.2668.82138.53
4p3v (M±δ)+R+F4.921.4765.8576.8784.9071.737.192.5649.7166.7578.9552.8410.523.3643.7957.1768.7392.89
4p3v (M±δ)+R+F+ENM4.661.4666.1277.1485.15112.606.652.5549.8766.9579.2081.989.963.3244.0257.4268.97139.19
", + "image_path": "442b33bf3f54da311f50ba99ac9fc2ee26501849307ece6ab58ac2408e40dd8c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "type": "text", + "content": "Table 2. Results for different solvers implemented in the PoseLib framework [27] on 12 scenes from PhotoTourism [22], 5 scenes from Cambridge Landmarks [23] and Aachen Day-Night v1.1 [58]. We mark the best and second best results. Runtimes are reported in ms for the whole RANSAC with early termination (0.9999 confidence, minimum 100 iterations) and epipolar threshold set to 5 px." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 464, + 276, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 276, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 276, + 476 + ], + "type": "text", + "content": "icance of each individual modification " + }, + { + "bbox": [ + 55, + 464, + 276, + 476 + ], + "type": "inline_equation", + "content": "(+R / + C / + ENM)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "text", + "content": "Limitations. As discussed, the accuracy of the proposed approximate solvers is scene dependent. However, after using the proposed ENM refitting, the scene-dependency is negligible. Especially for the " + }, + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "inline_equation", + "content": "(\\mathbb{M}\\pm \\delta)" + }, + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "text", + "content": " +ENM solver, we have not noticed performance drops for some specific scene geometries or camera configurations. While the two-view variants of our solvers " + }, + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}" + }, + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "inline_equation", + "content": "(\\mathbb{M}) / 4\\mathrm{p}" + }, + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "inline_equation", + "content": "(\\mathbb{M}\\pm \\delta)" + }, + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "text", + "content": " provide very similar results to the 5pt solver in two views, in this case they do not outperform the 5pt solver. In the three-view scenario, the better performance is achieved thanks to the proposed pose refinement " + }, + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "inline_equation", + "content": "(+\\mathbf{R})" + }, + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "text", + "content": " and filtering " + }, + { + "bbox": [ + 54, + 480, + 295, + 612 + ], + "type": "inline_equation", + "content": "(+\\mathbf{F})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "content": "The proposed ENM refitting cannot be applied to the " + }, + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 55, + 613, + 295, + 685 + ], + "type": "text", + "content": " (HC) solver [21], since this solver estimates the pose of all three cameras together. It is theoretically applicable to the solver proposed in [36]. However, here ENM would have needed to be run on hundreds of candidate poses corresponding to tens-to-hundreds of sampled epipoles. This" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 464, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 548 + ], + "type": "text", + "content": "is because for one fixed point on the curve of epipoles, the error of the sampled epipole is not bounded since the true epipole can lie anywhere on the curve, even outside the image. Thus, ENM might avoid local minima present in the second step of the original solver, but would significantly slow down the solver in the first step, making it much slower than our " + }, + { + "bbox": [ + 313, + 464, + 555, + 548 + ], + "type": "inline_equation", + "content": "4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 464, + 555, + 548 + ], + "type": "inline_equation", + "content": "(\\mathbf{M}) + \\mathsf{ENM} / 4\\mathrm{p}3\\mathrm{v}" + }, + { + "bbox": [ + 313, + 464, + 555, + 548 + ], + "type": "inline_equation", + "content": "(\\mathbf{M}\\pm \\delta)" + }, + { + "bbox": [ + 313, + 464, + 555, + 548 + ], + "type": "text", + "content": " +ENM solvers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "We consider the highly challenging problem of relative pose estimation of three calibrated cameras from four correspondences. We propose novel approaches that solve the problem by utilizing solvers based on approximate geometry. The best performing solver uses a simple, yet novel strategy by using mean coordinates of three input points and points in their vicinity as an approximate " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " correspondences. Extensive experiments show that our solvers achieve state-of-the-art performance on a large variety of real scenes. At the same time, our solvers are simple to implement, especially compared to the current state-of-the-art [21]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "This weakness also applies to [21], since the scene needs to be similar enough to the training scenes for the MLP-based classifier to work well." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21920" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 301 + ], + "type": "text", + "content": "Acknowledgements. This work was funded by the Czech Science Foundation (GACR) JUNIOR STAR Grant No. 22-23183M (supporting C.T., Y.D., and Z.K.), the Grant Agency of the Czech Technical University in Prague grant no. SGS23/173/OHK3/3T/13 (supporting C.T.), the EU NextGenerationEU through the Recovery and Resilience Plan for Slovakia under the project \"InnovAite Slovakia, Illuminating Pathways for AI-Driven Breakthroughs\" No. 09I02-03-V01-00029, the TERAIS project - a Horizon-Widera-2021 program of the European Union under the Grant agreement number 101079338 (supporting V.K. and Z.B.H.), and the EU Horizon 2020 project RICAIP grant agreement No. 857306 (supporting T.S.). Part of the research results was obtained using the computational resources procured in the national project National competence centre for high performance computing project code: 311070AKF2, funded by European Regional Development Fund, EU Structural Funds Informatization of society, Operational Program Integrated Infrastructure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 311, + 115, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 311, + 115, + 323 + ], + "spans": [ + { + "bbox": [ + 57, + 311, + 115, + 323 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 331, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 331, + 294, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 331, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 61, + 331, + 294, + 362 + ], + "type": "text", + "content": "[1] Chris Aholt and Luke Oeding. The ideal of the trifocal variety. Mathematics of Computation, 83(289):2553-2574, 2014. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 365, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 365, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 61, + 365, + 295, + 397 + ], + "type": "text", + "content": "[2] D. Barath and J. Matas. Graph-Cut RANSAC. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6733-6741, 2018. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 399, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 399, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 61, + 399, + 295, + 442 + ], + "type": "text", + "content": "[3] Martin Bujnak, Zuzana Kukelova, and Tomas Pajdla. A general solution to the p4p problem for camera with unknown focal length. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 445, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 294, + 476 + ], + "type": "text", + "content": "[4] Robert Castle, Georg Klein, and David W. Murray. Videorelate localization in multiple maps for wearable augmented reality. In ISWC, 2008. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 478, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 295, + 543 + ], + "type": "text", + "content": "[5] Chiang-Heng Chien, Hongyi Fan, Ahmad Abdelfattah, Elias Tsigaridas, Stanimire Tomov, and Benjamin Kimia. Gpubased homotopy continuation for minimal problems in computer vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15765-15776, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 294, + 578 + ], + "type": "text", + "content": "[6] Ondrej Chum, Jiří Matas, and Josef Kittler. Locally optimized ransac. In Pattern Recognition, pages 236-243. Springer Berlin Heidelberg, 2003. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 579, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 295, + 623 + ], + "type": "text", + "content": "[7] Andrea Porfiri Dal Cin, Timothy Duff, Luca Magri, and Tomas Pajdla. Minimal perspective autocalibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5064-5073, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 656 + ], + "type": "text", + "content": "[8] Toby Collins and Adrien Bartoli. Infinitesimal plane-based pose estimation. International Journal of Computer Vision, 109:252-286, 2014. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "text", + "content": "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 224-236, 2018. 6, 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[10] Yaqing Ding, Chiang-Heng Chien, Viktor Larsson, Karl Åström, and Benjamin Kimia. Minimal solutions to generalized three-view relative pose problem. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8156-8164, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "type": "text", + "content": "[11] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. Plmp-point-line minimal problems in complete multiview visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1675-1684, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 186, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 554, + 229 + ], + "type": "text", + "content": "[12] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomás Pajdla. " + }, + { + "bbox": [ + 317, + 186, + 554, + 229 + ], + "type": "inline_equation", + "content": "\\mathrm{Pl}_1\\mathfrak{p}" + }, + { + "bbox": [ + 317, + 186, + 554, + 229 + ], + "type": "text", + "content": " - point-line minimal problems under partial visibility in three views. In ECCV (26), pages 175-192. Springer, 2020. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 231, + 554, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 554, + 305 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 554, + 305 + ], + "type": "text", + "content": "[13] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H. Regan, David da Costa de Pinho, Elias P. Tsigaridas, Charles W. Wampler, Jonathan D. Hauenstein, Peter J. Giblin, Benjamin B. Kimia, Anton Leykin, and Tomás Pajdla. TRPLP - trifocal relative pose from lines at points. In CVPR, pages 12070-12080. Computer Vision Foundation / IEEE, 2020. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 308, + 554, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 554, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 554, + 351 + ], + "type": "text", + "content": "[14] M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6): 381-395, 1981. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "text", + "content": "[15] Liran Goshen and Ilan Shimshoni. Balanced exploration and exploitation model search for efficient epipolar geometry estimation. In Computer Vision - ECCV 2006, pages 151-164, Berlin, Heidelberg, 2006. Springer Berlin Heidelberg. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 399, + 554, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 454 + ], + "type": "text", + "content": "[16] Christian Häne, Lionel Heng, Gim Hee Lee, Friedrich Fraun-dorfer, Paul Furgale, Torsten Sattler, and Marc Pollefeys. 3d visual perception for self-driving cars using a multi-camera system: Calibration, mapping, localization, and obstacle detection. Image and Vision Computing, 68:14-27, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 456, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 476 + ], + "type": "text", + "content": "[17] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision, 2006. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 478, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 511 + ], + "type": "text", + "content": "[18] Joel A. Hesch and Stergios I. Roumeliotis. A direct least-squares (dls) method for pnp. In 2011 International Conference on Computer Vision, pages 383-390, 2011. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 513, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 554, + 555 + ], + "type": "text", + "content": "[19] R.J. Holt and A.N. Netravali. Uniqueness of solutions to three perspective views of four points. IEEE Transactions on Pattern Analysis and Machine Intelligence, 17(3):303-307, 1995. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 557, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 554, + 600 + ], + "type": "text", + "content": "[20] Radu Horaud, Fadi Dornaika, and Bart Lamiroy. Object pose: The link between weak perspective, parapspective, and full perspective. Int. J. Comput. Vision, 22(2):173-189, 1997. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 602, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 555, + 656 + ], + "type": "text", + "content": "[21] Petr Hruby, Timothy Duff, Anton Leykin, and Tomas Pajdla. Learning to solve hard minimal problems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5532-5542, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "type": "text", + "content": "[22] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 6, 7, 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21921" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[23] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Proceedings of the IEEE international conference on computer vision, pages 2938-2946, 2015. 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 163 + ], + "type": "text", + "content": "[24] J. Kileel. Minimal problems for the calibrated trifocal variety. SIAM Journal on Applied Algebra and Geometry, 1(1): 575-598, 2017. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 296, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 296, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 296, + 208 + ], + "type": "text", + "content": "[25] Z. Kukelova and T. Pajdla. A minimal solution to the autocalibration of radial distortion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2007), 2007. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 210, + 296, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 296, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 296, + 265 + ], + "type": "text", + "content": "[26] Zuzana Kukelova, Martin Bujnak, and Tomas Pajdla. Real-time solution to the absolute pose problem with unknown radial distortion and focal length. In Proceedings of the IEEE International Conference on Computer Vision, pages 2816-2823, 2013. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 266, + 296, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 296, + 289 + ], + "type": "text", + "content": "[27] Viktor Larsson. PoseLib - Minimal Solvers for Camera Pose Estimation, 2020. 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 291, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 296, + 335 + ], + "type": "text", + "content": "[28] Viktor Larsson, Kalle Åström, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Computer Vision and Pattern Recognition (CVPR), pages 820-829, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 337, + 296, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 296, + 381 + ], + "type": "text", + "content": "[29] Viktor Larsson, Torsten Sattler, Zuzana Kukelova, and Marc Pollefeys. Revisiting radial distortion absolute pose. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1062-1071, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 296, + 426 + ], + "type": "text", + "content": "[30] Karel Lebeda, Jiri Matas, and Ondrej Chum. Fixing the locally optimized ransac. In Proceedings of the British Machine Vision Conference 2012, pages 4961-4974. British Machine Vision Association, 2012. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 428, + 296, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 296, + 472 + ], + "type": "text", + "content": "[31] S. Leonardos, R. Tron, and K. Daniilidis. A metric parametrization for trifocal tensors with non-colinear pinholes. In Computer Vision and Pattern Recognition (CVPR), pages 259-267. IEEE, 2015. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 474, + 296, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 296, + 517 + ], + "type": "text", + "content": "[32] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In International Conference on Computer Vision (ICCV), pages 17627–17638, 2023. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 519, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 519, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 519, + 296, + 563 + ], + "type": "text", + "content": "[33] H. Christopher Longuet-Higgins. A method of obtaining the relative positions of 4 points from 3 perspective projections. In BMVC91, pages 86-94, London, 1991. Springer London. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 565, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 296, + 599 + ], + "type": "text", + "content": "[34] Evgeniy V. Martyushev. On some properties of calibrated trifocal tensors. Journal of Mathematical Imaging and Vision, 58(2):321-332, 2017. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 600, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 296, + 633 + ], + "type": "text", + "content": "[35] D. Nister. An efficient solution to the five-point relative pose problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26(6):756-770, 2004. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "type": "text", + "content": "[36] D. Nister and F. Schaffalitzky. Four points in two or three calibrated views: Theory and practice. International Journal of Computer Vision, 67(2):211-231, 2006. 1, 2, 3, 4, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[37] Denis Oberkampf, Daniel F DeMenton, and Larry S Davis. Iterative pose estimation using coplanar feature points. Computer Vision and Image Understanding, 63(3):495-511, 1996. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[38] Magnus Oskarsson, Andrew Zisserman, and Kalle Åström. Minimal projective reconstruction for combinations of points and lines in three views. Image and Vision Computing, 22 (10):777-785, 2004. British Machine Vision Computing 2002. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 173 + ], + "type": "text", + "content": "[39] Michal Perdoch, Jiri Matas, and Ondrej Chum. Epipolar geometry from two correspondences. In 18th International Conference on Pattern Recognition (ICPR'06), pages 215-219. IEEE, 2006. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 555, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 219 + ], + "type": "text", + "content": "[40] Mikael Persson and Klas Nordberg. Lambda twist: An accurate fast robust perspective three point (p3p) solver. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318-332, 2018. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 285 + ], + "type": "text", + "content": "[41] James Pritts, Ondrej Chum, and Jiri Matas. Approximate models for fast and accurate epipolar geometry estimation. In 28th International Conference on Image and Vision Computing New Zealand, IVCNZ 2013, Wellington, New Zealand, November 27-29, 2013, pages 106-111. IEEE, 2013. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 342 + ], + "type": "text", + "content": "[42] James Pritts, Zuzana Kukelova, Viktor Larsson, and Ondrej Chum. Radially-distorted conjugate translations. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 1993-2001. IEEE Computer Society, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 386 + ], + "type": "text", + "content": "[43] Long Quan. Invariants of six points and projective reconstruction from three uncalibrated images. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 17(1): 34-46, 1995. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 555, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 430 + ], + "type": "text", + "content": "[44] L. Quan, B. Triggs, and B. Mourrain. Some results on minimal euclidean reconstruction from four points. Journal of Mathematical Imaging and Vision, 24(3):341-348, 2006. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 432, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 555, + 476 + ], + "type": "text", + "content": "[45] R. Raguram, O. Chum, M. Pollefeys, J. Matas, and J.-M. Frahm. USAC: A universal framework for random sample consensus. IEEE Transactions on Pattern Recognition and Machine Intelligence, 35(8):2022-2038, 2013. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "type": "text", + "content": "[46] Volker Rodehorst. Evaluation of the metric trifocal tensor for relative three-view orientation. In Digital Proceedings, International Conference on the Applications of Computer Science and Mathematics in Architecture and Civil Engineering: July 20 - 22 2015, Bauhaus-University Weimar, 2017. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 533, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 555, + 578 + ], + "type": "text", + "content": "[47] T. Sattler, B. Leibe, and L. Kobbelt. Efficient & Effective Prioritized Matching for Large-Scale Image-Based Localization. IEEE Transactions on Pattern Recognition and Machine Intelligence, 39(9):1744-1756, 2016. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 578, + 555, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 555, + 601 + ], + "type": "text", + "content": "[48] D. Scaramuzza and F. Fraundorfer. Visual odometry [tutorial]. IEEE Robot. Automat. Mag., 18(4):80-92, 2011. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 555, + 645 + ], + "type": "text", + "content": "[49] Johannes L. Schonberger and Jan-Michael Frahm. Structure-From-Motion Revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104–4113, 2016. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 555, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 680 + ], + "type": "text", + "content": "[50] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 555, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 715 + ], + "type": "text", + "content": "[51] N. Snavely, S. M. Seitz, and R. Szeliski. Modeling the world from internet photo collections. International Journal Computer Vision, 80(2):189-210, 2008. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21922" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 342 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[52] Andrew J Sommese, Andrew J Sommese, and Charles W Wampler. The numerical solution of systems of polynomials arising in engineering and science. World Scientific Publishing Co. Pte. Ltd, Singapore, 2005. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[53] H. Stewenius, D. Nister, F. Kahl, and F. Schaffalitzky. A minimal solution for relative pose with unknown focal length. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2005), 2005. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "text", + "content": "[54] H. Stewenius, C. Engels, and D. Nister. Recent developments on direct relative orientation. ISPRS J. of Photogrammetry and Remote Sensing, 60:284-294, 2006. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "text", + "content": "[55] P. H. S. Torr and A. Zisserman. Robust parameterization and computation of the trifocal tensor. Image and Vision Computing, 15:591-605, 1997. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[56] G. Xu and N. Sugimoto. A linear algorithm for motion from three weak perspective images using euler angles. IEEE Transactions on Pattern Analysis and Machine Intelligence, 21(1):54-57, 1999. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 297 + ], + "type": "text", + "content": "[57] Zhengyou Zhang. Weak Perspective Projection, pages 877-883. Springer US, Boston, MA, 2014. 2, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[58] Zichao Zhang, Torsten Sattler, and Davide Scaramuzza. Reference pose generation for long-term visual localization via learned features and view synthesis. International Journal of Computer Vision, 129(4):821-844, 2021. 7, 8" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21923" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/f70c89aa-2e9b-4112-89fc-f4837e98e688_content_list.json b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/f70c89aa-2e9b-4112-89fc-f4837e98e688_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8afd160c68b0f2a2dc5f42e005831daff64f2061 --- /dev/null +++ b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/f70c89aa-2e9b-4112-89fc-f4837e98e688_content_list.json @@ -0,0 +1,1400 @@ +[ + { + "type": "text", + "text": "Precise Event Spotting in Sports Videos: Solving Long-Range Dependency and Class Imbalance", + "text_level": 1, + "bbox": [ + 102, + 128, + 893, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sanchayan Santra $^{1}$ Vishal Chudasama $^{1}$ Pankaj Wasnik $^{1*}$ Vineeth N Balasubramanian $^{2}$ $^{1}$ Sony Research India $^{2}$ Indian Institute of Technology Hyderabad", + "bbox": [ + 125, + 202, + 870, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sanchayan.santra, vishal.chudasamal, pankaj.wasnik}@sony.com, vineethnb@cse.iith.ac.in", + "bbox": [ + 114, + 241, + 880, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Precise Event Spotting (PES) aims to identify events and their class from long, untrimmed videos, particularly in sports. The main objective of PES is to detect the event at the exact moment it occurs. Existing methods mainly rely on features from a large pre-trained network, which may not be ideal for the task. Furthermore, these methods overlook the issue of imbalanced event class distribution present in the data, negatively impacting performance in challenging scenarios. This paper demonstrates that an appropriately designed network, trained end-to-end, can outperform state-of-the-art (SOTA) methods. Particularly, we propose a network with a convolutional spatial-temporal feature extractor enhanced with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) and a long-range temporal module. The ASTRM enhances the features with spatiotemporal information. Meanwhile, the long-range temporal module helps extract global context from the data by modeling long-range dependencies. To address the class imbalance issue, we introduce the Soft Instance Contrastive (Soft-IC) loss that promotes feature compactness and class separation. Extensive experiments show that the proposed method is efficient and outperforms the SOTA methods, specifically in more challenging settings.", + "bbox": [ + 88, + 321, + 485, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 700, + 220, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event spotting methods locate the occurrence of events in long, untrimmed videos. An event is typically defined by a starting and ending time, but this may not always be feasible, particularly in sports, where an event is a specific occurrence that occurs at a particular point in time, as dictated by the game's rules. For instance, a \"Goal\" in soccer is scored when the ball crosses the goal line, while a \"Boundary\" in cricket is made when the ball touches the boundary line. This process is referred to as Precise Event Spotting (PES), where PES methods aim to predict an event at the", + "bbox": [ + 89, + 724, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a145267ea273f846cb45b196f296fec5d4b7a2097d67da6ac81f028468589c81.jpg", + "image_caption": [ + "Figure 1. Temporal dependency of \"Shots on target\" event: The event is marked when the shot is taken, but whether the shot is at the target can only be known by looking at future frames. Here the shot is on target because the ball is saved by the goal keeper." + ], + "image_footnote": [], + "bbox": [ + 516, + 287, + 901, + 439 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4430d79e5bcdf58e0fabb04ee2b14382a55280ac93b69dbe00eacfcf07b0e97b.jpg", + "image_caption": [ + "Figure 2. Per class score analysis on a few classes of SoccerNet V2 dataset in tight setting. The proposed method outperforms state-of-the-art (SOTA) methods especially for classes with less number of samples." + ], + "image_footnote": [], + "bbox": [ + 517, + 513, + 903, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "exact frame it occurs (e.g., goal, boundary). This is crucial for generating game summaries, finding specific events, and editing lengthy matches. However, it poses significant challenges, requiring a deep understanding of the game and its key events.", + "bbox": [ + 511, + 762, + 906, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing PES methods mainly face two challenges. First is the long-range temporal dependency that arises due to the nature of events. To predict the event frames, observing just the nearby frames does not suffice; several other potentially", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 109, + 886, + 238, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3163", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "distant frames need to be observed to predict the event correctly. For example, \"Shots on target\" and \"Shots off target\" events of Soccer games are marked when the shot is taken. Still, the actual class is decided based on whether the ball went toward the goal or away from it (see Figure 1). This can be only detected if long-range dependencies are captured correctly. Second, there is an inherent imbalance in the number of samples among the classes (see Figure 2). Some events naturally occur more frequently than others. For example, in a soccer game, a \"Ball out of play\" is likely to occur more than a \"Yellow card\"; a \"Red-card\" and \"Yellow $\\rightarrow$ Red card\" are expected to occur even less often. Therefore, the same trend can be observed in datasets created using broadcast videos. Due to these issues, existing event spotting methods [6, 14] perform poorly in underrepresented classes such as \"Red card\" & \"Yellow $\\rightarrow$ Red card.\" This can be visualized in Figure 2. Handling these issues and accurately performing PES is an active area of research.", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the long-range temporal dependency issue, recent SOTA methods [24, 29] prefer to use the feature extracted using the pre-trained network of [35]. These methods also use transformer-based temporal modules due to their better sequence modeling capability in a global context. However, as stated in [18, 32], transformers also encounter issues with long videos. Self-attention struggles to learn appropriate weights for long videos due to the increase in the size of the weight matrix with the number of frames being processed. On the other hand, E2ESpot [14] proposed an end-to-end trainable feature extractor and a GRU-based module for temporal modeling. Recently proposed T-DEED [30] improves the performance of E2ESpot via temporal-discriminability enhancing encoder-decoder (T-DEED) module. In our proposed method, we handle this issue by enhancing the extracted features with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) and GRU-based temporal module.", + "bbox": [ + 89, + 364, + 482, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The class imbalance issue has not received sufficient attention in the literature on PES, resulting in lower performance on real-world datasets. The recent method ASTRA [29] is one of the few attempts to address this problem through a modified \"mixup\" augmentation. Although standard machine learning approaches, such as focal loss [16], are available, they are not suitable for every task, as this paper demonstrates. To address this challenge, we introduce a novel loss function called Soft Instance Contrastive (Soft-IC) loss. This loss function aims to mitigate class imbalance by promoting compact feature representations and improving class separability.", + "bbox": [ + 89, + 641, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primary goal of this paper is to address these issues and improve the accuracy of PES. To this end, we propose an end-to-end trainable network with a CNN-based spatiotemporal feature extractor and a GRU-based long-term temporal module. The spatio-temporal feature extractor in-", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cludes our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) to augment the features with spatial and temporal information. We chose to train the model end-to-end because it has been shown to perform better than a pre-trained frozen feature extractor [17, 26]. We have also employed the Adaptive Sharpness-Aware Minimization (ASAM) [15] to enable the trained model to perform well on unseen data. Through experiments, we demonstrate that the proposed framework obtains superior performance compared to SOTA methods, especially in tight settings where long-range dependency and class imbalance issues are prominent (see Figure 2). Our key contributions can be summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an end-to-end framework using a CNN-based spatio-temporal feature extractor and GRU-based temporal network to achieve precise event spotting while addressing long-range dependency and class imbalance.", + "- We introduce ASTRM, which enhances the features with spatial and temporal information.", + "- We also introduce Soft-IC loss to enforce compact features and improve class separability, addressing the class imbalance issue." + ], + "bbox": [ + 511, + 287, + 903, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 436, + 661, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The concept of precise sports event spotting first appeared in Einfalt et al. [7], where they considered temporally precise detection of various events related to long and triple jump. Later, SoccerNet [10] proposed a large-scale dataset of soccer videos to identify specific events or actions at exact timestamps. The work by [14] expanded this to precise event spotting (PES), introducing significant differences in the necessary precision of predictions. PES is constrained to the tolerance of a few frames only. Recently, several PES methods [6, 14, 24, 27, 29, 30, 35] have been proposed that focus on various sports videos. E2E-Spot [14] employed a CNN-based backbone with a Gated Shift Module (GSM) [25] for temporal modeling and a GRU [3] for producing per-frame prediction. Even though it is effective on the Tennis and Figure Skating datasets, it struggles on more challenging SoccerNet V2 dataset [5] due to its simplistic temporal information extraction scheme using GSM.", + "bbox": [ + 511, + 462, + 903, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon E2E-Spot [14], T-DEED [30] proposed improvements to address issues arising from tight tolerances in fast-paced sporting events and achieved SOTA performance on the Figure Skating and Figure Diving datasets. The method of [35] revamped the feature extraction using a group of five networks pre-trained on Kinetics-400 and fine-tuned on SoccerNet V2. Several other methods use the features extracted by this method instead of raw video frames. For instance, Soares et al. [24] utilized these features and developed a U-Net- and transformer-based trunk for making predictions. ASTRA [29] utilizes the same features but incorporates audio features for complementary in", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3164", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "formation. In contrast, COMEDIAN [6] operates in 3 stages to train their network on raw frames: 1) training only the spatial part; 2) training the spatial and temporal parts using knowledge distillation; 3) final tuning to optimize the classification scores. Recently, UGLF [27] proposed a unifying global-local (UGL) module to disentangle the scene content into global features and locally relevant scene entity features in conjunction with a Vision-Language model.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above-mentioned methods mainly addressed the long-range temporal dependency issue by utilizing pretrained features and transformers. They have not adapted their feature extractors to capture spatial and temporal information effectively. Additionally, these methods did not specifically deal with the class imbalance, which led to poor performance in underrepresented classes. In our framework, we propose ASTRM, which helps the feature extractor to augment the features with relevant spatial and temporal information. Furthermore, we introduce Soft-IC loss to address the class imbalance issue and employ the ASAM to train the proposed framework properly. With these advancements, the proposed framework outperforms existing event-spotting methods on various sports datasets, particularly in settings with tighter tolerances.", + "bbox": [ + 89, + 212, + 482, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Methodology", + "text_level": 1, + "bbox": [ + 89, + 450, + 308, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents our methodology in detail, starting with the problem statement and then presenting the details of the proposed modules as follows.", + "bbox": [ + 89, + 476, + 482, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement", + "text_level": 1, + "bbox": [ + 89, + 529, + 274, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The task of precise event spotting is defined as follows: Given a video with $N$ frames $\\mathbf{x}_1, \\mathbf{x}_2, \\ldots, \\mathbf{x}_N$ and a set of possible events $\\{c_1, \\ldots, c_K\\}$ , the goal is to predict frames where an event occurs along with its class: $(t, \\hat{y}_t) \\in \\mathbb{N} \\times \\{c_1, \\ldots, c_K\\}$ . A prediction is deemed correct if it falls within $\\delta$ frames of the actual occurrence and accurately identifies the event class. Since it is not possible to process the entire video at once, the processing is done in chunks of a predefined size. Each chunk (or clip), represented as $\\mathbf{X} \\in \\mathbb{R}^{C \\times T_s \\times H \\times W}$ , consists of $T_s$ frames. We compute predictions for each frame, represented as $\\mathbf{Y} \\in \\mathbb{R}^{T_s \\times K}$ , where $K$ is the number of classes. As the video is processed in segments, the predictions from each chunk of $T_s$ frames are aggregated and refined with a non-maximal suppression to get the final sparse prediction.", + "bbox": [ + 89, + 551, + 482, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Methodology", + "text_level": 1, + "bbox": [ + 89, + 786, + 228, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 3 shows the network design of our proposed framework, which takes a set of input frames and classifies whether each frame contains an event or not. The framework comprises three main components: the spatiotemporal feature extractor, the temporal block, and the classifier. In the spatio-temporal feature extractor, we", + "bbox": [ + 89, + 809, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "have selected the RegNetY [21] as the feature extractor due to its efficiency, surpassing commonly used networks like ResNet. This network is further enhanced with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM). This module is integrated into each bottleneck block, refining the features by incorporating both spatial and temporal information. Both of this information is necessary to correctly identify events, as validated by the experiments. The refined features are then passed to the temporal block, which enriches them with long-term temporal information. This long-term information is crucial because, in many situations, information from temporally distant frames is essential to accurately classify a frame. The spatio-temporal feature extractor alone can not accomplish this task. Finally, the feature maps are forwarded to the classifier, which is a simple linear layer that outputs the classification scores. Furthermore, we introduce the Soft Instance Contrastive (Soft-IC) loss to effectively address the imbalance among the class samples. This loss function operates by compacting the features and enhancing their separability. The key modules of the proposed framework are detailed in subsequent subsections.", + "bbox": [ + 511, + 90, + 903, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Adaptive Spatio-Temporal Refinement Module:", + "text_level": 1, + "bbox": [ + 511, + 441, + 898, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This module enhances the extracted features by incorporating both spatial and temporal information. It comprises three blocks: Local Spatial, Local Temporal, and Global Temporal. We include the Local Spatial block because local spatial features also offer important cues about temporal information. Relying solely on temporal information can overlook these valuable insights. The spatial component is inspired by [28], while the temporal aspect is adapted from [19]. The refinement is done as follows:", + "bbox": [ + 511, + 465, + 903, + 602 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi (\\mathbf {x}) = \\left(\\left(\\mathbf {x} \\odot \\left(1 + \\mathcal {F} _ {s} (\\mathbf {x})\\right)\\right) \\odot \\left(1 + \\mathcal {F} _ {t} (\\mathbf {x})\\right)\\right) * \\mathcal {G} _ {t} (\\mathbf {x}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 614, + 903, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ is the input feature to be enhanced. $\\mathcal{F}_s, \\mathcal{F}_t,$ and $\\mathcal{G}_t$ denote the operation corresponding to the local spatial, local temporal, and global temporal block, respectively. $\\odot$ denotes Hadamard Product, while $*$ denotes convolution operation. To elaborate, the input feature is weighted with information from the local spatial block and a residual connection, resulting in features enhanced with local spatial information. These enhanced features are further refined with local temporal information. Finally, spatially and temporally enhanced features are then convolved with the kernel obtained from the global temporal block to make local enhancements using \"global\" information. This dynamic kernel approach simplifies the network while maintaining effectiveness. Please refer to Figure 3 for a visual representation.", + "bbox": [ + 511, + 643, + 905, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the Local Spatial block, the input features are pooled using Max (i.e., $f_{MaxPool}$ ) and Average (i.e., $f_{AvgPool}$ )", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3165", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e46dd065db54d0faa2035b4b74c2249192712b2832c5b8d92f49c1880a13fb4a.jpg", + "image_caption": [ + "Figure 3. Network design of the proposed framework. The framework composed of a spatio-temporal feature extractor and a temporal block for capturing long-range dependency before the classifier. In each bottleneck block we add ASTRM after the first conv. ASTRM further enhances the features with local spatial, local temporal and global temporal information. The network is trained with Soft-IC loss in addition to the classification loss to handle class imbalance." + ], + "image_footnote": [], + "bbox": [ + 109, + 95, + 890, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pooling operations along the channel dimension. The pooled features are then concatenated and convolved with a $7 \\times 7$ filter (i.e., $f_{conv}^{7 \\times 7}$ ). The pooling operation on the channel axis helps emphasize informative regions. The subsequent convolution operation filters out non-informative portions from the features. Finally, the output is passed through a Sigmoid function (i.e., $\\sigma(\\cdot)$ ) to obtain the \"spatial\" feature weights (i.e., $\\mathcal{F}_s(\\mathbf{x})$ ) as follow:", + "bbox": [ + 89, + 474, + 483, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {s} (\\mathbf {x}) = \\sigma \\left(f _ {c o n v} ^ {7 \\times 7} \\left(\\left[ f _ {A v g P o o l} (\\mathbf {x}); f _ {M a x P o o l} (\\mathbf {x}) \\right]\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 599, + 482, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the Local Temporal block, we use 3D convolution operations, namely $f_{conv}^{3 \\times 1 \\times 1}$ and $f_{conv}^{1 \\times 1 \\times 1}$ , followed by the Sigmoid function, $\\sigma(\\cdot)$ , with a focus on the temporal dimension. Spatially pooled features are not used in this block because they do not account for the different temporal changes at various spatial locations. The locally temporal enhanced features are defined as below:", + "bbox": [ + 89, + 625, + 483, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {t} (\\mathbf {x}) = \\sigma \\left(f _ {\\text {c o n v}} ^ {1 \\times 1 \\times 1} \\left(\\mathrm {B N} \\left(\\operatorname {R e L U} \\left(f _ {\\text {c o n v}} ^ {3 \\times 1 \\times 1} (\\mathbf {x})\\right)\\right)\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 734, + 482, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the Global Temporal block, we adaptively create convolution kernels to improve the feature maps by considering temporal information across frames. First, the features are pooled in the spatial dimension using a global average pooling operation (i.e., $f_{GAP}$ ), and then fully connected layers (i.e., $f_{FC}$ ) are used to compute the kernel with a global view of the data with a softmax ( $S$ ) at the end. Mathematically, the kernel is formulated as below:", + "bbox": [ + 89, + 758, + 483, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {t} (\\mathbf {x}) = \\mathcal {S} \\left(f _ {F C} \\left(f _ {F C} \\left(f _ {G A P} (\\mathbf {x})\\right)\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Temporal Block:", + "text_level": 1, + "bbox": [ + 511, + 474, + 684, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The spatio-temporal feature extraction stage computes features containing local spatial, local temporal, and location-independent temporal information. Since we are interested in finding the temporal location of the events, we aggregate the features in the spatial dimension with average pooling before extracting long-range temporal information. We use temporal block because the feature extractor alone is not capable enough to capture the required long-range dependency, although it may help to some extent. As motivated in the introduction section, we need to capture long-range temporal information to accurately predict events with precision. For this task, we selected Bidirectional GRU (Bi-GRU) as the temporal block because it effectively models sequences without overfitting the input data. The Bi-GRU is capable of quickly adapting to the data while maintaining its generalizability. Further experiments confirm this, demonstrating that the Bi-GRU performs well compared to more complex temporal blocks.", + "bbox": [ + 511, + 500, + 906, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 Soft Instance Contrastive (Soft-IC) loss:", + "text_level": 1, + "bbox": [ + 511, + 784, + 844, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the class imbalance issue, we introduce a novel Soft-IC loss, which aims to make the features more compact and increase their separability. It is inspired by the Instance Contrastive loss (IC loss) [11] and is designed to work with mixup augmentation [33]. IC loss enhances the separation among the features by promoting its compactness. In case", + "bbox": [ + 511, + 810, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3166", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of class imbalance, features may deviate from their respective class center and potentially overlap with features from other classes. This can hinder the learning of a decision boundary and negatively impact performance [4]. By compacting the features, we indirectly address the challenges posed by class imbalance, a strategy supported by research in the field [31]. For a batch of $N$ features computed from the input, the IC Loss is calculated as described in [11].", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} (\\mathbf {z} _ {i}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 226, + 482, + 267 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\mathbf {z} _ {i}\\right) = \\frac {1}{\\left| M \\left(\\mathbf {c} _ {i}\\right) \\right|} \\sum_ {\\mathbf {z} _ {j} \\in M \\left(\\mathbf {c} _ {i}\\right)} \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j} / \\tau\\right)}{\\sum_ {\\mathbf {z} _ {\\mathrm {k}} \\in \\mathbf {A} \\left(\\mathbf {c} _ {\\mathrm {i}}\\right)} \\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {k} / \\tau\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 300, + 483, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_i$ is the label of $i$ -th sample, $\\tau$ is a temperature hyper-parameter, $M(\\mathbf{c}_i)$ denotes the memory bank of class $\\mathbf{c}_i$ and $A(\\mathbf{c}_i) = M\\backslash M(\\mathbf{c}_i)$ .", + "bbox": [ + 89, + 353, + 483, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we use mixup augmentation [33] to regularize the model and improve its generalization. In this, two samples (and their labels) are combined convexly to create a new sample (and its label) as follows.", + "bbox": [ + 89, + 400, + 483, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} = \\lambda x _ {i} + (1 - \\lambda) x _ {j}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 476, + 482, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} = \\lambda y _ {i} + (1 - \\lambda) y _ {j}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 494, + 482, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{i}$ and $x_{j}$ are two samples from the training data, and $y_{i}$ and $y_{j}$ are their corresponding one-hot labels. The parameter $\\lambda$ takes value in the range [0, 1]. Consequently, the label $(\\tilde{y})$ will not be in one-hot format; instead, it will reflect weighted contributions from each class.", + "bbox": [ + 89, + 526, + 483, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "IC Loss assumes that each sample has a single label. However, this assumption does not hold true when using mixup augmentation. To account for weight associated with each class in the label, we define Soft-IC loss as", + "bbox": [ + 89, + 604, + 483, + 664 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S I C} = \\frac {\\sum_ {\\left(\\mathbf {z} _ {i} , w _ {i j}\\right) , w _ {i j} \\neq 0} \\mathcal {L} _ {S I C} \\left(\\mathbf {z} _ {i} , w _ {i j}\\right)}{\\sum_ {\\left(\\mathbf {z} _ {i} , w _ {i j}\\right) , w _ {i j} \\neq 0} 1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 676, + 483, + 719 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S I C} (\\mathbf {z} _ {i}, \\omega) = \\frac {1}{\\omega | M (\\mathbf {c} _ {i}) |} \\times\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 750, + 295, + 782 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {(\\mathbf {z} _ {j}, w _ {j}) \\in M (\\mathbf {c} _ {i})} \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot w _ {j} \\mathbf {z} _ {j} / \\tau\\right)}{\\sum_ {(\\mathbf {z} _ {\\mathbf {k}} , w _ {k}) \\in A (\\mathbf {c} _ {i})} \\exp \\left(\\mathbf {z} _ {i} \\cdot w _ {k} \\mathbf {z} _ {k} / \\tau\\right)}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 785, + 483, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $w_{ij}$ is the weight associated with the label of class $c_i$ for sample $\\mathbf{z}_i$ . Additionally, in the memory bank $M(\\mathbf{c}_i)$ , we store the corresponding class weights along with features.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Strategy", + "text_level": 1, + "bbox": [ + 513, + 90, + 687, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed framework is trained end-to-end using a combination of two loss functions: Binary Cross Entropy loss $(\\mathcal{L}_{BCE})$ as the standard classification loss and the proposed Soft-IC loss $(\\mathcal{L}_{SIC})$ . The overall loss function $(\\mathcal{L}_{final})$ can be expressed as below:", + "bbox": [ + 511, + 113, + 906, + 188 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i n a l}} (\\cdot) = \\mathcal {L} _ {B C E} + \\lambda_ {S I C} \\cdot \\mathcal {L} _ {S I C}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 200, + 906, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\lambda_{SIC}$ denotes the weighting constant that adjusts the contributions of different loss functions. With the total loss, the proposed network is trained using the AdamW [20] optimizer with Adaptive Sharpness-Aware Minimization (ASAM) [15].", + "bbox": [ + 511, + 226, + 905, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Utilizing advanced optimization techniques can significantly impact the performance of the trained model, yet not much focus is given to this aspect. In [23], the authors utilized the Sharpness-Aware Minimization (SAM) [9] with the Adam optimizer, which significantly impacted the final scores. SAM is a technique that considers the \"sharpness\" of the loss manifold in each gradient descent step, resulting in improved generalization. In SAM, the sharpness is computed from a fixed-size neighborhood of the current location in the loss manifold. However, the fixed-size neighborhood is not always ideal for sharpness computation. Adaptive SAM (ASAM) [15] proposed a size-varying neighborhood based on the scale to improve upon it. Taking a cue from this, we propose to utilize ASAM with AdamW optimizer to optimize our proposed framework. Through ablation analysis, we validate the effectiveness of ASAM in boosting the performance of the proposed framework.", + "bbox": [ + 511, + 301, + 906, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Analysis", + "text_level": 1, + "bbox": [ + 511, + 571, + 728, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section provides detailed information about the implementations and datasets, followed by the evaluation and ablation study.", + "bbox": [ + 511, + 595, + 905, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 650, + 730, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We choose a pre-trained RegNetY200 initialized with ImageNet weights as the backbone in the spatio-temporal feature extractor. The model is trained using ASAM [15] with AdamW [20] as the base optimizer, setting $\\rho = 2$ . For training, a cosine annealing learning rate scheduler is employed with a warm-up of 3 epochs and an initial learning rate of 1e-5 during the warm-up. For the Soft-IC loss, a feature dimension of 128 and a memory bank size of 256 is used, with a loss weight (i.e., $\\lambda_{SIC}$ ) of 0.001. To enhance the training data, mixup [33] augmentation is applied with a mixing factor of $\\alpha = 0.1$ . Additionally, Gaussian blur is applied with a kernel size of 5 and a probability of 0.25. Brightness, contrast, and saturation adjustments are made with a factor ranging from 0.7 to 1.2, at a probability of 0.5. The proposed method is implemented in PyTorch and trained", + "bbox": [ + 511, + 674, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3167", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on 4 A100 GPUs with a batch size of 2 on each GPU. At a time, 128 frames are processed, constituting a single sample (clip). During training, clips are uniformly sampled from videos, while during evaluation, a sliding window with half overlap is utilized to extract the clips.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the efficacy of the proposed method, we conducted experiments using event spotting datasets outlined in [14] such as Tennis [34], Figure Skating (FS) [13], Fine Gym and the SoccerNet V2 [5] action spotting dataset. These datasets are described in the supplementary material.", + "bbox": [ + 89, + 167, + 480, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Metric Description", + "text_level": 1, + "bbox": [ + 89, + 252, + 359, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Like previous works, we calculate the average precision (AP) within a specified tolerance range, denoted as $\\delta$ . The AP is then calculated for each event class to calculate the mean average precision (mAP) with the specified $\\delta$ value as an evaluation metric. The $\\delta$ is defined based on the respective datasets. For the Tennis, Figure Skating, and FineGym datasets, which have very strict evaluation criteria, we report the results with $\\delta = 0,1$ , and 2. The SoccerNet V2 dataset has relaxed evaluation criteria and defines two metrics: tight-mAP and loose-mAP, with $\\delta = 1 - 4$ seconds and $\\delta = 5 - 60$ seconds, respectively. Although the metrics are less strict compared to others, the SoccerNet V2 dataset is more challenging; having a more stringent metric would make the task even harder.", + "bbox": [ + 89, + 275, + 482, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Baseline Methods", + "text_level": 1, + "bbox": [ + 89, + 496, + 261, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the SoccerNet V2 [5] dataset, we compare existing SOTA methods proposed in conjunction with the SoccerNet Action Spotting challenge. Some of the results are taken from the 2021 challenge leaderboard, including Baidu (TF) [35], and SpotFormer [2]. We have also compared the results with CNN-based SOTA methods like E2ESpot [14], T-DEED [30], UGLF [27] and transformer-base SOTA methods like Spivak [23], ASTRA [29], and COMEDIAN [6]. The results are generated using the checkpoints provided by the respective authors whenever possible.", + "bbox": [ + 89, + 518, + 482, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the Tennis, Figure Skating, and FineGym dataset, we compare our method with the baseline method E2E-Spot [14]. We have also included the results of some best-performing pre-trained and fine-tuned variants as defined in [14]. These include TSP [1] + MS-TCN [8], TSP [1] + AS-Former [32], VC-Spot [14], and 2D-VDP [12] + MS-TCN [8]. We also compare the result with the recent state-of-the-art T-DEED [30] method. Unless otherwise stated, the results are mainly generated from the checkpoint provided by the authors.", + "bbox": [ + 89, + 671, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Result Analysis on Various Sports Datasets", + "text_level": 1, + "bbox": [ + 89, + 830, + 455, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section compares the results obtained by our proposed method with existing SOTA methods on SoccerNet V2, Tennis, Figure Skating & FineGym datasets.", + "bbox": [ + 89, + 854, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Result Comparison on SoccerNet-V2 datasets in terms of mAP tight and loose settings. Here, the best-performing measures are highlighted with ** bold font and second best-performing is highlighted with ** underline font. * result obtained from the challenge leaderboard of 2021. ** indicate that the GFLOPs value is calculated from the temporal network only without feature extractors. † indicate result generated from the author provided checkpoint. ‡ indicate results obtained from official repo.", + "bbox": [ + 511, + 88, + 903, + 200 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5ae2e9bd6b28f08d050b1267e286851c453af8bc765888ae303fbdf660a691b2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelParams (M)GFLOPsTightLoose
Baidu (TF) [35]——47.05*73.77*
SpotFormer [2]——60.9076.10
E2E-Spot (RegNet-Y 200MF) [14]4.4639.6161.1973.25
E2E-Spot (RegNet-Y 800MF) [14]12.64151.4061.8274.05
Spivak [23]17.46461.8965.1078.50
ASTRA [29]44.338.83**66.6378.14
T-DEED (RegNet-Y 200MF) [30]16.3621.9639.43‡52.93‡
T-DEED (RegNet-Y 800MF) [30]46.2260.2521.57‡30.49‡
UGLF [27]4.4639.6163.5173.98‡
COMEDIAN (ViSwin-T) [6]70.12222.7671.33†77.07†
Proposed6.4660.2573.7479.11
", + "bbox": [ + 516, + 210, + 903, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SoccerNet V2: Table 1 compares the existing SOTA methods and the proposed approach on the SoccerNet V2 testing dataset. It is evident from the results that the proposed method has achieved notably strong performance in both tight and loose settings. The scores demonstrate significant improvement compared to both RegNet-Y 200MF and 800MF of E2E-Spot despite utilizing a similar framework. Additionally, there is a substantial improvement in the tight setting compared to the Baidu, Spivak, and ASTRA methods. These methods employ a pre-trained frozen larger backbone for feature extraction and transformer-based long-term temporal modules. COMEDIAN model utilizes a transformer network to achieve SOTA results in tight settings. However, the proposed method outperforms these methods without requiring a larger network. Notably, the proposed model achieves a $2.41\\%$ improvement in tight settings than the COMEDIAN model while having significantly fewer parameters and FLOPs.", + "bbox": [ + 511, + 383, + 906, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tennis, Figure Skating & FineGym: Table 2 shows the results for Tennis, Figure Skating settings, i.e., FS-Comp and FS-Perf, and FineGym with $\\delta = 0, 1$ and 2. In this case, the proposed method also outperforms existing methods with a noticeable improvement. For instance, in case of $\\delta = 0$ , the proposed method achieves $15.67\\%$ , $8.64\\%$ , $8.49\\%$ and $12.55\\%$ improvements over the best-performing E2ESpot (RegNet-Y 200MF) and T-DEED (RegNet-Y 200MF) methods in Tennis, FS-Comp & FS-Perf and FineGym datasets, respectively. Similarly, for $\\delta = 1$ case, the proposed method outperforms the previous best-performing T-DEED variants RegNet-Y 200MF and RegNet-Y 800MF by $1.52\\%$ & $1.4\\%$ improvement on FS-Comp & FS-Perf datasets, respectively. Additionally, we present the per-class score comparison on Tennis, Figure Skating, & FineGym datasets in Figure 4. Here, it can be seen that the perfor", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3168", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fae9d28f86b248af4ce89ff4cdeda5cecb94b1bdb32f77d2eda636bc20e5c498.jpg", + "table_caption": [ + "Table 2. Result Comparison on Tennis, FS-Comp, FS-Perf and FinGym datasets in terms of mAP for different delta values. Here, the best-performing measures are highlighted with bold font and second best-performing is highlighted with underline font. * indicate result generated using provided checkpoints." + ], + "table_footnote": [], + "table_body": "
ModelTennisFS-CompFS-PerfFineGym
δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2
TSP [1] + MS-TCN [8]—90.9095.10—72.4087.80—76.8089.90—40.5058.50
TSP [1] + ASFormer [32]—89.8095.50—77.7094.10—80.2094.50—38.8057.60
2D-VDP [12] + MS-TCN [8]————83.5096.20—85.2096.40———
VC-Spot (RegNet-Y 200MF) [14]—92.4096.00—61.8075.50—56.2075.30—18.7028.60
E2E-Spot (RegNet-Y 200MF) [14]45.34*96.1097.7022.08*81.0093.5022.33*85.1095.7010.01*47.9065.20
E2E-Spot (RegNet-Y 800MF) [14]44.64*95.16*97.72*23.95*83.4094.9022.17*83.3096.0010.15*50.11*67.23*
T-DEED (RegNet-Y 200MF) [30]43.4996.1797.9528.92*85.1591.7031.87*86.7996.0513.54*53.73*66.43*
T-DEED (RegNet-Y 800MF) [30]44.0896.4297.0527.59*84.7792.8631.25*88.1795.8712.42*54.00*66.70*
UGLF [27]——————————50.2067.80
Proposed61.0196.2197.7537.5486.6795.5340.3689.5797.3815.2452.3166.57
", + "bbox": [ + 94, + 130, + 903, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f32d2124f031badae1e4917ca47ac56f1efc47c2ab986de5f8598f64c566c5b8.jpg", + "image_caption": [ + "(a) Tennis dataset [34]" + ], + "image_footnote": [], + "bbox": [ + 89, + 313, + 460, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0cb9a6e62b8e5d77982a536536f8afed6881a95c08390c0822a854efaff96067.jpg", + "image_caption": [ + "(c) FS-Comp dataset [13]" + ], + "image_footnote": [], + "bbox": [ + 94, + 453, + 460, + 553 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97c5c9fe5513194355556825f9706982ba56a9a57f5c452e3e90c10fd0f96981.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 308, + 900, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da0e9b2119e7b616cc6ab177c3c9aff1a62cbeefd10a763a73c70549640a1e95.jpg", + "image_caption": [ + "(b) FineGym dataset [22]", + "(d) FS-Perf dataset [13]" + ], + "image_footnote": [], + "bbox": [ + 540, + 455, + 903, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance of the proposed method is not just good for a single class; it has performed consistently well across all the classes by a significant margin.", + "bbox": [ + 89, + 619, + 485, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 678, + 254, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies to validate the design choices made in the proposed framework. To ensure a fair comparison, we analyze the experimental results on SoccerNet V2.", + "bbox": [ + 89, + 703, + 482, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of the proposed modules: Table 3 shows the contribution of the proposed ASTRM, Soft-IC loss, and ASAM components to achieve SOTA results. We observe a significant improvement in scores when replacing GSM [14] with the proposed ASTRM, highlighting the importance of effective feature extraction. SAM [9], utilized in the Spivak method, further enhances the results by a notable margin. Furthermore, we have illustrated how ASAM outperforms the scores achieved by SAM. Although ASAM and Soft-IC loss improves the overall scores by a small", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/839dba5ad5e8504802bb00a446c93dbd06a3cdc7424ee79323f51a8a14495cfb.jpg", + "image_caption": [ + "Figure 4. Per-class score comparison on $\\delta = 0$ setting in terms of mAP for Tennis, FineGym, FS-Comp, and FS-Perf dataset. For FineGym, we report results for only 6 out of 32 classes, while results for all classes in the other datasets are reported.", + "Figure 5. Analysis on our method with & without Soft-IC loss." + ], + "image_footnote": [], + "bbox": [ + 535, + 617, + 885, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "amount, Soft-IC loss plays an important role in improving the scores of minority classes. This is illustrated with scores of individual classes in Figure 5. As can be seen, Soft-IC loss improves the scores of classes with less number of samples without affecting other classes. This emphasizes the contributions of all components to achieve SOTA results.", + "bbox": [ + 511, + 776, + 906, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Components of the proposed ASTRM: The proposed AS-TRM module uses 3 major components: local spatial, local", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3169", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fae8cecf3afa47fb811c975b1fcc567445e354404cd395f912e63468c77ed563.jpg", + "table_caption": [ + "Table 3. Ablation studies on various aspects of the proposed framework on SoccerNet V2 testing dataset in terms of mAP." + ], + "table_footnote": [], + "table_body": "
MethodTightLoose
Analysis on proposed modules
Baseline: E2E-Spot (RegNet-Y 200MF + GSM)61.1973.25
RegNet-Y 200MF + ASTRM68.6375.57
RegNet-Y 200MF + ASTRM + SAM71.8977.50
RegNet-Y 200MF + ASTRM + ASAM72.6578.58
RegNet-Y 200MF + ASTRM + ASAM + Focal Loss70.3676.86
RegNet-Y 200MF + ASTRM + ASAM + IC Loss72.0578.30
RegNet-Y 200MF + ASTRM + ASAM + Soft-IC loss73.7479.11
Analysis on proposed ASTRM module
With local spatial only69.9176.39
With local temporal only70.4176.79
With global temporal only71.6577.06
With local spatial + local temporal71.0677.10
With local spatial + global temporal71.8277.85
With local temporal + global temporal71.5576.93
Proposed ASTRM module73.7479.11
Analysis on Different Temporal networks
Baseline with Bi-GRU61.1973.25
Proposed with Bi-GRU73.7479.11
with Deformable Attention67.4372.15
with Bi-GRU 2 Layers67.8573.03
with Transformer (L1H8)65.8970.76
with Transformer (L2H8)67.6072.35
with Bi-LSTM68.1573.35
with MSTCN68.7573.50
Analysis on Clip Length
Clip length = 6443.8252.09
Clip length = 100 (Used in baseline E2E-Spot)71.2076.73
Clip length = 128 (proposed)73.7479.11
Clip length = 19271.7477.02
", + "bbox": [ + 91, + 116, + 478, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "temporal, and global temporal. Here, we highlight the importance of each component, and the corresponding findings are reported in Table 3. Notably, the score is at its lowest when utilizing only the spatial aspect of ASTRM; it fails to capture temporal information. Furthermore, the performance of local temporal is superior to using only the spatial aspect. Interestingly, we also observed that global temporal performs better than local temporal alone. This finding supports the notion that observing distant frames may be necessary to classify the current frame accurately. A similar pattern is seen when comparing local spatial + local temporal with local spatial + global temporal. Moreover, the inclusion of local spatial information leads to improved scores. Conversely, using only temporal information results in even lower scores, indicating the necessity of spatial information for understanding \"what is happening\" and temporal information for determining \"when it is happening.\" Overall, our observations highlight the need for spatial and local-global temporal information to identify events precisely and outperform other settings significantly.", + "bbox": [ + 89, + 564, + 482, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on Different Temporal modules In our proposed approach, we opted for the Bidirectional GRU (Bi-GRU) as", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the long-range dependency module in the temporal block after conducting various experiments with different potential temporal networks. Table 3 presents the results obtained from the proposed approach using different temporal networks on the SoccerNet V2 [5] test set. The results indicate that the Bi-GRU, as a temporal network, performs favorably compared to other temporal networks. Even more advanced temporal networks, such as Bi-LSTM, Transformer, Deformable Attention [18] and MSTCN [8], could not surpass the scores achieved by the Bi-GRU temporal network. The difference in scores is visible across all tolerances. It is also observed that the proposed method generates significantly better results than the baseline E2E-Spot method, which also uses Bi-GRU as the temporal block. For instance, the proposed method achieves $12.55\\%$ and $5.86\\%$ improvements over the baseline model in the tight and loose settings of the SoccerNet V2 dataset, respectively. Similarly, there is an improvement of $15.67\\%$ in the $\\delta = 0$ setting compared to the baseline on the Tennis dataset. This significant improvement is attributed to the proposed ASTRM module and the proposed Soft-IC loss function.", + "bbox": [ + 511, + 90, + 906, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Clip Length: Due to memory and computation constraints, it's not feasible to process all the frames simultaneously. Therefore, we process a fixed number of consecutive frames at a time, which we refer to as the \"clip length\". The value of the clip length affects the scores because it determines how many frames the network considers at once. While we might want to maximize the clip length, we need to find a balance between performance and clip length. The results in Table 3 support this idea. If the clip length is too short, it significantly impacts the results, while making it too long also has a negative effect on the results.", + "bbox": [ + 511, + 411, + 908, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion & Future Directions", + "text_level": 1, + "bbox": [ + 511, + 599, + 805, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes an end-to-end network for precise spotting of events in sports videos. It effectively tackles the long-range dependency and class imbalance issues and outperforms existing methods. To this extent, we introduce a new ASTRM module for enhancing spatio-temporal features. Further, to address the class imbalance issue, we propose Soft-IC loss, which enforces compact features and improves class separability. We further employed ASAM during training to ensure the model performs well on unseen data. While our method has advanced the state-of-the-art, especially in the more challenging cases with a relatively simple design, there is further scope for improvement on the two fronts we tried to address. Certain larger networks can improve the performance further but at the cost of computation. On the other hand, the spatial and temporal resolution at which the data is being processed will likely have an impact. The precise spotting may become easier with increased resolution, but it increases the processing burden.", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3170", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Humam Alwassel, Silvio Giancola, and Bernard Ghanem. Tsp: Temporally-sensitive pretraining of video encoders for localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3173-3183, 2021. 6, 7", + "[2] Mengqi Cao, Min Yang, Guozhen Zhang, Xiaotian Li, Yilu Wu, Gangshan Wu, and Limin Wang. Spotformer: A transformer-based framework for precise soccer action spotting. In 2022 IEEE 24th International Workshop on Multimedia Signal Processing (MMSP), pages 1-6. IEEE, 2022. 6", + "[3] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2", + "[4] Charika De Alvis et al. A survey of deep long-tail classification advancements. arXiv:2404.15593, 2024. 5", + "[5] Adrien Deliège, Anthony Cioppa, Silvio Giancola, Meisam J. Seikavandi, Jacob V. Dueholm, Kamal Nasrollahi, Bernard Ghanem, Thomas B. Moeslund, and Marc Van Droogenbroeck. Soccertnet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4508-4519, 2021. 2, 6, 8", + "[6] Julien Denize, Mykola Liashuha, Jaonary Rabarisoa, Astrid Orcesi, and Romain Hérault. Comedian: Self-supervised learning and knowledge distillation for action spotting using transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 530-540, 2024. 2, 3, 6", + "[7] Moritz Einfalt, Charles Dampeyrou, Dan Zecha, and Rainer Lienhart. Frame-level event detection in athletics videos with pose-based convolutional sequence networks. In Proceedings Proceedings of the 2nd International Workshop on Multimedia Content Analysis in Sports, page 42-50, New York, NY, USA, 2019. Association for Computing Machinery. 2", + "[8] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 6, 7, 8", + "[9] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021. 5, 7", + "[10] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccertnet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018. 2", + "[11] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9591-9600, 2022. 4, 5" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] James Hong, Matthew Fisher, Michael Gharbi, and Kayvon Fatahalian. Video pose distillation for few-shot, fine-grained sports action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9254-9263, 2021. 6, 7", + "[13] James Hong, Matthew Fisher, Michael Gharbi, and Kayvon Fatahalian. Video pose distillation for few-shot, fine-grained sports action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9254-9263, 2021. 6, 7", + "[14] James Hong, Haotian Zhang, Michael Gharbi, Matthew Fisher, and Kayvon Fatahalian. Spotting temporally precise, fine-grained events in video. In European Conference on Computer Vision, pages 33-51. Springer, 2022. 2, 6, 7", + "[15] Jungmin Kwon, Jeongseop Kim, Hyunseo Park, and In Kwon Choi. Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks. In Proceedings of the 38th International Conference on Machine Learning, pages 5905-5914. PMLR, 2021. 2, 5", + "[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 2", + "[17] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20010-20019, 2022. 2", + "[18] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2, 8", + "[19] Zhaoyang Liu et al. Tam: Temporal adaptive module for video recognition. In ICCV, pages 13708-13718, 2021. 3", + "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5", + "[21] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[22] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7", + "[23] Joao VB Soares and Avijit Shah. Action spotting using dense detection anchors revisited: Submission to the soccer-net challenge 2022. arXiv preprint arXiv:2206.07846, 2022. 5, 6", + "[24] Joao VB Soares, Avijit Shah, and Topojoy Biswas. Temporally precise action spotting in soccer videos using dense detection anchors. In 2022 IEEE International Conference on Image Processing (ICIP), pages 2796-2800. IEEE, 2022. 2", + "[25] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Gate-shift networks for video action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3171", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Matteo Tomei, Lorenzo Baraldi, Simone Calderara, Simone Bronzin, and Rita Cucchiara. Rms-net: Regression and masking for soccer event spotting. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7699-7706. IEEE, 2021. 2", + "[27] Kim Hoang Tran, Phuc Vuong Do, Ngoc Quoc Ly, and Ngan Le. Unifying global and local scene entities modelling for precise action spotting. arXiv preprint arXiv:2404.09951 (Accepted for IJCNN 2024), 2024. 2, 3, 6, 7", + "[28] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3", + "[29] Artur Xarles, Sergio Escalera, Thomas B Moeslund, and Albert Clapés. Astra: An action spotting transformer for soccer videos. In Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports, pages 93-102, 2023. 2, 6", + "[30] Artur Xarles, Sergio Escalera, Thomas B. Moeslund, and Albert Clapés. T-deed: Temporal-discriminability enhancer encoder-decoder for precise event spotting in sports videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3410-3419, 2024. 2, 6, 7", + "[31] Huiran Yan et al. Emphasizing feature inter-class separability for improving highly imbalanced overlapped data classification. Knowledge-Based Systems, 276:110745, 2023. 5", + "[32] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. 32nd British Machine Vision Conference, BMVC 2021, 2021. 2, 6, 7", + "[33] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 4, 5", + "[34] Haotian Zhang, Cristobal Sciutto, Maneesh Agrawala, and Kayvon Fatahalian. Vid2player: Controllable video sprites that behave and appear like professional tennis players. ACM Trans. Graph., 40(3), 2021. 6, 7", + "[35] Xin Zhou, Le Kang, Zhiyu Cheng, Bo He, and Jingyu Xin. Feature combination meets attention: Baidu soccer embeddings and transformer based temporal detection. arXiv preprint arXiv:2106.14447, 2021. 2, 6" + ], + "bbox": [ + 91, + 90, + 480, + 684 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3172", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/f70c89aa-2e9b-4112-89fc-f4837e98e688_model.json b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/f70c89aa-2e9b-4112-89fc-f4837e98e688_model.json new file mode 100644 index 0000000000000000000000000000000000000000..64c231f6f89b8ee3d14c9c8dac213d961ffc451b --- /dev/null +++ b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/f70c89aa-2e9b-4112-89fc-f4837e98e688_model.json @@ -0,0 +1,1837 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.13, + 0.895, + 0.175 + ], + "angle": 0, + "content": "Precise Event Spotting in Sports Videos: Solving Long-Range Dependency and Class Imbalance" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.203, + 0.871, + 0.24 + ], + "angle": 0, + "content": "Sanchayan Santra\\(^{1}\\) Vishal Chudasama\\(^{1}\\) Pankaj Wasnik\\(^{1*}\\) Vineeth N Balasubramanian\\(^{2}\\) \n\\(^{1}\\)Sony Research India \\(^{2}\\)Indian Institute of Technology Hyderabad" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.242, + 0.882, + 0.258 + ], + "angle": 0, + "content": "{sanchayan.santra, vishal.chudasamal, pankaj.wasnik}@sony.com, vineethnb@cse.iith.ac.in" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.486, + 0.673 + ], + "angle": 0, + "content": "Precise Event Spotting (PES) aims to identify events and their class from long, untrimmed videos, particularly in sports. The main objective of PES is to detect the event at the exact moment it occurs. Existing methods mainly rely on features from a large pre-trained network, which may not be ideal for the task. Furthermore, these methods overlook the issue of imbalanced event class distribution present in the data, negatively impacting performance in challenging scenarios. This paper demonstrates that an appropriately designed network, trained end-to-end, can outperform state-of-the-art (SOTA) methods. Particularly, we propose a network with a convolutional spatial-temporal feature extractor enhanced with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) and a long-range temporal module. The ASTRM enhances the features with spatiotemporal information. Meanwhile, the long-range temporal module helps extract global context from the data by modeling long-range dependencies. To address the class imbalance issue, we introduce the Soft Instance Contrastive (Soft-IC) loss that promotes feature compactness and class separation. Extensive experiments show that the proposed method is efficient and outperforms the SOTA methods, specifically in more challenging settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.701, + 0.222, + 0.716 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Event spotting methods locate the occurrence of events in long, untrimmed videos. An event is typically defined by a starting and ending time, but this may not always be feasible, particularly in sports, where an event is a specific occurrence that occurs at a particular point in time, as dictated by the game's rules. For instance, a \"Goal\" in soccer is scored when the ball crosses the goal line, while a \"Boundary\" in cricket is made when the ball touches the boundary line. This process is referred to as Precise Event Spotting (PES), where PES methods aim to predict an event at the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.289, + 0.902, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.444, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Figure 1. Temporal dependency of \"Shots on target\" event: The event is marked when the shot is taken, but whether the shot is at the target can only be known by looking at future frames. Here the shot is on target because the ball is saved by the goal keeper." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.514, + 0.905, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.677, + 0.909, + 0.734 + ], + "angle": 0, + "content": "Figure 2. Per class score analysis on a few classes of SoccerNet V2 dataset in tight setting. The proposed method outperforms state-of-the-art (SOTA) methods especially for classes with less number of samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.839 + ], + "angle": 0, + "content": "exact frame it occurs (e.g., goal, boundary). This is crucial for generating game summaries, finding specific events, and editing lengthy matches. However, it poses significant challenges, requiring a deep understanding of the game and its key events." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Existing PES methods mainly face two challenges. First is the long-range temporal dependency that arises due to the nature of events. To predict the event frames, observing just the nearby frames does not suffice; several other potentially" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.239, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3163" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.364 + ], + "angle": 0, + "content": "distant frames need to be observed to predict the event correctly. For example, \"Shots on target\" and \"Shots off target\" events of Soccer games are marked when the shot is taken. Still, the actual class is decided based on whether the ball went toward the goal or away from it (see Figure 1). This can be only detected if long-range dependencies are captured correctly. Second, there is an inherent imbalance in the number of samples among the classes (see Figure 2). Some events naturally occur more frequently than others. For example, in a soccer game, a \"Ball out of play\" is likely to occur more than a \"Yellow card\"; a \"Red-card\" and \"Yellow \\(\\rightarrow\\) Red card\" are expected to occur even less often. Therefore, the same trend can be observed in datasets created using broadcast videos. Due to these issues, existing event spotting methods [6, 14] perform poorly in underrepresented classes such as \"Red card\" & \"Yellow \\(\\rightarrow\\) Red card.\" This can be visualized in Figure 2. Handling these issues and accurately performing PES is an active area of research." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.483, + 0.638 + ], + "angle": 0, + "content": "To address the long-range temporal dependency issue, recent SOTA methods [24, 29] prefer to use the feature extracted using the pre-trained network of [35]. These methods also use transformer-based temporal modules due to their better sequence modeling capability in a global context. However, as stated in [18, 32], transformers also encounter issues with long videos. Self-attention struggles to learn appropriate weights for long videos due to the increase in the size of the weight matrix with the number of frames being processed. On the other hand, E2ESpot [14] proposed an end-to-end trainable feature extractor and a GRU-based module for temporal modeling. Recently proposed T-DEED [30] improves the performance of E2ESpot via temporal-discriminability enhancing encoder-decoder (T-DEED) module. In our proposed method, we handle this issue by enhancing the extracted features with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) and GRU-based temporal module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.822 + ], + "angle": 0, + "content": "The class imbalance issue has not received sufficient attention in the literature on PES, resulting in lower performance on real-world datasets. The recent method ASTRA [29] is one of the few attempts to address this problem through a modified \"mixup\" augmentation. Although standard machine learning approaches, such as focal loss [16], are available, they are not suitable for every task, as this paper demonstrates. To address this challenge, we introduce a novel loss function called Soft Instance Contrastive (Soft-IC) loss. This loss function aims to mitigate class imbalance by promoting compact feature representations and improving class separability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The primary goal of this paper is to address these issues and improve the accuracy of PES. To this end, we propose an end-to-end trainable network with a CNN-based spatiotemporal feature extractor and a GRU-based long-term temporal module. The spatio-temporal feature extractor in-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "cludes our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) to augment the features with spatial and temporal information. We chose to train the model end-to-end because it has been shown to perform better than a pre-trained frozen feature extractor [17, 26]. We have also employed the Adaptive Sharpness-Aware Minimization (ASAM) [15] to enable the trained model to perform well on unseen data. Through experiments, we demonstrate that the proposed framework obtains superior performance compared to SOTA methods, especially in tight settings where long-range dependency and class imbalance issues are prominent (see Figure 2). Our key contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.288, + 0.905, + 0.348 + ], + "angle": 0, + "content": "- We propose an end-to-end framework using a CNN-based spatio-temporal feature extractor and GRU-based temporal network to achieve precise event spotting while addressing long-range dependency and class imbalance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.905, + 0.378 + ], + "angle": 0, + "content": "- We introduce ASTRM, which enhances the features with spatial and temporal information." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.905, + 0.422 + ], + "angle": 0, + "content": "- We also introduce Soft-IC loss to enforce compact features and improve class separability, addressing the class imbalance issue." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.288, + 0.905, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.438, + 0.663, + 0.453 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.905, + 0.719 + ], + "angle": 0, + "content": "The concept of precise sports event spotting first appeared in Einfalt et al. [7], where they considered temporally precise detection of various events related to long and triple jump. Later, SoccerNet [10] proposed a large-scale dataset of soccer videos to identify specific events or actions at exact timestamps. The work by [14] expanded this to precise event spotting (PES), introducing significant differences in the necessary precision of predictions. PES is constrained to the tolerance of a few frames only. Recently, several PES methods [6, 14, 24, 27, 29, 30, 35] have been proposed that focus on various sports videos. E2E-Spot [14] employed a CNN-based backbone with a Gated Shift Module (GSM) [25] for temporal modeling and a GRU [3] for producing per-frame prediction. Even though it is effective on the Tennis and Figure Skating datasets, it struggles on more challenging SoccerNet V2 dataset [5] due to its simplistic temporal information extraction scheme using GSM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Building upon E2E-Spot [14], T-DEED [30] proposed improvements to address issues arising from tight tolerances in fast-paced sporting events and achieved SOTA performance on the Figure Skating and Figure Diving datasets. The method of [35] revamped the feature extraction using a group of five networks pre-trained on Kinetics-400 and fine-tuned on SoccerNet V2. Several other methods use the features extracted by this method instead of raw video frames. For instance, Soares et al. [24] utilized these features and developed a U-Net- and transformer-based trunk for making predictions. ASTRA [29] utilizes the same features but incorporates audio features for complementary in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3164" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "formation. In contrast, COMEDIAN [6] operates in 3 stages to train their network on raw frames: 1) training only the spatial part; 2) training the spatial and temporal parts using knowledge distillation; 3) final tuning to optimize the classification scores. Recently, UGLF [27] proposed a unifying global-local (UGL) module to disentangle the scene content into global features and locally relevant scene entity features in conjunction with a Vision-Language model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.483, + 0.439 + ], + "angle": 0, + "content": "The above-mentioned methods mainly addressed the long-range temporal dependency issue by utilizing pretrained features and transformers. They have not adapted their feature extractors to capture spatial and temporal information effectively. Additionally, these methods did not specifically deal with the class imbalance, which led to poor performance in underrepresented classes. In our framework, we propose ASTRM, which helps the feature extractor to augment the features with relevant spatial and temporal information. Furthermore, we introduce Soft-IC loss to address the class imbalance issue and employ the ASAM to train the proposed framework properly. With these advancements, the proposed framework outperforms existing event-spotting methods on various sports datasets, particularly in settings with tighter tolerances." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.451, + 0.31, + 0.468 + ], + "angle": 0, + "content": "3. Proposed Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.522 + ], + "angle": 0, + "content": "This section presents our methodology in detail, starting with the problem statement and then presenting the details of the proposed modules as follows." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.53, + 0.276, + 0.545 + ], + "angle": 0, + "content": "3.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.78 + ], + "angle": 0, + "content": "The task of precise event spotting is defined as follows: Given a video with \\(N\\) frames \\(\\mathbf{x}_1, \\mathbf{x}_2, \\ldots, \\mathbf{x}_N\\) and a set of possible events \\(\\{c_1, \\ldots, c_K\\}\\), the goal is to predict frames where an event occurs along with its class: \\((t, \\hat{y}_t) \\in \\mathbb{N} \\times \\{c_1, \\ldots, c_K\\}\\). A prediction is deemed correct if it falls within \\(\\delta\\) frames of the actual occurrence and accurately identifies the event class. Since it is not possible to process the entire video at once, the processing is done in chunks of a predefined size. Each chunk (or clip), represented as \\(\\mathbf{X} \\in \\mathbb{R}^{C \\times T_s \\times H \\times W}\\), consists of \\(T_s\\) frames. We compute predictions for each frame, represented as \\(\\mathbf{Y} \\in \\mathbb{R}^{T_s \\times K}\\), where \\(K\\) is the number of classes. As the video is processed in segments, the predictions from each chunk of \\(T_s\\) frames are aggregated and refined with a non-maximal suppression to get the final sparse prediction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.229, + 0.803 + ], + "angle": 0, + "content": "3.2. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Figure 3 shows the network design of our proposed framework, which takes a set of input frames and classifies whether each frame contains an event or not. The framework comprises three main components: the spatiotemporal feature extractor, the temporal block, and the classifier. In the spatio-temporal feature extractor, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.424 + ], + "angle": 0, + "content": "have selected the RegNetY [21] as the feature extractor due to its efficiency, surpassing commonly used networks like ResNet. This network is further enhanced with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM). This module is integrated into each bottleneck block, refining the features by incorporating both spatial and temporal information. Both of this information is necessary to correctly identify events, as validated by the experiments. The refined features are then passed to the temporal block, which enriches them with long-term temporal information. This long-term information is crucial because, in many situations, information from temporally distant frames is essential to accurately classify a frame. The spatio-temporal feature extractor alone can not accomplish this task. Finally, the feature maps are forwarded to the classifier, which is a simple linear layer that outputs the classification scores. Furthermore, we introduce the Soft Instance Contrastive (Soft-IC) loss to effectively address the imbalance among the class samples. This loss function operates by compacting the features and enhancing their separability. The key modules of the proposed framework are detailed in subsequent subsections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.443, + 0.899, + 0.459 + ], + "angle": 0, + "content": "3.2.1 Adaptive Spatio-Temporal Refinement Module:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.905, + 0.603 + ], + "angle": 0, + "content": "This module enhances the extracted features by incorporating both spatial and temporal information. It comprises three blocks: Local Spatial, Local Temporal, and Global Temporal. We include the Local Spatial block because local spatial features also offer important cues about temporal information. Relying solely on temporal information can overlook these valuable insights. The spatial component is inspired by [28], while the temporal aspect is adapted from [19]. The refinement is done as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.615, + 0.905, + 0.633 + ], + "angle": 0, + "content": "\\[\n\\Psi (\\mathbf {x}) = \\left(\\left(\\mathbf {x} \\odot \\left(1 + \\mathcal {F} _ {s} (\\mathbf {x})\\right)\\right) \\odot \\left(1 + \\mathcal {F} _ {t} (\\mathbf {x})\\right)\\right) * \\mathcal {G} _ {t} (\\mathbf {x}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.869 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) is the input feature to be enhanced. \\(\\mathcal{F}_s, \\mathcal{F}_t,\\) and \\(\\mathcal{G}_t\\) denote the operation corresponding to the local spatial, local temporal, and global temporal block, respectively. \\(\\odot\\) denotes Hadamard Product, while \\(*\\) denotes convolution operation. To elaborate, the input feature is weighted with information from the local spatial block and a residual connection, resulting in features enhanced with local spatial information. These enhanced features are further refined with local temporal information. Finally, spatially and temporally enhanced features are then convolved with the kernel obtained from the global temporal block to make local enhancements using \"global\" information. This dynamic kernel approach simplifies the network while maintaining effectiveness. Please refer to Figure 3 for a visual representation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "In the Local Spatial block, the input features are pooled using Max (i.e., \\( f_{MaxPool} \\)) and Average (i.e., \\( f_{AvgPool} \\))" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3165" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.096, + 0.891, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.393, + 0.908, + 0.449 + ], + "angle": 0, + "content": "Figure 3. Network design of the proposed framework. The framework composed of a spatio-temporal feature extractor and a temporal block for capturing long-range dependency before the classifier. In each bottleneck block we add ASTRM after the first conv. ASTRM further enhances the features with local spatial, local temporal and global temporal information. The network is trained with Soft-IC loss in addition to the classification loss to handle class imbalance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.596 + ], + "angle": 0, + "content": "pooling operations along the channel dimension. The pooled features are then concatenated and convolved with a \\(7 \\times 7\\) filter (i.e., \\(f_{conv}^{7 \\times 7}\\)). The pooling operation on the channel axis helps emphasize informative regions. The subsequent convolution operation filters out non-informative portions from the features. Finally, the output is passed through a Sigmoid function (i.e., \\(\\sigma(\\cdot)\\)) to obtain the \"spatial\" feature weights (i.e., \\(\\mathcal{F}_s(\\mathbf{x})\\)) as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.601, + 0.483, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {s} (\\mathbf {x}) = \\sigma \\left(f _ {c o n v} ^ {7 \\times 7} \\left(\\left[ f _ {A v g P o o l} (\\mathbf {x}); f _ {M a x P o o l} (\\mathbf {x}) \\right]\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.729 + ], + "angle": 0, + "content": "In the Local Temporal block, we use 3D convolution operations, namely \\( f_{conv}^{3 \\times 1 \\times 1} \\) and \\( f_{conv}^{1 \\times 1 \\times 1} \\), followed by the Sigmoid function, \\( \\sigma(\\cdot) \\), with a focus on the temporal dimension. Spatially pooled features are not used in this block because they do not account for the different temporal changes at various spatial locations. The locally temporal enhanced features are defined as below:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.735, + 0.483, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {t} (\\mathbf {x}) = \\sigma \\left(f _ {\\text {c o n v}} ^ {1 \\times 1 \\times 1} \\left(\\mathrm {B N} \\left(\\operatorname {R e L U} \\left(f _ {\\text {c o n v}} ^ {3 \\times 1 \\times 1} (\\mathbf {x})\\right)\\right)\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.484, + 0.879 + ], + "angle": 0, + "content": "In the Global Temporal block, we adaptively create convolution kernels to improve the feature maps by considering temporal information across frames. First, the features are pooled in the spatial dimension using a global average pooling operation (i.e., \\( f_{GAP} \\)), and then fully connected layers (i.e., \\( f_{FC} \\)) are used to compute the kernel with a global view of the data with a softmax (\\( S \\)) at the end. Mathematically, the kernel is formulated as below:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.885, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {t} (\\mathbf {x}) = \\mathcal {S} \\left(f _ {F C} \\left(f _ {F C} \\left(f _ {G A P} (\\mathbf {x})\\right)\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.475, + 0.685, + 0.49 + ], + "angle": 0, + "content": "3.2.2 Temporal Block:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.907, + 0.773 + ], + "angle": 0, + "content": "The spatio-temporal feature extraction stage computes features containing local spatial, local temporal, and location-independent temporal information. Since we are interested in finding the temporal location of the events, we aggregate the features in the spatial dimension with average pooling before extracting long-range temporal information. We use temporal block because the feature extractor alone is not capable enough to capture the required long-range dependency, although it may help to some extent. As motivated in the introduction section, we need to capture long-range temporal information to accurately predict events with precision. For this task, we selected Bidirectional GRU (Bi-GRU) as the temporal block because it effectively models sequences without overfitting the input data. The Bi-GRU is capable of quickly adapting to the data while maintaining its generalizability. Further experiments confirm this, demonstrating that the Bi-GRU performs well compared to more complex temporal blocks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.845, + 0.8 + ], + "angle": 0, + "content": "3.2.3 Soft Instance Contrastive (Soft-IC) loss:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To address the class imbalance issue, we introduce a novel Soft-IC loss, which aims to make the features more compact and increase their separability. It is inspired by the Instance Contrastive loss (IC loss) [11] and is designed to work with mixup augmentation [33]. IC loss enhances the separation among the features by promoting its compactness. In case" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3166" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "of class imbalance, features may deviate from their respective class center and potentially overlap with features from other classes. This can hinder the learning of a decision boundary and negatively impact performance [4]. By compacting the features, we indirectly address the challenges posed by class imbalance, a strategy supported by research in the field [31]. For a batch of \\( N \\) features computed from the input, the IC Loss is calculated as described in [11]." + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.227, + 0.483, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} (\\mathbf {z} _ {i}), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.301, + 0.484, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\mathbf {z} _ {i}\\right) = \\frac {1}{\\left| M \\left(\\mathbf {c} _ {i}\\right) \\right|} \\sum_ {\\mathbf {z} _ {j} \\in M \\left(\\mathbf {c} _ {i}\\right)} \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j} / \\tau\\right)}{\\sum_ {\\mathbf {z} _ {\\mathrm {k}} \\in \\mathbf {A} \\left(\\mathbf {c} _ {\\mathrm {i}}\\right)} \\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {k} / \\tau\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.484, + 0.399 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_i\\) is the label of \\(i\\)-th sample, \\(\\tau\\) is a temperature hyper-parameter, \\(M(\\mathbf{c}_i)\\) denotes the memory bank of class \\(\\mathbf{c}_i\\) and \\(A(\\mathbf{c}_i) = M\\backslash M(\\mathbf{c}_i)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.462 + ], + "angle": 0, + "content": "During training, we use mixup augmentation [33] to regularize the model and improve its generalization. In this, two samples (and their labels) are combined convexly to create a new sample (and its label) as follows." + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.477, + 0.483, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} = \\lambda x _ {i} + (1 - \\lambda) x _ {j}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.496, + 0.483, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} = \\lambda y _ {i} + (1 - \\lambda) y _ {j}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.603 + ], + "angle": 0, + "content": "where \\( x_{i} \\) and \\( x_{j} \\) are two samples from the training data, and \\( y_{i} \\) and \\( y_{j} \\) are their corresponding one-hot labels. The parameter \\( \\lambda \\) takes value in the range [0, 1]. Consequently, the label \\( (\\tilde{y}) \\) will not be in one-hot format; instead, it will reflect weighted contributions from each class." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.665 + ], + "angle": 0, + "content": "IC Loss assumes that each sample has a single label. However, this assumption does not hold true when using mixup augmentation. To account for weight associated with each class in the label, we define Soft-IC loss as" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.678, + 0.484, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S I C} = \\frac {\\sum_ {\\left(\\mathbf {z} _ {i} , w _ {i j}\\right) , w _ {i j} \\neq 0} \\mathcal {L} _ {S I C} \\left(\\mathbf {z} _ {i} , w _ {i j}\\right)}{\\sum_ {\\left(\\mathbf {z} _ {i} , w _ {i j}\\right) , w _ {i j} \\neq 0} 1}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.751, + 0.297, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S I C} (\\mathbf {z} _ {i}, \\omega) = \\frac {1}{\\omega | M (\\mathbf {c} _ {i}) |} \\times\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.786, + 0.484, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\sum_ {(\\mathbf {z} _ {j}, w _ {j}) \\in M (\\mathbf {c} _ {i})} \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot w _ {j} \\mathbf {z} _ {j} / \\tau\\right)}{\\sum_ {(\\mathbf {z} _ {\\mathbf {k}} , w _ {k}) \\in A (\\mathbf {c} _ {i})} \\exp \\left(\\mathbf {z} _ {i} \\cdot w _ {k} \\mathbf {z} _ {k} / \\tau\\right)}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where, \\( w_{ij} \\) is the weight associated with the label of class \\( c_i \\) for sample \\( \\mathbf{z}_i \\). Additionally, in the memory bank \\( M(\\mathbf{c}_i) \\), we store the corresponding class weights along with features." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.688, + 0.108 + ], + "angle": 0, + "content": "3.3. Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.114, + 0.907, + 0.189 + ], + "angle": 0, + "content": "The proposed framework is trained end-to-end using a combination of two loss functions: Binary Cross Entropy loss \\((\\mathcal{L}_{BCE})\\) as the standard classification loss and the proposed Soft-IC loss \\((\\mathcal{L}_{SIC})\\). The overall loss function \\((\\mathcal{L}_{final})\\) can be expressed as below:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.201, + 0.907, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i n a l}} (\\cdot) = \\mathcal {L} _ {B C E} + \\lambda_ {S I C} \\cdot \\mathcal {L} _ {S I C}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.227, + 0.906, + 0.302 + ], + "angle": 0, + "content": "Here, \\(\\lambda_{SIC}\\) denotes the weighting constant that adjusts the contributions of different loss functions. With the total loss, the proposed network is trained using the AdamW [20] optimizer with Adaptive Sharpness-Aware Minimization (ASAM) [15]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.907, + 0.56 + ], + "angle": 0, + "content": "Utilizing advanced optimization techniques can significantly impact the performance of the trained model, yet not much focus is given to this aspect. In [23], the authors utilized the Sharpness-Aware Minimization (SAM) [9] with the Adam optimizer, which significantly impacted the final scores. SAM is a technique that considers the \"sharpness\" of the loss manifold in each gradient descent step, resulting in improved generalization. In SAM, the sharpness is computed from a fixed-size neighborhood of the current location in the loss manifold. However, the fixed-size neighborhood is not always ideal for sharpness computation. Adaptive SAM (ASAM) [15] proposed a size-varying neighborhood based on the scale to improve upon it. Taking a cue from this, we propose to utilize ASAM with AdamW optimizer to optimize our proposed framework. Through ablation analysis, we validate the effectiveness of ASAM in boosting the performance of the proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.729, + 0.589 + ], + "angle": 0, + "content": "4. Experimental Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.643 + ], + "angle": 0, + "content": "This section provides detailed information about the implementations and datasets, followed by the evaluation and ablation study." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.731, + 0.667 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We choose a pre-trained RegNetY200 initialized with ImageNet weights as the backbone in the spatio-temporal feature extractor. The model is trained using ASAM [15] with AdamW [20] as the base optimizer, setting \\(\\rho = 2\\). For training, a cosine annealing learning rate scheduler is employed with a warm-up of 3 epochs and an initial learning rate of 1e-5 during the warm-up. For the Soft-IC loss, a feature dimension of 128 and a memory bank size of 256 is used, with a loss weight (i.e., \\(\\lambda_{SIC}\\)) of 0.001. To enhance the training data, mixup [33] augmentation is applied with a mixing factor of \\(\\alpha = 0.1\\). Additionally, Gaussian blur is applied with a kernel size of 5 and a probability of 0.25. Brightness, contrast, and saturation adjustments are made with a factor ranging from 0.7 to 1.2, at a probability of 0.5. The proposed method is implemented in PyTorch and trained" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3167" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.168 + ], + "angle": 0, + "content": "on 4 A100 GPUs with a batch size of 2 on each GPU. At a time, 128 frames are processed, constituting a single sample (clip). During training, clips are uniformly sampled from videos, while during evaluation, a sliding window with half overlap is utilized to extract the clips." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.482, + 0.243 + ], + "angle": 0, + "content": "To validate the efficacy of the proposed method, we conducted experiments using event spotting datasets outlined in [14] such as Tennis [34], Figure Skating (FS) [13], Fine Gym and the SoccerNet V2 [5] action spotting dataset. These datasets are described in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.253, + 0.361, + 0.269 + ], + "angle": 0, + "content": "4.2. Evaluation Metric Description" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.483, + 0.487 + ], + "angle": 0, + "content": "Like previous works, we calculate the average precision (AP) within a specified tolerance range, denoted as \\(\\delta\\). The AP is then calculated for each event class to calculate the mean average precision (mAP) with the specified \\(\\delta\\) value as an evaluation metric. The \\(\\delta\\) is defined based on the respective datasets. For the Tennis, Figure Skating, and FineGym datasets, which have very strict evaluation criteria, we report the results with \\(\\delta = 0,1\\), and 2. The SoccerNet V2 dataset has relaxed evaluation criteria and defines two metrics: tight-mAP and loose-mAP, with \\(\\delta = 1 - 4\\) seconds and \\(\\delta = 5 - 60\\) seconds, respectively. Although the metrics are less strict compared to others, the SoccerNet V2 dataset is more challenging; having a more stringent metric would make the task even harder." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.497, + 0.263, + 0.512 + ], + "angle": 0, + "content": "4.3. Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.672 + ], + "angle": 0, + "content": "For the SoccerNet V2 [5] dataset, we compare existing SOTA methods proposed in conjunction with the SoccerNet Action Spotting challenge. Some of the results are taken from the 2021 challenge leaderboard, including Baidu (TF) [35], and SpotFormer [2]. We have also compared the results with CNN-based SOTA methods like E2ESpot [14], T-DEED [30], UGLF [27] and transformer-base SOTA methods like Spivak [23], ASTRA [29], and COMEDIAN [6]. The results are generated using the checkpoints provided by the respective authors whenever possible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.822 + ], + "angle": 0, + "content": "For the Tennis, Figure Skating, and FineGym dataset, we compare our method with the baseline method E2E-Spot [14]. We have also included the results of some best-performing pre-trained and fine-tuned variants as defined in [14]. These include TSP [1] + MS-TCN [8], TSP [1] + AS-Former [32], VC-Spot [14], and 2D-VDP [12] + MS-TCN [8]. We also compare the result with the recent state-of-the-art T-DEED [30] method. Unless otherwise stated, the results are mainly generated from the checkpoint provided by the authors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.456, + 0.849 + ], + "angle": 0, + "content": "4.4. Result Analysis on Various Sports Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.482, + 0.901 + ], + "angle": 0, + "content": "This section compares the results obtained by our proposed method with existing SOTA methods on SoccerNet V2, Tennis, Figure Skating & FineGym datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.201 + ], + "angle": 0, + "content": "Table 1. Result Comparison on SoccerNet-V2 datasets in terms of mAP tight and loose settings. Here, the best-performing measures are highlighted with ** bold font and second best-performing is highlighted with ** underline font. * result obtained from the challenge leaderboard of 2021. ** indicate that the GFLOPs value is calculated from the temporal network only without feature extractors. † indicate result generated from the author provided checkpoint. ‡ indicate results obtained from official repo." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.212, + 0.905, + 0.351 + ], + "angle": 0, + "content": "
ModelParams (M)GFLOPsTightLoose
Baidu (TF) [35]——47.05*73.77*
SpotFormer [2]——60.9076.10
E2E-Spot (RegNet-Y 200MF) [14]4.4639.6161.1973.25
E2E-Spot (RegNet-Y 800MF) [14]12.64151.4061.8274.05
Spivak [23]17.46461.8965.1078.50
ASTRA [29]44.338.83**66.6378.14
T-DEED (RegNet-Y 200MF) [30]16.3621.9639.43‡52.93‡
T-DEED (RegNet-Y 800MF) [30]46.2260.2521.57‡30.49‡
UGLF [27]4.4639.6163.5173.98‡
COMEDIAN (ViSwin-T) [6]70.12222.7671.33†77.07†
Proposed6.4660.2573.7479.11
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.907, + 0.655 + ], + "angle": 0, + "content": "SoccerNet V2: Table 1 compares the existing SOTA methods and the proposed approach on the SoccerNet V2 testing dataset. It is evident from the results that the proposed method has achieved notably strong performance in both tight and loose settings. The scores demonstrate significant improvement compared to both RegNet-Y 200MF and 800MF of E2E-Spot despite utilizing a similar framework. Additionally, there is a substantial improvement in the tight setting compared to the Baidu, Spivak, and ASTRA methods. These methods employ a pre-trained frozen larger backbone for feature extraction and transformer-based long-term temporal modules. COMEDIAN model utilizes a transformer network to achieve SOTA results in tight settings. However, the proposed method outperforms these methods without requiring a larger network. Notably, the proposed model achieves a \\(2.41\\%\\) improvement in tight settings than the COMEDIAN model while having significantly fewer parameters and FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Tennis, Figure Skating & FineGym: Table 2 shows the results for Tennis, Figure Skating settings, i.e., FS-Comp and FS-Perf, and FineGym with \\(\\delta = 0, 1\\) and 2. In this case, the proposed method also outperforms existing methods with a noticeable improvement. For instance, in case of \\(\\delta = 0\\), the proposed method achieves \\(15.67\\%\\), \\(8.64\\%\\), \\(8.49\\%\\) and \\(12.55\\%\\) improvements over the best-performing E2ESpot (RegNet-Y 200MF) and T-DEED (RegNet-Y 200MF) methods in Tennis, FS-Comp & FS-Perf and FineGym datasets, respectively. Similarly, for \\(\\delta = 1\\) case, the proposed method outperforms the previous best-performing T-DEED variants RegNet-Y 200MF and RegNet-Y 800MF by \\(1.52\\%\\) & \\(1.4\\%\\) improvement on FS-Comp & FS-Perf datasets, respectively. Additionally, we present the per-class score comparison on Tennis, Figure Skating, & FineGym datasets in Figure 4. Here, it can be seen that the perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3168" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 2. Result Comparison on Tennis, FS-Comp, FS-Perf and FinGym datasets in terms of mAP for different delta values. Here, the best-performing measures are highlighted with bold font and second best-performing is highlighted with underline font. * indicate result generated using provided checkpoints." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.131, + 0.905, + 0.297 + ], + "angle": 0, + "content": "
ModelTennisFS-CompFS-PerfFineGym
δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2
TSP [1] + MS-TCN [8]—90.9095.10—72.4087.80—76.8089.90—40.5058.50
TSP [1] + ASFormer [32]—89.8095.50—77.7094.10—80.2094.50—38.8057.60
2D-VDP [12] + MS-TCN [8]————83.5096.20—85.2096.40———
VC-Spot (RegNet-Y 200MF) [14]—92.4096.00—61.8075.50—56.2075.30—18.7028.60
E2E-Spot (RegNet-Y 200MF) [14]45.34*96.1097.7022.08*81.0093.5022.33*85.1095.7010.01*47.9065.20
E2E-Spot (RegNet-Y 800MF) [14]44.64*95.16*97.72*23.95*83.4094.9022.17*83.3096.0010.15*50.11*67.23*
T-DEED (RegNet-Y 200MF) [30]43.4996.1797.9528.92*85.1591.7031.87*86.7996.0513.54*53.73*66.43*
T-DEED (RegNet-Y 800MF) [30]44.0896.4297.0527.59*84.7792.8631.25*88.1795.8712.42*54.00*66.70*
UGLF [27]——————————50.2067.80
Proposed61.0196.2197.7537.5486.6795.5340.3689.5797.3815.2452.3166.57
" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.314, + 0.462, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.443, + 0.331, + 0.455 + ], + "angle": 0, + "content": "(a) Tennis dataset [34]" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.454, + 0.462, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.556, + 0.338, + 0.567 + ], + "angle": 0, + "content": "(c) FS-Comp dataset [13]" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.309, + 0.901, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.442, + 0.783, + 0.456 + ], + "angle": 0, + "content": "(b) FineGym dataset [22]" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.457, + 0.904, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.556, + 0.778, + 0.567 + ], + "angle": 0, + "content": "(d) FS-Perf dataset [13]" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.579, + 0.908, + 0.608 + ], + "angle": 0, + "content": "Figure 4. Per-class score comparison on \\(\\delta = 0\\) setting in terms of mAP for Tennis, FineGym, FS-Comp, and FS-Perf dataset. For FineGym, we report results for only 6 out of 32 classes, while results for all classes in the other datasets are reported." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.486, + 0.669 + ], + "angle": 0, + "content": "mance of the proposed method is not just good for a single class; it has performed consistently well across all the classes by a significant margin." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.255, + 0.695 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.748 + ], + "angle": 0, + "content": "We conduct ablation studies to validate the design choices made in the proposed framework. To ensure a fair comparison, we analyze the experimental results on SoccerNet V2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Importance of the proposed modules: Table 3 shows the contribution of the proposed ASTRM, Soft-IC loss, and ASAM components to achieve SOTA results. We observe a significant improvement in scores when replacing GSM [14] with the proposed ASTRM, highlighting the importance of effective feature extraction. SAM [9], utilized in the Spivak method, further enhances the results by a notable margin. Furthermore, we have illustrated how ASAM outperforms the scores achieved by SAM. Although ASAM and Soft-IC loss improves the overall scores by a small" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.618, + 0.887, + 0.733 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.742, + 0.896, + 0.756 + ], + "angle": 0, + "content": "Figure 5. Analysis on our method with & without Soft-IC loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.907, + 0.868 + ], + "angle": 0, + "content": "amount, Soft-IC loss plays an important role in improving the scores of minority classes. This is illustrated with scores of individual classes in Figure 5. As can be seen, Soft-IC loss improves the scores of classes with less number of samples without affecting other classes. This emphasizes the contributions of all components to achieve SOTA results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Components of the proposed ASTRM: The proposed AS-TRM module uses 3 major components: local spatial, local" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3169" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 3. Ablation studies on various aspects of the proposed framework on SoccerNet V2 testing dataset in terms of mAP." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.117, + 0.48, + 0.546 + ], + "angle": 0, + "content": "
MethodTightLoose
Analysis on proposed modules
Baseline: E2E-Spot (RegNet-Y 200MF + GSM)61.1973.25
RegNet-Y 200MF + ASTRM68.6375.57
RegNet-Y 200MF + ASTRM + SAM71.8977.50
RegNet-Y 200MF + ASTRM + ASAM72.6578.58
RegNet-Y 200MF + ASTRM + ASAM + Focal Loss70.3676.86
RegNet-Y 200MF + ASTRM + ASAM + IC Loss72.0578.30
RegNet-Y 200MF + ASTRM + ASAM + Soft-IC loss73.7479.11
Analysis on proposed ASTRM module
With local spatial only69.9176.39
With local temporal only70.4176.79
With global temporal only71.6577.06
With local spatial + local temporal71.0677.10
With local spatial + global temporal71.8277.85
With local temporal + global temporal71.5576.93
Proposed ASTRM module73.7479.11
Analysis on Different Temporal networks
Baseline with Bi-GRU61.1973.25
Proposed with Bi-GRU73.7479.11
with Deformable Attention67.4372.15
with Bi-GRU 2 Layers67.8573.03
with Transformer (L1H8)65.8970.76
with Transformer (L2H8)67.6072.35
with Bi-LSTM68.1573.35
with MSTCN68.7573.50
Analysis on Clip Length
Clip length = 6443.8252.09
Clip length = 100 (Used in baseline E2E-Spot)71.2076.73
Clip length = 128 (proposed)73.7479.11
Clip length = 19271.7477.02
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.868 + ], + "angle": 0, + "content": "temporal, and global temporal. Here, we highlight the importance of each component, and the corresponding findings are reported in Table 3. Notably, the score is at its lowest when utilizing only the spatial aspect of ASTRM; it fails to capture temporal information. Furthermore, the performance of local temporal is superior to using only the spatial aspect. Interestingly, we also observed that global temporal performs better than local temporal alone. This finding supports the notion that observing distant frames may be necessary to classify the current frame accurately. A similar pattern is seen when comparing local spatial + local temporal with local spatial + global temporal. Moreover, the inclusion of local spatial information leads to improved scores. Conversely, using only temporal information results in even lower scores, indicating the necessity of spatial information for understanding \"what is happening\" and temporal information for determining \"when it is happening.\" Overall, our observations highlight the need for spatial and local-global temporal information to identify events precisely and outperform other settings significantly." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Analysis on Different Temporal modules In our proposed approach, we opted for the Bidirectional GRU (Bi-GRU) as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.409 + ], + "angle": 0, + "content": "the long-range dependency module in the temporal block after conducting various experiments with different potential temporal networks. Table 3 presents the results obtained from the proposed approach using different temporal networks on the SoccerNet V2 [5] test set. The results indicate that the Bi-GRU, as a temporal network, performs favorably compared to other temporal networks. Even more advanced temporal networks, such as Bi-LSTM, Transformer, Deformable Attention [18] and MSTCN [8], could not surpass the scores achieved by the Bi-GRU temporal network. The difference in scores is visible across all tolerances. It is also observed that the proposed method generates significantly better results than the baseline E2E-Spot method, which also uses Bi-GRU as the temporal block. For instance, the proposed method achieves \\(12.55\\%\\) and \\(5.86\\%\\) improvements over the baseline model in the tight and loose settings of the SoccerNet V2 dataset, respectively. Similarly, there is an improvement of \\(15.67\\%\\) in the \\(\\delta = 0\\) setting compared to the baseline on the Tennis dataset. This significant improvement is attributed to the proposed ASTRM module and the proposed Soft-IC loss function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.909, + 0.579 + ], + "angle": 0, + "content": "Clip Length: Due to memory and computation constraints, it's not feasible to process all the frames simultaneously. Therefore, we process a fixed number of consecutive frames at a time, which we refer to as the \"clip length\". The value of the clip length affects the scores because it determines how many frames the network considers at once. While we might want to maximize the clip length, we need to find a balance between performance and clip length. The results in Table 3 support this idea. If the clip length is too short, it significantly impacts the results, while making it too long also has a negative effect on the results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.807, + 0.617 + ], + "angle": 0, + "content": "5. Conclusion & Future Directions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This paper proposes an end-to-end network for precise spotting of events in sports videos. It effectively tackles the long-range dependency and class imbalance issues and outperforms existing methods. To this extent, we introduce a new ASTRM module for enhancing spatio-temporal features. Further, to address the class imbalance issue, we propose Soft-IC loss, which enforces compact features and improves class separability. We further employed ASAM during training to ensure the model performs well on unseen data. While our method has advanced the state-of-the-art, especially in the more challenging cases with a relatively simple design, there is further scope for improvement on the two fronts we tried to address. Certain larger networks can improve the performance further but at the cost of computation. On the other hand, the spatial and temporal resolution at which the data is being processed will likely have an impact. The precise spotting may become easier with increased resolution, but it increases the processing burden." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3170" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Humam Alwassel, Silvio Giancola, and Bernard Ghanem. Tsp: Temporally-sensitive pretraining of video encoders for localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3173-3183, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[2] Mengqi Cao, Min Yang, Guozhen Zhang, Xiaotian Li, Yilu Wu, Gangshan Wu, and Limin Wang. Spotformer: A transformer-based framework for precise soccer action spotting. In 2022 IEEE 24th International Workshop on Multimedia Signal Processing (MMSP), pages 1-6. IEEE, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[4] Charika De Alvis et al. A survey of deep long-tail classification advancements. arXiv:2404.15593, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.484, + 0.47 + ], + "angle": 0, + "content": "[5] Adrien Deliège, Anthony Cioppa, Silvio Giancola, Meisam J. Seikavandi, Jacob V. Dueholm, Kamal Nasrollahi, Bernard Ghanem, Thomas B. Moeslund, and Marc Van Droogenbroeck. Soccertnet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4508-4519, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.483, + 0.556 + ], + "angle": 0, + "content": "[6] Julien Denize, Mykola Liashuha, Jaonary Rabarisoa, Astrid Orcesi, and Romain Hérault. Comedian: Self-supervised learning and knowledge distillation for action spotting using transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 530-540, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[7] Moritz Einfalt, Charles Dampeyrou, Dan Zecha, and Rainer Lienhart. Frame-level event detection in athletics videos with pose-based convolutional sequence networks. In Proceedings Proceedings of the 2nd International Workshop on Multimedia Content Analysis in Sports, page 42-50, New York, NY, USA, 2019. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[8] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[9] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[10] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccertnet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9591-9600, 2022. 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[12] James Hong, Matthew Fisher, Michael Gharbi, and Kayvon Fatahalian. Video pose distillation for few-shot, fine-grained sports action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9254-9263, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[13] James Hong, Matthew Fisher, Michael Gharbi, and Kayvon Fatahalian. Video pose distillation for few-shot, fine-grained sports action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9254-9263, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[14] James Hong, Haotian Zhang, Michael Gharbi, Matthew Fisher, and Kayvon Fatahalian. Spotting temporally precise, fine-grained events in video. In European Conference on Computer Vision, pages 33-51. Springer, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[15] Jungmin Kwon, Jeongseop Kim, Hyunseo Park, and In Kwon Choi. Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks. In Proceedings of the 38th International Conference on Machine Learning, pages 5905-5914. PMLR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.908, + 0.474 + ], + "angle": 0, + "content": "[17] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20010-20019, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[18] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[19] Zhaoyang Liu et al. Tam: Temporal adaptive module for video recognition. In ICCV, pages 13708-13718, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.906, + 0.602 + ], + "angle": 0, + "content": "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[21] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[22] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[23] Joao VB Soares and Avijit Shah. Action spotting using dense detection anchors revisited: Submission to the soccer-net challenge 2022. arXiv preprint arXiv:2206.07846, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[24] Joao VB Soares, Avijit Shah, and Topojoy Biswas. Temporally precise action spotting in soccer videos using dense detection anchors. In 2022 IEEE International Conference on Image Processing (ICIP), pages 2796-2800. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Gate-shift networks for video action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "3171" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[26] Matteo Tomei, Lorenzo Baraldi, Simone Calderara, Simone Bronzin, and Rita Cucchiara. Rms-net: Regression and masking for soccer event spotting. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7699-7706. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[27] Kim Hoang Tran, Phuc Vuong Do, Ngoc Quoc Ly, and Ngan Le. Unifying global and local scene entities modelling for precise action spotting. arXiv preprint arXiv:2404.09951 (Accepted for IJCNN 2024), 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[28] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[29] Artur Xarles, Sergio Escalera, Thomas B Moeslund, and Albert Clapés. Astra: An action spotting transformer for soccer videos. In Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports, pages 93-102, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[30] Artur Xarles, Sergio Escalera, Thomas B. Moeslund, and Albert Clapés. T-deed: Temporal-discriminability enhancer encoder-decoder for precise event spotting in sports videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3410-3419, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.472 + ], + "angle": 0, + "content": "[31] Huiran Yan et al. Emphasizing feature inter-class separability for improving highly imbalanced overlapped data classification. Knowledge-Based Systems, 276:110745, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[32] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. 32nd British Machine Vision Conference, BMVC 2021, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[33] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[34] Haotian Zhang, Cristobal Sciutto, Maneesh Agrawala, and Kayvon Fatahalian. Vid2player: Controllable video sprites that behave and appear like professional tennis players. ACM Trans. Graph., 40(3), 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[35] Xin Zhou, Le Kang, Zhiyu Cheng, Bo He, and Jingyu Xin. Feature combination meets attention: Baidu soccer embeddings and transformer based temporal detection. arXiv preprint arXiv:2106.14447, 2021. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3172" + } + ] +] \ No newline at end of file diff --git a/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/f70c89aa-2e9b-4112-89fc-f4837e98e688_origin.pdf b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/f70c89aa-2e9b-4112-89fc-f4837e98e688_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..122dc5308dd1bb278e7c00b1338258bee78c4c60 --- /dev/null +++ b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/f70c89aa-2e9b-4112-89fc-f4837e98e688_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02e52b726eccfd00d91529c9b7ef1b1453ab86a77affcb8fcf19675b9903524e +size 2128896 diff --git a/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/full.md b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ec5711f2c2242aaecddc12cc5539a6fc9c8cccfa --- /dev/null +++ b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/full.md @@ -0,0 +1,276 @@ +# Precise Event Spotting in Sports Videos: Solving Long-Range Dependency and Class Imbalance + +Sanchayan Santra $^{1}$ Vishal Chudasama $^{1}$ Pankaj Wasnik $^{1*}$ Vineeth N Balasubramanian $^{2}$ $^{1}$ Sony Research India $^{2}$ Indian Institute of Technology Hyderabad + +{sanchayan.santra, vishal.chudasamal, pankaj.wasnik}@sony.com, vineethnb@cse.iith.ac.in + +# Abstract + +Precise Event Spotting (PES) aims to identify events and their class from long, untrimmed videos, particularly in sports. The main objective of PES is to detect the event at the exact moment it occurs. Existing methods mainly rely on features from a large pre-trained network, which may not be ideal for the task. Furthermore, these methods overlook the issue of imbalanced event class distribution present in the data, negatively impacting performance in challenging scenarios. This paper demonstrates that an appropriately designed network, trained end-to-end, can outperform state-of-the-art (SOTA) methods. Particularly, we propose a network with a convolutional spatial-temporal feature extractor enhanced with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) and a long-range temporal module. The ASTRM enhances the features with spatiotemporal information. Meanwhile, the long-range temporal module helps extract global context from the data by modeling long-range dependencies. To address the class imbalance issue, we introduce the Soft Instance Contrastive (Soft-IC) loss that promotes feature compactness and class separation. Extensive experiments show that the proposed method is efficient and outperforms the SOTA methods, specifically in more challenging settings. + +# 1. Introduction + +Event spotting methods locate the occurrence of events in long, untrimmed videos. An event is typically defined by a starting and ending time, but this may not always be feasible, particularly in sports, where an event is a specific occurrence that occurs at a particular point in time, as dictated by the game's rules. For instance, a "Goal" in soccer is scored when the ball crosses the goal line, while a "Boundary" in cricket is made when the ball touches the boundary line. This process is referred to as Precise Event Spotting (PES), where PES methods aim to predict an event at the + +![](images/a145267ea273f846cb45b196f296fec5d4b7a2097d67da6ac81f028468589c81.jpg) +Figure 1. Temporal dependency of "Shots on target" event: The event is marked when the shot is taken, but whether the shot is at the target can only be known by looking at future frames. Here the shot is on target because the ball is saved by the goal keeper. + +![](images/4430d79e5bcdf58e0fabb04ee2b14382a55280ac93b69dbe00eacfcf07b0e97b.jpg) +Figure 2. Per class score analysis on a few classes of SoccerNet V2 dataset in tight setting. The proposed method outperforms state-of-the-art (SOTA) methods especially for classes with less number of samples. + +exact frame it occurs (e.g., goal, boundary). This is crucial for generating game summaries, finding specific events, and editing lengthy matches. However, it poses significant challenges, requiring a deep understanding of the game and its key events. + +Existing PES methods mainly face two challenges. First is the long-range temporal dependency that arises due to the nature of events. To predict the event frames, observing just the nearby frames does not suffice; several other potentially + +distant frames need to be observed to predict the event correctly. For example, "Shots on target" and "Shots off target" events of Soccer games are marked when the shot is taken. Still, the actual class is decided based on whether the ball went toward the goal or away from it (see Figure 1). This can be only detected if long-range dependencies are captured correctly. Second, there is an inherent imbalance in the number of samples among the classes (see Figure 2). Some events naturally occur more frequently than others. For example, in a soccer game, a "Ball out of play" is likely to occur more than a "Yellow card"; a "Red-card" and "Yellow $\rightarrow$ Red card" are expected to occur even less often. Therefore, the same trend can be observed in datasets created using broadcast videos. Due to these issues, existing event spotting methods [6, 14] perform poorly in underrepresented classes such as "Red card" & "Yellow $\rightarrow$ Red card." This can be visualized in Figure 2. Handling these issues and accurately performing PES is an active area of research. + +To address the long-range temporal dependency issue, recent SOTA methods [24, 29] prefer to use the feature extracted using the pre-trained network of [35]. These methods also use transformer-based temporal modules due to their better sequence modeling capability in a global context. However, as stated in [18, 32], transformers also encounter issues with long videos. Self-attention struggles to learn appropriate weights for long videos due to the increase in the size of the weight matrix with the number of frames being processed. On the other hand, E2ESpot [14] proposed an end-to-end trainable feature extractor and a GRU-based module for temporal modeling. Recently proposed T-DEED [30] improves the performance of E2ESpot via temporal-discriminability enhancing encoder-decoder (T-DEED) module. In our proposed method, we handle this issue by enhancing the extracted features with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) and GRU-based temporal module. + +The class imbalance issue has not received sufficient attention in the literature on PES, resulting in lower performance on real-world datasets. The recent method ASTRA [29] is one of the few attempts to address this problem through a modified "mixup" augmentation. Although standard machine learning approaches, such as focal loss [16], are available, they are not suitable for every task, as this paper demonstrates. To address this challenge, we introduce a novel loss function called Soft Instance Contrastive (Soft-IC) loss. This loss function aims to mitigate class imbalance by promoting compact feature representations and improving class separability. + +The primary goal of this paper is to address these issues and improve the accuracy of PES. To this end, we propose an end-to-end trainable network with a CNN-based spatiotemporal feature extractor and a GRU-based long-term temporal module. The spatio-temporal feature extractor in- + +cludes our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) to augment the features with spatial and temporal information. We chose to train the model end-to-end because it has been shown to perform better than a pre-trained frozen feature extractor [17, 26]. We have also employed the Adaptive Sharpness-Aware Minimization (ASAM) [15] to enable the trained model to perform well on unseen data. Through experiments, we demonstrate that the proposed framework obtains superior performance compared to SOTA methods, especially in tight settings where long-range dependency and class imbalance issues are prominent (see Figure 2). Our key contributions can be summarized as follows: + +- We propose an end-to-end framework using a CNN-based spatio-temporal feature extractor and GRU-based temporal network to achieve precise event spotting while addressing long-range dependency and class imbalance. +- We introduce ASTRM, which enhances the features with spatial and temporal information. +- We also introduce Soft-IC loss to enforce compact features and improve class separability, addressing the class imbalance issue. + +# 2. Related Works + +The concept of precise sports event spotting first appeared in Einfalt et al. [7], where they considered temporally precise detection of various events related to long and triple jump. Later, SoccerNet [10] proposed a large-scale dataset of soccer videos to identify specific events or actions at exact timestamps. The work by [14] expanded this to precise event spotting (PES), introducing significant differences in the necessary precision of predictions. PES is constrained to the tolerance of a few frames only. Recently, several PES methods [6, 14, 24, 27, 29, 30, 35] have been proposed that focus on various sports videos. E2E-Spot [14] employed a CNN-based backbone with a Gated Shift Module (GSM) [25] for temporal modeling and a GRU [3] for producing per-frame prediction. Even though it is effective on the Tennis and Figure Skating datasets, it struggles on more challenging SoccerNet V2 dataset [5] due to its simplistic temporal information extraction scheme using GSM. + +Building upon E2E-Spot [14], T-DEED [30] proposed improvements to address issues arising from tight tolerances in fast-paced sporting events and achieved SOTA performance on the Figure Skating and Figure Diving datasets. The method of [35] revamped the feature extraction using a group of five networks pre-trained on Kinetics-400 and fine-tuned on SoccerNet V2. Several other methods use the features extracted by this method instead of raw video frames. For instance, Soares et al. [24] utilized these features and developed a U-Net- and transformer-based trunk for making predictions. ASTRA [29] utilizes the same features but incorporates audio features for complementary in + +formation. In contrast, COMEDIAN [6] operates in 3 stages to train their network on raw frames: 1) training only the spatial part; 2) training the spatial and temporal parts using knowledge distillation; 3) final tuning to optimize the classification scores. Recently, UGLF [27] proposed a unifying global-local (UGL) module to disentangle the scene content into global features and locally relevant scene entity features in conjunction with a Vision-Language model. + +The above-mentioned methods mainly addressed the long-range temporal dependency issue by utilizing pretrained features and transformers. They have not adapted their feature extractors to capture spatial and temporal information effectively. Additionally, these methods did not specifically deal with the class imbalance, which led to poor performance in underrepresented classes. In our framework, we propose ASTRM, which helps the feature extractor to augment the features with relevant spatial and temporal information. Furthermore, we introduce Soft-IC loss to address the class imbalance issue and employ the ASAM to train the proposed framework properly. With these advancements, the proposed framework outperforms existing event-spotting methods on various sports datasets, particularly in settings with tighter tolerances. + +# 3. Proposed Methodology + +This section presents our methodology in detail, starting with the problem statement and then presenting the details of the proposed modules as follows. + +# 3.1. Problem Statement + +The task of precise event spotting is defined as follows: Given a video with $N$ frames $\mathbf{x}_1, \mathbf{x}_2, \ldots, \mathbf{x}_N$ and a set of possible events $\{c_1, \ldots, c_K\}$ , the goal is to predict frames where an event occurs along with its class: $(t, \hat{y}_t) \in \mathbb{N} \times \{c_1, \ldots, c_K\}$ . A prediction is deemed correct if it falls within $\delta$ frames of the actual occurrence and accurately identifies the event class. Since it is not possible to process the entire video at once, the processing is done in chunks of a predefined size. Each chunk (or clip), represented as $\mathbf{X} \in \mathbb{R}^{C \times T_s \times H \times W}$ , consists of $T_s$ frames. We compute predictions for each frame, represented as $\mathbf{Y} \in \mathbb{R}^{T_s \times K}$ , where $K$ is the number of classes. As the video is processed in segments, the predictions from each chunk of $T_s$ frames are aggregated and refined with a non-maximal suppression to get the final sparse prediction. + +# 3.2. Methodology + +Figure 3 shows the network design of our proposed framework, which takes a set of input frames and classifies whether each frame contains an event or not. The framework comprises three main components: the spatiotemporal feature extractor, the temporal block, and the classifier. In the spatio-temporal feature extractor, we + +have selected the RegNetY [21] as the feature extractor due to its efficiency, surpassing commonly used networks like ResNet. This network is further enhanced with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM). This module is integrated into each bottleneck block, refining the features by incorporating both spatial and temporal information. Both of this information is necessary to correctly identify events, as validated by the experiments. The refined features are then passed to the temporal block, which enriches them with long-term temporal information. This long-term information is crucial because, in many situations, information from temporally distant frames is essential to accurately classify a frame. The spatio-temporal feature extractor alone can not accomplish this task. Finally, the feature maps are forwarded to the classifier, which is a simple linear layer that outputs the classification scores. Furthermore, we introduce the Soft Instance Contrastive (Soft-IC) loss to effectively address the imbalance among the class samples. This loss function operates by compacting the features and enhancing their separability. The key modules of the proposed framework are detailed in subsequent subsections. + +# 3.2.1 Adaptive Spatio-Temporal Refinement Module: + +This module enhances the extracted features by incorporating both spatial and temporal information. It comprises three blocks: Local Spatial, Local Temporal, and Global Temporal. We include the Local Spatial block because local spatial features also offer important cues about temporal information. Relying solely on temporal information can overlook these valuable insights. The spatial component is inspired by [28], while the temporal aspect is adapted from [19]. The refinement is done as follows: + +$$ +\Psi (\mathbf {x}) = \left(\left(\mathbf {x} \odot \left(1 + \mathcal {F} _ {s} (\mathbf {x})\right)\right) \odot \left(1 + \mathcal {F} _ {t} (\mathbf {x})\right)\right) * \mathcal {G} _ {t} (\mathbf {x}) \tag {1} +$$ + +where $\mathbf{x}$ is the input feature to be enhanced. $\mathcal{F}_s, \mathcal{F}_t,$ and $\mathcal{G}_t$ denote the operation corresponding to the local spatial, local temporal, and global temporal block, respectively. $\odot$ denotes Hadamard Product, while $*$ denotes convolution operation. To elaborate, the input feature is weighted with information from the local spatial block and a residual connection, resulting in features enhanced with local spatial information. These enhanced features are further refined with local temporal information. Finally, spatially and temporally enhanced features are then convolved with the kernel obtained from the global temporal block to make local enhancements using "global" information. This dynamic kernel approach simplifies the network while maintaining effectiveness. Please refer to Figure 3 for a visual representation. + +In the Local Spatial block, the input features are pooled using Max (i.e., $f_{MaxPool}$ ) and Average (i.e., $f_{AvgPool}$ ) + +![](images/e46dd065db54d0faa2035b4b74c2249192712b2832c5b8d92f49c1880a13fb4a.jpg) +Figure 3. Network design of the proposed framework. The framework composed of a spatio-temporal feature extractor and a temporal block for capturing long-range dependency before the classifier. In each bottleneck block we add ASTRM after the first conv. ASTRM further enhances the features with local spatial, local temporal and global temporal information. The network is trained with Soft-IC loss in addition to the classification loss to handle class imbalance. + +pooling operations along the channel dimension. The pooled features are then concatenated and convolved with a $7 \times 7$ filter (i.e., $f_{conv}^{7 \times 7}$ ). The pooling operation on the channel axis helps emphasize informative regions. The subsequent convolution operation filters out non-informative portions from the features. Finally, the output is passed through a Sigmoid function (i.e., $\sigma(\cdot)$ ) to obtain the "spatial" feature weights (i.e., $\mathcal{F}_s(\mathbf{x})$ ) as follow: + +$$ +\mathcal {F} _ {s} (\mathbf {x}) = \sigma \left(f _ {c o n v} ^ {7 \times 7} \left(\left[ f _ {A v g P o o l} (\mathbf {x}); f _ {M a x P o o l} (\mathbf {x}) \right]\right)\right). \tag {2} +$$ + +In the Local Temporal block, we use 3D convolution operations, namely $f_{conv}^{3 \times 1 \times 1}$ and $f_{conv}^{1 \times 1 \times 1}$ , followed by the Sigmoid function, $\sigma(\cdot)$ , with a focus on the temporal dimension. Spatially pooled features are not used in this block because they do not account for the different temporal changes at various spatial locations. The locally temporal enhanced features are defined as below: + +$$ +\mathcal {F} _ {t} (\mathbf {x}) = \sigma \left(f _ {\text {c o n v}} ^ {1 \times 1 \times 1} \left(\mathrm {B N} \left(\operatorname {R e L U} \left(f _ {\text {c o n v}} ^ {3 \times 1 \times 1} (\mathbf {x})\right)\right)\right)\right). \tag {3} +$$ + +In the Global Temporal block, we adaptively create convolution kernels to improve the feature maps by considering temporal information across frames. First, the features are pooled in the spatial dimension using a global average pooling operation (i.e., $f_{GAP}$ ), and then fully connected layers (i.e., $f_{FC}$ ) are used to compute the kernel with a global view of the data with a softmax ( $S$ ) at the end. Mathematically, the kernel is formulated as below: + +$$ +\mathcal {G} _ {t} (\mathbf {x}) = \mathcal {S} \left(f _ {F C} \left(f _ {F C} \left(f _ {G A P} (\mathbf {x})\right)\right)\right). \tag {4} +$$ + +# 3.2.2 Temporal Block: + +The spatio-temporal feature extraction stage computes features containing local spatial, local temporal, and location-independent temporal information. Since we are interested in finding the temporal location of the events, we aggregate the features in the spatial dimension with average pooling before extracting long-range temporal information. We use temporal block because the feature extractor alone is not capable enough to capture the required long-range dependency, although it may help to some extent. As motivated in the introduction section, we need to capture long-range temporal information to accurately predict events with precision. For this task, we selected Bidirectional GRU (Bi-GRU) as the temporal block because it effectively models sequences without overfitting the input data. The Bi-GRU is capable of quickly adapting to the data while maintaining its generalizability. Further experiments confirm this, demonstrating that the Bi-GRU performs well compared to more complex temporal blocks. + +# 3.2.3 Soft Instance Contrastive (Soft-IC) loss: + +To address the class imbalance issue, we introduce a novel Soft-IC loss, which aims to make the features more compact and increase their separability. It is inspired by the Instance Contrastive loss (IC loss) [11] and is designed to work with mixup augmentation [33]. IC loss enhances the separation among the features by promoting its compactness. In case + +of class imbalance, features may deviate from their respective class center and potentially overlap with features from other classes. This can hinder the learning of a decision boundary and negatively impact performance [4]. By compacting the features, we indirectly address the challenges posed by class imbalance, a strategy supported by research in the field [31]. For a batch of $N$ features computed from the input, the IC Loss is calculated as described in [11]. + +$$ +\mathcal {L} _ {I C} = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathcal {L} (\mathbf {z} _ {i}), \tag {5} +$$ + +$$ +\mathcal {L} \left(\mathbf {z} _ {i}\right) = \frac {1}{\left| M \left(\mathbf {c} _ {i}\right) \right|} \sum_ {\mathbf {z} _ {j} \in M \left(\mathbf {c} _ {i}\right)} \log \frac {\exp \left(\mathbf {z} _ {i} \cdot \mathbf {z} _ {j} / \tau\right)}{\sum_ {\mathbf {z} _ {\mathrm {k}} \in \mathbf {A} \left(\mathbf {c} _ {\mathrm {i}}\right)} \exp \left(\mathbf {z} _ {i} \cdot \mathbf {z} _ {k} / \tau\right)}, \tag {6} +$$ + +where $\mathbf{c}_i$ is the label of $i$ -th sample, $\tau$ is a temperature hyper-parameter, $M(\mathbf{c}_i)$ denotes the memory bank of class $\mathbf{c}_i$ and $A(\mathbf{c}_i) = M\backslash M(\mathbf{c}_i)$ . + +During training, we use mixup augmentation [33] to regularize the model and improve its generalization. In this, two samples (and their labels) are combined convexly to create a new sample (and its label) as follows. + +$$ +\tilde {x} = \lambda x _ {i} + (1 - \lambda) x _ {j}, \tag {7} +$$ + +$$ +\tilde {y} = \lambda y _ {i} + (1 - \lambda) y _ {j}, \tag {8} +$$ + +where $x_{i}$ and $x_{j}$ are two samples from the training data, and $y_{i}$ and $y_{j}$ are their corresponding one-hot labels. The parameter $\lambda$ takes value in the range [0, 1]. Consequently, the label $(\tilde{y})$ will not be in one-hot format; instead, it will reflect weighted contributions from each class. + +IC Loss assumes that each sample has a single label. However, this assumption does not hold true when using mixup augmentation. To account for weight associated with each class in the label, we define Soft-IC loss as + +$$ +\mathcal {L} _ {S I C} = \frac {\sum_ {\left(\mathbf {z} _ {i} , w _ {i j}\right) , w _ {i j} \neq 0} \mathcal {L} _ {S I C} \left(\mathbf {z} _ {i} , w _ {i j}\right)}{\sum_ {\left(\mathbf {z} _ {i} , w _ {i j}\right) , w _ {i j} \neq 0} 1}, \tag {9} +$$ + +$$ +\mathcal {L} _ {S I C} (\mathbf {z} _ {i}, \omega) = \frac {1}{\omega | M (\mathbf {c} _ {i}) |} \times +$$ + +$$ +\sum_ {(\mathbf {z} _ {j}, w _ {j}) \in M (\mathbf {c} _ {i})} \log \frac {\exp \left(\mathbf {z} _ {i} \cdot w _ {j} \mathbf {z} _ {j} / \tau\right)}{\sum_ {(\mathbf {z} _ {\mathbf {k}} , w _ {k}) \in A (\mathbf {c} _ {i})} \exp \left(\mathbf {z} _ {i} \cdot w _ {k} \mathbf {z} _ {k} / \tau\right)}. \tag {10} +$$ + +where, $w_{ij}$ is the weight associated with the label of class $c_i$ for sample $\mathbf{z}_i$ . Additionally, in the memory bank $M(\mathbf{c}_i)$ , we store the corresponding class weights along with features. + +# 3.3. Training Strategy + +The proposed framework is trained end-to-end using a combination of two loss functions: Binary Cross Entropy loss $(\mathcal{L}_{BCE})$ as the standard classification loss and the proposed Soft-IC loss $(\mathcal{L}_{SIC})$ . The overall loss function $(\mathcal{L}_{final})$ can be expressed as below: + +$$ +\mathcal {L} _ {\text {f i n a l}} (\cdot) = \mathcal {L} _ {B C E} + \lambda_ {S I C} \cdot \mathcal {L} _ {S I C}. \tag {11} +$$ + +Here, $\lambda_{SIC}$ denotes the weighting constant that adjusts the contributions of different loss functions. With the total loss, the proposed network is trained using the AdamW [20] optimizer with Adaptive Sharpness-Aware Minimization (ASAM) [15]. + +Utilizing advanced optimization techniques can significantly impact the performance of the trained model, yet not much focus is given to this aspect. In [23], the authors utilized the Sharpness-Aware Minimization (SAM) [9] with the Adam optimizer, which significantly impacted the final scores. SAM is a technique that considers the "sharpness" of the loss manifold in each gradient descent step, resulting in improved generalization. In SAM, the sharpness is computed from a fixed-size neighborhood of the current location in the loss manifold. However, the fixed-size neighborhood is not always ideal for sharpness computation. Adaptive SAM (ASAM) [15] proposed a size-varying neighborhood based on the scale to improve upon it. Taking a cue from this, we propose to utilize ASAM with AdamW optimizer to optimize our proposed framework. Through ablation analysis, we validate the effectiveness of ASAM in boosting the performance of the proposed framework. + +# 4. Experimental Analysis + +This section provides detailed information about the implementations and datasets, followed by the evaluation and ablation study. + +# 4.1. Implementation Details + +We choose a pre-trained RegNetY200 initialized with ImageNet weights as the backbone in the spatio-temporal feature extractor. The model is trained using ASAM [15] with AdamW [20] as the base optimizer, setting $\rho = 2$ . For training, a cosine annealing learning rate scheduler is employed with a warm-up of 3 epochs and an initial learning rate of 1e-5 during the warm-up. For the Soft-IC loss, a feature dimension of 128 and a memory bank size of 256 is used, with a loss weight (i.e., $\lambda_{SIC}$ ) of 0.001. To enhance the training data, mixup [33] augmentation is applied with a mixing factor of $\alpha = 0.1$ . Additionally, Gaussian blur is applied with a kernel size of 5 and a probability of 0.25. Brightness, contrast, and saturation adjustments are made with a factor ranging from 0.7 to 1.2, at a probability of 0.5. The proposed method is implemented in PyTorch and trained + +on 4 A100 GPUs with a batch size of 2 on each GPU. At a time, 128 frames are processed, constituting a single sample (clip). During training, clips are uniformly sampled from videos, while during evaluation, a sliding window with half overlap is utilized to extract the clips. + +To validate the efficacy of the proposed method, we conducted experiments using event spotting datasets outlined in [14] such as Tennis [34], Figure Skating (FS) [13], Fine Gym and the SoccerNet V2 [5] action spotting dataset. These datasets are described in the supplementary material. + +# 4.2. Evaluation Metric Description + +Like previous works, we calculate the average precision (AP) within a specified tolerance range, denoted as $\delta$ . The AP is then calculated for each event class to calculate the mean average precision (mAP) with the specified $\delta$ value as an evaluation metric. The $\delta$ is defined based on the respective datasets. For the Tennis, Figure Skating, and FineGym datasets, which have very strict evaluation criteria, we report the results with $\delta = 0,1$ , and 2. The SoccerNet V2 dataset has relaxed evaluation criteria and defines two metrics: tight-mAP and loose-mAP, with $\delta = 1 - 4$ seconds and $\delta = 5 - 60$ seconds, respectively. Although the metrics are less strict compared to others, the SoccerNet V2 dataset is more challenging; having a more stringent metric would make the task even harder. + +# 4.3. Baseline Methods + +For the SoccerNet V2 [5] dataset, we compare existing SOTA methods proposed in conjunction with the SoccerNet Action Spotting challenge. Some of the results are taken from the 2021 challenge leaderboard, including Baidu (TF) [35], and SpotFormer [2]. We have also compared the results with CNN-based SOTA methods like E2ESpot [14], T-DEED [30], UGLF [27] and transformer-base SOTA methods like Spivak [23], ASTRA [29], and COMEDIAN [6]. The results are generated using the checkpoints provided by the respective authors whenever possible. + +For the Tennis, Figure Skating, and FineGym dataset, we compare our method with the baseline method E2E-Spot [14]. We have also included the results of some best-performing pre-trained and fine-tuned variants as defined in [14]. These include TSP [1] + MS-TCN [8], TSP [1] + AS-Former [32], VC-Spot [14], and 2D-VDP [12] + MS-TCN [8]. We also compare the result with the recent state-of-the-art T-DEED [30] method. Unless otherwise stated, the results are mainly generated from the checkpoint provided by the authors. + +# 4.4. Result Analysis on Various Sports Datasets + +This section compares the results obtained by our proposed method with existing SOTA methods on SoccerNet V2, Tennis, Figure Skating & FineGym datasets. + +Table 1. Result Comparison on SoccerNet-V2 datasets in terms of mAP tight and loose settings. Here, the best-performing measures are highlighted with ** bold font and second best-performing is highlighted with ** underline font. * result obtained from the challenge leaderboard of 2021. ** indicate that the GFLOPs value is calculated from the temporal network only without feature extractors. † indicate result generated from the author provided checkpoint. ‡ indicate results obtained from official repo. + +
ModelParams (M)GFLOPsTightLoose
Baidu (TF) [35]——47.05*73.77*
SpotFormer [2]——60.9076.10
E2E-Spot (RegNet-Y 200MF) [14]4.4639.6161.1973.25
E2E-Spot (RegNet-Y 800MF) [14]12.64151.4061.8274.05
Spivak [23]17.46461.8965.1078.50
ASTRA [29]44.338.83**66.6378.14
T-DEED (RegNet-Y 200MF) [30]16.3621.9639.43‡52.93‡
T-DEED (RegNet-Y 800MF) [30]46.2260.2521.57‡30.49‡
UGLF [27]4.4639.6163.5173.98‡
COMEDIAN (ViSwin-T) [6]70.12222.7671.33†77.07†
Proposed6.4660.2573.7479.11
+ +SoccerNet V2: Table 1 compares the existing SOTA methods and the proposed approach on the SoccerNet V2 testing dataset. It is evident from the results that the proposed method has achieved notably strong performance in both tight and loose settings. The scores demonstrate significant improvement compared to both RegNet-Y 200MF and 800MF of E2E-Spot despite utilizing a similar framework. Additionally, there is a substantial improvement in the tight setting compared to the Baidu, Spivak, and ASTRA methods. These methods employ a pre-trained frozen larger backbone for feature extraction and transformer-based long-term temporal modules. COMEDIAN model utilizes a transformer network to achieve SOTA results in tight settings. However, the proposed method outperforms these methods without requiring a larger network. Notably, the proposed model achieves a $2.41\%$ improvement in tight settings than the COMEDIAN model while having significantly fewer parameters and FLOPs. + +Tennis, Figure Skating & FineGym: Table 2 shows the results for Tennis, Figure Skating settings, i.e., FS-Comp and FS-Perf, and FineGym with $\delta = 0, 1$ and 2. In this case, the proposed method also outperforms existing methods with a noticeable improvement. For instance, in case of $\delta = 0$ , the proposed method achieves $15.67\%$ , $8.64\%$ , $8.49\%$ and $12.55\%$ improvements over the best-performing E2ESpot (RegNet-Y 200MF) and T-DEED (RegNet-Y 200MF) methods in Tennis, FS-Comp & FS-Perf and FineGym datasets, respectively. Similarly, for $\delta = 1$ case, the proposed method outperforms the previous best-performing T-DEED variants RegNet-Y 200MF and RegNet-Y 800MF by $1.52\%$ & $1.4\%$ improvement on FS-Comp & FS-Perf datasets, respectively. Additionally, we present the per-class score comparison on Tennis, Figure Skating, & FineGym datasets in Figure 4. Here, it can be seen that the perfor + +Table 2. Result Comparison on Tennis, FS-Comp, FS-Perf and FinGym datasets in terms of mAP for different delta values. Here, the best-performing measures are highlighted with bold font and second best-performing is highlighted with underline font. * indicate result generated using provided checkpoints. + +
ModelTennisFS-CompFS-PerfFineGym
δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2
TSP [1] + MS-TCN [8]—90.9095.10—72.4087.80—76.8089.90—40.5058.50
TSP [1] + ASFormer [32]—89.8095.50—77.7094.10—80.2094.50—38.8057.60
2D-VDP [12] + MS-TCN [8]————83.5096.20—85.2096.40———
VC-Spot (RegNet-Y 200MF) [14]—92.4096.00—61.8075.50—56.2075.30—18.7028.60
E2E-Spot (RegNet-Y 200MF) [14]45.34*96.1097.7022.08*81.0093.5022.33*85.1095.7010.01*47.9065.20
E2E-Spot (RegNet-Y 800MF) [14]44.64*95.16*97.72*23.95*83.4094.9022.17*83.3096.0010.15*50.11*67.23*
T-DEED (RegNet-Y 200MF) [30]43.4996.1797.9528.92*85.1591.7031.87*86.7996.0513.54*53.73*66.43*
T-DEED (RegNet-Y 800MF) [30]44.0896.4297.0527.59*84.7792.8631.25*88.1795.8712.42*54.00*66.70*
UGLF [27]——————————50.2067.80
Proposed61.0196.2197.7537.5486.6795.5340.3689.5797.3815.2452.3166.57
+ +![](images/f32d2124f031badae1e4917ca47ac56f1efc47c2ab986de5f8598f64c566c5b8.jpg) +(a) Tennis dataset [34] + +![](images/0cb9a6e62b8e5d77982a536536f8afed6881a95c08390c0822a854efaff96067.jpg) +(c) FS-Comp dataset [13] + +![](images/97c5c9fe5513194355556825f9706982ba56a9a57f5c452e3e90c10fd0f96981.jpg) + +![](images/da0e9b2119e7b616cc6ab177c3c9aff1a62cbeefd10a763a73c70549640a1e95.jpg) +(b) FineGym dataset [22] +(d) FS-Perf dataset [13] + +mance of the proposed method is not just good for a single class; it has performed consistently well across all the classes by a significant margin. + +# 4.5. Ablation Studies + +We conduct ablation studies to validate the design choices made in the proposed framework. To ensure a fair comparison, we analyze the experimental results on SoccerNet V2. + +Importance of the proposed modules: Table 3 shows the contribution of the proposed ASTRM, Soft-IC loss, and ASAM components to achieve SOTA results. We observe a significant improvement in scores when replacing GSM [14] with the proposed ASTRM, highlighting the importance of effective feature extraction. SAM [9], utilized in the Spivak method, further enhances the results by a notable margin. Furthermore, we have illustrated how ASAM outperforms the scores achieved by SAM. Although ASAM and Soft-IC loss improves the overall scores by a small + +![](images/839dba5ad5e8504802bb00a446c93dbd06a3cdc7424ee79323f51a8a14495cfb.jpg) +Figure 4. Per-class score comparison on $\delta = 0$ setting in terms of mAP for Tennis, FineGym, FS-Comp, and FS-Perf dataset. For FineGym, we report results for only 6 out of 32 classes, while results for all classes in the other datasets are reported. +Figure 5. Analysis on our method with & without Soft-IC loss. + +amount, Soft-IC loss plays an important role in improving the scores of minority classes. This is illustrated with scores of individual classes in Figure 5. As can be seen, Soft-IC loss improves the scores of classes with less number of samples without affecting other classes. This emphasizes the contributions of all components to achieve SOTA results. + +Components of the proposed ASTRM: The proposed AS-TRM module uses 3 major components: local spatial, local + +Table 3. Ablation studies on various aspects of the proposed framework on SoccerNet V2 testing dataset in terms of mAP. + +
MethodTightLoose
Analysis on proposed modules
Baseline: E2E-Spot (RegNet-Y 200MF + GSM)61.1973.25
RegNet-Y 200MF + ASTRM68.6375.57
RegNet-Y 200MF + ASTRM + SAM71.8977.50
RegNet-Y 200MF + ASTRM + ASAM72.6578.58
RegNet-Y 200MF + ASTRM + ASAM + Focal Loss70.3676.86
RegNet-Y 200MF + ASTRM + ASAM + IC Loss72.0578.30
RegNet-Y 200MF + ASTRM + ASAM + Soft-IC loss73.7479.11
Analysis on proposed ASTRM module
With local spatial only69.9176.39
With local temporal only70.4176.79
With global temporal only71.6577.06
With local spatial + local temporal71.0677.10
With local spatial + global temporal71.8277.85
With local temporal + global temporal71.5576.93
Proposed ASTRM module73.7479.11
Analysis on Different Temporal networks
Baseline with Bi-GRU61.1973.25
Proposed with Bi-GRU73.7479.11
with Deformable Attention67.4372.15
with Bi-GRU 2 Layers67.8573.03
with Transformer (L1H8)65.8970.76
with Transformer (L2H8)67.6072.35
with Bi-LSTM68.1573.35
with MSTCN68.7573.50
Analysis on Clip Length
Clip length = 6443.8252.09
Clip length = 100 (Used in baseline E2E-Spot)71.2076.73
Clip length = 128 (proposed)73.7479.11
Clip length = 19271.7477.02
+ +temporal, and global temporal. Here, we highlight the importance of each component, and the corresponding findings are reported in Table 3. Notably, the score is at its lowest when utilizing only the spatial aspect of ASTRM; it fails to capture temporal information. Furthermore, the performance of local temporal is superior to using only the spatial aspect. Interestingly, we also observed that global temporal performs better than local temporal alone. This finding supports the notion that observing distant frames may be necessary to classify the current frame accurately. A similar pattern is seen when comparing local spatial + local temporal with local spatial + global temporal. Moreover, the inclusion of local spatial information leads to improved scores. Conversely, using only temporal information results in even lower scores, indicating the necessity of spatial information for understanding "what is happening" and temporal information for determining "when it is happening." Overall, our observations highlight the need for spatial and local-global temporal information to identify events precisely and outperform other settings significantly. + +Analysis on Different Temporal modules In our proposed approach, we opted for the Bidirectional GRU (Bi-GRU) as + +the long-range dependency module in the temporal block after conducting various experiments with different potential temporal networks. Table 3 presents the results obtained from the proposed approach using different temporal networks on the SoccerNet V2 [5] test set. The results indicate that the Bi-GRU, as a temporal network, performs favorably compared to other temporal networks. Even more advanced temporal networks, such as Bi-LSTM, Transformer, Deformable Attention [18] and MSTCN [8], could not surpass the scores achieved by the Bi-GRU temporal network. The difference in scores is visible across all tolerances. It is also observed that the proposed method generates significantly better results than the baseline E2E-Spot method, which also uses Bi-GRU as the temporal block. For instance, the proposed method achieves $12.55\%$ and $5.86\%$ improvements over the baseline model in the tight and loose settings of the SoccerNet V2 dataset, respectively. Similarly, there is an improvement of $15.67\%$ in the $\delta = 0$ setting compared to the baseline on the Tennis dataset. This significant improvement is attributed to the proposed ASTRM module and the proposed Soft-IC loss function. + +Clip Length: Due to memory and computation constraints, it's not feasible to process all the frames simultaneously. Therefore, we process a fixed number of consecutive frames at a time, which we refer to as the "clip length". The value of the clip length affects the scores because it determines how many frames the network considers at once. While we might want to maximize the clip length, we need to find a balance between performance and clip length. The results in Table 3 support this idea. If the clip length is too short, it significantly impacts the results, while making it too long also has a negative effect on the results. + +# 5. Conclusion & Future Directions + +This paper proposes an end-to-end network for precise spotting of events in sports videos. It effectively tackles the long-range dependency and class imbalance issues and outperforms existing methods. To this extent, we introduce a new ASTRM module for enhancing spatio-temporal features. Further, to address the class imbalance issue, we propose Soft-IC loss, which enforces compact features and improves class separability. We further employed ASAM during training to ensure the model performs well on unseen data. While our method has advanced the state-of-the-art, especially in the more challenging cases with a relatively simple design, there is further scope for improvement on the two fronts we tried to address. Certain larger networks can improve the performance further but at the cost of computation. On the other hand, the spatial and temporal resolution at which the data is being processed will likely have an impact. The precise spotting may become easier with increased resolution, but it increases the processing burden. + +# References + +[1] Humam Alwassel, Silvio Giancola, and Bernard Ghanem. Tsp: Temporally-sensitive pretraining of video encoders for localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3173-3183, 2021. 6, 7 +[2] Mengqi Cao, Min Yang, Guozhen Zhang, Xiaotian Li, Yilu Wu, Gangshan Wu, and Limin Wang. Spotformer: A transformer-based framework for precise soccer action spotting. In 2022 IEEE 24th International Workshop on Multimedia Signal Processing (MMSP), pages 1-6. IEEE, 2022. 6 +[3] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2 +[4] Charika De Alvis et al. A survey of deep long-tail classification advancements. arXiv:2404.15593, 2024. 5 +[5] Adrien Deliège, Anthony Cioppa, Silvio Giancola, Meisam J. Seikavandi, Jacob V. Dueholm, Kamal Nasrollahi, Bernard Ghanem, Thomas B. Moeslund, and Marc Van Droogenbroeck. Soccertnet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4508-4519, 2021. 2, 6, 8 +[6] Julien Denize, Mykola Liashuha, Jaonary Rabarisoa, Astrid Orcesi, and Romain Hérault. Comedian: Self-supervised learning and knowledge distillation for action spotting using transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 530-540, 2024. 2, 3, 6 +[7] Moritz Einfalt, Charles Dampeyrou, Dan Zecha, and Rainer Lienhart. Frame-level event detection in athletics videos with pose-based convolutional sequence networks. In Proceedings Proceedings of the 2nd International Workshop on Multimedia Content Analysis in Sports, page 42-50, New York, NY, USA, 2019. Association for Computing Machinery. 2 +[8] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 6, 7, 8 +[9] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021. 5, 7 +[10] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccertnet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018. 2 +[11] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9591-9600, 2022. 4, 5 + +[12] James Hong, Matthew Fisher, Michael Gharbi, and Kayvon Fatahalian. Video pose distillation for few-shot, fine-grained sports action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9254-9263, 2021. 6, 7 +[13] James Hong, Matthew Fisher, Michael Gharbi, and Kayvon Fatahalian. Video pose distillation for few-shot, fine-grained sports action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9254-9263, 2021. 6, 7 +[14] James Hong, Haotian Zhang, Michael Gharbi, Matthew Fisher, and Kayvon Fatahalian. Spotting temporally precise, fine-grained events in video. In European Conference on Computer Vision, pages 33-51. Springer, 2022. 2, 6, 7 +[15] Jungmin Kwon, Jeongseop Kim, Hyunseo Park, and In Kwon Choi. Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks. In Proceedings of the 38th International Conference on Machine Learning, pages 5905-5914. PMLR, 2021. 2, 5 +[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 2 +[17] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20010-20019, 2022. 2 +[18] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2, 8 +[19] Zhaoyang Liu et al. Tam: Temporal adaptive module for video recognition. In ICCV, pages 13708-13718, 2021. 3 +[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5 +[21] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[22] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7 +[23] Joao VB Soares and Avijit Shah. Action spotting using dense detection anchors revisited: Submission to the soccer-net challenge 2022. arXiv preprint arXiv:2206.07846, 2022. 5, 6 +[24] Joao VB Soares, Avijit Shah, and Topojoy Biswas. Temporally precise action spotting in soccer videos using dense detection anchors. In 2022 IEEE International Conference on Image Processing (ICIP), pages 2796-2800. IEEE, 2022. 2 +[25] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Gate-shift networks for video action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 + +[26] Matteo Tomei, Lorenzo Baraldi, Simone Calderara, Simone Bronzin, and Rita Cucchiara. Rms-net: Regression and masking for soccer event spotting. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7699-7706. IEEE, 2021. 2 +[27] Kim Hoang Tran, Phuc Vuong Do, Ngoc Quoc Ly, and Ngan Le. Unifying global and local scene entities modelling for precise action spotting. arXiv preprint arXiv:2404.09951 (Accepted for IJCNN 2024), 2024. 2, 3, 6, 7 +[28] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3 +[29] Artur Xarles, Sergio Escalera, Thomas B Moeslund, and Albert Clapés. Astra: An action spotting transformer for soccer videos. In Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports, pages 93-102, 2023. 2, 6 +[30] Artur Xarles, Sergio Escalera, Thomas B. Moeslund, and Albert Clapés. T-deed: Temporal-discriminability enhancer encoder-decoder for precise event spotting in sports videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3410-3419, 2024. 2, 6, 7 +[31] Huiran Yan et al. Emphasizing feature inter-class separability for improving highly imbalanced overlapped data classification. Knowledge-Based Systems, 276:110745, 2023. 5 +[32] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. 32nd British Machine Vision Conference, BMVC 2021, 2021. 2, 6, 7 +[33] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 4, 5 +[34] Haotian Zhang, Cristobal Sciutto, Maneesh Agrawala, and Kayvon Fatahalian. Vid2player: Controllable video sprites that behave and appear like professional tennis players. ACM Trans. Graph., 40(3), 2021. 6, 7 +[35] Xin Zhou, Le Kang, Zhiyu Cheng, Bo He, and Jingyu Xin. Feature combination meets attention: Baidu soccer embeddings and transformer based temporal detection. arXiv preprint arXiv:2106.14447, 2021. 2, 6 \ No newline at end of file diff --git a/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/images.zip b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5d0c238802341ea3028273be5c6d8b311cf976b7 --- /dev/null +++ b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5851d09ddd6eeedd01c26716b88ee020c14e906f577a7c6448ea4382281e0ef9 +size 580167 diff --git a/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/layout.json b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5e0993571ed82d2dcd848f9d179b1f7f147c08fb --- /dev/null +++ b/2025/Precise Event Spotting in Sports Videos_ Solving Long-Range Dependency and Class Imbalance/layout.json @@ -0,0 +1,7222 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 547, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 547, + 138 + ], + "type": "text", + "content": "Precise Event Spotting in Sports Videos: Solving Long-Range Dependency and Class Imbalance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "spans": [ + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "text", + "content": "Sanchayan Santra" + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "text", + "content": " Vishal Chudasama" + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "text", + "content": " Pankaj Wasnik" + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "text", + "content": " Vineeth N Balasubramanian" + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "text", + "content": "Sony Research India " + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 160, + 533, + 190 + ], + "type": "text", + "content": "Indian Institute of Technology Hyderabad" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 191, + 539, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 191, + 539, + 204 + ], + "spans": [ + { + "bbox": [ + 70, + 191, + 539, + 204 + ], + "type": "text", + "content": "{sanchayan.santra, vishal.chudasamal, pankaj.wasnik}@sony.com, vineethnb@cse.iith.ac.in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 255, + 297, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 297, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 297, + 533 + ], + "type": "text", + "content": "Precise Event Spotting (PES) aims to identify events and their class from long, untrimmed videos, particularly in sports. The main objective of PES is to detect the event at the exact moment it occurs. Existing methods mainly rely on features from a large pre-trained network, which may not be ideal for the task. Furthermore, these methods overlook the issue of imbalanced event class distribution present in the data, negatively impacting performance in challenging scenarios. This paper demonstrates that an appropriately designed network, trained end-to-end, can outperform state-of-the-art (SOTA) methods. Particularly, we propose a network with a convolutional spatial-temporal feature extractor enhanced with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) and a long-range temporal module. The ASTRM enhances the features with spatiotemporal information. Meanwhile, the long-range temporal module helps extract global context from the data by modeling long-range dependencies. To address the class imbalance issue, we introduce the Soft Instance Contrastive (Soft-IC) loss that promotes feature compactness and class separation. Extensive experiments show that the proposed method is efficient and outperforms the SOTA methods, specifically in more challenging settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 574, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 696 + ], + "type": "text", + "content": "Event spotting methods locate the occurrence of events in long, untrimmed videos. An event is typically defined by a starting and ending time, but this may not always be feasible, particularly in sports, where an event is a specific occurrence that occurs at a particular point in time, as dictated by the game's rules. For instance, a \"Goal\" in soccer is scored when the ball crosses the goal line, while a \"Boundary\" in cricket is made when the ball touches the boundary line. This process is referred to as Precise Event Spotting (PES), where PES methods aim to predict an event at the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 228, + 552, + 348 + ], + "blocks": [ + { + "bbox": [ + 316, + 228, + 552, + 348 + ], + "lines": [ + { + "bbox": [ + 316, + 228, + 552, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 552, + 348 + ], + "type": "image", + "image_path": "a145267ea273f846cb45b196f296fec5d4b7a2097d67da6ac81f028468589c81.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 351, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 396 + ], + "type": "text", + "content": "Figure 1. Temporal dependency of \"Shots on target\" event: The event is marked when the shot is taken, but whether the shot is at the target can only be known by looking at future frames. Here the shot is on target because the ball is saved by the goal keeper." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 407, + 553, + 529 + ], + "blocks": [ + { + "bbox": [ + 317, + 407, + 553, + 529 + ], + "lines": [ + { + "bbox": [ + 317, + 407, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 407, + 553, + 529 + ], + "type": "image", + "image_path": "4430d79e5bcdf58e0fabb04ee2b14382a55280ac93b69dbe00eacfcf07b0e97b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 536, + 556, + 581 + ], + "lines": [ + { + "bbox": [ + 313, + 536, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 556, + 581 + ], + "type": "text", + "content": "Figure 2. Per class score analysis on a few classes of SoccerNet V2 dataset in tight setting. The proposed method outperforms state-of-the-art (SOTA) methods especially for classes with less number of samples." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 604, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 664 + ], + "type": "text", + "content": "exact frame it occurs (e.g., goal, boundary). This is crucial for generating game summaries, finding specific events, and editing lengthy matches. However, it poses significant challenges, requiring a deep understanding of the game and its key events." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Existing PES methods mainly face two challenges. First is the long-range temporal dependency that arises due to the nature of events. To predict the event frames, observing just the nearby frames does not suffice; several other potentially" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 146, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 146, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 146, + 713 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3163" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "distant frames need to be observed to predict the event correctly. For example, \"Shots on target\" and \"Shots off target\" events of Soccer games are marked when the shot is taken. Still, the actual class is decided based on whether the ball went toward the goal or away from it (see Figure 1). This can be only detected if long-range dependencies are captured correctly. Second, there is an inherent imbalance in the number of samples among the classes (see Figure 2). Some events naturally occur more frequently than others. For example, in a soccer game, a \"Ball out of play\" is likely to occur more than a \"Yellow card\"; a \"Red-card\" and \"Yellow " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " Red card\" are expected to occur even less often. Therefore, the same trend can be observed in datasets created using broadcast videos. Due to these issues, existing event spotting methods [6, 14] perform poorly in underrepresented classes such as \"Red card\" & \"Yellow " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " Red card.\" This can be visualized in Figure 2. Handling these issues and accurately performing PES is an active area of research." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 295, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 505 + ], + "type": "text", + "content": "To address the long-range temporal dependency issue, recent SOTA methods [24, 29] prefer to use the feature extracted using the pre-trained network of [35]. These methods also use transformer-based temporal modules due to their better sequence modeling capability in a global context. However, as stated in [18, 32], transformers also encounter issues with long videos. Self-attention struggles to learn appropriate weights for long videos due to the increase in the size of the weight matrix with the number of frames being processed. On the other hand, E2ESpot [14] proposed an end-to-end trainable feature extractor and a GRU-based module for temporal modeling. Recently proposed T-DEED [30] improves the performance of E2ESpot via temporal-discriminability enhancing encoder-decoder (T-DEED) module. In our proposed method, we handle this issue by enhancing the extracted features with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) and GRU-based temporal module." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "content": "The class imbalance issue has not received sufficient attention in the literature on PES, resulting in lower performance on real-world datasets. The recent method ASTRA [29] is one of the few attempts to address this problem through a modified \"mixup\" augmentation. Although standard machine learning approaches, such as focal loss [16], are available, they are not suitable for every task, as this paper demonstrates. To address this challenge, we introduce a novel loss function called Soft Instance Contrastive (Soft-IC) loss. This loss function aims to mitigate class imbalance by promoting compact feature representations and improving class separability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "The primary goal of this paper is to address these issues and improve the accuracy of PES. To this end, we propose an end-to-end trainable network with a CNN-based spatiotemporal feature extractor and a GRU-based long-term temporal module. The spatio-temporal feature extractor in-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "cludes our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM) to augment the features with spatial and temporal information. We chose to train the model end-to-end because it has been shown to perform better than a pre-trained frozen feature extractor [17, 26]. We have also employed the Adaptive Sharpness-Aware Minimization (ASAM) [15] to enable the trained model to perform well on unseen data. Through experiments, we demonstrate that the proposed framework obtains superior performance compared to SOTA methods, especially in tight settings where long-range dependency and class imbalance issues are prominent (see Figure 2). Our key contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 228, + 553, + 334 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 228, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 553, + 275 + ], + "type": "text", + "content": "- We propose an end-to-end framework using a CNN-based spatio-temporal feature extractor and GRU-based temporal network to achieve precise event spotting while addressing long-range dependency and class imbalance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 275, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 553, + 299 + ], + "type": "text", + "content": "- We introduce ASTRM, which enhances the features with spatial and temporal information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 334 + ], + "type": "text", + "content": "- We also introduce Soft-IC loss to enforce compact features and improve class separability, addressing the class imbalance issue." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 346, + 405, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 346, + 405, + 358 + ], + "spans": [ + { + "bbox": [ + 314, + 346, + 405, + 358 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 366, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 553, + 569 + ], + "type": "text", + "content": "The concept of precise sports event spotting first appeared in Einfalt et al. [7], where they considered temporally precise detection of various events related to long and triple jump. Later, SoccerNet [10] proposed a large-scale dataset of soccer videos to identify specific events or actions at exact timestamps. The work by [14] expanded this to precise event spotting (PES), introducing significant differences in the necessary precision of predictions. PES is constrained to the tolerance of a few frames only. Recently, several PES methods [6, 14, 24, 27, 29, 30, 35] have been proposed that focus on various sports videos. E2E-Spot [14] employed a CNN-based backbone with a Gated Shift Module (GSM) [25] for temporal modeling and a GRU [3] for producing per-frame prediction. Even though it is effective on the Tennis and Figure Skating datasets, it struggles on more challenging SoccerNet V2 dataset [5] due to its simplistic temporal information extraction scheme using GSM." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "Building upon E2E-Spot [14], T-DEED [30] proposed improvements to address issues arising from tight tolerances in fast-paced sporting events and achieved SOTA performance on the Figure Skating and Figure Diving datasets. The method of [35] revamped the feature extraction using a group of five networks pre-trained on Kinetics-400 and fine-tuned on SoccerNet V2. Several other methods use the features extracted by this method instead of raw video frames. For instance, Soares et al. [24] utilized these features and developed a U-Net- and transformer-based trunk for making predictions. ASTRA [29] utilizes the same features but incorporates audio features for complementary in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3164" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "formation. In contrast, COMEDIAN [6] operates in 3 stages to train their network on raw frames: 1) training only the spatial part; 2) training the spatial and temporal parts using knowledge distillation; 3) final tuning to optimize the classification scores. Recently, UGLF [27] proposed a unifying global-local (UGL) module to disentangle the scene content into global features and locally relevant scene entity features in conjunction with a Vision-Language model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 295, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 347 + ], + "type": "text", + "content": "The above-mentioned methods mainly addressed the long-range temporal dependency issue by utilizing pretrained features and transformers. They have not adapted their feature extractors to capture spatial and temporal information effectively. Additionally, these methods did not specifically deal with the class imbalance, which led to poor performance in underrepresented classes. In our framework, we propose ASTRM, which helps the feature extractor to augment the features with relevant spatial and temporal information. Furthermore, we introduce Soft-IC loss to address the class imbalance issue and employ the ASAM to train the proposed framework properly. With these advancements, the proposed framework outperforms existing event-spotting methods on various sports datasets, particularly in settings with tighter tolerances." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 357, + 189, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 189, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 189, + 370 + ], + "type": "text", + "content": "3. Proposed Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 377, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 413 + ], + "type": "text", + "content": "This section presents our methodology in detail, starting with the problem statement and then presenting the details of the proposed modules as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 419, + 168, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 168, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 168, + 431 + ], + "type": "text", + "content": "3.1. Problem Statement" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "content": "The task of precise event spotting is defined as follows: Given a video with " + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1, \\mathbf{x}_2, \\ldots, \\mathbf{x}_N" + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "content": " and a set of possible events " + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\{c_1, \\ldots, c_K\\}" + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "content": ", the goal is to predict frames where an event occurs along with its class: " + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "inline_equation", + "content": "(t, \\hat{y}_t) \\in \\mathbb{N} \\times \\{c_1, \\ldots, c_K\\}" + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "content": ". A prediction is deemed correct if it falls within " + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "content": " frames of the actual occurrence and accurately identifies the event class. Since it is not possible to process the entire video at once, the processing is done in chunks of a predefined size. Each chunk (or clip), represented as " + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{C \\times T_s \\times H \\times W}" + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "content": ", consists of " + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "inline_equation", + "content": "T_s" + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "content": " frames. We compute predictions for each frame, represented as " + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{R}^{T_s \\times K}" + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "content": " is the number of classes. As the video is processed in segments, the predictions from each chunk of " + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "inline_equation", + "content": "T_s" + }, + { + "bbox": [ + 55, + 437, + 295, + 617 + ], + "type": "text", + "content": " frames are aggregated and refined with a non-maximal suppression to get the final sparse prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 623, + 140, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 140, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 140, + 635 + ], + "type": "text", + "content": "3.2. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": "Figure 3 shows the network design of our proposed framework, which takes a set of input frames and classifies whether each frame contains an event or not. The framework comprises three main components: the spatiotemporal feature extractor, the temporal block, and the classifier. In the spatio-temporal feature extractor, we" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "content": "have selected the RegNetY [21] as the feature extractor due to its efficiency, surpassing commonly used networks like ResNet. This network is further enhanced with our proposed Adaptive Spatio-Temporal Refinement Module (ASTRM). This module is integrated into each bottleneck block, refining the features by incorporating both spatial and temporal information. Both of this information is necessary to correctly identify events, as validated by the experiments. The refined features are then passed to the temporal block, which enriches them with long-term temporal information. This long-term information is crucial because, in many situations, information from temporally distant frames is essential to accurately classify a frame. The spatio-temporal feature extractor alone can not accomplish this task. Finally, the feature maps are forwarded to the classifier, which is a simple linear layer that outputs the classification scores. Furthermore, we introduce the Soft Instance Contrastive (Soft-IC) loss to effectively address the imbalance among the class samples. This loss function operates by compacting the features and enhancing their separability. The key modules of the proposed framework are detailed in subsequent subsections." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 350, + 550, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 550, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 550, + 363 + ], + "type": "text", + "content": "3.2.1 Adaptive Spatio-Temporal Refinement Module:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 369, + 553, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 553, + 477 + ], + "type": "text", + "content": "This module enhances the extracted features by incorporating both spatial and temporal information. It comprises three blocks: Local Spatial, Local Temporal, and Global Temporal. We include the Local Spatial block because local spatial features also offer important cues about temporal information. Relying solely on temporal information can overlook these valuable insights. The spatial component is inspired by [28], while the temporal aspect is adapted from [19]. The refinement is done as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 487, + 553, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 487, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 320, + 487, + 553, + 501 + ], + "type": "interline_equation", + "content": "\\Psi (\\mathbf {x}) = \\left(\\left(\\mathbf {x} \\odot \\left(1 + \\mathcal {F} _ {s} (\\mathbf {x})\\right)\\right) \\odot \\left(1 + \\mathcal {F} _ {t} (\\mathbf {x})\\right)\\right) * \\mathcal {G} _ {t} (\\mathbf {x}) \\tag {1}", + "image_path": "255973bd0549beb58eba9be589293c447ea3714ed612a54c07335f8b2b0229cb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "text", + "content": " is the input feature to be enhanced. " + }, + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s, \\mathcal{F}_t," + }, + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "text", + "content": " denote the operation corresponding to the local spatial, local temporal, and global temporal block, respectively. " + }, + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "text", + "content": " denotes Hadamard Product, while " + }, + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 313, + 510, + 554, + 688 + ], + "type": "text", + "content": " denotes convolution operation. To elaborate, the input feature is weighted with information from the local spatial block and a residual connection, resulting in features enhanced with local spatial information. These enhanced features are further refined with local temporal information. Finally, spatially and temporally enhanced features are then convolved with the kernel obtained from the global temporal block to make local enhancements using \"global\" information. This dynamic kernel approach simplifies the network while maintaining effectiveness. Please refer to Figure 3 for a visual representation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "In the Local Spatial block, the input features are pooled using Max (i.e., " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "f_{MaxPool}" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": ") and Average (i.e., " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "f_{AvgPool}" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3165" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 76, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 67, + 76, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 67, + 76, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 67, + 76, + 545, + 304 + ], + "type": "image", + "image_path": "e46dd065db54d0faa2035b4b74c2249192712b2832c5b8d92f49c1880a13fb4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 311, + 555, + 355 + ], + "lines": [ + { + "bbox": [ + 54, + 311, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 555, + 355 + ], + "type": "text", + "content": "Figure 3. Network design of the proposed framework. The framework composed of a spatio-temporal feature extractor and a temporal block for capturing long-range dependency before the classifier. In each bottleneck block we add ASTRM after the first conv. ASTRM further enhances the features with local spatial, local temporal and global temporal information. The network is trained with Soft-IC loss in addition to the classification loss to handle class imbalance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 376, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 472 + ], + "type": "text", + "content": "pooling operations along the channel dimension. The pooled features are then concatenated and convolved with a " + }, + { + "bbox": [ + 55, + 376, + 296, + 472 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 55, + 376, + 296, + 472 + ], + "type": "text", + "content": " filter (i.e., " + }, + { + "bbox": [ + 55, + 376, + 296, + 472 + ], + "type": "inline_equation", + "content": "f_{conv}^{7 \\times 7}" + }, + { + "bbox": [ + 55, + 376, + 296, + 472 + ], + "type": "text", + "content": "). The pooling operation on the channel axis helps emphasize informative regions. The subsequent convolution operation filters out non-informative portions from the features. Finally, the output is passed through a Sigmoid function (i.e., " + }, + { + "bbox": [ + 55, + 376, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 55, + 376, + 296, + 472 + ], + "type": "text", + "content": ") to obtain the \"spatial\" feature weights (i.e., " + }, + { + "bbox": [ + 55, + 376, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 376, + 296, + 472 + ], + "type": "text", + "content": ") as follow:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 475, + 295, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 475, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 66, + 475, + 295, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {s} (\\mathbf {x}) = \\sigma \\left(f _ {c o n v} ^ {7 \\times 7} \\left(\\left[ f _ {A v g P o o l} (\\mathbf {x}); f _ {M a x P o o l} (\\mathbf {x}) \\right]\\right)\\right). \\tag {2}", + "image_path": "5e8bfd3d549e83cdec8159d2765fef4005e3f8b32c27fa0a5ee4919d00e2f8eb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 495, + 296, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 577 + ], + "type": "text", + "content": "In the Local Temporal block, we use 3D convolution operations, namely " + }, + { + "bbox": [ + 55, + 495, + 296, + 577 + ], + "type": "inline_equation", + "content": "f_{conv}^{3 \\times 1 \\times 1}" + }, + { + "bbox": [ + 55, + 495, + 296, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 495, + 296, + 577 + ], + "type": "inline_equation", + "content": "f_{conv}^{1 \\times 1 \\times 1}" + }, + { + "bbox": [ + 55, + 495, + 296, + 577 + ], + "type": "text", + "content": ", followed by the Sigmoid function, " + }, + { + "bbox": [ + 55, + 495, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 55, + 495, + 296, + 577 + ], + "type": "text", + "content": ", with a focus on the temporal dimension. Spatially pooled features are not used in this block because they do not account for the different temporal changes at various spatial locations. The locally temporal enhanced features are defined as below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 582, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 582, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 68, + 582, + 295, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {t} (\\mathbf {x}) = \\sigma \\left(f _ {\\text {c o n v}} ^ {1 \\times 1 \\times 1} \\left(\\mathrm {B N} \\left(\\operatorname {R e L U} \\left(f _ {\\text {c o n v}} ^ {3 \\times 1 \\times 1} (\\mathbf {x})\\right)\\right)\\right)\\right). \\tag {3}", + "image_path": "f1a1803f0406705bd084296c6b0ad139f857dd7dd6526f602de1ade42eeb37d1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 601, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 696 + ], + "type": "text", + "content": "In the Global Temporal block, we adaptively create convolution kernels to improve the feature maps by considering temporal information across frames. First, the features are pooled in the spatial dimension using a global average pooling operation (i.e., " + }, + { + "bbox": [ + 55, + 601, + 296, + 696 + ], + "type": "inline_equation", + "content": "f_{GAP}" + }, + { + "bbox": [ + 55, + 601, + 296, + 696 + ], + "type": "text", + "content": "), and then fully connected layers (i.e., " + }, + { + "bbox": [ + 55, + 601, + 296, + 696 + ], + "type": "inline_equation", + "content": "f_{FC}" + }, + { + "bbox": [ + 55, + 601, + 296, + 696 + ], + "type": "text", + "content": ") are used to compute the kernel with a global view of the data with a softmax (" + }, + { + "bbox": [ + 55, + 601, + 296, + 696 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 601, + 296, + 696 + ], + "type": "text", + "content": ") at the end. Mathematically, the kernel is formulated as below:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 99, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {t} (\\mathbf {x}) = \\mathcal {S} \\left(f _ {F C} \\left(f _ {F C} \\left(f _ {G A P} (\\mathbf {x})\\right)\\right)\\right). \\tag {4}", + "image_path": "07e3451ad267fcc0abf013197477ea307a30472451f2ff417520cddd7f95eb94.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 376, + 419, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 419, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 419, + 388 + ], + "type": "text", + "content": "3.2.2 Temporal Block:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 612 + ], + "type": "text", + "content": "The spatio-temporal feature extraction stage computes features containing local spatial, local temporal, and location-independent temporal information. Since we are interested in finding the temporal location of the events, we aggregate the features in the spatial dimension with average pooling before extracting long-range temporal information. We use temporal block because the feature extractor alone is not capable enough to capture the required long-range dependency, although it may help to some extent. As motivated in the introduction section, we need to capture long-range temporal information to accurately predict events with precision. For this task, we selected Bidirectional GRU (Bi-GRU) as the temporal block because it effectively models sequences without overfitting the input data. The Bi-GRU is capable of quickly adapting to the data while maintaining its generalizability. Further experiments confirm this, demonstrating that the Bi-GRU performs well compared to more complex temporal blocks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 621, + 517, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 517, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 517, + 633 + ], + "type": "text", + "content": "3.2.3 Soft Instance Contrastive (Soft-IC) loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "To address the class imbalance issue, we introduce a novel Soft-IC loss, which aims to make the features more compact and increase their separability. It is inspired by the Instance Contrastive loss (IC loss) [11] and is designed to work with mixup augmentation [33]. IC loss enhances the separation among the features by promoting its compactness. In case" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3166" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "of class imbalance, features may deviate from their respective class center and potentially overlap with features from other classes. This can hinder the learning of a decision boundary and negatively impact performance [4]. By compacting the features, we indirectly address the challenges posed by class imbalance, a strategy supported by research in the field [31]. For a batch of " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " features computed from the input, the IC Loss is calculated as described in [11]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 130, + 179, + 295, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 179, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 130, + 179, + 295, + 212 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} (\\mathbf {z} _ {i}), \\tag {5}", + "image_path": "f36b9ea831d09d90727c71a756dc4034a35b4554c71549da395a895f744329e2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 238, + 296, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 296, + 280 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\mathbf {z} _ {i}\\right) = \\frac {1}{\\left| M \\left(\\mathbf {c} _ {i}\\right) \\right|} \\sum_ {\\mathbf {z} _ {j} \\in M \\left(\\mathbf {c} _ {i}\\right)} \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {j} / \\tau\\right)}{\\sum_ {\\mathbf {z} _ {\\mathrm {k}} \\in \\mathbf {A} \\left(\\mathbf {c} _ {\\mathrm {i}}\\right)} \\exp \\left(\\mathbf {z} _ {i} \\cdot \\mathbf {z} _ {k} / \\tau\\right)}, \\tag {6}", + "image_path": "ec73f3b3b64f2a5be9bfb66b650925ca81cc420f59687eeb6fcddf2904684cf3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": " is the label of " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": "-th sample, " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": " is a temperature hyper-parameter, " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "M(\\mathbf{c}_i)" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": " denotes the memory bank of class " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "inline_equation", + "content": "A(\\mathbf{c}_i) = M\\backslash M(\\mathbf{c}_i)" + }, + { + "bbox": [ + 55, + 280, + 296, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 365 + ], + "type": "text", + "content": "During training, we use mixup augmentation [33] to regularize the model and improve its generalization. In this, two samples (and their labels) are combined convexly to create a new sample (and its label) as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 377, + 295, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 377, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 118, + 377, + 295, + 390 + ], + "type": "interline_equation", + "content": "\\tilde {x} = \\lambda x _ {i} + (1 - \\lambda) x _ {j}, \\tag {7}", + "image_path": "482002f5e3f0285b7b23f769adad87df75c6ec9ce50344c33b9752664890a6c1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 392, + 295, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 392, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 122, + 392, + 295, + 405 + ], + "type": "interline_equation", + "content": "\\tilde {y} = \\lambda y _ {i} + (1 - \\lambda) y _ {j}, \\tag {8}", + "image_path": "7450cd6b41549f9eacc4a6e2ea5d0638da69b57ad02e67597f85fa7482a437fe.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": " are two samples from the training data, and " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": " are their corresponding one-hot labels. The parameter " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": " takes value in the range [0, 1]. Consequently, the label " + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "inline_equation", + "content": "(\\tilde{y})" + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": " will not be in one-hot format; instead, it will reflect weighted contributions from each class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 479, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 526 + ], + "type": "text", + "content": "IC Loss assumes that each sample has a single label. However, this assumption does not hold true when using mixup augmentation. To account for weight associated with each class in the label, we define Soft-IC loss as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 536, + 296, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 536, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 91, + 536, + 296, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S I C} = \\frac {\\sum_ {\\left(\\mathbf {z} _ {i} , w _ {i j}\\right) , w _ {i j} \\neq 0} \\mathcal {L} _ {S I C} \\left(\\mathbf {z} _ {i} , w _ {i j}\\right)}{\\sum_ {\\left(\\mathbf {z} _ {i} , w _ {i j}\\right) , w _ {i j} \\neq 0} 1}, \\tag {9}", + "image_path": "58687d3d906f4a30b5a812ad34b4c6ab16ebb908102e646f4f20fb2d11449c52.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 594, + 181, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 594, + 181, + 620 + ], + "spans": [ + { + "bbox": [ + 66, + 594, + 181, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S I C} (\\mathbf {z} _ {i}, \\omega) = \\frac {1}{\\omega | M (\\mathbf {c} _ {i}) |} \\times", + "image_path": "b9db07b34bbbb9fbd019df6df8cefeab5935639b5477dd00b26a2f41fd1abc14.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 72, + 622, + 296, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 622, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 72, + 622, + 296, + 666 + ], + "type": "interline_equation", + "content": "\\sum_ {(\\mathbf {z} _ {j}, w _ {j}) \\in M (\\mathbf {c} _ {i})} \\log \\frac {\\exp \\left(\\mathbf {z} _ {i} \\cdot w _ {j} \\mathbf {z} _ {j} / \\tau\\right)}{\\sum_ {(\\mathbf {z} _ {\\mathbf {k}} , w _ {k}) \\in A (\\mathbf {c} _ {i})} \\exp \\left(\\mathbf {z} _ {i} \\cdot w _ {k} \\mathbf {z} _ {k} / \\tau\\right)}. \\tag {10}", + "image_path": "7cd720d66bd572b0d470655cf13457b05f9b5d0822403db091b1c0e43e57c4a5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "w_{ij}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is the weight associated with the label of class " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " for sample " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". Additionally, in the memory bank " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "M(\\mathbf{c}_i)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", we store the corresponding class weights along with features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 72, + 421, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 421, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 421, + 85 + ], + "type": "text", + "content": "3.3. Training Strategy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 90, + 555, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 555, + 149 + ], + "type": "text", + "content": "The proposed framework is trained end-to-end using a combination of two loss functions: Binary Cross Entropy loss " + }, + { + "bbox": [ + 313, + 90, + 555, + 149 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{BCE})" + }, + { + "bbox": [ + 313, + 90, + 555, + 149 + ], + "type": "text", + "content": " as the standard classification loss and the proposed Soft-IC loss " + }, + { + "bbox": [ + 313, + 90, + 555, + 149 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{SIC})" + }, + { + "bbox": [ + 313, + 90, + 555, + 149 + ], + "type": "text", + "content": ". The overall loss function " + }, + { + "bbox": [ + 313, + 90, + 555, + 149 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{final})" + }, + { + "bbox": [ + 313, + 90, + 555, + 149 + ], + "type": "text", + "content": " can be expressed as below:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 159, + 555, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 159, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 361, + 159, + 555, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i n a l}} (\\cdot) = \\mathcal {L} _ {B C E} + \\lambda_ {S I C} \\cdot \\mathcal {L} _ {S I C}. \\tag {11}", + "image_path": "e627a5b79974e3e41cb376663218c044049dae70c4492bec4e9094f43aa67bf8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "inline_equation", + "content": "\\lambda_{SIC}" + }, + { + "bbox": [ + 313, + 179, + 554, + 239 + ], + "type": "text", + "content": " denotes the weighting constant that adjusts the contributions of different loss functions. With the total loss, the proposed network is trained using the AdamW [20] optimizer with Adaptive Sharpness-Aware Minimization (ASAM) [15]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 239, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 443 + ], + "type": "text", + "content": "Utilizing advanced optimization techniques can significantly impact the performance of the trained model, yet not much focus is given to this aspect. In [23], the authors utilized the Sharpness-Aware Minimization (SAM) [9] with the Adam optimizer, which significantly impacted the final scores. SAM is a technique that considers the \"sharpness\" of the loss manifold in each gradient descent step, resulting in improved generalization. In SAM, the sharpness is computed from a fixed-size neighborhood of the current location in the loss manifold. However, the fixed-size neighborhood is not always ideal for sharpness computation. Adaptive SAM (ASAM) [15] proposed a size-varying neighborhood based on the scale to improve upon it. Taking a cue from this, we propose to utilize ASAM with AdamW optimizer to optimize our proposed framework. Through ablation analysis, we validate the effectiveness of ASAM in boosting the performance of the proposed framework." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 453, + 446, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 446, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 446, + 466 + ], + "type": "text", + "content": "4. Experimental Analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 472, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 509 + ], + "type": "text", + "content": "This section provides detailed information about the implementations and datasets, followed by the evaluation and ablation study." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 515, + 447, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 447, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 447, + 528 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "content": "We choose a pre-trained RegNetY200 initialized with ImageNet weights as the backbone in the spatio-temporal feature extractor. The model is trained using ASAM [15] with AdamW [20] as the base optimizer, setting " + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\rho = 2" + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "content": ". For training, a cosine annealing learning rate scheduler is employed with a warm-up of 3 epochs and an initial learning rate of 1e-5 during the warm-up. For the Soft-IC loss, a feature dimension of 128 and a memory bank size of 256 is used, with a loss weight (i.e., " + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{SIC}" + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "content": ") of 0.001. To enhance the training data, mixup [33] augmentation is applied with a mixing factor of " + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "content": ". Additionally, Gaussian blur is applied with a kernel size of 5 and a probability of 0.25. Brightness, contrast, and saturation adjustments are made with a factor ranging from 0.7 to 1.2, at a probability of 0.5. The proposed method is implemented in PyTorch and trained" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3167" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "on 4 A100 GPUs with a batch size of 2 on each GPU. At a time, 128 frames are processed, constituting a single sample (clip). During training, clips are uniformly sampled from videos, while during evaluation, a sliding window with half overlap is utilized to extract the clips." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 192 + ], + "type": "text", + "content": "To validate the efficacy of the proposed method, we conducted experiments using event spotting datasets outlined in [14] such as Tennis [34], Figure Skating (FS) [13], Fine Gym and the SoccerNet V2 [5] action spotting dataset. These datasets are described in the supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 200, + 220, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 220, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 220, + 213 + ], + "type": "text", + "content": "4.2. Evaluation Metric Description" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "text", + "content": "Like previous works, we calculate the average precision (AP) within a specified tolerance range, denoted as " + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "text", + "content": ". The AP is then calculated for each event class to calculate the mean average precision (mAP) with the specified " + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "text", + "content": " value as an evaluation metric. The " + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "text", + "content": " is defined based on the respective datasets. For the Tennis, Figure Skating, and FineGym datasets, which have very strict evaluation criteria, we report the results with " + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "inline_equation", + "content": "\\delta = 0,1" + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "text", + "content": ", and 2. The SoccerNet V2 dataset has relaxed evaluation criteria and defines two metrics: tight-mAP and loose-mAP, with " + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "inline_equation", + "content": "\\delta = 1 - 4" + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "text", + "content": " seconds and " + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "inline_equation", + "content": "\\delta = 5 - 60" + }, + { + "bbox": [ + 55, + 218, + 295, + 385 + ], + "type": "text", + "content": " seconds, respectively. Although the metrics are less strict compared to others, the SoccerNet V2 dataset is more challenging; having a more stringent metric would make the task even harder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 393, + 160, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 160, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 160, + 405 + ], + "type": "text", + "content": "4.3. Baseline Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 411, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 532 + ], + "type": "text", + "content": "For the SoccerNet V2 [5] dataset, we compare existing SOTA methods proposed in conjunction with the SoccerNet Action Spotting challenge. Some of the results are taken from the 2021 challenge leaderboard, including Baidu (TF) [35], and SpotFormer [2]. We have also compared the results with CNN-based SOTA methods like E2ESpot [14], T-DEED [30], UGLF [27] and transformer-base SOTA methods like Spivak [23], ASTRA [29], and COMEDIAN [6]. The results are generated using the checkpoints provided by the respective authors whenever possible." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 651 + ], + "type": "text", + "content": "For the Tennis, Figure Skating, and FineGym dataset, we compare our method with the baseline method E2E-Spot [14]. We have also included the results of some best-performing pre-trained and fine-tuned variants as defined in [14]. These include TSP [1] + MS-TCN [8], TSP [1] + AS-Former [32], VC-Spot [14], and 2D-VDP [12] + MS-TCN [8]. We also compare the result with the recent state-of-the-art T-DEED [30] method. Unless otherwise stated, the results are mainly generated from the checkpoint provided by the authors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 658, + 279, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 279, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 279, + 672 + ], + "type": "text", + "content": "4.4. Result Analysis on Various Sports Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "content": "This section compares the results obtained by our proposed method with existing SOTA methods on SoccerNet V2, Tennis, Figure Skating & FineGym datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 70, + 553, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 159 + ], + "type": "text", + "content": "Table 1. Result Comparison on SoccerNet-V2 datasets in terms of mAP tight and loose settings. Here, the best-performing measures are highlighted with ** bold font and second best-performing is highlighted with ** underline font. * result obtained from the challenge leaderboard of 2021. ** indicate that the GFLOPs value is calculated from the temporal network only without feature extractors. † indicate result generated from the author provided checkpoint. ‡ indicate results obtained from official repo." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 167, + 553, + 277 + ], + "blocks": [ + { + "bbox": [ + 316, + 167, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 316, + 167, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 167, + 553, + 277 + ], + "type": "table", + "html": "
ModelParams (M)GFLOPsTightLoose
Baidu (TF) [35]——47.05*73.77*
SpotFormer [2]——60.9076.10
E2E-Spot (RegNet-Y 200MF) [14]4.4639.6161.1973.25
E2E-Spot (RegNet-Y 800MF) [14]12.64151.4061.8274.05
Spivak [23]17.46461.8965.1078.50
ASTRA [29]44.338.83**66.6378.14
T-DEED (RegNet-Y 200MF) [30]16.3621.9639.43‡52.93‡
T-DEED (RegNet-Y 800MF) [30]46.2260.2521.57‡30.49‡
UGLF [27]4.4639.6163.5173.98‡
COMEDIAN (ViSwin-T) [6]70.12222.7671.33†77.07†
Proposed6.4660.2573.7479.11
", + "image_path": "5ae2e9bd6b28f08d050b1267e286851c453af8bc765888ae303fbdf660a691b2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 304, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 518 + ], + "type": "text", + "content": "SoccerNet V2: Table 1 compares the existing SOTA methods and the proposed approach on the SoccerNet V2 testing dataset. It is evident from the results that the proposed method has achieved notably strong performance in both tight and loose settings. The scores demonstrate significant improvement compared to both RegNet-Y 200MF and 800MF of E2E-Spot despite utilizing a similar framework. Additionally, there is a substantial improvement in the tight setting compared to the Baidu, Spivak, and ASTRA methods. These methods employ a pre-trained frozen larger backbone for feature extraction and transformer-based long-term temporal modules. COMEDIAN model utilizes a transformer network to achieve SOTA results in tight settings. However, the proposed method outperforms these methods without requiring a larger network. Notably, the proposed model achieves a " + }, + { + "bbox": [ + 313, + 304, + 555, + 518 + ], + "type": "inline_equation", + "content": "2.41\\%" + }, + { + "bbox": [ + 313, + 304, + 555, + 518 + ], + "type": "text", + "content": " improvement in tight settings than the COMEDIAN model while having significantly fewer parameters and FLOPs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "Tennis, Figure Skating & FineGym: Table 2 shows the results for Tennis, Figure Skating settings, i.e., FS-Comp and FS-Perf, and FineGym with " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\delta = 0, 1" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " and 2. In this case, the proposed method also outperforms existing methods with a noticeable improvement. For instance, in case of " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\delta = 0" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": ", the proposed method achieves " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "15.67\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "8.64\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "8.49\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "12.55\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " improvements over the best-performing E2ESpot (RegNet-Y 200MF) and T-DEED (RegNet-Y 200MF) methods in Tennis, FS-Comp & FS-Perf and FineGym datasets, respectively. Similarly, for " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\delta = 1" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " case, the proposed method outperforms the previous best-performing T-DEED variants RegNet-Y 200MF and RegNet-Y 800MF by " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "1.52\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " improvement on FS-Comp & FS-Perf datasets, respectively. Additionally, we present the per-class score comparison on Tennis, Figure Skating, & FineGym datasets in Figure 4. Here, it can be seen that the perfor" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3168" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 103, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 2. Result Comparison on Tennis, FS-Comp, FS-Perf and FinGym datasets in terms of mAP for different delta values. Here, the best-performing measures are highlighted with bold font and second best-performing is highlighted with underline font. * indicate result generated using provided checkpoints." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 103, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 58, + 103, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 553, + 235 + ], + "type": "table", + "html": "
ModelTennisFS-CompFS-PerfFineGym
δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2δ = 0δ = 1δ = 2
TSP [1] + MS-TCN [8]—90.9095.10—72.4087.80—76.8089.90—40.5058.50
TSP [1] + ASFormer [32]—89.8095.50—77.7094.10—80.2094.50—38.8057.60
2D-VDP [12] + MS-TCN [8]————83.5096.20—85.2096.40———
VC-Spot (RegNet-Y 200MF) [14]—92.4096.00—61.8075.50—56.2075.30—18.7028.60
E2E-Spot (RegNet-Y 200MF) [14]45.34*96.1097.7022.08*81.0093.5022.33*85.1095.7010.01*47.9065.20
E2E-Spot (RegNet-Y 800MF) [14]44.64*95.16*97.72*23.95*83.4094.9022.17*83.3096.0010.15*50.11*67.23*
T-DEED (RegNet-Y 200MF) [30]43.4996.1797.9528.92*85.1591.7031.87*86.7996.0513.54*53.73*66.43*
T-DEED (RegNet-Y 800MF) [30]44.0896.4297.0527.59*84.7792.8631.25*88.1795.8712.42*54.00*66.70*
UGLF [27]——————————50.2067.80
Proposed61.0196.2197.7537.5486.6795.5340.3689.5797.3815.2452.3166.57
", + "image_path": "fae9d28f86b248af4ce89ff4cdeda5cecb94b1bdb32f77d2eda636bc20e5c498.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 55, + 248, + 282, + 350 + ], + "blocks": [ + { + "bbox": [ + 55, + 248, + 282, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 248, + 282, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 282, + 350 + ], + "type": "image", + "image_path": "f32d2124f031badae1e4917ca47ac56f1efc47c2ab986de5f8598f64c566c5b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 350, + 202, + 360 + ], + "lines": [ + { + "bbox": [ + 136, + 350, + 202, + 360 + ], + "spans": [ + { + "bbox": [ + 136, + 350, + 202, + 360 + ], + "type": "text", + "content": "(a) Tennis dataset [34]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 359, + 282, + 438 + ], + "blocks": [ + { + "bbox": [ + 58, + 359, + 282, + 438 + ], + "lines": [ + { + "bbox": [ + 58, + 359, + 282, + 438 + ], + "spans": [ + { + "bbox": [ + 58, + 359, + 282, + 438 + ], + "type": "image", + "image_path": "0cb9a6e62b8e5d77982a536536f8afed6881a95c08390c0822a854efaff96067.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 440, + 206, + 449 + ], + "lines": [ + { + "bbox": [ + 132, + 440, + 206, + 449 + ], + "spans": [ + { + "bbox": [ + 132, + 440, + 206, + 449 + ], + "type": "text", + "content": "(c) FS-Comp dataset [13]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 331, + 244, + 551, + 347 + ], + "blocks": [ + { + "bbox": [ + 331, + 244, + 551, + 347 + ], + "lines": [ + { + "bbox": [ + 331, + 244, + 551, + 347 + ], + "spans": [ + { + "bbox": [ + 331, + 244, + 551, + 347 + ], + "type": "image", + "image_path": "97c5c9fe5513194355556825f9706982ba56a9a57f5c452e3e90c10fd0f96981.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 331, + 361, + 553, + 438 + ], + "blocks": [ + { + "bbox": [ + 404, + 350, + 479, + 361 + ], + "lines": [ + { + "bbox": [ + 404, + 350, + 479, + 361 + ], + "spans": [ + { + "bbox": [ + 404, + 350, + 479, + 361 + ], + "type": "text", + "content": "(b) FineGym dataset [22]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 361, + 553, + 438 + ], + "lines": [ + { + "bbox": [ + 331, + 361, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 331, + 361, + 553, + 438 + ], + "type": "image", + "image_path": "da0e9b2119e7b616cc6ab177c3c9aff1a62cbeefd10a763a73c70549640a1e95.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 440, + 476, + 449 + ], + "lines": [ + { + "bbox": [ + 406, + 440, + 476, + 449 + ], + "spans": [ + { + "bbox": [ + 406, + 440, + 476, + 449 + ], + "type": "text", + "content": "(d) FS-Perf dataset [13]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 491, + 297, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 297, + 529 + ], + "type": "text", + "content": "mance of the proposed method is not just good for a single class; it has performed consistently well across all the classes by a significant margin." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 537, + 156, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 156, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 156, + 550 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 557, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 592 + ], + "type": "text", + "content": "We conduct ablation studies to validate the design choices made in the proposed framework. To ensure a fair comparison, we analyze the experimental results on SoccerNet V2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Importance of the proposed modules: Table 3 shows the contribution of the proposed ASTRM, Soft-IC loss, and ASAM components to achieve SOTA results. We observe a significant improvement in scores when replacing GSM [14] with the proposed ASTRM, highlighting the importance of effective feature extraction. SAM [9], utilized in the Spivak method, further enhances the results by a notable margin. Furthermore, we have illustrated how ASAM outperforms the scores achieved by SAM. Although ASAM and Soft-IC loss improves the overall scores by a small" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 328, + 489, + 542, + 580 + ], + "blocks": [ + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "lines": [ + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "type": "text", + "content": "Figure 4. Per-class score comparison on " + }, + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "type": "inline_equation", + "content": "\\delta = 0" + }, + { + "bbox": [ + 55, + 458, + 555, + 481 + ], + "type": "text", + "content": " setting in terms of mAP for Tennis, FineGym, FS-Comp, and FS-Perf dataset. For FineGym, we report results for only 6 out of 32 classes, while results for all classes in the other datasets are reported." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 489, + 542, + 580 + ], + "lines": [ + { + "bbox": [ + 328, + 489, + 542, + 580 + ], + "spans": [ + { + "bbox": [ + 328, + 489, + 542, + 580 + ], + "type": "image", + "image_path": "839dba5ad5e8504802bb00a446c93dbd06a3cdc7424ee79323f51a8a14495cfb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 587, + 548, + 598 + ], + "lines": [ + { + "bbox": [ + 319, + 587, + 548, + 598 + ], + "spans": [ + { + "bbox": [ + 319, + 587, + 548, + 598 + ], + "type": "text", + "content": "Figure 5. Analysis on our method with & without Soft-IC loss." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": "amount, Soft-IC loss plays an important role in improving the scores of minority classes. This is illustrated with scores of individual classes in Figure 5. As can be seen, Soft-IC loss improves the scores of classes with less number of samples without affecting other classes. This emphasizes the contributions of all components to achieve SOTA results." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Components of the proposed ASTRM: The proposed AS-TRM module uses 3 major components: local spatial, local" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3169" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 92, + 293, + 432 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 3. Ablation studies on various aspects of the proposed framework on SoccerNet V2 testing dataset in terms of mAP." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 92, + 293, + 432 + ], + "lines": [ + { + "bbox": [ + 56, + 92, + 293, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 92, + 293, + 432 + ], + "type": "table", + "html": "
MethodTightLoose
Analysis on proposed modules
Baseline: E2E-Spot (RegNet-Y 200MF + GSM)61.1973.25
RegNet-Y 200MF + ASTRM68.6375.57
RegNet-Y 200MF + ASTRM + SAM71.8977.50
RegNet-Y 200MF + ASTRM + ASAM72.6578.58
RegNet-Y 200MF + ASTRM + ASAM + Focal Loss70.3676.86
RegNet-Y 200MF + ASTRM + ASAM + IC Loss72.0578.30
RegNet-Y 200MF + ASTRM + ASAM + Soft-IC loss73.7479.11
Analysis on proposed ASTRM module
With local spatial only69.9176.39
With local temporal only70.4176.79
With global temporal only71.6577.06
With local spatial + local temporal71.0677.10
With local spatial + global temporal71.8277.85
With local temporal + global temporal71.5576.93
Proposed ASTRM module73.7479.11
Analysis on Different Temporal networks
Baseline with Bi-GRU61.1973.25
Proposed with Bi-GRU73.7479.11
with Deformable Attention67.4372.15
with Bi-GRU 2 Layers67.8573.03
with Transformer (L1H8)65.8970.76
with Transformer (L2H8)67.6072.35
with Bi-LSTM68.1573.35
with MSTCN68.7573.50
Analysis on Clip Length
Clip length = 6443.8252.09
Clip length = 100 (Used in baseline E2E-Spot)71.2076.73
Clip length = 128 (proposed)73.7479.11
Clip length = 19271.7477.02
", + "image_path": "fae8cecf3afa47fb811c975b1fcc567445e354404cd395f912e63468c77ed563.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 447, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 687 + ], + "type": "text", + "content": "temporal, and global temporal. Here, we highlight the importance of each component, and the corresponding findings are reported in Table 3. Notably, the score is at its lowest when utilizing only the spatial aspect of ASTRM; it fails to capture temporal information. Furthermore, the performance of local temporal is superior to using only the spatial aspect. Interestingly, we also observed that global temporal performs better than local temporal alone. This finding supports the notion that observing distant frames may be necessary to classify the current frame accurately. A similar pattern is seen when comparing local spatial + local temporal with local spatial + global temporal. Moreover, the inclusion of local spatial information leads to improved scores. Conversely, using only temporal information results in even lower scores, indicating the necessity of spatial information for understanding \"what is happening\" and temporal information for determining \"when it is happening.\" Overall, our observations highlight the need for spatial and local-global temporal information to identify events precisely and outperform other settings significantly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Analysis on Different Temporal modules In our proposed approach, we opted for the Bidirectional GRU (Bi-GRU) as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "content": "the long-range dependency module in the temporal block after conducting various experiments with different potential temporal networks. Table 3 presents the results obtained from the proposed approach using different temporal networks on the SoccerNet V2 [5] test set. The results indicate that the Bi-GRU, as a temporal network, performs favorably compared to other temporal networks. Even more advanced temporal networks, such as Bi-LSTM, Transformer, Deformable Attention [18] and MSTCN [8], could not surpass the scores achieved by the Bi-GRU temporal network. The difference in scores is visible across all tolerances. It is also observed that the proposed method generates significantly better results than the baseline E2E-Spot method, which also uses Bi-GRU as the temporal block. For instance, the proposed method achieves " + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "inline_equation", + "content": "12.55\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "inline_equation", + "content": "5.86\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "content": " improvements over the baseline model in the tight and loose settings of the SoccerNet V2 dataset, respectively. Similarly, there is an improvement of " + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "inline_equation", + "content": "15.67\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\delta = 0" + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "content": " setting compared to the baseline on the Tennis dataset. This significant improvement is attributed to the proposed ASTRM module and the proposed Soft-IC loss function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 326, + 556, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 556, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 556, + 458 + ], + "type": "text", + "content": "Clip Length: Due to memory and computation constraints, it's not feasible to process all the frames simultaneously. Therefore, we process a fixed number of consecutive frames at a time, which we refer to as the \"clip length\". The value of the clip length affects the scores because it determines how many frames the network considers at once. While we might want to maximize the clip length, we need to find a balance between performance and clip length. The results in Table 3 support this idea. If the clip length is too short, it significantly impacts the results, while making it too long also has a negative effect on the results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 475, + 493, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 493, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 493, + 488 + ], + "type": "text", + "content": "5. Conclusion & Future Directions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "This paper proposes an end-to-end network for precise spotting of events in sports videos. It effectively tackles the long-range dependency and class imbalance issues and outperforms existing methods. To this extent, we introduce a new ASTRM module for enhancing spatio-temporal features. Further, to address the class imbalance issue, we propose Soft-IC loss, which enforces compact features and improves class separability. We further employed ASAM during training to ensure the model performs well on unseen data. While our method has advanced the state-of-the-art, especially in the more challenging cases with a relatively simple design, there is further scope for improvement on the two fronts we tried to address. Certain larger networks can improve the performance further but at the cost of computation. On the other hand, the spatial and temporal resolution at which the data is being processed will likely have an impact. The precise spotting may become easier with increased resolution, but it increases the processing burden." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3170" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Humam Alwassel, Silvio Giancola, and Bernard Ghanem. Tsp: Temporally-sensitive pretraining of video encoders for localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3173-3183, 2021. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "type": "text", + "content": "[2] Mengqi Cao, Min Yang, Guozhen Zhang, Xiaotian Li, Yilu Wu, Gangshan Wu, and Limin Wang. Spotformer: A transformer-based framework for precise soccer action spotting. In 2022 IEEE 24th International Workshop on Multimedia Signal Processing (MMSP), pages 1-6. IEEE, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 216, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 295, + 258 + ], + "type": "text", + "content": "[3] Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 262, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 282 + ], + "type": "text", + "content": "[4] Charika De Alvis et al. A survey of deep long-tail classification advancements. arXiv:2404.15593, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 296, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 372 + ], + "type": "text", + "content": "[5] Adrien Deliège, Anthony Cioppa, Silvio Giancola, Meisam J. Seikavandi, Jacob V. Dueholm, Kamal Nasrollahi, Bernard Ghanem, Thomas B. Moeslund, and Marc Van Droogenbroeck. Soccertnet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 4508-4519, 2021. 2, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 375, + 295, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 295, + 440 + ], + "type": "text", + "content": "[6] Julien Denize, Mykola Liashuha, Jaonary Rabarisoa, Astrid Orcesi, and Romain Hérault. Comedian: Self-supervised learning and knowledge distillation for action spotting using transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 530-540, 2024. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 443, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 295, + 509 + ], + "type": "text", + "content": "[7] Moritz Einfalt, Charles Dampeyrou, Dan Zecha, and Rainer Lienhart. Frame-level event detection in athletics videos with pose-based convolutional sequence networks. In Proceedings Proceedings of the 2nd International Workshop on Multimedia Content Analysis in Sports, page 42-50, New York, NY, USA, 2019. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 510, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 295, + 554 + ], + "type": "text", + "content": "[8] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 556, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 295, + 600 + ], + "type": "text", + "content": "[9] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, 2021. 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "text", + "content": "[10] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccertnet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[11] Jiaming Han, Yuqiang Ren, Jian Ding, Xingjia Pan, Ke Yan, and Gui-Song Xia. Expanding low-density latent regions for open-set object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9591-9600, 2022. 4, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[12] James Hong, Matthew Fisher, Michael Gharbi, and Kayvon Fatahalian. Video pose distillation for few-shot, fine-grained sports action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9254-9263, 2021. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "type": "text", + "content": "[13] James Hong, Matthew Fisher, Michael Gharbi, and Kayvon Fatahalian. Video pose distillation for few-shot, fine-grained sports action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9254-9263, 2021. 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 186, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 554, + 229 + ], + "type": "text", + "content": "[14] James Hong, Haotian Zhang, Michael Gharbi, Matthew Fisher, and Kayvon Fatahalian. Spotting temporally precise, fine-grained events in video. In European Conference on Computer Vision, pages 33-51. Springer, 2022. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 231, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 554, + 285 + ], + "type": "text", + "content": "[15] Jungmin Kwon, Jeongseop Kim, Hyunseo Park, and In Kwon Choi. Asam: Adaptive sharpness-aware minimization for scale-invariant learning of deep neural networks. In Proceedings of the 38th International Conference on Machine Learning, pages 5905-5914. PMLR, 2021. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 286, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 329 + ], + "type": "text", + "content": "[16] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(2):318-327, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 331, + 555, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 555, + 375 + ], + "type": "text", + "content": "[17] Xiaolong Liu, Song Bai, and Xiang Bai. An empirical study of end-to-end temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20010-20019, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 376, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 420 + ], + "type": "text", + "content": "[18] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31:5427-5441, 2022. 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 422, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 443 + ], + "type": "text", + "content": "[19] Zhaoyang Liu et al. Tam: Temporal adaptive module for video recognition. In ICCV, pages 13708-13718, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 445, + 554, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 554, + 476 + ], + "type": "text", + "content": "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 478, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 554, + 521 + ], + "type": "text", + "content": "[21] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollar. Designing network design spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 523, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 554, + 567 + ], + "type": "text", + "content": "[22] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 568, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 554, + 612 + ], + "type": "text", + "content": "[23] Joao VB Soares and Avijit Shah. Action spotting using dense detection anchors revisited: Submission to the soccer-net challenge 2022. arXiv preprint arXiv:2206.07846, 2022. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 613, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 554, + 667 + ], + "type": "text", + "content": "[24] Joao VB Soares, Avijit Shah, and Topojoy Biswas. Temporally precise action spotting in soccer videos using dense detection anchors. In 2022 IEEE International Conference on Image Processing (ICIP), pages 2796-2800. IEEE, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[25] Swathikiran Sudhakaran, Sergio Escalera, and Oswald Lanz. Gate-shift networks for video action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3171" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 542 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[26] Matteo Tomei, Lorenzo Baraldi, Simone Calderara, Simone Bronzin, and Rita Cucchiara. Rms-net: Regression and masking for soccer event spotting. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7699-7706. IEEE, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[27] Kim Hoang Tran, Phuc Vuong Do, Ngoc Quoc Ly, and Ngan Le. Unifying global and local scene entities modelling for precise action spotting. arXiv preprint arXiv:2404.09951 (Accepted for IJCNN 2024), 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[28] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "text", + "content": "[29] Artur Xarles, Sergio Escalera, Thomas B Moeslund, and Albert Clapés. Astra: An action spotting transformer for soccer videos. In Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports, pages 93-102, 2023. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 339 + ], + "type": "text", + "content": "[30] Artur Xarles, Sergio Escalera, Thomas B. Moeslund, and Albert Clapés. T-deed: Temporal-discriminability enhancer encoder-decoder for precise event spotting in sports videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3410-3419, 2024. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "type": "text", + "content": "[31] Huiran Yan et al. Emphasizing feature inter-class separability for improving highly imbalanced overlapped data classification. Knowledge-Based Systems, 276:110745, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 407 + ], + "type": "text", + "content": "[32] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. 32nd British Machine Vision Conference, BMVC 2021, 2021. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 452 + ], + "type": "text", + "content": "[33] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 497 + ], + "type": "text", + "content": "[34] Haotian Zhang, Cristobal Sciutto, Maneesh Agrawala, and Kayvon Fatahalian. Vid2player: Controllable video sprites that behave and appear like professional tennis players. ACM Trans. Graph., 40(3), 2021. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "type": "text", + "content": "[35] Xin Zhou, Le Kang, Zhiyu Cheng, Bo He, and Jingyu Xin. Feature combination meets attention: Baidu soccer embeddings and transformer based temporal detection. arXiv preprint arXiv:2106.14447, 2021. 2, 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3172" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file